diff --git a/.gitattributes b/.gitattributes index 4019a9f4d26fd594fa0246f1ece7b970cd0a71af..a36998ba35cc11e393a28434fe2c2c7456c5e34a 100644 --- a/.gitattributes +++ b/.gitattributes @@ -3301,3 +3301,67 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text 2024/EASE-DETR_[[:space:]]Easing[[:space:]]the[[:space:]]Competition[[:space:]]among[[:space:]]Object[[:space:]]Queries/9ff0d4e0-f729-41a1-9284-a3462c62983a_origin.pdf filter=lfs diff=lfs merge=lfs -text 2024/ECLIPSE_[[:space:]]A[[:space:]]Resource-Efficient[[:space:]]Text-to-Image[[:space:]]Prior[[:space:]]for[[:space:]]Image[[:space:]]Generations/83a8980c-9503-4c5f-99c0-a6eb3b04e979_origin.pdf filter=lfs diff=lfs merge=lfs -text 2024/ECLIPSE_[[:space:]]Efficient[[:space:]]Continual[[:space:]]Learning[[:space:]]in[[:space:]]Panoptic[[:space:]]Segmentation[[:space:]]with[[:space:]]Visual[[:space:]]Prompt[[:space:]]Tuning/9f39aafa-e0ef-4ad0-a765-ad142df3abb7_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/ECoDepth_[[:space:]]Effective[[:space:]]Conditioning[[:space:]]of[[:space:]]Diffusion[[:space:]]Models[[:space:]]for[[:space:]]Monocular[[:space:]]Depth[[:space:]]Estimation/7e3c5935-f184-4aa8-a48b-682e7477329f_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/EFHQ_[[:space:]]Multi-purpose[[:space:]]ExtremePose-Face-HQ[[:space:]]dataset/65a10ed9-cef6-42e8-a141-8586b125aea4_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/EFormer_[[:space:]]Enhanced[[:space:]]Transformer[[:space:]]towards[[:space:]]Semantic-Contour[[:space:]]Features[[:space:]]of[[:space:]]Foreground[[:space:]]for[[:space:]]Portraits[[:space:]]Matting/652e59cf-51c6-401e-a44e-7a8d54a66599_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/EGTR_[[:space:]]Extracting[[:space:]]Graph[[:space:]]from[[:space:]]Transformer[[:space:]]for[[:space:]]Scene[[:space:]]Graph[[:space:]]Generation/174c2199-8aa8-4e45-8d2e-205eb2e2e099_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/EMAGE_[[:space:]]Towards[[:space:]]Unified[[:space:]]Holistic[[:space:]]Co-Speech[[:space:]]Gesture[[:space:]]Generation[[:space:]]via[[:space:]]Expressive[[:space:]]Masked[[:space:]]Audio[[:space:]]Gesture[[:space:]]Modeling/f8a85da7-9221-4a37-baf0-e5a3b46705e3_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/EMCAD_[[:space:]]Efficient[[:space:]]Multi-scale[[:space:]]Convolutional[[:space:]]Attention[[:space:]]Decoding[[:space:]]for[[:space:]]Medical[[:space:]]Image[[:space:]]Segmentation/32047b53-ccf5-41e9-bc53-4e712089f958_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/EMOPortraits_[[:space:]]Emotion-enhanced[[:space:]]Multimodal[[:space:]]One-shot[[:space:]]Head[[:space:]]Avatars/ab68d1cd-73a3-4a4c-9896-716aeaaff436_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/ERMVP_[[:space:]]Communication-Efficient[[:space:]]and[[:space:]]Collaboration-Robust[[:space:]]Multi-Vehicle[[:space:]]Perception[[:space:]]in[[:space:]]Challenging[[:space:]]Environments/197e9226-8724-4a9f-ac9a-1556459fa452_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/ES3_[[:space:]]Evolving[[:space:]]Self-Supervised[[:space:]]Learning[[:space:]]of[[:space:]]Robust[[:space:]]Audio-Visual[[:space:]]Speech[[:space:]]Representations/09b7366b-807f-4dfb-a9c8-45edf9472928_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/ESCAPE_[[:space:]]Encoding[[:space:]]Super-keypoints[[:space:]]for[[:space:]]Category-Agnostic[[:space:]]Pose[[:space:]]Estimation/64b3bbb9-9580-410a-adfe-c8a2bd5b8bca_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/ESR-NeRF_[[:space:]]Emissive[[:space:]]Source[[:space:]]Reconstruction[[:space:]]Using[[:space:]]LDR[[:space:]]Multi-view[[:space:]]Images/2c6d481f-7a23-410c-a62a-93255bbc656e_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/EVCap_[[:space:]]Retrieval-Augmented[[:space:]]Image[[:space:]]Captioning[[:space:]]with[[:space:]]External[[:space:]]Visual-Name[[:space:]]Memory[[:space:]]for[[:space:]]Open-World[[:space:]]Comprehension/003cc44f-e449-40f1-aec0-baf1c6235010_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/EVS-assisted[[:space:]]Joint[[:space:]]Deblurring[[:space:]]Rolling-Shutter[[:space:]]Correction[[:space:]]and[[:space:]]Video[[:space:]]Frame[[:space:]]Interpolation[[:space:]]through[[:space:]]Sensor[[:space:]]Inverse[[:space:]]Modeling/2babbe27-3090-44dc-8f9d-2d16b50cc5ab_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Each[[:space:]]Test[[:space:]]Image[[:space:]]Deserves[[:space:]]A[[:space:]]Specific[[:space:]]Prompt_[[:space:]]Continual[[:space:]]Test-Time[[:space:]]Adaptation[[:space:]]for[[:space:]]2D[[:space:]]Medical[[:space:]]Image[[:space:]]Segmentation/b547c627-f0bf-4b29-9f5d-10851e261fb1_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/EarthLoc_[[:space:]]Astronaut[[:space:]]Photography[[:space:]]Localization[[:space:]]by[[:space:]]Indexing[[:space:]]Earth[[:space:]]from[[:space:]]Space/09848c3c-ff6b-4e53-9236-ab0f624d8495_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/EasyDrag_[[:space:]]Efficient[[:space:]]Point-based[[:space:]]Manipulation[[:space:]]on[[:space:]]Diffusion[[:space:]]Models/3c97514b-8a67-4bdc-8795-3e7cfc02ceca_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Eclipse_[[:space:]]Disambiguating[[:space:]]Illumination[[:space:]]and[[:space:]]Materials[[:space:]]using[[:space:]]Unintended[[:space:]]Shadows/a9a99e7b-5bee-4200-8036-13e74c55e36e_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Edge-Aware[[:space:]]3D[[:space:]]Instance[[:space:]]Segmentation[[:space:]]Network[[:space:]]with[[:space:]]Intelligent[[:space:]]Semantic[[:space:]]Prior/ef85b90b-3f1b-4322-ac5a-c5d6ae60802c_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Edit[[:space:]]One[[:space:]]for[[:space:]]All_[[:space:]]Interactive[[:space:]]Batch[[:space:]]Image[[:space:]]Editing/d0962876-52e1-4fef-ac92-9ddbf4fa5230_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/EditGuard_[[:space:]]Versatile[[:space:]]Image[[:space:]]Watermarking[[:space:]]for[[:space:]]Tamper[[:space:]]Localization[[:space:]]and[[:space:]]Copyright[[:space:]]Protection/43cb9c2b-dd1a-4670-9150-d4345483687e_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Editable[[:space:]]Scene[[:space:]]Simulation[[:space:]]for[[:space:]]Autonomous[[:space:]]Driving[[:space:]]via[[:space:]]Collaborative[[:space:]]LLM-Agents/49eb43d9-92f1-4849-a7c5-3da61380ffc6_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Effective[[:space:]]Video[[:space:]]Mirror[[:space:]]Detection[[:space:]]with[[:space:]]Inconsistent[[:space:]]Motion[[:space:]]Cues/423d0156-aefb-4c1d-9839-9825d419ff81_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Efficient[[:space:]]3D[[:space:]]Implicit[[:space:]]Head[[:space:]]Avatar[[:space:]]with[[:space:]]Mesh-anchored[[:space:]]Hash[[:space:]]Table[[:space:]]Blendshapes/afaaa8c5-698e-4fcb-9bde-21aa2469de9a_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Efficient[[:space:]]Dataset[[:space:]]Distillation[[:space:]]via[[:space:]]Minimax[[:space:]]Diffusion/75146b44-b76a-42b0-8456-4c37446317b7_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Efficient[[:space:]]Deformable[[:space:]]ConvNets_[[:space:]]Rethinking[[:space:]]Dynamic[[:space:]]and[[:space:]]Sparse[[:space:]]Operator[[:space:]]for[[:space:]]Vision[[:space:]]Applications/d839f1b7-f6aa-4faa-bb5a-7c661a9aa4e5_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Efficient[[:space:]]Detection[[:space:]]of[[:space:]]Long[[:space:]]Consistent[[:space:]]Cycles[[:space:]]and[[:space:]]its[[:space:]]Application[[:space:]]to[[:space:]]Distributed[[:space:]]Synchronization/c0dd6b15-09a8-4eba-bcfb-24b8083dd375_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Efficient[[:space:]]Hyperparameter[[:space:]]Optimization[[:space:]]with[[:space:]]Adaptive[[:space:]]Fidelity[[:space:]]Identification/fa155004-89c9-4599-9ab7-9a5d6edc5fbc_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Efficient[[:space:]]LoFTR_[[:space:]]Semi-Dense[[:space:]]Local[[:space:]]Feature[[:space:]]Matching[[:space:]]with[[:space:]]Sparse-Like[[:space:]]Speed/14a55345-a1a3-4ac4-a5a2-7c153c54b832_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Efficient[[:space:]]Meshflow[[:space:]]and[[:space:]]Optical[[:space:]]Flow[[:space:]]Estimation[[:space:]]from[[:space:]]Event[[:space:]]Cameras/141a2ee1-d1ef-416d-8c61-bc462b1074b9_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Efficient[[:space:]]Model[[:space:]]Stealing[[:space:]]Defense[[:space:]]with[[:space:]]Noise[[:space:]]Transition[[:space:]]Matrix/5181a0c8-f9af-47c9-b829-4e394817b1e7_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Efficient[[:space:]]Multi-scale[[:space:]]Network[[:space:]]with[[:space:]]Learnable[[:space:]]Discrete[[:space:]]Wavelet[[:space:]]Transform[[:space:]]for[[:space:]]Blind[[:space:]]Motion[[:space:]]Deblurring/4af0ea1e-f85d-4e0b-ab29-04d268e0386c_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Efficient[[:space:]]Multitask[[:space:]]Dense[[:space:]]Predictor[[:space:]]via[[:space:]]Binarization/b91bc222-326c-4762-aabe-433ec8965e13_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Efficient[[:space:]]Privacy-Preserving[[:space:]]Visual[[:space:]]Localization[[:space:]]Using[[:space:]]3D[[:space:]]Ray[[:space:]]Clouds/61345f4c-613d-4e27-91d1-09fcd6a52a24_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Efficient[[:space:]]Scene[[:space:]]Recovery[[:space:]]Using[[:space:]]Luminous[[:space:]]Flux[[:space:]]Prior/0327d2f4-6c3a-4e74-a5a8-041807da336a_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Efficient[[:space:]]Solution[[:space:]]of[[:space:]]Point-Line[[:space:]]Absolute[[:space:]]Pose/8561d28f-e735-46cc-9618-c6c206bdafc2_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Efficient[[:space:]]Stitchable[[:space:]]Task[[:space:]]Adaptation/03126c3f-71ba-4532-921c-1c95d792e515_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Efficient[[:space:]]Test-Time[[:space:]]Adaptation[[:space:]]of[[:space:]]Vision-Language[[:space:]]Models/f86a8c24-7089-4003-88ca-fa1a5c70fe2d_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Efficient[[:space:]]Vision-Language[[:space:]]Pre-training[[:space:]]by[[:space:]]Cluster[[:space:]]Masking/24f9f648-410c-4f99-8615-dc6d0380efad_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Efficient[[:space:]]and[[:space:]]Effective[[:space:]]Weakly-Supervised[[:space:]]Action[[:space:]]Segmentation[[:space:]]via[[:space:]]Action-Transition-Aware[[:space:]]Boundary[[:space:]]Alignment/1ae3fdc0-77c3-4dc9-87e2-e3b0568d2efa_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/EfficientDreamer_[[:space:]]High-Fidelity[[:space:]]and[[:space:]]Robust[[:space:]]3D[[:space:]]Creation[[:space:]]via[[:space:]]Orthogonal-view[[:space:]]Diffusion[[:space:]]Priors/6a261e68-4855-488d-9a23-298bfd3be3c9_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/EfficientSAM_[[:space:]]Leveraged[[:space:]]Masked[[:space:]]Image[[:space:]]Pretraining[[:space:]]for[[:space:]]Efficient[[:space:]]Segment[[:space:]]Anything/f1e649e9-e5d4-48b3-8018-af1f58065ff2_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Efficiently[[:space:]]Assemble[[:space:]]Normalization[[:space:]]Layers[[:space:]]and[[:space:]]Regularization[[:space:]]for[[:space:]]Federated[[:space:]]Domain[[:space:]]Generalization/16ce0297-575e-48b3-acb7-4a9aa0af4fe6_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Ego-Exo4D_[[:space:]]Understanding[[:space:]]Skilled[[:space:]]Human[[:space:]]Activity[[:space:]]from[[:space:]]First-[[:space:]]and[[:space:]]Third-Person[[:space:]]Perspectives/742959dc-4498-434e-af9a-b1cc7f6e5bd2_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/EgoExoLearn_[[:space:]]A[[:space:]]Dataset[[:space:]]for[[:space:]]Bridging[[:space:]]Asynchronous[[:space:]]Ego-[[:space:]]and[[:space:]]Exo-centric[[:space:]]View[[:space:]]of[[:space:]]Procedural[[:space:]]Activities[[:space:]]in[[:space:]]Real[[:space:]]World/134e2089-370f-4ad6-9686-84a1d7af0ed4_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/EgoGen_[[:space:]]An[[:space:]]Egocentric[[:space:]]Synthetic[[:space:]]Data[[:space:]]Generator/717ad579-fa9d-4ce5-996d-77c046c2cdbf_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/EgoThink_[[:space:]]Evaluating[[:space:]]First-Person[[:space:]]Perspective[[:space:]]Thinking[[:space:]]Capability[[:space:]]of[[:space:]]Vision-Language[[:space:]]Models/2ff0d147-b907-475e-8aa3-5584b71f3014_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Egocentric[[:space:]]Whole-Body[[:space:]]Motion[[:space:]]Capture[[:space:]]with[[:space:]]FisheyeViT[[:space:]]and[[:space:]]Diffusion-Based[[:space:]]Motion[[:space:]]Refinement/90ece1d3-5ea4-4b56-9fe8-5a7d9a524eef_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/ElasticDiffusion_[[:space:]]Training-free[[:space:]]Arbitrary[[:space:]]Size[[:space:]]Image[[:space:]]Generation[[:space:]]through[[:space:]]Global-Local[[:space:]]Content[[:space:]]Separation/de40da54-b1b7-40fa-9aec-0f4f1bc4f406_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Elite360D_[[:space:]]Towards[[:space:]]Efficient[[:space:]]360[[:space:]]Depth[[:space:]]Estimation[[:space:]]via[[:space:]]Semantic-[[:space:]]and[[:space:]]Distance-Aware[[:space:]]Bi-Projection[[:space:]]Fusion/b933bca2-9ff4-42fc-9bd7-be64c39e1936_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Embodied[[:space:]]Multi-Modal[[:space:]]Agent[[:space:]]trained[[:space:]]by[[:space:]]an[[:space:]]LLM[[:space:]]from[[:space:]]a[[:space:]]Parallel[[:space:]]TextWorld/911fda92-445d-4ff5-8ef4-3ff0a7daf4bf_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/EmbodiedScan_[[:space:]]A[[:space:]]Holistic[[:space:]]Multi-Modal[[:space:]]3D[[:space:]]Perception[[:space:]]Suite[[:space:]]Towards[[:space:]]Embodied[[:space:]]AI/ab5e0a2a-2a5f-41d4-a08d-01fd4564b83b_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Embracing[[:space:]]Unimodal[[:space:]]Aleatoric[[:space:]]Uncertainty[[:space:]]for[[:space:]]Robust[[:space:]]Multimodal[[:space:]]Fusion/b864c4f0-5e7f-441f-a2a0-028202223b38_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Emergent[[:space:]]Open-Vocabulary[[:space:]]Semantic[[:space:]]Segmentation[[:space:]]from[[:space:]]Off-the-shelf[[:space:]]Vision-Language[[:space:]]Models/bf254b56-9ba7-4039-9f52-56365463bb00_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/EmoGen_[[:space:]]Emotional[[:space:]]Image[[:space:]]Content[[:space:]]Generation[[:space:]]with[[:space:]]Text-to-Image[[:space:]]Diffusion[[:space:]]Models/a0b8035c-47d5-4f15-8dd7-134ed07d94b6_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/EmoVIT_[[:space:]]Revolutionizing[[:space:]]Emotion[[:space:]]Insights[[:space:]]with[[:space:]]Visual[[:space:]]Instruction[[:space:]]Tuning/337f29a8-74df-4960-89a4-6eff2797f1cc_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Emotional[[:space:]]Speech-driven[[:space:]]3D[[:space:]]Body[[:space:]]Animation[[:space:]]via[[:space:]]Disentangled[[:space:]]Latent[[:space:]]Diffusion/8fe0dd5b-5118-4e0a-ac09-f70da9816cdc_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Empowering[[:space:]]Resampling[[:space:]]Operation[[:space:]]for[[:space:]]Ultra-High-Definition[[:space:]]Image[[:space:]]Enhancement[[:space:]]with[[:space:]]Model-Aware[[:space:]]Guidance/5e31bf7f-5e48-4030-90cc-b526f724e932_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Emu[[:space:]]Edit_[[:space:]]Precise[[:space:]]Image[[:space:]]Editing[[:space:]]via[[:space:]]Recognition[[:space:]]and[[:space:]]Generation[[:space:]]Tasks/b6371d78-25ca-4b11-84b1-a566ed9d62f2_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/En3D_[[:space:]]An[[:space:]]Enhanced[[:space:]]Generative[[:space:]]Model[[:space:]]for[[:space:]]Sculpting[[:space:]]3D[[:space:]]Humans[[:space:]]from[[:space:]]2D[[:space:]]Synthetic[[:space:]]Data/cb7c8896-3bc3-4bc4-a3cb-b2aedebff579_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/End-to-End[[:space:]]Spatio-Temporal[[:space:]]Action[[:space:]]Localisation[[:space:]]with[[:space:]]Video[[:space:]]Transformers/62e0dda5-0752-47db-aad5-cb1b83945b43_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/End-to-End[[:space:]]Temporal[[:space:]]Action[[:space:]]Detection[[:space:]]with[[:space:]]1B[[:space:]]Parameters[[:space:]]Across[[:space:]]1000[[:space:]]Frames/aed9af76-c260-41b9-be21-a2e11ea6ab0b_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Endow[[:space:]]SAM[[:space:]]with[[:space:]]Keen[[:space:]]Eyes_[[:space:]]Temporal-spatial[[:space:]]Prompt[[:space:]]Learning[[:space:]]for[[:space:]]Video[[:space:]]Camouflaged[[:space:]]Object[[:space:]]Detection/e8cb1a37-360c-44d3-92de-d83afed15e31_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Enhance[[:space:]]Image[[:space:]]Classification[[:space:]]via[[:space:]]Inter-Class[[:space:]]Image[[:space:]]Mixup[[:space:]]with[[:space:]]Diffusion[[:space:]]Model/d24eac86-8eb2-4146-81d8-2b9455b8e75f_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Enhanced[[:space:]]Motion-Text[[:space:]]Alignment[[:space:]]for[[:space:]]Image-to-Video[[:space:]]Transfer[[:space:]]Learning/1b431374-7f10-4d10-9e33-68850035e408_origin.pdf filter=lfs diff=lfs merge=lfs -text diff --git a/2024/ECoDepth_ Effective Conditioning of Diffusion Models for Monocular Depth Estimation/7e3c5935-f184-4aa8-a48b-682e7477329f_content_list.json b/2024/ECoDepth_ Effective Conditioning of Diffusion Models for Monocular Depth Estimation/7e3c5935-f184-4aa8-a48b-682e7477329f_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..9327fb0c7e84c9eebae99ff6821d922d16ad37de --- /dev/null +++ b/2024/ECoDepth_ Effective Conditioning of Diffusion Models for Monocular Depth Estimation/7e3c5935-f184-4aa8-a48b-682e7477329f_content_list.json @@ -0,0 +1,1329 @@ +[ + { + "type": "text", + "text": "ECoDepth: Effective Conditioning of Diffusion Models for Monocular Depth Estimation", + "text_level": 1, + "bbox": [ + 207, + 130, + 759, + 175 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Suraj Patni*", + "bbox": [ + 222, + 203, + 320, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Aradhye Agarwal*", + "bbox": [ + 408, + 204, + 557, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Chetan Arora", + "bbox": [ + 635, + 204, + 745, + 219 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Indian Institute of Technology Delhi", + "bbox": [ + 339, + 222, + 629, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "https://ecodepth-iitd.github.io", + "bbox": [ + 297, + 239, + 666, + 256 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 291, + 312, + 306 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In the absence of parallax cues, a learning based single image depth estimation (SIDE) model relies heavily on shading and contextual cues in the image. While this simplicity is attractive, it is necessary to train such models on large and varied datasets, which are difficult to capture. It has been shown that using embeddings from pretrained foundational models, such as CLIP, improves zero shot transfer in several applications. Taking inspiration from this, in our paper we explore the use of global image priors generated from a pre-trained ViT model to provide more detailed contextual information. We argue that the embedding vector from a ViT model, pre-trained on a large dataset, captures greater relevant information for SIDE than the usual route of generating pseudo image captions, followed by CLIP based text embeddings. Based on this idea, we propose a new SIDE model using a diffusion backbone which is conditioned on ViT embeddings. Our proposed design establishes a new state-of-the-art (SOTA) for SIDE on NYU Depth v2 dataset, achieving Abs Rel error of 0.059(14% improvement) compared to 0.069 by the current SOTA (VPD). And on KITTI dataset, achieving Sq Rel error of 0.139 (2% improvement) compared to 0.142 by the current SOTA (GED). For zero shot transfer with a model trained on NYU Depth v2, we report mean relative improvement of (20%, 23%, 81%, 25%) over NeWCRF on (Sun-RGBD, iBims1, DIODE, HyperSim) datasets, compared to (16%, 18%, 45%, 9%) by ZoEDepth. The code is available in our project page.", + "bbox": [ + 75, + 323, + 473, + 748 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 775, + 207, + 790 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Single Image Depth Estimation (SIDE) is the task of predicting per pixel depth using a single RGB image from a monocular camera. It is a fundamental problem in computer vision with applications in several domains, viz robotics, autonomous driving, and augmented reality. The problem is", + "bbox": [ + 75, + 801, + 468, + 878 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/f7274f44c4d22bc9eae8a19f69aa92c1c20646d54843a02397296650e2bfc12a.jpg", + "image_caption": [ + "(a) Sun-RGBD [42]" + ], + "image_footnote": [], + "bbox": [ + 500, + 290, + 696, + 330 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/214cfea1c3468673bbbfb8b1f65db15a952fafc13969355b377513969847d17f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 697, + 290, + 890, + 330 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/ff8d25236e5638817716f97d36604908ba6a66250a0304d48025d379b65c2095.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 501, + 344, + 694, + 383 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/f56327bab79fcd3f27c08fbef298e0b0e23fd81917afe234efe190f0d7e73e38.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 697, + 344, + 890, + 383 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/acfee1ae7e238a4d60c7914c77d3468b68bc4fb10c24754766404c5d2a5dd4a9.jpg", + "image_caption": [ + "(b) iBims1 [19]" + ], + "image_footnote": [], + "bbox": [ + 501, + 398, + 696, + 439 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/b6c9dc2b7d14f841f908bf571798f99779b924830b846d5fe98305823fd3225f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 697, + 398, + 890, + 439 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/5a9b6186ef62db411c91c5843145d756b45f98d60f13a0e2aa97b4f3d446142e.jpg", + "image_caption": [ + "(c) DIODE [49]", + "(d) HyperSim [35]", + "Figure 1. Qualitative results across four different datasets, demonstrating the zero-shot performance of our model trained only on the NYU Depth v2 dataset. Corresponding quantitative results are presented in Table 3. The first column displays RGB images, the second column depicts ground truth depth, and the third column showcases our model's predicted depths. Additional images for each dataset are available in the Supplementary Material." + ], + "image_footnote": [], + "bbox": [ + 501, + 452, + 696, + 493 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/8d5070d4cbc6c4ebcfef4196245823ef4f87ec26324f3274b07d2d886dd1ddc2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 697, + 452, + 890, + 493 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "typically formulated in two flavors: metric depth estimation (MDE), and relative depth estimation (RDE). As the names suggest, MDE deals with estimation in physical units such as meters, whereas, RDE techniques focus on relative depth only, and require a per-image affine transformation as post-processing to convert to a physical unit.", + "bbox": [ + 496, + 642, + 890, + 734 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Conventional geometric techniques for depth estimation typically rely on feature correspondence, parallax, and triangulation from two or more views. However, the problem becomes ill-posed for estimation from a single view. Intuitively, depth map is a 3D representation of the scene, whereas an RGB image is a 2D projection of the scene. Hence, it cannot be uniquely determined from a single RGB image. Therefore, learning-based SIDE models rely on visual cues like 'shape from shading' and other contextual priors for per-pixel depth prediction.", + "bbox": [ + 496, + 734, + 890, + 885 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "A data-driven approach makes the SIDE pipeline sim", + "bbox": [ + 517, + 885, + 890, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Equal contribution.", + "bbox": [ + 94, + 886, + 205, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "28285", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "pler, but also makes the learnt model dependent on the quality of training data. It has been observed that such models overfit on a particular training distribution/domain and fail to generalize on unseen data. This is especially true for MDE models when the range of depth in the training dataset is limited. Hence, training on multiple datasets, with wide variations in depth ranges has been proposed [4, 34].", + "bbox": [ + 75, + 90, + 470, + 196 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "On the other hand, development of large foundational models (LFMs) in recent years has altered the preferred design approach for many computer vision problems. These huge models are trained using extensive datasets of unlabeled images and learning objectives that are agnostic to specific tasks. The learnt embeddings from such pre-trained models have been shown to help generalization and zero-shot transfer in many applications. We are aware of at least two works for MDE in SIDE problem that have appeared in the last few months and make use of such foundational models. VPD [54] uses a text-to-image diffusion model pretrained on LAION-400M [39] dataset having large-scale image-text pairs as the backbone. The model prompts the denoising UNet with textual inputs to make the visual contents interact with the text prompts. Since the problem formulation doesn't include text description as the input, the model generates simple descriptions such as 'A photo of a {scene name}' based on the scene label given in the NYU Depth v2 dataset. Another work, TADP [20] improves upon VPD. Instead of simple image descriptions, TADP uses BLIP-2 [23] to generate image captions, and then uses CLIP [31] embeddings of the pseudo-caption to condition the diffusion model.", + "bbox": [ + 75, + 200, + 473, + 547 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We view the above works as providing robust semantic context to LFMs for the actual task at hand, which helps in visual recognition in general, as well as SIDE. While we do agree with the broad motivations of these works, the question that we ask is if pseudo-captions are the most effective way to provide the semantic context. Textual descriptions of an image typically focus on large salient objects and emphasize on their relationships. On the other hand, large vision models for image classification, typically contain representation for even smaller objects present in the scene. Even when a single object is present in the scene, the representations typically capture uncertainties and ambiguities inherent in the scene. Hence, we posit that using embeddings from a transformer model, pretrained on a large image dataset unrelated to the SIDE task, captures more relevant information, and is a better alternative to using pseudocaptions' CLIP embedding. Since diffusion-based models have shown their superiority for the dense prediction tasks in recent works [8, 16, 20, 38, 54]. Hence, we propose a diffusion backbone for our model, along with a novel CIDE module which employs ViT [7] to extract semantic context embeddings. These embeddings are subsequently utilized to condition the diffusion backbone.", + "bbox": [ + 75, + 553, + 472, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Contributions. The key contributions of this work are:", + "bbox": [ + 500, + 90, + 864, + 104 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "(1) We propose a new model for MDE in a SIDE task. The proposed model uses a conditional diffusion architecture with the semantic context being supplied through embeddings generated using a ViT model. Achieving a new state-of-the-art (SOTA) performance, our method outperforms existing approaches on benchmark datasets, including the NYU Depth v2 indoor and KITTI outdoor datasets. Notably, we report a significant improvement of $14\\%$ in absolute relative error, achieving 0.059 compared to the current SOTA (VPD) performance of 0.069 on NYU Depth v2. And we report an improvement of $2\\%$ in square relative error, achieving 0.139 compared to the current SOTA (GED) performance of 0.142 on KITTI.", + "(2) We show, qualitatively as well as quantitatively, that using ViT embeddings to provide semantic context is a better alternative to generating pseudo captions and then using its CLIP embeddings to condition a SIDE model. In contrast to TADP [20], which uses pseudo captions, but only achieves a RMSE of 0.225 on NYU Depth v2, we report a lower, and a new SOTA, error of 0.218 (VPD [54] achieves 0.254).", + "(3) We show that providing ViT conditioning, helps our model perform better in a zero shot transfer task. ZoEDepth [4], the current SOTA for zero-shot transfer, reports an improvement of (16%, 18%, 45%, 9%) over NeWCRF on (Sun-RGBD, iBims1, DIODE, HyperSim) datasets, after training their model on 12 other datasets and NYU Depth v2. In contrast, we only train on NYU Depth v2, and report a much larger improvement of (21%, 23%, 81%, 25%)." + ], + "bbox": [ + 500, + 106, + 893, + 574 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 585, + 640, + 601 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Traditional Methods. Earlier techniques for SIDE have used Markov Random Fields [47], non-parametric depth sampling [18], and structural similarity with prior depth map [13] to predict pixel-wise depth.", + "bbox": [ + 498, + 611, + 890, + 672 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Deep Learning Techniques for SIDE. Modern techniques have approached the problem as a dense regression problem. CNNs have been the dominant architecture for the SIDE in the last decade, with global-local network stack [9], and multi-scale [21], or encoder-decoder architecture [11] as some of the popular solution strategies. Recently, PixelFormer [1] used a transformer-based encoder-decoder architecture with skip connections from encoders to decoders. MIM [51] proposed masked image modeling as a general-purpose pre-training for geometric and motion tasks such as SIDE and pose estimation. Similarly, AiT [28] used mask augmentation and proposed soft tokens to generalize visual prediction tasks. While earlier works either focused on MDE for a specific dataset or RDE for generalization on multiple datasets, ZoEDepth [4] has proposed", + "bbox": [ + 498, + 674, + 893, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "28286", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/c7e530d0b4f2d38341472733a1549e75c90b00840b1be1a8bba2c60c1793cff7.jpg", + "image_caption": [ + "Figure 2. An overview of our proposed model: The latent representation of the input image undergoes a diffusion process, which is conditioned by our proposed CIDe module. Within the CIDe module, the input image is fed through the frozen ViT model. From this, a linear combination of the learnt embeddings is computed, which is transformed to generate a 768-dimensional contextual embedding. This embedding is utilized to condition the diffusion backbone. Subsequently, hierarchical feature maps are extracted from the UNet's decoder which are concatenated and processed through a depth regressor to generate the depth map." + ], + "image_footnote": [], + "bbox": [ + 80, + 92, + 880, + 300 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "a generalized method for MDE that performs well in zero-shot transfer. We outperform [4] by a large margin even when training on a single dataset - NYU Depth v2 for indoor scenes or KITTI for outdoor scenes. Whereas, [4] uses 12 datasets for pre-training and then fine-tunes on NYU Depth v2 or KITTI for zero shot transfer.", + "bbox": [ + 75, + 410, + 467, + 500 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Diffusion-based Methods with Pretraining on Large Datasets. Recently, many techniques for SIDE has used diffusion architectures. These techniques [8, 16, 20, 38, 54], exploit prior knowledge acquired by pretraining on large datasets like LAION-400M [39], which consists of 400 million image-text pairs. In contrast, depth datasets, such as NYU Depth v2 and KITTI, contain around 20-30 thousand image-depth pairs. DepthGen [38] and DDP [16] work on a noise-to-depthmap paradigm and use images for conditional guidance of the diffusion process. DepthGen employs self-supervised pretraining on tasks like colorization, inpainting, and JPEG artifact removal, followed by supervised training on indoor and outdoor datasets [6, 12, 26, 27, 48]. DDP [16] decouples the image encoder and map decoder, allowing the image encoder to run just once, while the lightweight map decoder is run multiple times. VPD [54] and TADP [20] use denoising UNet [37] as a backbone to extract the rich features at multiple scales. Also, they utilize text instead of image for conditioning the diffusion backbone.", + "bbox": [ + 75, + 503, + 467, + 805 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Vision Transformer for Scene Understanding. The transformer architecture was initially proposed for NLP tasks [50], but introduced to the computer vision community as Vision Transformer (ViT) in [7]. Prior to this, CNNs dominated computer vision problems due to their ability to capture spatial hierarchies in image data. However, such archi", + "bbox": [ + 75, + 809, + 467, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "tectures were constrained due to their ability to learn spatially localized features only. Transformer architecture has a weaker inductive bias and allows ViT to learn long-range dependencies, and robust, generalizable features. ViT architectures have replaced CNNs for SOTA performance on most computer vision tasks in recent years. We use a pretrained ViT model for providing semantic information to the diffusion backbone in our model.", + "bbox": [ + 498, + 410, + 890, + 530 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Proposed Methodology", + "text_level": 1, + "bbox": [ + 500, + 561, + 718, + 579 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Preliminaries", + "text_level": 1, + "bbox": [ + 500, + 590, + 640, + 606 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Problem Formulation. The objective of single image depth prediction task is to predict continuous values, denoted as $\\mathbf{y} \\in \\mathbb{R}^{H \\times W}$ , for every pixel present in the input RGB image, $\\mathbf{x} \\in [0,255]^{3 \\times H \\times W}$ . Here $H$ and $W$ represent the height and width respectively of the input image.", + "bbox": [ + 498, + 619, + 890, + 696 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Diffusion Model. Diffusion models are a class of generative models that progressively inject noise into the input data (forward pass) and then learn to reconstruct the original data in a reverse denoising process (reverse pass). There are three formulations of diffusion models: denoising diffusion probabilistic models (DDPMs) [14], score-based generative models [43, 44], and those based on stochastic differential equations [45, 46]. DDPMs are of relevance to our paper, and are described below. The architecture of DDPMs consists of two Markov chains: a forward chain that adds noise to the data, and a reverse chain that converts noise back to data by learning transition kernels parameterized by deep neural networks. Formally, the forward pass is modeled as", + "bbox": [ + 496, + 703, + 892, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "28287", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "a Markov process:", + "bbox": [ + 76, + 90, + 202, + 104 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbb {P} \\left(\\mathbf {z} _ {t} \\mid \\mathbf {z} _ {t - 1}\\right) = \\mathcal {N} \\left(\\mathbf {z} _ {t}; \\sqrt {1 - \\beta_ {t}} \\mathbf {z} _ {t - 1}, \\beta_ {t} \\boldsymbol {I}\\right). \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 122, + 113, + 468, + 140 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Here $\\mathbf{z}_t$ denotes the random variable at the $t^{\\mathrm{th}}$ time step, $\\mathcal{N}(\\mathbf{z};\\boldsymbol {\\mu},\\boldsymbol {\\sigma})$ denotes Gaussian probability distribution, and $\\beta_{t}$ is the noise schedule. The above equation leads to the analytic form of $\\mathbb{P}(\\mathbf{z}_t\\mid \\mathbf{z}_0),\\forall t\\in \\{0,1,\\dots ,T\\}$ :", + "bbox": [ + 76, + 147, + 468, + 208 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {z} _ {t} = \\sqrt {\\bar {\\beta} _ {t}} \\mathbf {z} _ {0} + \\sqrt {1 - \\bar {\\beta} _ {t}} \\boldsymbol {\\epsilon}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 181, + 215, + 468, + 242 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\bar{\\beta}_t = \\prod_{s=1}^t \\beta_s$ and $\\epsilon \\sim \\mathcal{N}(\\mathbf{0},\\mathbf{I})$ . The model then gradually removes noise by executing a learnable Markov chain in the reverse time direction, parameterized by a normal prior distribution $\\mathbb{P}(\\mathbf{z}_T) = \\mathcal{N}(\\mathbf{z}_T;0,\\mathbf{I})$ and a learnable transition kernel $\\mathbb{P}_{\\theta}(\\mathbf{z}_{t-1}|\\mathbf{z}_t)$ given by:", + "bbox": [ + 76, + 250, + 468, + 327 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbb {P} _ {\\theta} \\left(\\mathbf {z} _ {t - 1} \\mid \\mathbf {z} _ {t}\\right) = \\mathcal {N} \\left(\\mathbf {z} _ {t - 1}; \\mu_ {\\theta} \\left(\\mathbf {z} _ {t}, t\\right), \\Sigma_ {\\theta} \\left(\\mathbf {z} _ {t}, t\\right)\\right). \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 99, + 334, + 468, + 351 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The goal of the training process is to approximately match the reverse Markov chain with the actual time reversal of the forward Markov chain. Mathematically, parameter $\\theta$ is adjusted so that the joint distribution of the reverse Markov chain $\\mathbb{P}_{\\theta}(\\mathbf{z}_0,\\mathbf{z}_1,\\dots ,\\mathbf{z}_T)\\coloneqq \\mathbb{P}(\\mathbf{z}_T)\\prod_{t = 1}^T\\mathbb{P}_{\\theta}(\\mathbf{z}_{t - 1}\\mid \\mathbf{z}_t)$ closely approximates that of the forward Markov chain $\\mathbb{P}(\\mathbf{z}_0,\\mathbf{z}_1,\\dots ,\\mathbf{z}_T)\\coloneqq \\mathbb{P}(\\mathbf{z}_0)\\prod_{t = 1}^T\\mathbb{P}(\\mathbf{z}_t\\mid \\mathbf{z}_{t - 1})$ . This is achieved by minimizing the following loss:", + "bbox": [ + 76, + 359, + 468, + 479 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbb {E} _ {t \\sim \\mathcal {U} [ 1, T ], \\mathbf {z} _ {0} \\sim \\mathbb {P} (\\mathbf {z} _ {0}), \\epsilon \\sim \\mathcal {N} (0, I)} \\left[ \\| \\epsilon - \\epsilon_ {\\theta} (\\mathbf {z} _ {t}, t) \\| ^ {2} \\right], \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 102, + 487, + 468, + 513 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathbf{z}_t$ is computed using Eq. (2), and $\\epsilon_{\\theta}$ is predicted using a neural network, typically a UNet architecture [37]. In a conditional diffusion model $\\epsilon_{\\theta}(\\mathbf{z}_t,t)$ gets replaced by $\\epsilon_{\\theta}(\\mathbf{z}_t,t,\\mathcal{C})$ , where $\\mathcal{C}$ is a conditioning variable.", + "bbox": [ + 76, + 518, + 468, + 580 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Our Architecture", + "text_level": 1, + "bbox": [ + 76, + 587, + 246, + 602 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Image Encoder and Latent Diffusion. Diffusion models typically take large number of time steps to train, and are difficult to converge. Recently, [36] proposed a new diffusion with improved convergence properties and is called \"stable-diffusion\". The key idea is to perform the diffusion in latent space, with latent embedding learnt separately through a variational autoencoder (VAE). The Encoder of VAE first transforms the input image $\\mathbf{x}$ of size $(H,W)$ to latent space, then we follow latent diffusion formulation and utilize the UNet used in Stable Diffusion[36]. Utilizing latent diffusion formulation enables our architecture to capture multi-resolution features. Hence, we aggregate the feature maps from different layers of the UNet module (implementing conditional diffusion) by bringing them all to 1/4th resolution of the latent space, resulting in a feature map of size $8e\\times H / 32\\times W / 32$ .", + "bbox": [ + 76, + 611, + 468, + 851 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Exploiting Semantic Context with Conditional Diffusion. Recall that we formulated the single image depth estimation as a dense regression problem, predicting $\\mathbb{P}(\\mathbf{y}\\mid \\mathbf{x})$", + "bbox": [ + 76, + 854, + 468, + 901 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/c884c50e858aad2ed14417c86c14d50488599881dbd4715cd3e63339ec6572b7.jpg", + "image_caption": [ + "(a)" + ], + "image_footnote": [], + "bbox": [ + 598, + 88, + 629, + 215 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/57991546efa28fa9bd31fcb0cf70cb7c634b4e25f7be459feddfdac4e6a8ced9.jpg", + "image_caption": [ + "(b)", + "Figure 3. (a) Probabilistic graphical model corresponding to VPD. (b) The same corresponding to our formulation. Here, $\\mathcal{C}$ represents the semantic embedding derived from our CIDE module. This embedding is internally generated by passing $\\mathbf{x}$ through the ViT, resulting in $\\mathcal{E}$ . Subsequently, $\\mathcal{E}$ undergoes further processing to yield $\\mathcal{C}$ , which is then utilized in the conditional diffusion module implementing $\\mathbb{P}(\\mathbf{z}_0\\mid \\mathbf{z}_t,\\mathcal{C})$ . The output of the conditional diffusion module is fed into the Depth Regressor module within our architecture, implementing $\\mathbb{P}(\\mathbf{y}\\mid \\mathbf{z}_0)$ ." + ], + "image_footnote": [], + "bbox": [ + 746, + 88, + 808, + 215 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathbf{y}$ and $\\mathbf{x}$ denote output depth, and input image respectively. In our diffusion formulation, we predict $\\mathbf{z}_0$ which is then used to predict pixel-wise depth, $\\mathbf{y}$ . It has been shown that noise prediction in a diffusion model, $\\epsilon_{\\theta}$ , can be seen as predicting the gradient of the density function, $\\nabla_{\\mathbf{z}_t} \\log \\mathbb{P}(\\mathbf{z}_t)$ . Hence, overall architecture for SIDE using a diffusion architecture can be seen as factorizing the conditional probability, as shown in the probabilistic graph model in Fig. 3a. To utilize additional semantic context generated from a ViT model, we condition it on the ViT embeddings as shown in Fig. 3b. Mathematically, we model:", + "bbox": [ + 496, + 392, + 890, + 559 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbb {P} (\\mathbf {y} \\mid \\mathbf {x}, \\mathcal {E}) = \\mathbb {P} (\\mathbf {y} \\mid \\mathbf {z} _ {0}) \\mathbb {P} (\\mathbf {z} _ {0} \\mid \\mathbf {z} _ {t}, \\mathcal {C}) \\mathbb {P} (\\mathbf {z} _ {t} \\mid \\mathbf {x}) \\mathbb {P} (\\mathcal {C} \\mid \\mathbf {x}),\n$$\n", + "text_format": "latex", + "bbox": [ + 513, + 587, + 875, + 604 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where", + "bbox": [ + 500, + 608, + 544, + 619 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbb {P} (\\mathcal {C} \\mid \\mathbf {x}) = \\mathbb {P} (\\mathcal {C} \\mid \\mathcal {E}) \\mathbb {P} (\\mathcal {E} \\mid \\mathbf {x}). \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 532, + 626, + 890, + 643 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Here, the first term, $\\mathbb{P}(\\mathbf{y}\\mid \\mathbf{z}_0)$ , is implemented through the Depth Regressor module explained earlier. Similarly, $\\mathbb{P}(\\mathbf{z}_t\\mid \\mathbf{x})$ is implemented using the VAE's Encoder as described earlier. We generate conditional information, $\\mathcal{C}$ using our Comprehensive Image Detail Embedding module (hereafter CIDE, and explained below). The CIDE module takes $\\mathbf{x}$ as input and generates embedding vector $\\mathcal{C}$ (of dimension 768 in our design) as the output, thus, implementing $\\mathbb{P}(\\mathcal{C}\\mid \\mathbf{x})$ given in Eq. (5). We use $\\mathcal{E}$ to denote the embedding vector of a ViT module, and $\\mathbb{P}(\\mathcal{E}\\mid \\mathbf{x})$ is implemented through the ViT. $\\mathbb{P}(\\mathcal{C}\\mid \\mathcal{E})$ is implemented using downstream modules in CIDE consisting of learnable embeddings. The second term, $\\mathbb{P}(\\mathbf{z}_0\\mid \\mathbf{z}_t,\\mathcal{C})$ , is implemented using conditional diffusion,", + "bbox": [ + 496, + 656, + 890, + 866 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Comprehensive Image Detail Embedding (CIDE) Module. As described earlier, we believe using pseudo-captions", + "bbox": [ + 496, + 869, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "28288", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/85cecef06f97900f28fcbdaa57d7bdd2247ae7c37b92b6324ec104478fa24f85.jpg", + "table_caption": [ + "Table 1. Results on Indoor NYU Depth v2 [27] Dataset. Results that are bold perform best. $\\uparrow$ means the metric should be higher, $\\downarrow$ indicate lower is better. The evaluation uses an upper bound of 10 meters on the ground truth depth map. All the numbers for other works have been taken from the corresponding papers. For MIM, and ZoEDepth we have used SwinV2-L 1K, and ZoeDepth-M12-N versions respectively. We see an overall improvement against SOTA on all the metrics used for evaluation." + ], + "table_footnote": [], + "table_body": "
MethodVenueAbs Rel↓RMSE↓log10↓Sq Rel↓δ1↑δ2↑δ3↑
Eigen et al.[9]NIPS'140.1580.641--0.7690.9500.988
DORN[10]CVPR'180.1150.5090.051-0.8280.9650.992
SharpNet[32]ICCV'190.1390.5020.047-0.8360.9660.993
Chen et al.[5]IJCAI-190.1110.5140.048-0.8780.9770.994
BTS[22]Arxiv'190.1100.3920.0470.0660.8850.9780.994
AdaBins[2]CVPR'210.1030.3640.044-0.9030.9840.997
DPT[33]ICCV'210.1100.3570.045-0.9040.9880.998
P3Depth[30]CVPR'220.1040.3560.043-0.8980.9810.996
NeWCRFs[53]CVPR'220.0950.3340.0410.0450.9220.9920.998
SwinV2-B[24]CVPR'220.1330.4620.059-0.8190.9750.995
SwinV2-L[24]CVPR'220.1120.3810.051-0.8860.9840.997
Localbins[3]ECCV'220.0990.3570.042-0.9070.9870.998
Jun et al.[17]ECCV'220.0980.3550.042-0.9130.9870.998
PixelFormer[1]WACV'230.0900.3220.0390.0430.9290.9910.998
DDP[16]ICCV'230.0940.3290.040-0.9210.9900.998
MIM [51]CVPR'230.0830.2870.035-0.9490.9940.999
AiT[28]ICCV'230.0760.2750.033-0.9540.9940.999
ZoeDepth [4]Arxiv'230.0750.2700.0320.0300.9550.9950.999
VPD[54]ICCV'230.0690.2540.0300.0270.9640.9950.999
OursCVPR'240.0590.2180.0260.0130.9780.9970.999
", + "bbox": [ + 106, + 155, + 867, + 496 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/a76cc74dd06135df0021484b881dda524bb0070b423a6e292f03cff9e997573e.jpg", + "image_caption": [ + "Figure 4. Visual Comparison on NYU Depth v2 Indoor Dataset. Note, our method's ability to delineate objects in terms of their depth, such as the table lamp in Row 5, even when such information is absent from the ground truth depth map." + ], + "image_footnote": [], + "bbox": [ + 158, + 507, + 812, + 825 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "28289", + "bbox": [ + 478, + 945, + 521, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/7e3acb0e160e30d314f36226aad83e643f08cfb3b7d455b1bbed0204d4726d20.jpg", + "table_caption": [ + "Table 2. Performance on the Outdoor KITTI [12] Dataset. Please refer to the caption of Tab. 1 for notation details. For ZoEDepth results, we use the ZoeDepth-M12-K version following the authors' recommendation. In instances where results for certain methods were not reported in the respective works, denoted by “-”, and the code is unavailable, we were unable to generate the missing numbers. Despite the saturation of results on the outdoor KITTI dataset, our method consistently achieves comparable or superior performance to the state-of-the-art (SOTA) across all metrics. VPD[54] cannot be trained on KITTI dataset as it required a per image text label which is not present in KITTI. The symbol $\\dagger$ indicates that the method utilizes additional information beyond RGB." + ], + "table_footnote": [], + "table_body": "
MethodVenueAbs Rel↓Sq Rel↓RMSElog ↓RMSE↓δ1 ↑δ2 ↑δ3 ↑
Eigen et al.[9]NIPS'140.2031.5170.2826.3070.7020.8980.967
DORN[10]CVPR'180.0720.3070.1202.7270.9320.9840.994
BTS[22]Arxiv'190.0590.2410.0962.7560.9560.9930.998
AdaBins[2]CVPR'210.0670.1900.0882.9600.9490.9920.998
DPT[33]ICCV'210.060-0.0922.5730.9590.9950.996
P3Depth[30]CVPR'220.0710.2700.1032.8420.9530.9930.998
NeWCRFs[53]CVPR'220.0520.1550.0792.1290.9740.9970.999
PixelFormer[1]WACV'230.0510.1490.0772.0810.9760.9970.999
ZoeDepth [4]Arxiv'230.0540.1890.0832.4400.970.9960.999
DDP [16]ICCV'230.0500.1480.0762.0720.9750.9970.999
URCDC [40]ToM'230.0500.1420.0762.0320.9770.9970.999
IEBins [41]NeurIPS'230.0500.1420.0752.0110.9780.9980.999
MIM [51]CVPR'230.0500.1390.0751.9660.9770.9981.000
GEDepth[52]†ICCV'230.0480.1420.0762.0440.9760.9970.999
OursCVPR'240.0480.1390.0742.0390.9790.9981.000
", + "bbox": [ + 84, + 183, + 888, + 450 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/fd5626577cb0a6d7c11ad79b7f023d80f1eaefeb0fdc9d0368fc06a025dcee80.jpg", + "image_caption": [ + "Figure 5. Visual Comparison on KITTI Outdoor Dataset." + ], + "image_footnote": [], + "bbox": [ + 78, + 460, + 890, + 705 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "to generate the semantic context has limited utility, as the textual descriptions typically focus on large salient objects only. Instead we propose our CIDE module which use embeddings from a pre-trained ViT, and extract detailed semantic context from these embeddings. For this we take 1000 dimensional logit vector from ViT, and pass it through a two layer MLP which converts it to a 100 dimensional vector. Subsequently, we employ this vector to compute the linear combination of 100 learnable embeddings.", + "bbox": [ + 75, + 756, + 472, + 891 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "This resulting embedding undergoes a linear transformation to yield a semantic context vector of dimension 768, which is then passed to conditional diffusion module.", + "bbox": [ + 496, + 756, + 890, + 800 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Depth Regressor. The output feature map undergoes through an Upsampling Decoder, comprised of deconvolution layers, followed by a Depth Regressor. The Depth Regressor is essentially a two-layer convolutional neural network (CNN), with the initial layer having dimensions $\\mathrm{Conv}(3\\times 3)$ , 192, and the subsequent layer $\\mathrm{Conv}(3\\times 3)$ , 1.", + "bbox": [ + 496, + 809, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "28290", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/2e0b5f3c63f0cc8f262938f75cd81372b094e58cf9940af92c8e9463c4040b78.jpg", + "table_caption": [ + "Table 3. Quantitative results for zero-shot transfer to four unseen indoor datasets. mRI denotes the mean relative improvement with respect to NeWCRFs across all metrics (δ1, REL, RMSE). Evaluation depth is capped at 8m for SUN RGB-D, 10m for iBims and DIODE Indoor, and 80m for HyperSim. Best results are in bold, second best are underlined. Our mRI outperforms all methods across all datasets by a large margin. † denotes that ZoeD is trained on 12 datasets and our method is trained only on NYUv2." + ], + "table_footnote": [], + "table_body": "
MethodSUN RGB-DiBims-1 BenchmarkDIODE IndoorHyperSim
δ1↑REL↓RMSE↓mRIθ↑δ1↑REL↓RMSE↓mRIθ↑δ1↑REL↓RMSE↓mRIθ↑δ1↑REL↓RMSE↓mRIθ↑
BTS [22]0.7400.1720.515-14.2%0.5380.2310.919-6.9%0.2100.4181.9052.3%0.2250.4766.404-8.6%
AdaBins [2]0.7710.1590.476-7.0%0.5550.2120.901-2.1%0.1740.4431.963-7.2%0.2210.4836.546-10.5%
LocalBins [3]0.7770.1560.470-5.6%0.5580.2110.880-0.7%0.2290.4121.8537.1%0.2340.4686.362-6.6%
NeWCRFs [53]0.7980.1510.4240.0%0.5480.2060.8610.0%0.1870.4041.8670.0%0.2550.4426.0170.0%
VPD [54]0.8610.1210.35514.7%0.6270.1870.76711.5%0.4800.3921.29563.4%0.3330.5315.1118.5%
ZoeD-M12-N† [4]0.8640.1190.34616.0%0.6580.1690.71118.5%0.3760.3271.58845.0%0.2920.4105.7718.6%
Ours0.8850.1120.31920.5%0.6880.1630.66423.1%0.5450.3441.16481.3%0.3940.4424.73925.2%
", + "bbox": [ + 76, + 157, + 913, + 287 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/10a7e4cdf5394b729bfc2c2fda0f2f9fb4531171367e28fe6cc3a38ac46faaed.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 135, + 305, + 846, + 460 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f31058aa86b9b19c5376f9c5dabda1e36277891d933e7679fad2e97499e3bd57.jpg", + "image_caption": [ + "Figure 6. Visualization of improvements over VPD [54] in our model due to ViT embeddings passed down as conditional vectors (in blue and green). In the above images, ViT detects the desktop computer (first image) and table lamp (second image) with high probability, and they are thus better detected by our model. Additional visualizations are provided in the supplementary material." + ], + "image_footnote": [], + "bbox": [ + 135, + 478, + 844, + 638 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4. Experiments and Results", + "text_level": 1, + "bbox": [ + 76, + 718, + 310, + 734 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Datasets and Evaluation. We use NYU Depth v2 [27] and KITTI [12] as the primary datasets for training. The NYU Depth v2 dataset is a widely used indoor benchmark for monocular depth estimation, containing over 24k densely labeled pairs of RGB and depth images in the train set and 654 in the test set. The dataset covers a wide range of indoor scenes and includes challenging scenarios such as reflective surfaces, transparent objects, and occlusions. The ground truth depth maps are obtained using a structured light sensor and are provided at a resolution of $640 \\times 480$ .", + "bbox": [ + 75, + 750, + 470, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "KITTI dataset is a widely used outdoor benchmark for monocular depth estimation, containing over 24k densely labeled pairs of RGB and depth images. The dataset covers outdoor driving scenarios and includes varying lighting conditions, weather, and occlusions. The ground truth depth maps are obtained using a Velodyne LiDAR sensor and are provided at a resolution of $1242 \\times 375$ . To demonstrate generalizability, we evaluate zero-shot performance on the following datasets: Sun-RGBD [42], iBims1 [19], DIODE [49], and HyperSim [35]. Whereas, the main paper contains mostly quantitative, and only some representative", + "bbox": [ + 496, + 719, + 890, + 887 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "28291", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/5fdba6a9d50ed2fb62d14f37e682b07bbb0186f63efe75d3ebcd71721d942fd5.jpg", + "table_caption": [ + "Table 4. Effectiveness of different embeddings for guiding the diffusion process for depth estimation on NYU Depth v2 dataset. In the third row, rather than using pseudo caption embeddings generated from the scene label (as implemented by VPD [54]), we provide the one-hot vector representing the scene label as a condition to the diffusion model. We observe a slight improvement in the metrics, highlighting that information content in the caption is similar to that in one-hot label." + ], + "table_footnote": [], + "table_body": "
EmbeddingsRMSE↓Abs Rel↓δ1↑
Scene label emb. [54]0.2540.0690.964
Text caption emb. [20]0.2250.0620.976
One-hot vector emb.0.2440.0670.968
Proposed scene emb.0.2180.0590.978
", + "bbox": [ + 83, + 210, + 464, + 304 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "tive visual results, detailed visual results on each dataset are included in the supplementary.", + "bbox": [ + 76, + 329, + 468, + 359 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Implementation Details. Our model is implemented using PyTorch [29]. For optimization, we have used AdamW optimizer [25] with $\\beta_0$ values of 0.9 and 0.999, a batch size of 32, and a weight decay of 0.1. We train our model for 25 epochs for both KITTI and NYU Depth v2 datasets, with an initial learning rate of $3 \\times 10^{-5}$ . We first linearly increase learning rate to $5 \\times 10^{-4}$ , and then linearly decrease across training iterations. We use usual data augmentation techniques, including random hue addition, horizontal flipping, changing the image brightness, and Cut Depth [15]. Our model takes approximately 21 minutes per epoch to train using 8 NVIDIA A100 GPUs.", + "bbox": [ + 75, + 362, + 468, + 544 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.1. Comparison on Benchmark Datasets", + "text_level": 1, + "bbox": [ + 76, + 554, + 395, + 569 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Comparison on NYU Depth v2. Tab. 1 shows the comparison of our proposed method with SOTA methods on the indoor NYU Depth v2 dataset [27]. We achieve a new state of the art on this dataset. Our methods perform better than the previous SOTA ([54])) by a large margin of $14\\%$ in terms of RMSE. Fig. 4 provides a qualitative comparison on the dataset.", + "bbox": [ + 75, + 577, + 468, + 681 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Comparison on KITTI. Tab. 2 shows the comparison with various methods on the KITTI dataset [12]. Fig. 5 shows the qualitative results. Unlike NYU Depth v2 dataset, KITTI is an outdoor dataset. On this dataset also, we achieve similar or better performance than all existing state-of-the-art techniques.", + "bbox": [ + 75, + 685, + 468, + 777 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.2. Generalization and Zero Shot Transfer", + "text_level": 1, + "bbox": [ + 76, + 786, + 411, + 801 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Unlike state of the art on zero short transfer (ZoEDepth [4]), which requiring training on many datasets (12 in their case) for effective zero shot transfer, we show that our model generalizes well to other unseen dataset even when trained on a single NYU Depth v2 dataset. Tab. 3 shows quantitative results to back our findings.", + "bbox": [ + 75, + 809, + 468, + 902 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.3. Ablation Study", + "text_level": 1, + "bbox": [ + 500, + 90, + 653, + 107 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effect of Contextual Information. As highlighted in the motivation, a key observation of this study is the richness of information contained within the output probability vector of the ViT, surpassing the textual embeddings employed in current state-of-the-art (e.g., VPD [54]). To compare with the utilization of scene label information (as implemented by VPD), we construct the conditioning embedding as a one-hot vector using the scene label, and subsequently transform it using an MLP. As illustrated in Tab. 4, we observe a slight improvement over VPD, indicating that the information content in pseudo-captions resembles that of one-hot labels. Furthermore, our proposed method surpasses both the approaches.", + "bbox": [ + 496, + 113, + 890, + 311 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Qualitative Results. Perhaps the most important task would be to verify that use of rich probability vector instead of text embeddings actually results in an improvement in depth as a direct consequence. We do this by considering the top few objects predicted by the ViT and correlating this with the depth predicted at these objects. We hypothesise when a particular class say dog is predicted with a high probability (hence there must be a dog in the image), then the corresponding depth must also be more accurate. We show this in Fig. 6.", + "bbox": [ + 496, + 311, + 890, + 465 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 472, + 617, + 487 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We presented a new architecture block Comprehensive Image Detail Embedding (CIDE) module for robust monocular depth estimation in this paper. Our key idea is to highlight the limitations of using pseudo-captions to provide contextual information, and instead propose to use richer class-wise probability generated by a classification model, such as ViT. The motivation is that, while textual embedding typically highlights salient objects, class-wise probability vector preserves more details, including smaller objects in the background also. We implement the idea using proposed CIDE module cascaded with a conditional diffusion pipeline for monocular depth estimation. We demonstrated the effectiveness of our approach on several benchmark datasets and showed that it outperforms SOTA methods by a significant margin.", + "bbox": [ + 496, + 497, + 890, + 724 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgement. We thank Kartik Anand for assistance with some of the experiments. This work has been supported by funding through Staqu Technologies Pvt. Ltd., and Department of Science and Technology (DST), Government of India. The travel for Suraj Patni is supported through Research Acceleration funds of Department of Computer Science and Engineering, IIT Delhi.", + "bbox": [ + 496, + 727, + 890, + 834 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 500, + 847, + 596, + 863 + ], + "page_idx": 7 + }, + { + "type": "ref_text", + "text": "[1] Ashutosh Agarwal and Chetan Arora. Attention attention everywhere: Monocular depth prediction with skip attention.", + "bbox": [ + 506, + 872, + 890, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "28292", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 7 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 5861-5870, 2023. 2, 5, 6", + "[2] Shariq Farooq Bhat, Ibrahim Alhashim, and Peter Wonka. Adabins: Depth estimation using adaptive bins. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4009-4018, 2021. 5, 6, 7", + "[3] Shariq Farooq Bhat, Ibraheem Alhashim, and Peter Wonka. Localbins: Improving depth estimation by learning local distributions. In European Conference on Computer Vision, pages 480-496. Springer, 2022. 5, 7", + "[4] Shariq Farooq Bhat, Reiner Birkl, Diana Wofk, Peter Wonka, and Matthias Müller. Zoedepth: Zero-shot transfer by combining relative and metric depth, 2023. 2, 3, 5, 6, 7, 8", + "[5] Xiaotian Chen, Xuejin Chen, and Zheng-Jun Zha. Structure-aware residual pyramid network for monocular depth estimation. arXiv preprint arXiv:1907.06023, 2019. 5", + "[6] Angela Dai, Angel X Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5828-5839, 2017. 3", + "[7] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 2, 3", + "[8] Yiqun Duan, Xianda Guo, and Zheng Zhu. Diffusiondepth: Diffusion denoising approach for monocular depth estimation, 2023. 2, 3", + "[9] David Eigen, Christian Puhrsch, and Rob Fergus. Depth map prediction from a single image using a multi-scale deep network. Advances in neural information processing systems, 27, 2014. 2, 5, 6", + "[10] Huan Fu, Mingming Gong, Chaohui Wang, Kayhan Bat-manghelich, and Dacheng Tao. Deep ordinal regression network for monocular depth estimation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2002-2011, 2018. 5, 6", + "[11] Huan Fu, Mingming Gong, Chaohui Wang, Kayhan Bat-manghelich, and Dacheng Tao. Deep ordinal regression network for monocular depth estimation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2002-2011, 2018. 2", + "[12] Andreas Geiger, Philip Lenz, Christoph Stiller, and Raquel Urtasun. Vision meets robotics: The kitti dataset. International Journal of Robotics Research (IJRR), 2013. 3, 6, 7, 8", + "[13] José L. Herrera, Carlos R. del Blanco, and Narciso García. Automatic depth extraction from 2d images using a cluster-based learning framework. IEEE Transactions on Image Processing, 27(7):3288-3299, 2018. 2", + "[14] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 3" + ], + "bbox": [ + 78, + 90, + 470, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[15] Yasunori Ishii and Takayoshi Yamashita. Cutdepth: Edge-aware data augmentation in depth estimation. arXiv preprint arXiv:2107.07684, 2021. 8", + "[16] Yuanfeng Ji, Zhe Chen, Enze Xie, Lanqing Hong, Xihui Liu, Zhaoqiang Liu, Tong Lu, Zhenguo Li, and Ping Luo. Ddp: Diffusion model for dense visual prediction. arXiv preprint arXiv:2303.17559, 2023. 2, 3, 5, 6", + "[17] Jinyoung Jun, Jae-Han Lee, Chul Lee, and Chang-Su Kim. Depth map decomposition for monocular depth estimation. In European Conference on Computer Vision, pages 18-34. Springer, 2022. 5", + "[18] Kevin Karsch, Ce Liu, and Sing Bing Kang. Depth transfer: Depth extraction from video using non-parametric sampling. IEEE transactions on pattern analysis and machine intelligence, 36(11):2144-2158, 2014. 2", + "[19] Tobias Koch, Lukas Liebel, Friedrich Fraundorfer, and Marco Korner. Evaluation of cnn-based single-image depth estimation methods. In Proceedings of the European Conference on Computer Vision (ECCV) Workshops, pages 0–0, 2018. 1, 7", + "[20] Neehar Kondapaneni, Markus Marks, Manuel Knott, Rogério Guimarães, and Pietro Perona. Text-image alignment for diffusion-based perception. arXiv preprint arXiv:2310.00031, 2023. 2, 3, 8", + "[21] Iro Laina, Christian Rupprecht, Vasileios Belagiannis, Federico Tombari, and Nassir Navab. Deeper depth prediction with fully convolutional residual networks. In 2016 Fourth international conference on 3D vision (3DV), pages 239-248. IEEE, 2016. 2", + "[22] Jin Han Lee, Myung-Kyu Han, Dong Wook Ko, and Il Hong Suh. From big to small: Multi-scale local planar guidance for monocular depth estimation. arXiv preprint arXiv:1907.10326, 2019. 5, 6, 7", + "[23] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In International Conference on Machine Learning, pages 12888-12900. PMLR, 2022. 2", + "[24] Ze Liu, Han Hu, Yutong Lin, Zhuliang Yao, Zhenda Xie, Yixuan Wei, Jia Ning, Yue Cao, Zheng Zhang, Li Dong, et al. Swin transformer v2: Scaling up capacity and resolution. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12009-12019, 2022. 5", + "[25] Ilya Loshchilov and Frank Hutter. Decoupled Weight Decay Regularization. In ICML 2019. 8", + "[26] John McCormac, Ankur Handa, Stefan Leutenegger, and Andrew J Davison. Scelenet rgb-d: 5m photorealistic images of synthetic indoor trajectories with ground truth. arXiv preprint arXiv:1612.05079, 2016. 3", + "[27] Pushmeet Kohli Nathan Silberman, Derek Hoiem and Rob Fergus. Indoor segmentation and support inference from rgbd images. In ECCV, 2012. 3, 5, 7, 8", + "[28] Jia Ning, Chen Li, Zheng Zhang, Chunyu Wang, Zigang Geng, Qi Dai, Kun He, and Han Hu. All in tokens: Unifying output space of visual tasks via soft token. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 19900-19910, 2023. 2, 5" + ], + "bbox": [ + 501, + 92, + 892, + 901 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "28293", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[29] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, Alban Desmaison, Andreas Kopf, Edward Yang, Zachary DeVito, Martin Raison, Alykhan Tejani, Sasank Chilamkurthy, Benoit Steiner, Lu Fang, Junjie Bai, and Soumith Chintala. Pytorch: An imperative style, high-performance deep learning library. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2019. 8", + "[30] Vaishakh Patil, Christos Sakaridis, Alexander Liniger, and Luc Van Gool. P3depth: Monocular depth estimation with a piecewise planarity prior. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1610–1621, 2022. 5, 6", + "[31] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 2", + "[32] Michael Ramamonjisoa and Vincent Lepetit. Sharpnet: Fast and accurate recovery of occluding contours in monocular depth estimation. In Proceedings of the IEEE/CVF International Conference on Computer Vision Workshops, pages 0–0, 2019. 5", + "[33] René Ranftl, Alexey Bochkovskiy, and Vladlen Koltun. Vision transformers for dense prediction. In Proceedings of the IEEE/CVF international conference on computer vision, pages 12179-12188, 2021. 5, 6", + "[34] René Ranftl, Katrin Lasinger, David Hafner, Konrad Schindler, and Vladlen Koltun. Towards robust monocular depth estimation: Mixing datasets for zero-shot cross-dataset transfer. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(3), 2022. 2", + "[35] Mike Roberts, Jason Ramapuram, Anurag Ranjan, Atulit Kumar, Miguel Angel Bautista, Nathan Paczan, Russ Webb, and Joshua M. Susskind. Hypersim: A photorealistic synthetic dataset for holistic indoor scene understanding. In International Conference on Computer Vision (ICCV) 2021, 2021. 1, 7", + "[36] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 4", + "[37] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical Image Computing and Computer-Assisted Intervention-MICCAI 2015: 18th International Conference, Munich, Germany, October 5-9, 2015, Proceedings, Part III 18, pages 234-241. Springer, 2015. 3, 4", + "[38] Saurabh Saxena, Abhishek Kar, Mohammad Norouzi, and David J Fleet. Monocular depth estimation using diffusion models. arXiv preprint arXiv:2302.14816, 2023. 2, 3", + "[39] Christoph Schuhmann, Richard Vencu, Romain Beaumont, Robert Kaczmarczyk, Clayton Mullis, Aarush Katta, Theo Coombes, Jenia Jitsev, and Aran Komatsuzaki. Laion-400m:" + ], + "bbox": [ + 78, + 90, + 470, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Open dataset of clip-filtered 400 million image-text pairs. arXiv preprint arXiv:2111.02114, 2021. 2, 3", + "[40] Shuwei Shao, Zhongcai Pei, Weihai Chen, Ran Li, Zhong Liu, and Zhengguo Li. Urcdc-depth: Uncertainty rectified cross-distillation with cutflip for monocular depth estimation. arXiv preprint arXiv:2302.08149, 2023. 6", + "[41] Shuwei Shao, Zhongcai Pei, Xingming Wu, Zhong Liu, Weihai Chen, and Zhengguo Li. Iebins: Iterative elastic bins for monocular depth estimation. arXiv preprint arXiv:2309.14137, 2023. 6", + "[42] Shuran Song, Samuel P. Lichtenberg, and Jianxiong Xiao. Sun rgb-d: A rgb-d scene understanding benchmark suite. In 2015 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 567-576, 2015. 1, 7", + "[43] Yang Song and Stefano Ermon. Generative modeling by estimating gradients of the data distribution. Advances in neural information processing systems, 32, 2019. 3", + "[44] Yang Song and Stefano Ermon. Improved techniques for training score-based generative models. Advances in neural information processing systems, 33:12438-12448, 2020. 3", + "[45] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. arXiv preprint arXiv:2011.13456, 2020. 3", + "[46] Yang Song, Conor Durkan, Iain Murray, and Stefano Ermon. Maximum likelihood training of score-based diffusion models. Advances in Neural Information Processing Systems, 34: 1415-1428, 2021. 3", + "[47] M. Sun, A. Y. Ng, and A. Saxena. Make3d: Learning 3d scene structure from a single still image. IEEE Transactions on Pattern Analysis and Machine Intelligence, 31(05):824-840, 2009. 2", + "[48] Pei Sun, Henrik Kretzschmar, Xerxes Dotiwalla, Aurelien Chouard, Vijaysai Patnaik, Paul Tsui, James Guo, Yin Zhou, Yuning Chai, Benjamin Caine, et al. Scalability in perception for autonomous driving: Waymo open dataset. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2446-2454, 2020. 3", + "[49] Igor Vasiljevic, Nick Kolkin, Shanyi Zhang, Ruotian Luo, Haochen Wang, Falcon Z Dai, Andrea F Daniele, Mohammadreza Mostajabi, Steven Basart, Matthew R Walter, et al. Diode: A dense indoor and outdoor depth dataset. arXiv preprint arXiv:1908.00463, 2019. 1, 7", + "[50] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 3", + "[51] Zhenda Xie, Zigang Geng, Jingcheng Hu, Zheng Zhang, Han Hu, and Yue Cao. Revealing the dark secrets of masked image modeling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14475-14485, 2023. 2, 5, 6", + "[52] Xiaodong Yang, Zhuang Ma, Zhiyu Ji, and Zhe Ren. Gedgepth: Ground embedding for monocular depth estimation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12719-12727, 2023. 6", + "[53] Weihao Yuan, Xiaodong Gu, Zuozhuo Dai, Siyu Zhu, and Ping Tan. Newcrfs: Neural window fully-connected crfs for" + ], + "bbox": [ + 501, + 92, + 890, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "28294", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "monocular depth estimation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2022. 5, 6, 7", + "[54] Wenliang Zhao, Yongming Rao, Zuyan Liu, Benlin Liu, Jie Zhou, and Jiwen Lu. Unleashing text-to-image diffusion models for visual perception. ICCV, 2023. 2, 3, 5, 6, 7, 8" + ], + "bbox": [ + 78, + 90, + 468, + 186 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "28295", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2024/ECoDepth_ Effective Conditioning of Diffusion Models for Monocular Depth Estimation/7e3c5935-f184-4aa8-a48b-682e7477329f_model.json b/2024/ECoDepth_ Effective Conditioning of Diffusion Models for Monocular Depth Estimation/7e3c5935-f184-4aa8-a48b-682e7477329f_model.json new file mode 100644 index 0000000000000000000000000000000000000000..ec63b5a61ebe92c8e31de70dbe53aa79960f3f88 --- /dev/null +++ b/2024/ECoDepth_ Effective Conditioning of Diffusion Models for Monocular Depth Estimation/7e3c5935-f184-4aa8-a48b-682e7477329f_model.json @@ -0,0 +1,2015 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.209, + 0.131, + 0.761, + 0.176 + ], + "angle": 0, + "content": "ECoDepth: Effective Conditioning of Diffusion Models for Monocular Depth Estimation" + }, + { + "type": "text", + "bbox": [ + 0.223, + 0.204, + 0.321, + 0.222 + ], + "angle": 0, + "content": "Suraj Patni*" + }, + { + "type": "text", + "bbox": [ + 0.409, + 0.205, + 0.558, + 0.222 + ], + "angle": 0, + "content": "Aradhye Agarwal*" + }, + { + "type": "text", + "bbox": [ + 0.636, + 0.205, + 0.746, + 0.22 + ], + "angle": 0, + "content": "Chetan Arora" + }, + { + "type": "text", + "bbox": [ + 0.34, + 0.223, + 0.63, + 0.24 + ], + "angle": 0, + "content": "Indian Institute of Technology Delhi" + }, + { + "type": "text", + "bbox": [ + 0.299, + 0.241, + 0.667, + 0.257 + ], + "angle": 0, + "content": "https://ecodepth-iitd.github.io" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.292, + 0.313, + 0.307 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.324, + 0.474, + 0.749 + ], + "angle": 0, + "content": "In the absence of parallax cues, a learning based single image depth estimation (SIDE) model relies heavily on shading and contextual cues in the image. While this simplicity is attractive, it is necessary to train such models on large and varied datasets, which are difficult to capture. It has been shown that using embeddings from pretrained foundational models, such as CLIP, improves zero shot transfer in several applications. Taking inspiration from this, in our paper we explore the use of global image priors generated from a pre-trained ViT model to provide more detailed contextual information. We argue that the embedding vector from a ViT model, pre-trained on a large dataset, captures greater relevant information for SIDE than the usual route of generating pseudo image captions, followed by CLIP based text embeddings. Based on this idea, we propose a new SIDE model using a diffusion backbone which is conditioned on ViT embeddings. Our proposed design establishes a new state-of-the-art (SOTA) for SIDE on NYU Depth v2 dataset, achieving Abs Rel error of 0.059(14% improvement) compared to 0.069 by the current SOTA (VPD). And on KITTI dataset, achieving Sq Rel error of 0.139 (2% improvement) compared to 0.142 by the current SOTA (GED). For zero shot transfer with a model trained on NYU Depth v2, we report mean relative improvement of (20%, 23%, 81%, 25%) over NeWCRF on (Sun-RGBD, iBims1, DIODE, HyperSim) datasets, compared to (16%, 18%, 45%, 9%) by ZoEDepth. The code is available in our project page." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.776, + 0.208, + 0.791 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.802, + 0.47, + 0.879 + ], + "angle": 0, + "content": "Single Image Depth Estimation (SIDE) is the task of predicting per pixel depth using a single RGB image from a monocular camera. It is a fundamental problem in computer vision with applications in several domains, viz robotics, autonomous driving, and augmented reality. The problem is" + }, + { + "type": "image", + "bbox": [ + 0.501, + 0.291, + 0.697, + 0.331 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.698, + 0.291, + 0.891, + 0.331 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.65, + 0.334, + 0.745, + 0.344 + ], + "angle": 0, + "content": "(a) Sun-RGBD [42]" + }, + { + "type": "image", + "bbox": [ + 0.502, + 0.345, + 0.696, + 0.385 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.698, + 0.345, + 0.892, + 0.385 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.656, + 0.388, + 0.737, + 0.399 + ], + "angle": 0, + "content": "(b) iBims1 [19]" + }, + { + "type": "image", + "bbox": [ + 0.502, + 0.399, + 0.697, + 0.44 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.698, + 0.399, + 0.891, + 0.44 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.659, + 0.442, + 0.734, + 0.453 + ], + "angle": 0, + "content": "(c) DIODE [49]" + }, + { + "type": "image", + "bbox": [ + 0.502, + 0.453, + 0.697, + 0.494 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.698, + 0.453, + 0.891, + 0.494 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.649, + 0.496, + 0.745, + 0.508 + ], + "angle": 0, + "content": "(d) HyperSim [35]" + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.52, + 0.892, + 0.617 + ], + "angle": 0, + "content": "Figure 1. Qualitative results across four different datasets, demonstrating the zero-shot performance of our model trained only on the NYU Depth v2 dataset. Corresponding quantitative results are presented in Table 3. The first column displays RGB images, the second column depicts ground truth depth, and the third column showcases our model's predicted depths. Additional images for each dataset are available in the Supplementary Material." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.643, + 0.892, + 0.735 + ], + "angle": 0, + "content": "typically formulated in two flavors: metric depth estimation (MDE), and relative depth estimation (RDE). As the names suggest, MDE deals with estimation in physical units such as meters, whereas, RDE techniques focus on relative depth only, and require a per-image affine transformation as post-processing to convert to a physical unit." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.735, + 0.892, + 0.886 + ], + "angle": 0, + "content": "Conventional geometric techniques for depth estimation typically rely on feature correspondence, parallax, and triangulation from two or more views. However, the problem becomes ill-posed for estimation from a single view. Intuitively, depth map is a 3D representation of the scene, whereas an RGB image is a 2D projection of the scene. Hence, it cannot be uniquely determined from a single RGB image. Therefore, learning-based SIDE models rely on visual cues like 'shape from shading' and other contextual priors for per-pixel depth prediction." + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.886, + 0.892, + 0.902 + ], + "angle": 0, + "content": "A data-driven approach makes the SIDE pipeline sim" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.887, + 0.206, + 0.9 + ], + "angle": 0, + "content": "*Equal contribution." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "28285" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.471, + 0.198 + ], + "angle": 0, + "content": "pler, but also makes the learnt model dependent on the quality of training data. It has been observed that such models overfit on a particular training distribution/domain and fail to generalize on unseen data. This is especially true for MDE models when the range of depth in the training dataset is limited. Hence, training on multiple datasets, with wide variations in depth ranges has been proposed [4, 34]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.202, + 0.474, + 0.549 + ], + "angle": 0, + "content": "On the other hand, development of large foundational models (LFMs) in recent years has altered the preferred design approach for many computer vision problems. These huge models are trained using extensive datasets of unlabeled images and learning objectives that are agnostic to specific tasks. The learnt embeddings from such pre-trained models have been shown to help generalization and zero-shot transfer in many applications. We are aware of at least two works for MDE in SIDE problem that have appeared in the last few months and make use of such foundational models. VPD [54] uses a text-to-image diffusion model pretrained on LAION-400M [39] dataset having large-scale image-text pairs as the backbone. The model prompts the denoising UNet with textual inputs to make the visual contents interact with the text prompts. Since the problem formulation doesn't include text description as the input, the model generates simple descriptions such as 'A photo of a {scene name}' based on the scene label given in the NYU Depth v2 dataset. Another work, TADP [20] improves upon VPD. Instead of simple image descriptions, TADP uses BLIP-2 [23] to generate image captions, and then uses CLIP [31] embeddings of the pseudo-caption to condition the diffusion model." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.554, + 0.473, + 0.901 + ], + "angle": 0, + "content": "We view the above works as providing robust semantic context to LFMs for the actual task at hand, which helps in visual recognition in general, as well as SIDE. While we do agree with the broad motivations of these works, the question that we ask is if pseudo-captions are the most effective way to provide the semantic context. Textual descriptions of an image typically focus on large salient objects and emphasize on their relationships. On the other hand, large vision models for image classification, typically contain representation for even smaller objects present in the scene. Even when a single object is present in the scene, the representations typically capture uncertainties and ambiguities inherent in the scene. Hence, we posit that using embeddings from a transformer model, pretrained on a large image dataset unrelated to the SIDE task, captures more relevant information, and is a better alternative to using pseudocaptions' CLIP embedding. Since diffusion-based models have shown their superiority for the dense prediction tasks in recent works [8, 16, 20, 38, 54]. Hence, we propose a diffusion backbone for our model, along with a novel CIDE module which employs ViT [7] to extract semantic context embeddings. These embeddings are subsequently utilized to condition the diffusion backbone." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.092, + 0.866, + 0.106 + ], + "angle": 0, + "content": "Contributions. The key contributions of this work are:" + }, + { + "type": "text", + "bbox": [ + 0.501, + 0.107, + 0.894, + 0.318 + ], + "angle": 0, + "content": "(1) We propose a new model for MDE in a SIDE task. The proposed model uses a conditional diffusion architecture with the semantic context being supplied through embeddings generated using a ViT model. Achieving a new state-of-the-art (SOTA) performance, our method outperforms existing approaches on benchmark datasets, including the NYU Depth v2 indoor and KITTI outdoor datasets. Notably, we report a significant improvement of \\(14\\%\\) in absolute relative error, achieving 0.059 compared to the current SOTA (VPD) performance of 0.069 on NYU Depth v2. And we report an improvement of \\(2\\%\\) in square relative error, achieving 0.139 compared to the current SOTA (GED) performance of 0.142 on KITTI." + }, + { + "type": "text", + "bbox": [ + 0.501, + 0.318, + 0.894, + 0.438 + ], + "angle": 0, + "content": "(2) We show, qualitatively as well as quantitatively, that using ViT embeddings to provide semantic context is a better alternative to generating pseudo captions and then using its CLIP embeddings to condition a SIDE model. In contrast to TADP [20], which uses pseudo captions, but only achieves a RMSE of 0.225 on NYU Depth v2, we report a lower, and a new SOTA, error of 0.218 (VPD [54] achieves 0.254)." + }, + { + "type": "text", + "bbox": [ + 0.501, + 0.439, + 0.894, + 0.575 + ], + "angle": 0, + "content": "(3) We show that providing ViT conditioning, helps our model perform better in a zero shot transfer task. ZoEDepth [4], the current SOTA for zero-shot transfer, reports an improvement of (16%, 18%, 45%, 9%) over NeWCRF on (Sun-RGBD, iBims1, DIODE, HyperSim) datasets, after training their model on 12 other datasets and NYU Depth v2. In contrast, we only train on NYU Depth v2, and report a much larger improvement of (21%, 23%, 81%, 25%)." + }, + { + "type": "list", + "bbox": [ + 0.501, + 0.107, + 0.894, + 0.575 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.587, + 0.642, + 0.602 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.612, + 0.892, + 0.673 + ], + "angle": 0, + "content": "Traditional Methods. Earlier techniques for SIDE have used Markov Random Fields [47], non-parametric depth sampling [18], and structural similarity with prior depth map [13] to predict pixel-wise depth." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.675, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Deep Learning Techniques for SIDE. Modern techniques have approached the problem as a dense regression problem. CNNs have been the dominant architecture for the SIDE in the last decade, with global-local network stack [9], and multi-scale [21], or encoder-decoder architecture [11] as some of the popular solution strategies. Recently, PixelFormer [1] used a transformer-based encoder-decoder architecture with skip connections from encoders to decoders. MIM [51] proposed masked image modeling as a general-purpose pre-training for geometric and motion tasks such as SIDE and pose estimation. Similarly, AiT [28] used mask augmentation and proposed soft tokens to generalize visual prediction tasks. While earlier works either focused on MDE for a specific dataset or RDE for generalization on multiple datasets, ZoEDepth [4] has proposed" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.521, + 0.957 + ], + "angle": 0, + "content": "28286" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.081, + 0.093, + 0.882, + 0.301 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.315, + 0.893, + 0.384 + ], + "angle": 0, + "content": "Figure 2. An overview of our proposed model: The latent representation of the input image undergoes a diffusion process, which is conditioned by our proposed CIDe module. Within the CIDe module, the input image is fed through the frozen ViT model. From this, a linear combination of the learnt embeddings is computed, which is transformed to generate a 768-dimensional contextual embedding. This embedding is utilized to condition the diffusion backbone. Subsequently, hierarchical feature maps are extracted from the UNet's decoder which are concatenated and processed through a depth regressor to generate the depth map." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.411, + 0.468, + 0.501 + ], + "angle": 0, + "content": "a generalized method for MDE that performs well in zero-shot transfer. We outperform [4] by a large margin even when training on a single dataset - NYU Depth v2 for indoor scenes or KITTI for outdoor scenes. Whereas, [4] uses 12 datasets for pre-training and then fine-tunes on NYU Depth v2 or KITTI for zero shot transfer." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.505, + 0.468, + 0.806 + ], + "angle": 0, + "content": "Diffusion-based Methods with Pretraining on Large Datasets. Recently, many techniques for SIDE has used diffusion architectures. These techniques [8, 16, 20, 38, 54], exploit prior knowledge acquired by pretraining on large datasets like LAION-400M [39], which consists of 400 million image-text pairs. In contrast, depth datasets, such as NYU Depth v2 and KITTI, contain around 20-30 thousand image-depth pairs. DepthGen [38] and DDP [16] work on a noise-to-depthmap paradigm and use images for conditional guidance of the diffusion process. DepthGen employs self-supervised pretraining on tasks like colorization, inpainting, and JPEG artifact removal, followed by supervised training on indoor and outdoor datasets [6, 12, 26, 27, 48]. DDP [16] decouples the image encoder and map decoder, allowing the image encoder to run just once, while the lightweight map decoder is run multiple times. VPD [54] and TADP [20] use denoising UNet [37] as a backbone to extract the rich features at multiple scales. Also, they utilize text instead of image for conditioning the diffusion backbone." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.81, + 0.468, + 0.901 + ], + "angle": 0, + "content": "Vision Transformer for Scene Understanding. The transformer architecture was initially proposed for NLP tasks [50], but introduced to the computer vision community as Vision Transformer (ViT) in [7]. Prior to this, CNNs dominated computer vision problems due to their ability to capture spatial hierarchies in image data. However, such archi" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.411, + 0.892, + 0.531 + ], + "angle": 0, + "content": "tectures were constrained due to their ability to learn spatially localized features only. Transformer architecture has a weaker inductive bias and allows ViT to learn long-range dependencies, and robust, generalizable features. ViT architectures have replaced CNNs for SOTA performance on most computer vision tasks in recent years. We use a pretrained ViT model for providing semantic information to the diffusion backbone in our model." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.562, + 0.719, + 0.58 + ], + "angle": 0, + "content": "3. Proposed Methodology" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.592, + 0.642, + 0.607 + ], + "angle": 0, + "content": "3.1. Preliminaries" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.621, + 0.892, + 0.698 + ], + "angle": 0, + "content": "Problem Formulation. The objective of single image depth prediction task is to predict continuous values, denoted as \\(\\mathbf{y} \\in \\mathbb{R}^{H \\times W}\\), for every pixel present in the input RGB image, \\(\\mathbf{x} \\in [0,255]^{3 \\times H \\times W}\\). Here \\(H\\) and \\(W\\) represent the height and width respectively of the input image." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.704, + 0.893, + 0.901 + ], + "angle": 0, + "content": "Diffusion Model. Diffusion models are a class of generative models that progressively inject noise into the input data (forward pass) and then learn to reconstruct the original data in a reverse denoising process (reverse pass). There are three formulations of diffusion models: denoising diffusion probabilistic models (DDPMs) [14], score-based generative models [43, 44], and those based on stochastic differential equations [45, 46]. DDPMs are of relevance to our paper, and are described below. The architecture of DDPMs consists of two Markov chains: a forward chain that adds noise to the data, and a reverse chain that converts noise back to data by learning transition kernels parameterized by deep neural networks. Formally, the forward pass is modeled as" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "28287" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.204, + 0.106 + ], + "angle": 0, + "content": "a Markov process:" + }, + { + "type": "equation", + "bbox": [ + 0.123, + 0.114, + 0.47, + 0.141 + ], + "angle": 0, + "content": "\\[\n\\mathbb {P} \\left(\\mathbf {z} _ {t} \\mid \\mathbf {z} _ {t - 1}\\right) = \\mathcal {N} \\left(\\mathbf {z} _ {t}; \\sqrt {1 - \\beta_ {t}} \\mathbf {z} _ {t - 1}, \\beta_ {t} \\boldsymbol {I}\\right). \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.148, + 0.47, + 0.209 + ], + "angle": 0, + "content": "Here \\(\\mathbf{z}_t\\) denotes the random variable at the \\(t^{\\mathrm{th}}\\) time step, \\(\\mathcal{N}(\\mathbf{z};\\boldsymbol {\\mu},\\boldsymbol {\\sigma})\\) denotes Gaussian probability distribution, and \\(\\beta_{t}\\) is the noise schedule. The above equation leads to the analytic form of \\(\\mathbb{P}(\\mathbf{z}_t\\mid \\mathbf{z}_0),\\forall t\\in \\{0,1,\\dots ,T\\}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.182, + 0.217, + 0.47, + 0.243 + ], + "angle": 0, + "content": "\\[\n\\mathbf {z} _ {t} = \\sqrt {\\bar {\\beta} _ {t}} \\mathbf {z} _ {0} + \\sqrt {1 - \\bar {\\beta} _ {t}} \\boldsymbol {\\epsilon}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.25, + 0.469, + 0.328 + ], + "angle": 0, + "content": "where \\(\\bar{\\beta}_t = \\prod_{s=1}^t \\beta_s\\) and \\(\\epsilon \\sim \\mathcal{N}(\\mathbf{0},\\mathbf{I})\\). The model then gradually removes noise by executing a learnable Markov chain in the reverse time direction, parameterized by a normal prior distribution \\(\\mathbb{P}(\\mathbf{z}_T) = \\mathcal{N}(\\mathbf{z}_T;0,\\mathbf{I})\\) and a learnable transition kernel \\(\\mathbb{P}_{\\theta}(\\mathbf{z}_{t-1}|\\mathbf{z}_t)\\) given by:" + }, + { + "type": "equation", + "bbox": [ + 0.1, + 0.335, + 0.469, + 0.352 + ], + "angle": 0, + "content": "\\[\n\\mathbb {P} _ {\\theta} \\left(\\mathbf {z} _ {t - 1} \\mid \\mathbf {z} _ {t}\\right) = \\mathcal {N} \\left(\\mathbf {z} _ {t - 1}; \\mu_ {\\theta} \\left(\\mathbf {z} _ {t}, t\\right), \\Sigma_ {\\theta} \\left(\\mathbf {z} _ {t}, t\\right)\\right). \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.36, + 0.469, + 0.481 + ], + "angle": 0, + "content": "The goal of the training process is to approximately match the reverse Markov chain with the actual time reversal of the forward Markov chain. Mathematically, parameter \\(\\theta\\) is adjusted so that the joint distribution of the reverse Markov chain \\(\\mathbb{P}_{\\theta}(\\mathbf{z}_0,\\mathbf{z}_1,\\dots ,\\mathbf{z}_T)\\coloneqq \\mathbb{P}(\\mathbf{z}_T)\\prod_{t = 1}^T\\mathbb{P}_{\\theta}(\\mathbf{z}_{t - 1}\\mid \\mathbf{z}_t)\\) closely approximates that of the forward Markov chain \\(\\mathbb{P}(\\mathbf{z}_0,\\mathbf{z}_1,\\dots ,\\mathbf{z}_T)\\coloneqq \\mathbb{P}(\\mathbf{z}_0)\\prod_{t = 1}^T\\mathbb{P}(\\mathbf{z}_t\\mid \\mathbf{z}_{t - 1})\\). This is achieved by minimizing the following loss:" + }, + { + "type": "equation", + "bbox": [ + 0.103, + 0.488, + 0.469, + 0.514 + ], + "angle": 0, + "content": "\\[\n\\mathbb {E} _ {t \\sim \\mathcal {U} [ 1, T ], \\mathbf {z} _ {0} \\sim \\mathbb {P} (\\mathbf {z} _ {0}), \\epsilon \\sim \\mathcal {N} (0, I)} \\left[ \\| \\epsilon - \\epsilon_ {\\theta} (\\mathbf {z} _ {t}, t) \\| ^ {2} \\right], \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.52, + 0.469, + 0.581 + ], + "angle": 0, + "content": "where \\(\\mathbf{z}_t\\) is computed using Eq. (2), and \\(\\epsilon_{\\theta}\\) is predicted using a neural network, typically a UNet architecture [37]. In a conditional diffusion model \\(\\epsilon_{\\theta}(\\mathbf{z}_t,t)\\) gets replaced by \\(\\epsilon_{\\theta}(\\mathbf{z}_t,t,\\mathcal{C})\\), where \\(\\mathcal{C}\\) is a conditioning variable." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.588, + 0.248, + 0.603 + ], + "angle": 0, + "content": "3.2. Our Architecture" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.612, + 0.47, + 0.852 + ], + "angle": 0, + "content": "Image Encoder and Latent Diffusion. Diffusion models typically take large number of time steps to train, and are difficult to converge. Recently, [36] proposed a new diffusion with improved convergence properties and is called \"stable-diffusion\". The key idea is to perform the diffusion in latent space, with latent embedding learnt separately through a variational autoencoder (VAE). The Encoder of VAE first transforms the input image \\(\\mathbf{x}\\) of size \\((H,W)\\) to latent space, then we follow latent diffusion formulation and utilize the UNet used in Stable Diffusion[36]. Utilizing latent diffusion formulation enables our architecture to capture multi-resolution features. Hence, we aggregate the feature maps from different layers of the UNet module (implementing conditional diffusion) by bringing them all to 1/4th resolution of the latent space, resulting in a feature map of size \\(8e\\times H / 32\\times W / 32\\)." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.856, + 0.469, + 0.902 + ], + "angle": 0, + "content": "Exploiting Semantic Context with Conditional Diffusion. Recall that we formulated the single image depth estimation as a dense regression problem, predicting \\(\\mathbb{P}(\\mathbf{y}\\mid \\mathbf{x})\\)" + }, + { + "type": "image", + "bbox": [ + 0.599, + 0.089, + 0.63, + 0.216 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.607, + 0.219, + 0.621, + 0.228 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.747, + 0.089, + 0.81, + 0.216 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.773, + 0.219, + 0.787, + 0.228 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.241, + 0.892, + 0.366 + ], + "angle": 0, + "content": "Figure 3. (a) Probabilistic graphical model corresponding to VPD. (b) The same corresponding to our formulation. Here, \\(\\mathcal{C}\\) represents the semantic embedding derived from our CIDE module. This embedding is internally generated by passing \\(\\mathbf{x}\\) through the ViT, resulting in \\(\\mathcal{E}\\). Subsequently, \\(\\mathcal{E}\\) undergoes further processing to yield \\(\\mathcal{C}\\), which is then utilized in the conditional diffusion module implementing \\(\\mathbb{P}(\\mathbf{z}_0\\mid \\mathbf{z}_t,\\mathcal{C})\\). The output of the conditional diffusion module is fed into the Depth Regressor module within our architecture, implementing \\(\\mathbb{P}(\\mathbf{y}\\mid \\mathbf{z}_0)\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.393, + 0.892, + 0.56 + ], + "angle": 0, + "content": "where \\(\\mathbf{y}\\) and \\(\\mathbf{x}\\) denote output depth, and input image respectively. In our diffusion formulation, we predict \\(\\mathbf{z}_0\\) which is then used to predict pixel-wise depth, \\(\\mathbf{y}\\). It has been shown that noise prediction in a diffusion model, \\(\\epsilon_{\\theta}\\), can be seen as predicting the gradient of the density function, \\(\\nabla_{\\mathbf{z}_t} \\log \\mathbb{P}(\\mathbf{z}_t)\\). Hence, overall architecture for SIDE using a diffusion architecture can be seen as factorizing the conditional probability, as shown in the probabilistic graph model in Fig. 3a. To utilize additional semantic context generated from a ViT model, we condition it on the ViT embeddings as shown in Fig. 3b. Mathematically, we model:" + }, + { + "type": "equation", + "bbox": [ + 0.514, + 0.588, + 0.876, + 0.605 + ], + "angle": 0, + "content": "\\[\n\\mathbb {P} (\\mathbf {y} \\mid \\mathbf {x}, \\mathcal {E}) = \\mathbb {P} (\\mathbf {y} \\mid \\mathbf {z} _ {0}) \\mathbb {P} (\\mathbf {z} _ {0} \\mid \\mathbf {z} _ {t}, \\mathcal {C}) \\mathbb {P} (\\mathbf {z} _ {t} \\mid \\mathbf {x}) \\mathbb {P} (\\mathcal {C} \\mid \\mathbf {x}),\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.609, + 0.545, + 0.621 + ], + "angle": 0, + "content": "where" + }, + { + "type": "equation", + "bbox": [ + 0.533, + 0.627, + 0.891, + 0.644 + ], + "angle": 0, + "content": "\\[\n\\mathbb {P} (\\mathcal {C} \\mid \\mathbf {x}) = \\mathbb {P} (\\mathcal {C} \\mid \\mathcal {E}) \\mathbb {P} (\\mathcal {E} \\mid \\mathbf {x}). \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.657, + 0.892, + 0.867 + ], + "angle": 0, + "content": "Here, the first term, \\(\\mathbb{P}(\\mathbf{y}\\mid \\mathbf{z}_0)\\), is implemented through the Depth Regressor module explained earlier. Similarly, \\(\\mathbb{P}(\\mathbf{z}_t\\mid \\mathbf{x})\\) is implemented using the VAE's Encoder as described earlier. We generate conditional information, \\(\\mathcal{C}\\) using our Comprehensive Image Detail Embedding module (hereafter CIDE, and explained below). The CIDE module takes \\(\\mathbf{x}\\) as input and generates embedding vector \\(\\mathcal{C}\\) (of dimension 768 in our design) as the output, thus, implementing \\(\\mathbb{P}(\\mathcal{C}\\mid \\mathbf{x})\\) given in Eq. (5). We use \\(\\mathcal{E}\\) to denote the embedding vector of a ViT module, and \\(\\mathbb{P}(\\mathcal{E}\\mid \\mathbf{x})\\) is implemented through the ViT. \\(\\mathbb{P}(\\mathcal{C}\\mid \\mathcal{E})\\) is implemented using downstream modules in CIDE consisting of learnable embeddings. The second term, \\(\\mathbb{P}(\\mathbf{z}_0\\mid \\mathbf{z}_t,\\mathcal{C})\\), is implemented using conditional diffusion," + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.871, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Comprehensive Image Detail Embedding (CIDE) Module. As described earlier, we believe using pseudo-captions" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "28288" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.089, + 0.895, + 0.149 + ], + "angle": 0, + "content": "Table 1. Results on Indoor NYU Depth v2 [27] Dataset. Results that are bold perform best. \\(\\uparrow\\) means the metric should be higher, \\(\\downarrow\\) indicate lower is better. The evaluation uses an upper bound of 10 meters on the ground truth depth map. All the numbers for other works have been taken from the corresponding papers. For MIM, and ZoEDepth we have used SwinV2-L 1K, and ZoeDepth-M12-N versions respectively. We see an overall improvement against SOTA on all the metrics used for evaluation." + }, + { + "type": "table", + "bbox": [ + 0.107, + 0.156, + 0.869, + 0.497 + ], + "angle": 0, + "content": "
MethodVenueAbs Rel↓RMSE↓log10↓Sq Rel↓δ1↑δ2↑δ3↑
Eigen et al.[9]NIPS'140.1580.641--0.7690.9500.988
DORN[10]CVPR'180.1150.5090.051-0.8280.9650.992
SharpNet[32]ICCV'190.1390.5020.047-0.8360.9660.993
Chen et al.[5]IJCAI-190.1110.5140.048-0.8780.9770.994
BTS[22]Arxiv'190.1100.3920.0470.0660.8850.9780.994
AdaBins[2]CVPR'210.1030.3640.044-0.9030.9840.997
DPT[33]ICCV'210.1100.3570.045-0.9040.9880.998
P3Depth[30]CVPR'220.1040.3560.043-0.8980.9810.996
NeWCRFs[53]CVPR'220.0950.3340.0410.0450.9220.9920.998
SwinV2-B[24]CVPR'220.1330.4620.059-0.8190.9750.995
SwinV2-L[24]CVPR'220.1120.3810.051-0.8860.9840.997
Localbins[3]ECCV'220.0990.3570.042-0.9070.9870.998
Jun et al.[17]ECCV'220.0980.3550.042-0.9130.9870.998
PixelFormer[1]WACV'230.0900.3220.0390.0430.9290.9910.998
DDP[16]ICCV'230.0940.3290.040-0.9210.9900.998
MIM [51]CVPR'230.0830.2870.035-0.9490.9940.999
AiT[28]ICCV'230.0760.2750.033-0.9540.9940.999
ZoeDepth [4]Arxiv'230.0750.2700.0320.0300.9550.9950.999
VPD[54]ICCV'230.0690.2540.0300.0270.9640.9950.999
OursCVPR'240.0590.2180.0260.0130.9780.9970.999
" + }, + { + "type": "image", + "bbox": [ + 0.16, + 0.508, + 0.813, + 0.827 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.837, + 0.897, + 0.867 + ], + "angle": 0, + "content": "Figure 4. Visual Comparison on NYU Depth v2 Indoor Dataset. Note, our method's ability to delineate objects in terms of their depth, such as the table lamp in Row 5, even when such information is absent from the ground truth depth map." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.946, + 0.522, + 0.957 + ], + "angle": 0, + "content": "28289" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.089, + 0.895, + 0.175 + ], + "angle": 0, + "content": "Table 2. Performance on the Outdoor KITTI [12] Dataset. Please refer to the caption of Tab. 1 for notation details. For ZoEDepth results, we use the ZoeDepth-M12-K version following the authors' recommendation. In instances where results for certain methods were not reported in the respective works, denoted by “-”, and the code is unavailable, we were unable to generate the missing numbers. Despite the saturation of results on the outdoor KITTI dataset, our method consistently achieves comparable or superior performance to the state-of-the-art (SOTA) across all metrics. VPD[54] cannot be trained on KITTI dataset as it required a per image text label which is not present in KITTI. The symbol \\(\\dagger\\) indicates that the method utilizes additional information beyond RGB." + }, + { + "type": "table", + "bbox": [ + 0.085, + 0.184, + 0.889, + 0.451 + ], + "angle": 0, + "content": "
MethodVenueAbs Rel↓Sq Rel↓RMSElog ↓RMSE↓δ1 ↑δ2 ↑δ3 ↑
Eigen et al.[9]NIPS'140.2031.5170.2826.3070.7020.8980.967
DORN[10]CVPR'180.0720.3070.1202.7270.9320.9840.994
BTS[22]Arxiv'190.0590.2410.0962.7560.9560.9930.998
AdaBins[2]CVPR'210.0670.1900.0882.9600.9490.9920.998
DPT[33]ICCV'210.060-0.0922.5730.9590.9950.996
P3Depth[30]CVPR'220.0710.2700.1032.8420.9530.9930.998
NeWCRFs[53]CVPR'220.0520.1550.0792.1290.9740.9970.999
PixelFormer[1]WACV'230.0510.1490.0772.0810.9760.9970.999
ZoeDepth [4]Arxiv'230.0540.1890.0832.4400.970.9960.999
DDP [16]ICCV'230.0500.1480.0762.0720.9750.9970.999
URCDC [40]ToM'230.0500.1420.0762.0320.9770.9970.999
IEBins [41]NeurIPS'230.0500.1420.0752.0110.9780.9980.999
MIM [51]CVPR'230.0500.1390.0751.9660.9770.9981.000
GEDepth[52]†ICCV'230.0480.1420.0762.0440.9760.9970.999
OursCVPR'240.0480.1390.0742.0390.9790.9981.000
" + }, + { + "type": "image", + "bbox": [ + 0.08, + 0.462, + 0.891, + 0.707 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.303, + 0.715, + 0.666, + 0.73 + ], + "angle": 0, + "content": "Figure 5. Visual Comparison on KITTI Outdoor Dataset." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.757, + 0.473, + 0.892 + ], + "angle": 0, + "content": "to generate the semantic context has limited utility, as the textual descriptions typically focus on large salient objects only. Instead we propose our CIDE module which use embeddings from a pre-trained ViT, and extract detailed semantic context from these embeddings. For this we take 1000 dimensional logit vector from ViT, and pass it through a two layer MLP which converts it to a 100 dimensional vector. Subsequently, we employ this vector to compute the linear combination of 100 learnable embeddings." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.757, + 0.892, + 0.801 + ], + "angle": 0, + "content": "This resulting embedding undergoes a linear transformation to yield a semantic context vector of dimension 768, which is then passed to conditional diffusion module." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.81, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Depth Regressor. The output feature map undergoes through an Upsampling Decoder, comprised of deconvolution layers, followed by a Depth Regressor. The Depth Regressor is essentially a two-layer convolutional neural network (CNN), with the initial layer having dimensions \\(\\mathrm{Conv}(3\\times 3)\\), 192, and the subsequent layer \\(\\mathrm{Conv}(3\\times 3)\\), 1." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "28290" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.089, + 0.892, + 0.146 + ], + "angle": 0, + "content": "Table 3. Quantitative results for zero-shot transfer to four unseen indoor datasets. mRI denotes the mean relative improvement with respect to NeWCRFs across all metrics (δ1, REL, RMSE). Evaluation depth is capped at 8m for SUN RGB-D, 10m for iBims and DIODE Indoor, and 80m for HyperSim. Best results are in bold, second best are underlined. Our mRI outperforms all methods across all datasets by a large margin. † denotes that ZoeD is trained on 12 datasets and our method is trained only on NYUv2." + }, + { + "type": "table", + "bbox": [ + 0.078, + 0.158, + 0.915, + 0.288 + ], + "angle": 0, + "content": "
MethodSUN RGB-DiBims-1 BenchmarkDIODE IndoorHyperSim
δ1↑REL↓RMSE↓mRIθ↑δ1↑REL↓RMSE↓mRIθ↑δ1↑REL↓RMSE↓mRIθ↑δ1↑REL↓RMSE↓mRIθ↑
BTS [22]0.7400.1720.515-14.2%0.5380.2310.919-6.9%0.2100.4181.9052.3%0.2250.4766.404-8.6%
AdaBins [2]0.7710.1590.476-7.0%0.5550.2120.901-2.1%0.1740.4431.963-7.2%0.2210.4836.546-10.5%
LocalBins [3]0.7770.1560.470-5.6%0.5580.2110.880-0.7%0.2290.4121.8537.1%0.2340.4686.362-6.6%
NeWCRFs [53]0.7980.1510.4240.0%0.5480.2060.8610.0%0.1870.4041.8670.0%0.2550.4426.0170.0%
VPD [54]0.8610.1210.35514.7%0.6270.1870.76711.5%0.4800.3921.29563.4%0.3330.5315.1118.5%
ZoeD-M12-N† [4]0.8640.1190.34616.0%0.6580.1690.71118.5%0.3760.3271.58845.0%0.2920.4105.7718.6%
Ours0.8850.1120.31920.5%0.6880.1630.66423.1%0.5450.3441.16481.3%0.3940.4424.73925.2%
" + }, + { + "type": "image", + "bbox": [ + 0.137, + 0.306, + 0.847, + 0.462 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.137, + 0.479, + 0.846, + 0.639 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.652, + 0.89, + 0.694 + ], + "angle": 0, + "content": "Figure 6. Visualization of improvements over VPD [54] in our model due to ViT embeddings passed down as conditional vectors (in blue and green). In the above images, ViT detects the desktop computer (first image) and table lamp (second image) with high probability, and they are thus better detected by our model. Additional visualizations are provided in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.719, + 0.312, + 0.736 + ], + "angle": 0, + "content": "4. Experiments and Results" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.471, + 0.901 + ], + "angle": 0, + "content": "Datasets and Evaluation. We use NYU Depth v2 [27] and KITTI [12] as the primary datasets for training. The NYU Depth v2 dataset is a widely used indoor benchmark for monocular depth estimation, containing over 24k densely labeled pairs of RGB and depth images in the train set and 654 in the test set. The dataset covers a wide range of indoor scenes and includes challenging scenarios such as reflective surfaces, transparent objects, and occlusions. The ground truth depth maps are obtained using a structured light sensor and are provided at a resolution of \\(640 \\times 480\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.892, + 0.888 + ], + "angle": 0, + "content": "KITTI dataset is a widely used outdoor benchmark for monocular depth estimation, containing over 24k densely labeled pairs of RGB and depth images. The dataset covers outdoor driving scenarios and includes varying lighting conditions, weather, and occlusions. The ground truth depth maps are obtained using a Velodyne LiDAR sensor and are provided at a resolution of \\(1242 \\times 375\\). To demonstrate generalizability, we evaluate zero-shot performance on the following datasets: Sun-RGBD [42], iBims1 [19], DIODE [49], and HyperSim [35]. Whereas, the main paper contains mostly quantitative, and only some representative" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.946, + 0.518, + 0.957 + ], + "angle": 0, + "content": "28291" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.089, + 0.472, + 0.201 + ], + "angle": 0, + "content": "Table 4. Effectiveness of different embeddings for guiding the diffusion process for depth estimation on NYU Depth v2 dataset. In the third row, rather than using pseudo caption embeddings generated from the scene label (as implemented by VPD [54]), we provide the one-hot vector representing the scene label as a condition to the diffusion model. We observe a slight improvement in the metrics, highlighting that information content in the caption is similar to that in one-hot label." + }, + { + "type": "table", + "bbox": [ + 0.084, + 0.212, + 0.465, + 0.305 + ], + "angle": 0, + "content": "
EmbeddingsRMSE↓Abs Rel↓δ1↑
Scene label emb. [54]0.2540.0690.964
Text caption emb. [20]0.2250.0620.976
One-hot vector emb.0.2440.0670.968
Proposed scene emb.0.2180.0590.978
" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.33, + 0.47, + 0.361 + ], + "angle": 0, + "content": "tive visual results, detailed visual results on each dataset are included in the supplementary." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.363, + 0.47, + 0.545 + ], + "angle": 0, + "content": "Implementation Details. Our model is implemented using PyTorch [29]. For optimization, we have used AdamW optimizer [25] with \\(\\beta_0\\) values of 0.9 and 0.999, a batch size of 32, and a weight decay of 0.1. We train our model for 25 epochs for both KITTI and NYU Depth v2 datasets, with an initial learning rate of \\(3 \\times 10^{-5}\\). We first linearly increase learning rate to \\(5 \\times 10^{-4}\\), and then linearly decrease across training iterations. We use usual data augmentation techniques, including random hue addition, horizontal flipping, changing the image brightness, and Cut Depth [15]. Our model takes approximately 21 minutes per epoch to train using 8 NVIDIA A100 GPUs." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.555, + 0.396, + 0.57 + ], + "angle": 0, + "content": "4.1. Comparison on Benchmark Datasets" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.578, + 0.469, + 0.683 + ], + "angle": 0, + "content": "Comparison on NYU Depth v2. Tab. 1 shows the comparison of our proposed method with SOTA methods on the indoor NYU Depth v2 dataset [27]. We achieve a new state of the art on this dataset. Our methods perform better than the previous SOTA ([54])) by a large margin of \\(14\\%\\) in terms of RMSE. Fig. 4 provides a qualitative comparison on the dataset." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.686, + 0.469, + 0.779 + ], + "angle": 0, + "content": "Comparison on KITTI. Tab. 2 shows the comparison with various methods on the KITTI dataset [12]. Fig. 5 shows the qualitative results. Unlike NYU Depth v2 dataset, KITTI is an outdoor dataset. On this dataset also, we achieve similar or better performance than all existing state-of-the-art techniques." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.787, + 0.412, + 0.803 + ], + "angle": 0, + "content": "4.2. Generalization and Zero Shot Transfer" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.81, + 0.469, + 0.903 + ], + "angle": 0, + "content": "Unlike state of the art on zero short transfer (ZoEDepth [4]), which requiring training on many datasets (12 in their case) for effective zero shot transfer, we show that our model generalizes well to other unseen dataset even when trained on a single NYU Depth v2 dataset. Tab. 3 shows quantitative results to back our findings." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.091, + 0.655, + 0.108 + ], + "angle": 0, + "content": "4.3. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.114, + 0.892, + 0.312 + ], + "angle": 0, + "content": "Effect of Contextual Information. As highlighted in the motivation, a key observation of this study is the richness of information contained within the output probability vector of the ViT, surpassing the textual embeddings employed in current state-of-the-art (e.g., VPD [54]). To compare with the utilization of scene label information (as implemented by VPD), we construct the conditioning embedding as a one-hot vector using the scene label, and subsequently transform it using an MLP. As illustrated in Tab. 4, we observe a slight improvement over VPD, indicating that the information content in pseudo-captions resembles that of one-hot labels. Furthermore, our proposed method surpasses both the approaches." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.313, + 0.892, + 0.466 + ], + "angle": 0, + "content": "Qualitative Results. Perhaps the most important task would be to verify that use of rich probability vector instead of text embeddings actually results in an improvement in depth as a direct consequence. We do this by considering the top few objects predicted by the ViT and correlating this with the depth predicted at these objects. We hypothesise when a particular class say dog is predicted with a high probability (hence there must be a dog in the image), then the corresponding depth must also be more accurate. We show this in Fig. 6." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.473, + 0.619, + 0.488 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.498, + 0.892, + 0.725 + ], + "angle": 0, + "content": "We presented a new architecture block Comprehensive Image Detail Embedding (CIDE) module for robust monocular depth estimation in this paper. Our key idea is to highlight the limitations of using pseudo-captions to provide contextual information, and instead propose to use richer class-wise probability generated by a classification model, such as ViT. The motivation is that, while textual embedding typically highlights salient objects, class-wise probability vector preserves more details, including smaller objects in the background also. We implement the idea using proposed CIDE module cascaded with a conditional diffusion pipeline for monocular depth estimation. We demonstrated the effectiveness of our approach on several benchmark datasets and showed that it outperforms SOTA methods by a significant margin." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.728, + 0.892, + 0.835 + ], + "angle": 0, + "content": "Acknowledgement. We thank Kartik Anand for assistance with some of the experiments. This work has been supported by funding through Staqu Technologies Pvt. Ltd., and Department of Science and Technology (DST), Government of India. The travel for Suraj Patni is supported through Research Acceleration funds of Department of Computer Science and Engineering, IIT Delhi." + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.848, + 0.597, + 0.864 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.507, + 0.873, + 0.892, + 0.903 + ], + "angle": 0, + "content": "[1] Ashutosh Agarwal and Chetan Arora. Attention attention everywhere: Monocular depth prediction with skip attention." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "28292" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.092, + 0.47, + 0.134 + ], + "angle": 0, + "content": "In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 5861-5870, 2023. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.136, + 0.472, + 0.205 + ], + "angle": 0, + "content": "[2] Shariq Farooq Bhat, Ibrahim Alhashim, and Peter Wonka. Adabins: Depth estimation using adaptive bins. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4009-4018, 2021. 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.207, + 0.472, + 0.262 + ], + "angle": 0, + "content": "[3] Shariq Farooq Bhat, Ibraheem Alhashim, and Peter Wonka. Localbins: Improving depth estimation by learning local distributions. In European Conference on Computer Vision, pages 480-496. Springer, 2022. 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.264, + 0.472, + 0.305 + ], + "angle": 0, + "content": "[4] Shariq Farooq Bhat, Reiner Birkl, Diana Wofk, Peter Wonka, and Matthias Müller. Zoedepth: Zero-shot transfer by combining relative and metric depth, 2023. 2, 3, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.307, + 0.472, + 0.348 + ], + "angle": 0, + "content": "[5] Xiaotian Chen, Xuejin Chen, and Zheng-Jun Zha. Structure-aware residual pyramid network for monocular depth estimation. arXiv preprint arXiv:1907.06023, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.35, + 0.472, + 0.419 + ], + "angle": 0, + "content": "[6] Angela Dai, Angel X Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5828-5839, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.421, + 0.472, + 0.502 + ], + "angle": 0, + "content": "[7] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.504, + 0.472, + 0.545 + ], + "angle": 0, + "content": "[8] Yiqun Duan, Xianda Guo, and Zheng Zhu. Diffusiondepth: Diffusion denoising approach for monocular depth estimation, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.548, + 0.472, + 0.602 + ], + "angle": 0, + "content": "[9] David Eigen, Christian Puhrsch, and Rob Fergus. Depth map prediction from a single image using a multi-scale deep network. Advances in neural information processing systems, 27, 2014. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.604, + 0.472, + 0.673 + ], + "angle": 0, + "content": "[10] Huan Fu, Mingming Gong, Chaohui Wang, Kayhan Bat-manghelich, and Dacheng Tao. Deep ordinal regression network for monocular depth estimation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2002-2011, 2018. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.676, + 0.472, + 0.744 + ], + "angle": 0, + "content": "[11] Huan Fu, Mingming Gong, Chaohui Wang, Kayhan Bat-manghelich, and Dacheng Tao. Deep ordinal regression network for monocular depth estimation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2002-2011, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.746, + 0.472, + 0.8 + ], + "angle": 0, + "content": "[12] Andreas Geiger, Philip Lenz, Christoph Stiller, and Raquel Urtasun. Vision meets robotics: The kitti dataset. International Journal of Robotics Research (IJRR), 2013. 3, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.803, + 0.472, + 0.857 + ], + "angle": 0, + "content": "[13] José L. Herrera, Carlos R. del Blanco, and Narciso García. Automatic depth extraction from 2d images using a cluster-based learning framework. IEEE Transactions on Image Processing, 27(7):3288-3299, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.859, + 0.472, + 0.902 + ], + "angle": 0, + "content": "[14] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 3" + }, + { + "type": "list", + "bbox": [ + 0.079, + 0.092, + 0.472, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.893, + 0.134 + ], + "angle": 0, + "content": "[15] Yasunori Ishii and Takayoshi Yamashita. Cutdepth: Edge-aware data augmentation in depth estimation. arXiv preprint arXiv:2107.07684, 2021. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.136, + 0.893, + 0.19 + ], + "angle": 0, + "content": "[16] Yuanfeng Ji, Zhe Chen, Enze Xie, Lanqing Hong, Xihui Liu, Zhaoqiang Liu, Tong Lu, Zhenguo Li, and Ping Luo. Ddp: Diffusion model for dense visual prediction. arXiv preprint arXiv:2303.17559, 2023. 2, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.193, + 0.893, + 0.248 + ], + "angle": 0, + "content": "[17] Jinyoung Jun, Jae-Han Lee, Chul Lee, and Chang-Su Kim. Depth map decomposition for monocular depth estimation. In European Conference on Computer Vision, pages 18-34. Springer, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.25, + 0.893, + 0.305 + ], + "angle": 0, + "content": "[18] Kevin Karsch, Ce Liu, and Sing Bing Kang. Depth transfer: Depth extraction from video using non-parametric sampling. IEEE transactions on pattern analysis and machine intelligence, 36(11):2144-2158, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.307, + 0.893, + 0.374 + ], + "angle": 0, + "content": "[19] Tobias Koch, Lukas Liebel, Friedrich Fraundorfer, and Marco Korner. Evaluation of cnn-based single-image depth estimation methods. In Proceedings of the European Conference on Computer Vision (ECCV) Workshops, pages 0–0, 2018. 1, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.377, + 0.893, + 0.432 + ], + "angle": 0, + "content": "[20] Neehar Kondapaneni, Markus Marks, Manuel Knott, Rogério Guimarães, and Pietro Perona. Text-image alignment for diffusion-based perception. arXiv preprint arXiv:2310.00031, 2023. 2, 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.434, + 0.893, + 0.502 + ], + "angle": 0, + "content": "[21] Iro Laina, Christian Rupprecht, Vasileios Belagiannis, Federico Tombari, and Nassir Navab. Deeper depth prediction with fully convolutional residual networks. In 2016 Fourth international conference on 3D vision (3DV), pages 239-248. IEEE, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.504, + 0.893, + 0.559 + ], + "angle": 0, + "content": "[22] Jin Han Lee, Myung-Kyu Han, Dong Wook Ko, and Il Hong Suh. From big to small: Multi-scale local planar guidance for monocular depth estimation. arXiv preprint arXiv:1907.10326, 2019. 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.561, + 0.893, + 0.63 + ], + "angle": 0, + "content": "[23] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In International Conference on Machine Learning, pages 12888-12900. PMLR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.632, + 0.893, + 0.702 + ], + "angle": 0, + "content": "[24] Ze Liu, Han Hu, Yutong Lin, Zhuliang Yao, Zhenda Xie, Yixuan Wei, Jia Ning, Yue Cao, Zheng Zhang, Li Dong, et al. Swin transformer v2: Scaling up capacity and resolution. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12009-12019, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.704, + 0.893, + 0.731 + ], + "angle": 0, + "content": "[25] Ilya Loshchilov and Frank Hutter. Decoupled Weight Decay Regularization. In ICML 2019. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.733, + 0.893, + 0.788 + ], + "angle": 0, + "content": "[26] John McCormac, Ankur Handa, Stefan Leutenegger, and Andrew J Davison. Scelenet rgb-d: 5m photorealistic images of synthetic indoor trajectories with ground truth. arXiv preprint arXiv:1612.05079, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.789, + 0.893, + 0.83 + ], + "angle": 0, + "content": "[27] Pushmeet Kohli Nathan Silberman, Derek Hoiem and Rob Fergus. Indoor segmentation and support inference from rgbd images. In ECCV, 2012. 3, 5, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.832, + 0.893, + 0.902 + ], + "angle": 0, + "content": "[28] Jia Ning, Chen Li, Zheng Zhang, Chunyu Wang, Zigang Geng, Qi Dai, Kun He, and Han Hu. All in tokens: Unifying output space of visual tasks via soft token. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 19900-19910, 2023. 2, 5" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.893, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "28293" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.217 + ], + "angle": 0, + "content": "[29] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, Alban Desmaison, Andreas Kopf, Edward Yang, Zachary DeVito, Martin Raison, Alykhan Tejani, Sasank Chilamkurthy, Benoit Steiner, Lu Fang, Junjie Bai, and Soumith Chintala. Pytorch: An imperative style, high-performance deep learning library. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2019. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.219, + 0.47, + 0.288 + ], + "angle": 0, + "content": "[30] Vaishakh Patil, Christos Sakaridis, Alexander Liniger, and Luc Van Gool. P3depth: Monocular depth estimation with a piecewise planarity prior. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1610–1621, 2022. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.29, + 0.47, + 0.373 + ], + "angle": 0, + "content": "[31] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.375, + 0.469, + 0.444 + ], + "angle": 0, + "content": "[32] Michael Ramamonjisoa and Vincent Lepetit. Sharpnet: Fast and accurate recovery of occluding contours in monocular depth estimation. In Proceedings of the IEEE/CVF International Conference on Computer Vision Workshops, pages 0–0, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.446, + 0.471, + 0.502 + ], + "angle": 0, + "content": "[33] René Ranftl, Alexey Bochkovskiy, and Vladlen Koltun. Vision transformers for dense prediction. In Proceedings of the IEEE/CVF international conference on computer vision, pages 12179-12188, 2021. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.504, + 0.47, + 0.574 + ], + "angle": 0, + "content": "[34] René Ranftl, Katrin Lasinger, David Hafner, Konrad Schindler, and Vladlen Koltun. Towards robust monocular depth estimation: Mixing datasets for zero-shot cross-dataset transfer. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(3), 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.575, + 0.469, + 0.657 + ], + "angle": 0, + "content": "[35] Mike Roberts, Jason Ramapuram, Anurag Ranjan, Atulit Kumar, Miguel Angel Bautista, Nathan Paczan, Russ Webb, and Joshua M. Susskind. Hypersim: A photorealistic synthetic dataset for holistic indoor scene understanding. In International Conference on Computer Vision (ICCV) 2021, 2021. 1, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.66, + 0.47, + 0.73 + ], + "angle": 0, + "content": "[36] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.731, + 0.47, + 0.815 + ], + "angle": 0, + "content": "[37] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical Image Computing and Computer-Assisted Intervention-MICCAI 2015: 18th International Conference, Munich, Germany, October 5-9, 2015, Proceedings, Part III 18, pages 234-241. Springer, 2015. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.817, + 0.469, + 0.857 + ], + "angle": 0, + "content": "[38] Saurabh Saxena, Abhishek Kar, Mohammad Norouzi, and David J Fleet. Monocular depth estimation using diffusion models. arXiv preprint arXiv:2302.14816, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.859, + 0.469, + 0.9 + ], + "angle": 0, + "content": "[39] Christoph Schuhmann, Richard Vencu, Romain Beaumont, Robert Kaczmarczyk, Clayton Mullis, Aarush Katta, Theo Coombes, Jenia Jitsev, and Aran Komatsuzaki. Laion-400m:" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.471, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.89, + 0.12 + ], + "angle": 0, + "content": "Open dataset of clip-filtered 400 million image-text pairs. arXiv preprint arXiv:2111.02114, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.121, + 0.892, + 0.176 + ], + "angle": 0, + "content": "[40] Shuwei Shao, Zhongcai Pei, Weihai Chen, Ran Li, Zhong Liu, and Zhengguo Li. Urcdc-depth: Uncertainty rectified cross-distillation with cutflip for monocular depth estimation. arXiv preprint arXiv:2302.08149, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.177, + 0.892, + 0.232 + ], + "angle": 0, + "content": "[41] Shuwei Shao, Zhongcai Pei, Xingming Wu, Zhong Liu, Weihai Chen, and Zhengguo Li. Iebins: Iterative elastic bins for monocular depth estimation. arXiv preprint arXiv:2309.14137, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.233, + 0.892, + 0.288 + ], + "angle": 0, + "content": "[42] Shuran Song, Samuel P. Lichtenberg, and Jianxiong Xiao. Sun rgb-d: A rgb-d scene understanding benchmark suite. In 2015 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 567-576, 2015. 1, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.289, + 0.892, + 0.329 + ], + "angle": 0, + "content": "[43] Yang Song and Stefano Ermon. Generative modeling by estimating gradients of the data distribution. Advances in neural information processing systems, 32, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.33, + 0.892, + 0.372 + ], + "angle": 0, + "content": "[44] Yang Song and Stefano Ermon. Improved techniques for training score-based generative models. Advances in neural information processing systems, 33:12438-12448, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.372, + 0.892, + 0.427 + ], + "angle": 0, + "content": "[45] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. arXiv preprint arXiv:2011.13456, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.428, + 0.892, + 0.482 + ], + "angle": 0, + "content": "[46] Yang Song, Conor Durkan, Iain Murray, and Stefano Ermon. Maximum likelihood training of score-based diffusion models. Advances in Neural Information Processing Systems, 34: 1415-1428, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.484, + 0.892, + 0.537 + ], + "angle": 0, + "content": "[47] M. Sun, A. Y. Ng, and A. Saxena. Make3d: Learning 3d scene structure from a single still image. IEEE Transactions on Pattern Analysis and Machine Intelligence, 31(05):824-840, 2009. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.539, + 0.892, + 0.622 + ], + "angle": 0, + "content": "[48] Pei Sun, Henrik Kretzschmar, Xerxes Dotiwalla, Aurelien Chouard, Vijaysai Patnaik, Paul Tsui, James Guo, Yin Zhou, Yuning Chai, Benjamin Caine, et al. Scalability in perception for autonomous driving: Waymo open dataset. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2446-2454, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.623, + 0.892, + 0.692 + ], + "angle": 0, + "content": "[49] Igor Vasiljevic, Nick Kolkin, Shanyi Zhang, Ruotian Luo, Haochen Wang, Falcon Z Dai, Andrea F Daniele, Mohammadreza Mostajabi, Steven Basart, Matthew R Walter, et al. Diode: A dense indoor and outdoor depth dataset. arXiv preprint arXiv:1908.00463, 2019. 1, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.692, + 0.892, + 0.748 + ], + "angle": 0, + "content": "[50] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.749, + 0.892, + 0.816 + ], + "angle": 0, + "content": "[51] Zhenda Xie, Zigang Geng, Jingcheng Hu, Zheng Zhang, Han Hu, and Yue Cao. Revealing the dark secrets of masked image modeling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14475-14485, 2023. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.818, + 0.892, + 0.872 + ], + "angle": 0, + "content": "[52] Xiaodong Yang, Zhuang Ma, Zhiyu Ji, and Zhe Ren. Gedgepth: Ground embedding for monocular depth estimation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12719-12727, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.873, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[53] Weihao Yuan, Xiaodong Gu, Zuozhuo Dai, Siyu Zhu, and Ping Tan. Newcrfs: Neural window fully-connected crfs for" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "28294" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.11, + 0.092, + 0.47, + 0.134 + ], + "angle": 0, + "content": "monocular depth estimation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2022. 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.136, + 0.47, + 0.188 + ], + "angle": 0, + "content": "[54] Wenliang Zhao, Yongming Rao, Zuyan Liu, Benlin Liu, Jie Zhou, and Jiwen Lu. Unleashing text-to-image diffusion models for visual perception. ICCV, 2023. 2, 3, 5, 6, 7, 8" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.188 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.518, + 0.957 + ], + "angle": 0, + "content": "28295" + } + ] +] \ No newline at end of file diff --git a/2024/ECoDepth_ Effective Conditioning of Diffusion Models for Monocular Depth Estimation/7e3c5935-f184-4aa8-a48b-682e7477329f_origin.pdf b/2024/ECoDepth_ Effective Conditioning of Diffusion Models for Monocular Depth Estimation/7e3c5935-f184-4aa8-a48b-682e7477329f_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..504292e18881c511abe3a30d7d1d7594ef7a0cb5 --- /dev/null +++ b/2024/ECoDepth_ Effective Conditioning of Diffusion Models for Monocular Depth Estimation/7e3c5935-f184-4aa8-a48b-682e7477329f_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:82db315fa3070134234835799ec765d2db6701c5a0886c590e7f9dd7788b348d +size 3204672 diff --git a/2024/ECoDepth_ Effective Conditioning of Diffusion Models for Monocular Depth Estimation/full.md b/2024/ECoDepth_ Effective Conditioning of Diffusion Models for Monocular Depth Estimation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..23734d5a289c13c5f6b7b3ff58b330a8eced882a --- /dev/null +++ b/2024/ECoDepth_ Effective Conditioning of Diffusion Models for Monocular Depth Estimation/full.md @@ -0,0 +1,268 @@ +# ECoDepth: Effective Conditioning of Diffusion Models for Monocular Depth Estimation + +Suraj Patni* + +Aradhye Agarwal* + +Chetan Arora + +Indian Institute of Technology Delhi + +https://ecodepth-iitd.github.io + +# Abstract + +In the absence of parallax cues, a learning based single image depth estimation (SIDE) model relies heavily on shading and contextual cues in the image. While this simplicity is attractive, it is necessary to train such models on large and varied datasets, which are difficult to capture. It has been shown that using embeddings from pretrained foundational models, such as CLIP, improves zero shot transfer in several applications. Taking inspiration from this, in our paper we explore the use of global image priors generated from a pre-trained ViT model to provide more detailed contextual information. We argue that the embedding vector from a ViT model, pre-trained on a large dataset, captures greater relevant information for SIDE than the usual route of generating pseudo image captions, followed by CLIP based text embeddings. Based on this idea, we propose a new SIDE model using a diffusion backbone which is conditioned on ViT embeddings. Our proposed design establishes a new state-of-the-art (SOTA) for SIDE on NYU Depth v2 dataset, achieving Abs Rel error of 0.059(14% improvement) compared to 0.069 by the current SOTA (VPD). And on KITTI dataset, achieving Sq Rel error of 0.139 (2% improvement) compared to 0.142 by the current SOTA (GED). For zero shot transfer with a model trained on NYU Depth v2, we report mean relative improvement of (20%, 23%, 81%, 25%) over NeWCRF on (Sun-RGBD, iBims1, DIODE, HyperSim) datasets, compared to (16%, 18%, 45%, 9%) by ZoEDepth. The code is available in our project page. + +# 1. Introduction + +Single Image Depth Estimation (SIDE) is the task of predicting per pixel depth using a single RGB image from a monocular camera. It is a fundamental problem in computer vision with applications in several domains, viz robotics, autonomous driving, and augmented reality. The problem is + +![](images/f7274f44c4d22bc9eae8a19f69aa92c1c20646d54843a02397296650e2bfc12a.jpg) +(a) Sun-RGBD [42] + +![](images/214cfea1c3468673bbbfb8b1f65db15a952fafc13969355b377513969847d17f.jpg) + +![](images/ff8d25236e5638817716f97d36604908ba6a66250a0304d48025d379b65c2095.jpg) + +![](images/f56327bab79fcd3f27c08fbef298e0b0e23fd81917afe234efe190f0d7e73e38.jpg) + +![](images/acfee1ae7e238a4d60c7914c77d3468b68bc4fb10c24754766404c5d2a5dd4a9.jpg) +(b) iBims1 [19] + +![](images/b6c9dc2b7d14f841f908bf571798f99779b924830b846d5fe98305823fd3225f.jpg) + +![](images/5a9b6186ef62db411c91c5843145d756b45f98d60f13a0e2aa97b4f3d446142e.jpg) +(c) DIODE [49] +(d) HyperSim [35] +Figure 1. Qualitative results across four different datasets, demonstrating the zero-shot performance of our model trained only on the NYU Depth v2 dataset. Corresponding quantitative results are presented in Table 3. The first column displays RGB images, the second column depicts ground truth depth, and the third column showcases our model's predicted depths. Additional images for each dataset are available in the Supplementary Material. + +![](images/8d5070d4cbc6c4ebcfef4196245823ef4f87ec26324f3274b07d2d886dd1ddc2.jpg) + +typically formulated in two flavors: metric depth estimation (MDE), and relative depth estimation (RDE). As the names suggest, MDE deals with estimation in physical units such as meters, whereas, RDE techniques focus on relative depth only, and require a per-image affine transformation as post-processing to convert to a physical unit. + +Conventional geometric techniques for depth estimation typically rely on feature correspondence, parallax, and triangulation from two or more views. However, the problem becomes ill-posed for estimation from a single view. Intuitively, depth map is a 3D representation of the scene, whereas an RGB image is a 2D projection of the scene. Hence, it cannot be uniquely determined from a single RGB image. Therefore, learning-based SIDE models rely on visual cues like 'shape from shading' and other contextual priors for per-pixel depth prediction. + +A data-driven approach makes the SIDE pipeline sim + +pler, but also makes the learnt model dependent on the quality of training data. It has been observed that such models overfit on a particular training distribution/domain and fail to generalize on unseen data. This is especially true for MDE models when the range of depth in the training dataset is limited. Hence, training on multiple datasets, with wide variations in depth ranges has been proposed [4, 34]. + +On the other hand, development of large foundational models (LFMs) in recent years has altered the preferred design approach for many computer vision problems. These huge models are trained using extensive datasets of unlabeled images and learning objectives that are agnostic to specific tasks. The learnt embeddings from such pre-trained models have been shown to help generalization and zero-shot transfer in many applications. We are aware of at least two works for MDE in SIDE problem that have appeared in the last few months and make use of such foundational models. VPD [54] uses a text-to-image diffusion model pretrained on LAION-400M [39] dataset having large-scale image-text pairs as the backbone. The model prompts the denoising UNet with textual inputs to make the visual contents interact with the text prompts. Since the problem formulation doesn't include text description as the input, the model generates simple descriptions such as 'A photo of a {scene name}' based on the scene label given in the NYU Depth v2 dataset. Another work, TADP [20] improves upon VPD. Instead of simple image descriptions, TADP uses BLIP-2 [23] to generate image captions, and then uses CLIP [31] embeddings of the pseudo-caption to condition the diffusion model. + +We view the above works as providing robust semantic context to LFMs for the actual task at hand, which helps in visual recognition in general, as well as SIDE. While we do agree with the broad motivations of these works, the question that we ask is if pseudo-captions are the most effective way to provide the semantic context. Textual descriptions of an image typically focus on large salient objects and emphasize on their relationships. On the other hand, large vision models for image classification, typically contain representation for even smaller objects present in the scene. Even when a single object is present in the scene, the representations typically capture uncertainties and ambiguities inherent in the scene. Hence, we posit that using embeddings from a transformer model, pretrained on a large image dataset unrelated to the SIDE task, captures more relevant information, and is a better alternative to using pseudocaptions' CLIP embedding. Since diffusion-based models have shown their superiority for the dense prediction tasks in recent works [8, 16, 20, 38, 54]. Hence, we propose a diffusion backbone for our model, along with a novel CIDE module which employs ViT [7] to extract semantic context embeddings. These embeddings are subsequently utilized to condition the diffusion backbone. + +Contributions. The key contributions of this work are: + +(1) We propose a new model for MDE in a SIDE task. The proposed model uses a conditional diffusion architecture with the semantic context being supplied through embeddings generated using a ViT model. Achieving a new state-of-the-art (SOTA) performance, our method outperforms existing approaches on benchmark datasets, including the NYU Depth v2 indoor and KITTI outdoor datasets. Notably, we report a significant improvement of $14\%$ in absolute relative error, achieving 0.059 compared to the current SOTA (VPD) performance of 0.069 on NYU Depth v2. And we report an improvement of $2\%$ in square relative error, achieving 0.139 compared to the current SOTA (GED) performance of 0.142 on KITTI. +(2) We show, qualitatively as well as quantitatively, that using ViT embeddings to provide semantic context is a better alternative to generating pseudo captions and then using its CLIP embeddings to condition a SIDE model. In contrast to TADP [20], which uses pseudo captions, but only achieves a RMSE of 0.225 on NYU Depth v2, we report a lower, and a new SOTA, error of 0.218 (VPD [54] achieves 0.254). +(3) We show that providing ViT conditioning, helps our model perform better in a zero shot transfer task. ZoEDepth [4], the current SOTA for zero-shot transfer, reports an improvement of (16%, 18%, 45%, 9%) over NeWCRF on (Sun-RGBD, iBims1, DIODE, HyperSim) datasets, after training their model on 12 other datasets and NYU Depth v2. In contrast, we only train on NYU Depth v2, and report a much larger improvement of (21%, 23%, 81%, 25%). + +# 2. Related Work + +Traditional Methods. Earlier techniques for SIDE have used Markov Random Fields [47], non-parametric depth sampling [18], and structural similarity with prior depth map [13] to predict pixel-wise depth. + +Deep Learning Techniques for SIDE. Modern techniques have approached the problem as a dense regression problem. CNNs have been the dominant architecture for the SIDE in the last decade, with global-local network stack [9], and multi-scale [21], or encoder-decoder architecture [11] as some of the popular solution strategies. Recently, PixelFormer [1] used a transformer-based encoder-decoder architecture with skip connections from encoders to decoders. MIM [51] proposed masked image modeling as a general-purpose pre-training for geometric and motion tasks such as SIDE and pose estimation. Similarly, AiT [28] used mask augmentation and proposed soft tokens to generalize visual prediction tasks. While earlier works either focused on MDE for a specific dataset or RDE for generalization on multiple datasets, ZoEDepth [4] has proposed + +![](images/c7e530d0b4f2d38341472733a1549e75c90b00840b1be1a8bba2c60c1793cff7.jpg) +Figure 2. An overview of our proposed model: The latent representation of the input image undergoes a diffusion process, which is conditioned by our proposed CIDe module. Within the CIDe module, the input image is fed through the frozen ViT model. From this, a linear combination of the learnt embeddings is computed, which is transformed to generate a 768-dimensional contextual embedding. This embedding is utilized to condition the diffusion backbone. Subsequently, hierarchical feature maps are extracted from the UNet's decoder which are concatenated and processed through a depth regressor to generate the depth map. + +a generalized method for MDE that performs well in zero-shot transfer. We outperform [4] by a large margin even when training on a single dataset - NYU Depth v2 for indoor scenes or KITTI for outdoor scenes. Whereas, [4] uses 12 datasets for pre-training and then fine-tunes on NYU Depth v2 or KITTI for zero shot transfer. + +Diffusion-based Methods with Pretraining on Large Datasets. Recently, many techniques for SIDE has used diffusion architectures. These techniques [8, 16, 20, 38, 54], exploit prior knowledge acquired by pretraining on large datasets like LAION-400M [39], which consists of 400 million image-text pairs. In contrast, depth datasets, such as NYU Depth v2 and KITTI, contain around 20-30 thousand image-depth pairs. DepthGen [38] and DDP [16] work on a noise-to-depthmap paradigm and use images for conditional guidance of the diffusion process. DepthGen employs self-supervised pretraining on tasks like colorization, inpainting, and JPEG artifact removal, followed by supervised training on indoor and outdoor datasets [6, 12, 26, 27, 48]. DDP [16] decouples the image encoder and map decoder, allowing the image encoder to run just once, while the lightweight map decoder is run multiple times. VPD [54] and TADP [20] use denoising UNet [37] as a backbone to extract the rich features at multiple scales. Also, they utilize text instead of image for conditioning the diffusion backbone. + +Vision Transformer for Scene Understanding. The transformer architecture was initially proposed for NLP tasks [50], but introduced to the computer vision community as Vision Transformer (ViT) in [7]. Prior to this, CNNs dominated computer vision problems due to their ability to capture spatial hierarchies in image data. However, such archi + +tectures were constrained due to their ability to learn spatially localized features only. Transformer architecture has a weaker inductive bias and allows ViT to learn long-range dependencies, and robust, generalizable features. ViT architectures have replaced CNNs for SOTA performance on most computer vision tasks in recent years. We use a pretrained ViT model for providing semantic information to the diffusion backbone in our model. + +# 3. Proposed Methodology + +# 3.1. Preliminaries + +Problem Formulation. The objective of single image depth prediction task is to predict continuous values, denoted as $\mathbf{y} \in \mathbb{R}^{H \times W}$ , for every pixel present in the input RGB image, $\mathbf{x} \in [0,255]^{3 \times H \times W}$ . Here $H$ and $W$ represent the height and width respectively of the input image. + +Diffusion Model. Diffusion models are a class of generative models that progressively inject noise into the input data (forward pass) and then learn to reconstruct the original data in a reverse denoising process (reverse pass). There are three formulations of diffusion models: denoising diffusion probabilistic models (DDPMs) [14], score-based generative models [43, 44], and those based on stochastic differential equations [45, 46]. DDPMs are of relevance to our paper, and are described below. The architecture of DDPMs consists of two Markov chains: a forward chain that adds noise to the data, and a reverse chain that converts noise back to data by learning transition kernels parameterized by deep neural networks. Formally, the forward pass is modeled as + +a Markov process: + +$$ +\mathbb {P} \left(\mathbf {z} _ {t} \mid \mathbf {z} _ {t - 1}\right) = \mathcal {N} \left(\mathbf {z} _ {t}; \sqrt {1 - \beta_ {t}} \mathbf {z} _ {t - 1}, \beta_ {t} \boldsymbol {I}\right). \tag {1} +$$ + +Here $\mathbf{z}_t$ denotes the random variable at the $t^{\mathrm{th}}$ time step, $\mathcal{N}(\mathbf{z};\boldsymbol {\mu},\boldsymbol {\sigma})$ denotes Gaussian probability distribution, and $\beta_{t}$ is the noise schedule. The above equation leads to the analytic form of $\mathbb{P}(\mathbf{z}_t\mid \mathbf{z}_0),\forall t\in \{0,1,\dots ,T\}$ : + +$$ +\mathbf {z} _ {t} = \sqrt {\bar {\beta} _ {t}} \mathbf {z} _ {0} + \sqrt {1 - \bar {\beta} _ {t}} \boldsymbol {\epsilon}, \tag {2} +$$ + +where $\bar{\beta}_t = \prod_{s=1}^t \beta_s$ and $\epsilon \sim \mathcal{N}(\mathbf{0},\mathbf{I})$ . The model then gradually removes noise by executing a learnable Markov chain in the reverse time direction, parameterized by a normal prior distribution $\mathbb{P}(\mathbf{z}_T) = \mathcal{N}(\mathbf{z}_T;0,\mathbf{I})$ and a learnable transition kernel $\mathbb{P}_{\theta}(\mathbf{z}_{t-1}|\mathbf{z}_t)$ given by: + +$$ +\mathbb {P} _ {\theta} \left(\mathbf {z} _ {t - 1} \mid \mathbf {z} _ {t}\right) = \mathcal {N} \left(\mathbf {z} _ {t - 1}; \mu_ {\theta} \left(\mathbf {z} _ {t}, t\right), \Sigma_ {\theta} \left(\mathbf {z} _ {t}, t\right)\right). \tag {3} +$$ + +The goal of the training process is to approximately match the reverse Markov chain with the actual time reversal of the forward Markov chain. Mathematically, parameter $\theta$ is adjusted so that the joint distribution of the reverse Markov chain $\mathbb{P}_{\theta}(\mathbf{z}_0,\mathbf{z}_1,\dots ,\mathbf{z}_T)\coloneqq \mathbb{P}(\mathbf{z}_T)\prod_{t = 1}^T\mathbb{P}_{\theta}(\mathbf{z}_{t - 1}\mid \mathbf{z}_t)$ closely approximates that of the forward Markov chain $\mathbb{P}(\mathbf{z}_0,\mathbf{z}_1,\dots ,\mathbf{z}_T)\coloneqq \mathbb{P}(\mathbf{z}_0)\prod_{t = 1}^T\mathbb{P}(\mathbf{z}_t\mid \mathbf{z}_{t - 1})$ . This is achieved by minimizing the following loss: + +$$ +\mathbb {E} _ {t \sim \mathcal {U} [ 1, T ], \mathbf {z} _ {0} \sim \mathbb {P} (\mathbf {z} _ {0}), \epsilon \sim \mathcal {N} (0, I)} \left[ \| \epsilon - \epsilon_ {\theta} (\mathbf {z} _ {t}, t) \| ^ {2} \right], \tag {4} +$$ + +where $\mathbf{z}_t$ is computed using Eq. (2), and $\epsilon_{\theta}$ is predicted using a neural network, typically a UNet architecture [37]. In a conditional diffusion model $\epsilon_{\theta}(\mathbf{z}_t,t)$ gets replaced by $\epsilon_{\theta}(\mathbf{z}_t,t,\mathcal{C})$ , where $\mathcal{C}$ is a conditioning variable. + +# 3.2. Our Architecture + +Image Encoder and Latent Diffusion. Diffusion models typically take large number of time steps to train, and are difficult to converge. Recently, [36] proposed a new diffusion with improved convergence properties and is called "stable-diffusion". The key idea is to perform the diffusion in latent space, with latent embedding learnt separately through a variational autoencoder (VAE). The Encoder of VAE first transforms the input image $\mathbf{x}$ of size $(H,W)$ to latent space, then we follow latent diffusion formulation and utilize the UNet used in Stable Diffusion[36]. Utilizing latent diffusion formulation enables our architecture to capture multi-resolution features. Hence, we aggregate the feature maps from different layers of the UNet module (implementing conditional diffusion) by bringing them all to 1/4th resolution of the latent space, resulting in a feature map of size $8e\times H / 32\times W / 32$ . + +Exploiting Semantic Context with Conditional Diffusion. Recall that we formulated the single image depth estimation as a dense regression problem, predicting $\mathbb{P}(\mathbf{y}\mid \mathbf{x})$ + +![](images/c884c50e858aad2ed14417c86c14d50488599881dbd4715cd3e63339ec6572b7.jpg) +(a) + +![](images/57991546efa28fa9bd31fcb0cf70cb7c634b4e25f7be459feddfdac4e6a8ced9.jpg) +(b) +Figure 3. (a) Probabilistic graphical model corresponding to VPD. (b) The same corresponding to our formulation. Here, $\mathcal{C}$ represents the semantic embedding derived from our CIDE module. This embedding is internally generated by passing $\mathbf{x}$ through the ViT, resulting in $\mathcal{E}$ . Subsequently, $\mathcal{E}$ undergoes further processing to yield $\mathcal{C}$ , which is then utilized in the conditional diffusion module implementing $\mathbb{P}(\mathbf{z}_0\mid \mathbf{z}_t,\mathcal{C})$ . The output of the conditional diffusion module is fed into the Depth Regressor module within our architecture, implementing $\mathbb{P}(\mathbf{y}\mid \mathbf{z}_0)$ . + +where $\mathbf{y}$ and $\mathbf{x}$ denote output depth, and input image respectively. In our diffusion formulation, we predict $\mathbf{z}_0$ which is then used to predict pixel-wise depth, $\mathbf{y}$ . It has been shown that noise prediction in a diffusion model, $\epsilon_{\theta}$ , can be seen as predicting the gradient of the density function, $\nabla_{\mathbf{z}_t} \log \mathbb{P}(\mathbf{z}_t)$ . Hence, overall architecture for SIDE using a diffusion architecture can be seen as factorizing the conditional probability, as shown in the probabilistic graph model in Fig. 3a. To utilize additional semantic context generated from a ViT model, we condition it on the ViT embeddings as shown in Fig. 3b. Mathematically, we model: + +$$ +\mathbb {P} (\mathbf {y} \mid \mathbf {x}, \mathcal {E}) = \mathbb {P} (\mathbf {y} \mid \mathbf {z} _ {0}) \mathbb {P} (\mathbf {z} _ {0} \mid \mathbf {z} _ {t}, \mathcal {C}) \mathbb {P} (\mathbf {z} _ {t} \mid \mathbf {x}) \mathbb {P} (\mathcal {C} \mid \mathbf {x}), +$$ + +where + +$$ +\mathbb {P} (\mathcal {C} \mid \mathbf {x}) = \mathbb {P} (\mathcal {C} \mid \mathcal {E}) \mathbb {P} (\mathcal {E} \mid \mathbf {x}). \tag {5} +$$ + +Here, the first term, $\mathbb{P}(\mathbf{y}\mid \mathbf{z}_0)$ , is implemented through the Depth Regressor module explained earlier. Similarly, $\mathbb{P}(\mathbf{z}_t\mid \mathbf{x})$ is implemented using the VAE's Encoder as described earlier. We generate conditional information, $\mathcal{C}$ using our Comprehensive Image Detail Embedding module (hereafter CIDE, and explained below). The CIDE module takes $\mathbf{x}$ as input and generates embedding vector $\mathcal{C}$ (of dimension 768 in our design) as the output, thus, implementing $\mathbb{P}(\mathcal{C}\mid \mathbf{x})$ given in Eq. (5). We use $\mathcal{E}$ to denote the embedding vector of a ViT module, and $\mathbb{P}(\mathcal{E}\mid \mathbf{x})$ is implemented through the ViT. $\mathbb{P}(\mathcal{C}\mid \mathcal{E})$ is implemented using downstream modules in CIDE consisting of learnable embeddings. The second term, $\mathbb{P}(\mathbf{z}_0\mid \mathbf{z}_t,\mathcal{C})$ , is implemented using conditional diffusion, + +Comprehensive Image Detail Embedding (CIDE) Module. As described earlier, we believe using pseudo-captions + +Table 1. Results on Indoor NYU Depth v2 [27] Dataset. Results that are bold perform best. $\uparrow$ means the metric should be higher, $\downarrow$ indicate lower is better. The evaluation uses an upper bound of 10 meters on the ground truth depth map. All the numbers for other works have been taken from the corresponding papers. For MIM, and ZoEDepth we have used SwinV2-L 1K, and ZoeDepth-M12-N versions respectively. We see an overall improvement against SOTA on all the metrics used for evaluation. + +
MethodVenueAbs Rel↓RMSE↓log10↓Sq Rel↓δ1↑δ2↑δ3↑
Eigen et al.[9]NIPS'140.1580.641--0.7690.9500.988
DORN[10]CVPR'180.1150.5090.051-0.8280.9650.992
SharpNet[32]ICCV'190.1390.5020.047-0.8360.9660.993
Chen et al.[5]IJCAI-190.1110.5140.048-0.8780.9770.994
BTS[22]Arxiv'190.1100.3920.0470.0660.8850.9780.994
AdaBins[2]CVPR'210.1030.3640.044-0.9030.9840.997
DPT[33]ICCV'210.1100.3570.045-0.9040.9880.998
P3Depth[30]CVPR'220.1040.3560.043-0.8980.9810.996
NeWCRFs[53]CVPR'220.0950.3340.0410.0450.9220.9920.998
SwinV2-B[24]CVPR'220.1330.4620.059-0.8190.9750.995
SwinV2-L[24]CVPR'220.1120.3810.051-0.8860.9840.997
Localbins[3]ECCV'220.0990.3570.042-0.9070.9870.998
Jun et al.[17]ECCV'220.0980.3550.042-0.9130.9870.998
PixelFormer[1]WACV'230.0900.3220.0390.0430.9290.9910.998
DDP[16]ICCV'230.0940.3290.040-0.9210.9900.998
MIM [51]CVPR'230.0830.2870.035-0.9490.9940.999
AiT[28]ICCV'230.0760.2750.033-0.9540.9940.999
ZoeDepth [4]Arxiv'230.0750.2700.0320.0300.9550.9950.999
VPD[54]ICCV'230.0690.2540.0300.0270.9640.9950.999
OursCVPR'240.0590.2180.0260.0130.9780.9970.999
+ +![](images/a76cc74dd06135df0021484b881dda524bb0070b423a6e292f03cff9e997573e.jpg) +Figure 4. Visual Comparison on NYU Depth v2 Indoor Dataset. Note, our method's ability to delineate objects in terms of their depth, such as the table lamp in Row 5, even when such information is absent from the ground truth depth map. + +Table 2. Performance on the Outdoor KITTI [12] Dataset. Please refer to the caption of Tab. 1 for notation details. For ZoEDepth results, we use the ZoeDepth-M12-K version following the authors' recommendation. In instances where results for certain methods were not reported in the respective works, denoted by “-”, and the code is unavailable, we were unable to generate the missing numbers. Despite the saturation of results on the outdoor KITTI dataset, our method consistently achieves comparable or superior performance to the state-of-the-art (SOTA) across all metrics. VPD[54] cannot be trained on KITTI dataset as it required a per image text label which is not present in KITTI. The symbol $\dagger$ indicates that the method utilizes additional information beyond RGB. + +
MethodVenueAbs Rel↓Sq Rel↓RMSElog ↓RMSE↓δ1 ↑δ2 ↑δ3 ↑
Eigen et al.[9]NIPS'140.2031.5170.2826.3070.7020.8980.967
DORN[10]CVPR'180.0720.3070.1202.7270.9320.9840.994
BTS[22]Arxiv'190.0590.2410.0962.7560.9560.9930.998
AdaBins[2]CVPR'210.0670.1900.0882.9600.9490.9920.998
DPT[33]ICCV'210.060-0.0922.5730.9590.9950.996
P3Depth[30]CVPR'220.0710.2700.1032.8420.9530.9930.998
NeWCRFs[53]CVPR'220.0520.1550.0792.1290.9740.9970.999
PixelFormer[1]WACV'230.0510.1490.0772.0810.9760.9970.999
ZoeDepth [4]Arxiv'230.0540.1890.0832.4400.970.9960.999
DDP [16]ICCV'230.0500.1480.0762.0720.9750.9970.999
URCDC [40]ToM'230.0500.1420.0762.0320.9770.9970.999
IEBins [41]NeurIPS'230.0500.1420.0752.0110.9780.9980.999
MIM [51]CVPR'230.0500.1390.0751.9660.9770.9981.000
GEDepth[52]†ICCV'230.0480.1420.0762.0440.9760.9970.999
OursCVPR'240.0480.1390.0742.0390.9790.9981.000
+ +![](images/fd5626577cb0a6d7c11ad79b7f023d80f1eaefeb0fdc9d0368fc06a025dcee80.jpg) +Figure 5. Visual Comparison on KITTI Outdoor Dataset. + +to generate the semantic context has limited utility, as the textual descriptions typically focus on large salient objects only. Instead we propose our CIDE module which use embeddings from a pre-trained ViT, and extract detailed semantic context from these embeddings. For this we take 1000 dimensional logit vector from ViT, and pass it through a two layer MLP which converts it to a 100 dimensional vector. Subsequently, we employ this vector to compute the linear combination of 100 learnable embeddings. + +This resulting embedding undergoes a linear transformation to yield a semantic context vector of dimension 768, which is then passed to conditional diffusion module. + +Depth Regressor. The output feature map undergoes through an Upsampling Decoder, comprised of deconvolution layers, followed by a Depth Regressor. The Depth Regressor is essentially a two-layer convolutional neural network (CNN), with the initial layer having dimensions $\mathrm{Conv}(3\times 3)$ , 192, and the subsequent layer $\mathrm{Conv}(3\times 3)$ , 1. + +Table 3. Quantitative results for zero-shot transfer to four unseen indoor datasets. mRI denotes the mean relative improvement with respect to NeWCRFs across all metrics (δ1, REL, RMSE). Evaluation depth is capped at 8m for SUN RGB-D, 10m for iBims and DIODE Indoor, and 80m for HyperSim. Best results are in bold, second best are underlined. Our mRI outperforms all methods across all datasets by a large margin. † denotes that ZoeD is trained on 12 datasets and our method is trained only on NYUv2. + +
MethodSUN RGB-DiBims-1 BenchmarkDIODE IndoorHyperSim
δ1↑REL↓RMSE↓mRIθ↑δ1↑REL↓RMSE↓mRIθ↑δ1↑REL↓RMSE↓mRIθ↑δ1↑REL↓RMSE↓mRIθ↑
BTS [22]0.7400.1720.515-14.2%0.5380.2310.919-6.9%0.2100.4181.9052.3%0.2250.4766.404-8.6%
AdaBins [2]0.7710.1590.476-7.0%0.5550.2120.901-2.1%0.1740.4431.963-7.2%0.2210.4836.546-10.5%
LocalBins [3]0.7770.1560.470-5.6%0.5580.2110.880-0.7%0.2290.4121.8537.1%0.2340.4686.362-6.6%
NeWCRFs [53]0.7980.1510.4240.0%0.5480.2060.8610.0%0.1870.4041.8670.0%0.2550.4426.0170.0%
VPD [54]0.8610.1210.35514.7%0.6270.1870.76711.5%0.4800.3921.29563.4%0.3330.5315.1118.5%
ZoeD-M12-N† [4]0.8640.1190.34616.0%0.6580.1690.71118.5%0.3760.3271.58845.0%0.2920.4105.7718.6%
Ours0.8850.1120.31920.5%0.6880.1630.66423.1%0.5450.3441.16481.3%0.3940.4424.73925.2%
+ +![](images/10a7e4cdf5394b729bfc2c2fda0f2f9fb4531171367e28fe6cc3a38ac46faaed.jpg) + +![](images/f31058aa86b9b19c5376f9c5dabda1e36277891d933e7679fad2e97499e3bd57.jpg) +Figure 6. Visualization of improvements over VPD [54] in our model due to ViT embeddings passed down as conditional vectors (in blue and green). In the above images, ViT detects the desktop computer (first image) and table lamp (second image) with high probability, and they are thus better detected by our model. Additional visualizations are provided in the supplementary material. + +# 4. Experiments and Results + +Datasets and Evaluation. We use NYU Depth v2 [27] and KITTI [12] as the primary datasets for training. The NYU Depth v2 dataset is a widely used indoor benchmark for monocular depth estimation, containing over 24k densely labeled pairs of RGB and depth images in the train set and 654 in the test set. The dataset covers a wide range of indoor scenes and includes challenging scenarios such as reflective surfaces, transparent objects, and occlusions. The ground truth depth maps are obtained using a structured light sensor and are provided at a resolution of $640 \times 480$ . + +KITTI dataset is a widely used outdoor benchmark for monocular depth estimation, containing over 24k densely labeled pairs of RGB and depth images. The dataset covers outdoor driving scenarios and includes varying lighting conditions, weather, and occlusions. The ground truth depth maps are obtained using a Velodyne LiDAR sensor and are provided at a resolution of $1242 \times 375$ . To demonstrate generalizability, we evaluate zero-shot performance on the following datasets: Sun-RGBD [42], iBims1 [19], DIODE [49], and HyperSim [35]. Whereas, the main paper contains mostly quantitative, and only some representative + +Table 4. Effectiveness of different embeddings for guiding the diffusion process for depth estimation on NYU Depth v2 dataset. In the third row, rather than using pseudo caption embeddings generated from the scene label (as implemented by VPD [54]), we provide the one-hot vector representing the scene label as a condition to the diffusion model. We observe a slight improvement in the metrics, highlighting that information content in the caption is similar to that in one-hot label. + +
EmbeddingsRMSE↓Abs Rel↓δ1↑
Scene label emb. [54]0.2540.0690.964
Text caption emb. [20]0.2250.0620.976
One-hot vector emb.0.2440.0670.968
Proposed scene emb.0.2180.0590.978
+ +tive visual results, detailed visual results on each dataset are included in the supplementary. + +Implementation Details. Our model is implemented using PyTorch [29]. For optimization, we have used AdamW optimizer [25] with $\beta_0$ values of 0.9 and 0.999, a batch size of 32, and a weight decay of 0.1. We train our model for 25 epochs for both KITTI and NYU Depth v2 datasets, with an initial learning rate of $3 \times 10^{-5}$ . We first linearly increase learning rate to $5 \times 10^{-4}$ , and then linearly decrease across training iterations. We use usual data augmentation techniques, including random hue addition, horizontal flipping, changing the image brightness, and Cut Depth [15]. Our model takes approximately 21 minutes per epoch to train using 8 NVIDIA A100 GPUs. + +# 4.1. Comparison on Benchmark Datasets + +Comparison on NYU Depth v2. Tab. 1 shows the comparison of our proposed method with SOTA methods on the indoor NYU Depth v2 dataset [27]. We achieve a new state of the art on this dataset. Our methods perform better than the previous SOTA ([54])) by a large margin of $14\%$ in terms of RMSE. Fig. 4 provides a qualitative comparison on the dataset. + +Comparison on KITTI. Tab. 2 shows the comparison with various methods on the KITTI dataset [12]. Fig. 5 shows the qualitative results. Unlike NYU Depth v2 dataset, KITTI is an outdoor dataset. On this dataset also, we achieve similar or better performance than all existing state-of-the-art techniques. + +# 4.2. Generalization and Zero Shot Transfer + +Unlike state of the art on zero short transfer (ZoEDepth [4]), which requiring training on many datasets (12 in their case) for effective zero shot transfer, we show that our model generalizes well to other unseen dataset even when trained on a single NYU Depth v2 dataset. Tab. 3 shows quantitative results to back our findings. + +# 4.3. Ablation Study + +Effect of Contextual Information. As highlighted in the motivation, a key observation of this study is the richness of information contained within the output probability vector of the ViT, surpassing the textual embeddings employed in current state-of-the-art (e.g., VPD [54]). To compare with the utilization of scene label information (as implemented by VPD), we construct the conditioning embedding as a one-hot vector using the scene label, and subsequently transform it using an MLP. As illustrated in Tab. 4, we observe a slight improvement over VPD, indicating that the information content in pseudo-captions resembles that of one-hot labels. Furthermore, our proposed method surpasses both the approaches. + +Qualitative Results. Perhaps the most important task would be to verify that use of rich probability vector instead of text embeddings actually results in an improvement in depth as a direct consequence. We do this by considering the top few objects predicted by the ViT and correlating this with the depth predicted at these objects. We hypothesise when a particular class say dog is predicted with a high probability (hence there must be a dog in the image), then the corresponding depth must also be more accurate. We show this in Fig. 6. + +# 5. Conclusion + +We presented a new architecture block Comprehensive Image Detail Embedding (CIDE) module for robust monocular depth estimation in this paper. Our key idea is to highlight the limitations of using pseudo-captions to provide contextual information, and instead propose to use richer class-wise probability generated by a classification model, such as ViT. The motivation is that, while textual embedding typically highlights salient objects, class-wise probability vector preserves more details, including smaller objects in the background also. We implement the idea using proposed CIDE module cascaded with a conditional diffusion pipeline for monocular depth estimation. We demonstrated the effectiveness of our approach on several benchmark datasets and showed that it outperforms SOTA methods by a significant margin. + +Acknowledgement. We thank Kartik Anand for assistance with some of the experiments. This work has been supported by funding through Staqu Technologies Pvt. Ltd., and Department of Science and Technology (DST), Government of India. The travel for Suraj Patni is supported through Research Acceleration funds of Department of Computer Science and Engineering, IIT Delhi. + +# References + +[1] Ashutosh Agarwal and Chetan Arora. Attention attention everywhere: Monocular depth prediction with skip attention. + +In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 5861-5870, 2023. 2, 5, 6 +[2] Shariq Farooq Bhat, Ibrahim Alhashim, and Peter Wonka. Adabins: Depth estimation using adaptive bins. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4009-4018, 2021. 5, 6, 7 +[3] Shariq Farooq Bhat, Ibraheem Alhashim, and Peter Wonka. Localbins: Improving depth estimation by learning local distributions. In European Conference on Computer Vision, pages 480-496. Springer, 2022. 5, 7 +[4] Shariq Farooq Bhat, Reiner Birkl, Diana Wofk, Peter Wonka, and Matthias Müller. Zoedepth: Zero-shot transfer by combining relative and metric depth, 2023. 2, 3, 5, 6, 7, 8 +[5] Xiaotian Chen, Xuejin Chen, and Zheng-Jun Zha. Structure-aware residual pyramid network for monocular depth estimation. arXiv preprint arXiv:1907.06023, 2019. 5 +[6] Angela Dai, Angel X Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5828-5839, 2017. 3 +[7] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 2, 3 +[8] Yiqun Duan, Xianda Guo, and Zheng Zhu. Diffusiondepth: Diffusion denoising approach for monocular depth estimation, 2023. 2, 3 +[9] David Eigen, Christian Puhrsch, and Rob Fergus. Depth map prediction from a single image using a multi-scale deep network. Advances in neural information processing systems, 27, 2014. 2, 5, 6 +[10] Huan Fu, Mingming Gong, Chaohui Wang, Kayhan Bat-manghelich, and Dacheng Tao. Deep ordinal regression network for monocular depth estimation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2002-2011, 2018. 5, 6 +[11] Huan Fu, Mingming Gong, Chaohui Wang, Kayhan Bat-manghelich, and Dacheng Tao. Deep ordinal regression network for monocular depth estimation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2002-2011, 2018. 2 +[12] Andreas Geiger, Philip Lenz, Christoph Stiller, and Raquel Urtasun. Vision meets robotics: The kitti dataset. International Journal of Robotics Research (IJRR), 2013. 3, 6, 7, 8 +[13] José L. Herrera, Carlos R. del Blanco, and Narciso García. Automatic depth extraction from 2d images using a cluster-based learning framework. IEEE Transactions on Image Processing, 27(7):3288-3299, 2018. 2 +[14] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 3 + +[15] Yasunori Ishii and Takayoshi Yamashita. Cutdepth: Edge-aware data augmentation in depth estimation. arXiv preprint arXiv:2107.07684, 2021. 8 +[16] Yuanfeng Ji, Zhe Chen, Enze Xie, Lanqing Hong, Xihui Liu, Zhaoqiang Liu, Tong Lu, Zhenguo Li, and Ping Luo. Ddp: Diffusion model for dense visual prediction. arXiv preprint arXiv:2303.17559, 2023. 2, 3, 5, 6 +[17] Jinyoung Jun, Jae-Han Lee, Chul Lee, and Chang-Su Kim. Depth map decomposition for monocular depth estimation. In European Conference on Computer Vision, pages 18-34. Springer, 2022. 5 +[18] Kevin Karsch, Ce Liu, and Sing Bing Kang. Depth transfer: Depth extraction from video using non-parametric sampling. IEEE transactions on pattern analysis and machine intelligence, 36(11):2144-2158, 2014. 2 +[19] Tobias Koch, Lukas Liebel, Friedrich Fraundorfer, and Marco Korner. Evaluation of cnn-based single-image depth estimation methods. In Proceedings of the European Conference on Computer Vision (ECCV) Workshops, pages 0–0, 2018. 1, 7 +[20] Neehar Kondapaneni, Markus Marks, Manuel Knott, Rogério Guimarães, and Pietro Perona. Text-image alignment for diffusion-based perception. arXiv preprint arXiv:2310.00031, 2023. 2, 3, 8 +[21] Iro Laina, Christian Rupprecht, Vasileios Belagiannis, Federico Tombari, and Nassir Navab. Deeper depth prediction with fully convolutional residual networks. In 2016 Fourth international conference on 3D vision (3DV), pages 239-248. IEEE, 2016. 2 +[22] Jin Han Lee, Myung-Kyu Han, Dong Wook Ko, and Il Hong Suh. From big to small: Multi-scale local planar guidance for monocular depth estimation. arXiv preprint arXiv:1907.10326, 2019. 5, 6, 7 +[23] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In International Conference on Machine Learning, pages 12888-12900. PMLR, 2022. 2 +[24] Ze Liu, Han Hu, Yutong Lin, Zhuliang Yao, Zhenda Xie, Yixuan Wei, Jia Ning, Yue Cao, Zheng Zhang, Li Dong, et al. Swin transformer v2: Scaling up capacity and resolution. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12009-12019, 2022. 5 +[25] Ilya Loshchilov and Frank Hutter. Decoupled Weight Decay Regularization. In ICML 2019. 8 +[26] John McCormac, Ankur Handa, Stefan Leutenegger, and Andrew J Davison. Scelenet rgb-d: 5m photorealistic images of synthetic indoor trajectories with ground truth. arXiv preprint arXiv:1612.05079, 2016. 3 +[27] Pushmeet Kohli Nathan Silberman, Derek Hoiem and Rob Fergus. Indoor segmentation and support inference from rgbd images. In ECCV, 2012. 3, 5, 7, 8 +[28] Jia Ning, Chen Li, Zheng Zhang, Chunyu Wang, Zigang Geng, Qi Dai, Kun He, and Han Hu. All in tokens: Unifying output space of visual tasks via soft token. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 19900-19910, 2023. 2, 5 + +[29] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, Alban Desmaison, Andreas Kopf, Edward Yang, Zachary DeVito, Martin Raison, Alykhan Tejani, Sasank Chilamkurthy, Benoit Steiner, Lu Fang, Junjie Bai, and Soumith Chintala. Pytorch: An imperative style, high-performance deep learning library. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2019. 8 +[30] Vaishakh Patil, Christos Sakaridis, Alexander Liniger, and Luc Van Gool. P3depth: Monocular depth estimation with a piecewise planarity prior. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1610–1621, 2022. 5, 6 +[31] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 2 +[32] Michael Ramamonjisoa and Vincent Lepetit. Sharpnet: Fast and accurate recovery of occluding contours in monocular depth estimation. In Proceedings of the IEEE/CVF International Conference on Computer Vision Workshops, pages 0–0, 2019. 5 +[33] René Ranftl, Alexey Bochkovskiy, and Vladlen Koltun. Vision transformers for dense prediction. In Proceedings of the IEEE/CVF international conference on computer vision, pages 12179-12188, 2021. 5, 6 +[34] René Ranftl, Katrin Lasinger, David Hafner, Konrad Schindler, and Vladlen Koltun. Towards robust monocular depth estimation: Mixing datasets for zero-shot cross-dataset transfer. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(3), 2022. 2 +[35] Mike Roberts, Jason Ramapuram, Anurag Ranjan, Atulit Kumar, Miguel Angel Bautista, Nathan Paczan, Russ Webb, and Joshua M. Susskind. Hypersim: A photorealistic synthetic dataset for holistic indoor scene understanding. In International Conference on Computer Vision (ICCV) 2021, 2021. 1, 7 +[36] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 4 +[37] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical Image Computing and Computer-Assisted Intervention-MICCAI 2015: 18th International Conference, Munich, Germany, October 5-9, 2015, Proceedings, Part III 18, pages 234-241. Springer, 2015. 3, 4 +[38] Saurabh Saxena, Abhishek Kar, Mohammad Norouzi, and David J Fleet. Monocular depth estimation using diffusion models. arXiv preprint arXiv:2302.14816, 2023. 2, 3 +[39] Christoph Schuhmann, Richard Vencu, Romain Beaumont, Robert Kaczmarczyk, Clayton Mullis, Aarush Katta, Theo Coombes, Jenia Jitsev, and Aran Komatsuzaki. Laion-400m: + +Open dataset of clip-filtered 400 million image-text pairs. arXiv preprint arXiv:2111.02114, 2021. 2, 3 +[40] Shuwei Shao, Zhongcai Pei, Weihai Chen, Ran Li, Zhong Liu, and Zhengguo Li. Urcdc-depth: Uncertainty rectified cross-distillation with cutflip for monocular depth estimation. arXiv preprint arXiv:2302.08149, 2023. 6 +[41] Shuwei Shao, Zhongcai Pei, Xingming Wu, Zhong Liu, Weihai Chen, and Zhengguo Li. Iebins: Iterative elastic bins for monocular depth estimation. arXiv preprint arXiv:2309.14137, 2023. 6 +[42] Shuran Song, Samuel P. Lichtenberg, and Jianxiong Xiao. Sun rgb-d: A rgb-d scene understanding benchmark suite. In 2015 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 567-576, 2015. 1, 7 +[43] Yang Song and Stefano Ermon. Generative modeling by estimating gradients of the data distribution. Advances in neural information processing systems, 32, 2019. 3 +[44] Yang Song and Stefano Ermon. Improved techniques for training score-based generative models. Advances in neural information processing systems, 33:12438-12448, 2020. 3 +[45] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. arXiv preprint arXiv:2011.13456, 2020. 3 +[46] Yang Song, Conor Durkan, Iain Murray, and Stefano Ermon. Maximum likelihood training of score-based diffusion models. Advances in Neural Information Processing Systems, 34: 1415-1428, 2021. 3 +[47] M. Sun, A. Y. Ng, and A. Saxena. Make3d: Learning 3d scene structure from a single still image. IEEE Transactions on Pattern Analysis and Machine Intelligence, 31(05):824-840, 2009. 2 +[48] Pei Sun, Henrik Kretzschmar, Xerxes Dotiwalla, Aurelien Chouard, Vijaysai Patnaik, Paul Tsui, James Guo, Yin Zhou, Yuning Chai, Benjamin Caine, et al. Scalability in perception for autonomous driving: Waymo open dataset. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2446-2454, 2020. 3 +[49] Igor Vasiljevic, Nick Kolkin, Shanyi Zhang, Ruotian Luo, Haochen Wang, Falcon Z Dai, Andrea F Daniele, Mohammadreza Mostajabi, Steven Basart, Matthew R Walter, et al. Diode: A dense indoor and outdoor depth dataset. arXiv preprint arXiv:1908.00463, 2019. 1, 7 +[50] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 3 +[51] Zhenda Xie, Zigang Geng, Jingcheng Hu, Zheng Zhang, Han Hu, and Yue Cao. Revealing the dark secrets of masked image modeling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14475-14485, 2023. 2, 5, 6 +[52] Xiaodong Yang, Zhuang Ma, Zhiyu Ji, and Zhe Ren. Gedgepth: Ground embedding for monocular depth estimation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12719-12727, 2023. 6 +[53] Weihao Yuan, Xiaodong Gu, Zuozhuo Dai, Siyu Zhu, and Ping Tan. Newcrfs: Neural window fully-connected crfs for + +monocular depth estimation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2022. 5, 6, 7 +[54] Wenliang Zhao, Yongming Rao, Zuyan Liu, Benlin Liu, Jie Zhou, and Jiwen Lu. Unleashing text-to-image diffusion models for visual perception. ICCV, 2023. 2, 3, 5, 6, 7, 8 \ No newline at end of file diff --git a/2024/ECoDepth_ Effective Conditioning of Diffusion Models for Monocular Depth Estimation/images.zip b/2024/ECoDepth_ Effective Conditioning of Diffusion Models for Monocular Depth Estimation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..c0fe678fa1c9cb626e0727a7d56f1be43328b8e2 --- /dev/null +++ b/2024/ECoDepth_ Effective Conditioning of Diffusion Models for Monocular Depth Estimation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:863e04a34f1a7e85775c197f2d584e373498273cd760c211dff122e8b0aab040 +size 900243 diff --git a/2024/ECoDepth_ Effective Conditioning of Diffusion Models for Monocular Depth Estimation/layout.json b/2024/ECoDepth_ Effective Conditioning of Diffusion Models for Monocular Depth Estimation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..1427e26bdb0d45b5d93e63f62b8509d5d09191fa --- /dev/null +++ b/2024/ECoDepth_ Effective Conditioning of Diffusion Models for Monocular Depth Estimation/layout.json @@ -0,0 +1,7434 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 127, + 103, + 465, + 139 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 127, + 103, + 465, + 139 + ], + "spans": [ + { + "bbox": [ + 127, + 103, + 465, + 139 + ], + "type": "text", + "content": "ECoDepth: Effective Conditioning of Diffusion Models for Monocular Depth Estimation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 136, + 161, + 196, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 161, + 196, + 175 + ], + "spans": [ + { + "bbox": [ + 136, + 161, + 196, + 175 + ], + "type": "text", + "content": "Suraj Patni*" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 250, + 162, + 341, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 250, + 162, + 341, + 175 + ], + "spans": [ + { + "bbox": [ + 250, + 162, + 341, + 175 + ], + "type": "text", + "content": "Aradhye Agarwal*" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 389, + 162, + 456, + 174 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 389, + 162, + 456, + 174 + ], + "spans": [ + { + "bbox": [ + 389, + 162, + 456, + 174 + ], + "type": "text", + "content": "Chetan Arora" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 208, + 176, + 385, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 208, + 176, + 385, + 190 + ], + "spans": [ + { + "bbox": [ + 208, + 176, + 385, + 190 + ], + "type": "text", + "content": "Indian Institute of Technology Delhi" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 182, + 190, + 408, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 182, + 190, + 408, + 203 + ], + "spans": [ + { + "bbox": [ + 182, + 190, + 408, + 203 + ], + "type": "text", + "content": "https://ecodepth-iitd.github.io" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 143, + 231, + 191, + 243 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 231, + 191, + 243 + ], + "spans": [ + { + "bbox": [ + 143, + 231, + 191, + 243 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 256, + 290, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 256, + 290, + 593 + ], + "spans": [ + { + "bbox": [ + 46, + 256, + 290, + 593 + ], + "type": "text", + "content": "In the absence of parallax cues, a learning based single image depth estimation (SIDE) model relies heavily on shading and contextual cues in the image. While this simplicity is attractive, it is necessary to train such models on large and varied datasets, which are difficult to capture. It has been shown that using embeddings from pretrained foundational models, such as CLIP, improves zero shot transfer in several applications. Taking inspiration from this, in our paper we explore the use of global image priors generated from a pre-trained ViT model to provide more detailed contextual information. We argue that the embedding vector from a ViT model, pre-trained on a large dataset, captures greater relevant information for SIDE than the usual route of generating pseudo image captions, followed by CLIP based text embeddings. Based on this idea, we propose a new SIDE model using a diffusion backbone which is conditioned on ViT embeddings. Our proposed design establishes a new state-of-the-art (SOTA) for SIDE on NYU Depth v2 dataset, achieving Abs Rel error of 0.059(14% improvement) compared to 0.069 by the current SOTA (VPD). And on KITTI dataset, achieving Sq Rel error of 0.139 (2% improvement) compared to 0.142 by the current SOTA (GED). For zero shot transfer with a model trained on NYU Depth v2, we report mean relative improvement of (20%, 23%, 81%, 25%) over NeWCRF on (Sun-RGBD, iBims1, DIODE, HyperSim) datasets, compared to (16%, 18%, 45%, 9%) by ZoEDepth. The code is available in our project page." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 614, + 127, + 626 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 614, + 127, + 626 + ], + "spans": [ + { + "bbox": [ + 47, + 614, + 127, + 626 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 635, + 287, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 635, + 287, + 696 + ], + "spans": [ + { + "bbox": [ + 46, + 635, + 287, + 696 + ], + "type": "text", + "content": "Single Image Depth Estimation (SIDE) is the task of predicting per pixel depth using a single RGB image from a monocular camera. It is a fundamental problem in computer vision with applications in several domains, viz robotics, autonomous driving, and augmented reality. The problem is" + } + ] + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 306, + 230, + 426, + 262 + ], + "blocks": [ + { + "bbox": [ + 306, + 230, + 426, + 262 + ], + "lines": [ + { + "bbox": [ + 306, + 230, + 426, + 262 + ], + "spans": [ + { + "bbox": [ + 306, + 230, + 426, + 262 + ], + "type": "image", + "image_path": "f7274f44c4d22bc9eae8a19f69aa92c1c20646d54843a02397296650e2bfc12a.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 397, + 264, + 455, + 272 + ], + "lines": [ + { + "bbox": [ + 397, + 264, + 455, + 272 + ], + "spans": [ + { + "bbox": [ + 397, + 264, + 455, + 272 + ], + "type": "text", + "content": "(a) Sun-RGBD [42]" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 427, + 230, + 545, + 262 + ], + "blocks": [ + { + "bbox": [ + 427, + 230, + 545, + 262 + ], + "lines": [ + { + "bbox": [ + 427, + 230, + 545, + 262 + ], + "spans": [ + { + "bbox": [ + 427, + 230, + 545, + 262 + ], + "type": "image", + "image_path": "214cfea1c3468673bbbfb8b1f65db15a952fafc13969355b377513969847d17f.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 307, + 273, + 425, + 304 + ], + "blocks": [ + { + "bbox": [ + 307, + 273, + 425, + 304 + ], + "lines": [ + { + "bbox": [ + 307, + 273, + 425, + 304 + ], + "spans": [ + { + "bbox": [ + 307, + 273, + 425, + 304 + ], + "type": "image", + "image_path": "ff8d25236e5638817716f97d36604908ba6a66250a0304d48025d379b65c2095.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 427, + 273, + 545, + 304 + ], + "blocks": [ + { + "bbox": [ + 427, + 273, + 545, + 304 + ], + "lines": [ + { + "bbox": [ + 427, + 273, + 545, + 304 + ], + "spans": [ + { + "bbox": [ + 427, + 273, + 545, + 304 + ], + "type": "image", + "image_path": "f56327bab79fcd3f27c08fbef298e0b0e23fd81917afe234efe190f0d7e73e38.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 307, + 316, + 426, + 348 + ], + "blocks": [ + { + "bbox": [ + 401, + 307, + 451, + 316 + ], + "lines": [ + { + "bbox": [ + 401, + 307, + 451, + 316 + ], + "spans": [ + { + "bbox": [ + 401, + 307, + 451, + 316 + ], + "type": "text", + "content": "(b) iBims1 [19]" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 307, + 316, + 426, + 348 + ], + "lines": [ + { + "bbox": [ + 307, + 316, + 426, + 348 + ], + "spans": [ + { + "bbox": [ + 307, + 316, + 426, + 348 + ], + "type": "image", + "image_path": "acfee1ae7e238a4d60c7914c77d3468b68bc4fb10c24754766404c5d2a5dd4a9.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 427, + 316, + 545, + 348 + ], + "blocks": [ + { + "bbox": [ + 427, + 316, + 545, + 348 + ], + "lines": [ + { + "bbox": [ + 427, + 316, + 545, + 348 + ], + "spans": [ + { + "bbox": [ + 427, + 316, + 545, + 348 + ], + "type": "image", + "image_path": "b6c9dc2b7d14f841f908bf571798f99779b924830b846d5fe98305823fd3225f.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 307, + 358, + 426, + 391 + ], + "blocks": [ + { + "bbox": [ + 403, + 350, + 449, + 358 + ], + "lines": [ + { + "bbox": [ + 403, + 350, + 449, + 358 + ], + "spans": [ + { + "bbox": [ + 403, + 350, + 449, + 358 + ], + "type": "text", + "content": "(c) DIODE [49]" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 307, + 358, + 426, + 391 + ], + "lines": [ + { + "bbox": [ + 307, + 358, + 426, + 391 + ], + "spans": [ + { + "bbox": [ + 307, + 358, + 426, + 391 + ], + "type": "image", + "image_path": "5a9b6186ef62db411c91c5843145d756b45f98d60f13a0e2aa97b4f3d446142e.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 397, + 392, + 455, + 402 + ], + "lines": [ + { + "bbox": [ + 397, + 392, + 455, + 402 + ], + "spans": [ + { + "bbox": [ + 397, + 392, + 455, + 402 + ], + "type": "text", + "content": "(d) HyperSim [35]" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 304, + 411, + 545, + 488 + ], + "lines": [ + { + "bbox": [ + 304, + 411, + 545, + 488 + ], + "spans": [ + { + "bbox": [ + 304, + 411, + 545, + 488 + ], + "type": "text", + "content": "Figure 1. Qualitative results across four different datasets, demonstrating the zero-shot performance of our model trained only on the NYU Depth v2 dataset. Corresponding quantitative results are presented in Table 3. The first column displays RGB images, the second column depicts ground truth depth, and the third column showcases our model's predicted depths. Additional images for each dataset are available in the Supplementary Material." + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 427, + 358, + 545, + 391 + ], + "blocks": [ + { + "bbox": [ + 427, + 358, + 545, + 391 + ], + "lines": [ + { + "bbox": [ + 427, + 358, + 545, + 391 + ], + "spans": [ + { + "bbox": [ + 427, + 358, + 545, + 391 + ], + "type": "image", + "image_path": "8d5070d4cbc6c4ebcfef4196245823ef4f87ec26324f3274b07d2d886dd1ddc2.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + } + ], + "index": 22 + }, + { + "bbox": [ + 304, + 509, + 545, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 509, + 545, + 582 + ], + "spans": [ + { + "bbox": [ + 304, + 509, + 545, + 582 + ], + "type": "text", + "content": "typically formulated in two flavors: metric depth estimation (MDE), and relative depth estimation (RDE). As the names suggest, MDE deals with estimation in physical units such as meters, whereas, RDE techniques focus on relative depth only, and require a per-image affine transformation as post-processing to convert to a physical unit." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 304, + 582, + 545, + 701 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 582, + 545, + 701 + ], + "spans": [ + { + "bbox": [ + 304, + 582, + 545, + 701 + ], + "type": "text", + "content": "Conventional geometric techniques for depth estimation typically rely on feature correspondence, parallax, and triangulation from two or more views. However, the problem becomes ill-posed for estimation from a single view. Intuitively, depth map is a 3D representation of the scene, whereas an RGB image is a 2D projection of the scene. Hence, it cannot be uniquely determined from a single RGB image. Therefore, learning-based SIDE models rely on visual cues like 'shape from shading' and other contextual priors for per-pixel depth prediction." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 317, + 701, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 701, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 317, + 701, + 545, + 714 + ], + "type": "text", + "content": "A data-driven approach makes the SIDE pipeline sim" + } + ] + } + ], + "index": 27 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 702, + 126, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 702, + 126, + 712 + ], + "spans": [ + { + "bbox": [ + 58, + 702, + 126, + 712 + ], + "type": "text", + "content": "*Equal contribution." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "28285" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 288, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 288, + 156 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 288, + 156 + ], + "type": "text", + "content": "pler, but also makes the learnt model dependent on the quality of training data. It has been observed that such models overfit on a particular training distribution/domain and fail to generalize on unseen data. This is especially true for MDE models when the range of depth in the training dataset is limited. Hence, training on multiple datasets, with wide variations in depth ranges has been proposed [4, 34]." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 159, + 290, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 159, + 290, + 434 + ], + "spans": [ + { + "bbox": [ + 46, + 159, + 290, + 434 + ], + "type": "text", + "content": "On the other hand, development of large foundational models (LFMs) in recent years has altered the preferred design approach for many computer vision problems. These huge models are trained using extensive datasets of unlabeled images and learning objectives that are agnostic to specific tasks. The learnt embeddings from such pre-trained models have been shown to help generalization and zero-shot transfer in many applications. We are aware of at least two works for MDE in SIDE problem that have appeared in the last few months and make use of such foundational models. VPD [54] uses a text-to-image diffusion model pretrained on LAION-400M [39] dataset having large-scale image-text pairs as the backbone. The model prompts the denoising UNet with textual inputs to make the visual contents interact with the text prompts. Since the problem formulation doesn't include text description as the input, the model generates simple descriptions such as 'A photo of a {scene name}' based on the scene label given in the NYU Depth v2 dataset. Another work, TADP [20] improves upon VPD. Instead of simple image descriptions, TADP uses BLIP-2 [23] to generate image captions, and then uses CLIP [31] embeddings of the pseudo-caption to condition the diffusion model." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 438, + 289, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 438, + 289, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 438, + 289, + 713 + ], + "type": "text", + "content": "We view the above works as providing robust semantic context to LFMs for the actual task at hand, which helps in visual recognition in general, as well as SIDE. While we do agree with the broad motivations of these works, the question that we ask is if pseudo-captions are the most effective way to provide the semantic context. Textual descriptions of an image typically focus on large salient objects and emphasize on their relationships. On the other hand, large vision models for image classification, typically contain representation for even smaller objects present in the scene. Even when a single object is present in the scene, the representations typically capture uncertainties and ambiguities inherent in the scene. Hence, we posit that using embeddings from a transformer model, pretrained on a large image dataset unrelated to the SIDE task, captures more relevant information, and is a better alternative to using pseudocaptions' CLIP embedding. Since diffusion-based models have shown their superiority for the dense prediction tasks in recent works [8, 16, 20, 38, 54]. Hence, we propose a diffusion backbone for our model, along with a novel CIDE module which employs ViT [7] to extract semantic context embeddings. These embeddings are subsequently utilized to condition the diffusion backbone." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 306, + 72, + 529, + 83 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 529, + 83 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 529, + 83 + ], + "type": "text", + "content": "Contributions. The key contributions of this work are:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 306, + 84, + 547, + 455 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 306, + 84, + 547, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 84, + 547, + 251 + ], + "spans": [ + { + "bbox": [ + 306, + 84, + 547, + 251 + ], + "type": "text", + "content": "(1) We propose a new model for MDE in a SIDE task. The proposed model uses a conditional diffusion architecture with the semantic context being supplied through embeddings generated using a ViT model. Achieving a new state-of-the-art (SOTA) performance, our method outperforms existing approaches on benchmark datasets, including the NYU Depth v2 indoor and KITTI outdoor datasets. Notably, we report a significant improvement of " + }, + { + "bbox": [ + 306, + 84, + 547, + 251 + ], + "type": "inline_equation", + "content": "14\\%" + }, + { + "bbox": [ + 306, + 84, + 547, + 251 + ], + "type": "text", + "content": " in absolute relative error, achieving 0.059 compared to the current SOTA (VPD) performance of 0.069 on NYU Depth v2. And we report an improvement of " + }, + { + "bbox": [ + 306, + 84, + 547, + 251 + ], + "type": "inline_equation", + "content": "2\\%" + }, + { + "bbox": [ + 306, + 84, + 547, + 251 + ], + "type": "text", + "content": " in square relative error, achieving 0.139 compared to the current SOTA (GED) performance of 0.142 on KITTI." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 306, + 251, + 547, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 251, + 547, + 346 + ], + "spans": [ + { + "bbox": [ + 306, + 251, + 547, + 346 + ], + "type": "text", + "content": "(2) We show, qualitatively as well as quantitatively, that using ViT embeddings to provide semantic context is a better alternative to generating pseudo captions and then using its CLIP embeddings to condition a SIDE model. In contrast to TADP [20], which uses pseudo captions, but only achieves a RMSE of 0.225 on NYU Depth v2, we report a lower, and a new SOTA, error of 0.218 (VPD [54] achieves 0.254)." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 347, + 547, + 455 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 347, + 547, + 455 + ], + "spans": [ + { + "bbox": [ + 306, + 347, + 547, + 455 + ], + "type": "text", + "content": "(3) We show that providing ViT conditioning, helps our model perform better in a zero shot transfer task. ZoEDepth [4], the current SOTA for zero-shot transfer, reports an improvement of (16%, 18%, 45%, 9%) over NeWCRF on (Sun-RGBD, iBims1, DIODE, HyperSim) datasets, after training their model on 12 other datasets and NYU Depth v2. In contrast, we only train on NYU Depth v2, and report a much larger improvement of (21%, 23%, 81%, 25%)." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 464, + 392, + 476 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 464, + 392, + 476 + ], + "spans": [ + { + "bbox": [ + 306, + 464, + 392, + 476 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 484, + 545, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 484, + 545, + 533 + ], + "spans": [ + { + "bbox": [ + 305, + 484, + 545, + 533 + ], + "type": "text", + "content": "Traditional Methods. Earlier techniques for SIDE have used Markov Random Fields [47], non-parametric depth sampling [18], and structural similarity with prior depth map [13] to predict pixel-wise depth." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 534, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 534, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 305, + 534, + 547, + 715 + ], + "type": "text", + "content": "Deep Learning Techniques for SIDE. Modern techniques have approached the problem as a dense regression problem. CNNs have been the dominant architecture for the SIDE in the last decade, with global-local network stack [9], and multi-scale [21], or encoder-decoder architecture [11] as some of the popular solution strategies. Recently, PixelFormer [1] used a transformer-based encoder-decoder architecture with skip connections from encoders to decoders. MIM [51] proposed masked image modeling as a general-purpose pre-training for geometric and motion tasks such as SIDE and pose estimation. Similarly, AiT [28] used mask augmentation and proposed soft tokens to generalize visual prediction tasks. While earlier works either focused on MDE for a specific dataset or RDE for generalization on multiple datasets, ZoEDepth [4] has proposed" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "28286" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 49, + 73, + 539, + 238 + ], + "blocks": [ + { + "bbox": [ + 49, + 73, + 539, + 238 + ], + "lines": [ + { + "bbox": [ + 49, + 73, + 539, + 238 + ], + "spans": [ + { + "bbox": [ + 49, + 73, + 539, + 238 + ], + "type": "image", + "image_path": "c7e530d0b4f2d38341472733a1549e75c90b00840b1be1a8bba2c60c1793cff7.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 249, + 546, + 304 + ], + "lines": [ + { + "bbox": [ + 46, + 249, + 546, + 304 + ], + "spans": [ + { + "bbox": [ + 46, + 249, + 546, + 304 + ], + "type": "text", + "content": "Figure 2. An overview of our proposed model: The latent representation of the input image undergoes a diffusion process, which is conditioned by our proposed CIDe module. Within the CIDe module, the input image is fed through the frozen ViT model. From this, a linear combination of the learnt embeddings is computed, which is transformed to generate a 768-dimensional contextual embedding. This embedding is utilized to condition the diffusion backbone. Subsequently, hierarchical feature maps are extracted from the UNet's decoder which are concatenated and processed through a depth regressor to generate the depth map." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 325, + 286, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 325, + 286, + 396 + ], + "spans": [ + { + "bbox": [ + 46, + 325, + 286, + 396 + ], + "type": "text", + "content": "a generalized method for MDE that performs well in zero-shot transfer. We outperform [4] by a large margin even when training on a single dataset - NYU Depth v2 for indoor scenes or KITTI for outdoor scenes. Whereas, [4] uses 12 datasets for pre-training and then fine-tunes on NYU Depth v2 or KITTI for zero shot transfer." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 399, + 286, + 638 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 399, + 286, + 638 + ], + "spans": [ + { + "bbox": [ + 46, + 399, + 286, + 638 + ], + "type": "text", + "content": "Diffusion-based Methods with Pretraining on Large Datasets. Recently, many techniques for SIDE has used diffusion architectures. These techniques [8, 16, 20, 38, 54], exploit prior knowledge acquired by pretraining on large datasets like LAION-400M [39], which consists of 400 million image-text pairs. In contrast, depth datasets, such as NYU Depth v2 and KITTI, contain around 20-30 thousand image-depth pairs. DepthGen [38] and DDP [16] work on a noise-to-depthmap paradigm and use images for conditional guidance of the diffusion process. DepthGen employs self-supervised pretraining on tasks like colorization, inpainting, and JPEG artifact removal, followed by supervised training on indoor and outdoor datasets [6, 12, 26, 27, 48]. DDP [16] decouples the image encoder and map decoder, allowing the image encoder to run just once, while the lightweight map decoder is run multiple times. VPD [54] and TADP [20] use denoising UNet [37] as a backbone to extract the rich features at multiple scales. Also, they utilize text instead of image for conditioning the diffusion backbone." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 641, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 641, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 641, + 286, + 713 + ], + "type": "text", + "content": "Vision Transformer for Scene Understanding. The transformer architecture was initially proposed for NLP tasks [50], but introduced to the computer vision community as Vision Transformer (ViT) in [7]. Prior to this, CNNs dominated computer vision problems due to their ability to capture spatial hierarchies in image data. However, such archi" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 305, + 325, + 545, + 420 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 325, + 545, + 420 + ], + "spans": [ + { + "bbox": [ + 305, + 325, + 545, + 420 + ], + "type": "text", + "content": "tectures were constrained due to their ability to learn spatially localized features only. Transformer architecture has a weaker inductive bias and allows ViT to learn long-range dependencies, and robust, generalizable features. ViT architectures have replaced CNNs for SOTA performance on most computer vision tasks in recent years. We use a pretrained ViT model for providing semantic information to the diffusion backbone in our model." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 445, + 440, + 459 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 445, + 440, + 459 + ], + "spans": [ + { + "bbox": [ + 306, + 445, + 440, + 459 + ], + "type": "text", + "content": "3. Proposed Methodology" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 468, + 392, + 480 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 468, + 392, + 480 + ], + "spans": [ + { + "bbox": [ + 306, + 468, + 392, + 480 + ], + "type": "text", + "content": "3.1. Preliminaries" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 491, + 545, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 491, + 545, + 552 + ], + "spans": [ + { + "bbox": [ + 305, + 491, + 545, + 552 + ], + "type": "text", + "content": "Problem Formulation. The objective of single image depth prediction task is to predict continuous values, denoted as " + }, + { + "bbox": [ + 305, + 491, + 545, + 552 + ], + "type": "inline_equation", + "content": "\\mathbf{y} \\in \\mathbb{R}^{H \\times W}" + }, + { + "bbox": [ + 305, + 491, + 545, + 552 + ], + "type": "text", + "content": ", for every pixel present in the input RGB image, " + }, + { + "bbox": [ + 305, + 491, + 545, + 552 + ], + "type": "inline_equation", + "content": "\\mathbf{x} \\in [0,255]^{3 \\times H \\times W}" + }, + { + "bbox": [ + 305, + 491, + 545, + 552 + ], + "type": "text", + "content": ". Here " + }, + { + "bbox": [ + 305, + 491, + 545, + 552 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 305, + 491, + 545, + 552 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 491, + 545, + 552 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 305, + 491, + 545, + 552 + ], + "type": "text", + "content": " represent the height and width respectively of the input image." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 557, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 557, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 557, + 546, + 713 + ], + "type": "text", + "content": "Diffusion Model. Diffusion models are a class of generative models that progressively inject noise into the input data (forward pass) and then learn to reconstruct the original data in a reverse denoising process (reverse pass). There are three formulations of diffusion models: denoising diffusion probabilistic models (DDPMs) [14], score-based generative models [43, 44], and those based on stochastic differential equations [45, 46]. DDPMs are of relevance to our paper, and are described below. The architecture of DDPMs consists of two Markov chains: a forward chain that adds noise to the data, and a reverse chain that converts noise back to data by learning transition kernels parameterized by deep neural networks. Formally, the forward pass is modeled as" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "28287" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 124, + 83 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 124, + 83 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 124, + 83 + ], + "type": "text", + "content": "a Markov process:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 75, + 90, + 287, + 111 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 90, + 287, + 111 + ], + "spans": [ + { + "bbox": [ + 75, + 90, + 287, + 111 + ], + "type": "interline_equation", + "content": "\\mathbb {P} \\left(\\mathbf {z} _ {t} \\mid \\mathbf {z} _ {t - 1}\\right) = \\mathcal {N} \\left(\\mathbf {z} _ {t}; \\sqrt {1 - \\beta_ {t}} \\mathbf {z} _ {t - 1}, \\beta_ {t} \\boldsymbol {I}\\right). \\tag {1}", + "image_path": "200bd7d2a6559bc434e9181f57e368cfa9a2bdd385502235a02c39fa9b947c26.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 117, + 287, + 165 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 117, + 287, + 165 + ], + "spans": [ + { + "bbox": [ + 47, + 117, + 287, + 165 + ], + "type": "text", + "content": "Here " + }, + { + "bbox": [ + 47, + 117, + 287, + 165 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_t" + }, + { + "bbox": [ + 47, + 117, + 287, + 165 + ], + "type": "text", + "content": " denotes the random variable at the " + }, + { + "bbox": [ + 47, + 117, + 287, + 165 + ], + "type": "inline_equation", + "content": "t^{\\mathrm{th}}" + }, + { + "bbox": [ + 47, + 117, + 287, + 165 + ], + "type": "text", + "content": " time step, " + }, + { + "bbox": [ + 47, + 117, + 287, + 165 + ], + "type": "inline_equation", + "content": "\\mathcal{N}(\\mathbf{z};\\boldsymbol {\\mu},\\boldsymbol {\\sigma})" + }, + { + "bbox": [ + 47, + 117, + 287, + 165 + ], + "type": "text", + "content": " denotes Gaussian probability distribution, and " + }, + { + "bbox": [ + 47, + 117, + 287, + 165 + ], + "type": "inline_equation", + "content": "\\beta_{t}" + }, + { + "bbox": [ + 47, + 117, + 287, + 165 + ], + "type": "text", + "content": " is the noise schedule. The above equation leads to the analytic form of " + }, + { + "bbox": [ + 47, + 117, + 287, + 165 + ], + "type": "inline_equation", + "content": "\\mathbb{P}(\\mathbf{z}_t\\mid \\mathbf{z}_0),\\forall t\\in \\{0,1,\\dots ,T\\}" + }, + { + "bbox": [ + 47, + 117, + 287, + 165 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 111, + 171, + 287, + 192 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 111, + 171, + 287, + 192 + ], + "spans": [ + { + "bbox": [ + 111, + 171, + 287, + 192 + ], + "type": "interline_equation", + "content": "\\mathbf {z} _ {t} = \\sqrt {\\bar {\\beta} _ {t}} \\mathbf {z} _ {0} + \\sqrt {1 - \\bar {\\beta} _ {t}} \\boldsymbol {\\epsilon}, \\tag {2}", + "image_path": "480db9ccb49386e8f08ff88a5a7719d3e76d4fe2ab10fc530bbdc7212c4b6026.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 198, + 287, + 259 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 198, + 287, + 259 + ], + "spans": [ + { + "bbox": [ + 47, + 198, + 287, + 259 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 198, + 287, + 259 + ], + "type": "inline_equation", + "content": "\\bar{\\beta}_t = \\prod_{s=1}^t \\beta_s" + }, + { + "bbox": [ + 47, + 198, + 287, + 259 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 198, + 287, + 259 + ], + "type": "inline_equation", + "content": "\\epsilon \\sim \\mathcal{N}(\\mathbf{0},\\mathbf{I})" + }, + { + "bbox": [ + 47, + 198, + 287, + 259 + ], + "type": "text", + "content": ". The model then gradually removes noise by executing a learnable Markov chain in the reverse time direction, parameterized by a normal prior distribution " + }, + { + "bbox": [ + 47, + 198, + 287, + 259 + ], + "type": "inline_equation", + "content": "\\mathbb{P}(\\mathbf{z}_T) = \\mathcal{N}(\\mathbf{z}_T;0,\\mathbf{I})" + }, + { + "bbox": [ + 47, + 198, + 287, + 259 + ], + "type": "text", + "content": " and a learnable transition kernel " + }, + { + "bbox": [ + 47, + 198, + 287, + 259 + ], + "type": "inline_equation", + "content": "\\mathbb{P}_{\\theta}(\\mathbf{z}_{t-1}|\\mathbf{z}_t)" + }, + { + "bbox": [ + 47, + 198, + 287, + 259 + ], + "type": "text", + "content": " given by:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 61, + 265, + 287, + 278 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 265, + 287, + 278 + ], + "spans": [ + { + "bbox": [ + 61, + 265, + 287, + 278 + ], + "type": "interline_equation", + "content": "\\mathbb {P} _ {\\theta} \\left(\\mathbf {z} _ {t - 1} \\mid \\mathbf {z} _ {t}\\right) = \\mathcal {N} \\left(\\mathbf {z} _ {t - 1}; \\mu_ {\\theta} \\left(\\mathbf {z} _ {t}, t\\right), \\Sigma_ {\\theta} \\left(\\mathbf {z} _ {t}, t\\right)\\right). \\tag {3}", + "image_path": "a4fed940d439bb3075d7178276aaa97543b1571442b96fd548f282d1b821f06f.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 285, + 287, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 285, + 287, + 380 + ], + "spans": [ + { + "bbox": [ + 47, + 285, + 287, + 380 + ], + "type": "text", + "content": "The goal of the training process is to approximately match the reverse Markov chain with the actual time reversal of the forward Markov chain. Mathematically, parameter " + }, + { + "bbox": [ + 47, + 285, + 287, + 380 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 47, + 285, + 287, + 380 + ], + "type": "text", + "content": " is adjusted so that the joint distribution of the reverse Markov chain " + }, + { + "bbox": [ + 47, + 285, + 287, + 380 + ], + "type": "inline_equation", + "content": "\\mathbb{P}_{\\theta}(\\mathbf{z}_0,\\mathbf{z}_1,\\dots ,\\mathbf{z}_T)\\coloneqq \\mathbb{P}(\\mathbf{z}_T)\\prod_{t = 1}^T\\mathbb{P}_{\\theta}(\\mathbf{z}_{t - 1}\\mid \\mathbf{z}_t)" + }, + { + "bbox": [ + 47, + 285, + 287, + 380 + ], + "type": "text", + "content": " closely approximates that of the forward Markov chain " + }, + { + "bbox": [ + 47, + 285, + 287, + 380 + ], + "type": "inline_equation", + "content": "\\mathbb{P}(\\mathbf{z}_0,\\mathbf{z}_1,\\dots ,\\mathbf{z}_T)\\coloneqq \\mathbb{P}(\\mathbf{z}_0)\\prod_{t = 1}^T\\mathbb{P}(\\mathbf{z}_t\\mid \\mathbf{z}_{t - 1})" + }, + { + "bbox": [ + 47, + 285, + 287, + 380 + ], + "type": "text", + "content": ". This is achieved by minimizing the following loss:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 63, + 386, + 287, + 407 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 386, + 287, + 407 + ], + "spans": [ + { + "bbox": [ + 63, + 386, + 287, + 407 + ], + "type": "interline_equation", + "content": "\\mathbb {E} _ {t \\sim \\mathcal {U} [ 1, T ], \\mathbf {z} _ {0} \\sim \\mathbb {P} (\\mathbf {z} _ {0}), \\epsilon \\sim \\mathcal {N} (0, I)} \\left[ \\| \\epsilon - \\epsilon_ {\\theta} (\\mathbf {z} _ {t}, t) \\| ^ {2} \\right], \\tag {4}", + "image_path": "150d803d0ec8591c41e0d8914315f94d36a58eb39bc88ba483f420d9ef53f66d.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 411, + 287, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 411, + 287, + 460 + ], + "spans": [ + { + "bbox": [ + 47, + 411, + 287, + 460 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 411, + 287, + 460 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_t" + }, + { + "bbox": [ + 47, + 411, + 287, + 460 + ], + "type": "text", + "content": " is computed using Eq. (2), and " + }, + { + "bbox": [ + 47, + 411, + 287, + 460 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}" + }, + { + "bbox": [ + 47, + 411, + 287, + 460 + ], + "type": "text", + "content": " is predicted using a neural network, typically a UNet architecture [37]. In a conditional diffusion model " + }, + { + "bbox": [ + 47, + 411, + 287, + 460 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}(\\mathbf{z}_t,t)" + }, + { + "bbox": [ + 47, + 411, + 287, + 460 + ], + "type": "text", + "content": " gets replaced by " + }, + { + "bbox": [ + 47, + 411, + 287, + 460 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}(\\mathbf{z}_t,t,\\mathcal{C})" + }, + { + "bbox": [ + 47, + 411, + 287, + 460 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 47, + 411, + 287, + 460 + ], + "type": "inline_equation", + "content": "\\mathcal{C}" + }, + { + "bbox": [ + 47, + 411, + 287, + 460 + ], + "type": "text", + "content": " is a conditioning variable." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 465, + 151, + 477 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 465, + 151, + 477 + ], + "spans": [ + { + "bbox": [ + 47, + 465, + 151, + 477 + ], + "type": "text", + "content": "3.2. Our Architecture" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 484, + 287, + 674 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 484, + 287, + 674 + ], + "spans": [ + { + "bbox": [ + 47, + 484, + 287, + 674 + ], + "type": "text", + "content": "Image Encoder and Latent Diffusion. Diffusion models typically take large number of time steps to train, and are difficult to converge. Recently, [36] proposed a new diffusion with improved convergence properties and is called \"stable-diffusion\". The key idea is to perform the diffusion in latent space, with latent embedding learnt separately through a variational autoencoder (VAE). The Encoder of VAE first transforms the input image " + }, + { + "bbox": [ + 47, + 484, + 287, + 674 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 47, + 484, + 287, + 674 + ], + "type": "text", + "content": " of size " + }, + { + "bbox": [ + 47, + 484, + 287, + 674 + ], + "type": "inline_equation", + "content": "(H,W)" + }, + { + "bbox": [ + 47, + 484, + 287, + 674 + ], + "type": "text", + "content": " to latent space, then we follow latent diffusion formulation and utilize the UNet used in Stable Diffusion[36]. Utilizing latent diffusion formulation enables our architecture to capture multi-resolution features. Hence, we aggregate the feature maps from different layers of the UNet module (implementing conditional diffusion) by bringing them all to 1/4th resolution of the latent space, resulting in a feature map of size " + }, + { + "bbox": [ + 47, + 484, + 287, + 674 + ], + "type": "inline_equation", + "content": "8e\\times H / 32\\times W / 32" + }, + { + "bbox": [ + 47, + 484, + 287, + 674 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 677, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 677, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 677, + 287, + 714 + ], + "type": "text", + "content": "Exploiting Semantic Context with Conditional Diffusion. Recall that we formulated the single image depth estimation as a dense regression problem, predicting " + }, + { + "bbox": [ + 47, + 677, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\mathbb{P}(\\mathbf{y}\\mid \\mathbf{x})" + } + ] + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 366, + 70, + 385, + 171 + ], + "blocks": [ + { + "bbox": [ + 366, + 70, + 385, + 171 + ], + "lines": [ + { + "bbox": [ + 366, + 70, + 385, + 171 + ], + "spans": [ + { + "bbox": [ + 366, + 70, + 385, + 171 + ], + "type": "image", + "image_path": "c884c50e858aad2ed14417c86c14d50488599881dbd4715cd3e63339ec6572b7.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 371, + 173, + 380, + 180 + ], + "lines": [ + { + "bbox": [ + 371, + 173, + 380, + 180 + ], + "spans": [ + { + "bbox": [ + 371, + 173, + 380, + 180 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 457, + 70, + 495, + 171 + ], + "blocks": [ + { + "bbox": [ + 457, + 70, + 495, + 171 + ], + "lines": [ + { + "bbox": [ + 457, + 70, + 495, + 171 + ], + "spans": [ + { + "bbox": [ + 457, + 70, + 495, + 171 + ], + "type": "image", + "image_path": "57991546efa28fa9bd31fcb0cf70cb7c634b4e25f7be459feddfdac4e6a8ced9.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 473, + 173, + 481, + 180 + ], + "lines": [ + { + "bbox": [ + 473, + 173, + 481, + 180 + ], + "spans": [ + { + "bbox": [ + 473, + 173, + 481, + 180 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 304, + 190, + 545, + 289 + ], + "lines": [ + { + "bbox": [ + 304, + 190, + 545, + 289 + ], + "spans": [ + { + "bbox": [ + 304, + 190, + 545, + 289 + ], + "type": "text", + "content": "Figure 3. (a) Probabilistic graphical model corresponding to VPD. (b) The same corresponding to our formulation. Here, " + }, + { + "bbox": [ + 304, + 190, + 545, + 289 + ], + "type": "inline_equation", + "content": "\\mathcal{C}" + }, + { + "bbox": [ + 304, + 190, + 545, + 289 + ], + "type": "text", + "content": " represents the semantic embedding derived from our CIDE module. This embedding is internally generated by passing " + }, + { + "bbox": [ + 304, + 190, + 545, + 289 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 304, + 190, + 545, + 289 + ], + "type": "text", + "content": " through the ViT, resulting in " + }, + { + "bbox": [ + 304, + 190, + 545, + 289 + ], + "type": "inline_equation", + "content": "\\mathcal{E}" + }, + { + "bbox": [ + 304, + 190, + 545, + 289 + ], + "type": "text", + "content": ". Subsequently, " + }, + { + "bbox": [ + 304, + 190, + 545, + 289 + ], + "type": "inline_equation", + "content": "\\mathcal{E}" + }, + { + "bbox": [ + 304, + 190, + 545, + 289 + ], + "type": "text", + "content": " undergoes further processing to yield " + }, + { + "bbox": [ + 304, + 190, + 545, + 289 + ], + "type": "inline_equation", + "content": "\\mathcal{C}" + }, + { + "bbox": [ + 304, + 190, + 545, + 289 + ], + "type": "text", + "content": ", which is then utilized in the conditional diffusion module implementing " + }, + { + "bbox": [ + 304, + 190, + 545, + 289 + ], + "type": "inline_equation", + "content": "\\mathbb{P}(\\mathbf{z}_0\\mid \\mathbf{z}_t,\\mathcal{C})" + }, + { + "bbox": [ + 304, + 190, + 545, + 289 + ], + "type": "text", + "content": ". The output of the conditional diffusion module is fed into the Depth Regressor module within our architecture, implementing " + }, + { + "bbox": [ + 304, + 190, + 545, + 289 + ], + "type": "inline_equation", + "content": "\\mathbb{P}(\\mathbf{y}\\mid \\mathbf{z}_0)" + }, + { + "bbox": [ + 304, + 190, + 545, + 289 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 311, + 545, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 311, + 545, + 443 + ], + "spans": [ + { + "bbox": [ + 304, + 311, + 545, + 443 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 311, + 545, + 443 + ], + "type": "inline_equation", + "content": "\\mathbf{y}" + }, + { + "bbox": [ + 304, + 311, + 545, + 443 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 311, + 545, + 443 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 304, + 311, + 545, + 443 + ], + "type": "text", + "content": " denote output depth, and input image respectively. In our diffusion formulation, we predict " + }, + { + "bbox": [ + 304, + 311, + 545, + 443 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_0" + }, + { + "bbox": [ + 304, + 311, + 545, + 443 + ], + "type": "text", + "content": " which is then used to predict pixel-wise depth, " + }, + { + "bbox": [ + 304, + 311, + 545, + 443 + ], + "type": "inline_equation", + "content": "\\mathbf{y}" + }, + { + "bbox": [ + 304, + 311, + 545, + 443 + ], + "type": "text", + "content": ". It has been shown that noise prediction in a diffusion model, " + }, + { + "bbox": [ + 304, + 311, + 545, + 443 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}" + }, + { + "bbox": [ + 304, + 311, + 545, + 443 + ], + "type": "text", + "content": ", can be seen as predicting the gradient of the density function, " + }, + { + "bbox": [ + 304, + 311, + 545, + 443 + ], + "type": "inline_equation", + "content": "\\nabla_{\\mathbf{z}_t} \\log \\mathbb{P}(\\mathbf{z}_t)" + }, + { + "bbox": [ + 304, + 311, + 545, + 443 + ], + "type": "text", + "content": ". Hence, overall architecture for SIDE using a diffusion architecture can be seen as factorizing the conditional probability, as shown in the probabilistic graph model in Fig. 3a. To utilize additional semantic context generated from a ViT model, we condition it on the ViT embeddings as shown in Fig. 3b. Mathematically, we model:" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 314, + 465, + 536, + 479 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 465, + 536, + 479 + ], + "spans": [ + { + "bbox": [ + 314, + 465, + 536, + 479 + ], + "type": "interline_equation", + "content": "\\mathbb {P} (\\mathbf {y} \\mid \\mathbf {x}, \\mathcal {E}) = \\mathbb {P} (\\mathbf {y} \\mid \\mathbf {z} _ {0}) \\mathbb {P} (\\mathbf {z} _ {0} \\mid \\mathbf {z} _ {t}, \\mathcal {C}) \\mathbb {P} (\\mathbf {z} _ {t} \\mid \\mathbf {x}) \\mathbb {P} (\\mathcal {C} \\mid \\mathbf {x}),", + "image_path": "6dc85b7a09fbf94c8d7257d9b2624dfd756575a9eba5eef1a39a534b814a47cc.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 306, + 482, + 333, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 482, + 333, + 491 + ], + "spans": [ + { + "bbox": [ + 306, + 482, + 333, + 491 + ], + "type": "text", + "content": "where" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 326, + 496, + 545, + 510 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 496, + 545, + 510 + ], + "spans": [ + { + "bbox": [ + 326, + 496, + 545, + 510 + ], + "type": "interline_equation", + "content": "\\mathbb {P} (\\mathcal {C} \\mid \\mathbf {x}) = \\mathbb {P} (\\mathcal {C} \\mid \\mathcal {E}) \\mathbb {P} (\\mathcal {E} \\mid \\mathbf {x}). \\tag {5}", + "image_path": "0d1bd73234f0d5b120babbc9a9ced2df67bce9625cec8f3f7f4443751a24bc36.jpg" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 520, + 545, + 686 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 520, + 545, + 686 + ], + "spans": [ + { + "bbox": [ + 304, + 520, + 545, + 686 + ], + "type": "text", + "content": "Here, the first term, " + }, + { + "bbox": [ + 304, + 520, + 545, + 686 + ], + "type": "inline_equation", + "content": "\\mathbb{P}(\\mathbf{y}\\mid \\mathbf{z}_0)" + }, + { + "bbox": [ + 304, + 520, + 545, + 686 + ], + "type": "text", + "content": ", is implemented through the Depth Regressor module explained earlier. Similarly, " + }, + { + "bbox": [ + 304, + 520, + 545, + 686 + ], + "type": "inline_equation", + "content": "\\mathbb{P}(\\mathbf{z}_t\\mid \\mathbf{x})" + }, + { + "bbox": [ + 304, + 520, + 545, + 686 + ], + "type": "text", + "content": " is implemented using the VAE's Encoder as described earlier. We generate conditional information, " + }, + { + "bbox": [ + 304, + 520, + 545, + 686 + ], + "type": "inline_equation", + "content": "\\mathcal{C}" + }, + { + "bbox": [ + 304, + 520, + 545, + 686 + ], + "type": "text", + "content": " using our Comprehensive Image Detail Embedding module (hereafter CIDE, and explained below). The CIDE module takes " + }, + { + "bbox": [ + 304, + 520, + 545, + 686 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 304, + 520, + 545, + 686 + ], + "type": "text", + "content": " as input and generates embedding vector " + }, + { + "bbox": [ + 304, + 520, + 545, + 686 + ], + "type": "inline_equation", + "content": "\\mathcal{C}" + }, + { + "bbox": [ + 304, + 520, + 545, + 686 + ], + "type": "text", + "content": " (of dimension 768 in our design) as the output, thus, implementing " + }, + { + "bbox": [ + 304, + 520, + 545, + 686 + ], + "type": "inline_equation", + "content": "\\mathbb{P}(\\mathcal{C}\\mid \\mathbf{x})" + }, + { + "bbox": [ + 304, + 520, + 545, + 686 + ], + "type": "text", + "content": " given in Eq. (5). We use " + }, + { + "bbox": [ + 304, + 520, + 545, + 686 + ], + "type": "inline_equation", + "content": "\\mathcal{E}" + }, + { + "bbox": [ + 304, + 520, + 545, + 686 + ], + "type": "text", + "content": " to denote the embedding vector of a ViT module, and " + }, + { + "bbox": [ + 304, + 520, + 545, + 686 + ], + "type": "inline_equation", + "content": "\\mathbb{P}(\\mathcal{E}\\mid \\mathbf{x})" + }, + { + "bbox": [ + 304, + 520, + 545, + 686 + ], + "type": "text", + "content": " is implemented through the ViT. " + }, + { + "bbox": [ + 304, + 520, + 545, + 686 + ], + "type": "inline_equation", + "content": "\\mathbb{P}(\\mathcal{C}\\mid \\mathcal{E})" + }, + { + "bbox": [ + 304, + 520, + 545, + 686 + ], + "type": "text", + "content": " is implemented using downstream modules in CIDE consisting of learnable embeddings. The second term, " + }, + { + "bbox": [ + 304, + 520, + 545, + 686 + ], + "type": "inline_equation", + "content": "\\mathbb{P}(\\mathbf{z}_0\\mid \\mathbf{z}_t,\\mathcal{C})" + }, + { + "bbox": [ + 304, + 520, + 545, + 686 + ], + "type": "text", + "content": ", is implemented using conditional diffusion," + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 304, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 689, + 545, + 713 + ], + "type": "text", + "content": "Comprehensive Image Detail Embedding (CIDE) Module. As described earlier, we believe using pseudo-captions" + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "28288" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 65, + 123, + 531, + 393 + ], + "blocks": [ + { + "bbox": [ + 46, + 70, + 547, + 118 + ], + "lines": [ + { + "bbox": [ + 46, + 70, + 547, + 118 + ], + "spans": [ + { + "bbox": [ + 46, + 70, + 547, + 118 + ], + "type": "text", + "content": "Table 1. Results on Indoor NYU Depth v2 [27] Dataset. Results that are bold perform best. " + }, + { + "bbox": [ + 46, + 70, + 547, + 118 + ], + "type": "inline_equation", + "content": "\\uparrow" + }, + { + "bbox": [ + 46, + 70, + 547, + 118 + ], + "type": "text", + "content": " means the metric should be higher, " + }, + { + "bbox": [ + 46, + 70, + 547, + 118 + ], + "type": "inline_equation", + "content": "\\downarrow" + }, + { + "bbox": [ + 46, + 70, + 547, + 118 + ], + "type": "text", + "content": " indicate lower is better. The evaluation uses an upper bound of 10 meters on the ground truth depth map. All the numbers for other works have been taken from the corresponding papers. For MIM, and ZoEDepth we have used SwinV2-L 1K, and ZoeDepth-M12-N versions respectively. We see an overall improvement against SOTA on all the metrics used for evaluation." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 65, + 123, + 531, + 393 + ], + "lines": [ + { + "bbox": [ + 65, + 123, + 531, + 393 + ], + "spans": [ + { + "bbox": [ + 65, + 123, + 531, + 393 + ], + "type": "table", + "html": "
MethodVenueAbs Rel↓RMSE↓log10↓Sq Rel↓δ1↑δ2↑δ3↑
Eigen et al.[9]NIPS'140.1580.641--0.7690.9500.988
DORN[10]CVPR'180.1150.5090.051-0.8280.9650.992
SharpNet[32]ICCV'190.1390.5020.047-0.8360.9660.993
Chen et al.[5]IJCAI-190.1110.5140.048-0.8780.9770.994
BTS[22]Arxiv'190.1100.3920.0470.0660.8850.9780.994
AdaBins[2]CVPR'210.1030.3640.044-0.9030.9840.997
DPT[33]ICCV'210.1100.3570.045-0.9040.9880.998
P3Depth[30]CVPR'220.1040.3560.043-0.8980.9810.996
NeWCRFs[53]CVPR'220.0950.3340.0410.0450.9220.9920.998
SwinV2-B[24]CVPR'220.1330.4620.059-0.8190.9750.995
SwinV2-L[24]CVPR'220.1120.3810.051-0.8860.9840.997
Localbins[3]ECCV'220.0990.3570.042-0.9070.9870.998
Jun et al.[17]ECCV'220.0980.3550.042-0.9130.9870.998
PixelFormer[1]WACV'230.0900.3220.0390.0430.9290.9910.998
DDP[16]ICCV'230.0940.3290.040-0.9210.9900.998
MIM [51]CVPR'230.0830.2870.035-0.9490.9940.999
AiT[28]ICCV'230.0760.2750.033-0.9540.9940.999
ZoeDepth [4]Arxiv'230.0750.2700.0320.0300.9550.9950.999
VPD[54]ICCV'230.0690.2540.0300.0270.9640.9950.999
OursCVPR'240.0590.2180.0260.0130.9780.9970.999
", + "image_path": "85cecef06f97900f28fcbdaa57d7bdd2247ae7c37b92b6324ec104478fa24f85.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 97, + 402, + 497, + 654 + ], + "blocks": [ + { + "bbox": [ + 97, + 402, + 497, + 654 + ], + "lines": [ + { + "bbox": [ + 97, + 402, + 497, + 654 + ], + "spans": [ + { + "bbox": [ + 97, + 402, + 497, + 654 + ], + "type": "image", + "image_path": "a76cc74dd06135df0021484b881dda524bb0070b423a6e292f03cff9e997573e.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 662, + 548, + 686 + ], + "lines": [ + { + "bbox": [ + 46, + 662, + 548, + 686 + ], + "spans": [ + { + "bbox": [ + 46, + 662, + 548, + 686 + ], + "type": "text", + "content": "Figure 4. Visual Comparison on NYU Depth v2 Indoor Dataset. Note, our method's ability to delineate objects in terms of their depth, such as the table lamp in Row 5, even when such information is absent from the ground truth depth map." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 319, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 319, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 319, + 757 + ], + "type": "text", + "content": "28289" + } + ] + } + ], + "index": 4 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 52, + 145, + 544, + 357 + ], + "blocks": [ + { + "bbox": [ + 46, + 70, + 547, + 138 + ], + "lines": [ + { + "bbox": [ + 46, + 70, + 547, + 138 + ], + "spans": [ + { + "bbox": [ + 46, + 70, + 547, + 138 + ], + "type": "text", + "content": "Table 2. Performance on the Outdoor KITTI [12] Dataset. Please refer to the caption of Tab. 1 for notation details. For ZoEDepth results, we use the ZoeDepth-M12-K version following the authors' recommendation. In instances where results for certain methods were not reported in the respective works, denoted by “-”, and the code is unavailable, we were unable to generate the missing numbers. Despite the saturation of results on the outdoor KITTI dataset, our method consistently achieves comparable or superior performance to the state-of-the-art (SOTA) across all metrics. VPD[54] cannot be trained on KITTI dataset as it required a per image text label which is not present in KITTI. The symbol " + }, + { + "bbox": [ + 46, + 70, + 547, + 138 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 46, + 70, + 547, + 138 + ], + "type": "text", + "content": " indicates that the method utilizes additional information beyond RGB." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 52, + 145, + 544, + 357 + ], + "lines": [ + { + "bbox": [ + 52, + 145, + 544, + 357 + ], + "spans": [ + { + "bbox": [ + 52, + 145, + 544, + 357 + ], + "type": "table", + "html": "
MethodVenueAbs Rel↓Sq Rel↓RMSElog ↓RMSE↓δ1 ↑δ2 ↑δ3 ↑
Eigen et al.[9]NIPS'140.2031.5170.2826.3070.7020.8980.967
DORN[10]CVPR'180.0720.3070.1202.7270.9320.9840.994
BTS[22]Arxiv'190.0590.2410.0962.7560.9560.9930.998
AdaBins[2]CVPR'210.0670.1900.0882.9600.9490.9920.998
DPT[33]ICCV'210.060-0.0922.5730.9590.9950.996
P3Depth[30]CVPR'220.0710.2700.1032.8420.9530.9930.998
NeWCRFs[53]CVPR'220.0520.1550.0792.1290.9740.9970.999
PixelFormer[1]WACV'230.0510.1490.0772.0810.9760.9970.999
ZoeDepth [4]Arxiv'230.0540.1890.0832.4400.970.9960.999
DDP [16]ICCV'230.0500.1480.0762.0720.9750.9970.999
URCDC [40]ToM'230.0500.1420.0762.0320.9770.9970.999
IEBins [41]NeurIPS'230.0500.1420.0752.0110.9780.9980.999
MIM [51]CVPR'230.0500.1390.0751.9660.9770.9981.000
GEDepth[52]†ICCV'230.0480.1420.0762.0440.9760.9970.999
OursCVPR'240.0480.1390.0742.0390.9790.9981.000
", + "image_path": "7e3acb0e160e30d314f36226aad83e643f08cfb3b7d455b1bbed0204d4726d20.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 48, + 365, + 545, + 559 + ], + "blocks": [ + { + "bbox": [ + 48, + 365, + 545, + 559 + ], + "lines": [ + { + "bbox": [ + 48, + 365, + 545, + 559 + ], + "spans": [ + { + "bbox": [ + 48, + 365, + 545, + 559 + ], + "type": "image", + "image_path": "fd5626577cb0a6d7c11ad79b7f023d80f1eaefeb0fdc9d0368fc06a025dcee80.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 185, + 566, + 407, + 578 + ], + "lines": [ + { + "bbox": [ + 185, + 566, + 407, + 578 + ], + "spans": [ + { + "bbox": [ + 185, + 566, + 407, + 578 + ], + "type": "text", + "content": "Figure 5. Visual Comparison on KITTI Outdoor Dataset." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 599, + 289, + 706 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 599, + 289, + 706 + ], + "spans": [ + { + "bbox": [ + 46, + 599, + 289, + 706 + ], + "type": "text", + "content": "to generate the semantic context has limited utility, as the textual descriptions typically focus on large salient objects only. Instead we propose our CIDE module which use embeddings from a pre-trained ViT, and extract detailed semantic context from these embeddings. For this we take 1000 dimensional logit vector from ViT, and pass it through a two layer MLP which converts it to a 100 dimensional vector. Subsequently, we employ this vector to compute the linear combination of 100 learnable embeddings." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 599, + 545, + 634 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 599, + 545, + 634 + ], + "spans": [ + { + "bbox": [ + 304, + 599, + 545, + 634 + ], + "type": "text", + "content": "This resulting embedding undergoes a linear transformation to yield a semantic context vector of dimension 768, which is then passed to conditional diffusion module." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "type": "text", + "content": "Depth Regressor. The output feature map undergoes through an Upsampling Decoder, comprised of deconvolution layers, followed by a Depth Regressor. The Depth Regressor is essentially a two-layer convolutional neural network (CNN), with the initial layer having dimensions " + }, + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{Conv}(3\\times 3)" + }, + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "type": "text", + "content": ", 192, and the subsequent layer " + }, + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{Conv}(3\\times 3)" + }, + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "type": "text", + "content": ", 1." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "28290" + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 47, + 125, + 559, + 228 + ], + "blocks": [ + { + "bbox": [ + 46, + 70, + 545, + 115 + ], + "lines": [ + { + "bbox": [ + 46, + 70, + 545, + 115 + ], + "spans": [ + { + "bbox": [ + 46, + 70, + 545, + 115 + ], + "type": "text", + "content": "Table 3. Quantitative results for zero-shot transfer to four unseen indoor datasets. mRI denotes the mean relative improvement with respect to NeWCRFs across all metrics (δ1, REL, RMSE). Evaluation depth is capped at 8m for SUN RGB-D, 10m for iBims and DIODE Indoor, and 80m for HyperSim. Best results are in bold, second best are underlined. Our mRI outperforms all methods across all datasets by a large margin. † denotes that ZoeD is trained on 12 datasets and our method is trained only on NYUv2." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 47, + 125, + 559, + 228 + ], + "lines": [ + { + "bbox": [ + 47, + 125, + 559, + 228 + ], + "spans": [ + { + "bbox": [ + 47, + 125, + 559, + 228 + ], + "type": "table", + "html": "
MethodSUN RGB-DiBims-1 BenchmarkDIODE IndoorHyperSim
δ1↑REL↓RMSE↓mRIθ↑δ1↑REL↓RMSE↓mRIθ↑δ1↑REL↓RMSE↓mRIθ↑δ1↑REL↓RMSE↓mRIθ↑
BTS [22]0.7400.1720.515-14.2%0.5380.2310.919-6.9%0.2100.4181.9052.3%0.2250.4766.404-8.6%
AdaBins [2]0.7710.1590.476-7.0%0.5550.2120.901-2.1%0.1740.4431.963-7.2%0.2210.4836.546-10.5%
LocalBins [3]0.7770.1560.470-5.6%0.5580.2110.880-0.7%0.2290.4121.8537.1%0.2340.4686.362-6.6%
NeWCRFs [53]0.7980.1510.4240.0%0.5480.2060.8610.0%0.1870.4041.8670.0%0.2550.4426.0170.0%
VPD [54]0.8610.1210.35514.7%0.6270.1870.76711.5%0.4800.3921.29563.4%0.3330.5315.1118.5%
ZoeD-M12-N† [4]0.8640.1190.34616.0%0.6580.1690.71118.5%0.3760.3271.58845.0%0.2920.4105.7718.6%
Ours0.8850.1120.31920.5%0.6880.1630.66423.1%0.5450.3441.16481.3%0.3940.4424.73925.2%
", + "image_path": "2e0b5f3c63f0cc8f262938f75cd81372b094e58cf9940af92c8e9463c4040b78.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 83, + 242, + 518, + 365 + ], + "blocks": [ + { + "bbox": [ + 83, + 242, + 518, + 365 + ], + "lines": [ + { + "bbox": [ + 83, + 242, + 518, + 365 + ], + "spans": [ + { + "bbox": [ + 83, + 242, + 518, + 365 + ], + "type": "image", + "image_path": "10a7e4cdf5394b729bfc2c2fda0f2f9fb4531171367e28fe6cc3a38ac46faaed.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 83, + 379, + 517, + 506 + ], + "blocks": [ + { + "bbox": [ + 83, + 379, + 517, + 506 + ], + "lines": [ + { + "bbox": [ + 83, + 379, + 517, + 506 + ], + "spans": [ + { + "bbox": [ + 83, + 379, + 517, + 506 + ], + "type": "image", + "image_path": "f31058aa86b9b19c5376f9c5dabda1e36277891d933e7679fad2e97499e3bd57.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 516, + 544, + 549 + ], + "lines": [ + { + "bbox": [ + 46, + 516, + 544, + 549 + ], + "spans": [ + { + "bbox": [ + 46, + 516, + 544, + 549 + ], + "type": "text", + "content": "Figure 6. Visualization of improvements over VPD [54] in our model due to ViT embeddings passed down as conditional vectors (in blue and green). In the above images, ViT detects the desktop computer (first image) and table lamp (second image) with high probability, and they are thus better detected by our model. Additional visualizations are provided in the supplementary material." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 569, + 190, + 582 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 569, + 190, + 582 + ], + "spans": [ + { + "bbox": [ + 47, + 569, + 190, + 582 + ], + "type": "text", + "content": "4. Experiments and Results" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 594, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 288, + 713 + ], + "type": "text", + "content": "Datasets and Evaluation. We use NYU Depth v2 [27] and KITTI [12] as the primary datasets for training. The NYU Depth v2 dataset is a widely used indoor benchmark for monocular depth estimation, containing over 24k densely labeled pairs of RGB and depth images in the train set and 654 in the test set. The dataset covers a wide range of indoor scenes and includes challenging scenarios such as reflective surfaces, transparent objects, and occlusions. The ground truth depth maps are obtained using a structured light sensor and are provided at a resolution of " + }, + { + "bbox": [ + 46, + 594, + 288, + 713 + ], + "type": "inline_equation", + "content": "640 \\times 480" + }, + { + "bbox": [ + 46, + 594, + 288, + 713 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 570, + 545, + 703 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 545, + 703 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 545, + 703 + ], + "type": "text", + "content": "KITTI dataset is a widely used outdoor benchmark for monocular depth estimation, containing over 24k densely labeled pairs of RGB and depth images. The dataset covers outdoor driving scenarios and includes varying lighting conditions, weather, and occlusions. The ground truth depth maps are obtained using a Velodyne LiDAR sensor and are provided at a resolution of " + }, + { + "bbox": [ + 304, + 570, + 545, + 703 + ], + "type": "inline_equation", + "content": "1242 \\times 375" + }, + { + "bbox": [ + 304, + 570, + 545, + 703 + ], + "type": "text", + "content": ". To demonstrate generalizability, we evaluate zero-shot performance on the following datasets: Sun-RGBD [42], iBims1 [19], DIODE [49], and HyperSim [35]. Whereas, the main paper contains mostly quantitative, and only some representative" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "28291" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 51, + 167, + 284, + 241 + ], + "blocks": [ + { + "bbox": [ + 47, + 70, + 288, + 159 + ], + "lines": [ + { + "bbox": [ + 47, + 70, + 288, + 159 + ], + "spans": [ + { + "bbox": [ + 47, + 70, + 288, + 159 + ], + "type": "text", + "content": "Table 4. Effectiveness of different embeddings for guiding the diffusion process for depth estimation on NYU Depth v2 dataset. In the third row, rather than using pseudo caption embeddings generated from the scene label (as implemented by VPD [54]), we provide the one-hot vector representing the scene label as a condition to the diffusion model. We observe a slight improvement in the metrics, highlighting that information content in the caption is similar to that in one-hot label." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 51, + 167, + 284, + 241 + ], + "lines": [ + { + "bbox": [ + 51, + 167, + 284, + 241 + ], + "spans": [ + { + "bbox": [ + 51, + 167, + 284, + 241 + ], + "type": "table", + "html": "
EmbeddingsRMSE↓Abs Rel↓δ1↑
Scene label emb. [54]0.2540.0690.964
Text caption emb. [20]0.2250.0620.976
One-hot vector emb.0.2440.0670.968
Proposed scene emb.0.2180.0590.978
", + "image_path": "5fdba6a9d50ed2fb62d14f37e682b07bbb0186f63efe75d3ebcd71721d942fd5.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 261, + 287, + 285 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 261, + 287, + 285 + ], + "spans": [ + { + "bbox": [ + 47, + 261, + 287, + 285 + ], + "type": "text", + "content": "tive visual results, detailed visual results on each dataset are included in the supplementary." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 287, + 287, + 431 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 287, + 287, + 431 + ], + "spans": [ + { + "bbox": [ + 46, + 287, + 287, + 431 + ], + "type": "text", + "content": "Implementation Details. Our model is implemented using PyTorch [29]. For optimization, we have used AdamW optimizer [25] with " + }, + { + "bbox": [ + 46, + 287, + 287, + 431 + ], + "type": "inline_equation", + "content": "\\beta_0" + }, + { + "bbox": [ + 46, + 287, + 287, + 431 + ], + "type": "text", + "content": " values of 0.9 and 0.999, a batch size of 32, and a weight decay of 0.1. We train our model for 25 epochs for both KITTI and NYU Depth v2 datasets, with an initial learning rate of " + }, + { + "bbox": [ + 46, + 287, + 287, + 431 + ], + "type": "inline_equation", + "content": "3 \\times 10^{-5}" + }, + { + "bbox": [ + 46, + 287, + 287, + 431 + ], + "type": "text", + "content": ". We first linearly increase learning rate to " + }, + { + "bbox": [ + 46, + 287, + 287, + 431 + ], + "type": "inline_equation", + "content": "5 \\times 10^{-4}" + }, + { + "bbox": [ + 46, + 287, + 287, + 431 + ], + "type": "text", + "content": ", and then linearly decrease across training iterations. We use usual data augmentation techniques, including random hue addition, horizontal flipping, changing the image brightness, and Cut Depth [15]. Our model takes approximately 21 minutes per epoch to train using 8 NVIDIA A100 GPUs." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 439, + 242, + 451 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 439, + 242, + 451 + ], + "spans": [ + { + "bbox": [ + 47, + 439, + 242, + 451 + ], + "type": "text", + "content": "4.1. Comparison on Benchmark Datasets" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 457, + 287, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 457, + 287, + 540 + ], + "spans": [ + { + "bbox": [ + 46, + 457, + 287, + 540 + ], + "type": "text", + "content": "Comparison on NYU Depth v2. Tab. 1 shows the comparison of our proposed method with SOTA methods on the indoor NYU Depth v2 dataset [27]. We achieve a new state of the art on this dataset. Our methods perform better than the previous SOTA ([54])) by a large margin of " + }, + { + "bbox": [ + 46, + 457, + 287, + 540 + ], + "type": "inline_equation", + "content": "14\\%" + }, + { + "bbox": [ + 46, + 457, + 287, + 540 + ], + "type": "text", + "content": " in terms of RMSE. Fig. 4 provides a qualitative comparison on the dataset." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 543, + 287, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 543, + 287, + 616 + ], + "spans": [ + { + "bbox": [ + 46, + 543, + 287, + 616 + ], + "type": "text", + "content": "Comparison on KITTI. Tab. 2 shows the comparison with various methods on the KITTI dataset [12]. Fig. 5 shows the qualitative results. Unlike NYU Depth v2 dataset, KITTI is an outdoor dataset. On this dataset also, we achieve similar or better performance than all existing state-of-the-art techniques." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 623, + 252, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 623, + 252, + 635 + ], + "spans": [ + { + "bbox": [ + 47, + 623, + 252, + 635 + ], + "type": "text", + "content": "4.2. Generalization and Zero Shot Transfer" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 641, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 641, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 641, + 287, + 715 + ], + "type": "text", + "content": "Unlike state of the art on zero short transfer (ZoEDepth [4]), which requiring training on many datasets (12 in their case) for effective zero shot transfer, we show that our model generalizes well to other unseen dataset even when trained on a single NYU Depth v2 dataset. Tab. 3 shows quantitative results to back our findings." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 72, + 400, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 400, + 85 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 400, + 85 + ], + "type": "text", + "content": "4.3. Ablation Study" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 90, + 545, + 247 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 90, + 545, + 247 + ], + "spans": [ + { + "bbox": [ + 304, + 90, + 545, + 247 + ], + "type": "text", + "content": "Effect of Contextual Information. As highlighted in the motivation, a key observation of this study is the richness of information contained within the output probability vector of the ViT, surpassing the textual embeddings employed in current state-of-the-art (e.g., VPD [54]). To compare with the utilization of scene label information (as implemented by VPD), we construct the conditioning embedding as a one-hot vector using the scene label, and subsequently transform it using an MLP. As illustrated in Tab. 4, we observe a slight improvement over VPD, indicating that the information content in pseudo-captions resembles that of one-hot labels. Furthermore, our proposed method surpasses both the approaches." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 247, + 545, + 369 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 247, + 545, + 369 + ], + "spans": [ + { + "bbox": [ + 304, + 247, + 545, + 369 + ], + "type": "text", + "content": "Qualitative Results. Perhaps the most important task would be to verify that use of rich probability vector instead of text embeddings actually results in an improvement in depth as a direct consequence. We do this by considering the top few objects predicted by the ViT and correlating this with the depth predicted at these objects. We hypothesise when a particular class say dog is predicted with a high probability (hence there must be a dog in the image), then the corresponding depth must also be more accurate. We show this in Fig. 6." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 374, + 378, + 386 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 374, + 378, + 386 + ], + "spans": [ + { + "bbox": [ + 306, + 374, + 378, + 386 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 394, + 545, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 394, + 545, + 574 + ], + "spans": [ + { + "bbox": [ + 304, + 394, + 545, + 574 + ], + "type": "text", + "content": "We presented a new architecture block Comprehensive Image Detail Embedding (CIDE) module for robust monocular depth estimation in this paper. Our key idea is to highlight the limitations of using pseudo-captions to provide contextual information, and instead propose to use richer class-wise probability generated by a classification model, such as ViT. The motivation is that, while textual embedding typically highlights salient objects, class-wise probability vector preserves more details, including smaller objects in the background also. We implement the idea using proposed CIDE module cascaded with a conditional diffusion pipeline for monocular depth estimation. We demonstrated the effectiveness of our approach on several benchmark datasets and showed that it outperforms SOTA methods by a significant margin." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 576, + 545, + 661 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 576, + 545, + 661 + ], + "spans": [ + { + "bbox": [ + 304, + 576, + 545, + 661 + ], + "type": "text", + "content": "Acknowledgement. We thank Kartik Anand for assistance with some of the experiments. This work has been supported by funding through Staqu Technologies Pvt. Ltd., and Department of Science and Technology (DST), Government of India. The travel for Suraj Patni is supported through Research Acceleration funds of Department of Computer Science and Engineering, IIT Delhi." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 671, + 365, + 684 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 671, + 365, + 684 + ], + "spans": [ + { + "bbox": [ + 306, + 671, + 365, + 684 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 310, + 691, + 545, + 715 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 310, + 691, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 310, + 691, + 545, + 715 + ], + "type": "text", + "content": "[1] Ashutosh Agarwal and Chetan Arora. Attention attention everywhere: Monocular depth prediction with skip attention." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "28292" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 714 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 66, + 72, + 287, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 72, + 287, + 106 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 287, + 106 + ], + "type": "text", + "content": "In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 5861-5870, 2023. 2, 5, 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 53, + 107, + 288, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 107, + 288, + 162 + ], + "spans": [ + { + "bbox": [ + 53, + 107, + 288, + 162 + ], + "type": "text", + "content": "[2] Shariq Farooq Bhat, Ibrahim Alhashim, and Peter Wonka. Adabins: Depth estimation using adaptive bins. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4009-4018, 2021. 5, 6, 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 163, + 288, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 163, + 288, + 207 + ], + "spans": [ + { + "bbox": [ + 53, + 163, + 288, + 207 + ], + "type": "text", + "content": "[3] Shariq Farooq Bhat, Ibraheem Alhashim, and Peter Wonka. Localbins: Improving depth estimation by learning local distributions. In European Conference on Computer Vision, pages 480-496. Springer, 2022. 5, 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 209, + 288, + 241 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 209, + 288, + 241 + ], + "spans": [ + { + "bbox": [ + 53, + 209, + 288, + 241 + ], + "type": "text", + "content": "[4] Shariq Farooq Bhat, Reiner Birkl, Diana Wofk, Peter Wonka, and Matthias Müller. Zoedepth: Zero-shot transfer by combining relative and metric depth, 2023. 2, 3, 5, 6, 7, 8" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 243, + 288, + 275 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 243, + 288, + 275 + ], + "spans": [ + { + "bbox": [ + 53, + 243, + 288, + 275 + ], + "type": "text", + "content": "[5] Xiaotian Chen, Xuejin Chen, and Zheng-Jun Zha. Structure-aware residual pyramid network for monocular depth estimation. arXiv preprint arXiv:1907.06023, 2019. 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 277, + 288, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 277, + 288, + 331 + ], + "spans": [ + { + "bbox": [ + 53, + 277, + 288, + 331 + ], + "type": "text", + "content": "[6] Angela Dai, Angel X Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5828-5839, 2017. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 333, + 288, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 333, + 288, + 397 + ], + "spans": [ + { + "bbox": [ + 53, + 333, + 288, + 397 + ], + "type": "text", + "content": "[7] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 2, 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 399, + 288, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 399, + 288, + 431 + ], + "spans": [ + { + "bbox": [ + 53, + 399, + 288, + 431 + ], + "type": "text", + "content": "[8] Yiqun Duan, Xianda Guo, and Zheng Zhu. Diffusiondepth: Diffusion denoising approach for monocular depth estimation, 2023. 2, 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 434, + 288, + 476 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 434, + 288, + 476 + ], + "spans": [ + { + "bbox": [ + 53, + 434, + 288, + 476 + ], + "type": "text", + "content": "[9] David Eigen, Christian Puhrsch, and Rob Fergus. Depth map prediction from a single image using a multi-scale deep network. Advances in neural information processing systems, 27, 2014. 2, 5, 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 478, + 288, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 478, + 288, + 533 + ], + "spans": [ + { + "bbox": [ + 48, + 478, + 288, + 533 + ], + "type": "text", + "content": "[10] Huan Fu, Mingming Gong, Chaohui Wang, Kayhan Bat-manghelich, and Dacheng Tao. Deep ordinal regression network for monocular depth estimation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2002-2011, 2018. 5, 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 535, + 288, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 535, + 288, + 589 + ], + "spans": [ + { + "bbox": [ + 48, + 535, + 288, + 589 + ], + "type": "text", + "content": "[11] Huan Fu, Mingming Gong, Chaohui Wang, Kayhan Bat-manghelich, and Dacheng Tao. Deep ordinal regression network for monocular depth estimation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2002-2011, 2018. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 590, + 288, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 590, + 288, + 633 + ], + "spans": [ + { + "bbox": [ + 48, + 590, + 288, + 633 + ], + "type": "text", + "content": "[12] Andreas Geiger, Philip Lenz, Christoph Stiller, and Raquel Urtasun. Vision meets robotics: The kitti dataset. International Journal of Robotics Research (IJRR), 2013. 3, 6, 7, 8" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 635, + 288, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 635, + 288, + 678 + ], + "spans": [ + { + "bbox": [ + 48, + 635, + 288, + 678 + ], + "type": "text", + "content": "[13] José L. Herrera, Carlos R. del Blanco, and Narciso García. Automatic depth extraction from 2d images using a cluster-based learning framework. IEEE Transactions on Image Processing, 27(7):3288-3299, 2018. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 680, + 288, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 680, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 48, + 680, + 288, + 714 + ], + "type": "text", + "content": "[14] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 3" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 546, + 714 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 307, + 73, + 546, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 546, + 106 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 546, + 106 + ], + "type": "text", + "content": "[15] Yasunori Ishii and Takayoshi Yamashita. Cutdepth: Edge-aware data augmentation in depth estimation. arXiv preprint arXiv:2107.07684, 2021. 8" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 107, + 546, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 107, + 546, + 150 + ], + "spans": [ + { + "bbox": [ + 307, + 107, + 546, + 150 + ], + "type": "text", + "content": "[16] Yuanfeng Ji, Zhe Chen, Enze Xie, Lanqing Hong, Xihui Liu, Zhaoqiang Liu, Tong Lu, Zhenguo Li, and Ping Luo. Ddp: Diffusion model for dense visual prediction. arXiv preprint arXiv:2303.17559, 2023. 2, 3, 5, 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 152, + 546, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 152, + 546, + 196 + ], + "spans": [ + { + "bbox": [ + 307, + 152, + 546, + 196 + ], + "type": "text", + "content": "[17] Jinyoung Jun, Jae-Han Lee, Chul Lee, and Chang-Su Kim. Depth map decomposition for monocular depth estimation. In European Conference on Computer Vision, pages 18-34. Springer, 2022. 5" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 198, + 546, + 241 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 198, + 546, + 241 + ], + "spans": [ + { + "bbox": [ + 307, + 198, + 546, + 241 + ], + "type": "text", + "content": "[18] Kevin Karsch, Ce Liu, and Sing Bing Kang. Depth transfer: Depth extraction from video using non-parametric sampling. IEEE transactions on pattern analysis and machine intelligence, 36(11):2144-2158, 2014. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 243, + 546, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 243, + 546, + 296 + ], + "spans": [ + { + "bbox": [ + 307, + 243, + 546, + 296 + ], + "type": "text", + "content": "[19] Tobias Koch, Lukas Liebel, Friedrich Fraundorfer, and Marco Korner. Evaluation of cnn-based single-image depth estimation methods. In Proceedings of the European Conference on Computer Vision (ECCV) Workshops, pages 0–0, 2018. 1, 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 298, + 546, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 298, + 546, + 342 + ], + "spans": [ + { + "bbox": [ + 307, + 298, + 546, + 342 + ], + "type": "text", + "content": "[20] Neehar Kondapaneni, Markus Marks, Manuel Knott, Rogério Guimarães, and Pietro Perona. Text-image alignment for diffusion-based perception. arXiv preprint arXiv:2310.00031, 2023. 2, 3, 8" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 343, + 546, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 343, + 546, + 397 + ], + "spans": [ + { + "bbox": [ + 307, + 343, + 546, + 397 + ], + "type": "text", + "content": "[21] Iro Laina, Christian Rupprecht, Vasileios Belagiannis, Federico Tombari, and Nassir Navab. Deeper depth prediction with fully convolutional residual networks. In 2016 Fourth international conference on 3D vision (3DV), pages 239-248. IEEE, 2016. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 399, + 546, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 399, + 546, + 442 + ], + "spans": [ + { + "bbox": [ + 307, + 399, + 546, + 442 + ], + "type": "text", + "content": "[22] Jin Han Lee, Myung-Kyu Han, Dong Wook Ko, and Il Hong Suh. From big to small: Multi-scale local planar guidance for monocular depth estimation. arXiv preprint arXiv:1907.10326, 2019. 5, 6, 7" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 444, + 546, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 444, + 546, + 498 + ], + "spans": [ + { + "bbox": [ + 307, + 444, + 546, + 498 + ], + "type": "text", + "content": "[23] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In International Conference on Machine Learning, pages 12888-12900. PMLR, 2022. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 500, + 546, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 500, + 546, + 555 + ], + "spans": [ + { + "bbox": [ + 307, + 500, + 546, + 555 + ], + "type": "text", + "content": "[24] Ze Liu, Han Hu, Yutong Lin, Zhuliang Yao, Zhenda Xie, Yixuan Wei, Jia Ning, Yue Cao, Zheng Zhang, Li Dong, et al. Swin transformer v2: Scaling up capacity and resolution. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12009-12019, 2022. 5" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 557, + 546, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 557, + 546, + 578 + ], + "spans": [ + { + "bbox": [ + 307, + 557, + 546, + 578 + ], + "type": "text", + "content": "[25] Ilya Loshchilov and Frank Hutter. Decoupled Weight Decay Regularization. In ICML 2019. 8" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 580, + 546, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 580, + 546, + 624 + ], + "spans": [ + { + "bbox": [ + 307, + 580, + 546, + 624 + ], + "type": "text", + "content": "[26] John McCormac, Ankur Handa, Stefan Leutenegger, and Andrew J Davison. Scelenet rgb-d: 5m photorealistic images of synthetic indoor trajectories with ground truth. arXiv preprint arXiv:1612.05079, 2016. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 624, + 546, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 624, + 546, + 657 + ], + "spans": [ + { + "bbox": [ + 307, + 624, + 546, + 657 + ], + "type": "text", + "content": "[27] Pushmeet Kohli Nathan Silberman, Derek Hoiem and Rob Fergus. Indoor segmentation and support inference from rgbd images. In ECCV, 2012. 3, 5, 7, 8" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 658, + 546, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 658, + 546, + 714 + ], + "spans": [ + { + "bbox": [ + 307, + 658, + 546, + 714 + ], + "type": "text", + "content": "[28] Jia Ning, Chen Li, Zheng Zhang, Chunyu Wang, Zigang Geng, Qi Dai, Kun He, and Han Hu. All in tokens: Unifying output space of visual tasks via soft token. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 19900-19910, 2023. 2, 5" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "28293" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 171 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 171 + ], + "type": "text", + "content": "[29] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, Alban Desmaison, Andreas Kopf, Edward Yang, Zachary DeVito, Martin Raison, Alykhan Tejani, Sasank Chilamkurthy, Benoit Steiner, Lu Fang, Junjie Bai, and Soumith Chintala. Pytorch: An imperative style, high-performance deep learning library. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2019. 8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 173, + 287, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 173, + 287, + 228 + ], + "spans": [ + { + "bbox": [ + 48, + 173, + 287, + 228 + ], + "type": "text", + "content": "[30] Vaishakh Patil, Christos Sakaridis, Alexander Liniger, and Luc Van Gool. P3depth: Monocular depth estimation with a piecewise planarity prior. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1610–1621, 2022. 5, 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 229, + 287, + 295 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 229, + 287, + 295 + ], + "spans": [ + { + "bbox": [ + 48, + 229, + 287, + 295 + ], + "type": "text", + "content": "[31] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 297, + 287, + 351 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 297, + 287, + 351 + ], + "spans": [ + { + "bbox": [ + 48, + 297, + 287, + 351 + ], + "type": "text", + "content": "[32] Michael Ramamonjisoa and Vincent Lepetit. Sharpnet: Fast and accurate recovery of occluding contours in monocular depth estimation. In Proceedings of the IEEE/CVF International Conference on Computer Vision Workshops, pages 0–0, 2019. 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 353, + 288, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 353, + 288, + 397 + ], + "spans": [ + { + "bbox": [ + 48, + 353, + 288, + 397 + ], + "type": "text", + "content": "[33] René Ranftl, Alexey Bochkovskiy, and Vladlen Koltun. Vision transformers for dense prediction. In Proceedings of the IEEE/CVF international conference on computer vision, pages 12179-12188, 2021. 5, 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 399, + 287, + 454 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 399, + 287, + 454 + ], + "spans": [ + { + "bbox": [ + 48, + 399, + 287, + 454 + ], + "type": "text", + "content": "[34] René Ranftl, Katrin Lasinger, David Hafner, Konrad Schindler, and Vladlen Koltun. Towards robust monocular depth estimation: Mixing datasets for zero-shot cross-dataset transfer. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(3), 2022. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 455, + 287, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 455, + 287, + 520 + ], + "spans": [ + { + "bbox": [ + 48, + 455, + 287, + 520 + ], + "type": "text", + "content": "[35] Mike Roberts, Jason Ramapuram, Anurag Ranjan, Atulit Kumar, Miguel Angel Bautista, Nathan Paczan, Russ Webb, and Joshua M. Susskind. Hypersim: A photorealistic synthetic dataset for holistic indoor scene understanding. In International Conference on Computer Vision (ICCV) 2021, 2021. 1, 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 522, + 287, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 522, + 287, + 578 + ], + "spans": [ + { + "bbox": [ + 48, + 522, + 287, + 578 + ], + "type": "text", + "content": "[36] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 578, + 287, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 578, + 287, + 645 + ], + "spans": [ + { + "bbox": [ + 48, + 578, + 287, + 645 + ], + "type": "text", + "content": "[37] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical Image Computing and Computer-Assisted Intervention-MICCAI 2015: 18th International Conference, Munich, Germany, October 5-9, 2015, Proceedings, Part III 18, pages 234-241. Springer, 2015. 3, 4" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 647, + 287, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 647, + 287, + 678 + ], + "spans": [ + { + "bbox": [ + 48, + 647, + 287, + 678 + ], + "type": "text", + "content": "[38] Saurabh Saxena, Abhishek Kar, Mohammad Norouzi, and David J Fleet. Monocular depth estimation using diffusion models. arXiv preprint arXiv:2302.14816, 2023. 2, 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 680, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 680, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 680, + 287, + 712 + ], + "type": "text", + "content": "[39] Christoph Schuhmann, Richard Vencu, Romain Beaumont, Robert Kaczmarczyk, Clayton Mullis, Aarush Katta, Theo Coombes, Jenia Jitsev, and Aran Komatsuzaki. Laion-400m:" + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 326, + 73, + 544, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 544, + 95 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 544, + 95 + ], + "type": "text", + "content": "Open dataset of clip-filtered 400 million image-text pairs. arXiv preprint arXiv:2111.02114, 2021. 2, 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 307, + 95, + 545, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 95, + 545, + 139 + ], + "spans": [ + { + "bbox": [ + 307, + 95, + 545, + 139 + ], + "type": "text", + "content": "[40] Shuwei Shao, Zhongcai Pei, Weihai Chen, Ran Li, Zhong Liu, and Zhengguo Li. Urcdc-depth: Uncertainty rectified cross-distillation with cutflip for monocular depth estimation. arXiv preprint arXiv:2302.08149, 2023. 6" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 307, + 140, + 545, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 140, + 545, + 183 + ], + "spans": [ + { + "bbox": [ + 307, + 140, + 545, + 183 + ], + "type": "text", + "content": "[41] Shuwei Shao, Zhongcai Pei, Xingming Wu, Zhong Liu, Weihai Chen, and Zhengguo Li. Iebins: Iterative elastic bins for monocular depth estimation. arXiv preprint arXiv:2309.14137, 2023. 6" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 184, + 545, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 184, + 545, + 228 + ], + "spans": [ + { + "bbox": [ + 307, + 184, + 545, + 228 + ], + "type": "text", + "content": "[42] Shuran Song, Samuel P. Lichtenberg, and Jianxiong Xiao. Sun rgb-d: A rgb-d scene understanding benchmark suite. In 2015 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 567-576, 2015. 1, 7" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 228, + 545, + 260 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 228, + 545, + 260 + ], + "spans": [ + { + "bbox": [ + 307, + 228, + 545, + 260 + ], + "type": "text", + "content": "[43] Yang Song and Stefano Ermon. Generative modeling by estimating gradients of the data distribution. Advances in neural information processing systems, 32, 2019. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 261, + 545, + 294 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 261, + 545, + 294 + ], + "spans": [ + { + "bbox": [ + 307, + 261, + 545, + 294 + ], + "type": "text", + "content": "[44] Yang Song and Stefano Ermon. Improved techniques for training score-based generative models. Advances in neural information processing systems, 33:12438-12448, 2020. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 294, + 545, + 338 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 294, + 545, + 338 + ], + "spans": [ + { + "bbox": [ + 307, + 294, + 545, + 338 + ], + "type": "text", + "content": "[45] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. arXiv preprint arXiv:2011.13456, 2020. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 338, + 545, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 338, + 545, + 381 + ], + "spans": [ + { + "bbox": [ + 307, + 338, + 545, + 381 + ], + "type": "text", + "content": "[46] Yang Song, Conor Durkan, Iain Murray, and Stefano Ermon. Maximum likelihood training of score-based diffusion models. Advances in Neural Information Processing Systems, 34: 1415-1428, 2021. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 383, + 545, + 425 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 383, + 545, + 425 + ], + "spans": [ + { + "bbox": [ + 307, + 383, + 545, + 425 + ], + "type": "text", + "content": "[47] M. Sun, A. Y. Ng, and A. Saxena. Make3d: Learning 3d scene structure from a single still image. IEEE Transactions on Pattern Analysis and Machine Intelligence, 31(05):824-840, 2009. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 426, + 545, + 492 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 426, + 545, + 492 + ], + "spans": [ + { + "bbox": [ + 307, + 426, + 545, + 492 + ], + "type": "text", + "content": "[48] Pei Sun, Henrik Kretzschmar, Xerxes Dotiwalla, Aurelien Chouard, Vijaysai Patnaik, Paul Tsui, James Guo, Yin Zhou, Yuning Chai, Benjamin Caine, et al. Scalability in perception for autonomous driving: Waymo open dataset. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2446-2454, 2020. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 493, + 545, + 548 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 493, + 545, + 548 + ], + "spans": [ + { + "bbox": [ + 307, + 493, + 545, + 548 + ], + "type": "text", + "content": "[49] Igor Vasiljevic, Nick Kolkin, Shanyi Zhang, Ruotian Luo, Haochen Wang, Falcon Z Dai, Andrea F Daniele, Mohammadreza Mostajabi, Steven Basart, Matthew R Walter, et al. Diode: A dense indoor and outdoor depth dataset. arXiv preprint arXiv:1908.00463, 2019. 1, 7" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 548, + 545, + 592 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 548, + 545, + 592 + ], + "spans": [ + { + "bbox": [ + 307, + 548, + 545, + 592 + ], + "type": "text", + "content": "[50] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 593, + 545, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 593, + 545, + 646 + ], + "spans": [ + { + "bbox": [ + 307, + 593, + 545, + 646 + ], + "type": "text", + "content": "[51] Zhenda Xie, Zigang Geng, Jingcheng Hu, Zheng Zhang, Han Hu, and Yue Cao. Revealing the dark secrets of masked image modeling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14475-14485, 2023. 2, 5, 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 647, + 545, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 647, + 545, + 690 + ], + "spans": [ + { + "bbox": [ + 307, + 647, + 545, + 690 + ], + "type": "text", + "content": "[52] Xiaodong Yang, Zhuang Ma, Zhiyu Ji, and Zhe Ren. Gedgepth: Ground embedding for monocular depth estimation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12719-12727, 2023. 6" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 691, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 691, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 307, + 691, + 545, + 712 + ], + "type": "text", + "content": "[53] Weihao Yuan, Xiaodong Gu, Zuozhuo Dai, Siyu Zhu, and Ping Tan. Newcrfs: Neural window fully-connected crfs for" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "28294" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 148 + ], + "type": "list", + "angle": 0, + "index": 2, + "blocks": [ + { + "bbox": [ + 67, + 72, + 287, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 72, + 287, + 106 + ], + "spans": [ + { + "bbox": [ + 67, + 72, + 287, + 106 + ], + "type": "text", + "content": "monocular depth estimation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2022. 5, 6, 7" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 107, + 287, + 148 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 107, + 287, + 148 + ], + "spans": [ + { + "bbox": [ + 48, + 107, + 287, + 148 + ], + "type": "text", + "content": "[54] Wenliang Zhao, Yongming Rao, Zuyan Liu, Benlin Liu, Jie Zhou, and Jiwen Lu. Unleashing text-to-image diffusion models for visual perception. ICCV, 2023. 2, 3, 5, 6, 7, 8" + } + ] + } + ], + "index": 1 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "28295" + } + ] + } + ], + "index": 3 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/EFHQ_ Multi-purpose ExtremePose-Face-HQ dataset/65a10ed9-cef6-42e8-a141-8586b125aea4_content_list.json b/2024/EFHQ_ Multi-purpose ExtremePose-Face-HQ dataset/65a10ed9-cef6-42e8-a141-8586b125aea4_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..7d771f309be1fa7e0325e5fcbdeaf80026f82cc2 --- /dev/null +++ b/2024/EFHQ_ Multi-purpose ExtremePose-Face-HQ dataset/65a10ed9-cef6-42e8-a141-8586b125aea4_content_list.json @@ -0,0 +1,1586 @@ +[ + { + "type": "text", + "text": "EFHQ: Multi-purpose ExtremePose-Face-HQ dataset", + "text_level": 1, + "bbox": [ + 212, + 130, + 756, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Trung Tuan Dao $^{1,*}$ Duc Hong Vu $^{1,2,*}$ Cuong Pham $^{1,3}$ Anh Tran $^{1}$", + "bbox": [ + 200, + 189, + 766, + 208 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ VinAI Research, Vietnam $^{2}$ Miami University $^{3}$ Posts & Telecommunications Inst. of Tech., Vietnam", + "bbox": [ + 174, + 210, + 795, + 224 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{v.trungdt21,v.ducvh5,v.cuongpv11,v.anhtt152}@vinai.io vudh3@miamioh.edu cuongpv@ptit.edu.vn", + "bbox": [ + 140, + 229, + 828, + 243 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "*Equal contribution", + "bbox": [ + 423, + 244, + 542, + 260 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/cbf16de748627893751e10fb88a1c1970cd6edd74202d56555bea8261d2edfef.jpg", + "image_caption": [ + "Figure 1. Benefits of our proposed dataset (EFHQ). Standard large-scale facial datasets have most images at near frontal views, causing inferior performance of trained models on downstream tasks when dealing with extreme head poses. For instance, the trained 2D image generators and text-to-image ones often produce only near frontal faces, while the 3D face generators and face reenactment methods often show distorted outputs at profile views. The recently proposed dataset LPFF [47] partially handles that issue by providing complementary images at extreme head poses for only 2D and 3D image generation tasks. Our proposed dataset EFHQ provides high-quality extreme-pose images to complement a wide range of face-related tasks. It supports 2D and 3D image generation, with generally better diversity than LPFF. EFHQ also helps correct the outputs of text-to-image generation and face reenactment at extreme views. Finally, EFHQ provides a more challenging pose-based face verification benchmark to better assess the quality of face recognition networks." + ], + "image_footnote": [], + "bbox": [ + 143, + 292, + 831, + 465 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 597, + 313, + 612 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The existing facial datasets, while having plentiful images at near frontal views, lack images with extreme head poses, leading to the downgraded performance of deep learning models when dealing with profile or pitched faces. This work aims to address this gap by introducing a novel dataset named Extreme Pose Face High-Quality Dataset (EFHQ), which includes a maximum of 450k high-quality images of faces at extreme poses. To produce such a massive dataset, we utilize a novel and meticulous dataset processing pipeline to curate two publicly available datasets, VFHQ and CelebV-HQ, which contain many high-resolution face videos captured in various settings. Our dataset can complement existing datasets on various facial-related tasks, such as facial synthesis with 2D/3D-aware GAN, diffusion-based text-to-image face generation, and face reenactment. Specifically, training with EFHQ helps models generalize well across diverse poses, significantly improving performance in scenarios involving ex", + "bbox": [ + 73, + 628, + 473, + 902 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "treme views, confirmed by extensive experiments. Additionally, we utilize EFHQ to define a challenging cross-view face verification benchmark, in which the performance of SOTA face recognition models drops $5 - 37\\%$ compared to frontal-to-frontal scenarios, aiming to stimulate studies on face recognition under severe pose conditions in the wild.", + "bbox": [ + 496, + 599, + 895, + 689 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 500, + 722, + 632, + 738 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The human face is a central focus in computer vision, with extensive research dedicated to tasks like detection, recognition, generation, and manipulation [39, 42]. Numerous large-scale facial datasets[9, 12, 14, 24, 28, 30, 45, 48, 54, 55] have facilitated the face-related studies. Along with deep learning techniques, many tasks have made a considerable leap in performance in recent years. For instance, face recognition systems can recognize near-frontal faces in the wild with near-perfect accuracy [13, 18]. Facial generative models, such as [5, 15, 49], can synthesize realistic", + "bbox": [ + 496, + 750, + 893, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "22605", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "images indistinguishable from real faces. Face reenactment systems, including [41, 43, 51], can animate an input face based on poses and expressions from any driver video.", + "bbox": [ + 75, + 90, + 468, + 136 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Despite impressive results, the mentioned models produce inferior results when dealing with faces at extreme head poses. As illustrated in Fig. 1, faces synthesized by 3D face generation or face reenactment models are often distorted at profile views, due to the scarcity of extreme pose images in training datasets. Current extensive datasets for face generation leverage portrait images from Google or stock photo repositories like Flickr, which feature images with upfront views and minimal facial rotations. Face reenactment methods leverage video datasets such as Vox-Celeb [29], primarily featuring frames with frontal faces. The instability of pre-processing tools, particularly face alignment modules when handling extreme pose images, further reduces the proportion of those images in the existing datasets. This issue is critical since faces with large head rotations are common in real life, and dealing with them is unavoidable in practical systems.", + "bbox": [ + 75, + 138, + 470, + 396 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "This paper aims to fill that void by proposing a novel large-scale dataset of extreme-pose and high-quality facial images called ExtremePose-Face-HQ, or EFHQ for short. Our ambition is to build an \"one-for-all\" dataset that can supplement existing datasets on all mentioned face-related tasks. To achieve that goal, the dataset needs all the following properties: (1) extreme-pose, (2) large-scale, (3) high-quality, (4) in-the-wild, (5) having multiple images for each subject and having subject identity annotated. Given those requirements, we utilize a novel and meticulous dataset processing pipeline to filter and sample frames from two publicly available datasets, including VFHQ [48] and CelebV-HQ [54], which contain high-resolution face videos captured in various settings. This results in a dataset with 450k high-quality images of extreme poses, which are combined with frontal images of the same subject to use as image pairs. These image pairs can support face reenactment and face recognition tasks, while the extreme-pose images alone can complement the existing datasets of face generation training. Note that there is a recent attempt [47] to build a large-pose facial dataset similar to ours. However, that work relies on images crawled from Flickr, thus having limited size (19k images) and missing identity information. Hence, its applications are restricted, covering only several face-generation tasks, unlike our large-scale and multi-purpose dataset.", + "bbox": [ + 75, + 398, + 470, + 790 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We perform a comprehensive set of experiments to validate the effectiveness of EFHQ across three distinct face generation sub-tasks and face reenactment scenarios. Our dataset, EFHQ, when combined with standard training sets, significantly enhances the quality of synthesized images in highly profile or pitched views while maintaining synthesis quality in frontal views. EFHQ also provides a more", + "bbox": [ + 75, + 795, + 468, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "challenging pose-centric face verification benchmark. It reveals the vulnerability of many state-of-the-art face recognition networks when dealing with extreme-pose images, with TAR@FAR=1e-3 scores dropping 5-37%.", + "bbox": [ + 496, + 90, + 890, + 151 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In summary, our contributions include (1) a large-scale, high-quality, in-the-wild, extreme-posed dataset (EFHQ) designed to complement various face-related tasks, (2) a novel and meticulous dataset processing pipeline ensuring the creation and quality of the EFHQ dataset, and (3) a series of experiments, with plan-to-release models, demonstrating the usefulness of EFHQ in enhancing performance on a wide range of tasks when dealing with faces at extreme rotations. The final dataset will be publicly available as metadata files to not violate the two original dataset's licenses, alongside the pretrained networks.", + "bbox": [ + 496, + 152, + 892, + 316 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 332, + 640, + 347 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Synthetic Face Generation. Existing 2D-based models like StyleGAN [14, 15, 17] and 3D-aware models like EG3D [5] demonstrate impressive fidelity, especially for near-frontal views. Nevertheless, their performance significantly degrades when faced with extreme views, primarily due to insufficient training data for such scenarios. Datasets like FFHQ [14] exhibit a notable bias toward near-frontal poses, heavily limiting the range of poses synthesized by current 2D and 3D models despite the 3D-aware generators' ability to handle arbitrary poses.", + "bbox": [ + 496, + 364, + 890, + 515 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Diffusion-based text-to-image models [32, 33] can synthesize high-quality images from simple text inputs. However, these models lack mechanisms to control the generated pose. Consequently, the generated facial images remain in a frontal view despite terms like \"profile\" or \"looking-up\" in the input prompts. ControlNet[49] proposes a finetuning approach allowing these models to take additional context as conditional input. Namely, one can use landmarks to guide the generated face to the desired pose. However, the scarcity of profile-view data in training affects the model's performance, particularly in extreme angles. Consequently, effectively managing the long-tailed distribution of these extreme views becomes necessary to address this issue.", + "bbox": [ + 496, + 516, + 892, + 712 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Face Reenactment. Face reenactment involves transferring facial expressions and movements from a driving image to a source image. Recent advancements claim high-quality results, especially for near-frontal poses [31, 41, 43, 51]. However, pose collapse persists, particularly in keypoint-based models [41, 51]. Common datasets for this task are VoxCeleb1 [29] and HDTF [50], which primarily feature frontal talking face videos. Despite recent potentially high-quality datasets for reenactment such as VFHQ [48] and CelebV-HQ [54], the frontal-view bias persists. As a result, facial reenactment methods often produce artifacts and lower-quality outputs when faced with significant rotations.", + "bbox": [ + 496, + 719, + 892, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "22606", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 1 + }, + { + "type": "table", + "img_path": "images/beb48d3203d2a90d8ab0890319d86d2fe69993915e96900069cc7e6ecfa0a320.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
DatasetAttributesReady-to-use tasks
#imagesMultiple images per IDHigh QualityProfile ViewFace VerificationFace SynthesisFace Reenactment
AgeDB12K✓✓
CPLFW12K✓✓✓
CFP-FP14K✓✓✓
IJB-x138K✓✓✓
CelebA-HQ30K✓✓
FFHQ70K✓✓
LPFF19K✓✓✓
CelebV-HQ6M✓✓✓✓
VFHQ3M✓✓✓✓✓
VoxCeleb14M✓✓✓
Ours450,538✓✓✓✓✓✓
FFHQ+Ours*112,671✓✓✓
Vox+Ours*4.5M✓✓✓
", + "bbox": [ + 158, + 88, + 803, + 296 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Table 1. Representative comparison of key attributes and supported tasks across existing face datasets and our proposed EFHQ dataset. EFHQ provides large-scale, high-quality images with multiple views per identity, enabling diverse face-related tasks, including synthesis, reenactment, and verification. Combining EFHQ with existing datasets like FFHQ or VoxCeleb can augment available data for selected tasks. Asterisk (*) indicates a combined dataset with a subsampled version of EFHQ.", + "bbox": [ + 75, + 301, + 890, + 357 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Face Verification. Despite near-perfect accuracy on benchmark datasets[12, 20, 27, 28, 44], state-of-the-art face verification models[7, 13, 18] fall short in unconstrained environments due to a lack of pose diversity. While CPLFW[52] introduces 6000 cross-pose pairs, its small size and mix of pose samples hinder analyzing performance on specific pose scenarios such as frontal-frontal, frontal-profile, and profile-profile. More granular benchmarking on diverse pose pairs is crucial to identify failure modes and drive progress in robust face verification across poses.", + "bbox": [ + 75, + 369, + 468, + 521 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Extreme Pose Face Dataset. Tab. 1 summarizes existing datasets related to face verification, synthesis, and re-enactment, showing that existing face datasets have limitations for multi-task use. Recent efforts like LPFF[47] made progress by crawling from Flickr and processing for diverse views, resulting in a dataset of 19,590 images. The dataset could complement high-quality datasets like FFHQ[14] for generative models. However, due to its limited scale comprising only $20\\%$ of the combined dataset, the unbalanced distribution between frontal and profile view still needs further improvements. Additionally, since the images are from image hosting services, the dataset fails to provide the identity information necessary for other facial applications.", + "bbox": [ + 75, + 527, + 468, + 724 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Dataset Process", + "text_level": 1, + "bbox": [ + 76, + 739, + 232, + 756 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we discuss the process of producing our novel dataset EFHQ. To fulfill our goals, we carefully curated frames from the two recent facial video datasets, VFHQ[48] and CelebV-HQ[54], as seen in Fig. 2.", + "bbox": [ + 75, + 766, + 468, + 828 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Attributes Preparation", + "text_level": 1, + "bbox": [ + 76, + 838, + 287, + 854 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Attributes Extraction. To curate accurate facial poses and high-quality frames from existing datasets, we first ex", + "bbox": [ + 75, + 869, + 468, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "tract (1) face bounding boxes, (2) facial landmarks (5/68 keypoints), (3) image quality score, and (4) face identity. For bounding boxes and landmarks, we employ the popular RetinaFace[36] and SynergyNet[46]. We use HyperIQA [40] to grade image quality. For VFHQ, we only generate 68 landmarks and image quality scores since bounding boxes and 5 keypoints are provided. After extracting these attributes, we match them with existing annotations using an IoU-based Hungarian matching algorithm [21]. Regarding identity, VFHQ-wise, we source labels from the annotations. As CelebV-HQ lacks individual identity labels within the video, we use[35] to identify the identities associated with each bounding box.", + "bbox": [ + 496, + 369, + 890, + 566 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Due to the instability of head pose estimators when handling extreme-pose images, we strategically apply multiple pose estimators relying on various methodologies: (1) SynergyNet[46], a 3DMM-based model, (2) DirectMHP[53], a landmark-free joint head detector and pose estimator, and (3) an in-house FacePoseNet[6] improved through extensive data augmentation focused on extreme poses and extra training data. We then categorize each estimated pose into a hierarchical binning scheme illustrated in Fig. 2. Next, we gather bin predictions and perform majority voting to arrive at consensus labels. When no agreement emerges, we put it into the $5^{th}$ bin for \"confusing\" cases. As depicted in Fig. 3, this ensemble approach provides robustness in cases where a single estimator incorrectly categorizes a sample. Please refer to the supplementary for statistics of the extracted bins and the hyperparameters for each model used for attribute extraction.", + "bbox": [ + 496, + 570, + 890, + 828 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Annotation Review. Next, we manually review the results to ensure high-quality labels. Given the large dataset, we developed a graphical user interface tool to streamline the review process. Our focus is verifying the binning annota", + "bbox": [ + 496, + 839, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "22607", + "bbox": [ + 478, + 944, + 517, + 957 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/f254cce2dcbe27b3334c49712a16370a9280f177f1bfe4f41e39c028a6f522d5.jpg", + "image_caption": [ + "Figure 2. The pipeline of EFHQ dataset creation. Starting with high-quality videos from the VFHQ[48] and CelebV-HQ[54] datasets, single-frame attributes are extracted then manually reviews. Task-specific preprocessing is then applied to generate specialized versions of the dataset for tasks such as face generation, reenactment, and verification." + ], + "image_footnote": [], + "bbox": [ + 158, + 79, + 807, + 256 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/38c77dbb4c806449587b476c0b22ce3f7a86ea80b9d0b108ab1e63be1859d8d5.jpg", + "image_caption": [ + "Figure 3. Example cases where a pose estimator fails to categorize the sample to the correct bin." + ], + "image_footnote": [], + "bbox": [ + 96, + 321, + 447, + 448 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "tions, which most directly impact label accuracy. Concurrently, we randomly subsample the data based on pose angle and image quality to validate facial landmark quality further and discard cases of landmark prediction failure.", + "bbox": [ + 75, + 496, + 468, + 556 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. EFHQ dataset", + "text_level": 1, + "bbox": [ + 76, + 571, + 225, + 587 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our final EFHQ dataset comprises up to 450k frames with extreme poses extracted from approximately 5,000 clips. Most of the clips in our dataset include at least one frame with a frontal face and multiple frames with extreme pose angles. We also include extreme pose-only clips, representing profile-to-profile pose transfer cases.", + "bbox": [ + 75, + 595, + 468, + 688 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/caec04bb3476c53e27cfcf754fc7da82c83f7eb861fc3f48a721d325f51ab89b.jpg", + "image_caption": [ + "Figure 4. Pose distribution comparison between our sampled dataset and other datasets, including FFHQ[14], LPFF[47], CPLFW[52] and 40K random samples from VoxCeleb1 [29]. Our sampled dataset demonstrates greater pose diversity, with increased sample counts across high angle bins." + ], + "image_footnote": [], + "bbox": [ + 86, + 705, + 455, + 825 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Dataset for subtasks", + "text_level": 1, + "bbox": [ + 500, + 321, + 691, + 335 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Supplementary datasets for face generation. To address FFHQ's pose distribution gap, we compile a dataset encompassing diverse poses, varied identities, and FFHQ-comparable image quality from EFHQ. Additionally, we also embed an image brightness filter[3] in the sampling process, given that HyperIQA doesn't explicitly assess this aspect. The sampled images are then processed using the same pipeline in [5, 16] and manually reviewed. The aim is to improve coverage of under-represented poses without degrading frontal performance of trained generation models. This yields a dataset of 42,671 images with equitably distributed poses as visualized in Fig. 4. Compared to the original FFHQ distribution, yaw and pitch distributions are significantly improved.", + "bbox": [ + 496, + 358, + 890, + 570 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Additionally, for 3D-aware GAN subtasks, we extract camera parameters to serve as conditional pose information, following [5]. We convert these parameters into yaw, pitch angles to validate against our existing annotations. When disagreements emerge, we re-examine the images and filter if needed, to ensure accurate, quality labels.", + "bbox": [ + 496, + 575, + 890, + 667 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For diffusion-based text-to-image generation, a specialized dataset is curated to refine Stable Diffusion models[33] by accommodating landmark conditional input through ControlNet[49]. Based on our 3D-aware face generation dataset, we integrate landmark-based conditional images and tailored text prompts for each image. To capture fine facial nuances, we extract 478 facial landmarks and draw the condition image with connected edges between matched keypoints, following Mediapipe framework[26], better-conveying face expression and pose. For text prompts, detailed facial attributes like gender, race, and emotion are derived from existing labels or inferred via BLIP-2 pretrained captioning model [23] and a face attribute estimator [37]. The resulting prompt adheres to a structured format: \"A profile portrait image of a [emotion] [race] [gender].\"", + "bbox": [ + 496, + 674, + 892, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "22608", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/867f1367cf0d97dea13112a5cf0db97b2e1330fb65f9182a4c2e564667591df6.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelReference DatasetFID↓Recall↑
GFFHQFFHQ2.840.49
GFFHQ+LPFFFFHQ+LPFF3.430.44
GFFHQ+EFHQFFHQ+EFHQ3.440.46
GFFHQ+LPFFFFHQ+EFHQ8.260.43
GFFHQ+EFHQFFHQ+LPFF7.040.44
GFFHQ+EFHQFFHQ+EFHQ3.330.46
GFFHQ+EFHQFFHQ3.120.44
GFFHQ+EFHQFFHQ+LPFF4.470.42
", + "bbox": [ + 78, + 88, + 470, + 258 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Table 2. StyleGAN2-ADA models trained on different datasets. Lower FIDs indicate better fidelity, while higher Recalls indicate better diversity. The second and third blocks show comparisons between unconditional models trained on FFHQ+LPFF and FFHQ+EFHQ when evaluating on the same and cross-dataset settings, with the better metrics in bold.", + "bbox": [ + 75, + 263, + 468, + 348 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Supplementary dataset for face reenactment. As EFHQ is crafted, we can employ the entire dataset to complement established training datasets, e.g., VoxCeleb1 [29]. We also built an extra evaluation set with 1200 EFHQ clips.", + "bbox": [ + 75, + 354, + 468, + 417 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Benchmarking dataset for face verification. We curate a dataset covering three distinct scenarios: frontal-to-frontal, frontal-to-profile, and profile-to-profile. To enable rigorous benchmarking, we sample 10,000 pairs each for both negative and positive cases per scenario, resulting in a balanced benchmark dataset containing 60,000 image pairs. Images follow the same established preprocessing pipelines in [7]. Next, we filter misaligned images with RetinaFace[36], randomly review, and replace equivalent samples from our diverse corpus, if needed. Finally, to simulate varying image quality, we randomly applied downscaling and compression to samples in the dataset.", + "bbox": [ + 75, + 424, + 468, + 604 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Face Generation Subtask", + "text_level": 1, + "bbox": [ + 76, + 623, + 313, + 638 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. StyleGAN", + "text_level": 1, + "bbox": [ + 76, + 648, + 191, + 666 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Training details. We train StyleGAN2-ADA[15] models from scratch after combining FFHQ with our own curated dataset (denoted FFHQ+EFHQ) at $1024 \\times 1024$ resolution. We also train a conditional version, allowing direct control over frontal versus profile synthesis and enhancing model performance. Specifically, we categorize the dataset into two classes: \"frontal\" and \"profile\" based on pose angles. All models are trained for 6 days on 8 Nvidia A100 GPUs.", + "bbox": [ + 75, + 680, + 467, + 800 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Evaluation protocol. We utilize the Frechet Inception Distance[11] (FID) metric and improved Recall[22] to quantify the generation fidelity and diversity. Regarding the two unconditional models trained with the supplemented FFHQ+LPFF and FFHQ+EFHQ, we evaluate each model with both of these datasets. For conditional mod", + "bbox": [ + 75, + 809, + 467, + 898 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/3dcea56d5362a57eb679bd30469ce22064230f4b8d123e344efe25aaf2dc37ea.jpg", + "image_caption": [ + "(a) Samples from the model trained with FFHQ+EFHQ." + ], + "image_footnote": [], + "bbox": [ + 532, + 88, + 859, + 174 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/3c0d5c9941056f9827add44aae19a2e757bb2b54eaf548fa151058eafdd04f75.jpg", + "image_caption": [ + "(b) Profile-view comparison of LPFF (left) and EFHQ (right).", + "Figure 5. Qualitative result and comparison of generated samples, with truncation $\\psi = 0.7$ , from StyleGAN2-ADA training with other dataset and ours." + ], + "image_footnote": [], + "bbox": [ + 531, + 190, + 857, + 349 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "els, we evaluate performance on FFHQ, FFHQ+LPFF, and FFHQ+EFHQ to demonstrate the broader coverage of the model trained with our proposed dataset. For FFHQ, we generate 50K images with the \"frontal\" label, whereas for the other two datasets, we generate 50K following their label distributions. This comprehensive quantitative analysis examines the fidelity and variety of the generated images.", + "bbox": [ + 496, + 431, + 890, + 537 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Experimental results. We compared the performance of our proposed unconditional models $(G^{\\mathrm{FFHQ + EFHQ}})$ , against models trained on the FFHQ and FFHQ+LPFF datasets $(G^{\\mathrm{FFHQ}}$ and $G^{\\mathrm{FFHQ + LPFF}})$ . Models trained conditionally are specifically subscripted, e.g., $G_{c}^{\\mathrm{FFHQ + EFHQ}}$ . The qualitative results in Fig. 5 illustrate that our models generate high-quality frontal faces comparable to the FFHQ model while producing realistic and varied profile faces. Additionally, our method has less noise in extreme poses than the FFHQ+LPFF model. Specifically, Fig. 5b shows LPFF's generated face contains more noise and less photorealistic details than our model, such as the pixelated eye and nose patches. Meanwhile, ours achieves greater realism with smoother skin, well-defined features, noise reduction, and natural lighting. The quantitative results in Tab. 2 show that our unconditional model achieves competitive FID scores and improved Recall to other models, indicating similar fidelity but greater sample pose diversity. In the cross-examination (third block of Tab. 2), $G^{\\mathrm{FFHQ + EFHQ}}$ has both better FID and improved Recall on the opposite dataset than $G^{\\mathrm{FFHQ + LPFF}}$ . Finally, our conditional model demonstrates substantially improved FID scores over multiple reference datasets, highlighting its versatility.", + "bbox": [ + 496, + 553, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "22609", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/44f3155bf285746e0ba7966b8c0a6c4ed50d0cc4776b7446c00c1152e1796d93.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelReference DatasetFID↓Pose↓Depth↓ID↑
EFHQFFHQ/Trainset4.85e-40.2340.74
EFFHQFFHQ/Trainset3.75e-40.2880.74
EFFHQ+LPFFFFHQ6.26e-40.3160.77
EFFHQ+EFHQFFHQ3.94e-40.2240.77
EFFHQ+LPFFTrainset7.06e-40.3160.65
EFFHQ+EFHQTrainset4.64e-40.2230.65
", + "bbox": [ + 80, + 88, + 467, + 205 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/8e70a13b656bbde69089723c284a0af97e6cca6c1c51f5d55a61b3fe1f63ec4d.jpg", + "image_caption": [ + "Figure 6. Comparison between multiview generated samples, with truncation $\\psi = 0.8$ , of EG3D model trained with various datasets." + ], + "image_footnote": [], + "bbox": [ + 88, + 294, + 455, + 457 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. EG3D", + "text_level": 1, + "bbox": [ + 76, + 507, + 161, + 522 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Training details. We followed the same pipeline and hyper-parameters as the original EG3D model [5]. However, based on empirical results, we adjusted $\\gamma = 5$ . We also fixed the triplane implementation bug (XY, XZ, ZX) [1] to replicate the original idea in the papers (XY, XZ, ZY). We trained the model on 8 Nvidia A100 40GB GPUs for approximately 7 days until convergence.", + "bbox": [ + 75, + 537, + 467, + 643 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation Protocol. To evaluate the models, mirroring [5], we measure image quality with FID; multi-view consistency and identity preservation with mean cosine similarity of pre-trained face recognition model[7]; pose accuracy and geometry by calculating MSE against pseudo labels from [8]. We evaluate all models on 1,024 generated images, except for FID, which we compute on 50k images.", + "bbox": [ + 75, + 651, + 467, + 758 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Experimental results. To evaluate the benefits of our dataset, we trained the same backbone architecture on different datasets: FFHQ $(E^{\\mathrm{FFHQ}})$ , FFHQ+LPFF $(E^{\\mathrm{FFHQ} + \\mathrm{LPFF}})$ , and FFHQ+EFHQ (our model, $E^{\\mathrm{FFHQ} + \\mathrm{EFHQ}}$ ). The quantitative results are presented in Table 3, and qualitative examples are shown in Figure 6. Qualitatively, we observe that $E^{\\mathrm{FFHQ} + \\mathrm{EFHQ}}$ generates improved face shapes in extreme poses compared to $E^{\\mathrm{FFHQ}}$ , similar to the improvements seen with $E^{\\mathrm{FFHQ} + \\mathrm{LPFF}}$ . Quantitatively, the FID", + "bbox": [ + 75, + 763, + 467, + 901 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/7cff13cf3cc12be33b6c195c7d3f103525cc0601828cf4e62e89e6bc85fa9fff.jpg", + "table_caption": [ + "Table 3. EG3D models trained on different datasets. In the second and third blocks, we compare models trained on FFHQ+LPFF and FFHQ+EFHQ with the testset is either FFHQ or their training datasets; the better metrics are in bold. The best metrics on the FFHQ testset are in red, while the best metrics when evaluating on training dataset are underlined." + ], + "table_footnote": [], + "table_body": "
MetricsDatasetPretrained ControlNetOurs
NME↓LPFF0.0850.084
FID↓LPFF119.7483.39
FIDeyepatch ↓LPFF103.7749.45
NME↓EFHQ0.1060.081
FID↓EFHQ111.2381.90
FIDeyepatch ↓EFHQ137.4358.60
", + "bbox": [ + 500, + 88, + 888, + 217 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 4. Quantitative results of fine-tuning StableDiffusion v1.5 with ControlNet trained on various datasets, using Normalized Mean Error (NME), FID, and FID on eye region.", + "bbox": [ + 498, + 220, + 890, + 262 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "of $E^{\\mathrm{FFHQ + EFHQ}}$ is comparable to $E^{\\mathrm{FFHQ}}$ when evaluated on FFHQ, indicating no degradation in frontal view performance. Meanwhile, the competitive FID achieved by $E^{\\mathrm{FFHQ + EFHQ}}$ on the combined FFHQ+EFHQ dataset also suggests proficient modeling of profile views. Regarding identity consistency, we see a significant drop when the reference dataset contains profile-view images, due to the weak confidence of the pretrained face recognition model when handling profile-view images, as confirmed by its low average cosine similarity of around 0.6 on the simple frontal-to-profile verification set CFP-FP[34]. Furthermore, our model also outperforms in pose accuracy and geometry quality with both frontal and cross-pose dataset.", + "bbox": [ + 498, + 270, + 890, + 469 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/91c1c39d34e1ff469ab19168be92e10d78b69610e53bee1af6b1b4b5daca6041.jpg", + "image_caption": [ + "(a) Generated samples; condition image is in between samples." + ], + "image_footnote": [], + "bbox": [ + 513, + 479, + 694, + 578 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/e2849aa6187df629ee39998db75c92f3603e153a68b749d5b56f72afd4fcceeb.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 696, + 479, + 880, + 577 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/ea5572592c76c179b7a23caa798b0464f0d77a81bf0014c7a81b2a1c2f7a141b.jpg", + "image_caption": [ + "(b) Generated eye regions." + ], + "image_footnote": [], + "bbox": [ + 501, + 592, + 697, + 625 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/42ea89334a93b892c5276bce82ddb8d2426f524d79dcfa74a222ced094510ff8.jpg", + "image_caption": [ + "Figure 7. Comparison profile-view generated samples of pretrained ControlNet (left) and our fine-tuned ControlNet (right) with the prompt: \"A profile portrait image of a person\"." + ], + "image_footnote": [], + "bbox": [ + 699, + 592, + 893, + 625 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. ControlNet", + "text_level": 1, + "bbox": [ + 500, + 696, + 622, + 710 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Training details. We finetuned Stable Diffusion v1.5 [33] using ControlNet [49] on the EFHQ dataset to enhance its ability to generate extreme facial poses. We use AdamW optimizer [25] with learning rate $1 \\times e^{-5}$ , $\\beta_0 = 0.9$ , $\\beta_1 = 0.999$ , and $\\epsilon = 0.01$ . We train with batch size 24 on 1 Nvidia A100 40GB GPU in 2 days.", + "bbox": [ + 498, + 726, + 890, + 816 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation Protocol. To assess image qualities across diverse poses, we measure the Normalized Mean Error (NME) of 478-point facial landmarks between the generated and source image. Specifically, we evaluate performance across the full pose distribution in the reference", + "bbox": [ + 496, + 824, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "22610", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/0100f099836e506a48e568b5e3766bbb91f0c71b7470a5ec112ea03488b75695.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelTest Dataset (L1 ↓/ AKD ↓/ AED ↓)
VoxCeleb1EFHQ (Ours)
TPS(VoxCeleb10.041/1.214/0.1320.061/3.814/0.275
TPS(VoxCeleb1+EFHQ0.041/1.250/0.1330.051/1.963/0.214
LIA(VoxCeleb10.044/1.342/0.1380.059/3.807/0.266
LIA(VoxCeleb1+EFHQ0.044/1.398/0.1420.054/2.291/0.233
", + "bbox": [ + 81, + 88, + 470, + 175 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "dataset. For the prompt, we use: \"A profile portrait image of a person.\" with extra positive and negative keywords, further detailed in supplementary. Utilizing LPFF and EFHQ datasets as the reference datasets, we compute FID to assess synthesized face quality and extend evaluation to cropped eye regions due to unrealistic gazes observed from the original ControlNet model. Note that the models receive condition images created akin to its training data.", + "bbox": [ + 75, + 241, + 468, + 363 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Experimental results. Fig. 7a visually compares the baseline ControlNet OpenPose face and our fine-tuned model. Our fine-tuned model adeptly conditions diverse extreme poses, generating high-quality results with precise details while maintaining fidelity to the condition, such as mouth shape, and exhibiting minimal artifacts. These patterns indicate the superiority of the fine-tuned model over the baseline model in handling extreme viewing angles. Additionally, Fig. 7b contrasts the eye regions of both models. Our fine-tuned model accurately captures natural eye gaze and shape from condition images, a notable improvement over the baseline. Quantitative results in Tab. 4 consistently favor our fine-tuned model across all metrics. FID scores on full and eye patches exhibit significant enhancement, and NME performance remains stable on training and reference datasets (EFHQ and LPFF). These findings showcase the fine-tuned model's capability to generate high-quality, geometrically consistent faces across diverse poses.", + "bbox": [ + 75, + 369, + 470, + 642 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5. Face Reenactment Subtask", + "text_level": 1, + "bbox": [ + 76, + 656, + 328, + 672 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.1. Evaluation Protocol", + "text_level": 1, + "bbox": [ + 76, + 680, + 266, + 695 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We follow the same preprocessing and evaluation protocol as FOMM [38]. Specifically, all the frames are cropped to $256 \\times 256$ , and we use the first frame of a clip as the source image to reconstruct the remaining frames. In the case of the EFHQ test set, most source images have a frontal pose, while the target frames have an extreme pose. Some test cases only contain extreme pose to evaluate self-reconstruction on these cases. We run the experiments using two recent methods, Thin-Plate Spline Motion Model (TPS) [51] and Latent Image Animator (LIA) [43]. For training details, please refer to the supplementary. We select the following three metrics to compare the generated frames and the driving ground truth. $\\mathcal{L}_1$ measures the av", + "bbox": [ + 75, + 703, + 470, + 902 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/6d045a3c5d3768b18d0d1ca71db35c934026da6a519eeaea4da966eb56060a00.jpg", + "table_caption": [ + "Table 5. Quantitative results of two state-of-the-art reenactment methods (TPS and LIA) when training on VoxCeleb1 and the combined set VoxCeleb1+EFHQ. We present results evaluated either on VoxCeleb1 (left) or EFHQ (right) test sets." + ], + "table_footnote": [], + "table_body": "
ModelTrainsetCPLFWEFHQ
f2ff2pp2p
R18arfaceGlint360k0.800.940.570.370.880.06
R18arfaceMS1MV30.750.930.570.360.880.05
R50arfaceGlint360k0.910.970.770.200.920.05
R100arfaceGlint360k0.930.980.830.150.820.16
R100adafaceWebFace4M0.940.990.940.050.970.02
", + "bbox": [ + 503, + 88, + 888, + 195 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 6. Quantitative evaluation of state-of-the-art face recognition models on the cleansed CPLFW and the proposed benchmark dataset. We use the metric TAR@FAR=1e-3 (higher is better). Our EFHQ benchmark includes frontal-frontal (f2f), frontal-profile (f2p), and profile-profile (p2p) subsets. For f2p and p2p, we report the score drop when comparing to f2f in the red subscription. The lowest benchmark score for each model is in bold.", + "bbox": [ + 496, + 199, + 890, + 297 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "erage distance among pixel values, indicating reconstruction faithfulness. Average Keypoint Distance (AKD) calculates the mean distance between keypoints, thus assessing the pose accuracy. Average Euclidean Distance (AED) measures identity preservation by computing the average $\\mathcal{L}_2$ distance between identity embeddings.", + "bbox": [ + 496, + 306, + 890, + 398 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.2. Experimental results", + "text_level": 1, + "bbox": [ + 498, + 412, + 696, + 428 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We compare TPS and LIA models trained with our supplementary EFHQ dataset $(TPS^{\\mathrm{VoxCeleb1 + EFHQ}}$ and $LIA^{\\mathrm{VoxCeleb1 + EFHQ}})$ against models trained solely on VoxCeleb1 $(TPS^{\\mathrm{VoxCeleb1}}$ and $LIA^{\\mathrm{VoxCeleb1}})$ . Tab. 5 displays quantitative results, revealing that models trained with EFHQ maintain comparable performance when evaluated on VoxCeleb1, emphasizing high-quality performance in frontal-to-frontal reenactment scenarios. Notably, on the EFHQ test set, both EFHQ-trained models show significant improvement in all metrics, indicating enhanced extreme pose and identity reenactment performance. In Fig. 8, qualitative results prove that EFHQ-trained models excel in transferring facial expressions and motion across diverse poses. Models trained on EFHQ showcase improved image quality, reduced artifacts, and well-preserved shapes of driving faces. These results highlight our models' effectiveness in capturing and reproducing facial motions across varying poses while enhancing fidelity to the original facial features.", + "bbox": [ + 496, + 436, + 892, + 724 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "6. Face Recognition Subtask", + "text_level": 1, + "bbox": [ + 498, + 742, + 741, + 760 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "6.1. Evaluation Protocol", + "text_level": 1, + "bbox": [ + 500, + 768, + 689, + 784 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To evaluate face recognition performance on extreme poses, we selected two state-of-the-art methods: ArcFace[7] and AdaFace[18]. We further explored various backbone sizes (ResNet18, ResNet50, ResNet101)[10] and training datasets (MS1MV2[9], Glint360K[2], WebFace4M[55]) of increasing size. For metrics, we used True Acceptance Rate (TAR) at False Acceptance Rate (FAR) of $1 \\times e^{-3}$ .", + "bbox": [ + 496, + 795, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "22611", + "bbox": [ + 478, + 944, + 517, + 957 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/2c8117e4d7e86d126609d5accefe90974ed636e97faaf9d4022551304474fbb7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 168, + 99, + 264, + 172 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/f69d710555708f6551cef3b5ce48de1fde2a208e53247211237ab7c6fbae5fcb.jpg", + "image_caption": [ + "Source" + ], + "image_footnote": [], + "bbox": [ + 168, + 174, + 267, + 247 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/7d2912606c5b61da5c606f88306fd6e060536c5ef9c314d5d5374ab99e85b6c3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 277, + 99, + 374, + 172 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/61c559132cc4890d912246b645d7129a8dff94b65997700f60ef27bdef7545eb.jpg", + "image_caption": [ + "Figure 8. Comparison of performance for frontal-profile face reenactment on EFHQ. From left to right: source image, TPS on VoxCeleb1, TPS with EFHQ, LIA on VoxCeleb1, LIA with EFHQ, driving image. Additional EFHQ training data improves synthesis of extreme poses." + ], + "image_footnote": [], + "bbox": [ + 277, + 174, + 375, + 250 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/21f7cb68aa95c4ae50e04e75b16ffcc4a864c7551ef38fd4b4973bd55631b879.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 379, + 99, + 473, + 172 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/74afe71aafafe1944aae621c88642f991e2738f3e52b03082731ebdf1d978f1f.jpg", + "image_caption": [ + "TPS (VoxCeleb1)", + "TPS (Ours)" + ], + "image_footnote": [], + "bbox": [ + 379, + 174, + 473, + 250 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/2eb2f40499fafbf4fcb78e429ecb6a3f128bea37332e4a23ac5950ae3a760e86.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 488, + 99, + 586, + 172 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/13c7f831840d7ea36cbc2037bba4722099980fd3cb1850d9ce250d64185c67f0.jpg", + "image_caption": [ + "LIA (VoxCeleb1)" + ], + "image_footnote": [], + "bbox": [ + 488, + 174, + 586, + 247 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/6d928c801997e6e85bd8d4dd1dab3f7ab8b88397fe154b822f5f89eac285434f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 591, + 99, + 687, + 172 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/a37e832c4e7938ac2972ac1b7a72c3542f56a1a39355e19829977df27ba063d8.jpg", + "image_caption": [ + "LIA (Ours)" + ], + "image_footnote": [], + "bbox": [ + 591, + 174, + 687, + 247 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/cf4b1d5c92581dc96cd7ada2ad8e39c5f6b911a5e50861f172bbd3c22fc7a3d0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 702, + 99, + 797, + 172 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/227ecfb2eb75f91aa3aa305d5cb140c2676776284e15ef4c1a063540403f72ea.jpg", + "image_caption": [ + "Driving" + ], + "image_footnote": [], + "bbox": [ + 700, + 174, + 795, + 247 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6.2. Experimental results", + "text_level": 1, + "bbox": [ + 76, + 309, + 264, + 325 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Before assessing face recognition networks on EFHQ, we first clean the commonly-used version of CPLFW, processed by[7], by removing misaligned images (please refer to the Supplementary), and evaluate the models on it. Models performed well on this cleansed set, with minimal gaps between ArcFace and AdaFace.", + "bbox": [ + 75, + 332, + 468, + 422 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Our proposed evaluation benchmark involves three pose-based scenarios. Results in Tab. 6 reveal performance drops of $5 - 37\\%$ on the frontal-profile subset compared to frontal-frontal, highlighting the challenge of matching faces in different poses. Larger ResNet models outperformed ResNet18 across subsets, showcasing the advantages of increased capacity in handling pose variation. However, the ResNet100 model trained on Glint360K performed worse than smaller models on the profile-profile subset, indicating the need for sufficient profile data to prevent overfitting. Our results also indicate AdaFace's suitability for handling pose variations. In summary, pose mismatch brings significant challenges for face recognition, necessitating larger models trained on diverse data covering profile faces.", + "bbox": [ + 75, + 422, + 468, + 635 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "7. Discussion", + "text_level": 1, + "bbox": [ + 76, + 646, + 191, + 660 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/e17c16d7e1823307bbc687e505298164fa1f4d5649b691e9a0108114e17bb041.jpg", + "image_caption": [ + "Figure 9. User study results. We reported the percentage of times testers rank whose output is best or tied for each criterion." + ], + "image_footnote": [], + "bbox": [ + 96, + 667, + 446, + 813 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "User survey. We conduct user surveys to evaluate the face generation improvement of our method through human per", + "bbox": [ + 75, + 869, + 468, + 901 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "ceptual assessment. In this survey, we compare the outputs of the models mentioned in Secs. 4 and 5. The setup of this user survey is included in the supplementary. In all human evaluation results (Fig. 9), our models outperform previous ones, notably on identity and motion test of face reenactment models (74.80% and 64.22%, respectively).", + "bbox": [ + 496, + 310, + 890, + 401 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/5406b417968dda5059ac39836df61475119ae9f04474919d0df72f427ee2ccb5.jpg", + "image_caption": [ + "Figure 10. Representative examples of the raw LPFF dataset that got excluded from the final dataset, either from misdetections or pose filtering process. Red keypoints represent landmarks prediction of selected samples from LPFF, green keypoints represent landmarks prediction from our pipeline." + ], + "image_footnote": [], + "bbox": [ + 527, + 419, + 859, + 506 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Robustness of data processing pipeline. As shown in Fig. 10, our data processing pipeline can detect additional profile-view faces filtered out by the LPFF pipeline, likely due to misdetections or incorrect pose filtering. This confirms the robustness of our pipeline, and we hope it will enable the community to develop larger and higher-quality in-the-wild facial datasets.", + "bbox": [ + 496, + 594, + 890, + 699 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "8. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 720, + 617, + 734 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This work has introduced a large-scale, diverse facial dataset to address performance gaps between frontal and profile faces, powered by a novel and robust data processing pipeline. Crucially, we provided tailored sub-datasets for advancing essential tasks like face synthesis and reenactment. Moreover, our new face verification benchmark reveals gaps between techniques, granting valuable insights. Ultimately, we hope this high-quality, diverse facial dataset opens up new and exciting opportunities to push forward cross-pose tasks.", + "bbox": [ + 496, + 747, + 890, + 887 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "22612", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Fix for triplane projection $\\cdot$ Issue #67 $\\cdot$ NVlabs/eg3d. https://github.com/NVlabs/eg3d/issues/67, 2023.6", + "[2] Xiang An, Xuhan Zhu, Yuan Gao, Yang Xiao, Yongle Zhao, Ziyong Feng, Lan Wu, Bin Qin, Ming Zhang, Debing Zhang, et al. Partial fc: Training 10 million identities on a single machine. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1445-1449, 2021. 7", + "[3] Sergey Bezryadin, Pavel Bourov, and Dmitry Ilinih. Brightness calculation in digital image processing. International Symposium on Technologies for Digital Photo Fulfillment, 2007:10-15, 2007. 4, 1", + "[4] Z. Cao, G. Hidalgo Martinez, T. Simon, S. Wei, and Y. A. Sheikh. Openpose: Realtime multi-person 2d pose estimation using part affinity fields. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2019. 4", + "[5] Eric R. Chan, Connor Z. Lin, Matthew A. Chan, Koki Nagano, Boxiao Pan, Shalini De Mello, Orazio Gallo, Leonidas Guibas, Jonathan Tremblay, Sameh Khamis, Tero Karras, and Gordon Wetzstein. Efficient geometry-aware 3D generative adversarial networks. In CVPR, 2022. 1, 2, 4, 6", + "[6] Feng-Ju Chang, Anh Tuan Tran, Tal Hassner, Iacopo Masi, Ram Nevatia, and Gerard Medioni. Faceposenet: Making a case for landmark-free face alignment. In Proceedings of the IEEE International Conference on Computer Vision Workshops, pages 1599-1608, 2017. 3", + "[7] Jiankang Deng, Jia Guo, Niannan Xue, and Stefanos Zafeiriou. Arcface: Additive angular margin loss for deep face recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4690-4699, 2019. 3, 5, 6, 7, 8, 1, 2", + "[8] Yu Deng, Jiaolong Yang, Sicheng Xu, Dong Chen, Yunde Jia, and Xin Tong. Accurate 3d face reconstruction with weakly-supervised learning: From single image to image set. In IEEE Computer Vision and Pattern Recognition Workshops, 2019. 6, 3", + "[9] Yandong Guo, Lei Zhang, Yuxiao Hu, Xiaodong He, and Jianfeng Gao. Ms-celeb-1m: A dataset and benchmark for large-scale face recognition. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part III 14, pages 87-102. Springer, 2016. 1, 7", + "[10] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep Residual Learning for Image Recognition. In Proceedings of 2016 IEEE Conference on Computer Vision and Pattern Recognition, pages 770-778. IEEE, 2016. 7", + "[11] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2017. 5", + "[12] Gary B Huang, Marwan Mattar, Tamara Berg, and Eric Learned-Miller. Labeled faces in the wild: A database for studying face recognition in unconstrained environments. In Workshop on faces in 'Real-Life' Images: detection, alignment, and recognition, 2008. 1, 3" + ], + "bbox": [ + 78, + 114, + 470, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[13] Yuge Huang, Yuhan Wang, Ying Tai, Xiaoming Liu, Pengcheng Shen, Shaoxin Li, Jilin Li, and Feiyue Huang. Curricularface: adaptive curriculum learning loss for deep face recognition. In proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5901-5910, 2020. 1, 3", + "[14] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4401-4410, 2019. 1, 2, 3, 4", + "[15] Tero Karras, Miika Aittala, Janne Hellsten, Samuli Laine, Jaakko Lehtinen, and Timo Aila. Training generative adversarial networks with limited data. In Proc. NeurIPS, 2020. 1, 2, 5", + "[16] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of stylegan. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8110-8119, 2020. 4", + "[17] Tero Karras, Miika Aittala, Samuli Laine, Erik Härkönen, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Alias-free generative adversarial networks. In Proc. NeurIPS, 2021. 2", + "[18] Minchul Kim, Anil K Jain, and Xiaoming Liu. Adaface: Quality adaptive margin for face recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022. 1, 3, 7", + "[19] Davis E. King. Dlib-ml: A machine learning toolkit. J. Mach. Learn. Res., 10:1755-1758, 2009. 3", + "[20] Brendan F Klare, Ben Klein, Emma Taborsky, Austin Blanton, Jordan Cheney, Kristen Allen, Patrick Grother, Alan Mah, and Anil K Jain. Pushing the frontiers of unconstrained face detection and recognition: Iarpa janus benchmark a. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1931-1939, 2015. 3", + "[21] Harold W Kuhn. The hungarian method for the assignment problem. Naval research logistics quarterly, 2(1-2):83-97, 1955. 3", + "[22] Tuomas Kynkänniemi, Tero Karras, Samuli Laine, Jaakko Lehtinen, and Timo Aila. Improved precision and recall metric for assessing generative models. In Proceedings of the 33rd International Conference on Neural Information Processing Systems, Red Hook, NY, USA, 2019. Curran Associates Inc. 5", + "[23] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. BLIP-2: bootstrapping language-image pre-training with frozen image encoders and large language models. In ICML, 2023. 4", + "[24] Ziwei Liu, Ping Luo, Xiaogang Wang, and Xiaou Tang. Deep learning face attributes in the wild. In Proceedings of International Conference on Computer Vision (ICCV), 2015. 1", + "[25] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In International Conference on Learning Representations, 2017. 6", + "[26] Camillo Lugaresi, Jiuqiang Tang, Hadon Nash, Chris McClanahan, Esha Uboweja, Michael Hays, Fan Zhang, Chuo-Ling Chang, Ming Guang Yong, Juhyun Lee, Wan-Teh" + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "22613", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Chang, Wei Hua, Manfred Georg, and Matthias Grundmann. Mediapipe: A framework for building perception pipelines. CoRR, abs/1906.08172, 2019. 4", + "[27] Brianna Maze, Jocelyn Adams, James A Duncan, Nathan Kalka, Tim Miller, Charles Otto, Anil K Jain, W Tyler Niggel, Janet Anderson, Jordan Cheney, et al. Iarpa janus benchmark-c: Face dataset and protocol. In 2018 international conference on biometrics (ICB), pages 158-165. IEEE, 2018. 3", + "[28] Stylianos Moschoglou, Athanasios Papaioannou, Christos Sagonas, Jiankang Deng, Irene Kotsia, and Stefanos Zafeiriou. Agedb: the first manually collected, in-the-wild age database. In proceedings of the IEEE conference on computer vision and pattern recognition workshops, pages 51-59, 2017. 1, 3", + "[29] Arsha Nagrani, Joon Son Chung, and Andrew Zisserman. VoxCeleb: A large-scale speaker identification dataset. In Interspeech 2017. ISCA, 2017. 2, 4, 5", + "[30] Aaron Nech and Ira Kemelmacher-Shlizerman. Level playing field for million scale face recognition. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2017. 1", + "[31] Youxin Pang, Yong Zhang, Weize Quan, Yanbo Fan, Xiaodong Cun, Ying Shan, and Dong-Ming Yan. Dpe: Disentanglement of pose and expression for general video portrait editing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 427-436, 2023. 2", + "[32] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv e-prints, pages arXiv-2204, 2022. 2", + "[33] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 2, 4, 6", + "[34] Soumyadip Sengupta, Jun-Cheng Chen, Carlos Castillo, Vishal M Patel, Rama Chellappa, and David W Jacobs. Frontal to profile face verification in the wild. In 2016 IEEE winter conference on applications of computer vision (WACV), pages 1-9. IEEE, 2016. 6", + "[35] Sefik Ilkin Serengil and Alper Ozpinar. Lightface: A hybrid deep face recognition framework. In 2020 Innovations in Intelligent Systems and Applications Conference (ASYU), pages 23-27. IEEE, 2020. 3", + "[36] Sefik Ilkin Serengil and Alper Ozpinar. Hyperextended lightface: A facial attribute analysis framework. In 2021 International Conference on Engineering and Emerging Technologies (ICEET), pages 1-4. IEEE, 2021. 3, 5, 1, 2", + "[37] Sefik Ilkin Serengil and Alper Ozpinar. Hyperextended lightface: A facial attribute analysis framework. In 2021 International Conference on Engineering and Emerging Technologies (ICEET), pages 1-4. IEEE, 2021. 4", + "[38] Aliaksandr Siarohin, Stéphane Lathuilière, Sergey Tulyakov, Elisa Ricci, and Nicu Sebe. First order motion model for image animation. Advances in neural information processing systems, 32, 2019. 7, 2" + ], + "bbox": [ + 78, + 90, + 467, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[39] L Sirovich and M Kirby. Low-dimensional procedure for the characterization of human faces. Journal of the Optical Society of America A, 4(3):519-524, 1987. 1", + "[40] Shaolin Su, Qingsen Yan, Yu Zhu, Cheng Zhang, Xin Ge, Jinqiu Sun, and Yanning Zhang. Blindly assess image quality in the wild guided by a self-adaptive hyper network. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 3", + "[41] Jiale Tao, Biao Wang, Borun Xu, Tiezheng Ge, Yuning Jiang, Wen Li, and Lixin Duan. Structure-aware motion transfer with deformable anchor model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3637-3646, 2022. 2", + "[42] Matthew A Turk and Alex P Pentland. Face recognition using eigenfaces. In Proceedings. 1991 IEEE computer society conference on computer vision and pattern recognition, pages 586-587. IEEE Computer Society, 1991. 1", + "[43] Yaohui Wang, Di Yang, Francois Bremond, and Antitza Dantcheva. Latent image animator: Learning to animate images via latent space navigation. In ICLR 2022-The International Conference on Learning Representations, 2022. 2, 7, 3", + "[44] Cameron Whitelam, Emma Taborsky, Austin Blanton, Brianna Maze, Jocelyn Adams, Tim Miller, Nathan Kalka, Anil K Jain, James A Duncan, Kristen Allen, et al. Iarpa janus benchmark-b face dataset. In proceedings of the IEEE conference on computer vision and pattern recognition workshops, pages 90-98, 2017. 3", + "[45] L Wolf, T Hassner, and I Maoz. Face recognition in unconstrained videos with matched background similarity. In Proceedings of the 2011 IEEE Conference on Computer Vision and Pattern Recognition, pages 529-534, 2011. 1", + "[46] Cho-Ying Wu, Qiangeng Xu, and Ulrich Neumann. Synergy between 3dmm and 3d landmarks for accurate 3d facial geometry. In 2021 International Conference on 3D Vision (3DV), 2021. 3", + "[47] Yiqian Wu, Jing Zhang, Hongbo Fu, and Xiaogang Jin. Lpff: A portrait dataset for face generators across large poses. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 20327-20337, 2023. 1, 2, 3, 4", + "[48] Liangbin Xie, Xintao Wang, Honglun Zhang, Chao Dong, and Ying Shan. Vfhq: A high-quality dataset and benchmark for video face super-resolution. In The IEEE Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), 2022. 1, 2, 3, 4", + "[49] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3836-3847, 2023. 1, 2, 4, 6", + "[50] Zhimeng Zhang, Lincheng Li, Yu Ding, and Changjie Fan. Flow-guided one-shot talking face generation with a high-resolution audio-visual dataset. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3661–3670, 2021. 2", + "[51] Jian Zhao and Hui Zhang. Thin-plate spline motion model for image animation. In Proceedings of the IEEE/CVF Con" + ], + "bbox": [ + 501, + 92, + 890, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "22614", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "ference on Computer Vision and Pattern Recognition, pages 3657-3666, 2022. 2, 7, 3", + "[52] Tianyue Zheng and Weihong Deng. Cross-pose lfw: A database for studying cross-posed face recognition in unconstrained environments. Beijing University of Posts and Telecommunications, Tech. Rep, 5(7), 2018. 3, 4, 2", + "[53] Huayi Zhou, Fei Jiang, and Hongtao Lu. Directmhp: Direct 2d multi-person head pose estimation with full-range angles. arXiv preprint arXiv:2302.01110, 2023. 3", + "[54] Hao Zhu, Wayne Wu, Wentao Zhu, Liming Jiang, Siwei Tang, Li Zhang, Ziwei Liu, and Chen Change Loy. CelebV-HQ: A large-scale video facial attributes dataset. In ECCV, 2022. 1, 2, 3, 4", + "[55] Zheng Zhu, Guan Huang, Jiankang Deng, Yun Ye, Junjie Huang, Xinze Chen, Jiagang Zhu, Tian Yang, Jiwen Lu, Dalong Du, and Jie Zhou. Webface260m: A benchmark unveiling the power of million-scale deep face recognition. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 1, 7" + ], + "bbox": [ + 78, + 92, + 468, + 359 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "22615", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2024/EFHQ_ Multi-purpose ExtremePose-Face-HQ dataset/65a10ed9-cef6-42e8-a141-8586b125aea4_model.json b/2024/EFHQ_ Multi-purpose ExtremePose-Face-HQ dataset/65a10ed9-cef6-42e8-a141-8586b125aea4_model.json new file mode 100644 index 0000000000000000000000000000000000000000..625511926daafb370f23b6b869989ba61c962033 --- /dev/null +++ b/2024/EFHQ_ Multi-purpose ExtremePose-Face-HQ dataset/65a10ed9-cef6-42e8-a141-8586b125aea4_model.json @@ -0,0 +1,2279 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.214, + 0.131, + 0.758, + 0.154 + ], + "angle": 0, + "content": "EFHQ: Multi-purpose ExtremePose-Face-HQ dataset" + }, + { + "type": "text", + "bbox": [ + 0.202, + 0.19, + 0.767, + 0.209 + ], + "angle": 0, + "content": "Trung Tuan Dao\\(^{1,*}\\) Duc Hong Vu\\(^{1,2,*}\\) Cuong Pham\\(^{1,3}\\) Anh Tran\\(^{1}\\)" + }, + { + "type": "text", + "bbox": [ + 0.175, + 0.211, + 0.797, + 0.225 + ], + "angle": 0, + "content": "\\(^{1}\\)VinAI Research, Vietnam \\(^{2}\\)Miami University \\(^{3}\\)Posts & Telecommunications Inst. of Tech., Vietnam" + }, + { + "type": "text", + "bbox": [ + 0.142, + 0.23, + 0.83, + 0.244 + ], + "angle": 0, + "content": "{v.trungdt21,v.ducvh5,v.cuongpv11,v.anhtt152}@vinai.io vudh3@miamioh.edu cuongpv@ptit.edu.vn" + }, + { + "type": "text", + "bbox": [ + 0.424, + 0.246, + 0.543, + 0.261 + ], + "angle": 0, + "content": "*Equal contribution" + }, + { + "type": "image", + "bbox": [ + 0.145, + 0.293, + 0.832, + 0.466 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.473, + 0.896, + 0.586 + ], + "angle": 0, + "content": "Figure 1. Benefits of our proposed dataset (EFHQ). Standard large-scale facial datasets have most images at near frontal views, causing inferior performance of trained models on downstream tasks when dealing with extreme head poses. For instance, the trained 2D image generators and text-to-image ones often produce only near frontal faces, while the 3D face generators and face reenactment methods often show distorted outputs at profile views. The recently proposed dataset LPFF [47] partially handles that issue by providing complementary images at extreme head poses for only 2D and 3D image generation tasks. Our proposed dataset EFHQ provides high-quality extreme-pose images to complement a wide range of face-related tasks. It supports 2D and 3D image generation, with generally better diversity than LPFF. EFHQ also helps correct the outputs of text-to-image generation and face reenactment at extreme views. Finally, EFHQ provides a more challenging pose-based face verification benchmark to better assess the quality of face recognition networks." + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.598, + 0.314, + 0.613 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.629, + 0.474, + 0.903 + ], + "angle": 0, + "content": "The existing facial datasets, while having plentiful images at near frontal views, lack images with extreme head poses, leading to the downgraded performance of deep learning models when dealing with profile or pitched faces. This work aims to address this gap by introducing a novel dataset named Extreme Pose Face High-Quality Dataset (EFHQ), which includes a maximum of 450k high-quality images of faces at extreme poses. To produce such a massive dataset, we utilize a novel and meticulous dataset processing pipeline to curate two publicly available datasets, VFHQ and CelebV-HQ, which contain many high-resolution face videos captured in various settings. Our dataset can complement existing datasets on various facial-related tasks, such as facial synthesis with 2D/3D-aware GAN, diffusion-based text-to-image face generation, and face reenactment. Specifically, training with EFHQ helps models generalize well across diverse poses, significantly improving performance in scenarios involving ex" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.6, + 0.896, + 0.69 + ], + "angle": 0, + "content": "treme views, confirmed by extensive experiments. Additionally, we utilize EFHQ to define a challenging cross-view face verification benchmark, in which the performance of SOTA face recognition models drops \\(5 - 37\\%\\) compared to frontal-to-frontal scenarios, aiming to stimulate studies on face recognition under severe pose conditions in the wild." + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.723, + 0.633, + 0.739 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.895, + 0.903 + ], + "angle": 0, + "content": "The human face is a central focus in computer vision, with extensive research dedicated to tasks like detection, recognition, generation, and manipulation [39, 42]. Numerous large-scale facial datasets[9, 12, 14, 24, 28, 30, 45, 48, 54, 55] have facilitated the face-related studies. Along with deep learning techniques, many tasks have made a considerable leap in performance in recent years. For instance, face recognition systems can recognize near-frontal faces in the wild with near-perfect accuracy [13, 18]. Facial generative models, such as [5, 15, 49], can synthesize realistic" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "22605" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.137 + ], + "angle": 0, + "content": "images indistinguishable from real faces. Face reenactment systems, including [41, 43, 51], can animate an input face based on poses and expressions from any driver video." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.14, + 0.471, + 0.397 + ], + "angle": 0, + "content": "Despite impressive results, the mentioned models produce inferior results when dealing with faces at extreme head poses. As illustrated in Fig. 1, faces synthesized by 3D face generation or face reenactment models are often distorted at profile views, due to the scarcity of extreme pose images in training datasets. Current extensive datasets for face generation leverage portrait images from Google or stock photo repositories like Flickr, which feature images with upfront views and minimal facial rotations. Face reenactment methods leverage video datasets such as Vox-Celeb [29], primarily featuring frames with frontal faces. The instability of pre-processing tools, particularly face alignment modules when handling extreme pose images, further reduces the proportion of those images in the existing datasets. This issue is critical since faces with large head rotations are common in real life, and dealing with them is unavoidable in practical systems." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.399, + 0.471, + 0.791 + ], + "angle": 0, + "content": "This paper aims to fill that void by proposing a novel large-scale dataset of extreme-pose and high-quality facial images called ExtremePose-Face-HQ, or EFHQ for short. Our ambition is to build an \"one-for-all\" dataset that can supplement existing datasets on all mentioned face-related tasks. To achieve that goal, the dataset needs all the following properties: (1) extreme-pose, (2) large-scale, (3) high-quality, (4) in-the-wild, (5) having multiple images for each subject and having subject identity annotated. Given those requirements, we utilize a novel and meticulous dataset processing pipeline to filter and sample frames from two publicly available datasets, including VFHQ [48] and CelebV-HQ [54], which contain high-resolution face videos captured in various settings. This results in a dataset with 450k high-quality images of extreme poses, which are combined with frontal images of the same subject to use as image pairs. These image pairs can support face reenactment and face recognition tasks, while the extreme-pose images alone can complement the existing datasets of face generation training. Note that there is a recent attempt [47] to build a large-pose facial dataset similar to ours. However, that work relies on images crawled from Flickr, thus having limited size (19k images) and missing identity information. Hence, its applications are restricted, covering only several face-generation tasks, unlike our large-scale and multi-purpose dataset." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.796, + 0.47, + 0.903 + ], + "angle": 0, + "content": "We perform a comprehensive set of experiments to validate the effectiveness of EFHQ across three distinct face generation sub-tasks and face reenactment scenarios. Our dataset, EFHQ, when combined with standard training sets, significantly enhances the quality of synthesized images in highly profile or pitched views while maintaining synthesis quality in frontal views. EFHQ also provides a more" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.152 + ], + "angle": 0, + "content": "challenging pose-centric face verification benchmark. It reveals the vulnerability of many state-of-the-art face recognition networks when dealing with extreme-pose images, with TAR@FAR=1e-3 scores dropping 5-37%." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.153, + 0.893, + 0.318 + ], + "angle": 0, + "content": "In summary, our contributions include (1) a large-scale, high-quality, in-the-wild, extreme-posed dataset (EFHQ) designed to complement various face-related tasks, (2) a novel and meticulous dataset processing pipeline ensuring the creation and quality of the EFHQ dataset, and (3) a series of experiments, with plan-to-release models, demonstrating the usefulness of EFHQ in enhancing performance on a wide range of tasks when dealing with faces at extreme rotations. The final dataset will be publicly available as metadata files to not violate the two original dataset's licenses, alongside the pretrained networks." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.333, + 0.642, + 0.348 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.365, + 0.892, + 0.516 + ], + "angle": 0, + "content": "Synthetic Face Generation. Existing 2D-based models like StyleGAN [14, 15, 17] and 3D-aware models like EG3D [5] demonstrate impressive fidelity, especially for near-frontal views. Nevertheless, their performance significantly degrades when faced with extreme views, primarily due to insufficient training data for such scenarios. Datasets like FFHQ [14] exhibit a notable bias toward near-frontal poses, heavily limiting the range of poses synthesized by current 2D and 3D models despite the 3D-aware generators' ability to handle arbitrary poses." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.517, + 0.893, + 0.713 + ], + "angle": 0, + "content": "Diffusion-based text-to-image models [32, 33] can synthesize high-quality images from simple text inputs. However, these models lack mechanisms to control the generated pose. Consequently, the generated facial images remain in a frontal view despite terms like \"profile\" or \"looking-up\" in the input prompts. ControlNet[49] proposes a finetuning approach allowing these models to take additional context as conditional input. Namely, one can use landmarks to guide the generated face to the desired pose. However, the scarcity of profile-view data in training affects the model's performance, particularly in extreme angles. Consequently, effectively managing the long-tailed distribution of these extreme views becomes necessary to address this issue." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.893, + 0.901 + ], + "angle": 0, + "content": "Face Reenactment. Face reenactment involves transferring facial expressions and movements from a driving image to a source image. Recent advancements claim high-quality results, especially for near-frontal poses [31, 41, 43, 51]. However, pose collapse persists, particularly in keypoint-based models [41, 51]. Common datasets for this task are VoxCeleb1 [29] and HDTF [50], which primarily feature frontal talking face videos. Despite recent potentially high-quality datasets for reenactment such as VFHQ [48] and CelebV-HQ [54], the frontal-view bias persists. As a result, facial reenactment methods often produce artifacts and lower-quality outputs when faced with significant rotations." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.521, + 0.957 + ], + "angle": 0, + "content": "22606" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.16, + 0.089, + 0.805, + 0.297 + ], + "angle": 0, + "content": "
DatasetAttributesReady-to-use tasks
#imagesMultiple images per IDHigh QualityProfile ViewFace VerificationFace SynthesisFace Reenactment
AgeDB12K✓✓
CPLFW12K✓✓✓
CFP-FP14K✓✓✓
IJB-x138K✓✓✓
CelebA-HQ30K✓✓
FFHQ70K✓✓
LPFF19K✓✓✓
CelebV-HQ6M✓✓✓✓
VFHQ3M✓✓✓✓✓
VoxCeleb14M✓✓✓
Ours450,538✓✓✓✓✓✓
FFHQ+Ours*112,671✓✓✓
Vox+Ours*4.5M✓✓✓
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.302, + 0.892, + 0.358 + ], + "angle": 0, + "content": "Table 1. Representative comparison of key attributes and supported tasks across existing face datasets and our proposed EFHQ dataset. EFHQ provides large-scale, high-quality images with multiple views per identity, enabling diverse face-related tasks, including synthesis, reenactment, and verification. Combining EFHQ with existing datasets like FFHQ or VoxCeleb can augment available data for selected tasks. Asterisk (*) indicates a combined dataset with a subsampled version of EFHQ." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.37, + 0.47, + 0.522 + ], + "angle": 0, + "content": "Face Verification. Despite near-perfect accuracy on benchmark datasets[12, 20, 27, 28, 44], state-of-the-art face verification models[7, 13, 18] fall short in unconstrained environments due to a lack of pose diversity. While CPLFW[52] introduces 6000 cross-pose pairs, its small size and mix of pose samples hinder analyzing performance on specific pose scenarios such as frontal-frontal, frontal-profile, and profile-profile. More granular benchmarking on diverse pose pairs is crucial to identify failure modes and drive progress in robust face verification across poses." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.529, + 0.47, + 0.726 + ], + "angle": 0, + "content": "Extreme Pose Face Dataset. Tab. 1 summarizes existing datasets related to face verification, synthesis, and re-enactment, showing that existing face datasets have limitations for multi-task use. Recent efforts like LPFF[47] made progress by crawling from Flickr and processing for diverse views, resulting in a dataset of 19,590 images. The dataset could complement high-quality datasets like FFHQ[14] for generative models. However, due to its limited scale comprising only \\(20\\%\\) of the combined dataset, the unbalanced distribution between frontal and profile view still needs further improvements. Additionally, since the images are from image hosting services, the dataset fails to provide the identity information necessary for other facial applications." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.741, + 0.233, + 0.757 + ], + "angle": 0, + "content": "3. Dataset Process" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.767, + 0.469, + 0.829 + ], + "angle": 0, + "content": "In this section, we discuss the process of producing our novel dataset EFHQ. To fulfill our goals, we carefully curated frames from the two recent facial video datasets, VFHQ[48] and CelebV-HQ[54], as seen in Fig. 2." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.839, + 0.289, + 0.856 + ], + "angle": 0, + "content": "3.1. Attributes Preparation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.871, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Attributes Extraction. To curate accurate facial poses and high-quality frames from existing datasets, we first ex" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.37, + 0.892, + 0.567 + ], + "angle": 0, + "content": "tract (1) face bounding boxes, (2) facial landmarks (5/68 keypoints), (3) image quality score, and (4) face identity. For bounding boxes and landmarks, we employ the popular RetinaFace[36] and SynergyNet[46]. We use HyperIQA [40] to grade image quality. For VFHQ, we only generate 68 landmarks and image quality scores since bounding boxes and 5 keypoints are provided. After extracting these attributes, we match them with existing annotations using an IoU-based Hungarian matching algorithm [21]. Regarding identity, VFHQ-wise, we source labels from the annotations. As CelebV-HQ lacks individual identity labels within the video, we use[35] to identify the identities associated with each bounding box." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.571, + 0.892, + 0.829 + ], + "angle": 0, + "content": "Due to the instability of head pose estimators when handling extreme-pose images, we strategically apply multiple pose estimators relying on various methodologies: (1) SynergyNet[46], a 3DMM-based model, (2) DirectMHP[53], a landmark-free joint head detector and pose estimator, and (3) an in-house FacePoseNet[6] improved through extensive data augmentation focused on extreme poses and extra training data. We then categorize each estimated pose into a hierarchical binning scheme illustrated in Fig. 2. Next, we gather bin predictions and perform majority voting to arrive at consensus labels. When no agreement emerges, we put it into the \\(5^{th}\\) bin for \"confusing\" cases. As depicted in Fig. 3, this ensemble approach provides robustness in cases where a single estimator incorrectly categorizes a sample. Please refer to the supplementary for statistics of the extracted bins and the hyperparameters for each model used for attribute extraction." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Annotation Review. Next, we manually review the results to ensure high-quality labels. Given the large dataset, we developed a graphical user interface tool to streamline the review process. Our focus is verifying the binning annota" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.958 + ], + "angle": 0, + "content": "22607" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.16, + 0.08, + 0.808, + 0.257 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.262, + 0.895, + 0.304 + ], + "angle": 0, + "content": "Figure 2. The pipeline of EFHQ dataset creation. Starting with high-quality videos from the VFHQ[48] and CelebV-HQ[54] datasets, single-frame attributes are extracted then manually reviews. Task-specific preprocessing is then applied to generate specialized versions of the dataset for tasks such as face generation, reenactment, and verification." + }, + { + "type": "image", + "bbox": [ + 0.098, + 0.322, + 0.449, + 0.449 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.453, + 0.47, + 0.48 + ], + "angle": 0, + "content": "Figure 3. Example cases where a pose estimator fails to categorize the sample to the correct bin." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.497, + 0.47, + 0.558 + ], + "angle": 0, + "content": "tions, which most directly impact label accuracy. Concurrently, we randomly subsample the data based on pose angle and image quality to validate facial landmark quality further and discard cases of landmark prediction failure." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.572, + 0.226, + 0.588 + ], + "angle": 0, + "content": "3.2. EFHQ dataset" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.597, + 0.47, + 0.689 + ], + "angle": 0, + "content": "Our final EFHQ dataset comprises up to 450k frames with extreme poses extracted from approximately 5,000 clips. Most of the clips in our dataset include at least one frame with a frontal face and multiple frames with extreme pose angles. We also include extreme pose-only clips, representing profile-to-profile pose transfer cases." + }, + { + "type": "image", + "bbox": [ + 0.088, + 0.706, + 0.456, + 0.826 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.828, + 0.47, + 0.899 + ], + "angle": 0, + "content": "Figure 4. Pose distribution comparison between our sampled dataset and other datasets, including FFHQ[14], LPFF[47], CPLFW[52] and 40K random samples from VoxCeleb1 [29]. Our sampled dataset demonstrates greater pose diversity, with increased sample counts across high angle bins." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.323, + 0.692, + 0.337 + ], + "angle": 0, + "content": "3.3. Dataset for subtasks" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.359, + 0.892, + 0.571 + ], + "angle": 0, + "content": "Supplementary datasets for face generation. To address FFHQ's pose distribution gap, we compile a dataset encompassing diverse poses, varied identities, and FFHQ-comparable image quality from EFHQ. Additionally, we also embed an image brightness filter[3] in the sampling process, given that HyperIQA doesn't explicitly assess this aspect. The sampled images are then processed using the same pipeline in [5, 16] and manually reviewed. The aim is to improve coverage of under-represented poses without degrading frontal performance of trained generation models. This yields a dataset of 42,671 images with equitably distributed poses as visualized in Fig. 4. Compared to the original FFHQ distribution, yaw and pitch distributions are significantly improved." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.577, + 0.892, + 0.669 + ], + "angle": 0, + "content": "Additionally, for 3D-aware GAN subtasks, we extract camera parameters to serve as conditional pose information, following [5]. We convert these parameters into yaw, pitch angles to validate against our existing annotations. When disagreements emerge, we re-examine the images and filter if needed, to ensure accurate, quality labels." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.675, + 0.893, + 0.903 + ], + "angle": 0, + "content": "For diffusion-based text-to-image generation, a specialized dataset is curated to refine Stable Diffusion models[33] by accommodating landmark conditional input through ControlNet[49]. Based on our 3D-aware face generation dataset, we integrate landmark-based conditional images and tailored text prompts for each image. To capture fine facial nuances, we extract 478 facial landmarks and draw the condition image with connected edges between matched keypoints, following Mediapipe framework[26], better-conveying face expression and pose. For text prompts, detailed facial attributes like gender, race, and emotion are derived from existing labels or inferred via BLIP-2 pretrained captioning model [23] and a face attribute estimator [37]. The resulting prompt adheres to a structured format: \"A profile portrait image of a [emotion] [race] [gender].\"" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "22608" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.08, + 0.089, + 0.472, + 0.26 + ], + "angle": 0, + "content": "
ModelReference DatasetFID↓Recall↑
GFFHQFFHQ2.840.49
GFFHQ+LPFFFFHQ+LPFF3.430.44
GFFHQ+EFHQFFHQ+EFHQ3.440.46
GFFHQ+LPFFFFHQ+EFHQ8.260.43
GFFHQ+EFHQFFHQ+LPFF7.040.44
GFFHQ+EFHQFFHQ+EFHQ3.330.46
GFFHQ+EFHQFFHQ3.120.44
GFFHQ+EFHQFFHQ+LPFF4.470.42
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.265, + 0.47, + 0.349 + ], + "angle": 0, + "content": "Table 2. StyleGAN2-ADA models trained on different datasets. Lower FIDs indicate better fidelity, while higher Recalls indicate better diversity. The second and third blocks show comparisons between unconditional models trained on FFHQ+LPFF and FFHQ+EFHQ when evaluating on the same and cross-dataset settings, with the better metrics in bold." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.356, + 0.47, + 0.418 + ], + "angle": 0, + "content": "Supplementary dataset for face reenactment. As EFHQ is crafted, we can employ the entire dataset to complement established training datasets, e.g., VoxCeleb1 [29]. We also built an extra evaluation set with 1200 EFHQ clips." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.425, + 0.47, + 0.606 + ], + "angle": 0, + "content": "Benchmarking dataset for face verification. We curate a dataset covering three distinct scenarios: frontal-to-frontal, frontal-to-profile, and profile-to-profile. To enable rigorous benchmarking, we sample 10,000 pairs each for both negative and positive cases per scenario, resulting in a balanced benchmark dataset containing 60,000 image pairs. Images follow the same established preprocessing pipelines in [7]. Next, we filter misaligned images with RetinaFace[36], randomly review, and replace equivalent samples from our diverse corpus, if needed. Finally, to simulate varying image quality, we randomly applied downscaling and compression to samples in the dataset." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.624, + 0.314, + 0.64 + ], + "angle": 0, + "content": "4. Face Generation Subtask" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.649, + 0.192, + 0.667 + ], + "angle": 0, + "content": "4.1. StyleGAN" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.681, + 0.468, + 0.801 + ], + "angle": 0, + "content": "Training details. We train StyleGAN2-ADA[15] models from scratch after combining FFHQ with our own curated dataset (denoted FFHQ+EFHQ) at \\(1024 \\times 1024\\) resolution. We also train a conditional version, allowing direct control over frontal versus profile synthesis and enhancing model performance. Specifically, we categorize the dataset into two classes: \"frontal\" and \"profile\" based on pose angles. All models are trained for 6 days on 8 Nvidia A100 GPUs." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.81, + 0.468, + 0.9 + ], + "angle": 0, + "content": "Evaluation protocol. We utilize the Frechet Inception Distance[11] (FID) metric and improved Recall[22] to quantify the generation fidelity and diversity. Regarding the two unconditional models trained with the supplemented FFHQ+LPFF and FFHQ+EFHQ, we evaluate each model with both of these datasets. For conditional mod" + }, + { + "type": "image", + "bbox": [ + 0.533, + 0.089, + 0.86, + 0.175 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.548, + 0.178, + 0.845, + 0.19 + ], + "angle": 0, + "content": "(a) Samples from the model trained with FFHQ+EFHQ." + }, + { + "type": "image", + "bbox": [ + 0.532, + 0.191, + 0.858, + 0.35 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.531, + 0.351, + 0.858, + 0.365 + ], + "angle": 0, + "content": "(b) Profile-view comparison of LPFF (left) and EFHQ (right)." + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.369, + 0.892, + 0.41 + ], + "angle": 0, + "content": "Figure 5. Qualitative result and comparison of generated samples, with truncation \\(\\psi = 0.7\\), from StyleGAN2-ADA training with other dataset and ours." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.433, + 0.892, + 0.538 + ], + "angle": 0, + "content": "els, we evaluate performance on FFHQ, FFHQ+LPFF, and FFHQ+EFHQ to demonstrate the broader coverage of the model trained with our proposed dataset. For FFHQ, we generate 50K images with the \"frontal\" label, whereas for the other two datasets, we generate 50K following their label distributions. This comprehensive quantitative analysis examines the fidelity and variety of the generated images." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.554, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Experimental results. We compared the performance of our proposed unconditional models \\((G^{\\mathrm{FFHQ + EFHQ}})\\), against models trained on the FFHQ and FFHQ+LPFF datasets \\((G^{\\mathrm{FFHQ}}\\) and \\(G^{\\mathrm{FFHQ + LPFF}})\\). Models trained conditionally are specifically subscripted, e.g., \\(G_{c}^{\\mathrm{FFHQ + EFHQ}}\\). The qualitative results in Fig. 5 illustrate that our models generate high-quality frontal faces comparable to the FFHQ model while producing realistic and varied profile faces. Additionally, our method has less noise in extreme poses than the FFHQ+LPFF model. Specifically, Fig. 5b shows LPFF's generated face contains more noise and less photorealistic details than our model, such as the pixelated eye and nose patches. Meanwhile, ours achieves greater realism with smoother skin, well-defined features, noise reduction, and natural lighting. The quantitative results in Tab. 2 show that our unconditional model achieves competitive FID scores and improved Recall to other models, indicating similar fidelity but greater sample pose diversity. In the cross-examination (third block of Tab. 2), \\(G^{\\mathrm{FFHQ + EFHQ}}\\) has both better FID and improved Recall on the opposite dataset than \\(G^{\\mathrm{FFHQ + LPFF}}\\). Finally, our conditional model demonstrates substantially improved FID scores over multiple reference datasets, highlighting its versatility." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "22609" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.081, + 0.089, + 0.468, + 0.206 + ], + "angle": 0, + "content": "
ModelReference DatasetFID↓Pose↓Depth↓ID↑
EFHQFFHQ/Trainset4.85e-40.2340.74
EFFHQFFHQ/Trainset3.75e-40.2880.74
EFFHQ+LPFFFFHQ6.26e-40.3160.77
EFFHQ+EFHQFFHQ3.94e-40.2240.77
EFFHQ+LPFFTrainset7.06e-40.3160.65
EFFHQ+EFHQTrainset4.64e-40.2230.65
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.211, + 0.468, + 0.294 + ], + "angle": 0, + "content": "Table 3. EG3D models trained on different datasets. In the second and third blocks, we compare models trained on FFHQ+LPFF and FFHQ+EFHQ with the testset is either FFHQ or their training datasets; the better metrics are in bold. The best metrics on the FFHQ testset are in red, while the best metrics when evaluating on training dataset are underlined." + }, + { + "type": "image", + "bbox": [ + 0.089, + 0.295, + 0.457, + 0.458 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.461, + 0.468, + 0.502 + ], + "angle": 0, + "content": "Figure 6. Comparison between multiview generated samples, with truncation \\(\\psi = 0.8\\), of EG3D model trained with various datasets." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.508, + 0.163, + 0.523 + ], + "angle": 0, + "content": "4.2. EG3D" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.539, + 0.468, + 0.645 + ], + "angle": 0, + "content": "Training details. We followed the same pipeline and hyper-parameters as the original EG3D model [5]. However, based on empirical results, we adjusted \\(\\gamma = 5\\). We also fixed the triplane implementation bug (XY, XZ, ZX) [1] to replicate the original idea in the papers (XY, XZ, ZY). We trained the model on 8 Nvidia A100 40GB GPUs for approximately 7 days until convergence." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.652, + 0.468, + 0.759 + ], + "angle": 0, + "content": "Evaluation Protocol. To evaluate the models, mirroring [5], we measure image quality with FID; multi-view consistency and identity preservation with mean cosine similarity of pre-trained face recognition model[7]; pose accuracy and geometry by calculating MSE against pseudo labels from [8]. We evaluate all models on 1,024 generated images, except for FID, which we compute on 50k images." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.468, + 0.902 + ], + "angle": 0, + "content": "Experimental results. To evaluate the benefits of our dataset, we trained the same backbone architecture on different datasets: FFHQ \\((E^{\\mathrm{FFHQ}})\\), FFHQ+LPFF \\((E^{\\mathrm{FFHQ} + \\mathrm{LPFF}})\\), and FFHQ+EFHQ (our model, \\(E^{\\mathrm{FFHQ} + \\mathrm{EFHQ}}\\)). The quantitative results are presented in Table 3, and qualitative examples are shown in Figure 6. Qualitatively, we observe that \\(E^{\\mathrm{FFHQ} + \\mathrm{EFHQ}}\\) generates improved face shapes in extreme poses compared to \\(E^{\\mathrm{FFHQ}}\\), similar to the improvements seen with \\(E^{\\mathrm{FFHQ} + \\mathrm{LPFF}}\\). Quantitatively, the FID" + }, + { + "type": "table", + "bbox": [ + 0.501, + 0.089, + 0.889, + 0.218 + ], + "angle": 0, + "content": "
MetricsDatasetPretrained ControlNetOurs
NME↓LPFF0.0850.084
FID↓LPFF119.7483.39
FIDeyepatch ↓LPFF103.7749.45
NME↓EFHQ0.1060.081
FID↓EFHQ111.2381.90
FIDeyepatch ↓EFHQ137.4358.60
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.221, + 0.892, + 0.263 + ], + "angle": 0, + "content": "Table 4. Quantitative results of fine-tuning StableDiffusion v1.5 with ControlNet trained on various datasets, using Normalized Mean Error (NME), FID, and FID on eye region." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.271, + 0.892, + 0.47 + ], + "angle": 0, + "content": "of \\( E^{\\mathrm{FFHQ + EFHQ}} \\) is comparable to \\( E^{\\mathrm{FFHQ}} \\) when evaluated on FFHQ, indicating no degradation in frontal view performance. Meanwhile, the competitive FID achieved by \\( E^{\\mathrm{FFHQ + EFHQ}} \\) on the combined FFHQ+EFHQ dataset also suggests proficient modeling of profile views. Regarding identity consistency, we see a significant drop when the reference dataset contains profile-view images, due to the weak confidence of the pretrained face recognition model when handling profile-view images, as confirmed by its low average cosine similarity of around 0.6 on the simple frontal-to-profile verification set CFP-FP[34]. Furthermore, our model also outperforms in pose accuracy and geometry quality with both frontal and cross-pose dataset." + }, + { + "type": "image", + "bbox": [ + 0.514, + 0.481, + 0.696, + 0.579 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.697, + 0.481, + 0.882, + 0.578 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.531, + 0.58, + 0.862, + 0.592 + ], + "angle": 0, + "content": "(a) Generated samples; condition image is in between samples." + }, + { + "type": "image", + "bbox": [ + 0.502, + 0.593, + 0.699, + 0.626 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.7, + 0.593, + 0.894, + 0.626 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.628, + 0.628, + 0.768, + 0.641 + ], + "angle": 0, + "content": "(b) Generated eye regions." + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.646, + 0.892, + 0.689 + ], + "angle": 0, + "content": "Figure 7. Comparison profile-view generated samples of pretrained ControlNet (left) and our fine-tuned ControlNet (right) with the prompt: \"A profile portrait image of a person\"." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.698, + 0.623, + 0.712 + ], + "angle": 0, + "content": "4.3. ControlNet" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.727, + 0.892, + 0.818 + ], + "angle": 0, + "content": "Training details. We finetuned Stable Diffusion v1.5 [33] using ControlNet [49] on the EFHQ dataset to enhance its ability to generate extreme facial poses. We use AdamW optimizer [25] with learning rate \\( 1 \\times e^{-5} \\), \\( \\beta_0 = 0.9 \\), \\( \\beta_1 = 0.999 \\), and \\( \\epsilon = 0.01 \\). We train with batch size 24 on 1 Nvidia A100 40GB GPU in 2 days." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.825, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Evaluation Protocol. To assess image qualities across diverse poses, we measure the Normalized Mean Error (NME) of 478-point facial landmarks between the generated and source image. Specifically, we evaluate performance across the full pose distribution in the reference" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "22610" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.082, + 0.089, + 0.472, + 0.176 + ], + "angle": 0, + "content": "
ModelTest Dataset (L1 ↓/ AKD ↓/ AED ↓)
VoxCeleb1EFHQ (Ours)
TPS(VoxCeleb10.041/1.214/0.1320.061/3.814/0.275
TPS(VoxCeleb1+EFHQ0.041/1.250/0.1330.051/1.963/0.214
LIA(VoxCeleb10.044/1.342/0.1380.059/3.807/0.266
LIA(VoxCeleb1+EFHQ0.044/1.398/0.1420.054/2.291/0.233
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.181, + 0.47, + 0.237 + ], + "angle": 0, + "content": "Table 5. Quantitative results of two state-of-the-art reenactment methods (TPS and LIA) when training on VoxCeleb1 and the combined set VoxCeleb1+EFHQ. We present results evaluated either on VoxCeleb1 (left) or EFHQ (right) test sets." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.242, + 0.47, + 0.364 + ], + "angle": 0, + "content": "dataset. For the prompt, we use: \"A profile portrait image of a person.\" with extra positive and negative keywords, further detailed in supplementary. Utilizing LPFF and EFHQ datasets as the reference datasets, we compute FID to assess synthesized face quality and extend evaluation to cropped eye regions due to unrealistic gazes observed from the original ControlNet model. Note that the models receive condition images created akin to its training data." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.37, + 0.471, + 0.643 + ], + "angle": 0, + "content": "Experimental results. Fig. 7a visually compares the baseline ControlNet OpenPose face and our fine-tuned model. Our fine-tuned model adeptly conditions diverse extreme poses, generating high-quality results with precise details while maintaining fidelity to the condition, such as mouth shape, and exhibiting minimal artifacts. These patterns indicate the superiority of the fine-tuned model over the baseline model in handling extreme viewing angles. Additionally, Fig. 7b contrasts the eye regions of both models. Our fine-tuned model accurately captures natural eye gaze and shape from condition images, a notable improvement over the baseline. Quantitative results in Tab. 4 consistently favor our fine-tuned model across all metrics. FID scores on full and eye patches exhibit significant enhancement, and NME performance remains stable on training and reference datasets (EFHQ and LPFF). These findings showcase the fine-tuned model's capability to generate high-quality, geometrically consistent faces across diverse poses." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.657, + 0.329, + 0.673 + ], + "angle": 0, + "content": "5. Face Reenactment Subtask" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.681, + 0.267, + 0.696 + ], + "angle": 0, + "content": "5.1. Evaluation Protocol" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.704, + 0.471, + 0.903 + ], + "angle": 0, + "content": "We follow the same preprocessing and evaluation protocol as FOMM [38]. Specifically, all the frames are cropped to \\(256 \\times 256\\), and we use the first frame of a clip as the source image to reconstruct the remaining frames. In the case of the EFHQ test set, most source images have a frontal pose, while the target frames have an extreme pose. Some test cases only contain extreme pose to evaluate self-reconstruction on these cases. We run the experiments using two recent methods, Thin-Plate Spline Motion Model (TPS) [51] and Latent Image Animator (LIA) [43]. For training details, please refer to the supplementary. We select the following three metrics to compare the generated frames and the driving ground truth. \\(\\mathcal{L}_1\\) measures the av" + }, + { + "type": "table", + "bbox": [ + 0.504, + 0.089, + 0.89, + 0.196 + ], + "angle": 0, + "content": "
ModelTrainsetCPLFWEFHQ
f2ff2pp2p
R18arfaceGlint360k0.800.940.570.370.880.06
R18arfaceMS1MV30.750.930.570.360.880.05
R50arfaceGlint360k0.910.970.770.200.920.05
R100arfaceGlint360k0.930.980.830.150.820.16
R100adafaceWebFace4M0.940.990.940.050.970.02
" + }, + { + "type": "table_caption", + "bbox": [ + 0.498, + 0.2, + 0.892, + 0.299 + ], + "angle": 0, + "content": "Table 6. Quantitative evaluation of state-of-the-art face recognition models on the cleansed CPLFW and the proposed benchmark dataset. We use the metric TAR@FAR=1e-3 (higher is better). Our EFHQ benchmark includes frontal-frontal (f2f), frontal-profile (f2p), and profile-profile (p2p) subsets. For f2p and p2p, we report the score drop when comparing to f2f in the red subscription. The lowest benchmark score for each model is in bold." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.307, + 0.892, + 0.399 + ], + "angle": 0, + "content": "erage distance among pixel values, indicating reconstruction faithfulness. Average Keypoint Distance (AKD) calculates the mean distance between keypoints, thus assessing the pose accuracy. Average Euclidean Distance (AED) measures identity preservation by computing the average \\(\\mathcal{L}_2\\) distance between identity embeddings." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.413, + 0.697, + 0.429 + ], + "angle": 0, + "content": "5.2. Experimental results" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.437, + 0.893, + 0.725 + ], + "angle": 0, + "content": "We compare TPS and LIA models trained with our supplementary EFHQ dataset \\((TPS^{\\mathrm{VoxCeleb1 + EFHQ}}\\) and \\(LIA^{\\mathrm{VoxCeleb1 + EFHQ}})\\) against models trained solely on VoxCeleb1 \\((TPS^{\\mathrm{VoxCeleb1}}\\) and \\(LIA^{\\mathrm{VoxCeleb1}})\\). Tab. 5 displays quantitative results, revealing that models trained with EFHQ maintain comparable performance when evaluated on VoxCeleb1, emphasizing high-quality performance in frontal-to-frontal reenactment scenarios. Notably, on the EFHQ test set, both EFHQ-trained models show significant improvement in all metrics, indicating enhanced extreme pose and identity reenactment performance. In Fig. 8, qualitative results prove that EFHQ-trained models excel in transferring facial expressions and motion across diverse poses. Models trained on EFHQ showcase improved image quality, reduced artifacts, and well-preserved shapes of driving faces. These results highlight our models' effectiveness in capturing and reproducing facial motions across varying poses while enhancing fidelity to the original facial features." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.743, + 0.742, + 0.761 + ], + "angle": 0, + "content": "6. Face Recognition Subtask" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.77, + 0.691, + 0.785 + ], + "angle": 0, + "content": "6.1. Evaluation Protocol" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.796, + 0.892, + 0.901 + ], + "angle": 0, + "content": "To evaluate face recognition performance on extreme poses, we selected two state-of-the-art methods: ArcFace[7] and AdaFace[18]. We further explored various backbone sizes (ResNet18, ResNet50, ResNet101)[10] and training datasets (MS1MV2[9], Glint360K[2], WebFace4M[55]) of increasing size. For metrics, we used True Acceptance Rate (TAR) at False Acceptance Rate (FAR) of \\( 1 \\times e^{-3} \\)." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.958 + ], + "angle": 0, + "content": "22611" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.169, + 0.1, + 0.266, + 0.173 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.169, + 0.175, + 0.268, + 0.248 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.199, + 0.255, + 0.236, + 0.264 + ], + "angle": 0, + "content": "Source" + }, + { + "type": "image", + "bbox": [ + 0.279, + 0.1, + 0.375, + 0.173 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.279, + 0.175, + 0.376, + 0.25 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.285, + 0.254, + 0.372, + 0.265 + ], + "angle": 0, + "content": "TPS (VoxCeleb1)" + }, + { + "type": "image", + "bbox": [ + 0.38, + 0.1, + 0.475, + 0.173 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.38, + 0.175, + 0.475, + 0.25 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.398, + 0.255, + 0.458, + 0.266 + ], + "angle": 0, + "content": "TPS (Ours)" + }, + { + "type": "image", + "bbox": [ + 0.49, + 0.1, + 0.587, + 0.173 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.49, + 0.175, + 0.587, + 0.248 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.497, + 0.254, + 0.585, + 0.265 + ], + "angle": 0, + "content": "LIA (VoxCeleb1)" + }, + { + "type": "image", + "bbox": [ + 0.593, + 0.1, + 0.688, + 0.173 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.593, + 0.175, + 0.688, + 0.248 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.61, + 0.254, + 0.669, + 0.265 + ], + "angle": 0, + "content": "LIA (Ours)" + }, + { + "type": "image", + "bbox": [ + 0.703, + 0.1, + 0.798, + 0.173 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.702, + 0.175, + 0.797, + 0.248 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.73, + 0.254, + 0.771, + 0.266 + ], + "angle": 0, + "content": "Driving" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.27, + 0.892, + 0.3 + ], + "angle": 0, + "content": "Figure 8. Comparison of performance for frontal-profile face reenactment on EFHQ. From left to right: source image, TPS on VoxCeleb1, TPS with EFHQ, LIA on VoxCeleb1, LIA with EFHQ, driving image. Additional EFHQ training data improves synthesis of extreme poses." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.31, + 0.265, + 0.326 + ], + "angle": 0, + "content": "6.2. Experimental results" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.333, + 0.469, + 0.424 + ], + "angle": 0, + "content": "Before assessing face recognition networks on EFHQ, we first clean the commonly-used version of CPLFW, processed by[7], by removing misaligned images (please refer to the Supplementary), and evaluate the models on it. Models performed well on this cleansed set, with minimal gaps between ArcFace and AdaFace." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.424, + 0.47, + 0.636 + ], + "angle": 0, + "content": "Our proposed evaluation benchmark involves three pose-based scenarios. Results in Tab. 6 reveal performance drops of \\(5 - 37\\%\\) on the frontal-profile subset compared to frontal-frontal, highlighting the challenge of matching faces in different poses. Larger ResNet models outperformed ResNet18 across subsets, showcasing the advantages of increased capacity in handling pose variation. However, the ResNet100 model trained on Glint360K performed worse than smaller models on the profile-profile subset, indicating the need for sufficient profile data to prevent overfitting. Our results also indicate AdaFace's suitability for handling pose variations. In summary, pose mismatch brings significant challenges for face recognition, necessitating larger models trained on diverse data covering profile faces." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.647, + 0.192, + 0.661 + ], + "angle": 0, + "content": "7. Discussion" + }, + { + "type": "image", + "bbox": [ + 0.097, + 0.669, + 0.447, + 0.814 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.819, + 0.468, + 0.846 + ], + "angle": 0, + "content": "Figure 9. User study results. We reported the percentage of times testers rank whose output is best or tied for each criterion." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.871, + 0.469, + 0.902 + ], + "angle": 0, + "content": "User survey. We conduct user surveys to evaluate the face generation improvement of our method through human per" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.311, + 0.892, + 0.402 + ], + "angle": 0, + "content": "ceptual assessment. In this survey, we compare the outputs of the models mentioned in Secs. 4 and 5. The setup of this user survey is included in the supplementary. In all human evaluation results (Fig. 9), our models outperform previous ones, notably on identity and motion test of face reenactment models (74.80% and 64.22%, respectively)." + }, + { + "type": "image", + "bbox": [ + 0.529, + 0.42, + 0.861, + 0.507 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.511, + 0.892, + 0.58 + ], + "angle": 0, + "content": "Figure 10. Representative examples of the raw LPFF dataset that got excluded from the final dataset, either from misdetections or pose filtering process. Red keypoints represent landmarks prediction of selected samples from LPFF, green keypoints represent landmarks prediction from our pipeline." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.595, + 0.892, + 0.7 + ], + "angle": 0, + "content": "Robustness of data processing pipeline. As shown in Fig. 10, our data processing pipeline can detect additional profile-view faces filtered out by the LPFF pipeline, likely due to misdetections or incorrect pose filtering. This confirms the robustness of our pipeline, and we hope it will enable the community to develop larger and higher-quality in-the-wild facial datasets." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.721, + 0.619, + 0.736 + ], + "angle": 0, + "content": "8. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.748, + 0.892, + 0.888 + ], + "angle": 0, + "content": "This work has introduced a large-scale, diverse facial dataset to address performance gaps between frontal and profile faces, powered by a novel and robust data processing pipeline. Crucially, we provided tailored sub-datasets for advancing essential tasks like face synthesis and reenactment. Moreover, our new face verification benchmark reveals gaps between techniques, granting valuable insights. Ultimately, we hope this high-quality, diverse facial dataset opens up new and exciting opportunities to push forward cross-pose tasks." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "22612" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.156 + ], + "angle": 0, + "content": "[1] Fix for triplane projection \\(\\cdot\\) Issue #67 \\(\\cdot\\) NVlabs/eg3d. https://github.com/NVlabs/eg3d/issues/67, 2023.6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.157, + 0.472, + 0.227 + ], + "angle": 0, + "content": "[2] Xiang An, Xuhan Zhu, Yuan Gao, Yang Xiao, Yongle Zhao, Ziyong Feng, Lan Wu, Bin Qin, Ming Zhang, Debing Zhang, et al. Partial fc: Training 10 million identities on a single machine. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1445-1449, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.228, + 0.471, + 0.282 + ], + "angle": 0, + "content": "[3] Sergey Bezryadin, Pavel Bourov, and Dmitry Ilinih. Brightness calculation in digital image processing. International Symposium on Technologies for Digital Photo Fulfillment, 2007:10-15, 2007. 4, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.284, + 0.471, + 0.34 + ], + "angle": 0, + "content": "[4] Z. Cao, G. Hidalgo Martinez, T. Simon, S. Wei, and Y. A. Sheikh. Openpose: Realtime multi-person 2d pose estimation using part affinity fields. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2019. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.341, + 0.471, + 0.41 + ], + "angle": 0, + "content": "[5] Eric R. Chan, Connor Z. Lin, Matthew A. Chan, Koki Nagano, Boxiao Pan, Shalini De Mello, Orazio Gallo, Leonidas Guibas, Jonathan Tremblay, Sameh Khamis, Tero Karras, and Gordon Wetzstein. Efficient geometry-aware 3D generative adversarial networks. In CVPR, 2022. 1, 2, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.411, + 0.471, + 0.48 + ], + "angle": 0, + "content": "[6] Feng-Ju Chang, Anh Tuan Tran, Tal Hassner, Iacopo Masi, Ram Nevatia, and Gerard Medioni. Faceposenet: Making a case for landmark-free face alignment. In Proceedings of the IEEE International Conference on Computer Vision Workshops, pages 1599-1608, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.481, + 0.471, + 0.549 + ], + "angle": 0, + "content": "[7] Jiankang Deng, Jia Guo, Niannan Xue, and Stefanos Zafeiriou. Arcface: Additive angular margin loss for deep face recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4690-4699, 2019. 3, 5, 6, 7, 8, 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.551, + 0.471, + 0.62 + ], + "angle": 0, + "content": "[8] Yu Deng, Jiaolong Yang, Sicheng Xu, Dong Chen, Yunde Jia, and Xin Tong. Accurate 3d face reconstruction with weakly-supervised learning: From single image to image set. In IEEE Computer Vision and Pattern Recognition Workshops, 2019. 6, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.621, + 0.471, + 0.704 + ], + "angle": 0, + "content": "[9] Yandong Guo, Lei Zhang, Yuxiao Hu, Xiaodong He, and Jianfeng Gao. Ms-celeb-1m: A dataset and benchmark for large-scale face recognition. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part III 14, pages 87-102. Springer, 2016. 1, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.705, + 0.47, + 0.761 + ], + "angle": 0, + "content": "[10] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep Residual Learning for Image Recognition. In Proceedings of 2016 IEEE Conference on Computer Vision and Pattern Recognition, pages 770-778. IEEE, 2016. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.761, + 0.47, + 0.83 + ], + "angle": 0, + "content": "[11] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.831, + 0.47, + 0.901 + ], + "angle": 0, + "content": "[12] Gary B Huang, Marwan Mattar, Tamara Berg, and Eric Learned-Miller. Labeled faces in the wild: A database for studying face recognition in unconstrained environments. In Workshop on faces in 'Real-Life' Images: detection, alignment, and recognition, 2008. 1, 3" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.175 + ], + "angle": 0, + "content": "[13] Yuge Huang, Yuhan Wang, Ying Tai, Xiaoming Liu, Pengcheng Shen, Shaoxin Li, Jilin Li, and Feiyue Huang. Curricularface: adaptive curriculum learning loss for deep face recognition. In proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5901-5910, 2020. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.177, + 0.892, + 0.245 + ], + "angle": 0, + "content": "[14] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4401-4410, 2019. 1, 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.246, + 0.892, + 0.3 + ], + "angle": 0, + "content": "[15] Tero Karras, Miika Aittala, Janne Hellsten, Samuli Laine, Jaakko Lehtinen, and Timo Aila. Training generative adversarial networks with limited data. In Proc. NeurIPS, 2020. 1, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.301, + 0.892, + 0.37 + ], + "angle": 0, + "content": "[16] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of stylegan. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8110-8119, 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.371, + 0.892, + 0.413 + ], + "angle": 0, + "content": "[17] Tero Karras, Miika Aittala, Samuli Laine, Erik Härkönen, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Alias-free generative adversarial networks. In Proc. NeurIPS, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.414, + 0.892, + 0.468 + ], + "angle": 0, + "content": "[18] Minchul Kim, Anil K Jain, and Xiaoming Liu. Adaface: Quality adaptive margin for face recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022. 1, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.469, + 0.892, + 0.496 + ], + "angle": 0, + "content": "[19] Davis E. King. Dlib-ml: A machine learning toolkit. J. Mach. Learn. Res., 10:1755-1758, 2009. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.497, + 0.892, + 0.58 + ], + "angle": 0, + "content": "[20] Brendan F Klare, Ben Klein, Emma Taborsky, Austin Blanton, Jordan Cheney, Kristen Allen, Patrick Grother, Alan Mah, and Anil K Jain. Pushing the frontiers of unconstrained face detection and recognition: Iarpa janus benchmark a. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1931-1939, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.581, + 0.892, + 0.62 + ], + "angle": 0, + "content": "[21] Harold W Kuhn. The hungarian method for the assignment problem. Naval research logistics quarterly, 2(1-2):83-97, 1955. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.622, + 0.892, + 0.704 + ], + "angle": 0, + "content": "[22] Tuomas Kynkänniemi, Tero Karras, Samuli Laine, Jaakko Lehtinen, and Timo Aila. Improved precision and recall metric for assessing generative models. In Proceedings of the 33rd International Conference on Neural Information Processing Systems, Red Hook, NY, USA, 2019. Curran Associates Inc. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.705, + 0.892, + 0.761 + ], + "angle": 0, + "content": "[23] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. BLIP-2: bootstrapping language-image pre-training with frozen image encoders and large language models. In ICML, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.761, + 0.892, + 0.816 + ], + "angle": 0, + "content": "[24] Ziwei Liu, Ping Luo, Xiaogang Wang, and Xiaou Tang. Deep learning face attributes in the wild. In Proceedings of International Conference on Computer Vision (ICCV), 2015. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.817, + 0.892, + 0.859 + ], + "angle": 0, + "content": "[25] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In International Conference on Learning Representations, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.86, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[26] Camillo Lugaresi, Jiuqiang Tang, Hadon Nash, Chris McClanahan, Esha Uboweja, Michael Hays, Fan Zhang, Chuo-Ling Chang, Ming Guang Yong, Juhyun Lee, Wan-Teh" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "22613" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.092, + 0.468, + 0.134 + ], + "angle": 0, + "content": "Chang, Wei Hua, Manfred Georg, and Matthias Grundmann. Mediapipe: A framework for building perception pipelines. CoRR, abs/1906.08172, 2019. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.135, + 0.468, + 0.216 + ], + "angle": 0, + "content": "[27] Brianna Maze, Jocelyn Adams, James A Duncan, Nathan Kalka, Tim Miller, Charles Otto, Anil K Jain, W Tyler Niggel, Janet Anderson, Jordan Cheney, et al. Iarpa janus benchmark-c: Face dataset and protocol. In 2018 international conference on biometrics (ICB), pages 158-165. IEEE, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.218, + 0.468, + 0.3 + ], + "angle": 0, + "content": "[28] Stylianos Moschoglou, Athanasios Papaioannou, Christos Sagonas, Jiankang Deng, Irene Kotsia, and Stefanos Zafeiriou. Agedb: the first manually collected, in-the-wild age database. In proceedings of the IEEE conference on computer vision and pattern recognition workshops, pages 51-59, 2017. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.302, + 0.468, + 0.342 + ], + "angle": 0, + "content": "[29] Arsha Nagrani, Joon Son Chung, and Andrew Zisserman. VoxCeleb: A large-scale speaker identification dataset. In Interspeech 2017. ISCA, 2017. 2, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.344, + 0.468, + 0.399 + ], + "angle": 0, + "content": "[30] Aaron Nech and Ira Kemelmacher-Shlizerman. Level playing field for million scale face recognition. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.4, + 0.468, + 0.481 + ], + "angle": 0, + "content": "[31] Youxin Pang, Yong Zhang, Weize Quan, Yanbo Fan, Xiaodong Cun, Ying Shan, and Dong-Ming Yan. Dpe: Disentanglement of pose and expression for general video portrait editing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 427-436, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.483, + 0.468, + 0.536 + ], + "angle": 0, + "content": "[32] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv e-prints, pages arXiv-2204, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.539, + 0.468, + 0.607 + ], + "angle": 0, + "content": "[33] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 2, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.608, + 0.468, + 0.677 + ], + "angle": 0, + "content": "[34] Soumyadip Sengupta, Jun-Cheng Chen, Carlos Castillo, Vishal M Patel, Rama Chellappa, and David W Jacobs. Frontal to profile face verification in the wild. In 2016 IEEE winter conference on applications of computer vision (WACV), pages 1-9. IEEE, 2016. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.679, + 0.468, + 0.733 + ], + "angle": 0, + "content": "[35] Sefik Ilkin Serengil and Alper Ozpinar. Lightface: A hybrid deep face recognition framework. In 2020 Innovations in Intelligent Systems and Applications Conference (ASYU), pages 23-27. IEEE, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.735, + 0.468, + 0.789 + ], + "angle": 0, + "content": "[36] Sefik Ilkin Serengil and Alper Ozpinar. Hyperextended lightface: A facial attribute analysis framework. In 2021 International Conference on Engineering and Emerging Technologies (ICEET), pages 1-4. IEEE, 2021. 3, 5, 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.79, + 0.468, + 0.845 + ], + "angle": 0, + "content": "[37] Sefik Ilkin Serengil and Alper Ozpinar. Hyperextended lightface: A facial attribute analysis framework. In 2021 International Conference on Engineering and Emerging Technologies (ICEET), pages 1-4. IEEE, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.846, + 0.468, + 0.9 + ], + "angle": 0, + "content": "[38] Aliaksandr Siarohin, Stéphane Lathuilière, Sergey Tulyakov, Elisa Ricci, and Nicu Sebe. First order motion model for image animation. Advances in neural information processing systems, 32, 2019. 7, 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.468, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.134 + ], + "angle": 0, + "content": "[39] L Sirovich and M Kirby. Low-dimensional procedure for the characterization of human faces. Journal of the Optical Society of America A, 4(3):519-524, 1987. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.137, + 0.892, + 0.205 + ], + "angle": 0, + "content": "[40] Shaolin Su, Qingsen Yan, Yu Zhu, Cheng Zhang, Xin Ge, Jinqiu Sun, and Yanning Zhang. Blindly assess image quality in the wild guided by a self-adaptive hyper network. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.207, + 0.892, + 0.275 + ], + "angle": 0, + "content": "[41] Jiale Tao, Biao Wang, Borun Xu, Tiezheng Ge, Yuning Jiang, Wen Li, and Lixin Duan. Structure-aware motion transfer with deformable anchor model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3637-3646, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.278, + 0.892, + 0.332 + ], + "angle": 0, + "content": "[42] Matthew A Turk and Alex P Pentland. Face recognition using eigenfaces. In Proceedings. 1991 IEEE computer society conference on computer vision and pattern recognition, pages 586-587. IEEE Computer Society, 1991. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.334, + 0.892, + 0.401 + ], + "angle": 0, + "content": "[43] Yaohui Wang, Di Yang, Francois Bremond, and Antitza Dantcheva. Latent image animator: Learning to animate images via latent space navigation. In ICLR 2022-The International Conference on Learning Representations, 2022. 2, 7, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.405, + 0.892, + 0.488 + ], + "angle": 0, + "content": "[44] Cameron Whitelam, Emma Taborsky, Austin Blanton, Brianna Maze, Jocelyn Adams, Tim Miller, Nathan Kalka, Anil K Jain, James A Duncan, Kristen Allen, et al. Iarpa janus benchmark-b face dataset. In proceedings of the IEEE conference on computer vision and pattern recognition workshops, pages 90-98, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.49, + 0.892, + 0.545 + ], + "angle": 0, + "content": "[45] L Wolf, T Hassner, and I Maoz. Face recognition in unconstrained videos with matched background similarity. In Proceedings of the 2011 IEEE Conference on Computer Vision and Pattern Recognition, pages 529-534, 2011. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.547, + 0.892, + 0.602 + ], + "angle": 0, + "content": "[46] Cho-Ying Wu, Qiangeng Xu, and Ulrich Neumann. Synergy between 3dmm and 3d landmarks for accurate 3d facial geometry. In 2021 International Conference on 3D Vision (3DV), 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.604, + 0.892, + 0.671 + ], + "angle": 0, + "content": "[47] Yiqian Wu, Jing Zhang, Hongbo Fu, and Xiaogang Jin. Lpff: A portrait dataset for face generators across large poses. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 20327-20337, 2023. 1, 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.675, + 0.892, + 0.743 + ], + "angle": 0, + "content": "[48] Liangbin Xie, Xintao Wang, Honglun Zhang, Chao Dong, and Ying Shan. Vfhq: A high-quality dataset and benchmark for video face super-resolution. In The IEEE Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), 2022. 1, 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.746, + 0.892, + 0.801 + ], + "angle": 0, + "content": "[49] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3836-3847, 2023. 1, 2, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.803, + 0.892, + 0.871 + ], + "angle": 0, + "content": "[50] Zhimeng Zhang, Lincheng Li, Yu Ding, and Changjie Fan. Flow-guided one-shot talking face generation with a high-resolution audio-visual dataset. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3661–3670, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.873, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[51] Jian Zhao and Hui Zhang. Thin-plate spline motion model for image animation. In Proceedings of the IEEE/CVF Con" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "22614" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.093, + 0.468, + 0.119 + ], + "angle": 0, + "content": "ference on Computer Vision and Pattern Recognition, pages 3657-3666, 2022. 2, 7, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.122, + 0.469, + 0.177 + ], + "angle": 0, + "content": "[52] Tianyue Zheng and Weihong Deng. Cross-pose lfw: A database for studying cross-posed face recognition in unconstrained environments. Beijing University of Posts and Telecommunications, Tech. Rep, 5(7), 2018. 3, 4, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.179, + 0.469, + 0.219 + ], + "angle": 0, + "content": "[53] Huayi Zhou, Fei Jiang, and Hongtao Lu. Directmhp: Direct 2d multi-person head pose estimation with full-range angles. arXiv preprint arXiv:2302.01110, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.221, + 0.469, + 0.275 + ], + "angle": 0, + "content": "[54] Hao Zhu, Wayne Wu, Wentao Zhu, Liming Jiang, Siwei Tang, Li Zhang, Ziwei Liu, and Chen Change Loy. CelebV-HQ: A large-scale video facial attributes dataset. In ECCV, 2022. 1, 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.278, + 0.469, + 0.36 + ], + "angle": 0, + "content": "[55] Zheng Zhu, Guan Huang, Jiankang Deng, Yun Ye, Junjie Huang, Xinze Chen, Jiagang Zhu, Tian Yang, Jiwen Lu, Dalong Du, and Jie Zhou. Webface260m: A benchmark unveiling the power of million-scale deep face recognition. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 1, 7" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.093, + 0.469, + 0.36 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.518, + 0.957 + ], + "angle": 0, + "content": "22615" + } + ] +] \ No newline at end of file diff --git a/2024/EFHQ_ Multi-purpose ExtremePose-Face-HQ dataset/65a10ed9-cef6-42e8-a141-8586b125aea4_origin.pdf b/2024/EFHQ_ Multi-purpose ExtremePose-Face-HQ dataset/65a10ed9-cef6-42e8-a141-8586b125aea4_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..28c5e216b826da537f13a643f951727ea7fe6961 --- /dev/null +++ b/2024/EFHQ_ Multi-purpose ExtremePose-Face-HQ dataset/65a10ed9-cef6-42e8-a141-8586b125aea4_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:be3409fea9659bb691923937f02b2e8710da69be27354a78c9dae72d89611b27 +size 3811407 diff --git a/2024/EFHQ_ Multi-purpose ExtremePose-Face-HQ dataset/full.md b/2024/EFHQ_ Multi-purpose ExtremePose-Face-HQ dataset/full.md new file mode 100644 index 0000000000000000000000000000000000000000..96754beb73f5408e9948a8c9f8d7c7663136fa99 --- /dev/null +++ b/2024/EFHQ_ Multi-purpose ExtremePose-Face-HQ dataset/full.md @@ -0,0 +1,301 @@ +# EFHQ: Multi-purpose ExtremePose-Face-HQ dataset + +Trung Tuan Dao $^{1,*}$ Duc Hong Vu $^{1,2,*}$ Cuong Pham $^{1,3}$ Anh Tran $^{1}$ + +$^{1}$ VinAI Research, Vietnam $^{2}$ Miami University $^{3}$ Posts & Telecommunications Inst. of Tech., Vietnam + +{v.trungdt21,v.ducvh5,v.cuongpv11,v.anhtt152}@vinai.io vudh3@miamioh.edu cuongpv@ptit.edu.vn + +*Equal contribution + +![](images/cbf16de748627893751e10fb88a1c1970cd6edd74202d56555bea8261d2edfef.jpg) +Figure 1. Benefits of our proposed dataset (EFHQ). Standard large-scale facial datasets have most images at near frontal views, causing inferior performance of trained models on downstream tasks when dealing with extreme head poses. For instance, the trained 2D image generators and text-to-image ones often produce only near frontal faces, while the 3D face generators and face reenactment methods often show distorted outputs at profile views. The recently proposed dataset LPFF [47] partially handles that issue by providing complementary images at extreme head poses for only 2D and 3D image generation tasks. Our proposed dataset EFHQ provides high-quality extreme-pose images to complement a wide range of face-related tasks. It supports 2D and 3D image generation, with generally better diversity than LPFF. EFHQ also helps correct the outputs of text-to-image generation and face reenactment at extreme views. Finally, EFHQ provides a more challenging pose-based face verification benchmark to better assess the quality of face recognition networks. + +# Abstract + +The existing facial datasets, while having plentiful images at near frontal views, lack images with extreme head poses, leading to the downgraded performance of deep learning models when dealing with profile or pitched faces. This work aims to address this gap by introducing a novel dataset named Extreme Pose Face High-Quality Dataset (EFHQ), which includes a maximum of 450k high-quality images of faces at extreme poses. To produce such a massive dataset, we utilize a novel and meticulous dataset processing pipeline to curate two publicly available datasets, VFHQ and CelebV-HQ, which contain many high-resolution face videos captured in various settings. Our dataset can complement existing datasets on various facial-related tasks, such as facial synthesis with 2D/3D-aware GAN, diffusion-based text-to-image face generation, and face reenactment. Specifically, training with EFHQ helps models generalize well across diverse poses, significantly improving performance in scenarios involving ex + +treme views, confirmed by extensive experiments. Additionally, we utilize EFHQ to define a challenging cross-view face verification benchmark, in which the performance of SOTA face recognition models drops $5 - 37\%$ compared to frontal-to-frontal scenarios, aiming to stimulate studies on face recognition under severe pose conditions in the wild. + +# 1. Introduction + +The human face is a central focus in computer vision, with extensive research dedicated to tasks like detection, recognition, generation, and manipulation [39, 42]. Numerous large-scale facial datasets[9, 12, 14, 24, 28, 30, 45, 48, 54, 55] have facilitated the face-related studies. Along with deep learning techniques, many tasks have made a considerable leap in performance in recent years. For instance, face recognition systems can recognize near-frontal faces in the wild with near-perfect accuracy [13, 18]. Facial generative models, such as [5, 15, 49], can synthesize realistic + +images indistinguishable from real faces. Face reenactment systems, including [41, 43, 51], can animate an input face based on poses and expressions from any driver video. + +Despite impressive results, the mentioned models produce inferior results when dealing with faces at extreme head poses. As illustrated in Fig. 1, faces synthesized by 3D face generation or face reenactment models are often distorted at profile views, due to the scarcity of extreme pose images in training datasets. Current extensive datasets for face generation leverage portrait images from Google or stock photo repositories like Flickr, which feature images with upfront views and minimal facial rotations. Face reenactment methods leverage video datasets such as Vox-Celeb [29], primarily featuring frames with frontal faces. The instability of pre-processing tools, particularly face alignment modules when handling extreme pose images, further reduces the proportion of those images in the existing datasets. This issue is critical since faces with large head rotations are common in real life, and dealing with them is unavoidable in practical systems. + +This paper aims to fill that void by proposing a novel large-scale dataset of extreme-pose and high-quality facial images called ExtremePose-Face-HQ, or EFHQ for short. Our ambition is to build an "one-for-all" dataset that can supplement existing datasets on all mentioned face-related tasks. To achieve that goal, the dataset needs all the following properties: (1) extreme-pose, (2) large-scale, (3) high-quality, (4) in-the-wild, (5) having multiple images for each subject and having subject identity annotated. Given those requirements, we utilize a novel and meticulous dataset processing pipeline to filter and sample frames from two publicly available datasets, including VFHQ [48] and CelebV-HQ [54], which contain high-resolution face videos captured in various settings. This results in a dataset with 450k high-quality images of extreme poses, which are combined with frontal images of the same subject to use as image pairs. These image pairs can support face reenactment and face recognition tasks, while the extreme-pose images alone can complement the existing datasets of face generation training. Note that there is a recent attempt [47] to build a large-pose facial dataset similar to ours. However, that work relies on images crawled from Flickr, thus having limited size (19k images) and missing identity information. Hence, its applications are restricted, covering only several face-generation tasks, unlike our large-scale and multi-purpose dataset. + +We perform a comprehensive set of experiments to validate the effectiveness of EFHQ across three distinct face generation sub-tasks and face reenactment scenarios. Our dataset, EFHQ, when combined with standard training sets, significantly enhances the quality of synthesized images in highly profile or pitched views while maintaining synthesis quality in frontal views. EFHQ also provides a more + +challenging pose-centric face verification benchmark. It reveals the vulnerability of many state-of-the-art face recognition networks when dealing with extreme-pose images, with TAR@FAR=1e-3 scores dropping 5-37%. + +In summary, our contributions include (1) a large-scale, high-quality, in-the-wild, extreme-posed dataset (EFHQ) designed to complement various face-related tasks, (2) a novel and meticulous dataset processing pipeline ensuring the creation and quality of the EFHQ dataset, and (3) a series of experiments, with plan-to-release models, demonstrating the usefulness of EFHQ in enhancing performance on a wide range of tasks when dealing with faces at extreme rotations. The final dataset will be publicly available as metadata files to not violate the two original dataset's licenses, alongside the pretrained networks. + +# 2. Related Work + +Synthetic Face Generation. Existing 2D-based models like StyleGAN [14, 15, 17] and 3D-aware models like EG3D [5] demonstrate impressive fidelity, especially for near-frontal views. Nevertheless, their performance significantly degrades when faced with extreme views, primarily due to insufficient training data for such scenarios. Datasets like FFHQ [14] exhibit a notable bias toward near-frontal poses, heavily limiting the range of poses synthesized by current 2D and 3D models despite the 3D-aware generators' ability to handle arbitrary poses. + +Diffusion-based text-to-image models [32, 33] can synthesize high-quality images from simple text inputs. However, these models lack mechanisms to control the generated pose. Consequently, the generated facial images remain in a frontal view despite terms like "profile" or "looking-up" in the input prompts. ControlNet[49] proposes a finetuning approach allowing these models to take additional context as conditional input. Namely, one can use landmarks to guide the generated face to the desired pose. However, the scarcity of profile-view data in training affects the model's performance, particularly in extreme angles. Consequently, effectively managing the long-tailed distribution of these extreme views becomes necessary to address this issue. + +Face Reenactment. Face reenactment involves transferring facial expressions and movements from a driving image to a source image. Recent advancements claim high-quality results, especially for near-frontal poses [31, 41, 43, 51]. However, pose collapse persists, particularly in keypoint-based models [41, 51]. Common datasets for this task are VoxCeleb1 [29] and HDTF [50], which primarily feature frontal talking face videos. Despite recent potentially high-quality datasets for reenactment such as VFHQ [48] and CelebV-HQ [54], the frontal-view bias persists. As a result, facial reenactment methods often produce artifacts and lower-quality outputs when faced with significant rotations. + +
DatasetAttributesReady-to-use tasks
#imagesMultiple images per IDHigh QualityProfile ViewFace VerificationFace SynthesisFace Reenactment
AgeDB12K✓✓
CPLFW12K✓✓✓
CFP-FP14K✓✓✓
IJB-x138K✓✓✓
CelebA-HQ30K✓✓
FFHQ70K✓✓
LPFF19K✓✓✓
CelebV-HQ6M✓✓✓✓
VFHQ3M✓✓✓✓✓
VoxCeleb14M✓✓✓
Ours450,538✓✓✓✓✓✓
FFHQ+Ours*112,671✓✓✓
Vox+Ours*4.5M✓✓✓
+ +Table 1. Representative comparison of key attributes and supported tasks across existing face datasets and our proposed EFHQ dataset. EFHQ provides large-scale, high-quality images with multiple views per identity, enabling diverse face-related tasks, including synthesis, reenactment, and verification. Combining EFHQ with existing datasets like FFHQ or VoxCeleb can augment available data for selected tasks. Asterisk (*) indicates a combined dataset with a subsampled version of EFHQ. + +Face Verification. Despite near-perfect accuracy on benchmark datasets[12, 20, 27, 28, 44], state-of-the-art face verification models[7, 13, 18] fall short in unconstrained environments due to a lack of pose diversity. While CPLFW[52] introduces 6000 cross-pose pairs, its small size and mix of pose samples hinder analyzing performance on specific pose scenarios such as frontal-frontal, frontal-profile, and profile-profile. More granular benchmarking on diverse pose pairs is crucial to identify failure modes and drive progress in robust face verification across poses. + +Extreme Pose Face Dataset. Tab. 1 summarizes existing datasets related to face verification, synthesis, and re-enactment, showing that existing face datasets have limitations for multi-task use. Recent efforts like LPFF[47] made progress by crawling from Flickr and processing for diverse views, resulting in a dataset of 19,590 images. The dataset could complement high-quality datasets like FFHQ[14] for generative models. However, due to its limited scale comprising only $20\%$ of the combined dataset, the unbalanced distribution between frontal and profile view still needs further improvements. Additionally, since the images are from image hosting services, the dataset fails to provide the identity information necessary for other facial applications. + +# 3. Dataset Process + +In this section, we discuss the process of producing our novel dataset EFHQ. To fulfill our goals, we carefully curated frames from the two recent facial video datasets, VFHQ[48] and CelebV-HQ[54], as seen in Fig. 2. + +# 3.1. Attributes Preparation + +Attributes Extraction. To curate accurate facial poses and high-quality frames from existing datasets, we first ex + +tract (1) face bounding boxes, (2) facial landmarks (5/68 keypoints), (3) image quality score, and (4) face identity. For bounding boxes and landmarks, we employ the popular RetinaFace[36] and SynergyNet[46]. We use HyperIQA [40] to grade image quality. For VFHQ, we only generate 68 landmarks and image quality scores since bounding boxes and 5 keypoints are provided. After extracting these attributes, we match them with existing annotations using an IoU-based Hungarian matching algorithm [21]. Regarding identity, VFHQ-wise, we source labels from the annotations. As CelebV-HQ lacks individual identity labels within the video, we use[35] to identify the identities associated with each bounding box. + +Due to the instability of head pose estimators when handling extreme-pose images, we strategically apply multiple pose estimators relying on various methodologies: (1) SynergyNet[46], a 3DMM-based model, (2) DirectMHP[53], a landmark-free joint head detector and pose estimator, and (3) an in-house FacePoseNet[6] improved through extensive data augmentation focused on extreme poses and extra training data. We then categorize each estimated pose into a hierarchical binning scheme illustrated in Fig. 2. Next, we gather bin predictions and perform majority voting to arrive at consensus labels. When no agreement emerges, we put it into the $5^{th}$ bin for "confusing" cases. As depicted in Fig. 3, this ensemble approach provides robustness in cases where a single estimator incorrectly categorizes a sample. Please refer to the supplementary for statistics of the extracted bins and the hyperparameters for each model used for attribute extraction. + +Annotation Review. Next, we manually review the results to ensure high-quality labels. Given the large dataset, we developed a graphical user interface tool to streamline the review process. Our focus is verifying the binning annota + +![](images/f254cce2dcbe27b3334c49712a16370a9280f177f1bfe4f41e39c028a6f522d5.jpg) +Figure 2. The pipeline of EFHQ dataset creation. Starting with high-quality videos from the VFHQ[48] and CelebV-HQ[54] datasets, single-frame attributes are extracted then manually reviews. Task-specific preprocessing is then applied to generate specialized versions of the dataset for tasks such as face generation, reenactment, and verification. + +![](images/38c77dbb4c806449587b476c0b22ce3f7a86ea80b9d0b108ab1e63be1859d8d5.jpg) +Figure 3. Example cases where a pose estimator fails to categorize the sample to the correct bin. + +tions, which most directly impact label accuracy. Concurrently, we randomly subsample the data based on pose angle and image quality to validate facial landmark quality further and discard cases of landmark prediction failure. + +# 3.2. EFHQ dataset + +Our final EFHQ dataset comprises up to 450k frames with extreme poses extracted from approximately 5,000 clips. Most of the clips in our dataset include at least one frame with a frontal face and multiple frames with extreme pose angles. We also include extreme pose-only clips, representing profile-to-profile pose transfer cases. + +![](images/caec04bb3476c53e27cfcf754fc7da82c83f7eb861fc3f48a721d325f51ab89b.jpg) +Figure 4. Pose distribution comparison between our sampled dataset and other datasets, including FFHQ[14], LPFF[47], CPLFW[52] and 40K random samples from VoxCeleb1 [29]. Our sampled dataset demonstrates greater pose diversity, with increased sample counts across high angle bins. + +# 3.3. Dataset for subtasks + +Supplementary datasets for face generation. To address FFHQ's pose distribution gap, we compile a dataset encompassing diverse poses, varied identities, and FFHQ-comparable image quality from EFHQ. Additionally, we also embed an image brightness filter[3] in the sampling process, given that HyperIQA doesn't explicitly assess this aspect. The sampled images are then processed using the same pipeline in [5, 16] and manually reviewed. The aim is to improve coverage of under-represented poses without degrading frontal performance of trained generation models. This yields a dataset of 42,671 images with equitably distributed poses as visualized in Fig. 4. Compared to the original FFHQ distribution, yaw and pitch distributions are significantly improved. + +Additionally, for 3D-aware GAN subtasks, we extract camera parameters to serve as conditional pose information, following [5]. We convert these parameters into yaw, pitch angles to validate against our existing annotations. When disagreements emerge, we re-examine the images and filter if needed, to ensure accurate, quality labels. + +For diffusion-based text-to-image generation, a specialized dataset is curated to refine Stable Diffusion models[33] by accommodating landmark conditional input through ControlNet[49]. Based on our 3D-aware face generation dataset, we integrate landmark-based conditional images and tailored text prompts for each image. To capture fine facial nuances, we extract 478 facial landmarks and draw the condition image with connected edges between matched keypoints, following Mediapipe framework[26], better-conveying face expression and pose. For text prompts, detailed facial attributes like gender, race, and emotion are derived from existing labels or inferred via BLIP-2 pretrained captioning model [23] and a face attribute estimator [37]. The resulting prompt adheres to a structured format: "A profile portrait image of a [emotion] [race] [gender]." + +
ModelReference DatasetFID↓Recall↑
GFFHQFFHQ2.840.49
GFFHQ+LPFFFFHQ+LPFF3.430.44
GFFHQ+EFHQFFHQ+EFHQ3.440.46
GFFHQ+LPFFFFHQ+EFHQ8.260.43
GFFHQ+EFHQFFHQ+LPFF7.040.44
GFFHQ+EFHQFFHQ+EFHQ3.330.46
GFFHQ+EFHQFFHQ3.120.44
GFFHQ+EFHQFFHQ+LPFF4.470.42
+ +Table 2. StyleGAN2-ADA models trained on different datasets. Lower FIDs indicate better fidelity, while higher Recalls indicate better diversity. The second and third blocks show comparisons between unconditional models trained on FFHQ+LPFF and FFHQ+EFHQ when evaluating on the same and cross-dataset settings, with the better metrics in bold. + +Supplementary dataset for face reenactment. As EFHQ is crafted, we can employ the entire dataset to complement established training datasets, e.g., VoxCeleb1 [29]. We also built an extra evaluation set with 1200 EFHQ clips. + +Benchmarking dataset for face verification. We curate a dataset covering three distinct scenarios: frontal-to-frontal, frontal-to-profile, and profile-to-profile. To enable rigorous benchmarking, we sample 10,000 pairs each for both negative and positive cases per scenario, resulting in a balanced benchmark dataset containing 60,000 image pairs. Images follow the same established preprocessing pipelines in [7]. Next, we filter misaligned images with RetinaFace[36], randomly review, and replace equivalent samples from our diverse corpus, if needed. Finally, to simulate varying image quality, we randomly applied downscaling and compression to samples in the dataset. + +# 4. Face Generation Subtask + +# 4.1. StyleGAN + +Training details. We train StyleGAN2-ADA[15] models from scratch after combining FFHQ with our own curated dataset (denoted FFHQ+EFHQ) at $1024 \times 1024$ resolution. We also train a conditional version, allowing direct control over frontal versus profile synthesis and enhancing model performance. Specifically, we categorize the dataset into two classes: "frontal" and "profile" based on pose angles. All models are trained for 6 days on 8 Nvidia A100 GPUs. + +Evaluation protocol. We utilize the Frechet Inception Distance[11] (FID) metric and improved Recall[22] to quantify the generation fidelity and diversity. Regarding the two unconditional models trained with the supplemented FFHQ+LPFF and FFHQ+EFHQ, we evaluate each model with both of these datasets. For conditional mod + +![](images/3dcea56d5362a57eb679bd30469ce22064230f4b8d123e344efe25aaf2dc37ea.jpg) +(a) Samples from the model trained with FFHQ+EFHQ. + +![](images/3c0d5c9941056f9827add44aae19a2e757bb2b54eaf548fa151058eafdd04f75.jpg) +(b) Profile-view comparison of LPFF (left) and EFHQ (right). +Figure 5. Qualitative result and comparison of generated samples, with truncation $\psi = 0.7$ , from StyleGAN2-ADA training with other dataset and ours. + +els, we evaluate performance on FFHQ, FFHQ+LPFF, and FFHQ+EFHQ to demonstrate the broader coverage of the model trained with our proposed dataset. For FFHQ, we generate 50K images with the "frontal" label, whereas for the other two datasets, we generate 50K following their label distributions. This comprehensive quantitative analysis examines the fidelity and variety of the generated images. + +Experimental results. We compared the performance of our proposed unconditional models $(G^{\mathrm{FFHQ + EFHQ}})$ , against models trained on the FFHQ and FFHQ+LPFF datasets $(G^{\mathrm{FFHQ}}$ and $G^{\mathrm{FFHQ + LPFF}})$ . Models trained conditionally are specifically subscripted, e.g., $G_{c}^{\mathrm{FFHQ + EFHQ}}$ . The qualitative results in Fig. 5 illustrate that our models generate high-quality frontal faces comparable to the FFHQ model while producing realistic and varied profile faces. Additionally, our method has less noise in extreme poses than the FFHQ+LPFF model. Specifically, Fig. 5b shows LPFF's generated face contains more noise and less photorealistic details than our model, such as the pixelated eye and nose patches. Meanwhile, ours achieves greater realism with smoother skin, well-defined features, noise reduction, and natural lighting. The quantitative results in Tab. 2 show that our unconditional model achieves competitive FID scores and improved Recall to other models, indicating similar fidelity but greater sample pose diversity. In the cross-examination (third block of Tab. 2), $G^{\mathrm{FFHQ + EFHQ}}$ has both better FID and improved Recall on the opposite dataset than $G^{\mathrm{FFHQ + LPFF}}$ . Finally, our conditional model demonstrates substantially improved FID scores over multiple reference datasets, highlighting its versatility. + +
ModelReference DatasetFID↓Pose↓Depth↓ID↑
EFHQFFHQ/Trainset4.85e-40.2340.74
EFFHQFFHQ/Trainset3.75e-40.2880.74
EFFHQ+LPFFFFHQ6.26e-40.3160.77
EFFHQ+EFHQFFHQ3.94e-40.2240.77
EFFHQ+LPFFTrainset7.06e-40.3160.65
EFFHQ+EFHQTrainset4.64e-40.2230.65
+ +![](images/8e70a13b656bbde69089723c284a0af97e6cca6c1c51f5d55a61b3fe1f63ec4d.jpg) +Figure 6. Comparison between multiview generated samples, with truncation $\psi = 0.8$ , of EG3D model trained with various datasets. + +# 4.2. EG3D + +Training details. We followed the same pipeline and hyper-parameters as the original EG3D model [5]. However, based on empirical results, we adjusted $\gamma = 5$ . We also fixed the triplane implementation bug (XY, XZ, ZX) [1] to replicate the original idea in the papers (XY, XZ, ZY). We trained the model on 8 Nvidia A100 40GB GPUs for approximately 7 days until convergence. + +Evaluation Protocol. To evaluate the models, mirroring [5], we measure image quality with FID; multi-view consistency and identity preservation with mean cosine similarity of pre-trained face recognition model[7]; pose accuracy and geometry by calculating MSE against pseudo labels from [8]. We evaluate all models on 1,024 generated images, except for FID, which we compute on 50k images. + +Experimental results. To evaluate the benefits of our dataset, we trained the same backbone architecture on different datasets: FFHQ $(E^{\mathrm{FFHQ}})$ , FFHQ+LPFF $(E^{\mathrm{FFHQ} + \mathrm{LPFF}})$ , and FFHQ+EFHQ (our model, $E^{\mathrm{FFHQ} + \mathrm{EFHQ}}$ ). The quantitative results are presented in Table 3, and qualitative examples are shown in Figure 6. Qualitatively, we observe that $E^{\mathrm{FFHQ} + \mathrm{EFHQ}}$ generates improved face shapes in extreme poses compared to $E^{\mathrm{FFHQ}}$ , similar to the improvements seen with $E^{\mathrm{FFHQ} + \mathrm{LPFF}}$ . Quantitatively, the FID + +Table 3. EG3D models trained on different datasets. In the second and third blocks, we compare models trained on FFHQ+LPFF and FFHQ+EFHQ with the testset is either FFHQ or their training datasets; the better metrics are in bold. The best metrics on the FFHQ testset are in red, while the best metrics when evaluating on training dataset are underlined. + +
MetricsDatasetPretrained ControlNetOurs
NME↓LPFF0.0850.084
FID↓LPFF119.7483.39
FIDeyepatch ↓LPFF103.7749.45
NME↓EFHQ0.1060.081
FID↓EFHQ111.2381.90
FIDeyepatch ↓EFHQ137.4358.60
+ +Table 4. Quantitative results of fine-tuning StableDiffusion v1.5 with ControlNet trained on various datasets, using Normalized Mean Error (NME), FID, and FID on eye region. + +of $E^{\mathrm{FFHQ + EFHQ}}$ is comparable to $E^{\mathrm{FFHQ}}$ when evaluated on FFHQ, indicating no degradation in frontal view performance. Meanwhile, the competitive FID achieved by $E^{\mathrm{FFHQ + EFHQ}}$ on the combined FFHQ+EFHQ dataset also suggests proficient modeling of profile views. Regarding identity consistency, we see a significant drop when the reference dataset contains profile-view images, due to the weak confidence of the pretrained face recognition model when handling profile-view images, as confirmed by its low average cosine similarity of around 0.6 on the simple frontal-to-profile verification set CFP-FP[34]. Furthermore, our model also outperforms in pose accuracy and geometry quality with both frontal and cross-pose dataset. + +![](images/91c1c39d34e1ff469ab19168be92e10d78b69610e53bee1af6b1b4b5daca6041.jpg) +(a) Generated samples; condition image is in between samples. + +![](images/e2849aa6187df629ee39998db75c92f3603e153a68b749d5b56f72afd4fcceeb.jpg) + +![](images/ea5572592c76c179b7a23caa798b0464f0d77a81bf0014c7a81b2a1c2f7a141b.jpg) +(b) Generated eye regions. + +![](images/42ea89334a93b892c5276bce82ddb8d2426f524d79dcfa74a222ced094510ff8.jpg) +Figure 7. Comparison profile-view generated samples of pretrained ControlNet (left) and our fine-tuned ControlNet (right) with the prompt: "A profile portrait image of a person". + +# 4.3. ControlNet + +Training details. We finetuned Stable Diffusion v1.5 [33] using ControlNet [49] on the EFHQ dataset to enhance its ability to generate extreme facial poses. We use AdamW optimizer [25] with learning rate $1 \times e^{-5}$ , $\beta_0 = 0.9$ , $\beta_1 = 0.999$ , and $\epsilon = 0.01$ . We train with batch size 24 on 1 Nvidia A100 40GB GPU in 2 days. + +Evaluation Protocol. To assess image qualities across diverse poses, we measure the Normalized Mean Error (NME) of 478-point facial landmarks between the generated and source image. Specifically, we evaluate performance across the full pose distribution in the reference + +
ModelTest Dataset (L1 ↓/ AKD ↓/ AED ↓)
VoxCeleb1EFHQ (Ours)
TPS(VoxCeleb10.041/1.214/0.1320.061/3.814/0.275
TPS(VoxCeleb1+EFHQ0.041/1.250/0.1330.051/1.963/0.214
LIA(VoxCeleb10.044/1.342/0.1380.059/3.807/0.266
LIA(VoxCeleb1+EFHQ0.044/1.398/0.1420.054/2.291/0.233
+ +dataset. For the prompt, we use: "A profile portrait image of a person." with extra positive and negative keywords, further detailed in supplementary. Utilizing LPFF and EFHQ datasets as the reference datasets, we compute FID to assess synthesized face quality and extend evaluation to cropped eye regions due to unrealistic gazes observed from the original ControlNet model. Note that the models receive condition images created akin to its training data. + +Experimental results. Fig. 7a visually compares the baseline ControlNet OpenPose face and our fine-tuned model. Our fine-tuned model adeptly conditions diverse extreme poses, generating high-quality results with precise details while maintaining fidelity to the condition, such as mouth shape, and exhibiting minimal artifacts. These patterns indicate the superiority of the fine-tuned model over the baseline model in handling extreme viewing angles. Additionally, Fig. 7b contrasts the eye regions of both models. Our fine-tuned model accurately captures natural eye gaze and shape from condition images, a notable improvement over the baseline. Quantitative results in Tab. 4 consistently favor our fine-tuned model across all metrics. FID scores on full and eye patches exhibit significant enhancement, and NME performance remains stable on training and reference datasets (EFHQ and LPFF). These findings showcase the fine-tuned model's capability to generate high-quality, geometrically consistent faces across diverse poses. + +# 5. Face Reenactment Subtask + +# 5.1. Evaluation Protocol + +We follow the same preprocessing and evaluation protocol as FOMM [38]. Specifically, all the frames are cropped to $256 \times 256$ , and we use the first frame of a clip as the source image to reconstruct the remaining frames. In the case of the EFHQ test set, most source images have a frontal pose, while the target frames have an extreme pose. Some test cases only contain extreme pose to evaluate self-reconstruction on these cases. We run the experiments using two recent methods, Thin-Plate Spline Motion Model (TPS) [51] and Latent Image Animator (LIA) [43]. For training details, please refer to the supplementary. We select the following three metrics to compare the generated frames and the driving ground truth. $\mathcal{L}_1$ measures the av + +Table 5. Quantitative results of two state-of-the-art reenactment methods (TPS and LIA) when training on VoxCeleb1 and the combined set VoxCeleb1+EFHQ. We present results evaluated either on VoxCeleb1 (left) or EFHQ (right) test sets. + +
ModelTrainsetCPLFWEFHQ
f2ff2pp2p
R18arfaceGlint360k0.800.940.570.370.880.06
R18arfaceMS1MV30.750.930.570.360.880.05
R50arfaceGlint360k0.910.970.770.200.920.05
R100arfaceGlint360k0.930.980.830.150.820.16
R100adafaceWebFace4M0.940.990.940.050.970.02
+ +Table 6. Quantitative evaluation of state-of-the-art face recognition models on the cleansed CPLFW and the proposed benchmark dataset. We use the metric TAR@FAR=1e-3 (higher is better). Our EFHQ benchmark includes frontal-frontal (f2f), frontal-profile (f2p), and profile-profile (p2p) subsets. For f2p and p2p, we report the score drop when comparing to f2f in the red subscription. The lowest benchmark score for each model is in bold. + +erage distance among pixel values, indicating reconstruction faithfulness. Average Keypoint Distance (AKD) calculates the mean distance between keypoints, thus assessing the pose accuracy. Average Euclidean Distance (AED) measures identity preservation by computing the average $\mathcal{L}_2$ distance between identity embeddings. + +# 5.2. Experimental results + +We compare TPS and LIA models trained with our supplementary EFHQ dataset $(TPS^{\mathrm{VoxCeleb1 + EFHQ}}$ and $LIA^{\mathrm{VoxCeleb1 + EFHQ}})$ against models trained solely on VoxCeleb1 $(TPS^{\mathrm{VoxCeleb1}}$ and $LIA^{\mathrm{VoxCeleb1}})$ . Tab. 5 displays quantitative results, revealing that models trained with EFHQ maintain comparable performance when evaluated on VoxCeleb1, emphasizing high-quality performance in frontal-to-frontal reenactment scenarios. Notably, on the EFHQ test set, both EFHQ-trained models show significant improvement in all metrics, indicating enhanced extreme pose and identity reenactment performance. In Fig. 8, qualitative results prove that EFHQ-trained models excel in transferring facial expressions and motion across diverse poses. Models trained on EFHQ showcase improved image quality, reduced artifacts, and well-preserved shapes of driving faces. These results highlight our models' effectiveness in capturing and reproducing facial motions across varying poses while enhancing fidelity to the original facial features. + +# 6. Face Recognition Subtask + +# 6.1. Evaluation Protocol + +To evaluate face recognition performance on extreme poses, we selected two state-of-the-art methods: ArcFace[7] and AdaFace[18]. We further explored various backbone sizes (ResNet18, ResNet50, ResNet101)[10] and training datasets (MS1MV2[9], Glint360K[2], WebFace4M[55]) of increasing size. For metrics, we used True Acceptance Rate (TAR) at False Acceptance Rate (FAR) of $1 \times e^{-3}$ . + +![](images/2c8117e4d7e86d126609d5accefe90974ed636e97faaf9d4022551304474fbb7.jpg) + +![](images/f69d710555708f6551cef3b5ce48de1fde2a208e53247211237ab7c6fbae5fcb.jpg) +Source + +![](images/7d2912606c5b61da5c606f88306fd6e060536c5ef9c314d5d5374ab99e85b6c3.jpg) + +![](images/61c559132cc4890d912246b645d7129a8dff94b65997700f60ef27bdef7545eb.jpg) +Figure 8. Comparison of performance for frontal-profile face reenactment on EFHQ. From left to right: source image, TPS on VoxCeleb1, TPS with EFHQ, LIA on VoxCeleb1, LIA with EFHQ, driving image. Additional EFHQ training data improves synthesis of extreme poses. + +![](images/21f7cb68aa95c4ae50e04e75b16ffcc4a864c7551ef38fd4b4973bd55631b879.jpg) + +![](images/74afe71aafafe1944aae621c88642f991e2738f3e52b03082731ebdf1d978f1f.jpg) +TPS (VoxCeleb1) +TPS (Ours) + +![](images/2eb2f40499fafbf4fcb78e429ecb6a3f128bea37332e4a23ac5950ae3a760e86.jpg) + +![](images/13c7f831840d7ea36cbc2037bba4722099980fd3cb1850d9ce250d64185c67f0.jpg) +LIA (VoxCeleb1) + +![](images/6d928c801997e6e85bd8d4dd1dab3f7ab8b88397fe154b822f5f89eac285434f.jpg) + +![](images/a37e832c4e7938ac2972ac1b7a72c3542f56a1a39355e19829977df27ba063d8.jpg) +LIA (Ours) + +![](images/cf4b1d5c92581dc96cd7ada2ad8e39c5f6b911a5e50861f172bbd3c22fc7a3d0.jpg) + +![](images/227ecfb2eb75f91aa3aa305d5cb140c2676776284e15ef4c1a063540403f72ea.jpg) +Driving + +# 6.2. Experimental results + +Before assessing face recognition networks on EFHQ, we first clean the commonly-used version of CPLFW, processed by[7], by removing misaligned images (please refer to the Supplementary), and evaluate the models on it. Models performed well on this cleansed set, with minimal gaps between ArcFace and AdaFace. + +Our proposed evaluation benchmark involves three pose-based scenarios. Results in Tab. 6 reveal performance drops of $5 - 37\%$ on the frontal-profile subset compared to frontal-frontal, highlighting the challenge of matching faces in different poses. Larger ResNet models outperformed ResNet18 across subsets, showcasing the advantages of increased capacity in handling pose variation. However, the ResNet100 model trained on Glint360K performed worse than smaller models on the profile-profile subset, indicating the need for sufficient profile data to prevent overfitting. Our results also indicate AdaFace's suitability for handling pose variations. In summary, pose mismatch brings significant challenges for face recognition, necessitating larger models trained on diverse data covering profile faces. + +# 7. Discussion + +![](images/e17c16d7e1823307bbc687e505298164fa1f4d5649b691e9a0108114e17bb041.jpg) +Figure 9. User study results. We reported the percentage of times testers rank whose output is best or tied for each criterion. + +User survey. We conduct user surveys to evaluate the face generation improvement of our method through human per + +ceptual assessment. In this survey, we compare the outputs of the models mentioned in Secs. 4 and 5. The setup of this user survey is included in the supplementary. In all human evaluation results (Fig. 9), our models outperform previous ones, notably on identity and motion test of face reenactment models (74.80% and 64.22%, respectively). + +![](images/5406b417968dda5059ac39836df61475119ae9f04474919d0df72f427ee2ccb5.jpg) +Figure 10. Representative examples of the raw LPFF dataset that got excluded from the final dataset, either from misdetections or pose filtering process. Red keypoints represent landmarks prediction of selected samples from LPFF, green keypoints represent landmarks prediction from our pipeline. + +Robustness of data processing pipeline. As shown in Fig. 10, our data processing pipeline can detect additional profile-view faces filtered out by the LPFF pipeline, likely due to misdetections or incorrect pose filtering. This confirms the robustness of our pipeline, and we hope it will enable the community to develop larger and higher-quality in-the-wild facial datasets. + +# 8. Conclusion + +This work has introduced a large-scale, diverse facial dataset to address performance gaps between frontal and profile faces, powered by a novel and robust data processing pipeline. Crucially, we provided tailored sub-datasets for advancing essential tasks like face synthesis and reenactment. Moreover, our new face verification benchmark reveals gaps between techniques, granting valuable insights. Ultimately, we hope this high-quality, diverse facial dataset opens up new and exciting opportunities to push forward cross-pose tasks. + +# References + +[1] Fix for triplane projection $\cdot$ Issue #67 $\cdot$ NVlabs/eg3d. https://github.com/NVlabs/eg3d/issues/67, 2023.6 +[2] Xiang An, Xuhan Zhu, Yuan Gao, Yang Xiao, Yongle Zhao, Ziyong Feng, Lan Wu, Bin Qin, Ming Zhang, Debing Zhang, et al. Partial fc: Training 10 million identities on a single machine. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1445-1449, 2021. 7 +[3] Sergey Bezryadin, Pavel Bourov, and Dmitry Ilinih. Brightness calculation in digital image processing. International Symposium on Technologies for Digital Photo Fulfillment, 2007:10-15, 2007. 4, 1 +[4] Z. Cao, G. Hidalgo Martinez, T. Simon, S. Wei, and Y. A. Sheikh. Openpose: Realtime multi-person 2d pose estimation using part affinity fields. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2019. 4 +[5] Eric R. Chan, Connor Z. Lin, Matthew A. Chan, Koki Nagano, Boxiao Pan, Shalini De Mello, Orazio Gallo, Leonidas Guibas, Jonathan Tremblay, Sameh Khamis, Tero Karras, and Gordon Wetzstein. Efficient geometry-aware 3D generative adversarial networks. In CVPR, 2022. 1, 2, 4, 6 +[6] Feng-Ju Chang, Anh Tuan Tran, Tal Hassner, Iacopo Masi, Ram Nevatia, and Gerard Medioni. Faceposenet: Making a case for landmark-free face alignment. In Proceedings of the IEEE International Conference on Computer Vision Workshops, pages 1599-1608, 2017. 3 +[7] Jiankang Deng, Jia Guo, Niannan Xue, and Stefanos Zafeiriou. Arcface: Additive angular margin loss for deep face recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4690-4699, 2019. 3, 5, 6, 7, 8, 1, 2 +[8] Yu Deng, Jiaolong Yang, Sicheng Xu, Dong Chen, Yunde Jia, and Xin Tong. Accurate 3d face reconstruction with weakly-supervised learning: From single image to image set. In IEEE Computer Vision and Pattern Recognition Workshops, 2019. 6, 3 +[9] Yandong Guo, Lei Zhang, Yuxiao Hu, Xiaodong He, and Jianfeng Gao. Ms-celeb-1m: A dataset and benchmark for large-scale face recognition. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part III 14, pages 87-102. Springer, 2016. 1, 7 +[10] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep Residual Learning for Image Recognition. In Proceedings of 2016 IEEE Conference on Computer Vision and Pattern Recognition, pages 770-778. IEEE, 2016. 7 +[11] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2017. 5 +[12] Gary B Huang, Marwan Mattar, Tamara Berg, and Eric Learned-Miller. Labeled faces in the wild: A database for studying face recognition in unconstrained environments. In Workshop on faces in 'Real-Life' Images: detection, alignment, and recognition, 2008. 1, 3 + +[13] Yuge Huang, Yuhan Wang, Ying Tai, Xiaoming Liu, Pengcheng Shen, Shaoxin Li, Jilin Li, and Feiyue Huang. Curricularface: adaptive curriculum learning loss for deep face recognition. In proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5901-5910, 2020. 1, 3 +[14] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4401-4410, 2019. 1, 2, 3, 4 +[15] Tero Karras, Miika Aittala, Janne Hellsten, Samuli Laine, Jaakko Lehtinen, and Timo Aila. Training generative adversarial networks with limited data. In Proc. NeurIPS, 2020. 1, 2, 5 +[16] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of stylegan. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8110-8119, 2020. 4 +[17] Tero Karras, Miika Aittala, Samuli Laine, Erik Härkönen, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Alias-free generative adversarial networks. In Proc. NeurIPS, 2021. 2 +[18] Minchul Kim, Anil K Jain, and Xiaoming Liu. Adaface: Quality adaptive margin for face recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022. 1, 3, 7 +[19] Davis E. King. Dlib-ml: A machine learning toolkit. J. Mach. Learn. Res., 10:1755-1758, 2009. 3 +[20] Brendan F Klare, Ben Klein, Emma Taborsky, Austin Blanton, Jordan Cheney, Kristen Allen, Patrick Grother, Alan Mah, and Anil K Jain. Pushing the frontiers of unconstrained face detection and recognition: Iarpa janus benchmark a. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1931-1939, 2015. 3 +[21] Harold W Kuhn. The hungarian method for the assignment problem. Naval research logistics quarterly, 2(1-2):83-97, 1955. 3 +[22] Tuomas Kynkänniemi, Tero Karras, Samuli Laine, Jaakko Lehtinen, and Timo Aila. Improved precision and recall metric for assessing generative models. In Proceedings of the 33rd International Conference on Neural Information Processing Systems, Red Hook, NY, USA, 2019. Curran Associates Inc. 5 +[23] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. BLIP-2: bootstrapping language-image pre-training with frozen image encoders and large language models. In ICML, 2023. 4 +[24] Ziwei Liu, Ping Luo, Xiaogang Wang, and Xiaou Tang. Deep learning face attributes in the wild. In Proceedings of International Conference on Computer Vision (ICCV), 2015. 1 +[25] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In International Conference on Learning Representations, 2017. 6 +[26] Camillo Lugaresi, Jiuqiang Tang, Hadon Nash, Chris McClanahan, Esha Uboweja, Michael Hays, Fan Zhang, Chuo-Ling Chang, Ming Guang Yong, Juhyun Lee, Wan-Teh + +Chang, Wei Hua, Manfred Georg, and Matthias Grundmann. Mediapipe: A framework for building perception pipelines. CoRR, abs/1906.08172, 2019. 4 +[27] Brianna Maze, Jocelyn Adams, James A Duncan, Nathan Kalka, Tim Miller, Charles Otto, Anil K Jain, W Tyler Niggel, Janet Anderson, Jordan Cheney, et al. Iarpa janus benchmark-c: Face dataset and protocol. In 2018 international conference on biometrics (ICB), pages 158-165. IEEE, 2018. 3 +[28] Stylianos Moschoglou, Athanasios Papaioannou, Christos Sagonas, Jiankang Deng, Irene Kotsia, and Stefanos Zafeiriou. Agedb: the first manually collected, in-the-wild age database. In proceedings of the IEEE conference on computer vision and pattern recognition workshops, pages 51-59, 2017. 1, 3 +[29] Arsha Nagrani, Joon Son Chung, and Andrew Zisserman. VoxCeleb: A large-scale speaker identification dataset. In Interspeech 2017. ISCA, 2017. 2, 4, 5 +[30] Aaron Nech and Ira Kemelmacher-Shlizerman. Level playing field for million scale face recognition. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2017. 1 +[31] Youxin Pang, Yong Zhang, Weize Quan, Yanbo Fan, Xiaodong Cun, Ying Shan, and Dong-Ming Yan. Dpe: Disentanglement of pose and expression for general video portrait editing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 427-436, 2023. 2 +[32] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv e-prints, pages arXiv-2204, 2022. 2 +[33] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 2, 4, 6 +[34] Soumyadip Sengupta, Jun-Cheng Chen, Carlos Castillo, Vishal M Patel, Rama Chellappa, and David W Jacobs. Frontal to profile face verification in the wild. In 2016 IEEE winter conference on applications of computer vision (WACV), pages 1-9. IEEE, 2016. 6 +[35] Sefik Ilkin Serengil and Alper Ozpinar. Lightface: A hybrid deep face recognition framework. In 2020 Innovations in Intelligent Systems and Applications Conference (ASYU), pages 23-27. IEEE, 2020. 3 +[36] Sefik Ilkin Serengil and Alper Ozpinar. Hyperextended lightface: A facial attribute analysis framework. In 2021 International Conference on Engineering and Emerging Technologies (ICEET), pages 1-4. IEEE, 2021. 3, 5, 1, 2 +[37] Sefik Ilkin Serengil and Alper Ozpinar. Hyperextended lightface: A facial attribute analysis framework. In 2021 International Conference on Engineering and Emerging Technologies (ICEET), pages 1-4. IEEE, 2021. 4 +[38] Aliaksandr Siarohin, Stéphane Lathuilière, Sergey Tulyakov, Elisa Ricci, and Nicu Sebe. First order motion model for image animation. Advances in neural information processing systems, 32, 2019. 7, 2 + +[39] L Sirovich and M Kirby. Low-dimensional procedure for the characterization of human faces. Journal of the Optical Society of America A, 4(3):519-524, 1987. 1 +[40] Shaolin Su, Qingsen Yan, Yu Zhu, Cheng Zhang, Xin Ge, Jinqiu Sun, and Yanning Zhang. Blindly assess image quality in the wild guided by a self-adaptive hyper network. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 3 +[41] Jiale Tao, Biao Wang, Borun Xu, Tiezheng Ge, Yuning Jiang, Wen Li, and Lixin Duan. Structure-aware motion transfer with deformable anchor model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3637-3646, 2022. 2 +[42] Matthew A Turk and Alex P Pentland. Face recognition using eigenfaces. In Proceedings. 1991 IEEE computer society conference on computer vision and pattern recognition, pages 586-587. IEEE Computer Society, 1991. 1 +[43] Yaohui Wang, Di Yang, Francois Bremond, and Antitza Dantcheva. Latent image animator: Learning to animate images via latent space navigation. In ICLR 2022-The International Conference on Learning Representations, 2022. 2, 7, 3 +[44] Cameron Whitelam, Emma Taborsky, Austin Blanton, Brianna Maze, Jocelyn Adams, Tim Miller, Nathan Kalka, Anil K Jain, James A Duncan, Kristen Allen, et al. Iarpa janus benchmark-b face dataset. In proceedings of the IEEE conference on computer vision and pattern recognition workshops, pages 90-98, 2017. 3 +[45] L Wolf, T Hassner, and I Maoz. Face recognition in unconstrained videos with matched background similarity. In Proceedings of the 2011 IEEE Conference on Computer Vision and Pattern Recognition, pages 529-534, 2011. 1 +[46] Cho-Ying Wu, Qiangeng Xu, and Ulrich Neumann. Synergy between 3dmm and 3d landmarks for accurate 3d facial geometry. In 2021 International Conference on 3D Vision (3DV), 2021. 3 +[47] Yiqian Wu, Jing Zhang, Hongbo Fu, and Xiaogang Jin. Lpff: A portrait dataset for face generators across large poses. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 20327-20337, 2023. 1, 2, 3, 4 +[48] Liangbin Xie, Xintao Wang, Honglun Zhang, Chao Dong, and Ying Shan. Vfhq: A high-quality dataset and benchmark for video face super-resolution. In The IEEE Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), 2022. 1, 2, 3, 4 +[49] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3836-3847, 2023. 1, 2, 4, 6 +[50] Zhimeng Zhang, Lincheng Li, Yu Ding, and Changjie Fan. Flow-guided one-shot talking face generation with a high-resolution audio-visual dataset. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3661–3670, 2021. 2 +[51] Jian Zhao and Hui Zhang. Thin-plate spline motion model for image animation. In Proceedings of the IEEE/CVF Con + +ference on Computer Vision and Pattern Recognition, pages 3657-3666, 2022. 2, 7, 3 +[52] Tianyue Zheng and Weihong Deng. Cross-pose lfw: A database for studying cross-posed face recognition in unconstrained environments. Beijing University of Posts and Telecommunications, Tech. Rep, 5(7), 2018. 3, 4, 2 +[53] Huayi Zhou, Fei Jiang, and Hongtao Lu. Directmhp: Direct 2d multi-person head pose estimation with full-range angles. arXiv preprint arXiv:2302.01110, 2023. 3 +[54] Hao Zhu, Wayne Wu, Wentao Zhu, Liming Jiang, Siwei Tang, Li Zhang, Ziwei Liu, and Chen Change Loy. CelebV-HQ: A large-scale video facial attributes dataset. In ECCV, 2022. 1, 2, 3, 4 +[55] Zheng Zhu, Guan Huang, Jiankang Deng, Yun Ye, Junjie Huang, Xinze Chen, Jiagang Zhu, Tian Yang, Jiwen Lu, Dalong Du, and Jie Zhou. Webface260m: A benchmark unveiling the power of million-scale deep face recognition. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 1, 7 \ No newline at end of file diff --git a/2024/EFHQ_ Multi-purpose ExtremePose-Face-HQ dataset/images.zip b/2024/EFHQ_ Multi-purpose ExtremePose-Face-HQ dataset/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..64f301e54fbcf2f83e559846ee5b1575a848cfa7 --- /dev/null +++ b/2024/EFHQ_ Multi-purpose ExtremePose-Face-HQ dataset/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f7f02d32324440dae91cf8ef5d64edc18f0dcd5de29b3d16cdd6d700eabe0e64 +size 625399 diff --git a/2024/EFHQ_ Multi-purpose ExtremePose-Face-HQ dataset/layout.json b/2024/EFHQ_ Multi-purpose ExtremePose-Face-HQ dataset/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..080c9c62f24196d415f4188ef453c3133e57f454 --- /dev/null +++ b/2024/EFHQ_ Multi-purpose ExtremePose-Face-HQ dataset/layout.json @@ -0,0 +1,8035 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 130, + 103, + 463, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 103, + 463, + 121 + ], + "spans": [ + { + "bbox": [ + 130, + 103, + 463, + 121 + ], + "type": "text", + "content": "EFHQ: Multi-purpose ExtremePose-Face-HQ dataset" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 123, + 150, + 469, + 165 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 123, + 150, + 469, + 165 + ], + "spans": [ + { + "bbox": [ + 123, + 150, + 469, + 165 + ], + "type": "text", + "content": "Trung Tuan Dao" + }, + { + "bbox": [ + 123, + 150, + 469, + 165 + ], + "type": "inline_equation", + "content": "^{1,*}" + }, + { + "bbox": [ + 123, + 150, + 469, + 165 + ], + "type": "text", + "content": " Duc Hong Vu" + }, + { + "bbox": [ + 123, + 150, + 469, + 165 + ], + "type": "inline_equation", + "content": "^{1,2,*}" + }, + { + "bbox": [ + 123, + 150, + 469, + 165 + ], + "type": "text", + "content": " Cuong Pham" + }, + { + "bbox": [ + 123, + 150, + 469, + 165 + ], + "type": "inline_equation", + "content": "^{1,3}" + }, + { + "bbox": [ + 123, + 150, + 469, + 165 + ], + "type": "text", + "content": " Anh Tran" + }, + { + "bbox": [ + 123, + 150, + 469, + 165 + ], + "type": "inline_equation", + "content": "^{1}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 107, + 167, + 487, + 178 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 107, + 167, + 487, + 178 + ], + "spans": [ + { + "bbox": [ + 107, + 167, + 487, + 178 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 107, + 167, + 487, + 178 + ], + "type": "text", + "content": "VinAI Research, Vietnam " + }, + { + "bbox": [ + 107, + 167, + 487, + 178 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 107, + 167, + 487, + 178 + ], + "type": "text", + "content": "Miami University " + }, + { + "bbox": [ + 107, + 167, + 487, + 178 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 107, + 167, + 487, + 178 + ], + "type": "text", + "content": "Posts & Telecommunications Inst. of Tech., Vietnam" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 86, + 182, + 507, + 193 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 86, + 182, + 507, + 193 + ], + "spans": [ + { + "bbox": [ + 86, + 182, + 507, + 193 + ], + "type": "text", + "content": "{v.trungdt21,v.ducvh5,v.cuongpv11,v.anhtt152}@vinai.io vudh3@miamioh.edu cuongpv@ptit.edu.vn" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 259, + 194, + 332, + 206 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 259, + 194, + 332, + 206 + ], + "spans": [ + { + "bbox": [ + 259, + 194, + 332, + 206 + ], + "type": "text", + "content": "*Equal contribution" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 88, + 232, + 509, + 369 + ], + "blocks": [ + { + "bbox": [ + 88, + 232, + 509, + 369 + ], + "lines": [ + { + "bbox": [ + 88, + 232, + 509, + 369 + ], + "spans": [ + { + "bbox": [ + 88, + 232, + 509, + 369 + ], + "type": "image", + "image_path": "cbf16de748627893751e10fb88a1c1970cd6edd74202d56555bea8261d2edfef.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 374, + 548, + 464 + ], + "lines": [ + { + "bbox": [ + 46, + 374, + 548, + 464 + ], + "spans": [ + { + "bbox": [ + 46, + 374, + 548, + 464 + ], + "type": "text", + "content": "Figure 1. Benefits of our proposed dataset (EFHQ). Standard large-scale facial datasets have most images at near frontal views, causing inferior performance of trained models on downstream tasks when dealing with extreme head poses. For instance, the trained 2D image generators and text-to-image ones often produce only near frontal faces, while the 3D face generators and face reenactment methods often show distorted outputs at profile views. The recently proposed dataset LPFF [47] partially handles that issue by providing complementary images at extreme head poses for only 2D and 3D image generation tasks. Our proposed dataset EFHQ provides high-quality extreme-pose images to complement a wide range of face-related tasks. It supports 2D and 3D image generation, with generally better diversity than LPFF. EFHQ also helps correct the outputs of text-to-image generation and face reenactment at extreme views. Finally, EFHQ provides a more challenging pose-based face verification benchmark to better assess the quality of face recognition networks." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 143, + 473, + 192, + 485 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 473, + 192, + 485 + ], + "spans": [ + { + "bbox": [ + 143, + 473, + 192, + 485 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 45, + 498, + 290, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 498, + 290, + 715 + ], + "spans": [ + { + "bbox": [ + 45, + 498, + 290, + 715 + ], + "type": "text", + "content": "The existing facial datasets, while having plentiful images at near frontal views, lack images with extreme head poses, leading to the downgraded performance of deep learning models when dealing with profile or pitched faces. This work aims to address this gap by introducing a novel dataset named Extreme Pose Face High-Quality Dataset (EFHQ), which includes a maximum of 450k high-quality images of faces at extreme poses. To produce such a massive dataset, we utilize a novel and meticulous dataset processing pipeline to curate two publicly available datasets, VFHQ and CelebV-HQ, which contain many high-resolution face videos captured in various settings. Our dataset can complement existing datasets on various facial-related tasks, such as facial synthesis with 2D/3D-aware GAN, diffusion-based text-to-image face generation, and face reenactment. Specifically, training with EFHQ helps models generalize well across diverse poses, significantly improving performance in scenarios involving ex" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 475, + 548, + 546 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 475, + 548, + 546 + ], + "spans": [ + { + "bbox": [ + 304, + 475, + 548, + 546 + ], + "type": "text", + "content": "treme views, confirmed by extensive experiments. Additionally, we utilize EFHQ to define a challenging cross-view face verification benchmark, in which the performance of SOTA face recognition models drops " + }, + { + "bbox": [ + 304, + 475, + 548, + 546 + ], + "type": "inline_equation", + "content": "5 - 37\\%" + }, + { + "bbox": [ + 304, + 475, + 548, + 546 + ], + "type": "text", + "content": " compared to frontal-to-frontal scenarios, aiming to stimulate studies on face recognition under severe pose conditions in the wild." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 572, + 387, + 585 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 572, + 387, + 585 + ], + "spans": [ + { + "bbox": [ + 306, + 572, + 387, + 585 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "type": "text", + "content": "The human face is a central focus in computer vision, with extensive research dedicated to tasks like detection, recognition, generation, and manipulation [39, 42]. Numerous large-scale facial datasets[9, 12, 14, 24, 28, 30, 45, 48, 54, 55] have facilitated the face-related studies. Along with deep learning techniques, many tasks have made a considerable leap in performance in recent years. For instance, face recognition systems can recognize near-frontal faces in the wild with near-perfect accuracy [13, 18]. Facial generative models, such as [5, 15, 49], can synthesize realistic" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "22605" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 108 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 108 + ], + "type": "text", + "content": "images indistinguishable from real faces. Face reenactment systems, including [41, 43, 51], can animate an input face based on poses and expressions from any driver video." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 110, + 288, + 314 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 110, + 288, + 314 + ], + "spans": [ + { + "bbox": [ + 46, + 110, + 288, + 314 + ], + "type": "text", + "content": "Despite impressive results, the mentioned models produce inferior results when dealing with faces at extreme head poses. As illustrated in Fig. 1, faces synthesized by 3D face generation or face reenactment models are often distorted at profile views, due to the scarcity of extreme pose images in training datasets. Current extensive datasets for face generation leverage portrait images from Google or stock photo repositories like Flickr, which feature images with upfront views and minimal facial rotations. Face reenactment methods leverage video datasets such as Vox-Celeb [29], primarily featuring frames with frontal faces. The instability of pre-processing tools, particularly face alignment modules when handling extreme pose images, further reduces the proportion of those images in the existing datasets. This issue is critical since faces with large head rotations are common in real life, and dealing with them is unavoidable in practical systems." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 316, + 288, + 626 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 316, + 288, + 626 + ], + "spans": [ + { + "bbox": [ + 46, + 316, + 288, + 626 + ], + "type": "text", + "content": "This paper aims to fill that void by proposing a novel large-scale dataset of extreme-pose and high-quality facial images called ExtremePose-Face-HQ, or EFHQ for short. Our ambition is to build an \"one-for-all\" dataset that can supplement existing datasets on all mentioned face-related tasks. To achieve that goal, the dataset needs all the following properties: (1) extreme-pose, (2) large-scale, (3) high-quality, (4) in-the-wild, (5) having multiple images for each subject and having subject identity annotated. Given those requirements, we utilize a novel and meticulous dataset processing pipeline to filter and sample frames from two publicly available datasets, including VFHQ [48] and CelebV-HQ [54], which contain high-resolution face videos captured in various settings. This results in a dataset with 450k high-quality images of extreme poses, which are combined with frontal images of the same subject to use as image pairs. These image pairs can support face reenactment and face recognition tasks, while the extreme-pose images alone can complement the existing datasets of face generation training. Note that there is a recent attempt [47] to build a large-pose facial dataset similar to ours. However, that work relies on images crawled from Flickr, thus having limited size (19k images) and missing identity information. Hence, its applications are restricted, covering only several face-generation tasks, unlike our large-scale and multi-purpose dataset." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 630, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 630, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 630, + 287, + 715 + ], + "type": "text", + "content": "We perform a comprehensive set of experiments to validate the effectiveness of EFHQ across three distinct face generation sub-tasks and face reenactment scenarios. Our dataset, EFHQ, when combined with standard training sets, significantly enhances the quality of synthesized images in highly profile or pitched views while maintaining synthesis quality in frontal views. EFHQ also provides a more" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "text", + "content": "challenging pose-centric face verification benchmark. It reveals the vulnerability of many state-of-the-art face recognition networks when dealing with extreme-pose images, with TAR@FAR=1e-3 scores dropping 5-37%." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 121, + 546, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 121, + 546, + 251 + ], + "spans": [ + { + "bbox": [ + 304, + 121, + 546, + 251 + ], + "type": "text", + "content": "In summary, our contributions include (1) a large-scale, high-quality, in-the-wild, extreme-posed dataset (EFHQ) designed to complement various face-related tasks, (2) a novel and meticulous dataset processing pipeline ensuring the creation and quality of the EFHQ dataset, and (3) a series of experiments, with plan-to-release models, demonstrating the usefulness of EFHQ in enhancing performance on a wide range of tasks when dealing with faces at extreme rotations. The final dataset will be publicly available as metadata files to not violate the two original dataset's licenses, alongside the pretrained networks." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 263, + 392, + 275 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 263, + 392, + 275 + ], + "spans": [ + { + "bbox": [ + 306, + 263, + 392, + 275 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 289, + 545, + 408 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 289, + 545, + 408 + ], + "spans": [ + { + "bbox": [ + 304, + 289, + 545, + 408 + ], + "type": "text", + "content": "Synthetic Face Generation. Existing 2D-based models like StyleGAN [14, 15, 17] and 3D-aware models like EG3D [5] demonstrate impressive fidelity, especially for near-frontal views. Nevertheless, their performance significantly degrades when faced with extreme views, primarily due to insufficient training data for such scenarios. Datasets like FFHQ [14] exhibit a notable bias toward near-frontal poses, heavily limiting the range of poses synthesized by current 2D and 3D models despite the 3D-aware generators' ability to handle arbitrary poses." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 409, + 546, + 564 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 409, + 546, + 564 + ], + "spans": [ + { + "bbox": [ + 304, + 409, + 546, + 564 + ], + "type": "text", + "content": "Diffusion-based text-to-image models [32, 33] can synthesize high-quality images from simple text inputs. However, these models lack mechanisms to control the generated pose. Consequently, the generated facial images remain in a frontal view despite terms like \"profile\" or \"looking-up\" in the input prompts. ControlNet[49] proposes a finetuning approach allowing these models to take additional context as conditional input. Namely, one can use landmarks to guide the generated face to the desired pose. However, the scarcity of profile-view data in training affects the model's performance, particularly in extreme angles. Consequently, effectively managing the long-tailed distribution of these extreme views becomes necessary to address this issue." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 570, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 546, + 713 + ], + "type": "text", + "content": "Face Reenactment. Face reenactment involves transferring facial expressions and movements from a driving image to a source image. Recent advancements claim high-quality results, especially for near-frontal poses [31, 41, 43, 51]. However, pose collapse persists, particularly in keypoint-based models [41, 51]. Common datasets for this task are VoxCeleb1 [29] and HDTF [50], which primarily feature frontal talking face videos. Despite recent potentially high-quality datasets for reenactment such as VFHQ [48] and CelebV-HQ [54], the frontal-view bias persists. As a result, facial reenactment methods often produce artifacts and lower-quality outputs when faced with significant rotations." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "22606" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 97, + 70, + 492, + 235 + ], + "blocks": [ + { + "bbox": [ + 97, + 70, + 492, + 235 + ], + "lines": [ + { + "bbox": [ + 97, + 70, + 492, + 235 + ], + "spans": [ + { + "bbox": [ + 97, + 70, + 492, + 235 + ], + "type": "table", + "html": "
DatasetAttributesReady-to-use tasks
#imagesMultiple images per IDHigh QualityProfile ViewFace VerificationFace SynthesisFace Reenactment
AgeDB12K✓✓
CPLFW12K✓✓✓
CFP-FP14K✓✓✓
IJB-x138K✓✓✓
CelebA-HQ30K✓✓
FFHQ70K✓✓
LPFF19K✓✓✓
CelebV-HQ6M✓✓✓✓
VFHQ3M✓✓✓✓✓
VoxCeleb14M✓✓✓
Ours450,538✓✓✓✓✓✓
FFHQ+Ours*112,671✓✓✓
Vox+Ours*4.5M✓✓✓
", + "image_path": "beb48d3203d2a90d8ab0890319d86d2fe69993915e96900069cc7e6ecfa0a320.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 239, + 545, + 283 + ], + "lines": [ + { + "bbox": [ + 46, + 239, + 545, + 283 + ], + "spans": [ + { + "bbox": [ + 46, + 239, + 545, + 283 + ], + "type": "text", + "content": "Table 1. Representative comparison of key attributes and supported tasks across existing face datasets and our proposed EFHQ dataset. EFHQ provides large-scale, high-quality images with multiple views per identity, enabling diverse face-related tasks, including synthesis, reenactment, and verification. Combining EFHQ with existing datasets like FFHQ or VoxCeleb can augment available data for selected tasks. Asterisk (*) indicates a combined dataset with a subsampled version of EFHQ." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 293, + 287, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 293, + 287, + 413 + ], + "spans": [ + { + "bbox": [ + 46, + 293, + 287, + 413 + ], + "type": "text", + "content": "Face Verification. Despite near-perfect accuracy on benchmark datasets[12, 20, 27, 28, 44], state-of-the-art face verification models[7, 13, 18] fall short in unconstrained environments due to a lack of pose diversity. While CPLFW[52] introduces 6000 cross-pose pairs, its small size and mix of pose samples hinder analyzing performance on specific pose scenarios such as frontal-frontal, frontal-profile, and profile-profile. More granular benchmarking on diverse pose pairs is crucial to identify failure modes and drive progress in robust face verification across poses." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 418, + 287, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 418, + 287, + 574 + ], + "spans": [ + { + "bbox": [ + 46, + 418, + 287, + 574 + ], + "type": "text", + "content": "Extreme Pose Face Dataset. Tab. 1 summarizes existing datasets related to face verification, synthesis, and re-enactment, showing that existing face datasets have limitations for multi-task use. Recent efforts like LPFF[47] made progress by crawling from Flickr and processing for diverse views, resulting in a dataset of 19,590 images. The dataset could complement high-quality datasets like FFHQ[14] for generative models. However, due to its limited scale comprising only " + }, + { + "bbox": [ + 46, + 418, + 287, + 574 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 46, + 418, + 287, + 574 + ], + "type": "text", + "content": " of the combined dataset, the unbalanced distribution between frontal and profile view still needs further improvements. Additionally, since the images are from image hosting services, the dataset fails to provide the identity information necessary for other facial applications." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 586, + 142, + 599 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 586, + 142, + 599 + ], + "spans": [ + { + "bbox": [ + 47, + 586, + 142, + 599 + ], + "type": "text", + "content": "3. Dataset Process" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 607, + 287, + 656 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 607, + 287, + 656 + ], + "spans": [ + { + "bbox": [ + 46, + 607, + 287, + 656 + ], + "type": "text", + "content": "In this section, we discuss the process of producing our novel dataset EFHQ. To fulfill our goals, we carefully curated frames from the two recent facial video datasets, VFHQ[48] and CelebV-HQ[54], as seen in Fig. 2." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 664, + 176, + 677 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 664, + 176, + 677 + ], + "spans": [ + { + "bbox": [ + 47, + 664, + 176, + 677 + ], + "type": "text", + "content": "3.1. Attributes Preparation" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 689, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 689, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 689, + 287, + 713 + ], + "type": "text", + "content": "Attributes Extraction. To curate accurate facial poses and high-quality frames from existing datasets, we first ex" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 293, + 545, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 293, + 545, + 449 + ], + "spans": [ + { + "bbox": [ + 304, + 293, + 545, + 449 + ], + "type": "text", + "content": "tract (1) face bounding boxes, (2) facial landmarks (5/68 keypoints), (3) image quality score, and (4) face identity. For bounding boxes and landmarks, we employ the popular RetinaFace[36] and SynergyNet[46]. We use HyperIQA [40] to grade image quality. For VFHQ, we only generate 68 landmarks and image quality scores since bounding boxes and 5 keypoints are provided. After extracting these attributes, we match them with existing annotations using an IoU-based Hungarian matching algorithm [21]. Regarding identity, VFHQ-wise, we source labels from the annotations. As CelebV-HQ lacks individual identity labels within the video, we use[35] to identify the identities associated with each bounding box." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 452, + 545, + 656 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 452, + 545, + 656 + ], + "spans": [ + { + "bbox": [ + 304, + 452, + 545, + 656 + ], + "type": "text", + "content": "Due to the instability of head pose estimators when handling extreme-pose images, we strategically apply multiple pose estimators relying on various methodologies: (1) SynergyNet[46], a 3DMM-based model, (2) DirectMHP[53], a landmark-free joint head detector and pose estimator, and (3) an in-house FacePoseNet[6] improved through extensive data augmentation focused on extreme poses and extra training data. We then categorize each estimated pose into a hierarchical binning scheme illustrated in Fig. 2. Next, we gather bin predictions and perform majority voting to arrive at consensus labels. When no agreement emerges, we put it into the " + }, + { + "bbox": [ + 304, + 452, + 545, + 656 + ], + "type": "inline_equation", + "content": "5^{th}" + }, + { + "bbox": [ + 304, + 452, + 545, + 656 + ], + "type": "text", + "content": " bin for \"confusing\" cases. As depicted in Fig. 3, this ensemble approach provides robustness in cases where a single estimator incorrectly categorizes a sample. Please refer to the supplementary for statistics of the extracted bins and the hyperparameters for each model used for attribute extraction." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "content": "Annotation Review. Next, we manually review the results to ensure high-quality labels. Given the large dataset, we developed a graphical user interface tool to streamline the review process. Our focus is verifying the binning annota" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "type": "text", + "content": "22607" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 97, + 63, + 494, + 203 + ], + "blocks": [ + { + "bbox": [ + 97, + 63, + 494, + 203 + ], + "lines": [ + { + "bbox": [ + 97, + 63, + 494, + 203 + ], + "spans": [ + { + "bbox": [ + 97, + 63, + 494, + 203 + ], + "type": "image", + "image_path": "f254cce2dcbe27b3334c49712a16370a9280f177f1bfe4f41e39c028a6f522d5.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 207, + 547, + 240 + ], + "lines": [ + { + "bbox": [ + 46, + 207, + 547, + 240 + ], + "spans": [ + { + "bbox": [ + 46, + 207, + 547, + 240 + ], + "type": "text", + "content": "Figure 2. The pipeline of EFHQ dataset creation. Starting with high-quality videos from the VFHQ[48] and CelebV-HQ[54] datasets, single-frame attributes are extracted then manually reviews. Task-specific preprocessing is then applied to generate specialized versions of the dataset for tasks such as face generation, reenactment, and verification." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 59, + 255, + 274, + 355 + ], + "blocks": [ + { + "bbox": [ + 59, + 255, + 274, + 355 + ], + "lines": [ + { + "bbox": [ + 59, + 255, + 274, + 355 + ], + "spans": [ + { + "bbox": [ + 59, + 255, + 274, + 355 + ], + "type": "image", + "image_path": "38c77dbb4c806449587b476c0b22ce3f7a86ea80b9d0b108ab1e63be1859d8d5.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 47, + 358, + 287, + 380 + ], + "lines": [ + { + "bbox": [ + 47, + 358, + 287, + 380 + ], + "spans": [ + { + "bbox": [ + 47, + 358, + 287, + 380 + ], + "type": "text", + "content": "Figure 3. Example cases where a pose estimator fails to categorize the sample to the correct bin." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 393, + 287, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 393, + 287, + 441 + ], + "spans": [ + { + "bbox": [ + 46, + 393, + 287, + 441 + ], + "type": "text", + "content": "tions, which most directly impact label accuracy. Concurrently, we randomly subsample the data based on pose angle and image quality to validate facial landmark quality further and discard cases of landmark prediction failure." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 453, + 138, + 465 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 453, + 138, + 465 + ], + "spans": [ + { + "bbox": [ + 47, + 453, + 138, + 465 + ], + "type": "text", + "content": "3.2. EFHQ dataset" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 472, + 287, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 472, + 287, + 545 + ], + "spans": [ + { + "bbox": [ + 46, + 472, + 287, + 545 + ], + "type": "text", + "content": "Our final EFHQ dataset comprises up to 450k frames with extreme poses extracted from approximately 5,000 clips. Most of the clips in our dataset include at least one frame with a frontal face and multiple frames with extreme pose angles. We also include extreme pose-only clips, representing profile-to-profile pose transfer cases." + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 53, + 559, + 279, + 654 + ], + "blocks": [ + { + "bbox": [ + 53, + 559, + 279, + 654 + ], + "lines": [ + { + "bbox": [ + 53, + 559, + 279, + 654 + ], + "spans": [ + { + "bbox": [ + 53, + 559, + 279, + 654 + ], + "type": "image", + "image_path": "caec04bb3476c53e27cfcf754fc7da82c83f7eb861fc3f48a721d325f51ab89b.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 655, + 287, + 712 + ], + "lines": [ + { + "bbox": [ + 46, + 655, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 46, + 655, + 287, + 712 + ], + "type": "text", + "content": "Figure 4. Pose distribution comparison between our sampled dataset and other datasets, including FFHQ[14], LPFF[47], CPLFW[52] and 40K random samples from VoxCeleb1 [29]. Our sampled dataset demonstrates greater pose diversity, with increased sample counts across high angle bins." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 255, + 423, + 266 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 255, + 423, + 266 + ], + "spans": [ + { + "bbox": [ + 306, + 255, + 423, + 266 + ], + "type": "text", + "content": "3.3. Dataset for subtasks" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 284, + 545, + 452 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 284, + 545, + 452 + ], + "spans": [ + { + "bbox": [ + 304, + 284, + 545, + 452 + ], + "type": "text", + "content": "Supplementary datasets for face generation. To address FFHQ's pose distribution gap, we compile a dataset encompassing diverse poses, varied identities, and FFHQ-comparable image quality from EFHQ. Additionally, we also embed an image brightness filter[3] in the sampling process, given that HyperIQA doesn't explicitly assess this aspect. The sampled images are then processed using the same pipeline in [5, 16] and manually reviewed. The aim is to improve coverage of under-represented poses without degrading frontal performance of trained generation models. This yields a dataset of 42,671 images with equitably distributed poses as visualized in Fig. 4. Compared to the original FFHQ distribution, yaw and pitch distributions are significantly improved." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 456, + 545, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 456, + 545, + 529 + ], + "spans": [ + { + "bbox": [ + 304, + 456, + 545, + 529 + ], + "type": "text", + "content": "Additionally, for 3D-aware GAN subtasks, we extract camera parameters to serve as conditional pose information, following [5]. We convert these parameters into yaw, pitch angles to validate against our existing annotations. When disagreements emerge, we re-examine the images and filter if needed, to ensure accurate, quality labels." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 534, + 546, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 534, + 546, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 534, + 546, + 715 + ], + "type": "text", + "content": "For diffusion-based text-to-image generation, a specialized dataset is curated to refine Stable Diffusion models[33] by accommodating landmark conditional input through ControlNet[49]. Based on our 3D-aware face generation dataset, we integrate landmark-based conditional images and tailored text prompts for each image. To capture fine facial nuances, we extract 478 facial landmarks and draw the condition image with connected edges between matched keypoints, following Mediapipe framework[26], better-conveying face expression and pose. For text prompts, detailed facial attributes like gender, race, and emotion are derived from existing labels or inferred via BLIP-2 pretrained captioning model [23] and a face attribute estimator [37]. The resulting prompt adheres to a structured format: \"A profile portrait image of a [emotion] [race] [gender].\"" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "22608" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 48, + 70, + 288, + 205 + ], + "blocks": [ + { + "bbox": [ + 48, + 70, + 288, + 205 + ], + "lines": [ + { + "bbox": [ + 48, + 70, + 288, + 205 + ], + "spans": [ + { + "bbox": [ + 48, + 70, + 288, + 205 + ], + "type": "table", + "html": "
ModelReference DatasetFID↓Recall↑
GFFHQFFHQ2.840.49
GFFHQ+LPFFFFHQ+LPFF3.430.44
GFFHQ+EFHQFFHQ+EFHQ3.440.46
GFFHQ+LPFFFFHQ+EFHQ8.260.43
GFFHQ+EFHQFFHQ+LPFF7.040.44
GFFHQ+EFHQFFHQ+EFHQ3.330.46
GFFHQ+EFHQFFHQ3.120.44
GFFHQ+EFHQFFHQ+LPFF4.470.42
", + "image_path": "867f1367cf0d97dea13112a5cf0db97b2e1330fb65f9182a4c2e564667591df6.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 209, + 287, + 276 + ], + "lines": [ + { + "bbox": [ + 46, + 209, + 287, + 276 + ], + "spans": [ + { + "bbox": [ + 46, + 209, + 287, + 276 + ], + "type": "text", + "content": "Table 2. StyleGAN2-ADA models trained on different datasets. Lower FIDs indicate better fidelity, while higher Recalls indicate better diversity. The second and third blocks show comparisons between unconditional models trained on FFHQ+LPFF and FFHQ+EFHQ when evaluating on the same and cross-dataset settings, with the better metrics in bold." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 281, + 287, + 331 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 281, + 287, + 331 + ], + "spans": [ + { + "bbox": [ + 46, + 281, + 287, + 331 + ], + "type": "text", + "content": "Supplementary dataset for face reenactment. As EFHQ is crafted, we can employ the entire dataset to complement established training datasets, e.g., VoxCeleb1 [29]. We also built an extra evaluation set with 1200 EFHQ clips." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 336, + 287, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 336, + 287, + 479 + ], + "spans": [ + { + "bbox": [ + 46, + 336, + 287, + 479 + ], + "type": "text", + "content": "Benchmarking dataset for face verification. We curate a dataset covering three distinct scenarios: frontal-to-frontal, frontal-to-profile, and profile-to-profile. To enable rigorous benchmarking, we sample 10,000 pairs each for both negative and positive cases per scenario, resulting in a balanced benchmark dataset containing 60,000 image pairs. Images follow the same established preprocessing pipelines in [7]. Next, we filter misaligned images with RetinaFace[36], randomly review, and replace equivalent samples from our diverse corpus, if needed. Finally, to simulate varying image quality, we randomly applied downscaling and compression to samples in the dataset." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 494, + 192, + 506 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 494, + 192, + 506 + ], + "spans": [ + { + "bbox": [ + 47, + 494, + 192, + 506 + ], + "type": "text", + "content": "4. Face Generation Subtask" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 514, + 117, + 528 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 514, + 117, + 528 + ], + "spans": [ + { + "bbox": [ + 47, + 514, + 117, + 528 + ], + "type": "text", + "content": "4.1. StyleGAN" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 539, + 286, + 634 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 539, + 286, + 634 + ], + "spans": [ + { + "bbox": [ + 46, + 539, + 286, + 634 + ], + "type": "text", + "content": "Training details. We train StyleGAN2-ADA[15] models from scratch after combining FFHQ with our own curated dataset (denoted FFHQ+EFHQ) at " + }, + { + "bbox": [ + 46, + 539, + 286, + 634 + ], + "type": "inline_equation", + "content": "1024 \\times 1024" + }, + { + "bbox": [ + 46, + 539, + 286, + 634 + ], + "type": "text", + "content": " resolution. We also train a conditional version, allowing direct control over frontal versus profile synthesis and enhancing model performance. Specifically, we categorize the dataset into two classes: \"frontal\" and \"profile\" based on pose angles. All models are trained for 6 days on 8 Nvidia A100 GPUs." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 641, + 286, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 641, + 286, + 712 + ], + "spans": [ + { + "bbox": [ + 46, + 641, + 286, + 712 + ], + "type": "text", + "content": "Evaluation protocol. We utilize the Frechet Inception Distance[11] (FID) metric and improved Recall[22] to quantify the generation fidelity and diversity. Regarding the two unconditional models trained with the supplemented FFHQ+LPFF and FFHQ+EFHQ, we evaluate each model with both of these datasets. For conditional mod" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 326, + 70, + 526, + 138 + ], + "blocks": [ + { + "bbox": [ + 326, + 70, + 526, + 138 + ], + "lines": [ + { + "bbox": [ + 326, + 70, + 526, + 138 + ], + "spans": [ + { + "bbox": [ + 326, + 70, + 526, + 138 + ], + "type": "image", + "image_path": "3dcea56d5362a57eb679bd30469ce22064230f4b8d123e344efe25aaf2dc37ea.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 335, + 140, + 517, + 150 + ], + "lines": [ + { + "bbox": [ + 335, + 140, + 517, + 150 + ], + "spans": [ + { + "bbox": [ + 335, + 140, + 517, + 150 + ], + "type": "text", + "content": "(a) Samples from the model trained with FFHQ+EFHQ." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 325, + 151, + 525, + 277 + ], + "blocks": [ + { + "bbox": [ + 325, + 151, + 525, + 277 + ], + "lines": [ + { + "bbox": [ + 325, + 151, + 525, + 277 + ], + "spans": [ + { + "bbox": [ + 325, + 151, + 525, + 277 + ], + "type": "image", + "image_path": "3c0d5c9941056f9827add44aae19a2e757bb2b54eaf548fa151058eafdd04f75.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 324, + 277, + 525, + 289 + ], + "lines": [ + { + "bbox": [ + 324, + 277, + 525, + 289 + ], + "spans": [ + { + "bbox": [ + 324, + 277, + 525, + 289 + ], + "type": "text", + "content": "(b) Profile-view comparison of LPFF (left) and EFHQ (right)." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 305, + 292, + 545, + 324 + ], + "lines": [ + { + "bbox": [ + 305, + 292, + 545, + 324 + ], + "spans": [ + { + "bbox": [ + 305, + 292, + 545, + 324 + ], + "type": "text", + "content": "Figure 5. Qualitative result and comparison of generated samples, with truncation " + }, + { + "bbox": [ + 305, + 292, + 545, + 324 + ], + "type": "inline_equation", + "content": "\\psi = 0.7" + }, + { + "bbox": [ + 305, + 292, + 545, + 324 + ], + "type": "text", + "content": ", from StyleGAN2-ADA training with other dataset and ours." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 342, + 545, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 342, + 545, + 426 + ], + "spans": [ + { + "bbox": [ + 304, + 342, + 545, + 426 + ], + "type": "text", + "content": "els, we evaluate performance on FFHQ, FFHQ+LPFF, and FFHQ+EFHQ to demonstrate the broader coverage of the model trained with our proposed dataset. For FFHQ, we generate 50K images with the \"frontal\" label, whereas for the other two datasets, we generate 50K following their label distributions. This comprehensive quantitative analysis examines the fidelity and variety of the generated images." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 438, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 438, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 438, + 545, + 713 + ], + "type": "text", + "content": "Experimental results. We compared the performance of our proposed unconditional models " + }, + { + "bbox": [ + 304, + 438, + 545, + 713 + ], + "type": "inline_equation", + "content": "(G^{\\mathrm{FFHQ + EFHQ}})" + }, + { + "bbox": [ + 304, + 438, + 545, + 713 + ], + "type": "text", + "content": ", against models trained on the FFHQ and FFHQ+LPFF datasets " + }, + { + "bbox": [ + 304, + 438, + 545, + 713 + ], + "type": "inline_equation", + "content": "(G^{\\mathrm{FFHQ}}" + }, + { + "bbox": [ + 304, + 438, + 545, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 438, + 545, + 713 + ], + "type": "inline_equation", + "content": "G^{\\mathrm{FFHQ + LPFF}})" + }, + { + "bbox": [ + 304, + 438, + 545, + 713 + ], + "type": "text", + "content": ". Models trained conditionally are specifically subscripted, e.g., " + }, + { + "bbox": [ + 304, + 438, + 545, + 713 + ], + "type": "inline_equation", + "content": "G_{c}^{\\mathrm{FFHQ + EFHQ}}" + }, + { + "bbox": [ + 304, + 438, + 545, + 713 + ], + "type": "text", + "content": ". The qualitative results in Fig. 5 illustrate that our models generate high-quality frontal faces comparable to the FFHQ model while producing realistic and varied profile faces. Additionally, our method has less noise in extreme poses than the FFHQ+LPFF model. Specifically, Fig. 5b shows LPFF's generated face contains more noise and less photorealistic details than our model, such as the pixelated eye and nose patches. Meanwhile, ours achieves greater realism with smoother skin, well-defined features, noise reduction, and natural lighting. The quantitative results in Tab. 2 show that our unconditional model achieves competitive FID scores and improved Recall to other models, indicating similar fidelity but greater sample pose diversity. In the cross-examination (third block of Tab. 2), " + }, + { + "bbox": [ + 304, + 438, + 545, + 713 + ], + "type": "inline_equation", + "content": "G^{\\mathrm{FFHQ + EFHQ}}" + }, + { + "bbox": [ + 304, + 438, + 545, + 713 + ], + "type": "text", + "content": " has both better FID and improved Recall on the opposite dataset than " + }, + { + "bbox": [ + 304, + 438, + 545, + 713 + ], + "type": "inline_equation", + "content": "G^{\\mathrm{FFHQ + LPFF}}" + }, + { + "bbox": [ + 304, + 438, + 545, + 713 + ], + "type": "text", + "content": ". Finally, our conditional model demonstrates substantially improved FID scores over multiple reference datasets, highlighting its versatility." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "22609" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 49, + 70, + 286, + 163 + ], + "blocks": [ + { + "bbox": [ + 49, + 70, + 286, + 163 + ], + "lines": [ + { + "bbox": [ + 49, + 70, + 286, + 163 + ], + "spans": [ + { + "bbox": [ + 49, + 70, + 286, + 163 + ], + "type": "table", + "html": "
ModelReference DatasetFID↓Pose↓Depth↓ID↑
EFHQFFHQ/Trainset4.85e-40.2340.74
EFFHQFFHQ/Trainset3.75e-40.2880.74
EFFHQ+LPFFFFHQ6.26e-40.3160.77
EFFHQ+EFHQFFHQ3.94e-40.2240.77
EFFHQ+LPFFTrainset7.06e-40.3160.65
EFFHQ+EFHQTrainset4.64e-40.2230.65
", + "image_path": "44f3155bf285746e0ba7966b8c0a6c4ed50d0cc4776b7446c00c1152e1796d93.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 54, + 233, + 279, + 362 + ], + "blocks": [ + { + "bbox": [ + 54, + 233, + 279, + 362 + ], + "lines": [ + { + "bbox": [ + 54, + 233, + 279, + 362 + ], + "spans": [ + { + "bbox": [ + 54, + 233, + 279, + 362 + ], + "type": "image", + "image_path": "8e70a13b656bbde69089723c284a0af97e6cca6c1c51f5d55a61b3fe1f63ec4d.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 365, + 286, + 397 + ], + "lines": [ + { + "bbox": [ + 46, + 365, + 286, + 397 + ], + "spans": [ + { + "bbox": [ + 46, + 365, + 286, + 397 + ], + "type": "text", + "content": "Figure 6. Comparison between multiview generated samples, with truncation " + }, + { + "bbox": [ + 46, + 365, + 286, + 397 + ], + "type": "inline_equation", + "content": "\\psi = 0.8" + }, + { + "bbox": [ + 46, + 365, + 286, + 397 + ], + "type": "text", + "content": ", of EG3D model trained with various datasets." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 402, + 99, + 414 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 402, + 99, + 414 + ], + "spans": [ + { + "bbox": [ + 47, + 402, + 99, + 414 + ], + "type": "text", + "content": "4.2. EG3D" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 426, + 286, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 426, + 286, + 510 + ], + "spans": [ + { + "bbox": [ + 46, + 426, + 286, + 510 + ], + "type": "text", + "content": "Training details. We followed the same pipeline and hyper-parameters as the original EG3D model [5]. However, based on empirical results, we adjusted " + }, + { + "bbox": [ + 46, + 426, + 286, + 510 + ], + "type": "inline_equation", + "content": "\\gamma = 5" + }, + { + "bbox": [ + 46, + 426, + 286, + 510 + ], + "type": "text", + "content": ". We also fixed the triplane implementation bug (XY, XZ, ZX) [1] to replicate the original idea in the papers (XY, XZ, ZY). We trained the model on 8 Nvidia A100 40GB GPUs for approximately 7 days until convergence." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 516, + 286, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 516, + 286, + 601 + ], + "spans": [ + { + "bbox": [ + 46, + 516, + 286, + 601 + ], + "type": "text", + "content": "Evaluation Protocol. To evaluate the models, mirroring [5], we measure image quality with FID; multi-view consistency and identity preservation with mean cosine similarity of pre-trained face recognition model[7]; pose accuracy and geometry by calculating MSE against pseudo labels from [8]. We evaluate all models on 1,024 generated images, except for FID, which we compute on 50k images." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 605, + 286, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 286, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 286, + 714 + ], + "type": "text", + "content": "Experimental results. To evaluate the benefits of our dataset, we trained the same backbone architecture on different datasets: FFHQ " + }, + { + "bbox": [ + 46, + 605, + 286, + 714 + ], + "type": "inline_equation", + "content": "(E^{\\mathrm{FFHQ}})" + }, + { + "bbox": [ + 46, + 605, + 286, + 714 + ], + "type": "text", + "content": ", FFHQ+LPFF " + }, + { + "bbox": [ + 46, + 605, + 286, + 714 + ], + "type": "inline_equation", + "content": "(E^{\\mathrm{FFHQ} + \\mathrm{LPFF}})" + }, + { + "bbox": [ + 46, + 605, + 286, + 714 + ], + "type": "text", + "content": ", and FFHQ+EFHQ (our model, " + }, + { + "bbox": [ + 46, + 605, + 286, + 714 + ], + "type": "inline_equation", + "content": "E^{\\mathrm{FFHQ} + \\mathrm{EFHQ}}" + }, + { + "bbox": [ + 46, + 605, + 286, + 714 + ], + "type": "text", + "content": "). The quantitative results are presented in Table 3, and qualitative examples are shown in Figure 6. Qualitatively, we observe that " + }, + { + "bbox": [ + 46, + 605, + 286, + 714 + ], + "type": "inline_equation", + "content": "E^{\\mathrm{FFHQ} + \\mathrm{EFHQ}}" + }, + { + "bbox": [ + 46, + 605, + 286, + 714 + ], + "type": "text", + "content": " generates improved face shapes in extreme poses compared to " + }, + { + "bbox": [ + 46, + 605, + 286, + 714 + ], + "type": "inline_equation", + "content": "E^{\\mathrm{FFHQ}}" + }, + { + "bbox": [ + 46, + 605, + 286, + 714 + ], + "type": "text", + "content": ", similar to the improvements seen with " + }, + { + "bbox": [ + 46, + 605, + 286, + 714 + ], + "type": "inline_equation", + "content": "E^{\\mathrm{FFHQ} + \\mathrm{LPFF}}" + }, + { + "bbox": [ + 46, + 605, + 286, + 714 + ], + "type": "text", + "content": ". Quantitatively, the FID" + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 306, + 70, + 544, + 172 + ], + "blocks": [ + { + "bbox": [ + 47, + 167, + 286, + 232 + ], + "lines": [ + { + "bbox": [ + 47, + 167, + 286, + 232 + ], + "spans": [ + { + "bbox": [ + 47, + 167, + 286, + 232 + ], + "type": "text", + "content": "Table 3. EG3D models trained on different datasets. In the second and third blocks, we compare models trained on FFHQ+LPFF and FFHQ+EFHQ with the testset is either FFHQ or their training datasets; the better metrics are in bold. The best metrics on the FFHQ testset are in red, while the best metrics when evaluating on training dataset are underlined." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 306, + 70, + 544, + 172 + ], + "lines": [ + { + "bbox": [ + 306, + 70, + 544, + 172 + ], + "spans": [ + { + "bbox": [ + 306, + 70, + 544, + 172 + ], + "type": "table", + "html": "
MetricsDatasetPretrained ControlNetOurs
NME↓LPFF0.0850.084
FID↓LPFF119.7483.39
FIDeyepatch ↓LPFF103.7749.45
NME↓EFHQ0.1060.081
FID↓EFHQ111.2381.90
FIDeyepatch ↓EFHQ137.4358.60
", + "image_path": "7cff13cf3cc12be33b6c195c7d3f103525cc0601828cf4e62e89e6bc85fa9fff.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 175, + 545, + 208 + ], + "lines": [ + { + "bbox": [ + 305, + 175, + 545, + 208 + ], + "spans": [ + { + "bbox": [ + 305, + 175, + 545, + 208 + ], + "type": "text", + "content": "Table 4. Quantitative results of fine-tuning StableDiffusion v1.5 with ControlNet trained on various datasets, using Normalized Mean Error (NME), FID, and FID on eye region." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 305, + 214, + 545, + 372 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 214, + 545, + 372 + ], + "spans": [ + { + "bbox": [ + 305, + 214, + 545, + 372 + ], + "type": "text", + "content": "of " + }, + { + "bbox": [ + 305, + 214, + 545, + 372 + ], + "type": "inline_equation", + "content": "E^{\\mathrm{FFHQ + EFHQ}}" + }, + { + "bbox": [ + 305, + 214, + 545, + 372 + ], + "type": "text", + "content": " is comparable to " + }, + { + "bbox": [ + 305, + 214, + 545, + 372 + ], + "type": "inline_equation", + "content": "E^{\\mathrm{FFHQ}}" + }, + { + "bbox": [ + 305, + 214, + 545, + 372 + ], + "type": "text", + "content": " when evaluated on FFHQ, indicating no degradation in frontal view performance. Meanwhile, the competitive FID achieved by " + }, + { + "bbox": [ + 305, + 214, + 545, + 372 + ], + "type": "inline_equation", + "content": "E^{\\mathrm{FFHQ + EFHQ}}" + }, + { + "bbox": [ + 305, + 214, + 545, + 372 + ], + "type": "text", + "content": " on the combined FFHQ+EFHQ dataset also suggests proficient modeling of profile views. Regarding identity consistency, we see a significant drop when the reference dataset contains profile-view images, due to the weak confidence of the pretrained face recognition model when handling profile-view images, as confirmed by its low average cosine similarity of around 0.6 on the simple frontal-to-profile verification set CFP-FP[34]. Furthermore, our model also outperforms in pose accuracy and geometry quality with both frontal and cross-pose dataset." + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 314, + 380, + 425, + 458 + ], + "blocks": [ + { + "bbox": [ + 314, + 380, + 425, + 458 + ], + "lines": [ + { + "bbox": [ + 314, + 380, + 425, + 458 + ], + "spans": [ + { + "bbox": [ + 314, + 380, + 425, + 458 + ], + "type": "image", + "image_path": "91c1c39d34e1ff469ab19168be92e10d78b69610e53bee1af6b1b4b5daca6041.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 324, + 459, + 527, + 468 + ], + "lines": [ + { + "bbox": [ + 324, + 459, + 527, + 468 + ], + "spans": [ + { + "bbox": [ + 324, + 459, + 527, + 468 + ], + "type": "text", + "content": "(a) Generated samples; condition image is in between samples." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 426, + 380, + 539, + 457 + ], + "blocks": [ + { + "bbox": [ + 426, + 380, + 539, + 457 + ], + "lines": [ + { + "bbox": [ + 426, + 380, + 539, + 457 + ], + "spans": [ + { + "bbox": [ + 426, + 380, + 539, + 457 + ], + "type": "image", + "image_path": "e2849aa6187df629ee39998db75c92f3603e153a68b749d5b56f72afd4fcceeb.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 307, + 469, + 427, + 495 + ], + "blocks": [ + { + "bbox": [ + 307, + 469, + 427, + 495 + ], + "lines": [ + { + "bbox": [ + 307, + 469, + 427, + 495 + ], + "spans": [ + { + "bbox": [ + 307, + 469, + 427, + 495 + ], + "type": "image", + "image_path": "ea5572592c76c179b7a23caa798b0464f0d77a81bf0014c7a81b2a1c2f7a141b.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 384, + 497, + 470, + 507 + ], + "lines": [ + { + "bbox": [ + 384, + 497, + 470, + 507 + ], + "spans": [ + { + "bbox": [ + 384, + 497, + 470, + 507 + ], + "type": "text", + "content": "(b) Generated eye regions." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 428, + 469, + 547, + 495 + ], + "blocks": [ + { + "bbox": [ + 428, + 469, + 547, + 495 + ], + "lines": [ + { + "bbox": [ + 428, + 469, + 547, + 495 + ], + "spans": [ + { + "bbox": [ + 428, + 469, + 547, + 495 + ], + "type": "image", + "image_path": "42ea89334a93b892c5276bce82ddb8d2426f524d79dcfa74a222ced094510ff8.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 511, + 545, + 545 + ], + "lines": [ + { + "bbox": [ + 305, + 511, + 545, + 545 + ], + "spans": [ + { + "bbox": [ + 305, + 511, + 545, + 545 + ], + "type": "text", + "content": "Figure 7. Comparison profile-view generated samples of pretrained ControlNet (left) and our fine-tuned ControlNet (right) with the prompt: \"A profile portrait image of a person\"." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 552, + 381, + 563 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 552, + 381, + 563 + ], + "spans": [ + { + "bbox": [ + 306, + 552, + 381, + 563 + ], + "type": "text", + "content": "4.3. ControlNet" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 305, + 575, + 545, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 575, + 545, + 647 + ], + "spans": [ + { + "bbox": [ + 305, + 575, + 545, + 647 + ], + "type": "text", + "content": "Training details. We finetuned Stable Diffusion v1.5 [33] using ControlNet [49] on the EFHQ dataset to enhance its ability to generate extreme facial poses. We use AdamW optimizer [25] with learning rate " + }, + { + "bbox": [ + 305, + 575, + 545, + 647 + ], + "type": "inline_equation", + "content": "1 \\times e^{-5}" + }, + { + "bbox": [ + 305, + 575, + 545, + 647 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 305, + 575, + 545, + 647 + ], + "type": "inline_equation", + "content": "\\beta_0 = 0.9" + }, + { + "bbox": [ + 305, + 575, + 545, + 647 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 305, + 575, + 545, + 647 + ], + "type": "inline_equation", + "content": "\\beta_1 = 0.999" + }, + { + "bbox": [ + 305, + 575, + 545, + 647 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 305, + 575, + 545, + 647 + ], + "type": "inline_equation", + "content": "\\epsilon = 0.01" + }, + { + "bbox": [ + 305, + 575, + 545, + 647 + ], + "type": "text", + "content": ". We train with batch size 24 on 1 Nvidia A100 40GB GPU in 2 days." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "type": "text", + "content": "Evaluation Protocol. To assess image qualities across diverse poses, we measure the Normalized Mean Error (NME) of 478-point facial landmarks between the generated and source image. Specifically, we evaluate performance across the full pose distribution in the reference" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "22610" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 50, + 70, + 288, + 139 + ], + "blocks": [ + { + "bbox": [ + 50, + 70, + 288, + 139 + ], + "lines": [ + { + "bbox": [ + 50, + 70, + 288, + 139 + ], + "spans": [ + { + "bbox": [ + 50, + 70, + 288, + 139 + ], + "type": "table", + "html": "
ModelTest Dataset (L1 ↓/ AKD ↓/ AED ↓)
VoxCeleb1EFHQ (Ours)
TPS(VoxCeleb10.041/1.214/0.1320.061/3.814/0.275
TPS(VoxCeleb1+EFHQ0.041/1.250/0.1330.051/1.963/0.214
LIA(VoxCeleb10.044/1.342/0.1380.059/3.807/0.266
LIA(VoxCeleb1+EFHQ0.044/1.398/0.1420.054/2.291/0.233
", + "image_path": "0100f099836e506a48e568b5e3766bbb91f0c71b7470a5ec112ea03488b75695.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 191, + 287, + 288 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 191, + 287, + 288 + ], + "spans": [ + { + "bbox": [ + 46, + 191, + 287, + 288 + ], + "type": "text", + "content": "dataset. For the prompt, we use: \"A profile portrait image of a person.\" with extra positive and negative keywords, further detailed in supplementary. Utilizing LPFF and EFHQ datasets as the reference datasets, we compute FID to assess synthesized face quality and extend evaluation to cropped eye regions due to unrealistic gazes observed from the original ControlNet model. Note that the models receive condition images created akin to its training data." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 293, + 288, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 293, + 288, + 509 + ], + "spans": [ + { + "bbox": [ + 46, + 293, + 288, + 509 + ], + "type": "text", + "content": "Experimental results. Fig. 7a visually compares the baseline ControlNet OpenPose face and our fine-tuned model. Our fine-tuned model adeptly conditions diverse extreme poses, generating high-quality results with precise details while maintaining fidelity to the condition, such as mouth shape, and exhibiting minimal artifacts. These patterns indicate the superiority of the fine-tuned model over the baseline model in handling extreme viewing angles. Additionally, Fig. 7b contrasts the eye regions of both models. Our fine-tuned model accurately captures natural eye gaze and shape from condition images, a notable improvement over the baseline. Quantitative results in Tab. 4 consistently favor our fine-tuned model across all metrics. FID scores on full and eye patches exhibit significant enhancement, and NME performance remains stable on training and reference datasets (EFHQ and LPFF). These findings showcase the fine-tuned model's capability to generate high-quality, geometrically consistent faces across diverse poses." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 520, + 201, + 533 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 520, + 201, + 533 + ], + "spans": [ + { + "bbox": [ + 47, + 520, + 201, + 533 + ], + "type": "text", + "content": "5. Face Reenactment Subtask" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 539, + 163, + 551 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 539, + 163, + 551 + ], + "spans": [ + { + "bbox": [ + 47, + 539, + 163, + 551 + ], + "type": "text", + "content": "5.1. Evaluation Protocol" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 557, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 557, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 557, + 288, + 715 + ], + "type": "text", + "content": "We follow the same preprocessing and evaluation protocol as FOMM [38]. Specifically, all the frames are cropped to " + }, + { + "bbox": [ + 46, + 557, + 288, + 715 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 46, + 557, + 288, + 715 + ], + "type": "text", + "content": ", and we use the first frame of a clip as the source image to reconstruct the remaining frames. In the case of the EFHQ test set, most source images have a frontal pose, while the target frames have an extreme pose. Some test cases only contain extreme pose to evaluate self-reconstruction on these cases. We run the experiments using two recent methods, Thin-Plate Spline Motion Model (TPS) [51] and Latent Image Animator (LIA) [43]. For training details, please refer to the supplementary. We select the following three metrics to compare the generated frames and the driving ground truth. " + }, + { + "bbox": [ + 46, + 557, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_1" + }, + { + "bbox": [ + 46, + 557, + 288, + 715 + ], + "type": "text", + "content": " measures the av" + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 308, + 70, + 544, + 155 + ], + "blocks": [ + { + "bbox": [ + 46, + 143, + 287, + 187 + ], + "lines": [ + { + "bbox": [ + 46, + 143, + 287, + 187 + ], + "spans": [ + { + "bbox": [ + 46, + 143, + 287, + 187 + ], + "type": "text", + "content": "Table 5. Quantitative results of two state-of-the-art reenactment methods (TPS and LIA) when training on VoxCeleb1 and the combined set VoxCeleb1+EFHQ. We present results evaluated either on VoxCeleb1 (left) or EFHQ (right) test sets." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 308, + 70, + 544, + 155 + ], + "lines": [ + { + "bbox": [ + 308, + 70, + 544, + 155 + ], + "spans": [ + { + "bbox": [ + 308, + 70, + 544, + 155 + ], + "type": "table", + "html": "
ModelTrainsetCPLFWEFHQ
f2ff2pp2p
R18arfaceGlint360k0.800.940.570.370.880.06
R18arfaceMS1MV30.750.930.570.360.880.05
R50arfaceGlint360k0.910.970.770.200.920.05
R100arfaceGlint360k0.930.980.830.150.820.16
R100adafaceWebFace4M0.940.990.940.050.970.02
", + "image_path": "6d045a3c5d3768b18d0d1ca71db35c934026da6a519eeaea4da966eb56060a00.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 158, + 545, + 236 + ], + "lines": [ + { + "bbox": [ + 304, + 158, + 545, + 236 + ], + "spans": [ + { + "bbox": [ + 304, + 158, + 545, + 236 + ], + "type": "text", + "content": "Table 6. Quantitative evaluation of state-of-the-art face recognition models on the cleansed CPLFW and the proposed benchmark dataset. We use the metric TAR@FAR=1e-3 (higher is better). Our EFHQ benchmark includes frontal-frontal (f2f), frontal-profile (f2p), and profile-profile (p2p) subsets. For f2p and p2p, we report the score drop when comparing to f2f in the red subscription. The lowest benchmark score for each model is in bold." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 243, + 545, + 316 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 243, + 545, + 316 + ], + "spans": [ + { + "bbox": [ + 304, + 243, + 545, + 316 + ], + "type": "text", + "content": "erage distance among pixel values, indicating reconstruction faithfulness. Average Keypoint Distance (AKD) calculates the mean distance between keypoints, thus assessing the pose accuracy. Average Euclidean Distance (AED) measures identity preservation by computing the average " + }, + { + "bbox": [ + 304, + 243, + 545, + 316 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_2" + }, + { + "bbox": [ + 304, + 243, + 545, + 316 + ], + "type": "text", + "content": " distance between identity embeddings." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 327, + 426, + 339 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 327, + 426, + 339 + ], + "spans": [ + { + "bbox": [ + 305, + 327, + 426, + 339 + ], + "type": "text", + "content": "5.2. Experimental results" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 346, + 546, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 346, + 546, + 574 + ], + "spans": [ + { + "bbox": [ + 304, + 346, + 546, + 574 + ], + "type": "text", + "content": "We compare TPS and LIA models trained with our supplementary EFHQ dataset " + }, + { + "bbox": [ + 304, + 346, + 546, + 574 + ], + "type": "inline_equation", + "content": "(TPS^{\\mathrm{VoxCeleb1 + EFHQ}}" + }, + { + "bbox": [ + 304, + 346, + 546, + 574 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 346, + 546, + 574 + ], + "type": "inline_equation", + "content": "LIA^{\\mathrm{VoxCeleb1 + EFHQ}})" + }, + { + "bbox": [ + 304, + 346, + 546, + 574 + ], + "type": "text", + "content": " against models trained solely on VoxCeleb1 " + }, + { + "bbox": [ + 304, + 346, + 546, + 574 + ], + "type": "inline_equation", + "content": "(TPS^{\\mathrm{VoxCeleb1}}" + }, + { + "bbox": [ + 304, + 346, + 546, + 574 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 346, + 546, + 574 + ], + "type": "inline_equation", + "content": "LIA^{\\mathrm{VoxCeleb1}})" + }, + { + "bbox": [ + 304, + 346, + 546, + 574 + ], + "type": "text", + "content": ". Tab. 5 displays quantitative results, revealing that models trained with EFHQ maintain comparable performance when evaluated on VoxCeleb1, emphasizing high-quality performance in frontal-to-frontal reenactment scenarios. Notably, on the EFHQ test set, both EFHQ-trained models show significant improvement in all metrics, indicating enhanced extreme pose and identity reenactment performance. In Fig. 8, qualitative results prove that EFHQ-trained models excel in transferring facial expressions and motion across diverse poses. Models trained on EFHQ showcase improved image quality, reduced artifacts, and well-preserved shapes of driving faces. These results highlight our models' effectiveness in capturing and reproducing facial motions across varying poses while enhancing fidelity to the original facial features." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 588, + 454, + 602 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 588, + 454, + 602 + ], + "spans": [ + { + "bbox": [ + 305, + 588, + 454, + 602 + ], + "type": "text", + "content": "6. Face Recognition Subtask" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 609, + 422, + 621 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 609, + 422, + 621 + ], + "spans": [ + { + "bbox": [ + 306, + 609, + 422, + 621 + ], + "type": "text", + "content": "6.1. Evaluation Protocol" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "content": "To evaluate face recognition performance on extreme poses, we selected two state-of-the-art methods: ArcFace[7] and AdaFace[18]. We further explored various backbone sizes (ResNet18, ResNet50, ResNet101)[10] and training datasets (MS1MV2[9], Glint360K[2], WebFace4M[55]) of increasing size. For metrics, we used True Acceptance Rate (TAR) at False Acceptance Rate (FAR) of " + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "inline_equation", + "content": "1 \\times e^{-3}" + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "type": "text", + "content": "22611" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 103, + 79, + 162, + 137 + ], + "blocks": [ + { + "bbox": [ + 103, + 79, + 162, + 137 + ], + "lines": [ + { + "bbox": [ + 103, + 79, + 162, + 137 + ], + "spans": [ + { + "bbox": [ + 103, + 79, + 162, + 137 + ], + "type": "image", + "image_path": "2c8117e4d7e86d126609d5accefe90974ed636e97faaf9d4022551304474fbb7.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 103, + 138, + 164, + 196 + ], + "blocks": [ + { + "bbox": [ + 103, + 138, + 164, + 196 + ], + "lines": [ + { + "bbox": [ + 103, + 138, + 164, + 196 + ], + "spans": [ + { + "bbox": [ + 103, + 138, + 164, + 196 + ], + "type": "image", + "image_path": "f69d710555708f6551cef3b5ce48de1fde2a208e53247211237ab7c6fbae5fcb.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 121, + 201, + 144, + 209 + ], + "lines": [ + { + "bbox": [ + 121, + 201, + 144, + 209 + ], + "spans": [ + { + "bbox": [ + 121, + 201, + 144, + 209 + ], + "type": "text", + "content": "Source" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 170, + 79, + 229, + 137 + ], + "blocks": [ + { + "bbox": [ + 170, + 79, + 229, + 137 + ], + "lines": [ + { + "bbox": [ + 170, + 79, + 229, + 137 + ], + "spans": [ + { + "bbox": [ + 170, + 79, + 229, + 137 + ], + "type": "image", + "image_path": "7d2912606c5b61da5c606f88306fd6e060536c5ef9c314d5d5374ab99e85b6c3.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 170, + 138, + 230, + 198 + ], + "blocks": [ + { + "bbox": [ + 170, + 138, + 230, + 198 + ], + "lines": [ + { + "bbox": [ + 170, + 138, + 230, + 198 + ], + "spans": [ + { + "bbox": [ + 170, + 138, + 230, + 198 + ], + "type": "image", + "image_path": "61c559132cc4890d912246b645d7129a8dff94b65997700f60ef27bdef7545eb.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 213, + 545, + 237 + ], + "lines": [ + { + "bbox": [ + 46, + 213, + 545, + 237 + ], + "spans": [ + { + "bbox": [ + 46, + 213, + 545, + 237 + ], + "type": "text", + "content": "Figure 8. Comparison of performance for frontal-profile face reenactment on EFHQ. From left to right: source image, TPS on VoxCeleb1, TPS with EFHQ, LIA on VoxCeleb1, LIA with EFHQ, driving image. Additional EFHQ training data improves synthesis of extreme poses." + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 232, + 79, + 290, + 137 + ], + "blocks": [ + { + "bbox": [ + 232, + 79, + 290, + 137 + ], + "lines": [ + { + "bbox": [ + 232, + 79, + 290, + 137 + ], + "spans": [ + { + "bbox": [ + 232, + 79, + 290, + 137 + ], + "type": "image", + "image_path": "21f7cb68aa95c4ae50e04e75b16ffcc4a864c7551ef38fd4b4973bd55631b879.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 232, + 138, + 290, + 198 + ], + "blocks": [ + { + "bbox": [ + 174, + 201, + 227, + 209 + ], + "lines": [ + { + "bbox": [ + 174, + 201, + 227, + 209 + ], + "spans": [ + { + "bbox": [ + 174, + 201, + 227, + 209 + ], + "type": "text", + "content": "TPS (VoxCeleb1)" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 232, + 138, + 290, + 198 + ], + "lines": [ + { + "bbox": [ + 232, + 138, + 290, + 198 + ], + "spans": [ + { + "bbox": [ + 232, + 138, + 290, + 198 + ], + "type": "image", + "image_path": "74afe71aafafe1944aae621c88642f991e2738f3e52b03082731ebdf1d978f1f.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 243, + 201, + 280, + 210 + ], + "lines": [ + { + "bbox": [ + 243, + 201, + 280, + 210 + ], + "spans": [ + { + "bbox": [ + 243, + 201, + 280, + 210 + ], + "type": "text", + "content": "TPS (Ours)" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 299, + 79, + 359, + 137 + ], + "blocks": [ + { + "bbox": [ + 299, + 79, + 359, + 137 + ], + "lines": [ + { + "bbox": [ + 299, + 79, + 359, + 137 + ], + "spans": [ + { + "bbox": [ + 299, + 79, + 359, + 137 + ], + "type": "image", + "image_path": "2eb2f40499fafbf4fcb78e429ecb6a3f128bea37332e4a23ac5950ae3a760e86.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 299, + 138, + 359, + 196 + ], + "blocks": [ + { + "bbox": [ + 299, + 138, + 359, + 196 + ], + "lines": [ + { + "bbox": [ + 299, + 138, + 359, + 196 + ], + "spans": [ + { + "bbox": [ + 299, + 138, + 359, + 196 + ], + "type": "image", + "image_path": "13c7f831840d7ea36cbc2037bba4722099980fd3cb1850d9ce250d64185c67f0.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 201, + 358, + 209 + ], + "lines": [ + { + "bbox": [ + 304, + 201, + 358, + 209 + ], + "spans": [ + { + "bbox": [ + 304, + 201, + 358, + 209 + ], + "type": "text", + "content": "LIA (VoxCeleb1)" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 362, + 79, + 421, + 137 + ], + "blocks": [ + { + "bbox": [ + 362, + 79, + 421, + 137 + ], + "lines": [ + { + "bbox": [ + 362, + 79, + 421, + 137 + ], + "spans": [ + { + "bbox": [ + 362, + 79, + 421, + 137 + ], + "type": "image", + "image_path": "6d928c801997e6e85bd8d4dd1dab3f7ab8b88397fe154b822f5f89eac285434f.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 362, + 138, + 421, + 196 + ], + "blocks": [ + { + "bbox": [ + 362, + 138, + 421, + 196 + ], + "lines": [ + { + "bbox": [ + 362, + 138, + 421, + 196 + ], + "spans": [ + { + "bbox": [ + 362, + 138, + 421, + 196 + ], + "type": "image", + "image_path": "a37e832c4e7938ac2972ac1b7a72c3542f56a1a39355e19829977df27ba063d8.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 373, + 201, + 409, + 209 + ], + "lines": [ + { + "bbox": [ + 373, + 201, + 409, + 209 + ], + "spans": [ + { + "bbox": [ + 373, + 201, + 409, + 209 + ], + "type": "text", + "content": "LIA (Ours)" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 430, + 79, + 488, + 137 + ], + "blocks": [ + { + "bbox": [ + 430, + 79, + 488, + 137 + ], + "lines": [ + { + "bbox": [ + 430, + 79, + 488, + 137 + ], + "spans": [ + { + "bbox": [ + 430, + 79, + 488, + 137 + ], + "type": "image", + "image_path": "cf4b1d5c92581dc96cd7ada2ad8e39c5f6b911a5e50861f172bbd3c22fc7a3d0.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 429, + 138, + 487, + 196 + ], + "blocks": [ + { + "bbox": [ + 429, + 138, + 487, + 196 + ], + "lines": [ + { + "bbox": [ + 429, + 138, + 487, + 196 + ], + "spans": [ + { + "bbox": [ + 429, + 138, + 487, + 196 + ], + "type": "image", + "image_path": "227ecfb2eb75f91aa3aa305d5cb140c2676776284e15ef4c1a063540403f72ea.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 446, + 201, + 471, + 210 + ], + "lines": [ + { + "bbox": [ + 446, + 201, + 471, + 210 + ], + "spans": [ + { + "bbox": [ + 446, + 201, + 471, + 210 + ], + "type": "text", + "content": "Driving" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "bbox": [ + 47, + 245, + 162, + 258 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 245, + 162, + 258 + ], + "spans": [ + { + "bbox": [ + 47, + 245, + 162, + 258 + ], + "type": "text", + "content": "6.2. Experimental results" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 46, + 263, + 287, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 263, + 287, + 335 + ], + "spans": [ + { + "bbox": [ + 46, + 263, + 287, + 335 + ], + "type": "text", + "content": "Before assessing face recognition networks on EFHQ, we first clean the commonly-used version of CPLFW, processed by[7], by removing misaligned images (please refer to the Supplementary), and evaluate the models on it. Models performed well on this cleansed set, with minimal gaps between ArcFace and AdaFace." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 46, + 335, + 287, + 503 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 335, + 287, + 503 + ], + "spans": [ + { + "bbox": [ + 46, + 335, + 287, + 503 + ], + "type": "text", + "content": "Our proposed evaluation benchmark involves three pose-based scenarios. Results in Tab. 6 reveal performance drops of " + }, + { + "bbox": [ + 46, + 335, + 287, + 503 + ], + "type": "inline_equation", + "content": "5 - 37\\%" + }, + { + "bbox": [ + 46, + 335, + 287, + 503 + ], + "type": "text", + "content": " on the frontal-profile subset compared to frontal-frontal, highlighting the challenge of matching faces in different poses. Larger ResNet models outperformed ResNet18 across subsets, showcasing the advantages of increased capacity in handling pose variation. However, the ResNet100 model trained on Glint360K performed worse than smaller models on the profile-profile subset, indicating the need for sufficient profile data to prevent overfitting. Our results also indicate AdaFace's suitability for handling pose variations. In summary, pose mismatch brings significant challenges for face recognition, necessitating larger models trained on diverse data covering profile faces." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 47, + 512, + 117, + 523 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 512, + 117, + 523 + ], + "spans": [ + { + "bbox": [ + 47, + 512, + 117, + 523 + ], + "type": "text", + "content": "7. Discussion" + } + ] + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 59, + 529, + 273, + 644 + ], + "blocks": [ + { + "bbox": [ + 59, + 529, + 273, + 644 + ], + "lines": [ + { + "bbox": [ + 59, + 529, + 273, + 644 + ], + "spans": [ + { + "bbox": [ + 59, + 529, + 273, + 644 + ], + "type": "image", + "image_path": "e17c16d7e1823307bbc687e505298164fa1f4d5649b691e9a0108114e17bb041.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 648, + 286, + 670 + ], + "lines": [ + { + "bbox": [ + 46, + 648, + 286, + 670 + ], + "spans": [ + { + "bbox": [ + 46, + 648, + 286, + 670 + ], + "type": "text", + "content": "Figure 9. User study results. We reported the percentage of times testers rank whose output is best or tied for each criterion." + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + } + ], + "index": 23 + }, + { + "bbox": [ + 46, + 689, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 689, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 689, + 287, + 714 + ], + "type": "text", + "content": "User survey. We conduct user surveys to evaluate the face generation improvement of our method through human per" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 304, + 246, + 545, + 318 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 246, + 545, + 318 + ], + "spans": [ + { + "bbox": [ + 304, + 246, + 545, + 318 + ], + "type": "text", + "content": "ceptual assessment. In this survey, we compare the outputs of the models mentioned in Secs. 4 and 5. The setup of this user survey is included in the supplementary. In all human evaluation results (Fig. 9), our models outperform previous ones, notably on identity and motion test of face reenactment models (74.80% and 64.22%, respectively)." + } + ] + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 323, + 332, + 526, + 401 + ], + "blocks": [ + { + "bbox": [ + 323, + 332, + 526, + 401 + ], + "lines": [ + { + "bbox": [ + 323, + 332, + 526, + 401 + ], + "spans": [ + { + "bbox": [ + 323, + 332, + 526, + 401 + ], + "type": "image", + "image_path": "5406b417968dda5059ac39836df61475119ae9f04474919d0df72f427ee2ccb5.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 404, + 545, + 459 + ], + "lines": [ + { + "bbox": [ + 304, + 404, + 545, + 459 + ], + "spans": [ + { + "bbox": [ + 304, + 404, + 545, + 459 + ], + "type": "text", + "content": "Figure 10. Representative examples of the raw LPFF dataset that got excluded from the final dataset, either from misdetections or pose filtering process. Red keypoints represent landmarks prediction of selected samples from LPFF, green keypoints represent landmarks prediction from our pipeline." + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_caption" + } + ], + "index": 27 + }, + { + "bbox": [ + 304, + 471, + 545, + 554 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 471, + 545, + 554 + ], + "spans": [ + { + "bbox": [ + 304, + 471, + 545, + 554 + ], + "type": "text", + "content": "Robustness of data processing pipeline. As shown in Fig. 10, our data processing pipeline can detect additional profile-view faces filtered out by the LPFF pipeline, likely due to misdetections or incorrect pose filtering. This confirms the robustness of our pipeline, and we hope it will enable the community to develop larger and higher-quality in-the-wild facial datasets." + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 306, + 571, + 378, + 582 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 571, + 378, + 582 + ], + "spans": [ + { + "bbox": [ + 306, + 571, + 378, + 582 + ], + "type": "text", + "content": "8. Conclusion" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 304, + 592, + 545, + 703 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 592, + 545, + 703 + ], + "spans": [ + { + "bbox": [ + 304, + 592, + 545, + 703 + ], + "type": "text", + "content": "This work has introduced a large-scale, diverse facial dataset to address performance gaps between frontal and profile faces, powered by a novel and robust data processing pipeline. Crucially, we provided tailored sub-datasets for advancing essential tasks like face synthesis and reenactment. Moreover, our new face verification benchmark reveals gaps between techniques, granting valuable insights. Ultimately, we hope this high-quality, diverse facial dataset opens up new and exciting opportunities to push forward cross-pose tasks." + } + ] + } + ], + "index": 31 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "22612" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 123 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 123 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 123 + ], + "type": "text", + "content": "[1] Fix for triplane projection " + }, + { + "bbox": [ + 53, + 91, + 287, + 123 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 53, + 91, + 287, + 123 + ], + "type": "text", + "content": " Issue #67 " + }, + { + "bbox": [ + 53, + 91, + 287, + 123 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 53, + 91, + 287, + 123 + ], + "type": "text", + "content": " NVlabs/eg3d. https://github.com/NVlabs/eg3d/issues/67, 2023.6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 124, + 288, + 179 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 124, + 288, + 179 + ], + "spans": [ + { + "bbox": [ + 53, + 124, + 288, + 179 + ], + "type": "text", + "content": "[2] Xiang An, Xuhan Zhu, Yuan Gao, Yang Xiao, Yongle Zhao, Ziyong Feng, Lan Wu, Bin Qin, Ming Zhang, Debing Zhang, et al. Partial fc: Training 10 million identities on a single machine. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1445-1449, 2021. 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 180, + 288, + 223 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 180, + 288, + 223 + ], + "spans": [ + { + "bbox": [ + 53, + 180, + 288, + 223 + ], + "type": "text", + "content": "[3] Sergey Bezryadin, Pavel Bourov, and Dmitry Ilinih. Brightness calculation in digital image processing. International Symposium on Technologies for Digital Photo Fulfillment, 2007:10-15, 2007. 4, 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 224, + 288, + 269 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 224, + 288, + 269 + ], + "spans": [ + { + "bbox": [ + 53, + 224, + 288, + 269 + ], + "type": "text", + "content": "[4] Z. Cao, G. Hidalgo Martinez, T. Simon, S. Wei, and Y. A. Sheikh. Openpose: Realtime multi-person 2d pose estimation using part affinity fields. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2019. 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 270, + 288, + 324 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 270, + 288, + 324 + ], + "spans": [ + { + "bbox": [ + 53, + 270, + 288, + 324 + ], + "type": "text", + "content": "[5] Eric R. Chan, Connor Z. Lin, Matthew A. Chan, Koki Nagano, Boxiao Pan, Shalini De Mello, Orazio Gallo, Leonidas Guibas, Jonathan Tremblay, Sameh Khamis, Tero Karras, and Gordon Wetzstein. Efficient geometry-aware 3D generative adversarial networks. In CVPR, 2022. 1, 2, 4, 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 325, + 288, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 325, + 288, + 380 + ], + "spans": [ + { + "bbox": [ + 53, + 325, + 288, + 380 + ], + "type": "text", + "content": "[6] Feng-Ju Chang, Anh Tuan Tran, Tal Hassner, Iacopo Masi, Ram Nevatia, and Gerard Medioni. Faceposenet: Making a case for landmark-free face alignment. In Proceedings of the IEEE International Conference on Computer Vision Workshops, pages 1599-1608, 2017. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 380, + 288, + 434 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 380, + 288, + 434 + ], + "spans": [ + { + "bbox": [ + 53, + 380, + 288, + 434 + ], + "type": "text", + "content": "[7] Jiankang Deng, Jia Guo, Niannan Xue, and Stefanos Zafeiriou. Arcface: Additive angular margin loss for deep face recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4690-4699, 2019. 3, 5, 6, 7, 8, 1, 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 436, + 288, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 436, + 288, + 491 + ], + "spans": [ + { + "bbox": [ + 53, + 436, + 288, + 491 + ], + "type": "text", + "content": "[8] Yu Deng, Jiaolong Yang, Sicheng Xu, Dong Chen, Yunde Jia, and Xin Tong. Accurate 3d face reconstruction with weakly-supervised learning: From single image to image set. In IEEE Computer Vision and Pattern Recognition Workshops, 2019. 6, 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 491, + 288, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 491, + 288, + 557 + ], + "spans": [ + { + "bbox": [ + 53, + 491, + 288, + 557 + ], + "type": "text", + "content": "[9] Yandong Guo, Lei Zhang, Yuxiao Hu, Xiaodong He, and Jianfeng Gao. Ms-celeb-1m: A dataset and benchmark for large-scale face recognition. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part III 14, pages 87-102. Springer, 2016. 1, 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 558, + 287, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 558, + 287, + 602 + ], + "spans": [ + { + "bbox": [ + 48, + 558, + 287, + 602 + ], + "type": "text", + "content": "[10] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep Residual Learning for Image Recognition. In Proceedings of 2016 IEEE Conference on Computer Vision and Pattern Recognition, pages 770-778. IEEE, 2016. 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 602, + 287, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 602, + 287, + 657 + ], + "spans": [ + { + "bbox": [ + 48, + 602, + 287, + 657 + ], + "type": "text", + "content": "[11] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2017. 5" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 658, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 287, + 713 + ], + "type": "text", + "content": "[12] Gary B Huang, Marwan Mattar, Tamara Berg, and Eric Learned-Miller. Labeled faces in the wild: A database for studying face recognition in unconstrained environments. In Workshop on faces in 'Real-Life' Images: detection, alignment, and recognition, 2008. 1, 3" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 138 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 138 + ], + "type": "text", + "content": "[13] Yuge Huang, Yuhan Wang, Ying Tai, Xiaoming Liu, Pengcheng Shen, Shaoxin Li, Jilin Li, and Feiyue Huang. Curricularface: adaptive curriculum learning loss for deep face recognition. In proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5901-5910, 2020. 1, 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 140, + 545, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 140, + 545, + 194 + ], + "spans": [ + { + "bbox": [ + 307, + 140, + 545, + 194 + ], + "type": "text", + "content": "[14] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4401-4410, 2019. 1, 2, 3, 4" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 194, + 545, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 194, + 545, + 237 + ], + "spans": [ + { + "bbox": [ + 307, + 194, + 545, + 237 + ], + "type": "text", + "content": "[15] Tero Karras, Miika Aittala, Janne Hellsten, Samuli Laine, Jaakko Lehtinen, and Timo Aila. Training generative adversarial networks with limited data. In Proc. NeurIPS, 2020. 1, 2, 5" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 238, + 545, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 238, + 545, + 293 + ], + "spans": [ + { + "bbox": [ + 307, + 238, + 545, + 293 + ], + "type": "text", + "content": "[16] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of stylegan. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8110-8119, 2020. 4" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 293, + 545, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 293, + 545, + 327 + ], + "spans": [ + { + "bbox": [ + 307, + 293, + 545, + 327 + ], + "type": "text", + "content": "[17] Tero Karras, Miika Aittala, Samuli Laine, Erik Härkönen, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Alias-free generative adversarial networks. In Proc. NeurIPS, 2021. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 327, + 545, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 327, + 545, + 370 + ], + "spans": [ + { + "bbox": [ + 307, + 327, + 545, + 370 + ], + "type": "text", + "content": "[18] Minchul Kim, Anil K Jain, and Xiaoming Liu. Adaface: Quality adaptive margin for face recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022. 1, 3, 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 371, + 545, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 371, + 545, + 392 + ], + "spans": [ + { + "bbox": [ + 307, + 371, + 545, + 392 + ], + "type": "text", + "content": "[19] Davis E. King. Dlib-ml: A machine learning toolkit. J. Mach. Learn. Res., 10:1755-1758, 2009. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 393, + 545, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 393, + 545, + 459 + ], + "spans": [ + { + "bbox": [ + 307, + 393, + 545, + 459 + ], + "type": "text", + "content": "[20] Brendan F Klare, Ben Klein, Emma Taborsky, Austin Blanton, Jordan Cheney, Kristen Allen, Patrick Grother, Alan Mah, and Anil K Jain. Pushing the frontiers of unconstrained face detection and recognition: Iarpa janus benchmark a. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1931-1939, 2015. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 460, + 545, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 460, + 545, + 491 + ], + "spans": [ + { + "bbox": [ + 307, + 460, + 545, + 491 + ], + "type": "text", + "content": "[21] Harold W Kuhn. The hungarian method for the assignment problem. Naval research logistics quarterly, 2(1-2):83-97, 1955. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 492, + 545, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 492, + 545, + 557 + ], + "spans": [ + { + "bbox": [ + 307, + 492, + 545, + 557 + ], + "type": "text", + "content": "[22] Tuomas Kynkänniemi, Tero Karras, Samuli Laine, Jaakko Lehtinen, and Timo Aila. Improved precision and recall metric for assessing generative models. In Proceedings of the 33rd International Conference on Neural Information Processing Systems, Red Hook, NY, USA, 2019. Curran Associates Inc. 5" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 558, + 545, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 558, + 545, + 602 + ], + "spans": [ + { + "bbox": [ + 307, + 558, + 545, + 602 + ], + "type": "text", + "content": "[23] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. BLIP-2: bootstrapping language-image pre-training with frozen image encoders and large language models. In ICML, 2023. 4" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 602, + 545, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 602, + 545, + 646 + ], + "spans": [ + { + "bbox": [ + 307, + 602, + 545, + 646 + ], + "type": "text", + "content": "[24] Ziwei Liu, Ping Luo, Xiaogang Wang, and Xiaou Tang. Deep learning face attributes in the wild. In Proceedings of International Conference on Computer Vision (ICCV), 2015. 1" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 647, + 545, + 680 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 647, + 545, + 680 + ], + "spans": [ + { + "bbox": [ + 307, + 647, + 545, + 680 + ], + "type": "text", + "content": "[25] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In International Conference on Learning Representations, 2017. 6" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 681, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 681, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 681, + 545, + 713 + ], + "type": "text", + "content": "[26] Camillo Lugaresi, Jiuqiang Tang, Hadon Nash, Chris McClanahan, Esha Uboweja, Michael Hays, Fan Zhang, Chuo-Ling Chang, Ming Guang Yong, Juhyun Lee, Wan-Teh" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "22613" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 286, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 66, + 72, + 286, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 72, + 286, + 106 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 286, + 106 + ], + "type": "text", + "content": "Chang, Wei Hua, Manfred Georg, and Matthias Grundmann. Mediapipe: A framework for building perception pipelines. CoRR, abs/1906.08172, 2019. 4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 106, + 286, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 106, + 286, + 171 + ], + "spans": [ + { + "bbox": [ + 48, + 106, + 286, + 171 + ], + "type": "text", + "content": "[27] Brianna Maze, Jocelyn Adams, James A Duncan, Nathan Kalka, Tim Miller, Charles Otto, Anil K Jain, W Tyler Niggel, Janet Anderson, Jordan Cheney, et al. Iarpa janus benchmark-c: Face dataset and protocol. In 2018 international conference on biometrics (ICB), pages 158-165. IEEE, 2018. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 172, + 286, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 172, + 286, + 237 + ], + "spans": [ + { + "bbox": [ + 49, + 172, + 286, + 237 + ], + "type": "text", + "content": "[28] Stylianos Moschoglou, Athanasios Papaioannou, Christos Sagonas, Jiankang Deng, Irene Kotsia, and Stefanos Zafeiriou. Agedb: the first manually collected, in-the-wild age database. In proceedings of the IEEE conference on computer vision and pattern recognition workshops, pages 51-59, 2017. 1, 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 239, + 286, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 239, + 286, + 270 + ], + "spans": [ + { + "bbox": [ + 49, + 239, + 286, + 270 + ], + "type": "text", + "content": "[29] Arsha Nagrani, Joon Son Chung, and Andrew Zisserman. VoxCeleb: A large-scale speaker identification dataset. In Interspeech 2017. ISCA, 2017. 2, 4, 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 49, + 272, + 286, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 272, + 286, + 316 + ], + "spans": [ + { + "bbox": [ + 49, + 272, + 286, + 316 + ], + "type": "text", + "content": "[30] Aaron Nech and Ira Kemelmacher-Shlizerman. Level playing field for million scale face recognition. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2017. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 49, + 316, + 286, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 316, + 286, + 380 + ], + "spans": [ + { + "bbox": [ + 49, + 316, + 286, + 380 + ], + "type": "text", + "content": "[31] Youxin Pang, Yong Zhang, Weize Quan, Yanbo Fan, Xiaodong Cun, Ying Shan, and Dong-Ming Yan. Dpe: Disentanglement of pose and expression for general video portrait editing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 427-436, 2023. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 49, + 382, + 286, + 424 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 382, + 286, + 424 + ], + "spans": [ + { + "bbox": [ + 49, + 382, + 286, + 424 + ], + "type": "text", + "content": "[32] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv e-prints, pages arXiv-2204, 2022. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 49, + 426, + 286, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 426, + 286, + 480 + ], + "spans": [ + { + "bbox": [ + 49, + 426, + 286, + 480 + ], + "type": "text", + "content": "[33] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 2, 4, 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 49, + 481, + 286, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 481, + 286, + 536 + ], + "spans": [ + { + "bbox": [ + 49, + 481, + 286, + 536 + ], + "type": "text", + "content": "[34] Soumyadip Sengupta, Jun-Cheng Chen, Carlos Castillo, Vishal M Patel, Rama Chellappa, and David W Jacobs. Frontal to profile face verification in the wild. In 2016 IEEE winter conference on applications of computer vision (WACV), pages 1-9. IEEE, 2016. 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 49, + 537, + 286, + 580 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 537, + 286, + 580 + ], + "spans": [ + { + "bbox": [ + 49, + 537, + 286, + 580 + ], + "type": "text", + "content": "[35] Sefik Ilkin Serengil and Alper Ozpinar. Lightface: A hybrid deep face recognition framework. In 2020 Innovations in Intelligent Systems and Applications Conference (ASYU), pages 23-27. IEEE, 2020. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 49, + 582, + 286, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 582, + 286, + 624 + ], + "spans": [ + { + "bbox": [ + 49, + 582, + 286, + 624 + ], + "type": "text", + "content": "[36] Sefik Ilkin Serengil and Alper Ozpinar. Hyperextended lightface: A facial attribute analysis framework. In 2021 International Conference on Engineering and Emerging Technologies (ICEET), pages 1-4. IEEE, 2021. 3, 5, 1, 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 49, + 625, + 286, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 625, + 286, + 669 + ], + "spans": [ + { + "bbox": [ + 49, + 625, + 286, + 669 + ], + "type": "text", + "content": "[37] Sefik Ilkin Serengil and Alper Ozpinar. Hyperextended lightface: A facial attribute analysis framework. In 2021 International Conference on Engineering and Emerging Technologies (ICEET), pages 1-4. IEEE, 2021. 4" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 49, + 670, + 286, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 670, + 286, + 712 + ], + "spans": [ + { + "bbox": [ + 49, + 670, + 286, + 712 + ], + "type": "text", + "content": "[38] Aliaksandr Siarohin, Stéphane Lathuilière, Sergey Tulyakov, Elisa Ricci, and Nicu Sebe. First order motion model for image animation. Advances in neural information processing systems, 32, 2019. 7, 2" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "text", + "content": "[39] L Sirovich and M Kirby. Low-dimensional procedure for the characterization of human faces. Journal of the Optical Society of America A, 4(3):519-524, 1987. 1" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 108, + 545, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 108, + 545, + 162 + ], + "spans": [ + { + "bbox": [ + 307, + 108, + 545, + 162 + ], + "type": "text", + "content": "[40] Shaolin Su, Qingsen Yan, Yu Zhu, Cheng Zhang, Xin Ge, Jinqiu Sun, and Yanning Zhang. Blindly assess image quality in the wild guided by a self-adaptive hyper network. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 163, + 545, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 163, + 545, + 217 + ], + "spans": [ + { + "bbox": [ + 307, + 163, + 545, + 217 + ], + "type": "text", + "content": "[41] Jiale Tao, Biao Wang, Borun Xu, Tiezheng Ge, Yuning Jiang, Wen Li, and Lixin Duan. Structure-aware motion transfer with deformable anchor model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3637-3646, 2022. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 220, + 545, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 220, + 545, + 262 + ], + "spans": [ + { + "bbox": [ + 307, + 220, + 545, + 262 + ], + "type": "text", + "content": "[42] Matthew A Turk and Alex P Pentland. Face recognition using eigenfaces. In Proceedings. 1991 IEEE computer society conference on computer vision and pattern recognition, pages 586-587. IEEE Computer Society, 1991. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 264, + 545, + 317 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 264, + 545, + 317 + ], + "spans": [ + { + "bbox": [ + 308, + 264, + 545, + 317 + ], + "type": "text", + "content": "[43] Yaohui Wang, Di Yang, Francois Bremond, and Antitza Dantcheva. Latent image animator: Learning to animate images via latent space navigation. In ICLR 2022-The International Conference on Learning Representations, 2022. 2, 7, 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 320, + 545, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 320, + 545, + 386 + ], + "spans": [ + { + "bbox": [ + 308, + 320, + 545, + 386 + ], + "type": "text", + "content": "[44] Cameron Whitelam, Emma Taborsky, Austin Blanton, Brianna Maze, Jocelyn Adams, Tim Miller, Nathan Kalka, Anil K Jain, James A Duncan, Kristen Allen, et al. Iarpa janus benchmark-b face dataset. In proceedings of the IEEE conference on computer vision and pattern recognition workshops, pages 90-98, 2017. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 388, + 545, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 388, + 545, + 431 + ], + "spans": [ + { + "bbox": [ + 308, + 388, + 545, + 431 + ], + "type": "text", + "content": "[45] L Wolf, T Hassner, and I Maoz. Face recognition in unconstrained videos with matched background similarity. In Proceedings of the 2011 IEEE Conference on Computer Vision and Pattern Recognition, pages 529-534, 2011. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 433, + 545, + 476 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 433, + 545, + 476 + ], + "spans": [ + { + "bbox": [ + 308, + 433, + 545, + 476 + ], + "type": "text", + "content": "[46] Cho-Ying Wu, Qiangeng Xu, and Ulrich Neumann. Synergy between 3dmm and 3d landmarks for accurate 3d facial geometry. In 2021 International Conference on 3D Vision (3DV), 2021. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 478, + 545, + 531 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 478, + 545, + 531 + ], + "spans": [ + { + "bbox": [ + 308, + 478, + 545, + 531 + ], + "type": "text", + "content": "[47] Yiqian Wu, Jing Zhang, Hongbo Fu, and Xiaogang Jin. Lpff: A portrait dataset for face generators across large poses. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 20327-20337, 2023. 1, 2, 3, 4" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 534, + 545, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 534, + 545, + 588 + ], + "spans": [ + { + "bbox": [ + 308, + 534, + 545, + 588 + ], + "type": "text", + "content": "[48] Liangbin Xie, Xintao Wang, Honglun Zhang, Chao Dong, and Ying Shan. Vfhq: A high-quality dataset and benchmark for video face super-resolution. In The IEEE Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), 2022. 1, 2, 3, 4" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 590, + 545, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 590, + 545, + 634 + ], + "spans": [ + { + "bbox": [ + 308, + 590, + 545, + 634 + ], + "type": "text", + "content": "[49] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3836-3847, 2023. 1, 2, 4, 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 635, + 545, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 635, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 308, + 635, + 545, + 689 + ], + "type": "text", + "content": "[50] Zhimeng Zhang, Lincheng Li, Yu Ding, and Changjie Fan. Flow-guided one-shot talking face generation with a high-resolution audio-visual dataset. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3661–3670, 2021. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 691, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 691, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 308, + 691, + 545, + 712 + ], + "type": "text", + "content": "[51] Jian Zhao and Hui Zhang. Thin-plate spline motion model for image animation. In Proceedings of the IEEE/CVF Con" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "22614" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 73, + 287, + 285 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 66, + 73, + 286, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 73, + 286, + 94 + ], + "spans": [ + { + "bbox": [ + 66, + 73, + 286, + 94 + ], + "type": "text", + "content": "ference on Computer Vision and Pattern Recognition, pages 3657-3666, 2022. 2, 7, 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 96, + 287, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 96, + 287, + 140 + ], + "spans": [ + { + "bbox": [ + 48, + 96, + 287, + 140 + ], + "type": "text", + "content": "[52] Tianyue Zheng and Weihong Deng. Cross-pose lfw: A database for studying cross-posed face recognition in unconstrained environments. Beijing University of Posts and Telecommunications, Tech. Rep, 5(7), 2018. 3, 4, 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 141, + 287, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 141, + 287, + 173 + ], + "spans": [ + { + "bbox": [ + 48, + 141, + 287, + 173 + ], + "type": "text", + "content": "[53] Huayi Zhou, Fei Jiang, and Hongtao Lu. Directmhp: Direct 2d multi-person head pose estimation with full-range angles. arXiv preprint arXiv:2302.01110, 2023. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 175, + 287, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 175, + 287, + 217 + ], + "spans": [ + { + "bbox": [ + 48, + 175, + 287, + 217 + ], + "type": "text", + "content": "[54] Hao Zhu, Wayne Wu, Wentao Zhu, Liming Jiang, Siwei Tang, Li Zhang, Ziwei Liu, and Chen Change Loy. CelebV-HQ: A large-scale video facial attributes dataset. In ECCV, 2022. 1, 2, 3, 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 220, + 287, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 220, + 287, + 285 + ], + "spans": [ + { + "bbox": [ + 48, + 220, + 287, + 285 + ], + "type": "text", + "content": "[55] Zheng Zhu, Guan Huang, Jiankang Deng, Yun Ye, Junjie Huang, Xinze Chen, Jiagang Zhu, Tian Yang, Jiwen Lu, Dalong Du, and Jie Zhou. Webface260m: A benchmark unveiling the power of million-scale deep face recognition. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 1, 7" + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "22615" + } + ] + } + ], + "index": 6 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/EFormer_ Enhanced Transformer towards Semantic-Contour Features of Foreground for Portraits Matting/652e59cf-51c6-401e-a44e-7a8d54a66599_content_list.json b/2024/EFormer_ Enhanced Transformer towards Semantic-Contour Features of Foreground for Portraits Matting/652e59cf-51c6-401e-a44e-7a8d54a66599_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..1d2c928108e2ecc3651fe56376c49da8e75ea21a --- /dev/null +++ b/2024/EFormer_ Enhanced Transformer towards Semantic-Contour Features of Foreground for Portraits Matting/652e59cf-51c6-401e-a44e-7a8d54a66599_content_list.json @@ -0,0 +1,1576 @@ +[ + { + "type": "text", + "text": "EFormer: Enhanced Transformer towards Semantic-Contour Features of Foreground for Portraits Matting", + "text_level": 1, + "bbox": [ + 114, + 128, + 854, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Zitao Wang Qiguang Miao* Yue Xi* Peipei Zhao", + "bbox": [ + 259, + 203, + 709, + 222 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Xidian University, China", + "bbox": [ + 408, + 224, + 558, + 238 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{21181214174@stu.，qgmiao@，xiyue@，zhpp2023@}xidian.edu.cn", + "bbox": [ + 218, + 241, + 740, + 256 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 291, + 313, + 306 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The portrait matting task aims to extract an alpha matte with complete semantics and finely detailed contours. In comparison to CNN-based approaches, transformers with self-attention module have a better capacity to capture long-range dependencies and low-frequency semantic information of a portrait. However, recent research shows that the self-attention mechanism struggles with modeling high-frequency contour information and capturing fine contour details, which can lead to bias while predicting the portrait's contours. To deal with this issue, we propose EFormer to enhance the model's attention towards both the low-frequency semantic and high-frequency contour features. For the high-frequency contours, our research demonstrates that cross-attention module between different resolutions can guide our model to allocate attention appropriately to these contour regions. Supported by this, we can successfully extract the high-frequency detail information around the portrait's contours, which were previously ignored by self-attention. Based on the cross-attention module, we further build a semantic and contour detector (SCD) to accurately capture both the low-frequency semantic and high-frequency contour features. And we design a contour-edge extraction branch and semantic extraction branch to extract refined high-frequency contour features and complete low-frequency semantic information, respectively. Finally, we fuse the two kinds of features and leverage the segmentation head to generate a predicted portrait matte. Experiments on VideoMatte240K (JPEG SD Format) and Adobe Image Matting (AIM) datasets demonstrate that EFormer outperforms previous portrait matte methods.", + "bbox": [ + 75, + 323, + 473, + 777 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 805, + 209, + 821 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Portrait matting aims to extract a precise alpha matte of persons from natural images. It has gathered great momentum in vision applications, including but not limited to photo", + "bbox": [ + 76, + 830, + 468, + 878 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/c9937fb43e03a7633aee3ae5e531b87101ba94ad97cd2f40462a3b2430c6fa14.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 501, + 290, + 890, + 473 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/3b06508ac7a09bba239e870787572c7c30df658ab82ad01b9fb750bf1c4e71f8.jpg", + "image_caption": [ + "Figure 1. (I) Visualization with Grad-CAM of output from each layer within different transformer blocks. (a) The features captured from Self-attention and MLP layers in the original transformer block. (b) In comparison, the features captured from each layer in the proposed transformer block. (II) Comparisons of different transformer blocks. (c) The original transformer block in ViT. (d) The proposed transformer block for portrait matte." + ], + "image_footnote": [], + "bbox": [ + 517, + 477, + 890, + 656 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "editing and background replacement.", + "bbox": [ + 500, + 794, + 746, + 809 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Existing methods[20, 30, 33, 35] attempt to estimate a portrait matte with Convolutional Neural Networks (CNNs) under the guidance of a pre-determined trimap, which is a three-class map that indicates the foreground, background, and unknown region. However, it is time-consuming and labor-intensive to obtain abundant trimaps by manually an-", + "bbox": [ + 496, + 810, + 893, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding authors.", + "bbox": [ + 94, + 886, + 227, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "3880", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "notating images. Some methods eliminate the trimap, such as background matting series[21, 28] uses a separate background image instead of the trimap as an auxiliary reference for model prediction. However, in practical cases, it is necessary to provide two aligned images for the model, which include one with only the background and the other one including people. An alternative solution[17, 22, 23] is to rely solely on RGB images as input for predicting portrait matte. Although the previous models have made strides in portrait matte, there is a constraint on the receptive field in CNN-based approaches.", + "bbox": [ + 75, + 90, + 472, + 257 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In contrast, Vision Transformer (ViT)[7] and Swin Transformer (Swin) [24] for semantic segmentation employ a self-attention mechanism, which conducts interactions among all pixels in an image to obtain a global receptive field. Previous research [25] shows that ViT and its variants can capture the low-frequency components in images effectively, such as global shapes and structures of a scene or object. However, they usually neglect high-frequency ones in images, such as edges and textures. They work similarly to low-pass filters.", + "bbox": [ + 75, + 258, + 472, + 411 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "As shown in Fig. 1(a), we can observe the transformer block with self-attention clearly pays little attention to the portrait's high-frequency contour regions. We believe that this phenomenon is attributed to the excessive propagation of global information because global attention is performed in each transformer block during the calculation of multiple blocks. So, the global information is propagated throughout the model in this iterative calculation, ultimately leading it to only focus on low-frequency semantic features with stronger clustering in portraits. Over time, the model gradually ignores the discrete high-frequency detail features around the contours. Therefore, it is desired to capture low- and high-frequency components simultaneously to improve the segmentation performance.", + "bbox": [ + 75, + 412, + 472, + 625 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To achieve this goal, Inception Transformer (IFormer)[29] splits all channels into a convolutional path and self-attention path, respectively, to capture high and low frequencies. However, it is required for a channel ratio of each block at different levels in the IFormer. These channel ratios typically require manual adjustment, resulting in significant uncertainty when aiming for optimal model performance. To enhance the model's stability and adaptive capability, we propose a novel architecture EFormer without manually adjusted parameters. In EFormer, we use the cross-attention layer between different resolution features, as shown in Fig. 1(d), which is replaced by convolution to handle high-frequency detail features.", + "bbox": [ + 75, + 626, + 472, + 823 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our research further demonstrates that the cross-attention module between different resolutions can formulate a more reasonable attention allocation mechanism, which can accurately capture contour features, as shown in Fig. 1(b). Particularly, EFormer with a cross-attention layer", + "bbox": [ + 75, + 825, + 472, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "not only can not lose the semantic information within the portrait but also can successfully capture the high-frequency details around the portrait's contours, which were previously ignored by self-attention. It is worth noting that high-frequency components only around the boundary between foreground and background are critical to the improvement of segmentation performance. Therefore, we follow a coarse-to-fine manner. Initially, EFormer needs to locate all high-frequency regions in the image. Then, it filters and extracts the high-frequency regions around the foreground contours.", + "bbox": [ + 496, + 90, + 890, + 257 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Based on the cross-attention module, we design a semantic and contour detector (SCD) in EFormer, which cascades the cross-attention layer and self-attention layer to sequentially locate the contour and semantic features. Firstly, the model adjusts its attention to contour features with the guidance of a cross-attention module. Secondly, the self-attention module captures semantic information within a portrait's contours and reverses filtering and correcting the contour information output by the cross-attention layer. This allows the model to gradually capture and match the high-frequency contour information and the low-frequency semantic information. Then, we use multi-layer perceptron (MLP) to build the contour-edge extraction branch (CEEB) and the semantic extraction branch (SEB), respectively. With this support, we independently purify the contour flow and the semantic flow to obtain finer contour information and more comprehensive portrait semantic information, as shown in Fig. 1(b). Finally, we fuse the contour features and semantic features and send them to the segmentation head to estimate the portrait matte. We perform extensive experiments on VideoMatte240K (JPEG SD Format)[21], Adobe Image Matting (AIM)[35], and BG10K. Comparisons with other state-of-the-art models and ablation studies verify that our model performs better than previous works.", + "bbox": [ + 496, + 258, + 892, + 619 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To summarize, our contributions are as follows:", + "bbox": [ + 517, + 621, + 834, + 636 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose EFormer: an approach that can enhance the transformer's attention towards both semantic and contour features of the foreground.", + "- We further build a semantic and contour detector (SCD) to accurately capture the semantic and contour features and design two separate extraction branches to purify the contour flow and the semantic flow.", + "- We leverage the cross-attention module between different resolution features in transformer block to autonomously capture and extract high-frequency detail features, which are sparsely distributed along the foreground contour." + ], + "bbox": [ + 496, + 638, + 890, + 804 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 500, + 820, + 650, + 835 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Image Matting", + "text_level": 1, + "bbox": [ + 500, + 845, + 651, + 863 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Image matting is an essential task in the field of computer vision that aims to accurately estimate the foreground in an", + "bbox": [ + 500, + 869, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "3881", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "image. Mathematically speaking, an image $I$ is a combination of an unknown foreground image $F$ and a background image $B$ with a probability coefficient alpha matte maps $\\alpha$ .", + "bbox": [ + 75, + 90, + 468, + 137 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nI = \\alpha F + (1 - \\alpha) B \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 199, + 147, + 468, + 164 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Previous image matting solutions have predominantly focused on low-level features, such as color cues[9, 10, 14, 15] or propagation[1, 2, 19], to distinguish the transition areas between foregrounds and backgrounds. However, such traditional matting algorithms commonly struggle to perform properly in complex scenes.", + "bbox": [ + 75, + 174, + 468, + 263 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "With the significant advancements making in deep learning, numerous methods based on convolutional neural networks (CNNs) are proposed, leading to notable successes. Some approaches[36, 37, 39] incorporate auxiliary trimap supervisions to enhance matting performance, while other methods[22, 23] leverage trimap-free solutions to estimate alpha matts from image feature maps using an end-to-end segmentation network. In the background matting series[21, 28], an auxiliary input of the background image is used to compute alpha matte. Furthermore, MODNet[17] proposes an end-to-end manner without any auxiliary input. Although prior works based on CNN kernels have improved the accuracy of image matting, the prediction of models is in the limited receptive field.", + "bbox": [ + 75, + 265, + 468, + 474 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. Vision Transformer", + "text_level": 1, + "bbox": [ + 76, + 484, + 264, + 500 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Different from CNNs, vision transformers with self-attention mechanisms can capture long-term dependencies. Therefore, in the realm of computer vision, vision transformers have gained significant attention in recent times. For pixel-level prediction tasks, such as segmentation, SETR[38], SegFormer[34] and DPT[27] apply transformer as an encoder to attain feature maps of images. Their performances demonstrate the transformer with self-attention is capable of building complete contextual information through global interactions. However, recent studies[25, 29] indicate that transformer \"yet is incompetent in capturing high frequencies that predominantly convey local information\".", + "bbox": [ + 75, + 508, + 468, + 702 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Compared to self-attention, cross-attention has more potential to induce customized features. Max-Deeplab[31], MaskFormer[5, 6] and SeMask[13] use query-based methods through cross-attention, which are inspired by DETR[3]. They view segmentation as a set prediction problem. U-Transformer[26] and EUT[8] modify the cross-attention in the original transformer to leverage the information from the encoder, allowing a fine spatial recovery in the decoder. The cross-resolution attention employed by RTFormer[32] enables the gathering of comprehensive contextual information for high-resolution features.", + "bbox": [ + 75, + 704, + 468, + 869 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this work, we propose a strategy that using of crossattention between different resolutions guides the model to", + "bbox": [ + 76, + 869, + 468, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "autonomously locate and capture high-frequency features, particularly those near the portrait's contours. Contrary to IFormer[29], this method eliminates the need for predefined parameters to enhance the model's adaptive capabilities. At the same time, this approach can significantly improve the model's performance in capturing contour details.", + "bbox": [ + 496, + 90, + 890, + 181 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.3. Refinement for Segmentation", + "text_level": 1, + "bbox": [ + 498, + 191, + 759, + 207 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In the field of portrait segmentation, refining the high-frequency details near the contours is crucial. Many prior studies on segmentation refinement depend on convolutional networks or MLPs that are designed specifically for this purpose.", + "bbox": [ + 496, + 214, + 890, + 289 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "PointRend[18], based on the coarse predictions of Mask R-CNN[12], calculates confidence scores as the point-wise uncertainty measure. And it focuses on the points with the higher uncertainty according to set hyperparameters. Finally, PointRend uses a shared MLP to refine the labels of the selected points. Although PointRend has implemented segmentation refinement to some extent, the model still requires manual adjustment of certain hyperparameters to assist in filtering uncertain points. Transfiner[16] processes detected error-prone and incoherent regions with transformer, which include high-frequency regions strewn along object boundaries. However, during this procedure, excessive focus is placed on the local details and incoherent regions, without effectively associating the local features of the object with the overall semantic information.", + "bbox": [ + 496, + 290, + 892, + 515 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We believe that locating effective detail distribution areas requires information interaction between local details and global semantics. To achieve this, we design a semantic and contour detector that adopts a cascade structure of cross-attention and self-attention. This approach uses global semantic information garnered from self-attention to autonomously select out local details situated at the borders of the contours. On this base, we further utilize the MLP-based extraction branch to purify contour details, thereby significantly enhancing the effect of refinement.", + "bbox": [ + 496, + 516, + 890, + 667 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Model", + "text_level": 1, + "bbox": [ + 500, + 681, + 578, + 696 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we introduce the process of establishing the complete model architecture of EFormer, which comprises the backbone for encoder, the transformer block for decoder, and the prediction stage, as shown in Fig. 2.", + "bbox": [ + 496, + 705, + 890, + 767 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Backbone for Encoder", + "text_level": 1, + "bbox": [ + 500, + 776, + 709, + 791 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The CNN backbone takes RGB images $I \\in \\mathbb{R}^{B \\times 3 \\times H \\times W}$ as input, where $B$ is the number of images, to obtain a pyramid of feature maps $F = \\left\\{F_{enc\\frac{1}{4}}, F_{enc\\frac{1}{8}}, F_{enc\\frac{1}{16}}\\right\\}$ with dimension $F_{enc\\frac{1}{i}} \\in \\mathbb{R}^{B \\times C_i \\times \\frac{H}{i} \\times \\frac{W}{i}}$ . As we all know, in the pyramid of feature maps, shallow features have high resolution and rich details, while deep features have low reso", + "bbox": [ + 496, + 797, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3882", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/19dc18b25a5ce7108cc5db14d5b7e0dec35cccd2f842cb026d2fcf2d6a35cae3.jpg", + "image_caption": [ + "Figure 2. The Architecture of the EFormer. It contains a backbone for encoder, a transformer block for decoder, and the prediction stage, where the transformer block includes a semantic and contour detector, a semantic extraction branch, and a contour-edge extraction branch." + ], + "image_footnote": [], + "bbox": [ + 99, + 85, + 874, + 393 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "lation but more concentrated semantic information. Based on this, our research makes the cross-attention mechanism to filter the high-frequency contour features from high-resolution features, using low-frequency semantic information in low-resolution features. However, since the large resolution of feature maps would bring large computation cost, we use the $F_{enc\\frac{1}{8}}$ as $F_{HR}$ (high-resolution features) and use the $F_{enc\\frac{1}{16}}$ as $F_{LR}$ (low-resolution features) in our base model. The impact of other choices of $F_{HR}$ and $F_{LR}$ on model performance is demonstrated in ablation experiments Sec. 4.4. As shown in Fig. 2, the $F_{HR}$ serves as the source of contour flow, providing high-frequency contour detail features for the following processes. Correspondingly, the $F_{LR}$ serves as the source of semantic flow, providing low-frequency semantic information. Then, the $F_{HR}$ and $F_{LR}$ are projected to same dimension $C$ and respectively flattened into high-resolution feature embedding as $F_{HR}^{em}$ and low-resolution feature embedding as $F_{LR}^{em}$ , where $F_{HR}^{em} \\in R^{N\\times B\\times C}$ , $F_{LR}^{em} \\in R^{N\\times B\\times C}$ , and $N = \\frac{H}{8}\\times \\frac{W}{8}$ . Both of them are utilized as inputs for the transformer block.", + "bbox": [ + 75, + 458, + 472, + 760 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Transformer Block for Decoder", + "text_level": 1, + "bbox": [ + 76, + 771, + 356, + 786 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We propose a novel transformer block in the model decoding stage to enhance the ability to capture and extract visual features. It not only does not reduce the focus on low-frequency semantic features but also significantly increases the attention to high-frequency contour features. Its detailed architecture is depicted in Fig. 3. The novel transformer block first uses a Semantic and Contour De", + "bbox": [ + 75, + 794, + 468, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "tector (SCD) to locate and capture high-frequency contour features and low-frequency semantic features. To achieve this, we cascade a cross-attention layer (CA) and a self-attention layer (SA) in the SCD. Then, as shown in Fig. 3, a Contour-Edge Extraction Branch (CEEB) and a Semantic Extraction Branch (SEB) are further built to handle high-frequency contour feature flow and low-frequency semantic feature flow respectively. These branches can separately purify more refined high-frequency contour features and extract more complete low-frequency semantic features. This ensures an outcome that is optimal for both types of features.", + "bbox": [ + 496, + 458, + 890, + 637 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Semantic and Contour Detector (SCD). There is a fact that low-frequency semantic features far outperform high-frequency contour detail features in terms of quantity and distribution density. So, in the global information propagation process, the low-frequency semantic information easily dominates the feature representations of the original transformer block with self-attention module.", + "bbox": [ + 496, + 641, + 890, + 746 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To prevent the loss of high-frequency contour detail information, we first use a cross-attention layer with different resolution features rather than a self-attention layer. The cross-attention layer can capture high-frequency detail features with high semantic correlation in high-resolution features, based on the low-frequency semantic information in low-resolution features. Technically, after getting $F_{HR}^{em}$ and $F_{LR}^{em}$ , we use Layer Norm (LN) processing them to maintain consistency. Due to the sensitivity of the attention mechanism towards positional information, we use absolute and", + "bbox": [ + 496, + 750, + 892, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "3883", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/8b731e62de48786a607962e324bfde9d2bf61d970716c65951847e6993d27c69.jpg", + "image_caption": [ + "Figure 3. Our transformer block. It includes a Semantic and Contour Detector (SCD), a Semantic Extraction Branch (SEB), and a Contour-Edge Extraction Branch (CEEB). The detailed composition of the module is shown in the figure." + ], + "image_footnote": [], + "bbox": [ + 89, + 88, + 459, + 526 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "learnable position encoding (PE) in the SCD. The input $K$ , $Q$ , $V$ of the cross-attention layer (CA) is calculated as", + "bbox": [ + 75, + 618, + 468, + 648 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nK = L N \\left(F _ {H R} ^ {e m}\\right) + P E \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 189, + 657, + 468, + 676 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nQ = L N \\left(F _ {L R} ^ {e m}\\right) + P E \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 191, + 684, + 468, + 700 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nV = L N \\left(F _ {L R} ^ {e m} + F _ {H R} ^ {e m}\\right) \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 187, + 705, + 468, + 722 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "By performing cross-attention calculations using features of different resolutions, it is possible to automatically locate and capture the discrete contours along the portrait and allocate more attention to corresponding positions. The contour-edge features can be derived from the output of the CA.", + "bbox": [ + 75, + 728, + 468, + 816 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nF _ {\\text {c o n t o u r - e d g e}} = C A (K, Q, V) \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 161, + 818, + 468, + 835 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Although CA pays extra attention to contour-edge features, it still captures the overall semantics. By the residual connection between CA and $V$ , both the semantic and contour-edge features of the portrait are enhanced, especially the", + "bbox": [ + 75, + 840, + 468, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "contour-edge features.", + "bbox": [ + 500, + 92, + 651, + 106 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nF _ {\\text {e n h a n c e}} = C A (K, Q, V) + V \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 584, + 116, + 890, + 133 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Since semantic information is still present, we use the self-attention layer (SA) after CA. SA can locate the semantic feature within the portrait and contours, thereby emphasizing the semantic attributes at their respective positions. After using LN to normalize $F_{\\text{enhance}}$ , the input $K', Q', V'$ of the SA is calculated as", + "bbox": [ + 498, + 143, + 893, + 233 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nK ^ {\\prime} = Q ^ {\\prime} = L N \\left(F _ {\\text {e n h a n c e}}\\right) + P E \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 578, + 243, + 890, + 261 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nV ^ {\\prime} = L N \\left(F _ {\\text {e n h a n c e}}\\right) \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 622, + 270, + 890, + 287 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nF _ {\\text {c o n t o u r - s e m a n t i c}} = S A \\left(K ^ {\\prime}, Q ^ {\\prime}, V ^ {\\prime}\\right) + V ^ {\\prime} \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 544, + 292, + 890, + 309 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Through SA filtering the semantic and contour features, the residual connection between $V'$ and the output of SA, we can produce $F_{contour - semantic}$ as the output of the semantic and contour detector.", + "bbox": [ + 498, + 315, + 890, + 373 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Feature Extraction Branches. After SCD, we use multi-layer perceptron (MLP) to build a Contour-Edge Extraction Branch (CEEB) and a Semantic Extraction Branch (SEB). To prevent the loss of high-frequency contour detail information, we aggregate $F_{contour - semantic}$ getting from SCD and $F_{HR}^{em}$ in the contour flow. After getting the merged feature normalized by LN, we employ CEEB to purify it, in order to further extract and refine high-frequency contour details individually. Similarly, to obtain more comprehensive and consistent portrait semantic features separately, we merge $F_{contour - semantic}$ and $F_{LR}^{em}$ in the semantic flow. Then, we normalize the aggregated feature by applying LN and use SEB to filter it.", + "bbox": [ + 498, + 375, + 892, + 571 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nF _ {\\text {c o n t o u r}} = M L P \\left(L N \\left(F _ {H R} ^ {\\text {e m}} + F _ {\\text {c o n t o u r - s e m a n t i c}}\\right)\\right) \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 514, + 580, + 890, + 612 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nF _ {\\text {s e m a n t i c}} = M L P \\left(L N \\left(F _ {L R} ^ {\\text {e m}} + F _ {\\text {c o n t o u r - s e m a n t i c}}\\right)\\right) \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 511, + 612, + 890, + 643 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Prediction Stage", + "text_level": 1, + "bbox": [ + 500, + 652, + 663, + 667 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "After getting key characteristics $F_{semantic} \\in R^{N \\times B \\times C}$ and $F_{contour} \\in R^{N \\times B \\times C}$ from transformer block, where $N = \\frac{H}{8} \\times \\frac{W}{8}$ , we transform them into $F_{semantic}' \\in R^{B \\times C \\times \\frac{H}{8} \\times \\frac{W}{8}}$ and $F_{contour}' \\in R^{B \\times C \\times \\frac{H}{8} \\times \\frac{W}{8}}$ by converting the vector into the matrix. To achieve improved outcomes in portrait segmentation, we propose fusing the two kinds of features and ultimately acquiring the predicted portrait matte from the segmentation head module.", + "bbox": [ + 500, + 674, + 890, + 797 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Fuse. We attain the semantic feature through the residual connection between $F_{LR}^{\\prime}$ and $F_{semantic}^{\\prime}$ of semantic flow. Similarly, we get the contour feature from contour flow and subsequently merge the $F_{semantic}^{\\prime}$ and $F_{contour}^{\\prime}$ to generate the fused features $F_{contour - semantic}^{\\prime}$ .", + "bbox": [ + 498, + 797, + 890, + 876 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nF _ {\\text {s e m a n t i c}} ^ {\\prime} = \\operatorname {C o n v} \\left(F _ {\\text {s e m a n t i c}} ^ {\\prime} + F _ {L R} ^ {\\prime}\\right) \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 562, + 883, + 890, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "3884", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nF _ {\\text {c o n t o u r}} ^ {\\prime} = \\operatorname {C o n v} \\left(F _ {\\text {c o n t o u r}} ^ {\\prime} + F _ {H R} ^ {\\prime}\\right) \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 147, + 89, + 467, + 108 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nF _ {\\text {c o n t o u r - s e m a n t i c}} ^ {\\prime} = \\operatorname {F u s e} \\left(F _ {\\text {s e m a n t i c}} ^ {\\prime} + F _ {\\text {c o n t o u r}} ^ {\\prime}\\right) \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 84, + 113, + 467, + 132 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Segmentation Head. Finally, We gather $F_{contour - semantic}^{\\prime}$ and $F_{enc\\frac{1}{4}}$ with higher resolution to predict portrait matte in 1/4 scale of the input image and upscale it to the original scale through bi-linear interpolation as the output of our model.", + "bbox": [ + 75, + 138, + 467, + 213 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nm a t t e = H e a d \\left(F _ {c o n t o u r - s e m a n t i c} ^ {\\prime} + F _ {e n c \\frac {1}{4}}\\right) \\tag {15}\n$$\n", + "text_format": "latex", + "bbox": [ + 101, + 222, + 467, + 248 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 76, + 258, + 209, + 276 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Dataset and Evaluation", + "text_level": 1, + "bbox": [ + 76, + 284, + 292, + 299 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Dataset. We utilize composite training data from two sources: the foreground image dataset VideoMatte240K (JPEG SD Format)[21] and Adobe Image Matting (AIM)[35]. For the background images, we select BG10K following the approach of BGMv2[21], which is a collection of photographs depicting various life scenes without any human portraits. We select foregrounds from VideoMatte240K-JPEG-SD and AIM, and backgrounds from BG10K to composite image datasets. We split VideoMatte240K-JPEG-SD into 234,982/3,007/2720 image sets with image resolutions of $224 \\times 224$ and $512 \\times 288$ for training, validating, and testing our model. Similarly, BG10K is split into 9000/1000 image sets, and AIM is split into 214/10 image sets for training and testing, where AIM only includes the portrait images with resolutions at $512 \\times 512$ . Finally, we use the aforementioned testing datasets to compare and evaluate the performance of EFormer against other models.", + "bbox": [ + 75, + 306, + 468, + 578 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation. We mainly consider portrait matting accuracy for evaluation. For portrait matting accuracy, we use Mean Absolute Difference (MAD), Mean Squared Error (MSE), Gradient (Grad), and Connectivity (Conn) as evaluation metrics. We also scale MAD, MSE, Grad, and Conn by $10^{3}$ , $10^{3}$ , $10^{-3}$ , and $10^{-3}$ respectively, for convenience of reference. For all these metrics, the lower number represents better performance.", + "bbox": [ + 75, + 580, + 467, + 700 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Implementation Details", + "text_level": 1, + "bbox": [ + 76, + 710, + 294, + 726 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Training Settings. For training the network, we use a single RTX 3090 GPU with batch size at 24. The optimizer is AdamW and the initial learning rate is set to $10^{-4}$ . All models in ablation studies are trained for 25 epochs, with the learning rate decaying by a factor of 0.8 every 5 epochs. Additionally, to augment the data, each input image is subjected to random horizontal flipping.", + "bbox": [ + 75, + 733, + 467, + 839 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Backbone. The backbone we use for the network is the pretrained ResNet50[11], since it is the most frequently used backbone in prior works. We use the implementation and weights from torchvision.", + "bbox": [ + 75, + 839, + 467, + 900 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/2b29dd07027dfc431e484454ec4b4fcbcaf97a2dea64be6672f24670cdbd512d.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelMAD↓MSE↓Grad↓Conn↓
DeepLabv3[4]14.47009.67008.55001.6900
MODnet[17]10.39005.65002.02001.0400
BGMv2[21]4.18581.79341.30110.5326
RVM[22]5.99001.17001.10000.3400
EFormer2.30970.66370.45800.2740
", + "bbox": [ + 513, + 88, + 879, + 195 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/ea643b03dda1a2cfc98ef9e9a8cf8ac9e1fce6990aeb2dc38ccb6b152968a860.jpg", + "table_caption": [ + "Table 1. Comparisons on the test set of VideoMatte240K-JPEG-SD. Bold indicates the best performance among these models under the inputs with the same resolution at $512 \\times 288$ ." + ], + "table_footnote": [], + "table_body": "
ModelMAD↓MSE↓Grad↓Conn↓
DeepLabv3[4]29.6423.7820.177.71
MODnet[17]21.6614.275.375.23
BGMv2[21]44.6139.085.5411.60
RVM[22]14.848.934.353.83
EFormer7.472.132.831.90
", + "bbox": [ + 519, + 260, + 870, + 367 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 2. Comparisons on the test set of AIM. Bold indicates the best performance among these models under the inputs with the same resolution at $512 \\times 512$ .", + "bbox": [ + 498, + 378, + 890, + 419 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Transformer Block for Decoder. The transformer block is based on the transformer decoder of DETR. We continuously use four transformer blocks in the decoder of the network. We set the channel $C = 256$ and the number of attention heads $M = 8$ for each multi-head attention module.", + "bbox": [ + 496, + 443, + 890, + 532 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Prediction Stage. After obtaining the semantic and contour feature maps of the portrait from the stacked transformer blocks, we fuse both of them. Then, the alpha mattes predicted by the segmentation head are upsampled to the original size of the input image with bi-linear interpolation.", + "bbox": [ + 496, + 534, + 890, + 609 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Loss function. To compute losses during the training stage, we use a Binary CrossEntropy Loss function.", + "bbox": [ + 498, + 609, + 888, + 640 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nL _ {b c e} = - (g \\times \\log (m) + (1 - g) \\times \\log (1 - m)) \\tag {16}\n$$\n", + "text_format": "latex", + "bbox": [ + 516, + 648, + 888, + 665 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Where, $m$ is the prediction of the network, $g$ is the ground truth.", + "bbox": [ + 498, + 672, + 890, + 702 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Comparisons to State-of-the-art Methods", + "text_level": 1, + "bbox": [ + 500, + 710, + 851, + 726 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To demonstrate the progressiveness of EFormer, we compare it against the latest state-of-the-art trimap-free portrait matting solutions. These are including DeepLabV3[4] with ResNet101[11] backbone, BGMv2[21] with ResNet50[11] backbone, MODNet[17], and RVM[22]. Our assessment is based on the testing datasets of VideoMatte240K-JPEG-SD and AIM, respectively. As the training files for MODNet are not available, we use its official weights. Furthermore, since BGMv2, DeepLabV3, and RVM have already been trained on all datasets, our method can be compared with them in a fair manner.", + "bbox": [ + 496, + 734, + 890, + 898 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "3885", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/a44cfc197d899a897cdae63917124a1992d807f4b3ea9c1a1cb1af45bf9a7ef0.jpg", + "image_caption": [ + "Figure 4. Visualization of portrait matte predictions from MODnet, BGMv2, and EFormer under challenging image from the test set of VideoMatte240K (JPEG SD Format). Our model shows a better ability to distinguish ambiguous foreground contours, as indicated by the red box. Please zoom in to view the detailed information." + ], + "image_footnote": [], + "bbox": [ + 83, + 88, + 893, + 428 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/d66d74013182b05d955a7fc314b2f295b3303d5a4d703b858ad43c84fa974857.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelMAD↓MSE↓Grad↓Conn↓
DeepLabv3[4]18.158614.76236.00152.8824
MODnet[17]13.14008.96562.60282.1543
BGMv2[21]7.04712.35431.56211.0651
RVM[22]6.13572.26470.65230.2866
EFormer4.13571.42800.37600.1753
", + "bbox": [ + 84, + 501, + 460, + 609 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "As shown in the Tab. 1, Tab. 2 and Tab. 3, our model exhibits superior performance in portrait matte tasks. It has a stronger capability to distinguish the contour of a portrait. And it can finely segment edge details (e.g. fingers), as exemplified in Fig. 4.", + "bbox": [ + 75, + 686, + 468, + 763 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Ablation Studies", + "text_level": 1, + "bbox": [ + 76, + 771, + 240, + 786 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Training Epochs. In the ablation study with an image resolution of $224 \\times 224$ , we conduct 25 training epochs. Every 5 epochs of training, we evaluate the training effect of the model. At the same time, the learning rate is reduced by 0.8 times. The experimental results indicate that our model achieves optimal performance in the 20 epoch, as shown in the Tab. 4.", + "bbox": [ + 75, + 794, + 470, + 898 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/3e01d395210f49963e3779f01492181715d5c54a39bd96f04650def1402a3c47.jpg", + "table_caption": [ + "Table 3. Comparisons on the test set of VideoMatte240K-JPEG-SD. Bold indicates the best performance among these models under the inputs with the same resolution at $224 \\times 224$ ." + ], + "table_footnote": [], + "table_body": "
MAD↓MSE↓Grad↓Conn↓
5-epoch4.39531.48240.40890.1865
10-epoch4.31081.47660.39080.1834
15-epoch4.28671.52380.40870.1833
20-epoch4.13571.42800.37600.1753
25-epoch4.18111.47970.38580.1779
", + "bbox": [ + 532, + 501, + 859, + 609 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 4. Ablation Study on the Training Epochs with the test set of VideoMatte240K-JPEG-SD (resolution at $224 \\times 224$ ). **Bold** indicates the best performance among these models. We use the result of 20-epoch training as a strong baseline model.", + "bbox": [ + 496, + 619, + 893, + 675 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The Functional Role of Each Module. To demonstrate the effectiveness of the model in capturing and extracting contour details and portrait semantics, we analyze the outputs of CA, SA, CEEB, and SEB, and utilize Class Activation Map (CAM) to visualize the attention allocation of each layer's features, as shown in Fig. 5.", + "bbox": [ + 496, + 703, + 890, + 792 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "From the first line of Fig. 5, it can be evident that after the cross-attention layer, the model clearly strengthens its attention to the contours of the portrait. Following the contour prior, the self-attention layer captures the portrait's semantics thoroughly. CEEB and SEB models provide sharper contour and more comprehensive semantic information.", + "bbox": [ + 496, + 794, + 890, + 883 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Comparing the second and first lines of Fig. 5, it is ob", + "bbox": [ + 517, + 885, + 890, + 901 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "3886", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f4a86092d25db2d4a4c9dd6b16240410a2a59c5e7287d4d91926926071d90d7f.jpg", + "image_caption": [ + "Figure 5. We use Grad-CAM to visualize the attention distribution of various features in CA, SA, CEEB, and SEB. The model in the first row contains four complete modules. For comparison, the SA layer and CA layer were removed from the models in the second and third rows, respectively." + ], + "image_footnote": [], + "bbox": [ + 80, + 89, + 467, + 275 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/9284b5286ab0833f03b0deb33b734389a968a9090ebdffccfb98e06530372592.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
CASAMAD↓MSE↓Grad↓Conn↓
✓4.27621.53440.40920.1823
✓4.17101.43540.37730.1769
✓✓4.13571.42800.37600.1753
", + "bbox": [ + 89, + 373, + 455, + 450 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "served that removing the self-attention layer considerably weakens the semantic information extracted by SEB. In the comparison between the third and first lines of Fig. 5, it can be indicated that removing the cross-attention layer significantly weakens the contour information extracted by CEEB.", + "bbox": [ + 75, + 547, + 468, + 625 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "After qualitative analysis, the prediction results of all the above models are evaluated. It is concluded that the model with CA, SA, CEEB, and SEB is both effective and rational, as demonstrated in the Tab. 5.", + "bbox": [ + 75, + 625, + 468, + 684 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "As is well known, when it comes to extracting features for portrait segmentation tasks, semantic information plays a significant role, in contrast, contour details account for a relatively small proportion. Owing to this, as shown in the Tab. 5 and Fig. 5, the model without SA leads to a significant increase in all metrics due to the lack of amount of semantic information, while the model without CA experiences a slight increase in all indicators due to the loss of contour details. In order to fully harness the potential of the vision transformer, we integrate both CA and SA to ensure semantic completeness and enhance attention towards contour details, thereby achieving more comprehensive and fine-grained segmentation results.", + "bbox": [ + 75, + 686, + 468, + 883 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The Selection of $F_{HR}$ and $F_{LR}$ . In our base model,", + "bbox": [ + 76, + 885, + 468, + 901 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/e71b74d17dbb4b9a3f2efbc7a347c6138f07d21bd6f697e21c3171dcf0fce96a.jpg", + "table_caption": [ + "Table 5. Ablation Study on the Functional Role of CA and SA with the test set of VideoMatte240K-JPEG-SD (resolution at $224 \\times 224$ ). **Bold** indicates the best performance among these models." + ], + "table_footnote": [], + "table_body": "
FHRFLRMAD↓MSE↓Grad↓Conn↓
Fenc1/8Fenc1/164.13571.42800.37600.1753
Fenc1/4Fenc1/83.99701.33430.33280.1679
Fenc1/4Fenc1/163.97301.29490.32150.1665
", + "bbox": [ + 500, + 88, + 890, + 170 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 6. Ablation Study on the Selection of $F_{HR}$ and $F_{LR}$ with the test set of VideoMatte240K-JPEG-SD (resolution at $224 \\times 224$ ). Bold indicates the best performance among these models.", + "bbox": [ + 498, + 180, + 890, + 223 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "considering the large resolution of feature maps would bring large computation cost, we use the $F_{enc}^{\\frac{1}{8}}$ as $F_{HR}$ (high-resolution features) and use the $F_{enc}^{\\frac{1}{16}}$ as $F_{LR}$ (low-resolution features). Then the $F_{HR}$ serves as the source of contour flow, providing high-frequency contour detail features. Correspondingly, the $F_{LR}$ serves as the source of semantic flow, providing low-frequency semantic information. As shown in the Tab. 6, after using $F_{enc}^{\\frac{1}{4}}$ with higher resolution as $F_{HR}$ , our model can capture more high-frequency contour detail features and its performance is better. And after using $F_{enc}^{\\frac{1}{16}}$ with lower resolution as $F_{LR}$ , our model can use more concentrated low-frequency semantic information, which makes all indicators optimal.", + "bbox": [ + 496, + 258, + 890, + 455 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 484, + 617, + 500 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we propose a strategy that use of cross-attention module between different resolutions guides the EFormer to autonomously locate and capture high-frequency contour features. Without manual parameter tuning, the proposed method can adaptability capture high-frequency details neglected by previous transformer-based methods. Based on the cross-attention module, we further construct a semantic and contour detector (SCD), which can enhance the transformer's attention to both the high-frequency contour and low-frequency semantic features. Additionally, we build a contour-edge extraction branch (CEEB) and a semantic extraction branch (SEB) to extract finer high-frequency contour details and more comprehensive portrait's low-frequency semantic information. Extensive experiments show that our method outperforms previous portrait matting solutions and exhibits robust performance during testing and inference.", + "bbox": [ + 496, + 515, + 890, + 772 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgement", + "text_level": 1, + "bbox": [ + 500, + 800, + 660, + 816 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The work was jointly supported by the National Natural Science Foundations of China under grant No.62272364, the Teaching Reform Project of Shaanxi Higher Continuing Education under Grant No.21XJZ004.", + "bbox": [ + 498, + 829, + 890, + 887 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "3887", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Yagiz Aksoy, Tunc Ozan Aydin, and Marc Pollefeys. Designing effective inter-pixel information flow for natural image matting. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 29-37, 2017. 3", + "[2] Yagiz Aksoy, Tae-Hyun Oh, Sylvain Paris, Marc Pollefeys, and Wojciech Matusik. Semantic soft segmentation. ACM Transactions on Graphics (TOG), 37(4):1-13, 2018. 3", + "[3] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In European conference on computer vision, pages 213-229. Springer, 2020. 3", + "[4] Liang-Chieh Chen, George Papandreou, Florian Schroff, and Hartwig Adam. Rethinking atrous convolution for semantic image segmentation. arXiv preprint arXiv:1706.05587, 2017. 6, 7", + "[5] Bowen Cheng, Alex Schwing, and Alexander Kirillov. Per-pixel classification is not all you need for semantic segmentation. Advances in Neural Information Processing Systems, 34:17864-17875, 2021. 3", + "[6] Bowen Cheng, Ishan Misra, Alexander G Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1290–1299, 2022. 3", + "[7] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 2", + "[8] Dazhao Du, Bing Su, Yu Li, Zhongang Qi, Lingyu Si, and Ying Shan. Efficient u-transformer with boundary-aware loss for action segmentation. arXiv preprint arXiv:2205.13425, 2022. 3", + "[9] Xiaoxue Feng, Xiaohui Liang, and Zili Zhang. A cluster sampling method for image matting via sparse coding. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part II 14, pages 204-219. Springer, 2016. 3", + "[10] Kaiming He, Christoph Rhemann, Carsten Rother, Xiaou Tang, and Jian Sun. A global sampling method for alpha matting. In CVPR 2011, pages 2049-2056. IEEE, 2011. 3", + "[11] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 6", + "[12] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In Proceedings of the IEEE international conference on computer vision, pages 2961-2969, 2017. 3", + "[13] Jitesh Jain, Anukriti Singh, Nikita Orlov, Zilong Huang, Jiachen Li, Steven Walton, and Humphrey Shi. Semask: Semantically masked transformers for semantic segmentation. arXiv preprint arXiv:2112.12782, 2021. 3", + "[14] Jubin Johnson, Ehsan Shahrian Varnousfaderani, Hisham Cholakkal, and Deepu Rajan. Sparse coding for alpha mat" + ], + "bbox": [ + 78, + 114, + 470, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "ting. IEEE Transactions on Image Processing, 25(7):3032-3043, 2016. 3", + "[15] Levent Karacan, Aykut Erdem, and Erkut Erdem. Image matting with kl-divergence based sparse sampling. In Proceedings of the IEEE international conference on computer vision, pages 424-432, 2015. 3", + "[16] Lei Ke, Martin Danelljan, Xia Li, Yu-Wing Tai, Chi-Keung Tang, and Fisher Yu. Mask transfiner for high-quality instance segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4412–4421, 2022. 3", + "[17] Zhanghan Ke, Jiayu Sun, Kaican Li, Qiong Yan, and Rynson WH Lau. Modnet: Real-time trimap-free portrait matting via objective decomposition. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1140-1147, 2022. 2, 3, 6, 7", + "[18] Alexander Kirillov, Yuxin Wu, Kaiming He, and Ross Girshick. Pointrend: Image segmentation as rendering. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9799-9808, 2020. 3", + "[19] Anat Levin, Alex Rav-Acha, and Dani Lischinski. Spectral matting. IEEE transactions on pattern analysis and machine intelligence, 30(10):1699-1712, 2008. 3", + "[20] Yaoyi Li and Hongtao Lu. Natural image matting via guided contextual attention. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 11450-11457, 2020. 1", + "[21] Shanchuan Lin, Andrey Ryabtsev, Soumyadip Sengupta, Brian L Curless, Steven M Seitz, and Ira Kemelmacher-Shlizerman. Real-time high-resolution background matting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8762-8771, 2021. 2, 3, 6, 7", + "[22] Shanchuan Lin, Linjie Yang, Imran Saleemi, and Soumyadip Sengupta. Robust high-resolution video matting with temporal guidance. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 238-247, 2022. 2, 3, 6, 7", + "[23] Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dólar. Focal loss for dense object detection. In Proceedings of the IEEE international conference on computer vision, pages 2980-2988, 2017. 2, 3", + "[24] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10012-10022, 2021. 2", + "[25] Namuk Park and Songkuk Kim. How do vision transformers work? arXiv preprint arXiv:2202.06709, 2022. 2, 3", + "[26] Olivier Petit, Nicolas Thome, Clement Rambour, Loic Themyr, Toby Collins, and Luc Soler. U-net transformer: Self and cross attention for medical image segmentation. In Machine Learning in Medical Imaging: 12th International Workshop, MLMI 2021, Held in Conjunction with MICCAI 2021, Strasbourg, France, September 27, 2021, Proceedings 12, pages 267-276. Springer, 2021. 3", + "[27] René Ranftl, Alexey Bochkovskiy, and Vladlen Koltun. Vision transformers for dense prediction. In Proceedings of" + ], + "bbox": [ + 503, + 92, + 890, + 901 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "3888", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "the IEEE/CVF international conference on computer vision, pages 12179-12188, 2021.3", + "[28] Soumyadip Sengupta, Vivek Jayaram, Brian Curless, Steven M Seitz, and Ira Kemelmacher-Shlizerman. Background matting: The world is your green screen. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2291-2300, 2020. 2, 3", + "[29] Chenyang Si, Weihao Yu, Pan Zhou, Yichen Zhou, Xinchao Wang, and Shuicheng Yan. Inception transformer. Advances in Neural Information Processing Systems, 35:23495-23509, 2022. 2, 3", + "[30] Yanan Sun, Guanzhi Wang, Qiao Gu, Chi-Keung Tang, and Yu-Wing Tai. Deep video matting via spatio-temporal alignment and aggregation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6975–6984, 2021. 1", + "[31] Huiyu Wang, Yukun Zhu, Hartwig Adam, Alan Yuille, and Liang-Chieh Chen. Max-deeplab: End-to-end panoptic segmentation with mask transformers. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5463-5474, 2021. 3", + "[32] Jian Wang, Chenhui Gou, Qiman Wu, Haocheng Feng, Junyu Han, Errui Ding, and Jingdong Wang. Rrformer: Efficient design for real-time semantic segmentation with transformer. Advances in Neural Information Processing Systems, 35:7423-7436, 2022. 3", + "[33] Yu Wang, Yi Niu, Peiyong Duan, Jianwei Lin, and Yuanjie Zheng. Deep propagation based image matting. In IJCAI, pages 999-1006, 2018. 1", + "[34] Enze Xie, Wenhai Wang, Zhiding Yu, Anima Anandkumar, Jose M Alvarez, and Ping Luo. Segformer: Simple and efficient design for semantic segmentation with transformers. Advances in Neural Information Processing Systems, 34:12077-12090, 2021. 3", + "[35] Ning Xu, Brian Price, Scott Cohen, and Thomas Huang. Deep image matting. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2970-2979, 2017. 1, 2, 6", + "[36] Haichao Yu, Ning Xu, Zilong Huang, Yuqian Zhou, and Humphrey Shi. High-resolution deep image matting. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 3217-3224, 2021. 3", + "[37] Yunke Zhang, Chi Wang, Miaomiao Cui, Peiran Ren, Xuansong Xie, Xian-Sheng Hua, Hujun Bao, Qixing Huang, and Weiwei Xu. Attention-guided temporally coherent video object matting. In Proceedings of the 29th ACM International Conference on Multimedia, pages 5128-5137, 2021. 3", + "[38] Sixiao Zheng, Jiachen Lu, Hengshuang Zhao, Xiatian Zhu, Zekun Luo, Yabiao Wang, Yanwei Fu, Jianfeng Feng, Tao Xiang, Philip HS Torr, et al. Rethinking semantic segmentation from a sequence-to-sequence perspective with transformers. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6881-6890, 2021. 3", + "[39] Bingke Zhu, Yingying Chen, Jinqiao Wang, Si Liu, Bo Zhang, and Ming Tang. Fast deep matting for portrait animation on mobile phone. In Proceedings of the 25th ACM international conference on Multimedia, pages 297-305, 2017. 3" + ], + "bbox": [ + 78, + 92, + 468, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "3889", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2024/EFormer_ Enhanced Transformer towards Semantic-Contour Features of Foreground for Portraits Matting/652e59cf-51c6-401e-a44e-7a8d54a66599_model.json b/2024/EFormer_ Enhanced Transformer towards Semantic-Contour Features of Foreground for Portraits Matting/652e59cf-51c6-401e-a44e-7a8d54a66599_model.json new file mode 100644 index 0000000000000000000000000000000000000000..f7663b023577ba541956591224ee5450101519fe --- /dev/null +++ b/2024/EFormer_ Enhanced Transformer towards Semantic-Contour Features of Foreground for Portraits Matting/652e59cf-51c6-401e-a44e-7a8d54a66599_model.json @@ -0,0 +1,2035 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.115, + 0.13, + 0.856, + 0.177 + ], + "angle": 0, + "content": "EFormer: Enhanced Transformer towards Semantic-Contour Features of Foreground for Portraits Matting" + }, + { + "type": "text", + "bbox": [ + 0.26, + 0.204, + 0.71, + 0.223 + ], + "angle": 0, + "content": "Zitao Wang Qiguang Miao* Yue Xi* Peipei Zhao" + }, + { + "type": "text", + "bbox": [ + 0.41, + 0.225, + 0.56, + 0.239 + ], + "angle": 0, + "content": "Xidian University, China" + }, + { + "type": "text", + "bbox": [ + 0.22, + 0.242, + 0.741, + 0.257 + ], + "angle": 0, + "content": "{21181214174@stu.，qgmiao@，xiyue@，zhpp2023@}xidian.edu.cn" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.292, + 0.314, + 0.308 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.324, + 0.474, + 0.778 + ], + "angle": 0, + "content": "The portrait matting task aims to extract an alpha matte with complete semantics and finely detailed contours. In comparison to CNN-based approaches, transformers with self-attention module have a better capacity to capture long-range dependencies and low-frequency semantic information of a portrait. However, recent research shows that the self-attention mechanism struggles with modeling high-frequency contour information and capturing fine contour details, which can lead to bias while predicting the portrait's contours. To deal with this issue, we propose EFormer to enhance the model's attention towards both the low-frequency semantic and high-frequency contour features. For the high-frequency contours, our research demonstrates that cross-attention module between different resolutions can guide our model to allocate attention appropriately to these contour regions. Supported by this, we can successfully extract the high-frequency detail information around the portrait's contours, which were previously ignored by self-attention. Based on the cross-attention module, we further build a semantic and contour detector (SCD) to accurately capture both the low-frequency semantic and high-frequency contour features. And we design a contour-edge extraction branch and semantic extraction branch to extract refined high-frequency contour features and complete low-frequency semantic information, respectively. Finally, we fuse the two kinds of features and leverage the segmentation head to generate a predicted portrait matte. Experiments on VideoMatte240K (JPEG SD Format) and Adobe Image Matting (AIM) datasets demonstrate that EFormer outperforms previous portrait matte methods." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.806, + 0.21, + 0.822 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.832, + 0.47, + 0.879 + ], + "angle": 0, + "content": "Portrait matting aims to extract a precise alpha matte of persons from natural images. It has gathered great momentum in vision applications, including but not limited to photo" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.291, + 0.892, + 0.474 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.519, + 0.478, + 0.892, + 0.657 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.671, + 0.894, + 0.77 + ], + "angle": 0, + "content": "Figure 1. (I) Visualization with Grad-CAM of output from each layer within different transformer blocks. (a) The features captured from Self-attention and MLP layers in the original transformer block. (b) In comparison, the features captured from each layer in the proposed transformer block. (II) Comparisons of different transformer blocks. (c) The original transformer block in ViT. (d) The proposed transformer block for portrait matte." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.795, + 0.747, + 0.81 + ], + "angle": 0, + "content": "editing and background replacement." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.811, + 0.894, + 0.903 + ], + "angle": 0, + "content": "Existing methods[20, 30, 33, 35] attempt to estimate a portrait matte with Convolutional Neural Networks (CNNs) under the guidance of a pre-determined trimap, which is a three-class map that indicates the foreground, background, and unknown region. However, it is time-consuming and labor-intensive to obtain abundant trimaps by manually an-" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.887, + 0.228, + 0.901 + ], + "angle": 0, + "content": "*Corresponding authors." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3880" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.258 + ], + "angle": 0, + "content": "notating images. Some methods eliminate the trimap, such as background matting series[21, 28] uses a separate background image instead of the trimap as an auxiliary reference for model prediction. However, in practical cases, it is necessary to provide two aligned images for the model, which include one with only the background and the other one including people. An alternative solution[17, 22, 23] is to rely solely on RGB images as input for predicting portrait matte. Although the previous models have made strides in portrait matte, there is a constraint on the receptive field in CNN-based approaches." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.26, + 0.473, + 0.412 + ], + "angle": 0, + "content": "In contrast, Vision Transformer (ViT)[7] and Swin Transformer (Swin) [24] for semantic segmentation employ a self-attention mechanism, which conducts interactions among all pixels in an image to obtain a global receptive field. Previous research [25] shows that ViT and its variants can capture the low-frequency components in images effectively, such as global shapes and structures of a scene or object. However, they usually neglect high-frequency ones in images, such as edges and textures. They work similarly to low-pass filters." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.413, + 0.473, + 0.625 + ], + "angle": 0, + "content": "As shown in Fig. 1(a), we can observe the transformer block with self-attention clearly pays little attention to the portrait's high-frequency contour regions. We believe that this phenomenon is attributed to the excessive propagation of global information because global attention is performed in each transformer block during the calculation of multiple blocks. So, the global information is propagated throughout the model in this iterative calculation, ultimately leading it to only focus on low-frequency semantic features with stronger clustering in portraits. Over time, the model gradually ignores the discrete high-frequency detail features around the contours. Therefore, it is desired to capture low- and high-frequency components simultaneously to improve the segmentation performance." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.627, + 0.473, + 0.824 + ], + "angle": 0, + "content": "To achieve this goal, Inception Transformer (IFormer)[29] splits all channels into a convolutional path and self-attention path, respectively, to capture high and low frequencies. However, it is required for a channel ratio of each block at different levels in the IFormer. These channel ratios typically require manual adjustment, resulting in significant uncertainty when aiming for optimal model performance. To enhance the model's stability and adaptive capability, we propose a novel architecture EFormer without manually adjusted parameters. In EFormer, we use the cross-attention layer between different resolution features, as shown in Fig. 1(d), which is replaced by convolution to handle high-frequency detail features." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.826, + 0.473, + 0.903 + ], + "angle": 0, + "content": "Our research further demonstrates that the cross-attention module between different resolutions can formulate a more reasonable attention allocation mechanism, which can accurately capture contour features, as shown in Fig. 1(b). Particularly, EFormer with a cross-attention layer" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.258 + ], + "angle": 0, + "content": "not only can not lose the semantic information within the portrait but also can successfully capture the high-frequency details around the portrait's contours, which were previously ignored by self-attention. It is worth noting that high-frequency components only around the boundary between foreground and background are critical to the improvement of segmentation performance. Therefore, we follow a coarse-to-fine manner. Initially, EFormer needs to locate all high-frequency regions in the image. Then, it filters and extracts the high-frequency regions around the foreground contours." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.259, + 0.893, + 0.621 + ], + "angle": 0, + "content": "Based on the cross-attention module, we design a semantic and contour detector (SCD) in EFormer, which cascades the cross-attention layer and self-attention layer to sequentially locate the contour and semantic features. Firstly, the model adjusts its attention to contour features with the guidance of a cross-attention module. Secondly, the self-attention module captures semantic information within a portrait's contours and reverses filtering and correcting the contour information output by the cross-attention layer. This allows the model to gradually capture and match the high-frequency contour information and the low-frequency semantic information. Then, we use multi-layer perceptron (MLP) to build the contour-edge extraction branch (CEEB) and the semantic extraction branch (SEB), respectively. With this support, we independently purify the contour flow and the semantic flow to obtain finer contour information and more comprehensive portrait semantic information, as shown in Fig. 1(b). Finally, we fuse the contour features and semantic features and send them to the segmentation head to estimate the portrait matte. We perform extensive experiments on VideoMatte240K (JPEG SD Format)[21], Adobe Image Matting (AIM)[35], and BG10K. Comparisons with other state-of-the-art models and ablation studies verify that our model performs better than previous works." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.622, + 0.836, + 0.637 + ], + "angle": 0, + "content": "To summarize, our contributions are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.639, + 0.892, + 0.683 + ], + "angle": 0, + "content": "- We propose EFormer: an approach that can enhance the transformer's attention towards both semantic and contour features of the foreground." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.684, + 0.892, + 0.743 + ], + "angle": 0, + "content": "- We further build a semantic and contour detector (SCD) to accurately capture the semantic and contour features and design two separate extraction branches to purify the contour flow and the semantic flow." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.744, + 0.892, + 0.805 + ], + "angle": 0, + "content": "- We leverage the cross-attention module between different resolution features in transformer block to autonomously capture and extract high-frequency detail features, which are sparsely distributed along the foreground contour." + }, + { + "type": "list", + "bbox": [ + 0.498, + 0.639, + 0.892, + 0.805 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.821, + 0.651, + 0.836 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.847, + 0.652, + 0.864 + ], + "angle": 0, + "content": "2.1. Image Matting" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Image matting is an essential task in the field of computer vision that aims to accurately estimate the foreground in an" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "3881" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.138 + ], + "angle": 0, + "content": "image. Mathematically speaking, an image \\(I\\) is a combination of an unknown foreground image \\(F\\) and a background image \\(B\\) with a probability coefficient alpha matte maps \\(\\alpha\\)." + }, + { + "type": "equation", + "bbox": [ + 0.2, + 0.148, + 0.469, + 0.165 + ], + "angle": 0, + "content": "\\[\nI = \\alpha F + (1 - \\alpha) B \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.175, + 0.469, + 0.265 + ], + "angle": 0, + "content": "Previous image matting solutions have predominantly focused on low-level features, such as color cues[9, 10, 14, 15] or propagation[1, 2, 19], to distinguish the transition areas between foregrounds and backgrounds. However, such traditional matting algorithms commonly struggle to perform properly in complex scenes." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.266, + 0.47, + 0.476 + ], + "angle": 0, + "content": "With the significant advancements making in deep learning, numerous methods based on convolutional neural networks (CNNs) are proposed, leading to notable successes. Some approaches[36, 37, 39] incorporate auxiliary trimap supervisions to enhance matting performance, while other methods[22, 23] leverage trimap-free solutions to estimate alpha matts from image feature maps using an end-to-end segmentation network. In the background matting series[21, 28], an auxiliary input of the background image is used to compute alpha matte. Furthermore, MODNet[17] proposes an end-to-end manner without any auxiliary input. Although prior works based on CNN kernels have improved the accuracy of image matting, the prediction of models is in the limited receptive field." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.485, + 0.266, + 0.5 + ], + "angle": 0, + "content": "2.2. Vision Transformer" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.509, + 0.469, + 0.703 + ], + "angle": 0, + "content": "Different from CNNs, vision transformers with self-attention mechanisms can capture long-term dependencies. Therefore, in the realm of computer vision, vision transformers have gained significant attention in recent times. For pixel-level prediction tasks, such as segmentation, SETR[38], SegFormer[34] and DPT[27] apply transformer as an encoder to attain feature maps of images. Their performances demonstrate the transformer with self-attention is capable of building complete contextual information through global interactions. However, recent studies[25, 29] indicate that transformer \"yet is incompetent in capturing high frequencies that predominantly convey local information\"." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.705, + 0.469, + 0.87 + ], + "angle": 0, + "content": "Compared to self-attention, cross-attention has more potential to induce customized features. Max-Deeplab[31], MaskFormer[5, 6] and SeMask[13] use query-based methods through cross-attention, which are inspired by DETR[3]. They view segmentation as a set prediction problem. U-Transformer[26] and EUT[8] modify the cross-attention in the original transformer to leverage the information from the encoder, allowing a fine spatial recovery in the decoder. The cross-resolution attention employed by RTFormer[32] enables the gathering of comprehensive contextual information for high-resolution features." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.469, + 0.901 + ], + "angle": 0, + "content": "In this work, we propose a strategy that using of crossattention between different resolutions guides the model to" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.182 + ], + "angle": 0, + "content": "autonomously locate and capture high-frequency features, particularly those near the portrait's contours. Contrary to IFormer[29], this method eliminates the need for predefined parameters to enhance the model's adaptive capabilities. At the same time, this approach can significantly improve the model's performance in capturing contour details." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.192, + 0.761, + 0.208 + ], + "angle": 0, + "content": "2.3. Refinement for Segmentation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.215, + 0.892, + 0.29 + ], + "angle": 0, + "content": "In the field of portrait segmentation, refining the high-frequency details near the contours is crucial. Many prior studies on segmentation refinement depend on convolutional networks or MLPs that are designed specifically for this purpose." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.291, + 0.893, + 0.516 + ], + "angle": 0, + "content": "PointRend[18], based on the coarse predictions of Mask R-CNN[12], calculates confidence scores as the point-wise uncertainty measure. And it focuses on the points with the higher uncertainty according to set hyperparameters. Finally, PointRend uses a shared MLP to refine the labels of the selected points. Although PointRend has implemented segmentation refinement to some extent, the model still requires manual adjustment of certain hyperparameters to assist in filtering uncertain points. Transfiner[16] processes detected error-prone and incoherent regions with transformer, which include high-frequency regions strewn along object boundaries. However, during this procedure, excessive focus is placed on the local details and incoherent regions, without effectively associating the local features of the object with the overall semantic information." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.517, + 0.892, + 0.669 + ], + "angle": 0, + "content": "We believe that locating effective detail distribution areas requires information interaction between local details and global semantics. To achieve this, we design a semantic and contour detector that adopts a cascade structure of cross-attention and self-attention. This approach uses global semantic information garnered from self-attention to autonomously select out local details situated at the borders of the contours. On this base, we further utilize the MLP-based extraction branch to purify contour details, thereby significantly enhancing the effect of refinement." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.682, + 0.579, + 0.697 + ], + "angle": 0, + "content": "3. Model" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.707, + 0.892, + 0.768 + ], + "angle": 0, + "content": "In this section, we introduce the process of establishing the complete model architecture of EFormer, which comprises the backbone for encoder, the transformer block for decoder, and the prediction stage, as shown in Fig. 2." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.777, + 0.71, + 0.792 + ], + "angle": 0, + "content": "3.1. Backbone for Encoder" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.799, + 0.892, + 0.901 + ], + "angle": 0, + "content": "The CNN backbone takes RGB images \\( I \\in \\mathbb{R}^{B \\times 3 \\times H \\times W} \\) as input, where \\( B \\) is the number of images, to obtain a pyramid of feature maps \\( F = \\left\\{F_{enc\\frac{1}{4}}, F_{enc\\frac{1}{8}}, F_{enc\\frac{1}{16}}\\right\\} \\) with dimension \\( F_{enc\\frac{1}{i}} \\in \\mathbb{R}^{B \\times C_i \\times \\frac{H}{i} \\times \\frac{W}{i}} \\). As we all know, in the pyramid of feature maps, shallow features have high resolution and rich details, while deep features have low reso" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3882" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.1, + 0.087, + 0.875, + 0.395 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.404, + 0.893, + 0.434 + ], + "angle": 0, + "content": "Figure 2. The Architecture of the EFormer. It contains a backbone for encoder, a transformer block for decoder, and the prediction stage, where the transformer block includes a semantic and contour detector, a semantic extraction branch, and a contour-edge extraction branch." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.459, + 0.473, + 0.761 + ], + "angle": 0, + "content": "lation but more concentrated semantic information. Based on this, our research makes the cross-attention mechanism to filter the high-frequency contour features from high-resolution features, using low-frequency semantic information in low-resolution features. However, since the large resolution of feature maps would bring large computation cost, we use the \\( F_{enc\\frac{1}{8}} \\) as \\( F_{HR} \\) (high-resolution features) and use the \\( F_{enc\\frac{1}{16}} \\) as \\( F_{LR} \\) (low-resolution features) in our base model. The impact of other choices of \\( F_{HR} \\) and \\( F_{LR} \\) on model performance is demonstrated in ablation experiments Sec. 4.4. As shown in Fig. 2, the \\( F_{HR} \\) serves as the source of contour flow, providing high-frequency contour detail features for the following processes. Correspondingly, the \\( F_{LR} \\) serves as the source of semantic flow, providing low-frequency semantic information. Then, the \\( F_{HR} \\) and \\( F_{LR} \\) are projected to same dimension \\( C \\) and respectively flattened into high-resolution feature embedding as \\( F_{HR}^{em} \\) and low-resolution feature embedding as \\( F_{LR}^{em} \\), where \\( F_{HR}^{em} \\in R^{N\\times B\\times C} \\), \\( F_{LR}^{em} \\in R^{N\\times B\\times C} \\), and \\( N = \\frac{H}{8}\\times \\frac{W}{8} \\). Both of them are utilized as inputs for the transformer block." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.772, + 0.357, + 0.787 + ], + "angle": 0, + "content": "3.2. Transformer Block for Decoder" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.795, + 0.47, + 0.901 + ], + "angle": 0, + "content": "We propose a novel transformer block in the model decoding stage to enhance the ability to capture and extract visual features. It not only does not reduce the focus on low-frequency semantic features but also significantly increases the attention to high-frequency contour features. Its detailed architecture is depicted in Fig. 3. The novel transformer block first uses a Semantic and Contour De" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.459, + 0.892, + 0.638 + ], + "angle": 0, + "content": "tector (SCD) to locate and capture high-frequency contour features and low-frequency semantic features. To achieve this, we cascade a cross-attention layer (CA) and a self-attention layer (SA) in the SCD. Then, as shown in Fig. 3, a Contour-Edge Extraction Branch (CEEB) and a Semantic Extraction Branch (SEB) are further built to handle high-frequency contour feature flow and low-frequency semantic feature flow respectively. These branches can separately purify more refined high-frequency contour features and extract more complete low-frequency semantic features. This ensures an outcome that is optimal for both types of features." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.642, + 0.892, + 0.747 + ], + "angle": 0, + "content": "Semantic and Contour Detector (SCD). There is a fact that low-frequency semantic features far outperform high-frequency contour detail features in terms of quantity and distribution density. So, in the global information propagation process, the low-frequency semantic information easily dominates the feature representations of the original transformer block with self-attention module." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.893, + 0.901 + ], + "angle": 0, + "content": "To prevent the loss of high-frequency contour detail information, we first use a cross-attention layer with different resolution features rather than a self-attention layer. The cross-attention layer can capture high-frequency detail features with high semantic correlation in high-resolution features, based on the low-frequency semantic information in low-resolution features. Technically, after getting \\( F_{HR}^{em} \\) and \\( F_{LR}^{em} \\), we use Layer Norm (LN) processing them to maintain consistency. Due to the sensitivity of the attention mechanism towards positional information, we use absolute and" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3883" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.091, + 0.089, + 0.46, + 0.527 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.538, + 0.47, + 0.595 + ], + "angle": 0, + "content": "Figure 3. Our transformer block. It includes a Semantic and Contour Detector (SCD), a Semantic Extraction Branch (SEB), and a Contour-Edge Extraction Branch (CEEB). The detailed composition of the module is shown in the figure." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.619, + 0.47, + 0.65 + ], + "angle": 0, + "content": "learnable position encoding (PE) in the SCD. The input \\( K \\), \\( Q \\), \\( V \\) of the cross-attention layer (CA) is calculated as" + }, + { + "type": "equation", + "bbox": [ + 0.19, + 0.659, + 0.469, + 0.677 + ], + "angle": 0, + "content": "\\[\nK = L N \\left(F _ {H R} ^ {e m}\\right) + P E \\tag {2}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.192, + 0.685, + 0.469, + 0.702 + ], + "angle": 0, + "content": "\\[\nQ = L N \\left(F _ {L R} ^ {e m}\\right) + P E \\tag {3}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.188, + 0.707, + 0.469, + 0.723 + ], + "angle": 0, + "content": "\\[\nV = L N \\left(F _ {L R} ^ {e m} + F _ {H R} ^ {e m}\\right) \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.729, + 0.47, + 0.817 + ], + "angle": 0, + "content": "By performing cross-attention calculations using features of different resolutions, it is possible to automatically locate and capture the discrete contours along the portrait and allocate more attention to corresponding positions. The contour-edge features can be derived from the output of the CA." + }, + { + "type": "equation", + "bbox": [ + 0.162, + 0.819, + 0.469, + 0.836 + ], + "angle": 0, + "content": "\\[\nF _ {\\text {c o n t o u r - e d g e}} = C A (K, Q, V) \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.841, + 0.47, + 0.902 + ], + "angle": 0, + "content": "Although CA pays extra attention to contour-edge features, it still captures the overall semantics. By the residual connection between CA and \\(V\\), both the semantic and contour-edge features of the portrait are enhanced, especially the" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.093, + 0.652, + 0.107 + ], + "angle": 0, + "content": "contour-edge features." + }, + { + "type": "equation", + "bbox": [ + 0.585, + 0.117, + 0.892, + 0.134 + ], + "angle": 0, + "content": "\\[\nF _ {\\text {e n h a n c e}} = C A (K, Q, V) + V \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.144, + 0.894, + 0.234 + ], + "angle": 0, + "content": "Since semantic information is still present, we use the self-attention layer (SA) after CA. SA can locate the semantic feature within the portrait and contours, thereby emphasizing the semantic attributes at their respective positions. After using LN to normalize \\( F_{\\text{enhance}} \\), the input \\( K', Q', V' \\) of the SA is calculated as" + }, + { + "type": "equation", + "bbox": [ + 0.58, + 0.244, + 0.892, + 0.262 + ], + "angle": 0, + "content": "\\[\nK ^ {\\prime} = Q ^ {\\prime} = L N \\left(F _ {\\text {e n h a n c e}}\\right) + P E \\tag {7}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.624, + 0.271, + 0.891, + 0.288 + ], + "angle": 0, + "content": "\\[\nV ^ {\\prime} = L N \\left(F _ {\\text {e n h a n c e}}\\right) \\tag {8}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.545, + 0.293, + 0.891, + 0.31 + ], + "angle": 0, + "content": "\\[\nF _ {\\text {c o n t o u r - s e m a n t i c}} = S A \\left(K ^ {\\prime}, Q ^ {\\prime}, V ^ {\\prime}\\right) + V ^ {\\prime} \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.316, + 0.892, + 0.374 + ], + "angle": 0, + "content": "Through SA filtering the semantic and contour features, the residual connection between \\( V' \\) and the output of SA, we can produce \\( F_{contour - semantic} \\) as the output of the semantic and contour detector." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.376, + 0.893, + 0.572 + ], + "angle": 0, + "content": "Feature Extraction Branches. After SCD, we use multi-layer perceptron (MLP) to build a Contour-Edge Extraction Branch (CEEB) and a Semantic Extraction Branch (SEB). To prevent the loss of high-frequency contour detail information, we aggregate \\( F_{contour - semantic} \\) getting from SCD and \\( F_{HR}^{em} \\) in the contour flow. After getting the merged feature normalized by LN, we employ CEEB to purify it, in order to further extract and refine high-frequency contour details individually. Similarly, to obtain more comprehensive and consistent portrait semantic features separately, we merge \\( F_{contour - semantic} \\) and \\( F_{LR}^{em} \\) in the semantic flow. Then, we normalize the aggregated feature by applying LN and use SEB to filter it." + }, + { + "type": "equation", + "bbox": [ + 0.515, + 0.582, + 0.891, + 0.613 + ], + "angle": 0, + "content": "\\[\nF _ {\\text {c o n t o u r}} = M L P \\left(L N \\left(F _ {H R} ^ {\\text {e m}} + F _ {\\text {c o n t o u r - s e m a n t i c}}\\right)\\right) \\tag {10}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.513, + 0.613, + 0.891, + 0.644 + ], + "angle": 0, + "content": "\\[\nF _ {\\text {s e m a n t i c}} = M L P \\left(L N \\left(F _ {L R} ^ {\\text {e m}} + F _ {\\text {c o n t o u r - s e m a n t i c}}\\right)\\right) \\tag {11}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.653, + 0.664, + 0.669 + ], + "angle": 0, + "content": "3.3. Prediction Stage" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.675, + 0.892, + 0.798 + ], + "angle": 0, + "content": "After getting key characteristics \\( F_{semantic} \\in R^{N \\times B \\times C} \\) and \\( F_{contour} \\in R^{N \\times B \\times C} \\) from transformer block, where \\( N = \\frac{H}{8} \\times \\frac{W}{8} \\), we transform them into \\( F_{semantic}' \\in R^{B \\times C \\times \\frac{H}{8} \\times \\frac{W}{8}} \\) and \\( F_{contour}' \\in R^{B \\times C \\times \\frac{H}{8} \\times \\frac{W}{8}} \\) by converting the vector into the matrix. To achieve improved outcomes in portrait segmentation, we propose fusing the two kinds of features and ultimately acquiring the predicted portrait matte from the segmentation head module." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.799, + 0.892, + 0.877 + ], + "angle": 0, + "content": "Fuse. We attain the semantic feature through the residual connection between \\( F_{LR}^{\\prime} \\) and \\( F_{semantic}^{\\prime} \\) of semantic flow. Similarly, we get the contour feature from contour flow and subsequently merge the \\( F_{semantic}^{\\prime} \\) and \\( F_{contour}^{\\prime} \\) to generate the fused features \\( F_{contour - semantic}^{\\prime} \\)." + }, + { + "type": "equation", + "bbox": [ + 0.563, + 0.885, + 0.892, + 0.903 + ], + "angle": 0, + "content": "\\[\nF _ {\\text {s e m a n t i c}} ^ {\\prime} = \\operatorname {C o n v} \\left(F _ {\\text {s e m a n t i c}} ^ {\\prime} + F _ {L R} ^ {\\prime}\\right) \\tag {12}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3884" + } + ], + [ + { + "type": "equation", + "bbox": [ + 0.148, + 0.09, + 0.468, + 0.109 + ], + "angle": 0, + "content": "\\[\nF _ {\\text {c o n t o u r}} ^ {\\prime} = \\operatorname {C o n v} \\left(F _ {\\text {c o n t o u r}} ^ {\\prime} + F _ {H R} ^ {\\prime}\\right) \\tag {13}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.086, + 0.114, + 0.468, + 0.133 + ], + "angle": 0, + "content": "\\[\nF _ {\\text {c o n t o u r - s e m a n t i c}} ^ {\\prime} = \\operatorname {F u s e} \\left(F _ {\\text {s e m a n t i c}} ^ {\\prime} + F _ {\\text {c o n t o u r}} ^ {\\prime}\\right) \\tag {14}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.139, + 0.468, + 0.214 + ], + "angle": 0, + "content": "Segmentation Head. Finally, We gather \\( F_{contour - semantic}^{\\prime} \\) and \\( F_{enc\\frac{1}{4}} \\) with higher resolution to predict portrait matte in 1/4 scale of the input image and upscale it to the original scale through bi-linear interpolation as the output of our model." + }, + { + "type": "equation", + "bbox": [ + 0.102, + 0.223, + 0.468, + 0.249 + ], + "angle": 0, + "content": "\\[\nm a t t e = H e a d \\left(F _ {c o n t o u r - s e m a n t i c} ^ {\\prime} + F _ {e n c \\frac {1}{4}}\\right) \\tag {15}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.26, + 0.21, + 0.277 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.285, + 0.294, + 0.3 + ], + "angle": 0, + "content": "4.1. Dataset and Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.308, + 0.47, + 0.579 + ], + "angle": 0, + "content": "Dataset. We utilize composite training data from two sources: the foreground image dataset VideoMatte240K (JPEG SD Format)[21] and Adobe Image Matting (AIM)[35]. For the background images, we select BG10K following the approach of BGMv2[21], which is a collection of photographs depicting various life scenes without any human portraits. We select foregrounds from VideoMatte240K-JPEG-SD and AIM, and backgrounds from BG10K to composite image datasets. We split VideoMatte240K-JPEG-SD into 234,982/3,007/2720 image sets with image resolutions of \\(224 \\times 224\\) and \\(512 \\times 288\\) for training, validating, and testing our model. Similarly, BG10K is split into 9000/1000 image sets, and AIM is split into 214/10 image sets for training and testing, where AIM only includes the portrait images with resolutions at \\(512 \\times 512\\). Finally, we use the aforementioned testing datasets to compare and evaluate the performance of EFormer against other models." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.581, + 0.468, + 0.701 + ], + "angle": 0, + "content": "Evaluation. We mainly consider portrait matting accuracy for evaluation. For portrait matting accuracy, we use Mean Absolute Difference (MAD), Mean Squared Error (MSE), Gradient (Grad), and Connectivity (Conn) as evaluation metrics. We also scale MAD, MSE, Grad, and Conn by \\(10^{3}\\), \\(10^{3}\\), \\(10^{-3}\\), and \\(10^{-3}\\) respectively, for convenience of reference. For all these metrics, the lower number represents better performance." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.711, + 0.295, + 0.727 + ], + "angle": 0, + "content": "4.2. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.734, + 0.468, + 0.84 + ], + "angle": 0, + "content": "Training Settings. For training the network, we use a single RTX 3090 GPU with batch size at 24. The optimizer is AdamW and the initial learning rate is set to \\(10^{-4}\\). All models in ablation studies are trained for 25 epochs, with the learning rate decaying by a factor of 0.8 every 5 epochs. Additionally, to augment the data, each input image is subjected to random horizontal flipping." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.468, + 0.901 + ], + "angle": 0, + "content": "Backbone. The backbone we use for the network is the pretrained ResNet50[11], since it is the most frequently used backbone in prior works. We use the implementation and weights from torchvision." + }, + { + "type": "table", + "bbox": [ + 0.514, + 0.089, + 0.88, + 0.196 + ], + "angle": 0, + "content": "
ModelMAD↓MSE↓Grad↓Conn↓
DeepLabv3[4]14.47009.67008.55001.6900
MODnet[17]10.39005.65002.02001.0400
BGMv2[21]4.18581.79341.30110.5326
RVM[22]5.99001.17001.10000.3400
EFormer2.30970.66370.45800.2740
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.206, + 0.892, + 0.248 + ], + "angle": 0, + "content": "Table 1. Comparisons on the test set of VideoMatte240K-JPEG-SD. Bold indicates the best performance among these models under the inputs with the same resolution at \\(512 \\times 288\\)." + }, + { + "type": "table", + "bbox": [ + 0.521, + 0.261, + 0.871, + 0.368 + ], + "angle": 0, + "content": "
ModelMAD↓MSE↓Grad↓Conn↓
DeepLabv3[4]29.6423.7820.177.71
MODnet[17]21.6614.275.375.23
BGMv2[21]44.6139.085.5411.60
RVM[22]14.848.934.353.83
EFormer7.472.132.831.90
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.379, + 0.892, + 0.42 + ], + "angle": 0, + "content": "Table 2. Comparisons on the test set of AIM. Bold indicates the best performance among these models under the inputs with the same resolution at \\(512 \\times 512\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.444, + 0.892, + 0.533 + ], + "angle": 0, + "content": "Transformer Block for Decoder. The transformer block is based on the transformer decoder of DETR. We continuously use four transformer blocks in the decoder of the network. We set the channel \\( C = 256 \\) and the number of attention heads \\( M = 8 \\) for each multi-head attention module." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.535, + 0.892, + 0.61 + ], + "angle": 0, + "content": "Prediction Stage. After obtaining the semantic and contour feature maps of the portrait from the stacked transformer blocks, we fuse both of them. Then, the alpha mattes predicted by the segmentation head are upsampled to the original size of the input image with bi-linear interpolation." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.611, + 0.89, + 0.641 + ], + "angle": 0, + "content": "Loss function. To compute losses during the training stage, we use a Binary CrossEntropy Loss function." + }, + { + "type": "equation", + "bbox": [ + 0.517, + 0.649, + 0.89, + 0.666 + ], + "angle": 0, + "content": "\\[\nL _ {b c e} = - (g \\times \\log (m) + (1 - g) \\times \\log (1 - m)) \\tag {16}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.673, + 0.892, + 0.703 + ], + "angle": 0, + "content": "Where, \\( m \\) is the prediction of the network, \\( g \\) is the ground truth." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.712, + 0.852, + 0.727 + ], + "angle": 0, + "content": "4.3. Comparisons to State-of-the-art Methods" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.735, + 0.892, + 0.9 + ], + "angle": 0, + "content": "To demonstrate the progressiveness of EFormer, we compare it against the latest state-of-the-art trimap-free portrait matting solutions. These are including DeepLabV3[4] with ResNet101[11] backbone, BGMv2[21] with ResNet50[11] backbone, MODNet[17], and RVM[22]. Our assessment is based on the testing datasets of VideoMatte240K-JPEG-SD and AIM, respectively. As the training files for MODNet are not available, we use its official weights. Furthermore, since BGMv2, DeepLabV3, and RVM have already been trained on all datasets, our method can be compared with them in a fair manner." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3885" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.084, + 0.089, + 0.895, + 0.429 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.436, + 0.895, + 0.478 + ], + "angle": 0, + "content": "Figure 4. Visualization of portrait matte predictions from MODnet, BGMv2, and EFormer under challenging image from the test set of VideoMatte240K (JPEG SD Format). Our model shows a better ability to distinguish ambiguous foreground contours, as indicated by the red box. Please zoom in to view the detailed information." + }, + { + "type": "table", + "bbox": [ + 0.086, + 0.502, + 0.462, + 0.61 + ], + "angle": 0, + "content": "
ModelMAD↓MSE↓Grad↓Conn↓
DeepLabv3[4]18.158614.76236.00152.8824
MODnet[17]13.14008.96562.60282.1543
BGMv2[21]7.04712.35431.56211.0651
RVM[22]6.13572.26470.65230.2866
EFormer4.13571.42800.37600.1753
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.62, + 0.47, + 0.662 + ], + "angle": 0, + "content": "Table 3. Comparisons on the test set of VideoMatte240K-JPEG-SD. Bold indicates the best performance among these models under the inputs with the same resolution at \\(224 \\times 224\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.688, + 0.47, + 0.764 + ], + "angle": 0, + "content": "As shown in the Tab. 1, Tab. 2 and Tab. 3, our model exhibits superior performance in portrait matte tasks. It has a stronger capability to distinguish the contour of a portrait. And it can finely segment edge details (e.g. fingers), as exemplified in Fig. 4." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.772, + 0.241, + 0.787 + ], + "angle": 0, + "content": "4.4. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.795, + 0.471, + 0.9 + ], + "angle": 0, + "content": "Training Epochs. In the ablation study with an image resolution of \\(224 \\times 224\\), we conduct 25 training epochs. Every 5 epochs of training, we evaluate the training effect of the model. At the same time, the learning rate is reduced by 0.8 times. The experimental results indicate that our model achieves optimal performance in the 20 epoch, as shown in the Tab. 4." + }, + { + "type": "table", + "bbox": [ + 0.534, + 0.502, + 0.86, + 0.61 + ], + "angle": 0, + "content": "
MAD↓MSE↓Grad↓Conn↓
5-epoch4.39531.48240.40890.1865
10-epoch4.31081.47660.39080.1834
15-epoch4.28671.52380.40870.1833
20-epoch4.13571.42800.37600.1753
25-epoch4.18111.47970.38580.1779
" + }, + { + "type": "table_caption", + "bbox": [ + 0.498, + 0.62, + 0.894, + 0.676 + ], + "angle": 0, + "content": "Table 4. Ablation Study on the Training Epochs with the test set of VideoMatte240K-JPEG-SD (resolution at \\(224 \\times 224\\)). **Bold** indicates the best performance among these models. We use the result of 20-epoch training as a strong baseline model." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.704, + 0.892, + 0.794 + ], + "angle": 0, + "content": "The Functional Role of Each Module. To demonstrate the effectiveness of the model in capturing and extracting contour details and portrait semantics, we analyze the outputs of CA, SA, CEEB, and SEB, and utilize Class Activation Map (CAM) to visualize the attention allocation of each layer's features, as shown in Fig. 5." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.795, + 0.892, + 0.885 + ], + "angle": 0, + "content": "From the first line of Fig. 5, it can be evident that after the cross-attention layer, the model clearly strengthens its attention to the contours of the portrait. Following the contour prior, the self-attention layer captures the portrait's semantics thoroughly. CEEB and SEB models provide sharper contour and more comprehensive semantic information." + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.886, + 0.892, + 0.902 + ], + "angle": 0, + "content": "Comparing the second and first lines of Fig. 5, it is ob" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "3886" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.081, + 0.09, + 0.468, + 0.276 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.287, + 0.472, + 0.358 + ], + "angle": 0, + "content": "Figure 5. We use Grad-CAM to visualize the attention distribution of various features in CA, SA, CEEB, and SEB. The model in the first row contains four complete modules. For comparison, the SA layer and CA layer were removed from the models in the second and third rows, respectively." + }, + { + "type": "table", + "bbox": [ + 0.091, + 0.374, + 0.456, + 0.452 + ], + "angle": 0, + "content": "
CASAMAD↓MSE↓Grad↓Conn↓
✓4.27621.53440.40920.1823
✓4.17101.43540.37730.1769
✓✓4.13571.42800.37600.1753
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.462, + 0.47, + 0.517 + ], + "angle": 0, + "content": "Table 5. Ablation Study on the Functional Role of CA and SA with the test set of VideoMatte240K-JPEG-SD (resolution at \\(224 \\times 224\\)). **Bold** indicates the best performance among these models." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.548, + 0.469, + 0.625 + ], + "angle": 0, + "content": "served that removing the self-attention layer considerably weakens the semantic information extracted by SEB. In the comparison between the third and first lines of Fig. 5, it can be indicated that removing the cross-attention layer significantly weakens the contour information extracted by CEEB." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.626, + 0.469, + 0.685 + ], + "angle": 0, + "content": "After qualitative analysis, the prediction results of all the above models are evaluated. It is concluded that the model with CA, SA, CEEB, and SEB is both effective and rational, as demonstrated in the Tab. 5." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.688, + 0.47, + 0.884 + ], + "angle": 0, + "content": "As is well known, when it comes to extracting features for portrait segmentation tasks, semantic information plays a significant role, in contrast, contour details account for a relatively small proportion. Owing to this, as shown in the Tab. 5 and Fig. 5, the model without SA leads to a significant increase in all metrics due to the lack of amount of semantic information, while the model without CA experiences a slight increase in all indicators due to the loss of contour details. In order to fully harness the potential of the vision transformer, we integrate both CA and SA to ensure semantic completeness and enhance attention towards contour details, thereby achieving more comprehensive and fine-grained segmentation results." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.886, + 0.469, + 0.902 + ], + "angle": 0, + "content": "The Selection of \\(F_{HR}\\) and \\(F_{LR}\\). In our base model," + }, + { + "type": "table", + "bbox": [ + 0.501, + 0.089, + 0.892, + 0.171 + ], + "angle": 0, + "content": "
FHRFLRMAD↓MSE↓Grad↓Conn↓
Fenc1/8Fenc1/164.13571.42800.37600.1753
Fenc1/4Fenc1/83.99701.33430.33280.1679
Fenc1/4Fenc1/163.97301.29490.32150.1665
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.181, + 0.892, + 0.224 + ], + "angle": 0, + "content": "Table 6. Ablation Study on the Selection of \\(F_{HR}\\) and \\(F_{LR}\\) with the test set of VideoMatte240K-JPEG-SD (resolution at \\(224 \\times 224\\)). Bold indicates the best performance among these models." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.26, + 0.892, + 0.457 + ], + "angle": 0, + "content": "considering the large resolution of feature maps would bring large computation cost, we use the \\( F_{enc}^{\\frac{1}{8}} \\) as \\( F_{HR} \\) (high-resolution features) and use the \\( F_{enc}^{\\frac{1}{16}} \\) as \\( F_{LR} \\) (low-resolution features). Then the \\( F_{HR} \\) serves as the source of contour flow, providing high-frequency contour detail features. Correspondingly, the \\( F_{LR} \\) serves as the source of semantic flow, providing low-frequency semantic information. As shown in the Tab. 6, after using \\( F_{enc}^{\\frac{1}{4}} \\) with higher resolution as \\( F_{HR} \\), our model can capture more high-frequency contour detail features and its performance is better. And after using \\( F_{enc}^{\\frac{1}{16}} \\) with lower resolution as \\( F_{LR} \\), our model can use more concentrated low-frequency semantic information, which makes all indicators optimal." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.485, + 0.619, + 0.501 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.516, + 0.892, + 0.773 + ], + "angle": 0, + "content": "In this paper, we propose a strategy that use of cross-attention module between different resolutions guides the EFormer to autonomously locate and capture high-frequency contour features. Without manual parameter tuning, the proposed method can adaptability capture high-frequency details neglected by previous transformer-based methods. Based on the cross-attention module, we further construct a semantic and contour detector (SCD), which can enhance the transformer's attention to both the high-frequency contour and low-frequency semantic features. Additionally, we build a contour-edge extraction branch (CEEB) and a semantic extraction branch (SEB) to extract finer high-frequency contour details and more comprehensive portrait's low-frequency semantic information. Extensive experiments show that our method outperforms previous portrait matting solutions and exhibits robust performance during testing and inference." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.801, + 0.661, + 0.818 + ], + "angle": 0, + "content": "Acknowledgement" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.83, + 0.892, + 0.888 + ], + "angle": 0, + "content": "The work was jointly supported by the National Natural Science Foundations of China under grant No.62272364, the Teaching Reform Project of Shaanxi Higher Continuing Education under Grant No.21XJZ004." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3887" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.171 + ], + "angle": 0, + "content": "[1] Yagiz Aksoy, Tunc Ozan Aydin, and Marc Pollefeys. Designing effective inter-pixel information flow for natural image matting. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 29-37, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.173, + 0.472, + 0.214 + ], + "angle": 0, + "content": "[2] Yagiz Aksoy, Tae-Hyun Oh, Sylvain Paris, Marc Pollefeys, and Wojciech Matusik. Semantic soft segmentation. ACM Transactions on Graphics (TOG), 37(4):1-13, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.216, + 0.47, + 0.272 + ], + "angle": 0, + "content": "[3] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In European conference on computer vision, pages 213-229. Springer, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.274, + 0.47, + 0.327 + ], + "angle": 0, + "content": "[4] Liang-Chieh Chen, George Papandreou, Florian Schroff, and Hartwig Adam. Rethinking atrous convolution for semantic image segmentation. arXiv preprint arXiv:1706.05587, 2017. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.33, + 0.47, + 0.385 + ], + "angle": 0, + "content": "[5] Bowen Cheng, Alex Schwing, and Alexander Kirillov. Per-pixel classification is not all you need for semantic segmentation. Advances in Neural Information Processing Systems, 34:17864-17875, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.388, + 0.47, + 0.457 + ], + "angle": 0, + "content": "[6] Bowen Cheng, Ishan Misra, Alexander G Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1290–1299, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.459, + 0.47, + 0.541 + ], + "angle": 0, + "content": "[7] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.544, + 0.47, + 0.598 + ], + "angle": 0, + "content": "[8] Dazhao Du, Bing Su, Yu Li, Zhongang Qi, Lingyu Si, and Ying Shan. Efficient u-transformer with boundary-aware loss for action segmentation. arXiv preprint arXiv:2205.13425, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.601, + 0.47, + 0.67 + ], + "angle": 0, + "content": "[9] Xiaoxue Feng, Xiaohui Liang, and Zili Zhang. A cluster sampling method for image matting via sparse coding. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part II 14, pages 204-219. Springer, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.672, + 0.469, + 0.714 + ], + "angle": 0, + "content": "[10] Kaiming He, Christoph Rhemann, Carsten Rother, Xiaou Tang, and Jian Sun. A global sampling method for alpha matting. In CVPR 2011, pages 2049-2056. IEEE, 2011. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.716, + 0.469, + 0.77 + ], + "angle": 0, + "content": "[11] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.772, + 0.469, + 0.814 + ], + "angle": 0, + "content": "[12] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In Proceedings of the IEEE international conference on computer vision, pages 2961-2969, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.816, + 0.469, + 0.871 + ], + "angle": 0, + "content": "[13] Jitesh Jain, Anukriti Singh, Nikita Orlov, Zilong Huang, Jiachen Li, Steven Walton, and Humphrey Shi. Semask: Semantically masked transformers for semantic segmentation. arXiv preprint arXiv:2112.12782, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.469, + 0.902 + ], + "angle": 0, + "content": "[14] Jubin Johnson, Ehsan Shahrian Varnousfaderani, Hisham Cholakkal, and Deepu Rajan. Sparse coding for alpha mat" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.472, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "ting. IEEE Transactions on Image Processing, 25(7):3032-3043, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.122, + 0.892, + 0.177 + ], + "angle": 0, + "content": "[15] Levent Karacan, Aykut Erdem, and Erkut Erdem. Image matting with kl-divergence based sparse sampling. In Proceedings of the IEEE international conference on computer vision, pages 424-432, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.179, + 0.892, + 0.247 + ], + "angle": 0, + "content": "[16] Lei Ke, Martin Danelljan, Xia Li, Yu-Wing Tai, Chi-Keung Tang, and Fisher Yu. Mask transfiner for high-quality instance segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4412–4421, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.249, + 0.892, + 0.318 + ], + "angle": 0, + "content": "[17] Zhanghan Ke, Jiayu Sun, Kaican Li, Qiong Yan, and Rynson WH Lau. Modnet: Real-time trimap-free portrait matting via objective decomposition. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1140-1147, 2022. 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.32, + 0.892, + 0.375 + ], + "angle": 0, + "content": "[18] Alexander Kirillov, Yuxin Wu, Kaiming He, and Ross Girshick. Pointrend: Image segmentation as rendering. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9799-9808, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.377, + 0.892, + 0.418 + ], + "angle": 0, + "content": "[19] Anat Levin, Alex Rav-Acha, and Dani Lischinski. Spectral matting. IEEE transactions on pattern analysis and machine intelligence, 30(10):1699-1712, 2008. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.42, + 0.892, + 0.462 + ], + "angle": 0, + "content": "[20] Yaoyi Li and Hongtao Lu. Natural image matting via guided contextual attention. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 11450-11457, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.463, + 0.892, + 0.545 + ], + "angle": 0, + "content": "[21] Shanchuan Lin, Andrey Ryabtsev, Soumyadip Sengupta, Brian L Curless, Steven M Seitz, and Ira Kemelmacher-Shlizerman. Real-time high-resolution background matting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8762-8771, 2021. 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.547, + 0.892, + 0.616 + ], + "angle": 0, + "content": "[22] Shanchuan Lin, Linjie Yang, Imran Saleemi, and Soumyadip Sengupta. Robust high-resolution video matting with temporal guidance. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 238-247, 2022. 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.618, + 0.892, + 0.673 + ], + "angle": 0, + "content": "[23] Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dólar. Focal loss for dense object detection. In Proceedings of the IEEE international conference on computer vision, pages 2980-2988, 2017. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.675, + 0.892, + 0.745 + ], + "angle": 0, + "content": "[24] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10012-10022, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.746, + 0.892, + 0.773 + ], + "angle": 0, + "content": "[25] Namuk Park and Songkuk Kim. How do vision transformers work? arXiv preprint arXiv:2202.06709, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.775, + 0.892, + 0.871 + ], + "angle": 0, + "content": "[26] Olivier Petit, Nicolas Thome, Clement Rambour, Loic Themyr, Toby Collins, and Luc Soler. U-net transformer: Self and cross attention for medical image segmentation. In Machine Learning in Medical Imaging: 12th International Workshop, MLMI 2021, Held in Conjunction with MICCAI 2021, Strasbourg, France, September 27, 2021, Proceedings 12, pages 267-276. Springer, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.873, + 0.892, + 0.902 + ], + "angle": 0, + "content": "[27] René Ranftl, Alexey Bochkovskiy, and Vladlen Koltun. Vision transformers for dense prediction. In Proceedings of" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.892, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "3888" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.11, + 0.093, + 0.468, + 0.12 + ], + "angle": 0, + "content": "the IEEE/CVF international conference on computer vision, pages 12179-12188, 2021.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.122, + 0.469, + 0.189 + ], + "angle": 0, + "content": "[28] Soumyadip Sengupta, Vivek Jayaram, Brian Curless, Steven M Seitz, and Ira Kemelmacher-Shlizerman. Background matting: The world is your green screen. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2291-2300, 2020. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.191, + 0.469, + 0.244 + ], + "angle": 0, + "content": "[29] Chenyang Si, Weihao Yu, Pan Zhou, Yichen Zhou, Xinchao Wang, and Shuicheng Yan. Inception transformer. Advances in Neural Information Processing Systems, 35:23495-23509, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.247, + 0.469, + 0.314 + ], + "angle": 0, + "content": "[30] Yanan Sun, Guanzhi Wang, Qiao Gu, Chi-Keung Tang, and Yu-Wing Tai. Deep video matting via spatio-temporal alignment and aggregation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6975–6984, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.317, + 0.469, + 0.385 + ], + "angle": 0, + "content": "[31] Huiyu Wang, Yukun Zhu, Hartwig Adam, Alan Yuille, and Liang-Chieh Chen. Max-deeplab: End-to-end panoptic segmentation with mask transformers. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5463-5474, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.386, + 0.469, + 0.453 + ], + "angle": 0, + "content": "[32] Jian Wang, Chenhui Gou, Qiman Wu, Haocheng Feng, Junyu Han, Errui Ding, and Jingdong Wang. Rrformer: Efficient design for real-time semantic segmentation with transformer. Advances in Neural Information Processing Systems, 35:7423-7436, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.456, + 0.469, + 0.496 + ], + "angle": 0, + "content": "[33] Yu Wang, Yi Niu, Peiyong Duan, Jianwei Lin, and Yuanjie Zheng. Deep propagation based image matting. In IJCAI, pages 999-1006, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.498, + 0.469, + 0.564 + ], + "angle": 0, + "content": "[34] Enze Xie, Wenhai Wang, Zhiding Yu, Anima Anandkumar, Jose M Alvarez, and Ping Luo. Segformer: Simple and efficient design for semantic segmentation with transformers. Advances in Neural Information Processing Systems, 34:12077-12090, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.568, + 0.469, + 0.621 + ], + "angle": 0, + "content": "[35] Ning Xu, Brian Price, Scott Cohen, and Thomas Huang. Deep image matting. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2970-2979, 2017. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.624, + 0.469, + 0.678 + ], + "angle": 0, + "content": "[36] Haichao Yu, Ning Xu, Zilong Huang, Yuqian Zhou, and Humphrey Shi. High-resolution deep image matting. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 3217-3224, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.679, + 0.469, + 0.747 + ], + "angle": 0, + "content": "[37] Yunke Zhang, Chi Wang, Miaomiao Cui, Peiran Ren, Xuansong Xie, Xian-Sheng Hua, Hujun Bao, Qixing Huang, and Weiwei Xu. Attention-guided temporally coherent video object matting. In Proceedings of the 29th ACM International Conference on Multimedia, pages 5128-5137, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.749, + 0.469, + 0.843 + ], + "angle": 0, + "content": "[38] Sixiao Zheng, Jiachen Lu, Hengshuang Zhao, Xiatian Zhu, Zekun Luo, Yabiao Wang, Yanwei Fu, Jianfeng Feng, Tao Xiang, Philip HS Torr, et al. Rethinking semantic segmentation from a sequence-to-sequence perspective with transformers. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6881-6890, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.846, + 0.469, + 0.9 + ], + "angle": 0, + "content": "[39] Bingke Zhu, Yingying Chen, Jinqiao Wang, Si Liu, Bo Zhang, and Ming Tang. Fast deep matting for portrait animation on mobile phone. In Proceedings of the 25th ACM international conference on Multimedia, pages 297-305, 2017. 3" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.093, + 0.469, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.956 + ], + "angle": 0, + "content": "3889" + } + ] +] \ No newline at end of file diff --git a/2024/EFormer_ Enhanced Transformer towards Semantic-Contour Features of Foreground for Portraits Matting/652e59cf-51c6-401e-a44e-7a8d54a66599_origin.pdf b/2024/EFormer_ Enhanced Transformer towards Semantic-Contour Features of Foreground for Portraits Matting/652e59cf-51c6-401e-a44e-7a8d54a66599_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..eb821e528a1c99efd440268311392477a6a883d3 --- /dev/null +++ b/2024/EFormer_ Enhanced Transformer towards Semantic-Contour Features of Foreground for Portraits Matting/652e59cf-51c6-401e-a44e-7a8d54a66599_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:470ef394bca49ec8b7d6948b4f4feedb5f7ea565943a75289eaddda24fbf4a91 +size 7257131 diff --git a/2024/EFormer_ Enhanced Transformer towards Semantic-Contour Features of Foreground for Portraits Matting/full.md b/2024/EFormer_ Enhanced Transformer towards Semantic-Contour Features of Foreground for Portraits Matting/full.md new file mode 100644 index 0000000000000000000000000000000000000000..01d9b76f907e09e336db730a88e6b08fb8de8a67 --- /dev/null +++ b/2024/EFormer_ Enhanced Transformer towards Semantic-Contour Features of Foreground for Portraits Matting/full.md @@ -0,0 +1,318 @@ +# EFormer: Enhanced Transformer towards Semantic-Contour Features of Foreground for Portraits Matting + +Zitao Wang Qiguang Miao* Yue Xi* Peipei Zhao + +Xidian University, China + +{21181214174@stu.，qgmiao@，xiyue@，zhpp2023@}xidian.edu.cn + +# Abstract + +The portrait matting task aims to extract an alpha matte with complete semantics and finely detailed contours. In comparison to CNN-based approaches, transformers with self-attention module have a better capacity to capture long-range dependencies and low-frequency semantic information of a portrait. However, recent research shows that the self-attention mechanism struggles with modeling high-frequency contour information and capturing fine contour details, which can lead to bias while predicting the portrait's contours. To deal with this issue, we propose EFormer to enhance the model's attention towards both the low-frequency semantic and high-frequency contour features. For the high-frequency contours, our research demonstrates that cross-attention module between different resolutions can guide our model to allocate attention appropriately to these contour regions. Supported by this, we can successfully extract the high-frequency detail information around the portrait's contours, which were previously ignored by self-attention. Based on the cross-attention module, we further build a semantic and contour detector (SCD) to accurately capture both the low-frequency semantic and high-frequency contour features. And we design a contour-edge extraction branch and semantic extraction branch to extract refined high-frequency contour features and complete low-frequency semantic information, respectively. Finally, we fuse the two kinds of features and leverage the segmentation head to generate a predicted portrait matte. Experiments on VideoMatte240K (JPEG SD Format) and Adobe Image Matting (AIM) datasets demonstrate that EFormer outperforms previous portrait matte methods. + +# 1. Introduction + +Portrait matting aims to extract a precise alpha matte of persons from natural images. It has gathered great momentum in vision applications, including but not limited to photo + +![](images/c9937fb43e03a7633aee3ae5e531b87101ba94ad97cd2f40462a3b2430c6fa14.jpg) + +![](images/3b06508ac7a09bba239e870787572c7c30df658ab82ad01b9fb750bf1c4e71f8.jpg) +Figure 1. (I) Visualization with Grad-CAM of output from each layer within different transformer blocks. (a) The features captured from Self-attention and MLP layers in the original transformer block. (b) In comparison, the features captured from each layer in the proposed transformer block. (II) Comparisons of different transformer blocks. (c) The original transformer block in ViT. (d) The proposed transformer block for portrait matte. + +editing and background replacement. + +Existing methods[20, 30, 33, 35] attempt to estimate a portrait matte with Convolutional Neural Networks (CNNs) under the guidance of a pre-determined trimap, which is a three-class map that indicates the foreground, background, and unknown region. However, it is time-consuming and labor-intensive to obtain abundant trimaps by manually an- + +notating images. Some methods eliminate the trimap, such as background matting series[21, 28] uses a separate background image instead of the trimap as an auxiliary reference for model prediction. However, in practical cases, it is necessary to provide two aligned images for the model, which include one with only the background and the other one including people. An alternative solution[17, 22, 23] is to rely solely on RGB images as input for predicting portrait matte. Although the previous models have made strides in portrait matte, there is a constraint on the receptive field in CNN-based approaches. + +In contrast, Vision Transformer (ViT)[7] and Swin Transformer (Swin) [24] for semantic segmentation employ a self-attention mechanism, which conducts interactions among all pixels in an image to obtain a global receptive field. Previous research [25] shows that ViT and its variants can capture the low-frequency components in images effectively, such as global shapes and structures of a scene or object. However, they usually neglect high-frequency ones in images, such as edges and textures. They work similarly to low-pass filters. + +As shown in Fig. 1(a), we can observe the transformer block with self-attention clearly pays little attention to the portrait's high-frequency contour regions. We believe that this phenomenon is attributed to the excessive propagation of global information because global attention is performed in each transformer block during the calculation of multiple blocks. So, the global information is propagated throughout the model in this iterative calculation, ultimately leading it to only focus on low-frequency semantic features with stronger clustering in portraits. Over time, the model gradually ignores the discrete high-frequency detail features around the contours. Therefore, it is desired to capture low- and high-frequency components simultaneously to improve the segmentation performance. + +To achieve this goal, Inception Transformer (IFormer)[29] splits all channels into a convolutional path and self-attention path, respectively, to capture high and low frequencies. However, it is required for a channel ratio of each block at different levels in the IFormer. These channel ratios typically require manual adjustment, resulting in significant uncertainty when aiming for optimal model performance. To enhance the model's stability and adaptive capability, we propose a novel architecture EFormer without manually adjusted parameters. In EFormer, we use the cross-attention layer between different resolution features, as shown in Fig. 1(d), which is replaced by convolution to handle high-frequency detail features. + +Our research further demonstrates that the cross-attention module between different resolutions can formulate a more reasonable attention allocation mechanism, which can accurately capture contour features, as shown in Fig. 1(b). Particularly, EFormer with a cross-attention layer + +not only can not lose the semantic information within the portrait but also can successfully capture the high-frequency details around the portrait's contours, which were previously ignored by self-attention. It is worth noting that high-frequency components only around the boundary between foreground and background are critical to the improvement of segmentation performance. Therefore, we follow a coarse-to-fine manner. Initially, EFormer needs to locate all high-frequency regions in the image. Then, it filters and extracts the high-frequency regions around the foreground contours. + +Based on the cross-attention module, we design a semantic and contour detector (SCD) in EFormer, which cascades the cross-attention layer and self-attention layer to sequentially locate the contour and semantic features. Firstly, the model adjusts its attention to contour features with the guidance of a cross-attention module. Secondly, the self-attention module captures semantic information within a portrait's contours and reverses filtering and correcting the contour information output by the cross-attention layer. This allows the model to gradually capture and match the high-frequency contour information and the low-frequency semantic information. Then, we use multi-layer perceptron (MLP) to build the contour-edge extraction branch (CEEB) and the semantic extraction branch (SEB), respectively. With this support, we independently purify the contour flow and the semantic flow to obtain finer contour information and more comprehensive portrait semantic information, as shown in Fig. 1(b). Finally, we fuse the contour features and semantic features and send them to the segmentation head to estimate the portrait matte. We perform extensive experiments on VideoMatte240K (JPEG SD Format)[21], Adobe Image Matting (AIM)[35], and BG10K. Comparisons with other state-of-the-art models and ablation studies verify that our model performs better than previous works. + +To summarize, our contributions are as follows: + +- We propose EFormer: an approach that can enhance the transformer's attention towards both semantic and contour features of the foreground. +- We further build a semantic and contour detector (SCD) to accurately capture the semantic and contour features and design two separate extraction branches to purify the contour flow and the semantic flow. +- We leverage the cross-attention module between different resolution features in transformer block to autonomously capture and extract high-frequency detail features, which are sparsely distributed along the foreground contour. + +# 2. Related Works + +# 2.1. Image Matting + +Image matting is an essential task in the field of computer vision that aims to accurately estimate the foreground in an + +image. Mathematically speaking, an image $I$ is a combination of an unknown foreground image $F$ and a background image $B$ with a probability coefficient alpha matte maps $\alpha$ . + +$$ +I = \alpha F + (1 - \alpha) B \tag {1} +$$ + +Previous image matting solutions have predominantly focused on low-level features, such as color cues[9, 10, 14, 15] or propagation[1, 2, 19], to distinguish the transition areas between foregrounds and backgrounds. However, such traditional matting algorithms commonly struggle to perform properly in complex scenes. + +With the significant advancements making in deep learning, numerous methods based on convolutional neural networks (CNNs) are proposed, leading to notable successes. Some approaches[36, 37, 39] incorporate auxiliary trimap supervisions to enhance matting performance, while other methods[22, 23] leverage trimap-free solutions to estimate alpha matts from image feature maps using an end-to-end segmentation network. In the background matting series[21, 28], an auxiliary input of the background image is used to compute alpha matte. Furthermore, MODNet[17] proposes an end-to-end manner without any auxiliary input. Although prior works based on CNN kernels have improved the accuracy of image matting, the prediction of models is in the limited receptive field. + +# 2.2. Vision Transformer + +Different from CNNs, vision transformers with self-attention mechanisms can capture long-term dependencies. Therefore, in the realm of computer vision, vision transformers have gained significant attention in recent times. For pixel-level prediction tasks, such as segmentation, SETR[38], SegFormer[34] and DPT[27] apply transformer as an encoder to attain feature maps of images. Their performances demonstrate the transformer with self-attention is capable of building complete contextual information through global interactions. However, recent studies[25, 29] indicate that transformer "yet is incompetent in capturing high frequencies that predominantly convey local information". + +Compared to self-attention, cross-attention has more potential to induce customized features. Max-Deeplab[31], MaskFormer[5, 6] and SeMask[13] use query-based methods through cross-attention, which are inspired by DETR[3]. They view segmentation as a set prediction problem. U-Transformer[26] and EUT[8] modify the cross-attention in the original transformer to leverage the information from the encoder, allowing a fine spatial recovery in the decoder. The cross-resolution attention employed by RTFormer[32] enables the gathering of comprehensive contextual information for high-resolution features. + +In this work, we propose a strategy that using of crossattention between different resolutions guides the model to + +autonomously locate and capture high-frequency features, particularly those near the portrait's contours. Contrary to IFormer[29], this method eliminates the need for predefined parameters to enhance the model's adaptive capabilities. At the same time, this approach can significantly improve the model's performance in capturing contour details. + +# 2.3. Refinement for Segmentation + +In the field of portrait segmentation, refining the high-frequency details near the contours is crucial. Many prior studies on segmentation refinement depend on convolutional networks or MLPs that are designed specifically for this purpose. + +PointRend[18], based on the coarse predictions of Mask R-CNN[12], calculates confidence scores as the point-wise uncertainty measure. And it focuses on the points with the higher uncertainty according to set hyperparameters. Finally, PointRend uses a shared MLP to refine the labels of the selected points. Although PointRend has implemented segmentation refinement to some extent, the model still requires manual adjustment of certain hyperparameters to assist in filtering uncertain points. Transfiner[16] processes detected error-prone and incoherent regions with transformer, which include high-frequency regions strewn along object boundaries. However, during this procedure, excessive focus is placed on the local details and incoherent regions, without effectively associating the local features of the object with the overall semantic information. + +We believe that locating effective detail distribution areas requires information interaction between local details and global semantics. To achieve this, we design a semantic and contour detector that adopts a cascade structure of cross-attention and self-attention. This approach uses global semantic information garnered from self-attention to autonomously select out local details situated at the borders of the contours. On this base, we further utilize the MLP-based extraction branch to purify contour details, thereby significantly enhancing the effect of refinement. + +# 3. Model + +In this section, we introduce the process of establishing the complete model architecture of EFormer, which comprises the backbone for encoder, the transformer block for decoder, and the prediction stage, as shown in Fig. 2. + +# 3.1. Backbone for Encoder + +The CNN backbone takes RGB images $I \in \mathbb{R}^{B \times 3 \times H \times W}$ as input, where $B$ is the number of images, to obtain a pyramid of feature maps $F = \left\{F_{enc\frac{1}{4}}, F_{enc\frac{1}{8}}, F_{enc\frac{1}{16}}\right\}$ with dimension $F_{enc\frac{1}{i}} \in \mathbb{R}^{B \times C_i \times \frac{H}{i} \times \frac{W}{i}}$ . As we all know, in the pyramid of feature maps, shallow features have high resolution and rich details, while deep features have low reso + +![](images/19dc18b25a5ce7108cc5db14d5b7e0dec35cccd2f842cb026d2fcf2d6a35cae3.jpg) +Figure 2. The Architecture of the EFormer. It contains a backbone for encoder, a transformer block for decoder, and the prediction stage, where the transformer block includes a semantic and contour detector, a semantic extraction branch, and a contour-edge extraction branch. + +lation but more concentrated semantic information. Based on this, our research makes the cross-attention mechanism to filter the high-frequency contour features from high-resolution features, using low-frequency semantic information in low-resolution features. However, since the large resolution of feature maps would bring large computation cost, we use the $F_{enc\frac{1}{8}}$ as $F_{HR}$ (high-resolution features) and use the $F_{enc\frac{1}{16}}$ as $F_{LR}$ (low-resolution features) in our base model. The impact of other choices of $F_{HR}$ and $F_{LR}$ on model performance is demonstrated in ablation experiments Sec. 4.4. As shown in Fig. 2, the $F_{HR}$ serves as the source of contour flow, providing high-frequency contour detail features for the following processes. Correspondingly, the $F_{LR}$ serves as the source of semantic flow, providing low-frequency semantic information. Then, the $F_{HR}$ and $F_{LR}$ are projected to same dimension $C$ and respectively flattened into high-resolution feature embedding as $F_{HR}^{em}$ and low-resolution feature embedding as $F_{LR}^{em}$ , where $F_{HR}^{em} \in R^{N\times B\times C}$ , $F_{LR}^{em} \in R^{N\times B\times C}$ , and $N = \frac{H}{8}\times \frac{W}{8}$ . Both of them are utilized as inputs for the transformer block. + +# 3.2. Transformer Block for Decoder + +We propose a novel transformer block in the model decoding stage to enhance the ability to capture and extract visual features. It not only does not reduce the focus on low-frequency semantic features but also significantly increases the attention to high-frequency contour features. Its detailed architecture is depicted in Fig. 3. The novel transformer block first uses a Semantic and Contour De + +tector (SCD) to locate and capture high-frequency contour features and low-frequency semantic features. To achieve this, we cascade a cross-attention layer (CA) and a self-attention layer (SA) in the SCD. Then, as shown in Fig. 3, a Contour-Edge Extraction Branch (CEEB) and a Semantic Extraction Branch (SEB) are further built to handle high-frequency contour feature flow and low-frequency semantic feature flow respectively. These branches can separately purify more refined high-frequency contour features and extract more complete low-frequency semantic features. This ensures an outcome that is optimal for both types of features. + +Semantic and Contour Detector (SCD). There is a fact that low-frequency semantic features far outperform high-frequency contour detail features in terms of quantity and distribution density. So, in the global information propagation process, the low-frequency semantic information easily dominates the feature representations of the original transformer block with self-attention module. + +To prevent the loss of high-frequency contour detail information, we first use a cross-attention layer with different resolution features rather than a self-attention layer. The cross-attention layer can capture high-frequency detail features with high semantic correlation in high-resolution features, based on the low-frequency semantic information in low-resolution features. Technically, after getting $F_{HR}^{em}$ and $F_{LR}^{em}$ , we use Layer Norm (LN) processing them to maintain consistency. Due to the sensitivity of the attention mechanism towards positional information, we use absolute and + +![](images/8b731e62de48786a607962e324bfde9d2bf61d970716c65951847e6993d27c69.jpg) +Figure 3. Our transformer block. It includes a Semantic and Contour Detector (SCD), a Semantic Extraction Branch (SEB), and a Contour-Edge Extraction Branch (CEEB). The detailed composition of the module is shown in the figure. + +learnable position encoding (PE) in the SCD. The input $K$ , $Q$ , $V$ of the cross-attention layer (CA) is calculated as + +$$ +K = L N \left(F _ {H R} ^ {e m}\right) + P E \tag {2} +$$ + +$$ +Q = L N \left(F _ {L R} ^ {e m}\right) + P E \tag {3} +$$ + +$$ +V = L N \left(F _ {L R} ^ {e m} + F _ {H R} ^ {e m}\right) \tag {4} +$$ + +By performing cross-attention calculations using features of different resolutions, it is possible to automatically locate and capture the discrete contours along the portrait and allocate more attention to corresponding positions. The contour-edge features can be derived from the output of the CA. + +$$ +F _ {\text {c o n t o u r - e d g e}} = C A (K, Q, V) \tag {5} +$$ + +Although CA pays extra attention to contour-edge features, it still captures the overall semantics. By the residual connection between CA and $V$ , both the semantic and contour-edge features of the portrait are enhanced, especially the + +contour-edge features. + +$$ +F _ {\text {e n h a n c e}} = C A (K, Q, V) + V \tag {6} +$$ + +Since semantic information is still present, we use the self-attention layer (SA) after CA. SA can locate the semantic feature within the portrait and contours, thereby emphasizing the semantic attributes at their respective positions. After using LN to normalize $F_{\text{enhance}}$ , the input $K', Q', V'$ of the SA is calculated as + +$$ +K ^ {\prime} = Q ^ {\prime} = L N \left(F _ {\text {e n h a n c e}}\right) + P E \tag {7} +$$ + +$$ +V ^ {\prime} = L N \left(F _ {\text {e n h a n c e}}\right) \tag {8} +$$ + +$$ +F _ {\text {c o n t o u r - s e m a n t i c}} = S A \left(K ^ {\prime}, Q ^ {\prime}, V ^ {\prime}\right) + V ^ {\prime} \tag {9} +$$ + +Through SA filtering the semantic and contour features, the residual connection between $V'$ and the output of SA, we can produce $F_{contour - semantic}$ as the output of the semantic and contour detector. + +Feature Extraction Branches. After SCD, we use multi-layer perceptron (MLP) to build a Contour-Edge Extraction Branch (CEEB) and a Semantic Extraction Branch (SEB). To prevent the loss of high-frequency contour detail information, we aggregate $F_{contour - semantic}$ getting from SCD and $F_{HR}^{em}$ in the contour flow. After getting the merged feature normalized by LN, we employ CEEB to purify it, in order to further extract and refine high-frequency contour details individually. Similarly, to obtain more comprehensive and consistent portrait semantic features separately, we merge $F_{contour - semantic}$ and $F_{LR}^{em}$ in the semantic flow. Then, we normalize the aggregated feature by applying LN and use SEB to filter it. + +$$ +F _ {\text {c o n t o u r}} = M L P \left(L N \left(F _ {H R} ^ {\text {e m}} + F _ {\text {c o n t o u r - s e m a n t i c}}\right)\right) \tag {10} +$$ + +$$ +F _ {\text {s e m a n t i c}} = M L P \left(L N \left(F _ {L R} ^ {\text {e m}} + F _ {\text {c o n t o u r - s e m a n t i c}}\right)\right) \tag {11} +$$ + +# 3.3. Prediction Stage + +After getting key characteristics $F_{semantic} \in R^{N \times B \times C}$ and $F_{contour} \in R^{N \times B \times C}$ from transformer block, where $N = \frac{H}{8} \times \frac{W}{8}$ , we transform them into $F_{semantic}' \in R^{B \times C \times \frac{H}{8} \times \frac{W}{8}}$ and $F_{contour}' \in R^{B \times C \times \frac{H}{8} \times \frac{W}{8}}$ by converting the vector into the matrix. To achieve improved outcomes in portrait segmentation, we propose fusing the two kinds of features and ultimately acquiring the predicted portrait matte from the segmentation head module. + +Fuse. We attain the semantic feature through the residual connection between $F_{LR}^{\prime}$ and $F_{semantic}^{\prime}$ of semantic flow. Similarly, we get the contour feature from contour flow and subsequently merge the $F_{semantic}^{\prime}$ and $F_{contour}^{\prime}$ to generate the fused features $F_{contour - semantic}^{\prime}$ . + +$$ +F _ {\text {s e m a n t i c}} ^ {\prime} = \operatorname {C o n v} \left(F _ {\text {s e m a n t i c}} ^ {\prime} + F _ {L R} ^ {\prime}\right) \tag {12} +$$ + +$$ +F _ {\text {c o n t o u r}} ^ {\prime} = \operatorname {C o n v} \left(F _ {\text {c o n t o u r}} ^ {\prime} + F _ {H R} ^ {\prime}\right) \tag {13} +$$ + +$$ +F _ {\text {c o n t o u r - s e m a n t i c}} ^ {\prime} = \operatorname {F u s e} \left(F _ {\text {s e m a n t i c}} ^ {\prime} + F _ {\text {c o n t o u r}} ^ {\prime}\right) \tag {14} +$$ + +Segmentation Head. Finally, We gather $F_{contour - semantic}^{\prime}$ and $F_{enc\frac{1}{4}}$ with higher resolution to predict portrait matte in 1/4 scale of the input image and upscale it to the original scale through bi-linear interpolation as the output of our model. + +$$ +m a t t e = H e a d \left(F _ {c o n t o u r - s e m a n t i c} ^ {\prime} + F _ {e n c \frac {1}{4}}\right) \tag {15} +$$ + +# 4. Experiments + +# 4.1. Dataset and Evaluation + +Dataset. We utilize composite training data from two sources: the foreground image dataset VideoMatte240K (JPEG SD Format)[21] and Adobe Image Matting (AIM)[35]. For the background images, we select BG10K following the approach of BGMv2[21], which is a collection of photographs depicting various life scenes without any human portraits. We select foregrounds from VideoMatte240K-JPEG-SD and AIM, and backgrounds from BG10K to composite image datasets. We split VideoMatte240K-JPEG-SD into 234,982/3,007/2720 image sets with image resolutions of $224 \times 224$ and $512 \times 288$ for training, validating, and testing our model. Similarly, BG10K is split into 9000/1000 image sets, and AIM is split into 214/10 image sets for training and testing, where AIM only includes the portrait images with resolutions at $512 \times 512$ . Finally, we use the aforementioned testing datasets to compare and evaluate the performance of EFormer against other models. + +Evaluation. We mainly consider portrait matting accuracy for evaluation. For portrait matting accuracy, we use Mean Absolute Difference (MAD), Mean Squared Error (MSE), Gradient (Grad), and Connectivity (Conn) as evaluation metrics. We also scale MAD, MSE, Grad, and Conn by $10^{3}$ , $10^{3}$ , $10^{-3}$ , and $10^{-3}$ respectively, for convenience of reference. For all these metrics, the lower number represents better performance. + +# 4.2. Implementation Details + +Training Settings. For training the network, we use a single RTX 3090 GPU with batch size at 24. The optimizer is AdamW and the initial learning rate is set to $10^{-4}$ . All models in ablation studies are trained for 25 epochs, with the learning rate decaying by a factor of 0.8 every 5 epochs. Additionally, to augment the data, each input image is subjected to random horizontal flipping. + +Backbone. The backbone we use for the network is the pretrained ResNet50[11], since it is the most frequently used backbone in prior works. We use the implementation and weights from torchvision. + +
ModelMAD↓MSE↓Grad↓Conn↓
DeepLabv3[4]14.47009.67008.55001.6900
MODnet[17]10.39005.65002.02001.0400
BGMv2[21]4.18581.79341.30110.5326
RVM[22]5.99001.17001.10000.3400
EFormer2.30970.66370.45800.2740
+ +Table 1. Comparisons on the test set of VideoMatte240K-JPEG-SD. Bold indicates the best performance among these models under the inputs with the same resolution at $512 \times 288$ . + +
ModelMAD↓MSE↓Grad↓Conn↓
DeepLabv3[4]29.6423.7820.177.71
MODnet[17]21.6614.275.375.23
BGMv2[21]44.6139.085.5411.60
RVM[22]14.848.934.353.83
EFormer7.472.132.831.90
+ +Table 2. Comparisons on the test set of AIM. Bold indicates the best performance among these models under the inputs with the same resolution at $512 \times 512$ . + +Transformer Block for Decoder. The transformer block is based on the transformer decoder of DETR. We continuously use four transformer blocks in the decoder of the network. We set the channel $C = 256$ and the number of attention heads $M = 8$ for each multi-head attention module. + +Prediction Stage. After obtaining the semantic and contour feature maps of the portrait from the stacked transformer blocks, we fuse both of them. Then, the alpha mattes predicted by the segmentation head are upsampled to the original size of the input image with bi-linear interpolation. + +Loss function. To compute losses during the training stage, we use a Binary CrossEntropy Loss function. + +$$ +L _ {b c e} = - (g \times \log (m) + (1 - g) \times \log (1 - m)) \tag {16} +$$ + +Where, $m$ is the prediction of the network, $g$ is the ground truth. + +# 4.3. Comparisons to State-of-the-art Methods + +To demonstrate the progressiveness of EFormer, we compare it against the latest state-of-the-art trimap-free portrait matting solutions. These are including DeepLabV3[4] with ResNet101[11] backbone, BGMv2[21] with ResNet50[11] backbone, MODNet[17], and RVM[22]. Our assessment is based on the testing datasets of VideoMatte240K-JPEG-SD and AIM, respectively. As the training files for MODNet are not available, we use its official weights. Furthermore, since BGMv2, DeepLabV3, and RVM have already been trained on all datasets, our method can be compared with them in a fair manner. + +![](images/a44cfc197d899a897cdae63917124a1992d807f4b3ea9c1a1cb1af45bf9a7ef0.jpg) +Figure 4. Visualization of portrait matte predictions from MODnet, BGMv2, and EFormer under challenging image from the test set of VideoMatte240K (JPEG SD Format). Our model shows a better ability to distinguish ambiguous foreground contours, as indicated by the red box. Please zoom in to view the detailed information. + +
ModelMAD↓MSE↓Grad↓Conn↓
DeepLabv3[4]18.158614.76236.00152.8824
MODnet[17]13.14008.96562.60282.1543
BGMv2[21]7.04712.35431.56211.0651
RVM[22]6.13572.26470.65230.2866
EFormer4.13571.42800.37600.1753
+ +As shown in the Tab. 1, Tab. 2 and Tab. 3, our model exhibits superior performance in portrait matte tasks. It has a stronger capability to distinguish the contour of a portrait. And it can finely segment edge details (e.g. fingers), as exemplified in Fig. 4. + +# 4.4. Ablation Studies + +Training Epochs. In the ablation study with an image resolution of $224 \times 224$ , we conduct 25 training epochs. Every 5 epochs of training, we evaluate the training effect of the model. At the same time, the learning rate is reduced by 0.8 times. The experimental results indicate that our model achieves optimal performance in the 20 epoch, as shown in the Tab. 4. + +Table 3. Comparisons on the test set of VideoMatte240K-JPEG-SD. Bold indicates the best performance among these models under the inputs with the same resolution at $224 \times 224$ . + +
MAD↓MSE↓Grad↓Conn↓
5-epoch4.39531.48240.40890.1865
10-epoch4.31081.47660.39080.1834
15-epoch4.28671.52380.40870.1833
20-epoch4.13571.42800.37600.1753
25-epoch4.18111.47970.38580.1779
+ +Table 4. Ablation Study on the Training Epochs with the test set of VideoMatte240K-JPEG-SD (resolution at $224 \times 224$ ). **Bold** indicates the best performance among these models. We use the result of 20-epoch training as a strong baseline model. + +The Functional Role of Each Module. To demonstrate the effectiveness of the model in capturing and extracting contour details and portrait semantics, we analyze the outputs of CA, SA, CEEB, and SEB, and utilize Class Activation Map (CAM) to visualize the attention allocation of each layer's features, as shown in Fig. 5. + +From the first line of Fig. 5, it can be evident that after the cross-attention layer, the model clearly strengthens its attention to the contours of the portrait. Following the contour prior, the self-attention layer captures the portrait's semantics thoroughly. CEEB and SEB models provide sharper contour and more comprehensive semantic information. + +Comparing the second and first lines of Fig. 5, it is ob + +![](images/f4a86092d25db2d4a4c9dd6b16240410a2a59c5e7287d4d91926926071d90d7f.jpg) +Figure 5. We use Grad-CAM to visualize the attention distribution of various features in CA, SA, CEEB, and SEB. The model in the first row contains four complete modules. For comparison, the SA layer and CA layer were removed from the models in the second and third rows, respectively. + +
CASAMAD↓MSE↓Grad↓Conn↓
✓4.27621.53440.40920.1823
✓4.17101.43540.37730.1769
✓✓4.13571.42800.37600.1753
+ +served that removing the self-attention layer considerably weakens the semantic information extracted by SEB. In the comparison between the third and first lines of Fig. 5, it can be indicated that removing the cross-attention layer significantly weakens the contour information extracted by CEEB. + +After qualitative analysis, the prediction results of all the above models are evaluated. It is concluded that the model with CA, SA, CEEB, and SEB is both effective and rational, as demonstrated in the Tab. 5. + +As is well known, when it comes to extracting features for portrait segmentation tasks, semantic information plays a significant role, in contrast, contour details account for a relatively small proportion. Owing to this, as shown in the Tab. 5 and Fig. 5, the model without SA leads to a significant increase in all metrics due to the lack of amount of semantic information, while the model without CA experiences a slight increase in all indicators due to the loss of contour details. In order to fully harness the potential of the vision transformer, we integrate both CA and SA to ensure semantic completeness and enhance attention towards contour details, thereby achieving more comprehensive and fine-grained segmentation results. + +The Selection of $F_{HR}$ and $F_{LR}$ . In our base model, + +Table 5. Ablation Study on the Functional Role of CA and SA with the test set of VideoMatte240K-JPEG-SD (resolution at $224 \times 224$ ). **Bold** indicates the best performance among these models. + +
FHRFLRMAD↓MSE↓Grad↓Conn↓
Fenc1/8Fenc1/164.13571.42800.37600.1753
Fenc1/4Fenc1/83.99701.33430.33280.1679
Fenc1/4Fenc1/163.97301.29490.32150.1665
+ +Table 6. Ablation Study on the Selection of $F_{HR}$ and $F_{LR}$ with the test set of VideoMatte240K-JPEG-SD (resolution at $224 \times 224$ ). Bold indicates the best performance among these models. + +considering the large resolution of feature maps would bring large computation cost, we use the $F_{enc}^{\frac{1}{8}}$ as $F_{HR}$ (high-resolution features) and use the $F_{enc}^{\frac{1}{16}}$ as $F_{LR}$ (low-resolution features). Then the $F_{HR}$ serves as the source of contour flow, providing high-frequency contour detail features. Correspondingly, the $F_{LR}$ serves as the source of semantic flow, providing low-frequency semantic information. As shown in the Tab. 6, after using $F_{enc}^{\frac{1}{4}}$ with higher resolution as $F_{HR}$ , our model can capture more high-frequency contour detail features and its performance is better. And after using $F_{enc}^{\frac{1}{16}}$ with lower resolution as $F_{LR}$ , our model can use more concentrated low-frequency semantic information, which makes all indicators optimal. + +# 5. Conclusion + +In this paper, we propose a strategy that use of cross-attention module between different resolutions guides the EFormer to autonomously locate and capture high-frequency contour features. Without manual parameter tuning, the proposed method can adaptability capture high-frequency details neglected by previous transformer-based methods. Based on the cross-attention module, we further construct a semantic and contour detector (SCD), which can enhance the transformer's attention to both the high-frequency contour and low-frequency semantic features. Additionally, we build a contour-edge extraction branch (CEEB) and a semantic extraction branch (SEB) to extract finer high-frequency contour details and more comprehensive portrait's low-frequency semantic information. Extensive experiments show that our method outperforms previous portrait matting solutions and exhibits robust performance during testing and inference. + +# Acknowledgement + +The work was jointly supported by the National Natural Science Foundations of China under grant No.62272364, the Teaching Reform Project of Shaanxi Higher Continuing Education under Grant No.21XJZ004. + +# References + +[1] Yagiz Aksoy, Tunc Ozan Aydin, and Marc Pollefeys. Designing effective inter-pixel information flow for natural image matting. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 29-37, 2017. 3 +[2] Yagiz Aksoy, Tae-Hyun Oh, Sylvain Paris, Marc Pollefeys, and Wojciech Matusik. Semantic soft segmentation. ACM Transactions on Graphics (TOG), 37(4):1-13, 2018. 3 +[3] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In European conference on computer vision, pages 213-229. Springer, 2020. 3 +[4] Liang-Chieh Chen, George Papandreou, Florian Schroff, and Hartwig Adam. Rethinking atrous convolution for semantic image segmentation. arXiv preprint arXiv:1706.05587, 2017. 6, 7 +[5] Bowen Cheng, Alex Schwing, and Alexander Kirillov. Per-pixel classification is not all you need for semantic segmentation. Advances in Neural Information Processing Systems, 34:17864-17875, 2021. 3 +[6] Bowen Cheng, Ishan Misra, Alexander G Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1290–1299, 2022. 3 +[7] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 2 +[8] Dazhao Du, Bing Su, Yu Li, Zhongang Qi, Lingyu Si, and Ying Shan. Efficient u-transformer with boundary-aware loss for action segmentation. arXiv preprint arXiv:2205.13425, 2022. 3 +[9] Xiaoxue Feng, Xiaohui Liang, and Zili Zhang. A cluster sampling method for image matting via sparse coding. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part II 14, pages 204-219. Springer, 2016. 3 +[10] Kaiming He, Christoph Rhemann, Carsten Rother, Xiaou Tang, and Jian Sun. A global sampling method for alpha matting. In CVPR 2011, pages 2049-2056. IEEE, 2011. 3 +[11] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 6 +[12] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In Proceedings of the IEEE international conference on computer vision, pages 2961-2969, 2017. 3 +[13] Jitesh Jain, Anukriti Singh, Nikita Orlov, Zilong Huang, Jiachen Li, Steven Walton, and Humphrey Shi. Semask: Semantically masked transformers for semantic segmentation. arXiv preprint arXiv:2112.12782, 2021. 3 +[14] Jubin Johnson, Ehsan Shahrian Varnousfaderani, Hisham Cholakkal, and Deepu Rajan. Sparse coding for alpha mat + +ting. IEEE Transactions on Image Processing, 25(7):3032-3043, 2016. 3 +[15] Levent Karacan, Aykut Erdem, and Erkut Erdem. Image matting with kl-divergence based sparse sampling. In Proceedings of the IEEE international conference on computer vision, pages 424-432, 2015. 3 +[16] Lei Ke, Martin Danelljan, Xia Li, Yu-Wing Tai, Chi-Keung Tang, and Fisher Yu. Mask transfiner for high-quality instance segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4412–4421, 2022. 3 +[17] Zhanghan Ke, Jiayu Sun, Kaican Li, Qiong Yan, and Rynson WH Lau. Modnet: Real-time trimap-free portrait matting via objective decomposition. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1140-1147, 2022. 2, 3, 6, 7 +[18] Alexander Kirillov, Yuxin Wu, Kaiming He, and Ross Girshick. Pointrend: Image segmentation as rendering. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9799-9808, 2020. 3 +[19] Anat Levin, Alex Rav-Acha, and Dani Lischinski. Spectral matting. IEEE transactions on pattern analysis and machine intelligence, 30(10):1699-1712, 2008. 3 +[20] Yaoyi Li and Hongtao Lu. Natural image matting via guided contextual attention. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 11450-11457, 2020. 1 +[21] Shanchuan Lin, Andrey Ryabtsev, Soumyadip Sengupta, Brian L Curless, Steven M Seitz, and Ira Kemelmacher-Shlizerman. Real-time high-resolution background matting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8762-8771, 2021. 2, 3, 6, 7 +[22] Shanchuan Lin, Linjie Yang, Imran Saleemi, and Soumyadip Sengupta. Robust high-resolution video matting with temporal guidance. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 238-247, 2022. 2, 3, 6, 7 +[23] Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dólar. Focal loss for dense object detection. In Proceedings of the IEEE international conference on computer vision, pages 2980-2988, 2017. 2, 3 +[24] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10012-10022, 2021. 2 +[25] Namuk Park and Songkuk Kim. How do vision transformers work? arXiv preprint arXiv:2202.06709, 2022. 2, 3 +[26] Olivier Petit, Nicolas Thome, Clement Rambour, Loic Themyr, Toby Collins, and Luc Soler. U-net transformer: Self and cross attention for medical image segmentation. In Machine Learning in Medical Imaging: 12th International Workshop, MLMI 2021, Held in Conjunction with MICCAI 2021, Strasbourg, France, September 27, 2021, Proceedings 12, pages 267-276. Springer, 2021. 3 +[27] René Ranftl, Alexey Bochkovskiy, and Vladlen Koltun. Vision transformers for dense prediction. In Proceedings of + +the IEEE/CVF international conference on computer vision, pages 12179-12188, 2021.3 +[28] Soumyadip Sengupta, Vivek Jayaram, Brian Curless, Steven M Seitz, and Ira Kemelmacher-Shlizerman. Background matting: The world is your green screen. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2291-2300, 2020. 2, 3 +[29] Chenyang Si, Weihao Yu, Pan Zhou, Yichen Zhou, Xinchao Wang, and Shuicheng Yan. Inception transformer. Advances in Neural Information Processing Systems, 35:23495-23509, 2022. 2, 3 +[30] Yanan Sun, Guanzhi Wang, Qiao Gu, Chi-Keung Tang, and Yu-Wing Tai. Deep video matting via spatio-temporal alignment and aggregation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6975–6984, 2021. 1 +[31] Huiyu Wang, Yukun Zhu, Hartwig Adam, Alan Yuille, and Liang-Chieh Chen. Max-deeplab: End-to-end panoptic segmentation with mask transformers. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5463-5474, 2021. 3 +[32] Jian Wang, Chenhui Gou, Qiman Wu, Haocheng Feng, Junyu Han, Errui Ding, and Jingdong Wang. Rrformer: Efficient design for real-time semantic segmentation with transformer. Advances in Neural Information Processing Systems, 35:7423-7436, 2022. 3 +[33] Yu Wang, Yi Niu, Peiyong Duan, Jianwei Lin, and Yuanjie Zheng. Deep propagation based image matting. In IJCAI, pages 999-1006, 2018. 1 +[34] Enze Xie, Wenhai Wang, Zhiding Yu, Anima Anandkumar, Jose M Alvarez, and Ping Luo. Segformer: Simple and efficient design for semantic segmentation with transformers. Advances in Neural Information Processing Systems, 34:12077-12090, 2021. 3 +[35] Ning Xu, Brian Price, Scott Cohen, and Thomas Huang. Deep image matting. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2970-2979, 2017. 1, 2, 6 +[36] Haichao Yu, Ning Xu, Zilong Huang, Yuqian Zhou, and Humphrey Shi. High-resolution deep image matting. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 3217-3224, 2021. 3 +[37] Yunke Zhang, Chi Wang, Miaomiao Cui, Peiran Ren, Xuansong Xie, Xian-Sheng Hua, Hujun Bao, Qixing Huang, and Weiwei Xu. Attention-guided temporally coherent video object matting. In Proceedings of the 29th ACM International Conference on Multimedia, pages 5128-5137, 2021. 3 +[38] Sixiao Zheng, Jiachen Lu, Hengshuang Zhao, Xiatian Zhu, Zekun Luo, Yabiao Wang, Yanwei Fu, Jianfeng Feng, Tao Xiang, Philip HS Torr, et al. Rethinking semantic segmentation from a sequence-to-sequence perspective with transformers. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6881-6890, 2021. 3 +[39] Bingke Zhu, Yingying Chen, Jinqiao Wang, Si Liu, Bo Zhang, and Ming Tang. Fast deep matting for portrait animation on mobile phone. In Proceedings of the 25th ACM international conference on Multimedia, pages 297-305, 2017. 3 \ No newline at end of file diff --git a/2024/EFormer_ Enhanced Transformer towards Semantic-Contour Features of Foreground for Portraits Matting/images.zip b/2024/EFormer_ Enhanced Transformer towards Semantic-Contour Features of Foreground for Portraits Matting/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..f273b5dda99819cf26cbf15100ef707d906cddbb --- /dev/null +++ b/2024/EFormer_ Enhanced Transformer towards Semantic-Contour Features of Foreground for Portraits Matting/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c099c386318137ef5f3e4a8c16184ed3601bd5298d3b1897237a6d3c51933f41 +size 540061 diff --git a/2024/EFormer_ Enhanced Transformer towards Semantic-Contour Features of Foreground for Portraits Matting/layout.json b/2024/EFormer_ Enhanced Transformer towards Semantic-Contour Features of Foreground for Portraits Matting/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..a02f5c89cc6001cac07f88450cb72b46707e3e35 --- /dev/null +++ b/2024/EFormer_ Enhanced Transformer towards Semantic-Contour Features of Foreground for Portraits Matting/layout.json @@ -0,0 +1,7905 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 70, + 102, + 523, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 102, + 523, + 140 + ], + "spans": [ + { + "bbox": [ + 70, + 102, + 523, + 140 + ], + "type": "text", + "content": "EFormer: Enhanced Transformer towards Semantic-Contour Features of Foreground for Portraits Matting" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 159, + 161, + 434, + 176 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 161, + 434, + 176 + ], + "spans": [ + { + "bbox": [ + 159, + 161, + 434, + 176 + ], + "type": "text", + "content": "Zitao Wang Qiguang Miao* Yue Xi* Peipei Zhao" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 250, + 178, + 342, + 189 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 250, + 178, + 342, + 189 + ], + "spans": [ + { + "bbox": [ + 250, + 178, + 342, + 189 + ], + "type": "text", + "content": "Xidian University, China" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 134, + 191, + 453, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 191, + 453, + 203 + ], + "spans": [ + { + "bbox": [ + 134, + 191, + 453, + 203 + ], + "type": "text", + "content": "{21181214174@stu.，qgmiao@，xiyue@，zhpp2023@}xidian.edu.cn" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 143, + 231, + 192, + 243 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 231, + 192, + 243 + ], + "spans": [ + { + "bbox": [ + 143, + 231, + 192, + 243 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 256, + 290, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 256, + 290, + 616 + ], + "spans": [ + { + "bbox": [ + 46, + 256, + 290, + 616 + ], + "type": "text", + "content": "The portrait matting task aims to extract an alpha matte with complete semantics and finely detailed contours. In comparison to CNN-based approaches, transformers with self-attention module have a better capacity to capture long-range dependencies and low-frequency semantic information of a portrait. However, recent research shows that the self-attention mechanism struggles with modeling high-frequency contour information and capturing fine contour details, which can lead to bias while predicting the portrait's contours. To deal with this issue, we propose EFormer to enhance the model's attention towards both the low-frequency semantic and high-frequency contour features. For the high-frequency contours, our research demonstrates that cross-attention module between different resolutions can guide our model to allocate attention appropriately to these contour regions. Supported by this, we can successfully extract the high-frequency detail information around the portrait's contours, which were previously ignored by self-attention. Based on the cross-attention module, we further build a semantic and contour detector (SCD) to accurately capture both the low-frequency semantic and high-frequency contour features. And we design a contour-edge extraction branch and semantic extraction branch to extract refined high-frequency contour features and complete low-frequency semantic information, respectively. Finally, we fuse the two kinds of features and leverage the segmentation head to generate a predicted portrait matte. Experiments on VideoMatte240K (JPEG SD Format) and Adobe Image Matting (AIM) datasets demonstrate that EFormer outperforms previous portrait matte methods." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 638, + 128, + 651 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 638, + 128, + 651 + ], + "spans": [ + { + "bbox": [ + 47, + 638, + 128, + 651 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 658, + 287, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 658, + 287, + 696 + ], + "spans": [ + { + "bbox": [ + 47, + 658, + 287, + 696 + ], + "type": "text", + "content": "Portrait matting aims to extract a precise alpha matte of persons from natural images. It has gathered great momentum in vision applications, including but not limited to photo" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 307, + 230, + 545, + 375 + ], + "blocks": [ + { + "bbox": [ + 307, + 230, + 545, + 375 + ], + "lines": [ + { + "bbox": [ + 307, + 230, + 545, + 375 + ], + "spans": [ + { + "bbox": [ + 307, + 230, + 545, + 375 + ], + "type": "image", + "image_path": "c9937fb43e03a7633aee3ae5e531b87101ba94ad97cd2f40462a3b2430c6fa14.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 317, + 378, + 545, + 520 + ], + "blocks": [ + { + "bbox": [ + 317, + 378, + 545, + 520 + ], + "lines": [ + { + "bbox": [ + 317, + 378, + 545, + 520 + ], + "spans": [ + { + "bbox": [ + 317, + 378, + 545, + 520 + ], + "type": "image", + "image_path": "3b06508ac7a09bba239e870787572c7c30df658ab82ad01b9fb750bf1c4e71f8.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 531, + 547, + 609 + ], + "lines": [ + { + "bbox": [ + 304, + 531, + 547, + 609 + ], + "spans": [ + { + "bbox": [ + 304, + 531, + 547, + 609 + ], + "type": "text", + "content": "Figure 1. (I) Visualization with Grad-CAM of output from each layer within different transformer blocks. (a) The features captured from Self-attention and MLP layers in the original transformer block. (b) In comparison, the features captured from each layer in the proposed transformer block. (II) Comparisons of different transformer blocks. (c) The original transformer block in ViT. (d) The proposed transformer block for portrait matte." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 629, + 457, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 629, + 457, + 641 + ], + "spans": [ + { + "bbox": [ + 306, + 629, + 457, + 641 + ], + "type": "text", + "content": "editing and background replacement." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 642, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 642, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 642, + 547, + 715 + ], + "type": "text", + "content": "Existing methods[20, 30, 33, 35] attempt to estimate a portrait matte with Convolutional Neural Networks (CNNs) under the guidance of a pre-determined trimap, which is a three-class map that indicates the foreground, background, and unknown region. However, it is time-consuming and labor-intensive to obtain abundant trimaps by manually an-" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 702, + 139, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 702, + 139, + 713 + ], + "spans": [ + { + "bbox": [ + 58, + 702, + 139, + 713 + ], + "type": "text", + "content": "*Corresponding authors." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3880" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 204 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 204 + ], + "type": "text", + "content": "notating images. Some methods eliminate the trimap, such as background matting series[21, 28] uses a separate background image instead of the trimap as an auxiliary reference for model prediction. However, in practical cases, it is necessary to provide two aligned images for the model, which include one with only the background and the other one including people. An alternative solution[17, 22, 23] is to rely solely on RGB images as input for predicting portrait matte. Although the previous models have made strides in portrait matte, there is a constraint on the receptive field in CNN-based approaches." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 205, + 289, + 326 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 205, + 289, + 326 + ], + "spans": [ + { + "bbox": [ + 46, + 205, + 289, + 326 + ], + "type": "text", + "content": "In contrast, Vision Transformer (ViT)[7] and Swin Transformer (Swin) [24] for semantic segmentation employ a self-attention mechanism, which conducts interactions among all pixels in an image to obtain a global receptive field. Previous research [25] shows that ViT and its variants can capture the low-frequency components in images effectively, such as global shapes and structures of a scene or object. However, they usually neglect high-frequency ones in images, such as edges and textures. They work similarly to low-pass filters." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 327, + 289, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 327, + 289, + 495 + ], + "spans": [ + { + "bbox": [ + 46, + 327, + 289, + 495 + ], + "type": "text", + "content": "As shown in Fig. 1(a), we can observe the transformer block with self-attention clearly pays little attention to the portrait's high-frequency contour regions. We believe that this phenomenon is attributed to the excessive propagation of global information because global attention is performed in each transformer block during the calculation of multiple blocks. So, the global information is propagated throughout the model in this iterative calculation, ultimately leading it to only focus on low-frequency semantic features with stronger clustering in portraits. Over time, the model gradually ignores the discrete high-frequency detail features around the contours. Therefore, it is desired to capture low- and high-frequency components simultaneously to improve the segmentation performance." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 496, + 289, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 496, + 289, + 652 + ], + "spans": [ + { + "bbox": [ + 46, + 496, + 289, + 652 + ], + "type": "text", + "content": "To achieve this goal, Inception Transformer (IFormer)[29] splits all channels into a convolutional path and self-attention path, respectively, to capture high and low frequencies. However, it is required for a channel ratio of each block at different levels in the IFormer. These channel ratios typically require manual adjustment, resulting in significant uncertainty when aiming for optimal model performance. To enhance the model's stability and adaptive capability, we propose a novel architecture EFormer without manually adjusted parameters. In EFormer, we use the cross-attention layer between different resolution features, as shown in Fig. 1(d), which is replaced by convolution to handle high-frequency detail features." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 654, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 654, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 654, + 289, + 715 + ], + "type": "text", + "content": "Our research further demonstrates that the cross-attention module between different resolutions can formulate a more reasonable attention allocation mechanism, which can accurately capture contour features, as shown in Fig. 1(b). Particularly, EFormer with a cross-attention layer" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 72, + 545, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 204 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 204 + ], + "type": "text", + "content": "not only can not lose the semantic information within the portrait but also can successfully capture the high-frequency details around the portrait's contours, which were previously ignored by self-attention. It is worth noting that high-frequency components only around the boundary between foreground and background are critical to the improvement of segmentation performance. Therefore, we follow a coarse-to-fine manner. Initially, EFormer needs to locate all high-frequency regions in the image. Then, it filters and extracts the high-frequency regions around the foreground contours." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 205, + 546, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 205, + 546, + 491 + ], + "spans": [ + { + "bbox": [ + 304, + 205, + 546, + 491 + ], + "type": "text", + "content": "Based on the cross-attention module, we design a semantic and contour detector (SCD) in EFormer, which cascades the cross-attention layer and self-attention layer to sequentially locate the contour and semantic features. Firstly, the model adjusts its attention to contour features with the guidance of a cross-attention module. Secondly, the self-attention module captures semantic information within a portrait's contours and reverses filtering and correcting the contour information output by the cross-attention layer. This allows the model to gradually capture and match the high-frequency contour information and the low-frequency semantic information. Then, we use multi-layer perceptron (MLP) to build the contour-edge extraction branch (CEEB) and the semantic extraction branch (SEB), respectively. With this support, we independently purify the contour flow and the semantic flow to obtain finer contour information and more comprehensive portrait semantic information, as shown in Fig. 1(b). Finally, we fuse the contour features and semantic features and send them to the segmentation head to estimate the portrait matte. We perform extensive experiments on VideoMatte240K (JPEG SD Format)[21], Adobe Image Matting (AIM)[35], and BG10K. Comparisons with other state-of-the-art models and ablation studies verify that our model performs better than previous works." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 317, + 492, + 511, + 504 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 492, + 511, + 504 + ], + "spans": [ + { + "bbox": [ + 317, + 492, + 511, + 504 + ], + "type": "text", + "content": "To summarize, our contributions are as follows:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 506, + 545, + 637 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 304, + 506, + 545, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 506, + 545, + 540 + ], + "spans": [ + { + "bbox": [ + 304, + 506, + 545, + 540 + ], + "type": "text", + "content": "- We propose EFormer: an approach that can enhance the transformer's attention towards both semantic and contour features of the foreground." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 541, + 545, + 588 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 541, + 545, + 588 + ], + "spans": [ + { + "bbox": [ + 304, + 541, + 545, + 588 + ], + "type": "text", + "content": "- We further build a semantic and contour detector (SCD) to accurately capture the semantic and contour features and design two separate extraction branches to purify the contour flow and the semantic flow." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 589, + 545, + 637 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 589, + 545, + 637 + ], + "spans": [ + { + "bbox": [ + 304, + 589, + 545, + 637 + ], + "type": "text", + "content": "- We leverage the cross-attention module between different resolution features in transformer block to autonomously capture and extract high-frequency detail features, which are sparsely distributed along the foreground contour." + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 650, + 398, + 662 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 650, + 398, + 662 + ], + "spans": [ + { + "bbox": [ + 306, + 650, + 398, + 662 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 670, + 399, + 684 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 670, + 399, + 684 + ], + "spans": [ + { + "bbox": [ + 306, + 670, + 399, + 684 + ], + "type": "text", + "content": "2.1. Image Matting" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "content": "Image matting is an essential task in the field of computer vision that aims to accurately estimate the foreground in an" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "3881" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 109 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 109 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 109 + ], + "type": "text", + "content": "image. Mathematically speaking, an image " + }, + { + "bbox": [ + 46, + 72, + 287, + 109 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 46, + 72, + 287, + 109 + ], + "type": "text", + "content": " is a combination of an unknown foreground image " + }, + { + "bbox": [ + 46, + 72, + 287, + 109 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 46, + 72, + 287, + 109 + ], + "type": "text", + "content": " and a background image " + }, + { + "bbox": [ + 46, + 72, + 287, + 109 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 46, + 72, + 287, + 109 + ], + "type": "text", + "content": " with a probability coefficient alpha matte maps " + }, + { + "bbox": [ + 46, + 72, + 287, + 109 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 46, + 72, + 287, + 109 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 122, + 117, + 287, + 130 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 122, + 117, + 287, + 130 + ], + "spans": [ + { + "bbox": [ + 122, + 117, + 287, + 130 + ], + "type": "interline_equation", + "content": "I = \\alpha F + (1 - \\alpha) B \\tag {1}", + "image_path": "3fea930b38b9d3f3a531a1aebefde57371eee284e250cfe2b2c32ad0031fbbc6.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 138, + 287, + 209 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 138, + 287, + 209 + ], + "spans": [ + { + "bbox": [ + 46, + 138, + 287, + 209 + ], + "type": "text", + "content": "Previous image matting solutions have predominantly focused on low-level features, such as color cues[9, 10, 14, 15] or propagation[1, 2, 19], to distinguish the transition areas between foregrounds and backgrounds. However, such traditional matting algorithms commonly struggle to perform properly in complex scenes." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 210, + 287, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 210, + 287, + 376 + ], + "spans": [ + { + "bbox": [ + 46, + 210, + 287, + 376 + ], + "type": "text", + "content": "With the significant advancements making in deep learning, numerous methods based on convolutional neural networks (CNNs) are proposed, leading to notable successes. Some approaches[36, 37, 39] incorporate auxiliary trimap supervisions to enhance matting performance, while other methods[22, 23] leverage trimap-free solutions to estimate alpha matts from image feature maps using an end-to-end segmentation network. In the background matting series[21, 28], an auxiliary input of the background image is used to compute alpha matte. Furthermore, MODNet[17] proposes an end-to-end manner without any auxiliary input. Although prior works based on CNN kernels have improved the accuracy of image matting, the prediction of models is in the limited receptive field." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 384, + 162, + 396 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 384, + 162, + 396 + ], + "spans": [ + { + "bbox": [ + 47, + 384, + 162, + 396 + ], + "type": "text", + "content": "2.2. Vision Transformer" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 403, + 287, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 403, + 287, + 556 + ], + "spans": [ + { + "bbox": [ + 46, + 403, + 287, + 556 + ], + "type": "text", + "content": "Different from CNNs, vision transformers with self-attention mechanisms can capture long-term dependencies. Therefore, in the realm of computer vision, vision transformers have gained significant attention in recent times. For pixel-level prediction tasks, such as segmentation, SETR[38], SegFormer[34] and DPT[27] apply transformer as an encoder to attain feature maps of images. Their performances demonstrate the transformer with self-attention is capable of building complete contextual information through global interactions. However, recent studies[25, 29] indicate that transformer \"yet is incompetent in capturing high frequencies that predominantly convey local information\"." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 558, + 287, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 558, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 46, + 558, + 287, + 689 + ], + "type": "text", + "content": "Compared to self-attention, cross-attention has more potential to induce customized features. Max-Deeplab[31], MaskFormer[5, 6] and SeMask[13] use query-based methods through cross-attention, which are inspired by DETR[3]. They view segmentation as a set prediction problem. U-Transformer[26] and EUT[8] modify the cross-attention in the original transformer to leverage the information from the encoder, allowing a fine spatial recovery in the decoder. The cross-resolution attention employed by RTFormer[32] enables the gathering of comprehensive contextual information for high-resolution features." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "content": "In this work, we propose a strategy that using of crossattention between different resolutions guides the model to" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "text", + "content": "autonomously locate and capture high-frequency features, particularly those near the portrait's contours. Contrary to IFormer[29], this method eliminates the need for predefined parameters to enhance the model's adaptive capabilities. At the same time, this approach can significantly improve the model's performance in capturing contour details." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 152, + 465, + 164 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 152, + 465, + 164 + ], + "spans": [ + { + "bbox": [ + 305, + 152, + 465, + 164 + ], + "type": "text", + "content": "2.3. Refinement for Segmentation" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 170, + 545, + 229 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 170, + 545, + 229 + ], + "spans": [ + { + "bbox": [ + 304, + 170, + 545, + 229 + ], + "type": "text", + "content": "In the field of portrait segmentation, refining the high-frequency details near the contours is crucial. Many prior studies on segmentation refinement depend on convolutional networks or MLPs that are designed specifically for this purpose." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 230, + 546, + 408 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 230, + 546, + 408 + ], + "spans": [ + { + "bbox": [ + 304, + 230, + 546, + 408 + ], + "type": "text", + "content": "PointRend[18], based on the coarse predictions of Mask R-CNN[12], calculates confidence scores as the point-wise uncertainty measure. And it focuses on the points with the higher uncertainty according to set hyperparameters. Finally, PointRend uses a shared MLP to refine the labels of the selected points. Although PointRend has implemented segmentation refinement to some extent, the model still requires manual adjustment of certain hyperparameters to assist in filtering uncertain points. Transfiner[16] processes detected error-prone and incoherent regions with transformer, which include high-frequency regions strewn along object boundaries. However, during this procedure, excessive focus is placed on the local details and incoherent regions, without effectively associating the local features of the object with the overall semantic information." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 409, + 545, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 409, + 545, + 529 + ], + "spans": [ + { + "bbox": [ + 304, + 409, + 545, + 529 + ], + "type": "text", + "content": "We believe that locating effective detail distribution areas requires information interaction between local details and global semantics. To achieve this, we design a semantic and contour detector that adopts a cascade structure of cross-attention and self-attention. This approach uses global semantic information garnered from self-attention to autonomously select out local details situated at the borders of the contours. On this base, we further utilize the MLP-based extraction branch to purify contour details, thereby significantly enhancing the effect of refinement." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 540, + 354, + 552 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 540, + 354, + 552 + ], + "spans": [ + { + "bbox": [ + 306, + 540, + 354, + 552 + ], + "type": "text", + "content": "3. Model" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 559, + 545, + 608 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 559, + 545, + 608 + ], + "spans": [ + { + "bbox": [ + 304, + 559, + 545, + 608 + ], + "type": "text", + "content": "In this section, we introduce the process of establishing the complete model architecture of EFormer, which comprises the backbone for encoder, the transformer block for decoder, and the prediction stage, as shown in Fig. 2." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 615, + 434, + 627 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 615, + 434, + 627 + ], + "spans": [ + { + "bbox": [ + 306, + 615, + 434, + 627 + ], + "type": "text", + "content": "3.1. Backbone for Encoder" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 632, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 632, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 632, + 545, + 713 + ], + "type": "text", + "content": "The CNN backbone takes RGB images " + }, + { + "bbox": [ + 304, + 632, + 545, + 713 + ], + "type": "inline_equation", + "content": "I \\in \\mathbb{R}^{B \\times 3 \\times H \\times W}" + }, + { + "bbox": [ + 304, + 632, + 545, + 713 + ], + "type": "text", + "content": " as input, where " + }, + { + "bbox": [ + 304, + 632, + 545, + 713 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 304, + 632, + 545, + 713 + ], + "type": "text", + "content": " is the number of images, to obtain a pyramid of feature maps " + }, + { + "bbox": [ + 304, + 632, + 545, + 713 + ], + "type": "inline_equation", + "content": "F = \\left\\{F_{enc\\frac{1}{4}}, F_{enc\\frac{1}{8}}, F_{enc\\frac{1}{16}}\\right\\}" + }, + { + "bbox": [ + 304, + 632, + 545, + 713 + ], + "type": "text", + "content": " with dimension " + }, + { + "bbox": [ + 304, + 632, + 545, + 713 + ], + "type": "inline_equation", + "content": "F_{enc\\frac{1}{i}} \\in \\mathbb{R}^{B \\times C_i \\times \\frac{H}{i} \\times \\frac{W}{i}}" + }, + { + "bbox": [ + 304, + 632, + 545, + 713 + ], + "type": "text", + "content": ". As we all know, in the pyramid of feature maps, shallow features have high resolution and rich details, while deep features have low reso" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3882" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 61, + 68, + 535, + 312 + ], + "blocks": [ + { + "bbox": [ + 61, + 68, + 535, + 312 + ], + "lines": [ + { + "bbox": [ + 61, + 68, + 535, + 312 + ], + "spans": [ + { + "bbox": [ + 61, + 68, + 535, + 312 + ], + "type": "image", + "image_path": "19dc18b25a5ce7108cc5db14d5b7e0dec35cccd2f842cb026d2fcf2d6a35cae3.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 319, + 546, + 343 + ], + "lines": [ + { + "bbox": [ + 46, + 319, + 546, + 343 + ], + "spans": [ + { + "bbox": [ + 46, + 319, + 546, + 343 + ], + "type": "text", + "content": "Figure 2. The Architecture of the EFormer. It contains a backbone for encoder, a transformer block for decoder, and the prediction stage, where the transformer block includes a semantic and contour detector, a semantic extraction branch, and a contour-edge extraction branch." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 363, + 289, + 602 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 363, + 289, + 602 + ], + "spans": [ + { + "bbox": [ + 46, + 363, + 289, + 602 + ], + "type": "text", + "content": "lation but more concentrated semantic information. Based on this, our research makes the cross-attention mechanism to filter the high-frequency contour features from high-resolution features, using low-frequency semantic information in low-resolution features. However, since the large resolution of feature maps would bring large computation cost, we use the " + }, + { + "bbox": [ + 46, + 363, + 289, + 602 + ], + "type": "inline_equation", + "content": "F_{enc\\frac{1}{8}}" + }, + { + "bbox": [ + 46, + 363, + 289, + 602 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 46, + 363, + 289, + 602 + ], + "type": "inline_equation", + "content": "F_{HR}" + }, + { + "bbox": [ + 46, + 363, + 289, + 602 + ], + "type": "text", + "content": " (high-resolution features) and use the " + }, + { + "bbox": [ + 46, + 363, + 289, + 602 + ], + "type": "inline_equation", + "content": "F_{enc\\frac{1}{16}}" + }, + { + "bbox": [ + 46, + 363, + 289, + 602 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 46, + 363, + 289, + 602 + ], + "type": "inline_equation", + "content": "F_{LR}" + }, + { + "bbox": [ + 46, + 363, + 289, + 602 + ], + "type": "text", + "content": " (low-resolution features) in our base model. The impact of other choices of " + }, + { + "bbox": [ + 46, + 363, + 289, + 602 + ], + "type": "inline_equation", + "content": "F_{HR}" + }, + { + "bbox": [ + 46, + 363, + 289, + 602 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 363, + 289, + 602 + ], + "type": "inline_equation", + "content": "F_{LR}" + }, + { + "bbox": [ + 46, + 363, + 289, + 602 + ], + "type": "text", + "content": " on model performance is demonstrated in ablation experiments Sec. 4.4. As shown in Fig. 2, the " + }, + { + "bbox": [ + 46, + 363, + 289, + 602 + ], + "type": "inline_equation", + "content": "F_{HR}" + }, + { + "bbox": [ + 46, + 363, + 289, + 602 + ], + "type": "text", + "content": " serves as the source of contour flow, providing high-frequency contour detail features for the following processes. Correspondingly, the " + }, + { + "bbox": [ + 46, + 363, + 289, + 602 + ], + "type": "inline_equation", + "content": "F_{LR}" + }, + { + "bbox": [ + 46, + 363, + 289, + 602 + ], + "type": "text", + "content": " serves as the source of semantic flow, providing low-frequency semantic information. Then, the " + }, + { + "bbox": [ + 46, + 363, + 289, + 602 + ], + "type": "inline_equation", + "content": "F_{HR}" + }, + { + "bbox": [ + 46, + 363, + 289, + 602 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 363, + 289, + 602 + ], + "type": "inline_equation", + "content": "F_{LR}" + }, + { + "bbox": [ + 46, + 363, + 289, + 602 + ], + "type": "text", + "content": " are projected to same dimension " + }, + { + "bbox": [ + 46, + 363, + 289, + 602 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 46, + 363, + 289, + 602 + ], + "type": "text", + "content": " and respectively flattened into high-resolution feature embedding as " + }, + { + "bbox": [ + 46, + 363, + 289, + 602 + ], + "type": "inline_equation", + "content": "F_{HR}^{em}" + }, + { + "bbox": [ + 46, + 363, + 289, + 602 + ], + "type": "text", + "content": " and low-resolution feature embedding as " + }, + { + "bbox": [ + 46, + 363, + 289, + 602 + ], + "type": "inline_equation", + "content": "F_{LR}^{em}" + }, + { + "bbox": [ + 46, + 363, + 289, + 602 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 363, + 289, + 602 + ], + "type": "inline_equation", + "content": "F_{HR}^{em} \\in R^{N\\times B\\times C}" + }, + { + "bbox": [ + 46, + 363, + 289, + 602 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 363, + 289, + 602 + ], + "type": "inline_equation", + "content": "F_{LR}^{em} \\in R^{N\\times B\\times C}" + }, + { + "bbox": [ + 46, + 363, + 289, + 602 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 363, + 289, + 602 + ], + "type": "inline_equation", + "content": "N = \\frac{H}{8}\\times \\frac{W}{8}" + }, + { + "bbox": [ + 46, + 363, + 289, + 602 + ], + "type": "text", + "content": ". Both of them are utilized as inputs for the transformer block." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 611, + 218, + 623 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 611, + 218, + 623 + ], + "spans": [ + { + "bbox": [ + 47, + 611, + 218, + 623 + ], + "type": "text", + "content": "3.2. Transformer Block for Decoder" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 629, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 629, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 629, + 287, + 713 + ], + "type": "text", + "content": "We propose a novel transformer block in the model decoding stage to enhance the ability to capture and extract visual features. It not only does not reduce the focus on low-frequency semantic features but also significantly increases the attention to high-frequency contour features. Its detailed architecture is depicted in Fig. 3. The novel transformer block first uses a Semantic and Contour De" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 363, + 545, + 505 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 363, + 545, + 505 + ], + "spans": [ + { + "bbox": [ + 304, + 363, + 545, + 505 + ], + "type": "text", + "content": "tector (SCD) to locate and capture high-frequency contour features and low-frequency semantic features. To achieve this, we cascade a cross-attention layer (CA) and a self-attention layer (SA) in the SCD. Then, as shown in Fig. 3, a Contour-Edge Extraction Branch (CEEB) and a Semantic Extraction Branch (SEB) are further built to handle high-frequency contour feature flow and low-frequency semantic feature flow respectively. These branches can separately purify more refined high-frequency contour features and extract more complete low-frequency semantic features. This ensures an outcome that is optimal for both types of features." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 508, + 545, + 591 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 508, + 545, + 591 + ], + "spans": [ + { + "bbox": [ + 304, + 508, + 545, + 591 + ], + "type": "text", + "content": "Semantic and Contour Detector (SCD). There is a fact that low-frequency semantic features far outperform high-frequency contour detail features in terms of quantity and distribution density. So, in the global information propagation process, the low-frequency semantic information easily dominates the feature representations of the original transformer block with self-attention module." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 594, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 546, + 713 + ], + "type": "text", + "content": "To prevent the loss of high-frequency contour detail information, we first use a cross-attention layer with different resolution features rather than a self-attention layer. The cross-attention layer can capture high-frequency detail features with high semantic correlation in high-resolution features, based on the low-frequency semantic information in low-resolution features. Technically, after getting " + }, + { + "bbox": [ + 304, + 594, + 546, + 713 + ], + "type": "inline_equation", + "content": "F_{HR}^{em}" + }, + { + "bbox": [ + 304, + 594, + 546, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 594, + 546, + 713 + ], + "type": "inline_equation", + "content": "F_{LR}^{em}" + }, + { + "bbox": [ + 304, + 594, + 546, + 713 + ], + "type": "text", + "content": ", we use Layer Norm (LN) processing them to maintain consistency. Due to the sensitivity of the attention mechanism towards positional information, we use absolute and" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3883" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 55, + 70, + 281, + 417 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 281, + 417 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 281, + 417 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 281, + 417 + ], + "type": "image", + "image_path": "8b731e62de48786a607962e324bfde9d2bf61d970716c65951847e6993d27c69.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 426, + 287, + 471 + ], + "lines": [ + { + "bbox": [ + 46, + 426, + 287, + 471 + ], + "spans": [ + { + "bbox": [ + 46, + 426, + 287, + 471 + ], + "type": "text", + "content": "Figure 3. Our transformer block. It includes a Semantic and Contour Detector (SCD), a Semantic Extraction Branch (SEB), and a Contour-Edge Extraction Branch (CEEB). The detailed composition of the module is shown in the figure." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 490, + 287, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 490, + 287, + 514 + ], + "spans": [ + { + "bbox": [ + 46, + 490, + 287, + 514 + ], + "type": "text", + "content": "learnable position encoding (PE) in the SCD. The input " + }, + { + "bbox": [ + 46, + 490, + 287, + 514 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 46, + 490, + 287, + 514 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 490, + 287, + 514 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 46, + 490, + 287, + 514 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 490, + 287, + 514 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 46, + 490, + 287, + 514 + ], + "type": "text", + "content": " of the cross-attention layer (CA) is calculated as" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 116, + 521, + 287, + 536 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 116, + 521, + 287, + 536 + ], + "spans": [ + { + "bbox": [ + 116, + 521, + 287, + 536 + ], + "type": "interline_equation", + "content": "K = L N \\left(F _ {H R} ^ {e m}\\right) + P E \\tag {2}", + "image_path": "4c49fc820fa02334fd209f498cc00bd40210c931f617a601217344304d9aa2c9.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 117, + 542, + 287, + 555 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 117, + 542, + 287, + 555 + ], + "spans": [ + { + "bbox": [ + 117, + 542, + 287, + 555 + ], + "type": "interline_equation", + "content": "Q = L N \\left(F _ {L R} ^ {e m}\\right) + P E \\tag {3}", + "image_path": "d922f956e126fc53e9c7ee91a4f655916efed9c2370e7ca23525979287a76132.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 115, + 559, + 287, + 572 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 115, + 559, + 287, + 572 + ], + "spans": [ + { + "bbox": [ + 115, + 559, + 287, + 572 + ], + "type": "interline_equation", + "content": "V = L N \\left(F _ {L R} ^ {e m} + F _ {H R} ^ {e m}\\right) \\tag {4}", + "image_path": "659403ccbb92a1bb76b7861d7df735755c5f0d42a2440f2949be721f6bd54cdd.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 577, + 287, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 577, + 287, + 647 + ], + "spans": [ + { + "bbox": [ + 46, + 577, + 287, + 647 + ], + "type": "text", + "content": "By performing cross-attention calculations using features of different resolutions, it is possible to automatically locate and capture the discrete contours along the portrait and allocate more attention to corresponding positions. The contour-edge features can be derived from the output of the CA." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 99, + 648, + 287, + 662 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 99, + 648, + 287, + 662 + ], + "spans": [ + { + "bbox": [ + 99, + 648, + 287, + 662 + ], + "type": "interline_equation", + "content": "F _ {\\text {c o n t o u r - e d g e}} = C A (K, Q, V) \\tag {5}", + "image_path": "4cc26dc00f1e6b4e120c3f4915bd2e1ff11861f2e9451744338b852243cff055.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 666, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 666, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 666, + 287, + 714 + ], + "type": "text", + "content": "Although CA pays extra attention to contour-edge features, it still captures the overall semantics. By the residual connection between CA and " + }, + { + "bbox": [ + 46, + 666, + 287, + 714 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 46, + 666, + 287, + 714 + ], + "type": "text", + "content": ", both the semantic and contour-edge features of the portrait are enhanced, especially the" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 73, + 399, + 84 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 73, + 399, + 84 + ], + "spans": [ + { + "bbox": [ + 306, + 73, + 399, + 84 + ], + "type": "text", + "content": "contour-edge features." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 358, + 92, + 545, + 106 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 358, + 92, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 358, + 92, + 545, + 106 + ], + "type": "interline_equation", + "content": "F _ {\\text {e n h a n c e}} = C A (K, Q, V) + V \\tag {6}", + "image_path": "921a82ee03c7898d298e9f42fd99b7de4e4f20a17dccc9480c1b4177cb2d7265.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 114, + 547, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 114, + 547, + 185 + ], + "spans": [ + { + "bbox": [ + 305, + 114, + 547, + 185 + ], + "type": "text", + "content": "Since semantic information is still present, we use the self-attention layer (SA) after CA. SA can locate the semantic feature within the portrait and contours, thereby emphasizing the semantic attributes at their respective positions. After using LN to normalize " + }, + { + "bbox": [ + 305, + 114, + 547, + 185 + ], + "type": "inline_equation", + "content": "F_{\\text{enhance}}" + }, + { + "bbox": [ + 305, + 114, + 547, + 185 + ], + "type": "text", + "content": ", the input " + }, + { + "bbox": [ + 305, + 114, + 547, + 185 + ], + "type": "inline_equation", + "content": "K', Q', V'" + }, + { + "bbox": [ + 305, + 114, + 547, + 185 + ], + "type": "text", + "content": " of the SA is calculated as" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 354, + 193, + 545, + 207 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 354, + 193, + 545, + 207 + ], + "spans": [ + { + "bbox": [ + 354, + 193, + 545, + 207 + ], + "type": "interline_equation", + "content": "K ^ {\\prime} = Q ^ {\\prime} = L N \\left(F _ {\\text {e n h a n c e}}\\right) + P E \\tag {7}", + "image_path": "4d09a6c07d1acdc7cf2a44e242c18891b1485dfcf9f9d5ac5b96aacae4f55582.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 381, + 214, + 545, + 228 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 381, + 214, + 545, + 228 + ], + "spans": [ + { + "bbox": [ + 381, + 214, + 545, + 228 + ], + "type": "interline_equation", + "content": "V ^ {\\prime} = L N \\left(F _ {\\text {e n h a n c e}}\\right) \\tag {8}", + "image_path": "026acaa0454415164063e5c81998eca3a5078b63823219a22fc6485f519b877d.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 333, + 232, + 545, + 245 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 232, + 545, + 245 + ], + "spans": [ + { + "bbox": [ + 333, + 232, + 545, + 245 + ], + "type": "interline_equation", + "content": "F _ {\\text {c o n t o u r - s e m a n t i c}} = S A \\left(K ^ {\\prime}, Q ^ {\\prime}, V ^ {\\prime}\\right) + V ^ {\\prime} \\tag {9}", + "image_path": "87a864922831bbaa97005944f3b54edf31c39b10d0e405f40fd2ec3bb95578e2.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 250, + 545, + 296 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 250, + 545, + 296 + ], + "spans": [ + { + "bbox": [ + 305, + 250, + 545, + 296 + ], + "type": "text", + "content": "Through SA filtering the semantic and contour features, the residual connection between " + }, + { + "bbox": [ + 305, + 250, + 545, + 296 + ], + "type": "inline_equation", + "content": "V'" + }, + { + "bbox": [ + 305, + 250, + 545, + 296 + ], + "type": "text", + "content": " and the output of SA, we can produce " + }, + { + "bbox": [ + 305, + 250, + 545, + 296 + ], + "type": "inline_equation", + "content": "F_{contour - semantic}" + }, + { + "bbox": [ + 305, + 250, + 545, + 296 + ], + "type": "text", + "content": " as the output of the semantic and contour detector." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 305, + 297, + 546, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 297, + 546, + 453 + ], + "spans": [ + { + "bbox": [ + 305, + 297, + 546, + 453 + ], + "type": "text", + "content": "Feature Extraction Branches. After SCD, we use multi-layer perceptron (MLP) to build a Contour-Edge Extraction Branch (CEEB) and a Semantic Extraction Branch (SEB). To prevent the loss of high-frequency contour detail information, we aggregate " + }, + { + "bbox": [ + 305, + 297, + 546, + 453 + ], + "type": "inline_equation", + "content": "F_{contour - semantic}" + }, + { + "bbox": [ + 305, + 297, + 546, + 453 + ], + "type": "text", + "content": " getting from SCD and " + }, + { + "bbox": [ + 305, + 297, + 546, + 453 + ], + "type": "inline_equation", + "content": "F_{HR}^{em}" + }, + { + "bbox": [ + 305, + 297, + 546, + 453 + ], + "type": "text", + "content": " in the contour flow. After getting the merged feature normalized by LN, we employ CEEB to purify it, in order to further extract and refine high-frequency contour details individually. Similarly, to obtain more comprehensive and consistent portrait semantic features separately, we merge " + }, + { + "bbox": [ + 305, + 297, + 546, + 453 + ], + "type": "inline_equation", + "content": "F_{contour - semantic}" + }, + { + "bbox": [ + 305, + 297, + 546, + 453 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 297, + 546, + 453 + ], + "type": "inline_equation", + "content": "F_{LR}^{em}" + }, + { + "bbox": [ + 305, + 297, + 546, + 453 + ], + "type": "text", + "content": " in the semantic flow. Then, we normalize the aggregated feature by applying LN and use SEB to filter it." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 315, + 460, + 545, + 485 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 460, + 545, + 485 + ], + "spans": [ + { + "bbox": [ + 315, + 460, + 545, + 485 + ], + "type": "interline_equation", + "content": "F _ {\\text {c o n t o u r}} = M L P \\left(L N \\left(F _ {H R} ^ {\\text {e m}} + F _ {\\text {c o n t o u r - s e m a n t i c}}\\right)\\right) \\tag {10}", + "image_path": "e9a763e3d19f560cadfc8ca9e271d25a77cf376be381a1c1878d2aec91b93623.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 485, + 545, + 510 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 485, + 545, + 510 + ], + "spans": [ + { + "bbox": [ + 313, + 485, + 545, + 510 + ], + "type": "interline_equation", + "content": "F _ {\\text {s e m a n t i c}} = M L P \\left(L N \\left(F _ {L R} ^ {\\text {e m}} + F _ {\\text {c o n t o u r - s e m a n t i c}}\\right)\\right) \\tag {11}", + "image_path": "919faa5c3862fb002c897a16f6a9ea3a0721929dd518abd44c0287aed082ed76.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 306, + 517, + 406, + 529 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 517, + 406, + 529 + ], + "spans": [ + { + "bbox": [ + 306, + 517, + 406, + 529 + ], + "type": "text", + "content": "3.3. Prediction Stage" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 306, + 534, + 545, + 632 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 534, + 545, + 632 + ], + "spans": [ + { + "bbox": [ + 306, + 534, + 545, + 632 + ], + "type": "text", + "content": "After getting key characteristics " + }, + { + "bbox": [ + 306, + 534, + 545, + 632 + ], + "type": "inline_equation", + "content": "F_{semantic} \\in R^{N \\times B \\times C}" + }, + { + "bbox": [ + 306, + 534, + 545, + 632 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 306, + 534, + 545, + 632 + ], + "type": "inline_equation", + "content": "F_{contour} \\in R^{N \\times B \\times C}" + }, + { + "bbox": [ + 306, + 534, + 545, + 632 + ], + "type": "text", + "content": " from transformer block, where " + }, + { + "bbox": [ + 306, + 534, + 545, + 632 + ], + "type": "inline_equation", + "content": "N = \\frac{H}{8} \\times \\frac{W}{8}" + }, + { + "bbox": [ + 306, + 534, + 545, + 632 + ], + "type": "text", + "content": ", we transform them into " + }, + { + "bbox": [ + 306, + 534, + 545, + 632 + ], + "type": "inline_equation", + "content": "F_{semantic}' \\in R^{B \\times C \\times \\frac{H}{8} \\times \\frac{W}{8}}" + }, + { + "bbox": [ + 306, + 534, + 545, + 632 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 306, + 534, + 545, + 632 + ], + "type": "inline_equation", + "content": "F_{contour}' \\in R^{B \\times C \\times \\frac{H}{8} \\times \\frac{W}{8}}" + }, + { + "bbox": [ + 306, + 534, + 545, + 632 + ], + "type": "text", + "content": " by converting the vector into the matrix. To achieve improved outcomes in portrait segmentation, we propose fusing the two kinds of features and ultimately acquiring the predicted portrait matte from the segmentation head module." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 305, + 632, + 545, + 694 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 632, + 545, + 694 + ], + "spans": [ + { + "bbox": [ + 305, + 632, + 545, + 694 + ], + "type": "text", + "content": "Fuse. We attain the semantic feature through the residual connection between " + }, + { + "bbox": [ + 305, + 632, + 545, + 694 + ], + "type": "inline_equation", + "content": "F_{LR}^{\\prime}" + }, + { + "bbox": [ + 305, + 632, + 545, + 694 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 632, + 545, + 694 + ], + "type": "inline_equation", + "content": "F_{semantic}^{\\prime}" + }, + { + "bbox": [ + 305, + 632, + 545, + 694 + ], + "type": "text", + "content": " of semantic flow. Similarly, we get the contour feature from contour flow and subsequently merge the " + }, + { + "bbox": [ + 305, + 632, + 545, + 694 + ], + "type": "inline_equation", + "content": "F_{semantic}^{\\prime}" + }, + { + "bbox": [ + 305, + 632, + 545, + 694 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 632, + 545, + 694 + ], + "type": "inline_equation", + "content": "F_{contour}^{\\prime}" + }, + { + "bbox": [ + 305, + 632, + 545, + 694 + ], + "type": "text", + "content": " to generate the fused features " + }, + { + "bbox": [ + 305, + 632, + 545, + 694 + ], + "type": "inline_equation", + "content": "F_{contour - semantic}^{\\prime}" + }, + { + "bbox": [ + 305, + 632, + 545, + 694 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 344, + 700, + 545, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 344, + 700, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 344, + 700, + 545, + 715 + ], + "type": "interline_equation", + "content": "F _ {\\text {s e m a n t i c}} ^ {\\prime} = \\operatorname {C o n v} \\left(F _ {\\text {s e m a n t i c}} ^ {\\prime} + F _ {L R} ^ {\\prime}\\right) \\tag {12}", + "image_path": "1e4a72544e9f79f0874750d7ad2c9f156164a440b97785ea4a1ab136d10594f8.jpg" + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3884" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 90, + 71, + 286, + 86 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 90, + 71, + 286, + 86 + ], + "spans": [ + { + "bbox": [ + 90, + 71, + 286, + 86 + ], + "type": "interline_equation", + "content": "F _ {\\text {c o n t o u r}} ^ {\\prime} = \\operatorname {C o n v} \\left(F _ {\\text {c o n t o u r}} ^ {\\prime} + F _ {H R} ^ {\\prime}\\right) \\tag {13}", + "image_path": "b1b60f2a5332519aefe1b1e5a69fc81f54c1ceb01ddb5e0e4cca93bc7bc65378.jpg" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 52, + 90, + 286, + 105 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 52, + 90, + 286, + 105 + ], + "spans": [ + { + "bbox": [ + 52, + 90, + 286, + 105 + ], + "type": "interline_equation", + "content": "F _ {\\text {c o n t o u r - s e m a n t i c}} ^ {\\prime} = \\operatorname {F u s e} \\left(F _ {\\text {s e m a n t i c}} ^ {\\prime} + F _ {\\text {c o n t o u r}} ^ {\\prime}\\right) \\tag {14}", + "image_path": "af3987d74861ba534347e64c22cb262dde5cd990acbbecc0f8d34a75bb8cbd01.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 110, + 286, + 169 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 110, + 286, + 169 + ], + "spans": [ + { + "bbox": [ + 46, + 110, + 286, + 169 + ], + "type": "text", + "content": "Segmentation Head. Finally, We gather " + }, + { + "bbox": [ + 46, + 110, + 286, + 169 + ], + "type": "inline_equation", + "content": "F_{contour - semantic}^{\\prime}" + }, + { + "bbox": [ + 46, + 110, + 286, + 169 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 110, + 286, + 169 + ], + "type": "inline_equation", + "content": "F_{enc\\frac{1}{4}}" + }, + { + "bbox": [ + 46, + 110, + 286, + 169 + ], + "type": "text", + "content": " with higher resolution to predict portrait matte in 1/4 scale of the input image and upscale it to the original scale through bi-linear interpolation as the output of our model." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 176, + 286, + 197 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 176, + 286, + 197 + ], + "spans": [ + { + "bbox": [ + 62, + 176, + 286, + 197 + ], + "type": "interline_equation", + "content": "m a t t e = H e a d \\left(F _ {c o n t o u r - s e m a n t i c} ^ {\\prime} + F _ {e n c \\frac {1}{4}}\\right) \\tag {15}", + "image_path": "d164252b4edf04b04fd21a56793c335203834cfa3c48dab361edcf62e47e9530.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 205, + 128, + 219 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 205, + 128, + 219 + ], + "spans": [ + { + "bbox": [ + 47, + 205, + 128, + 219 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 225, + 179, + 237 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 225, + 179, + 237 + ], + "spans": [ + { + "bbox": [ + 47, + 225, + 179, + 237 + ], + "type": "text", + "content": "4.1. Dataset and Evaluation" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 243, + 287, + 458 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 243, + 287, + 458 + ], + "spans": [ + { + "bbox": [ + 46, + 243, + 287, + 458 + ], + "type": "text", + "content": "Dataset. We utilize composite training data from two sources: the foreground image dataset VideoMatte240K (JPEG SD Format)[21] and Adobe Image Matting (AIM)[35]. For the background images, we select BG10K following the approach of BGMv2[21], which is a collection of photographs depicting various life scenes without any human portraits. We select foregrounds from VideoMatte240K-JPEG-SD and AIM, and backgrounds from BG10K to composite image datasets. We split VideoMatte240K-JPEG-SD into 234,982/3,007/2720 image sets with image resolutions of " + }, + { + "bbox": [ + 46, + 243, + 287, + 458 + ], + "type": "inline_equation", + "content": "224 \\times 224" + }, + { + "bbox": [ + 46, + 243, + 287, + 458 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 243, + 287, + 458 + ], + "type": "inline_equation", + "content": "512 \\times 288" + }, + { + "bbox": [ + 46, + 243, + 287, + 458 + ], + "type": "text", + "content": " for training, validating, and testing our model. Similarly, BG10K is split into 9000/1000 image sets, and AIM is split into 214/10 image sets for training and testing, where AIM only includes the portrait images with resolutions at " + }, + { + "bbox": [ + 46, + 243, + 287, + 458 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 46, + 243, + 287, + 458 + ], + "type": "text", + "content": ". Finally, we use the aforementioned testing datasets to compare and evaluate the performance of EFormer against other models." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 460, + 286, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 460, + 286, + 555 + ], + "spans": [ + { + "bbox": [ + 46, + 460, + 286, + 555 + ], + "type": "text", + "content": "Evaluation. We mainly consider portrait matting accuracy for evaluation. For portrait matting accuracy, we use Mean Absolute Difference (MAD), Mean Squared Error (MSE), Gradient (Grad), and Connectivity (Conn) as evaluation metrics. We also scale MAD, MSE, Grad, and Conn by " + }, + { + "bbox": [ + 46, + 460, + 286, + 555 + ], + "type": "inline_equation", + "content": "10^{3}" + }, + { + "bbox": [ + 46, + 460, + 286, + 555 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 460, + 286, + 555 + ], + "type": "inline_equation", + "content": "10^{3}" + }, + { + "bbox": [ + 46, + 460, + 286, + 555 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 460, + 286, + 555 + ], + "type": "inline_equation", + "content": "10^{-3}" + }, + { + "bbox": [ + 46, + 460, + 286, + 555 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 460, + 286, + 555 + ], + "type": "inline_equation", + "content": "10^{-3}" + }, + { + "bbox": [ + 46, + 460, + 286, + 555 + ], + "type": "text", + "content": " respectively, for convenience of reference. For all these metrics, the lower number represents better performance." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 563, + 180, + 575 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 563, + 180, + 575 + ], + "spans": [ + { + "bbox": [ + 47, + 563, + 180, + 575 + ], + "type": "text", + "content": "4.2. Implementation Details" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 581, + 286, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 581, + 286, + 665 + ], + "spans": [ + { + "bbox": [ + 46, + 581, + 286, + 665 + ], + "type": "text", + "content": "Training Settings. For training the network, we use a single RTX 3090 GPU with batch size at 24. The optimizer is AdamW and the initial learning rate is set to " + }, + { + "bbox": [ + 46, + 581, + 286, + 665 + ], + "type": "inline_equation", + "content": "10^{-4}" + }, + { + "bbox": [ + 46, + 581, + 286, + 665 + ], + "type": "text", + "content": ". All models in ablation studies are trained for 25 epochs, with the learning rate decaying by a factor of 0.8 every 5 epochs. Additionally, to augment the data, each input image is subjected to random horizontal flipping." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 665, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 286, + 713 + ], + "type": "text", + "content": "Backbone. The backbone we use for the network is the pretrained ResNet50[11], since it is the most frequently used backbone in prior works. We use the implementation and weights from torchvision." + } + ] + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 314, + 70, + 538, + 155 + ], + "blocks": [ + { + "bbox": [ + 314, + 70, + 538, + 155 + ], + "lines": [ + { + "bbox": [ + 314, + 70, + 538, + 155 + ], + "spans": [ + { + "bbox": [ + 314, + 70, + 538, + 155 + ], + "type": "table", + "html": "
ModelMAD↓MSE↓Grad↓Conn↓
DeepLabv3[4]14.47009.67008.55001.6900
MODnet[17]10.39005.65002.02001.0400
BGMv2[21]4.18581.79341.30110.5326
RVM[22]5.99001.17001.10000.3400
EFormer2.30970.66370.45800.2740
", + "image_path": "2b29dd07027dfc431e484454ec4b4fcbcaf97a2dea64be6672f24670cdbd512d.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "type": "table", + "bbox": [ + 318, + 206, + 533, + 291 + ], + "blocks": [ + { + "bbox": [ + 305, + 163, + 545, + 196 + ], + "lines": [ + { + "bbox": [ + 305, + 163, + 545, + 196 + ], + "spans": [ + { + "bbox": [ + 305, + 163, + 545, + 196 + ], + "type": "text", + "content": "Table 1. Comparisons on the test set of VideoMatte240K-JPEG-SD. Bold indicates the best performance among these models under the inputs with the same resolution at " + }, + { + "bbox": [ + 305, + 163, + 545, + 196 + ], + "type": "inline_equation", + "content": "512 \\times 288" + }, + { + "bbox": [ + 305, + 163, + 545, + 196 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 318, + 206, + 533, + 291 + ], + "lines": [ + { + "bbox": [ + 318, + 206, + 533, + 291 + ], + "spans": [ + { + "bbox": [ + 318, + 206, + 533, + 291 + ], + "type": "table", + "html": "
ModelMAD↓MSE↓Grad↓Conn↓
DeepLabv3[4]29.6423.7820.177.71
MODnet[17]21.6614.275.375.23
BGMv2[21]44.6139.085.5411.60
RVM[22]14.848.934.353.83
EFormer7.472.132.831.90
", + "image_path": "ea643b03dda1a2cfc98ef9e9a8cf8ac9e1fce6990aeb2dc38ccb6b152968a860.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_body" + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 300, + 545, + 332 + ], + "lines": [ + { + "bbox": [ + 305, + 300, + 545, + 332 + ], + "spans": [ + { + "bbox": [ + 305, + 300, + 545, + 332 + ], + "type": "text", + "content": "Table 2. Comparisons on the test set of AIM. Bold indicates the best performance among these models under the inputs with the same resolution at " + }, + { + "bbox": [ + 305, + 300, + 545, + 332 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 305, + 300, + 545, + 332 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 351, + 545, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 351, + 545, + 422 + ], + "spans": [ + { + "bbox": [ + 304, + 351, + 545, + 422 + ], + "type": "text", + "content": "Transformer Block for Decoder. The transformer block is based on the transformer decoder of DETR. We continuously use four transformer blocks in the decoder of the network. We set the channel " + }, + { + "bbox": [ + 304, + 351, + 545, + 422 + ], + "type": "inline_equation", + "content": "C = 256" + }, + { + "bbox": [ + 304, + 351, + 545, + 422 + ], + "type": "text", + "content": " and the number of attention heads " + }, + { + "bbox": [ + 304, + 351, + 545, + 422 + ], + "type": "inline_equation", + "content": "M = 8" + }, + { + "bbox": [ + 304, + 351, + 545, + 422 + ], + "type": "text", + "content": " for each multi-head attention module." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 423, + 545, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 423, + 545, + 483 + ], + "spans": [ + { + "bbox": [ + 304, + 423, + 545, + 483 + ], + "type": "text", + "content": "Prediction Stage. After obtaining the semantic and contour feature maps of the portrait from the stacked transformer blocks, we fuse both of them. Then, the alpha mattes predicted by the segmentation head are upsampled to the original size of the input image with bi-linear interpolation." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 305, + 483, + 544, + 507 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 483, + 544, + 507 + ], + "spans": [ + { + "bbox": [ + 305, + 483, + 544, + 507 + ], + "type": "text", + "content": "Loss function. To compute losses during the training stage, we use a Binary CrossEntropy Loss function." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 514, + 544, + 527 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 514, + 544, + 527 + ], + "spans": [ + { + "bbox": [ + 316, + 514, + 544, + 527 + ], + "type": "interline_equation", + "content": "L _ {b c e} = - (g \\times \\log (m) + (1 - g) \\times \\log (1 - m)) \\tag {16}", + "image_path": "d5d049283bf62a195f9e264cbcb77d5046547144c5a17a2f022da8c1b2e1af2c.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 305, + 533, + 545, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 533, + 545, + 556 + ], + "spans": [ + { + "bbox": [ + 305, + 533, + 545, + 556 + ], + "type": "text", + "content": "Where, " + }, + { + "bbox": [ + 305, + 533, + 545, + 556 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 305, + 533, + 545, + 556 + ], + "type": "text", + "content": " is the prediction of the network, " + }, + { + "bbox": [ + 305, + 533, + 545, + 556 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 305, + 533, + 545, + 556 + ], + "type": "text", + "content": " is the ground truth." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 306, + 563, + 521, + 575 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 563, + 521, + 575 + ], + "spans": [ + { + "bbox": [ + 306, + 563, + 521, + 575 + ], + "type": "text", + "content": "4.3. Comparisons to State-of-the-art Methods" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 582, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 582, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 582, + 545, + 712 + ], + "type": "text", + "content": "To demonstrate the progressiveness of EFormer, we compare it against the latest state-of-the-art trimap-free portrait matting solutions. These are including DeepLabV3[4] with ResNet101[11] backbone, BGMv2[21] with ResNet50[11] backbone, MODNet[17], and RVM[22]. Our assessment is based on the testing datasets of VideoMatte240K-JPEG-SD and AIM, respectively. As the training files for MODNet are not available, we use its official weights. Furthermore, since BGMv2, DeepLabV3, and RVM have already been trained on all datasets, our method can be compared with them in a fair manner." + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3885" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 51, + 70, + 547, + 339 + ], + "blocks": [ + { + "bbox": [ + 51, + 70, + 547, + 339 + ], + "lines": [ + { + "bbox": [ + 51, + 70, + 547, + 339 + ], + "spans": [ + { + "bbox": [ + 51, + 70, + 547, + 339 + ], + "type": "image", + "image_path": "a44cfc197d899a897cdae63917124a1992d807f4b3ea9c1a1cb1af45bf9a7ef0.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 345, + 547, + 378 + ], + "lines": [ + { + "bbox": [ + 46, + 345, + 547, + 378 + ], + "spans": [ + { + "bbox": [ + 46, + 345, + 547, + 378 + ], + "type": "text", + "content": "Figure 4. Visualization of portrait matte predictions from MODnet, BGMv2, and EFormer under challenging image from the test set of VideoMatte240K (JPEG SD Format). Our model shows a better ability to distinguish ambiguous foreground contours, as indicated by the red box. Please zoom in to view the detailed information." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 52, + 397, + 282, + 483 + ], + "blocks": [ + { + "bbox": [ + 52, + 397, + 282, + 483 + ], + "lines": [ + { + "bbox": [ + 52, + 397, + 282, + 483 + ], + "spans": [ + { + "bbox": [ + 52, + 397, + 282, + 483 + ], + "type": "table", + "html": "
ModelMAD↓MSE↓Grad↓Conn↓
DeepLabv3[4]18.158614.76236.00152.8824
MODnet[17]13.14008.96562.60282.1543
BGMv2[21]7.04712.35431.56211.0651
RVM[22]6.13572.26470.65230.2866
EFormer4.13571.42800.37600.1753
", + "image_path": "d66d74013182b05d955a7fc314b2f295b3303d5a4d703b858ad43c84fa974857.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 544, + 287, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 544, + 287, + 605 + ], + "spans": [ + { + "bbox": [ + 46, + 544, + 287, + 605 + ], + "type": "text", + "content": "As shown in the Tab. 1, Tab. 2 and Tab. 3, our model exhibits superior performance in portrait matte tasks. It has a stronger capability to distinguish the contour of a portrait. And it can finely segment edge details (e.g. fingers), as exemplified in Fig. 4." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 611, + 147, + 623 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 611, + 147, + 623 + ], + "spans": [ + { + "bbox": [ + 47, + 611, + 147, + 623 + ], + "type": "text", + "content": "4.4. Ablation Studies" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 629, + 288, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 629, + 288, + 712 + ], + "spans": [ + { + "bbox": [ + 46, + 629, + 288, + 712 + ], + "type": "text", + "content": "Training Epochs. In the ablation study with an image resolution of " + }, + { + "bbox": [ + 46, + 629, + 288, + 712 + ], + "type": "inline_equation", + "content": "224 \\times 224" + }, + { + "bbox": [ + 46, + 629, + 288, + 712 + ], + "type": "text", + "content": ", we conduct 25 training epochs. Every 5 epochs of training, we evaluate the training effect of the model. At the same time, the learning rate is reduced by 0.8 times. The experimental results indicate that our model achieves optimal performance in the 20 epoch, as shown in the Tab. 4." + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 326, + 397, + 526, + 483 + ], + "blocks": [ + { + "bbox": [ + 46, + 491, + 287, + 524 + ], + "lines": [ + { + "bbox": [ + 46, + 491, + 287, + 524 + ], + "spans": [ + { + "bbox": [ + 46, + 491, + 287, + 524 + ], + "type": "text", + "content": "Table 3. Comparisons on the test set of VideoMatte240K-JPEG-SD. Bold indicates the best performance among these models under the inputs with the same resolution at " + }, + { + "bbox": [ + 46, + 491, + 287, + 524 + ], + "type": "inline_equation", + "content": "224 \\times 224" + }, + { + "bbox": [ + 46, + 491, + 287, + 524 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 326, + 397, + 526, + 483 + ], + "lines": [ + { + "bbox": [ + 326, + 397, + 526, + 483 + ], + "spans": [ + { + "bbox": [ + 326, + 397, + 526, + 483 + ], + "type": "table", + "html": "
MAD↓MSE↓Grad↓Conn↓
5-epoch4.39531.48240.40890.1865
10-epoch4.31081.47660.39080.1834
15-epoch4.28671.52380.40870.1833
20-epoch4.13571.42800.37600.1753
25-epoch4.18111.47970.38580.1779
", + "image_path": "3e01d395210f49963e3779f01492181715d5c54a39bd96f04650def1402a3c47.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 491, + 547, + 535 + ], + "lines": [ + { + "bbox": [ + 304, + 491, + 547, + 535 + ], + "spans": [ + { + "bbox": [ + 304, + 491, + 547, + 535 + ], + "type": "text", + "content": "Table 4. Ablation Study on the Training Epochs with the test set of VideoMatte240K-JPEG-SD (resolution at " + }, + { + "bbox": [ + 304, + 491, + 547, + 535 + ], + "type": "inline_equation", + "content": "224 \\times 224" + }, + { + "bbox": [ + 304, + 491, + 547, + 535 + ], + "type": "text", + "content": "). **Bold** indicates the best performance among these models. We use the result of 20-epoch training as a strong baseline model." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 557, + 545, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 557, + 545, + 628 + ], + "spans": [ + { + "bbox": [ + 304, + 557, + 545, + 628 + ], + "type": "text", + "content": "The Functional Role of Each Module. To demonstrate the effectiveness of the model in capturing and extracting contour details and portrait semantics, we analyze the outputs of CA, SA, CEEB, and SEB, and utilize Class Activation Map (CAM) to visualize the attention allocation of each layer's features, as shown in Fig. 5." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 629, + 545, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 629, + 545, + 700 + ], + "spans": [ + { + "bbox": [ + 304, + 629, + 545, + 700 + ], + "type": "text", + "content": "From the first line of Fig. 5, it can be evident that after the cross-attention layer, the model clearly strengthens its attention to the contours of the portrait. Following the contour prior, the self-attention layer captures the portrait's semantics thoroughly. CEEB and SEB models provide sharper contour and more comprehensive semantic information." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 317, + 701, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 701, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 317, + 701, + 545, + 714 + ], + "type": "text", + "content": "Comparing the second and first lines of Fig. 5, it is ob" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3886" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 49, + 71, + 286, + 218 + ], + "blocks": [ + { + "bbox": [ + 49, + 71, + 286, + 218 + ], + "lines": [ + { + "bbox": [ + 49, + 71, + 286, + 218 + ], + "spans": [ + { + "bbox": [ + 49, + 71, + 286, + 218 + ], + "type": "image", + "image_path": "f4a86092d25db2d4a4c9dd6b16240410a2a59c5e7287d4d91926926071d90d7f.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 227, + 288, + 283 + ], + "lines": [ + { + "bbox": [ + 46, + 227, + 288, + 283 + ], + "spans": [ + { + "bbox": [ + 46, + 227, + 288, + 283 + ], + "type": "text", + "content": "Figure 5. We use Grad-CAM to visualize the attention distribution of various features in CA, SA, CEEB, and SEB. The model in the first row contains four complete modules. For comparison, the SA layer and CA layer were removed from the models in the second and third rows, respectively." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 55, + 296, + 279, + 357 + ], + "blocks": [ + { + "bbox": [ + 55, + 296, + 279, + 357 + ], + "lines": [ + { + "bbox": [ + 55, + 296, + 279, + 357 + ], + "spans": [ + { + "bbox": [ + 55, + 296, + 279, + 357 + ], + "type": "table", + "html": "
CASAMAD↓MSE↓Grad↓Conn↓
✓4.27621.53440.40920.1823
✓4.17101.43540.37730.1769
✓✓4.13571.42800.37600.1753
", + "image_path": "9284b5286ab0833f03b0deb33b734389a968a9090ebdffccfb98e06530372592.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 434, + 287, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 434, + 287, + 495 + ], + "spans": [ + { + "bbox": [ + 46, + 434, + 287, + 495 + ], + "type": "text", + "content": "served that removing the self-attention layer considerably weakens the semantic information extracted by SEB. In the comparison between the third and first lines of Fig. 5, it can be indicated that removing the cross-attention layer significantly weakens the contour information extracted by CEEB." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 495, + 287, + 542 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 495, + 287, + 542 + ], + "spans": [ + { + "bbox": [ + 46, + 495, + 287, + 542 + ], + "type": "text", + "content": "After qualitative analysis, the prediction results of all the above models are evaluated. It is concluded that the model with CA, SA, CEEB, and SEB is both effective and rational, as demonstrated in the Tab. 5." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 544, + 287, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 544, + 287, + 700 + ], + "spans": [ + { + "bbox": [ + 46, + 544, + 287, + 700 + ], + "type": "text", + "content": "As is well known, when it comes to extracting features for portrait segmentation tasks, semantic information plays a significant role, in contrast, contour details account for a relatively small proportion. Owing to this, as shown in the Tab. 5 and Fig. 5, the model without SA leads to a significant increase in all metrics due to the lack of amount of semantic information, while the model without CA experiences a slight increase in all indicators due to the loss of contour details. In order to fully harness the potential of the vision transformer, we integrate both CA and SA to ensure semantic completeness and enhance attention towards contour details, thereby achieving more comprehensive and fine-grained segmentation results." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 701, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 701, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 701, + 287, + 714 + ], + "type": "text", + "content": "The Selection of " + }, + { + "bbox": [ + 47, + 701, + 287, + 714 + ], + "type": "inline_equation", + "content": "F_{HR}" + }, + { + "bbox": [ + 47, + 701, + 287, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 701, + 287, + 714 + ], + "type": "inline_equation", + "content": "F_{LR}" + }, + { + "bbox": [ + 47, + 701, + 287, + 714 + ], + "type": "text", + "content": ". In our base model," + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 306, + 70, + 545, + 135 + ], + "blocks": [ + { + "bbox": [ + 46, + 365, + 287, + 409 + ], + "lines": [ + { + "bbox": [ + 46, + 365, + 287, + 409 + ], + "spans": [ + { + "bbox": [ + 46, + 365, + 287, + 409 + ], + "type": "text", + "content": "Table 5. Ablation Study on the Functional Role of CA and SA with the test set of VideoMatte240K-JPEG-SD (resolution at " + }, + { + "bbox": [ + 46, + 365, + 287, + 409 + ], + "type": "inline_equation", + "content": "224 \\times 224" + }, + { + "bbox": [ + 46, + 365, + 287, + 409 + ], + "type": "text", + "content": "). **Bold** indicates the best performance among these models." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 306, + 70, + 545, + 135 + ], + "lines": [ + { + "bbox": [ + 306, + 70, + 545, + 135 + ], + "spans": [ + { + "bbox": [ + 306, + 70, + 545, + 135 + ], + "type": "table", + "html": "
FHRFLRMAD↓MSE↓Grad↓Conn↓
Fenc1/8Fenc1/164.13571.42800.37600.1753
Fenc1/4Fenc1/83.99701.33430.33280.1679
Fenc1/4Fenc1/163.97301.29490.32150.1665
", + "image_path": "e71b74d17dbb4b9a3f2efbc7a347c6138f07d21bd6f697e21c3171dcf0fce96a.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 143, + 545, + 177 + ], + "lines": [ + { + "bbox": [ + 305, + 143, + 545, + 177 + ], + "spans": [ + { + "bbox": [ + 305, + 143, + 545, + 177 + ], + "type": "text", + "content": "Table 6. Ablation Study on the Selection of " + }, + { + "bbox": [ + 305, + 143, + 545, + 177 + ], + "type": "inline_equation", + "content": "F_{HR}" + }, + { + "bbox": [ + 305, + 143, + 545, + 177 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 143, + 545, + 177 + ], + "type": "inline_equation", + "content": "F_{LR}" + }, + { + "bbox": [ + 305, + 143, + 545, + 177 + ], + "type": "text", + "content": " with the test set of VideoMatte240K-JPEG-SD (resolution at " + }, + { + "bbox": [ + 305, + 143, + 545, + 177 + ], + "type": "inline_equation", + "content": "224 \\times 224" + }, + { + "bbox": [ + 305, + 143, + 545, + 177 + ], + "type": "text", + "content": "). Bold indicates the best performance among these models." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 205, + 545, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 205, + 545, + 361 + ], + "spans": [ + { + "bbox": [ + 304, + 205, + 545, + 361 + ], + "type": "text", + "content": "considering the large resolution of feature maps would bring large computation cost, we use the " + }, + { + "bbox": [ + 304, + 205, + 545, + 361 + ], + "type": "inline_equation", + "content": "F_{enc}^{\\frac{1}{8}}" + }, + { + "bbox": [ + 304, + 205, + 545, + 361 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 304, + 205, + 545, + 361 + ], + "type": "inline_equation", + "content": "F_{HR}" + }, + { + "bbox": [ + 304, + 205, + 545, + 361 + ], + "type": "text", + "content": " (high-resolution features) and use the " + }, + { + "bbox": [ + 304, + 205, + 545, + 361 + ], + "type": "inline_equation", + "content": "F_{enc}^{\\frac{1}{16}}" + }, + { + "bbox": [ + 304, + 205, + 545, + 361 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 304, + 205, + 545, + 361 + ], + "type": "inline_equation", + "content": "F_{LR}" + }, + { + "bbox": [ + 304, + 205, + 545, + 361 + ], + "type": "text", + "content": " (low-resolution features). Then the " + }, + { + "bbox": [ + 304, + 205, + 545, + 361 + ], + "type": "inline_equation", + "content": "F_{HR}" + }, + { + "bbox": [ + 304, + 205, + 545, + 361 + ], + "type": "text", + "content": " serves as the source of contour flow, providing high-frequency contour detail features. Correspondingly, the " + }, + { + "bbox": [ + 304, + 205, + 545, + 361 + ], + "type": "inline_equation", + "content": "F_{LR}" + }, + { + "bbox": [ + 304, + 205, + 545, + 361 + ], + "type": "text", + "content": " serves as the source of semantic flow, providing low-frequency semantic information. As shown in the Tab. 6, after using " + }, + { + "bbox": [ + 304, + 205, + 545, + 361 + ], + "type": "inline_equation", + "content": "F_{enc}^{\\frac{1}{4}}" + }, + { + "bbox": [ + 304, + 205, + 545, + 361 + ], + "type": "text", + "content": " with higher resolution as " + }, + { + "bbox": [ + 304, + 205, + 545, + 361 + ], + "type": "inline_equation", + "content": "F_{HR}" + }, + { + "bbox": [ + 304, + 205, + 545, + 361 + ], + "type": "text", + "content": ", our model can capture more high-frequency contour detail features and its performance is better. And after using " + }, + { + "bbox": [ + 304, + 205, + 545, + 361 + ], + "type": "inline_equation", + "content": "F_{enc}^{\\frac{1}{16}}" + }, + { + "bbox": [ + 304, + 205, + 545, + 361 + ], + "type": "text", + "content": " with lower resolution as " + }, + { + "bbox": [ + 304, + 205, + 545, + 361 + ], + "type": "inline_equation", + "content": "F_{LR}" + }, + { + "bbox": [ + 304, + 205, + 545, + 361 + ], + "type": "text", + "content": ", our model can use more concentrated low-frequency semantic information, which makes all indicators optimal." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 384, + 378, + 396 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 384, + 378, + 396 + ], + "spans": [ + { + "bbox": [ + 306, + 384, + 378, + 396 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 408, + 545, + 612 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 408, + 545, + 612 + ], + "spans": [ + { + "bbox": [ + 304, + 408, + 545, + 612 + ], + "type": "text", + "content": "In this paper, we propose a strategy that use of cross-attention module between different resolutions guides the EFormer to autonomously locate and capture high-frequency contour features. Without manual parameter tuning, the proposed method can adaptability capture high-frequency details neglected by previous transformer-based methods. Based on the cross-attention module, we further construct a semantic and contour detector (SCD), which can enhance the transformer's attention to both the high-frequency contour and low-frequency semantic features. Additionally, we build a contour-edge extraction branch (CEEB) and a semantic extraction branch (SEB) to extract finer high-frequency contour details and more comprehensive portrait's low-frequency semantic information. Extensive experiments show that our method outperforms previous portrait matting solutions and exhibits robust performance during testing and inference." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 634, + 404, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 634, + 404, + 647 + ], + "spans": [ + { + "bbox": [ + 306, + 634, + 404, + 647 + ], + "type": "text", + "content": "Acknowledgement" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 657, + 545, + 703 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 657, + 545, + 703 + ], + "spans": [ + { + "bbox": [ + 305, + 657, + 545, + 703 + ], + "type": "text", + "content": "The work was jointly supported by the National Natural Science Foundations of China under grant No.62272364, the Teaching Reform Project of Shaanxi Higher Continuing Education under Grant No.21XJZ004." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3887" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 714 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 135 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 135 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 135 + ], + "type": "text", + "content": "[1] Yagiz Aksoy, Tunc Ozan Aydin, and Marc Pollefeys. Designing effective inter-pixel information flow for natural image matting. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 29-37, 2017. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 137, + 288, + 169 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 137, + 288, + 169 + ], + "spans": [ + { + "bbox": [ + 53, + 137, + 288, + 169 + ], + "type": "text", + "content": "[2] Yagiz Aksoy, Tae-Hyun Oh, Sylvain Paris, Marc Pollefeys, and Wojciech Matusik. Semantic soft segmentation. ACM Transactions on Graphics (TOG), 37(4):1-13, 2018. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 171, + 287, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 171, + 287, + 215 + ], + "spans": [ + { + "bbox": [ + 53, + 171, + 287, + 215 + ], + "type": "text", + "content": "[3] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In European conference on computer vision, pages 213-229. Springer, 2020. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 217, + 287, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 217, + 287, + 258 + ], + "spans": [ + { + "bbox": [ + 53, + 217, + 287, + 258 + ], + "type": "text", + "content": "[4] Liang-Chieh Chen, George Papandreou, Florian Schroff, and Hartwig Adam. Rethinking atrous convolution for semantic image segmentation. arXiv preprint arXiv:1706.05587, 2017. 6, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 261, + 287, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 261, + 287, + 304 + ], + "spans": [ + { + "bbox": [ + 53, + 261, + 287, + 304 + ], + "type": "text", + "content": "[5] Bowen Cheng, Alex Schwing, and Alexander Kirillov. Per-pixel classification is not all you need for semantic segmentation. Advances in Neural Information Processing Systems, 34:17864-17875, 2021. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 307, + 287, + 361 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 307, + 287, + 361 + ], + "spans": [ + { + "bbox": [ + 53, + 307, + 287, + 361 + ], + "type": "text", + "content": "[6] Bowen Cheng, Ishan Misra, Alexander G Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1290–1299, 2022. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 363, + 287, + 428 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 363, + 287, + 428 + ], + "spans": [ + { + "bbox": [ + 53, + 363, + 287, + 428 + ], + "type": "text", + "content": "[7] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 430, + 287, + 473 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 430, + 287, + 473 + ], + "spans": [ + { + "bbox": [ + 53, + 430, + 287, + 473 + ], + "type": "text", + "content": "[8] Dazhao Du, Bing Su, Yu Li, Zhongang Qi, Lingyu Si, and Ying Shan. Efficient u-transformer with boundary-aware loss for action segmentation. arXiv preprint arXiv:2205.13425, 2022. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 475, + 287, + 530 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 475, + 287, + 530 + ], + "spans": [ + { + "bbox": [ + 53, + 475, + 287, + 530 + ], + "type": "text", + "content": "[9] Xiaoxue Feng, Xiaohui Liang, and Zili Zhang. A cluster sampling method for image matting via sparse coding. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part II 14, pages 204-219. Springer, 2016. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 532, + 287, + 565 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 532, + 287, + 565 + ], + "spans": [ + { + "bbox": [ + 48, + 532, + 287, + 565 + ], + "type": "text", + "content": "[10] Kaiming He, Christoph Rhemann, Carsten Rother, Xiaou Tang, and Jian Sun. A global sampling method for alpha matting. In CVPR 2011, pages 2049-2056. IEEE, 2011. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 567, + 287, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 567, + 287, + 609 + ], + "spans": [ + { + "bbox": [ + 48, + 567, + 287, + 609 + ], + "type": "text", + "content": "[11] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 611, + 287, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 611, + 287, + 644 + ], + "spans": [ + { + "bbox": [ + 48, + 611, + 287, + 644 + ], + "type": "text", + "content": "[12] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In Proceedings of the IEEE international conference on computer vision, pages 2961-2969, 2017. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 646, + 287, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 646, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 646, + 287, + 689 + ], + "type": "text", + "content": "[13] Jitesh Jain, Anukriti Singh, Nikita Orlov, Zilong Huang, Jiachen Li, Steven Walton, and Humphrey Shi. Semask: Semantically masked transformers for semantic segmentation. arXiv preprint arXiv:2112.12782, 2021. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 691, + 287, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 287, + 714 + ], + "type": "text", + "content": "[14] Jubin Johnson, Ehsan Shahrian Varnousfaderani, Hisham Cholakkal, and Deepu Rajan. Sparse coding for alpha mat" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 545, + 714 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "text", + "content": "ting. IEEE Transactions on Image Processing, 25(7):3032-3043, 2016. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 96, + 545, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 96, + 545, + 140 + ], + "spans": [ + { + "bbox": [ + 308, + 96, + 545, + 140 + ], + "type": "text", + "content": "[15] Levent Karacan, Aykut Erdem, and Erkut Erdem. Image matting with kl-divergence based sparse sampling. In Proceedings of the IEEE international conference on computer vision, pages 424-432, 2015. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 141, + 545, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 141, + 545, + 195 + ], + "spans": [ + { + "bbox": [ + 308, + 141, + 545, + 195 + ], + "type": "text", + "content": "[16] Lei Ke, Martin Danelljan, Xia Li, Yu-Wing Tai, Chi-Keung Tang, and Fisher Yu. Mask transfiner for high-quality instance segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4412–4421, 2022. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 197, + 545, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 197, + 545, + 251 + ], + "spans": [ + { + "bbox": [ + 308, + 197, + 545, + 251 + ], + "type": "text", + "content": "[17] Zhanghan Ke, Jiayu Sun, Kaican Li, Qiong Yan, and Rynson WH Lau. Modnet: Real-time trimap-free portrait matting via objective decomposition. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1140-1147, 2022. 2, 3, 6, 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 253, + 545, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 253, + 545, + 297 + ], + "spans": [ + { + "bbox": [ + 308, + 253, + 545, + 297 + ], + "type": "text", + "content": "[18] Alexander Kirillov, Yuxin Wu, Kaiming He, and Ross Girshick. Pointrend: Image segmentation as rendering. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9799-9808, 2020. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 298, + 545, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 298, + 545, + 331 + ], + "spans": [ + { + "bbox": [ + 308, + 298, + 545, + 331 + ], + "type": "text", + "content": "[19] Anat Levin, Alex Rav-Acha, and Dani Lischinski. Spectral matting. IEEE transactions on pattern analysis and machine intelligence, 30(10):1699-1712, 2008. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 332, + 545, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 332, + 545, + 365 + ], + "spans": [ + { + "bbox": [ + 308, + 332, + 545, + 365 + ], + "type": "text", + "content": "[20] Yaoyi Li and Hongtao Lu. Natural image matting via guided contextual attention. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 11450-11457, 2020. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 366, + 545, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 366, + 545, + 431 + ], + "spans": [ + { + "bbox": [ + 308, + 366, + 545, + 431 + ], + "type": "text", + "content": "[21] Shanchuan Lin, Andrey Ryabtsev, Soumyadip Sengupta, Brian L Curless, Steven M Seitz, and Ira Kemelmacher-Shlizerman. Real-time high-resolution background matting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8762-8771, 2021. 2, 3, 6, 7" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 433, + 545, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 433, + 545, + 487 + ], + "spans": [ + { + "bbox": [ + 308, + 433, + 545, + 487 + ], + "type": "text", + "content": "[22] Shanchuan Lin, Linjie Yang, Imran Saleemi, and Soumyadip Sengupta. Robust high-resolution video matting with temporal guidance. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 238-247, 2022. 2, 3, 6, 7" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 489, + 545, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 489, + 545, + 533 + ], + "spans": [ + { + "bbox": [ + 308, + 489, + 545, + 533 + ], + "type": "text", + "content": "[23] Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dólar. Focal loss for dense object detection. In Proceedings of the IEEE international conference on computer vision, pages 2980-2988, 2017. 2, 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 534, + 545, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 534, + 545, + 590 + ], + "spans": [ + { + "bbox": [ + 308, + 534, + 545, + 590 + ], + "type": "text", + "content": "[24] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10012-10022, 2021. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 590, + 545, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 590, + 545, + 612 + ], + "spans": [ + { + "bbox": [ + 308, + 590, + 545, + 612 + ], + "type": "text", + "content": "[25] Namuk Park and Songkuk Kim. How do vision transformers work? arXiv preprint arXiv:2202.06709, 2022. 2, 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 613, + 545, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 613, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 308, + 613, + 545, + 689 + ], + "type": "text", + "content": "[26] Olivier Petit, Nicolas Thome, Clement Rambour, Loic Themyr, Toby Collins, and Luc Soler. U-net transformer: Self and cross attention for medical image segmentation. In Machine Learning in Medical Imaging: 12th International Workshop, MLMI 2021, Held in Conjunction with MICCAI 2021, Strasbourg, France, September 27, 2021, Proceedings 12, pages 267-276. Springer, 2021. 3" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 691, + 545, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 691, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 308, + 691, + 545, + 714 + ], + "type": "text", + "content": "[27] René Ranftl, Alexey Bochkovskiy, and Vladlen Koltun. Vision transformers for dense prediction. In Proceedings of" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3888" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 73, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 67, + 73, + 286, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 73, + 286, + 95 + ], + "spans": [ + { + "bbox": [ + 67, + 73, + 286, + 95 + ], + "type": "text", + "content": "the IEEE/CVF international conference on computer vision, pages 12179-12188, 2021.3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 96, + 287, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 96, + 287, + 149 + ], + "spans": [ + { + "bbox": [ + 48, + 96, + 287, + 149 + ], + "type": "text", + "content": "[28] Soumyadip Sengupta, Vivek Jayaram, Brian Curless, Steven M Seitz, and Ira Kemelmacher-Shlizerman. Background matting: The world is your green screen. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2291-2300, 2020. 2, 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 151, + 287, + 193 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 151, + 287, + 193 + ], + "spans": [ + { + "bbox": [ + 49, + 151, + 287, + 193 + ], + "type": "text", + "content": "[29] Chenyang Si, Weihao Yu, Pan Zhou, Yichen Zhou, Xinchao Wang, and Shuicheng Yan. Inception transformer. Advances in Neural Information Processing Systems, 35:23495-23509, 2022. 2, 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 195, + 287, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 195, + 287, + 248 + ], + "spans": [ + { + "bbox": [ + 49, + 195, + 287, + 248 + ], + "type": "text", + "content": "[30] Yanan Sun, Guanzhi Wang, Qiao Gu, Chi-Keung Tang, and Yu-Wing Tai. Deep video matting via spatio-temporal alignment and aggregation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6975–6984, 2021. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 49, + 251, + 287, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 251, + 287, + 304 + ], + "spans": [ + { + "bbox": [ + 49, + 251, + 287, + 304 + ], + "type": "text", + "content": "[31] Huiyu Wang, Yukun Zhu, Hartwig Adam, Alan Yuille, and Liang-Chieh Chen. Max-deeplab: End-to-end panoptic segmentation with mask transformers. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5463-5474, 2021. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 49, + 305, + 287, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 305, + 287, + 358 + ], + "spans": [ + { + "bbox": [ + 49, + 305, + 287, + 358 + ], + "type": "text", + "content": "[32] Jian Wang, Chenhui Gou, Qiman Wu, Haocheng Feng, Junyu Han, Errui Ding, and Jingdong Wang. Rrformer: Efficient design for real-time semantic segmentation with transformer. Advances in Neural Information Processing Systems, 35:7423-7436, 2022. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 49, + 361, + 287, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 361, + 287, + 392 + ], + "spans": [ + { + "bbox": [ + 49, + 361, + 287, + 392 + ], + "type": "text", + "content": "[33] Yu Wang, Yi Niu, Peiyong Duan, Jianwei Lin, and Yuanjie Zheng. Deep propagation based image matting. In IJCAI, pages 999-1006, 2018. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 49, + 394, + 287, + 446 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 394, + 287, + 446 + ], + "spans": [ + { + "bbox": [ + 49, + 394, + 287, + 446 + ], + "type": "text", + "content": "[34] Enze Xie, Wenhai Wang, Zhiding Yu, Anima Anandkumar, Jose M Alvarez, and Ping Luo. Segformer: Simple and efficient design for semantic segmentation with transformers. Advances in Neural Information Processing Systems, 34:12077-12090, 2021. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 49, + 449, + 287, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 449, + 287, + 491 + ], + "spans": [ + { + "bbox": [ + 49, + 449, + 287, + 491 + ], + "type": "text", + "content": "[35] Ning Xu, Brian Price, Scott Cohen, and Thomas Huang. Deep image matting. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2970-2979, 2017. 1, 2, 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 49, + 494, + 287, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 494, + 287, + 536 + ], + "spans": [ + { + "bbox": [ + 49, + 494, + 287, + 536 + ], + "type": "text", + "content": "[36] Haichao Yu, Ning Xu, Zilong Huang, Yuqian Zhou, and Humphrey Shi. High-resolution deep image matting. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 3217-3224, 2021. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 49, + 537, + 287, + 591 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 537, + 287, + 591 + ], + "spans": [ + { + "bbox": [ + 49, + 537, + 287, + 591 + ], + "type": "text", + "content": "[37] Yunke Zhang, Chi Wang, Miaomiao Cui, Peiran Ren, Xuansong Xie, Xian-Sheng Hua, Hujun Bao, Qixing Huang, and Weiwei Xu. Attention-guided temporally coherent video object matting. In Proceedings of the 29th ACM International Conference on Multimedia, pages 5128-5137, 2021. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 49, + 593, + 287, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 593, + 287, + 667 + ], + "spans": [ + { + "bbox": [ + 49, + 593, + 287, + 667 + ], + "type": "text", + "content": "[38] Sixiao Zheng, Jiachen Lu, Hengshuang Zhao, Xiatian Zhu, Zekun Luo, Yabiao Wang, Yanwei Fu, Jianfeng Feng, Tao Xiang, Philip HS Torr, et al. Rethinking semantic segmentation from a sequence-to-sequence perspective with transformers. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6881-6890, 2021. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 49, + 670, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 670, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 49, + 670, + 287, + 712 + ], + "type": "text", + "content": "[39] Bingke Zhu, Yingying Chen, Jinqiao Wang, Si Liu, Bo Zhang, and Ming Tang. Fast deep matting for portrait animation on mobile phone. In Proceedings of the 25th ACM international conference on Multimedia, pages 297-305, 2017. 3" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "3889" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/EGTR_ Extracting Graph from Transformer for Scene Graph Generation/174c2199-8aa8-4e45-8d2e-205eb2e2e099_content_list.json b/2024/EGTR_ Extracting Graph from Transformer for Scene Graph Generation/174c2199-8aa8-4e45-8d2e-205eb2e2e099_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..7c01f0e450b32bb779622408cc8de7bb97019bf1 --- /dev/null +++ b/2024/EGTR_ Extracting Graph from Transformer for Scene Graph Generation/174c2199-8aa8-4e45-8d2e-205eb2e2e099_content_list.json @@ -0,0 +1,1703 @@ +[ + { + "type": "text", + "text": "EGTR: Extracting Graph from Transformer for Scene Graph Generation", + "text_level": 1, + "bbox": [ + 112, + 130, + 854, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Jinbae Im1 JeongYeon Nam1 Nokyung Park1,2,3* Hyungmin Lee2 Seunghyun Park1 \n1NAVER Cloud AI 2NAVER 3Korea University \n{jinbae.im, jy.nam, nokyung.park99, hyungmin.lee, seung.park}@navercorp.com", + "bbox": [ + 122, + 179, + 843, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 268, + 313, + 284 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Scene Graph Generation (SGG) is a challenging task of detecting objects and predicting relationships between objects. After DETR was developed, one-stage SGG models based on a one-stage object detector have been actively studied. However, complex modeling is used to predict the relationship between objects, and the inherent relationship between object queries learned in the multi-head self-attention of the object detector has been neglected. We propose a lightweight one-stage SGG model that extracts the relation graph from the various relationships learned in the multi-head self-attention layers of the DETR decoder. By fully utilizing the self-attention by-products, the relation graph can be extracted effectively with a shallow relation extraction head. Considering the dependency of the relation extraction task on the object detection task, we propose a novel relation smoothing technique that adjusts the relation label adaptively according to the quality of the detected objects. By the relation smoothing, the model is trained according to the continuous curriculum that focuses on object detection task at the beginning of training and performs multi-task learning as the object detection performance gradually improves. Furthermore, we propose a connectivity prediction task that predicts whether a relation exists between object pairs as an auxiliary task of the relation extraction. We demonstrate the effectiveness and efficiency of our method for the Visual Genome and Open Image V6 datasets. Our code is publicly available at https://github.com/naver-ai/egtr.", + "bbox": [ + 75, + 301, + 473, + 724 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 756, + 209, + 771 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Scene Graph Generation (SGG) [8] aims to generate a scene graph that represents objects as nodes and relationships between objects as edges from an image as shown in Fig. 1b. SGG is a challenging task as it is required to go beyond simply detecting objects but predicting the relationships between them based on a comprehensive understanding of the", + "bbox": [ + 75, + 782, + 468, + 875 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/da086494a3493dff7e071faf8fea070928900aae618d983c549510da6ad11f8c.jpg", + "image_caption": [ + "(a) Visual Genome example" + ], + "image_footnote": [], + "bbox": [ + 506, + 271, + 746, + 395 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/9fcc74a969bc741c22cf5e31e09c3ed6a0b106a111765290803f5a92ea5dbd50.jpg", + "image_caption": [ + "(b) Scene Graph" + ], + "image_footnote": [], + "bbox": [ + 764, + 268, + 883, + 323 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/d3d669486085177c87b65b065112e8bb0b5abe8f60a9d5e1549828477fc1cbc7.jpg", + "image_caption": [ + "(c) Attention Graph", + "Figure 1. Motivation. SGG task aims to predict scene graph (Fig. 1b) with objects as nodes and relations as edges. We draw a plausible attention graph (Fig. 1c) simply by connecting objects with high attention weights to edges from the self-attention layers of the pre-trained DETR. It shows the potential for the self-attention from the object detector to contain rich information that aids in predicting the relations of the scene graph." + ], + "image_footnote": [], + "bbox": [ + 767, + 343, + 883, + 397 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "scene. Since a scene graph provides structural information of the image, it can be used for various vision tasks that require a higher level of understanding and reasoning about images such as image captioning [5, 10, 41], image retrieval [8, 28], and visual question answering [14, 43].", + "bbox": [ + 496, + 535, + 892, + 611 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Most previous studies [16, 19, 33, 39, 42] took two-stage SGG approaches that detect objects first and then predict their relations. However, these approaches incurred high computational costs and risks of error propagation from the object detection. To address the drawbacks of the two-stage approaches, one-stage SGG models that perform object detection and relation prediction at once [3, 9, 17, 20, 26, 29, 35] have been studied recently, leveraging one-stage object detectors such as DETR [1]. Since edges in a scene graph can be represented as subject-predicate-object triplets, many studies embraced triplet-based approaches as shown in Figs. 2a and 2b. However, object-triplet detection models [3, 17] required sophisticated triplet detectors to obtain necessary information for triplet queries from the object detector. Moreover, triplet detection models [9, 35] lacked the ability to detect objects without relation such as \"tree\" in Fig. 1b, by focusing solely on the triplet detection without an object detector. Considering that objects without relations account for more than", + "bbox": [ + 496, + 613, + 893, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Most work was done during the internship at Naver Cloud AI.", + "bbox": [ + 94, + 886, + 442, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "24229", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/eeef0756f110b56051eef1ee222a4df8ecbb04eaa0125bb5bc46ca221ac47ed5.jpg", + "image_caption": [ + "(a) Object-Triplet Detection Models" + ], + "image_footnote": [], + "bbox": [ + 107, + 89, + 349, + 205 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/17fbf1da620e9dc2c721e805545b7c7f43c92cb1db5cdc6e1494d2fbdc0930f5.jpg", + "image_caption": [ + "(b) Triplet Detection Models", + "Figure 2. Comparison with existing one-stage SGG models. (a) Object-Triplet Detection Models introduce additional triplet queries and a triplet detector to the object detector. The triplet detector requires additional modules to incorporate information from the object detector into the triplet queries. (b) Triplet Detection Models focus on detecting triplets directly without an object detector. Objects without relations may not be detected. (c) Relation Extraction Models extract relations from the object detector without a separate triplet detector. In particular, ours extracts relations more effectively by utilizing by-products from the self-attention of the object detector." + ], + "image_footnote": [], + "bbox": [ + 385, + 90, + 571, + 205 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/1200dc7f83785412b59a2e3649fd9623129020a4a233efc8ba94d41c3b085407.jpg", + "image_caption": [ + "(c) Relation Extraction Models" + ], + "image_footnote": [], + "bbox": [ + 616, + 90, + 859, + 205 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "$42\\%$ in Visual Genome [12] data, their priority lies in detecting a sub-graph rather than the complete scene graph.", + "bbox": [ + 75, + 309, + 467, + 339 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address the shortcomings of the existing one-stage SGG models, we focus on the relationships between objects inherent in the object detector. As shown in Fig. 1a, objects are related to each other. For instance, when a horse appears in a scene, a person is likely to appear in the scene, and clothing such as hats, jackets, and pants often depicts the person's current situation. From this intuition, there has been a long belief [2, 7, 22] that modeling relationship or context between objects would be beneficial to object detection task. Accordingly, recent one-stage object detectors [1, 32, 46] have incorporated self-attention layers [36] to implicitly model the relationships among the object queries. We hypothesize that self-attention between object queries learned in a one-stage object detector might contain valuable information for predicting triplet outputs. In our preliminary investigation, we are able to extract a plausible attention graph by simply connecting two object queries with high attention weights from the pre-trained DETR, as shown in Fig. 1c. It shows the potential that the attention weights between object queries can be interpreted as relations between them.", + "bbox": [ + 73, + 340, + 468, + 656 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "From the findings, we propose a lightweight one-stage scene graph generator EGTR, which stands for Extracting Graph from TRansformer. We design the model to comprehensively leverage the by-products of the object detector, eliminating the need for a separate triplet detector, as depicted in Fig. 2c. From the multi-head self-attention layers of the object detector, we regard an attention query and key, where their relations are learned in the attention weights, as the subject entity and object entity, respectively. Subsequently, we leverage a shallow classifier to predict relations between them. Due to the abundant information about the relationships among the objects present in the by-products derived from all layers of self-attention, we can effectively extract the scene graph.", + "bbox": [ + 75, + 657, + 468, + 869 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Since the relation extraction task is dependent on the object detection task, we speculate that performing relation ex", + "bbox": [ + 75, + 869, + 468, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "traction without sufficiently learned representations of the object queries might be harmful. Therefore, we devise a novel adaptive smoothing technique that smoothes the value of the ground truth relation label based on the object detection performance. With the adaptive smoothing, the model is trained with a continuous curriculum that initially focuses on object detection and gradually performs multi-task learning. Furthermore, we propose a connectivity prediction task as an auxiliary task for relation extraction, aiming to predict the existence of any relationship between a subject entity and an object entity. This auxiliary task facilitates the acquisition of representations for the relation extraction.", + "bbox": [ + 496, + 309, + 890, + 489 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To verify the effectiveness of the proposed method, we conduct experiments on two representative SGG datasets: Visual Genome [12] and Open Images V6 [13]. By actively utilizing the by-products of the object detector, EGTR shows the best object detection performance and comparable triplet detection performance, with the fewest parameters and the fastest inference speed.", + "bbox": [ + 496, + 492, + 890, + 597 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our main contributions can be summarized as follows:", + "bbox": [ + 517, + 599, + 880, + 612 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose EGTR that generates scene graphs efficiently and effectively by utilizing the multi-head self-attention by-products derived from the object detector.", + "- We present adaptive smoothing, enabling effective multitask learning for both object detection and relation extraction. In addition, the proposed connectivity prediction offers clues to the relation extraction.", + "- Our comprehensive experiments show the superiority of the proposed model framework and the effectiveness of the devised training techniques." + ], + "bbox": [ + 500, + 614, + 893, + 766 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 782, + 640, + 799 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "SGG models can be categorized into two groups: two-stage models and one-stage models. For two-stage models [4, 11, 15, 16, 19, 23-25, 30, 33, 37, 39, 42, 45], separate object detection model and relation prediction model are trained sequentially. They usually detect $N$ objects from the off-the-shelf object detector such as Faster R-CNN [27],", + "bbox": [ + 496, + 810, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "24230", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "and then all possible combinations of the detected objects are fed into the relation prediction model to predict the relations between each object pair. Although they showed high relation extraction performance, they have the inherent limitation that the object detector that helps with relation extraction is trained separately, resulting in a significant increase in model complexity.", + "bbox": [ + 75, + 90, + 470, + 196 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "As for the one-stage models [3, 9, 17, 20, 26, 29, 35], the object detection and relation prediction are trained in an end-to-end manner. Early studies [20, 26] proposed fully convolutional SGG models and took pixel-based approaches. After DETR [1] brought a huge success as a Transformer [36]-based one-stage object detector, many one-stage SGG studies are based on one-stage object detectors [1, 32, 46]. They efficiently modeled SGG by introducing object queries or triplet queries. We categorize them into three distinct groups: (a) object-triplet detection models, (b) triplet prediction models, and (c) relation extraction models, as shown in Fig. 2.", + "bbox": [ + 75, + 200, + 470, + 383 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Object-Triplet Detection Models. Object-triplet detection models are characterized by introducing additional triplet queries and building a triplet predictor on top of the object detector as shown in Fig. 2a. RelTR [3] introduced paired subject queries and object queries, and SGTR [17] introduced compositional queries decoupled into subjects, objects, and predicates. As the introduced queries are initialized without prior cues from the object detector, the triplet predictor requires modules to incorporate information from the outputs of the object detector and modules for triplet queries to exchange information with each other. It led to the triplet predictor having an intricate structure. In contrast, we refrain from introducing additional queries and regard the attention queries and attention keys, whose relationships are learned in the object detector, as subject and object queries, respectively.", + "bbox": [ + 75, + 392, + 468, + 633 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Triplet Detection Models. Triplet detection models directly detect triplets using triplet queries without an object detector, as shown in Fig. 2b. Iterative SGG [9] introduced subject, object, and predicate queries and modeled the conditional dependencies among them with separate subject, object, and predicate multi-layer Transformer decoders. Inspired by Sparse R-CNN [32], Structured Sparse R-CNN (SSR-CNN) [35] designed triplet queries consisting of subject box, object box, subject, object, and predicate queries. Since it was difficult to train the model only with sparse triplet annotations, SSR-CNN introduced somewhat intricate training details: it detected object pairs with a Siamese Sparse RCNN model and auxiliary queries and performed additional triplet matching using the detected object pairs as a pseudo label. Although they show excellent triplet detection performance, the model architecture became even more intricate to detect subjects and objects separately since", + "bbox": [ + 75, + 643, + 470, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "an explicit object detector is not used. Last but not least, the triplet prediction models focus on detecting a sub-graph composed only of objects with relations, overlooking objects in an image that lack explicit relations.", + "bbox": [ + 496, + 90, + 890, + 151 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Relation Extraction Models. Relation extraction models extract a scene graph using a lightweight relation predictor without separate triplet queries or triplet detector. Relationformer [29] added a special \"[rln]\" token to capture global information in conjunction with object queries. They concatenated the final hidden representations of the object query pairs and the relation token, followed by a shallow fully connected network for relation prediction. In this work, in conjunction with the final hidden representations, we use the inherent relationship information among the object queries learned in the multi-head self-attention layer of the object detector. Furthermore, we propose training techniques to boost multi-task learning, leading to significantly improved performance with architectural simplicity.", + "bbox": [ + 496, + 156, + 890, + 369 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 500, + 382, + 591, + 398 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Preliminaries", + "text_level": 1, + "bbox": [ + 500, + 407, + 640, + 422 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we first introduce the formulation of the SGG task and then provide a brief review of the one-stage object detector that serves as the basis for our study.", + "bbox": [ + 496, + 431, + 890, + 476 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1.1 Formulation", + "text_level": 1, + "bbox": [ + 500, + 494, + 642, + 510 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Scene graph generation is a task of generating a scene graph $\\mathcal{G} = \\{\\mathcal{V},\\mathcal{E}\\}$ from an image, where $\\mathcal{V}$ denotes the node set consisting of objects and $\\mathcal{E}$ denotes the edge set that represents the relation between objects. Each object entity $v_{i}\\in \\mathcal{V}$ has an object category label $v_{i}^{c}$ from a set of object categories $\\mathcal{C}_v$ and box coordinates $v_{i}^{b}$ . Each relation $e_j\\in \\mathcal{E}$ represents the $j$ -th triplet $(s_j,p_j,o_j)$ , where subject $s_j$ and object $o_j$ indicate related object entities and predicate $p_j$ has a relation category label $p_j^c$ from a set of predicate categories $\\mathcal{C}_p$ . Generating $\\mathcal{V}$ and $\\mathcal{E}$ correspond to object detection and relation extraction, respectively.", + "bbox": [ + 496, + 518, + 890, + 686 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1.2 One-stage Object Detector", + "text_level": 1, + "bbox": [ + 500, + 705, + 740, + 720 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our model framework is deeply influenced by DETR [1], a one-stage object detector. In DETR, representations for input images are learned through CNN and Transformer [36] encoder. Transformer decoder enhances object queries by utilizing self- and cross-attention mechanisms. The final object detection heads predict object category labels and box coordinates from the contextualized object queries.", + "bbox": [ + 496, + 728, + 890, + 835 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Backbone. CNN backbone generates $C$ -dimensional feature map $F \\in \\mathbb{R}^{C \\times H_F \\times W_F}$ from an input image $x_{\\mathrm{img}} \\in \\mathbb{R}^{3 \\times H_0 \\times W_0}$ . For ResNet-50 backbone, $H_F$ and $W_F$ are set to $H_0 / 32$ and $W_0 / 32$ , respectively.", + "bbox": [ + 496, + 839, + 890, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "24231", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/8ecf05244754a50b2be1aece4d420418de000403424f1f62df4891ed67db2042.jpg", + "image_caption": [ + "Figure 3. The overall architecture of EGTR. We present a novel lightweight relation extractor, EGTR, which fully utilizes the self-attention of the DETR decoder. We extract query and key representations from each self-attention layer and concatenate them pairwise to represent relations between them. Additionally, we leverage the last hidden representation in the same manner. To effectively aggregate information, we apply a gated sum and then predict relation with a shallow relation head." + ], + "image_footnote": [], + "bbox": [ + 78, + 88, + 704, + 280 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/051d3620529018711d7ac2cc82099c6a645af5e74a4b87f1aedd5638cdb66c02.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 709, + 93, + 885, + 183 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/cbe093ab64a6516b8c1b4273701a8dbd74801876a7784e78ebbac8263356f574.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 714, + 188, + 885, + 279 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Transformer Encoder. Transformer encoder enhances image representations based on multi-head self-attention. After applying a $1 \\times 1$ convolution to reduce dimension from $C$ to $d_{\\mathrm{model}}$ , the feature map is flattened to get an input sequence with length $H_{F}W_{F}$ for the Transformer encoder. Additional positional encodings are used to reflect the spatial information of the feature maps.", + "bbox": [ + 75, + 367, + 470, + 474 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Transformer Decoder. Transformer decoder takes $N$ object queries as input and returns their representations. Each object query detects an individual object, and $N$ is usually set large enough to cover all objects in the image. Through alternating self-attention and cross-attention layers, object queries learn features of object candidates in the input image. Note that causal attention mask is not applied to the self-attention layer; therefore, fully connected $N \\times N$ attention weights among the $N$ object queries are learned from the multi-head self-attention as follows:", + "bbox": [ + 75, + 479, + 468, + 630 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nA _ {h} ^ {l} = f \\left(Q _ {h} ^ {l}, K _ {h} ^ {l}\\right) = \\operatorname {s o f t m a x} \\left(Q _ {h} ^ {l} K _ {h} ^ {T} / \\sqrt {d _ {\\text {h e a d}}}\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 102, + 641, + 468, + 664 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $A_h^l \\in \\mathbb{R}^{N \\times N}$ denotes the attention weights from the $h$ -th head in the $l$ -th layer. $Q_h^l \\in \\mathbb{R}^{N \\times d_{\\mathrm{head}}}$ and $K_h^l \\in \\mathbb{R}^{N \\times d_{\\mathrm{head}}}$ are attention queries and keys, respectively.", + "bbox": [ + 76, + 674, + 470, + 722 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Object Detection Heads. Object detection heads detect $N$ object candidates $\\{\\hat{v}_i\\}_{i = 1}^N$ from the last layer representations of the object queries $Z^{L}$ . For each object query, a linear layer is used to predict the object category label $\\hat{v}_i^c\\in \\mathbb{R}^{N\\times |\\mathcal{C}_v|}$ , and three-layer perceptron with ReLU activation is used to obtain the box coordinates $\\hat{v}_i^b\\in \\mathbb{R}^{N\\times 4}$ .", + "bbox": [ + 76, + 727, + 468, + 820 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Object Detection Loss. To match $N$ detected object candidates $\\{\\hat{v}_i\\}_{i=1}^N$ and $M$ ground truth objects $\\{v_i\\}_{i=1}^M$ , Carion et al. [1] pad the ground truth objects with $\\phi$ (no object) and find the best permutation of predicted objects that minimizes total bipartite matching costs. From the permuted", + "bbox": [ + 75, + 824, + 470, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "predictions $\\{\\hat{v}_i^{\\prime}\\}_{i = 1}^{N}$ , the loss is calculated as follows:", + "bbox": [ + 498, + 366, + 852, + 383 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {o d}} = \\Sigma_ {i = 1} ^ {N} \\left[ \\lambda_ {c} \\mathcal {L} _ {c} \\left(\\hat {v} _ {i} ^ {\\prime c}, v _ {i} ^ {c}\\right) + \\mathbb {1} _ {v _ {i} ^ {c} \\neq \\phi} \\left(\\lambda_ {b} \\mathcal {L} _ {b} \\left(\\hat {v} _ {i} ^ {\\prime b}, v _ {i} ^ {b}\\right)\\right) \\right], \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 511, + 393, + 890, + 414 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathcal{L}_c$ and $\\mathcal{L}_b$ denote loss function for category labels and box coordinates, respectively.", + "bbox": [ + 498, + 425, + 890, + 455 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. EGTR", + "text_level": 1, + "bbox": [ + 500, + 467, + 588, + 482 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We propose a novel lightweight relation extractor, EGTR, which exploits the self-attention of DETR decoder, as depicted in Fig. 3. Since the self-attention weights in Eq. (1) contain $N \\times N$ bidirectional relationships among the $N$ object queries, our relation extractor aims to extract the predicate information from the self-attention weights in the entire $L$ layers, by considering the attention queries and keys as subjects and objects, respectively.", + "bbox": [ + 496, + 491, + 890, + 612 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In order to preserve rich information learned in the self-attention layer, we devise another relation function $f$ between the attention queries and keys instead of the dot product attention in Eq. (1). As concatenation preserves representations completely, we concatenate the representations on all $N \\times N$ pairs of attention queries and keys as shown in Fig. 3 (a) to get the relation representations of the $l$ -th layer $R_{a}^{l} \\in \\mathbb{R}^{N \\times N \\times 2d_{\\mathrm{model}}}$ . Before pairwise concatenation, to help queries and keys play the role of subjects and objects, a linear projection is added as follows:", + "bbox": [ + 496, + 613, + 892, + 763 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nR _ {a} ^ {l} = \\left[ Q ^ {l} W _ {S} ^ {l}; K ^ {l} W _ {O} ^ {l} \\right], \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 616, + 776, + 890, + 795 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $Q^{l}\\in \\mathbb{R}^{N\\times d_{\\mathrm{model}}}$ and $K^l\\in \\mathbb{R}^{N\\times d_{\\mathrm{model}}}$ refer to attention queries and keys of the $l$ -th layer, and $[\\cdot ;\\cdot ]$ denotes the pairwise concatenation. $W_{S}^{l}$ and $W_{O}^{l}$ are linear weights of shape $\\mathbb{R}^{d_{\\mathrm{model}}\\times d_{\\mathrm{model}}}$ .", + "bbox": [ + 496, + 806, + 890, + 868 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We also utilize the last layer representations of the object queries $Z^{L} \\in \\mathbb{R}^{N \\times d_{\\mathrm{model}}}$ , which are used for the object", + "bbox": [ + 496, + 869, + 890, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "24232", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "detection in the same manner:", + "bbox": [ + 76, + 90, + 279, + 104 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nR _ {z} = \\left[ Z ^ {L} W _ {S}; Z ^ {L} W _ {O} \\right], \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 191, + 116, + 468, + 133 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $W_{S}, W_{O}$ are linear weights of shape $\\mathbb{R}^{d_{\\mathrm{model}} \\times d_{\\mathrm{model}}}$ . To effectively use the various information learned in all layers, we introduce a gating mechanism as follows:", + "bbox": [ + 76, + 146, + 468, + 191 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\ng _ {a} ^ {l} = \\sigma \\left(R _ {a} ^ {l} W _ {G}\\right), g _ {z} = \\sigma \\left(R _ {z} W _ {G}\\right), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 156, + 200, + 468, + 220 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $g_{a}^{l}$ and $g_{z} \\in \\mathbb{R}^{N \\times N \\times 1}$ represent the gate values obtained through a single linear layer $W_{G} \\in \\mathbb{R}^{2d_{\\mathrm{model}} \\times 1}$ for $R_{a}^{l}$ and $R_{z}$ , respectively. Finally, we extract the relation graph from the gated summation of the relation representations across all layers as follows:", + "bbox": [ + 76, + 229, + 470, + 308 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {G} = \\sigma \\left(\\mathrm {M L P} _ {\\mathrm {r e l}} \\left(\\Sigma_ {l = 1} ^ {L} \\left(g _ {a} ^ {l} * R _ {a} ^ {l}\\right) + g _ {z} * R _ {z}\\right)\\right), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 125, + 332, + 468, + 352 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\hat{G} \\in \\mathbb{R}^{N \\times N \\times |\\mathcal{C}_p|}$ denotes predicted relation graph and $\\mathrm{MLP}_{\\mathrm{rel}}$ is a three-layer perceptron with ReLU activation. Note that we use sigmoid function $\\sigma$ so that multiple relationships can exist between objects.", + "bbox": [ + 76, + 363, + 468, + 425 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Learning and Inference", + "text_level": 1, + "bbox": [ + 76, + 434, + 294, + 450 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To train EGTR, we perform multi-task learning. In addition to object detection and relation extraction, we devise connectivity prediction, an auxiliary task for relation extraction. The overall loss for the framework is as follows:", + "bbox": [ + 76, + 457, + 468, + 516 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\mathcal {L} _ {\\mathrm {o d}} + \\lambda_ {\\mathrm {r e l}} \\mathcal {L} _ {\\mathrm {r e l}} + \\lambda_ {\\mathrm {c o n}} \\mathcal {L} _ {\\mathrm {c o n}}, \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 168, + 530, + 468, + 546 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathcal{L}_{\\mathrm{od}}$ is object detection loss in Eq. (2). $\\mathcal{L}_{\\mathrm{rel}}$ and $\\mathcal{L}_{\\mathrm{con}}$ are loss functions for the relation extraction and connectivity prediction, respectively. Since explicit object detection loss is used, EGTR has the capability to detect all nodes in the scene graph. The details of the loss for each task are provided below.", + "bbox": [ + 76, + 556, + 468, + 647 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3.1 Relation Extraction", + "text_level": 1, + "bbox": [ + 76, + 665, + 269, + 679 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We use binary cross-entropy loss for the relation extraction. To match predicted graph $\\hat{G} \\in \\mathbb{R}^{N \\times N \\times |\\mathcal{C}_p|}$ and ground truth triplet set $\\mathcal{E}$ , we first encode $\\mathcal{E}$ as a one-hot ground truth graph $G \\in \\mathbb{R}^{N \\times N \\times |\\mathcal{C}_p|}$ by padding the regions that do not correspond to relations between the ground truth objects as zero. Then, we permute indices of the predicted graph using the permutation found in the object detection. From the permuted graph $\\hat{G}'$ , the loss for the relation extraction is calculated as $\\mathcal{L}_{\\mathrm{rel}} = \\mathcal{L}_r(\\hat{G}', G)$ .", + "bbox": [ + 76, + 689, + 468, + 824 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "However, since the size of the graph is proportional to the square of $N$ , the sparsity of the ground truth graph $G$ is too severe. For instance, the density is only $10^{-14}$ when $N$ is set to 200 for the Visual Genome [12] validation dataset. Therefore, we divide $G$ into three regions: (1) GT region,", + "bbox": [ + 76, + 825, + 468, + 901 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/8b4b9a348b44f666252103d9ddd29526093a28ee47d8308828587aeab359d27e.jpg", + "image_caption": [], + "image_footnote": [ + "girl - on - horse", + "jirl - riding" + ], + "bbox": [ + 519, + 94, + 676, + 220 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/150501b135af7d479b690a4d0f111b17fd90658d9f4ed8d284146a7bbe9e14b7.jpg", + "image_caption": [ + "(subject - predicate - object)", + "non-matching region", + "Figure 4. Example of graph $G$ region. The target graph has a shape of $N \\times N \\times |\\mathcal{V}_c|$ . The graph is severely sparse since the number of object queries $N$ is set large enough to cover objects in the image. We split $G$ into GT, negative, and non-matching regions." + ], + "image_footnote": [ + "girl - wearing - pant" + ], + "bbox": [ + 676, + 92, + 877, + 257 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "(2) negative region, and (3) non-matching region, as shown in Fig. 4. The GT region indicates the ground truth triplets where the value of $G$ is 1. The negative region consists of the triplets in which both subjects and objects are composed of the ground truth objects, but no relation exists between them. The non-matching region represents the zero-padded region. It is paired with a region consisting of the object candidates that do not match the ground truth objects and match $\\phi$ in Eq. (2). For each region, we apply different techniques to effectively train the relation extraction with object detection as subsequently described.", + "bbox": [ + 498, + 364, + 892, + 532 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Adaptive Smoothing. We propose a novel adaptive smoothing for the GT region. For $G_{ijk}$ belonging to the GT region, the model is trained to predict the $k$ -th predicate category between subject entity $v_i$ and object entity $v_j$ . However, since $\\hat{v}_i'$ and $\\hat{v}_j'$ , which match $v_i$ and $v_j$ respectively, do not have enough representations about the corresponding ground truth object at the beginning of the training, it may not be appropriate to predict the probability of the predicate as 1. Moreover, even when the object detection performance is reasonably assured, the detection performance may still vary for individual object candidates. Therefore, we reflect the detection performance of each object candidate on the relation label via adaptive smoothing.", + "bbox": [ + 498, + 537, + 890, + 733 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We first measure the uncertainty of each object candidate with the corresponding bipartite matching cost. For object candidate $\\hat{v}_i'$ , we define the uncertainty as follows:", + "bbox": [ + 498, + 733, + 890, + 777 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nu _ {i} = \\sigma \\left(\\operatorname {c o s t} _ {i} - \\operatorname {c o s t} _ {\\min } + \\sigma^ {- 1} (\\alpha)\\right), \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 576, + 785, + 890, + 803 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathrm{cost}_i$ denotes the matching cost and $\\mathrm{cost}_{\\mathrm{min}}$ indicates the matching cost when $\\hat{v}_i^\\prime$ perfectly matches $v_{i}$ . $\\alpha$ is a nonnegative hyperparameter representing the minimum uncertainty. We set the value of $G_{ijk}$ to $(1 - u_{i})(1 - u_{j})$ taking the uncertainty into account. By using the relation label adjusted by the uncertainty, the multi-task learning of object", + "bbox": [ + 498, + 810, + 892, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "24233", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "detection and relation extraction is dynamically regulated according to the quality of the detected objects.", + "bbox": [ + 76, + 90, + 468, + 121 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Negative and Non-matching Sampling. Rather than employing all negatives, we sample them from the negative region. Inspired by the hard negative mining introduced by Liu et al. [21], we sort all negatives based on the predicted relation score $\\hat{G}_{ijk}^{\\prime}$ and choose the top $k_{\\mathrm{neg}} \\times |\\mathcal{E}|$ most challenging negatives. Similarly, we extract $k_{\\mathrm{non}} \\times |\\mathcal{E}|$ hard samples from the non-matching region. As the non-matching region typically encompasses a substantial part of the graph $G$ , this approach notably diminishes sparsity.", + "bbox": [ + 76, + 126, + 467, + 262 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.3.2 Connectivity Prediction", + "text_level": 1, + "bbox": [ + 76, + 281, + 294, + 295 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Influenced by Graph-RCNN [40] that predicted a relatedness to prune object pairs, we propose a connectivity prediction that predicts whether at least one edge between two object nodes exists for the relation extraction. We get a connectivity graph $\\hat{E} \\in \\mathbb{R}^{N \\times N \\times 1}$ in a similar way to get the relation graph in Eq. (6) with the same relation source representations. Instead of $\\mathrm{MLP}_{\\mathrm{rel}}$ for multi-label prediction, we use another $\\mathrm{MLP}_{\\mathrm{con}}$ for binary classification. We calculate the binary cross entropy loss from the permutated connectivity graph as follows: $\\mathcal{L}_{\\mathrm{con}} = \\mathcal{L}_r(\\hat{E}', E)$ .", + "bbox": [ + 75, + 305, + 467, + 455 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.3.3 Inference", + "text_level": 1, + "bbox": [ + 76, + 474, + 196, + 488 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For model inference, we get triplet scores by multiplying predicate score $\\hat{G}_{ijk}$ by the corresponding class scores of $\\hat{v}_i^c$ and $\\hat{v}_j^c$ . Note that we set $\\hat{G}_{iik}$ to 0 to prevent self-connections in which the subject and object are the same entity. In addition, we enhance the triplet scores by utilizing the connectivity score $\\hat{E}_{ijk}$ . By multiplying the connectivity score, we can effectively filter out the triplets that are not likely to have a relation between the subject and object.", + "bbox": [ + 76, + 500, + 467, + 621 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 76, + 636, + 207, + 652 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Datasets and Evaluation Settings", + "text_level": 1, + "bbox": [ + 76, + 660, + 364, + 676 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We conduct experiments on two SGG datasets. We describe the datasets and evaluation settings for each dataset. Detailed settings for each dataset are presented in the supplementary materials.", + "bbox": [ + 76, + 684, + 467, + 744 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Visual Genome. Visual Genome [12] is the most representative SGG dataset, consisting of 57K training images, 5K validation images, and 26K test images. We follow the popular Visual Genome split, retaining the most frequent 150 object categories and 50 relation categories. We adopt Scene Graph Detection (SGDet) evaluation settings and report Recall@k (R@k) which is class agnostic and mean Recall@k (mR@k) that aggregates the recalls for each predicate category. Following Motifs [42], these metrics are measures with graph constraint, which means each", + "bbox": [ + 75, + 750, + 467, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "object pair can have a single predicate category. Since they are only related to triplet detection of the scene graph, we report AP50 that evaluates the detection performance of all objects appearing in the scene. Furthermore, we report the number of model parameters and Frames Per Second (FPS) to measure efficiency.", + "bbox": [ + 496, + 90, + 890, + 181 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Open Image V6. Open Image V6 [13] is also widely used dataset, comprising of 126K images for training, 2K for validation, and 5K for test set. It includes 601 object categories and 30 relation categories. For evaluation, we adopt both recall and weighted mean AP (wmAP) following standard settings. For recall evaluation, micro-R@50 is adopted. The WMAP is evaluated with two settings: $\\mathrm{wmAP}_{\\mathrm{rel}}$ for predicting boxes of subject entity and object entity separately and $\\mathrm{wmAP}_{\\mathrm{phr}}$ for predicting a union box of them. The final score is calculated by $0.2 \\times \\text{micro-R} @50 + 0.4 \\times \\mathrm{wmAP}_{\\mathrm{rel}} + 0.4 \\times \\mathrm{wmAP}_{\\mathrm{phr}}$ .", + "bbox": [ + 496, + 186, + 890, + 354 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Implementation Details", + "text_level": 1, + "bbox": [ + 500, + 366, + 715, + 381 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We employ Deformable DETR [46] that improves the convergence speed of the DETR with ResNet-50 [6] as a backbone. It is worth noting that our approach can be extended to any object detector that incorporates self-attention mechanisms between object features, including models like DETR [1], Sparse R-CNN [32], and others. We follow the configurations of the original Deformable DETR, except that we use only 200 object queries. To speed up convergence, we first train the object detector with the target dataset and subsequently train the SGG task using the pretrained object detector, similar to prior work [17]. To calculate the overall loss in Equation 7, $\\lambda_{\\mathrm{rel}}$ is set to 15. $\\lambda_{\\mathrm{con}}$ is set to 30 and 90 for the Visual Genome and Open Image V6, respectively. We set $\\alpha$ as $10^{-14}$ for adaptive smoothing and both $k_{\\mathrm{neg}}$ and $k_{\\mathrm{non}}$ are set to 80.", + "bbox": [ + 496, + 388, + 890, + 617 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Results", + "text_level": 1, + "bbox": [ + 500, + 628, + 591, + 642 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We present quantitative results and perform a comparative analysis of our proposed framework with representative SGG models. Additional experimental results are included in the supplementary materials.", + "bbox": [ + 496, + 652, + 890, + 713 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Visual Genome. Visual Genome results are shown in Tab. 1. Our proposed method demonstrates competitive performance with the current one-stage SGG models that have 1.5 to 6.5 times larger parameters with the fastest inference speed. In particular, our method achieves the highest object detection performance and shows comparable performance in the triplet detection compared to SSR-CNN [35], the state-of-the-art method. The results demonstrate that our method can generate scene graphs in an efficient and effective way by exploiting the by-products of the object detector. By applying logit adjustment [35], a technique for predicting tail predicate classes well, our", + "bbox": [ + 496, + 719, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "24234", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/a16fe17cf96c9225171bc6281cd360eb2c9e469b302579803d440aef71ae7091.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Model# params (M)FPSAP50R@20R@50R@100mR@20mR@50mR@100
two-stageIMP (EBM) [31, 39]322.22.028.118.125.931.22.84.25.4
VTransE [44]312.33.5-24.531.335.55.16.88.0
Motifs [42]369.91.928.125.132.136.94.15.56.8
VCTree [33]361.50.828.124.831.836.14.96.67.7
VCTree (TDE) [33, 34]361.30.828.114.019.423.26.99.311.1
VCTree (EBM) [31, 33]372.5-28.124.231.435.95.77.79.1
GPS-Net [19]----31.135.9-6.78.6
BGNN [16]341.91.729.023.331.035.87.510.712.6
one-stageFCSGG [20]87.16.028.516.121.325.12.73.64.2
RelTR [7]63.713.426.421.227.5-6.810.8-
SGTR [17]117.16.225.4-24.628.4-12.015.2
Relationformer [29]92.98.526.322.228.431.34.69.310.7
Iterative SGG [9]93.56.027.7†-29.732.1-8.08.8
SSR-CNN [35]274.34.023.8†25.832.736.96.18.410.0
SSR-CNN [35] LA,τ=0.3274.34.023.8†18.423.326.513.517.921.4
EGTR (Ours)42.514.730.823.530.234.35.57.910.1
EGTR (Ours) LA,τ=0.742.514.730.815.718.720.512.117.821.7
EGTR (Ours) LA,τ=0.542.514.730.819.724.226.711.017.121.4
EGTR (Ours) LA,τ=0.342.514.730.822.428.231.78.814.018.3
", + "bbox": [ + 76, + 88, + 893, + 356 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/7064226fa7dbe87e3156582cfc815f2a9dc9c4a9cd4fd7effeb5b3b0549607d9.jpg", + "table_caption": [ + "Table 1. Graph-Constraint results on the test set of Visual Genome dataset. We report the results of representative two-stage SGG models based on the Faster R-CNN [27] object detector with a ResNeXt-101-FPN [18, 38] backbone and concurrent one-stage SGG models. Among the one-stage models, the best is highlighted in bold, and the second-best is indicated with underlining. LA denotes logit adjustment proposed in SSR-CNN [35]. $\\dagger$ represents that we concatenate the predicted subjects and objects and then apply non-maximum-suppression with a threshold of 0.3. Italic indicates the evaluation of metrics using publicly available model checkpoints. FPS is measured with a single V100 for images resized to a minimum of 600 for the shortest side and a maximum of 1000 for the longest side." + ], + "table_footnote": [], + "table_body": "
Modelscoremicro-R@50wmAPrelwmAPphr
Motifs [42]38.971.629.931.6
VCTree [33]40.274.134.233.1
GPS-Net [19]41.774.832.934.0
BGNN [16]42.175.033.534.2
RelTR [7]43.071.734.237.5
SGTR [17]42.359.937.038.7
SSR-CNN [35]49.476.741.543.6
EGTR (Ours)48.675.042.041.9
", + "bbox": [ + 76, + 457, + 468, + 595 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 2. Results on test set of Open Image V6. The score is a weighted sum of micro-R@50, $\\mathrm{wmAP}_{\\mathrm{rel}}$ , and $\\mathrm{wmAP}_{\\mathrm{phr}}$ .", + "bbox": [ + 76, + 602, + 468, + 632 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "method performs favorably in the trade-off between $\\mathrm{R}@\\mathbf{k}$ and $\\mathrm{mR}@\\mathbf{k}$ , showing superiority over existing state-of-the-art models. Note that for triplet detection models [9, 35] without an explicit object detector, AP50 is measured by applying non-maximum-suppression (NMS) to the union of predicted subjects and objects set. AP50 performance is discussed further in the section 4.5.", + "bbox": [ + 75, + 643, + 468, + 750 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Open Image V6. As depicted in Tab. 2, the experiments conducted on the Open Image V6 dataset also demonstrate competitive performance, underscoring the effectiveness and robustness of our method across different datasets.", + "bbox": [ + 75, + 753, + 468, + 816 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Ablation Studies", + "text_level": 1, + "bbox": [ + 76, + 825, + 241, + 842 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We analyze the effects of our model components. All of the ablation studies are done with the Visual Genome dataset.", + "bbox": [ + 76, + 849, + 468, + 878 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Relation Sources. In Tab. 3, we investigate the source of", + "bbox": [ + 76, + 885, + 468, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "relations used for the relation extractor. To verify the benefits of using attention queries and keys, we change $Q^l$ and $K^l$ in Eq. (3) to $Z^l$ , the hidden states of each layer. Remarkably, when simply using the hidden states of all layers, the triplet detection performance decreases. Surprisingly, the performance is lower than that of using only $R_z$ from the last layer. On the other hand, when only attention by-products are used as a relation source, the performance is similar to that of using only hidden states of the last layer, the most contextualized representations for detecting objects. Incorporating both attention by-products and final hidden states yields improved performance by amalgamating varied information. The results underscore that the attention by-products of the object detectors contain rich information for relation extraction.", + "bbox": [ + 496, + 460, + 890, + 686 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Training Techniques. In Tab. 4, we ablate proposed techniques. Without any techniques, our model framework shows low performance. When each technique is applied, it substantially improves both R@50 and mR@50, and the best performance is achieved when all techniques are combined. The results demonstrate that our proposed techniques help to train the model framework effectively. Additional experimental results are provided in the supplementary materials.", + "bbox": [ + 496, + 695, + 890, + 832 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Sampling Methodology. In Tab. 5, we investigate the impact of hard sampling methods for the negative and non-matching regions. Hard negatives and hard non-matchings show a trade-off that improves the performance of R@50", + "bbox": [ + 496, + 839, + 893, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "24235", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/130b3fcdca977aff0793d48b4bd73ce8c49708db37c5250079679f4f3392cdfc.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Rl^l a sourceRl^l aRzR@50mR@50
Ql & Kl✓✓30.27.9
Zl✓✓29.67.4
-✓29.97.6
Ql & Kl✓29.87.7
", + "bbox": [ + 78, + 88, + 470, + 167 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/9da1d5b0b3751bc9139a7ea1b3f68db2ac24a2b648ee5e40c17b42a22563ab5c.jpg", + "table_caption": [ + "Table 3. Ablation study on relation source." + ], + "table_footnote": [], + "table_body": "
adaptive smoothing\\( \\mathcal{L}_{\\text{con}} \\)samplingR@50mR@50
26.65.3
✓28.36.5
✓29.67.0
✓28.97.1
✓✓✓30.27.9
", + "bbox": [ + 78, + 203, + 468, + 292 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/ca8a4a2118b2bf6c879fa9b0eeed7043ab378d3428d9275d525d53208d7986e0.jpg", + "table_caption": [ + "Table 4. Ablation study on proposed techniques." + ], + "table_footnote": [], + "table_body": "
hard negativehard non-matchingR@50mR@50
✓✓30.27.9
✓30.07.1
✓29.67.7
", + "bbox": [ + 78, + 328, + 468, + 393 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "and mR@50, respectively, and additional performance improvements are achieved for both measures when they are used together. For hard negatives, we speculate that R@50 performance is improved by selecting negatives for the tail predicate class, which is likely to exist in reality but is not in the annotation. For hard non-matchings, mR@50 performance seems to have improved by choosing object candidates that are akin to the ground truth objects but unmatched with them. This encourages the model to predict diverse predicate classes by preventing duplicate object pairs, which are likely to predict head predicate classes, from being predicted.", + "bbox": [ + 75, + 428, + 468, + 609 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.5. Discussion", + "text_level": 1, + "bbox": [ + 76, + 619, + 194, + 633 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Regarding the highest AP50 score, we posit that it is because our proposed method focuses on not only objects with relations but also objects without any relations. To validate this, we calculate AP50 for two subsets of the ground-truth objects: those with relations and those without relations. We compare our method with two triplet detection-based models. Tab. 6 illustrates that our method achieves significantly high $\\mathrm{AP50}_{\\mathrm{no - rel}}$ . Considering that triplet detection performance is related to $\\mathrm{AP50}_{\\mathrm{rel}}$ , while triplet detection-based models only focus on objects with relations to detect triplets well, our method is capable of extracting the complete scene graph, including objects without relations.", + "bbox": [ + 75, + 643, + 468, + 824 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "For further analysis, we visualize bounding boxes for subjects and objects from the top 20 predictions in Fig. 5. For SSR-CNN [35], which directly predicts triplets, we observe a significant number of overlapping bounding boxes. On the other hand, our proposed model leverages predic", + "bbox": [ + 75, + 824, + 468, + 901 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/e58fc5b81d9148d9476052a5decff82b2473d1214d9d6f90d01e0685faefade3.jpg", + "table_caption": [ + "Table 5. Ablation study on sampling methods." + ], + "table_footnote": [], + "table_body": "
ModelAP50\\( AP50_{\\text{rel}} \\)\\( AP50_{\\text{no-rel}} \\)
Iterative SGG [9]†27.724.37.8
SSR-CNN [35]†23.820.27.4
EGTR (Ours)30.824.310.7
", + "bbox": [ + 500, + 88, + 890, + 152 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 6. AP50 for two subsets of objects. $\\mathrm{AP50}_{\\mathrm{rel}}$ and $\\mathrm{AP50}_{\\mathrm{no - rel}}$ denote scores evaluated only on objects having at least one edge and no edge, respectively. Note that each measure operates on its unique scale since the ground-truth objects used for each measure differ. $\\dagger$ indicates that additional NMS is applied to the union of subjects and objects set.", + "bbox": [ + 498, + 160, + 893, + 244 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/24bd9ff20bfc213d9d6ec0c5211af0c429e35efa9df27fb0358e28fccd39745f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 501, + 257, + 694, + 358 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/18ae33dd39cb2090a7c41cafda42b6943398cbcb2bce6166315195b836fe3f0f.jpg", + "image_caption": [ + "(a) SSR-CNN [35]" + ], + "image_footnote": [], + "bbox": [ + 501, + 358, + 694, + 455 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/46306f91c2dbb914d9f6359cae5848fe4882b7587c39fc49c4b75c459bcd9dfb.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 694, + 258, + 885, + 358 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/bcd320527c7035d3e2f40585be25959ccfdf58231515632ecd35636a8f686e10.jpg", + "image_caption": [ + "(b) EGTR (Ours)", + "Figure 5. Comparison of detected subjects and objects. We select the top 20 predictions of SSR-CNN and EGTR and visualize bounding boxes of detected subjects and objects." + ], + "image_footnote": [], + "bbox": [ + 694, + 358, + 885, + 455 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "tions from an explicit object detector, reducing redundant predictions. Furthermore, our method detects various subjects and objects that appear in the scene rather than focusing on objects that are more likely to have relations.", + "bbox": [ + 498, + 532, + 890, + 592 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 606, + 617, + 621 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this study, we propose the lightweight one-stage scene graph generator EGTR. We significantly reduce the model complexity by harnessing the relationships among the object queries learned from the self-attention layers of the object decoder. Furthermore, we devise a novel adaptive smoothing technique that helps multi-task learning of object detection and relation extraction by adjusting the relation label according to the object detection performance. As an auxiliary task of relation extraction, connectivity prediction contributes to the effective learning of EGTR. We conduct extensive experiments and ablation studies, and the results demonstrate that EGTR achieves the highest object detection performance and competitive triplet detection capabilities with the fastest inference speed.", + "bbox": [ + 496, + 631, + 890, + 843 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements. We thank Taeho Kil, Geewook Kim, and the anonymous reviewers for their insightful comments and suggestions.", + "bbox": [ + 498, + 847, + 890, + 893 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "24236", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In ECCV, pages 213-229, 2020. 1, 2, 3, 4, 6", + "[2] Zhe Chen, Shaoli Huang, and Dacheng Tao. Context refinement for object detection. In ECCV, pages 71-86, 2018. 2", + "[3] Yuren Cong, Michael Ying Yang, and Bodo Rosenhahn. Reltr: Relation transformer for scene graph generation. IEEE TPAMI, 2023. 1, 3", + "[4] Naina Dhingra, Florian Ritter, and Andreas Kunz. Bgt-net: Bidirectional gru transformer network for scene graph generation. In CVPR, pages 2150-2159, 2021. 2", + "[5] Lizhao Gao, Bo Wang, and Wenmin Wang. Image captioning with scene-graph based semantic concepts. In ICMLC, pages 225-229, 2018. 1", + "[6] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, pages 770-778, 2016. 6", + "[7] Han Hu, Jiayuan Gu, Zheng Zhang, Jifeng Dai, and Yichen Wei. Relation networks for object detection. In CVPR, pages 3588-3597, 2018. 2, 7", + "[8] Justin Johnson, Ranjay Krishna, Michael Stark, Li-Jia Li, David Shamma, Michael Bernstein, and Li Fei-Fei. Image retrieval using scene graphs. In CVPR, pages 3668-3678, 2015. 1", + "[9] Siddhesh Khandelwal and Leonid Sigal. Iterative scene graph generation. In NeurIPS, 2022. 1, 3, 7, 8", + "[10] Dong-Jin Kim, Jinsoo Choi, Tae-Hyun Oh, and In So Kweon. Dense relational captioning: Triple-stream networks for relationship-based captioning. In CVPR, pages 6271-6280, 2019. 1", + "[11] Rajat Koner, Suprosanna Shit, and Volker Tresp. Relation transformer network. ECCV, pages 422-439, 2022. 2", + "[12] Ranjay Krishna, Yuke Zhu, Oliver Groth, Justin Johnson, Kenji Hata, Joshua Kravitz, Stephanie Chen, Yannis Kalantidis, Li-Jia Li, David A Shamma, et al. Visual genome: Connecting language and vision using crowdsourced dense image annotations. IJCV, 123:32-73, 2017. 2, 5, 6", + "[13] Alina Kuznetsova, Hassan Rom, Neil Alldrin, Jasper Uijlings, Ivan Krasin, Jordi Pont-Tuset, Shahab Kamali, Stefan Popov, Matteo Malloci, Alexander Kolesnikov, et al. The open images dataset v4: Unified image classification, object detection, and visual relationship detection at scale. IJCV, 128(7):1956-1981, 2020. 2, 6", + "[14] Linjie Li, Zhe Gan, Yu Cheng, and Jingjing Liu. Relation-aware graph attention network for visual question answering. In ICCV, pages 10313-10322, 2019. 1", + "[15] Lin Li, Guikun Chen, Jun Xiao, Yi Yang, Chunping Wang, and Long Chen. Compositional feature augmentation for unbiased scene graph generation. In ICCV, pages 21685-21695, 2023. 2", + "[16] Rongjie Li, Songyang Zhang, Bo Wan, and Xuming He. Bipartite graph network with adaptive message passing for unbiased scene graph generation. In CVPR, pages 11109-11119, 2021. 1, 2, 7" + ], + "bbox": [ + 78, + 114, + 470, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[17] Rongjie Li, Songyang Zhang, and Xuming He. Sgtr: End-to-end scene graph generation with transformer. In CVPR, pages 19486-19496, 2022. 1, 3, 6, 7", + "[18] Tsung-Yi Lin, Piotr Dollar, Ross Girshick, Kaiming He, Bharath Hariharan, and Serge Belongie. Feature pyramid networks for object detection. In CVPR, pages 2117-2125, 2017. 7", + "[19] Xin Lin, Changxing Ding, Jinquan Zeng, and Dacheng Tao. Gps-net: Graph property sensing network for scene graph generation. In CVPR, pages 3746-3753, 2020. 1, 2, 7", + "[20] Hengyue Liu, Ning Yan, Masood Mortazavi, and Bir Bhanu. Fully convolutional scene graph generation. In CVPR, pages 11546-11556, 2021. 1, 3, 7", + "[21] Wei Liu, Dragomir Anguelov, Dumitru Erhan, Christian Szegedy, Scott Reed, Cheng-Yang Fu, and Alexander C Berg. Ssd: Single shot multibox detector. In ECCV, pages 21-37, 2016. 6", + "[22] Yong Liu, Ruiping Wang, Shiguang Shan, and Xilin Chen. Structure inference net: Object detection using scene-level context and instance-level relationships. In CVPR, pages 6985-6994, 2018. 2", + "[23] Cewu Lu, Ranjay Krishna, Michael Bernstein, and Li Fei-Fei. Visual relationship detection with language priors. In ECCV, pages 852–869, 2016. 2", + "[24] Yichao Lu, Himanshu Rai, Jason Chang, Boris Knyazev, Guangwei Yu, Shashank Shekhar, Graham W Taylor, and Maksims Volkovs. Context-aware scene graph generation with seq2seq transformers. In ICCV, pages 15931-15941, 2021.", + "[25] Yukuan Min, Among Wu, and Cheng Deng. Environment-invariant curriculum relation learning for fine-grained scene graph generation. In ICCV, pages 13296-13307, 2023. 2", + "[26] Alejandro Newell and Jia Deng. Pixels to graphs by associative embedding. NeurIPS, 30, 2017. 1, 3", + "[27] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. NeurIPS, 28, 2015. 2, 7", + "[28] Sebastian Schuster, Ranjay Krishna, Angel Chang, Li Fei-Fei, and Christopher D Manning. Generating semantically precise scene graphs from textual descriptions for improved image retrieval. In Proceedings of the fourth workshop on vision and language, pages 70–80. Association for Computational Linguistics, 2015. 1", + "[29] Suprosanna Shit, Rajat Koner, Bastian Wittmann, Johannes Paetzold, Ivan Ezhov, Hongwei Li, Jiazhen Pan, Sahand Sharifzadeh, Georgios Kaissis, Volker Tresp, et al. Relationformer: A unified framework for image-to-graph generation. In ECCV, pages 422-439, 2022. 1, 3, 7", + "[30] Gopika Sudhakaran, Devendra Singh Dhami, Kristian Kersting, and Stefan Roth. Vision relation transformer for unbiased scene graph generation. In ICCV, pages 21882-21893, 2023. 2", + "[31] Mohammed Suhail, Abhay Mittal, Behjat Siddiquie, Chris Broaddus, Jayan Eledath, Gerard Medioni, and Leonid Sigal. Energy-based learning for scene graph generation. In CVPR, pages 13936-13945, 2021. 7" + ], + "bbox": [ + 501, + 92, + 890, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "24237", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[32] Peize Sun, Rufeng Zhang, Yi Jiang, Tao Kong, Chenfeng Xu, Wei Zhan, Masayoshi Tomizuka, Lei Li, Zehuan Yuan, Changhu Wang, and Ping Luo. Sparse r-cnn: End-to-end object detection with learnable proposals. In CVPR, pages 14454-14463, 2021. 2, 3, 6", + "[33] Kaihua Tang, Hanwang Zhang, Baoyuan Wu, Wenhan Luo, and Wei Liu. Learning to compose dynamic tree structures for visual contexts. In CVPR, pages 6619-6628, 2019. 1, 2, 7", + "[34] Kaihua Tang, Yulei Niu, Jianqiang Huang, Jiaxin Shi, and Hanwang Zhang. Unbiased scene graph generation from biased training. In CVPR, pages 3716-3725, 2020. 7", + "[35] Yao Teng and Limin Wang. Structured sparse r-cnn for direct scene graph generation. In CVPR, pages 19437-19446, 2022. 1, 3, 6, 7, 8", + "[36] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. NeurIPS, 30, 2017. 2, 3", + "[37] Sanghyun Woo, Dahun Kim, Donghyeon Cho, and In So Kweon. Linknet: Relational embedding for scene graph. NeurIPS, 31, 2018. 2", + "[38] Saining Xie, Ross Girshick, Piotr Dollar, Zhuowen Tu, and Kaiming He. Aggregated residual transformations for deep neural networks. In CVPR, pages 1492-1500, 2017. 7", + "[39] Danfei Xu, Yuke Zhu, Christopher B Choy, and Li Fei-Fei. Scene graph generation by iterative message passing. In CVPR, pages 5410–5419, 2017. 1, 2, 7", + "[40] Jianwei Yang, Jiasen Lu, Stefan Lee, Dhruv Batra, and Devi Parikh. Graph r-cnn for scene graph generation. In ECCV, pages 670-685, 2018. 6", + "[41] Xu Yang, Kaihua Tang, Hanwang Zhang, and Jianfei Cai. Auto-encoding scene graphs for image captioning. In CVPR, pages 10685–10694, 2019. 1", + "[42] Rowan Zellers, Mark Yatskar, Sam Thomson, and Yejin Choi. Neural motifs: Scene graph parsing with global context. In CVPR, pages 5831-5840, 2018. 1, 2, 6, 7", + "[43] Cheng Zhang, Wei-Lun Chao, and Dong Xuan. An empirical study on leveraging scene graphs for visual question answering. In BMVC, 2019. 1", + "[44] Hanwang Zhang, Zawlin Kyaw, Shih-Fu Chang, and Tat-Seng Chua. Visual translation embedding network for visual relation detection. In CVPR, page 3107-3115, 2017. 7", + "[45] Chaofan Zheng, Xinyu Lyu, Yuyu Guo, Pengpeng Zeng, Jingkuan Song, and Lianli Gao. Learning to generate scene graph from head to tail. In ICME, pages 1-6. IEEE, 2022. 2", + "[46] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable detr: Deformable transformers for end-to-end object detection. In ICLR, 2021. 2, 3, 6" + ], + "bbox": [ + 78, + 90, + 468, + 787 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "24238", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2024/EGTR_ Extracting Graph from Transformer for Scene Graph Generation/174c2199-8aa8-4e45-8d2e-205eb2e2e099_model.json b/2024/EGTR_ Extracting Graph from Transformer for Scene Graph Generation/174c2199-8aa8-4e45-8d2e-205eb2e2e099_model.json new file mode 100644 index 0000000000000000000000000000000000000000..793f3b340832ade5743a378ce531d786238dd824 --- /dev/null +++ b/2024/EGTR_ Extracting Graph from Transformer for Scene Graph Generation/174c2199-8aa8-4e45-8d2e-205eb2e2e099_model.json @@ -0,0 +1,2343 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.131, + 0.856, + 0.154 + ], + "angle": 0, + "content": "EGTR: Extracting Graph from Transformer for Scene Graph Generation" + }, + { + "type": "text", + "bbox": [ + 0.123, + 0.18, + 0.844, + 0.24 + ], + "angle": 0, + "content": "Jinbae Im1 JeongYeon Nam1 Nokyung Park1,2,3* Hyungmin Lee2 Seunghyun Park1 \n1NAVER Cloud AI 2NAVER 3Korea University \n{jinbae.im, jy.nam, nokyung.park99, hyungmin.lee, seung.park}@navercorp.com" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.269, + 0.314, + 0.285 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.302, + 0.474, + 0.726 + ], + "angle": 0, + "content": "Scene Graph Generation (SGG) is a challenging task of detecting objects and predicting relationships between objects. After DETR was developed, one-stage SGG models based on a one-stage object detector have been actively studied. However, complex modeling is used to predict the relationship between objects, and the inherent relationship between object queries learned in the multi-head self-attention of the object detector has been neglected. We propose a lightweight one-stage SGG model that extracts the relation graph from the various relationships learned in the multi-head self-attention layers of the DETR decoder. By fully utilizing the self-attention by-products, the relation graph can be extracted effectively with a shallow relation extraction head. Considering the dependency of the relation extraction task on the object detection task, we propose a novel relation smoothing technique that adjusts the relation label adaptively according to the quality of the detected objects. By the relation smoothing, the model is trained according to the continuous curriculum that focuses on object detection task at the beginning of training and performs multi-task learning as the object detection performance gradually improves. Furthermore, we propose a connectivity prediction task that predicts whether a relation exists between object pairs as an auxiliary task of the relation extraction. We demonstrate the effectiveness and efficiency of our method for the Visual Genome and Open Image V6 datasets. Our code is publicly available at https://github.com/naver-ai/egtr." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.757, + 0.21, + 0.772 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.783, + 0.47, + 0.875 + ], + "angle": 0, + "content": "Scene Graph Generation (SGG) [8] aims to generate a scene graph that represents objects as nodes and relationships between objects as edges from an image as shown in Fig. 1b. SGG is a challenging task as it is required to go beyond simply detecting objects but predicting the relationships between them based on a comprehensive understanding of the" + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.272, + 0.747, + 0.396 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.552, + 0.401, + 0.702, + 0.413 + ], + "angle": 0, + "content": "(a) Visual Genome example" + }, + { + "type": "image", + "bbox": [ + 0.765, + 0.269, + 0.884, + 0.324 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.782, + 0.326, + 0.871, + 0.338 + ], + "angle": 0, + "content": "(b) Scene Graph" + }, + { + "type": "image", + "bbox": [ + 0.768, + 0.344, + 0.885, + 0.398 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.774, + 0.402, + 0.88, + 0.414 + ], + "angle": 0, + "content": "(c) Attention Graph" + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.42, + 0.893, + 0.518 + ], + "angle": 0, + "content": "Figure 1. Motivation. SGG task aims to predict scene graph (Fig. 1b) with objects as nodes and relations as edges. We draw a plausible attention graph (Fig. 1c) simply by connecting objects with high attention weights to edges from the self-attention layers of the pre-trained DETR. It shows the potential for the self-attention from the object detector to contain rich information that aids in predicting the relations of the scene graph." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.536, + 0.893, + 0.612 + ], + "angle": 0, + "content": "scene. Since a scene graph provides structural information of the image, it can be used for various vision tasks that require a higher level of understanding and reasoning about images such as image captioning [5, 10, 41], image retrieval [8, 28], and visual question answering [14, 43]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.614, + 0.895, + 0.901 + ], + "angle": 0, + "content": "Most previous studies [16, 19, 33, 39, 42] took two-stage SGG approaches that detect objects first and then predict their relations. However, these approaches incurred high computational costs and risks of error propagation from the object detection. To address the drawbacks of the two-stage approaches, one-stage SGG models that perform object detection and relation prediction at once [3, 9, 17, 20, 26, 29, 35] have been studied recently, leveraging one-stage object detectors such as DETR [1]. Since edges in a scene graph can be represented as subject-predicate-object triplets, many studies embraced triplet-based approaches as shown in Figs. 2a and 2b. However, object-triplet detection models [3, 17] required sophisticated triplet detectors to obtain necessary information for triplet queries from the object detector. Moreover, triplet detection models [9, 35] lacked the ability to detect objects without relation such as \"tree\" in Fig. 1b, by focusing solely on the triplet detection without an object detector. Considering that objects without relations account for more than" + }, + { + "type": "page_footnote", + "bbox": [ + 0.095, + 0.887, + 0.444, + 0.9 + ], + "angle": 0, + "content": "* Most work was done during the internship at Naver Cloud AI." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "24229" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.109, + 0.09, + 0.351, + 0.207 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.133, + 0.211, + 0.325, + 0.223 + ], + "angle": 0, + "content": "(a) Object-Triplet Detection Models" + }, + { + "type": "image", + "bbox": [ + 0.387, + 0.091, + 0.572, + 0.206 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.407, + 0.211, + 0.56, + 0.223 + ], + "angle": 0, + "content": "(b) Triplet Detection Models" + }, + { + "type": "image", + "bbox": [ + 0.617, + 0.092, + 0.861, + 0.206 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.656, + 0.211, + 0.821, + 0.222 + ], + "angle": 0, + "content": "(c) Relation Extraction Models" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.228, + 0.894, + 0.299 + ], + "angle": 0, + "content": "Figure 2. Comparison with existing one-stage SGG models. (a) Object-Triplet Detection Models introduce additional triplet queries and a triplet detector to the object detector. The triplet detector requires additional modules to incorporate information from the object detector into the triplet queries. (b) Triplet Detection Models focus on detecting triplets directly without an object detector. Objects without relations may not be detected. (c) Relation Extraction Models extract relations from the object detector without a separate triplet detector. In particular, ours extracts relations more effectively by utilizing by-products from the self-attention of the object detector." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.31, + 0.468, + 0.34 + ], + "angle": 0, + "content": "\\(42\\%\\) in Visual Genome [12] data, their priority lies in detecting a sub-graph rather than the complete scene graph." + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.342, + 0.47, + 0.657 + ], + "angle": 0, + "content": "To address the shortcomings of the existing one-stage SGG models, we focus on the relationships between objects inherent in the object detector. As shown in Fig. 1a, objects are related to each other. For instance, when a horse appears in a scene, a person is likely to appear in the scene, and clothing such as hats, jackets, and pants often depicts the person's current situation. From this intuition, there has been a long belief [2, 7, 22] that modeling relationship or context between objects would be beneficial to object detection task. Accordingly, recent one-stage object detectors [1, 32, 46] have incorporated self-attention layers [36] to implicitly model the relationships among the object queries. We hypothesize that self-attention between object queries learned in a one-stage object detector might contain valuable information for predicting triplet outputs. In our preliminary investigation, we are able to extract a plausible attention graph by simply connecting two object queries with high attention weights from the pre-trained DETR, as shown in Fig. 1c. It shows the potential that the attention weights between object queries can be interpreted as relations between them." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.659, + 0.47, + 0.87 + ], + "angle": 0, + "content": "From the findings, we propose a lightweight one-stage scene graph generator EGTR, which stands for Extracting Graph from TRansformer. We design the model to comprehensively leverage the by-products of the object detector, eliminating the need for a separate triplet detector, as depicted in Fig. 2c. From the multi-head self-attention layers of the object detector, we regard an attention query and key, where their relations are learned in the attention weights, as the subject entity and object entity, respectively. Subsequently, we leverage a shallow classifier to predict relations between them. Due to the abundant information about the relationships among the objects present in the by-products derived from all layers of self-attention, we can effectively extract the scene graph." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.871, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Since the relation extraction task is dependent on the object detection task, we speculate that performing relation ex" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.31, + 0.892, + 0.491 + ], + "angle": 0, + "content": "traction without sufficiently learned representations of the object queries might be harmful. Therefore, we devise a novel adaptive smoothing technique that smoothes the value of the ground truth relation label based on the object detection performance. With the adaptive smoothing, the model is trained with a continuous curriculum that initially focuses on object detection and gradually performs multi-task learning. Furthermore, we propose a connectivity prediction task as an auxiliary task for relation extraction, aiming to predict the existence of any relationship between a subject entity and an object entity. This auxiliary task facilitates the acquisition of representations for the relation extraction." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.493, + 0.892, + 0.598 + ], + "angle": 0, + "content": "To verify the effectiveness of the proposed method, we conduct experiments on two representative SGG datasets: Visual Genome [12] and Open Images V6 [13]. By actively utilizing the by-products of the object detector, EGTR shows the best object detection performance and comparable triplet detection performance, with the fewest parameters and the fastest inference speed." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.6, + 0.882, + 0.613 + ], + "angle": 0, + "content": "Our main contributions can be summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.616, + 0.891, + 0.66 + ], + "angle": 0, + "content": "- We propose EGTR that generates scene graphs efficiently and effectively by utilizing the multi-head self-attention by-products derived from the object detector." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.661, + 0.892, + 0.72 + ], + "angle": 0, + "content": "- We present adaptive smoothing, enabling effective multitask learning for both object detection and relation extraction. In addition, the proposed connectivity prediction offers clues to the relation extraction." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.722, + 0.894, + 0.767 + ], + "angle": 0, + "content": "- Our comprehensive experiments show the superiority of the proposed model framework and the effectiveness of the devised training techniques." + }, + { + "type": "list", + "bbox": [ + 0.5, + 0.616, + 0.894, + 0.767 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.784, + 0.642, + 0.8 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.811, + 0.892, + 0.901 + ], + "angle": 0, + "content": "SGG models can be categorized into two groups: two-stage models and one-stage models. For two-stage models [4, 11, 15, 16, 19, 23-25, 30, 33, 37, 39, 42, 45], separate object detection model and relation prediction model are trained sequentially. They usually detect \\(N\\) objects from the off-the-shelf object detector such as Faster R-CNN [27]," + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "24230" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.471, + 0.198 + ], + "angle": 0, + "content": "and then all possible combinations of the detected objects are fed into the relation prediction model to predict the relations between each object pair. Although they showed high relation extraction performance, they have the inherent limitation that the object detector that helps with relation extraction is trained separately, resulting in a significant increase in model complexity." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.202, + 0.471, + 0.384 + ], + "angle": 0, + "content": "As for the one-stage models [3, 9, 17, 20, 26, 29, 35], the object detection and relation prediction are trained in an end-to-end manner. Early studies [20, 26] proposed fully convolutional SGG models and took pixel-based approaches. After DETR [1] brought a huge success as a Transformer [36]-based one-stage object detector, many one-stage SGG studies are based on one-stage object detectors [1, 32, 46]. They efficiently modeled SGG by introducing object queries or triplet queries. We categorize them into three distinct groups: (a) object-triplet detection models, (b) triplet prediction models, and (c) relation extraction models, as shown in Fig. 2." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.393, + 0.47, + 0.635 + ], + "angle": 0, + "content": "Object-Triplet Detection Models. Object-triplet detection models are characterized by introducing additional triplet queries and building a triplet predictor on top of the object detector as shown in Fig. 2a. RelTR [3] introduced paired subject queries and object queries, and SGTR [17] introduced compositional queries decoupled into subjects, objects, and predicates. As the introduced queries are initialized without prior cues from the object detector, the triplet predictor requires modules to incorporate information from the outputs of the object detector and modules for triplet queries to exchange information with each other. It led to the triplet predictor having an intricate structure. In contrast, we refrain from introducing additional queries and regard the attention queries and attention keys, whose relationships are learned in the object detector, as subject and object queries, respectively." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.644, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Triplet Detection Models. Triplet detection models directly detect triplets using triplet queries without an object detector, as shown in Fig. 2b. Iterative SGG [9] introduced subject, object, and predicate queries and modeled the conditional dependencies among them with separate subject, object, and predicate multi-layer Transformer decoders. Inspired by Sparse R-CNN [32], Structured Sparse R-CNN (SSR-CNN) [35] designed triplet queries consisting of subject box, object box, subject, object, and predicate queries. Since it was difficult to train the model only with sparse triplet annotations, SSR-CNN introduced somewhat intricate training details: it detected object pairs with a Siamese Sparse RCNN model and auxiliary queries and performed additional triplet matching using the detected object pairs as a pseudo label. Although they show excellent triplet detection performance, the model architecture became even more intricate to detect subjects and objects separately since" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.152 + ], + "angle": 0, + "content": "an explicit object detector is not used. Last but not least, the triplet prediction models focus on detecting a sub-graph composed only of objects with relations, overlooking objects in an image that lack explicit relations." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.157, + 0.892, + 0.37 + ], + "angle": 0, + "content": "Relation Extraction Models. Relation extraction models extract a scene graph using a lightweight relation predictor without separate triplet queries or triplet detector. Relationformer [29] added a special \"[rln]\" token to capture global information in conjunction with object queries. They concatenated the final hidden representations of the object query pairs and the relation token, followed by a shallow fully connected network for relation prediction. In this work, in conjunction with the final hidden representations, we use the inherent relationship information among the object queries learned in the multi-head self-attention layer of the object detector. Furthermore, we propose training techniques to boost multi-task learning, leading to significantly improved performance with architectural simplicity." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.383, + 0.593, + 0.399 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.408, + 0.642, + 0.423 + ], + "angle": 0, + "content": "3.1. Preliminaries" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.432, + 0.892, + 0.477 + ], + "angle": 0, + "content": "In this section, we first introduce the formulation of the SGG task and then provide a brief review of the one-stage object detector that serves as the basis for our study." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.496, + 0.643, + 0.511 + ], + "angle": 0, + "content": "3.1.1 Formulation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.52, + 0.892, + 0.687 + ], + "angle": 0, + "content": "Scene graph generation is a task of generating a scene graph \\(\\mathcal{G} = \\{\\mathcal{V},\\mathcal{E}\\}\\) from an image, where \\(\\mathcal{V}\\) denotes the node set consisting of objects and \\(\\mathcal{E}\\) denotes the edge set that represents the relation between objects. Each object entity \\(v_{i}\\in \\mathcal{V}\\) has an object category label \\(v_{i}^{c}\\) from a set of object categories \\(\\mathcal{C}_v\\) and box coordinates \\(v_{i}^{b}\\). Each relation \\(e_j\\in \\mathcal{E}\\) represents the \\(j\\)-th triplet \\((s_j,p_j,o_j)\\), where subject \\(s_j\\) and object \\(o_j\\) indicate related object entities and predicate \\(p_j\\) has a relation category label \\(p_j^c\\) from a set of predicate categories \\(\\mathcal{C}_p\\). Generating \\(\\mathcal{V}\\) and \\(\\mathcal{E}\\) correspond to object detection and relation extraction, respectively." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.706, + 0.741, + 0.722 + ], + "angle": 0, + "content": "3.1.2 One-stage Object Detector" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.729, + 0.892, + 0.837 + ], + "angle": 0, + "content": "Our model framework is deeply influenced by DETR [1], a one-stage object detector. In DETR, representations for input images are learned through CNN and Transformer [36] encoder. Transformer decoder enhances object queries by utilizing self- and cross-attention mechanisms. The final object detection heads predict object category labels and box coordinates from the contextualized object queries." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.892, + 0.902 + ], + "angle": 0, + "content": "Backbone. CNN backbone generates \\(C\\)-dimensional feature map \\(F \\in \\mathbb{R}^{C \\times H_F \\times W_F}\\) from an input image \\(x_{\\mathrm{img}} \\in \\mathbb{R}^{3 \\times H_0 \\times W_0}\\). For ResNet-50 backbone, \\(H_F\\) and \\(W_F\\) are set to \\(H_0 / 32\\) and \\(W_0 / 32\\), respectively." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "24231" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.08, + 0.089, + 0.705, + 0.281 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.71, + 0.094, + 0.887, + 0.184 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.715, + 0.189, + 0.887, + 0.28 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.299, + 0.895, + 0.356 + ], + "angle": 0, + "content": "Figure 3. The overall architecture of EGTR. We present a novel lightweight relation extractor, EGTR, which fully utilizes the self-attention of the DETR decoder. We extract query and key representations from each self-attention layer and concatenate them pairwise to represent relations between them. Additionally, we leverage the last hidden representation in the same manner. To effectively aggregate information, we apply a gated sum and then predict relation with a shallow relation head." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.368, + 0.471, + 0.475 + ], + "angle": 0, + "content": "Transformer Encoder. Transformer encoder enhances image representations based on multi-head self-attention. After applying a \\(1 \\times 1\\) convolution to reduce dimension from \\(C\\) to \\(d_{\\mathrm{model}}\\), the feature map is flattened to get an input sequence with length \\(H_{F}W_{F}\\) for the Transformer encoder. Additional positional encodings are used to reflect the spatial information of the feature maps." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.48, + 0.47, + 0.631 + ], + "angle": 0, + "content": "Transformer Decoder. Transformer decoder takes \\(N\\) object queries as input and returns their representations. Each object query detects an individual object, and \\(N\\) is usually set large enough to cover all objects in the image. Through alternating self-attention and cross-attention layers, object queries learn features of object candidates in the input image. Note that causal attention mask is not applied to the self-attention layer; therefore, fully connected \\(N \\times N\\) attention weights among the \\(N\\) object queries are learned from the multi-head self-attention as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.103, + 0.642, + 0.47, + 0.665 + ], + "angle": 0, + "content": "\\[\nA _ {h} ^ {l} = f \\left(Q _ {h} ^ {l}, K _ {h} ^ {l}\\right) = \\operatorname {s o f t m a x} \\left(Q _ {h} ^ {l} K _ {h} ^ {T} / \\sqrt {d _ {\\text {h e a d}}}\\right), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.675, + 0.471, + 0.723 + ], + "angle": 0, + "content": "where \\(A_h^l \\in \\mathbb{R}^{N \\times N}\\) denotes the attention weights from the \\(h\\)-th head in the \\(l\\)-th layer. \\(Q_h^l \\in \\mathbb{R}^{N \\times d_{\\mathrm{head}}}\\) and \\(K_h^l \\in \\mathbb{R}^{N \\times d_{\\mathrm{head}}}\\) are attention queries and keys, respectively." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.728, + 0.47, + 0.821 + ], + "angle": 0, + "content": "Object Detection Heads. Object detection heads detect \\(N\\) object candidates \\(\\{\\hat{v}_i\\}_{i = 1}^N\\) from the last layer representations of the object queries \\(Z^{L}\\). For each object query, a linear layer is used to predict the object category label \\(\\hat{v}_i^c\\in \\mathbb{R}^{N\\times |\\mathcal{C}_v|}\\), and three-layer perceptron with ReLU activation is used to obtain the box coordinates \\(\\hat{v}_i^b\\in \\mathbb{R}^{N\\times 4}\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Object Detection Loss. To match \\(N\\) detected object candidates \\(\\{\\hat{v}_i\\}_{i=1}^N\\) and \\(M\\) ground truth objects \\(\\{v_i\\}_{i=1}^M\\), Carion et al. [1] pad the ground truth objects with \\(\\phi\\) (no object) and find the best permutation of predicted objects that minimizes total bipartite matching costs. From the permuted" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.367, + 0.854, + 0.385 + ], + "angle": 0, + "content": "predictions \\(\\{\\hat{v}_i^{\\prime}\\}_{i = 1}^{N}\\) , the loss is calculated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.512, + 0.395, + 0.892, + 0.415 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {o d}} = \\Sigma_ {i = 1} ^ {N} \\left[ \\lambda_ {c} \\mathcal {L} _ {c} \\left(\\hat {v} _ {i} ^ {\\prime c}, v _ {i} ^ {c}\\right) + \\mathbb {1} _ {v _ {i} ^ {c} \\neq \\phi} \\left(\\lambda_ {b} \\mathcal {L} _ {b} \\left(\\hat {v} _ {i} ^ {\\prime b}, v _ {i} ^ {b}\\right)\\right) \\right], \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.426, + 0.892, + 0.457 + ], + "angle": 0, + "content": "where \\(\\mathcal{L}_c\\) and \\(\\mathcal{L}_b\\) denote loss function for category labels and box coordinates, respectively." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.468, + 0.589, + 0.483 + ], + "angle": 0, + "content": "3.2. EGTR" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.492, + 0.892, + 0.613 + ], + "angle": 0, + "content": "We propose a novel lightweight relation extractor, EGTR, which exploits the self-attention of DETR decoder, as depicted in Fig. 3. Since the self-attention weights in Eq. (1) contain \\( N \\times N \\) bidirectional relationships among the \\( N \\) object queries, our relation extractor aims to extract the predicate information from the self-attention weights in the entire \\( L \\) layers, by considering the attention queries and keys as subjects and objects, respectively." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.614, + 0.893, + 0.765 + ], + "angle": 0, + "content": "In order to preserve rich information learned in the self-attention layer, we devise another relation function \\( f \\) between the attention queries and keys instead of the dot product attention in Eq. (1). As concatenation preserves representations completely, we concatenate the representations on all \\( N \\times N \\) pairs of attention queries and keys as shown in Fig. 3 (a) to get the relation representations of the \\( l \\)-th layer \\( R_{a}^{l} \\in \\mathbb{R}^{N \\times N \\times 2d_{\\mathrm{model}}} \\). Before pairwise concatenation, to help queries and keys play the role of subjects and objects, a linear projection is added as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.617, + 0.777, + 0.892, + 0.796 + ], + "angle": 0, + "content": "\\[\nR _ {a} ^ {l} = \\left[ Q ^ {l} W _ {S} ^ {l}; K ^ {l} W _ {O} ^ {l} \\right], \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.808, + 0.892, + 0.869 + ], + "angle": 0, + "content": "where \\(Q^{l}\\in \\mathbb{R}^{N\\times d_{\\mathrm{model}}}\\) and \\(K^l\\in \\mathbb{R}^{N\\times d_{\\mathrm{model}}}\\) refer to attention queries and keys of the \\(l\\) -th layer, and \\([\\cdot ;\\cdot ]\\) denotes the pairwise concatenation. \\(W_{S}^{l}\\) and \\(W_{O}^{l}\\) are linear weights of shape \\(\\mathbb{R}^{d_{\\mathrm{model}}\\times d_{\\mathrm{model}}}\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.871, + 0.892, + 0.902 + ], + "angle": 0, + "content": "We also utilize the last layer representations of the object queries \\( Z^{L} \\in \\mathbb{R}^{N \\times d_{\\mathrm{model}}} \\), which are used for the object" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "24232" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.078, + 0.092, + 0.28, + 0.106 + ], + "angle": 0, + "content": "detection in the same manner:" + }, + { + "type": "equation", + "bbox": [ + 0.192, + 0.117, + 0.47, + 0.135 + ], + "angle": 0, + "content": "\\[\nR _ {z} = \\left[ Z ^ {L} W _ {S}; Z ^ {L} W _ {O} \\right], \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.147, + 0.47, + 0.193 + ], + "angle": 0, + "content": "where \\( W_{S}, W_{O} \\) are linear weights of shape \\( \\mathbb{R}^{d_{\\mathrm{model}} \\times d_{\\mathrm{model}}} \\). To effectively use the various information learned in all layers, we introduce a gating mechanism as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.158, + 0.202, + 0.469, + 0.222 + ], + "angle": 0, + "content": "\\[\ng _ {a} ^ {l} = \\sigma \\left(R _ {a} ^ {l} W _ {G}\\right), g _ {z} = \\sigma \\left(R _ {z} W _ {G}\\right), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.231, + 0.471, + 0.309 + ], + "angle": 0, + "content": "where \\( g_{a}^{l} \\) and \\( g_{z} \\in \\mathbb{R}^{N \\times N \\times 1} \\) represent the gate values obtained through a single linear layer \\( W_{G} \\in \\mathbb{R}^{2d_{\\mathrm{model}} \\times 1} \\) for \\( R_{a}^{l} \\) and \\( R_{z} \\), respectively. Finally, we extract the relation graph from the gated summation of the relation representations across all layers as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.126, + 0.333, + 0.47, + 0.353 + ], + "angle": 0, + "content": "\\[\n\\hat {G} = \\sigma \\left(\\mathrm {M L P} _ {\\mathrm {r e l}} \\left(\\Sigma_ {l = 1} ^ {L} \\left(g _ {a} ^ {l} * R _ {a} ^ {l}\\right) + g _ {z} * R _ {z}\\right)\\right), \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.364, + 0.469, + 0.426 + ], + "angle": 0, + "content": "where \\(\\hat{G} \\in \\mathbb{R}^{N \\times N \\times |\\mathcal{C}_p|}\\) denotes predicted relation graph and \\(\\mathrm{MLP}_{\\mathrm{rel}}\\) is a three-layer perceptron with ReLU activation. Note that we use sigmoid function \\(\\sigma\\) so that multiple relationships can exist between objects." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.435, + 0.295, + 0.451 + ], + "angle": 0, + "content": "3.3. Learning and Inference" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.458, + 0.469, + 0.517 + ], + "angle": 0, + "content": "To train EGTR, we perform multi-task learning. In addition to object detection and relation extraction, we devise connectivity prediction, an auxiliary task for relation extraction. The overall loss for the framework is as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.169, + 0.531, + 0.469, + 0.547 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\mathcal {L} _ {\\mathrm {o d}} + \\lambda_ {\\mathrm {r e l}} \\mathcal {L} _ {\\mathrm {r e l}} + \\lambda_ {\\mathrm {c o n}} \\mathcal {L} _ {\\mathrm {c o n}}, \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.557, + 0.469, + 0.648 + ], + "angle": 0, + "content": "where \\(\\mathcal{L}_{\\mathrm{od}}\\) is object detection loss in Eq. (2). \\(\\mathcal{L}_{\\mathrm{rel}}\\) and \\(\\mathcal{L}_{\\mathrm{con}}\\) are loss functions for the relation extraction and connectivity prediction, respectively. Since explicit object detection loss is used, EGTR has the capability to detect all nodes in the scene graph. The details of the loss for each task are provided below." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.666, + 0.27, + 0.68 + ], + "angle": 0, + "content": "3.3.1 Relation Extraction" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.69, + 0.469, + 0.825 + ], + "angle": 0, + "content": "We use binary cross-entropy loss for the relation extraction. To match predicted graph \\(\\hat{G} \\in \\mathbb{R}^{N \\times N \\times |\\mathcal{C}_p|}\\) and ground truth triplet set \\(\\mathcal{E}\\), we first encode \\(\\mathcal{E}\\) as a one-hot ground truth graph \\(G \\in \\mathbb{R}^{N \\times N \\times |\\mathcal{C}_p|}\\) by padding the regions that do not correspond to relations between the ground truth objects as zero. Then, we permute indices of the predicted graph using the permutation found in the object detection. From the permuted graph \\(\\hat{G}'\\), the loss for the relation extraction is calculated as \\(\\mathcal{L}_{\\mathrm{rel}} = \\mathcal{L}_r(\\hat{G}', G)\\)." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.826, + 0.47, + 0.902 + ], + "angle": 0, + "content": "However, since the size of the graph is proportional to the square of \\(N\\), the sparsity of the ground truth graph \\(G\\) is too severe. For instance, the density is only \\(10^{-14}\\) when \\(N\\) is set to 200 for the Visual Genome [12] validation dataset. Therefore, we divide \\(G\\) into three regions: (1) GT region," + }, + { + "type": "image", + "bbox": [ + 0.52, + 0.095, + 0.677, + 0.221 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.522, + 0.227, + 0.676, + 0.239 + ], + "angle": 0, + "content": "(subject - predicate - object)" + }, + { + "type": "image_footnote", + "bbox": [ + 0.555, + 0.24, + 0.642, + 0.25 + ], + "angle": 0, + "content": "girl - on - horse" + }, + { + "type": "image_footnote", + "bbox": [ + 0.555, + 0.25, + 0.61, + 0.261 + ], + "angle": 0, + "content": "jirl - riding" + }, + { + "type": "image_footnote", + "bbox": [ + 0.546, + 0.262, + 0.652, + 0.274 + ], + "angle": 0, + "content": "girl - wearing - pant" + }, + { + "type": "list", + "bbox": [ + 0.546, + 0.24, + 0.652, + 0.274 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.677, + 0.093, + 0.878, + 0.258 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.683, + 0.26, + 0.789, + 0.274 + ], + "angle": 0, + "content": "non-matching region" + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.287, + 0.893, + 0.357 + ], + "angle": 0, + "content": "Figure 4. Example of graph \\( G \\) region. The target graph has a shape of \\( N \\times N \\times |\\mathcal{V}_c| \\). The graph is severely sparse since the number of object queries \\( N \\) is set large enough to cover objects in the image. We split \\( G \\) into GT, negative, and non-matching regions." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.366, + 0.893, + 0.533 + ], + "angle": 0, + "content": "(2) negative region, and (3) non-matching region, as shown in Fig. 4. The GT region indicates the ground truth triplets where the value of \\( G \\) is 1. The negative region consists of the triplets in which both subjects and objects are composed of the ground truth objects, but no relation exists between them. The non-matching region represents the zero-padded region. It is paired with a region consisting of the object candidates that do not match the ground truth objects and match \\( \\phi \\) in Eq. (2). For each region, we apply different techniques to effectively train the relation extraction with object detection as subsequently described." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.538, + 0.892, + 0.734 + ], + "angle": 0, + "content": "Adaptive Smoothing. We propose a novel adaptive smoothing for the GT region. For \\( G_{ijk} \\) belonging to the GT region, the model is trained to predict the \\( k \\)-th predicate category between subject entity \\( v_i \\) and object entity \\( v_j \\). However, since \\( \\hat{v}_i' \\) and \\( \\hat{v}_j' \\), which match \\( v_i \\) and \\( v_j \\) respectively, do not have enough representations about the corresponding ground truth object at the beginning of the training, it may not be appropriate to predict the probability of the predicate as 1. Moreover, even when the object detection performance is reasonably assured, the detection performance may still vary for individual object candidates. Therefore, we reflect the detection performance of each object candidate on the relation label via adaptive smoothing." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.734, + 0.892, + 0.779 + ], + "angle": 0, + "content": "We first measure the uncertainty of each object candidate with the corresponding bipartite matching cost. For object candidate \\(\\hat{v}_i'\\), we define the uncertainty as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.578, + 0.786, + 0.891, + 0.804 + ], + "angle": 0, + "content": "\\[\nu _ {i} = \\sigma \\left(\\operatorname {c o s t} _ {i} - \\operatorname {c o s t} _ {\\min } + \\sigma^ {- 1} (\\alpha)\\right), \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.811, + 0.893, + 0.902 + ], + "angle": 0, + "content": "where \\(\\mathrm{cost}_i\\) denotes the matching cost and \\(\\mathrm{cost}_{\\mathrm{min}}\\) indicates the matching cost when \\(\\hat{v}_i^\\prime\\) perfectly matches \\(v_{i}\\) . \\(\\alpha\\) is a nonnegative hyperparameter representing the minimum uncertainty. We set the value of \\(G_{ijk}\\) to \\((1 - u_{i})(1 - u_{j})\\) taking the uncertainty into account. By using the relation label adjusted by the uncertainty, the multi-task learning of object" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "24233" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.47, + 0.122 + ], + "angle": 0, + "content": "detection and relation extraction is dynamically regulated according to the quality of the detected objects." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.127, + 0.468, + 0.263 + ], + "angle": 0, + "content": "Negative and Non-matching Sampling. Rather than employing all negatives, we sample them from the negative region. Inspired by the hard negative mining introduced by Liu et al. [21], we sort all negatives based on the predicted relation score \\(\\hat{G}_{ijk}^{\\prime}\\) and choose the top \\(k_{\\mathrm{neg}} \\times |\\mathcal{E}|\\) most challenging negatives. Similarly, we extract \\(k_{\\mathrm{non}} \\times |\\mathcal{E}|\\) hard samples from the non-matching region. As the non-matching region typically encompasses a substantial part of the graph \\(G\\), this approach notably diminishes sparsity." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.282, + 0.295, + 0.296 + ], + "angle": 0, + "content": "3.3.2 Connectivity Prediction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.306, + 0.468, + 0.457 + ], + "angle": 0, + "content": "Influenced by Graph-RCNN [40] that predicted a relatedness to prune object pairs, we propose a connectivity prediction that predicts whether at least one edge between two object nodes exists for the relation extraction. We get a connectivity graph \\(\\hat{E} \\in \\mathbb{R}^{N \\times N \\times 1}\\) in a similar way to get the relation graph in Eq. (6) with the same relation source representations. Instead of \\(\\mathrm{MLP}_{\\mathrm{rel}}\\) for multi-label prediction, we use another \\(\\mathrm{MLP}_{\\mathrm{con}}\\) for binary classification. We calculate the binary cross entropy loss from the permutated connectivity graph as follows: \\(\\mathcal{L}_{\\mathrm{con}} = \\mathcal{L}_r(\\hat{E}', E)\\)." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.476, + 0.197, + 0.489 + ], + "angle": 0, + "content": "3.3.3 Inference" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.5, + 0.468, + 0.622 + ], + "angle": 0, + "content": "For model inference, we get triplet scores by multiplying predicate score \\(\\hat{G}_{ijk}\\) by the corresponding class scores of \\(\\hat{v}_i^c\\) and \\(\\hat{v}_j^c\\). Note that we set \\(\\hat{G}_{iik}\\) to 0 to prevent self-connections in which the subject and object are the same entity. In addition, we enhance the triplet scores by utilizing the connectivity score \\(\\hat{E}_{ijk}\\). By multiplying the connectivity score, we can effectively filter out the triplets that are not likely to have a relation between the subject and object." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.637, + 0.208, + 0.653 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.661, + 0.365, + 0.677 + ], + "angle": 0, + "content": "4.1. Datasets and Evaluation Settings" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.685, + 0.468, + 0.745 + ], + "angle": 0, + "content": "We conduct experiments on two SGG datasets. We describe the datasets and evaluation settings for each dataset. Detailed settings for each dataset are presented in the supplementary materials." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.468, + 0.901 + ], + "angle": 0, + "content": "Visual Genome. Visual Genome [12] is the most representative SGG dataset, consisting of 57K training images, 5K validation images, and 26K test images. We follow the popular Visual Genome split, retaining the most frequent 150 object categories and 50 relation categories. We adopt Scene Graph Detection (SGDet) evaluation settings and report Recall@k (R@k) which is class agnostic and mean Recall@k (mR@k) that aggregates the recalls for each predicate category. Following Motifs [42], these metrics are measures with graph constraint, which means each" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.182 + ], + "angle": 0, + "content": "object pair can have a single predicate category. Since they are only related to triplet detection of the scene graph, we report AP50 that evaluates the detection performance of all objects appearing in the scene. Furthermore, we report the number of model parameters and Frames Per Second (FPS) to measure efficiency." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.188, + 0.892, + 0.355 + ], + "angle": 0, + "content": "Open Image V6. Open Image V6 [13] is also widely used dataset, comprising of 126K images for training, 2K for validation, and 5K for test set. It includes 601 object categories and 30 relation categories. For evaluation, we adopt both recall and weighted mean AP (wmAP) following standard settings. For recall evaluation, micro-R@50 is adopted. The WMAP is evaluated with two settings: \\(\\mathrm{wmAP}_{\\mathrm{rel}}\\) for predicting boxes of subject entity and object entity separately and \\(\\mathrm{wmAP}_{\\mathrm{phr}}\\) for predicting a union box of them. The final score is calculated by \\(0.2 \\times \\text{micro-R} @50 + 0.4 \\times \\mathrm{wmAP}_{\\mathrm{rel}} + 0.4 \\times \\mathrm{wmAP}_{\\mathrm{phr}}\\)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.367, + 0.716, + 0.382 + ], + "angle": 0, + "content": "4.2. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.39, + 0.892, + 0.618 + ], + "angle": 0, + "content": "We employ Deformable DETR [46] that improves the convergence speed of the DETR with ResNet-50 [6] as a backbone. It is worth noting that our approach can be extended to any object detector that incorporates self-attention mechanisms between object features, including models like DETR [1], Sparse R-CNN [32], and others. We follow the configurations of the original Deformable DETR, except that we use only 200 object queries. To speed up convergence, we first train the object detector with the target dataset and subsequently train the SGG task using the pretrained object detector, similar to prior work [17]. To calculate the overall loss in Equation 7, \\(\\lambda_{\\mathrm{rel}}\\) is set to 15. \\(\\lambda_{\\mathrm{con}}\\) is set to 30 and 90 for the Visual Genome and Open Image V6, respectively. We set \\(\\alpha\\) as \\(10^{-14}\\) for adaptive smoothing and both \\(k_{\\mathrm{neg}}\\) and \\(k_{\\mathrm{non}}\\) are set to 80." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.63, + 0.593, + 0.643 + ], + "angle": 0, + "content": "4.3. Results" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.654, + 0.892, + 0.714 + ], + "angle": 0, + "content": "We present quantitative results and perform a comparative analysis of our proposed framework with representative SGG models. Additional experimental results are included in the supplementary materials." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Visual Genome. Visual Genome results are shown in Tab. 1. Our proposed method demonstrates competitive performance with the current one-stage SGG models that have 1.5 to 6.5 times larger parameters with the fastest inference speed. In particular, our method achieves the highest object detection performance and shows comparable performance in the triplet detection compared to SSR-CNN [35], the state-of-the-art method. The results demonstrate that our method can generate scene graphs in an efficient and effective way by exploiting the by-products of the object detector. By applying logit adjustment [35], a technique for predicting tail predicate classes well, our" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "24234" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.078, + 0.089, + 0.895, + 0.357 + ], + "angle": 0, + "content": "
Model# params (M)FPSAP50R@20R@50R@100mR@20mR@50mR@100
two-stageIMP (EBM) [31, 39]322.22.028.118.125.931.22.84.25.4
VTransE [44]312.33.5-24.531.335.55.16.88.0
Motifs [42]369.91.928.125.132.136.94.15.56.8
VCTree [33]361.50.828.124.831.836.14.96.67.7
VCTree (TDE) [33, 34]361.30.828.114.019.423.26.99.311.1
VCTree (EBM) [31, 33]372.5-28.124.231.435.95.77.79.1
GPS-Net [19]----31.135.9-6.78.6
BGNN [16]341.91.729.023.331.035.87.510.712.6
one-stageFCSGG [20]87.16.028.516.121.325.12.73.64.2
RelTR [7]63.713.426.421.227.5-6.810.8-
SGTR [17]117.16.225.4-24.628.4-12.015.2
Relationformer [29]92.98.526.322.228.431.34.69.310.7
Iterative SGG [9]93.56.027.7†-29.732.1-8.08.8
SSR-CNN [35]274.34.023.8†25.832.736.96.18.410.0
SSR-CNN [35] LA,τ=0.3274.34.023.8†18.423.326.513.517.921.4
EGTR (Ours)42.514.730.823.530.234.35.57.910.1
EGTR (Ours) LA,τ=0.742.514.730.815.718.720.512.117.821.7
EGTR (Ours) LA,τ=0.542.514.730.819.724.226.711.017.121.4
EGTR (Ours) LA,τ=0.342.514.730.822.428.231.78.814.018.3
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.365, + 0.895, + 0.45 + ], + "angle": 0, + "content": "Table 1. Graph-Constraint results on the test set of Visual Genome dataset. We report the results of representative two-stage SGG models based on the Faster R-CNN [27] object detector with a ResNeXt-101-FPN [18, 38] backbone and concurrent one-stage SGG models. Among the one-stage models, the best is highlighted in bold, and the second-best is indicated with underlining. LA denotes logit adjustment proposed in SSR-CNN [35]. \\(\\dagger\\) represents that we concatenate the predicted subjects and objects and then apply non-maximum-suppression with a threshold of 0.3. Italic indicates the evaluation of metrics using publicly available model checkpoints. FPS is measured with a single V100 for images resized to a minimum of 600 for the shortest side and a maximum of 1000 for the longest side." + }, + { + "type": "table", + "bbox": [ + 0.078, + 0.458, + 0.47, + 0.596 + ], + "angle": 0, + "content": "
Modelscoremicro-R@50wmAPrelwmAPphr
Motifs [42]38.971.629.931.6
VCTree [33]40.274.134.233.1
GPS-Net [19]41.774.832.934.0
BGNN [16]42.175.033.534.2
RelTR [7]43.071.734.237.5
SGTR [17]42.359.937.038.7
SSR-CNN [35]49.476.741.543.6
EGTR (Ours)48.675.042.041.9
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.603, + 0.47, + 0.633 + ], + "angle": 0, + "content": "Table 2. Results on test set of Open Image V6. The score is a weighted sum of micro-R@50, \\(\\mathrm{wmAP}_{\\mathrm{rel}}\\), and \\(\\mathrm{wmAP}_{\\mathrm{phr}}\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.645, + 0.47, + 0.75 + ], + "angle": 0, + "content": "method performs favorably in the trade-off between \\(\\mathrm{R}@\\mathbf{k}\\) and \\(\\mathrm{mR}@\\mathbf{k}\\), showing superiority over existing state-of-the-art models. Note that for triplet detection models [9, 35] without an explicit object detector, AP50 is measured by applying non-maximum-suppression (NMS) to the union of predicted subjects and objects set. AP50 performance is discussed further in the section 4.5." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.755, + 0.47, + 0.817 + ], + "angle": 0, + "content": "Open Image V6. As depicted in Tab. 2, the experiments conducted on the Open Image V6 dataset also demonstrate competitive performance, underscoring the effectiveness and robustness of our method across different datasets." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.827, + 0.242, + 0.843 + ], + "angle": 0, + "content": "4.4. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.85, + 0.469, + 0.88 + ], + "angle": 0, + "content": "We analyze the effects of our model components. All of the ablation studies are done with the Visual Genome dataset." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.886, + 0.47, + 0.901 + ], + "angle": 0, + "content": "Relation Sources. In Tab. 3, we investigate the source of" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.461, + 0.892, + 0.688 + ], + "angle": 0, + "content": "relations used for the relation extractor. To verify the benefits of using attention queries and keys, we change \\( Q^l \\) and \\( K^l \\) in Eq. (3) to \\( Z^l \\), the hidden states of each layer. Remarkably, when simply using the hidden states of all layers, the triplet detection performance decreases. Surprisingly, the performance is lower than that of using only \\( R_z \\) from the last layer. On the other hand, when only attention by-products are used as a relation source, the performance is similar to that of using only hidden states of the last layer, the most contextualized representations for detecting objects. Incorporating both attention by-products and final hidden states yields improved performance by amalgamating varied information. The results underscore that the attention by-products of the object detectors contain rich information for relation extraction." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.696, + 0.892, + 0.833 + ], + "angle": 0, + "content": "Training Techniques. In Tab. 4, we ablate proposed techniques. Without any techniques, our model framework shows low performance. When each technique is applied, it substantially improves both R@50 and mR@50, and the best performance is achieved when all techniques are combined. The results demonstrate that our proposed techniques help to train the model framework effectively. Additional experimental results are provided in the supplementary materials." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Sampling Methodology. In Tab. 5, we investigate the impact of hard sampling methods for the negative and non-matching regions. Hard negatives and hard non-matchings show a trade-off that improves the performance of R@50" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "24235" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.079, + 0.089, + 0.472, + 0.168 + ], + "angle": 0, + "content": "
Rl^l a sourceRl^l aRzR@50mR@50
Ql & Kl✓✓30.27.9
Zl✓✓29.67.4
-✓29.97.6
Ql & Kl✓29.87.7
" + }, + { + "type": "table_caption", + "bbox": [ + 0.14, + 0.175, + 0.408, + 0.189 + ], + "angle": 0, + "content": "Table 3. Ablation study on relation source." + }, + { + "type": "table", + "bbox": [ + 0.079, + 0.204, + 0.47, + 0.294 + ], + "angle": 0, + "content": "
adaptive smoothing\\( \\mathcal{L}_{\\text{con}} \\)samplingR@50mR@50
26.65.3
✓28.36.5
✓29.67.0
✓28.97.1
✓✓✓30.27.9
" + }, + { + "type": "table_caption", + "bbox": [ + 0.122, + 0.301, + 0.425, + 0.315 + ], + "angle": 0, + "content": "Table 4. Ablation study on proposed techniques." + }, + { + "type": "table", + "bbox": [ + 0.079, + 0.329, + 0.47, + 0.394 + ], + "angle": 0, + "content": "
hard negativehard non-matchingR@50mR@50
✓✓30.27.9
✓30.07.1
✓29.67.7
" + }, + { + "type": "table_caption", + "bbox": [ + 0.129, + 0.402, + 0.418, + 0.416 + ], + "angle": 0, + "content": "Table 5. Ablation study on sampling methods." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.429, + 0.47, + 0.61 + ], + "angle": 0, + "content": "and mR@50, respectively, and additional performance improvements are achieved for both measures when they are used together. For hard negatives, we speculate that R@50 performance is improved by selecting negatives for the tail predicate class, which is likely to exist in reality but is not in the annotation. For hard non-matchings, mR@50 performance seems to have improved by choosing object candidates that are akin to the ground truth objects but unmatched with them. This encourages the model to predict diverse predicate classes by preventing duplicate object pairs, which are likely to predict head predicate classes, from being predicted." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.621, + 0.196, + 0.635 + ], + "angle": 0, + "content": "4.5. Discussion" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.644, + 0.469, + 0.825 + ], + "angle": 0, + "content": "Regarding the highest AP50 score, we posit that it is because our proposed method focuses on not only objects with relations but also objects without any relations. To validate this, we calculate AP50 for two subsets of the ground-truth objects: those with relations and those without relations. We compare our method with two triplet detection-based models. Tab. 6 illustrates that our method achieves significantly high \\(\\mathrm{AP50}_{\\mathrm{no - rel}}\\). Considering that triplet detection performance is related to \\(\\mathrm{AP50}_{\\mathrm{rel}}\\), while triplet detection-based models only focus on objects with relations to detect triplets well, our method is capable of extracting the complete scene graph, including objects without relations." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.47, + 0.902 + ], + "angle": 0, + "content": "For further analysis, we visualize bounding boxes for subjects and objects from the top 20 predictions in Fig. 5. For SSR-CNN [35], which directly predicts triplets, we observe a significant number of overlapping bounding boxes. On the other hand, our proposed model leverages predic" + }, + { + "type": "table", + "bbox": [ + 0.5, + 0.089, + 0.892, + 0.154 + ], + "angle": 0, + "content": "
ModelAP50\\( AP50_{\\text{rel}} \\)\\( AP50_{\\text{no-rel}} \\)
Iterative SGG [9]†27.724.37.8
SSR-CNN [35]†23.820.27.4
EGTR (Ours)30.824.310.7
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.161, + 0.894, + 0.245 + ], + "angle": 0, + "content": "Table 6. AP50 for two subsets of objects. \\(\\mathrm{AP50}_{\\mathrm{rel}}\\) and \\(\\mathrm{AP50}_{\\mathrm{no - rel}}\\) denote scores evaluated only on objects having at least one edge and no edge, respectively. Note that each measure operates on its unique scale since the ground-truth objects used for each measure differ. \\(\\dagger\\) indicates that additional NMS is applied to the union of subjects and objects set." + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.258, + 0.695, + 0.359 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.359, + 0.695, + 0.457 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.548, + 0.46, + 0.648, + 0.472 + ], + "angle": 0, + "content": "(a) SSR-CNN [35]" + }, + { + "type": "image", + "bbox": [ + 0.696, + 0.259, + 0.887, + 0.359 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.696, + 0.359, + 0.887, + 0.457 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.746, + 0.46, + 0.837, + 0.472 + ], + "angle": 0, + "content": "(b) EGTR (Ours)" + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.478, + 0.892, + 0.52 + ], + "angle": 0, + "content": "Figure 5. Comparison of detected subjects and objects. We select the top 20 predictions of SSR-CNN and EGTR and visualize bounding boxes of detected subjects and objects." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.533, + 0.892, + 0.593 + ], + "angle": 0, + "content": "tions from an explicit object detector, reducing redundant predictions. Furthermore, our method detects various subjects and objects that appear in the scene rather than focusing on objects that are more likely to have relations." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.607, + 0.619, + 0.622 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.632, + 0.892, + 0.844 + ], + "angle": 0, + "content": "In this study, we propose the lightweight one-stage scene graph generator EGTR. We significantly reduce the model complexity by harnessing the relationships among the object queries learned from the self-attention layers of the object decoder. Furthermore, we devise a novel adaptive smoothing technique that helps multi-task learning of object detection and relation extraction by adjusting the relation label according to the object detection performance. As an auxiliary task of relation extraction, connectivity prediction contributes to the effective learning of EGTR. We conduct extensive experiments and ablation studies, and the results demonstrate that EGTR achieves the highest object detection performance and competitive triplet detection capabilities with the fastest inference speed." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.848, + 0.892, + 0.894 + ], + "angle": 0, + "content": "Acknowledgements. We thank Taeho Kil, Geewook Kim, and the anonymous reviewers for their insightful comments and suggestions." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.946, + 0.521, + 0.957 + ], + "angle": 0, + "content": "24236" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.17 + ], + "angle": 0, + "content": "[1] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In ECCV, pages 213-229, 2020. 1, 2, 3, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.172, + 0.47, + 0.2 + ], + "angle": 0, + "content": "[2] Zhe Chen, Shaoli Huang, and Dacheng Tao. Context refinement for object detection. In ECCV, pages 71-86, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.202, + 0.471, + 0.242 + ], + "angle": 0, + "content": "[3] Yuren Cong, Michael Ying Yang, and Bodo Rosenhahn. Reltr: Relation transformer for scene graph generation. IEEE TPAMI, 2023. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.245, + 0.47, + 0.286 + ], + "angle": 0, + "content": "[4] Naina Dhingra, Florian Ritter, and Andreas Kunz. Bgt-net: Bidirectional gru transformer network for scene graph generation. In CVPR, pages 2150-2159, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.288, + 0.47, + 0.328 + ], + "angle": 0, + "content": "[5] Lizhao Gao, Bo Wang, and Wenmin Wang. Image captioning with scene-graph based semantic concepts. In ICMLC, pages 225-229, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.331, + 0.47, + 0.373 + ], + "angle": 0, + "content": "[6] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, pages 770-778, 2016. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.375, + 0.47, + 0.415 + ], + "angle": 0, + "content": "[7] Han Hu, Jiayuan Gu, Zheng Zhang, Jifeng Dai, and Yichen Wei. Relation networks for object detection. In CVPR, pages 3588-3597, 2018. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.418, + 0.47, + 0.471 + ], + "angle": 0, + "content": "[8] Justin Johnson, Ranjay Krishna, Michael Stark, Li-Jia Li, David Shamma, Michael Bernstein, and Li Fei-Fei. Image retrieval using scene graphs. In CVPR, pages 3668-3678, 2015. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.474, + 0.47, + 0.502 + ], + "angle": 0, + "content": "[9] Siddhesh Khandelwal and Leonid Sigal. Iterative scene graph generation. In NeurIPS, 2022. 1, 3, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.504, + 0.47, + 0.558 + ], + "angle": 0, + "content": "[10] Dong-Jin Kim, Jinsoo Choi, Tae-Hyun Oh, and In So Kweon. Dense relational captioning: Triple-stream networks for relationship-based captioning. In CVPR, pages 6271-6280, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.561, + 0.47, + 0.588 + ], + "angle": 0, + "content": "[11] Rajat Koner, Suprosanna Shit, and Volker Tresp. Relation transformer network. ECCV, pages 422-439, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.59, + 0.47, + 0.659 + ], + "angle": 0, + "content": "[12] Ranjay Krishna, Yuke Zhu, Oliver Groth, Justin Johnson, Kenji Hata, Joshua Kravitz, Stephanie Chen, Yannis Kalantidis, Li-Jia Li, David A Shamma, et al. Visual genome: Connecting language and vision using crowdsourced dense image annotations. IJCV, 123:32-73, 2017. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.661, + 0.47, + 0.743 + ], + "angle": 0, + "content": "[13] Alina Kuznetsova, Hassan Rom, Neil Alldrin, Jasper Uijlings, Ivan Krasin, Jordi Pont-Tuset, Shahab Kamali, Stefan Popov, Matteo Malloci, Alexander Kolesnikov, et al. The open images dataset v4: Unified image classification, object detection, and visual relationship detection at scale. IJCV, 128(7):1956-1981, 2020. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.746, + 0.47, + 0.787 + ], + "angle": 0, + "content": "[14] Linjie Li, Zhe Gan, Yu Cheng, and Jingjing Liu. Relation-aware graph attention network for visual question answering. In ICCV, pages 10313-10322, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.789, + 0.47, + 0.843 + ], + "angle": 0, + "content": "[15] Lin Li, Guikun Chen, Jun Xiao, Yi Yang, Chunping Wang, and Long Chen. Compositional feature augmentation for unbiased scene graph generation. In ICCV, pages 21685-21695, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.846, + 0.47, + 0.9 + ], + "angle": 0, + "content": "[16] Rongjie Li, Songyang Zhang, Bo Wan, and Xuming He. Bipartite graph network with adaptive message passing for unbiased scene graph generation. In CVPR, pages 11109-11119, 2021. 1, 2, 7" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.471, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.135 + ], + "angle": 0, + "content": "[17] Rongjie Li, Songyang Zhang, and Xuming He. Sgtr: End-to-end scene graph generation with transformer. In CVPR, pages 19486-19496, 2022. 1, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.137, + 0.892, + 0.19 + ], + "angle": 0, + "content": "[18] Tsung-Yi Lin, Piotr Dollar, Ross Girshick, Kaiming He, Bharath Hariharan, and Serge Belongie. Feature pyramid networks for object detection. In CVPR, pages 2117-2125, 2017. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.194, + 0.892, + 0.236 + ], + "angle": 0, + "content": "[19] Xin Lin, Changxing Ding, Jinquan Zeng, and Dacheng Tao. Gps-net: Graph property sensing network for scene graph generation. In CVPR, pages 3746-3753, 2020. 1, 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.239, + 0.892, + 0.279 + ], + "angle": 0, + "content": "[20] Hengyue Liu, Ning Yan, Masood Mortazavi, and Bir Bhanu. Fully convolutional scene graph generation. In CVPR, pages 11546-11556, 2021. 1, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.282, + 0.892, + 0.336 + ], + "angle": 0, + "content": "[21] Wei Liu, Dragomir Anguelov, Dumitru Erhan, Christian Szegedy, Scott Reed, Cheng-Yang Fu, and Alexander C Berg. Ssd: Single shot multibox detector. In ECCV, pages 21-37, 2016. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.34, + 0.892, + 0.394 + ], + "angle": 0, + "content": "[22] Yong Liu, Ruiping Wang, Shiguang Shan, and Xilin Chen. Structure inference net: Object detection using scene-level context and instance-level relationships. In CVPR, pages 6985-6994, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.398, + 0.892, + 0.439 + ], + "angle": 0, + "content": "[23] Cewu Lu, Ranjay Krishna, Michael Bernstein, and Li Fei-Fei. Visual relationship detection with language priors. In ECCV, pages 852–869, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.442, + 0.892, + 0.509 + ], + "angle": 0, + "content": "[24] Yichao Lu, Himanshu Rai, Jason Chang, Boris Knyazev, Guangwei Yu, Shashank Shekhar, Graham W Taylor, and Maksims Volkovs. Context-aware scene graph generation with seq2seq transformers. In ICCV, pages 15931-15941, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.513, + 0.892, + 0.555 + ], + "angle": 0, + "content": "[25] Yukuan Min, Among Wu, and Cheng Deng. Environment-invariant curriculum relation learning for fine-grained scene graph generation. In ICCV, pages 13296-13307, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.557, + 0.892, + 0.585 + ], + "angle": 0, + "content": "[26] Alejandro Newell and Jia Deng. Pixels to graphs by associative embedding. NeurIPS, 30, 2017. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.587, + 0.892, + 0.628 + ], + "angle": 0, + "content": "[27] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. NeurIPS, 28, 2015. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.631, + 0.892, + 0.714 + ], + "angle": 0, + "content": "[28] Sebastian Schuster, Ranjay Krishna, Angel Chang, Li Fei-Fei, and Christopher D Manning. Generating semantically precise scene graphs from textual descriptions for improved image retrieval. In Proceedings of the fourth workshop on vision and language, pages 70–80. Association for Computational Linguistics, 2015. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.716, + 0.892, + 0.786 + ], + "angle": 0, + "content": "[29] Suprosanna Shit, Rajat Koner, Bastian Wittmann, Johannes Paetzold, Ivan Ezhov, Hongwei Li, Jiazhen Pan, Sahand Sharifzadeh, Georgios Kaissis, Volker Tresp, et al. Relationformer: A unified framework for image-to-graph generation. In ECCV, pages 422-439, 2022. 1, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.788, + 0.892, + 0.842 + ], + "angle": 0, + "content": "[30] Gopika Sudhakaran, Devendra Singh Dhami, Kristian Kersting, and Stefan Roth. Vision relation transformer for unbiased scene graph generation. In ICCV, pages 21882-21893, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.845, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[31] Mohammed Suhail, Abhay Mittal, Behjat Siddiquie, Chris Broaddus, Jayan Eledath, Gerard Medioni, and Leonid Sigal. Energy-based learning for scene graph generation. In CVPR, pages 13936-13945, 2021. 7" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "24237" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.161 + ], + "angle": 0, + "content": "[32] Peize Sun, Rufeng Zhang, Yi Jiang, Tao Kong, Chenfeng Xu, Wei Zhan, Masayoshi Tomizuka, Lei Li, Zehuan Yuan, Changhu Wang, and Ping Luo. Sparse r-cnn: End-to-end object detection with learnable proposals. In CVPR, pages 14454-14463, 2021. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.163, + 0.47, + 0.217 + ], + "angle": 0, + "content": "[33] Kaihua Tang, Hanwang Zhang, Baoyuan Wu, Wenhan Luo, and Wei Liu. Learning to compose dynamic tree structures for visual contexts. In CVPR, pages 6619-6628, 2019. 1, 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.219, + 0.47, + 0.262 + ], + "angle": 0, + "content": "[34] Kaihua Tang, Yulei Niu, Jianqiang Huang, Jiaxin Shi, and Hanwang Zhang. Unbiased scene graph generation from biased training. In CVPR, pages 3716-3725, 2020. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.263, + 0.47, + 0.303 + ], + "angle": 0, + "content": "[35] Yao Teng and Limin Wang. Structured sparse r-cnn for direct scene graph generation. In CVPR, pages 19437-19446, 2022. 1, 3, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.305, + 0.47, + 0.358 + ], + "angle": 0, + "content": "[36] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. NeurIPS, 30, 2017. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.362, + 0.47, + 0.403 + ], + "angle": 0, + "content": "[37] Sanghyun Woo, Dahun Kim, Donghyeon Cho, and In So Kweon. Linknet: Relational embedding for scene graph. NeurIPS, 31, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.405, + 0.47, + 0.446 + ], + "angle": 0, + "content": "[38] Saining Xie, Ross Girshick, Piotr Dollar, Zhuowen Tu, and Kaiming He. Aggregated residual transformations for deep neural networks. In CVPR, pages 1492-1500, 2017. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.448, + 0.47, + 0.489 + ], + "angle": 0, + "content": "[39] Danfei Xu, Yuke Zhu, Christopher B Choy, and Li Fei-Fei. Scene graph generation by iterative message passing. In CVPR, pages 5410–5419, 2017. 1, 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.49, + 0.47, + 0.532 + ], + "angle": 0, + "content": "[40] Jianwei Yang, Jiasen Lu, Stefan Lee, Dhruv Batra, and Devi Parikh. Graph r-cnn for scene graph generation. In ECCV, pages 670-685, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.533, + 0.47, + 0.574 + ], + "angle": 0, + "content": "[41] Xu Yang, Kaihua Tang, Hanwang Zhang, and Jianfei Cai. Auto-encoding scene graphs for image captioning. In CVPR, pages 10685–10694, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.576, + 0.47, + 0.617 + ], + "angle": 0, + "content": "[42] Rowan Zellers, Mark Yatskar, Sam Thomson, and Yejin Choi. Neural motifs: Scene graph parsing with global context. In CVPR, pages 5831-5840, 2018. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.619, + 0.47, + 0.66 + ], + "angle": 0, + "content": "[43] Cheng Zhang, Wei-Lun Chao, and Dong Xuan. An empirical study on leveraging scene graphs for visual question answering. In BMVC, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.661, + 0.47, + 0.702 + ], + "angle": 0, + "content": "[44] Hanwang Zhang, Zawlin Kyaw, Shih-Fu Chang, and Tat-Seng Chua. Visual translation embedding network for visual relation detection. In CVPR, page 3107-3115, 2017. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.704, + 0.47, + 0.746 + ], + "angle": 0, + "content": "[45] Chaofan Zheng, Xinyu Lyu, Yuyu Guo, Pengpeng Zeng, Jingkuan Song, and Lianli Gao. Learning to generate scene graph from head to tail. In ICME, pages 1-6. IEEE, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.747, + 0.47, + 0.788 + ], + "angle": 0, + "content": "[46] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable detr: Deformable transformers for end-to-end object detection. In ICLR, 2021. 2, 3, 6" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.788 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "24238" + } + ] +] \ No newline at end of file diff --git a/2024/EGTR_ Extracting Graph from Transformer for Scene Graph Generation/174c2199-8aa8-4e45-8d2e-205eb2e2e099_origin.pdf b/2024/EGTR_ Extracting Graph from Transformer for Scene Graph Generation/174c2199-8aa8-4e45-8d2e-205eb2e2e099_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..f47d741fc0e6562f25ff1f767159996e8d16fe43 --- /dev/null +++ b/2024/EGTR_ Extracting Graph from Transformer for Scene Graph Generation/174c2199-8aa8-4e45-8d2e-205eb2e2e099_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:48415d99182105a59d0b4f6135ceef5c08353ce230019ec47ed4a7853f3f29d7 +size 3379274 diff --git a/2024/EGTR_ Extracting Graph from Transformer for Scene Graph Generation/full.md b/2024/EGTR_ Extracting Graph from Transformer for Scene Graph Generation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..db501dc96ec5c81d6c3b9aaafc2d6c0335f5fc09 --- /dev/null +++ b/2024/EGTR_ Extracting Graph from Transformer for Scene Graph Generation/full.md @@ -0,0 +1,340 @@ +# EGTR: Extracting Graph from Transformer for Scene Graph Generation + +Jinbae Im1 JeongYeon Nam1 Nokyung Park1,2,3* Hyungmin Lee2 Seunghyun Park1 +1NAVER Cloud AI 2NAVER 3Korea University +{jinbae.im, jy.nam, nokyung.park99, hyungmin.lee, seung.park}@navercorp.com + +# Abstract + +Scene Graph Generation (SGG) is a challenging task of detecting objects and predicting relationships between objects. After DETR was developed, one-stage SGG models based on a one-stage object detector have been actively studied. However, complex modeling is used to predict the relationship between objects, and the inherent relationship between object queries learned in the multi-head self-attention of the object detector has been neglected. We propose a lightweight one-stage SGG model that extracts the relation graph from the various relationships learned in the multi-head self-attention layers of the DETR decoder. By fully utilizing the self-attention by-products, the relation graph can be extracted effectively with a shallow relation extraction head. Considering the dependency of the relation extraction task on the object detection task, we propose a novel relation smoothing technique that adjusts the relation label adaptively according to the quality of the detected objects. By the relation smoothing, the model is trained according to the continuous curriculum that focuses on object detection task at the beginning of training and performs multi-task learning as the object detection performance gradually improves. Furthermore, we propose a connectivity prediction task that predicts whether a relation exists between object pairs as an auxiliary task of the relation extraction. We demonstrate the effectiveness and efficiency of our method for the Visual Genome and Open Image V6 datasets. Our code is publicly available at https://github.com/naver-ai/egtr. + +# 1. Introduction + +Scene Graph Generation (SGG) [8] aims to generate a scene graph that represents objects as nodes and relationships between objects as edges from an image as shown in Fig. 1b. SGG is a challenging task as it is required to go beyond simply detecting objects but predicting the relationships between them based on a comprehensive understanding of the + +![](images/da086494a3493dff7e071faf8fea070928900aae618d983c549510da6ad11f8c.jpg) +(a) Visual Genome example + +![](images/9fcc74a969bc741c22cf5e31e09c3ed6a0b106a111765290803f5a92ea5dbd50.jpg) +(b) Scene Graph + +![](images/d3d669486085177c87b65b065112e8bb0b5abe8f60a9d5e1549828477fc1cbc7.jpg) +(c) Attention Graph +Figure 1. Motivation. SGG task aims to predict scene graph (Fig. 1b) with objects as nodes and relations as edges. We draw a plausible attention graph (Fig. 1c) simply by connecting objects with high attention weights to edges from the self-attention layers of the pre-trained DETR. It shows the potential for the self-attention from the object detector to contain rich information that aids in predicting the relations of the scene graph. + +scene. Since a scene graph provides structural information of the image, it can be used for various vision tasks that require a higher level of understanding and reasoning about images such as image captioning [5, 10, 41], image retrieval [8, 28], and visual question answering [14, 43]. + +Most previous studies [16, 19, 33, 39, 42] took two-stage SGG approaches that detect objects first and then predict their relations. However, these approaches incurred high computational costs and risks of error propagation from the object detection. To address the drawbacks of the two-stage approaches, one-stage SGG models that perform object detection and relation prediction at once [3, 9, 17, 20, 26, 29, 35] have been studied recently, leveraging one-stage object detectors such as DETR [1]. Since edges in a scene graph can be represented as subject-predicate-object triplets, many studies embraced triplet-based approaches as shown in Figs. 2a and 2b. However, object-triplet detection models [3, 17] required sophisticated triplet detectors to obtain necessary information for triplet queries from the object detector. Moreover, triplet detection models [9, 35] lacked the ability to detect objects without relation such as "tree" in Fig. 1b, by focusing solely on the triplet detection without an object detector. Considering that objects without relations account for more than + +![](images/eeef0756f110b56051eef1ee222a4df8ecbb04eaa0125bb5bc46ca221ac47ed5.jpg) +(a) Object-Triplet Detection Models + +![](images/17fbf1da620e9dc2c721e805545b7c7f43c92cb1db5cdc6e1494d2fbdc0930f5.jpg) +(b) Triplet Detection Models +Figure 2. Comparison with existing one-stage SGG models. (a) Object-Triplet Detection Models introduce additional triplet queries and a triplet detector to the object detector. The triplet detector requires additional modules to incorporate information from the object detector into the triplet queries. (b) Triplet Detection Models focus on detecting triplets directly without an object detector. Objects without relations may not be detected. (c) Relation Extraction Models extract relations from the object detector without a separate triplet detector. In particular, ours extracts relations more effectively by utilizing by-products from the self-attention of the object detector. + +![](images/1200dc7f83785412b59a2e3649fd9623129020a4a233efc8ba94d41c3b085407.jpg) +(c) Relation Extraction Models + +$42\%$ in Visual Genome [12] data, their priority lies in detecting a sub-graph rather than the complete scene graph. + +To address the shortcomings of the existing one-stage SGG models, we focus on the relationships between objects inherent in the object detector. As shown in Fig. 1a, objects are related to each other. For instance, when a horse appears in a scene, a person is likely to appear in the scene, and clothing such as hats, jackets, and pants often depicts the person's current situation. From this intuition, there has been a long belief [2, 7, 22] that modeling relationship or context between objects would be beneficial to object detection task. Accordingly, recent one-stage object detectors [1, 32, 46] have incorporated self-attention layers [36] to implicitly model the relationships among the object queries. We hypothesize that self-attention between object queries learned in a one-stage object detector might contain valuable information for predicting triplet outputs. In our preliminary investigation, we are able to extract a plausible attention graph by simply connecting two object queries with high attention weights from the pre-trained DETR, as shown in Fig. 1c. It shows the potential that the attention weights between object queries can be interpreted as relations between them. + +From the findings, we propose a lightweight one-stage scene graph generator EGTR, which stands for Extracting Graph from TRansformer. We design the model to comprehensively leverage the by-products of the object detector, eliminating the need for a separate triplet detector, as depicted in Fig. 2c. From the multi-head self-attention layers of the object detector, we regard an attention query and key, where their relations are learned in the attention weights, as the subject entity and object entity, respectively. Subsequently, we leverage a shallow classifier to predict relations between them. Due to the abundant information about the relationships among the objects present in the by-products derived from all layers of self-attention, we can effectively extract the scene graph. + +Since the relation extraction task is dependent on the object detection task, we speculate that performing relation ex + +traction without sufficiently learned representations of the object queries might be harmful. Therefore, we devise a novel adaptive smoothing technique that smoothes the value of the ground truth relation label based on the object detection performance. With the adaptive smoothing, the model is trained with a continuous curriculum that initially focuses on object detection and gradually performs multi-task learning. Furthermore, we propose a connectivity prediction task as an auxiliary task for relation extraction, aiming to predict the existence of any relationship between a subject entity and an object entity. This auxiliary task facilitates the acquisition of representations for the relation extraction. + +To verify the effectiveness of the proposed method, we conduct experiments on two representative SGG datasets: Visual Genome [12] and Open Images V6 [13]. By actively utilizing the by-products of the object detector, EGTR shows the best object detection performance and comparable triplet detection performance, with the fewest parameters and the fastest inference speed. + +Our main contributions can be summarized as follows: + +- We propose EGTR that generates scene graphs efficiently and effectively by utilizing the multi-head self-attention by-products derived from the object detector. +- We present adaptive smoothing, enabling effective multitask learning for both object detection and relation extraction. In addition, the proposed connectivity prediction offers clues to the relation extraction. +- Our comprehensive experiments show the superiority of the proposed model framework and the effectiveness of the devised training techniques. + +# 2. Related Work + +SGG models can be categorized into two groups: two-stage models and one-stage models. For two-stage models [4, 11, 15, 16, 19, 23-25, 30, 33, 37, 39, 42, 45], separate object detection model and relation prediction model are trained sequentially. They usually detect $N$ objects from the off-the-shelf object detector such as Faster R-CNN [27], + +and then all possible combinations of the detected objects are fed into the relation prediction model to predict the relations between each object pair. Although they showed high relation extraction performance, they have the inherent limitation that the object detector that helps with relation extraction is trained separately, resulting in a significant increase in model complexity. + +As for the one-stage models [3, 9, 17, 20, 26, 29, 35], the object detection and relation prediction are trained in an end-to-end manner. Early studies [20, 26] proposed fully convolutional SGG models and took pixel-based approaches. After DETR [1] brought a huge success as a Transformer [36]-based one-stage object detector, many one-stage SGG studies are based on one-stage object detectors [1, 32, 46]. They efficiently modeled SGG by introducing object queries or triplet queries. We categorize them into three distinct groups: (a) object-triplet detection models, (b) triplet prediction models, and (c) relation extraction models, as shown in Fig. 2. + +Object-Triplet Detection Models. Object-triplet detection models are characterized by introducing additional triplet queries and building a triplet predictor on top of the object detector as shown in Fig. 2a. RelTR [3] introduced paired subject queries and object queries, and SGTR [17] introduced compositional queries decoupled into subjects, objects, and predicates. As the introduced queries are initialized without prior cues from the object detector, the triplet predictor requires modules to incorporate information from the outputs of the object detector and modules for triplet queries to exchange information with each other. It led to the triplet predictor having an intricate structure. In contrast, we refrain from introducing additional queries and regard the attention queries and attention keys, whose relationships are learned in the object detector, as subject and object queries, respectively. + +Triplet Detection Models. Triplet detection models directly detect triplets using triplet queries without an object detector, as shown in Fig. 2b. Iterative SGG [9] introduced subject, object, and predicate queries and modeled the conditional dependencies among them with separate subject, object, and predicate multi-layer Transformer decoders. Inspired by Sparse R-CNN [32], Structured Sparse R-CNN (SSR-CNN) [35] designed triplet queries consisting of subject box, object box, subject, object, and predicate queries. Since it was difficult to train the model only with sparse triplet annotations, SSR-CNN introduced somewhat intricate training details: it detected object pairs with a Siamese Sparse RCNN model and auxiliary queries and performed additional triplet matching using the detected object pairs as a pseudo label. Although they show excellent triplet detection performance, the model architecture became even more intricate to detect subjects and objects separately since + +an explicit object detector is not used. Last but not least, the triplet prediction models focus on detecting a sub-graph composed only of objects with relations, overlooking objects in an image that lack explicit relations. + +Relation Extraction Models. Relation extraction models extract a scene graph using a lightweight relation predictor without separate triplet queries or triplet detector. Relationformer [29] added a special "[rln]" token to capture global information in conjunction with object queries. They concatenated the final hidden representations of the object query pairs and the relation token, followed by a shallow fully connected network for relation prediction. In this work, in conjunction with the final hidden representations, we use the inherent relationship information among the object queries learned in the multi-head self-attention layer of the object detector. Furthermore, we propose training techniques to boost multi-task learning, leading to significantly improved performance with architectural simplicity. + +# 3. Method + +# 3.1. Preliminaries + +In this section, we first introduce the formulation of the SGG task and then provide a brief review of the one-stage object detector that serves as the basis for our study. + +# 3.1.1 Formulation + +Scene graph generation is a task of generating a scene graph $\mathcal{G} = \{\mathcal{V},\mathcal{E}\}$ from an image, where $\mathcal{V}$ denotes the node set consisting of objects and $\mathcal{E}$ denotes the edge set that represents the relation between objects. Each object entity $v_{i}\in \mathcal{V}$ has an object category label $v_{i}^{c}$ from a set of object categories $\mathcal{C}_v$ and box coordinates $v_{i}^{b}$ . Each relation $e_j\in \mathcal{E}$ represents the $j$ -th triplet $(s_j,p_j,o_j)$ , where subject $s_j$ and object $o_j$ indicate related object entities and predicate $p_j$ has a relation category label $p_j^c$ from a set of predicate categories $\mathcal{C}_p$ . Generating $\mathcal{V}$ and $\mathcal{E}$ correspond to object detection and relation extraction, respectively. + +# 3.1.2 One-stage Object Detector + +Our model framework is deeply influenced by DETR [1], a one-stage object detector. In DETR, representations for input images are learned through CNN and Transformer [36] encoder. Transformer decoder enhances object queries by utilizing self- and cross-attention mechanisms. The final object detection heads predict object category labels and box coordinates from the contextualized object queries. + +Backbone. CNN backbone generates $C$ -dimensional feature map $F \in \mathbb{R}^{C \times H_F \times W_F}$ from an input image $x_{\mathrm{img}} \in \mathbb{R}^{3 \times H_0 \times W_0}$ . For ResNet-50 backbone, $H_F$ and $W_F$ are set to $H_0 / 32$ and $W_0 / 32$ , respectively. + +![](images/8ecf05244754a50b2be1aece4d420418de000403424f1f62df4891ed67db2042.jpg) +Figure 3. The overall architecture of EGTR. We present a novel lightweight relation extractor, EGTR, which fully utilizes the self-attention of the DETR decoder. We extract query and key representations from each self-attention layer and concatenate them pairwise to represent relations between them. Additionally, we leverage the last hidden representation in the same manner. To effectively aggregate information, we apply a gated sum and then predict relation with a shallow relation head. + +![](images/051d3620529018711d7ac2cc82099c6a645af5e74a4b87f1aedd5638cdb66c02.jpg) + +![](images/cbe093ab64a6516b8c1b4273701a8dbd74801876a7784e78ebbac8263356f574.jpg) + +Transformer Encoder. Transformer encoder enhances image representations based on multi-head self-attention. After applying a $1 \times 1$ convolution to reduce dimension from $C$ to $d_{\mathrm{model}}$ , the feature map is flattened to get an input sequence with length $H_{F}W_{F}$ for the Transformer encoder. Additional positional encodings are used to reflect the spatial information of the feature maps. + +Transformer Decoder. Transformer decoder takes $N$ object queries as input and returns their representations. Each object query detects an individual object, and $N$ is usually set large enough to cover all objects in the image. Through alternating self-attention and cross-attention layers, object queries learn features of object candidates in the input image. Note that causal attention mask is not applied to the self-attention layer; therefore, fully connected $N \times N$ attention weights among the $N$ object queries are learned from the multi-head self-attention as follows: + +$$ +A _ {h} ^ {l} = f \left(Q _ {h} ^ {l}, K _ {h} ^ {l}\right) = \operatorname {s o f t m a x} \left(Q _ {h} ^ {l} K _ {h} ^ {T} / \sqrt {d _ {\text {h e a d}}}\right), \tag {1} +$$ + +where $A_h^l \in \mathbb{R}^{N \times N}$ denotes the attention weights from the $h$ -th head in the $l$ -th layer. $Q_h^l \in \mathbb{R}^{N \times d_{\mathrm{head}}}$ and $K_h^l \in \mathbb{R}^{N \times d_{\mathrm{head}}}$ are attention queries and keys, respectively. + +Object Detection Heads. Object detection heads detect $N$ object candidates $\{\hat{v}_i\}_{i = 1}^N$ from the last layer representations of the object queries $Z^{L}$ . For each object query, a linear layer is used to predict the object category label $\hat{v}_i^c\in \mathbb{R}^{N\times |\mathcal{C}_v|}$ , and three-layer perceptron with ReLU activation is used to obtain the box coordinates $\hat{v}_i^b\in \mathbb{R}^{N\times 4}$ . + +Object Detection Loss. To match $N$ detected object candidates $\{\hat{v}_i\}_{i=1}^N$ and $M$ ground truth objects $\{v_i\}_{i=1}^M$ , Carion et al. [1] pad the ground truth objects with $\phi$ (no object) and find the best permutation of predicted objects that minimizes total bipartite matching costs. From the permuted + +predictions $\{\hat{v}_i^{\prime}\}_{i = 1}^{N}$ , the loss is calculated as follows: + +$$ +\mathcal {L} _ {\mathrm {o d}} = \Sigma_ {i = 1} ^ {N} \left[ \lambda_ {c} \mathcal {L} _ {c} \left(\hat {v} _ {i} ^ {\prime c}, v _ {i} ^ {c}\right) + \mathbb {1} _ {v _ {i} ^ {c} \neq \phi} \left(\lambda_ {b} \mathcal {L} _ {b} \left(\hat {v} _ {i} ^ {\prime b}, v _ {i} ^ {b}\right)\right) \right], \tag {2} +$$ + +where $\mathcal{L}_c$ and $\mathcal{L}_b$ denote loss function for category labels and box coordinates, respectively. + +# 3.2. EGTR + +We propose a novel lightweight relation extractor, EGTR, which exploits the self-attention of DETR decoder, as depicted in Fig. 3. Since the self-attention weights in Eq. (1) contain $N \times N$ bidirectional relationships among the $N$ object queries, our relation extractor aims to extract the predicate information from the self-attention weights in the entire $L$ layers, by considering the attention queries and keys as subjects and objects, respectively. + +In order to preserve rich information learned in the self-attention layer, we devise another relation function $f$ between the attention queries and keys instead of the dot product attention in Eq. (1). As concatenation preserves representations completely, we concatenate the representations on all $N \times N$ pairs of attention queries and keys as shown in Fig. 3 (a) to get the relation representations of the $l$ -th layer $R_{a}^{l} \in \mathbb{R}^{N \times N \times 2d_{\mathrm{model}}}$ . Before pairwise concatenation, to help queries and keys play the role of subjects and objects, a linear projection is added as follows: + +$$ +R _ {a} ^ {l} = \left[ Q ^ {l} W _ {S} ^ {l}; K ^ {l} W _ {O} ^ {l} \right], \tag {3} +$$ + +where $Q^{l}\in \mathbb{R}^{N\times d_{\mathrm{model}}}$ and $K^l\in \mathbb{R}^{N\times d_{\mathrm{model}}}$ refer to attention queries and keys of the $l$ -th layer, and $[\cdot ;\cdot ]$ denotes the pairwise concatenation. $W_{S}^{l}$ and $W_{O}^{l}$ are linear weights of shape $\mathbb{R}^{d_{\mathrm{model}}\times d_{\mathrm{model}}}$ . + +We also utilize the last layer representations of the object queries $Z^{L} \in \mathbb{R}^{N \times d_{\mathrm{model}}}$ , which are used for the object + +detection in the same manner: + +$$ +R _ {z} = \left[ Z ^ {L} W _ {S}; Z ^ {L} W _ {O} \right], \tag {4} +$$ + +where $W_{S}, W_{O}$ are linear weights of shape $\mathbb{R}^{d_{\mathrm{model}} \times d_{\mathrm{model}}}$ . To effectively use the various information learned in all layers, we introduce a gating mechanism as follows: + +$$ +g _ {a} ^ {l} = \sigma \left(R _ {a} ^ {l} W _ {G}\right), g _ {z} = \sigma \left(R _ {z} W _ {G}\right), \tag {5} +$$ + +where $g_{a}^{l}$ and $g_{z} \in \mathbb{R}^{N \times N \times 1}$ represent the gate values obtained through a single linear layer $W_{G} \in \mathbb{R}^{2d_{\mathrm{model}} \times 1}$ for $R_{a}^{l}$ and $R_{z}$ , respectively. Finally, we extract the relation graph from the gated summation of the relation representations across all layers as follows: + +$$ +\hat {G} = \sigma \left(\mathrm {M L P} _ {\mathrm {r e l}} \left(\Sigma_ {l = 1} ^ {L} \left(g _ {a} ^ {l} * R _ {a} ^ {l}\right) + g _ {z} * R _ {z}\right)\right), \tag {6} +$$ + +where $\hat{G} \in \mathbb{R}^{N \times N \times |\mathcal{C}_p|}$ denotes predicted relation graph and $\mathrm{MLP}_{\mathrm{rel}}$ is a three-layer perceptron with ReLU activation. Note that we use sigmoid function $\sigma$ so that multiple relationships can exist between objects. + +# 3.3. Learning and Inference + +To train EGTR, we perform multi-task learning. In addition to object detection and relation extraction, we devise connectivity prediction, an auxiliary task for relation extraction. The overall loss for the framework is as follows: + +$$ +\mathcal {L} = \mathcal {L} _ {\mathrm {o d}} + \lambda_ {\mathrm {r e l}} \mathcal {L} _ {\mathrm {r e l}} + \lambda_ {\mathrm {c o n}} \mathcal {L} _ {\mathrm {c o n}}, \tag {7} +$$ + +where $\mathcal{L}_{\mathrm{od}}$ is object detection loss in Eq. (2). $\mathcal{L}_{\mathrm{rel}}$ and $\mathcal{L}_{\mathrm{con}}$ are loss functions for the relation extraction and connectivity prediction, respectively. Since explicit object detection loss is used, EGTR has the capability to detect all nodes in the scene graph. The details of the loss for each task are provided below. + +# 3.3.1 Relation Extraction + +We use binary cross-entropy loss for the relation extraction. To match predicted graph $\hat{G} \in \mathbb{R}^{N \times N \times |\mathcal{C}_p|}$ and ground truth triplet set $\mathcal{E}$ , we first encode $\mathcal{E}$ as a one-hot ground truth graph $G \in \mathbb{R}^{N \times N \times |\mathcal{C}_p|}$ by padding the regions that do not correspond to relations between the ground truth objects as zero. Then, we permute indices of the predicted graph using the permutation found in the object detection. From the permuted graph $\hat{G}'$ , the loss for the relation extraction is calculated as $\mathcal{L}_{\mathrm{rel}} = \mathcal{L}_r(\hat{G}', G)$ . + +However, since the size of the graph is proportional to the square of $N$ , the sparsity of the ground truth graph $G$ is too severe. For instance, the density is only $10^{-14}$ when $N$ is set to 200 for the Visual Genome [12] validation dataset. Therefore, we divide $G$ into three regions: (1) GT region, + +![](images/8b4b9a348b44f666252103d9ddd29526093a28ee47d8308828587aeab359d27e.jpg) +girl - on - horse +jirl - riding + +(subject - predicate - object) +non-matching region +Figure 4. Example of graph $G$ region. The target graph has a shape of $N \times N \times |\mathcal{V}_c|$ . The graph is severely sparse since the number of object queries $N$ is set large enough to cover objects in the image. We split $G$ into GT, negative, and non-matching regions. +![](images/150501b135af7d479b690a4d0f111b17fd90658d9f4ed8d284146a7bbe9e14b7.jpg) +girl - wearing - pant + +(2) negative region, and (3) non-matching region, as shown in Fig. 4. The GT region indicates the ground truth triplets where the value of $G$ is 1. The negative region consists of the triplets in which both subjects and objects are composed of the ground truth objects, but no relation exists between them. The non-matching region represents the zero-padded region. It is paired with a region consisting of the object candidates that do not match the ground truth objects and match $\phi$ in Eq. (2). For each region, we apply different techniques to effectively train the relation extraction with object detection as subsequently described. + +Adaptive Smoothing. We propose a novel adaptive smoothing for the GT region. For $G_{ijk}$ belonging to the GT region, the model is trained to predict the $k$ -th predicate category between subject entity $v_i$ and object entity $v_j$ . However, since $\hat{v}_i'$ and $\hat{v}_j'$ , which match $v_i$ and $v_j$ respectively, do not have enough representations about the corresponding ground truth object at the beginning of the training, it may not be appropriate to predict the probability of the predicate as 1. Moreover, even when the object detection performance is reasonably assured, the detection performance may still vary for individual object candidates. Therefore, we reflect the detection performance of each object candidate on the relation label via adaptive smoothing. + +We first measure the uncertainty of each object candidate with the corresponding bipartite matching cost. For object candidate $\hat{v}_i'$ , we define the uncertainty as follows: + +$$ +u _ {i} = \sigma \left(\operatorname {c o s t} _ {i} - \operatorname {c o s t} _ {\min } + \sigma^ {- 1} (\alpha)\right), \tag {8} +$$ + +where $\mathrm{cost}_i$ denotes the matching cost and $\mathrm{cost}_{\mathrm{min}}$ indicates the matching cost when $\hat{v}_i^\prime$ perfectly matches $v_{i}$ . $\alpha$ is a nonnegative hyperparameter representing the minimum uncertainty. We set the value of $G_{ijk}$ to $(1 - u_{i})(1 - u_{j})$ taking the uncertainty into account. By using the relation label adjusted by the uncertainty, the multi-task learning of object + +detection and relation extraction is dynamically regulated according to the quality of the detected objects. + +Negative and Non-matching Sampling. Rather than employing all negatives, we sample them from the negative region. Inspired by the hard negative mining introduced by Liu et al. [21], we sort all negatives based on the predicted relation score $\hat{G}_{ijk}^{\prime}$ and choose the top $k_{\mathrm{neg}} \times |\mathcal{E}|$ most challenging negatives. Similarly, we extract $k_{\mathrm{non}} \times |\mathcal{E}|$ hard samples from the non-matching region. As the non-matching region typically encompasses a substantial part of the graph $G$ , this approach notably diminishes sparsity. + +# 3.3.2 Connectivity Prediction + +Influenced by Graph-RCNN [40] that predicted a relatedness to prune object pairs, we propose a connectivity prediction that predicts whether at least one edge between two object nodes exists for the relation extraction. We get a connectivity graph $\hat{E} \in \mathbb{R}^{N \times N \times 1}$ in a similar way to get the relation graph in Eq. (6) with the same relation source representations. Instead of $\mathrm{MLP}_{\mathrm{rel}}$ for multi-label prediction, we use another $\mathrm{MLP}_{\mathrm{con}}$ for binary classification. We calculate the binary cross entropy loss from the permutated connectivity graph as follows: $\mathcal{L}_{\mathrm{con}} = \mathcal{L}_r(\hat{E}', E)$ . + +# 3.3.3 Inference + +For model inference, we get triplet scores by multiplying predicate score $\hat{G}_{ijk}$ by the corresponding class scores of $\hat{v}_i^c$ and $\hat{v}_j^c$ . Note that we set $\hat{G}_{iik}$ to 0 to prevent self-connections in which the subject and object are the same entity. In addition, we enhance the triplet scores by utilizing the connectivity score $\hat{E}_{ijk}$ . By multiplying the connectivity score, we can effectively filter out the triplets that are not likely to have a relation between the subject and object. + +# 4. Experiments + +# 4.1. Datasets and Evaluation Settings + +We conduct experiments on two SGG datasets. We describe the datasets and evaluation settings for each dataset. Detailed settings for each dataset are presented in the supplementary materials. + +Visual Genome. Visual Genome [12] is the most representative SGG dataset, consisting of 57K training images, 5K validation images, and 26K test images. We follow the popular Visual Genome split, retaining the most frequent 150 object categories and 50 relation categories. We adopt Scene Graph Detection (SGDet) evaluation settings and report Recall@k (R@k) which is class agnostic and mean Recall@k (mR@k) that aggregates the recalls for each predicate category. Following Motifs [42], these metrics are measures with graph constraint, which means each + +object pair can have a single predicate category. Since they are only related to triplet detection of the scene graph, we report AP50 that evaluates the detection performance of all objects appearing in the scene. Furthermore, we report the number of model parameters and Frames Per Second (FPS) to measure efficiency. + +Open Image V6. Open Image V6 [13] is also widely used dataset, comprising of 126K images for training, 2K for validation, and 5K for test set. It includes 601 object categories and 30 relation categories. For evaluation, we adopt both recall and weighted mean AP (wmAP) following standard settings. For recall evaluation, micro-R@50 is adopted. The WMAP is evaluated with two settings: $\mathrm{wmAP}_{\mathrm{rel}}$ for predicting boxes of subject entity and object entity separately and $\mathrm{wmAP}_{\mathrm{phr}}$ for predicting a union box of them. The final score is calculated by $0.2 \times \text{micro-R} @50 + 0.4 \times \mathrm{wmAP}_{\mathrm{rel}} + 0.4 \times \mathrm{wmAP}_{\mathrm{phr}}$ . + +# 4.2. Implementation Details + +We employ Deformable DETR [46] that improves the convergence speed of the DETR with ResNet-50 [6] as a backbone. It is worth noting that our approach can be extended to any object detector that incorporates self-attention mechanisms between object features, including models like DETR [1], Sparse R-CNN [32], and others. We follow the configurations of the original Deformable DETR, except that we use only 200 object queries. To speed up convergence, we first train the object detector with the target dataset and subsequently train the SGG task using the pretrained object detector, similar to prior work [17]. To calculate the overall loss in Equation 7, $\lambda_{\mathrm{rel}}$ is set to 15. $\lambda_{\mathrm{con}}$ is set to 30 and 90 for the Visual Genome and Open Image V6, respectively. We set $\alpha$ as $10^{-14}$ for adaptive smoothing and both $k_{\mathrm{neg}}$ and $k_{\mathrm{non}}$ are set to 80. + +# 4.3. Results + +We present quantitative results and perform a comparative analysis of our proposed framework with representative SGG models. Additional experimental results are included in the supplementary materials. + +Visual Genome. Visual Genome results are shown in Tab. 1. Our proposed method demonstrates competitive performance with the current one-stage SGG models that have 1.5 to 6.5 times larger parameters with the fastest inference speed. In particular, our method achieves the highest object detection performance and shows comparable performance in the triplet detection compared to SSR-CNN [35], the state-of-the-art method. The results demonstrate that our method can generate scene graphs in an efficient and effective way by exploiting the by-products of the object detector. By applying logit adjustment [35], a technique for predicting tail predicate classes well, our + +
Model# params (M)FPSAP50R@20R@50R@100mR@20mR@50mR@100
two-stageIMP (EBM) [31, 39]322.22.028.118.125.931.22.84.25.4
VTransE [44]312.33.5-24.531.335.55.16.88.0
Motifs [42]369.91.928.125.132.136.94.15.56.8
VCTree [33]361.50.828.124.831.836.14.96.67.7
VCTree (TDE) [33, 34]361.30.828.114.019.423.26.99.311.1
VCTree (EBM) [31, 33]372.5-28.124.231.435.95.77.79.1
GPS-Net [19]----31.135.9-6.78.6
BGNN [16]341.91.729.023.331.035.87.510.712.6
one-stageFCSGG [20]87.16.028.516.121.325.12.73.64.2
RelTR [7]63.713.426.421.227.5-6.810.8-
SGTR [17]117.16.225.4-24.628.4-12.015.2
Relationformer [29]92.98.526.322.228.431.34.69.310.7
Iterative SGG [9]93.56.027.7†-29.732.1-8.08.8
SSR-CNN [35]274.34.023.8†25.832.736.96.18.410.0
SSR-CNN [35] LA,τ=0.3274.34.023.8†18.423.326.513.517.921.4
EGTR (Ours)42.514.730.823.530.234.35.57.910.1
EGTR (Ours) LA,τ=0.742.514.730.815.718.720.512.117.821.7
EGTR (Ours) LA,τ=0.542.514.730.819.724.226.711.017.121.4
EGTR (Ours) LA,τ=0.342.514.730.822.428.231.78.814.018.3
+ +Table 1. Graph-Constraint results on the test set of Visual Genome dataset. We report the results of representative two-stage SGG models based on the Faster R-CNN [27] object detector with a ResNeXt-101-FPN [18, 38] backbone and concurrent one-stage SGG models. Among the one-stage models, the best is highlighted in bold, and the second-best is indicated with underlining. LA denotes logit adjustment proposed in SSR-CNN [35]. $\dagger$ represents that we concatenate the predicted subjects and objects and then apply non-maximum-suppression with a threshold of 0.3. Italic indicates the evaluation of metrics using publicly available model checkpoints. FPS is measured with a single V100 for images resized to a minimum of 600 for the shortest side and a maximum of 1000 for the longest side. + +
Modelscoremicro-R@50wmAPrelwmAPphr
Motifs [42]38.971.629.931.6
VCTree [33]40.274.134.233.1
GPS-Net [19]41.774.832.934.0
BGNN [16]42.175.033.534.2
RelTR [7]43.071.734.237.5
SGTR [17]42.359.937.038.7
SSR-CNN [35]49.476.741.543.6
EGTR (Ours)48.675.042.041.9
+ +Table 2. Results on test set of Open Image V6. The score is a weighted sum of micro-R@50, $\mathrm{wmAP}_{\mathrm{rel}}$ , and $\mathrm{wmAP}_{\mathrm{phr}}$ . + +method performs favorably in the trade-off between $\mathrm{R}@\mathbf{k}$ and $\mathrm{mR}@\mathbf{k}$ , showing superiority over existing state-of-the-art models. Note that for triplet detection models [9, 35] without an explicit object detector, AP50 is measured by applying non-maximum-suppression (NMS) to the union of predicted subjects and objects set. AP50 performance is discussed further in the section 4.5. + +Open Image V6. As depicted in Tab. 2, the experiments conducted on the Open Image V6 dataset also demonstrate competitive performance, underscoring the effectiveness and robustness of our method across different datasets. + +# 4.4. Ablation Studies + +We analyze the effects of our model components. All of the ablation studies are done with the Visual Genome dataset. + +Relation Sources. In Tab. 3, we investigate the source of + +relations used for the relation extractor. To verify the benefits of using attention queries and keys, we change $Q^l$ and $K^l$ in Eq. (3) to $Z^l$ , the hidden states of each layer. Remarkably, when simply using the hidden states of all layers, the triplet detection performance decreases. Surprisingly, the performance is lower than that of using only $R_z$ from the last layer. On the other hand, when only attention by-products are used as a relation source, the performance is similar to that of using only hidden states of the last layer, the most contextualized representations for detecting objects. Incorporating both attention by-products and final hidden states yields improved performance by amalgamating varied information. The results underscore that the attention by-products of the object detectors contain rich information for relation extraction. + +Training Techniques. In Tab. 4, we ablate proposed techniques. Without any techniques, our model framework shows low performance. When each technique is applied, it substantially improves both R@50 and mR@50, and the best performance is achieved when all techniques are combined. The results demonstrate that our proposed techniques help to train the model framework effectively. Additional experimental results are provided in the supplementary materials. + +Sampling Methodology. In Tab. 5, we investigate the impact of hard sampling methods for the negative and non-matching regions. Hard negatives and hard non-matchings show a trade-off that improves the performance of R@50 + +
Rl^l a sourceRl^l aRzR@50mR@50
Ql & Kl✓✓30.27.9
Zl✓✓29.67.4
-✓29.97.6
Ql & Kl✓29.87.7
+ +Table 3. Ablation study on relation source. + +
adaptive smoothing\( \mathcal{L}_{\text{con}} \)samplingR@50mR@50
26.65.3
✓28.36.5
✓29.67.0
✓28.97.1
✓✓✓30.27.9
+ +Table 4. Ablation study on proposed techniques. + +
hard negativehard non-matchingR@50mR@50
✓✓30.27.9
✓30.07.1
✓29.67.7
+ +and mR@50, respectively, and additional performance improvements are achieved for both measures when they are used together. For hard negatives, we speculate that R@50 performance is improved by selecting negatives for the tail predicate class, which is likely to exist in reality but is not in the annotation. For hard non-matchings, mR@50 performance seems to have improved by choosing object candidates that are akin to the ground truth objects but unmatched with them. This encourages the model to predict diverse predicate classes by preventing duplicate object pairs, which are likely to predict head predicate classes, from being predicted. + +# 4.5. Discussion + +Regarding the highest AP50 score, we posit that it is because our proposed method focuses on not only objects with relations but also objects without any relations. To validate this, we calculate AP50 for two subsets of the ground-truth objects: those with relations and those without relations. We compare our method with two triplet detection-based models. Tab. 6 illustrates that our method achieves significantly high $\mathrm{AP50}_{\mathrm{no - rel}}$ . Considering that triplet detection performance is related to $\mathrm{AP50}_{\mathrm{rel}}$ , while triplet detection-based models only focus on objects with relations to detect triplets well, our method is capable of extracting the complete scene graph, including objects without relations. + +For further analysis, we visualize bounding boxes for subjects and objects from the top 20 predictions in Fig. 5. For SSR-CNN [35], which directly predicts triplets, we observe a significant number of overlapping bounding boxes. On the other hand, our proposed model leverages predic + +Table 5. Ablation study on sampling methods. + +
ModelAP50\( AP50_{\text{rel}} \)\( AP50_{\text{no-rel}} \)
Iterative SGG [9]†27.724.37.8
SSR-CNN [35]†23.820.27.4
EGTR (Ours)30.824.310.7
+ +Table 6. AP50 for two subsets of objects. $\mathrm{AP50}_{\mathrm{rel}}$ and $\mathrm{AP50}_{\mathrm{no - rel}}$ denote scores evaluated only on objects having at least one edge and no edge, respectively. Note that each measure operates on its unique scale since the ground-truth objects used for each measure differ. $\dagger$ indicates that additional NMS is applied to the union of subjects and objects set. + +![](images/24bd9ff20bfc213d9d6ec0c5211af0c429e35efa9df27fb0358e28fccd39745f.jpg) + +![](images/18ae33dd39cb2090a7c41cafda42b6943398cbcb2bce6166315195b836fe3f0f.jpg) +(a) SSR-CNN [35] + +![](images/46306f91c2dbb914d9f6359cae5848fe4882b7587c39fc49c4b75c459bcd9dfb.jpg) + +![](images/bcd320527c7035d3e2f40585be25959ccfdf58231515632ecd35636a8f686e10.jpg) +(b) EGTR (Ours) +Figure 5. Comparison of detected subjects and objects. We select the top 20 predictions of SSR-CNN and EGTR and visualize bounding boxes of detected subjects and objects. + +tions from an explicit object detector, reducing redundant predictions. Furthermore, our method detects various subjects and objects that appear in the scene rather than focusing on objects that are more likely to have relations. + +# 5. Conclusion + +In this study, we propose the lightweight one-stage scene graph generator EGTR. We significantly reduce the model complexity by harnessing the relationships among the object queries learned from the self-attention layers of the object decoder. Furthermore, we devise a novel adaptive smoothing technique that helps multi-task learning of object detection and relation extraction by adjusting the relation label according to the object detection performance. As an auxiliary task of relation extraction, connectivity prediction contributes to the effective learning of EGTR. We conduct extensive experiments and ablation studies, and the results demonstrate that EGTR achieves the highest object detection performance and competitive triplet detection capabilities with the fastest inference speed. + +Acknowledgements. We thank Taeho Kil, Geewook Kim, and the anonymous reviewers for their insightful comments and suggestions. + +# References + +[1] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In ECCV, pages 213-229, 2020. 1, 2, 3, 4, 6 +[2] Zhe Chen, Shaoli Huang, and Dacheng Tao. Context refinement for object detection. In ECCV, pages 71-86, 2018. 2 +[3] Yuren Cong, Michael Ying Yang, and Bodo Rosenhahn. Reltr: Relation transformer for scene graph generation. IEEE TPAMI, 2023. 1, 3 +[4] Naina Dhingra, Florian Ritter, and Andreas Kunz. Bgt-net: Bidirectional gru transformer network for scene graph generation. In CVPR, pages 2150-2159, 2021. 2 +[5] Lizhao Gao, Bo Wang, and Wenmin Wang. Image captioning with scene-graph based semantic concepts. In ICMLC, pages 225-229, 2018. 1 +[6] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, pages 770-778, 2016. 6 +[7] Han Hu, Jiayuan Gu, Zheng Zhang, Jifeng Dai, and Yichen Wei. Relation networks for object detection. In CVPR, pages 3588-3597, 2018. 2, 7 +[8] Justin Johnson, Ranjay Krishna, Michael Stark, Li-Jia Li, David Shamma, Michael Bernstein, and Li Fei-Fei. Image retrieval using scene graphs. In CVPR, pages 3668-3678, 2015. 1 +[9] Siddhesh Khandelwal and Leonid Sigal. Iterative scene graph generation. In NeurIPS, 2022. 1, 3, 7, 8 +[10] Dong-Jin Kim, Jinsoo Choi, Tae-Hyun Oh, and In So Kweon. Dense relational captioning: Triple-stream networks for relationship-based captioning. In CVPR, pages 6271-6280, 2019. 1 +[11] Rajat Koner, Suprosanna Shit, and Volker Tresp. Relation transformer network. ECCV, pages 422-439, 2022. 2 +[12] Ranjay Krishna, Yuke Zhu, Oliver Groth, Justin Johnson, Kenji Hata, Joshua Kravitz, Stephanie Chen, Yannis Kalantidis, Li-Jia Li, David A Shamma, et al. Visual genome: Connecting language and vision using crowdsourced dense image annotations. IJCV, 123:32-73, 2017. 2, 5, 6 +[13] Alina Kuznetsova, Hassan Rom, Neil Alldrin, Jasper Uijlings, Ivan Krasin, Jordi Pont-Tuset, Shahab Kamali, Stefan Popov, Matteo Malloci, Alexander Kolesnikov, et al. The open images dataset v4: Unified image classification, object detection, and visual relationship detection at scale. IJCV, 128(7):1956-1981, 2020. 2, 6 +[14] Linjie Li, Zhe Gan, Yu Cheng, and Jingjing Liu. Relation-aware graph attention network for visual question answering. In ICCV, pages 10313-10322, 2019. 1 +[15] Lin Li, Guikun Chen, Jun Xiao, Yi Yang, Chunping Wang, and Long Chen. Compositional feature augmentation for unbiased scene graph generation. In ICCV, pages 21685-21695, 2023. 2 +[16] Rongjie Li, Songyang Zhang, Bo Wan, and Xuming He. Bipartite graph network with adaptive message passing for unbiased scene graph generation. In CVPR, pages 11109-11119, 2021. 1, 2, 7 + +[17] Rongjie Li, Songyang Zhang, and Xuming He. Sgtr: End-to-end scene graph generation with transformer. In CVPR, pages 19486-19496, 2022. 1, 3, 6, 7 +[18] Tsung-Yi Lin, Piotr Dollar, Ross Girshick, Kaiming He, Bharath Hariharan, and Serge Belongie. Feature pyramid networks for object detection. In CVPR, pages 2117-2125, 2017. 7 +[19] Xin Lin, Changxing Ding, Jinquan Zeng, and Dacheng Tao. Gps-net: Graph property sensing network for scene graph generation. In CVPR, pages 3746-3753, 2020. 1, 2, 7 +[20] Hengyue Liu, Ning Yan, Masood Mortazavi, and Bir Bhanu. Fully convolutional scene graph generation. In CVPR, pages 11546-11556, 2021. 1, 3, 7 +[21] Wei Liu, Dragomir Anguelov, Dumitru Erhan, Christian Szegedy, Scott Reed, Cheng-Yang Fu, and Alexander C Berg. Ssd: Single shot multibox detector. In ECCV, pages 21-37, 2016. 6 +[22] Yong Liu, Ruiping Wang, Shiguang Shan, and Xilin Chen. Structure inference net: Object detection using scene-level context and instance-level relationships. In CVPR, pages 6985-6994, 2018. 2 +[23] Cewu Lu, Ranjay Krishna, Michael Bernstein, and Li Fei-Fei. Visual relationship detection with language priors. In ECCV, pages 852–869, 2016. 2 +[24] Yichao Lu, Himanshu Rai, Jason Chang, Boris Knyazev, Guangwei Yu, Shashank Shekhar, Graham W Taylor, and Maksims Volkovs. Context-aware scene graph generation with seq2seq transformers. In ICCV, pages 15931-15941, 2021. +[25] Yukuan Min, Among Wu, and Cheng Deng. Environment-invariant curriculum relation learning for fine-grained scene graph generation. In ICCV, pages 13296-13307, 2023. 2 +[26] Alejandro Newell and Jia Deng. Pixels to graphs by associative embedding. NeurIPS, 30, 2017. 1, 3 +[27] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. NeurIPS, 28, 2015. 2, 7 +[28] Sebastian Schuster, Ranjay Krishna, Angel Chang, Li Fei-Fei, and Christopher D Manning. Generating semantically precise scene graphs from textual descriptions for improved image retrieval. In Proceedings of the fourth workshop on vision and language, pages 70–80. Association for Computational Linguistics, 2015. 1 +[29] Suprosanna Shit, Rajat Koner, Bastian Wittmann, Johannes Paetzold, Ivan Ezhov, Hongwei Li, Jiazhen Pan, Sahand Sharifzadeh, Georgios Kaissis, Volker Tresp, et al. Relationformer: A unified framework for image-to-graph generation. In ECCV, pages 422-439, 2022. 1, 3, 7 +[30] Gopika Sudhakaran, Devendra Singh Dhami, Kristian Kersting, and Stefan Roth. Vision relation transformer for unbiased scene graph generation. In ICCV, pages 21882-21893, 2023. 2 +[31] Mohammed Suhail, Abhay Mittal, Behjat Siddiquie, Chris Broaddus, Jayan Eledath, Gerard Medioni, and Leonid Sigal. Energy-based learning for scene graph generation. In CVPR, pages 13936-13945, 2021. 7 + +[32] Peize Sun, Rufeng Zhang, Yi Jiang, Tao Kong, Chenfeng Xu, Wei Zhan, Masayoshi Tomizuka, Lei Li, Zehuan Yuan, Changhu Wang, and Ping Luo. Sparse r-cnn: End-to-end object detection with learnable proposals. In CVPR, pages 14454-14463, 2021. 2, 3, 6 +[33] Kaihua Tang, Hanwang Zhang, Baoyuan Wu, Wenhan Luo, and Wei Liu. Learning to compose dynamic tree structures for visual contexts. In CVPR, pages 6619-6628, 2019. 1, 2, 7 +[34] Kaihua Tang, Yulei Niu, Jianqiang Huang, Jiaxin Shi, and Hanwang Zhang. Unbiased scene graph generation from biased training. In CVPR, pages 3716-3725, 2020. 7 +[35] Yao Teng and Limin Wang. Structured sparse r-cnn for direct scene graph generation. In CVPR, pages 19437-19446, 2022. 1, 3, 6, 7, 8 +[36] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. NeurIPS, 30, 2017. 2, 3 +[37] Sanghyun Woo, Dahun Kim, Donghyeon Cho, and In So Kweon. Linknet: Relational embedding for scene graph. NeurIPS, 31, 2018. 2 +[38] Saining Xie, Ross Girshick, Piotr Dollar, Zhuowen Tu, and Kaiming He. Aggregated residual transformations for deep neural networks. In CVPR, pages 1492-1500, 2017. 7 +[39] Danfei Xu, Yuke Zhu, Christopher B Choy, and Li Fei-Fei. Scene graph generation by iterative message passing. In CVPR, pages 5410–5419, 2017. 1, 2, 7 +[40] Jianwei Yang, Jiasen Lu, Stefan Lee, Dhruv Batra, and Devi Parikh. Graph r-cnn for scene graph generation. In ECCV, pages 670-685, 2018. 6 +[41] Xu Yang, Kaihua Tang, Hanwang Zhang, and Jianfei Cai. Auto-encoding scene graphs for image captioning. In CVPR, pages 10685–10694, 2019. 1 +[42] Rowan Zellers, Mark Yatskar, Sam Thomson, and Yejin Choi. Neural motifs: Scene graph parsing with global context. In CVPR, pages 5831-5840, 2018. 1, 2, 6, 7 +[43] Cheng Zhang, Wei-Lun Chao, and Dong Xuan. An empirical study on leveraging scene graphs for visual question answering. In BMVC, 2019. 1 +[44] Hanwang Zhang, Zawlin Kyaw, Shih-Fu Chang, and Tat-Seng Chua. Visual translation embedding network for visual relation detection. In CVPR, page 3107-3115, 2017. 7 +[45] Chaofan Zheng, Xinyu Lyu, Yuyu Guo, Pengpeng Zeng, Jingkuan Song, and Lianli Gao. Learning to generate scene graph from head to tail. In ICME, pages 1-6. IEEE, 2022. 2 +[46] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable detr: Deformable transformers for end-to-end object detection. In ICLR, 2021. 2, 3, 6 \ No newline at end of file diff --git a/2024/EGTR_ Extracting Graph from Transformer for Scene Graph Generation/images.zip b/2024/EGTR_ Extracting Graph from Transformer for Scene Graph Generation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..b0a3e875333d07a724f12e3f9b309cb9f9e37416 --- /dev/null +++ b/2024/EGTR_ Extracting Graph from Transformer for Scene Graph Generation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:dbb00bc197d6b92493feb5957d82b027f749dddc3f93b2bf8cd49b8f8b50cbcb +size 510266 diff --git a/2024/EGTR_ Extracting Graph from Transformer for Scene Graph Generation/layout.json b/2024/EGTR_ Extracting Graph from Transformer for Scene Graph Generation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..f45997776cde64fc5aba59ae984c84dd8285532b --- /dev/null +++ b/2024/EGTR_ Extracting Graph from Transformer for Scene Graph Generation/layout.json @@ -0,0 +1,10206 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 69, + 103, + 523, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 103, + 523, + 121 + ], + "spans": [ + { + "bbox": [ + 69, + 103, + 523, + 121 + ], + "type": "text", + "content": "EGTR: Extracting Graph from Transformer for Scene Graph Generation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 75, + 142, + 516, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 142, + 516, + 190 + ], + "spans": [ + { + "bbox": [ + 75, + 142, + 516, + 190 + ], + "type": "text", + "content": "Jinbae Im1 JeongYeon Nam1 Nokyung Park1,2,3* Hyungmin Lee2 Seunghyun Park1 \n1NAVER Cloud AI 2NAVER 3Korea University \n{jinbae.im, jy.nam, nokyung.park99, hyungmin.lee, seung.park}@navercorp.com" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "spans": [ + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 239, + 290, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 239, + 290, + 574 + ], + "spans": [ + { + "bbox": [ + 46, + 239, + 290, + 574 + ], + "type": "text", + "content": "Scene Graph Generation (SGG) is a challenging task of detecting objects and predicting relationships between objects. After DETR was developed, one-stage SGG models based on a one-stage object detector have been actively studied. However, complex modeling is used to predict the relationship between objects, and the inherent relationship between object queries learned in the multi-head self-attention of the object detector has been neglected. We propose a lightweight one-stage SGG model that extracts the relation graph from the various relationships learned in the multi-head self-attention layers of the DETR decoder. By fully utilizing the self-attention by-products, the relation graph can be extracted effectively with a shallow relation extraction head. Considering the dependency of the relation extraction task on the object detection task, we propose a novel relation smoothing technique that adjusts the relation label adaptively according to the quality of the detected objects. By the relation smoothing, the model is trained according to the continuous curriculum that focuses on object detection task at the beginning of training and performs multi-task learning as the object detection performance gradually improves. Furthermore, we propose a connectivity prediction task that predicts whether a relation exists between object pairs as an auxiliary task of the relation extraction. We demonstrate the effectiveness and efficiency of our method for the Visual Genome and Open Image V6 datasets. Our code is publicly available at https://github.com/naver-ai/egtr." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 599, + 128, + 611 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 599, + 128, + 611 + ], + "spans": [ + { + "bbox": [ + 47, + 599, + 128, + 611 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 620, + 287, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 620, + 287, + 693 + ], + "spans": [ + { + "bbox": [ + 46, + 620, + 287, + 693 + ], + "type": "text", + "content": "Scene Graph Generation (SGG) [8] aims to generate a scene graph that represents objects as nodes and relationships between objects as edges from an image as shown in Fig. 1b. SGG is a challenging task as it is required to go beyond simply detecting objects but predicting the relationships between them based on a comprehensive understanding of the" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 310, + 215, + 457, + 313 + ], + "blocks": [ + { + "bbox": [ + 310, + 215, + 457, + 313 + ], + "lines": [ + { + "bbox": [ + 310, + 215, + 457, + 313 + ], + "spans": [ + { + "bbox": [ + 310, + 215, + 457, + 313 + ], + "type": "image", + "image_path": "da086494a3493dff7e071faf8fea070928900aae618d983c549510da6ad11f8c.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 337, + 317, + 429, + 327 + ], + "lines": [ + { + "bbox": [ + 337, + 317, + 429, + 327 + ], + "spans": [ + { + "bbox": [ + 337, + 317, + 429, + 327 + ], + "type": "text", + "content": "(a) Visual Genome example" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 468, + 213, + 541, + 256 + ], + "blocks": [ + { + "bbox": [ + 468, + 213, + 541, + 256 + ], + "lines": [ + { + "bbox": [ + 468, + 213, + 541, + 256 + ], + "spans": [ + { + "bbox": [ + 468, + 213, + 541, + 256 + ], + "type": "image", + "image_path": "9fcc74a969bc741c22cf5e31e09c3ed6a0b106a111765290803f5a92ea5dbd50.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 478, + 258, + 533, + 267 + ], + "lines": [ + { + "bbox": [ + 478, + 258, + 533, + 267 + ], + "spans": [ + { + "bbox": [ + 478, + 258, + 533, + 267 + ], + "type": "text", + "content": "(b) Scene Graph" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 470, + 272, + 541, + 315 + ], + "blocks": [ + { + "bbox": [ + 470, + 272, + 541, + 315 + ], + "lines": [ + { + "bbox": [ + 470, + 272, + 541, + 315 + ], + "spans": [ + { + "bbox": [ + 470, + 272, + 541, + 315 + ], + "type": "image", + "image_path": "d3d669486085177c87b65b065112e8bb0b5abe8f60a9d5e1549828477fc1cbc7.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 473, + 318, + 538, + 327 + ], + "lines": [ + { + "bbox": [ + 473, + 318, + 538, + 327 + ], + "spans": [ + { + "bbox": [ + 473, + 318, + 538, + 327 + ], + "type": "text", + "content": "(c) Attention Graph" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 304, + 332, + 546, + 410 + ], + "lines": [ + { + "bbox": [ + 304, + 332, + 546, + 410 + ], + "spans": [ + { + "bbox": [ + 304, + 332, + 546, + 410 + ], + "type": "text", + "content": "Figure 1. Motivation. SGG task aims to predict scene graph (Fig. 1b) with objects as nodes and relations as edges. We draw a plausible attention graph (Fig. 1c) simply by connecting objects with high attention weights to edges from the self-attention layers of the pre-trained DETR. It shows the potential for the self-attention from the object detector to contain rich information that aids in predicting the relations of the scene graph." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 424, + 546, + 484 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 424, + 546, + 484 + ], + "spans": [ + { + "bbox": [ + 304, + 424, + 546, + 484 + ], + "type": "text", + "content": "scene. Since a scene graph provides structural information of the image, it can be used for various vision tasks that require a higher level of understanding and reasoning about images such as image captioning [5, 10, 41], image retrieval [8, 28], and visual question answering [14, 43]." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 486, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 486, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 486, + 547, + 713 + ], + "type": "text", + "content": "Most previous studies [16, 19, 33, 39, 42] took two-stage SGG approaches that detect objects first and then predict their relations. However, these approaches incurred high computational costs and risks of error propagation from the object detection. To address the drawbacks of the two-stage approaches, one-stage SGG models that perform object detection and relation prediction at once [3, 9, 17, 20, 26, 29, 35] have been studied recently, leveraging one-stage object detectors such as DETR [1]. Since edges in a scene graph can be represented as subject-predicate-object triplets, many studies embraced triplet-based approaches as shown in Figs. 2a and 2b. However, object-triplet detection models [3, 17] required sophisticated triplet detectors to obtain necessary information for triplet queries from the object detector. Moreover, triplet detection models [9, 35] lacked the ability to detect objects without relation such as \"tree\" in Fig. 1b, by focusing solely on the triplet detection without an object detector. Considering that objects without relations account for more than" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 702, + 271, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 702, + 271, + 712 + ], + "spans": [ + { + "bbox": [ + 58, + 702, + 271, + 712 + ], + "type": "text", + "content": "* Most work was done during the internship at Naver Cloud AI." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "24229" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 66, + 71, + 214, + 163 + ], + "blocks": [ + { + "bbox": [ + 66, + 71, + 214, + 163 + ], + "lines": [ + { + "bbox": [ + 66, + 71, + 214, + 163 + ], + "spans": [ + { + "bbox": [ + 66, + 71, + 214, + 163 + ], + "type": "image", + "image_path": "eeef0756f110b56051eef1ee222a4df8ecbb04eaa0125bb5bc46ca221ac47ed5.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 81, + 167, + 198, + 176 + ], + "lines": [ + { + "bbox": [ + 81, + 167, + 198, + 176 + ], + "spans": [ + { + "bbox": [ + 81, + 167, + 198, + 176 + ], + "type": "text", + "content": "(a) Object-Triplet Detection Models" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 236, + 72, + 350, + 163 + ], + "blocks": [ + { + "bbox": [ + 236, + 72, + 350, + 163 + ], + "lines": [ + { + "bbox": [ + 236, + 72, + 350, + 163 + ], + "spans": [ + { + "bbox": [ + 236, + 72, + 350, + 163 + ], + "type": "image", + "image_path": "17fbf1da620e9dc2c721e805545b7c7f43c92cb1db5cdc6e1494d2fbdc0930f5.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 249, + 167, + 342, + 176 + ], + "lines": [ + { + "bbox": [ + 249, + 167, + 342, + 176 + ], + "spans": [ + { + "bbox": [ + 249, + 167, + 342, + 176 + ], + "type": "text", + "content": "(b) Triplet Detection Models" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 180, + 547, + 236 + ], + "lines": [ + { + "bbox": [ + 46, + 180, + 547, + 236 + ], + "spans": [ + { + "bbox": [ + 46, + 180, + 547, + 236 + ], + "type": "text", + "content": "Figure 2. Comparison with existing one-stage SGG models. (a) Object-Triplet Detection Models introduce additional triplet queries and a triplet detector to the object detector. The triplet detector requires additional modules to incorporate information from the object detector into the triplet queries. (b) Triplet Detection Models focus on detecting triplets directly without an object detector. Objects without relations may not be detected. (c) Relation Extraction Models extract relations from the object detector without a separate triplet detector. In particular, ours extracts relations more effectively by utilizing by-products from the self-attention of the object detector." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 377, + 72, + 526, + 163 + ], + "blocks": [ + { + "bbox": [ + 377, + 72, + 526, + 163 + ], + "lines": [ + { + "bbox": [ + 377, + 72, + 526, + 163 + ], + "spans": [ + { + "bbox": [ + 377, + 72, + 526, + 163 + ], + "type": "image", + "image_path": "1200dc7f83785412b59a2e3649fd9623129020a4a233efc8ba94d41c3b085407.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 401, + 167, + 502, + 175 + ], + "lines": [ + { + "bbox": [ + 401, + 167, + 502, + 175 + ], + "spans": [ + { + "bbox": [ + 401, + 167, + 502, + 175 + ], + "type": "text", + "content": "(c) Relation Extraction Models" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 245, + 286, + 269 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 245, + 286, + 269 + ], + "spans": [ + { + "bbox": [ + 46, + 245, + 286, + 269 + ], + "type": "inline_equation", + "content": "42\\%" + }, + { + "bbox": [ + 46, + 245, + 286, + 269 + ], + "type": "text", + "content": " in Visual Genome [12] data, their priority lies in detecting a sub-graph rather than the complete scene graph." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 45, + 270, + 287, + 520 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 270, + 287, + 520 + ], + "spans": [ + { + "bbox": [ + 45, + 270, + 287, + 520 + ], + "type": "text", + "content": "To address the shortcomings of the existing one-stage SGG models, we focus on the relationships between objects inherent in the object detector. As shown in Fig. 1a, objects are related to each other. For instance, when a horse appears in a scene, a person is likely to appear in the scene, and clothing such as hats, jackets, and pants often depicts the person's current situation. From this intuition, there has been a long belief [2, 7, 22] that modeling relationship or context between objects would be beneficial to object detection task. Accordingly, recent one-stage object detectors [1, 32, 46] have incorporated self-attention layers [36] to implicitly model the relationships among the object queries. We hypothesize that self-attention between object queries learned in a one-stage object detector might contain valuable information for predicting triplet outputs. In our preliminary investigation, we are able to extract a plausible attention graph by simply connecting two object queries with high attention weights from the pre-trained DETR, as shown in Fig. 1c. It shows the potential that the attention weights between object queries can be interpreted as relations between them." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 521, + 287, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 521, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 46, + 521, + 287, + 689 + ], + "type": "text", + "content": "From the findings, we propose a lightweight one-stage scene graph generator EGTR, which stands for Extracting Graph from TRansformer. We design the model to comprehensively leverage the by-products of the object detector, eliminating the need for a separate triplet detector, as depicted in Fig. 2c. From the multi-head self-attention layers of the object detector, we regard an attention query and key, where their relations are learned in the attention weights, as the subject entity and object entity, respectively. Subsequently, we leverage a shallow classifier to predict relations between them. Due to the abundant information about the relationships among the objects present in the by-products derived from all layers of self-attention, we can effectively extract the scene graph." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 689, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 689, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 689, + 287, + 713 + ], + "type": "text", + "content": "Since the relation extraction task is dependent on the object detection task, we speculate that performing relation ex" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 245, + 545, + 388 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 245, + 545, + 388 + ], + "spans": [ + { + "bbox": [ + 304, + 245, + 545, + 388 + ], + "type": "text", + "content": "traction without sufficiently learned representations of the object queries might be harmful. Therefore, we devise a novel adaptive smoothing technique that smoothes the value of the ground truth relation label based on the object detection performance. With the adaptive smoothing, the model is trained with a continuous curriculum that initially focuses on object detection and gradually performs multi-task learning. Furthermore, we propose a connectivity prediction task as an auxiliary task for relation extraction, aiming to predict the existence of any relationship between a subject entity and an object entity. This auxiliary task facilitates the acquisition of representations for the relation extraction." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 390, + 545, + 473 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 390, + 545, + 473 + ], + "spans": [ + { + "bbox": [ + 304, + 390, + 545, + 473 + ], + "type": "text", + "content": "To verify the effectiveness of the proposed method, we conduct experiments on two representative SGG datasets: Visual Genome [12] and Open Images V6 [13]. By actively utilizing the by-products of the object detector, EGTR shows the best object detection performance and comparable triplet detection performance, with the fewest parameters and the fastest inference speed." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 317, + 475, + 539, + 485 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 475, + 539, + 485 + ], + "spans": [ + { + "bbox": [ + 317, + 475, + 539, + 485 + ], + "type": "text", + "content": "Our main contributions can be summarized as follows:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 487, + 547, + 607 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 306, + 487, + 545, + 522 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 487, + 545, + 522 + ], + "spans": [ + { + "bbox": [ + 306, + 487, + 545, + 522 + ], + "type": "text", + "content": "- We propose EGTR that generates scene graphs efficiently and effectively by utilizing the multi-head self-attention by-products derived from the object detector." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 523, + 545, + 570 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 523, + 545, + 570 + ], + "spans": [ + { + "bbox": [ + 306, + 523, + 545, + 570 + ], + "type": "text", + "content": "- We present adaptive smoothing, enabling effective multitask learning for both object detection and relation extraction. In addition, the proposed connectivity prediction offers clues to the relation extraction." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 571, + 547, + 607 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 571, + 547, + 607 + ], + "spans": [ + { + "bbox": [ + 306, + 571, + 547, + 607 + ], + "type": "text", + "content": "- Our comprehensive experiments show the superiority of the proposed model framework and the effectiveness of the devised training techniques." + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 620, + 392, + 633 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 620, + 392, + 633 + ], + "spans": [ + { + "bbox": [ + 306, + 620, + 392, + 633 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "type": "text", + "content": "SGG models can be categorized into two groups: two-stage models and one-stage models. For two-stage models [4, 11, 15, 16, 19, 23-25, 30, 33, 37, 39, 42, 45], separate object detection model and relation prediction model are trained sequentially. They usually detect " + }, + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "type": "text", + "content": " objects from the off-the-shelf object detector such as Faster R-CNN [27]," + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "24230" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 288, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 288, + 156 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 288, + 156 + ], + "type": "text", + "content": "and then all possible combinations of the detected objects are fed into the relation prediction model to predict the relations between each object pair. Although they showed high relation extraction performance, they have the inherent limitation that the object detector that helps with relation extraction is trained separately, resulting in a significant increase in model complexity." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 159, + 288, + 304 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 159, + 288, + 304 + ], + "spans": [ + { + "bbox": [ + 46, + 159, + 288, + 304 + ], + "type": "text", + "content": "As for the one-stage models [3, 9, 17, 20, 26, 29, 35], the object detection and relation prediction are trained in an end-to-end manner. Early studies [20, 26] proposed fully convolutional SGG models and took pixel-based approaches. After DETR [1] brought a huge success as a Transformer [36]-based one-stage object detector, many one-stage SGG studies are based on one-stage object detectors [1, 32, 46]. They efficiently modeled SGG by introducing object queries or triplet queries. We categorize them into three distinct groups: (a) object-triplet detection models, (b) triplet prediction models, and (c) relation extraction models, as shown in Fig. 2." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 311, + 287, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 311, + 287, + 502 + ], + "spans": [ + { + "bbox": [ + 46, + 311, + 287, + 502 + ], + "type": "text", + "content": "Object-Triplet Detection Models. Object-triplet detection models are characterized by introducing additional triplet queries and building a triplet predictor on top of the object detector as shown in Fig. 2a. RelTR [3] introduced paired subject queries and object queries, and SGTR [17] introduced compositional queries decoupled into subjects, objects, and predicates. As the introduced queries are initialized without prior cues from the object detector, the triplet predictor requires modules to incorporate information from the outputs of the object detector and modules for triplet queries to exchange information with each other. It led to the triplet predictor having an intricate structure. In contrast, we refrain from introducing additional queries and regard the attention queries and attention keys, whose relationships are learned in the object detector, as subject and object queries, respectively." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 510, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 510, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 510, + 288, + 714 + ], + "type": "text", + "content": "Triplet Detection Models. Triplet detection models directly detect triplets using triplet queries without an object detector, as shown in Fig. 2b. Iterative SGG [9] introduced subject, object, and predicate queries and modeled the conditional dependencies among them with separate subject, object, and predicate multi-layer Transformer decoders. Inspired by Sparse R-CNN [32], Structured Sparse R-CNN (SSR-CNN) [35] designed triplet queries consisting of subject box, object box, subject, object, and predicate queries. Since it was difficult to train the model only with sparse triplet annotations, SSR-CNN introduced somewhat intricate training details: it detected object pairs with a Siamese Sparse RCNN model and auxiliary queries and performed additional triplet matching using the detected object pairs as a pseudo label. Although they show excellent triplet detection performance, the model architecture became even more intricate to detect subjects and objects separately since" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "text", + "content": "an explicit object detector is not used. Last but not least, the triplet prediction models focus on detecting a sub-graph composed only of objects with relations, overlooking objects in an image that lack explicit relations." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 124, + 545, + 293 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 124, + 545, + 293 + ], + "spans": [ + { + "bbox": [ + 304, + 124, + 545, + 293 + ], + "type": "text", + "content": "Relation Extraction Models. Relation extraction models extract a scene graph using a lightweight relation predictor without separate triplet queries or triplet detector. Relationformer [29] added a special \"[rln]\" token to capture global information in conjunction with object queries. They concatenated the final hidden representations of the object query pairs and the relation token, followed by a shallow fully connected network for relation prediction. In this work, in conjunction with the final hidden representations, we use the inherent relationship information among the object queries learned in the multi-head self-attention layer of the object detector. Furthermore, we propose training techniques to boost multi-task learning, leading to significantly improved performance with architectural simplicity." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 303, + 362, + 316 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 303, + 362, + 316 + ], + "spans": [ + { + "bbox": [ + 306, + 303, + 362, + 316 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 323, + 392, + 335 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 323, + 392, + 335 + ], + "spans": [ + { + "bbox": [ + 306, + 323, + 392, + 335 + ], + "type": "text", + "content": "3.1. Preliminaries" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 342, + 545, + 377 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 342, + 545, + 377 + ], + "spans": [ + { + "bbox": [ + 304, + 342, + 545, + 377 + ], + "type": "text", + "content": "In this section, we first introduce the formulation of the SGG task and then provide a brief review of the one-stage object detector that serves as the basis for our study." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 392, + 393, + 404 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 392, + 393, + 404 + ], + "spans": [ + { + "bbox": [ + 306, + 392, + 393, + 404 + ], + "type": "text", + "content": "3.1.1 Formulation" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 411, + 545, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 411, + 545, + 544 + ], + "spans": [ + { + "bbox": [ + 304, + 411, + 545, + 544 + ], + "type": "text", + "content": "Scene graph generation is a task of generating a scene graph " + }, + { + "bbox": [ + 304, + 411, + 545, + 544 + ], + "type": "inline_equation", + "content": "\\mathcal{G} = \\{\\mathcal{V},\\mathcal{E}\\}" + }, + { + "bbox": [ + 304, + 411, + 545, + 544 + ], + "type": "text", + "content": " from an image, where " + }, + { + "bbox": [ + 304, + 411, + 545, + 544 + ], + "type": "inline_equation", + "content": "\\mathcal{V}" + }, + { + "bbox": [ + 304, + 411, + 545, + 544 + ], + "type": "text", + "content": " denotes the node set consisting of objects and " + }, + { + "bbox": [ + 304, + 411, + 545, + 544 + ], + "type": "inline_equation", + "content": "\\mathcal{E}" + }, + { + "bbox": [ + 304, + 411, + 545, + 544 + ], + "type": "text", + "content": " denotes the edge set that represents the relation between objects. Each object entity " + }, + { + "bbox": [ + 304, + 411, + 545, + 544 + ], + "type": "inline_equation", + "content": "v_{i}\\in \\mathcal{V}" + }, + { + "bbox": [ + 304, + 411, + 545, + 544 + ], + "type": "text", + "content": " has an object category label " + }, + { + "bbox": [ + 304, + 411, + 545, + 544 + ], + "type": "inline_equation", + "content": "v_{i}^{c}" + }, + { + "bbox": [ + 304, + 411, + 545, + 544 + ], + "type": "text", + "content": " from a set of object categories " + }, + { + "bbox": [ + 304, + 411, + 545, + 544 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_v" + }, + { + "bbox": [ + 304, + 411, + 545, + 544 + ], + "type": "text", + "content": " and box coordinates " + }, + { + "bbox": [ + 304, + 411, + 545, + 544 + ], + "type": "inline_equation", + "content": "v_{i}^{b}" + }, + { + "bbox": [ + 304, + 411, + 545, + 544 + ], + "type": "text", + "content": ". Each relation " + }, + { + "bbox": [ + 304, + 411, + 545, + 544 + ], + "type": "inline_equation", + "content": "e_j\\in \\mathcal{E}" + }, + { + "bbox": [ + 304, + 411, + 545, + 544 + ], + "type": "text", + "content": " represents the " + }, + { + "bbox": [ + 304, + 411, + 545, + 544 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 304, + 411, + 545, + 544 + ], + "type": "text", + "content": "-th triplet " + }, + { + "bbox": [ + 304, + 411, + 545, + 544 + ], + "type": "inline_equation", + "content": "(s_j,p_j,o_j)" + }, + { + "bbox": [ + 304, + 411, + 545, + 544 + ], + "type": "text", + "content": ", where subject " + }, + { + "bbox": [ + 304, + 411, + 545, + 544 + ], + "type": "inline_equation", + "content": "s_j" + }, + { + "bbox": [ + 304, + 411, + 545, + 544 + ], + "type": "text", + "content": " and object " + }, + { + "bbox": [ + 304, + 411, + 545, + 544 + ], + "type": "inline_equation", + "content": "o_j" + }, + { + "bbox": [ + 304, + 411, + 545, + 544 + ], + "type": "text", + "content": " indicate related object entities and predicate " + }, + { + "bbox": [ + 304, + 411, + 545, + 544 + ], + "type": "inline_equation", + "content": "p_j" + }, + { + "bbox": [ + 304, + 411, + 545, + 544 + ], + "type": "text", + "content": " has a relation category label " + }, + { + "bbox": [ + 304, + 411, + 545, + 544 + ], + "type": "inline_equation", + "content": "p_j^c" + }, + { + "bbox": [ + 304, + 411, + 545, + 544 + ], + "type": "text", + "content": " from a set of predicate categories " + }, + { + "bbox": [ + 304, + 411, + 545, + 544 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_p" + }, + { + "bbox": [ + 304, + 411, + 545, + 544 + ], + "type": "text", + "content": ". Generating " + }, + { + "bbox": [ + 304, + 411, + 545, + 544 + ], + "type": "inline_equation", + "content": "\\mathcal{V}" + }, + { + "bbox": [ + 304, + 411, + 545, + 544 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 411, + 545, + 544 + ], + "type": "inline_equation", + "content": "\\mathcal{E}" + }, + { + "bbox": [ + 304, + 411, + 545, + 544 + ], + "type": "text", + "content": " correspond to object detection and relation extraction, respectively." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 559, + 453, + 571 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 559, + 453, + 571 + ], + "spans": [ + { + "bbox": [ + 306, + 559, + 453, + 571 + ], + "type": "text", + "content": "3.1.2 One-stage Object Detector" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 577, + 545, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 577, + 545, + 662 + ], + "spans": [ + { + "bbox": [ + 304, + 577, + 545, + 662 + ], + "type": "text", + "content": "Our model framework is deeply influenced by DETR [1], a one-stage object detector. In DETR, representations for input images are learned through CNN and Transformer [36] encoder. Transformer decoder enhances object queries by utilizing self- and cross-attention mechanisms. The final object detection heads predict object category labels and box coordinates from the contextualized object queries." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 665, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 545, + 714 + ], + "type": "text", + "content": "Backbone. CNN backbone generates " + }, + { + "bbox": [ + 304, + 665, + 545, + 714 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 304, + 665, + 545, + 714 + ], + "type": "text", + "content": "-dimensional feature map " + }, + { + "bbox": [ + 304, + 665, + 545, + 714 + ], + "type": "inline_equation", + "content": "F \\in \\mathbb{R}^{C \\times H_F \\times W_F}" + }, + { + "bbox": [ + 304, + 665, + 545, + 714 + ], + "type": "text", + "content": " from an input image " + }, + { + "bbox": [ + 304, + 665, + 545, + 714 + ], + "type": "inline_equation", + "content": "x_{\\mathrm{img}} \\in \\mathbb{R}^{3 \\times H_0 \\times W_0}" + }, + { + "bbox": [ + 304, + 665, + 545, + 714 + ], + "type": "text", + "content": ". For ResNet-50 backbone, " + }, + { + "bbox": [ + 304, + 665, + 545, + 714 + ], + "type": "inline_equation", + "content": "H_F" + }, + { + "bbox": [ + 304, + 665, + 545, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 665, + 545, + 714 + ], + "type": "inline_equation", + "content": "W_F" + }, + { + "bbox": [ + 304, + 665, + 545, + 714 + ], + "type": "text", + "content": " are set to " + }, + { + "bbox": [ + 304, + 665, + 545, + 714 + ], + "type": "inline_equation", + "content": "H_0 / 32" + }, + { + "bbox": [ + 304, + 665, + 545, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 665, + 545, + 714 + ], + "type": "inline_equation", + "content": "W_0 / 32" + }, + { + "bbox": [ + 304, + 665, + 545, + 714 + ], + "type": "text", + "content": ", respectively." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "24231" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 48, + 70, + 431, + 222 + ], + "blocks": [ + { + "bbox": [ + 48, + 70, + 431, + 222 + ], + "lines": [ + { + "bbox": [ + 48, + 70, + 431, + 222 + ], + "spans": [ + { + "bbox": [ + 48, + 70, + 431, + 222 + ], + "type": "image", + "image_path": "8ecf05244754a50b2be1aece4d420418de000403424f1f62df4891ed67db2042.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 236, + 547, + 281 + ], + "lines": [ + { + "bbox": [ + 46, + 236, + 547, + 281 + ], + "spans": [ + { + "bbox": [ + 46, + 236, + 547, + 281 + ], + "type": "text", + "content": "Figure 3. The overall architecture of EGTR. We present a novel lightweight relation extractor, EGTR, which fully utilizes the self-attention of the DETR decoder. We extract query and key representations from each self-attention layer and concatenate them pairwise to represent relations between them. Additionally, we leverage the last hidden representation in the same manner. To effectively aggregate information, we apply a gated sum and then predict relation with a shallow relation head." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 434, + 74, + 542, + 145 + ], + "blocks": [ + { + "bbox": [ + 434, + 74, + 542, + 145 + ], + "lines": [ + { + "bbox": [ + 434, + 74, + 542, + 145 + ], + "spans": [ + { + "bbox": [ + 434, + 74, + 542, + 145 + ], + "type": "image", + "image_path": "051d3620529018711d7ac2cc82099c6a645af5e74a4b87f1aedd5638cdb66c02.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 437, + 149, + 542, + 221 + ], + "blocks": [ + { + "bbox": [ + 437, + 149, + 542, + 221 + ], + "lines": [ + { + "bbox": [ + 437, + 149, + 542, + 221 + ], + "spans": [ + { + "bbox": [ + 437, + 149, + 542, + 221 + ], + "type": "image", + "image_path": "cbe093ab64a6516b8c1b4273701a8dbd74801876a7784e78ebbac8263356f574.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 291, + 288, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 291, + 288, + 376 + ], + "spans": [ + { + "bbox": [ + 46, + 291, + 288, + 376 + ], + "type": "text", + "content": "Transformer Encoder. Transformer encoder enhances image representations based on multi-head self-attention. After applying a " + }, + { + "bbox": [ + 46, + 291, + 288, + 376 + ], + "type": "inline_equation", + "content": "1 \\times 1" + }, + { + "bbox": [ + 46, + 291, + 288, + 376 + ], + "type": "text", + "content": " convolution to reduce dimension from " + }, + { + "bbox": [ + 46, + 291, + 288, + 376 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 46, + 291, + 288, + 376 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 46, + 291, + 288, + 376 + ], + "type": "inline_equation", + "content": "d_{\\mathrm{model}}" + }, + { + "bbox": [ + 46, + 291, + 288, + 376 + ], + "type": "text", + "content": ", the feature map is flattened to get an input sequence with length " + }, + { + "bbox": [ + 46, + 291, + 288, + 376 + ], + "type": "inline_equation", + "content": "H_{F}W_{F}" + }, + { + "bbox": [ + 46, + 291, + 288, + 376 + ], + "type": "text", + "content": " for the Transformer encoder. Additional positional encodings are used to reflect the spatial information of the feature maps." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 380, + 287, + 499 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 380, + 287, + 499 + ], + "spans": [ + { + "bbox": [ + 46, + 380, + 287, + 499 + ], + "type": "text", + "content": "Transformer Decoder. Transformer decoder takes " + }, + { + "bbox": [ + 46, + 380, + 287, + 499 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 380, + 287, + 499 + ], + "type": "text", + "content": " object queries as input and returns their representations. Each object query detects an individual object, and " + }, + { + "bbox": [ + 46, + 380, + 287, + 499 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 380, + 287, + 499 + ], + "type": "text", + "content": " is usually set large enough to cover all objects in the image. Through alternating self-attention and cross-attention layers, object queries learn features of object candidates in the input image. Note that causal attention mask is not applied to the self-attention layer; therefore, fully connected " + }, + { + "bbox": [ + 46, + 380, + 287, + 499 + ], + "type": "inline_equation", + "content": "N \\times N" + }, + { + "bbox": [ + 46, + 380, + 287, + 499 + ], + "type": "text", + "content": " attention weights among the " + }, + { + "bbox": [ + 46, + 380, + 287, + 499 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 380, + 287, + 499 + ], + "type": "text", + "content": " object queries are learned from the multi-head self-attention as follows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 63, + 508, + 287, + 526 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 508, + 287, + 526 + ], + "spans": [ + { + "bbox": [ + 63, + 508, + 287, + 526 + ], + "type": "interline_equation", + "content": "A _ {h} ^ {l} = f \\left(Q _ {h} ^ {l}, K _ {h} ^ {l}\\right) = \\operatorname {s o f t m a x} \\left(Q _ {h} ^ {l} K _ {h} ^ {T} / \\sqrt {d _ {\\text {h e a d}}}\\right), \\tag {1}", + "image_path": "af3263b3dc8e661d2a05a46d931316e793b8b3f28d1b0f504384c096c276c56b.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 534, + 288, + 572 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 534, + 288, + 572 + ], + "spans": [ + { + "bbox": [ + 47, + 534, + 288, + 572 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 534, + 288, + 572 + ], + "type": "inline_equation", + "content": "A_h^l \\in \\mathbb{R}^{N \\times N}" + }, + { + "bbox": [ + 47, + 534, + 288, + 572 + ], + "type": "text", + "content": " denotes the attention weights from the " + }, + { + "bbox": [ + 47, + 534, + 288, + 572 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 47, + 534, + 288, + 572 + ], + "type": "text", + "content": "-th head in the " + }, + { + "bbox": [ + 47, + 534, + 288, + 572 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 47, + 534, + 288, + 572 + ], + "type": "text", + "content": "-th layer. " + }, + { + "bbox": [ + 47, + 534, + 288, + 572 + ], + "type": "inline_equation", + "content": "Q_h^l \\in \\mathbb{R}^{N \\times d_{\\mathrm{head}}}" + }, + { + "bbox": [ + 47, + 534, + 288, + 572 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 534, + 288, + 572 + ], + "type": "inline_equation", + "content": "K_h^l \\in \\mathbb{R}^{N \\times d_{\\mathrm{head}}}" + }, + { + "bbox": [ + 47, + 534, + 288, + 572 + ], + "type": "text", + "content": " are attention queries and keys, respectively." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 576, + 287, + 650 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 576, + 287, + 650 + ], + "spans": [ + { + "bbox": [ + 47, + 576, + 287, + 650 + ], + "type": "text", + "content": "Object Detection Heads. Object detection heads detect " + }, + { + "bbox": [ + 47, + 576, + 287, + 650 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 47, + 576, + 287, + 650 + ], + "type": "text", + "content": " object candidates " + }, + { + "bbox": [ + 47, + 576, + 287, + 650 + ], + "type": "inline_equation", + "content": "\\{\\hat{v}_i\\}_{i = 1}^N" + }, + { + "bbox": [ + 47, + 576, + 287, + 650 + ], + "type": "text", + "content": " from the last layer representations of the object queries " + }, + { + "bbox": [ + 47, + 576, + 287, + 650 + ], + "type": "inline_equation", + "content": "Z^{L}" + }, + { + "bbox": [ + 47, + 576, + 287, + 650 + ], + "type": "text", + "content": ". For each object query, a linear layer is used to predict the object category label " + }, + { + "bbox": [ + 47, + 576, + 287, + 650 + ], + "type": "inline_equation", + "content": "\\hat{v}_i^c\\in \\mathbb{R}^{N\\times |\\mathcal{C}_v|}" + }, + { + "bbox": [ + 47, + 576, + 287, + 650 + ], + "type": "text", + "content": ", and three-layer perceptron with ReLU activation is used to obtain the box coordinates " + }, + { + "bbox": [ + 47, + 576, + 287, + 650 + ], + "type": "inline_equation", + "content": "\\hat{v}_i^b\\in \\mathbb{R}^{N\\times 4}" + }, + { + "bbox": [ + 47, + 576, + 287, + 650 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "type": "text", + "content": "Object Detection Loss. To match " + }, + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "type": "text", + "content": " detected object candidates " + }, + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\{\\hat{v}_i\\}_{i=1}^N" + }, + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "type": "text", + "content": " ground truth objects " + }, + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\{v_i\\}_{i=1}^M" + }, + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "type": "text", + "content": ", Carion et al. [1] pad the ground truth objects with " + }, + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "type": "text", + "content": " (no object) and find the best permutation of predicted objects that minimizes total bipartite matching costs. From the permuted" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 290, + 522, + 304 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 290, + 522, + 304 + ], + "spans": [ + { + "bbox": [ + 305, + 290, + 522, + 304 + ], + "type": "text", + "content": "predictions " + }, + { + "bbox": [ + 305, + 290, + 522, + 304 + ], + "type": "inline_equation", + "content": "\\{\\hat{v}_i^{\\prime}\\}_{i = 1}^{N}" + }, + { + "bbox": [ + 305, + 290, + 522, + 304 + ], + "type": "text", + "content": " , the loss is calculated as follows:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 312, + 545, + 328 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 312, + 545, + 328 + ], + "spans": [ + { + "bbox": [ + 313, + 312, + 545, + 328 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {o d}} = \\Sigma_ {i = 1} ^ {N} \\left[ \\lambda_ {c} \\mathcal {L} _ {c} \\left(\\hat {v} _ {i} ^ {\\prime c}, v _ {i} ^ {c}\\right) + \\mathbb {1} _ {v _ {i} ^ {c} \\neq \\phi} \\left(\\lambda_ {b} \\mathcal {L} _ {b} \\left(\\hat {v} _ {i} ^ {\\prime b}, v _ {i} ^ {b}\\right)\\right) \\right], \\tag {2}", + "image_path": "0e64ca0fc91f6276c9cf792d4b24f36c40a6f2128445961e6a4df1be853fcb32.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 337, + 545, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 337, + 545, + 361 + ], + "spans": [ + { + "bbox": [ + 305, + 337, + 545, + 361 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 337, + 545, + 361 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_c" + }, + { + "bbox": [ + 305, + 337, + 545, + 361 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 337, + 545, + 361 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_b" + }, + { + "bbox": [ + 305, + 337, + 545, + 361 + ], + "type": "text", + "content": " denote loss function for category labels and box coordinates, respectively." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 370, + 360, + 382 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 370, + 360, + 382 + ], + "spans": [ + { + "bbox": [ + 306, + 370, + 360, + 382 + ], + "type": "text", + "content": "3.2. EGTR" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 389, + 545, + 485 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 389, + 545, + 485 + ], + "spans": [ + { + "bbox": [ + 304, + 389, + 545, + 485 + ], + "type": "text", + "content": "We propose a novel lightweight relation extractor, EGTR, which exploits the self-attention of DETR decoder, as depicted in Fig. 3. Since the self-attention weights in Eq. (1) contain " + }, + { + "bbox": [ + 304, + 389, + 545, + 485 + ], + "type": "inline_equation", + "content": "N \\times N" + }, + { + "bbox": [ + 304, + 389, + 545, + 485 + ], + "type": "text", + "content": " bidirectional relationships among the " + }, + { + "bbox": [ + 304, + 389, + 545, + 485 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 389, + 545, + 485 + ], + "type": "text", + "content": " object queries, our relation extractor aims to extract the predicate information from the self-attention weights in the entire " + }, + { + "bbox": [ + 304, + 389, + 545, + 485 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 304, + 389, + 545, + 485 + ], + "type": "text", + "content": " layers, by considering the attention queries and keys as subjects and objects, respectively." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 486, + 546, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 486, + 546, + 605 + ], + "spans": [ + { + "bbox": [ + 304, + 486, + 546, + 605 + ], + "type": "text", + "content": "In order to preserve rich information learned in the self-attention layer, we devise another relation function " + }, + { + "bbox": [ + 304, + 486, + 546, + 605 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 304, + 486, + 546, + 605 + ], + "type": "text", + "content": " between the attention queries and keys instead of the dot product attention in Eq. (1). As concatenation preserves representations completely, we concatenate the representations on all " + }, + { + "bbox": [ + 304, + 486, + 546, + 605 + ], + "type": "inline_equation", + "content": "N \\times N" + }, + { + "bbox": [ + 304, + 486, + 546, + 605 + ], + "type": "text", + "content": " pairs of attention queries and keys as shown in Fig. 3 (a) to get the relation representations of the " + }, + { + "bbox": [ + 304, + 486, + 546, + 605 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 304, + 486, + 546, + 605 + ], + "type": "text", + "content": "-th layer " + }, + { + "bbox": [ + 304, + 486, + 546, + 605 + ], + "type": "inline_equation", + "content": "R_{a}^{l} \\in \\mathbb{R}^{N \\times N \\times 2d_{\\mathrm{model}}}" + }, + { + "bbox": [ + 304, + 486, + 546, + 605 + ], + "type": "text", + "content": ". Before pairwise concatenation, to help queries and keys play the role of subjects and objects, a linear projection is added as follows:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 377, + 615, + 545, + 630 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 377, + 615, + 545, + 630 + ], + "spans": [ + { + "bbox": [ + 377, + 615, + 545, + 630 + ], + "type": "interline_equation", + "content": "R _ {a} ^ {l} = \\left[ Q ^ {l} W _ {S} ^ {l}; K ^ {l} W _ {O} ^ {l} \\right], \\tag {3}", + "image_path": "9a7973120aaa87802992aa6c8931c1ef5a9304881ad73708de629131b7a0f88a.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 639, + 545, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 639, + 545, + 688 + ], + "spans": [ + { + "bbox": [ + 304, + 639, + 545, + 688 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 639, + 545, + 688 + ], + "type": "inline_equation", + "content": "Q^{l}\\in \\mathbb{R}^{N\\times d_{\\mathrm{model}}}" + }, + { + "bbox": [ + 304, + 639, + 545, + 688 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 639, + 545, + 688 + ], + "type": "inline_equation", + "content": "K^l\\in \\mathbb{R}^{N\\times d_{\\mathrm{model}}}" + }, + { + "bbox": [ + 304, + 639, + 545, + 688 + ], + "type": "text", + "content": " refer to attention queries and keys of the " + }, + { + "bbox": [ + 304, + 639, + 545, + 688 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 304, + 639, + 545, + 688 + ], + "type": "text", + "content": " -th layer, and " + }, + { + "bbox": [ + 304, + 639, + 545, + 688 + ], + "type": "inline_equation", + "content": "[\\cdot ;\\cdot ]" + }, + { + "bbox": [ + 304, + 639, + 545, + 688 + ], + "type": "text", + "content": " denotes the pairwise concatenation. " + }, + { + "bbox": [ + 304, + 639, + 545, + 688 + ], + "type": "inline_equation", + "content": "W_{S}^{l}" + }, + { + "bbox": [ + 304, + 639, + 545, + 688 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 639, + 545, + 688 + ], + "type": "inline_equation", + "content": "W_{O}^{l}" + }, + { + "bbox": [ + 304, + 639, + 545, + 688 + ], + "type": "text", + "content": " are linear weights of shape " + }, + { + "bbox": [ + 304, + 639, + 545, + 688 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^{d_{\\mathrm{model}}\\times d_{\\mathrm{model}}}" + }, + { + "bbox": [ + 304, + 639, + 545, + 688 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 689, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 689, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 689, + 545, + 714 + ], + "type": "text", + "content": "We also utilize the last layer representations of the object queries " + }, + { + "bbox": [ + 304, + 689, + 545, + 714 + ], + "type": "inline_equation", + "content": "Z^{L} \\in \\mathbb{R}^{N \\times d_{\\mathrm{model}}}" + }, + { + "bbox": [ + 304, + 689, + 545, + 714 + ], + "type": "text", + "content": ", which are used for the object" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "24232" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 171, + 83 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 171, + 83 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 171, + 83 + ], + "type": "text", + "content": "detection in the same manner:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 117, + 92, + 287, + 106 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 117, + 92, + 287, + 106 + ], + "spans": [ + { + "bbox": [ + 117, + 92, + 287, + 106 + ], + "type": "interline_equation", + "content": "R _ {z} = \\left[ Z ^ {L} W _ {S}; Z ^ {L} W _ {O} \\right], \\tag {4}", + "image_path": "8bed13360791e8bc5879ab4f2abcea26da06c467c9676bcf1424e221b5df1597.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 116, + 287, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 116, + 287, + 152 + ], + "spans": [ + { + "bbox": [ + 47, + 116, + 287, + 152 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 116, + 287, + 152 + ], + "type": "inline_equation", + "content": "W_{S}, W_{O}" + }, + { + "bbox": [ + 47, + 116, + 287, + 152 + ], + "type": "text", + "content": " are linear weights of shape " + }, + { + "bbox": [ + 47, + 116, + 287, + 152 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^{d_{\\mathrm{model}} \\times d_{\\mathrm{model}}}" + }, + { + "bbox": [ + 47, + 116, + 287, + 152 + ], + "type": "text", + "content": ". To effectively use the various information learned in all layers, we introduce a gating mechanism as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 96, + 159, + 287, + 175 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 96, + 159, + 287, + 175 + ], + "spans": [ + { + "bbox": [ + 96, + 159, + 287, + 175 + ], + "type": "interline_equation", + "content": "g _ {a} ^ {l} = \\sigma \\left(R _ {a} ^ {l} W _ {G}\\right), g _ {z} = \\sigma \\left(R _ {z} W _ {G}\\right), \\tag {5}", + "image_path": "f5eb7d499c40d1ede47e1f4946b3ec32206691955f1ad0a1a08c5b4c573febc3.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 182, + 288, + 244 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 182, + 288, + 244 + ], + "spans": [ + { + "bbox": [ + 47, + 182, + 288, + 244 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 182, + 288, + 244 + ], + "type": "inline_equation", + "content": "g_{a}^{l}" + }, + { + "bbox": [ + 47, + 182, + 288, + 244 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 182, + 288, + 244 + ], + "type": "inline_equation", + "content": "g_{z} \\in \\mathbb{R}^{N \\times N \\times 1}" + }, + { + "bbox": [ + 47, + 182, + 288, + 244 + ], + "type": "text", + "content": " represent the gate values obtained through a single linear layer " + }, + { + "bbox": [ + 47, + 182, + 288, + 244 + ], + "type": "inline_equation", + "content": "W_{G} \\in \\mathbb{R}^{2d_{\\mathrm{model}} \\times 1}" + }, + { + "bbox": [ + 47, + 182, + 288, + 244 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 47, + 182, + 288, + 244 + ], + "type": "inline_equation", + "content": "R_{a}^{l}" + }, + { + "bbox": [ + 47, + 182, + 288, + 244 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 182, + 288, + 244 + ], + "type": "inline_equation", + "content": "R_{z}" + }, + { + "bbox": [ + 47, + 182, + 288, + 244 + ], + "type": "text", + "content": ", respectively. Finally, we extract the relation graph from the gated summation of the relation representations across all layers as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 77, + 263, + 287, + 279 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 263, + 287, + 279 + ], + "spans": [ + { + "bbox": [ + 77, + 263, + 287, + 279 + ], + "type": "interline_equation", + "content": "\\hat {G} = \\sigma \\left(\\mathrm {M L P} _ {\\mathrm {r e l}} \\left(\\Sigma_ {l = 1} ^ {L} \\left(g _ {a} ^ {l} * R _ {a} ^ {l}\\right) + g _ {z} * R _ {z}\\right)\\right), \\tag {6}", + "image_path": "94257f826935c04d3eafb51fadf10296b1220db2d1ff6c62e3c826e8b2328d4a.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 288, + 287, + 337 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 288, + 287, + 337 + ], + "spans": [ + { + "bbox": [ + 47, + 288, + 287, + 337 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 288, + 287, + 337 + ], + "type": "inline_equation", + "content": "\\hat{G} \\in \\mathbb{R}^{N \\times N \\times |\\mathcal{C}_p|}" + }, + { + "bbox": [ + 47, + 288, + 287, + 337 + ], + "type": "text", + "content": " denotes predicted relation graph and " + }, + { + "bbox": [ + 47, + 288, + 287, + 337 + ], + "type": "inline_equation", + "content": "\\mathrm{MLP}_{\\mathrm{rel}}" + }, + { + "bbox": [ + 47, + 288, + 287, + 337 + ], + "type": "text", + "content": " is a three-layer perceptron with ReLU activation. Note that we use sigmoid function " + }, + { + "bbox": [ + 47, + 288, + 287, + 337 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 47, + 288, + 287, + 337 + ], + "type": "text", + "content": " so that multiple relationships can exist between objects." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 344, + 180, + 357 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 344, + 180, + 357 + ], + "spans": [ + { + "bbox": [ + 47, + 344, + 180, + 357 + ], + "type": "text", + "content": "3.3. Learning and Inference" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 362, + 287, + 409 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 362, + 287, + 409 + ], + "spans": [ + { + "bbox": [ + 47, + 362, + 287, + 409 + ], + "type": "text", + "content": "To train EGTR, we perform multi-task learning. In addition to object detection and relation extraction, we devise connectivity prediction, an auxiliary task for relation extraction. The overall loss for the framework is as follows:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 103, + 420, + 287, + 433 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 103, + 420, + 287, + 433 + ], + "spans": [ + { + "bbox": [ + 103, + 420, + 287, + 433 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\mathcal {L} _ {\\mathrm {o d}} + \\lambda_ {\\mathrm {r e l}} \\mathcal {L} _ {\\mathrm {r e l}} + \\lambda_ {\\mathrm {c o n}} \\mathcal {L} _ {\\mathrm {c o n}}, \\tag {7}", + "image_path": "da31118e5c0cb2b594b684a5c439aebcd3e7705588297bb28fd4523eb27e2292.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 441, + 287, + 513 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 441, + 287, + 513 + ], + "spans": [ + { + "bbox": [ + 47, + 441, + 287, + 513 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 441, + 287, + 513 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{od}}" + }, + { + "bbox": [ + 47, + 441, + 287, + 513 + ], + "type": "text", + "content": " is object detection loss in Eq. (2). " + }, + { + "bbox": [ + 47, + 441, + 287, + 513 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{rel}}" + }, + { + "bbox": [ + 47, + 441, + 287, + 513 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 441, + 287, + 513 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{con}}" + }, + { + "bbox": [ + 47, + 441, + 287, + 513 + ], + "type": "text", + "content": " are loss functions for the relation extraction and connectivity prediction, respectively. Since explicit object detection loss is used, EGTR has the capability to detect all nodes in the scene graph. The details of the loss for each task are provided below." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 527, + 165, + 538 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 527, + 165, + 538 + ], + "spans": [ + { + "bbox": [ + 47, + 527, + 165, + 538 + ], + "type": "text", + "content": "3.3.1 Relation Extraction" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 47, + 546, + 287, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 546, + 287, + 653 + ], + "spans": [ + { + "bbox": [ + 47, + 546, + 287, + 653 + ], + "type": "text", + "content": "We use binary cross-entropy loss for the relation extraction. To match predicted graph " + }, + { + "bbox": [ + 47, + 546, + 287, + 653 + ], + "type": "inline_equation", + "content": "\\hat{G} \\in \\mathbb{R}^{N \\times N \\times |\\mathcal{C}_p|}" + }, + { + "bbox": [ + 47, + 546, + 287, + 653 + ], + "type": "text", + "content": " and ground truth triplet set " + }, + { + "bbox": [ + 47, + 546, + 287, + 653 + ], + "type": "inline_equation", + "content": "\\mathcal{E}" + }, + { + "bbox": [ + 47, + 546, + 287, + 653 + ], + "type": "text", + "content": ", we first encode " + }, + { + "bbox": [ + 47, + 546, + 287, + 653 + ], + "type": "inline_equation", + "content": "\\mathcal{E}" + }, + { + "bbox": [ + 47, + 546, + 287, + 653 + ], + "type": "text", + "content": " as a one-hot ground truth graph " + }, + { + "bbox": [ + 47, + 546, + 287, + 653 + ], + "type": "inline_equation", + "content": "G \\in \\mathbb{R}^{N \\times N \\times |\\mathcal{C}_p|}" + }, + { + "bbox": [ + 47, + 546, + 287, + 653 + ], + "type": "text", + "content": " by padding the regions that do not correspond to relations between the ground truth objects as zero. Then, we permute indices of the predicted graph using the permutation found in the object detection. From the permuted graph " + }, + { + "bbox": [ + 47, + 546, + 287, + 653 + ], + "type": "inline_equation", + "content": "\\hat{G}'" + }, + { + "bbox": [ + 47, + 546, + 287, + 653 + ], + "type": "text", + "content": ", the loss for the relation extraction is calculated as " + }, + { + "bbox": [ + 47, + 546, + 287, + 653 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{rel}} = \\mathcal{L}_r(\\hat{G}', G)" + }, + { + "bbox": [ + 47, + 546, + 287, + 653 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 47, + 654, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 654, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 654, + 287, + 714 + ], + "type": "text", + "content": "However, since the size of the graph is proportional to the square of " + }, + { + "bbox": [ + 47, + 654, + 287, + 714 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 47, + 654, + 287, + 714 + ], + "type": "text", + "content": ", the sparsity of the ground truth graph " + }, + { + "bbox": [ + 47, + 654, + 287, + 714 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 47, + 654, + 287, + 714 + ], + "type": "text", + "content": " is too severe. For instance, the density is only " + }, + { + "bbox": [ + 47, + 654, + 287, + 714 + ], + "type": "inline_equation", + "content": "10^{-14}" + }, + { + "bbox": [ + 47, + 654, + 287, + 714 + ], + "type": "text", + "content": " when " + }, + { + "bbox": [ + 47, + 654, + 287, + 714 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 47, + 654, + 287, + 714 + ], + "type": "text", + "content": " is set to 200 for the Visual Genome [12] validation dataset. Therefore, we divide " + }, + { + "bbox": [ + 47, + 654, + 287, + 714 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 47, + 654, + 287, + 714 + ], + "type": "text", + "content": " into three regions: (1) GT region," + } + ] + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 318, + 75, + 414, + 175 + ], + "blocks": [ + { + "bbox": [ + 318, + 75, + 414, + 175 + ], + "lines": [ + { + "bbox": [ + 318, + 75, + 414, + 175 + ], + "spans": [ + { + "bbox": [ + 318, + 75, + 414, + 175 + ], + "type": "image", + "image_path": "8b4b9a348b44f666252103d9ddd29526093a28ee47d8308828587aeab359d27e.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 339, + 190, + 392, + 198 + ], + "lines": [ + { + "bbox": [ + 339, + 190, + 392, + 198 + ], + "spans": [ + { + "bbox": [ + 339, + 190, + 392, + 198 + ], + "type": "text", + "content": "girl - on - horse" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 339, + 198, + 373, + 206 + ], + "lines": [ + { + "bbox": [ + 339, + 198, + 373, + 206 + ], + "spans": [ + { + "bbox": [ + 339, + 198, + 373, + 206 + ], + "type": "text", + "content": "jirl - riding" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_footnote" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 414, + 73, + 537, + 204 + ], + "blocks": [ + { + "bbox": [ + 319, + 179, + 413, + 189 + ], + "lines": [ + { + "bbox": [ + 319, + 179, + 413, + 189 + ], + "spans": [ + { + "bbox": [ + 319, + 179, + 413, + 189 + ], + "type": "text", + "content": "(subject - predicate - object)" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 334, + 207, + 399, + 217 + ], + "lines": [ + { + "bbox": [ + 334, + 207, + 399, + 217 + ], + "spans": [ + { + "bbox": [ + 334, + 207, + 399, + 217 + ], + "type": "text", + "content": "girl - wearing - pant" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 414, + 73, + 537, + 204 + ], + "lines": [ + { + "bbox": [ + 414, + 73, + 537, + 204 + ], + "spans": [ + { + "bbox": [ + 414, + 73, + 537, + 204 + ], + "type": "image", + "image_path": "150501b135af7d479b690a4d0f111b17fd90658d9f4ed8d284146a7bbe9e14b7.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 417, + 205, + 482, + 217 + ], + "lines": [ + { + "bbox": [ + 417, + 205, + 482, + 217 + ], + "spans": [ + { + "bbox": [ + 417, + 205, + 482, + 217 + ], + "type": "text", + "content": "non-matching region" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 305, + 227, + 546, + 282 + ], + "lines": [ + { + "bbox": [ + 305, + 227, + 546, + 282 + ], + "spans": [ + { + "bbox": [ + 305, + 227, + 546, + 282 + ], + "type": "text", + "content": "Figure 4. Example of graph " + }, + { + "bbox": [ + 305, + 227, + 546, + 282 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 305, + 227, + 546, + 282 + ], + "type": "text", + "content": " region. The target graph has a shape of " + }, + { + "bbox": [ + 305, + 227, + 546, + 282 + ], + "type": "inline_equation", + "content": "N \\times N \\times |\\mathcal{V}_c|" + }, + { + "bbox": [ + 305, + 227, + 546, + 282 + ], + "type": "text", + "content": ". The graph is severely sparse since the number of object queries " + }, + { + "bbox": [ + 305, + 227, + 546, + 282 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 305, + 227, + 546, + 282 + ], + "type": "text", + "content": " is set large enough to cover objects in the image. We split " + }, + { + "bbox": [ + 305, + 227, + 546, + 282 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 305, + 227, + 546, + 282 + ], + "type": "text", + "content": " into GT, negative, and non-matching regions." + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + } + ], + "index": 20 + }, + { + "bbox": [ + 305, + 289, + 546, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 289, + 546, + 422 + ], + "spans": [ + { + "bbox": [ + 305, + 289, + 546, + 422 + ], + "type": "text", + "content": "(2) negative region, and (3) non-matching region, as shown in Fig. 4. The GT region indicates the ground truth triplets where the value of " + }, + { + "bbox": [ + 305, + 289, + 546, + 422 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 305, + 289, + 546, + 422 + ], + "type": "text", + "content": " is 1. The negative region consists of the triplets in which both subjects and objects are composed of the ground truth objects, but no relation exists between them. The non-matching region represents the zero-padded region. It is paired with a region consisting of the object candidates that do not match the ground truth objects and match " + }, + { + "bbox": [ + 305, + 289, + 546, + 422 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 305, + 289, + 546, + 422 + ], + "type": "text", + "content": " in Eq. (2). For each region, we apply different techniques to effectively train the relation extraction with object detection as subsequently described." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 305, + 426, + 545, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 426, + 545, + 581 + ], + "spans": [ + { + "bbox": [ + 305, + 426, + 545, + 581 + ], + "type": "text", + "content": "Adaptive Smoothing. We propose a novel adaptive smoothing for the GT region. For " + }, + { + "bbox": [ + 305, + 426, + 545, + 581 + ], + "type": "inline_equation", + "content": "G_{ijk}" + }, + { + "bbox": [ + 305, + 426, + 545, + 581 + ], + "type": "text", + "content": " belonging to the GT region, the model is trained to predict the " + }, + { + "bbox": [ + 305, + 426, + 545, + 581 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 305, + 426, + 545, + 581 + ], + "type": "text", + "content": "-th predicate category between subject entity " + }, + { + "bbox": [ + 305, + 426, + 545, + 581 + ], + "type": "inline_equation", + "content": "v_i" + }, + { + "bbox": [ + 305, + 426, + 545, + 581 + ], + "type": "text", + "content": " and object entity " + }, + { + "bbox": [ + 305, + 426, + 545, + 581 + ], + "type": "inline_equation", + "content": "v_j" + }, + { + "bbox": [ + 305, + 426, + 545, + 581 + ], + "type": "text", + "content": ". However, since " + }, + { + "bbox": [ + 305, + 426, + 545, + 581 + ], + "type": "inline_equation", + "content": "\\hat{v}_i'" + }, + { + "bbox": [ + 305, + 426, + 545, + 581 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 426, + 545, + 581 + ], + "type": "inline_equation", + "content": "\\hat{v}_j'" + }, + { + "bbox": [ + 305, + 426, + 545, + 581 + ], + "type": "text", + "content": ", which match " + }, + { + "bbox": [ + 305, + 426, + 545, + 581 + ], + "type": "inline_equation", + "content": "v_i" + }, + { + "bbox": [ + 305, + 426, + 545, + 581 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 426, + 545, + 581 + ], + "type": "inline_equation", + "content": "v_j" + }, + { + "bbox": [ + 305, + 426, + 545, + 581 + ], + "type": "text", + "content": " respectively, do not have enough representations about the corresponding ground truth object at the beginning of the training, it may not be appropriate to predict the probability of the predicate as 1. Moreover, even when the object detection performance is reasonably assured, the detection performance may still vary for individual object candidates. Therefore, we reflect the detection performance of each object candidate on the relation label via adaptive smoothing." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 305, + 581, + 545, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 581, + 545, + 616 + ], + "spans": [ + { + "bbox": [ + 305, + 581, + 545, + 616 + ], + "type": "text", + "content": "We first measure the uncertainty of each object candidate with the corresponding bipartite matching cost. For object candidate " + }, + { + "bbox": [ + 305, + 581, + 545, + 616 + ], + "type": "inline_equation", + "content": "\\hat{v}_i'" + }, + { + "bbox": [ + 305, + 581, + 545, + 616 + ], + "type": "text", + "content": ", we define the uncertainty as follows:" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 353, + 622, + 545, + 636 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 353, + 622, + 545, + 636 + ], + "spans": [ + { + "bbox": [ + 353, + 622, + 545, + 636 + ], + "type": "interline_equation", + "content": "u _ {i} = \\sigma \\left(\\operatorname {c o s t} _ {i} - \\operatorname {c o s t} _ {\\min } + \\sigma^ {- 1} (\\alpha)\\right), \\tag {8}", + "image_path": "f96628f46464ffcfc451f91c09b46cf3b7fdba245c6d1bab450ce80d491821fa.jpg" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 305, + 642, + 546, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 642, + 546, + 714 + ], + "spans": [ + { + "bbox": [ + 305, + 642, + 546, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 642, + 546, + 714 + ], + "type": "inline_equation", + "content": "\\mathrm{cost}_i" + }, + { + "bbox": [ + 305, + 642, + 546, + 714 + ], + "type": "text", + "content": " denotes the matching cost and " + }, + { + "bbox": [ + 305, + 642, + 546, + 714 + ], + "type": "inline_equation", + "content": "\\mathrm{cost}_{\\mathrm{min}}" + }, + { + "bbox": [ + 305, + 642, + 546, + 714 + ], + "type": "text", + "content": " indicates the matching cost when " + }, + { + "bbox": [ + 305, + 642, + 546, + 714 + ], + "type": "inline_equation", + "content": "\\hat{v}_i^\\prime" + }, + { + "bbox": [ + 305, + 642, + 546, + 714 + ], + "type": "text", + "content": " perfectly matches " + }, + { + "bbox": [ + 305, + 642, + 546, + 714 + ], + "type": "inline_equation", + "content": "v_{i}" + }, + { + "bbox": [ + 305, + 642, + 546, + 714 + ], + "type": "text", + "content": " . " + }, + { + "bbox": [ + 305, + 642, + 546, + 714 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 305, + 642, + 546, + 714 + ], + "type": "text", + "content": " is a nonnegative hyperparameter representing the minimum uncertainty. We set the value of " + }, + { + "bbox": [ + 305, + 642, + 546, + 714 + ], + "type": "inline_equation", + "content": "G_{ijk}" + }, + { + "bbox": [ + 305, + 642, + 546, + 714 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 305, + 642, + 546, + 714 + ], + "type": "inline_equation", + "content": "(1 - u_{i})(1 - u_{j})" + }, + { + "bbox": [ + 305, + 642, + 546, + 714 + ], + "type": "text", + "content": " taking the uncertainty into account. By using the relation label adjusted by the uncertainty, the multi-task learning of object" + } + ] + } + ], + "index": 27 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "24233" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 287, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 287, + 96 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 287, + 96 + ], + "type": "text", + "content": "detection and relation extraction is dynamically regulated according to the quality of the detected objects." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 100, + 286, + 208 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 100, + 286, + 208 + ], + "spans": [ + { + "bbox": [ + 47, + 100, + 286, + 208 + ], + "type": "text", + "content": "Negative and Non-matching Sampling. Rather than employing all negatives, we sample them from the negative region. Inspired by the hard negative mining introduced by Liu et al. [21], we sort all negatives based on the predicted relation score " + }, + { + "bbox": [ + 47, + 100, + 286, + 208 + ], + "type": "inline_equation", + "content": "\\hat{G}_{ijk}^{\\prime}" + }, + { + "bbox": [ + 47, + 100, + 286, + 208 + ], + "type": "text", + "content": " and choose the top " + }, + { + "bbox": [ + 47, + 100, + 286, + 208 + ], + "type": "inline_equation", + "content": "k_{\\mathrm{neg}} \\times |\\mathcal{E}|" + }, + { + "bbox": [ + 47, + 100, + 286, + 208 + ], + "type": "text", + "content": " most challenging negatives. Similarly, we extract " + }, + { + "bbox": [ + 47, + 100, + 286, + 208 + ], + "type": "inline_equation", + "content": "k_{\\mathrm{non}} \\times |\\mathcal{E}|" + }, + { + "bbox": [ + 47, + 100, + 286, + 208 + ], + "type": "text", + "content": " hard samples from the non-matching region. As the non-matching region typically encompasses a substantial part of the graph " + }, + { + "bbox": [ + 47, + 100, + 286, + 208 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 47, + 100, + 286, + 208 + ], + "type": "text", + "content": ", this approach notably diminishes sparsity." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 223, + 180, + 234 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 223, + 180, + 234 + ], + "spans": [ + { + "bbox": [ + 47, + 223, + 180, + 234 + ], + "type": "text", + "content": "3.3.2 Connectivity Prediction" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 242, + 286, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 242, + 286, + 361 + ], + "spans": [ + { + "bbox": [ + 46, + 242, + 286, + 361 + ], + "type": "text", + "content": "Influenced by Graph-RCNN [40] that predicted a relatedness to prune object pairs, we propose a connectivity prediction that predicts whether at least one edge between two object nodes exists for the relation extraction. We get a connectivity graph " + }, + { + "bbox": [ + 46, + 242, + 286, + 361 + ], + "type": "inline_equation", + "content": "\\hat{E} \\in \\mathbb{R}^{N \\times N \\times 1}" + }, + { + "bbox": [ + 46, + 242, + 286, + 361 + ], + "type": "text", + "content": " in a similar way to get the relation graph in Eq. (6) with the same relation source representations. Instead of " + }, + { + "bbox": [ + 46, + 242, + 286, + 361 + ], + "type": "inline_equation", + "content": "\\mathrm{MLP}_{\\mathrm{rel}}" + }, + { + "bbox": [ + 46, + 242, + 286, + 361 + ], + "type": "text", + "content": " for multi-label prediction, we use another " + }, + { + "bbox": [ + 46, + 242, + 286, + 361 + ], + "type": "inline_equation", + "content": "\\mathrm{MLP}_{\\mathrm{con}}" + }, + { + "bbox": [ + 46, + 242, + 286, + 361 + ], + "type": "text", + "content": " for binary classification. We calculate the binary cross entropy loss from the permutated connectivity graph as follows: " + }, + { + "bbox": [ + 46, + 242, + 286, + 361 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{con}} = \\mathcal{L}_r(\\hat{E}', E)" + }, + { + "bbox": [ + 46, + 242, + 286, + 361 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 376, + 120, + 387 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 376, + 120, + 387 + ], + "spans": [ + { + "bbox": [ + 47, + 376, + 120, + 387 + ], + "type": "text", + "content": "3.3.3 Inference" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 396, + 286, + 492 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 396, + 286, + 492 + ], + "spans": [ + { + "bbox": [ + 47, + 396, + 286, + 492 + ], + "type": "text", + "content": "For model inference, we get triplet scores by multiplying predicate score " + }, + { + "bbox": [ + 47, + 396, + 286, + 492 + ], + "type": "inline_equation", + "content": "\\hat{G}_{ijk}" + }, + { + "bbox": [ + 47, + 396, + 286, + 492 + ], + "type": "text", + "content": " by the corresponding class scores of " + }, + { + "bbox": [ + 47, + 396, + 286, + 492 + ], + "type": "inline_equation", + "content": "\\hat{v}_i^c" + }, + { + "bbox": [ + 47, + 396, + 286, + 492 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 396, + 286, + 492 + ], + "type": "inline_equation", + "content": "\\hat{v}_j^c" + }, + { + "bbox": [ + 47, + 396, + 286, + 492 + ], + "type": "text", + "content": ". Note that we set " + }, + { + "bbox": [ + 47, + 396, + 286, + 492 + ], + "type": "inline_equation", + "content": "\\hat{G}_{iik}" + }, + { + "bbox": [ + 47, + 396, + 286, + 492 + ], + "type": "text", + "content": " to 0 to prevent self-connections in which the subject and object are the same entity. In addition, we enhance the triplet scores by utilizing the connectivity score " + }, + { + "bbox": [ + 47, + 396, + 286, + 492 + ], + "type": "inline_equation", + "content": "\\hat{E}_{ijk}" + }, + { + "bbox": [ + 47, + 396, + 286, + 492 + ], + "type": "text", + "content": ". By multiplying the connectivity score, we can effectively filter out the triplets that are not likely to have a relation between the subject and object." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 504, + 127, + 517 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 504, + 127, + 517 + ], + "spans": [ + { + "bbox": [ + 47, + 504, + 127, + 517 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 523, + 223, + 536 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 523, + 223, + 536 + ], + "spans": [ + { + "bbox": [ + 47, + 523, + 223, + 536 + ], + "type": "text", + "content": "4.1. Datasets and Evaluation Settings" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 542, + 286, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 542, + 286, + 590 + ], + "spans": [ + { + "bbox": [ + 47, + 542, + 286, + 590 + ], + "type": "text", + "content": "We conduct experiments on two SGG datasets. We describe the datasets and evaluation settings for each dataset. Detailed settings for each dataset are presented in the supplementary materials." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 594, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 286, + 713 + ], + "type": "text", + "content": "Visual Genome. Visual Genome [12] is the most representative SGG dataset, consisting of 57K training images, 5K validation images, and 26K test images. We follow the popular Visual Genome split, retaining the most frequent 150 object categories and 50 relation categories. We adopt Scene Graph Detection (SGDet) evaluation settings and report Recall@k (R@k) which is class agnostic and mean Recall@k (mR@k) that aggregates the recalls for each predicate category. Following Motifs [42], these metrics are measures with graph constraint, which means each" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "text", + "content": "object pair can have a single predicate category. Since they are only related to triplet detection of the scene graph, we report AP50 that evaluates the detection performance of all objects appearing in the scene. Furthermore, we report the number of model parameters and Frames Per Second (FPS) to measure efficiency." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 148, + 545, + 281 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 148, + 545, + 281 + ], + "spans": [ + { + "bbox": [ + 304, + 148, + 545, + 281 + ], + "type": "text", + "content": "Open Image V6. Open Image V6 [13] is also widely used dataset, comprising of 126K images for training, 2K for validation, and 5K for test set. It includes 601 object categories and 30 relation categories. For evaluation, we adopt both recall and weighted mean AP (wmAP) following standard settings. For recall evaluation, micro-R@50 is adopted. The WMAP is evaluated with two settings: " + }, + { + "bbox": [ + 304, + 148, + 545, + 281 + ], + "type": "inline_equation", + "content": "\\mathrm{wmAP}_{\\mathrm{rel}}" + }, + { + "bbox": [ + 304, + 148, + 545, + 281 + ], + "type": "text", + "content": " for predicting boxes of subject entity and object entity separately and " + }, + { + "bbox": [ + 304, + 148, + 545, + 281 + ], + "type": "inline_equation", + "content": "\\mathrm{wmAP}_{\\mathrm{phr}}" + }, + { + "bbox": [ + 304, + 148, + 545, + 281 + ], + "type": "text", + "content": " for predicting a union box of them. The final score is calculated by " + }, + { + "bbox": [ + 304, + 148, + 545, + 281 + ], + "type": "inline_equation", + "content": "0.2 \\times \\text{micro-R} @50 + 0.4 \\times \\mathrm{wmAP}_{\\mathrm{rel}} + 0.4 \\times \\mathrm{wmAP}_{\\mathrm{phr}}" + }, + { + "bbox": [ + 304, + 148, + 545, + 281 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 290, + 438, + 302 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 290, + 438, + 302 + ], + "spans": [ + { + "bbox": [ + 306, + 290, + 438, + 302 + ], + "type": "text", + "content": "4.2. Implementation Details" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 308, + 545, + 489 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 308, + 545, + 489 + ], + "spans": [ + { + "bbox": [ + 304, + 308, + 545, + 489 + ], + "type": "text", + "content": "We employ Deformable DETR [46] that improves the convergence speed of the DETR with ResNet-50 [6] as a backbone. It is worth noting that our approach can be extended to any object detector that incorporates self-attention mechanisms between object features, including models like DETR [1], Sparse R-CNN [32], and others. We follow the configurations of the original Deformable DETR, except that we use only 200 object queries. To speed up convergence, we first train the object detector with the target dataset and subsequently train the SGG task using the pretrained object detector, similar to prior work [17]. To calculate the overall loss in Equation 7, " + }, + { + "bbox": [ + 304, + 308, + 545, + 489 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{rel}}" + }, + { + "bbox": [ + 304, + 308, + 545, + 489 + ], + "type": "text", + "content": " is set to 15. " + }, + { + "bbox": [ + 304, + 308, + 545, + 489 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{con}}" + }, + { + "bbox": [ + 304, + 308, + 545, + 489 + ], + "type": "text", + "content": " is set to 30 and 90 for the Visual Genome and Open Image V6, respectively. We set " + }, + { + "bbox": [ + 304, + 308, + 545, + 489 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 304, + 308, + 545, + 489 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 304, + 308, + 545, + 489 + ], + "type": "inline_equation", + "content": "10^{-14}" + }, + { + "bbox": [ + 304, + 308, + 545, + 489 + ], + "type": "text", + "content": " for adaptive smoothing and both " + }, + { + "bbox": [ + 304, + 308, + 545, + 489 + ], + "type": "inline_equation", + "content": "k_{\\mathrm{neg}}" + }, + { + "bbox": [ + 304, + 308, + 545, + 489 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 308, + 545, + 489 + ], + "type": "inline_equation", + "content": "k_{\\mathrm{non}}" + }, + { + "bbox": [ + 304, + 308, + 545, + 489 + ], + "type": "text", + "content": " are set to 80." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 498, + 362, + 509 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 498, + 362, + 509 + ], + "spans": [ + { + "bbox": [ + 306, + 498, + 362, + 509 + ], + "type": "text", + "content": "4.3. Results" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 517, + 545, + 565 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 517, + 545, + 565 + ], + "spans": [ + { + "bbox": [ + 304, + 517, + 545, + 565 + ], + "type": "text", + "content": "We present quantitative results and perform a comparative analysis of our proposed framework with representative SGG models. Additional experimental results are included in the supplementary materials." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "text", + "content": "Visual Genome. Visual Genome results are shown in Tab. 1. Our proposed method demonstrates competitive performance with the current one-stage SGG models that have 1.5 to 6.5 times larger parameters with the fastest inference speed. In particular, our method achieves the highest object detection performance and shows comparable performance in the triplet detection compared to SSR-CNN [35], the state-of-the-art method. The results demonstrate that our method can generate scene graphs in an efficient and effective way by exploiting the by-products of the object detector. By applying logit adjustment [35], a technique for predicting tail predicate classes well, our" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "24234" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 47, + 70, + 547, + 282 + ], + "blocks": [ + { + "bbox": [ + 47, + 70, + 547, + 282 + ], + "lines": [ + { + "bbox": [ + 47, + 70, + 547, + 282 + ], + "spans": [ + { + "bbox": [ + 47, + 70, + 547, + 282 + ], + "type": "table", + "html": "
Model# params (M)FPSAP50R@20R@50R@100mR@20mR@50mR@100
two-stageIMP (EBM) [31, 39]322.22.028.118.125.931.22.84.25.4
VTransE [44]312.33.5-24.531.335.55.16.88.0
Motifs [42]369.91.928.125.132.136.94.15.56.8
VCTree [33]361.50.828.124.831.836.14.96.67.7
VCTree (TDE) [33, 34]361.30.828.114.019.423.26.99.311.1
VCTree (EBM) [31, 33]372.5-28.124.231.435.95.77.79.1
GPS-Net [19]----31.135.9-6.78.6
BGNN [16]341.91.729.023.331.035.87.510.712.6
one-stageFCSGG [20]87.16.028.516.121.325.12.73.64.2
RelTR [7]63.713.426.421.227.5-6.810.8-
SGTR [17]117.16.225.4-24.628.4-12.015.2
Relationformer [29]92.98.526.322.228.431.34.69.310.7
Iterative SGG [9]93.56.027.7†-29.732.1-8.08.8
SSR-CNN [35]274.34.023.8†25.832.736.96.18.410.0
SSR-CNN [35] LA,τ=0.3274.34.023.8†18.423.326.513.517.921.4
EGTR (Ours)42.514.730.823.530.234.35.57.910.1
EGTR (Ours) LA,τ=0.742.514.730.815.718.720.512.117.821.7
EGTR (Ours) LA,τ=0.542.514.730.819.724.226.711.017.121.4
EGTR (Ours) LA,τ=0.342.514.730.822.428.231.78.814.018.3
", + "image_path": "a16fe17cf96c9225171bc6281cd360eb2c9e469b302579803d440aef71ae7091.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 47, + 362, + 287, + 472 + ], + "blocks": [ + { + "bbox": [ + 46, + 289, + 547, + 356 + ], + "lines": [ + { + "bbox": [ + 46, + 289, + 547, + 356 + ], + "spans": [ + { + "bbox": [ + 46, + 289, + 547, + 356 + ], + "type": "text", + "content": "Table 1. Graph-Constraint results on the test set of Visual Genome dataset. We report the results of representative two-stage SGG models based on the Faster R-CNN [27] object detector with a ResNeXt-101-FPN [18, 38] backbone and concurrent one-stage SGG models. Among the one-stage models, the best is highlighted in bold, and the second-best is indicated with underlining. LA denotes logit adjustment proposed in SSR-CNN [35]. " + }, + { + "bbox": [ + 46, + 289, + 547, + 356 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 46, + 289, + 547, + 356 + ], + "type": "text", + "content": " represents that we concatenate the predicted subjects and objects and then apply non-maximum-suppression with a threshold of 0.3. Italic indicates the evaluation of metrics using publicly available model checkpoints. FPS is measured with a single V100 for images resized to a minimum of 600 for the shortest side and a maximum of 1000 for the longest side." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 47, + 362, + 287, + 472 + ], + "lines": [ + { + "bbox": [ + 47, + 362, + 287, + 472 + ], + "spans": [ + { + "bbox": [ + 47, + 362, + 287, + 472 + ], + "type": "table", + "html": "
Modelscoremicro-R@50wmAPrelwmAPphr
Motifs [42]38.971.629.931.6
VCTree [33]40.274.134.233.1
GPS-Net [19]41.774.832.934.0
BGNN [16]42.175.033.534.2
RelTR [7]43.071.734.237.5
SGTR [17]42.359.937.038.7
SSR-CNN [35]49.476.741.543.6
EGTR (Ours)48.675.042.041.9
", + "image_path": "7064226fa7dbe87e3156582cfc815f2a9dc9c4a9cd4fd7effeb5b3b0549607d9.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 477, + 287, + 501 + ], + "lines": [ + { + "bbox": [ + 47, + 477, + 287, + 501 + ], + "spans": [ + { + "bbox": [ + 47, + 477, + 287, + 501 + ], + "type": "text", + "content": "Table 2. Results on test set of Open Image V6. The score is a weighted sum of micro-R@50, " + }, + { + "bbox": [ + 47, + 477, + 287, + 501 + ], + "type": "inline_equation", + "content": "\\mathrm{wmAP}_{\\mathrm{rel}}" + }, + { + "bbox": [ + 47, + 477, + 287, + 501 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 47, + 477, + 287, + 501 + ], + "type": "inline_equation", + "content": "\\mathrm{wmAP}_{\\mathrm{phr}}" + }, + { + "bbox": [ + 47, + 477, + 287, + 501 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 510, + 287, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 510, + 287, + 594 + ], + "spans": [ + { + "bbox": [ + 46, + 510, + 287, + 594 + ], + "type": "text", + "content": "method performs favorably in the trade-off between " + }, + { + "bbox": [ + 46, + 510, + 287, + 594 + ], + "type": "inline_equation", + "content": "\\mathrm{R}@\\mathbf{k}" + }, + { + "bbox": [ + 46, + 510, + 287, + 594 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 510, + 287, + 594 + ], + "type": "inline_equation", + "content": "\\mathrm{mR}@\\mathbf{k}" + }, + { + "bbox": [ + 46, + 510, + 287, + 594 + ], + "type": "text", + "content": ", showing superiority over existing state-of-the-art models. Note that for triplet detection models [9, 35] without an explicit object detector, AP50 is measured by applying non-maximum-suppression (NMS) to the union of predicted subjects and objects set. AP50 performance is discussed further in the section 4.5." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 597, + 287, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 597, + 287, + 647 + ], + "spans": [ + { + "bbox": [ + 46, + 597, + 287, + 647 + ], + "type": "text", + "content": "Open Image V6. As depicted in Tab. 2, the experiments conducted on the Open Image V6 dataset also demonstrate competitive performance, underscoring the effectiveness and robustness of our method across different datasets." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 654, + 148, + 667 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 654, + 148, + 667 + ], + "spans": [ + { + "bbox": [ + 47, + 654, + 148, + 667 + ], + "type": "text", + "content": "4.4. Ablation Studies" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 673, + 287, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 673, + 287, + 696 + ], + "spans": [ + { + "bbox": [ + 47, + 673, + 287, + 696 + ], + "type": "text", + "content": "We analyze the effects of our model components. All of the ablation studies are done with the Visual Genome dataset." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 701, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 701, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 701, + 287, + 713 + ], + "type": "text", + "content": "Relation Sources. In Tab. 3, we investigate the source of" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 365, + 545, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 365, + 545, + 544 + ], + "spans": [ + { + "bbox": [ + 304, + 365, + 545, + 544 + ], + "type": "text", + "content": "relations used for the relation extractor. To verify the benefits of using attention queries and keys, we change " + }, + { + "bbox": [ + 304, + 365, + 545, + 544 + ], + "type": "inline_equation", + "content": "Q^l" + }, + { + "bbox": [ + 304, + 365, + 545, + 544 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 365, + 545, + 544 + ], + "type": "inline_equation", + "content": "K^l" + }, + { + "bbox": [ + 304, + 365, + 545, + 544 + ], + "type": "text", + "content": " in Eq. (3) to " + }, + { + "bbox": [ + 304, + 365, + 545, + 544 + ], + "type": "inline_equation", + "content": "Z^l" + }, + { + "bbox": [ + 304, + 365, + 545, + 544 + ], + "type": "text", + "content": ", the hidden states of each layer. Remarkably, when simply using the hidden states of all layers, the triplet detection performance decreases. Surprisingly, the performance is lower than that of using only " + }, + { + "bbox": [ + 304, + 365, + 545, + 544 + ], + "type": "inline_equation", + "content": "R_z" + }, + { + "bbox": [ + 304, + 365, + 545, + 544 + ], + "type": "text", + "content": " from the last layer. On the other hand, when only attention by-products are used as a relation source, the performance is similar to that of using only hidden states of the last layer, the most contextualized representations for detecting objects. Incorporating both attention by-products and final hidden states yields improved performance by amalgamating varied information. The results underscore that the attention by-products of the object detectors contain rich information for relation extraction." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 551, + 545, + 659 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 551, + 545, + 659 + ], + "spans": [ + { + "bbox": [ + 304, + 551, + 545, + 659 + ], + "type": "text", + "content": "Training Techniques. In Tab. 4, we ablate proposed techniques. Without any techniques, our model framework shows low performance. When each technique is applied, it substantially improves both R@50 and mR@50, and the best performance is achieved when all techniques are combined. The results demonstrate that our proposed techniques help to train the model framework effectively. Additional experimental results are provided in the supplementary materials." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 665, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 547, + 715 + ], + "type": "text", + "content": "Sampling Methodology. In Tab. 5, we investigate the impact of hard sampling methods for the negative and non-matching regions. Hard negatives and hard non-matchings show a trade-off that improves the performance of R@50" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "24235" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 48, + 70, + 288, + 133 + ], + "blocks": [ + { + "bbox": [ + 48, + 70, + 288, + 133 + ], + "lines": [ + { + "bbox": [ + 48, + 70, + 288, + 133 + ], + "spans": [ + { + "bbox": [ + 48, + 70, + 288, + 133 + ], + "type": "table", + "html": "
Rl^l a sourceRl^l aRzR@50mR@50
Ql & Kl✓✓30.27.9
Zl✓✓29.67.4
-✓29.97.6
Ql & Kl✓29.87.7
", + "image_path": "130b3fcdca977aff0793d48b4bd73ce8c49708db37c5250079679f4f3392cdfc.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 48, + 161, + 287, + 232 + ], + "blocks": [ + { + "bbox": [ + 85, + 138, + 249, + 149 + ], + "lines": [ + { + "bbox": [ + 85, + 138, + 249, + 149 + ], + "spans": [ + { + "bbox": [ + 85, + 138, + 249, + 149 + ], + "type": "text", + "content": "Table 3. Ablation study on relation source." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 48, + 161, + 287, + 232 + ], + "lines": [ + { + "bbox": [ + 48, + 161, + 287, + 232 + ], + "spans": [ + { + "bbox": [ + 48, + 161, + 287, + 232 + ], + "type": "table", + "html": "
adaptive smoothing\\( \\mathcal{L}_{\\text{con}} \\)samplingR@50mR@50
26.65.3
✓28.36.5
✓29.67.0
✓28.97.1
✓✓✓30.27.9
", + "image_path": "9da1d5b0b3751bc9139a7ea1b3f68db2ac24a2b648ee5e40c17b42a22563ab5c.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 48, + 260, + 287, + 312 + ], + "blocks": [ + { + "bbox": [ + 74, + 238, + 260, + 249 + ], + "lines": [ + { + "bbox": [ + 74, + 238, + 260, + 249 + ], + "spans": [ + { + "bbox": [ + 74, + 238, + 260, + 249 + ], + "type": "text", + "content": "Table 4. Ablation study on proposed techniques." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 48, + 260, + 287, + 312 + ], + "lines": [ + { + "bbox": [ + 48, + 260, + 287, + 312 + ], + "spans": [ + { + "bbox": [ + 48, + 260, + 287, + 312 + ], + "type": "table", + "html": "
hard negativehard non-matchingR@50mR@50
✓✓30.27.9
✓30.07.1
✓29.67.7
", + "image_path": "ca8a4a2118b2bf6c879fa9b0eeed7043ab378d3428d9275d525d53208d7986e0.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 339, + 287, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 339, + 287, + 483 + ], + "spans": [ + { + "bbox": [ + 46, + 339, + 287, + 483 + ], + "type": "text", + "content": "and mR@50, respectively, and additional performance improvements are achieved for both measures when they are used together. For hard negatives, we speculate that R@50 performance is improved by selecting negatives for the tail predicate class, which is likely to exist in reality but is not in the annotation. For hard non-matchings, mR@50 performance seems to have improved by choosing object candidates that are akin to the ground truth objects but unmatched with them. This encourages the model to predict diverse predicate classes by preventing duplicate object pairs, which are likely to predict head predicate classes, from being predicted." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 491, + 119, + 502 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 491, + 119, + 502 + ], + "spans": [ + { + "bbox": [ + 47, + 491, + 119, + 502 + ], + "type": "text", + "content": "4.5. Discussion" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 510, + 287, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 510, + 287, + 653 + ], + "spans": [ + { + "bbox": [ + 46, + 510, + 287, + 653 + ], + "type": "text", + "content": "Regarding the highest AP50 score, we posit that it is because our proposed method focuses on not only objects with relations but also objects without any relations. To validate this, we calculate AP50 for two subsets of the ground-truth objects: those with relations and those without relations. We compare our method with two triplet detection-based models. Tab. 6 illustrates that our method achieves significantly high " + }, + { + "bbox": [ + 46, + 510, + 287, + 653 + ], + "type": "inline_equation", + "content": "\\mathrm{AP50}_{\\mathrm{no - rel}}" + }, + { + "bbox": [ + 46, + 510, + 287, + 653 + ], + "type": "text", + "content": ". Considering that triplet detection performance is related to " + }, + { + "bbox": [ + 46, + 510, + 287, + 653 + ], + "type": "inline_equation", + "content": "\\mathrm{AP50}_{\\mathrm{rel}}" + }, + { + "bbox": [ + 46, + 510, + 287, + 653 + ], + "type": "text", + "content": ", while triplet detection-based models only focus on objects with relations to detect triplets well, our method is capable of extracting the complete scene graph, including objects without relations." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 653, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 287, + 714 + ], + "type": "text", + "content": "For further analysis, we visualize bounding boxes for subjects and objects from the top 20 predictions in Fig. 5. For SSR-CNN [35], which directly predicts triplets, we observe a significant number of overlapping bounding boxes. On the other hand, our proposed model leverages predic" + } + ] + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 306, + 70, + 545, + 121 + ], + "blocks": [ + { + "bbox": [ + 78, + 318, + 255, + 329 + ], + "lines": [ + { + "bbox": [ + 78, + 318, + 255, + 329 + ], + "spans": [ + { + "bbox": [ + 78, + 318, + 255, + 329 + ], + "type": "text", + "content": "Table 5. Ablation study on sampling methods." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 306, + 70, + 545, + 121 + ], + "lines": [ + { + "bbox": [ + 306, + 70, + 545, + 121 + ], + "spans": [ + { + "bbox": [ + 306, + 70, + 545, + 121 + ], + "type": "table", + "html": "
ModelAP50\\( AP50_{\\text{rel}} \\)\\( AP50_{\\text{no-rel}} \\)
Iterative SGG [9]†27.724.37.8
SSR-CNN [35]†23.820.27.4
EGTR (Ours)30.824.310.7
", + "image_path": "e58fc5b81d9148d9476052a5decff82b2473d1214d9d6f90d01e0685faefade3.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 127, + 547, + 194 + ], + "lines": [ + { + "bbox": [ + 305, + 127, + 547, + 194 + ], + "spans": [ + { + "bbox": [ + 305, + 127, + 547, + 194 + ], + "type": "text", + "content": "Table 6. AP50 for two subsets of objects. " + }, + { + "bbox": [ + 305, + 127, + 547, + 194 + ], + "type": "inline_equation", + "content": "\\mathrm{AP50}_{\\mathrm{rel}}" + }, + { + "bbox": [ + 305, + 127, + 547, + 194 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 127, + 547, + 194 + ], + "type": "inline_equation", + "content": "\\mathrm{AP50}_{\\mathrm{no - rel}}" + }, + { + "bbox": [ + 305, + 127, + 547, + 194 + ], + "type": "text", + "content": " denote scores evaluated only on objects having at least one edge and no edge, respectively. Note that each measure operates on its unique scale since the ground-truth objects used for each measure differ. " + }, + { + "bbox": [ + 305, + 127, + 547, + 194 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 305, + 127, + 547, + 194 + ], + "type": "text", + "content": " indicates that additional NMS is applied to the union of subjects and objects set." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 307, + 204, + 425, + 284 + ], + "blocks": [ + { + "bbox": [ + 307, + 204, + 425, + 284 + ], + "lines": [ + { + "bbox": [ + 307, + 204, + 425, + 284 + ], + "spans": [ + { + "bbox": [ + 307, + 204, + 425, + 284 + ], + "type": "image", + "image_path": "24bd9ff20bfc213d9d6ec0c5211af0c429e35efa9df27fb0358e28fccd39745f.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 307, + 284, + 425, + 361 + ], + "blocks": [ + { + "bbox": [ + 307, + 284, + 425, + 361 + ], + "lines": [ + { + "bbox": [ + 307, + 284, + 425, + 361 + ], + "spans": [ + { + "bbox": [ + 307, + 284, + 425, + 361 + ], + "type": "image", + "image_path": "18ae33dd39cb2090a7c41cafda42b6943398cbcb2bce6166315195b836fe3f0f.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 335, + 364, + 396, + 373 + ], + "lines": [ + { + "bbox": [ + 335, + 364, + 396, + 373 + ], + "spans": [ + { + "bbox": [ + 335, + 364, + 396, + 373 + ], + "type": "text", + "content": "(a) SSR-CNN [35]" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 425, + 205, + 542, + 284 + ], + "blocks": [ + { + "bbox": [ + 425, + 205, + 542, + 284 + ], + "lines": [ + { + "bbox": [ + 425, + 205, + 542, + 284 + ], + "spans": [ + { + "bbox": [ + 425, + 205, + 542, + 284 + ], + "type": "image", + "image_path": "46306f91c2dbb914d9f6359cae5848fe4882b7587c39fc49c4b75c459bcd9dfb.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 425, + 284, + 542, + 361 + ], + "blocks": [ + { + "bbox": [ + 425, + 284, + 542, + 361 + ], + "lines": [ + { + "bbox": [ + 425, + 284, + 542, + 361 + ], + "spans": [ + { + "bbox": [ + 425, + 284, + 542, + 361 + ], + "type": "image", + "image_path": "bcd320527c7035d3e2f40585be25959ccfdf58231515632ecd35636a8f686e10.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 456, + 364, + 512, + 373 + ], + "lines": [ + { + "bbox": [ + 456, + 364, + 512, + 373 + ], + "spans": [ + { + "bbox": [ + 456, + 364, + 512, + 373 + ], + "type": "text", + "content": "(b) EGTR (Ours)" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 305, + 378, + 545, + 411 + ], + "lines": [ + { + "bbox": [ + 305, + 378, + 545, + 411 + ], + "spans": [ + { + "bbox": [ + 305, + 378, + 545, + 411 + ], + "type": "text", + "content": "Figure 5. Comparison of detected subjects and objects. We select the top 20 predictions of SSR-CNN and EGTR and visualize bounding boxes of detected subjects and objects." + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "bbox": [ + 305, + 422, + 545, + 469 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 422, + 545, + 469 + ], + "spans": [ + { + "bbox": [ + 305, + 422, + 545, + 469 + ], + "type": "text", + "content": "tions from an explicit object detector, reducing redundant predictions. Furthermore, our method detects various subjects and objects that appear in the scene rather than focusing on objects that are more likely to have relations." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 306, + 480, + 378, + 492 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 480, + 378, + 492 + ], + "spans": [ + { + "bbox": [ + 306, + 480, + 378, + 492 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 500, + 545, + 668 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 500, + 545, + 668 + ], + "spans": [ + { + "bbox": [ + 304, + 500, + 545, + 668 + ], + "type": "text", + "content": "In this study, we propose the lightweight one-stage scene graph generator EGTR. We significantly reduce the model complexity by harnessing the relationships among the object queries learned from the self-attention layers of the object decoder. Furthermore, we devise a novel adaptive smoothing technique that helps multi-task learning of object detection and relation extraction by adjusting the relation label according to the object detection performance. As an auxiliary task of relation extraction, connectivity prediction contributes to the effective learning of EGTR. We conduct extensive experiments and ablation studies, and the results demonstrate that EGTR achieves the highest object detection performance and competitive triplet detection capabilities with the fastest inference speed." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 305, + 671, + 545, + 708 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 671, + 545, + 708 + ], + "spans": [ + { + "bbox": [ + 305, + 671, + 545, + 708 + ], + "type": "text", + "content": "Acknowledgements. We thank Taeho Kil, Geewook Kim, and the anonymous reviewers for their insightful comments and suggestions." + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "24236" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 134 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 134 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 134 + ], + "type": "text", + "content": "[1] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In ECCV, pages 213-229, 2020. 1, 2, 3, 4, 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 136, + 287, + 158 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 136, + 287, + 158 + ], + "spans": [ + { + "bbox": [ + 53, + 136, + 287, + 158 + ], + "type": "text", + "content": "[2] Zhe Chen, Shaoli Huang, and Dacheng Tao. Context refinement for object detection. In ECCV, pages 71-86, 2018. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 159, + 288, + 191 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 159, + 288, + 191 + ], + "spans": [ + { + "bbox": [ + 53, + 159, + 288, + 191 + ], + "type": "text", + "content": "[3] Yuren Cong, Michael Ying Yang, and Bodo Rosenhahn. Reltr: Relation transformer for scene graph generation. IEEE TPAMI, 2023. 1, 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 194, + 287, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 194, + 287, + 226 + ], + "spans": [ + { + "bbox": [ + 53, + 194, + 287, + 226 + ], + "type": "text", + "content": "[4] Naina Dhingra, Florian Ritter, and Andreas Kunz. Bgt-net: Bidirectional gru transformer network for scene graph generation. In CVPR, pages 2150-2159, 2021. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 228, + 287, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 228, + 287, + 259 + ], + "spans": [ + { + "bbox": [ + 53, + 228, + 287, + 259 + ], + "type": "text", + "content": "[5] Lizhao Gao, Bo Wang, and Wenmin Wang. Image captioning with scene-graph based semantic concepts. In ICMLC, pages 225-229, 2018. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 262, + 287, + 295 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 262, + 287, + 295 + ], + "spans": [ + { + "bbox": [ + 53, + 262, + 287, + 295 + ], + "type": "text", + "content": "[6] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, pages 770-778, 2016. 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 297, + 287, + 328 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 297, + 287, + 328 + ], + "spans": [ + { + "bbox": [ + 53, + 297, + 287, + 328 + ], + "type": "text", + "content": "[7] Han Hu, Jiayuan Gu, Zheng Zhang, Jifeng Dai, and Yichen Wei. Relation networks for object detection. In CVPR, pages 3588-3597, 2018. 2, 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 331, + 287, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 331, + 287, + 373 + ], + "spans": [ + { + "bbox": [ + 53, + 331, + 287, + 373 + ], + "type": "text", + "content": "[8] Justin Johnson, Ranjay Krishna, Michael Stark, Li-Jia Li, David Shamma, Michael Bernstein, and Li Fei-Fei. Image retrieval using scene graphs. In CVPR, pages 3668-3678, 2015. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 375, + 287, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 375, + 287, + 397 + ], + "spans": [ + { + "bbox": [ + 53, + 375, + 287, + 397 + ], + "type": "text", + "content": "[9] Siddhesh Khandelwal and Leonid Sigal. Iterative scene graph generation. In NeurIPS, 2022. 1, 3, 7, 8" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 399, + 287, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 399, + 287, + 441 + ], + "spans": [ + { + "bbox": [ + 48, + 399, + 287, + 441 + ], + "type": "text", + "content": "[10] Dong-Jin Kim, Jinsoo Choi, Tae-Hyun Oh, and In So Kweon. Dense relational captioning: Triple-stream networks for relationship-based captioning. In CVPR, pages 6271-6280, 2019. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 444, + 287, + 465 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 444, + 287, + 465 + ], + "spans": [ + { + "bbox": [ + 48, + 444, + 287, + 465 + ], + "type": "text", + "content": "[11] Rajat Koner, Suprosanna Shit, and Volker Tresp. Relation transformer network. ECCV, pages 422-439, 2022. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 467, + 287, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 467, + 287, + 521 + ], + "spans": [ + { + "bbox": [ + 48, + 467, + 287, + 521 + ], + "type": "text", + "content": "[12] Ranjay Krishna, Yuke Zhu, Oliver Groth, Justin Johnson, Kenji Hata, Joshua Kravitz, Stephanie Chen, Yannis Kalantidis, Li-Jia Li, David A Shamma, et al. Visual genome: Connecting language and vision using crowdsourced dense image annotations. IJCV, 123:32-73, 2017. 2, 5, 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 523, + 287, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 523, + 287, + 588 + ], + "spans": [ + { + "bbox": [ + 48, + 523, + 287, + 588 + ], + "type": "text", + "content": "[13] Alina Kuznetsova, Hassan Rom, Neil Alldrin, Jasper Uijlings, Ivan Krasin, Jordi Pont-Tuset, Shahab Kamali, Stefan Popov, Matteo Malloci, Alexander Kolesnikov, et al. The open images dataset v4: Unified image classification, object detection, and visual relationship detection at scale. IJCV, 128(7):1956-1981, 2020. 2, 6" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 590, + 287, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 590, + 287, + 623 + ], + "spans": [ + { + "bbox": [ + 48, + 590, + 287, + 623 + ], + "type": "text", + "content": "[14] Linjie Li, Zhe Gan, Yu Cheng, and Jingjing Liu. Relation-aware graph attention network for visual question answering. In ICCV, pages 10313-10322, 2019. 1" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 624, + 287, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 624, + 287, + 667 + ], + "spans": [ + { + "bbox": [ + 48, + 624, + 287, + 667 + ], + "type": "text", + "content": "[15] Lin Li, Guikun Chen, Jun Xiao, Yi Yang, Chunping Wang, and Long Chen. Compositional feature augmentation for unbiased scene graph generation. In ICCV, pages 21685-21695, 2023. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "type": "text", + "content": "[16] Rongjie Li, Songyang Zhang, Bo Wan, and Xuming He. Bipartite graph network with adaptive message passing for unbiased scene graph generation. In CVPR, pages 11109-11119, 2021. 1, 2, 7" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 33, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "text", + "content": "[17] Rongjie Li, Songyang Zhang, and Xuming He. Sgtr: End-to-end scene graph generation with transformer. In CVPR, pages 19486-19496, 2022. 1, 3, 6, 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 108, + 545, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 108, + 545, + 150 + ], + "spans": [ + { + "bbox": [ + 307, + 108, + 545, + 150 + ], + "type": "text", + "content": "[18] Tsung-Yi Lin, Piotr Dollar, Ross Girshick, Kaiming He, Bharath Hariharan, and Serge Belongie. Feature pyramid networks for object detection. In CVPR, pages 2117-2125, 2017. 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 153, + 545, + 186 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 153, + 545, + 186 + ], + "spans": [ + { + "bbox": [ + 307, + 153, + 545, + 186 + ], + "type": "text", + "content": "[19] Xin Lin, Changxing Ding, Jinquan Zeng, and Dacheng Tao. Gps-net: Graph property sensing network for scene graph generation. In CVPR, pages 3746-3753, 2020. 1, 2, 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 189, + 545, + 220 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 189, + 545, + 220 + ], + "spans": [ + { + "bbox": [ + 307, + 189, + 545, + 220 + ], + "type": "text", + "content": "[20] Hengyue Liu, Ning Yan, Masood Mortazavi, and Bir Bhanu. Fully convolutional scene graph generation. In CVPR, pages 11546-11556, 2021. 1, 3, 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 223, + 545, + 266 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 223, + 545, + 266 + ], + "spans": [ + { + "bbox": [ + 307, + 223, + 545, + 266 + ], + "type": "text", + "content": "[21] Wei Liu, Dragomir Anguelov, Dumitru Erhan, Christian Szegedy, Scott Reed, Cheng-Yang Fu, and Alexander C Berg. Ssd: Single shot multibox detector. In ECCV, pages 21-37, 2016. 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 269, + 545, + 312 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 269, + 545, + 312 + ], + "spans": [ + { + "bbox": [ + 307, + 269, + 545, + 312 + ], + "type": "text", + "content": "[22] Yong Liu, Ruiping Wang, Shiguang Shan, and Xilin Chen. Structure inference net: Object detection using scene-level context and instance-level relationships. In CVPR, pages 6985-6994, 2018. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 315, + 545, + 347 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 315, + 545, + 347 + ], + "spans": [ + { + "bbox": [ + 307, + 315, + 545, + 347 + ], + "type": "text", + "content": "[23] Cewu Lu, Ranjay Krishna, Michael Bernstein, and Li Fei-Fei. Visual relationship detection with language priors. In ECCV, pages 852–869, 2016. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 350, + 545, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 350, + 545, + 403 + ], + "spans": [ + { + "bbox": [ + 307, + 350, + 545, + 403 + ], + "type": "text", + "content": "[24] Yichao Lu, Himanshu Rai, Jason Chang, Boris Knyazev, Guangwei Yu, Shashank Shekhar, Graham W Taylor, and Maksims Volkovs. Context-aware scene graph generation with seq2seq transformers. In ICCV, pages 15931-15941, 2021." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 406, + 545, + 439 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 406, + 545, + 439 + ], + "spans": [ + { + "bbox": [ + 307, + 406, + 545, + 439 + ], + "type": "text", + "content": "[25] Yukuan Min, Among Wu, and Cheng Deng. Environment-invariant curriculum relation learning for fine-grained scene graph generation. In ICCV, pages 13296-13307, 2023. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 441, + 545, + 463 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 441, + 545, + 463 + ], + "spans": [ + { + "bbox": [ + 307, + 441, + 545, + 463 + ], + "type": "text", + "content": "[26] Alejandro Newell and Jia Deng. Pixels to graphs by associative embedding. NeurIPS, 30, 2017. 1, 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 464, + 545, + 497 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 464, + 545, + 497 + ], + "spans": [ + { + "bbox": [ + 307, + 464, + 545, + 497 + ], + "type": "text", + "content": "[27] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. NeurIPS, 28, 2015. 2, 7" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 499, + 545, + 565 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 499, + 545, + 565 + ], + "spans": [ + { + "bbox": [ + 307, + 499, + 545, + 565 + ], + "type": "text", + "content": "[28] Sebastian Schuster, Ranjay Krishna, Angel Chang, Li Fei-Fei, and Christopher D Manning. Generating semantically precise scene graphs from textual descriptions for improved image retrieval. In Proceedings of the fourth workshop on vision and language, pages 70–80. Association for Computational Linguistics, 2015. 1" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 567, + 545, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 567, + 545, + 622 + ], + "spans": [ + { + "bbox": [ + 307, + 567, + 545, + 622 + ], + "type": "text", + "content": "[29] Suprosanna Shit, Rajat Koner, Bastian Wittmann, Johannes Paetzold, Ivan Ezhov, Hongwei Li, Jiazhen Pan, Sahand Sharifzadeh, Georgios Kaissis, Volker Tresp, et al. Relationformer: A unified framework for image-to-graph generation. In ECCV, pages 422-439, 2022. 1, 3, 7" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 307, + 624, + 545, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 624, + 545, + 666 + ], + "spans": [ + { + "bbox": [ + 307, + 624, + 545, + 666 + ], + "type": "text", + "content": "[30] Gopika Sudhakaran, Devendra Singh Dhami, Kristian Kersting, and Stefan Roth. Vision relation transformer for unbiased scene graph generation. In ICCV, pages 21882-21893, 2023. 2" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 307, + 669, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 669, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 307, + 669, + 545, + 712 + ], + "type": "text", + "content": "[31] Mohammed Suhail, Abhay Mittal, Behjat Siddiquie, Chris Broaddus, Jayan Eledath, Gerard Medioni, and Leonid Sigal. Energy-based learning for scene graph generation. In CVPR, pages 13936-13945, 2021. 7" + } + ] + } + ], + "index": 32 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "24237" + } + ] + } + ], + "index": 34 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 624 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "text", + "content": "[32] Peize Sun, Rufeng Zhang, Yi Jiang, Tao Kong, Chenfeng Xu, Wei Zhan, Masayoshi Tomizuka, Lei Li, Zehuan Yuan, Changhu Wang, and Ping Luo. Sparse r-cnn: End-to-end object detection with learnable proposals. In CVPR, pages 14454-14463, 2021. 2, 3, 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 129, + 287, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 129, + 287, + 171 + ], + "spans": [ + { + "bbox": [ + 48, + 129, + 287, + 171 + ], + "type": "text", + "content": "[33] Kaihua Tang, Hanwang Zhang, Baoyuan Wu, Wenhan Luo, and Wei Liu. Learning to compose dynamic tree structures for visual contexts. In CVPR, pages 6619-6628, 2019. 1, 2, 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 173, + 287, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 173, + 287, + 207 + ], + "spans": [ + { + "bbox": [ + 48, + 173, + 287, + 207 + ], + "type": "text", + "content": "[34] Kaihua Tang, Yulei Niu, Jianqiang Huang, Jiaxin Shi, and Hanwang Zhang. Unbiased scene graph generation from biased training. In CVPR, pages 3716-3725, 2020. 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 208, + 287, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 208, + 287, + 239 + ], + "spans": [ + { + "bbox": [ + 48, + 208, + 287, + 239 + ], + "type": "text", + "content": "[35] Yao Teng and Limin Wang. Structured sparse r-cnn for direct scene graph generation. In CVPR, pages 19437-19446, 2022. 1, 3, 6, 7, 8" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 241, + 287, + 283 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 241, + 287, + 283 + ], + "spans": [ + { + "bbox": [ + 48, + 241, + 287, + 283 + ], + "type": "text", + "content": "[36] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. NeurIPS, 30, 2017. 2, 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 286, + 287, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 286, + 287, + 319 + ], + "spans": [ + { + "bbox": [ + 48, + 286, + 287, + 319 + ], + "type": "text", + "content": "[37] Sanghyun Woo, Dahun Kim, Donghyeon Cho, and In So Kweon. Linknet: Relational embedding for scene graph. NeurIPS, 31, 2018. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 320, + 287, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 320, + 287, + 353 + ], + "spans": [ + { + "bbox": [ + 48, + 320, + 287, + 353 + ], + "type": "text", + "content": "[38] Saining Xie, Ross Girshick, Piotr Dollar, Zhuowen Tu, and Kaiming He. Aggregated residual transformations for deep neural networks. In CVPR, pages 1492-1500, 2017. 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 354, + 287, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 354, + 287, + 387 + ], + "spans": [ + { + "bbox": [ + 48, + 354, + 287, + 387 + ], + "type": "text", + "content": "[39] Danfei Xu, Yuke Zhu, Christopher B Choy, and Li Fei-Fei. Scene graph generation by iterative message passing. In CVPR, pages 5410–5419, 2017. 1, 2, 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 388, + 287, + 421 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 388, + 287, + 421 + ], + "spans": [ + { + "bbox": [ + 48, + 388, + 287, + 421 + ], + "type": "text", + "content": "[40] Jianwei Yang, Jiasen Lu, Stefan Lee, Dhruv Batra, and Devi Parikh. Graph r-cnn for scene graph generation. In ECCV, pages 670-685, 2018. 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 422, + 287, + 454 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 422, + 287, + 454 + ], + "spans": [ + { + "bbox": [ + 48, + 422, + 287, + 454 + ], + "type": "text", + "content": "[41] Xu Yang, Kaihua Tang, Hanwang Zhang, and Jianfei Cai. Auto-encoding scene graphs for image captioning. In CVPR, pages 10685–10694, 2019. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 456, + 287, + 488 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 456, + 287, + 488 + ], + "spans": [ + { + "bbox": [ + 48, + 456, + 287, + 488 + ], + "type": "text", + "content": "[42] Rowan Zellers, Mark Yatskar, Sam Thomson, and Yejin Choi. Neural motifs: Scene graph parsing with global context. In CVPR, pages 5831-5840, 2018. 1, 2, 6, 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 490, + 287, + 522 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 490, + 287, + 522 + ], + "spans": [ + { + "bbox": [ + 48, + 490, + 287, + 522 + ], + "type": "text", + "content": "[43] Cheng Zhang, Wei-Lun Chao, and Dong Xuan. An empirical study on leveraging scene graphs for visual question answering. In BMVC, 2019. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 523, + 287, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 523, + 287, + 555 + ], + "spans": [ + { + "bbox": [ + 48, + 523, + 287, + 555 + ], + "type": "text", + "content": "[44] Hanwang Zhang, Zawlin Kyaw, Shih-Fu Chang, and Tat-Seng Chua. Visual translation embedding network for visual relation detection. In CVPR, page 3107-3115, 2017. 7" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 557, + 287, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 557, + 287, + 590 + ], + "spans": [ + { + "bbox": [ + 48, + 557, + 287, + 590 + ], + "type": "text", + "content": "[45] Chaofan Zheng, Xinyu Lyu, Yuyu Guo, Pengpeng Zeng, Jingkuan Song, and Lianli Gao. Learning to generate scene graph from head to tail. In ICME, pages 1-6. IEEE, 2022. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 591, + 287, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 591, + 287, + 624 + ], + "spans": [ + { + "bbox": [ + 48, + 591, + 287, + 624 + ], + "type": "text", + "content": "[46] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable detr: Deformable transformers for end-to-end object detection. In ICLR, 2021. 2, 3, 6" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "24238" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/EMAGE_ Towards Unified Holistic Co-Speech Gesture Generation via Expressive Masked Audio Gesture Modeling/f8a85da7-9221-4a37-baf0-e5a3b46705e3_content_list.json b/2024/EMAGE_ Towards Unified Holistic Co-Speech Gesture Generation via Expressive Masked Audio Gesture Modeling/f8a85da7-9221-4a37-baf0-e5a3b46705e3_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..89068e82435bc28509a1d0cfbe4c3a4f1578ca34 --- /dev/null +++ b/2024/EMAGE_ Towards Unified Holistic Co-Speech Gesture Generation via Expressive Masked Audio Gesture Modeling/f8a85da7-9221-4a37-baf0-e5a3b46705e3_content_list.json @@ -0,0 +1,1976 @@ +[ + { + "type": "text", + "text": "EMAGE: Towards Unified Holistic Co-Speech Gesture Generation via Expressive Masked Audio Gesture Modeling", + "text_level": 1, + "bbox": [ + 130, + 128, + 841, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Haiyang Liu $^{1*}$ Zihao Zhu $^{2*}$ Giorgio Becherini $^{4}$ Yichen Peng $^{5}$ Mingyang Su $^{6}$ You Zhou $^{3}$ Xuefei Zhe $^{3}$ Naoya Iwamoto $^{3}$ Bo Zheng $^{3}$ Michael J. Black $^{3}$", + "bbox": [ + 143, + 200, + 821, + 241 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ The University of Tokyo $^{2}$ Keio University $^{3}$ Independent Researcher $^{4}$ Max Planck Institute for Intelligent Systems $^{5}$ JAIST $^{6}$ Tsinghua University", + "bbox": [ + 155, + 255, + 789, + 292 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/77aa6d75a5af7f91c15f89f767785f7436b00b41c1d6074cbaa3e215dafb7c19.jpg", + "image_caption": [ + "Figure 1. EMAGE. We present a Masked Audio-Conditioned Gesture Modeling framework, along with a new holistic gesture dataset, BEAT2 (BEAT-SMPLX-FLAME), for jointly generating facial expressions, local body dynamics, hand movements and global translations, conditioned on audio and a partially or completely masked gestures. The gray denotes visible gestures, and blue represents our outputs." + ], + "image_footnote": [], + "bbox": [ + 116, + 327, + 857, + 505 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 565, + 313, + 583 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We propose EMAGE, a framework to generate full-body human gestures from audio and masked gestures, encompassing facial, local body, hands, and global movements. To achieve this, we first introduce BEAT2 (BEATSMLPX-FLAME), a new mesh-level holistic co-speech dataset. BEAT2 combines a MoShed SMPL-X body with FLAME head parameters and further refines the modeling of head, neck, and finger movements, offering a community-standardized, high-quality 3D motion captured dataset. EMAGE leverages masked body gesture priors during training to boost inference performance. It involves a Masked Audio Gesture Transformer, facilitating joint training on audio-to-gesture generation and masked gesture reconstruction to effectively encode audio and body gesture hints. Encoded body hints from masked gestures are then separately employed to generate facial and body movements. Moreover, EMAGE adaptively merges speech features from the audio's rhythm and content and utilizes four compositional VQ-VAEs to enhance the results' fidelity and diversity. Experiments demonstrate that EMAGE generates", + "bbox": [ + 75, + 598, + 473, + 902 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "holistic gestures with state-of-the-art performance and is flexible in accepting predefined spatial-temporal gesture inputs, generating complete, audio-synchronized results. Our code and dataset are available.1", + "bbox": [ + 496, + 566, + 893, + 628 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 500, + 657, + 633, + 672 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The path towards full-body co-speech gesture generation presents many remaining challenges. Despite the development of various baselines, e.g., audio conditioned models for face [15, 20, 61], body [13, 25, 67], and hand movements [33, 39, 41], along with a few attempts at merged models [28, 65], the limited availability of comprehensive data and models poses an ongoing obstacle. To make progress in this direction, we present a new framework that can incorporate partial spatial-temporal predefined gestures and autonomously complete the remaining frames in synchronization with the audio. This provides a new tool for creating realistic animations of digital humans", + "bbox": [ + 496, + 681, + 893, + 866 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Equal Contribution. \n1https://pantomatrix.github.io/EMAGE/", + "bbox": [ + 522, + 875, + 810, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "1144", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 0 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[57, 63, 64, 71].", + "To achieve this, we first require a comprehensive gesture dataset. While several datasets are available for audio-to-body [67], audio-to-body&face [28], body-to-hands [39, 41], and common actions [16, 46], integrating them poses challenges due to varying data formats and evaluation metrics across different sub-areas. This motivates us to establish a unified, community-standard benchmark for training and evaluating holistic gestures. We build upon the BEAT dataset [39], which provides the most comprehensive mocap co-speech gesture dataset with diverse modalities. Despite its size and richness, the data in BEAT is not presented in a standardized format. The dataset comprises iPhone ARKit blendshape weights and Vicon skeletons, posing challenges when using the data for training; e.g., the lack of a mesh representation prevents the use of a vertex loss [19, 52]. Additionally, conducting direct human perceptual evaluations on these diverse modalities is inherently complex. On the other hand, SMPL-X [50] and FLAME [35] are common mesh standards for the academic community for other human dynamics-related tasks, e.g., action [52] and talking head generation [14]. Driven by the motivation to facilitate the sharing of knowledge across different tasks, we present BEAT-SMPLX-FLAME (BEAT2), which consists of two main components: i) Refined SMPL-X body shape and pose parameters from MoSh++ [44, 46] with hard-coded physical priors. ii) Optimized high-quality FLAME head parameters. By integrating SMPL-X's body and FLAME's head, BEAT2 facilitates a comprehensive training and evaluation benchmark across multiple sub-areas of co-speech human animation generation.", + "With the holistic data from BEAT2, we aim to enhance the coherence of individual body parts while ensuring accurate cross-modal alignment between audio and motion; see Figure 1. This leads to the design of Expressive Masked Audio-conditioned GEsture modeling, denoted as EMAGE, a spatial and temporal transformer-based framework. EMAGE first aggregates spatial features from masked body joints. Then, it reconstructs the latent space of pretrained gestures with a switchable gesture temporal self-attention and audio-gesture cross-attention. The selection of different forward paths enables modeling the gesture-to-gesture and audio-to-gesture prior separately and effectively. Once the reconstructed latent features are obtained, EMAGE decodes local face and body gestures from four compositional pretrained Vector Quantized-Variational AutoEncoders (VQ-VAEs) and decodes global translations from a pretrained Global Motion Predictor. The compositional VQ-VAEs are key to preserving audio-related motions during the reconstruction. Through these designs, EMAGE achieves state-of-the-art performance in generating body and face gestures. It takes as input audio and partially initialized gestures, and recovers audio" + ], + "bbox": [ + 76, + 90, + 472, + 902 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "synchronized, coherent gestures. Additionally, we show how EMAGE can flexibly incorporate additional, nonholistic, datasets to improve results.", + "Overall, our contributions are as follows: (1) We release BEAT2, a representation-unified, mesh-level dataset that leverages MoShed SMPL-X body and FLAME head parameters. (2) We propose EMAGE, a simple yet effective holistic gesture generation framework that generates coherent gestures with partial gesture and audio priors. (3) EMAGE achieves state-of-the-art performance in generating both body and face gestures with only four frames of seed gestures. (4) EMAGE showcases the use of additional non-holistic gesture datasets for training, e.g., Trinity and AMASS, to further improve the fidelity and diversity of the results, effectively leveraging data from different datasets." + ], + "bbox": [ + 496, + 90, + 893, + 318 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 330, + 640, + 347 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Co-Speech Animation Datasets are categorized into two types; see Table 1: pseudo-labeled (PGT) and motion-captured (mocap). For PGT, the Speech2Gesture Dataset [25] utilizes OpenPose [11] to extract 2D poses from News and Teaching videos. Subsequent works extend the dataset to 3D pose [28] and SMPL-X [65]. The TED Dataset [66], which extracts 2D pose from TED videos, has been extended to include 3D pose [67], fingers [40], and SMPL-X [45]. Similarly, 2D, 3D landmarks, and meshes are estimated from multi-view recorded videos [60] as PGT for talking-head generation. Although pseudo-labeled approaches could theoretically extract infinite data, their accuracy is limited. For instance, the average error (in mm) for the SOTA monocular 3D pose estimation algorithm [23] on the Human3.6M dataset [30] is 33.4, while for Vicon mocap it is 0.142 [47]. For mocap datasets, Trinity [3] features one male actor with 4h of data, and TalkingWithHands [32] collects conversation scenarios for two speakers. ZEEG [24] considers one speaker with 12 styles. For the face, 3D scanning is accurate but limited in quantity due to cost, e.g., less than 3h for BIWI [21], VOCA [14], and MeshTalk [56]. This leads to a balance between performance and quantity in the ARKit dataset [51]. The aforementioned datasets address the body and face separately. On the other hand, BEAT [39] includes both 3D pose body and ARKit [8] facial blendshape weights for the first time. However, BEAT lacks mesh data.", + "Co-Speech Animation Models are categorized based on their outputs; see Table 2. In addition to the baselines and datasets [14, 25, 32, 48, 56, 67], several frameworks for body gestures [2, 33, 53, 54] have improved the performance of the baselines. They are trained and evaluated by selecting specific upper body joints [4, 5, 38-40, 49, 70], or all body joints [62, 72]. Recent improvements in facial gestures drive the vertices with transformers and discrete face priors [19, 61]. However, these methods only address ei" + ], + "bbox": [ + 496, + 357, + 893, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "1145", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "table", + "img_path": "images/66e6b8d9cd0bc32106e05b2751c2989dd9fc80f1283577f7cd48d24a19ffe185.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Trinity Mocap 2017 [22]S2G S2G-2D 2019 [25]Seq2Seq TED-2D 2019 [66]TWH Mocap 2019 [32]Trimodal TED-3D 2020 [67]VOCA Scan 2020 [14]MeshTalk Scan 2021 [56]Hahibie et al. S2G-3D 2021 [28]HA2G TED-3D+ 2022 [40]BEAT Mocap 2022 [39]ZEEG Mocap 2022 [24]Yoon. et al. TED-SMPL 2023 [45]Talkshow S2G-SMPL 2023 [65]BEAT2 Ours 2024
Head-2D---3D Scan3D Scan3D-ARKit--PGT-MeshMC-Mesh
Upper Body3D2D2D3D3D--3D3D3D3DPGT-MeshPGT-MeshMC-Mesh
Hands3D2D-3D---3D3D3D3DPGT-MeshPGT-MeshMC-Mesh
Lower Body3D--3D-----3D3D--MC-Mesh
Global Motion3D--3D-----3D3D--MC-Mesh
Duration (hours)4609720970.523833764302760
", + "bbox": [ + 81, + 87, + 883, + 170 + ], + "page_idx": 2 + }, + { + "type": "table", + "img_path": "images/e89e22ee13f61b8b34cf4be94f8611a9fa64537145333d53cd95adc63906479a.jpg", + "table_caption": [ + "Table 1. Comparison of Co-Speech Gesture Datasets. We summarize gesture and face datasets in talking scenarios. 'PGT' and 'MC' denote Pseudo Ground Truth and Motion Capture, respectively. The BEAT2 (Ours) dataset is the largest dataset for motion-captured data, providing holistic, academic community standard mesh-level information." + ], + "table_footnote": [], + "table_body": "
S2GTriModalB2HTWHHahibie et al.DisCoHA2GFace Former 2022[19]RhythmicBEAT CaMN 2022[39]Diff Gesture 2023[72]Code Talker 2023[61]Talk Show 2023[65]DiffStyle Gesture 2023[62]Body Former 2023[49]Lively Speaker 2023[70]EMAEG (Ours) 2024
2019 [25]2020 [67]2020[48]2021[32]2021[28]2022[38]2022[40]2022[5]2022[5]-------Masked
Input Gestures--BodyBody------------
Head---M--S---SS---C
UpperSS-MSC-SCS-CSSSC
Hands--SSMSC--CS-CS-SC
Lower + Global------------S--C
", + "bbox": [ + 81, + 228, + 883, + 296 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Table 2. Comparison of Co-Speech Gesture Models. We compare with previous methods for generating face or body motion trained on co-speech datasets. The first row lists their inputs, and the subsequent rows list their outputs, respectively. Different decoder designs are denoted by the initials 'S' for Single, 'M' for Multiple, and 'C' for Cascaded. To the best of our knowledge, EMAGE (Ours) is the first to accept audio and partially or completely masked gestures, generating full-body audio-synchronized results.", + "bbox": [ + 75, + 301, + 893, + 357 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "ther the face or body. Applying these methods to the full body will yield sub-optimal results as audio is differently correlated with face and body dynamics. Most similar to our work, Habibie et al. [28] employ a single audio encoder and multiple decoders for generating face and body gestures. TalkSHOW [65] demonstrates the advantage of separating the audio encoders and decoding the body and hands using quantized codes in an auto-regressive manner. However, it lacks lower body and global motion, and there is no shared information for body and facial gesture generation. Moreover, it cannot accept partially masked body hints due to the design of a fully auto-regressive model.", + "bbox": [ + 75, + 383, + 470, + 565 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Masked Representation Learning was first demonstrated to be effective in Natural Language Processing, with BERT-based models [17, 31, 42] boosting the performance of learned word embeddings in downstream tasks through a combination of masked language modeling and transformer architecture. Subsequently, Masked AutoEncoders [29] expanded masked image modeling to computer vision by removing and inpainting image patches. This concept of masked representation learning has since been employed in other modalities, e.g., video [12, 18, 43], audio [26, 36, 37], and point clouds [55, 68, 69]. Most related to our work, MotionBERT [73] proposes a spatial-temporal transformer to learn a robust motion representation for classification-based tasks by masking 2D pose. In contrast to their method, we target robust motion features for conditional motion generation tasks, requiring a balance in training between multiple modalities, e.g., audio and gesture.", + "bbox": [ + 75, + 566, + 470, + 825 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. BEAT2", + "text_level": 1, + "bbox": [ + 76, + 843, + 163, + 858 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "This section introduces how we obtain unified mesh-level data, i.e., SMPL-X [50] and FLAME [35] parameters, from", + "bbox": [ + 75, + 869, + 470, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "the original BEAT dataset [39]. BEAT utilizes a Vicon motion capture system [47] and releases 78-joint skeleton-level Biovision Hierarchy (BVH) [10] motion files. Their facial capture system uses ARKit [8] with a depth camera on the iPhone 12 Pro, extracting $\\mathbb{R}^{51}$ blendshape weights. These blendshapes are designed based on the Facial Action Coding System (FACS), which is widely adopted. However, both the motion and facial data lack mesh-level details (see Figure 2), e.g., shapes and vertices.", + "bbox": [ + 496, + 383, + 893, + 521 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Body Parameters Initialization via MoSh++", + "text_level": 1, + "bbox": [ + 498, + 542, + 870, + 559 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We initialize the SMPL-X body shape and pose parameters from BEAT mocap marker data using MoSh++ [46]. Given the captured markers positions $\\mathbf{m} \\in \\mathbb{R}^{T \\times K \\times 3}$ , predefined markers position offsets $\\mathbf{d} \\in \\mathbb{R}^{K \\times 3}$ , and user-defined vertices-to-markers function $\\mathcal{H}$ , we aim to obtain body shape $\\beta \\in \\mathbb{R}^{300}$ , poses $\\theta \\in \\mathbb{R}^{T \\times 55 \\times 3}$ , and translation parameters $\\gamma \\in \\mathbb{R}^{T \\times 3}$ . The optimization uses a differentiable surface vertex mapping function $S(\\beta, \\theta, \\gamma)$ and vertex normal function $\\mathcal{N}(\\beta, \\theta_0, \\gamma_0)$ . For each frame, the latent marker $\\tilde{\\mathbf{m}} \\in \\mathbb{R}^{T \\times K \\times 3}$ is calculated as", + "bbox": [ + 496, + 569, + 890, + 720 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {\\mathbf {m}} _ {i} \\equiv \\mathcal {S} _ {\\mathcal {H}} (\\beta , \\theta_ {i}, \\gamma_ {i}) + \\mathbf {d N} _ {\\mathcal {H}} (\\beta , \\theta_ {i}, \\gamma_ {i}). \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 565, + 742, + 890, + 760 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We first select 12 frames to optimize and fix $\\mathbf{d}$ and $\\beta$ , then optimize $\\theta_{i}$ , $\\gamma_{i}$ for $i \\in (1:T)$ by minimizing the loss terms based on $\\|\\tilde{\\mathbf{m}}_i - \\mathbf{m}_i\\|^2$ , including Data Term, Surface Distance Energy, Marker Initialization Regularization, Pose and Shape Priors, Velocity Constancy, and Soft-Tissue Term (see supplementary materials). The overall objective function is the weighted sum of these terms to balance accuracy and plausibility.", + "bbox": [ + 496, + 779, + 893, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "1146", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/1a29c1ea8463bf00c2c3f184a817a4f1c708628a8ef1d47d76d7f19f87bcfc5d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 127, + 90, + 197, + 220 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/0d3097e295ab5f39d373170046e3ede29bd44caca6a90f7a1e16e8cbbc74f8e2.jpg", + "image_caption": [ + "BEAT" + ], + "image_footnote": [], + "bbox": [ + 133, + 224, + 197, + 273 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/3cf8e0bff276205e84046de305c5ed3d778dd92008b38ecc417fbe3aa563ac59.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 202, + 92, + 297, + 220 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/9c115889ca59e8c11153c9150f32df01371750dbb710af599fab43cfc7f8e73d.jpg", + "image_caption": [ + "Retargeted" + ], + "image_footnote": [], + "bbox": [ + 202, + 224, + 297, + 273 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/0b2f09f2374956161351194f45e2e2532c9d05df8cf0f908d5ce265ba34ba94c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 308, + 92, + 398, + 220 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/f8f4640aefd7e52843ec5ee7f210087f3efb995a8208587356d71d31ebcdc3c4.jpg", + "image_caption": [ + "Moshed", + "Figure 2. Comparison of Data between BEAT2 and Others. BEAT-SMPLX-FLAME presents a new mesh-level, motion-captured, holistic co-speech gesture dataset with 60h of data. Left: We compare our refined SMPL-X body parameters (denoted as Refined MoSh) with the original BEAT skeleton [39], Retargeted from AutoRegPro, and initial results of $\\mathrm{Mosh}++$ [50]. The refined results show correct neck flexion, appropriate head and neck shape ratios, and detailed finger representations. Right: Visualization of blendshape weights from the original BEAT dataset [39] with ARKit's template, Wrapped-based, and handcrafted optimization. Our final handcrafted FLAME blendshape-based optimization demonstrates both accurate lip movement details and natural mouth shapes." + ], + "image_footnote": [], + "bbox": [ + 308, + 224, + 403, + 273 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/3c1b9aa5d1182a6c0f9d831ead367f4923e1919823cf79b76dca4f59fb4552c0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 408, + 92, + 506, + 220 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/998179254e8dbd0a0cd08de2c91b898bf38655be992f223fb73fe5128e9c24bd.jpg", + "image_caption": [ + "Refined Moshed (Ours)" + ], + "image_footnote": [], + "bbox": [ + 408, + 223, + 506, + 273 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/2e2280169812212b1d236e2a6a3e66b5ab26bd14574a2bb1d9abdb43d01b8b65.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 524, + 95, + 612, + 220 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/0823b4da28a4d4e58bba88dec90cc1bb5f4de875f981e5c1269d82ed5df385a6.jpg", + "image_caption": [ + "BEAT" + ], + "image_footnote": [], + "bbox": [ + 524, + 224, + 617, + 273 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/0a5a6dee70f74b4a51ebca6acff78622320dfb2cf09a947e6ba026a25c523992.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 620, + 98, + 718, + 220 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/4e2e0a6d3965907893b8e602a4bd4865e5ecea413958654dd3f083be29da6d78.jpg", + "image_caption": [ + "Wrapped Optimum" + ], + "image_footnote": [], + "bbox": [ + 620, + 224, + 722, + 273 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/aa583dca0754a84b1df2c39c7b1147b4db1387c0d69b19fcce8f8d91e9e9466b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 725, + 98, + 826, + 220 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/3da42aa9860e70ce3b694a6054783d9f404c5a344355f108f4502a33f7839eda.jpg", + "image_caption": [ + "Handcraft Optimum (Ours)" + ], + "image_footnote": [], + "bbox": [ + 727, + 224, + 831, + 273 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Body Parameters Refinement", + "text_level": 1, + "bbox": [ + 76, + 393, + 339, + 409 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "MoSh++ produces unnatural head shapes due to the fact that head markers were worn on a helmet and it sometimes produces unnatural finger poses. Consequently, we refine body shape and pose parameters with three simple yet effective hard-coded physical rules. i) The neck and head length should approximate 1/7 of the total length of the body. ii) Fingers, except for the thumb, should not bend backward. iii) We employ the Kolmogorov-Smirnov test, which reveals the data is similar to a normal distribution. Subsequently, we apply a Gaussian truncation approach, where all data points falling outside the $3\\sigma$ range are adjusted to align with the $3\\sigma$ threshold and blended with the adjacent 10 frames. We compare our refined body parameters (denoted as MoSh Refined) with Original BEAT, Retargeted SMPL-X, and MoSh SMPL-X in Figure 2.", + "bbox": [ + 75, + 422, + 472, + 650 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Blendshape Weights to FLAME Parameters", + "text_level": 1, + "bbox": [ + 76, + 675, + 450, + 691 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Given the ARKit blendshape weights $\\mathbf{b}_{\\mathrm{ARKit}} \\in \\mathbb{R}^{T \\times 51}$ , we aim to derive a transformation matrix $\\mathbf{W} \\in \\mathbb{R}^{51 \\times 103}$ that maps these into FLAME parameters $\\mathbf{b}_{\\mathrm{FLAME}} \\in \\mathbb{R}^{T \\times (100 + 3)}$ , where 100 represents the number of dimensions for expression parameters and 3 for jaw movement. Due to discrepancies between the iPhone's and FLAME's template mesh topology, simply optimizing FLAME parameters by minimizing the geometric error between the ARKit template vertices and the wrapped FLAME vertices does not yield satisfactory results. Instead, we release a set of handcrafted blendshape templates $\\mathbf{v}_{\\mathrm{template}} \\in \\mathbb{R}^{52}$ on FLAME following ARKit's FACS configuration. This approach allows direct driving of the FLAME topology ver", + "bbox": [ + 75, + 703, + 472, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "tices $\\mathbf{v}\\in \\mathbb{R}^{T\\times 5023\\times 3}$ from the given blendshape weights,", + "bbox": [ + 498, + 392, + 879, + 410 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {v} = \\mathbf {v} _ {\\text {t e m p l a t e}} ^ {0} + \\sum_ {j = 1} ^ {5 1} \\mathbf {b} _ {\\text {A R K i t}, j} \\cdot \\mathbf {v} _ {\\text {t e m p l a t e}} ^ {j} \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 571, + 422, + 890, + 465 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathbf{b}_{\\mathrm{ARKit},j}$ is the weight of the $j$ -th ARKit blend-shape and $\\mathbf{v}_{\\mathrm{template}}^j$ is the vertex position of the FLAME model template corresponding to the $j$ -th blendshape. The term $\\mathbf{v}_{\\mathrm{template}}^0$ represents the original template vertex position of the FLAME model. We optimize $\\mathbf{W}$ by minimizing $\\| \\tilde{\\mathbf{v}}_j - \\mathbf{v}_j\\| _2$ , where $\\tilde{\\mathbf{v}}$ is obtained from FLAME's LBS $\\mathcal{V}(\\mathbf{b}_{\\mathrm{FLAME}})$ . The comparisons between ARKit data, wrap-based, and our approach are shown in Figure 2.", + "bbox": [ + 498, + 477, + 890, + 601 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. IMAGE", + "text_level": 1, + "bbox": [ + 500, + 617, + 598, + 632 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We introduce the details of EMAGE, Expressive Masked Audio-Conditioned GEsture Modeling (see Figure 3). Given gestures $\\mathbf{g} \\in \\mathbb{R}^{T \\times (55 \\times 6 + 100 + 4 + 3)}$ , representing 55 joints in Rot6D, $\\mathbb{R}^{100}$ FLAME parameters, $\\mathbb{R}^4$ foot contact labels, $\\mathbb{R}^3$ global translations, and speech audio $\\mathbf{a} \\in \\mathbb{R}^{T \\times sk}$ , where $sk = sr_{\\mathrm{audio}} / fps_{\\mathrm{gestures}}$ , EMAGE jointly optimizes masked gesture reconstruction and audio-conditioned gesture generation. This optimization enhances performance during inference and enables the use of partially masked gestures to complete holistic gestures. To achieve this, we first model the quantized latent space in Sec. 4.1, following [61, 65]. We then design a separated speech audio encoder using Content Rhythm Self-Attention (CRA) as detailed in Sec. 4.2. Subsequently, we learn the body hints from masked gestures via the Masked Audio Gesture Transformer, explained in Sec. 4.3. Finally, the decoding strategy for different body segments is discussed in Sec. 4.4.", + "bbox": [ + 496, + 643, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "1147", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/03d3d4b036b0ca4357b992552ab291dfb862bda77305f32207eae0b1502e2b37.jpg", + "image_caption": [ + "Figure 3. EMAGE leverages two training paths: Masked Gesture Reconstruction (MG2G) and Audio-Conditioned Gesture Generation (A2G). The MG2G path focuses on encoding robust body hints through a spatial-temporal transformer gesture encoder and cross-attention gesture decoder. In contrast, the A2G path utilizes these body hints, and separated audio encoders to decode pretrained face and body latent features. A key component in this process is a switchable cross-attention layer, crucial for merging body hints and audio features. This fusion allows the features to be effectively disentangled and utilized for gesture decoding. Once the gesture latent features are reconstructed, EMAGE employs a pretrained VQ-Decoder to decode face and local body motions. Additionally, a pretrained Global Motion Predictor is used to estimate global body translations, further enhancing the model's capability to generate realistic and coherent gestures." + ], + "image_footnote": [], + "bbox": [ + 117, + 94, + 867, + 272 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/fcfbe7a0829c0edac6c9bb4d65277fe8bae5b40a7eca41f562d382b9fffccf6f.jpg", + "image_caption": [ + "Figure 4. Details of CRA and Pretrained VQ-VAEs. Left: Content Rhythm Attention fuses speech rhythm (onset and amplitude) with content (pretrained word embeddings from text scripts) adaptively. This results in a preference for either content or rhythm in specific frames, which encourages the generation of semantical-aware gestures. Right: We pretrain four compositional VQ-VAEs by reconstructing face, upper body, hands and lower body separately to disentangle audio-agnostic gestures explicitly." + ], + "image_footnote": [], + "bbox": [ + 151, + 404, + 271, + 518 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/20c78e60af2f8a1e4a54cfc9be2acb3cffddc5a758e0ff797f181e0b8b1facf5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 276, + 402, + 395, + 518 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Compositional Discrete Face and Body Prior", + "text_level": 1, + "bbox": [ + 76, + 672, + 454, + 690 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We model full-body gestures in the separated quantized latent space (Figure 4) for several reasons. Similar to [65], the body and hands improves the diversity of the results. However, we additionally separate the face and lower body due to their differing correlations with audio, i.e., using a single VQ-VAE to encode both the upper and lower body may lead the model to overlook gestures that occur less frequently - regardless of their relation to the audio. Specifically, a single model might focus only on recovering lower body motion, neglecting elbow movements in a speaker who is constantly walking during the conversation.", + "bbox": [ + 75, + 700, + 467, + 866 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The separated quantized latent space $Q = \\{\\mathbf{q}_{\\mathrm{f}}, \\mathbf{q}_{u}, \\mathbf{q}_{\\mathrm{h}}, \\mathbf{q}_{\\mathrm{l}}\\}$ for the Face $\\mathbf{g}_{\\mathrm{f}} \\in \\mathbb{R}^{T \\times 106}$ , Upper", + "bbox": [ + 76, + 869, + 468, + 902 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "body $\\mathbf{g}_{\\mathrm{u}} \\in \\mathbb{R}^{T \\times 78}$ , Hands $\\mathbf{g}_{\\mathrm{h}} \\in \\mathbb{R}^{T \\times 180}$ , and Lower body $\\mathbf{g}_{\\mathrm{l}} \\in \\mathbb{R}^{T \\times (54 + 4)}$ are from four VQ-VAEs. The VQ-VAE for each is optimized by jointly optimizing the following loss terms,", + "bbox": [ + 498, + 402, + 890, + 462 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nq _ {i} = \\underset {q _ {i} \\in \\mathbf {q}} {\\arg \\min } \\| z _ {j} - q _ {i} \\| ^ {2} \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 612, + 464, + 890, + 491 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {\\mathrm {V Q - V A E}} = \\mathcal {L} _ {r e c} (\\mathbf {g}, \\hat {\\mathbf {g}}) + \\mathcal {L} _ {\\mathrm {v e l}} \\left(\\mathbf {g} ^ {\\prime}, \\hat {\\mathbf {g}} ^ {\\prime}\\right) + \\mathcal {L} _ {\\mathrm {a c c}} \\left(\\mathbf {g} ^ {\\prime \\prime}, \\hat {\\mathbf {g}} ^ {\\prime \\prime}\\right) \\\\ + \\left\\| \\operatorname {s g} [ \\mathbf {z} ] - \\mathbf {q} \\right\\| ^ {2} + \\left\\| \\mathbf {z} - \\operatorname {s g} [ \\mathbf {q} ] \\right\\| ^ {2}, \\tag {4} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 519, + 512, + 890, + 551 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathbf{z}$ is the encoded $\\mathbf{g}$ with a temporal window size $w = 1$ . $\\mathcal{L}_{\\mathrm{rec}}$ , $\\mathcal{L}_{\\mathrm{vel}}$ , and $\\mathcal{L}_{\\mathrm{acc}}$ are Geodesic [58] and L1 losses, sg is a stop gradient operation, we set the weight of the commitment loss [59] (the last term) to 1 in this paper.", + "bbox": [ + 498, + 563, + 890, + 625 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Content and Rhythm Self-Attention", + "text_level": 1, + "bbox": [ + 500, + 633, + 810, + 650 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Given the speech audio, s, inspired by [4], we employ onset o and amplitude a as explicit audio rhythm, alongside the pretrained embeddings [9] e from transcripts as content. Different from previous approaches, which typically add the rhythm and content features, we leverage self-attention to merge these features adaptively. This approach is driven by the observation that for specific frames, the gestures are more related to content (semantic-aware) or rhythm (beat-aware). The rhythm and content features are first encoded into time-aligned features $\\mathbf{r}_{1:T}$ and $\\mathbf{c}_{1:T}$ , using a Temporal Convolutional Network (TCN) and linear mapping, respectively. For each time step $t\\in \\{1,\\dots,T\\}$ , we merge the rhythm and content features by:", + "bbox": [ + 496, + 657, + 890, + 854 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathbf {f} _ {1: T} = \\alpha \\times \\mathbf {r} _ {1: T} + (1 - \\alpha) \\times \\mathbf {c} _ {1: T}, \\tag {5} \\\\ \\alpha = \\operatorname {S o f t m a x} (\\mathcal {A T} (\\mathbf {r} _ {1: T}, \\mathbf {c} _ {1: T})), \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 576, + 867, + 890, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "1148", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/b5ffedfff937a78fa40fca7cb43e4752e73f12773787d02e95ac5a687967d09f.jpg", + "image_caption": [ + "Figure 5. Comparison of Forward Path Designs. Straightforward fusion module (a) merges audio features without refined body features and recombines audio features based only on position embedding. The Self-Attention decoder module (b), adopted in previous MLM models [17, 31], is limited for tasks requiring auto-regressive inference. Our design (c) considers effective audio feature fusion and auto-regressive decoding." + ], + "image_footnote": [], + "bbox": [ + 84, + 90, + 207, + 234 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/1cacf29fb5e6624bf438a25165fccad902e7718042aa87ac7732c73e36efd625.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 212, + 90, + 334, + 234 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/32cfc4eb4ef2edd24fa8beeb6c6b37217aecf0976a8d36ece3aaf17f2d0afc5b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 338, + 90, + 454, + 236 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\mathcal{A}\\mathcal{T}$ is a 2-layer MLP. We apply two separate CRA encoders for the face and body.", + "bbox": [ + 75, + 361, + 468, + 391 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Masked Audio-Conditioned Gesture Modeling", + "text_level": 1, + "bbox": [ + 76, + 398, + 467, + 416 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We propose a Masked Audio Gesture Transformer to leverage different training paths (see Fig. 5 for the motivation behind the architecture designs). Given temporal and spatially masked gestures $\\overline{\\mathbf{g}}\\in \\mathbb{R}^{T\\times 337}$ , we first replace the masked tokens with learned masked embeddings $e_{\\mathrm{mask}}\\in \\mathbb{R}^{256}$ , as the value zero still represents specific motion content, e.g., a T-pose. We linearly increase the ratio of masked joints and frames from 0 to $95\\%$ according to the training epochs.", + "bbox": [ + 75, + 422, + 468, + 542 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "A spatial convolutional $\\mathcal{SC}$ encoder initially summarizes the spatial information, and compress the spatial feature into $\\mathbb{R}^{T\\times 256}$ . We then employ a temporal self-attention (without feed-forward) $\\mathcal{TSA}$ to refine the summarized spatial features", + "bbox": [ + 75, + 544, + 468, + 616 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {h} = \\mathcal {T S A} (\\mathcal {S C} (\\overline {{\\mathbf {g}}}) + \\mathbf {p} _ {t}), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 186, + 618, + 468, + 633 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\mathbf{h} \\in \\mathbb{R}^{T \\times 512}$ represents encoded body hints, and $\\mathbf{p}_t$ is the sum of a learned speaker embedding and PPE [19]. Then, a straightforward temporal cross-attention Transformer decoder $\\mathcal{TCA}\\mathcal{T}$ is adopted for the reconstructed latent $\\hat{\\mathbf{q}}_{\\mathrm{mg2g}}$ ,", + "bbox": [ + 75, + 638, + 468, + 715 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathbf {q}} _ {\\mathrm {m g} 2 \\mathrm {g}} = \\mathcal {T C A T} (\\mathbf {h} + \\mathbf {p} _ {t}). \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 184, + 715, + 468, + 731 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We minimize the L1 distance in the latent space,", + "bbox": [ + 76, + 736, + 397, + 751 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {m g} 2 \\mathrm {g}} = \\| \\hat {\\mathbf {q}} _ {\\mathrm {m g} 2 \\mathrm {g}} - \\mathbf {q} \\|. \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 196, + 760, + 468, + 777 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Masked gesture reconstruction encodes effective body hints, with the key being to leverage these body hints for gesture generation. We employ a selective fusion for audio and body hints via a temporal cross-attention $\\mathcal{TCA}$ . Subsequently, we use the merged audio-gesture features for audio-conditioned gesture latent reconstruction,", + "bbox": [ + 75, + 785, + 468, + 875 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathbf {q}} _ {\\mathrm {a} 2 \\mathrm {g}} = \\mathcal {T C A T} \\left(\\mathcal {T C A} \\left(\\mathbf {h} + \\mathbf {p} _ {t}, \\mathbf {f} _ {\\text {b o d y}}\\right), \\overline {{\\mathbf {g}}} + \\mathbf {p} _ {t}\\right). \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 119, + 885, + 468, + 902 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/1bd2e05a0ddf11f82d1922cf9aa70f0429d1d52a384c056486df192d8633e7ff.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Top.-BTop.-FBodyFace
VOCA[14]-FLAME-38.3 ± 5.63
AMASS[46]SMPL-XFLAME42.0 ± 3.60-
TalkSHOW[65]SMPL-XSMPL-X14.4 ± 2.1926.1 ± 6.42
BEAT2 (Ours)SMPL-XFLAME43.6 ± 3.3835.7 ± 5.91
", + "bbox": [ + 522, + 88, + 867, + 143 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 3. User Preference Win Rate between Existing Datasets. 'Top' denotes the topology of the mesh. The results show our BEAT2 dataset outperforms the existing PGT dataset [65] in both body and face aspects. It also performs slightly better on body and lower on face when compared with previous mocap [46] and 3D scan-based datasets [14], respectively.", + "bbox": [ + 498, + 150, + 890, + 233 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We optimize both the latent code class classification cross-entropy loss $\\mathcal{L}_{\\mathrm{a2g - rec}}$ and the latent reconstruction loss $\\mathcal{L}_{\\mathrm{a2g - c1s}}$ to encourage diverse results.", + "bbox": [ + 498, + 251, + 890, + 297 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.4. Face and Translation Decoding", + "text_level": 1, + "bbox": [ + 500, + 305, + 774, + 321 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Considering that the face is weakly related to body motion, applying the same operation - recombining the audio features based on body hints - is not reasonable. Therefore, we directly concatenate the masked body hints with the audio features for the final decoding of facial latent,", + "bbox": [ + 498, + 328, + 890, + 402 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathbf {q}} _ {\\mathrm {f}} = \\mathcal {T} \\mathcal {C A T} \\left(\\mathbf {f} _ {\\text {f a c e}} \\oplus \\mathbf {h}, \\mathbf {p} _ {t}\\right). \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 599, + 412, + 890, + 429 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Once we have obtained the local lower body motion $\\tilde{\\mathbf{g}}_l\\in$ $\\mathbb{R}^{T\\times (54 + 4)}$ from the VQ-Decoder, we estimate the global translations $\\tilde{\\mathbf{t}}\\in \\mathbb{R}^{T\\times 3}$ with a pretrained Global Motion Predictor $\\tilde{\\mathbf{t}} = \\mathcal{G}(\\tilde{\\mathbf{g}}_l)$ , which significantly reduces foot sliding.", + "bbox": [ + 498, + 438, + 890, + 500 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 500, + 511, + 630, + 527 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We separate the evaluation into two categories: dataset quality and model ability. After removing data with low finger quality from BEAT, BEAT2 is reduced to 60 hours. We further split it into BEAT2-Standard (27 hours) and BEAT2-Additional (33 hours), based on the type of speech and conversation sections [39]. While acted gestures in speech sections are more diverse and expressive, spontaneous gestures in conversational sections tend to be more natural yet less varied. We report the results on BEAT2-Standard Speaker2 with an $85\\% / 7.5\\% / 7.5\\%$ train/val/test split.", + "bbox": [ + 496, + 537, + 890, + 688 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1. Dataset Quality Evaluation", + "text_level": 1, + "bbox": [ + 500, + 696, + 743, + 712 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We compare our dataset with the state-of-the-art pseudo ground truth (PGT) dataset, TalkSHOW [65], for both face and body. Additionally, we compare it with the AMASS [46] dataset for the body and VOCA [14] for the face as references. Due to the varying lengths of sequences in each dataset, we sample 100 comparison pairs, each with an equal duration ranging from 2 to 4 seconds. In a perceptual study, each participant evaluates a random set of 40 pairs during a 10-minute session by selecting the sequence they consider to have the best captured quality. In total, 60 participants were invited. It is important to note that participants are instructed to compare only the upper body results,", + "bbox": [ + 496, + 719, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "1149", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "... In my spare time ... ... if I feel ok ...", + "bbox": [ + 137, + 112, + 380, + 126 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "... I will go for a walk or hike in nature. I try to organize something", + "bbox": [ + 413, + 112, + 830, + 127 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/2c125761dd9cae7c4576e912ae7f448a11376383f3058ef27dd260925960ebc0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 135, + 130, + 256, + 199 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/cdb9bcf28bebd72500a014eecdaa75e44c0ee0d34d340bf78adbc004c8817654.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 277, + 130, + 395, + 199 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/bae710f5cf209e06c839b96c23eb6e0f4e8f3ae177079db6f6be7e4871ae7a24.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 418, + 130, + 539, + 199 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c1c57afefb94e35ad11b6de08d396967a3fa5faca4bb4d28920f253113b08c23.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 560, + 131, + 679, + 199 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/428f0f0053095cf331f75a2ee392f3922abd1e4ccefa68d335d26cba72ee5c79.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 707, + 132, + 825, + 199 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/0e571774cfd0d6a43f3e034ffbc64ca28c24d7885bf47b0ac76ba27000ba1cb3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 130, + 200, + 271, + 273 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/6ac6fc9288d18bedb3545a7e62f6007ef4f7c006b15af8b108bf094d1ad7880d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 284, + 200, + 395, + 272 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/5f6862fca91f66b90bd24f08769f95c06dfa042121411312abe32d886c6e6098.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 418, + 200, + 539, + 273 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/331dc7ed8f3b856ee1e11e203e33219dc639d6472ce45416330386342a33996a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 560, + 200, + 676, + 273 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/fd3a413577dc8a0433bca6a0d076bf3a615e42cf9e70297b2b0c83b45ad34d34.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 697, + 202, + 818, + 273 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d5f6d5f60c872172245a54c48b4f67219ea87e02dbdecbd6b9a8a9f105928657.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 137, + 276, + 251, + 345 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/50f47b1881a4f057c10224c1452d9152a3706b0b2443921ac8022fe89d6e58c5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 281, + 275, + 390, + 345 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/783ca256b44282fda31bf3021b3fb18df2884846cbacbae94c1c54303861afca.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 416, + 275, + 535, + 345 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/7e6a2efeb4f1b1f039e4b4103e35d22570d17ac002aa7b76b5737e7ad19b9a58.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 549, + 273, + 676, + 345 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/3b5bb70b6c5c3fbf791f74724eb0bfd3ce4c78419828d7f4465313421d034016.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 694, + 273, + 825, + 347 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/6870b2a16ab908948f4c9a29bfe7b7a29afcad6abcba862bc36131580f62c209.jpg", + "image_caption": [ + "Figure 6. Subjective Examples. Top to down: GroundTruth, Generated W/O Body Hints, Generated With Body Hints, Visible Body Hints, EMAGE generates diverse, semantic-aware and audio synchronized body gestures, e.g. raise the both hands for \"spare time\", relaxing gestures for \"hike in nature\". Besides, as shown in the third and bottom rows, EMAGE is flexible to accept non-audio-synchronized body hints on any frames and joints, to guide or customize the generated gestures explicitly, e.g., repeat a similar motion for raising hands, change the walking orientations, etc. Note that in this figure, The gray joints of generated results are not copies of visible hints." + ], + "image_footnote": [], + "bbox": [ + 132, + 348, + 259, + 415 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/6d0dd81e175b8e70e7d2cbaa3207c1c29eec10056fee93189c6f6b8833554f74.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 277, + 348, + 395, + 415 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/fd6f3eee973bb1ecc58436590341bf1e98b4a7419d4e84fc2f7ed5c9472f580e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 416, + 347, + 535, + 415 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/fe89c02a59451a9e1f4d73c03f7d1acf521b5838104d33ec8fb5dc3acbf675a8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 550, + 345, + 663, + 421 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/87fdaad99801ac186d6b6a5c6f959383750e13087ad63f5e0cb766c482d4920a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 692, + 348, + 831, + 420 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/faacea25f48cb1d51502542f3db4f1848e16a4bbb87af4b0c969e82300ef07f1.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
FGD ↓BC ↑Diversity ↑MSE ↓LVD ↓
FaceFormer[19]---7.7877.593
CodeTalker[61]---8.0267.766
S2G[25]28.154.6835.971--
Trimodal[67]12.415.9337.724--
HA2G[40]12.326.7798.626--
DisCo[38]9.4176.4399.912--
CaMN[39]6.6446.76910.86--
DiffStyleGesture[62]8.8117.24111.49--
Habibie et al.[28]9.0407.7168.2138.6148.043
TalkSHOW[65]6.2096.94713.477.7917.771
EMAGE (Ours)5.5127.72413.067.6807.556
", + "bbox": [ + 81, + 517, + 465, + 661 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "as TalkSHOW contains only the upper body. The results are shown in Table 3.", + "bbox": [ + 76, + 753, + 468, + 784 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.2. Model Ability Evaluation", + "text_level": 1, + "bbox": [ + 76, + 795, + 307, + 811 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We report results using BEATv1.3 and BEATv2. The latter has facial data refined by animators and hand data filtered by annotators.", + "bbox": [ + 76, + 819, + 468, + 864 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Metrics. We adopt FGD[67] to evaluate the realism of the body gestures. Then, we measure Diversity[33] by calcu", + "bbox": [ + 76, + 869, + 468, + 901 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/897c87d22557e913147a4e036f9b9610a58050783ae0433b85d3687dcc096855.jpg", + "table_caption": [ + "Table 4. Quantitative evaluation on BEATv2. We report FGD $\\times 10^{-1}$ , BC $\\times 10^{-1}$ , Diversity, MSE $\\times 10^{-8}$ , and LVD $\\times 10^{-5}$ . For body gestures, EMAGE significantly improves the FGD, indicating that the generated results are closer to the GT. This shows the effectiveness of body hints from masked gesture modeling." + ], + "table_footnote": [], + "table_body": "
HolisticBodyFace
Habibie et al.[28]12.4 ± 3.7015.9 ± 6.4910.8 ± 3.19
TalkSHOW[65]34.9 ± 5.7940.4 ± 8.2233.2 ± 6.03
EMAGE (Ours)52.7 ± 7.9144.7 ± 8.6856.0 ± 7.80
", + "bbox": [ + 531, + 518, + 859, + 566 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 5. User Preference Win Rate On Generated Results. The results indicate that our generated outcomes are perceived as more realistic and believable, with a $14\\%$ and $23\\%$ higher user preference for body and face gestures, respectively.", + "bbox": [ + 498, + 571, + 890, + 628 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "lating the average L1 distance between multiple body gesture clips, and use BC [34] to assess speech-motion synchrony. For the face, we calculate the vertex MSE[61] to measure the positional distance and the vertex L1 difference LVD [65] between the GT and generated facial vertices.", + "bbox": [ + 498, + 670, + 890, + 746 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Compared Methods. We first compare our method with representative state-of-the-art methods in body gesture generation [25, 38-40, 62, 67] and talking head generation [19, 61] by reproducing their methods for body and face, respectively. In addition, we reproduce two previous state-of-the-art holistic pipelines, Habibie et al. [28] and TalkSHOW [65], whose original implementations are limited to the upper body. We add a lower-body decoder to Habibie et al. and a lower-body VQ-VAE to TalkSHOW.", + "bbox": [ + 496, + 772, + 892, + 907 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "1150", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ec4acad0cc10bb45d1f30e78a3cac08c72cd08e88bef35a1db35c54e8a1d34f8.jpg", + "image_caption": [ + "Figure 7. Comparison of Generated Facial Movements. Compared with previous state-of-the-art talking face generation methods FaceFormer [19] and CodeTalker [61] as well as holistic gesture generation methods Habibie et al. [28] and TalkSHOW [65]. Note that CodeTalker has higher MSE than IMAGE on BEATv2 (Table 4, lower is better) but is subjectively realistic. IMAGE gets good lip motions by leveraging both the face model and masked body hints." + ], + "image_footnote": [], + "bbox": [ + 96, + 87, + 436, + 256 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.2.1 Quantitative and Qualitative Analysis", + "text_level": 1, + "bbox": [ + 76, + 383, + 393, + 398 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "As shown in Table 4, with a four-frame seed pose, our method outperforms previous SOTA algorithms. For qualitative results, see Figures 6 and 7. Furthermore, we conduct a perceptual study. Maintaining the same setup of 60 participants, each participant evaluates 40 pairs of 10-second results do decide which is most believable; this gives the win rate shown in Table 5.", + "bbox": [ + 75, + 401, + 470, + 506 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.2.2 Ablation Analysis", + "text_level": 1, + "bbox": [ + 76, + 527, + 256, + 542 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Performance of Baseline. As shown in Table 6, we start with a teacher-force transformer-based baseline. This baseline, inspired by FaceFormer [19], adopts a 1-layer cross-attention transformer decoder and replaces the audio features from Wav2Vec2 [6] with our custom TCN [7] and trainable word embeddings [9].", + "bbox": [ + 75, + 551, + 468, + 643 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effect of Multiple VQ-VAEs. Simply applying one VQ-VAE [27, 59] for full-body movements, including the face, decreases performance in facial movements. This is because the VQ-VAE is trained to minimize the average loss for the full body, and some speakers' most frequent movements are not related to the audio. Implementing separated VQ-VAEs allows the model to better leverage the advantages of discrete priors.", + "bbox": [ + 75, + 643, + 470, + 763 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effect of Content Rhythm Self-Attention. Adaptive fusion of rhythm and content features shows improvement in both FGD and Alignment. It selectively focuses the current motion more on rhythm or content features based on the training data's distribution. Moreover, we observed more semantic-aware results when applying CRA.", + "bbox": [ + 75, + 763, + 468, + 854 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effect of Masked Body Gesture Hints. The improvement across all objective metrics demonstrates that our model effectively leverages spatial-temporal gesture priors, reducing", + "bbox": [ + 75, + 854, + 470, + 901 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/b5e3474bf149b65160747babbc2213f218c7780bfb8d61b4f14ff101b36b7dd3.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
FGD ↓BC ↑Diversity ↑MSE ↓LVD ↓
Ground Truth06.89613.07400
Reconstruction3.9136.75813.1450.8416.389
Baseline13.0806.9418.31451.4429.317
+ VQVAE9.7876.67310.6241.6199.473
+ 4 VQVAE7.3976.69812.5441.2438.938
+ CRA6.8336.78312.6761.1868.715
+ Masked Hints5.4236.79413.0571.1809.015
", + "bbox": [ + 506, + 87, + 885, + 193 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/3082b043011fb1ad1fe944210d5da70299672d88000c5ade9d11baf995a271b7.jpg", + "table_caption": [ + "Table 6. Abliation Analysis on BEATv1.3" + ], + "table_footnote": [], + "table_body": "
bodyhandsfaceFGD ↓BC ↑Diversity ↑
BESTX✓✓✓5.4236.79413.057
+ Trinity[22]✓5.3196.84313.346
+ AMASS[46]✓✓5.1746.76914.318
", + "bbox": [ + 521, + 231, + 870, + 281 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 7. Training IMAGE on Multiple Datasets. IMAGE demonstrates flexibility by training on multiple datasets, even when only a subset of holistic gestures is available. This approach further improves the objective metrics on the BEATv1.3 test set.", + "bbox": [ + 498, + 285, + 890, + 340 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "the likelihood of incorrect gesture sampling. Furthermore, masked gesture modeling is key to enabling the network to accept predefined gestures in specific frames.", + "bbox": [ + 498, + 362, + 890, + 409 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.2.3 Ability of Multi-Dataset Training.", + "text_level": 1, + "bbox": [ + 500, + 431, + 789, + 448 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We demonstrate that EMAGE can effectively combine multiple non-holistic datasets for training by jointly training with the Trinity [22] and AMASS [46] datasets, by using only the upper body and audio pairs from Trinity and only body and hands from AMASS. Separate VQ-VAEs are trained for BEAT2, Trinity, and AMASS, with separate MLP heads implemented for codebook classification. The results in Table 7 show that incorporating data improves performance on the BEATv1.3 test set.", + "bbox": [ + 496, + 460, + 890, + 597 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 612, + 617, + 628 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this work, we present IMAGE, a framework designed to accept partial gestures as input for completing audio-synchronized holistic gestures. It demonstrates that leveraging masked gesture reconstruction can significantly enhance the performance of audio-conditioned gesture generation. Furthermore, the design of IMAGE enables the training on multiple datasets, further improving performance. Alongside IMAGE, we release BEAT2, the largest multi-modal gesture dataset consistent with SMPL-X and FLAME. We hope that BEAT2 will contribute to knowledge and model sharing across various subareas.", + "bbox": [ + 496, + 641, + 890, + 806 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Disclosures: While You Zhou, Xuefei Zhe, Naoya Iwamoto, and Bo Zheng are employees of Huawei Tokyo Research Center, this work was done on their own time with the approval of their employer. MJB CoI: https://files.is.tue.mpg.de/black/CoI_CVPR_2024.txt", + "bbox": [ + 496, + 809, + 890, + 898 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "1151", + "bbox": [ + 483, + 944, + 513, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Kfir Aberman, Peizhuo Li, Dani Lischinski, Olga Sorkine-Hornung, Daniel Cohen-Or, and Baoquan Chen. Skeleton-aware networks for deep motion retargeting. ACM Transactions on Graphics (TOG), 39(4):62-1, 2020. 2", + "[2] Chaitanya Ahuja, Dong Won Lee, Yukiko I Nakano, and Louis-Philippe Morency. Style transfer for co-speech gesture animation: A multi-speaker conditional-mixture approach. In European Conference on Computer Vision, pages 248–265. Springer, 2020. 2", + "[3] Simon Alexanderson, Gustav Eje Henter, Taras Kucherenko, and Jonas Beskow. Style-controllable speech-driven gesture synthesis using normalising flows. In Computer Graphics Forum, pages 487-496. Wiley Online Library, 2020. 2", + "[4] Tenglong Ao, Zeyi Zhang, and Libin Liu. Gesture diffu-clip: Gesture diffusion model with clip latents. ACM Trans. Graph. 2, 5", + "[5] Tenglong Ao, Qingzhe Gao, Yuke Lou, Baoquan Chen, and Libin Liu. Rhythmic gesticulator: Rhythm-aware co-speech gesture synthesis with hierarchical neural embeddings. ACM Transactions on Graphics (TOG), 41(6):1-19, 2022. 2, 3", + "[6] Alexei Baevski, Yuhao Zhou, Abdelrahman Mohamed, and Michael Auli. wav2vec 2.0: A framework for self-supervised learning of speech representations. Advances in neural information processing systems, 33:12449-12460, 2020. 8", + "[7] Shaojie Bai, J Zico Kolter, and Vladlen Koltun. An empirical evaluation of generic convolutional and recurrent networks for sequence modeling. arxiv 2018. arXiv preprint arXiv:1803.01271, 2, 1803. 8", + "[8] Gilad Baruch, Zhuoyuan Chen, Afshin Dehghan, Tal Dimry, Yuri Feigin, Peter Fu, Thomas Gebauer, Brandon Joffe, Daniel Kurz, Arik Schwartz, and Elad Shulman. ARK-scenes - a diverse real-world dataset for 3d indoor scene understanding using mobile RGB-d data. In Thirty-fifth Conference on Neural Information Processing Systems Datasets and Benchmarks Track (Round 1), 2021. 2, 3", + "[9] Piotr Bojanowski, Edouard Grave, Armand Joulin, and Tomas Mikolov. Enriching word vectors with subword information. Transactions of the association for computational linguistics, 5:135-146, 2017. 5, 8", + "[10] Biovision BVH. Biovision bvh. In https://research.cs.wisc.edu/graphics /Courses/ cs-838-1999/Jeff/BVH.html, 1999. 3", + "[11] Zhe Cao, Gines Hidalgo, Tomas Simon, Shih-En Wei, and Yaser Sheikh. Openpose: realtime multi-person 2d pose estimation using part affinity fields. IEEE transactions on pattern analysis and machine intelligence, 43(1):172-186, 2019. 2", + "[12] Zhe Chen, Yuchen Duan, Wenhai Wang, Junjun He, Tong Lu, Jifeng Dai, and Yu Qiao. Vision transformer adapter for dense predictions. arXiv preprint arXiv:2205.08534, 2022. 3", + "[13] Kiran Chhatre, Radek Danček, Nikos Athanasiou, Giorgio Becherini, Christopher Peters, Michael J Black, and Timo Bolkart. Emotional speech-driven 3d body animation via disentangled latent diffusion. arXiv preprint arXiv:2312.04466, 2023. 1" + ], + "bbox": [ + 78, + 114, + 470, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[14] Daniel Cudeiro, Timo Bolkart, Cassidy Laidlaw, Anurag Ranjan, and Michael Black. Capture, learning, and synthesis of 3D speaking styles. Computer Vision and Pattern Recognition (CVPR), pages 10101-10111, 2019. 2, 3, 6", + "[15] Radek Dančecek, Kiran Chhatre, Shashank Tripathi, Yandong Wen, Michael Black, and Timo Bolkart. Emotional speech-driven animation with content-emotion disentanglement. ACM, 2023. 1", + "[16] Anna Deichler, Kiran Chhatre, Christopher Peters, and Jonas Beskow. Spatio-temporal priors in 3d human motion. 2021. 2", + "[17] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018. 3, 6", + "[18] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 3", + "[19] Yingruo Fan, Zhaojiang Lin, Jun Saito, Wenping Wang, and Taku Komura. Faceformer: Speech-driven 3d facial animation with transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2, 3, 6, 7, 8", + "[20] Yingruo Fan, Zhaojiang Lin, Jun Saito, Wenping Wang, and Taku Komura. Faceformer: Speech-driven 3d facial animation with transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18770-18780, 2022. 1", + "[21] Gabriele Fanelli, Juergen Gall, Harald Romsdorfer, Thibaut Weise, and Luc Van Gool. A 3-d audio-visual corpus of affective communication. IEEE Transactions on Multimedia, 12(6):591-598, 2010. 2", + "[22] Ylva Ferstl and Rachel McDonnell. Investigating the use of recurrent motion modelling for speech gesture generation. In Proceedings of the 18th International Conference on Intelligent Virtual Agents, pages 93-98, 2018. 3, 8", + "[23] Erik Gartner, Mykhaylo Andriluka, Erwin Coumans, and Cristian Sminchisescu. Differentiable dynamics for articulated 3d human motion reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13190-13200, 2022. 2", + "[24] Saeed Ghorbani, Ylva Ferstl, Daniel Holden, Nikolaus F. Troje, and Marc-Andre Carbonneau. Zeroeggs: Zero-shot example-based gesture generation from speech. Computer Graphics Forum, 42(1):206-216, 2023. 2, 3", + "[25] Shiry Ginosar, Amir Bar, Gefen Kohavi, Caroline Chan, Andrew Owens, and Jitendra Malik. Learning individual styles of conversational gesture. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3497-3506, 2019. 1, 2, 3, 7", + "[26] Yuan Gong, Yu-An Chung, and James Glass. Ast: Audio spectrogram transformer. arXiv preprint arXiv:2104.01778, 2021.3" + ], + "bbox": [ + 501, + 92, + 890, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "1152", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[27] Chuan Guo, Xinxin Zuo, Sen Wang, and Li Cheng. Tm2t: Stochastic and tokenized modeling for the reciprocal generation of 3d human motions and texts. In European Conference on Computer Vision, pages 580-597. Springer, 2022. 8, 4", + "[28] Ikhsanul Habibie, Weipeng Xu, Dushyant Mehta, Lingjie Liu, Hans-Peter Seidel, Gerard Pons-Moll, Mohamed Elgharib, and Christian Theobalt. Learning speech-driven 3d conversational gestures from video. arXiv preprint arXiv:2102.06837, 2021. 1, 2, 3, 7, 8", + "[29] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16000-16009, 2022. 3", + "[30] Catalin Ionescu, Dragos Papava, Vlad Olaru, and Cristian Sminchisescu. Human3. 6m: Large scale datasets and predictive methods for 3d human sensing in natural environments. IEEE transactions on pattern analysis and machine intelligence, 36(7):1325-1339, 2013. 2", + "[31] Zhenzhong Lan, Mingda Chen, Sebastian Goodman, Kevin Gimpel, Piyush Sharma, and Radu Soricut. Albert: A lite bert for self-supervised learning of language representations. arXiv preprint arXiv:1909.11942, 2019. 3, 6", + "[32] Gilwoo Lee, Zhiwei Deng, Shugao Ma, Takaaki Shiratori, Siddhartha S Srinivasa, and Yaser Sheikh. Talking with hands $16.2\\mathrm{m}$ : A large-scale dataset of synchronized body-finger motion and audio for conversational motion analysis and synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 763-772, 2019. 2, 3", + "[33] Jing Li, Di Kang, Wenjie Pei, Xuefei Zhe, Ying Zhang, Zhenyu He, and Linchao Bao. Audio2gestures: Generating diverse gestures from speech audio with conditional variational autoencoders. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11293-11302, 2021. 1, 2, 7, 3", + "[34] Ruilong Li, Shan Yang, David A Ross, and Angjoo Kanazawa. Ai choreographer: Music conditioned 3d dance generation with aist++. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13401-13412, 2021. 7, 2", + "[35] Tianye Li, Timo Bolkart, Michael. J. Black, Hao Li, and Javier Romero. Learning a model of facial shape and expression from 4D scans. ACM Transactions on Graphics, (Proc. SIGGRAPH Asia), 36(6):194:1-194:17, 2017. 2, 3", + "[36] Haiyang Liu and Cheng Zhang. Reinforcement learning based neural architecture search for audio tagging. In 2020 International Joint Conference on Neural Networks (IJCNN), pages 1-8. IEEE, 2020. 3", + "[37] Haiyang Liu and Jihan Zhang. Improving ultrasound tongue image reconstruction from lip images using self-supervised learning and attention mechanism. arXiv preprint arXiv:2106.11769, 2021. 3", + "[38] Haiyang Liu, Naoya Iwamoto, Zihao Zhu, Zhengqing Li, You Zhou, Elif Bozkurt, and Bo Zheng. Disco: Disentangled implicit content and rhythm learning for diverse co-speech gestures synthesis. In Proceedings of the 30th ACM Interna-" + ], + "bbox": [ + 78, + 90, + 470, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "tional Conference on Multimedia, pages 3764-3773, 2022. 2,3,7", + "[39] Haiyang Liu, Zihao Zhu, Naoya Iwamoto, Yichen Peng, Zhengqing Li, You Zhou, Elif Bozkurt, and Bo Zheng. Beat: A large-scale semantic and emotional multi-modal dataset for conversational gestures synthesis. arXiv preprint arXiv:2203.05297, 2022. 1, 2, 3, 4, 6, 7", + "[40] Xian Liu, Qianyi Wu, Hang Zhou, Yinghao Xu, Rui Qian, Xinyi Lin, Xiaowei Zhou, Wayne Wu, Bo Dai, and Bolei Zhou. Learning hierarchical cross-modal association for cospeech gesture generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10462-10472, 2022. 2, 3, 7", + "[41] Xian Liu, Qianyi Wu, Hang Zhou, Yinghao Xu, Rui Qian, Xinyi Lin, Xiaowei Zhou, Wayne Wu, Bo Dai, and Bolei Zhou. Learning hierarchical cross-modal association for cospeech gesture generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10462-10472, 2022. 1, 2", + "[42] Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. Roberta: A robustly optimized bert pretraining approach. arXiv preprint arXiv:1907.11692, 2019.3", + "[43] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10012-10022, 2021. 3", + "[44] Matthew M. Loper, Naureen Mahmood, and Michael J. Black. MoSh: Motion and shape capture from sparse markers. ACM Transactions on Graphics, (Proc. SIGGRAPH Asia), 33(6):220:1-220:13, 2014. 2", + "[45] Shuhong Lu, Youngwoo Yoon, and Andrew Feng. Co-speech gesture synthesis using discrete gesture token learning. arXiv preprint arXiv:2303.12822, 2023. 2, 3", + "[46] Naureen Mahmood, Nima Ghorbani, Nikolaus F. Troje, Gerard Pons-Moll, and Michael J. Black. AMASS: Archive of motion capture as surface shapes. In International Conference on Computer Vision, pages 5442-5451, 2019. 2, 3, 6, 8", + "[47] Tim Massey. Vicon study of dynamic object tracking accuracy. In https://www.vicon.com/resources/blog/vicon-study-of-dynamic-object-tracking-accuracy/, 2020. 2, 3", + "[48] Evonne Ng, Shiry Ginosar, Trevor Darrell, and Hanbyul Joo. Body2hands: Learning to infer 3d hands from conversational gesture body dynamics. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11865-11874, 2021. 2, 3", + "[49] Kunkun Pang, Dafei Qin, Yingruo Fan, Julian Habekost, Takaaki Shiratori, Junichi Yamagishi, and Taku Komura. Bodyformer: Semantics-guided 3d body gesture synthesis with transformer. ACM Transactions on Graphics (TOG), 42(4):1-12, 2023. 2, 3", + "[50] Georgios Pavlakos, Vasileios Choutas, Nima Ghorbani, Timo Bolkart, Ahmed A. A. Osman, Dimitrios Tzionas, and Michael J. Black. Expressive body capture: 3D hands, face," + ], + "bbox": [ + 501, + 92, + 893, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "1153", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "and body from a single image. In Proceedings IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), pages 10975-10985, 2019. 2, 3, 4", + "[51] Ziqiao Peng, Haoyu Wu, Zhenbo Song, Hao Xu, Xiangyu Zhu, Jun He, Hongyan Liu, and Zhaoxin Fan. Emotalk: Speech-driven emotional disentanglement for 3d face animation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 20687-20697, 2023. 2", + "[52] Mathis Petrovich, Michael J Black, and Gül Varol. Action-conditioned 3d human motion synthesis with transformer vae. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10985-10995, 2021. 2", + "[53] Xingqun Qi, Chen Liu, Lincheng Li, Jie Hou, Haoran Xin, and Xin Yu. Emotiongesture: Audio-driven diverse emotional co-speech 3d gesture generation. arXiv preprint arXiv:2305.18891, 2023. 2", + "[54] Xingqun Qi, Jiahao Pan, Peng Li, Ruibin Yuan, Xiaowei Chi, Mengfei Li, Wenhan Luo, Wei Xue, Shanghang Zhang, Qifeng Liu, et al. Weakly-supervised emotion transition learning for diverse 3d co-speech gesture generation. arXiv preprint arXiv:2311.17532, 2023. 2", + "[55] Guocheng Qian, Xingdi Zhang, Abdullah Hamdi, and Bernard Ghanem. Pix4point: Image pretrained transformers for 3d point cloud understanding. 2022. 3", + "[56] Alexander Richard, Michael Zollhöfer, Yandong Wen, Fernando De la Torre, and Yaser Sheikh. Meshtalk: 3d face animation from speech using cross-modality disentanglement. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1173-1182, 2021. 2, 3", + "[57] Kaede Shiohara, Xingchao Yang, and Takafumi Taketomi. Blendface: Re-designing identity encoders for face-swapping. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7634-7644, 2023. 2", + "[58] Tommi Tykkälä, Cédric Audras, and Andrew I Comport. Direct iterative closest point for real-time visual odometry. In 2011 IEEE International Conference on Computer Vision Workshops (ICCV Workshops), pages 2050-2056. IEEE, 2011. 5", + "[59] Aaron Van Den Oord, Oriol Vinyals, et al. Neural discrete representation learning. Advances in neural information processing systems, 30, 2017. 5, 8", + "[60] Kaisiyuan Wang, Qianyi Wu, Linsen Song, Zhuoqian Yang, Wayne Wu, Chen Qian, Ran He, Yu Qiao, and Chen Change Loy. Mead: A large-scale audio-visual dataset for emotional talking-face generation. In ECCV, 2020. 2", + "[61] Jinbo Xing, Menghan Xia, Yuechen Zhang, Xiaodong Cun, Jue Wang, and Tien-Tsin Wong. Codetalker: Speech-driven 3d facial animation with discrete motion prior. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12780-12790, 2023. 1, 2, 3, 4, 7, 8", + "[62] Sicheng Yang, Zhiyong Wu, Minglei Li, Zhensong Zhang, Lei Hao, Weihong Bao, Ming Cheng, and Long Xiao. Diffusestylegesture: Stylized audio-driven co-speech gesture generation with diffusion models. arXiv preprint arXiv:2305.04919, 2023. 2, 3, 7" + ], + "bbox": [ + 78, + 90, + 470, + 898 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[63] Xingchao Yang and Takafumi Taketomi. Bareskinnet: Demakeup and de-lighting via 3d face reconstruction. In Computer Graphics Forum, pages 623-634. Wiley Online Library, 2022. 2", + "[64] Xingchao Yang, Takafumi Taketomi, and Yoshihiro Kanamori. Makeup extraction of 3d representation via illumination-aware image decomposition. In Computer Graphics Forum, pages 293-307. Wiley Online Library, 2023. 2", + "[65] Hongwei Yi, Hualin Liang, Yifei Liu, Qiong Cao, Yandong Wen, Timo Bolkart, Dacheng Tao, and Michael J Black. Generating holistic 3d human motion from speech. In CVPR, 2023. 1, 2, 3, 4, 5, 6, 7, 8", + "[66] Youngwoo Yoon, Woo-Ri Ko, Minsu Jang, Jaeyeon Lee, Jae-hong Kim, and Geehyuk Lee. Robots learn social skills: End-to-end learning of co-speech gesture generation for humanoid robots. In 2019 International Conference on Robotics and Automation (ICRA), pages 4303-4309. IEEE, 2019. 2, 3", + "[67] Youngwoo Yoon, Bok Cha, Joo-Haeng Lee, Minsu Jang, Jaeyeon Lee, Jaehong Kim, and Geehyuk Lee. Speech gesture generation from the trimodal context of text, audio, and speaker identity. ACM Transactions on Graphics (TOG), 39 (6):1-16, 2020. 1, 2, 3, 7", + "[68] Xumin Yu, Lulu Tang, Yongming Rao, Tiejun Huang, Jie Zhou, and Jiwen Lu. Point-bert: Pre-training 3d point cloud transformers with masked point modeling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19313-19322, 2022. 3", + "[69] Hengshuang Zhao, Li Jiang, Jiaya Jia, Philip HS Torr, and Vladlen Koltun. Point transformer. In Proceedings of the IEEE/CVF international conference on computer vision, pages 16259-16268, 2021. 3", + "[70] Yihao Zhi, Xiaodong Cun, Xuelin Chen, Xi Shen, Wen Guo, Shaoli Huang, and Shenghua Gao. Livelyspeaker: Towards semantic-aware co-speech gesture generation. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 20807-20817, 2023. 2, 3", + "[71] Yang Zhou, Jimei Yang, Dingzeyu Li, Jun Saito, Deepali Aneja, and Evangelos Kalogerakis. Audio-driven neural gesture reenactment with video motion graphs. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3418-3428, 2022. 2", + "[72] Lingting Zhu, Xian Liu, Xuanyu Liu, Rui Qian, Ziwei Liu, and Lequan Yu. Taming diffusion models for audio-driven co-speech gesture generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10544-10553, 2023. 2, 3", + "[73] Wentao Zhu, Xiaoxuan Ma, Zhaoyang Liu, Libin Liu, Wayne Wu, and Yizhou Wang. Motionbert: A unified perspective on learning human motion representations. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2023. 3" + ], + "bbox": [ + 501, + 92, + 890, + 835 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "1154", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2024/EMAGE_ Towards Unified Holistic Co-Speech Gesture Generation via Expressive Masked Audio Gesture Modeling/f8a85da7-9221-4a37-baf0-e5a3b46705e3_model.json b/2024/EMAGE_ Towards Unified Holistic Co-Speech Gesture Generation via Expressive Masked Audio Gesture Modeling/f8a85da7-9221-4a37-baf0-e5a3b46705e3_model.json new file mode 100644 index 0000000000000000000000000000000000000000..7998894870c64aa6a7319130d8a5d2c2de10cc5e --- /dev/null +++ b/2024/EMAGE_ Towards Unified Holistic Co-Speech Gesture Generation via Expressive Masked Audio Gesture Modeling/f8a85da7-9221-4a37-baf0-e5a3b46705e3_model.json @@ -0,0 +1,2818 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.132, + 0.13, + 0.842, + 0.178 + ], + "angle": 0, + "content": "EMAGE: Towards Unified Holistic Co-Speech Gesture Generation via Expressive Masked Audio Gesture Modeling" + }, + { + "type": "text", + "bbox": [ + 0.145, + 0.202, + 0.823, + 0.242 + ], + "angle": 0, + "content": "Haiyang Liu\\(^{1*}\\) Zihao Zhu\\(^{2*}\\) Giorgio Becherini\\(^{4}\\) Yichen Peng\\(^{5}\\) Mingyang Su\\(^{6}\\) You Zhou\\(^{3}\\) Xuefei Zhe\\(^{3}\\) Naoya Iwamoto\\(^{3}\\) Bo Zheng\\(^{3}\\) Michael J. Black\\(^{3}\\)" + }, + { + "type": "text", + "bbox": [ + 0.156, + 0.256, + 0.79, + 0.294 + ], + "angle": 0, + "content": "\\(^{1}\\)The University of Tokyo \\(^{2}\\)Keio University \\(^{3}\\)Independent Researcher \\(^{4}\\)Max Planck Institute for Intelligent Systems \\(^{5}\\)JAIST \\(^{6}\\)Tsinghua University" + }, + { + "type": "image", + "bbox": [ + 0.117, + 0.328, + 0.859, + 0.506 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.075, + 0.511, + 0.896, + 0.556 + ], + "angle": 0, + "content": "Figure 1. EMAGE. We present a Masked Audio-Conditioned Gesture Modeling framework, along with a new holistic gesture dataset, BEAT2 (BEAT-SMPLX-FLAME), for jointly generating facial expressions, local body dynamics, hand movements and global translations, conditioned on audio and a partially or completely masked gestures. The gray denotes visible gestures, and blue represents our outputs." + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.566, + 0.315, + 0.584 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.599, + 0.474, + 0.903 + ], + "angle": 0, + "content": "We propose EMAGE, a framework to generate full-body human gestures from audio and masked gestures, encompassing facial, local body, hands, and global movements. To achieve this, we first introduce BEAT2 (BEATSMLPX-FLAME), a new mesh-level holistic co-speech dataset. BEAT2 combines a MoShed SMPL-X body with FLAME head parameters and further refines the modeling of head, neck, and finger movements, offering a community-standardized, high-quality 3D motion captured dataset. EMAGE leverages masked body gesture priors during training to boost inference performance. It involves a Masked Audio Gesture Transformer, facilitating joint training on audio-to-gesture generation and masked gesture reconstruction to effectively encode audio and body gesture hints. Encoded body hints from masked gestures are then separately employed to generate facial and body movements. Moreover, EMAGE adaptively merges speech features from the audio's rhythm and content and utilizes four compositional VQ-VAEs to enhance the results' fidelity and diversity. Experiments demonstrate that EMAGE generates" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.567, + 0.895, + 0.629 + ], + "angle": 0, + "content": "holistic gestures with state-of-the-art performance and is flexible in accepting predefined spatial-temporal gesture inputs, generating complete, audio-synchronized results. Our code and dataset are available.1" + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.658, + 0.634, + 0.673 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.683, + 0.895, + 0.867 + ], + "angle": 0, + "content": "The path towards full-body co-speech gesture generation presents many remaining challenges. Despite the development of various baselines, e.g., audio conditioned models for face [15, 20, 61], body [13, 25, 67], and hand movements [33, 39, 41], along with a few attempts at merged models [28, 65], the limited availability of comprehensive data and models poses an ongoing obstacle. To make progress in this direction, we present a new framework that can incorporate partial spatial-temporal predefined gestures and autonomously complete the remaining frames in synchronization with the audio. This provides a new tool for creating realistic animations of digital humans" + }, + { + "type": "page_footnote", + "bbox": [ + 0.524, + 0.875, + 0.811, + 0.901 + ], + "angle": 0, + "content": "*Equal Contribution. \n1https://pantomatrix.github.io/EMAGE/" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.957 + ], + "angle": 0, + "content": "1144" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.078, + 0.092, + 0.186, + 0.106 + ], + "angle": 0, + "content": "[57, 63, 64, 71]." + }, + { + "type": "ref_text", + "bbox": [ + 0.077, + 0.112, + 0.473, + 0.565 + ], + "angle": 0, + "content": "To achieve this, we first require a comprehensive gesture dataset. While several datasets are available for audio-to-body [67], audio-to-body&face [28], body-to-hands [39, 41], and common actions [16, 46], integrating them poses challenges due to varying data formats and evaluation metrics across different sub-areas. This motivates us to establish a unified, community-standard benchmark for training and evaluating holistic gestures. We build upon the BEAT dataset [39], which provides the most comprehensive mocap co-speech gesture dataset with diverse modalities. Despite its size and richness, the data in BEAT is not presented in a standardized format. The dataset comprises iPhone ARKit blendshape weights and Vicon skeletons, posing challenges when using the data for training; e.g., the lack of a mesh representation prevents the use of a vertex loss [19, 52]. Additionally, conducting direct human perceptual evaluations on these diverse modalities is inherently complex. On the other hand, SMPL-X [50] and FLAME [35] are common mesh standards for the academic community for other human dynamics-related tasks, e.g., action [52] and talking head generation [14]. Driven by the motivation to facilitate the sharing of knowledge across different tasks, we present BEAT-SMPLX-FLAME (BEAT2), which consists of two main components: i) Refined SMPL-X body shape and pose parameters from MoSh++ [44, 46] with hard-coded physical priors. ii) Optimized high-quality FLAME head parameters. By integrating SMPL-X's body and FLAME's head, BEAT2 facilitates a comprehensive training and evaluation benchmark across multiple sub-areas of co-speech human animation generation." + }, + { + "type": "ref_text", + "bbox": [ + 0.077, + 0.569, + 0.473, + 0.903 + ], + "angle": 0, + "content": "With the holistic data from BEAT2, we aim to enhance the coherence of individual body parts while ensuring accurate cross-modal alignment between audio and motion; see Figure 1. This leads to the design of Expressive Masked Audio-conditioned GEsture modeling, denoted as EMAGE, a spatial and temporal transformer-based framework. EMAGE first aggregates spatial features from masked body joints. Then, it reconstructs the latent space of pretrained gestures with a switchable gesture temporal self-attention and audio-gesture cross-attention. The selection of different forward paths enables modeling the gesture-to-gesture and audio-to-gesture prior separately and effectively. Once the reconstructed latent features are obtained, EMAGE decodes local face and body gestures from four compositional pretrained Vector Quantized-Variational AutoEncoders (VQ-VAEs) and decodes global translations from a pretrained Global Motion Predictor. The compositional VQ-VAEs are key to preserving audio-related motions during the reconstruction. Through these designs, EMAGE achieves state-of-the-art performance in generating body and face gestures. It takes as input audio and partially initialized gestures, and recovers audio" + }, + { + "type": "list", + "bbox": [ + 0.077, + 0.092, + 0.473, + 0.903 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.137 + ], + "angle": 0, + "content": "synchronized, coherent gestures. Additionally, we show how EMAGE can flexibly incorporate additional, nonholistic, datasets to improve results." + }, + { + "type": "ref_text", + "bbox": [ + 0.498, + 0.137, + 0.895, + 0.319 + ], + "angle": 0, + "content": "Overall, our contributions are as follows: (1) We release BEAT2, a representation-unified, mesh-level dataset that leverages MoShed SMPL-X body and FLAME head parameters. (2) We propose EMAGE, a simple yet effective holistic gesture generation framework that generates coherent gestures with partial gesture and audio priors. (3) EMAGE achieves state-of-the-art performance in generating both body and face gestures with only four frames of seed gestures. (4) EMAGE showcases the use of additional non-holistic gesture datasets for training, e.g., Trinity and AMASS, to further improve the fidelity and diversity of the results, effectively leveraging data from different datasets." + }, + { + "type": "list", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.319 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.332, + 0.642, + 0.348 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "ref_text", + "bbox": [ + 0.498, + 0.358, + 0.895, + 0.763 + ], + "angle": 0, + "content": "Co-Speech Animation Datasets are categorized into two types; see Table 1: pseudo-labeled (PGT) and motion-captured (mocap). For PGT, the Speech2Gesture Dataset [25] utilizes OpenPose [11] to extract 2D poses from News and Teaching videos. Subsequent works extend the dataset to 3D pose [28] and SMPL-X [65]. The TED Dataset [66], which extracts 2D pose from TED videos, has been extended to include 3D pose [67], fingers [40], and SMPL-X [45]. Similarly, 2D, 3D landmarks, and meshes are estimated from multi-view recorded videos [60] as PGT for talking-head generation. Although pseudo-labeled approaches could theoretically extract infinite data, their accuracy is limited. For instance, the average error (in mm) for the SOTA monocular 3D pose estimation algorithm [23] on the Human3.6M dataset [30] is 33.4, while for Vicon mocap it is 0.142 [47]. For mocap datasets, Trinity [3] features one male actor with 4h of data, and TalkingWithHands [32] collects conversation scenarios for two speakers. ZEEG [24] considers one speaker with 12 styles. For the face, 3D scanning is accurate but limited in quantity due to cost, e.g., less than 3h for BIWI [21], VOCA [14], and MeshTalk [56]. This leads to a balance between performance and quantity in the ARKit dataset [51]. The aforementioned datasets address the body and face separately. On the other hand, BEAT [39] includes both 3D pose body and ARKit [8] facial blendshape weights for the first time. However, BEAT lacks mesh data." + }, + { + "type": "ref_text", + "bbox": [ + 0.498, + 0.765, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Co-Speech Animation Models are categorized based on their outputs; see Table 2. In addition to the baselines and datasets [14, 25, 32, 48, 56, 67], several frameworks for body gestures [2, 33, 53, 54] have improved the performance of the baselines. They are trained and evaluated by selecting specific upper body joints [4, 5, 38-40, 49, 70], or all body joints [62, 72]. Recent improvements in facial gestures drive the vertices with transformers and discrete face priors [19, 61]. However, these methods only address ei" + }, + { + "type": "list", + "bbox": [ + 0.498, + 0.358, + 0.895, + 0.903 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "1145" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.082, + 0.088, + 0.885, + 0.171 + ], + "angle": 0, + "content": "
Trinity Mocap 2017 [22]S2G S2G-2D 2019 [25]Seq2Seq TED-2D 2019 [66]TWH Mocap 2019 [32]Trimodal TED-3D 2020 [67]VOCA Scan 2020 [14]MeshTalk Scan 2021 [56]Hahibie et al. S2G-3D 2021 [28]HA2G TED-3D+ 2022 [40]BEAT Mocap 2022 [39]ZEEG Mocap 2022 [24]Yoon. et al. TED-SMPL 2023 [45]Talkshow S2G-SMPL 2023 [65]BEAT2 Ours 2024
Head-2D---3D Scan3D Scan3D-ARKit--PGT-MeshMC-Mesh
Upper Body3D2D2D3D3D--3D3D3D3DPGT-MeshPGT-MeshMC-Mesh
Hands3D2D-3D---3D3D3D3DPGT-MeshPGT-MeshMC-Mesh
Lower Body3D--3D-----3D3D--MC-Mesh
Global Motion3D--3D-----3D3D--MC-Mesh
Duration (hours)4609720970.523833764302760
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.175, + 0.893, + 0.218 + ], + "angle": 0, + "content": "Table 1. Comparison of Co-Speech Gesture Datasets. We summarize gesture and face datasets in talking scenarios. 'PGT' and 'MC' denote Pseudo Ground Truth and Motion Capture, respectively. The BEAT2 (Ours) dataset is the largest dataset for motion-captured data, providing holistic, academic community standard mesh-level information." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.229, + 0.885, + 0.297 + ], + "angle": 0, + "content": "
S2GTriModalB2HTWHHahibie et al.DisCoHA2GFace Former 2022[19]RhythmicBEAT CaMN 2022[39]Diff Gesture 2023[72]Code Talker 2023[61]Talk Show 2023[65]DiffStyle Gesture 2023[62]Body Former 2023[49]Lively Speaker 2023[70]EMAEG (Ours) 2024
2019 [25]2020 [67]2020[48]2021[32]2021[28]2022[38]2022[40]2022[5]2022[5]-------Masked
Input Gestures--BodyBody------------
Head---M--S---SS---C
UpperSS-MSC-SCS-CSSSC
Hands--SSMSC--CS-CS-SC
Lower + Global------------S--C
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.302, + 0.894, + 0.358 + ], + "angle": 0, + "content": "Table 2. Comparison of Co-Speech Gesture Models. We compare with previous methods for generating face or body motion trained on co-speech datasets. The first row lists their inputs, and the subsequent rows list their outputs, respectively. Different decoder designs are denoted by the initials 'S' for Single, 'M' for Multiple, and 'C' for Cascaded. To the best of our knowledge, EMAGE (Ours) is the first to accept audio and partially or completely masked gestures, generating full-body audio-synchronized results." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.385, + 0.471, + 0.566 + ], + "angle": 0, + "content": "ther the face or body. Applying these methods to the full body will yield sub-optimal results as audio is differently correlated with face and body dynamics. Most similar to our work, Habibie et al. [28] employ a single audio encoder and multiple decoders for generating face and body gestures. TalkSHOW [65] demonstrates the advantage of separating the audio encoders and decoding the body and hands using quantized codes in an auto-regressive manner. However, it lacks lower body and global motion, and there is no shared information for body and facial gesture generation. Moreover, it cannot accept partially masked body hints due to the design of a fully auto-regressive model." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.568, + 0.471, + 0.826 + ], + "angle": 0, + "content": "Masked Representation Learning was first demonstrated to be effective in Natural Language Processing, with BERT-based models [17, 31, 42] boosting the performance of learned word embeddings in downstream tasks through a combination of masked language modeling and transformer architecture. Subsequently, Masked AutoEncoders [29] expanded masked image modeling to computer vision by removing and inpainting image patches. This concept of masked representation learning has since been employed in other modalities, e.g., video [12, 18, 43], audio [26, 36, 37], and point clouds [55, 68, 69]. Most related to our work, MotionBERT [73] proposes a spatial-temporal transformer to learn a robust motion representation for classification-based tasks by masking 2D pose. In contrast to their method, we target robust motion features for conditional motion generation tasks, requiring a balance in training between multiple modalities, e.g., audio and gesture." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.844, + 0.165, + 0.859 + ], + "angle": 0, + "content": "3. BEAT2" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.871, + 0.471, + 0.903 + ], + "angle": 0, + "content": "This section introduces how we obtain unified mesh-level data, i.e., SMPL-X [50] and FLAME [35] parameters, from" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.384, + 0.895, + 0.522 + ], + "angle": 0, + "content": "the original BEAT dataset [39]. BEAT utilizes a Vicon motion capture system [47] and releases 78-joint skeleton-level Biovision Hierarchy (BVH) [10] motion files. Their facial capture system uses ARKit [8] with a depth camera on the iPhone 12 Pro, extracting \\(\\mathbb{R}^{51}\\) blendshape weights. These blendshapes are designed based on the Facial Action Coding System (FACS), which is widely adopted. However, both the motion and facial data lack mesh-level details (see Figure 2), e.g., shapes and vertices." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.543, + 0.871, + 0.56 + ], + "angle": 0, + "content": "3.1. Body Parameters Initialization via MoSh++" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.57, + 0.892, + 0.722 + ], + "angle": 0, + "content": "We initialize the SMPL-X body shape and pose parameters from BEAT mocap marker data using MoSh++ [46]. Given the captured markers positions \\(\\mathbf{m} \\in \\mathbb{R}^{T \\times K \\times 3}\\), predefined markers position offsets \\(\\mathbf{d} \\in \\mathbb{R}^{K \\times 3}\\), and user-defined vertices-to-markers function \\(\\mathcal{H}\\), we aim to obtain body shape \\(\\beta \\in \\mathbb{R}^{300}\\), poses \\(\\theta \\in \\mathbb{R}^{T \\times 55 \\times 3}\\), and translation parameters \\(\\gamma \\in \\mathbb{R}^{T \\times 3}\\). The optimization uses a differentiable surface vertex mapping function \\(S(\\beta, \\theta, \\gamma)\\) and vertex normal function \\(\\mathcal{N}(\\beta, \\theta_0, \\gamma_0)\\). For each frame, the latent marker \\(\\tilde{\\mathbf{m}} \\in \\mathbb{R}^{T \\times K \\times 3}\\) is calculated as" + }, + { + "type": "equation", + "bbox": [ + 0.566, + 0.743, + 0.892, + 0.761 + ], + "angle": 0, + "content": "\\[\n\\tilde {\\mathbf {m}} _ {i} \\equiv \\mathcal {S} _ {\\mathcal {H}} (\\beta , \\theta_ {i}, \\gamma_ {i}) + \\mathbf {d N} _ {\\mathcal {H}} (\\beta , \\theta_ {i}, \\gamma_ {i}). \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.78, + 0.895, + 0.903 + ], + "angle": 0, + "content": "We first select 12 frames to optimize and fix \\(\\mathbf{d}\\) and \\(\\beta\\), then optimize \\(\\theta_{i}\\), \\(\\gamma_{i}\\) for \\(i \\in (1:T)\\) by minimizing the loss terms based on \\(\\|\\tilde{\\mathbf{m}}_i - \\mathbf{m}_i\\|^2\\), including Data Term, Surface Distance Energy, Marker Initialization Regularization, Pose and Shape Priors, Velocity Constancy, and Soft-Tissue Term (see supplementary materials). The overall objective function is the weighted sum of these terms to balance accuracy and plausibility." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.957 + ], + "angle": 0, + "content": "1146" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.129, + 0.092, + 0.199, + 0.222 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.134, + 0.225, + 0.198, + 0.275 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.14, + 0.277, + 0.171, + 0.286 + ], + "angle": 0, + "content": "BEAT" + }, + { + "type": "image", + "bbox": [ + 0.203, + 0.093, + 0.299, + 0.222 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.204, + 0.225, + 0.298, + 0.275 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.23, + 0.278, + 0.279, + 0.287 + ], + "angle": 0, + "content": "Retargeted" + }, + { + "type": "image", + "bbox": [ + 0.309, + 0.093, + 0.399, + 0.222 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.31, + 0.225, + 0.405, + 0.275 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.338, + 0.277, + 0.376, + 0.286 + ], + "angle": 0, + "content": "Moshed" + }, + { + "type": "image", + "bbox": [ + 0.41, + 0.093, + 0.508, + 0.222 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.41, + 0.224, + 0.508, + 0.275 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.409, + 0.277, + 0.514, + 0.287 + ], + "angle": 0, + "content": "Refined Moshed (Ours)" + }, + { + "type": "image", + "bbox": [ + 0.525, + 0.097, + 0.613, + 0.222 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.525, + 0.225, + 0.618, + 0.274 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.557, + 0.277, + 0.587, + 0.286 + ], + "angle": 0, + "content": "BEAT" + }, + { + "type": "image", + "bbox": [ + 0.621, + 0.099, + 0.719, + 0.222 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.622, + 0.225, + 0.723, + 0.274 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.63, + 0.277, + 0.716, + 0.287 + ], + "angle": 0, + "content": "Wrapped Optimum" + }, + { + "type": "image", + "bbox": [ + 0.727, + 0.099, + 0.828, + 0.222 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.728, + 0.225, + 0.832, + 0.275 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.724, + 0.277, + 0.845, + 0.287 + ], + "angle": 0, + "content": "Handcraft Optimum (Ours)" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.292, + 0.894, + 0.377 + ], + "angle": 0, + "content": "Figure 2. Comparison of Data between BEAT2 and Others. BEAT-SMPLX-FLAME presents a new mesh-level, motion-captured, holistic co-speech gesture dataset with 60h of data. Left: We compare our refined SMPL-X body parameters (denoted as Refined MoSh) with the original BEAT skeleton [39], Retargeted from AutoRegPro, and initial results of \\(\\mathrm{Mosh}++\\) [50]. The refined results show correct neck flexion, appropriate head and neck shape ratios, and detailed finger representations. Right: Visualization of blendshape weights from the original BEAT dataset [39] with ARKit's template, Wrapped-based, and handcrafted optimization. Our final handcrafted FLAME blendshape-based optimization demonstrates both accurate lip movement details and natural mouth shapes." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.394, + 0.34, + 0.41 + ], + "angle": 0, + "content": "3.2. Body Parameters Refinement" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.423, + 0.473, + 0.651 + ], + "angle": 0, + "content": "MoSh++ produces unnatural head shapes due to the fact that head markers were worn on a helmet and it sometimes produces unnatural finger poses. Consequently, we refine body shape and pose parameters with three simple yet effective hard-coded physical rules. i) The neck and head length should approximate 1/7 of the total length of the body. ii) Fingers, except for the thumb, should not bend backward. iii) We employ the Kolmogorov-Smirnov test, which reveals the data is similar to a normal distribution. Subsequently, we apply a Gaussian truncation approach, where all data points falling outside the \\(3\\sigma\\) range are adjusted to align with the \\(3\\sigma\\) threshold and blended with the adjacent 10 frames. We compare our refined body parameters (denoted as MoSh Refined) with Original BEAT, Retargeted SMPL-X, and MoSh SMPL-X in Figure 2." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.676, + 0.452, + 0.692 + ], + "angle": 0, + "content": "3.3. Blendshape Weights to FLAME Parameters" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.704, + 0.473, + 0.903 + ], + "angle": 0, + "content": "Given the ARKit blendshape weights \\(\\mathbf{b}_{\\mathrm{ARKit}} \\in \\mathbb{R}^{T \\times 51}\\), we aim to derive a transformation matrix \\(\\mathbf{W} \\in \\mathbb{R}^{51 \\times 103}\\) that maps these into FLAME parameters \\(\\mathbf{b}_{\\mathrm{FLAME}} \\in \\mathbb{R}^{T \\times (100 + 3)}\\), where 100 represents the number of dimensions for expression parameters and 3 for jaw movement. Due to discrepancies between the iPhone's and FLAME's template mesh topology, simply optimizing FLAME parameters by minimizing the geometric error between the ARKit template vertices and the wrapped FLAME vertices does not yield satisfactory results. Instead, we release a set of handcrafted blendshape templates \\(\\mathbf{v}_{\\mathrm{template}} \\in \\mathbb{R}^{52}\\) on FLAME following ARKit's FACS configuration. This approach allows direct driving of the FLAME topology ver" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.393, + 0.88, + 0.411 + ], + "angle": 0, + "content": "tices \\(\\mathbf{v}\\in \\mathbb{R}^{T\\times 5023\\times 3}\\) from the given blendshape weights," + }, + { + "type": "equation", + "bbox": [ + 0.572, + 0.423, + 0.892, + 0.466 + ], + "angle": 0, + "content": "\\[\n\\mathbf {v} = \\mathbf {v} _ {\\text {t e m p l a t e}} ^ {0} + \\sum_ {j = 1} ^ {5 1} \\mathbf {b} _ {\\text {A R K i t}, j} \\cdot \\mathbf {v} _ {\\text {t e m p l a t e}} ^ {j} \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.478, + 0.892, + 0.602 + ], + "angle": 0, + "content": "where \\(\\mathbf{b}_{\\mathrm{ARKit},j}\\) is the weight of the \\(j\\)-th ARKit blend-shape and \\(\\mathbf{v}_{\\mathrm{template}}^j\\) is the vertex position of the FLAME model template corresponding to the \\(j\\)-th blendshape. The term \\(\\mathbf{v}_{\\mathrm{template}}^0\\) represents the original template vertex position of the FLAME model. We optimize \\(\\mathbf{W}\\) by minimizing \\(\\| \\tilde{\\mathbf{v}}_j - \\mathbf{v}_j\\| _2\\), where \\(\\tilde{\\mathbf{v}}\\) is obtained from FLAME's LBS \\(\\mathcal{V}(\\mathbf{b}_{\\mathrm{FLAME}})\\). The comparisons between ARKit data, wrap-based, and our approach are shown in Figure 2." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.618, + 0.599, + 0.633 + ], + "angle": 0, + "content": "4. IMAGE" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.644, + 0.892, + 0.901 + ], + "angle": 0, + "content": "We introduce the details of EMAGE, Expressive Masked Audio-Conditioned GEsture Modeling (see Figure 3). Given gestures \\(\\mathbf{g} \\in \\mathbb{R}^{T \\times (55 \\times 6 + 100 + 4 + 3)}\\), representing 55 joints in Rot6D, \\(\\mathbb{R}^{100}\\) FLAME parameters, \\(\\mathbb{R}^4\\) foot contact labels, \\(\\mathbb{R}^3\\) global translations, and speech audio \\(\\mathbf{a} \\in \\mathbb{R}^{T \\times sk}\\), where \\(sk = sr_{\\mathrm{audio}} / fps_{\\mathrm{gestures}}\\), EMAGE jointly optimizes masked gesture reconstruction and audio-conditioned gesture generation. This optimization enhances performance during inference and enables the use of partially masked gestures to complete holistic gestures. To achieve this, we first model the quantized latent space in Sec. 4.1, following [61, 65]. We then design a separated speech audio encoder using Content Rhythm Self-Attention (CRA) as detailed in Sec. 4.2. Subsequently, we learn the body hints from masked gestures via the Masked Audio Gesture Transformer, explained in Sec. 4.3. Finally, the decoding strategy for different body segments is discussed in Sec. 4.4." + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1147" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.119, + 0.095, + 0.868, + 0.273 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.28, + 0.892, + 0.378 + ], + "angle": 0, + "content": "Figure 3. EMAGE leverages two training paths: Masked Gesture Reconstruction (MG2G) and Audio-Conditioned Gesture Generation (A2G). The MG2G path focuses on encoding robust body hints through a spatial-temporal transformer gesture encoder and cross-attention gesture decoder. In contrast, the A2G path utilizes these body hints, and separated audio encoders to decode pretrained face and body latent features. A key component in this process is a switchable cross-attention layer, crucial for merging body hints and audio features. This fusion allows the features to be effectively disentangled and utilized for gesture decoding. Once the gesture latent features are reconstructed, EMAGE employs a pretrained VQ-Decoder to decode face and local body motions. Additionally, a pretrained Global Motion Predictor is used to estimate global body translations, further enhancing the model's capability to generate realistic and coherent gestures." + }, + { + "type": "image", + "bbox": [ + 0.153, + 0.405, + 0.272, + 0.52 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.277, + 0.404, + 0.396, + 0.52 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.529, + 0.47, + 0.643 + ], + "angle": 0, + "content": "Figure 4. Details of CRA and Pretrained VQ-VAEs. Left: Content Rhythm Attention fuses speech rhythm (onset and amplitude) with content (pretrained word embeddings from text scripts) adaptively. This results in a preference for either content or rhythm in specific frames, which encourages the generation of semantical-aware gestures. Right: We pretrain four compositional VQ-VAEs by reconstructing face, upper body, hands and lower body separately to disentangle audio-agnostic gestures explicitly." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.674, + 0.455, + 0.691 + ], + "angle": 0, + "content": "4.1. Compositional Discrete Face and Body Prior" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.701, + 0.468, + 0.867 + ], + "angle": 0, + "content": "We model full-body gestures in the separated quantized latent space (Figure 4) for several reasons. Similar to [65], the body and hands improves the diversity of the results. However, we additionally separate the face and lower body due to their differing correlations with audio, i.e., using a single VQ-VAE to encode both the upper and lower body may lead the model to overlook gestures that occur less frequently - regardless of their relation to the audio. Specifically, a single model might focus only on recovering lower body motion, neglecting elbow movements in a speaker who is constantly walking during the conversation." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.871, + 0.47, + 0.903 + ], + "angle": 0, + "content": "The separated quantized latent space \\( Q = \\{\\mathbf{q}_{\\mathrm{f}}, \\mathbf{q}_{u}, \\mathbf{q}_{\\mathrm{h}}, \\mathbf{q}_{\\mathrm{l}}\\} \\) for the Face \\( \\mathbf{g}_{\\mathrm{f}} \\in \\mathbb{R}^{T \\times 106} \\), Upper" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.403, + 0.892, + 0.463 + ], + "angle": 0, + "content": "body \\(\\mathbf{g}_{\\mathrm{u}} \\in \\mathbb{R}^{T \\times 78}\\), Hands \\(\\mathbf{g}_{\\mathrm{h}} \\in \\mathbb{R}^{T \\times 180}\\), and Lower body \\(\\mathbf{g}_{\\mathrm{l}} \\in \\mathbb{R}^{T \\times (54 + 4)}\\) are from four VQ-VAEs. The VQ-VAE for each is optimized by jointly optimizing the following loss terms," + }, + { + "type": "equation", + "bbox": [ + 0.613, + 0.465, + 0.892, + 0.492 + ], + "angle": 0, + "content": "\\[\nq _ {i} = \\underset {q _ {i} \\in \\mathbf {q}} {\\arg \\min } \\| z _ {j} - q _ {i} \\| ^ {2} \\tag {3}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.521, + 0.513, + 0.892, + 0.552 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} _ {\\mathrm {V Q - V A E}} = \\mathcal {L} _ {r e c} (\\mathbf {g}, \\hat {\\mathbf {g}}) + \\mathcal {L} _ {\\mathrm {v e l}} \\left(\\mathbf {g} ^ {\\prime}, \\hat {\\mathbf {g}} ^ {\\prime}\\right) + \\mathcal {L} _ {\\mathrm {a c c}} \\left(\\mathbf {g} ^ {\\prime \\prime}, \\hat {\\mathbf {g}} ^ {\\prime \\prime}\\right) \\\\ + \\left\\| \\operatorname {s g} [ \\mathbf {z} ] - \\mathbf {q} \\right\\| ^ {2} + \\left\\| \\mathbf {z} - \\operatorname {s g} [ \\mathbf {q} ] \\right\\| ^ {2}, \\tag {4} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.564, + 0.892, + 0.625 + ], + "angle": 0, + "content": "where \\(\\mathbf{z}\\) is the encoded \\(\\mathbf{g}\\) with a temporal window size \\(w = 1\\). \\(\\mathcal{L}_{\\mathrm{rec}}\\), \\(\\mathcal{L}_{\\mathrm{vel}}\\), and \\(\\mathcal{L}_{\\mathrm{acc}}\\) are Geodesic [58] and L1 losses, sg is a stop gradient operation, we set the weight of the commitment loss [59] (the last term) to 1 in this paper." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.635, + 0.812, + 0.651 + ], + "angle": 0, + "content": "4.2. Content and Rhythm Self-Attention" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.659, + 0.892, + 0.856 + ], + "angle": 0, + "content": "Given the speech audio, s, inspired by [4], we employ onset o and amplitude a as explicit audio rhythm, alongside the pretrained embeddings [9] e from transcripts as content. Different from previous approaches, which typically add the rhythm and content features, we leverage self-attention to merge these features adaptively. This approach is driven by the observation that for specific frames, the gestures are more related to content (semantic-aware) or rhythm (beat-aware). The rhythm and content features are first encoded into time-aligned features \\(\\mathbf{r}_{1:T}\\) and \\(\\mathbf{c}_{1:T}\\), using a Temporal Convolutional Network (TCN) and linear mapping, respectively. For each time step \\(t\\in \\{1,\\dots,T\\}\\), we merge the rhythm and content features by:" + }, + { + "type": "equation", + "bbox": [ + 0.577, + 0.868, + 0.892, + 0.904 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathbf {f} _ {1: T} = \\alpha \\times \\mathbf {r} _ {1: T} + (1 - \\alpha) \\times \\mathbf {c} _ {1: T}, \\tag {5} \\\\ \\alpha = \\operatorname {S o f t m a x} (\\mathcal {A T} (\\mathbf {r} _ {1: T}, \\mathbf {c} _ {1: T})), \\\\ \\end{array}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "1148" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.086, + 0.092, + 0.209, + 0.236 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.213, + 0.092, + 0.335, + 0.236 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.339, + 0.092, + 0.455, + 0.237 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.24, + 0.47, + 0.337 + ], + "angle": 0, + "content": "Figure 5. Comparison of Forward Path Designs. Straightforward fusion module (a) merges audio features without refined body features and recombines audio features based only on position embedding. The Self-Attention decoder module (b), adopted in previous MLM models [17, 31], is limited for tasks requiring auto-regressive inference. Our design (c) considers effective audio feature fusion and auto-regressive decoding." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.362, + 0.469, + 0.392 + ], + "angle": 0, + "content": "where \\(\\mathcal{A}\\mathcal{T}\\) is a 2-layer MLP. We apply two separate CRA encoders for the face and body." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.4, + 0.468, + 0.417 + ], + "angle": 0, + "content": "4.3. Masked Audio-Conditioned Gesture Modeling" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.423, + 0.469, + 0.544 + ], + "angle": 0, + "content": "We propose a Masked Audio Gesture Transformer to leverage different training paths (see Fig. 5 for the motivation behind the architecture designs). Given temporal and spatially masked gestures \\(\\overline{\\mathbf{g}}\\in \\mathbb{R}^{T\\times 337}\\), we first replace the masked tokens with learned masked embeddings \\(e_{\\mathrm{mask}}\\in \\mathbb{R}^{256}\\), as the value zero still represents specific motion content, e.g., a T-pose. We linearly increase the ratio of masked joints and frames from 0 to \\(95\\%\\) according to the training epochs." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.545, + 0.469, + 0.617 + ], + "angle": 0, + "content": "A spatial convolutional \\(\\mathcal{SC}\\) encoder initially summarizes the spatial information, and compress the spatial feature into \\(\\mathbb{R}^{T\\times 256}\\). We then employ a temporal self-attention (without feed-forward) \\(\\mathcal{TSA}\\) to refine the summarized spatial features" + }, + { + "type": "equation", + "bbox": [ + 0.187, + 0.619, + 0.469, + 0.635 + ], + "angle": 0, + "content": "\\[\n\\mathbf {h} = \\mathcal {T S A} (\\mathcal {S C} (\\overline {{\\mathbf {g}}}) + \\mathbf {p} _ {t}), \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.64, + 0.469, + 0.716 + ], + "angle": 0, + "content": "where \\(\\mathbf{h} \\in \\mathbb{R}^{T \\times 512}\\) represents encoded body hints, and \\(\\mathbf{p}_t\\) is the sum of a learned speaker embedding and PPE [19]. Then, a straightforward temporal cross-attention Transformer decoder \\(\\mathcal{TCA}\\mathcal{T}\\) is adopted for the reconstructed latent \\(\\hat{\\mathbf{q}}_{\\mathrm{mg2g}}\\)," + }, + { + "type": "equation", + "bbox": [ + 0.185, + 0.716, + 0.469, + 0.732 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathbf {q}} _ {\\mathrm {m g} 2 \\mathrm {g}} = \\mathcal {T C A T} (\\mathbf {h} + \\mathbf {p} _ {t}). \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.737, + 0.398, + 0.752 + ], + "angle": 0, + "content": "We minimize the L1 distance in the latent space," + }, + { + "type": "equation", + "bbox": [ + 0.197, + 0.761, + 0.469, + 0.778 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {m g} 2 \\mathrm {g}} = \\| \\hat {\\mathbf {q}} _ {\\mathrm {m g} 2 \\mathrm {g}} - \\mathbf {q} \\|. \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.786, + 0.469, + 0.876 + ], + "angle": 0, + "content": "Masked gesture reconstruction encodes effective body hints, with the key being to leverage these body hints for gesture generation. We employ a selective fusion for audio and body hints via a temporal cross-attention \\(\\mathcal{TCA}\\). Subsequently, we use the merged audio-gesture features for audio-conditioned gesture latent reconstruction," + }, + { + "type": "equation", + "bbox": [ + 0.12, + 0.886, + 0.469, + 0.903 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathbf {q}} _ {\\mathrm {a} 2 \\mathrm {g}} = \\mathcal {T C A T} \\left(\\mathcal {T C A} \\left(\\mathbf {h} + \\mathbf {p} _ {t}, \\mathbf {f} _ {\\text {b o d y}}\\right), \\overline {{\\mathbf {g}}} + \\mathbf {p} _ {t}\\right). \\tag {9}\n\\]" + }, + { + "type": "table", + "bbox": [ + 0.523, + 0.089, + 0.868, + 0.145 + ], + "angle": 0, + "content": "
Top.-BTop.-FBodyFace
VOCA[14]-FLAME-38.3 ± 5.63
AMASS[46]SMPL-XFLAME42.0 ± 3.60-
TalkSHOW[65]SMPL-XSMPL-X14.4 ± 2.1926.1 ± 6.42
BEAT2 (Ours)SMPL-XFLAME43.6 ± 3.3835.7 ± 5.91
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.151, + 0.892, + 0.234 + ], + "angle": 0, + "content": "Table 3. User Preference Win Rate between Existing Datasets. 'Top' denotes the topology of the mesh. The results show our BEAT2 dataset outperforms the existing PGT dataset [65] in both body and face aspects. It also performs slightly better on body and lower on face when compared with previous mocap [46] and 3D scan-based datasets [14], respectively." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.252, + 0.892, + 0.298 + ], + "angle": 0, + "content": "We optimize both the latent code class classification cross-entropy loss \\(\\mathcal{L}_{\\mathrm{a2g - rec}}\\) and the latent reconstruction loss \\(\\mathcal{L}_{\\mathrm{a2g - c1s}}\\) to encourage diverse results." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.306, + 0.776, + 0.322 + ], + "angle": 0, + "content": "4.4. Face and Translation Decoding" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.329, + 0.892, + 0.404 + ], + "angle": 0, + "content": "Considering that the face is weakly related to body motion, applying the same operation - recombining the audio features based on body hints - is not reasonable. Therefore, we directly concatenate the masked body hints with the audio features for the final decoding of facial latent," + }, + { + "type": "equation", + "bbox": [ + 0.601, + 0.414, + 0.891, + 0.43 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathbf {q}} _ {\\mathrm {f}} = \\mathcal {T} \\mathcal {C A T} \\left(\\mathbf {f} _ {\\text {f a c e}} \\oplus \\mathbf {h}, \\mathbf {p} _ {t}\\right). \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.439, + 0.892, + 0.5 + ], + "angle": 0, + "content": "Once we have obtained the local lower body motion \\(\\tilde{\\mathbf{g}}_l\\in\\) \\(\\mathbb{R}^{T\\times (54 + 4)}\\) from the VQ-Decoder, we estimate the global translations \\(\\tilde{\\mathbf{t}}\\in \\mathbb{R}^{T\\times 3}\\) with a pretrained Global Motion Predictor \\(\\tilde{\\mathbf{t}} = \\mathcal{G}(\\tilde{\\mathbf{g}}_l)\\), which significantly reduces foot sliding." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.512, + 0.632, + 0.529 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.538, + 0.892, + 0.689 + ], + "angle": 0, + "content": "We separate the evaluation into two categories: dataset quality and model ability. After removing data with low finger quality from BEAT, BEAT2 is reduced to 60 hours. We further split it into BEAT2-Standard (27 hours) and BEAT2-Additional (33 hours), based on the type of speech and conversation sections [39]. While acted gestures in speech sections are more diverse and expressive, spontaneous gestures in conversational sections tend to be more natural yet less varied. We report the results on BEAT2-Standard Speaker2 with an \\(85\\% / 7.5\\% / 7.5\\%\\) train/val/test split." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.697, + 0.745, + 0.713 + ], + "angle": 0, + "content": "5.1. Dataset Quality Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.892, + 0.901 + ], + "angle": 0, + "content": "We compare our dataset with the state-of-the-art pseudo ground truth (PGT) dataset, TalkSHOW [65], for both face and body. Additionally, we compare it with the AMASS [46] dataset for the body and VOCA [14] for the face as references. Due to the varying lengths of sequences in each dataset, we sample 100 comparison pairs, each with an equal duration ranging from 2 to 4 seconds. In a perceptual study, each participant evaluates a random set of 40 pairs during a 10-minute session by selecting the sequence they consider to have the best captured quality. In total, 60 participants were invited. It is important to note that participants are instructed to compare only the upper body results," + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "1149" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.138, + 0.113, + 0.382, + 0.127 + ], + "angle": 0, + "content": "... In my spare time ... ... if I feel ok ..." + }, + { + "type": "text", + "bbox": [ + 0.415, + 0.113, + 0.831, + 0.128 + ], + "angle": 0, + "content": "... I will go for a walk or hike in nature. I try to organize something" + }, + { + "type": "image", + "bbox": [ + 0.136, + 0.131, + 0.258, + 0.2 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.278, + 0.131, + 0.396, + 0.2 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.419, + 0.131, + 0.54, + 0.2 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.561, + 0.132, + 0.68, + 0.2 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.709, + 0.133, + 0.826, + 0.2 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.131, + 0.202, + 0.272, + 0.274 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.285, + 0.202, + 0.397, + 0.273 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.419, + 0.202, + 0.54, + 0.274 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.561, + 0.202, + 0.678, + 0.274 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.699, + 0.203, + 0.819, + 0.274 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.138, + 0.277, + 0.252, + 0.346 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.282, + 0.276, + 0.391, + 0.346 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.418, + 0.276, + 0.536, + 0.346 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.55, + 0.275, + 0.678, + 0.346 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.695, + 0.275, + 0.826, + 0.348 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.133, + 0.349, + 0.261, + 0.416 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.279, + 0.349, + 0.397, + 0.416 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.418, + 0.348, + 0.537, + 0.416 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.552, + 0.347, + 0.664, + 0.422 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.693, + 0.349, + 0.832, + 0.421 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.427, + 0.894, + 0.497 + ], + "angle": 0, + "content": "Figure 6. Subjective Examples. Top to down: GroundTruth, Generated W/O Body Hints, Generated With Body Hints, Visible Body Hints, EMAGE generates diverse, semantic-aware and audio synchronized body gestures, e.g. raise the both hands for \"spare time\", relaxing gestures for \"hike in nature\". Besides, as shown in the third and bottom rows, EMAGE is flexible to accept non-audio-synchronized body hints on any frames and joints, to guide or customize the generated gestures explicitly, e.g., repeat a similar motion for raising hands, change the walking orientations, etc. Note that in this figure, The gray joints of generated results are not copies of visible hints." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.518, + 0.466, + 0.662 + ], + "angle": 0, + "content": "
FGD ↓BC ↑Diversity ↑MSE ↓LVD ↓
FaceFormer[19]---7.7877.593
CodeTalker[61]---8.0267.766
S2G[25]28.154.6835.971--
Trimodal[67]12.415.9337.724--
HA2G[40]12.326.7798.626--
DisCo[38]9.4176.4399.912--
CaMN[39]6.6446.76910.86--
DiffStyleGesture[62]8.8117.24111.49--
Habibie et al.[28]9.0407.7168.2138.6148.043
TalkSHOW[65]6.2096.94713.477.7917.771
EMAGE (Ours)5.5127.72413.067.6807.556
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.665, + 0.469, + 0.736 + ], + "angle": 0, + "content": "Table 4. Quantitative evaluation on BEATv2. We report FGD \\(\\times 10^{-1}\\), BC \\(\\times 10^{-1}\\), Diversity, MSE \\(\\times 10^{-8}\\), and LVD \\(\\times 10^{-5}\\). For body gestures, EMAGE significantly improves the FGD, indicating that the generated results are closer to the GT. This shows the effectiveness of body hints from masked gesture modeling." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.755, + 0.469, + 0.785 + ], + "angle": 0, + "content": "as TalkSHOW contains only the upper body. The results are shown in Table 3." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.796, + 0.308, + 0.813 + ], + "angle": 0, + "content": "5.2. Model Ability Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.82, + 0.469, + 0.866 + ], + "angle": 0, + "content": "We report results using BEATv1.3 and BEATv2. The latter has facial data refined by animators and hand data filtered by annotators." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.469, + 0.902 + ], + "angle": 0, + "content": "Metrics. We adopt FGD[67] to evaluate the realism of the body gestures. Then, we measure Diversity[33] by calcu" + }, + { + "type": "table", + "bbox": [ + 0.532, + 0.519, + 0.861, + 0.568 + ], + "angle": 0, + "content": "
HolisticBodyFace
Habibie et al.[28]12.4 ± 3.7015.9 ± 6.4910.8 ± 3.19
TalkSHOW[65]34.9 ± 5.7940.4 ± 8.2233.2 ± 6.03
EMAGE (Ours)52.7 ± 7.9144.7 ± 8.6856.0 ± 7.80
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.572, + 0.892, + 0.629 + ], + "angle": 0, + "content": "Table 5. User Preference Win Rate On Generated Results. The results indicate that our generated outcomes are perceived as more realistic and believable, with a \\(14\\%\\) and \\(23\\%\\) higher user preference for body and face gestures, respectively." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.671, + 0.892, + 0.747 + ], + "angle": 0, + "content": "lating the average L1 distance between multiple body gesture clips, and use BC [34] to assess speech-motion synchrony. For the face, we calculate the vertex MSE[61] to measure the positional distance and the vertex L1 difference LVD [65] between the GT and generated facial vertices." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.773, + 0.893, + 0.909 + ], + "angle": 0, + "content": "Compared Methods. We first compare our method with representative state-of-the-art methods in body gesture generation [25, 38-40, 62, 67] and talking head generation [19, 61] by reproducing their methods for body and face, respectively. In addition, we reproduce two previous state-of-the-art holistic pipelines, Habibie et al. [28] and TalkSHOW [65], whose original implementations are limited to the upper body. We add a lower-body decoder to Habibie et al. and a lower-body VQ-VAE to TalkSHOW." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1150" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.097, + 0.088, + 0.437, + 0.257 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.261, + 0.473, + 0.372 + ], + "angle": 0, + "content": "Figure 7. Comparison of Generated Facial Movements. Compared with previous state-of-the-art talking face generation methods FaceFormer [19] and CodeTalker [61] as well as holistic gesture generation methods Habibie et al. [28] and TalkSHOW [65]. Note that CodeTalker has higher MSE than IMAGE on BEATv2 (Table 4, lower is better) but is subjectively realistic. IMAGE gets good lip motions by leveraging both the face model and masked body hints." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.385, + 0.395, + 0.4 + ], + "angle": 0, + "content": "5.2.1 Quantitative and Qualitative Analysis" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.402, + 0.471, + 0.507 + ], + "angle": 0, + "content": "As shown in Table 4, with a four-frame seed pose, our method outperforms previous SOTA algorithms. For qualitative results, see Figures 6 and 7. Furthermore, we conduct a perceptual study. Maintaining the same setup of 60 participants, each participant evaluates 40 pairs of 10-second results do decide which is most believable; this gives the win rate shown in Table 5." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.528, + 0.257, + 0.543 + ], + "angle": 0, + "content": "5.2.2 Ablation Analysis" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.553, + 0.47, + 0.644 + ], + "angle": 0, + "content": "Performance of Baseline. As shown in Table 6, we start with a teacher-force transformer-based baseline. This baseline, inspired by FaceFormer [19], adopts a 1-layer cross-attention transformer decoder and replaces the audio features from Wav2Vec2 [6] with our custom TCN [7] and trainable word embeddings [9]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.644, + 0.471, + 0.764 + ], + "angle": 0, + "content": "Effect of Multiple VQ-VAEs. Simply applying one VQ-VAE [27, 59] for full-body movements, including the face, decreases performance in facial movements. This is because the VQ-VAE is trained to minimize the average loss for the full body, and some speakers' most frequent movements are not related to the audio. Implementing separated VQ-VAEs allows the model to better leverage the advantages of discrete priors." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.47, + 0.856 + ], + "angle": 0, + "content": "Effect of Content Rhythm Self-Attention. Adaptive fusion of rhythm and content features shows improvement in both FGD and Alignment. It selectively focuses the current motion more on rhythm or content features based on the training data's distribution. Moreover, we observed more semantic-aware results when applying CRA." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Effect of Masked Body Gesture Hints. The improvement across all objective metrics demonstrates that our model effectively leverages spatial-temporal gesture priors, reducing" + }, + { + "type": "table", + "bbox": [ + 0.507, + 0.088, + 0.887, + 0.194 + ], + "angle": 0, + "content": "
FGD ↓BC ↑Diversity ↑MSE ↓LVD ↓
Ground Truth06.89613.07400
Reconstruction3.9136.75813.1450.8416.389
Baseline13.0806.9418.31451.4429.317
+ VQVAE9.7876.67310.6241.6199.473
+ 4 VQVAE7.3976.69812.5441.2438.938
+ CRA6.8336.78312.6761.1868.715
+ Masked Hints5.4236.79413.0571.1809.015
" + }, + { + "type": "table_caption", + "bbox": [ + 0.567, + 0.199, + 0.826, + 0.212 + ], + "angle": 0, + "content": "Table 6. Abliation Analysis on BEATv1.3" + }, + { + "type": "table", + "bbox": [ + 0.522, + 0.232, + 0.871, + 0.282 + ], + "angle": 0, + "content": "
bodyhandsfaceFGD ↓BC ↑Diversity ↑
BESTX✓✓✓5.4236.79413.057
+ Trinity[22]✓5.3196.84313.346
+ AMASS[46]✓✓5.1746.76914.318
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.286, + 0.892, + 0.342 + ], + "angle": 0, + "content": "Table 7. Training IMAGE on Multiple Datasets. IMAGE demonstrates flexibility by training on multiple datasets, even when only a subset of holistic gestures is available. This approach further improves the objective metrics on the BEATv1.3 test set." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.363, + 0.892, + 0.41 + ], + "angle": 0, + "content": "the likelihood of incorrect gesture sampling. Furthermore, masked gesture modeling is key to enabling the network to accept predefined gestures in specific frames." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.433, + 0.79, + 0.449 + ], + "angle": 0, + "content": "5.2.3 Ability of Multi-Dataset Training." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.462, + 0.892, + 0.598 + ], + "angle": 0, + "content": "We demonstrate that EMAGE can effectively combine multiple non-holistic datasets for training by jointly training with the Trinity [22] and AMASS [46] datasets, by using only the upper body and audio pairs from Trinity and only body and hands from AMASS. Separate VQ-VAEs are trained for BEAT2, Trinity, and AMASS, with separate MLP heads implemented for codebook classification. The results in Table 7 show that incorporating data improves performance on the BEATv1.3 test set." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.613, + 0.619, + 0.629 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.642, + 0.892, + 0.807 + ], + "angle": 0, + "content": "In this work, we present IMAGE, a framework designed to accept partial gestures as input for completing audio-synchronized holistic gestures. It demonstrates that leveraging masked gesture reconstruction can significantly enhance the performance of audio-conditioned gesture generation. Furthermore, the design of IMAGE enables the training on multiple datasets, further improving performance. Alongside IMAGE, we release BEAT2, the largest multi-modal gesture dataset consistent with SMPL-X and FLAME. We hope that BEAT2 will contribute to knowledge and model sharing across various subareas." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.81, + 0.892, + 0.9 + ], + "angle": 0, + "content": "Disclosures: While You Zhou, Xuefei Zhe, Naoya Iwamoto, and Bo Zheng are employees of Huawei Tokyo Research Center, this work was done on their own time with the approval of their employer. MJB CoI: https://files.is.tue.mpg.de/black/CoI_CVPR_2024.txt" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "1151" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.17 + ], + "angle": 0, + "content": "[1] Kfir Aberman, Peizhuo Li, Dani Lischinski, Olga Sorkine-Hornung, Daniel Cohen-Or, and Baoquan Chen. Skeleton-aware networks for deep motion retargeting. ACM Transactions on Graphics (TOG), 39(4):62-1, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.171, + 0.472, + 0.24 + ], + "angle": 0, + "content": "[2] Chaitanya Ahuja, Dong Won Lee, Yukiko I Nakano, and Louis-Philippe Morency. Style transfer for co-speech gesture animation: A multi-speaker conditional-mixture approach. In European Conference on Computer Vision, pages 248–265. Springer, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.242, + 0.47, + 0.297 + ], + "angle": 0, + "content": "[3] Simon Alexanderson, Gustav Eje Henter, Taras Kucherenko, and Jonas Beskow. Style-controllable speech-driven gesture synthesis using normalising flows. In Computer Graphics Forum, pages 487-496. Wiley Online Library, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.298, + 0.468, + 0.338 + ], + "angle": 0, + "content": "[4] Tenglong Ao, Zeyi Zhang, and Libin Liu. Gesture diffu-clip: Gesture diffusion model with clip latents. ACM Trans. Graph. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.34, + 0.47, + 0.396 + ], + "angle": 0, + "content": "[5] Tenglong Ao, Qingzhe Gao, Yuke Lou, Baoquan Chen, and Libin Liu. Rhythmic gesticulator: Rhythm-aware co-speech gesture synthesis with hierarchical neural embeddings. ACM Transactions on Graphics (TOG), 41(6):1-19, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.397, + 0.468, + 0.451 + ], + "angle": 0, + "content": "[6] Alexei Baevski, Yuhao Zhou, Abdelrahman Mohamed, and Michael Auli. wav2vec 2.0: A framework for self-supervised learning of speech representations. Advances in neural information processing systems, 33:12449-12460, 2020. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.453, + 0.468, + 0.507 + ], + "angle": 0, + "content": "[7] Shaojie Bai, J Zico Kolter, and Vladlen Koltun. An empirical evaluation of generic convolutional and recurrent networks for sequence modeling. arxiv 2018. arXiv preprint arXiv:1803.01271, 2, 1803. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.509, + 0.468, + 0.605 + ], + "angle": 0, + "content": "[8] Gilad Baruch, Zhuoyuan Chen, Afshin Dehghan, Tal Dimry, Yuri Feigin, Peter Fu, Thomas Gebauer, Brandon Joffe, Daniel Kurz, Arik Schwartz, and Elad Shulman. ARK-scenes - a diverse real-world dataset for 3d indoor scene understanding using mobile RGB-d data. In Thirty-fifth Conference on Neural Information Processing Systems Datasets and Benchmarks Track (Round 1), 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.607, + 0.468, + 0.661 + ], + "angle": 0, + "content": "[9] Piotr Bojanowski, Edouard Grave, Armand Joulin, and Tomas Mikolov. Enriching word vectors with subword information. Transactions of the association for computational linguistics, 5:135-146, 2017. 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.663, + 0.468, + 0.704 + ], + "angle": 0, + "content": "[10] Biovision BVH. Biovision bvh. In https://research.cs.wisc.edu/graphics /Courses/ cs-838-1999/Jeff/BVH.html, 1999. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.705, + 0.468, + 0.773 + ], + "angle": 0, + "content": "[11] Zhe Cao, Gines Hidalgo, Tomas Simon, Shih-En Wei, and Yaser Sheikh. Openpose: realtime multi-person 2d pose estimation using part affinity fields. IEEE transactions on pattern analysis and machine intelligence, 43(1):172-186, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.775, + 0.468, + 0.829 + ], + "angle": 0, + "content": "[12] Zhe Chen, Yuchen Duan, Wenhai Wang, Junjun He, Tong Lu, Jifeng Dai, and Yu Qiao. Vision transformer adapter for dense predictions. arXiv preprint arXiv:2205.08534, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.831, + 0.468, + 0.899 + ], + "angle": 0, + "content": "[13] Kiran Chhatre, Radek Danček, Nikos Athanasiou, Giorgio Becherini, Christopher Peters, Michael J Black, and Timo Bolkart. Emotional speech-driven 3d body animation via disentangled latent diffusion. arXiv preprint arXiv:2312.04466, 2023. 1" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.472, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.147 + ], + "angle": 0, + "content": "[14] Daniel Cudeiro, Timo Bolkart, Cassidy Laidlaw, Anurag Ranjan, and Michael Black. Capture, learning, and synthesis of 3D speaking styles. Computer Vision and Pattern Recognition (CVPR), pages 10101-10111, 2019. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.151, + 0.892, + 0.205 + ], + "angle": 0, + "content": "[15] Radek Dančecek, Kiran Chhatre, Shashank Tripathi, Yandong Wen, Michael Black, and Timo Bolkart. Emotional speech-driven animation with content-emotion disentanglement. ACM, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.209, + 0.892, + 0.248 + ], + "angle": 0, + "content": "[16] Anna Deichler, Kiran Chhatre, Christopher Peters, and Jonas Beskow. Spatio-temporal priors in 3d human motion. 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.253, + 0.892, + 0.307 + ], + "angle": 0, + "content": "[17] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.311, + 0.892, + 0.393 + ], + "angle": 0, + "content": "[18] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.397, + 0.892, + 0.466 + ], + "angle": 0, + "content": "[19] Yingruo Fan, Zhaojiang Lin, Jun Saito, Wenping Wang, and Taku Komura. Faceformer: Speech-driven 3d facial animation with transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2, 3, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.469, + 0.892, + 0.538 + ], + "angle": 0, + "content": "[20] Yingruo Fan, Zhaojiang Lin, Jun Saito, Wenping Wang, and Taku Komura. Faceformer: Speech-driven 3d facial animation with transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18770-18780, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.541, + 0.892, + 0.595 + ], + "angle": 0, + "content": "[21] Gabriele Fanelli, Juergen Gall, Harald Romsdorfer, Thibaut Weise, and Luc Van Gool. A 3-d audio-visual corpus of affective communication. IEEE Transactions on Multimedia, 12(6):591-598, 2010. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.599, + 0.892, + 0.655 + ], + "angle": 0, + "content": "[22] Ylva Ferstl and Rachel McDonnell. Investigating the use of recurrent motion modelling for speech gesture generation. In Proceedings of the 18th International Conference on Intelligent Virtual Agents, pages 93-98, 2018. 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.657, + 0.892, + 0.726 + ], + "angle": 0, + "content": "[23] Erik Gartner, Mykhaylo Andriluka, Erwin Coumans, and Cristian Sminchisescu. Differentiable dynamics for articulated 3d human motion reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13190-13200, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.729, + 0.892, + 0.784 + ], + "angle": 0, + "content": "[24] Saeed Ghorbani, Ylva Ferstl, Daniel Holden, Nikolaus F. Troje, and Marc-Andre Carbonneau. Zeroeggs: Zero-shot example-based gesture generation from speech. Computer Graphics Forum, 42(1):206-216, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.787, + 0.892, + 0.857 + ], + "angle": 0, + "content": "[25] Shiry Ginosar, Amir Bar, Gefen Kohavi, Caroline Chan, Andrew Owens, and Jitendra Malik. Learning individual styles of conversational gesture. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3497-3506, 2019. 1, 2, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.859, + 0.892, + 0.899 + ], + "angle": 0, + "content": "[26] Yuan Gong, Yu-An Chung, and James Glass. Ast: Audio spectrogram transformer. arXiv preprint arXiv:2104.01778, 2021.3" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1152" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.148 + ], + "angle": 0, + "content": "[27] Chuan Guo, Xinxin Zuo, Sen Wang, and Li Cheng. Tm2t: Stochastic and tokenized modeling for the reciprocal generation of 3d human motions and texts. In European Conference on Computer Vision, pages 580-597. Springer, 2022. 8, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.149, + 0.471, + 0.219 + ], + "angle": 0, + "content": "[28] Ikhsanul Habibie, Weipeng Xu, Dushyant Mehta, Lingjie Liu, Hans-Peter Seidel, Gerard Pons-Moll, Mohamed Elgharib, and Christian Theobalt. Learning speech-driven 3d conversational gestures from video. arXiv preprint arXiv:2102.06837, 2021. 1, 2, 3, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.221, + 0.471, + 0.289 + ], + "angle": 0, + "content": "[29] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16000-16009, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.291, + 0.471, + 0.36 + ], + "angle": 0, + "content": "[30] Catalin Ionescu, Dragos Papava, Vlad Olaru, and Cristian Sminchisescu. Human3. 6m: Large scale datasets and predictive methods for 3d human sensing in natural environments. IEEE transactions on pattern analysis and machine intelligence, 36(7):1325-1339, 2013. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.362, + 0.47, + 0.417 + ], + "angle": 0, + "content": "[31] Zhenzhong Lan, Mingda Chen, Sebastian Goodman, Kevin Gimpel, Piyush Sharma, and Radu Soricut. Albert: A lite bert for self-supervised learning of language representations. arXiv preprint arXiv:1909.11942, 2019. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.419, + 0.471, + 0.516 + ], + "angle": 0, + "content": "[32] Gilwoo Lee, Zhiwei Deng, Shugao Ma, Takaaki Shiratori, Siddhartha S Srinivasa, and Yaser Sheikh. Talking with hands \\(16.2\\mathrm{m}\\): A large-scale dataset of synchronized body-finger motion and audio for conversational motion analysis and synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 763-772, 2019. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.518, + 0.471, + 0.601 + ], + "angle": 0, + "content": "[33] Jing Li, Di Kang, Wenjie Pei, Xuefei Zhe, Ying Zhang, Zhenyu He, and Linchao Bao. Audio2gestures: Generating diverse gestures from speech audio with conditional variational autoencoders. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11293-11302, 2021. 1, 2, 7, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.603, + 0.471, + 0.671 + ], + "angle": 0, + "content": "[34] Ruilong Li, Shan Yang, David A Ross, and Angjoo Kanazawa. Ai choreographer: Music conditioned 3d dance generation with aist++. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13401-13412, 2021. 7, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.673, + 0.47, + 0.73 + ], + "angle": 0, + "content": "[35] Tianye Li, Timo Bolkart, Michael. J. Black, Hao Li, and Javier Romero. Learning a model of facial shape and expression from 4D scans. ACM Transactions on Graphics, (Proc. SIGGRAPH Asia), 36(6):194:1-194:17, 2017. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.731, + 0.47, + 0.787 + ], + "angle": 0, + "content": "[36] Haiyang Liu and Cheng Zhang. Reinforcement learning based neural architecture search for audio tagging. In 2020 International Joint Conference on Neural Networks (IJCNN), pages 1-8. IEEE, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.788, + 0.47, + 0.843 + ], + "angle": 0, + "content": "[37] Haiyang Liu and Jihan Zhang. Improving ultrasound tongue image reconstruction from lip images using self-supervised learning and attention mechanism. arXiv preprint arXiv:2106.11769, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.845, + 0.47, + 0.901 + ], + "angle": 0, + "content": "[38] Haiyang Liu, Naoya Iwamoto, Zihao Zhu, Zhengqing Li, You Zhou, Elif Bozkurt, and Bo Zheng. Disco: Disentangled implicit content and rhythm learning for diverse co-speech gestures synthesis. In Proceedings of the 30th ACM Interna-" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.471, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.891, + 0.12 + ], + "angle": 0, + "content": "tional Conference on Multimedia, pages 3764-3773, 2022. 2,3,7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.122, + 0.894, + 0.19 + ], + "angle": 0, + "content": "[39] Haiyang Liu, Zihao Zhu, Naoya Iwamoto, Yichen Peng, Zhengqing Li, You Zhou, Elif Bozkurt, and Bo Zheng. Beat: A large-scale semantic and emotional multi-modal dataset for conversational gestures synthesis. arXiv preprint arXiv:2203.05297, 2022. 1, 2, 3, 4, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.193, + 0.894, + 0.275 + ], + "angle": 0, + "content": "[40] Xian Liu, Qianyi Wu, Hang Zhou, Yinghao Xu, Rui Qian, Xinyi Lin, Xiaowei Zhou, Wayne Wu, Bo Dai, and Bolei Zhou. Learning hierarchical cross-modal association for cospeech gesture generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10462-10472, 2022. 2, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.277, + 0.893, + 0.36 + ], + "angle": 0, + "content": "[41] Xian Liu, Qianyi Wu, Hang Zhou, Yinghao Xu, Rui Qian, Xinyi Lin, Xiaowei Zhou, Wayne Wu, Bo Dai, and Bolei Zhou. Learning hierarchical cross-modal association for cospeech gesture generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10462-10472, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.362, + 0.893, + 0.43 + ], + "angle": 0, + "content": "[42] Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. Roberta: A robustly optimized bert pretraining approach. arXiv preprint arXiv:1907.11692, 2019.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.433, + 0.893, + 0.502 + ], + "angle": 0, + "content": "[43] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10012-10022, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.504, + 0.893, + 0.559 + ], + "angle": 0, + "content": "[44] Matthew M. Loper, Naureen Mahmood, and Michael J. Black. MoSh: Motion and shape capture from sparse markers. ACM Transactions on Graphics, (Proc. SIGGRAPH Asia), 33(6):220:1-220:13, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.561, + 0.893, + 0.602 + ], + "angle": 0, + "content": "[45] Shuhong Lu, Youngwoo Yoon, and Andrew Feng. Co-speech gesture synthesis using discrete gesture token learning. arXiv preprint arXiv:2303.12822, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.604, + 0.893, + 0.671 + ], + "angle": 0, + "content": "[46] Naureen Mahmood, Nima Ghorbani, Nikolaus F. Troje, Gerard Pons-Moll, and Michael J. Black. AMASS: Archive of motion capture as surface shapes. In International Conference on Computer Vision, pages 5442-5451, 2019. 2, 3, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.675, + 0.893, + 0.716 + ], + "angle": 0, + "content": "[47] Tim Massey. Vicon study of dynamic object tracking accuracy. In https://www.vicon.com/resources/blog/vicon-study-of-dynamic-object-tracking-accuracy/, 2020. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.718, + 0.893, + 0.786 + ], + "angle": 0, + "content": "[48] Evonne Ng, Shiry Ginosar, Trevor Darrell, and Hanbyul Joo. Body2hands: Learning to infer 3d hands from conversational gesture body dynamics. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11865-11874, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.788, + 0.893, + 0.857 + ], + "angle": 0, + "content": "[49] Kunkun Pang, Dafei Qin, Yingruo Fan, Julian Habekost, Takaaki Shiratori, Junichi Yamagishi, and Taku Komura. Bodyformer: Semantics-guided 3d body gesture synthesis with transformer. ACM Transactions on Graphics (TOG), 42(4):1-12, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.859, + 0.893, + 0.901 + ], + "angle": 0, + "content": "[50] Georgios Pavlakos, Vasileios Choutas, Nima Ghorbani, Timo Bolkart, Ahmed A. A. Osman, Dimitrios Tzionas, and Michael J. Black. Expressive body capture: 3D hands, face," + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1153" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.092, + 0.47, + 0.134 + ], + "angle": 0, + "content": "and body from a single image. In Proceedings IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), pages 10975-10985, 2019. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.137, + 0.47, + 0.207 + ], + "angle": 0, + "content": "[51] Ziqiao Peng, Haoyu Wu, Zhenbo Song, Hao Xu, Xiangyu Zhu, Jun He, Hongyan Liu, and Zhaoxin Fan. Emotalk: Speech-driven emotional disentanglement for 3d face animation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 20687-20697, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.209, + 0.471, + 0.264 + ], + "angle": 0, + "content": "[52] Mathis Petrovich, Michael J Black, and Gül Varol. Action-conditioned 3d human motion synthesis with transformer vae. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10985-10995, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.267, + 0.47, + 0.322 + ], + "angle": 0, + "content": "[53] Xingqun Qi, Chen Liu, Lincheng Li, Jie Hou, Haoran Xin, and Xin Yu. Emotiongesture: Audio-driven diverse emotional co-speech 3d gesture generation. arXiv preprint arXiv:2305.18891, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.325, + 0.47, + 0.395 + ], + "angle": 0, + "content": "[54] Xingqun Qi, Jiahao Pan, Peng Li, Ruibin Yuan, Xiaowei Chi, Mengfei Li, Wenhan Luo, Wei Xue, Shanghang Zhang, Qifeng Liu, et al. Weakly-supervised emotion transition learning for diverse 3d co-speech gesture generation. arXiv preprint arXiv:2311.17532, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.397, + 0.47, + 0.438 + ], + "angle": 0, + "content": "[55] Guocheng Qian, Xingdi Zhang, Abdullah Hamdi, and Bernard Ghanem. Pix4point: Image pretrained transformers for 3d point cloud understanding. 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.442, + 0.47, + 0.511 + ], + "angle": 0, + "content": "[56] Alexander Richard, Michael Zollhöfer, Yandong Wen, Fernando De la Torre, and Yaser Sheikh. Meshtalk: 3d face animation from speech using cross-modality disentanglement. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1173-1182, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.513, + 0.47, + 0.569 + ], + "angle": 0, + "content": "[57] Kaede Shiohara, Xingchao Yang, and Takafumi Taketomi. Blendface: Re-designing identity encoders for face-swapping. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7634-7644, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.572, + 0.47, + 0.64 + ], + "angle": 0, + "content": "[58] Tommi Tykkälä, Cédric Audras, and Andrew I Comport. Direct iterative closest point for real-time visual odometry. In 2011 IEEE International Conference on Computer Vision Workshops (ICCV Workshops), pages 2050-2056. IEEE, 2011. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.643, + 0.47, + 0.685 + ], + "angle": 0, + "content": "[59] Aaron Van Den Oord, Oriol Vinyals, et al. Neural discrete representation learning. Advances in neural information processing systems, 30, 2017. 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.688, + 0.47, + 0.743 + ], + "angle": 0, + "content": "[60] Kaisiyuan Wang, Qianyi Wu, Linsen Song, Zhuoqian Yang, Wayne Wu, Chen Qian, Ran He, Yu Qiao, and Chen Change Loy. Mead: A large-scale audio-visual dataset for emotional talking-face generation. In ECCV, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.746, + 0.47, + 0.828 + ], + "angle": 0, + "content": "[61] Jinbo Xing, Menghan Xia, Yuechen Zhang, Xiaodong Cun, Jue Wang, and Tien-Tsin Wong. Codetalker: Speech-driven 3d facial animation with discrete motion prior. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12780-12790, 2023. 1, 2, 3, 4, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.832, + 0.47, + 0.9 + ], + "angle": 0, + "content": "[62] Sicheng Yang, Zhiyong Wu, Minglei Li, Zhensong Zhang, Lei Hao, Weihong Bao, Ming Cheng, and Long Xiao. Diffusestylegesture: Stylized audio-driven co-speech gesture generation with diffusion models. arXiv preprint arXiv:2305.04919, 2023. 2, 3, 7" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.471, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.147 + ], + "angle": 0, + "content": "[63] Xingchao Yang and Takafumi Taketomi. Bareskinnet: Demakeup and de-lighting via 3d face reconstruction. In Computer Graphics Forum, pages 623-634. Wiley Online Library, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.149, + 0.892, + 0.217 + ], + "angle": 0, + "content": "[64] Xingchao Yang, Takafumi Taketomi, and Yoshihiro Kanamori. Makeup extraction of 3d representation via illumination-aware image decomposition. In Computer Graphics Forum, pages 293-307. Wiley Online Library, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.22, + 0.892, + 0.275 + ], + "angle": 0, + "content": "[65] Hongwei Yi, Hualin Liang, Yifei Liu, Qiong Cao, Yandong Wen, Timo Bolkart, Dacheng Tao, and Michael J Black. Generating holistic 3d human motion from speech. In CVPR, 2023. 1, 2, 3, 4, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.277, + 0.892, + 0.357 + ], + "angle": 0, + "content": "[66] Youngwoo Yoon, Woo-Ri Ko, Minsu Jang, Jaeyeon Lee, Jae-hong Kim, and Geehyuk Lee. Robots learn social skills: End-to-end learning of co-speech gesture generation for humanoid robots. In 2019 International Conference on Robotics and Automation (ICRA), pages 4303-4309. IEEE, 2019. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.361, + 0.892, + 0.43 + ], + "angle": 0, + "content": "[67] Youngwoo Yoon, Bok Cha, Joo-Haeng Lee, Minsu Jang, Jaeyeon Lee, Jaehong Kim, and Geehyuk Lee. Speech gesture generation from the trimodal context of text, audio, and speaker identity. ACM Transactions on Graphics (TOG), 39 (6):1-16, 2020. 1, 2, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.432, + 0.892, + 0.5 + ], + "angle": 0, + "content": "[68] Xumin Yu, Lulu Tang, Yongming Rao, Tiejun Huang, Jie Zhou, and Jiwen Lu. Point-bert: Pre-training 3d point cloud transformers with masked point modeling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19313-19322, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.502, + 0.892, + 0.557 + ], + "angle": 0, + "content": "[69] Hengshuang Zhao, Li Jiang, Jiaya Jia, Philip HS Torr, and Vladlen Koltun. Point transformer. In Proceedings of the IEEE/CVF international conference on computer vision, pages 16259-16268, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.559, + 0.892, + 0.627 + ], + "angle": 0, + "content": "[70] Yihao Zhi, Xiaodong Cun, Xuelin Chen, Xi Shen, Wen Guo, Shaoli Huang, and Shenghua Gao. Livelyspeaker: Towards semantic-aware co-speech gesture generation. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 20807-20817, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.629, + 0.892, + 0.697 + ], + "angle": 0, + "content": "[71] Yang Zhou, Jimei Yang, Dingzeyu Li, Jun Saito, Deepali Aneja, and Evangelos Kalogerakis. Audio-driven neural gesture reenactment with video motion graphs. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3418-3428, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.7, + 0.892, + 0.768 + ], + "angle": 0, + "content": "[72] Lingting Zhu, Xian Liu, Xuanyu Liu, Rui Qian, Ziwei Liu, and Lequan Yu. Taming diffusion models for audio-driven co-speech gesture generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10544-10553, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.77, + 0.892, + 0.837 + ], + "angle": 0, + "content": "[73] Wentao Zhu, Xiaoxuan Ma, Zhaoyang Liu, Libin Liu, Wayne Wu, and Yizhou Wang. Motionbert: A unified perspective on learning human motion representations. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2023. 3" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.837 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "1154" + } + ] +] \ No newline at end of file diff --git a/2024/EMAGE_ Towards Unified Holistic Co-Speech Gesture Generation via Expressive Masked Audio Gesture Modeling/f8a85da7-9221-4a37-baf0-e5a3b46705e3_origin.pdf b/2024/EMAGE_ Towards Unified Holistic Co-Speech Gesture Generation via Expressive Masked Audio Gesture Modeling/f8a85da7-9221-4a37-baf0-e5a3b46705e3_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..df487e0d8e78dd6927bd69e6adaf7fc9d4ee4909 --- /dev/null +++ b/2024/EMAGE_ Towards Unified Holistic Co-Speech Gesture Generation via Expressive Masked Audio Gesture Modeling/f8a85da7-9221-4a37-baf0-e5a3b46705e3_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:63c7969f18cbc47e99af485c52313d8c199c32c0e5487a5241335591e7d46854 +size 4858502 diff --git a/2024/EMAGE_ Towards Unified Holistic Co-Speech Gesture Generation via Expressive Masked Audio Gesture Modeling/full.md b/2024/EMAGE_ Towards Unified Holistic Co-Speech Gesture Generation via Expressive Masked Audio Gesture Modeling/full.md new file mode 100644 index 0000000000000000000000000000000000000000..a4cb7b47f13fb3d18f53657cde20c4af1d869086 --- /dev/null +++ b/2024/EMAGE_ Towards Unified Holistic Co-Speech Gesture Generation via Expressive Masked Audio Gesture Modeling/full.md @@ -0,0 +1,394 @@ +# EMAGE: Towards Unified Holistic Co-Speech Gesture Generation via Expressive Masked Audio Gesture Modeling + +Haiyang Liu $^{1*}$ Zihao Zhu $^{2*}$ Giorgio Becherini $^{4}$ Yichen Peng $^{5}$ Mingyang Su $^{6}$ You Zhou $^{3}$ Xuefei Zhe $^{3}$ Naoya Iwamoto $^{3}$ Bo Zheng $^{3}$ Michael J. Black $^{3}$ + +$^{1}$ The University of Tokyo $^{2}$ Keio University $^{3}$ Independent Researcher $^{4}$ Max Planck Institute for Intelligent Systems $^{5}$ JAIST $^{6}$ Tsinghua University + +![](images/77aa6d75a5af7f91c15f89f767785f7436b00b41c1d6074cbaa3e215dafb7c19.jpg) +Figure 1. EMAGE. We present a Masked Audio-Conditioned Gesture Modeling framework, along with a new holistic gesture dataset, BEAT2 (BEAT-SMPLX-FLAME), for jointly generating facial expressions, local body dynamics, hand movements and global translations, conditioned on audio and a partially or completely masked gestures. The gray denotes visible gestures, and blue represents our outputs. + +# Abstract + +We propose EMAGE, a framework to generate full-body human gestures from audio and masked gestures, encompassing facial, local body, hands, and global movements. To achieve this, we first introduce BEAT2 (BEATSMLPX-FLAME), a new mesh-level holistic co-speech dataset. BEAT2 combines a MoShed SMPL-X body with FLAME head parameters and further refines the modeling of head, neck, and finger movements, offering a community-standardized, high-quality 3D motion captured dataset. EMAGE leverages masked body gesture priors during training to boost inference performance. It involves a Masked Audio Gesture Transformer, facilitating joint training on audio-to-gesture generation and masked gesture reconstruction to effectively encode audio and body gesture hints. Encoded body hints from masked gestures are then separately employed to generate facial and body movements. Moreover, EMAGE adaptively merges speech features from the audio's rhythm and content and utilizes four compositional VQ-VAEs to enhance the results' fidelity and diversity. Experiments demonstrate that EMAGE generates + +holistic gestures with state-of-the-art performance and is flexible in accepting predefined spatial-temporal gesture inputs, generating complete, audio-synchronized results. Our code and dataset are available.1 + +# 1. Introduction + +The path towards full-body co-speech gesture generation presents many remaining challenges. Despite the development of various baselines, e.g., audio conditioned models for face [15, 20, 61], body [13, 25, 67], and hand movements [33, 39, 41], along with a few attempts at merged models [28, 65], the limited availability of comprehensive data and models poses an ongoing obstacle. To make progress in this direction, we present a new framework that can incorporate partial spatial-temporal predefined gestures and autonomously complete the remaining frames in synchronization with the audio. This provides a new tool for creating realistic animations of digital humans + +[57, 63, 64, 71]. +To achieve this, we first require a comprehensive gesture dataset. While several datasets are available for audio-to-body [67], audio-to-body&face [28], body-to-hands [39, 41], and common actions [16, 46], integrating them poses challenges due to varying data formats and evaluation metrics across different sub-areas. This motivates us to establish a unified, community-standard benchmark for training and evaluating holistic gestures. We build upon the BEAT dataset [39], which provides the most comprehensive mocap co-speech gesture dataset with diverse modalities. Despite its size and richness, the data in BEAT is not presented in a standardized format. The dataset comprises iPhone ARKit blendshape weights and Vicon skeletons, posing challenges when using the data for training; e.g., the lack of a mesh representation prevents the use of a vertex loss [19, 52]. Additionally, conducting direct human perceptual evaluations on these diverse modalities is inherently complex. On the other hand, SMPL-X [50] and FLAME [35] are common mesh standards for the academic community for other human dynamics-related tasks, e.g., action [52] and talking head generation [14]. Driven by the motivation to facilitate the sharing of knowledge across different tasks, we present BEAT-SMPLX-FLAME (BEAT2), which consists of two main components: i) Refined SMPL-X body shape and pose parameters from MoSh++ [44, 46] with hard-coded physical priors. ii) Optimized high-quality FLAME head parameters. By integrating SMPL-X's body and FLAME's head, BEAT2 facilitates a comprehensive training and evaluation benchmark across multiple sub-areas of co-speech human animation generation. +With the holistic data from BEAT2, we aim to enhance the coherence of individual body parts while ensuring accurate cross-modal alignment between audio and motion; see Figure 1. This leads to the design of Expressive Masked Audio-conditioned GEsture modeling, denoted as EMAGE, a spatial and temporal transformer-based framework. EMAGE first aggregates spatial features from masked body joints. Then, it reconstructs the latent space of pretrained gestures with a switchable gesture temporal self-attention and audio-gesture cross-attention. The selection of different forward paths enables modeling the gesture-to-gesture and audio-to-gesture prior separately and effectively. Once the reconstructed latent features are obtained, EMAGE decodes local face and body gestures from four compositional pretrained Vector Quantized-Variational AutoEncoders (VQ-VAEs) and decodes global translations from a pretrained Global Motion Predictor. The compositional VQ-VAEs are key to preserving audio-related motions during the reconstruction. Through these designs, EMAGE achieves state-of-the-art performance in generating body and face gestures. It takes as input audio and partially initialized gestures, and recovers audio + +synchronized, coherent gestures. Additionally, we show how EMAGE can flexibly incorporate additional, nonholistic, datasets to improve results. +Overall, our contributions are as follows: (1) We release BEAT2, a representation-unified, mesh-level dataset that leverages MoShed SMPL-X body and FLAME head parameters. (2) We propose EMAGE, a simple yet effective holistic gesture generation framework that generates coherent gestures with partial gesture and audio priors. (3) EMAGE achieves state-of-the-art performance in generating both body and face gestures with only four frames of seed gestures. (4) EMAGE showcases the use of additional non-holistic gesture datasets for training, e.g., Trinity and AMASS, to further improve the fidelity and diversity of the results, effectively leveraging data from different datasets. + +# 2. Related Work + +Co-Speech Animation Datasets are categorized into two types; see Table 1: pseudo-labeled (PGT) and motion-captured (mocap). For PGT, the Speech2Gesture Dataset [25] utilizes OpenPose [11] to extract 2D poses from News and Teaching videos. Subsequent works extend the dataset to 3D pose [28] and SMPL-X [65]. The TED Dataset [66], which extracts 2D pose from TED videos, has been extended to include 3D pose [67], fingers [40], and SMPL-X [45]. Similarly, 2D, 3D landmarks, and meshes are estimated from multi-view recorded videos [60] as PGT for talking-head generation. Although pseudo-labeled approaches could theoretically extract infinite data, their accuracy is limited. For instance, the average error (in mm) for the SOTA monocular 3D pose estimation algorithm [23] on the Human3.6M dataset [30] is 33.4, while for Vicon mocap it is 0.142 [47]. For mocap datasets, Trinity [3] features one male actor with 4h of data, and TalkingWithHands [32] collects conversation scenarios for two speakers. ZEEG [24] considers one speaker with 12 styles. For the face, 3D scanning is accurate but limited in quantity due to cost, e.g., less than 3h for BIWI [21], VOCA [14], and MeshTalk [56]. This leads to a balance between performance and quantity in the ARKit dataset [51]. The aforementioned datasets address the body and face separately. On the other hand, BEAT [39] includes both 3D pose body and ARKit [8] facial blendshape weights for the first time. However, BEAT lacks mesh data. +Co-Speech Animation Models are categorized based on their outputs; see Table 2. In addition to the baselines and datasets [14, 25, 32, 48, 56, 67], several frameworks for body gestures [2, 33, 53, 54] have improved the performance of the baselines. They are trained and evaluated by selecting specific upper body joints [4, 5, 38-40, 49, 70], or all body joints [62, 72]. Recent improvements in facial gestures drive the vertices with transformers and discrete face priors [19, 61]. However, these methods only address ei + +
Trinity Mocap 2017 [22]S2G S2G-2D 2019 [25]Seq2Seq TED-2D 2019 [66]TWH Mocap 2019 [32]Trimodal TED-3D 2020 [67]VOCA Scan 2020 [14]MeshTalk Scan 2021 [56]Hahibie et al. S2G-3D 2021 [28]HA2G TED-3D+ 2022 [40]BEAT Mocap 2022 [39]ZEEG Mocap 2022 [24]Yoon. et al. TED-SMPL 2023 [45]Talkshow S2G-SMPL 2023 [65]BEAT2 Ours 2024
Head-2D---3D Scan3D Scan3D-ARKit--PGT-MeshMC-Mesh
Upper Body3D2D2D3D3D--3D3D3D3DPGT-MeshPGT-MeshMC-Mesh
Hands3D2D-3D---3D3D3D3DPGT-MeshPGT-MeshMC-Mesh
Lower Body3D--3D-----3D3D--MC-Mesh
Global Motion3D--3D-----3D3D--MC-Mesh
Duration (hours)4609720970.523833764302760
+ +Table 1. Comparison of Co-Speech Gesture Datasets. We summarize gesture and face datasets in talking scenarios. 'PGT' and 'MC' denote Pseudo Ground Truth and Motion Capture, respectively. The BEAT2 (Ours) dataset is the largest dataset for motion-captured data, providing holistic, academic community standard mesh-level information. + +
S2GTriModalB2HTWHHahibie et al.DisCoHA2GFace Former 2022[19]RhythmicBEAT CaMN 2022[39]Diff Gesture 2023[72]Code Talker 2023[61]Talk Show 2023[65]DiffStyle Gesture 2023[62]Body Former 2023[49]Lively Speaker 2023[70]EMAEG (Ours) 2024
2019 [25]2020 [67]2020[48]2021[32]2021[28]2022[38]2022[40]2022[5]2022[5]-------Masked
Input Gestures--BodyBody------------
Head---M--S---SS---C
UpperSS-MSC-SCS-CSSSC
Hands--SSMSC--CS-CS-SC
Lower + Global------------S--C
+ +Table 2. Comparison of Co-Speech Gesture Models. We compare with previous methods for generating face or body motion trained on co-speech datasets. The first row lists their inputs, and the subsequent rows list their outputs, respectively. Different decoder designs are denoted by the initials 'S' for Single, 'M' for Multiple, and 'C' for Cascaded. To the best of our knowledge, EMAGE (Ours) is the first to accept audio and partially or completely masked gestures, generating full-body audio-synchronized results. + +ther the face or body. Applying these methods to the full body will yield sub-optimal results as audio is differently correlated with face and body dynamics. Most similar to our work, Habibie et al. [28] employ a single audio encoder and multiple decoders for generating face and body gestures. TalkSHOW [65] demonstrates the advantage of separating the audio encoders and decoding the body and hands using quantized codes in an auto-regressive manner. However, it lacks lower body and global motion, and there is no shared information for body and facial gesture generation. Moreover, it cannot accept partially masked body hints due to the design of a fully auto-regressive model. + +Masked Representation Learning was first demonstrated to be effective in Natural Language Processing, with BERT-based models [17, 31, 42] boosting the performance of learned word embeddings in downstream tasks through a combination of masked language modeling and transformer architecture. Subsequently, Masked AutoEncoders [29] expanded masked image modeling to computer vision by removing and inpainting image patches. This concept of masked representation learning has since been employed in other modalities, e.g., video [12, 18, 43], audio [26, 36, 37], and point clouds [55, 68, 69]. Most related to our work, MotionBERT [73] proposes a spatial-temporal transformer to learn a robust motion representation for classification-based tasks by masking 2D pose. In contrast to their method, we target robust motion features for conditional motion generation tasks, requiring a balance in training between multiple modalities, e.g., audio and gesture. + +# 3. BEAT2 + +This section introduces how we obtain unified mesh-level data, i.e., SMPL-X [50] and FLAME [35] parameters, from + +the original BEAT dataset [39]. BEAT utilizes a Vicon motion capture system [47] and releases 78-joint skeleton-level Biovision Hierarchy (BVH) [10] motion files. Their facial capture system uses ARKit [8] with a depth camera on the iPhone 12 Pro, extracting $\mathbb{R}^{51}$ blendshape weights. These blendshapes are designed based on the Facial Action Coding System (FACS), which is widely adopted. However, both the motion and facial data lack mesh-level details (see Figure 2), e.g., shapes and vertices. + +# 3.1. Body Parameters Initialization via MoSh++ + +We initialize the SMPL-X body shape and pose parameters from BEAT mocap marker data using MoSh++ [46]. Given the captured markers positions $\mathbf{m} \in \mathbb{R}^{T \times K \times 3}$ , predefined markers position offsets $\mathbf{d} \in \mathbb{R}^{K \times 3}$ , and user-defined vertices-to-markers function $\mathcal{H}$ , we aim to obtain body shape $\beta \in \mathbb{R}^{300}$ , poses $\theta \in \mathbb{R}^{T \times 55 \times 3}$ , and translation parameters $\gamma \in \mathbb{R}^{T \times 3}$ . The optimization uses a differentiable surface vertex mapping function $S(\beta, \theta, \gamma)$ and vertex normal function $\mathcal{N}(\beta, \theta_0, \gamma_0)$ . For each frame, the latent marker $\tilde{\mathbf{m}} \in \mathbb{R}^{T \times K \times 3}$ is calculated as + +$$ +\tilde {\mathbf {m}} _ {i} \equiv \mathcal {S} _ {\mathcal {H}} (\beta , \theta_ {i}, \gamma_ {i}) + \mathbf {d N} _ {\mathcal {H}} (\beta , \theta_ {i}, \gamma_ {i}). \tag {1} +$$ + +We first select 12 frames to optimize and fix $\mathbf{d}$ and $\beta$ , then optimize $\theta_{i}$ , $\gamma_{i}$ for $i \in (1:T)$ by minimizing the loss terms based on $\|\tilde{\mathbf{m}}_i - \mathbf{m}_i\|^2$ , including Data Term, Surface Distance Energy, Marker Initialization Regularization, Pose and Shape Priors, Velocity Constancy, and Soft-Tissue Term (see supplementary materials). The overall objective function is the weighted sum of these terms to balance accuracy and plausibility. + +![](images/1a29c1ea8463bf00c2c3f184a817a4f1c708628a8ef1d47d76d7f19f87bcfc5d.jpg) + +![](images/0d3097e295ab5f39d373170046e3ede29bd44caca6a90f7a1e16e8cbbc74f8e2.jpg) +BEAT + +![](images/3cf8e0bff276205e84046de305c5ed3d778dd92008b38ecc417fbe3aa563ac59.jpg) + +![](images/9c115889ca59e8c11153c9150f32df01371750dbb710af599fab43cfc7f8e73d.jpg) +Retargeted + +![](images/0b2f09f2374956161351194f45e2e2532c9d05df8cf0f908d5ce265ba34ba94c.jpg) + +![](images/f8f4640aefd7e52843ec5ee7f210087f3efb995a8208587356d71d31ebcdc3c4.jpg) +Moshed +Figure 2. Comparison of Data between BEAT2 and Others. BEAT-SMPLX-FLAME presents a new mesh-level, motion-captured, holistic co-speech gesture dataset with 60h of data. Left: We compare our refined SMPL-X body parameters (denoted as Refined MoSh) with the original BEAT skeleton [39], Retargeted from AutoRegPro, and initial results of $\mathrm{Mosh}++$ [50]. The refined results show correct neck flexion, appropriate head and neck shape ratios, and detailed finger representations. Right: Visualization of blendshape weights from the original BEAT dataset [39] with ARKit's template, Wrapped-based, and handcrafted optimization. Our final handcrafted FLAME blendshape-based optimization demonstrates both accurate lip movement details and natural mouth shapes. + +![](images/3c1b9aa5d1182a6c0f9d831ead367f4923e1919823cf79b76dca4f59fb4552c0.jpg) + +![](images/998179254e8dbd0a0cd08de2c91b898bf38655be992f223fb73fe5128e9c24bd.jpg) +Refined Moshed (Ours) + +![](images/2e2280169812212b1d236e2a6a3e66b5ab26bd14574a2bb1d9abdb43d01b8b65.jpg) + +![](images/0823b4da28a4d4e58bba88dec90cc1bb5f4de875f981e5c1269d82ed5df385a6.jpg) +BEAT + +![](images/0a5a6dee70f74b4a51ebca6acff78622320dfb2cf09a947e6ba026a25c523992.jpg) + +![](images/4e2e0a6d3965907893b8e602a4bd4865e5ecea413958654dd3f083be29da6d78.jpg) +Wrapped Optimum + +![](images/aa583dca0754a84b1df2c39c7b1147b4db1387c0d69b19fcce8f8d91e9e9466b.jpg) + +![](images/3da42aa9860e70ce3b694a6054783d9f404c5a344355f108f4502a33f7839eda.jpg) +Handcraft Optimum (Ours) + +# 3.2. Body Parameters Refinement + +MoSh++ produces unnatural head shapes due to the fact that head markers were worn on a helmet and it sometimes produces unnatural finger poses. Consequently, we refine body shape and pose parameters with three simple yet effective hard-coded physical rules. i) The neck and head length should approximate 1/7 of the total length of the body. ii) Fingers, except for the thumb, should not bend backward. iii) We employ the Kolmogorov-Smirnov test, which reveals the data is similar to a normal distribution. Subsequently, we apply a Gaussian truncation approach, where all data points falling outside the $3\sigma$ range are adjusted to align with the $3\sigma$ threshold and blended with the adjacent 10 frames. We compare our refined body parameters (denoted as MoSh Refined) with Original BEAT, Retargeted SMPL-X, and MoSh SMPL-X in Figure 2. + +# 3.3. Blendshape Weights to FLAME Parameters + +Given the ARKit blendshape weights $\mathbf{b}_{\mathrm{ARKit}} \in \mathbb{R}^{T \times 51}$ , we aim to derive a transformation matrix $\mathbf{W} \in \mathbb{R}^{51 \times 103}$ that maps these into FLAME parameters $\mathbf{b}_{\mathrm{FLAME}} \in \mathbb{R}^{T \times (100 + 3)}$ , where 100 represents the number of dimensions for expression parameters and 3 for jaw movement. Due to discrepancies between the iPhone's and FLAME's template mesh topology, simply optimizing FLAME parameters by minimizing the geometric error between the ARKit template vertices and the wrapped FLAME vertices does not yield satisfactory results. Instead, we release a set of handcrafted blendshape templates $\mathbf{v}_{\mathrm{template}} \in \mathbb{R}^{52}$ on FLAME following ARKit's FACS configuration. This approach allows direct driving of the FLAME topology ver + +tices $\mathbf{v}\in \mathbb{R}^{T\times 5023\times 3}$ from the given blendshape weights, + +$$ +\mathbf {v} = \mathbf {v} _ {\text {t e m p l a t e}} ^ {0} + \sum_ {j = 1} ^ {5 1} \mathbf {b} _ {\text {A R K i t}, j} \cdot \mathbf {v} _ {\text {t e m p l a t e}} ^ {j} \tag {2} +$$ + +where $\mathbf{b}_{\mathrm{ARKit},j}$ is the weight of the $j$ -th ARKit blend-shape and $\mathbf{v}_{\mathrm{template}}^j$ is the vertex position of the FLAME model template corresponding to the $j$ -th blendshape. The term $\mathbf{v}_{\mathrm{template}}^0$ represents the original template vertex position of the FLAME model. We optimize $\mathbf{W}$ by minimizing $\| \tilde{\mathbf{v}}_j - \mathbf{v}_j\| _2$ , where $\tilde{\mathbf{v}}$ is obtained from FLAME's LBS $\mathcal{V}(\mathbf{b}_{\mathrm{FLAME}})$ . The comparisons between ARKit data, wrap-based, and our approach are shown in Figure 2. + +# 4. IMAGE + +We introduce the details of EMAGE, Expressive Masked Audio-Conditioned GEsture Modeling (see Figure 3). Given gestures $\mathbf{g} \in \mathbb{R}^{T \times (55 \times 6 + 100 + 4 + 3)}$ , representing 55 joints in Rot6D, $\mathbb{R}^{100}$ FLAME parameters, $\mathbb{R}^4$ foot contact labels, $\mathbb{R}^3$ global translations, and speech audio $\mathbf{a} \in \mathbb{R}^{T \times sk}$ , where $sk = sr_{\mathrm{audio}} / fps_{\mathrm{gestures}}$ , EMAGE jointly optimizes masked gesture reconstruction and audio-conditioned gesture generation. This optimization enhances performance during inference and enables the use of partially masked gestures to complete holistic gestures. To achieve this, we first model the quantized latent space in Sec. 4.1, following [61, 65]. We then design a separated speech audio encoder using Content Rhythm Self-Attention (CRA) as detailed in Sec. 4.2. Subsequently, we learn the body hints from masked gestures via the Masked Audio Gesture Transformer, explained in Sec. 4.3. Finally, the decoding strategy for different body segments is discussed in Sec. 4.4. + +![](images/03d3d4b036b0ca4357b992552ab291dfb862bda77305f32207eae0b1502e2b37.jpg) +Figure 3. EMAGE leverages two training paths: Masked Gesture Reconstruction (MG2G) and Audio-Conditioned Gesture Generation (A2G). The MG2G path focuses on encoding robust body hints through a spatial-temporal transformer gesture encoder and cross-attention gesture decoder. In contrast, the A2G path utilizes these body hints, and separated audio encoders to decode pretrained face and body latent features. A key component in this process is a switchable cross-attention layer, crucial for merging body hints and audio features. This fusion allows the features to be effectively disentangled and utilized for gesture decoding. Once the gesture latent features are reconstructed, EMAGE employs a pretrained VQ-Decoder to decode face and local body motions. Additionally, a pretrained Global Motion Predictor is used to estimate global body translations, further enhancing the model's capability to generate realistic and coherent gestures. + +![](images/fcfbe7a0829c0edac6c9bb4d65277fe8bae5b40a7eca41f562d382b9fffccf6f.jpg) +Figure 4. Details of CRA and Pretrained VQ-VAEs. Left: Content Rhythm Attention fuses speech rhythm (onset and amplitude) with content (pretrained word embeddings from text scripts) adaptively. This results in a preference for either content or rhythm in specific frames, which encourages the generation of semantical-aware gestures. Right: We pretrain four compositional VQ-VAEs by reconstructing face, upper body, hands and lower body separately to disentangle audio-agnostic gestures explicitly. + +![](images/20c78e60af2f8a1e4a54cfc9be2acb3cffddc5a758e0ff797f181e0b8b1facf5.jpg) + +# 4.1. Compositional Discrete Face and Body Prior + +We model full-body gestures in the separated quantized latent space (Figure 4) for several reasons. Similar to [65], the body and hands improves the diversity of the results. However, we additionally separate the face and lower body due to their differing correlations with audio, i.e., using a single VQ-VAE to encode both the upper and lower body may lead the model to overlook gestures that occur less frequently - regardless of their relation to the audio. Specifically, a single model might focus only on recovering lower body motion, neglecting elbow movements in a speaker who is constantly walking during the conversation. + +The separated quantized latent space $Q = \{\mathbf{q}_{\mathrm{f}}, \mathbf{q}_{u}, \mathbf{q}_{\mathrm{h}}, \mathbf{q}_{\mathrm{l}}\}$ for the Face $\mathbf{g}_{\mathrm{f}} \in \mathbb{R}^{T \times 106}$ , Upper + +body $\mathbf{g}_{\mathrm{u}} \in \mathbb{R}^{T \times 78}$ , Hands $\mathbf{g}_{\mathrm{h}} \in \mathbb{R}^{T \times 180}$ , and Lower body $\mathbf{g}_{\mathrm{l}} \in \mathbb{R}^{T \times (54 + 4)}$ are from four VQ-VAEs. The VQ-VAE for each is optimized by jointly optimizing the following loss terms, + +$$ +q _ {i} = \underset {q _ {i} \in \mathbf {q}} {\arg \min } \| z _ {j} - q _ {i} \| ^ {2} \tag {3} +$$ + +$$ +\begin{array}{l} \mathcal {L} _ {\mathrm {V Q - V A E}} = \mathcal {L} _ {r e c} (\mathbf {g}, \hat {\mathbf {g}}) + \mathcal {L} _ {\mathrm {v e l}} \left(\mathbf {g} ^ {\prime}, \hat {\mathbf {g}} ^ {\prime}\right) + \mathcal {L} _ {\mathrm {a c c}} \left(\mathbf {g} ^ {\prime \prime}, \hat {\mathbf {g}} ^ {\prime \prime}\right) \\ + \left\| \operatorname {s g} [ \mathbf {z} ] - \mathbf {q} \right\| ^ {2} + \left\| \mathbf {z} - \operatorname {s g} [ \mathbf {q} ] \right\| ^ {2}, \tag {4} \\ \end{array} +$$ + +where $\mathbf{z}$ is the encoded $\mathbf{g}$ with a temporal window size $w = 1$ . $\mathcal{L}_{\mathrm{rec}}$ , $\mathcal{L}_{\mathrm{vel}}$ , and $\mathcal{L}_{\mathrm{acc}}$ are Geodesic [58] and L1 losses, sg is a stop gradient operation, we set the weight of the commitment loss [59] (the last term) to 1 in this paper. + +# 4.2. Content and Rhythm Self-Attention + +Given the speech audio, s, inspired by [4], we employ onset o and amplitude a as explicit audio rhythm, alongside the pretrained embeddings [9] e from transcripts as content. Different from previous approaches, which typically add the rhythm and content features, we leverage self-attention to merge these features adaptively. This approach is driven by the observation that for specific frames, the gestures are more related to content (semantic-aware) or rhythm (beat-aware). The rhythm and content features are first encoded into time-aligned features $\mathbf{r}_{1:T}$ and $\mathbf{c}_{1:T}$ , using a Temporal Convolutional Network (TCN) and linear mapping, respectively. For each time step $t\in \{1,\dots,T\}$ , we merge the rhythm and content features by: + +$$ +\begin{array}{l} \mathbf {f} _ {1: T} = \alpha \times \mathbf {r} _ {1: T} + (1 - \alpha) \times \mathbf {c} _ {1: T}, \tag {5} \\ \alpha = \operatorname {S o f t m a x} (\mathcal {A T} (\mathbf {r} _ {1: T}, \mathbf {c} _ {1: T})), \\ \end{array} +$$ + +![](images/b5ffedfff937a78fa40fca7cb43e4752e73f12773787d02e95ac5a687967d09f.jpg) +Figure 5. Comparison of Forward Path Designs. Straightforward fusion module (a) merges audio features without refined body features and recombines audio features based only on position embedding. The Self-Attention decoder module (b), adopted in previous MLM models [17, 31], is limited for tasks requiring auto-regressive inference. Our design (c) considers effective audio feature fusion and auto-regressive decoding. + +![](images/1cacf29fb5e6624bf438a25165fccad902e7718042aa87ac7732c73e36efd625.jpg) + +![](images/32cfc4eb4ef2edd24fa8beeb6c6b37217aecf0976a8d36ece3aaf17f2d0afc5b.jpg) + +where $\mathcal{A}\mathcal{T}$ is a 2-layer MLP. We apply two separate CRA encoders for the face and body. + +# 4.3. Masked Audio-Conditioned Gesture Modeling + +We propose a Masked Audio Gesture Transformer to leverage different training paths (see Fig. 5 for the motivation behind the architecture designs). Given temporal and spatially masked gestures $\overline{\mathbf{g}}\in \mathbb{R}^{T\times 337}$ , we first replace the masked tokens with learned masked embeddings $e_{\mathrm{mask}}\in \mathbb{R}^{256}$ , as the value zero still represents specific motion content, e.g., a T-pose. We linearly increase the ratio of masked joints and frames from 0 to $95\%$ according to the training epochs. + +A spatial convolutional $\mathcal{SC}$ encoder initially summarizes the spatial information, and compress the spatial feature into $\mathbb{R}^{T\times 256}$ . We then employ a temporal self-attention (without feed-forward) $\mathcal{TSA}$ to refine the summarized spatial features + +$$ +\mathbf {h} = \mathcal {T S A} (\mathcal {S C} (\overline {{\mathbf {g}}}) + \mathbf {p} _ {t}), \tag {6} +$$ + +where $\mathbf{h} \in \mathbb{R}^{T \times 512}$ represents encoded body hints, and $\mathbf{p}_t$ is the sum of a learned speaker embedding and PPE [19]. Then, a straightforward temporal cross-attention Transformer decoder $\mathcal{TCA}\mathcal{T}$ is adopted for the reconstructed latent $\hat{\mathbf{q}}_{\mathrm{mg2g}}$ , + +$$ +\hat {\mathbf {q}} _ {\mathrm {m g} 2 \mathrm {g}} = \mathcal {T C A T} (\mathbf {h} + \mathbf {p} _ {t}). \tag {7} +$$ + +We minimize the L1 distance in the latent space, + +$$ +\mathcal {L} _ {\mathrm {m g} 2 \mathrm {g}} = \| \hat {\mathbf {q}} _ {\mathrm {m g} 2 \mathrm {g}} - \mathbf {q} \|. \tag {8} +$$ + +Masked gesture reconstruction encodes effective body hints, with the key being to leverage these body hints for gesture generation. We employ a selective fusion for audio and body hints via a temporal cross-attention $\mathcal{TCA}$ . Subsequently, we use the merged audio-gesture features for audio-conditioned gesture latent reconstruction, + +$$ +\hat {\mathbf {q}} _ {\mathrm {a} 2 \mathrm {g}} = \mathcal {T C A T} \left(\mathcal {T C A} \left(\mathbf {h} + \mathbf {p} _ {t}, \mathbf {f} _ {\text {b o d y}}\right), \overline {{\mathbf {g}}} + \mathbf {p} _ {t}\right). \tag {9} +$$ + +
Top.-BTop.-FBodyFace
VOCA[14]-FLAME-38.3 ± 5.63
AMASS[46]SMPL-XFLAME42.0 ± 3.60-
TalkSHOW[65]SMPL-XSMPL-X14.4 ± 2.1926.1 ± 6.42
BEAT2 (Ours)SMPL-XFLAME43.6 ± 3.3835.7 ± 5.91
+ +Table 3. User Preference Win Rate between Existing Datasets. 'Top' denotes the topology of the mesh. The results show our BEAT2 dataset outperforms the existing PGT dataset [65] in both body and face aspects. It also performs slightly better on body and lower on face when compared with previous mocap [46] and 3D scan-based datasets [14], respectively. + +We optimize both the latent code class classification cross-entropy loss $\mathcal{L}_{\mathrm{a2g - rec}}$ and the latent reconstruction loss $\mathcal{L}_{\mathrm{a2g - c1s}}$ to encourage diverse results. + +# 4.4. Face and Translation Decoding + +Considering that the face is weakly related to body motion, applying the same operation - recombining the audio features based on body hints - is not reasonable. Therefore, we directly concatenate the masked body hints with the audio features for the final decoding of facial latent, + +$$ +\hat {\mathbf {q}} _ {\mathrm {f}} = \mathcal {T} \mathcal {C A T} \left(\mathbf {f} _ {\text {f a c e}} \oplus \mathbf {h}, \mathbf {p} _ {t}\right). \tag {10} +$$ + +Once we have obtained the local lower body motion $\tilde{\mathbf{g}}_l\in$ $\mathbb{R}^{T\times (54 + 4)}$ from the VQ-Decoder, we estimate the global translations $\tilde{\mathbf{t}}\in \mathbb{R}^{T\times 3}$ with a pretrained Global Motion Predictor $\tilde{\mathbf{t}} = \mathcal{G}(\tilde{\mathbf{g}}_l)$ , which significantly reduces foot sliding. + +# 5. Experiments + +We separate the evaluation into two categories: dataset quality and model ability. After removing data with low finger quality from BEAT, BEAT2 is reduced to 60 hours. We further split it into BEAT2-Standard (27 hours) and BEAT2-Additional (33 hours), based on the type of speech and conversation sections [39]. While acted gestures in speech sections are more diverse and expressive, spontaneous gestures in conversational sections tend to be more natural yet less varied. We report the results on BEAT2-Standard Speaker2 with an $85\% / 7.5\% / 7.5\%$ train/val/test split. + +# 5.1. Dataset Quality Evaluation + +We compare our dataset with the state-of-the-art pseudo ground truth (PGT) dataset, TalkSHOW [65], for both face and body. Additionally, we compare it with the AMASS [46] dataset for the body and VOCA [14] for the face as references. Due to the varying lengths of sequences in each dataset, we sample 100 comparison pairs, each with an equal duration ranging from 2 to 4 seconds. In a perceptual study, each participant evaluates a random set of 40 pairs during a 10-minute session by selecting the sequence they consider to have the best captured quality. In total, 60 participants were invited. It is important to note that participants are instructed to compare only the upper body results, + +... In my spare time ... ... if I feel ok ... + +... I will go for a walk or hike in nature. I try to organize something + +![](images/2c125761dd9cae7c4576e912ae7f448a11376383f3058ef27dd260925960ebc0.jpg) + +![](images/cdb9bcf28bebd72500a014eecdaa75e44c0ee0d34d340bf78adbc004c8817654.jpg) + +![](images/bae710f5cf209e06c839b96c23eb6e0f4e8f3ae177079db6f6be7e4871ae7a24.jpg) + +![](images/c1c57afefb94e35ad11b6de08d396967a3fa5faca4bb4d28920f253113b08c23.jpg) + +![](images/428f0f0053095cf331f75a2ee392f3922abd1e4ccefa68d335d26cba72ee5c79.jpg) + +![](images/0e571774cfd0d6a43f3e034ffbc64ca28c24d7885bf47b0ac76ba27000ba1cb3.jpg) + +![](images/6ac6fc9288d18bedb3545a7e62f6007ef4f7c006b15af8b108bf094d1ad7880d.jpg) + +![](images/5f6862fca91f66b90bd24f08769f95c06dfa042121411312abe32d886c6e6098.jpg) + +![](images/331dc7ed8f3b856ee1e11e203e33219dc639d6472ce45416330386342a33996a.jpg) + +![](images/fd3a413577dc8a0433bca6a0d076bf3a615e42cf9e70297b2b0c83b45ad34d34.jpg) + +![](images/d5f6d5f60c872172245a54c48b4f67219ea87e02dbdecbd6b9a8a9f105928657.jpg) + +![](images/50f47b1881a4f057c10224c1452d9152a3706b0b2443921ac8022fe89d6e58c5.jpg) + +![](images/783ca256b44282fda31bf3021b3fb18df2884846cbacbae94c1c54303861afca.jpg) + +![](images/7e6a2efeb4f1b1f039e4b4103e35d22570d17ac002aa7b76b5737e7ad19b9a58.jpg) + +![](images/3b5bb70b6c5c3fbf791f74724eb0bfd3ce4c78419828d7f4465313421d034016.jpg) + +![](images/6870b2a16ab908948f4c9a29bfe7b7a29afcad6abcba862bc36131580f62c209.jpg) +Figure 6. Subjective Examples. Top to down: GroundTruth, Generated W/O Body Hints, Generated With Body Hints, Visible Body Hints, EMAGE generates diverse, semantic-aware and audio synchronized body gestures, e.g. raise the both hands for "spare time", relaxing gestures for "hike in nature". Besides, as shown in the third and bottom rows, EMAGE is flexible to accept non-audio-synchronized body hints on any frames and joints, to guide or customize the generated gestures explicitly, e.g., repeat a similar motion for raising hands, change the walking orientations, etc. Note that in this figure, The gray joints of generated results are not copies of visible hints. + +![](images/6d0dd81e175b8e70e7d2cbaa3207c1c29eec10056fee93189c6f6b8833554f74.jpg) + +![](images/fd6f3eee973bb1ecc58436590341bf1e98b4a7419d4e84fc2f7ed5c9472f580e.jpg) + +![](images/fe89c02a59451a9e1f4d73c03f7d1acf521b5838104d33ec8fb5dc3acbf675a8.jpg) + +![](images/87fdaad99801ac186d6b6a5c6f959383750e13087ad63f5e0cb766c482d4920a.jpg) + +
FGD ↓BC ↑Diversity ↑MSE ↓LVD ↓
FaceFormer[19]---7.7877.593
CodeTalker[61]---8.0267.766
S2G[25]28.154.6835.971--
Trimodal[67]12.415.9337.724--
HA2G[40]12.326.7798.626--
DisCo[38]9.4176.4399.912--
CaMN[39]6.6446.76910.86--
DiffStyleGesture[62]8.8117.24111.49--
Habibie et al.[28]9.0407.7168.2138.6148.043
TalkSHOW[65]6.2096.94713.477.7917.771
EMAGE (Ours)5.5127.72413.067.6807.556
+ +as TalkSHOW contains only the upper body. The results are shown in Table 3. + +# 5.2. Model Ability Evaluation + +We report results using BEATv1.3 and BEATv2. The latter has facial data refined by animators and hand data filtered by annotators. + +Metrics. We adopt FGD[67] to evaluate the realism of the body gestures. Then, we measure Diversity[33] by calcu + +Table 4. Quantitative evaluation on BEATv2. We report FGD $\times 10^{-1}$ , BC $\times 10^{-1}$ , Diversity, MSE $\times 10^{-8}$ , and LVD $\times 10^{-5}$ . For body gestures, EMAGE significantly improves the FGD, indicating that the generated results are closer to the GT. This shows the effectiveness of body hints from masked gesture modeling. + +
HolisticBodyFace
Habibie et al.[28]12.4 ± 3.7015.9 ± 6.4910.8 ± 3.19
TalkSHOW[65]34.9 ± 5.7940.4 ± 8.2233.2 ± 6.03
EMAGE (Ours)52.7 ± 7.9144.7 ± 8.6856.0 ± 7.80
+ +Table 5. User Preference Win Rate On Generated Results. The results indicate that our generated outcomes are perceived as more realistic and believable, with a $14\%$ and $23\%$ higher user preference for body and face gestures, respectively. + +lating the average L1 distance between multiple body gesture clips, and use BC [34] to assess speech-motion synchrony. For the face, we calculate the vertex MSE[61] to measure the positional distance and the vertex L1 difference LVD [65] between the GT and generated facial vertices. + +Compared Methods. We first compare our method with representative state-of-the-art methods in body gesture generation [25, 38-40, 62, 67] and talking head generation [19, 61] by reproducing their methods for body and face, respectively. In addition, we reproduce two previous state-of-the-art holistic pipelines, Habibie et al. [28] and TalkSHOW [65], whose original implementations are limited to the upper body. We add a lower-body decoder to Habibie et al. and a lower-body VQ-VAE to TalkSHOW. + +![](images/ec4acad0cc10bb45d1f30e78a3cac08c72cd08e88bef35a1db35c54e8a1d34f8.jpg) +Figure 7. Comparison of Generated Facial Movements. Compared with previous state-of-the-art talking face generation methods FaceFormer [19] and CodeTalker [61] as well as holistic gesture generation methods Habibie et al. [28] and TalkSHOW [65]. Note that CodeTalker has higher MSE than IMAGE on BEATv2 (Table 4, lower is better) but is subjectively realistic. IMAGE gets good lip motions by leveraging both the face model and masked body hints. + +# 5.2.1 Quantitative and Qualitative Analysis + +As shown in Table 4, with a four-frame seed pose, our method outperforms previous SOTA algorithms. For qualitative results, see Figures 6 and 7. Furthermore, we conduct a perceptual study. Maintaining the same setup of 60 participants, each participant evaluates 40 pairs of 10-second results do decide which is most believable; this gives the win rate shown in Table 5. + +# 5.2.2 Ablation Analysis + +Performance of Baseline. As shown in Table 6, we start with a teacher-force transformer-based baseline. This baseline, inspired by FaceFormer [19], adopts a 1-layer cross-attention transformer decoder and replaces the audio features from Wav2Vec2 [6] with our custom TCN [7] and trainable word embeddings [9]. + +Effect of Multiple VQ-VAEs. Simply applying one VQ-VAE [27, 59] for full-body movements, including the face, decreases performance in facial movements. This is because the VQ-VAE is trained to minimize the average loss for the full body, and some speakers' most frequent movements are not related to the audio. Implementing separated VQ-VAEs allows the model to better leverage the advantages of discrete priors. + +Effect of Content Rhythm Self-Attention. Adaptive fusion of rhythm and content features shows improvement in both FGD and Alignment. It selectively focuses the current motion more on rhythm or content features based on the training data's distribution. Moreover, we observed more semantic-aware results when applying CRA. + +Effect of Masked Body Gesture Hints. The improvement across all objective metrics demonstrates that our model effectively leverages spatial-temporal gesture priors, reducing + +
FGD ↓BC ↑Diversity ↑MSE ↓LVD ↓
Ground Truth06.89613.07400
Reconstruction3.9136.75813.1450.8416.389
Baseline13.0806.9418.31451.4429.317
+ VQVAE9.7876.67310.6241.6199.473
+ 4 VQVAE7.3976.69812.5441.2438.938
+ CRA6.8336.78312.6761.1868.715
+ Masked Hints5.4236.79413.0571.1809.015
+ +Table 6. Abliation Analysis on BEATv1.3 + +
bodyhandsfaceFGD ↓BC ↑Diversity ↑
BESTX✓✓✓5.4236.79413.057
+ Trinity[22]✓5.3196.84313.346
+ AMASS[46]✓✓5.1746.76914.318
+ +Table 7. Training IMAGE on Multiple Datasets. IMAGE demonstrates flexibility by training on multiple datasets, even when only a subset of holistic gestures is available. This approach further improves the objective metrics on the BEATv1.3 test set. + +the likelihood of incorrect gesture sampling. Furthermore, masked gesture modeling is key to enabling the network to accept predefined gestures in specific frames. + +# 5.2.3 Ability of Multi-Dataset Training. + +We demonstrate that EMAGE can effectively combine multiple non-holistic datasets for training by jointly training with the Trinity [22] and AMASS [46] datasets, by using only the upper body and audio pairs from Trinity and only body and hands from AMASS. Separate VQ-VAEs are trained for BEAT2, Trinity, and AMASS, with separate MLP heads implemented for codebook classification. The results in Table 7 show that incorporating data improves performance on the BEATv1.3 test set. + +# 6. Conclusion + +In this work, we present IMAGE, a framework designed to accept partial gestures as input for completing audio-synchronized holistic gestures. It demonstrates that leveraging masked gesture reconstruction can significantly enhance the performance of audio-conditioned gesture generation. Furthermore, the design of IMAGE enables the training on multiple datasets, further improving performance. Alongside IMAGE, we release BEAT2, the largest multi-modal gesture dataset consistent with SMPL-X and FLAME. We hope that BEAT2 will contribute to knowledge and model sharing across various subareas. + +Disclosures: While You Zhou, Xuefei Zhe, Naoya Iwamoto, and Bo Zheng are employees of Huawei Tokyo Research Center, this work was done on their own time with the approval of their employer. MJB CoI: https://files.is.tue.mpg.de/black/CoI_CVPR_2024.txt + +# References + +[1] Kfir Aberman, Peizhuo Li, Dani Lischinski, Olga Sorkine-Hornung, Daniel Cohen-Or, and Baoquan Chen. Skeleton-aware networks for deep motion retargeting. ACM Transactions on Graphics (TOG), 39(4):62-1, 2020. 2 +[2] Chaitanya Ahuja, Dong Won Lee, Yukiko I Nakano, and Louis-Philippe Morency. Style transfer for co-speech gesture animation: A multi-speaker conditional-mixture approach. In European Conference on Computer Vision, pages 248–265. Springer, 2020. 2 +[3] Simon Alexanderson, Gustav Eje Henter, Taras Kucherenko, and Jonas Beskow. Style-controllable speech-driven gesture synthesis using normalising flows. In Computer Graphics Forum, pages 487-496. Wiley Online Library, 2020. 2 +[4] Tenglong Ao, Zeyi Zhang, and Libin Liu. Gesture diffu-clip: Gesture diffusion model with clip latents. ACM Trans. Graph. 2, 5 +[5] Tenglong Ao, Qingzhe Gao, Yuke Lou, Baoquan Chen, and Libin Liu. Rhythmic gesticulator: Rhythm-aware co-speech gesture synthesis with hierarchical neural embeddings. ACM Transactions on Graphics (TOG), 41(6):1-19, 2022. 2, 3 +[6] Alexei Baevski, Yuhao Zhou, Abdelrahman Mohamed, and Michael Auli. wav2vec 2.0: A framework for self-supervised learning of speech representations. Advances in neural information processing systems, 33:12449-12460, 2020. 8 +[7] Shaojie Bai, J Zico Kolter, and Vladlen Koltun. An empirical evaluation of generic convolutional and recurrent networks for sequence modeling. arxiv 2018. arXiv preprint arXiv:1803.01271, 2, 1803. 8 +[8] Gilad Baruch, Zhuoyuan Chen, Afshin Dehghan, Tal Dimry, Yuri Feigin, Peter Fu, Thomas Gebauer, Brandon Joffe, Daniel Kurz, Arik Schwartz, and Elad Shulman. ARK-scenes - a diverse real-world dataset for 3d indoor scene understanding using mobile RGB-d data. In Thirty-fifth Conference on Neural Information Processing Systems Datasets and Benchmarks Track (Round 1), 2021. 2, 3 +[9] Piotr Bojanowski, Edouard Grave, Armand Joulin, and Tomas Mikolov. Enriching word vectors with subword information. Transactions of the association for computational linguistics, 5:135-146, 2017. 5, 8 +[10] Biovision BVH. Biovision bvh. In https://research.cs.wisc.edu/graphics /Courses/ cs-838-1999/Jeff/BVH.html, 1999. 3 +[11] Zhe Cao, Gines Hidalgo, Tomas Simon, Shih-En Wei, and Yaser Sheikh. Openpose: realtime multi-person 2d pose estimation using part affinity fields. IEEE transactions on pattern analysis and machine intelligence, 43(1):172-186, 2019. 2 +[12] Zhe Chen, Yuchen Duan, Wenhai Wang, Junjun He, Tong Lu, Jifeng Dai, and Yu Qiao. Vision transformer adapter for dense predictions. arXiv preprint arXiv:2205.08534, 2022. 3 +[13] Kiran Chhatre, Radek Danček, Nikos Athanasiou, Giorgio Becherini, Christopher Peters, Michael J Black, and Timo Bolkart. Emotional speech-driven 3d body animation via disentangled latent diffusion. arXiv preprint arXiv:2312.04466, 2023. 1 + +[14] Daniel Cudeiro, Timo Bolkart, Cassidy Laidlaw, Anurag Ranjan, and Michael Black. Capture, learning, and synthesis of 3D speaking styles. Computer Vision and Pattern Recognition (CVPR), pages 10101-10111, 2019. 2, 3, 6 +[15] Radek Dančecek, Kiran Chhatre, Shashank Tripathi, Yandong Wen, Michael Black, and Timo Bolkart. Emotional speech-driven animation with content-emotion disentanglement. ACM, 2023. 1 +[16] Anna Deichler, Kiran Chhatre, Christopher Peters, and Jonas Beskow. Spatio-temporal priors in 3d human motion. 2021. 2 +[17] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018. 3, 6 +[18] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 3 +[19] Yingruo Fan, Zhaojiang Lin, Jun Saito, Wenping Wang, and Taku Komura. Faceformer: Speech-driven 3d facial animation with transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2, 3, 6, 7, 8 +[20] Yingruo Fan, Zhaojiang Lin, Jun Saito, Wenping Wang, and Taku Komura. Faceformer: Speech-driven 3d facial animation with transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18770-18780, 2022. 1 +[21] Gabriele Fanelli, Juergen Gall, Harald Romsdorfer, Thibaut Weise, and Luc Van Gool. A 3-d audio-visual corpus of affective communication. IEEE Transactions on Multimedia, 12(6):591-598, 2010. 2 +[22] Ylva Ferstl and Rachel McDonnell. Investigating the use of recurrent motion modelling for speech gesture generation. In Proceedings of the 18th International Conference on Intelligent Virtual Agents, pages 93-98, 2018. 3, 8 +[23] Erik Gartner, Mykhaylo Andriluka, Erwin Coumans, and Cristian Sminchisescu. Differentiable dynamics for articulated 3d human motion reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13190-13200, 2022. 2 +[24] Saeed Ghorbani, Ylva Ferstl, Daniel Holden, Nikolaus F. Troje, and Marc-Andre Carbonneau. Zeroeggs: Zero-shot example-based gesture generation from speech. Computer Graphics Forum, 42(1):206-216, 2023. 2, 3 +[25] Shiry Ginosar, Amir Bar, Gefen Kohavi, Caroline Chan, Andrew Owens, and Jitendra Malik. Learning individual styles of conversational gesture. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3497-3506, 2019. 1, 2, 3, 7 +[26] Yuan Gong, Yu-An Chung, and James Glass. Ast: Audio spectrogram transformer. arXiv preprint arXiv:2104.01778, 2021.3 + +[27] Chuan Guo, Xinxin Zuo, Sen Wang, and Li Cheng. Tm2t: Stochastic and tokenized modeling for the reciprocal generation of 3d human motions and texts. In European Conference on Computer Vision, pages 580-597. Springer, 2022. 8, 4 +[28] Ikhsanul Habibie, Weipeng Xu, Dushyant Mehta, Lingjie Liu, Hans-Peter Seidel, Gerard Pons-Moll, Mohamed Elgharib, and Christian Theobalt. Learning speech-driven 3d conversational gestures from video. arXiv preprint arXiv:2102.06837, 2021. 1, 2, 3, 7, 8 +[29] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16000-16009, 2022. 3 +[30] Catalin Ionescu, Dragos Papava, Vlad Olaru, and Cristian Sminchisescu. Human3. 6m: Large scale datasets and predictive methods for 3d human sensing in natural environments. IEEE transactions on pattern analysis and machine intelligence, 36(7):1325-1339, 2013. 2 +[31] Zhenzhong Lan, Mingda Chen, Sebastian Goodman, Kevin Gimpel, Piyush Sharma, and Radu Soricut. Albert: A lite bert for self-supervised learning of language representations. arXiv preprint arXiv:1909.11942, 2019. 3, 6 +[32] Gilwoo Lee, Zhiwei Deng, Shugao Ma, Takaaki Shiratori, Siddhartha S Srinivasa, and Yaser Sheikh. Talking with hands $16.2\mathrm{m}$ : A large-scale dataset of synchronized body-finger motion and audio for conversational motion analysis and synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 763-772, 2019. 2, 3 +[33] Jing Li, Di Kang, Wenjie Pei, Xuefei Zhe, Ying Zhang, Zhenyu He, and Linchao Bao. Audio2gestures: Generating diverse gestures from speech audio with conditional variational autoencoders. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11293-11302, 2021. 1, 2, 7, 3 +[34] Ruilong Li, Shan Yang, David A Ross, and Angjoo Kanazawa. Ai choreographer: Music conditioned 3d dance generation with aist++. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13401-13412, 2021. 7, 2 +[35] Tianye Li, Timo Bolkart, Michael. J. Black, Hao Li, and Javier Romero. Learning a model of facial shape and expression from 4D scans. ACM Transactions on Graphics, (Proc. SIGGRAPH Asia), 36(6):194:1-194:17, 2017. 2, 3 +[36] Haiyang Liu and Cheng Zhang. Reinforcement learning based neural architecture search for audio tagging. In 2020 International Joint Conference on Neural Networks (IJCNN), pages 1-8. IEEE, 2020. 3 +[37] Haiyang Liu and Jihan Zhang. Improving ultrasound tongue image reconstruction from lip images using self-supervised learning and attention mechanism. arXiv preprint arXiv:2106.11769, 2021. 3 +[38] Haiyang Liu, Naoya Iwamoto, Zihao Zhu, Zhengqing Li, You Zhou, Elif Bozkurt, and Bo Zheng. Disco: Disentangled implicit content and rhythm learning for diverse co-speech gestures synthesis. In Proceedings of the 30th ACM Interna- + +tional Conference on Multimedia, pages 3764-3773, 2022. 2,3,7 +[39] Haiyang Liu, Zihao Zhu, Naoya Iwamoto, Yichen Peng, Zhengqing Li, You Zhou, Elif Bozkurt, and Bo Zheng. Beat: A large-scale semantic and emotional multi-modal dataset for conversational gestures synthesis. arXiv preprint arXiv:2203.05297, 2022. 1, 2, 3, 4, 6, 7 +[40] Xian Liu, Qianyi Wu, Hang Zhou, Yinghao Xu, Rui Qian, Xinyi Lin, Xiaowei Zhou, Wayne Wu, Bo Dai, and Bolei Zhou. Learning hierarchical cross-modal association for cospeech gesture generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10462-10472, 2022. 2, 3, 7 +[41] Xian Liu, Qianyi Wu, Hang Zhou, Yinghao Xu, Rui Qian, Xinyi Lin, Xiaowei Zhou, Wayne Wu, Bo Dai, and Bolei Zhou. Learning hierarchical cross-modal association for cospeech gesture generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10462-10472, 2022. 1, 2 +[42] Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. Roberta: A robustly optimized bert pretraining approach. arXiv preprint arXiv:1907.11692, 2019.3 +[43] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10012-10022, 2021. 3 +[44] Matthew M. Loper, Naureen Mahmood, and Michael J. Black. MoSh: Motion and shape capture from sparse markers. ACM Transactions on Graphics, (Proc. SIGGRAPH Asia), 33(6):220:1-220:13, 2014. 2 +[45] Shuhong Lu, Youngwoo Yoon, and Andrew Feng. Co-speech gesture synthesis using discrete gesture token learning. arXiv preprint arXiv:2303.12822, 2023. 2, 3 +[46] Naureen Mahmood, Nima Ghorbani, Nikolaus F. Troje, Gerard Pons-Moll, and Michael J. Black. AMASS: Archive of motion capture as surface shapes. In International Conference on Computer Vision, pages 5442-5451, 2019. 2, 3, 6, 8 +[47] Tim Massey. Vicon study of dynamic object tracking accuracy. In https://www.vicon.com/resources/blog/vicon-study-of-dynamic-object-tracking-accuracy/, 2020. 2, 3 +[48] Evonne Ng, Shiry Ginosar, Trevor Darrell, and Hanbyul Joo. Body2hands: Learning to infer 3d hands from conversational gesture body dynamics. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11865-11874, 2021. 2, 3 +[49] Kunkun Pang, Dafei Qin, Yingruo Fan, Julian Habekost, Takaaki Shiratori, Junichi Yamagishi, and Taku Komura. Bodyformer: Semantics-guided 3d body gesture synthesis with transformer. ACM Transactions on Graphics (TOG), 42(4):1-12, 2023. 2, 3 +[50] Georgios Pavlakos, Vasileios Choutas, Nima Ghorbani, Timo Bolkart, Ahmed A. A. Osman, Dimitrios Tzionas, and Michael J. Black. Expressive body capture: 3D hands, face, + +and body from a single image. In Proceedings IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), pages 10975-10985, 2019. 2, 3, 4 +[51] Ziqiao Peng, Haoyu Wu, Zhenbo Song, Hao Xu, Xiangyu Zhu, Jun He, Hongyan Liu, and Zhaoxin Fan. Emotalk: Speech-driven emotional disentanglement for 3d face animation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 20687-20697, 2023. 2 +[52] Mathis Petrovich, Michael J Black, and Gül Varol. Action-conditioned 3d human motion synthesis with transformer vae. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10985-10995, 2021. 2 +[53] Xingqun Qi, Chen Liu, Lincheng Li, Jie Hou, Haoran Xin, and Xin Yu. Emotiongesture: Audio-driven diverse emotional co-speech 3d gesture generation. arXiv preprint arXiv:2305.18891, 2023. 2 +[54] Xingqun Qi, Jiahao Pan, Peng Li, Ruibin Yuan, Xiaowei Chi, Mengfei Li, Wenhan Luo, Wei Xue, Shanghang Zhang, Qifeng Liu, et al. Weakly-supervised emotion transition learning for diverse 3d co-speech gesture generation. arXiv preprint arXiv:2311.17532, 2023. 2 +[55] Guocheng Qian, Xingdi Zhang, Abdullah Hamdi, and Bernard Ghanem. Pix4point: Image pretrained transformers for 3d point cloud understanding. 2022. 3 +[56] Alexander Richard, Michael Zollhöfer, Yandong Wen, Fernando De la Torre, and Yaser Sheikh. Meshtalk: 3d face animation from speech using cross-modality disentanglement. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1173-1182, 2021. 2, 3 +[57] Kaede Shiohara, Xingchao Yang, and Takafumi Taketomi. Blendface: Re-designing identity encoders for face-swapping. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7634-7644, 2023. 2 +[58] Tommi Tykkälä, Cédric Audras, and Andrew I Comport. Direct iterative closest point for real-time visual odometry. In 2011 IEEE International Conference on Computer Vision Workshops (ICCV Workshops), pages 2050-2056. IEEE, 2011. 5 +[59] Aaron Van Den Oord, Oriol Vinyals, et al. Neural discrete representation learning. Advances in neural information processing systems, 30, 2017. 5, 8 +[60] Kaisiyuan Wang, Qianyi Wu, Linsen Song, Zhuoqian Yang, Wayne Wu, Chen Qian, Ran He, Yu Qiao, and Chen Change Loy. Mead: A large-scale audio-visual dataset for emotional talking-face generation. In ECCV, 2020. 2 +[61] Jinbo Xing, Menghan Xia, Yuechen Zhang, Xiaodong Cun, Jue Wang, and Tien-Tsin Wong. Codetalker: Speech-driven 3d facial animation with discrete motion prior. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12780-12790, 2023. 1, 2, 3, 4, 7, 8 +[62] Sicheng Yang, Zhiyong Wu, Minglei Li, Zhensong Zhang, Lei Hao, Weihong Bao, Ming Cheng, and Long Xiao. Diffusestylegesture: Stylized audio-driven co-speech gesture generation with diffusion models. arXiv preprint arXiv:2305.04919, 2023. 2, 3, 7 + +[63] Xingchao Yang and Takafumi Taketomi. Bareskinnet: Demakeup and de-lighting via 3d face reconstruction. In Computer Graphics Forum, pages 623-634. Wiley Online Library, 2022. 2 +[64] Xingchao Yang, Takafumi Taketomi, and Yoshihiro Kanamori. Makeup extraction of 3d representation via illumination-aware image decomposition. In Computer Graphics Forum, pages 293-307. Wiley Online Library, 2023. 2 +[65] Hongwei Yi, Hualin Liang, Yifei Liu, Qiong Cao, Yandong Wen, Timo Bolkart, Dacheng Tao, and Michael J Black. Generating holistic 3d human motion from speech. In CVPR, 2023. 1, 2, 3, 4, 5, 6, 7, 8 +[66] Youngwoo Yoon, Woo-Ri Ko, Minsu Jang, Jaeyeon Lee, Jae-hong Kim, and Geehyuk Lee. Robots learn social skills: End-to-end learning of co-speech gesture generation for humanoid robots. In 2019 International Conference on Robotics and Automation (ICRA), pages 4303-4309. IEEE, 2019. 2, 3 +[67] Youngwoo Yoon, Bok Cha, Joo-Haeng Lee, Minsu Jang, Jaeyeon Lee, Jaehong Kim, and Geehyuk Lee. Speech gesture generation from the trimodal context of text, audio, and speaker identity. ACM Transactions on Graphics (TOG), 39 (6):1-16, 2020. 1, 2, 3, 7 +[68] Xumin Yu, Lulu Tang, Yongming Rao, Tiejun Huang, Jie Zhou, and Jiwen Lu. Point-bert: Pre-training 3d point cloud transformers with masked point modeling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19313-19322, 2022. 3 +[69] Hengshuang Zhao, Li Jiang, Jiaya Jia, Philip HS Torr, and Vladlen Koltun. Point transformer. In Proceedings of the IEEE/CVF international conference on computer vision, pages 16259-16268, 2021. 3 +[70] Yihao Zhi, Xiaodong Cun, Xuelin Chen, Xi Shen, Wen Guo, Shaoli Huang, and Shenghua Gao. Livelyspeaker: Towards semantic-aware co-speech gesture generation. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 20807-20817, 2023. 2, 3 +[71] Yang Zhou, Jimei Yang, Dingzeyu Li, Jun Saito, Deepali Aneja, and Evangelos Kalogerakis. Audio-driven neural gesture reenactment with video motion graphs. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3418-3428, 2022. 2 +[72] Lingting Zhu, Xian Liu, Xuanyu Liu, Rui Qian, Ziwei Liu, and Lequan Yu. Taming diffusion models for audio-driven co-speech gesture generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10544-10553, 2023. 2, 3 +[73] Wentao Zhu, Xiaoxuan Ma, Zhaoyang Liu, Libin Liu, Wayne Wu, and Yizhou Wang. Motionbert: A unified perspective on learning human motion representations. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2023. 3 \ No newline at end of file diff --git a/2024/EMAGE_ Towards Unified Holistic Co-Speech Gesture Generation via Expressive Masked Audio Gesture Modeling/images.zip b/2024/EMAGE_ Towards Unified Holistic Co-Speech Gesture Generation via Expressive Masked Audio Gesture Modeling/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..52d6952d06daa14cea5ccc590b34de07894da13b --- /dev/null +++ b/2024/EMAGE_ Towards Unified Holistic Co-Speech Gesture Generation via Expressive Masked Audio Gesture Modeling/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9e22bf9b2357d77a2b649b18fefa1f07a46110c1b49e9db0788c106ab73de39e +size 588813 diff --git a/2024/EMAGE_ Towards Unified Holistic Co-Speech Gesture Generation via Expressive Masked Audio Gesture Modeling/layout.json b/2024/EMAGE_ Towards Unified Holistic Co-Speech Gesture Generation via Expressive Masked Audio Gesture Modeling/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..e58ffae740608c341c9b40a19455a5e763487dd9 --- /dev/null +++ b/2024/EMAGE_ Towards Unified Holistic Co-Speech Gesture Generation via Expressive Masked Audio Gesture Modeling/layout.json @@ -0,0 +1,10813 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 80, + 102, + 515, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 80, + 102, + 515, + 140 + ], + "spans": [ + { + "bbox": [ + 80, + 102, + 515, + 140 + ], + "type": "text", + "content": "EMAGE: Towards Unified Holistic Co-Speech Gesture Generation via Expressive Masked Audio Gesture Modeling" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 88, + 159, + 503, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 88, + 159, + 503, + 191 + ], + "spans": [ + { + "bbox": [ + 88, + 159, + 503, + 191 + ], + "type": "text", + "content": "Haiyang Liu" + }, + { + "bbox": [ + 88, + 159, + 503, + 191 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 88, + 159, + 503, + 191 + ], + "type": "text", + "content": " Zihao Zhu" + }, + { + "bbox": [ + 88, + 159, + 503, + 191 + ], + "type": "inline_equation", + "content": "^{2*}" + }, + { + "bbox": [ + 88, + 159, + 503, + 191 + ], + "type": "text", + "content": " Giorgio Becherini" + }, + { + "bbox": [ + 88, + 159, + 503, + 191 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 88, + 159, + 503, + 191 + ], + "type": "text", + "content": " Yichen Peng" + }, + { + "bbox": [ + 88, + 159, + 503, + 191 + ], + "type": "inline_equation", + "content": "^{5}" + }, + { + "bbox": [ + 88, + 159, + 503, + 191 + ], + "type": "text", + "content": " Mingyang Su" + }, + { + "bbox": [ + 88, + 159, + 503, + 191 + ], + "type": "inline_equation", + "content": "^{6}" + }, + { + "bbox": [ + 88, + 159, + 503, + 191 + ], + "type": "text", + "content": " You Zhou" + }, + { + "bbox": [ + 88, + 159, + 503, + 191 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 88, + 159, + 503, + 191 + ], + "type": "text", + "content": " Xuefei Zhe" + }, + { + "bbox": [ + 88, + 159, + 503, + 191 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 88, + 159, + 503, + 191 + ], + "type": "text", + "content": " Naoya Iwamoto" + }, + { + "bbox": [ + 88, + 159, + 503, + 191 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 88, + 159, + 503, + 191 + ], + "type": "text", + "content": " Bo Zheng" + }, + { + "bbox": [ + 88, + 159, + 503, + 191 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 88, + 159, + 503, + 191 + ], + "type": "text", + "content": " Michael J. Black" + }, + { + "bbox": [ + 88, + 159, + 503, + 191 + ], + "type": "inline_equation", + "content": "^{3}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 95, + 202, + 483, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 95, + 202, + 483, + 232 + ], + "spans": [ + { + "bbox": [ + 95, + 202, + 483, + 232 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 95, + 202, + 483, + 232 + ], + "type": "text", + "content": "The University of Tokyo " + }, + { + "bbox": [ + 95, + 202, + 483, + 232 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 95, + 202, + 483, + 232 + ], + "type": "text", + "content": "Keio University " + }, + { + "bbox": [ + 95, + 202, + 483, + 232 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 95, + 202, + 483, + 232 + ], + "type": "text", + "content": "Independent Researcher " + }, + { + "bbox": [ + 95, + 202, + 483, + 232 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 95, + 202, + 483, + 232 + ], + "type": "text", + "content": "Max Planck Institute for Intelligent Systems " + }, + { + "bbox": [ + 95, + 202, + 483, + 232 + ], + "type": "inline_equation", + "content": "^{5}" + }, + { + "bbox": [ + 95, + 202, + 483, + 232 + ], + "type": "text", + "content": "JAIST " + }, + { + "bbox": [ + 95, + 202, + 483, + 232 + ], + "type": "inline_equation", + "content": "^{6}" + }, + { + "bbox": [ + 95, + 202, + 483, + 232 + ], + "type": "text", + "content": "Tsinghua University" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 71, + 259, + 525, + 400 + ], + "blocks": [ + { + "bbox": [ + 71, + 259, + 525, + 400 + ], + "lines": [ + { + "bbox": [ + 71, + 259, + 525, + 400 + ], + "spans": [ + { + "bbox": [ + 71, + 259, + 525, + 400 + ], + "type": "image", + "image_path": "77aa6d75a5af7f91c15f89f767785f7436b00b41c1d6074cbaa3e215dafb7c19.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 45, + 404, + 548, + 440 + ], + "lines": [ + { + "bbox": [ + 45, + 404, + 548, + 440 + ], + "spans": [ + { + "bbox": [ + 45, + 404, + 548, + 440 + ], + "type": "text", + "content": "Figure 1. EMAGE. We present a Masked Audio-Conditioned Gesture Modeling framework, along with a new holistic gesture dataset, BEAT2 (BEAT-SMPLX-FLAME), for jointly generating facial expressions, local body dynamics, hand movements and global translations, conditioned on audio and a partially or completely masked gestures. The gray denotes visible gestures, and blue represents our outputs." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 143, + 448, + 192, + 462 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 448, + 192, + 462 + ], + "spans": [ + { + "bbox": [ + 143, + 448, + 192, + 462 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 474, + 290, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 474, + 290, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 474, + 290, + 715 + ], + "type": "text", + "content": "We propose EMAGE, a framework to generate full-body human gestures from audio and masked gestures, encompassing facial, local body, hands, and global movements. To achieve this, we first introduce BEAT2 (BEATSMLPX-FLAME), a new mesh-level holistic co-speech dataset. BEAT2 combines a MoShed SMPL-X body with FLAME head parameters and further refines the modeling of head, neck, and finger movements, offering a community-standardized, high-quality 3D motion captured dataset. EMAGE leverages masked body gesture priors during training to boost inference performance. It involves a Masked Audio Gesture Transformer, facilitating joint training on audio-to-gesture generation and masked gesture reconstruction to effectively encode audio and body gesture hints. Encoded body hints from masked gestures are then separately employed to generate facial and body movements. Moreover, EMAGE adaptively merges speech features from the audio's rhythm and content and utilizes four compositional VQ-VAEs to enhance the results' fidelity and diversity. Experiments demonstrate that EMAGE generates" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 449, + 547, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 449, + 547, + 498 + ], + "spans": [ + { + "bbox": [ + 304, + 449, + 547, + 498 + ], + "type": "text", + "content": "holistic gestures with state-of-the-art performance and is flexible in accepting predefined spatial-temporal gesture inputs, generating complete, audio-synchronized results. Our code and dataset are available.1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 521, + 388, + 533 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 521, + 388, + 533 + ], + "spans": [ + { + "bbox": [ + 306, + 521, + 388, + 533 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 540, + 547, + 686 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 540, + 547, + 686 + ], + "spans": [ + { + "bbox": [ + 304, + 540, + 547, + 686 + ], + "type": "text", + "content": "The path towards full-body co-speech gesture generation presents many remaining challenges. Despite the development of various baselines, e.g., audio conditioned models for face [15, 20, 61], body [13, 25, 67], and hand movements [33, 39, 41], along with a few attempts at merged models [28, 65], the limited availability of comprehensive data and models poses an ongoing obstacle. To make progress in this direction, we present a new framework that can incorporate partial spatial-temporal predefined gestures and autonomously complete the remaining frames in synchronization with the audio. This provides a new tool for creating realistic animations of digital humans" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 320, + 693, + 496, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 693, + 496, + 713 + ], + "spans": [ + { + "bbox": [ + 320, + 693, + 496, + 713 + ], + "type": "text", + "content": "*Equal Contribution. \n1https://pantomatrix.github.io/EMAGE/" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "1144" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 289, + 715 + ], + "type": "list", + "angle": 0, + "index": 3, + "blocks": [ + { + "bbox": [ + 47, + 72, + 113, + 83 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 113, + 83 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 113, + 83 + ], + "type": "text", + "content": "[57, 63, 64, 71]." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 88, + 289, + 447 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 88, + 289, + 447 + ], + "spans": [ + { + "bbox": [ + 47, + 88, + 289, + 447 + ], + "type": "text", + "content": "To achieve this, we first require a comprehensive gesture dataset. While several datasets are available for audio-to-body [67], audio-to-body&face [28], body-to-hands [39, 41], and common actions [16, 46], integrating them poses challenges due to varying data formats and evaluation metrics across different sub-areas. This motivates us to establish a unified, community-standard benchmark for training and evaluating holistic gestures. We build upon the BEAT dataset [39], which provides the most comprehensive mocap co-speech gesture dataset with diverse modalities. Despite its size and richness, the data in BEAT is not presented in a standardized format. The dataset comprises iPhone ARKit blendshape weights and Vicon skeletons, posing challenges when using the data for training; e.g., the lack of a mesh representation prevents the use of a vertex loss [19, 52]. Additionally, conducting direct human perceptual evaluations on these diverse modalities is inherently complex. On the other hand, SMPL-X [50] and FLAME [35] are common mesh standards for the academic community for other human dynamics-related tasks, e.g., action [52] and talking head generation [14]. Driven by the motivation to facilitate the sharing of knowledge across different tasks, we present BEAT-SMPLX-FLAME (BEAT2), which consists of two main components: i) Refined SMPL-X body shape and pose parameters from MoSh++ [44, 46] with hard-coded physical priors. ii) Optimized high-quality FLAME head parameters. By integrating SMPL-X's body and FLAME's head, BEAT2 facilitates a comprehensive training and evaluation benchmark across multiple sub-areas of co-speech human animation generation." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 450, + 289, + 715 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 450, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 47, + 450, + 289, + 715 + ], + "type": "text", + "content": "With the holistic data from BEAT2, we aim to enhance the coherence of individual body parts while ensuring accurate cross-modal alignment between audio and motion; see Figure 1. This leads to the design of Expressive Masked Audio-conditioned GEsture modeling, denoted as EMAGE, a spatial and temporal transformer-based framework. EMAGE first aggregates spatial features from masked body joints. Then, it reconstructs the latent space of pretrained gestures with a switchable gesture temporal self-attention and audio-gesture cross-attention. The selection of different forward paths enables modeling the gesture-to-gesture and audio-to-gesture prior separately and effectively. Once the reconstructed latent features are obtained, EMAGE decodes local face and body gestures from four compositional pretrained Vector Quantized-Variational AutoEncoders (VQ-VAEs) and decodes global translations from a pretrained Global Motion Predictor. The compositional VQ-VAEs are key to preserving audio-related motions during the reconstruction. Through these designs, EMAGE achieves state-of-the-art performance in generating body and face gestures. It takes as input audio and partially initialized gestures, and recovers audio" + } + ] + } + ], + "index": 2 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 304, + 72, + 547, + 252 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 304, + 72, + 545, + 108 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 108 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 108 + ], + "type": "text", + "content": "synchronized, coherent gestures. Additionally, we show how EMAGE can flexibly incorporate additional, nonholistic, datasets to improve results." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 108, + 547, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 108, + 547, + 252 + ], + "spans": [ + { + "bbox": [ + 304, + 108, + 547, + 252 + ], + "type": "text", + "content": "Overall, our contributions are as follows: (1) We release BEAT2, a representation-unified, mesh-level dataset that leverages MoShed SMPL-X body and FLAME head parameters. (2) We propose EMAGE, a simple yet effective holistic gesture generation framework that generates coherent gestures with partial gesture and audio priors. (3) EMAGE achieves state-of-the-art performance in generating both body and face gestures with only four frames of seed gestures. (4) EMAGE showcases the use of additional non-holistic gesture datasets for training, e.g., Trinity and AMASS, to further improve the fidelity and diversity of the results, effectively leveraging data from different datasets." + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 306, + 262, + 392, + 275 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 262, + 392, + 275 + ], + "spans": [ + { + "bbox": [ + 306, + 262, + 392, + 275 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 283, + 547, + 715 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 304, + 283, + 547, + 604 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 283, + 547, + 604 + ], + "spans": [ + { + "bbox": [ + 304, + 283, + 547, + 604 + ], + "type": "text", + "content": "Co-Speech Animation Datasets are categorized into two types; see Table 1: pseudo-labeled (PGT) and motion-captured (mocap). For PGT, the Speech2Gesture Dataset [25] utilizes OpenPose [11] to extract 2D poses from News and Teaching videos. Subsequent works extend the dataset to 3D pose [28] and SMPL-X [65]. The TED Dataset [66], which extracts 2D pose from TED videos, has been extended to include 3D pose [67], fingers [40], and SMPL-X [45]. Similarly, 2D, 3D landmarks, and meshes are estimated from multi-view recorded videos [60] as PGT for talking-head generation. Although pseudo-labeled approaches could theoretically extract infinite data, their accuracy is limited. For instance, the average error (in mm) for the SOTA monocular 3D pose estimation algorithm [23] on the Human3.6M dataset [30] is 33.4, while for Vicon mocap it is 0.142 [47]. For mocap datasets, Trinity [3] features one male actor with 4h of data, and TalkingWithHands [32] collects conversation scenarios for two speakers. ZEEG [24] considers one speaker with 12 styles. For the face, 3D scanning is accurate but limited in quantity due to cost, e.g., less than 3h for BIWI [21], VOCA [14], and MeshTalk [56]. This leads to a balance between performance and quantity in the ARKit dataset [51]. The aforementioned datasets address the body and face separately. On the other hand, BEAT [39] includes both 3D pose body and ARKit [8] facial blendshape weights for the first time. However, BEAT lacks mesh data." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "type": "text", + "content": "Co-Speech Animation Models are categorized based on their outputs; see Table 2. In addition to the baselines and datasets [14, 25, 32, 48, 56, 67], several frameworks for body gestures [2, 33, 53, 54] have improved the performance of the baselines. They are trained and evaluated by selecting specific upper body joints [4, 5, 38-40, 49, 70], or all body joints [62, 72]. Recent improvements in facial gestures drive the vertices with transformers and discrete face priors [19, 61]. However, these methods only address ei" + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "1145" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 50, + 69, + 541, + 135 + ], + "blocks": [ + { + "bbox": [ + 50, + 69, + 541, + 135 + ], + "lines": [ + { + "bbox": [ + 50, + 69, + 541, + 135 + ], + "spans": [ + { + "bbox": [ + 50, + 69, + 541, + 135 + ], + "type": "table", + "html": "
Trinity Mocap 2017 [22]S2G S2G-2D 2019 [25]Seq2Seq TED-2D 2019 [66]TWH Mocap 2019 [32]Trimodal TED-3D 2020 [67]VOCA Scan 2020 [14]MeshTalk Scan 2021 [56]Hahibie et al. S2G-3D 2021 [28]HA2G TED-3D+ 2022 [40]BEAT Mocap 2022 [39]ZEEG Mocap 2022 [24]Yoon. et al. TED-SMPL 2023 [45]Talkshow S2G-SMPL 2023 [65]BEAT2 Ours 2024
Head-2D---3D Scan3D Scan3D-ARKit--PGT-MeshMC-Mesh
Upper Body3D2D2D3D3D--3D3D3D3DPGT-MeshPGT-MeshMC-Mesh
Hands3D2D-3D---3D3D3D3DPGT-MeshPGT-MeshMC-Mesh
Lower Body3D--3D-----3D3D--MC-Mesh
Global Motion3D--3D-----3D3D--MC-Mesh
Duration (hours)4609720970.523833764302760
", + "image_path": "66e6b8d9cd0bc32106e05b2751c2989dd9fc80f1283577f7cd48d24a19ffe185.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 50, + 181, + 541, + 235 + ], + "blocks": [ + { + "bbox": [ + 46, + 138, + 546, + 172 + ], + "lines": [ + { + "bbox": [ + 46, + 138, + 546, + 172 + ], + "spans": [ + { + "bbox": [ + 46, + 138, + 546, + 172 + ], + "type": "text", + "content": "Table 1. Comparison of Co-Speech Gesture Datasets. We summarize gesture and face datasets in talking scenarios. 'PGT' and 'MC' denote Pseudo Ground Truth and Motion Capture, respectively. The BEAT2 (Ours) dataset is the largest dataset for motion-captured data, providing holistic, academic community standard mesh-level information." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 50, + 181, + 541, + 235 + ], + "lines": [ + { + "bbox": [ + 50, + 181, + 541, + 235 + ], + "spans": [ + { + "bbox": [ + 50, + 181, + 541, + 235 + ], + "type": "table", + "html": "
S2GTriModalB2HTWHHahibie et al.DisCoHA2GFace Former 2022[19]RhythmicBEAT CaMN 2022[39]Diff Gesture 2023[72]Code Talker 2023[61]Talk Show 2023[65]DiffStyle Gesture 2023[62]Body Former 2023[49]Lively Speaker 2023[70]EMAEG (Ours) 2024
2019 [25]2020 [67]2020[48]2021[32]2021[28]2022[38]2022[40]2022[5]2022[5]-------Masked
Input Gestures--BodyBody------------
Head---M--S---SS---C
UpperSS-MSC-SCS-CSSSC
Hands--SSMSC--CS-CS-SC
Lower + Global------------S--C
", + "image_path": "e89e22ee13f61b8b34cf4be94f8611a9fa64537145333d53cd95adc63906479a.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 239, + 547, + 283 + ], + "lines": [ + { + "bbox": [ + 46, + 239, + 547, + 283 + ], + "spans": [ + { + "bbox": [ + 46, + 239, + 547, + 283 + ], + "type": "text", + "content": "Table 2. Comparison of Co-Speech Gesture Models. We compare with previous methods for generating face or body motion trained on co-speech datasets. The first row lists their inputs, and the subsequent rows list their outputs, respectively. Different decoder designs are denoted by the initials 'S' for Single, 'M' for Multiple, and 'C' for Cascaded. To the best of our knowledge, EMAGE (Ours) is the first to accept audio and partially or completely masked gestures, generating full-body audio-synchronized results." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 304, + 288, + 448 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 304, + 288, + 448 + ], + "spans": [ + { + "bbox": [ + 46, + 304, + 288, + 448 + ], + "type": "text", + "content": "ther the face or body. Applying these methods to the full body will yield sub-optimal results as audio is differently correlated with face and body dynamics. Most similar to our work, Habibie et al. [28] employ a single audio encoder and multiple decoders for generating face and body gestures. TalkSHOW [65] demonstrates the advantage of separating the audio encoders and decoding the body and hands using quantized codes in an auto-regressive manner. However, it lacks lower body and global motion, and there is no shared information for body and facial gesture generation. Moreover, it cannot accept partially masked body hints due to the design of a fully auto-regressive model." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 449, + 288, + 654 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 449, + 288, + 654 + ], + "spans": [ + { + "bbox": [ + 46, + 449, + 288, + 654 + ], + "type": "text", + "content": "Masked Representation Learning was first demonstrated to be effective in Natural Language Processing, with BERT-based models [17, 31, 42] boosting the performance of learned word embeddings in downstream tasks through a combination of masked language modeling and transformer architecture. Subsequently, Masked AutoEncoders [29] expanded masked image modeling to computer vision by removing and inpainting image patches. This concept of masked representation learning has since been employed in other modalities, e.g., video [12, 18, 43], audio [26, 36, 37], and point clouds [55, 68, 69]. Most related to our work, MotionBERT [73] proposes a spatial-temporal transformer to learn a robust motion representation for classification-based tasks by masking 2D pose. In contrast to their method, we target robust motion features for conditional motion generation tasks, requiring a balance in training between multiple modalities, e.g., audio and gesture." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 668, + 100, + 680 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 668, + 100, + 680 + ], + "spans": [ + { + "bbox": [ + 47, + 668, + 100, + 680 + ], + "type": "text", + "content": "3. BEAT2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 689, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 689, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 689, + 288, + 715 + ], + "type": "text", + "content": "This section introduces how we obtain unified mesh-level data, i.e., SMPL-X [50] and FLAME [35] parameters, from" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 304, + 547, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 304, + 547, + 413 + ], + "spans": [ + { + "bbox": [ + 304, + 304, + 547, + 413 + ], + "type": "text", + "content": "the original BEAT dataset [39]. BEAT utilizes a Vicon motion capture system [47] and releases 78-joint skeleton-level Biovision Hierarchy (BVH) [10] motion files. Their facial capture system uses ARKit [8] with a depth camera on the iPhone 12 Pro, extracting " + }, + { + "bbox": [ + 304, + 304, + 547, + 413 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^{51}" + }, + { + "bbox": [ + 304, + 304, + 547, + 413 + ], + "type": "text", + "content": " blendshape weights. These blendshapes are designed based on the Facial Action Coding System (FACS), which is widely adopted. However, both the motion and facial data lack mesh-level details (see Figure 2), e.g., shapes and vertices." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 430, + 533, + 443 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 430, + 533, + 443 + ], + "spans": [ + { + "bbox": [ + 305, + 430, + 533, + 443 + ], + "type": "text", + "content": "3.1. Body Parameters Initialization via MoSh++" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 451, + 545, + 571 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 451, + 545, + 571 + ], + "spans": [ + { + "bbox": [ + 304, + 451, + 545, + 571 + ], + "type": "text", + "content": "We initialize the SMPL-X body shape and pose parameters from BEAT mocap marker data using MoSh++ [46]. Given the captured markers positions " + }, + { + "bbox": [ + 304, + 451, + 545, + 571 + ], + "type": "inline_equation", + "content": "\\mathbf{m} \\in \\mathbb{R}^{T \\times K \\times 3}" + }, + { + "bbox": [ + 304, + 451, + 545, + 571 + ], + "type": "text", + "content": ", predefined markers position offsets " + }, + { + "bbox": [ + 304, + 451, + 545, + 571 + ], + "type": "inline_equation", + "content": "\\mathbf{d} \\in \\mathbb{R}^{K \\times 3}" + }, + { + "bbox": [ + 304, + 451, + 545, + 571 + ], + "type": "text", + "content": ", and user-defined vertices-to-markers function " + }, + { + "bbox": [ + 304, + 451, + 545, + 571 + ], + "type": "inline_equation", + "content": "\\mathcal{H}" + }, + { + "bbox": [ + 304, + 451, + 545, + 571 + ], + "type": "text", + "content": ", we aim to obtain body shape " + }, + { + "bbox": [ + 304, + 451, + 545, + 571 + ], + "type": "inline_equation", + "content": "\\beta \\in \\mathbb{R}^{300}" + }, + { + "bbox": [ + 304, + 451, + 545, + 571 + ], + "type": "text", + "content": ", poses " + }, + { + "bbox": [ + 304, + 451, + 545, + 571 + ], + "type": "inline_equation", + "content": "\\theta \\in \\mathbb{R}^{T \\times 55 \\times 3}" + }, + { + "bbox": [ + 304, + 451, + 545, + 571 + ], + "type": "text", + "content": ", and translation parameters " + }, + { + "bbox": [ + 304, + 451, + 545, + 571 + ], + "type": "inline_equation", + "content": "\\gamma \\in \\mathbb{R}^{T \\times 3}" + }, + { + "bbox": [ + 304, + 451, + 545, + 571 + ], + "type": "text", + "content": ". The optimization uses a differentiable surface vertex mapping function " + }, + { + "bbox": [ + 304, + 451, + 545, + 571 + ], + "type": "inline_equation", + "content": "S(\\beta, \\theta, \\gamma)" + }, + { + "bbox": [ + 304, + 451, + 545, + 571 + ], + "type": "text", + "content": " and vertex normal function " + }, + { + "bbox": [ + 304, + 451, + 545, + 571 + ], + "type": "inline_equation", + "content": "\\mathcal{N}(\\beta, \\theta_0, \\gamma_0)" + }, + { + "bbox": [ + 304, + 451, + 545, + 571 + ], + "type": "text", + "content": ". For each frame, the latent marker " + }, + { + "bbox": [ + 304, + 451, + 545, + 571 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{m}} \\in \\mathbb{R}^{T \\times K \\times 3}" + }, + { + "bbox": [ + 304, + 451, + 545, + 571 + ], + "type": "text", + "content": " is calculated as" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 346, + 588, + 545, + 602 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 346, + 588, + 545, + 602 + ], + "spans": [ + { + "bbox": [ + 346, + 588, + 545, + 602 + ], + "type": "interline_equation", + "content": "\\tilde {\\mathbf {m}} _ {i} \\equiv \\mathcal {S} _ {\\mathcal {H}} (\\beta , \\theta_ {i}, \\gamma_ {i}) + \\mathbf {d N} _ {\\mathcal {H}} (\\beta , \\theta_ {i}, \\gamma_ {i}). \\tag {1}", + "image_path": "983173b01e6d9811f392dc6b1235be9c0aa2493ad6ea04535fc55b59e05a7fcd.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 617, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 617, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 617, + 547, + 715 + ], + "type": "text", + "content": "We first select 12 frames to optimize and fix " + }, + { + "bbox": [ + 304, + 617, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{d}" + }, + { + "bbox": [ + 304, + 617, + 547, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 617, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 304, + 617, + 547, + 715 + ], + "type": "text", + "content": ", then optimize " + }, + { + "bbox": [ + 304, + 617, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\theta_{i}" + }, + { + "bbox": [ + 304, + 617, + 547, + 715 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 617, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\gamma_{i}" + }, + { + "bbox": [ + 304, + 617, + 547, + 715 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 304, + 617, + 547, + 715 + ], + "type": "inline_equation", + "content": "i \\in (1:T)" + }, + { + "bbox": [ + 304, + 617, + 547, + 715 + ], + "type": "text", + "content": " by minimizing the loss terms based on " + }, + { + "bbox": [ + 304, + 617, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\|\\tilde{\\mathbf{m}}_i - \\mathbf{m}_i\\|^2" + }, + { + "bbox": [ + 304, + 617, + 547, + 715 + ], + "type": "text", + "content": ", including Data Term, Surface Distance Energy, Marker Initialization Regularization, Pose and Shape Priors, Velocity Constancy, and Soft-Tissue Term (see supplementary materials). The overall objective function is the weighted sum of these terms to balance accuracy and plausibility." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "1146" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 78, + 72, + 121, + 175 + ], + "blocks": [ + { + "bbox": [ + 78, + 72, + 121, + 175 + ], + "lines": [ + { + "bbox": [ + 78, + 72, + 121, + 175 + ], + "spans": [ + { + "bbox": [ + 78, + 72, + 121, + 175 + ], + "type": "image", + "image_path": "1a29c1ea8463bf00c2c3f184a817a4f1c708628a8ef1d47d76d7f19f87bcfc5d.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 82, + 178, + 121, + 217 + ], + "blocks": [ + { + "bbox": [ + 82, + 178, + 121, + 217 + ], + "lines": [ + { + "bbox": [ + 82, + 178, + 121, + 217 + ], + "spans": [ + { + "bbox": [ + 82, + 178, + 121, + 217 + ], + "type": "image", + "image_path": "0d3097e295ab5f39d373170046e3ede29bd44caca6a90f7a1e16e8cbbc74f8e2.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 85, + 219, + 104, + 226 + ], + "lines": [ + { + "bbox": [ + 85, + 219, + 104, + 226 + ], + "spans": [ + { + "bbox": [ + 85, + 219, + 104, + 226 + ], + "type": "text", + "content": "BEAT" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 124, + 73, + 182, + 175 + ], + "blocks": [ + { + "bbox": [ + 124, + 73, + 182, + 175 + ], + "lines": [ + { + "bbox": [ + 124, + 73, + 182, + 175 + ], + "spans": [ + { + "bbox": [ + 124, + 73, + 182, + 175 + ], + "type": "image", + "image_path": "3cf8e0bff276205e84046de305c5ed3d778dd92008b38ecc417fbe3aa563ac59.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 124, + 178, + 182, + 217 + ], + "blocks": [ + { + "bbox": [ + 124, + 178, + 182, + 217 + ], + "lines": [ + { + "bbox": [ + 124, + 178, + 182, + 217 + ], + "spans": [ + { + "bbox": [ + 124, + 178, + 182, + 217 + ], + "type": "image", + "image_path": "9c115889ca59e8c11153c9150f32df01371750dbb710af599fab43cfc7f8e73d.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 140, + 220, + 170, + 227 + ], + "lines": [ + { + "bbox": [ + 140, + 220, + 170, + 227 + ], + "spans": [ + { + "bbox": [ + 140, + 220, + 170, + 227 + ], + "type": "text", + "content": "Retargeted" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 189, + 73, + 244, + 175 + ], + "blocks": [ + { + "bbox": [ + 189, + 73, + 244, + 175 + ], + "lines": [ + { + "bbox": [ + 189, + 73, + 244, + 175 + ], + "spans": [ + { + "bbox": [ + 189, + 73, + 244, + 175 + ], + "type": "image", + "image_path": "0b2f09f2374956161351194f45e2e2532c9d05df8cf0f908d5ce265ba34ba94c.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 189, + 178, + 247, + 217 + ], + "blocks": [ + { + "bbox": [ + 189, + 178, + 247, + 217 + ], + "lines": [ + { + "bbox": [ + 189, + 178, + 247, + 217 + ], + "spans": [ + { + "bbox": [ + 189, + 178, + 247, + 217 + ], + "type": "image", + "image_path": "f8f4640aefd7e52843ec5ee7f210087f3efb995a8208587356d71d31ebcdc3c4.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 206, + 219, + 230, + 226 + ], + "lines": [ + { + "bbox": [ + 206, + 219, + 230, + 226 + ], + "spans": [ + { + "bbox": [ + 206, + 219, + 230, + 226 + ], + "type": "text", + "content": "Moshed" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 231, + 547, + 298 + ], + "lines": [ + { + "bbox": [ + 46, + 231, + 547, + 298 + ], + "spans": [ + { + "bbox": [ + 46, + 231, + 547, + 298 + ], + "type": "text", + "content": "Figure 2. Comparison of Data between BEAT2 and Others. BEAT-SMPLX-FLAME presents a new mesh-level, motion-captured, holistic co-speech gesture dataset with 60h of data. Left: We compare our refined SMPL-X body parameters (denoted as Refined MoSh) with the original BEAT skeleton [39], Retargeted from AutoRegPro, and initial results of " + }, + { + "bbox": [ + 46, + 231, + 547, + 298 + ], + "type": "inline_equation", + "content": "\\mathrm{Mosh}++" + }, + { + "bbox": [ + 46, + 231, + 547, + 298 + ], + "type": "text", + "content": " [50]. The refined results show correct neck flexion, appropriate head and neck shape ratios, and detailed finger representations. Right: Visualization of blendshape weights from the original BEAT dataset [39] with ARKit's template, Wrapped-based, and handcrafted optimization. Our final handcrafted FLAME blendshape-based optimization demonstrates both accurate lip movement details and natural mouth shapes." + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 250, + 73, + 310, + 175 + ], + "blocks": [ + { + "bbox": [ + 250, + 73, + 310, + 175 + ], + "lines": [ + { + "bbox": [ + 250, + 73, + 310, + 175 + ], + "spans": [ + { + "bbox": [ + 250, + 73, + 310, + 175 + ], + "type": "image", + "image_path": "3c1b9aa5d1182a6c0f9d831ead367f4923e1919823cf79b76dca4f59fb4552c0.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 250, + 177, + 310, + 217 + ], + "blocks": [ + { + "bbox": [ + 250, + 177, + 310, + 217 + ], + "lines": [ + { + "bbox": [ + 250, + 177, + 310, + 217 + ], + "spans": [ + { + "bbox": [ + 250, + 177, + 310, + 217 + ], + "type": "image", + "image_path": "998179254e8dbd0a0cd08de2c91b898bf38655be992f223fb73fe5128e9c24bd.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 250, + 219, + 314, + 227 + ], + "lines": [ + { + "bbox": [ + 250, + 219, + 314, + 227 + ], + "spans": [ + { + "bbox": [ + 250, + 219, + 314, + 227 + ], + "type": "text", + "content": "Refined Moshed (Ours)" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 321, + 76, + 375, + 175 + ], + "blocks": [ + { + "bbox": [ + 321, + 76, + 375, + 175 + ], + "lines": [ + { + "bbox": [ + 321, + 76, + 375, + 175 + ], + "spans": [ + { + "bbox": [ + 321, + 76, + 375, + 175 + ], + "type": "image", + "image_path": "2e2280169812212b1d236e2a6a3e66b5ab26bd14574a2bb1d9abdb43d01b8b65.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 321, + 178, + 378, + 217 + ], + "blocks": [ + { + "bbox": [ + 321, + 178, + 378, + 217 + ], + "lines": [ + { + "bbox": [ + 321, + 178, + 378, + 217 + ], + "spans": [ + { + "bbox": [ + 321, + 178, + 378, + 217 + ], + "type": "image", + "image_path": "0823b4da28a4d4e58bba88dec90cc1bb5f4de875f981e5c1269d82ed5df385a6.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 340, + 219, + 359, + 226 + ], + "lines": [ + { + "bbox": [ + 340, + 219, + 359, + 226 + ], + "spans": [ + { + "bbox": [ + 340, + 219, + 359, + 226 + ], + "type": "text", + "content": "BEAT" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 380, + 78, + 440, + 175 + ], + "blocks": [ + { + "bbox": [ + 380, + 78, + 440, + 175 + ], + "lines": [ + { + "bbox": [ + 380, + 78, + 440, + 175 + ], + "spans": [ + { + "bbox": [ + 380, + 78, + 440, + 175 + ], + "type": "image", + "image_path": "0a5a6dee70f74b4a51ebca6acff78622320dfb2cf09a947e6ba026a25c523992.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 380, + 178, + 442, + 217 + ], + "blocks": [ + { + "bbox": [ + 380, + 178, + 442, + 217 + ], + "lines": [ + { + "bbox": [ + 380, + 178, + 442, + 217 + ], + "spans": [ + { + "bbox": [ + 380, + 178, + 442, + 217 + ], + "type": "image", + "image_path": "4e2e0a6d3965907893b8e602a4bd4865e5ecea413958654dd3f083be29da6d78.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 385, + 219, + 438, + 227 + ], + "lines": [ + { + "bbox": [ + 385, + 219, + 438, + 227 + ], + "spans": [ + { + "bbox": [ + 385, + 219, + 438, + 227 + ], + "type": "text", + "content": "Wrapped Optimum" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 444, + 78, + 506, + 175 + ], + "blocks": [ + { + "bbox": [ + 444, + 78, + 506, + 175 + ], + "lines": [ + { + "bbox": [ + 444, + 78, + 506, + 175 + ], + "spans": [ + { + "bbox": [ + 444, + 78, + 506, + 175 + ], + "type": "image", + "image_path": "aa583dca0754a84b1df2c39c7b1147b4db1387c0d69b19fcce8f8d91e9e9466b.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 445, + 178, + 509, + 217 + ], + "blocks": [ + { + "bbox": [ + 445, + 178, + 509, + 217 + ], + "lines": [ + { + "bbox": [ + 445, + 178, + 509, + 217 + ], + "spans": [ + { + "bbox": [ + 445, + 178, + 509, + 217 + ], + "type": "image", + "image_path": "3da42aa9860e70ce3b694a6054783d9f404c5a344355f108f4502a33f7839eda.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 443, + 219, + 517, + 227 + ], + "lines": [ + { + "bbox": [ + 443, + 219, + 517, + 227 + ], + "spans": [ + { + "bbox": [ + 443, + 219, + 517, + 227 + ], + "type": "text", + "content": "Handcraft Optimum (Ours)" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "bbox": [ + 47, + 312, + 208, + 324 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 312, + 208, + 324 + ], + "spans": [ + { + "bbox": [ + 47, + 312, + 208, + 324 + ], + "type": "text", + "content": "3.2. Body Parameters Refinement" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 46, + 335, + 289, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 335, + 289, + 515 + ], + "spans": [ + { + "bbox": [ + 46, + 335, + 289, + 515 + ], + "type": "text", + "content": "MoSh++ produces unnatural head shapes due to the fact that head markers were worn on a helmet and it sometimes produces unnatural finger poses. Consequently, we refine body shape and pose parameters with three simple yet effective hard-coded physical rules. i) The neck and head length should approximate 1/7 of the total length of the body. ii) Fingers, except for the thumb, should not bend backward. iii) We employ the Kolmogorov-Smirnov test, which reveals the data is similar to a normal distribution. Subsequently, we apply a Gaussian truncation approach, where all data points falling outside the " + }, + { + "bbox": [ + 46, + 335, + 289, + 515 + ], + "type": "inline_equation", + "content": "3\\sigma" + }, + { + "bbox": [ + 46, + 335, + 289, + 515 + ], + "type": "text", + "content": " range are adjusted to align with the " + }, + { + "bbox": [ + 46, + 335, + 289, + 515 + ], + "type": "inline_equation", + "content": "3\\sigma" + }, + { + "bbox": [ + 46, + 335, + 289, + 515 + ], + "type": "text", + "content": " threshold and blended with the adjacent 10 frames. We compare our refined body parameters (denoted as MoSh Refined) with Original BEAT, Retargeted SMPL-X, and MoSh SMPL-X in Figure 2." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 47, + 535, + 276, + 548 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 535, + 276, + 548 + ], + "spans": [ + { + "bbox": [ + 47, + 535, + 276, + 548 + ], + "type": "text", + "content": "3.3. Blendshape Weights to FLAME Parameters" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 46, + 557, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 557, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 557, + 289, + 715 + ], + "type": "text", + "content": "Given the ARKit blendshape weights " + }, + { + "bbox": [ + 46, + 557, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{b}_{\\mathrm{ARKit}} \\in \\mathbb{R}^{T \\times 51}" + }, + { + "bbox": [ + 46, + 557, + 289, + 715 + ], + "type": "text", + "content": ", we aim to derive a transformation matrix " + }, + { + "bbox": [ + 46, + 557, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{W} \\in \\mathbb{R}^{51 \\times 103}" + }, + { + "bbox": [ + 46, + 557, + 289, + 715 + ], + "type": "text", + "content": " that maps these into FLAME parameters " + }, + { + "bbox": [ + 46, + 557, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{b}_{\\mathrm{FLAME}} \\in \\mathbb{R}^{T \\times (100 + 3)}" + }, + { + "bbox": [ + 46, + 557, + 289, + 715 + ], + "type": "text", + "content": ", where 100 represents the number of dimensions for expression parameters and 3 for jaw movement. Due to discrepancies between the iPhone's and FLAME's template mesh topology, simply optimizing FLAME parameters by minimizing the geometric error between the ARKit template vertices and the wrapped FLAME vertices does not yield satisfactory results. Instead, we release a set of handcrafted blendshape templates " + }, + { + "bbox": [ + 46, + 557, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_{\\mathrm{template}} \\in \\mathbb{R}^{52}" + }, + { + "bbox": [ + 46, + 557, + 289, + 715 + ], + "type": "text", + "content": " on FLAME following ARKit's FACS configuration. This approach allows direct driving of the FLAME topology ver" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 305, + 311, + 538, + 325 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 311, + 538, + 325 + ], + "spans": [ + { + "bbox": [ + 305, + 311, + 538, + 325 + ], + "type": "text", + "content": "tices " + }, + { + "bbox": [ + 305, + 311, + 538, + 325 + ], + "type": "inline_equation", + "content": "\\mathbf{v}\\in \\mathbb{R}^{T\\times 5023\\times 3}" + }, + { + "bbox": [ + 305, + 311, + 538, + 325 + ], + "type": "text", + "content": " from the given blendshape weights," + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 350, + 335, + 545, + 369 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 350, + 335, + 545, + 369 + ], + "spans": [ + { + "bbox": [ + 350, + 335, + 545, + 369 + ], + "type": "interline_equation", + "content": "\\mathbf {v} = \\mathbf {v} _ {\\text {t e m p l a t e}} ^ {0} + \\sum_ {j = 1} ^ {5 1} \\mathbf {b} _ {\\text {A R K i t}, j} \\cdot \\mathbf {v} _ {\\text {t e m p l a t e}} ^ {j} \\tag {2}", + "image_path": "61de684f215c676e2cdbc6720e9cce83ce43d5ba635241d5c826091efa435686.jpg" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 305, + 378, + 545, + 476 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 378, + 545, + 476 + ], + "spans": [ + { + "bbox": [ + 305, + 378, + 545, + 476 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 378, + 545, + 476 + ], + "type": "inline_equation", + "content": "\\mathbf{b}_{\\mathrm{ARKit},j}" + }, + { + "bbox": [ + 305, + 378, + 545, + 476 + ], + "type": "text", + "content": " is the weight of the " + }, + { + "bbox": [ + 305, + 378, + 545, + 476 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 305, + 378, + 545, + 476 + ], + "type": "text", + "content": "-th ARKit blend-shape and " + }, + { + "bbox": [ + 305, + 378, + 545, + 476 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_{\\mathrm{template}}^j" + }, + { + "bbox": [ + 305, + 378, + 545, + 476 + ], + "type": "text", + "content": " is the vertex position of the FLAME model template corresponding to the " + }, + { + "bbox": [ + 305, + 378, + 545, + 476 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 305, + 378, + 545, + 476 + ], + "type": "text", + "content": "-th blendshape. The term " + }, + { + "bbox": [ + 305, + 378, + 545, + 476 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_{\\mathrm{template}}^0" + }, + { + "bbox": [ + 305, + 378, + 545, + 476 + ], + "type": "text", + "content": " represents the original template vertex position of the FLAME model. We optimize " + }, + { + "bbox": [ + 305, + 378, + 545, + 476 + ], + "type": "inline_equation", + "content": "\\mathbf{W}" + }, + { + "bbox": [ + 305, + 378, + 545, + 476 + ], + "type": "text", + "content": " by minimizing " + }, + { + "bbox": [ + 305, + 378, + 545, + 476 + ], + "type": "inline_equation", + "content": "\\| \\tilde{\\mathbf{v}}_j - \\mathbf{v}_j\\| _2" + }, + { + "bbox": [ + 305, + 378, + 545, + 476 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 305, + 378, + 545, + 476 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{v}}" + }, + { + "bbox": [ + 305, + 378, + 545, + 476 + ], + "type": "text", + "content": " is obtained from FLAME's LBS " + }, + { + "bbox": [ + 305, + 378, + 545, + 476 + ], + "type": "inline_equation", + "content": "\\mathcal{V}(\\mathbf{b}_{\\mathrm{FLAME}})" + }, + { + "bbox": [ + 305, + 378, + 545, + 476 + ], + "type": "text", + "content": ". The comparisons between ARKit data, wrap-based, and our approach are shown in Figure 2." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 306, + 489, + 366, + 501 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 489, + 366, + 501 + ], + "spans": [ + { + "bbox": [ + 306, + 489, + 366, + 501 + ], + "type": "text", + "content": "4. IMAGE" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "text", + "content": "We introduce the details of EMAGE, Expressive Masked Audio-Conditioned GEsture Modeling (see Figure 3). Given gestures " + }, + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{g} \\in \\mathbb{R}^{T \\times (55 \\times 6 + 100 + 4 + 3)}" + }, + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "text", + "content": ", representing 55 joints in Rot6D, " + }, + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^{100}" + }, + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "text", + "content": " FLAME parameters, " + }, + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^4" + }, + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "text", + "content": " foot contact labels, " + }, + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^3" + }, + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "text", + "content": " global translations, and speech audio " + }, + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{a} \\in \\mathbb{R}^{T \\times sk}" + }, + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "inline_equation", + "content": "sk = sr_{\\mathrm{audio}} / fps_{\\mathrm{gestures}}" + }, + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "text", + "content": ", EMAGE jointly optimizes masked gesture reconstruction and audio-conditioned gesture generation. This optimization enhances performance during inference and enables the use of partially masked gestures to complete holistic gestures. To achieve this, we first model the quantized latent space in Sec. 4.1, following [61, 65]. We then design a separated speech audio encoder using Content Rhythm Self-Attention (CRA) as detailed in Sec. 4.2. Subsequently, we learn the body hints from masked gestures via the Masked Audio Gesture Transformer, explained in Sec. 4.3. Finally, the decoding strategy for different body segments is discussed in Sec. 4.4." + } + ] + } + ], + "index": 30 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1147" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 72, + 75, + 531, + 216 + ], + "blocks": [ + { + "bbox": [ + 72, + 75, + 531, + 216 + ], + "lines": [ + { + "bbox": [ + 72, + 75, + 531, + 216 + ], + "spans": [ + { + "bbox": [ + 72, + 75, + 531, + 216 + ], + "type": "image", + "image_path": "03d3d4b036b0ca4357b992552ab291dfb862bda77305f32207eae0b1502e2b37.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 221, + 545, + 299 + ], + "lines": [ + { + "bbox": [ + 46, + 221, + 545, + 299 + ], + "spans": [ + { + "bbox": [ + 46, + 221, + 545, + 299 + ], + "type": "text", + "content": "Figure 3. EMAGE leverages two training paths: Masked Gesture Reconstruction (MG2G) and Audio-Conditioned Gesture Generation (A2G). The MG2G path focuses on encoding robust body hints through a spatial-temporal transformer gesture encoder and cross-attention gesture decoder. In contrast, the A2G path utilizes these body hints, and separated audio encoders to decode pretrained face and body latent features. A key component in this process is a switchable cross-attention layer, crucial for merging body hints and audio features. This fusion allows the features to be effectively disentangled and utilized for gesture decoding. Once the gesture latent features are reconstructed, EMAGE employs a pretrained VQ-Decoder to decode face and local body motions. Additionally, a pretrained Global Motion Predictor is used to estimate global body translations, further enhancing the model's capability to generate realistic and coherent gestures." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 93, + 320, + 166, + 411 + ], + "blocks": [ + { + "bbox": [ + 93, + 320, + 166, + 411 + ], + "lines": [ + { + "bbox": [ + 93, + 320, + 166, + 411 + ], + "spans": [ + { + "bbox": [ + 93, + 320, + 166, + 411 + ], + "type": "image", + "image_path": "fcfbe7a0829c0edac6c9bb4d65277fe8bae5b40a7eca41f562d382b9fffccf6f.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 418, + 287, + 509 + ], + "lines": [ + { + "bbox": [ + 46, + 418, + 287, + 509 + ], + "spans": [ + { + "bbox": [ + 46, + 418, + 287, + 509 + ], + "type": "text", + "content": "Figure 4. Details of CRA and Pretrained VQ-VAEs. Left: Content Rhythm Attention fuses speech rhythm (onset and amplitude) with content (pretrained word embeddings from text scripts) adaptively. This results in a preference for either content or rhythm in specific frames, which encourages the generation of semantical-aware gestures. Right: We pretrain four compositional VQ-VAEs by reconstructing face, upper body, hands and lower body separately to disentangle audio-agnostic gestures explicitly." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 169, + 319, + 242, + 411 + ], + "blocks": [ + { + "bbox": [ + 169, + 319, + 242, + 411 + ], + "lines": [ + { + "bbox": [ + 169, + 319, + 242, + 411 + ], + "spans": [ + { + "bbox": [ + 169, + 319, + 242, + 411 + ], + "type": "image", + "image_path": "20c78e60af2f8a1e4a54cfc9be2acb3cffddc5a758e0ff797f181e0b8b1facf5.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 533, + 278, + 547 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 533, + 278, + 547 + ], + "spans": [ + { + "bbox": [ + 47, + 533, + 278, + 547 + ], + "type": "text", + "content": "4.1. Compositional Discrete Face and Body Prior" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 555, + 286, + 686 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 555, + 286, + 686 + ], + "spans": [ + { + "bbox": [ + 46, + 555, + 286, + 686 + ], + "type": "text", + "content": "We model full-body gestures in the separated quantized latent space (Figure 4) for several reasons. Similar to [65], the body and hands improves the diversity of the results. However, we additionally separate the face and lower body due to their differing correlations with audio, i.e., using a single VQ-VAE to encode both the upper and lower body may lead the model to overlook gestures that occur less frequently - regardless of their relation to the audio. Specifically, a single model might focus only on recovering lower body motion, neglecting elbow movements in a speaker who is constantly walking during the conversation." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 689, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 287, + 715 + ], + "type": "text", + "content": "The separated quantized latent space " + }, + { + "bbox": [ + 47, + 689, + 287, + 715 + ], + "type": "inline_equation", + "content": "Q = \\{\\mathbf{q}_{\\mathrm{f}}, \\mathbf{q}_{u}, \\mathbf{q}_{\\mathrm{h}}, \\mathbf{q}_{\\mathrm{l}}\\}" + }, + { + "bbox": [ + 47, + 689, + 287, + 715 + ], + "type": "text", + "content": " for the Face " + }, + { + "bbox": [ + 47, + 689, + 287, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{g}_{\\mathrm{f}} \\in \\mathbb{R}^{T \\times 106}" + }, + { + "bbox": [ + 47, + 689, + 287, + 715 + ], + "type": "text", + "content": ", Upper" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 319, + 545, + 366 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 319, + 545, + 366 + ], + "spans": [ + { + "bbox": [ + 305, + 319, + 545, + 366 + ], + "type": "text", + "content": "body " + }, + { + "bbox": [ + 305, + 319, + 545, + 366 + ], + "type": "inline_equation", + "content": "\\mathbf{g}_{\\mathrm{u}} \\in \\mathbb{R}^{T \\times 78}" + }, + { + "bbox": [ + 305, + 319, + 545, + 366 + ], + "type": "text", + "content": ", Hands " + }, + { + "bbox": [ + 305, + 319, + 545, + 366 + ], + "type": "inline_equation", + "content": "\\mathbf{g}_{\\mathrm{h}} \\in \\mathbb{R}^{T \\times 180}" + }, + { + "bbox": [ + 305, + 319, + 545, + 366 + ], + "type": "text", + "content": ", and Lower body " + }, + { + "bbox": [ + 305, + 319, + 545, + 366 + ], + "type": "inline_equation", + "content": "\\mathbf{g}_{\\mathrm{l}} \\in \\mathbb{R}^{T \\times (54 + 4)}" + }, + { + "bbox": [ + 305, + 319, + 545, + 366 + ], + "type": "text", + "content": " are from four VQ-VAEs. The VQ-VAE for each is optimized by jointly optimizing the following loss terms," + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 375, + 368, + 545, + 389 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 375, + 368, + 545, + 389 + ], + "spans": [ + { + "bbox": [ + 375, + 368, + 545, + 389 + ], + "type": "interline_equation", + "content": "q _ {i} = \\underset {q _ {i} \\in \\mathbf {q}} {\\arg \\min } \\| z _ {j} - q _ {i} \\| ^ {2} \\tag {3}", + "image_path": "ada54492864ff63ef8b72040b25c5fe89fc87f6d93874e866da01f198fbb3f83.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 318, + 406, + 545, + 437 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 406, + 545, + 437 + ], + "spans": [ + { + "bbox": [ + 318, + 406, + 545, + 437 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} _ {\\mathrm {V Q - V A E}} = \\mathcal {L} _ {r e c} (\\mathbf {g}, \\hat {\\mathbf {g}}) + \\mathcal {L} _ {\\mathrm {v e l}} \\left(\\mathbf {g} ^ {\\prime}, \\hat {\\mathbf {g}} ^ {\\prime}\\right) + \\mathcal {L} _ {\\mathrm {a c c}} \\left(\\mathbf {g} ^ {\\prime \\prime}, \\hat {\\mathbf {g}} ^ {\\prime \\prime}\\right) \\\\ + \\left\\| \\operatorname {s g} [ \\mathbf {z} ] - \\mathbf {q} \\right\\| ^ {2} + \\left\\| \\mathbf {z} - \\operatorname {s g} [ \\mathbf {q} ] \\right\\| ^ {2}, \\tag {4} \\\\ \\end{array}", + "image_path": "d1c60e200107878a5131b6bccca19de70a2ad9afc3ead02cbb483b5c1214113f.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 446, + 545, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 446, + 545, + 495 + ], + "spans": [ + { + "bbox": [ + 305, + 446, + 545, + 495 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 446, + 545, + 495 + ], + "type": "inline_equation", + "content": "\\mathbf{z}" + }, + { + "bbox": [ + 305, + 446, + 545, + 495 + ], + "type": "text", + "content": " is the encoded " + }, + { + "bbox": [ + 305, + 446, + 545, + 495 + ], + "type": "inline_equation", + "content": "\\mathbf{g}" + }, + { + "bbox": [ + 305, + 446, + 545, + 495 + ], + "type": "text", + "content": " with a temporal window size " + }, + { + "bbox": [ + 305, + 446, + 545, + 495 + ], + "type": "inline_equation", + "content": "w = 1" + }, + { + "bbox": [ + 305, + 446, + 545, + 495 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 305, + 446, + 545, + 495 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{rec}}" + }, + { + "bbox": [ + 305, + 446, + 545, + 495 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 305, + 446, + 545, + 495 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{vel}}" + }, + { + "bbox": [ + 305, + 446, + 545, + 495 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 305, + 446, + 545, + 495 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{acc}}" + }, + { + "bbox": [ + 305, + 446, + 545, + 495 + ], + "type": "text", + "content": " are Geodesic [58] and L1 losses, sg is a stop gradient operation, we set the weight of the commitment loss [59] (the last term) to 1 in this paper." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 502, + 496, + 515 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 502, + 496, + 515 + ], + "spans": [ + { + "bbox": [ + 306, + 502, + 496, + 515 + ], + "type": "text", + "content": "4.2. Content and Rhythm Self-Attention" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 521, + 545, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 521, + 545, + 677 + ], + "spans": [ + { + "bbox": [ + 304, + 521, + 545, + 677 + ], + "type": "text", + "content": "Given the speech audio, s, inspired by [4], we employ onset o and amplitude a as explicit audio rhythm, alongside the pretrained embeddings [9] e from transcripts as content. Different from previous approaches, which typically add the rhythm and content features, we leverage self-attention to merge these features adaptively. This approach is driven by the observation that for specific frames, the gestures are more related to content (semantic-aware) or rhythm (beat-aware). The rhythm and content features are first encoded into time-aligned features " + }, + { + "bbox": [ + 304, + 521, + 545, + 677 + ], + "type": "inline_equation", + "content": "\\mathbf{r}_{1:T}" + }, + { + "bbox": [ + 304, + 521, + 545, + 677 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 521, + 545, + 677 + ], + "type": "inline_equation", + "content": "\\mathbf{c}_{1:T}" + }, + { + "bbox": [ + 304, + 521, + 545, + 677 + ], + "type": "text", + "content": ", using a Temporal Convolutional Network (TCN) and linear mapping, respectively. For each time step " + }, + { + "bbox": [ + 304, + 521, + 545, + 677 + ], + "type": "inline_equation", + "content": "t\\in \\{1,\\dots,T\\}" + }, + { + "bbox": [ + 304, + 521, + 545, + 677 + ], + "type": "text", + "content": ", we merge the rhythm and content features by:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 353, + 687, + 545, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 353, + 687, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 353, + 687, + 545, + 715 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathbf {f} _ {1: T} = \\alpha \\times \\mathbf {r} _ {1: T} + (1 - \\alpha) \\times \\mathbf {c} _ {1: T}, \\tag {5} \\\\ \\alpha = \\operatorname {S o f t m a x} (\\mathcal {A T} (\\mathbf {r} _ {1: T}, \\mathbf {c} _ {1: T})), \\\\ \\end{array}", + "image_path": "33d19fc663aae5e0b49510e413a7eef573fec4214420a0b44ae46583ee72b2bf.jpg" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1148" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 52, + 72, + 127, + 186 + ], + "blocks": [ + { + "bbox": [ + 52, + 72, + 127, + 186 + ], + "lines": [ + { + "bbox": [ + 52, + 72, + 127, + 186 + ], + "spans": [ + { + "bbox": [ + 52, + 72, + 127, + 186 + ], + "type": "image", + "image_path": "b5ffedfff937a78fa40fca7cb43e4752e73f12773787d02e95ac5a687967d09f.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 190, + 287, + 266 + ], + "lines": [ + { + "bbox": [ + 46, + 190, + 287, + 266 + ], + "spans": [ + { + "bbox": [ + 46, + 190, + 287, + 266 + ], + "type": "text", + "content": "Figure 5. Comparison of Forward Path Designs. Straightforward fusion module (a) merges audio features without refined body features and recombines audio features based only on position embedding. The Self-Attention decoder module (b), adopted in previous MLM models [17, 31], is limited for tasks requiring auto-regressive inference. Our design (c) considers effective audio feature fusion and auto-regressive decoding." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 130, + 72, + 205, + 186 + ], + "blocks": [ + { + "bbox": [ + 130, + 72, + 205, + 186 + ], + "lines": [ + { + "bbox": [ + 130, + 72, + 205, + 186 + ], + "spans": [ + { + "bbox": [ + 130, + 72, + 205, + 186 + ], + "type": "image", + "image_path": "1cacf29fb5e6624bf438a25165fccad902e7718042aa87ac7732c73e36efd625.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 207, + 72, + 278, + 187 + ], + "blocks": [ + { + "bbox": [ + 207, + 72, + 278, + 187 + ], + "lines": [ + { + "bbox": [ + 207, + 72, + 278, + 187 + ], + "spans": [ + { + "bbox": [ + 207, + 72, + 278, + 187 + ], + "type": "image", + "image_path": "32cfc4eb4ef2edd24fa8beeb6c6b37217aecf0976a8d36ece3aaf17f2d0afc5b.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 286, + 287, + 310 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 286, + 287, + 310 + ], + "spans": [ + { + "bbox": [ + 46, + 286, + 287, + 310 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 286, + 287, + 310 + ], + "type": "inline_equation", + "content": "\\mathcal{A}\\mathcal{T}" + }, + { + "bbox": [ + 46, + 286, + 287, + 310 + ], + "type": "text", + "content": " is a 2-layer MLP. We apply two separate CRA encoders for the face and body." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 316, + 286, + 330 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 316, + 286, + 330 + ], + "spans": [ + { + "bbox": [ + 47, + 316, + 286, + 330 + ], + "type": "text", + "content": "4.3. Masked Audio-Conditioned Gesture Modeling" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 335, + 287, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 335, + 287, + 430 + ], + "spans": [ + { + "bbox": [ + 46, + 335, + 287, + 430 + ], + "type": "text", + "content": "We propose a Masked Audio Gesture Transformer to leverage different training paths (see Fig. 5 for the motivation behind the architecture designs). Given temporal and spatially masked gestures " + }, + { + "bbox": [ + 46, + 335, + 287, + 430 + ], + "type": "inline_equation", + "content": "\\overline{\\mathbf{g}}\\in \\mathbb{R}^{T\\times 337}" + }, + { + "bbox": [ + 46, + 335, + 287, + 430 + ], + "type": "text", + "content": ", we first replace the masked tokens with learned masked embeddings " + }, + { + "bbox": [ + 46, + 335, + 287, + 430 + ], + "type": "inline_equation", + "content": "e_{\\mathrm{mask}}\\in \\mathbb{R}^{256}" + }, + { + "bbox": [ + 46, + 335, + 287, + 430 + ], + "type": "text", + "content": ", as the value zero still represents specific motion content, e.g., a T-pose. We linearly increase the ratio of masked joints and frames from 0 to " + }, + { + "bbox": [ + 46, + 335, + 287, + 430 + ], + "type": "inline_equation", + "content": "95\\%" + }, + { + "bbox": [ + 46, + 335, + 287, + 430 + ], + "type": "text", + "content": " according to the training epochs." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 431, + 287, + 488 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 431, + 287, + 488 + ], + "spans": [ + { + "bbox": [ + 46, + 431, + 287, + 488 + ], + "type": "text", + "content": "A spatial convolutional " + }, + { + "bbox": [ + 46, + 431, + 287, + 488 + ], + "type": "inline_equation", + "content": "\\mathcal{SC}" + }, + { + "bbox": [ + 46, + 431, + 287, + 488 + ], + "type": "text", + "content": " encoder initially summarizes the spatial information, and compress the spatial feature into " + }, + { + "bbox": [ + 46, + 431, + 287, + 488 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^{T\\times 256}" + }, + { + "bbox": [ + 46, + 431, + 287, + 488 + ], + "type": "text", + "content": ". We then employ a temporal self-attention (without feed-forward) " + }, + { + "bbox": [ + 46, + 431, + 287, + 488 + ], + "type": "inline_equation", + "content": "\\mathcal{TSA}" + }, + { + "bbox": [ + 46, + 431, + 287, + 488 + ], + "type": "text", + "content": " to refine the summarized spatial features" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 114, + 490, + 287, + 502 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 114, + 490, + 287, + 502 + ], + "spans": [ + { + "bbox": [ + 114, + 490, + 287, + 502 + ], + "type": "interline_equation", + "content": "\\mathbf {h} = \\mathcal {T S A} (\\mathcal {S C} (\\overline {{\\mathbf {g}}}) + \\mathbf {p} _ {t}), \\tag {6}", + "image_path": "d0d1ead427ec2b3a7a7908da3877ab1ab7805f15fe9a5338e7c77712dc95ba02.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 506, + 287, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 506, + 287, + 567 + ], + "spans": [ + { + "bbox": [ + 46, + 506, + 287, + 567 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 506, + 287, + 567 + ], + "type": "inline_equation", + "content": "\\mathbf{h} \\in \\mathbb{R}^{T \\times 512}" + }, + { + "bbox": [ + 46, + 506, + 287, + 567 + ], + "type": "text", + "content": " represents encoded body hints, and " + }, + { + "bbox": [ + 46, + 506, + 287, + 567 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_t" + }, + { + "bbox": [ + 46, + 506, + 287, + 567 + ], + "type": "text", + "content": " is the sum of a learned speaker embedding and PPE [19]. Then, a straightforward temporal cross-attention Transformer decoder " + }, + { + "bbox": [ + 46, + 506, + 287, + 567 + ], + "type": "inline_equation", + "content": "\\mathcal{TCA}\\mathcal{T}" + }, + { + "bbox": [ + 46, + 506, + 287, + 567 + ], + "type": "text", + "content": " is adopted for the reconstructed latent " + }, + { + "bbox": [ + 46, + 506, + 287, + 567 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{q}}_{\\mathrm{mg2g}}" + }, + { + "bbox": [ + 46, + 506, + 287, + 567 + ], + "type": "text", + "content": "," + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 113, + 567, + 287, + 579 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 113, + 567, + 287, + 579 + ], + "spans": [ + { + "bbox": [ + 113, + 567, + 287, + 579 + ], + "type": "interline_equation", + "content": "\\hat {\\mathbf {q}} _ {\\mathrm {m g} 2 \\mathrm {g}} = \\mathcal {T C A T} (\\mathbf {h} + \\mathbf {p} _ {t}). \\tag {7}", + "image_path": "41861e7f3ca1d4ba0449fd7327d5bd3ee860d97320323937800f401a589bf49e.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 583, + 243, + 595 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 583, + 243, + 595 + ], + "spans": [ + { + "bbox": [ + 47, + 583, + 243, + 595 + ], + "type": "text", + "content": "We minimize the L1 distance in the latent space," + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 120, + 602, + 287, + 616 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 120, + 602, + 287, + 616 + ], + "spans": [ + { + "bbox": [ + 120, + 602, + 287, + 616 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {m g} 2 \\mathrm {g}} = \\| \\hat {\\mathbf {q}} _ {\\mathrm {m g} 2 \\mathrm {g}} - \\mathbf {q} \\|. \\tag {8}", + "image_path": "113a25c15de35aab4037803fc24ecc7d3b0262a9720c4838993b74a4c5e50781.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 46, + 622, + 287, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 622, + 287, + 693 + ], + "spans": [ + { + "bbox": [ + 46, + 622, + 287, + 693 + ], + "type": "text", + "content": "Masked gesture reconstruction encodes effective body hints, with the key being to leverage these body hints for gesture generation. We employ a selective fusion for audio and body hints via a temporal cross-attention " + }, + { + "bbox": [ + 46, + 622, + 287, + 693 + ], + "type": "inline_equation", + "content": "\\mathcal{TCA}" + }, + { + "bbox": [ + 46, + 622, + 287, + 693 + ], + "type": "text", + "content": ". Subsequently, we use the merged audio-gesture features for audio-conditioned gesture latent reconstruction," + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 73, + 701, + 287, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 73, + 701, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 73, + 701, + 287, + 715 + ], + "type": "interline_equation", + "content": "\\hat {\\mathbf {q}} _ {\\mathrm {a} 2 \\mathrm {g}} = \\mathcal {T C A T} \\left(\\mathcal {T C A} \\left(\\mathbf {h} + \\mathbf {p} _ {t}, \\mathbf {f} _ {\\text {b o d y}}\\right), \\overline {{\\mathbf {g}}} + \\mathbf {p} _ {t}\\right). \\tag {9}", + "image_path": "b971fda6f0a17e850f7d833b4221541370eea3585e9f386590deedf3c002cac0.jpg" + } + ] + } + ], + "index": 14 + }, + { + "type": "table", + "bbox": [ + 320, + 70, + 531, + 114 + ], + "blocks": [ + { + "bbox": [ + 320, + 70, + 531, + 114 + ], + "lines": [ + { + "bbox": [ + 320, + 70, + 531, + 114 + ], + "spans": [ + { + "bbox": [ + 320, + 70, + 531, + 114 + ], + "type": "table", + "html": "
Top.-BTop.-FBodyFace
VOCA[14]-FLAME-38.3 ± 5.63
AMASS[46]SMPL-XFLAME42.0 ± 3.60-
TalkSHOW[65]SMPL-XSMPL-X14.4 ± 2.1926.1 ± 6.42
BEAT2 (Ours)SMPL-XFLAME43.6 ± 3.3835.7 ± 5.91
", + "image_path": "1bd2e05a0ddf11f82d1922cf9aa70f0429d1d52a384c056486df192d8633e7ff.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "table_body" + } + ], + "index": 15 + }, + { + "bbox": [ + 305, + 119, + 545, + 185 + ], + "lines": [ + { + "bbox": [ + 305, + 119, + 545, + 185 + ], + "spans": [ + { + "bbox": [ + 305, + 119, + 545, + 185 + ], + "type": "text", + "content": "Table 3. User Preference Win Rate between Existing Datasets. 'Top' denotes the topology of the mesh. The results show our BEAT2 dataset outperforms the existing PGT dataset [65] in both body and face aspects. It also performs slightly better on body and lower on face when compared with previous mocap [46] and 3D scan-based datasets [14], respectively." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 305, + 199, + 545, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 199, + 545, + 236 + ], + "spans": [ + { + "bbox": [ + 305, + 199, + 545, + 236 + ], + "type": "text", + "content": "We optimize both the latent code class classification cross-entropy loss " + }, + { + "bbox": [ + 305, + 199, + 545, + 236 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{a2g - rec}}" + }, + { + "bbox": [ + 305, + 199, + 545, + 236 + ], + "type": "text", + "content": " and the latent reconstruction loss " + }, + { + "bbox": [ + 305, + 199, + 545, + 236 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{a2g - c1s}}" + }, + { + "bbox": [ + 305, + 199, + 545, + 236 + ], + "type": "text", + "content": " to encourage diverse results." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 306, + 242, + 474, + 255 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 242, + 474, + 255 + ], + "spans": [ + { + "bbox": [ + 306, + 242, + 474, + 255 + ], + "type": "text", + "content": "4.4. Face and Translation Decoding" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 305, + 260, + 545, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 260, + 545, + 319 + ], + "spans": [ + { + "bbox": [ + 305, + 260, + 545, + 319 + ], + "type": "text", + "content": "Considering that the face is weakly related to body motion, applying the same operation - recombining the audio features based on body hints - is not reasonable. Therefore, we directly concatenate the masked body hints with the audio features for the final decoding of facial latent," + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 367, + 327, + 545, + 340 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 367, + 327, + 545, + 340 + ], + "spans": [ + { + "bbox": [ + 367, + 327, + 545, + 340 + ], + "type": "interline_equation", + "content": "\\hat {\\mathbf {q}} _ {\\mathrm {f}} = \\mathcal {T} \\mathcal {C A T} \\left(\\mathbf {f} _ {\\text {f a c e}} \\oplus \\mathbf {h}, \\mathbf {p} _ {t}\\right). \\tag {10}", + "image_path": "6fca1ef6e2cbdec72c70fd2f3e29f95270aba498b20ea008669d9a0ebf7cf5cb.jpg" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 305, + 347, + 545, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 347, + 545, + 396 + ], + "spans": [ + { + "bbox": [ + 305, + 347, + 545, + 396 + ], + "type": "text", + "content": "Once we have obtained the local lower body motion " + }, + { + "bbox": [ + 305, + 347, + 545, + 396 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{g}}_l\\in" + }, + { + "bbox": [ + 305, + 347, + 545, + 396 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^{T\\times (54 + 4)}" + }, + { + "bbox": [ + 305, + 347, + 545, + 396 + ], + "type": "text", + "content": " from the VQ-Decoder, we estimate the global translations " + }, + { + "bbox": [ + 305, + 347, + 545, + 396 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{t}}\\in \\mathbb{R}^{T\\times 3}" + }, + { + "bbox": [ + 305, + 347, + 545, + 396 + ], + "type": "text", + "content": " with a pretrained Global Motion Predictor " + }, + { + "bbox": [ + 305, + 347, + 545, + 396 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{t}} = \\mathcal{G}(\\tilde{\\mathbf{g}}_l)" + }, + { + "bbox": [ + 305, + 347, + 545, + 396 + ], + "type": "text", + "content": ", which significantly reduces foot sliding." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 306, + 405, + 386, + 418 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 405, + 386, + 418 + ], + "spans": [ + { + "bbox": [ + 306, + 405, + 386, + 418 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 304, + 426, + 545, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 426, + 545, + 545 + ], + "spans": [ + { + "bbox": [ + 304, + 426, + 545, + 545 + ], + "type": "text", + "content": "We separate the evaluation into two categories: dataset quality and model ability. After removing data with low finger quality from BEAT, BEAT2 is reduced to 60 hours. We further split it into BEAT2-Standard (27 hours) and BEAT2-Additional (33 hours), based on the type of speech and conversation sections [39]. While acted gestures in speech sections are more diverse and expressive, spontaneous gestures in conversational sections tend to be more natural yet less varied. We report the results on BEAT2-Standard Speaker2 with an " + }, + { + "bbox": [ + 304, + 426, + 545, + 545 + ], + "type": "inline_equation", + "content": "85\\% / 7.5\\% / 7.5\\%" + }, + { + "bbox": [ + 304, + 426, + 545, + 545 + ], + "type": "text", + "content": " train/val/test split." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 306, + 552, + 455, + 564 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 552, + 455, + 564 + ], + "spans": [ + { + "bbox": [ + 306, + 552, + 455, + 564 + ], + "type": "text", + "content": "5.1. Dataset Quality Evaluation" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "text", + "content": "We compare our dataset with the state-of-the-art pseudo ground truth (PGT) dataset, TalkSHOW [65], for both face and body. Additionally, we compare it with the AMASS [46] dataset for the body and VOCA [14] for the face as references. Due to the varying lengths of sequences in each dataset, we sample 100 comparison pairs, each with an equal duration ranging from 2 to 4 seconds. In a perceptual study, each participant evaluates a random set of 40 pairs during a 10-minute session by selecting the sequence they consider to have the best captured quality. In total, 60 participants were invited. It is important to note that participants are instructed to compare only the upper body results," + } + ] + } + ], + "index": 25 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1149" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 84, + 89, + 233, + 100 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 84, + 89, + 233, + 100 + ], + "spans": [ + { + "bbox": [ + 84, + 89, + 233, + 100 + ], + "type": "text", + "content": "... In my spare time ... ... if I feel ok ..." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 253, + 89, + 508, + 101 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 253, + 89, + 508, + 101 + ], + "spans": [ + { + "bbox": [ + 253, + 89, + 508, + 101 + ], + "type": "text", + "content": "... I will go for a walk or hike in nature. I try to organize something" + } + ] + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 83, + 103, + 157, + 158 + ], + "blocks": [ + { + "bbox": [ + 83, + 103, + 157, + 158 + ], + "lines": [ + { + "bbox": [ + 83, + 103, + 157, + 158 + ], + "spans": [ + { + "bbox": [ + 83, + 103, + 157, + 158 + ], + "type": "image", + "image_path": "2c125761dd9cae7c4576e912ae7f448a11376383f3058ef27dd260925960ebc0.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 170, + 103, + 242, + 158 + ], + "blocks": [ + { + "bbox": [ + 170, + 103, + 242, + 158 + ], + "lines": [ + { + "bbox": [ + 170, + 103, + 242, + 158 + ], + "spans": [ + { + "bbox": [ + 170, + 103, + 242, + 158 + ], + "type": "image", + "image_path": "cdb9bcf28bebd72500a014eecdaa75e44c0ee0d34d340bf78adbc004c8817654.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 256, + 103, + 330, + 158 + ], + "blocks": [ + { + "bbox": [ + 256, + 103, + 330, + 158 + ], + "lines": [ + { + "bbox": [ + 256, + 103, + 330, + 158 + ], + "spans": [ + { + "bbox": [ + 256, + 103, + 330, + 158 + ], + "type": "image", + "image_path": "bae710f5cf209e06c839b96c23eb6e0f4e8f3ae177079db6f6be7e4871ae7a24.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 343, + 104, + 416, + 158 + ], + "blocks": [ + { + "bbox": [ + 343, + 104, + 416, + 158 + ], + "lines": [ + { + "bbox": [ + 343, + 104, + 416, + 158 + ], + "spans": [ + { + "bbox": [ + 343, + 104, + 416, + 158 + ], + "type": "image", + "image_path": "c1c57afefb94e35ad11b6de08d396967a3fa5faca4bb4d28920f253113b08c23.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 433, + 105, + 505, + 158 + ], + "blocks": [ + { + "bbox": [ + 433, + 105, + 505, + 158 + ], + "lines": [ + { + "bbox": [ + 433, + 105, + 505, + 158 + ], + "spans": [ + { + "bbox": [ + 433, + 105, + 505, + 158 + ], + "type": "image", + "image_path": "428f0f0053095cf331f75a2ee392f3922abd1e4ccefa68d335d26cba72ee5c79.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 80, + 159, + 166, + 217 + ], + "blocks": [ + { + "bbox": [ + 80, + 159, + 166, + 217 + ], + "lines": [ + { + "bbox": [ + 80, + 159, + 166, + 217 + ], + "spans": [ + { + "bbox": [ + 80, + 159, + 166, + 217 + ], + "type": "image", + "image_path": "0e571774cfd0d6a43f3e034ffbc64ca28c24d7885bf47b0ac76ba27000ba1cb3.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 174, + 159, + 242, + 216 + ], + "blocks": [ + { + "bbox": [ + 174, + 159, + 242, + 216 + ], + "lines": [ + { + "bbox": [ + 174, + 159, + 242, + 216 + ], + "spans": [ + { + "bbox": [ + 174, + 159, + 242, + 216 + ], + "type": "image", + "image_path": "6ac6fc9288d18bedb3545a7e62f6007ef4f7c006b15af8b108bf094d1ad7880d.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 256, + 159, + 330, + 217 + ], + "blocks": [ + { + "bbox": [ + 256, + 159, + 330, + 217 + ], + "lines": [ + { + "bbox": [ + 256, + 159, + 330, + 217 + ], + "spans": [ + { + "bbox": [ + 256, + 159, + 330, + 217 + ], + "type": "image", + "image_path": "5f6862fca91f66b90bd24f08769f95c06dfa042121411312abe32d886c6e6098.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 343, + 159, + 414, + 217 + ], + "blocks": [ + { + "bbox": [ + 343, + 159, + 414, + 217 + ], + "lines": [ + { + "bbox": [ + 343, + 159, + 414, + 217 + ], + "spans": [ + { + "bbox": [ + 343, + 159, + 414, + 217 + ], + "type": "image", + "image_path": "331dc7ed8f3b856ee1e11e203e33219dc639d6472ce45416330386342a33996a.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 427, + 160, + 501, + 217 + ], + "blocks": [ + { + "bbox": [ + 427, + 160, + 501, + 217 + ], + "lines": [ + { + "bbox": [ + 427, + 160, + 501, + 217 + ], + "spans": [ + { + "bbox": [ + 427, + 160, + 501, + 217 + ], + "type": "image", + "image_path": "fd3a413577dc8a0433bca6a0d076bf3a615e42cf9e70297b2b0c83b45ad34d34.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 84, + 219, + 154, + 274 + ], + "blocks": [ + { + "bbox": [ + 84, + 219, + 154, + 274 + ], + "lines": [ + { + "bbox": [ + 84, + 219, + 154, + 274 + ], + "spans": [ + { + "bbox": [ + 84, + 219, + 154, + 274 + ], + "type": "image", + "image_path": "d5f6d5f60c872172245a54c48b4f67219ea87e02dbdecbd6b9a8a9f105928657.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 172, + 218, + 239, + 274 + ], + "blocks": [ + { + "bbox": [ + 172, + 218, + 239, + 274 + ], + "lines": [ + { + "bbox": [ + 172, + 218, + 239, + 274 + ], + "spans": [ + { + "bbox": [ + 172, + 218, + 239, + 274 + ], + "type": "image", + "image_path": "50f47b1881a4f057c10224c1452d9152a3706b0b2443921ac8022fe89d6e58c5.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 255, + 218, + 328, + 274 + ], + "blocks": [ + { + "bbox": [ + 255, + 218, + 328, + 274 + ], + "lines": [ + { + "bbox": [ + 255, + 218, + 328, + 274 + ], + "spans": [ + { + "bbox": [ + 255, + 218, + 328, + 274 + ], + "type": "image", + "image_path": "783ca256b44282fda31bf3021b3fb18df2884846cbacbae94c1c54303861afca.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 336, + 217, + 414, + 274 + ], + "blocks": [ + { + "bbox": [ + 336, + 217, + 414, + 274 + ], + "lines": [ + { + "bbox": [ + 336, + 217, + 414, + 274 + ], + "spans": [ + { + "bbox": [ + 336, + 217, + 414, + 274 + ], + "type": "image", + "image_path": "7e6a2efeb4f1b1f039e4b4103e35d22570d17ac002aa7b76b5737e7ad19b9a58.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 425, + 217, + 505, + 275 + ], + "blocks": [ + { + "bbox": [ + 425, + 217, + 505, + 275 + ], + "lines": [ + { + "bbox": [ + 425, + 217, + 505, + 275 + ], + "spans": [ + { + "bbox": [ + 425, + 217, + 505, + 275 + ], + "type": "image", + "image_path": "3b5bb70b6c5c3fbf791f74724eb0bfd3ce4c78419828d7f4465313421d034016.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 81, + 276, + 159, + 329 + ], + "blocks": [ + { + "bbox": [ + 81, + 276, + 159, + 329 + ], + "lines": [ + { + "bbox": [ + 81, + 276, + 159, + 329 + ], + "spans": [ + { + "bbox": [ + 81, + 276, + 159, + 329 + ], + "type": "image", + "image_path": "6870b2a16ab908948f4c9a29bfe7b7a29afcad6abcba862bc36131580f62c209.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 338, + 547, + 393 + ], + "lines": [ + { + "bbox": [ + 46, + 338, + 547, + 393 + ], + "spans": [ + { + "bbox": [ + 46, + 338, + 547, + 393 + ], + "type": "text", + "content": "Figure 6. Subjective Examples. Top to down: GroundTruth, Generated W/O Body Hints, Generated With Body Hints, Visible Body Hints, EMAGE generates diverse, semantic-aware and audio synchronized body gestures, e.g. raise the both hands for \"spare time\", relaxing gestures for \"hike in nature\". Besides, as shown in the third and bottom rows, EMAGE is flexible to accept non-audio-synchronized body hints on any frames and joints, to guide or customize the generated gestures explicitly, e.g., repeat a similar motion for raising hands, change the walking orientations, etc. Note that in this figure, The gray joints of generated results are not copies of visible hints." + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 170, + 276, + 242, + 329 + ], + "blocks": [ + { + "bbox": [ + 170, + 276, + 242, + 329 + ], + "lines": [ + { + "bbox": [ + 170, + 276, + 242, + 329 + ], + "spans": [ + { + "bbox": [ + 170, + 276, + 242, + 329 + ], + "type": "image", + "image_path": "6d0dd81e175b8e70e7d2cbaa3207c1c29eec10056fee93189c6f6b8833554f74.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 255, + 275, + 328, + 329 + ], + "blocks": [ + { + "bbox": [ + 255, + 275, + 328, + 329 + ], + "lines": [ + { + "bbox": [ + 255, + 275, + 328, + 329 + ], + "spans": [ + { + "bbox": [ + 255, + 275, + 328, + 329 + ], + "type": "image", + "image_path": "fd6f3eee973bb1ecc58436590341bf1e98b4a7419d4e84fc2f7ed5c9472f580e.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 337, + 274, + 406, + 334 + ], + "blocks": [ + { + "bbox": [ + 337, + 274, + 406, + 334 + ], + "lines": [ + { + "bbox": [ + 337, + 274, + 406, + 334 + ], + "spans": [ + { + "bbox": [ + 337, + 274, + 406, + 334 + ], + "type": "image", + "image_path": "fe89c02a59451a9e1f4d73c03f7d1acf521b5838104d33ec8fb5dc3acbf675a8.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 424, + 276, + 509, + 333 + ], + "blocks": [ + { + "bbox": [ + 424, + 276, + 509, + 333 + ], + "lines": [ + { + "bbox": [ + 424, + 276, + 509, + 333 + ], + "spans": [ + { + "bbox": [ + 424, + 276, + 509, + 333 + ], + "type": "image", + "image_path": "87fdaad99801ac186d6b6a5c6f959383750e13087ad63f5e0cb766c482d4920a.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + } + ], + "index": 21 + }, + { + "type": "table", + "bbox": [ + 50, + 410, + 285, + 524 + ], + "blocks": [ + { + "bbox": [ + 50, + 410, + 285, + 524 + ], + "lines": [ + { + "bbox": [ + 50, + 410, + 285, + 524 + ], + "spans": [ + { + "bbox": [ + 50, + 410, + 285, + 524 + ], + "type": "table", + "html": "
FGD ↓BC ↑Diversity ↑MSE ↓LVD ↓
FaceFormer[19]---7.7877.593
CodeTalker[61]---8.0267.766
S2G[25]28.154.6835.971--
Trimodal[67]12.415.9337.724--
HA2G[40]12.326.7798.626--
DisCo[38]9.4176.4399.912--
CaMN[39]6.6446.76910.86--
DiffStyleGesture[62]8.8117.24111.49--
Habibie et al.[28]9.0407.7168.2138.6148.043
TalkSHOW[65]6.2096.94713.477.7917.771
EMAGE (Ours)5.5127.72413.067.6807.556
", + "image_path": "faacea25f48cb1d51502542f3db4f1848e16a4bbb87af4b0c969e82300ef07f1.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "table_body" + } + ], + "index": 23 + }, + { + "bbox": [ + 47, + 597, + 287, + 621 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 597, + 287, + 621 + ], + "spans": [ + { + "bbox": [ + 47, + 597, + 287, + 621 + ], + "type": "text", + "content": "as TalkSHOW contains only the upper body. The results are shown in Table 3." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 47, + 630, + 188, + 643 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 630, + 188, + 643 + ], + "spans": [ + { + "bbox": [ + 47, + 630, + 188, + 643 + ], + "type": "text", + "content": "5.2. Model Ability Evaluation" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 47, + 649, + 287, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 649, + 287, + 685 + ], + "spans": [ + { + "bbox": [ + 47, + 649, + 287, + 685 + ], + "type": "text", + "content": "We report results using BEATv1.3 and BEATv2. The latter has facial data refined by animators and hand data filtered by annotators." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "type": "text", + "content": "Metrics. We adopt FGD[67] to evaluate the realism of the body gestures. Then, we measure Diversity[33] by calcu" + } + ] + } + ], + "index": 28 + }, + { + "type": "table", + "bbox": [ + 325, + 411, + 526, + 449 + ], + "blocks": [ + { + "bbox": [ + 47, + 526, + 287, + 582 + ], + "lines": [ + { + "bbox": [ + 47, + 526, + 287, + 582 + ], + "spans": [ + { + "bbox": [ + 47, + 526, + 287, + 582 + ], + "type": "text", + "content": "Table 4. Quantitative evaluation on BEATv2. We report FGD " + }, + { + "bbox": [ + 47, + 526, + 287, + 582 + ], + "type": "inline_equation", + "content": "\\times 10^{-1}" + }, + { + "bbox": [ + 47, + 526, + 287, + 582 + ], + "type": "text", + "content": ", BC " + }, + { + "bbox": [ + 47, + 526, + 287, + 582 + ], + "type": "inline_equation", + "content": "\\times 10^{-1}" + }, + { + "bbox": [ + 47, + 526, + 287, + 582 + ], + "type": "text", + "content": ", Diversity, MSE " + }, + { + "bbox": [ + 47, + 526, + 287, + 582 + ], + "type": "inline_equation", + "content": "\\times 10^{-8}" + }, + { + "bbox": [ + 47, + 526, + 287, + 582 + ], + "type": "text", + "content": ", and LVD " + }, + { + "bbox": [ + 47, + 526, + 287, + 582 + ], + "type": "inline_equation", + "content": "\\times 10^{-5}" + }, + { + "bbox": [ + 47, + 526, + 287, + 582 + ], + "type": "text", + "content": ". For body gestures, EMAGE significantly improves the FGD, indicating that the generated results are closer to the GT. This shows the effectiveness of body hints from masked gesture modeling." + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 325, + 411, + 526, + 449 + ], + "lines": [ + { + "bbox": [ + 325, + 411, + 526, + 449 + ], + "spans": [ + { + "bbox": [ + 325, + 411, + 526, + 449 + ], + "type": "table", + "html": "
HolisticBodyFace
Habibie et al.[28]12.4 ± 3.7015.9 ± 6.4910.8 ± 3.19
TalkSHOW[65]34.9 ± 5.7940.4 ± 8.2233.2 ± 6.03
EMAGE (Ours)52.7 ± 7.9144.7 ± 8.6856.0 ± 7.80
", + "image_path": "897c87d22557e913147a4e036f9b9610a58050783ae0433b85d3687dcc096855.jpg" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "table_body" + } + ], + "index": 29 + }, + { + "bbox": [ + 305, + 453, + 545, + 498 + ], + "lines": [ + { + "bbox": [ + 305, + 453, + 545, + 498 + ], + "spans": [ + { + "bbox": [ + 305, + 453, + 545, + 498 + ], + "type": "text", + "content": "Table 5. User Preference Win Rate On Generated Results. The results indicate that our generated outcomes are perceived as more realistic and believable, with a " + }, + { + "bbox": [ + 305, + 453, + 545, + 498 + ], + "type": "inline_equation", + "content": "14\\%" + }, + { + "bbox": [ + 305, + 453, + 545, + 498 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 453, + 545, + 498 + ], + "type": "inline_equation", + "content": "23\\%" + }, + { + "bbox": [ + 305, + 453, + 545, + 498 + ], + "type": "text", + "content": " higher user preference for body and face gestures, respectively." + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 305, + 531, + 545, + 591 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 531, + 545, + 591 + ], + "spans": [ + { + "bbox": [ + 305, + 531, + 545, + 591 + ], + "type": "text", + "content": "lating the average L1 distance between multiple body gesture clips, and use BC [34] to assess speech-motion synchrony. For the face, we calculate the vertex MSE[61] to measure the positional distance and the vertex L1 difference LVD [65] between the GT and generated facial vertices." + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 304, + 612, + 546, + 719 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 612, + 546, + 719 + ], + "spans": [ + { + "bbox": [ + 304, + 612, + 546, + 719 + ], + "type": "text", + "content": "Compared Methods. We first compare our method with representative state-of-the-art methods in body gesture generation [25, 38-40, 62, 67] and talking head generation [19, 61] by reproducing their methods for body and face, respectively. In addition, we reproduce two previous state-of-the-art holistic pipelines, Habibie et al. [28] and TalkSHOW [65], whose original implementations are limited to the upper body. We add a lower-body decoder to Habibie et al. and a lower-body VQ-VAE to TalkSHOW." + } + ] + } + ], + "index": 32 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "1150" + } + ] + } + ], + "index": 33 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 69, + 267, + 203 + ], + "blocks": [ + { + "bbox": [ + 59, + 69, + 267, + 203 + ], + "lines": [ + { + "bbox": [ + 59, + 69, + 267, + 203 + ], + "spans": [ + { + "bbox": [ + 59, + 69, + 267, + 203 + ], + "type": "image", + "image_path": "ec4acad0cc10bb45d1f30e78a3cac08c72cd08e88bef35a1db35c54e8a1d34f8.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 206, + 289, + 294 + ], + "lines": [ + { + "bbox": [ + 46, + 206, + 289, + 294 + ], + "spans": [ + { + "bbox": [ + 46, + 206, + 289, + 294 + ], + "type": "text", + "content": "Figure 7. Comparison of Generated Facial Movements. Compared with previous state-of-the-art talking face generation methods FaceFormer [19] and CodeTalker [61] as well as holistic gesture generation methods Habibie et al. [28] and TalkSHOW [65]. Note that CodeTalker has higher MSE than IMAGE on BEATv2 (Table 4, lower is better) but is subjectively realistic. IMAGE gets good lip motions by leveraging both the face model and masked body hints." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 304, + 241, + 316 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 304, + 241, + 316 + ], + "spans": [ + { + "bbox": [ + 47, + 304, + 241, + 316 + ], + "type": "text", + "content": "5.2.1 Quantitative and Qualitative Analysis" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 318, + 288, + 401 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 318, + 288, + 401 + ], + "spans": [ + { + "bbox": [ + 46, + 318, + 288, + 401 + ], + "type": "text", + "content": "As shown in Table 4, with a four-frame seed pose, our method outperforms previous SOTA algorithms. For qualitative results, see Figures 6 and 7. Furthermore, we conduct a perceptual study. Maintaining the same setup of 60 participants, each participant evaluates 40 pairs of 10-second results do decide which is most believable; this gives the win rate shown in Table 5." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 418, + 157, + 430 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 418, + 157, + 430 + ], + "spans": [ + { + "bbox": [ + 47, + 418, + 157, + 430 + ], + "type": "text", + "content": "5.2.2 Ablation Analysis" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 437, + 287, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 437, + 287, + 510 + ], + "spans": [ + { + "bbox": [ + 46, + 437, + 287, + 510 + ], + "type": "text", + "content": "Performance of Baseline. As shown in Table 6, we start with a teacher-force transformer-based baseline. This baseline, inspired by FaceFormer [19], adopts a 1-layer cross-attention transformer decoder and replaces the audio features from Wav2Vec2 [6] with our custom TCN [7] and trainable word embeddings [9]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 510, + 288, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 510, + 288, + 605 + ], + "spans": [ + { + "bbox": [ + 46, + 510, + 288, + 605 + ], + "type": "text", + "content": "Effect of Multiple VQ-VAEs. Simply applying one VQ-VAE [27, 59] for full-body movements, including the face, decreases performance in facial movements. This is because the VQ-VAE is trained to minimize the average loss for the full body, and some speakers' most frequent movements are not related to the audio. Implementing separated VQ-VAEs allows the model to better leverage the advantages of discrete priors." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 605, + 287, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 287, + 677 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 287, + 677 + ], + "type": "text", + "content": "Effect of Content Rhythm Self-Attention. Adaptive fusion of rhythm and content features shows improvement in both FGD and Alignment. It selectively focuses the current motion more on rhythm or content features based on the training data's distribution. Moreover, we observed more semantic-aware results when applying CRA." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 677, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 288, + 714 + ], + "type": "text", + "content": "Effect of Masked Body Gesture Hints. The improvement across all objective metrics demonstrates that our model effectively leverages spatial-temporal gesture priors, reducing" + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 310, + 69, + 542, + 153 + ], + "blocks": [ + { + "bbox": [ + 310, + 69, + 542, + 153 + ], + "lines": [ + { + "bbox": [ + 310, + 69, + 542, + 153 + ], + "spans": [ + { + "bbox": [ + 310, + 69, + 542, + 153 + ], + "type": "table", + "html": "
FGD ↓BC ↑Diversity ↑MSE ↓LVD ↓
Ground Truth06.89613.07400
Reconstruction3.9136.75813.1450.8416.389
Baseline13.0806.9418.31451.4429.317
+ VQVAE9.7876.67310.6241.6199.473
+ 4 VQVAE7.3976.69812.5441.2438.938
+ CRA6.8336.78312.6761.1868.715
+ Masked Hints5.4236.79413.0571.1809.015
", + "image_path": "b5e3474bf149b65160747babbc2213f218c7780bfb8d61b4f14ff101b36b7dd3.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 319, + 183, + 533, + 223 + ], + "blocks": [ + { + "bbox": [ + 347, + 157, + 505, + 167 + ], + "lines": [ + { + "bbox": [ + 347, + 157, + 505, + 167 + ], + "spans": [ + { + "bbox": [ + 347, + 157, + 505, + 167 + ], + "type": "text", + "content": "Table 6. Abliation Analysis on BEATv1.3" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 319, + 183, + 533, + 223 + ], + "lines": [ + { + "bbox": [ + 319, + 183, + 533, + 223 + ], + "spans": [ + { + "bbox": [ + 319, + 183, + 533, + 223 + ], + "type": "table", + "html": "
bodyhandsfaceFGD ↓BC ↑Diversity ↑
BESTX✓✓✓5.4236.79413.057
+ Trinity[22]✓5.3196.84313.346
+ AMASS[46]✓✓5.1746.76914.318
", + "image_path": "3082b043011fb1ad1fe944210d5da70299672d88000c5ade9d11baf995a271b7.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 226, + 545, + 270 + ], + "lines": [ + { + "bbox": [ + 305, + 226, + 545, + 270 + ], + "spans": [ + { + "bbox": [ + 305, + 226, + 545, + 270 + ], + "type": "text", + "content": "Table 7. Training IMAGE on Multiple Datasets. IMAGE demonstrates flexibility by training on multiple datasets, even when only a subset of holistic gestures is available. This approach further improves the objective metrics on the BEATv1.3 test set." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 305, + 287, + 545, + 324 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 287, + 545, + 324 + ], + "spans": [ + { + "bbox": [ + 305, + 287, + 545, + 324 + ], + "type": "text", + "content": "the likelihood of incorrect gesture sampling. Furthermore, masked gesture modeling is key to enabling the network to accept predefined gestures in specific frames." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 342, + 483, + 355 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 342, + 483, + 355 + ], + "spans": [ + { + "bbox": [ + 306, + 342, + 483, + 355 + ], + "type": "text", + "content": "5.2.3 Ability of Multi-Dataset Training." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 365, + 545, + 473 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 365, + 545, + 473 + ], + "spans": [ + { + "bbox": [ + 304, + 365, + 545, + 473 + ], + "type": "text", + "content": "We demonstrate that EMAGE can effectively combine multiple non-holistic datasets for training by jointly training with the Trinity [22] and AMASS [46] datasets, by using only the upper body and audio pairs from Trinity and only body and hands from AMASS. Separate VQ-VAEs are trained for BEAT2, Trinity, and AMASS, with separate MLP heads implemented for codebook classification. The results in Table 7 show that incorporating data improves performance on the BEATv1.3 test set." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 485, + 378, + 498 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 485, + 378, + 498 + ], + "spans": [ + { + "bbox": [ + 306, + 485, + 378, + 498 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 508, + 545, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 508, + 545, + 639 + ], + "spans": [ + { + "bbox": [ + 304, + 508, + 545, + 639 + ], + "type": "text", + "content": "In this work, we present IMAGE, a framework designed to accept partial gestures as input for completing audio-synchronized holistic gestures. It demonstrates that leveraging masked gesture reconstruction can significantly enhance the performance of audio-conditioned gesture generation. Furthermore, the design of IMAGE enables the training on multiple datasets, further improving performance. Alongside IMAGE, we release BEAT2, the largest multi-modal gesture dataset consistent with SMPL-X and FLAME. We hope that BEAT2 will contribute to knowledge and model sharing across various subareas." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 641, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 641, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 641, + 545, + 712 + ], + "type": "text", + "content": "Disclosures: While You Zhou, Xuefei Zhe, Naoya Iwamoto, and Bo Zheng are employees of Huawei Tokyo Research Center, this work was done on their own time with the approval of their employer. MJB CoI: https://files.is.tue.mpg.de/black/CoI_CVPR_2024.txt" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 314, + 757 + ], + "type": "text", + "content": "1151" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 134 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 134 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 134 + ], + "type": "text", + "content": "[1] Kfir Aberman, Peizhuo Li, Dani Lischinski, Olga Sorkine-Hornung, Daniel Cohen-Or, and Baoquan Chen. Skeleton-aware networks for deep motion retargeting. ACM Transactions on Graphics (TOG), 39(4):62-1, 2020. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 135, + 288, + 190 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 135, + 288, + 190 + ], + "spans": [ + { + "bbox": [ + 53, + 135, + 288, + 190 + ], + "type": "text", + "content": "[2] Chaitanya Ahuja, Dong Won Lee, Yukiko I Nakano, and Louis-Philippe Morency. Style transfer for co-speech gesture animation: A multi-speaker conditional-mixture approach. In European Conference on Computer Vision, pages 248–265. Springer, 2020. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 191, + 287, + 235 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 191, + 287, + 235 + ], + "spans": [ + { + "bbox": [ + 53, + 191, + 287, + 235 + ], + "type": "text", + "content": "[3] Simon Alexanderson, Gustav Eje Henter, Taras Kucherenko, and Jonas Beskow. Style-controllable speech-driven gesture synthesis using normalising flows. In Computer Graphics Forum, pages 487-496. Wiley Online Library, 2020. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 236, + 286, + 267 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 236, + 286, + 267 + ], + "spans": [ + { + "bbox": [ + 53, + 236, + 286, + 267 + ], + "type": "text", + "content": "[4] Tenglong Ao, Zeyi Zhang, and Libin Liu. Gesture diffu-clip: Gesture diffusion model with clip latents. ACM Trans. Graph. 2, 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 269, + 287, + 313 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 269, + 287, + 313 + ], + "spans": [ + { + "bbox": [ + 53, + 269, + 287, + 313 + ], + "type": "text", + "content": "[5] Tenglong Ao, Qingzhe Gao, Yuke Lou, Baoquan Chen, and Libin Liu. Rhythmic gesticulator: Rhythm-aware co-speech gesture synthesis with hierarchical neural embeddings. ACM Transactions on Graphics (TOG), 41(6):1-19, 2022. 2, 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 314, + 286, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 314, + 286, + 357 + ], + "spans": [ + { + "bbox": [ + 53, + 314, + 286, + 357 + ], + "type": "text", + "content": "[6] Alexei Baevski, Yuhao Zhou, Abdelrahman Mohamed, and Michael Auli. wav2vec 2.0: A framework for self-supervised learning of speech representations. Advances in neural information processing systems, 33:12449-12460, 2020. 8" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 358, + 286, + 401 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 358, + 286, + 401 + ], + "spans": [ + { + "bbox": [ + 53, + 358, + 286, + 401 + ], + "type": "text", + "content": "[7] Shaojie Bai, J Zico Kolter, and Vladlen Koltun. An empirical evaluation of generic convolutional and recurrent networks for sequence modeling. arxiv 2018. arXiv preprint arXiv:1803.01271, 2, 1803. 8" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 403, + 286, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 403, + 286, + 479 + ], + "spans": [ + { + "bbox": [ + 53, + 403, + 286, + 479 + ], + "type": "text", + "content": "[8] Gilad Baruch, Zhuoyuan Chen, Afshin Dehghan, Tal Dimry, Yuri Feigin, Peter Fu, Thomas Gebauer, Brandon Joffe, Daniel Kurz, Arik Schwartz, and Elad Shulman. ARK-scenes - a diverse real-world dataset for 3d indoor scene understanding using mobile RGB-d data. In Thirty-fifth Conference on Neural Information Processing Systems Datasets and Benchmarks Track (Round 1), 2021. 2, 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 480, + 286, + 523 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 480, + 286, + 523 + ], + "spans": [ + { + "bbox": [ + 53, + 480, + 286, + 523 + ], + "type": "text", + "content": "[9] Piotr Bojanowski, Edouard Grave, Armand Joulin, and Tomas Mikolov. Enriching word vectors with subword information. Transactions of the association for computational linguistics, 5:135-146, 2017. 5, 8" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 525, + 286, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 525, + 286, + 557 + ], + "spans": [ + { + "bbox": [ + 48, + 525, + 286, + 557 + ], + "type": "text", + "content": "[10] Biovision BVH. Biovision bvh. In https://research.cs.wisc.edu/graphics /Courses/ cs-838-1999/Jeff/BVH.html, 1999. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 558, + 286, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 558, + 286, + 612 + ], + "spans": [ + { + "bbox": [ + 48, + 558, + 286, + 612 + ], + "type": "text", + "content": "[11] Zhe Cao, Gines Hidalgo, Tomas Simon, Shih-En Wei, and Yaser Sheikh. Openpose: realtime multi-person 2d pose estimation using part affinity fields. IEEE transactions on pattern analysis and machine intelligence, 43(1):172-186, 2019. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 613, + 286, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 613, + 286, + 656 + ], + "spans": [ + { + "bbox": [ + 48, + 613, + 286, + 656 + ], + "type": "text", + "content": "[12] Zhe Chen, Yuchen Duan, Wenhai Wang, Junjun He, Tong Lu, Jifeng Dai, and Yu Qiao. Vision transformer adapter for dense predictions. arXiv preprint arXiv:2205.08534, 2022. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 658, + 286, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 286, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 286, + 712 + ], + "type": "text", + "content": "[13] Kiran Chhatre, Radek Danček, Nikos Athanasiou, Giorgio Becherini, Christopher Peters, Michael J Black, and Timo Bolkart. Emotional speech-driven 3d body animation via disentangled latent diffusion. arXiv preprint arXiv:2312.04466, 2023. 1" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "text", + "content": "[14] Daniel Cudeiro, Timo Bolkart, Cassidy Laidlaw, Anurag Ranjan, and Michael Black. Capture, learning, and synthesis of 3D speaking styles. Computer Vision and Pattern Recognition (CVPR), pages 10101-10111, 2019. 2, 3, 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 119, + 545, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 119, + 545, + 162 + ], + "spans": [ + { + "bbox": [ + 307, + 119, + 545, + 162 + ], + "type": "text", + "content": "[15] Radek Dančecek, Kiran Chhatre, Shashank Tripathi, Yandong Wen, Michael Black, and Timo Bolkart. Emotional speech-driven animation with content-emotion disentanglement. ACM, 2023. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 165, + 545, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 165, + 545, + 196 + ], + "spans": [ + { + "bbox": [ + 307, + 165, + 545, + 196 + ], + "type": "text", + "content": "[16] Anna Deichler, Kiran Chhatre, Christopher Peters, and Jonas Beskow. Spatio-temporal priors in 3d human motion. 2021. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 200, + 545, + 243 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 200, + 545, + 243 + ], + "spans": [ + { + "bbox": [ + 307, + 200, + 545, + 243 + ], + "type": "text", + "content": "[17] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018. 3, 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 246, + 545, + 311 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 246, + 545, + 311 + ], + "spans": [ + { + "bbox": [ + 307, + 246, + 545, + 311 + ], + "type": "text", + "content": "[18] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 314, + 545, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 314, + 545, + 369 + ], + "spans": [ + { + "bbox": [ + 307, + 314, + 545, + 369 + ], + "type": "text", + "content": "[19] Yingruo Fan, Zhaojiang Lin, Jun Saito, Wenping Wang, and Taku Komura. Faceformer: Speech-driven 3d facial animation with transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2, 3, 6, 7, 8" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 371, + 545, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 371, + 545, + 426 + ], + "spans": [ + { + "bbox": [ + 307, + 371, + 545, + 426 + ], + "type": "text", + "content": "[20] Yingruo Fan, Zhaojiang Lin, Jun Saito, Wenping Wang, and Taku Komura. Faceformer: Speech-driven 3d facial animation with transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18770-18780, 2022. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 428, + 545, + 471 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 428, + 545, + 471 + ], + "spans": [ + { + "bbox": [ + 307, + 428, + 545, + 471 + ], + "type": "text", + "content": "[21] Gabriele Fanelli, Juergen Gall, Harald Romsdorfer, Thibaut Weise, and Luc Van Gool. A 3-d audio-visual corpus of affective communication. IEEE Transactions on Multimedia, 12(6):591-598, 2010. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 474, + 545, + 518 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 474, + 545, + 518 + ], + "spans": [ + { + "bbox": [ + 307, + 474, + 545, + 518 + ], + "type": "text", + "content": "[22] Ylva Ferstl and Rachel McDonnell. Investigating the use of recurrent motion modelling for speech gesture generation. In Proceedings of the 18th International Conference on Intelligent Virtual Agents, pages 93-98, 2018. 3, 8" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 520, + 545, + 574 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 520, + 545, + 574 + ], + "spans": [ + { + "bbox": [ + 307, + 520, + 545, + 574 + ], + "type": "text", + "content": "[23] Erik Gartner, Mykhaylo Andriluka, Erwin Coumans, and Cristian Sminchisescu. Differentiable dynamics for articulated 3d human motion reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13190-13200, 2022. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 577, + 545, + 620 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 577, + 545, + 620 + ], + "spans": [ + { + "bbox": [ + 307, + 577, + 545, + 620 + ], + "type": "text", + "content": "[24] Saeed Ghorbani, Ylva Ferstl, Daniel Holden, Nikolaus F. Troje, and Marc-Andre Carbonneau. Zeroeggs: Zero-shot example-based gesture generation from speech. Computer Graphics Forum, 42(1):206-216, 2023. 2, 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 623, + 545, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 623, + 545, + 678 + ], + "spans": [ + { + "bbox": [ + 307, + 623, + 545, + 678 + ], + "type": "text", + "content": "[25] Shiry Ginosar, Amir Bar, Gefen Kohavi, Caroline Chan, Andrew Owens, and Jitendra Malik. Learning individual styles of conversational gesture. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3497-3506, 2019. 1, 2, 3, 7" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 680, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 680, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 307, + 680, + 545, + 712 + ], + "type": "text", + "content": "[26] Yuan Gong, Yu-An Chung, and James Glass. Ast: Audio spectrogram transformer. arXiv preprint arXiv:2104.01778, 2021.3" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1152" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 117 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 117 + ], + "type": "text", + "content": "[27] Chuan Guo, Xinxin Zuo, Sen Wang, and Li Cheng. Tm2t: Stochastic and tokenized modeling for the reciprocal generation of 3d human motions and texts. In European Conference on Computer Vision, pages 580-597. Springer, 2022. 8, 4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 118, + 288, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 118, + 288, + 173 + ], + "spans": [ + { + "bbox": [ + 49, + 118, + 288, + 173 + ], + "type": "text", + "content": "[28] Ikhsanul Habibie, Weipeng Xu, Dushyant Mehta, Lingjie Liu, Hans-Peter Seidel, Gerard Pons-Moll, Mohamed Elgharib, and Christian Theobalt. Learning speech-driven 3d conversational gestures from video. arXiv preprint arXiv:2102.06837, 2021. 1, 2, 3, 7, 8" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 175, + 288, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 175, + 288, + 228 + ], + "spans": [ + { + "bbox": [ + 48, + 175, + 288, + 228 + ], + "type": "text", + "content": "[29] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16000-16009, 2022. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 230, + 288, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 230, + 288, + 285 + ], + "spans": [ + { + "bbox": [ + 48, + 230, + 288, + 285 + ], + "type": "text", + "content": "[30] Catalin Ionescu, Dragos Papava, Vlad Olaru, and Cristian Sminchisescu. Human3. 6m: Large scale datasets and predictive methods for 3d human sensing in natural environments. IEEE transactions on pattern analysis and machine intelligence, 36(7):1325-1339, 2013. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 286, + 287, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 286, + 287, + 330 + ], + "spans": [ + { + "bbox": [ + 48, + 286, + 287, + 330 + ], + "type": "text", + "content": "[31] Zhenzhong Lan, Mingda Chen, Sebastian Goodman, Kevin Gimpel, Piyush Sharma, and Radu Soricut. Albert: A lite bert for self-supervised learning of language representations. arXiv preprint arXiv:1909.11942, 2019. 3, 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 331, + 288, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 331, + 288, + 408 + ], + "spans": [ + { + "bbox": [ + 48, + 331, + 288, + 408 + ], + "type": "text", + "content": "[32] Gilwoo Lee, Zhiwei Deng, Shugao Ma, Takaaki Shiratori, Siddhartha S Srinivasa, and Yaser Sheikh. Talking with hands " + }, + { + "bbox": [ + 48, + 331, + 288, + 408 + ], + "type": "inline_equation", + "content": "16.2\\mathrm{m}" + }, + { + "bbox": [ + 48, + 331, + 288, + 408 + ], + "type": "text", + "content": ": A large-scale dataset of synchronized body-finger motion and audio for conversational motion analysis and synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 763-772, 2019. 2, 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 410, + 288, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 410, + 288, + 475 + ], + "spans": [ + { + "bbox": [ + 48, + 410, + 288, + 475 + ], + "type": "text", + "content": "[33] Jing Li, Di Kang, Wenjie Pei, Xuefei Zhe, Ying Zhang, Zhenyu He, and Linchao Bao. Audio2gestures: Generating diverse gestures from speech audio with conditional variational autoencoders. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11293-11302, 2021. 1, 2, 7, 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 477, + 288, + 531 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 477, + 288, + 531 + ], + "spans": [ + { + "bbox": [ + 48, + 477, + 288, + 531 + ], + "type": "text", + "content": "[34] Ruilong Li, Shan Yang, David A Ross, and Angjoo Kanazawa. Ai choreographer: Music conditioned 3d dance generation with aist++. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13401-13412, 2021. 7, 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 533, + 287, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 533, + 287, + 578 + ], + "spans": [ + { + "bbox": [ + 48, + 533, + 287, + 578 + ], + "type": "text", + "content": "[35] Tianye Li, Timo Bolkart, Michael. J. Black, Hao Li, and Javier Romero. Learning a model of facial shape and expression from 4D scans. ACM Transactions on Graphics, (Proc. SIGGRAPH Asia), 36(6):194:1-194:17, 2017. 2, 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 578, + 287, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 578, + 287, + 623 + ], + "spans": [ + { + "bbox": [ + 48, + 578, + 287, + 623 + ], + "type": "text", + "content": "[36] Haiyang Liu and Cheng Zhang. Reinforcement learning based neural architecture search for audio tagging. In 2020 International Joint Conference on Neural Networks (IJCNN), pages 1-8. IEEE, 2020. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 624, + 287, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 624, + 287, + 667 + ], + "spans": [ + { + "bbox": [ + 48, + 624, + 287, + 667 + ], + "type": "text", + "content": "[37] Haiyang Liu and Jihan Zhang. Improving ultrasound tongue image reconstruction from lip images using self-supervised learning and attention mechanism. arXiv preprint arXiv:2106.11769, 2021. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 669, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 669, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 669, + 287, + 713 + ], + "type": "text", + "content": "[38] Haiyang Liu, Naoya Iwamoto, Zihao Zhu, Zhengqing Li, You Zhou, Elif Bozkurt, and Bo Zheng. Disco: Disentangled implicit content and rhythm learning for diverse co-speech gestures synthesis. In Proceedings of the 30th ACM Interna-" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 713 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "type": "text", + "content": "tional Conference on Multimedia, pages 3764-3773, 2022. 2,3,7" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 307, + 96, + 547, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 96, + 547, + 150 + ], + "spans": [ + { + "bbox": [ + 307, + 96, + 547, + 150 + ], + "type": "text", + "content": "[39] Haiyang Liu, Zihao Zhu, Naoya Iwamoto, Yichen Peng, Zhengqing Li, You Zhou, Elif Bozkurt, and Bo Zheng. Beat: A large-scale semantic and emotional multi-modal dataset for conversational gestures synthesis. arXiv preprint arXiv:2203.05297, 2022. 1, 2, 3, 4, 6, 7" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 152, + 547, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 152, + 547, + 217 + ], + "spans": [ + { + "bbox": [ + 307, + 152, + 547, + 217 + ], + "type": "text", + "content": "[40] Xian Liu, Qianyi Wu, Hang Zhou, Yinghao Xu, Rui Qian, Xinyi Lin, Xiaowei Zhou, Wayne Wu, Bo Dai, and Bolei Zhou. Learning hierarchical cross-modal association for cospeech gesture generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10462-10472, 2022. 2, 3, 7" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 219, + 546, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 219, + 546, + 285 + ], + "spans": [ + { + "bbox": [ + 307, + 219, + 546, + 285 + ], + "type": "text", + "content": "[41] Xian Liu, Qianyi Wu, Hang Zhou, Yinghao Xu, Rui Qian, Xinyi Lin, Xiaowei Zhou, Wayne Wu, Bo Dai, and Bolei Zhou. Learning hierarchical cross-modal association for cospeech gesture generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10462-10472, 2022. 1, 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 286, + 546, + 340 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 286, + 546, + 340 + ], + "spans": [ + { + "bbox": [ + 308, + 286, + 546, + 340 + ], + "type": "text", + "content": "[42] Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. Roberta: A robustly optimized bert pretraining approach. arXiv preprint arXiv:1907.11692, 2019.3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 342, + 546, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 342, + 546, + 397 + ], + "spans": [ + { + "bbox": [ + 308, + 342, + 546, + 397 + ], + "type": "text", + "content": "[43] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10012-10022, 2021. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 399, + 546, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 399, + 546, + 442 + ], + "spans": [ + { + "bbox": [ + 308, + 399, + 546, + 442 + ], + "type": "text", + "content": "[44] Matthew M. Loper, Naureen Mahmood, and Michael J. Black. MoSh: Motion and shape capture from sparse markers. ACM Transactions on Graphics, (Proc. SIGGRAPH Asia), 33(6):220:1-220:13, 2014. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 444, + 546, + 476 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 444, + 546, + 476 + ], + "spans": [ + { + "bbox": [ + 308, + 444, + 546, + 476 + ], + "type": "text", + "content": "[45] Shuhong Lu, Youngwoo Yoon, and Andrew Feng. Co-speech gesture synthesis using discrete gesture token learning. arXiv preprint arXiv:2303.12822, 2023. 2, 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 478, + 546, + 531 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 478, + 546, + 531 + ], + "spans": [ + { + "bbox": [ + 308, + 478, + 546, + 531 + ], + "type": "text", + "content": "[46] Naureen Mahmood, Nima Ghorbani, Nikolaus F. Troje, Gerard Pons-Moll, and Michael J. Black. AMASS: Archive of motion capture as surface shapes. In International Conference on Computer Vision, pages 5442-5451, 2019. 2, 3, 6, 8" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 534, + 546, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 534, + 546, + 567 + ], + "spans": [ + { + "bbox": [ + 308, + 534, + 546, + 567 + ], + "type": "text", + "content": "[47] Tim Massey. Vicon study of dynamic object tracking accuracy. In https://www.vicon.com/resources/blog/vicon-study-of-dynamic-object-tracking-accuracy/, 2020. 2, 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 568, + 546, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 568, + 546, + 622 + ], + "spans": [ + { + "bbox": [ + 308, + 568, + 546, + 622 + ], + "type": "text", + "content": "[48] Evonne Ng, Shiry Ginosar, Trevor Darrell, and Hanbyul Joo. Body2hands: Learning to infer 3d hands from conversational gesture body dynamics. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11865-11874, 2021. 2, 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 624, + 546, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 624, + 546, + 678 + ], + "spans": [ + { + "bbox": [ + 308, + 624, + 546, + 678 + ], + "type": "text", + "content": "[49] Kunkun Pang, Dafei Qin, Yingruo Fan, Julian Habekost, Takaaki Shiratori, Junichi Yamagishi, and Taku Komura. Bodyformer: Semantics-guided 3d body gesture synthesis with transformer. ACM Transactions on Graphics (TOG), 42(4):1-12, 2023. 2, 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 680, + 546, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 680, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 680, + 546, + 713 + ], + "type": "text", + "content": "[50] Georgios Pavlakos, Vasileios Choutas, Nima Ghorbani, Timo Bolkart, Ahmed A. A. Osman, Dimitrios Tzionas, and Michael J. Black. Expressive body capture: 3D hands, face," + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1153" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 66, + 72, + 287, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 72, + 287, + 106 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 287, + 106 + ], + "type": "text", + "content": "and body from a single image. In Proceedings IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), pages 10975-10985, 2019. 2, 3, 4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 108, + 287, + 163 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 108, + 287, + 163 + ], + "spans": [ + { + "bbox": [ + 48, + 108, + 287, + 163 + ], + "type": "text", + "content": "[51] Ziqiao Peng, Haoyu Wu, Zhenbo Song, Hao Xu, Xiangyu Zhu, Jun He, Hongyan Liu, and Zhaoxin Fan. Emotalk: Speech-driven emotional disentanglement for 3d face animation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 20687-20697, 2023. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 165, + 288, + 209 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 165, + 288, + 209 + ], + "spans": [ + { + "bbox": [ + 48, + 165, + 288, + 209 + ], + "type": "text", + "content": "[52] Mathis Petrovich, Michael J Black, and Gül Varol. Action-conditioned 3d human motion synthesis with transformer vae. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10985-10995, 2021. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 211, + 287, + 255 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 211, + 287, + 255 + ], + "spans": [ + { + "bbox": [ + 48, + 211, + 287, + 255 + ], + "type": "text", + "content": "[53] Xingqun Qi, Chen Liu, Lincheng Li, Jie Hou, Haoran Xin, and Xin Yu. Emotiongesture: Audio-driven diverse emotional co-speech 3d gesture generation. arXiv preprint arXiv:2305.18891, 2023. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 257, + 287, + 312 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 257, + 287, + 312 + ], + "spans": [ + { + "bbox": [ + 48, + 257, + 287, + 312 + ], + "type": "text", + "content": "[54] Xingqun Qi, Jiahao Pan, Peng Li, Ruibin Yuan, Xiaowei Chi, Mengfei Li, Wenhan Luo, Wei Xue, Shanghang Zhang, Qifeng Liu, et al. Weakly-supervised emotion transition learning for diverse 3d co-speech gesture generation. arXiv preprint arXiv:2311.17532, 2023. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 314, + 287, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 314, + 287, + 346 + ], + "spans": [ + { + "bbox": [ + 48, + 314, + 287, + 346 + ], + "type": "text", + "content": "[55] Guocheng Qian, Xingdi Zhang, Abdullah Hamdi, and Bernard Ghanem. Pix4point: Image pretrained transformers for 3d point cloud understanding. 2022. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 350, + 287, + 404 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 350, + 287, + 404 + ], + "spans": [ + { + "bbox": [ + 48, + 350, + 287, + 404 + ], + "type": "text", + "content": "[56] Alexander Richard, Michael Zollhöfer, Yandong Wen, Fernando De la Torre, and Yaser Sheikh. Meshtalk: 3d face animation from speech using cross-modality disentanglement. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1173-1182, 2021. 2, 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 406, + 287, + 450 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 406, + 287, + 450 + ], + "spans": [ + { + "bbox": [ + 48, + 406, + 287, + 450 + ], + "type": "text", + "content": "[57] Kaede Shiohara, Xingchao Yang, and Takafumi Taketomi. Blendface: Re-designing identity encoders for face-swapping. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7634-7644, 2023. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 453, + 287, + 506 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 453, + 287, + 506 + ], + "spans": [ + { + "bbox": [ + 48, + 453, + 287, + 506 + ], + "type": "text", + "content": "[58] Tommi Tykkälä, Cédric Audras, and Andrew I Comport. Direct iterative closest point for real-time visual odometry. In 2011 IEEE International Conference on Computer Vision Workshops (ICCV Workshops), pages 2050-2056. IEEE, 2011. 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 509, + 287, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 509, + 287, + 542 + ], + "spans": [ + { + "bbox": [ + 48, + 509, + 287, + 542 + ], + "type": "text", + "content": "[59] Aaron Van Den Oord, Oriol Vinyals, et al. Neural discrete representation learning. Advances in neural information processing systems, 30, 2017. 5, 8" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 544, + 287, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 544, + 287, + 588 + ], + "spans": [ + { + "bbox": [ + 48, + 544, + 287, + 588 + ], + "type": "text", + "content": "[60] Kaisiyuan Wang, Qianyi Wu, Linsen Song, Zhuoqian Yang, Wayne Wu, Chen Qian, Ran He, Yu Qiao, and Chen Change Loy. Mead: A large-scale audio-visual dataset for emotional talking-face generation. In ECCV, 2020. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 590, + 287, + 655 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 590, + 287, + 655 + ], + "spans": [ + { + "bbox": [ + 48, + 590, + 287, + 655 + ], + "type": "text", + "content": "[61] Jinbo Xing, Menghan Xia, Yuechen Zhang, Xiaodong Cun, Jue Wang, and Tien-Tsin Wong. Codetalker: Speech-driven 3d facial animation with discrete motion prior. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12780-12790, 2023. 1, 2, 3, 4, 7, 8" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 658, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 287, + 712 + ], + "type": "text", + "content": "[62] Sicheng Yang, Zhiyong Wu, Minglei Li, Zhensong Zhang, Lei Hao, Weihong Bao, Ming Cheng, and Long Xiao. Diffusestylegesture: Stylized audio-driven co-speech gesture generation with diffusion models. arXiv preprint arXiv:2305.04919, 2023. 2, 3, 7" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 662 + ], + "type": "list", + "angle": 0, + "index": 25, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "text", + "content": "[63] Xingchao Yang and Takafumi Taketomi. Bareskinnet: Demakeup and de-lighting via 3d face reconstruction. In Computer Graphics Forum, pages 623-634. Wiley Online Library, 2022. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 118, + 545, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 545, + 171 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 545, + 171 + ], + "type": "text", + "content": "[64] Xingchao Yang, Takafumi Taketomi, and Yoshihiro Kanamori. Makeup extraction of 3d representation via illumination-aware image decomposition. In Computer Graphics Forum, pages 293-307. Wiley Online Library, 2023. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 174, + 545, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 174, + 545, + 217 + ], + "spans": [ + { + "bbox": [ + 307, + 174, + 545, + 217 + ], + "type": "text", + "content": "[65] Hongwei Yi, Hualin Liang, Yifei Liu, Qiong Cao, Yandong Wen, Timo Bolkart, Dacheng Tao, and Michael J Black. Generating holistic 3d human motion from speech. In CVPR, 2023. 1, 2, 3, 4, 5, 6, 7, 8" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 219, + 545, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 219, + 545, + 282 + ], + "spans": [ + { + "bbox": [ + 307, + 219, + 545, + 282 + ], + "type": "text", + "content": "[66] Youngwoo Yoon, Woo-Ri Ko, Minsu Jang, Jaeyeon Lee, Jae-hong Kim, and Geehyuk Lee. Robots learn social skills: End-to-end learning of co-speech gesture generation for humanoid robots. In 2019 International Conference on Robotics and Automation (ICRA), pages 4303-4309. IEEE, 2019. 2, 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 285, + 545, + 340 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 285, + 545, + 340 + ], + "spans": [ + { + "bbox": [ + 307, + 285, + 545, + 340 + ], + "type": "text", + "content": "[67] Youngwoo Yoon, Bok Cha, Joo-Haeng Lee, Minsu Jang, Jaeyeon Lee, Jaehong Kim, and Geehyuk Lee. Speech gesture generation from the trimodal context of text, audio, and speaker identity. ACM Transactions on Graphics (TOG), 39 (6):1-16, 2020. 1, 2, 3, 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 342, + 545, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 342, + 545, + 396 + ], + "spans": [ + { + "bbox": [ + 307, + 342, + 545, + 396 + ], + "type": "text", + "content": "[68] Xumin Yu, Lulu Tang, Yongming Rao, Tiejun Huang, Jie Zhou, and Jiwen Lu. Point-bert: Pre-training 3d point cloud transformers with masked point modeling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19313-19322, 2022. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 397, + 545, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 397, + 545, + 441 + ], + "spans": [ + { + "bbox": [ + 307, + 397, + 545, + 441 + ], + "type": "text", + "content": "[69] Hengshuang Zhao, Li Jiang, Jiaya Jia, Philip HS Torr, and Vladlen Koltun. Point transformer. In Proceedings of the IEEE/CVF international conference on computer vision, pages 16259-16268, 2021. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 442, + 545, + 496 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 442, + 545, + 496 + ], + "spans": [ + { + "bbox": [ + 307, + 442, + 545, + 496 + ], + "type": "text", + "content": "[70] Yihao Zhi, Xiaodong Cun, Xuelin Chen, Xi Shen, Wen Guo, Shaoli Huang, and Shenghua Gao. Livelyspeaker: Towards semantic-aware co-speech gesture generation. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 20807-20817, 2023. 2, 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 498, + 545, + 552 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 498, + 545, + 552 + ], + "spans": [ + { + "bbox": [ + 307, + 498, + 545, + 552 + ], + "type": "text", + "content": "[71] Yang Zhou, Jimei Yang, Dingzeyu Li, Jun Saito, Deepali Aneja, and Evangelos Kalogerakis. Audio-driven neural gesture reenactment with video motion graphs. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3418-3428, 2022. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 554, + 545, + 608 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 554, + 545, + 608 + ], + "spans": [ + { + "bbox": [ + 307, + 554, + 545, + 608 + ], + "type": "text", + "content": "[72] Lingting Zhu, Xian Liu, Xuanyu Liu, Rui Qian, Ziwei Liu, and Lequan Yu. Taming diffusion models for audio-driven co-speech gesture generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10544-10553, 2023. 2, 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 609, + 545, + 662 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 609, + 545, + 662 + ], + "spans": [ + { + "bbox": [ + 307, + 609, + 545, + 662 + ], + "type": "text", + "content": "[73] Wentao Zhu, Xiaoxuan Ma, Zhaoyang Liu, Libin Liu, Wayne Wu, and Yizhou Wang. Motionbert: A unified perspective on learning human motion representations. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2023. 3" + } + ] + } + ], + "index": 24 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1154" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/EMCAD_ Efficient Multi-scale Convolutional Attention Decoding for Medical Image Segmentation/32047b53-ccf5-41e9-bc53-4e712089f958_content_list.json b/2024/EMCAD_ Efficient Multi-scale Convolutional Attention Decoding for Medical Image Segmentation/32047b53-ccf5-41e9-bc53-4e712089f958_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..a5b3b2cac64aa3b78b676f995a45e54c0ef7b81a --- /dev/null +++ b/2024/EMCAD_ Efficient Multi-scale Convolutional Attention Decoding for Medical Image Segmentation/32047b53-ccf5-41e9-bc53-4e712089f958_content_list.json @@ -0,0 +1,1618 @@ +[ + { + "type": "text", + "text": "EMCAD: Efficient Multi-scale Convolutional Attention Decoding for Medical Image Segmentation", + "text_level": 1, + "bbox": [ + 96, + 128, + 875, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Md Mostafijur Rahman, Mustafa Munir, and Radu Marculescu", + "bbox": [ + 236, + 203, + 732, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The University of Texas at Austin", + "bbox": [ + 351, + 222, + 617, + 238 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Austin, Texas, USA", + "bbox": [ + 405, + 239, + 565, + 255 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "mostafijur.rahman, mmunir, radum@utexas.edu", + "bbox": [ + 289, + 258, + 673, + 273 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 309, + 313, + 325 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "An efficient and effective decoding mechanism is crucial in medical image segmentation, especially in scenarios with limited computational resources. However, these decoding mechanisms usually come with high computational costs. To address this concern, we introduce EMCAD, a new efficient multi-scale convolutional attention decoder, designed to optimize both performance and computational efficiency. EMCAD leverages a unique multi-scale depth-wise convolution block, significantly enhancing feature maps through multi-scale convolutions. EMCAD also employs channel, spatial, and grouped (large-kernel) gated attention mechanisms, which are highly effective at capturing intricate spatial relationships while focusing on salient regions. By employing group and depth-wise convolution, EMCAD is very efficient and scales well (e.g., only 1.91M parameters and 0.381G FLOPs are needed when using a standard encoder). Our rigorous evaluations across 12 datasets that belong to six medical image segmentation tasks reveal that EMCAD achieves state-of-the-art (SOTA) performance with $79.4\\%$ and $80.3\\%$ reduction in #Params and #FLOPs, respectively. Moreover, EMCAD's adaptability to different encoders and versatility across segmentation tasks further establish EMCAD as a promising tool, advancing the field towards more efficient and accurate medical image analysis. Our implementation is available at https://github.com/SLDGroup/EMCAD.", + "bbox": [ + 76, + 340, + 472, + 734 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 753, + 209, + 768 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In the realm of medical diagnostics and therapeutic strategies, automated segmentation of medical images is vital, as it classifies pixels to identify critical regions such as lesions, tumors, or entire organs. A variety of U-shaped convolutional neural network (CNN) architectures [20, 24, 37, 41, 44, 62], notably UNet [44], UNet++ [62], UNet3+ [24], and nnU-Net [19], have become standard techniques for this purpose, achieving high-quality, high-resolution segmen", + "bbox": [ + 75, + 779, + 470, + 902 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "tation output. Attention mechanisms [12, 17, 20, 41, 57] have also been integrated into these models to enhance feature maps and improve pixel-level classification. Although attention-based models have shown improved performance, they still face significant challenges due to the computationally expensive convolutional blocks that are typically used in conjunction with attention mechanisms.", + "bbox": [ + 496, + 310, + 893, + 415 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recently, vision transformers [18] have shown promise in medical image segmentation tasks [5, 8, 17, 42, 43, 52, 54, 61] by capturing long-range dependencies among pixels through Self-attention (SA) mechanisms. Hierarchical vision transformers like Swin [34], PVT [55, 56], MaxViT [49], MERIT [43], ConvFormer [33], and MetaFormer [59] have been introduced to further improve the performance in this field. While the SA excels at capturing global information, it is less adept at understanding the local spatial context [13, 28]. To address this limitation, some approaches have integrated local convolutional attention within the decoders to better grasp spatial details. Nevertheless, these methods can still be computationally demanding because they frequently employ costly convolutional blocks. This limits their applicability to real-world scenarios where computational resources are restricted.", + "bbox": [ + 496, + 416, + 893, + 656 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To address the aforementioned limitations, we introduce EMCAD, an efficient multi-scale convolutional attention decoding using a new multi-scale depth-wise convolution block. More precisely, EMCAD enhances the feature maps via efficient multi-scale convolutions, while incorporating complex spatial relationships and local attention through the use of channel, spatial, and grouped (large-kernel) gated attention mechanisms. Our contributions are as follows:", + "bbox": [ + 496, + 657, + 893, + 777 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "- New Efficient Multi-scale Convolutional Decoder: We introduce an efficient multi-scale cascaded fullyconvolutional attention decoder (EMCAD) for 2D medical image segmentation; this takes the multi-stage features of vision encoders and progressively enhances the multi-scale and multi-resolution spatial representations. EMCAD has only 0.506M parameters and 0.11G FLOPs for a tiny encoder with #channels = [32, 64, 160, 256],", + "bbox": [ + 500, + 779, + 893, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "11769", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/3af94799489a0efcba7e12ed56d3a523e2bdb917ee5fa27705063e9a6672a011.jpg", + "image_caption": [ + "Figure 1. Average DICE scores vs. #FLOPs for different methods over 10 binary medical image segmentation datasets. As shown, our approaches (PVT-EMCAD-B0 and PVT-EMCAD-B2) have the lowest #FLOPs, yet the highest DICE scores." + ], + "image_footnote": [], + "bbox": [ + 109, + 89, + 441, + 273 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "while it has 1.91M parameters and 0.381G FLOPs for a standard encoder with #channels = [64, 128, 320, 512].", + "bbox": [ + 89, + 347, + 468, + 378 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Efficient Multi-scale Convolutional Attention Module: We introduce MSCAM, a new efficient multi-scale convolutional attention module that performs depth-wise convolutions at multiple scales; this refines the feature maps produced by vision encoders and enables capturing multi-scale salient features by suppressing irrelevant regions. The use of depth-wise convolutions makes MSCAM very efficient.", + "- Large-kernel Grouped Attention Gate: We introduce a new grouped attention gate to fuse refined features with the features from skip connections. By using larger kernel $(3 \\times 3)$ group convolutions instead of point-wise convolutions in the design, we capture salient features in a larger local context with less computation.", + "- Improved Performance: We empirically show that EM-CAD can be used with any hierarchical vision encoder (e.g., PVTv2-B0, PVTv2-B2 [56]), while significantly improving the performance of 2D medical image segmentation. EMCAD produces better results than SOTA methods with a significantly lower computational cost (as shown in Figure 1) on 12 medical image segmentation benchmarks that belong to six different tasks." + ], + "bbox": [ + 76, + 378, + 468, + 710 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The remaining of this paper is organized as follows: Section 2 summarizes related work. Section 3 describes the proposed method. Section 4 explains our experimental setup and results on 12 medical image segmentation benchmarks. Section 5 covers different ablation experiments. Lastly, Section 6 concludes the paper.", + "bbox": [ + 75, + 710, + 468, + 803 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 76, + 819, + 218, + 834 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Vision encoders", + "text_level": 1, + "bbox": [ + 76, + 845, + 235, + 859 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Convolutional Neural Networks (CNNs) [21-23, 32, 35, 45-48] have been foundational as encoders due to their pro", + "bbox": [ + 75, + 869, + 468, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "ficiency in handling spatial relationships in images. More precisely, AlexNet [32] and VGG [46] pave the way, leveraging deep layers of convolutions to extract features progressively. GoogleNet [47] introduces the inception module, allowing more efficient computation of representations across various scales. ResNet [21] introduces residual connections, enabling the training of networks with substantially more layers by addressing the vanishing gradients problem. MobileNets [22, 45] bring CNNs to mobile devices through lightweight, depth-wise separable convolutions. EfficientNet [48] introduces a scalable architectural design to CNNs with compound scaling. Although CNNs are pivotal for many vision applications, they generally lack the ability to capture long-range dependencies within images due to their inherent local receptive fields.", + "bbox": [ + 496, + 90, + 890, + 316 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recently, Vision Transformers (ViTs), pioneered by Dosovitskiy et al. [18], enabled the learning of long-range relationships among pixels using Self-attention (SA). Since then, ViTs have been enhanced by integrating CNN features [49, 56], developing novel self-attention (SA) blocks [34, 49], and introducing new architectural designs [55, 58]. The Swin Transformer [34] incorporates a sliding window attention mechanism, while SegFormer [58] leverages MixFFN blocks for hierarchical structures. PVT [55] uses spatial reduction attention, refined in PVTv2 [56] with overlapping patch embedding and a linear complexity attention layer. MaxViT [49] introduces a multi-axis self-attention to form a hierarchical CNN-transformer encoder. Although ViTs address the CNNs limitation in capturing long-range pixel dependencies [21-23, 32, 35, 45-48], they face challenges in capturing the local spatial relationships among pixels. In this paper, we aim to overcome these limitations by introducing a new multi-scale cascaded attention decoder that refines feature maps and incorporates local attention using a multi-scale convolutional attention module.", + "bbox": [ + 496, + 319, + 892, + 621 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Medical image segmentation", + "text_level": 1, + "bbox": [ + 500, + 633, + 756, + 648 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Medical image segmentation involves pixel-wise classification to identify various anatomical structures like lesions, tumors, or organs within different imaging modalities such as endoscopy, MRI, or CT scans [8]. U-shaped networks [7, 19, 24, 26, 37, 41, 44, 62] are particularly favored due to their simple but effective encoder-decoder design. The UNet [44] pioneered this approach with its use of skip connections to fuse features at different resolution stages. UNet++ [62] evolves this design by incorporating nested encoder-decoder pathways with dense skip connections. Expanding on these ideas, UNet $3+$ [24] introduces comprehensive skip pathways that facilitate full-scale feature integration. Further advancement comes with DC-UNet [37], which integrates a multi-resolution convolution scheme and residual paths into its skip connections. The DeepLab series, including DeepLabv3 [10] and DeepLabv3+ [11], in", + "bbox": [ + 496, + 659, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "11770", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "produce atrous convolutions and spatial pyramid pooling to handle multi-scale information. SegNet [2] uses pooling indices to upsample feature maps, preserving the boundary details. nnU-Net [19] automatically configures hyperparameters based on the specific dataset characteristics, using standard 2D and 3D UNets. Collectively, these U-shaped models have become a benchmark for success in the domain of medical image segmentation.", + "bbox": [ + 75, + 90, + 468, + 210 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Recently, vision transformers have emerged as a formidable force in medical image segmentation, harnessing the ability to capture pixel relationships at global scales [5, 8, 17, 42, 43, 52, 58, 61]. TransUNet [8] presents a novel blend of CNNs for local feature extraction and transformers for global context, enhancing both local and global feature capture. Swin-Unet [5] extends this by incorporating Swin Transformer blocks [34] into a U-shaped model for both encoding and decoding processes. Building on these concepts, MERIT [43] introduces a multi-scale hierarchical transformer, which employs SA across different window sizes, thus enhancing the model capacity to capture multiscale features critical for medical image segmentation.", + "bbox": [ + 75, + 212, + 470, + 409 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The integration of attention mechanisms has been investigated within CNNs [20, 41] and transformer-based systems [17] for enhancing medical image segmentation. PraNet [20] employs a reverse attention strategy for feature refinement. PolypPVT [17] leverages PVTv2 [56] as its backbone encoder and incorporates CBAM [57] within its decoding stages. The CASCADE [42] presents a novel cascaded decoder, combining channel [23] and spatial [9] attention to refine features at multiple stages, extracted from a transformer encoder, culminating in high-resolution segmentation outputs. While CASCADE achieves notable performance in segmenting medical images by integrating local and global insights from transformers, it is computationally inefficient due to the use of triple $3 \\times 3$ convolution layers at each decoder stage. In addition to this, it uses single-scale convolutions during decoding. Our new proposal involves the adoption of multi-scale depth-wise convolutions to mitigate these constraints.", + "bbox": [ + 75, + 410, + 470, + 680 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Methodology", + "text_level": 1, + "bbox": [ + 76, + 690, + 212, + 708 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we first introduce our new EMCAD decoder and then explain two transformer-based architectures (i.e., PVT-EMCAD-B0 and PVT-EMCAD-B2) incorporating our proposed decoder.", + "bbox": [ + 75, + 715, + 468, + 777 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Efficient multi-scale convolutional attention decoding (EMCAD)", + "text_level": 1, + "bbox": [ + 76, + 785, + 468, + 816 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we introduce our efficient multi-scale convolutional decoding (EMCAD) to process the multi-stage features extracted from pretrained hierarchical vision encoders for high-resolution semantic segmentation. As shown in Figure 2(b), EMCAD consists of efficient multi", + "bbox": [ + 75, + 825, + 468, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "scale convolutional attention modules (MSCAMs) to robustly enhance the feature maps, large-kernel grouped attention gates (LGAGs) to refine feature maps fusing with the skip connection via gated attention mechanism, efficient up-convolution blocks (EUCBs) for up-sampling followed by enhancement of feature maps, and segmentation heads (SHs) to produce the segmentation outputs.", + "bbox": [ + 496, + 90, + 890, + 196 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "More specifically, we use four MSCAMs to refine pyramid features (i.e., $X1$ , $X2$ , $X3$ , $X4$ in Figure 2) extracted from the four stages of the encoder. After each MSCAM, we use an SH to produce a segmentation map of that stage. Subsequently, we upscale the refined feature maps using EUCBs and add them to the outputs from the corresponding LGAGs. Finally, we add four different segmentation maps to produce the final segmentation output. Different modules of our decoder are described next.", + "bbox": [ + 496, + 199, + 892, + 335 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1.1 Large-kernel grouped attention gate (LGAG)", + "text_level": 1, + "bbox": [ + 498, + 361, + 867, + 378 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We introduce a new large-kernel grouped attention gate (LGAG) to progressively combine feature maps with attention coefficients, which are learned by the network to allow higher activation of relevant features and suppression of irrelevant ones. This process employs a gating signal derived from higher-level features to control the flow of information across different stages of the network, thus enhancing its precision for medical image segmentation. Unlike Attention UNet [41] which uses $1 \\times 1$ convolution to process gating signal $g$ (features from skip connections) and input feature map $x$ (upsampled features), in our $q_{att}(.)$ function, we process $g$ and $x$ by applying separate $3 \\times 3$ group convolutions $GC_{g}(.)$ and $GC_{x}(.)$ , respectively. These convolved features are then normalized using batch normalization $(BN(.))$ [27] and merged through element-wise addition. The resultant feature map is activated through a ReLU $(R(.))$ layer [39]. Afterward, we apply a $1 \\times 1$ convolution $(C(.))$ followed by $BN(.)$ layer to get a single channel feature map. We then pass the resultant single-channel feature map through a Sigmoid $(\\sigma(.))$ activation function to yield the attention coefficients. The output of this transformation is used to scale the input feature $x$ through elementwise multiplication, producing the attention-gated feature $LGAG(g, x)$ . The $LGAG(.)$ (Figure 2(g)) can be formulated as in Equations 1 and 2:", + "bbox": [ + 496, + 391, + 892, + 768 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nq _ {a t t} (g, x) = R \\left(B N \\left(G C _ {g} (g) + B N \\left(G C _ {x} (x)\\right)\\right)\\right) \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 519, + 782, + 890, + 799 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nL G A G (g, x) = x * \\sigma \\left(B N \\left(C \\left(q _ {a t t} (g, x)\\right)\\right)\\right) \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 549, + 832, + 890, + 848 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Due to using $3 \\times 3$ kernel group convolutions in $q_{att}(.)$ , our LGAG captures comparatively larger spatial contexts with less computational cost.", + "bbox": [ + 496, + 859, + 892, + 905 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "11771", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/680916196f86f563262f187b1e9498563b3b45b7b797b45690ae9a0e76ac7c8c.jpg", + "image_caption": [ + "Figure 2. Hierarchical encoder with newly proposed EMCAD decoder architecture. (a) CNN or transformer encoder with four hierarchical stages, (b) EMCAD decoder, (c) Efficient up-convolution block (EUCB), (d) Multi-scale convolutional attention module (MSCAM), (e) Multi-scale convolution block (MSCB), (f) Multi-scale (parallel) depth-wise convolution (MSDC), (g) Large-kernel grouped attention gate (LGAG), (h) Channel attention block (CAB), and (i) Spatial attention block (SAB). X1, X2, X3, and X4 are the features from the four stages of the hierarchical encoder. p1, p2, p3, and p4 are output segmentation maps from four stages of our decoder." + ], + "image_footnote": [], + "bbox": [ + 117, + 87, + 854, + 354 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1.2 Multi-scale convolutional attention module (MSCAM)", + "text_level": 1, + "bbox": [ + 76, + 436, + 470, + 465 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We introduce an efficient multi-scale convolutional attention module to refine the feature maps. MSCAM consists of a channel attention block $(CAB(\\cdot))$ to put emphasis on pertinent channels, a spatial attention block [9] $(SAB(\\cdot))$ to capture the local contextual information, and an efficient multi-scale convolution block $(MSCB(\\cdot))$ to enhance the feature maps preserving contextual relationships. The $MSCAM(\\cdot)$ (Figure 2(d)) is given in Equation 3:", + "bbox": [ + 75, + 474, + 468, + 597 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nM S C A M (x) = M S C B (S A B (C A B (x))) \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 124, + 604, + 468, + 622 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $x$ is the input tensor. Due to using depth-wise convolution in multiple scales, our MSCAM is more effective with significantly lower computational cost than the convolutional attention module (CAM) proposed in [42].", + "bbox": [ + 75, + 628, + 468, + 689 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Multi-scale Convolution Block (MSCB): We introduce an efficient multi-scale convolution block to enhance the features generated by our cascaded expanding path. In our MSCB, we follow the design of the inverted residual block (IRB) of MobileNetV2 [45]. However, unlike IRB, our MSCB performs depth-wise convolution at multiple scales and uses channel_shuffle [60] to shuffle channels across groups. More specifically, in our MSCB, we first expand the number of channels (i.e., expansion_factor = 2) using a point-wise $(1 \\times 1)$ convolution layers $PWC_{1}(\\cdot)$ followed by a batch normalization layer $BN(\\cdot)$ and a ReLU6 [31] activation layer $R6(\\cdot)$ . We then use a multi-scale depth-wise convolution $MSDC(\\cdot)$ to capture both multi-scale and multi-resolution contexts. As depth-wise convolution overlooks", + "bbox": [ + 75, + 689, + 470, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "the relationships among channels, we use a channel_shuffle operation to incorporate relationships among channels. Afterward, we use another point-wise convolution $PWC_{2}(.)$ followed by a $BN(.)$ to transform back the original #channels, which also encodes dependency among channels. The $MSCB(\\cdot)$ (Figure 2(e)) is formulated as in Equation 4:", + "bbox": [ + 498, + 436, + 890, + 529 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nM S C B (x) = B N \\left(P W C _ {2} \\left(C S \\left(M S D C \\left(R 6 \\left(B N \\left(P W C _ {1} (x)\\right)\\right)\\right)\\right)\\right)\\right) \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 511, + 537, + 890, + 551 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where parallel $MSDC(.)$ (Figure 2(f)) for different kernel sizes $(KS)$ can be formulated using Equation 5:", + "bbox": [ + 498, + 560, + 890, + 592 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nM S D C (x) = \\sum_ {k s \\in K S} D W C B _ {k s} (x) \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 571, + 599, + 890, + 617 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $DWCB_{ks}(x) = R6(BN(DWC_{ks}(x)))$ . Here, $DWC_{ks}()$ is a depth-wise convolution with the kernel size $ks$ . $BN()$ and $R6()$ are batch normalization and ReLU6 activation, respectively. Additionally, our sequential $MSDC()$ uses the recursively updated input $x$ , where the input $x$ is residually connected to the previous $DWCB_{ks}()$ for better regularization as in Equation 6:", + "bbox": [ + 498, + 625, + 890, + 731 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nx = x + D W C B _ {k s} (x) \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 614, + 739, + 890, + 756 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Channel Attention Block (CAB): We use channel attention block to assign different levels of importance to each channel, thus emphasizing more relevant features while suppressing less useful ones. Basically, the CAB identifies which feature maps to focus on (and then refine them). Following [57], in CAB, we first apply the adaptive maximum pooling $(P_{m}(\\cdot))$ and adaptive average pooling $(P_{a}(\\cdot))$ to the spatial dimensions (i.e., height and width) to extract the most significant feature of the entire feature map per", + "bbox": [ + 496, + 763, + 890, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "11772", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "channel. Then, for each pooled feature map, we reduce the number of channels $r = 1 / 16$ times separately using a point-wise convolution $(C_1(\\cdot))$ followed by a ReLU activation $(R)$ . Afterward, we recover the original channels using another point-wise convolution $(C_2(\\cdot))$ . We then add both recovered feature maps and apply Sigmoid $(\\sigma)$ activation to estimate attention weights. Finally, we incorporate these weights to input $x$ using the Hadamard product $(\\otimes)$ . The $CAB(\\cdot)$ (Figure 2(h)) is defined using Equation 7:", + "bbox": [ + 75, + 90, + 472, + 227 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nC A B (x) = \\sigma \\left(C _ {2} \\left(R \\left(C _ {1} \\left(P _ {m} (x)\\right)\\right)\\right) + C _ {2} \\left(R \\left(C _ {1} \\left(P _ {a} (x)\\right)\\right)\\right)\\right) * x \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 94, + 237, + 468, + 252 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Spatial Attention Block (SAB): We use spatial attention to mimic the attentional processes of the human brain by focusing on specific parts of an input image. Basically, the SAB determines where to focus in a feature map; then it enhances those features. This process enhances the model's ability to recognize and respond to relevant spatial features, which is crucial for image segmentation where the context and location of objects significantly influence the output. In SAB, we first pool maximum $(Ch_{max}(\\cdot))$ and average $(Ch_{avg}(\\cdot))$ values along the channel dimension to pay attention to local features. Then, we use a large kernel (i.e., $7\\times 7$ as in [17]) convolution layer to enhance local contextual relationships among features. Afterward, we apply the Sigmoid activation $(\\sigma)$ to calculate attention weights. Finally, we feed these weights to the input $x$ (using Hadamard product $(\\otimes)$ to attend information in a more targeted way. The $SAB(.)$ (Figure 2(i)) is defined using Equation 8:", + "bbox": [ + 75, + 260, + 472, + 496 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nS A B (x) = \\sigma \\left(L K C \\left(\\left[ C h _ {m a x} (x), C h _ {a v g} (x) \\right]\\right)\\right) * x \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 114, + 506, + 468, + 522 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.1.3 Efficient up-convolution block (EUCB)", + "text_level": 1, + "bbox": [ + 76, + 534, + 401, + 550 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We use an efficient up-convolution block to progressively upsample the feature maps of the current stage to match the dimension and resolution of the feature maps from the next skip connection. The EUCB first uses an UpSampling $Up(\\cdot)$ with scale-factor 2 to upscale the feature maps. Then, it enhances the upscaled feature maps by applying a $3 \\times 3$ depth-wise convolution $DWC(\\cdot)$ followed by a $BN(\\cdot)$ and a $ReLU(\\cdot)$ activation. Finally, a $1 \\times 1$ convolution $C_{1 \\times 1}(\\cdot)$ is used to reduce the #channels to match with the next stage. The $EUCB(\\cdot)$ (Figure 2(c)) is formulated as in Equation 9:", + "bbox": [ + 75, + 559, + 472, + 710 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nE U C B (x) = C _ {1 \\times 1} \\left(R e L U \\left(B N \\left(D W C (U p (x))\\right)\\right)\\right) \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 89, + 737, + 468, + 753 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Due to using depth-wise convolution instead of $3 \\times 3$ convolution, our EUCB is very efficient.", + "bbox": [ + 76, + 762, + 468, + 792 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.1.4 Segmentation head (SH)", + "text_level": 1, + "bbox": [ + 76, + 800, + 302, + 816 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We use segmentation heads to produce the segmentation outputs from the refined feature maps of four stages of the decoder. The SH layer applies a $1 \\times 1$ convolution $Conv_{1 \\times 1}(\\cdot)$ to the refined feature maps having $ch_i$ channels ( $ch_i$ is the #channels in the feature map of stage $i$ ) and", + "bbox": [ + 76, + 825, + 472, + 902 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "produces output with #channels equal to #classes in target dataset for multi-class but 1 channel for binary segmentation. The $SH(\\cdot)$ is formulated as in Equation 10:", + "bbox": [ + 496, + 90, + 892, + 137 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nS H (x) = \\operatorname {C o n v} _ {1 \\times 1} (x) \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 614, + 142, + 892, + 159 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2. Overall architecture", + "text_level": 1, + "bbox": [ + 498, + 169, + 692, + 184 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To show the generalization, effectiveness, and ability to process multi-scale features for medical image segmentation, we integrate our EMCAD decoder alongside tiny (PVTv2-B0) and standard (PVTv2-B2) networks of PVTv2 [56]. However, our decoder is adaptable and seamlessly compatible with other hierarchical backbone networks.", + "bbox": [ + 496, + 191, + 890, + 281 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "PVTv2 differs from conventional transformer patch embedding modules by applying convolutional operations for consistent spatial information capture. Using PVTv2-b0 (Tiny) and PVTv2-b2 (Standard) encoders [56], we develop the PVT-EMCAD-B0 and PVT-EMCAD-B2 architectures. To adopt PVTv2, we first extract the features (X1, X2, X3, and X4) from four layers and feed them (i.e., X4 in the upsample path and X3, X2, X1 in the skip connections) into our EMCAD decoder as shown in Figure 2(a-b). EMCAD then processes them and produces four segmentation maps that correspond to the four stages of the encoder network.", + "bbox": [ + 496, + 282, + 892, + 449 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Multi-stage loss and outputs aggregation", + "text_level": 1, + "bbox": [ + 498, + 453, + 848, + 469 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our EMCAD decoder's four segmentation heads produce four prediction maps $p_1, p_2, p_3$ , and $p_4$ across its stages.", + "bbox": [ + 496, + 476, + 890, + 506 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Loss aggregation: We adopt a combinatorial approach to loss combination called MUTATION, inspired by the work of MERIT [43] for multi-class segmentation. This involves calculating the loss for all possible combinations of predictions derived from 4 heads, totaling $2^{4} - 1 = 15$ unique predictions, and then summing these losses. We focus on minimizing this cumulative combinatorial loss during the training process. For binary segmentation, we optimize the additive loss like [42] with an additional term $\\mathcal{L}_{p_1 + p_2 + p_3 + p_4}$ as in Equation 11:", + "bbox": [ + 496, + 507, + 892, + 657 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {t o t a l}} = \\alpha \\mathcal {L} _ {p _ {1}} + \\beta \\mathcal {L} _ {p _ {2}} + \\gamma \\mathcal {L} _ {p _ {3}} + \\zeta \\mathcal {L} _ {p _ {4}} + \\delta \\mathcal {L} _ {p _ {1} + p _ {2} + p _ {3} + p _ {4}} \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 517, + 665, + 890, + 680 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathcal{L}_{p_1},\\mathcal{L}_{p_2},\\mathcal{L}_{p_3}$ , and $\\mathcal{L}_{p_4}$ are the losses of each individual prediction maps. $\\alpha = \\beta = \\gamma = \\zeta = \\delta = 1.0$ are the weights assigned to each loss.", + "bbox": [ + 496, + 681, + 890, + 727 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Output segmentation maps aggregation: We consider the prediction map, $p_4$ , from the last stage of our decoder as the final segmentation map. Then, we obtain the final segmentation output by employing a Sigmoid function for binary or a Softmax function for multi-class segmentation.", + "bbox": [ + 496, + 728, + 890, + 803 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 498, + 815, + 633, + 832 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this section, we present the details of our implementation followed by a comparative analysis of our PVT-EMCAD-B0 and PVT-EMCAD-B2 against SOTA methods. Datasets and evaluation metrics are in Supplementary Section 7.", + "bbox": [ + 496, + 840, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "11773", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/5013ef6128dea5916bb2dcbb8344229ac8db20e6b47ae03b79b7b457e9a3f673.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Methods#Params#FLOPsPolypSkin LesionCellBUSIAvg.
ClinicColonETISKvasirBKAIISIC17ISIC18DSB18EM
UNet [44]34.53M65.53G92.1183.9576.8582.8785.0583.0786.6792.2395.4674.0485.23
UNet++ [62]9.16M34.65G92.1787.8877.4083.3684.0782.9887.4691.9795.4874.7685.75
AttnUNet [41]34.88M66.64G92.2086.4676.8483.4984.0783.6687.0592.2295.5574.4885.60
DeepLabv3+ [10]39.76M14.92G93.2491.9290.7389.0689.7483.8488.6492.1494.9676.8189.11
PraNet [20]32.55M6.93G91.7189.1683.8484.8285.5683.0388.5689.8992.3775.1486.41
CaraNet [38]46.64M11.48G94.0891.1990.2589.7489.7185.0290.1889.1592.7877.3488.94
UACANet-L [30]69.16M31.51G94.1691.0289.7790.1790.3583.7289.7688.8689.2876.9688.41
SSFormer-L [54]66.22M17.28G94.1892.1190.1691.4791.1485.2890.2592.0394.9578.7690.03
PolypPVT [17]25.11M5.30G94.1391.5389.9391.5691.1785.5690.3690.6994.4079.3589.87
TransUNet [8]105.32M38.52G93.9091.6387.7991.0889.1785.0089.1692.0495.2778.3089.33
SwinUNet [5]27.17M6.2G92.4289.2785.1089.5987.6183.9789.2691.0394.4777.3888.01
TransFuse [61]143.74M82.71G93.6290.3586.9190.2487.4784.8989.6290.8594.3579.3688.77
UNeXt [50]1.47M0.57G90.2083.8474.0377.8877.9382.7487.7886.0193.8174.7182.89
PVT-CASCADE [42]34.12M7.62G94.5391.6091.0392.0592.1485.5090.4192.3595.4279.2190.42
PVT-EMCAD-B0 (Ours)3.92M0.84G94.6091.7191.6591.9591.3085.6790.7092.4695.3579.8090.52
PVT-EMCAD-B2 (Ours)26.76M5.6G95.2192.3192.2992.7592.9685.9590.9692.7495.5380.2591.10
", + "bbox": [ + 80, + 89, + 890, + 324 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/fdfb3380337452f6615e0035d575276581729e149378c3ca5dcfb7cce8888e93.jpg", + "table_caption": [ + "Table 1. Results of binary medical image segmentation (i.e., polyp, skin lesion, cell, and breast cancer). We reproduce the results of SOTA methods using their publicly available implementation with our train-val-test splits of 80:10:10. #FLOPs of all the methods are reported for $256 \\times 256$ inputs, except Swin-UNet ( $224 \\times 224$ ). All results are averaged over five runs. Best results are shown in bold." + ], + "table_footnote": [], + "table_body": "
ArchitecturesAverageAortaGBKLKRLiverPCSPSM
DICE↑HD95↓mIoU↑
UNet [44]70.1144.6959.3984.0056.7072.4162.6486.9848.7381.4867.96
AttnUNet [41]71.7034.4761.3882.6161.9476.0770.4287.5446.7080.6767.66
R50+UNet [8]74.6836.87-84.1862.8479.1971.2993.3548.2384.4173.92
R50+AttnUNet [8]75.5736.97-55.9263.9179.2072.7193.5649.3787.1974.95
SSFormer [54]78.0125.7267.2382.7863.7480.7278.1193.5361.5387.0776.61
PolypPVT [17]78.0825.6167.4382.3466.1481.2173.7894.3759.3488.0579.4
TransUNet [8]77.6126.967.3286.5660.4380.5478.5394.3358.4787.0675.00
SwinUNet [5]77.5827.3266.8881.7665.9582.3279.2293.7353.8188.0475.79
MT-UNet [53]78.5926.59-87.9264.9981.4777.2993.0659.4687.7576.81
MISSformer [25]81.9618.20-86.9968.6585.2182.0094.4165.6791.9280.81
PVT-CASCADE [42]81.0620.2370.8883.0170.5982.2380.3794.0864.4390.183.69
TransCASCADE [42]82.6817.3473.4886.6368.4887.6684.5694.4365.3390.7983.52
PVT-EMCAD-B0 (Ours)81.9717.3972.6487.2166.6287.4883.9694.5762.0092.6681.22
PVT-EMCAD-B2 (Ours)83.6315.6874.6588.1468.8788.0884.1095.2668.5192.1783.92
", + "bbox": [ + 93, + 377, + 877, + 619 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 2. Results of abdomen organ segmentation on Synapse Multi-organ dataset. DICE scores are reported for individual organs. Results of UNet, AttnUNet, PolypPVT, SSFormerPVT, TransUNet, and SwinUNet are taken from [42]. $\\uparrow (\\downarrow)$ denotes the higher (lower) the better. $^{\\prime} - ^{\\prime}$ means missing data from the source. EMCAD results are averaged over five runs. Best results are shown in bold.", + "bbox": [ + 76, + 623, + 893, + 666 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Implementation details", + "text_level": 1, + "bbox": [ + 76, + 676, + 290, + 694 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We implement our network and conduct experiments using Pytorch 1.11.0 on a single NVIDIA RTX A6000 GPU with 48GB of memory. We utilize ImageNet [16] pretrained PVTv2-b0 and PVTv2-b2 [56] as encoders. In the MSDC of our decoder, we set the multi-scale kernels [1,3,5] through an ablation study. We use the parallel arrangement of depth-wise convolutions in all experiments. Our models are trained using the AdamW optimizer [36] with a learning rate and weight decay of $1e - 4$ . We generally train for 200 epochs with a batch size of 16, except for Synapse multi-organ (300 epochs, batch size 6) and ACDC cardiac organ (400 epochs, batch size 12), saving the best model based on the DICE score. We resize images", + "bbox": [ + 75, + 703, + 470, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "to $352 \\times 352$ and use a multi-scale $\\{0.75, 1.0, 1.25\\}$ training strategy with a gradient clip limit of 0.5 for ClinicDB [3], Kvasir [29], ColonDB [51], ETIS [51], BKAI [40], ISIC17 [15], and ISIC18 [15], while we resize images to $256 \\times 256$ for BUSI [1], EM [6], and DSB18 [4]. For Synapse and ACDC datasets, images are resized to $224 \\times 224$ , with random rotation and flipping augmentations, optimizing a combined Cross-entropy (0.3) and DICE (0.7) loss. For binary segmentation, we utilize the combined weighted BinaryCrossEntropy (BCE) and weighted IoU loss function.", + "bbox": [ + 496, + 676, + 890, + 830 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Results", + "text_level": 1, + "bbox": [ + 500, + 844, + 591, + 858 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We compare our architectures (i.e., PVT-EMCAD-B0 and PVT-EMCAD-B2) with SOTA CNN and transformer-based", + "bbox": [ + 500, + 869, + 893, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "11774", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/68e95713cdd7aa1342c98b587e762e015c98e20ad9f521d47ceae1aa69432ee7.jpg", + "image_caption": [ + "Figure 3. Average DICE scores vs. #Params for different methods over 10 binary medical image segmentation datasets. As shown, our proposed approaches (PVT-EMCAD-B0 and PVT-EMCAD-B2) have the fewest parameters, yet the highest DICE scores." + ], + "image_footnote": [], + "bbox": [ + 109, + 89, + 439, + 277 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/cc9b5fe31b2429853d4989b4ae359486428262b5afc60e7caacd71fa24d74474.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodsAvg. DICERVMyoLV
R50+UNet [8]87.5587.1080.6394.92
R50+AttnUNet [8]86.7587.5879.2093.47
ViT+CUP [8]81.4581.4670.7192.18
R50+ViT+CUP [8]87.5786.0781.8894.75
TransUNet [8]89.7186.6787.2795.18
SwinUNet [5]88.0785.7784.4294.03
MT-UNet [53]90.4386.6489.0495.62
MISSformer [25]90.8689.5588.0494.99
PVT-CASCADE [42]91.4689.9788.995.50
TransCASCade [42]91.6390.2589.1495.50
Cascaded MERIT [43]91.8590.2389.5395.80
PVT-EMCAD-B0 (Ours)91.34±0.289.3788.9995.65
PVT-EMCAD-B2 (Ours)92.12±0.290.6589.6896.02
", + "bbox": [ + 78, + 343, + 472, + 547 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "segmentation methods on 12 datasets that belong to six medical image segmentation tasks. Qualitative results are in the Supplementary Section 7.3.", + "bbox": [ + 75, + 606, + 468, + 654 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2.1 Results of binary medical image segmentation", + "text_level": 1, + "bbox": [ + 75, + 664, + 450, + 681 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Results for different methods on 10 binary medical image segmentation datasets are shown in Table 1 and Figure 1. Our PVT-EMCAD-B2 attains the highest average DICE score $(91.10\\%)$ with only $26.76\\mathrm{M}$ parameters and 5.6G FLOPs. The multi-scale depth-wise convolution in our EMCAD decoder, combined with the transformer encoder, contributes to these performance gains.", + "bbox": [ + 75, + 688, + 468, + 794 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Polyp segmentation: Table 1 reveals that our PVT-EMCAD-B2 surpasses all SOTA methods in five polyp segmentation datasets. PVT-EMCAD-B2 achieves DICE score improvements of $1.08\\%$ , $0.78\\%$ , $2.36\\%$ , $1.19\\%$ , and $1.79\\%$ over PolypPVT in ClinicDB, ColonDB, ETIS, Kvasir, and BKAI-IGI, despite having slightly more parameters and FLOPs. The smallest model UNeXt, ex", + "bbox": [ + 75, + 795, + 470, + 901 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/2d653adeec59f708f8225a0f18e1acdfba7cd70d71d3dab6b3fe686aa3f65563.jpg", + "table_caption": [ + "Table 3. Results of cardiac organ segmentation on ACDC dataset. DICE scores $(\\%)$ are reported for individual organs. We get the results of SwinUNet from [42]. Best results are shown in bold." + ], + "table_footnote": [], + "table_body": "
Components#FLOPs(G)#Params (M)Avg DICE
CascadedLGAGMSCAM224256
NoNoNo00080.10±0.2
YesNoNo0.1000.1310.22481.08±0.2
YesYesNo0.1080.1410.23581.92±0.2
YesNoYes0.3730.4871.89882.86±0.3
YesYesYes0.3810.4981.9183.63±0.3
", + "bbox": [ + 501, + 89, + 893, + 186 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 4. Effect of different components of EMCAD with PVTv2-b2 encoder on Synapse multi-organ dataset. #FLOPs are reported for input resolution of $224 \\times 224$ and $256 \\times 256$ . All results are averaged over five runs. Best results are shown in bold.", + "bbox": [ + 496, + 189, + 893, + 247 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "hibits the worst performance in all five polyp segmentation datasets. Our smaller model with only 3.92M parameters and 0.84G FLOPs also outperforms all the methods except PVT-CASCADE (in Kvasir and BKAI-IGH) and SSFormer-L (in ColonDB), which achieve the best performance among SOTA methods. In conclusion, our PVT-EMCAD-B2 achieves the new SOTA results in these five polyp segmentation datasets.", + "bbox": [ + 496, + 255, + 890, + 376 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Skin lesion segmentation: Table 1 shows PVT-EMCAD-B2's strong performance on ISIC17 and ISIC18 skin lesion segmentation datasets, achieving DICE scores of $85.95\\%$ and $90.96\\%$ , surpassing DeepLabV3+ by $2.11\\%$ and $2.32\\%$ . It also beats the nearest method PVT-CASCADE by $0.45\\%$ and $0.55\\%$ in ISIC17 and ISIC18, respectively, though our decoder is significantly more efficient than CASCADE. Our PVT-EMCAD-B0 also shows huge potential in point care applications like skin lesion segmentation with only 3.92M parameters and 0.84G FLOPs.", + "bbox": [ + 496, + 377, + 892, + 527 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Cell segmentation: To evaluate our method's effectiveness in biological imaging, we use DSB18 [4] for cell nuclei and EM [6] for cell structure segmentation. As Table 1 indicates, our PVT-EMCAD-B2 sets a SOTA benchmark in cell nuclei segmentation on DSB18, outperforming DeepLabv3+, TransFuse, and PVT-CASCADE. On the EM dataset, PVT-EMCAD-B2 secures the second-best DICE score (95.53%), offering significantly lower computational costs than the top-performing AttnUNet (95.55%).", + "bbox": [ + 496, + 529, + 893, + 665 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Breast cancer segmentation: We conduct experiments on the BUSI dataset for breast cancer segmentation in ultrasound images. Our PVT-EMCAD-B2 achieves the SOTA DICE score $(80.25\\%)$ on this dataset. Furthermore, our PVT-EMCAD-B0 outperforms the computationally similar method UNeXt by a notable margin of $5.54\\%$ .", + "bbox": [ + 496, + 665, + 893, + 756 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2.2 Results of abdomen organ segmentation", + "text_level": 1, + "bbox": [ + 498, + 770, + 831, + 786 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 2 shows that our PVT-EMCAD-B2 excels in abdomen organ segmentation on the Synapse multi-organ dataset, achieving the highest average DICE score of $83.63\\%$ and surpassing all SOTA CNN- and transformer-based methods. It outperforms PVT-CASCADE by $2.57\\%$ in DICE score and 4.55 in HD95 distance, indicating superior organ boundary location. Our EMCAD decoder boosts individ", + "bbox": [ + 496, + 794, + 893, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "11775", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/960dbdb51d3a1ddc506c2d15b7d43fb9faa404d4b0f8e5c2be86685d76f5ca3f.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Conv. kernels[1][3][5][1,3][3,3][1,3,5][3,3,3][3,5,7][1,3,5,7][1,3,5,7,9]
Synapse82.4382.7982.7482.9882.8183.6382.9283.1183.5783.34
ClinicDB94.8194.9094.9895.1395.0695.2195.1595.0395.1895.07
", + "bbox": [ + 109, + 88, + 857, + 147 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/795666650d4a54a2666e7382c68cfd032201b42d5155ab078918ba97932efa6c.jpg", + "table_caption": [ + "Table 5. Effect of multi-scale kernels in the depth-wise convolution of MSDC on ClinicDB and Synapse multi-organ datasets. We use the PVTv2-b2 encoder for these experiments. All results are averaged over five runs. Best results are highlighted in bold." + ], + "table_footnote": [ + "Table 6. Comparison with the baseline decoder on Synapse Multi-organ dataset. We only report the #FLOPs (with input resolution of $224 \\times 224$ ) and the #parameters of the decoders. All the results are averaged over five runs. Best results are shown in bold." + ], + "table_body": "
EncodersDecoders#FLOPs(G)#Params(M)DICE (%)
PVTv2-B0CASCADE0.4392.3280.54
PVTv2-B0EMCAD (Ours)0.1100.50781.97
PVTv2-B2CASCADE1.939.2782.78
PVTv2-B2EMCAD (Ours)0.3811.9183.63
", + "bbox": [ + 78, + 189, + 470, + 267 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "ual organ segmentation, significantly outperforming SOTA methods on six of eight organs.", + "bbox": [ + 76, + 330, + 467, + 361 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.2.3 Results of cardiac organ segmentation", + "text_level": 1, + "bbox": [ + 76, + 364, + 395, + 380 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 3 shows the DICE scores of our PVT-EMCAD-B2 and PVT-EMCAD-B0 along with other SOTA methods, on the MRI images of the ACDC dataset for cardiac organ segmentation. Our PVT-EMCAD-B2 achieves the highest average DICE score of $92.12\\%$ , thus improving about $0.27\\%$ over Cascaded MERIT though our network has significantly lower computational cost. Besides, PVT-EMCAD-B2 has better DICE scores in all three organ segmentations.", + "bbox": [ + 75, + 388, + 467, + 508 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Ablation Studies", + "text_level": 1, + "bbox": [ + 76, + 518, + 240, + 534 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this section, we conduct ablation studies to explore different aspects of our architectures and the experimental framework. More ablations are in Supplementary Section 8.", + "bbox": [ + 75, + 542, + 467, + 588 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.1. Effect of different components of EMCAD", + "text_level": 1, + "bbox": [ + 76, + 592, + 436, + 608 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We conduct a set of experiments on the Synapse multi-organ dataset to understand the effect of different components of our EMCAD decoder. We start with only the encoder and add different modules such as Cascaded structure, LGAG, and MSCAM to understand their effect. Table 4 exhibits that the cascaded structure of the decoder helps to improve performance over the non-cascaded one. The incorporation of LGAG and MSCAM improves performance, however, MSCAM proves to be more effective. When both the LGAG and MSCAM modules are used together, it produces the best DICE score of $83.63\\%$ . It is also evident that there is about $3.53\\%$ improvement in the DICE score with an additional 0.381G FLOPs and 1.91M parameters.", + "bbox": [ + 75, + 616, + 467, + 811 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.2. Effect of multi-scale kernels in MSCAM", + "text_level": 1, + "bbox": [ + 76, + 816, + 419, + 832 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We have conducted another set of experiments on Synapse multi-organ and ClinicDB datasets to understand the effect of different multi-scale kernels used for depth-wise convolutions in MSDC. Table 5 reports these results which show", + "bbox": [ + 75, + 839, + 467, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "that performance improves from $1 \\times 1$ to $3 \\times 3$ kernel. When $1 \\times 1$ kernel is used together with $3 \\times 3$ it improves more than when using them alone. However, when two $3 \\times 3$ kernels are used together, performance drops. The incorporation of a $5 \\times 5$ kernel with $1 \\times 1$ and $3 \\times 3$ kernels further improves the performance and it achieves the best results in both Synapse multi-organ and ClinicDB datasets. If we add additional larger kernels (e.g., $7 \\times 7$ , $9 \\times 9$ ), the performance of both datasets drops. Based on these empirical observations, we choose [1, 3, 5] kernels in all our experiments.", + "bbox": [ + 496, + 191, + 890, + 343 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.3. Comparison with the baseline decoder", + "text_level": 1, + "bbox": [ + 498, + 352, + 828, + 368 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In Table 6, we report the experimental results with the computational complexity of our EMCAD decoder and a baseline decoder, namely CASCADE. From Table 6, we can see that our EMCAD decoder with PVTv2-b2 requires $80.3\\%$ fewer FLOPs and $79.4\\%$ fewer parameters to outperform (by $0.85\\%$ ) the respective CASCADE decoder. Similarly, our EMCAD decoder with PVTv2-B0 achieves $1.43\\%$ better DICE score than the CASCADE decoder with $78.1\\%$ fewer parameters and $74.9\\%$ fewer FLOPs.", + "bbox": [ + 496, + 376, + 890, + 512 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusions", + "text_level": 1, + "bbox": [ + 500, + 522, + 625, + 537 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we have presented EMCAD, a new and efficient multi-scale convolutional attention decoder designed for multi-stage feature aggregation and refinement in medical image segmentation. EMCAD employs a multi-scale depth-wise convolution block, which is key for capturing diverse scale information within feature maps, a critical factor for precision in medical image segmentation. This design choice, using depth-wise convolutions instead of standard $3 \\times 3$ convolution blocks, makes EMCAD notably efficient.", + "bbox": [ + 496, + 547, + 890, + 683 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Our experiments reveal that EMCAD surpasses the recent CASCADE decoder in DICE scores with $79.4\\%$ fewer parameters and $80.3\\%$ less FLOPs. Our extensive experiments also confirm EMCAD's superior performance compared to SOTA methods across 12 public datasets covering six different 2D medical image segmentation tasks. EMCAD's compatibility with smaller encoders makes it an excellent fit for point-of-care applications while maintaining high performance. We anticipate that our EMCAD decoder will be a valuable asset in enhancing a variety of medical image segmentation and semantic segmentation tasks.", + "bbox": [ + 496, + 685, + 890, + 849 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements: This work is supported in part by the NSF grant CNS 2007284, and in part by the iMAGiNE Consortium (https://imagine.utexas.edu/).", + "bbox": [ + 496, + 851, + 890, + 893 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "11776", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Walid Al-Dhabyani, Mohammed Gomaa, Hussien Khaled, and Aly Fahmy. Dataset of breast ultrasound images. Data in brief, 28:104863, 2020. 6, 1", + "[2] Vijay Badrinarayanan, Alex Kendall, and Roberto Cipolla. Segnet: A deep convolutional encoder-decoder architecture for image segmentation. IEEE Trans. Pattern Anal. Mach. Intell., 39(12):2481-2495, 2017. 3", + "[3] Jorge Bernal, F Javier Sánchez, Gloria Fernández-Esparrach, Debora Gil, Cristina Rodríguez, and Fernando Vilarino. Wm-dova maps for accurate polyp highlighting in colonoscopy: Validation vs. saliency maps from physicians. Comput. Med. Imaging Graph., 43:99-111, 2015. 6, 1", + "[4] Juan C Caicedo, Allen Goodman, Kyle W Karhohs, Beth A Cimini, Jeanelle Ackerman, Marzieh Haghighi, CherKeng Heng, Tim Becker, Minh Doan, Claire McQuin, et al. Nucleus segmentation across imaging experiments: the 2018 data science bowl. Nature methods, 16(12):1247-1253, 2019. 6, 7, 1", + "[5] Hu Cao, Yueyue Wang, Joy Chen, Dongsheng Jiang, Xiaopeng Zhang, Qi Tian, and Manning Wang. Swin-unet: Unet-like pure transformer for medical image segmentation. arXiv preprint arXiv:2105.05537, 2021. 1, 3, 6, 7", + "[6] Albert Cardona, Stephan Saalfeld, Stephan Preibisch, Benjamin Schmid, Anchi Cheng, Jim Pulokas, Pavel Tomancak, and Volker Hartenstein. An integrated micro-and macroarchitectural analysis of the drosophila brain by computer-assisted serial section electron microscopy. PLoS biology, 8(10):e1000502, 2010. 6, 7, 1", + "[7] Gongping Chen, Lei Li, Yu Dai, Jianxun Zhang, and Moi Hoon Yap. Aau-net: an adaptive attention u-net for breast lesions segmentation in ultrasound images. IEEE Trans. Med. Imaging, 2022. 2", + "[8] Jieneng Chen, Yongyi Lu, Qihang Yu, Xiangde Luo, Ehsan Adeli, Yan Wang, Le Lu, Alan L Yuille, and Yuyin Zhou. Transunet: Transformers make strong encoders for medical image segmentation. arXiv preprint arXiv:2102.04306, 2021. 1, 2, 3, 6, 7", + "[9] Long Chen, Hanwang Zhang, Jun Xiao, Liqiang Nie, Jian Shao, Wei Liu, and Tat-Seng Chua. Sca-cnn: Spatial and channel-wise attention in convolutional networks for image captioning. In IEEE Conf. Comput. Vis. Pattern Recog., pages 5659-5667, 2017. 3, 4", + "[10] Liang-Chieh Chen, George Papandreou, Iasonas Kokkinos, Kevin Murphy, and Alan L Yuille. Deeplab: Semantic image segmentation with deep convolutional nets, atrous convolution, and fully connected crfs. IEEE Trans. Pattern Anal. Mach. Intell., 40(4):834-848, 2017. 2, 6", + "[11] Liang-Chieh Chen, Yukun Zhu, George Papandreou, Florian Schroff, and Hartwig Adam. Encoder-decoder with atrous separable convolution for semantic image segmentation. In Eur. Conf. Comput. Vis., pages 801-818, 2018. 2", + "[12] Shuhan Chen, Xiuli Tan, Ben Wang, and Xuelong Hu. Reverse attention for salient object detection. In Eur. Conf. Comput. Vis., pages 234-250, 2018. 1", + "[13] Xiangxiang Chu, Zhi Tian, Bo Zhang, Xinlong Wang, Xiaolin Wei, Huaxia Xia, and Chunhua Shen. Conditional po" + ], + "bbox": [ + 78, + 114, + 468, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "sitional encodings for vision transformers. arXiv preprint arXiv:2102.10882, 2021. 1", + "[14] Noel Codella, Veronica Rotemberg, Philipp Tschandl, M Emre Celebi, Stephen Dusza, David Gutman, Brian Helba, Aadi Kalloo, Konstantinos Liopyris, Michael Marchetti, et al. Skin lesion analysis toward melanoma detection 2018: A challenge hosted by the international skin imaging collaboration (isic). arXiv preprint arXiv:1902.03368, 2019. 1", + "[15] Noel CF Codella, David Gutman, M Emre Celebi, Brian Helba, Michael A Marchetti, Stephen W Dusza, Aadi Kalloo, Konstantinos Liopyris, Nabin Mishra, Harald Kittler, et al. Skin lesion analysis toward melanoma detection: A challenge at the 2017 international symposium on biomedical imaging (isbi), hosted by the international skin imaging collaboration (isic). In IEEE Int. Symp. Biomed. Imaging, pages 168-172. IEEE, 2018. 6, 1", + "[16] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In IEEE Conf. Comput. Vis. Pattern Recog., pages 248-255. IEEE, 2009. 6", + "[17] Bo Dong, Wenhai Wang, Deng-Ping Fan, Jinpeng Li, Huazhu Fu, and Ling Shao. Polyp-pvt: Polyp segmentation with pyramid vision transformers. arXiv preprint arXiv:2108.06932, 2021. 1, 3, 5, 6", + "[18] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 1, 2", + "[19] Isensee et al. nnu-net: a self-configuring method for deep learning-based biomedical image segmentation. Nature methods, 18(2):203-211, 2021. 1, 2, 3", + "[20] Deng-Ping Fan, Ge-Peng Ji, Tao Zhou, Geng Chen, Huazhu Fu, Jianbing Shen, and Ling Shao. Pranet: Parallel reverse attention network for polyp segmentation. In Int. Conf. Med. Image Comput. Comput. Assist. Interv., pages 263-273. Springer, 2020. 1, 3, 6", + "[21] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In IEEE Conf. Comput. Vis. Pattern Recog., pages 770-778, 2016. 2", + "[22] Andrew G Howard, Menglong Zhu, Bo Chen, Dmitry Kalenichenko, Weijun Wang, Tobias Weyand, Marco Andreetto, and Hartwig Adam. Mobilenets: Efficient convolutional neural networks for mobile vision applications. arXiv preprint arXiv:1704.04861, 2017. 2", + "[23] Jie Hu, Li Shen, and Gang Sun. Squeeze-and-excitation networks. In IEEE Conf. Comput. Vis. Pattern Recog., pages 7132-7141, 2018. 2, 3", + "[24] Huimin Huang, Lanfen Lin, Ruofeng Tong, Hongjie Hu, Qiaowei Zhang, Yutaro Iwamoto, Xianhua Han, Yen-Wei Chen, and Jian Wu. Unet 3+: A full-scale connected unet for medical image segmentation. In ICASSP, pages 1055–1059. IEEE, 2020. 1, 2", + "[25] Xiaohong Huang, Zhifang Deng, Dandan Li, and Xueguang Yuan. Missformer: An effective medical image segmentation transformer. arXiv preprint arXiv:2109.07162, 2021. 6, 7" + ], + "bbox": [ + 503, + 92, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "11777", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[26] Nabil Ibtehaz and Daisuke Kihara. Acc-unet: A completely convolutional unet model for the 2020s. In Int. Conf. Med. Image Comput. Comput. Assist. Interv., pages 692-702. Springer, 2023. 2", + "[27] Sergey Ioffe and Christian Szegedy. Batch normalization: Accelerating deep network training by reducing internal covariate shift. In Int. Conf. Mach. Learn., pages 448-456. pmlr, 2015. 3", + "[28] Md Amirul Islam, Sen Jia, and Neil DB Bruce. How much position information do convolutional neural networks encode? arXiv preprint arXiv:2001.08248, 2020. 1", + "[29] Debesh Jha, Pia H Smedsrud, Michael A Riegler, Pål Halvorsen, Thomas de Lange, Dag Johansen, and Håvard D Johansen. Kvasir-seg: A segmented polyp dataset. In Int. Conf. Multimedia Model., pages 451-462. Springer, 2020. 6, 1", + "[30] Taehun Kim, Hyemin Lee, and Daijin Kim. Uacanet: Uncertainty augmented context attention for polyp segmentation. In ACM Int. Conf. Multimedia, pages 2167-2175, 2021. 6", + "[31] Alex Krizhevsky and Geoff Hinton. Convolutional deep belief networks on CIFar-10. Unpublished manuscript, 40(7): 1-9, 2010. 4", + "[32] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convolutional neural networks. Adv. Neural Inform. Process. Syst., 25, 2012. 2", + "[33] Xian Lin, Zengqiang Yan, Xianbo Deng, Chuansheng Zheng, and Li Yu. Convformer: Plug-and-play cnn-style transformers for improving medical image segmentation. In Int. Conf. Med. Image Comput. Comput. Assist. Interv., pages 642-651. Springer, 2023. 1", + "[34] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Int. Conf. Comput. Vis., pages 10012-10022, 2021. 1, 2, 3", + "[35] Zhuang Liu, Hanzi Mao, Chao-Yuan Wu, Christoph Feichtenhofer, Trevor Darrell, and Saining Xie. A convnet for the 2020s. In IEEE Conf. Comput. Vis. Pattern Recog., pages 11976-11986, 2022. 2", + "[36] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 6", + "[37] Ange Lou, Shuyue Guan, and Murray Loew. Dc-unet: rethinking the u-net architecture with dual channel efficient cnn for medical image segmentation. In Med. Imaging 2021: Image Process., pages 758-768. SPIE, 2021. 1, 2", + "[38] Ange Lou, Shuyue Guan, Hanseok Ko, and Murray H Loew. Caranet: context axial reverse attention network for segmentation of small medical objects. In Med. Imaging 2022: Image Process., pages 81-92. SPIE, 2022. 6", + "[39] Vinod Nair and Geoffrey E Hinton. Rectified linear units improve restricted boltzmann machines. In Int. Conf. Mach. Learn., pages 807-814, 2010. 3", + "[40] Phan Ngoc Lan, Nguyen Sy An, Dao Viet Hang, Dao Van Long, Tran Quang Trung, Nguyen Thi Thuy, and Dinh Viet Sang. Neounet: Towards accurate colon polyp segmentation and neoplasm detection. In Adv. Vis. Comput. - Int. Symp., pages 15-28. Springer, 2021. 6, 1" + ], + "bbox": [ + 78, + 90, + 468, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[41] Ozan Oktay, Jo Schlemper, Loic Le Folgoc, Matthew Lee, Mattias Heinrich, Kazunari Misawa, Kensaku Mori, Steven McDonagh, Nils Y Hammerla, Bernhard Kainz, et al. Attention u-net: Learning where to look for the pancreas. arXiv preprint arXiv:1804.03999, 2018. 1, 2, 3, 6", + "[42] Md Mostafijur Rahman and Radu Marculescu. Medical image segmentation via cascaded attention decoding. In IEEE/CVF Winter Conf. Appl. Comput. Vis., pages 6222-6231, 2023. 1, 3, 4, 5, 6, 7", + "[43] Md Mostafijur Rahman and Radu Marculescu. Multi-scale hierarchical vision transformer with cascaded attention decoding for medical image segmentation. In Med. Imaging Deep Learn., 2023. 1, 3, 5, 7", + "[44] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Int. Conf. Med. Image Comput. Comput. Assist. Interv., pages 234-241. Springer, 2015. 1, 2, 6", + "[45] Mark Sandler, Andrew Howard, Menglong Zhu, Andrey Zh-moginov, and Liang-Chieh Chen. Mobilenetv2: Inverted residuals and linear bottlenecks. In IEEE Conf. Comput. Vis. Pattern Recog., pages 4510-4520, 2018. 2, 4", + "[46] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556, 2014. 2", + "[47] Christian Szegedy, Wei Liu, Yangqing Jia, Pierre Sermanet, Scott Reed, Dragomir Anguelov, Dumitru Erhan, Vincent Vanhoucke, and Andrew Rabinovich. Going deeper with convolutions. In IEEE Conf. Comput. Vis. Pattern Recog., pages 1-9, 2015. 2", + "[48] Mingxing Tan and Quoc Le. Efficientnet: Rethinking model scaling for convolutional neural networks. In Int. Conf. Mach. Learn., pages 6105-6114. PMLR, 2019. 2", + "[49] Zhengzhong Tu, Hossein Talebi, Han Zhang, Feng Yang, Peyman Milanfar, Alan Bovik, and Yinxiao Li. Maxvit: Multi-axis vision transformer. In Eur. Conf. Comput. Vis., pages 459-479. Springer, 2022. 1, 2", + "[50] Jeya Maria Jose Valanarasu and Vishal M Patel. Unext: Mlp-based rapid medical image segmentation network. In Int. Conf. Med. Image Comput. Comput. Assist. Interv., pages 23-33. Springer, 2022. 6, 1", + "[51] David Vázquez, Jorge Bernal, F Javier Sánchez, Gloria Fernández-Esparrach, Antonio M López, Adriana Romero, Michal Drozdzal, and Aaron Courville. A benchmark for endoluminal scene segmentation of colonoscopy images. J. Healthc. Eng., 2017, 2017. 6, 1", + "[52] Haonan Wang, Peng Cao, Jiaqi Wang, and Osmar R Zaiane. Uctransnet: rethinking the skip connections in u-net from a channel-wise perspective with transformer. In AAAI, pages 2441-2449, 2022. 1, 3", + "[53] Hongyi Wang, Shiao Xie, Lanfen Lin, Yutaro Iwamoto, Xian-Hua Han, Yen-Wei Chen, and Ruofeng Tong. Mixed transformer u-net for medical image segmentation. In ICASSP, pages 2390–2394. IEEE, 2022. 6, 7", + "[54] Jinfeng Wang, Qiming Huang, Feilong Tang, Jia Meng, Jionglong Su, and Sifan Song. Stepwise feature fusion: Local guides global. arXiv preprint arXiv:2203.03635, 2022. 1, 6" + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "11778", + "bbox": [ + 480, + 945, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[55] Wenhai Wang, Enze Xie, Xiang Li, Deng-Ping Fan, Kaitao Song, Ding Liang, Tong Lu, Ping Luo, and Ling Shao. Pyramid vision transformer: A versatile backbone for dense prediction without convolutions. In Int. Conf. Comput. Vis., pages 568-578, 2021. 1, 2", + "[56] Wenhai Wang, Enze Xie, Xiang Li, Deng-Ping Fan, Kaitao Song, Ding Liang, Tong Lu, Ping Luo, and Ling Shao. Pvt v2: Improved baselines with pyramid vision transformer. Comput. Vis. Media, 8(3):415-424, 2022. 1, 2, 3, 5, 6", + "[57] Sanghyun Woo, Jongchan Park, Joon-Young Lee, and In So Kweon. Cbam: Convolutional block attention module. In Eur. Conf. Comput. Vis., pages 3–19, 2018. 1, 3, 4", + "[58] Enze Xie, Wenhai Wang, Zhiding Yu, Anima Anandkumar, Jose M Alvarez, and Ping Luo. Segformer: Simple and efficient design for semantic segmentation with transformers. Adv. Neural Inform. Process. Syst., 34:12077-12090, 2021. 2, 3", + "[59] Weihao Yu, Mi Luo, Pan Zhou, Chenyang Si, Yichen Zhou, Xinchao Wang, Jiashi Feng, and Shuicheng Yan. Metaformer is actually what you need for vision. In IEEE Conf. Comput. Vis. Pattern Recog., pages 10819-10829, 2022. 1", + "[60] Xiangyu Zhang, Xinyu Zhou, Mengxiao Lin, and Jian Sun. Shufflenet: An extremely efficient convolutional neural network for mobile devices. In IEEE Conf. Comput. Vis. Pattern Recog., pages 6848-6856, 2018. 4", + "[61] Yundong Zhang, Huiye Liu, and Qiang Hu. Transfuse: Fusing transformers and cnns for medical image segmentation. In Int. Conf. Med. Image Comput. Comput. Assist. Interv., pages 14-24. Springer, 2021. 1, 3, 6", + "[62] Zongwei Zhou, Md Mahfuzur Rahman Siddiquee, Nima Tajbakhsh, and Jianming Liang. Unet++: A nested u-net architecture for medical image segmentation. In Deep Learn. Med. Image Anal. Multimodal Learn. Clin. Decis. Support, pages 3-11. Springer, 2018. 1, 2, 6" + ], + "bbox": [ + 78, + 90, + 468, + 571 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11779", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2024/EMCAD_ Efficient Multi-scale Convolutional Attention Decoding for Medical Image Segmentation/32047b53-ccf5-41e9-bc53-4e712089f958_model.json b/2024/EMCAD_ Efficient Multi-scale Convolutional Attention Decoding for Medical Image Segmentation/32047b53-ccf5-41e9-bc53-4e712089f958_model.json new file mode 100644 index 0000000000000000000000000000000000000000..b5d920406f31a2f1f5ed48014b19bc05748cfcf3 --- /dev/null +++ b/2024/EMCAD_ Efficient Multi-scale Convolutional Attention Decoding for Medical Image Segmentation/32047b53-ccf5-41e9-bc53-4e712089f958_model.json @@ -0,0 +1,2290 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.098, + 0.13, + 0.877, + 0.178 + ], + "angle": 0, + "content": "EMCAD: Efficient Multi-scale Convolutional Attention Decoding for Medical Image Segmentation" + }, + { + "type": "text", + "bbox": [ + 0.237, + 0.204, + 0.733, + 0.221 + ], + "angle": 0, + "content": "Md Mostafijur Rahman, Mustafa Munir, and Radu Marculescu" + }, + { + "type": "text", + "bbox": [ + 0.352, + 0.223, + 0.619, + 0.239 + ], + "angle": 0, + "content": "The University of Texas at Austin" + }, + { + "type": "text", + "bbox": [ + 0.406, + 0.24, + 0.566, + 0.256 + ], + "angle": 0, + "content": "Austin, Texas, USA" + }, + { + "type": "text", + "bbox": [ + 0.29, + 0.26, + 0.674, + 0.275 + ], + "angle": 0, + "content": "mostafijur.rahman, mmunir, radum@utexas.edu" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.31, + 0.314, + 0.326 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.341, + 0.473, + 0.735 + ], + "angle": 0, + "content": "An efficient and effective decoding mechanism is crucial in medical image segmentation, especially in scenarios with limited computational resources. However, these decoding mechanisms usually come with high computational costs. To address this concern, we introduce EMCAD, a new efficient multi-scale convolutional attention decoder, designed to optimize both performance and computational efficiency. EMCAD leverages a unique multi-scale depth-wise convolution block, significantly enhancing feature maps through multi-scale convolutions. EMCAD also employs channel, spatial, and grouped (large-kernel) gated attention mechanisms, which are highly effective at capturing intricate spatial relationships while focusing on salient regions. By employing group and depth-wise convolution, EMCAD is very efficient and scales well (e.g., only 1.91M parameters and 0.381G FLOPs are needed when using a standard encoder). Our rigorous evaluations across 12 datasets that belong to six medical image segmentation tasks reveal that EMCAD achieves state-of-the-art (SOTA) performance with \\(79.4\\%\\) and \\(80.3\\%\\) reduction in #Params and #FLOPs, respectively. Moreover, EMCAD's adaptability to different encoders and versatility across segmentation tasks further establish EMCAD as a promising tool, advancing the field towards more efficient and accurate medical image analysis. Our implementation is available at https://github.com/SLDGroup/EMCAD." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.755, + 0.21, + 0.77 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.78, + 0.471, + 0.904 + ], + "angle": 0, + "content": "In the realm of medical diagnostics and therapeutic strategies, automated segmentation of medical images is vital, as it classifies pixels to identify critical regions such as lesions, tumors, or entire organs. A variety of U-shaped convolutional neural network (CNN) architectures [20, 24, 37, 41, 44, 62], notably UNet [44], UNet++ [62], UNet3+ [24], and nnU-Net [19], have become standard techniques for this purpose, achieving high-quality, high-resolution segmen" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.311, + 0.895, + 0.416 + ], + "angle": 0, + "content": "tation output. Attention mechanisms [12, 17, 20, 41, 57] have also been integrated into these models to enhance feature maps and improve pixel-level classification. Although attention-based models have shown improved performance, they still face significant challenges due to the computationally expensive convolutional blocks that are typically used in conjunction with attention mechanisms." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.417, + 0.895, + 0.657 + ], + "angle": 0, + "content": "Recently, vision transformers [18] have shown promise in medical image segmentation tasks [5, 8, 17, 42, 43, 52, 54, 61] by capturing long-range dependencies among pixels through Self-attention (SA) mechanisms. Hierarchical vision transformers like Swin [34], PVT [55, 56], MaxViT [49], MERIT [43], ConvFormer [33], and MetaFormer [59] have been introduced to further improve the performance in this field. While the SA excels at capturing global information, it is less adept at understanding the local spatial context [13, 28]. To address this limitation, some approaches have integrated local convolutional attention within the decoders to better grasp spatial details. Nevertheless, these methods can still be computationally demanding because they frequently employ costly convolutional blocks. This limits their applicability to real-world scenarios where computational resources are restricted." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.659, + 0.895, + 0.779 + ], + "angle": 0, + "content": "To address the aforementioned limitations, we introduce EMCAD, an efficient multi-scale convolutional attention decoding using a new multi-scale depth-wise convolution block. More precisely, EMCAD enhances the feature maps via efficient multi-scale convolutions, while incorporating complex spatial relationships and local attention through the use of channel, spatial, and grouped (large-kernel) gated attention mechanisms. Our contributions are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.78, + 0.895, + 0.902 + ], + "angle": 0, + "content": "- New Efficient Multi-scale Convolutional Decoder: We introduce an efficient multi-scale cascaded fullyconvolutional attention decoder (EMCAD) for 2D medical image segmentation; this takes the multi-stage features of vision encoders and progressively enhances the multi-scale and multi-resolution spatial representations. EMCAD has only 0.506M parameters and 0.11G FLOPs for a tiny encoder with #channels = [32, 64, 160, 256]," + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "11769" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.11, + 0.09, + 0.442, + 0.274 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.278, + 0.472, + 0.334 + ], + "angle": 0, + "content": "Figure 1. Average DICE scores vs. #FLOPs for different methods over 10 binary medical image segmentation datasets. As shown, our approaches (PVT-EMCAD-B0 and PVT-EMCAD-B2) have the lowest #FLOPs, yet the highest DICE scores." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.348, + 0.47, + 0.379 + ], + "angle": 0, + "content": "while it has 1.91M parameters and 0.381G FLOPs for a standard encoder with #channels = [64, 128, 320, 512]." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.38, + 0.47, + 0.499 + ], + "angle": 0, + "content": "- Efficient Multi-scale Convolutional Attention Module: We introduce MSCAM, a new efficient multi-scale convolutional attention module that performs depth-wise convolutions at multiple scales; this refines the feature maps produced by vision encoders and enables capturing multi-scale salient features by suppressing irrelevant regions. The use of depth-wise convolutions makes MSCAM very efficient." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.5, + 0.47, + 0.59 + ], + "angle": 0, + "content": "- Large-kernel Grouped Attention Gate: We introduce a new grouped attention gate to fuse refined features with the features from skip connections. By using larger kernel \\((3 \\times 3)\\) group convolutions instead of point-wise convolutions in the design, we capture salient features in a larger local context with less computation." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.59, + 0.47, + 0.711 + ], + "angle": 0, + "content": "- Improved Performance: We empirically show that EM-CAD can be used with any hierarchical vision encoder (e.g., PVTv2-B0, PVTv2-B2 [56]), while significantly improving the performance of 2D medical image segmentation. EMCAD produces better results than SOTA methods with a significantly lower computational cost (as shown in Figure 1) on 12 medical image segmentation benchmarks that belong to six different tasks." + }, + { + "type": "list", + "bbox": [ + 0.078, + 0.38, + 0.47, + 0.711 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.712, + 0.469, + 0.804 + ], + "angle": 0, + "content": "The remaining of this paper is organized as follows: Section 2 summarizes related work. Section 3 describes the proposed method. Section 4 explains our experimental setup and results on 12 medical image segmentation benchmarks. Section 5 covers different ablation experiments. Lastly, Section 6 concludes the paper." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.82, + 0.22, + 0.835 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.846, + 0.236, + 0.861 + ], + "angle": 0, + "content": "2.1. Vision encoders" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.871, + 0.47, + 0.902 + ], + "angle": 0, + "content": "Convolutional Neural Networks (CNNs) [21-23, 32, 35, 45-48] have been foundational as encoders due to their pro" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.318 + ], + "angle": 0, + "content": "ficiency in handling spatial relationships in images. More precisely, AlexNet [32] and VGG [46] pave the way, leveraging deep layers of convolutions to extract features progressively. GoogleNet [47] introduces the inception module, allowing more efficient computation of representations across various scales. ResNet [21] introduces residual connections, enabling the training of networks with substantially more layers by addressing the vanishing gradients problem. MobileNets [22, 45] bring CNNs to mobile devices through lightweight, depth-wise separable convolutions. EfficientNet [48] introduces a scalable architectural design to CNNs with compound scaling. Although CNNs are pivotal for many vision applications, they generally lack the ability to capture long-range dependencies within images due to their inherent local receptive fields." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.32, + 0.893, + 0.622 + ], + "angle": 0, + "content": "Recently, Vision Transformers (ViTs), pioneered by Dosovitskiy et al. [18], enabled the learning of long-range relationships among pixels using Self-attention (SA). Since then, ViTs have been enhanced by integrating CNN features [49, 56], developing novel self-attention (SA) blocks [34, 49], and introducing new architectural designs [55, 58]. The Swin Transformer [34] incorporates a sliding window attention mechanism, while SegFormer [58] leverages MixFFN blocks for hierarchical structures. PVT [55] uses spatial reduction attention, refined in PVTv2 [56] with overlapping patch embedding and a linear complexity attention layer. MaxViT [49] introduces a multi-axis self-attention to form a hierarchical CNN-transformer encoder. Although ViTs address the CNNs limitation in capturing long-range pixel dependencies [21-23, 32, 35, 45-48], they face challenges in capturing the local spatial relationships among pixels. In this paper, we aim to overcome these limitations by introducing a new multi-scale cascaded attention decoder that refines feature maps and incorporates local attention using a multi-scale convolutional attention module." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.634, + 0.757, + 0.65 + ], + "angle": 0, + "content": "2.2. Medical image segmentation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.66, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Medical image segmentation involves pixel-wise classification to identify various anatomical structures like lesions, tumors, or organs within different imaging modalities such as endoscopy, MRI, or CT scans [8]. U-shaped networks [7, 19, 24, 26, 37, 41, 44, 62] are particularly favored due to their simple but effective encoder-decoder design. The UNet [44] pioneered this approach with its use of skip connections to fuse features at different resolution stages. UNet++ [62] evolves this design by incorporating nested encoder-decoder pathways with dense skip connections. Expanding on these ideas, UNet \\(3+\\) [24] introduces comprehensive skip pathways that facilitate full-scale feature integration. Further advancement comes with DC-UNet [37], which integrates a multi-resolution convolution scheme and residual paths into its skip connections. The DeepLab series, including DeepLabv3 [10] and DeepLabv3+ [11], in" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "11770" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.212 + ], + "angle": 0, + "content": "produce atrous convolutions and spatial pyramid pooling to handle multi-scale information. SegNet [2] uses pooling indices to upsample feature maps, preserving the boundary details. nnU-Net [19] automatically configures hyperparameters based on the specific dataset characteristics, using standard 2D and 3D UNets. Collectively, these U-shaped models have become a benchmark for success in the domain of medical image segmentation." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.213, + 0.471, + 0.41 + ], + "angle": 0, + "content": "Recently, vision transformers have emerged as a formidable force in medical image segmentation, harnessing the ability to capture pixel relationships at global scales [5, 8, 17, 42, 43, 52, 58, 61]. TransUNet [8] presents a novel blend of CNNs for local feature extraction and transformers for global context, enhancing both local and global feature capture. Swin-Unet [5] extends this by incorporating Swin Transformer blocks [34] into a U-shaped model for both encoding and decoding processes. Building on these concepts, MERIT [43] introduces a multi-scale hierarchical transformer, which employs SA across different window sizes, thus enhancing the model capacity to capture multiscale features critical for medical image segmentation." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.411, + 0.471, + 0.681 + ], + "angle": 0, + "content": "The integration of attention mechanisms has been investigated within CNNs [20, 41] and transformer-based systems [17] for enhancing medical image segmentation. PraNet [20] employs a reverse attention strategy for feature refinement. PolypPVT [17] leverages PVTv2 [56] as its backbone encoder and incorporates CBAM [57] within its decoding stages. The CASCADE [42] presents a novel cascaded decoder, combining channel [23] and spatial [9] attention to refine features at multiple stages, extracted from a transformer encoder, culminating in high-resolution segmentation outputs. While CASCADE achieves notable performance in segmenting medical images by integrating local and global insights from transformers, it is computationally inefficient due to the use of triple \\(3 \\times 3\\) convolution layers at each decoder stage. In addition to this, it uses single-scale convolutions during decoding. Our new proposal involves the adoption of multi-scale depth-wise convolutions to mitigate these constraints." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.691, + 0.214, + 0.709 + ], + "angle": 0, + "content": "3. Methodology" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.717, + 0.47, + 0.779 + ], + "angle": 0, + "content": "In this section, we first introduce our new EMCAD decoder and then explain two transformer-based architectures (i.e., PVT-EMCAD-B0 and PVT-EMCAD-B2) incorporating our proposed decoder." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.786, + 0.47, + 0.817 + ], + "angle": 0, + "content": "3.1. Efficient multi-scale convolutional attention decoding (EMCAD)" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.826, + 0.47, + 0.901 + ], + "angle": 0, + "content": "In this section, we introduce our efficient multi-scale convolutional decoding (EMCAD) to process the multi-stage features extracted from pretrained hierarchical vision encoders for high-resolution semantic segmentation. As shown in Figure 2(b), EMCAD consists of efficient multi" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.198 + ], + "angle": 0, + "content": "scale convolutional attention modules (MSCAMs) to robustly enhance the feature maps, large-kernel grouped attention gates (LGAGs) to refine feature maps fusing with the skip connection via gated attention mechanism, efficient up-convolution blocks (EUCBs) for up-sampling followed by enhancement of feature maps, and segmentation heads (SHs) to produce the segmentation outputs." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.2, + 0.893, + 0.336 + ], + "angle": 0, + "content": "More specifically, we use four MSCAMs to refine pyramid features (i.e., \\(X1\\), \\(X2\\), \\(X3\\), \\(X4\\) in Figure 2) extracted from the four stages of the encoder. After each MSCAM, we use an SH to produce a segmentation map of that stage. Subsequently, we upscale the refined feature maps using EUCBs and add them to the outputs from the corresponding LGAGs. Finally, we add four different segmentation maps to produce the final segmentation output. Different modules of our decoder are described next." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.362, + 0.869, + 0.379 + ], + "angle": 0, + "content": "3.1.1 Large-kernel grouped attention gate (LGAG)" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.392, + 0.893, + 0.769 + ], + "angle": 0, + "content": "We introduce a new large-kernel grouped attention gate (LGAG) to progressively combine feature maps with attention coefficients, which are learned by the network to allow higher activation of relevant features and suppression of irrelevant ones. This process employs a gating signal derived from higher-level features to control the flow of information across different stages of the network, thus enhancing its precision for medical image segmentation. Unlike Attention UNet [41] which uses \\( 1 \\times 1 \\) convolution to process gating signal \\( g \\) (features from skip connections) and input feature map \\( x \\) (upsampled features), in our \\( q_{att}(.) \\) function, we process \\( g \\) and \\( x \\) by applying separate \\( 3 \\times 3 \\) group convolutions \\( GC_{g}(.) \\) and \\( GC_{x}(.) \\), respectively. These convolved features are then normalized using batch normalization \\( (BN(.)) \\) [27] and merged through element-wise addition. The resultant feature map is activated through a ReLU \\( (R(.)) \\) layer [39]. Afterward, we apply a \\( 1 \\times 1 \\) convolution \\( (C(.)) \\) followed by \\( BN(.) \\) layer to get a single channel feature map. We then pass the resultant single-channel feature map through a Sigmoid \\( (\\sigma(.)) \\) activation function to yield the attention coefficients. The output of this transformation is used to scale the input feature \\( x \\) through elementwise multiplication, producing the attention-gated feature \\( LGAG(g, x) \\). The \\( LGAG(.) \\) (Figure 2(g)) can be formulated as in Equations 1 and 2:" + }, + { + "type": "equation", + "bbox": [ + 0.52, + 0.783, + 0.892, + 0.8 + ], + "angle": 0, + "content": "\\[\nq _ {a t t} (g, x) = R \\left(B N \\left(G C _ {g} (g) + B N \\left(G C _ {x} (x)\\right)\\right)\\right) \\tag {1}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.55, + 0.833, + 0.892, + 0.849 + ], + "angle": 0, + "content": "\\[\nL G A G (g, x) = x * \\sigma \\left(B N \\left(C \\left(q _ {a t t} (g, x)\\right)\\right)\\right) \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.86, + 0.893, + 0.906 + ], + "angle": 0, + "content": "Due to using \\(3 \\times 3\\) kernel group convolutions in \\(q_{att}(.)\\), our LGAG captures comparatively larger spatial contexts with less computational cost." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "11771" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.118, + 0.088, + 0.855, + 0.355 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.358, + 0.895, + 0.43 + ], + "angle": 0, + "content": "Figure 2. Hierarchical encoder with newly proposed EMCAD decoder architecture. (a) CNN or transformer encoder with four hierarchical stages, (b) EMCAD decoder, (c) Efficient up-convolution block (EUCB), (d) Multi-scale convolutional attention module (MSCAM), (e) Multi-scale convolution block (MSCB), (f) Multi-scale (parallel) depth-wise convolution (MSDC), (g) Large-kernel grouped attention gate (LGAG), (h) Channel attention block (CAB), and (i) Spatial attention block (SAB). X1, X2, X3, and X4 are the features from the four stages of the hierarchical encoder. p1, p2, p3, and p4 are output segmentation maps from four stages of our decoder." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.437, + 0.471, + 0.467 + ], + "angle": 0, + "content": "3.1.2 Multi-scale convolutional attention module (MSCAM)" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.475, + 0.47, + 0.598 + ], + "angle": 0, + "content": "We introduce an efficient multi-scale convolutional attention module to refine the feature maps. MSCAM consists of a channel attention block \\((CAB(\\cdot))\\) to put emphasis on pertinent channels, a spatial attention block [9] \\((SAB(\\cdot))\\) to capture the local contextual information, and an efficient multi-scale convolution block \\((MSCB(\\cdot))\\) to enhance the feature maps preserving contextual relationships. The \\(MSCAM(\\cdot)\\) (Figure 2(d)) is given in Equation 3:" + }, + { + "type": "equation", + "bbox": [ + 0.125, + 0.605, + 0.47, + 0.623 + ], + "angle": 0, + "content": "\\[\nM S C A M (x) = M S C B (S A B (C A B (x))) \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.629, + 0.47, + 0.69 + ], + "angle": 0, + "content": "where \\( x \\) is the input tensor. Due to using depth-wise convolution in multiple scales, our MSCAM is more effective with significantly lower computational cost than the convolutional attention module (CAM) proposed in [42]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.69, + 0.471, + 0.901 + ], + "angle": 0, + "content": "Multi-scale Convolution Block (MSCB): We introduce an efficient multi-scale convolution block to enhance the features generated by our cascaded expanding path. In our MSCB, we follow the design of the inverted residual block (IRB) of MobileNetV2 [45]. However, unlike IRB, our MSCB performs depth-wise convolution at multiple scales and uses channel_shuffle [60] to shuffle channels across groups. More specifically, in our MSCB, we first expand the number of channels (i.e., expansion_factor = 2) using a point-wise \\((1 \\times 1)\\) convolution layers \\(PWC_{1}(\\cdot)\\) followed by a batch normalization layer \\(BN(\\cdot)\\) and a ReLU6 [31] activation layer \\(R6(\\cdot)\\). We then use a multi-scale depth-wise convolution \\(MSDC(\\cdot)\\) to capture both multi-scale and multi-resolution contexts. As depth-wise convolution overlooks" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.437, + 0.892, + 0.53 + ], + "angle": 0, + "content": "the relationships among channels, we use a channel_shuffle operation to incorporate relationships among channels. Afterward, we use another point-wise convolution \\(PWC_{2}(.)\\) followed by a \\(BN(.)\\) to transform back the original #channels, which also encodes dependency among channels. The \\(MSCB(\\cdot)\\) (Figure 2(e)) is formulated as in Equation 4:" + }, + { + "type": "equation", + "bbox": [ + 0.513, + 0.538, + 0.892, + 0.553 + ], + "angle": 0, + "content": "\\[\nM S C B (x) = B N \\left(P W C _ {2} \\left(C S \\left(M S D C \\left(R 6 \\left(B N \\left(P W C _ {1} (x)\\right)\\right)\\right)\\right)\\right)\\right) \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.561, + 0.892, + 0.593 + ], + "angle": 0, + "content": "where parallel \\(MSDC(.)\\) (Figure 2(f)) for different kernel sizes \\((KS)\\) can be formulated using Equation 5:" + }, + { + "type": "equation", + "bbox": [ + 0.572, + 0.601, + 0.892, + 0.618 + ], + "angle": 0, + "content": "\\[\nM S D C (x) = \\sum_ {k s \\in K S} D W C B _ {k s} (x) \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.625, + 0.892, + 0.732 + ], + "angle": 0, + "content": "where \\( DWCB_{ks}(x) = R6(BN(DWC_{ks}(x))) \\). Here, \\( DWC_{ks}() \\) is a depth-wise convolution with the kernel size \\( ks \\). \\( BN() \\) and \\( R6() \\) are batch normalization and ReLU6 activation, respectively. Additionally, our sequential \\( MSDC() \\) uses the recursively updated input \\( x \\), where the input \\( x \\) is residually connected to the previous \\( DWCB_{ks}() \\) for better regularization as in Equation 6:" + }, + { + "type": "equation", + "bbox": [ + 0.616, + 0.74, + 0.892, + 0.757 + ], + "angle": 0, + "content": "\\[\nx = x + D W C B _ {k s} (x) \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.892, + 0.903 + ], + "angle": 0, + "content": "Channel Attention Block (CAB): We use channel attention block to assign different levels of importance to each channel, thus emphasizing more relevant features while suppressing less useful ones. Basically, the CAB identifies which feature maps to focus on (and then refine them). Following [57], in CAB, we first apply the adaptive maximum pooling \\((P_{m}(\\cdot))\\) and adaptive average pooling \\((P_{a}(\\cdot))\\) to the spatial dimensions (i.e., height and width) to extract the most significant feature of the entire feature map per" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "11772" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.228 + ], + "angle": 0, + "content": "channel. Then, for each pooled feature map, we reduce the number of channels \\( r = 1 / 16 \\) times separately using a point-wise convolution \\( (C_1(\\cdot)) \\) followed by a ReLU activation \\( (R) \\). Afterward, we recover the original channels using another point-wise convolution \\( (C_2(\\cdot)) \\). We then add both recovered feature maps and apply Sigmoid \\( (\\sigma) \\) activation to estimate attention weights. Finally, we incorporate these weights to input \\( x \\) using the Hadamard product \\( (\\otimes) \\). The \\( CAB(\\cdot) \\) (Figure 2(h)) is defined using Equation 7:" + }, + { + "type": "equation", + "bbox": [ + 0.095, + 0.238, + 0.47, + 0.253 + ], + "angle": 0, + "content": "\\[\nC A B (x) = \\sigma \\left(C _ {2} \\left(R \\left(C _ {1} \\left(P _ {m} (x)\\right)\\right)\\right) + C _ {2} \\left(R \\left(C _ {1} \\left(P _ {a} (x)\\right)\\right)\\right)\\right) * x \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.261, + 0.473, + 0.497 + ], + "angle": 0, + "content": "Spatial Attention Block (SAB): We use spatial attention to mimic the attentional processes of the human brain by focusing on specific parts of an input image. Basically, the SAB determines where to focus in a feature map; then it enhances those features. This process enhances the model's ability to recognize and respond to relevant spatial features, which is crucial for image segmentation where the context and location of objects significantly influence the output. In SAB, we first pool maximum \\((Ch_{max}(\\cdot))\\) and average \\((Ch_{avg}(\\cdot))\\) values along the channel dimension to pay attention to local features. Then, we use a large kernel (i.e., \\(7\\times 7\\) as in [17]) convolution layer to enhance local contextual relationships among features. Afterward, we apply the Sigmoid activation \\((\\sigma)\\) to calculate attention weights. Finally, we feed these weights to the input \\(x\\) (using Hadamard product \\((\\otimes)\\) to attend information in a more targeted way. The \\(SAB(.)\\) (Figure 2(i)) is defined using Equation 8:" + }, + { + "type": "equation", + "bbox": [ + 0.115, + 0.507, + 0.47, + 0.523 + ], + "angle": 0, + "content": "\\[\nS A B (x) = \\sigma \\left(L K C \\left(\\left[ C h _ {m a x} (x), C h _ {a v g} (x) \\right]\\right)\\right) * x \\tag {8}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.535, + 0.402, + 0.551 + ], + "angle": 0, + "content": "3.1.3 Efficient up-convolution block (EUCB)" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.56, + 0.473, + 0.711 + ], + "angle": 0, + "content": "We use an efficient up-convolution block to progressively upsample the feature maps of the current stage to match the dimension and resolution of the feature maps from the next skip connection. The EUCB first uses an UpSampling \\(Up(\\cdot)\\) with scale-factor 2 to upscale the feature maps. Then, it enhances the upscaled feature maps by applying a \\(3 \\times 3\\) depth-wise convolution \\(DWC(\\cdot)\\) followed by a \\(BN(\\cdot)\\) and a \\(ReLU(\\cdot)\\) activation. Finally, a \\(1 \\times 1\\) convolution \\(C_{1 \\times 1}(\\cdot)\\) is used to reduce the #channels to match with the next stage. The \\(EUCB(\\cdot)\\) (Figure 2(c)) is formulated as in Equation 9:" + }, + { + "type": "equation", + "bbox": [ + 0.091, + 0.738, + 0.47, + 0.755 + ], + "angle": 0, + "content": "\\[\nE U C B (x) = C _ {1 \\times 1} \\left(R e L U \\left(B N \\left(D W C (U p (x))\\right)\\right)\\right) \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.763, + 0.47, + 0.793 + ], + "angle": 0, + "content": "Due to using depth-wise convolution instead of \\(3 \\times 3\\) convolution, our EUCB is very efficient." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.801, + 0.303, + 0.817 + ], + "angle": 0, + "content": "3.1.4 Segmentation head (SH)" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.826, + 0.473, + 0.903 + ], + "angle": 0, + "content": "We use segmentation heads to produce the segmentation outputs from the refined feature maps of four stages of the decoder. The SH layer applies a \\(1 \\times 1\\) convolution \\(Conv_{1 \\times 1}(\\cdot)\\) to the refined feature maps having \\(ch_i\\) channels (\\(ch_i\\) is the #channels in the feature map of stage \\(i\\)) and" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.893, + 0.138 + ], + "angle": 0, + "content": "produces output with #channels equal to #classes in target dataset for multi-class but 1 channel for binary segmentation. The \\( SH(\\cdot) \\) is formulated as in Equation 10:" + }, + { + "type": "equation", + "bbox": [ + 0.616, + 0.143, + 0.893, + 0.16 + ], + "angle": 0, + "content": "\\[\nS H (x) = \\operatorname {C o n v} _ {1 \\times 1} (x) \\tag {10}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.17, + 0.694, + 0.185 + ], + "angle": 0, + "content": "3.2. Overall architecture" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.193, + 0.892, + 0.282 + ], + "angle": 0, + "content": "To show the generalization, effectiveness, and ability to process multi-scale features for medical image segmentation, we integrate our EMCAD decoder alongside tiny (PVTv2-B0) and standard (PVTv2-B2) networks of PVTv2 [56]. However, our decoder is adaptable and seamlessly compatible with other hierarchical backbone networks." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.284, + 0.893, + 0.45 + ], + "angle": 0, + "content": "PVTv2 differs from conventional transformer patch embedding modules by applying convolutional operations for consistent spatial information capture. Using PVTv2-b0 (Tiny) and PVTv2-b2 (Standard) encoders [56], we develop the PVT-EMCAD-B0 and PVT-EMCAD-B2 architectures. To adopt PVTv2, we first extract the features (X1, X2, X3, and X4) from four layers and feed them (i.e., X4 in the upsample path and X3, X2, X1 in the skip connections) into our EMCAD decoder as shown in Figure 2(a-b). EMCAD then processes them and produces four segmentation maps that correspond to the four stages of the encoder network." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.454, + 0.849, + 0.47 + ], + "angle": 0, + "content": "3.3. Multi-stage loss and outputs aggregation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.477, + 0.892, + 0.507 + ], + "angle": 0, + "content": "Our EMCAD decoder's four segmentation heads produce four prediction maps \\( p_1, p_2, p_3 \\), and \\( p_4 \\) across its stages." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.508, + 0.893, + 0.659 + ], + "angle": 0, + "content": "Loss aggregation: We adopt a combinatorial approach to loss combination called MUTATION, inspired by the work of MERIT [43] for multi-class segmentation. This involves calculating the loss for all possible combinations of predictions derived from 4 heads, totaling \\(2^{4} - 1 = 15\\) unique predictions, and then summing these losses. We focus on minimizing this cumulative combinatorial loss during the training process. For binary segmentation, we optimize the additive loss like [42] with an additional term \\(\\mathcal{L}_{p_1 + p_2 + p_3 + p_4}\\) as in Equation 11:" + }, + { + "type": "equation", + "bbox": [ + 0.518, + 0.666, + 0.891, + 0.681 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {t o t a l}} = \\alpha \\mathcal {L} _ {p _ {1}} + \\beta \\mathcal {L} _ {p _ {2}} + \\gamma \\mathcal {L} _ {p _ {3}} + \\zeta \\mathcal {L} _ {p _ {4}} + \\delta \\mathcal {L} _ {p _ {1} + p _ {2} + p _ {3} + p _ {4}} \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.683, + 0.892, + 0.728 + ], + "angle": 0, + "content": "where \\(\\mathcal{L}_{p_1},\\mathcal{L}_{p_2},\\mathcal{L}_{p_3}\\) , and \\(\\mathcal{L}_{p_4}\\) are the losses of each individual prediction maps. \\(\\alpha = \\beta = \\gamma = \\zeta = \\delta = 1.0\\) are the weights assigned to each loss." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.729, + 0.892, + 0.804 + ], + "angle": 0, + "content": "Output segmentation maps aggregation: We consider the prediction map, \\( p_4 \\), from the last stage of our decoder as the final segmentation map. Then, we obtain the final segmentation output by employing a Sigmoid function for binary or a Softmax function for multi-class segmentation." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.816, + 0.634, + 0.833 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.841, + 0.892, + 0.901 + ], + "angle": 0, + "content": "In this section, we present the details of our implementation followed by a comparative analysis of our PVT-EMCAD-B0 and PVT-EMCAD-B2 against SOTA methods. Datasets and evaluation metrics are in Supplementary Section 7." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "11773" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.081, + 0.09, + 0.892, + 0.325 + ], + "angle": 0, + "content": "
Methods#Params#FLOPsPolypSkin LesionCellBUSIAvg.
ClinicColonETISKvasirBKAIISIC17ISIC18DSB18EM
UNet [44]34.53M65.53G92.1183.9576.8582.8785.0583.0786.6792.2395.4674.0485.23
UNet++ [62]9.16M34.65G92.1787.8877.4083.3684.0782.9887.4691.9795.4874.7685.75
AttnUNet [41]34.88M66.64G92.2086.4676.8483.4984.0783.6687.0592.2295.5574.4885.60
DeepLabv3+ [10]39.76M14.92G93.2491.9290.7389.0689.7483.8488.6492.1494.9676.8189.11
PraNet [20]32.55M6.93G91.7189.1683.8484.8285.5683.0388.5689.8992.3775.1486.41
CaraNet [38]46.64M11.48G94.0891.1990.2589.7489.7185.0290.1889.1592.7877.3488.94
UACANet-L [30]69.16M31.51G94.1691.0289.7790.1790.3583.7289.7688.8689.2876.9688.41
SSFormer-L [54]66.22M17.28G94.1892.1190.1691.4791.1485.2890.2592.0394.9578.7690.03
PolypPVT [17]25.11M5.30G94.1391.5389.9391.5691.1785.5690.3690.6994.4079.3589.87
TransUNet [8]105.32M38.52G93.9091.6387.7991.0889.1785.0089.1692.0495.2778.3089.33
SwinUNet [5]27.17M6.2G92.4289.2785.1089.5987.6183.9789.2691.0394.4777.3888.01
TransFuse [61]143.74M82.71G93.6290.3586.9190.2487.4784.8989.6290.8594.3579.3688.77
UNeXt [50]1.47M0.57G90.2083.8474.0377.8877.9382.7487.7886.0193.8174.7182.89
PVT-CASCADE [42]34.12M7.62G94.5391.6091.0392.0592.1485.5090.4192.3595.4279.2190.42
PVT-EMCAD-B0 (Ours)3.92M0.84G94.6091.7191.6591.9591.3085.6790.7092.4695.3579.8090.52
PVT-EMCAD-B2 (Ours)26.76M5.6G95.2192.3192.2992.7592.9685.9590.9692.7495.5380.2591.10
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.328, + 0.894, + 0.37 + ], + "angle": 0, + "content": "Table 1. Results of binary medical image segmentation (i.e., polyp, skin lesion, cell, and breast cancer). We reproduce the results of SOTA methods using their publicly available implementation with our train-val-test splits of 80:10:10. #FLOPs of all the methods are reported for \\(256 \\times 256\\) inputs, except Swin-UNet (\\(224 \\times 224\\)). All results are averaged over five runs. Best results are shown in bold." + }, + { + "type": "table", + "bbox": [ + 0.094, + 0.378, + 0.879, + 0.621 + ], + "angle": 0, + "content": "
ArchitecturesAverageAortaGBKLKRLiverPCSPSM
DICE↑HD95↓mIoU↑
UNet [44]70.1144.6959.3984.0056.7072.4162.6486.9848.7381.4867.96
AttnUNet [41]71.7034.4761.3882.6161.9476.0770.4287.5446.7080.6767.66
R50+UNet [8]74.6836.87-84.1862.8479.1971.2993.3548.2384.4173.92
R50+AttnUNet [8]75.5736.97-55.9263.9179.2072.7193.5649.3787.1974.95
SSFormer [54]78.0125.7267.2382.7863.7480.7278.1193.5361.5387.0776.61
PolypPVT [17]78.0825.6167.4382.3466.1481.2173.7894.3759.3488.0579.4
TransUNet [8]77.6126.967.3286.5660.4380.5478.5394.3358.4787.0675.00
SwinUNet [5]77.5827.3266.8881.7665.9582.3279.2293.7353.8188.0475.79
MT-UNet [53]78.5926.59-87.9264.9981.4777.2993.0659.4687.7576.81
MISSformer [25]81.9618.20-86.9968.6585.2182.0094.4165.6791.9280.81
PVT-CASCADE [42]81.0620.2370.8883.0170.5982.2380.3794.0864.4390.183.69
TransCASCADE [42]82.6817.3473.4886.6368.4887.6684.5694.4365.3390.7983.52
PVT-EMCAD-B0 (Ours)81.9717.3972.6487.2166.6287.4883.9694.5762.0092.6681.22
PVT-EMCAD-B2 (Ours)83.6315.6874.6588.1468.8788.0884.1095.2668.5192.1783.92
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.624, + 0.895, + 0.667 + ], + "angle": 0, + "content": "Table 2. Results of abdomen organ segmentation on Synapse Multi-organ dataset. DICE scores are reported for individual organs. Results of UNet, AttnUNet, PolypPVT, SSFormerPVT, TransUNet, and SwinUNet are taken from [42]. \\(\\uparrow (\\downarrow)\\) denotes the higher (lower) the better. \\(^{\\prime} - ^{\\prime}\\) means missing data from the source. EMCAD results are averaged over five runs. Best results are shown in bold." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.678, + 0.292, + 0.695 + ], + "angle": 0, + "content": "4.1. Implementation details" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.704, + 0.471, + 0.902 + ], + "angle": 0, + "content": "We implement our network and conduct experiments using Pytorch 1.11.0 on a single NVIDIA RTX A6000 GPU with 48GB of memory. We utilize ImageNet [16] pretrained PVTv2-b0 and PVTv2-b2 [56] as encoders. In the MSDC of our decoder, we set the multi-scale kernels [1,3,5] through an ablation study. We use the parallel arrangement of depth-wise convolutions in all experiments. Our models are trained using the AdamW optimizer [36] with a learning rate and weight decay of \\(1e - 4\\). We generally train for 200 epochs with a batch size of 16, except for Synapse multi-organ (300 epochs, batch size 6) and ACDC cardiac organ (400 epochs, batch size 12), saving the best model based on the DICE score. We resize images" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.678, + 0.892, + 0.831 + ], + "angle": 0, + "content": "to \\(352 \\times 352\\) and use a multi-scale \\(\\{0.75, 1.0, 1.25\\}\\) training strategy with a gradient clip limit of 0.5 for ClinicDB [3], Kvasir [29], ColonDB [51], ETIS [51], BKAI [40], ISIC17 [15], and ISIC18 [15], while we resize images to \\(256 \\times 256\\) for BUSI [1], EM [6], and DSB18 [4]. For Synapse and ACDC datasets, images are resized to \\(224 \\times 224\\), with random rotation and flipping augmentations, optimizing a combined Cross-entropy (0.3) and DICE (0.7) loss. For binary segmentation, we utilize the combined weighted BinaryCrossEntropy (BCE) and weighted IoU loss function." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.845, + 0.593, + 0.859 + ], + "angle": 0, + "content": "4.2. Results" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.894, + 0.901 + ], + "angle": 0, + "content": "We compare our architectures (i.e., PVT-EMCAD-B0 and PVT-EMCAD-B2) with SOTA CNN and transformer-based" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "11774" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.11, + 0.09, + 0.44, + 0.279 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.28, + 0.472, + 0.337 + ], + "angle": 0, + "content": "Figure 3. Average DICE scores vs. #Params for different methods over 10 binary medical image segmentation datasets. As shown, our proposed approaches (PVT-EMCAD-B0 and PVT-EMCAD-B2) have the fewest parameters, yet the highest DICE scores." + }, + { + "type": "table", + "bbox": [ + 0.079, + 0.344, + 0.473, + 0.548 + ], + "angle": 0, + "content": "
MethodsAvg. DICERVMyoLV
R50+UNet [8]87.5587.1080.6394.92
R50+AttnUNet [8]86.7587.5879.2093.47
ViT+CUP [8]81.4581.4670.7192.18
R50+ViT+CUP [8]87.5786.0781.8894.75
TransUNet [8]89.7186.6787.2795.18
SwinUNet [5]88.0785.7784.4294.03
MT-UNet [53]90.4386.6489.0495.62
MISSformer [25]90.8689.5588.0494.99
PVT-CASCADE [42]91.4689.9788.995.50
TransCASCade [42]91.6390.2589.1495.50
Cascaded MERIT [43]91.8590.2389.5395.80
PVT-EMCAD-B0 (Ours)91.34±0.289.3788.9995.65
PVT-EMCAD-B2 (Ours)92.12±0.290.6589.6896.02
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.552, + 0.47, + 0.595 + ], + "angle": 0, + "content": "Table 3. Results of cardiac organ segmentation on ACDC dataset. DICE scores \\((\\%)\\) are reported for individual organs. We get the results of SwinUNet from [42]. Best results are shown in bold." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.607, + 0.47, + 0.655 + ], + "angle": 0, + "content": "segmentation methods on 12 datasets that belong to six medical image segmentation tasks. Qualitative results are in the Supplementary Section 7.3." + }, + { + "type": "title", + "bbox": [ + 0.076, + 0.665, + 0.452, + 0.683 + ], + "angle": 0, + "content": "4.2.1 Results of binary medical image segmentation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.689, + 0.47, + 0.795 + ], + "angle": 0, + "content": "Results for different methods on 10 binary medical image segmentation datasets are shown in Table 1 and Figure 1. Our PVT-EMCAD-B2 attains the highest average DICE score \\((91.10\\%)\\) with only \\(26.76\\mathrm{M}\\) parameters and 5.6G FLOPs. The multi-scale depth-wise convolution in our EMCAD decoder, combined with the transformer encoder, contributes to these performance gains." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.796, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Polyp segmentation: Table 1 reveals that our PVT-EMCAD-B2 surpasses all SOTA methods in five polyp segmentation datasets. PVT-EMCAD-B2 achieves DICE score improvements of \\(1.08\\%\\), \\(0.78\\%\\), \\(2.36\\%\\), \\(1.19\\%\\), and \\(1.79\\%\\) over PolypPVT in ClinicDB, ColonDB, ETIS, Kvasir, and BKAI-IGI, despite having slightly more parameters and FLOPs. The smallest model UNeXt, ex" + }, + { + "type": "table", + "bbox": [ + 0.503, + 0.09, + 0.895, + 0.187 + ], + "angle": 0, + "content": "
Components#FLOPs(G)#Params (M)Avg DICE
CascadedLGAGMSCAM224256
NoNoNo00080.10±0.2
YesNoNo0.1000.1310.22481.08±0.2
YesYesNo0.1080.1410.23581.92±0.2
YesNoYes0.3730.4871.89882.86±0.3
YesYesYes0.3810.4981.9183.63±0.3
" + }, + { + "type": "table_caption", + "bbox": [ + 0.498, + 0.19, + 0.895, + 0.248 + ], + "angle": 0, + "content": "Table 4. Effect of different components of EMCAD with PVTv2-b2 encoder on Synapse multi-organ dataset. #FLOPs are reported for input resolution of \\(224 \\times 224\\) and \\(256 \\times 256\\). All results are averaged over five runs. Best results are shown in bold." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.256, + 0.892, + 0.377 + ], + "angle": 0, + "content": "hibits the worst performance in all five polyp segmentation datasets. Our smaller model with only 3.92M parameters and 0.84G FLOPs also outperforms all the methods except PVT-CASCADE (in Kvasir and BKAI-IGH) and SSFormer-L (in ColonDB), which achieve the best performance among SOTA methods. In conclusion, our PVT-EMCAD-B2 achieves the new SOTA results in these five polyp segmentation datasets." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.378, + 0.893, + 0.529 + ], + "angle": 0, + "content": "Skin lesion segmentation: Table 1 shows PVT-EMCAD-B2's strong performance on ISIC17 and ISIC18 skin lesion segmentation datasets, achieving DICE scores of \\(85.95\\%\\) and \\(90.96\\%\\), surpassing DeepLabV3+ by \\(2.11\\%\\) and \\(2.32\\%\\). It also beats the nearest method PVT-CASCADE by \\(0.45\\%\\) and \\(0.55\\%\\) in ISIC17 and ISIC18, respectively, though our decoder is significantly more efficient than CASCADE. Our PVT-EMCAD-B0 also shows huge potential in point care applications like skin lesion segmentation with only 3.92M parameters and 0.84G FLOPs." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.53, + 0.894, + 0.666 + ], + "angle": 0, + "content": "Cell segmentation: To evaluate our method's effectiveness in biological imaging, we use DSB18 [4] for cell nuclei and EM [6] for cell structure segmentation. As Table 1 indicates, our PVT-EMCAD-B2 sets a SOTA benchmark in cell nuclei segmentation on DSB18, outperforming DeepLabv3+, TransFuse, and PVT-CASCADE. On the EM dataset, PVT-EMCAD-B2 secures the second-best DICE score (95.53%), offering significantly lower computational costs than the top-performing AttnUNet (95.55%)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.666, + 0.894, + 0.757 + ], + "angle": 0, + "content": "Breast cancer segmentation: We conduct experiments on the BUSI dataset for breast cancer segmentation in ultrasound images. Our PVT-EMCAD-B2 achieves the SOTA DICE score \\((80.25\\%)\\) on this dataset. Furthermore, our PVT-EMCAD-B0 outperforms the computationally similar method UNeXt by a notable margin of \\(5.54\\%\\)." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.771, + 0.833, + 0.787 + ], + "angle": 0, + "content": "4.2.2 Results of abdomen organ segmentation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.795, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Table 2 shows that our PVT-EMCAD-B2 excels in abdomen organ segmentation on the Synapse multi-organ dataset, achieving the highest average DICE score of \\(83.63\\%\\) and surpassing all SOTA CNN- and transformer-based methods. It outperforms PVT-CASCADE by \\(2.57\\%\\) in DICE score and 4.55 in HD95 distance, indicating superior organ boundary location. Our EMCAD decoder boosts individ" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "11775" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.111, + 0.089, + 0.858, + 0.148 + ], + "angle": 0, + "content": "
Conv. kernels[1][3][5][1,3][3,3][1,3,5][3,3,3][3,5,7][1,3,5,7][1,3,5,7,9]
Synapse82.4382.7982.7482.9882.8183.6382.9283.1183.5783.34
ClinicDB94.8194.9094.9895.1395.0695.2195.1595.0395.1895.07
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.152, + 0.892, + 0.181 + ], + "angle": 0, + "content": "Table 5. Effect of multi-scale kernels in the depth-wise convolution of MSDC on ClinicDB and Synapse multi-organ datasets. We use the PVTv2-b2 encoder for these experiments. All results are averaged over five runs. Best results are highlighted in bold." + }, + { + "type": "table", + "bbox": [ + 0.079, + 0.19, + 0.472, + 0.268 + ], + "angle": 0, + "content": "
EncodersDecoders#FLOPs(G)#Params(M)DICE (%)
PVTv2-B0CASCADE0.4392.3280.54
PVTv2-B0EMCAD (Ours)0.1100.50781.97
PVTv2-B2CASCADE1.939.2782.78
PVTv2-B2EMCAD (Ours)0.3811.9183.63
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.077, + 0.27, + 0.468, + 0.326 + ], + "angle": 0, + "content": "Table 6. Comparison with the baseline decoder on Synapse Multi-organ dataset. We only report the #FLOPs (with input resolution of \\(224 \\times 224\\)) and the #parameters of the decoders. All the results are averaged over five runs. Best results are shown in bold." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.332, + 0.468, + 0.362 + ], + "angle": 0, + "content": "ual organ segmentation, significantly outperforming SOTA methods on six of eight organs." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.366, + 0.396, + 0.381 + ], + "angle": 0, + "content": "4.2.3 Results of cardiac organ segmentation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.389, + 0.468, + 0.51 + ], + "angle": 0, + "content": "Table 3 shows the DICE scores of our PVT-EMCAD-B2 and PVT-EMCAD-B0 along with other SOTA methods, on the MRI images of the ACDC dataset for cardiac organ segmentation. Our PVT-EMCAD-B2 achieves the highest average DICE score of \\(92.12\\%\\), thus improving about \\(0.27\\%\\) over Cascaded MERIT though our network has significantly lower computational cost. Besides, PVT-EMCAD-B2 has better DICE scores in all three organ segmentations." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.519, + 0.241, + 0.535 + ], + "angle": 0, + "content": "5. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.544, + 0.468, + 0.589 + ], + "angle": 0, + "content": "In this section, we conduct ablation studies to explore different aspects of our architectures and the experimental framework. More ablations are in Supplementary Section 8." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.593, + 0.437, + 0.609 + ], + "angle": 0, + "content": "5.1. Effect of different components of EMCAD" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.617, + 0.468, + 0.813 + ], + "angle": 0, + "content": "We conduct a set of experiments on the Synapse multi-organ dataset to understand the effect of different components of our EMCAD decoder. We start with only the encoder and add different modules such as Cascaded structure, LGAG, and MSCAM to understand their effect. Table 4 exhibits that the cascaded structure of the decoder helps to improve performance over the non-cascaded one. The incorporation of LGAG and MSCAM improves performance, however, MSCAM proves to be more effective. When both the LGAG and MSCAM modules are used together, it produces the best DICE score of \\(83.63\\%\\). It is also evident that there is about \\(3.53\\%\\) improvement in the DICE score with an additional 0.381G FLOPs and 1.91M parameters." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.818, + 0.421, + 0.833 + ], + "angle": 0, + "content": "5.2. Effect of multi-scale kernels in MSCAM" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.468, + 0.901 + ], + "angle": 0, + "content": "We have conducted another set of experiments on Synapse multi-organ and ClinicDB datasets to understand the effect of different multi-scale kernels used for depth-wise convolutions in MSDC. Table 5 reports these results which show" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.193, + 0.892, + 0.344 + ], + "angle": 0, + "content": "that performance improves from \\(1 \\times 1\\) to \\(3 \\times 3\\) kernel. When \\(1 \\times 1\\) kernel is used together with \\(3 \\times 3\\) it improves more than when using them alone. However, when two \\(3 \\times 3\\) kernels are used together, performance drops. The incorporation of a \\(5 \\times 5\\) kernel with \\(1 \\times 1\\) and \\(3 \\times 3\\) kernels further improves the performance and it achieves the best results in both Synapse multi-organ and ClinicDB datasets. If we add additional larger kernels (e.g., \\(7 \\times 7\\), \\(9 \\times 9\\)), the performance of both datasets drops. Based on these empirical observations, we choose [1, 3, 5] kernels in all our experiments." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.353, + 0.83, + 0.369 + ], + "angle": 0, + "content": "5.3. Comparison with the baseline decoder" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.377, + 0.892, + 0.513 + ], + "angle": 0, + "content": "In Table 6, we report the experimental results with the computational complexity of our EMCAD decoder and a baseline decoder, namely CASCADE. From Table 6, we can see that our EMCAD decoder with PVTv2-b2 requires \\(80.3\\%\\) fewer FLOPs and \\(79.4\\%\\) fewer parameters to outperform (by \\(0.85\\%\\)) the respective CASCADE decoder. Similarly, our EMCAD decoder with PVTv2-B0 achieves \\(1.43\\%\\) better DICE score than the CASCADE decoder with \\(78.1\\%\\) fewer parameters and \\(74.9\\%\\) fewer FLOPs." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.523, + 0.627, + 0.538 + ], + "angle": 0, + "content": "6. Conclusions" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.549, + 0.892, + 0.684 + ], + "angle": 0, + "content": "In this paper, we have presented EMCAD, a new and efficient multi-scale convolutional attention decoder designed for multi-stage feature aggregation and refinement in medical image segmentation. EMCAD employs a multi-scale depth-wise convolution block, which is key for capturing diverse scale information within feature maps, a critical factor for precision in medical image segmentation. This design choice, using depth-wise convolutions instead of standard \\(3 \\times 3\\) convolution blocks, makes EMCAD notably efficient." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.686, + 0.892, + 0.851 + ], + "angle": 0, + "content": "Our experiments reveal that EMCAD surpasses the recent CASCADE decoder in DICE scores with \\(79.4\\%\\) fewer parameters and \\(80.3\\%\\) less FLOPs. Our extensive experiments also confirm EMCAD's superior performance compared to SOTA methods across 12 public datasets covering six different 2D medical image segmentation tasks. EMCAD's compatibility with smaller encoders makes it an excellent fit for point-of-care applications while maintaining high performance. We anticipate that our EMCAD decoder will be a valuable asset in enhancing a variety of medical image segmentation and semantic segmentation tasks." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.852, + 0.892, + 0.894 + ], + "angle": 0, + "content": "Acknowledgements: This work is supported in part by the NSF grant CNS 2007284, and in part by the iMAGiNE Consortium (https://imagine.utexas.edu/)." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "11776" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.105 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.469, + 0.156 + ], + "angle": 0, + "content": "[1] Walid Al-Dhabyani, Mohammed Gomaa, Hussien Khaled, and Aly Fahmy. Dataset of breast ultrasound images. Data in brief, 28:104863, 2020. 6, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.158, + 0.469, + 0.212 + ], + "angle": 0, + "content": "[2] Vijay Badrinarayanan, Alex Kendall, and Roberto Cipolla. Segnet: A deep convolutional encoder-decoder architecture for image segmentation. IEEE Trans. Pattern Anal. Mach. Intell., 39(12):2481-2495, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.214, + 0.469, + 0.282 + ], + "angle": 0, + "content": "[3] Jorge Bernal, F Javier Sánchez, Gloria Fernández-Esparrach, Debora Gil, Cristina Rodríguez, and Fernando Vilarino. Wm-dova maps for accurate polyp highlighting in colonoscopy: Validation vs. saliency maps from physicians. Comput. Med. Imaging Graph., 43:99-111, 2015. 6, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.284, + 0.469, + 0.366 + ], + "angle": 0, + "content": "[4] Juan C Caicedo, Allen Goodman, Kyle W Karhohs, Beth A Cimini, Jeanelle Ackerman, Marzieh Haghighi, CherKeng Heng, Tim Becker, Minh Doan, Claire McQuin, et al. Nucleus segmentation across imaging experiments: the 2018 data science bowl. Nature methods, 16(12):1247-1253, 2019. 6, 7, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.368, + 0.469, + 0.422 + ], + "angle": 0, + "content": "[5] Hu Cao, Yueyue Wang, Joy Chen, Dongsheng Jiang, Xiaopeng Zhang, Qi Tian, and Manning Wang. Swin-unet: Unet-like pure transformer for medical image segmentation. arXiv preprint arXiv:2105.05537, 2021. 1, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.424, + 0.469, + 0.506 + ], + "angle": 0, + "content": "[6] Albert Cardona, Stephan Saalfeld, Stephan Preibisch, Benjamin Schmid, Anchi Cheng, Jim Pulokas, Pavel Tomancak, and Volker Hartenstein. An integrated micro-and macroarchitectural analysis of the drosophila brain by computer-assisted serial section electron microscopy. PLoS biology, 8(10):e1000502, 2010. 6, 7, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.508, + 0.469, + 0.563 + ], + "angle": 0, + "content": "[7] Gongping Chen, Lei Li, Yu Dai, Jianxun Zhang, and Moi Hoon Yap. Aau-net: an adaptive attention u-net for breast lesions segmentation in ultrasound images. IEEE Trans. Med. Imaging, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.565, + 0.469, + 0.632 + ], + "angle": 0, + "content": "[8] Jieneng Chen, Yongyi Lu, Qihang Yu, Xiangde Luo, Ehsan Adeli, Yan Wang, Le Lu, Alan L Yuille, and Yuyin Zhou. Transunet: Transformers make strong encoders for medical image segmentation. arXiv preprint arXiv:2102.04306, 2021. 1, 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.634, + 0.469, + 0.702 + ], + "angle": 0, + "content": "[9] Long Chen, Hanwang Zhang, Jun Xiao, Liqiang Nie, Jian Shao, Wei Liu, and Tat-Seng Chua. Sca-cnn: Spatial and channel-wise attention in convolutional networks for image captioning. In IEEE Conf. Comput. Vis. Pattern Recog., pages 5659-5667, 2017. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.704, + 0.469, + 0.772 + ], + "angle": 0, + "content": "[10] Liang-Chieh Chen, George Papandreou, Iasonas Kokkinos, Kevin Murphy, and Alan L Yuille. Deeplab: Semantic image segmentation with deep convolutional nets, atrous convolution, and fully connected crfs. IEEE Trans. Pattern Anal. Mach. Intell., 40(4):834-848, 2017. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.775, + 0.469, + 0.829 + ], + "angle": 0, + "content": "[11] Liang-Chieh Chen, Yukun Zhu, George Papandreou, Florian Schroff, and Hartwig Adam. Encoder-decoder with atrous separable convolution for semantic image segmentation. In Eur. Conf. Comput. Vis., pages 801-818, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.831, + 0.469, + 0.871 + ], + "angle": 0, + "content": "[12] Shuhan Chen, Xiuli Tan, Ben Wang, and Xuelong Hu. Reverse attention for salient object detection. In Eur. Conf. Comput. Vis., pages 234-250, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[13] Xiangxiang Chu, Zhi Tian, Bo Zhang, Xinlong Wang, Xiaolin Wei, Huaxia Xia, and Chunhua Shen. Conditional po" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.469, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.892, + 0.119 + ], + "angle": 0, + "content": "sitional encodings for vision transformers. arXiv preprint arXiv:2102.10882, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.121, + 0.892, + 0.216 + ], + "angle": 0, + "content": "[14] Noel Codella, Veronica Rotemberg, Philipp Tschandl, M Emre Celebi, Stephen Dusza, David Gutman, Brian Helba, Aadi Kalloo, Konstantinos Liopyris, Michael Marchetti, et al. Skin lesion analysis toward melanoma detection 2018: A challenge hosted by the international skin imaging collaboration (isic). arXiv preprint arXiv:1902.03368, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.218, + 0.892, + 0.328 + ], + "angle": 0, + "content": "[15] Noel CF Codella, David Gutman, M Emre Celebi, Brian Helba, Michael A Marchetti, Stephen W Dusza, Aadi Kalloo, Konstantinos Liopyris, Nabin Mishra, Harald Kittler, et al. Skin lesion analysis toward melanoma detection: A challenge at the 2017 international symposium on biomedical imaging (isbi), hosted by the international skin imaging collaboration (isic). In IEEE Int. Symp. Biomed. Imaging, pages 168-172. IEEE, 2018. 6, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.329, + 0.892, + 0.383 + ], + "angle": 0, + "content": "[16] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In IEEE Conf. Comput. Vis. Pattern Recog., pages 248-255. IEEE, 2009. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.385, + 0.892, + 0.439 + ], + "angle": 0, + "content": "[17] Bo Dong, Wenhai Wang, Deng-Ping Fan, Jinpeng Li, Huazhu Fu, and Ling Shao. Polyp-pvt: Polyp segmentation with pyramid vision transformers. arXiv preprint arXiv:2108.06932, 2021. 1, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.441, + 0.892, + 0.522 + ], + "angle": 0, + "content": "[18] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.524, + 0.892, + 0.565 + ], + "angle": 0, + "content": "[19] Isensee et al. nnu-net: a self-configuring method for deep learning-based biomedical image segmentation. Nature methods, 18(2):203-211, 2021. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.567, + 0.892, + 0.634 + ], + "angle": 0, + "content": "[20] Deng-Ping Fan, Ge-Peng Ji, Tao Zhou, Geng Chen, Huazhu Fu, Jianbing Shen, and Ling Shao. Pranet: Parallel reverse attention network for polyp segmentation. In Int. Conf. Med. Image Comput. Comput. Assist. Interv., pages 263-273. Springer, 2020. 1, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.636, + 0.892, + 0.677 + ], + "angle": 0, + "content": "[21] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In IEEE Conf. Comput. Vis. Pattern Recog., pages 770-778, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.679, + 0.892, + 0.747 + ], + "angle": 0, + "content": "[22] Andrew G Howard, Menglong Zhu, Bo Chen, Dmitry Kalenichenko, Weijun Wang, Tobias Weyand, Marco Andreetto, and Hartwig Adam. Mobilenets: Efficient convolutional neural networks for mobile vision applications. arXiv preprint arXiv:1704.04861, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.748, + 0.892, + 0.788 + ], + "angle": 0, + "content": "[23] Jie Hu, Li Shen, and Gang Sun. Squeeze-and-excitation networks. In IEEE Conf. Comput. Vis. Pattern Recog., pages 7132-7141, 2018. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.79, + 0.892, + 0.857 + ], + "angle": 0, + "content": "[24] Huimin Huang, Lanfen Lin, Ruofeng Tong, Hongjie Hu, Qiaowei Zhang, Yutaro Iwamoto, Xianhua Han, Yen-Wei Chen, and Jian Wu. Unet 3+: A full-scale connected unet for medical image segmentation. In ICASSP, pages 1055–1059. IEEE, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.859, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[25] Xiaohong Huang, Zhifang Deng, Dandan Li, and Xueguang Yuan. Missformer: An effective medical image segmentation transformer. arXiv preprint arXiv:2109.07162, 2021. 6, 7" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.956 + ], + "angle": 0, + "content": "11777" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.147 + ], + "angle": 0, + "content": "[26] Nabil Ibtehaz and Daisuke Kihara. Acc-unet: A completely convolutional unet model for the 2020s. In Int. Conf. Med. Image Comput. Comput. Assist. Interv., pages 692-702. Springer, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.151, + 0.47, + 0.207 + ], + "angle": 0, + "content": "[27] Sergey Ioffe and Christian Szegedy. Batch normalization: Accelerating deep network training by reducing internal covariate shift. In Int. Conf. Mach. Learn., pages 448-456. pmlr, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.209, + 0.47, + 0.25 + ], + "angle": 0, + "content": "[28] Md Amirul Islam, Sen Jia, and Neil DB Bruce. How much position information do convolutional neural networks encode? arXiv preprint arXiv:2001.08248, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.252, + 0.47, + 0.32 + ], + "angle": 0, + "content": "[29] Debesh Jha, Pia H Smedsrud, Michael A Riegler, Pål Halvorsen, Thomas de Lange, Dag Johansen, and Håvard D Johansen. Kvasir-seg: A segmented polyp dataset. In Int. Conf. Multimedia Model., pages 451-462. Springer, 2020. 6, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.324, + 0.47, + 0.365 + ], + "angle": 0, + "content": "[30] Taehun Kim, Hyemin Lee, and Daijin Kim. Uacanet: Uncertainty augmented context attention for polyp segmentation. In ACM Int. Conf. Multimedia, pages 2167-2175, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.368, + 0.469, + 0.407 + ], + "angle": 0, + "content": "[31] Alex Krizhevsky and Geoff Hinton. Convolutional deep belief networks on CIFar-10. Unpublished manuscript, 40(7): 1-9, 2010. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.412, + 0.469, + 0.453 + ], + "angle": 0, + "content": "[32] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convolutional neural networks. Adv. Neural Inform. Process. Syst., 25, 2012. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.456, + 0.469, + 0.524 + ], + "angle": 0, + "content": "[33] Xian Lin, Zengqiang Yan, Xianbo Deng, Chuansheng Zheng, and Li Yu. Convformer: Plug-and-play cnn-style transformers for improving medical image segmentation. In Int. Conf. Med. Image Comput. Comput. Assist. Interv., pages 642-651. Springer, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.527, + 0.469, + 0.582 + ], + "angle": 0, + "content": "[34] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Int. Conf. Comput. Vis., pages 10012-10022, 2021. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.585, + 0.469, + 0.638 + ], + "angle": 0, + "content": "[35] Zhuang Liu, Hanzi Mao, Chao-Yuan Wu, Christoph Feichtenhofer, Trevor Darrell, and Saining Xie. A convnet for the 2020s. In IEEE Conf. Comput. Vis. Pattern Recog., pages 11976-11986, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.643, + 0.469, + 0.67 + ], + "angle": 0, + "content": "[36] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.673, + 0.469, + 0.727 + ], + "angle": 0, + "content": "[37] Ange Lou, Shuyue Guan, and Murray Loew. Dc-unet: rethinking the u-net architecture with dual channel efficient cnn for medical image segmentation. In Med. Imaging 2021: Image Process., pages 758-768. SPIE, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.731, + 0.469, + 0.785 + ], + "angle": 0, + "content": "[38] Ange Lou, Shuyue Guan, Hanseok Ko, and Murray H Loew. Caranet: context axial reverse attention network for segmentation of small medical objects. In Med. Imaging 2022: Image Process., pages 81-92. SPIE, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.788, + 0.469, + 0.83 + ], + "angle": 0, + "content": "[39] Vinod Nair and Geoffrey E Hinton. Rectified linear units improve restricted boltzmann machines. In Int. Conf. Mach. Learn., pages 807-814, 2010. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.832, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[40] Phan Ngoc Lan, Nguyen Sy An, Dao Viet Hang, Dao Van Long, Tran Quang Trung, Nguyen Thi Thuy, and Dinh Viet Sang. Neounet: Towards accurate colon polyp segmentation and neoplasm detection. In Adv. Vis. Comput. - Int. Symp., pages 15-28. Springer, 2021. 6, 1" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.162 + ], + "angle": 0, + "content": "[41] Ozan Oktay, Jo Schlemper, Loic Le Folgoc, Matthew Lee, Mattias Heinrich, Kazunari Misawa, Kensaku Mori, Steven McDonagh, Nils Y Hammerla, Bernhard Kainz, et al. Attention u-net: Learning where to look for the pancreas. arXiv preprint arXiv:1804.03999, 2018. 1, 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.165, + 0.892, + 0.219 + ], + "angle": 0, + "content": "[42] Md Mostafijur Rahman and Radu Marculescu. Medical image segmentation via cascaded attention decoding. In IEEE/CVF Winter Conf. Appl. Comput. Vis., pages 6222-6231, 2023. 1, 3, 4, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.223, + 0.892, + 0.277 + ], + "angle": 0, + "content": "[43] Md Mostafijur Rahman and Radu Marculescu. Multi-scale hierarchical vision transformer with cascaded attention decoding for medical image segmentation. In Med. Imaging Deep Learn., 2023. 1, 3, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.281, + 0.892, + 0.336 + ], + "angle": 0, + "content": "[44] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Int. Conf. Med. Image Comput. Comput. Assist. Interv., pages 234-241. Springer, 2015. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.339, + 0.892, + 0.393 + ], + "angle": 0, + "content": "[45] Mark Sandler, Andrew Howard, Menglong Zhu, Andrey Zh-moginov, and Liang-Chieh Chen. Mobilenetv2: Inverted residuals and linear bottlenecks. In IEEE Conf. Comput. Vis. Pattern Recog., pages 4510-4520, 2018. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.397, + 0.892, + 0.438 + ], + "angle": 0, + "content": "[46] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.441, + 0.892, + 0.509 + ], + "angle": 0, + "content": "[47] Christian Szegedy, Wei Liu, Yangqing Jia, Pierre Sermanet, Scott Reed, Dragomir Anguelov, Dumitru Erhan, Vincent Vanhoucke, and Andrew Rabinovich. Going deeper with convolutions. In IEEE Conf. Comput. Vis. Pattern Recog., pages 1-9, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.512, + 0.892, + 0.553 + ], + "angle": 0, + "content": "[48] Mingxing Tan and Quoc Le. Efficientnet: Rethinking model scaling for convolutional neural networks. In Int. Conf. Mach. Learn., pages 6105-6114. PMLR, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.557, + 0.892, + 0.611 + ], + "angle": 0, + "content": "[49] Zhengzhong Tu, Hossein Talebi, Han Zhang, Feng Yang, Peyman Milanfar, Alan Bovik, and Yinxiao Li. Maxvit: Multi-axis vision transformer. In Eur. Conf. Comput. Vis., pages 459-479. Springer, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.614, + 0.892, + 0.669 + ], + "angle": 0, + "content": "[50] Jeya Maria Jose Valanarasu and Vishal M Patel. Unext: Mlp-based rapid medical image segmentation network. In Int. Conf. Med. Image Comput. Comput. Assist. Interv., pages 23-33. Springer, 2022. 6, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.673, + 0.892, + 0.74 + ], + "angle": 0, + "content": "[51] David Vázquez, Jorge Bernal, F Javier Sánchez, Gloria Fernández-Esparrach, Antonio M López, Adriana Romero, Michal Drozdzal, and Aaron Courville. A benchmark for endoluminal scene segmentation of colonoscopy images. J. Healthc. Eng., 2017, 2017. 6, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.744, + 0.892, + 0.798 + ], + "angle": 0, + "content": "[52] Haonan Wang, Peng Cao, Jiaqi Wang, and Osmar R Zaiane. Uctransnet: rethinking the skip connections in u-net from a channel-wise perspective with transformer. In AAAI, pages 2441-2449, 2022. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.802, + 0.892, + 0.856 + ], + "angle": 0, + "content": "[53] Hongyi Wang, Shiao Xie, Lanfen Lin, Yutaro Iwamoto, Xian-Hua Han, Yen-Wei Chen, and Ruofeng Tong. Mixed transformer u-net for medical image segmentation. In ICASSP, pages 2390–2394. IEEE, 2022. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.859, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[54] Jinfeng Wang, Qiming Huang, Feilong Tang, Jia Meng, Jionglong Su, and Sifan Song. Stepwise feature fusion: Local guides global. arXiv preprint arXiv:2203.03635, 2022. 1, 6" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.52, + 0.956 + ], + "angle": 0, + "content": "11778" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.161 + ], + "angle": 0, + "content": "[55] Wenhai Wang, Enze Xie, Xiang Li, Deng-Ping Fan, Kaitao Song, Ding Liang, Tong Lu, Ping Luo, and Ling Shao. Pyramid vision transformer: A versatile backbone for dense prediction without convolutions. In Int. Conf. Comput. Vis., pages 568-578, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.164, + 0.47, + 0.219 + ], + "angle": 0, + "content": "[56] Wenhai Wang, Enze Xie, Xiang Li, Deng-Ping Fan, Kaitao Song, Ding Liang, Tong Lu, Ping Luo, and Ling Shao. Pvt v2: Improved baselines with pyramid vision transformer. Comput. Vis. Media, 8(3):415-424, 2022. 1, 2, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.221, + 0.47, + 0.262 + ], + "angle": 0, + "content": "[57] Sanghyun Woo, Jongchan Park, Joon-Young Lee, and In So Kweon. Cbam: Convolutional block attention module. In Eur. Conf. Comput. Vis., pages 3–19, 2018. 1, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.264, + 0.47, + 0.33 + ], + "angle": 0, + "content": "[58] Enze Xie, Wenhai Wang, Zhiding Yu, Anima Anandkumar, Jose M Alvarez, and Ping Luo. Segformer: Simple and efficient design for semantic segmentation with transformers. Adv. Neural Inform. Process. Syst., 34:12077-12090, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.333, + 0.47, + 0.388 + ], + "angle": 0, + "content": "[59] Weihao Yu, Mi Luo, Pan Zhou, Chenyang Si, Yichen Zhou, Xinchao Wang, Jiashi Feng, and Shuicheng Yan. Metaformer is actually what you need for vision. In IEEE Conf. Comput. Vis. Pattern Recog., pages 10819-10829, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.391, + 0.47, + 0.445 + ], + "angle": 0, + "content": "[60] Xiangyu Zhang, Xinyu Zhou, Mengxiao Lin, and Jian Sun. Shufflenet: An extremely efficient convolutional neural network for mobile devices. In IEEE Conf. Comput. Vis. Pattern Recog., pages 6848-6856, 2018. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.447, + 0.469, + 0.502 + ], + "angle": 0, + "content": "[61] Yundong Zhang, Huiye Liu, and Qiang Hu. Transfuse: Fusing transformers and cnns for medical image segmentation. In Int. Conf. Med. Image Comput. Comput. Assist. Interv., pages 14-24. Springer, 2021. 1, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.504, + 0.469, + 0.573 + ], + "angle": 0, + "content": "[62] Zongwei Zhou, Md Mahfuzur Rahman Siddiquee, Nima Tajbakhsh, and Jianming Liang. Unet++: A nested u-net architecture for medical image segmentation. In Deep Learn. Med. Image Anal. Multimodal Learn. Clin. Decis. Support, pages 3-11. Springer, 2018. 1, 2, 6" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.573 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.956 + ], + "angle": 0, + "content": "11779" + } + ] +] \ No newline at end of file diff --git a/2024/EMCAD_ Efficient Multi-scale Convolutional Attention Decoding for Medical Image Segmentation/32047b53-ccf5-41e9-bc53-4e712089f958_origin.pdf b/2024/EMCAD_ Efficient Multi-scale Convolutional Attention Decoding for Medical Image Segmentation/32047b53-ccf5-41e9-bc53-4e712089f958_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..c8a08acdd06fe829a748dedfa53666ed4afb0bca --- /dev/null +++ b/2024/EMCAD_ Efficient Multi-scale Convolutional Attention Decoding for Medical Image Segmentation/32047b53-ccf5-41e9-bc53-4e712089f958_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8a18399b13fcea07f75053cd8135ed365e90e488b3c0e39d0c73e750565af542 +size 1077730 diff --git a/2024/EMCAD_ Efficient Multi-scale Convolutional Attention Decoding for Medical Image Segmentation/full.md b/2024/EMCAD_ Efficient Multi-scale Convolutional Attention Decoding for Medical Image Segmentation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..aab39b9fc62f9c62bf75eb26d902250a2eae93a1 --- /dev/null +++ b/2024/EMCAD_ Efficient Multi-scale Convolutional Attention Decoding for Medical Image Segmentation/full.md @@ -0,0 +1,332 @@ +# EMCAD: Efficient Multi-scale Convolutional Attention Decoding for Medical Image Segmentation + +Md Mostafijur Rahman, Mustafa Munir, and Radu Marculescu + +The University of Texas at Austin + +Austin, Texas, USA + +mostafijur.rahman, mmunir, radum@utexas.edu + +# Abstract + +An efficient and effective decoding mechanism is crucial in medical image segmentation, especially in scenarios with limited computational resources. However, these decoding mechanisms usually come with high computational costs. To address this concern, we introduce EMCAD, a new efficient multi-scale convolutional attention decoder, designed to optimize both performance and computational efficiency. EMCAD leverages a unique multi-scale depth-wise convolution block, significantly enhancing feature maps through multi-scale convolutions. EMCAD also employs channel, spatial, and grouped (large-kernel) gated attention mechanisms, which are highly effective at capturing intricate spatial relationships while focusing on salient regions. By employing group and depth-wise convolution, EMCAD is very efficient and scales well (e.g., only 1.91M parameters and 0.381G FLOPs are needed when using a standard encoder). Our rigorous evaluations across 12 datasets that belong to six medical image segmentation tasks reveal that EMCAD achieves state-of-the-art (SOTA) performance with $79.4\%$ and $80.3\%$ reduction in #Params and #FLOPs, respectively. Moreover, EMCAD's adaptability to different encoders and versatility across segmentation tasks further establish EMCAD as a promising tool, advancing the field towards more efficient and accurate medical image analysis. Our implementation is available at https://github.com/SLDGroup/EMCAD. + +# 1. Introduction + +In the realm of medical diagnostics and therapeutic strategies, automated segmentation of medical images is vital, as it classifies pixels to identify critical regions such as lesions, tumors, or entire organs. A variety of U-shaped convolutional neural network (CNN) architectures [20, 24, 37, 41, 44, 62], notably UNet [44], UNet++ [62], UNet3+ [24], and nnU-Net [19], have become standard techniques for this purpose, achieving high-quality, high-resolution segmen + +tation output. Attention mechanisms [12, 17, 20, 41, 57] have also been integrated into these models to enhance feature maps and improve pixel-level classification. Although attention-based models have shown improved performance, they still face significant challenges due to the computationally expensive convolutional blocks that are typically used in conjunction with attention mechanisms. + +Recently, vision transformers [18] have shown promise in medical image segmentation tasks [5, 8, 17, 42, 43, 52, 54, 61] by capturing long-range dependencies among pixels through Self-attention (SA) mechanisms. Hierarchical vision transformers like Swin [34], PVT [55, 56], MaxViT [49], MERIT [43], ConvFormer [33], and MetaFormer [59] have been introduced to further improve the performance in this field. While the SA excels at capturing global information, it is less adept at understanding the local spatial context [13, 28]. To address this limitation, some approaches have integrated local convolutional attention within the decoders to better grasp spatial details. Nevertheless, these methods can still be computationally demanding because they frequently employ costly convolutional blocks. This limits their applicability to real-world scenarios where computational resources are restricted. + +To address the aforementioned limitations, we introduce EMCAD, an efficient multi-scale convolutional attention decoding using a new multi-scale depth-wise convolution block. More precisely, EMCAD enhances the feature maps via efficient multi-scale convolutions, while incorporating complex spatial relationships and local attention through the use of channel, spatial, and grouped (large-kernel) gated attention mechanisms. Our contributions are as follows: + +- New Efficient Multi-scale Convolutional Decoder: We introduce an efficient multi-scale cascaded fullyconvolutional attention decoder (EMCAD) for 2D medical image segmentation; this takes the multi-stage features of vision encoders and progressively enhances the multi-scale and multi-resolution spatial representations. EMCAD has only 0.506M parameters and 0.11G FLOPs for a tiny encoder with #channels = [32, 64, 160, 256], + +![](images/3af94799489a0efcba7e12ed56d3a523e2bdb917ee5fa27705063e9a6672a011.jpg) +Figure 1. Average DICE scores vs. #FLOPs for different methods over 10 binary medical image segmentation datasets. As shown, our approaches (PVT-EMCAD-B0 and PVT-EMCAD-B2) have the lowest #FLOPs, yet the highest DICE scores. + +while it has 1.91M parameters and 0.381G FLOPs for a standard encoder with #channels = [64, 128, 320, 512]. + +- Efficient Multi-scale Convolutional Attention Module: We introduce MSCAM, a new efficient multi-scale convolutional attention module that performs depth-wise convolutions at multiple scales; this refines the feature maps produced by vision encoders and enables capturing multi-scale salient features by suppressing irrelevant regions. The use of depth-wise convolutions makes MSCAM very efficient. +- Large-kernel Grouped Attention Gate: We introduce a new grouped attention gate to fuse refined features with the features from skip connections. By using larger kernel $(3 \times 3)$ group convolutions instead of point-wise convolutions in the design, we capture salient features in a larger local context with less computation. +- Improved Performance: We empirically show that EM-CAD can be used with any hierarchical vision encoder (e.g., PVTv2-B0, PVTv2-B2 [56]), while significantly improving the performance of 2D medical image segmentation. EMCAD produces better results than SOTA methods with a significantly lower computational cost (as shown in Figure 1) on 12 medical image segmentation benchmarks that belong to six different tasks. + +The remaining of this paper is organized as follows: Section 2 summarizes related work. Section 3 describes the proposed method. Section 4 explains our experimental setup and results on 12 medical image segmentation benchmarks. Section 5 covers different ablation experiments. Lastly, Section 6 concludes the paper. + +# 2. Related Work + +# 2.1. Vision encoders + +Convolutional Neural Networks (CNNs) [21-23, 32, 35, 45-48] have been foundational as encoders due to their pro + +ficiency in handling spatial relationships in images. More precisely, AlexNet [32] and VGG [46] pave the way, leveraging deep layers of convolutions to extract features progressively. GoogleNet [47] introduces the inception module, allowing more efficient computation of representations across various scales. ResNet [21] introduces residual connections, enabling the training of networks with substantially more layers by addressing the vanishing gradients problem. MobileNets [22, 45] bring CNNs to mobile devices through lightweight, depth-wise separable convolutions. EfficientNet [48] introduces a scalable architectural design to CNNs with compound scaling. Although CNNs are pivotal for many vision applications, they generally lack the ability to capture long-range dependencies within images due to their inherent local receptive fields. + +Recently, Vision Transformers (ViTs), pioneered by Dosovitskiy et al. [18], enabled the learning of long-range relationships among pixels using Self-attention (SA). Since then, ViTs have been enhanced by integrating CNN features [49, 56], developing novel self-attention (SA) blocks [34, 49], and introducing new architectural designs [55, 58]. The Swin Transformer [34] incorporates a sliding window attention mechanism, while SegFormer [58] leverages MixFFN blocks for hierarchical structures. PVT [55] uses spatial reduction attention, refined in PVTv2 [56] with overlapping patch embedding and a linear complexity attention layer. MaxViT [49] introduces a multi-axis self-attention to form a hierarchical CNN-transformer encoder. Although ViTs address the CNNs limitation in capturing long-range pixel dependencies [21-23, 32, 35, 45-48], they face challenges in capturing the local spatial relationships among pixels. In this paper, we aim to overcome these limitations by introducing a new multi-scale cascaded attention decoder that refines feature maps and incorporates local attention using a multi-scale convolutional attention module. + +# 2.2. Medical image segmentation + +Medical image segmentation involves pixel-wise classification to identify various anatomical structures like lesions, tumors, or organs within different imaging modalities such as endoscopy, MRI, or CT scans [8]. U-shaped networks [7, 19, 24, 26, 37, 41, 44, 62] are particularly favored due to their simple but effective encoder-decoder design. The UNet [44] pioneered this approach with its use of skip connections to fuse features at different resolution stages. UNet++ [62] evolves this design by incorporating nested encoder-decoder pathways with dense skip connections. Expanding on these ideas, UNet $3+$ [24] introduces comprehensive skip pathways that facilitate full-scale feature integration. Further advancement comes with DC-UNet [37], which integrates a multi-resolution convolution scheme and residual paths into its skip connections. The DeepLab series, including DeepLabv3 [10] and DeepLabv3+ [11], in + +produce atrous convolutions and spatial pyramid pooling to handle multi-scale information. SegNet [2] uses pooling indices to upsample feature maps, preserving the boundary details. nnU-Net [19] automatically configures hyperparameters based on the specific dataset characteristics, using standard 2D and 3D UNets. Collectively, these U-shaped models have become a benchmark for success in the domain of medical image segmentation. + +Recently, vision transformers have emerged as a formidable force in medical image segmentation, harnessing the ability to capture pixel relationships at global scales [5, 8, 17, 42, 43, 52, 58, 61]. TransUNet [8] presents a novel blend of CNNs for local feature extraction and transformers for global context, enhancing both local and global feature capture. Swin-Unet [5] extends this by incorporating Swin Transformer blocks [34] into a U-shaped model for both encoding and decoding processes. Building on these concepts, MERIT [43] introduces a multi-scale hierarchical transformer, which employs SA across different window sizes, thus enhancing the model capacity to capture multiscale features critical for medical image segmentation. + +The integration of attention mechanisms has been investigated within CNNs [20, 41] and transformer-based systems [17] for enhancing medical image segmentation. PraNet [20] employs a reverse attention strategy for feature refinement. PolypPVT [17] leverages PVTv2 [56] as its backbone encoder and incorporates CBAM [57] within its decoding stages. The CASCADE [42] presents a novel cascaded decoder, combining channel [23] and spatial [9] attention to refine features at multiple stages, extracted from a transformer encoder, culminating in high-resolution segmentation outputs. While CASCADE achieves notable performance in segmenting medical images by integrating local and global insights from transformers, it is computationally inefficient due to the use of triple $3 \times 3$ convolution layers at each decoder stage. In addition to this, it uses single-scale convolutions during decoding. Our new proposal involves the adoption of multi-scale depth-wise convolutions to mitigate these constraints. + +# 3. Methodology + +In this section, we first introduce our new EMCAD decoder and then explain two transformer-based architectures (i.e., PVT-EMCAD-B0 and PVT-EMCAD-B2) incorporating our proposed decoder. + +# 3.1. Efficient multi-scale convolutional attention decoding (EMCAD) + +In this section, we introduce our efficient multi-scale convolutional decoding (EMCAD) to process the multi-stage features extracted from pretrained hierarchical vision encoders for high-resolution semantic segmentation. As shown in Figure 2(b), EMCAD consists of efficient multi + +scale convolutional attention modules (MSCAMs) to robustly enhance the feature maps, large-kernel grouped attention gates (LGAGs) to refine feature maps fusing with the skip connection via gated attention mechanism, efficient up-convolution blocks (EUCBs) for up-sampling followed by enhancement of feature maps, and segmentation heads (SHs) to produce the segmentation outputs. + +More specifically, we use four MSCAMs to refine pyramid features (i.e., $X1$ , $X2$ , $X3$ , $X4$ in Figure 2) extracted from the four stages of the encoder. After each MSCAM, we use an SH to produce a segmentation map of that stage. Subsequently, we upscale the refined feature maps using EUCBs and add them to the outputs from the corresponding LGAGs. Finally, we add four different segmentation maps to produce the final segmentation output. Different modules of our decoder are described next. + +# 3.1.1 Large-kernel grouped attention gate (LGAG) + +We introduce a new large-kernel grouped attention gate (LGAG) to progressively combine feature maps with attention coefficients, which are learned by the network to allow higher activation of relevant features and suppression of irrelevant ones. This process employs a gating signal derived from higher-level features to control the flow of information across different stages of the network, thus enhancing its precision for medical image segmentation. Unlike Attention UNet [41] which uses $1 \times 1$ convolution to process gating signal $g$ (features from skip connections) and input feature map $x$ (upsampled features), in our $q_{att}(.)$ function, we process $g$ and $x$ by applying separate $3 \times 3$ group convolutions $GC_{g}(.)$ and $GC_{x}(.)$ , respectively. These convolved features are then normalized using batch normalization $(BN(.))$ [27] and merged through element-wise addition. The resultant feature map is activated through a ReLU $(R(.))$ layer [39]. Afterward, we apply a $1 \times 1$ convolution $(C(.))$ followed by $BN(.)$ layer to get a single channel feature map. We then pass the resultant single-channel feature map through a Sigmoid $(\sigma(.))$ activation function to yield the attention coefficients. The output of this transformation is used to scale the input feature $x$ through elementwise multiplication, producing the attention-gated feature $LGAG(g, x)$ . The $LGAG(.)$ (Figure 2(g)) can be formulated as in Equations 1 and 2: + +$$ +q _ {a t t} (g, x) = R \left(B N \left(G C _ {g} (g) + B N \left(G C _ {x} (x)\right)\right)\right) \tag {1} +$$ + +$$ +L G A G (g, x) = x * \sigma \left(B N \left(C \left(q _ {a t t} (g, x)\right)\right)\right) \tag {2} +$$ + +Due to using $3 \times 3$ kernel group convolutions in $q_{att}(.)$ , our LGAG captures comparatively larger spatial contexts with less computational cost. + +![](images/680916196f86f563262f187b1e9498563b3b45b7b797b45690ae9a0e76ac7c8c.jpg) +Figure 2. Hierarchical encoder with newly proposed EMCAD decoder architecture. (a) CNN or transformer encoder with four hierarchical stages, (b) EMCAD decoder, (c) Efficient up-convolution block (EUCB), (d) Multi-scale convolutional attention module (MSCAM), (e) Multi-scale convolution block (MSCB), (f) Multi-scale (parallel) depth-wise convolution (MSDC), (g) Large-kernel grouped attention gate (LGAG), (h) Channel attention block (CAB), and (i) Spatial attention block (SAB). X1, X2, X3, and X4 are the features from the four stages of the hierarchical encoder. p1, p2, p3, and p4 are output segmentation maps from four stages of our decoder. + +# 3.1.2 Multi-scale convolutional attention module (MSCAM) + +We introduce an efficient multi-scale convolutional attention module to refine the feature maps. MSCAM consists of a channel attention block $(CAB(\cdot))$ to put emphasis on pertinent channels, a spatial attention block [9] $(SAB(\cdot))$ to capture the local contextual information, and an efficient multi-scale convolution block $(MSCB(\cdot))$ to enhance the feature maps preserving contextual relationships. The $MSCAM(\cdot)$ (Figure 2(d)) is given in Equation 3: + +$$ +M S C A M (x) = M S C B (S A B (C A B (x))) \tag {3} +$$ + +where $x$ is the input tensor. Due to using depth-wise convolution in multiple scales, our MSCAM is more effective with significantly lower computational cost than the convolutional attention module (CAM) proposed in [42]. + +Multi-scale Convolution Block (MSCB): We introduce an efficient multi-scale convolution block to enhance the features generated by our cascaded expanding path. In our MSCB, we follow the design of the inverted residual block (IRB) of MobileNetV2 [45]. However, unlike IRB, our MSCB performs depth-wise convolution at multiple scales and uses channel_shuffle [60] to shuffle channels across groups. More specifically, in our MSCB, we first expand the number of channels (i.e., expansion_factor = 2) using a point-wise $(1 \times 1)$ convolution layers $PWC_{1}(\cdot)$ followed by a batch normalization layer $BN(\cdot)$ and a ReLU6 [31] activation layer $R6(\cdot)$ . We then use a multi-scale depth-wise convolution $MSDC(\cdot)$ to capture both multi-scale and multi-resolution contexts. As depth-wise convolution overlooks + +the relationships among channels, we use a channel_shuffle operation to incorporate relationships among channels. Afterward, we use another point-wise convolution $PWC_{2}(.)$ followed by a $BN(.)$ to transform back the original #channels, which also encodes dependency among channels. The $MSCB(\cdot)$ (Figure 2(e)) is formulated as in Equation 4: + +$$ +M S C B (x) = B N \left(P W C _ {2} \left(C S \left(M S D C \left(R 6 \left(B N \left(P W C _ {1} (x)\right)\right)\right)\right)\right)\right) \tag {4} +$$ + +where parallel $MSDC(.)$ (Figure 2(f)) for different kernel sizes $(KS)$ can be formulated using Equation 5: + +$$ +M S D C (x) = \sum_ {k s \in K S} D W C B _ {k s} (x) \tag {5} +$$ + +where $DWCB_{ks}(x) = R6(BN(DWC_{ks}(x)))$ . Here, $DWC_{ks}()$ is a depth-wise convolution with the kernel size $ks$ . $BN()$ and $R6()$ are batch normalization and ReLU6 activation, respectively. Additionally, our sequential $MSDC()$ uses the recursively updated input $x$ , where the input $x$ is residually connected to the previous $DWCB_{ks}()$ for better regularization as in Equation 6: + +$$ +x = x + D W C B _ {k s} (x) \tag {6} +$$ + +Channel Attention Block (CAB): We use channel attention block to assign different levels of importance to each channel, thus emphasizing more relevant features while suppressing less useful ones. Basically, the CAB identifies which feature maps to focus on (and then refine them). Following [57], in CAB, we first apply the adaptive maximum pooling $(P_{m}(\cdot))$ and adaptive average pooling $(P_{a}(\cdot))$ to the spatial dimensions (i.e., height and width) to extract the most significant feature of the entire feature map per + +channel. Then, for each pooled feature map, we reduce the number of channels $r = 1 / 16$ times separately using a point-wise convolution $(C_1(\cdot))$ followed by a ReLU activation $(R)$ . Afterward, we recover the original channels using another point-wise convolution $(C_2(\cdot))$ . We then add both recovered feature maps and apply Sigmoid $(\sigma)$ activation to estimate attention weights. Finally, we incorporate these weights to input $x$ using the Hadamard product $(\otimes)$ . The $CAB(\cdot)$ (Figure 2(h)) is defined using Equation 7: + +$$ +C A B (x) = \sigma \left(C _ {2} \left(R \left(C _ {1} \left(P _ {m} (x)\right)\right)\right) + C _ {2} \left(R \left(C _ {1} \left(P _ {a} (x)\right)\right)\right)\right) * x \tag {7} +$$ + +Spatial Attention Block (SAB): We use spatial attention to mimic the attentional processes of the human brain by focusing on specific parts of an input image. Basically, the SAB determines where to focus in a feature map; then it enhances those features. This process enhances the model's ability to recognize and respond to relevant spatial features, which is crucial for image segmentation where the context and location of objects significantly influence the output. In SAB, we first pool maximum $(Ch_{max}(\cdot))$ and average $(Ch_{avg}(\cdot))$ values along the channel dimension to pay attention to local features. Then, we use a large kernel (i.e., $7\times 7$ as in [17]) convolution layer to enhance local contextual relationships among features. Afterward, we apply the Sigmoid activation $(\sigma)$ to calculate attention weights. Finally, we feed these weights to the input $x$ (using Hadamard product $(\otimes)$ to attend information in a more targeted way. The $SAB(.)$ (Figure 2(i)) is defined using Equation 8: + +$$ +S A B (x) = \sigma \left(L K C \left(\left[ C h _ {m a x} (x), C h _ {a v g} (x) \right]\right)\right) * x \tag {8} +$$ + +# 3.1.3 Efficient up-convolution block (EUCB) + +We use an efficient up-convolution block to progressively upsample the feature maps of the current stage to match the dimension and resolution of the feature maps from the next skip connection. The EUCB first uses an UpSampling $Up(\cdot)$ with scale-factor 2 to upscale the feature maps. Then, it enhances the upscaled feature maps by applying a $3 \times 3$ depth-wise convolution $DWC(\cdot)$ followed by a $BN(\cdot)$ and a $ReLU(\cdot)$ activation. Finally, a $1 \times 1$ convolution $C_{1 \times 1}(\cdot)$ is used to reduce the #channels to match with the next stage. The $EUCB(\cdot)$ (Figure 2(c)) is formulated as in Equation 9: + +$$ +E U C B (x) = C _ {1 \times 1} \left(R e L U \left(B N \left(D W C (U p (x))\right)\right)\right) \tag {9} +$$ + +Due to using depth-wise convolution instead of $3 \times 3$ convolution, our EUCB is very efficient. + +# 3.1.4 Segmentation head (SH) + +We use segmentation heads to produce the segmentation outputs from the refined feature maps of four stages of the decoder. The SH layer applies a $1 \times 1$ convolution $Conv_{1 \times 1}(\cdot)$ to the refined feature maps having $ch_i$ channels ( $ch_i$ is the #channels in the feature map of stage $i$ ) and + +produces output with #channels equal to #classes in target dataset for multi-class but 1 channel for binary segmentation. The $SH(\cdot)$ is formulated as in Equation 10: + +$$ +S H (x) = \operatorname {C o n v} _ {1 \times 1} (x) \tag {10} +$$ + +# 3.2. Overall architecture + +To show the generalization, effectiveness, and ability to process multi-scale features for medical image segmentation, we integrate our EMCAD decoder alongside tiny (PVTv2-B0) and standard (PVTv2-B2) networks of PVTv2 [56]. However, our decoder is adaptable and seamlessly compatible with other hierarchical backbone networks. + +PVTv2 differs from conventional transformer patch embedding modules by applying convolutional operations for consistent spatial information capture. Using PVTv2-b0 (Tiny) and PVTv2-b2 (Standard) encoders [56], we develop the PVT-EMCAD-B0 and PVT-EMCAD-B2 architectures. To adopt PVTv2, we first extract the features (X1, X2, X3, and X4) from four layers and feed them (i.e., X4 in the upsample path and X3, X2, X1 in the skip connections) into our EMCAD decoder as shown in Figure 2(a-b). EMCAD then processes them and produces four segmentation maps that correspond to the four stages of the encoder network. + +# 3.3. Multi-stage loss and outputs aggregation + +Our EMCAD decoder's four segmentation heads produce four prediction maps $p_1, p_2, p_3$ , and $p_4$ across its stages. + +Loss aggregation: We adopt a combinatorial approach to loss combination called MUTATION, inspired by the work of MERIT [43] for multi-class segmentation. This involves calculating the loss for all possible combinations of predictions derived from 4 heads, totaling $2^{4} - 1 = 15$ unique predictions, and then summing these losses. We focus on minimizing this cumulative combinatorial loss during the training process. For binary segmentation, we optimize the additive loss like [42] with an additional term $\mathcal{L}_{p_1 + p_2 + p_3 + p_4}$ as in Equation 11: + +$$ +\mathcal {L} _ {\text {t o t a l}} = \alpha \mathcal {L} _ {p _ {1}} + \beta \mathcal {L} _ {p _ {2}} + \gamma \mathcal {L} _ {p _ {3}} + \zeta \mathcal {L} _ {p _ {4}} + \delta \mathcal {L} _ {p _ {1} + p _ {2} + p _ {3} + p _ {4}} \tag {11} +$$ + +where $\mathcal{L}_{p_1},\mathcal{L}_{p_2},\mathcal{L}_{p_3}$ , and $\mathcal{L}_{p_4}$ are the losses of each individual prediction maps. $\alpha = \beta = \gamma = \zeta = \delta = 1.0$ are the weights assigned to each loss. + +Output segmentation maps aggregation: We consider the prediction map, $p_4$ , from the last stage of our decoder as the final segmentation map. Then, we obtain the final segmentation output by employing a Sigmoid function for binary or a Softmax function for multi-class segmentation. + +# 4. Experiments + +In this section, we present the details of our implementation followed by a comparative analysis of our PVT-EMCAD-B0 and PVT-EMCAD-B2 against SOTA methods. Datasets and evaluation metrics are in Supplementary Section 7. + +
Methods#Params#FLOPsPolypSkin LesionCellBUSIAvg.
ClinicColonETISKvasirBKAIISIC17ISIC18DSB18EM
UNet [44]34.53M65.53G92.1183.9576.8582.8785.0583.0786.6792.2395.4674.0485.23
UNet++ [62]9.16M34.65G92.1787.8877.4083.3684.0782.9887.4691.9795.4874.7685.75
AttnUNet [41]34.88M66.64G92.2086.4676.8483.4984.0783.6687.0592.2295.5574.4885.60
DeepLabv3+ [10]39.76M14.92G93.2491.9290.7389.0689.7483.8488.6492.1494.9676.8189.11
PraNet [20]32.55M6.93G91.7189.1683.8484.8285.5683.0388.5689.8992.3775.1486.41
CaraNet [38]46.64M11.48G94.0891.1990.2589.7489.7185.0290.1889.1592.7877.3488.94
UACANet-L [30]69.16M31.51G94.1691.0289.7790.1790.3583.7289.7688.8689.2876.9688.41
SSFormer-L [54]66.22M17.28G94.1892.1190.1691.4791.1485.2890.2592.0394.9578.7690.03
PolypPVT [17]25.11M5.30G94.1391.5389.9391.5691.1785.5690.3690.6994.4079.3589.87
TransUNet [8]105.32M38.52G93.9091.6387.7991.0889.1785.0089.1692.0495.2778.3089.33
SwinUNet [5]27.17M6.2G92.4289.2785.1089.5987.6183.9789.2691.0394.4777.3888.01
TransFuse [61]143.74M82.71G93.6290.3586.9190.2487.4784.8989.6290.8594.3579.3688.77
UNeXt [50]1.47M0.57G90.2083.8474.0377.8877.9382.7487.7886.0193.8174.7182.89
PVT-CASCADE [42]34.12M7.62G94.5391.6091.0392.0592.1485.5090.4192.3595.4279.2190.42
PVT-EMCAD-B0 (Ours)3.92M0.84G94.6091.7191.6591.9591.3085.6790.7092.4695.3579.8090.52
PVT-EMCAD-B2 (Ours)26.76M5.6G95.2192.3192.2992.7592.9685.9590.9692.7495.5380.2591.10
+ +Table 1. Results of binary medical image segmentation (i.e., polyp, skin lesion, cell, and breast cancer). We reproduce the results of SOTA methods using their publicly available implementation with our train-val-test splits of 80:10:10. #FLOPs of all the methods are reported for $256 \times 256$ inputs, except Swin-UNet ( $224 \times 224$ ). All results are averaged over five runs. Best results are shown in bold. + +
ArchitecturesAverageAortaGBKLKRLiverPCSPSM
DICE↑HD95↓mIoU↑
UNet [44]70.1144.6959.3984.0056.7072.4162.6486.9848.7381.4867.96
AttnUNet [41]71.7034.4761.3882.6161.9476.0770.4287.5446.7080.6767.66
R50+UNet [8]74.6836.87-84.1862.8479.1971.2993.3548.2384.4173.92
R50+AttnUNet [8]75.5736.97-55.9263.9179.2072.7193.5649.3787.1974.95
SSFormer [54]78.0125.7267.2382.7863.7480.7278.1193.5361.5387.0776.61
PolypPVT [17]78.0825.6167.4382.3466.1481.2173.7894.3759.3488.0579.4
TransUNet [8]77.6126.967.3286.5660.4380.5478.5394.3358.4787.0675.00
SwinUNet [5]77.5827.3266.8881.7665.9582.3279.2293.7353.8188.0475.79
MT-UNet [53]78.5926.59-87.9264.9981.4777.2993.0659.4687.7576.81
MISSformer [25]81.9618.20-86.9968.6585.2182.0094.4165.6791.9280.81
PVT-CASCADE [42]81.0620.2370.8883.0170.5982.2380.3794.0864.4390.183.69
TransCASCADE [42]82.6817.3473.4886.6368.4887.6684.5694.4365.3390.7983.52
PVT-EMCAD-B0 (Ours)81.9717.3972.6487.2166.6287.4883.9694.5762.0092.6681.22
PVT-EMCAD-B2 (Ours)83.6315.6874.6588.1468.8788.0884.1095.2668.5192.1783.92
+ +Table 2. Results of abdomen organ segmentation on Synapse Multi-organ dataset. DICE scores are reported for individual organs. Results of UNet, AttnUNet, PolypPVT, SSFormerPVT, TransUNet, and SwinUNet are taken from [42]. $\uparrow (\downarrow)$ denotes the higher (lower) the better. $^{\prime} - ^{\prime}$ means missing data from the source. EMCAD results are averaged over five runs. Best results are shown in bold. + +# 4.1. Implementation details + +We implement our network and conduct experiments using Pytorch 1.11.0 on a single NVIDIA RTX A6000 GPU with 48GB of memory. We utilize ImageNet [16] pretrained PVTv2-b0 and PVTv2-b2 [56] as encoders. In the MSDC of our decoder, we set the multi-scale kernels [1,3,5] through an ablation study. We use the parallel arrangement of depth-wise convolutions in all experiments. Our models are trained using the AdamW optimizer [36] with a learning rate and weight decay of $1e - 4$ . We generally train for 200 epochs with a batch size of 16, except for Synapse multi-organ (300 epochs, batch size 6) and ACDC cardiac organ (400 epochs, batch size 12), saving the best model based on the DICE score. We resize images + +to $352 \times 352$ and use a multi-scale $\{0.75, 1.0, 1.25\}$ training strategy with a gradient clip limit of 0.5 for ClinicDB [3], Kvasir [29], ColonDB [51], ETIS [51], BKAI [40], ISIC17 [15], and ISIC18 [15], while we resize images to $256 \times 256$ for BUSI [1], EM [6], and DSB18 [4]. For Synapse and ACDC datasets, images are resized to $224 \times 224$ , with random rotation and flipping augmentations, optimizing a combined Cross-entropy (0.3) and DICE (0.7) loss. For binary segmentation, we utilize the combined weighted BinaryCrossEntropy (BCE) and weighted IoU loss function. + +# 4.2. Results + +We compare our architectures (i.e., PVT-EMCAD-B0 and PVT-EMCAD-B2) with SOTA CNN and transformer-based + +![](images/68e95713cdd7aa1342c98b587e762e015c98e20ad9f521d47ceae1aa69432ee7.jpg) +Figure 3. Average DICE scores vs. #Params for different methods over 10 binary medical image segmentation datasets. As shown, our proposed approaches (PVT-EMCAD-B0 and PVT-EMCAD-B2) have the fewest parameters, yet the highest DICE scores. + +
MethodsAvg. DICERVMyoLV
R50+UNet [8]87.5587.1080.6394.92
R50+AttnUNet [8]86.7587.5879.2093.47
ViT+CUP [8]81.4581.4670.7192.18
R50+ViT+CUP [8]87.5786.0781.8894.75
TransUNet [8]89.7186.6787.2795.18
SwinUNet [5]88.0785.7784.4294.03
MT-UNet [53]90.4386.6489.0495.62
MISSformer [25]90.8689.5588.0494.99
PVT-CASCADE [42]91.4689.9788.995.50
TransCASCade [42]91.6390.2589.1495.50
Cascaded MERIT [43]91.8590.2389.5395.80
PVT-EMCAD-B0 (Ours)91.34±0.289.3788.9995.65
PVT-EMCAD-B2 (Ours)92.12±0.290.6589.6896.02
+ +segmentation methods on 12 datasets that belong to six medical image segmentation tasks. Qualitative results are in the Supplementary Section 7.3. + +# 4.2.1 Results of binary medical image segmentation + +Results for different methods on 10 binary medical image segmentation datasets are shown in Table 1 and Figure 1. Our PVT-EMCAD-B2 attains the highest average DICE score $(91.10\%)$ with only $26.76\mathrm{M}$ parameters and 5.6G FLOPs. The multi-scale depth-wise convolution in our EMCAD decoder, combined with the transformer encoder, contributes to these performance gains. + +Polyp segmentation: Table 1 reveals that our PVT-EMCAD-B2 surpasses all SOTA methods in five polyp segmentation datasets. PVT-EMCAD-B2 achieves DICE score improvements of $1.08\%$ , $0.78\%$ , $2.36\%$ , $1.19\%$ , and $1.79\%$ over PolypPVT in ClinicDB, ColonDB, ETIS, Kvasir, and BKAI-IGI, despite having slightly more parameters and FLOPs. The smallest model UNeXt, ex + +Table 3. Results of cardiac organ segmentation on ACDC dataset. DICE scores $(\%)$ are reported for individual organs. We get the results of SwinUNet from [42]. Best results are shown in bold. + +
Components#FLOPs(G)#Params (M)Avg DICE
CascadedLGAGMSCAM224256
NoNoNo00080.10±0.2
YesNoNo0.1000.1310.22481.08±0.2
YesYesNo0.1080.1410.23581.92±0.2
YesNoYes0.3730.4871.89882.86±0.3
YesYesYes0.3810.4981.9183.63±0.3
+ +Table 4. Effect of different components of EMCAD with PVTv2-b2 encoder on Synapse multi-organ dataset. #FLOPs are reported for input resolution of $224 \times 224$ and $256 \times 256$ . All results are averaged over five runs. Best results are shown in bold. + +hibits the worst performance in all five polyp segmentation datasets. Our smaller model with only 3.92M parameters and 0.84G FLOPs also outperforms all the methods except PVT-CASCADE (in Kvasir and BKAI-IGH) and SSFormer-L (in ColonDB), which achieve the best performance among SOTA methods. In conclusion, our PVT-EMCAD-B2 achieves the new SOTA results in these five polyp segmentation datasets. + +Skin lesion segmentation: Table 1 shows PVT-EMCAD-B2's strong performance on ISIC17 and ISIC18 skin lesion segmentation datasets, achieving DICE scores of $85.95\%$ and $90.96\%$ , surpassing DeepLabV3+ by $2.11\%$ and $2.32\%$ . It also beats the nearest method PVT-CASCADE by $0.45\%$ and $0.55\%$ in ISIC17 and ISIC18, respectively, though our decoder is significantly more efficient than CASCADE. Our PVT-EMCAD-B0 also shows huge potential in point care applications like skin lesion segmentation with only 3.92M parameters and 0.84G FLOPs. + +Cell segmentation: To evaluate our method's effectiveness in biological imaging, we use DSB18 [4] for cell nuclei and EM [6] for cell structure segmentation. As Table 1 indicates, our PVT-EMCAD-B2 sets a SOTA benchmark in cell nuclei segmentation on DSB18, outperforming DeepLabv3+, TransFuse, and PVT-CASCADE. On the EM dataset, PVT-EMCAD-B2 secures the second-best DICE score (95.53%), offering significantly lower computational costs than the top-performing AttnUNet (95.55%). + +Breast cancer segmentation: We conduct experiments on the BUSI dataset for breast cancer segmentation in ultrasound images. Our PVT-EMCAD-B2 achieves the SOTA DICE score $(80.25\%)$ on this dataset. Furthermore, our PVT-EMCAD-B0 outperforms the computationally similar method UNeXt by a notable margin of $5.54\%$ . + +# 4.2.2 Results of abdomen organ segmentation + +Table 2 shows that our PVT-EMCAD-B2 excels in abdomen organ segmentation on the Synapse multi-organ dataset, achieving the highest average DICE score of $83.63\%$ and surpassing all SOTA CNN- and transformer-based methods. It outperforms PVT-CASCADE by $2.57\%$ in DICE score and 4.55 in HD95 distance, indicating superior organ boundary location. Our EMCAD decoder boosts individ + +
Conv. kernels[1][3][5][1,3][3,3][1,3,5][3,3,3][3,5,7][1,3,5,7][1,3,5,7,9]
Synapse82.4382.7982.7482.9882.8183.6382.9283.1183.5783.34
ClinicDB94.8194.9094.9895.1395.0695.2195.1595.0395.1895.07
+ +Table 5. Effect of multi-scale kernels in the depth-wise convolution of MSDC on ClinicDB and Synapse multi-organ datasets. We use the PVTv2-b2 encoder for these experiments. All results are averaged over five runs. Best results are highlighted in bold. + +
EncodersDecoders#FLOPs(G)#Params(M)DICE (%)
PVTv2-B0CASCADE0.4392.3280.54
PVTv2-B0EMCAD (Ours)0.1100.50781.97
PVTv2-B2CASCADE1.939.2782.78
PVTv2-B2EMCAD (Ours)0.3811.9183.63
+ +Table 6. Comparison with the baseline decoder on Synapse Multi-organ dataset. We only report the #FLOPs (with input resolution of $224 \times 224$ ) and the #parameters of the decoders. All the results are averaged over five runs. Best results are shown in bold. + +ual organ segmentation, significantly outperforming SOTA methods on six of eight organs. + +# 4.2.3 Results of cardiac organ segmentation + +Table 3 shows the DICE scores of our PVT-EMCAD-B2 and PVT-EMCAD-B0 along with other SOTA methods, on the MRI images of the ACDC dataset for cardiac organ segmentation. Our PVT-EMCAD-B2 achieves the highest average DICE score of $92.12\%$ , thus improving about $0.27\%$ over Cascaded MERIT though our network has significantly lower computational cost. Besides, PVT-EMCAD-B2 has better DICE scores in all three organ segmentations. + +# 5. Ablation Studies + +In this section, we conduct ablation studies to explore different aspects of our architectures and the experimental framework. More ablations are in Supplementary Section 8. + +# 5.1. Effect of different components of EMCAD + +We conduct a set of experiments on the Synapse multi-organ dataset to understand the effect of different components of our EMCAD decoder. We start with only the encoder and add different modules such as Cascaded structure, LGAG, and MSCAM to understand their effect. Table 4 exhibits that the cascaded structure of the decoder helps to improve performance over the non-cascaded one. The incorporation of LGAG and MSCAM improves performance, however, MSCAM proves to be more effective. When both the LGAG and MSCAM modules are used together, it produces the best DICE score of $83.63\%$ . It is also evident that there is about $3.53\%$ improvement in the DICE score with an additional 0.381G FLOPs and 1.91M parameters. + +# 5.2. Effect of multi-scale kernels in MSCAM + +We have conducted another set of experiments on Synapse multi-organ and ClinicDB datasets to understand the effect of different multi-scale kernels used for depth-wise convolutions in MSDC. Table 5 reports these results which show + +that performance improves from $1 \times 1$ to $3 \times 3$ kernel. When $1 \times 1$ kernel is used together with $3 \times 3$ it improves more than when using them alone. However, when two $3 \times 3$ kernels are used together, performance drops. The incorporation of a $5 \times 5$ kernel with $1 \times 1$ and $3 \times 3$ kernels further improves the performance and it achieves the best results in both Synapse multi-organ and ClinicDB datasets. If we add additional larger kernels (e.g., $7 \times 7$ , $9 \times 9$ ), the performance of both datasets drops. Based on these empirical observations, we choose [1, 3, 5] kernels in all our experiments. + +# 5.3. Comparison with the baseline decoder + +In Table 6, we report the experimental results with the computational complexity of our EMCAD decoder and a baseline decoder, namely CASCADE. From Table 6, we can see that our EMCAD decoder with PVTv2-b2 requires $80.3\%$ fewer FLOPs and $79.4\%$ fewer parameters to outperform (by $0.85\%$ ) the respective CASCADE decoder. Similarly, our EMCAD decoder with PVTv2-B0 achieves $1.43\%$ better DICE score than the CASCADE decoder with $78.1\%$ fewer parameters and $74.9\%$ fewer FLOPs. + +# 6. Conclusions + +In this paper, we have presented EMCAD, a new and efficient multi-scale convolutional attention decoder designed for multi-stage feature aggregation and refinement in medical image segmentation. EMCAD employs a multi-scale depth-wise convolution block, which is key for capturing diverse scale information within feature maps, a critical factor for precision in medical image segmentation. This design choice, using depth-wise convolutions instead of standard $3 \times 3$ convolution blocks, makes EMCAD notably efficient. + +Our experiments reveal that EMCAD surpasses the recent CASCADE decoder in DICE scores with $79.4\%$ fewer parameters and $80.3\%$ less FLOPs. Our extensive experiments also confirm EMCAD's superior performance compared to SOTA methods across 12 public datasets covering six different 2D medical image segmentation tasks. EMCAD's compatibility with smaller encoders makes it an excellent fit for point-of-care applications while maintaining high performance. We anticipate that our EMCAD decoder will be a valuable asset in enhancing a variety of medical image segmentation and semantic segmentation tasks. + +Acknowledgements: This work is supported in part by the NSF grant CNS 2007284, and in part by the iMAGiNE Consortium (https://imagine.utexas.edu/). + +# References + +[1] Walid Al-Dhabyani, Mohammed Gomaa, Hussien Khaled, and Aly Fahmy. Dataset of breast ultrasound images. Data in brief, 28:104863, 2020. 6, 1 +[2] Vijay Badrinarayanan, Alex Kendall, and Roberto Cipolla. Segnet: A deep convolutional encoder-decoder architecture for image segmentation. IEEE Trans. Pattern Anal. Mach. Intell., 39(12):2481-2495, 2017. 3 +[3] Jorge Bernal, F Javier Sánchez, Gloria Fernández-Esparrach, Debora Gil, Cristina Rodríguez, and Fernando Vilarino. Wm-dova maps for accurate polyp highlighting in colonoscopy: Validation vs. saliency maps from physicians. Comput. Med. Imaging Graph., 43:99-111, 2015. 6, 1 +[4] Juan C Caicedo, Allen Goodman, Kyle W Karhohs, Beth A Cimini, Jeanelle Ackerman, Marzieh Haghighi, CherKeng Heng, Tim Becker, Minh Doan, Claire McQuin, et al. Nucleus segmentation across imaging experiments: the 2018 data science bowl. Nature methods, 16(12):1247-1253, 2019. 6, 7, 1 +[5] Hu Cao, Yueyue Wang, Joy Chen, Dongsheng Jiang, Xiaopeng Zhang, Qi Tian, and Manning Wang. Swin-unet: Unet-like pure transformer for medical image segmentation. arXiv preprint arXiv:2105.05537, 2021. 1, 3, 6, 7 +[6] Albert Cardona, Stephan Saalfeld, Stephan Preibisch, Benjamin Schmid, Anchi Cheng, Jim Pulokas, Pavel Tomancak, and Volker Hartenstein. An integrated micro-and macroarchitectural analysis of the drosophila brain by computer-assisted serial section electron microscopy. PLoS biology, 8(10):e1000502, 2010. 6, 7, 1 +[7] Gongping Chen, Lei Li, Yu Dai, Jianxun Zhang, and Moi Hoon Yap. Aau-net: an adaptive attention u-net for breast lesions segmentation in ultrasound images. IEEE Trans. Med. Imaging, 2022. 2 +[8] Jieneng Chen, Yongyi Lu, Qihang Yu, Xiangde Luo, Ehsan Adeli, Yan Wang, Le Lu, Alan L Yuille, and Yuyin Zhou. Transunet: Transformers make strong encoders for medical image segmentation. arXiv preprint arXiv:2102.04306, 2021. 1, 2, 3, 6, 7 +[9] Long Chen, Hanwang Zhang, Jun Xiao, Liqiang Nie, Jian Shao, Wei Liu, and Tat-Seng Chua. Sca-cnn: Spatial and channel-wise attention in convolutional networks for image captioning. In IEEE Conf. Comput. Vis. Pattern Recog., pages 5659-5667, 2017. 3, 4 +[10] Liang-Chieh Chen, George Papandreou, Iasonas Kokkinos, Kevin Murphy, and Alan L Yuille. Deeplab: Semantic image segmentation with deep convolutional nets, atrous convolution, and fully connected crfs. IEEE Trans. Pattern Anal. Mach. Intell., 40(4):834-848, 2017. 2, 6 +[11] Liang-Chieh Chen, Yukun Zhu, George Papandreou, Florian Schroff, and Hartwig Adam. Encoder-decoder with atrous separable convolution for semantic image segmentation. In Eur. Conf. Comput. Vis., pages 801-818, 2018. 2 +[12] Shuhan Chen, Xiuli Tan, Ben Wang, and Xuelong Hu. Reverse attention for salient object detection. In Eur. Conf. Comput. Vis., pages 234-250, 2018. 1 +[13] Xiangxiang Chu, Zhi Tian, Bo Zhang, Xinlong Wang, Xiaolin Wei, Huaxia Xia, and Chunhua Shen. Conditional po + +sitional encodings for vision transformers. arXiv preprint arXiv:2102.10882, 2021. 1 +[14] Noel Codella, Veronica Rotemberg, Philipp Tschandl, M Emre Celebi, Stephen Dusza, David Gutman, Brian Helba, Aadi Kalloo, Konstantinos Liopyris, Michael Marchetti, et al. Skin lesion analysis toward melanoma detection 2018: A challenge hosted by the international skin imaging collaboration (isic). arXiv preprint arXiv:1902.03368, 2019. 1 +[15] Noel CF Codella, David Gutman, M Emre Celebi, Brian Helba, Michael A Marchetti, Stephen W Dusza, Aadi Kalloo, Konstantinos Liopyris, Nabin Mishra, Harald Kittler, et al. Skin lesion analysis toward melanoma detection: A challenge at the 2017 international symposium on biomedical imaging (isbi), hosted by the international skin imaging collaboration (isic). In IEEE Int. Symp. Biomed. Imaging, pages 168-172. IEEE, 2018. 6, 1 +[16] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In IEEE Conf. Comput. Vis. Pattern Recog., pages 248-255. IEEE, 2009. 6 +[17] Bo Dong, Wenhai Wang, Deng-Ping Fan, Jinpeng Li, Huazhu Fu, and Ling Shao. Polyp-pvt: Polyp segmentation with pyramid vision transformers. arXiv preprint arXiv:2108.06932, 2021. 1, 3, 5, 6 +[18] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 1, 2 +[19] Isensee et al. nnu-net: a self-configuring method for deep learning-based biomedical image segmentation. Nature methods, 18(2):203-211, 2021. 1, 2, 3 +[20] Deng-Ping Fan, Ge-Peng Ji, Tao Zhou, Geng Chen, Huazhu Fu, Jianbing Shen, and Ling Shao. Pranet: Parallel reverse attention network for polyp segmentation. In Int. Conf. Med. Image Comput. Comput. Assist. Interv., pages 263-273. Springer, 2020. 1, 3, 6 +[21] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In IEEE Conf. Comput. Vis. Pattern Recog., pages 770-778, 2016. 2 +[22] Andrew G Howard, Menglong Zhu, Bo Chen, Dmitry Kalenichenko, Weijun Wang, Tobias Weyand, Marco Andreetto, and Hartwig Adam. Mobilenets: Efficient convolutional neural networks for mobile vision applications. arXiv preprint arXiv:1704.04861, 2017. 2 +[23] Jie Hu, Li Shen, and Gang Sun. Squeeze-and-excitation networks. In IEEE Conf. Comput. Vis. Pattern Recog., pages 7132-7141, 2018. 2, 3 +[24] Huimin Huang, Lanfen Lin, Ruofeng Tong, Hongjie Hu, Qiaowei Zhang, Yutaro Iwamoto, Xianhua Han, Yen-Wei Chen, and Jian Wu. Unet 3+: A full-scale connected unet for medical image segmentation. In ICASSP, pages 1055–1059. IEEE, 2020. 1, 2 +[25] Xiaohong Huang, Zhifang Deng, Dandan Li, and Xueguang Yuan. Missformer: An effective medical image segmentation transformer. arXiv preprint arXiv:2109.07162, 2021. 6, 7 + +[26] Nabil Ibtehaz and Daisuke Kihara. Acc-unet: A completely convolutional unet model for the 2020s. In Int. Conf. Med. Image Comput. Comput. Assist. Interv., pages 692-702. Springer, 2023. 2 +[27] Sergey Ioffe and Christian Szegedy. Batch normalization: Accelerating deep network training by reducing internal covariate shift. In Int. Conf. Mach. Learn., pages 448-456. pmlr, 2015. 3 +[28] Md Amirul Islam, Sen Jia, and Neil DB Bruce. How much position information do convolutional neural networks encode? arXiv preprint arXiv:2001.08248, 2020. 1 +[29] Debesh Jha, Pia H Smedsrud, Michael A Riegler, Pål Halvorsen, Thomas de Lange, Dag Johansen, and Håvard D Johansen. Kvasir-seg: A segmented polyp dataset. In Int. Conf. Multimedia Model., pages 451-462. Springer, 2020. 6, 1 +[30] Taehun Kim, Hyemin Lee, and Daijin Kim. Uacanet: Uncertainty augmented context attention for polyp segmentation. In ACM Int. Conf. Multimedia, pages 2167-2175, 2021. 6 +[31] Alex Krizhevsky and Geoff Hinton. Convolutional deep belief networks on CIFar-10. Unpublished manuscript, 40(7): 1-9, 2010. 4 +[32] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convolutional neural networks. Adv. Neural Inform. Process. Syst., 25, 2012. 2 +[33] Xian Lin, Zengqiang Yan, Xianbo Deng, Chuansheng Zheng, and Li Yu. Convformer: Plug-and-play cnn-style transformers for improving medical image segmentation. In Int. Conf. Med. Image Comput. Comput. Assist. Interv., pages 642-651. Springer, 2023. 1 +[34] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Int. Conf. Comput. Vis., pages 10012-10022, 2021. 1, 2, 3 +[35] Zhuang Liu, Hanzi Mao, Chao-Yuan Wu, Christoph Feichtenhofer, Trevor Darrell, and Saining Xie. A convnet for the 2020s. In IEEE Conf. Comput. Vis. Pattern Recog., pages 11976-11986, 2022. 2 +[36] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 6 +[37] Ange Lou, Shuyue Guan, and Murray Loew. Dc-unet: rethinking the u-net architecture with dual channel efficient cnn for medical image segmentation. In Med. Imaging 2021: Image Process., pages 758-768. SPIE, 2021. 1, 2 +[38] Ange Lou, Shuyue Guan, Hanseok Ko, and Murray H Loew. Caranet: context axial reverse attention network for segmentation of small medical objects. In Med. Imaging 2022: Image Process., pages 81-92. SPIE, 2022. 6 +[39] Vinod Nair and Geoffrey E Hinton. Rectified linear units improve restricted boltzmann machines. In Int. Conf. Mach. Learn., pages 807-814, 2010. 3 +[40] Phan Ngoc Lan, Nguyen Sy An, Dao Viet Hang, Dao Van Long, Tran Quang Trung, Nguyen Thi Thuy, and Dinh Viet Sang. Neounet: Towards accurate colon polyp segmentation and neoplasm detection. In Adv. Vis. Comput. - Int. Symp., pages 15-28. Springer, 2021. 6, 1 + +[41] Ozan Oktay, Jo Schlemper, Loic Le Folgoc, Matthew Lee, Mattias Heinrich, Kazunari Misawa, Kensaku Mori, Steven McDonagh, Nils Y Hammerla, Bernhard Kainz, et al. Attention u-net: Learning where to look for the pancreas. arXiv preprint arXiv:1804.03999, 2018. 1, 2, 3, 6 +[42] Md Mostafijur Rahman and Radu Marculescu. Medical image segmentation via cascaded attention decoding. In IEEE/CVF Winter Conf. Appl. Comput. Vis., pages 6222-6231, 2023. 1, 3, 4, 5, 6, 7 +[43] Md Mostafijur Rahman and Radu Marculescu. Multi-scale hierarchical vision transformer with cascaded attention decoding for medical image segmentation. In Med. Imaging Deep Learn., 2023. 1, 3, 5, 7 +[44] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Int. Conf. Med. Image Comput. Comput. Assist. Interv., pages 234-241. Springer, 2015. 1, 2, 6 +[45] Mark Sandler, Andrew Howard, Menglong Zhu, Andrey Zh-moginov, and Liang-Chieh Chen. Mobilenetv2: Inverted residuals and linear bottlenecks. In IEEE Conf. Comput. Vis. Pattern Recog., pages 4510-4520, 2018. 2, 4 +[46] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556, 2014. 2 +[47] Christian Szegedy, Wei Liu, Yangqing Jia, Pierre Sermanet, Scott Reed, Dragomir Anguelov, Dumitru Erhan, Vincent Vanhoucke, and Andrew Rabinovich. Going deeper with convolutions. In IEEE Conf. Comput. Vis. Pattern Recog., pages 1-9, 2015. 2 +[48] Mingxing Tan and Quoc Le. Efficientnet: Rethinking model scaling for convolutional neural networks. In Int. Conf. Mach. Learn., pages 6105-6114. PMLR, 2019. 2 +[49] Zhengzhong Tu, Hossein Talebi, Han Zhang, Feng Yang, Peyman Milanfar, Alan Bovik, and Yinxiao Li. Maxvit: Multi-axis vision transformer. In Eur. Conf. Comput. Vis., pages 459-479. Springer, 2022. 1, 2 +[50] Jeya Maria Jose Valanarasu and Vishal M Patel. Unext: Mlp-based rapid medical image segmentation network. In Int. Conf. Med. Image Comput. Comput. Assist. Interv., pages 23-33. Springer, 2022. 6, 1 +[51] David Vázquez, Jorge Bernal, F Javier Sánchez, Gloria Fernández-Esparrach, Antonio M López, Adriana Romero, Michal Drozdzal, and Aaron Courville. A benchmark for endoluminal scene segmentation of colonoscopy images. J. Healthc. Eng., 2017, 2017. 6, 1 +[52] Haonan Wang, Peng Cao, Jiaqi Wang, and Osmar R Zaiane. Uctransnet: rethinking the skip connections in u-net from a channel-wise perspective with transformer. In AAAI, pages 2441-2449, 2022. 1, 3 +[53] Hongyi Wang, Shiao Xie, Lanfen Lin, Yutaro Iwamoto, Xian-Hua Han, Yen-Wei Chen, and Ruofeng Tong. Mixed transformer u-net for medical image segmentation. In ICASSP, pages 2390–2394. IEEE, 2022. 6, 7 +[54] Jinfeng Wang, Qiming Huang, Feilong Tang, Jia Meng, Jionglong Su, and Sifan Song. Stepwise feature fusion: Local guides global. arXiv preprint arXiv:2203.03635, 2022. 1, 6 + +[55] Wenhai Wang, Enze Xie, Xiang Li, Deng-Ping Fan, Kaitao Song, Ding Liang, Tong Lu, Ping Luo, and Ling Shao. Pyramid vision transformer: A versatile backbone for dense prediction without convolutions. In Int. Conf. Comput. Vis., pages 568-578, 2021. 1, 2 +[56] Wenhai Wang, Enze Xie, Xiang Li, Deng-Ping Fan, Kaitao Song, Ding Liang, Tong Lu, Ping Luo, and Ling Shao. Pvt v2: Improved baselines with pyramid vision transformer. Comput. Vis. Media, 8(3):415-424, 2022. 1, 2, 3, 5, 6 +[57] Sanghyun Woo, Jongchan Park, Joon-Young Lee, and In So Kweon. Cbam: Convolutional block attention module. In Eur. Conf. Comput. Vis., pages 3–19, 2018. 1, 3, 4 +[58] Enze Xie, Wenhai Wang, Zhiding Yu, Anima Anandkumar, Jose M Alvarez, and Ping Luo. Segformer: Simple and efficient design for semantic segmentation with transformers. Adv. Neural Inform. Process. Syst., 34:12077-12090, 2021. 2, 3 +[59] Weihao Yu, Mi Luo, Pan Zhou, Chenyang Si, Yichen Zhou, Xinchao Wang, Jiashi Feng, and Shuicheng Yan. Metaformer is actually what you need for vision. In IEEE Conf. Comput. Vis. Pattern Recog., pages 10819-10829, 2022. 1 +[60] Xiangyu Zhang, Xinyu Zhou, Mengxiao Lin, and Jian Sun. Shufflenet: An extremely efficient convolutional neural network for mobile devices. In IEEE Conf. Comput. Vis. Pattern Recog., pages 6848-6856, 2018. 4 +[61] Yundong Zhang, Huiye Liu, and Qiang Hu. Transfuse: Fusing transformers and cnns for medical image segmentation. In Int. Conf. Med. Image Comput. Comput. Assist. Interv., pages 14-24. Springer, 2021. 1, 3, 6 +[62] Zongwei Zhou, Md Mahfuzur Rahman Siddiquee, Nima Tajbakhsh, and Jianming Liang. Unet++: A nested u-net architecture for medical image segmentation. In Deep Learn. Med. Image Anal. Multimodal Learn. Clin. Decis. Support, pages 3-11. Springer, 2018. 1, 2, 6 \ No newline at end of file diff --git a/2024/EMCAD_ Efficient Multi-scale Convolutional Attention Decoding for Medical Image Segmentation/images.zip b/2024/EMCAD_ Efficient Multi-scale Convolutional Attention Decoding for Medical Image Segmentation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..8a079be0ed1dadbed8cabf113ffa0cb0ea7e607c --- /dev/null +++ b/2024/EMCAD_ Efficient Multi-scale Convolutional Attention Decoding for Medical Image Segmentation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2d41e832c1e30177b3d0d37c9c5b80d8bd020394b77ae8b7b06630db65a850b4 +size 666005 diff --git a/2024/EMCAD_ Efficient Multi-scale Convolutional Attention Decoding for Medical Image Segmentation/layout.json b/2024/EMCAD_ Efficient Multi-scale Convolutional Attention Decoding for Medical Image Segmentation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..0e41acf0bba339c5273cf7405e531d276ef2df20 --- /dev/null +++ b/2024/EMCAD_ Efficient Multi-scale Convolutional Attention Decoding for Medical Image Segmentation/layout.json @@ -0,0 +1,9856 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 59, + 102, + 536, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 102, + 536, + 140 + ], + "spans": [ + { + "bbox": [ + 59, + 102, + 536, + 140 + ], + "type": "text", + "content": "EMCAD: Efficient Multi-scale Convolutional Attention Decoding for Medical Image Segmentation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 145, + 161, + 448, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 161, + 448, + 175 + ], + "spans": [ + { + "bbox": [ + 145, + 161, + 448, + 175 + ], + "type": "text", + "content": "Md Mostafijur Rahman, Mustafa Munir, and Radu Marculescu" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 215, + 176, + 378, + 189 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 215, + 176, + 378, + 189 + ], + "spans": [ + { + "bbox": [ + 215, + 176, + 378, + 189 + ], + "type": "text", + "content": "The University of Texas at Austin" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 248, + 190, + 346, + 202 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 248, + 190, + 346, + 202 + ], + "spans": [ + { + "bbox": [ + 248, + 190, + 346, + 202 + ], + "type": "text", + "content": "Austin, Texas, USA" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 177, + 205, + 412, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 177, + 205, + 412, + 217 + ], + "spans": [ + { + "bbox": [ + 177, + 205, + 412, + 217 + ], + "type": "text", + "content": "mostafijur.rahman, mmunir, radum@utexas.edu" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 143, + 245, + 192, + 258 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 245, + 192, + 258 + ], + "spans": [ + { + "bbox": [ + 143, + 245, + 192, + 258 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 270, + 289, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 270, + 289, + 582 + ], + "spans": [ + { + "bbox": [ + 47, + 270, + 289, + 582 + ], + "type": "text", + "content": "An efficient and effective decoding mechanism is crucial in medical image segmentation, especially in scenarios with limited computational resources. However, these decoding mechanisms usually come with high computational costs. To address this concern, we introduce EMCAD, a new efficient multi-scale convolutional attention decoder, designed to optimize both performance and computational efficiency. EMCAD leverages a unique multi-scale depth-wise convolution block, significantly enhancing feature maps through multi-scale convolutions. EMCAD also employs channel, spatial, and grouped (large-kernel) gated attention mechanisms, which are highly effective at capturing intricate spatial relationships while focusing on salient regions. By employing group and depth-wise convolution, EMCAD is very efficient and scales well (e.g., only 1.91M parameters and 0.381G FLOPs are needed when using a standard encoder). Our rigorous evaluations across 12 datasets that belong to six medical image segmentation tasks reveal that EMCAD achieves state-of-the-art (SOTA) performance with " + }, + { + "bbox": [ + 47, + 270, + 289, + 582 + ], + "type": "inline_equation", + "content": "79.4\\%" + }, + { + "bbox": [ + 47, + 270, + 289, + 582 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 270, + 289, + 582 + ], + "type": "inline_equation", + "content": "80.3\\%" + }, + { + "bbox": [ + 47, + 270, + 289, + 582 + ], + "type": "text", + "content": " reduction in #Params and #FLOPs, respectively. Moreover, EMCAD's adaptability to different encoders and versatility across segmentation tasks further establish EMCAD as a promising tool, advancing the field towards more efficient and accurate medical image analysis. Our implementation is available at https://github.com/SLDGroup/EMCAD." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 597, + 128, + 609 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 597, + 128, + 609 + ], + "spans": [ + { + "bbox": [ + 47, + 597, + 128, + 609 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 617, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 617, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 617, + 288, + 715 + ], + "type": "text", + "content": "In the realm of medical diagnostics and therapeutic strategies, automated segmentation of medical images is vital, as it classifies pixels to identify critical regions such as lesions, tumors, or entire organs. A variety of U-shaped convolutional neural network (CNN) architectures [20, 24, 37, 41, 44, 62], notably UNet [44], UNet++ [62], UNet3+ [24], and nnU-Net [19], have become standard techniques for this purpose, achieving high-quality, high-resolution segmen" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 246, + 547, + 329 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 246, + 547, + 329 + ], + "spans": [ + { + "bbox": [ + 304, + 246, + 547, + 329 + ], + "type": "text", + "content": "tation output. Attention mechanisms [12, 17, 20, 41, 57] have also been integrated into these models to enhance feature maps and improve pixel-level classification. Although attention-based models have shown improved performance, they still face significant challenges due to the computationally expensive convolutional blocks that are typically used in conjunction with attention mechanisms." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 330, + 547, + 520 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 330, + 547, + 520 + ], + "spans": [ + { + "bbox": [ + 304, + 330, + 547, + 520 + ], + "type": "text", + "content": "Recently, vision transformers [18] have shown promise in medical image segmentation tasks [5, 8, 17, 42, 43, 52, 54, 61] by capturing long-range dependencies among pixels through Self-attention (SA) mechanisms. Hierarchical vision transformers like Swin [34], PVT [55, 56], MaxViT [49], MERIT [43], ConvFormer [33], and MetaFormer [59] have been introduced to further improve the performance in this field. While the SA excels at capturing global information, it is less adept at understanding the local spatial context [13, 28]. To address this limitation, some approaches have integrated local convolutional attention within the decoders to better grasp spatial details. Nevertheless, these methods can still be computationally demanding because they frequently employ costly convolutional blocks. This limits their applicability to real-world scenarios where computational resources are restricted." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 521, + 547, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 521, + 547, + 616 + ], + "spans": [ + { + "bbox": [ + 304, + 521, + 547, + 616 + ], + "type": "text", + "content": "To address the aforementioned limitations, we introduce EMCAD, an efficient multi-scale convolutional attention decoding using a new multi-scale depth-wise convolution block. More precisely, EMCAD enhances the feature maps via efficient multi-scale convolutions, while incorporating complex spatial relationships and local attention through the use of channel, spatial, and grouped (large-kernel) gated attention mechanisms. Our contributions are as follows:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 617, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 617, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 306, + 617, + 547, + 714 + ], + "type": "text", + "content": "- New Efficient Multi-scale Convolutional Decoder: We introduce an efficient multi-scale cascaded fullyconvolutional attention decoder (EMCAD) for 2D medical image segmentation; this takes the multi-stage features of vision encoders and progressively enhances the multi-scale and multi-resolution spatial representations. EMCAD has only 0.506M parameters and 0.11G FLOPs for a tiny encoder with #channels = [32, 64, 160, 256]," + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "11769" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 67, + 71, + 270, + 217 + ], + "blocks": [ + { + "bbox": [ + 67, + 71, + 270, + 217 + ], + "lines": [ + { + "bbox": [ + 67, + 71, + 270, + 217 + ], + "spans": [ + { + "bbox": [ + 67, + 71, + 270, + 217 + ], + "type": "image", + "image_path": "3af94799489a0efcba7e12ed56d3a523e2bdb917ee5fa27705063e9a6672a011.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 220, + 288, + 264 + ], + "lines": [ + { + "bbox": [ + 46, + 220, + 288, + 264 + ], + "spans": [ + { + "bbox": [ + 46, + 220, + 288, + 264 + ], + "type": "text", + "content": "Figure 1. Average DICE scores vs. #FLOPs for different methods over 10 binary medical image segmentation datasets. As shown, our approaches (PVT-EMCAD-B0 and PVT-EMCAD-B2) have the lowest #FLOPs, yet the highest DICE scores." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 275, + 287, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 275, + 287, + 300 + ], + "spans": [ + { + "bbox": [ + 55, + 275, + 287, + 300 + ], + "type": "text", + "content": "while it has 1.91M parameters and 0.381G FLOPs for a standard encoder with #channels = [64, 128, 320, 512]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 300, + 287, + 563 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 47, + 300, + 287, + 395 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 300, + 287, + 395 + ], + "spans": [ + { + "bbox": [ + 47, + 300, + 287, + 395 + ], + "type": "text", + "content": "- Efficient Multi-scale Convolutional Attention Module: We introduce MSCAM, a new efficient multi-scale convolutional attention module that performs depth-wise convolutions at multiple scales; this refines the feature maps produced by vision encoders and enables capturing multi-scale salient features by suppressing irrelevant regions. The use of depth-wise convolutions makes MSCAM very efficient." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 396, + 287, + 467 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 396, + 287, + 467 + ], + "spans": [ + { + "bbox": [ + 47, + 396, + 287, + 467 + ], + "type": "text", + "content": "- Large-kernel Grouped Attention Gate: We introduce a new grouped attention gate to fuse refined features with the features from skip connections. By using larger kernel " + }, + { + "bbox": [ + 47, + 396, + 287, + 467 + ], + "type": "inline_equation", + "content": "(3 \\times 3)" + }, + { + "bbox": [ + 47, + 396, + 287, + 467 + ], + "type": "text", + "content": " group convolutions instead of point-wise convolutions in the design, we capture salient features in a larger local context with less computation." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 467, + 287, + 563 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 467, + 287, + 563 + ], + "spans": [ + { + "bbox": [ + 47, + 467, + 287, + 563 + ], + "type": "text", + "content": "- Improved Performance: We empirically show that EM-CAD can be used with any hierarchical vision encoder (e.g., PVTv2-B0, PVTv2-B2 [56]), while significantly improving the performance of 2D medical image segmentation. EMCAD produces better results than SOTA methods with a significantly lower computational cost (as shown in Figure 1) on 12 medical image segmentation benchmarks that belong to six different tasks." + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 46, + 563, + 287, + 636 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 563, + 287, + 636 + ], + "spans": [ + { + "bbox": [ + 46, + 563, + 287, + 636 + ], + "type": "text", + "content": "The remaining of this paper is organized as follows: Section 2 summarizes related work. Section 3 describes the proposed method. Section 4 explains our experimental setup and results on 12 medical image segmentation benchmarks. Section 5 covers different ablation experiments. Lastly, Section 6 concludes the paper." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 649, + 134, + 661 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 649, + 134, + 661 + ], + "spans": [ + { + "bbox": [ + 47, + 649, + 134, + 661 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 670, + 144, + 681 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 670, + 144, + 681 + ], + "spans": [ + { + "bbox": [ + 47, + 670, + 144, + 681 + ], + "type": "text", + "content": "2.1. Vision encoders" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 689, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 689, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 689, + 287, + 714 + ], + "type": "text", + "content": "Convolutional Neural Networks (CNNs) [21-23, 32, 35, 45-48] have been foundational as encoders due to their pro" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 72, + 545, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 251 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 251 + ], + "type": "text", + "content": "ficiency in handling spatial relationships in images. More precisely, AlexNet [32] and VGG [46] pave the way, leveraging deep layers of convolutions to extract features progressively. GoogleNet [47] introduces the inception module, allowing more efficient computation of representations across various scales. ResNet [21] introduces residual connections, enabling the training of networks with substantially more layers by addressing the vanishing gradients problem. MobileNets [22, 45] bring CNNs to mobile devices through lightweight, depth-wise separable convolutions. EfficientNet [48] introduces a scalable architectural design to CNNs with compound scaling. Although CNNs are pivotal for many vision applications, they generally lack the ability to capture long-range dependencies within images due to their inherent local receptive fields." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 253, + 546, + 492 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 253, + 546, + 492 + ], + "spans": [ + { + "bbox": [ + 304, + 253, + 546, + 492 + ], + "type": "text", + "content": "Recently, Vision Transformers (ViTs), pioneered by Dosovitskiy et al. [18], enabled the learning of long-range relationships among pixels using Self-attention (SA). Since then, ViTs have been enhanced by integrating CNN features [49, 56], developing novel self-attention (SA) blocks [34, 49], and introducing new architectural designs [55, 58]. The Swin Transformer [34] incorporates a sliding window attention mechanism, while SegFormer [58] leverages MixFFN blocks for hierarchical structures. PVT [55] uses spatial reduction attention, refined in PVTv2 [56] with overlapping patch embedding and a linear complexity attention layer. MaxViT [49] introduces a multi-axis self-attention to form a hierarchical CNN-transformer encoder. Although ViTs address the CNNs limitation in capturing long-range pixel dependencies [21-23, 32, 35, 45-48], they face challenges in capturing the local spatial relationships among pixels. In this paper, we aim to overcome these limitations by introducing a new multi-scale cascaded attention decoder that refines feature maps and incorporates local attention using a multi-scale convolutional attention module." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 502, + 463, + 514 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 502, + 463, + 514 + ], + "spans": [ + { + "bbox": [ + 306, + 502, + 463, + 514 + ], + "type": "text", + "content": "2.2. Medical image segmentation" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 522, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 522, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 522, + 545, + 713 + ], + "type": "text", + "content": "Medical image segmentation involves pixel-wise classification to identify various anatomical structures like lesions, tumors, or organs within different imaging modalities such as endoscopy, MRI, or CT scans [8]. U-shaped networks [7, 19, 24, 26, 37, 41, 44, 62] are particularly favored due to their simple but effective encoder-decoder design. The UNet [44] pioneered this approach with its use of skip connections to fuse features at different resolution stages. UNet++ [62] evolves this design by incorporating nested encoder-decoder pathways with dense skip connections. Expanding on these ideas, UNet " + }, + { + "bbox": [ + 304, + 522, + 545, + 713 + ], + "type": "inline_equation", + "content": "3+" + }, + { + "bbox": [ + 304, + 522, + 545, + 713 + ], + "type": "text", + "content": " [24] introduces comprehensive skip pathways that facilitate full-scale feature integration. Further advancement comes with DC-UNet [37], which integrates a multi-resolution convolution scheme and residual paths into its skip connections. The DeepLab series, including DeepLabv3 [10] and DeepLabv3+ [11], in" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "11770" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 167 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 167 + ], + "type": "text", + "content": "produce atrous convolutions and spatial pyramid pooling to handle multi-scale information. SegNet [2] uses pooling indices to upsample feature maps, preserving the boundary details. nnU-Net [19] automatically configures hyperparameters based on the specific dataset characteristics, using standard 2D and 3D UNets. Collectively, these U-shaped models have become a benchmark for success in the domain of medical image segmentation." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 168, + 288, + 324 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 168, + 288, + 324 + ], + "spans": [ + { + "bbox": [ + 46, + 168, + 288, + 324 + ], + "type": "text", + "content": "Recently, vision transformers have emerged as a formidable force in medical image segmentation, harnessing the ability to capture pixel relationships at global scales [5, 8, 17, 42, 43, 52, 58, 61]. TransUNet [8] presents a novel blend of CNNs for local feature extraction and transformers for global context, enhancing both local and global feature capture. Swin-Unet [5] extends this by incorporating Swin Transformer blocks [34] into a U-shaped model for both encoding and decoding processes. Building on these concepts, MERIT [43] introduces a multi-scale hierarchical transformer, which employs SA across different window sizes, thus enhancing the model capacity to capture multiscale features critical for medical image segmentation." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 325, + 288, + 539 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 325, + 288, + 539 + ], + "spans": [ + { + "bbox": [ + 46, + 325, + 288, + 539 + ], + "type": "text", + "content": "The integration of attention mechanisms has been investigated within CNNs [20, 41] and transformer-based systems [17] for enhancing medical image segmentation. PraNet [20] employs a reverse attention strategy for feature refinement. PolypPVT [17] leverages PVTv2 [56] as its backbone encoder and incorporates CBAM [57] within its decoding stages. The CASCADE [42] presents a novel cascaded decoder, combining channel [23] and spatial [9] attention to refine features at multiple stages, extracted from a transformer encoder, culminating in high-resolution segmentation outputs. While CASCADE achieves notable performance in segmenting medical images by integrating local and global insights from transformers, it is computationally inefficient due to the use of triple " + }, + { + "bbox": [ + 46, + 325, + 288, + 539 + ], + "type": "inline_equation", + "content": "3 \\times 3" + }, + { + "bbox": [ + 46, + 325, + 288, + 539 + ], + "type": "text", + "content": " convolution layers at each decoder stage. In addition to this, it uses single-scale convolutions during decoding. Our new proposal involves the adoption of multi-scale depth-wise convolutions to mitigate these constraints." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 547, + 130, + 561 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 547, + 130, + 561 + ], + "spans": [ + { + "bbox": [ + 47, + 547, + 130, + 561 + ], + "type": "text", + "content": "3. Methodology" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 567, + 287, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 567, + 287, + 616 + ], + "spans": [ + { + "bbox": [ + 46, + 567, + 287, + 616 + ], + "type": "text", + "content": "In this section, we first introduce our new EMCAD decoder and then explain two transformer-based architectures (i.e., PVT-EMCAD-B0 and PVT-EMCAD-B2) incorporating our proposed decoder." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 622, + 287, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 622, + 287, + 647 + ], + "spans": [ + { + "bbox": [ + 47, + 622, + 287, + 647 + ], + "type": "text", + "content": "3.1. Efficient multi-scale convolutional attention decoding (EMCAD)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 654, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 654, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 654, + 287, + 713 + ], + "type": "text", + "content": "In this section, we introduce our efficient multi-scale convolutional decoding (EMCAD) to process the multi-stage features extracted from pretrained hierarchical vision encoders for high-resolution semantic segmentation. As shown in Figure 2(b), EMCAD consists of efficient multi" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "text", + "content": "scale convolutional attention modules (MSCAMs) to robustly enhance the feature maps, large-kernel grouped attention gates (LGAGs) to refine feature maps fusing with the skip connection via gated attention mechanism, efficient up-convolution blocks (EUCBs) for up-sampling followed by enhancement of feature maps, and segmentation heads (SHs) to produce the segmentation outputs." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 158, + 546, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 158, + 546, + 266 + ], + "spans": [ + { + "bbox": [ + 304, + 158, + 546, + 266 + ], + "type": "text", + "content": "More specifically, we use four MSCAMs to refine pyramid features (i.e., " + }, + { + "bbox": [ + 304, + 158, + 546, + 266 + ], + "type": "inline_equation", + "content": "X1" + }, + { + "bbox": [ + 304, + 158, + 546, + 266 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 158, + 546, + 266 + ], + "type": "inline_equation", + "content": "X2" + }, + { + "bbox": [ + 304, + 158, + 546, + 266 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 158, + 546, + 266 + ], + "type": "inline_equation", + "content": "X3" + }, + { + "bbox": [ + 304, + 158, + 546, + 266 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 158, + 546, + 266 + ], + "type": "inline_equation", + "content": "X4" + }, + { + "bbox": [ + 304, + 158, + 546, + 266 + ], + "type": "text", + "content": " in Figure 2) extracted from the four stages of the encoder. After each MSCAM, we use an SH to produce a segmentation map of that stage. Subsequently, we upscale the refined feature maps using EUCBs and add them to the outputs from the corresponding LGAGs. Finally, we add four different segmentation maps to produce the final segmentation output. Different modules of our decoder are described next." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 286, + 531, + 300 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 286, + 531, + 300 + ], + "spans": [ + { + "bbox": [ + 305, + 286, + 531, + 300 + ], + "type": "text", + "content": "3.1.1 Large-kernel grouped attention gate (LGAG)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 310, + 546, + 609 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 310, + 546, + 609 + ], + "spans": [ + { + "bbox": [ + 304, + 310, + 546, + 609 + ], + "type": "text", + "content": "We introduce a new large-kernel grouped attention gate (LGAG) to progressively combine feature maps with attention coefficients, which are learned by the network to allow higher activation of relevant features and suppression of irrelevant ones. This process employs a gating signal derived from higher-level features to control the flow of information across different stages of the network, thus enhancing its precision for medical image segmentation. Unlike Attention UNet [41] which uses " + }, + { + "bbox": [ + 304, + 310, + 546, + 609 + ], + "type": "inline_equation", + "content": "1 \\times 1" + }, + { + "bbox": [ + 304, + 310, + 546, + 609 + ], + "type": "text", + "content": " convolution to process gating signal " + }, + { + "bbox": [ + 304, + 310, + 546, + 609 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 304, + 310, + 546, + 609 + ], + "type": "text", + "content": " (features from skip connections) and input feature map " + }, + { + "bbox": [ + 304, + 310, + 546, + 609 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 304, + 310, + 546, + 609 + ], + "type": "text", + "content": " (upsampled features), in our " + }, + { + "bbox": [ + 304, + 310, + 546, + 609 + ], + "type": "inline_equation", + "content": "q_{att}(.)" + }, + { + "bbox": [ + 304, + 310, + 546, + 609 + ], + "type": "text", + "content": " function, we process " + }, + { + "bbox": [ + 304, + 310, + 546, + 609 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 304, + 310, + 546, + 609 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 310, + 546, + 609 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 304, + 310, + 546, + 609 + ], + "type": "text", + "content": " by applying separate " + }, + { + "bbox": [ + 304, + 310, + 546, + 609 + ], + "type": "inline_equation", + "content": "3 \\times 3" + }, + { + "bbox": [ + 304, + 310, + 546, + 609 + ], + "type": "text", + "content": " group convolutions " + }, + { + "bbox": [ + 304, + 310, + 546, + 609 + ], + "type": "inline_equation", + "content": "GC_{g}(.)" + }, + { + "bbox": [ + 304, + 310, + 546, + 609 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 310, + 546, + 609 + ], + "type": "inline_equation", + "content": "GC_{x}(.)" + }, + { + "bbox": [ + 304, + 310, + 546, + 609 + ], + "type": "text", + "content": ", respectively. These convolved features are then normalized using batch normalization " + }, + { + "bbox": [ + 304, + 310, + 546, + 609 + ], + "type": "inline_equation", + "content": "(BN(.))" + }, + { + "bbox": [ + 304, + 310, + 546, + 609 + ], + "type": "text", + "content": " [27] and merged through element-wise addition. The resultant feature map is activated through a ReLU " + }, + { + "bbox": [ + 304, + 310, + 546, + 609 + ], + "type": "inline_equation", + "content": "(R(.))" + }, + { + "bbox": [ + 304, + 310, + 546, + 609 + ], + "type": "text", + "content": " layer [39]. Afterward, we apply a " + }, + { + "bbox": [ + 304, + 310, + 546, + 609 + ], + "type": "inline_equation", + "content": "1 \\times 1" + }, + { + "bbox": [ + 304, + 310, + 546, + 609 + ], + "type": "text", + "content": " convolution " + }, + { + "bbox": [ + 304, + 310, + 546, + 609 + ], + "type": "inline_equation", + "content": "(C(.))" + }, + { + "bbox": [ + 304, + 310, + 546, + 609 + ], + "type": "text", + "content": " followed by " + }, + { + "bbox": [ + 304, + 310, + 546, + 609 + ], + "type": "inline_equation", + "content": "BN(.)" + }, + { + "bbox": [ + 304, + 310, + 546, + 609 + ], + "type": "text", + "content": " layer to get a single channel feature map. We then pass the resultant single-channel feature map through a Sigmoid " + }, + { + "bbox": [ + 304, + 310, + 546, + 609 + ], + "type": "inline_equation", + "content": "(\\sigma(.))" + }, + { + "bbox": [ + 304, + 310, + 546, + 609 + ], + "type": "text", + "content": " activation function to yield the attention coefficients. The output of this transformation is used to scale the input feature " + }, + { + "bbox": [ + 304, + 310, + 546, + 609 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 304, + 310, + 546, + 609 + ], + "type": "text", + "content": " through elementwise multiplication, producing the attention-gated feature " + }, + { + "bbox": [ + 304, + 310, + 546, + 609 + ], + "type": "inline_equation", + "content": "LGAG(g, x)" + }, + { + "bbox": [ + 304, + 310, + 546, + 609 + ], + "type": "text", + "content": ". The " + }, + { + "bbox": [ + 304, + 310, + 546, + 609 + ], + "type": "inline_equation", + "content": "LGAG(.)" + }, + { + "bbox": [ + 304, + 310, + 546, + 609 + ], + "type": "text", + "content": " (Figure 2(g)) can be formulated as in Equations 1 and 2:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 318, + 620, + 545, + 633 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 620, + 545, + 633 + ], + "spans": [ + { + "bbox": [ + 318, + 620, + 545, + 633 + ], + "type": "interline_equation", + "content": "q _ {a t t} (g, x) = R \\left(B N \\left(G C _ {g} (g) + B N \\left(G C _ {x} (x)\\right)\\right)\\right) \\tag {1}", + "image_path": "86256c366d113ce8774271977487167f30be9956249cbb3d3e3b9d3a5972ab35.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 336, + 659, + 545, + 672 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 336, + 659, + 545, + 672 + ], + "spans": [ + { + "bbox": [ + 336, + 659, + 545, + 672 + ], + "type": "interline_equation", + "content": "L G A G (g, x) = x * \\sigma \\left(B N \\left(C \\left(q _ {a t t} (g, x)\\right)\\right)\\right) \\tag {2}", + "image_path": "dacafda2b2933ba3c947ac896469ae896d268f040a2c50b5f0f4fb461e710be8.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 681, + 546, + 717 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 681, + 546, + 717 + ], + "spans": [ + { + "bbox": [ + 304, + 681, + 546, + 717 + ], + "type": "text", + "content": "Due to using " + }, + { + "bbox": [ + 304, + 681, + 546, + 717 + ], + "type": "inline_equation", + "content": "3 \\times 3" + }, + { + "bbox": [ + 304, + 681, + 546, + 717 + ], + "type": "text", + "content": " kernel group convolutions in " + }, + { + "bbox": [ + 304, + 681, + 546, + 717 + ], + "type": "inline_equation", + "content": "q_{att}(.)" + }, + { + "bbox": [ + 304, + 681, + 546, + 717 + ], + "type": "text", + "content": ", our LGAG captures comparatively larger spatial contexts with less computational cost." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "11771" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 72, + 69, + 523, + 281 + ], + "blocks": [ + { + "bbox": [ + 72, + 69, + 523, + 281 + ], + "lines": [ + { + "bbox": [ + 72, + 69, + 523, + 281 + ], + "spans": [ + { + "bbox": [ + 72, + 69, + 523, + 281 + ], + "type": "image", + "image_path": "680916196f86f563262f187b1e9498563b3b45b7b797b45690ae9a0e76ac7c8c.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 283, + 547, + 340 + ], + "lines": [ + { + "bbox": [ + 46, + 283, + 547, + 340 + ], + "spans": [ + { + "bbox": [ + 46, + 283, + 547, + 340 + ], + "type": "text", + "content": "Figure 2. Hierarchical encoder with newly proposed EMCAD decoder architecture. (a) CNN or transformer encoder with four hierarchical stages, (b) EMCAD decoder, (c) Efficient up-convolution block (EUCB), (d) Multi-scale convolutional attention module (MSCAM), (e) Multi-scale convolution block (MSCB), (f) Multi-scale (parallel) depth-wise convolution (MSDC), (g) Large-kernel grouped attention gate (LGAG), (h) Channel attention block (CAB), and (i) Spatial attention block (SAB). X1, X2, X3, and X4 are the features from the four stages of the hierarchical encoder. p1, p2, p3, and p4 are output segmentation maps from four stages of our decoder." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 346, + 288, + 369 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 346, + 288, + 369 + ], + "spans": [ + { + "bbox": [ + 47, + 346, + 288, + 369 + ], + "type": "text", + "content": "3.1.2 Multi-scale convolutional attention module (MSCAM)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 376, + 287, + 473 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 376, + 287, + 473 + ], + "spans": [ + { + "bbox": [ + 46, + 376, + 287, + 473 + ], + "type": "text", + "content": "We introduce an efficient multi-scale convolutional attention module to refine the feature maps. MSCAM consists of a channel attention block " + }, + { + "bbox": [ + 46, + 376, + 287, + 473 + ], + "type": "inline_equation", + "content": "(CAB(\\cdot))" + }, + { + "bbox": [ + 46, + 376, + 287, + 473 + ], + "type": "text", + "content": " to put emphasis on pertinent channels, a spatial attention block [9] " + }, + { + "bbox": [ + 46, + 376, + 287, + 473 + ], + "type": "inline_equation", + "content": "(SAB(\\cdot))" + }, + { + "bbox": [ + 46, + 376, + 287, + 473 + ], + "type": "text", + "content": " to capture the local contextual information, and an efficient multi-scale convolution block " + }, + { + "bbox": [ + 46, + 376, + 287, + 473 + ], + "type": "inline_equation", + "content": "(MSCB(\\cdot))" + }, + { + "bbox": [ + 46, + 376, + 287, + 473 + ], + "type": "text", + "content": " to enhance the feature maps preserving contextual relationships. The " + }, + { + "bbox": [ + 46, + 376, + 287, + 473 + ], + "type": "inline_equation", + "content": "MSCAM(\\cdot)" + }, + { + "bbox": [ + 46, + 376, + 287, + 473 + ], + "type": "text", + "content": " (Figure 2(d)) is given in Equation 3:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 76, + 479, + 287, + 493 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 479, + 287, + 493 + ], + "spans": [ + { + "bbox": [ + 76, + 479, + 287, + 493 + ], + "type": "interline_equation", + "content": "M S C A M (x) = M S C B (S A B (C A B (x))) \\tag {3}", + "image_path": "645dca5013c8826351841395e4889bb72779627bd79bd391764bb0f2113e32dc.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 498, + 287, + 546 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 498, + 287, + 546 + ], + "spans": [ + { + "bbox": [ + 46, + 498, + 287, + 546 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 498, + 287, + 546 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 46, + 498, + 287, + 546 + ], + "type": "text", + "content": " is the input tensor. Due to using depth-wise convolution in multiple scales, our MSCAM is more effective with significantly lower computational cost than the convolutional attention module (CAM) proposed in [42]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 546, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 546, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 546, + 288, + 713 + ], + "type": "text", + "content": "Multi-scale Convolution Block (MSCB): We introduce an efficient multi-scale convolution block to enhance the features generated by our cascaded expanding path. In our MSCB, we follow the design of the inverted residual block (IRB) of MobileNetV2 [45]. However, unlike IRB, our MSCB performs depth-wise convolution at multiple scales and uses channel_shuffle [60] to shuffle channels across groups. More specifically, in our MSCB, we first expand the number of channels (i.e., expansion_factor = 2) using a point-wise " + }, + { + "bbox": [ + 46, + 546, + 288, + 713 + ], + "type": "inline_equation", + "content": "(1 \\times 1)" + }, + { + "bbox": [ + 46, + 546, + 288, + 713 + ], + "type": "text", + "content": " convolution layers " + }, + { + "bbox": [ + 46, + 546, + 288, + 713 + ], + "type": "inline_equation", + "content": "PWC_{1}(\\cdot)" + }, + { + "bbox": [ + 46, + 546, + 288, + 713 + ], + "type": "text", + "content": " followed by a batch normalization layer " + }, + { + "bbox": [ + 46, + 546, + 288, + 713 + ], + "type": "inline_equation", + "content": "BN(\\cdot)" + }, + { + "bbox": [ + 46, + 546, + 288, + 713 + ], + "type": "text", + "content": " and a ReLU6 [31] activation layer " + }, + { + "bbox": [ + 46, + 546, + 288, + 713 + ], + "type": "inline_equation", + "content": "R6(\\cdot)" + }, + { + "bbox": [ + 46, + 546, + 288, + 713 + ], + "type": "text", + "content": ". We then use a multi-scale depth-wise convolution " + }, + { + "bbox": [ + 46, + 546, + 288, + 713 + ], + "type": "inline_equation", + "content": "MSDC(\\cdot)" + }, + { + "bbox": [ + 46, + 546, + 288, + 713 + ], + "type": "text", + "content": " to capture both multi-scale and multi-resolution contexts. As depth-wise convolution overlooks" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 346, + 545, + 419 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 346, + 545, + 419 + ], + "spans": [ + { + "bbox": [ + 305, + 346, + 545, + 419 + ], + "type": "text", + "content": "the relationships among channels, we use a channel_shuffle operation to incorporate relationships among channels. Afterward, we use another point-wise convolution " + }, + { + "bbox": [ + 305, + 346, + 545, + 419 + ], + "type": "inline_equation", + "content": "PWC_{2}(.)" + }, + { + "bbox": [ + 305, + 346, + 545, + 419 + ], + "type": "text", + "content": " followed by a " + }, + { + "bbox": [ + 305, + 346, + 545, + 419 + ], + "type": "inline_equation", + "content": "BN(.)" + }, + { + "bbox": [ + 305, + 346, + 545, + 419 + ], + "type": "text", + "content": " to transform back the original #channels, which also encodes dependency among channels. The " + }, + { + "bbox": [ + 305, + 346, + 545, + 419 + ], + "type": "inline_equation", + "content": "MSCB(\\cdot)" + }, + { + "bbox": [ + 305, + 346, + 545, + 419 + ], + "type": "text", + "content": " (Figure 2(e)) is formulated as in Equation 4:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 426, + 545, + 437 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 426, + 545, + 437 + ], + "spans": [ + { + "bbox": [ + 313, + 426, + 545, + 437 + ], + "type": "interline_equation", + "content": "M S C B (x) = B N \\left(P W C _ {2} \\left(C S \\left(M S D C \\left(R 6 \\left(B N \\left(P W C _ {1} (x)\\right)\\right)\\right)\\right)\\right)\\right) \\tag {4}", + "image_path": "7946ce2bb380363371427dca03ccdbd62214f4f1af499ca576ded53ac81248c7.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 444, + 545, + 469 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 444, + 545, + 469 + ], + "spans": [ + { + "bbox": [ + 305, + 444, + 545, + 469 + ], + "type": "text", + "content": "where parallel " + }, + { + "bbox": [ + 305, + 444, + 545, + 469 + ], + "type": "inline_equation", + "content": "MSDC(.)" + }, + { + "bbox": [ + 305, + 444, + 545, + 469 + ], + "type": "text", + "content": " (Figure 2(f)) for different kernel sizes " + }, + { + "bbox": [ + 305, + 444, + 545, + 469 + ], + "type": "inline_equation", + "content": "(KS)" + }, + { + "bbox": [ + 305, + 444, + 545, + 469 + ], + "type": "text", + "content": " can be formulated using Equation 5:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 350, + 475, + 545, + 489 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 350, + 475, + 545, + 489 + ], + "spans": [ + { + "bbox": [ + 350, + 475, + 545, + 489 + ], + "type": "interline_equation", + "content": "M S D C (x) = \\sum_ {k s \\in K S} D W C B _ {k s} (x) \\tag {5}", + "image_path": "62d1c6f6466ab1d5f2d5a9cb85eeffcbc3d7db0c45acf49adfcb6ab433b23587.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 495, + 545, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 495, + 545, + 579 + ], + "spans": [ + { + "bbox": [ + 305, + 495, + 545, + 579 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 495, + 545, + 579 + ], + "type": "inline_equation", + "content": "DWCB_{ks}(x) = R6(BN(DWC_{ks}(x)))" + }, + { + "bbox": [ + 305, + 495, + 545, + 579 + ], + "type": "text", + "content": ". Here, " + }, + { + "bbox": [ + 305, + 495, + 545, + 579 + ], + "type": "inline_equation", + "content": "DWC_{ks}()" + }, + { + "bbox": [ + 305, + 495, + 545, + 579 + ], + "type": "text", + "content": " is a depth-wise convolution with the kernel size " + }, + { + "bbox": [ + 305, + 495, + 545, + 579 + ], + "type": "inline_equation", + "content": "ks" + }, + { + "bbox": [ + 305, + 495, + 545, + 579 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 305, + 495, + 545, + 579 + ], + "type": "inline_equation", + "content": "BN()" + }, + { + "bbox": [ + 305, + 495, + 545, + 579 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 495, + 545, + 579 + ], + "type": "inline_equation", + "content": "R6()" + }, + { + "bbox": [ + 305, + 495, + 545, + 579 + ], + "type": "text", + "content": " are batch normalization and ReLU6 activation, respectively. Additionally, our sequential " + }, + { + "bbox": [ + 305, + 495, + 545, + 579 + ], + "type": "inline_equation", + "content": "MSDC()" + }, + { + "bbox": [ + 305, + 495, + 545, + 579 + ], + "type": "text", + "content": " uses the recursively updated input " + }, + { + "bbox": [ + 305, + 495, + 545, + 579 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 305, + 495, + 545, + 579 + ], + "type": "text", + "content": ", where the input " + }, + { + "bbox": [ + 305, + 495, + 545, + 579 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 305, + 495, + 545, + 579 + ], + "type": "text", + "content": " is residually connected to the previous " + }, + { + "bbox": [ + 305, + 495, + 545, + 579 + ], + "type": "inline_equation", + "content": "DWCB_{ks}()" + }, + { + "bbox": [ + 305, + 495, + 545, + 579 + ], + "type": "text", + "content": " for better regularization as in Equation 6:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 376, + 586, + 545, + 599 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 376, + 586, + 545, + 599 + ], + "spans": [ + { + "bbox": [ + 376, + 586, + 545, + 599 + ], + "type": "interline_equation", + "content": "x = x + D W C B _ {k s} (x) \\tag {6}", + "image_path": "db28e23b5b2b8f6fd1cb4d5d1b154d61e31dce0046e281d92c8a056fac915aea.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 605, + 545, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 545, + 715 + ], + "type": "text", + "content": "Channel Attention Block (CAB): We use channel attention block to assign different levels of importance to each channel, thus emphasizing more relevant features while suppressing less useful ones. Basically, the CAB identifies which feature maps to focus on (and then refine them). Following [57], in CAB, we first apply the adaptive maximum pooling " + }, + { + "bbox": [ + 304, + 605, + 545, + 715 + ], + "type": "inline_equation", + "content": "(P_{m}(\\cdot))" + }, + { + "bbox": [ + 304, + 605, + 545, + 715 + ], + "type": "text", + "content": " and adaptive average pooling " + }, + { + "bbox": [ + 304, + 605, + 545, + 715 + ], + "type": "inline_equation", + "content": "(P_{a}(\\cdot))" + }, + { + "bbox": [ + 304, + 605, + 545, + 715 + ], + "type": "text", + "content": " to the spatial dimensions (i.e., height and width) to extract the most significant feature of the entire feature map per" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "11772" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "type": "text", + "content": "channel. Then, for each pooled feature map, we reduce the number of channels " + }, + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "type": "inline_equation", + "content": "r = 1 / 16" + }, + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "type": "text", + "content": " times separately using a point-wise convolution " + }, + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "type": "inline_equation", + "content": "(C_1(\\cdot))" + }, + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "type": "text", + "content": " followed by a ReLU activation " + }, + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "type": "inline_equation", + "content": "(R)" + }, + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "type": "text", + "content": ". Afterward, we recover the original channels using another point-wise convolution " + }, + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "type": "inline_equation", + "content": "(C_2(\\cdot))" + }, + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "type": "text", + "content": ". We then add both recovered feature maps and apply Sigmoid " + }, + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "type": "inline_equation", + "content": "(\\sigma)" + }, + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "type": "text", + "content": " activation to estimate attention weights. Finally, we incorporate these weights to input " + }, + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "type": "text", + "content": " using the Hadamard product " + }, + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "type": "inline_equation", + "content": "(\\otimes)" + }, + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "type": "text", + "content": ". The " + }, + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "type": "inline_equation", + "content": "CAB(\\cdot)" + }, + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "type": "text", + "content": " (Figure 2(h)) is defined using Equation 7:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 58, + 188, + 287, + 200 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 188, + 287, + 200 + ], + "spans": [ + { + "bbox": [ + 58, + 188, + 287, + 200 + ], + "type": "interline_equation", + "content": "C A B (x) = \\sigma \\left(C _ {2} \\left(R \\left(C _ {1} \\left(P _ {m} (x)\\right)\\right)\\right) + C _ {2} \\left(R \\left(C _ {1} \\left(P _ {a} (x)\\right)\\right)\\right)\\right) * x \\tag {7}", + "image_path": "2426ba884c1fe9ecb1481897f3e87c15df7cc7d71f3364c7b3ad774b11b17c24.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 206, + 289, + 393 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 206, + 289, + 393 + ], + "spans": [ + { + "bbox": [ + 46, + 206, + 289, + 393 + ], + "type": "text", + "content": "Spatial Attention Block (SAB): We use spatial attention to mimic the attentional processes of the human brain by focusing on specific parts of an input image. Basically, the SAB determines where to focus in a feature map; then it enhances those features. This process enhances the model's ability to recognize and respond to relevant spatial features, which is crucial for image segmentation where the context and location of objects significantly influence the output. In SAB, we first pool maximum " + }, + { + "bbox": [ + 46, + 206, + 289, + 393 + ], + "type": "inline_equation", + "content": "(Ch_{max}(\\cdot))" + }, + { + "bbox": [ + 46, + 206, + 289, + 393 + ], + "type": "text", + "content": " and average " + }, + { + "bbox": [ + 46, + 206, + 289, + 393 + ], + "type": "inline_equation", + "content": "(Ch_{avg}(\\cdot))" + }, + { + "bbox": [ + 46, + 206, + 289, + 393 + ], + "type": "text", + "content": " values along the channel dimension to pay attention to local features. Then, we use a large kernel (i.e., " + }, + { + "bbox": [ + 46, + 206, + 289, + 393 + ], + "type": "inline_equation", + "content": "7\\times 7" + }, + { + "bbox": [ + 46, + 206, + 289, + 393 + ], + "type": "text", + "content": " as in [17]) convolution layer to enhance local contextual relationships among features. Afterward, we apply the Sigmoid activation " + }, + { + "bbox": [ + 46, + 206, + 289, + 393 + ], + "type": "inline_equation", + "content": "(\\sigma)" + }, + { + "bbox": [ + 46, + 206, + 289, + 393 + ], + "type": "text", + "content": " to calculate attention weights. Finally, we feed these weights to the input " + }, + { + "bbox": [ + 46, + 206, + 289, + 393 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 46, + 206, + 289, + 393 + ], + "type": "text", + "content": " (using Hadamard product " + }, + { + "bbox": [ + 46, + 206, + 289, + 393 + ], + "type": "inline_equation", + "content": "(\\otimes)" + }, + { + "bbox": [ + 46, + 206, + 289, + 393 + ], + "type": "text", + "content": " to attend information in a more targeted way. The " + }, + { + "bbox": [ + 46, + 206, + 289, + 393 + ], + "type": "inline_equation", + "content": "SAB(.)" + }, + { + "bbox": [ + 46, + 206, + 289, + 393 + ], + "type": "text", + "content": " (Figure 2(i)) is defined using Equation 8:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 70, + 401, + 287, + 414 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 401, + 287, + 414 + ], + "spans": [ + { + "bbox": [ + 70, + 401, + 287, + 414 + ], + "type": "interline_equation", + "content": "S A B (x) = \\sigma \\left(L K C \\left(\\left[ C h _ {m a x} (x), C h _ {a v g} (x) \\right]\\right)\\right) * x \\tag {8}", + "image_path": "4c3313c7bd0b804ba6be465740f2cfc4b8d9b1e93b3ab85fd031f3af2e6225f3.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 423, + 246, + 436 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 423, + 246, + 436 + ], + "spans": [ + { + "bbox": [ + 47, + 423, + 246, + 436 + ], + "type": "text", + "content": "3.1.3 Efficient up-convolution block (EUCB)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 443, + 289, + 563 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 443, + 289, + 563 + ], + "spans": [ + { + "bbox": [ + 46, + 443, + 289, + 563 + ], + "type": "text", + "content": "We use an efficient up-convolution block to progressively upsample the feature maps of the current stage to match the dimension and resolution of the feature maps from the next skip connection. The EUCB first uses an UpSampling " + }, + { + "bbox": [ + 46, + 443, + 289, + 563 + ], + "type": "inline_equation", + "content": "Up(\\cdot)" + }, + { + "bbox": [ + 46, + 443, + 289, + 563 + ], + "type": "text", + "content": " with scale-factor 2 to upscale the feature maps. Then, it enhances the upscaled feature maps by applying a " + }, + { + "bbox": [ + 46, + 443, + 289, + 563 + ], + "type": "inline_equation", + "content": "3 \\times 3" + }, + { + "bbox": [ + 46, + 443, + 289, + 563 + ], + "type": "text", + "content": " depth-wise convolution " + }, + { + "bbox": [ + 46, + 443, + 289, + 563 + ], + "type": "inline_equation", + "content": "DWC(\\cdot)" + }, + { + "bbox": [ + 46, + 443, + 289, + 563 + ], + "type": "text", + "content": " followed by a " + }, + { + "bbox": [ + 46, + 443, + 289, + 563 + ], + "type": "inline_equation", + "content": "BN(\\cdot)" + }, + { + "bbox": [ + 46, + 443, + 289, + 563 + ], + "type": "text", + "content": " and a " + }, + { + "bbox": [ + 46, + 443, + 289, + 563 + ], + "type": "inline_equation", + "content": "ReLU(\\cdot)" + }, + { + "bbox": [ + 46, + 443, + 289, + 563 + ], + "type": "text", + "content": " activation. Finally, a " + }, + { + "bbox": [ + 46, + 443, + 289, + 563 + ], + "type": "inline_equation", + "content": "1 \\times 1" + }, + { + "bbox": [ + 46, + 443, + 289, + 563 + ], + "type": "text", + "content": " convolution " + }, + { + "bbox": [ + 46, + 443, + 289, + 563 + ], + "type": "inline_equation", + "content": "C_{1 \\times 1}(\\cdot)" + }, + { + "bbox": [ + 46, + 443, + 289, + 563 + ], + "type": "text", + "content": " is used to reduce the #channels to match with the next stage. The " + }, + { + "bbox": [ + 46, + 443, + 289, + 563 + ], + "type": "inline_equation", + "content": "EUCB(\\cdot)" + }, + { + "bbox": [ + 46, + 443, + 289, + 563 + ], + "type": "text", + "content": " (Figure 2(c)) is formulated as in Equation 9:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 584, + 287, + 597 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 584, + 287, + 597 + ], + "spans": [ + { + "bbox": [ + 55, + 584, + 287, + 597 + ], + "type": "interline_equation", + "content": "E U C B (x) = C _ {1 \\times 1} \\left(R e L U \\left(B N \\left(D W C (U p (x))\\right)\\right)\\right) \\tag {9}", + "image_path": "5c9171e99ce0fb54e57286980d1b94cf82713624388205d5bc634ade3ccaaaf0.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 604, + 287, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 604, + 287, + 628 + ], + "spans": [ + { + "bbox": [ + 47, + 604, + 287, + 628 + ], + "type": "text", + "content": "Due to using depth-wise convolution instead of " + }, + { + "bbox": [ + 47, + 604, + 287, + 628 + ], + "type": "inline_equation", + "content": "3 \\times 3" + }, + { + "bbox": [ + 47, + 604, + 287, + 628 + ], + "type": "text", + "content": " convolution, our EUCB is very efficient." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 634, + 185, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 634, + 185, + 647 + ], + "spans": [ + { + "bbox": [ + 47, + 634, + 185, + 647 + ], + "type": "text", + "content": "3.1.4 Segmentation head (SH)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 654, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 654, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 47, + 654, + 289, + 715 + ], + "type": "text", + "content": "We use segmentation heads to produce the segmentation outputs from the refined feature maps of four stages of the decoder. The SH layer applies a " + }, + { + "bbox": [ + 47, + 654, + 289, + 715 + ], + "type": "inline_equation", + "content": "1 \\times 1" + }, + { + "bbox": [ + 47, + 654, + 289, + 715 + ], + "type": "text", + "content": " convolution " + }, + { + "bbox": [ + 47, + 654, + 289, + 715 + ], + "type": "inline_equation", + "content": "Conv_{1 \\times 1}(\\cdot)" + }, + { + "bbox": [ + 47, + 654, + 289, + 715 + ], + "type": "text", + "content": " to the refined feature maps having " + }, + { + "bbox": [ + 47, + 654, + 289, + 715 + ], + "type": "inline_equation", + "content": "ch_i" + }, + { + "bbox": [ + 47, + 654, + 289, + 715 + ], + "type": "text", + "content": " channels (" + }, + { + "bbox": [ + 47, + 654, + 289, + 715 + ], + "type": "inline_equation", + "content": "ch_i" + }, + { + "bbox": [ + 47, + 654, + 289, + 715 + ], + "type": "text", + "content": " is the #channels in the feature map of stage " + }, + { + "bbox": [ + 47, + 654, + 289, + 715 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 47, + 654, + 289, + 715 + ], + "type": "text", + "content": ") and" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 72, + 546, + 109 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 546, + 109 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 546, + 109 + ], + "type": "text", + "content": "produces output with #channels equal to #classes in target dataset for multi-class but 1 channel for binary segmentation. The " + }, + { + "bbox": [ + 304, + 72, + 546, + 109 + ], + "type": "inline_equation", + "content": "SH(\\cdot)" + }, + { + "bbox": [ + 304, + 72, + 546, + 109 + ], + "type": "text", + "content": " is formulated as in Equation 10:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 376, + 113, + 546, + 126 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 376, + 113, + 546, + 126 + ], + "spans": [ + { + "bbox": [ + 376, + 113, + 546, + 126 + ], + "type": "interline_equation", + "content": "S H (x) = \\operatorname {C o n v} _ {1 \\times 1} (x) \\tag {10}", + "image_path": "444d61214c2a79b30e62bdcfd11d2414c5543529e9bded95c1c17f4007e40a48.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 134, + 424, + 146 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 134, + 424, + 146 + ], + "spans": [ + { + "bbox": [ + 305, + 134, + 424, + 146 + ], + "type": "text", + "content": "3.2. Overall architecture" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 152, + 545, + 223 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 152, + 545, + 223 + ], + "spans": [ + { + "bbox": [ + 304, + 152, + 545, + 223 + ], + "type": "text", + "content": "To show the generalization, effectiveness, and ability to process multi-scale features for medical image segmentation, we integrate our EMCAD decoder alongside tiny (PVTv2-B0) and standard (PVTv2-B2) networks of PVTv2 [56]. However, our decoder is adaptable and seamlessly compatible with other hierarchical backbone networks." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 224, + 546, + 356 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 224, + 546, + 356 + ], + "spans": [ + { + "bbox": [ + 304, + 224, + 546, + 356 + ], + "type": "text", + "content": "PVTv2 differs from conventional transformer patch embedding modules by applying convolutional operations for consistent spatial information capture. Using PVTv2-b0 (Tiny) and PVTv2-b2 (Standard) encoders [56], we develop the PVT-EMCAD-B0 and PVT-EMCAD-B2 architectures. To adopt PVTv2, we first extract the features (X1, X2, X3, and X4) from four layers and feed them (i.e., X4 in the upsample path and X3, X2, X1 in the skip connections) into our EMCAD decoder as shown in Figure 2(a-b). EMCAD then processes them and produces four segmentation maps that correspond to the four stages of the encoder network." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 359, + 519, + 372 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 359, + 519, + 372 + ], + "spans": [ + { + "bbox": [ + 305, + 359, + 519, + 372 + ], + "type": "text", + "content": "3.3. Multi-stage loss and outputs aggregation" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 377, + 545, + 401 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 377, + 545, + 401 + ], + "spans": [ + { + "bbox": [ + 304, + 377, + 545, + 401 + ], + "type": "text", + "content": "Our EMCAD decoder's four segmentation heads produce four prediction maps " + }, + { + "bbox": [ + 304, + 377, + 545, + 401 + ], + "type": "inline_equation", + "content": "p_1, p_2, p_3" + }, + { + "bbox": [ + 304, + 377, + 545, + 401 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 377, + 545, + 401 + ], + "type": "inline_equation", + "content": "p_4" + }, + { + "bbox": [ + 304, + 377, + 545, + 401 + ], + "type": "text", + "content": " across its stages." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 402, + 546, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 402, + 546, + 521 + ], + "spans": [ + { + "bbox": [ + 304, + 402, + 546, + 521 + ], + "type": "text", + "content": "Loss aggregation: We adopt a combinatorial approach to loss combination called MUTATION, inspired by the work of MERIT [43] for multi-class segmentation. This involves calculating the loss for all possible combinations of predictions derived from 4 heads, totaling " + }, + { + "bbox": [ + 304, + 402, + 546, + 521 + ], + "type": "inline_equation", + "content": "2^{4} - 1 = 15" + }, + { + "bbox": [ + 304, + 402, + 546, + 521 + ], + "type": "text", + "content": " unique predictions, and then summing these losses. We focus on minimizing this cumulative combinatorial loss during the training process. For binary segmentation, we optimize the additive loss like [42] with an additional term " + }, + { + "bbox": [ + 304, + 402, + 546, + 521 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{p_1 + p_2 + p_3 + p_4}" + }, + { + "bbox": [ + 304, + 402, + 546, + 521 + ], + "type": "text", + "content": " as in Equation 11:" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 527, + 545, + 539 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 527, + 545, + 539 + ], + "spans": [ + { + "bbox": [ + 317, + 527, + 545, + 539 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {t o t a l}} = \\alpha \\mathcal {L} _ {p _ {1}} + \\beta \\mathcal {L} _ {p _ {2}} + \\gamma \\mathcal {L} _ {p _ {3}} + \\zeta \\mathcal {L} _ {p _ {4}} + \\delta \\mathcal {L} _ {p _ {1} + p _ {2} + p _ {3} + p _ {4}} \\tag {11}", + "image_path": "20f967d71b508202d3a8cf11bc7fa6eaf4ede8b065cd215f9cec98f6c4682032.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 540, + 545, + 576 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 540, + 545, + 576 + ], + "spans": [ + { + "bbox": [ + 304, + 540, + 545, + 576 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 540, + 545, + 576 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{p_1},\\mathcal{L}_{p_2},\\mathcal{L}_{p_3}" + }, + { + "bbox": [ + 304, + 540, + 545, + 576 + ], + "type": "text", + "content": " , and " + }, + { + "bbox": [ + 304, + 540, + 545, + 576 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{p_4}" + }, + { + "bbox": [ + 304, + 540, + 545, + 576 + ], + "type": "text", + "content": " are the losses of each individual prediction maps. " + }, + { + "bbox": [ + 304, + 540, + 545, + 576 + ], + "type": "inline_equation", + "content": "\\alpha = \\beta = \\gamma = \\zeta = \\delta = 1.0" + }, + { + "bbox": [ + 304, + 540, + 545, + 576 + ], + "type": "text", + "content": " are the weights assigned to each loss." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 577, + 545, + 636 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 577, + 545, + 636 + ], + "spans": [ + { + "bbox": [ + 304, + 577, + 545, + 636 + ], + "type": "text", + "content": "Output segmentation maps aggregation: We consider the prediction map, " + }, + { + "bbox": [ + 304, + 577, + 545, + 636 + ], + "type": "inline_equation", + "content": "p_4" + }, + { + "bbox": [ + 304, + 577, + 545, + 636 + ], + "type": "text", + "content": ", from the last stage of our decoder as the final segmentation map. Then, we obtain the final segmentation output by employing a Sigmoid function for binary or a Softmax function for multi-class segmentation." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 305, + 646, + 388, + 659 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 646, + 388, + 659 + ], + "spans": [ + { + "bbox": [ + 305, + 646, + 388, + 659 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 304, + 666, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 666, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 666, + 545, + 713 + ], + "type": "text", + "content": "In this section, we present the details of our implementation followed by a comparative analysis of our PVT-EMCAD-B0 and PVT-EMCAD-B2 against SOTA methods. Datasets and evaluation metrics are in Supplementary Section 7." + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "11773" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 49, + 71, + 545, + 257 + ], + "blocks": [ + { + "bbox": [ + 49, + 71, + 545, + 257 + ], + "lines": [ + { + "bbox": [ + 49, + 71, + 545, + 257 + ], + "spans": [ + { + "bbox": [ + 49, + 71, + 545, + 257 + ], + "type": "table", + "html": "
Methods#Params#FLOPsPolypSkin LesionCellBUSIAvg.
ClinicColonETISKvasirBKAIISIC17ISIC18DSB18EM
UNet [44]34.53M65.53G92.1183.9576.8582.8785.0583.0786.6792.2395.4674.0485.23
UNet++ [62]9.16M34.65G92.1787.8877.4083.3684.0782.9887.4691.9795.4874.7685.75
AttnUNet [41]34.88M66.64G92.2086.4676.8483.4984.0783.6687.0592.2295.5574.4885.60
DeepLabv3+ [10]39.76M14.92G93.2491.9290.7389.0689.7483.8488.6492.1494.9676.8189.11
PraNet [20]32.55M6.93G91.7189.1683.8484.8285.5683.0388.5689.8992.3775.1486.41
CaraNet [38]46.64M11.48G94.0891.1990.2589.7489.7185.0290.1889.1592.7877.3488.94
UACANet-L [30]69.16M31.51G94.1691.0289.7790.1790.3583.7289.7688.8689.2876.9688.41
SSFormer-L [54]66.22M17.28G94.1892.1190.1691.4791.1485.2890.2592.0394.9578.7690.03
PolypPVT [17]25.11M5.30G94.1391.5389.9391.5691.1785.5690.3690.6994.4079.3589.87
TransUNet [8]105.32M38.52G93.9091.6387.7991.0889.1785.0089.1692.0495.2778.3089.33
SwinUNet [5]27.17M6.2G92.4289.2785.1089.5987.6183.9789.2691.0394.4777.3888.01
TransFuse [61]143.74M82.71G93.6290.3586.9190.2487.4784.8989.6290.8594.3579.3688.77
UNeXt [50]1.47M0.57G90.2083.8474.0377.8877.9382.7487.7886.0193.8174.7182.89
PVT-CASCADE [42]34.12M7.62G94.5391.6091.0392.0592.1485.5090.4192.3595.4279.2190.42
PVT-EMCAD-B0 (Ours)3.92M0.84G94.6091.7191.6591.9591.3085.6790.7092.4695.3579.8090.52
PVT-EMCAD-B2 (Ours)26.76M5.6G95.2192.3192.2992.7592.9685.9590.9692.7495.5380.2591.10
", + "image_path": "5013ef6128dea5916bb2dcbb8344229ac8db20e6b47ae03b79b7b457e9a3f673.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 57, + 299, + 537, + 491 + ], + "blocks": [ + { + "bbox": [ + 47, + 259, + 547, + 293 + ], + "lines": [ + { + "bbox": [ + 47, + 259, + 547, + 293 + ], + "spans": [ + { + "bbox": [ + 47, + 259, + 547, + 293 + ], + "type": "text", + "content": "Table 1. Results of binary medical image segmentation (i.e., polyp, skin lesion, cell, and breast cancer). We reproduce the results of SOTA methods using their publicly available implementation with our train-val-test splits of 80:10:10. #FLOPs of all the methods are reported for " + }, + { + "bbox": [ + 47, + 259, + 547, + 293 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 47, + 259, + 547, + 293 + ], + "type": "text", + "content": " inputs, except Swin-UNet (" + }, + { + "bbox": [ + 47, + 259, + 547, + 293 + ], + "type": "inline_equation", + "content": "224 \\times 224" + }, + { + "bbox": [ + 47, + 259, + 547, + 293 + ], + "type": "text", + "content": "). All results are averaged over five runs. Best results are shown in bold." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 57, + 299, + 537, + 491 + ], + "lines": [ + { + "bbox": [ + 57, + 299, + 537, + 491 + ], + "spans": [ + { + "bbox": [ + 57, + 299, + 537, + 491 + ], + "type": "table", + "html": "
ArchitecturesAverageAortaGBKLKRLiverPCSPSM
DICE↑HD95↓mIoU↑
UNet [44]70.1144.6959.3984.0056.7072.4162.6486.9848.7381.4867.96
AttnUNet [41]71.7034.4761.3882.6161.9476.0770.4287.5446.7080.6767.66
R50+UNet [8]74.6836.87-84.1862.8479.1971.2993.3548.2384.4173.92
R50+AttnUNet [8]75.5736.97-55.9263.9179.2072.7193.5649.3787.1974.95
SSFormer [54]78.0125.7267.2382.7863.7480.7278.1193.5361.5387.0776.61
PolypPVT [17]78.0825.6167.4382.3466.1481.2173.7894.3759.3488.0579.4
TransUNet [8]77.6126.967.3286.5660.4380.5478.5394.3358.4787.0675.00
SwinUNet [5]77.5827.3266.8881.7665.9582.3279.2293.7353.8188.0475.79
MT-UNet [53]78.5926.59-87.9264.9981.4777.2993.0659.4687.7576.81
MISSformer [25]81.9618.20-86.9968.6585.2182.0094.4165.6791.9280.81
PVT-CASCADE [42]81.0620.2370.8883.0170.5982.2380.3794.0864.4390.183.69
TransCASCADE [42]82.6817.3473.4886.6368.4887.6684.5694.4365.3390.7983.52
PVT-EMCAD-B0 (Ours)81.9717.3972.6487.2166.6287.4883.9694.5762.0092.6681.22
PVT-EMCAD-B2 (Ours)83.6315.6874.6588.1468.8788.0884.1095.2668.5192.1783.92
", + "image_path": "fdfb3380337452f6615e0035d575276581729e149378c3ca5dcfb7cce8888e93.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 494, + 547, + 528 + ], + "lines": [ + { + "bbox": [ + 47, + 494, + 547, + 528 + ], + "spans": [ + { + "bbox": [ + 47, + 494, + 547, + 528 + ], + "type": "text", + "content": "Table 2. Results of abdomen organ segmentation on Synapse Multi-organ dataset. DICE scores are reported for individual organs. Results of UNet, AttnUNet, PolypPVT, SSFormerPVT, TransUNet, and SwinUNet are taken from [42]. " + }, + { + "bbox": [ + 47, + 494, + 547, + 528 + ], + "type": "inline_equation", + "content": "\\uparrow (\\downarrow)" + }, + { + "bbox": [ + 47, + 494, + 547, + 528 + ], + "type": "text", + "content": " denotes the higher (lower) the better. " + }, + { + "bbox": [ + 47, + 494, + 547, + 528 + ], + "type": "inline_equation", + "content": "^{\\prime} - ^{\\prime}" + }, + { + "bbox": [ + 47, + 494, + 547, + 528 + ], + "type": "text", + "content": " means missing data from the source. EMCAD results are averaged over five runs. Best results are shown in bold." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 47, + 536, + 178, + 550 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 536, + 178, + 550 + ], + "spans": [ + { + "bbox": [ + 47, + 536, + 178, + 550 + ], + "type": "text", + "content": "4.1. Implementation details" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 557, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 557, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 557, + 288, + 714 + ], + "type": "text", + "content": "We implement our network and conduct experiments using Pytorch 1.11.0 on a single NVIDIA RTX A6000 GPU with 48GB of memory. We utilize ImageNet [16] pretrained PVTv2-b0 and PVTv2-b2 [56] as encoders. In the MSDC of our decoder, we set the multi-scale kernels [1,3,5] through an ablation study. We use the parallel arrangement of depth-wise convolutions in all experiments. Our models are trained using the AdamW optimizer [36] with a learning rate and weight decay of " + }, + { + "bbox": [ + 46, + 557, + 288, + 714 + ], + "type": "inline_equation", + "content": "1e - 4" + }, + { + "bbox": [ + 46, + 557, + 288, + 714 + ], + "type": "text", + "content": ". We generally train for 200 epochs with a batch size of 16, except for Synapse multi-organ (300 epochs, batch size 6) and ACDC cardiac organ (400 epochs, batch size 12), saving the best model based on the DICE score. We resize images" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 536, + 545, + 658 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 536, + 545, + 658 + ], + "spans": [ + { + "bbox": [ + 304, + 536, + 545, + 658 + ], + "type": "text", + "content": "to " + }, + { + "bbox": [ + 304, + 536, + 545, + 658 + ], + "type": "inline_equation", + "content": "352 \\times 352" + }, + { + "bbox": [ + 304, + 536, + 545, + 658 + ], + "type": "text", + "content": " and use a multi-scale " + }, + { + "bbox": [ + 304, + 536, + 545, + 658 + ], + "type": "inline_equation", + "content": "\\{0.75, 1.0, 1.25\\}" + }, + { + "bbox": [ + 304, + 536, + 545, + 658 + ], + "type": "text", + "content": " training strategy with a gradient clip limit of 0.5 for ClinicDB [3], Kvasir [29], ColonDB [51], ETIS [51], BKAI [40], ISIC17 [15], and ISIC18 [15], while we resize images to " + }, + { + "bbox": [ + 304, + 536, + 545, + 658 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 304, + 536, + 545, + 658 + ], + "type": "text", + "content": " for BUSI [1], EM [6], and DSB18 [4]. For Synapse and ACDC datasets, images are resized to " + }, + { + "bbox": [ + 304, + 536, + 545, + 658 + ], + "type": "inline_equation", + "content": "224 \\times 224" + }, + { + "bbox": [ + 304, + 536, + 545, + 658 + ], + "type": "text", + "content": ", with random rotation and flipping augmentations, optimizing a combined Cross-entropy (0.3) and DICE (0.7) loss. For binary segmentation, we utilize the combined weighted BinaryCrossEntropy (BCE) and weighted IoU loss function." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 669, + 362, + 680 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 669, + 362, + 680 + ], + "spans": [ + { + "bbox": [ + 306, + 669, + 362, + 680 + ], + "type": "text", + "content": "4.2. Results" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 689, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 547, + 713 + ], + "type": "text", + "content": "We compare our architectures (i.e., PVT-EMCAD-B0 and PVT-EMCAD-B2) with SOTA CNN and transformer-based" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "11774" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 67, + 71, + 269, + 220 + ], + "blocks": [ + { + "bbox": [ + 67, + 71, + 269, + 220 + ], + "lines": [ + { + "bbox": [ + 67, + 71, + 269, + 220 + ], + "spans": [ + { + "bbox": [ + 67, + 71, + 269, + 220 + ], + "type": "image", + "image_path": "68e95713cdd7aa1342c98b587e762e015c98e20ad9f521d47ceae1aa69432ee7.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 221, + 288, + 266 + ], + "lines": [ + { + "bbox": [ + 46, + 221, + 288, + 266 + ], + "spans": [ + { + "bbox": [ + 46, + 221, + 288, + 266 + ], + "type": "text", + "content": "Figure 3. Average DICE scores vs. #Params for different methods over 10 binary medical image segmentation datasets. As shown, our proposed approaches (PVT-EMCAD-B0 and PVT-EMCAD-B2) have the fewest parameters, yet the highest DICE scores." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 48, + 272, + 289, + 434 + ], + "blocks": [ + { + "bbox": [ + 48, + 272, + 289, + 434 + ], + "lines": [ + { + "bbox": [ + 48, + 272, + 289, + 434 + ], + "spans": [ + { + "bbox": [ + 48, + 272, + 289, + 434 + ], + "type": "table", + "html": "
MethodsAvg. DICERVMyoLV
R50+UNet [8]87.5587.1080.6394.92
R50+AttnUNet [8]86.7587.5879.2093.47
ViT+CUP [8]81.4581.4670.7192.18
R50+ViT+CUP [8]87.5786.0781.8894.75
TransUNet [8]89.7186.6787.2795.18
SwinUNet [5]88.0785.7784.4294.03
MT-UNet [53]90.4386.6489.0495.62
MISSformer [25]90.8689.5588.0494.99
PVT-CASCADE [42]91.4689.9788.995.50
TransCASCade [42]91.6390.2589.1495.50
Cascaded MERIT [43]91.8590.2389.5395.80
PVT-EMCAD-B0 (Ours)91.34±0.289.3788.9995.65
PVT-EMCAD-B2 (Ours)92.12±0.290.6589.6896.02
", + "image_path": "cc9b5fe31b2429853d4989b4ae359486428262b5afc60e7caacd71fa24d74474.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 480, + 287, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 480, + 287, + 518 + ], + "spans": [ + { + "bbox": [ + 46, + 480, + 287, + 518 + ], + "type": "text", + "content": "segmentation methods on 12 datasets that belong to six medical image segmentation tasks. Qualitative results are in the Supplementary Section 7.3." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 526, + 276, + 540 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 526, + 276, + 540 + ], + "spans": [ + { + "bbox": [ + 46, + 526, + 276, + 540 + ], + "type": "text", + "content": "4.2.1 Results of binary medical image segmentation" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 545, + 287, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 545, + 287, + 629 + ], + "spans": [ + { + "bbox": [ + 46, + 545, + 287, + 629 + ], + "type": "text", + "content": "Results for different methods on 10 binary medical image segmentation datasets are shown in Table 1 and Figure 1. Our PVT-EMCAD-B2 attains the highest average DICE score " + }, + { + "bbox": [ + 46, + 545, + 287, + 629 + ], + "type": "inline_equation", + "content": "(91.10\\%)" + }, + { + "bbox": [ + 46, + 545, + 287, + 629 + ], + "type": "text", + "content": " with only " + }, + { + "bbox": [ + 46, + 545, + 287, + 629 + ], + "type": "inline_equation", + "content": "26.76\\mathrm{M}" + }, + { + "bbox": [ + 46, + 545, + 287, + 629 + ], + "type": "text", + "content": " parameters and 5.6G FLOPs. The multi-scale depth-wise convolution in our EMCAD decoder, combined with the transformer encoder, contributes to these performance gains." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 630, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 630, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 630, + 288, + 714 + ], + "type": "text", + "content": "Polyp segmentation: Table 1 reveals that our PVT-EMCAD-B2 surpasses all SOTA methods in five polyp segmentation datasets. PVT-EMCAD-B2 achieves DICE score improvements of " + }, + { + "bbox": [ + 46, + 630, + 288, + 714 + ], + "type": "inline_equation", + "content": "1.08\\%" + }, + { + "bbox": [ + 46, + 630, + 288, + 714 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 630, + 288, + 714 + ], + "type": "inline_equation", + "content": "0.78\\%" + }, + { + "bbox": [ + 46, + 630, + 288, + 714 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 630, + 288, + 714 + ], + "type": "inline_equation", + "content": "2.36\\%" + }, + { + "bbox": [ + 46, + 630, + 288, + 714 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 630, + 288, + 714 + ], + "type": "inline_equation", + "content": "1.19\\%" + }, + { + "bbox": [ + 46, + 630, + 288, + 714 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 630, + 288, + 714 + ], + "type": "inline_equation", + "content": "1.79\\%" + }, + { + "bbox": [ + 46, + 630, + 288, + 714 + ], + "type": "text", + "content": " over PolypPVT in ClinicDB, ColonDB, ETIS, Kvasir, and BKAI-IGI, despite having slightly more parameters and FLOPs. The smallest model UNeXt, ex" + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 307, + 71, + 547, + 148 + ], + "blocks": [ + { + "bbox": [ + 46, + 437, + 287, + 471 + ], + "lines": [ + { + "bbox": [ + 46, + 437, + 287, + 471 + ], + "spans": [ + { + "bbox": [ + 46, + 437, + 287, + 471 + ], + "type": "text", + "content": "Table 3. Results of cardiac organ segmentation on ACDC dataset. DICE scores " + }, + { + "bbox": [ + 46, + 437, + 287, + 471 + ], + "type": "inline_equation", + "content": "(\\%)" + }, + { + "bbox": [ + 46, + 437, + 287, + 471 + ], + "type": "text", + "content": " are reported for individual organs. We get the results of SwinUNet from [42]. Best results are shown in bold." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 307, + 71, + 547, + 148 + ], + "lines": [ + { + "bbox": [ + 307, + 71, + 547, + 148 + ], + "spans": [ + { + "bbox": [ + 307, + 71, + 547, + 148 + ], + "type": "table", + "html": "
Components#FLOPs(G)#Params (M)Avg DICE
CascadedLGAGMSCAM224256
NoNoNo00080.10±0.2
YesNoNo0.1000.1310.22481.08±0.2
YesYesNo0.1080.1410.23581.92±0.2
YesNoYes0.3730.4871.89882.86±0.3
YesYesYes0.3810.4981.9183.63±0.3
", + "image_path": "2d653adeec59f708f8225a0f18e1acdfba7cd70d71d3dab6b3fe686aa3f65563.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 150, + 547, + 196 + ], + "lines": [ + { + "bbox": [ + 304, + 150, + 547, + 196 + ], + "spans": [ + { + "bbox": [ + 304, + 150, + 547, + 196 + ], + "type": "text", + "content": "Table 4. Effect of different components of EMCAD with PVTv2-b2 encoder on Synapse multi-organ dataset. #FLOPs are reported for input resolution of " + }, + { + "bbox": [ + 304, + 150, + 547, + 196 + ], + "type": "inline_equation", + "content": "224 \\times 224" + }, + { + "bbox": [ + 304, + 150, + 547, + 196 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 150, + 547, + 196 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 304, + 150, + 547, + 196 + ], + "type": "text", + "content": ". All results are averaged over five runs. Best results are shown in bold." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 202, + 545, + 298 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 202, + 545, + 298 + ], + "spans": [ + { + "bbox": [ + 304, + 202, + 545, + 298 + ], + "type": "text", + "content": "hibits the worst performance in all five polyp segmentation datasets. Our smaller model with only 3.92M parameters and 0.84G FLOPs also outperforms all the methods except PVT-CASCADE (in Kvasir and BKAI-IGH) and SSFormer-L (in ColonDB), which achieve the best performance among SOTA methods. In conclusion, our PVT-EMCAD-B2 achieves the new SOTA results in these five polyp segmentation datasets." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 299, + 546, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 299, + 546, + 418 + ], + "spans": [ + { + "bbox": [ + 304, + 299, + 546, + 418 + ], + "type": "text", + "content": "Skin lesion segmentation: Table 1 shows PVT-EMCAD-B2's strong performance on ISIC17 and ISIC18 skin lesion segmentation datasets, achieving DICE scores of " + }, + { + "bbox": [ + 304, + 299, + 546, + 418 + ], + "type": "inline_equation", + "content": "85.95\\%" + }, + { + "bbox": [ + 304, + 299, + 546, + 418 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 299, + 546, + 418 + ], + "type": "inline_equation", + "content": "90.96\\%" + }, + { + "bbox": [ + 304, + 299, + 546, + 418 + ], + "type": "text", + "content": ", surpassing DeepLabV3+ by " + }, + { + "bbox": [ + 304, + 299, + 546, + 418 + ], + "type": "inline_equation", + "content": "2.11\\%" + }, + { + "bbox": [ + 304, + 299, + 546, + 418 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 299, + 546, + 418 + ], + "type": "inline_equation", + "content": "2.32\\%" + }, + { + "bbox": [ + 304, + 299, + 546, + 418 + ], + "type": "text", + "content": ". It also beats the nearest method PVT-CASCADE by " + }, + { + "bbox": [ + 304, + 299, + 546, + 418 + ], + "type": "inline_equation", + "content": "0.45\\%" + }, + { + "bbox": [ + 304, + 299, + 546, + 418 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 299, + 546, + 418 + ], + "type": "inline_equation", + "content": "0.55\\%" + }, + { + "bbox": [ + 304, + 299, + 546, + 418 + ], + "type": "text", + "content": " in ISIC17 and ISIC18, respectively, though our decoder is significantly more efficient than CASCADE. Our PVT-EMCAD-B0 also shows huge potential in point care applications like skin lesion segmentation with only 3.92M parameters and 0.84G FLOPs." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 419, + 547, + 527 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 419, + 547, + 527 + ], + "spans": [ + { + "bbox": [ + 304, + 419, + 547, + 527 + ], + "type": "text", + "content": "Cell segmentation: To evaluate our method's effectiveness in biological imaging, we use DSB18 [4] for cell nuclei and EM [6] for cell structure segmentation. As Table 1 indicates, our PVT-EMCAD-B2 sets a SOTA benchmark in cell nuclei segmentation on DSB18, outperforming DeepLabv3+, TransFuse, and PVT-CASCADE. On the EM dataset, PVT-EMCAD-B2 secures the second-best DICE score (95.53%), offering significantly lower computational costs than the top-performing AttnUNet (95.55%)." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 527, + 547, + 599 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 527, + 547, + 599 + ], + "spans": [ + { + "bbox": [ + 304, + 527, + 547, + 599 + ], + "type": "text", + "content": "Breast cancer segmentation: We conduct experiments on the BUSI dataset for breast cancer segmentation in ultrasound images. Our PVT-EMCAD-B2 achieves the SOTA DICE score " + }, + { + "bbox": [ + 304, + 527, + 547, + 599 + ], + "type": "inline_equation", + "content": "(80.25\\%)" + }, + { + "bbox": [ + 304, + 527, + 547, + 599 + ], + "type": "text", + "content": " on this dataset. Furthermore, our PVT-EMCAD-B0 outperforms the computationally similar method UNeXt by a notable margin of " + }, + { + "bbox": [ + 304, + 527, + 547, + 599 + ], + "type": "inline_equation", + "content": "5.54\\%" + }, + { + "bbox": [ + 304, + 527, + 547, + 599 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 610, + 509, + 623 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 610, + 509, + 623 + ], + "spans": [ + { + "bbox": [ + 305, + 610, + 509, + 623 + ], + "type": "text", + "content": "4.2.2 Results of abdomen organ segmentation" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 629, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 629, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 629, + 547, + 715 + ], + "type": "text", + "content": "Table 2 shows that our PVT-EMCAD-B2 excels in abdomen organ segmentation on the Synapse multi-organ dataset, achieving the highest average DICE score of " + }, + { + "bbox": [ + 304, + 629, + 547, + 715 + ], + "type": "inline_equation", + "content": "83.63\\%" + }, + { + "bbox": [ + 304, + 629, + 547, + 715 + ], + "type": "text", + "content": " and surpassing all SOTA CNN- and transformer-based methods. It outperforms PVT-CASCADE by " + }, + { + "bbox": [ + 304, + 629, + 547, + 715 + ], + "type": "inline_equation", + "content": "2.57\\%" + }, + { + "bbox": [ + 304, + 629, + 547, + 715 + ], + "type": "text", + "content": " in DICE score and 4.55 in HD95 distance, indicating superior organ boundary location. Our EMCAD decoder boosts individ" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "11775" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 67, + 70, + 525, + 117 + ], + "blocks": [ + { + "bbox": [ + 67, + 70, + 525, + 117 + ], + "lines": [ + { + "bbox": [ + 67, + 70, + 525, + 117 + ], + "spans": [ + { + "bbox": [ + 67, + 70, + 525, + 117 + ], + "type": "table", + "html": "
Conv. kernels[1][3][5][1,3][3,3][1,3,5][3,3,3][3,5,7][1,3,5,7][1,3,5,7,9]
Synapse82.4382.7982.7482.9882.8183.6382.9283.1183.5783.34
ClinicDB94.8194.9094.9895.1395.0695.2195.1595.0395.1895.07
", + "image_path": "960dbdb51d3a1ddc506c2d15b7d43fb9faa404d4b0f8e5c2be86685d76f5ca3f.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 48, + 150, + 288, + 212 + ], + "blocks": [ + { + "bbox": [ + 47, + 120, + 545, + 143 + ], + "lines": [ + { + "bbox": [ + 47, + 120, + 545, + 143 + ], + "spans": [ + { + "bbox": [ + 47, + 120, + 545, + 143 + ], + "type": "text", + "content": "Table 5. Effect of multi-scale kernels in the depth-wise convolution of MSDC on ClinicDB and Synapse multi-organ datasets. We use the PVTv2-b2 encoder for these experiments. All results are averaged over five runs. Best results are highlighted in bold." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 48, + 150, + 288, + 212 + ], + "lines": [ + { + "bbox": [ + 48, + 150, + 288, + 212 + ], + "spans": [ + { + "bbox": [ + 48, + 150, + 288, + 212 + ], + "type": "table", + "html": "
EncodersDecoders#FLOPs(G)#Params(M)DICE (%)
PVTv2-B0CASCADE0.4392.3280.54
PVTv2-B0EMCAD (Ours)0.1100.50781.97
PVTv2-B2CASCADE1.939.2782.78
PVTv2-B2EMCAD (Ours)0.3811.9183.63
", + "image_path": "795666650d4a54a2666e7382c68cfd032201b42d5155ab078918ba97932efa6c.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 47, + 213, + 286, + 258 + ], + "lines": [ + { + "bbox": [ + 47, + 213, + 286, + 258 + ], + "spans": [ + { + "bbox": [ + 47, + 213, + 286, + 258 + ], + "type": "text", + "content": "Table 6. Comparison with the baseline decoder on Synapse Multi-organ dataset. We only report the #FLOPs (with input resolution of " + }, + { + "bbox": [ + 47, + 213, + 286, + 258 + ], + "type": "inline_equation", + "content": "224 \\times 224" + }, + { + "bbox": [ + 47, + 213, + 286, + 258 + ], + "type": "text", + "content": ") and the #parameters of the decoders. All the results are averaged over five runs. Best results are shown in bold." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 262, + 286, + 286 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 262, + 286, + 286 + ], + "spans": [ + { + "bbox": [ + 47, + 262, + 286, + 286 + ], + "type": "text", + "content": "ual organ segmentation, significantly outperforming SOTA methods on six of eight organs." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 289, + 242, + 301 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 289, + 242, + 301 + ], + "spans": [ + { + "bbox": [ + 47, + 289, + 242, + 301 + ], + "type": "text", + "content": "4.2.3 Results of cardiac organ segmentation" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 308, + 286, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 308, + 286, + 403 + ], + "spans": [ + { + "bbox": [ + 46, + 308, + 286, + 403 + ], + "type": "text", + "content": "Table 3 shows the DICE scores of our PVT-EMCAD-B2 and PVT-EMCAD-B0 along with other SOTA methods, on the MRI images of the ACDC dataset for cardiac organ segmentation. Our PVT-EMCAD-B2 achieves the highest average DICE score of " + }, + { + "bbox": [ + 46, + 308, + 286, + 403 + ], + "type": "inline_equation", + "content": "92.12\\%" + }, + { + "bbox": [ + 46, + 308, + 286, + 403 + ], + "type": "text", + "content": ", thus improving about " + }, + { + "bbox": [ + 46, + 308, + 286, + 403 + ], + "type": "inline_equation", + "content": "0.27\\%" + }, + { + "bbox": [ + 46, + 308, + 286, + 403 + ], + "type": "text", + "content": " over Cascaded MERIT though our network has significantly lower computational cost. Besides, PVT-EMCAD-B2 has better DICE scores in all three organ segmentations." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 411, + 147, + 423 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 411, + 147, + 423 + ], + "spans": [ + { + "bbox": [ + 47, + 411, + 147, + 423 + ], + "type": "text", + "content": "5. Ablation Studies" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 430, + 286, + 466 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 430, + 286, + 466 + ], + "spans": [ + { + "bbox": [ + 46, + 430, + 286, + 466 + ], + "type": "text", + "content": "In this section, we conduct ablation studies to explore different aspects of our architectures and the experimental framework. More ablations are in Supplementary Section 8." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 469, + 267, + 482 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 469, + 267, + 482 + ], + "spans": [ + { + "bbox": [ + 47, + 469, + 267, + 482 + ], + "type": "text", + "content": "5.1. Effect of different components of EMCAD" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 488, + 286, + 643 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 488, + 286, + 643 + ], + "spans": [ + { + "bbox": [ + 46, + 488, + 286, + 643 + ], + "type": "text", + "content": "We conduct a set of experiments on the Synapse multi-organ dataset to understand the effect of different components of our EMCAD decoder. We start with only the encoder and add different modules such as Cascaded structure, LGAG, and MSCAM to understand their effect. Table 4 exhibits that the cascaded structure of the decoder helps to improve performance over the non-cascaded one. The incorporation of LGAG and MSCAM improves performance, however, MSCAM proves to be more effective. When both the LGAG and MSCAM modules are used together, it produces the best DICE score of " + }, + { + "bbox": [ + 46, + 488, + 286, + 643 + ], + "type": "inline_equation", + "content": "83.63\\%" + }, + { + "bbox": [ + 46, + 488, + 286, + 643 + ], + "type": "text", + "content": ". It is also evident that there is about " + }, + { + "bbox": [ + 46, + 488, + 286, + 643 + ], + "type": "inline_equation", + "content": "3.53\\%" + }, + { + "bbox": [ + 46, + 488, + 286, + 643 + ], + "type": "text", + "content": " improvement in the DICE score with an additional 0.381G FLOPs and 1.91M parameters." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 647, + 257, + 659 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 647, + 257, + 659 + ], + "spans": [ + { + "bbox": [ + 47, + 647, + 257, + 659 + ], + "type": "text", + "content": "5.2. Effect of multi-scale kernels in MSCAM" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 665, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 286, + 713 + ], + "type": "text", + "content": "We have conducted another set of experiments on Synapse multi-organ and ClinicDB datasets to understand the effect of different multi-scale kernels used for depth-wise convolutions in MSDC. Table 5 reports these results which show" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 152, + 545, + 272 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 152, + 545, + 272 + ], + "spans": [ + { + "bbox": [ + 304, + 152, + 545, + 272 + ], + "type": "text", + "content": "that performance improves from " + }, + { + "bbox": [ + 304, + 152, + 545, + 272 + ], + "type": "inline_equation", + "content": "1 \\times 1" + }, + { + "bbox": [ + 304, + 152, + 545, + 272 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 304, + 152, + 545, + 272 + ], + "type": "inline_equation", + "content": "3 \\times 3" + }, + { + "bbox": [ + 304, + 152, + 545, + 272 + ], + "type": "text", + "content": " kernel. When " + }, + { + "bbox": [ + 304, + 152, + 545, + 272 + ], + "type": "inline_equation", + "content": "1 \\times 1" + }, + { + "bbox": [ + 304, + 152, + 545, + 272 + ], + "type": "text", + "content": " kernel is used together with " + }, + { + "bbox": [ + 304, + 152, + 545, + 272 + ], + "type": "inline_equation", + "content": "3 \\times 3" + }, + { + "bbox": [ + 304, + 152, + 545, + 272 + ], + "type": "text", + "content": " it improves more than when using them alone. However, when two " + }, + { + "bbox": [ + 304, + 152, + 545, + 272 + ], + "type": "inline_equation", + "content": "3 \\times 3" + }, + { + "bbox": [ + 304, + 152, + 545, + 272 + ], + "type": "text", + "content": " kernels are used together, performance drops. The incorporation of a " + }, + { + "bbox": [ + 304, + 152, + 545, + 272 + ], + "type": "inline_equation", + "content": "5 \\times 5" + }, + { + "bbox": [ + 304, + 152, + 545, + 272 + ], + "type": "text", + "content": " kernel with " + }, + { + "bbox": [ + 304, + 152, + 545, + 272 + ], + "type": "inline_equation", + "content": "1 \\times 1" + }, + { + "bbox": [ + 304, + 152, + 545, + 272 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 152, + 545, + 272 + ], + "type": "inline_equation", + "content": "3 \\times 3" + }, + { + "bbox": [ + 304, + 152, + 545, + 272 + ], + "type": "text", + "content": " kernels further improves the performance and it achieves the best results in both Synapse multi-organ and ClinicDB datasets. If we add additional larger kernels (e.g., " + }, + { + "bbox": [ + 304, + 152, + 545, + 272 + ], + "type": "inline_equation", + "content": "7 \\times 7" + }, + { + "bbox": [ + 304, + 152, + 545, + 272 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 152, + 545, + 272 + ], + "type": "inline_equation", + "content": "9 \\times 9" + }, + { + "bbox": [ + 304, + 152, + 545, + 272 + ], + "type": "text", + "content": "), the performance of both datasets drops. Based on these empirical observations, we choose [1, 3, 5] kernels in all our experiments." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 279, + 507, + 292 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 279, + 507, + 292 + ], + "spans": [ + { + "bbox": [ + 305, + 279, + 507, + 292 + ], + "type": "text", + "content": "5.3. Comparison with the baseline decoder" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 298, + 545, + 406 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 298, + 545, + 406 + ], + "spans": [ + { + "bbox": [ + 304, + 298, + 545, + 406 + ], + "type": "text", + "content": "In Table 6, we report the experimental results with the computational complexity of our EMCAD decoder and a baseline decoder, namely CASCADE. From Table 6, we can see that our EMCAD decoder with PVTv2-b2 requires " + }, + { + "bbox": [ + 304, + 298, + 545, + 406 + ], + "type": "inline_equation", + "content": "80.3\\%" + }, + { + "bbox": [ + 304, + 298, + 545, + 406 + ], + "type": "text", + "content": " fewer FLOPs and " + }, + { + "bbox": [ + 304, + 298, + 545, + 406 + ], + "type": "inline_equation", + "content": "79.4\\%" + }, + { + "bbox": [ + 304, + 298, + 545, + 406 + ], + "type": "text", + "content": " fewer parameters to outperform (by " + }, + { + "bbox": [ + 304, + 298, + 545, + 406 + ], + "type": "inline_equation", + "content": "0.85\\%" + }, + { + "bbox": [ + 304, + 298, + 545, + 406 + ], + "type": "text", + "content": ") the respective CASCADE decoder. Similarly, our EMCAD decoder with PVTv2-B0 achieves " + }, + { + "bbox": [ + 304, + 298, + 545, + 406 + ], + "type": "inline_equation", + "content": "1.43\\%" + }, + { + "bbox": [ + 304, + 298, + 545, + 406 + ], + "type": "text", + "content": " better DICE score than the CASCADE decoder with " + }, + { + "bbox": [ + 304, + 298, + 545, + 406 + ], + "type": "inline_equation", + "content": "78.1\\%" + }, + { + "bbox": [ + 304, + 298, + 545, + 406 + ], + "type": "text", + "content": " fewer parameters and " + }, + { + "bbox": [ + 304, + 298, + 545, + 406 + ], + "type": "inline_equation", + "content": "74.9\\%" + }, + { + "bbox": [ + 304, + 298, + 545, + 406 + ], + "type": "text", + "content": " fewer FLOPs." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 414, + 383, + 426 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 414, + 383, + 426 + ], + "spans": [ + { + "bbox": [ + 306, + 414, + 383, + 426 + ], + "type": "text", + "content": "6. Conclusions" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 434, + 545, + 541 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 434, + 545, + 541 + ], + "spans": [ + { + "bbox": [ + 304, + 434, + 545, + 541 + ], + "type": "text", + "content": "In this paper, we have presented EMCAD, a new and efficient multi-scale convolutional attention decoder designed for multi-stage feature aggregation and refinement in medical image segmentation. EMCAD employs a multi-scale depth-wise convolution block, which is key for capturing diverse scale information within feature maps, a critical factor for precision in medical image segmentation. This design choice, using depth-wise convolutions instead of standard " + }, + { + "bbox": [ + 304, + 434, + 545, + 541 + ], + "type": "inline_equation", + "content": "3 \\times 3" + }, + { + "bbox": [ + 304, + 434, + 545, + 541 + ], + "type": "text", + "content": " convolution blocks, makes EMCAD notably efficient." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 543, + 545, + 673 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 543, + 545, + 673 + ], + "spans": [ + { + "bbox": [ + 304, + 543, + 545, + 673 + ], + "type": "text", + "content": "Our experiments reveal that EMCAD surpasses the recent CASCADE decoder in DICE scores with " + }, + { + "bbox": [ + 304, + 543, + 545, + 673 + ], + "type": "inline_equation", + "content": "79.4\\%" + }, + { + "bbox": [ + 304, + 543, + 545, + 673 + ], + "type": "text", + "content": " fewer parameters and " + }, + { + "bbox": [ + 304, + 543, + 545, + 673 + ], + "type": "inline_equation", + "content": "80.3\\%" + }, + { + "bbox": [ + 304, + 543, + 545, + 673 + ], + "type": "text", + "content": " less FLOPs. Our extensive experiments also confirm EMCAD's superior performance compared to SOTA methods across 12 public datasets covering six different 2D medical image segmentation tasks. EMCAD's compatibility with smaller encoders makes it an excellent fit for point-of-care applications while maintaining high performance. We anticipate that our EMCAD decoder will be a valuable asset in enhancing a variety of medical image segmentation and semantic segmentation tasks." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 674, + 545, + 708 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 674, + 545, + 708 + ], + "spans": [ + { + "bbox": [ + 304, + 674, + 545, + 708 + ], + "type": "text", + "content": "Acknowledgements: This work is supported in part by the NSF grant CNS 2007284, and in part by the iMAGiNE Consortium (https://imagine.utexas.edu/)." + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "11776" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 123 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 123 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 123 + ], + "type": "text", + "content": "[1] Walid Al-Dhabyani, Mohammed Gomaa, Hussien Khaled, and Aly Fahmy. Dataset of breast ultrasound images. Data in brief, 28:104863, 2020. 6, 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 125, + 287, + 167 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 125, + 287, + 167 + ], + "spans": [ + { + "bbox": [ + 53, + 125, + 287, + 167 + ], + "type": "text", + "content": "[2] Vijay Badrinarayanan, Alex Kendall, and Roberto Cipolla. Segnet: A deep convolutional encoder-decoder architecture for image segmentation. IEEE Trans. Pattern Anal. Mach. Intell., 39(12):2481-2495, 2017. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 169, + 287, + 223 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 169, + 287, + 223 + ], + "spans": [ + { + "bbox": [ + 53, + 169, + 287, + 223 + ], + "type": "text", + "content": "[3] Jorge Bernal, F Javier Sánchez, Gloria Fernández-Esparrach, Debora Gil, Cristina Rodríguez, and Fernando Vilarino. Wm-dova maps for accurate polyp highlighting in colonoscopy: Validation vs. saliency maps from physicians. Comput. Med. Imaging Graph., 43:99-111, 2015. 6, 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 224, + 287, + 289 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 224, + 287, + 289 + ], + "spans": [ + { + "bbox": [ + 53, + 224, + 287, + 289 + ], + "type": "text", + "content": "[4] Juan C Caicedo, Allen Goodman, Kyle W Karhohs, Beth A Cimini, Jeanelle Ackerman, Marzieh Haghighi, CherKeng Heng, Tim Becker, Minh Doan, Claire McQuin, et al. Nucleus segmentation across imaging experiments: the 2018 data science bowl. Nature methods, 16(12):1247-1253, 2019. 6, 7, 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 291, + 287, + 334 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 291, + 287, + 334 + ], + "spans": [ + { + "bbox": [ + 53, + 291, + 287, + 334 + ], + "type": "text", + "content": "[5] Hu Cao, Yueyue Wang, Joy Chen, Dongsheng Jiang, Xiaopeng Zhang, Qi Tian, and Manning Wang. Swin-unet: Unet-like pure transformer for medical image segmentation. arXiv preprint arXiv:2105.05537, 2021. 1, 3, 6, 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 335, + 287, + 400 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 335, + 287, + 400 + ], + "spans": [ + { + "bbox": [ + 53, + 335, + 287, + 400 + ], + "type": "text", + "content": "[6] Albert Cardona, Stephan Saalfeld, Stephan Preibisch, Benjamin Schmid, Anchi Cheng, Jim Pulokas, Pavel Tomancak, and Volker Hartenstein. An integrated micro-and macroarchitectural analysis of the drosophila brain by computer-assisted serial section electron microscopy. PLoS biology, 8(10):e1000502, 2010. 6, 7, 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 402, + 287, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 402, + 287, + 445 + ], + "spans": [ + { + "bbox": [ + 53, + 402, + 287, + 445 + ], + "type": "text", + "content": "[7] Gongping Chen, Lei Li, Yu Dai, Jianxun Zhang, and Moi Hoon Yap. Aau-net: an adaptive attention u-net for breast lesions segmentation in ultrasound images. IEEE Trans. Med. Imaging, 2022. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 447, + 287, + 500 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 447, + 287, + 500 + ], + "spans": [ + { + "bbox": [ + 53, + 447, + 287, + 500 + ], + "type": "text", + "content": "[8] Jieneng Chen, Yongyi Lu, Qihang Yu, Xiangde Luo, Ehsan Adeli, Yan Wang, Le Lu, Alan L Yuille, and Yuyin Zhou. Transunet: Transformers make strong encoders for medical image segmentation. arXiv preprint arXiv:2102.04306, 2021. 1, 2, 3, 6, 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 502, + 287, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 502, + 287, + 555 + ], + "spans": [ + { + "bbox": [ + 53, + 502, + 287, + 555 + ], + "type": "text", + "content": "[9] Long Chen, Hanwang Zhang, Jun Xiao, Liqiang Nie, Jian Shao, Wei Liu, and Tat-Seng Chua. Sca-cnn: Spatial and channel-wise attention in convolutional networks for image captioning. In IEEE Conf. Comput. Vis. Pattern Recog., pages 5659-5667, 2017. 3, 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 557, + 287, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 557, + 287, + 611 + ], + "spans": [ + { + "bbox": [ + 48, + 557, + 287, + 611 + ], + "type": "text", + "content": "[10] Liang-Chieh Chen, George Papandreou, Iasonas Kokkinos, Kevin Murphy, and Alan L Yuille. Deeplab: Semantic image segmentation with deep convolutional nets, atrous convolution, and fully connected crfs. IEEE Trans. Pattern Anal. Mach. Intell., 40(4):834-848, 2017. 2, 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 613, + 287, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 613, + 287, + 656 + ], + "spans": [ + { + "bbox": [ + 48, + 613, + 287, + 656 + ], + "type": "text", + "content": "[11] Liang-Chieh Chen, Yukun Zhu, George Papandreou, Florian Schroff, and Hartwig Adam. Encoder-decoder with atrous separable convolution for semantic image segmentation. In Eur. Conf. Comput. Vis., pages 801-818, 2018. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 658, + 287, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 287, + 689 + ], + "type": "text", + "content": "[12] Shuhan Chen, Xiuli Tan, Ben Wang, and Xuelong Hu. Reverse attention for salient object detection. In Eur. Conf. Comput. Vis., pages 234-250, 2018. 1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "type": "text", + "content": "[13] Xiangxiang Chu, Zhi Tian, Bo Zhang, Xinlong Wang, Xiaolin Wei, Huaxia Xia, and Chunhua Shen. Conditional po" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 327, + 73, + 545, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 545, + 94 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 545, + 94 + ], + "type": "text", + "content": "sitional encodings for vision transformers. arXiv preprint arXiv:2102.10882, 2021. 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 95, + 545, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 95, + 545, + 171 + ], + "spans": [ + { + "bbox": [ + 308, + 95, + 545, + 171 + ], + "type": "text", + "content": "[14] Noel Codella, Veronica Rotemberg, Philipp Tschandl, M Emre Celebi, Stephen Dusza, David Gutman, Brian Helba, Aadi Kalloo, Konstantinos Liopyris, Michael Marchetti, et al. Skin lesion analysis toward melanoma detection 2018: A challenge hosted by the international skin imaging collaboration (isic). arXiv preprint arXiv:1902.03368, 2019. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 172, + 545, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 172, + 545, + 259 + ], + "spans": [ + { + "bbox": [ + 308, + 172, + 545, + 259 + ], + "type": "text", + "content": "[15] Noel CF Codella, David Gutman, M Emre Celebi, Brian Helba, Michael A Marchetti, Stephen W Dusza, Aadi Kalloo, Konstantinos Liopyris, Nabin Mishra, Harald Kittler, et al. Skin lesion analysis toward melanoma detection: A challenge at the 2017 international symposium on biomedical imaging (isbi), hosted by the international skin imaging collaboration (isic). In IEEE Int. Symp. Biomed. Imaging, pages 168-172. IEEE, 2018. 6, 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 260, + 545, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 260, + 545, + 303 + ], + "spans": [ + { + "bbox": [ + 308, + 260, + 545, + 303 + ], + "type": "text", + "content": "[16] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In IEEE Conf. Comput. Vis. Pattern Recog., pages 248-255. IEEE, 2009. 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 304, + 545, + 347 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 304, + 545, + 347 + ], + "spans": [ + { + "bbox": [ + 308, + 304, + 545, + 347 + ], + "type": "text", + "content": "[17] Bo Dong, Wenhai Wang, Deng-Ping Fan, Jinpeng Li, Huazhu Fu, and Ling Shao. Polyp-pvt: Polyp segmentation with pyramid vision transformers. arXiv preprint arXiv:2108.06932, 2021. 1, 3, 5, 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 349, + 545, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 349, + 545, + 413 + ], + "spans": [ + { + "bbox": [ + 308, + 349, + 545, + 413 + ], + "type": "text", + "content": "[18] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 1, 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 415, + 545, + 447 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 415, + 545, + 447 + ], + "spans": [ + { + "bbox": [ + 308, + 415, + 545, + 447 + ], + "type": "text", + "content": "[19] Isensee et al. nnu-net: a self-configuring method for deep learning-based biomedical image segmentation. Nature methods, 18(2):203-211, 2021. 1, 2, 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 449, + 545, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 449, + 545, + 502 + ], + "spans": [ + { + "bbox": [ + 308, + 449, + 545, + 502 + ], + "type": "text", + "content": "[20] Deng-Ping Fan, Ge-Peng Ji, Tao Zhou, Geng Chen, Huazhu Fu, Jianbing Shen, and Ling Shao. Pranet: Parallel reverse attention network for polyp segmentation. In Int. Conf. Med. Image Comput. Comput. Assist. Interv., pages 263-273. Springer, 2020. 1, 3, 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 503, + 545, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 503, + 545, + 536 + ], + "spans": [ + { + "bbox": [ + 308, + 503, + 545, + 536 + ], + "type": "text", + "content": "[21] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In IEEE Conf. Comput. Vis. Pattern Recog., pages 770-778, 2016. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 537, + 545, + 591 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 537, + 545, + 591 + ], + "spans": [ + { + "bbox": [ + 308, + 537, + 545, + 591 + ], + "type": "text", + "content": "[22] Andrew G Howard, Menglong Zhu, Bo Chen, Dmitry Kalenichenko, Weijun Wang, Tobias Weyand, Marco Andreetto, and Hartwig Adam. Mobilenets: Efficient convolutional neural networks for mobile vision applications. arXiv preprint arXiv:1704.04861, 2017. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 592, + 545, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 592, + 545, + 624 + ], + "spans": [ + { + "bbox": [ + 308, + 592, + 545, + 624 + ], + "type": "text", + "content": "[23] Jie Hu, Li Shen, and Gang Sun. Squeeze-and-excitation networks. In IEEE Conf. Comput. Vis. Pattern Recog., pages 7132-7141, 2018. 2, 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 625, + 545, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 625, + 545, + 678 + ], + "spans": [ + { + "bbox": [ + 308, + 625, + 545, + 678 + ], + "type": "text", + "content": "[24] Huimin Huang, Lanfen Lin, Ruofeng Tong, Hongjie Hu, Qiaowei Zhang, Yutaro Iwamoto, Xianhua Han, Yen-Wei Chen, and Jian Wu. Unet 3+: A full-scale connected unet for medical image segmentation. In ICASSP, pages 1055–1059. IEEE, 2020. 1, 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 680, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 680, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 680, + 545, + 713 + ], + "type": "text", + "content": "[25] Xiaohong Huang, Zhifang Deng, Dandan Li, and Xueguang Yuan. Missformer: An effective medical image segmentation transformer. arXiv preprint arXiv:2109.07162, 2021. 6, 7" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "11777" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "text", + "content": "[26] Nabil Ibtehaz and Daisuke Kihara. Acc-unet: A completely convolutional unet model for the 2020s. In Int. Conf. Med. Image Comput. Comput. Assist. Interv., pages 692-702. Springer, 2023. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 119, + 287, + 163 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 119, + 287, + 163 + ], + "spans": [ + { + "bbox": [ + 48, + 119, + 287, + 163 + ], + "type": "text", + "content": "[27] Sergey Ioffe and Christian Szegedy. Batch normalization: Accelerating deep network training by reducing internal covariate shift. In Int. Conf. Mach. Learn., pages 448-456. pmlr, 2015. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 165, + 287, + 198 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 165, + 287, + 198 + ], + "spans": [ + { + "bbox": [ + 48, + 165, + 287, + 198 + ], + "type": "text", + "content": "[28] Md Amirul Islam, Sen Jia, and Neil DB Bruce. How much position information do convolutional neural networks encode? arXiv preprint arXiv:2001.08248, 2020. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 199, + 287, + 253 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 199, + 287, + 253 + ], + "spans": [ + { + "bbox": [ + 48, + 199, + 287, + 253 + ], + "type": "text", + "content": "[29] Debesh Jha, Pia H Smedsrud, Michael A Riegler, Pål Halvorsen, Thomas de Lange, Dag Johansen, and Håvard D Johansen. Kvasir-seg: A segmented polyp dataset. In Int. Conf. Multimedia Model., pages 451-462. Springer, 2020. 6, 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 256, + 287, + 289 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 256, + 287, + 289 + ], + "spans": [ + { + "bbox": [ + 48, + 256, + 287, + 289 + ], + "type": "text", + "content": "[30] Taehun Kim, Hyemin Lee, and Daijin Kim. Uacanet: Uncertainty augmented context attention for polyp segmentation. In ACM Int. Conf. Multimedia, pages 2167-2175, 2021. 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 291, + 287, + 322 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 291, + 287, + 322 + ], + "spans": [ + { + "bbox": [ + 48, + 291, + 287, + 322 + ], + "type": "text", + "content": "[31] Alex Krizhevsky and Geoff Hinton. Convolutional deep belief networks on CIFar-10. Unpublished manuscript, 40(7): 1-9, 2010. 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 326, + 287, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 326, + 287, + 358 + ], + "spans": [ + { + "bbox": [ + 48, + 326, + 287, + 358 + ], + "type": "text", + "content": "[32] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convolutional neural networks. Adv. Neural Inform. Process. Syst., 25, 2012. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 361, + 287, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 361, + 287, + 415 + ], + "spans": [ + { + "bbox": [ + 48, + 361, + 287, + 415 + ], + "type": "text", + "content": "[33] Xian Lin, Zengqiang Yan, Xianbo Deng, Chuansheng Zheng, and Li Yu. Convformer: Plug-and-play cnn-style transformers for improving medical image segmentation. In Int. Conf. Med. Image Comput. Comput. Assist. Interv., pages 642-651. Springer, 2023. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 417, + 287, + 460 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 417, + 287, + 460 + ], + "spans": [ + { + "bbox": [ + 48, + 417, + 287, + 460 + ], + "type": "text", + "content": "[34] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Int. Conf. Comput. Vis., pages 10012-10022, 2021. 1, 2, 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 463, + 287, + 505 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 463, + 287, + 505 + ], + "spans": [ + { + "bbox": [ + 48, + 463, + 287, + 505 + ], + "type": "text", + "content": "[35] Zhuang Liu, Hanzi Mao, Chao-Yuan Wu, Christoph Feichtenhofer, Trevor Darrell, and Saining Xie. A convnet for the 2020s. In IEEE Conf. Comput. Vis. Pattern Recog., pages 11976-11986, 2022. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 509, + 287, + 530 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 509, + 287, + 530 + ], + "spans": [ + { + "bbox": [ + 48, + 509, + 287, + 530 + ], + "type": "text", + "content": "[36] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 533, + 287, + 575 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 533, + 287, + 575 + ], + "spans": [ + { + "bbox": [ + 48, + 533, + 287, + 575 + ], + "type": "text", + "content": "[37] Ange Lou, Shuyue Guan, and Murray Loew. Dc-unet: rethinking the u-net architecture with dual channel efficient cnn for medical image segmentation. In Med. Imaging 2021: Image Process., pages 758-768. SPIE, 2021. 1, 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 578, + 287, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 578, + 287, + 621 + ], + "spans": [ + { + "bbox": [ + 48, + 578, + 287, + 621 + ], + "type": "text", + "content": "[38] Ange Lou, Shuyue Guan, Hanseok Ko, and Murray H Loew. Caranet: context axial reverse attention network for segmentation of small medical objects. In Med. Imaging 2022: Image Process., pages 81-92. SPIE, 2022. 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 624, + 287, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 624, + 287, + 657 + ], + "spans": [ + { + "bbox": [ + 48, + 624, + 287, + 657 + ], + "type": "text", + "content": "[39] Vinod Nair and Geoffrey E Hinton. Rectified linear units improve restricted boltzmann machines. In Int. Conf. Mach. Learn., pages 807-814, 2010. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 658, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 287, + 713 + ], + "type": "text", + "content": "[40] Phan Ngoc Lan, Nguyen Sy An, Dao Viet Hang, Dao Van Long, Tran Quang Trung, Nguyen Thi Thuy, and Dinh Viet Sang. Neounet: Towards accurate colon polyp segmentation and neoplasm detection. In Adv. Vis. Comput. - Int. Symp., pages 15-28. Springer, 2021. 6, 1" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 128 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 128 + ], + "type": "text", + "content": "[41] Ozan Oktay, Jo Schlemper, Loic Le Folgoc, Matthew Lee, Mattias Heinrich, Kazunari Misawa, Kensaku Mori, Steven McDonagh, Nils Y Hammerla, Bernhard Kainz, et al. Attention u-net: Learning where to look for the pancreas. arXiv preprint arXiv:1804.03999, 2018. 1, 2, 3, 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 130, + 545, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 130, + 545, + 173 + ], + "spans": [ + { + "bbox": [ + 307, + 130, + 545, + 173 + ], + "type": "text", + "content": "[42] Md Mostafijur Rahman and Radu Marculescu. Medical image segmentation via cascaded attention decoding. In IEEE/CVF Winter Conf. Appl. Comput. Vis., pages 6222-6231, 2023. 1, 3, 4, 5, 6, 7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 176, + 545, + 219 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 176, + 545, + 219 + ], + "spans": [ + { + "bbox": [ + 307, + 176, + 545, + 219 + ], + "type": "text", + "content": "[43] Md Mostafijur Rahman and Radu Marculescu. Multi-scale hierarchical vision transformer with cascaded attention decoding for medical image segmentation. In Med. Imaging Deep Learn., 2023. 1, 3, 5, 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 222, + 545, + 266 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 222, + 545, + 266 + ], + "spans": [ + { + "bbox": [ + 307, + 222, + 545, + 266 + ], + "type": "text", + "content": "[44] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Int. Conf. Med. Image Comput. Comput. Assist. Interv., pages 234-241. Springer, 2015. 1, 2, 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 268, + 545, + 311 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 268, + 545, + 311 + ], + "spans": [ + { + "bbox": [ + 307, + 268, + 545, + 311 + ], + "type": "text", + "content": "[45] Mark Sandler, Andrew Howard, Menglong Zhu, Andrey Zh-moginov, and Liang-Chieh Chen. Mobilenetv2: Inverted residuals and linear bottlenecks. In IEEE Conf. Comput. Vis. Pattern Recog., pages 4510-4520, 2018. 2, 4" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 314, + 545, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 314, + 545, + 346 + ], + "spans": [ + { + "bbox": [ + 307, + 314, + 545, + 346 + ], + "type": "text", + "content": "[46] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556, 2014. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 349, + 545, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 349, + 545, + 403 + ], + "spans": [ + { + "bbox": [ + 307, + 349, + 545, + 403 + ], + "type": "text", + "content": "[47] Christian Szegedy, Wei Liu, Yangqing Jia, Pierre Sermanet, Scott Reed, Dragomir Anguelov, Dumitru Erhan, Vincent Vanhoucke, and Andrew Rabinovich. Going deeper with convolutions. In IEEE Conf. Comput. Vis. Pattern Recog., pages 1-9, 2015. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 405, + 545, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 405, + 545, + 437 + ], + "spans": [ + { + "bbox": [ + 307, + 405, + 545, + 437 + ], + "type": "text", + "content": "[48] Mingxing Tan and Quoc Le. Efficientnet: Rethinking model scaling for convolutional neural networks. In Int. Conf. Mach. Learn., pages 6105-6114. PMLR, 2019. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 441, + 545, + 483 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 441, + 545, + 483 + ], + "spans": [ + { + "bbox": [ + 307, + 441, + 545, + 483 + ], + "type": "text", + "content": "[49] Zhengzhong Tu, Hossein Talebi, Han Zhang, Feng Yang, Peyman Milanfar, Alan Bovik, and Yinxiao Li. Maxvit: Multi-axis vision transformer. In Eur. Conf. Comput. Vis., pages 459-479. Springer, 2022. 1, 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 486, + 545, + 529 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 486, + 545, + 529 + ], + "spans": [ + { + "bbox": [ + 307, + 486, + 545, + 529 + ], + "type": "text", + "content": "[50] Jeya Maria Jose Valanarasu and Vishal M Patel. Unext: Mlp-based rapid medical image segmentation network. In Int. Conf. Med. Image Comput. Comput. Assist. Interv., pages 23-33. Springer, 2022. 6, 1" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 533, + 545, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 533, + 545, + 586 + ], + "spans": [ + { + "bbox": [ + 307, + 533, + 545, + 586 + ], + "type": "text", + "content": "[51] David Vázquez, Jorge Bernal, F Javier Sánchez, Gloria Fernández-Esparrach, Antonio M López, Adriana Romero, Michal Drozdzal, and Aaron Courville. A benchmark for endoluminal scene segmentation of colonoscopy images. J. Healthc. Eng., 2017, 2017. 6, 1" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 589, + 545, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 589, + 545, + 632 + ], + "spans": [ + { + "bbox": [ + 307, + 589, + 545, + 632 + ], + "type": "text", + "content": "[52] Haonan Wang, Peng Cao, Jiaqi Wang, and Osmar R Zaiane. Uctransnet: rethinking the skip connections in u-net from a channel-wise perspective with transformer. In AAAI, pages 2441-2449, 2022. 1, 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 635, + 545, + 677 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 635, + 545, + 677 + ], + "spans": [ + { + "bbox": [ + 307, + 635, + 545, + 677 + ], + "type": "text", + "content": "[53] Hongyi Wang, Shiao Xie, Lanfen Lin, Yutaro Iwamoto, Xian-Hua Han, Yen-Wei Chen, and Ruofeng Tong. Mixed transformer u-net for medical image segmentation. In ICASSP, pages 2390–2394. IEEE, 2022. 6, 7" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 680, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 680, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 680, + 545, + 713 + ], + "type": "text", + "content": "[54] Jinfeng Wang, Qiming Huang, Feilong Tang, Jia Meng, Jionglong Su, and Sifan Song. Stepwise feature fusion: Local guides global. arXiv preprint arXiv:2203.03635, 2022. 1, 6" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "text", + "content": "11778" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 453 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "text", + "content": "[55] Wenhai Wang, Enze Xie, Xiang Li, Deng-Ping Fan, Kaitao Song, Ding Liang, Tong Lu, Ping Luo, and Ling Shao. Pyramid vision transformer: A versatile backbone for dense prediction without convolutions. In Int. Conf. Comput. Vis., pages 568-578, 2021. 1, 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 129, + 287, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 129, + 287, + 173 + ], + "spans": [ + { + "bbox": [ + 49, + 129, + 287, + 173 + ], + "type": "text", + "content": "[56] Wenhai Wang, Enze Xie, Xiang Li, Deng-Ping Fan, Kaitao Song, Ding Liang, Tong Lu, Ping Luo, and Ling Shao. Pvt v2: Improved baselines with pyramid vision transformer. Comput. Vis. Media, 8(3):415-424, 2022. 1, 2, 3, 5, 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 175, + 287, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 175, + 287, + 207 + ], + "spans": [ + { + "bbox": [ + 48, + 175, + 287, + 207 + ], + "type": "text", + "content": "[57] Sanghyun Woo, Jongchan Park, Joon-Young Lee, and In So Kweon. Cbam: Convolutional block attention module. In Eur. Conf. Comput. Vis., pages 3–19, 2018. 1, 3, 4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 209, + 287, + 261 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 209, + 287, + 261 + ], + "spans": [ + { + "bbox": [ + 48, + 209, + 287, + 261 + ], + "type": "text", + "content": "[58] Enze Xie, Wenhai Wang, Zhiding Yu, Anima Anandkumar, Jose M Alvarez, and Ping Luo. Segformer: Simple and efficient design for semantic segmentation with transformers. Adv. Neural Inform. Process. Syst., 34:12077-12090, 2021. 2, 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 263, + 287, + 307 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 263, + 287, + 307 + ], + "spans": [ + { + "bbox": [ + 48, + 263, + 287, + 307 + ], + "type": "text", + "content": "[59] Weihao Yu, Mi Luo, Pan Zhou, Chenyang Si, Yichen Zhou, Xinchao Wang, Jiashi Feng, and Shuicheng Yan. Metaformer is actually what you need for vision. In IEEE Conf. Comput. Vis. Pattern Recog., pages 10819-10829, 2022. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 309, + 287, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 309, + 287, + 352 + ], + "spans": [ + { + "bbox": [ + 48, + 309, + 287, + 352 + ], + "type": "text", + "content": "[60] Xiangyu Zhang, Xinyu Zhou, Mengxiao Lin, and Jian Sun. Shufflenet: An extremely efficient convolutional neural network for mobile devices. In IEEE Conf. Comput. Vis. Pattern Recog., pages 6848-6856, 2018. 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 354, + 287, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 354, + 287, + 397 + ], + "spans": [ + { + "bbox": [ + 48, + 354, + 287, + 397 + ], + "type": "text", + "content": "[61] Yundong Zhang, Huiye Liu, and Qiang Hu. Transfuse: Fusing transformers and cnns for medical image segmentation. In Int. Conf. Med. Image Comput. Comput. Assist. Interv., pages 14-24. Springer, 2021. 1, 3, 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 399, + 287, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 399, + 287, + 453 + ], + "spans": [ + { + "bbox": [ + 48, + 399, + 287, + 453 + ], + "type": "text", + "content": "[62] Zongwei Zhou, Md Mahfuzur Rahman Siddiquee, Nima Tajbakhsh, and Jianming Liang. Unet++: A nested u-net architecture for medical image segmentation. In Deep Learn. Med. Image Anal. Multimodal Learn. Clin. Decis. Support, pages 3-11. Springer, 2018. 1, 2, 6" + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "11779" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/EMOPortraits_ Emotion-enhanced Multimodal One-shot Head Avatars/ab68d1cd-73a3-4a4c-9896-716aeaaff436_content_list.json b/2024/EMOPortraits_ Emotion-enhanced Multimodal One-shot Head Avatars/ab68d1cd-73a3-4a4c-9896-716aeaaff436_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..4e14bab3014990af868abf73ec1d921f538bbca2 --- /dev/null +++ b/2024/EMOPortraits_ Emotion-enhanced Multimodal One-shot Head Avatars/ab68d1cd-73a3-4a4c-9896-716aeaaff436_content_list.json @@ -0,0 +1,2133 @@ +[ + { + "type": "text", + "text": "EMOPortraits: Emotion-enhanced Multimodal One-shot Head Avatars", + "text_level": 1, + "bbox": [ + 125, + 130, + 844, + 150 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Nikita Drobyshev", + "bbox": [ + 86, + 180, + 230, + 199 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Antoni Bigata Casademunt \nStavros Petridis", + "bbox": [ + 271, + 181, + 486, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Konstantinos Vougioukas \nMaja Pantic", + "bbox": [ + 519, + 181, + 728, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Zoe Landgraf", + "bbox": [ + 769, + 181, + 880, + 199 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{n.drobyshev, a/bigata-casademunt22, k.vougioukas, zoe.landgraf15, stavros.petridis04, m.pantic}@imperial.ac.uk Imperial College London", + "bbox": [ + 220, + 219, + 751, + 270 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/5571c03b59a22946cb276d1deded7196a8c084b040541ce529bbee5da0acb26d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 122, + 304, + 271, + 416 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/6f78ae2acda6416384be5581a2bc2ae87413212cd8bfa3feb9df5869a0f8e39f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 269, + 305, + 415, + 416 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/867efe1bddaf1887f13140417c41e46dc4c66e055a39bb4c7e189e5a87cba5b5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 415, + 305, + 560, + 416 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/c12038d2254398ecb4c7ba204508b41706cc611de0c85c955e62dc16a81d8eb4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 560, + 305, + 704, + 416 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/272336b8074a1488e724b3c0fbb531059f7b99cb096e5d5bc5ee91f1570dcb0f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 704, + 305, + 844, + 416 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/e1eb172f4eb38c6ece771d39f7a6007b8dcd48c8729563a1872fa82a97e69526.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 122, + 420, + 269, + 527 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/d5150c201a58cf952e956203832059c28533b90018e0a66348c1f3f500048536.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 269, + 420, + 413, + 527 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/316eae8164a7f14a42016f25d69f3acd05748b74e70f6c167f52db9458ece090.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 413, + 420, + 560, + 527 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/479f825123fe7a77d8b249bb5eb77206881d58fc1981346f540388e496db4655.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 560, + 420, + 699, + 527 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/8da0d6fcd24f450ef1006655f2860be5194a1760923b15e36c23cfbcd55b3cd1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 699, + 420, + 844, + 527 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/1d3be78635c31aa04786c174e323af741e93526bbfde5ad9472c0a7afda2dfaf.jpg", + "image_caption": [ + "Source" + ], + "image_footnote": [], + "bbox": [ + 122, + 530, + 269, + 642 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/a2ab0f9b76c3a8b179b604978468a109567c35f2493df4b12487308d887b0ea9.jpg", + "image_caption": [ + "Driver" + ], + "image_footnote": [], + "bbox": [ + 269, + 530, + 413, + 642 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/7af6d1668fe3d655129ce10d232c0993eb2693facbf0ba6bfeccb40e09db4e40.jpg", + "image_caption": [ + "Output" + ], + "image_footnote": [], + "bbox": [ + 413, + 530, + 560, + 642 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/37fa260d97e9e5d2b12fdb7dff66ab9aa5c86ba957ebcccf9309e28ed674827c.jpg", + "image_caption": [ + "Driver" + ], + "image_footnote": [], + "bbox": [ + 560, + 530, + 699, + 642 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/6f169bc4bce76438b2a929bd4d15373bc33cedd96bc3fcaca9948d0a78a551f7.jpg", + "image_caption": [ + "Output", + "Figure 1. Selected animation results for image-driven mode of our method." + ], + "image_footnote": [], + "bbox": [ + 699, + 530, + 844, + 642 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 691, + 313, + 707 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Head avatars animated by visual signals have gained popularity, particularly in cross-driving synthesis where the driver differs from the animated character, a challenging but highly practical approach. The recently presented MegaPortraits model has demonstrated state-of-the-art results in this domain. We conduct a deep examination and evaluation of this model, with a particular focus on its latent space for facial expression descriptors, and uncover several limitations with its ability to express intense face motions. To address these limitations, we propose substantial changes in both training pipeline and model architecture, to introduce our EMOPortraits model, where we:", + "bbox": [ + 75, + 719, + 472, + 900 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Enhance the model's capability to faithfully support intense, asymmetric face expressions, setting a new state-of-the-art result in the emotion transfer task, surpassing previous methods in both metrics and quality.", + "bbox": [ + 496, + 712, + 893, + 772 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Incorporate speech-driven mode to our model, achieving top-tier performance in audio-driven facial animation, making it possible to drive source identity through diverse modalities, including visual signal, audio, or a blend of both.", + "bbox": [ + 496, + 776, + 893, + 837 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Furthermore, we propose a novel multi-view video dataset featuring a wide range of intense and asymmetric facial expressions, filling the gap with absence of such data in existing datasets.", + "bbox": [ + 496, + 840, + 893, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "8498", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 78, + 90, + 207, + 104 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "1.1. Representation of facial expression", + "text_level": 1, + "bbox": [ + 78, + 114, + 379, + 131 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Advancements in neural head technologies now enable the creation of realistic avatars from a few or even one image, with the latter being crucial when only one source image is available. Cross-driving synthesis, where avatars are animated with different identities is a key technique for virtual reality, filmmaking, photo animation, etc. However, accurate transfer of facial expressions, especially those that are intense and uneven, remains a substantial challenge in avatar animation, particularly for cross-driving synthesis. While previous research focused mainly on preserving identity and transferring moderate facial motions, our work also seeks to accurately drive high-intensity and asymmetric expressions.", + "bbox": [ + 76, + 138, + 468, + 318 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our research is built on and extends the MegaPortraits [9] model, notable for state-of-the-art results in cross-driving synthesis. An additional advantage using this method is that, unlike many avatar systems that depend on limited [3, 25] predefined motion descriptors, MegaPortraits, learns expression representations from scratch, allowing for greater adaptability to a wider range of expressions. We investigate the latent expression spaces and training methods of MegaPortraits to enhance its ability to depict a broad spectrum of facial expressions. Our comprehensive analysis reveals that, while the original model shows limited effectiveness in intense motion representing, it has significant potential for improvement through targeted architectural modifications, adjustments in the training approach, and the integration of our novel dataset.", + "bbox": [ + 76, + 320, + 470, + 545 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "1.2. Integrating speech driving", + "text_level": 1, + "bbox": [ + 76, + 556, + 315, + 571 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We integrate speech in our model to be used either as a complementary or an alternative driver, a crucial aspect for applications like virtual assistance or mixed reality when a primary visual signal is absent.", + "bbox": [ + 76, + 580, + 468, + 638 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "By improving disentanglement of facial expression latent space, we remove disturbing information from face motion descriptors and emphasize components that solely control lip movements. Based on this, we formulated a novel loss that helps to achieve desirable results. Furthermore, our method can generate plausible head rotations and blinks, which appear natural and enhance its applicability across various tasks. Thus, our final method can drive source identity through image, audio, or a mix of them.", + "bbox": [ + 76, + 641, + 468, + 776 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "1.3. FEED: Facial Extreme Emotions Dataset", + "text_level": 1, + "bbox": [ + 76, + 786, + 426, + 801 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Public data scarcity is a common obstacle in deep learning, the topic of human facial expressions and movements is not an exception. In particular, we note a lack of high quality video data capturing a wide range of facial expressions, with existing datasets not going beyond basic facial actions shown at Fig. 2. To address this, we collect a new dataset that", + "bbox": [ + 76, + 810, + 468, + 898 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "includes basic expressions and also captures complex movements like blinks, winks, and head and tongue movements, along with varied extreme expressions which are difficult or impossible to categorize through basic facial actions. Given this variety in expressions in our dataset, we believe that it will be a valuable resource for research in human emotions and facial reconstruction fields. In summary, our main contributions are as follows:", + "bbox": [ + 496, + 90, + 893, + 210 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We introduce our new EMOPortraits model for one-shot head avatars synthesis, that is capable of transferring intense facial expression, showing state-of-the-art results in cross-driving synthesis. We achieve this through careful latent facial expression space development as well as novel losses and a minimal amount of domain-specific data ( $\\sim$ 0.1% of the train set).", + "- We integrate a speech-driving mode in our model, that demonstrates cutting-edge results in speech-driven animations. It functions effectively alongside visual signals or independently, also generating realistic head rotations and eye blinks.", + "- We present a unique multi-view dataset that spans a broad spectrum of extreme facial expressions, filling the gap of absence of such data in existing datasets." + ], + "bbox": [ + 500, + 223, + 890, + 449 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 476, + 640, + 491 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Visual-driven head avatars", + "text_level": 1, + "bbox": [ + 500, + 506, + 741, + 520 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In recent years, the domain of neural avatars has branched into two prominent sub-fields: person-specific and person-agnostic avatars. While person-specific methodologies [10, 11, 30, 38, 39] excel in delivering stunning realism and motion fidelity for a particular individual, they encounter challenges representing an arbitrary person. Moreover, these approach demands multiple frames of training data, distinct training for each avatar, and can struggle to replicate motions not encountered during training.", + "bbox": [ + 496, + 532, + 893, + 669 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Person-agnostic methods, on the other hand, don't need train or fine-tuning for each new person and present an alternative approach to talking-head synthesis. Earlier works in this domain generated avatars in a few-shot technique [3, 33], while subsequent studies introduced one-shot capabilities [2, 8, 9, 22, 29, 31, 32, 34]. Many of these works employ predefined motion representations, such as 3DMM's blendshapes [8, 15, 31, 32]. In contrast, some learned latent emotion representation from scratch [3, 9, 22, 34, 35]. This latter approach holds potential for better motion representation as in such settings, expression descriptors become entirely trainable and free from inheriting the limitations typical for blendshapes as was shown in [3, 25]. In our work, we adopt this strategy, remarkably improving upon the methodology used in MegaPortraits[9].", + "bbox": [ + 496, + 674, + 893, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "8499", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "table", + "img_path": "images/66a3d422459b2df9f0a60e5e7a3cd577b47d5f2fcd46c4308a5bbed92a3374f7.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodBFEsFFED dataset internal structure
Head rotationsWinksEyes moveAsym. em.Tongue em.Extreme em.
ModeMild/StrongAxes/randomMild/StrongRandomStrongStrongExtreme
# of part.21/23212321232323
Avg. length1:38/1.520:580:490:331:411:533:02
", + "bbox": [ + 89, + 88, + 457, + 127 + ], + "page_idx": 2 + }, + { + "type": "table", + "img_path": "images/578c2462fd48de5b379bc63eeefef14e809e14ca69ec9aef8856235a65654fea.jpg", + "table_caption": [ + "Table 1. The tasks we asked our participants to perform for our FEED dataset. Here BFEs stands for basic facial expressions, shown in Fig. 2" + ], + "table_footnote": [], + "table_body": "
DatasetDatasets comparison
#Actors#ViewsBasic exp.Strong exp.Extrme & Assym. exp.Tongue exp.Resolution
SAVESS41✓✘✘✘1280 x 1024
RAVGESS241✓✘✘✘1920 x 1080
CREMA-D911✓✘✘✘1280 x 720
MMI251✓✓✘✘1920 x 1080
MEAD487✓✓✘✘1920 x 1080
Ours (FEED)233✓✓✓✓3840 x 2160
", + "bbox": [ + 89, + 195, + 457, + 257 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Table 2. Comparisons with modern, high-quality audio-visual datasets created in controlled settings.", + "bbox": [ + 76, + 267, + 467, + 295 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. Face expression datasets", + "text_level": 1, + "bbox": [ + 76, + 324, + 302, + 339 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Early image datasets[1, 19] predominantly offered annotated face expression data corresponded to up to 8 basic emotions. However, these datasets are not well-suited for training head avatars, which necessitate video or multi-view data.", + "bbox": [ + 76, + 348, + 468, + 407 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The SAVEE [14] dataset captures facial expressions in speech, but involves just 4 actors. The MMI [20] dataset, more expansive in actor participation, offers spontaneous moderate-intensity expression but is restricted to single-view sequences. RAVDESS [18] distinguishes itself by capturing two motion intensities, but its limited recordings challenge broad applicability. CREMA-D [4] encompasses three expression intensities, and though MEAD [28] provides detailed multi-view data across three intensities, it's limited to the standard eight expression groups. Our novel FEED dataset aims to address the scientific community's demand for high-quality multi-view facial expression videos outside the standard categories: Joy, Fear, Sadness, Disgust, Anger Contempt and Surprise.", + "bbox": [ + 75, + 409, + 470, + 621 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.3. Speech-driven head avatars", + "text_level": 1, + "bbox": [ + 76, + 633, + 326, + 648 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "While numerous studies introduce speech-driven avatars [12, 21, 24, 31, 36, 39, 40], few excel in producing high-quality talking heads with authentic rotations, blinks, and the capability to use both visual and audio inputs. For instance, Wav2Lip [21] aims at re-dubbing videos with accurate lip motions, but often falls short in realism with using single images. Diffused heads [24] struggles to generate long sequences and doesn't provide access to pose control. MakeItTalk [40] animates facial landmarks in a speaker-specific manner, yet struggles with head pose controls due to its non-reliance on 3D. PC-AVS [39] offers a solution, but demands a driving video for pose modulation. Newly developed models such as SadTalker [36] and StyleHEAT [31] have shown promising outcomes in generating high-quality talking heads. We conducted a quantitative evaluation of our audio-driven mode compared to the aforementioned models.", + "bbox": [ + 75, + 659, + 472, + 900 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/e582360968f3b94ebaa9e6716b25e04a6f5eb78834c851fa58aea8caced7cf0b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 529, + 87, + 867, + 193 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/33d7247f6f7c8a4a0d29ea41ed87dcebf4c56350ecb13484916dc2b21c5c7926.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 527, + 196, + 867, + 292 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/a0ffc5d39905b35e0710031436f2b4a76dde5773b3d2205c14a51ddcbc729cf5.jpg", + "image_caption": [ + "Figure 2. Illustration of the problem in publicly available face expression data and selected examples from our FEED dataset." + ], + "image_footnote": [], + "bbox": [ + 524, + 294, + 867, + 369 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. FEED dataset", + "text_level": 1, + "bbox": [ + 500, + 441, + 640, + 455 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "As previously mentioned, the current publicly available datasets with human face expression videos fall short in capturing a broad range of facial manipulations (see illustration at Fig. 2). In our view, this gap could stem from the complexity of designing clear tasks for participants, coupled with the reality that not everyone is ready for performing extreme expressions on camera. To bridge this gap, our pioneering FEED dataset is designed to meet the scientific community's need for high-quality, multi-view emotion videos that extend beyond standard emotions. It contains various expressions, including strong asymmetric ones, tongue and cheeks movements, winks, head rotations, eye movements and more nuanced gestures. ;", + "bbox": [ + 496, + 470, + 893, + 667 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our dataset consists of 520 multi-view videos of 23 subjects, captured with 3 cameras. As extreme face motions are complex and their perception can be heavily influenced by subtle differences, we use a high resolution of 4k for all video, capturing the whole face up to the level of individual hair strands and wrinkles, as shown in Fig. 2. For more examples, please refer to the supplementary materials. Our participants were asked to perform 7 tasks (see Tab. 1) to cover as many facial and head movements as possible.", + "bbox": [ + 496, + 670, + 893, + 806 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In our comparison with other expression datasets detailed in Tab. 2, the MEAD dataset [28] emerges as the closest to ours regarding viewpoints and maximal expression intensity. However, despite MEAD's larger number of actors and viewpoints, it offers significantly less variety of facial expressions and have lower resolution than our dataset.", + "bbox": [ + 496, + 810, + 893, + 898 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "8500", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4. Expression enhancement", + "text_level": 1, + "bbox": [ + 76, + 89, + 308, + 107 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We found that the MegaPortraits [9] model fails to transfer intense expressions correctly, as shown in Fig. 9. Using our FEED dataset of strong and uneven facial expression to fine-tune a pre-trained MegaPortraits model didn't enhance final results as shown at Tab. 4. Training from scratch on FEED leads to fast overfitting due to the small number of identities presented in the dataset compared to VoxCeleb2 [7] used in [9] (23 vs $\\sim 5000$ ). Our method effectively injects this small dataset into training, yielding desirable results. We list our key findings here, for all details, see supplementary materials. Our model's scheme, displayed in Fig. 4, shares similarities with [9], leading us to occasionally refer to certain elements of our scheme in the context of [9]'s pipeline.", + "bbox": [ + 75, + 114, + 473, + 311 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1. Latent expression space", + "text_level": 1, + "bbox": [ + 76, + 321, + 295, + 338 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We begin by exploring MegaPortraits' latent expression descriptors' space $(\\mathbf{z}_{s / d}$ in Fig. 4), which is crucial for expression transfer, through PCA analysis. Inspired by [16], which demonstrated that the area under the cumulative explained variance curve of singular values (denoted here as $\\mathbf{AUC}_{\\mathbf{z}}$ ) can serve as an effective metric for dimensionality collapse in some latent spaces, we applied this measure to our study.", + "bbox": [ + 75, + 344, + 470, + 450 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "This metric allows us to forecast model performance: greater collapse suggests reduced entropy in the latent space, which correlates with lower representational quality. High quality of the expression space, produced by $\\mathbf{E}_{\\text {motion }}$ (as shown in Fig. 4), is crucial to capture and differentiate between the nuances of strong facial expressions. Our ablation study confirms (see Tab. 4) that there is a notable correlation between how broad and isotropic model's latent space is (expressed by $\\mathbf{AUC}_{\\mathbf{z}}$ ) and its final performance. Visual comparison is available in supplementary materials. As shown in Fig. 3, our model's latent space outperforms MegaPortraits in expression representation ability. This is supported by both visual (Fig. 9) and quantitative analysis (Tab. 3).", + "bbox": [ + 75, + 450, + 473, + 648 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/43e0be9fd62862a686fc0a3749085dde5d7228aefbfb924c5e39ad2a1a627ee5.jpg", + "image_caption": [ + "Figure 3. Comparison of latent spaces. Left plot shows that our model's latent space is wider and exhibits more even variance distribution. Also, as shown on the right plot, a greater number of principal components are involved in capturing variance across various thresholds. This implies a more robust representational capacity of expression space compared to [9]. The VoxCeleb2 test set was used for both plots." + ], + "image_footnote": [], + "bbox": [ + 96, + 660, + 450, + 789 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/cfaa5cbc7efaf0aaec7497921709bfe2ffc0e67e9a55e614e647ce8545e2ac06.jpg", + "image_caption": [ + "Figure 4. Method Overview. We use $\\mathbf{E}_{\\mathrm{app}}$ to extract volume features $\\mathbf{V}_s$ and a global descriptor $\\mathbf{e}_s$ from the source image. Then $\\mathbf{E}_{\\mathrm{motion}}$ or $\\mathbf{E}_{\\mathrm{audio}}$ generates motion representations from source and driver, including head rotations $\\mathbf{R}_{s / d}$ , translations $\\mathbf{t}_{s / d}$ , and expression descriptors $\\mathbf{z}_{s / d}$ . Using them, we predict warpings $\\mathbf{w}_{s\\rightarrow}$ and $\\mathbf{w}_{\\rightarrow d}$ . First warping and $\\mathbf{G}_{3D}$ transform $\\mathbf{V}_s$ into a canonical volume $\\mathbf{V}_s^C$ by removing the source motions. Second warping and $\\mathbf{G}_{2D}$ imposes the driver's motions and renders the final image." + ], + "image_footnote": [], + "bbox": [ + 539, + 87, + 849, + 200 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Both metrics shown in Fig. 3 were vital in our model's development since they highly correlate with the final model's ability to transfer intense expressions. This was particularly useful for finding adjustments described in Sec. 4.2 and Sec. 4.3, as early visual results during training are not very telling, but changes in the latent space are noticeable after a few epochs. We also found that in MegaPortraits, just a few principal components significantly affect variance, with only 18 components making up $99\\%$ (Fig. 3), suggesting their $512\\mathbf{z}$ 's dimension is oversized. For our model, we found a 128-dimensional $\\mathbf{z}$ is optimal, offering solid performance and reducing overfitting in imbalanced data, as our ablation study reveals. We calculate the Explained Variance $(\\mathrm{EV}_i)$ for the $i^{th}$ component and $(\\mathbf{AUC}_{\\mathbf{z}})$ using equations shown in Eq. (1), which involves standardizing the vector set $\\mathcal{Z}$ for PCA and sorting the eigenvalues $\\lambda$ in descending order.", + "bbox": [ + 496, + 335, + 893, + 579 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {Z} _ {\\mathrm {s t d}} = \\frac {\\mathcal {Z} - \\bar {\\mathcal {Z}}}{\\sigma (\\bar {\\mathcal {Z}})}, \\quad \\boldsymbol {\\lambda} = \\operatorname {s o r t e d} _ {\\mathrm {d e s c}} \\left(\\operatorname {e i g} \\left(\\frac {1}{n - 1} \\mathbf {Z} _ {\\mathrm {s t d}} ^ {\\top} \\mathbf {Z} _ {\\mathrm {s t d}}\\right)\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 506, + 593, + 893, + 633 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {E V} _ {i} = \\frac {\\lambda_ {i}}{\\sum_ {j = 1} ^ {p} \\lambda_ {j}}, \\quad \\mathbf {A U C _ {z}} = \\frac {1}{d} \\frac {\\sum_ {i = 1} ^ {d} \\sum_ {k = 1} ^ {i} \\lambda_ {k}}{\\sum_ {j = 1} ^ {d} \\lambda_ {j}},\n$$\n", + "text_format": "latex", + "bbox": [ + 506, + 627, + 820, + 662 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.2. Canonical volume", + "text_level": 1, + "bbox": [ + 500, + 666, + 673, + 679 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Despite being designed to exclude face expression details, we found that canonical volume $(\\mathbf{V}^C$ in Fig. 4) in MegaPortraits actually retains significant expression information from the source image, contributing to poor translation of intense expressions, as shown in our experiment detailed in Fig. 5. In this experiment, using portraits with varied expressions, we assessed the expression translation accuracy and how expression intensity affects the $\\mathbf{V}^C$ s. Our findings suggest that the canonical volumes in MegaPortraits are not truly neutral. As confirmed by our ablation study (Tab. 4), creating an expression-free $\\mathbf{V}^C$ is essential for precise translation of intense motions. We believe that a more neutral canonical volume improves tractability and effectiveness in expression translation tasks.", + "bbox": [ + 496, + 688, + 893, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "8501", + "bbox": [ + 480, + 944, + 513, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/65fd3cb501e87e562244924ab7734643dc949462bd6d5e71948526318373e41c.jpg", + "image_caption": [ + "Figure 5. Canonical volumes $(\\mathbf{V}^C)$ in MegaPortraits [9] are not expression-neutral. To show it we use three portraits: A (regular expression), B (intense expression), D (regular expression, new identity). Case 1 visualizes poor results in transferring B's intense expression to A using a self-driving mode, contrasting with the effective reconstruction of B when used as both driver and source in case 3. This discrepancy, indicative of expression leakage into the canonical volume, is further quantified by a $43\\%$ relative difference in $\\mathbf{V}^C$ s between B and C, contrary to expectations of their similarity for same identity. In cross-driving generation case 2, using B as the driver and D as the source yielded poor results. However, in case 4, after using additive operations on source canonical volumes, the output expression is much closer to driver than in case 2. This manipulation again confirms the significant retention of expression information in canonical volumes." + ], + "image_footnote": [], + "bbox": [ + 89, + 88, + 465, + 210 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To overcome this issue, we propose to match canonical volumes from the different images of the same person $(\\mathbf{V}_{s^n}^C,\\mathbf{V}_{d^n}^C)$ during training as following:", + "bbox": [ + 75, + 448, + 470, + 494 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {C V}} ^ {\\mathrm {n}} = \\mathcal {L} _ {\\mathrm {M A E}} \\left(\\mathbf {V} _ {s ^ {n}} ^ {C}, \\mathbf {V} _ {d ^ {n}} ^ {C}\\right), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 186, + 505, + 468, + 526 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "This loss ensures $\\mathbf{V}^C$ remains stable and expression-independent, crucial for translating intense expressions, as shown in our ablation study (Tab. 4).", + "bbox": [ + 75, + 529, + 472, + 575 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.3. Source-driver mismatch loss", + "text_level": 1, + "bbox": [ + 76, + 582, + 331, + 597 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In addition to maintaining expression-free $\\mathbf{V}^{C}$ , it's vital to remove all identity information from the expression vector $\\mathbf{z}$ . While contrastive losses address this in [9], our experiments indicate they are insufficient to prevent overfitting in our imbalanced data scenario, where emotionally intense images represent only $0.1\\%$ of our training set but are sampled $25\\%$ of the time. We introduce a novel self-supervised loss which mitigates identity information in latent expression vectors:", + "bbox": [ + 75, + 604, + 470, + 727 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {s d m} (\\mathbf {z} _ {s}, \\mathbf {z} _ {d}) = w * \\max (0, \\cos (\\mathbf {z} _ {s}, \\mathbf {z} _ {d}) - m a r g i n), \\\\ w; \\text {m a r g i n} = \\left\\{ \\begin{array}{l l} 1; 0. 5 & \\text {i f} \\mathbf {z} _ {s}, \\mathbf {z} _ {d} \\text {a r e f r o m V C 2} \\\\ 1 0; 0. 2 5 & \\text {i f} \\mathbf {z} _ {s}, \\mathbf {z} _ {d} \\text {a r e f r o m F E E D} \\end{array} \\right. \\tag {3} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 84, + 744, + 468, + 805 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Here, $\\mathbf{z}_s, \\mathbf{z}_d$ are emotion vectors from source and driver images, $w, \\text{margin}$ adjust the loss's intensity and strictness respectively. We increase $w$ for FEED due to its higher sampling rate and decrease margin for more assured expression variance in $s, d$ . Refer to Fig. 6 for a visual summary of all self-supervised losses enhancing our latent emotion space.", + "bbox": [ + 75, + 810, + 472, + 902 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/4168394fbcbb61f875f732ea64d8e9e00b593fc9b8b7198c39c30b7289b588f7.jpg", + "image_caption": [ + "Figure 6. The visualisation of our self-supervised losses. $\\mathcal{L}_{\\mathrm{cos}}$ is a contrastive loss, similar to one used in [9] but uses different pairs as described in Sec. 5.1. Our novel $\\mathcal{L}_{sdm}$ is designed to prevent overfitting in an imbalanced dataset." + ], + "image_footnote": [], + "bbox": [ + 540, + 87, + 854, + 335 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5. Incorporating speech-driving mode", + "text_level": 1, + "bbox": [ + 500, + 426, + 820, + 445 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.1. Latent space disentanglement", + "text_level": 1, + "bbox": [ + 500, + 452, + 764, + 468 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "During our latent space analysis described in Sec. 4, we found that combining expression from one driver with the head pose from another resulted in poor performance for MegaPortraits model, as illustrated in Fig. 7 (bottom row). Indeed, the base model lacks a mechanism to intentionally prevent head pose leakage. Disentangled expression latent space, besides expanding the model's use cases, plays the crucial role for the speech-driving mode. When the latent vector $\\mathbf{z}$ is entangled with head pose data, predicting it from speech becomes challenging because speech lacks head rotation information, unlike images. Moreover, we would like to have a full control over head rotations during speech mode. We were able to make our latent space disentangled by changing the way we sample images for $\\mathcal{L}_{\\mathrm{cos}}$ Eq. (5).", + "bbox": [ + 496, + 476, + 893, + 688 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The MegaPortraits model uses $\\mathcal{L}_{\\mathrm{cos}}$ - a modified Large Margin Cosine Loss [26] to mitigate the transfer of appearance features to expression descriptor (latent vector $\\mathbf{z}$ on Fig. 4), which is crucial for cross-driving mode. Without this, differences in appearance between the source and driver, like hairstyle or skin tone, leak from driver to output image.", + "bbox": [ + 496, + 688, + 893, + 779 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For computing this loss, the authors utilize a supplementary source-driving pair $(\\mathbf{x}_{s^{*}}$ and $\\mathbf{x}_{d^{*}})$ from a different video with another identity, ensuring a distinct appearance from the current $\\mathbf{x}_s,\\mathbf{x}_d$ pair. The base model is then employed to produce cross-reenacted image $(\\hat{\\mathbf{x}}_{s^{*}\\rightarrow d} = \\mathbf{G}_{\\mathrm{base}}(\\mathbf{x}_{s^{*}},\\mathbf{x}_d))$ Concurrently, they determine a separate motion descriptor, $\\mathbf{z}_{d^{*}} = \\mathbf{E}_{\\mathrm{motion}}(\\mathbf{x}_{d^{*}})$ . Descriptors $\\mathbf{z}_{* \\rightarrow d}$ from the respective forward passes are also used for positive and negative pairs.", + "bbox": [ + 496, + 780, + 895, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "8502", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/2fc771a3720ccaa6e2cde67511df1976c6cbeed5e2acdb994a7fe463b911909b.jpg", + "image_caption": [ + "Figure 7. The illustration of disentanglement problem. Images in the right row indicate the expression descriptor $\\mathbf{z}$ entangled with head pose for MegaPortraits, whereas we avoid it using changes proposed in Sec. 5.1." + ], + "image_footnote": [], + "bbox": [ + 80, + 88, + 467, + 205 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Our innovation here is a new sampling strategy: for each $\\mathbf{x}_s$ , $\\mathbf{x}_d$ pair, we sample one more random additional pair $\\mathbf{x}_{s^m}$ and $\\mathbf{x}_{d^m}$ apart from $\\mathbf{x}_{s^*}$ and $\\mathbf{x}_{d^*}$ . We then apply the model to produce the following cross-reenacted moved image using source identity from $\\mathbf{x}_{s^*}$ , desired head pose from $\\mathbf{x}_{s^m}$ and emotions from $\\mathbf{x}_d$ , represented as: $\\hat{\\mathbf{x}}_{s^{*m}\\rightarrow d} = \\mathbf{G}_{\\mathrm{base}}(\\mathbf{x}_{s^{*m}},\\mathbf{x}_d)$ . This creates a positive pair with the same desired emotion but varying head poses.", + "bbox": [ + 75, + 297, + 468, + 417 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Motion descriptors are organized into positive pairs $\\mathcal{P}$ for alignment, and negative pairs $\\mathcal{N}$ for non-alignment: $\\mathcal{P} = \\left\\{(\\mathbf{z}_{s\\rightarrow d},\\mathbf{z}_d),(\\mathbf{z}_{s^*\\rightarrow d},\\mathbf{z}_d),(\\mathbf{z}_{s^{*m}\\rightarrow d},\\mathbf{z}_d)\\right\\}$ , and $\\mathcal{N} = \\left\\{(\\mathbf{z}_{s\\rightarrow d},\\mathbf{z}_{d^*}),(\\mathbf{z}_{s^*\\rightarrow d},\\mathbf{z}_{d^*}),(\\mathbf{z}_{s^{*m}\\rightarrow d},\\mathbf{z}_{d^*})\\right\\}$ (see Fig. 6). These pairs are used to calculate the following cosine distance:", + "bbox": [ + 76, + 417, + 470, + 507 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nd \\left(\\mathbf {z} _ {i}, \\mathbf {z} _ {j}\\right) = s \\cdot \\left(\\left\\langle \\mathbf {z} _ {i}, \\mathbf {z} _ {j} \\right\\rangle - m\\right), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 169, + 508, + 468, + 526 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where both $s$ and $m$ are hyperparameters. This distance is then used to calculate a large margin cosine loss (Cos-Face) [27]:", + "bbox": [ + 75, + 530, + 468, + 575 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\cos} = - \\sum_ {(\\mathbf {z} _ {k}, \\mathbf {z} _ {l}) \\in \\mathcal {P}} \\log \\frac {\\exp \\left\\{d \\left(\\mathbf {z} _ {k} , \\mathbf {z} _ {l}\\right) \\right\\}}{\\exp \\left\\{d \\left(\\mathbf {z} _ {k} , \\mathbf {z} _ {l}\\right) \\right\\} + \\sum_ {(\\mathbf {z} _ {i} , \\mathbf {z} _ {j}) \\in \\mathcal {N}} \\exp \\left\\{d \\left(\\mathbf {z} _ {i} , \\mathbf {z} _ {j}\\right) \\right\\}} \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 86, + 595, + 468, + 637 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "This loss prevents the head pose from leaking into the embeddings as shown in Fig. 7 (upper row)", + "bbox": [ + 75, + 643, + 468, + 672 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "By doing this, we achieved the capability to interpret $\\mathbf{z}_i$ as a latent vector encapsulating the emotional content from an image $\\mathbf{x}_i$ . This advancement opened the door for an intriguing possibility during inference: the prediction of this vector could potentially be derived from an audio signal instead of a facial image using the motion encoder $\\mathbf{E}_{\\mathrm{motion}}$ , effectively transforming the model into a speech-driven system. The critical challenge here is aligning the audio input with the pretrained latent space. We addressed this by introducing an additional audio encoder, $\\mathbf{E}_{\\mathrm{aud}}$ , and employing $\\mathbf{E}_{\\mathrm{motion}}$ as a teacher model for this new encoder. A crucial factor in achieving remarkable accuracy in lip synchronization was our ability to isolate specific components within the expression latent space, that are solely responsible for mouth movements. We describe it in the next subsection Sec. 5.2.", + "bbox": [ + 75, + 674, + 470, + 898 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/426d2459ddac66dc468bd55fd6056ae9f8154ca624c50c1e768277a3b7502cea.jpg", + "image_caption": [ + "Figure 8. In the upper row, the first principal component from a random subset of expression vectors $\\mathcal{Z}$ affects not just the mouth but also blinks and gaze, making it difficult to interpret and isolate mouth movement components, needed for training $\\mathbf{E}_{\\mathrm{aud}}$ . However, as shown in the bottom row, using $\\mathcal{Z}_m$ from Sec. 5.2, the first principal component, $\\mathbf{v}_1$ , focused on mouth movements, is effectively isolated, $\\mathbf{z}_0$ - latent vector, corresponded to the central image." + ], + "image_footnote": [], + "bbox": [ + 498, + 88, + 890, + 227 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.2. Mouth movements", + "text_level": 1, + "bbox": [ + 500, + 356, + 679, + 371 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To train the audio encoder $\\mathbf{E}_{\\mathrm{aud}}$ for predicting $\\mathbf{z}$ from speech, we chose to use the motion encoder $\\mathbf{E}_{\\mathrm{motion}}$ as a teacher model. However, direct matching vectors $\\mathbf{z}_i^{\\mathrm{aud}}$ from $\\mathbf{E}_{\\mathrm{aud}}$ and pseudo-ground truth $\\mathbf{z}_i$ from $\\mathbf{E}_{\\mathrm{motion}}$ gives poor results (see Tab. 6). We believe that the reason behind this is that speech-derived $\\mathbf{z}_i^{\\mathrm{aud}}$ , while able to capture lip movements, struggles with other facial motions, especially in the upper face. Thus, we focused on isolating mouth movement components in the expression latent space using PCA analysis.", + "bbox": [ + 496, + 378, + 893, + 516 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Employing PCA on a broad array of expression vectors $\\mathcal{Z}$ , generated from images with varied facial emotions, yields principal components with their explained variances $(\\mathrm{EV}_i)$ . Altering these components, particularly those with high $\\mathrm{EV}_i$ , significantly changes facial expressions in an image. However, they typically affect a combination of facial features—Fig. 8 (upper row), not isolated areas like the mouth or eyes.", + "bbox": [ + 496, + 516, + 893, + 621 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To focus on mouth movements, we created a unique set of $\\mathcal{Z}_m$ from just one video of a person performing mouth manipulations and manually edited it to minimize upper face movements. This involved using a still upper face from the first frame in all subsequent frames, ensuring that the principal components mainly represented mouth movements. Applying PCA of $\\mathcal{Z}_m$ reveals that most principal components are responsible for solely mouth movements. For illustration of the first component, see Fig. 8 (bottom row). Based on distilled components, we introduce mouth PCA mouth loss:", + "bbox": [ + 496, + 622, + 893, + 772 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {P C A}} \\left(v _ {i}, v _ {j}, n\\right) = \\frac {1}{n} \\sum_ {k = 1} ^ {n} \\left| \\mathrm {P C} _ {\\mathbf {v} _ {i}} (k) - \\mathrm {P C} _ {\\mathbf {v} _ {j}} (k) \\right| \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 539, + 779, + 890, + 819 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Where $n$ is the number of principal components to be considered, $\\mathrm{PC}_{\\mathbf{z}_i}(k)$ and $\\mathrm{PC}_{\\mathbf{z}_j}(k)$ are the $k$ -th principal components of vectors $\\mathbf{z}_i$ and $\\mathbf{z}_j$ respectively. Implementing this loss has demonstrably enhanced the quality and accuracy of our results, as our ablation study confirms (see Tab. 6).", + "bbox": [ + 496, + 824, + 893, + 898 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "8503", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "6. Experiments", + "text_level": 1, + "bbox": [ + 76, + 90, + 209, + 107 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In this section, we describe experiments we set for comparison for both our image-driven and speech-driven mode. Please see implementation details such as all used losses, component's sizes and architecture details, training procedure, data preprocess in supplementary materials.", + "bbox": [ + 75, + 114, + 470, + 191 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "6.1. Image-driven comparison", + "text_level": 1, + "bbox": [ + 76, + 196, + 310, + 214 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Methods. We evaluated our image-driven model alongside various one-shot models. For this comparison, we included models like NOFA [32] and StyleHEAT [31], which use 3DMM expression blendshapes without disentanglement training, and models with trainable latent face motion representations like FOMM [22], UVA [23], MetaPortrait [35] and MegaPortraits [9]. It's noteworthy that NOFA and UVA require per-source optimization, unlike the others, including our model.", + "bbox": [ + 75, + 220, + 468, + 354 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Evaluation metrics and data. For our evaluation, we use 100 random, non-child, single-person images from the FFHQ dataset as sources. From the MEAD dataset, we chose 100 images covering all high-intensity emotions except neutral, and another 100 from the FEED dataset's extreme emotion task, featuring 5 identities not seen during training of our model. Each source image was paired with 20 driving images from these sets, totaling 2000 pairs.", + "bbox": [ + 75, + 356, + 468, + 477 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "For assessing reenactment quality, we employ various metrics including the Frechet Inception Distance (FID) [13] to measure the distributional discrepancy between synthetic and real images. Cosine similarity (CSIM) from a face recognition network [5] quantifies the identity preservation in generated images. Additionally, we conducted user studies to determine preferences regarding motion (UMTN) and appearance (UAPP) preservation, presenting participants with triplets—either a driving or source image and two method outputs—and asking them to choose the one with better motion or appearance preservation. Our user study includes 34 participants and around 300 unique questions. Both metrics numerically reflect preference percentages. As our visual results Fig. 9, and quantitative metrics Tab. 3 show, our method demonstrates exceptional ability in translating strong and extreme expressions, notably surpassing other image-driven avatar techniques in user preference for facial expression translation (UMTN) and FID scores.", + "bbox": [ + 75, + 478, + 470, + 750 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "6.2. Speech-driven comparison", + "text_level": 1, + "bbox": [ + 76, + 756, + 316, + 772 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Methods. To assess our model's speech-driven capabilities, we compared it with StyleHEAT [31], which also operates using both image and audio inputs, and audio-exclusive models including SadTalker [36], MakeItTalk [40], and PC-AVS [39].", + "bbox": [ + 75, + 779, + 468, + 854 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Evaluation metrics and data. We employ the evaluation procedure described in [12]. We measure video realism using Frechet Inception Distance (FID) [13]. We also extract", + "bbox": [ + 76, + 854, + 470, + 900 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/f723b0a62eb5b4733b04b9bdc559c91c2d518e61d1a5e9187c6db9f9ecf35458.jpg", + "table_caption": [], + "table_footnote": [ + "Table 3. Our method notably outperforms others in the FID score and strongly leads in the user preference metric for face expression translation (UMTN). It excels in reliably translating strong and extreme expressions from the driving image, distinguishing it as the most capable among the compared methods. Additionally, our approach maintains the source image's shape and appearance (measured by CSIM and UAPP metrics) on par with MegaPortraits [9] when using strong and regular expressions from the MEAD dataset as the driving signal. However, we observe a minor deviation in identity preservation metrics for extreme emotions (using the FEED dataset). This difference may arise because our method attempts to transfer asymmetric facial expressions, while MegaPortraits might modify the source image less, or not at all, with a challenging driver (as depicted in Fig. 5, case 1), thereby achieving better identity preservation. UVA was excluded from the MEAD-driven user study due to the authors were able to provide only part of the test data. Also, MetaPortrait was excluded from both user studies because of late data provision by authors, offering only the basic model results, without their super-resolution module outputs." + ], + "table_body": "
Driver:MEADFEED
MethodFID↓CSIM↑UMTN↑UAPP↑FID↓CSIM↑UMTN↑UAPP↑
MetaPortrait [35]77.50.64**78.80.66**
NOFA [32]69.20.65**69.80.663.88.8
FOMM [22]84.40.562.40.385.90.551.80.5
UVA [17]78.70.680.96.479.50.680.69.4
StyleHEAT [31]71.90.633.73.372.40.612.63.8
MegaPortraits [9]61.10.7322.746.662.80.7316.742.8
Our (EMOPortraits)59.60.7470.343.460.20.7074.634.7
", + "bbox": [ + 504, + 89, + 890, + 176 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/873156c9e6afd47be52e4d5a99fb0aa75c8f06eb0509809b0b49c314487dafaa.jpg", + "table_caption": [], + "table_footnote": [ + "Table 4. Our ablation study examines key aspects of the image-driven mode. After fine-tuning MegaPortraits on the FEED dataset, we observed a decline in identity preservation due to overfitting on FEED identities, without any improvement in emotion translation. The lack of $\\mathcal{L}_{CV}^{n}$ primarily impacts the ability to translate extreme expression, as indicated by UMTN. Additionally, not using $\\mathcal{L}_{sdm}$ or setting $dim(\\mathbf{z}) = 512$ results in identity leakage, particularly evident in the latter scenario. For this ablation, we use the part of our driven by MEAD dataset and described in Sec. 6.1" + ], + "table_body": "
MethodFID↓CSIM↑UMTN↓UAPP↓AUCz↓
MegaPortraits fine-tuned62.30.638.512.40.89
Ours w/o LcnCV61.40.679.217.90.88
Ours w/o Lsdm61.80.5911.36.70.85
Ours dim(z) = 51263.40.4425.61.10.77
Ours59.60.7445.461.90.80
", + "bbox": [ + 509, + 465, + 883, + 534 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "facial landmarks from outputs of all methods, frontalized, and normalized to position the mouth edges at $(-1,0)$ and $(1,0)$ . We compute mean absolute error (MAE) for both predicted mouth landmark positions $(M_P)$ and velocities $(M_V)$ , along with errors in facial geometry $(F_P)$ and velocity $(F_V)$ . For our evaluation, we use 100 randomly selected video sequences from the HDTF dataset [37], each up to 30 seconds. As the results show in Tab. 5, our speech-driven model performs on par with leading methods, excelling in realism and facial dynamics. SyncNet [6] is used to assess lip sync quality, providing offset (temporal misalignment between the audio and video) and confidence (certainty of audio-visual alignment) scores.", + "bbox": [ + 496, + 703, + 893, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "8504", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/78926baf7339f356df241bf8a88b63a9b83f4b6c3649de1e1ee938d86b29bfe5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 112, + 87, + 194, + 151 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/249eb01f4131fdddf201108d1aabb80575d66581bff4a918301a2418cf1ced01.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 194, + 88, + 276, + 151 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/e6833fe36b387c24e85728fd42f64f6c1ebddbb0976d1cc0ff663463ff80b708.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 277, + 88, + 359, + 151 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/babf75fa3d53bebbcf1c449f3409b684032610a97ed8f57c97e1cac8db6ce0b7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 362, + 88, + 441, + 151 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/69919e8e1c6efec27db2fa17703209ebafd9fb8767da1ce3158618272fad3ff7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 442, + 88, + 522, + 151 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/4e727456ce94a10c5cb0cf85dd00e0f157e05f4b06e02049f397c50584441bdd.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 524, + 88, + 606, + 151 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/40604b2aa996fc6d2acd23a2db2b3411973eea072538acf29f635c8bee37d1b6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 609, + 88, + 687, + 151 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/414e09e5425d5e35225c67a5ef070930a8035b17cd7948c4f042cab6758ae06a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 691, + 88, + 771, + 151 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/fed734f68f0698bb790e0edb0feec83072cb67a77fe1500b92938e93eb4d9317.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 772, + 88, + 854, + 151 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/3404ad765f09bb27d600e94b7dfd8336954bec7b7f517c9d336de8e7692620ea.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 112, + 157, + 192, + 220 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/5d483df94f00a8f21a401b78a622031777a572b55a0ad53f3c54126f06199994.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 194, + 157, + 276, + 220 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/2f5865bfe29388b182d03ed4d7a9d27dd59feab9fd7f2c0fb1094a9e3561bc51.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 277, + 157, + 359, + 220 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/514cacf7cd21b690fc389cf3ac74b73b37cf9698264c21cad61806b3e19a0ea3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 361, + 157, + 439, + 220 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/6d511aa5d5a396393d77990f87760f332acc794d63f47c209639a816d5ea6435.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 442, + 157, + 522, + 220 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/70d3e059bb62413c66fd955421186f77044e156d0e45ee6c2b8d47b983c07b88.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 524, + 157, + 606, + 220 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/18926e0a02b7ae432452adbefe702e8868b7a75fe3646dec0b6b5d9e908bf0d5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 609, + 157, + 687, + 220 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/b9020c9c3663e6a403c4af94a5a9fa47af6149c2161bc18ebf0b4b4174ee17e0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 689, + 157, + 771, + 220 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/18cada0a23f99cc471fbaa33df513dc3bdc55efddbbf3f720784a087bb0d7786.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 772, + 157, + 854, + 220 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/abdd874750024f19fafebca20a06b1f70aa928706d99df413ba354979d75ba5c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 112, + 229, + 192, + 292 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/8c349ae9a21dac090ccda0bff743883591aeeca9e5450b427eac42b730eaa9f6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 194, + 229, + 276, + 292 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/3c10c4d85aa3cca3c2b78f3e2c34e7a1a4bfbaa64af74b33787d9f6fe596d827.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 277, + 229, + 359, + 292 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/17e23469abdc5c7ab7a6a5550d1010eb601a5984b5074e13f879ce44ad308ecd.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 361, + 229, + 439, + 292 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/6a6831ddbb77f1d43713ce4c0cf619c551296f51f84ded434ab7b7241a9f5c62.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 442, + 229, + 522, + 292 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/fd0efa99b095cb5c8d77d101660f6c7bb18b5ee8d5d2d7cb16d3cd628893f63d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 524, + 229, + 606, + 292 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/73780beddb0fb11c3d6d1430fac4b29508bdfa7ada34735177e0e1aa0d8d7c4c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 609, + 229, + 687, + 292 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/be464290184a7120b5a3faa800bfa8fea8bed08c5546356fdabcb1efb9302981.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 689, + 229, + 771, + 292 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/455b641de64cefbaf69d4b86b1a3d711d7191a180fdb1111b6aad142ca9b34d4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 772, + 229, + 854, + 292 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/c20f63e03b90141305ab3f748fe6e81e7db5f7761224f2e498b225b5bd0fbdb8.jpg", + "image_caption": [ + "Source" + ], + "image_footnote": [], + "bbox": [ + 112, + 301, + 192, + 363 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/914f03cb4787c3219044fec13bdedab69d72544c0aaf586e94c11589a47cc118.jpg", + "image_caption": [ + "Driver" + ], + "image_footnote": [], + "bbox": [ + 194, + 301, + 276, + 363 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/2a4a11262eaefe3893566bd9f65ab0cf40f2dd90dbc5eb22abf893187532cf4e.jpg", + "image_caption": [ + "FOMM [22]" + ], + "image_footnote": [], + "bbox": [ + 277, + 301, + 357, + 363 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/167c2eae9f425be2922aad926fe782a8fae893bef8708cefbea97fb2b42f3c59.jpg", + "image_caption": [ + "UVA [17]" + ], + "image_footnote": [], + "bbox": [ + 359, + 301, + 439, + 363 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/8a024063cda079847e308c3b84df1cd3d92d6782b38b156413a97c18dec990a1.jpg", + "image_caption": [ + "StyleHEAT [31]" + ], + "image_footnote": [], + "bbox": [ + 442, + 301, + 522, + 363 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/ca68c0f5dd22b90c2d4067ccb3bef63902c3770d8b9a7a684fa690295510ba96.jpg", + "image_caption": [ + "MetaPortrait [35]", + "Figure 9. A qualitative comparison of head avatar systems in cross-reenactment scenario." + ], + "image_footnote": [], + "bbox": [ + 524, + 301, + 606, + 363 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/2f6050cbd5ae51b790b8714d4b1c4c863640bf93ba51abf7f465583f4c63ce65.jpg", + "image_caption": [ + "NOFA [32]" + ], + "image_footnote": [], + "bbox": [ + 609, + 301, + 687, + 363 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/a60b6865628d9d29031984af0b311f49aa111e0d10c11263007409dd4919132c.jpg", + "image_caption": [ + "MegaPortraits [9]" + ], + "image_footnote": [], + "bbox": [ + 691, + 301, + 771, + 363 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/5cd2ec3d01fca131da0568ab9ce41e09a269a4f1cdc110e312097c112488ba8f.jpg", + "image_caption": [ + "Ours" + ], + "image_footnote": [], + "bbox": [ + 772, + 301, + 854, + 363 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/234c49cf3f3f200a5accfefb0aaf8eed6b3d111522a888153bde86d510c0a6e8.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodFID↓\\( F_P \\downarrow \\)\\( F_V \\downarrow \\)\\( M_P \\downarrow \\)\\( M_V \\downarrow \\)Sync. offset ↓Sync. conf. ↑
StyleHEAT [31]62.41.340.412.981.670.417.22
PC-AVS [39]71.52.780.512.561.043.798.42
SadTalker [36]55.60.910.352.121.170.277.58
MakeIRTalk [40]63.32.150.383.111.592.985.17
Ours (EMOPortraits)28.50.820.332.531.380.075.77
", + "bbox": [ + 81, + 431, + 467, + 489 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/a33ea7fa3dd7e2f88d704e8bd1b0a93c9948c355710e8868de8a970281300f33.jpg", + "table_caption": [ + "Table 5. Our model matches other leading speech-driven models, exceeding in realism (FID), facial geometry and velocity, but is slightly less accurate in mouth landmark positions and velocities. We also score best in audio-video sync offset, but lower in confidence. We believe that while SyncNet metrics are still helpful, they may not fully represent true synchronization quality, as generated videos can sometimes score higher than ground truth without necessarily being better synchronized. In our experimentation, ground truth videos scored an offset of 1.13 (which is worse than our method, SadTalker, and StyleHEAT) and a confidence score of 8.35 (less than PC-AVS). For qualitative analysis, please refer to our supplementary materials." + ], + "table_footnote": [], + "table_body": "
MethodFID↓MP↓MV↓FP↓FV↓
Ours w/o disentanglement44.70.980.455.782.55
Ours w/o LPCA29.30.840.353.131.88
Ours with MAE(\\( \\mathbf{z}_i, \\hat{\\mathbf{z}}_i^{\\text{speech}} \\))28.20.840.342.611.45
Ours28.50.820.332.531.38
", + "bbox": [ + 81, + 705, + 467, + 773 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 6. Ablation study of our speech-driven mode, where in the first line, we show that it notably underperforms without head pose - face expression disentanglement of the latent space (see Sec. 5.1). Removing $\\mathcal{L}_{\\mathrm{PCA}}$ leads to a significant reduction in generating realistic mouth movements. Replacing $\\mathcal{L}_{\\mathrm{PCA}}$ with $\\mathrm{MAE}(\\mathbf{z}_i,\\hat{\\mathbf{z}}_i^{\\mathrm{speech}})$ causes a noticeable, decline in mouth movement generation. This suggests that identifying mouth-related PCA components was beneficial in enhancing the outcome.", + "bbox": [ + 75, + 785, + 470, + 896 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6.3. Ablation study", + "text_level": 1, + "bbox": [ + 500, + 426, + 650, + 441 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We conducted an extensive ablation study to evaluate the contributions of individual components within our method. For our main model, we present the evaluation of the importance of the proposed source-driver unmatched loss, canonical volume loss, and change of $dim(\\mathbf{z})$ from 512 in original model to 128 Tab. 4. For our speech-driven mode, we show how vital the disentanglement of the latent space for speech-drive ability, how $\\mathcal{L}_{\\mathrm{PCA}}$ is better than naive matching of PCA components from random $\\mathcal{Z}$ and how it helps match speech and latent expression space Tab. 6. For more ablation analysis, plots and visual comparison, please refer to our supplementary.", + "bbox": [ + 496, + 449, + 895, + 632 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "7. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 647, + 619, + 662 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we introduce EmoPortraits, a novel method for creating neural avatars with superior performance in image-driven, cross-identity emotion translation. Our speech-driven mode makes it possible to drive the facial animation through multiple conditions (video, audio, head motion). We collected FEED dataset which, we believe, will be a valuable asset for researchers in diverse human-centered studies.", + "bbox": [ + 496, + 672, + 893, + 777 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "However, our method has some limitations. It doesn't generate the avatar's body or shoulders, limiting some use cases. We currently integrate our output with a source image body. Additionally, the model sometimes struggles with accurate expression translation and underperforms with extensive head rotation. These challenges are crucial for future enhancements and remain central to our ongoing research efforts.", + "bbox": [ + 496, + 779, + 895, + 898 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8505", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Emad Barsoum, Cha Zhang, Cristian Canton Ferrer, and Zhengyou Zhang. Training deep networks for facial expression recognition with crowd-sourced label distribution. In Proceedings of the 18th ACM international conference on multimodal interaction, pages 279-283, 2016. 3", + "[2] Stella Bounareli, Christos Tzelepis, Vasileios Argyriou, Ioannis Patras, and Georgios Tzimiropoulos. Hyperreenact: One-shot reenactment via jointly learning to refine and retarget faces, 2023. 2", + "[3] Egor Burkov, Igor Pasechnik, Artur Grigorev, and Victor Lempitsky. Neural head reenactment with latent pose descriptors. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). IEEE, 2020. 2", + "[4] Houwei Cao, David G Cooper, Michael K Keutmann, Ruben C Gur, Ani Nenkova, and Ragini Verma. Crema-d: Crowd-sourced emotional multimodal actors dataset. IEEE transactions on affective computing, 5(4):377-390, 2014. 3", + "[5] Qiong Cao, Li Shen, Weidi Xie, Omkar M. Parkhi, and Andrew Zisserman. Vggface2: A dataset for recognising faces across pose and age, 2018. 7", + "[6] Joon Son Chung and Andrew Zisserman. Out of time: automated lip sync in the wild. In Computer Vision-ACCV 2016 Workshops: ACCV 2016 International Workshops, Taipei, Taiwan, November 20-24, 2016, Revised Selected Papers, Part II 13, pages 251-263. Springer, 2017. 7", + "[7] Joon Son Chung, Arsha Nagrani, and Andrew Zisserman. Voxceleb2: Deep speaker recognition. In INTERSPEECH, 2018. 4", + "[8] Michail Christos Doukas, Stefanos Zafeiriou, and Viktoriia Sharmanska. Headgan: One-shot neural head synthesis and editing, 2021. 2", + "[9] Nikita Drobyshev, Jenya Chelishev, Taras Khakhulin, Aleksei Ivakhnenko, Victor Lempitsky, and Egor Zakharov. Megaportraits: One-shot megapixel neural head avatars, 2023. 2, 4, 5, 7, 8", + "[10] Guy Gafni, Justus Thies, Michael Zollhöfer, and Matthias Nießner. Dynamic neural radiance fields for monocular 4d facial avatar reconstruction, 2020. 2", + "[11] Philip-William Grassal, Malte Prinzler, Titus Leistner, Carsten Rother, Matthias Nießner, and Justus Thies. Neural head avatars from monocular rgb videos, 2022. 2", + "[12] Siddharth Gururani, Arun Mallya, Ting-Chun Wang, Rafael Valle, and Ming-Yu Liu. Space: Speech-driven portrait animation with controllable expression, 2022. 3, 7", + "[13] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. In Advances in Neural Information Processing Systems, 2017. 7", + "[14] Philip Jackson and SJUoSG Haq. Surrey audio-visual expressed emotion (savee) database. University of Surrey: Guildford, UK, 2014. 3", + "[15] Taras Khakhulin, Vanessa Sklyarova, Victor Lempitsky, and Egor Zakharov. Realistic one-shot mesh-based head avatars, 2022. 2" + ], + "bbox": [ + 78, + 114, + 470, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[16] Alexander C. Li, Alexei A. Efros, and Deepak Pathak. Understanding collapse in non-contrastive siamese representation learning, 2022. 4", + "[17] Xueting Li, Shalini De Mello, Sifei Liu, Koki Nagano, Umar Iqbal, and Jan Kautz. Generalizable one-shot neural head avatar, 2023. 7, 8", + "[18] Steven R Livingstone and Frank A Russo. The ryerson audiovisual database of emotional speech and song (ravdess): A dynamic, multimodal set of facial and vocal expressions in north american english. PloS one, 13(5):e0196391, 2018. 3", + "[19] Patrick Lucey, Jeffrey F Cohn, Takeo Kanade, Jason Saragih, Zara Ambadar, and Iain Matthews. The extended cohn-kanade dataset (ck+): A complete dataset for action unit and emotion-specified expression. In 2010 iiee computer society conference on computer vision and pattern recognition-workshops, pages 94-101. IEEE, 2010. 3", + "[20] Maja Pantic, Michel Valstar, Ron Rademaker, and Ludo Maat. Web-based database for facial expression analysis. In 2005 IEEE international conference on multimedia and Expo, pages 5-pp. IEEE, 2005. 3", + "[21] KR Prajwal, Rudrabha Mukhopadhyay, Vinay P Namboodiri, and CV Jawahar. A lip sync expert is all you need for speech to lip generation in the wild. In Proceedings of the 28th ACM international conference on multimedia, pages 484-492, 2020. 3", + "[22] Aliaksandr Siarohin, Stéphane Lathuilière, S. Tulyakov, Elisa Ricci, and N. Sebe. First order motion model for image animation. ArXiv, abs/2003.00196, 2019. 2, 7, 8", + "[23] Aliaksandr Siarohin, Willi Menapace, Ivan Skorokhodov, Kyle Olszewski, Hsin-Ying Lee, Jian Ren, Mengelei Chai, and Sergey Tulyakov. Unsupervised volumetric animation. arXiv preprint arXiv:2301.11326, 2023. 7", + "[24] Michal Stypulkowski, Konstantinos Vougioukas, Sen He, Maciej Zieba, Stavros Petridis, and Maja Pantic. Diffused heads: Diffusion models beat gans on talking-face generation, 2023. 3", + "[25] Mani Kumar Tellamekala, Ömer Ümer, Björn W. Schuller, Elisabeth Andre, Timo Giesbrecht, and Michel Valstar. Are 3d face shapes expressive enough for recognising continuous emotions and action unit intensities?, 2023. 2", + "[26] Hao Wang, Yitong Wang, Zheng Zhou, Xing Ji, Dihong Gong, Jingchao Zhou, Zhifeng Li, and Wei Liu. Cosface: Large margin cosine loss for deep face recognition, 2018. 5", + "[27] H. Wang, Yitong Wang, Zheng Zhou, Xing Ji, Zhifeng Li, Dihong Gong, Jin Zhou, and Wenyu Liu. Cosface: Large margin cosine loss for deep face recognition. 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5265-5274, 2018. 6", + "[28] Kaisiyuan Wang, Qianyi Wu, Linsen Song, Zhuoqian Yang, Wayne Wu, Chen Qian, Ran He, Yu Qiao, and Chen Change Loy. Mead: A large-scale audio-visual dataset for emotional talking-face generation. In European Conference on Computer Vision, pages 700–717. Springer, 2020. 3", + "[29] Ting-Chun Wang, Arun Mallya, and Ming-Yu Liu. One-shot free-view neural talking-head synthesis for video conferencing. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2021. 2" + ], + "bbox": [ + 501, + 92, + 893, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "8506", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[30] Yuelang Xu, Lizhen Wang, Xiaochen Zhao, Hongwen Zhang, and Yebin Liu. AvatarMAV: Fast 3d head avatar reconstruction using motion-aware neural voxels. In *Special Interest Group on Computer Graphics and Interactive Techniques Conference Conference Proceedings*. ACM, 2023. 2", + "[31] Fei Yin, Yong Zhang, Xiaodong Cun, Mingdeng Cao, Yanbo Fan, Xuan Wang, Qingyan Bai, Baoyuan Wu, Jue Wang, and Yujiu Yang. Styleheat: One-shot high-resolution editable talking face generation via pre-trained stylegan, 2022. 2, 3, 7, 8", + "[32] Wangbo Yu, Yanbo Fan, Yong Zhang, Xuan Wang, Fei Yin, Yunpeng Bai, Yan-Pei Cao, Ying Shan, Yang Wu, Zhongqian Sun, and Baoyuan Wu. Nofa: Nerf-based one-shot facial avatar reconstruction, 2023. 2, 7, 8", + "[33] Egor Zakharov, Aliaksandra Shysheya, Egor Burkov, and Victor Lempitsky. Few-shot adversarial learning of realistic neural talking head models, 2019. 2", + "[34] Egor Zakharov, Aleksei Ivakhnenko, Aliaksandra Shysheya, and Victor Lempitsky. Fast bi-layer neural synthesis of one-shot realistic head avatars, 2020. 2", + "[35] Bowen Zhang, Chenyang Qi, Pan Zhang, Bo Zhang, Hsiang-Tao Wu, Dong Chen, Qifeng Chen, Yong Wang, and Fang Wen. Metaportrait: Identity-preserving talking head generation with fast personalized adaptation, 2023. 2, 7, 8", + "[36] Wenxuan Zhang, Xiaodong Cun, Xuan Wang, Yong Zhang, Xi Shen, Yu Guo, Ying Shan, and Fei Wang. Sadtalker: Learning realistic 3d motion coefficients for stylized audio-driven single image talking face animation, 2023. 3, 7, 8", + "[37] Zhimeng Zhang, Lincheng Li, Yu Ding, and Changjie Fan. Flow-guided one-shot talking face generation with a high-resolution audio-visual dataset. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3661–3670, 2021. 7", + "[38] Yufeng Zheng, Victoria Fernández Abrevaya, Marcel C. Bühler, Xu Chen, Michael J. Black, and Otmar Hilliges. I m avatar: Implicit morphable head avatars from videos, 2022. 2", + "[39] Hang Zhou, Yasheng Sun, Wayne Wu, Chen Change Loy, Xiaogang Wang, and Ziwei Liu. Pose-controllable talking face generation by implicitly modularized audio-visual representation, 2021. 2, 3, 7, 8", + "[40] Yang Zhou, Xintong Han, Eli Shechtman, Jose Echevarria, Evangelos Kalogerakis, and Dingzeyu Li. MakeltTalk. ACM Transactions on Graphics, 39(6):1-15, 2020. 3, 7, 8" + ], + "bbox": [ + 78, + 90, + 470, + 699 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "8507", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2024/EMOPortraits_ Emotion-enhanced Multimodal One-shot Head Avatars/ab68d1cd-73a3-4a4c-9896-716aeaaff436_model.json b/2024/EMOPortraits_ Emotion-enhanced Multimodal One-shot Head Avatars/ab68d1cd-73a3-4a4c-9896-716aeaaff436_model.json new file mode 100644 index 0000000000000000000000000000000000000000..2eef6387b73684545a4d9ef570afe35afac65b7d --- /dev/null +++ b/2024/EMOPortraits_ Emotion-enhanced Multimodal One-shot Head Avatars/ab68d1cd-73a3-4a4c-9896-716aeaaff436_model.json @@ -0,0 +1,2651 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.127, + 0.131, + 0.845, + 0.151 + ], + "angle": 0, + "content": "EMOPortraits: Emotion-enhanced Multimodal One-shot Head Avatars" + }, + { + "type": "text", + "bbox": [ + 0.088, + 0.181, + 0.232, + 0.2 + ], + "angle": 0, + "content": "Nikita Drobyshev" + }, + { + "type": "text", + "bbox": [ + 0.272, + 0.182, + 0.488, + 0.216 + ], + "angle": 0, + "content": "Antoni Bigata Casademunt \nStavros Petridis" + }, + { + "type": "text", + "bbox": [ + 0.52, + 0.182, + 0.73, + 0.217 + ], + "angle": 0, + "content": "Konstantinos Vougioukas \nMaja Pantic" + }, + { + "type": "text", + "bbox": [ + 0.77, + 0.182, + 0.882, + 0.2 + ], + "angle": 0, + "content": "Zoe Landgraf" + }, + { + "type": "text", + "bbox": [ + 0.221, + 0.22, + 0.752, + 0.271 + ], + "angle": 0, + "content": "{n.drobyshev, a/bigata-casademunt22, k.vougioukas, zoe.landgraf15, stavros.petridis04, m.pantic}@imperial.ac.uk Imperial College London" + }, + { + "type": "image", + "bbox": [ + 0.123, + 0.305, + 0.272, + 0.417 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.271, + 0.306, + 0.416, + 0.417 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.416, + 0.306, + 0.562, + 0.417 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.562, + 0.306, + 0.705, + 0.417 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.705, + 0.306, + 0.845, + 0.417 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.124, + 0.421, + 0.271, + 0.528 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.27, + 0.421, + 0.415, + 0.528 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.415, + 0.421, + 0.562, + 0.528 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.562, + 0.421, + 0.7, + 0.528 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.7, + 0.421, + 0.845, + 0.528 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.124, + 0.531, + 0.271, + 0.643 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.173, + 0.644, + 0.219, + 0.655 + ], + "angle": 0, + "content": "Source" + }, + { + "type": "image", + "bbox": [ + 0.27, + 0.531, + 0.415, + 0.643 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.333, + 0.644, + 0.372, + 0.654 + ], + "angle": 0, + "content": "Driver" + }, + { + "type": "image", + "bbox": [ + 0.415, + 0.531, + 0.562, + 0.643 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.467, + 0.644, + 0.51, + 0.655 + ], + "angle": 0, + "content": "Output" + }, + { + "type": "image", + "bbox": [ + 0.562, + 0.531, + 0.7, + 0.643 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.614, + 0.644, + 0.654, + 0.655 + ], + "angle": 0, + "content": "Driver" + }, + { + "type": "image", + "bbox": [ + 0.7, + 0.531, + 0.846, + 0.643 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.753, + 0.644, + 0.797, + 0.655 + ], + "angle": 0, + "content": "Output" + }, + { + "type": "image_caption", + "bbox": [ + 0.26, + 0.665, + 0.709, + 0.68 + ], + "angle": 0, + "content": "Figure 1. Selected animation results for image-driven mode of our method." + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.693, + 0.314, + 0.708 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.72, + 0.473, + 0.901 + ], + "angle": 0, + "content": "Head avatars animated by visual signals have gained popularity, particularly in cross-driving synthesis where the driver differs from the animated character, a challenging but highly practical approach. The recently presented MegaPortraits model has demonstrated state-of-the-art results in this domain. We conduct a deep examination and evaluation of this model, with a particular focus on its latent space for facial expression descriptors, and uncover several limitations with its ability to express intense face motions. To address these limitations, we propose substantial changes in both training pipeline and model architecture, to introduce our EMOPortraits model, where we:" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.713, + 0.895, + 0.773 + ], + "angle": 0, + "content": "Enhance the model's capability to faithfully support intense, asymmetric face expressions, setting a new state-of-the-art result in the emotion transfer task, surpassing previous methods in both metrics and quality." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.777, + 0.895, + 0.838 + ], + "angle": 0, + "content": "Incorporate speech-driven mode to our model, achieving top-tier performance in audio-driven facial animation, making it possible to drive source identity through diverse modalities, including visual signal, audio, or a blend of both." + }, + { + "type": "text", + "bbox": [ + 0.497, + 0.841, + 0.895, + 0.901 + ], + "angle": 0, + "content": "Furthermore, we propose a novel multi-view video dataset featuring a wide range of intense and asymmetric facial expressions, filling the gap with absence of such data in existing datasets." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8498" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.079, + 0.091, + 0.208, + 0.105 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "title", + "bbox": [ + 0.079, + 0.115, + 0.38, + 0.132 + ], + "angle": 0, + "content": "1.1. Representation of facial expression" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.139, + 0.47, + 0.319 + ], + "angle": 0, + "content": "Advancements in neural head technologies now enable the creation of realistic avatars from a few or even one image, with the latter being crucial when only one source image is available. Cross-driving synthesis, where avatars are animated with different identities is a key technique for virtual reality, filmmaking, photo animation, etc. However, accurate transfer of facial expressions, especially those that are intense and uneven, remains a substantial challenge in avatar animation, particularly for cross-driving synthesis. While previous research focused mainly on preserving identity and transferring moderate facial motions, our work also seeks to accurately drive high-intensity and asymmetric expressions." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.321, + 0.471, + 0.546 + ], + "angle": 0, + "content": "Our research is built on and extends the MegaPortraits [9] model, notable for state-of-the-art results in cross-driving synthesis. An additional advantage using this method is that, unlike many avatar systems that depend on limited [3, 25] predefined motion descriptors, MegaPortraits, learns expression representations from scratch, allowing for greater adaptability to a wider range of expressions. We investigate the latent expression spaces and training methods of MegaPortraits to enhance its ability to depict a broad spectrum of facial expressions. Our comprehensive analysis reveals that, while the original model shows limited effectiveness in intense motion representing, it has significant potential for improvement through targeted architectural modifications, adjustments in the training approach, and the integration of our novel dataset." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.557, + 0.316, + 0.573 + ], + "angle": 0, + "content": "1.2. Integrating speech driving" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.581, + 0.469, + 0.64 + ], + "angle": 0, + "content": "We integrate speech in our model to be used either as a complementary or an alternative driver, a crucial aspect for applications like virtual assistance or mixed reality when a primary visual signal is absent." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.642, + 0.47, + 0.777 + ], + "angle": 0, + "content": "By improving disentanglement of facial expression latent space, we remove disturbing information from face motion descriptors and emphasize components that solely control lip movements. Based on this, we formulated a novel loss that helps to achieve desirable results. Furthermore, our method can generate plausible head rotations and blinks, which appear natural and enhance its applicability across various tasks. Thus, our final method can drive source identity through image, audio, or a mix of them." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.787, + 0.428, + 0.803 + ], + "angle": 0, + "content": "1.3. FEED: Facial Extreme Emotions Dataset" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.811, + 0.47, + 0.9 + ], + "angle": 0, + "content": "Public data scarcity is a common obstacle in deep learning, the topic of human facial expressions and movements is not an exception. In particular, we note a lack of high quality video data capturing a wide range of facial expressions, with existing datasets not going beyond basic facial actions shown at Fig. 2. To address this, we collect a new dataset that" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.212 + ], + "angle": 0, + "content": "includes basic expressions and also captures complex movements like blinks, winks, and head and tongue movements, along with varied extreme expressions which are difficult or impossible to categorize through basic facial actions. Given this variety in expressions in our dataset, we believe that it will be a valuable resource for research in human emotions and facial reconstruction fields. In summary, our main contributions are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.224, + 0.892, + 0.329 + ], + "angle": 0, + "content": "- We introduce our new EMOPortraits model for one-shot head avatars synthesis, that is capable of transferring intense facial expression, showing state-of-the-art results in cross-driving synthesis. We achieve this through careful latent facial expression space development as well as novel losses and a minimal amount of domain-specific data (\\(\\sim\\) 0.1% of the train set)." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.33, + 0.892, + 0.405 + ], + "angle": 0, + "content": "- We integrate a speech-driving mode in our model, that demonstrates cutting-edge results in speech-driven animations. It functions effectively alongside visual signals or independently, also generating realistic head rotations and eye blinks." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.405, + 0.892, + 0.45 + ], + "angle": 0, + "content": "- We present a unique multi-view dataset that spans a broad spectrum of extreme facial expressions, filling the gap of absence of such data in existing datasets." + }, + { + "type": "list", + "bbox": [ + 0.5, + 0.224, + 0.892, + 0.45 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.477, + 0.641, + 0.492 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.507, + 0.742, + 0.521 + ], + "angle": 0, + "content": "2.1. Visual-driven head avatars" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.534, + 0.895, + 0.67 + ], + "angle": 0, + "content": "In recent years, the domain of neural avatars has branched into two prominent sub-fields: person-specific and person-agnostic avatars. While person-specific methodologies [10, 11, 30, 38, 39] excel in delivering stunning realism and motion fidelity for a particular individual, they encounter challenges representing an arbitrary person. Moreover, these approach demands multiple frames of training data, distinct training for each avatar, and can struggle to replicate motions not encountered during training." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.675, + 0.895, + 0.901 + ], + "angle": 0, + "content": "Person-agnostic methods, on the other hand, don't need train or fine-tuning for each new person and present an alternative approach to talking-head synthesis. Earlier works in this domain generated avatars in a few-shot technique [3, 33], while subsequent studies introduced one-shot capabilities [2, 8, 9, 22, 29, 31, 32, 34]. Many of these works employ predefined motion representations, such as 3DMM's blendshapes [8, 15, 31, 32]. In contrast, some learned latent emotion representation from scratch [3, 9, 22, 34, 35]. This latter approach holds potential for better motion representation as in such settings, expression descriptors become entirely trainable and free from inheriting the limitations typical for blendshapes as was shown in [3, 25]. In our work, we adopt this strategy, remarkably improving upon the methodology used in MegaPortraits[9]." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8499" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.091, + 0.089, + 0.458, + 0.128 + ], + "angle": 0, + "content": "
MethodBFEsFFED dataset internal structure
Head rotationsWinksEyes moveAsym. em.Tongue em.Extreme em.
ModeMild/StrongAxes/randomMild/StrongRandomStrongStrongExtreme
# of part.21/23212321232323
Avg. length1:38/1.520:580:490:331:411:533:02
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.137, + 0.47, + 0.179 + ], + "angle": 0, + "content": "Table 1. The tasks we asked our participants to perform for our FEED dataset. Here BFEs stands for basic facial expressions, shown in Fig. 2" + }, + { + "type": "table", + "bbox": [ + 0.09, + 0.196, + 0.458, + 0.258 + ], + "angle": 0, + "content": "
DatasetDatasets comparison
#Actors#ViewsBasic exp.Strong exp.Extrme & Assym. exp.Tongue exp.Resolution
SAVESS41✓✘✘✘1280 x 1024
RAVGESS241✓✘✘✘1920 x 1080
CREMA-D911✓✘✘✘1280 x 720
MMI251✓✓✘✘1920 x 1080
MEAD487✓✓✘✘1920 x 1080
Ours (FEED)233✓✓✓✓3840 x 2160
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.268, + 0.468, + 0.296 + ], + "angle": 0, + "content": "Table 2. Comparisons with modern, high-quality audio-visual datasets created in controlled settings." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.325, + 0.303, + 0.34 + ], + "angle": 0, + "content": "2.2. Face expression datasets" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.349, + 0.47, + 0.409 + ], + "angle": 0, + "content": "Early image datasets[1, 19] predominantly offered annotated face expression data corresponded to up to 8 basic emotions. However, these datasets are not well-suited for training head avatars, which necessitate video or multi-view data." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.41, + 0.471, + 0.622 + ], + "angle": 0, + "content": "The SAVEE [14] dataset captures facial expressions in speech, but involves just 4 actors. The MMI [20] dataset, more expansive in actor participation, offers spontaneous moderate-intensity expression but is restricted to single-view sequences. RAVDESS [18] distinguishes itself by capturing two motion intensities, but its limited recordings challenge broad applicability. CREMA-D [4] encompasses three expression intensities, and though MEAD [28] provides detailed multi-view data across three intensities, it's limited to the standard eight expression groups. Our novel FEED dataset aims to address the scientific community's demand for high-quality multi-view facial expression videos outside the standard categories: Joy, Fear, Sadness, Disgust, Anger Contempt and Surprise." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.635, + 0.327, + 0.65 + ], + "angle": 0, + "content": "2.3. Speech-driven head avatars" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.66, + 0.473, + 0.901 + ], + "angle": 0, + "content": "While numerous studies introduce speech-driven avatars [12, 21, 24, 31, 36, 39, 40], few excel in producing high-quality talking heads with authentic rotations, blinks, and the capability to use both visual and audio inputs. For instance, Wav2Lip [21] aims at re-dubbing videos with accurate lip motions, but often falls short in realism with using single images. Diffused heads [24] struggles to generate long sequences and doesn't provide access to pose control. MakeItTalk [40] animates facial landmarks in a speaker-specific manner, yet struggles with head pose controls due to its non-reliance on 3D. PC-AVS [39] offers a solution, but demands a driving video for pose modulation. Newly developed models such as SadTalker [36] and StyleHEAT [31] have shown promising outcomes in generating high-quality talking heads. We conducted a quantitative evaluation of our audio-driven mode compared to the aforementioned models." + }, + { + "type": "image", + "bbox": [ + 0.53, + 0.088, + 0.868, + 0.194 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.528, + 0.197, + 0.868, + 0.294 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.526, + 0.295, + 0.868, + 0.371 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.381, + 0.892, + 0.409 + ], + "angle": 0, + "content": "Figure 2. Illustration of the problem in publicly available face expression data and selected examples from our FEED dataset." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.442, + 0.642, + 0.457 + ], + "angle": 0, + "content": "3. FEED dataset" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.471, + 0.895, + 0.668 + ], + "angle": 0, + "content": "As previously mentioned, the current publicly available datasets with human face expression videos fall short in capturing a broad range of facial manipulations (see illustration at Fig. 2). In our view, this gap could stem from the complexity of designing clear tasks for participants, coupled with the reality that not everyone is ready for performing extreme expressions on camera. To bridge this gap, our pioneering FEED dataset is designed to meet the scientific community's need for high-quality, multi-view emotion videos that extend beyond standard emotions. It contains various expressions, including strong asymmetric ones, tongue and cheeks movements, winks, head rotations, eye movements and more nuanced gestures. ;" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.671, + 0.895, + 0.807 + ], + "angle": 0, + "content": "Our dataset consists of 520 multi-view videos of 23 subjects, captured with 3 cameras. As extreme face motions are complex and their perception can be heavily influenced by subtle differences, we use a high resolution of 4k for all video, capturing the whole face up to the level of individual hair strands and wrinkles, as shown in Fig. 2. For more examples, please refer to the supplementary materials. Our participants were asked to perform 7 tasks (see Tab. 1) to cover as many facial and head movements as possible." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.811, + 0.895, + 0.9 + ], + "angle": 0, + "content": "In our comparison with other expression datasets detailed in Tab. 2, the MEAD dataset [28] emerges as the closest to ours regarding viewpoints and maximal expression intensity. However, despite MEAD's larger number of actors and viewpoints, it offers significantly less variety of facial expressions and have lower resolution than our dataset." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8500" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.077, + 0.09, + 0.31, + 0.108 + ], + "angle": 0, + "content": "4. Expression enhancement" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.115, + 0.474, + 0.313 + ], + "angle": 0, + "content": "We found that the MegaPortraits [9] model fails to transfer intense expressions correctly, as shown in Fig. 9. Using our FEED dataset of strong and uneven facial expression to fine-tune a pre-trained MegaPortraits model didn't enhance final results as shown at Tab. 4. Training from scratch on FEED leads to fast overfitting due to the small number of identities presented in the dataset compared to VoxCeleb2 [7] used in [9] (23 vs \\(\\sim 5000\\)). Our method effectively injects this small dataset into training, yielding desirable results. We list our key findings here, for all details, see supplementary materials. Our model's scheme, displayed in Fig. 4, shares similarities with [9], leading us to occasionally refer to certain elements of our scheme in the context of [9]'s pipeline." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.322, + 0.296, + 0.339 + ], + "angle": 0, + "content": "4.1. Latent expression space" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.345, + 0.471, + 0.451 + ], + "angle": 0, + "content": "We begin by exploring MegaPortraits' latent expression descriptors' space \\((\\mathbf{z}_{s / d}\\) in Fig. 4), which is crucial for expression transfer, through PCA analysis. Inspired by [16], which demonstrated that the area under the cumulative explained variance curve of singular values (denoted here as \\(\\mathbf{AUC}_{\\mathbf{z}}\\)) can serve as an effective metric for dimensionality collapse in some latent spaces, we applied this measure to our study." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.452, + 0.474, + 0.649 + ], + "angle": 0, + "content": "This metric allows us to forecast model performance: greater collapse suggests reduced entropy in the latent space, which correlates with lower representational quality. High quality of the expression space, produced by \\(\\mathbf{E}_{\\text {motion }}\\) (as shown in Fig. 4), is crucial to capture and differentiate between the nuances of strong facial expressions. Our ablation study confirms (see Tab. 4) that there is a notable correlation between how broad and isotropic model's latent space is (expressed by \\(\\mathbf{AUC}_{\\mathbf{z}}\\)) and its final performance. Visual comparison is available in supplementary materials. As shown in Fig. 3, our model's latent space outperforms MegaPortraits in expression representation ability. This is supported by both visual (Fig. 9) and quantitative analysis (Tab. 3)." + }, + { + "type": "image", + "bbox": [ + 0.097, + 0.661, + 0.451, + 0.79 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.8, + 0.471, + 0.897 + ], + "angle": 0, + "content": "Figure 3. Comparison of latent spaces. Left plot shows that our model's latent space is wider and exhibits more even variance distribution. Also, as shown on the right plot, a greater number of principal components are involved in capturing variance across various thresholds. This implies a more robust representational capacity of expression space compared to [9]. The VoxCeleb2 test set was used for both plots." + }, + { + "type": "image", + "bbox": [ + 0.54, + 0.088, + 0.85, + 0.201 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.209, + 0.895, + 0.321 + ], + "angle": 0, + "content": "Figure 4. Method Overview. We use \\(\\mathbf{E}_{\\mathrm{app}}\\) to extract volume features \\(\\mathbf{V}_s\\) and a global descriptor \\(\\mathbf{e}_s\\) from the source image. Then \\(\\mathbf{E}_{\\mathrm{motion}}\\) or \\(\\mathbf{E}_{\\mathrm{audio}}\\) generates motion representations from source and driver, including head rotations \\(\\mathbf{R}_{s / d}\\), translations \\(\\mathbf{t}_{s / d}\\), and expression descriptors \\(\\mathbf{z}_{s / d}\\). Using them, we predict warpings \\(\\mathbf{w}_{s\\rightarrow}\\) and \\(\\mathbf{w}_{\\rightarrow d}\\). First warping and \\(\\mathbf{G}_{3D}\\) transform \\(\\mathbf{V}_s\\) into a canonical volume \\(\\mathbf{V}_s^C\\) by removing the source motions. Second warping and \\(\\mathbf{G}_{2D}\\) imposes the driver's motions and renders the final image." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.337, + 0.895, + 0.58 + ], + "angle": 0, + "content": "Both metrics shown in Fig. 3 were vital in our model's development since they highly correlate with the final model's ability to transfer intense expressions. This was particularly useful for finding adjustments described in Sec. 4.2 and Sec. 4.3, as early visual results during training are not very telling, but changes in the latent space are noticeable after a few epochs. We also found that in MegaPortraits, just a few principal components significantly affect variance, with only 18 components making up \\(99\\%\\) (Fig. 3), suggesting their \\(512\\mathbf{z}\\)'s dimension is oversized. For our model, we found a 128-dimensional \\(\\mathbf{z}\\) is optimal, offering solid performance and reducing overfitting in imbalanced data, as our ablation study reveals. We calculate the Explained Variance \\((\\mathrm{EV}_i)\\) for the \\(i^{th}\\) component and \\((\\mathbf{AUC}_{\\mathbf{z}})\\) using equations shown in Eq. (1), which involves standardizing the vector set \\(\\mathcal{Z}\\) for PCA and sorting the eigenvalues \\(\\lambda\\) in descending order." + }, + { + "type": "equation", + "bbox": [ + 0.508, + 0.594, + 0.894, + 0.634 + ], + "angle": 0, + "content": "\\[\n\\mathbf {Z} _ {\\mathrm {s t d}} = \\frac {\\mathcal {Z} - \\bar {\\mathcal {Z}}}{\\sigma (\\bar {\\mathcal {Z}})}, \\quad \\boldsymbol {\\lambda} = \\operatorname {s o r t e d} _ {\\mathrm {d e s c}} \\left(\\operatorname {e i g} \\left(\\frac {1}{n - 1} \\mathbf {Z} _ {\\mathrm {s t d}} ^ {\\top} \\mathbf {Z} _ {\\mathrm {s t d}}\\right)\\right), \\tag {1}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.508, + 0.628, + 0.821, + 0.663 + ], + "angle": 0, + "content": "\\[\n\\mathrm {E V} _ {i} = \\frac {\\lambda_ {i}}{\\sum_ {j = 1} ^ {p} \\lambda_ {j}}, \\quad \\mathbf {A U C _ {z}} = \\frac {1}{d} \\frac {\\sum_ {i = 1} ^ {d} \\sum_ {k = 1} ^ {i} \\lambda_ {k}}{\\sum_ {j = 1} ^ {d} \\lambda_ {j}},\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.667, + 0.674, + 0.68 + ], + "angle": 0, + "content": "4.2. Canonical volume" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.689, + 0.895, + 0.901 + ], + "angle": 0, + "content": "Despite being designed to exclude face expression details, we found that canonical volume \\((\\mathbf{V}^C\\) in Fig. 4) in MegaPortraits actually retains significant expression information from the source image, contributing to poor translation of intense expressions, as shown in our experiment detailed in Fig. 5. In this experiment, using portraits with varied expressions, we assessed the expression translation accuracy and how expression intensity affects the \\(\\mathbf{V}^C\\)s. Our findings suggest that the canonical volumes in MegaPortraits are not truly neutral. As confirmed by our ablation study (Tab. 4), creating an expression-free \\(\\mathbf{V}^C\\) is essential for precise translation of intense motions. We believe that a more neutral canonical volume improves tractability and effectiveness in expression translation tasks." + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "8501" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.091, + 0.089, + 0.466, + 0.212 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.223, + 0.473, + 0.432 + ], + "angle": 0, + "content": "Figure 5. Canonical volumes \\((\\mathbf{V}^C)\\) in MegaPortraits [9] are not expression-neutral. To show it we use three portraits: A (regular expression), B (intense expression), D (regular expression, new identity). Case 1 visualizes poor results in transferring B's intense expression to A using a self-driving mode, contrasting with the effective reconstruction of B when used as both driver and source in case 3. This discrepancy, indicative of expression leakage into the canonical volume, is further quantified by a \\(43\\%\\) relative difference in \\(\\mathbf{V}^C\\)s between B and C, contrary to expectations of their similarity for same identity. In cross-driving generation case 2, using B as the driver and D as the source yielded poor results. However, in case 4, after using additive operations on source canonical volumes, the output expression is much closer to driver than in case 2. This manipulation again confirms the significant retention of expression information in canonical volumes." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.449, + 0.472, + 0.496 + ], + "angle": 0, + "content": "To overcome this issue, we propose to match canonical volumes from the different images of the same person \\((\\mathbf{V}_{s^n}^C,\\mathbf{V}_{d^n}^C)\\) during training as following:" + }, + { + "type": "equation", + "bbox": [ + 0.187, + 0.506, + 0.47, + 0.527 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {C V}} ^ {\\mathrm {n}} = \\mathcal {L} _ {\\mathrm {M A E}} \\left(\\mathbf {V} _ {s ^ {n}} ^ {C}, \\mathbf {V} _ {d ^ {n}} ^ {C}\\right), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.53, + 0.473, + 0.577 + ], + "angle": 0, + "content": "This loss ensures \\(\\mathbf{V}^C\\) remains stable and expression-independent, crucial for translating intense expressions, as shown in our ablation study (Tab. 4)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.583, + 0.333, + 0.598 + ], + "angle": 0, + "content": "4.3. Source-driver mismatch loss" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.605, + 0.471, + 0.728 + ], + "angle": 0, + "content": "In addition to maintaining expression-free \\(\\mathbf{V}^{C}\\), it's vital to remove all identity information from the expression vector \\(\\mathbf{z}\\). While contrastive losses address this in [9], our experiments indicate they are insufficient to prevent overfitting in our imbalanced data scenario, where emotionally intense images represent only \\(0.1\\%\\) of our training set but are sampled \\(25\\%\\) of the time. We introduce a novel self-supervised loss which mitigates identity information in latent expression vectors:" + }, + { + "type": "equation", + "bbox": [ + 0.086, + 0.746, + 0.47, + 0.806 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} _ {s d m} (\\mathbf {z} _ {s}, \\mathbf {z} _ {d}) = w * \\max (0, \\cos (\\mathbf {z} _ {s}, \\mathbf {z} _ {d}) - m a r g i n), \\\\ w; \\text {m a r g i n} = \\left\\{ \\begin{array}{l l} 1; 0. 5 & \\text {i f} \\mathbf {z} _ {s}, \\mathbf {z} _ {d} \\text {a r e f r o m V C 2} \\\\ 1 0; 0. 2 5 & \\text {i f} \\mathbf {z} _ {s}, \\mathbf {z} _ {d} \\text {a r e f r o m F E E D} \\end{array} \\right. \\tag {3} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.811, + 0.473, + 0.903 + ], + "angle": 0, + "content": "Here, \\(\\mathbf{z}_s, \\mathbf{z}_d\\) are emotion vectors from source and driver images, \\(w, \\text{margin}\\) adjust the loss's intensity and strictness respectively. We increase \\(w\\) for FEED due to its higher sampling rate and decrease margin for more assured expression variance in \\(s, d\\). Refer to Fig. 6 for a visual summary of all self-supervised losses enhancing our latent emotion space." + }, + { + "type": "image", + "bbox": [ + 0.541, + 0.088, + 0.855, + 0.337 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.345, + 0.895, + 0.402 + ], + "angle": 0, + "content": "Figure 6. The visualisation of our self-supervised losses. \\(\\mathcal{L}_{\\mathrm{cos}}\\) is a contrastive loss, similar to one used in [9] but uses different pairs as described in Sec. 5.1. Our novel \\(\\mathcal{L}_{sdm}\\) is designed to prevent overfitting in an imbalanced dataset." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.427, + 0.821, + 0.446 + ], + "angle": 0, + "content": "5. Incorporating speech-driving mode" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.453, + 0.766, + 0.469 + ], + "angle": 0, + "content": "5.1. Latent space disentanglement" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.477, + 0.895, + 0.689 + ], + "angle": 0, + "content": "During our latent space analysis described in Sec. 4, we found that combining expression from one driver with the head pose from another resulted in poor performance for MegaPortraits model, as illustrated in Fig. 7 (bottom row). Indeed, the base model lacks a mechanism to intentionally prevent head pose leakage. Disentangled expression latent space, besides expanding the model's use cases, plays the crucial role for the speech-driving mode. When the latent vector \\(\\mathbf{z}\\) is entangled with head pose data, predicting it from speech becomes challenging because speech lacks head rotation information, unlike images. Moreover, we would like to have a full control over head rotations during speech mode. We were able to make our latent space disentangled by changing the way we sample images for \\(\\mathcal{L}_{\\mathrm{cos}}\\) Eq. (5)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.689, + 0.895, + 0.78 + ], + "angle": 0, + "content": "The MegaPortraits model uses \\(\\mathcal{L}_{\\mathrm{cos}}\\) - a modified Large Margin Cosine Loss [26] to mitigate the transfer of appearance features to expression descriptor (latent vector \\(\\mathbf{z}\\) on Fig. 4), which is crucial for cross-driving mode. Without this, differences in appearance between the source and driver, like hairstyle or skin tone, leak from driver to output image." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.781, + 0.896, + 0.902 + ], + "angle": 0, + "content": "For computing this loss, the authors utilize a supplementary source-driving pair \\((\\mathbf{x}_{s^{*}}\\) and \\(\\mathbf{x}_{d^{*}})\\) from a different video with another identity, ensuring a distinct appearance from the current \\(\\mathbf{x}_s,\\mathbf{x}_d\\) pair. The base model is then employed to produce cross-reenacted image \\((\\hat{\\mathbf{x}}_{s^{*}\\rightarrow d} = \\mathbf{G}_{\\mathrm{base}}(\\mathbf{x}_{s^{*}},\\mathbf{x}_d))\\) Concurrently, they determine a separate motion descriptor, \\(\\mathbf{z}_{d^{*}} = \\mathbf{E}_{\\mathrm{motion}}(\\mathbf{x}_{d^{*}})\\). Descriptors \\(\\mathbf{z}_{* \\rightarrow d}\\) from the respective forward passes are also used for positive and negative pairs." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8502" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.081, + 0.089, + 0.468, + 0.207 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.218, + 0.47, + 0.273 + ], + "angle": 0, + "content": "Figure 7. The illustration of disentanglement problem. Images in the right row indicate the expression descriptor \\(\\mathbf{z}\\) entangled with head pose for MegaPortraits, whereas we avoid it using changes proposed in Sec. 5.1." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.298, + 0.47, + 0.418 + ], + "angle": 0, + "content": "Our innovation here is a new sampling strategy: for each \\(\\mathbf{x}_s\\), \\(\\mathbf{x}_d\\) pair, we sample one more random additional pair \\(\\mathbf{x}_{s^m}\\) and \\(\\mathbf{x}_{d^m}\\) apart from \\(\\mathbf{x}_{s^*}\\) and \\(\\mathbf{x}_{d^*}\\). We then apply the model to produce the following cross-reenacted moved image using source identity from \\(\\mathbf{x}_{s^*}\\), desired head pose from \\(\\mathbf{x}_{s^m}\\) and emotions from \\(\\mathbf{x}_d\\), represented as: \\(\\hat{\\mathbf{x}}_{s^{*m}\\rightarrow d} = \\mathbf{G}_{\\mathrm{base}}(\\mathbf{x}_{s^{*m}},\\mathbf{x}_d)\\). This creates a positive pair with the same desired emotion but varying head poses." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.419, + 0.472, + 0.508 + ], + "angle": 0, + "content": "Motion descriptors are organized into positive pairs \\(\\mathcal{P}\\) for alignment, and negative pairs \\(\\mathcal{N}\\) for non-alignment: \\(\\mathcal{P} = \\left\\{(\\mathbf{z}_{s\\rightarrow d},\\mathbf{z}_d),(\\mathbf{z}_{s^*\\rightarrow d},\\mathbf{z}_d),(\\mathbf{z}_{s^{*m}\\rightarrow d},\\mathbf{z}_d)\\right\\}\\), and \\(\\mathcal{N} = \\left\\{(\\mathbf{z}_{s\\rightarrow d},\\mathbf{z}_{d^*}),(\\mathbf{z}_{s^*\\rightarrow d},\\mathbf{z}_{d^*}),(\\mathbf{z}_{s^{*m}\\rightarrow d},\\mathbf{z}_{d^*})\\right\\}\\) (see Fig. 6). These pairs are used to calculate the following cosine distance:" + }, + { + "type": "equation", + "bbox": [ + 0.17, + 0.51, + 0.47, + 0.527 + ], + "angle": 0, + "content": "\\[\nd \\left(\\mathbf {z} _ {i}, \\mathbf {z} _ {j}\\right) = s \\cdot \\left(\\left\\langle \\mathbf {z} _ {i}, \\mathbf {z} _ {j} \\right\\rangle - m\\right), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.531, + 0.47, + 0.577 + ], + "angle": 0, + "content": "where both \\(s\\) and \\(m\\) are hyperparameters. This distance is then used to calculate a large margin cosine loss (Cos-Face) [27]:" + }, + { + "type": "equation", + "bbox": [ + 0.088, + 0.596, + 0.47, + 0.638 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\cos} = - \\sum_ {(\\mathbf {z} _ {k}, \\mathbf {z} _ {l}) \\in \\mathcal {P}} \\log \\frac {\\exp \\left\\{d \\left(\\mathbf {z} _ {k} , \\mathbf {z} _ {l}\\right) \\right\\}}{\\exp \\left\\{d \\left(\\mathbf {z} _ {k} , \\mathbf {z} _ {l}\\right) \\right\\} + \\sum_ {(\\mathbf {z} _ {i} , \\mathbf {z} _ {j}) \\in \\mathcal {N}} \\exp \\left\\{d \\left(\\mathbf {z} _ {i} , \\mathbf {z} _ {j}\\right) \\right\\}} \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.645, + 0.469, + 0.674 + ], + "angle": 0, + "content": "This loss prevents the head pose from leaking into the embeddings as shown in Fig. 7 (upper row)" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.675, + 0.471, + 0.9 + ], + "angle": 0, + "content": "By doing this, we achieved the capability to interpret \\(\\mathbf{z}_i\\) as a latent vector encapsulating the emotional content from an image \\(\\mathbf{x}_i\\). This advancement opened the door for an intriguing possibility during inference: the prediction of this vector could potentially be derived from an audio signal instead of a facial image using the motion encoder \\(\\mathbf{E}_{\\mathrm{motion}}\\), effectively transforming the model into a speech-driven system. The critical challenge here is aligning the audio input with the pretrained latent space. We addressed this by introducing an additional audio encoder, \\(\\mathbf{E}_{\\mathrm{aud}}\\), and employing \\(\\mathbf{E}_{\\mathrm{motion}}\\) as a teacher model for this new encoder. A crucial factor in achieving remarkable accuracy in lip synchronization was our ability to isolate specific components within the expression latent space, that are solely responsible for mouth movements. We describe it in the next subsection Sec. 5.2." + }, + { + "type": "image", + "bbox": [ + 0.499, + 0.089, + 0.892, + 0.228 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.237, + 0.894, + 0.335 + ], + "angle": 0, + "content": "Figure 8. In the upper row, the first principal component from a random subset of expression vectors \\(\\mathcal{Z}\\) affects not just the mouth but also blinks and gaze, making it difficult to interpret and isolate mouth movement components, needed for training \\(\\mathbf{E}_{\\mathrm{aud}}\\). However, as shown in the bottom row, using \\(\\mathcal{Z}_m\\) from Sec. 5.2, the first principal component, \\(\\mathbf{v}_1\\), focused on mouth movements, is effectively isolated, \\(\\mathbf{z}_0\\) - latent vector, corresponded to the central image." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.357, + 0.68, + 0.372 + ], + "angle": 0, + "content": "5.2. Mouth movements" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.38, + 0.895, + 0.517 + ], + "angle": 0, + "content": "To train the audio encoder \\(\\mathbf{E}_{\\mathrm{aud}}\\) for predicting \\(\\mathbf{z}\\) from speech, we chose to use the motion encoder \\(\\mathbf{E}_{\\mathrm{motion}}\\) as a teacher model. However, direct matching vectors \\(\\mathbf{z}_i^{\\mathrm{aud}}\\) from \\(\\mathbf{E}_{\\mathrm{aud}}\\) and pseudo-ground truth \\(\\mathbf{z}_i\\) from \\(\\mathbf{E}_{\\mathrm{motion}}\\) gives poor results (see Tab. 6). We believe that the reason behind this is that speech-derived \\(\\mathbf{z}_i^{\\mathrm{aud}}\\), while able to capture lip movements, struggles with other facial motions, especially in the upper face. Thus, we focused on isolating mouth movement components in the expression latent space using PCA analysis." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.517, + 0.895, + 0.622 + ], + "angle": 0, + "content": "Employing PCA on a broad array of expression vectors \\(\\mathcal{Z}\\), generated from images with varied facial emotions, yields principal components with their explained variances \\((\\mathrm{EV}_i)\\). Altering these components, particularly those with high \\(\\mathrm{EV}_i\\), significantly changes facial expressions in an image. However, they typically affect a combination of facial features—Fig. 8 (upper row), not isolated areas like the mouth or eyes." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.623, + 0.894, + 0.773 + ], + "angle": 0, + "content": "To focus on mouth movements, we created a unique set of \\(\\mathcal{Z}_m\\) from just one video of a person performing mouth manipulations and manually edited it to minimize upper face movements. This involved using a still upper face from the first frame in all subsequent frames, ensuring that the principal components mainly represented mouth movements. Applying PCA of \\(\\mathcal{Z}_m\\) reveals that most principal components are responsible for solely mouth movements. For illustration of the first component, see Fig. 8 (bottom row). Based on distilled components, we introduce mouth PCA mouth loss:" + }, + { + "type": "equation", + "bbox": [ + 0.54, + 0.78, + 0.892, + 0.82 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {P C A}} \\left(v _ {i}, v _ {j}, n\\right) = \\frac {1}{n} \\sum_ {k = 1} ^ {n} \\left| \\mathrm {P C} _ {\\mathbf {v} _ {i}} (k) - \\mathrm {P C} _ {\\mathbf {v} _ {j}} (k) \\right| \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.825, + 0.894, + 0.9 + ], + "angle": 0, + "content": "Where \\( n \\) is the number of principal components to be considered, \\( \\mathrm{PC}_{\\mathbf{z}_i}(k) \\) and \\( \\mathrm{PC}_{\\mathbf{z}_j}(k) \\) are the \\( k \\)-th principal components of vectors \\( \\mathbf{z}_i \\) and \\( \\mathbf{z}_j \\) respectively. Implementing this loss has demonstrably enhanced the quality and accuracy of our results, as our ablation study confirms (see Tab. 6)." + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8503" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.078, + 0.091, + 0.21, + 0.108 + ], + "angle": 0, + "content": "6. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.115, + 0.471, + 0.192 + ], + "angle": 0, + "content": "In this section, we describe experiments we set for comparison for both our image-driven and speech-driven mode. Please see implementation details such as all used losses, component's sizes and architecture details, training procedure, data preprocess in supplementary materials." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.198, + 0.312, + 0.215 + ], + "angle": 0, + "content": "6.1. Image-driven comparison" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.221, + 0.47, + 0.356 + ], + "angle": 0, + "content": "Methods. We evaluated our image-driven model alongside various one-shot models. For this comparison, we included models like NOFA [32] and StyleHEAT [31], which use 3DMM expression blendshapes without disentanglement training, and models with trainable latent face motion representations like FOMM [22], UVA [23], MetaPortrait [35] and MegaPortraits [9]. It's noteworthy that NOFA and UVA require per-source optimization, unlike the others, including our model." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.357, + 0.47, + 0.478 + ], + "angle": 0, + "content": "Evaluation metrics and data. For our evaluation, we use 100 random, non-child, single-person images from the FFHQ dataset as sources. From the MEAD dataset, we chose 100 images covering all high-intensity emotions except neutral, and another 100 from the FEED dataset's extreme emotion task, featuring 5 identities not seen during training of our model. Each source image was paired with 20 driving images from these sets, totaling 2000 pairs." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.479, + 0.471, + 0.75 + ], + "angle": 0, + "content": "For assessing reenactment quality, we employ various metrics including the Frechet Inception Distance (FID) [13] to measure the distributional discrepancy between synthetic and real images. Cosine similarity (CSIM) from a face recognition network [5] quantifies the identity preservation in generated images. Additionally, we conducted user studies to determine preferences regarding motion (UMTN) and appearance (UAPP) preservation, presenting participants with triplets—either a driving or source image and two method outputs—and asking them to choose the one with better motion or appearance preservation. Our user study includes 34 participants and around 300 unique questions. Both metrics numerically reflect preference percentages. As our visual results Fig. 9, and quantitative metrics Tab. 3 show, our method demonstrates exceptional ability in translating strong and extreme expressions, notably surpassing other image-driven avatar techniques in user preference for facial expression translation (UMTN) and FID scores." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.757, + 0.318, + 0.773 + ], + "angle": 0, + "content": "6.2. Speech-driven comparison" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.78, + 0.47, + 0.855 + ], + "angle": 0, + "content": "Methods. To assess our model's speech-driven capabilities, we compared it with StyleHEAT [31], which also operates using both image and audio inputs, and audio-exclusive models including SadTalker [36], MakeItTalk [40], and PC-AVS [39]." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.856, + 0.471, + 0.901 + ], + "angle": 0, + "content": "Evaluation metrics and data. We employ the evaluation procedure described in [12]. We measure video realism using Frechet Inception Distance (FID) [13]. We also extract" + }, + { + "type": "table", + "bbox": [ + 0.505, + 0.09, + 0.891, + 0.177 + ], + "angle": 0, + "content": "
Driver:MEADFEED
MethodFID↓CSIM↑UMTN↑UAPP↑FID↓CSIM↑UMTN↑UAPP↑
MetaPortrait [35]77.50.64**78.80.66**
NOFA [32]69.20.65**69.80.663.88.8
FOMM [22]84.40.562.40.385.90.551.80.5
UVA [17]78.70.680.96.479.50.680.69.4
StyleHEAT [31]71.90.633.73.372.40.612.63.8
MegaPortraits [9]61.10.7322.746.662.80.7316.742.8
Our (EMOPortraits)59.60.7470.343.460.20.7074.634.7
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.499, + 0.18, + 0.894, + 0.444 + ], + "angle": 0, + "content": "Table 3. Our method notably outperforms others in the FID score and strongly leads in the user preference metric for face expression translation (UMTN). It excels in reliably translating strong and extreme expressions from the driving image, distinguishing it as the most capable among the compared methods. Additionally, our approach maintains the source image's shape and appearance (measured by CSIM and UAPP metrics) on par with MegaPortraits [9] when using strong and regular expressions from the MEAD dataset as the driving signal. However, we observe a minor deviation in identity preservation metrics for extreme emotions (using the FEED dataset). This difference may arise because our method attempts to transfer asymmetric facial expressions, while MegaPortraits might modify the source image less, or not at all, with a challenging driver (as depicted in Fig. 5, case 1), thereby achieving better identity preservation. UVA was excluded from the MEAD-driven user study due to the authors were able to provide only part of the test data. Also, MetaPortrait was excluded from both user studies because of late data provision by authors, offering only the basic model results, without their super-resolution module outputs." + }, + { + "type": "table", + "bbox": [ + 0.511, + 0.466, + 0.885, + 0.535 + ], + "angle": 0, + "content": "
MethodFID↓CSIM↑UMTN↓UAPP↓AUCz↓
MegaPortraits fine-tuned62.30.638.512.40.89
Ours w/o LcnCV61.40.679.217.90.88
Ours w/o Lsdm61.80.5911.36.70.85
Ours dim(z) = 51263.40.4425.61.10.77
Ours59.60.7445.461.90.80
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.499, + 0.545, + 0.894, + 0.67 + ], + "angle": 0, + "content": "Table 4. Our ablation study examines key aspects of the image-driven mode. After fine-tuning MegaPortraits on the FEED dataset, we observed a decline in identity preservation due to overfitting on FEED identities, without any improvement in emotion translation. The lack of \\(\\mathcal{L}_{CV}^{n}\\) primarily impacts the ability to translate extreme expression, as indicated by UMTN. Additionally, not using \\(\\mathcal{L}_{sdm}\\) or setting \\(dim(\\mathbf{z}) = 512\\) results in identity leakage, particularly evident in the latter scenario. For this ablation, we use the part of our driven by MEAD dataset and described in Sec. 6.1" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.704, + 0.895, + 0.901 + ], + "angle": 0, + "content": "facial landmarks from outputs of all methods, frontalized, and normalized to position the mouth edges at \\((-1,0)\\) and \\((1,0)\\). We compute mean absolute error (MAE) for both predicted mouth landmark positions \\((M_P)\\) and velocities \\((M_V)\\), along with errors in facial geometry \\((F_P)\\) and velocity \\((F_V)\\). For our evaluation, we use 100 randomly selected video sequences from the HDTF dataset [37], each up to 30 seconds. As the results show in Tab. 5, our speech-driven model performs on par with leading methods, excelling in realism and facial dynamics. SyncNet [6] is used to assess lip sync quality, providing offset (temporal misalignment between the audio and video) and confidence (certainty of audio-visual alignment) scores." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8504" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.113, + 0.088, + 0.195, + 0.152 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.196, + 0.089, + 0.277, + 0.152 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.278, + 0.089, + 0.36, + 0.152 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.364, + 0.089, + 0.442, + 0.152 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.443, + 0.089, + 0.523, + 0.152 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.526, + 0.089, + 0.607, + 0.152 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.61, + 0.089, + 0.689, + 0.152 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.692, + 0.089, + 0.772, + 0.152 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.774, + 0.089, + 0.855, + 0.152 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.113, + 0.159, + 0.194, + 0.222 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.196, + 0.159, + 0.277, + 0.222 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.278, + 0.159, + 0.36, + 0.222 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.362, + 0.159, + 0.441, + 0.222 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.443, + 0.159, + 0.524, + 0.222 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.526, + 0.159, + 0.607, + 0.222 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.61, + 0.159, + 0.689, + 0.222 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.691, + 0.159, + 0.772, + 0.222 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.774, + 0.159, + 0.855, + 0.222 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.113, + 0.231, + 0.194, + 0.294 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.196, + 0.231, + 0.277, + 0.294 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.278, + 0.231, + 0.36, + 0.294 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.362, + 0.231, + 0.441, + 0.294 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.443, + 0.231, + 0.524, + 0.294 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.526, + 0.231, + 0.607, + 0.294 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.61, + 0.231, + 0.689, + 0.294 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.691, + 0.231, + 0.772, + 0.294 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.774, + 0.231, + 0.855, + 0.294 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.113, + 0.302, + 0.194, + 0.364 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.138, + 0.372, + 0.168, + 0.381 + ], + "angle": 0, + "content": "Source" + }, + { + "type": "image", + "bbox": [ + 0.196, + 0.302, + 0.277, + 0.364 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.221, + 0.372, + 0.25, + 0.38 + ], + "angle": 0, + "content": "Driver" + }, + { + "type": "image", + "bbox": [ + 0.278, + 0.302, + 0.359, + 0.364 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.295, + 0.372, + 0.342, + 0.38 + ], + "angle": 0, + "content": "FOMM [22]" + }, + { + "type": "image", + "bbox": [ + 0.361, + 0.302, + 0.441, + 0.364 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.382, + 0.372, + 0.42, + 0.381 + ], + "angle": 0, + "content": "UVA [17]" + }, + { + "type": "image", + "bbox": [ + 0.443, + 0.302, + 0.524, + 0.364 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.453, + 0.372, + 0.515, + 0.381 + ], + "angle": 0, + "content": "StyleHEAT [31]" + }, + { + "type": "image", + "bbox": [ + 0.526, + 0.302, + 0.607, + 0.364 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.532, + 0.372, + 0.6, + 0.381 + ], + "angle": 0, + "content": "MetaPortrait [35]" + }, + { + "type": "image", + "bbox": [ + 0.61, + 0.302, + 0.689, + 0.364 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.627, + 0.372, + 0.67, + 0.381 + ], + "angle": 0, + "content": "NOFA [32]" + }, + { + "type": "image", + "bbox": [ + 0.692, + 0.302, + 0.772, + 0.364 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.697, + 0.372, + 0.766, + 0.381 + ], + "angle": 0, + "content": "MegaPortraits [9]" + }, + { + "type": "image", + "bbox": [ + 0.774, + 0.302, + 0.855, + 0.364 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.803, + 0.372, + 0.825, + 0.38 + ], + "angle": 0, + "content": "Ours" + }, + { + "type": "image_caption", + "bbox": [ + 0.221, + 0.386, + 0.747, + 0.4 + ], + "angle": 0, + "content": "Figure 9. A qualitative comparison of head avatar systems in cross-reenactment scenario." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.433, + 0.468, + 0.49 + ], + "angle": 0, + "content": "
MethodFID↓\\( F_P \\downarrow \\)\\( F_V \\downarrow \\)\\( M_P \\downarrow \\)\\( M_V \\downarrow \\)Sync. offset ↓Sync. conf. ↑
StyleHEAT [31]62.41.340.412.981.670.417.22
PC-AVS [39]71.52.780.512.561.043.798.42
SadTalker [36]55.60.910.352.121.170.277.58
MakeIRTalk [40]63.32.150.383.111.592.985.17
Ours (EMOPortraits)28.50.820.332.531.380.075.77
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.501, + 0.473, + 0.669 + ], + "angle": 0, + "content": "Table 5. Our model matches other leading speech-driven models, exceeding in realism (FID), facial geometry and velocity, but is slightly less accurate in mouth landmark positions and velocities. We also score best in audio-video sync offset, but lower in confidence. We believe that while SyncNet metrics are still helpful, they may not fully represent true synchronization quality, as generated videos can sometimes score higher than ground truth without necessarily being better synchronized. In our experimentation, ground truth videos scored an offset of 1.13 (which is worse than our method, SadTalker, and StyleHEAT) and a confidence score of 8.35 (less than PC-AVS). For qualitative analysis, please refer to our supplementary materials." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.707, + 0.468, + 0.774 + ], + "angle": 0, + "content": "
MethodFID↓MP↓MV↓FP↓FV↓
Ours w/o disentanglement44.70.980.455.782.55
Ours w/o LPCA29.30.840.353.131.88
Ours with MAE(\\( \\mathbf{z}_i, \\hat{\\mathbf{z}}_i^{\\text{speech}} \\))28.20.840.342.611.45
Ours28.50.820.332.531.38
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.786, + 0.472, + 0.897 + ], + "angle": 0, + "content": "Table 6. Ablation study of our speech-driven mode, where in the first line, we show that it notably underperforms without head pose - face expression disentanglement of the latent space (see Sec. 5.1). Removing \\(\\mathcal{L}_{\\mathrm{PCA}}\\) leads to a significant reduction in generating realistic mouth movements. Replacing \\(\\mathcal{L}_{\\mathrm{PCA}}\\) with \\(\\mathrm{MAE}(\\mathbf{z}_i,\\hat{\\mathbf{z}}_i^{\\mathrm{speech}})\\) causes a noticeable, decline in mouth movement generation. This suggests that identifying mouth-related PCA components was beneficial in enhancing the outcome." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.427, + 0.651, + 0.443 + ], + "angle": 0, + "content": "6.3. Ablation study" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.45, + 0.896, + 0.633 + ], + "angle": 0, + "content": "We conducted an extensive ablation study to evaluate the contributions of individual components within our method. For our main model, we present the evaluation of the importance of the proposed source-driver unmatched loss, canonical volume loss, and change of \\( dim(\\mathbf{z}) \\) from 512 in original model to 128 Tab. 4. For our speech-driven mode, we show how vital the disentanglement of the latent space for speech-drive ability, how \\( \\mathcal{L}_{\\mathrm{PCA}} \\) is better than naive matching of PCA components from random \\( \\mathcal{Z} \\) and how it helps match speech and latent expression space Tab. 6. For more ablation analysis, plots and visual comparison, please refer to our supplementary." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.648, + 0.62, + 0.663 + ], + "angle": 0, + "content": "7. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.674, + 0.895, + 0.779 + ], + "angle": 0, + "content": "In this paper, we introduce EmoPortraits, a novel method for creating neural avatars with superior performance in image-driven, cross-identity emotion translation. Our speech-driven mode makes it possible to drive the facial animation through multiple conditions (video, audio, head motion). We collected FEED dataset which, we believe, will be a valuable asset for researchers in diverse human-centered studies." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.78, + 0.896, + 0.9 + ], + "angle": 0, + "content": "However, our method has some limitations. It doesn't generate the avatar's body or shoulders, limiting some use cases. We currently integrate our output with a source image body. Additionally, the model sometimes struggles with accurate expression translation and underperforms with extensive head rotation. These challenges are crucial for future enhancements and remain central to our ongoing research efforts." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8505" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.184 + ], + "angle": 0, + "content": "[1] Emad Barsoum, Cha Zhang, Cristian Canton Ferrer, and Zhengyou Zhang. Training deep networks for facial expression recognition with crowd-sourced label distribution. In Proceedings of the 18th ACM international conference on multimodal interaction, pages 279-283, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.187, + 0.472, + 0.241 + ], + "angle": 0, + "content": "[2] Stella Bounareli, Christos Tzelepis, Vasileios Argyriou, Ioannis Patras, and Georgios Tzimiropoulos. Hyperreenact: One-shot reenactment via jointly learning to refine and retarget faces, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.244, + 0.472, + 0.298 + ], + "angle": 0, + "content": "[3] Egor Burkov, Igor Pasechnik, Artur Grigorev, and Victor Lempitsky. Neural head reenactment with latent pose descriptors. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). IEEE, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.301, + 0.471, + 0.356 + ], + "angle": 0, + "content": "[4] Houwei Cao, David G Cooper, Michael K Keutmann, Ruben C Gur, Ani Nenkova, and Ragini Verma. Crema-d: Crowd-sourced emotional multimodal actors dataset. IEEE transactions on affective computing, 5(4):377-390, 2014. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.358, + 0.471, + 0.399 + ], + "angle": 0, + "content": "[5] Qiong Cao, Li Shen, Weidi Xie, Omkar M. Parkhi, and Andrew Zisserman. Vggface2: A dataset for recognising faces across pose and age, 2018. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.401, + 0.471, + 0.47 + ], + "angle": 0, + "content": "[6] Joon Son Chung and Andrew Zisserman. Out of time: automated lip sync in the wild. In Computer Vision-ACCV 2016 Workshops: ACCV 2016 International Workshops, Taipei, Taiwan, November 20-24, 2016, Revised Selected Papers, Part II 13, pages 251-263. Springer, 2017. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.472, + 0.471, + 0.512 + ], + "angle": 0, + "content": "[7] Joon Son Chung, Arsha Nagrani, and Andrew Zisserman. Voxceleb2: Deep speaker recognition. In INTERSPEECH, 2018. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.515, + 0.471, + 0.557 + ], + "angle": 0, + "content": "[8] Michail Christos Doukas, Stefanos Zafeiriou, and Viktoriia Sharmanska. Headgan: One-shot neural head synthesis and editing, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.559, + 0.471, + 0.613 + ], + "angle": 0, + "content": "[9] Nikita Drobyshev, Jenya Chelishev, Taras Khakhulin, Aleksei Ivakhnenko, Victor Lempitsky, and Egor Zakharov. Megaportraits: One-shot megapixel neural head avatars, 2023. 2, 4, 5, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.616, + 0.47, + 0.656 + ], + "angle": 0, + "content": "[10] Guy Gafni, Justus Thies, Michael Zollhöfer, and Matthias Nießner. Dynamic neural radiance fields for monocular 4d facial avatar reconstruction, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.659, + 0.47, + 0.699 + ], + "angle": 0, + "content": "[11] Philip-William Grassal, Malte Prinzler, Titus Leistner, Carsten Rother, Matthias Nießner, and Justus Thies. Neural head avatars from monocular rgb videos, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.702, + 0.47, + 0.743 + ], + "angle": 0, + "content": "[12] Siddharth Gururani, Arun Mallya, Ting-Chun Wang, Rafael Valle, and Ming-Yu Liu. Space: Speech-driven portrait animation with controllable expression, 2022. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.746, + 0.471, + 0.813 + ], + "angle": 0, + "content": "[13] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. In Advances in Neural Information Processing Systems, 2017. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.816, + 0.471, + 0.857 + ], + "angle": 0, + "content": "[14] Philip Jackson and SJUoSG Haq. Surrey audio-visual expressed emotion (savee) database. University of Surrey: Guildford, UK, 2014. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.859, + 0.471, + 0.899 + ], + "angle": 0, + "content": "[15] Taras Khakhulin, Vanessa Sklyarova, Victor Lempitsky, and Egor Zakharov. Realistic one-shot mesh-based head avatars, 2022. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.472, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.135 + ], + "angle": 0, + "content": "[16] Alexander C. Li, Alexei A. Efros, and Deepak Pathak. Understanding collapse in non-contrastive siamese representation learning, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.137, + 0.892, + 0.176 + ], + "angle": 0, + "content": "[17] Xueting Li, Shalini De Mello, Sifei Liu, Koki Nagano, Umar Iqbal, and Jan Kautz. Generalizable one-shot neural head avatar, 2023. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.179, + 0.894, + 0.234 + ], + "angle": 0, + "content": "[18] Steven R Livingstone and Frank A Russo. The ryerson audiovisual database of emotional speech and song (ravdess): A dynamic, multimodal set of facial and vocal expressions in north american english. PloS one, 13(5):e0196391, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.236, + 0.894, + 0.319 + ], + "angle": 0, + "content": "[19] Patrick Lucey, Jeffrey F Cohn, Takeo Kanade, Jason Saragih, Zara Ambadar, and Iain Matthews. The extended cohn-kanade dataset (ck+): A complete dataset for action unit and emotion-specified expression. In 2010 iiee computer society conference on computer vision and pattern recognition-workshops, pages 94-101. IEEE, 2010. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.321, + 0.894, + 0.376 + ], + "angle": 0, + "content": "[20] Maja Pantic, Michel Valstar, Ron Rademaker, and Ludo Maat. Web-based database for facial expression analysis. In 2005 IEEE international conference on multimedia and Expo, pages 5-pp. IEEE, 2005. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.378, + 0.892, + 0.445 + ], + "angle": 0, + "content": "[21] KR Prajwal, Rudrabha Mukhopadhyay, Vinay P Namboodiri, and CV Jawahar. A lip sync expert is all you need for speech to lip generation in the wild. In Proceedings of the 28th ACM international conference on multimedia, pages 484-492, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.448, + 0.892, + 0.489 + ], + "angle": 0, + "content": "[22] Aliaksandr Siarohin, Stéphane Lathuilière, S. Tulyakov, Elisa Ricci, and N. Sebe. First order motion model for image animation. ArXiv, abs/2003.00196, 2019. 2, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.491, + 0.892, + 0.546 + ], + "angle": 0, + "content": "[23] Aliaksandr Siarohin, Willi Menapace, Ivan Skorokhodov, Kyle Olszewski, Hsin-Ying Lee, Jian Ren, Mengelei Chai, and Sergey Tulyakov. Unsupervised volumetric animation. arXiv preprint arXiv:2301.11326, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.548, + 0.894, + 0.601 + ], + "angle": 0, + "content": "[24] Michal Stypulkowski, Konstantinos Vougioukas, Sen He, Maciej Zieba, Stavros Petridis, and Maja Pantic. Diffused heads: Diffusion models beat gans on talking-face generation, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.604, + 0.892, + 0.659 + ], + "angle": 0, + "content": "[25] Mani Kumar Tellamekala, Ömer Ümer, Björn W. Schuller, Elisabeth Andre, Timo Giesbrecht, and Michel Valstar. Are 3d face shapes expressive enough for recognising continuous emotions and action unit intensities?, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.661, + 0.892, + 0.703 + ], + "angle": 0, + "content": "[26] Hao Wang, Yitong Wang, Zheng Zhou, Xing Ji, Dihong Gong, Jingchao Zhou, Zhifeng Li, and Wei Liu. Cosface: Large margin cosine loss for deep face recognition, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.705, + 0.892, + 0.773 + ], + "angle": 0, + "content": "[27] H. Wang, Yitong Wang, Zheng Zhou, Xing Ji, Zhifeng Li, Dihong Gong, Jin Zhou, and Wenyu Liu. Cosface: Large margin cosine loss for deep face recognition. 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5265-5274, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.775, + 0.892, + 0.845 + ], + "angle": 0, + "content": "[28] Kaisiyuan Wang, Qianyi Wu, Linsen Song, Zhuoqian Yang, Wayne Wu, Chen Qian, Ran He, Yu Qiao, and Chen Change Loy. Mead: A large-scale audio-visual dataset for emotional talking-face generation. In European Conference on Computer Vision, pages 700–717. Springer, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.847, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[29] Ting-Chun Wang, Arun Mallya, and Ming-Yu Liu. One-shot free-view neural talking-head synthesis for video conferencing. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2021. 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8506" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.161 + ], + "angle": 0, + "content": "[30] Yuelang Xu, Lizhen Wang, Xiaochen Zhao, Hongwen Zhang, and Yebin Liu. AvatarMAV: Fast 3d head avatar reconstruction using motion-aware neural voxels. In *Special Interest Group on Computer Graphics and Interactive Techniques Conference Conference Proceedings*. ACM, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.164, + 0.472, + 0.232 + ], + "angle": 0, + "content": "[31] Fei Yin, Yong Zhang, Xiaodong Cun, Mingdeng Cao, Yanbo Fan, Xuan Wang, Qingyan Bai, Baoyuan Wu, Jue Wang, and Yujiu Yang. Styleheat: One-shot high-resolution editable talking face generation via pre-trained stylegan, 2022. 2, 3, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.234, + 0.472, + 0.288 + ], + "angle": 0, + "content": "[32] Wangbo Yu, Yanbo Fan, Yong Zhang, Xuan Wang, Fei Yin, Yunpeng Bai, Yan-Pei Cao, Ying Shan, Yang Wu, Zhongqian Sun, and Baoyuan Wu. Nofa: Nerf-based one-shot facial avatar reconstruction, 2023. 2, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.291, + 0.472, + 0.331 + ], + "angle": 0, + "content": "[33] Egor Zakharov, Aliaksandra Shysheya, Egor Burkov, and Victor Lempitsky. Few-shot adversarial learning of realistic neural talking head models, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.333, + 0.472, + 0.374 + ], + "angle": 0, + "content": "[34] Egor Zakharov, Aleksei Ivakhnenko, Aliaksandra Shysheya, and Victor Lempitsky. Fast bi-layer neural synthesis of one-shot realistic head avatars, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.376, + 0.472, + 0.432 + ], + "angle": 0, + "content": "[35] Bowen Zhang, Chenyang Qi, Pan Zhang, Bo Zhang, Hsiang-Tao Wu, Dong Chen, Qifeng Chen, Yong Wang, and Fang Wen. Metaportrait: Identity-preserving talking head generation with fast personalized adaptation, 2023. 2, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.434, + 0.472, + 0.488 + ], + "angle": 0, + "content": "[36] Wenxuan Zhang, Xiaodong Cun, Xuan Wang, Yong Zhang, Xi Shen, Yu Guo, Ying Shan, and Fei Wang. Sadtalker: Learning realistic 3d motion coefficients for stylized audio-driven single image talking face animation, 2023. 3, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.49, + 0.472, + 0.558 + ], + "angle": 0, + "content": "[37] Zhimeng Zhang, Lincheng Li, Yu Ding, and Changjie Fan. Flow-guided one-shot talking face generation with a high-resolution audio-visual dataset. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3661–3670, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.56, + 0.472, + 0.601 + ], + "angle": 0, + "content": "[38] Yufeng Zheng, Victoria Fernández Abrevaya, Marcel C. Bühler, Xu Chen, Michael J. Black, and Otmar Hilliges. I m avatar: Implicit morphable head avatars from videos, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.603, + 0.472, + 0.657 + ], + "angle": 0, + "content": "[39] Hang Zhou, Yasheng Sun, Wayne Wu, Chen Change Loy, Xiaogang Wang, and Ziwei Liu. Pose-controllable talking face generation by implicitly modularized audio-visual representation, 2021. 2, 3, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.66, + 0.472, + 0.7 + ], + "angle": 0, + "content": "[40] Yang Zhou, Xintong Han, Eli Shechtman, Jose Echevarria, Evangelos Kalogerakis, and Dingzeyu Li. MakeltTalk. ACM Transactions on Graphics, 39(6):1-15, 2020. 3, 7, 8" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.7 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "8507" + } + ] +] \ No newline at end of file diff --git a/2024/EMOPortraits_ Emotion-enhanced Multimodal One-shot Head Avatars/ab68d1cd-73a3-4a4c-9896-716aeaaff436_origin.pdf b/2024/EMOPortraits_ Emotion-enhanced Multimodal One-shot Head Avatars/ab68d1cd-73a3-4a4c-9896-716aeaaff436_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..ff9051d994ffd7e42d3b6b1bcbf106bf448ec705 --- /dev/null +++ b/2024/EMOPortraits_ Emotion-enhanced Multimodal One-shot Head Avatars/ab68d1cd-73a3-4a4c-9896-716aeaaff436_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:57c622f292fb82ba78bc221eb442f36191520e00ccc87bbdbba9030f7bced5a3 +size 2935821 diff --git a/2024/EMOPortraits_ Emotion-enhanced Multimodal One-shot Head Avatars/full.md b/2024/EMOPortraits_ Emotion-enhanced Multimodal One-shot Head Avatars/full.md new file mode 100644 index 0000000000000000000000000000000000000000..3095e51f6722ea34915833e3007b747d906be6ce --- /dev/null +++ b/2024/EMOPortraits_ Emotion-enhanced Multimodal One-shot Head Avatars/full.md @@ -0,0 +1,399 @@ +# EMOPortraits: Emotion-enhanced Multimodal One-shot Head Avatars + +Nikita Drobyshev + +Antoni Bigata Casademunt +Stavros Petridis + +Konstantinos Vougioukas +Maja Pantic + +Zoe Landgraf + +{n.drobyshev, a/bigata-casademunt22, k.vougioukas, zoe.landgraf15, stavros.petridis04, m.pantic}@imperial.ac.uk Imperial College London + +![](images/5571c03b59a22946cb276d1deded7196a8c084b040541ce529bbee5da0acb26d.jpg) + +![](images/6f78ae2acda6416384be5581a2bc2ae87413212cd8bfa3feb9df5869a0f8e39f.jpg) + +![](images/867efe1bddaf1887f13140417c41e46dc4c66e055a39bb4c7e189e5a87cba5b5.jpg) + +![](images/c12038d2254398ecb4c7ba204508b41706cc611de0c85c955e62dc16a81d8eb4.jpg) + +![](images/272336b8074a1488e724b3c0fbb531059f7b99cb096e5d5bc5ee91f1570dcb0f.jpg) + +![](images/e1eb172f4eb38c6ece771d39f7a6007b8dcd48c8729563a1872fa82a97e69526.jpg) + +![](images/d5150c201a58cf952e956203832059c28533b90018e0a66348c1f3f500048536.jpg) + +![](images/316eae8164a7f14a42016f25d69f3acd05748b74e70f6c167f52db9458ece090.jpg) + +![](images/479f825123fe7a77d8b249bb5eb77206881d58fc1981346f540388e496db4655.jpg) + +![](images/8da0d6fcd24f450ef1006655f2860be5194a1760923b15e36c23cfbcd55b3cd1.jpg) + +![](images/1d3be78635c31aa04786c174e323af741e93526bbfde5ad9472c0a7afda2dfaf.jpg) +Source + +![](images/a2ab0f9b76c3a8b179b604978468a109567c35f2493df4b12487308d887b0ea9.jpg) +Driver + +![](images/7af6d1668fe3d655129ce10d232c0993eb2693facbf0ba6bfeccb40e09db4e40.jpg) +Output + +![](images/37fa260d97e9e5d2b12fdb7dff66ab9aa5c86ba957ebcccf9309e28ed674827c.jpg) +Driver + +![](images/6f169bc4bce76438b2a929bd4d15373bc33cedd96bc3fcaca9948d0a78a551f7.jpg) +Output +Figure 1. Selected animation results for image-driven mode of our method. + +# Abstract + +Head avatars animated by visual signals have gained popularity, particularly in cross-driving synthesis where the driver differs from the animated character, a challenging but highly practical approach. The recently presented MegaPortraits model has demonstrated state-of-the-art results in this domain. We conduct a deep examination and evaluation of this model, with a particular focus on its latent space for facial expression descriptors, and uncover several limitations with its ability to express intense face motions. To address these limitations, we propose substantial changes in both training pipeline and model architecture, to introduce our EMOPortraits model, where we: + +Enhance the model's capability to faithfully support intense, asymmetric face expressions, setting a new state-of-the-art result in the emotion transfer task, surpassing previous methods in both metrics and quality. + +Incorporate speech-driven mode to our model, achieving top-tier performance in audio-driven facial animation, making it possible to drive source identity through diverse modalities, including visual signal, audio, or a blend of both. + +Furthermore, we propose a novel multi-view video dataset featuring a wide range of intense and asymmetric facial expressions, filling the gap with absence of such data in existing datasets. + +# 1. Introduction + +# 1.1. Representation of facial expression + +Advancements in neural head technologies now enable the creation of realistic avatars from a few or even one image, with the latter being crucial when only one source image is available. Cross-driving synthesis, where avatars are animated with different identities is a key technique for virtual reality, filmmaking, photo animation, etc. However, accurate transfer of facial expressions, especially those that are intense and uneven, remains a substantial challenge in avatar animation, particularly for cross-driving synthesis. While previous research focused mainly on preserving identity and transferring moderate facial motions, our work also seeks to accurately drive high-intensity and asymmetric expressions. + +Our research is built on and extends the MegaPortraits [9] model, notable for state-of-the-art results in cross-driving synthesis. An additional advantage using this method is that, unlike many avatar systems that depend on limited [3, 25] predefined motion descriptors, MegaPortraits, learns expression representations from scratch, allowing for greater adaptability to a wider range of expressions. We investigate the latent expression spaces and training methods of MegaPortraits to enhance its ability to depict a broad spectrum of facial expressions. Our comprehensive analysis reveals that, while the original model shows limited effectiveness in intense motion representing, it has significant potential for improvement through targeted architectural modifications, adjustments in the training approach, and the integration of our novel dataset. + +# 1.2. Integrating speech driving + +We integrate speech in our model to be used either as a complementary or an alternative driver, a crucial aspect for applications like virtual assistance or mixed reality when a primary visual signal is absent. + +By improving disentanglement of facial expression latent space, we remove disturbing information from face motion descriptors and emphasize components that solely control lip movements. Based on this, we formulated a novel loss that helps to achieve desirable results. Furthermore, our method can generate plausible head rotations and blinks, which appear natural and enhance its applicability across various tasks. Thus, our final method can drive source identity through image, audio, or a mix of them. + +# 1.3. FEED: Facial Extreme Emotions Dataset + +Public data scarcity is a common obstacle in deep learning, the topic of human facial expressions and movements is not an exception. In particular, we note a lack of high quality video data capturing a wide range of facial expressions, with existing datasets not going beyond basic facial actions shown at Fig. 2. To address this, we collect a new dataset that + +includes basic expressions and also captures complex movements like blinks, winks, and head and tongue movements, along with varied extreme expressions which are difficult or impossible to categorize through basic facial actions. Given this variety in expressions in our dataset, we believe that it will be a valuable resource for research in human emotions and facial reconstruction fields. In summary, our main contributions are as follows: + +- We introduce our new EMOPortraits model for one-shot head avatars synthesis, that is capable of transferring intense facial expression, showing state-of-the-art results in cross-driving synthesis. We achieve this through careful latent facial expression space development as well as novel losses and a minimal amount of domain-specific data ( $\sim$ 0.1% of the train set). +- We integrate a speech-driving mode in our model, that demonstrates cutting-edge results in speech-driven animations. It functions effectively alongside visual signals or independently, also generating realistic head rotations and eye blinks. +- We present a unique multi-view dataset that spans a broad spectrum of extreme facial expressions, filling the gap of absence of such data in existing datasets. + +# 2. Related Work + +# 2.1. Visual-driven head avatars + +In recent years, the domain of neural avatars has branched into two prominent sub-fields: person-specific and person-agnostic avatars. While person-specific methodologies [10, 11, 30, 38, 39] excel in delivering stunning realism and motion fidelity for a particular individual, they encounter challenges representing an arbitrary person. Moreover, these approach demands multiple frames of training data, distinct training for each avatar, and can struggle to replicate motions not encountered during training. + +Person-agnostic methods, on the other hand, don't need train or fine-tuning for each new person and present an alternative approach to talking-head synthesis. Earlier works in this domain generated avatars in a few-shot technique [3, 33], while subsequent studies introduced one-shot capabilities [2, 8, 9, 22, 29, 31, 32, 34]. Many of these works employ predefined motion representations, such as 3DMM's blendshapes [8, 15, 31, 32]. In contrast, some learned latent emotion representation from scratch [3, 9, 22, 34, 35]. This latter approach holds potential for better motion representation as in such settings, expression descriptors become entirely trainable and free from inheriting the limitations typical for blendshapes as was shown in [3, 25]. In our work, we adopt this strategy, remarkably improving upon the methodology used in MegaPortraits[9]. + +
MethodBFEsFFED dataset internal structure
Head rotationsWinksEyes moveAsym. em.Tongue em.Extreme em.
ModeMild/StrongAxes/randomMild/StrongRandomStrongStrongExtreme
# of part.21/23212321232323
Avg. length1:38/1.520:580:490:331:411:533:02
+ +Table 1. The tasks we asked our participants to perform for our FEED dataset. Here BFEs stands for basic facial expressions, shown in Fig. 2 + +
DatasetDatasets comparison
#Actors#ViewsBasic exp.Strong exp.Extrme & Assym. exp.Tongue exp.Resolution
SAVESS41✓✘✘✘1280 x 1024
RAVGESS241✓✘✘✘1920 x 1080
CREMA-D911✓✘✘✘1280 x 720
MMI251✓✓✘✘1920 x 1080
MEAD487✓✓✘✘1920 x 1080
Ours (FEED)233✓✓✓✓3840 x 2160
+ +Table 2. Comparisons with modern, high-quality audio-visual datasets created in controlled settings. + +# 2.2. Face expression datasets + +Early image datasets[1, 19] predominantly offered annotated face expression data corresponded to up to 8 basic emotions. However, these datasets are not well-suited for training head avatars, which necessitate video or multi-view data. + +The SAVEE [14] dataset captures facial expressions in speech, but involves just 4 actors. The MMI [20] dataset, more expansive in actor participation, offers spontaneous moderate-intensity expression but is restricted to single-view sequences. RAVDESS [18] distinguishes itself by capturing two motion intensities, but its limited recordings challenge broad applicability. CREMA-D [4] encompasses three expression intensities, and though MEAD [28] provides detailed multi-view data across three intensities, it's limited to the standard eight expression groups. Our novel FEED dataset aims to address the scientific community's demand for high-quality multi-view facial expression videos outside the standard categories: Joy, Fear, Sadness, Disgust, Anger Contempt and Surprise. + +# 2.3. Speech-driven head avatars + +While numerous studies introduce speech-driven avatars [12, 21, 24, 31, 36, 39, 40], few excel in producing high-quality talking heads with authentic rotations, blinks, and the capability to use both visual and audio inputs. For instance, Wav2Lip [21] aims at re-dubbing videos with accurate lip motions, but often falls short in realism with using single images. Diffused heads [24] struggles to generate long sequences and doesn't provide access to pose control. MakeItTalk [40] animates facial landmarks in a speaker-specific manner, yet struggles with head pose controls due to its non-reliance on 3D. PC-AVS [39] offers a solution, but demands a driving video for pose modulation. Newly developed models such as SadTalker [36] and StyleHEAT [31] have shown promising outcomes in generating high-quality talking heads. We conducted a quantitative evaluation of our audio-driven mode compared to the aforementioned models. + +![](images/e582360968f3b94ebaa9e6716b25e04a6f5eb78834c851fa58aea8caced7cf0b.jpg) + +![](images/33d7247f6f7c8a4a0d29ea41ed87dcebf4c56350ecb13484916dc2b21c5c7926.jpg) + +![](images/a0ffc5d39905b35e0710031436f2b4a76dde5773b3d2205c14a51ddcbc729cf5.jpg) +Figure 2. Illustration of the problem in publicly available face expression data and selected examples from our FEED dataset. + +# 3. FEED dataset + +As previously mentioned, the current publicly available datasets with human face expression videos fall short in capturing a broad range of facial manipulations (see illustration at Fig. 2). In our view, this gap could stem from the complexity of designing clear tasks for participants, coupled with the reality that not everyone is ready for performing extreme expressions on camera. To bridge this gap, our pioneering FEED dataset is designed to meet the scientific community's need for high-quality, multi-view emotion videos that extend beyond standard emotions. It contains various expressions, including strong asymmetric ones, tongue and cheeks movements, winks, head rotations, eye movements and more nuanced gestures. ; + +Our dataset consists of 520 multi-view videos of 23 subjects, captured with 3 cameras. As extreme face motions are complex and their perception can be heavily influenced by subtle differences, we use a high resolution of 4k for all video, capturing the whole face up to the level of individual hair strands and wrinkles, as shown in Fig. 2. For more examples, please refer to the supplementary materials. Our participants were asked to perform 7 tasks (see Tab. 1) to cover as many facial and head movements as possible. + +In our comparison with other expression datasets detailed in Tab. 2, the MEAD dataset [28] emerges as the closest to ours regarding viewpoints and maximal expression intensity. However, despite MEAD's larger number of actors and viewpoints, it offers significantly less variety of facial expressions and have lower resolution than our dataset. + +# 4. Expression enhancement + +We found that the MegaPortraits [9] model fails to transfer intense expressions correctly, as shown in Fig. 9. Using our FEED dataset of strong and uneven facial expression to fine-tune a pre-trained MegaPortraits model didn't enhance final results as shown at Tab. 4. Training from scratch on FEED leads to fast overfitting due to the small number of identities presented in the dataset compared to VoxCeleb2 [7] used in [9] (23 vs $\sim 5000$ ). Our method effectively injects this small dataset into training, yielding desirable results. We list our key findings here, for all details, see supplementary materials. Our model's scheme, displayed in Fig. 4, shares similarities with [9], leading us to occasionally refer to certain elements of our scheme in the context of [9]'s pipeline. + +# 4.1. Latent expression space + +We begin by exploring MegaPortraits' latent expression descriptors' space $(\mathbf{z}_{s / d}$ in Fig. 4), which is crucial for expression transfer, through PCA analysis. Inspired by [16], which demonstrated that the area under the cumulative explained variance curve of singular values (denoted here as $\mathbf{AUC}_{\mathbf{z}}$ ) can serve as an effective metric for dimensionality collapse in some latent spaces, we applied this measure to our study. + +This metric allows us to forecast model performance: greater collapse suggests reduced entropy in the latent space, which correlates with lower representational quality. High quality of the expression space, produced by $\mathbf{E}_{\text {motion }}$ (as shown in Fig. 4), is crucial to capture and differentiate between the nuances of strong facial expressions. Our ablation study confirms (see Tab. 4) that there is a notable correlation between how broad and isotropic model's latent space is (expressed by $\mathbf{AUC}_{\mathbf{z}}$ ) and its final performance. Visual comparison is available in supplementary materials. As shown in Fig. 3, our model's latent space outperforms MegaPortraits in expression representation ability. This is supported by both visual (Fig. 9) and quantitative analysis (Tab. 3). + +![](images/43e0be9fd62862a686fc0a3749085dde5d7228aefbfb924c5e39ad2a1a627ee5.jpg) +Figure 3. Comparison of latent spaces. Left plot shows that our model's latent space is wider and exhibits more even variance distribution. Also, as shown on the right plot, a greater number of principal components are involved in capturing variance across various thresholds. This implies a more robust representational capacity of expression space compared to [9]. The VoxCeleb2 test set was used for both plots. + +![](images/cfaa5cbc7efaf0aaec7497921709bfe2ffc0e67e9a55e614e647ce8545e2ac06.jpg) +Figure 4. Method Overview. We use $\mathbf{E}_{\mathrm{app}}$ to extract volume features $\mathbf{V}_s$ and a global descriptor $\mathbf{e}_s$ from the source image. Then $\mathbf{E}_{\mathrm{motion}}$ or $\mathbf{E}_{\mathrm{audio}}$ generates motion representations from source and driver, including head rotations $\mathbf{R}_{s / d}$ , translations $\mathbf{t}_{s / d}$ , and expression descriptors $\mathbf{z}_{s / d}$ . Using them, we predict warpings $\mathbf{w}_{s\rightarrow}$ and $\mathbf{w}_{\rightarrow d}$ . First warping and $\mathbf{G}_{3D}$ transform $\mathbf{V}_s$ into a canonical volume $\mathbf{V}_s^C$ by removing the source motions. Second warping and $\mathbf{G}_{2D}$ imposes the driver's motions and renders the final image. + +Both metrics shown in Fig. 3 were vital in our model's development since they highly correlate with the final model's ability to transfer intense expressions. This was particularly useful for finding adjustments described in Sec. 4.2 and Sec. 4.3, as early visual results during training are not very telling, but changes in the latent space are noticeable after a few epochs. We also found that in MegaPortraits, just a few principal components significantly affect variance, with only 18 components making up $99\%$ (Fig. 3), suggesting their $512\mathbf{z}$ 's dimension is oversized. For our model, we found a 128-dimensional $\mathbf{z}$ is optimal, offering solid performance and reducing overfitting in imbalanced data, as our ablation study reveals. We calculate the Explained Variance $(\mathrm{EV}_i)$ for the $i^{th}$ component and $(\mathbf{AUC}_{\mathbf{z}})$ using equations shown in Eq. (1), which involves standardizing the vector set $\mathcal{Z}$ for PCA and sorting the eigenvalues $\lambda$ in descending order. + +$$ +\mathbf {Z} _ {\mathrm {s t d}} = \frac {\mathcal {Z} - \bar {\mathcal {Z}}}{\sigma (\bar {\mathcal {Z}})}, \quad \boldsymbol {\lambda} = \operatorname {s o r t e d} _ {\mathrm {d e s c}} \left(\operatorname {e i g} \left(\frac {1}{n - 1} \mathbf {Z} _ {\mathrm {s t d}} ^ {\top} \mathbf {Z} _ {\mathrm {s t d}}\right)\right), \tag {1} +$$ + +$$ +\mathrm {E V} _ {i} = \frac {\lambda_ {i}}{\sum_ {j = 1} ^ {p} \lambda_ {j}}, \quad \mathbf {A U C _ {z}} = \frac {1}{d} \frac {\sum_ {i = 1} ^ {d} \sum_ {k = 1} ^ {i} \lambda_ {k}}{\sum_ {j = 1} ^ {d} \lambda_ {j}}, +$$ + +# 4.2. Canonical volume + +Despite being designed to exclude face expression details, we found that canonical volume $(\mathbf{V}^C$ in Fig. 4) in MegaPortraits actually retains significant expression information from the source image, contributing to poor translation of intense expressions, as shown in our experiment detailed in Fig. 5. In this experiment, using portraits with varied expressions, we assessed the expression translation accuracy and how expression intensity affects the $\mathbf{V}^C$ s. Our findings suggest that the canonical volumes in MegaPortraits are not truly neutral. As confirmed by our ablation study (Tab. 4), creating an expression-free $\mathbf{V}^C$ is essential for precise translation of intense motions. We believe that a more neutral canonical volume improves tractability and effectiveness in expression translation tasks. + +![](images/65fd3cb501e87e562244924ab7734643dc949462bd6d5e71948526318373e41c.jpg) +Figure 5. Canonical volumes $(\mathbf{V}^C)$ in MegaPortraits [9] are not expression-neutral. To show it we use three portraits: A (regular expression), B (intense expression), D (regular expression, new identity). Case 1 visualizes poor results in transferring B's intense expression to A using a self-driving mode, contrasting with the effective reconstruction of B when used as both driver and source in case 3. This discrepancy, indicative of expression leakage into the canonical volume, is further quantified by a $43\%$ relative difference in $\mathbf{V}^C$ s between B and C, contrary to expectations of their similarity for same identity. In cross-driving generation case 2, using B as the driver and D as the source yielded poor results. However, in case 4, after using additive operations on source canonical volumes, the output expression is much closer to driver than in case 2. This manipulation again confirms the significant retention of expression information in canonical volumes. + +To overcome this issue, we propose to match canonical volumes from the different images of the same person $(\mathbf{V}_{s^n}^C,\mathbf{V}_{d^n}^C)$ during training as following: + +$$ +\mathcal {L} _ {\mathrm {C V}} ^ {\mathrm {n}} = \mathcal {L} _ {\mathrm {M A E}} \left(\mathbf {V} _ {s ^ {n}} ^ {C}, \mathbf {V} _ {d ^ {n}} ^ {C}\right), \tag {2} +$$ + +This loss ensures $\mathbf{V}^C$ remains stable and expression-independent, crucial for translating intense expressions, as shown in our ablation study (Tab. 4). + +# 4.3. Source-driver mismatch loss + +In addition to maintaining expression-free $\mathbf{V}^{C}$ , it's vital to remove all identity information from the expression vector $\mathbf{z}$ . While contrastive losses address this in [9], our experiments indicate they are insufficient to prevent overfitting in our imbalanced data scenario, where emotionally intense images represent only $0.1\%$ of our training set but are sampled $25\%$ of the time. We introduce a novel self-supervised loss which mitigates identity information in latent expression vectors: + +$$ +\begin{array}{l} \mathcal {L} _ {s d m} (\mathbf {z} _ {s}, \mathbf {z} _ {d}) = w * \max (0, \cos (\mathbf {z} _ {s}, \mathbf {z} _ {d}) - m a r g i n), \\ w; \text {m a r g i n} = \left\{ \begin{array}{l l} 1; 0. 5 & \text {i f} \mathbf {z} _ {s}, \mathbf {z} _ {d} \text {a r e f r o m V C 2} \\ 1 0; 0. 2 5 & \text {i f} \mathbf {z} _ {s}, \mathbf {z} _ {d} \text {a r e f r o m F E E D} \end{array} \right. \tag {3} \\ \end{array} +$$ + +Here, $\mathbf{z}_s, \mathbf{z}_d$ are emotion vectors from source and driver images, $w, \text{margin}$ adjust the loss's intensity and strictness respectively. We increase $w$ for FEED due to its higher sampling rate and decrease margin for more assured expression variance in $s, d$ . Refer to Fig. 6 for a visual summary of all self-supervised losses enhancing our latent emotion space. + +![](images/4168394fbcbb61f875f732ea64d8e9e00b593fc9b8b7198c39c30b7289b588f7.jpg) +Figure 6. The visualisation of our self-supervised losses. $\mathcal{L}_{\mathrm{cos}}$ is a contrastive loss, similar to one used in [9] but uses different pairs as described in Sec. 5.1. Our novel $\mathcal{L}_{sdm}$ is designed to prevent overfitting in an imbalanced dataset. + +# 5. Incorporating speech-driving mode + +# 5.1. Latent space disentanglement + +During our latent space analysis described in Sec. 4, we found that combining expression from one driver with the head pose from another resulted in poor performance for MegaPortraits model, as illustrated in Fig. 7 (bottom row). Indeed, the base model lacks a mechanism to intentionally prevent head pose leakage. Disentangled expression latent space, besides expanding the model's use cases, plays the crucial role for the speech-driving mode. When the latent vector $\mathbf{z}$ is entangled with head pose data, predicting it from speech becomes challenging because speech lacks head rotation information, unlike images. Moreover, we would like to have a full control over head rotations during speech mode. We were able to make our latent space disentangled by changing the way we sample images for $\mathcal{L}_{\mathrm{cos}}$ Eq. (5). + +The MegaPortraits model uses $\mathcal{L}_{\mathrm{cos}}$ - a modified Large Margin Cosine Loss [26] to mitigate the transfer of appearance features to expression descriptor (latent vector $\mathbf{z}$ on Fig. 4), which is crucial for cross-driving mode. Without this, differences in appearance between the source and driver, like hairstyle or skin tone, leak from driver to output image. + +For computing this loss, the authors utilize a supplementary source-driving pair $(\mathbf{x}_{s^{*}}$ and $\mathbf{x}_{d^{*}})$ from a different video with another identity, ensuring a distinct appearance from the current $\mathbf{x}_s,\mathbf{x}_d$ pair. The base model is then employed to produce cross-reenacted image $(\hat{\mathbf{x}}_{s^{*}\rightarrow d} = \mathbf{G}_{\mathrm{base}}(\mathbf{x}_{s^{*}},\mathbf{x}_d))$ Concurrently, they determine a separate motion descriptor, $\mathbf{z}_{d^{*}} = \mathbf{E}_{\mathrm{motion}}(\mathbf{x}_{d^{*}})$ . Descriptors $\mathbf{z}_{* \rightarrow d}$ from the respective forward passes are also used for positive and negative pairs. + +![](images/2fc771a3720ccaa6e2cde67511df1976c6cbeed5e2acdb994a7fe463b911909b.jpg) +Figure 7. The illustration of disentanglement problem. Images in the right row indicate the expression descriptor $\mathbf{z}$ entangled with head pose for MegaPortraits, whereas we avoid it using changes proposed in Sec. 5.1. + +Our innovation here is a new sampling strategy: for each $\mathbf{x}_s$ , $\mathbf{x}_d$ pair, we sample one more random additional pair $\mathbf{x}_{s^m}$ and $\mathbf{x}_{d^m}$ apart from $\mathbf{x}_{s^*}$ and $\mathbf{x}_{d^*}$ . We then apply the model to produce the following cross-reenacted moved image using source identity from $\mathbf{x}_{s^*}$ , desired head pose from $\mathbf{x}_{s^m}$ and emotions from $\mathbf{x}_d$ , represented as: $\hat{\mathbf{x}}_{s^{*m}\rightarrow d} = \mathbf{G}_{\mathrm{base}}(\mathbf{x}_{s^{*m}},\mathbf{x}_d)$ . This creates a positive pair with the same desired emotion but varying head poses. + +Motion descriptors are organized into positive pairs $\mathcal{P}$ for alignment, and negative pairs $\mathcal{N}$ for non-alignment: $\mathcal{P} = \left\{(\mathbf{z}_{s\rightarrow d},\mathbf{z}_d),(\mathbf{z}_{s^*\rightarrow d},\mathbf{z}_d),(\mathbf{z}_{s^{*m}\rightarrow d},\mathbf{z}_d)\right\}$ , and $\mathcal{N} = \left\{(\mathbf{z}_{s\rightarrow d},\mathbf{z}_{d^*}),(\mathbf{z}_{s^*\rightarrow d},\mathbf{z}_{d^*}),(\mathbf{z}_{s^{*m}\rightarrow d},\mathbf{z}_{d^*})\right\}$ (see Fig. 6). These pairs are used to calculate the following cosine distance: + +$$ +d \left(\mathbf {z} _ {i}, \mathbf {z} _ {j}\right) = s \cdot \left(\left\langle \mathbf {z} _ {i}, \mathbf {z} _ {j} \right\rangle - m\right), \tag {4} +$$ + +where both $s$ and $m$ are hyperparameters. This distance is then used to calculate a large margin cosine loss (Cos-Face) [27]: + +$$ +\mathcal {L} _ {\cos} = - \sum_ {(\mathbf {z} _ {k}, \mathbf {z} _ {l}) \in \mathcal {P}} \log \frac {\exp \left\{d \left(\mathbf {z} _ {k} , \mathbf {z} _ {l}\right) \right\}}{\exp \left\{d \left(\mathbf {z} _ {k} , \mathbf {z} _ {l}\right) \right\} + \sum_ {(\mathbf {z} _ {i} , \mathbf {z} _ {j}) \in \mathcal {N}} \exp \left\{d \left(\mathbf {z} _ {i} , \mathbf {z} _ {j}\right) \right\}} \tag {5} +$$ + +This loss prevents the head pose from leaking into the embeddings as shown in Fig. 7 (upper row) + +By doing this, we achieved the capability to interpret $\mathbf{z}_i$ as a latent vector encapsulating the emotional content from an image $\mathbf{x}_i$ . This advancement opened the door for an intriguing possibility during inference: the prediction of this vector could potentially be derived from an audio signal instead of a facial image using the motion encoder $\mathbf{E}_{\mathrm{motion}}$ , effectively transforming the model into a speech-driven system. The critical challenge here is aligning the audio input with the pretrained latent space. We addressed this by introducing an additional audio encoder, $\mathbf{E}_{\mathrm{aud}}$ , and employing $\mathbf{E}_{\mathrm{motion}}$ as a teacher model for this new encoder. A crucial factor in achieving remarkable accuracy in lip synchronization was our ability to isolate specific components within the expression latent space, that are solely responsible for mouth movements. We describe it in the next subsection Sec. 5.2. + +![](images/426d2459ddac66dc468bd55fd6056ae9f8154ca624c50c1e768277a3b7502cea.jpg) +Figure 8. In the upper row, the first principal component from a random subset of expression vectors $\mathcal{Z}$ affects not just the mouth but also blinks and gaze, making it difficult to interpret and isolate mouth movement components, needed for training $\mathbf{E}_{\mathrm{aud}}$ . However, as shown in the bottom row, using $\mathcal{Z}_m$ from Sec. 5.2, the first principal component, $\mathbf{v}_1$ , focused on mouth movements, is effectively isolated, $\mathbf{z}_0$ - latent vector, corresponded to the central image. + +# 5.2. Mouth movements + +To train the audio encoder $\mathbf{E}_{\mathrm{aud}}$ for predicting $\mathbf{z}$ from speech, we chose to use the motion encoder $\mathbf{E}_{\mathrm{motion}}$ as a teacher model. However, direct matching vectors $\mathbf{z}_i^{\mathrm{aud}}$ from $\mathbf{E}_{\mathrm{aud}}$ and pseudo-ground truth $\mathbf{z}_i$ from $\mathbf{E}_{\mathrm{motion}}$ gives poor results (see Tab. 6). We believe that the reason behind this is that speech-derived $\mathbf{z}_i^{\mathrm{aud}}$ , while able to capture lip movements, struggles with other facial motions, especially in the upper face. Thus, we focused on isolating mouth movement components in the expression latent space using PCA analysis. + +Employing PCA on a broad array of expression vectors $\mathcal{Z}$ , generated from images with varied facial emotions, yields principal components with their explained variances $(\mathrm{EV}_i)$ . Altering these components, particularly those with high $\mathrm{EV}_i$ , significantly changes facial expressions in an image. However, they typically affect a combination of facial features—Fig. 8 (upper row), not isolated areas like the mouth or eyes. + +To focus on mouth movements, we created a unique set of $\mathcal{Z}_m$ from just one video of a person performing mouth manipulations and manually edited it to minimize upper face movements. This involved using a still upper face from the first frame in all subsequent frames, ensuring that the principal components mainly represented mouth movements. Applying PCA of $\mathcal{Z}_m$ reveals that most principal components are responsible for solely mouth movements. For illustration of the first component, see Fig. 8 (bottom row). Based on distilled components, we introduce mouth PCA mouth loss: + +$$ +\mathcal {L} _ {\mathrm {P C A}} \left(v _ {i}, v _ {j}, n\right) = \frac {1}{n} \sum_ {k = 1} ^ {n} \left| \mathrm {P C} _ {\mathbf {v} _ {i}} (k) - \mathrm {P C} _ {\mathbf {v} _ {j}} (k) \right| \tag {6} +$$ + +Where $n$ is the number of principal components to be considered, $\mathrm{PC}_{\mathbf{z}_i}(k)$ and $\mathrm{PC}_{\mathbf{z}_j}(k)$ are the $k$ -th principal components of vectors $\mathbf{z}_i$ and $\mathbf{z}_j$ respectively. Implementing this loss has demonstrably enhanced the quality and accuracy of our results, as our ablation study confirms (see Tab. 6). + +# 6. Experiments + +In this section, we describe experiments we set for comparison for both our image-driven and speech-driven mode. Please see implementation details such as all used losses, component's sizes and architecture details, training procedure, data preprocess in supplementary materials. + +# 6.1. Image-driven comparison + +Methods. We evaluated our image-driven model alongside various one-shot models. For this comparison, we included models like NOFA [32] and StyleHEAT [31], which use 3DMM expression blendshapes without disentanglement training, and models with trainable latent face motion representations like FOMM [22], UVA [23], MetaPortrait [35] and MegaPortraits [9]. It's noteworthy that NOFA and UVA require per-source optimization, unlike the others, including our model. + +Evaluation metrics and data. For our evaluation, we use 100 random, non-child, single-person images from the FFHQ dataset as sources. From the MEAD dataset, we chose 100 images covering all high-intensity emotions except neutral, and another 100 from the FEED dataset's extreme emotion task, featuring 5 identities not seen during training of our model. Each source image was paired with 20 driving images from these sets, totaling 2000 pairs. + +For assessing reenactment quality, we employ various metrics including the Frechet Inception Distance (FID) [13] to measure the distributional discrepancy between synthetic and real images. Cosine similarity (CSIM) from a face recognition network [5] quantifies the identity preservation in generated images. Additionally, we conducted user studies to determine preferences regarding motion (UMTN) and appearance (UAPP) preservation, presenting participants with triplets—either a driving or source image and two method outputs—and asking them to choose the one with better motion or appearance preservation. Our user study includes 34 participants and around 300 unique questions. Both metrics numerically reflect preference percentages. As our visual results Fig. 9, and quantitative metrics Tab. 3 show, our method demonstrates exceptional ability in translating strong and extreme expressions, notably surpassing other image-driven avatar techniques in user preference for facial expression translation (UMTN) and FID scores. + +# 6.2. Speech-driven comparison + +Methods. To assess our model's speech-driven capabilities, we compared it with StyleHEAT [31], which also operates using both image and audio inputs, and audio-exclusive models including SadTalker [36], MakeItTalk [40], and PC-AVS [39]. + +Evaluation metrics and data. We employ the evaluation procedure described in [12]. We measure video realism using Frechet Inception Distance (FID) [13]. We also extract + +
Driver:MEADFEED
MethodFID↓CSIM↑UMTN↑UAPP↑FID↓CSIM↑UMTN↑UAPP↑
MetaPortrait [35]77.50.64**78.80.66**
NOFA [32]69.20.65**69.80.663.88.8
FOMM [22]84.40.562.40.385.90.551.80.5
UVA [17]78.70.680.96.479.50.680.69.4
StyleHEAT [31]71.90.633.73.372.40.612.63.8
MegaPortraits [9]61.10.7322.746.662.80.7316.742.8
Our (EMOPortraits)59.60.7470.343.460.20.7074.634.7
+ +Table 3. Our method notably outperforms others in the FID score and strongly leads in the user preference metric for face expression translation (UMTN). It excels in reliably translating strong and extreme expressions from the driving image, distinguishing it as the most capable among the compared methods. Additionally, our approach maintains the source image's shape and appearance (measured by CSIM and UAPP metrics) on par with MegaPortraits [9] when using strong and regular expressions from the MEAD dataset as the driving signal. However, we observe a minor deviation in identity preservation metrics for extreme emotions (using the FEED dataset). This difference may arise because our method attempts to transfer asymmetric facial expressions, while MegaPortraits might modify the source image less, or not at all, with a challenging driver (as depicted in Fig. 5, case 1), thereby achieving better identity preservation. UVA was excluded from the MEAD-driven user study due to the authors were able to provide only part of the test data. Also, MetaPortrait was excluded from both user studies because of late data provision by authors, offering only the basic model results, without their super-resolution module outputs. + +
MethodFID↓CSIM↑UMTN↓UAPP↓AUCz↓
MegaPortraits fine-tuned62.30.638.512.40.89
Ours w/o LcnCV61.40.679.217.90.88
Ours w/o Lsdm61.80.5911.36.70.85
Ours dim(z) = 51263.40.4425.61.10.77
Ours59.60.7445.461.90.80
+ +Table 4. Our ablation study examines key aspects of the image-driven mode. After fine-tuning MegaPortraits on the FEED dataset, we observed a decline in identity preservation due to overfitting on FEED identities, without any improvement in emotion translation. The lack of $\mathcal{L}_{CV}^{n}$ primarily impacts the ability to translate extreme expression, as indicated by UMTN. Additionally, not using $\mathcal{L}_{sdm}$ or setting $dim(\mathbf{z}) = 512$ results in identity leakage, particularly evident in the latter scenario. For this ablation, we use the part of our driven by MEAD dataset and described in Sec. 6.1 + +facial landmarks from outputs of all methods, frontalized, and normalized to position the mouth edges at $(-1,0)$ and $(1,0)$ . We compute mean absolute error (MAE) for both predicted mouth landmark positions $(M_P)$ and velocities $(M_V)$ , along with errors in facial geometry $(F_P)$ and velocity $(F_V)$ . For our evaluation, we use 100 randomly selected video sequences from the HDTF dataset [37], each up to 30 seconds. As the results show in Tab. 5, our speech-driven model performs on par with leading methods, excelling in realism and facial dynamics. SyncNet [6] is used to assess lip sync quality, providing offset (temporal misalignment between the audio and video) and confidence (certainty of audio-visual alignment) scores. + +![](images/78926baf7339f356df241bf8a88b63a9b83f4b6c3649de1e1ee938d86b29bfe5.jpg) + +![](images/249eb01f4131fdddf201108d1aabb80575d66581bff4a918301a2418cf1ced01.jpg) + +![](images/e6833fe36b387c24e85728fd42f64f6c1ebddbb0976d1cc0ff663463ff80b708.jpg) + +![](images/babf75fa3d53bebbcf1c449f3409b684032610a97ed8f57c97e1cac8db6ce0b7.jpg) + +![](images/69919e8e1c6efec27db2fa17703209ebafd9fb8767da1ce3158618272fad3ff7.jpg) + +![](images/4e727456ce94a10c5cb0cf85dd00e0f157e05f4b06e02049f397c50584441bdd.jpg) + +![](images/40604b2aa996fc6d2acd23a2db2b3411973eea072538acf29f635c8bee37d1b6.jpg) + +![](images/414e09e5425d5e35225c67a5ef070930a8035b17cd7948c4f042cab6758ae06a.jpg) + +![](images/fed734f68f0698bb790e0edb0feec83072cb67a77fe1500b92938e93eb4d9317.jpg) + +![](images/3404ad765f09bb27d600e94b7dfd8336954bec7b7f517c9d336de8e7692620ea.jpg) + +![](images/5d483df94f00a8f21a401b78a622031777a572b55a0ad53f3c54126f06199994.jpg) + +![](images/2f5865bfe29388b182d03ed4d7a9d27dd59feab9fd7f2c0fb1094a9e3561bc51.jpg) + +![](images/514cacf7cd21b690fc389cf3ac74b73b37cf9698264c21cad61806b3e19a0ea3.jpg) + +![](images/6d511aa5d5a396393d77990f87760f332acc794d63f47c209639a816d5ea6435.jpg) + +![](images/70d3e059bb62413c66fd955421186f77044e156d0e45ee6c2b8d47b983c07b88.jpg) + +![](images/18926e0a02b7ae432452adbefe702e8868b7a75fe3646dec0b6b5d9e908bf0d5.jpg) + +![](images/b9020c9c3663e6a403c4af94a5a9fa47af6149c2161bc18ebf0b4b4174ee17e0.jpg) + +![](images/18cada0a23f99cc471fbaa33df513dc3bdc55efddbbf3f720784a087bb0d7786.jpg) + +![](images/abdd874750024f19fafebca20a06b1f70aa928706d99df413ba354979d75ba5c.jpg) + +![](images/8c349ae9a21dac090ccda0bff743883591aeeca9e5450b427eac42b730eaa9f6.jpg) + +![](images/3c10c4d85aa3cca3c2b78f3e2c34e7a1a4bfbaa64af74b33787d9f6fe596d827.jpg) + +![](images/17e23469abdc5c7ab7a6a5550d1010eb601a5984b5074e13f879ce44ad308ecd.jpg) + +![](images/6a6831ddbb77f1d43713ce4c0cf619c551296f51f84ded434ab7b7241a9f5c62.jpg) + +![](images/fd0efa99b095cb5c8d77d101660f6c7bb18b5ee8d5d2d7cb16d3cd628893f63d.jpg) + +![](images/73780beddb0fb11c3d6d1430fac4b29508bdfa7ada34735177e0e1aa0d8d7c4c.jpg) + +![](images/be464290184a7120b5a3faa800bfa8fea8bed08c5546356fdabcb1efb9302981.jpg) + +![](images/455b641de64cefbaf69d4b86b1a3d711d7191a180fdb1111b6aad142ca9b34d4.jpg) + +![](images/c20f63e03b90141305ab3f748fe6e81e7db5f7761224f2e498b225b5bd0fbdb8.jpg) +Source + +![](images/914f03cb4787c3219044fec13bdedab69d72544c0aaf586e94c11589a47cc118.jpg) +Driver + +![](images/2a4a11262eaefe3893566bd9f65ab0cf40f2dd90dbc5eb22abf893187532cf4e.jpg) +FOMM [22] + +![](images/167c2eae9f425be2922aad926fe782a8fae893bef8708cefbea97fb2b42f3c59.jpg) +UVA [17] + +![](images/8a024063cda079847e308c3b84df1cd3d92d6782b38b156413a97c18dec990a1.jpg) +StyleHEAT [31] + +![](images/ca68c0f5dd22b90c2d4067ccb3bef63902c3770d8b9a7a684fa690295510ba96.jpg) +MetaPortrait [35] +Figure 9. A qualitative comparison of head avatar systems in cross-reenactment scenario. + +![](images/2f6050cbd5ae51b790b8714d4b1c4c863640bf93ba51abf7f465583f4c63ce65.jpg) +NOFA [32] + +![](images/a60b6865628d9d29031984af0b311f49aa111e0d10c11263007409dd4919132c.jpg) +MegaPortraits [9] + +![](images/5cd2ec3d01fca131da0568ab9ce41e09a269a4f1cdc110e312097c112488ba8f.jpg) +Ours + +
MethodFID↓\( F_P \downarrow \)\( F_V \downarrow \)\( M_P \downarrow \)\( M_V \downarrow \)Sync. offset ↓Sync. conf. ↑
StyleHEAT [31]62.41.340.412.981.670.417.22
PC-AVS [39]71.52.780.512.561.043.798.42
SadTalker [36]55.60.910.352.121.170.277.58
MakeIRTalk [40]63.32.150.383.111.592.985.17
Ours (EMOPortraits)28.50.820.332.531.380.075.77
+ +Table 5. Our model matches other leading speech-driven models, exceeding in realism (FID), facial geometry and velocity, but is slightly less accurate in mouth landmark positions and velocities. We also score best in audio-video sync offset, but lower in confidence. We believe that while SyncNet metrics are still helpful, they may not fully represent true synchronization quality, as generated videos can sometimes score higher than ground truth without necessarily being better synchronized. In our experimentation, ground truth videos scored an offset of 1.13 (which is worse than our method, SadTalker, and StyleHEAT) and a confidence score of 8.35 (less than PC-AVS). For qualitative analysis, please refer to our supplementary materials. + +
MethodFID↓MP↓MV↓FP↓FV↓
Ours w/o disentanglement44.70.980.455.782.55
Ours w/o LPCA29.30.840.353.131.88
Ours with MAE(\( \mathbf{z}_i, \hat{\mathbf{z}}_i^{\text{speech}} \))28.20.840.342.611.45
Ours28.50.820.332.531.38
+ +Table 6. Ablation study of our speech-driven mode, where in the first line, we show that it notably underperforms without head pose - face expression disentanglement of the latent space (see Sec. 5.1). Removing $\mathcal{L}_{\mathrm{PCA}}$ leads to a significant reduction in generating realistic mouth movements. Replacing $\mathcal{L}_{\mathrm{PCA}}$ with $\mathrm{MAE}(\mathbf{z}_i,\hat{\mathbf{z}}_i^{\mathrm{speech}})$ causes a noticeable, decline in mouth movement generation. This suggests that identifying mouth-related PCA components was beneficial in enhancing the outcome. + +# 6.3. Ablation study + +We conducted an extensive ablation study to evaluate the contributions of individual components within our method. For our main model, we present the evaluation of the importance of the proposed source-driver unmatched loss, canonical volume loss, and change of $dim(\mathbf{z})$ from 512 in original model to 128 Tab. 4. For our speech-driven mode, we show how vital the disentanglement of the latent space for speech-drive ability, how $\mathcal{L}_{\mathrm{PCA}}$ is better than naive matching of PCA components from random $\mathcal{Z}$ and how it helps match speech and latent expression space Tab. 6. For more ablation analysis, plots and visual comparison, please refer to our supplementary. + +# 7. Conclusion + +In this paper, we introduce EmoPortraits, a novel method for creating neural avatars with superior performance in image-driven, cross-identity emotion translation. Our speech-driven mode makes it possible to drive the facial animation through multiple conditions (video, audio, head motion). We collected FEED dataset which, we believe, will be a valuable asset for researchers in diverse human-centered studies. + +However, our method has some limitations. It doesn't generate the avatar's body or shoulders, limiting some use cases. We currently integrate our output with a source image body. Additionally, the model sometimes struggles with accurate expression translation and underperforms with extensive head rotation. These challenges are crucial for future enhancements and remain central to our ongoing research efforts. + +# References + +[1] Emad Barsoum, Cha Zhang, Cristian Canton Ferrer, and Zhengyou Zhang. Training deep networks for facial expression recognition with crowd-sourced label distribution. In Proceedings of the 18th ACM international conference on multimodal interaction, pages 279-283, 2016. 3 +[2] Stella Bounareli, Christos Tzelepis, Vasileios Argyriou, Ioannis Patras, and Georgios Tzimiropoulos. Hyperreenact: One-shot reenactment via jointly learning to refine and retarget faces, 2023. 2 +[3] Egor Burkov, Igor Pasechnik, Artur Grigorev, and Victor Lempitsky. Neural head reenactment with latent pose descriptors. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). IEEE, 2020. 2 +[4] Houwei Cao, David G Cooper, Michael K Keutmann, Ruben C Gur, Ani Nenkova, and Ragini Verma. Crema-d: Crowd-sourced emotional multimodal actors dataset. IEEE transactions on affective computing, 5(4):377-390, 2014. 3 +[5] Qiong Cao, Li Shen, Weidi Xie, Omkar M. Parkhi, and Andrew Zisserman. Vggface2: A dataset for recognising faces across pose and age, 2018. 7 +[6] Joon Son Chung and Andrew Zisserman. Out of time: automated lip sync in the wild. In Computer Vision-ACCV 2016 Workshops: ACCV 2016 International Workshops, Taipei, Taiwan, November 20-24, 2016, Revised Selected Papers, Part II 13, pages 251-263. Springer, 2017. 7 +[7] Joon Son Chung, Arsha Nagrani, and Andrew Zisserman. Voxceleb2: Deep speaker recognition. In INTERSPEECH, 2018. 4 +[8] Michail Christos Doukas, Stefanos Zafeiriou, and Viktoriia Sharmanska. Headgan: One-shot neural head synthesis and editing, 2021. 2 +[9] Nikita Drobyshev, Jenya Chelishev, Taras Khakhulin, Aleksei Ivakhnenko, Victor Lempitsky, and Egor Zakharov. Megaportraits: One-shot megapixel neural head avatars, 2023. 2, 4, 5, 7, 8 +[10] Guy Gafni, Justus Thies, Michael Zollhöfer, and Matthias Nießner. Dynamic neural radiance fields for monocular 4d facial avatar reconstruction, 2020. 2 +[11] Philip-William Grassal, Malte Prinzler, Titus Leistner, Carsten Rother, Matthias Nießner, and Justus Thies. Neural head avatars from monocular rgb videos, 2022. 2 +[12] Siddharth Gururani, Arun Mallya, Ting-Chun Wang, Rafael Valle, and Ming-Yu Liu. Space: Speech-driven portrait animation with controllable expression, 2022. 3, 7 +[13] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. In Advances in Neural Information Processing Systems, 2017. 7 +[14] Philip Jackson and SJUoSG Haq. Surrey audio-visual expressed emotion (savee) database. University of Surrey: Guildford, UK, 2014. 3 +[15] Taras Khakhulin, Vanessa Sklyarova, Victor Lempitsky, and Egor Zakharov. Realistic one-shot mesh-based head avatars, 2022. 2 + +[16] Alexander C. Li, Alexei A. Efros, and Deepak Pathak. Understanding collapse in non-contrastive siamese representation learning, 2022. 4 +[17] Xueting Li, Shalini De Mello, Sifei Liu, Koki Nagano, Umar Iqbal, and Jan Kautz. Generalizable one-shot neural head avatar, 2023. 7, 8 +[18] Steven R Livingstone and Frank A Russo. The ryerson audiovisual database of emotional speech and song (ravdess): A dynamic, multimodal set of facial and vocal expressions in north american english. PloS one, 13(5):e0196391, 2018. 3 +[19] Patrick Lucey, Jeffrey F Cohn, Takeo Kanade, Jason Saragih, Zara Ambadar, and Iain Matthews. The extended cohn-kanade dataset (ck+): A complete dataset for action unit and emotion-specified expression. In 2010 iiee computer society conference on computer vision and pattern recognition-workshops, pages 94-101. IEEE, 2010. 3 +[20] Maja Pantic, Michel Valstar, Ron Rademaker, and Ludo Maat. Web-based database for facial expression analysis. In 2005 IEEE international conference on multimedia and Expo, pages 5-pp. IEEE, 2005. 3 +[21] KR Prajwal, Rudrabha Mukhopadhyay, Vinay P Namboodiri, and CV Jawahar. A lip sync expert is all you need for speech to lip generation in the wild. In Proceedings of the 28th ACM international conference on multimedia, pages 484-492, 2020. 3 +[22] Aliaksandr Siarohin, Stéphane Lathuilière, S. Tulyakov, Elisa Ricci, and N. Sebe. First order motion model for image animation. ArXiv, abs/2003.00196, 2019. 2, 7, 8 +[23] Aliaksandr Siarohin, Willi Menapace, Ivan Skorokhodov, Kyle Olszewski, Hsin-Ying Lee, Jian Ren, Mengelei Chai, and Sergey Tulyakov. Unsupervised volumetric animation. arXiv preprint arXiv:2301.11326, 2023. 7 +[24] Michal Stypulkowski, Konstantinos Vougioukas, Sen He, Maciej Zieba, Stavros Petridis, and Maja Pantic. Diffused heads: Diffusion models beat gans on talking-face generation, 2023. 3 +[25] Mani Kumar Tellamekala, Ömer Ümer, Björn W. Schuller, Elisabeth Andre, Timo Giesbrecht, and Michel Valstar. Are 3d face shapes expressive enough for recognising continuous emotions and action unit intensities?, 2023. 2 +[26] Hao Wang, Yitong Wang, Zheng Zhou, Xing Ji, Dihong Gong, Jingchao Zhou, Zhifeng Li, and Wei Liu. Cosface: Large margin cosine loss for deep face recognition, 2018. 5 +[27] H. Wang, Yitong Wang, Zheng Zhou, Xing Ji, Zhifeng Li, Dihong Gong, Jin Zhou, and Wenyu Liu. Cosface: Large margin cosine loss for deep face recognition. 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5265-5274, 2018. 6 +[28] Kaisiyuan Wang, Qianyi Wu, Linsen Song, Zhuoqian Yang, Wayne Wu, Chen Qian, Ran He, Yu Qiao, and Chen Change Loy. Mead: A large-scale audio-visual dataset for emotional talking-face generation. In European Conference on Computer Vision, pages 700–717. Springer, 2020. 3 +[29] Ting-Chun Wang, Arun Mallya, and Ming-Yu Liu. One-shot free-view neural talking-head synthesis for video conferencing. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2021. 2 + +[30] Yuelang Xu, Lizhen Wang, Xiaochen Zhao, Hongwen Zhang, and Yebin Liu. AvatarMAV: Fast 3d head avatar reconstruction using motion-aware neural voxels. In *Special Interest Group on Computer Graphics and Interactive Techniques Conference Conference Proceedings*. ACM, 2023. 2 +[31] Fei Yin, Yong Zhang, Xiaodong Cun, Mingdeng Cao, Yanbo Fan, Xuan Wang, Qingyan Bai, Baoyuan Wu, Jue Wang, and Yujiu Yang. Styleheat: One-shot high-resolution editable talking face generation via pre-trained stylegan, 2022. 2, 3, 7, 8 +[32] Wangbo Yu, Yanbo Fan, Yong Zhang, Xuan Wang, Fei Yin, Yunpeng Bai, Yan-Pei Cao, Ying Shan, Yang Wu, Zhongqian Sun, and Baoyuan Wu. Nofa: Nerf-based one-shot facial avatar reconstruction, 2023. 2, 7, 8 +[33] Egor Zakharov, Aliaksandra Shysheya, Egor Burkov, and Victor Lempitsky. Few-shot adversarial learning of realistic neural talking head models, 2019. 2 +[34] Egor Zakharov, Aleksei Ivakhnenko, Aliaksandra Shysheya, and Victor Lempitsky. Fast bi-layer neural synthesis of one-shot realistic head avatars, 2020. 2 +[35] Bowen Zhang, Chenyang Qi, Pan Zhang, Bo Zhang, Hsiang-Tao Wu, Dong Chen, Qifeng Chen, Yong Wang, and Fang Wen. Metaportrait: Identity-preserving talking head generation with fast personalized adaptation, 2023. 2, 7, 8 +[36] Wenxuan Zhang, Xiaodong Cun, Xuan Wang, Yong Zhang, Xi Shen, Yu Guo, Ying Shan, and Fei Wang. Sadtalker: Learning realistic 3d motion coefficients for stylized audio-driven single image talking face animation, 2023. 3, 7, 8 +[37] Zhimeng Zhang, Lincheng Li, Yu Ding, and Changjie Fan. Flow-guided one-shot talking face generation with a high-resolution audio-visual dataset. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3661–3670, 2021. 7 +[38] Yufeng Zheng, Victoria Fernández Abrevaya, Marcel C. Bühler, Xu Chen, Michael J. Black, and Otmar Hilliges. I m avatar: Implicit morphable head avatars from videos, 2022. 2 +[39] Hang Zhou, Yasheng Sun, Wayne Wu, Chen Change Loy, Xiaogang Wang, and Ziwei Liu. Pose-controllable talking face generation by implicitly modularized audio-visual representation, 2021. 2, 3, 7, 8 +[40] Yang Zhou, Xintong Han, Eli Shechtman, Jose Echevarria, Evangelos Kalogerakis, and Dingzeyu Li. MakeltTalk. ACM Transactions on Graphics, 39(6):1-15, 2020. 3, 7, 8 \ No newline at end of file diff --git a/2024/EMOPortraits_ Emotion-enhanced Multimodal One-shot Head Avatars/images.zip b/2024/EMOPortraits_ Emotion-enhanced Multimodal One-shot Head Avatars/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..33800ff9d3a01f2948325e0b87df6384d11a4b24 --- /dev/null +++ b/2024/EMOPortraits_ Emotion-enhanced Multimodal One-shot Head Avatars/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c295bbdd840ed661f28e73a7687010051a1f3e519cdd975a1b149536af22ed86 +size 672662 diff --git a/2024/EMOPortraits_ Emotion-enhanced Multimodal One-shot Head Avatars/layout.json b/2024/EMOPortraits_ Emotion-enhanced Multimodal One-shot Head Avatars/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..81ad9b161386804a765dae1454e5a9ddeba155ed --- /dev/null +++ b/2024/EMOPortraits_ Emotion-enhanced Multimodal One-shot Head Avatars/layout.json @@ -0,0 +1,11072 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 77, + 103, + 517, + 119 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 103, + 517, + 119 + ], + "spans": [ + { + "bbox": [ + 77, + 103, + 517, + 119 + ], + "type": "text", + "content": "EMOPortraits: Emotion-enhanced Multimodal One-shot Head Avatars" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 143, + 141, + 158 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 143, + 141, + 158 + ], + "spans": [ + { + "bbox": [ + 53, + 143, + 141, + 158 + ], + "type": "text", + "content": "Nikita Drobyshev" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 166, + 144, + 298, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 144, + 298, + 171 + ], + "spans": [ + { + "bbox": [ + 166, + 144, + 298, + 171 + ], + "type": "text", + "content": "Antoni Bigata Casademunt \nStavros Petridis" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 318, + 144, + 446, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 144, + 446, + 171 + ], + "spans": [ + { + "bbox": [ + 318, + 144, + 446, + 171 + ], + "type": "text", + "content": "Konstantinos Vougioukas \nMaja Pantic" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 471, + 144, + 539, + 158 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 471, + 144, + 539, + 158 + ], + "spans": [ + { + "bbox": [ + 471, + 144, + 539, + 158 + ], + "type": "text", + "content": "Zoe Landgraf" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 135, + 174, + 460, + 214 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 174, + 460, + 214 + ], + "spans": [ + { + "bbox": [ + 135, + 174, + 460, + 214 + ], + "type": "text", + "content": "{n.drobyshev, a/bigata-casademunt22, k.vougioukas, zoe.landgraf15, stavros.petridis04, m.pantic}@imperial.ac.uk Imperial College London" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 75, + 241, + 166, + 330 + ], + "blocks": [ + { + "bbox": [ + 75, + 241, + 166, + 330 + ], + "lines": [ + { + "bbox": [ + 75, + 241, + 166, + 330 + ], + "spans": [ + { + "bbox": [ + 75, + 241, + 166, + 330 + ], + "type": "image", + "image_path": "5571c03b59a22946cb276d1deded7196a8c084b040541ce529bbee5da0acb26d.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 165, + 242, + 254, + 330 + ], + "blocks": [ + { + "bbox": [ + 165, + 242, + 254, + 330 + ], + "lines": [ + { + "bbox": [ + 165, + 242, + 254, + 330 + ], + "spans": [ + { + "bbox": [ + 165, + 242, + 254, + 330 + ], + "type": "image", + "image_path": "6f78ae2acda6416384be5581a2bc2ae87413212cd8bfa3feb9df5869a0f8e39f.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 254, + 242, + 343, + 330 + ], + "blocks": [ + { + "bbox": [ + 254, + 242, + 343, + 330 + ], + "lines": [ + { + "bbox": [ + 254, + 242, + 343, + 330 + ], + "spans": [ + { + "bbox": [ + 254, + 242, + 343, + 330 + ], + "type": "image", + "image_path": "867efe1bddaf1887f13140417c41e46dc4c66e055a39bb4c7e189e5a87cba5b5.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 343, + 242, + 431, + 330 + ], + "blocks": [ + { + "bbox": [ + 343, + 242, + 431, + 330 + ], + "lines": [ + { + "bbox": [ + 343, + 242, + 431, + 330 + ], + "spans": [ + { + "bbox": [ + 343, + 242, + 431, + 330 + ], + "type": "image", + "image_path": "c12038d2254398ecb4c7ba204508b41706cc611de0c85c955e62dc16a81d8eb4.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 431, + 242, + 517, + 330 + ], + "blocks": [ + { + "bbox": [ + 431, + 242, + 517, + 330 + ], + "lines": [ + { + "bbox": [ + 431, + 242, + 517, + 330 + ], + "spans": [ + { + "bbox": [ + 431, + 242, + 517, + 330 + ], + "type": "image", + "image_path": "272336b8074a1488e724b3c0fbb531059f7b99cb096e5d5bc5ee91f1570dcb0f.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 75, + 333, + 165, + 418 + ], + "blocks": [ + { + "bbox": [ + 75, + 333, + 165, + 418 + ], + "lines": [ + { + "bbox": [ + 75, + 333, + 165, + 418 + ], + "spans": [ + { + "bbox": [ + 75, + 333, + 165, + 418 + ], + "type": "image", + "image_path": "e1eb172f4eb38c6ece771d39f7a6007b8dcd48c8729563a1872fa82a97e69526.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 165, + 333, + 253, + 418 + ], + "blocks": [ + { + "bbox": [ + 165, + 333, + 253, + 418 + ], + "lines": [ + { + "bbox": [ + 165, + 333, + 253, + 418 + ], + "spans": [ + { + "bbox": [ + 165, + 333, + 253, + 418 + ], + "type": "image", + "image_path": "d5150c201a58cf952e956203832059c28533b90018e0a66348c1f3f500048536.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 253, + 333, + 343, + 418 + ], + "blocks": [ + { + "bbox": [ + 253, + 333, + 343, + 418 + ], + "lines": [ + { + "bbox": [ + 253, + 333, + 343, + 418 + ], + "spans": [ + { + "bbox": [ + 253, + 333, + 343, + 418 + ], + "type": "image", + "image_path": "316eae8164a7f14a42016f25d69f3acd05748b74e70f6c167f52db9458ece090.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 343, + 333, + 428, + 418 + ], + "blocks": [ + { + "bbox": [ + 343, + 333, + 428, + 418 + ], + "lines": [ + { + "bbox": [ + 343, + 333, + 428, + 418 + ], + "spans": [ + { + "bbox": [ + 343, + 333, + 428, + 418 + ], + "type": "image", + "image_path": "479f825123fe7a77d8b249bb5eb77206881d58fc1981346f540388e496db4655.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 428, + 333, + 517, + 418 + ], + "blocks": [ + { + "bbox": [ + 428, + 333, + 517, + 418 + ], + "lines": [ + { + "bbox": [ + 428, + 333, + 517, + 418 + ], + "spans": [ + { + "bbox": [ + 428, + 333, + 517, + 418 + ], + "type": "image", + "image_path": "8da0d6fcd24f450ef1006655f2860be5194a1760923b15e36c23cfbcd55b3cd1.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 75, + 420, + 165, + 509 + ], + "blocks": [ + { + "bbox": [ + 75, + 420, + 165, + 509 + ], + "lines": [ + { + "bbox": [ + 75, + 420, + 165, + 509 + ], + "spans": [ + { + "bbox": [ + 75, + 420, + 165, + 509 + ], + "type": "image", + "image_path": "1d3be78635c31aa04786c174e323af741e93526bbfde5ad9472c0a7afda2dfaf.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 105, + 510, + 134, + 518 + ], + "lines": [ + { + "bbox": [ + 105, + 510, + 134, + 518 + ], + "spans": [ + { + "bbox": [ + 105, + 510, + 134, + 518 + ], + "type": "text", + "content": "Source" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 165, + 420, + 253, + 509 + ], + "blocks": [ + { + "bbox": [ + 165, + 420, + 253, + 509 + ], + "lines": [ + { + "bbox": [ + 165, + 420, + 253, + 509 + ], + "spans": [ + { + "bbox": [ + 165, + 420, + 253, + 509 + ], + "type": "image", + "image_path": "a2ab0f9b76c3a8b179b604978468a109567c35f2493df4b12487308d887b0ea9.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 203, + 510, + 227, + 517 + ], + "lines": [ + { + "bbox": [ + 203, + 510, + 227, + 517 + ], + "spans": [ + { + "bbox": [ + 203, + 510, + 227, + 517 + ], + "type": "text", + "content": "Driver" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 253, + 420, + 343, + 509 + ], + "blocks": [ + { + "bbox": [ + 253, + 420, + 343, + 509 + ], + "lines": [ + { + "bbox": [ + 253, + 420, + 343, + 509 + ], + "spans": [ + { + "bbox": [ + 253, + 420, + 343, + 509 + ], + "type": "image", + "image_path": "7af6d1668fe3d655129ce10d232c0993eb2693facbf0ba6bfeccb40e09db4e40.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 285, + 510, + 312, + 518 + ], + "lines": [ + { + "bbox": [ + 285, + 510, + 312, + 518 + ], + "spans": [ + { + "bbox": [ + 285, + 510, + 312, + 518 + ], + "type": "text", + "content": "Output" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 343, + 420, + 428, + 509 + ], + "blocks": [ + { + "bbox": [ + 343, + 420, + 428, + 509 + ], + "lines": [ + { + "bbox": [ + 343, + 420, + 428, + 509 + ], + "spans": [ + { + "bbox": [ + 343, + 420, + 428, + 509 + ], + "type": "image", + "image_path": "37fa260d97e9e5d2b12fdb7dff66ab9aa5c86ba957ebcccf9309e28ed674827c.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 375, + 510, + 400, + 518 + ], + "lines": [ + { + "bbox": [ + 375, + 510, + 400, + 518 + ], + "spans": [ + { + "bbox": [ + 375, + 510, + 400, + 518 + ], + "type": "text", + "content": "Driver" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 428, + 420, + 517, + 509 + ], + "blocks": [ + { + "bbox": [ + 428, + 420, + 517, + 509 + ], + "lines": [ + { + "bbox": [ + 428, + 420, + 517, + 509 + ], + "spans": [ + { + "bbox": [ + 428, + 420, + 517, + 509 + ], + "type": "image", + "image_path": "6f169bc4bce76438b2a929bd4d15373bc33cedd96bc3fcaca9948d0a78a551f7.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 460, + 510, + 487, + 518 + ], + "lines": [ + { + "bbox": [ + 460, + 510, + 487, + 518 + ], + "spans": [ + { + "bbox": [ + 460, + 510, + 487, + 518 + ], + "type": "text", + "content": "Output" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 159, + 526, + 433, + 538 + ], + "lines": [ + { + "bbox": [ + 159, + 526, + 433, + 538 + ], + "spans": [ + { + "bbox": [ + 159, + 526, + 433, + 538 + ], + "type": "text", + "content": "Figure 1. Selected animation results for image-driven mode of our method." + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_caption" + } + ], + "index": 26 + }, + { + "bbox": [ + 143, + 548, + 192, + 560 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 548, + 192, + 560 + ], + "spans": [ + { + "bbox": [ + 143, + 548, + 192, + 560 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 46, + 570, + 289, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 570, + 289, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 570, + 289, + 713 + ], + "type": "text", + "content": "Head avatars animated by visual signals have gained popularity, particularly in cross-driving synthesis where the driver differs from the animated character, a challenging but highly practical approach. The recently presented MegaPortraits model has demonstrated state-of-the-art results in this domain. We conduct a deep examination and evaluation of this model, with a particular focus on its latent space for facial expression descriptors, and uncover several limitations with its ability to express intense face motions. To address these limitations, we propose substantial changes in both training pipeline and model architecture, to introduce our EMOPortraits model, where we:" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 304, + 564, + 547, + 612 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 564, + 547, + 612 + ], + "spans": [ + { + "bbox": [ + 304, + 564, + 547, + 612 + ], + "type": "text", + "content": "Enhance the model's capability to faithfully support intense, asymmetric face expressions, setting a new state-of-the-art result in the emotion transfer task, surpassing previous methods in both metrics and quality." + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 304, + 615, + 547, + 663 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 615, + 547, + 663 + ], + "spans": [ + { + "bbox": [ + 304, + 615, + 547, + 663 + ], + "type": "text", + "content": "Incorporate speech-driven mode to our model, achieving top-tier performance in audio-driven facial animation, making it possible to drive source identity through diverse modalities, including visual signal, audio, or a blend of both." + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 304, + 666, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 666, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 666, + 547, + 713 + ], + "type": "text", + "content": "Furthermore, we propose a novel multi-view video dataset featuring a wide range of intense and asymmetric facial expressions, filling the gap with absence of such data in existing datasets." + } + ] + } + ], + "index": 33 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8498" + } + ] + } + ], + "index": 34 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 127, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 127, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 127, + 83 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 232, + 104 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 91, + 232, + 104 + ], + "spans": [ + { + "bbox": [ + 48, + 91, + 232, + 104 + ], + "type": "text", + "content": "1.1. Representation of facial expression" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 110, + 287, + 252 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 110, + 287, + 252 + ], + "spans": [ + { + "bbox": [ + 47, + 110, + 287, + 252 + ], + "type": "text", + "content": "Advancements in neural head technologies now enable the creation of realistic avatars from a few or even one image, with the latter being crucial when only one source image is available. Cross-driving synthesis, where avatars are animated with different identities is a key technique for virtual reality, filmmaking, photo animation, etc. However, accurate transfer of facial expressions, especially those that are intense and uneven, remains a substantial challenge in avatar animation, particularly for cross-driving synthesis. While previous research focused mainly on preserving identity and transferring moderate facial motions, our work also seeks to accurately drive high-intensity and asymmetric expressions." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 254, + 288, + 432 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 254, + 288, + 432 + ], + "spans": [ + { + "bbox": [ + 47, + 254, + 288, + 432 + ], + "type": "text", + "content": "Our research is built on and extends the MegaPortraits [9] model, notable for state-of-the-art results in cross-driving synthesis. An additional advantage using this method is that, unlike many avatar systems that depend on limited [3, 25] predefined motion descriptors, MegaPortraits, learns expression representations from scratch, allowing for greater adaptability to a wider range of expressions. We investigate the latent expression spaces and training methods of MegaPortraits to enhance its ability to depict a broad spectrum of facial expressions. Our comprehensive analysis reveals that, while the original model shows limited effectiveness in intense motion representing, it has significant potential for improvement through targeted architectural modifications, adjustments in the training approach, and the integration of our novel dataset." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 441, + 193, + 453 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 441, + 193, + 453 + ], + "spans": [ + { + "bbox": [ + 47, + 441, + 193, + 453 + ], + "type": "text", + "content": "1.2. Integrating speech driving" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 460, + 287, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 460, + 287, + 506 + ], + "spans": [ + { + "bbox": [ + 47, + 460, + 287, + 506 + ], + "type": "text", + "content": "We integrate speech in our model to be used either as a complementary or an alternative driver, a crucial aspect for applications like virtual assistance or mixed reality when a primary visual signal is absent." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 508, + 287, + 615 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 508, + 287, + 615 + ], + "spans": [ + { + "bbox": [ + 47, + 508, + 287, + 615 + ], + "type": "text", + "content": "By improving disentanglement of facial expression latent space, we remove disturbing information from face motion descriptors and emphasize components that solely control lip movements. Based on this, we formulated a novel loss that helps to achieve desirable results. Furthermore, our method can generate plausible head rotations and blinks, which appear natural and enhance its applicability across various tasks. Thus, our final method can drive source identity through image, audio, or a mix of them." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 623, + 261, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 623, + 261, + 635 + ], + "spans": [ + { + "bbox": [ + 47, + 623, + 261, + 635 + ], + "type": "text", + "content": "1.3. FEED: Facial Extreme Emotions Dataset" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 642, + 287, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 642, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 47, + 642, + 287, + 712 + ], + "type": "text", + "content": "Public data scarcity is a common obstacle in deep learning, the topic of human facial expressions and movements is not an exception. In particular, we note a lack of high quality video data capturing a wide range of facial expressions, with existing datasets not going beyond basic facial actions shown at Fig. 2. To address this, we collect a new dataset that" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 72, + 547, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 167 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 167 + ], + "type": "text", + "content": "includes basic expressions and also captures complex movements like blinks, winks, and head and tongue movements, along with varied extreme expressions which are difficult or impossible to categorize through basic facial actions. Given this variety in expressions in our dataset, we believe that it will be a valuable resource for research in human emotions and facial reconstruction fields. In summary, our main contributions are as follows:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 177, + 545, + 356 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 306, + 177, + 545, + 260 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 177, + 545, + 260 + ], + "spans": [ + { + "bbox": [ + 306, + 177, + 545, + 260 + ], + "type": "text", + "content": "- We introduce our new EMOPortraits model for one-shot head avatars synthesis, that is capable of transferring intense facial expression, showing state-of-the-art results in cross-driving synthesis. We achieve this through careful latent facial expression space development as well as novel losses and a minimal amount of domain-specific data (" + }, + { + "bbox": [ + 306, + 177, + 545, + 260 + ], + "type": "inline_equation", + "content": "\\sim" + }, + { + "bbox": [ + 306, + 177, + 545, + 260 + ], + "type": "text", + "content": " 0.1% of the train set)." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 261, + 545, + 320 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 261, + 545, + 320 + ], + "spans": [ + { + "bbox": [ + 306, + 261, + 545, + 320 + ], + "type": "text", + "content": "- We integrate a speech-driving mode in our model, that demonstrates cutting-edge results in speech-driven animations. It functions effectively alongside visual signals or independently, also generating realistic head rotations and eye blinks." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 320, + 545, + 356 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 320, + 545, + 356 + ], + "spans": [ + { + "bbox": [ + 306, + 320, + 545, + 356 + ], + "type": "text", + "content": "- We present a unique multi-view dataset that spans a broad spectrum of extreme facial expressions, filling the gap of absence of such data in existing datasets." + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 377, + 392, + 389 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 377, + 392, + 389 + ], + "spans": [ + { + "bbox": [ + 306, + 377, + 392, + 389 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 401, + 454, + 412 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 401, + 454, + 412 + ], + "spans": [ + { + "bbox": [ + 306, + 401, + 454, + 412 + ], + "type": "text", + "content": "2.1. Visual-driven head avatars" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 422, + 547, + 530 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 422, + 547, + 530 + ], + "spans": [ + { + "bbox": [ + 304, + 422, + 547, + 530 + ], + "type": "text", + "content": "In recent years, the domain of neural avatars has branched into two prominent sub-fields: person-specific and person-agnostic avatars. While person-specific methodologies [10, 11, 30, 38, 39] excel in delivering stunning realism and motion fidelity for a particular individual, they encounter challenges representing an arbitrary person. Moreover, these approach demands multiple frames of training data, distinct training for each avatar, and can struggle to replicate motions not encountered during training." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 534, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 534, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 534, + 547, + 713 + ], + "type": "text", + "content": "Person-agnostic methods, on the other hand, don't need train or fine-tuning for each new person and present an alternative approach to talking-head synthesis. Earlier works in this domain generated avatars in a few-shot technique [3, 33], while subsequent studies introduced one-shot capabilities [2, 8, 9, 22, 29, 31, 32, 34]. Many of these works employ predefined motion representations, such as 3DMM's blendshapes [8, 15, 31, 32]. In contrast, some learned latent emotion representation from scratch [3, 9, 22, 34, 35]. This latter approach holds potential for better motion representation as in such settings, expression descriptors become entirely trainable and free from inheriting the limitations typical for blendshapes as was shown in [3, 25]. In our work, we adopt this strategy, remarkably improving upon the methodology used in MegaPortraits[9]." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8499" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 55, + 70, + 280, + 101 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 280, + 101 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 280, + 101 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 280, + 101 + ], + "type": "table", + "html": "
MethodBFEsFFED dataset internal structure
Head rotationsWinksEyes moveAsym. em.Tongue em.Extreme em.
ModeMild/StrongAxes/randomMild/StrongRandomStrongStrongExtreme
# of part.21/23212321232323
Avg. length1:38/1.520:580:490:331:411:533:02
", + "image_path": "66a3d422459b2df9f0a60e5e7a3cd577b47d5f2fcd46c4308a5bbed92a3374f7.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 55, + 155, + 280, + 204 + ], + "blocks": [ + { + "bbox": [ + 47, + 108, + 287, + 141 + ], + "lines": [ + { + "bbox": [ + 47, + 108, + 287, + 141 + ], + "spans": [ + { + "bbox": [ + 47, + 108, + 287, + 141 + ], + "type": "text", + "content": "Table 1. The tasks we asked our participants to perform for our FEED dataset. Here BFEs stands for basic facial expressions, shown in Fig. 2" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 55, + 155, + 280, + 204 + ], + "lines": [ + { + "bbox": [ + 55, + 155, + 280, + 204 + ], + "spans": [ + { + "bbox": [ + 55, + 155, + 280, + 204 + ], + "type": "table", + "html": "
DatasetDatasets comparison
#Actors#ViewsBasic exp.Strong exp.Extrme & Assym. exp.Tongue exp.Resolution
SAVESS41✓✘✘✘1280 x 1024
RAVGESS241✓✘✘✘1920 x 1080
CREMA-D911✓✘✘✘1280 x 720
MMI251✓✓✘✘1920 x 1080
MEAD487✓✓✘✘1920 x 1080
Ours (FEED)233✓✓✓✓3840 x 2160
", + "image_path": "578c2462fd48de5b379bc63eeefef14e809e14ca69ec9aef8856235a65654fea.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 212, + 286, + 234 + ], + "lines": [ + { + "bbox": [ + 47, + 212, + 286, + 234 + ], + "spans": [ + { + "bbox": [ + 47, + 212, + 286, + 234 + ], + "type": "text", + "content": "Table 2. Comparisons with modern, high-quality audio-visual datasets created in controlled settings." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 47, + 257, + 185, + 269 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 257, + 185, + 269 + ], + "spans": [ + { + "bbox": [ + 47, + 257, + 185, + 269 + ], + "type": "text", + "content": "2.2. Face expression datasets" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 276, + 287, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 276, + 287, + 323 + ], + "spans": [ + { + "bbox": [ + 47, + 276, + 287, + 323 + ], + "type": "text", + "content": "Early image datasets[1, 19] predominantly offered annotated face expression data corresponded to up to 8 basic emotions. However, these datasets are not well-suited for training head avatars, which necessitate video or multi-view data." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 324, + 288, + 492 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 324, + 288, + 492 + ], + "spans": [ + { + "bbox": [ + 46, + 324, + 288, + 492 + ], + "type": "text", + "content": "The SAVEE [14] dataset captures facial expressions in speech, but involves just 4 actors. The MMI [20] dataset, more expansive in actor participation, offers spontaneous moderate-intensity expression but is restricted to single-view sequences. RAVDESS [18] distinguishes itself by capturing two motion intensities, but its limited recordings challenge broad applicability. CREMA-D [4] encompasses three expression intensities, and though MEAD [28] provides detailed multi-view data across three intensities, it's limited to the standard eight expression groups. Our novel FEED dataset aims to address the scientific community's demand for high-quality multi-view facial expression videos outside the standard categories: Joy, Fear, Sadness, Disgust, Anger Contempt and Surprise." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 502, + 200, + 514 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 502, + 200, + 514 + ], + "spans": [ + { + "bbox": [ + 47, + 502, + 200, + 514 + ], + "type": "text", + "content": "2.3. Speech-driven head avatars" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 522, + 289, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 522, + 289, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 522, + 289, + 713 + ], + "type": "text", + "content": "While numerous studies introduce speech-driven avatars [12, 21, 24, 31, 36, 39, 40], few excel in producing high-quality talking heads with authentic rotations, blinks, and the capability to use both visual and audio inputs. For instance, Wav2Lip [21] aims at re-dubbing videos with accurate lip motions, but often falls short in realism with using single images. Diffused heads [24] struggles to generate long sequences and doesn't provide access to pose control. MakeItTalk [40] animates facial landmarks in a speaker-specific manner, yet struggles with head pose controls due to its non-reliance on 3D. PC-AVS [39] offers a solution, but demands a driving video for pose modulation. Newly developed models such as SadTalker [36] and StyleHEAT [31] have shown promising outcomes in generating high-quality talking heads. We conducted a quantitative evaluation of our audio-driven mode compared to the aforementioned models." + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 324, + 69, + 531, + 153 + ], + "blocks": [ + { + "bbox": [ + 324, + 69, + 531, + 153 + ], + "lines": [ + { + "bbox": [ + 324, + 69, + 531, + 153 + ], + "spans": [ + { + "bbox": [ + 324, + 69, + 531, + 153 + ], + "type": "image", + "image_path": "e582360968f3b94ebaa9e6716b25e04a6f5eb78834c851fa58aea8caced7cf0b.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 323, + 156, + 531, + 232 + ], + "blocks": [ + { + "bbox": [ + 323, + 156, + 531, + 232 + ], + "lines": [ + { + "bbox": [ + 323, + 156, + 531, + 232 + ], + "spans": [ + { + "bbox": [ + 323, + 156, + 531, + 232 + ], + "type": "image", + "image_path": "33d7247f6f7c8a4a0d29ea41ed87dcebf4c56350ecb13484916dc2b21c5c7926.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 321, + 233, + 531, + 293 + ], + "blocks": [ + { + "bbox": [ + 321, + 233, + 531, + 293 + ], + "lines": [ + { + "bbox": [ + 321, + 233, + 531, + 293 + ], + "spans": [ + { + "bbox": [ + 321, + 233, + 531, + 293 + ], + "type": "image", + "image_path": "a0ffc5d39905b35e0710031436f2b4a76dde5773b3d2205c14a51ddcbc729cf5.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 301, + 545, + 323 + ], + "lines": [ + { + "bbox": [ + 305, + 301, + 545, + 323 + ], + "spans": [ + { + "bbox": [ + 305, + 301, + 545, + 323 + ], + "type": "text", + "content": "Figure 2. Illustration of the problem in publicly available face expression data and selected examples from our FEED dataset." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 350, + 392, + 361 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 350, + 392, + 361 + ], + "spans": [ + { + "bbox": [ + 306, + 350, + 392, + 361 + ], + "type": "text", + "content": "3. FEED dataset" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 373, + 547, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 373, + 547, + 529 + ], + "spans": [ + { + "bbox": [ + 304, + 373, + 547, + 529 + ], + "type": "text", + "content": "As previously mentioned, the current publicly available datasets with human face expression videos fall short in capturing a broad range of facial manipulations (see illustration at Fig. 2). In our view, this gap could stem from the complexity of designing clear tasks for participants, coupled with the reality that not everyone is ready for performing extreme expressions on camera. To bridge this gap, our pioneering FEED dataset is designed to meet the scientific community's need for high-quality, multi-view emotion videos that extend beyond standard emotions. It contains various expressions, including strong asymmetric ones, tongue and cheeks movements, winks, head rotations, eye movements and more nuanced gestures. ;" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 531, + 547, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 531, + 547, + 639 + ], + "spans": [ + { + "bbox": [ + 304, + 531, + 547, + 639 + ], + "type": "text", + "content": "Our dataset consists of 520 multi-view videos of 23 subjects, captured with 3 cameras. As extreme face motions are complex and their perception can be heavily influenced by subtle differences, we use a high resolution of 4k for all video, capturing the whole face up to the level of individual hair strands and wrinkles, as shown in Fig. 2. For more examples, please refer to the supplementary materials. Our participants were asked to perform 7 tasks (see Tab. 1) to cover as many facial and head movements as possible." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 642, + 547, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 642, + 547, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 642, + 547, + 712 + ], + "type": "text", + "content": "In our comparison with other expression datasets detailed in Tab. 2, the MEAD dataset [28] emerges as the closest to ours regarding viewpoints and maximal expression intensity. However, despite MEAD's larger number of actors and viewpoints, it offers significantly less variety of facial expressions and have lower resolution than our dataset." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8500" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 71, + 189, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 71, + 189, + 85 + ], + "spans": [ + { + "bbox": [ + 47, + 71, + 189, + 85 + ], + "type": "text", + "content": "4. Expression enhancement" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 91, + 290, + 247 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 91, + 290, + 247 + ], + "spans": [ + { + "bbox": [ + 46, + 91, + 290, + 247 + ], + "type": "text", + "content": "We found that the MegaPortraits [9] model fails to transfer intense expressions correctly, as shown in Fig. 9. Using our FEED dataset of strong and uneven facial expression to fine-tune a pre-trained MegaPortraits model didn't enhance final results as shown at Tab. 4. Training from scratch on FEED leads to fast overfitting due to the small number of identities presented in the dataset compared to VoxCeleb2 [7] used in [9] (23 vs " + }, + { + "bbox": [ + 46, + 91, + 290, + 247 + ], + "type": "inline_equation", + "content": "\\sim 5000" + }, + { + "bbox": [ + 46, + 91, + 290, + 247 + ], + "type": "text", + "content": "). Our method effectively injects this small dataset into training, yielding desirable results. We list our key findings here, for all details, see supplementary materials. Our model's scheme, displayed in Fig. 4, shares similarities with [9], leading us to occasionally refer to certain elements of our scheme in the context of [9]'s pipeline." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 255, + 181, + 268 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 255, + 181, + 268 + ], + "spans": [ + { + "bbox": [ + 47, + 255, + 181, + 268 + ], + "type": "text", + "content": "4.1. Latent expression space" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 273, + 288, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 273, + 288, + 357 + ], + "spans": [ + { + "bbox": [ + 46, + 273, + 288, + 357 + ], + "type": "text", + "content": "We begin by exploring MegaPortraits' latent expression descriptors' space " + }, + { + "bbox": [ + 46, + 273, + 288, + 357 + ], + "type": "inline_equation", + "content": "(\\mathbf{z}_{s / d}" + }, + { + "bbox": [ + 46, + 273, + 288, + 357 + ], + "type": "text", + "content": " in Fig. 4), which is crucial for expression transfer, through PCA analysis. Inspired by [16], which demonstrated that the area under the cumulative explained variance curve of singular values (denoted here as " + }, + { + "bbox": [ + 46, + 273, + 288, + 357 + ], + "type": "inline_equation", + "content": "\\mathbf{AUC}_{\\mathbf{z}}" + }, + { + "bbox": [ + 46, + 273, + 288, + 357 + ], + "type": "text", + "content": ") can serve as an effective metric for dimensionality collapse in some latent spaces, we applied this measure to our study." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 357, + 290, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 357, + 290, + 514 + ], + "spans": [ + { + "bbox": [ + 46, + 357, + 290, + 514 + ], + "type": "text", + "content": "This metric allows us to forecast model performance: greater collapse suggests reduced entropy in the latent space, which correlates with lower representational quality. High quality of the expression space, produced by " + }, + { + "bbox": [ + 46, + 357, + 290, + 514 + ], + "type": "inline_equation", + "content": "\\mathbf{E}_{\\text {motion }}" + }, + { + "bbox": [ + 46, + 357, + 290, + 514 + ], + "type": "text", + "content": " (as shown in Fig. 4), is crucial to capture and differentiate between the nuances of strong facial expressions. Our ablation study confirms (see Tab. 4) that there is a notable correlation between how broad and isotropic model's latent space is (expressed by " + }, + { + "bbox": [ + 46, + 357, + 290, + 514 + ], + "type": "inline_equation", + "content": "\\mathbf{AUC}_{\\mathbf{z}}" + }, + { + "bbox": [ + 46, + 357, + 290, + 514 + ], + "type": "text", + "content": ") and its final performance. Visual comparison is available in supplementary materials. As shown in Fig. 3, our model's latent space outperforms MegaPortraits in expression representation ability. This is supported by both visual (Fig. 9) and quantitative analysis (Tab. 3)." + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 59, + 523, + 276, + 625 + ], + "blocks": [ + { + "bbox": [ + 59, + 523, + 276, + 625 + ], + "lines": [ + { + "bbox": [ + 59, + 523, + 276, + 625 + ], + "spans": [ + { + "bbox": [ + 59, + 523, + 276, + 625 + ], + "type": "image", + "image_path": "43e0be9fd62862a686fc0a3749085dde5d7228aefbfb924c5e39ad2a1a627ee5.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 633, + 288, + 710 + ], + "lines": [ + { + "bbox": [ + 46, + 633, + 288, + 710 + ], + "spans": [ + { + "bbox": [ + 46, + 633, + 288, + 710 + ], + "type": "text", + "content": "Figure 3. Comparison of latent spaces. Left plot shows that our model's latent space is wider and exhibits more even variance distribution. Also, as shown on the right plot, a greater number of principal components are involved in capturing variance across various thresholds. This implies a more robust representational capacity of expression space compared to [9]. The VoxCeleb2 test set was used for both plots." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 330, + 69, + 520, + 159 + ], + "blocks": [ + { + "bbox": [ + 330, + 69, + 520, + 159 + ], + "lines": [ + { + "bbox": [ + 330, + 69, + 520, + 159 + ], + "spans": [ + { + "bbox": [ + 330, + 69, + 520, + 159 + ], + "type": "image", + "image_path": "cfaa5cbc7efaf0aaec7497921709bfe2ffc0e67e9a55e614e647ce8545e2ac06.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 165, + 547, + 254 + ], + "lines": [ + { + "bbox": [ + 304, + 165, + 547, + 254 + ], + "spans": [ + { + "bbox": [ + 304, + 165, + 547, + 254 + ], + "type": "text", + "content": "Figure 4. Method Overview. We use " + }, + { + "bbox": [ + 304, + 165, + 547, + 254 + ], + "type": "inline_equation", + "content": "\\mathbf{E}_{\\mathrm{app}}" + }, + { + "bbox": [ + 304, + 165, + 547, + 254 + ], + "type": "text", + "content": " to extract volume features " + }, + { + "bbox": [ + 304, + 165, + 547, + 254 + ], + "type": "inline_equation", + "content": "\\mathbf{V}_s" + }, + { + "bbox": [ + 304, + 165, + 547, + 254 + ], + "type": "text", + "content": " and a global descriptor " + }, + { + "bbox": [ + 304, + 165, + 547, + 254 + ], + "type": "inline_equation", + "content": "\\mathbf{e}_s" + }, + { + "bbox": [ + 304, + 165, + 547, + 254 + ], + "type": "text", + "content": " from the source image. Then " + }, + { + "bbox": [ + 304, + 165, + 547, + 254 + ], + "type": "inline_equation", + "content": "\\mathbf{E}_{\\mathrm{motion}}" + }, + { + "bbox": [ + 304, + 165, + 547, + 254 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 304, + 165, + 547, + 254 + ], + "type": "inline_equation", + "content": "\\mathbf{E}_{\\mathrm{audio}}" + }, + { + "bbox": [ + 304, + 165, + 547, + 254 + ], + "type": "text", + "content": " generates motion representations from source and driver, including head rotations " + }, + { + "bbox": [ + 304, + 165, + 547, + 254 + ], + "type": "inline_equation", + "content": "\\mathbf{R}_{s / d}" + }, + { + "bbox": [ + 304, + 165, + 547, + 254 + ], + "type": "text", + "content": ", translations " + }, + { + "bbox": [ + 304, + 165, + 547, + 254 + ], + "type": "inline_equation", + "content": "\\mathbf{t}_{s / d}" + }, + { + "bbox": [ + 304, + 165, + 547, + 254 + ], + "type": "text", + "content": ", and expression descriptors " + }, + { + "bbox": [ + 304, + 165, + 547, + 254 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_{s / d}" + }, + { + "bbox": [ + 304, + 165, + 547, + 254 + ], + "type": "text", + "content": ". Using them, we predict warpings " + }, + { + "bbox": [ + 304, + 165, + 547, + 254 + ], + "type": "inline_equation", + "content": "\\mathbf{w}_{s\\rightarrow}" + }, + { + "bbox": [ + 304, + 165, + 547, + 254 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 165, + 547, + 254 + ], + "type": "inline_equation", + "content": "\\mathbf{w}_{\\rightarrow d}" + }, + { + "bbox": [ + 304, + 165, + 547, + 254 + ], + "type": "text", + "content": ". First warping and " + }, + { + "bbox": [ + 304, + 165, + 547, + 254 + ], + "type": "inline_equation", + "content": "\\mathbf{G}_{3D}" + }, + { + "bbox": [ + 304, + 165, + 547, + 254 + ], + "type": "text", + "content": " transform " + }, + { + "bbox": [ + 304, + 165, + 547, + 254 + ], + "type": "inline_equation", + "content": "\\mathbf{V}_s" + }, + { + "bbox": [ + 304, + 165, + 547, + 254 + ], + "type": "text", + "content": " into a canonical volume " + }, + { + "bbox": [ + 304, + 165, + 547, + 254 + ], + "type": "inline_equation", + "content": "\\mathbf{V}_s^C" + }, + { + "bbox": [ + 304, + 165, + 547, + 254 + ], + "type": "text", + "content": " by removing the source motions. Second warping and " + }, + { + "bbox": [ + 304, + 165, + 547, + 254 + ], + "type": "inline_equation", + "content": "\\mathbf{G}_{2D}" + }, + { + "bbox": [ + 304, + 165, + 547, + 254 + ], + "type": "text", + "content": " imposes the driver's motions and renders the final image." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 266, + 547, + 459 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 266, + 547, + 459 + ], + "spans": [ + { + "bbox": [ + 304, + 266, + 547, + 459 + ], + "type": "text", + "content": "Both metrics shown in Fig. 3 were vital in our model's development since they highly correlate with the final model's ability to transfer intense expressions. This was particularly useful for finding adjustments described in Sec. 4.2 and Sec. 4.3, as early visual results during training are not very telling, but changes in the latent space are noticeable after a few epochs. We also found that in MegaPortraits, just a few principal components significantly affect variance, with only 18 components making up " + }, + { + "bbox": [ + 304, + 266, + 547, + 459 + ], + "type": "inline_equation", + "content": "99\\%" + }, + { + "bbox": [ + 304, + 266, + 547, + 459 + ], + "type": "text", + "content": " (Fig. 3), suggesting their " + }, + { + "bbox": [ + 304, + 266, + 547, + 459 + ], + "type": "inline_equation", + "content": "512\\mathbf{z}" + }, + { + "bbox": [ + 304, + 266, + 547, + 459 + ], + "type": "text", + "content": "'s dimension is oversized. For our model, we found a 128-dimensional " + }, + { + "bbox": [ + 304, + 266, + 547, + 459 + ], + "type": "inline_equation", + "content": "\\mathbf{z}" + }, + { + "bbox": [ + 304, + 266, + 547, + 459 + ], + "type": "text", + "content": " is optimal, offering solid performance and reducing overfitting in imbalanced data, as our ablation study reveals. We calculate the Explained Variance " + }, + { + "bbox": [ + 304, + 266, + 547, + 459 + ], + "type": "inline_equation", + "content": "(\\mathrm{EV}_i)" + }, + { + "bbox": [ + 304, + 266, + 547, + 459 + ], + "type": "text", + "content": " for the " + }, + { + "bbox": [ + 304, + 266, + 547, + 459 + ], + "type": "inline_equation", + "content": "i^{th}" + }, + { + "bbox": [ + 304, + 266, + 547, + 459 + ], + "type": "text", + "content": " component and " + }, + { + "bbox": [ + 304, + 266, + 547, + 459 + ], + "type": "inline_equation", + "content": "(\\mathbf{AUC}_{\\mathbf{z}})" + }, + { + "bbox": [ + 304, + 266, + 547, + 459 + ], + "type": "text", + "content": " using equations shown in Eq. (1), which involves standardizing the vector set " + }, + { + "bbox": [ + 304, + 266, + 547, + 459 + ], + "type": "inline_equation", + "content": "\\mathcal{Z}" + }, + { + "bbox": [ + 304, + 266, + 547, + 459 + ], + "type": "text", + "content": " for PCA and sorting the eigenvalues " + }, + { + "bbox": [ + 304, + 266, + 547, + 459 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 304, + 266, + 547, + 459 + ], + "type": "text", + "content": " in descending order." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 310, + 470, + 547, + 502 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 310, + 470, + 547, + 502 + ], + "spans": [ + { + "bbox": [ + 310, + 470, + 547, + 502 + ], + "type": "interline_equation", + "content": "\\mathbf {Z} _ {\\mathrm {s t d}} = \\frac {\\mathcal {Z} - \\bar {\\mathcal {Z}}}{\\sigma (\\bar {\\mathcal {Z}})}, \\quad \\boldsymbol {\\lambda} = \\operatorname {s o r t e d} _ {\\mathrm {d e s c}} \\left(\\operatorname {e i g} \\left(\\frac {1}{n - 1} \\mathbf {Z} _ {\\mathrm {s t d}} ^ {\\top} \\mathbf {Z} _ {\\mathrm {s t d}}\\right)\\right), \\tag {1}", + "image_path": "34df5eeeebb9f1fc19a8a5e5e4924bacfe98c8ed957fb363ab138fa9f35f9822.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 310, + 497, + 502, + 525 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 310, + 497, + 502, + 525 + ], + "spans": [ + { + "bbox": [ + 310, + 497, + 502, + 525 + ], + "type": "interline_equation", + "content": "\\mathrm {E V} _ {i} = \\frac {\\lambda_ {i}}{\\sum_ {j = 1} ^ {p} \\lambda_ {j}}, \\quad \\mathbf {A U C _ {z}} = \\frac {1}{d} \\frac {\\sum_ {i = 1} ^ {d} \\sum_ {k = 1} ^ {i} \\lambda_ {k}}{\\sum_ {j = 1} ^ {d} \\lambda_ {j}},", + "image_path": "b25376b5cd1366d72edbeb93be280a23fb1ad9195da2fb50952fef1211c2e7fa.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 528, + 412, + 538 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 528, + 412, + 538 + ], + "spans": [ + { + "bbox": [ + 306, + 528, + 412, + 538 + ], + "type": "text", + "content": "4.2. Canonical volume" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 545, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 545, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 545, + 547, + 713 + ], + "type": "text", + "content": "Despite being designed to exclude face expression details, we found that canonical volume " + }, + { + "bbox": [ + 304, + 545, + 547, + 713 + ], + "type": "inline_equation", + "content": "(\\mathbf{V}^C" + }, + { + "bbox": [ + 304, + 545, + 547, + 713 + ], + "type": "text", + "content": " in Fig. 4) in MegaPortraits actually retains significant expression information from the source image, contributing to poor translation of intense expressions, as shown in our experiment detailed in Fig. 5. In this experiment, using portraits with varied expressions, we assessed the expression translation accuracy and how expression intensity affects the " + }, + { + "bbox": [ + 304, + 545, + 547, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{V}^C" + }, + { + "bbox": [ + 304, + 545, + 547, + 713 + ], + "type": "text", + "content": "s. Our findings suggest that the canonical volumes in MegaPortraits are not truly neutral. As confirmed by our ablation study (Tab. 4), creating an expression-free " + }, + { + "bbox": [ + 304, + 545, + 547, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{V}^C" + }, + { + "bbox": [ + 304, + 545, + 547, + 713 + ], + "type": "text", + "content": " is essential for precise translation of intense motions. We believe that a more neutral canonical volume improves tractability and effectiveness in expression translation tasks." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 314, + 757 + ], + "type": "text", + "content": "8501" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 55, + 70, + 285, + 167 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 285, + 167 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 285, + 167 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 285, + 167 + ], + "type": "image", + "image_path": "65fd3cb501e87e562244924ab7734643dc949462bd6d5e71948526318373e41c.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 176, + 289, + 342 + ], + "lines": [ + { + "bbox": [ + 46, + 176, + 289, + 342 + ], + "spans": [ + { + "bbox": [ + 46, + 176, + 289, + 342 + ], + "type": "text", + "content": "Figure 5. Canonical volumes " + }, + { + "bbox": [ + 46, + 176, + 289, + 342 + ], + "type": "inline_equation", + "content": "(\\mathbf{V}^C)" + }, + { + "bbox": [ + 46, + 176, + 289, + 342 + ], + "type": "text", + "content": " in MegaPortraits [9] are not expression-neutral. To show it we use three portraits: A (regular expression), B (intense expression), D (regular expression, new identity). Case 1 visualizes poor results in transferring B's intense expression to A using a self-driving mode, contrasting with the effective reconstruction of B when used as both driver and source in case 3. This discrepancy, indicative of expression leakage into the canonical volume, is further quantified by a " + }, + { + "bbox": [ + 46, + 176, + 289, + 342 + ], + "type": "inline_equation", + "content": "43\\%" + }, + { + "bbox": [ + 46, + 176, + 289, + 342 + ], + "type": "text", + "content": " relative difference in " + }, + { + "bbox": [ + 46, + 176, + 289, + 342 + ], + "type": "inline_equation", + "content": "\\mathbf{V}^C" + }, + { + "bbox": [ + 46, + 176, + 289, + 342 + ], + "type": "text", + "content": "s between B and C, contrary to expectations of their similarity for same identity. In cross-driving generation case 2, using B as the driver and D as the source yielded poor results. However, in case 4, after using additive operations on source canonical volumes, the output expression is much closer to driver than in case 2. This manipulation again confirms the significant retention of expression information in canonical volumes." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 355, + 288, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 355, + 288, + 392 + ], + "spans": [ + { + "bbox": [ + 46, + 355, + 288, + 392 + ], + "type": "text", + "content": "To overcome this issue, we propose to match canonical volumes from the different images of the same person " + }, + { + "bbox": [ + 46, + 355, + 288, + 392 + ], + "type": "inline_equation", + "content": "(\\mathbf{V}_{s^n}^C,\\mathbf{V}_{d^n}^C)" + }, + { + "bbox": [ + 46, + 355, + 288, + 392 + ], + "type": "text", + "content": " during training as following:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 114, + 400, + 287, + 417 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 114, + 400, + 287, + 417 + ], + "spans": [ + { + "bbox": [ + 114, + 400, + 287, + 417 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {C V}} ^ {\\mathrm {n}} = \\mathcal {L} _ {\\mathrm {M A E}} \\left(\\mathbf {V} _ {s ^ {n}} ^ {C}, \\mathbf {V} _ {d ^ {n}} ^ {C}\\right), \\tag {2}", + "image_path": "c94e3cf7128c483e476915a129e0e215fe9f8911e920ac88bda379ac56eb16be.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 419, + 289, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 419, + 289, + 456 + ], + "spans": [ + { + "bbox": [ + 46, + 419, + 289, + 456 + ], + "type": "text", + "content": "This loss ensures " + }, + { + "bbox": [ + 46, + 419, + 289, + 456 + ], + "type": "inline_equation", + "content": "\\mathbf{V}^C" + }, + { + "bbox": [ + 46, + 419, + 289, + 456 + ], + "type": "text", + "content": " remains stable and expression-independent, crucial for translating intense expressions, as shown in our ablation study (Tab. 4)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 461, + 203, + 473 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 461, + 203, + 473 + ], + "spans": [ + { + "bbox": [ + 47, + 461, + 203, + 473 + ], + "type": "text", + "content": "4.3. Source-driver mismatch loss" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 479, + 288, + 576 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 479, + 288, + 576 + ], + "spans": [ + { + "bbox": [ + 46, + 479, + 288, + 576 + ], + "type": "text", + "content": "In addition to maintaining expression-free " + }, + { + "bbox": [ + 46, + 479, + 288, + 576 + ], + "type": "inline_equation", + "content": "\\mathbf{V}^{C}" + }, + { + "bbox": [ + 46, + 479, + 288, + 576 + ], + "type": "text", + "content": ", it's vital to remove all identity information from the expression vector " + }, + { + "bbox": [ + 46, + 479, + 288, + 576 + ], + "type": "inline_equation", + "content": "\\mathbf{z}" + }, + { + "bbox": [ + 46, + 479, + 288, + 576 + ], + "type": "text", + "content": ". While contrastive losses address this in [9], our experiments indicate they are insufficient to prevent overfitting in our imbalanced data scenario, where emotionally intense images represent only " + }, + { + "bbox": [ + 46, + 479, + 288, + 576 + ], + "type": "inline_equation", + "content": "0.1\\%" + }, + { + "bbox": [ + 46, + 479, + 288, + 576 + ], + "type": "text", + "content": " of our training set but are sampled " + }, + { + "bbox": [ + 46, + 479, + 288, + 576 + ], + "type": "inline_equation", + "content": "25\\%" + }, + { + "bbox": [ + 46, + 479, + 288, + 576 + ], + "type": "text", + "content": " of the time. We introduce a novel self-supervised loss which mitigates identity information in latent expression vectors:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 52, + 590, + 287, + 638 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 52, + 590, + 287, + 638 + ], + "spans": [ + { + "bbox": [ + 52, + 590, + 287, + 638 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} _ {s d m} (\\mathbf {z} _ {s}, \\mathbf {z} _ {d}) = w * \\max (0, \\cos (\\mathbf {z} _ {s}, \\mathbf {z} _ {d}) - m a r g i n), \\\\ w; \\text {m a r g i n} = \\left\\{ \\begin{array}{l l} 1; 0. 5 & \\text {i f} \\mathbf {z} _ {s}, \\mathbf {z} _ {d} \\text {a r e f r o m V C 2} \\\\ 1 0; 0. 2 5 & \\text {i f} \\mathbf {z} _ {s}, \\mathbf {z} _ {d} \\text {a r e f r o m F E E D} \\end{array} \\right. \\tag {3} \\\\ \\end{array}", + "image_path": "0a401f4490bad78c78e3a5049ed2855d896a35ef7995f778f0e3b4168cb3c5cf.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 642, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 642, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 642, + 289, + 715 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 46, + 642, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_s, \\mathbf{z}_d" + }, + { + "bbox": [ + 46, + 642, + 289, + 715 + ], + "type": "text", + "content": " are emotion vectors from source and driver images, " + }, + { + "bbox": [ + 46, + 642, + 289, + 715 + ], + "type": "inline_equation", + "content": "w, \\text{margin}" + }, + { + "bbox": [ + 46, + 642, + 289, + 715 + ], + "type": "text", + "content": " adjust the loss's intensity and strictness respectively. We increase " + }, + { + "bbox": [ + 46, + 642, + 289, + 715 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 46, + 642, + 289, + 715 + ], + "type": "text", + "content": " for FEED due to its higher sampling rate and decrease margin for more assured expression variance in " + }, + { + "bbox": [ + 46, + 642, + 289, + 715 + ], + "type": "inline_equation", + "content": "s, d" + }, + { + "bbox": [ + 46, + 642, + 289, + 715 + ], + "type": "text", + "content": ". Refer to Fig. 6 for a visual summary of all self-supervised losses enhancing our latent emotion space." + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 331, + 69, + 523, + 266 + ], + "blocks": [ + { + "bbox": [ + 331, + 69, + 523, + 266 + ], + "lines": [ + { + "bbox": [ + 331, + 69, + 523, + 266 + ], + "spans": [ + { + "bbox": [ + 331, + 69, + 523, + 266 + ], + "type": "image", + "image_path": "4168394fbcbb61f875f732ea64d8e9e00b593fc9b8b7198c39c30b7289b588f7.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 273, + 547, + 318 + ], + "lines": [ + { + "bbox": [ + 305, + 273, + 547, + 318 + ], + "spans": [ + { + "bbox": [ + 305, + 273, + 547, + 318 + ], + "type": "text", + "content": "Figure 6. The visualisation of our self-supervised losses. " + }, + { + "bbox": [ + 305, + 273, + 547, + 318 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{cos}}" + }, + { + "bbox": [ + 305, + 273, + 547, + 318 + ], + "type": "text", + "content": " is a contrastive loss, similar to one used in [9] but uses different pairs as described in Sec. 5.1. Our novel " + }, + { + "bbox": [ + 305, + 273, + 547, + 318 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{sdm}" + }, + { + "bbox": [ + 305, + 273, + 547, + 318 + ], + "type": "text", + "content": " is designed to prevent overfitting in an imbalanced dataset." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 338, + 502, + 353 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 338, + 502, + 353 + ], + "spans": [ + { + "bbox": [ + 306, + 338, + 502, + 353 + ], + "type": "text", + "content": "5. Incorporating speech-driving mode" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 358, + 468, + 371 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 358, + 468, + 371 + ], + "spans": [ + { + "bbox": [ + 306, + 358, + 468, + 371 + ], + "type": "text", + "content": "5.1. Latent space disentanglement" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 377, + 547, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 377, + 547, + 545 + ], + "spans": [ + { + "bbox": [ + 304, + 377, + 547, + 545 + ], + "type": "text", + "content": "During our latent space analysis described in Sec. 4, we found that combining expression from one driver with the head pose from another resulted in poor performance for MegaPortraits model, as illustrated in Fig. 7 (bottom row). Indeed, the base model lacks a mechanism to intentionally prevent head pose leakage. Disentangled expression latent space, besides expanding the model's use cases, plays the crucial role for the speech-driving mode. When the latent vector " + }, + { + "bbox": [ + 304, + 377, + 547, + 545 + ], + "type": "inline_equation", + "content": "\\mathbf{z}" + }, + { + "bbox": [ + 304, + 377, + 547, + 545 + ], + "type": "text", + "content": " is entangled with head pose data, predicting it from speech becomes challenging because speech lacks head rotation information, unlike images. Moreover, we would like to have a full control over head rotations during speech mode. We were able to make our latent space disentangled by changing the way we sample images for " + }, + { + "bbox": [ + 304, + 377, + 547, + 545 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{cos}}" + }, + { + "bbox": [ + 304, + 377, + 547, + 545 + ], + "type": "text", + "content": " Eq. (5)." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 545, + 547, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 545, + 547, + 617 + ], + "spans": [ + { + "bbox": [ + 304, + 545, + 547, + 617 + ], + "type": "text", + "content": "The MegaPortraits model uses " + }, + { + "bbox": [ + 304, + 545, + 547, + 617 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{cos}}" + }, + { + "bbox": [ + 304, + 545, + 547, + 617 + ], + "type": "text", + "content": " - a modified Large Margin Cosine Loss [26] to mitigate the transfer of appearance features to expression descriptor (latent vector " + }, + { + "bbox": [ + 304, + 545, + 547, + 617 + ], + "type": "inline_equation", + "content": "\\mathbf{z}" + }, + { + "bbox": [ + 304, + 545, + 547, + 617 + ], + "type": "text", + "content": " on Fig. 4), which is crucial for cross-driving mode. Without this, differences in appearance between the source and driver, like hairstyle or skin tone, leak from driver to output image." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 618, + 548, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 618, + 548, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 618, + 548, + 714 + ], + "type": "text", + "content": "For computing this loss, the authors utilize a supplementary source-driving pair " + }, + { + "bbox": [ + 304, + 618, + 548, + 714 + ], + "type": "inline_equation", + "content": "(\\mathbf{x}_{s^{*}}" + }, + { + "bbox": [ + 304, + 618, + 548, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 618, + 548, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{d^{*}})" + }, + { + "bbox": [ + 304, + 618, + 548, + 714 + ], + "type": "text", + "content": " from a different video with another identity, ensuring a distinct appearance from the current " + }, + { + "bbox": [ + 304, + 618, + 548, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_s,\\mathbf{x}_d" + }, + { + "bbox": [ + 304, + 618, + 548, + 714 + ], + "type": "text", + "content": " pair. The base model is then employed to produce cross-reenacted image " + }, + { + "bbox": [ + 304, + 618, + 548, + 714 + ], + "type": "inline_equation", + "content": "(\\hat{\\mathbf{x}}_{s^{*}\\rightarrow d} = \\mathbf{G}_{\\mathrm{base}}(\\mathbf{x}_{s^{*}},\\mathbf{x}_d))" + }, + { + "bbox": [ + 304, + 618, + 548, + 714 + ], + "type": "text", + "content": " Concurrently, they determine a separate motion descriptor, " + }, + { + "bbox": [ + 304, + 618, + 548, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_{d^{*}} = \\mathbf{E}_{\\mathrm{motion}}(\\mathbf{x}_{d^{*}})" + }, + { + "bbox": [ + 304, + 618, + 548, + 714 + ], + "type": "text", + "content": ". Descriptors " + }, + { + "bbox": [ + 304, + 618, + 548, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_{* \\rightarrow d}" + }, + { + "bbox": [ + 304, + 618, + 548, + 714 + ], + "type": "text", + "content": " from the respective forward passes are also used for positive and negative pairs." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8502" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 49, + 70, + 286, + 163 + ], + "blocks": [ + { + "bbox": [ + 49, + 70, + 286, + 163 + ], + "lines": [ + { + "bbox": [ + 49, + 70, + 286, + 163 + ], + "spans": [ + { + "bbox": [ + 49, + 70, + 286, + 163 + ], + "type": "image", + "image_path": "2fc771a3720ccaa6e2cde67511df1976c6cbeed5e2acdb994a7fe463b911909b.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 172, + 287, + 216 + ], + "lines": [ + { + "bbox": [ + 46, + 172, + 287, + 216 + ], + "spans": [ + { + "bbox": [ + 46, + 172, + 287, + 216 + ], + "type": "text", + "content": "Figure 7. The illustration of disentanglement problem. Images in the right row indicate the expression descriptor " + }, + { + "bbox": [ + 46, + 172, + 287, + 216 + ], + "type": "inline_equation", + "content": "\\mathbf{z}" + }, + { + "bbox": [ + 46, + 172, + 287, + 216 + ], + "type": "text", + "content": " entangled with head pose for MegaPortraits, whereas we avoid it using changes proposed in Sec. 5.1." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 236, + 287, + 331 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 236, + 287, + 331 + ], + "spans": [ + { + "bbox": [ + 46, + 236, + 287, + 331 + ], + "type": "text", + "content": "Our innovation here is a new sampling strategy: for each " + }, + { + "bbox": [ + 46, + 236, + 287, + 331 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_s" + }, + { + "bbox": [ + 46, + 236, + 287, + 331 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 236, + 287, + 331 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_d" + }, + { + "bbox": [ + 46, + 236, + 287, + 331 + ], + "type": "text", + "content": " pair, we sample one more random additional pair " + }, + { + "bbox": [ + 46, + 236, + 287, + 331 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{s^m}" + }, + { + "bbox": [ + 46, + 236, + 287, + 331 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 236, + 287, + 331 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{d^m}" + }, + { + "bbox": [ + 46, + 236, + 287, + 331 + ], + "type": "text", + "content": " apart from " + }, + { + "bbox": [ + 46, + 236, + 287, + 331 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{s^*}" + }, + { + "bbox": [ + 46, + 236, + 287, + 331 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 236, + 287, + 331 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{d^*}" + }, + { + "bbox": [ + 46, + 236, + 287, + 331 + ], + "type": "text", + "content": ". We then apply the model to produce the following cross-reenacted moved image using source identity from " + }, + { + "bbox": [ + 46, + 236, + 287, + 331 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{s^*}" + }, + { + "bbox": [ + 46, + 236, + 287, + 331 + ], + "type": "text", + "content": ", desired head pose from " + }, + { + "bbox": [ + 46, + 236, + 287, + 331 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{s^m}" + }, + { + "bbox": [ + 46, + 236, + 287, + 331 + ], + "type": "text", + "content": " and emotions from " + }, + { + "bbox": [ + 46, + 236, + 287, + 331 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_d" + }, + { + "bbox": [ + 46, + 236, + 287, + 331 + ], + "type": "text", + "content": ", represented as: " + }, + { + "bbox": [ + 46, + 236, + 287, + 331 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{x}}_{s^{*m}\\rightarrow d} = \\mathbf{G}_{\\mathrm{base}}(\\mathbf{x}_{s^{*m}},\\mathbf{x}_d)" + }, + { + "bbox": [ + 46, + 236, + 287, + 331 + ], + "type": "text", + "content": ". This creates a positive pair with the same desired emotion but varying head poses." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 331, + 288, + 402 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 331, + 288, + 402 + ], + "spans": [ + { + "bbox": [ + 47, + 331, + 288, + 402 + ], + "type": "text", + "content": "Motion descriptors are organized into positive pairs " + }, + { + "bbox": [ + 47, + 331, + 288, + 402 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 47, + 331, + 288, + 402 + ], + "type": "text", + "content": " for alignment, and negative pairs " + }, + { + "bbox": [ + 47, + 331, + 288, + 402 + ], + "type": "inline_equation", + "content": "\\mathcal{N}" + }, + { + "bbox": [ + 47, + 331, + 288, + 402 + ], + "type": "text", + "content": " for non-alignment: " + }, + { + "bbox": [ + 47, + 331, + 288, + 402 + ], + "type": "inline_equation", + "content": "\\mathcal{P} = \\left\\{(\\mathbf{z}_{s\\rightarrow d},\\mathbf{z}_d),(\\mathbf{z}_{s^*\\rightarrow d},\\mathbf{z}_d),(\\mathbf{z}_{s^{*m}\\rightarrow d},\\mathbf{z}_d)\\right\\}" + }, + { + "bbox": [ + 47, + 331, + 288, + 402 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 47, + 331, + 288, + 402 + ], + "type": "inline_equation", + "content": "\\mathcal{N} = \\left\\{(\\mathbf{z}_{s\\rightarrow d},\\mathbf{z}_{d^*}),(\\mathbf{z}_{s^*\\rightarrow d},\\mathbf{z}_{d^*}),(\\mathbf{z}_{s^{*m}\\rightarrow d},\\mathbf{z}_{d^*})\\right\\}" + }, + { + "bbox": [ + 47, + 331, + 288, + 402 + ], + "type": "text", + "content": " (see Fig. 6). These pairs are used to calculate the following cosine distance:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 104, + 403, + 287, + 417 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 104, + 403, + 287, + 417 + ], + "spans": [ + { + "bbox": [ + 104, + 403, + 287, + 417 + ], + "type": "interline_equation", + "content": "d \\left(\\mathbf {z} _ {i}, \\mathbf {z} _ {j}\\right) = s \\cdot \\left(\\left\\langle \\mathbf {z} _ {i}, \\mathbf {z} _ {j} \\right\\rangle - m\\right), \\tag {4}", + "image_path": "ac3dd235f99a08bf263283455503e22cc393348fad303e29ec90144c08b56c60.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 420, + 287, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 420, + 287, + 456 + ], + "spans": [ + { + "bbox": [ + 46, + 420, + 287, + 456 + ], + "type": "text", + "content": "where both " + }, + { + "bbox": [ + 46, + 420, + 287, + 456 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 46, + 420, + 287, + 456 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 420, + 287, + 456 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 46, + 420, + 287, + 456 + ], + "type": "text", + "content": " are hyperparameters. This distance is then used to calculate a large margin cosine loss (Cos-Face) [27]:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 472, + 287, + 505 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 472, + 287, + 505 + ], + "spans": [ + { + "bbox": [ + 53, + 472, + 287, + 505 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\cos} = - \\sum_ {(\\mathbf {z} _ {k}, \\mathbf {z} _ {l}) \\in \\mathcal {P}} \\log \\frac {\\exp \\left\\{d \\left(\\mathbf {z} _ {k} , \\mathbf {z} _ {l}\\right) \\right\\}}{\\exp \\left\\{d \\left(\\mathbf {z} _ {k} , \\mathbf {z} _ {l}\\right) \\right\\} + \\sum_ {(\\mathbf {z} _ {i} , \\mathbf {z} _ {j}) \\in \\mathcal {N}} \\exp \\left\\{d \\left(\\mathbf {z} _ {i} , \\mathbf {z} _ {j}\\right) \\right\\}} \\tag {5}", + "image_path": "2d6f85c07830f66c0571acd3af8da897e59644347e0f9279ce5a1e63f1bb2aff.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 510, + 287, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 510, + 287, + 533 + ], + "spans": [ + { + "bbox": [ + 46, + 510, + 287, + 533 + ], + "type": "text", + "content": "This loss prevents the head pose from leaking into the embeddings as shown in Fig. 7 (upper row)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 534, + 288, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 534, + 288, + 712 + ], + "spans": [ + { + "bbox": [ + 46, + 534, + 288, + 712 + ], + "type": "text", + "content": "By doing this, we achieved the capability to interpret " + }, + { + "bbox": [ + 46, + 534, + 288, + 712 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_i" + }, + { + "bbox": [ + 46, + 534, + 288, + 712 + ], + "type": "text", + "content": " as a latent vector encapsulating the emotional content from an image " + }, + { + "bbox": [ + 46, + 534, + 288, + 712 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_i" + }, + { + "bbox": [ + 46, + 534, + 288, + 712 + ], + "type": "text", + "content": ". This advancement opened the door for an intriguing possibility during inference: the prediction of this vector could potentially be derived from an audio signal instead of a facial image using the motion encoder " + }, + { + "bbox": [ + 46, + 534, + 288, + 712 + ], + "type": "inline_equation", + "content": "\\mathbf{E}_{\\mathrm{motion}}" + }, + { + "bbox": [ + 46, + 534, + 288, + 712 + ], + "type": "text", + "content": ", effectively transforming the model into a speech-driven system. The critical challenge here is aligning the audio input with the pretrained latent space. We addressed this by introducing an additional audio encoder, " + }, + { + "bbox": [ + 46, + 534, + 288, + 712 + ], + "type": "inline_equation", + "content": "\\mathbf{E}_{\\mathrm{aud}}" + }, + { + "bbox": [ + 46, + 534, + 288, + 712 + ], + "type": "text", + "content": ", and employing " + }, + { + "bbox": [ + 46, + 534, + 288, + 712 + ], + "type": "inline_equation", + "content": "\\mathbf{E}_{\\mathrm{motion}}" + }, + { + "bbox": [ + 46, + 534, + 288, + 712 + ], + "type": "text", + "content": " as a teacher model for this new encoder. A crucial factor in achieving remarkable accuracy in lip synchronization was our ability to isolate specific components within the expression latent space, that are solely responsible for mouth movements. We describe it in the next subsection Sec. 5.2." + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 305, + 70, + 545, + 180 + ], + "blocks": [ + { + "bbox": [ + 305, + 70, + 545, + 180 + ], + "lines": [ + { + "bbox": [ + 305, + 70, + 545, + 180 + ], + "spans": [ + { + "bbox": [ + 305, + 70, + 545, + 180 + ], + "type": "image", + "image_path": "426d2459ddac66dc468bd55fd6056ae9f8154ca624c50c1e768277a3b7502cea.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 187, + 547, + 265 + ], + "lines": [ + { + "bbox": [ + 304, + 187, + 547, + 265 + ], + "spans": [ + { + "bbox": [ + 304, + 187, + 547, + 265 + ], + "type": "text", + "content": "Figure 8. In the upper row, the first principal component from a random subset of expression vectors " + }, + { + "bbox": [ + 304, + 187, + 547, + 265 + ], + "type": "inline_equation", + "content": "\\mathcal{Z}" + }, + { + "bbox": [ + 304, + 187, + 547, + 265 + ], + "type": "text", + "content": " affects not just the mouth but also blinks and gaze, making it difficult to interpret and isolate mouth movement components, needed for training " + }, + { + "bbox": [ + 304, + 187, + 547, + 265 + ], + "type": "inline_equation", + "content": "\\mathbf{E}_{\\mathrm{aud}}" + }, + { + "bbox": [ + 304, + 187, + 547, + 265 + ], + "type": "text", + "content": ". However, as shown in the bottom row, using " + }, + { + "bbox": [ + 304, + 187, + 547, + 265 + ], + "type": "inline_equation", + "content": "\\mathcal{Z}_m" + }, + { + "bbox": [ + 304, + 187, + 547, + 265 + ], + "type": "text", + "content": " from Sec. 5.2, the first principal component, " + }, + { + "bbox": [ + 304, + 187, + 547, + 265 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_1" + }, + { + "bbox": [ + 304, + 187, + 547, + 265 + ], + "type": "text", + "content": ", focused on mouth movements, is effectively isolated, " + }, + { + "bbox": [ + 304, + 187, + 547, + 265 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_0" + }, + { + "bbox": [ + 304, + 187, + 547, + 265 + ], + "type": "text", + "content": " - latent vector, corresponded to the central image." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 282, + 416, + 294 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 282, + 416, + 294 + ], + "spans": [ + { + "bbox": [ + 306, + 282, + 416, + 294 + ], + "type": "text", + "content": "5.2. Mouth movements" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 300, + 547, + 409 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 300, + 547, + 409 + ], + "spans": [ + { + "bbox": [ + 304, + 300, + 547, + 409 + ], + "type": "text", + "content": "To train the audio encoder " + }, + { + "bbox": [ + 304, + 300, + 547, + 409 + ], + "type": "inline_equation", + "content": "\\mathbf{E}_{\\mathrm{aud}}" + }, + { + "bbox": [ + 304, + 300, + 547, + 409 + ], + "type": "text", + "content": " for predicting " + }, + { + "bbox": [ + 304, + 300, + 547, + 409 + ], + "type": "inline_equation", + "content": "\\mathbf{z}" + }, + { + "bbox": [ + 304, + 300, + 547, + 409 + ], + "type": "text", + "content": " from speech, we chose to use the motion encoder " + }, + { + "bbox": [ + 304, + 300, + 547, + 409 + ], + "type": "inline_equation", + "content": "\\mathbf{E}_{\\mathrm{motion}}" + }, + { + "bbox": [ + 304, + 300, + 547, + 409 + ], + "type": "text", + "content": " as a teacher model. However, direct matching vectors " + }, + { + "bbox": [ + 304, + 300, + 547, + 409 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_i^{\\mathrm{aud}}" + }, + { + "bbox": [ + 304, + 300, + 547, + 409 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 304, + 300, + 547, + 409 + ], + "type": "inline_equation", + "content": "\\mathbf{E}_{\\mathrm{aud}}" + }, + { + "bbox": [ + 304, + 300, + 547, + 409 + ], + "type": "text", + "content": " and pseudo-ground truth " + }, + { + "bbox": [ + 304, + 300, + 547, + 409 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_i" + }, + { + "bbox": [ + 304, + 300, + 547, + 409 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 304, + 300, + 547, + 409 + ], + "type": "inline_equation", + "content": "\\mathbf{E}_{\\mathrm{motion}}" + }, + { + "bbox": [ + 304, + 300, + 547, + 409 + ], + "type": "text", + "content": " gives poor results (see Tab. 6). We believe that the reason behind this is that speech-derived " + }, + { + "bbox": [ + 304, + 300, + 547, + 409 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_i^{\\mathrm{aud}}" + }, + { + "bbox": [ + 304, + 300, + 547, + 409 + ], + "type": "text", + "content": ", while able to capture lip movements, struggles with other facial motions, especially in the upper face. Thus, we focused on isolating mouth movement components in the expression latent space using PCA analysis." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 409, + 547, + 492 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 409, + 547, + 492 + ], + "spans": [ + { + "bbox": [ + 304, + 409, + 547, + 492 + ], + "type": "text", + "content": "Employing PCA on a broad array of expression vectors " + }, + { + "bbox": [ + 304, + 409, + 547, + 492 + ], + "type": "inline_equation", + "content": "\\mathcal{Z}" + }, + { + "bbox": [ + 304, + 409, + 547, + 492 + ], + "type": "text", + "content": ", generated from images with varied facial emotions, yields principal components with their explained variances " + }, + { + "bbox": [ + 304, + 409, + 547, + 492 + ], + "type": "inline_equation", + "content": "(\\mathrm{EV}_i)" + }, + { + "bbox": [ + 304, + 409, + 547, + 492 + ], + "type": "text", + "content": ". Altering these components, particularly those with high " + }, + { + "bbox": [ + 304, + 409, + 547, + 492 + ], + "type": "inline_equation", + "content": "\\mathrm{EV}_i" + }, + { + "bbox": [ + 304, + 409, + 547, + 492 + ], + "type": "text", + "content": ", significantly changes facial expressions in an image. However, they typically affect a combination of facial features—Fig. 8 (upper row), not isolated areas like the mouth or eyes." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 493, + 547, + 612 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 493, + 547, + 612 + ], + "spans": [ + { + "bbox": [ + 304, + 493, + 547, + 612 + ], + "type": "text", + "content": "To focus on mouth movements, we created a unique set of " + }, + { + "bbox": [ + 304, + 493, + 547, + 612 + ], + "type": "inline_equation", + "content": "\\mathcal{Z}_m" + }, + { + "bbox": [ + 304, + 493, + 547, + 612 + ], + "type": "text", + "content": " from just one video of a person performing mouth manipulations and manually edited it to minimize upper face movements. This involved using a still upper face from the first frame in all subsequent frames, ensuring that the principal components mainly represented mouth movements. Applying PCA of " + }, + { + "bbox": [ + 304, + 493, + 547, + 612 + ], + "type": "inline_equation", + "content": "\\mathcal{Z}_m" + }, + { + "bbox": [ + 304, + 493, + 547, + 612 + ], + "type": "text", + "content": " reveals that most principal components are responsible for solely mouth movements. For illustration of the first component, see Fig. 8 (bottom row). Based on distilled components, we introduce mouth PCA mouth loss:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 330, + 617, + 545, + 649 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 330, + 617, + 545, + 649 + ], + "spans": [ + { + "bbox": [ + 330, + 617, + 545, + 649 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {P C A}} \\left(v _ {i}, v _ {j}, n\\right) = \\frac {1}{n} \\sum_ {k = 1} ^ {n} \\left| \\mathrm {P C} _ {\\mathbf {v} _ {i}} (k) - \\mathrm {P C} _ {\\mathbf {v} _ {j}} (k) \\right| \\tag {6}", + "image_path": "eb8b3d66cd0da9570593d63f42e9e7b37a3612d22bb8d5cf4411267d3f017ffc.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 653, + 547, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 653, + 547, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 653, + 547, + 712 + ], + "type": "text", + "content": "Where " + }, + { + "bbox": [ + 304, + 653, + 547, + 712 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 304, + 653, + 547, + 712 + ], + "type": "text", + "content": " is the number of principal components to be considered, " + }, + { + "bbox": [ + 304, + 653, + 547, + 712 + ], + "type": "inline_equation", + "content": "\\mathrm{PC}_{\\mathbf{z}_i}(k)" + }, + { + "bbox": [ + 304, + 653, + 547, + 712 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 653, + 547, + 712 + ], + "type": "inline_equation", + "content": "\\mathrm{PC}_{\\mathbf{z}_j}(k)" + }, + { + "bbox": [ + 304, + 653, + 547, + 712 + ], + "type": "text", + "content": " are the " + }, + { + "bbox": [ + 304, + 653, + 547, + 712 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 653, + 547, + 712 + ], + "type": "text", + "content": "-th principal components of vectors " + }, + { + "bbox": [ + 304, + 653, + 547, + 712 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_i" + }, + { + "bbox": [ + 304, + 653, + 547, + 712 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 653, + 547, + 712 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_j" + }, + { + "bbox": [ + 304, + 653, + 547, + 712 + ], + "type": "text", + "content": " respectively. Implementing this loss has demonstrably enhanced the quality and accuracy of our results, as our ablation study confirms (see Tab. 6)." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "8503" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 128, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 128, + 85 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 128, + 85 + ], + "type": "text", + "content": "6. Experiments" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 91, + 288, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 91, + 288, + 152 + ], + "spans": [ + { + "bbox": [ + 46, + 91, + 288, + 152 + ], + "type": "text", + "content": "In this section, we describe experiments we set for comparison for both our image-driven and speech-driven mode. Please see implementation details such as all used losses, component's sizes and architecture details, training procedure, data preprocess in supplementary materials." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 156, + 190, + 170 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 156, + 190, + 170 + ], + "spans": [ + { + "bbox": [ + 47, + 156, + 190, + 170 + ], + "type": "text", + "content": "6.1. Image-driven comparison" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 175, + 287, + 281 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 175, + 287, + 281 + ], + "spans": [ + { + "bbox": [ + 46, + 175, + 287, + 281 + ], + "type": "text", + "content": "Methods. We evaluated our image-driven model alongside various one-shot models. For this comparison, we included models like NOFA [32] and StyleHEAT [31], which use 3DMM expression blendshapes without disentanglement training, and models with trainable latent face motion representations like FOMM [22], UVA [23], MetaPortrait [35] and MegaPortraits [9]. It's noteworthy that NOFA and UVA require per-source optimization, unlike the others, including our model." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 282, + 287, + 378 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 282, + 287, + 378 + ], + "spans": [ + { + "bbox": [ + 46, + 282, + 287, + 378 + ], + "type": "text", + "content": "Evaluation metrics and data. For our evaluation, we use 100 random, non-child, single-person images from the FFHQ dataset as sources. From the MEAD dataset, we chose 100 images covering all high-intensity emotions except neutral, and another 100 from the FEED dataset's extreme emotion task, featuring 5 identities not seen during training of our model. Each source image was paired with 20 driving images from these sets, totaling 2000 pairs." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 379, + 288, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 379, + 288, + 594 + ], + "spans": [ + { + "bbox": [ + 46, + 379, + 288, + 594 + ], + "type": "text", + "content": "For assessing reenactment quality, we employ various metrics including the Frechet Inception Distance (FID) [13] to measure the distributional discrepancy between synthetic and real images. Cosine similarity (CSIM) from a face recognition network [5] quantifies the identity preservation in generated images. Additionally, we conducted user studies to determine preferences regarding motion (UMTN) and appearance (UAPP) preservation, presenting participants with triplets—either a driving or source image and two method outputs—and asking them to choose the one with better motion or appearance preservation. Our user study includes 34 participants and around 300 unique questions. Both metrics numerically reflect preference percentages. As our visual results Fig. 9, and quantitative metrics Tab. 3 show, our method demonstrates exceptional ability in translating strong and extreme expressions, notably surpassing other image-driven avatar techniques in user preference for facial expression translation (UMTN) and FID scores." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 599, + 194, + 612 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 599, + 194, + 612 + ], + "spans": [ + { + "bbox": [ + 47, + 599, + 194, + 612 + ], + "type": "text", + "content": "6.2. Speech-driven comparison" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 617, + 287, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 617, + 287, + 677 + ], + "spans": [ + { + "bbox": [ + 46, + 617, + 287, + 677 + ], + "type": "text", + "content": "Methods. To assess our model's speech-driven capabilities, we compared it with StyleHEAT [31], which also operates using both image and audio inputs, and audio-exclusive models including SadTalker [36], MakeItTalk [40], and PC-AVS [39]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 677, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 677, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 677, + 288, + 713 + ], + "type": "text", + "content": "Evaluation metrics and data. We employ the evaluation procedure described in [12]. We measure video realism using Frechet Inception Distance (FID) [13]. We also extract" + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 309, + 71, + 545, + 140 + ], + "blocks": [ + { + "bbox": [ + 309, + 71, + 545, + 140 + ], + "lines": [ + { + "bbox": [ + 309, + 71, + 545, + 140 + ], + "spans": [ + { + "bbox": [ + 309, + 71, + 545, + 140 + ], + "type": "table", + "html": "
Driver:MEADFEED
MethodFID↓CSIM↑UMTN↑UAPP↑FID↓CSIM↑UMTN↑UAPP↑
MetaPortrait [35]77.50.64**78.80.66**
NOFA [32]69.20.65**69.80.663.88.8
FOMM [22]84.40.562.40.385.90.551.80.5
UVA [17]78.70.680.96.479.50.680.69.4
StyleHEAT [31]71.90.633.73.372.40.612.63.8
MegaPortraits [9]61.10.7322.746.662.80.7316.742.8
Our (EMOPortraits)59.60.7470.343.460.20.7074.634.7
", + "image_path": "f723b0a62eb5b4733b04b9bdc559c91c2d518e61d1a5e9187c6db9f9ecf35458.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 305, + 142, + 547, + 351 + ], + "lines": [ + { + "bbox": [ + 305, + 142, + 547, + 351 + ], + "spans": [ + { + "bbox": [ + 305, + 142, + 547, + 351 + ], + "type": "text", + "content": "Table 3. Our method notably outperforms others in the FID score and strongly leads in the user preference metric for face expression translation (UMTN). It excels in reliably translating strong and extreme expressions from the driving image, distinguishing it as the most capable among the compared methods. Additionally, our approach maintains the source image's shape and appearance (measured by CSIM and UAPP metrics) on par with MegaPortraits [9] when using strong and regular expressions from the MEAD dataset as the driving signal. However, we observe a minor deviation in identity preservation metrics for extreme emotions (using the FEED dataset). This difference may arise because our method attempts to transfer asymmetric facial expressions, while MegaPortraits might modify the source image less, or not at all, with a challenging driver (as depicted in Fig. 5, case 1), thereby achieving better identity preservation. UVA was excluded from the MEAD-driven user study due to the authors were able to provide only part of the test data. Also, MetaPortrait was excluded from both user studies because of late data provision by authors, offering only the basic model results, without their super-resolution module outputs." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 312, + 369, + 541, + 423 + ], + "blocks": [ + { + "bbox": [ + 312, + 369, + 541, + 423 + ], + "lines": [ + { + "bbox": [ + 312, + 369, + 541, + 423 + ], + "spans": [ + { + "bbox": [ + 312, + 369, + 541, + 423 + ], + "type": "table", + "html": "
MethodFID↓CSIM↑UMTN↓UAPP↓AUCz↓
MegaPortraits fine-tuned62.30.638.512.40.89
Ours w/o LcnCV61.40.679.217.90.88
Ours w/o Lsdm61.80.5911.36.70.85
Ours dim(z) = 51263.40.4425.61.10.77
Ours59.60.7445.461.90.80
", + "image_path": "873156c9e6afd47be52e4d5a99fb0aa75c8f06eb0509809b0b49c314487dafaa.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 305, + 431, + 547, + 530 + ], + "lines": [ + { + "bbox": [ + 305, + 431, + 547, + 530 + ], + "spans": [ + { + "bbox": [ + 305, + 431, + 547, + 530 + ], + "type": "text", + "content": "Table 4. Our ablation study examines key aspects of the image-driven mode. After fine-tuning MegaPortraits on the FEED dataset, we observed a decline in identity preservation due to overfitting on FEED identities, without any improvement in emotion translation. The lack of " + }, + { + "bbox": [ + 305, + 431, + 547, + 530 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{CV}^{n}" + }, + { + "bbox": [ + 305, + 431, + 547, + 530 + ], + "type": "text", + "content": " primarily impacts the ability to translate extreme expression, as indicated by UMTN. Additionally, not using " + }, + { + "bbox": [ + 305, + 431, + 547, + 530 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{sdm}" + }, + { + "bbox": [ + 305, + 431, + 547, + 530 + ], + "type": "text", + "content": " or setting " + }, + { + "bbox": [ + 305, + 431, + 547, + 530 + ], + "type": "inline_equation", + "content": "dim(\\mathbf{z}) = 512" + }, + { + "bbox": [ + 305, + 431, + 547, + 530 + ], + "type": "text", + "content": " results in identity leakage, particularly evident in the latter scenario. For this ablation, we use the part of our driven by MEAD dataset and described in Sec. 6.1" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 557, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 557, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 557, + 547, + 713 + ], + "type": "text", + "content": "facial landmarks from outputs of all methods, frontalized, and normalized to position the mouth edges at " + }, + { + "bbox": [ + 304, + 557, + 547, + 713 + ], + "type": "inline_equation", + "content": "(-1,0)" + }, + { + "bbox": [ + 304, + 557, + 547, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 557, + 547, + 713 + ], + "type": "inline_equation", + "content": "(1,0)" + }, + { + "bbox": [ + 304, + 557, + 547, + 713 + ], + "type": "text", + "content": ". We compute mean absolute error (MAE) for both predicted mouth landmark positions " + }, + { + "bbox": [ + 304, + 557, + 547, + 713 + ], + "type": "inline_equation", + "content": "(M_P)" + }, + { + "bbox": [ + 304, + 557, + 547, + 713 + ], + "type": "text", + "content": " and velocities " + }, + { + "bbox": [ + 304, + 557, + 547, + 713 + ], + "type": "inline_equation", + "content": "(M_V)" + }, + { + "bbox": [ + 304, + 557, + 547, + 713 + ], + "type": "text", + "content": ", along with errors in facial geometry " + }, + { + "bbox": [ + 304, + 557, + 547, + 713 + ], + "type": "inline_equation", + "content": "(F_P)" + }, + { + "bbox": [ + 304, + 557, + 547, + 713 + ], + "type": "text", + "content": " and velocity " + }, + { + "bbox": [ + 304, + 557, + 547, + 713 + ], + "type": "inline_equation", + "content": "(F_V)" + }, + { + "bbox": [ + 304, + 557, + 547, + 713 + ], + "type": "text", + "content": ". For our evaluation, we use 100 randomly selected video sequences from the HDTF dataset [37], each up to 30 seconds. As the results show in Tab. 5, our speech-driven model performs on par with leading methods, excelling in realism and facial dynamics. SyncNet [6] is used to assess lip sync quality, providing offset (temporal misalignment between the audio and video) and confidence (certainty of audio-visual alignment) scores." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8504" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 69, + 69, + 119, + 120 + ], + "blocks": [ + { + "bbox": [ + 69, + 69, + 119, + 120 + ], + "lines": [ + { + "bbox": [ + 69, + 69, + 119, + 120 + ], + "spans": [ + { + "bbox": [ + 69, + 69, + 119, + 120 + ], + "type": "image", + "image_path": "78926baf7339f356df241bf8a88b63a9b83f4b6c3649de1e1ee938d86b29bfe5.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 119, + 70, + 169, + 120 + ], + "blocks": [ + { + "bbox": [ + 119, + 70, + 169, + 120 + ], + "lines": [ + { + "bbox": [ + 119, + 70, + 169, + 120 + ], + "spans": [ + { + "bbox": [ + 119, + 70, + 169, + 120 + ], + "type": "image", + "image_path": "249eb01f4131fdddf201108d1aabb80575d66581bff4a918301a2418cf1ced01.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 170, + 70, + 220, + 120 + ], + "blocks": [ + { + "bbox": [ + 170, + 70, + 220, + 120 + ], + "lines": [ + { + "bbox": [ + 170, + 70, + 220, + 120 + ], + "spans": [ + { + "bbox": [ + 170, + 70, + 220, + 120 + ], + "type": "image", + "image_path": "e6833fe36b387c24e85728fd42f64f6c1ebddbb0976d1cc0ff663463ff80b708.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 222, + 70, + 270, + 120 + ], + "blocks": [ + { + "bbox": [ + 222, + 70, + 270, + 120 + ], + "lines": [ + { + "bbox": [ + 222, + 70, + 270, + 120 + ], + "spans": [ + { + "bbox": [ + 222, + 70, + 270, + 120 + ], + "type": "image", + "image_path": "babf75fa3d53bebbcf1c449f3409b684032610a97ed8f57c97e1cac8db6ce0b7.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 271, + 70, + 320, + 120 + ], + "blocks": [ + { + "bbox": [ + 271, + 70, + 320, + 120 + ], + "lines": [ + { + "bbox": [ + 271, + 70, + 320, + 120 + ], + "spans": [ + { + "bbox": [ + 271, + 70, + 320, + 120 + ], + "type": "image", + "image_path": "69919e8e1c6efec27db2fa17703209ebafd9fb8767da1ce3158618272fad3ff7.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 321, + 70, + 371, + 120 + ], + "blocks": [ + { + "bbox": [ + 321, + 70, + 371, + 120 + ], + "lines": [ + { + "bbox": [ + 321, + 70, + 371, + 120 + ], + "spans": [ + { + "bbox": [ + 321, + 70, + 371, + 120 + ], + "type": "image", + "image_path": "4e727456ce94a10c5cb0cf85dd00e0f157e05f4b06e02049f397c50584441bdd.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 373, + 70, + 421, + 120 + ], + "blocks": [ + { + "bbox": [ + 373, + 70, + 421, + 120 + ], + "lines": [ + { + "bbox": [ + 373, + 70, + 421, + 120 + ], + "spans": [ + { + "bbox": [ + 373, + 70, + 421, + 120 + ], + "type": "image", + "image_path": "40604b2aa996fc6d2acd23a2db2b3411973eea072538acf29f635c8bee37d1b6.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 423, + 70, + 472, + 120 + ], + "blocks": [ + { + "bbox": [ + 423, + 70, + 472, + 120 + ], + "lines": [ + { + "bbox": [ + 423, + 70, + 472, + 120 + ], + "spans": [ + { + "bbox": [ + 423, + 70, + 472, + 120 + ], + "type": "image", + "image_path": "414e09e5425d5e35225c67a5ef070930a8035b17cd7948c4f042cab6758ae06a.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 473, + 70, + 523, + 120 + ], + "blocks": [ + { + "bbox": [ + 473, + 70, + 523, + 120 + ], + "lines": [ + { + "bbox": [ + 473, + 70, + 523, + 120 + ], + "spans": [ + { + "bbox": [ + 473, + 70, + 523, + 120 + ], + "type": "image", + "image_path": "fed734f68f0698bb790e0edb0feec83072cb67a77fe1500b92938e93eb4d9317.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 69, + 125, + 118, + 175 + ], + "blocks": [ + { + "bbox": [ + 69, + 125, + 118, + 175 + ], + "lines": [ + { + "bbox": [ + 69, + 125, + 118, + 175 + ], + "spans": [ + { + "bbox": [ + 69, + 125, + 118, + 175 + ], + "type": "image", + "image_path": "3404ad765f09bb27d600e94b7dfd8336954bec7b7f517c9d336de8e7692620ea.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 119, + 125, + 169, + 175 + ], + "blocks": [ + { + "bbox": [ + 119, + 125, + 169, + 175 + ], + "lines": [ + { + "bbox": [ + 119, + 125, + 169, + 175 + ], + "spans": [ + { + "bbox": [ + 119, + 125, + 169, + 175 + ], + "type": "image", + "image_path": "5d483df94f00a8f21a401b78a622031777a572b55a0ad53f3c54126f06199994.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 170, + 125, + 220, + 175 + ], + "blocks": [ + { + "bbox": [ + 170, + 125, + 220, + 175 + ], + "lines": [ + { + "bbox": [ + 170, + 125, + 220, + 175 + ], + "spans": [ + { + "bbox": [ + 170, + 125, + 220, + 175 + ], + "type": "image", + "image_path": "2f5865bfe29388b182d03ed4d7a9d27dd59feab9fd7f2c0fb1094a9e3561bc51.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 221, + 125, + 269, + 175 + ], + "blocks": [ + { + "bbox": [ + 221, + 125, + 269, + 175 + ], + "lines": [ + { + "bbox": [ + 221, + 125, + 269, + 175 + ], + "spans": [ + { + "bbox": [ + 221, + 125, + 269, + 175 + ], + "type": "image", + "image_path": "514cacf7cd21b690fc389cf3ac74b73b37cf9698264c21cad61806b3e19a0ea3.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 271, + 125, + 320, + 175 + ], + "blocks": [ + { + "bbox": [ + 271, + 125, + 320, + 175 + ], + "lines": [ + { + "bbox": [ + 271, + 125, + 320, + 175 + ], + "spans": [ + { + "bbox": [ + 271, + 125, + 320, + 175 + ], + "type": "image", + "image_path": "6d511aa5d5a396393d77990f87760f332acc794d63f47c209639a816d5ea6435.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 321, + 125, + 371, + 175 + ], + "blocks": [ + { + "bbox": [ + 321, + 125, + 371, + 175 + ], + "lines": [ + { + "bbox": [ + 321, + 125, + 371, + 175 + ], + "spans": [ + { + "bbox": [ + 321, + 125, + 371, + 175 + ], + "type": "image", + "image_path": "70d3e059bb62413c66fd955421186f77044e156d0e45ee6c2b8d47b983c07b88.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 373, + 125, + 421, + 175 + ], + "blocks": [ + { + "bbox": [ + 373, + 125, + 421, + 175 + ], + "lines": [ + { + "bbox": [ + 373, + 125, + 421, + 175 + ], + "spans": [ + { + "bbox": [ + 373, + 125, + 421, + 175 + ], + "type": "image", + "image_path": "18926e0a02b7ae432452adbefe702e8868b7a75fe3646dec0b6b5d9e908bf0d5.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 422, + 125, + 472, + 175 + ], + "blocks": [ + { + "bbox": [ + 422, + 125, + 472, + 175 + ], + "lines": [ + { + "bbox": [ + 422, + 125, + 472, + 175 + ], + "spans": [ + { + "bbox": [ + 422, + 125, + 472, + 175 + ], + "type": "image", + "image_path": "b9020c9c3663e6a403c4af94a5a9fa47af6149c2161bc18ebf0b4b4174ee17e0.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 473, + 125, + 523, + 175 + ], + "blocks": [ + { + "bbox": [ + 473, + 125, + 523, + 175 + ], + "lines": [ + { + "bbox": [ + 473, + 125, + 523, + 175 + ], + "spans": [ + { + "bbox": [ + 473, + 125, + 523, + 175 + ], + "type": "image", + "image_path": "18cada0a23f99cc471fbaa33df513dc3bdc55efddbbf3f720784a087bb0d7786.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 69, + 182, + 118, + 232 + ], + "blocks": [ + { + "bbox": [ + 69, + 182, + 118, + 232 + ], + "lines": [ + { + "bbox": [ + 69, + 182, + 118, + 232 + ], + "spans": [ + { + "bbox": [ + 69, + 182, + 118, + 232 + ], + "type": "image", + "image_path": "abdd874750024f19fafebca20a06b1f70aa928706d99df413ba354979d75ba5c.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 119, + 182, + 169, + 232 + ], + "blocks": [ + { + "bbox": [ + 119, + 182, + 169, + 232 + ], + "lines": [ + { + "bbox": [ + 119, + 182, + 169, + 232 + ], + "spans": [ + { + "bbox": [ + 119, + 182, + 169, + 232 + ], + "type": "image", + "image_path": "8c349ae9a21dac090ccda0bff743883591aeeca9e5450b427eac42b730eaa9f6.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 170, + 182, + 220, + 232 + ], + "blocks": [ + { + "bbox": [ + 170, + 182, + 220, + 232 + ], + "lines": [ + { + "bbox": [ + 170, + 182, + 220, + 232 + ], + "spans": [ + { + "bbox": [ + 170, + 182, + 220, + 232 + ], + "type": "image", + "image_path": "3c10c4d85aa3cca3c2b78f3e2c34e7a1a4bfbaa64af74b33787d9f6fe596d827.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 221, + 182, + 269, + 232 + ], + "blocks": [ + { + "bbox": [ + 221, + 182, + 269, + 232 + ], + "lines": [ + { + "bbox": [ + 221, + 182, + 269, + 232 + ], + "spans": [ + { + "bbox": [ + 221, + 182, + 269, + 232 + ], + "type": "image", + "image_path": "17e23469abdc5c7ab7a6a5550d1010eb601a5984b5074e13f879ce44ad308ecd.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 271, + 182, + 320, + 232 + ], + "blocks": [ + { + "bbox": [ + 271, + 182, + 320, + 232 + ], + "lines": [ + { + "bbox": [ + 271, + 182, + 320, + 232 + ], + "spans": [ + { + "bbox": [ + 271, + 182, + 320, + 232 + ], + "type": "image", + "image_path": "6a6831ddbb77f1d43713ce4c0cf619c551296f51f84ded434ab7b7241a9f5c62.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 321, + 182, + 371, + 232 + ], + "blocks": [ + { + "bbox": [ + 321, + 182, + 371, + 232 + ], + "lines": [ + { + "bbox": [ + 321, + 182, + 371, + 232 + ], + "spans": [ + { + "bbox": [ + 321, + 182, + 371, + 232 + ], + "type": "image", + "image_path": "fd0efa99b095cb5c8d77d101660f6c7bb18b5ee8d5d2d7cb16d3cd628893f63d.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 373, + 182, + 421, + 232 + ], + "blocks": [ + { + "bbox": [ + 373, + 182, + 421, + 232 + ], + "lines": [ + { + "bbox": [ + 373, + 182, + 421, + 232 + ], + "spans": [ + { + "bbox": [ + 373, + 182, + 421, + 232 + ], + "type": "image", + "image_path": "73780beddb0fb11c3d6d1430fac4b29508bdfa7ada34735177e0e1aa0d8d7c4c.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 422, + 182, + 472, + 232 + ], + "blocks": [ + { + "bbox": [ + 422, + 182, + 472, + 232 + ], + "lines": [ + { + "bbox": [ + 422, + 182, + 472, + 232 + ], + "spans": [ + { + "bbox": [ + 422, + 182, + 472, + 232 + ], + "type": "image", + "image_path": "be464290184a7120b5a3faa800bfa8fea8bed08c5546356fdabcb1efb9302981.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 473, + 182, + 523, + 232 + ], + "blocks": [ + { + "bbox": [ + 473, + 182, + 523, + 232 + ], + "lines": [ + { + "bbox": [ + 473, + 182, + 523, + 232 + ], + "spans": [ + { + "bbox": [ + 473, + 182, + 523, + 232 + ], + "type": "image", + "image_path": "455b641de64cefbaf69d4b86b1a3d711d7191a180fdb1111b6aad142ca9b34d4.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 69, + 239, + 118, + 288 + ], + "blocks": [ + { + "bbox": [ + 69, + 239, + 118, + 288 + ], + "lines": [ + { + "bbox": [ + 69, + 239, + 118, + 288 + ], + "spans": [ + { + "bbox": [ + 69, + 239, + 118, + 288 + ], + "type": "image", + "image_path": "c20f63e03b90141305ab3f748fe6e81e7db5f7761224f2e498b225b5bd0fbdb8.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 84, + 294, + 102, + 301 + ], + "lines": [ + { + "bbox": [ + 84, + 294, + 102, + 301 + ], + "spans": [ + { + "bbox": [ + 84, + 294, + 102, + 301 + ], + "type": "text", + "content": "Source" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_caption" + } + ], + "index": 27 + }, + { + "type": "image", + "bbox": [ + 119, + 239, + 169, + 288 + ], + "blocks": [ + { + "bbox": [ + 119, + 239, + 169, + 288 + ], + "lines": [ + { + "bbox": [ + 119, + 239, + 169, + 288 + ], + "spans": [ + { + "bbox": [ + 119, + 239, + 169, + 288 + ], + "type": "image", + "image_path": "914f03cb4787c3219044fec13bdedab69d72544c0aaf586e94c11589a47cc118.jpg" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 135, + 294, + 153, + 300 + ], + "lines": [ + { + "bbox": [ + 135, + 294, + 153, + 300 + ], + "spans": [ + { + "bbox": [ + 135, + 294, + 153, + 300 + ], + "type": "text", + "content": "Driver" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_caption" + } + ], + "index": 29 + }, + { + "type": "image", + "bbox": [ + 170, + 239, + 219, + 288 + ], + "blocks": [ + { + "bbox": [ + 170, + 239, + 219, + 288 + ], + "lines": [ + { + "bbox": [ + 170, + 239, + 219, + 288 + ], + "spans": [ + { + "bbox": [ + 170, + 239, + 219, + 288 + ], + "type": "image", + "image_path": "2a4a11262eaefe3893566bd9f65ab0cf40f2dd90dbc5eb22abf893187532cf4e.jpg" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 180, + 294, + 209, + 300 + ], + "lines": [ + { + "bbox": [ + 180, + 294, + 209, + 300 + ], + "spans": [ + { + "bbox": [ + 180, + 294, + 209, + 300 + ], + "type": "text", + "content": "FOMM [22]" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_caption" + } + ], + "index": 31 + }, + { + "type": "image", + "bbox": [ + 220, + 239, + 269, + 288 + ], + "blocks": [ + { + "bbox": [ + 220, + 239, + 269, + 288 + ], + "lines": [ + { + "bbox": [ + 220, + 239, + 269, + 288 + ], + "spans": [ + { + "bbox": [ + 220, + 239, + 269, + 288 + ], + "type": "image", + "image_path": "167c2eae9f425be2922aad926fe782a8fae893bef8708cefbea97fb2b42f3c59.jpg" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 233, + 294, + 257, + 301 + ], + "lines": [ + { + "bbox": [ + 233, + 294, + 257, + 301 + ], + "spans": [ + { + "bbox": [ + 233, + 294, + 257, + 301 + ], + "type": "text", + "content": "UVA [17]" + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_caption" + } + ], + "index": 33 + }, + { + "type": "image", + "bbox": [ + 271, + 239, + 320, + 288 + ], + "blocks": [ + { + "bbox": [ + 271, + 239, + 320, + 288 + ], + "lines": [ + { + "bbox": [ + 271, + 239, + 320, + 288 + ], + "spans": [ + { + "bbox": [ + 271, + 239, + 320, + 288 + ], + "type": "image", + "image_path": "8a024063cda079847e308c3b84df1cd3d92d6782b38b156413a97c18dec990a1.jpg" + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 277, + 294, + 315, + 301 + ], + "lines": [ + { + "bbox": [ + 277, + 294, + 315, + 301 + ], + "spans": [ + { + "bbox": [ + 277, + 294, + 315, + 301 + ], + "type": "text", + "content": "StyleHEAT [31]" + } + ] + } + ], + "index": 36, + "angle": 0, + "type": "image_caption" + } + ], + "index": 35 + }, + { + "type": "image", + "bbox": [ + 321, + 239, + 371, + 288 + ], + "blocks": [ + { + "bbox": [ + 321, + 239, + 371, + 288 + ], + "lines": [ + { + "bbox": [ + 321, + 239, + 371, + 288 + ], + "spans": [ + { + "bbox": [ + 321, + 239, + 371, + 288 + ], + "type": "image", + "image_path": "ca68c0f5dd22b90c2d4067ccb3bef63902c3770d8b9a7a684fa690295510ba96.jpg" + } + ] + } + ], + "index": 37, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 325, + 294, + 367, + 301 + ], + "lines": [ + { + "bbox": [ + 325, + 294, + 367, + 301 + ], + "spans": [ + { + "bbox": [ + 325, + 294, + 367, + 301 + ], + "type": "text", + "content": "MetaPortrait [35]" + } + ] + } + ], + "index": 38, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 135, + 305, + 457, + 316 + ], + "lines": [ + { + "bbox": [ + 135, + 305, + 457, + 316 + ], + "spans": [ + { + "bbox": [ + 135, + 305, + 457, + 316 + ], + "type": "text", + "content": "Figure 9. A qualitative comparison of head avatar systems in cross-reenactment scenario." + } + ] + } + ], + "index": 45, + "angle": 0, + "type": "image_caption" + } + ], + "index": 37 + }, + { + "type": "image", + "bbox": [ + 373, + 239, + 421, + 288 + ], + "blocks": [ + { + "bbox": [ + 373, + 239, + 421, + 288 + ], + "lines": [ + { + "bbox": [ + 373, + 239, + 421, + 288 + ], + "spans": [ + { + "bbox": [ + 373, + 239, + 421, + 288 + ], + "type": "image", + "image_path": "2f6050cbd5ae51b790b8714d4b1c4c863640bf93ba51abf7f465583f4c63ce65.jpg" + } + ] + } + ], + "index": 39, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 383, + 294, + 410, + 301 + ], + "lines": [ + { + "bbox": [ + 383, + 294, + 410, + 301 + ], + "spans": [ + { + "bbox": [ + 383, + 294, + 410, + 301 + ], + "type": "text", + "content": "NOFA [32]" + } + ] + } + ], + "index": 40, + "angle": 0, + "type": "image_caption" + } + ], + "index": 39 + }, + { + "type": "image", + "bbox": [ + 423, + 239, + 472, + 288 + ], + "blocks": [ + { + "bbox": [ + 423, + 239, + 472, + 288 + ], + "lines": [ + { + "bbox": [ + 423, + 239, + 472, + 288 + ], + "spans": [ + { + "bbox": [ + 423, + 239, + 472, + 288 + ], + "type": "image", + "image_path": "a60b6865628d9d29031984af0b311f49aa111e0d10c11263007409dd4919132c.jpg" + } + ] + } + ], + "index": 41, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 426, + 294, + 468, + 301 + ], + "lines": [ + { + "bbox": [ + 426, + 294, + 468, + 301 + ], + "spans": [ + { + "bbox": [ + 426, + 294, + 468, + 301 + ], + "type": "text", + "content": "MegaPortraits [9]" + } + ] + } + ], + "index": 42, + "angle": 0, + "type": "image_caption" + } + ], + "index": 41 + }, + { + "type": "image", + "bbox": [ + 473, + 239, + 523, + 288 + ], + "blocks": [ + { + "bbox": [ + 473, + 239, + 523, + 288 + ], + "lines": [ + { + "bbox": [ + 473, + 239, + 523, + 288 + ], + "spans": [ + { + "bbox": [ + 473, + 239, + 523, + 288 + ], + "type": "image", + "image_path": "5cd2ec3d01fca131da0568ab9ce41e09a269a4f1cdc110e312097c112488ba8f.jpg" + } + ] + } + ], + "index": 43, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 491, + 294, + 504, + 300 + ], + "lines": [ + { + "bbox": [ + 491, + 294, + 504, + 300 + ], + "spans": [ + { + "bbox": [ + 491, + 294, + 504, + 300 + ], + "type": "text", + "content": "Ours" + } + ] + } + ], + "index": 44, + "angle": 0, + "type": "image_caption" + } + ], + "index": 43 + }, + { + "type": "table", + "bbox": [ + 50, + 342, + 286, + 388 + ], + "blocks": [ + { + "bbox": [ + 50, + 342, + 286, + 388 + ], + "lines": [ + { + "bbox": [ + 50, + 342, + 286, + 388 + ], + "spans": [ + { + "bbox": [ + 50, + 342, + 286, + 388 + ], + "type": "table", + "html": "
MethodFID↓\\( F_P \\downarrow \\)\\( F_V \\downarrow \\)\\( M_P \\downarrow \\)\\( M_V \\downarrow \\)Sync. offset ↓Sync. conf. ↑
StyleHEAT [31]62.41.340.412.981.670.417.22
PC-AVS [39]71.52.780.512.561.043.798.42
SadTalker [36]55.60.910.352.121.170.277.58
MakeIRTalk [40]63.32.150.383.111.592.985.17
Ours (EMOPortraits)28.50.820.332.531.380.075.77
", + "image_path": "234c49cf3f3f200a5accfefb0aaf8eed6b3d111522a888153bde86d510c0a6e8.jpg" + } + ] + } + ], + "index": 46, + "angle": 0, + "type": "table_body" + } + ], + "index": 46 + }, + { + "type": "table", + "bbox": [ + 50, + 559, + 286, + 613 + ], + "blocks": [ + { + "bbox": [ + 46, + 396, + 289, + 529 + ], + "lines": [ + { + "bbox": [ + 46, + 396, + 289, + 529 + ], + "spans": [ + { + "bbox": [ + 46, + 396, + 289, + 529 + ], + "type": "text", + "content": "Table 5. Our model matches other leading speech-driven models, exceeding in realism (FID), facial geometry and velocity, but is slightly less accurate in mouth landmark positions and velocities. We also score best in audio-video sync offset, but lower in confidence. We believe that while SyncNet metrics are still helpful, they may not fully represent true synchronization quality, as generated videos can sometimes score higher than ground truth without necessarily being better synchronized. In our experimentation, ground truth videos scored an offset of 1.13 (which is worse than our method, SadTalker, and StyleHEAT) and a confidence score of 8.35 (less than PC-AVS). For qualitative analysis, please refer to our supplementary materials." + } + ] + } + ], + "index": 47, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 50, + 559, + 286, + 613 + ], + "lines": [ + { + "bbox": [ + 50, + 559, + 286, + 613 + ], + "spans": [ + { + "bbox": [ + 50, + 559, + 286, + 613 + ], + "type": "table", + "html": "
MethodFID↓MP↓MV↓FP↓FV↓
Ours w/o disentanglement44.70.980.455.782.55
Ours w/o LPCA29.30.840.353.131.88
Ours with MAE(\\( \\mathbf{z}_i, \\hat{\\mathbf{z}}_i^{\\text{speech}} \\))28.20.840.342.611.45
Ours28.50.820.332.531.38
", + "image_path": "a33ea7fa3dd7e2f88d704e8bd1b0a93c9948c355710e8868de8a970281300f33.jpg" + } + ] + } + ], + "index": 48, + "angle": 0, + "type": "table_body" + } + ], + "index": 48 + }, + { + "bbox": [ + 46, + 622, + 288, + 710 + ], + "lines": [ + { + "bbox": [ + 46, + 622, + 288, + 710 + ], + "spans": [ + { + "bbox": [ + 46, + 622, + 288, + 710 + ], + "type": "text", + "content": "Table 6. Ablation study of our speech-driven mode, where in the first line, we show that it notably underperforms without head pose - face expression disentanglement of the latent space (see Sec. 5.1). Removing " + }, + { + "bbox": [ + 46, + 622, + 288, + 710 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{PCA}}" + }, + { + "bbox": [ + 46, + 622, + 288, + 710 + ], + "type": "text", + "content": " leads to a significant reduction in generating realistic mouth movements. Replacing " + }, + { + "bbox": [ + 46, + 622, + 288, + 710 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{PCA}}" + }, + { + "bbox": [ + 46, + 622, + 288, + 710 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 46, + 622, + 288, + 710 + ], + "type": "inline_equation", + "content": "\\mathrm{MAE}(\\mathbf{z}_i,\\hat{\\mathbf{z}}_i^{\\mathrm{speech}})" + }, + { + "bbox": [ + 46, + 622, + 288, + 710 + ], + "type": "text", + "content": " causes a noticeable, decline in mouth movement generation. This suggests that identifying mouth-related PCA components was beneficial in enhancing the outcome." + } + ] + } + ], + "index": 49, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 306, + 338, + 398, + 350 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 338, + 398, + 350 + ], + "spans": [ + { + "bbox": [ + 306, + 338, + 398, + 350 + ], + "type": "text", + "content": "6.3. Ablation study" + } + ] + } + ], + "index": 50 + }, + { + "bbox": [ + 304, + 356, + 548, + 501 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 356, + 548, + 501 + ], + "spans": [ + { + "bbox": [ + 304, + 356, + 548, + 501 + ], + "type": "text", + "content": "We conducted an extensive ablation study to evaluate the contributions of individual components within our method. For our main model, we present the evaluation of the importance of the proposed source-driver unmatched loss, canonical volume loss, and change of " + }, + { + "bbox": [ + 304, + 356, + 548, + 501 + ], + "type": "inline_equation", + "content": "dim(\\mathbf{z})" + }, + { + "bbox": [ + 304, + 356, + 548, + 501 + ], + "type": "text", + "content": " from 512 in original model to 128 Tab. 4. For our speech-driven mode, we show how vital the disentanglement of the latent space for speech-drive ability, how " + }, + { + "bbox": [ + 304, + 356, + 548, + 501 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{PCA}}" + }, + { + "bbox": [ + 304, + 356, + 548, + 501 + ], + "type": "text", + "content": " is better than naive matching of PCA components from random " + }, + { + "bbox": [ + 304, + 356, + 548, + 501 + ], + "type": "inline_equation", + "content": "\\mathcal{Z}" + }, + { + "bbox": [ + 304, + 356, + 548, + 501 + ], + "type": "text", + "content": " and how it helps match speech and latent expression space Tab. 6. For more ablation analysis, plots and visual comparison, please refer to our supplementary." + } + ] + } + ], + "index": 51 + }, + { + "bbox": [ + 306, + 513, + 379, + 525 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 513, + 379, + 525 + ], + "spans": [ + { + "bbox": [ + 306, + 513, + 379, + 525 + ], + "type": "text", + "content": "7. Conclusion" + } + ] + } + ], + "index": 52 + }, + { + "bbox": [ + 304, + 533, + 547, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 533, + 547, + 616 + ], + "spans": [ + { + "bbox": [ + 304, + 533, + 547, + 616 + ], + "type": "text", + "content": "In this paper, we introduce EmoPortraits, a novel method for creating neural avatars with superior performance in image-driven, cross-identity emotion translation. Our speech-driven mode makes it possible to drive the facial animation through multiple conditions (video, audio, head motion). We collected FEED dataset which, we believe, will be a valuable asset for researchers in diverse human-centered studies." + } + ] + } + ], + "index": 53 + }, + { + "bbox": [ + 304, + 617, + 548, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 617, + 548, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 617, + 548, + 712 + ], + "type": "text", + "content": "However, our method has some limitations. It doesn't generate the avatar's body or shoulders, limiting some use cases. We currently integrate our output with a source image body. Additionally, the model sometimes struggles with accurate expression translation and underperforms with extensive head rotation. These challenges are crucial for future enhancements and remain central to our ongoing research efforts." + } + ] + } + ], + "index": 54 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8505" + } + ] + } + ], + "index": 55 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "type": "text", + "content": "[1] Emad Barsoum, Cha Zhang, Cristian Canton Ferrer, and Zhengyou Zhang. Training deep networks for facial expression recognition with crowd-sourced label distribution. In Proceedings of the 18th ACM international conference on multimodal interaction, pages 279-283, 2016. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 148, + 288, + 190 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 148, + 288, + 190 + ], + "spans": [ + { + "bbox": [ + 53, + 148, + 288, + 190 + ], + "type": "text", + "content": "[2] Stella Bounareli, Christos Tzelepis, Vasileios Argyriou, Ioannis Patras, and Georgios Tzimiropoulos. Hyperreenact: One-shot reenactment via jointly learning to refine and retarget faces, 2023. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 193, + 288, + 236 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 193, + 288, + 236 + ], + "spans": [ + { + "bbox": [ + 53, + 193, + 288, + 236 + ], + "type": "text", + "content": "[3] Egor Burkov, Igor Pasechnik, Artur Grigorev, and Victor Lempitsky. Neural head reenactment with latent pose descriptors. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). IEEE, 2020. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 238, + 288, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 238, + 288, + 281 + ], + "spans": [ + { + "bbox": [ + 53, + 238, + 288, + 281 + ], + "type": "text", + "content": "[4] Houwei Cao, David G Cooper, Michael K Keutmann, Ruben C Gur, Ani Nenkova, and Ragini Verma. Crema-d: Crowd-sourced emotional multimodal actors dataset. IEEE transactions on affective computing, 5(4):377-390, 2014. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 283, + 288, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 283, + 288, + 316 + ], + "spans": [ + { + "bbox": [ + 53, + 283, + 288, + 316 + ], + "type": "text", + "content": "[5] Qiong Cao, Li Shen, Weidi Xie, Omkar M. Parkhi, and Andrew Zisserman. Vggface2: A dataset for recognising faces across pose and age, 2018. 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 317, + 288, + 372 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 317, + 288, + 372 + ], + "spans": [ + { + "bbox": [ + 53, + 317, + 288, + 372 + ], + "type": "text", + "content": "[6] Joon Son Chung and Andrew Zisserman. Out of time: automated lip sync in the wild. In Computer Vision-ACCV 2016 Workshops: ACCV 2016 International Workshops, Taipei, Taiwan, November 20-24, 2016, Revised Selected Papers, Part II 13, pages 251-263. Springer, 2017. 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 373, + 288, + 405 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 373, + 288, + 405 + ], + "spans": [ + { + "bbox": [ + 53, + 373, + 288, + 405 + ], + "type": "text", + "content": "[7] Joon Son Chung, Arsha Nagrani, and Andrew Zisserman. Voxceleb2: Deep speaker recognition. In INTERSPEECH, 2018. 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 407, + 288, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 407, + 288, + 441 + ], + "spans": [ + { + "bbox": [ + 53, + 407, + 288, + 441 + ], + "type": "text", + "content": "[8] Michail Christos Doukas, Stefanos Zafeiriou, and Viktoriia Sharmanska. Headgan: One-shot neural head synthesis and editing, 2021. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 442, + 288, + 485 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 442, + 288, + 485 + ], + "spans": [ + { + "bbox": [ + 53, + 442, + 288, + 485 + ], + "type": "text", + "content": "[9] Nikita Drobyshev, Jenya Chelishev, Taras Khakhulin, Aleksei Ivakhnenko, Victor Lempitsky, and Egor Zakharov. Megaportraits: One-shot megapixel neural head avatars, 2023. 2, 4, 5, 7, 8" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 487, + 287, + 519 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 487, + 287, + 519 + ], + "spans": [ + { + "bbox": [ + 48, + 487, + 287, + 519 + ], + "type": "text", + "content": "[10] Guy Gafni, Justus Thies, Michael Zollhöfer, and Matthias Nießner. Dynamic neural radiance fields for monocular 4d facial avatar reconstruction, 2020. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 521, + 287, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 521, + 287, + 553 + ], + "spans": [ + { + "bbox": [ + 48, + 521, + 287, + 553 + ], + "type": "text", + "content": "[11] Philip-William Grassal, Malte Prinzler, Titus Leistner, Carsten Rother, Matthias Nießner, and Justus Thies. Neural head avatars from monocular rgb videos, 2022. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 555, + 287, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 555, + 287, + 588 + ], + "spans": [ + { + "bbox": [ + 48, + 555, + 287, + 588 + ], + "type": "text", + "content": "[12] Siddharth Gururani, Arun Mallya, Ting-Chun Wang, Rafael Valle, and Ming-Yu Liu. Space: Speech-driven portrait animation with controllable expression, 2022. 3, 7" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 590, + 288, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 590, + 288, + 643 + ], + "spans": [ + { + "bbox": [ + 48, + 590, + 288, + 643 + ], + "type": "text", + "content": "[13] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. In Advances in Neural Information Processing Systems, 2017. 7" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 646, + 288, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 646, + 288, + 678 + ], + "spans": [ + { + "bbox": [ + 48, + 646, + 288, + 678 + ], + "type": "text", + "content": "[14] Philip Jackson and SJUoSG Haq. Surrey audio-visual expressed emotion (savee) database. University of Surrey: Guildford, UK, 2014. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 680, + 288, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 680, + 288, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 680, + 288, + 712 + ], + "type": "text", + "content": "[15] Taras Khakhulin, Vanessa Sklyarova, Victor Lempitsky, and Egor Zakharov. Realistic one-shot mesh-based head avatars, 2022. 2" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 713 + ], + "type": "list", + "angle": 0, + "index": 31, + "blocks": [ + { + "bbox": [ + 307, + 73, + 547, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 547, + 106 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 547, + 106 + ], + "type": "text", + "content": "[16] Alexander C. Li, Alexei A. Efros, and Deepak Pathak. Understanding collapse in non-contrastive siamese representation learning, 2022. 4" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 108, + 545, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 108, + 545, + 139 + ], + "spans": [ + { + "bbox": [ + 307, + 108, + 545, + 139 + ], + "type": "text", + "content": "[17] Xueting Li, Shalini De Mello, Sifei Liu, Koki Nagano, Umar Iqbal, and Jan Kautz. Generalizable one-shot neural head avatar, 2023. 7, 8" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 141, + 547, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 141, + 547, + 185 + ], + "spans": [ + { + "bbox": [ + 307, + 141, + 547, + 185 + ], + "type": "text", + "content": "[18] Steven R Livingstone and Frank A Russo. The ryerson audiovisual database of emotional speech and song (ravdess): A dynamic, multimodal set of facial and vocal expressions in north american english. PloS one, 13(5):e0196391, 2018. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 186, + 547, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 186, + 547, + 252 + ], + "spans": [ + { + "bbox": [ + 307, + 186, + 547, + 252 + ], + "type": "text", + "content": "[19] Patrick Lucey, Jeffrey F Cohn, Takeo Kanade, Jason Saragih, Zara Ambadar, and Iain Matthews. The extended cohn-kanade dataset (ck+): A complete dataset for action unit and emotion-specified expression. In 2010 iiee computer society conference on computer vision and pattern recognition-workshops, pages 94-101. IEEE, 2010. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 254, + 547, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 254, + 547, + 297 + ], + "spans": [ + { + "bbox": [ + 307, + 254, + 547, + 297 + ], + "type": "text", + "content": "[20] Maja Pantic, Michel Valstar, Ron Rademaker, and Ludo Maat. Web-based database for facial expression analysis. In 2005 IEEE international conference on multimedia and Expo, pages 5-pp. IEEE, 2005. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 299, + 545, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 299, + 545, + 352 + ], + "spans": [ + { + "bbox": [ + 307, + 299, + 545, + 352 + ], + "type": "text", + "content": "[21] KR Prajwal, Rudrabha Mukhopadhyay, Vinay P Namboodiri, and CV Jawahar. A lip sync expert is all you need for speech to lip generation in the wild. In Proceedings of the 28th ACM international conference on multimedia, pages 484-492, 2020. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 354, + 545, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 354, + 545, + 387 + ], + "spans": [ + { + "bbox": [ + 307, + 354, + 545, + 387 + ], + "type": "text", + "content": "[22] Aliaksandr Siarohin, Stéphane Lathuilière, S. Tulyakov, Elisa Ricci, and N. Sebe. First order motion model for image animation. ArXiv, abs/2003.00196, 2019. 2, 7, 8" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 388, + 545, + 432 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 388, + 545, + 432 + ], + "spans": [ + { + "bbox": [ + 307, + 388, + 545, + 432 + ], + "type": "text", + "content": "[23] Aliaksandr Siarohin, Willi Menapace, Ivan Skorokhodov, Kyle Olszewski, Hsin-Ying Lee, Jian Ren, Mengelei Chai, and Sergey Tulyakov. Unsupervised volumetric animation. arXiv preprint arXiv:2301.11326, 2023. 7" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 434, + 547, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 434, + 547, + 475 + ], + "spans": [ + { + "bbox": [ + 307, + 434, + 547, + 475 + ], + "type": "text", + "content": "[24] Michal Stypulkowski, Konstantinos Vougioukas, Sen He, Maciej Zieba, Stavros Petridis, and Maja Pantic. Diffused heads: Diffusion models beat gans on talking-face generation, 2023. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 478, + 545, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 478, + 545, + 521 + ], + "spans": [ + { + "bbox": [ + 307, + 478, + 545, + 521 + ], + "type": "text", + "content": "[25] Mani Kumar Tellamekala, Ömer Ümer, Björn W. Schuller, Elisabeth Andre, Timo Giesbrecht, and Michel Valstar. Are 3d face shapes expressive enough for recognising continuous emotions and action unit intensities?, 2023. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 523, + 545, + 556 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 523, + 545, + 556 + ], + "spans": [ + { + "bbox": [ + 307, + 523, + 545, + 556 + ], + "type": "text", + "content": "[26] Hao Wang, Yitong Wang, Zheng Zhou, Xing Ji, Dihong Gong, Jingchao Zhou, Zhifeng Li, and Wei Liu. Cosface: Large margin cosine loss for deep face recognition, 2018. 5" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 558, + 545, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 558, + 545, + 612 + ], + "spans": [ + { + "bbox": [ + 307, + 558, + 545, + 612 + ], + "type": "text", + "content": "[27] H. Wang, Yitong Wang, Zheng Zhou, Xing Ji, Zhifeng Li, Dihong Gong, Jin Zhou, and Wenyu Liu. Cosface: Large margin cosine loss for deep face recognition. 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5265-5274, 2018. 6" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 613, + 545, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 613, + 545, + 669 + ], + "spans": [ + { + "bbox": [ + 307, + 613, + 545, + 669 + ], + "type": "text", + "content": "[28] Kaisiyuan Wang, Qianyi Wu, Linsen Song, Zhuoqian Yang, Wayne Wu, Chen Qian, Ran He, Yu Qiao, and Chen Change Loy. Mead: A large-scale audio-visual dataset for emotional talking-face generation. In European Conference on Computer Vision, pages 700–717. Springer, 2020. 3" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 670, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 670, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 670, + 545, + 713 + ], + "type": "text", + "content": "[29] Ting-Chun Wang, Arun Mallya, and Ming-Yu Liu. One-shot free-view neural talking-head synthesis for video conferencing. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2021. 2" + } + ] + } + ], + "index": 30 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "8506" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 554 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "text", + "content": "[30] Yuelang Xu, Lizhen Wang, Xiaochen Zhao, Hongwen Zhang, and Yebin Liu. AvatarMAV: Fast 3d head avatar reconstruction using motion-aware neural voxels. In *Special Interest Group on Computer Graphics and Interactive Techniques Conference Conference Proceedings*. ACM, 2023. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 129, + 288, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 129, + 288, + 183 + ], + "spans": [ + { + "bbox": [ + 48, + 129, + 288, + 183 + ], + "type": "text", + "content": "[31] Fei Yin, Yong Zhang, Xiaodong Cun, Mingdeng Cao, Yanbo Fan, Xuan Wang, Qingyan Bai, Baoyuan Wu, Jue Wang, and Yujiu Yang. Styleheat: One-shot high-resolution editable talking face generation via pre-trained stylegan, 2022. 2, 3, 7, 8" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 185, + 288, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 185, + 288, + 228 + ], + "spans": [ + { + "bbox": [ + 48, + 185, + 288, + 228 + ], + "type": "text", + "content": "[32] Wangbo Yu, Yanbo Fan, Yong Zhang, Xuan Wang, Fei Yin, Yunpeng Bai, Yan-Pei Cao, Ying Shan, Yang Wu, Zhongqian Sun, and Baoyuan Wu. Nofa: Nerf-based one-shot facial avatar reconstruction, 2023. 2, 7, 8" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 230, + 288, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 230, + 288, + 262 + ], + "spans": [ + { + "bbox": [ + 48, + 230, + 288, + 262 + ], + "type": "text", + "content": "[33] Egor Zakharov, Aliaksandra Shysheya, Egor Burkov, and Victor Lempitsky. Few-shot adversarial learning of realistic neural talking head models, 2019. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 263, + 288, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 263, + 288, + 296 + ], + "spans": [ + { + "bbox": [ + 48, + 263, + 288, + 296 + ], + "type": "text", + "content": "[34] Egor Zakharov, Aleksei Ivakhnenko, Aliaksandra Shysheya, and Victor Lempitsky. Fast bi-layer neural synthesis of one-shot realistic head avatars, 2020. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 297, + 288, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 297, + 288, + 342 + ], + "spans": [ + { + "bbox": [ + 48, + 297, + 288, + 342 + ], + "type": "text", + "content": "[35] Bowen Zhang, Chenyang Qi, Pan Zhang, Bo Zhang, Hsiang-Tao Wu, Dong Chen, Qifeng Chen, Yong Wang, and Fang Wen. Metaportrait: Identity-preserving talking head generation with fast personalized adaptation, 2023. 2, 7, 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 343, + 288, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 343, + 288, + 386 + ], + "spans": [ + { + "bbox": [ + 48, + 343, + 288, + 386 + ], + "type": "text", + "content": "[36] Wenxuan Zhang, Xiaodong Cun, Xuan Wang, Yong Zhang, Xi Shen, Yu Guo, Ying Shan, and Fei Wang. Sadtalker: Learning realistic 3d motion coefficients for stylized audio-driven single image talking face animation, 2023. 3, 7, 8" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 388, + 288, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 388, + 288, + 441 + ], + "spans": [ + { + "bbox": [ + 48, + 388, + 288, + 441 + ], + "type": "text", + "content": "[37] Zhimeng Zhang, Lincheng Li, Yu Ding, and Changjie Fan. Flow-guided one-shot talking face generation with a high-resolution audio-visual dataset. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3661–3670, 2021. 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 443, + 288, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 443, + 288, + 475 + ], + "spans": [ + { + "bbox": [ + 48, + 443, + 288, + 475 + ], + "type": "text", + "content": "[38] Yufeng Zheng, Victoria Fernández Abrevaya, Marcel C. Bühler, Xu Chen, Michael J. Black, and Otmar Hilliges. I m avatar: Implicit morphable head avatars from videos, 2022. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 477, + 288, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 477, + 288, + 520 + ], + "spans": [ + { + "bbox": [ + 48, + 477, + 288, + 520 + ], + "type": "text", + "content": "[39] Hang Zhou, Yasheng Sun, Wayne Wu, Chen Change Loy, Xiaogang Wang, and Ziwei Liu. Pose-controllable talking face generation by implicitly modularized audio-visual representation, 2021. 2, 3, 7, 8" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 522, + 288, + 554 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 522, + 288, + 554 + ], + "spans": [ + { + "bbox": [ + 48, + 522, + 288, + 554 + ], + "type": "text", + "content": "[40] Yang Zhou, Xintong Han, Eli Shechtman, Jose Echevarria, Evangelos Kalogerakis, and Dingzeyu Li. MakeltTalk. ACM Transactions on Graphics, 39(6):1-15, 2020. 3, 7, 8" + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8507" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/ERMVP_ Communication-Efficient and Collaboration-Robust Multi-Vehicle Perception in Challenging Environments/197e9226-8724-4a9f-ac9a-1556459fa452_content_list.json b/2024/ERMVP_ Communication-Efficient and Collaboration-Robust Multi-Vehicle Perception in Challenging Environments/197e9226-8724-4a9f-ac9a-1556459fa452_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..8e2a0c320f17e80a5f3d9266c402553b5d34975a --- /dev/null +++ b/2024/ERMVP_ Communication-Efficient and Collaboration-Robust Multi-Vehicle Perception in Challenging Environments/197e9226-8724-4a9f-ac9a-1556459fa452_content_list.json @@ -0,0 +1,1450 @@ +[ + { + "type": "text", + "text": "ERMVP: Communication-Efficient and Collaboration-Robust Multi-Vehicle Perception in Challenging Environments", + "text_level": 1, + "bbox": [ + 100, + 161, + 868, + 203 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Jingyu Zhang $^{1}$ Kun Yang $^{1}$ Yilei Wang $^{1}$ Hanqi Wang $^{1}$ Peng Sun $^{2,*}$ Liang Song $^{1,*}$ $^{1}$ Academy for Engineering and Technology, Fudan University $^{2}$ Duke Kunshan University \n{jingyuzhang22, yileiwang23}@m.fudan.edu.cn, songl@m.fudan.edu.cn", + "bbox": [ + 131, + 226, + 836, + 277 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 308, + 312, + 324 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Collaborative perception enhances perception performance by enabling autonomous vehicles to exchange complementary information. Despite its potential to revolutionize the mobile industry, challenges in various environments, such as communication bandwidth limitations, localization errors and information aggregation inefficiencies, hinder its implementation in practical applications. In this work, we propose ERMVP, a communication-Efficient and collaboration-Robust Multi-Vehicle Perception method in challenging environments. Specifically, ERMVP has three distinct strengths: i) It utilizes the hierarchical feature sampling strategy to abstract a representative set of feature vectors, using less communication overhead for efficient communication; ii) It employs the sparse consensus features to execute precise spatial location calibrations, effectively mitigating the implications of vehicle localization errors; iii) A pioneering feature fusion and interaction paradigm is introduced to integrate holistic spatial semantics among different vehicles and data sources. To thoroughly validate our method, we conduct extensive experiments on real-world and simulated datasets. The results demonstrate that the proposed ERMVP is significantly superior to the state-of-the-art collaborative perception methods.", + "bbox": [ + 75, + 338, + 472, + 657 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 77, + 684, + 208, + 698 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Autonomous vehicles are widely recognized as a valuable means to enhance road safety and traffic efficiency. Equipped with lidar, cameras, and other sensors, these vehicles are capable of accurately sensing their surroundings to ensure safe and reliable operation. However, the single-vehicle perception system has inevitable drawbacks [26, 39], such as a limited sensor field of view that can be easily obstructed and the challenge of detecting distant objects due to sparse and low-resolution data. Recently, the", + "bbox": [ + 75, + 707, + 468, + 834 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "advances in vehicle-to-vehicle (V2V) communication technologies [12, 17, 36] and deep learning [4-6, 10, 16, 24, 45] have spurred innovation and progress in the collaborative perception technology. This technology allows connected autonomous vehicles (CAVs) to share sensory data, leading to more comprehensive environmental perception.", + "bbox": [ + 499, + 309, + 890, + 394 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Although collaborative perception technology shows great potential in transforming the mobility industry, its practical application faces several challenges, including communication bandwidth limitations [15, 35], localization errors [11, 25] and information aggregation inefficiencies [38, 47]. In practical situations, wireless communication resource and reliability constraints severely hamper the efficacy of delay-sensitive collaborative perception. While recent works [15, 42] have achieved a balance between perception performance and communication bandwidth through well-designed mechanisms, these methods have their limitations because they primarily considering information compression over spatial redundancy. This narrow focus exacerbates performance degradation at high compression ratios.", + "bbox": [ + 497, + 394, + 892, + 600 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Further, complex dynamic environments lead to localization errors, which result in inaccurate relative transform estimates and spatial feature misalignment. This relative pose noise produces misleading features that adversely affect the effectiveness of collaborative perception. Existing methods [25, 34] attempt to optimize the overall pose through intensive computation, but the high latency makes them unsuitable for real-time dynamic perception. Meanwhile, collaborative methods [3, 38, 41, 42, 46] only focus on aggregated information, but overlook the inherent perceptual strengths of ego vehicle. This paradigm is vulnerable to the perturbations introduced by collaborative noise, including asynchronous motion blur and inaccurate projections. Such drawback becomes a bottleneck for achieving optimal perceptual performance. In contrast, ego-centric features may contain locally accurate spatial location information that is not affected by collaborative noise. Therefore, a priority for establishing a pragmatic collaborative perception system is to effectively overcome the above challenges.", + "bbox": [ + 497, + 603, + 894, + 866 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 127, + 49, + 200, + 85 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 252, + 48, + 791, + 87 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding authors. Our code is available at https://github.com/Terry9a/ERMVP.", + "bbox": [ + 75, + 842, + 468, + 866 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "12575", + "bbox": [ + 465, + 887, + 505, + 897 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/17285042ad72afee503d390d9bf80d325dc5985f785ca8f0682e6d8e69e83e0c.jpg", + "image_caption": [ + "Figure 1. The overall architecture of the proposed framework. The framework consists of six phases: metadata sharing and feature extraction, filter and merge feature sampling, feature spatial calibration, attention-based feature fusion, accuracy enhanced feature interaction and detection decoders. The details of each individual component are illustrated in Section 3." + ], + "image_footnote": [], + "bbox": [ + 77, + 122, + 890, + 371 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Based on these observations, we propose ERMVP, a communication-efficient and collaboration-robust multi-vehicle perception method in challenging environments. From Figure 1, ERMVP uses four proposed core components to tackle existing challenges jointly. Specifically, (i) we first design an advanced filter and merge feature sampling strategy to tackle the limitations of wireless communication resources. This strategy considers both inter-class and intra-class redundancy relationships to abstract a refined set of feature vectors from the redundant features, using less communication overhead for efficient communication. (ii) Second, we introduce a plug-and-play feature spatial calibration module to mitigate the implications of vehicle localization errors. This module ingeniously utilizes consensus sparse foreground features to align the relative pose relationships between ego-vehicle and collaborators without any precise pose supervision. (iii) Furthermore, we present a pioneering feature fusion and interaction paradigm to integrate holistic spatial semantics. This paradigm comprises two key components: The first is an attention-based feature fusion module that alternates between local and global attention to fuse heterogeneous information from different vehicles. The second is an accuracy enhanced feature interaction strategy that leverages the accurate positional information inherent in ego-centric features to enhance the rich semantic information provided by fused features. Through these tailored components, ERMVP represents a significant advancement towards pragmatic collaborative perception. To validate the effectiveness of the ER-MVP, we conduct extensive experiments on two collaborative 3D object detection datasets including V2V4Real [44]", + "bbox": [ + 75, + 435, + 472, + 865 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "and OPV2V [43]. Comprehensive experimental results demonstrate that our method outperforms previous state-of-the-art methods under the bandwidth-limited noisy setting. The main contributions can be summarized as follows:", + "bbox": [ + 497, + 435, + 890, + 491 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We present ERMVP, a communication-efficient and collaboration-robust multi-vehicle perception method, which addresses the communication bandwidth limitations, localization errors and information aggregation efficiency challenges.", + "- We develop a filter and merge feature sampling strategy to enhance communication efficiency, a feature spatial calibration module for accurate spatial feature alignment, and two information aggregation components to optimize the fusion process.", + "- We conduct extensive experiments on both real-world and simulated datasets. The results show the superiority of our method and the necessity of the proposed components." + ], + "bbox": [ + 499, + 491, + 890, + 682 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 499, + 697, + 640, + 711 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Multi-Agent Communication", + "text_level": 1, + "bbox": [ + 499, + 720, + 757, + 733 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Communication has played a pivotal role in the development of robust multi-agent systems. Early multi-agent communication [20, 29, 32] relied on predefined protocols and heuristics to regulate communication between agents. However, these fixed methods are unsuitable in complex and dynamic settings. Advances in deep learning have inspired the development of advanced strategies for information exchange and collaboration among agents. For example, MAGIC [27] used a graph attention encoder to solve the", + "bbox": [ + 497, + 741, + 890, + 866 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "12576", + "bbox": [ + 465, + 887, + 505, + 897 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "problem of when and to whom to send messages. TMC [49] utilized temporal smoothing to ensure efficient communication. EC-MARL [2] tackled high-dimensional continuous control and partially observable states by introducing a rapid communication protocol to resolve task dilemmas. In comparison, our research focuses on LiDAR-based collaborative 3D object detection tasks in complex driving scenarios. We propose a hierarchical feature sampling strategy to achieve efficient communication across agents.", + "bbox": [ + 75, + 125, + 472, + 251 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. Collaborative Perception", + "text_level": 1, + "bbox": [ + 77, + 261, + 305, + 275 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Factors including limited fields of view of sensors and physical obstructions in the environment can adversely impact the perception capabilities of individual agents. To address these challenges, collaborative perception within multi-agent systems has become a key technology. Utilizing newly available datasets [22, 43, 44], several novel collaborative perception methods have been proposed. For example, V2VNet [38] incorporated a graph neural network to fuse data from different agents. V2X-ViT [42] introduced a transformer architecture that combined information from vehicles and infrastructures. Where2comm [15] used feature spatial heterogeneity to reduce bandwidth utilization by transmitting sparse feature maps. CoBEVT [41] proposed the first multi-camera-based collaborative perception framework and designed the fused axis attention module to enable multi-view interactions. In this work, we propose an innovative feature calibration and interaction method that ensures robust vehicle collaboration in high-noise dynamic environments.", + "bbox": [ + 75, + 282, + 472, + 545 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 77, + 560, + 169, + 574 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "This section introduces our efficient and robust collaborative perception method. Figure 1 shows the general procedure of our method, which is divided into six phases. We will detail each phase in the following introduction.", + "bbox": [ + 75, + 584, + 468, + 641 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Metadata Sharing and Feature Extraction", + "text_level": 1, + "bbox": [ + 75, + 650, + 436, + 666 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In multi-vehicle collaboration scenarios, one of the CAVs, known as the ego vehicle, constructs a communication graph. In this graph, the ego vehicle serves as the requester, while other vehicles within its communication range act as supporters. The ego vehicle broadcasts its metadata information, including position, heading, speed, and more. Supporters, upon receiving this metadata, project their local point cloud observations into the ego vehicle's coordinate system. Each vehicle then encodes these transformed point clouds into bird's eye view (BEV) features, yielding a visual representation. Given the $i$ -th vehicle local observations $\\mathbf{X}_i$ , the extracted features are represented as $\\mathbf{F}_i = \\Phi_{\\mathrm{enc}}(\\mathbf{X}_i) \\in \\mathbb{R}^{H \\times W \\times C}$ , where $\\Phi_{\\mathrm{enc}}(\\cdot)$ denotes the PointPillar [19] encoder shared by all vehicles and $H$ , $W$ ,", + "bbox": [ + 75, + 672, + 472, + 866 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "and $C$ stand for the height, width, and channel of the feature map, respectively. Then, these extracted features are fed into the filter and merge feature sampling module.", + "bbox": [ + 497, + 125, + 890, + 168 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Filter and Merge Feature Sampling", + "text_level": 1, + "bbox": [ + 497, + 174, + 808, + 190 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Previous works have utilized well-designed mechanisms such as information entropy communication selection [37] and spatial heterogeneity map [15, 35] to reduce the required transmission bandwidth. However, these methods primarily focus on the inter-class redundancy between foreground and background features, ignoring the intra-class redundancy among features, which results in sub-optimal compression. To address this gap, we introduce an advanced Filtering and Merging Feature Sampling strategy (FMS). This strategy considers both inter-class and intra-class redundancy relationships, efficiently extracting a concise and distinctive set of feature vectors from the original feature maps, thus reducing communication overheads more effectively. FMS is composed of two core components as follows.", + "bbox": [ + 497, + 195, + 890, + 402 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Filter Sampler. In object detection, foreground areas containing objects are more significant than the background areas. Therefore, we implement the idea of reducing spatial redundancy into a feature filter sampler module, aiming to preserve perceptually important yet sparse sets of feature vectors. Since explicitly learning a binary sampler is infeasible, we develop a confidence filter strategy. Initially, a detection confidence map is generated for the feature map. It reflects the perceptual importance of different spatial areas, with higher levels indicating potential object areas and lower levels typically denoting redundant background areas. For the feature map $\\pmb{F}_i$ of the $i$ -th vehicle, its confidence map $C_i$ is defined as:", + "bbox": [ + 497, + 403, + 892, + 583 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {C} _ {i} = \\Phi_ {\\text {c o n - g e n}} \\left(\\boldsymbol {F} _ {i}\\right) \\in [ 0, 1 ] ^ {H \\times W}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 581, + 591, + 890, + 609 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\Phi_{\\mathrm{con\\_gen}(\\cdot)}$ represents the confidence generation network with detection decoder structure. Then the confidence map is thresholded, followed by non-maximum suppression, resulting in a binary mask $B$ . Utilizing this binary mask, we proceed to preserve the sparse foreground features $\\widetilde{\\pmb{F}}_i = B\\odot \\pmb{F}_i$ . In order to cope with the dynamically changing environmental conditions and fully ensure the robustness of the system, the filtering threshold can be varied with the sensor data and the network state. We set $\\alpha$ as the filter rate and the number of remaining feature vectors is $L = \\alpha \\times HW$ .", + "bbox": [ + 497, + 617, + 890, + 767 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Merge Sampler. Upon extracting a detailed foreground feature vector set with the filter sampler, we employ the merge sampler for additional optimization, refining similar or repetitive foreground feature vectors through weighted merging. The process is divided into three stages: information-driven feature grouping, attention-inspired feature merging, and index-based feature reconstruction.", + "bbox": [ + 497, + 769, + 890, + 866 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "12577", + "bbox": [ + 465, + 887, + 505, + 897 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "(a) Information-Driven Feature Grouping. Initially, a variant of the nearest neighbor clustering algorithm is applied to group the foreground feature vector set. Given a set of feature vector $\\widetilde{\\boldsymbol{F}}_i = [x_1,x_2,\\dots ,x_L]^\\top$ and cluster center $X_{c}$ , we compute the indicator $\\delta_{i}$ for each feature vector. $\\delta_{i}$ is calculated as the minimum feature distance minus the average pixel distance to any other cluster center vector, expressed as:", + "bbox": [ + 75, + 125, + 472, + 238 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\delta_ {i} = \\min _ {j: x _ {j} \\in X _ {c}} \\left(\\| x _ {i} - x _ {j} \\| _ {2} ^ {2} - \\gamma \\| p (x _ {i}), p (x _ {j}) \\| _ {2} ^ {2}\\right), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 99, + 248, + 468, + 273 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\delta_{i}$ denotes to which cluster the feature vector $x_{i}$ should belong. $p(\\cdot)$ means getting the position of the vector and $\\gamma$ is a hyperparameter. Subsequently, we can divide all the feature vectors in $\\widetilde{\\boldsymbol{F}}_i$ into $K$ clusters represented by $G = \\{G_{1},G_{2},\\dots ,G_{K}\\}$ . The total number of clusters is determined by dynamic clustering ratio $\\beta$ , and is calculated as $K = \\beta \\times L$ .", + "bbox": [ + 75, + 283, + 468, + 380 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "(b) Attention-Inspired Feature Merging. A straightforward strategy for merging feature vectors is to average of each feature vector within a cluster. However, this scheme can be severely affected by outlier feature vectors. Drawing inspiration from attention mechanisms, we utilize the confidence score as a guide to quantify the significance of each feature. Thus, the merged feature vector $\\tilde{x}_i$ for the $i$ th cluster $G_i$ is computed as:", + "bbox": [ + 75, + 381, + 468, + 491 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\widetilde {x} _ {i} = \\frac {\\sum_ {j \\in G _ {i}} c _ {j} x _ {j}}{\\sum_ {j \\in G _ {i}} c _ {j}}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 205, + 501, + 468, + 536 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $c_{j}$ and $x_{j}$ represent the confidence score and the original feature vector, respectively. In the end, we obtain a final set of feature vectors $\\mathbf{Z}_i = \\{\\widetilde{x}_1,\\widetilde{x}_2,\\dots ,\\widetilde{x}_K\\}$ that is necessary for transmission.", + "bbox": [ + 75, + 545, + 468, + 600 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "(c) Index-Based Feature Reconstruction. During the feature grouping and merging process, each feature vector is allocated to a cluster, and every cluster is represented by a merged vector. We maintain a record of the index correspondence between original and merged feature vectors. Utilizing this index record, the ego-vehicle ensures that the merged feature vectors are mapped to their corresponding positions, leading to the reconstruction of feature maps.", + "bbox": [ + 75, + 600, + 468, + 712 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Feature Spatial Calibration", + "text_level": 1, + "bbox": [ + 77, + 720, + 324, + 735 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Localization errors [30, 31] may lead to the misalignment of feature maps among vehicles. Such misalignment causes the ego-vehicle to misinterpret the object's location, resulting in sub-optimal perception, as depicted in the upper right corner of Figure 2. To address this challenge, we introduce the Feature Spatial Calibration module (FSC) to facilitate precise feature alignment, as shown in Figure 2. The core idea of FSC is that overlapping viewpoints allow multiple vehicles to detect partially the same objects. The process", + "bbox": [ + 75, + 741, + 468, + 866 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/7d58f4ddb6a0733b531a8909f7caf793b1c101be3a55015892f12e2ce2992a07.jpg", + "image_caption": [ + "Figure 2. Illustration of the proposed FSC. By employing FSC, misaligned matching regions are corrected at the feature level, leading to more accurate predictions." + ], + "image_footnote": [], + "bbox": [ + 502, + 125, + 890, + 252 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "involves three phases: consensus matching, geometric verification, and error modulation.", + "bbox": [ + 497, + 312, + 890, + 339 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Consensus Matching. After reconstructing the received features, the ego vehicle generates sparse feature maps. Within these maps, areas rich in information signify potential target regions, which correspond to the proposed matching areas. We denote the proposed matching regions of the ego vehicle as $P$ and those identified by the collaborative vehicle under noisy pose conditions as $Q$ . Utilizing $P$ and $Q$ , a weighted bipartite graph is constructed, wherein the weight of each edge is determined by the distance between the nodes, encapsulated in a cost matrix. The matching process is then converted into a linear assignment task, with the objective of identifying a matching result with the lowest cumulative edge weight. This procedure yields the matching pairs $M$ .", + "bbox": [ + 497, + 340, + 890, + 533 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Geometric Verification. Invalid matches may occur due to objects located in exclusive zones and detection noise. To tackle this issue, we utilize RANSAC to filter and sift a consistent set of matches aligned with expected geometric transformations. Initially, a random matching subset $M_{s} \\subseteq M$ is selected, and then the transformation matrix $\\Gamma_{s}$ is computed using singular value decomposition. When $\\Gamma_{s}$ is applied to all pairs within $M_{s}$ , and if the posttransformation distance between these pairs remains below the threshold $\\eta$ , the set is considered correctly aligned. The threshold $\\eta$ reflects the allowable localization error within the original collaborative framework. This process is iteratively conducted to identify the optimal transformation matrix that correlates with the maximum number of correct matches. Ultimately, a optimal refined transformation matrix $\\Gamma_{r}$ is obtained and applied in subsequent spatial calibration operations, yielding the aligned feature $\\widetilde{Z}_{j} = \\Gamma_{r}Z_{j}$ .", + "bbox": [ + 497, + 534, + 890, + 769 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Error Modulation. To enhance the adaptability of the calibration method in various environments, we incorporate an error modulation strategy. This strategy aims to achieve a balance between localization errors and the estimation errors that emerge from the calibration process. It measures the overlap ratios between ego and collaborative features in both their adjusted and original states. Subsequently, fea", + "bbox": [ + 497, + 770, + 890, + 866 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "12578", + "bbox": [ + 465, + 887, + 504, + 897 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/d42d10cfc3c0721cb0a1f26165bbe6d235184a7bc9b1bd0e8df2217e6ae379c9.jpg", + "image_caption": [ + "Figure 3. A example of local and global Attention in AFF. This shows how AFF processes both 3D local windows and the global query token for position-level and context-aware aggregation." + ], + "image_footnote": [], + "bbox": [ + 82, + 122, + 463, + 252 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "tures demonstrating the most optimal overlap are selected as the final input.", + "bbox": [ + 75, + 318, + 468, + 347 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Attention-based Feature Fusion", + "text_level": 1, + "bbox": [ + 75, + 353, + 356, + 368 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In multi-vehicle collaboration scenarios, vehicles are able to capture heterogeneous information from different spatial regions. In order to efficiently fuse perceptual features from multiple vehicles, we propose an Attention-based Feature Fusion method (AFF). As shown in Figure 3, AFF utilizes alternating local and global attention to achieve position-level accurate matching in occlusion-variable traffic scenes and to capture the global semantics attention of road topology and traffic states. Specifically, we stack all the vehicles' features to $S \\in \\mathbb{R}^{N \\times H \\times W \\times C}$ , where $N$ is the number of vehicles. The features are divided into 3D non-overlapping windows, each of size $N \\times P \\times P$ . The shape of the partition tensor is $\\left(\\frac{H}{P} \\times \\frac{W}{P}, N \\times P^2, C\\right)$ . We perform 3D local self attention on the local window and implement region interaction within the window. In contrast, global attention goes beyond the limitations of the local view. It uses the extracted global query token and is shared between all windows to interact with the local key and value representations, thus helping to capture long-range dependencies of the features. Inspired by the efficiency of [13, 33], the global query token $g$ is generated as follows:", + "bbox": [ + 75, + 375, + 470, + 667 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {g} = \\mathcal {P} \\left(\\mathcal {C} _ {1 \\times 1} \\left(\\mathcal {S E} (\\mathcal {G} (\\mathcal {D} (\\mathbf {x}))) + \\mathbf {x}\\right)\\right), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 149, + 675, + 468, + 692 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathcal{D}$ , $\\mathcal{G}$ , $\\mathcal{SE}$ , $\\mathcal{C}_{1\\times 1}$ and $\\mathcal{P}$ denote the depth-wise separable convolution, GELU activation function, Squeeze-and-Excitation operation, $1\\times 1$ convolution and max pooling, respectively. These designs enable the global query token to provide advantages such as inductive bias and modeling of inter-channel dependencies. We then construct our proposed AFF module by combining this local and global attention with typical designs of Transformers, including LayerNorm [1], MLPs [9], and skip-connections [14]. This module allows the system to analyze spatial correlations from a local view while also capturing global feature responses, ensuring efficient and precise perception in dy", + "bbox": [ + 75, + 700, + 468, + 866 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/1f09b47ad9f51cbbee866c451bf65bbf30029ce7fdb5145e4bb85834b8ec315b.jpg", + "image_caption": [ + "Figure 4. The architecture of the proposed AFI component." + ], + "image_footnote": [], + "bbox": [ + 509, + 124, + 884, + 254 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "namic, complex, and occlusion-variable traffic scenarios. Finally, we obtain the fused features $\\pmb{H}_i$ .", + "bbox": [ + 497, + 306, + 889, + 333 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.5. Accuracy Enhanced Feature Interaction", + "text_level": 1, + "bbox": [ + 499, + 350, + 843, + 365 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Previous works [15, 21, 38, 42, 43] have demonstrated that fused features can provide richer semantic information, thereby enhancing perceptual performance. However, they may be affected by collaborative noise, such as asynchronous motion blur and inaccurate projections, which can compromise accurate position information and become a bottleneck to the optimal realization of perceptual performance. Ego-centric features may contain locally critical spatial location information without being affected by collaborative noise. To this end, we propose a novel Accuracy Enhanced Feature Interaction (AEI) strategy that leverages the accurate positional information inherent in ego-centric features to enhance the rich semantic information provided by collaborative fused features. Firstly, we design a context cross attention module that is tailored for feature interaction, as shown in Figure 4. It considers the features from the ego-vehicle as queries, and obtains the context of the query location from the fused features as the key and value to achieve cross-attention. The computed output of the spatial location $(i,j)$ is as follows:", + "bbox": [ + 497, + 372, + 890, + 650 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\ny _ {i j} = \\sum_ {a, b \\in \\mathcal {N} _ {k} (i, j)} \\operatorname {s o f t m a x} \\left(q _ {i j} ^ {\\top} k _ {a b}\\right) v _ {a b}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 563, + 665, + 890, + 698 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathcal{N}_k(i,j)$ is context position of $(i,j)$ and $q_{ij}, k_{ab}, v_{ab}$ represent the ego query, context key and value, respectively. As the relevant features of the same object are often located in similar locations across different feature maps, we prioritize the local context details of the query location. Additionally, considering that only a few positions within the sparse feature map hold significant information, focusing on these context-rich areas is computationally efficient. Following this, we employ the split attention [48] to adaptively fuse information from multiple branches, generating the enhanced output feature $\\widetilde{H}_i$ .", + "bbox": [ + 497, + 712, + 892, + 866 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "12579", + "bbox": [ + 465, + 887, + 505, + 897 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/1de46cb1d7557a613601596d0c8251c9c9eee305c37b4dae042ecda1c4e0817a.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
DatasetV2V4RealOPV2V
Noise Level σt/σr (m/°)0.0/0.00.4/0.40.8/0.80.0/0.00.4/0.40.8/0.8
No Collaboration41.64/23.4341.64/23.4341.64/23.4373.25/58.2273.25/58.2273.25/58.22
Late Fusion56.80/27.6240.81/13.3126.65/10.6686.46/79.2176.56/40.0145.57/21.00
F-Cooper [3]61.12/32.3651.71/24.4841.90/19.9187.39/79.3979.33/40.1252.91/16.90
AttFuse [7]64.41/34.3258.42/30.0249.37/25.4490.50/81.8083.54/52.4368.37/40.96
Where2comm [15]63.69/34.7958.48/28.0949.77/24.2590.46/84.2282.17/56.4773.95/45.86
DiscoNet [21]64.19/35.2557.71/28.4149.19/24.9489.58/81.4583.25/53.2261.37/30.53
V2VNet [38]65.70/35.3862.11/30.6354.73/25.5791.35/82.4385.43/54.1670.43/31.28
CoAlign [25]64.06/36.6059.58/31.5253.02/26.6090.93/84.2884.48/57.4573.95/48.86
V2X-ViT [42]66.42/37.3463.33/32.3857.15/28.2291.74/83.3182.91/54.7361.70/26.07
CoBEVT [41]66.01/37.3658.63/29.0949.22/23.6691.71/85.9885.49/60.6464.63/29.59
ERMVP (Ours)67.66/42.9765.01/40.4560.88/35.0192.18/85.5985.67/66.3277.13/59.15
", + "bbox": [ + 78, + 122, + 892, + 309 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1. Overall performance on V2V4Real and OPV2V datasets with pose noises. The results are reported in AP@0.5/0.7.", + "bbox": [ + 117, + 319, + 848, + 332 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.6. Decoder and Loss", + "text_level": 1, + "bbox": [ + 77, + 339, + 250, + 353 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Based on the final fused feature $\\widetilde{H}_i$ , we use the detection decoder to generate the final prediction output $Y_{i} = \\Phi_{\\mathrm{dec}}(\\widetilde{H}_{i})$ . Each position of $Y_{i}$ represents a rotated box with classes (c, x, y, h, w, $\\cos \\alpha$ , $\\sin \\alpha$ ), corresponding to class confidence, position, size, and angle. These objects are the final output of the proposed collaborative perception system. Following existing work [19], we adopt the smooth $L_{1}$ loss for regression and focal loss [23] for classification.", + "bbox": [ + 75, + 362, + 468, + 476 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 77, + 496, + 208, + 511 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Datasets and Evaluation Metrics", + "text_level": 1, + "bbox": [ + 77, + 521, + 363, + 535 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Datasets. We validate the proposed ERMVP in the task of LiDAR-based 3D object detection on two benchmark datasets including V2V4real [44] and OPV2V [43]. V2V4real [44] is the first large-scale real-world dataset on V2V perception, collected by two vehicles equipped with multi-modal sensors driving together in a variety of scenarios. It covers a driving area of 410 kilometers, including 20K frames of point clouds, with training/validation/test sets containing 14210/2000/3986 frames, respectively. OPV2V [43] is a large-scale vehicle-to-vehicle collaborative perception dataset, simulated by OpenCDA [40] and Carla [8]. It contains 73 different scenarios, in which different numbers (2 to 7) of collaborative vehicles appear together, each equipped with a lidar sensor and 4 cameras. The dataset contains a total of 11,464 frames of point clouds and RGB images. The training/validation/test sets contain 6374, 1980, and 2170 frames, respectively.", + "bbox": [ + 75, + 545, + 468, + 793 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation Metrics. We adopt the Average Precision (AP) at Intersection-over-Union (IoU) thresholds of 0.5 and 0.7 to evaluate the detection performance. The calculation format of communication volume in [15] is used to count the message size by byte in the log scale with base 2.", + "bbox": [ + 75, + 796, + 468, + 866 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Implementation Details", + "text_level": 1, + "bbox": [ + 499, + 339, + 717, + 353 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We implement the proposed ERMVP and comparison model on the Pytorch toolbox [28], and train them on a NVIDIA GeForceRTX 3090 GPU using the Adam optimizer [18]. The initial learning rate is $2\\mathrm{e} - 3$ and it decays every 15 epochs with a factor of 0.1. All models are trained with 60 epochs and the batch size is set to 2. Early stopping is used to find the best epoch. We also add normal point cloud data augmentation for all experiments, including scaling, rotation, and flipping. All detection models are based on PointPillar [19] backbone to extract 2D features from point clouds and a $0.4\\mathrm{m}$ width/length is used for each voxel. AFF component has 3 encoded layers and a window size of 8 for both local and global attention. The balance hyperparameter $\\gamma$ is 0.1 and the error tolerance threshold $\\eta$ is 0.25. Each vehicle has a communication range of 70 m based on [42], while vehicles outside of this broadcasting radius will be ignored. To simulate the localization and heading errors, we add Gaussian noise with a standard deviation of $\\sigma_t$ for localization errors and $\\sigma_r$ for heading errors.", + "bbox": [ + 497, + 362, + 890, + 625 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3.Quantitative Evaluation", + "text_level": 1, + "bbox": [ + 499, + 636, + 721, + 650 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Comparison of Detection Performance. Table 1 shows the performance comparison results of 3D detection on two datasets. We use the No Collaboration method as a baseline, which only uses ego vehicle's point clouds data without collaboration. Late Fusion allows vehicles to exchange detected outputs and utilizes non-maximum suppression to produce the final result. For intermediate fusion strategies, we evaluate the existing state-of-the-art (SOTA) methods: AttFuse [43], F-Cooper [3], V2VNet [38], DiscoNet [21], V2X-ViT [42], CoAlign (only fusion) [25], CoBEVT [41] and Where2comm [15]. The proposed ER-MVP achieves average improvements of $12.48\\% / 18.47\\%$ and $6.18\\% / 10.02\\%$ on two datasets in AP@0.5/0.7 compared with No Collaboration and Late Fusion methods, demonstrating the superiority of our collaborative method.", + "bbox": [ + 497, + 657, + 890, + 866 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "12580", + "bbox": [ + 465, + 887, + 505, + 897 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/a05276fe44b9498839d47f184cc6c82d59fe0ebf90f874dbf5bd3a548c86c415.jpg", + "image_caption": [ + "(a) V2V4Real" + ], + "image_footnote": [ + "Late Fusion FCooper V2VNet V2X-VIT CoBEVT -Where2comm -ERMVP" + ], + "bbox": [ + 80, + 122, + 282, + 251 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/0d727aed98d8c717ad512713bc262dc41c0e12fc4ed41f3bfc943faddb1c8b02.jpg", + "image_caption": [ + "Figure 5. Collaborative perception performance comparison on the V2V4Real and OPV2V with varying communication volume." + ], + "image_footnote": [], + "bbox": [ + 282, + 122, + 482, + 250 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f88bea390d557ccfa8148f6e1357029b30be1ff3a78f8f3b8d9d1900be3dafa5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 482, + 122, + 682, + 250 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/517d86dc9d3835c14d405265d54458bd70a95d2611f426bab979bfa02ba7bde2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 682, + 122, + 880, + 250 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Meanwhile, it is superior to the SOTA collaborative perception method in both real-world and simulated scenarios: improves the SOTA collaborative performance by $1.24\\%$ and $5.61\\%$ on V2V4Real datasets in AP@0.5/0.7, and by $0.44\\%$ on OPV2V datasets in AP@0.5. These results fully demonstrate the superiority of the ERMVP collaboration paradigm.", + "bbox": [ + 75, + 321, + 468, + 419 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/aaa3e111652a9c67b6d7667c976a0c1cf556ed7598929d5b2d4b2facb1d4c5f2.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
DatasetV2V4RealOPV2V
Noise Level σt(m)0.51.00.51.0
Late Fusion11.0210.1353.8838.80
F-Cooper [3]24.6018.4132.2412.71
AttFuse [43]28.1024.3642.9535.53
DiscoNet [21]28.3624.0545.3624.96
V2VNet [38]29.5723.6747.4825.82
Where2comm [15]28.0223.5346.5945.59
CoAlign [21]29.8425.8249.0946.66
V2X-ViT [42]30.9923.6744.8321.77
CoBEVT [41]24.6722.0141.9126.05
ERMVP (Ours)39.2633.9477.2576.09
", + "bbox": [ + 78, + 429, + 455, + 600 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Robustness to Localization and Heading Errors. To evaluate the sensitivity of existing methods to localization and heading errors, we use the same noise setting as [42] and conduct extensive experiments on two datasets. As shown in Tables 1 and 2, under ideal settings, Late Fusion and some advanced intermediate fusion methods can detect object vehicles with high accuracy, but their detection accuracy rapidly decreases as the standard deviation of errors increases. For example, when the localization and heading errors of OPV2V dataset are over $0.4\\mathrm{m}$ and $0.4^{\\circ}$ , the Disconet [21] and F-Cooper [3] methods fail and perform even worse than No Collaboration. Our method exceeds previous SOTA models at all noise levels and consistently outperforms the No Collaboration baseline, clearly demonstrating the robustness of ERMVP to pose errors. It is noteworthy that for localization errors, ERMVP can maintain consistent", + "bbox": [ + 75, + 644, + 470, + 865 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "accuracy levels in high noise environments, showing superior robustness. When the localization error is $1.0\\mathrm{m}$ , it outperforms the second-best method by $29.43\\%$ and $8.08\\%$ in AP@0.7 on OPV2V and V2V4Real datasets. The reasonable explanations are: (i) feature spatial calibration aligns mismatched collaborative features; (ii) accuracy enhanced feature interaction reduces performance degradation caused by collaborative noise.", + "bbox": [ + 497, + 321, + 890, + 431 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Comparison of Communication Volume. To achieve pragmatic collaborative perception, it is crucial to evaluate the perceptual performance under different communication volumes. Figure 5 shows the results of the performance comparison under different bandwidth consumption conditions. Note that we do not consider any additional model/data/feature compression for a fair comparison. The results indicate that ERMVP achieves excellent perceptual performance and bandwidth consumption trade-off under all communication bandwidth conditions, consistently outperforming Where2comm [15]. At the same time, it achieves the same detection performance as the previous model [3, 38, 41, 42] with less communication volume on both real-world and simulated datasets.", + "bbox": [ + 497, + 432, + 892, + 625 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Ablation Studies", + "text_level": 1, + "bbox": [ + 499, + 631, + 663, + 644 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/69bd4eb33a169fcd8834f564b380acc1a803aa83bd6f57a98cc16dd8763cdf52.jpg", + "table_caption": [ + "Table 2. Detection AP@0.7 on V2V4Real and OPV2V datasets with localization error." + ], + "table_footnote": [], + "table_body": "
AFFFSCAEIV2V4RealOPV2V
51.71/24.4879.33/40.12
✓62.24/37.7084.30/61.43
✓✓64.36/39.5984.98/62.67
✓✓✓65.01/40.4585.67/66.32
", + "bbox": [ + 500, + 653, + 885, + 726 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 3. Ablation study results of the proposed core components on the both datasets with noise level of 0.4/0.4. AFF: Attention-based Feature Fusion; FSC: Feature Spatial Calibration; AEI: Accuracy Enhanced Feature Interaction.", + "bbox": [ + 497, + 736, + 890, + 788 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Effect of Core Components. Table 3 details the contribution of each core components in our ERMVP framework. The base model is the naive position-wise maximum fusion method. We then assess the impact of each component by sequentially introducing i) attention-based feature", + "bbox": [ + 497, + 796, + 890, + 866 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "12581", + "bbox": [ + 465, + 887, + 504, + 897 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/280d2b5c3bb6552ed83f6681ca5e9ab65f17fefd5ae8bbd0b8c2945e3095852f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 87, + 124, + 284, + 210 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/e770195a486e1bf921acc083d6a82eecd9c532e4557f8586c9120e99f4b1c3dd.jpg", + "image_caption": [ + "(a) V2VNet" + ], + "image_footnote": [], + "bbox": [ + 87, + 212, + 282, + 296 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/efd2f53dc9a564fb9dde0a5412ab504de41b34a537d009f58505a6de2f4afd40.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 285, + 124, + 482, + 210 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/492842c4ba1259fbc06256306e088f9a970ff6743f2cbd8d52a3c6f44fd3ba2c.jpg", + "image_caption": [ + "(b) CoBEVT" + ], + "image_footnote": [], + "bbox": [ + 285, + 212, + 482, + 295 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/aefb2e149c7c44ff1a1486876c71877e2004bf57de183a250fdfb7e67f496884.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 484, + 124, + 680, + 210 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/f4481015af92b97fb51c9f6721b939b1b571c6f3d179fb796d81053fc607f9bc.jpg", + "image_caption": [ + "(c) V2X-ViT", + "Figure 6. Qualitative comparison results in real-world scenarios from the V2V4Real dataset with noise level of 0.4/0.4. Green and red boxes denote the ground truths and detection results, respectively." + ], + "image_footnote": [], + "bbox": [ + 485, + 212, + 680, + 295 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/bba47526b06f4a957dc1c460977993b17132c871bd63b9b4bf3549c7e3f00d7f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 682, + 124, + 882, + 210 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/cdb2cc3bba2d1bd3841d0f2755d0b7379f1b9c64db023463ded86abc01c94515.jpg", + "image_caption": [ + "(d)ERMVP" + ], + "image_footnote": [], + "bbox": [ + 684, + 212, + 880, + 295 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "fusion (AFF), ii) feature spatial calibration (FSC), iii) accuracy enhanced feature interaction (AEF). The consistent rise in detection results over both datasets demonstrates the effectiveness of each introduced component. Notably, integrating all three components boosts detection performance by $13.3\\%$ and $15.97\\%$ on the V2V4Real dataset for AP@0.5 and AP@0.7, respectively.", + "bbox": [ + 75, + 355, + 468, + 454 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/b92763145d0f8cd9c12d9dd5457a2982c98ae21f38860f7e31289167e94214e1.jpg", + "image_caption": [ + "Figure 7. Hyperparameter analysis of $\\alpha$ and $\\beta$" + ], + "image_footnote": [], + "bbox": [ + 89, + 463, + 267, + 572 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/d71922f3b4b2034784cf84180fa4aae8983c8e9ae7ee30960da71fb8f0ab96ef.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 280, + 463, + 458, + 572 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Hyperparameters $\\alpha$ and $\\beta$ . We explore the impact of the hyperparameters $\\alpha$ and $\\beta$ (as described in Section 3.2), which jointly determine the communication volume. Specifically, the feature compression ratio is represented as $\\alpha \\times \\beta$ . The communication volume, in turn, is expressed as the product of the original feature size and the compression rate. By modulating the values of $\\alpha$ and $\\beta$ within the interval of $10\\%$ to $80\\%$ , we chart the resultant object detection performance (measured by AP@0.7) in Figure 7. As expected, a drop in either $\\alpha$ or $\\beta$ will impair ERMVP's efficacy. Yet, it's worth noting that ERMVP's performance generally exhibits robustness to these hyperparameters. The performance decline remains fairly minimal even under significant compression, such as a factor of 100. The results demonstrate the effectiveness of our proposed filter and merge feature sampling module. Therefore, ERMVP establishes an appropriate and effective communication paradigm among vehicles, ensuring an optimal tradeoff between perception and communication.", + "bbox": [ + 75, + 603, + 472, + 866 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.5. Qualitative Evaluation", + "text_level": 1, + "bbox": [ + 499, + 355, + 709, + 369 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Figure 6 displays the detection results of V2VNet [38], CoBEVT [41], V2X-ViT [42], and ERMVP in two scenarios. Clearly, ERMVP offers more precise and comprehensive detection than the prior SOTA methods. To begin with, ERMVP produces a greater number of predicted bounding boxes that are well aligned with the ground truths, while other methods show significant discrepancies. This demonstrates ERMVP's robustness, particularly in high-noise environments. Moreover, ERMVP detects more dynamic objects (more ground truth bounding boxes find matches). This suggests that ERMVP can effectively combine the inputs from nearby vehicles, leading to a thorough scene representation. Overall, these qualitative assessments confirm ERMVP's strengths in delivering accurate and comprehensive perception, especially in challenging conditions.", + "bbox": [ + 497, + 376, + 890, + 583 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 499, + 591, + 616, + 605 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we present ERMVP, a communication-efficient and collaboration-robust multi-vehicle perception method in challenging environments. We introduce a filter and merge feature sampling strategy to efficient communication, a feature spatial calibration module to align features in fine-grid and two spatial information aggregation components to solve the bottleneck of the fusion method. Extensive experiments prove the superiority of ERMVP and the effectiveness of our components. The feasibility of our approach for other modalities will be explored in future work.", + "bbox": [ + 497, + 614, + 890, + 752 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments", + "text_level": 1, + "bbox": [ + 499, + 760, + 658, + 776 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This work is supported by the Shanghai Key Research Laboratory of NSAI, the Innovation Platform for Academicians of Hainan Province, Haikou, 570228, the Specific Research Fund of the Innovation Platform for Academicians of Hainan Province under Grant YSPTZX202314, and NFSC under grant 62250410368.", + "bbox": [ + 497, + 783, + 890, + 866 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "12582", + "bbox": [ + 465, + 887, + 504, + 897 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 124, + 173, + 138 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Jimmy Lei Ba, Jamie Ryan Kiros, and Geoffrey E Hinton. Layer normalization. arXiv preprint arXiv:1607.06450, 2016. 5", + "[2] Marwa Chafii, Salmane Naoumi, Reda Alami, Ebtesam Al-mazrouei, Mehdi Bennis, and Merouane Debbah. Emergent communication in multi-agent reinforcement learning for future wireless networks. arXiv preprint arXiv:2309.06021, 2023. 3", + "[3] Qi Chen, Xu Ma, Sihai Tang, Jingda Guo, Qing Yang, and Song Fu. F-cooper: Feature based cooperative perception for autonomous vehicle edge computing system using 3d point clouds. In Proceedings of the 4th ACM/IEEE Symposium on Edge Computing, pages 88-100, 2019. 1, 6, 7", + "[4] Zhaoyu Chen, Bo Li, Shuang Wu, Jianghe Xu, Shouhong Ding, and Wenqiang Zhang. Shape matters: deformable patch attack. In European conference on computer vision, pages 529-548. Springer, 2022. 1", + "[5] Zhaoyu Chen, Bo Li, Shuang Wu, Shouhong Ding, and Wenqiang Zhang. Query-efficient decision-based black-box patch attack. IEEE Transactions on Information Forensics and Security, 2023.", + "[6] Zhaoyu Chen, Bo Li, Shuang Wu, Kaixun Jiang, Shouhong Ding, and Wenqiang Zhang. Content-based unrestricted adversarial attack. Advances in Neural Information Processing Systems, 36, 2024. 1", + "[7] Yimian Dai, Fabian Gieseke, Stefan Oehmcke, Yiquan Wu, and Kobus Barnard. Attentional feature fusion. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 3560-3569, 2021. 6", + "[8] Alexey Dosovitskiy, German Ros, Felipe Codevilla, Antonio Lopez, and Vladlen Koltun. Carla: An open urban driving simulator. In Conference on robot learning, pages 1-16. PMLR, 2017. 6", + "[9] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 5", + "[10] Ian Goodfellow, Yoshua Bengio, and Aaron Courville. Deep learning. MIT press, 2016. 1", + "[11] Jiaming Gu, Jingyu Zhang, Muyang Zhang, Weiliang Meng, Shibiao Xu, Jiguang Zhang, and Xiaopeng Zhang. Feaco: Reaching robust feature-level consensus in noisy pose conditions. In Proceedings of the 31st ACM International Conference on Multimedia, page 3628-3636, New York, NY, USA, 2023. Association for Computing Machinery. 1", + "[12] Monowar Hasan, Sibin Mohan, Takayuki Shimizu, and Hongsheng Lu. Securing vehicle-to-everything (v2x) communication platforms. IEEE Transactions on Intelligent Vehicles, 5(4):693-713, 2020. 1", + "[13] Ali Hatamizadeh, Hongxu Yin, Greg Heinrich, Jan Kautz, and Pavlo Molchanov. Global context vision transformers. In International Conference on Machine Learning, pages 12633-12646. PMLR, 2023. 5" + ], + "bbox": [ + 78, + 147, + 468, + 864 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[14] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 5", + "[15] Yue Hu, Shaoheng Fang and Zixing Lei, Yiqi Zhong, and Siheng Chen. Where2comm: Communication-efficient collaborative perception via spatial confidence maps. In Thirty-sixth Conference on Neural Information Processing Systems (NeurIPS), 2022. 1, 3, 5, 6, 7", + "[16] Kaixun Jiang, Zhaoyu Chen, Hao Huang, Jiafeng Wang, Dingkang Yang, Bo Li, Yan Wang, and Wenqiang Zhang. Efficient decision-based black-box patch attacks on video recognition. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 4379-4389, 2023. 1", + "[17] Hamza Khan, Sumudu Samarakoon, and Mehdi Bennis. Enhancing video streaming in vehicular networks via resource slicing. IEEE Transactions on Vehicular Technology, 69(4): 3513-3522, 2020. 1", + "[18] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In International Conference on Learning Representations (ICLR), 2015. 6", + "[19] Alex H Lang, Sourabh Vora, Holger Caesar, Lubing Zhou, Jiong Yang, and Oscar Beijbom. Pointpillars: Fast encoders for object detection from point clouds. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12697-12705, 2019. 3, 6", + "[20] Yiming Li, Bir Bhanu, and Wei Lin. Auction protocol for camera active control. In 2010 IEEE International Conference on Image Processing, pages 4325-4328. IEEE, 2010. 2", + "[21] Yiming Li, Shunli Ren, Pengxiang Wu, Siheng Chen, Chen Feng, and Wenjun Zhang. Learning distilled collaboration graph for multi-agent perception. Advances in Neural Information Processing Systems, 34:29541-29552, 2021. 5, 6, 7", + "[22] Yiming Li, Ziyan An, Zixun Wang, Yiqi Zhong, Siheng Chen, and Chen Feng. V2x-sim: A virtual collaborative perception dataset for autonomous driving. arXiv preprint arXiv:2202.08449, 2022. 3", + "[23] Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dollar. Focal loss for dense object detection. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), 2017. 6", + "[24] Yang Liu, Jing Liu, Kun Yang, Bobo Ju, Siao Liu, Yuzheng Wang, Dingkang Yang, Peng Sun, and Liang Song. Ampnet: Appearance-motion prototype network assisted automatic video anomaly detection system. IEEE Transactions on Industrial Informatics, 20(2):2843-2855, 2024. 1", + "[25] Yifan Lu, Quanhao Li, Baoan Liu, Mehrdad Dianati, Chen Feng, Siheng Chen, and Yanfeng Wang. Robust collaborative 3d object detection in presence of pose errors. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 4812-4818. IEEE, 2023. 1, 6", + "[26] Zonglin Meng, Xin Xia, Runsheng Xu, Wei Liu, and Jiaqi Ma. Hydro-3d: Hybrid object detection and tracking for cooperative perception using 3d lidar. IEEE Transactions on Intelligent Vehicles, 2023. 1" + ], + "bbox": [ + 502, + 125, + 890, + 865 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "12583", + "bbox": [ + 465, + 887, + 504, + 897 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[27] Yaru Niu, Rohan R Paleja, and Matthew C Gombolay. Multiagent graph-attention communication and teaming. In AA-MAS, pages 964–973, 2021. 2", + "[28] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. Advances in neural information processing systems, 32, 2019. 6", + "[29] Faisal Qureshi and Demetri Terzopoulos. Smart camera networks in virtual reality. Proceedings of the IEEE, 96(10): 1640-1656, 2008. 2", + "[30] Qingyu Song, Changan Wang, Zhengkai Jiang, Yabiao Wang, Ying Tai, Chengjie Wang, Jilin Li, Feiyue Huang, and Yang Wu. Rethinking counting and localization in crowds: A purely point-based framework. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3365-3374, 2021. 4", + "[31] Zhiying Song, Fuxi Wen, Hailiang Zhang, and Jun Li. An efficient and robust object-level cooperative perception framework for connected and automated driving. arXiv preprint arXiv:2210.06289, 2022. 4", + "[32] Ming Tan. Multi-agent reinforcement learning: Independent vs. cooperative agents. In Proceedings of the tenth international conference on machine learning, pages 330-337, 1993. 2", + "[33] Mingxing Tan and Quoc Le. Efficientnetv2: Smaller models and faster training. In International conference on machine learning, pages 10096-10106. PMLR, 2021. 5", + "[34] Nicholas Vadivelu, Mengye Ren, James Tu, Jingkang Wang, and Raquel Urtasun. Learning to communicate and correct pose errors. In Conference on Robot Learning, pages 1195-1210. PMLR, 2021. 1", + "[35] Binglu Wang, Lei Zhang, Zhaozhong Wang, Yongqiang Zhao, and Tianfei Zhou. Core: Cooperative reconstruction for multi-agent perception. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8710-8720, 2023. 1, 3", + "[36] Jian Wang, Yameng Shao, Yuming Ge, and Rundong Yu. A survey of vehicle to everything (v2x) testing. Sensors, 19(2): 334, 2019. 1", + "[37] Tianhang Wang, Guang Chen, Kai Chen, Zhengfa Liu, Bo Zhang, Alois Knoll, and Changjun Jiang. Umc: A unified bandwidth-efficient and multi-resolution based collaborative perception framework. arXiv preprint arXiv:2303.12400, 2023. 3", + "[38] Tsun-Hsuan Wang, Sivabanan Manivasagam, Ming Liang, Bin Yang, Wenyuan Zeng, and Raquel Urtasun. V2vnet: Vehicle-to-vehicle communication for joint perception and prediction. In European Conference on Computer Vision, pages 605–621. Springer, 2020. 1, 3, 5, 6, 7, 8", + "[39] Qianyi Wu, Kaisiyuan Wang, Kejie Li, Jianmin Zheng, and Jianfei Cai. Objectsdf++: Improved object-compositional neural implicit surfaces. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 21764-21774, 2023. 1" + ], + "bbox": [ + 78, + 125, + 468, + 865 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[40] Runsheng Xu, Yi Guo, Xu Han, Xin Xia, Hao Xiang, and Jiaqi Ma. Opencda: an open cooperative driving automation framework integrated with co-simulation. In 2021 IEEE International Intelligent Transportation Systems Conference (ITSC), pages 1155-1162. IEEE, 2021. 6", + "[41] Runsheng Xu, Zhengzhong Tu, Hao Xiang, Wei Shao, Bolei Zhou, and Jiaqi Ma. Cobevt: Cooperative bird's eye view semantic segmentation with sparse transformers, 2022. 1, 3, 6, 7, 8", + "[42] Runsheng Xu, Hao Xiang, Zhengzhong Tu, Xin Xia, Ming-Hsuan Yang, and Jiaqi Ma. V2x-vit: Vehicle-to-everything cooperative perception with vision transformer. In European Conference on Computer Vision, page 107-124. Springer, 2022. 1, 3, 5, 6, 7, 8", + "[43] Runsheng Xu, Hao Xiang, Xin Xia, Xu Han, Jinlong Li, and Jiaqi Ma. Opv2v: An open benchmark dataset and fusion pipeline for perception with vehicle-to-vehicle communication. In 2022 International Conference on Robotics and Automation (ICRA), pages 2583–2589. IEEE, 2022. 2, 3, 5, 6, 7", + "[44] Runsheng Xu, Xin Xia, Jinlong Li, Hanzhao Li, Shuo Zhang, Zhengzhong Tu, Zonglin Meng, Hao Xiang, Xiaoyu Dong, Rui Song, et al. V2v4real: A real-world large-scale dataset for vehicle-to-vehicle cooperative perception. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13712–13722, 2023. 2, 3, 6", + "[45] Kun Yang, Peng Sun, Jieyu Lin, Azzedine Boukerche, and Liang Song. A novel distributed task scheduling framework for supporting vehicular edge intelligence. In 2022 IEEE 42nd International Conference on Distributed Computing Systems (ICDCS), pages 972-982. IEEE, 2022. 1", + "[46] Kun Yang, Dingkang Yang, Jingyu Zhang, Mingcheng Li, Yang Liu, Jing Liu, Hanqi Wang, Peng Sun, and Liang Song. Spatio-temporal domain awareness for multi-agent collaborative perception, 2023. 1", + "[47] Kun Yang, Dingkang Yang, Jingyu Zhang, Hanqi Wang, Peng Sun, and Liang Song. What2comm: Towards communication-efficient collaborative perception via feature decoupling. In Proceedings of the 31st ACM International Conference on Multimedia, page 7686-7695, New York, NY, USA, 2023. Association for Computing Machinery. 1", + "[48] Hang Zhang, Chongruo Wu, Zhongyue Zhang, Yi Zhu, Haibin Lin, Zhi Zhang, Yue Sun, Tong He, Jonas Mueller, R. Manmatha, Mu Li, and Alexander Smola. Resnest: Split-attention networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, pages 2736-2746, 2022. 5", + "[49] Sai Qian Zhang, Qi Zhang, and Jieyu Lin. Succinct and robust multi-agent communication with temporal message control. Advances in Neural Information Processing Systems, 33:17271-17282, 2020. 3" + ], + "bbox": [ + 502, + 125, + 890, + 781 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "12584", + "bbox": [ + 465, + 887, + 504, + 897 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2024/ERMVP_ Communication-Efficient and Collaboration-Robust Multi-Vehicle Perception in Challenging Environments/197e9226-8724-4a9f-ac9a-1556459fa452_model.json b/2024/ERMVP_ Communication-Efficient and Collaboration-Robust Multi-Vehicle Perception in Challenging Environments/197e9226-8724-4a9f-ac9a-1556459fa452_model.json new file mode 100644 index 0000000000000000000000000000000000000000..e5888ce1da2cf16b698198a65c3a3314555b8bcd --- /dev/null +++ b/2024/ERMVP_ Communication-Efficient and Collaboration-Robust Multi-Vehicle Perception in Challenging Environments/197e9226-8724-4a9f-ac9a-1556459fa452_model.json @@ -0,0 +1,2035 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.129, + 0.051, + 0.201, + 0.086 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.253, + 0.049, + 0.793, + 0.088 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.102, + 0.162, + 0.87, + 0.204 + ], + "angle": 0, + "content": "ERMVP: Communication-Efficient and Collaboration-Robust Multi-Vehicle Perception in Challenging Environments" + }, + { + "type": "text", + "bbox": [ + 0.132, + 0.228, + 0.838, + 0.278 + ], + "angle": 0, + "content": "Jingyu Zhang\\(^{1}\\) Kun Yang\\(^{1}\\) Yilei Wang\\(^{1}\\) Hanqi Wang\\(^{1}\\) Peng Sun\\(^{2,*}\\) Liang Song\\(^{1,*}\\) \n\\(^{1}\\)Academy for Engineering and Technology, Fudan University \\(^{2}\\)Duke Kunshan University \n{jingyuzhang22, yileiwang23}@m.fudan.edu.cn, songl@m.fudan.edu.cn" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.309, + 0.314, + 0.325 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.339, + 0.473, + 0.658 + ], + "angle": 0, + "content": "Collaborative perception enhances perception performance by enabling autonomous vehicles to exchange complementary information. Despite its potential to revolutionize the mobile industry, challenges in various environments, such as communication bandwidth limitations, localization errors and information aggregation inefficiencies, hinder its implementation in practical applications. In this work, we propose ERMVP, a communication-Efficient and collaboration-Robust Multi-Vehicle Perception method in challenging environments. Specifically, ERMVP has three distinct strengths: i) It utilizes the hierarchical feature sampling strategy to abstract a representative set of feature vectors, using less communication overhead for efficient communication; ii) It employs the sparse consensus features to execute precise spatial location calibrations, effectively mitigating the implications of vehicle localization errors; iii) A pioneering feature fusion and interaction paradigm is introduced to integrate holistic spatial semantics among different vehicles and data sources. To thoroughly validate our method, we conduct extensive experiments on real-world and simulated datasets. The results demonstrate that the proposed ERMVP is significantly superior to the state-of-the-art collaborative perception methods." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.685, + 0.21, + 0.699 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.709, + 0.47, + 0.835 + ], + "angle": 0, + "content": "Autonomous vehicles are widely recognized as a valuable means to enhance road safety and traffic efficiency. Equipped with lidar, cameras, and other sensors, these vehicles are capable of accurately sensing their surroundings to ensure safe and reliable operation. However, the single-vehicle perception system has inevitable drawbacks [26, 39], such as a limited sensor field of view that can be easily obstructed and the challenge of detecting distant objects due to sparse and low-resolution data. Recently, the" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.31, + 0.892, + 0.395 + ], + "angle": 0, + "content": "advances in vehicle-to-vehicle (V2V) communication technologies [12, 17, 36] and deep learning [4-6, 10, 16, 24, 45] have spurred innovation and progress in the collaborative perception technology. This technology allows connected autonomous vehicles (CAVs) to share sensory data, leading to more comprehensive environmental perception." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.395, + 0.893, + 0.602 + ], + "angle": 0, + "content": "Although collaborative perception technology shows great potential in transforming the mobility industry, its practical application faces several challenges, including communication bandwidth limitations [15, 35], localization errors [11, 25] and information aggregation inefficiencies [38, 47]. In practical situations, wireless communication resource and reliability constraints severely hamper the efficacy of delay-sensitive collaborative perception. While recent works [15, 42] have achieved a balance between perception performance and communication bandwidth through well-designed mechanisms, these methods have their limitations because they primarily considering information compression over spatial redundancy. This narrow focus exacerbates performance degradation at high compression ratios." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.604, + 0.895, + 0.868 + ], + "angle": 0, + "content": "Further, complex dynamic environments lead to localization errors, which result in inaccurate relative transform estimates and spatial feature misalignment. This relative pose noise produces misleading features that adversely affect the effectiveness of collaborative perception. Existing methods [25, 34] attempt to optimize the overall pose through intensive computation, but the high latency makes them unsuitable for real-time dynamic perception. Meanwhile, collaborative methods [3, 38, 41, 42, 46] only focus on aggregated information, but overlook the inherent perceptual strengths of ego vehicle. This paradigm is vulnerable to the perturbations introduced by collaborative noise, including asynchronous motion blur and inaccurate projections. Such drawback becomes a bottleneck for achieving optimal perceptual performance. In contrast, ego-centric features may contain locally accurate spatial location information that is not affected by collaborative noise. Therefore, a priority for establishing a pragmatic collaborative perception system is to effectively overcome the above challenges." + }, + { + "type": "page_footnote", + "bbox": [ + 0.076, + 0.843, + 0.47, + 0.867 + ], + "angle": 0, + "content": "*Corresponding authors. Our code is available at https://github.com/Terry9a/ERMVP." + }, + { + "type": "page_number", + "bbox": [ + 0.466, + 0.888, + 0.506, + 0.899 + ], + "angle": 0, + "content": "12575" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.078, + 0.123, + 0.891, + 0.372 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.382, + 0.893, + 0.421 + ], + "angle": 0, + "content": "Figure 1. The overall architecture of the proposed framework. The framework consists of six phases: metadata sharing and feature extraction, filter and merge feature sampling, feature spatial calibration, attention-based feature fusion, accuracy enhanced feature interaction and detection decoders. The details of each individual component are illustrated in Section 3." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.437, + 0.473, + 0.866 + ], + "angle": 0, + "content": "Based on these observations, we propose ERMVP, a communication-efficient and collaboration-robust multi-vehicle perception method in challenging environments. From Figure 1, ERMVP uses four proposed core components to tackle existing challenges jointly. Specifically, (i) we first design an advanced filter and merge feature sampling strategy to tackle the limitations of wireless communication resources. This strategy considers both inter-class and intra-class redundancy relationships to abstract a refined set of feature vectors from the redundant features, using less communication overhead for efficient communication. (ii) Second, we introduce a plug-and-play feature spatial calibration module to mitigate the implications of vehicle localization errors. This module ingeniously utilizes consensus sparse foreground features to align the relative pose relationships between ego-vehicle and collaborators without any precise pose supervision. (iii) Furthermore, we present a pioneering feature fusion and interaction paradigm to integrate holistic spatial semantics. This paradigm comprises two key components: The first is an attention-based feature fusion module that alternates between local and global attention to fuse heterogeneous information from different vehicles. The second is an accuracy enhanced feature interaction strategy that leverages the accurate positional information inherent in ego-centric features to enhance the rich semantic information provided by fused features. Through these tailored components, ERMVP represents a significant advancement towards pragmatic collaborative perception. To validate the effectiveness of the ER-MVP, we conduct extensive experiments on two collaborative 3D object detection datasets including V2V4Real [44]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.436, + 0.892, + 0.492 + ], + "angle": 0, + "content": "and OPV2V [43]. Comprehensive experimental results demonstrate that our method outperforms previous state-of-the-art methods under the bandwidth-limited noisy setting. The main contributions can be summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.492, + 0.891, + 0.56 + ], + "angle": 0, + "content": "- We present ERMVP, a communication-efficient and collaboration-robust multi-vehicle perception method, which addresses the communication bandwidth limitations, localization errors and information aggregation efficiency challenges." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.561, + 0.892, + 0.629 + ], + "angle": 0, + "content": "- We develop a filter and merge feature sampling strategy to enhance communication efficiency, a feature spatial calibration module for accurate spatial feature alignment, and two information aggregation components to optimize the fusion process." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.63, + 0.892, + 0.684 + ], + "angle": 0, + "content": "- We conduct extensive experiments on both real-world and simulated datasets. The results show the superiority of our method and the necessity of the proposed components." + }, + { + "type": "list", + "bbox": [ + 0.5, + 0.492, + 0.892, + 0.684 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.698, + 0.641, + 0.712 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.721, + 0.759, + 0.735 + ], + "angle": 0, + "content": "2.1. Multi-Agent Communication" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.742, + 0.892, + 0.867 + ], + "angle": 0, + "content": "Communication has played a pivotal role in the development of robust multi-agent systems. Early multi-agent communication [20, 29, 32] relied on predefined protocols and heuristics to regulate communication between agents. However, these fixed methods are unsuitable in complex and dynamic settings. Advances in deep learning have inspired the development of advanced strategies for information exchange and collaboration among agents. For example, MAGIC [27] used a graph attention encoder to solve the" + }, + { + "type": "page_number", + "bbox": [ + 0.466, + 0.888, + 0.506, + 0.899 + ], + "angle": 0, + "content": "12576" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.126, + 0.473, + 0.252 + ], + "angle": 0, + "content": "problem of when and to whom to send messages. TMC [49] utilized temporal smoothing to ensure efficient communication. EC-MARL [2] tackled high-dimensional continuous control and partially observable states by introducing a rapid communication protocol to resolve task dilemmas. In comparison, our research focuses on LiDAR-based collaborative 3D object detection tasks in complex driving scenarios. We propose a hierarchical feature sampling strategy to achieve efficient communication across agents." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.262, + 0.307, + 0.276 + ], + "angle": 0, + "content": "2.2. Collaborative Perception" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.283, + 0.473, + 0.546 + ], + "angle": 0, + "content": "Factors including limited fields of view of sensors and physical obstructions in the environment can adversely impact the perception capabilities of individual agents. To address these challenges, collaborative perception within multi-agent systems has become a key technology. Utilizing newly available datasets [22, 43, 44], several novel collaborative perception methods have been proposed. For example, V2VNet [38] incorporated a graph neural network to fuse data from different agents. V2X-ViT [42] introduced a transformer architecture that combined information from vehicles and infrastructures. Where2comm [15] used feature spatial heterogeneity to reduce bandwidth utilization by transmitting sparse feature maps. CoBEVT [41] proposed the first multi-camera-based collaborative perception framework and designed the fused axis attention module to enable multi-view interactions. In this work, we propose an innovative feature calibration and interaction method that ensures robust vehicle collaboration in high-noise dynamic environments." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.561, + 0.17, + 0.575 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.585, + 0.47, + 0.642 + ], + "angle": 0, + "content": "This section introduces our efficient and robust collaborative perception method. Figure 1 shows the general procedure of our method, which is divided into six phases. We will detail each phase in the following introduction." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.651, + 0.437, + 0.667 + ], + "angle": 0, + "content": "3.1. Metadata Sharing and Feature Extraction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.673, + 0.473, + 0.868 + ], + "angle": 0, + "content": "In multi-vehicle collaboration scenarios, one of the CAVs, known as the ego vehicle, constructs a communication graph. In this graph, the ego vehicle serves as the requester, while other vehicles within its communication range act as supporters. The ego vehicle broadcasts its metadata information, including position, heading, speed, and more. Supporters, upon receiving this metadata, project their local point cloud observations into the ego vehicle's coordinate system. Each vehicle then encodes these transformed point clouds into bird's eye view (BEV) features, yielding a visual representation. Given the \\(i\\)-th vehicle local observations \\(\\mathbf{X}_i\\), the extracted features are represented as \\(\\mathbf{F}_i = \\Phi_{\\mathrm{enc}}(\\mathbf{X}_i) \\in \\mathbb{R}^{H \\times W \\times C}\\), where \\(\\Phi_{\\mathrm{enc}}(\\cdot)\\) denotes the PointPillar [19] encoder shared by all vehicles and \\(H\\), \\(W\\)," + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.126, + 0.892, + 0.169 + ], + "angle": 0, + "content": "and \\( C \\) stand for the height, width, and channel of the feature map, respectively. Then, these extracted features are fed into the filter and merge feature sampling module." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.175, + 0.809, + 0.191 + ], + "angle": 0, + "content": "3.2. Filter and Merge Feature Sampling" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.197, + 0.892, + 0.403 + ], + "angle": 0, + "content": "Previous works have utilized well-designed mechanisms such as information entropy communication selection [37] and spatial heterogeneity map [15, 35] to reduce the required transmission bandwidth. However, these methods primarily focus on the inter-class redundancy between foreground and background features, ignoring the intra-class redundancy among features, which results in sub-optimal compression. To address this gap, we introduce an advanced Filtering and Merging Feature Sampling strategy (FMS). This strategy considers both inter-class and intra-class redundancy relationships, efficiently extracting a concise and distinctive set of feature vectors from the original feature maps, thus reducing communication overheads more effectively. FMS is composed of two core components as follows." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.404, + 0.893, + 0.584 + ], + "angle": 0, + "content": "Filter Sampler. In object detection, foreground areas containing objects are more significant than the background areas. Therefore, we implement the idea of reducing spatial redundancy into a feature filter sampler module, aiming to preserve perceptually important yet sparse sets of feature vectors. Since explicitly learning a binary sampler is infeasible, we develop a confidence filter strategy. Initially, a detection confidence map is generated for the feature map. It reflects the perceptual importance of different spatial areas, with higher levels indicating potential object areas and lower levels typically denoting redundant background areas. For the feature map \\( \\pmb{F}_i \\) of the \\( i \\)-th vehicle, its confidence map \\( C_i \\) is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.583, + 0.592, + 0.891, + 0.61 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {C} _ {i} = \\Phi_ {\\text {c o n - g e n}} \\left(\\boldsymbol {F} _ {i}\\right) \\in [ 0, 1 ] ^ {H \\times W}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.618, + 0.892, + 0.768 + ], + "angle": 0, + "content": "where \\(\\Phi_{\\mathrm{con\\_gen}(\\cdot)}\\) represents the confidence generation network with detection decoder structure. Then the confidence map is thresholded, followed by non-maximum suppression, resulting in a binary mask \\(B\\). Utilizing this binary mask, we proceed to preserve the sparse foreground features \\(\\widetilde{\\pmb{F}}_i = B\\odot \\pmb{F}_i\\). In order to cope with the dynamically changing environmental conditions and fully ensure the robustness of the system, the filtering threshold can be varied with the sensor data and the network state. We set \\(\\alpha\\) as the filter rate and the number of remaining feature vectors is \\(L = \\alpha \\times HW\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.77, + 0.892, + 0.868 + ], + "angle": 0, + "content": "Merge Sampler. Upon extracting a detailed foreground feature vector set with the filter sampler, we employ the merge sampler for additional optimization, refining similar or repetitive foreground feature vectors through weighted merging. The process is divided into three stages: information-driven feature grouping, attention-inspired feature merging, and index-based feature reconstruction." + }, + { + "type": "page_number", + "bbox": [ + 0.466, + 0.888, + 0.506, + 0.899 + ], + "angle": 0, + "content": "12577" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.126, + 0.473, + 0.239 + ], + "angle": 0, + "content": "(a) Information-Driven Feature Grouping. Initially, a variant of the nearest neighbor clustering algorithm is applied to group the foreground feature vector set. Given a set of feature vector \\(\\widetilde{\\boldsymbol{F}}_i = [x_1,x_2,\\dots ,x_L]^\\top\\) and cluster center \\(X_{c}\\), we compute the indicator \\(\\delta_{i}\\) for each feature vector. \\(\\delta_{i}\\) is calculated as the minimum feature distance minus the average pixel distance to any other cluster center vector, expressed as:" + }, + { + "type": "equation", + "bbox": [ + 0.1, + 0.249, + 0.47, + 0.274 + ], + "angle": 0, + "content": "\\[\n\\delta_ {i} = \\min _ {j: x _ {j} \\in X _ {c}} \\left(\\| x _ {i} - x _ {j} \\| _ {2} ^ {2} - \\gamma \\| p (x _ {i}), p (x _ {j}) \\| _ {2} ^ {2}\\right), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.285, + 0.469, + 0.381 + ], + "angle": 0, + "content": "where \\(\\delta_{i}\\) denotes to which cluster the feature vector \\(x_{i}\\) should belong. \\(p(\\cdot)\\) means getting the position of the vector and \\(\\gamma\\) is a hyperparameter. Subsequently, we can divide all the feature vectors in \\(\\widetilde{\\boldsymbol{F}}_i\\) into \\(K\\) clusters represented by \\(G = \\{G_{1},G_{2},\\dots ,G_{K}\\}\\). The total number of clusters is determined by dynamic clustering ratio \\(\\beta\\), and is calculated as \\(K = \\beta \\times L\\)." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.382, + 0.47, + 0.492 + ], + "angle": 0, + "content": "(b) Attention-Inspired Feature Merging. A straightforward strategy for merging feature vectors is to average of each feature vector within a cluster. However, this scheme can be severely affected by outlier feature vectors. Drawing inspiration from attention mechanisms, we utilize the confidence score as a guide to quantify the significance of each feature. Thus, the merged feature vector \\(\\tilde{x}_i\\) for the \\(i\\)th cluster \\(G_i\\) is computed as:" + }, + { + "type": "equation", + "bbox": [ + 0.206, + 0.502, + 0.47, + 0.537 + ], + "angle": 0, + "content": "\\[\n\\widetilde {x} _ {i} = \\frac {\\sum_ {j \\in G _ {i}} c _ {j} x _ {j}}{\\sum_ {j \\in G _ {i}} c _ {j}}, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.546, + 0.469, + 0.601 + ], + "angle": 0, + "content": "where \\(c_{j}\\) and \\(x_{j}\\) represent the confidence score and the original feature vector, respectively. In the end, we obtain a final set of feature vectors \\(\\mathbf{Z}_i = \\{\\widetilde{x}_1,\\widetilde{x}_2,\\dots ,\\widetilde{x}_K\\}\\) that is necessary for transmission." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.602, + 0.469, + 0.713 + ], + "angle": 0, + "content": "(c) Index-Based Feature Reconstruction. During the feature grouping and merging process, each feature vector is allocated to a cluster, and every cluster is represented by a merged vector. We maintain a record of the index correspondence between original and merged feature vectors. Utilizing this index record, the ego-vehicle ensures that the merged feature vectors are mapped to their corresponding positions, leading to the reconstruction of feature maps." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.721, + 0.325, + 0.736 + ], + "angle": 0, + "content": "3.3. Feature Spatial Calibration" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.742, + 0.469, + 0.867 + ], + "angle": 0, + "content": "Localization errors [30, 31] may lead to the misalignment of feature maps among vehicles. Such misalignment causes the ego-vehicle to misinterpret the object's location, resulting in sub-optimal perception, as depicted in the upper right corner of Figure 2. To address this challenge, we introduce the Feature Spatial Calibration module (FSC) to facilitate precise feature alignment, as shown in Figure 2. The core idea of FSC is that overlapping viewpoints allow multiple vehicles to detect partially the same objects. The process" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.126, + 0.891, + 0.254 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.265, + 0.892, + 0.304 + ], + "angle": 0, + "content": "Figure 2. Illustration of the proposed FSC. By employing FSC, misaligned matching regions are corrected at the feature level, leading to more accurate predictions." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.313, + 0.891, + 0.34 + ], + "angle": 0, + "content": "involves three phases: consensus matching, geometric verification, and error modulation." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.341, + 0.892, + 0.534 + ], + "angle": 0, + "content": "Consensus Matching. After reconstructing the received features, the ego vehicle generates sparse feature maps. Within these maps, areas rich in information signify potential target regions, which correspond to the proposed matching areas. We denote the proposed matching regions of the ego vehicle as \\( P \\) and those identified by the collaborative vehicle under noisy pose conditions as \\( Q \\). Utilizing \\( P \\) and \\( Q \\), a weighted bipartite graph is constructed, wherein the weight of each edge is determined by the distance between the nodes, encapsulated in a cost matrix. The matching process is then converted into a linear assignment task, with the objective of identifying a matching result with the lowest cumulative edge weight. This procedure yields the matching pairs \\( M \\)." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.535, + 0.892, + 0.77 + ], + "angle": 0, + "content": "Geometric Verification. Invalid matches may occur due to objects located in exclusive zones and detection noise. To tackle this issue, we utilize RANSAC to filter and sift a consistent set of matches aligned with expected geometric transformations. Initially, a random matching subset \\(M_{s} \\subseteq M\\) is selected, and then the transformation matrix \\(\\Gamma_{s}\\) is computed using singular value decomposition. When \\(\\Gamma_{s}\\) is applied to all pairs within \\(M_{s}\\), and if the posttransformation distance between these pairs remains below the threshold \\(\\eta\\), the set is considered correctly aligned. The threshold \\(\\eta\\) reflects the allowable localization error within the original collaborative framework. This process is iteratively conducted to identify the optimal transformation matrix that correlates with the maximum number of correct matches. Ultimately, a optimal refined transformation matrix \\(\\Gamma_{r}\\) is obtained and applied in subsequent spatial calibration operations, yielding the aligned feature \\(\\widetilde{Z}_{j} = \\Gamma_{r}Z_{j}\\)." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.771, + 0.892, + 0.867 + ], + "angle": 0, + "content": "Error Modulation. To enhance the adaptability of the calibration method in various environments, we incorporate an error modulation strategy. This strategy aims to achieve a balance between localization errors and the estimation errors that emerge from the calibration process. It measures the overlap ratios between ego and collaborative features in both their adjusted and original states. Subsequently, fea" + }, + { + "type": "page_number", + "bbox": [ + 0.466, + 0.888, + 0.505, + 0.899 + ], + "angle": 0, + "content": "12578" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.084, + 0.123, + 0.465, + 0.254 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.263, + 0.47, + 0.303 + ], + "angle": 0, + "content": "Figure 3. A example of local and global Attention in AFF. This shows how AFF processes both 3D local windows and the global query token for position-level and context-aware aggregation." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.319, + 0.47, + 0.348 + ], + "angle": 0, + "content": "tures demonstrating the most optimal overlap are selected as the final input." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.355, + 0.357, + 0.369 + ], + "angle": 0, + "content": "3.4. Attention-based Feature Fusion" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.376, + 0.471, + 0.668 + ], + "angle": 0, + "content": "In multi-vehicle collaboration scenarios, vehicles are able to capture heterogeneous information from different spatial regions. In order to efficiently fuse perceptual features from multiple vehicles, we propose an Attention-based Feature Fusion method (AFF). As shown in Figure 3, AFF utilizes alternating local and global attention to achieve position-level accurate matching in occlusion-variable traffic scenes and to capture the global semantics attention of road topology and traffic states. Specifically, we stack all the vehicles' features to \\( S \\in \\mathbb{R}^{N \\times H \\times W \\times C} \\), where \\( N \\) is the number of vehicles. The features are divided into 3D non-overlapping windows, each of size \\( N \\times P \\times P \\). The shape of the partition tensor is \\( \\left(\\frac{H}{P} \\times \\frac{W}{P}, N \\times P^2, C\\right) \\). We perform 3D local self attention on the local window and implement region interaction within the window. In contrast, global attention goes beyond the limitations of the local view. It uses the extracted global query token and is shared between all windows to interact with the local key and value representations, thus helping to capture long-range dependencies of the features. Inspired by the efficiency of [13, 33], the global query token \\( g \\) is generated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.151, + 0.676, + 0.47, + 0.693 + ], + "angle": 0, + "content": "\\[\n\\mathbf {g} = \\mathcal {P} \\left(\\mathcal {C} _ {1 \\times 1} \\left(\\mathcal {S E} (\\mathcal {G} (\\mathcal {D} (\\mathbf {x}))) + \\mathbf {x}\\right)\\right), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.701, + 0.47, + 0.868 + ], + "angle": 0, + "content": "where \\(\\mathcal{D}\\), \\(\\mathcal{G}\\), \\(\\mathcal{SE}\\), \\(\\mathcal{C}_{1\\times 1}\\) and \\(\\mathcal{P}\\) denote the depth-wise separable convolution, GELU activation function, Squeeze-and-Excitation operation, \\(1\\times 1\\) convolution and max pooling, respectively. These designs enable the global query token to provide advantages such as inductive bias and modeling of inter-channel dependencies. We then construct our proposed AFF module by combining this local and global attention with typical designs of Transformers, including LayerNorm [1], MLPs [9], and skip-connections [14]. This module allows the system to analyze spatial correlations from a local view while also capturing global feature responses, ensuring efficient and precise perception in dy" + }, + { + "type": "image", + "bbox": [ + 0.51, + 0.125, + 0.885, + 0.255 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.518, + 0.265, + 0.871, + 0.278 + ], + "angle": 0, + "content": "Figure 4. The architecture of the proposed AFI component." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.307, + 0.89, + 0.334 + ], + "angle": 0, + "content": "namic, complex, and occlusion-variable traffic scenarios. Finally, we obtain the fused features \\( \\pmb{H}_i \\)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.351, + 0.844, + 0.366 + ], + "angle": 0, + "content": "3.5. Accuracy Enhanced Feature Interaction" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.374, + 0.892, + 0.652 + ], + "angle": 0, + "content": "Previous works [15, 21, 38, 42, 43] have demonstrated that fused features can provide richer semantic information, thereby enhancing perceptual performance. However, they may be affected by collaborative noise, such as asynchronous motion blur and inaccurate projections, which can compromise accurate position information and become a bottleneck to the optimal realization of perceptual performance. Ego-centric features may contain locally critical spatial location information without being affected by collaborative noise. To this end, we propose a novel Accuracy Enhanced Feature Interaction (AEI) strategy that leverages the accurate positional information inherent in ego-centric features to enhance the rich semantic information provided by collaborative fused features. Firstly, we design a context cross attention module that is tailored for feature interaction, as shown in Figure 4. It considers the features from the ego-vehicle as queries, and obtains the context of the query location from the fused features as the key and value to achieve cross-attention. The computed output of the spatial location \\((i,j)\\) is as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.564, + 0.666, + 0.892, + 0.699 + ], + "angle": 0, + "content": "\\[\ny _ {i j} = \\sum_ {a, b \\in \\mathcal {N} _ {k} (i, j)} \\operatorname {s o f t m a x} \\left(q _ {i j} ^ {\\top} k _ {a b}\\right) v _ {a b}, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.713, + 0.893, + 0.868 + ], + "angle": 0, + "content": "where \\(\\mathcal{N}_k(i,j)\\) is context position of \\((i,j)\\) and \\(q_{ij}, k_{ab}, v_{ab}\\) represent the ego query, context key and value, respectively. As the relevant features of the same object are often located in similar locations across different feature maps, we prioritize the local context details of the query location. Additionally, considering that only a few positions within the sparse feature map hold significant information, focusing on these context-rich areas is computationally efficient. Following this, we employ the split attention [48] to adaptively fuse information from multiple branches, generating the enhanced output feature \\(\\widetilde{H}_i\\)." + }, + { + "type": "page_number", + "bbox": [ + 0.466, + 0.888, + 0.506, + 0.899 + ], + "angle": 0, + "content": "12579" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.08, + 0.123, + 0.894, + 0.31 + ], + "angle": 0, + "content": "
DatasetV2V4RealOPV2V
Noise Level σt/σr (m/°)0.0/0.00.4/0.40.8/0.80.0/0.00.4/0.40.8/0.8
No Collaboration41.64/23.4341.64/23.4341.64/23.4373.25/58.2273.25/58.2273.25/58.22
Late Fusion56.80/27.6240.81/13.3126.65/10.6686.46/79.2176.56/40.0145.57/21.00
F-Cooper [3]61.12/32.3651.71/24.4841.90/19.9187.39/79.3979.33/40.1252.91/16.90
AttFuse [7]64.41/34.3258.42/30.0249.37/25.4490.50/81.8083.54/52.4368.37/40.96
Where2comm [15]63.69/34.7958.48/28.0949.77/24.2590.46/84.2282.17/56.4773.95/45.86
DiscoNet [21]64.19/35.2557.71/28.4149.19/24.9489.58/81.4583.25/53.2261.37/30.53
V2VNet [38]65.70/35.3862.11/30.6354.73/25.5791.35/82.4385.43/54.1670.43/31.28
CoAlign [25]64.06/36.6059.58/31.5253.02/26.6090.93/84.2884.48/57.4573.95/48.86
V2X-ViT [42]66.42/37.3463.33/32.3857.15/28.2291.74/83.3182.91/54.7361.70/26.07
CoBEVT [41]66.01/37.3658.63/29.0949.22/23.6691.71/85.9885.49/60.6464.63/29.59
ERMVP (Ours)67.66/42.9765.01/40.4560.88/35.0192.18/85.5985.67/66.3277.13/59.15
" + }, + { + "type": "table_caption", + "bbox": [ + 0.118, + 0.32, + 0.85, + 0.333 + ], + "angle": 0, + "content": "Table 1. Overall performance on V2V4Real and OPV2V datasets with pose noises. The results are reported in AP@0.5/0.7." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.34, + 0.251, + 0.354 + ], + "angle": 0, + "content": "3.6. Decoder and Loss" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.363, + 0.47, + 0.477 + ], + "angle": 0, + "content": "Based on the final fused feature \\(\\widetilde{H}_i\\), we use the detection decoder to generate the final prediction output \\(Y_{i} = \\Phi_{\\mathrm{dec}}(\\widetilde{H}_{i})\\). Each position of \\(Y_{i}\\) represents a rotated box with classes (c, x, y, h, w, \\(\\cos \\alpha\\), \\(\\sin \\alpha\\)), corresponding to class confidence, position, size, and angle. These objects are the final output of the proposed collaborative perception system. Following existing work [19], we adopt the smooth \\(L_{1}\\) loss for regression and focal loss [23] for classification." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.497, + 0.21, + 0.512 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.522, + 0.364, + 0.536 + ], + "angle": 0, + "content": "4.1. Datasets and Evaluation Metrics" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.546, + 0.47, + 0.794 + ], + "angle": 0, + "content": "Datasets. We validate the proposed ERMVP in the task of LiDAR-based 3D object detection on two benchmark datasets including V2V4real [44] and OPV2V [43]. V2V4real [44] is the first large-scale real-world dataset on V2V perception, collected by two vehicles equipped with multi-modal sensors driving together in a variety of scenarios. It covers a driving area of 410 kilometers, including 20K frames of point clouds, with training/validation/test sets containing 14210/2000/3986 frames, respectively. OPV2V [43] is a large-scale vehicle-to-vehicle collaborative perception dataset, simulated by OpenCDA [40] and Carla [8]. It contains 73 different scenarios, in which different numbers (2 to 7) of collaborative vehicles appear together, each equipped with a lidar sensor and 4 cameras. The dataset contains a total of 11,464 frames of point clouds and RGB images. The training/validation/test sets contain 6374, 1980, and 2170 frames, respectively." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.797, + 0.469, + 0.867 + ], + "angle": 0, + "content": "Evaluation Metrics. We adopt the Average Precision (AP) at Intersection-over-Union (IoU) thresholds of 0.5 and 0.7 to evaluate the detection performance. The calculation format of communication volume in [15] is used to count the message size by byte in the log scale with base 2." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.34, + 0.719, + 0.355 + ], + "angle": 0, + "content": "4.2. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.363, + 0.892, + 0.626 + ], + "angle": 0, + "content": "We implement the proposed ERMVP and comparison model on the Pytorch toolbox [28], and train them on a NVIDIA GeForceRTX 3090 GPU using the Adam optimizer [18]. The initial learning rate is \\(2\\mathrm{e} - 3\\) and it decays every 15 epochs with a factor of 0.1. All models are trained with 60 epochs and the batch size is set to 2. Early stopping is used to find the best epoch. We also add normal point cloud data augmentation for all experiments, including scaling, rotation, and flipping. All detection models are based on PointPillar [19] backbone to extract 2D features from point clouds and a \\(0.4\\mathrm{m}\\) width/length is used for each voxel. AFF component has 3 encoded layers and a window size of 8 for both local and global attention. The balance hyperparameter \\(\\gamma\\) is 0.1 and the error tolerance threshold \\(\\eta\\) is 0.25. Each vehicle has a communication range of 70 m based on [42], while vehicles outside of this broadcasting radius will be ignored. To simulate the localization and heading errors, we add Gaussian noise with a standard deviation of \\(\\sigma_t\\) for localization errors and \\(\\sigma_r\\) for heading errors." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.637, + 0.722, + 0.652 + ], + "angle": 0, + "content": "4.3.Quantitative Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.659, + 0.892, + 0.867 + ], + "angle": 0, + "content": "Comparison of Detection Performance. Table 1 shows the performance comparison results of 3D detection on two datasets. We use the No Collaboration method as a baseline, which only uses ego vehicle's point clouds data without collaboration. Late Fusion allows vehicles to exchange detected outputs and utilizes non-maximum suppression to produce the final result. For intermediate fusion strategies, we evaluate the existing state-of-the-art (SOTA) methods: AttFuse [43], F-Cooper [3], V2VNet [38], DiscoNet [21], V2X-ViT [42], CoAlign (only fusion) [25], CoBEVT [41] and Where2comm [15]. The proposed ER-MVP achieves average improvements of \\(12.48\\% / 18.47\\%\\) and \\(6.18\\% / 10.02\\%\\) on two datasets in AP@0.5/0.7 compared with No Collaboration and Late Fusion methods, demonstrating the superiority of our collaborative method." + }, + { + "type": "page_number", + "bbox": [ + 0.466, + 0.888, + 0.506, + 0.899 + ], + "angle": 0, + "content": "12580" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.123, + 0.283, + 0.252 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.284, + 0.123, + 0.484, + 0.251 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.484, + 0.123, + 0.683, + 0.251 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.683, + 0.123, + 0.882, + 0.251 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.259, + 0.256, + 0.341, + 0.267 + ], + "angle": 0, + "content": "(a) V2V4Real" + }, + { + "type": "image_footnote", + "bbox": [ + 0.137, + 0.273, + 0.855, + 0.291 + ], + "angle": 0, + "content": "Late Fusion FCooper V2VNet V2X-VIT CoBEVT -Where2comm -ERMVP" + }, + { + "type": "image_caption", + "bbox": [ + 0.102, + 0.298, + 0.866, + 0.312 + ], + "angle": 0, + "content": "Figure 5. Collaborative perception performance comparison on the V2V4Real and OPV2V with varying communication volume." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.322, + 0.47, + 0.42 + ], + "angle": 0, + "content": "Meanwhile, it is superior to the SOTA collaborative perception method in both real-world and simulated scenarios: improves the SOTA collaborative performance by \\(1.24\\%\\) and \\(5.61\\%\\) on V2V4Real datasets in AP@0.5/0.7, and by \\(0.44\\%\\) on OPV2V datasets in AP@0.5. These results fully demonstrate the superiority of the ERMVP collaboration paradigm." + }, + { + "type": "table", + "bbox": [ + 0.08, + 0.43, + 0.456, + 0.602 + ], + "angle": 0, + "content": "
DatasetV2V4RealOPV2V
Noise Level σt(m)0.51.00.51.0
Late Fusion11.0210.1353.8838.80
F-Cooper [3]24.6018.4132.2412.71
AttFuse [43]28.1024.3642.9535.53
DiscoNet [21]28.3624.0545.3624.96
V2VNet [38]29.5723.6747.4825.82
Where2comm [15]28.0223.5346.5945.59
CoAlign [21]29.8425.8249.0946.66
V2X-ViT [42]30.9923.6744.8321.77
CoBEVT [41]24.6722.0141.9126.05
ERMVP (Ours)39.2633.9477.2576.09
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.611, + 0.47, + 0.637 + ], + "angle": 0, + "content": "Table 2. Detection AP@0.7 on V2V4Real and OPV2V datasets with localization error." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.646, + 0.471, + 0.866 + ], + "angle": 0, + "content": "Robustness to Localization and Heading Errors. To evaluate the sensitivity of existing methods to localization and heading errors, we use the same noise setting as [42] and conduct extensive experiments on two datasets. As shown in Tables 1 and 2, under ideal settings, Late Fusion and some advanced intermediate fusion methods can detect object vehicles with high accuracy, but their detection accuracy rapidly decreases as the standard deviation of errors increases. For example, when the localization and heading errors of OPV2V dataset are over \\(0.4\\mathrm{m}\\) and \\(0.4^{\\circ}\\), the Disconet [21] and F-Cooper [3] methods fail and perform even worse than No Collaboration. Our method exceeds previous SOTA models at all noise levels and consistently outperforms the No Collaboration baseline, clearly demonstrating the robustness of ERMVP to pose errors. It is noteworthy that for localization errors, ERMVP can maintain consistent" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.322, + 0.892, + 0.432 + ], + "angle": 0, + "content": "accuracy levels in high noise environments, showing superior robustness. When the localization error is \\(1.0\\mathrm{m}\\), it outperforms the second-best method by \\(29.43\\%\\) and \\(8.08\\%\\) in AP@0.7 on OPV2V and V2V4Real datasets. The reasonable explanations are: (i) feature spatial calibration aligns mismatched collaborative features; (ii) accuracy enhanced feature interaction reduces performance degradation caused by collaborative noise." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.433, + 0.893, + 0.626 + ], + "angle": 0, + "content": "Comparison of Communication Volume. To achieve pragmatic collaborative perception, it is crucial to evaluate the perceptual performance under different communication volumes. Figure 5 shows the results of the performance comparison under different bandwidth consumption conditions. Note that we do not consider any additional model/data/feature compression for a fair comparison. The results indicate that ERMVP achieves excellent perceptual performance and bandwidth consumption trade-off under all communication bandwidth conditions, consistently outperforming Where2comm [15]. At the same time, it achieves the same detection performance as the previous model [3, 38, 41, 42] with less communication volume on both real-world and simulated datasets." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.633, + 0.665, + 0.646 + ], + "angle": 0, + "content": "4.4. Ablation Studies" + }, + { + "type": "table", + "bbox": [ + 0.501, + 0.654, + 0.887, + 0.728 + ], + "angle": 0, + "content": "
AFFFSCAEIV2V4RealOPV2V
51.71/24.4879.33/40.12
✓62.24/37.7084.30/61.43
✓✓64.36/39.5984.98/62.67
✓✓✓65.01/40.4585.67/66.32
" + }, + { + "type": "table_caption", + "bbox": [ + 0.498, + 0.737, + 0.892, + 0.789 + ], + "angle": 0, + "content": "Table 3. Ablation study results of the proposed core components on the both datasets with noise level of 0.4/0.4. AFF: Attention-based Feature Fusion; FSC: Feature Spatial Calibration; AEI: Accuracy Enhanced Feature Interaction." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.797, + 0.892, + 0.868 + ], + "angle": 0, + "content": "Effect of Core Components. Table 3 details the contribution of each core components in our ERMVP framework. The base model is the naive position-wise maximum fusion method. We then assess the impact of each component by sequentially introducing i) attention-based feature" + }, + { + "type": "page_number", + "bbox": [ + 0.466, + 0.888, + 0.505, + 0.899 + ], + "angle": 0, + "content": "12581" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.088, + 0.125, + 0.285, + 0.211 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.088, + 0.213, + 0.284, + 0.297 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.14, + 0.301, + 0.218, + 0.312 + ], + "angle": 0, + "content": "(a) V2VNet" + }, + { + "type": "image", + "bbox": [ + 0.287, + 0.125, + 0.483, + 0.211 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.287, + 0.213, + 0.484, + 0.296 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.34, + 0.301, + 0.424, + 0.312 + ], + "angle": 0, + "content": "(b) CoBEVT" + }, + { + "type": "image", + "bbox": [ + 0.485, + 0.125, + 0.682, + 0.211 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.486, + 0.213, + 0.682, + 0.296 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.54, + 0.301, + 0.621, + 0.312 + ], + "angle": 0, + "content": "(c) V2X-ViT" + }, + { + "type": "image", + "bbox": [ + 0.684, + 0.125, + 0.883, + 0.211 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.685, + 0.213, + 0.882, + 0.296 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.738, + 0.301, + 0.818, + 0.312 + ], + "angle": 0, + "content": "(d)ERMVP" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.322, + 0.893, + 0.349 + ], + "angle": 0, + "content": "Figure 6. Qualitative comparison results in real-world scenarios from the V2V4Real dataset with noise level of 0.4/0.4. Green and red boxes denote the ground truths and detection results, respectively." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.356, + 0.47, + 0.455 + ], + "angle": 0, + "content": "fusion (AFF), ii) feature spatial calibration (FSC), iii) accuracy enhanced feature interaction (AEF). The consistent rise in detection results over both datasets demonstrates the effectiveness of each introduced component. Notably, integrating all three components boosts detection performance by \\(13.3\\%\\) and \\(15.97\\%\\) on the V2V4Real dataset for AP@0.5 and AP@0.7, respectively." + }, + { + "type": "image", + "bbox": [ + 0.09, + 0.464, + 0.268, + 0.573 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.281, + 0.464, + 0.46, + 0.573 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.135, + 0.582, + 0.411, + 0.595 + ], + "angle": 0, + "content": "Figure 7. Hyperparameter analysis of \\(\\alpha\\) and \\(\\beta\\)" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.604, + 0.473, + 0.868 + ], + "angle": 0, + "content": "Hyperparameters \\(\\alpha\\) and \\(\\beta\\). We explore the impact of the hyperparameters \\(\\alpha\\) and \\(\\beta\\) (as described in Section 3.2), which jointly determine the communication volume. Specifically, the feature compression ratio is represented as \\(\\alpha \\times \\beta\\). The communication volume, in turn, is expressed as the product of the original feature size and the compression rate. By modulating the values of \\(\\alpha\\) and \\(\\beta\\) within the interval of \\(10\\%\\) to \\(80\\%\\), we chart the resultant object detection performance (measured by AP@0.7) in Figure 7. As expected, a drop in either \\(\\alpha\\) or \\(\\beta\\) will impair ERMVP's efficacy. Yet, it's worth noting that ERMVP's performance generally exhibits robustness to these hyperparameters. The performance decline remains fairly minimal even under significant compression, such as a factor of 100. The results demonstrate the effectiveness of our proposed filter and merge feature sampling module. Therefore, ERMVP establishes an appropriate and effective communication paradigm among vehicles, ensuring an optimal tradeoff between perception and communication." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.356, + 0.71, + 0.37 + ], + "angle": 0, + "content": "4.5. Qualitative Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.377, + 0.892, + 0.584 + ], + "angle": 0, + "content": "Figure 6 displays the detection results of V2VNet [38], CoBEVT [41], V2X-ViT [42], and ERMVP in two scenarios. Clearly, ERMVP offers more precise and comprehensive detection than the prior SOTA methods. To begin with, ERMVP produces a greater number of predicted bounding boxes that are well aligned with the ground truths, while other methods show significant discrepancies. This demonstrates ERMVP's robustness, particularly in high-noise environments. Moreover, ERMVP detects more dynamic objects (more ground truth bounding boxes find matches). This suggests that ERMVP can effectively combine the inputs from nearby vehicles, leading to a thorough scene representation. Overall, these qualitative assessments confirm ERMVP's strengths in delivering accurate and comprehensive perception, especially in challenging conditions." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.592, + 0.618, + 0.606 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.615, + 0.892, + 0.754 + ], + "angle": 0, + "content": "In this paper, we present ERMVP, a communication-efficient and collaboration-robust multi-vehicle perception method in challenging environments. We introduce a filter and merge feature sampling strategy to efficient communication, a feature spatial calibration module to align features in fine-grid and two spatial information aggregation components to solve the bottleneck of the fusion method. Extensive experiments prove the superiority of ERMVP and the effectiveness of our components. The feasibility of our approach for other modalities will be explored in future work." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.761, + 0.659, + 0.777 + ], + "angle": 0, + "content": "Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.784, + 0.892, + 0.867 + ], + "angle": 0, + "content": "This work is supported by the Shanghai Key Research Laboratory of NSAI, the Innovation Platform for Academicians of Hainan Province, Haikou, 570228, the Specific Research Fund of the Innovation Platform for Academicians of Hainan Province under Grant YSPTZX202314, and NFSC under grant 62250410368." + }, + { + "type": "page_number", + "bbox": [ + 0.466, + 0.888, + 0.505, + 0.898 + ], + "angle": 0, + "content": "12582" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.125, + 0.174, + 0.139 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.148, + 0.469, + 0.185 + ], + "angle": 0, + "content": "[1] Jimmy Lei Ba, Jamie Ryan Kiros, and Geoffrey E Hinton. Layer normalization. arXiv preprint arXiv:1607.06450, 2016. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.188, + 0.469, + 0.25 + ], + "angle": 0, + "content": "[2] Marwa Chafii, Salmane Naoumi, Reda Alami, Ebtesam Al-mazrouei, Mehdi Bennis, and Merouane Debbah. Emergent communication in multi-agent reinforcement learning for future wireless networks. arXiv preprint arXiv:2309.06021, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.253, + 0.469, + 0.316 + ], + "angle": 0, + "content": "[3] Qi Chen, Xu Ma, Sihai Tang, Jingda Guo, Qing Yang, and Song Fu. F-cooper: Feature based cooperative perception for autonomous vehicle edge computing system using 3d point clouds. In Proceedings of the 4th ACM/IEEE Symposium on Edge Computing, pages 88-100, 2019. 1, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.318, + 0.469, + 0.369 + ], + "angle": 0, + "content": "[4] Zhaoyu Chen, Bo Li, Shuang Wu, Jianghe Xu, Shouhong Ding, and Wenqiang Zhang. Shape matters: deformable patch attack. In European conference on computer vision, pages 529-548. Springer, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.371, + 0.469, + 0.421 + ], + "angle": 0, + "content": "[5] Zhaoyu Chen, Bo Li, Shuang Wu, Shouhong Ding, and Wenqiang Zhang. Query-efficient decision-based black-box patch attack. IEEE Transactions on Information Forensics and Security, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.423, + 0.469, + 0.473 + ], + "angle": 0, + "content": "[6] Zhaoyu Chen, Bo Li, Shuang Wu, Kaixun Jiang, Shouhong Ding, and Wenqiang Zhang. Content-based unrestricted adversarial attack. Advances in Neural Information Processing Systems, 36, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.476, + 0.47, + 0.526 + ], + "angle": 0, + "content": "[7] Yimian Dai, Fabian Gieseke, Stefan Oehmcke, Yiquan Wu, and Kobus Barnard. Attentional feature fusion. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 3560-3569, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.528, + 0.469, + 0.578 + ], + "angle": 0, + "content": "[8] Alexey Dosovitskiy, German Ros, Felipe Codevilla, Antonio Lopez, and Vladlen Koltun. Carla: An open urban driving simulator. In Conference on robot learning, pages 1-16. PMLR, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.581, + 0.469, + 0.655 + ], + "angle": 0, + "content": "[9] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.659, + 0.469, + 0.684 + ], + "angle": 0, + "content": "[10] Ian Goodfellow, Yoshua Bengio, and Aaron Courville. Deep learning. MIT press, 2016. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.686, + 0.469, + 0.761 + ], + "angle": 0, + "content": "[11] Jiaming Gu, Jingyu Zhang, Muyang Zhang, Weiliang Meng, Shibiao Xu, Jiguang Zhang, and Xiaopeng Zhang. Feaco: Reaching robust feature-level consensus in noisy pose conditions. In Proceedings of the 31st ACM International Conference on Multimedia, page 3628-3636, New York, NY, USA, 2023. Association for Computing Machinery. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.764, + 0.469, + 0.813 + ], + "angle": 0, + "content": "[12] Monowar Hasan, Sibin Mohan, Takayuki Shimizu, and Hongsheng Lu. Securing vehicle-to-everything (v2x) communication platforms. IEEE Transactions on Intelligent Vehicles, 5(4):693-713, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.816, + 0.469, + 0.865 + ], + "angle": 0, + "content": "[13] Ali Hatamizadeh, Hongxu Yin, Greg Heinrich, Jan Kautz, and Pavlo Molchanov. Global context vision transformers. In International Conference on Machine Learning, pages 12633-12646. PMLR, 2023. 5" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.148, + 0.47, + 0.865 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.127, + 0.892, + 0.178 + ], + "angle": 0, + "content": "[14] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.18, + 0.892, + 0.242 + ], + "angle": 0, + "content": "[15] Yue Hu, Shaoheng Fang and Zixing Lei, Yiqi Zhong, and Siheng Chen. Where2comm: Communication-efficient collaborative perception via spatial confidence maps. In Thirty-sixth Conference on Neural Information Processing Systems (NeurIPS), 2022. 1, 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.245, + 0.892, + 0.318 + ], + "angle": 0, + "content": "[16] Kaixun Jiang, Zhaoyu Chen, Hao Huang, Jiafeng Wang, Dingkang Yang, Bo Li, Yan Wang, and Wenqiang Zhang. Efficient decision-based black-box patch attacks on video recognition. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 4379-4389, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.322, + 0.892, + 0.371 + ], + "angle": 0, + "content": "[17] Hamza Khan, Sumudu Samarakoon, and Mehdi Bennis. Enhancing video streaming in vehicular networks via resource slicing. IEEE Transactions on Vehicular Technology, 69(4): 3513-3522, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.374, + 0.892, + 0.412 + ], + "angle": 0, + "content": "[18] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In International Conference on Learning Representations (ICLR), 2015. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.414, + 0.892, + 0.476 + ], + "angle": 0, + "content": "[19] Alex H Lang, Sourabh Vora, Holger Caesar, Lubing Zhou, Jiong Yang, and Oscar Beijbom. Pointpillars: Fast encoders for object detection from point clouds. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12697-12705, 2019. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.478, + 0.892, + 0.527 + ], + "angle": 0, + "content": "[20] Yiming Li, Bir Bhanu, and Wei Lin. Auction protocol for camera active control. In 2010 IEEE International Conference on Image Processing, pages 4325-4328. IEEE, 2010. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.531, + 0.892, + 0.58 + ], + "angle": 0, + "content": "[21] Yiming Li, Shunli Ren, Pengxiang Wu, Siheng Chen, Chen Feng, and Wenjun Zhang. Learning distilled collaboration graph for multi-agent perception. Advances in Neural Information Processing Systems, 34:29541-29552, 2021. 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.582, + 0.892, + 0.631 + ], + "angle": 0, + "content": "[22] Yiming Li, Ziyan An, Zixun Wang, Yiqi Zhong, Siheng Chen, and Chen Feng. V2x-sim: A virtual collaborative perception dataset for autonomous driving. arXiv preprint arXiv:2202.08449, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.634, + 0.892, + 0.684 + ], + "angle": 0, + "content": "[23] Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dollar. Focal loss for dense object detection. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.686, + 0.892, + 0.749 + ], + "angle": 0, + "content": "[24] Yang Liu, Jing Liu, Kun Yang, Bobo Ju, Siao Liu, Yuzheng Wang, Dingkang Yang, Peng Sun, and Liang Song. Ampnet: Appearance-motion prototype network assisted automatic video anomaly detection system. IEEE Transactions on Industrial Informatics, 20(2):2843-2855, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.751, + 0.892, + 0.813 + ], + "angle": 0, + "content": "[25] Yifan Lu, Quanhao Li, Baoan Liu, Mehrdad Dianati, Chen Feng, Siheng Chen, and Yanfeng Wang. Robust collaborative 3d object detection in presence of pose errors. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 4812-4818. IEEE, 2023. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.816, + 0.892, + 0.866 + ], + "angle": 0, + "content": "[26] Zonglin Meng, Xin Xia, Runsheng Xu, Wei Liu, and Jiaqi Ma. Hydro-3d: Hybrid object detection and tracking for cooperative perception using 3d lidar. IEEE Transactions on Intelligent Vehicles, 2023. 1" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.127, + 0.892, + 0.866 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.467, + 0.888, + 0.505, + 0.899 + ], + "angle": 0, + "content": "12583" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.127, + 0.47, + 0.165 + ], + "angle": 0, + "content": "[27] Yaru Niu, Rohan R Paleja, and Matthew C Gombolay. Multiagent graph-attention communication and teaming. In AA-MAS, pages 964–973, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.168, + 0.47, + 0.243 + ], + "angle": 0, + "content": "[28] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. Advances in neural information processing systems, 32, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.246, + 0.469, + 0.284 + ], + "angle": 0, + "content": "[29] Faisal Qureshi and Demetri Terzopoulos. Smart camera networks in virtual reality. Proceedings of the IEEE, 96(10): 1640-1656, 2008. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.287, + 0.469, + 0.363 + ], + "angle": 0, + "content": "[30] Qingyu Song, Changan Wang, Zhengkai Jiang, Yabiao Wang, Ying Tai, Chengjie Wang, Jilin Li, Feiyue Huang, and Yang Wu. Rethinking counting and localization in crowds: A purely point-based framework. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3365-3374, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.365, + 0.469, + 0.415 + ], + "angle": 0, + "content": "[31] Zhiying Song, Fuxi Wen, Hailiang Zhang, and Jun Li. An efficient and robust object-level cooperative perception framework for connected and automated driving. arXiv preprint arXiv:2210.06289, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.419, + 0.469, + 0.468 + ], + "angle": 0, + "content": "[32] Ming Tan. Multi-agent reinforcement learning: Independent vs. cooperative agents. In Proceedings of the tenth international conference on machine learning, pages 330-337, 1993. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.471, + 0.469, + 0.51 + ], + "angle": 0, + "content": "[33] Mingxing Tan and Quoc Le. Efficientnetv2: Smaller models and faster training. In International conference on machine learning, pages 10096-10106. PMLR, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.512, + 0.469, + 0.562 + ], + "angle": 0, + "content": "[34] Nicholas Vadivelu, Mengye Ren, James Tu, Jingkang Wang, and Raquel Urtasun. Learning to communicate and correct pose errors. In Conference on Robot Learning, pages 1195-1210. PMLR, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.565, + 0.469, + 0.628 + ], + "angle": 0, + "content": "[35] Binglu Wang, Lei Zhang, Zhaozhong Wang, Yongqiang Zhao, and Tianfei Zhou. Core: Cooperative reconstruction for multi-agent perception. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8710-8720, 2023. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.631, + 0.469, + 0.668 + ], + "angle": 0, + "content": "[36] Jian Wang, Yameng Shao, Yuming Ge, and Rundong Yu. A survey of vehicle to everything (v2x) testing. Sensors, 19(2): 334, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.671, + 0.469, + 0.734 + ], + "angle": 0, + "content": "[37] Tianhang Wang, Guang Chen, Kai Chen, Zhengfa Liu, Bo Zhang, Alois Knoll, and Changjun Jiang. Umc: A unified bandwidth-efficient and multi-resolution based collaborative perception framework. arXiv preprint arXiv:2303.12400, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.737, + 0.469, + 0.801 + ], + "angle": 0, + "content": "[38] Tsun-Hsuan Wang, Sivabanan Manivasagam, Ming Liang, Bin Yang, Wenyuan Zeng, and Raquel Urtasun. V2vnet: Vehicle-to-vehicle communication for joint perception and prediction. In European Conference on Computer Vision, pages 605–621. Springer, 2020. 1, 3, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.804, + 0.469, + 0.866 + ], + "angle": 0, + "content": "[39] Qianyi Wu, Kaisiyuan Wang, Kejie Li, Jianmin Zheng, and Jianfei Cai. Objectsdf++: Improved object-compositional neural implicit surfaces. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 21764-21774, 2023. 1" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.127, + 0.47, + 0.866 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.127, + 0.892, + 0.19 + ], + "angle": 0, + "content": "[40] Runsheng Xu, Yi Guo, Xu Han, Xin Xia, Hao Xiang, and Jiaqi Ma. Opencda: an open cooperative driving automation framework integrated with co-simulation. In 2021 IEEE International Intelligent Transportation Systems Conference (ITSC), pages 1155-1162. IEEE, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.192, + 0.892, + 0.242 + ], + "angle": 0, + "content": "[41] Runsheng Xu, Zhengzhong Tu, Hao Xiang, Wei Shao, Bolei Zhou, and Jiaqi Ma. Cobevt: Cooperative bird's eye view semantic segmentation with sparse transformers, 2022. 1, 3, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.244, + 0.892, + 0.306 + ], + "angle": 0, + "content": "[42] Runsheng Xu, Hao Xiang, Zhengzhong Tu, Xin Xia, Ming-Hsuan Yang, and Jiaqi Ma. V2x-vit: Vehicle-to-everything cooperative perception with vision transformer. In European Conference on Computer Vision, page 107-124. Springer, 2022. 1, 3, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.308, + 0.892, + 0.381 + ], + "angle": 0, + "content": "[43] Runsheng Xu, Hao Xiang, Xin Xia, Xu Han, Jinlong Li, and Jiaqi Ma. Opv2v: An open benchmark dataset and fusion pipeline for perception with vehicle-to-vehicle communication. In 2022 International Conference on Robotics and Automation (ICRA), pages 2583–2589. IEEE, 2022. 2, 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.385, + 0.892, + 0.461 + ], + "angle": 0, + "content": "[44] Runsheng Xu, Xin Xia, Jinlong Li, Hanzhao Li, Shuo Zhang, Zhengzhong Tu, Zonglin Meng, Hao Xiang, Xiaoyu Dong, Rui Song, et al. V2v4real: A real-world large-scale dataset for vehicle-to-vehicle cooperative perception. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13712–13722, 2023. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.462, + 0.892, + 0.526 + ], + "angle": 0, + "content": "[45] Kun Yang, Peng Sun, Jieyu Lin, Azzedine Boukerche, and Liang Song. A novel distributed task scheduling framework for supporting vehicular edge intelligence. In 2022 IEEE 42nd International Conference on Distributed Computing Systems (ICDCS), pages 972-982. IEEE, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.528, + 0.892, + 0.577 + ], + "angle": 0, + "content": "[46] Kun Yang, Dingkang Yang, Jingyu Zhang, Mingcheng Li, Yang Liu, Jing Liu, Hanqi Wang, Peng Sun, and Liang Song. Spatio-temporal domain awareness for multi-agent collaborative perception, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.579, + 0.892, + 0.654 + ], + "angle": 0, + "content": "[47] Kun Yang, Dingkang Yang, Jingyu Zhang, Hanqi Wang, Peng Sun, and Liang Song. What2comm: Towards communication-efficient collaborative perception via feature decoupling. In Proceedings of the 31st ACM International Conference on Multimedia, page 7686-7695, New York, NY, USA, 2023. Association for Computing Machinery. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.655, + 0.892, + 0.731 + ], + "angle": 0, + "content": "[48] Hang Zhang, Chongruo Wu, Zhongyue Zhang, Yi Zhu, Haibin Lin, Zhi Zhang, Yue Sun, Tong He, Jonas Mueller, R. Manmatha, Mu Li, and Alexander Smola. Resnest: Split-attention networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, pages 2736-2746, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.733, + 0.892, + 0.782 + ], + "angle": 0, + "content": "[49] Sai Qian Zhang, Qi Zhang, and Jieyu Lin. Succinct and robust multi-agent communication with temporal message control. Advances in Neural Information Processing Systems, 33:17271-17282, 2020. 3" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.127, + 0.892, + 0.782 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.467, + 0.888, + 0.505, + 0.899 + ], + "angle": 0, + "content": "12584" + } + ] +] \ No newline at end of file diff --git a/2024/ERMVP_ Communication-Efficient and Collaboration-Robust Multi-Vehicle Perception in Challenging Environments/197e9226-8724-4a9f-ac9a-1556459fa452_origin.pdf b/2024/ERMVP_ Communication-Efficient and Collaboration-Robust Multi-Vehicle Perception in Challenging Environments/197e9226-8724-4a9f-ac9a-1556459fa452_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..1a46d3cba99e90e454022ea85253da7d30545c07 --- /dev/null +++ b/2024/ERMVP_ Communication-Efficient and Collaboration-Robust Multi-Vehicle Perception in Challenging Environments/197e9226-8724-4a9f-ac9a-1556459fa452_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ecb2465d7f16469e301765389a7c4da1400282ac38680a92e63c4c85621f3e62 +size 1332501 diff --git a/2024/ERMVP_ Communication-Efficient and Collaboration-Robust Multi-Vehicle Perception in Challenging Environments/full.md b/2024/ERMVP_ Communication-Efficient and Collaboration-Robust Multi-Vehicle Perception in Challenging Environments/full.md new file mode 100644 index 0000000000000000000000000000000000000000..d5af3e8986d505cb0ee3531dfdcc1908a66f97ce --- /dev/null +++ b/2024/ERMVP_ Communication-Efficient and Collaboration-Robust Multi-Vehicle Perception in Challenging Environments/full.md @@ -0,0 +1,280 @@ +# ERMVP: Communication-Efficient and Collaboration-Robust Multi-Vehicle Perception in Challenging Environments + +Jingyu Zhang $^{1}$ Kun Yang $^{1}$ Yilei Wang $^{1}$ Hanqi Wang $^{1}$ Peng Sun $^{2,*}$ Liang Song $^{1,*}$ $^{1}$ Academy for Engineering and Technology, Fudan University $^{2}$ Duke Kunshan University +{jingyuzhang22, yileiwang23}@m.fudan.edu.cn, songl@m.fudan.edu.cn + +# Abstract + +Collaborative perception enhances perception performance by enabling autonomous vehicles to exchange complementary information. Despite its potential to revolutionize the mobile industry, challenges in various environments, such as communication bandwidth limitations, localization errors and information aggregation inefficiencies, hinder its implementation in practical applications. In this work, we propose ERMVP, a communication-Efficient and collaboration-Robust Multi-Vehicle Perception method in challenging environments. Specifically, ERMVP has three distinct strengths: i) It utilizes the hierarchical feature sampling strategy to abstract a representative set of feature vectors, using less communication overhead for efficient communication; ii) It employs the sparse consensus features to execute precise spatial location calibrations, effectively mitigating the implications of vehicle localization errors; iii) A pioneering feature fusion and interaction paradigm is introduced to integrate holistic spatial semantics among different vehicles and data sources. To thoroughly validate our method, we conduct extensive experiments on real-world and simulated datasets. The results demonstrate that the proposed ERMVP is significantly superior to the state-of-the-art collaborative perception methods. + +# 1. Introduction + +Autonomous vehicles are widely recognized as a valuable means to enhance road safety and traffic efficiency. Equipped with lidar, cameras, and other sensors, these vehicles are capable of accurately sensing their surroundings to ensure safe and reliable operation. However, the single-vehicle perception system has inevitable drawbacks [26, 39], such as a limited sensor field of view that can be easily obstructed and the challenge of detecting distant objects due to sparse and low-resolution data. Recently, the + +advances in vehicle-to-vehicle (V2V) communication technologies [12, 17, 36] and deep learning [4-6, 10, 16, 24, 45] have spurred innovation and progress in the collaborative perception technology. This technology allows connected autonomous vehicles (CAVs) to share sensory data, leading to more comprehensive environmental perception. + +Although collaborative perception technology shows great potential in transforming the mobility industry, its practical application faces several challenges, including communication bandwidth limitations [15, 35], localization errors [11, 25] and information aggregation inefficiencies [38, 47]. In practical situations, wireless communication resource and reliability constraints severely hamper the efficacy of delay-sensitive collaborative perception. While recent works [15, 42] have achieved a balance between perception performance and communication bandwidth through well-designed mechanisms, these methods have their limitations because they primarily considering information compression over spatial redundancy. This narrow focus exacerbates performance degradation at high compression ratios. + +Further, complex dynamic environments lead to localization errors, which result in inaccurate relative transform estimates and spatial feature misalignment. This relative pose noise produces misleading features that adversely affect the effectiveness of collaborative perception. Existing methods [25, 34] attempt to optimize the overall pose through intensive computation, but the high latency makes them unsuitable for real-time dynamic perception. Meanwhile, collaborative methods [3, 38, 41, 42, 46] only focus on aggregated information, but overlook the inherent perceptual strengths of ego vehicle. This paradigm is vulnerable to the perturbations introduced by collaborative noise, including asynchronous motion blur and inaccurate projections. Such drawback becomes a bottleneck for achieving optimal perceptual performance. In contrast, ego-centric features may contain locally accurate spatial location information that is not affected by collaborative noise. Therefore, a priority for establishing a pragmatic collaborative perception system is to effectively overcome the above challenges. + +![](images/17285042ad72afee503d390d9bf80d325dc5985f785ca8f0682e6d8e69e83e0c.jpg) +Figure 1. The overall architecture of the proposed framework. The framework consists of six phases: metadata sharing and feature extraction, filter and merge feature sampling, feature spatial calibration, attention-based feature fusion, accuracy enhanced feature interaction and detection decoders. The details of each individual component are illustrated in Section 3. + +Based on these observations, we propose ERMVP, a communication-efficient and collaboration-robust multi-vehicle perception method in challenging environments. From Figure 1, ERMVP uses four proposed core components to tackle existing challenges jointly. Specifically, (i) we first design an advanced filter and merge feature sampling strategy to tackle the limitations of wireless communication resources. This strategy considers both inter-class and intra-class redundancy relationships to abstract a refined set of feature vectors from the redundant features, using less communication overhead for efficient communication. (ii) Second, we introduce a plug-and-play feature spatial calibration module to mitigate the implications of vehicle localization errors. This module ingeniously utilizes consensus sparse foreground features to align the relative pose relationships between ego-vehicle and collaborators without any precise pose supervision. (iii) Furthermore, we present a pioneering feature fusion and interaction paradigm to integrate holistic spatial semantics. This paradigm comprises two key components: The first is an attention-based feature fusion module that alternates between local and global attention to fuse heterogeneous information from different vehicles. The second is an accuracy enhanced feature interaction strategy that leverages the accurate positional information inherent in ego-centric features to enhance the rich semantic information provided by fused features. Through these tailored components, ERMVP represents a significant advancement towards pragmatic collaborative perception. To validate the effectiveness of the ER-MVP, we conduct extensive experiments on two collaborative 3D object detection datasets including V2V4Real [44] + +and OPV2V [43]. Comprehensive experimental results demonstrate that our method outperforms previous state-of-the-art methods under the bandwidth-limited noisy setting. The main contributions can be summarized as follows: + +- We present ERMVP, a communication-efficient and collaboration-robust multi-vehicle perception method, which addresses the communication bandwidth limitations, localization errors and information aggregation efficiency challenges. +- We develop a filter and merge feature sampling strategy to enhance communication efficiency, a feature spatial calibration module for accurate spatial feature alignment, and two information aggregation components to optimize the fusion process. +- We conduct extensive experiments on both real-world and simulated datasets. The results show the superiority of our method and the necessity of the proposed components. + +# 2. Related Work + +# 2.1. Multi-Agent Communication + +Communication has played a pivotal role in the development of robust multi-agent systems. Early multi-agent communication [20, 29, 32] relied on predefined protocols and heuristics to regulate communication between agents. However, these fixed methods are unsuitable in complex and dynamic settings. Advances in deep learning have inspired the development of advanced strategies for information exchange and collaboration among agents. For example, MAGIC [27] used a graph attention encoder to solve the + +problem of when and to whom to send messages. TMC [49] utilized temporal smoothing to ensure efficient communication. EC-MARL [2] tackled high-dimensional continuous control and partially observable states by introducing a rapid communication protocol to resolve task dilemmas. In comparison, our research focuses on LiDAR-based collaborative 3D object detection tasks in complex driving scenarios. We propose a hierarchical feature sampling strategy to achieve efficient communication across agents. + +# 2.2. Collaborative Perception + +Factors including limited fields of view of sensors and physical obstructions in the environment can adversely impact the perception capabilities of individual agents. To address these challenges, collaborative perception within multi-agent systems has become a key technology. Utilizing newly available datasets [22, 43, 44], several novel collaborative perception methods have been proposed. For example, V2VNet [38] incorporated a graph neural network to fuse data from different agents. V2X-ViT [42] introduced a transformer architecture that combined information from vehicles and infrastructures. Where2comm [15] used feature spatial heterogeneity to reduce bandwidth utilization by transmitting sparse feature maps. CoBEVT [41] proposed the first multi-camera-based collaborative perception framework and designed the fused axis attention module to enable multi-view interactions. In this work, we propose an innovative feature calibration and interaction method that ensures robust vehicle collaboration in high-noise dynamic environments. + +# 3. Method + +This section introduces our efficient and robust collaborative perception method. Figure 1 shows the general procedure of our method, which is divided into six phases. We will detail each phase in the following introduction. + +# 3.1. Metadata Sharing and Feature Extraction + +In multi-vehicle collaboration scenarios, one of the CAVs, known as the ego vehicle, constructs a communication graph. In this graph, the ego vehicle serves as the requester, while other vehicles within its communication range act as supporters. The ego vehicle broadcasts its metadata information, including position, heading, speed, and more. Supporters, upon receiving this metadata, project their local point cloud observations into the ego vehicle's coordinate system. Each vehicle then encodes these transformed point clouds into bird's eye view (BEV) features, yielding a visual representation. Given the $i$ -th vehicle local observations $\mathbf{X}_i$ , the extracted features are represented as $\mathbf{F}_i = \Phi_{\mathrm{enc}}(\mathbf{X}_i) \in \mathbb{R}^{H \times W \times C}$ , where $\Phi_{\mathrm{enc}}(\cdot)$ denotes the PointPillar [19] encoder shared by all vehicles and $H$ , $W$ , + +and $C$ stand for the height, width, and channel of the feature map, respectively. Then, these extracted features are fed into the filter and merge feature sampling module. + +# 3.2. Filter and Merge Feature Sampling + +Previous works have utilized well-designed mechanisms such as information entropy communication selection [37] and spatial heterogeneity map [15, 35] to reduce the required transmission bandwidth. However, these methods primarily focus on the inter-class redundancy between foreground and background features, ignoring the intra-class redundancy among features, which results in sub-optimal compression. To address this gap, we introduce an advanced Filtering and Merging Feature Sampling strategy (FMS). This strategy considers both inter-class and intra-class redundancy relationships, efficiently extracting a concise and distinctive set of feature vectors from the original feature maps, thus reducing communication overheads more effectively. FMS is composed of two core components as follows. + +Filter Sampler. In object detection, foreground areas containing objects are more significant than the background areas. Therefore, we implement the idea of reducing spatial redundancy into a feature filter sampler module, aiming to preserve perceptually important yet sparse sets of feature vectors. Since explicitly learning a binary sampler is infeasible, we develop a confidence filter strategy. Initially, a detection confidence map is generated for the feature map. It reflects the perceptual importance of different spatial areas, with higher levels indicating potential object areas and lower levels typically denoting redundant background areas. For the feature map $\pmb{F}_i$ of the $i$ -th vehicle, its confidence map $C_i$ is defined as: + +$$ +\boldsymbol {C} _ {i} = \Phi_ {\text {c o n - g e n}} \left(\boldsymbol {F} _ {i}\right) \in [ 0, 1 ] ^ {H \times W}, \tag {1} +$$ + +where $\Phi_{\mathrm{con\_gen}(\cdot)}$ represents the confidence generation network with detection decoder structure. Then the confidence map is thresholded, followed by non-maximum suppression, resulting in a binary mask $B$ . Utilizing this binary mask, we proceed to preserve the sparse foreground features $\widetilde{\pmb{F}}_i = B\odot \pmb{F}_i$ . In order to cope with the dynamically changing environmental conditions and fully ensure the robustness of the system, the filtering threshold can be varied with the sensor data and the network state. We set $\alpha$ as the filter rate and the number of remaining feature vectors is $L = \alpha \times HW$ . + +Merge Sampler. Upon extracting a detailed foreground feature vector set with the filter sampler, we employ the merge sampler for additional optimization, refining similar or repetitive foreground feature vectors through weighted merging. The process is divided into three stages: information-driven feature grouping, attention-inspired feature merging, and index-based feature reconstruction. + +(a) Information-Driven Feature Grouping. Initially, a variant of the nearest neighbor clustering algorithm is applied to group the foreground feature vector set. Given a set of feature vector $\widetilde{\boldsymbol{F}}_i = [x_1,x_2,\dots ,x_L]^\top$ and cluster center $X_{c}$ , we compute the indicator $\delta_{i}$ for each feature vector. $\delta_{i}$ is calculated as the minimum feature distance minus the average pixel distance to any other cluster center vector, expressed as: + +$$ +\delta_ {i} = \min _ {j: x _ {j} \in X _ {c}} \left(\| x _ {i} - x _ {j} \| _ {2} ^ {2} - \gamma \| p (x _ {i}), p (x _ {j}) \| _ {2} ^ {2}\right), \tag {2} +$$ + +where $\delta_{i}$ denotes to which cluster the feature vector $x_{i}$ should belong. $p(\cdot)$ means getting the position of the vector and $\gamma$ is a hyperparameter. Subsequently, we can divide all the feature vectors in $\widetilde{\boldsymbol{F}}_i$ into $K$ clusters represented by $G = \{G_{1},G_{2},\dots ,G_{K}\}$ . The total number of clusters is determined by dynamic clustering ratio $\beta$ , and is calculated as $K = \beta \times L$ . + +(b) Attention-Inspired Feature Merging. A straightforward strategy for merging feature vectors is to average of each feature vector within a cluster. However, this scheme can be severely affected by outlier feature vectors. Drawing inspiration from attention mechanisms, we utilize the confidence score as a guide to quantify the significance of each feature. Thus, the merged feature vector $\tilde{x}_i$ for the $i$ th cluster $G_i$ is computed as: + +$$ +\widetilde {x} _ {i} = \frac {\sum_ {j \in G _ {i}} c _ {j} x _ {j}}{\sum_ {j \in G _ {i}} c _ {j}}, \tag {3} +$$ + +where $c_{j}$ and $x_{j}$ represent the confidence score and the original feature vector, respectively. In the end, we obtain a final set of feature vectors $\mathbf{Z}_i = \{\widetilde{x}_1,\widetilde{x}_2,\dots ,\widetilde{x}_K\}$ that is necessary for transmission. + +(c) Index-Based Feature Reconstruction. During the feature grouping and merging process, each feature vector is allocated to a cluster, and every cluster is represented by a merged vector. We maintain a record of the index correspondence between original and merged feature vectors. Utilizing this index record, the ego-vehicle ensures that the merged feature vectors are mapped to their corresponding positions, leading to the reconstruction of feature maps. + +# 3.3. Feature Spatial Calibration + +Localization errors [30, 31] may lead to the misalignment of feature maps among vehicles. Such misalignment causes the ego-vehicle to misinterpret the object's location, resulting in sub-optimal perception, as depicted in the upper right corner of Figure 2. To address this challenge, we introduce the Feature Spatial Calibration module (FSC) to facilitate precise feature alignment, as shown in Figure 2. The core idea of FSC is that overlapping viewpoints allow multiple vehicles to detect partially the same objects. The process + +![](images/7d58f4ddb6a0733b531a8909f7caf793b1c101be3a55015892f12e2ce2992a07.jpg) +Figure 2. Illustration of the proposed FSC. By employing FSC, misaligned matching regions are corrected at the feature level, leading to more accurate predictions. + +involves three phases: consensus matching, geometric verification, and error modulation. + +Consensus Matching. After reconstructing the received features, the ego vehicle generates sparse feature maps. Within these maps, areas rich in information signify potential target regions, which correspond to the proposed matching areas. We denote the proposed matching regions of the ego vehicle as $P$ and those identified by the collaborative vehicle under noisy pose conditions as $Q$ . Utilizing $P$ and $Q$ , a weighted bipartite graph is constructed, wherein the weight of each edge is determined by the distance between the nodes, encapsulated in a cost matrix. The matching process is then converted into a linear assignment task, with the objective of identifying a matching result with the lowest cumulative edge weight. This procedure yields the matching pairs $M$ . + +Geometric Verification. Invalid matches may occur due to objects located in exclusive zones and detection noise. To tackle this issue, we utilize RANSAC to filter and sift a consistent set of matches aligned with expected geometric transformations. Initially, a random matching subset $M_{s} \subseteq M$ is selected, and then the transformation matrix $\Gamma_{s}$ is computed using singular value decomposition. When $\Gamma_{s}$ is applied to all pairs within $M_{s}$ , and if the posttransformation distance between these pairs remains below the threshold $\eta$ , the set is considered correctly aligned. The threshold $\eta$ reflects the allowable localization error within the original collaborative framework. This process is iteratively conducted to identify the optimal transformation matrix that correlates with the maximum number of correct matches. Ultimately, a optimal refined transformation matrix $\Gamma_{r}$ is obtained and applied in subsequent spatial calibration operations, yielding the aligned feature $\widetilde{Z}_{j} = \Gamma_{r}Z_{j}$ . + +Error Modulation. To enhance the adaptability of the calibration method in various environments, we incorporate an error modulation strategy. This strategy aims to achieve a balance between localization errors and the estimation errors that emerge from the calibration process. It measures the overlap ratios between ego and collaborative features in both their adjusted and original states. Subsequently, fea + +![](images/d42d10cfc3c0721cb0a1f26165bbe6d235184a7bc9b1bd0e8df2217e6ae379c9.jpg) +Figure 3. A example of local and global Attention in AFF. This shows how AFF processes both 3D local windows and the global query token for position-level and context-aware aggregation. + +tures demonstrating the most optimal overlap are selected as the final input. + +# 3.4. Attention-based Feature Fusion + +In multi-vehicle collaboration scenarios, vehicles are able to capture heterogeneous information from different spatial regions. In order to efficiently fuse perceptual features from multiple vehicles, we propose an Attention-based Feature Fusion method (AFF). As shown in Figure 3, AFF utilizes alternating local and global attention to achieve position-level accurate matching in occlusion-variable traffic scenes and to capture the global semantics attention of road topology and traffic states. Specifically, we stack all the vehicles' features to $S \in \mathbb{R}^{N \times H \times W \times C}$ , where $N$ is the number of vehicles. The features are divided into 3D non-overlapping windows, each of size $N \times P \times P$ . The shape of the partition tensor is $\left(\frac{H}{P} \times \frac{W}{P}, N \times P^2, C\right)$ . We perform 3D local self attention on the local window and implement region interaction within the window. In contrast, global attention goes beyond the limitations of the local view. It uses the extracted global query token and is shared between all windows to interact with the local key and value representations, thus helping to capture long-range dependencies of the features. Inspired by the efficiency of [13, 33], the global query token $g$ is generated as follows: + +$$ +\mathbf {g} = \mathcal {P} \left(\mathcal {C} _ {1 \times 1} \left(\mathcal {S E} (\mathcal {G} (\mathcal {D} (\mathbf {x}))) + \mathbf {x}\right)\right), \tag {4} +$$ + +where $\mathcal{D}$ , $\mathcal{G}$ , $\mathcal{SE}$ , $\mathcal{C}_{1\times 1}$ and $\mathcal{P}$ denote the depth-wise separable convolution, GELU activation function, Squeeze-and-Excitation operation, $1\times 1$ convolution and max pooling, respectively. These designs enable the global query token to provide advantages such as inductive bias and modeling of inter-channel dependencies. We then construct our proposed AFF module by combining this local and global attention with typical designs of Transformers, including LayerNorm [1], MLPs [9], and skip-connections [14]. This module allows the system to analyze spatial correlations from a local view while also capturing global feature responses, ensuring efficient and precise perception in dy + +![](images/1f09b47ad9f51cbbee866c451bf65bbf30029ce7fdb5145e4bb85834b8ec315b.jpg) +Figure 4. The architecture of the proposed AFI component. + +namic, complex, and occlusion-variable traffic scenarios. Finally, we obtain the fused features $\pmb{H}_i$ . + +# 3.5. Accuracy Enhanced Feature Interaction + +Previous works [15, 21, 38, 42, 43] have demonstrated that fused features can provide richer semantic information, thereby enhancing perceptual performance. However, they may be affected by collaborative noise, such as asynchronous motion blur and inaccurate projections, which can compromise accurate position information and become a bottleneck to the optimal realization of perceptual performance. Ego-centric features may contain locally critical spatial location information without being affected by collaborative noise. To this end, we propose a novel Accuracy Enhanced Feature Interaction (AEI) strategy that leverages the accurate positional information inherent in ego-centric features to enhance the rich semantic information provided by collaborative fused features. Firstly, we design a context cross attention module that is tailored for feature interaction, as shown in Figure 4. It considers the features from the ego-vehicle as queries, and obtains the context of the query location from the fused features as the key and value to achieve cross-attention. The computed output of the spatial location $(i,j)$ is as follows: + +$$ +y _ {i j} = \sum_ {a, b \in \mathcal {N} _ {k} (i, j)} \operatorname {s o f t m a x} \left(q _ {i j} ^ {\top} k _ {a b}\right) v _ {a b}, \tag {5} +$$ + +where $\mathcal{N}_k(i,j)$ is context position of $(i,j)$ and $q_{ij}, k_{ab}, v_{ab}$ represent the ego query, context key and value, respectively. As the relevant features of the same object are often located in similar locations across different feature maps, we prioritize the local context details of the query location. Additionally, considering that only a few positions within the sparse feature map hold significant information, focusing on these context-rich areas is computationally efficient. Following this, we employ the split attention [48] to adaptively fuse information from multiple branches, generating the enhanced output feature $\widetilde{H}_i$ . + +
DatasetV2V4RealOPV2V
Noise Level σt/σr (m/°)0.0/0.00.4/0.40.8/0.80.0/0.00.4/0.40.8/0.8
No Collaboration41.64/23.4341.64/23.4341.64/23.4373.25/58.2273.25/58.2273.25/58.22
Late Fusion56.80/27.6240.81/13.3126.65/10.6686.46/79.2176.56/40.0145.57/21.00
F-Cooper [3]61.12/32.3651.71/24.4841.90/19.9187.39/79.3979.33/40.1252.91/16.90
AttFuse [7]64.41/34.3258.42/30.0249.37/25.4490.50/81.8083.54/52.4368.37/40.96
Where2comm [15]63.69/34.7958.48/28.0949.77/24.2590.46/84.2282.17/56.4773.95/45.86
DiscoNet [21]64.19/35.2557.71/28.4149.19/24.9489.58/81.4583.25/53.2261.37/30.53
V2VNet [38]65.70/35.3862.11/30.6354.73/25.5791.35/82.4385.43/54.1670.43/31.28
CoAlign [25]64.06/36.6059.58/31.5253.02/26.6090.93/84.2884.48/57.4573.95/48.86
V2X-ViT [42]66.42/37.3463.33/32.3857.15/28.2291.74/83.3182.91/54.7361.70/26.07
CoBEVT [41]66.01/37.3658.63/29.0949.22/23.6691.71/85.9885.49/60.6464.63/29.59
ERMVP (Ours)67.66/42.9765.01/40.4560.88/35.0192.18/85.5985.67/66.3277.13/59.15
+ +Table 1. Overall performance on V2V4Real and OPV2V datasets with pose noises. The results are reported in AP@0.5/0.7. + +# 3.6. Decoder and Loss + +Based on the final fused feature $\widetilde{H}_i$ , we use the detection decoder to generate the final prediction output $Y_{i} = \Phi_{\mathrm{dec}}(\widetilde{H}_{i})$ . Each position of $Y_{i}$ represents a rotated box with classes (c, x, y, h, w, $\cos \alpha$ , $\sin \alpha$ ), corresponding to class confidence, position, size, and angle. These objects are the final output of the proposed collaborative perception system. Following existing work [19], we adopt the smooth $L_{1}$ loss for regression and focal loss [23] for classification. + +# 4. Experiments + +# 4.1. Datasets and Evaluation Metrics + +Datasets. We validate the proposed ERMVP in the task of LiDAR-based 3D object detection on two benchmark datasets including V2V4real [44] and OPV2V [43]. V2V4real [44] is the first large-scale real-world dataset on V2V perception, collected by two vehicles equipped with multi-modal sensors driving together in a variety of scenarios. It covers a driving area of 410 kilometers, including 20K frames of point clouds, with training/validation/test sets containing 14210/2000/3986 frames, respectively. OPV2V [43] is a large-scale vehicle-to-vehicle collaborative perception dataset, simulated by OpenCDA [40] and Carla [8]. It contains 73 different scenarios, in which different numbers (2 to 7) of collaborative vehicles appear together, each equipped with a lidar sensor and 4 cameras. The dataset contains a total of 11,464 frames of point clouds and RGB images. The training/validation/test sets contain 6374, 1980, and 2170 frames, respectively. + +Evaluation Metrics. We adopt the Average Precision (AP) at Intersection-over-Union (IoU) thresholds of 0.5 and 0.7 to evaluate the detection performance. The calculation format of communication volume in [15] is used to count the message size by byte in the log scale with base 2. + +# 4.2. Implementation Details + +We implement the proposed ERMVP and comparison model on the Pytorch toolbox [28], and train them on a NVIDIA GeForceRTX 3090 GPU using the Adam optimizer [18]. The initial learning rate is $2\mathrm{e} - 3$ and it decays every 15 epochs with a factor of 0.1. All models are trained with 60 epochs and the batch size is set to 2. Early stopping is used to find the best epoch. We also add normal point cloud data augmentation for all experiments, including scaling, rotation, and flipping. All detection models are based on PointPillar [19] backbone to extract 2D features from point clouds and a $0.4\mathrm{m}$ width/length is used for each voxel. AFF component has 3 encoded layers and a window size of 8 for both local and global attention. The balance hyperparameter $\gamma$ is 0.1 and the error tolerance threshold $\eta$ is 0.25. Each vehicle has a communication range of 70 m based on [42], while vehicles outside of this broadcasting radius will be ignored. To simulate the localization and heading errors, we add Gaussian noise with a standard deviation of $\sigma_t$ for localization errors and $\sigma_r$ for heading errors. + +# 4.3.Quantitative Evaluation + +Comparison of Detection Performance. Table 1 shows the performance comparison results of 3D detection on two datasets. We use the No Collaboration method as a baseline, which only uses ego vehicle's point clouds data without collaboration. Late Fusion allows vehicles to exchange detected outputs and utilizes non-maximum suppression to produce the final result. For intermediate fusion strategies, we evaluate the existing state-of-the-art (SOTA) methods: AttFuse [43], F-Cooper [3], V2VNet [38], DiscoNet [21], V2X-ViT [42], CoAlign (only fusion) [25], CoBEVT [41] and Where2comm [15]. The proposed ER-MVP achieves average improvements of $12.48\% / 18.47\%$ and $6.18\% / 10.02\%$ on two datasets in AP@0.5/0.7 compared with No Collaboration and Late Fusion methods, demonstrating the superiority of our collaborative method. + +(a) V2V4Real +![](images/a05276fe44b9498839d47f184cc6c82d59fe0ebf90f874dbf5bd3a548c86c415.jpg) +Late Fusion FCooper V2VNet V2X-VIT CoBEVT -Where2comm -ERMVP + +![](images/0d727aed98d8c717ad512713bc262dc41c0e12fc4ed41f3bfc943faddb1c8b02.jpg) +Figure 5. Collaborative perception performance comparison on the V2V4Real and OPV2V with varying communication volume. + +![](images/f88bea390d557ccfa8148f6e1357029b30be1ff3a78f8f3b8d9d1900be3dafa5.jpg) + +![](images/517d86dc9d3835c14d405265d54458bd70a95d2611f426bab979bfa02ba7bde2.jpg) + +Meanwhile, it is superior to the SOTA collaborative perception method in both real-world and simulated scenarios: improves the SOTA collaborative performance by $1.24\%$ and $5.61\%$ on V2V4Real datasets in AP@0.5/0.7, and by $0.44\%$ on OPV2V datasets in AP@0.5. These results fully demonstrate the superiority of the ERMVP collaboration paradigm. + +
DatasetV2V4RealOPV2V
Noise Level σt(m)0.51.00.51.0
Late Fusion11.0210.1353.8838.80
F-Cooper [3]24.6018.4132.2412.71
AttFuse [43]28.1024.3642.9535.53
DiscoNet [21]28.3624.0545.3624.96
V2VNet [38]29.5723.6747.4825.82
Where2comm [15]28.0223.5346.5945.59
CoAlign [21]29.8425.8249.0946.66
V2X-ViT [42]30.9923.6744.8321.77
CoBEVT [41]24.6722.0141.9126.05
ERMVP (Ours)39.2633.9477.2576.09
+ +Robustness to Localization and Heading Errors. To evaluate the sensitivity of existing methods to localization and heading errors, we use the same noise setting as [42] and conduct extensive experiments on two datasets. As shown in Tables 1 and 2, under ideal settings, Late Fusion and some advanced intermediate fusion methods can detect object vehicles with high accuracy, but their detection accuracy rapidly decreases as the standard deviation of errors increases. For example, when the localization and heading errors of OPV2V dataset are over $0.4\mathrm{m}$ and $0.4^{\circ}$ , the Disconet [21] and F-Cooper [3] methods fail and perform even worse than No Collaboration. Our method exceeds previous SOTA models at all noise levels and consistently outperforms the No Collaboration baseline, clearly demonstrating the robustness of ERMVP to pose errors. It is noteworthy that for localization errors, ERMVP can maintain consistent + +accuracy levels in high noise environments, showing superior robustness. When the localization error is $1.0\mathrm{m}$ , it outperforms the second-best method by $29.43\%$ and $8.08\%$ in AP@0.7 on OPV2V and V2V4Real datasets. The reasonable explanations are: (i) feature spatial calibration aligns mismatched collaborative features; (ii) accuracy enhanced feature interaction reduces performance degradation caused by collaborative noise. + +Comparison of Communication Volume. To achieve pragmatic collaborative perception, it is crucial to evaluate the perceptual performance under different communication volumes. Figure 5 shows the results of the performance comparison under different bandwidth consumption conditions. Note that we do not consider any additional model/data/feature compression for a fair comparison. The results indicate that ERMVP achieves excellent perceptual performance and bandwidth consumption trade-off under all communication bandwidth conditions, consistently outperforming Where2comm [15]. At the same time, it achieves the same detection performance as the previous model [3, 38, 41, 42] with less communication volume on both real-world and simulated datasets. + +# 4.4. Ablation Studies + +Table 2. Detection AP@0.7 on V2V4Real and OPV2V datasets with localization error. + +
AFFFSCAEIV2V4RealOPV2V
51.71/24.4879.33/40.12
✓62.24/37.7084.30/61.43
✓✓64.36/39.5984.98/62.67
✓✓✓65.01/40.4585.67/66.32
+ +Table 3. Ablation study results of the proposed core components on the both datasets with noise level of 0.4/0.4. AFF: Attention-based Feature Fusion; FSC: Feature Spatial Calibration; AEI: Accuracy Enhanced Feature Interaction. + +Effect of Core Components. Table 3 details the contribution of each core components in our ERMVP framework. The base model is the naive position-wise maximum fusion method. We then assess the impact of each component by sequentially introducing i) attention-based feature + +![](images/280d2b5c3bb6552ed83f6681ca5e9ab65f17fefd5ae8bbd0b8c2945e3095852f.jpg) + +![](images/e770195a486e1bf921acc083d6a82eecd9c532e4557f8586c9120e99f4b1c3dd.jpg) +(a) V2VNet + +![](images/efd2f53dc9a564fb9dde0a5412ab504de41b34a537d009f58505a6de2f4afd40.jpg) + +![](images/492842c4ba1259fbc06256306e088f9a970ff6743f2cbd8d52a3c6f44fd3ba2c.jpg) +(b) CoBEVT + +![](images/aefb2e149c7c44ff1a1486876c71877e2004bf57de183a250fdfb7e67f496884.jpg) + +![](images/f4481015af92b97fb51c9f6721b939b1b571c6f3d179fb796d81053fc607f9bc.jpg) +(c) V2X-ViT +Figure 6. Qualitative comparison results in real-world scenarios from the V2V4Real dataset with noise level of 0.4/0.4. Green and red boxes denote the ground truths and detection results, respectively. + +![](images/bba47526b06f4a957dc1c460977993b17132c871bd63b9b4bf3549c7e3f00d7f.jpg) + +![](images/cdb2cc3bba2d1bd3841d0f2755d0b7379f1b9c64db023463ded86abc01c94515.jpg) +(d)ERMVP + +fusion (AFF), ii) feature spatial calibration (FSC), iii) accuracy enhanced feature interaction (AEF). The consistent rise in detection results over both datasets demonstrates the effectiveness of each introduced component. Notably, integrating all three components boosts detection performance by $13.3\%$ and $15.97\%$ on the V2V4Real dataset for AP@0.5 and AP@0.7, respectively. + +![](images/b92763145d0f8cd9c12d9dd5457a2982c98ae21f38860f7e31289167e94214e1.jpg) +Figure 7. Hyperparameter analysis of $\alpha$ and $\beta$ + +![](images/d71922f3b4b2034784cf84180fa4aae8983c8e9ae7ee30960da71fb8f0ab96ef.jpg) + +Hyperparameters $\alpha$ and $\beta$ . We explore the impact of the hyperparameters $\alpha$ and $\beta$ (as described in Section 3.2), which jointly determine the communication volume. Specifically, the feature compression ratio is represented as $\alpha \times \beta$ . The communication volume, in turn, is expressed as the product of the original feature size and the compression rate. By modulating the values of $\alpha$ and $\beta$ within the interval of $10\%$ to $80\%$ , we chart the resultant object detection performance (measured by AP@0.7) in Figure 7. As expected, a drop in either $\alpha$ or $\beta$ will impair ERMVP's efficacy. Yet, it's worth noting that ERMVP's performance generally exhibits robustness to these hyperparameters. The performance decline remains fairly minimal even under significant compression, such as a factor of 100. The results demonstrate the effectiveness of our proposed filter and merge feature sampling module. Therefore, ERMVP establishes an appropriate and effective communication paradigm among vehicles, ensuring an optimal tradeoff between perception and communication. + +# 4.5. Qualitative Evaluation + +Figure 6 displays the detection results of V2VNet [38], CoBEVT [41], V2X-ViT [42], and ERMVP in two scenarios. Clearly, ERMVP offers more precise and comprehensive detection than the prior SOTA methods. To begin with, ERMVP produces a greater number of predicted bounding boxes that are well aligned with the ground truths, while other methods show significant discrepancies. This demonstrates ERMVP's robustness, particularly in high-noise environments. Moreover, ERMVP detects more dynamic objects (more ground truth bounding boxes find matches). This suggests that ERMVP can effectively combine the inputs from nearby vehicles, leading to a thorough scene representation. Overall, these qualitative assessments confirm ERMVP's strengths in delivering accurate and comprehensive perception, especially in challenging conditions. + +# 5. Conclusion + +In this paper, we present ERMVP, a communication-efficient and collaboration-robust multi-vehicle perception method in challenging environments. We introduce a filter and merge feature sampling strategy to efficient communication, a feature spatial calibration module to align features in fine-grid and two spatial information aggregation components to solve the bottleneck of the fusion method. Extensive experiments prove the superiority of ERMVP and the effectiveness of our components. The feasibility of our approach for other modalities will be explored in future work. + +# Acknowledgments + +This work is supported by the Shanghai Key Research Laboratory of NSAI, the Innovation Platform for Academicians of Hainan Province, Haikou, 570228, the Specific Research Fund of the Innovation Platform for Academicians of Hainan Province under Grant YSPTZX202314, and NFSC under grant 62250410368. + +# References + +[1] Jimmy Lei Ba, Jamie Ryan Kiros, and Geoffrey E Hinton. Layer normalization. arXiv preprint arXiv:1607.06450, 2016. 5 +[2] Marwa Chafii, Salmane Naoumi, Reda Alami, Ebtesam Al-mazrouei, Mehdi Bennis, and Merouane Debbah. Emergent communication in multi-agent reinforcement learning for future wireless networks. arXiv preprint arXiv:2309.06021, 2023. 3 +[3] Qi Chen, Xu Ma, Sihai Tang, Jingda Guo, Qing Yang, and Song Fu. F-cooper: Feature based cooperative perception for autonomous vehicle edge computing system using 3d point clouds. In Proceedings of the 4th ACM/IEEE Symposium on Edge Computing, pages 88-100, 2019. 1, 6, 7 +[4] Zhaoyu Chen, Bo Li, Shuang Wu, Jianghe Xu, Shouhong Ding, and Wenqiang Zhang. Shape matters: deformable patch attack. In European conference on computer vision, pages 529-548. Springer, 2022. 1 +[5] Zhaoyu Chen, Bo Li, Shuang Wu, Shouhong Ding, and Wenqiang Zhang. Query-efficient decision-based black-box patch attack. IEEE Transactions on Information Forensics and Security, 2023. +[6] Zhaoyu Chen, Bo Li, Shuang Wu, Kaixun Jiang, Shouhong Ding, and Wenqiang Zhang. Content-based unrestricted adversarial attack. Advances in Neural Information Processing Systems, 36, 2024. 1 +[7] Yimian Dai, Fabian Gieseke, Stefan Oehmcke, Yiquan Wu, and Kobus Barnard. Attentional feature fusion. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 3560-3569, 2021. 6 +[8] Alexey Dosovitskiy, German Ros, Felipe Codevilla, Antonio Lopez, and Vladlen Koltun. Carla: An open urban driving simulator. In Conference on robot learning, pages 1-16. PMLR, 2017. 6 +[9] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 5 +[10] Ian Goodfellow, Yoshua Bengio, and Aaron Courville. Deep learning. MIT press, 2016. 1 +[11] Jiaming Gu, Jingyu Zhang, Muyang Zhang, Weiliang Meng, Shibiao Xu, Jiguang Zhang, and Xiaopeng Zhang. Feaco: Reaching robust feature-level consensus in noisy pose conditions. In Proceedings of the 31st ACM International Conference on Multimedia, page 3628-3636, New York, NY, USA, 2023. Association for Computing Machinery. 1 +[12] Monowar Hasan, Sibin Mohan, Takayuki Shimizu, and Hongsheng Lu. Securing vehicle-to-everything (v2x) communication platforms. IEEE Transactions on Intelligent Vehicles, 5(4):693-713, 2020. 1 +[13] Ali Hatamizadeh, Hongxu Yin, Greg Heinrich, Jan Kautz, and Pavlo Molchanov. Global context vision transformers. In International Conference on Machine Learning, pages 12633-12646. PMLR, 2023. 5 + +[14] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 5 +[15] Yue Hu, Shaoheng Fang and Zixing Lei, Yiqi Zhong, and Siheng Chen. Where2comm: Communication-efficient collaborative perception via spatial confidence maps. In Thirty-sixth Conference on Neural Information Processing Systems (NeurIPS), 2022. 1, 3, 5, 6, 7 +[16] Kaixun Jiang, Zhaoyu Chen, Hao Huang, Jiafeng Wang, Dingkang Yang, Bo Li, Yan Wang, and Wenqiang Zhang. Efficient decision-based black-box patch attacks on video recognition. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 4379-4389, 2023. 1 +[17] Hamza Khan, Sumudu Samarakoon, and Mehdi Bennis. Enhancing video streaming in vehicular networks via resource slicing. IEEE Transactions on Vehicular Technology, 69(4): 3513-3522, 2020. 1 +[18] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In International Conference on Learning Representations (ICLR), 2015. 6 +[19] Alex H Lang, Sourabh Vora, Holger Caesar, Lubing Zhou, Jiong Yang, and Oscar Beijbom. Pointpillars: Fast encoders for object detection from point clouds. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12697-12705, 2019. 3, 6 +[20] Yiming Li, Bir Bhanu, and Wei Lin. Auction protocol for camera active control. In 2010 IEEE International Conference on Image Processing, pages 4325-4328. IEEE, 2010. 2 +[21] Yiming Li, Shunli Ren, Pengxiang Wu, Siheng Chen, Chen Feng, and Wenjun Zhang. Learning distilled collaboration graph for multi-agent perception. Advances in Neural Information Processing Systems, 34:29541-29552, 2021. 5, 6, 7 +[22] Yiming Li, Ziyan An, Zixun Wang, Yiqi Zhong, Siheng Chen, and Chen Feng. V2x-sim: A virtual collaborative perception dataset for autonomous driving. arXiv preprint arXiv:2202.08449, 2022. 3 +[23] Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dollar. Focal loss for dense object detection. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), 2017. 6 +[24] Yang Liu, Jing Liu, Kun Yang, Bobo Ju, Siao Liu, Yuzheng Wang, Dingkang Yang, Peng Sun, and Liang Song. Ampnet: Appearance-motion prototype network assisted automatic video anomaly detection system. IEEE Transactions on Industrial Informatics, 20(2):2843-2855, 2024. 1 +[25] Yifan Lu, Quanhao Li, Baoan Liu, Mehrdad Dianati, Chen Feng, Siheng Chen, and Yanfeng Wang. Robust collaborative 3d object detection in presence of pose errors. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 4812-4818. IEEE, 2023. 1, 6 +[26] Zonglin Meng, Xin Xia, Runsheng Xu, Wei Liu, and Jiaqi Ma. Hydro-3d: Hybrid object detection and tracking for cooperative perception using 3d lidar. IEEE Transactions on Intelligent Vehicles, 2023. 1 + +[27] Yaru Niu, Rohan R Paleja, and Matthew C Gombolay. Multiagent graph-attention communication and teaming. In AA-MAS, pages 964–973, 2021. 2 +[28] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. Advances in neural information processing systems, 32, 2019. 6 +[29] Faisal Qureshi and Demetri Terzopoulos. Smart camera networks in virtual reality. Proceedings of the IEEE, 96(10): 1640-1656, 2008. 2 +[30] Qingyu Song, Changan Wang, Zhengkai Jiang, Yabiao Wang, Ying Tai, Chengjie Wang, Jilin Li, Feiyue Huang, and Yang Wu. Rethinking counting and localization in crowds: A purely point-based framework. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3365-3374, 2021. 4 +[31] Zhiying Song, Fuxi Wen, Hailiang Zhang, and Jun Li. An efficient and robust object-level cooperative perception framework for connected and automated driving. arXiv preprint arXiv:2210.06289, 2022. 4 +[32] Ming Tan. Multi-agent reinforcement learning: Independent vs. cooperative agents. In Proceedings of the tenth international conference on machine learning, pages 330-337, 1993. 2 +[33] Mingxing Tan and Quoc Le. Efficientnetv2: Smaller models and faster training. In International conference on machine learning, pages 10096-10106. PMLR, 2021. 5 +[34] Nicholas Vadivelu, Mengye Ren, James Tu, Jingkang Wang, and Raquel Urtasun. Learning to communicate and correct pose errors. In Conference on Robot Learning, pages 1195-1210. PMLR, 2021. 1 +[35] Binglu Wang, Lei Zhang, Zhaozhong Wang, Yongqiang Zhao, and Tianfei Zhou. Core: Cooperative reconstruction for multi-agent perception. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8710-8720, 2023. 1, 3 +[36] Jian Wang, Yameng Shao, Yuming Ge, and Rundong Yu. A survey of vehicle to everything (v2x) testing. Sensors, 19(2): 334, 2019. 1 +[37] Tianhang Wang, Guang Chen, Kai Chen, Zhengfa Liu, Bo Zhang, Alois Knoll, and Changjun Jiang. Umc: A unified bandwidth-efficient and multi-resolution based collaborative perception framework. arXiv preprint arXiv:2303.12400, 2023. 3 +[38] Tsun-Hsuan Wang, Sivabanan Manivasagam, Ming Liang, Bin Yang, Wenyuan Zeng, and Raquel Urtasun. V2vnet: Vehicle-to-vehicle communication for joint perception and prediction. In European Conference on Computer Vision, pages 605–621. Springer, 2020. 1, 3, 5, 6, 7, 8 +[39] Qianyi Wu, Kaisiyuan Wang, Kejie Li, Jianmin Zheng, and Jianfei Cai. Objectsdf++: Improved object-compositional neural implicit surfaces. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 21764-21774, 2023. 1 + +[40] Runsheng Xu, Yi Guo, Xu Han, Xin Xia, Hao Xiang, and Jiaqi Ma. Opencda: an open cooperative driving automation framework integrated with co-simulation. In 2021 IEEE International Intelligent Transportation Systems Conference (ITSC), pages 1155-1162. IEEE, 2021. 6 +[41] Runsheng Xu, Zhengzhong Tu, Hao Xiang, Wei Shao, Bolei Zhou, and Jiaqi Ma. Cobevt: Cooperative bird's eye view semantic segmentation with sparse transformers, 2022. 1, 3, 6, 7, 8 +[42] Runsheng Xu, Hao Xiang, Zhengzhong Tu, Xin Xia, Ming-Hsuan Yang, and Jiaqi Ma. V2x-vit: Vehicle-to-everything cooperative perception with vision transformer. In European Conference on Computer Vision, page 107-124. Springer, 2022. 1, 3, 5, 6, 7, 8 +[43] Runsheng Xu, Hao Xiang, Xin Xia, Xu Han, Jinlong Li, and Jiaqi Ma. Opv2v: An open benchmark dataset and fusion pipeline for perception with vehicle-to-vehicle communication. In 2022 International Conference on Robotics and Automation (ICRA), pages 2583–2589. IEEE, 2022. 2, 3, 5, 6, 7 +[44] Runsheng Xu, Xin Xia, Jinlong Li, Hanzhao Li, Shuo Zhang, Zhengzhong Tu, Zonglin Meng, Hao Xiang, Xiaoyu Dong, Rui Song, et al. V2v4real: A real-world large-scale dataset for vehicle-to-vehicle cooperative perception. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13712–13722, 2023. 2, 3, 6 +[45] Kun Yang, Peng Sun, Jieyu Lin, Azzedine Boukerche, and Liang Song. A novel distributed task scheduling framework for supporting vehicular edge intelligence. In 2022 IEEE 42nd International Conference on Distributed Computing Systems (ICDCS), pages 972-982. IEEE, 2022. 1 +[46] Kun Yang, Dingkang Yang, Jingyu Zhang, Mingcheng Li, Yang Liu, Jing Liu, Hanqi Wang, Peng Sun, and Liang Song. Spatio-temporal domain awareness for multi-agent collaborative perception, 2023. 1 +[47] Kun Yang, Dingkang Yang, Jingyu Zhang, Hanqi Wang, Peng Sun, and Liang Song. What2comm: Towards communication-efficient collaborative perception via feature decoupling. In Proceedings of the 31st ACM International Conference on Multimedia, page 7686-7695, New York, NY, USA, 2023. Association for Computing Machinery. 1 +[48] Hang Zhang, Chongruo Wu, Zhongyue Zhang, Yi Zhu, Haibin Lin, Zhi Zhang, Yue Sun, Tong He, Jonas Mueller, R. Manmatha, Mu Li, and Alexander Smola. Resnest: Split-attention networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, pages 2736-2746, 2022. 5 +[49] Sai Qian Zhang, Qi Zhang, and Jieyu Lin. Succinct and robust multi-agent communication with temporal message control. Advances in Neural Information Processing Systems, 33:17271-17282, 2020. 3 \ No newline at end of file diff --git a/2024/ERMVP_ Communication-Efficient and Collaboration-Robust Multi-Vehicle Perception in Challenging Environments/images.zip b/2024/ERMVP_ Communication-Efficient and Collaboration-Robust Multi-Vehicle Perception in Challenging Environments/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..ef6baeed8461ae49155ce6993ceacf9f4b8ecd37 --- /dev/null +++ b/2024/ERMVP_ Communication-Efficient and Collaboration-Robust Multi-Vehicle Perception in Challenging Environments/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e939236d3cfabb74b721c68771c8ca4ad916b2610319aef2fd107ebc229a78fe +size 524448 diff --git a/2024/ERMVP_ Communication-Efficient and Collaboration-Robust Multi-Vehicle Perception in Challenging Environments/layout.json b/2024/ERMVP_ Communication-Efficient and Collaboration-Robust Multi-Vehicle Perception in Challenging Environments/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..d3aca316747d1237e6c0121c05d188bfe57e931e --- /dev/null +++ b/2024/ERMVP_ Communication-Efficient and Collaboration-Robust Multi-Vehicle Perception in Challenging Environments/layout.json @@ -0,0 +1,8480 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 60, + 136, + 517, + 171 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 60, + 136, + 517, + 171 + ], + "spans": [ + { + "bbox": [ + 60, + 136, + 517, + 171 + ], + "type": "text", + "content": "ERMVP: Communication-Efficient and Collaboration-Robust Multi-Vehicle Perception in Challenging Environments" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 78, + 191, + 498, + 234 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 78, + 191, + 498, + 234 + ], + "spans": [ + { + "bbox": [ + 78, + 191, + 498, + 234 + ], + "type": "text", + "content": "Jingyu Zhang" + }, + { + "bbox": [ + 78, + 191, + 498, + 234 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 78, + 191, + 498, + 234 + ], + "type": "text", + "content": " Kun Yang" + }, + { + "bbox": [ + 78, + 191, + 498, + 234 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 78, + 191, + 498, + 234 + ], + "type": "text", + "content": " Yilei Wang" + }, + { + "bbox": [ + 78, + 191, + 498, + 234 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 78, + 191, + 498, + 234 + ], + "type": "text", + "content": " Hanqi Wang" + }, + { + "bbox": [ + 78, + 191, + 498, + 234 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 78, + 191, + 498, + 234 + ], + "type": "text", + "content": " Peng Sun" + }, + { + "bbox": [ + 78, + 191, + 498, + 234 + ], + "type": "inline_equation", + "content": "^{2,*}" + }, + { + "bbox": [ + 78, + 191, + 498, + 234 + ], + "type": "text", + "content": " Liang Song" + }, + { + "bbox": [ + 78, + 191, + 498, + 234 + ], + "type": "inline_equation", + "content": "^{1,*}" + }, + { + "bbox": [ + 78, + 191, + 498, + 234 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 78, + 191, + 498, + 234 + ], + "type": "text", + "content": "Academy for Engineering and Technology, Fudan University " + }, + { + "bbox": [ + 78, + 191, + 498, + 234 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 78, + 191, + 498, + 234 + ], + "type": "text", + "content": "Duke Kunshan University \n{jingyuzhang22, yileiwang23}@m.fudan.edu.cn, songl@m.fudan.edu.cn" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 139, + 260, + 186, + 273 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 260, + 186, + 273 + ], + "spans": [ + { + "bbox": [ + 139, + 260, + 186, + 273 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 45, + 285, + 281, + 554 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 285, + 281, + 554 + ], + "spans": [ + { + "bbox": [ + 45, + 285, + 281, + 554 + ], + "type": "text", + "content": "Collaborative perception enhances perception performance by enabling autonomous vehicles to exchange complementary information. Despite its potential to revolutionize the mobile industry, challenges in various environments, such as communication bandwidth limitations, localization errors and information aggregation inefficiencies, hinder its implementation in practical applications. In this work, we propose ERMVP, a communication-Efficient and collaboration-Robust Multi-Vehicle Perception method in challenging environments. Specifically, ERMVP has three distinct strengths: i) It utilizes the hierarchical feature sampling strategy to abstract a representative set of feature vectors, using less communication overhead for efficient communication; ii) It employs the sparse consensus features to execute precise spatial location calibrations, effectively mitigating the implications of vehicle localization errors; iii) A pioneering feature fusion and interaction paradigm is introduced to integrate holistic spatial semantics among different vehicles and data sources. To thoroughly validate our method, we conduct extensive experiments on real-world and simulated datasets. The results demonstrate that the proposed ERMVP is significantly superior to the state-of-the-art collaborative perception methods." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 576, + 124, + 588 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 576, + 124, + 588 + ], + "spans": [ + { + "bbox": [ + 46, + 576, + 124, + 588 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 45, + 596, + 279, + 703 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 596, + 279, + 703 + ], + "spans": [ + { + "bbox": [ + 45, + 596, + 279, + 703 + ], + "type": "text", + "content": "Autonomous vehicles are widely recognized as a valuable means to enhance road safety and traffic efficiency. Equipped with lidar, cameras, and other sensors, these vehicles are capable of accurately sensing their surroundings to ensure safe and reliable operation. However, the single-vehicle perception system has inevitable drawbacks [26, 39], such as a limited sensor field of view that can be easily obstructed and the challenge of detecting distant objects due to sparse and low-resolution data. Recently, the" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 297, + 261, + 530, + 332 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 261, + 530, + 332 + ], + "spans": [ + { + "bbox": [ + 297, + 261, + 530, + 332 + ], + "type": "text", + "content": "advances in vehicle-to-vehicle (V2V) communication technologies [12, 17, 36] and deep learning [4-6, 10, 16, 24, 45] have spurred innovation and progress in the collaborative perception technology. This technology allows connected autonomous vehicles (CAVs) to share sensory data, leading to more comprehensive environmental perception." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 296, + 332, + 531, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 332, + 531, + 506 + ], + "spans": [ + { + "bbox": [ + 296, + 332, + 531, + 506 + ], + "type": "text", + "content": "Although collaborative perception technology shows great potential in transforming the mobility industry, its practical application faces several challenges, including communication bandwidth limitations [15, 35], localization errors [11, 25] and information aggregation inefficiencies [38, 47]. In practical situations, wireless communication resource and reliability constraints severely hamper the efficacy of delay-sensitive collaborative perception. While recent works [15, 42] have achieved a balance between perception performance and communication bandwidth through well-designed mechanisms, these methods have their limitations because they primarily considering information compression over spatial redundancy. This narrow focus exacerbates performance degradation at high compression ratios." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 296, + 508, + 532, + 730 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 508, + 532, + 730 + ], + "spans": [ + { + "bbox": [ + 296, + 508, + 532, + 730 + ], + "type": "text", + "content": "Further, complex dynamic environments lead to localization errors, which result in inaccurate relative transform estimates and spatial feature misalignment. This relative pose noise produces misleading features that adversely affect the effectiveness of collaborative perception. Existing methods [25, 34] attempt to optimize the overall pose through intensive computation, but the high latency makes them unsuitable for real-time dynamic perception. Meanwhile, collaborative methods [3, 38, 41, 42, 46] only focus on aggregated information, but overlook the inherent perceptual strengths of ego vehicle. This paradigm is vulnerable to the perturbations introduced by collaborative noise, including asynchronous motion blur and inaccurate projections. Such drawback becomes a bottleneck for achieving optimal perceptual performance. In contrast, ego-centric features may contain locally accurate spatial location information that is not affected by collaborative noise. Therefore, a priority for establishing a pragmatic collaborative perception system is to effectively overcome the above challenges." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 76, + 42, + 119, + 72 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 42, + 119, + 72 + ], + "spans": [ + { + "bbox": [ + 76, + 42, + 119, + 72 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 150, + 41, + 471, + 74 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 150, + 41, + 471, + 74 + ], + "spans": [ + { + "bbox": [ + 150, + 41, + 471, + 74 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 45, + 709, + 279, + 730 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 709, + 279, + 730 + ], + "spans": [ + { + "bbox": [ + 45, + 709, + 279, + 730 + ], + "type": "text", + "content": "*Corresponding authors. Our code is available at https://github.com/Terry9a/ERMVP." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 277, + 747, + 301, + 756 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 277, + 747, + 301, + 756 + ], + "spans": [ + { + "bbox": [ + 277, + 747, + 301, + 756 + ], + "type": "text", + "content": "12575" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 595, + 842 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 46, + 103, + 530, + 313 + ], + "blocks": [ + { + "bbox": [ + 46, + 103, + 530, + 313 + ], + "lines": [ + { + "bbox": [ + 46, + 103, + 530, + 313 + ], + "spans": [ + { + "bbox": [ + 46, + 103, + 530, + 313 + ], + "type": "image", + "image_path": "17285042ad72afee503d390d9bf80d325dc5985f785ca8f0682e6d8e69e83e0c.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 45, + 321, + 531, + 354 + ], + "lines": [ + { + "bbox": [ + 45, + 321, + 531, + 354 + ], + "spans": [ + { + "bbox": [ + 45, + 321, + 531, + 354 + ], + "type": "text", + "content": "Figure 1. The overall architecture of the proposed framework. The framework consists of six phases: metadata sharing and feature extraction, filter and merge feature sampling, feature spatial calibration, attention-based feature fusion, accuracy enhanced feature interaction and detection decoders. The details of each individual component are illustrated in Section 3." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 45, + 367, + 281, + 729 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 367, + 281, + 729 + ], + "spans": [ + { + "bbox": [ + 45, + 367, + 281, + 729 + ], + "type": "text", + "content": "Based on these observations, we propose ERMVP, a communication-efficient and collaboration-robust multi-vehicle perception method in challenging environments. From Figure 1, ERMVP uses four proposed core components to tackle existing challenges jointly. Specifically, (i) we first design an advanced filter and merge feature sampling strategy to tackle the limitations of wireless communication resources. This strategy considers both inter-class and intra-class redundancy relationships to abstract a refined set of feature vectors from the redundant features, using less communication overhead for efficient communication. (ii) Second, we introduce a plug-and-play feature spatial calibration module to mitigate the implications of vehicle localization errors. This module ingeniously utilizes consensus sparse foreground features to align the relative pose relationships between ego-vehicle and collaborators without any precise pose supervision. (iii) Furthermore, we present a pioneering feature fusion and interaction paradigm to integrate holistic spatial semantics. This paradigm comprises two key components: The first is an attention-based feature fusion module that alternates between local and global attention to fuse heterogeneous information from different vehicles. The second is an accuracy enhanced feature interaction strategy that leverages the accurate positional information inherent in ego-centric features to enhance the rich semantic information provided by fused features. Through these tailored components, ERMVP represents a significant advancement towards pragmatic collaborative perception. To validate the effectiveness of the ER-MVP, we conduct extensive experiments on two collaborative 3D object detection datasets including V2V4Real [44]" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 296, + 367, + 530, + 414 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 367, + 530, + 414 + ], + "spans": [ + { + "bbox": [ + 296, + 367, + 530, + 414 + ], + "type": "text", + "content": "and OPV2V [43]. Comprehensive experimental results demonstrate that our method outperforms previous state-of-the-art methods under the bandwidth-limited noisy setting. The main contributions can be summarized as follows:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 297, + 414, + 530, + 575 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 297, + 414, + 530, + 471 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 414, + 530, + 471 + ], + "spans": [ + { + "bbox": [ + 297, + 414, + 530, + 471 + ], + "type": "text", + "content": "- We present ERMVP, a communication-efficient and collaboration-robust multi-vehicle perception method, which addresses the communication bandwidth limitations, localization errors and information aggregation efficiency challenges." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 297, + 472, + 530, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 472, + 530, + 529 + ], + "spans": [ + { + "bbox": [ + 297, + 472, + 530, + 529 + ], + "type": "text", + "content": "- We develop a filter and merge feature sampling strategy to enhance communication efficiency, a feature spatial calibration module for accurate spatial feature alignment, and two information aggregation components to optimize the fusion process." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 297, + 530, + 530, + 575 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 530, + 530, + 575 + ], + "spans": [ + { + "bbox": [ + 297, + 530, + 530, + 575 + ], + "type": "text", + "content": "- We conduct extensive experiments on both real-world and simulated datasets. The results show the superiority of our method and the necessity of the proposed components." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 297, + 587, + 381, + 599 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 587, + 381, + 599 + ], + "spans": [ + { + "bbox": [ + 297, + 587, + 381, + 599 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 297, + 607, + 451, + 618 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 607, + 451, + 618 + ], + "spans": [ + { + "bbox": [ + 297, + 607, + 451, + 618 + ], + "type": "text", + "content": "2.1. Multi-Agent Communication" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 296, + 624, + 530, + 730 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 624, + 530, + 730 + ], + "spans": [ + { + "bbox": [ + 296, + 624, + 530, + 730 + ], + "type": "text", + "content": "Communication has played a pivotal role in the development of robust multi-agent systems. Early multi-agent communication [20, 29, 32] relied on predefined protocols and heuristics to regulate communication between agents. However, these fixed methods are unsuitable in complex and dynamic settings. Advances in deep learning have inspired the development of advanced strategies for information exchange and collaboration among agents. For example, MAGIC [27] used a graph attention encoder to solve the" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 277, + 747, + 301, + 756 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 277, + 747, + 301, + 756 + ], + "spans": [ + { + "bbox": [ + 277, + 747, + 301, + 756 + ], + "type": "text", + "content": "12576" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 595, + 842 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 45, + 106, + 281, + 212 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 106, + 281, + 212 + ], + "spans": [ + { + "bbox": [ + 45, + 106, + 281, + 212 + ], + "type": "text", + "content": "problem of when and to whom to send messages. TMC [49] utilized temporal smoothing to ensure efficient communication. EC-MARL [2] tackled high-dimensional continuous control and partially observable states by introducing a rapid communication protocol to resolve task dilemmas. In comparison, our research focuses on LiDAR-based collaborative 3D object detection tasks in complex driving scenarios. We propose a hierarchical feature sampling strategy to achieve efficient communication across agents." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 220, + 182, + 232 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 220, + 182, + 232 + ], + "spans": [ + { + "bbox": [ + 46, + 220, + 182, + 232 + ], + "type": "text", + "content": "2.2. Collaborative Perception" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 45, + 238, + 281, + 459 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 238, + 281, + 459 + ], + "spans": [ + { + "bbox": [ + 45, + 238, + 281, + 459 + ], + "type": "text", + "content": "Factors including limited fields of view of sensors and physical obstructions in the environment can adversely impact the perception capabilities of individual agents. To address these challenges, collaborative perception within multi-agent systems has become a key technology. Utilizing newly available datasets [22, 43, 44], several novel collaborative perception methods have been proposed. For example, V2VNet [38] incorporated a graph neural network to fuse data from different agents. V2X-ViT [42] introduced a transformer architecture that combined information from vehicles and infrastructures. Where2comm [15] used feature spatial heterogeneity to reduce bandwidth utilization by transmitting sparse feature maps. CoBEVT [41] proposed the first multi-camera-based collaborative perception framework and designed the fused axis attention module to enable multi-view interactions. In this work, we propose an innovative feature calibration and interaction method that ensures robust vehicle collaboration in high-noise dynamic environments." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 472, + 101, + 484 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 472, + 101, + 484 + ], + "spans": [ + { + "bbox": [ + 46, + 472, + 101, + 484 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 45, + 492, + 279, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 492, + 279, + 540 + ], + "spans": [ + { + "bbox": [ + 45, + 492, + 279, + 540 + ], + "type": "text", + "content": "This section introduces our efficient and robust collaborative perception method. Figure 1 shows the general procedure of our method, which is divided into six phases. We will detail each phase in the following introduction." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 45, + 548, + 260, + 561 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 548, + 260, + 561 + ], + "spans": [ + { + "bbox": [ + 45, + 548, + 260, + 561 + ], + "type": "text", + "content": "3.1. Metadata Sharing and Feature Extraction" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 45, + 566, + 281, + 730 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 566, + 281, + 730 + ], + "spans": [ + { + "bbox": [ + 45, + 566, + 281, + 730 + ], + "type": "text", + "content": "In multi-vehicle collaboration scenarios, one of the CAVs, known as the ego vehicle, constructs a communication graph. In this graph, the ego vehicle serves as the requester, while other vehicles within its communication range act as supporters. The ego vehicle broadcasts its metadata information, including position, heading, speed, and more. Supporters, upon receiving this metadata, project their local point cloud observations into the ego vehicle's coordinate system. Each vehicle then encodes these transformed point clouds into bird's eye view (BEV) features, yielding a visual representation. Given the " + }, + { + "bbox": [ + 45, + 566, + 281, + 730 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 45, + 566, + 281, + 730 + ], + "type": "text", + "content": "-th vehicle local observations " + }, + { + "bbox": [ + 45, + 566, + 281, + 730 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_i" + }, + { + "bbox": [ + 45, + 566, + 281, + 730 + ], + "type": "text", + "content": ", the extracted features are represented as " + }, + { + "bbox": [ + 45, + 566, + 281, + 730 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_i = \\Phi_{\\mathrm{enc}}(\\mathbf{X}_i) \\in \\mathbb{R}^{H \\times W \\times C}" + }, + { + "bbox": [ + 45, + 566, + 281, + 730 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 45, + 566, + 281, + 730 + ], + "type": "inline_equation", + "content": "\\Phi_{\\mathrm{enc}}(\\cdot)" + }, + { + "bbox": [ + 45, + 566, + 281, + 730 + ], + "type": "text", + "content": " denotes the PointPillar [19] encoder shared by all vehicles and " + }, + { + "bbox": [ + 45, + 566, + 281, + 730 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 45, + 566, + 281, + 730 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 45, + 566, + 281, + 730 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 45, + 566, + 281, + 730 + ], + "type": "text", + "content": "," + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 296, + 106, + 530, + 142 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 106, + 530, + 142 + ], + "spans": [ + { + "bbox": [ + 296, + 106, + 530, + 142 + ], + "type": "text", + "content": "and " + }, + { + "bbox": [ + 296, + 106, + 530, + 142 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 296, + 106, + 530, + 142 + ], + "type": "text", + "content": " stand for the height, width, and channel of the feature map, respectively. Then, these extracted features are fed into the filter and merge feature sampling module." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 296, + 147, + 481, + 160 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 147, + 481, + 160 + ], + "spans": [ + { + "bbox": [ + 296, + 147, + 481, + 160 + ], + "type": "text", + "content": "3.2. Filter and Merge Feature Sampling" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 296, + 165, + 530, + 339 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 165, + 530, + 339 + ], + "spans": [ + { + "bbox": [ + 296, + 165, + 530, + 339 + ], + "type": "text", + "content": "Previous works have utilized well-designed mechanisms such as information entropy communication selection [37] and spatial heterogeneity map [15, 35] to reduce the required transmission bandwidth. However, these methods primarily focus on the inter-class redundancy between foreground and background features, ignoring the intra-class redundancy among features, which results in sub-optimal compression. To address this gap, we introduce an advanced Filtering and Merging Feature Sampling strategy (FMS). This strategy considers both inter-class and intra-class redundancy relationships, efficiently extracting a concise and distinctive set of feature vectors from the original feature maps, thus reducing communication overheads more effectively. FMS is composed of two core components as follows." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 296, + 340, + 531, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 340, + 531, + 491 + ], + "spans": [ + { + "bbox": [ + 296, + 340, + 531, + 491 + ], + "type": "text", + "content": "Filter Sampler. In object detection, foreground areas containing objects are more significant than the background areas. Therefore, we implement the idea of reducing spatial redundancy into a feature filter sampler module, aiming to preserve perceptually important yet sparse sets of feature vectors. Since explicitly learning a binary sampler is infeasible, we develop a confidence filter strategy. Initially, a detection confidence map is generated for the feature map. It reflects the perceptual importance of different spatial areas, with higher levels indicating potential object areas and lower levels typically denoting redundant background areas. For the feature map " + }, + { + "bbox": [ + 296, + 340, + 531, + 491 + ], + "type": "inline_equation", + "content": "\\pmb{F}_i" + }, + { + "bbox": [ + 296, + 340, + 531, + 491 + ], + "type": "text", + "content": " of the " + }, + { + "bbox": [ + 296, + 340, + 531, + 491 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 296, + 340, + 531, + 491 + ], + "type": "text", + "content": "-th vehicle, its confidence map " + }, + { + "bbox": [ + 296, + 340, + 531, + 491 + ], + "type": "inline_equation", + "content": "C_i" + }, + { + "bbox": [ + 296, + 340, + 531, + 491 + ], + "type": "text", + "content": " is defined as:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 346, + 498, + 530, + 513 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 346, + 498, + 530, + 513 + ], + "spans": [ + { + "bbox": [ + 346, + 498, + 530, + 513 + ], + "type": "interline_equation", + "content": "\\boldsymbol {C} _ {i} = \\Phi_ {\\text {c o n - g e n}} \\left(\\boldsymbol {F} _ {i}\\right) \\in [ 0, 1 ] ^ {H \\times W}, \\tag {1}", + "image_path": "8a09e87c55b500d291562a7ebf35612e323c6552110b7d698d53953224a60855.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 296, + 520, + 530, + 646 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 520, + 530, + 646 + ], + "spans": [ + { + "bbox": [ + 296, + 520, + 530, + 646 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 296, + 520, + 530, + 646 + ], + "type": "inline_equation", + "content": "\\Phi_{\\mathrm{con\\_gen}(\\cdot)}" + }, + { + "bbox": [ + 296, + 520, + 530, + 646 + ], + "type": "text", + "content": " represents the confidence generation network with detection decoder structure. Then the confidence map is thresholded, followed by non-maximum suppression, resulting in a binary mask " + }, + { + "bbox": [ + 296, + 520, + 530, + 646 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 296, + 520, + 530, + 646 + ], + "type": "text", + "content": ". Utilizing this binary mask, we proceed to preserve the sparse foreground features " + }, + { + "bbox": [ + 296, + 520, + 530, + 646 + ], + "type": "inline_equation", + "content": "\\widetilde{\\pmb{F}}_i = B\\odot \\pmb{F}_i" + }, + { + "bbox": [ + 296, + 520, + 530, + 646 + ], + "type": "text", + "content": ". In order to cope with the dynamically changing environmental conditions and fully ensure the robustness of the system, the filtering threshold can be varied with the sensor data and the network state. We set " + }, + { + "bbox": [ + 296, + 520, + 530, + 646 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 296, + 520, + 530, + 646 + ], + "type": "text", + "content": " as the filter rate and the number of remaining feature vectors is " + }, + { + "bbox": [ + 296, + 520, + 530, + 646 + ], + "type": "inline_equation", + "content": "L = \\alpha \\times HW" + }, + { + "bbox": [ + 296, + 520, + 530, + 646 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 296, + 648, + 530, + 730 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 648, + 530, + 730 + ], + "spans": [ + { + "bbox": [ + 296, + 648, + 530, + 730 + ], + "type": "text", + "content": "Merge Sampler. Upon extracting a detailed foreground feature vector set with the filter sampler, we employ the merge sampler for additional optimization, refining similar or repetitive foreground feature vectors through weighted merging. The process is divided into three stages: information-driven feature grouping, attention-inspired feature merging, and index-based feature reconstruction." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 277, + 747, + 301, + 756 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 277, + 747, + 301, + 756 + ], + "spans": [ + { + "bbox": [ + 277, + 747, + 301, + 756 + ], + "type": "text", + "content": "12577" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 595, + 842 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 45, + 106, + 281, + 201 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 106, + 281, + 201 + ], + "spans": [ + { + "bbox": [ + 45, + 106, + 281, + 201 + ], + "type": "text", + "content": "(a) Information-Driven Feature Grouping. Initially, a variant of the nearest neighbor clustering algorithm is applied to group the foreground feature vector set. Given a set of feature vector " + }, + { + "bbox": [ + 45, + 106, + 281, + 201 + ], + "type": "inline_equation", + "content": "\\widetilde{\\boldsymbol{F}}_i = [x_1,x_2,\\dots ,x_L]^\\top" + }, + { + "bbox": [ + 45, + 106, + 281, + 201 + ], + "type": "text", + "content": " and cluster center " + }, + { + "bbox": [ + 45, + 106, + 281, + 201 + ], + "type": "inline_equation", + "content": "X_{c}" + }, + { + "bbox": [ + 45, + 106, + 281, + 201 + ], + "type": "text", + "content": ", we compute the indicator " + }, + { + "bbox": [ + 45, + 106, + 281, + 201 + ], + "type": "inline_equation", + "content": "\\delta_{i}" + }, + { + "bbox": [ + 45, + 106, + 281, + 201 + ], + "type": "text", + "content": " for each feature vector. " + }, + { + "bbox": [ + 45, + 106, + 281, + 201 + ], + "type": "inline_equation", + "content": "\\delta_{i}" + }, + { + "bbox": [ + 45, + 106, + 281, + 201 + ], + "type": "text", + "content": " is calculated as the minimum feature distance minus the average pixel distance to any other cluster center vector, expressed as:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 59, + 209, + 279, + 230 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 209, + 279, + 230 + ], + "spans": [ + { + "bbox": [ + 59, + 209, + 279, + 230 + ], + "type": "interline_equation", + "content": "\\delta_ {i} = \\min _ {j: x _ {j} \\in X _ {c}} \\left(\\| x _ {i} - x _ {j} \\| _ {2} ^ {2} - \\gamma \\| p (x _ {i}), p (x _ {j}) \\| _ {2} ^ {2}\\right), \\tag {2}", + "image_path": "b318f1665f213f040e53be880d3771245ed1e13681847141090cda97fb550a95.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 45, + 239, + 279, + 320 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 239, + 279, + 320 + ], + "spans": [ + { + "bbox": [ + 45, + 239, + 279, + 320 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 45, + 239, + 279, + 320 + ], + "type": "inline_equation", + "content": "\\delta_{i}" + }, + { + "bbox": [ + 45, + 239, + 279, + 320 + ], + "type": "text", + "content": " denotes to which cluster the feature vector " + }, + { + "bbox": [ + 45, + 239, + 279, + 320 + ], + "type": "inline_equation", + "content": "x_{i}" + }, + { + "bbox": [ + 45, + 239, + 279, + 320 + ], + "type": "text", + "content": " should belong. " + }, + { + "bbox": [ + 45, + 239, + 279, + 320 + ], + "type": "inline_equation", + "content": "p(\\cdot)" + }, + { + "bbox": [ + 45, + 239, + 279, + 320 + ], + "type": "text", + "content": " means getting the position of the vector and " + }, + { + "bbox": [ + 45, + 239, + 279, + 320 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 45, + 239, + 279, + 320 + ], + "type": "text", + "content": " is a hyperparameter. Subsequently, we can divide all the feature vectors in " + }, + { + "bbox": [ + 45, + 239, + 279, + 320 + ], + "type": "inline_equation", + "content": "\\widetilde{\\boldsymbol{F}}_i" + }, + { + "bbox": [ + 45, + 239, + 279, + 320 + ], + "type": "text", + "content": " into " + }, + { + "bbox": [ + 45, + 239, + 279, + 320 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 45, + 239, + 279, + 320 + ], + "type": "text", + "content": " clusters represented by " + }, + { + "bbox": [ + 45, + 239, + 279, + 320 + ], + "type": "inline_equation", + "content": "G = \\{G_{1},G_{2},\\dots ,G_{K}\\}" + }, + { + "bbox": [ + 45, + 239, + 279, + 320 + ], + "type": "text", + "content": ". The total number of clusters is determined by dynamic clustering ratio " + }, + { + "bbox": [ + 45, + 239, + 279, + 320 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 45, + 239, + 279, + 320 + ], + "type": "text", + "content": ", and is calculated as " + }, + { + "bbox": [ + 45, + 239, + 279, + 320 + ], + "type": "inline_equation", + "content": "K = \\beta \\times L" + }, + { + "bbox": [ + 45, + 239, + 279, + 320 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 45, + 321, + 279, + 414 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 321, + 279, + 414 + ], + "spans": [ + { + "bbox": [ + 45, + 321, + 279, + 414 + ], + "type": "text", + "content": "(b) Attention-Inspired Feature Merging. A straightforward strategy for merging feature vectors is to average of each feature vector within a cluster. However, this scheme can be severely affected by outlier feature vectors. Drawing inspiration from attention mechanisms, we utilize the confidence score as a guide to quantify the significance of each feature. Thus, the merged feature vector " + }, + { + "bbox": [ + 45, + 321, + 279, + 414 + ], + "type": "inline_equation", + "content": "\\tilde{x}_i" + }, + { + "bbox": [ + 45, + 321, + 279, + 414 + ], + "type": "text", + "content": " for the " + }, + { + "bbox": [ + 45, + 321, + 279, + 414 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 45, + 321, + 279, + 414 + ], + "type": "text", + "content": "th cluster " + }, + { + "bbox": [ + 45, + 321, + 279, + 414 + ], + "type": "inline_equation", + "content": "G_i" + }, + { + "bbox": [ + 45, + 321, + 279, + 414 + ], + "type": "text", + "content": " is computed as:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 122, + 422, + 279, + 452 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 122, + 422, + 279, + 452 + ], + "spans": [ + { + "bbox": [ + 122, + 422, + 279, + 452 + ], + "type": "interline_equation", + "content": "\\widetilde {x} _ {i} = \\frac {\\sum_ {j \\in G _ {i}} c _ {j} x _ {j}}{\\sum_ {j \\in G _ {i}} c _ {j}}, \\tag {3}", + "image_path": "d46991d06acdaf0135b1a2b10e7826d61c8203e53d651242fbc31b01b6a9c486.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 45, + 459, + 279, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 459, + 279, + 506 + ], + "spans": [ + { + "bbox": [ + 45, + 459, + 279, + 506 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 45, + 459, + 279, + 506 + ], + "type": "inline_equation", + "content": "c_{j}" + }, + { + "bbox": [ + 45, + 459, + 279, + 506 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 45, + 459, + 279, + 506 + ], + "type": "inline_equation", + "content": "x_{j}" + }, + { + "bbox": [ + 45, + 459, + 279, + 506 + ], + "type": "text", + "content": " represent the confidence score and the original feature vector, respectively. In the end, we obtain a final set of feature vectors " + }, + { + "bbox": [ + 45, + 459, + 279, + 506 + ], + "type": "inline_equation", + "content": "\\mathbf{Z}_i = \\{\\widetilde{x}_1,\\widetilde{x}_2,\\dots ,\\widetilde{x}_K\\}" + }, + { + "bbox": [ + 45, + 459, + 279, + 506 + ], + "type": "text", + "content": " that is necessary for transmission." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 45, + 506, + 279, + 600 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 506, + 279, + 600 + ], + "spans": [ + { + "bbox": [ + 45, + 506, + 279, + 600 + ], + "type": "text", + "content": "(c) Index-Based Feature Reconstruction. During the feature grouping and merging process, each feature vector is allocated to a cluster, and every cluster is represented by a merged vector. We maintain a record of the index correspondence between original and merged feature vectors. Utilizing this index record, the ego-vehicle ensures that the merged feature vectors are mapped to their corresponding positions, leading to the reconstruction of feature maps." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 607, + 193, + 619 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 607, + 193, + 619 + ], + "spans": [ + { + "bbox": [ + 46, + 607, + 193, + 619 + ], + "type": "text", + "content": "3.3. Feature Spatial Calibration" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 45, + 624, + 279, + 730 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 624, + 279, + 730 + ], + "spans": [ + { + "bbox": [ + 45, + 624, + 279, + 730 + ], + "type": "text", + "content": "Localization errors [30, 31] may lead to the misalignment of feature maps among vehicles. Such misalignment causes the ego-vehicle to misinterpret the object's location, resulting in sub-optimal perception, as depicted in the upper right corner of Figure 2. To address this challenge, we introduce the Feature Spatial Calibration module (FSC) to facilitate precise feature alignment, as shown in Figure 2. The core idea of FSC is that overlapping viewpoints allow multiple vehicles to detect partially the same objects. The process" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 299, + 106, + 530, + 213 + ], + "blocks": [ + { + "bbox": [ + 299, + 106, + 530, + 213 + ], + "lines": [ + { + "bbox": [ + 299, + 106, + 530, + 213 + ], + "spans": [ + { + "bbox": [ + 299, + 106, + 530, + 213 + ], + "type": "image", + "image_path": "7d58f4ddb6a0733b531a8909f7caf793b1c101be3a55015892f12e2ce2992a07.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 296, + 223, + 530, + 255 + ], + "lines": [ + { + "bbox": [ + 296, + 223, + 530, + 255 + ], + "spans": [ + { + "bbox": [ + 296, + 223, + 530, + 255 + ], + "type": "text", + "content": "Figure 2. Illustration of the proposed FSC. By employing FSC, misaligned matching regions are corrected at the feature level, leading to more accurate predictions." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 296, + 263, + 530, + 286 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 263, + 530, + 286 + ], + "spans": [ + { + "bbox": [ + 296, + 263, + 530, + 286 + ], + "type": "text", + "content": "involves three phases: consensus matching, geometric verification, and error modulation." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 296, + 287, + 530, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 287, + 530, + 449 + ], + "spans": [ + { + "bbox": [ + 296, + 287, + 530, + 449 + ], + "type": "text", + "content": "Consensus Matching. After reconstructing the received features, the ego vehicle generates sparse feature maps. Within these maps, areas rich in information signify potential target regions, which correspond to the proposed matching areas. We denote the proposed matching regions of the ego vehicle as " + }, + { + "bbox": [ + 296, + 287, + 530, + 449 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 296, + 287, + 530, + 449 + ], + "type": "text", + "content": " and those identified by the collaborative vehicle under noisy pose conditions as " + }, + { + "bbox": [ + 296, + 287, + 530, + 449 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 296, + 287, + 530, + 449 + ], + "type": "text", + "content": ". Utilizing " + }, + { + "bbox": [ + 296, + 287, + 530, + 449 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 296, + 287, + 530, + 449 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 296, + 287, + 530, + 449 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 296, + 287, + 530, + 449 + ], + "type": "text", + "content": ", a weighted bipartite graph is constructed, wherein the weight of each edge is determined by the distance between the nodes, encapsulated in a cost matrix. The matching process is then converted into a linear assignment task, with the objective of identifying a matching result with the lowest cumulative edge weight. This procedure yields the matching pairs " + }, + { + "bbox": [ + 296, + 287, + 530, + 449 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 296, + 287, + 530, + 449 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 296, + 450, + 530, + 648 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 450, + 530, + 648 + ], + "spans": [ + { + "bbox": [ + 296, + 450, + 530, + 648 + ], + "type": "text", + "content": "Geometric Verification. Invalid matches may occur due to objects located in exclusive zones and detection noise. To tackle this issue, we utilize RANSAC to filter and sift a consistent set of matches aligned with expected geometric transformations. Initially, a random matching subset " + }, + { + "bbox": [ + 296, + 450, + 530, + 648 + ], + "type": "inline_equation", + "content": "M_{s} \\subseteq M" + }, + { + "bbox": [ + 296, + 450, + 530, + 648 + ], + "type": "text", + "content": " is selected, and then the transformation matrix " + }, + { + "bbox": [ + 296, + 450, + 530, + 648 + ], + "type": "inline_equation", + "content": "\\Gamma_{s}" + }, + { + "bbox": [ + 296, + 450, + 530, + 648 + ], + "type": "text", + "content": " is computed using singular value decomposition. When " + }, + { + "bbox": [ + 296, + 450, + 530, + 648 + ], + "type": "inline_equation", + "content": "\\Gamma_{s}" + }, + { + "bbox": [ + 296, + 450, + 530, + 648 + ], + "type": "text", + "content": " is applied to all pairs within " + }, + { + "bbox": [ + 296, + 450, + 530, + 648 + ], + "type": "inline_equation", + "content": "M_{s}" + }, + { + "bbox": [ + 296, + 450, + 530, + 648 + ], + "type": "text", + "content": ", and if the posttransformation distance between these pairs remains below the threshold " + }, + { + "bbox": [ + 296, + 450, + 530, + 648 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 296, + 450, + 530, + 648 + ], + "type": "text", + "content": ", the set is considered correctly aligned. The threshold " + }, + { + "bbox": [ + 296, + 450, + 530, + 648 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 296, + 450, + 530, + 648 + ], + "type": "text", + "content": " reflects the allowable localization error within the original collaborative framework. This process is iteratively conducted to identify the optimal transformation matrix that correlates with the maximum number of correct matches. Ultimately, a optimal refined transformation matrix " + }, + { + "bbox": [ + 296, + 450, + 530, + 648 + ], + "type": "inline_equation", + "content": "\\Gamma_{r}" + }, + { + "bbox": [ + 296, + 450, + 530, + 648 + ], + "type": "text", + "content": " is obtained and applied in subsequent spatial calibration operations, yielding the aligned feature " + }, + { + "bbox": [ + 296, + 450, + 530, + 648 + ], + "type": "inline_equation", + "content": "\\widetilde{Z}_{j} = \\Gamma_{r}Z_{j}" + }, + { + "bbox": [ + 296, + 450, + 530, + 648 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 296, + 649, + 530, + 730 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 649, + 530, + 730 + ], + "spans": [ + { + "bbox": [ + 296, + 649, + 530, + 730 + ], + "type": "text", + "content": "Error Modulation. To enhance the adaptability of the calibration method in various environments, we incorporate an error modulation strategy. This strategy aims to achieve a balance between localization errors and the estimation errors that emerge from the calibration process. It measures the overlap ratios between ego and collaborative features in both their adjusted and original states. Subsequently, fea" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 277, + 747, + 300, + 756 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 277, + 747, + 300, + 756 + ], + "spans": [ + { + "bbox": [ + 277, + 747, + 300, + 756 + ], + "type": "text", + "content": "12578" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 595, + 842 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 49, + 103, + 276, + 213 + ], + "blocks": [ + { + "bbox": [ + 49, + 103, + 276, + 213 + ], + "lines": [ + { + "bbox": [ + 49, + 103, + 276, + 213 + ], + "spans": [ + { + "bbox": [ + 49, + 103, + 276, + 213 + ], + "type": "image", + "image_path": "d42d10cfc3c0721cb0a1f26165bbe6d235184a7bc9b1bd0e8df2217e6ae379c9.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 45, + 221, + 279, + 255 + ], + "lines": [ + { + "bbox": [ + 45, + 221, + 279, + 255 + ], + "spans": [ + { + "bbox": [ + 45, + 221, + 279, + 255 + ], + "type": "text", + "content": "Figure 3. A example of local and global Attention in AFF. This shows how AFF processes both 3D local windows and the global query token for position-level and context-aware aggregation." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 45, + 268, + 279, + 293 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 268, + 279, + 293 + ], + "spans": [ + { + "bbox": [ + 45, + 268, + 279, + 293 + ], + "type": "text", + "content": "tures demonstrating the most optimal overlap are selected as the final input." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 45, + 298, + 212, + 310 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 298, + 212, + 310 + ], + "spans": [ + { + "bbox": [ + 45, + 298, + 212, + 310 + ], + "type": "text", + "content": "3.4. Attention-based Feature Fusion" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 45, + 316, + 280, + 562 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 316, + 280, + 562 + ], + "spans": [ + { + "bbox": [ + 45, + 316, + 280, + 562 + ], + "type": "text", + "content": "In multi-vehicle collaboration scenarios, vehicles are able to capture heterogeneous information from different spatial regions. In order to efficiently fuse perceptual features from multiple vehicles, we propose an Attention-based Feature Fusion method (AFF). As shown in Figure 3, AFF utilizes alternating local and global attention to achieve position-level accurate matching in occlusion-variable traffic scenes and to capture the global semantics attention of road topology and traffic states. Specifically, we stack all the vehicles' features to " + }, + { + "bbox": [ + 45, + 316, + 280, + 562 + ], + "type": "inline_equation", + "content": "S \\in \\mathbb{R}^{N \\times H \\times W \\times C}" + }, + { + "bbox": [ + 45, + 316, + 280, + 562 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 45, + 316, + 280, + 562 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 45, + 316, + 280, + 562 + ], + "type": "text", + "content": " is the number of vehicles. The features are divided into 3D non-overlapping windows, each of size " + }, + { + "bbox": [ + 45, + 316, + 280, + 562 + ], + "type": "inline_equation", + "content": "N \\times P \\times P" + }, + { + "bbox": [ + 45, + 316, + 280, + 562 + ], + "type": "text", + "content": ". The shape of the partition tensor is " + }, + { + "bbox": [ + 45, + 316, + 280, + 562 + ], + "type": "inline_equation", + "content": "\\left(\\frac{H}{P} \\times \\frac{W}{P}, N \\times P^2, C\\right)" + }, + { + "bbox": [ + 45, + 316, + 280, + 562 + ], + "type": "text", + "content": ". We perform 3D local self attention on the local window and implement region interaction within the window. In contrast, global attention goes beyond the limitations of the local view. It uses the extracted global query token and is shared between all windows to interact with the local key and value representations, thus helping to capture long-range dependencies of the features. Inspired by the efficiency of [13, 33], the global query token " + }, + { + "bbox": [ + 45, + 316, + 280, + 562 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 45, + 316, + 280, + 562 + ], + "type": "text", + "content": " is generated as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 89, + 569, + 279, + 583 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 569, + 279, + 583 + ], + "spans": [ + { + "bbox": [ + 89, + 569, + 279, + 583 + ], + "type": "interline_equation", + "content": "\\mathbf {g} = \\mathcal {P} \\left(\\mathcal {C} _ {1 \\times 1} \\left(\\mathcal {S E} (\\mathcal {G} (\\mathcal {D} (\\mathbf {x}))) + \\mathbf {x}\\right)\\right), \\tag {4}", + "image_path": "7d3cf72ee7c75cc0a5f79d32b646b752237eb8d0a916862be1ef6aee8a477cc8.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 45, + 590, + 279, + 730 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 590, + 279, + 730 + ], + "spans": [ + { + "bbox": [ + 45, + 590, + 279, + 730 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 45, + 590, + 279, + 730 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 45, + 590, + 279, + 730 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 45, + 590, + 279, + 730 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 45, + 590, + 279, + 730 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 45, + 590, + 279, + 730 + ], + "type": "inline_equation", + "content": "\\mathcal{SE}" + }, + { + "bbox": [ + 45, + 590, + 279, + 730 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 45, + 590, + 279, + 730 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_{1\\times 1}" + }, + { + "bbox": [ + 45, + 590, + 279, + 730 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 45, + 590, + 279, + 730 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 45, + 590, + 279, + 730 + ], + "type": "text", + "content": " denote the depth-wise separable convolution, GELU activation function, Squeeze-and-Excitation operation, " + }, + { + "bbox": [ + 45, + 590, + 279, + 730 + ], + "type": "inline_equation", + "content": "1\\times 1" + }, + { + "bbox": [ + 45, + 590, + 279, + 730 + ], + "type": "text", + "content": " convolution and max pooling, respectively. These designs enable the global query token to provide advantages such as inductive bias and modeling of inter-channel dependencies. We then construct our proposed AFF module by combining this local and global attention with typical designs of Transformers, including LayerNorm [1], MLPs [9], and skip-connections [14]. This module allows the system to analyze spatial correlations from a local view while also capturing global feature responses, ensuring efficient and precise perception in dy" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 303, + 105, + 526, + 214 + ], + "blocks": [ + { + "bbox": [ + 303, + 105, + 526, + 214 + ], + "lines": [ + { + "bbox": [ + 303, + 105, + 526, + 214 + ], + "spans": [ + { + "bbox": [ + 303, + 105, + 526, + 214 + ], + "type": "image", + "image_path": "1f09b47ad9f51cbbee866c451bf65bbf30029ce7fdb5145e4bb85834b8ec315b.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 308, + 223, + 518, + 234 + ], + "lines": [ + { + "bbox": [ + 308, + 223, + 518, + 234 + ], + "spans": [ + { + "bbox": [ + 308, + 223, + 518, + 234 + ], + "type": "text", + "content": "Figure 4. The architecture of the proposed AFI component." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 296, + 258, + 529, + 281 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 258, + 529, + 281 + ], + "spans": [ + { + "bbox": [ + 296, + 258, + 529, + 281 + ], + "type": "text", + "content": "namic, complex, and occlusion-variable traffic scenarios. Finally, we obtain the fused features " + }, + { + "bbox": [ + 296, + 258, + 529, + 281 + ], + "type": "inline_equation", + "content": "\\pmb{H}_i" + }, + { + "bbox": [ + 296, + 258, + 529, + 281 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 297, + 295, + 502, + 308 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 295, + 502, + 308 + ], + "spans": [ + { + "bbox": [ + 297, + 295, + 502, + 308 + ], + "type": "text", + "content": "3.5. Accuracy Enhanced Feature Interaction" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 296, + 314, + 530, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 314, + 530, + 548 + ], + "spans": [ + { + "bbox": [ + 296, + 314, + 530, + 548 + ], + "type": "text", + "content": "Previous works [15, 21, 38, 42, 43] have demonstrated that fused features can provide richer semantic information, thereby enhancing perceptual performance. However, they may be affected by collaborative noise, such as asynchronous motion blur and inaccurate projections, which can compromise accurate position information and become a bottleneck to the optimal realization of perceptual performance. Ego-centric features may contain locally critical spatial location information without being affected by collaborative noise. To this end, we propose a novel Accuracy Enhanced Feature Interaction (AEI) strategy that leverages the accurate positional information inherent in ego-centric features to enhance the rich semantic information provided by collaborative fused features. Firstly, we design a context cross attention module that is tailored for feature interaction, as shown in Figure 4. It considers the features from the ego-vehicle as queries, and obtains the context of the query location from the fused features as the key and value to achieve cross-attention. The computed output of the spatial location " + }, + { + "bbox": [ + 296, + 314, + 530, + 548 + ], + "type": "inline_equation", + "content": "(i,j)" + }, + { + "bbox": [ + 296, + 314, + 530, + 548 + ], + "type": "text", + "content": " is as follows:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 335, + 560, + 530, + 588 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 560, + 530, + 588 + ], + "spans": [ + { + "bbox": [ + 335, + 560, + 530, + 588 + ], + "type": "interline_equation", + "content": "y _ {i j} = \\sum_ {a, b \\in \\mathcal {N} _ {k} (i, j)} \\operatorname {s o f t m a x} \\left(q _ {i j} ^ {\\top} k _ {a b}\\right) v _ {a b}, \\tag {5}", + "image_path": "1f0251cca0b9da86efc9ca24d9b756f8b5932d86b8f92f486bf5393506646137.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 296, + 600, + 531, + 730 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 600, + 531, + 730 + ], + "spans": [ + { + "bbox": [ + 296, + 600, + 531, + 730 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 296, + 600, + 531, + 730 + ], + "type": "inline_equation", + "content": "\\mathcal{N}_k(i,j)" + }, + { + "bbox": [ + 296, + 600, + 531, + 730 + ], + "type": "text", + "content": " is context position of " + }, + { + "bbox": [ + 296, + 600, + 531, + 730 + ], + "type": "inline_equation", + "content": "(i,j)" + }, + { + "bbox": [ + 296, + 600, + 531, + 730 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 296, + 600, + 531, + 730 + ], + "type": "inline_equation", + "content": "q_{ij}, k_{ab}, v_{ab}" + }, + { + "bbox": [ + 296, + 600, + 531, + 730 + ], + "type": "text", + "content": " represent the ego query, context key and value, respectively. As the relevant features of the same object are often located in similar locations across different feature maps, we prioritize the local context details of the query location. Additionally, considering that only a few positions within the sparse feature map hold significant information, focusing on these context-rich areas is computationally efficient. Following this, we employ the split attention [48] to adaptively fuse information from multiple branches, generating the enhanced output feature " + }, + { + "bbox": [ + 296, + 600, + 531, + 730 + ], + "type": "inline_equation", + "content": "\\widetilde{H}_i" + }, + { + "bbox": [ + 296, + 600, + 531, + 730 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 277, + 747, + 301, + 756 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 277, + 747, + 301, + 756 + ], + "spans": [ + { + "bbox": [ + 277, + 747, + 301, + 756 + ], + "type": "text", + "content": "12579" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 595, + 842 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 47, + 103, + 531, + 261 + ], + "blocks": [ + { + "bbox": [ + 47, + 103, + 531, + 261 + ], + "lines": [ + { + "bbox": [ + 47, + 103, + 531, + 261 + ], + "spans": [ + { + "bbox": [ + 47, + 103, + 531, + 261 + ], + "type": "table", + "html": "
DatasetV2V4RealOPV2V
Noise Level σt/σr (m/°)0.0/0.00.4/0.40.8/0.80.0/0.00.4/0.40.8/0.8
No Collaboration41.64/23.4341.64/23.4341.64/23.4373.25/58.2273.25/58.2273.25/58.22
Late Fusion56.80/27.6240.81/13.3126.65/10.6686.46/79.2176.56/40.0145.57/21.00
F-Cooper [3]61.12/32.3651.71/24.4841.90/19.9187.39/79.3979.33/40.1252.91/16.90
AttFuse [7]64.41/34.3258.42/30.0249.37/25.4490.50/81.8083.54/52.4368.37/40.96
Where2comm [15]63.69/34.7958.48/28.0949.77/24.2590.46/84.2282.17/56.4773.95/45.86
DiscoNet [21]64.19/35.2557.71/28.4149.19/24.9489.58/81.4583.25/53.2261.37/30.53
V2VNet [38]65.70/35.3862.11/30.6354.73/25.5791.35/82.4385.43/54.1670.43/31.28
CoAlign [25]64.06/36.6059.58/31.5253.02/26.6090.93/84.2884.48/57.4573.95/48.86
V2X-ViT [42]66.42/37.3463.33/32.3857.15/28.2291.74/83.3182.91/54.7361.70/26.07
CoBEVT [41]66.01/37.3658.63/29.0949.22/23.6691.71/85.9885.49/60.6464.63/29.59
ERMVP (Ours)67.66/42.9765.01/40.4560.88/35.0192.18/85.5985.67/66.3277.13/59.15
", + "image_path": "1de46cb1d7557a613601596d0c8251c9c9eee305c37b4dae042ecda1c4e0817a.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 70, + 269, + 505, + 280 + ], + "lines": [ + { + "bbox": [ + 70, + 269, + 505, + 280 + ], + "spans": [ + { + "bbox": [ + 70, + 269, + 505, + 280 + ], + "type": "text", + "content": "Table 1. Overall performance on V2V4Real and OPV2V datasets with pose noises. The results are reported in AP@0.5/0.7." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 286, + 149, + 298 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 286, + 149, + 298 + ], + "spans": [ + { + "bbox": [ + 46, + 286, + 149, + 298 + ], + "type": "text", + "content": "3.6. Decoder and Loss" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 45, + 305, + 279, + 401 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 305, + 279, + 401 + ], + "spans": [ + { + "bbox": [ + 45, + 305, + 279, + 401 + ], + "type": "text", + "content": "Based on the final fused feature " + }, + { + "bbox": [ + 45, + 305, + 279, + 401 + ], + "type": "inline_equation", + "content": "\\widetilde{H}_i" + }, + { + "bbox": [ + 45, + 305, + 279, + 401 + ], + "type": "text", + "content": ", we use the detection decoder to generate the final prediction output " + }, + { + "bbox": [ + 45, + 305, + 279, + 401 + ], + "type": "inline_equation", + "content": "Y_{i} = \\Phi_{\\mathrm{dec}}(\\widetilde{H}_{i})" + }, + { + "bbox": [ + 45, + 305, + 279, + 401 + ], + "type": "text", + "content": ". Each position of " + }, + { + "bbox": [ + 45, + 305, + 279, + 401 + ], + "type": "inline_equation", + "content": "Y_{i}" + }, + { + "bbox": [ + 45, + 305, + 279, + 401 + ], + "type": "text", + "content": " represents a rotated box with classes (c, x, y, h, w, " + }, + { + "bbox": [ + 45, + 305, + 279, + 401 + ], + "type": "inline_equation", + "content": "\\cos \\alpha" + }, + { + "bbox": [ + 45, + 305, + 279, + 401 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 45, + 305, + 279, + 401 + ], + "type": "inline_equation", + "content": "\\sin \\alpha" + }, + { + "bbox": [ + 45, + 305, + 279, + 401 + ], + "type": "text", + "content": "), corresponding to class confidence, position, size, and angle. These objects are the final output of the proposed collaborative perception system. Following existing work [19], we adopt the smooth " + }, + { + "bbox": [ + 45, + 305, + 279, + 401 + ], + "type": "inline_equation", + "content": "L_{1}" + }, + { + "bbox": [ + 45, + 305, + 279, + 401 + ], + "type": "text", + "content": " loss for regression and focal loss [23] for classification." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 418, + 124, + 431 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 418, + 124, + 431 + ], + "spans": [ + { + "bbox": [ + 46, + 418, + 124, + 431 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 439, + 216, + 451 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 439, + 216, + 451 + ], + "spans": [ + { + "bbox": [ + 46, + 439, + 216, + 451 + ], + "type": "text", + "content": "4.1. Datasets and Evaluation Metrics" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 45, + 459, + 279, + 668 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 459, + 279, + 668 + ], + "spans": [ + { + "bbox": [ + 45, + 459, + 279, + 668 + ], + "type": "text", + "content": "Datasets. We validate the proposed ERMVP in the task of LiDAR-based 3D object detection on two benchmark datasets including V2V4real [44] and OPV2V [43]. V2V4real [44] is the first large-scale real-world dataset on V2V perception, collected by two vehicles equipped with multi-modal sensors driving together in a variety of scenarios. It covers a driving area of 410 kilometers, including 20K frames of point clouds, with training/validation/test sets containing 14210/2000/3986 frames, respectively. OPV2V [43] is a large-scale vehicle-to-vehicle collaborative perception dataset, simulated by OpenCDA [40] and Carla [8]. It contains 73 different scenarios, in which different numbers (2 to 7) of collaborative vehicles appear together, each equipped with a lidar sensor and 4 cameras. The dataset contains a total of 11,464 frames of point clouds and RGB images. The training/validation/test sets contain 6374, 1980, and 2170 frames, respectively." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 45, + 671, + 279, + 730 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 671, + 279, + 730 + ], + "spans": [ + { + "bbox": [ + 45, + 671, + 279, + 730 + ], + "type": "text", + "content": "Evaluation Metrics. We adopt the Average Precision (AP) at Intersection-over-Union (IoU) thresholds of 0.5 and 0.7 to evaluate the detection performance. The calculation format of communication volume in [15] is used to count the message size by byte in the log scale with base 2." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 297, + 286, + 427, + 298 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 286, + 427, + 298 + ], + "spans": [ + { + "bbox": [ + 297, + 286, + 427, + 298 + ], + "type": "text", + "content": "4.2. Implementation Details" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 296, + 305, + 530, + 527 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 305, + 530, + 527 + ], + "spans": [ + { + "bbox": [ + 296, + 305, + 530, + 527 + ], + "type": "text", + "content": "We implement the proposed ERMVP and comparison model on the Pytorch toolbox [28], and train them on a NVIDIA GeForceRTX 3090 GPU using the Adam optimizer [18]. The initial learning rate is " + }, + { + "bbox": [ + 296, + 305, + 530, + 527 + ], + "type": "inline_equation", + "content": "2\\mathrm{e} - 3" + }, + { + "bbox": [ + 296, + 305, + 530, + 527 + ], + "type": "text", + "content": " and it decays every 15 epochs with a factor of 0.1. All models are trained with 60 epochs and the batch size is set to 2. Early stopping is used to find the best epoch. We also add normal point cloud data augmentation for all experiments, including scaling, rotation, and flipping. All detection models are based on PointPillar [19] backbone to extract 2D features from point clouds and a " + }, + { + "bbox": [ + 296, + 305, + 530, + 527 + ], + "type": "inline_equation", + "content": "0.4\\mathrm{m}" + }, + { + "bbox": [ + 296, + 305, + 530, + 527 + ], + "type": "text", + "content": " width/length is used for each voxel. AFF component has 3 encoded layers and a window size of 8 for both local and global attention. The balance hyperparameter " + }, + { + "bbox": [ + 296, + 305, + 530, + 527 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 296, + 305, + 530, + 527 + ], + "type": "text", + "content": " is 0.1 and the error tolerance threshold " + }, + { + "bbox": [ + 296, + 305, + 530, + 527 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 296, + 305, + 530, + 527 + ], + "type": "text", + "content": " is 0.25. Each vehicle has a communication range of 70 m based on [42], while vehicles outside of this broadcasting radius will be ignored. To simulate the localization and heading errors, we add Gaussian noise with a standard deviation of " + }, + { + "bbox": [ + 296, + 305, + 530, + 527 + ], + "type": "inline_equation", + "content": "\\sigma_t" + }, + { + "bbox": [ + 296, + 305, + 530, + 527 + ], + "type": "text", + "content": " for localization errors and " + }, + { + "bbox": [ + 296, + 305, + 530, + 527 + ], + "type": "inline_equation", + "content": "\\sigma_r" + }, + { + "bbox": [ + 296, + 305, + 530, + 527 + ], + "type": "text", + "content": " for heading errors." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 297, + 536, + 429, + 548 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 536, + 429, + 548 + ], + "spans": [ + { + "bbox": [ + 297, + 536, + 429, + 548 + ], + "type": "text", + "content": "4.3.Quantitative Evaluation" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 296, + 554, + 530, + 730 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 554, + 530, + 730 + ], + "spans": [ + { + "bbox": [ + 296, + 554, + 530, + 730 + ], + "type": "text", + "content": "Comparison of Detection Performance. Table 1 shows the performance comparison results of 3D detection on two datasets. We use the No Collaboration method as a baseline, which only uses ego vehicle's point clouds data without collaboration. Late Fusion allows vehicles to exchange detected outputs and utilizes non-maximum suppression to produce the final result. For intermediate fusion strategies, we evaluate the existing state-of-the-art (SOTA) methods: AttFuse [43], F-Cooper [3], V2VNet [38], DiscoNet [21], V2X-ViT [42], CoAlign (only fusion) [25], CoBEVT [41] and Where2comm [15]. The proposed ER-MVP achieves average improvements of " + }, + { + "bbox": [ + 296, + 554, + 530, + 730 + ], + "type": "inline_equation", + "content": "12.48\\% / 18.47\\%" + }, + { + "bbox": [ + 296, + 554, + 530, + 730 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 296, + 554, + 530, + 730 + ], + "type": "inline_equation", + "content": "6.18\\% / 10.02\\%" + }, + { + "bbox": [ + 296, + 554, + 530, + 730 + ], + "type": "text", + "content": " on two datasets in AP@0.5/0.7 compared with No Collaboration and Late Fusion methods, demonstrating the superiority of our collaborative method." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 277, + 747, + 301, + 756 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 277, + 747, + 301, + 756 + ], + "spans": [ + { + "bbox": [ + 277, + 747, + 301, + 756 + ], + "type": "text", + "content": "12580" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 595, + 842 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 48, + 103, + 168, + 212 + ], + "blocks": [ + { + "bbox": [ + 48, + 103, + 168, + 212 + ], + "lines": [ + { + "bbox": [ + 48, + 103, + 168, + 212 + ], + "spans": [ + { + "bbox": [ + 48, + 103, + 168, + 212 + ], + "type": "image", + "image_path": "a05276fe44b9498839d47f184cc6c82d59fe0ebf90f874dbf5bd3a548c86c415.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 154, + 215, + 202, + 224 + ], + "lines": [ + { + "bbox": [ + 154, + 215, + 202, + 224 + ], + "spans": [ + { + "bbox": [ + 154, + 215, + 202, + 224 + ], + "type": "text", + "content": "(a) V2V4Real" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 81, + 229, + 508, + 245 + ], + "lines": [ + { + "bbox": [ + 81, + 229, + 508, + 245 + ], + "spans": [ + { + "bbox": [ + 81, + 229, + 508, + 245 + ], + "type": "text", + "content": "Late Fusion FCooper V2VNet V2X-VIT CoBEVT -Where2comm -ERMVP" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_footnote" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 168, + 103, + 287, + 211 + ], + "blocks": [ + { + "bbox": [ + 168, + 103, + 287, + 211 + ], + "lines": [ + { + "bbox": [ + 168, + 103, + 287, + 211 + ], + "spans": [ + { + "bbox": [ + 168, + 103, + 287, + 211 + ], + "type": "image", + "image_path": "0d727aed98d8c717ad512713bc262dc41c0e12fc4ed41f3bfc943faddb1c8b02.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 60, + 250, + 515, + 262 + ], + "lines": [ + { + "bbox": [ + 60, + 250, + 515, + 262 + ], + "spans": [ + { + "bbox": [ + 60, + 250, + 515, + 262 + ], + "type": "text", + "content": "Figure 5. Collaborative perception performance comparison on the V2V4Real and OPV2V with varying communication volume." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 287, + 103, + 406, + 211 + ], + "blocks": [ + { + "bbox": [ + 287, + 103, + 406, + 211 + ], + "lines": [ + { + "bbox": [ + 287, + 103, + 406, + 211 + ], + "spans": [ + { + "bbox": [ + 287, + 103, + 406, + 211 + ], + "type": "image", + "image_path": "f88bea390d557ccfa8148f6e1357029b30be1ff3a78f8f3b8d9d1900be3dafa5.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 406, + 103, + 524, + 211 + ], + "blocks": [ + { + "bbox": [ + 406, + 103, + 524, + 211 + ], + "lines": [ + { + "bbox": [ + 406, + 103, + 524, + 211 + ], + "spans": [ + { + "bbox": [ + 406, + 103, + 524, + 211 + ], + "type": "image", + "image_path": "517d86dc9d3835c14d405265d54458bd70a95d2611f426bab979bfa02ba7bde2.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 45, + 271, + 279, + 353 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 271, + 279, + 353 + ], + "spans": [ + { + "bbox": [ + 45, + 271, + 279, + 353 + ], + "type": "text", + "content": "Meanwhile, it is superior to the SOTA collaborative perception method in both real-world and simulated scenarios: improves the SOTA collaborative performance by " + }, + { + "bbox": [ + 45, + 271, + 279, + 353 + ], + "type": "inline_equation", + "content": "1.24\\%" + }, + { + "bbox": [ + 45, + 271, + 279, + 353 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 45, + 271, + 279, + 353 + ], + "type": "inline_equation", + "content": "5.61\\%" + }, + { + "bbox": [ + 45, + 271, + 279, + 353 + ], + "type": "text", + "content": " on V2V4Real datasets in AP@0.5/0.7, and by " + }, + { + "bbox": [ + 45, + 271, + 279, + 353 + ], + "type": "inline_equation", + "content": "0.44\\%" + }, + { + "bbox": [ + 45, + 271, + 279, + 353 + ], + "type": "text", + "content": " on OPV2V datasets in AP@0.5. These results fully demonstrate the superiority of the ERMVP collaboration paradigm." + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 47, + 362, + 271, + 506 + ], + "blocks": [ + { + "bbox": [ + 47, + 362, + 271, + 506 + ], + "lines": [ + { + "bbox": [ + 47, + 362, + 271, + 506 + ], + "spans": [ + { + "bbox": [ + 47, + 362, + 271, + 506 + ], + "type": "table", + "html": "
DatasetV2V4RealOPV2V
Noise Level σt(m)0.51.00.51.0
Late Fusion11.0210.1353.8838.80
F-Cooper [3]24.6018.4132.2412.71
AttFuse [43]28.1024.3642.9535.53
DiscoNet [21]28.3624.0545.3624.96
V2VNet [38]29.5723.6747.4825.82
Where2comm [15]28.0223.5346.5945.59
CoAlign [21]29.8425.8249.0946.66
V2X-ViT [42]30.9923.6744.8321.77
CoBEVT [41]24.6722.0141.9126.05
ERMVP (Ours)39.2633.9477.2576.09
", + "image_path": "aaa3e111652a9c67b6d7667c976a0c1cf556ed7598929d5b2d4b2facb1d4c5f2.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 45, + 543, + 280, + 729 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 543, + 280, + 729 + ], + "spans": [ + { + "bbox": [ + 45, + 543, + 280, + 729 + ], + "type": "text", + "content": "Robustness to Localization and Heading Errors. To evaluate the sensitivity of existing methods to localization and heading errors, we use the same noise setting as [42] and conduct extensive experiments on two datasets. As shown in Tables 1 and 2, under ideal settings, Late Fusion and some advanced intermediate fusion methods can detect object vehicles with high accuracy, but their detection accuracy rapidly decreases as the standard deviation of errors increases. For example, when the localization and heading errors of OPV2V dataset are over " + }, + { + "bbox": [ + 45, + 543, + 280, + 729 + ], + "type": "inline_equation", + "content": "0.4\\mathrm{m}" + }, + { + "bbox": [ + 45, + 543, + 280, + 729 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 45, + 543, + 280, + 729 + ], + "type": "inline_equation", + "content": "0.4^{\\circ}" + }, + { + "bbox": [ + 45, + 543, + 280, + 729 + ], + "type": "text", + "content": ", the Disconet [21] and F-Cooper [3] methods fail and perform even worse than No Collaboration. Our method exceeds previous SOTA models at all noise levels and consistently outperforms the No Collaboration baseline, clearly demonstrating the robustness of ERMVP to pose errors. It is noteworthy that for localization errors, ERMVP can maintain consistent" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 296, + 271, + 530, + 363 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 271, + 530, + 363 + ], + "spans": [ + { + "bbox": [ + 296, + 271, + 530, + 363 + ], + "type": "text", + "content": "accuracy levels in high noise environments, showing superior robustness. When the localization error is " + }, + { + "bbox": [ + 296, + 271, + 530, + 363 + ], + "type": "inline_equation", + "content": "1.0\\mathrm{m}" + }, + { + "bbox": [ + 296, + 271, + 530, + 363 + ], + "type": "text", + "content": ", it outperforms the second-best method by " + }, + { + "bbox": [ + 296, + 271, + 530, + 363 + ], + "type": "inline_equation", + "content": "29.43\\%" + }, + { + "bbox": [ + 296, + 271, + 530, + 363 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 296, + 271, + 530, + 363 + ], + "type": "inline_equation", + "content": "8.08\\%" + }, + { + "bbox": [ + 296, + 271, + 530, + 363 + ], + "type": "text", + "content": " in AP@0.7 on OPV2V and V2V4Real datasets. The reasonable explanations are: (i) feature spatial calibration aligns mismatched collaborative features; (ii) accuracy enhanced feature interaction reduces performance degradation caused by collaborative noise." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 296, + 364, + 531, + 527 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 364, + 531, + 527 + ], + "spans": [ + { + "bbox": [ + 296, + 364, + 531, + 527 + ], + "type": "text", + "content": "Comparison of Communication Volume. To achieve pragmatic collaborative perception, it is crucial to evaluate the perceptual performance under different communication volumes. Figure 5 shows the results of the performance comparison under different bandwidth consumption conditions. Note that we do not consider any additional model/data/feature compression for a fair comparison. The results indicate that ERMVP achieves excellent perceptual performance and bandwidth consumption trade-off under all communication bandwidth conditions, consistently outperforming Where2comm [15]. At the same time, it achieves the same detection performance as the previous model [3, 38, 41, 42] with less communication volume on both real-world and simulated datasets." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 297, + 532, + 395, + 543 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 532, + 395, + 543 + ], + "spans": [ + { + "bbox": [ + 297, + 532, + 395, + 543 + ], + "type": "text", + "content": "4.4. Ablation Studies" + } + ] + } + ], + "index": 13 + }, + { + "type": "table", + "bbox": [ + 298, + 550, + 527, + 612 + ], + "blocks": [ + { + "bbox": [ + 45, + 514, + 279, + 536 + ], + "lines": [ + { + "bbox": [ + 45, + 514, + 279, + 536 + ], + "spans": [ + { + "bbox": [ + 45, + 514, + 279, + 536 + ], + "type": "text", + "content": "Table 2. Detection AP@0.7 on V2V4Real and OPV2V datasets with localization error." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 298, + 550, + 527, + 612 + ], + "lines": [ + { + "bbox": [ + 298, + 550, + 527, + 612 + ], + "spans": [ + { + "bbox": [ + 298, + 550, + 527, + 612 + ], + "type": "table", + "html": "
AFFFSCAEIV2V4RealOPV2V
51.71/24.4879.33/40.12
✓62.24/37.7084.30/61.43
✓✓64.36/39.5984.98/62.67
✓✓✓65.01/40.4585.67/66.32
", + "image_path": "69bd4eb33a169fcd8834f564b380acc1a803aa83bd6f57a98cc16dd8763cdf52.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "table_body" + } + ], + "index": 14 + }, + { + "bbox": [ + 296, + 620, + 530, + 664 + ], + "lines": [ + { + "bbox": [ + 296, + 620, + 530, + 664 + ], + "spans": [ + { + "bbox": [ + 296, + 620, + 530, + 664 + ], + "type": "text", + "content": "Table 3. Ablation study results of the proposed core components on the both datasets with noise level of 0.4/0.4. AFF: Attention-based Feature Fusion; FSC: Feature Spatial Calibration; AEI: Accuracy Enhanced Feature Interaction." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 296, + 671, + 530, + 730 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 671, + 530, + 730 + ], + "spans": [ + { + "bbox": [ + 296, + 671, + 530, + 730 + ], + "type": "text", + "content": "Effect of Core Components. Table 3 details the contribution of each core components in our ERMVP framework. The base model is the naive position-wise maximum fusion method. We then assess the impact of each component by sequentially introducing i) attention-based feature" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 277, + 747, + 300, + 756 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 277, + 747, + 300, + 756 + ], + "spans": [ + { + "bbox": [ + 277, + 747, + 300, + 756 + ], + "type": "text", + "content": "12581" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 595, + 842 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 52, + 105, + 169, + 177 + ], + "blocks": [ + { + "bbox": [ + 52, + 105, + 169, + 177 + ], + "lines": [ + { + "bbox": [ + 52, + 105, + 169, + 177 + ], + "spans": [ + { + "bbox": [ + 52, + 105, + 169, + 177 + ], + "type": "image", + "image_path": "280d2b5c3bb6552ed83f6681ca5e9ab65f17fefd5ae8bbd0b8c2945e3095852f.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 52, + 179, + 168, + 250 + ], + "blocks": [ + { + "bbox": [ + 52, + 179, + 168, + 250 + ], + "lines": [ + { + "bbox": [ + 52, + 179, + 168, + 250 + ], + "spans": [ + { + "bbox": [ + 52, + 179, + 168, + 250 + ], + "type": "image", + "image_path": "e770195a486e1bf921acc083d6a82eecd9c532e4557f8586c9120e99f4b1c3dd.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 83, + 253, + 129, + 262 + ], + "lines": [ + { + "bbox": [ + 83, + 253, + 129, + 262 + ], + "spans": [ + { + "bbox": [ + 83, + 253, + 129, + 262 + ], + "type": "text", + "content": "(a) V2VNet" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 170, + 105, + 287, + 177 + ], + "blocks": [ + { + "bbox": [ + 170, + 105, + 287, + 177 + ], + "lines": [ + { + "bbox": [ + 170, + 105, + 287, + 177 + ], + "spans": [ + { + "bbox": [ + 170, + 105, + 287, + 177 + ], + "type": "image", + "image_path": "efd2f53dc9a564fb9dde0a5412ab504de41b34a537d009f58505a6de2f4afd40.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 170, + 179, + 287, + 249 + ], + "blocks": [ + { + "bbox": [ + 170, + 179, + 287, + 249 + ], + "lines": [ + { + "bbox": [ + 170, + 179, + 287, + 249 + ], + "spans": [ + { + "bbox": [ + 170, + 179, + 287, + 249 + ], + "type": "image", + "image_path": "492842c4ba1259fbc06256306e088f9a970ff6743f2cbd8d52a3c6f44fd3ba2c.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 202, + 253, + 252, + 262 + ], + "lines": [ + { + "bbox": [ + 202, + 253, + 252, + 262 + ], + "spans": [ + { + "bbox": [ + 202, + 253, + 252, + 262 + ], + "type": "text", + "content": "(b) CoBEVT" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 288, + 105, + 405, + 177 + ], + "blocks": [ + { + "bbox": [ + 288, + 105, + 405, + 177 + ], + "lines": [ + { + "bbox": [ + 288, + 105, + 405, + 177 + ], + "spans": [ + { + "bbox": [ + 288, + 105, + 405, + 177 + ], + "type": "image", + "image_path": "aefb2e149c7c44ff1a1486876c71877e2004bf57de183a250fdfb7e67f496884.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 289, + 179, + 405, + 249 + ], + "blocks": [ + { + "bbox": [ + 289, + 179, + 405, + 249 + ], + "lines": [ + { + "bbox": [ + 289, + 179, + 405, + 249 + ], + "spans": [ + { + "bbox": [ + 289, + 179, + 405, + 249 + ], + "type": "image", + "image_path": "f4481015af92b97fb51c9f6721b939b1b571c6f3d179fb796d81053fc607f9bc.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 321, + 253, + 369, + 262 + ], + "lines": [ + { + "bbox": [ + 321, + 253, + 369, + 262 + ], + "spans": [ + { + "bbox": [ + 321, + 253, + 369, + 262 + ], + "type": "text", + "content": "(c) V2X-ViT" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 45, + 271, + 531, + 293 + ], + "lines": [ + { + "bbox": [ + 45, + 271, + 531, + 293 + ], + "spans": [ + { + "bbox": [ + 45, + 271, + 531, + 293 + ], + "type": "text", + "content": "Figure 6. Qualitative comparison results in real-world scenarios from the V2V4Real dataset with noise level of 0.4/0.4. Green and red boxes denote the ground truths and detection results, respectively." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 406, + 105, + 525, + 177 + ], + "blocks": [ + { + "bbox": [ + 406, + 105, + 525, + 177 + ], + "lines": [ + { + "bbox": [ + 406, + 105, + 525, + 177 + ], + "spans": [ + { + "bbox": [ + 406, + 105, + 525, + 177 + ], + "type": "image", + "image_path": "bba47526b06f4a957dc1c460977993b17132c871bd63b9b4bf3549c7e3f00d7f.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 407, + 179, + 524, + 249 + ], + "blocks": [ + { + "bbox": [ + 407, + 179, + 524, + 249 + ], + "lines": [ + { + "bbox": [ + 407, + 179, + 524, + 249 + ], + "spans": [ + { + "bbox": [ + 407, + 179, + 524, + 249 + ], + "type": "image", + "image_path": "cdb2cc3bba2d1bd3841d0f2755d0b7379f1b9c64db023463ded86abc01c94515.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 439, + 253, + 486, + 262 + ], + "lines": [ + { + "bbox": [ + 439, + 253, + 486, + 262 + ], + "spans": [ + { + "bbox": [ + 439, + 253, + 486, + 262 + ], + "type": "text", + "content": "(d)ERMVP" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 45, + 299, + 279, + 383 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 299, + 279, + 383 + ], + "spans": [ + { + "bbox": [ + 45, + 299, + 279, + 383 + ], + "type": "text", + "content": "fusion (AFF), ii) feature spatial calibration (FSC), iii) accuracy enhanced feature interaction (AEF). The consistent rise in detection results over both datasets demonstrates the effectiveness of each introduced component. Notably, integrating all three components boosts detection performance by " + }, + { + "bbox": [ + 45, + 299, + 279, + 383 + ], + "type": "inline_equation", + "content": "13.3\\%" + }, + { + "bbox": [ + 45, + 299, + 279, + 383 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 45, + 299, + 279, + 383 + ], + "type": "inline_equation", + "content": "15.97\\%" + }, + { + "bbox": [ + 45, + 299, + 279, + 383 + ], + "type": "text", + "content": " on the V2V4Real dataset for AP@0.5 and AP@0.7, respectively." + } + ] + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 53, + 390, + 159, + 482 + ], + "blocks": [ + { + "bbox": [ + 53, + 390, + 159, + 482 + ], + "lines": [ + { + "bbox": [ + 53, + 390, + 159, + 482 + ], + "spans": [ + { + "bbox": [ + 53, + 390, + 159, + 482 + ], + "type": "image", + "image_path": "b92763145d0f8cd9c12d9dd5457a2982c98ae21f38860f7e31289167e94214e1.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 80, + 490, + 244, + 500 + ], + "lines": [ + { + "bbox": [ + 80, + 490, + 244, + 500 + ], + "spans": [ + { + "bbox": [ + 80, + 490, + 244, + 500 + ], + "type": "text", + "content": "Figure 7. Hyperparameter analysis of " + }, + { + "bbox": [ + 80, + 490, + 244, + 500 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 80, + 490, + 244, + 500 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 80, + 490, + 244, + 500 + ], + "type": "inline_equation", + "content": "\\beta" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 167, + 390, + 273, + 482 + ], + "blocks": [ + { + "bbox": [ + 167, + 390, + 273, + 482 + ], + "lines": [ + { + "bbox": [ + 167, + 390, + 273, + 482 + ], + "spans": [ + { + "bbox": [ + 167, + 390, + 273, + 482 + ], + "type": "image", + "image_path": "d71922f3b4b2034784cf84180fa4aae8983c8e9ae7ee30960da71fb8f0ab96ef.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "bbox": [ + 45, + 508, + 281, + 730 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 508, + 281, + 730 + ], + "spans": [ + { + "bbox": [ + 45, + 508, + 281, + 730 + ], + "type": "text", + "content": "Hyperparameters " + }, + { + "bbox": [ + 45, + 508, + 281, + 730 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 45, + 508, + 281, + 730 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 45, + 508, + 281, + 730 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 45, + 508, + 281, + 730 + ], + "type": "text", + "content": ". We explore the impact of the hyperparameters " + }, + { + "bbox": [ + 45, + 508, + 281, + 730 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 45, + 508, + 281, + 730 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 45, + 508, + 281, + 730 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 45, + 508, + 281, + 730 + ], + "type": "text", + "content": " (as described in Section 3.2), which jointly determine the communication volume. Specifically, the feature compression ratio is represented as " + }, + { + "bbox": [ + 45, + 508, + 281, + 730 + ], + "type": "inline_equation", + "content": "\\alpha \\times \\beta" + }, + { + "bbox": [ + 45, + 508, + 281, + 730 + ], + "type": "text", + "content": ". The communication volume, in turn, is expressed as the product of the original feature size and the compression rate. By modulating the values of " + }, + { + "bbox": [ + 45, + 508, + 281, + 730 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 45, + 508, + 281, + 730 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 45, + 508, + 281, + 730 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 45, + 508, + 281, + 730 + ], + "type": "text", + "content": " within the interval of " + }, + { + "bbox": [ + 45, + 508, + 281, + 730 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 45, + 508, + 281, + 730 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 45, + 508, + 281, + 730 + ], + "type": "inline_equation", + "content": "80\\%" + }, + { + "bbox": [ + 45, + 508, + 281, + 730 + ], + "type": "text", + "content": ", we chart the resultant object detection performance (measured by AP@0.7) in Figure 7. As expected, a drop in either " + }, + { + "bbox": [ + 45, + 508, + 281, + 730 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 45, + 508, + 281, + 730 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 45, + 508, + 281, + 730 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 45, + 508, + 281, + 730 + ], + "type": "text", + "content": " will impair ERMVP's efficacy. Yet, it's worth noting that ERMVP's performance generally exhibits robustness to these hyperparameters. The performance decline remains fairly minimal even under significant compression, such as a factor of 100. The results demonstrate the effectiveness of our proposed filter and merge feature sampling module. Therefore, ERMVP establishes an appropriate and effective communication paradigm among vehicles, ensuring an optimal tradeoff between perception and communication." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 297, + 299, + 422, + 311 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 299, + 422, + 311 + ], + "spans": [ + { + "bbox": [ + 297, + 299, + 422, + 311 + ], + "type": "text", + "content": "4.5. Qualitative Evaluation" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 296, + 317, + 530, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 317, + 530, + 491 + ], + "spans": [ + { + "bbox": [ + 296, + 317, + 530, + 491 + ], + "type": "text", + "content": "Figure 6 displays the detection results of V2VNet [38], CoBEVT [41], V2X-ViT [42], and ERMVP in two scenarios. Clearly, ERMVP offers more precise and comprehensive detection than the prior SOTA methods. To begin with, ERMVP produces a greater number of predicted bounding boxes that are well aligned with the ground truths, while other methods show significant discrepancies. This demonstrates ERMVP's robustness, particularly in high-noise environments. Moreover, ERMVP detects more dynamic objects (more ground truth bounding boxes find matches). This suggests that ERMVP can effectively combine the inputs from nearby vehicles, leading to a thorough scene representation. Overall, these qualitative assessments confirm ERMVP's strengths in delivering accurate and comprehensive perception, especially in challenging conditions." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 297, + 498, + 367, + 510 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 498, + 367, + 510 + ], + "spans": [ + { + "bbox": [ + 297, + 498, + 367, + 510 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 296, + 517, + 530, + 634 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 517, + 530, + 634 + ], + "spans": [ + { + "bbox": [ + 296, + 517, + 530, + 634 + ], + "type": "text", + "content": "In this paper, we present ERMVP, a communication-efficient and collaboration-robust multi-vehicle perception method in challenging environments. We introduce a filter and merge feature sampling strategy to efficient communication, a feature spatial calibration module to align features in fine-grid and two spatial information aggregation components to solve the bottleneck of the fusion method. Extensive experiments prove the superiority of ERMVP and the effectiveness of our components. The feasibility of our approach for other modalities will be explored in future work." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 297, + 640, + 392, + 654 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 640, + 392, + 654 + ], + "spans": [ + { + "bbox": [ + 297, + 640, + 392, + 654 + ], + "type": "text", + "content": "Acknowledgments" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 296, + 660, + 530, + 730 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 660, + 530, + 730 + ], + "spans": [ + { + "bbox": [ + 296, + 660, + 530, + 730 + ], + "type": "text", + "content": "This work is supported by the Shanghai Key Research Laboratory of NSAI, the Innovation Platform for Academicians of Hainan Province, Haikou, 570228, the Specific Research Fund of the Innovation Platform for Academicians of Hainan Province under Grant YSPTZX202314, and NFSC under grant 62250410368." + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 277, + 747, + 300, + 756 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 277, + 747, + 300, + 756 + ], + "spans": [ + { + "bbox": [ + 277, + 747, + 300, + 756 + ], + "type": "text", + "content": "12582" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 595, + 842 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 105, + 103, + 117 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 105, + 103, + 117 + ], + "spans": [ + { + "bbox": [ + 47, + 105, + 103, + 117 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 124, + 279, + 728 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 51, + 124, + 279, + 155 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 51, + 124, + 279, + 155 + ], + "spans": [ + { + "bbox": [ + 51, + 124, + 279, + 155 + ], + "type": "text", + "content": "[1] Jimmy Lei Ba, Jamie Ryan Kiros, and Geoffrey E Hinton. Layer normalization. arXiv preprint arXiv:1607.06450, 2016. 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 51, + 158, + 279, + 210 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 51, + 158, + 279, + 210 + ], + "spans": [ + { + "bbox": [ + 51, + 158, + 279, + 210 + ], + "type": "text", + "content": "[2] Marwa Chafii, Salmane Naoumi, Reda Alami, Ebtesam Al-mazrouei, Mehdi Bennis, and Merouane Debbah. Emergent communication in multi-agent reinforcement learning for future wireless networks. arXiv preprint arXiv:2309.06021, 2023. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 51, + 213, + 279, + 266 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 51, + 213, + 279, + 266 + ], + "spans": [ + { + "bbox": [ + 51, + 213, + 279, + 266 + ], + "type": "text", + "content": "[3] Qi Chen, Xu Ma, Sihai Tang, Jingda Guo, Qing Yang, and Song Fu. F-cooper: Feature based cooperative perception for autonomous vehicle edge computing system using 3d point clouds. In Proceedings of the 4th ACM/IEEE Symposium on Edge Computing, pages 88-100, 2019. 1, 6, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 51, + 267, + 279, + 310 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 51, + 267, + 279, + 310 + ], + "spans": [ + { + "bbox": [ + 51, + 267, + 279, + 310 + ], + "type": "text", + "content": "[4] Zhaoyu Chen, Bo Li, Shuang Wu, Jianghe Xu, Shouhong Ding, and Wenqiang Zhang. Shape matters: deformable patch attack. In European conference on computer vision, pages 529-548. Springer, 2022. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 51, + 312, + 279, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 51, + 312, + 279, + 354 + ], + "spans": [ + { + "bbox": [ + 51, + 312, + 279, + 354 + ], + "type": "text", + "content": "[5] Zhaoyu Chen, Bo Li, Shuang Wu, Shouhong Ding, and Wenqiang Zhang. Query-efficient decision-based black-box patch attack. IEEE Transactions on Information Forensics and Security, 2023." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 51, + 356, + 279, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 51, + 356, + 279, + 398 + ], + "spans": [ + { + "bbox": [ + 51, + 356, + 279, + 398 + ], + "type": "text", + "content": "[6] Zhaoyu Chen, Bo Li, Shuang Wu, Kaixun Jiang, Shouhong Ding, and Wenqiang Zhang. Content-based unrestricted adversarial attack. Advances in Neural Information Processing Systems, 36, 2024. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 51, + 400, + 279, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 51, + 400, + 279, + 442 + ], + "spans": [ + { + "bbox": [ + 51, + 400, + 279, + 442 + ], + "type": "text", + "content": "[7] Yimian Dai, Fabian Gieseke, Stefan Oehmcke, Yiquan Wu, and Kobus Barnard. Attentional feature fusion. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 3560-3569, 2021. 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 51, + 444, + 279, + 486 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 51, + 444, + 279, + 486 + ], + "spans": [ + { + "bbox": [ + 51, + 444, + 279, + 486 + ], + "type": "text", + "content": "[8] Alexey Dosovitskiy, German Ros, Felipe Codevilla, Antonio Lopez, and Vladlen Koltun. Carla: An open urban driving simulator. In Conference on robot learning, pages 1-16. PMLR, 2017. 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 51, + 489, + 279, + 551 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 51, + 489, + 279, + 551 + ], + "spans": [ + { + "bbox": [ + 51, + 489, + 279, + 551 + ], + "type": "text", + "content": "[9] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 554, + 279, + 575 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 554, + 279, + 575 + ], + "spans": [ + { + "bbox": [ + 47, + 554, + 279, + 575 + ], + "type": "text", + "content": "[10] Ian Goodfellow, Yoshua Bengio, and Aaron Courville. Deep learning. MIT press, 2016. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 577, + 279, + 640 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 577, + 279, + 640 + ], + "spans": [ + { + "bbox": [ + 47, + 577, + 279, + 640 + ], + "type": "text", + "content": "[11] Jiaming Gu, Jingyu Zhang, Muyang Zhang, Weiliang Meng, Shibiao Xu, Jiguang Zhang, and Xiaopeng Zhang. Feaco: Reaching robust feature-level consensus in noisy pose conditions. In Proceedings of the 31st ACM International Conference on Multimedia, page 3628-3636, New York, NY, USA, 2023. Association for Computing Machinery. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 47, + 643, + 279, + 684 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 643, + 279, + 684 + ], + "spans": [ + { + "bbox": [ + 47, + 643, + 279, + 684 + ], + "type": "text", + "content": "[12] Monowar Hasan, Sibin Mohan, Takayuki Shimizu, and Hongsheng Lu. Securing vehicle-to-everything (v2x) communication platforms. IEEE Transactions on Intelligent Vehicles, 5(4):693-713, 2020. 1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 47, + 687, + 279, + 728 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 687, + 279, + 728 + ], + "spans": [ + { + "bbox": [ + 47, + 687, + 279, + 728 + ], + "type": "text", + "content": "[13] Ali Hatamizadeh, Hongxu Yin, Greg Heinrich, Jan Kautz, and Pavlo Molchanov. Global context vision transformers. In International Conference on Machine Learning, pages 12633-12646. PMLR, 2023. 5" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 299, + 106, + 530, + 729 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 299, + 106, + 530, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 106, + 530, + 149 + ], + "spans": [ + { + "bbox": [ + 299, + 106, + 530, + 149 + ], + "type": "text", + "content": "[14] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 5" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 299, + 151, + 530, + 203 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 151, + 530, + 203 + ], + "spans": [ + { + "bbox": [ + 299, + 151, + 530, + 203 + ], + "type": "text", + "content": "[15] Yue Hu, Shaoheng Fang and Zixing Lei, Yiqi Zhong, and Siheng Chen. Where2comm: Communication-efficient collaborative perception via spatial confidence maps. In Thirty-sixth Conference on Neural Information Processing Systems (NeurIPS), 2022. 1, 3, 5, 6, 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 299, + 206, + 530, + 267 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 206, + 530, + 267 + ], + "spans": [ + { + "bbox": [ + 299, + 206, + 530, + 267 + ], + "type": "text", + "content": "[16] Kaixun Jiang, Zhaoyu Chen, Hao Huang, Jiafeng Wang, Dingkang Yang, Bo Li, Yan Wang, and Wenqiang Zhang. Efficient decision-based black-box patch attacks on video recognition. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 4379-4389, 2023. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 299, + 271, + 530, + 312 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 271, + 530, + 312 + ], + "spans": [ + { + "bbox": [ + 299, + 271, + 530, + 312 + ], + "type": "text", + "content": "[17] Hamza Khan, Sumudu Samarakoon, and Mehdi Bennis. Enhancing video streaming in vehicular networks via resource slicing. IEEE Transactions on Vehicular Technology, 69(4): 3513-3522, 2020. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 299, + 314, + 530, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 314, + 530, + 346 + ], + "spans": [ + { + "bbox": [ + 299, + 314, + 530, + 346 + ], + "type": "text", + "content": "[18] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In International Conference on Learning Representations (ICLR), 2015. 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 299, + 348, + 530, + 400 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 348, + 530, + 400 + ], + "spans": [ + { + "bbox": [ + 299, + 348, + 530, + 400 + ], + "type": "text", + "content": "[19] Alex H Lang, Sourabh Vora, Holger Caesar, Lubing Zhou, Jiong Yang, and Oscar Beijbom. Pointpillars: Fast encoders for object detection from point clouds. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12697-12705, 2019. 3, 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 299, + 402, + 530, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 402, + 530, + 443 + ], + "spans": [ + { + "bbox": [ + 299, + 402, + 530, + 443 + ], + "type": "text", + "content": "[20] Yiming Li, Bir Bhanu, and Wei Lin. Auction protocol for camera active control. In 2010 IEEE International Conference on Image Processing, pages 4325-4328. IEEE, 2010. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 299, + 447, + 530, + 488 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 447, + 530, + 488 + ], + "spans": [ + { + "bbox": [ + 299, + 447, + 530, + 488 + ], + "type": "text", + "content": "[21] Yiming Li, Shunli Ren, Pengxiang Wu, Siheng Chen, Chen Feng, and Wenjun Zhang. Learning distilled collaboration graph for multi-agent perception. Advances in Neural Information Processing Systems, 34:29541-29552, 2021. 5, 6, 7" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 299, + 490, + 530, + 531 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 490, + 530, + 531 + ], + "spans": [ + { + "bbox": [ + 299, + 490, + 530, + 531 + ], + "type": "text", + "content": "[22] Yiming Li, Ziyan An, Zixun Wang, Yiqi Zhong, Siheng Chen, and Chen Feng. V2x-sim: A virtual collaborative perception dataset for autonomous driving. arXiv preprint arXiv:2202.08449, 2022. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 299, + 533, + 530, + 575 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 533, + 530, + 575 + ], + "spans": [ + { + "bbox": [ + 299, + 533, + 530, + 575 + ], + "type": "text", + "content": "[23] Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dollar. Focal loss for dense object detection. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), 2017. 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 299, + 577, + 530, + 630 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 577, + 530, + 630 + ], + "spans": [ + { + "bbox": [ + 299, + 577, + 530, + 630 + ], + "type": "text", + "content": "[24] Yang Liu, Jing Liu, Kun Yang, Bobo Ju, Siao Liu, Yuzheng Wang, Dingkang Yang, Peng Sun, and Liang Song. Ampnet: Appearance-motion prototype network assisted automatic video anomaly detection system. IEEE Transactions on Industrial Informatics, 20(2):2843-2855, 2024. 1" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 299, + 632, + 530, + 684 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 632, + 530, + 684 + ], + "spans": [ + { + "bbox": [ + 299, + 632, + 530, + 684 + ], + "type": "text", + "content": "[25] Yifan Lu, Quanhao Li, Baoan Liu, Mehrdad Dianati, Chen Feng, Siheng Chen, and Yanfeng Wang. Robust collaborative 3d object detection in presence of pose errors. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 4812-4818. IEEE, 2023. 1, 6" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 299, + 687, + 530, + 729 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 687, + 530, + 729 + ], + "spans": [ + { + "bbox": [ + 299, + 687, + 530, + 729 + ], + "type": "text", + "content": "[26] Zonglin Meng, Xin Xia, Runsheng Xu, Wei Liu, and Jiaqi Ma. Hydro-3d: Hybrid object detection and tracking for cooperative perception using 3d lidar. IEEE Transactions on Intelligent Vehicles, 2023. 1" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 277, + 747, + 300, + 756 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 277, + 747, + 300, + 756 + ], + "spans": [ + { + "bbox": [ + 277, + 747, + 300, + 756 + ], + "type": "text", + "content": "12583" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 595, + 842 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 106, + 279, + 729 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 47, + 106, + 279, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 106, + 279, + 138 + ], + "spans": [ + { + "bbox": [ + 47, + 106, + 279, + 138 + ], + "type": "text", + "content": "[27] Yaru Niu, Rohan R Paleja, and Matthew C Gombolay. Multiagent graph-attention communication and teaming. In AA-MAS, pages 964–973, 2021. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 141, + 279, + 204 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 141, + 279, + 204 + ], + "spans": [ + { + "bbox": [ + 47, + 141, + 279, + 204 + ], + "type": "text", + "content": "[28] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. Advances in neural information processing systems, 32, 2019. 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 207, + 279, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 207, + 279, + 239 + ], + "spans": [ + { + "bbox": [ + 47, + 207, + 279, + 239 + ], + "type": "text", + "content": "[29] Faisal Qureshi and Demetri Terzopoulos. Smart camera networks in virtual reality. Proceedings of the IEEE, 96(10): 1640-1656, 2008. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 241, + 279, + 305 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 241, + 279, + 305 + ], + "spans": [ + { + "bbox": [ + 47, + 241, + 279, + 305 + ], + "type": "text", + "content": "[30] Qingyu Song, Changan Wang, Zhengkai Jiang, Yabiao Wang, Ying Tai, Chengjie Wang, Jilin Li, Feiyue Huang, and Yang Wu. Rethinking counting and localization in crowds: A purely point-based framework. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3365-3374, 2021. 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 307, + 279, + 349 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 307, + 279, + 349 + ], + "spans": [ + { + "bbox": [ + 47, + 307, + 279, + 349 + ], + "type": "text", + "content": "[31] Zhiying Song, Fuxi Wen, Hailiang Zhang, and Jun Li. An efficient and robust object-level cooperative perception framework for connected and automated driving. arXiv preprint arXiv:2210.06289, 2022. 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 352, + 279, + 394 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 352, + 279, + 394 + ], + "spans": [ + { + "bbox": [ + 47, + 352, + 279, + 394 + ], + "type": "text", + "content": "[32] Ming Tan. Multi-agent reinforcement learning: Independent vs. cooperative agents. In Proceedings of the tenth international conference on machine learning, pages 330-337, 1993. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 396, + 279, + 429 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 396, + 279, + 429 + ], + "spans": [ + { + "bbox": [ + 47, + 396, + 279, + 429 + ], + "type": "text", + "content": "[33] Mingxing Tan and Quoc Le. Efficientnetv2: Smaller models and faster training. In International conference on machine learning, pages 10096-10106. PMLR, 2021. 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 431, + 279, + 473 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 431, + 279, + 473 + ], + "spans": [ + { + "bbox": [ + 47, + 431, + 279, + 473 + ], + "type": "text", + "content": "[34] Nicholas Vadivelu, Mengye Ren, James Tu, Jingkang Wang, and Raquel Urtasun. Learning to communicate and correct pose errors. In Conference on Robot Learning, pages 1195-1210. PMLR, 2021. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 475, + 279, + 528 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 475, + 279, + 528 + ], + "spans": [ + { + "bbox": [ + 47, + 475, + 279, + 528 + ], + "type": "text", + "content": "[35] Binglu Wang, Lei Zhang, Zhaozhong Wang, Yongqiang Zhao, and Tianfei Zhou. Core: Cooperative reconstruction for multi-agent perception. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8710-8720, 2023. 1, 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 531, + 279, + 562 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 531, + 279, + 562 + ], + "spans": [ + { + "bbox": [ + 47, + 531, + 279, + 562 + ], + "type": "text", + "content": "[36] Jian Wang, Yameng Shao, Yuming Ge, and Rundong Yu. A survey of vehicle to everything (v2x) testing. Sensors, 19(2): 334, 2019. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 564, + 279, + 618 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 564, + 279, + 618 + ], + "spans": [ + { + "bbox": [ + 47, + 564, + 279, + 618 + ], + "type": "text", + "content": "[37] Tianhang Wang, Guang Chen, Kai Chen, Zhengfa Liu, Bo Zhang, Alois Knoll, and Changjun Jiang. Umc: A unified bandwidth-efficient and multi-resolution based collaborative perception framework. arXiv preprint arXiv:2303.12400, 2023. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 620, + 279, + 674 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 620, + 279, + 674 + ], + "spans": [ + { + "bbox": [ + 47, + 620, + 279, + 674 + ], + "type": "text", + "content": "[38] Tsun-Hsuan Wang, Sivabanan Manivasagam, Ming Liang, Bin Yang, Wenyuan Zeng, and Raquel Urtasun. V2vnet: Vehicle-to-vehicle communication for joint perception and prediction. In European Conference on Computer Vision, pages 605–621. Springer, 2020. 1, 3, 5, 6, 7, 8" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 47, + 676, + 279, + 729 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 676, + 279, + 729 + ], + "spans": [ + { + "bbox": [ + 47, + 676, + 279, + 729 + ], + "type": "text", + "content": "[39] Qianyi Wu, Kaisiyuan Wang, Kejie Li, Jianmin Zheng, and Jianfei Cai. Objectsdf++: Improved object-compositional neural implicit surfaces. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 21764-21774, 2023. 1" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 299, + 106, + 530, + 658 + ], + "type": "list", + "angle": 0, + "index": 24, + "blocks": [ + { + "bbox": [ + 299, + 106, + 530, + 159 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 106, + 530, + 159 + ], + "spans": [ + { + "bbox": [ + 299, + 106, + 530, + 159 + ], + "type": "text", + "content": "[40] Runsheng Xu, Yi Guo, Xu Han, Xin Xia, Hao Xiang, and Jiaqi Ma. Opencda: an open cooperative driving automation framework integrated with co-simulation. In 2021 IEEE International Intelligent Transportation Systems Conference (ITSC), pages 1155-1162. IEEE, 2021. 6" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 299, + 161, + 530, + 203 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 161, + 530, + 203 + ], + "spans": [ + { + "bbox": [ + 299, + 161, + 530, + 203 + ], + "type": "text", + "content": "[41] Runsheng Xu, Zhengzhong Tu, Hao Xiang, Wei Shao, Bolei Zhou, and Jiaqi Ma. Cobevt: Cooperative bird's eye view semantic segmentation with sparse transformers, 2022. 1, 3, 6, 7, 8" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 299, + 205, + 530, + 257 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 205, + 530, + 257 + ], + "spans": [ + { + "bbox": [ + 299, + 205, + 530, + 257 + ], + "type": "text", + "content": "[42] Runsheng Xu, Hao Xiang, Zhengzhong Tu, Xin Xia, Ming-Hsuan Yang, and Jiaqi Ma. V2x-vit: Vehicle-to-everything cooperative perception with vision transformer. In European Conference on Computer Vision, page 107-124. Springer, 2022. 1, 3, 5, 6, 7, 8" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 299, + 259, + 530, + 320 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 259, + 530, + 320 + ], + "spans": [ + { + "bbox": [ + 299, + 259, + 530, + 320 + ], + "type": "text", + "content": "[43] Runsheng Xu, Hao Xiang, Xin Xia, Xu Han, Jinlong Li, and Jiaqi Ma. Opv2v: An open benchmark dataset and fusion pipeline for perception with vehicle-to-vehicle communication. In 2022 International Conference on Robotics and Automation (ICRA), pages 2583–2589. IEEE, 2022. 2, 3, 5, 6, 7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 299, + 324, + 530, + 388 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 324, + 530, + 388 + ], + "spans": [ + { + "bbox": [ + 299, + 324, + 530, + 388 + ], + "type": "text", + "content": "[44] Runsheng Xu, Xin Xia, Jinlong Li, Hanzhao Li, Shuo Zhang, Zhengzhong Tu, Zonglin Meng, Hao Xiang, Xiaoyu Dong, Rui Song, et al. V2v4real: A real-world large-scale dataset for vehicle-to-vehicle cooperative perception. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13712–13722, 2023. 2, 3, 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 299, + 389, + 530, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 389, + 530, + 442 + ], + "spans": [ + { + "bbox": [ + 299, + 389, + 530, + 442 + ], + "type": "text", + "content": "[45] Kun Yang, Peng Sun, Jieyu Lin, Azzedine Boukerche, and Liang Song. A novel distributed task scheduling framework for supporting vehicular edge intelligence. In 2022 IEEE 42nd International Conference on Distributed Computing Systems (ICDCS), pages 972-982. IEEE, 2022. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 299, + 444, + 530, + 485 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 444, + 530, + 485 + ], + "spans": [ + { + "bbox": [ + 299, + 444, + 530, + 485 + ], + "type": "text", + "content": "[46] Kun Yang, Dingkang Yang, Jingyu Zhang, Mingcheng Li, Yang Liu, Jing Liu, Hanqi Wang, Peng Sun, and Liang Song. Spatio-temporal domain awareness for multi-agent collaborative perception, 2023. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 299, + 487, + 530, + 550 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 487, + 530, + 550 + ], + "spans": [ + { + "bbox": [ + 299, + 487, + 530, + 550 + ], + "type": "text", + "content": "[47] Kun Yang, Dingkang Yang, Jingyu Zhang, Hanqi Wang, Peng Sun, and Liang Song. What2comm: Towards communication-efficient collaborative perception via feature decoupling. In Proceedings of the 31st ACM International Conference on Multimedia, page 7686-7695, New York, NY, USA, 2023. Association for Computing Machinery. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 299, + 551, + 530, + 615 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 551, + 530, + 615 + ], + "spans": [ + { + "bbox": [ + 299, + 551, + 530, + 615 + ], + "type": "text", + "content": "[48] Hang Zhang, Chongruo Wu, Zhongyue Zhang, Yi Zhu, Haibin Lin, Zhi Zhang, Yue Sun, Tong He, Jonas Mueller, R. Manmatha, Mu Li, and Alexander Smola. Resnest: Split-attention networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, pages 2736-2746, 2022. 5" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 299, + 617, + 530, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 617, + 530, + 658 + ], + "spans": [ + { + "bbox": [ + 299, + 617, + 530, + 658 + ], + "type": "text", + "content": "[49] Sai Qian Zhang, Qi Zhang, and Jieyu Lin. Succinct and robust multi-agent communication with temporal message control. Advances in Neural Information Processing Systems, 33:17271-17282, 2020. 3" + } + ] + } + ], + "index": 23 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 277, + 747, + 300, + 756 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 277, + 747, + 300, + 756 + ], + "spans": [ + { + "bbox": [ + 277, + 747, + 300, + 756 + ], + "type": "text", + "content": "12584" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 595, + 842 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/ES3_ Evolving Self-Supervised Learning of Robust Audio-Visual Speech Representations/09b7366b-807f-4dfb-a9c8-45edf9472928_content_list.json b/2024/ES3_ Evolving Self-Supervised Learning of Robust Audio-Visual Speech Representations/09b7366b-807f-4dfb-a9c8-45edf9472928_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..fed2e99a8acd038d9fe1785b5d023cd69cef01e8 --- /dev/null +++ b/2024/ES3_ Evolving Self-Supervised Learning of Robust Audio-Visual Speech Representations/09b7366b-807f-4dfb-a9c8-45edf9472928_content_list.json @@ -0,0 +1,1590 @@ +[ + { + "type": "text", + "text": "$\\mathbf{ES}^{3}$ : Evolving Self-Supervised Learning of Robust Audio-Visual Speech Representations", + "text_level": 1, + "bbox": [ + 243, + 128, + 725, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yuanhang Zhang $^{1,2}$ , Shuang Yang $^{1,2}$ , Shiguang Shan $^{1,2}$ , Xilin Chen $^{1,2}$", + "bbox": [ + 209, + 202, + 759, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1Key Laboratory of Intelligent Information Processing of Chinese Academy of Sciences (CAS), Institute of Computing Technology, CAS, Beijing 100190, China", + "bbox": [ + 107, + 220, + 861, + 256 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ University of Chinese Academy of Sciences, Beijing 100049, China", + "bbox": [ + 212, + 256, + 756, + 273 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "zhangyuanhang15@mails.ucas.ac.cn, {shuang.yang, sgshan, xlchen}@ict.ac.cn", + "bbox": [ + 183, + 276, + 781, + 292 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 325, + 313, + 343 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We propose a novel strategy, $ES^3$ , for self-supervised learning of robust audio-visual speech representations from unlabeled talking face videos. While many recent approaches for this task primarily rely on guiding the learning process using the audio modality alone to capture information shared between audio and video, we reframe the problem as the acquisition of shared, unique (modality-specific) and synergistic speech information to address the inherent asymmetry between the modalities. Based on this formulation, we propose a novel \"evolving\" strategy that progressively builds joint audio-visual speech representations that are strong for both uni-modal (audio & visual) and bimodal (audio-visual) speech. First, we leverage the more easily learnable audio modality to initialize audio and visual representations by capturing audio-unique and shared speech information. Next, we incorporate video-unique speech information and bootstrap the audio-visual representations on top of the previously acquired shared knowledge. Finally, we maximize the total audio-visual speech information, including synergistic information to obtain robust and comprehensive representations. We implement $ES^3$ as a simple Siamese framework and experiments on both English benchmarks and a newly contributed large-scale Mandarin dataset show its effectiveness. In particular, on LRS2-BBC, our smallest model is on par with SoTA models with only 1/2 parameters and 1/8 unlabeled data (223h).", + "bbox": [ + 76, + 358, + 472, + 752 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 768, + 209, + 786 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The task of learning audio-visual speech representations has gained increasing importance due to its potential to support a wide range of relevant applications. Building on the success in other domains [9, 11, 12, 14, 22, 31], several effective self-supervised approaches have recently emerged for this task by pre-training models on paired or mixed audio, visual and text data with pretext tasks like speech unit", + "bbox": [ + 75, + 794, + 470, + 901 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "prediction [30, 59, 80]. However, learning representations in the audio-visual speech domain faces unique challenges brought by the inherent asymmetry between the two modalities involved. First, audio and video often convey distinct and complementary types of speech information. For example, $\\mathsf{m} /$ and $\\mathsf{n} /$ sound similar, but are visually different; on the contrary, $\\mathsf{p} /$ and $\\mathsf{b} /$ are visually similar but can be distinguished with audio [70]. Even more interesting is the presence of synergistic information – information that can further enhance speech perception [55] but only emerges when audio and video are considered together, and cannot be inferred when either modality is absent. Second, they manifest different amounts of noise and information. Specifically, in most existing audio-visual speech datasets, audio is characterized by high sampling rate and relatively clean speech samples, making audio representations readily learnable from these datasets. In contrast, the visual modality often presents several challenging factors, including missing (e.g. due to self-occlusion in profile faces or low image quality), redundant (e.g. due to appearance variations, makeup and beards etc.) and ambiguous (e.g. due to the lower sampling rate and the presence of homophones) visual information. Their presence complicates learning meaningful and robust representations solely from video. The difficulties encountered in the continuous efforts to improve the performance of visual speech recognition (VSR) [15, 24, 40, 41] models illustrate this persistent obstacle. Additionally, these differences between the two modalities lead to an imbalance between their learning dynamics, requiring additional manual settings [3, 39, 42, 61].", + "bbox": [ + 496, + 327, + 893, + 782 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this work, we break down the problem of learning audio-visual speech representations into the learning of three different types of information: shared (across modalities), unique (modality-specific) and synergistic speech information. Based on our formulation, we propose a novel strategy called $\\mathrm{ES^3}$ (short for Evolving Self-Supervised Learning of Robust Audio-Visual Speech Representations)", + "bbox": [ + 496, + 794, + 893, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "27069", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "to tackle the asymmetry between the two modalities for learning robust audio-visual speech representations. Specifically, our $\\mathrm{ES}^3$ employs a three-stage progressive learning approach. Initially, it focuses on acquiring the easily learnable audio-unique and shared speech information. Then, it addresses the more challenging visual-unique speech information while incorporating it into the initial audio-visual joint representation. Lastly, it optimizes the final audio-visual representations by leveraging synergy, on top of the acquired shared and modality-unique information.", + "bbox": [ + 75, + 90, + 472, + 242 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our contributions can be summarized as follows:", + "bbox": [ + 96, + 243, + 421, + 257 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose a new formulation for audio-visual speech representation learning by decomposing the task into the acquisition of shared (across modalities), unique (modality-specific) and synergistic speech information.", + "- We propose a novel strategy, $\\mathrm{ES}^3$ to progressively learn the three types of information in a self-supervised manner with a simple Siamese framework. This evolving strategy significantly alleviates the previous burden of tuning balancing parameters.", + "- Extensive experiments on two widely used benchmarks show that $\\mathrm{ES}^3$ rivals previous methods in auditory, visual and audio-visual speech recognition performance.", + "- We also contribute a new large-scale Mandarin audiovisual speech dataset named CAS-VSR-S101 to the community for evaluation of similar methods." + ], + "bbox": [ + 76, + 258, + 468, + 484 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 76, + 496, + 218, + 511 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Self-supervised learning from multi-modal data. The arguably most successful paradigm for self-supervised multimodal learning in the past few years is contrastive learning, which exploits the common information encoded in different modalities [68], such as audio (A) [6, 7], image (I) [53, 66], video (V) [5, 8, 32, 67] and text (T) [4, 16, 43, 44, 49, 63]. These approaches focus on learning semantically aligned multi-modal representations that are effective for specific downstream tasks like cross-modal retrieval, but discard modality-unique information that is useful for other uni-modal tasks. To solve this problem, Gong et al. [25] combine contrastive learning with the masked prediction objective [28] to learn strong audio-visual representations suitable for both retrieval and joint classification. Liang et al. [35] propose a new algorithm, FACTORCL to learn self-supervised multi-modal representations that capture both shared and unique information and verify its effectiveness on multiple tasks. In this paper, we explicitly address capturing unique, shared and synergistic information for uni- and multi-modal downstream tasks.", + "bbox": [ + 75, + 521, + 472, + 823 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Learning audio-visual speech representations. A main challenge of learning audio-visual speech representation is properly handling the aforementioned asymmetry between audio and video. An established approach is to leverage the easily learnable audio modality as a teacher to guide", + "bbox": [ + 75, + 825, + 472, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "visual or audio-visual speech representation learning. Implementing additional measures to mitigate audio dominance and promote visual features during learning is vital for this paradigm. For instance, Shi et al. [59] and Zhang et al. [76] use different masking strengths for the two modalities, apply modality dropout, and enforce simple audio features by only using a weak linear projection layer. Others have explored freezing pre-trained audio representations [2, 38, 78]. However, such an audio-centric approach is sub-optimal and inefficient. First, it limits the learned representations to capturing only shared, i.e. modality-invariant information, or at most, shared and audio-unique information. Recently, a few attempts have been made to go beyond merely using audio as the anchor modality. Haliassos et al. [27] combine multiple audio-visual prediction tasks in place of purely audio-centric ones, and treats A-V asymmetry by having the visual stream predict only auditory but not visual targets. Lian et al. [34] perform a series of trials to determine an optimal schedule for audio and video input combination. Second, the process of exhaustively testing and identifying the optimal set of balancing parameters such as masking ratios, modality schedules and modality dropout rate is extremely time-consuming and computationally demanding. In contrast, our newly proposed evolving strategy identifies different types of task-relevant information and learns them progressively, while avoiding pitfalls and computational burdens (see supp).", + "bbox": [ + 496, + 90, + 893, + 500 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3. Our Proposed ES3", + "text_level": 1, + "bbox": [ + 500, + 503, + 679, + 522 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3.1. Preliminaries", + "text_level": 1, + "bbox": [ + 500, + 530, + 640, + 544 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The goal of $\\mathrm{ES}^3$ is to learn audio-visual speech representations from unlabeled talking face videos. The inputs to our $\\mathrm{ES}^3$ are paired audio-visual data $X_{a,v} = \\{(X_{a,i},X_{v,i})\\mid i = 1,2,\\ldots ,N\\}$ . Each of the $N$ samples consists of a sequence of audio frames $X_{a,i} = \\operatorname {LogFBank}(\\mathbf{a}_i)\\in \\mathbb{R}^{F\\times T_{a,i}}$ and a temporally synchronized video sequence $X_{v,i} = \\mathbf{v}_i\\in \\mathbb{R}^{C\\times H\\times W\\times T_{v,i}}$ showing the lips of a speaker, where $F$ is the filter bank dimension, $C$ is the number of video channels, $T_{a,i}$ and $T_{v,i}$ are the total number of audio and video frames in the $i$ -th sample, and $H$ and $W$ are the width and height of each frame. Audio and video are sampled at 100 and 25 FPS, i.e. for a 1-second clip, $T_{a} = 4T_{v} = 100$ . The goal is to learn a $d$ -dimensional contextualized representation $\\gamma \\in \\mathbb{R}^{d\\times T_v}$ capturing vital speech and context information within the two modality spaces.", + "bbox": [ + 496, + 551, + 893, + 779 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Base architecture. The base architecture used in this paper is a relatively simple one and can be easily extended if needed. It includes two modality-specific encoders $\\varphi_{a}$ and $\\varphi_{v}$ , instantiated as a 2-layer CNN-based audio filter bank encoder and a lightweight VGG-(2+1)D video encoder [57, 58]. Here, $\\varphi_{a}$ downsamples $X_{a}$ by 4x to match the video frame rate; $\\varphi_{v}$ has a temporal receptive field of 5 frames, or 200ms at 25 FPS, same as the most com", + "bbox": [ + 496, + 780, + 893, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "27070", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "monly adopted 3D-ResNet18 encoder [41, 59, 64], but has much fewer parameters. A modality-agnostic Transformer-based [71] modality fusion encoder $\\theta$ equipped with rotary positional embeddings (RoPE) [65] follows the modality-specific encoders $\\varphi_{a}$ and $\\varphi_{v}$ . To allow $\\theta$ to accept either audio, video, or audio-visual features, we adopt the simple sum operation to compute a fused audio-visual representation $F_{\\mathrm{fused}}$ from modality features $F_{m} = \\varphi_{m}(X_{m})$ , where $m \\in \\{a, v\\}$ , to allow easy handling of missing modalities:", + "bbox": [ + 75, + 90, + 472, + 227 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nF _ {\\text {f u s e d}} = \\left\\{ \\begin{array}{l l} F _ {a} + F _ {v} & \\text {i f i n p u t i s a u d i o - v i s u a l ,} \\\\ F _ {a} + \\mathbf {0} = F _ {a} & \\text {i f i n p u t i s a u d i o ,} \\\\ F _ {v} + \\mathbf {0} = F _ {v} & \\text {i f i n p u t i s v i d e o .} \\end{array} \\right. \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 94, + 234, + 468, + 289 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The Transformer encoder $\\theta$ then yields the final contextual features $E = \\theta (F_{\\mathrm{fused}})$", + "bbox": [ + 75, + 296, + 468, + 325 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Based on this base architecture, we build our Siamese based learning framework as shown in Fig. 1(a). A student model (left) is guided by a teacher model (right), which share the same base architecture. The student and teacher accept different versions of the same speech inputs (e.g. masked vs unmasked). The teacher model updates its parameters in an exponential moving average (EMA) manner, and the student is trained to regress the teacher outputs in general. In the subsequent sections, our objective is to optimize the student model with the base architecture to obtain the parameterized audio-visual speech representations $\\gamma (\\mathbf{a},\\mathbf{v};\\varphi_{a},\\varphi_{v},\\theta) = E$", + "bbox": [ + 75, + 325, + 470, + 508 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Problem Reformulation", + "text_level": 1, + "bbox": [ + 76, + 513, + 295, + 529 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In the audio-visual speech representation learning task, audio and video inputs can be taken as two different views of the same underlying speech signal, carrying different types of information. The relationship between the speech information in the two modalities can be visualized as the bottom-left Venn diagram in Fig. 1(b). Owing to the asymmetry between the two modalities, there exists both modality-unique speech information, i.e. audio- and video-unique information $U_{a}$ and $U_{v}$ . There is also speech information $R$ shared between the two modalities, since they share the same underlying speech production process. At the same time, information that only emerges when both audio and video modality are present is the synergistic information, which we denote by $S$ . Our overall objective is to learn a representation that maximizes the total information $R + U_{a} + U_{v} + S$ for bi-modal (audio-visual) speech representations, as well as $R + U_{a}$ and $R + U_{v}$ for uni-modal (audio & visual) speech representations.", + "bbox": [ + 75, + 537, + 468, + 809 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "With this in mind, we reframe the task of learning audiovisual speech representations as a mutual information maximization task. Specifically, the two modalities, corresponding to source variables $X_{a}$ and $X_{v}$ , have mutual information with the target speech content variable $Y$ , containing information useful for downstream tasks that we", + "bbox": [ + 75, + 810, + 470, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "wish to capture. Denote the joint mutual information between the sources and the target by $I(X_{a},X_{v};Y)$ . The problem of learning robust audio-visual speech representations reduces to finding the set of parameters that maximizes $I(\\gamma (X_a),\\gamma (X_v);Y) = I(E_{av};Y)^1$ . Similarly, the problem of learning robust audio and visual speech representations can be viewed as maximizing the mutual information $I(\\gamma (X_a);Y) = I(E_a;Y)$ and $I(\\gamma (X_v);Y) = I(E_v;Y)$ , respectively. Taking advantage of the Partial Information Decomposition (PID) framework [73], we can precisely decompose the multivariate mutual information $I(E_{av};Y),I(E_a;Y)$ and $I(E_{v};Y)$ into the previous types of information as:", + "bbox": [ + 496, + 90, + 893, + 287 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nR + U _ {a} = I \\left(E _ {a}; Y\\right), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 640, + 300, + 890, + 316 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nR + U _ {v} = I \\left(E _ {v}; Y\\right), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 642, + 319, + 890, + 335 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nU _ {a} + S = I \\left(E _ {a}; Y \\mid X _ {v}\\right), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 642, + 338, + 890, + 353 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nU _ {v} + S = I \\left(E _ {v}; Y \\mid X _ {a}\\right), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 642, + 357, + 890, + 372 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nR - S = I \\left(E _ {a}; E _ {v}; Y\\right), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 650, + 376, + 890, + 392 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nR + U _ {a} + U _ {v} + S = I \\left(E _ {a v}; Y\\right), \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 571, + 395, + 890, + 411 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $R, U_{a}, U_{v}$ and $S$ are non-negative.", + "bbox": [ + 498, + 422, + 772, + 436 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To optimize the above mutual information terms with a neural network parameterized by $\\varphi_{a},\\varphi_{v}$ and $\\theta$ we divide the optimization process into three progressive and nonconflicting stages. In stage 1, we learn audio-unique information $U_{a}$ and shared information $R$ from the informative audio modality to maximize $I(E_{a};Y)$ . This process also helps to bootstrap visual representations by learning the shared information $R$ . In stage 2, we turn to learning visual unique information $U_{v}$ by maximizing $I(E_{v};Y) = R + U_{v}$ , which is now easier with the visual information captured via $R$ in stage 1. At the same time, we inject the learned visual information into the audio-visual joint representations to initialize $I(E_{av};Y)$ . Finally, in stage 3 we optimize the joint audio-visual representation on top of the previously learned information by maximizing $I(E_{av};Y)$ .", + "bbox": [ + 496, + 438, + 890, + 667 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.3. The Complete Learning Framework", + "text_level": 1, + "bbox": [ + 498, + 675, + 815, + 691 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To learn network parameters that maximize the mutual information terms $I(E;Y)$ , we leverage the popular self-supervised masked prediction task based on the simple Siamese framework to support the learning process via self-distillation. Inspired by the recent work of Baevski et al. [12], we create multiple masked copies for each training sample, encode the unmasked regions with the Transformer encoder, and train the student to recover contextualized targets of the masked regions with a lightweight 1D convolutional decoder, which are generated by an online EMA teacher. This task can be used to learn any mutual infor", + "bbox": [ + 496, + 698, + 890, + 864 + ], + "page_idx": 2 + }, + { + "type": "page_footnote", + "text": "1This equality holds since the architecture implemented in this work only models additive modality synergy stemming from sum fusion.", + "bbox": [ + 500, + 875, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "27071", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/3e1e85f97e7651037acf05269d30db860a88065913e5f669cb25d5be1075bfef.jpg", + "image_caption": [ + "Figure 1. Overview of the proposed method, $\\mathbf{ES}^3$ . (a) We use a Siamese framework with an EMA teacher to implement an iteration-free training procedure. (b) Venn diagrams showing how the three types of information $(R, U, \\text{and } S)$ are acquired in an evolving learning process. (c) Illustration of self-distillation with online clustering corresponding to $\\mathcal{L}_{\\mathrm{distill}}(F_{\\mathrm{tea}}, F_{\\mathrm{stu}})$ ." + ], + "image_footnote": [], + "bbox": [ + 83, + 92, + 890, + 385 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "mation $I(E_{;}Y)$ when $Y$ is the downstream speech understanding task. For example, training on auditory or visual speech acts as a proxy to maximize $I(E_{a};Y)$ or $I(E_{v};Y)$ . Self-distillation with online clustering [36]. We perform distillation from the right-hand side teacher to the left-hand side student in Fig. 1(a). Specifically, as illustrated in Fig. 1(c), for each layer $k$ of the top $K$ Transformer layers in $\\theta$ with hidden dimension $D$ , we learn a codebook $\\mathbf{Q}^{k} = \\{\\mathbf{q}_{1}^{k},\\dots,\\mathbf{q}_{Q}^{k}\\}$ with $Q$ codewords $\\mathbf{q}_i^k\\in \\mathbb{R}^D$ , which is updated as follows: for each entry $j$ , a set $\\tilde{\\mathbf{Z}}_{\\mathrm{tea}}^k$ of the teacher output features $\\tilde{\\mathbf{z}}_{\\mathrm{tea}}^k$ closest to the current representation of $q$ in the codebook is found as", + "bbox": [ + 75, + 438, + 468, + 622 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {\\mathbf {Z}} _ {\\mathrm {t e a}} ^ {k} = \\left\\{\\tilde {\\boldsymbol {z}} _ {\\mathrm {t e a}} ^ {k} \\Bigg | q = \\underset {j \\in \\{1, 2, \\dots , Q \\}} {\\arg \\min } | | \\tilde {\\boldsymbol {z}} _ {\\mathrm {t e a}} ^ {k} - \\mathbf {q} _ {j} ^ {k} | | _ {2} \\right\\}, \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 99, + 632, + 468, + 667 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where the set indices $q$ are used as discrete pseudo-labels to train the student model. Codewords are then updated in an EMA fashion as a weighted sum of the embeddings in $\\tilde{\\mathbf{Z}}_{\\mathrm{tea}}^k$ with decay rate $\\tau_{\\mathrm{code}}$ . The associated loss function is", + "bbox": [ + 76, + 676, + 468, + 738 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {d i s t i l l}} \\left(F _ {\\text {t e a}}, F _ {\\text {s t u}}\\right) = \\sum_ {t \\in \\mathbb {M}} \\sum_ {k \\in (K - N, K ]} \\log p _ {\\psi_ {k}} \\left(q _ {t} \\mid \\boldsymbol {z} _ {t} ^ {K}\\right), \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 93, + 747, + 468, + 792 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $F_{\\mathrm{tea}}$ and $F_{\\mathrm{stu}}$ are teacher and student features, $\\mathbb{M}$ is the set of masked time steps, $\\psi_{k}$ is the softmax activations of linear projections for each target layer $k$ , $q_{t}$ is the codeword index for time step $t$ of the teacher $\\tilde{\\theta}$ , and $z_{t}^{K}$ is the output of the $K$ -th and final layer of the student model $\\theta$ . The teacher and the online clustering process do not require gradients. EMA teacher parameterization. To reduce GPU memory", + "bbox": [ + 75, + 795, + 468, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "usage, we apply EMA updates to the Transformer encoder $\\theta$ , but not the earlier modality-specific encoders $\\varphi_{a}$ and $\\varphi_{v}$ , inspired by [50]. As shown in Fig. 1, the teacher model weights $\\tilde{\\theta}$ are computed as an exponential moving average of the student model weights $\\theta$ :", + "bbox": [ + 498, + 439, + 890, + 513 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {\\theta} \\leftarrow \\eta \\tilde {\\theta} + (1 - \\eta) \\theta , \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 625, + 523, + 890, + 542 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\eta = \\min \\{\\eta_{\\max}, (1 + \\text{step}) / (10 + \\text{step})\\}$ is a momentum parameter that slowly increases over time until a pre-determined value $\\eta_{\\max}$ is reached.", + "bbox": [ + 498, + 554, + 890, + 599 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.4. The Learning Process", + "text_level": 1, + "bbox": [ + 500, + 609, + 702, + 626 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We now elaborate on individual learning stages, based on the above Siamese framework and distillation loss.", + "bbox": [ + 498, + 633, + 890, + 662 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Stage 1: Learning audio-unique information and shared information. In stage 1, we perform distillation from un-masked audio teacher model to the masked audio student model, yielding audio-unique information $U_{a}$ and shared information $R$ from audio. The distillation loss is:", + "bbox": [ + 498, + 662, + 890, + 737 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {a} \\rightarrow \\mathrm {a} ^ {\\prime}} = \\mathcal {L} _ {\\text {d i s t i l l}} \\left(E _ {a}, E _ {a} ^ {\\prime}\\right). \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 606, + 751, + 890, + 767 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Being readily learnable, audio has the potential to provide most of $R$ for learning shared information. However, optimizing $I(E_{a};Y)$ alone cannot help visual learning. Noting that the InfoNCE loss between two variables has been widely used to maximize a lower bound on the mutual information, we adopt the InfoNCE loss to learn shared information corresponding to $I(E_{a};E_{v})$ to bootstrap visual learning with the shared information learned from audio.", + "bbox": [ + 496, + 779, + 890, + 898 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "27072", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Specifically, we apply it to a fine-grained frame-wise alignment task, where temporally aligned features from the same video are treated as positives, and unaligned ones negatives [19]. Since phonemes can appear repeatedly within the same video clip, leading to false negatives, we use a negative margin following Park et al. [47], which alleviates the effect of falsely matched pairs by enforcing a looser decision boundary. The objective function involved is:", + "bbox": [ + 75, + 90, + 472, + 212 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {N C E}} (H) = - \\frac {1}{n} \\sum_ {i = 1} ^ {n} \\log \\frac {e ^ {(H _ {i , i} - m) / \\tau}}{e ^ {(H _ {i , i} - m) / \\tau} + \\sum_ {i \\neq j} e ^ {H _ {i , j} / \\tau}}, \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 86, + 222, + 468, + 273 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $H$ is a matrix whose $(i,j)$ -th element represents the cosine similarity between $L^2$ -normalized audio and visual features at time step $i$ and $j$ , respectively, $\\tau$ is a learnable temperature parameter, and $m$ is a pre-determined margin. To promote the learning of shared features between the two modalities, we apply a symmetrized version of the above NCE loss as:", + "bbox": [ + 75, + 273, + 472, + 378 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {a} \\leftrightarrow \\mathrm {v}} = \\frac {1}{2} \\left(\\mathcal {L} _ {\\mathrm {a} \\rightarrow \\mathrm {v}} + \\mathcal {L} _ {\\mathrm {v} \\rightarrow \\mathrm {a}}\\right) = \\frac {1}{2} \\left(\\mathcal {L} _ {\\mathrm {N C E}} \\left(S _ {a v}\\right) + \\mathcal {L} _ {\\mathrm {N C E}} \\left(S _ {v a}\\right)\\right). \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 76, + 386, + 473, + 426 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Here the similarity matrices $S_{av} = \\langle F_a, F_v \\rangle / (||F_a|| \\cdot ||F_v||)$ and $S_{va} = S_{av}^{\\top}$ are computed with features $F_a$ and $F_v$ from modality-specific encoders $\\varphi_a$ and $\\varphi_v$ instead of the contextual features from the Transformer encoder $\\theta$ to avoid learning positional shortcuts. The total loss in stage 1 is", + "bbox": [ + 75, + 426, + 468, + 503 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {s t a g e 1}} = \\mathcal {L} _ {\\mathrm {a} \\rightarrow \\mathrm {a} ^ {\\prime}} + \\mathcal {L} _ {\\mathrm {a} \\leftrightarrow \\mathrm {v}}. \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 181, + 515, + 468, + 531 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Stage 2: Learning video-unique information. In stage 2, we proceed to learn video-unique information, which is now much easier than learning from scratch since it has been bootstrapped with shared knowledge from the audio in the previous stage. The audio encoder $\\varphi_{a}$ does not receive any gradients in this stage and is essentially frozen, since it is expected to have already acquired the audio-specific knowledge well in the previous stage. We adopt a similar distillation loss as stage 1:", + "bbox": [ + 75, + 537, + 472, + 675 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {v} \\rightarrow \\mathrm {v} ^ {\\prime}} = \\mathcal {L} _ {\\text {d i s t i l l}} \\left(E _ {v}, E _ {v} ^ {\\prime}\\right). \\tag {15}\n$$\n", + "text_format": "latex", + "bbox": [ + 183, + 685, + 468, + 702 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "On top of the learned visual-unique information, we also perform distillation from unmasked video features to the joint features of unmasked audio-visual inputs in this stage to prepare for the learning of the final robust audio-visual representations in the next stage. For this purpose, the loss in this stage is set as:", + "bbox": [ + 75, + 707, + 468, + 797 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {a v} \\rightarrow \\mathrm {v}} = \\mathcal {L} _ {\\text {d i s t i l l}} \\left(E _ {v}, E _ {a v}\\right), \\tag {16}\n$$\n", + "text_format": "latex", + "bbox": [ + 179, + 809, + 468, + 825 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {s t a g e 2}} = \\mathcal {L} _ {\\mathrm {v} \\rightarrow \\mathrm {v} ^ {\\prime}} + \\mathcal {L} _ {\\mathrm {a v} \\rightarrow \\mathrm {v}}. \\tag {17}\n$$\n", + "text_format": "latex", + "bbox": [ + 179, + 828, + 468, + 844 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To avoid repeated learning of only shared information between A and V again, the cross-modal loss $\\mathcal{L}_{\\mathrm{av}\\to \\mathrm{v}}$ is asymmetric, strictly one-way from AV to V, and a learnable tem", + "bbox": [ + 75, + 854, + 468, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "perature $\\tau_{v}$ Eq. (16) is introduced to the visual student outputs, i.e., we replace $\\psi_{k}$ in Eq. (9) with softmax activations of temperature-scaled linear projections of $E_{v}$ .", + "bbox": [ + 496, + 90, + 890, + 136 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Stage 3: Learning synergistic information. Finally, we aim to learn synergistic audio-visual information which previous methods have not explicitly addressed. Our ultimate goal is to maximize total task-relevant multi-modal information $I(E_{av};Y)$ , i.e. the sum of $R$ , $U_{a}$ , $U_{v}$ and $S$ , so that the representations are useful for both uni-modal and multi-modal tasks. Naively maximizing this objective without prior stages is problematic because the network tends to be biased towards learning audio-based representations (i.e. prioritizing $R$ and $U_{a}$ ) without considering the visual modality. By learning most of $R$ , $U_{a}$ and $U_{v}$ through the previous stages, our method is less susceptible to learning biased representations in the final stage. At this stage, we always mask identical positions in the audio and video inputs together so that the masked region never contains any unimodal information that can collapse or degrade training. To better recover the masked inputs, the model must cleverly leverage both modalities, which requires learning synergistic information that is not well-captured in prior stages. The loss function in stage 3 can be written as:", + "bbox": [ + 496, + 136, + 892, + 438 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {s t a g e 3}} = \\mathcal {L} _ {\\mathrm {a v} \\rightarrow \\mathrm {a v} ^ {\\prime}} = \\mathcal {L} _ {\\text {d i s t i l l}} \\left(E _ {a v} ^ {\\prime}, E _ {a v}\\right). \\tag {18}\n$$\n", + "text_format": "latex", + "bbox": [ + 557, + 448, + 890, + 465 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "By progressively acquiring the three types of useful information, we build strong and robust audio-visual speech representations for both uni- and bi-modal speech, which we shall show in subsequent experiments.", + "bbox": [ + 496, + 469, + 890, + 529 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.5. Fine-tuning Stage", + "text_level": 1, + "bbox": [ + 500, + 537, + 671, + 554 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "After pre-training, we add an extra linear layer after the fusion encoder $\\theta$ to fine-tune the model end-to-end with the Lattice-Free Maximum Mutual Information (LF-MMI) criterion [26, 51], which is much more lightweight than seq2seq Encoder-Decoder models and allows highly efficient decoding. We adopt a \"compact\" [33] 1-state hidden Markov model (HMM) topology, using $k2^2$ for differentiable Weighted Finite State Transducer (WFST)-based training. We use 624 bi-char units from text-based clustering [77] for English and tonal phone-level targets (initials and finals) for Chinese.", + "bbox": [ + 496, + 561, + 890, + 727 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 500, + 739, + 632, + 757 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Datasets", + "text_level": 1, + "bbox": [ + 500, + 763, + 601, + 779 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "LRS2-BBC [3]. LRS2-BBC is a 224-hour transcribed audio-visual English speech dataset collected from programs of various genres broadcast by the BBC. It is challenging due to the presence of extreme poses, as well as a large and diverse vocabulary. We use the official train / validation split (see supplementary materials) for evaluation.", + "bbox": [ + 496, + 787, + 890, + 878 + ], + "page_idx": 4 + }, + { + "type": "page_footnote", + "text": "2https://github.com/k2-fsa/k2", + "bbox": [ + 516, + 886, + 712, + 898 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "27073", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "LRS3-TED [1]. LRS3-TED is another audio-visual English speech dataset collected from TED and TEDx talks, with over 5.5k speakers and 439 hours of data. It covers diverse topics and is widely adopted for evaluating audio-visual speech representation learning methods for being the largest publicly available one of its kind. An official train / validation split is not given, so we use the widely adopted one prepared by Shi et al. [59], same as [27, 34, 80].", + "bbox": [ + 75, + 90, + 468, + 210 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "CAS-VSR-S101. We also collect a new large-scale, in-the-wild Mandarin dataset with 101.1 hours of data. The videos are sourced from broadcast news and conversational programs in Chinese, covering a highly diverse set of topics, speakers and filming conditions (see supplementary).", + "bbox": [ + 75, + 212, + 468, + 287 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Experimental Setup", + "text_level": 1, + "bbox": [ + 76, + 297, + 266, + 314 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "More implementation details can be found in the supplementary materials.", + "bbox": [ + 75, + 321, + 467, + 351 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "High- and low-resource setups. On LRS3-TED we follow the widely adopted setup of Shi et al. [59], which excludes the pre-train set during low-resource fine-tuning. To validate data scaling properties, previous works used VoxCeleb2-En [59], which is a 1326-hour subset of VoxCeleb2 [18] in English. However, not only are experiments of this scale computationally expensive, but more importantly, VoxCeleb2 is no longer available from its official source as of Feb. 2023. Therefore, we create a similar scaling protocol, using LRS3-TED (433h) as a pre-training source and LRS2-BBC (223h) for fine-tuning. For low-resource fine-tuning, we only use the 28-hour train subset of LRS2-BBC, while for high-resource fine-tuning, we include the pre-train subset, similar to [17].", + "bbox": [ + 75, + 352, + 467, + 564 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation. Decoding is performed using beam search with a beam size of 50. The common evaluation metrics for ASR, VSR and AVSR are Word Error Rate (WER) and Character Error Rate (CER) for English and Chinese, respectively. As our bi-char LF-MMI models are trained on verbatim transcriptions for ease of lexicon construction, we report WERs both without and with using Whisper-style text normalization [54] (in parentheses) to compensate for innocuous errors related to text normalization.", + "bbox": [ + 75, + 564, + 467, + 700 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Results on LRS2-BBC and LRS3-TED", + "text_level": 1, + "bbox": [ + 76, + 710, + 408, + 724 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Low-resource setup. Low-resource results on LRS2-BBC and LRS3-TED can be found in Tabs. 1 and 2. Our smallest BASE* encoder is modality-agnostic, only half the size of the one in [27], which uses two separate encoders to perform AVSR. On LRS3-TED, although we employ the simple LF-MMI instead of cross-entropy (CE) based seq2seq criterion for fine-tuning, our BASE model rivals or achieves SoTA in all cases. When scaling to a LARGE model, we observe significant improvements in both ASR and AVSR. Meanwhile, VSR performance remains comparable to SoTA, with a moderate $2\\%$ gap. This may sug", + "bbox": [ + 75, + 734, + 467, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "gest that with our current VGG-based visual encoder $\\varphi_v$ , the benefits of scaling up the Transformer primarily manifest in ASR and AVSR.", + "bbox": [ + 496, + 90, + 890, + 135 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "High-resource setup. High-resource results on LRS2-BBC and LRS3-TED can be found in Tabs. 1 and 3. On LRS2-BBC, with a 46M model and relying on its own 223h data for pre-training and fine-tuning, we achieve $30.3\\% / 3.6\\% / 2.9\\%$ WER for VSR, ASR and AVSR. Notably, these results surpass a highly competitive supervised method [40] and other self-supervised methods that incorporate additional data. This indicates that our progressive learning approach can effectively exploit visual information. Furthermore, when we introduce additional 433h unlabeled data, we achieve even better performance: $29.8\\% / 3.0\\% / 2.4\\%$ WER, which outperforms other methods that employ the same amount of data. When we switch to the LARGE model, performance rivals SoTA methods that employ 4x data (1759h) for pre-training. Similarly, we achieve or rival SoTA performance on LRS3-TED.", + "bbox": [ + 496, + 137, + 890, + 378 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.4. Results on CAS-VSR-S101", + "text_level": 1, + "bbox": [ + 498, + 390, + 741, + 404 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We also evaluate on our newly collected CAS-VSR-S101 dataset, which is smaller than both of the two previous datasets and in an even more ambiguous language (Chinese has more homophonous syllables and words than English). As shown in Tab. 4, $\\mathrm{ES}^3$ still works out-of-the-box under this challenging setup, and yields noticeable performance gains over a strong fully supervised baseline [39]. Interestingly, our fully supervised baseline model beats [39] in ASR and AVSR but slightly lags behind in VSR, again hinting that our strategy can further benefit from more complex visual modeling architectures.", + "bbox": [ + 496, + 412, + 890, + 580 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/d657ac7401c12c04bb6620d820470a30a41816cc760ec936a66a4bb09363df4e.jpg", + "table_caption": [ + "Table 4. Results on CAS-VSR-S101. Here we use a BASE* model for the fully-supervised baseline and $\\mathrm{ES}^3$ . The self-supervised models are pre-trained on training data and then finetuned for 50 epochs with the labels. The numbers before and after slashes correspond to CERs on the validation and test sets." + ], + "table_footnote": [], + "table_body": "
MethodsCriterionEncoder Size#EpochsVSRASRAVSR
Supervised
BaselineLF-MMI46M10060.0 / 57.415.4 / 12.614.6 / 12.1
E2E Conformer [39]CTC+CE79M10058.6 / 56.516.6 / 13.016.0 / 12.6
Self-supervised
ES3 Stage 161.2 / 59.014.7 / 11.9-
ES3 Stage 2LF-MMI46M5059.1 / 56.7-13.9 / 11.4
ES3 Stage 358.1 / 55.614.4 / 11.613.4 / 11.0
", + "bbox": [ + 503, + 657, + 890, + 732 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.5. Analysis and Discussion", + "text_level": 1, + "bbox": [ + 500, + 739, + 718, + 756 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Representation quality probing. In Tab. 5, we demonstrate the effect of each stage on LRS2-BBC. ASR performance between stages 1 and 3 are close, indicating that stage 1 indeed acquires most of audio-unique information. Comparing the VSR results between stages 1 and 2 reveals a significant gap, suggesting that stage 2 also effectively learns visual-unique information, hence the minimal differences between the final VSR result and those of stage 2. Moreover, comparing the results of stage 1-2 and stage", + "bbox": [ + 496, + 763, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "27074", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/927360e44f344c6270da4c7c7364195099ad0468b9eeae0f16e16f1fb37ff05a.jpg", + "table_caption": [ + "Table 1. Results on LRS2-BBC. We pre-train a BASE* model with 223h unlabeled data from LRS2-BBC, as well as BASE and LARGE models with 433h unlabeled data (LRS3-TED) to demonstrate scaling properties. *: external language models." + ], + "table_footnote": [ + "${}^{1}$ Not end-to-end; requires a GMM-HMM alignment stage. ${}^{2}$ Uses an additional ASR model trained on LibriSpeech (960h).", + "3 Reproduced by Zhu et al. [80]. 4 Uses additional 3846h audio, 452h audio-text and 600M text data.", + "5 Uses additional 60000h audio data and 1.28M unlabeled images." + ], + "table_body": "
MethodsUnlabeled AV dataLabeled DataEncoder SizeCriterionVSRASRAVSR
Supervised
Yu et al. [75]-1519h-LF-MMI148.96.75.9
Ma et al. [39]*-380h79MCTC+CE37.93.93.7
Ma et al. [41]*-818h186MCTC+CE27.92.6-
Semi-supervised
Afouras et al. [2]*777h223h2-CTC51.3--
Ma et al. [41]*2630h818h186MCTC+CE14.61.51.5
Self-supervised (BASE and BASE* models)
AV-HuBERT [59, 60]1759h223h103MCE31.23-3.63
VATLM [80]1759h4223h107MCE30.6-2.9
RAVeN [27]433h223h97MCTC+CE32.13.9-
Pan et al. [45]-5380h399MCTC+CE43.22.72.6
ES3(ours)223h28h46M40.2 (39.1)6.0 (5.0)5.7 (4.9)
433h28h102MLF-MMI39.3 (38.2)5.5 (4.8)5.1 (4.1)
223h223h46M31.4 (30.3)4.3 (3.6)3.8 (2.9)
433h223h102M30.7 (29.8)3.4 (3.0)3.2 (2.4)
Self-supervised (LARGE models)
ES3(ours)433h28h317MLF-MMI36.4 (35.4)5.2 (4.4)4.7 (4.0)
433h223h317M26.7 (25.8)3.1 (2.5)3.1 (2.5)
AV-HuBERT [59, 60]1759h223h325MCE25.53-2.53
VATLM [80]1759h4223h332MCE24.3-2.3
RAVeN [27]1759h223h671MCTC+CE23.22.5-
", + "bbox": [ + 161, + 121, + 810, + 393 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/2dcdb6385a0d37b67d50823791dac44d7c521cec9ae185d63940dae40105ddae.jpg", + "table_caption": [ + "Table 2. Low-resource results on LRS3-TED. We pre-train a BASE and LARGE model with 433h unlabeled data." + ], + "table_footnote": [ + "$^{1}$ Reproduced by Lian et al. [34]. $^{2}$ Uses additional 3846h audio, 452h audio-text and 600M text data.", + "3 Zhang et al. [76] inject noise during pre-training and obtain better fine-tuning results even with its base model AV-HuBERT $(47.1\\%)$ ." + ], + "table_body": "
MethodsUnlabeled AV dataLabeled DataEncoder SizeCriterionVSRASRAVSR
Self-supervised (BASE models)
AV-HuBERT [59, 60]433h30h103MCE51.84.9\\( {4.7}^{1} \\)
VATLM [80]\\( {433}{\\mathrm{\\;h}}^{2} \\)30h107MCE48.0-3.6
RAVEn [27]433h30h97MCTC+CE47.04.7-
\\( {\\mathrm{{AV2vec}}}^{\\left\\lbrack {76}\\right\\rbrack }{}^{3} \\)433h30h103MCE45.0-5.8
AV-data2vec [34]433h30h103MCE45.24.44.2
\\( {\\mathbf{{ES}}}^{3} \\) (ours)433h30h102MLF-MMI45.5 (44.7)3.9 (3.3)3.6 (3.0)
Self-supervised (LARGE models)
AV-HuBERT433h30h325MCE44.84.5\\( {4.2}^{1} \\)
AV-data2vec433h30h325MCE40.53.73.4
\\( {\\mathbf{{ES}}}^{3} \\) (ours)433h30h317MLF-MMI43.5 (42.5)3.8 (2.9)2.9 (2.3)
", + "bbox": [ + 161, + 454, + 810, + 597 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3 demonstrates that the joint information learning in stage 3, which also incorporates synergistic information, benefits both uni-modal (ASR & VSR) and bi-modal (AVSR) tasks. Similar trends can also be observed on the more challenging Mandarin dataset CAS-VSR-S101 in Tab. 4, further supporting the motivation of our proposed strategy.", + "bbox": [ + 75, + 627, + 467, + 719 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 5. Representation quality probing on LRS2-BBC. Results are reported on the validation set using the BASE* model under the low-resource setup for each pre-training stage.", + "bbox": [ + 75, + 720, + 467, + 762 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/4b2ca0b3d50677d50b9757e16e0ca0bb8b8248f847b4005e729ff730ba474000.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Pre-training StageVSRASRAVSR
Stage 158.3 (57.3)11.5 (10.1)-
Stage 248.7 (47.3)-12.9 (11.5)
Stage 3 (Full ES³)48.1 (46.9)11.4 (9.8)10.1 (8.6)
", + "bbox": [ + 109, + 767, + 439, + 819 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Effect of automatically balancing shared and unique information learning. It is known that in a hardness-aware loss function, the temperature rules the strength of penalties on hard negative samples [72]. We track the automatically learned temperature parameter $\\tau$ used for cross-modal", + "bbox": [ + 75, + 825, + 467, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "learning objectives $\\mathcal{L}_{\\mathrm{a}\\leftrightarrow \\mathrm{v}}$ in Eq. (12) and $\\mathcal{L}_{\\mathrm{av}\\rightarrow \\mathrm{v}}$ in Eq. (16) when pre-training on LRS2-BBC, as shown in Fig. 2. In stage 1, $\\tau$ continues to fall until it reaches $\\tau_{\\mathrm{min}}$ , while in stage 2 it briefly drops and then rises throughout training. The trend is in line with our intuition: in stage 1, audio information is easily learnable, so the model can fully exploit shared information by lowering $\\tau$ without sacrificing audio-unique information learning; in stage 2, after learning some residual shared information, an increasing $\\tau$ progressively relaxes $\\mathcal{L}_{\\mathrm{av}\\rightarrow \\mathrm{v}}$ to yield more freedom to $\\mathcal{L}_{\\mathrm{v}\\rightarrow \\mathrm{v}'}$ , i.e. video-unique information learning.", + "bbox": [ + 496, + 627, + 890, + 794 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/dc65de78e349c393d51fed49ea503110b75c3317341ac1413d53fd9f30ec9c74.jpg", + "image_caption": [ + "Figure 2. The temperature parameter paces learning automatically." + ], + "image_footnote": [], + "bbox": [ + 560, + 805, + 826, + 904 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "27075", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/7e5f9b935296553ba314447ad2237026f97cb39e4de41d77568a11274c9bd0ea.jpg", + "table_caption": [ + "Table 3. High-resource results on LRS3-TED. We fine-tune on 433h labeled data of LRS3. †: uses test-time augmentation. *: uses external language models. ‡: noise injection during pre-training. We list more methods for comparison in the supplementary material." + ], + "table_footnote": [ + "1 Uses an additional ASR model trained on LibriSpeech (960h).", + "2 Reproduced by Lian et al. [34]." + ], + "table_body": "
MethodsYearUnlabeled AV dataLabeled DataBackboneEncoder SizeCriterionVSRASRAVSR
Supervised
Ma et al. [39]*2021-595hConformer79MCTC+CE43.32.32.3
Prajwal et al. [52]†*2022-698hTransformer32MCE40.6--
Ma et al. [40]*2022-818hConformer79MCTC+CE34.7--
Semi-Supervised
Shillingford et al. [62]*2019-3886hRNN-CTC55.1--
Makino et al. [42]2019-31khRNN43MTransducer33.64.84.5
Afouras et al. [2]*2020344h433h1Jasper (CNN)-CTC59.8--
Serdyuk et al. [57]2021-90khTransformer-Transducer25.9-2.3
Ma et al. [40]*2022641h818hConformer79MCTC+CE31.5--
Serdyuk et al. [58]2022-90khConformer-Transducer17.01.61.6
Ma et al. [41]*20232630h818hConformer186MCTC+CE19.11.00.9
Chang et al. [15]2023-100khConformer98MTransducer12.8-0.9
Self-supervised (Base Models)
AV-HuBERT [59, 60]2022433h433hTransformer103MCE44.03.02.82
AV2vec [76]‡2023433h433hTransformer103MCE39.9-2.6
RAVEn [27]2023433h433hTransformer97MCTC+CE39.12.2-
ES3 (ours)2023433h433hTransformer102MLF-MMI40.3 (39.2)2.9 (2.4)2.5 (2.0)
AV-data2vec [34]2023433h433hTransformer103MCE39.02.01.8
Self-supervised (Large Models)
AV-HuBERT [59, 60]2022433h433hTransformer325MCE41.62.72.52
AV-data2vec [34]2023433h433hTransformer325MCE37.41.91.7
ES3 (ours)2023433h433hTransformer317MLF-MMI37.1 (36.7)2.8 (2.1)2.1 (1.7)
", + "bbox": [ + 140, + 122, + 831, + 371 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Visualization of our learned psuedo-label sequences. We randomly select a sample from the LRS3-TED validation set and visualize the pseudo-label sequences, i.e. code indices $q$ generated by the audio, visual and audio-visual EMA teacher $\\tilde{\\theta}$ in each stage. Ground truth phones for each audio frame (10ms) are derived by forced alignment [79]. Fig. 3 shows that our learned codebook yields high-quality pseudo-labels that strongly correlate with phonemic units.", + "bbox": [ + 75, + 388, + 468, + 508 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/5e00eb854a35c01a20449a847b6613fbe7a3476ae06be938215eb0055ab4b0f6.jpg", + "image_caption": [ + "Figure 3. Pseudo-label sequences from each stage. Example of pseudo-labels generated from the last Transformer layer for the utterance \"so what does that mean\" by the audio, visual, and audio-visual teachers in each stage (BASE model). At the bottom we show the spectrogram for reference. Please note that our pseudolabels are derived at 40ms intervals, limited by video frame rate." + ], + "image_footnote": [], + "bbox": [ + 78, + 515, + 462, + 633 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Training efficiency comparison with SoTA methods. Learning audio-visual speech representations has been extremely compute-intensive. Here we briefly compare $\\mathrm{ES}^3$ with prominent SoTA methods. Pre-training a BASE/LARGE model on LRS3-TED using our $\\mathrm{ES}^3$ for 100 epochs per stage only takes about 6/8.3 days in total with just 8 NVIDIA A40 GPUs. In contrast, AV-data2vec is trained for 1000k steps using 64 NVIDIA V100 GPUs for 4-5/6-7 days using BASE/LARGE with a batch size of 1280/2560s, or equivalently, 821/1642 epochs. They also experiment extensively with hyperparameters such as modality schedulers, which further increases the cost of", + "bbox": [ + 75, + 719, + 470, + 901 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "training. RAVEn performs an equivalent of 6.25/50 days of training on 8 A100s with BASE*/LARGE. It uses separate encoders for audio and video, so the resulting model cannot perform audio-visual fusion out-of-the-box. This also requires keeping track of four sets of parameters, which is very memory-intensive. By focusing on designing a clever learning strategy, we circumvent such manual tuning altogether and greatly reduce training costs. Finally, we use the simple LF-MMI criterion, and expect better performance from tuning beam size, including a larger external language model (see supplementary), or switching to the common heavy Attention-based Encoder-Decoder model.", + "bbox": [ + 496, + 388, + 893, + 569 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusions", + "text_level": 1, + "bbox": [ + 500, + 575, + 627, + 590 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We present a novel strategy, $\\mathrm{ES}^3$ , for self-supervised learning of audio-visual speech representations. Unlike existing methods that primarily focus on learning information shared between audio and video, we take an \"evolving\" approach that also captures unique and synergistic information. Our experimental results demonstrate that $\\mathrm{ES}^3$ indeed progressively learns unique, shared and synergistic information. Moreover, in terms of downstream performance, it rivals SoTA methods in terms of ASR, VSR and AVSR accuracy without burdensome manual balancing efforts.", + "bbox": [ + 496, + 595, + 890, + 747 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments. This work is partially supported by the National Natural Science Foundation of China (No. 62276247, 62076250). We are grateful to the ICT computing platform, the Nanjing Institute of InforSuperBahn, and Capitalonline Data Service Co., Ltd for providing additional compute resources. We also thank Chenhao Wang, Mingmin Yang and Jingyun Xiao for helping to prepare CAS-VSR-S101, Bingquan Xia and Songtao Luo for help with Conformer and seq2seq baselines, Bingquan Xia and Feixiang Wang for help during the rebuttal phase, and Feixiang Wang and Zhaoxin Yuan for an early version of Fig. 1.", + "bbox": [ + 496, + 747, + 892, + 901 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "27076", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 119, + 173, + 136 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Triantafyllos Afouras, Joon Son Chung, and Andrew Zisserman. LRS3-TED: a large-scale dataset for visual speech recognition. CoRR, abs/1809.00496, 2018. 6", + "[2] Triantafyllos Afouras, Joon Son Chung, and Andrew Zisserman. ASR is all you need: Cross-modal distillation for lip reading. In ICASSP, pages 2143-2147. IEEE, 2020. 2, 7, 8, 4, 5", + "[3] Triantafyllos Afouras, Joon Son Chung, Andrew W. Senior, Oriol Vinyals, and Andrew Zisserman. Deep audio-visual speech recognition. IEEE Trans. Pattern Anal. Mach. Intell., 44(12):8717-8727, 2022. 1, 5, 4", + "[4] Jean-Baptiste Alayrac, Adrià Recasens, Rosalia Schneider, Relja Arandjelovic, Jason Ramapuram, Jeffrey De Fauw, Lucas Smaira, Sander Dieleman, and Andrew Zisserman. Self-supervised multimodal versatile networks. In NeurlPS, 2020. 2", + "[5] Humam Alwassel, Dhruv Mahajan, Bruno Korbar, Lorenzo Torresani, Bernard Ghanem, and Du Tran. Self-supervised learning by cross-modal audio-video clustering. In NeurIPS, 2020. 2", + "[6] Relja Arandjelovic and Andrew Zisserman. Look, listen and learn. In ICCV, pages 609-617. IEEE Computer Society, 2017. 2", + "[7] Relja Arandjelovic and Andrew Zisserman. Objects that sound. In ECCV(1), pages 451-466. Springer, 2018. 2", + "[8] Yuki Markus Asano, Mandela Patrick, Christian Rupprecht, and Andrea Vedaldi. Labelling unlabelled videos from scratch with multi-modal self-supervision. In NeurIPS, 2020. 2", + "[9] Alexei Baevski, Michael Auli, and Abdelrahman Mohamed. Effectiveness of self-supervised pre-training for speech recognition. CoRR, abs/1911.03912, 2019. 1", + "[10] Alexei Baevski, Yuhao Zhou, Abdelrahman Mohamed, and Michael Auli. wav2vec 2.0: A framework for self-supervised learning of speech representations. In NeurIPS, 2020. 3", + "[11] Alexei Baevski, Wei-Ning Hsu, Qiantong Xu, Arun Babu, Jiatao Gu, and Michael Auli. data2vec: A general framework for self-supervised learning in speech, vision and language. In ICML, pages 1298-1312. PMLR, 2022. 1", + "[12] Alexei Baevski, Arun Babu, Wei-Ning Hsu, and Michael Auli. Efficient self-supervised learning with contextualized target representations for vision, speech and language. In ICML, pages 1416-1429. PMLR, 2023. 1, 3", + "[13] Adrian Bulat and Georgios Tzimiropoulos. How far are we from solving the 2D & 3D face alignment problem? (and a dataset of 230,000 3D facial landmarks). In ICCV, pages 1021-1030. IEEE Computer Society, 2017. 2", + "[14] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision Transformers. In ICCV, pages 9630-9640. IEEE, 2021. 1", + "[15] Oscar Chang, Hank Liao, Dmitriy Serdyuk, Ankit Shahy, and Olivier Siohan. Conformer is all you need for visual" + ], + "bbox": [ + 78, + 145, + 470, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "speech recognition. In ICASSP, pages 10136-10140. IEEE, 2024. 1, 8, 5", + "[16] Brian Chen, Andrew Rouditchenko, Kevin Duarte, Hilde Kuehne, Samuel Thomas, Angie W. Boggust, Rameswar Panda, Brian Kingsbury, Rogério Feris, David Harwath, James R. Glass, Michael Picheny, and Shih-Fu Chang. Multimodal clustering networks for self-supervised learning from unlabeled videos. In ICCV, pages 7992-8001. IEEE, 2021. 2", + "[17] Xize Cheng, Tao Jin, Linjun Li, Wang Lin, Xinyu Duan, and Zhou Zhao. OpenSR: Open-modality speech recognition via maintaining multi-modality alignment. In ACL (1), pages 6592-6607. Association for Computational Linguistics, 2023. 6", + "[18] Joon Son Chung, Arsha Nagrani, and Andrew Zisserman. VoxCeleb2: Deep speaker recognition. In INTERSPEECH, pages 1086-1090. ISCA, 2018. 6", + "[19] Soo-Whan Chung, Joon Son Chung, and Hong-Goo Kang. Perfect match: Improved cross-modal embeddings for audiovisual synchronisation. In ICASSP, pages 3965–3969. IEEE, 2019. 5", + "[20] Tri Dao. FlashAttention-2: Faster attention with better parallelism and work partitioning. CoRR, abs/2307.08691, 2023. 2", + "[21] Tri Dao, Daniel Y. Fu, Stefano Ermon, Atri Rudra, and Christopher Ré. FlashAttention: Fast and memory-efficient exact attention with IO-awareness. In NeurIPS, 2022. 2", + "[22] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. BERT: Pre-training of deep bidirectional transformers for language understanding. In NAACL-HLT (1), pages 4171–4186. Association for Computational Linguistics, 2019. 1, 2", + "[23] Angela Fan, Edouard Grave, and Armand Joulin. Reducing Transformer depth on demand with structured dropout. In ICLR. OpenReview.net, 2020. 2", + "[24] Dalu Feng, Shuang Yang, and Shiguang Shan. An efficient software for building lip reading models without pains. In ICME Workshops, pages 1-2. IEEE, 2021. 1", + "[25] Yuan Gong, Andrew Rouditchenko, Alexander H. Liu, David Harwath, Leonid Karlinsky, Hilde Kuehne, and James R. Glass. Contrastive audio-visual masked autoencoder. In ICLR. OpenReview.net, 2023. 2", + "[26] Hossein Hadian, Hossein Sameti, Daniel Povey, and Sanjeev Khudanpur. End-to-end speech recognition using lattice-free MMI. In INTERSPEECH, pages 12-16. ISCA, 2018. 5", + "[27] Alexandros Haliassos, Pingchuan Ma, Rodrigo Mira, Stavros Petridis, and Maja Pantic. Jointly learning visual and auditory speech representations from raw data. In ICLR. Open-Review.net, 2023. 2, 6, 7, 8, 1, 4, 5", + "[28] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, and Ross B. Girshick. Masked autoencoders are scalable vision learners. In CVPR, pages 15979-15988. IEEE, 2022. 2", + "[29] François Hernandez, Vincent Nguyen, Sahar Ghannay, Natalia A. Tomashenko, and Yannick Esteve. TED-LIUM 3: Twice as much data and corpus repartition for experiments on speaker adaptation. In SPECOM, pages 198-208. Springer, 2018. 2" + ], + "bbox": [ + 503, + 92, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "27077", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[30] Wei-Ning Hsu and Bowen Shi. u-HuBERT: Unified mixed-modal speech pretraining and zero-shot transfer to unlabeled modality. In NeurIPS, 2022. 1", + "[31] Wei-Ning Hsu, Benjamin Bolte, Yao-Hung Hubert Tsai, Kushal Lakhotia, Ruslan Salakhutdinov, and Abdelrahman Mohamed. HuBERT: Self-supervised speech representation learning by masked prediction of hidden units. IEEE ACM Trans. Audio Speech Lang. Process., 29:3451-3460, 2021. 1", + "[32] Bruno Korbar, Du Tran, and Lorenzo Torresani. Cooperative learning of audio and video models from self-supervised synchronization. In NeurIPS, pages 7774-7785, 2018. 2", + "[33] Aleksandr Laptev, Somshubra Majumdar, and Boris Ginsburg. CTC variations through new WFST topologies. In INTERSPEECH, pages 1041-1045. ISCA, 2022. 5", + "[34] Jiachen Lian, Alexei Baevski, Wei-Ning Hsu, and Michael Auli. AV-data2vec: Self-supervised learning of audio-visual speech representations with contextualized target representations. In ASRU. IEEE, 2023. 2, 6, 7, 8, 1, 5", + "[35] Paul Pu Liang, Zihao Deng, Martin Q. Ma, James Y. Zou, Louis-Philippe Morency, and Ruslan Salakhutdinov. Factorized contrastive learning: Going beyond multi-view redundancy. In NeurIPS, 2023. 2", + "[36] Alexander H. Liu, Heng-Jui Chang, Michael Auli, Wei-Ning Hsu, and Jim Glass. DinoSR: Self-distillation and online clustering for self-supervised speech representation learning. In NeurIPS, 2023. 4, 2", + "[37] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In *ICLR (Poster)*. OpenReview.net, 2019. 2", + "[38] Pingchuan Ma, Rodrigo Mira, Stavros Petridis, Bjorn W. Schuller, and Maja Pantic. LiRA: Learning visual speech representations from audio through self-supervision. In *Interspeech*, pages 3011–3015. ISCA, 2021. 2", + "[39] Pingchuan Ma, Stavros Petridis, and Maja Pantic. End-to-end audio-visual speech recognition with Conformers. In ICASSP, pages 7613-7617. IEEE, 2021. 1, 6, 7, 8, 4, 5", + "[40] Pingchuan Ma, Stavros Petridis, and Maja Pantic. Visual speech recognition for multiple languages in the wild. Nat. Mac. Intell., 4(11):930-939, 2022. 1, 6, 8, 2, 4, 5", + "[41] Pingchuan Ma, Alexandros Haliassos, Adriana Fernandez-Lopez, Honglie Chen, Stavros Petridis, and Maja Pantic. Auto-AVSR: Audio-Visual speech recognition with automatic labels. In ICASSP, pages 1-5. IEEE, 2023. 1, 3, 7, 8, 4, 5", + "[42] Takaki Makino, Hank Liao, Yannis M. Assael, Brendan Shillingford, Basilio Garcia, Otavio Braga, and Olivier Siohan. Recurrent neural network transducer for audio-visual speech recognition. In ASRU, pages 905–912. IEEE, 2019. 1, 8, 5", + "[43] Antoine Miech, Dimitri Zhukov, Jean-Baptiste Alayrac, Makarand Tapaswi, Ivan Laptev, and Josef Sivic. HowTo100M: Learning a text-video embedding by watching hundred million narrated video clips. In ICCV, pages 2630-2640. IEEE, 2019. 2", + "[44] Arsha Nagrani, Chen Sun, David Ross, Rahul Sukthankar, Cordelia Schmid, and Andrew Zisserman. Speech2Action: Cross-modal supervision for action recognition. In CVPR, pages 10314-10323. Computer Vision Foundation / IEEE, 2020. 2" + ], + "bbox": [ + 78, + 90, + 470, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[45] Xichen Pan, Peiyu Chen, Yichen Gong, Helong Zhou, Xinbing Wang, and Zhouhan Lin. Leveraging unimodal self-supervised learning for multimodal audio-visual speech recognition. In ACL (1), pages 4491–4503. Association for Computational Linguistics, 2022. 7, 4", + "[46] Daniel S. Park, William Chan, Yu Zhang, Chung-Cheng Chiu, Barret Zoph, Ekin D. Cubuk, and Quoc V. Le. SpecAugment: A simple data augmentation method for automatic speech recognition. In INTERSPEECH, pages 2613-2617. ISCA, 2019. 2", + "[47] Sooyoung Park, Arda Senocak, and Joon Son Chung. MarginNCE: Robust sound localization with a negative margin. In ICASSP, pages 1-5. IEEE, 2023. 5, 2", + "[48] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, Alban Desmaison, Andreas Köpf, Edward Z. Yang, Zachary DeVito, Martin Raison, Alykhan Tejani, Sasank Chilamkurthy, Benoit Steiner, Lu Fang, Junjie Bai, and Soumith Chintala. PyTorch: An imperative style, high-performance deep learning library. In NeurIPS, pages 8024-8035, 2019. 2", + "[49] Mandela Patrick, Yuki Markus Asano, Polina Kuznetsova, Ruth Fong, João F. Henriques, Geoffrey Zweig, and Andrea Vedaldi. On compositions of transformations in contrastive self-supervised learning. In ICCV, pages 9557-9567. IEEE, 2021. 2", + "[50] Trung Pham, Chaoning Zhang, Axi Niu, Kang Zhang, and Chang D. Yoo. On the pros and cons of momentum encoder in self-supervised visual representation learning. CoRR, abs/2208.05744, 2022. 4", + "[51] Daniel Povey, Vijayaditya Peddinti, Daniel Galvez, Pegah Ghahremani, Vimal Manohar, Xingyu Na, Yiming Wang, and Sanjeev Khudanpur. Purely sequence-trained neural networks for ASR based on lattice-free MMI. In INTER-SPEECH, pages 2751-2755. ISCA, 2016. 5", + "[52] K. R. Prajwal, Triantafyllos Afouras, and Andrew Zisserman. Sub-word level lip reading with visual attention. In CVPR, pages 5152-5162. IEEE, 2022. 8, 4, 5", + "[53] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In ICML, pages 8748-8763. PMLR, 2021. 2", + "[54] Alec Radford, Jong Wook Kim, Tao Xu, Greg Brockman, Christine McLeavey, and Ilya Sutskever. Robust speech recognition via large-scale weak supervision. In ICML, pages 28492-28518. PMLR, 2023. 6", + "[55] Jordi Robert-Ribes, Jean-Luc Schwartz, Tahar Lallouache, and Pierre Escudier. Complementarity and synergy in b-modal speech: Auditory, visual, and audio-visual identification of french oral vowels in noise. The Journal of the Acoustical Society of America, 103(6):3677-3689, 1998. 1", + "[56] Steffen Schneider, Alexei Baevski, Ronan Collobert, and Michael Auli. wav2vec: Unsupervised pre-training for speech recognition. In INTERSPEECH, pages 3465-3469. ISCA, 2019. 3" + ], + "bbox": [ + 501, + 90, + 892, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "27078", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[57] Dmitriy Serdyuk, Otavio Braga, and Olivier Siohan. Audiovisual speech recognition is worth $32 \\times 32 \\times 8$ voxels. In ASRU, pages 796-802. IEEE, 2021. 2, 8, 5", + "[58] Dmitriy Serdyuk, Otavio Braga, and Olivier Siohan. Transformer-based video front-ends for audio-visual speech recognition for single and multi-person video. In INTER-SPEECH, pages 2833–2837. ISCA, 2022. 2, 8, 5", + "[59] Bowen Shi, Wei-Ning Hsu, Kushal Lakhotia, and Abdelrahman Mohamed. Learning audio-visual speech representation by masked multimodal cluster prediction. In ICLR. OpenReview.net, 2022. 1, 2, 3, 6, 7, 8, 4, 5", + "[60] Bowen Shi, Wei-Ning Hsu, and Abdelrahman Mohamed. Robust self-supervised audio-visual speech recognition. In INTERSPEECH, pages 2118–2122. ISCA, 2022. 7, 8, 4, 5", + "[61] Bowen Shi, Abdelrahman Mohamed, and Wei-Ning Hsu. Learning lip-based audio-visual speaker embeddings with AV-HuBERT. In INTERSPEECH, pages 4785–4789. ISCA, 2022. 1", + "[62] Brendan Shillingford, Yannis M. Assael, Matthew W. Hoffman, Thomas Paine, Cian Hughes, Utsav Prabhu, Hank Liao, Hasim Sak, Kanishka Rao, Lorrayne Bennett, Marie Mulville, Misha Denil, Ben Coppin, Ben Laurie, Andrew W. Senior, and Nando de Freitas. Large-scale visual speech recognition. In INTERSPEECH, pages 4135-4139. ISCA, 2019. 8, 5", + "[63] Nina Shvetsova, Brian Chen, Andrew Rouditchenko, Samuel Thomas, Brian Kingsbury, Rogério Feris, David Harwath, James R. Glass, and Hilde Kuehne. Everything at once - multi-modal fusion Transformer for video retrieval. In CVPR, pages 19988-19997. IEEE, 2022. 2", + "[64] Themos Stafylakis and Georgios Tzimiropoulos. Combining residual networks with LSTMs for lipreading. In INTER-SPEECH, pages 3652-3656. ISCA, 2017. 3", + "[65] Jianlin Su, Yu Lu, Shengfeng Pan, Bo Wen, and Yunfeng Liu. RoFormer: Enhanced transformer with rotary position embedding. CoRR, abs/2104.09864, 2021. 3", + "[66] Weijie Su, Xizhou Zhu, Yue Cao, Bin Li, Lewei Lu, Furu Wei, and Jifeng Dai. VL-BERT: Pre-training of generic visual-linguistic representations. In ICLR. OpenReview.net, 2020. 2", + "[67] Chen Sun, Austin Myers, Carl Vondrick, Kevin Murphy, and Cordelia Schmid. VideoBERT: A joint model for video and language representation learning. In ICCV, pages 7463-7472. IEEE, 2019. 2", + "[68] Yonglong Tian, Dilip Krishnan, and Phillip Isola. Contrastive multiview coding. In ECCV (11), pages 776-794. Springer, 2020. 2", + "[69] Hugo Touvron, Matthieu Cord, Alexandre Sablayrolles, Gabriel Synnaeve, and Hervé Jégou. Going deeper with image Transformers. In ICCV, pages 32-42. IEEE, 2021. 2", + "[70] Kristin J Van Engen, Avanti Dey, Mitchell S Sommers, and Jonathan E Peelle. Audiovisual speech perception: Moving beyond mcgurk. The Journal of the Acoustical Society of America, 152(6):3216-3225, 2022. 1", + "[71] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In NIPS, pages 5998-6008, 2017. 3, 2" + ], + "bbox": [ + 78, + 92, + 467, + 898 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[72] Feng Wang and Huaping Liu. Understanding the behaviour of contrastive loss. In CVPR, pages 2495-2504. Computer Vision Foundation / IEEE, 2021. 7", + "[73] Paul L. Williams and Randall D. Beer. Nonnegative decomposition of multivariate information. CoRR, abs/1004.2515, 2010. 3", + "[74] Bo Xu, Cheng Lu, Yandong Guo, and Jacob Wang. Discriminative multi-modality speech recognition. In CVPR, pages 14421-14430. Computer Vision Foundation / IEEE, 2020. 5", + "[75] Jianwei Yu, Shi-Xiong Zhang, Jian Wu, Shahram Ghorbani, Bo Wu, Shiyin Kang, Shansong Liu, Xunying Liu, Helen Meng, and Dong Yu. Audio-visual recognition of overlapped speech for the LRS2 dataset. In ICASSP, pages 6984–6988. IEEE, 2020. 7, 4", + "[76] Jing-Xuan Zhang, Genshun Wan, Zhen-Hua Ling, Jia Pan, Jianqing Gao, and Cong Liu. Self-supervised audiovisual speech representations learning by multimodal self-distillation. In ICASSP, pages 1-5. IEEE, 2023. 2, 7, 8, 1, 5", + "[77] Xiaohui Zhang, Vimal Manohar, David Zhang, Frank Zhang, Yangyang Shi, Nayan Singhal, Julian Chan, Fuchun Peng, Yatharth Saraf, and Mike Seltzer. On lattice-free boosted MMI training of HMM and CTC-based full-context ASR models. In ASRU, pages 1026-1033. IEEE, 2021. 5", + "[78] Ya Zhao, Rui Xu, Xinchao Wang, Peng Hou, Haihong Tang, and Mingli Song. Hearing lips: Improving lip reading by distilling speech recognizers. In AAAI, pages 6917-6924. AAAI Press, 2020. 2", + "[79] Jian Zhu, Cong Zhang, and David Jurgens. Phone-to-audio alignment without text: A semi-supervised approach. In ICASSP, pages 8167-8171. IEEE, 2022. 8", + "[80] Qiushi Zhu, Long Zhou, Ziqiang Zhang, Shujie Liu, Binxing Jiao, Jie Zhang, Lirong Dai, Daxin Jiang, Jinyu Li, and Furu Wei. VATLM: Visual-Audio-Text pre-training with unified masked prediction for speech representation learning. IEEE Transactions on Multimedia, pages 1–11, 2023. 1, 6, 7, 4, 5" + ], + "bbox": [ + 501, + 92, + 890, + 599 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "27079", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2024/ES3_ Evolving Self-Supervised Learning of Robust Audio-Visual Speech Representations/09b7366b-807f-4dfb-a9c8-45edf9472928_model.json b/2024/ES3_ Evolving Self-Supervised Learning of Robust Audio-Visual Speech Representations/09b7366b-807f-4dfb-a9c8-45edf9472928_model.json new file mode 100644 index 0000000000000000000000000000000000000000..bc80a16d41574d2bf152699224cf780112e44c64 --- /dev/null +++ b/2024/ES3_ Evolving Self-Supervised Learning of Robust Audio-Visual Speech Representations/09b7366b-807f-4dfb-a9c8-45edf9472928_model.json @@ -0,0 +1,2543 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.0, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.245, + 0.13, + 0.726, + 0.177 + ], + "angle": 0, + "content": "\\(\\mathbf{ES}^{3}\\): Evolving Self-Supervised Learning of Robust Audio-Visual Speech Representations" + }, + { + "type": "text", + "bbox": [ + 0.21, + 0.203, + 0.76, + 0.221 + ], + "angle": 0, + "content": "Yuanhang Zhang\\(^{1,2}\\), Shuang Yang\\(^{1,2}\\), Shiguang Shan\\(^{1,2}\\), Xilin Chen\\(^{1,2}\\)" + }, + { + "type": "text", + "bbox": [ + 0.109, + 0.221, + 0.862, + 0.257 + ], + "angle": 0, + "content": "1Key Laboratory of Intelligent Information Processing of Chinese Academy of Sciences (CAS), Institute of Computing Technology, CAS, Beijing 100190, China" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.257, + 0.758, + 0.275 + ], + "angle": 0, + "content": "\\(^{2}\\)University of Chinese Academy of Sciences, Beijing 100049, China" + }, + { + "type": "text", + "bbox": [ + 0.184, + 0.277, + 0.782, + 0.293 + ], + "angle": 0, + "content": "zhangyuanhang15@mails.ucas.ac.cn, {shuang.yang, sgshan, xlchen}@ict.ac.cn" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.327, + 0.314, + 0.344 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.359, + 0.473, + 0.753 + ], + "angle": 0, + "content": "We propose a novel strategy, \\( ES^3 \\), for self-supervised learning of robust audio-visual speech representations from unlabeled talking face videos. While many recent approaches for this task primarily rely on guiding the learning process using the audio modality alone to capture information shared between audio and video, we reframe the problem as the acquisition of shared, unique (modality-specific) and synergistic speech information to address the inherent asymmetry between the modalities. Based on this formulation, we propose a novel \"evolving\" strategy that progressively builds joint audio-visual speech representations that are strong for both uni-modal (audio & visual) and bimodal (audio-visual) speech. First, we leverage the more easily learnable audio modality to initialize audio and visual representations by capturing audio-unique and shared speech information. Next, we incorporate video-unique speech information and bootstrap the audio-visual representations on top of the previously acquired shared knowledge. Finally, we maximize the total audio-visual speech information, including synergistic information to obtain robust and comprehensive representations. We implement \\( ES^3 \\) as a simple Siamese framework and experiments on both English benchmarks and a newly contributed large-scale Mandarin dataset show its effectiveness. In particular, on LRS2-BBC, our smallest model is on par with SoTA models with only 1/2 parameters and 1/8 unlabeled data (223h)." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.77, + 0.21, + 0.787 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.795, + 0.471, + 0.902 + ], + "angle": 0, + "content": "The task of learning audio-visual speech representations has gained increasing importance due to its potential to support a wide range of relevant applications. Building on the success in other domains [9, 11, 12, 14, 22, 31], several effective self-supervised approaches have recently emerged for this task by pre-training models on paired or mixed audio, visual and text data with pretext tasks like speech unit" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.328, + 0.895, + 0.784 + ], + "angle": 0, + "content": "prediction [30, 59, 80]. However, learning representations in the audio-visual speech domain faces unique challenges brought by the inherent asymmetry between the two modalities involved. First, audio and video often convey distinct and complementary types of speech information. For example, \\( \\mathsf{m} / \\) and \\( \\mathsf{n} / \\) sound similar, but are visually different; on the contrary, \\( \\mathsf{p} / \\) and \\( \\mathsf{b} / \\) are visually similar but can be distinguished with audio [70]. Even more interesting is the presence of synergistic information – information that can further enhance speech perception [55] but only emerges when audio and video are considered together, and cannot be inferred when either modality is absent. Second, they manifest different amounts of noise and information. Specifically, in most existing audio-visual speech datasets, audio is characterized by high sampling rate and relatively clean speech samples, making audio representations readily learnable from these datasets. In contrast, the visual modality often presents several challenging factors, including missing (e.g. due to self-occlusion in profile faces or low image quality), redundant (e.g. due to appearance variations, makeup and beards etc.) and ambiguous (e.g. due to the lower sampling rate and the presence of homophones) visual information. Their presence complicates learning meaningful and robust representations solely from video. The difficulties encountered in the continuous efforts to improve the performance of visual speech recognition (VSR) [15, 24, 40, 41] models illustrate this persistent obstacle. Additionally, these differences between the two modalities lead to an imbalance between their learning dynamics, requiring additional manual settings [3, 39, 42, 61]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.795, + 0.895, + 0.903 + ], + "angle": 0, + "content": "In this work, we break down the problem of learning audio-visual speech representations into the learning of three different types of information: shared (across modalities), unique (modality-specific) and synergistic speech information. Based on our formulation, we propose a novel strategy called \\(\\mathrm{ES^3}\\) (short for Evolving Self-Supervised Learning of Robust Audio-Visual Speech Representations)" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "27069" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.243 + ], + "angle": 0, + "content": "to tackle the asymmetry between the two modalities for learning robust audio-visual speech representations. Specifically, our \\(\\mathrm{ES}^3\\) employs a three-stage progressive learning approach. Initially, it focuses on acquiring the easily learnable audio-unique and shared speech information. Then, it addresses the more challenging visual-unique speech information while incorporating it into the initial audio-visual joint representation. Lastly, it optimizes the final audio-visual representations by leveraging synergy, on top of the acquired shared and modality-unique information." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.244, + 0.423, + 0.258 + ], + "angle": 0, + "content": "Our contributions can be summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.26, + 0.47, + 0.319 + ], + "angle": 0, + "content": "- We propose a new formulation for audio-visual speech representation learning by decomposing the task into the acquisition of shared (across modalities), unique (modality-specific) and synergistic speech information." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.32, + 0.469, + 0.395 + ], + "angle": 0, + "content": "- We propose a novel strategy, \\(\\mathrm{ES}^3\\) to progressively learn the three types of information in a self-supervised manner with a simple Siamese framework. This evolving strategy significantly alleviates the previous burden of tuning balancing parameters." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.396, + 0.469, + 0.44 + ], + "angle": 0, + "content": "- Extensive experiments on two widely used benchmarks show that \\(\\mathrm{ES}^3\\) rivals previous methods in auditory, visual and audio-visual speech recognition performance." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.441, + 0.469, + 0.485 + ], + "angle": 0, + "content": "- We also contribute a new large-scale Mandarin audiovisual speech dataset named CAS-VSR-S101 to the community for evaluation of similar methods." + }, + { + "type": "list", + "bbox": [ + 0.078, + 0.26, + 0.47, + 0.485 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.497, + 0.22, + 0.512 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.522, + 0.473, + 0.824 + ], + "angle": 0, + "content": "Self-supervised learning from multi-modal data. The arguably most successful paradigm for self-supervised multimodal learning in the past few years is contrastive learning, which exploits the common information encoded in different modalities [68], such as audio (A) [6, 7], image (I) [53, 66], video (V) [5, 8, 32, 67] and text (T) [4, 16, 43, 44, 49, 63]. These approaches focus on learning semantically aligned multi-modal representations that are effective for specific downstream tasks like cross-modal retrieval, but discard modality-unique information that is useful for other uni-modal tasks. To solve this problem, Gong et al. [25] combine contrastive learning with the masked prediction objective [28] to learn strong audio-visual representations suitable for both retrieval and joint classification. Liang et al. [35] propose a new algorithm, FACTORCL to learn self-supervised multi-modal representations that capture both shared and unique information and verify its effectiveness on multiple tasks. In this paper, we explicitly address capturing unique, shared and synergistic information for uni- and multi-modal downstream tasks." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.826, + 0.473, + 0.903 + ], + "angle": 0, + "content": "Learning audio-visual speech representations. A main challenge of learning audio-visual speech representation is properly handling the aforementioned asymmetry between audio and video. An established approach is to leverage the easily learnable audio modality as a teacher to guide" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.501 + ], + "angle": 0, + "content": "visual or audio-visual speech representation learning. Implementing additional measures to mitigate audio dominance and promote visual features during learning is vital for this paradigm. For instance, Shi et al. [59] and Zhang et al. [76] use different masking strengths for the two modalities, apply modality dropout, and enforce simple audio features by only using a weak linear projection layer. Others have explored freezing pre-trained audio representations [2, 38, 78]. However, such an audio-centric approach is sub-optimal and inefficient. First, it limits the learned representations to capturing only shared, i.e. modality-invariant information, or at most, shared and audio-unique information. Recently, a few attempts have been made to go beyond merely using audio as the anchor modality. Haliassos et al. [27] combine multiple audio-visual prediction tasks in place of purely audio-centric ones, and treats A-V asymmetry by having the visual stream predict only auditory but not visual targets. Lian et al. [34] perform a series of trials to determine an optimal schedule for audio and video input combination. Second, the process of exhaustively testing and identifying the optimal set of balancing parameters such as masking ratios, modality schedules and modality dropout rate is extremely time-consuming and computationally demanding. In contrast, our newly proposed evolving strategy identifies different types of task-relevant information and learns them progressively, while avoiding pitfalls and computational burdens (see supp)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.505, + 0.681, + 0.523 + ], + "angle": 0, + "content": "3. Our Proposed ES3" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.531, + 0.642, + 0.545 + ], + "angle": 0, + "content": "3.1. Preliminaries" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.553, + 0.895, + 0.78 + ], + "angle": 0, + "content": "The goal of \\(\\mathrm{ES}^3\\) is to learn audio-visual speech representations from unlabeled talking face videos. The inputs to our \\(\\mathrm{ES}^3\\) are paired audio-visual data \\(X_{a,v} = \\{(X_{a,i},X_{v,i})\\mid i = 1,2,\\ldots ,N\\}\\). Each of the \\(N\\) samples consists of a sequence of audio frames \\(X_{a,i} = \\operatorname {LogFBank}(\\mathbf{a}_i)\\in \\mathbb{R}^{F\\times T_{a,i}}\\) and a temporally synchronized video sequence \\(X_{v,i} = \\mathbf{v}_i\\in \\mathbb{R}^{C\\times H\\times W\\times T_{v,i}}\\) showing the lips of a speaker, where \\(F\\) is the filter bank dimension, \\(C\\) is the number of video channels, \\(T_{a,i}\\) and \\(T_{v,i}\\) are the total number of audio and video frames in the \\(i\\)-th sample, and \\(H\\) and \\(W\\) are the width and height of each frame. Audio and video are sampled at 100 and 25 FPS, i.e. for a 1-second clip, \\(T_{a} = 4T_{v} = 100\\). The goal is to learn a \\(d\\)-dimensional contextualized representation \\(\\gamma \\in \\mathbb{R}^{d\\times T_v}\\) capturing vital speech and context information within the two modality spaces." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.781, + 0.895, + 0.901 + ], + "angle": 0, + "content": "Base architecture. The base architecture used in this paper is a relatively simple one and can be easily extended if needed. It includes two modality-specific encoders \\(\\varphi_{a}\\) and \\(\\varphi_{v}\\), instantiated as a 2-layer CNN-based audio filter bank encoder and a lightweight VGG-(2+1)D video encoder [57, 58]. Here, \\(\\varphi_{a}\\) downsamples \\(X_{a}\\) by 4x to match the video frame rate; \\(\\varphi_{v}\\) has a temporal receptive field of 5 frames, or 200ms at 25 FPS, same as the most com" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "27070" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.228 + ], + "angle": 0, + "content": "monly adopted 3D-ResNet18 encoder [41, 59, 64], but has much fewer parameters. A modality-agnostic Transformer-based [71] modality fusion encoder \\(\\theta\\) equipped with rotary positional embeddings (RoPE) [65] follows the modality-specific encoders \\(\\varphi_{a}\\) and \\(\\varphi_{v}\\). To allow \\(\\theta\\) to accept either audio, video, or audio-visual features, we adopt the simple sum operation to compute a fused audio-visual representation \\(F_{\\mathrm{fused}}\\) from modality features \\(F_{m} = \\varphi_{m}(X_{m})\\), where \\(m \\in \\{a, v\\}\\), to allow easy handling of missing modalities:" + }, + { + "type": "equation", + "bbox": [ + 0.096, + 0.236, + 0.47, + 0.29 + ], + "angle": 0, + "content": "\\[\nF _ {\\text {f u s e d}} = \\left\\{ \\begin{array}{l l} F _ {a} + F _ {v} & \\text {i f i n p u t i s a u d i o - v i s u a l ,} \\\\ F _ {a} + \\mathbf {0} = F _ {a} & \\text {i f i n p u t i s a u d i o ,} \\\\ F _ {v} + \\mathbf {0} = F _ {v} & \\text {i f i n p u t i s v i d e o .} \\end{array} \\right. \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.297, + 0.47, + 0.327 + ], + "angle": 0, + "content": "The Transformer encoder \\(\\theta\\) then yields the final contextual features \\(E = \\theta (F_{\\mathrm{fused}})\\)" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.327, + 0.472, + 0.51 + ], + "angle": 0, + "content": "Based on this base architecture, we build our Siamese based learning framework as shown in Fig. 1(a). A student model (left) is guided by a teacher model (right), which share the same base architecture. The student and teacher accept different versions of the same speech inputs (e.g. masked vs unmasked). The teacher model updates its parameters in an exponential moving average (EMA) manner, and the student is trained to regress the teacher outputs in general. In the subsequent sections, our objective is to optimize the student model with the base architecture to obtain the parameterized audio-visual speech representations \\(\\gamma (\\mathbf{a},\\mathbf{v};\\varphi_{a},\\varphi_{v},\\theta) = E\\)" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.515, + 0.296, + 0.53 + ], + "angle": 0, + "content": "3.2. Problem Reformulation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.539, + 0.47, + 0.81 + ], + "angle": 0, + "content": "In the audio-visual speech representation learning task, audio and video inputs can be taken as two different views of the same underlying speech signal, carrying different types of information. The relationship between the speech information in the two modalities can be visualized as the bottom-left Venn diagram in Fig. 1(b). Owing to the asymmetry between the two modalities, there exists both modality-unique speech information, i.e. audio- and video-unique information \\( U_{a} \\) and \\( U_{v} \\). There is also speech information \\( R \\) shared between the two modalities, since they share the same underlying speech production process. At the same time, information that only emerges when both audio and video modality are present is the synergistic information, which we denote by \\( S \\). Our overall objective is to learn a representation that maximizes the total information \\( R + U_{a} + U_{v} + S \\) for bi-modal (audio-visual) speech representations, as well as \\( R + U_{a} \\) and \\( R + U_{v} \\) for uni-modal (audio & visual) speech representations." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.811, + 0.471, + 0.902 + ], + "angle": 0, + "content": "With this in mind, we reframe the task of learning audiovisual speech representations as a mutual information maximization task. Specifically, the two modalities, corresponding to source variables \\( X_{a} \\) and \\( X_{v} \\), have mutual information with the target speech content variable \\( Y \\), containing information useful for downstream tasks that we" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.288 + ], + "angle": 0, + "content": "wish to capture. Denote the joint mutual information between the sources and the target by \\( I(X_{a},X_{v};Y) \\). The problem of learning robust audio-visual speech representations reduces to finding the set of parameters that maximizes \\( I(\\gamma (X_a),\\gamma (X_v);Y) = I(E_{av};Y)^1 \\). Similarly, the problem of learning robust audio and visual speech representations can be viewed as maximizing the mutual information \\( I(\\gamma (X_a);Y) = I(E_a;Y) \\) and \\( I(\\gamma (X_v);Y) = I(E_v;Y) \\), respectively. Taking advantage of the Partial Information Decomposition (PID) framework [73], we can precisely decompose the multivariate mutual information \\( I(E_{av};Y),I(E_a;Y) \\) and \\( I(E_{v};Y) \\) into the previous types of information as:" + }, + { + "type": "equation", + "bbox": [ + 0.641, + 0.301, + 0.891, + 0.317 + ], + "angle": 0, + "content": "\\[\nR + U _ {a} = I \\left(E _ {a}; Y\\right), \\tag {2}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.643, + 0.32, + 0.891, + 0.336 + ], + "angle": 0, + "content": "\\[\nR + U _ {v} = I \\left(E _ {v}; Y\\right), \\tag {3}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.643, + 0.339, + 0.891, + 0.354 + ], + "angle": 0, + "content": "\\[\nU _ {a} + S = I \\left(E _ {a}; Y \\mid X _ {v}\\right), \\tag {4}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.643, + 0.358, + 0.891, + 0.373 + ], + "angle": 0, + "content": "\\[\nU _ {v} + S = I \\left(E _ {v}; Y \\mid X _ {a}\\right), \\tag {5}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.651, + 0.377, + 0.891, + 0.393 + ], + "angle": 0, + "content": "\\[\nR - S = I \\left(E _ {a}; E _ {v}; Y\\right), \\tag {6}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.573, + 0.396, + 0.891, + 0.412 + ], + "angle": 0, + "content": "\\[\nR + U _ {a} + U _ {v} + S = I \\left(E _ {a v}; Y\\right), \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.424, + 0.774, + 0.438 + ], + "angle": 0, + "content": "where \\(R, U_{a}, U_{v}\\) and \\(S\\) are non-negative." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.439, + 0.892, + 0.668 + ], + "angle": 0, + "content": "To optimize the above mutual information terms with a neural network parameterized by \\(\\varphi_{a},\\varphi_{v}\\) and \\(\\theta\\) we divide the optimization process into three progressive and nonconflicting stages. In stage 1, we learn audio-unique information \\(U_{a}\\) and shared information \\(R\\) from the informative audio modality to maximize \\(I(E_{a};Y)\\). This process also helps to bootstrap visual representations by learning the shared information \\(R\\). In stage 2, we turn to learning visual unique information \\(U_{v}\\) by maximizing \\(I(E_{v};Y) = R + U_{v}\\), which is now easier with the visual information captured via \\(R\\) in stage 1. At the same time, we inject the learned visual information into the audio-visual joint representations to initialize \\(I(E_{av};Y)\\). Finally, in stage 3 we optimize the joint audio-visual representation on top of the previously learned information by maximizing \\(I(E_{av};Y)\\)." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.676, + 0.816, + 0.692 + ], + "angle": 0, + "content": "3.3. The Complete Learning Framework" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.699, + 0.892, + 0.866 + ], + "angle": 0, + "content": "To learn network parameters that maximize the mutual information terms \\( I(E;Y) \\), we leverage the popular self-supervised masked prediction task based on the simple Siamese framework to support the learning process via self-distillation. Inspired by the recent work of Baevski et al. [12], we create multiple masked copies for each training sample, encode the unmasked regions with the Transformer encoder, and train the student to recover contextualized targets of the masked regions with a lightweight 1D convolutional decoder, which are generated by an online EMA teacher. This task can be used to learn any mutual infor" + }, + { + "type": "page_footnote", + "bbox": [ + 0.5, + 0.875, + 0.892, + 0.901 + ], + "angle": 0, + "content": "1This equality holds since the architecture implemented in this work only models additive modality synergy stemming from sum fusion." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "27071" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.084, + 0.093, + 0.892, + 0.386 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.388, + 0.894, + 0.43 + ], + "angle": 0, + "content": "Figure 1. Overview of the proposed method, \\(\\mathbf{ES}^3\\). (a) We use a Siamese framework with an EMA teacher to implement an iteration-free training procedure. (b) Venn diagrams showing how the three types of information \\((R, U, \\text{and } S)\\) are acquired in an evolving learning process. (c) Illustration of self-distillation with online clustering corresponding to \\(\\mathcal{L}_{\\mathrm{distill}}(F_{\\mathrm{tea}}, F_{\\mathrm{stu}})\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.439, + 0.47, + 0.623 + ], + "angle": 0, + "content": "mation \\(I(E_{;}Y)\\) when \\(Y\\) is the downstream speech understanding task. For example, training on auditory or visual speech acts as a proxy to maximize \\(I(E_{a};Y)\\) or \\(I(E_{v};Y)\\). Self-distillation with online clustering [36]. We perform distillation from the right-hand side teacher to the left-hand side student in Fig. 1(a). Specifically, as illustrated in Fig. 1(c), for each layer \\(k\\) of the top \\(K\\) Transformer layers in \\(\\theta\\) with hidden dimension \\(D\\), we learn a codebook \\(\\mathbf{Q}^{k} = \\{\\mathbf{q}_{1}^{k},\\dots,\\mathbf{q}_{Q}^{k}\\}\\) with \\(Q\\) codewords \\(\\mathbf{q}_i^k\\in \\mathbb{R}^D\\), which is updated as follows: for each entry \\(j\\), a set \\(\\tilde{\\mathbf{Z}}_{\\mathrm{tea}}^k\\) of the teacher output features \\(\\tilde{\\mathbf{z}}_{\\mathrm{tea}}^k\\) closest to the current representation of \\(q\\) in the codebook is found as" + }, + { + "type": "equation", + "bbox": [ + 0.101, + 0.633, + 0.47, + 0.669 + ], + "angle": 0, + "content": "\\[\n\\tilde {\\mathbf {Z}} _ {\\mathrm {t e a}} ^ {k} = \\left\\{\\tilde {\\boldsymbol {z}} _ {\\mathrm {t e a}} ^ {k} \\Bigg | q = \\underset {j \\in \\{1, 2, \\dots , Q \\}} {\\arg \\min } | | \\tilde {\\boldsymbol {z}} _ {\\mathrm {t e a}} ^ {k} - \\mathbf {q} _ {j} ^ {k} | | _ {2} \\right\\}, \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.678, + 0.47, + 0.739 + ], + "angle": 0, + "content": "where the set indices \\(q\\) are used as discrete pseudo-labels to train the student model. Codewords are then updated in an EMA fashion as a weighted sum of the embeddings in \\(\\tilde{\\mathbf{Z}}_{\\mathrm{tea}}^k\\) with decay rate \\(\\tau_{\\mathrm{code}}\\). The associated loss function is" + }, + { + "type": "equation", + "bbox": [ + 0.094, + 0.748, + 0.469, + 0.794 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {d i s t i l l}} \\left(F _ {\\text {t e a}}, F _ {\\text {s t u}}\\right) = \\sum_ {t \\in \\mathbb {M}} \\sum_ {k \\in (K - N, K ]} \\log p _ {\\psi_ {k}} \\left(q _ {t} \\mid \\boldsymbol {z} _ {t} ^ {K}\\right), \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.796, + 0.47, + 0.901 + ], + "angle": 0, + "content": "where \\(F_{\\mathrm{tea}}\\) and \\(F_{\\mathrm{stu}}\\) are teacher and student features, \\(\\mathbb{M}\\) is the set of masked time steps, \\(\\psi_{k}\\) is the softmax activations of linear projections for each target layer \\(k\\), \\(q_{t}\\) is the codeword index for time step \\(t\\) of the teacher \\(\\tilde{\\theta}\\), and \\(z_{t}^{K}\\) is the output of the \\(K\\)-th and final layer of the student model \\(\\theta\\). The teacher and the online clustering process do not require gradients. EMA teacher parameterization. To reduce GPU memory" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.44, + 0.892, + 0.515 + ], + "angle": 0, + "content": "usage, we apply EMA updates to the Transformer encoder \\(\\theta\\), but not the earlier modality-specific encoders \\(\\varphi_{a}\\) and \\(\\varphi_{v}\\), inspired by [50]. As shown in Fig. 1, the teacher model weights \\(\\tilde{\\theta}\\) are computed as an exponential moving average of the student model weights \\(\\theta\\):" + }, + { + "type": "equation", + "bbox": [ + 0.626, + 0.525, + 0.892, + 0.544 + ], + "angle": 0, + "content": "\\[\n\\tilde {\\theta} \\leftarrow \\eta \\tilde {\\theta} + (1 - \\eta) \\theta , \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.555, + 0.892, + 0.601 + ], + "angle": 0, + "content": "where \\(\\eta = \\min \\{\\eta_{\\max}, (1 + \\text{step}) / (10 + \\text{step})\\}\\) is a momentum parameter that slowly increases over time until a pre-determined value \\(\\eta_{\\max}\\) is reached." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.611, + 0.704, + 0.627 + ], + "angle": 0, + "content": "3.4. The Learning Process" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.634, + 0.892, + 0.663 + ], + "angle": 0, + "content": "We now elaborate on individual learning stages, based on the above Siamese framework and distillation loss." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.664, + 0.892, + 0.738 + ], + "angle": 0, + "content": "Stage 1: Learning audio-unique information and shared information. In stage 1, we perform distillation from un-masked audio teacher model to the masked audio student model, yielding audio-unique information \\( U_{a} \\) and shared information \\( R \\) from audio. The distillation loss is:" + }, + { + "type": "equation", + "bbox": [ + 0.607, + 0.752, + 0.891, + 0.768 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {a} \\rightarrow \\mathrm {a} ^ {\\prime}} = \\mathcal {L} _ {\\text {d i s t i l l}} \\left(E _ {a}, E _ {a} ^ {\\prime}\\right). \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.78, + 0.892, + 0.9 + ], + "angle": 0, + "content": "Being readily learnable, audio has the potential to provide most of \\( R \\) for learning shared information. However, optimizing \\( I(E_{a};Y) \\) alone cannot help visual learning. Noting that the InfoNCE loss between two variables has been widely used to maximize a lower bound on the mutual information, we adopt the InfoNCE loss to learn shared information corresponding to \\( I(E_{a};E_{v}) \\) to bootstrap visual learning with the shared information learned from audio." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "27072" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.213 + ], + "angle": 0, + "content": "Specifically, we apply it to a fine-grained frame-wise alignment task, where temporally aligned features from the same video are treated as positives, and unaligned ones negatives [19]. Since phonemes can appear repeatedly within the same video clip, leading to false negatives, we use a negative margin following Park et al. [47], which alleviates the effect of falsely matched pairs by enforcing a looser decision boundary. The objective function involved is:" + }, + { + "type": "equation", + "bbox": [ + 0.088, + 0.223, + 0.47, + 0.274 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {N C E}} (H) = - \\frac {1}{n} \\sum_ {i = 1} ^ {n} \\log \\frac {e ^ {(H _ {i , i} - m) / \\tau}}{e ^ {(H _ {i , i} - m) / \\tau} + \\sum_ {i \\neq j} e ^ {H _ {i , j} / \\tau}}, \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.275, + 0.473, + 0.38 + ], + "angle": 0, + "content": "where \\(H\\) is a matrix whose \\((i,j)\\)-th element represents the cosine similarity between \\(L^2\\)-normalized audio and visual features at time step \\(i\\) and \\(j\\), respectively, \\(\\tau\\) is a learnable temperature parameter, and \\(m\\) is a pre-determined margin. To promote the learning of shared features between the two modalities, we apply a symmetrized version of the above NCE loss as:" + }, + { + "type": "equation", + "bbox": [ + 0.078, + 0.387, + 0.475, + 0.427 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {a} \\leftrightarrow \\mathrm {v}} = \\frac {1}{2} \\left(\\mathcal {L} _ {\\mathrm {a} \\rightarrow \\mathrm {v}} + \\mathcal {L} _ {\\mathrm {v} \\rightarrow \\mathrm {a}}\\right) = \\frac {1}{2} \\left(\\mathcal {L} _ {\\mathrm {N C E}} \\left(S _ {a v}\\right) + \\mathcal {L} _ {\\mathrm {N C E}} \\left(S _ {v a}\\right)\\right). \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.428, + 0.47, + 0.504 + ], + "angle": 0, + "content": "Here the similarity matrices \\( S_{av} = \\langle F_a, F_v \\rangle / (||F_a|| \\cdot ||F_v||) \\) and \\( S_{va} = S_{av}^{\\top} \\) are computed with features \\( F_a \\) and \\( F_v \\) from modality-specific encoders \\( \\varphi_a \\) and \\( \\varphi_v \\) instead of the contextual features from the Transformer encoder \\( \\theta \\) to avoid learning positional shortcuts. The total loss in stage 1 is" + }, + { + "type": "equation", + "bbox": [ + 0.182, + 0.516, + 0.47, + 0.532 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {s t a g e 1}} = \\mathcal {L} _ {\\mathrm {a} \\rightarrow \\mathrm {a} ^ {\\prime}} + \\mathcal {L} _ {\\mathrm {a} \\leftrightarrow \\mathrm {v}}. \\tag {14}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.539, + 0.473, + 0.676 + ], + "angle": 0, + "content": "Stage 2: Learning video-unique information. In stage 2, we proceed to learn video-unique information, which is now much easier than learning from scratch since it has been bootstrapped with shared knowledge from the audio in the previous stage. The audio encoder \\(\\varphi_{a}\\) does not receive any gradients in this stage and is essentially frozen, since it is expected to have already acquired the audio-specific knowledge well in the previous stage. We adopt a similar distillation loss as stage 1:" + }, + { + "type": "equation", + "bbox": [ + 0.184, + 0.686, + 0.469, + 0.703 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {v} \\rightarrow \\mathrm {v} ^ {\\prime}} = \\mathcal {L} _ {\\text {d i s t i l l}} \\left(E _ {v}, E _ {v} ^ {\\prime}\\right). \\tag {15}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.708, + 0.47, + 0.799 + ], + "angle": 0, + "content": "On top of the learned visual-unique information, we also perform distillation from unmasked video features to the joint features of unmasked audio-visual inputs in this stage to prepare for the learning of the final robust audio-visual representations in the next stage. For this purpose, the loss in this stage is set as:" + }, + { + "type": "equation", + "bbox": [ + 0.18, + 0.81, + 0.469, + 0.826 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {a v} \\rightarrow \\mathrm {v}} = \\mathcal {L} _ {\\text {d i s t i l l}} \\left(E _ {v}, E _ {a v}\\right), \\tag {16}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.18, + 0.829, + 0.469, + 0.845 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {s t a g e 2}} = \\mathcal {L} _ {\\mathrm {v} \\rightarrow \\mathrm {v} ^ {\\prime}} + \\mathcal {L} _ {\\mathrm {a v} \\rightarrow \\mathrm {v}}. \\tag {17}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.47, + 0.901 + ], + "angle": 0, + "content": "To avoid repeated learning of only shared information between A and V again, the cross-modal loss \\(\\mathcal{L}_{\\mathrm{av}\\to \\mathrm{v}}\\) is asymmetric, strictly one-way from AV to V, and a learnable tem" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.137 + ], + "angle": 0, + "content": "perature \\(\\tau_{v}\\) Eq. (16) is introduced to the visual student outputs, i.e., we replace \\(\\psi_{k}\\) in Eq. (9) with softmax activations of temperature-scaled linear projections of \\(E_{v}\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.137, + 0.893, + 0.439 + ], + "angle": 0, + "content": "Stage 3: Learning synergistic information. Finally, we aim to learn synergistic audio-visual information which previous methods have not explicitly addressed. Our ultimate goal is to maximize total task-relevant multi-modal information \\( I(E_{av};Y) \\), i.e. the sum of \\( R \\), \\( U_{a} \\), \\( U_{v} \\) and \\( S \\), so that the representations are useful for both uni-modal and multi-modal tasks. Naively maximizing this objective without prior stages is problematic because the network tends to be biased towards learning audio-based representations (i.e. prioritizing \\( R \\) and \\( U_{a} \\)) without considering the visual modality. By learning most of \\( R \\), \\( U_{a} \\) and \\( U_{v} \\) through the previous stages, our method is less susceptible to learning biased representations in the final stage. At this stage, we always mask identical positions in the audio and video inputs together so that the masked region never contains any unimodal information that can collapse or degrade training. To better recover the masked inputs, the model must cleverly leverage both modalities, which requires learning synergistic information that is not well-captured in prior stages. The loss function in stage 3 can be written as:" + }, + { + "type": "equation", + "bbox": [ + 0.558, + 0.449, + 0.892, + 0.466 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {s t a g e 3}} = \\mathcal {L} _ {\\mathrm {a v} \\rightarrow \\mathrm {a v} ^ {\\prime}} = \\mathcal {L} _ {\\text {d i s t i l l}} \\left(E _ {a v} ^ {\\prime}, E _ {a v}\\right). \\tag {18}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.47, + 0.892, + 0.53 + ], + "angle": 0, + "content": "By progressively acquiring the three types of useful information, we build strong and robust audio-visual speech representations for both uni- and bi-modal speech, which we shall show in subsequent experiments." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.539, + 0.673, + 0.555 + ], + "angle": 0, + "content": "3.5. Fine-tuning Stage" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.562, + 0.892, + 0.728 + ], + "angle": 0, + "content": "After pre-training, we add an extra linear layer after the fusion encoder \\(\\theta\\) to fine-tune the model end-to-end with the Lattice-Free Maximum Mutual Information (LF-MMI) criterion [26, 51], which is much more lightweight than seq2seq Encoder-Decoder models and allows highly efficient decoding. We adopt a \"compact\" [33] 1-state hidden Markov model (HMM) topology, using \\(k2^2\\) for differentiable Weighted Finite State Transducer (WFST)-based training. We use 624 bi-char units from text-based clustering [77] for English and tonal phone-level targets (initials and finals) for Chinese." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.74, + 0.633, + 0.758 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.765, + 0.602, + 0.78 + ], + "angle": 0, + "content": "4.1. Datasets" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.788, + 0.892, + 0.88 + ], + "angle": 0, + "content": "LRS2-BBC [3]. LRS2-BBC is a 224-hour transcribed audio-visual English speech dataset collected from programs of various genres broadcast by the BBC. It is challenging due to the presence of extreme poses, as well as a large and diverse vocabulary. We use the official train / validation split (see supplementary materials) for evaluation." + }, + { + "type": "page_footnote", + "bbox": [ + 0.517, + 0.887, + 0.714, + 0.9 + ], + "angle": 0, + "content": "2https://github.com/k2-fsa/k2" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "27073" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.091, + 0.47, + 0.212 + ], + "angle": 0, + "content": "LRS3-TED [1]. LRS3-TED is another audio-visual English speech dataset collected from TED and TEDx talks, with over 5.5k speakers and 439 hours of data. It covers diverse topics and is widely adopted for evaluating audio-visual speech representation learning methods for being the largest publicly available one of its kind. An official train / validation split is not given, so we use the widely adopted one prepared by Shi et al. [59], same as [27, 34, 80]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.213, + 0.47, + 0.289 + ], + "angle": 0, + "content": "CAS-VSR-S101. We also collect a new large-scale, in-the-wild Mandarin dataset with 101.1 hours of data. The videos are sourced from broadcast news and conversational programs in Chinese, covering a highly diverse set of topics, speakers and filming conditions (see supplementary)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.299, + 0.267, + 0.315 + ], + "angle": 0, + "content": "4.2. Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.323, + 0.468, + 0.352 + ], + "angle": 0, + "content": "More implementation details can be found in the supplementary materials." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.353, + 0.468, + 0.565 + ], + "angle": 0, + "content": "High- and low-resource setups. On LRS3-TED we follow the widely adopted setup of Shi et al. [59], which excludes the pre-train set during low-resource fine-tuning. To validate data scaling properties, previous works used VoxCeleb2-En [59], which is a 1326-hour subset of VoxCeleb2 [18] in English. However, not only are experiments of this scale computationally expensive, but more importantly, VoxCeleb2 is no longer available from its official source as of Feb. 2023. Therefore, we create a similar scaling protocol, using LRS3-TED (433h) as a pre-training source and LRS2-BBC (223h) for fine-tuning. For low-resource fine-tuning, we only use the 28-hour train subset of LRS2-BBC, while for high-resource fine-tuning, we include the pre-train subset, similar to [17]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.565, + 0.468, + 0.701 + ], + "angle": 0, + "content": "Evaluation. Decoding is performed using beam search with a beam size of 50. The common evaluation metrics for ASR, VSR and AVSR are Word Error Rate (WER) and Character Error Rate (CER) for English and Chinese, respectively. As our bi-char LF-MMI models are trained on verbatim transcriptions for ease of lexicon construction, we report WERs both without and with using Whisper-style text normalization [54] (in parentheses) to compensate for innocuous errors related to text normalization." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.711, + 0.409, + 0.726 + ], + "angle": 0, + "content": "4.3. Results on LRS2-BBC and LRS3-TED" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.735, + 0.468, + 0.901 + ], + "angle": 0, + "content": "Low-resource setup. Low-resource results on LRS2-BBC and LRS3-TED can be found in Tabs. 1 and 2. Our smallest BASE* encoder is modality-agnostic, only half the size of the one in [27], which uses two separate encoders to perform AVSR. On LRS3-TED, although we employ the simple LF-MMI instead of cross-entropy (CE) based seq2seq criterion for fine-tuning, our BASE model rivals or achieves SoTA in all cases. When scaling to a LARGE model, we observe significant improvements in both ASR and AVSR. Meanwhile, VSR performance remains comparable to SoTA, with a moderate \\(2\\%\\) gap. This may sug" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.136 + ], + "angle": 0, + "content": "gest that with our current VGG-based visual encoder \\(\\varphi_v\\), the benefits of scaling up the Transformer primarily manifest in ASR and AVSR." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.138, + 0.892, + 0.379 + ], + "angle": 0, + "content": "High-resource setup. High-resource results on LRS2-BBC and LRS3-TED can be found in Tabs. 1 and 3. On LRS2-BBC, with a 46M model and relying on its own 223h data for pre-training and fine-tuning, we achieve \\(30.3\\% / 3.6\\% / 2.9\\%\\) WER for VSR, ASR and AVSR. Notably, these results surpass a highly competitive supervised method [40] and other self-supervised methods that incorporate additional data. This indicates that our progressive learning approach can effectively exploit visual information. Furthermore, when we introduce additional 433h unlabeled data, we achieve even better performance: \\(29.8\\% / 3.0\\% / 2.4\\%\\) WER, which outperforms other methods that employ the same amount of data. When we switch to the LARGE model, performance rivals SoTA methods that employ 4x data (1759h) for pre-training. Similarly, we achieve or rival SoTA performance on LRS3-TED." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.391, + 0.742, + 0.405 + ], + "angle": 0, + "content": "4.4. Results on CAS-VSR-S101" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.414, + 0.892, + 0.581 + ], + "angle": 0, + "content": "We also evaluate on our newly collected CAS-VSR-S101 dataset, which is smaller than both of the two previous datasets and in an even more ambiguous language (Chinese has more homophonous syllables and words than English). As shown in Tab. 4, \\(\\mathrm{ES}^3\\) still works out-of-the-box under this challenging setup, and yields noticeable performance gains over a strong fully supervised baseline [39]. Interestingly, our fully supervised baseline model beats [39] in ASR and AVSR but slightly lags behind in VSR, again hinting that our strategy can further benefit from more complex visual modeling architectures." + }, + { + "type": "table_caption", + "bbox": [ + 0.498, + 0.585, + 0.892, + 0.655 + ], + "angle": 0, + "content": "Table 4. Results on CAS-VSR-S101. Here we use a BASE* model for the fully-supervised baseline and \\(\\mathrm{ES}^3\\). The self-supervised models are pre-trained on training data and then finetuned for 50 epochs with the labels. The numbers before and after slashes correspond to CERs on the validation and test sets." + }, + { + "type": "table", + "bbox": [ + 0.504, + 0.659, + 0.892, + 0.733 + ], + "angle": 0, + "content": "
MethodsCriterionEncoder Size#EpochsVSRASRAVSR
Supervised
BaselineLF-MMI46M10060.0 / 57.415.4 / 12.614.6 / 12.1
E2E Conformer [39]CTC+CE79M10058.6 / 56.516.6 / 13.016.0 / 12.6
Self-supervised
ES3 Stage 161.2 / 59.014.7 / 11.9-
ES3 Stage 2LF-MMI46M5059.1 / 56.7-13.9 / 11.4
ES3 Stage 358.1 / 55.614.4 / 11.613.4 / 11.0
" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.741, + 0.72, + 0.757 + ], + "angle": 0, + "content": "4.5. Analysis and Discussion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Representation quality probing. In Tab. 5, we demonstrate the effect of each stage on LRS2-BBC. ASR performance between stages 1 and 3 are close, indicating that stage 1 indeed acquires most of audio-unique information. Comparing the VSR results between stages 1 and 2 reveals a significant gap, suggesting that stage 2 also effectively learns visual-unique information, hence the minimal differences between the final VSR result and those of stage 2. Moreover, comparing the results of stage 1-2 and stage" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "27074" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.089, + 0.892, + 0.118 + ], + "angle": 0, + "content": "Table 1. Results on LRS2-BBC. We pre-train a BASE* model with 223h unlabeled data from LRS2-BBC, as well as BASE and LARGE models with 433h unlabeled data (LRS3-TED) to demonstrate scaling properties. *: external language models." + }, + { + "type": "table", + "bbox": [ + 0.162, + 0.122, + 0.812, + 0.394 + ], + "angle": 0, + "content": "
MethodsUnlabeled AV dataLabeled DataEncoder SizeCriterionVSRASRAVSR
Supervised
Yu et al. [75]-1519h-LF-MMI148.96.75.9
Ma et al. [39]*-380h79MCTC+CE37.93.93.7
Ma et al. [41]*-818h186MCTC+CE27.92.6-
Semi-supervised
Afouras et al. [2]*777h223h2-CTC51.3--
Ma et al. [41]*2630h818h186MCTC+CE14.61.51.5
Self-supervised (BASE and BASE* models)
AV-HuBERT [59, 60]1759h223h103MCE31.23-3.63
VATLM [80]1759h4223h107MCE30.6-2.9
RAVeN [27]433h223h97MCTC+CE32.13.9-
Pan et al. [45]-5380h399MCTC+CE43.22.72.6
ES3(ours)223h28h46M40.2 (39.1)6.0 (5.0)5.7 (4.9)
433h28h102MLF-MMI39.3 (38.2)5.5 (4.8)5.1 (4.1)
223h223h46M31.4 (30.3)4.3 (3.6)3.8 (2.9)
433h223h102M30.7 (29.8)3.4 (3.0)3.2 (2.4)
Self-supervised (LARGE models)
ES3(ours)433h28h317MLF-MMI36.4 (35.4)5.2 (4.4)4.7 (4.0)
433h223h317M26.7 (25.8)3.1 (2.5)3.1 (2.5)
AV-HuBERT [59, 60]1759h223h325MCE25.53-2.53
VATLM [80]1759h4223h332MCE24.3-2.3
RAVeN [27]1759h223h671MCTC+CE23.22.5-
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.088, + 0.396, + 0.767, + 0.408 + ], + "angle": 0, + "content": "\\( {}^{1} \\) Not end-to-end; requires a GMM-HMM alignment stage. \\( {}^{2} \\) Uses an additional ASR model trained on LibriSpeech (960h)." + }, + { + "type": "table_footnote", + "bbox": [ + 0.088, + 0.408, + 0.651, + 0.42 + ], + "angle": 0, + "content": "3 Reproduced by Zhu et al. [80]. 4 Uses additional 3846h audio, 452h audio-text and 600M text data." + }, + { + "type": "table_footnote", + "bbox": [ + 0.089, + 0.42, + 0.437, + 0.432 + ], + "angle": 0, + "content": "5 Uses additional 60000h audio data and 1.28M unlabeled images." + }, + { + "type": "list", + "bbox": [ + 0.088, + 0.396, + 0.767, + 0.432 + ], + "angle": 0, + "content": null + }, + { + "type": "table_caption", + "bbox": [ + 0.145, + 0.436, + 0.822, + 0.449 + ], + "angle": 0, + "content": "Table 2. Low-resource results on LRS3-TED. We pre-train a BASE and LARGE model with 433h unlabeled data." + }, + { + "type": "table", + "bbox": [ + 0.163, + 0.455, + 0.812, + 0.598 + ], + "angle": 0, + "content": "
MethodsUnlabeled AV dataLabeled DataEncoder SizeCriterionVSRASRAVSR
Self-supervised (BASE models)
AV-HuBERT [59, 60]433h30h103MCE51.84.9\\( {4.7}^{1} \\)
VATLM [80]\\( {433}{\\mathrm{\\;h}}^{2} \\)30h107MCE48.0-3.6
RAVEn [27]433h30h97MCTC+CE47.04.7-
\\( {\\mathrm{{AV2vec}}}^{\\left\\lbrack {76}\\right\\rbrack }{}^{3} \\)433h30h103MCE45.0-5.8
AV-data2vec [34]433h30h103MCE45.24.44.2
\\( {\\mathbf{{ES}}}^{3} \\) (ours)433h30h102MLF-MMI45.5 (44.7)3.9 (3.3)3.6 (3.0)
Self-supervised (LARGE models)
AV-HuBERT433h30h325MCE44.84.5\\( {4.2}^{1} \\)
AV-data2vec433h30h325MCE40.53.73.4
\\( {\\mathbf{{ES}}}^{3} \\) (ours)433h30h317MLF-MMI43.5 (42.5)3.8 (2.9)2.9 (2.3)
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.088, + 0.6, + 0.654, + 0.612 + ], + "angle": 0, + "content": "\\(^{1}\\) Reproduced by Lian et al. [34]. \\(^{2}\\) Uses additional 3846h audio, 452h audio-text and 600M text data." + }, + { + "type": "table_footnote", + "bbox": [ + 0.088, + 0.612, + 0.794, + 0.625 + ], + "angle": 0, + "content": "3 Zhang et al. [76] inject noise during pre-training and obtain better fine-tuning results even with its base model AV-HuBERT \\((47.1\\%)\\)." + }, + { + "type": "list", + "bbox": [ + 0.088, + 0.6, + 0.794, + 0.625 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.628, + 0.468, + 0.72 + ], + "angle": 0, + "content": "3 demonstrates that the joint information learning in stage 3, which also incorporates synergistic information, benefits both uni-modal (ASR & VSR) and bi-modal (AVSR) tasks. Similar trends can also be observed on the more challenging Mandarin dataset CAS-VSR-S101 in Tab. 4, further supporting the motivation of our proposed strategy." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.721, + 0.468, + 0.763 + ], + "angle": 0, + "content": "Table 5. Representation quality probing on LRS2-BBC. Results are reported on the validation set using the BASE* model under the low-resource setup for each pre-training stage." + }, + { + "type": "table", + "bbox": [ + 0.11, + 0.768, + 0.441, + 0.82 + ], + "angle": 0, + "content": "
Pre-training StageVSRASRAVSR
Stage 158.3 (57.3)11.5 (10.1)-
Stage 248.7 (47.3)-12.9 (11.5)
Stage 3 (Full ES³)48.1 (46.9)11.4 (9.8)10.1 (8.6)
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.826, + 0.468, + 0.901 + ], + "angle": 0, + "content": "Effect of automatically balancing shared and unique information learning. It is known that in a hardness-aware loss function, the temperature rules the strength of penalties on hard negative samples [72]. We track the automatically learned temperature parameter \\(\\tau\\) used for cross-modal" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.628, + 0.892, + 0.795 + ], + "angle": 0, + "content": "learning objectives \\(\\mathcal{L}_{\\mathrm{a}\\leftrightarrow \\mathrm{v}}\\) in Eq. (12) and \\(\\mathcal{L}_{\\mathrm{av}\\rightarrow \\mathrm{v}}\\) in Eq. (16) when pre-training on LRS2-BBC, as shown in Fig. 2. In stage 1, \\(\\tau\\) continues to fall until it reaches \\(\\tau_{\\mathrm{min}}\\), while in stage 2 it briefly drops and then rises throughout training. The trend is in line with our intuition: in stage 1, audio information is easily learnable, so the model can fully exploit shared information by lowering \\(\\tau\\) without sacrificing audio-unique information learning; in stage 2, after learning some residual shared information, an increasing \\(\\tau\\) progressively relaxes \\(\\mathcal{L}_{\\mathrm{av}\\rightarrow \\mathrm{v}}\\) to yield more freedom to \\(\\mathcal{L}_{\\mathrm{v}\\rightarrow \\mathrm{v}'}\\), i.e. video-unique information learning." + }, + { + "type": "image", + "bbox": [ + 0.561, + 0.806, + 0.828, + 0.905 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.5, + 0.906, + 0.891, + 0.92 + ], + "angle": 0, + "content": "Figure 2. The temperature parameter paces learning automatically." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "27075" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.09, + 0.892, + 0.119 + ], + "angle": 0, + "content": "Table 3. High-resource results on LRS3-TED. We fine-tune on 433h labeled data of LRS3. †: uses test-time augmentation. *: uses external language models. ‡: noise injection during pre-training. We list more methods for comparison in the supplementary material." + }, + { + "type": "table", + "bbox": [ + 0.142, + 0.123, + 0.832, + 0.372 + ], + "angle": 0, + "content": "
MethodsYearUnlabeled AV dataLabeled DataBackboneEncoder SizeCriterionVSRASRAVSR
Supervised
Ma et al. [39]*2021-595hConformer79MCTC+CE43.32.32.3
Prajwal et al. [52]†*2022-698hTransformer32MCE40.6--
Ma et al. [40]*2022-818hConformer79MCTC+CE34.7--
Semi-Supervised
Shillingford et al. [62]*2019-3886hRNN-CTC55.1--
Makino et al. [42]2019-31khRNN43MTransducer33.64.84.5
Afouras et al. [2]*2020344h433h1Jasper (CNN)-CTC59.8--
Serdyuk et al. [57]2021-90khTransformer-Transducer25.9-2.3
Ma et al. [40]*2022641h818hConformer79MCTC+CE31.5--
Serdyuk et al. [58]2022-90khConformer-Transducer17.01.61.6
Ma et al. [41]*20232630h818hConformer186MCTC+CE19.11.00.9
Chang et al. [15]2023-100khConformer98MTransducer12.8-0.9
Self-supervised (Base Models)
AV-HuBERT [59, 60]2022433h433hTransformer103MCE44.03.02.82
AV2vec [76]‡2023433h433hTransformer103MCE39.9-2.6
RAVEn [27]2023433h433hTransformer97MCTC+CE39.12.2-
ES3 (ours)2023433h433hTransformer102MLF-MMI40.3 (39.2)2.9 (2.4)2.5 (2.0)
AV-data2vec [34]2023433h433hTransformer103MCE39.02.01.8
Self-supervised (Large Models)
AV-HuBERT [59, 60]2022433h433hTransformer325MCE41.62.72.52
AV-data2vec [34]2023433h433hTransformer325MCE37.41.91.7
ES3 (ours)2023433h433hTransformer317MLF-MMI37.1 (36.7)2.8 (2.1)2.1 (1.7)
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.088, + 0.372, + 0.424, + 0.385 + ], + "angle": 0, + "content": "1 Uses an additional ASR model trained on LibriSpeech (960h)." + }, + { + "type": "table_footnote", + "bbox": [ + 0.457, + 0.372, + 0.632, + 0.385 + ], + "angle": 0, + "content": "2 Reproduced by Lian et al. [34]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.389, + 0.47, + 0.51 + ], + "angle": 0, + "content": "Visualization of our learned psuedo-label sequences. We randomly select a sample from the LRS3-TED validation set and visualize the pseudo-label sequences, i.e. code indices \\( q \\) generated by the audio, visual and audio-visual EMA teacher \\( \\tilde{\\theta} \\) in each stage. Ground truth phones for each audio frame (10ms) are derived by forced alignment [79]. Fig. 3 shows that our learned codebook yields high-quality pseudo-labels that strongly correlate with phonemic units." + }, + { + "type": "image", + "bbox": [ + 0.08, + 0.516, + 0.463, + 0.634 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.634, + 0.47, + 0.718 + ], + "angle": 0, + "content": "Figure 3. Pseudo-label sequences from each stage. Example of pseudo-labels generated from the last Transformer layer for the utterance \"so what does that mean\" by the audio, visual, and audio-visual teachers in each stage (BASE model). At the bottom we show the spectrogram for reference. Please note that our pseudolabels are derived at 40ms intervals, limited by video frame rate." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.72, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Training efficiency comparison with SoTA methods. Learning audio-visual speech representations has been extremely compute-intensive. Here we briefly compare \\(\\mathrm{ES}^3\\) with prominent SoTA methods. Pre-training a BASE/LARGE model on LRS3-TED using our \\(\\mathrm{ES}^3\\) for 100 epochs per stage only takes about 6/8.3 days in total with just 8 NVIDIA A40 GPUs. In contrast, AV-data2vec is trained for 1000k steps using 64 NVIDIA V100 GPUs for 4-5/6-7 days using BASE/LARGE with a batch size of 1280/2560s, or equivalently, 821/1642 epochs. They also experiment extensively with hyperparameters such as modality schedulers, which further increases the cost of" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.389, + 0.895, + 0.57 + ], + "angle": 0, + "content": "training. RAVEn performs an equivalent of 6.25/50 days of training on 8 A100s with BASE*/LARGE. It uses separate encoders for audio and video, so the resulting model cannot perform audio-visual fusion out-of-the-box. This also requires keeping track of four sets of parameters, which is very memory-intensive. By focusing on designing a clever learning strategy, we circumvent such manual tuning altogether and greatly reduce training costs. Finally, we use the simple LF-MMI criterion, and expect better performance from tuning beam size, including a larger external language model (see supplementary), or switching to the common heavy Attention-based Encoder-Decoder model." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.576, + 0.628, + 0.592 + ], + "angle": 0, + "content": "5. Conclusions" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.596, + 0.892, + 0.748 + ], + "angle": 0, + "content": "We present a novel strategy, \\( \\mathrm{ES}^3 \\), for self-supervised learning of audio-visual speech representations. Unlike existing methods that primarily focus on learning information shared between audio and video, we take an \"evolving\" approach that also captures unique and synergistic information. Our experimental results demonstrate that \\( \\mathrm{ES}^3 \\) indeed progressively learns unique, shared and synergistic information. Moreover, in terms of downstream performance, it rivals SoTA methods in terms of ASR, VSR and AVSR accuracy without burdensome manual balancing efforts." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.748, + 0.893, + 0.902 + ], + "angle": 0, + "content": "Acknowledgments. This work is partially supported by the National Natural Science Foundation of China (No. 62276247, 62076250). We are grateful to the ICT computing platform, the Nanjing Institute of InforSuperBahn, and Capitalonline Data Service Co., Ltd for providing additional compute resources. We also thank Chenhao Wang, Mingmin Yang and Jingyun Xiao for helping to prepare CAS-VSR-S101, Bingquan Xia and Songtao Luo for help with Conformer and seq2seq baselines, Bingquan Xia and Feixiang Wang for help during the rebuttal phase, and Feixiang Wang and Zhaoxin Yuan for an early version of Fig. 1." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "27076" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.121, + 0.174, + 0.137 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.146, + 0.47, + 0.187 + ], + "angle": 0, + "content": "[1] Triantafyllos Afouras, Joon Son Chung, and Andrew Zisserman. LRS3-TED: a large-scale dataset for visual speech recognition. CoRR, abs/1809.00496, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.189, + 0.472, + 0.245 + ], + "angle": 0, + "content": "[2] Triantafyllos Afouras, Joon Son Chung, and Andrew Zisserman. ASR is all you need: Cross-modal distillation for lip reading. In ICASSP, pages 2143-2147. IEEE, 2020. 2, 7, 8, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.247, + 0.47, + 0.302 + ], + "angle": 0, + "content": "[3] Triantafyllos Afouras, Joon Son Chung, Andrew W. Senior, Oriol Vinyals, and Andrew Zisserman. Deep audio-visual speech recognition. IEEE Trans. Pattern Anal. Mach. Intell., 44(12):8717-8727, 2022. 1, 5, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.304, + 0.47, + 0.371 + ], + "angle": 0, + "content": "[4] Jean-Baptiste Alayrac, Adrià Recasens, Rosalia Schneider, Relja Arandjelovic, Jason Ramapuram, Jeffrey De Fauw, Lucas Smaira, Sander Dieleman, and Andrew Zisserman. Self-supervised multimodal versatile networks. In NeurlPS, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.374, + 0.47, + 0.428 + ], + "angle": 0, + "content": "[5] Humam Alwassel, Dhruv Mahajan, Bruno Korbar, Lorenzo Torresani, Bernard Ghanem, and Du Tran. Self-supervised learning by cross-modal audio-video clustering. In NeurIPS, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.431, + 0.47, + 0.471 + ], + "angle": 0, + "content": "[6] Relja Arandjelovic and Andrew Zisserman. Look, listen and learn. In ICCV, pages 609-617. IEEE Computer Society, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.474, + 0.47, + 0.502 + ], + "angle": 0, + "content": "[7] Relja Arandjelovic and Andrew Zisserman. Objects that sound. In ECCV(1), pages 451-466. Springer, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.504, + 0.47, + 0.557 + ], + "angle": 0, + "content": "[8] Yuki Markus Asano, Mandela Patrick, Christian Rupprecht, and Andrea Vedaldi. Labelling unlabelled videos from scratch with multi-modal self-supervision. In NeurIPS, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.56, + 0.47, + 0.602 + ], + "angle": 0, + "content": "[9] Alexei Baevski, Michael Auli, and Abdelrahman Mohamed. Effectiveness of self-supervised pre-training for speech recognition. CoRR, abs/1911.03912, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.604, + 0.47, + 0.644 + ], + "angle": 0, + "content": "[10] Alexei Baevski, Yuhao Zhou, Abdelrahman Mohamed, and Michael Auli. wav2vec 2.0: A framework for self-supervised learning of speech representations. In NeurIPS, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.646, + 0.47, + 0.701 + ], + "angle": 0, + "content": "[11] Alexei Baevski, Wei-Ning Hsu, Qiantong Xu, Arun Babu, Jiatao Gu, and Michael Auli. data2vec: A general framework for self-supervised learning in speech, vision and language. In ICML, pages 1298-1312. PMLR, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.703, + 0.47, + 0.758 + ], + "angle": 0, + "content": "[12] Alexei Baevski, Arun Babu, Wei-Ning Hsu, and Michael Auli. Efficient self-supervised learning with contextualized target representations for vision, speech and language. In ICML, pages 1416-1429. PMLR, 2023. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.76, + 0.47, + 0.815 + ], + "angle": 0, + "content": "[13] Adrian Bulat and Georgios Tzimiropoulos. How far are we from solving the 2D & 3D face alignment problem? (and a dataset of 230,000 3D facial landmarks). In ICCV, pages 1021-1030. IEEE Computer Society, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.817, + 0.47, + 0.871 + ], + "angle": 0, + "content": "[14] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision Transformers. In ICCV, pages 9630-9640. IEEE, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.47, + 0.901 + ], + "angle": 0, + "content": "[15] Oscar Chang, Hank Liao, Dmitriy Serdyuk, Ankit Shahy, and Olivier Siohan. Conformer is all you need for visual" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.146, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.892, + 0.119 + ], + "angle": 0, + "content": "speech recognition. In ICASSP, pages 10136-10140. IEEE, 2024. 1, 8, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.121, + 0.892, + 0.216 + ], + "angle": 0, + "content": "[16] Brian Chen, Andrew Rouditchenko, Kevin Duarte, Hilde Kuehne, Samuel Thomas, Angie W. Boggust, Rameswar Panda, Brian Kingsbury, Rogério Feris, David Harwath, James R. Glass, Michael Picheny, and Shih-Fu Chang. Multimodal clustering networks for self-supervised learning from unlabeled videos. In ICCV, pages 7992-8001. IEEE, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.218, + 0.892, + 0.286 + ], + "angle": 0, + "content": "[17] Xize Cheng, Tao Jin, Linjun Li, Wang Lin, Xinyu Duan, and Zhou Zhao. OpenSR: Open-modality speech recognition via maintaining multi-modality alignment. In ACL (1), pages 6592-6607. Association for Computational Linguistics, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.288, + 0.892, + 0.328 + ], + "angle": 0, + "content": "[18] Joon Son Chung, Arsha Nagrani, and Andrew Zisserman. VoxCeleb2: Deep speaker recognition. In INTERSPEECH, pages 1086-1090. ISCA, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.33, + 0.892, + 0.383 + ], + "angle": 0, + "content": "[19] Soo-Whan Chung, Joon Son Chung, and Hong-Goo Kang. Perfect match: Improved cross-modal embeddings for audiovisual synchronisation. In ICASSP, pages 3965–3969. IEEE, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.385, + 0.892, + 0.425 + ], + "angle": 0, + "content": "[20] Tri Dao. FlashAttention-2: Faster attention with better parallelism and work partitioning. CoRR, abs/2307.08691, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.427, + 0.892, + 0.468 + ], + "angle": 0, + "content": "[21] Tri Dao, Daniel Y. Fu, Stefano Ermon, Atri Rudra, and Christopher Ré. FlashAttention: Fast and memory-efficient exact attention with IO-awareness. In NeurIPS, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.47, + 0.892, + 0.537 + ], + "angle": 0, + "content": "[22] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. BERT: Pre-training of deep bidirectional transformers for language understanding. In NAACL-HLT (1), pages 4171–4186. Association for Computational Linguistics, 2019. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.539, + 0.892, + 0.58 + ], + "angle": 0, + "content": "[23] Angela Fan, Edouard Grave, and Armand Joulin. Reducing Transformer depth on demand with structured dropout. In ICLR. OpenReview.net, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.581, + 0.892, + 0.622 + ], + "angle": 0, + "content": "[24] Dalu Feng, Shuang Yang, and Shiguang Shan. An efficient software for building lip reading models without pains. In ICME Workshops, pages 1-2. IEEE, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.623, + 0.892, + 0.677 + ], + "angle": 0, + "content": "[25] Yuan Gong, Andrew Rouditchenko, Alexander H. Liu, David Harwath, Leonid Karlinsky, Hilde Kuehne, and James R. Glass. Contrastive audio-visual masked autoencoder. In ICLR. OpenReview.net, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.679, + 0.892, + 0.719 + ], + "angle": 0, + "content": "[26] Hossein Hadian, Hossein Sameti, Daniel Povey, and Sanjeev Khudanpur. End-to-end speech recognition using lattice-free MMI. In INTERSPEECH, pages 12-16. ISCA, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.721, + 0.892, + 0.775 + ], + "angle": 0, + "content": "[27] Alexandros Haliassos, Pingchuan Ma, Rodrigo Mira, Stavros Petridis, and Maja Pantic. Jointly learning visual and auditory speech representations from raw data. In ICLR. Open-Review.net, 2023. 2, 6, 7, 8, 1, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.777, + 0.892, + 0.83 + ], + "angle": 0, + "content": "[28] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, and Ross B. Girshick. Masked autoencoders are scalable vision learners. In CVPR, pages 15979-15988. IEEE, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.832, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[29] François Hernandez, Vincent Nguyen, Sahar Ghannay, Natalia A. Tomashenko, and Yannick Esteve. TED-LIUM 3: Twice as much data and corpus repartition for experiments on speaker adaptation. In SPECOM, pages 198-208. Springer, 2018. 2" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "27077" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.134 + ], + "angle": 0, + "content": "[30] Wei-Ning Hsu and Bowen Shi. u-HuBERT: Unified mixed-modal speech pretraining and zero-shot transfer to unlabeled modality. In NeurIPS, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.134, + 0.471, + 0.205 + ], + "angle": 0, + "content": "[31] Wei-Ning Hsu, Benjamin Bolte, Yao-Hung Hubert Tsai, Kushal Lakhotia, Ruslan Salakhutdinov, and Abdelrahman Mohamed. HuBERT: Self-supervised speech representation learning by masked prediction of hidden units. IEEE ACM Trans. Audio Speech Lang. Process., 29:3451-3460, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.205, + 0.471, + 0.247 + ], + "angle": 0, + "content": "[32] Bruno Korbar, Du Tran, and Lorenzo Torresani. Cooperative learning of audio and video models from self-supervised synchronization. In NeurIPS, pages 7774-7785, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.247, + 0.471, + 0.288 + ], + "angle": 0, + "content": "[33] Aleksandr Laptev, Somshubra Majumdar, and Boris Ginsburg. CTC variations through new WFST topologies. In INTERSPEECH, pages 1041-1045. ISCA, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.288, + 0.471, + 0.343 + ], + "angle": 0, + "content": "[34] Jiachen Lian, Alexei Baevski, Wei-Ning Hsu, and Michael Auli. AV-data2vec: Self-supervised learning of audio-visual speech representations with contextualized target representations. In ASRU. IEEE, 2023. 2, 6, 7, 8, 1, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.343, + 0.471, + 0.399 + ], + "angle": 0, + "content": "[35] Paul Pu Liang, Zihao Deng, Martin Q. Ma, James Y. Zou, Louis-Philippe Morency, and Ruslan Salakhutdinov. Factorized contrastive learning: Going beyond multi-view redundancy. In NeurIPS, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.4, + 0.471, + 0.455 + ], + "angle": 0, + "content": "[36] Alexander H. Liu, Heng-Jui Chang, Michael Auli, Wei-Ning Hsu, and Jim Glass. DinoSR: Self-distillation and online clustering for self-supervised speech representation learning. In NeurIPS, 2023. 4, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.455, + 0.471, + 0.484 + ], + "angle": 0, + "content": "[37] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In *ICLR (Poster)*. OpenReview.net, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.484, + 0.471, + 0.539 + ], + "angle": 0, + "content": "[38] Pingchuan Ma, Rodrigo Mira, Stavros Petridis, Bjorn W. Schuller, and Maja Pantic. LiRA: Learning visual speech representations from audio through self-supervision. In *Interspeech*, pages 3011–3015. ISCA, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.539, + 0.471, + 0.581 + ], + "angle": 0, + "content": "[39] Pingchuan Ma, Stavros Petridis, and Maja Pantic. End-to-end audio-visual speech recognition with Conformers. In ICASSP, pages 7613-7617. IEEE, 2021. 1, 6, 7, 8, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.581, + 0.471, + 0.623 + ], + "angle": 0, + "content": "[40] Pingchuan Ma, Stavros Petridis, and Maja Pantic. Visual speech recognition for multiple languages in the wild. Nat. Mac. Intell., 4(11):930-939, 2022. 1, 6, 8, 2, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.623, + 0.471, + 0.692 + ], + "angle": 0, + "content": "[41] Pingchuan Ma, Alexandros Haliassos, Adriana Fernandez-Lopez, Honglie Chen, Stavros Petridis, and Maja Pantic. Auto-AVSR: Audio-Visual speech recognition with automatic labels. In ICASSP, pages 1-5. IEEE, 2023. 1, 3, 7, 8, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.692, + 0.471, + 0.761 + ], + "angle": 0, + "content": "[42] Takaki Makino, Hank Liao, Yannis M. Assael, Brendan Shillingford, Basilio Garcia, Otavio Braga, and Olivier Siohan. Recurrent neural network transducer for audio-visual speech recognition. In ASRU, pages 905–912. IEEE, 2019. 1, 8, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.761, + 0.471, + 0.832 + ], + "angle": 0, + "content": "[43] Antoine Miech, Dimitri Zhukov, Jean-Baptiste Alayrac, Makarand Tapaswi, Ivan Laptev, and Josef Sivic. HowTo100M: Learning a text-video embedding by watching hundred million narrated video clips. In ICCV, pages 2630-2640. IEEE, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.832, + 0.471, + 0.9 + ], + "angle": 0, + "content": "[44] Arsha Nagrani, Chen Sun, David Ross, Rahul Sukthankar, Cordelia Schmid, and Andrew Zisserman. Speech2Action: Cross-modal supervision for action recognition. In CVPR, pages 10314-10323. Computer Vision Foundation / IEEE, 2020. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.471, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.092, + 0.893, + 0.163 + ], + "angle": 0, + "content": "[45] Xichen Pan, Peiyu Chen, Yichen Gong, Helong Zhou, Xinbing Wang, and Zhouhan Lin. Leveraging unimodal self-supervised learning for multimodal audio-visual speech recognition. In ACL (1), pages 4491–4503. Association for Computational Linguistics, 2022. 7, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.163, + 0.893, + 0.233 + ], + "angle": 0, + "content": "[46] Daniel S. Park, William Chan, Yu Zhang, Chung-Cheng Chiu, Barret Zoph, Ekin D. Cubuk, and Quoc V. Le. SpecAugment: A simple data augmentation method for automatic speech recognition. In INTERSPEECH, pages 2613-2617. ISCA, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.234, + 0.893, + 0.277 + ], + "angle": 0, + "content": "[47] Sooyoung Park, Arda Senocak, and Joon Son Chung. MarginNCE: Robust sound localization with a negative margin. In ICASSP, pages 1-5. IEEE, 2023. 5, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.278, + 0.893, + 0.389 + ], + "angle": 0, + "content": "[48] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, Alban Desmaison, Andreas Köpf, Edward Z. Yang, Zachary DeVito, Martin Raison, Alykhan Tejani, Sasank Chilamkurthy, Benoit Steiner, Lu Fang, Junjie Bai, and Soumith Chintala. PyTorch: An imperative style, high-performance deep learning library. In NeurIPS, pages 8024-8035, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.39, + 0.893, + 0.459 + ], + "angle": 0, + "content": "[49] Mandela Patrick, Yuki Markus Asano, Polina Kuznetsova, Ruth Fong, João F. Henriques, Geoffrey Zweig, and Andrea Vedaldi. On compositions of transformations in contrastive self-supervised learning. In ICCV, pages 9557-9567. IEEE, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.46, + 0.893, + 0.516 + ], + "angle": 0, + "content": "[50] Trung Pham, Chaoning Zhang, Axi Niu, Kang Zhang, and Chang D. Yoo. On the pros and cons of momentum encoder in self-supervised visual representation learning. CoRR, abs/2208.05744, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.518, + 0.893, + 0.588 + ], + "angle": 0, + "content": "[51] Daniel Povey, Vijayaditya Peddinti, Daniel Galvez, Pegah Ghahremani, Vimal Manohar, Xingyu Na, Yiming Wang, and Sanjeev Khudanpur. Purely sequence-trained neural networks for ASR based on lattice-free MMI. In INTER-SPEECH, pages 2751-2755. ISCA, 2016. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.589, + 0.893, + 0.631 + ], + "angle": 0, + "content": "[52] K. R. Prajwal, Triantafyllos Afouras, and Andrew Zisserman. Sub-word level lip reading with visual attention. In CVPR, pages 5152-5162. IEEE, 2022. 8, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.632, + 0.893, + 0.716 + ], + "angle": 0, + "content": "[53] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In ICML, pages 8748-8763. PMLR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.717, + 0.893, + 0.773 + ], + "angle": 0, + "content": "[54] Alec Radford, Jong Wook Kim, Tao Xu, Greg Brockman, Christine McLeavey, and Ilya Sutskever. Robust speech recognition via large-scale weak supervision. In ICML, pages 28492-28518. PMLR, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.774, + 0.893, + 0.845 + ], + "angle": 0, + "content": "[55] Jordi Robert-Ribes, Jean-Luc Schwartz, Tahar Lallouache, and Pierre Escudier. Complementarity and synergy in b-modal speech: Auditory, visual, and audio-visual identification of french oral vowels in noise. The Journal of the Acoustical Society of America, 103(6):3677-3689, 1998. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.845, + 0.893, + 0.9 + ], + "angle": 0, + "content": "[56] Steffen Schneider, Alexei Baevski, Ronan Collobert, and Michael Auli. wav2vec: Unsupervised pre-training for speech recognition. In INTERSPEECH, pages 3465-3469. ISCA, 2019. 3" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.092, + 0.893, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "27078" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.093, + 0.468, + 0.134 + ], + "angle": 0, + "content": "[57] Dmitriy Serdyuk, Otavio Braga, and Olivier Siohan. Audiovisual speech recognition is worth \\(32 \\times 32 \\times 8\\) voxels. In ASRU, pages 796-802. IEEE, 2021. 2, 8, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.134, + 0.468, + 0.19 + ], + "angle": 0, + "content": "[58] Dmitriy Serdyuk, Otavio Braga, and Olivier Siohan. Transformer-based video front-ends for audio-visual speech recognition for single and multi-person video. In INTER-SPEECH, pages 2833–2837. ISCA, 2022. 2, 8, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.19, + 0.468, + 0.245 + ], + "angle": 0, + "content": "[59] Bowen Shi, Wei-Ning Hsu, Kushal Lakhotia, and Abdelrahman Mohamed. Learning audio-visual speech representation by masked multimodal cluster prediction. In ICLR. OpenReview.net, 2022. 1, 2, 3, 6, 7, 8, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.246, + 0.468, + 0.287 + ], + "angle": 0, + "content": "[60] Bowen Shi, Wei-Ning Hsu, and Abdelrahman Mohamed. Robust self-supervised audio-visual speech recognition. In INTERSPEECH, pages 2118–2122. ISCA, 2022. 7, 8, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.288, + 0.468, + 0.342 + ], + "angle": 0, + "content": "[61] Bowen Shi, Abdelrahman Mohamed, and Wei-Ning Hsu. Learning lip-based audio-visual speaker embeddings with AV-HuBERT. In INTERSPEECH, pages 4785–4789. ISCA, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.343, + 0.468, + 0.439 + ], + "angle": 0, + "content": "[62] Brendan Shillingford, Yannis M. Assael, Matthew W. Hoffman, Thomas Paine, Cian Hughes, Utsav Prabhu, Hank Liao, Hasim Sak, Kanishka Rao, Lorrayne Bennett, Marie Mulville, Misha Denil, Ben Coppin, Ben Laurie, Andrew W. Senior, and Nando de Freitas. Large-scale visual speech recognition. In INTERSPEECH, pages 4135-4139. ISCA, 2019. 8, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.441, + 0.468, + 0.51 + ], + "angle": 0, + "content": "[63] Nina Shvetsova, Brian Chen, Andrew Rouditchenko, Samuel Thomas, Brian Kingsbury, Rogério Feris, David Harwath, James R. Glass, and Hilde Kuehne. Everything at once - multi-modal fusion Transformer for video retrieval. In CVPR, pages 19988-19997. IEEE, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.511, + 0.468, + 0.552 + ], + "angle": 0, + "content": "[64] Themos Stafylakis and Georgios Tzimiropoulos. Combining residual networks with LSTMs for lipreading. In INTER-SPEECH, pages 3652-3656. ISCA, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.553, + 0.468, + 0.594 + ], + "angle": 0, + "content": "[65] Jianlin Su, Yu Lu, Shengfeng Pan, Bo Wen, and Yunfeng Liu. RoFormer: Enhanced transformer with rotary position embedding. CoRR, abs/2104.09864, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.595, + 0.468, + 0.649 + ], + "angle": 0, + "content": "[66] Weijie Su, Xizhou Zhu, Yue Cao, Bin Li, Lewei Lu, Furu Wei, and Jifeng Dai. VL-BERT: Pre-training of generic visual-linguistic representations. In ICLR. OpenReview.net, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.65, + 0.468, + 0.704 + ], + "angle": 0, + "content": "[67] Chen Sun, Austin Myers, Carl Vondrick, Kevin Murphy, and Cordelia Schmid. VideoBERT: A joint model for video and language representation learning. In ICCV, pages 7463-7472. IEEE, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.706, + 0.468, + 0.748 + ], + "angle": 0, + "content": "[68] Yonglong Tian, Dilip Krishnan, and Phillip Isola. Contrastive multiview coding. In ECCV (11), pages 776-794. Springer, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.749, + 0.468, + 0.789 + ], + "angle": 0, + "content": "[69] Hugo Touvron, Matthieu Cord, Alexandre Sablayrolles, Gabriel Synnaeve, and Hervé Jégou. Going deeper with image Transformers. In ICCV, pages 32-42. IEEE, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.79, + 0.468, + 0.844 + ], + "angle": 0, + "content": "[70] Kristin J Van Engen, Avanti Dey, Mitchell S Sommers, and Jonathan E Peelle. Audiovisual speech perception: Moving beyond mcgurk. The Journal of the Acoustical Society of America, 152(6):3216-3225, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.845, + 0.468, + 0.9 + ], + "angle": 0, + "content": "[71] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In NIPS, pages 5998-6008, 2017. 3, 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.093, + 0.468, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.134 + ], + "angle": 0, + "content": "[72] Feng Wang and Huaping Liu. Understanding the behaviour of contrastive loss. In CVPR, pages 2495-2504. Computer Vision Foundation / IEEE, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.136, + 0.892, + 0.176 + ], + "angle": 0, + "content": "[73] Paul L. Williams and Randall D. Beer. Nonnegative decomposition of multivariate information. CoRR, abs/1004.2515, 2010. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.179, + 0.892, + 0.22 + ], + "angle": 0, + "content": "[74] Bo Xu, Cheng Lu, Yandong Guo, and Jacob Wang. Discriminative multi-modality speech recognition. In CVPR, pages 14421-14430. Computer Vision Foundation / IEEE, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.221, + 0.892, + 0.29 + ], + "angle": 0, + "content": "[75] Jianwei Yu, Shi-Xiong Zhang, Jian Wu, Shahram Ghorbani, Bo Wu, Shiyin Kang, Shansong Liu, Xunying Liu, Helen Meng, and Dong Yu. Audio-visual recognition of overlapped speech for the LRS2 dataset. In ICASSP, pages 6984–6988. IEEE, 2020. 7, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.291, + 0.892, + 0.359 + ], + "angle": 0, + "content": "[76] Jing-Xuan Zhang, Genshun Wan, Zhen-Hua Ling, Jia Pan, Jianqing Gao, and Cong Liu. Self-supervised audiovisual speech representations learning by multimodal self-distillation. In ICASSP, pages 1-5. IEEE, 2023. 2, 7, 8, 1, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.361, + 0.892, + 0.432 + ], + "angle": 0, + "content": "[77] Xiaohui Zhang, Vimal Manohar, David Zhang, Frank Zhang, Yangyang Shi, Nayan Singhal, Julian Chan, Fuchun Peng, Yatharth Saraf, and Mike Seltzer. On lattice-free boosted MMI training of HMM and CTC-based full-context ASR models. In ASRU, pages 1026-1033. IEEE, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.433, + 0.892, + 0.487 + ], + "angle": 0, + "content": "[78] Ya Zhao, Rui Xu, Xinchao Wang, Peng Hou, Haihong Tang, and Mingli Song. Hearing lips: Improving lip reading by distilling speech recognizers. In AAAI, pages 6917-6924. AAAI Press, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.489, + 0.892, + 0.53 + ], + "angle": 0, + "content": "[79] Jian Zhu, Cong Zhang, and David Jurgens. Phone-to-audio alignment without text: A semi-supervised approach. In ICASSP, pages 8167-8171. IEEE, 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.531, + 0.892, + 0.601 + ], + "angle": 0, + "content": "[80] Qiushi Zhu, Long Zhou, Ziqiang Zhang, Shujie Liu, Binxing Jiao, Jie Zhang, Lirong Dai, Daxin Jiang, Jinyu Li, and Furu Wei. VATLM: Visual-Audio-Text pre-training with unified masked prediction for speech representation learning. IEEE Transactions on Multimedia, pages 1–11, 2023. 1, 6, 7, 4, 5" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.601 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "27079" + } + ] +] \ No newline at end of file diff --git a/2024/ES3_ Evolving Self-Supervised Learning of Robust Audio-Visual Speech Representations/09b7366b-807f-4dfb-a9c8-45edf9472928_origin.pdf b/2024/ES3_ Evolving Self-Supervised Learning of Robust Audio-Visual Speech Representations/09b7366b-807f-4dfb-a9c8-45edf9472928_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..e5d5376a34638eec35d5b3dcb5a5b0402fb21ffa --- /dev/null +++ b/2024/ES3_ Evolving Self-Supervised Learning of Robust Audio-Visual Speech Representations/09b7366b-807f-4dfb-a9c8-45edf9472928_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:57a06e1772a41cb703a0befd9ea9fbc6f25df6d73b2ee9eb34149a3ebd39bcc1 +size 693886 diff --git a/2024/ES3_ Evolving Self-Supervised Learning of Robust Audio-Visual Speech Representations/full.md b/2024/ES3_ Evolving Self-Supervised Learning of Robust Audio-Visual Speech Representations/full.md new file mode 100644 index 0000000000000000000000000000000000000000..86da80ba84409d395b3f97e5b8394e0da5ee178b --- /dev/null +++ b/2024/ES3_ Evolving Self-Supervised Learning of Robust Audio-Visual Speech Representations/full.md @@ -0,0 +1,360 @@ +# $\mathbf{ES}^{3}$ : Evolving Self-Supervised Learning of Robust Audio-Visual Speech Representations + +Yuanhang Zhang $^{1,2}$ , Shuang Yang $^{1,2}$ , Shiguang Shan $^{1,2}$ , Xilin Chen $^{1,2}$ + +1Key Laboratory of Intelligent Information Processing of Chinese Academy of Sciences (CAS), Institute of Computing Technology, CAS, Beijing 100190, China + +$^{2}$ University of Chinese Academy of Sciences, Beijing 100049, China + +zhangyuanhang15@mails.ucas.ac.cn, {shuang.yang, sgshan, xlchen}@ict.ac.cn + +# Abstract + +We propose a novel strategy, $ES^3$ , for self-supervised learning of robust audio-visual speech representations from unlabeled talking face videos. While many recent approaches for this task primarily rely on guiding the learning process using the audio modality alone to capture information shared between audio and video, we reframe the problem as the acquisition of shared, unique (modality-specific) and synergistic speech information to address the inherent asymmetry between the modalities. Based on this formulation, we propose a novel "evolving" strategy that progressively builds joint audio-visual speech representations that are strong for both uni-modal (audio & visual) and bimodal (audio-visual) speech. First, we leverage the more easily learnable audio modality to initialize audio and visual representations by capturing audio-unique and shared speech information. Next, we incorporate video-unique speech information and bootstrap the audio-visual representations on top of the previously acquired shared knowledge. Finally, we maximize the total audio-visual speech information, including synergistic information to obtain robust and comprehensive representations. We implement $ES^3$ as a simple Siamese framework and experiments on both English benchmarks and a newly contributed large-scale Mandarin dataset show its effectiveness. In particular, on LRS2-BBC, our smallest model is on par with SoTA models with only 1/2 parameters and 1/8 unlabeled data (223h). + +# 1. Introduction + +The task of learning audio-visual speech representations has gained increasing importance due to its potential to support a wide range of relevant applications. Building on the success in other domains [9, 11, 12, 14, 22, 31], several effective self-supervised approaches have recently emerged for this task by pre-training models on paired or mixed audio, visual and text data with pretext tasks like speech unit + +prediction [30, 59, 80]. However, learning representations in the audio-visual speech domain faces unique challenges brought by the inherent asymmetry between the two modalities involved. First, audio and video often convey distinct and complementary types of speech information. For example, $\mathsf{m} /$ and $\mathsf{n} /$ sound similar, but are visually different; on the contrary, $\mathsf{p} /$ and $\mathsf{b} /$ are visually similar but can be distinguished with audio [70]. Even more interesting is the presence of synergistic information – information that can further enhance speech perception [55] but only emerges when audio and video are considered together, and cannot be inferred when either modality is absent. Second, they manifest different amounts of noise and information. Specifically, in most existing audio-visual speech datasets, audio is characterized by high sampling rate and relatively clean speech samples, making audio representations readily learnable from these datasets. In contrast, the visual modality often presents several challenging factors, including missing (e.g. due to self-occlusion in profile faces or low image quality), redundant (e.g. due to appearance variations, makeup and beards etc.) and ambiguous (e.g. due to the lower sampling rate and the presence of homophones) visual information. Their presence complicates learning meaningful and robust representations solely from video. The difficulties encountered in the continuous efforts to improve the performance of visual speech recognition (VSR) [15, 24, 40, 41] models illustrate this persistent obstacle. Additionally, these differences between the two modalities lead to an imbalance between their learning dynamics, requiring additional manual settings [3, 39, 42, 61]. + +In this work, we break down the problem of learning audio-visual speech representations into the learning of three different types of information: shared (across modalities), unique (modality-specific) and synergistic speech information. Based on our formulation, we propose a novel strategy called $\mathrm{ES^3}$ (short for Evolving Self-Supervised Learning of Robust Audio-Visual Speech Representations) + +to tackle the asymmetry between the two modalities for learning robust audio-visual speech representations. Specifically, our $\mathrm{ES}^3$ employs a three-stage progressive learning approach. Initially, it focuses on acquiring the easily learnable audio-unique and shared speech information. Then, it addresses the more challenging visual-unique speech information while incorporating it into the initial audio-visual joint representation. Lastly, it optimizes the final audio-visual representations by leveraging synergy, on top of the acquired shared and modality-unique information. + +Our contributions can be summarized as follows: + +- We propose a new formulation for audio-visual speech representation learning by decomposing the task into the acquisition of shared (across modalities), unique (modality-specific) and synergistic speech information. +- We propose a novel strategy, $\mathrm{ES}^3$ to progressively learn the three types of information in a self-supervised manner with a simple Siamese framework. This evolving strategy significantly alleviates the previous burden of tuning balancing parameters. +- Extensive experiments on two widely used benchmarks show that $\mathrm{ES}^3$ rivals previous methods in auditory, visual and audio-visual speech recognition performance. +- We also contribute a new large-scale Mandarin audiovisual speech dataset named CAS-VSR-S101 to the community for evaluation of similar methods. + +# 2. Related Work + +Self-supervised learning from multi-modal data. The arguably most successful paradigm for self-supervised multimodal learning in the past few years is contrastive learning, which exploits the common information encoded in different modalities [68], such as audio (A) [6, 7], image (I) [53, 66], video (V) [5, 8, 32, 67] and text (T) [4, 16, 43, 44, 49, 63]. These approaches focus on learning semantically aligned multi-modal representations that are effective for specific downstream tasks like cross-modal retrieval, but discard modality-unique information that is useful for other uni-modal tasks. To solve this problem, Gong et al. [25] combine contrastive learning with the masked prediction objective [28] to learn strong audio-visual representations suitable for both retrieval and joint classification. Liang et al. [35] propose a new algorithm, FACTORCL to learn self-supervised multi-modal representations that capture both shared and unique information and verify its effectiveness on multiple tasks. In this paper, we explicitly address capturing unique, shared and synergistic information for uni- and multi-modal downstream tasks. + +Learning audio-visual speech representations. A main challenge of learning audio-visual speech representation is properly handling the aforementioned asymmetry between audio and video. An established approach is to leverage the easily learnable audio modality as a teacher to guide + +visual or audio-visual speech representation learning. Implementing additional measures to mitigate audio dominance and promote visual features during learning is vital for this paradigm. For instance, Shi et al. [59] and Zhang et al. [76] use different masking strengths for the two modalities, apply modality dropout, and enforce simple audio features by only using a weak linear projection layer. Others have explored freezing pre-trained audio representations [2, 38, 78]. However, such an audio-centric approach is sub-optimal and inefficient. First, it limits the learned representations to capturing only shared, i.e. modality-invariant information, or at most, shared and audio-unique information. Recently, a few attempts have been made to go beyond merely using audio as the anchor modality. Haliassos et al. [27] combine multiple audio-visual prediction tasks in place of purely audio-centric ones, and treats A-V asymmetry by having the visual stream predict only auditory but not visual targets. Lian et al. [34] perform a series of trials to determine an optimal schedule for audio and video input combination. Second, the process of exhaustively testing and identifying the optimal set of balancing parameters such as masking ratios, modality schedules and modality dropout rate is extremely time-consuming and computationally demanding. In contrast, our newly proposed evolving strategy identifies different types of task-relevant information and learns them progressively, while avoiding pitfalls and computational burdens (see supp). + +# 3. Our Proposed ES3 + +# 3.1. Preliminaries + +The goal of $\mathrm{ES}^3$ is to learn audio-visual speech representations from unlabeled talking face videos. The inputs to our $\mathrm{ES}^3$ are paired audio-visual data $X_{a,v} = \{(X_{a,i},X_{v,i})\mid i = 1,2,\ldots ,N\}$ . Each of the $N$ samples consists of a sequence of audio frames $X_{a,i} = \operatorname {LogFBank}(\mathbf{a}_i)\in \mathbb{R}^{F\times T_{a,i}}$ and a temporally synchronized video sequence $X_{v,i} = \mathbf{v}_i\in \mathbb{R}^{C\times H\times W\times T_{v,i}}$ showing the lips of a speaker, where $F$ is the filter bank dimension, $C$ is the number of video channels, $T_{a,i}$ and $T_{v,i}$ are the total number of audio and video frames in the $i$ -th sample, and $H$ and $W$ are the width and height of each frame. Audio and video are sampled at 100 and 25 FPS, i.e. for a 1-second clip, $T_{a} = 4T_{v} = 100$ . The goal is to learn a $d$ -dimensional contextualized representation $\gamma \in \mathbb{R}^{d\times T_v}$ capturing vital speech and context information within the two modality spaces. + +Base architecture. The base architecture used in this paper is a relatively simple one and can be easily extended if needed. It includes two modality-specific encoders $\varphi_{a}$ and $\varphi_{v}$ , instantiated as a 2-layer CNN-based audio filter bank encoder and a lightweight VGG-(2+1)D video encoder [57, 58]. Here, $\varphi_{a}$ downsamples $X_{a}$ by 4x to match the video frame rate; $\varphi_{v}$ has a temporal receptive field of 5 frames, or 200ms at 25 FPS, same as the most com + +monly adopted 3D-ResNet18 encoder [41, 59, 64], but has much fewer parameters. A modality-agnostic Transformer-based [71] modality fusion encoder $\theta$ equipped with rotary positional embeddings (RoPE) [65] follows the modality-specific encoders $\varphi_{a}$ and $\varphi_{v}$ . To allow $\theta$ to accept either audio, video, or audio-visual features, we adopt the simple sum operation to compute a fused audio-visual representation $F_{\mathrm{fused}}$ from modality features $F_{m} = \varphi_{m}(X_{m})$ , where $m \in \{a, v\}$ , to allow easy handling of missing modalities: + +$$ +F _ {\text {f u s e d}} = \left\{ \begin{array}{l l} F _ {a} + F _ {v} & \text {i f i n p u t i s a u d i o - v i s u a l ,} \\ F _ {a} + \mathbf {0} = F _ {a} & \text {i f i n p u t i s a u d i o ,} \\ F _ {v} + \mathbf {0} = F _ {v} & \text {i f i n p u t i s v i d e o .} \end{array} \right. \tag {1} +$$ + +The Transformer encoder $\theta$ then yields the final contextual features $E = \theta (F_{\mathrm{fused}})$ + +Based on this base architecture, we build our Siamese based learning framework as shown in Fig. 1(a). A student model (left) is guided by a teacher model (right), which share the same base architecture. The student and teacher accept different versions of the same speech inputs (e.g. masked vs unmasked). The teacher model updates its parameters in an exponential moving average (EMA) manner, and the student is trained to regress the teacher outputs in general. In the subsequent sections, our objective is to optimize the student model with the base architecture to obtain the parameterized audio-visual speech representations $\gamma (\mathbf{a},\mathbf{v};\varphi_{a},\varphi_{v},\theta) = E$ + +# 3.2. Problem Reformulation + +In the audio-visual speech representation learning task, audio and video inputs can be taken as two different views of the same underlying speech signal, carrying different types of information. The relationship between the speech information in the two modalities can be visualized as the bottom-left Venn diagram in Fig. 1(b). Owing to the asymmetry between the two modalities, there exists both modality-unique speech information, i.e. audio- and video-unique information $U_{a}$ and $U_{v}$ . There is also speech information $R$ shared between the two modalities, since they share the same underlying speech production process. At the same time, information that only emerges when both audio and video modality are present is the synergistic information, which we denote by $S$ . Our overall objective is to learn a representation that maximizes the total information $R + U_{a} + U_{v} + S$ for bi-modal (audio-visual) speech representations, as well as $R + U_{a}$ and $R + U_{v}$ for uni-modal (audio & visual) speech representations. + +With this in mind, we reframe the task of learning audiovisual speech representations as a mutual information maximization task. Specifically, the two modalities, corresponding to source variables $X_{a}$ and $X_{v}$ , have mutual information with the target speech content variable $Y$ , containing information useful for downstream tasks that we + +wish to capture. Denote the joint mutual information between the sources and the target by $I(X_{a},X_{v};Y)$ . The problem of learning robust audio-visual speech representations reduces to finding the set of parameters that maximizes $I(\gamma (X_a),\gamma (X_v);Y) = I(E_{av};Y)^1$ . Similarly, the problem of learning robust audio and visual speech representations can be viewed as maximizing the mutual information $I(\gamma (X_a);Y) = I(E_a;Y)$ and $I(\gamma (X_v);Y) = I(E_v;Y)$ , respectively. Taking advantage of the Partial Information Decomposition (PID) framework [73], we can precisely decompose the multivariate mutual information $I(E_{av};Y),I(E_a;Y)$ and $I(E_{v};Y)$ into the previous types of information as: + +$$ +R + U _ {a} = I \left(E _ {a}; Y\right), \tag {2} +$$ + +$$ +R + U _ {v} = I \left(E _ {v}; Y\right), \tag {3} +$$ + +$$ +U _ {a} + S = I \left(E _ {a}; Y \mid X _ {v}\right), \tag {4} +$$ + +$$ +U _ {v} + S = I \left(E _ {v}; Y \mid X _ {a}\right), \tag {5} +$$ + +$$ +R - S = I \left(E _ {a}; E _ {v}; Y\right), \tag {6} +$$ + +$$ +R + U _ {a} + U _ {v} + S = I \left(E _ {a v}; Y\right), \tag {7} +$$ + +where $R, U_{a}, U_{v}$ and $S$ are non-negative. + +To optimize the above mutual information terms with a neural network parameterized by $\varphi_{a},\varphi_{v}$ and $\theta$ we divide the optimization process into three progressive and nonconflicting stages. In stage 1, we learn audio-unique information $U_{a}$ and shared information $R$ from the informative audio modality to maximize $I(E_{a};Y)$ . This process also helps to bootstrap visual representations by learning the shared information $R$ . In stage 2, we turn to learning visual unique information $U_{v}$ by maximizing $I(E_{v};Y) = R + U_{v}$ , which is now easier with the visual information captured via $R$ in stage 1. At the same time, we inject the learned visual information into the audio-visual joint representations to initialize $I(E_{av};Y)$ . Finally, in stage 3 we optimize the joint audio-visual representation on top of the previously learned information by maximizing $I(E_{av};Y)$ . + +# 3.3. The Complete Learning Framework + +To learn network parameters that maximize the mutual information terms $I(E;Y)$ , we leverage the popular self-supervised masked prediction task based on the simple Siamese framework to support the learning process via self-distillation. Inspired by the recent work of Baevski et al. [12], we create multiple masked copies for each training sample, encode the unmasked regions with the Transformer encoder, and train the student to recover contextualized targets of the masked regions with a lightweight 1D convolutional decoder, which are generated by an online EMA teacher. This task can be used to learn any mutual infor + +![](images/3e1e85f97e7651037acf05269d30db860a88065913e5f669cb25d5be1075bfef.jpg) +Figure 1. Overview of the proposed method, $\mathbf{ES}^3$ . (a) We use a Siamese framework with an EMA teacher to implement an iteration-free training procedure. (b) Venn diagrams showing how the three types of information $(R, U, \text{and } S)$ are acquired in an evolving learning process. (c) Illustration of self-distillation with online clustering corresponding to $\mathcal{L}_{\mathrm{distill}}(F_{\mathrm{tea}}, F_{\mathrm{stu}})$ . + +mation $I(E_{;}Y)$ when $Y$ is the downstream speech understanding task. For example, training on auditory or visual speech acts as a proxy to maximize $I(E_{a};Y)$ or $I(E_{v};Y)$ . Self-distillation with online clustering [36]. We perform distillation from the right-hand side teacher to the left-hand side student in Fig. 1(a). Specifically, as illustrated in Fig. 1(c), for each layer $k$ of the top $K$ Transformer layers in $\theta$ with hidden dimension $D$ , we learn a codebook $\mathbf{Q}^{k} = \{\mathbf{q}_{1}^{k},\dots,\mathbf{q}_{Q}^{k}\}$ with $Q$ codewords $\mathbf{q}_i^k\in \mathbb{R}^D$ , which is updated as follows: for each entry $j$ , a set $\tilde{\mathbf{Z}}_{\mathrm{tea}}^k$ of the teacher output features $\tilde{\mathbf{z}}_{\mathrm{tea}}^k$ closest to the current representation of $q$ in the codebook is found as + +$$ +\tilde {\mathbf {Z}} _ {\mathrm {t e a}} ^ {k} = \left\{\tilde {\boldsymbol {z}} _ {\mathrm {t e a}} ^ {k} \Bigg | q = \underset {j \in \{1, 2, \dots , Q \}} {\arg \min } | | \tilde {\boldsymbol {z}} _ {\mathrm {t e a}} ^ {k} - \mathbf {q} _ {j} ^ {k} | | _ {2} \right\}, \tag {8} +$$ + +where the set indices $q$ are used as discrete pseudo-labels to train the student model. Codewords are then updated in an EMA fashion as a weighted sum of the embeddings in $\tilde{\mathbf{Z}}_{\mathrm{tea}}^k$ with decay rate $\tau_{\mathrm{code}}$ . The associated loss function is + +$$ +\mathcal {L} _ {\text {d i s t i l l}} \left(F _ {\text {t e a}}, F _ {\text {s t u}}\right) = \sum_ {t \in \mathbb {M}} \sum_ {k \in (K - N, K ]} \log p _ {\psi_ {k}} \left(q _ {t} \mid \boldsymbol {z} _ {t} ^ {K}\right), \tag {9} +$$ + +where $F_{\mathrm{tea}}$ and $F_{\mathrm{stu}}$ are teacher and student features, $\mathbb{M}$ is the set of masked time steps, $\psi_{k}$ is the softmax activations of linear projections for each target layer $k$ , $q_{t}$ is the codeword index for time step $t$ of the teacher $\tilde{\theta}$ , and $z_{t}^{K}$ is the output of the $K$ -th and final layer of the student model $\theta$ . The teacher and the online clustering process do not require gradients. EMA teacher parameterization. To reduce GPU memory + +usage, we apply EMA updates to the Transformer encoder $\theta$ , but not the earlier modality-specific encoders $\varphi_{a}$ and $\varphi_{v}$ , inspired by [50]. As shown in Fig. 1, the teacher model weights $\tilde{\theta}$ are computed as an exponential moving average of the student model weights $\theta$ : + +$$ +\tilde {\theta} \leftarrow \eta \tilde {\theta} + (1 - \eta) \theta , \tag {10} +$$ + +where $\eta = \min \{\eta_{\max}, (1 + \text{step}) / (10 + \text{step})\}$ is a momentum parameter that slowly increases over time until a pre-determined value $\eta_{\max}$ is reached. + +# 3.4. The Learning Process + +We now elaborate on individual learning stages, based on the above Siamese framework and distillation loss. + +Stage 1: Learning audio-unique information and shared information. In stage 1, we perform distillation from un-masked audio teacher model to the masked audio student model, yielding audio-unique information $U_{a}$ and shared information $R$ from audio. The distillation loss is: + +$$ +\mathcal {L} _ {\mathrm {a} \rightarrow \mathrm {a} ^ {\prime}} = \mathcal {L} _ {\text {d i s t i l l}} \left(E _ {a}, E _ {a} ^ {\prime}\right). \tag {11} +$$ + +Being readily learnable, audio has the potential to provide most of $R$ for learning shared information. However, optimizing $I(E_{a};Y)$ alone cannot help visual learning. Noting that the InfoNCE loss between two variables has been widely used to maximize a lower bound on the mutual information, we adopt the InfoNCE loss to learn shared information corresponding to $I(E_{a};E_{v})$ to bootstrap visual learning with the shared information learned from audio. + +Specifically, we apply it to a fine-grained frame-wise alignment task, where temporally aligned features from the same video are treated as positives, and unaligned ones negatives [19]. Since phonemes can appear repeatedly within the same video clip, leading to false negatives, we use a negative margin following Park et al. [47], which alleviates the effect of falsely matched pairs by enforcing a looser decision boundary. The objective function involved is: + +$$ +\mathcal {L} _ {\mathrm {N C E}} (H) = - \frac {1}{n} \sum_ {i = 1} ^ {n} \log \frac {e ^ {(H _ {i , i} - m) / \tau}}{e ^ {(H _ {i , i} - m) / \tau} + \sum_ {i \neq j} e ^ {H _ {i , j} / \tau}}, \tag {12} +$$ + +where $H$ is a matrix whose $(i,j)$ -th element represents the cosine similarity between $L^2$ -normalized audio and visual features at time step $i$ and $j$ , respectively, $\tau$ is a learnable temperature parameter, and $m$ is a pre-determined margin. To promote the learning of shared features between the two modalities, we apply a symmetrized version of the above NCE loss as: + +$$ +\mathcal {L} _ {\mathrm {a} \leftrightarrow \mathrm {v}} = \frac {1}{2} \left(\mathcal {L} _ {\mathrm {a} \rightarrow \mathrm {v}} + \mathcal {L} _ {\mathrm {v} \rightarrow \mathrm {a}}\right) = \frac {1}{2} \left(\mathcal {L} _ {\mathrm {N C E}} \left(S _ {a v}\right) + \mathcal {L} _ {\mathrm {N C E}} \left(S _ {v a}\right)\right). \tag {13} +$$ + +Here the similarity matrices $S_{av} = \langle F_a, F_v \rangle / (||F_a|| \cdot ||F_v||)$ and $S_{va} = S_{av}^{\top}$ are computed with features $F_a$ and $F_v$ from modality-specific encoders $\varphi_a$ and $\varphi_v$ instead of the contextual features from the Transformer encoder $\theta$ to avoid learning positional shortcuts. The total loss in stage 1 is + +$$ +\mathcal {L} _ {\text {s t a g e 1}} = \mathcal {L} _ {\mathrm {a} \rightarrow \mathrm {a} ^ {\prime}} + \mathcal {L} _ {\mathrm {a} \leftrightarrow \mathrm {v}}. \tag {14} +$$ + +Stage 2: Learning video-unique information. In stage 2, we proceed to learn video-unique information, which is now much easier than learning from scratch since it has been bootstrapped with shared knowledge from the audio in the previous stage. The audio encoder $\varphi_{a}$ does not receive any gradients in this stage and is essentially frozen, since it is expected to have already acquired the audio-specific knowledge well in the previous stage. We adopt a similar distillation loss as stage 1: + +$$ +\mathcal {L} _ {\mathrm {v} \rightarrow \mathrm {v} ^ {\prime}} = \mathcal {L} _ {\text {d i s t i l l}} \left(E _ {v}, E _ {v} ^ {\prime}\right). \tag {15} +$$ + +On top of the learned visual-unique information, we also perform distillation from unmasked video features to the joint features of unmasked audio-visual inputs in this stage to prepare for the learning of the final robust audio-visual representations in the next stage. For this purpose, the loss in this stage is set as: + +$$ +\mathcal {L} _ {\mathrm {a v} \rightarrow \mathrm {v}} = \mathcal {L} _ {\text {d i s t i l l}} \left(E _ {v}, E _ {a v}\right), \tag {16} +$$ + +$$ +\mathcal {L} _ {\text {s t a g e 2}} = \mathcal {L} _ {\mathrm {v} \rightarrow \mathrm {v} ^ {\prime}} + \mathcal {L} _ {\mathrm {a v} \rightarrow \mathrm {v}}. \tag {17} +$$ + +To avoid repeated learning of only shared information between A and V again, the cross-modal loss $\mathcal{L}_{\mathrm{av}\to \mathrm{v}}$ is asymmetric, strictly one-way from AV to V, and a learnable tem + +perature $\tau_{v}$ Eq. (16) is introduced to the visual student outputs, i.e., we replace $\psi_{k}$ in Eq. (9) with softmax activations of temperature-scaled linear projections of $E_{v}$ . + +Stage 3: Learning synergistic information. Finally, we aim to learn synergistic audio-visual information which previous methods have not explicitly addressed. Our ultimate goal is to maximize total task-relevant multi-modal information $I(E_{av};Y)$ , i.e. the sum of $R$ , $U_{a}$ , $U_{v}$ and $S$ , so that the representations are useful for both uni-modal and multi-modal tasks. Naively maximizing this objective without prior stages is problematic because the network tends to be biased towards learning audio-based representations (i.e. prioritizing $R$ and $U_{a}$ ) without considering the visual modality. By learning most of $R$ , $U_{a}$ and $U_{v}$ through the previous stages, our method is less susceptible to learning biased representations in the final stage. At this stage, we always mask identical positions in the audio and video inputs together so that the masked region never contains any unimodal information that can collapse or degrade training. To better recover the masked inputs, the model must cleverly leverage both modalities, which requires learning synergistic information that is not well-captured in prior stages. The loss function in stage 3 can be written as: + +$$ +\mathcal {L} _ {\text {s t a g e 3}} = \mathcal {L} _ {\mathrm {a v} \rightarrow \mathrm {a v} ^ {\prime}} = \mathcal {L} _ {\text {d i s t i l l}} \left(E _ {a v} ^ {\prime}, E _ {a v}\right). \tag {18} +$$ + +By progressively acquiring the three types of useful information, we build strong and robust audio-visual speech representations for both uni- and bi-modal speech, which we shall show in subsequent experiments. + +# 3.5. Fine-tuning Stage + +After pre-training, we add an extra linear layer after the fusion encoder $\theta$ to fine-tune the model end-to-end with the Lattice-Free Maximum Mutual Information (LF-MMI) criterion [26, 51], which is much more lightweight than seq2seq Encoder-Decoder models and allows highly efficient decoding. We adopt a "compact" [33] 1-state hidden Markov model (HMM) topology, using $k2^2$ for differentiable Weighted Finite State Transducer (WFST)-based training. We use 624 bi-char units from text-based clustering [77] for English and tonal phone-level targets (initials and finals) for Chinese. + +# 4. Experiments + +# 4.1. Datasets + +LRS2-BBC [3]. LRS2-BBC is a 224-hour transcribed audio-visual English speech dataset collected from programs of various genres broadcast by the BBC. It is challenging due to the presence of extreme poses, as well as a large and diverse vocabulary. We use the official train / validation split (see supplementary materials) for evaluation. + +LRS3-TED [1]. LRS3-TED is another audio-visual English speech dataset collected from TED and TEDx talks, with over 5.5k speakers and 439 hours of data. It covers diverse topics and is widely adopted for evaluating audio-visual speech representation learning methods for being the largest publicly available one of its kind. An official train / validation split is not given, so we use the widely adopted one prepared by Shi et al. [59], same as [27, 34, 80]. + +CAS-VSR-S101. We also collect a new large-scale, in-the-wild Mandarin dataset with 101.1 hours of data. The videos are sourced from broadcast news and conversational programs in Chinese, covering a highly diverse set of topics, speakers and filming conditions (see supplementary). + +# 4.2. Experimental Setup + +More implementation details can be found in the supplementary materials. + +High- and low-resource setups. On LRS3-TED we follow the widely adopted setup of Shi et al. [59], which excludes the pre-train set during low-resource fine-tuning. To validate data scaling properties, previous works used VoxCeleb2-En [59], which is a 1326-hour subset of VoxCeleb2 [18] in English. However, not only are experiments of this scale computationally expensive, but more importantly, VoxCeleb2 is no longer available from its official source as of Feb. 2023. Therefore, we create a similar scaling protocol, using LRS3-TED (433h) as a pre-training source and LRS2-BBC (223h) for fine-tuning. For low-resource fine-tuning, we only use the 28-hour train subset of LRS2-BBC, while for high-resource fine-tuning, we include the pre-train subset, similar to [17]. + +Evaluation. Decoding is performed using beam search with a beam size of 50. The common evaluation metrics for ASR, VSR and AVSR are Word Error Rate (WER) and Character Error Rate (CER) for English and Chinese, respectively. As our bi-char LF-MMI models are trained on verbatim transcriptions for ease of lexicon construction, we report WERs both without and with using Whisper-style text normalization [54] (in parentheses) to compensate for innocuous errors related to text normalization. + +# 4.3. Results on LRS2-BBC and LRS3-TED + +Low-resource setup. Low-resource results on LRS2-BBC and LRS3-TED can be found in Tabs. 1 and 2. Our smallest BASE* encoder is modality-agnostic, only half the size of the one in [27], which uses two separate encoders to perform AVSR. On LRS3-TED, although we employ the simple LF-MMI instead of cross-entropy (CE) based seq2seq criterion for fine-tuning, our BASE model rivals or achieves SoTA in all cases. When scaling to a LARGE model, we observe significant improvements in both ASR and AVSR. Meanwhile, VSR performance remains comparable to SoTA, with a moderate $2\%$ gap. This may sug + +gest that with our current VGG-based visual encoder $\varphi_v$ , the benefits of scaling up the Transformer primarily manifest in ASR and AVSR. + +High-resource setup. High-resource results on LRS2-BBC and LRS3-TED can be found in Tabs. 1 and 3. On LRS2-BBC, with a 46M model and relying on its own 223h data for pre-training and fine-tuning, we achieve $30.3\% / 3.6\% / 2.9\%$ WER for VSR, ASR and AVSR. Notably, these results surpass a highly competitive supervised method [40] and other self-supervised methods that incorporate additional data. This indicates that our progressive learning approach can effectively exploit visual information. Furthermore, when we introduce additional 433h unlabeled data, we achieve even better performance: $29.8\% / 3.0\% / 2.4\%$ WER, which outperforms other methods that employ the same amount of data. When we switch to the LARGE model, performance rivals SoTA methods that employ 4x data (1759h) for pre-training. Similarly, we achieve or rival SoTA performance on LRS3-TED. + +# 4.4. Results on CAS-VSR-S101 + +We also evaluate on our newly collected CAS-VSR-S101 dataset, which is smaller than both of the two previous datasets and in an even more ambiguous language (Chinese has more homophonous syllables and words than English). As shown in Tab. 4, $\mathrm{ES}^3$ still works out-of-the-box under this challenging setup, and yields noticeable performance gains over a strong fully supervised baseline [39]. Interestingly, our fully supervised baseline model beats [39] in ASR and AVSR but slightly lags behind in VSR, again hinting that our strategy can further benefit from more complex visual modeling architectures. + +Table 4. Results on CAS-VSR-S101. Here we use a BASE* model for the fully-supervised baseline and $\mathrm{ES}^3$ . The self-supervised models are pre-trained on training data and then finetuned for 50 epochs with the labels. The numbers before and after slashes correspond to CERs on the validation and test sets. + +
MethodsCriterionEncoder Size#EpochsVSRASRAVSR
Supervised
BaselineLF-MMI46M10060.0 / 57.415.4 / 12.614.6 / 12.1
E2E Conformer [39]CTC+CE79M10058.6 / 56.516.6 / 13.016.0 / 12.6
Self-supervised
ES3 Stage 161.2 / 59.014.7 / 11.9-
ES3 Stage 2LF-MMI46M5059.1 / 56.7-13.9 / 11.4
ES3 Stage 358.1 / 55.614.4 / 11.613.4 / 11.0
+ +# 4.5. Analysis and Discussion + +Representation quality probing. In Tab. 5, we demonstrate the effect of each stage on LRS2-BBC. ASR performance between stages 1 and 3 are close, indicating that stage 1 indeed acquires most of audio-unique information. Comparing the VSR results between stages 1 and 2 reveals a significant gap, suggesting that stage 2 also effectively learns visual-unique information, hence the minimal differences between the final VSR result and those of stage 2. Moreover, comparing the results of stage 1-2 and stage + +Table 1. Results on LRS2-BBC. We pre-train a BASE* model with 223h unlabeled data from LRS2-BBC, as well as BASE and LARGE models with 433h unlabeled data (LRS3-TED) to demonstrate scaling properties. *: external language models. + +
MethodsUnlabeled AV dataLabeled DataEncoder SizeCriterionVSRASRAVSR
Supervised
Yu et al. [75]-1519h-LF-MMI148.96.75.9
Ma et al. [39]*-380h79MCTC+CE37.93.93.7
Ma et al. [41]*-818h186MCTC+CE27.92.6-
Semi-supervised
Afouras et al. [2]*777h223h2-CTC51.3--
Ma et al. [41]*2630h818h186MCTC+CE14.61.51.5
Self-supervised (BASE and BASE* models)
AV-HuBERT [59, 60]1759h223h103MCE31.23-3.63
VATLM [80]1759h4223h107MCE30.6-2.9
RAVeN [27]433h223h97MCTC+CE32.13.9-
Pan et al. [45]-5380h399MCTC+CE43.22.72.6
ES3(ours)223h28h46M40.2 (39.1)6.0 (5.0)5.7 (4.9)
433h28h102MLF-MMI39.3 (38.2)5.5 (4.8)5.1 (4.1)
223h223h46M31.4 (30.3)4.3 (3.6)3.8 (2.9)
433h223h102M30.7 (29.8)3.4 (3.0)3.2 (2.4)
Self-supervised (LARGE models)
ES3(ours)433h28h317MLF-MMI36.4 (35.4)5.2 (4.4)4.7 (4.0)
433h223h317M26.7 (25.8)3.1 (2.5)3.1 (2.5)
AV-HuBERT [59, 60]1759h223h325MCE25.53-2.53
VATLM [80]1759h4223h332MCE24.3-2.3
RAVeN [27]1759h223h671MCTC+CE23.22.5-
+ +${}^{1}$ Not end-to-end; requires a GMM-HMM alignment stage. ${}^{2}$ Uses an additional ASR model trained on LibriSpeech (960h). +3 Reproduced by Zhu et al. [80]. 4 Uses additional 3846h audio, 452h audio-text and 600M text data. +5 Uses additional 60000h audio data and 1.28M unlabeled images. + +Table 2. Low-resource results on LRS3-TED. We pre-train a BASE and LARGE model with 433h unlabeled data. + +
MethodsUnlabeled AV dataLabeled DataEncoder SizeCriterionVSRASRAVSR
Self-supervised (BASE models)
AV-HuBERT [59, 60]433h30h103MCE51.84.9\( {4.7}^{1} \)
VATLM [80]\( {433}{\mathrm{\;h}}^{2} \)30h107MCE48.0-3.6
RAVEn [27]433h30h97MCTC+CE47.04.7-
\( {\mathrm{{AV2vec}}}^{\left\lbrack {76}\right\rbrack }{}^{3} \)433h30h103MCE45.0-5.8
AV-data2vec [34]433h30h103MCE45.24.44.2
\( {\mathbf{{ES}}}^{3} \) (ours)433h30h102MLF-MMI45.5 (44.7)3.9 (3.3)3.6 (3.0)
Self-supervised (LARGE models)
AV-HuBERT433h30h325MCE44.84.5\( {4.2}^{1} \)
AV-data2vec433h30h325MCE40.53.73.4
\( {\mathbf{{ES}}}^{3} \) (ours)433h30h317MLF-MMI43.5 (42.5)3.8 (2.9)2.9 (2.3)
+ +$^{1}$ Reproduced by Lian et al. [34]. $^{2}$ Uses additional 3846h audio, 452h audio-text and 600M text data. +3 Zhang et al. [76] inject noise during pre-training and obtain better fine-tuning results even with its base model AV-HuBERT $(47.1\%)$ . + +3 demonstrates that the joint information learning in stage 3, which also incorporates synergistic information, benefits both uni-modal (ASR & VSR) and bi-modal (AVSR) tasks. Similar trends can also be observed on the more challenging Mandarin dataset CAS-VSR-S101 in Tab. 4, further supporting the motivation of our proposed strategy. + +Table 5. Representation quality probing on LRS2-BBC. Results are reported on the validation set using the BASE* model under the low-resource setup for each pre-training stage. + +
Pre-training StageVSRASRAVSR
Stage 158.3 (57.3)11.5 (10.1)-
Stage 248.7 (47.3)-12.9 (11.5)
Stage 3 (Full ES³)48.1 (46.9)11.4 (9.8)10.1 (8.6)
+ +Effect of automatically balancing shared and unique information learning. It is known that in a hardness-aware loss function, the temperature rules the strength of penalties on hard negative samples [72]. We track the automatically learned temperature parameter $\tau$ used for cross-modal + +learning objectives $\mathcal{L}_{\mathrm{a}\leftrightarrow \mathrm{v}}$ in Eq. (12) and $\mathcal{L}_{\mathrm{av}\rightarrow \mathrm{v}}$ in Eq. (16) when pre-training on LRS2-BBC, as shown in Fig. 2. In stage 1, $\tau$ continues to fall until it reaches $\tau_{\mathrm{min}}$ , while in stage 2 it briefly drops and then rises throughout training. The trend is in line with our intuition: in stage 1, audio information is easily learnable, so the model can fully exploit shared information by lowering $\tau$ without sacrificing audio-unique information learning; in stage 2, after learning some residual shared information, an increasing $\tau$ progressively relaxes $\mathcal{L}_{\mathrm{av}\rightarrow \mathrm{v}}$ to yield more freedom to $\mathcal{L}_{\mathrm{v}\rightarrow \mathrm{v}'}$ , i.e. video-unique information learning. + +![](images/dc65de78e349c393d51fed49ea503110b75c3317341ac1413d53fd9f30ec9c74.jpg) +Figure 2. The temperature parameter paces learning automatically. + +Table 3. High-resource results on LRS3-TED. We fine-tune on 433h labeled data of LRS3. †: uses test-time augmentation. *: uses external language models. ‡: noise injection during pre-training. We list more methods for comparison in the supplementary material. + +
MethodsYearUnlabeled AV dataLabeled DataBackboneEncoder SizeCriterionVSRASRAVSR
Supervised
Ma et al. [39]*2021-595hConformer79MCTC+CE43.32.32.3
Prajwal et al. [52]†*2022-698hTransformer32MCE40.6--
Ma et al. [40]*2022-818hConformer79MCTC+CE34.7--
Semi-Supervised
Shillingford et al. [62]*2019-3886hRNN-CTC55.1--
Makino et al. [42]2019-31khRNN43MTransducer33.64.84.5
Afouras et al. [2]*2020344h433h1Jasper (CNN)-CTC59.8--
Serdyuk et al. [57]2021-90khTransformer-Transducer25.9-2.3
Ma et al. [40]*2022641h818hConformer79MCTC+CE31.5--
Serdyuk et al. [58]2022-90khConformer-Transducer17.01.61.6
Ma et al. [41]*20232630h818hConformer186MCTC+CE19.11.00.9
Chang et al. [15]2023-100khConformer98MTransducer12.8-0.9
Self-supervised (Base Models)
AV-HuBERT [59, 60]2022433h433hTransformer103MCE44.03.02.82
AV2vec [76]‡2023433h433hTransformer103MCE39.9-2.6
RAVEn [27]2023433h433hTransformer97MCTC+CE39.12.2-
ES3 (ours)2023433h433hTransformer102MLF-MMI40.3 (39.2)2.9 (2.4)2.5 (2.0)
AV-data2vec [34]2023433h433hTransformer103MCE39.02.01.8
Self-supervised (Large Models)
AV-HuBERT [59, 60]2022433h433hTransformer325MCE41.62.72.52
AV-data2vec [34]2023433h433hTransformer325MCE37.41.91.7
ES3 (ours)2023433h433hTransformer317MLF-MMI37.1 (36.7)2.8 (2.1)2.1 (1.7)
+ +1 Uses an additional ASR model trained on LibriSpeech (960h). +2 Reproduced by Lian et al. [34]. + +Visualization of our learned psuedo-label sequences. We randomly select a sample from the LRS3-TED validation set and visualize the pseudo-label sequences, i.e. code indices $q$ generated by the audio, visual and audio-visual EMA teacher $\tilde{\theta}$ in each stage. Ground truth phones for each audio frame (10ms) are derived by forced alignment [79]. Fig. 3 shows that our learned codebook yields high-quality pseudo-labels that strongly correlate with phonemic units. + +![](images/5e00eb854a35c01a20449a847b6613fbe7a3476ae06be938215eb0055ab4b0f6.jpg) +Figure 3. Pseudo-label sequences from each stage. Example of pseudo-labels generated from the last Transformer layer for the utterance "so what does that mean" by the audio, visual, and audio-visual teachers in each stage (BASE model). At the bottom we show the spectrogram for reference. Please note that our pseudolabels are derived at 40ms intervals, limited by video frame rate. + +Training efficiency comparison with SoTA methods. Learning audio-visual speech representations has been extremely compute-intensive. Here we briefly compare $\mathrm{ES}^3$ with prominent SoTA methods. Pre-training a BASE/LARGE model on LRS3-TED using our $\mathrm{ES}^3$ for 100 epochs per stage only takes about 6/8.3 days in total with just 8 NVIDIA A40 GPUs. In contrast, AV-data2vec is trained for 1000k steps using 64 NVIDIA V100 GPUs for 4-5/6-7 days using BASE/LARGE with a batch size of 1280/2560s, or equivalently, 821/1642 epochs. They also experiment extensively with hyperparameters such as modality schedulers, which further increases the cost of + +training. RAVEn performs an equivalent of 6.25/50 days of training on 8 A100s with BASE*/LARGE. It uses separate encoders for audio and video, so the resulting model cannot perform audio-visual fusion out-of-the-box. This also requires keeping track of four sets of parameters, which is very memory-intensive. By focusing on designing a clever learning strategy, we circumvent such manual tuning altogether and greatly reduce training costs. Finally, we use the simple LF-MMI criterion, and expect better performance from tuning beam size, including a larger external language model (see supplementary), or switching to the common heavy Attention-based Encoder-Decoder model. + +# 5. Conclusions + +We present a novel strategy, $\mathrm{ES}^3$ , for self-supervised learning of audio-visual speech representations. Unlike existing methods that primarily focus on learning information shared between audio and video, we take an "evolving" approach that also captures unique and synergistic information. Our experimental results demonstrate that $\mathrm{ES}^3$ indeed progressively learns unique, shared and synergistic information. Moreover, in terms of downstream performance, it rivals SoTA methods in terms of ASR, VSR and AVSR accuracy without burdensome manual balancing efforts. + +Acknowledgments. This work is partially supported by the National Natural Science Foundation of China (No. 62276247, 62076250). We are grateful to the ICT computing platform, the Nanjing Institute of InforSuperBahn, and Capitalonline Data Service Co., Ltd for providing additional compute resources. We also thank Chenhao Wang, Mingmin Yang and Jingyun Xiao for helping to prepare CAS-VSR-S101, Bingquan Xia and Songtao Luo for help with Conformer and seq2seq baselines, Bingquan Xia and Feixiang Wang for help during the rebuttal phase, and Feixiang Wang and Zhaoxin Yuan for an early version of Fig. 1. + +# References + +[1] Triantafyllos Afouras, Joon Son Chung, and Andrew Zisserman. LRS3-TED: a large-scale dataset for visual speech recognition. CoRR, abs/1809.00496, 2018. 6 +[2] Triantafyllos Afouras, Joon Son Chung, and Andrew Zisserman. ASR is all you need: Cross-modal distillation for lip reading. In ICASSP, pages 2143-2147. IEEE, 2020. 2, 7, 8, 4, 5 +[3] Triantafyllos Afouras, Joon Son Chung, Andrew W. Senior, Oriol Vinyals, and Andrew Zisserman. Deep audio-visual speech recognition. IEEE Trans. Pattern Anal. Mach. Intell., 44(12):8717-8727, 2022. 1, 5, 4 +[4] Jean-Baptiste Alayrac, Adrià Recasens, Rosalia Schneider, Relja Arandjelovic, Jason Ramapuram, Jeffrey De Fauw, Lucas Smaira, Sander Dieleman, and Andrew Zisserman. Self-supervised multimodal versatile networks. In NeurlPS, 2020. 2 +[5] Humam Alwassel, Dhruv Mahajan, Bruno Korbar, Lorenzo Torresani, Bernard Ghanem, and Du Tran. Self-supervised learning by cross-modal audio-video clustering. In NeurIPS, 2020. 2 +[6] Relja Arandjelovic and Andrew Zisserman. Look, listen and learn. In ICCV, pages 609-617. IEEE Computer Society, 2017. 2 +[7] Relja Arandjelovic and Andrew Zisserman. Objects that sound. In ECCV(1), pages 451-466. Springer, 2018. 2 +[8] Yuki Markus Asano, Mandela Patrick, Christian Rupprecht, and Andrea Vedaldi. Labelling unlabelled videos from scratch with multi-modal self-supervision. In NeurIPS, 2020. 2 +[9] Alexei Baevski, Michael Auli, and Abdelrahman Mohamed. Effectiveness of self-supervised pre-training for speech recognition. CoRR, abs/1911.03912, 2019. 1 +[10] Alexei Baevski, Yuhao Zhou, Abdelrahman Mohamed, and Michael Auli. wav2vec 2.0: A framework for self-supervised learning of speech representations. In NeurIPS, 2020. 3 +[11] Alexei Baevski, Wei-Ning Hsu, Qiantong Xu, Arun Babu, Jiatao Gu, and Michael Auli. data2vec: A general framework for self-supervised learning in speech, vision and language. In ICML, pages 1298-1312. PMLR, 2022. 1 +[12] Alexei Baevski, Arun Babu, Wei-Ning Hsu, and Michael Auli. Efficient self-supervised learning with contextualized target representations for vision, speech and language. In ICML, pages 1416-1429. PMLR, 2023. 1, 3 +[13] Adrian Bulat and Georgios Tzimiropoulos. How far are we from solving the 2D & 3D face alignment problem? (and a dataset of 230,000 3D facial landmarks). In ICCV, pages 1021-1030. IEEE Computer Society, 2017. 2 +[14] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision Transformers. In ICCV, pages 9630-9640. IEEE, 2021. 1 +[15] Oscar Chang, Hank Liao, Dmitriy Serdyuk, Ankit Shahy, and Olivier Siohan. Conformer is all you need for visual + +speech recognition. In ICASSP, pages 10136-10140. IEEE, 2024. 1, 8, 5 +[16] Brian Chen, Andrew Rouditchenko, Kevin Duarte, Hilde Kuehne, Samuel Thomas, Angie W. Boggust, Rameswar Panda, Brian Kingsbury, Rogério Feris, David Harwath, James R. Glass, Michael Picheny, and Shih-Fu Chang. Multimodal clustering networks for self-supervised learning from unlabeled videos. In ICCV, pages 7992-8001. IEEE, 2021. 2 +[17] Xize Cheng, Tao Jin, Linjun Li, Wang Lin, Xinyu Duan, and Zhou Zhao. OpenSR: Open-modality speech recognition via maintaining multi-modality alignment. In ACL (1), pages 6592-6607. Association for Computational Linguistics, 2023. 6 +[18] Joon Son Chung, Arsha Nagrani, and Andrew Zisserman. VoxCeleb2: Deep speaker recognition. In INTERSPEECH, pages 1086-1090. ISCA, 2018. 6 +[19] Soo-Whan Chung, Joon Son Chung, and Hong-Goo Kang. Perfect match: Improved cross-modal embeddings for audiovisual synchronisation. In ICASSP, pages 3965–3969. IEEE, 2019. 5 +[20] Tri Dao. FlashAttention-2: Faster attention with better parallelism and work partitioning. CoRR, abs/2307.08691, 2023. 2 +[21] Tri Dao, Daniel Y. Fu, Stefano Ermon, Atri Rudra, and Christopher Ré. FlashAttention: Fast and memory-efficient exact attention with IO-awareness. In NeurIPS, 2022. 2 +[22] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. BERT: Pre-training of deep bidirectional transformers for language understanding. In NAACL-HLT (1), pages 4171–4186. Association for Computational Linguistics, 2019. 1, 2 +[23] Angela Fan, Edouard Grave, and Armand Joulin. Reducing Transformer depth on demand with structured dropout. In ICLR. OpenReview.net, 2020. 2 +[24] Dalu Feng, Shuang Yang, and Shiguang Shan. An efficient software for building lip reading models without pains. In ICME Workshops, pages 1-2. IEEE, 2021. 1 +[25] Yuan Gong, Andrew Rouditchenko, Alexander H. Liu, David Harwath, Leonid Karlinsky, Hilde Kuehne, and James R. Glass. Contrastive audio-visual masked autoencoder. In ICLR. OpenReview.net, 2023. 2 +[26] Hossein Hadian, Hossein Sameti, Daniel Povey, and Sanjeev Khudanpur. End-to-end speech recognition using lattice-free MMI. In INTERSPEECH, pages 12-16. ISCA, 2018. 5 +[27] Alexandros Haliassos, Pingchuan Ma, Rodrigo Mira, Stavros Petridis, and Maja Pantic. Jointly learning visual and auditory speech representations from raw data. In ICLR. Open-Review.net, 2023. 2, 6, 7, 8, 1, 4, 5 +[28] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, and Ross B. Girshick. Masked autoencoders are scalable vision learners. In CVPR, pages 15979-15988. IEEE, 2022. 2 +[29] François Hernandez, Vincent Nguyen, Sahar Ghannay, Natalia A. Tomashenko, and Yannick Esteve. TED-LIUM 3: Twice as much data and corpus repartition for experiments on speaker adaptation. In SPECOM, pages 198-208. Springer, 2018. 2 + +[30] Wei-Ning Hsu and Bowen Shi. u-HuBERT: Unified mixed-modal speech pretraining and zero-shot transfer to unlabeled modality. In NeurIPS, 2022. 1 +[31] Wei-Ning Hsu, Benjamin Bolte, Yao-Hung Hubert Tsai, Kushal Lakhotia, Ruslan Salakhutdinov, and Abdelrahman Mohamed. HuBERT: Self-supervised speech representation learning by masked prediction of hidden units. IEEE ACM Trans. Audio Speech Lang. Process., 29:3451-3460, 2021. 1 +[32] Bruno Korbar, Du Tran, and Lorenzo Torresani. Cooperative learning of audio and video models from self-supervised synchronization. In NeurIPS, pages 7774-7785, 2018. 2 +[33] Aleksandr Laptev, Somshubra Majumdar, and Boris Ginsburg. CTC variations through new WFST topologies. In INTERSPEECH, pages 1041-1045. ISCA, 2022. 5 +[34] Jiachen Lian, Alexei Baevski, Wei-Ning Hsu, and Michael Auli. AV-data2vec: Self-supervised learning of audio-visual speech representations with contextualized target representations. In ASRU. IEEE, 2023. 2, 6, 7, 8, 1, 5 +[35] Paul Pu Liang, Zihao Deng, Martin Q. Ma, James Y. Zou, Louis-Philippe Morency, and Ruslan Salakhutdinov. Factorized contrastive learning: Going beyond multi-view redundancy. In NeurIPS, 2023. 2 +[36] Alexander H. Liu, Heng-Jui Chang, Michael Auli, Wei-Ning Hsu, and Jim Glass. DinoSR: Self-distillation and online clustering for self-supervised speech representation learning. In NeurIPS, 2023. 4, 2 +[37] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In *ICLR (Poster)*. OpenReview.net, 2019. 2 +[38] Pingchuan Ma, Rodrigo Mira, Stavros Petridis, Bjorn W. Schuller, and Maja Pantic. LiRA: Learning visual speech representations from audio through self-supervision. In *Interspeech*, pages 3011–3015. ISCA, 2021. 2 +[39] Pingchuan Ma, Stavros Petridis, and Maja Pantic. End-to-end audio-visual speech recognition with Conformers. In ICASSP, pages 7613-7617. IEEE, 2021. 1, 6, 7, 8, 4, 5 +[40] Pingchuan Ma, Stavros Petridis, and Maja Pantic. Visual speech recognition for multiple languages in the wild. Nat. Mac. Intell., 4(11):930-939, 2022. 1, 6, 8, 2, 4, 5 +[41] Pingchuan Ma, Alexandros Haliassos, Adriana Fernandez-Lopez, Honglie Chen, Stavros Petridis, and Maja Pantic. Auto-AVSR: Audio-Visual speech recognition with automatic labels. In ICASSP, pages 1-5. IEEE, 2023. 1, 3, 7, 8, 4, 5 +[42] Takaki Makino, Hank Liao, Yannis M. Assael, Brendan Shillingford, Basilio Garcia, Otavio Braga, and Olivier Siohan. Recurrent neural network transducer for audio-visual speech recognition. In ASRU, pages 905–912. IEEE, 2019. 1, 8, 5 +[43] Antoine Miech, Dimitri Zhukov, Jean-Baptiste Alayrac, Makarand Tapaswi, Ivan Laptev, and Josef Sivic. HowTo100M: Learning a text-video embedding by watching hundred million narrated video clips. In ICCV, pages 2630-2640. IEEE, 2019. 2 +[44] Arsha Nagrani, Chen Sun, David Ross, Rahul Sukthankar, Cordelia Schmid, and Andrew Zisserman. Speech2Action: Cross-modal supervision for action recognition. In CVPR, pages 10314-10323. Computer Vision Foundation / IEEE, 2020. 2 + +[45] Xichen Pan, Peiyu Chen, Yichen Gong, Helong Zhou, Xinbing Wang, and Zhouhan Lin. Leveraging unimodal self-supervised learning for multimodal audio-visual speech recognition. In ACL (1), pages 4491–4503. Association for Computational Linguistics, 2022. 7, 4 +[46] Daniel S. Park, William Chan, Yu Zhang, Chung-Cheng Chiu, Barret Zoph, Ekin D. Cubuk, and Quoc V. Le. SpecAugment: A simple data augmentation method for automatic speech recognition. In INTERSPEECH, pages 2613-2617. ISCA, 2019. 2 +[47] Sooyoung Park, Arda Senocak, and Joon Son Chung. MarginNCE: Robust sound localization with a negative margin. In ICASSP, pages 1-5. IEEE, 2023. 5, 2 +[48] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, Alban Desmaison, Andreas Köpf, Edward Z. Yang, Zachary DeVito, Martin Raison, Alykhan Tejani, Sasank Chilamkurthy, Benoit Steiner, Lu Fang, Junjie Bai, and Soumith Chintala. PyTorch: An imperative style, high-performance deep learning library. In NeurIPS, pages 8024-8035, 2019. 2 +[49] Mandela Patrick, Yuki Markus Asano, Polina Kuznetsova, Ruth Fong, João F. Henriques, Geoffrey Zweig, and Andrea Vedaldi. On compositions of transformations in contrastive self-supervised learning. In ICCV, pages 9557-9567. IEEE, 2021. 2 +[50] Trung Pham, Chaoning Zhang, Axi Niu, Kang Zhang, and Chang D. Yoo. On the pros and cons of momentum encoder in self-supervised visual representation learning. CoRR, abs/2208.05744, 2022. 4 +[51] Daniel Povey, Vijayaditya Peddinti, Daniel Galvez, Pegah Ghahremani, Vimal Manohar, Xingyu Na, Yiming Wang, and Sanjeev Khudanpur. Purely sequence-trained neural networks for ASR based on lattice-free MMI. In INTER-SPEECH, pages 2751-2755. ISCA, 2016. 5 +[52] K. R. Prajwal, Triantafyllos Afouras, and Andrew Zisserman. Sub-word level lip reading with visual attention. In CVPR, pages 5152-5162. IEEE, 2022. 8, 4, 5 +[53] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In ICML, pages 8748-8763. PMLR, 2021. 2 +[54] Alec Radford, Jong Wook Kim, Tao Xu, Greg Brockman, Christine McLeavey, and Ilya Sutskever. Robust speech recognition via large-scale weak supervision. In ICML, pages 28492-28518. PMLR, 2023. 6 +[55] Jordi Robert-Ribes, Jean-Luc Schwartz, Tahar Lallouache, and Pierre Escudier. Complementarity and synergy in b-modal speech: Auditory, visual, and audio-visual identification of french oral vowels in noise. The Journal of the Acoustical Society of America, 103(6):3677-3689, 1998. 1 +[56] Steffen Schneider, Alexei Baevski, Ronan Collobert, and Michael Auli. wav2vec: Unsupervised pre-training for speech recognition. In INTERSPEECH, pages 3465-3469. ISCA, 2019. 3 + +[57] Dmitriy Serdyuk, Otavio Braga, and Olivier Siohan. Audiovisual speech recognition is worth $32 \times 32 \times 8$ voxels. In ASRU, pages 796-802. IEEE, 2021. 2, 8, 5 +[58] Dmitriy Serdyuk, Otavio Braga, and Olivier Siohan. Transformer-based video front-ends for audio-visual speech recognition for single and multi-person video. In INTER-SPEECH, pages 2833–2837. ISCA, 2022. 2, 8, 5 +[59] Bowen Shi, Wei-Ning Hsu, Kushal Lakhotia, and Abdelrahman Mohamed. Learning audio-visual speech representation by masked multimodal cluster prediction. In ICLR. OpenReview.net, 2022. 1, 2, 3, 6, 7, 8, 4, 5 +[60] Bowen Shi, Wei-Ning Hsu, and Abdelrahman Mohamed. Robust self-supervised audio-visual speech recognition. In INTERSPEECH, pages 2118–2122. ISCA, 2022. 7, 8, 4, 5 +[61] Bowen Shi, Abdelrahman Mohamed, and Wei-Ning Hsu. Learning lip-based audio-visual speaker embeddings with AV-HuBERT. In INTERSPEECH, pages 4785–4789. ISCA, 2022. 1 +[62] Brendan Shillingford, Yannis M. Assael, Matthew W. Hoffman, Thomas Paine, Cian Hughes, Utsav Prabhu, Hank Liao, Hasim Sak, Kanishka Rao, Lorrayne Bennett, Marie Mulville, Misha Denil, Ben Coppin, Ben Laurie, Andrew W. Senior, and Nando de Freitas. Large-scale visual speech recognition. In INTERSPEECH, pages 4135-4139. ISCA, 2019. 8, 5 +[63] Nina Shvetsova, Brian Chen, Andrew Rouditchenko, Samuel Thomas, Brian Kingsbury, Rogério Feris, David Harwath, James R. Glass, and Hilde Kuehne. Everything at once - multi-modal fusion Transformer for video retrieval. In CVPR, pages 19988-19997. IEEE, 2022. 2 +[64] Themos Stafylakis and Georgios Tzimiropoulos. Combining residual networks with LSTMs for lipreading. In INTER-SPEECH, pages 3652-3656. ISCA, 2017. 3 +[65] Jianlin Su, Yu Lu, Shengfeng Pan, Bo Wen, and Yunfeng Liu. RoFormer: Enhanced transformer with rotary position embedding. CoRR, abs/2104.09864, 2021. 3 +[66] Weijie Su, Xizhou Zhu, Yue Cao, Bin Li, Lewei Lu, Furu Wei, and Jifeng Dai. VL-BERT: Pre-training of generic visual-linguistic representations. In ICLR. OpenReview.net, 2020. 2 +[67] Chen Sun, Austin Myers, Carl Vondrick, Kevin Murphy, and Cordelia Schmid. VideoBERT: A joint model for video and language representation learning. In ICCV, pages 7463-7472. IEEE, 2019. 2 +[68] Yonglong Tian, Dilip Krishnan, and Phillip Isola. Contrastive multiview coding. In ECCV (11), pages 776-794. Springer, 2020. 2 +[69] Hugo Touvron, Matthieu Cord, Alexandre Sablayrolles, Gabriel Synnaeve, and Hervé Jégou. Going deeper with image Transformers. In ICCV, pages 32-42. IEEE, 2021. 2 +[70] Kristin J Van Engen, Avanti Dey, Mitchell S Sommers, and Jonathan E Peelle. Audiovisual speech perception: Moving beyond mcgurk. The Journal of the Acoustical Society of America, 152(6):3216-3225, 2022. 1 +[71] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In NIPS, pages 5998-6008, 2017. 3, 2 + +[72] Feng Wang and Huaping Liu. Understanding the behaviour of contrastive loss. In CVPR, pages 2495-2504. Computer Vision Foundation / IEEE, 2021. 7 +[73] Paul L. Williams and Randall D. Beer. Nonnegative decomposition of multivariate information. CoRR, abs/1004.2515, 2010. 3 +[74] Bo Xu, Cheng Lu, Yandong Guo, and Jacob Wang. Discriminative multi-modality speech recognition. In CVPR, pages 14421-14430. Computer Vision Foundation / IEEE, 2020. 5 +[75] Jianwei Yu, Shi-Xiong Zhang, Jian Wu, Shahram Ghorbani, Bo Wu, Shiyin Kang, Shansong Liu, Xunying Liu, Helen Meng, and Dong Yu. Audio-visual recognition of overlapped speech for the LRS2 dataset. In ICASSP, pages 6984–6988. IEEE, 2020. 7, 4 +[76] Jing-Xuan Zhang, Genshun Wan, Zhen-Hua Ling, Jia Pan, Jianqing Gao, and Cong Liu. Self-supervised audiovisual speech representations learning by multimodal self-distillation. In ICASSP, pages 1-5. IEEE, 2023. 2, 7, 8, 1, 5 +[77] Xiaohui Zhang, Vimal Manohar, David Zhang, Frank Zhang, Yangyang Shi, Nayan Singhal, Julian Chan, Fuchun Peng, Yatharth Saraf, and Mike Seltzer. On lattice-free boosted MMI training of HMM and CTC-based full-context ASR models. In ASRU, pages 1026-1033. IEEE, 2021. 5 +[78] Ya Zhao, Rui Xu, Xinchao Wang, Peng Hou, Haihong Tang, and Mingli Song. Hearing lips: Improving lip reading by distilling speech recognizers. In AAAI, pages 6917-6924. AAAI Press, 2020. 2 +[79] Jian Zhu, Cong Zhang, and David Jurgens. Phone-to-audio alignment without text: A semi-supervised approach. In ICASSP, pages 8167-8171. IEEE, 2022. 8 +[80] Qiushi Zhu, Long Zhou, Ziqiang Zhang, Shujie Liu, Binxing Jiao, Jie Zhang, Lirong Dai, Daxin Jiang, Jinyu Li, and Furu Wei. VATLM: Visual-Audio-Text pre-training with unified masked prediction for speech representation learning. IEEE Transactions on Multimedia, pages 1–11, 2023. 1, 6, 7, 4, 5 \ No newline at end of file diff --git a/2024/ES3_ Evolving Self-Supervised Learning of Robust Audio-Visual Speech Representations/images.zip b/2024/ES3_ Evolving Self-Supervised Learning of Robust Audio-Visual Speech Representations/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..db0cc6b8da8b2b5875db05f78b79a3d19d53df20 --- /dev/null +++ b/2024/ES3_ Evolving Self-Supervised Learning of Robust Audio-Visual Speech Representations/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a52dbf831e3e38e16474b8aa781be59c12f50c46cb4318a86ae4ed776c23f876 +size 549125 diff --git a/2024/ES3_ Evolving Self-Supervised Learning of Robust Audio-Visual Speech Representations/layout.json b/2024/ES3_ Evolving Self-Supervised Learning of Robust Audio-Visual Speech Representations/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..4f8a7bd51aa48e92637e71a2347e9b67e3a06751 --- /dev/null +++ b/2024/ES3_ Evolving Self-Supervised Learning of Robust Audio-Visual Speech Representations/layout.json @@ -0,0 +1,11114 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 149, + 102, + 444, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 149, + 102, + 444, + 140 + ], + "spans": [ + { + "bbox": [ + 149, + 102, + 444, + 140 + ], + "type": "inline_equation", + "content": "\\mathbf{ES}^{3}" + }, + { + "bbox": [ + 149, + 102, + 444, + 140 + ], + "type": "text", + "content": ": Evolving Self-Supervised Learning of Robust Audio-Visual Speech Representations" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 128, + 160, + 465, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 128, + 160, + 465, + 175 + ], + "spans": [ + { + "bbox": [ + 128, + 160, + 465, + 175 + ], + "type": "text", + "content": "Yuanhang Zhang" + }, + { + "bbox": [ + 128, + 160, + 465, + 175 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 128, + 160, + 465, + 175 + ], + "type": "text", + "content": ", Shuang Yang" + }, + { + "bbox": [ + 128, + 160, + 465, + 175 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 128, + 160, + 465, + 175 + ], + "type": "text", + "content": ", Shiguang Shan" + }, + { + "bbox": [ + 128, + 160, + 465, + 175 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 128, + 160, + 465, + 175 + ], + "type": "text", + "content": ", Xilin Chen" + }, + { + "bbox": [ + 128, + 160, + 465, + 175 + ], + "type": "inline_equation", + "content": "^{1,2}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 66, + 175, + 527, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 175, + 527, + 203 + ], + "spans": [ + { + "bbox": [ + 66, + 175, + 527, + 203 + ], + "type": "text", + "content": "1Key Laboratory of Intelligent Information Processing of Chinese Academy of Sciences (CAS), Institute of Computing Technology, CAS, Beijing 100190, China" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 203, + 463, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 203, + 463, + 217 + ], + "spans": [ + { + "bbox": [ + 130, + 203, + 463, + 217 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 130, + 203, + 463, + 217 + ], + "type": "text", + "content": "University of Chinese Academy of Sciences, Beijing 100049, China" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 112, + 219, + 478, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 112, + 219, + 478, + 232 + ], + "spans": [ + { + "bbox": [ + 112, + 219, + 478, + 232 + ], + "type": "text", + "content": "zhangyuanhang15@mails.ucas.ac.cn, {shuang.yang, sgshan, xlchen}@ict.ac.cn" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 143, + 258, + 192, + 272 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 258, + 192, + 272 + ], + "spans": [ + { + "bbox": [ + 143, + 258, + 192, + 272 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 284, + 289, + 596 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 284, + 289, + 596 + ], + "spans": [ + { + "bbox": [ + 47, + 284, + 289, + 596 + ], + "type": "text", + "content": "We propose a novel strategy, " + }, + { + "bbox": [ + 47, + 284, + 289, + 596 + ], + "type": "inline_equation", + "content": "ES^3" + }, + { + "bbox": [ + 47, + 284, + 289, + 596 + ], + "type": "text", + "content": ", for self-supervised learning of robust audio-visual speech representations from unlabeled talking face videos. While many recent approaches for this task primarily rely on guiding the learning process using the audio modality alone to capture information shared between audio and video, we reframe the problem as the acquisition of shared, unique (modality-specific) and synergistic speech information to address the inherent asymmetry between the modalities. Based on this formulation, we propose a novel \"evolving\" strategy that progressively builds joint audio-visual speech representations that are strong for both uni-modal (audio & visual) and bimodal (audio-visual) speech. First, we leverage the more easily learnable audio modality to initialize audio and visual representations by capturing audio-unique and shared speech information. Next, we incorporate video-unique speech information and bootstrap the audio-visual representations on top of the previously acquired shared knowledge. Finally, we maximize the total audio-visual speech information, including synergistic information to obtain robust and comprehensive representations. We implement " + }, + { + "bbox": [ + 47, + 284, + 289, + 596 + ], + "type": "inline_equation", + "content": "ES^3" + }, + { + "bbox": [ + 47, + 284, + 289, + 596 + ], + "type": "text", + "content": " as a simple Siamese framework and experiments on both English benchmarks and a newly contributed large-scale Mandarin dataset show its effectiveness. In particular, on LRS2-BBC, our smallest model is on par with SoTA models with only 1/2 parameters and 1/8 unlabeled data (223h)." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 609, + 128, + 623 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 609, + 128, + 623 + ], + "spans": [ + { + "bbox": [ + 47, + 609, + 128, + 623 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 629, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 629, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 629, + 288, + 714 + ], + "type": "text", + "content": "The task of learning audio-visual speech representations has gained increasing importance due to its potential to support a wide range of relevant applications. Building on the success in other domains [9, 11, 12, 14, 22, 31], several effective self-supervised approaches have recently emerged for this task by pre-training models on paired or mixed audio, visual and text data with pretext tasks like speech unit" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 259, + 547, + 620 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 259, + 547, + 620 + ], + "spans": [ + { + "bbox": [ + 304, + 259, + 547, + 620 + ], + "type": "text", + "content": "prediction [30, 59, 80]. However, learning representations in the audio-visual speech domain faces unique challenges brought by the inherent asymmetry between the two modalities involved. First, audio and video often convey distinct and complementary types of speech information. For example, " + }, + { + "bbox": [ + 304, + 259, + 547, + 620 + ], + "type": "inline_equation", + "content": "\\mathsf{m} /" + }, + { + "bbox": [ + 304, + 259, + 547, + 620 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 259, + 547, + 620 + ], + "type": "inline_equation", + "content": "\\mathsf{n} /" + }, + { + "bbox": [ + 304, + 259, + 547, + 620 + ], + "type": "text", + "content": " sound similar, but are visually different; on the contrary, " + }, + { + "bbox": [ + 304, + 259, + 547, + 620 + ], + "type": "inline_equation", + "content": "\\mathsf{p} /" + }, + { + "bbox": [ + 304, + 259, + 547, + 620 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 259, + 547, + 620 + ], + "type": "inline_equation", + "content": "\\mathsf{b} /" + }, + { + "bbox": [ + 304, + 259, + 547, + 620 + ], + "type": "text", + "content": " are visually similar but can be distinguished with audio [70]. Even more interesting is the presence of synergistic information – information that can further enhance speech perception [55] but only emerges when audio and video are considered together, and cannot be inferred when either modality is absent. Second, they manifest different amounts of noise and information. Specifically, in most existing audio-visual speech datasets, audio is characterized by high sampling rate and relatively clean speech samples, making audio representations readily learnable from these datasets. In contrast, the visual modality often presents several challenging factors, including missing (e.g. due to self-occlusion in profile faces or low image quality), redundant (e.g. due to appearance variations, makeup and beards etc.) and ambiguous (e.g. due to the lower sampling rate and the presence of homophones) visual information. Their presence complicates learning meaningful and robust representations solely from video. The difficulties encountered in the continuous efforts to improve the performance of visual speech recognition (VSR) [15, 24, 40, 41] models illustrate this persistent obstacle. Additionally, these differences between the two modalities lead to an imbalance between their learning dynamics, requiring additional manual settings [3, 39, 42, 61]." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 629, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 629, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 629, + 547, + 715 + ], + "type": "text", + "content": "In this work, we break down the problem of learning audio-visual speech representations into the learning of three different types of information: shared (across modalities), unique (modality-specific) and synergistic speech information. Based on our formulation, we propose a novel strategy called " + }, + { + "bbox": [ + 304, + 629, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\mathrm{ES^3}" + }, + { + "bbox": [ + 304, + 629, + 547, + 715 + ], + "type": "text", + "content": " (short for Evolving Self-Supervised Learning of Robust Audio-Visual Speech Representations)" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "27069" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 192 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 192 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 192 + ], + "type": "text", + "content": "to tackle the asymmetry between the two modalities for learning robust audio-visual speech representations. Specifically, our " + }, + { + "bbox": [ + 46, + 72, + 289, + 192 + ], + "type": "inline_equation", + "content": "\\mathrm{ES}^3" + }, + { + "bbox": [ + 46, + 72, + 289, + 192 + ], + "type": "text", + "content": " employs a three-stage progressive learning approach. Initially, it focuses on acquiring the easily learnable audio-unique and shared speech information. Then, it addresses the more challenging visual-unique speech information while incorporating it into the initial audio-visual joint representation. Lastly, it optimizes the final audio-visual representations by leveraging synergy, on top of the acquired shared and modality-unique information." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 59, + 193, + 258, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 193, + 258, + 204 + ], + "spans": [ + { + "bbox": [ + 59, + 193, + 258, + 204 + ], + "type": "text", + "content": "Our contributions can be summarized as follows:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 205, + 287, + 384 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 47, + 205, + 287, + 252 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 205, + 287, + 252 + ], + "spans": [ + { + "bbox": [ + 47, + 205, + 287, + 252 + ], + "type": "text", + "content": "- We propose a new formulation for audio-visual speech representation learning by decomposing the task into the acquisition of shared (across modalities), unique (modality-specific) and synergistic speech information." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 253, + 287, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 253, + 287, + 312 + ], + "spans": [ + { + "bbox": [ + 47, + 253, + 287, + 312 + ], + "type": "text", + "content": "- We propose a novel strategy, " + }, + { + "bbox": [ + 47, + 253, + 287, + 312 + ], + "type": "inline_equation", + "content": "\\mathrm{ES}^3" + }, + { + "bbox": [ + 47, + 253, + 287, + 312 + ], + "type": "text", + "content": " to progressively learn the three types of information in a self-supervised manner with a simple Siamese framework. This evolving strategy significantly alleviates the previous burden of tuning balancing parameters." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 313, + 287, + 348 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 313, + 287, + 348 + ], + "spans": [ + { + "bbox": [ + 47, + 313, + 287, + 348 + ], + "type": "text", + "content": "- Extensive experiments on two widely used benchmarks show that " + }, + { + "bbox": [ + 47, + 313, + 287, + 348 + ], + "type": "inline_equation", + "content": "\\mathrm{ES}^3" + }, + { + "bbox": [ + 47, + 313, + 287, + 348 + ], + "type": "text", + "content": " rivals previous methods in auditory, visual and audio-visual speech recognition performance." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 349, + 287, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 349, + 287, + 384 + ], + "spans": [ + { + "bbox": [ + 47, + 349, + 287, + 384 + ], + "type": "text", + "content": "- We also contribute a new large-scale Mandarin audiovisual speech dataset named CAS-VSR-S101 to the community for evaluation of similar methods." + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 47, + 393, + 134, + 405 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 393, + 134, + 405 + ], + "spans": [ + { + "bbox": [ + 47, + 393, + 134, + 405 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 413, + 289, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 413, + 289, + 652 + ], + "spans": [ + { + "bbox": [ + 46, + 413, + 289, + 652 + ], + "type": "text", + "content": "Self-supervised learning from multi-modal data. The arguably most successful paradigm for self-supervised multimodal learning in the past few years is contrastive learning, which exploits the common information encoded in different modalities [68], such as audio (A) [6, 7], image (I) [53, 66], video (V) [5, 8, 32, 67] and text (T) [4, 16, 43, 44, 49, 63]. These approaches focus on learning semantically aligned multi-modal representations that are effective for specific downstream tasks like cross-modal retrieval, but discard modality-unique information that is useful for other uni-modal tasks. To solve this problem, Gong et al. [25] combine contrastive learning with the masked prediction objective [28] to learn strong audio-visual representations suitable for both retrieval and joint classification. Liang et al. [35] propose a new algorithm, FACTORCL to learn self-supervised multi-modal representations that capture both shared and unique information and verify its effectiveness on multiple tasks. In this paper, we explicitly address capturing unique, shared and synergistic information for uni- and multi-modal downstream tasks." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 654, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 654, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 654, + 289, + 715 + ], + "type": "text", + "content": "Learning audio-visual speech representations. A main challenge of learning audio-visual speech representation is properly handling the aforementioned asymmetry between audio and video. An established approach is to leverage the easily learnable audio modality as a teacher to guide" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 72, + 547, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 396 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 396 + ], + "type": "text", + "content": "visual or audio-visual speech representation learning. Implementing additional measures to mitigate audio dominance and promote visual features during learning is vital for this paradigm. For instance, Shi et al. [59] and Zhang et al. [76] use different masking strengths for the two modalities, apply modality dropout, and enforce simple audio features by only using a weak linear projection layer. Others have explored freezing pre-trained audio representations [2, 38, 78]. However, such an audio-centric approach is sub-optimal and inefficient. First, it limits the learned representations to capturing only shared, i.e. modality-invariant information, or at most, shared and audio-unique information. Recently, a few attempts have been made to go beyond merely using audio as the anchor modality. Haliassos et al. [27] combine multiple audio-visual prediction tasks in place of purely audio-centric ones, and treats A-V asymmetry by having the visual stream predict only auditory but not visual targets. Lian et al. [34] perform a series of trials to determine an optimal schedule for audio and video input combination. Second, the process of exhaustively testing and identifying the optimal set of balancing parameters such as masking ratios, modality schedules and modality dropout rate is extremely time-consuming and computationally demanding. In contrast, our newly proposed evolving strategy identifies different types of task-relevant information and learns them progressively, while avoiding pitfalls and computational burdens (see supp)." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 399, + 416, + 414 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 399, + 416, + 414 + ], + "spans": [ + { + "bbox": [ + 306, + 399, + 416, + 414 + ], + "type": "text", + "content": "3. Our Proposed ES3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 420, + 392, + 431 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 420, + 392, + 431 + ], + "spans": [ + { + "bbox": [ + 306, + 420, + 392, + 431 + ], + "type": "text", + "content": "3.1. Preliminaries" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 437, + 547, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 437, + 547, + 617 + ], + "spans": [ + { + "bbox": [ + 304, + 437, + 547, + 617 + ], + "type": "text", + "content": "The goal of " + }, + { + "bbox": [ + 304, + 437, + 547, + 617 + ], + "type": "inline_equation", + "content": "\\mathrm{ES}^3" + }, + { + "bbox": [ + 304, + 437, + 547, + 617 + ], + "type": "text", + "content": " is to learn audio-visual speech representations from unlabeled talking face videos. The inputs to our " + }, + { + "bbox": [ + 304, + 437, + 547, + 617 + ], + "type": "inline_equation", + "content": "\\mathrm{ES}^3" + }, + { + "bbox": [ + 304, + 437, + 547, + 617 + ], + "type": "text", + "content": " are paired audio-visual data " + }, + { + "bbox": [ + 304, + 437, + 547, + 617 + ], + "type": "inline_equation", + "content": "X_{a,v} = \\{(X_{a,i},X_{v,i})\\mid i = 1,2,\\ldots ,N\\}" + }, + { + "bbox": [ + 304, + 437, + 547, + 617 + ], + "type": "text", + "content": ". Each of the " + }, + { + "bbox": [ + 304, + 437, + 547, + 617 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 437, + 547, + 617 + ], + "type": "text", + "content": " samples consists of a sequence of audio frames " + }, + { + "bbox": [ + 304, + 437, + 547, + 617 + ], + "type": "inline_equation", + "content": "X_{a,i} = \\operatorname {LogFBank}(\\mathbf{a}_i)\\in \\mathbb{R}^{F\\times T_{a,i}}" + }, + { + "bbox": [ + 304, + 437, + 547, + 617 + ], + "type": "text", + "content": " and a temporally synchronized video sequence " + }, + { + "bbox": [ + 304, + 437, + 547, + 617 + ], + "type": "inline_equation", + "content": "X_{v,i} = \\mathbf{v}_i\\in \\mathbb{R}^{C\\times H\\times W\\times T_{v,i}}" + }, + { + "bbox": [ + 304, + 437, + 547, + 617 + ], + "type": "text", + "content": " showing the lips of a speaker, where " + }, + { + "bbox": [ + 304, + 437, + 547, + 617 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 304, + 437, + 547, + 617 + ], + "type": "text", + "content": " is the filter bank dimension, " + }, + { + "bbox": [ + 304, + 437, + 547, + 617 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 304, + 437, + 547, + 617 + ], + "type": "text", + "content": " is the number of video channels, " + }, + { + "bbox": [ + 304, + 437, + 547, + 617 + ], + "type": "inline_equation", + "content": "T_{a,i}" + }, + { + "bbox": [ + 304, + 437, + 547, + 617 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 437, + 547, + 617 + ], + "type": "inline_equation", + "content": "T_{v,i}" + }, + { + "bbox": [ + 304, + 437, + 547, + 617 + ], + "type": "text", + "content": " are the total number of audio and video frames in the " + }, + { + "bbox": [ + 304, + 437, + 547, + 617 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 437, + 547, + 617 + ], + "type": "text", + "content": "-th sample, and " + }, + { + "bbox": [ + 304, + 437, + 547, + 617 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 304, + 437, + 547, + 617 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 437, + 547, + 617 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 304, + 437, + 547, + 617 + ], + "type": "text", + "content": " are the width and height of each frame. Audio and video are sampled at 100 and 25 FPS, i.e. for a 1-second clip, " + }, + { + "bbox": [ + 304, + 437, + 547, + 617 + ], + "type": "inline_equation", + "content": "T_{a} = 4T_{v} = 100" + }, + { + "bbox": [ + 304, + 437, + 547, + 617 + ], + "type": "text", + "content": ". The goal is to learn a " + }, + { + "bbox": [ + 304, + 437, + 547, + 617 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 304, + 437, + 547, + 617 + ], + "type": "text", + "content": "-dimensional contextualized representation " + }, + { + "bbox": [ + 304, + 437, + 547, + 617 + ], + "type": "inline_equation", + "content": "\\gamma \\in \\mathbb{R}^{d\\times T_v}" + }, + { + "bbox": [ + 304, + 437, + 547, + 617 + ], + "type": "text", + "content": " capturing vital speech and context information within the two modality spaces." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 618, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 618, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 618, + 547, + 713 + ], + "type": "text", + "content": "Base architecture. The base architecture used in this paper is a relatively simple one and can be easily extended if needed. It includes two modality-specific encoders " + }, + { + "bbox": [ + 304, + 618, + 547, + 713 + ], + "type": "inline_equation", + "content": "\\varphi_{a}" + }, + { + "bbox": [ + 304, + 618, + 547, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 618, + 547, + 713 + ], + "type": "inline_equation", + "content": "\\varphi_{v}" + }, + { + "bbox": [ + 304, + 618, + 547, + 713 + ], + "type": "text", + "content": ", instantiated as a 2-layer CNN-based audio filter bank encoder and a lightweight VGG-(2+1)D video encoder [57, 58]. Here, " + }, + { + "bbox": [ + 304, + 618, + 547, + 713 + ], + "type": "inline_equation", + "content": "\\varphi_{a}" + }, + { + "bbox": [ + 304, + 618, + 547, + 713 + ], + "type": "text", + "content": " downsamples " + }, + { + "bbox": [ + 304, + 618, + 547, + 713 + ], + "type": "inline_equation", + "content": "X_{a}" + }, + { + "bbox": [ + 304, + 618, + 547, + 713 + ], + "type": "text", + "content": " by 4x to match the video frame rate; " + }, + { + "bbox": [ + 304, + 618, + 547, + 713 + ], + "type": "inline_equation", + "content": "\\varphi_{v}" + }, + { + "bbox": [ + 304, + 618, + 547, + 713 + ], + "type": "text", + "content": " has a temporal receptive field of 5 frames, or 200ms at 25 FPS, same as the most com" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "27070" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "type": "text", + "content": "monly adopted 3D-ResNet18 encoder [41, 59, 64], but has much fewer parameters. A modality-agnostic Transformer-based [71] modality fusion encoder " + }, + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "type": "text", + "content": " equipped with rotary positional embeddings (RoPE) [65] follows the modality-specific encoders " + }, + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "type": "inline_equation", + "content": "\\varphi_{a}" + }, + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "type": "inline_equation", + "content": "\\varphi_{v}" + }, + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "type": "text", + "content": ". To allow " + }, + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "type": "text", + "content": " to accept either audio, video, or audio-visual features, we adopt the simple sum operation to compute a fused audio-visual representation " + }, + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "type": "inline_equation", + "content": "F_{\\mathrm{fused}}" + }, + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "type": "text", + "content": " from modality features " + }, + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "type": "inline_equation", + "content": "F_{m} = \\varphi_{m}(X_{m})" + }, + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "type": "inline_equation", + "content": "m \\in \\{a, v\\}" + }, + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "type": "text", + "content": ", to allow easy handling of missing modalities:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 58, + 186, + 287, + 229 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 186, + 287, + 229 + ], + "spans": [ + { + "bbox": [ + 58, + 186, + 287, + 229 + ], + "type": "interline_equation", + "content": "F _ {\\text {f u s e d}} = \\left\\{ \\begin{array}{l l} F _ {a} + F _ {v} & \\text {i f i n p u t i s a u d i o - v i s u a l ,} \\\\ F _ {a} + \\mathbf {0} = F _ {a} & \\text {i f i n p u t i s a u d i o ,} \\\\ F _ {v} + \\mathbf {0} = F _ {v} & \\text {i f i n p u t i s v i d e o .} \\end{array} \\right. \\tag {1}", + "image_path": "2d6b51fcfa3ebac3561d180ce5e17aa9d00494d60ec39553148077e96a4e17db.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 235, + 287, + 258 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 235, + 287, + 258 + ], + "spans": [ + { + "bbox": [ + 46, + 235, + 287, + 258 + ], + "type": "text", + "content": "The Transformer encoder " + }, + { + "bbox": [ + 46, + 235, + 287, + 258 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 46, + 235, + 287, + 258 + ], + "type": "text", + "content": " then yields the final contextual features " + }, + { + "bbox": [ + 46, + 235, + 287, + 258 + ], + "type": "inline_equation", + "content": "E = \\theta (F_{\\mathrm{fused}})" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 258, + 288, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 258, + 288, + 403 + ], + "spans": [ + { + "bbox": [ + 46, + 258, + 288, + 403 + ], + "type": "text", + "content": "Based on this base architecture, we build our Siamese based learning framework as shown in Fig. 1(a). A student model (left) is guided by a teacher model (right), which share the same base architecture. The student and teacher accept different versions of the same speech inputs (e.g. masked vs unmasked). The teacher model updates its parameters in an exponential moving average (EMA) manner, and the student is trained to regress the teacher outputs in general. In the subsequent sections, our objective is to optimize the student model with the base architecture to obtain the parameterized audio-visual speech representations " + }, + { + "bbox": [ + 46, + 258, + 288, + 403 + ], + "type": "inline_equation", + "content": "\\gamma (\\mathbf{a},\\mathbf{v};\\varphi_{a},\\varphi_{v},\\theta) = E" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 407, + 181, + 419 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 407, + 181, + 419 + ], + "spans": [ + { + "bbox": [ + 47, + 407, + 181, + 419 + ], + "type": "text", + "content": "3.2. Problem Reformulation" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 426, + 287, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 426, + 287, + 641 + ], + "spans": [ + { + "bbox": [ + 46, + 426, + 287, + 641 + ], + "type": "text", + "content": "In the audio-visual speech representation learning task, audio and video inputs can be taken as two different views of the same underlying speech signal, carrying different types of information. The relationship between the speech information in the two modalities can be visualized as the bottom-left Venn diagram in Fig. 1(b). Owing to the asymmetry between the two modalities, there exists both modality-unique speech information, i.e. audio- and video-unique information " + }, + { + "bbox": [ + 46, + 426, + 287, + 641 + ], + "type": "inline_equation", + "content": "U_{a}" + }, + { + "bbox": [ + 46, + 426, + 287, + 641 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 426, + 287, + 641 + ], + "type": "inline_equation", + "content": "U_{v}" + }, + { + "bbox": [ + 46, + 426, + 287, + 641 + ], + "type": "text", + "content": ". There is also speech information " + }, + { + "bbox": [ + 46, + 426, + 287, + 641 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 46, + 426, + 287, + 641 + ], + "type": "text", + "content": " shared between the two modalities, since they share the same underlying speech production process. At the same time, information that only emerges when both audio and video modality are present is the synergistic information, which we denote by " + }, + { + "bbox": [ + 46, + 426, + 287, + 641 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 46, + 426, + 287, + 641 + ], + "type": "text", + "content": ". Our overall objective is to learn a representation that maximizes the total information " + }, + { + "bbox": [ + 46, + 426, + 287, + 641 + ], + "type": "inline_equation", + "content": "R + U_{a} + U_{v} + S" + }, + { + "bbox": [ + 46, + 426, + 287, + 641 + ], + "type": "text", + "content": " for bi-modal (audio-visual) speech representations, as well as " + }, + { + "bbox": [ + 46, + 426, + 287, + 641 + ], + "type": "inline_equation", + "content": "R + U_{a}" + }, + { + "bbox": [ + 46, + 426, + 287, + 641 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 426, + 287, + 641 + ], + "type": "inline_equation", + "content": "R + U_{v}" + }, + { + "bbox": [ + 46, + 426, + 287, + 641 + ], + "type": "text", + "content": " for uni-modal (audio & visual) speech representations." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 642, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 642, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 642, + 288, + 714 + ], + "type": "text", + "content": "With this in mind, we reframe the task of learning audiovisual speech representations as a mutual information maximization task. Specifically, the two modalities, corresponding to source variables " + }, + { + "bbox": [ + 46, + 642, + 288, + 714 + ], + "type": "inline_equation", + "content": "X_{a}" + }, + { + "bbox": [ + 46, + 642, + 288, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 642, + 288, + 714 + ], + "type": "inline_equation", + "content": "X_{v}" + }, + { + "bbox": [ + 46, + 642, + 288, + 714 + ], + "type": "text", + "content": ", have mutual information with the target speech content variable " + }, + { + "bbox": [ + 46, + 642, + 288, + 714 + ], + "type": "inline_equation", + "content": "Y" + }, + { + "bbox": [ + 46, + 642, + 288, + 714 + ], + "type": "text", + "content": ", containing information useful for downstream tasks that we" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 72, + 547, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 228 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 228 + ], + "type": "text", + "content": "wish to capture. Denote the joint mutual information between the sources and the target by " + }, + { + "bbox": [ + 304, + 72, + 547, + 228 + ], + "type": "inline_equation", + "content": "I(X_{a},X_{v};Y)" + }, + { + "bbox": [ + 304, + 72, + 547, + 228 + ], + "type": "text", + "content": ". The problem of learning robust audio-visual speech representations reduces to finding the set of parameters that maximizes " + }, + { + "bbox": [ + 304, + 72, + 547, + 228 + ], + "type": "inline_equation", + "content": "I(\\gamma (X_a),\\gamma (X_v);Y) = I(E_{av};Y)^1" + }, + { + "bbox": [ + 304, + 72, + 547, + 228 + ], + "type": "text", + "content": ". Similarly, the problem of learning robust audio and visual speech representations can be viewed as maximizing the mutual information " + }, + { + "bbox": [ + 304, + 72, + 547, + 228 + ], + "type": "inline_equation", + "content": "I(\\gamma (X_a);Y) = I(E_a;Y)" + }, + { + "bbox": [ + 304, + 72, + 547, + 228 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 72, + 547, + 228 + ], + "type": "inline_equation", + "content": "I(\\gamma (X_v);Y) = I(E_v;Y)" + }, + { + "bbox": [ + 304, + 72, + 547, + 228 + ], + "type": "text", + "content": ", respectively. Taking advantage of the Partial Information Decomposition (PID) framework [73], we can precisely decompose the multivariate mutual information " + }, + { + "bbox": [ + 304, + 72, + 547, + 228 + ], + "type": "inline_equation", + "content": "I(E_{av};Y),I(E_a;Y)" + }, + { + "bbox": [ + 304, + 72, + 547, + 228 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 72, + 547, + 228 + ], + "type": "inline_equation", + "content": "I(E_{v};Y)" + }, + { + "bbox": [ + 304, + 72, + 547, + 228 + ], + "type": "text", + "content": " into the previous types of information as:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 392, + 238, + 545, + 251 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 392, + 238, + 545, + 251 + ], + "spans": [ + { + "bbox": [ + 392, + 238, + 545, + 251 + ], + "type": "interline_equation", + "content": "R + U _ {a} = I \\left(E _ {a}; Y\\right), \\tag {2}", + "image_path": "dee8d73d6b22635cf24d527dfb29bba3e87c7090a4eab607958f23ff0d749237.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 393, + 253, + 545, + 266 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 393, + 253, + 545, + 266 + ], + "spans": [ + { + "bbox": [ + 393, + 253, + 545, + 266 + ], + "type": "interline_equation", + "content": "R + U _ {v} = I \\left(E _ {v}; Y\\right), \\tag {3}", + "image_path": "0e26e3120cf83e2ba665729bf934b25918c9da99609531bf24098039e8c0d315.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 393, + 268, + 545, + 280 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 393, + 268, + 545, + 280 + ], + "spans": [ + { + "bbox": [ + 393, + 268, + 545, + 280 + ], + "type": "interline_equation", + "content": "U _ {a} + S = I \\left(E _ {a}; Y \\mid X _ {v}\\right), \\tag {4}", + "image_path": "6d59431d7a109121aaa81768ecf1d6bc19620d1518f24677af6a8dff3c5d5e06.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 393, + 283, + 545, + 295 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 393, + 283, + 545, + 295 + ], + "spans": [ + { + "bbox": [ + 393, + 283, + 545, + 295 + ], + "type": "interline_equation", + "content": "U _ {v} + S = I \\left(E _ {v}; Y \\mid X _ {a}\\right), \\tag {5}", + "image_path": "7d8bcafd34698422ce1ebda7bb96aa36450f43233e76058eb85929bbbc6f8101.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 398, + 298, + 545, + 311 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 398, + 298, + 545, + 311 + ], + "spans": [ + { + "bbox": [ + 398, + 298, + 545, + 311 + ], + "type": "interline_equation", + "content": "R - S = I \\left(E _ {a}; E _ {v}; Y\\right), \\tag {6}", + "image_path": "dc0b2bec245ffdb4918fd4b0d143febba4dbc4ca1e88b00f23c5d119dfdc5ece.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 350, + 313, + 545, + 326 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 350, + 313, + 545, + 326 + ], + "spans": [ + { + "bbox": [ + 350, + 313, + 545, + 326 + ], + "type": "interline_equation", + "content": "R + U _ {a} + U _ {v} + S = I \\left(E _ {a v}; Y\\right), \\tag {7}", + "image_path": "1f222fca7588844b808a8bdcde3ffc81e486203a6d205c864ca4c82c9e412457.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 335, + 473, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 335, + 473, + 346 + ], + "spans": [ + { + "bbox": [ + 305, + 335, + 473, + 346 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 335, + 473, + 346 + ], + "type": "inline_equation", + "content": "R, U_{a}, U_{v}" + }, + { + "bbox": [ + 305, + 335, + 473, + 346 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 335, + 473, + 346 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 305, + 335, + 473, + 346 + ], + "type": "text", + "content": " are non-negative." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 347, + 545, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 347, + 545, + 529 + ], + "spans": [ + { + "bbox": [ + 304, + 347, + 545, + 529 + ], + "type": "text", + "content": "To optimize the above mutual information terms with a neural network parameterized by " + }, + { + "bbox": [ + 304, + 347, + 545, + 529 + ], + "type": "inline_equation", + "content": "\\varphi_{a},\\varphi_{v}" + }, + { + "bbox": [ + 304, + 347, + 545, + 529 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 347, + 545, + 529 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 304, + 347, + 545, + 529 + ], + "type": "text", + "content": " we divide the optimization process into three progressive and nonconflicting stages. In stage 1, we learn audio-unique information " + }, + { + "bbox": [ + 304, + 347, + 545, + 529 + ], + "type": "inline_equation", + "content": "U_{a}" + }, + { + "bbox": [ + 304, + 347, + 545, + 529 + ], + "type": "text", + "content": " and shared information " + }, + { + "bbox": [ + 304, + 347, + 545, + 529 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 304, + 347, + 545, + 529 + ], + "type": "text", + "content": " from the informative audio modality to maximize " + }, + { + "bbox": [ + 304, + 347, + 545, + 529 + ], + "type": "inline_equation", + "content": "I(E_{a};Y)" + }, + { + "bbox": [ + 304, + 347, + 545, + 529 + ], + "type": "text", + "content": ". This process also helps to bootstrap visual representations by learning the shared information " + }, + { + "bbox": [ + 304, + 347, + 545, + 529 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 304, + 347, + 545, + 529 + ], + "type": "text", + "content": ". In stage 2, we turn to learning visual unique information " + }, + { + "bbox": [ + 304, + 347, + 545, + 529 + ], + "type": "inline_equation", + "content": "U_{v}" + }, + { + "bbox": [ + 304, + 347, + 545, + 529 + ], + "type": "text", + "content": " by maximizing " + }, + { + "bbox": [ + 304, + 347, + 545, + 529 + ], + "type": "inline_equation", + "content": "I(E_{v};Y) = R + U_{v}" + }, + { + "bbox": [ + 304, + 347, + 545, + 529 + ], + "type": "text", + "content": ", which is now easier with the visual information captured via " + }, + { + "bbox": [ + 304, + 347, + 545, + 529 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 304, + 347, + 545, + 529 + ], + "type": "text", + "content": " in stage 1. At the same time, we inject the learned visual information into the audio-visual joint representations to initialize " + }, + { + "bbox": [ + 304, + 347, + 545, + 529 + ], + "type": "inline_equation", + "content": "I(E_{av};Y)" + }, + { + "bbox": [ + 304, + 347, + 545, + 529 + ], + "type": "text", + "content": ". Finally, in stage 3 we optimize the joint audio-visual representation on top of the previously learned information by maximizing " + }, + { + "bbox": [ + 304, + 347, + 545, + 529 + ], + "type": "inline_equation", + "content": "I(E_{av};Y)" + }, + { + "bbox": [ + 304, + 347, + 545, + 529 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 305, + 535, + 499, + 548 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 535, + 499, + 548 + ], + "spans": [ + { + "bbox": [ + 305, + 535, + 499, + 548 + ], + "type": "text", + "content": "3.3. The Complete Learning Framework" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 553, + 545, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 553, + 545, + 685 + ], + "spans": [ + { + "bbox": [ + 304, + 553, + 545, + 685 + ], + "type": "text", + "content": "To learn network parameters that maximize the mutual information terms " + }, + { + "bbox": [ + 304, + 553, + 545, + 685 + ], + "type": "inline_equation", + "content": "I(E;Y)" + }, + { + "bbox": [ + 304, + 553, + 545, + 685 + ], + "type": "text", + "content": ", we leverage the popular self-supervised masked prediction task based on the simple Siamese framework to support the learning process via self-distillation. Inspired by the recent work of Baevski et al. [12], we create multiple masked copies for each training sample, encode the unmasked regions with the Transformer encoder, and train the student to recover contextualized targets of the masked regions with a lightweight 1D convolutional decoder, which are generated by an online EMA teacher. This task can be used to learn any mutual infor" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 306, + 693, + 545, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 693, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 693, + 545, + 713 + ], + "type": "text", + "content": "1This equality holds since the architecture implemented in this work only models additive modality synergy stemming from sum fusion." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "27071" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 51, + 73, + 545, + 305 + ], + "blocks": [ + { + "bbox": [ + 51, + 73, + 545, + 305 + ], + "lines": [ + { + "bbox": [ + 51, + 73, + 545, + 305 + ], + "spans": [ + { + "bbox": [ + 51, + 73, + 545, + 305 + ], + "type": "image", + "image_path": "3e1e85f97e7651037acf05269d30db860a88065913e5f669cb25d5be1075bfef.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 307, + 547, + 340 + ], + "lines": [ + { + "bbox": [ + 46, + 307, + 547, + 340 + ], + "spans": [ + { + "bbox": [ + 46, + 307, + 547, + 340 + ], + "type": "text", + "content": "Figure 1. Overview of the proposed method, " + }, + { + "bbox": [ + 46, + 307, + 547, + 340 + ], + "type": "inline_equation", + "content": "\\mathbf{ES}^3" + }, + { + "bbox": [ + 46, + 307, + 547, + 340 + ], + "type": "text", + "content": ". (a) We use a Siamese framework with an EMA teacher to implement an iteration-free training procedure. (b) Venn diagrams showing how the three types of information " + }, + { + "bbox": [ + 46, + 307, + 547, + 340 + ], + "type": "inline_equation", + "content": "(R, U, \\text{and } S)" + }, + { + "bbox": [ + 46, + 307, + 547, + 340 + ], + "type": "text", + "content": " are acquired in an evolving learning process. (c) Illustration of self-distillation with online clustering corresponding to " + }, + { + "bbox": [ + 46, + 307, + 547, + 340 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{distill}}(F_{\\mathrm{tea}}, F_{\\mathrm{stu}})" + }, + { + "bbox": [ + 46, + 307, + 547, + 340 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 347, + 287, + 493 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 347, + 287, + 493 + ], + "spans": [ + { + "bbox": [ + 46, + 347, + 287, + 493 + ], + "type": "text", + "content": "mation " + }, + { + "bbox": [ + 46, + 347, + 287, + 493 + ], + "type": "inline_equation", + "content": "I(E_{;}Y)" + }, + { + "bbox": [ + 46, + 347, + 287, + 493 + ], + "type": "text", + "content": " when " + }, + { + "bbox": [ + 46, + 347, + 287, + 493 + ], + "type": "inline_equation", + "content": "Y" + }, + { + "bbox": [ + 46, + 347, + 287, + 493 + ], + "type": "text", + "content": " is the downstream speech understanding task. For example, training on auditory or visual speech acts as a proxy to maximize " + }, + { + "bbox": [ + 46, + 347, + 287, + 493 + ], + "type": "inline_equation", + "content": "I(E_{a};Y)" + }, + { + "bbox": [ + 46, + 347, + 287, + 493 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 46, + 347, + 287, + 493 + ], + "type": "inline_equation", + "content": "I(E_{v};Y)" + }, + { + "bbox": [ + 46, + 347, + 287, + 493 + ], + "type": "text", + "content": ". Self-distillation with online clustering [36]. We perform distillation from the right-hand side teacher to the left-hand side student in Fig. 1(a). Specifically, as illustrated in Fig. 1(c), for each layer " + }, + { + "bbox": [ + 46, + 347, + 287, + 493 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 347, + 287, + 493 + ], + "type": "text", + "content": " of the top " + }, + { + "bbox": [ + 46, + 347, + 287, + 493 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 46, + 347, + 287, + 493 + ], + "type": "text", + "content": " Transformer layers in " + }, + { + "bbox": [ + 46, + 347, + 287, + 493 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 46, + 347, + 287, + 493 + ], + "type": "text", + "content": " with hidden dimension " + }, + { + "bbox": [ + 46, + 347, + 287, + 493 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 46, + 347, + 287, + 493 + ], + "type": "text", + "content": ", we learn a codebook " + }, + { + "bbox": [ + 46, + 347, + 287, + 493 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}^{k} = \\{\\mathbf{q}_{1}^{k},\\dots,\\mathbf{q}_{Q}^{k}\\}" + }, + { + "bbox": [ + 46, + 347, + 287, + 493 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 46, + 347, + 287, + 493 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 46, + 347, + 287, + 493 + ], + "type": "text", + "content": " codewords " + }, + { + "bbox": [ + 46, + 347, + 287, + 493 + ], + "type": "inline_equation", + "content": "\\mathbf{q}_i^k\\in \\mathbb{R}^D" + }, + { + "bbox": [ + 46, + 347, + 287, + 493 + ], + "type": "text", + "content": ", which is updated as follows: for each entry " + }, + { + "bbox": [ + 46, + 347, + 287, + 493 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 46, + 347, + 287, + 493 + ], + "type": "text", + "content": ", a set " + }, + { + "bbox": [ + 46, + 347, + 287, + 493 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{Z}}_{\\mathrm{tea}}^k" + }, + { + "bbox": [ + 46, + 347, + 287, + 493 + ], + "type": "text", + "content": " of the teacher output features " + }, + { + "bbox": [ + 46, + 347, + 287, + 493 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{z}}_{\\mathrm{tea}}^k" + }, + { + "bbox": [ + 46, + 347, + 287, + 493 + ], + "type": "text", + "content": " closest to the current representation of " + }, + { + "bbox": [ + 46, + 347, + 287, + 493 + ], + "type": "inline_equation", + "content": "q" + }, + { + "bbox": [ + 46, + 347, + 287, + 493 + ], + "type": "text", + "content": " in the codebook is found as" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 501, + 287, + 529 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 501, + 287, + 529 + ], + "spans": [ + { + "bbox": [ + 61, + 501, + 287, + 529 + ], + "type": "interline_equation", + "content": "\\tilde {\\mathbf {Z}} _ {\\mathrm {t e a}} ^ {k} = \\left\\{\\tilde {\\boldsymbol {z}} _ {\\mathrm {t e a}} ^ {k} \\Bigg | q = \\underset {j \\in \\{1, 2, \\dots , Q \\}} {\\arg \\min } | | \\tilde {\\boldsymbol {z}} _ {\\mathrm {t e a}} ^ {k} - \\mathbf {q} _ {j} ^ {k} | | _ {2} \\right\\}, \\tag {8}", + "image_path": "a3313f02ab043dc0e5cad2935b6a4e93ac50eb83c5dca42175b3d4f0bc1a301d.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 536, + 287, + 585 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 536, + 287, + 585 + ], + "spans": [ + { + "bbox": [ + 47, + 536, + 287, + 585 + ], + "type": "text", + "content": "where the set indices " + }, + { + "bbox": [ + 47, + 536, + 287, + 585 + ], + "type": "inline_equation", + "content": "q" + }, + { + "bbox": [ + 47, + 536, + 287, + 585 + ], + "type": "text", + "content": " are used as discrete pseudo-labels to train the student model. Codewords are then updated in an EMA fashion as a weighted sum of the embeddings in " + }, + { + "bbox": [ + 47, + 536, + 287, + 585 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{Z}}_{\\mathrm{tea}}^k" + }, + { + "bbox": [ + 47, + 536, + 287, + 585 + ], + "type": "text", + "content": " with decay rate " + }, + { + "bbox": [ + 47, + 536, + 287, + 585 + ], + "type": "inline_equation", + "content": "\\tau_{\\mathrm{code}}" + }, + { + "bbox": [ + 47, + 536, + 287, + 585 + ], + "type": "text", + "content": ". The associated loss function is" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 57, + 592, + 287, + 628 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 592, + 287, + 628 + ], + "spans": [ + { + "bbox": [ + 57, + 592, + 287, + 628 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {d i s t i l l}} \\left(F _ {\\text {t e a}}, F _ {\\text {s t u}}\\right) = \\sum_ {t \\in \\mathbb {M}} \\sum_ {k \\in (K - N, K ]} \\log p _ {\\psi_ {k}} \\left(q _ {t} \\mid \\boldsymbol {z} _ {t} ^ {K}\\right), \\tag {9}", + "image_path": "dd13260eb1faa2d417593f0e5e4569900a02c9930115ceec555ccd954d024925.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 630, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 630, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 630, + 287, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 630, + 287, + 713 + ], + "type": "inline_equation", + "content": "F_{\\mathrm{tea}}" + }, + { + "bbox": [ + 46, + 630, + 287, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 630, + 287, + 713 + ], + "type": "inline_equation", + "content": "F_{\\mathrm{stu}}" + }, + { + "bbox": [ + 46, + 630, + 287, + 713 + ], + "type": "text", + "content": " are teacher and student features, " + }, + { + "bbox": [ + 46, + 630, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\mathbb{M}" + }, + { + "bbox": [ + 46, + 630, + 287, + 713 + ], + "type": "text", + "content": " is the set of masked time steps, " + }, + { + "bbox": [ + 46, + 630, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\psi_{k}" + }, + { + "bbox": [ + 46, + 630, + 287, + 713 + ], + "type": "text", + "content": " is the softmax activations of linear projections for each target layer " + }, + { + "bbox": [ + 46, + 630, + 287, + 713 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 630, + 287, + 713 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 630, + 287, + 713 + ], + "type": "inline_equation", + "content": "q_{t}" + }, + { + "bbox": [ + 46, + 630, + 287, + 713 + ], + "type": "text", + "content": " is the codeword index for time step " + }, + { + "bbox": [ + 46, + 630, + 287, + 713 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 630, + 287, + 713 + ], + "type": "text", + "content": " of the teacher " + }, + { + "bbox": [ + 46, + 630, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\tilde{\\theta}" + }, + { + "bbox": [ + 46, + 630, + 287, + 713 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 630, + 287, + 713 + ], + "type": "inline_equation", + "content": "z_{t}^{K}" + }, + { + "bbox": [ + 46, + 630, + 287, + 713 + ], + "type": "text", + "content": " is the output of the " + }, + { + "bbox": [ + 46, + 630, + 287, + 713 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 46, + 630, + 287, + 713 + ], + "type": "text", + "content": "-th and final layer of the student model " + }, + { + "bbox": [ + 46, + 630, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 46, + 630, + 287, + 713 + ], + "type": "text", + "content": ". The teacher and the online clustering process do not require gradients. EMA teacher parameterization. To reduce GPU memory" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 348, + 545, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 348, + 545, + 407 + ], + "spans": [ + { + "bbox": [ + 305, + 348, + 545, + 407 + ], + "type": "text", + "content": "usage, we apply EMA updates to the Transformer encoder " + }, + { + "bbox": [ + 305, + 348, + 545, + 407 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 305, + 348, + 545, + 407 + ], + "type": "text", + "content": ", but not the earlier modality-specific encoders " + }, + { + "bbox": [ + 305, + 348, + 545, + 407 + ], + "type": "inline_equation", + "content": "\\varphi_{a}" + }, + { + "bbox": [ + 305, + 348, + 545, + 407 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 348, + 545, + 407 + ], + "type": "inline_equation", + "content": "\\varphi_{v}" + }, + { + "bbox": [ + 305, + 348, + 545, + 407 + ], + "type": "text", + "content": ", inspired by [50]. As shown in Fig. 1, the teacher model weights " + }, + { + "bbox": [ + 305, + 348, + 545, + 407 + ], + "type": "inline_equation", + "content": "\\tilde{\\theta}" + }, + { + "bbox": [ + 305, + 348, + 545, + 407 + ], + "type": "text", + "content": " are computed as an exponential moving average of the student model weights " + }, + { + "bbox": [ + 305, + 348, + 545, + 407 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 305, + 348, + 545, + 407 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 383, + 415, + 545, + 430 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 383, + 415, + 545, + 430 + ], + "spans": [ + { + "bbox": [ + 383, + 415, + 545, + 430 + ], + "type": "interline_equation", + "content": "\\tilde {\\theta} \\leftarrow \\eta \\tilde {\\theta} + (1 - \\eta) \\theta , \\tag {10}", + "image_path": "ad25f5bec9b86b1f39dc0ff7d0aca1df6c4889ed0b25518cc8961bc24900961d.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 439, + 545, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 439, + 545, + 475 + ], + "spans": [ + { + "bbox": [ + 305, + 439, + 545, + 475 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 439, + 545, + 475 + ], + "type": "inline_equation", + "content": "\\eta = \\min \\{\\eta_{\\max}, (1 + \\text{step}) / (10 + \\text{step})\\}" + }, + { + "bbox": [ + 305, + 439, + 545, + 475 + ], + "type": "text", + "content": " is a momentum parameter that slowly increases over time until a pre-determined value " + }, + { + "bbox": [ + 305, + 439, + 545, + 475 + ], + "type": "inline_equation", + "content": "\\eta_{\\max}" + }, + { + "bbox": [ + 305, + 439, + 545, + 475 + ], + "type": "text", + "content": " is reached." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 483, + 430, + 496 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 483, + 430, + 496 + ], + "spans": [ + { + "bbox": [ + 306, + 483, + 430, + 496 + ], + "type": "text", + "content": "3.4. The Learning Process" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 502, + 545, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 502, + 545, + 525 + ], + "spans": [ + { + "bbox": [ + 305, + 502, + 545, + 525 + ], + "type": "text", + "content": "We now elaborate on individual learning stages, based on the above Siamese framework and distillation loss." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 525, + 545, + 584 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 525, + 545, + 584 + ], + "spans": [ + { + "bbox": [ + 305, + 525, + 545, + 584 + ], + "type": "text", + "content": "Stage 1: Learning audio-unique information and shared information. In stage 1, we perform distillation from un-masked audio teacher model to the masked audio student model, yielding audio-unique information " + }, + { + "bbox": [ + 305, + 525, + 545, + 584 + ], + "type": "inline_equation", + "content": "U_{a}" + }, + { + "bbox": [ + 305, + 525, + 545, + 584 + ], + "type": "text", + "content": " and shared information " + }, + { + "bbox": [ + 305, + 525, + 545, + 584 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 305, + 525, + 545, + 584 + ], + "type": "text", + "content": " from audio. The distillation loss is:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 371, + 595, + 545, + 608 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 371, + 595, + 545, + 608 + ], + "spans": [ + { + "bbox": [ + 371, + 595, + 545, + 608 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {a} \\rightarrow \\mathrm {a} ^ {\\prime}} = \\mathcal {L} _ {\\text {d i s t i l l}} \\left(E _ {a}, E _ {a} ^ {\\prime}\\right). \\tag {11}", + "image_path": "6eab835547d9c634e630217d824571deeb990b252cb121a43e84a7813be50d1c.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 617, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 617, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 617, + 545, + 712 + ], + "type": "text", + "content": "Being readily learnable, audio has the potential to provide most of " + }, + { + "bbox": [ + 304, + 617, + 545, + 712 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 304, + 617, + 545, + 712 + ], + "type": "text", + "content": " for learning shared information. However, optimizing " + }, + { + "bbox": [ + 304, + 617, + 545, + 712 + ], + "type": "inline_equation", + "content": "I(E_{a};Y)" + }, + { + "bbox": [ + 304, + 617, + 545, + 712 + ], + "type": "text", + "content": " alone cannot help visual learning. Noting that the InfoNCE loss between two variables has been widely used to maximize a lower bound on the mutual information, we adopt the InfoNCE loss to learn shared information corresponding to " + }, + { + "bbox": [ + 304, + 617, + 545, + 712 + ], + "type": "inline_equation", + "content": "I(E_{a};E_{v})" + }, + { + "bbox": [ + 304, + 617, + 545, + 712 + ], + "type": "text", + "content": " to bootstrap visual learning with the shared information learned from audio." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "27072" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 168 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 168 + ], + "type": "text", + "content": "Specifically, we apply it to a fine-grained frame-wise alignment task, where temporally aligned features from the same video are treated as positives, and unaligned ones negatives [19]. Since phonemes can appear repeatedly within the same video clip, leading to false negatives, we use a negative margin following Park et al. [47], which alleviates the effect of falsely matched pairs by enforcing a looser decision boundary. The objective function involved is:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 53, + 176, + 287, + 217 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 176, + 287, + 217 + ], + "spans": [ + { + "bbox": [ + 53, + 176, + 287, + 217 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {N C E}} (H) = - \\frac {1}{n} \\sum_ {i = 1} ^ {n} \\log \\frac {e ^ {(H _ {i , i} - m) / \\tau}}{e ^ {(H _ {i , i} - m) / \\tau} + \\sum_ {i \\neq j} e ^ {H _ {i , j} / \\tau}}, \\tag {12}", + "image_path": "bc194bb9cb6cf3ef02b2070819960350ba6d08d07a932bde7d6e163e016bf084.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 217, + 289, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 217, + 289, + 300 + ], + "spans": [ + { + "bbox": [ + 46, + 217, + 289, + 300 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 217, + 289, + 300 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 46, + 217, + 289, + 300 + ], + "type": "text", + "content": " is a matrix whose " + }, + { + "bbox": [ + 46, + 217, + 289, + 300 + ], + "type": "inline_equation", + "content": "(i,j)" + }, + { + "bbox": [ + 46, + 217, + 289, + 300 + ], + "type": "text", + "content": "-th element represents the cosine similarity between " + }, + { + "bbox": [ + 46, + 217, + 289, + 300 + ], + "type": "inline_equation", + "content": "L^2" + }, + { + "bbox": [ + 46, + 217, + 289, + 300 + ], + "type": "text", + "content": "-normalized audio and visual features at time step " + }, + { + "bbox": [ + 46, + 217, + 289, + 300 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 217, + 289, + 300 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 217, + 289, + 300 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 46, + 217, + 289, + 300 + ], + "type": "text", + "content": ", respectively, " + }, + { + "bbox": [ + 46, + 217, + 289, + 300 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 46, + 217, + 289, + 300 + ], + "type": "text", + "content": " is a learnable temperature parameter, and " + }, + { + "bbox": [ + 46, + 217, + 289, + 300 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 46, + 217, + 289, + 300 + ], + "type": "text", + "content": " is a pre-determined margin. To promote the learning of shared features between the two modalities, we apply a symmetrized version of the above NCE loss as:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 306, + 290, + 338 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 306, + 290, + 338 + ], + "spans": [ + { + "bbox": [ + 47, + 306, + 290, + 338 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {a} \\leftrightarrow \\mathrm {v}} = \\frac {1}{2} \\left(\\mathcal {L} _ {\\mathrm {a} \\rightarrow \\mathrm {v}} + \\mathcal {L} _ {\\mathrm {v} \\rightarrow \\mathrm {a}}\\right) = \\frac {1}{2} \\left(\\mathcal {L} _ {\\mathrm {N C E}} \\left(S _ {a v}\\right) + \\mathcal {L} _ {\\mathrm {N C E}} \\left(S _ {v a}\\right)\\right). \\tag {13}", + "image_path": "b750855adc8ad0d0a29b8717b89ef68ab35c0ae53b73c014fe13e5213382451e.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 338, + 287, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 338, + 287, + 399 + ], + "spans": [ + { + "bbox": [ + 46, + 338, + 287, + 399 + ], + "type": "text", + "content": "Here the similarity matrices " + }, + { + "bbox": [ + 46, + 338, + 287, + 399 + ], + "type": "inline_equation", + "content": "S_{av} = \\langle F_a, F_v \\rangle / (||F_a|| \\cdot ||F_v||)" + }, + { + "bbox": [ + 46, + 338, + 287, + 399 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 338, + 287, + 399 + ], + "type": "inline_equation", + "content": "S_{va} = S_{av}^{\\top}" + }, + { + "bbox": [ + 46, + 338, + 287, + 399 + ], + "type": "text", + "content": " are computed with features " + }, + { + "bbox": [ + 46, + 338, + 287, + 399 + ], + "type": "inline_equation", + "content": "F_a" + }, + { + "bbox": [ + 46, + 338, + 287, + 399 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 338, + 287, + 399 + ], + "type": "inline_equation", + "content": "F_v" + }, + { + "bbox": [ + 46, + 338, + 287, + 399 + ], + "type": "text", + "content": " from modality-specific encoders " + }, + { + "bbox": [ + 46, + 338, + 287, + 399 + ], + "type": "inline_equation", + "content": "\\varphi_a" + }, + { + "bbox": [ + 46, + 338, + 287, + 399 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 338, + 287, + 399 + ], + "type": "inline_equation", + "content": "\\varphi_v" + }, + { + "bbox": [ + 46, + 338, + 287, + 399 + ], + "type": "text", + "content": " instead of the contextual features from the Transformer encoder " + }, + { + "bbox": [ + 46, + 338, + 287, + 399 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 46, + 338, + 287, + 399 + ], + "type": "text", + "content": " to avoid learning positional shortcuts. The total loss in stage 1 is" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 111, + 408, + 287, + 421 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 111, + 408, + 287, + 421 + ], + "spans": [ + { + "bbox": [ + 111, + 408, + 287, + 421 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {s t a g e 1}} = \\mathcal {L} _ {\\mathrm {a} \\rightarrow \\mathrm {a} ^ {\\prime}} + \\mathcal {L} _ {\\mathrm {a} \\leftrightarrow \\mathrm {v}}. \\tag {14}", + "image_path": "ec2e1d478ec9b97b49bd71214b5c46742e888e0974769c366f33f5aa23863c04.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 426, + 289, + 535 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 426, + 289, + 535 + ], + "spans": [ + { + "bbox": [ + 46, + 426, + 289, + 535 + ], + "type": "text", + "content": "Stage 2: Learning video-unique information. In stage 2, we proceed to learn video-unique information, which is now much easier than learning from scratch since it has been bootstrapped with shared knowledge from the audio in the previous stage. The audio encoder " + }, + { + "bbox": [ + 46, + 426, + 289, + 535 + ], + "type": "inline_equation", + "content": "\\varphi_{a}" + }, + { + "bbox": [ + 46, + 426, + 289, + 535 + ], + "type": "text", + "content": " does not receive any gradients in this stage and is essentially frozen, since it is expected to have already acquired the audio-specific knowledge well in the previous stage. We adopt a similar distillation loss as stage 1:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 112, + 543, + 287, + 556 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 112, + 543, + 287, + 556 + ], + "spans": [ + { + "bbox": [ + 112, + 543, + 287, + 556 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {v} \\rightarrow \\mathrm {v} ^ {\\prime}} = \\mathcal {L} _ {\\text {d i s t i l l}} \\left(E _ {v}, E _ {v} ^ {\\prime}\\right). \\tag {15}", + "image_path": "2e1882e6018a2fdca3489d91c948a35ba4ad30fa8aab897b55df9224e42bd6a9.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 560, + 287, + 632 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 560, + 287, + 632 + ], + "spans": [ + { + "bbox": [ + 46, + 560, + 287, + 632 + ], + "type": "text", + "content": "On top of the learned visual-unique information, we also perform distillation from unmasked video features to the joint features of unmasked audio-visual inputs in this stage to prepare for the learning of the final robust audio-visual representations in the next stage. For this purpose, the loss in this stage is set as:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 110, + 641, + 287, + 654 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 110, + 641, + 287, + 654 + ], + "spans": [ + { + "bbox": [ + 110, + 641, + 287, + 654 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {a v} \\rightarrow \\mathrm {v}} = \\mathcal {L} _ {\\text {d i s t i l l}} \\left(E _ {v}, E _ {a v}\\right), \\tag {16}", + "image_path": "07e7d81896900b189e23cf324c68a04bb934cee12fbb77f2338418c0f5fffcb1.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 110, + 656, + 287, + 669 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 110, + 656, + 287, + 669 + ], + "spans": [ + { + "bbox": [ + 110, + 656, + 287, + 669 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {s t a g e 2}} = \\mathcal {L} _ {\\mathrm {v} \\rightarrow \\mathrm {v} ^ {\\prime}} + \\mathcal {L} _ {\\mathrm {a v} \\rightarrow \\mathrm {v}}. \\tag {17}", + "image_path": "3f6eea897506cff6c83e3de23fb9d9f94619c45e67f6185873f8e1bbd3ca55ab.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "text", + "content": "To avoid repeated learning of only shared information between A and V again, the cross-modal loss " + }, + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{av}\\to \\mathrm{v}}" + }, + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "text", + "content": " is asymmetric, strictly one-way from AV to V, and a learnable tem" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 72, + 545, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 108 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 108 + ], + "type": "text", + "content": "perature " + }, + { + "bbox": [ + 304, + 72, + 545, + 108 + ], + "type": "inline_equation", + "content": "\\tau_{v}" + }, + { + "bbox": [ + 304, + 72, + 545, + 108 + ], + "type": "text", + "content": " Eq. (16) is introduced to the visual student outputs, i.e., we replace " + }, + { + "bbox": [ + 304, + 72, + 545, + 108 + ], + "type": "inline_equation", + "content": "\\psi_{k}" + }, + { + "bbox": [ + 304, + 72, + 545, + 108 + ], + "type": "text", + "content": " in Eq. (9) with softmax activations of temperature-scaled linear projections of " + }, + { + "bbox": [ + 304, + 72, + 545, + 108 + ], + "type": "inline_equation", + "content": "E_{v}" + }, + { + "bbox": [ + 304, + 72, + 545, + 108 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 108, + 546, + 347 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 108, + 546, + 347 + ], + "spans": [ + { + "bbox": [ + 304, + 108, + 546, + 347 + ], + "type": "text", + "content": "Stage 3: Learning synergistic information. Finally, we aim to learn synergistic audio-visual information which previous methods have not explicitly addressed. Our ultimate goal is to maximize total task-relevant multi-modal information " + }, + { + "bbox": [ + 304, + 108, + 546, + 347 + ], + "type": "inline_equation", + "content": "I(E_{av};Y)" + }, + { + "bbox": [ + 304, + 108, + 546, + 347 + ], + "type": "text", + "content": ", i.e. the sum of " + }, + { + "bbox": [ + 304, + 108, + 546, + 347 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 304, + 108, + 546, + 347 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 108, + 546, + 347 + ], + "type": "inline_equation", + "content": "U_{a}" + }, + { + "bbox": [ + 304, + 108, + 546, + 347 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 108, + 546, + 347 + ], + "type": "inline_equation", + "content": "U_{v}" + }, + { + "bbox": [ + 304, + 108, + 546, + 347 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 108, + 546, + 347 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 304, + 108, + 546, + 347 + ], + "type": "text", + "content": ", so that the representations are useful for both uni-modal and multi-modal tasks. Naively maximizing this objective without prior stages is problematic because the network tends to be biased towards learning audio-based representations (i.e. prioritizing " + }, + { + "bbox": [ + 304, + 108, + 546, + 347 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 304, + 108, + 546, + 347 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 108, + 546, + 347 + ], + "type": "inline_equation", + "content": "U_{a}" + }, + { + "bbox": [ + 304, + 108, + 546, + 347 + ], + "type": "text", + "content": ") without considering the visual modality. By learning most of " + }, + { + "bbox": [ + 304, + 108, + 546, + 347 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 304, + 108, + 546, + 347 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 108, + 546, + 347 + ], + "type": "inline_equation", + "content": "U_{a}" + }, + { + "bbox": [ + 304, + 108, + 546, + 347 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 108, + 546, + 347 + ], + "type": "inline_equation", + "content": "U_{v}" + }, + { + "bbox": [ + 304, + 108, + 546, + 347 + ], + "type": "text", + "content": " through the previous stages, our method is less susceptible to learning biased representations in the final stage. At this stage, we always mask identical positions in the audio and video inputs together so that the masked region never contains any unimodal information that can collapse or degrade training. To better recover the masked inputs, the model must cleverly leverage both modalities, which requires learning synergistic information that is not well-captured in prior stages. The loss function in stage 3 can be written as:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 341, + 355, + 545, + 369 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 341, + 355, + 545, + 369 + ], + "spans": [ + { + "bbox": [ + 341, + 355, + 545, + 369 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {s t a g e 3}} = \\mathcal {L} _ {\\mathrm {a v} \\rightarrow \\mathrm {a v} ^ {\\prime}} = \\mathcal {L} _ {\\text {d i s t i l l}} \\left(E _ {a v} ^ {\\prime}, E _ {a v}\\right). \\tag {18}", + "image_path": "b010764e938a9cffac5f7c223323c51d1789f2919fd4f271c739ac64d1c2ed68.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 372, + 545, + 419 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 372, + 545, + 419 + ], + "spans": [ + { + "bbox": [ + 304, + 372, + 545, + 419 + ], + "type": "text", + "content": "By progressively acquiring the three types of useful information, we build strong and robust audio-visual speech representations for both uni- and bi-modal speech, which we shall show in subsequent experiments." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 426, + 411, + 439 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 426, + 411, + 439 + ], + "spans": [ + { + "bbox": [ + 306, + 426, + 411, + 439 + ], + "type": "text", + "content": "3.5. Fine-tuning Stage" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 445, + 545, + 576 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 445, + 545, + 576 + ], + "spans": [ + { + "bbox": [ + 304, + 445, + 545, + 576 + ], + "type": "text", + "content": "After pre-training, we add an extra linear layer after the fusion encoder " + }, + { + "bbox": [ + 304, + 445, + 545, + 576 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 304, + 445, + 545, + 576 + ], + "type": "text", + "content": " to fine-tune the model end-to-end with the Lattice-Free Maximum Mutual Information (LF-MMI) criterion [26, 51], which is much more lightweight than seq2seq Encoder-Decoder models and allows highly efficient decoding. We adopt a \"compact\" [33] 1-state hidden Markov model (HMM) topology, using " + }, + { + "bbox": [ + 304, + 445, + 545, + 576 + ], + "type": "inline_equation", + "content": "k2^2" + }, + { + "bbox": [ + 304, + 445, + 545, + 576 + ], + "type": "text", + "content": " for differentiable Weighted Finite State Transducer (WFST)-based training. We use 624 bi-char units from text-based clustering [77] for English and tonal phone-level targets (initials and finals) for Chinese." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 306, + 586, + 387, + 600 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 586, + 387, + 600 + ], + "spans": [ + { + "bbox": [ + 306, + 586, + 387, + 600 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 306, + 605, + 368, + 617 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 605, + 368, + 617 + ], + "spans": [ + { + "bbox": [ + 306, + 605, + 368, + 617 + ], + "type": "text", + "content": "4.1. Datasets" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 624, + 545, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 624, + 545, + 696 + ], + "spans": [ + { + "bbox": [ + 304, + 624, + 545, + 696 + ], + "type": "text", + "content": "LRS2-BBC [3]. LRS2-BBC is a 224-hour transcribed audio-visual English speech dataset collected from programs of various genres broadcast by the BBC. It is challenging due to the presence of extreme poses, as well as a large and diverse vocabulary. We use the official train / validation split (see supplementary materials) for evaluation." + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 316, + 702, + 436, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 702, + 436, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 702, + 436, + 712 + ], + "type": "text", + "content": "2https://github.com/k2-fsa/k2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "27073" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 167 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 167 + ], + "type": "text", + "content": "LRS3-TED [1]. LRS3-TED is another audio-visual English speech dataset collected from TED and TEDx talks, with over 5.5k speakers and 439 hours of data. It covers diverse topics and is widely adopted for evaluating audio-visual speech representation learning methods for being the largest publicly available one of its kind. An official train / validation split is not given, so we use the widely adopted one prepared by Shi et al. [59], same as [27, 34, 80]." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 168, + 287, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 168, + 287, + 228 + ], + "spans": [ + { + "bbox": [ + 46, + 168, + 287, + 228 + ], + "type": "text", + "content": "CAS-VSR-S101. We also collect a new large-scale, in-the-wild Mandarin dataset with 101.1 hours of data. The videos are sourced from broadcast news and conversational programs in Chinese, covering a highly diverse set of topics, speakers and filming conditions (see supplementary)." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 236, + 163, + 249 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 236, + 163, + 249 + ], + "spans": [ + { + "bbox": [ + 47, + 236, + 163, + 249 + ], + "type": "text", + "content": "4.2. Experimental Setup" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 255, + 286, + 278 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 255, + 286, + 278 + ], + "spans": [ + { + "bbox": [ + 46, + 255, + 286, + 278 + ], + "type": "text", + "content": "More implementation details can be found in the supplementary materials." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 279, + 286, + 447 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 279, + 286, + 447 + ], + "spans": [ + { + "bbox": [ + 46, + 279, + 286, + 447 + ], + "type": "text", + "content": "High- and low-resource setups. On LRS3-TED we follow the widely adopted setup of Shi et al. [59], which excludes the pre-train set during low-resource fine-tuning. To validate data scaling properties, previous works used VoxCeleb2-En [59], which is a 1326-hour subset of VoxCeleb2 [18] in English. However, not only are experiments of this scale computationally expensive, but more importantly, VoxCeleb2 is no longer available from its official source as of Feb. 2023. Therefore, we create a similar scaling protocol, using LRS3-TED (433h) as a pre-training source and LRS2-BBC (223h) for fine-tuning. For low-resource fine-tuning, we only use the 28-hour train subset of LRS2-BBC, while for high-resource fine-tuning, we include the pre-train subset, similar to [17]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 447, + 286, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 447, + 286, + 555 + ], + "spans": [ + { + "bbox": [ + 46, + 447, + 286, + 555 + ], + "type": "text", + "content": "Evaluation. Decoding is performed using beam search with a beam size of 50. The common evaluation metrics for ASR, VSR and AVSR are Word Error Rate (WER) and Character Error Rate (CER) for English and Chinese, respectively. As our bi-char LF-MMI models are trained on verbatim transcriptions for ease of lexicon construction, we report WERs both without and with using Whisper-style text normalization [54] (in parentheses) to compensate for innocuous errors related to text normalization." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 563, + 250, + 574 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 563, + 250, + 574 + ], + "spans": [ + { + "bbox": [ + 47, + 563, + 250, + 574 + ], + "type": "text", + "content": "4.3. Results on LRS2-BBC and LRS3-TED" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 582, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 582, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 582, + 286, + 713 + ], + "type": "text", + "content": "Low-resource setup. Low-resource results on LRS2-BBC and LRS3-TED can be found in Tabs. 1 and 2. Our smallest BASE* encoder is modality-agnostic, only half the size of the one in [27], which uses two separate encoders to perform AVSR. On LRS3-TED, although we employ the simple LF-MMI instead of cross-entropy (CE) based seq2seq criterion for fine-tuning, our BASE model rivals or achieves SoTA in all cases. When scaling to a LARGE model, we observe significant improvements in both ASR and AVSR. Meanwhile, VSR performance remains comparable to SoTA, with a moderate " + }, + { + "bbox": [ + 46, + 582, + 286, + 713 + ], + "type": "inline_equation", + "content": "2\\%" + }, + { + "bbox": [ + 46, + 582, + 286, + 713 + ], + "type": "text", + "content": " gap. This may sug" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 72, + 545, + 107 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 107 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 107 + ], + "type": "text", + "content": "gest that with our current VGG-based visual encoder " + }, + { + "bbox": [ + 304, + 72, + 545, + 107 + ], + "type": "inline_equation", + "content": "\\varphi_v" + }, + { + "bbox": [ + 304, + 72, + 545, + 107 + ], + "type": "text", + "content": ", the benefits of scaling up the Transformer primarily manifest in ASR and AVSR." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 109, + 545, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 109, + 545, + 300 + ], + "spans": [ + { + "bbox": [ + 304, + 109, + 545, + 300 + ], + "type": "text", + "content": "High-resource setup. High-resource results on LRS2-BBC and LRS3-TED can be found in Tabs. 1 and 3. On LRS2-BBC, with a 46M model and relying on its own 223h data for pre-training and fine-tuning, we achieve " + }, + { + "bbox": [ + 304, + 109, + 545, + 300 + ], + "type": "inline_equation", + "content": "30.3\\% / 3.6\\% / 2.9\\%" + }, + { + "bbox": [ + 304, + 109, + 545, + 300 + ], + "type": "text", + "content": " WER for VSR, ASR and AVSR. Notably, these results surpass a highly competitive supervised method [40] and other self-supervised methods that incorporate additional data. This indicates that our progressive learning approach can effectively exploit visual information. Furthermore, when we introduce additional 433h unlabeled data, we achieve even better performance: " + }, + { + "bbox": [ + 304, + 109, + 545, + 300 + ], + "type": "inline_equation", + "content": "29.8\\% / 3.0\\% / 2.4\\%" + }, + { + "bbox": [ + 304, + 109, + 545, + 300 + ], + "type": "text", + "content": " WER, which outperforms other methods that employ the same amount of data. When we switch to the LARGE model, performance rivals SoTA methods that employ 4x data (1759h) for pre-training. Similarly, we achieve or rival SoTA performance on LRS3-TED." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 309, + 454, + 320 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 309, + 454, + 320 + ], + "spans": [ + { + "bbox": [ + 305, + 309, + 454, + 320 + ], + "type": "text", + "content": "4.4. Results on CAS-VSR-S101" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 327, + 545, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 327, + 545, + 460 + ], + "spans": [ + { + "bbox": [ + 304, + 327, + 545, + 460 + ], + "type": "text", + "content": "We also evaluate on our newly collected CAS-VSR-S101 dataset, which is smaller than both of the two previous datasets and in an even more ambiguous language (Chinese has more homophonous syllables and words than English). As shown in Tab. 4, " + }, + { + "bbox": [ + 304, + 327, + 545, + 460 + ], + "type": "inline_equation", + "content": "\\mathrm{ES}^3" + }, + { + "bbox": [ + 304, + 327, + 545, + 460 + ], + "type": "text", + "content": " still works out-of-the-box under this challenging setup, and yields noticeable performance gains over a strong fully supervised baseline [39]. Interestingly, our fully supervised baseline model beats [39] in ASR and AVSR but slightly lags behind in VSR, again hinting that our strategy can further benefit from more complex visual modeling architectures." + } + ] + } + ], + "index": 11 + }, + { + "type": "table", + "bbox": [ + 308, + 521, + 545, + 580 + ], + "blocks": [ + { + "bbox": [ + 304, + 463, + 545, + 518 + ], + "lines": [ + { + "bbox": [ + 304, + 463, + 545, + 518 + ], + "spans": [ + { + "bbox": [ + 304, + 463, + 545, + 518 + ], + "type": "text", + "content": "Table 4. Results on CAS-VSR-S101. Here we use a BASE* model for the fully-supervised baseline and " + }, + { + "bbox": [ + 304, + 463, + 545, + 518 + ], + "type": "inline_equation", + "content": "\\mathrm{ES}^3" + }, + { + "bbox": [ + 304, + 463, + 545, + 518 + ], + "type": "text", + "content": ". The self-supervised models are pre-trained on training data and then finetuned for 50 epochs with the labels. The numbers before and after slashes correspond to CERs on the validation and test sets." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 308, + 521, + 545, + 580 + ], + "lines": [ + { + "bbox": [ + 308, + 521, + 545, + 580 + ], + "spans": [ + { + "bbox": [ + 308, + 521, + 545, + 580 + ], + "type": "table", + "html": "
MethodsCriterionEncoder Size#EpochsVSRASRAVSR
Supervised
BaselineLF-MMI46M10060.0 / 57.415.4 / 12.614.6 / 12.1
E2E Conformer [39]CTC+CE79M10058.6 / 56.516.6 / 13.016.0 / 12.6
Self-supervised
ES3 Stage 161.2 / 59.014.7 / 11.9-
ES3 Stage 2LF-MMI46M5059.1 / 56.7-13.9 / 11.4
ES3 Stage 358.1 / 55.614.4 / 11.613.4 / 11.0
", + "image_path": "d657ac7401c12c04bb6620d820470a30a41816cc760ec936a66a4bb09363df4e.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_body" + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 586, + 440, + 599 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 586, + 440, + 599 + ], + "spans": [ + { + "bbox": [ + 306, + 586, + 440, + 599 + ], + "type": "text", + "content": "4.5. Analysis and Discussion" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": "Representation quality probing. In Tab. 5, we demonstrate the effect of each stage on LRS2-BBC. ASR performance between stages 1 and 3 are close, indicating that stage 1 indeed acquires most of audio-unique information. Comparing the VSR results between stages 1 and 2 reveals a significant gap, suggesting that stage 2 also effectively learns visual-unique information, hence the minimal differences between the final VSR result and those of stage 2. Moreover, comparing the results of stage 1-2 and stage" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "27074" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 99, + 96, + 496, + 312 + ], + "blocks": [ + { + "bbox": [ + 46, + 70, + 545, + 93 + ], + "lines": [ + { + "bbox": [ + 46, + 70, + 545, + 93 + ], + "spans": [ + { + "bbox": [ + 46, + 70, + 545, + 93 + ], + "type": "text", + "content": "Table 1. Results on LRS2-BBC. We pre-train a BASE* model with 223h unlabeled data from LRS2-BBC, as well as BASE and LARGE models with 433h unlabeled data (LRS3-TED) to demonstrate scaling properties. *: external language models." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 99, + 96, + 496, + 312 + ], + "lines": [ + { + "bbox": [ + 99, + 96, + 496, + 312 + ], + "spans": [ + { + "bbox": [ + 99, + 96, + 496, + 312 + ], + "type": "table", + "html": "
MethodsUnlabeled AV dataLabeled DataEncoder SizeCriterionVSRASRAVSR
Supervised
Yu et al. [75]-1519h-LF-MMI148.96.75.9
Ma et al. [39]*-380h79MCTC+CE37.93.93.7
Ma et al. [41]*-818h186MCTC+CE27.92.6-
Semi-supervised
Afouras et al. [2]*777h223h2-CTC51.3--
Ma et al. [41]*2630h818h186MCTC+CE14.61.51.5
Self-supervised (BASE and BASE* models)
AV-HuBERT [59, 60]1759h223h103MCE31.23-3.63
VATLM [80]1759h4223h107MCE30.6-2.9
RAVeN [27]433h223h97MCTC+CE32.13.9-
Pan et al. [45]-5380h399MCTC+CE43.22.72.6
ES3(ours)223h28h46M40.2 (39.1)6.0 (5.0)5.7 (4.9)
433h28h102MLF-MMI39.3 (38.2)5.5 (4.8)5.1 (4.1)
223h223h46M31.4 (30.3)4.3 (3.6)3.8 (2.9)
433h223h102M30.7 (29.8)3.4 (3.0)3.2 (2.4)
Self-supervised (LARGE models)
ES3(ours)433h28h317MLF-MMI36.4 (35.4)5.2 (4.4)4.7 (4.0)
433h223h317M26.7 (25.8)3.1 (2.5)3.1 (2.5)
AV-HuBERT [59, 60]1759h223h325MCE25.53-2.53
VATLM [80]1759h4223h332MCE24.3-2.3
RAVeN [27]1759h223h671MCTC+CE23.22.5-
", + "image_path": "927360e44f344c6270da4c7c7364195099ad0468b9eeae0f16e16f1fb37ff05a.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 53, + 313, + 469, + 323 + ], + "lines": [ + { + "bbox": [ + 53, + 313, + 469, + 323 + ], + "spans": [ + { + "bbox": [ + 53, + 313, + 469, + 323 + ], + "type": "inline_equation", + "content": "{}^{1}" + }, + { + "bbox": [ + 53, + 313, + 469, + 323 + ], + "type": "text", + "content": " Not end-to-end; requires a GMM-HMM alignment stage. " + }, + { + "bbox": [ + 53, + 313, + 469, + 323 + ], + "type": "inline_equation", + "content": "{}^{2}" + }, + { + "bbox": [ + 53, + 313, + 469, + 323 + ], + "type": "text", + "content": " Uses an additional ASR model trained on LibriSpeech (960h)." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_footnote" + }, + { + "bbox": [ + 53, + 323, + 398, + 332 + ], + "lines": [ + { + "bbox": [ + 53, + 323, + 398, + 332 + ], + "spans": [ + { + "bbox": [ + 53, + 323, + 398, + 332 + ], + "type": "text", + "content": "3 Reproduced by Zhu et al. [80]. 4 Uses additional 3846h audio, 452h audio-text and 600M text data." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_footnote" + }, + { + "bbox": [ + 54, + 332, + 267, + 342 + ], + "lines": [ + { + "bbox": [ + 54, + 332, + 267, + 342 + ], + "spans": [ + { + "bbox": [ + 54, + 332, + 267, + 342 + ], + "type": "text", + "content": "5 Uses additional 60000h audio data and 1.28M unlabeled images." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 99, + 360, + 496, + 473 + ], + "blocks": [ + { + "bbox": [ + 88, + 345, + 503, + 355 + ], + "lines": [ + { + "bbox": [ + 88, + 345, + 503, + 355 + ], + "spans": [ + { + "bbox": [ + 88, + 345, + 503, + 355 + ], + "type": "text", + "content": "Table 2. Low-resource results on LRS3-TED. We pre-train a BASE and LARGE model with 433h unlabeled data." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 99, + 360, + 496, + 473 + ], + "lines": [ + { + "bbox": [ + 99, + 360, + 496, + 473 + ], + "spans": [ + { + "bbox": [ + 99, + 360, + 496, + 473 + ], + "type": "table", + "html": "
MethodsUnlabeled AV dataLabeled DataEncoder SizeCriterionVSRASRAVSR
Self-supervised (BASE models)
AV-HuBERT [59, 60]433h30h103MCE51.84.9\\( {4.7}^{1} \\)
VATLM [80]\\( {433}{\\mathrm{\\;h}}^{2} \\)30h107MCE48.0-3.6
RAVEn [27]433h30h97MCTC+CE47.04.7-
\\( {\\mathrm{{AV2vec}}}^{\\left\\lbrack {76}\\right\\rbrack }{}^{3} \\)433h30h103MCE45.0-5.8
AV-data2vec [34]433h30h103MCE45.24.44.2
\\( {\\mathbf{{ES}}}^{3} \\) (ours)433h30h102MLF-MMI45.5 (44.7)3.9 (3.3)3.6 (3.0)
Self-supervised (LARGE models)
AV-HuBERT433h30h325MCE44.84.5\\( {4.2}^{1} \\)
AV-data2vec433h30h325MCE40.53.73.4
\\( {\\mathbf{{ES}}}^{3} \\) (ours)433h30h317MLF-MMI43.5 (42.5)3.8 (2.9)2.9 (2.3)
", + "image_path": "2dcdb6385a0d37b67d50823791dac44d7c521cec9ae185d63940dae40105ddae.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 53, + 475, + 400, + 484 + ], + "lines": [ + { + "bbox": [ + 53, + 475, + 400, + 484 + ], + "spans": [ + { + "bbox": [ + 53, + 475, + 400, + 484 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 53, + 475, + 400, + 484 + ], + "type": "text", + "content": " Reproduced by Lian et al. [34]. " + }, + { + "bbox": [ + 53, + 475, + 400, + 484 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 53, + 475, + 400, + 484 + ], + "type": "text", + "content": " Uses additional 3846h audio, 452h audio-text and 600M text data." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_footnote" + }, + { + "bbox": [ + 53, + 484, + 485, + 495 + ], + "lines": [ + { + "bbox": [ + 53, + 484, + 485, + 495 + ], + "spans": [ + { + "bbox": [ + 53, + 484, + 485, + 495 + ], + "type": "text", + "content": "3 Zhang et al. [76] inject noise during pre-training and obtain better fine-tuning results even with its base model AV-HuBERT " + }, + { + "bbox": [ + 53, + 484, + 485, + 495 + ], + "type": "inline_equation", + "content": "(47.1\\%)" + }, + { + "bbox": [ + 53, + 484, + 485, + 495 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 497, + 286, + 570 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 497, + 286, + 570 + ], + "spans": [ + { + "bbox": [ + 46, + 497, + 286, + 570 + ], + "type": "text", + "content": "3 demonstrates that the joint information learning in stage 3, which also incorporates synergistic information, benefits both uni-modal (ASR & VSR) and bi-modal (AVSR) tasks. Similar trends can also be observed on the more challenging Mandarin dataset CAS-VSR-S101 in Tab. 4, further supporting the motivation of our proposed strategy." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 571, + 286, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 571, + 286, + 604 + ], + "spans": [ + { + "bbox": [ + 46, + 571, + 286, + 604 + ], + "type": "text", + "content": "Table 5. Representation quality probing on LRS2-BBC. Results are reported on the validation set using the BASE* model under the low-resource setup for each pre-training stage." + } + ] + } + ], + "index": 12 + }, + { + "type": "table", + "bbox": [ + 67, + 608, + 269, + 649 + ], + "blocks": [ + { + "bbox": [ + 67, + 608, + 269, + 649 + ], + "lines": [ + { + "bbox": [ + 67, + 608, + 269, + 649 + ], + "spans": [ + { + "bbox": [ + 67, + 608, + 269, + 649 + ], + "type": "table", + "html": "
Pre-training StageVSRASRAVSR
Stage 158.3 (57.3)11.5 (10.1)-
Stage 248.7 (47.3)-12.9 (11.5)
Stage 3 (Full ES³)48.1 (46.9)11.4 (9.8)10.1 (8.6)
", + "image_path": "4b2ca0b3d50677d50b9757e16e0ca0bb8b8248f847b4005e729ff730ba474000.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_body" + } + ], + "index": 13 + }, + { + "bbox": [ + 46, + 654, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 654, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 654, + 286, + 713 + ], + "type": "text", + "content": "Effect of automatically balancing shared and unique information learning. It is known that in a hardness-aware loss function, the temperature rules the strength of penalties on hard negative samples [72]. We track the automatically learned temperature parameter " + }, + { + "bbox": [ + 46, + 654, + 286, + 713 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 46, + 654, + 286, + 713 + ], + "type": "text", + "content": " used for cross-modal" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 497, + 545, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 497, + 545, + 629 + ], + "spans": [ + { + "bbox": [ + 304, + 497, + 545, + 629 + ], + "type": "text", + "content": "learning objectives " + }, + { + "bbox": [ + 304, + 497, + 545, + 629 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{a}\\leftrightarrow \\mathrm{v}}" + }, + { + "bbox": [ + 304, + 497, + 545, + 629 + ], + "type": "text", + "content": " in Eq. (12) and " + }, + { + "bbox": [ + 304, + 497, + 545, + 629 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{av}\\rightarrow \\mathrm{v}}" + }, + { + "bbox": [ + 304, + 497, + 545, + 629 + ], + "type": "text", + "content": " in Eq. (16) when pre-training on LRS2-BBC, as shown in Fig. 2. In stage 1, " + }, + { + "bbox": [ + 304, + 497, + 545, + 629 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 304, + 497, + 545, + 629 + ], + "type": "text", + "content": " continues to fall until it reaches " + }, + { + "bbox": [ + 304, + 497, + 545, + 629 + ], + "type": "inline_equation", + "content": "\\tau_{\\mathrm{min}}" + }, + { + "bbox": [ + 304, + 497, + 545, + 629 + ], + "type": "text", + "content": ", while in stage 2 it briefly drops and then rises throughout training. The trend is in line with our intuition: in stage 1, audio information is easily learnable, so the model can fully exploit shared information by lowering " + }, + { + "bbox": [ + 304, + 497, + 545, + 629 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 304, + 497, + 545, + 629 + ], + "type": "text", + "content": " without sacrificing audio-unique information learning; in stage 2, after learning some residual shared information, an increasing " + }, + { + "bbox": [ + 304, + 497, + 545, + 629 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 304, + 497, + 545, + 629 + ], + "type": "text", + "content": " progressively relaxes " + }, + { + "bbox": [ + 304, + 497, + 545, + 629 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{av}\\rightarrow \\mathrm{v}}" + }, + { + "bbox": [ + 304, + 497, + 545, + 629 + ], + "type": "text", + "content": " to yield more freedom to " + }, + { + "bbox": [ + 304, + 497, + 545, + 629 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{v}\\rightarrow \\mathrm{v}'}" + }, + { + "bbox": [ + 304, + 497, + 545, + 629 + ], + "type": "text", + "content": ", i.e. video-unique information learning." + } + ] + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 343, + 638, + 506, + 716 + ], + "blocks": [ + { + "bbox": [ + 343, + 638, + 506, + 716 + ], + "lines": [ + { + "bbox": [ + 343, + 638, + 506, + 716 + ], + "spans": [ + { + "bbox": [ + 343, + 638, + 506, + 716 + ], + "type": "image", + "image_path": "dc65de78e349c393d51fed49ea503110b75c3317341ac1413d53fd9f30ec9c74.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 306, + 717, + 545, + 728 + ], + "lines": [ + { + "bbox": [ + 306, + 717, + 545, + 728 + ], + "spans": [ + { + "bbox": [ + 306, + 717, + 545, + 728 + ], + "type": "text", + "content": "Figure 2. The temperature parameter paces learning automatically." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "27075" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 86, + 97, + 509, + 294 + ], + "blocks": [ + { + "bbox": [ + 46, + 71, + 545, + 94 + ], + "lines": [ + { + "bbox": [ + 46, + 71, + 545, + 94 + ], + "spans": [ + { + "bbox": [ + 46, + 71, + 545, + 94 + ], + "type": "text", + "content": "Table 3. High-resource results on LRS3-TED. We fine-tune on 433h labeled data of LRS3. †: uses test-time augmentation. *: uses external language models. ‡: noise injection during pre-training. We list more methods for comparison in the supplementary material." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 86, + 97, + 509, + 294 + ], + "lines": [ + { + "bbox": [ + 86, + 97, + 509, + 294 + ], + "spans": [ + { + "bbox": [ + 86, + 97, + 509, + 294 + ], + "type": "table", + "html": "
MethodsYearUnlabeled AV dataLabeled DataBackboneEncoder SizeCriterionVSRASRAVSR
Supervised
Ma et al. [39]*2021-595hConformer79MCTC+CE43.32.32.3
Prajwal et al. [52]†*2022-698hTransformer32MCE40.6--
Ma et al. [40]*2022-818hConformer79MCTC+CE34.7--
Semi-Supervised
Shillingford et al. [62]*2019-3886hRNN-CTC55.1--
Makino et al. [42]2019-31khRNN43MTransducer33.64.84.5
Afouras et al. [2]*2020344h433h1Jasper (CNN)-CTC59.8--
Serdyuk et al. [57]2021-90khTransformer-Transducer25.9-2.3
Ma et al. [40]*2022641h818hConformer79MCTC+CE31.5--
Serdyuk et al. [58]2022-90khConformer-Transducer17.01.61.6
Ma et al. [41]*20232630h818hConformer186MCTC+CE19.11.00.9
Chang et al. [15]2023-100khConformer98MTransducer12.8-0.9
Self-supervised (Base Models)
AV-HuBERT [59, 60]2022433h433hTransformer103MCE44.03.02.82
AV2vec [76]‡2023433h433hTransformer103MCE39.9-2.6
RAVEn [27]2023433h433hTransformer97MCTC+CE39.12.2-
ES3 (ours)2023433h433hTransformer102MLF-MMI40.3 (39.2)2.9 (2.4)2.5 (2.0)
AV-data2vec [34]2023433h433hTransformer103MCE39.02.01.8
Self-supervised (Large Models)
AV-HuBERT [59, 60]2022433h433hTransformer325MCE41.62.72.52
AV-data2vec [34]2023433h433hTransformer325MCE37.41.91.7
ES3 (ours)2023433h433hTransformer317MLF-MMI37.1 (36.7)2.8 (2.1)2.1 (1.7)
", + "image_path": "7e5f9b935296553ba314447ad2237026f97cb39e4de41d77568a11274c9bd0ea.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 53, + 294, + 259, + 304 + ], + "lines": [ + { + "bbox": [ + 53, + 294, + 259, + 304 + ], + "spans": [ + { + "bbox": [ + 53, + 294, + 259, + 304 + ], + "type": "text", + "content": "1 Uses an additional ASR model trained on LibriSpeech (960h)." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_footnote" + }, + { + "bbox": [ + 279, + 294, + 386, + 304 + ], + "lines": [ + { + "bbox": [ + 279, + 294, + 386, + 304 + ], + "spans": [ + { + "bbox": [ + 279, + 294, + 386, + 304 + ], + "type": "text", + "content": "2 Reproduced by Lian et al. [34]." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 308, + 287, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 308, + 287, + 403 + ], + "spans": [ + { + "bbox": [ + 46, + 308, + 287, + 403 + ], + "type": "text", + "content": "Visualization of our learned psuedo-label sequences. We randomly select a sample from the LRS3-TED validation set and visualize the pseudo-label sequences, i.e. code indices " + }, + { + "bbox": [ + 46, + 308, + 287, + 403 + ], + "type": "inline_equation", + "content": "q" + }, + { + "bbox": [ + 46, + 308, + 287, + 403 + ], + "type": "text", + "content": " generated by the audio, visual and audio-visual EMA teacher " + }, + { + "bbox": [ + 46, + 308, + 287, + 403 + ], + "type": "inline_equation", + "content": "\\tilde{\\theta}" + }, + { + "bbox": [ + 46, + 308, + 287, + 403 + ], + "type": "text", + "content": " in each stage. Ground truth phones for each audio frame (10ms) are derived by forced alignment [79]. Fig. 3 shows that our learned codebook yields high-quality pseudo-labels that strongly correlate with phonemic units." + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 48, + 408, + 283, + 502 + ], + "blocks": [ + { + "bbox": [ + 48, + 408, + 283, + 502 + ], + "lines": [ + { + "bbox": [ + 48, + 408, + 283, + 502 + ], + "spans": [ + { + "bbox": [ + 48, + 408, + 283, + 502 + ], + "type": "image", + "image_path": "5e00eb854a35c01a20449a847b6613fbe7a3476ae06be938215eb0055ab4b0f6.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 502, + 287, + 568 + ], + "lines": [ + { + "bbox": [ + 46, + 502, + 287, + 568 + ], + "spans": [ + { + "bbox": [ + 46, + 502, + 287, + 568 + ], + "type": "text", + "content": "Figure 3. Pseudo-label sequences from each stage. Example of pseudo-labels generated from the last Transformer layer for the utterance \"so what does that mean\" by the audio, visual, and audio-visual teachers in each stage (BASE model). At the bottom we show the spectrogram for reference. Please note that our pseudolabels are derived at 40ms intervals, limited by video frame rate." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "type": "text", + "content": "Training efficiency comparison with SoTA methods. Learning audio-visual speech representations has been extremely compute-intensive. Here we briefly compare " + }, + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\mathrm{ES}^3" + }, + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "type": "text", + "content": " with prominent SoTA methods. Pre-training a BASE/LARGE model on LRS3-TED using our " + }, + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\mathrm{ES}^3" + }, + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "type": "text", + "content": " for 100 epochs per stage only takes about 6/8.3 days in total with just 8 NVIDIA A40 GPUs. In contrast, AV-data2vec is trained for 1000k steps using 64 NVIDIA V100 GPUs for 4-5/6-7 days using BASE/LARGE with a batch size of 1280/2560s, or equivalently, 821/1642 epochs. They also experiment extensively with hyperparameters such as modality schedulers, which further increases the cost of" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 308, + 547, + 451 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 308, + 547, + 451 + ], + "spans": [ + { + "bbox": [ + 304, + 308, + 547, + 451 + ], + "type": "text", + "content": "training. RAVEn performs an equivalent of 6.25/50 days of training on 8 A100s with BASE*/LARGE. It uses separate encoders for audio and video, so the resulting model cannot perform audio-visual fusion out-of-the-box. This also requires keeping track of four sets of parameters, which is very memory-intensive. By focusing on designing a clever learning strategy, we circumvent such manual tuning altogether and greatly reduce training costs. Finally, we use the simple LF-MMI criterion, and expect better performance from tuning beam size, including a larger external language model (see supplementary), or switching to the common heavy Attention-based Encoder-Decoder model." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 456, + 384, + 468 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 456, + 384, + 468 + ], + "spans": [ + { + "bbox": [ + 306, + 456, + 384, + 468 + ], + "type": "text", + "content": "5. Conclusions" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 472, + 545, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 472, + 545, + 592 + ], + "spans": [ + { + "bbox": [ + 304, + 472, + 545, + 592 + ], + "type": "text", + "content": "We present a novel strategy, " + }, + { + "bbox": [ + 304, + 472, + 545, + 592 + ], + "type": "inline_equation", + "content": "\\mathrm{ES}^3" + }, + { + "bbox": [ + 304, + 472, + 545, + 592 + ], + "type": "text", + "content": ", for self-supervised learning of audio-visual speech representations. Unlike existing methods that primarily focus on learning information shared between audio and video, we take an \"evolving\" approach that also captures unique and synergistic information. Our experimental results demonstrate that " + }, + { + "bbox": [ + 304, + 472, + 545, + 592 + ], + "type": "inline_equation", + "content": "\\mathrm{ES}^3" + }, + { + "bbox": [ + 304, + 472, + 545, + 592 + ], + "type": "text", + "content": " indeed progressively learns unique, shared and synergistic information. Moreover, in terms of downstream performance, it rivals SoTA methods in terms of ASR, VSR and AVSR accuracy without burdensome manual balancing efforts." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 592, + 546, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 592, + 546, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 592, + 546, + 714 + ], + "type": "text", + "content": "Acknowledgments. This work is partially supported by the National Natural Science Foundation of China (No. 62276247, 62076250). We are grateful to the ICT computing platform, the Nanjing Institute of InforSuperBahn, and Capitalonline Data Service Co., Ltd for providing additional compute resources. We also thank Chenhao Wang, Mingmin Yang and Jingyun Xiao for helping to prepare CAS-VSR-S101, Bingquan Xia and Songtao Luo for help with Conformer and seq2seq baselines, Bingquan Xia and Feixiang Wang for help during the rebuttal phase, and Feixiang Wang and Zhaoxin Yuan for an early version of Fig. 1." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "27076" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 95, + 106, + 108 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 95, + 106, + 108 + ], + "spans": [ + { + "bbox": [ + 48, + 95, + 106, + 108 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 115, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 53, + 115, + 287, + 148 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 115, + 287, + 148 + ], + "spans": [ + { + "bbox": [ + 53, + 115, + 287, + 148 + ], + "type": "text", + "content": "[1] Triantafyllos Afouras, Joon Son Chung, and Andrew Zisserman. LRS3-TED: a large-scale dataset for visual speech recognition. CoRR, abs/1809.00496, 2018. 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 149, + 288, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 149, + 288, + 194 + ], + "spans": [ + { + "bbox": [ + 53, + 149, + 288, + 194 + ], + "type": "text", + "content": "[2] Triantafyllos Afouras, Joon Son Chung, and Andrew Zisserman. ASR is all you need: Cross-modal distillation for lip reading. In ICASSP, pages 2143-2147. IEEE, 2020. 2, 7, 8, 4, 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 195, + 287, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 195, + 287, + 239 + ], + "spans": [ + { + "bbox": [ + 53, + 195, + 287, + 239 + ], + "type": "text", + "content": "[3] Triantafyllos Afouras, Joon Son Chung, Andrew W. Senior, Oriol Vinyals, and Andrew Zisserman. Deep audio-visual speech recognition. IEEE Trans. Pattern Anal. Mach. Intell., 44(12):8717-8727, 2022. 1, 5, 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 240, + 287, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 240, + 287, + 293 + ], + "spans": [ + { + "bbox": [ + 53, + 240, + 287, + 293 + ], + "type": "text", + "content": "[4] Jean-Baptiste Alayrac, Adrià Recasens, Rosalia Schneider, Relja Arandjelovic, Jason Ramapuram, Jeffrey De Fauw, Lucas Smaira, Sander Dieleman, and Andrew Zisserman. Self-supervised multimodal versatile networks. In NeurlPS, 2020. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 296, + 287, + 338 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 296, + 287, + 338 + ], + "spans": [ + { + "bbox": [ + 53, + 296, + 287, + 338 + ], + "type": "text", + "content": "[5] Humam Alwassel, Dhruv Mahajan, Bruno Korbar, Lorenzo Torresani, Bernard Ghanem, and Du Tran. Self-supervised learning by cross-modal audio-video clustering. In NeurIPS, 2020. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 341, + 287, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 341, + 287, + 373 + ], + "spans": [ + { + "bbox": [ + 53, + 341, + 287, + 373 + ], + "type": "text", + "content": "[6] Relja Arandjelovic and Andrew Zisserman. Look, listen and learn. In ICCV, pages 609-617. IEEE Computer Society, 2017. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 375, + 287, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 375, + 287, + 397 + ], + "spans": [ + { + "bbox": [ + 53, + 375, + 287, + 397 + ], + "type": "text", + "content": "[7] Relja Arandjelovic and Andrew Zisserman. Objects that sound. In ECCV(1), pages 451-466. Springer, 2018. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 399, + 287, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 399, + 287, + 441 + ], + "spans": [ + { + "bbox": [ + 53, + 399, + 287, + 441 + ], + "type": "text", + "content": "[8] Yuki Markus Asano, Mandela Patrick, Christian Rupprecht, and Andrea Vedaldi. Labelling unlabelled videos from scratch with multi-modal self-supervision. In NeurIPS, 2020. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 443, + 287, + 476 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 443, + 287, + 476 + ], + "spans": [ + { + "bbox": [ + 53, + 443, + 287, + 476 + ], + "type": "text", + "content": "[9] Alexei Baevski, Michael Auli, and Abdelrahman Mohamed. Effectiveness of self-supervised pre-training for speech recognition. CoRR, abs/1911.03912, 2019. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 478, + 287, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 478, + 287, + 510 + ], + "spans": [ + { + "bbox": [ + 48, + 478, + 287, + 510 + ], + "type": "text", + "content": "[10] Alexei Baevski, Yuhao Zhou, Abdelrahman Mohamed, and Michael Auli. wav2vec 2.0: A framework for self-supervised learning of speech representations. In NeurIPS, 2020. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 511, + 287, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 511, + 287, + 555 + ], + "spans": [ + { + "bbox": [ + 48, + 511, + 287, + 555 + ], + "type": "text", + "content": "[11] Alexei Baevski, Wei-Ning Hsu, Qiantong Xu, Arun Babu, Jiatao Gu, and Michael Auli. data2vec: A general framework for self-supervised learning in speech, vision and language. In ICML, pages 1298-1312. PMLR, 2022. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 556, + 287, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 556, + 287, + 600 + ], + "spans": [ + { + "bbox": [ + 48, + 556, + 287, + 600 + ], + "type": "text", + "content": "[12] Alexei Baevski, Arun Babu, Wei-Ning Hsu, and Michael Auli. Efficient self-supervised learning with contextualized target representations for vision, speech and language. In ICML, pages 1416-1429. PMLR, 2023. 1, 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 601, + 287, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 601, + 287, + 645 + ], + "spans": [ + { + "bbox": [ + 48, + 601, + 287, + 645 + ], + "type": "text", + "content": "[13] Adrian Bulat and Georgios Tzimiropoulos. How far are we from solving the 2D & 3D face alignment problem? (and a dataset of 230,000 3D facial landmarks). In ICCV, pages 1021-1030. IEEE Computer Society, 2017. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 647, + 287, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 647, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 647, + 287, + 689 + ], + "type": "text", + "content": "[14] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision Transformers. In ICCV, pages 9630-9640. IEEE, 2021. 1" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "type": "text", + "content": "[15] Oscar Chang, Hank Liao, Dmitriy Serdyuk, Ankit Shahy, and Olivier Siohan. Conformer is all you need for visual" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 32, + "blocks": [ + { + "bbox": [ + 327, + 73, + 545, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 545, + 94 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 545, + 94 + ], + "type": "text", + "content": "speech recognition. In ICASSP, pages 10136-10140. IEEE, 2024. 1, 8, 5" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 95, + 545, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 95, + 545, + 171 + ], + "spans": [ + { + "bbox": [ + 308, + 95, + 545, + 171 + ], + "type": "text", + "content": "[16] Brian Chen, Andrew Rouditchenko, Kevin Duarte, Hilde Kuehne, Samuel Thomas, Angie W. Boggust, Rameswar Panda, Brian Kingsbury, Rogério Feris, David Harwath, James R. Glass, Michael Picheny, and Shih-Fu Chang. Multimodal clustering networks for self-supervised learning from unlabeled videos. In ICCV, pages 7992-8001. IEEE, 2021. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 172, + 545, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 172, + 545, + 226 + ], + "spans": [ + { + "bbox": [ + 308, + 172, + 545, + 226 + ], + "type": "text", + "content": "[17] Xize Cheng, Tao Jin, Linjun Li, Wang Lin, Xinyu Duan, and Zhou Zhao. OpenSR: Open-modality speech recognition via maintaining multi-modality alignment. In ACL (1), pages 6592-6607. Association for Computational Linguistics, 2023. 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 228, + 545, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 228, + 545, + 259 + ], + "spans": [ + { + "bbox": [ + 308, + 228, + 545, + 259 + ], + "type": "text", + "content": "[18] Joon Son Chung, Arsha Nagrani, and Andrew Zisserman. VoxCeleb2: Deep speaker recognition. In INTERSPEECH, pages 1086-1090. ISCA, 2018. 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 261, + 545, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 261, + 545, + 303 + ], + "spans": [ + { + "bbox": [ + 308, + 261, + 545, + 303 + ], + "type": "text", + "content": "[19] Soo-Whan Chung, Joon Son Chung, and Hong-Goo Kang. Perfect match: Improved cross-modal embeddings for audiovisual synchronisation. In ICASSP, pages 3965–3969. IEEE, 2019. 5" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 304, + 545, + 336 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 304, + 545, + 336 + ], + "spans": [ + { + "bbox": [ + 308, + 304, + 545, + 336 + ], + "type": "text", + "content": "[20] Tri Dao. FlashAttention-2: Faster attention with better parallelism and work partitioning. CoRR, abs/2307.08691, 2023. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 338, + 545, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 338, + 545, + 370 + ], + "spans": [ + { + "bbox": [ + 308, + 338, + 545, + 370 + ], + "type": "text", + "content": "[21] Tri Dao, Daniel Y. Fu, Stefano Ermon, Atri Rudra, and Christopher Ré. FlashAttention: Fast and memory-efficient exact attention with IO-awareness. In NeurIPS, 2022. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 372, + 545, + 425 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 372, + 545, + 425 + ], + "spans": [ + { + "bbox": [ + 308, + 372, + 545, + 425 + ], + "type": "text", + "content": "[22] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. BERT: Pre-training of deep bidirectional transformers for language understanding. In NAACL-HLT (1), pages 4171–4186. Association for Computational Linguistics, 2019. 1, 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 426, + 545, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 426, + 545, + 459 + ], + "spans": [ + { + "bbox": [ + 308, + 426, + 545, + 459 + ], + "type": "text", + "content": "[23] Angela Fan, Edouard Grave, and Armand Joulin. Reducing Transformer depth on demand with structured dropout. In ICLR. OpenReview.net, 2020. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 460, + 545, + 492 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 460, + 545, + 492 + ], + "spans": [ + { + "bbox": [ + 308, + 460, + 545, + 492 + ], + "type": "text", + "content": "[24] Dalu Feng, Shuang Yang, and Shiguang Shan. An efficient software for building lip reading models without pains. In ICME Workshops, pages 1-2. IEEE, 2021. 1" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 493, + 545, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 493, + 545, + 536 + ], + "spans": [ + { + "bbox": [ + 308, + 493, + 545, + 536 + ], + "type": "text", + "content": "[25] Yuan Gong, Andrew Rouditchenko, Alexander H. Liu, David Harwath, Leonid Karlinsky, Hilde Kuehne, and James R. Glass. Contrastive audio-visual masked autoencoder. In ICLR. OpenReview.net, 2023. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 537, + 545, + 569 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 537, + 545, + 569 + ], + "spans": [ + { + "bbox": [ + 308, + 537, + 545, + 569 + ], + "type": "text", + "content": "[26] Hossein Hadian, Hossein Sameti, Daniel Povey, and Sanjeev Khudanpur. End-to-end speech recognition using lattice-free MMI. In INTERSPEECH, pages 12-16. ISCA, 2018. 5" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 571, + 545, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 571, + 545, + 613 + ], + "spans": [ + { + "bbox": [ + 308, + 571, + 545, + 613 + ], + "type": "text", + "content": "[27] Alexandros Haliassos, Pingchuan Ma, Rodrigo Mira, Stavros Petridis, and Maja Pantic. Jointly learning visual and auditory speech representations from raw data. In ICLR. Open-Review.net, 2023. 2, 6, 7, 8, 1, 4, 5" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 308, + 615, + 545, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 615, + 545, + 657 + ], + "spans": [ + { + "bbox": [ + 308, + 615, + 545, + 657 + ], + "type": "text", + "content": "[28] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, and Ross B. Girshick. Masked autoencoders are scalable vision learners. In CVPR, pages 15979-15988. IEEE, 2022. 2" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 308, + 658, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 658, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 658, + 545, + 713 + ], + "type": "text", + "content": "[29] François Hernandez, Vincent Nguyen, Sahar Ghannay, Natalia A. Tomashenko, and Yannick Esteve. TED-LIUM 3: Twice as much data and corpus repartition for experiments on speaker adaptation. In SPECOM, pages 198-208. Springer, 2018. 2" + } + ] + } + ], + "index": 31 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "27077" + } + ] + } + ], + "index": 33 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "type": "text", + "content": "[30] Wei-Ning Hsu and Bowen Shi. u-HuBERT: Unified mixed-modal speech pretraining and zero-shot transfer to unlabeled modality. In NeurIPS, 2022. 1" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 106, + 288, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 106, + 288, + 162 + ], + "spans": [ + { + "bbox": [ + 48, + 106, + 288, + 162 + ], + "type": "text", + "content": "[31] Wei-Ning Hsu, Benjamin Bolte, Yao-Hung Hubert Tsai, Kushal Lakhotia, Ruslan Salakhutdinov, and Abdelrahman Mohamed. HuBERT: Self-supervised speech representation learning by masked prediction of hidden units. IEEE ACM Trans. Audio Speech Lang. Process., 29:3451-3460, 2021. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 162, + 288, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 162, + 288, + 195 + ], + "spans": [ + { + "bbox": [ + 48, + 162, + 288, + 195 + ], + "type": "text", + "content": "[32] Bruno Korbar, Du Tran, and Lorenzo Torresani. Cooperative learning of audio and video models from self-supervised synchronization. In NeurIPS, pages 7774-7785, 2018. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 195, + 288, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 195, + 288, + 228 + ], + "spans": [ + { + "bbox": [ + 48, + 195, + 288, + 228 + ], + "type": "text", + "content": "[33] Aleksandr Laptev, Somshubra Majumdar, and Boris Ginsburg. CTC variations through new WFST topologies. In INTERSPEECH, pages 1041-1045. ISCA, 2022. 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 228, + 288, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 228, + 288, + 271 + ], + "spans": [ + { + "bbox": [ + 48, + 228, + 288, + 271 + ], + "type": "text", + "content": "[34] Jiachen Lian, Alexei Baevski, Wei-Ning Hsu, and Michael Auli. AV-data2vec: Self-supervised learning of audio-visual speech representations with contextualized target representations. In ASRU. IEEE, 2023. 2, 6, 7, 8, 1, 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 271, + 288, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 271, + 288, + 316 + ], + "spans": [ + { + "bbox": [ + 48, + 271, + 288, + 316 + ], + "type": "text", + "content": "[35] Paul Pu Liang, Zihao Deng, Martin Q. Ma, James Y. Zou, Louis-Philippe Morency, and Ruslan Salakhutdinov. Factorized contrastive learning: Going beyond multi-view redundancy. In NeurIPS, 2023. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 316, + 288, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 316, + 288, + 360 + ], + "spans": [ + { + "bbox": [ + 48, + 316, + 288, + 360 + ], + "type": "text", + "content": "[36] Alexander H. Liu, Heng-Jui Chang, Michael Auli, Wei-Ning Hsu, and Jim Glass. DinoSR: Self-distillation and online clustering for self-supervised speech representation learning. In NeurIPS, 2023. 4, 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 360, + 288, + 383 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 360, + 288, + 383 + ], + "spans": [ + { + "bbox": [ + 48, + 360, + 288, + 383 + ], + "type": "text", + "content": "[37] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In *ICLR (Poster)*. OpenReview.net, 2019. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 383, + 288, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 383, + 288, + 426 + ], + "spans": [ + { + "bbox": [ + 48, + 383, + 288, + 426 + ], + "type": "text", + "content": "[38] Pingchuan Ma, Rodrigo Mira, Stavros Petridis, Bjorn W. Schuller, and Maja Pantic. LiRA: Learning visual speech representations from audio through self-supervision. In *Interspeech*, pages 3011–3015. ISCA, 2021. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 426, + 288, + 460 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 426, + 288, + 460 + ], + "spans": [ + { + "bbox": [ + 48, + 426, + 288, + 460 + ], + "type": "text", + "content": "[39] Pingchuan Ma, Stavros Petridis, and Maja Pantic. End-to-end audio-visual speech recognition with Conformers. In ICASSP, pages 7613-7617. IEEE, 2021. 1, 6, 7, 8, 4, 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 460, + 288, + 493 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 460, + 288, + 493 + ], + "spans": [ + { + "bbox": [ + 48, + 460, + 288, + 493 + ], + "type": "text", + "content": "[40] Pingchuan Ma, Stavros Petridis, and Maja Pantic. Visual speech recognition for multiple languages in the wild. Nat. Mac. Intell., 4(11):930-939, 2022. 1, 6, 8, 2, 4, 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 493, + 288, + 548 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 493, + 288, + 548 + ], + "spans": [ + { + "bbox": [ + 48, + 493, + 288, + 548 + ], + "type": "text", + "content": "[41] Pingchuan Ma, Alexandros Haliassos, Adriana Fernandez-Lopez, Honglie Chen, Stavros Petridis, and Maja Pantic. Auto-AVSR: Audio-Visual speech recognition with automatic labels. In ICASSP, pages 1-5. IEEE, 2023. 1, 3, 7, 8, 4, 5" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 548, + 288, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 548, + 288, + 602 + ], + "spans": [ + { + "bbox": [ + 48, + 548, + 288, + 602 + ], + "type": "text", + "content": "[42] Takaki Makino, Hank Liao, Yannis M. Assael, Brendan Shillingford, Basilio Garcia, Otavio Braga, and Olivier Siohan. Recurrent neural network transducer for audio-visual speech recognition. In ASRU, pages 905–912. IEEE, 2019. 1, 8, 5" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 602, + 288, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 602, + 288, + 658 + ], + "spans": [ + { + "bbox": [ + 48, + 602, + 288, + 658 + ], + "type": "text", + "content": "[43] Antoine Miech, Dimitri Zhukov, Jean-Baptiste Alayrac, Makarand Tapaswi, Ivan Laptev, and Josef Sivic. HowTo100M: Learning a text-video embedding by watching hundred million narrated video clips. In ICCV, pages 2630-2640. IEEE, 2019. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 658, + 288, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 288, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 288, + 712 + ], + "type": "text", + "content": "[44] Arsha Nagrani, Chen Sun, David Ross, Rahul Sukthankar, Cordelia Schmid, and Andrew Zisserman. Speech2Action: Cross-modal supervision for action recognition. In CVPR, pages 10314-10323. Computer Vision Foundation / IEEE, 2020. 2" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 72, + 546, + 712 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 307, + 72, + 546, + 129 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 72, + 546, + 129 + ], + "spans": [ + { + "bbox": [ + 307, + 72, + 546, + 129 + ], + "type": "text", + "content": "[45] Xichen Pan, Peiyu Chen, Yichen Gong, Helong Zhou, Xinbing Wang, and Zhouhan Lin. Leveraging unimodal self-supervised learning for multimodal audio-visual speech recognition. In ACL (1), pages 4491–4503. Association for Computational Linguistics, 2022. 7, 4" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 129, + 546, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 129, + 546, + 184 + ], + "spans": [ + { + "bbox": [ + 307, + 129, + 546, + 184 + ], + "type": "text", + "content": "[46] Daniel S. Park, William Chan, Yu Zhang, Chung-Cheng Chiu, Barret Zoph, Ekin D. Cubuk, and Quoc V. Le. SpecAugment: A simple data augmentation method for automatic speech recognition. In INTERSPEECH, pages 2613-2617. ISCA, 2019. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 185, + 546, + 219 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 185, + 546, + 219 + ], + "spans": [ + { + "bbox": [ + 307, + 185, + 546, + 219 + ], + "type": "text", + "content": "[47] Sooyoung Park, Arda Senocak, and Joon Son Chung. MarginNCE: Robust sound localization with a negative margin. In ICASSP, pages 1-5. IEEE, 2023. 5, 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 220, + 546, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 220, + 546, + 308 + ], + "spans": [ + { + "bbox": [ + 307, + 220, + 546, + 308 + ], + "type": "text", + "content": "[48] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, Alban Desmaison, Andreas Köpf, Edward Z. Yang, Zachary DeVito, Martin Raison, Alykhan Tejani, Sasank Chilamkurthy, Benoit Steiner, Lu Fang, Junjie Bai, and Soumith Chintala. PyTorch: An imperative style, high-performance deep learning library. In NeurIPS, pages 8024-8035, 2019. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 308, + 546, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 308, + 546, + 363 + ], + "spans": [ + { + "bbox": [ + 307, + 308, + 546, + 363 + ], + "type": "text", + "content": "[49] Mandela Patrick, Yuki Markus Asano, Polina Kuznetsova, Ruth Fong, João F. Henriques, Geoffrey Zweig, and Andrea Vedaldi. On compositions of transformations in contrastive self-supervised learning. In ICCV, pages 9557-9567. IEEE, 2021. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 364, + 546, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 364, + 546, + 408 + ], + "spans": [ + { + "bbox": [ + 307, + 364, + 546, + 408 + ], + "type": "text", + "content": "[50] Trung Pham, Chaoning Zhang, Axi Niu, Kang Zhang, and Chang D. Yoo. On the pros and cons of momentum encoder in self-supervised visual representation learning. CoRR, abs/2208.05744, 2022. 4" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 410, + 546, + 465 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 410, + 546, + 465 + ], + "spans": [ + { + "bbox": [ + 307, + 410, + 546, + 465 + ], + "type": "text", + "content": "[51] Daniel Povey, Vijayaditya Peddinti, Daniel Galvez, Pegah Ghahremani, Vimal Manohar, Xingyu Na, Yiming Wang, and Sanjeev Khudanpur. Purely sequence-trained neural networks for ASR based on lattice-free MMI. In INTER-SPEECH, pages 2751-2755. ISCA, 2016. 5" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 466, + 546, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 466, + 546, + 499 + ], + "spans": [ + { + "bbox": [ + 307, + 466, + 546, + 499 + ], + "type": "text", + "content": "[52] K. R. Prajwal, Triantafyllos Afouras, and Andrew Zisserman. Sub-word level lip reading with visual attention. In CVPR, pages 5152-5162. IEEE, 2022. 8, 4, 5" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 500, + 546, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 500, + 546, + 567 + ], + "spans": [ + { + "bbox": [ + 307, + 500, + 546, + 567 + ], + "type": "text", + "content": "[53] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In ICML, pages 8748-8763. PMLR, 2021. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 567, + 546, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 567, + 546, + 612 + ], + "spans": [ + { + "bbox": [ + 307, + 567, + 546, + 612 + ], + "type": "text", + "content": "[54] Alec Radford, Jong Wook Kim, Tao Xu, Greg Brockman, Christine McLeavey, and Ilya Sutskever. Robust speech recognition via large-scale weak supervision. In ICML, pages 28492-28518. PMLR, 2023. 6" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 613, + 546, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 613, + 546, + 669 + ], + "spans": [ + { + "bbox": [ + 307, + 613, + 546, + 669 + ], + "type": "text", + "content": "[55] Jordi Robert-Ribes, Jean-Luc Schwartz, Tahar Lallouache, and Pierre Escudier. Complementarity and synergy in b-modal speech: Auditory, visual, and audio-visual identification of french oral vowels in noise. The Journal of the Acoustical Society of America, 103(6):3677-3689, 1998. 1" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 669, + 546, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 669, + 546, + 712 + ], + "spans": [ + { + "bbox": [ + 307, + 669, + 546, + 712 + ], + "type": "text", + "content": "[56] Steffen Schneider, Alexei Baevski, Ronan Collobert, and Michael Auli. wav2vec: Unsupervised pre-training for speech recognition. In INTERSPEECH, pages 3465-3469. ISCA, 2019. 3" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "27078" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 73, + 286, + 712 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 48, + 73, + 286, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 73, + 286, + 106 + ], + "spans": [ + { + "bbox": [ + 48, + 73, + 286, + 106 + ], + "type": "text", + "content": "[57] Dmitriy Serdyuk, Otavio Braga, and Olivier Siohan. Audiovisual speech recognition is worth " + }, + { + "bbox": [ + 48, + 73, + 286, + 106 + ], + "type": "inline_equation", + "content": "32 \\times 32 \\times 8" + }, + { + "bbox": [ + 48, + 73, + 286, + 106 + ], + "type": "text", + "content": " voxels. In ASRU, pages 796-802. IEEE, 2021. 2, 8, 5" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 106, + 286, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 106, + 286, + 150 + ], + "spans": [ + { + "bbox": [ + 48, + 106, + 286, + 150 + ], + "type": "text", + "content": "[58] Dmitriy Serdyuk, Otavio Braga, and Olivier Siohan. Transformer-based video front-ends for audio-visual speech recognition for single and multi-person video. In INTER-SPEECH, pages 2833–2837. ISCA, 2022. 2, 8, 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 150, + 286, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 150, + 286, + 194 + ], + "spans": [ + { + "bbox": [ + 48, + 150, + 286, + 194 + ], + "type": "text", + "content": "[59] Bowen Shi, Wei-Ning Hsu, Kushal Lakhotia, and Abdelrahman Mohamed. Learning audio-visual speech representation by masked multimodal cluster prediction. In ICLR. OpenReview.net, 2022. 1, 2, 3, 6, 7, 8, 4, 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 194, + 286, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 194, + 286, + 227 + ], + "spans": [ + { + "bbox": [ + 48, + 194, + 286, + 227 + ], + "type": "text", + "content": "[60] Bowen Shi, Wei-Ning Hsu, and Abdelrahman Mohamed. Robust self-supervised audio-visual speech recognition. In INTERSPEECH, pages 2118–2122. ISCA, 2022. 7, 8, 4, 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 228, + 286, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 228, + 286, + 270 + ], + "spans": [ + { + "bbox": [ + 48, + 228, + 286, + 270 + ], + "type": "text", + "content": "[61] Bowen Shi, Abdelrahman Mohamed, and Wei-Ning Hsu. Learning lip-based audio-visual speaker embeddings with AV-HuBERT. In INTERSPEECH, pages 4785–4789. ISCA, 2022. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 271, + 286, + 347 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 271, + 286, + 347 + ], + "spans": [ + { + "bbox": [ + 48, + 271, + 286, + 347 + ], + "type": "text", + "content": "[62] Brendan Shillingford, Yannis M. Assael, Matthew W. Hoffman, Thomas Paine, Cian Hughes, Utsav Prabhu, Hank Liao, Hasim Sak, Kanishka Rao, Lorrayne Bennett, Marie Mulville, Misha Denil, Ben Coppin, Ben Laurie, Andrew W. Senior, and Nando de Freitas. Large-scale visual speech recognition. In INTERSPEECH, pages 4135-4139. ISCA, 2019. 8, 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 349, + 286, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 349, + 286, + 403 + ], + "spans": [ + { + "bbox": [ + 48, + 349, + 286, + 403 + ], + "type": "text", + "content": "[63] Nina Shvetsova, Brian Chen, Andrew Rouditchenko, Samuel Thomas, Brian Kingsbury, Rogério Feris, David Harwath, James R. Glass, and Hilde Kuehne. Everything at once - multi-modal fusion Transformer for video retrieval. In CVPR, pages 19988-19997. IEEE, 2022. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 404, + 286, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 404, + 286, + 437 + ], + "spans": [ + { + "bbox": [ + 48, + 404, + 286, + 437 + ], + "type": "text", + "content": "[64] Themos Stafylakis and Georgios Tzimiropoulos. Combining residual networks with LSTMs for lipreading. In INTER-SPEECH, pages 3652-3656. ISCA, 2017. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 437, + 286, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 437, + 286, + 470 + ], + "spans": [ + { + "bbox": [ + 48, + 437, + 286, + 470 + ], + "type": "text", + "content": "[65] Jianlin Su, Yu Lu, Shengfeng Pan, Bo Wen, and Yunfeng Liu. RoFormer: Enhanced transformer with rotary position embedding. CoRR, abs/2104.09864, 2021. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 471, + 286, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 471, + 286, + 514 + ], + "spans": [ + { + "bbox": [ + 48, + 471, + 286, + 514 + ], + "type": "text", + "content": "[66] Weijie Su, Xizhou Zhu, Yue Cao, Bin Li, Lewei Lu, Furu Wei, and Jifeng Dai. VL-BERT: Pre-training of generic visual-linguistic representations. In ICLR. OpenReview.net, 2020. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 514, + 286, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 514, + 286, + 557 + ], + "spans": [ + { + "bbox": [ + 48, + 514, + 286, + 557 + ], + "type": "text", + "content": "[67] Chen Sun, Austin Myers, Carl Vondrick, Kevin Murphy, and Cordelia Schmid. VideoBERT: A joint model for video and language representation learning. In ICCV, pages 7463-7472. IEEE, 2019. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 559, + 286, + 592 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 559, + 286, + 592 + ], + "spans": [ + { + "bbox": [ + 48, + 559, + 286, + 592 + ], + "type": "text", + "content": "[68] Yonglong Tian, Dilip Krishnan, and Phillip Isola. Contrastive multiview coding. In ECCV (11), pages 776-794. Springer, 2020. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 593, + 286, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 593, + 286, + 624 + ], + "spans": [ + { + "bbox": [ + 48, + 593, + 286, + 624 + ], + "type": "text", + "content": "[69] Hugo Touvron, Matthieu Cord, Alexandre Sablayrolles, Gabriel Synnaeve, and Hervé Jégou. Going deeper with image Transformers. In ICCV, pages 32-42. IEEE, 2021. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 625, + 286, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 625, + 286, + 668 + ], + "spans": [ + { + "bbox": [ + 48, + 625, + 286, + 668 + ], + "type": "text", + "content": "[70] Kristin J Van Engen, Avanti Dey, Mitchell S Sommers, and Jonathan E Peelle. Audiovisual speech perception: Moving beyond mcgurk. The Journal of the Acoustical Society of America, 152(6):3216-3225, 2022. 1" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 669, + 286, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 669, + 286, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 669, + 286, + 712 + ], + "type": "text", + "content": "[71] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In NIPS, pages 5998-6008, 2017. 3, 2" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 475 + ], + "type": "list", + "angle": 0, + "index": 25, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "text", + "content": "[72] Feng Wang and Huaping Liu. Understanding the behaviour of contrastive loss. In CVPR, pages 2495-2504. Computer Vision Foundation / IEEE, 2021. 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 107, + 545, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 107, + 545, + 139 + ], + "spans": [ + { + "bbox": [ + 307, + 107, + 545, + 139 + ], + "type": "text", + "content": "[73] Paul L. Williams and Randall D. Beer. Nonnegative decomposition of multivariate information. CoRR, abs/1004.2515, 2010. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 141, + 545, + 174 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 141, + 545, + 174 + ], + "spans": [ + { + "bbox": [ + 307, + 141, + 545, + 174 + ], + "type": "text", + "content": "[74] Bo Xu, Cheng Lu, Yandong Guo, and Jacob Wang. Discriminative multi-modality speech recognition. In CVPR, pages 14421-14430. Computer Vision Foundation / IEEE, 2020. 5" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 175, + 545, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 175, + 545, + 229 + ], + "spans": [ + { + "bbox": [ + 307, + 175, + 545, + 229 + ], + "type": "text", + "content": "[75] Jianwei Yu, Shi-Xiong Zhang, Jian Wu, Shahram Ghorbani, Bo Wu, Shiyin Kang, Shansong Liu, Xunying Liu, Helen Meng, and Dong Yu. Audio-visual recognition of overlapped speech for the LRS2 dataset. In ICASSP, pages 6984–6988. IEEE, 2020. 7, 4" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 230, + 545, + 284 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 230, + 545, + 284 + ], + "spans": [ + { + "bbox": [ + 308, + 230, + 545, + 284 + ], + "type": "text", + "content": "[76] Jing-Xuan Zhang, Genshun Wan, Zhen-Hua Ling, Jia Pan, Jianqing Gao, and Cong Liu. Self-supervised audiovisual speech representations learning by multimodal self-distillation. In ICASSP, pages 1-5. IEEE, 2023. 2, 7, 8, 1, 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 285, + 545, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 285, + 545, + 342 + ], + "spans": [ + { + "bbox": [ + 308, + 285, + 545, + 342 + ], + "type": "text", + "content": "[77] Xiaohui Zhang, Vimal Manohar, David Zhang, Frank Zhang, Yangyang Shi, Nayan Singhal, Julian Chan, Fuchun Peng, Yatharth Saraf, and Mike Seltzer. On lattice-free boosted MMI training of HMM and CTC-based full-context ASR models. In ASRU, pages 1026-1033. IEEE, 2021. 5" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 342, + 545, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 342, + 545, + 385 + ], + "spans": [ + { + "bbox": [ + 308, + 342, + 545, + 385 + ], + "type": "text", + "content": "[78] Ya Zhao, Rui Xu, Xinchao Wang, Peng Hou, Haihong Tang, and Mingli Song. Hearing lips: Improving lip reading by distilling speech recognizers. In AAAI, pages 6917-6924. AAAI Press, 2020. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 387, + 545, + 419 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 387, + 545, + 419 + ], + "spans": [ + { + "bbox": [ + 308, + 387, + 545, + 419 + ], + "type": "text", + "content": "[79] Jian Zhu, Cong Zhang, and David Jurgens. Phone-to-audio alignment without text: A semi-supervised approach. In ICASSP, pages 8167-8171. IEEE, 2022. 8" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 420, + 545, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 420, + 545, + 475 + ], + "spans": [ + { + "bbox": [ + 308, + 420, + 545, + 475 + ], + "type": "text", + "content": "[80] Qiushi Zhu, Long Zhou, Ziqiang Zhang, Shujie Liu, Binxing Jiao, Jie Zhang, Lirong Dai, Daxin Jiang, Jinyu Li, and Furu Wei. VATLM: Visual-Audio-Text pre-training with unified masked prediction for speech representation learning. IEEE Transactions on Multimedia, pages 1–11, 2023. 1, 6, 7, 4, 5" + } + ] + } + ], + "index": 24 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "27079" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/ESCAPE_ Encoding Super-keypoints for Category-Agnostic Pose Estimation/64b3bbb9-9580-410a-adfe-c8a2bd5b8bca_content_list.json b/2024/ESCAPE_ Encoding Super-keypoints for Category-Agnostic Pose Estimation/64b3bbb9-9580-410a-adfe-c8a2bd5b8bca_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..b53993903a5de87b9eb50eb01b9c67efb66f902e --- /dev/null +++ b/2024/ESCAPE_ Encoding Super-keypoints for Category-Agnostic Pose Estimation/64b3bbb9-9580-410a-adfe-c8a2bd5b8bca_content_list.json @@ -0,0 +1,1667 @@ +[ + { + "type": "text", + "text": "ESCAPE: Encoding Super-keypoints for Category-Agnostic Pose Estimation", + "text_level": 1, + "bbox": [ + 99, + 130, + 870, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Khoi Duc Nguyen1* \n1University of Wisconsin-Madison \nkhoi.nguyen@wisc.edu lichen@u", + "bbox": [ + 155, + 179, + 459, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Chen Li $^{2}$ Gim Hee Lee $^{2}$ $^{2}$ National University of Singapore \n.nus.edu gimhee.lee@comp.nus.edu.", + "bbox": [ + 462, + 179, + 789, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 268, + 313, + 284 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this paper, we tackle the task of category-agnostic pose estimation (CAPE), which aims to predict poses for objects of any category with few annotated samples. Previous works either rely on local matching between features of support and query samples or require support keypoint identifier. The former is prone to overfitting due to its sensitivity to sparse samples, while the latter is impractical for the open-world nature of the task. To overcome these limitations, we propose ESCAPE - a Bayesian framework that learns a prior over the features of keypoints. The prior can be expressed as a mixture of super-keypoints, each being a high-level abstract keypoint that captures the statistics of semantically related keypoints from different categories. We estimate the super-keypoints from base categories and use them in adaptation to novel categories. The adaptation to an unseen category involves two steps: first, we match each novel keypoint to a related super-keypoint; and second, we transfer the knowledge encoded in the matched super-keypoints to the novel keypoints. For the first step, we propose a learnable matching network to capture the relationship between the novel keypoints and the super-keypoints, resulting in a more reliable matching. ESCAPE mitigates overfitting by directly transferring learned knowledge to novel categories while it does not use keypoint identifiers. We achieve state-of-the-art performance on the standard MP-100 benchmark. Our code is available at https://github.com/khoiucd/escape-tgt.", + "bbox": [ + 75, + 301, + 473, + 710 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 756, + 209, + 771 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "2D pose estimation aims to locate the semantic keypoints of an object of interest from an input image. This task is crucial for many visual understanding tasks, such as understanding human/animal behavior or interpreting the surrounding environment in autonomous driving. Existing works focus on detecting keypoints of only one specific", + "bbox": [ + 75, + 782, + 468, + 875 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/c7d77b36b6441bbb24345f270f639eeb9aea1b2ddc508e68a47a3c9de239fe77.jpg", + "image_caption": [ + "Figure 1. We encode super-keypoint 'left knee' from the 'left knee' of base animals and use the super-keypoint in adaptation to the left knee of novel animal." + ], + "image_footnote": [], + "bbox": [ + 506, + 270, + 883, + 381 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "(super-)category, such as human, animal, or vehicle. They are inefficient in the sense that they need to train a separate model for each application. Moreover, annotating a large number of samples for some categories, e.g., rare species, is too expensive, if not impossible, and thus hinders the usefulness of the deep learning approach. A unified model that can quickly adapt to keypoints of unseen categories is desirable. Addressing this, Xu et al. [42] propose the task of category-agnostic pose estimation (CAPE) in which a unified model is trained to detect the poses of objects from any category with limited supervision.", + "bbox": [ + 496, + 462, + 892, + 628 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Previous works attempts to compute keypoint prototypes from support and query samples and use them to locate keypoint positions on the query image via a matching step. Xu et al. [42] introduce POMNet, a transformer-based network, to capture the interactions between support and query images into the keypoint prototypes. POMNet is prone to overfitting since the keypoint prototypes heavily rely on very few, typically one or five, support samples. Recently, Shi et al. [31] propose a two-stage framework, called CapeFormer, to further calibrate the matching results. Although effective, CapeFormer uses support keypoint identifier to alleviate the ambiguity of keypoints with similar appearances. The keypoint identifier requires additional effort to consistently annotate the keypoints of different categories. This might prohibit their application in scenarios where keypoint identifiers are not available.", + "bbox": [ + 496, + 628, + 893, + 869 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this work, we aim to improve the representation capacity of the keypoint prototypes by leveraging the knowl-", + "bbox": [ + 498, + 869, + 893, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Work done as an intern at NUS.", + "bbox": [ + 94, + 886, + 271, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "23491", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "edge learned from the base categories. We propose a Bayesian framework that models a prior over the keypoint prototypes. This allows us to analytically encode the knowledge from the base categories and explicitly transfer it to the novel ones. Specifically, we introduce the notion of super-keypoint, a high-level abstract keypoint that models the distribution of semantically related keypoints from different categories. This is based on the observation that keypoints from different categories can be closely related. For example, keypoints corresponding to the left knee of many animals are visually similar and can be grouped into a super-keypoint 'left knee'. Encoding the super-keypoint 'left knee' from many base animals can help generalize to the left knee of an unseen animal (see Fig. 1). The super-keypoints are discovered and estimated automatically from the keypoints of base categories, whose keypoint distributions are well estimated from the abundant number of samples. The adaptation to an unseen category involves two steps: first, we match each novel keypoint to a related super-keypoint; and second, we estimate the keypoint prototypes by considering both the features from the support samples and the prior knowledge encoded in the matched super-keypoints. The first step requires finding an optimal bipartite matching between the set of novel keypoints and the learned super-keypoints. To solve this, we leverage a learnable matching network to capture the structural relations between the novel keypoints and the super-keypoints before making predictions.", + "bbox": [ + 76, + 90, + 472, + 513 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We call our framework Encoding Super-keypoints for Category-Agnostic Pose Estimation, dubbed ESCAPE. ESCAPE prevents the estimated keypoints from overfitting to the sparse support samples by allowing knowledge from the abundant base categories to be explicitly transferred to the novel one during adaptation. Furthermore, ESCAPE does not use keypoint identifiers to alleviate keypoint ambiguity. Experiments show that our method outperform state-of-the-art methods on the standard MP-100 benchmark, assuming keypoint identifiers are not available.", + "bbox": [ + 76, + 513, + 472, + 666 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related works", + "text_level": 1, + "bbox": [ + 76, + 679, + 222, + 694 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2D pose estimation. 2D pose estimation aims to estimate the positions of a set of semantic keypoints of an object from its image. This task holds significant importance across various applications, such as human pose estimation for understanding human behaviors [19, 24, 40, 41], animal pose estimation for wildlife conservation [11, 17, 20, 46], and more. Existing methods can be roughly divided into regression-based methods [3, 26, 34, 35, 37, 39] and heatmap-based methods [1, 2, 4, 5, 18, 33]. The former directly predicts the coordinates of keypoints, while the latter infers heatmaps and takes the peak positions as it predictions. However, these methods are primarily designed for estimating poses for a single (super-)category with a fixed", + "bbox": [ + 76, + 704, + 472, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "set of semantic keypoints, and they typically require access to large-scale datasets.", + "bbox": [ + 498, + 90, + 890, + 121 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Few-shot learning. Few-shot learning focuses on learning novel concepts from few examples [14, 45]. While few-shot learning has been extensively studied in domains such as image classification [16] and video classification [25], its application in 2D pose estimation remains relatively under-explored. Our notion of super-keypoints is related to the Bayesian approach in few-shot learning literature [29, 43, 48]. Salakhutdinov et al. [29] introduce super-categories that play role as higher-level abstract concepts for categories and encode distributions over category parameters. Zhang et al. [48] design a Bayesian framework to learn a single super-category that induces a prior over parameters for all categories; such an approach limits model flexibility and is not suitable for CAPE due to the diversity of keypoints. Yang et al. [43] observe that similar categories have nearly identical statistics and propose enhancing novel categories with similar base categories. However, it is not trivial to adopt this method for CAPE due to the lack of a structural way to measure similarity between categories with semantically and quantitatively different keypoints.", + "bbox": [ + 496, + 138, + 893, + 441 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Semantic correspondence. Semantic correspondence refers to the semantic matching between pixels or keypoints between images [6]. A semantic correspondence model typically includes a feature extractor that generates dense features from images and establishes the correspondence using a similarity function. Oquab et al. [27] show that self-supervised ViT models can produce expressive features for finding semantic correspondence. Recently, Tang et al. [36] observe that semantic correspondence naturally emerges as an ability of diffusion models. These models can effectively solve CAPE by matching query pixels to keypoints in support images. Since ESCAPE operates in the feature space, it complements semantic correspondence models. Integrating ESCAPE with an off-the-shelf feature extractor consistently enhances its performance in CAPE tasks.", + "bbox": [ + 496, + 458, + 893, + 700 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3. Background", + "text_level": 1, + "bbox": [ + 500, + 712, + 627, + 729 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3.1. Category-agnostic pose estimation", + "text_level": 1, + "bbox": [ + 500, + 737, + 799, + 753 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Category-agnostic pose estimation (CAPE) [42] seeks an unified model to estimate poses from any category given a few annotated samples. Each category $y$ has a set of $T_{y}$ keypoints of interest: $J_{y} = \\{j_{r}\\}_{r = 1}^{T_{y}}$ . Given an image $\\mathbf{I}$ of $y$ , the model's objective is to accurately predict the coordinates $P = \\{\\mathbf{p}_r\\}_{r = 1}^{T_y}$ of these keypoints. The setting of CAPE consists of two disjoint sets of categories that are $C_{base}$ for training the model and $C_{novel}$ for testing its generalization capabilities. While the base categories have abundant", + "bbox": [ + 496, + 760, + 893, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "23492", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "training data, the novel categories have very few training samples. Specifically, a training set $\\mathcal{D}_{base}^{y} = \\{(I^{i},P^{i})\\}_{i = 1}^{n_{y}}$ ( $n_{y}$ is large) of each base category $y$ from $C_{base}$ is available for training the model. The testing stage includes a large number of episodes (or tasks) sampled from the novel categories. In each episode, a support set $S_{novel}^{y} = \\{(\\mathbf{I}^{i},P^{i})\\}_{i = 1}^{K}$ of $K$ annotated samples from $y\\in C_{novel}$ is given to the model for adaptation, and the model performance is evaluated on a set $\\mathcal{Q}_{novel}^{y} = \\{(\\mathbf{I}^{i},P^{i})\\}_{i = K + 1}^{K + Q}$ of $Q$ query samples.", + "bbox": [ + 76, + 90, + 472, + 243 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Support keypoint identifier. The MP-100 dataset is composed of existing widely used datasets, many of which impose an ordering over the keypoints of their categories. Shi et al [31] refer to this ordering information as support keypoint identifier and encode it into the keypoint features to alleviate the ambiguity of novel keypoints. However, incorporating keypoint identifiers may limit the applicability of CAPE to real-world scenarios where novel keypoints can be freely defined. In this paper, we focus on the challenging yet practical settings for CAPE that do not assume the order of keypoints to be available. We refer readers to [13] for detailed discussion of the support keypoint identifier.", + "bbox": [ + 75, + 258, + 472, + 441 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Simple baseline for CAPE", + "text_level": 1, + "bbox": [ + 76, + 446, + 313, + 464 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we introduce a simple yet effective baseline for CAPE. The baseline includes a feature extractor $f_{\\theta}$ that extracts dense features from input images. We denote the local feature extracted at a position $\\mathbf{p}$ of an input image $\\mathbf{I}$ as $f_{\\theta}(\\mathbf{I})[\\mathbf{p}] \\in \\mathbb{R}^{d}$ . We aim to train $f_{\\theta}$ so that it possesses strong semantic correspondence for the keypoints of interest.", + "bbox": [ + 76, + 470, + 468, + 561 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "For each base category $y$ with keypoints of interest $J_{y}$ we define a set of prototypes $\\mathbf{W}^y = \\{\\mathbf{w}_r\\}_{r = 1}^{T_y}$ ; each prototype $\\mathbf{w}_r\\in \\mathbb{R}^d$ is responsible for predicting a heatmap $\\mathbf{H}_r$ for the $r^{\\mathrm{th}}$ keypoint in $J_{y}$ as follows:", + "bbox": [ + 76, + 561, + 468, + 625 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {H} _ {r} [ \\mathbf {p} ] = \\left\\langle f _ {\\theta} (\\mathbf {I}) [ \\mathbf {p} ], \\mathbf {w} _ {r} \\right\\rangle , \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 189, + 631, + 468, + 648 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\langle \\cdot, \\cdot \\rangle$ denotes the inner product, and $\\mathbf{p}$ varies over the spatial dimensions. The peak position in the heatmap is then the prediction for that keypoint.", + "bbox": [ + 76, + 652, + 468, + 698 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We train the feature extractor $f_{\\theta}$ and the base prototypes $\\{\\mathbf{W}^y\\}_{y \\in C_{base}}$ on the base training data $\\bigcup_{y \\in C_{base}} \\{\\mathcal{D}_{base}^y\\}$ with the standard Mean Squared Error (MSE) loss [23, 38, 41] between the predicted heatmaps and the ground-truth heatmaps.", + "bbox": [ + 76, + 699, + 468, + 773 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To adapt to a novel category $y \\in C_{\\text{novel}}$ with support and query sets $(S_{\\text{novel}}^y, Q_{\\text{novel}}^y)$ , we freeze the feature extractor and compute the prototypes $\\mathbf{W}^y = \\{\\mathbf{w}_r\\}_{r=1}^{T_y}$ for novel keypoints as the means of keypoint features extracted from the support set:", + "bbox": [ + 76, + 773, + 468, + 853 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {w} _ {r} = \\frac {1}{K} \\sum_ {\\left(\\mathbf {I} ^ {i}, P ^ {i}\\right) \\in \\mathcal {S} _ {n o v e l} ^ {y}} f _ {\\theta} \\left(\\mathbf {I} ^ {i}\\right) \\left[ \\mathbf {p} _ {r} ^ {i} \\right]. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 156, + 857, + 468, + 897 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/f8eae37c575b5467bca7dbb859695fd1c8f55767793bbe4f3cd04f09a43d6b62.jpg", + "image_caption": [ + "(a) zebra" + ], + "image_footnote": [], + "bbox": [ + 522, + 89, + 692, + 189 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/c6f86616d56748aa2372eaadbe603d88a655f0a4cc99e7d596c27f32b5dbe6e8.jpg", + "image_caption": [ + "(b) horse" + ], + "image_footnote": [], + "bbox": [ + 700, + 89, + 870, + 189 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/d658db9c6c42f13179b34da3b519c74e5545a9181ab05ec8cc45d3c4c53af1d8.jpg", + "image_caption": [ + "Figure 2. t-SNE visualization of distributions of 7 keypoints of zebra and horse. Different colors represent different keypoints. $\\times$ denotes keypoint features extracted from images. (a) shows extracted features (denoted as $\\times$ ) and distribution means (denoted as $\\triangle$ ) for keypoints of zebra. (b) shows extracted features (denoted as $\\times$ ) and distribution means (denoted as $\\circ$ ) for keypoints of horse.", + "Figure 3. Super-keypoint modeling. Keypoint 'right back paw of zebra' and keypoint 'right back paw of horse' belong to the same super-keypoint 'right back paw'. The keypoints distribute closely around their super-keypoint." + ], + "image_footnote": [], + "bbox": [ + 532, + 321, + 859, + 417 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The computed prototypes are then used to predict keypoint heatmaps for query images as in Eq. 1.", + "bbox": [ + 496, + 511, + 888, + 541 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our simple baseline provides results comparable to POMNet [42], which has more complicated matching network. The baseline, however, has the same problem of overfitting as POMNet and CapeFormer due to its heavy reliance on the sparse support samples. Building on our simple baseline, we propose a Bayesian framework that allows us to estimate better prototypes for the novel keypoints, while keeping the feature extractor the same. Specifically, we seek to equip the baseline with strong a prior to prevent the keypoint prototypes from overfitting to the support set.", + "bbox": [ + 496, + 542, + 890, + 693 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4. Prior as a mixture of super-keypoints", + "text_level": 1, + "bbox": [ + 498, + 708, + 834, + 726 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We notice that keypoints of different categories can be visually similar, and hence their features should follow the same distribution. To verify our hypothesis, we use the feature extractor from our baseline model to extract features at the positions of seven keypoints in images of 'zebra' (a base category) and 'horse' (an unseen category). Fig. 2 (a) and (b) show keypoints features extracted from zebra and horse images respectively. For each keypoint, we also visualize the mean of its distribution (denoted as $\\triangle$ for keypoint of zebra and $\\bigcirc$ for keypoint of horse). As shown in the figure, although the model has not seen any horse sample during", + "bbox": [ + 496, + 734, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "23493", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "training, the keypoint features of horse still align well with those of zebra. This alignment suggests that capturing the distributions of zebra keypoints facilitates generalization to horse keypoints. From this observation, we introduce super-keypoints to model the connections between keypoints of horse and zebra. For instance, the keypoint 'right back paw of zebra' and the keypoint 'right back paw of horse' belong to the same super-keypoint 'right back paw'.", + "bbox": [ + 75, + 90, + 472, + 212 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1. Super-keypoints modeling", + "text_level": 1, + "bbox": [ + 76, + 222, + 316, + 239 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We introduce the notion of super-keypoint to define a prior over a group of visually similar keypoints from different categories. Let us consider the $r^{\\mathrm{th}}$ keypoint of a category $y$ . Given an image $\\mathbf{I}$ of $y$ , we denote $\\mathbf{o}_r \\coloneqq f_\\theta(\\mathbf{I})[\\mathbf{p}_r]$ as the local feature extracted at the position of the $r^{\\mathrm{th}}$ keypoint in $\\mathbf{I}$ . We assume that the keypoint feature $\\mathbf{o}_r$ follows a Gaussian distribution with a mean $\\zeta_r$ and covariance $\\phi^2 I$ . We further assume that the mean of the keypoint distribution $\\zeta_r$ obeys a Gaussian distribution with a mean $\\mu_{z_r}$ and covariance $\\sigma^2 I$ , where $\\mu_{z_r}$ represent the super-keypoint of the $r^{\\mathrm{th}}$ keypoint. Formally,", + "bbox": [ + 75, + 244, + 472, + 412 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {o} _ {r} \\sim \\mathcal {N} (\\mathbf {o} | \\zeta_ {r}; \\phi^ {2}), \\quad \\text {a n d} \\quad \\zeta_ {r} \\sim \\mathcal {N} (\\zeta | \\mu_ {z _ {r}}; \\sigma^ {2}). \\qquad (3)\n$$\n", + "text_format": "latex", + "bbox": [ + 102, + 422, + 468, + 441 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Fig. 3 illustrates our super-keypoint modeling. Intuitively, $\\mu$ is the representation of a super-keypoint in the feature space, and its keypoint members distribute closely nearby. Our goal is to discover and estimate a set of superkeypoints from keypoints of base categories so that the super-keypoints can encode the statistics of groups of related keypoints across different categories. When adapting to a novel category, the statistics encoded in the superkeypoints are directly transferred to the prototypes of novel keypoints, preventing our model from overfitting to the sparse support set.", + "bbox": [ + 75, + 453, + 472, + 619 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.2. Super-keypoints discovering and learning", + "text_level": 1, + "bbox": [ + 76, + 630, + 431, + 647 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We obtain the set of super-keypoints by applying a clustering method to the keypoints of all base categories. Specifically, given the trained feature extractor $f_{\\theta}$ from the baseline, we extract the features of keypoints from all base categories. We then estimate the distribution mean of each keypoint in the base categories as follows:", + "bbox": [ + 75, + 652, + 468, + 744 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\zeta_ {r} ^ {y} = \\frac {1}{| \\mathcal {D} _ {\\text {b a s e}} ^ {y} |} \\sum_ {(\\mathbf {I} ^ {i}, P ^ {i}) \\in \\mathcal {D} _ {\\text {b a s e}} ^ {y}} f _ {\\theta} (\\mathbf {I} ^ {i}) [ \\mathbf {p} _ {r} ^ {i} ], \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 143, + 755, + 468, + 796 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\zeta_r^y$ represents the distribution mean of the $r^{\\mathrm{th}}$ keypoint of category $y\\in C_{base}$ . The set of keypoints from all base categories is then $\\bigcup_{y\\in C_{base}}\\{\\zeta_r^y\\}_{r = 1}^{T_y}$ . In this work, we use the k-nearest neighbor based density peaks clustering [9, 47], which does not require the number of clusters to be known in advance. Note that keypoints of the same", + "bbox": [ + 75, + 806, + 470, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "category are distinct, so they cannot belong to the same super-keypoint. A detailed description of the clustering is provided in the Appendix. Finally, we estimate each super-keypoint representation $\\mu$ as the empirical mean of distribution means of keypoints within the cluster, resulting in a set of $L$ super-keypoints: $\\mathbf{M} = \\{\\mu_1,\\mu_2,\\dots ,\\mu_L\\}$ .", + "bbox": [ + 496, + 90, + 892, + 183 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "5. Novel category adaptation", + "text_level": 1, + "bbox": [ + 498, + 198, + 745, + 215 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this section, we describe how to adapt to a novel category $y$ whose support and query sets are $(\\mathcal{S}_{novel}^{y},\\mathcal{Q}_{novel}^{y})$ . Our goal is to estimate the prototypes that well represent the distributions of the novel keypoints from a few support samples. In particular, for the $r^{\\mathrm{th}}$ keypoint of $y$ , we want the estimated prototype $\\mathbf{w}_r$ to be close to the true distribution mean $\\zeta_r$ as much as possible. With this interpretation, the estimated prototype for a novel keypoint in the baseline (Eq. 2) can be viewed as maximum likelihood estimate of the mean of the keypoint distribution. It is clear that the baseline is prone to overfitting since the number of support samples is extremely limited.", + "bbox": [ + 496, + 224, + 893, + 406 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "5.1. Prototype estimate via maximum a posteriori", + "text_level": 1, + "bbox": [ + 498, + 417, + 883, + 435 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We propose to alleviate the overfitting by letting the statistics captured in the super-keypoints to be transferred to the novel keypoints. Specifically, consider the $r^{\\mathrm{th}}$ keypoint of $y$ , we compute its prototype $\\mathbf{w}_r$ as maximum a posterior estimate of $\\zeta_r$ :", + "bbox": [ + 496, + 443, + 893, + 518 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {w} _ {r} ^ {M A P} = \\underset {\\zeta_ {r}} {\\operatorname {a r g m a x}} \\log p \\left(\\zeta_ {r} \\mid \\left\\{\\mathbf {o} _ {r} ^ {i} \\right\\} _ {i = 1} ^ {K}, \\mathbf {M}\\right), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 555, + 530, + 890, + 560 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathbf{o}_r^i = f_\\theta (\\mathbf{I}^i)[\\mathbf{p}_r^i ]$ is the feature of the $r^{\\mathrm{th}}$ keypoint extracted from the $i^{\\mathrm{th}}$ support sample in $\\mathcal{S}_{novel}^{y}$ , and $\\mathbf{M} = \\{\\mu_1,\\mu_2,\\dots ,\\mu_L\\}$ is the set of learned super-keypoints.", + "bbox": [ + 498, + 571, + 890, + 619 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "However, when adapting to a novel category, the superkeypoints of the novel keypoints are not available in general. We do not know which super-keypoint the $r^{\\mathrm{th}}$ novel keypoint should belong to. To this extent, we adopt the expectation maximization (EM) algorithm [21], which first matches the $r^{\\mathrm{th}}$ keypoint with the 'best' super-keypoint from M (the expectation step) and then optimizes for the prototype with the prior encoded in the matched super-keypoint (the maximization step). Below, we summarize the two steps; detailed derivation of the (EM) algorithm is provided in the Appendix.", + "bbox": [ + 496, + 619, + 893, + 787 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Expectation step. Let $z_{r}$ be the super-keypoint assignment of the $r^{\\mathrm{th}}$ keypoint of $y$ . We align the keypoint with the most likely super-keypoint $z_{r}^{\\star}$ , i.e., the $z_{r}$ that maximizes $\\log p(\\zeta_{r}, z_{r}|\\{\\mathbf{o}_{r}^{i}\\}_{i=1}^{K}, \\mathbf{M})$ . Note that keypoints of the same category cannot belong to the same super-keypoint. Therefore, we further adopt the Hungarian algorithm [15]", + "bbox": [ + 496, + 810, + 893, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "23494", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/4ab82e4fd947cf03830e88ec00198e6ff4371112cabfe569843345bd3b004f0b.jpg", + "image_caption": [ + "Figure 4. Illustration of our adaptation procedure for a novel category. We leverage the EM algorithm to probabilistically optimize for the keypoint prototypes, considering both the support samples and the learned super-keypoints. We first align the novel keypoints with the different super-keypoints via a learnable matching network (expectation step). We then let the statistics encoded in the matched super-keypoints to be transferred directly to the keypoint prototypes (maximization step). The keypoint prototypes can be used to estimate poses for query samples." + ], + "image_footnote": [], + "bbox": [ + 138, + 85, + 836, + 402 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "to optimally align the $T_{y}$ keypoints of $y$ with distinct superkeypoints $\\{\\mu_{z_r^*}\\}_{r = 1}^{T_y}\\subset \\mathbf{M}$ . We refer to this matching procedure as simple matching for the rest of the paper.", + "bbox": [ + 75, + 510, + 470, + 559 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Maximization step. Now, we know the 'best' superkeypoint for each keypoint of $y$ . We can compute their prototypes analytically as follows:", + "bbox": [ + 75, + 577, + 468, + 625 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathbf {w} _ {r} ^ {M A P} = \\underset {\\mathbf {w} _ {r}} {\\operatorname {a r g m a x}} \\log p \\left(\\mathbf {w} _ {r} | \\left\\{\\mathbf {o} _ {r} ^ {i} \\right\\} _ {i = 1} ^ {K}, \\mu_ {z _ {r} ^ {\\star}}\\right) (6) \\\\ = \\alpha \\frac {1}{K} \\sum_ {i = 1} ^ {K} \\mathbf {o} _ {r} ^ {i} + (1 - \\alpha) \\mu_ {z _ {r} ^ {\\star}}, (7) \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 129, + 648, + 468, + 719 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\alpha \\in (0,1)$ depends on $\\phi$ and $\\sigma$ . Since the distribution mean of a keypoint must distribute near its super-keypoint, the second term in Eq. 7 has an effect of pulling the estimated prototype closer to the true mean of the keypoint distribution, diminishing the dependence of the prototype on the support samples.", + "bbox": [ + 75, + 731, + 470, + 823 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.2. Learnable matching network", + "text_level": 1, + "bbox": [ + 76, + 830, + 336, + 848 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Matching novel keypoints with semantically related superkeypoints is crucial for the success of our method. A correct super-keypoint can pull the estimated prototype closer", + "bbox": [ + 75, + 854, + 470, + 902 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "to the distribution mean of the keypoint under consideration, while an incorrect super-keypoint pulls the prototype away. However, as later shown in the experiments, the simple matching in the expectation step fails to produce reliable super-keypoints assignments. We argue that this is because the simple matching treats the novel keypoints independently, discarding their structural relations. It ignores the keypoint coordinates in the support samples, and so the model cannot reason about poses of the object of interest.", + "bbox": [ + 496, + 510, + 890, + 647 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To overcome these limitations, we replace the simple matching in the expectation step with a learnable matching network to simultaneously produce matching scores between keypoints in an image and the learned super keypoints. The design is inspired by SuperGlue [30], a matching framework that has shown great performance in aligning two sequences of features.", + "bbox": [ + 496, + 647, + 892, + 753 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Input features. Our matching network takes two sequences of features as input and produces the matching score matrix for their elements. The first sequence is augmented features of $T_{y}$ keypoints extracted from a support image of $y$ . The second sequence is a set of features for the $L$ super-keypoints. To encode features for the keypoints, we consider not only their visual appearance but also their coordinates within the support image. Specifically, given", + "bbox": [ + 496, + 780, + 893, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "23495", + "bbox": [ + 478, + 944, + 517, + 957 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/81785062c73240156340f985726695a0a06563cb629eeadba3a834f28fa4e7a1.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodSplit 1Split 2Split 3Split 4Split 5Mean PCKSplit 1Split 2Split 3Split 4Split 5Mean PCK
ProtoNet [32]46.0540.8449.1343.3444.5444.7860.3153.5161.9258.4458.6158.56
MAML [10]68.1454.7264.1963.2457.2061.5070.0355.9863.2164.7958.4762.50
Finetune [22]70.6057.0466.0665.0059.2063.5871.6757.8466.7666.5360.2464.61
POMNet [42]84.2378.2578.1778.6879.1779.7084.7279.6178.0080.3880.8580.71
CapeFormer [31]84.1379.9079.5879.7880.3480.7590.4686.4785.8086.4786.7187.19
ESCAPE86.8982.5581.2581.7281.3282.7491.4187.4385.3387.2786.7687.63
", + "bbox": [ + 122, + 88, + 848, + 186 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1. Comparison with state-of-the-art works on MP-100 dataset under 1-shot and 5-shot settings. ${}^{ \\dagger }$ denotes keypoint identifiers.", + "bbox": [ + 96, + 198, + 869, + 212 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "support image $\\mathbf{I}\\sim S_{novel}^{y}$ , the keypoint features are:", + "bbox": [ + 76, + 239, + 421, + 257 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {u} _ {r} = f _ {\\gamma} (\\mathbf {I}) \\left[ \\mathbf {p} _ {r} \\right] + P E _ {\\mathrm {g e o}} \\left(\\mathbf {p} _ {r}\\right), \\text {f o r} r = \\{1, \\dots , T _ {y} \\}, \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 86, + 265, + 468, + 282 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $f_{\\gamma}$ is a CNN feature extractor, and $PE_{\\mathrm{geo}}$ embeds a coordinate into the feature space. By considering the keypoint coordinates, our model can reason about the pose of the object in the support image.", + "bbox": [ + 76, + 292, + 468, + 353 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "At the other end, the features for the super-keypoints are defined as:", + "bbox": [ + 76, + 353, + 468, + 382 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {v} _ {z} = \\omega_ {z} + P E _ {\\mathrm {a b s}} (z), \\quad \\text {f o r} z \\in \\{1, \\dots , L \\}, \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 122, + 393, + 468, + 411 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\omega_{z}$ is a learnable feature vector for the $z^{\\mathrm{th}}$ superkeypoint, and $PE_{\\mathrm{abs}}$ embeds an index $(1,\\dots ,L)$ into the feature space. Note that the indexes of the super-keypoints are obtained from the clustering step described in Sec. 4.2.", + "bbox": [ + 76, + 422, + 468, + 483 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Feature aggregation with attention. The two sequences of features $\\{\\mathbf{u}_r\\}_{r=1}^{T_y}$ and $\\{\\mathbf{v}_z\\}_{z=1}^L$ are then passed to a neural network for information aggregation. The network comprises $U$ blocks of self-attention and cross-attention to model interactions both within and between the elements of the sequences. Results are two sequences of refined features: $\\{\\mathbf{u}_r^\\star\\}_{r=1}^{T_y}$ and $\\{\\mathbf{v}_z^\\star\\}_{z=1}^L$ . The predicted matching scores matrix is then $\\mathbf{S} \\in \\mathbb{R}^{T_y \\times L}$ , with each element $\\mathbf{S}_{r,z} = \\langle \\mathbf{u}_r^\\star, \\mathbf{v}_z^\\star \\rangle$ measuring the matching score between the $r^{\\text{th}}$ keypoint and the $z^{\\text{th}}$ super-keypoint. An illustration of the matching network is provided at the bottom of Fig. 4.", + "bbox": [ + 76, + 501, + 468, + 672 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Matching as optimal transport problem. Since the keypoints within the same category must be matched to different super-keypoints, our goal is to find a maximum score bipartite matching between the two sequences. Following SuperGlue [30], we replace the Hungarian algorithm with the differentiable Sinkhorn algorithm [7] for optimal transport. It takes the matching score matrix $\\mathbf{S}$ as input and produces a matrix $\\mathbf{Q} \\in \\mathbb{R}^{T_y \\times L}$ where each element $\\mathbf{Q}_{r,z}$ represents the probability that the $r^{\\text{th}}$ keypoint is matched with the $z^{\\text{th}}$ super-keypoint. We train the network end-to-end on the keypoints of base categories, whose super-keypoints assignments are available from the clustering step. Details of the Sinkhorn algorithm and the training procedure of the matching network are provided in the Appendix.", + "bbox": [ + 76, + 689, + 468, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Incorporating learnable matching network to novel category adaptation. We integrate the learned matching network into the novel category adaptation. Specifically, for every support image $\\mathbf{I}^i\\in S_{novel}^y$ , we compute the matching score matrix $\\mathbf{S}^i$ from its keypoints and the learned super keypoints. The corresponding optimal assignment matrix is then denoted as $\\mathbf{Q}^i$ . The final prototype for the $r^{\\mathrm{th}}$ keypoint are then estimated as follows:", + "bbox": [ + 496, + 239, + 890, + 359 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {w} _ {r} ^ {\\text {E S C A P E}} = \\alpha \\frac {1}{K} \\sum_ {i = 1} ^ {K} \\mathbf {o} _ {r} ^ {i} + (1 - \\alpha) \\frac {L}{K} \\sum_ {i = 1} ^ {K} \\sum_ {z = 1} ^ {L} \\mathbf {Q} _ {r, z} ^ {i} \\mu_ {z}. \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 514, + 369, + 890, + 426 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Here we keep the soft assignment form of $\\mathbf{Q}^i$ instead of the hard assignment to model uncertainty. Our adaptation procedure is shown in Fig. 4.", + "bbox": [ + 496, + 441, + 890, + 488 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "6. Experiments", + "text_level": 1, + "bbox": [ + 500, + 503, + 632, + 522 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We perform experiments on Multi-category Pose (MP-100) proposed in [42]. The dataset consists of more than 20,000 data instances of 100 categories from 8 super-categories including human hand, human face, human body, animal body, animal face, clothes, furniture, and vehicle. We also follow the five splits from [42] for standard CAPE settings. Each split divides the dataset into 70 training categories, 10 validation categories, and 20 testing categories. The training set is used as base categories, while validation and testing sets are novel categories for validation and testing, respectively. Please refer to [42] for more details.", + "bbox": [ + 496, + 530, + 890, + 696 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "6.1. Experimental details", + "text_level": 1, + "bbox": [ + 500, + 709, + 694, + 724 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Unless otherwise specified, for fair comparisons with previous works, we utilize ResNet-50 [12] pretrained on ImageNet [8] as the feature extractor. Also following [42], we crop the object of interest from the original image using the ground-truth bounding box, and resize it to a fixed size of $256 \\times 256$ .", + "bbox": [ + 496, + 733, + 890, + 823 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Probability of correct keypoint [44] with a threshold of 0.2 (PCK@0.2) is the evaluation metric for MP-100. During testing, we generate 60,000 episodes evenly distributed across the test categories and report the average PCK@0.2 over these episodes.", + "bbox": [ + 496, + 824, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "23496", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/2dcb167bf1df6d2a3f0848cbf991a77c10385557f296ffba52d16b4497e2d3cc.jpg", + "image_caption": [ + "super-keypoint 1" + ], + "image_footnote": [], + "bbox": [ + 99, + 101, + 299, + 162 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/dc6d2dd56b6ec5f1f181475c2fe1c16ab7c667692324d7245460aef242154200.jpg", + "image_caption": [ + "super-keypoint 4" + ], + "image_footnote": [], + "bbox": [ + 99, + 176, + 300, + 238 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/7ab2abb81057dd1bd795501fd6a87507585c83a7247f6571d56827b2d5c89b82.jpg", + "image_caption": [ + "super-keypoint 2", + "arctic wolf" + ], + "image_footnote": [], + "bbox": [ + 316, + 101, + 382, + 152 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/5becc95d4bbc53fd37ef419c7d35c1458503d22e41b4db10616afa3c288c4d7f.jpg", + "image_caption": [ + "sealion", + "ferret" + ], + "image_footnote": [], + "bbox": [ + 383, + 101, + 517, + 152 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/5821cad53eb20a3b71bf969b5d87fbd067c5a4483afdb5659f47b9b272ac65db.jpg", + "image_caption": [ + "grebe" + ], + "image_footnote": [], + "bbox": [ + 316, + 176, + 383, + 227 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f936ba5b207f2eafb9bfbd02d131c92a99c29b07ac9f24e7ca7d93197999e3e9.jpg", + "image_caption": [ + "super-keypoint 5", + "kingfisher" + ], + "image_footnote": [], + "bbox": [ + 385, + 176, + 450, + 227 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/86b51a155c2dfd9e02781812436417f8818af38079354f12e4f54474ad7c9ef9.jpg", + "image_caption": [ + "wren", + "gull" + ], + "image_footnote": [], + "bbox": [ + 452, + 176, + 584, + 229 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/745a79b6731c5e06dd435e94e3e370fba06a59ba037462611dc234c26267ccdd.jpg", + "image_caption": [ + "super-keypoint 3" + ], + "image_footnote": [], + "bbox": [ + 535, + 101, + 867, + 162 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/529764b53507c5dd148173d39691175d465e94ba850e670a31e493d169546f71.jpg", + "image_caption": [ + "long sleeved dress" + ], + "image_footnote": [], + "bbox": [ + 599, + 176, + 669, + 228 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/864cc0f355e8c5fca8876a7a91a61925f3d5cb322aee6c66dfdbd2e2c0198f53.jpg", + "image_caption": [ + "super-keypoint 6", + "vest", + "vest dress long sleeved shirt" + ], + "image_footnote": [], + "bbox": [ + 671, + 176, + 867, + 228 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/6959ff7436ec326b22dd0585880a33e7240f3a16460d31ddf18e3e5e0dc94065.jpg", + "image_caption": [ + "Figure 5. Keypoints of four different super-keypoints discovered from the base categories. Below each keypoint is the name of its category." + ], + "image_footnote": [], + "bbox": [ + 84, + 282, + 230, + 416 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d2ef5f652fa5ababd2292e8a42bea4f21d8b7ae7fae169876860d639a7579ddf.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 236, + 282, + 883, + 426 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/1ab32c0f7de342004916f7973eb8bc0093513bf88ac4d5cfbe4f91092111ae67.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 84, + 441, + 232, + 579 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/4855aa92067acae24b11129676c42da4b62c1843c6aebf796e3bb39043d4ddc9.jpg", + "image_caption": [ + "Figure 6. Qualitative comparison between simple matching and learnable matching network in aligning novel keypoints with superkeypoints. We show one base keypoint member of each matched super-keypoint." + ], + "image_footnote": [], + "bbox": [ + 236, + 438, + 883, + 580 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "6.2. Comparison with previous methods", + "text_level": 1, + "bbox": [ + 76, + 646, + 383, + 662 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We compare ESCAPE against previous methods, including Prototypical Networks [32], MAML [10], Finetune [22], POMNet [32] and CapeFormer [31]. Note that we remove the support keypoint identifier from CapeFormer. The results are shown in Tab. 1. We can see that ESCAPE outperforms other methods by around $2.5\\%$ and $0.5\\%$ in 1-shot and 5-shot settings, respectively, establishing a new state-of-the-art on the MP-100 benchmark.", + "bbox": [ + 75, + 669, + 468, + 789 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "6.3. Ablation study", + "text_level": 1, + "bbox": [ + 76, + 801, + 225, + 816 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Super-keypoints discovering. We perform clustering on the keypoints of base categories to form super-keypoints. To verify that we can obtain meaningful super-keypoints from the clustering process, we show some keypoints members of a few super-keypoints in Fig. 5. We see that each super-", + "bbox": [ + 75, + 824, + 468, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "keypoint represents a meaningful concept. For example, super-keypoint 1 consists of 'left knee' of deer, antelope and zebra. super-keypoint 4 consists of 'right rear wheel' of SUV, bus, and car, while super-keypoint 6 is a landmark on the collars of different clothes categories.", + "bbox": [ + 496, + 647, + 890, + 723 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Novel keypoints and super-keypoints matching. We qualitatively compare the performance of the simple matching and the learnable matching network in aligning novel keypoints with the learned super-keypoints. Fig. 6 shows the matching results for two novel categories, namely squirrel and woodpecker. The left side of the figure shows support images of the categories, while the right side shows the super-keypoints aligned by the simple matching and the learnable matching network. For each super-keypoint, we only show one base keypoints member. Compared with the", + "bbox": [ + 496, + 750, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "23497", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/aa60340e1a78f9c67d8affcdc0acba12cc20edca8eae4e52c31994de4afb1dd8.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodPCK
MLE83.00 ± 0.16
MAP83.88 ± 0.15
ESCAPE86.89 ± 0.14
", + "bbox": [ + 78, + 90, + 243, + 156 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/fdbe59bb7d561149dda8a5cb4dc0653b91e9b35ba954a7d249348f359ca5e96a.jpg", + "table_caption": [ + "Table 2. Results of different prototype estimation methods, namely MLE, MAP and ESCAPE." + ], + "table_footnote": [], + "table_body": "
Feature extractorMethodPCK
DIFT [36]MLE60.61
DIFT [36]ESCAPE62.47
DINOv2 [27]MLE88.31
DINOv2 [27]ESCAPE89.14
", + "bbox": [ + 259, + 88, + 470, + 157 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 3. Results of integrating ESCAPE to feature extractors of different semantic correspondence models", + "bbox": [ + 256, + 169, + 468, + 223 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "simple matching, our learnable matching network provides more reliable super-keypoints for the novel keypoints. In particular, the simple matching aligns keypoints of squirrel with super-keypoints of keypoints of kingfisher, sling and suv, which are irrelevant to the squirrel. On the other hand, the learnable matching network aligns the novel keypoints with their actual concepts, e.g., the 'left elbow of squirrel' is matched with the super-keypoint of the 'left elbow of rat'.", + "bbox": [ + 75, + 258, + 468, + 380 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Ablation on novel category adaptation. In Tab. 2, we study the performance of different methods to estimate keypoint prototypes for a novel category, namely MLE, MAP and ESCAPE. MLE is our baseline, which estimate keypoint prototypes as means of support features as in Eq. 2. MAP and ESCAPE compute keypoint prototypes as maximum a posterior estimate with the EM algorithm. The difference between MAP and ESCAPE is that MAP (Eq. 7) uses the simple matching in the expectation step, whereas ESCAPE (Eq. 10) leverages the learnable matching network. We observe that incorporating super-keypoints into the estimation of keypoint prototypes helps improve the overall performance. Specifically, the MAP outperforms MLE by roughly $1\\%$ . Regardless ESCAPE, better novel keypoints and super-keypoints alignments from the learnable matching network immediately translate into $3\\%$ improvements over the MAP counterpart.", + "bbox": [ + 75, + 421, + 468, + 679 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Ablation on different feature extractors. Feature extractors from semantic correspondence models can effectively solve CAPE since they can generate dense features for support and query images. In Tab. 3, we study the compatibility of ESCAPE with feature extractors from different semantic correspondence models, namely DIFT [36] and DINoV2 [27]. We take the pretrained stable diffusion 2.1 [28] as the backbone for DIFT, whereas we use ViT-ViT-S/14 for DINoV2. As seen, ESCAPE consistently boosts the performance of the considered semantic correspondence models. This shows that ESCAPE can benefit from future research for better representation.", + "bbox": [ + 75, + 719, + 467, + 900 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/cf502b8569da0778b0813a397dc07a58cc3c1dc4da2c3c3847660a35c76a943e.jpg", + "image_caption": [ + "Figure 7. We qualitatively compare pose estimation of Cape-Former and ESCAPE under 1-shot setting." + ], + "image_footnote": [], + "bbox": [ + 504, + 88, + 888, + 321 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6.4. Qualitative results", + "text_level": 1, + "bbox": [ + 500, + 392, + 676, + 407 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Figure 7 compares the pose estimation performance of CapeFormer and ESCAPE. Overall, ESCAPE gives more accurate pose predictions compared to CapeFormer. In particular, CapeFormer is overfitting to the support image of the sheep category. The predicted positions of the left front and back paws by CapeFormer are geometrically close to each other as they appear in the support image. In contrast, since ESCAPE allows learned knowledge about base animals to be transferred to the novel keypoints, it recovers the pose of the query sheep beautifully.", + "bbox": [ + 496, + 417, + 890, + 569 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "7. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 588, + 617, + 604 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we address the task of category-agnostic pose estimation by introducing ESCAPE, a Bayesian framework that induces a prior over the keypoint features of different categories. This prior is a mixture of super-keypoints, with each super-keypoint encoding a distribution of a group of related keypoints. ESCAPE enables the transfer of knowledge from base categories to the novel ones, thereby preventing the query predictions from overfitting to sparse support samples. ESCAPE offers several advantages over previous works: it exhibits less overfitting to support samples, does not rely on support keypoint identifiers, and is model-agnostic.", + "bbox": [ + 496, + 616, + 890, + 797 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgement", + "text_level": 1, + "bbox": [ + 500, + 816, + 658, + 834 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This research is supported in part by the National Research Foundation, Singapore under its Competitive Research Program Award NRF-CRP23-2019-0003.", + "bbox": [ + 496, + 844, + 890, + 898 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "23498", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Yuanhao Cai, Zhicheng Wang, Zhengxiong Luo, Binyi Yin, Angang Du, Haoqian Wang, Xiangyu Zhang, Xinyu Zhou, Erjin Zhou, and Jian Sun. Learning delicate local representations for multi-person pose estimation. In European Conference on Computer Vision, pages 455-472. Springer, 2020. 2", + "[2] Zhe Cao, Tomas Simon, Shih-En Wei, and Yaser Sheikh. Realtime multi-person 2d pose estimation using part affinity fields. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7291-7299, 2017. 2", + "[3] Joao Carreira, Pulkit Agrawal, Katerina Fragkiadaki, and Jitendra Malik. Human pose estimation with iterative error feedback. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4733-4742, 2016. 2", + "[4] Yilun Chen, Zhicheng Wang, Yuxiang Peng, Zhiqiang Zhang, Gang Yu, and Jian Sun. Cascaded pyramid network for multi-person pose estimation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7103-7112, 2018. 2", + "[5] Bowen Cheng, Bin Xiao, Jingdong Wang, Honghui Shi, Thomas S Huang, and Lei Zhang. Higherhrnet: Scale-aware representation learning for bottom-up human pose estimation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5386-5395, 2020. 2", + "[6] Christopher B Choy, JunYoung Gwak, Silvio Savarese, and Manmohan Chandraker. Universal correspondence network. Advances in neural information processing systems, 29, 2016. 2", + "[7] Marco Cuturi. Sinkhorn distances: Lightspeed computation of optimal transport. Advances in neural information processing systems, 26, 2013. 6", + "[8] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 6", + "[9] Mingjing Du, Shifei Ding, and Hongjie Jia. Study on density peaks clustering based on k-nearest neighbors and principal component analysis. Knowledge-Based Systems, 99: 135-145, 2016. 4", + "[10] Chelsea Finn, Pieter Abbeel, and Sergey Levine. Model-agnostic meta-learning for fast adaptation of deep networks. In International conference on machine learning, pages 1126–1135. PMLR, 2017. 6, 7", + "[11] Jacob M Graving, Daniel Chae, Hemal Naik, Liang Li, Benjamin Koger, Blair R Costelloe, and Iain D Couzin. Deepposekit, a software toolkit for fast and robust animal pose estimation using deep learning. *Elife*, 8:e47994, 2019. 2", + "[12] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 6", + "[13] Or Hirschorn and Shai Avidan. Pose anything: A graph-based approach for category-agnostic pose estimation. arXiv preprint arXiv:2311.17891, 2023. 3" + ], + "bbox": [ + 78, + 114, + 470, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[14] Timothy Hospedales, Antreas Antoniou, Paul Micaelli, and Amos Storkey. Meta-learning in neural networks: A survey. IEEE transactions on pattern analysis and machine intelligence, 44(9):5149-5169, 2021. 2", + "[15] Harold W Kuhn. The hungarian method for the assignment problem. Naval research logistics quarterly, 2(1-2):83-97, 1955. 4", + "[16] Duong Le, Khoi Duc Nguyen, Khoi Nguyen, Quoc-Huy Tran, Rang Nguyen, and Binh-Son Hua. Poodle: Improving few-shot learning via penalizing out-of-distribution samples. Advances in Neural Information Processing Systems, 34:23942-23955, 2021. 2", + "[17] Shuyuan Li, Jianguo Li, Hanlin Tang, Rui Qian, and Weiyao Lin. Atrw: a benchmark for amur tiger re-identification in the wild. arXiv preprint arXiv:1906.05586, 2019. 2", + "[18] Wenbo Li, Zhicheng Wang, Binyi Yin, Qixiang Peng, Yuming Du, Tianzi Xiao, Gang Yu, Hongtao Lu, Yichen Wei, and Jian Sun. Rethinking on multi-stage networks for human pose estimation. arXiv preprint arXiv:1901.00148, 2019. 2", + "[19] Jiang-Jiang Liu, Qibin Hou, Ming-Ming Cheng, Changhu Wang, and Jiashi Feng. Improving convolutional networks with self-calibrated convolutions. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10096-10105, 2020. 2", + "[20] Alexander Mathis, Thomas Biasi, Steffen Schneider, Mert Yuksekgonul, Byron Rogers, Matthias Bethge, and Mackenzie W Mathis. Pretraining boosts out-of-domain robustness for pose estimation. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1859-1868, 2021. 2", + "[21] Kevin P. Murphy. Probabilistic Machine Learning: Advanced Topics. MIT Press, 2023. 4", + "[22] Akihiro Nakamura and Tatsuya Harada. Revisiting fine-tuning for few-shot learning. arXiv preprint arXiv:1910.00216, 2019. 6, 7", + "[23] Alejandro Newell, Kaiyu Yang, and Jia Deng. Stacked hourglass networks for human pose estimation. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part VIII 14, pages 483-499. Springer, 2016. 3", + "[24] Alejandro Newell, Zhiao Huang, and Jia Deng. Associative embedding: End-to-end learning for joint detection and grouping. Advances in neural information processing systems, 30, 2017. 2", + "[25] Khoi D Nguyen, Quoc-Huy Tran, Khoi Nguyen, Binh-Son Hua, and Rang Nguyen. Inductive and transductive few-shot video classification via appearance and temporal alignments. In European Conference on Computer Vision, pages 471-487. Springer, 2022. 2", + "[26] Xuecheng Nie, Jiashi Feng, Jianfeng Zhang, and Shuicheng Yan. Single-stage multi-person pose machines. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6951-6960, 2019. 2", + "[27] Maxime Oquab, Timothée Darcet, Théo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023. 2, 8" + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "23499", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[28] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 8", + "[29] Ruslan Salakhutdinov, Joshua Tenenbaum, and Antonio Torralba. One-shot learning with a hierarchical nonparametric bayesian model. In Proceedings of ICML Workshop on Unsupervised and Transfer Learning, pages 195-206. JMLR Workshop and Conference Proceedings, 2012. 2", + "[30] Paul-Edouard Sarlin, Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. Superglue: Learning feature matching with graph neural networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4938–4947, 2020. 5, 6", + "[31] Min Shi, Zihao Huang, Xianzheng Ma, Xiaowei Hu, and Zhiguo Cao. Matching is not enough: A two-stage framework for category-agnostic pose estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7308-7317, 2023. 1, 3, 6, 7", + "[32] Jake Snell, Kevin Swersky, and Richard Zemel. Prototypical networks for few-shot learning. Advances in neural information processing systems, 30, 2017. 6, 7", + "[33] Ke Sun, Bin Xiao, Dong Liu, and Jingdong Wang. Deep high-resolution representation learning for human pose estimation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5693-5703, 2019. 2", + "[34] Xiao Sun, Jiaxiang Shang, Shuang Liang, and Yichen Wei. Compositional human pose regression. In Proceedings of the IEEE International Conference on Computer Vision, pages 2602-2611, 2017. 2", + "[35] Xiao Sun, Bin Xiao, Fangyin Wei, Shuang Liang, and Yichen Wei. Integral human pose regression. In Proceedings of the European conference on computer vision (ECCV), pages 529-545, 2018. 2", + "[36] Luming Tang, Menglin Jia, Qianqian Wang, Cheng Perng Phoo, and Bharath Hariharan. Emergent correspondence from image diffusion. Advances in Neural Information Processing Systems, 36:1363-1389, 2023. 2, 8", + "[37] Zhi Tian, Hao Chen, and Chunhua Shen. Direct pose: Direct end-to-end multi-person pose estimation. arXiv preprint arXiv:1911.07451, 2019. 2", + "[38] Jonathan J. Tompson, Arjun Jain, Yann LeCun, and Christoph Bregler. Joint training of a convolutional network and a graphical model for human pose estimation. Advances in neural information processing systems, 27, 2014. 3", + "[39] Alexander Toshev and Christian Szegedy. Deeppose: Human pose estimation via deep neural networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1653-1660, 2014. 2", + "[40] Jingdong Wang, Ke Sun, Tianheng Cheng, Borui Jiang, Chaorui Deng, Yang Zhao, Dong Liu, Yadong Mu, Mingkui Tan, Xinggang Wang, et al. Deep high-resolution representation learning for visual recognition. IEEE transactions on pattern analysis and machine intelligence, 43(10):3349-3364, 2020. 2" + ], + "bbox": [ + 78, + 90, + 470, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[41] Bin Xiao, Haiping Wu, and Yichen Wei. Simple baselines for human pose estimation and tracking. In Proceedings of the European conference on computer vision (ECCV), pages 466-481, 2018. 2, 3", + "[42] Lumin Xu, Sheng Jin, Wang Zeng, Wentao Liu, Chen Qian, Wanli Ouyang, Ping Luo, and Xiaogang Wang. Pose for everything: Towards category-agnostic pose estimation. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23–27, 2022, Proceedings, Part VI, pages 398–416. Springer, 2022. 1, 2, 3, 6", + "[43] Shuo Yang, Lu Liu, and Min Xu. Free lunch for few-shot learning: Distribution calibration. arXiv preprint arXiv:2101.06395, 2021. 2", + "[44] Yi Yang and Deva Ramanan. Articulated human detection with flexible mixtures of parts. IEEE transactions on pattern analysis and machine intelligence, 35(12):2878-2890, 2012. 6", + "[45] Wang Yaqing, Yao Quanming, T Kwok James, and M Ni Lionel. Generalizing from a few examples: A survey on few-shot learning. ACM computing surveys, 53(3):1-34, 2020. 2", + "[46] Hang Yu, Yufei Xu, Jing Zhang, Wei Zhao, Ziyu Guan, and Dacheng Tao. Ap-10k: A benchmark for animal pose estimation in the wild. arXiv preprint arXiv:2108.12617, 2021. 2", + "[47] Wang Zeng, Sheng Jin, Wentao Liu, Chen Qian, Ping Luo, Wanli Ouyang, and Xiaogang Wang. Not all tokens are equal: Human-centric visual analysis via token clustering transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11101-11111, 2022. 4", + "[48] Xueting Zhang, Debin Meng, Henry Gouk, and Timothy M Hospedales. Shallow bayesian meta learning for real-world few-shot recognition. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 651-660, 2021. 2" + ], + "bbox": [ + 501, + 90, + 893, + 598 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "23500", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2024/ESCAPE_ Encoding Super-keypoints for Category-Agnostic Pose Estimation/64b3bbb9-9580-410a-adfe-c8a2bd5b8bca_model.json b/2024/ESCAPE_ Encoding Super-keypoints for Category-Agnostic Pose Estimation/64b3bbb9-9580-410a-adfe-c8a2bd5b8bca_model.json new file mode 100644 index 0000000000000000000000000000000000000000..a16f02b5ee2919e57935fbc9631268ae1dc55559 --- /dev/null +++ b/2024/ESCAPE_ Encoding Super-keypoints for Category-Agnostic Pose Estimation/64b3bbb9-9580-410a-adfe-c8a2bd5b8bca_model.json @@ -0,0 +1,2321 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.1, + 0.131, + 0.872, + 0.153 + ], + "angle": 0, + "content": "ESCAPE: Encoding Super-keypoints for Category-Agnostic Pose Estimation" + }, + { + "type": "text", + "bbox": [ + 0.156, + 0.18, + 0.46, + 0.234 + ], + "angle": 0, + "content": "Khoi Duc Nguyen1* \n1University of Wisconsin-Madison \nkhoi.nguyen@wisc.edu lichen@u" + }, + { + "type": "text", + "bbox": [ + 0.464, + 0.18, + 0.79, + 0.234 + ], + "angle": 0, + "content": "Chen Li\\(^{2}\\) Gim Hee Lee\\(^{2}\\) \n\\(^{2}\\)National University of Singapore \n.nus.edu gimhee.lee@comp.nus.edu." + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.269, + 0.314, + 0.285 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.302, + 0.474, + 0.712 + ], + "angle": 0, + "content": "In this paper, we tackle the task of category-agnostic pose estimation (CAPE), which aims to predict poses for objects of any category with few annotated samples. Previous works either rely on local matching between features of support and query samples or require support keypoint identifier. The former is prone to overfitting due to its sensitivity to sparse samples, while the latter is impractical for the open-world nature of the task. To overcome these limitations, we propose ESCAPE - a Bayesian framework that learns a prior over the features of keypoints. The prior can be expressed as a mixture of super-keypoints, each being a high-level abstract keypoint that captures the statistics of semantically related keypoints from different categories. We estimate the super-keypoints from base categories and use them in adaptation to novel categories. The adaptation to an unseen category involves two steps: first, we match each novel keypoint to a related super-keypoint; and second, we transfer the knowledge encoded in the matched super-keypoints to the novel keypoints. For the first step, we propose a learnable matching network to capture the relationship between the novel keypoints and the super-keypoints, resulting in a more reliable matching. ESCAPE mitigates overfitting by directly transferring learned knowledge to novel categories while it does not use keypoint identifiers. We achieve state-of-the-art performance on the standard MP-100 benchmark. Our code is available at https://github.com/khoiucd/escape-tgt." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.757, + 0.21, + 0.772 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.783, + 0.47, + 0.875 + ], + "angle": 0, + "content": "2D pose estimation aims to locate the semantic keypoints of an object of interest from an input image. This task is crucial for many visual understanding tasks, such as understanding human/animal behavior or interpreting the surrounding environment in autonomous driving. Existing works focus on detecting keypoints of only one specific" + }, + { + "type": "image", + "bbox": [ + 0.507, + 0.271, + 0.885, + 0.382 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.394, + 0.893, + 0.436 + ], + "angle": 0, + "content": "Figure 1. We encode super-keypoint 'left knee' from the 'left knee' of base animals and use the super-keypoint in adaptation to the left knee of novel animal." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.463, + 0.893, + 0.629 + ], + "angle": 0, + "content": "(super-)category, such as human, animal, or vehicle. They are inefficient in the sense that they need to train a separate model for each application. Moreover, annotating a large number of samples for some categories, e.g., rare species, is too expensive, if not impossible, and thus hinders the usefulness of the deep learning approach. A unified model that can quickly adapt to keypoints of unseen categories is desirable. Addressing this, Xu et al. [42] propose the task of category-agnostic pose estimation (CAPE) in which a unified model is trained to detect the poses of objects from any category with limited supervision." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.63, + 0.895, + 0.87 + ], + "angle": 0, + "content": "Previous works attempts to compute keypoint prototypes from support and query samples and use them to locate keypoint positions on the query image via a matching step. Xu et al. [42] introduce POMNet, a transformer-based network, to capture the interactions between support and query images into the keypoint prototypes. POMNet is prone to overfitting since the keypoint prototypes heavily rely on very few, typically one or five, support samples. Recently, Shi et al. [31] propose a two-stage framework, called CapeFormer, to further calibrate the matching results. Although effective, CapeFormer uses support keypoint identifier to alleviate the ambiguity of keypoints with similar appearances. The keypoint identifier requires additional effort to consistently annotate the keypoints of different categories. This might prohibit their application in scenarios where keypoint identifiers are not available." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.871, + 0.894, + 0.902 + ], + "angle": 0, + "content": "In this work, we aim to improve the representation capacity of the keypoint prototypes by leveraging the knowl-" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.887, + 0.272, + 0.9 + ], + "angle": 0, + "content": "*Work done as an intern at NUS." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "23491" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.473, + 0.514 + ], + "angle": 0, + "content": "edge learned from the base categories. We propose a Bayesian framework that models a prior over the keypoint prototypes. This allows us to analytically encode the knowledge from the base categories and explicitly transfer it to the novel ones. Specifically, we introduce the notion of super-keypoint, a high-level abstract keypoint that models the distribution of semantically related keypoints from different categories. This is based on the observation that keypoints from different categories can be closely related. For example, keypoints corresponding to the left knee of many animals are visually similar and can be grouped into a super-keypoint 'left knee'. Encoding the super-keypoint 'left knee' from many base animals can help generalize to the left knee of an unseen animal (see Fig. 1). The super-keypoints are discovered and estimated automatically from the keypoints of base categories, whose keypoint distributions are well estimated from the abundant number of samples. The adaptation to an unseen category involves two steps: first, we match each novel keypoint to a related super-keypoint; and second, we estimate the keypoint prototypes by considering both the features from the support samples and the prior knowledge encoded in the matched super-keypoints. The first step requires finding an optimal bipartite matching between the set of novel keypoints and the learned super-keypoints. To solve this, we leverage a learnable matching network to capture the structural relations between the novel keypoints and the super-keypoints before making predictions." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.515, + 0.473, + 0.667 + ], + "angle": 0, + "content": "We call our framework Encoding Super-keypoints for Category-Agnostic Pose Estimation, dubbed ESCAPE. ESCAPE prevents the estimated keypoints from overfitting to the sparse support samples by allowing knowledge from the abundant base categories to be explicitly transferred to the novel one during adaptation. Furthermore, ESCAPE does not use keypoint identifiers to alleviate keypoint ambiguity. Experiments show that our method outperform state-of-the-art methods on the standard MP-100 benchmark, assuming keypoint identifiers are not available." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.68, + 0.223, + 0.695 + ], + "angle": 0, + "content": "2. Related works" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.705, + 0.473, + 0.903 + ], + "angle": 0, + "content": "2D pose estimation. 2D pose estimation aims to estimate the positions of a set of semantic keypoints of an object from its image. This task holds significant importance across various applications, such as human pose estimation for understanding human behaviors [19, 24, 40, 41], animal pose estimation for wildlife conservation [11, 17, 20, 46], and more. Existing methods can be roughly divided into regression-based methods [3, 26, 34, 35, 37, 39] and heatmap-based methods [1, 2, 4, 5, 18, 33]. The former directly predicts the coordinates of keypoints, while the latter infers heatmaps and takes the peak positions as it predictions. However, these methods are primarily designed for estimating poses for a single (super-)category with a fixed" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.892, + 0.122 + ], + "angle": 0, + "content": "set of semantic keypoints, and they typically require access to large-scale datasets." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.139, + 0.895, + 0.442 + ], + "angle": 0, + "content": "Few-shot learning. Few-shot learning focuses on learning novel concepts from few examples [14, 45]. While few-shot learning has been extensively studied in domains such as image classification [16] and video classification [25], its application in 2D pose estimation remains relatively under-explored. Our notion of super-keypoints is related to the Bayesian approach in few-shot learning literature [29, 43, 48]. Salakhutdinov et al. [29] introduce super-categories that play role as higher-level abstract concepts for categories and encode distributions over category parameters. Zhang et al. [48] design a Bayesian framework to learn a single super-category that induces a prior over parameters for all categories; such an approach limits model flexibility and is not suitable for CAPE due to the diversity of keypoints. Yang et al. [43] observe that similar categories have nearly identical statistics and propose enhancing novel categories with similar base categories. However, it is not trivial to adopt this method for CAPE due to the lack of a structural way to measure similarity between categories with semantically and quantitatively different keypoints." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.459, + 0.895, + 0.701 + ], + "angle": 0, + "content": "Semantic correspondence. Semantic correspondence refers to the semantic matching between pixels or keypoints between images [6]. A semantic correspondence model typically includes a feature extractor that generates dense features from images and establishes the correspondence using a similarity function. Oquab et al. [27] show that self-supervised ViT models can produce expressive features for finding semantic correspondence. Recently, Tang et al. [36] observe that semantic correspondence naturally emerges as an ability of diffusion models. These models can effectively solve CAPE by matching query pixels to keypoints in support images. Since ESCAPE operates in the feature space, it complements semantic correspondence models. Integrating ESCAPE with an off-the-shelf feature extractor consistently enhances its performance in CAPE tasks." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.713, + 0.629, + 0.73 + ], + "angle": 0, + "content": "3. Background" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.738, + 0.8, + 0.755 + ], + "angle": 0, + "content": "3.1. Category-agnostic pose estimation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.761, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Category-agnostic pose estimation (CAPE) [42] seeks an unified model to estimate poses from any category given a few annotated samples. Each category \\(y\\) has a set of \\(T_{y}\\) keypoints of interest: \\(J_{y} = \\{j_{r}\\}_{r = 1}^{T_{y}}\\). Given an image \\(\\mathbf{I}\\) of \\(y\\), the model's objective is to accurately predict the coordinates \\(P = \\{\\mathbf{p}_r\\}_{r = 1}^{T_y}\\) of these keypoints. The setting of CAPE consists of two disjoint sets of categories that are \\(C_{base}\\) for training the model and \\(C_{novel}\\) for testing its generalization capabilities. While the base categories have abundant" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "23492" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.473, + 0.244 + ], + "angle": 0, + "content": "training data, the novel categories have very few training samples. Specifically, a training set \\(\\mathcal{D}_{base}^{y} = \\{(I^{i},P^{i})\\}_{i = 1}^{n_{y}}\\) (\\(n_{y}\\) is large) of each base category \\(y\\) from \\(C_{base}\\) is available for training the model. The testing stage includes a large number of episodes (or tasks) sampled from the novel categories. In each episode, a support set \\(S_{novel}^{y} = \\{(\\mathbf{I}^{i},P^{i})\\}_{i = 1}^{K}\\) of \\(K\\) annotated samples from \\(y\\in C_{novel}\\) is given to the model for adaptation, and the model performance is evaluated on a set \\(\\mathcal{Q}_{novel}^{y} = \\{(\\mathbf{I}^{i},P^{i})\\}_{i = K + 1}^{K + Q}\\) of \\(Q\\) query samples." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.26, + 0.473, + 0.442 + ], + "angle": 0, + "content": "Support keypoint identifier. The MP-100 dataset is composed of existing widely used datasets, many of which impose an ordering over the keypoints of their categories. Shi et al [31] refer to this ordering information as support keypoint identifier and encode it into the keypoint features to alleviate the ambiguity of novel keypoints. However, incorporating keypoint identifiers may limit the applicability of CAPE to real-world scenarios where novel keypoints can be freely defined. In this paper, we focus on the challenging yet practical settings for CAPE that do not assume the order of keypoints to be available. We refer readers to [13] for detailed discussion of the support keypoint identifier." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.448, + 0.315, + 0.465 + ], + "angle": 0, + "content": "3.2. Simple baseline for CAPE" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.472, + 0.47, + 0.562 + ], + "angle": 0, + "content": "In this section, we introduce a simple yet effective baseline for CAPE. The baseline includes a feature extractor \\( f_{\\theta} \\) that extracts dense features from input images. We denote the local feature extracted at a position \\( \\mathbf{p} \\) of an input image \\( \\mathbf{I} \\) as \\( f_{\\theta}(\\mathbf{I})[\\mathbf{p}] \\in \\mathbb{R}^{d} \\). We aim to train \\( f_{\\theta} \\) so that it possesses strong semantic correspondence for the keypoints of interest." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.563, + 0.47, + 0.626 + ], + "angle": 0, + "content": "For each base category \\(y\\) with keypoints of interest \\(J_{y}\\) we define a set of prototypes \\(\\mathbf{W}^y = \\{\\mathbf{w}_r\\}_{r = 1}^{T_y}\\); each prototype \\(\\mathbf{w}_r\\in \\mathbb{R}^d\\) is responsible for predicting a heatmap \\(\\mathbf{H}_r\\) for the \\(r^{\\mathrm{th}}\\) keypoint in \\(J_{y}\\) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.19, + 0.632, + 0.469, + 0.649 + ], + "angle": 0, + "content": "\\[\n\\mathbf {H} _ {r} [ \\mathbf {p} ] = \\left\\langle f _ {\\theta} (\\mathbf {I}) [ \\mathbf {p} ], \\mathbf {w} _ {r} \\right\\rangle , \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.654, + 0.47, + 0.699 + ], + "angle": 0, + "content": "where \\(\\langle \\cdot, \\cdot \\rangle\\) denotes the inner product, and \\(\\mathbf{p}\\) varies over the spatial dimensions. The peak position in the heatmap is then the prediction for that keypoint." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.7, + 0.469, + 0.774 + ], + "angle": 0, + "content": "We train the feature extractor \\( f_{\\theta} \\) and the base prototypes \\( \\{\\mathbf{W}^y\\}_{y \\in C_{base}} \\) on the base training data \\( \\bigcup_{y \\in C_{base}} \\{\\mathcal{D}_{base}^y\\} \\) with the standard Mean Squared Error (MSE) loss [23, 38, 41] between the predicted heatmaps and the ground-truth heatmaps." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.775, + 0.47, + 0.854 + ], + "angle": 0, + "content": "To adapt to a novel category \\(y \\in C_{\\text{novel}}\\) with support and query sets \\((S_{\\text{novel}}^y, Q_{\\text{novel}}^y)\\), we freeze the feature extractor and compute the prototypes \\(\\mathbf{W}^y = \\{\\mathbf{w}_r\\}_{r=1}^{T_y}\\) for novel keypoints as the means of keypoint features extracted from the support set:" + }, + { + "type": "equation", + "bbox": [ + 0.158, + 0.858, + 0.47, + 0.898 + ], + "angle": 0, + "content": "\\[\n\\mathbf {w} _ {r} = \\frac {1}{K} \\sum_ {\\left(\\mathbf {I} ^ {i}, P ^ {i}\\right) \\in \\mathcal {S} _ {n o v e l} ^ {y}} f _ {\\theta} \\left(\\mathbf {I} ^ {i}\\right) \\left[ \\mathbf {p} _ {r} ^ {i} \\right]. \\tag {2}\n\\]" + }, + { + "type": "image", + "bbox": [ + 0.523, + 0.09, + 0.693, + 0.19 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.573, + 0.193, + 0.628, + 0.204 + ], + "angle": 0, + "content": "(a) zebra" + }, + { + "type": "image", + "bbox": [ + 0.702, + 0.09, + 0.871, + 0.19 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.757, + 0.193, + 0.81, + 0.204 + ], + "angle": 0, + "content": "(b) horse" + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.216, + 0.894, + 0.312 + ], + "angle": 0, + "content": "Figure 2. t-SNE visualization of distributions of 7 keypoints of zebra and horse. Different colors represent different keypoints. \\(\\times\\) denotes keypoint features extracted from images. (a) shows extracted features (denoted as \\(\\times\\)) and distribution means (denoted as \\(\\triangle\\)) for keypoints of zebra. (b) shows extracted features (denoted as \\(\\times\\)) and distribution means (denoted as \\(\\circ\\)) for keypoints of horse." + }, + { + "type": "image", + "bbox": [ + 0.534, + 0.322, + 0.861, + 0.419 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.428, + 0.894, + 0.485 + ], + "angle": 0, + "content": "Figure 3. Super-keypoint modeling. Keypoint 'right back paw of zebra' and keypoint 'right back paw of horse' belong to the same super-keypoint 'right back paw'. The keypoints distribute closely around their super-keypoint." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.512, + 0.89, + 0.542 + ], + "angle": 0, + "content": "The computed prototypes are then used to predict keypoint heatmaps for query images as in Eq. 1." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.543, + 0.892, + 0.694 + ], + "angle": 0, + "content": "Our simple baseline provides results comparable to POMNet [42], which has more complicated matching network. The baseline, however, has the same problem of overfitting as POMNet and CapeFormer due to its heavy reliance on the sparse support samples. Building on our simple baseline, we propose a Bayesian framework that allows us to estimate better prototypes for the novel keypoints, while keeping the feature extractor the same. Specifically, we seek to equip the baseline with strong a prior to prevent the keypoint prototypes from overfitting to the support set." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.709, + 0.836, + 0.727 + ], + "angle": 0, + "content": "4. Prior as a mixture of super-keypoints" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.735, + 0.892, + 0.901 + ], + "angle": 0, + "content": "We notice that keypoints of different categories can be visually similar, and hence their features should follow the same distribution. To verify our hypothesis, we use the feature extractor from our baseline model to extract features at the positions of seven keypoints in images of 'zebra' (a base category) and 'horse' (an unseen category). Fig. 2 (a) and (b) show keypoints features extracted from zebra and horse images respectively. For each keypoint, we also visualize the mean of its distribution (denoted as \\(\\triangle\\) for keypoint of zebra and \\(\\bigcirc\\) for keypoint of horse). As shown in the figure, although the model has not seen any horse sample during" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "23493" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.213 + ], + "angle": 0, + "content": "training, the keypoint features of horse still align well with those of zebra. This alignment suggests that capturing the distributions of zebra keypoints facilitates generalization to horse keypoints. From this observation, we introduce super-keypoints to model the connections between keypoints of horse and zebra. For instance, the keypoint 'right back paw of zebra' and the keypoint 'right back paw of horse' belong to the same super-keypoint 'right back paw'." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.223, + 0.317, + 0.24 + ], + "angle": 0, + "content": "4.1. Super-keypoints modeling" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.246, + 0.473, + 0.413 + ], + "angle": 0, + "content": "We introduce the notion of super-keypoint to define a prior over a group of visually similar keypoints from different categories. Let us consider the \\( r^{\\mathrm{th}} \\) keypoint of a category \\( y \\). Given an image \\( \\mathbf{I} \\) of \\( y \\), we denote \\( \\mathbf{o}_r \\coloneqq f_\\theta(\\mathbf{I})[\\mathbf{p}_r] \\) as the local feature extracted at the position of the \\( r^{\\mathrm{th}} \\) keypoint in \\( \\mathbf{I} \\). We assume that the keypoint feature \\( \\mathbf{o}_r \\) follows a Gaussian distribution with a mean \\( \\zeta_r \\) and covariance \\( \\phi^2 I \\). We further assume that the mean of the keypoint distribution \\( \\zeta_r \\) obeys a Gaussian distribution with a mean \\( \\mu_{z_r} \\) and covariance \\( \\sigma^2 I \\), where \\( \\mu_{z_r} \\) represent the super-keypoint of the \\( r^{\\mathrm{th}} \\) keypoint. Formally," + }, + { + "type": "equation", + "bbox": [ + 0.104, + 0.424, + 0.47, + 0.443 + ], + "angle": 0, + "content": "\\[\n\\mathbf {o} _ {r} \\sim \\mathcal {N} (\\mathbf {o} | \\zeta_ {r}; \\phi^ {2}), \\quad \\text {a n d} \\quad \\zeta_ {r} \\sim \\mathcal {N} (\\zeta | \\mu_ {z _ {r}}; \\sigma^ {2}). \\qquad (3)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.454, + 0.473, + 0.621 + ], + "angle": 0, + "content": "Fig. 3 illustrates our super-keypoint modeling. Intuitively, \\(\\mu\\) is the representation of a super-keypoint in the feature space, and its keypoint members distribute closely nearby. Our goal is to discover and estimate a set of superkeypoints from keypoints of base categories so that the super-keypoints can encode the statistics of groups of related keypoints across different categories. When adapting to a novel category, the statistics encoded in the superkeypoints are directly transferred to the prototypes of novel keypoints, preventing our model from overfitting to the sparse support set." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.631, + 0.433, + 0.648 + ], + "angle": 0, + "content": "4.2. Super-keypoints discovering and learning" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.654, + 0.47, + 0.745 + ], + "angle": 0, + "content": "We obtain the set of super-keypoints by applying a clustering method to the keypoints of all base categories. Specifically, given the trained feature extractor \\( f_{\\theta} \\) from the baseline, we extract the features of keypoints from all base categories. We then estimate the distribution mean of each keypoint in the base categories as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.145, + 0.756, + 0.47, + 0.797 + ], + "angle": 0, + "content": "\\[\n\\zeta_ {r} ^ {y} = \\frac {1}{| \\mathcal {D} _ {\\text {b a s e}} ^ {y} |} \\sum_ {(\\mathbf {I} ^ {i}, P ^ {i}) \\in \\mathcal {D} _ {\\text {b a s e}} ^ {y}} f _ {\\theta} (\\mathbf {I} ^ {i}) [ \\mathbf {p} _ {r} ^ {i} ], \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.808, + 0.471, + 0.903 + ], + "angle": 0, + "content": "where \\(\\zeta_r^y\\) represents the distribution mean of the \\(r^{\\mathrm{th}}\\) keypoint of category \\(y\\in C_{base}\\). The set of keypoints from all base categories is then \\(\\bigcup_{y\\in C_{base}}\\{\\zeta_r^y\\}_{r = 1}^{T_y}\\). In this work, we use the k-nearest neighbor based density peaks clustering [9, 47], which does not require the number of clusters to be known in advance. Note that keypoints of the same" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.893, + 0.184 + ], + "angle": 0, + "content": "category are distinct, so they cannot belong to the same super-keypoint. A detailed description of the clustering is provided in the Appendix. Finally, we estimate each super-keypoint representation \\(\\mu\\) as the empirical mean of distribution means of keypoints within the cluster, resulting in a set of \\(L\\) super-keypoints: \\(\\mathbf{M} = \\{\\mu_1,\\mu_2,\\dots ,\\mu_L\\}\\)." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.199, + 0.746, + 0.217 + ], + "angle": 0, + "content": "5. Novel category adaptation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.225, + 0.895, + 0.407 + ], + "angle": 0, + "content": "In this section, we describe how to adapt to a novel category \\(y\\) whose support and query sets are \\((\\mathcal{S}_{novel}^{y},\\mathcal{Q}_{novel}^{y})\\). Our goal is to estimate the prototypes that well represent the distributions of the novel keypoints from a few support samples. In particular, for the \\(r^{\\mathrm{th}}\\) keypoint of \\(y\\), we want the estimated prototype \\(\\mathbf{w}_r\\) to be close to the true distribution mean \\(\\zeta_r\\) as much as possible. With this interpretation, the estimated prototype for a novel keypoint in the baseline (Eq. 2) can be viewed as maximum likelihood estimate of the mean of the keypoint distribution. It is clear that the baseline is prone to overfitting since the number of support samples is extremely limited." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.419, + 0.884, + 0.436 + ], + "angle": 0, + "content": "5.1. Prototype estimate via maximum a posteriori" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.444, + 0.894, + 0.52 + ], + "angle": 0, + "content": "We propose to alleviate the overfitting by letting the statistics captured in the super-keypoints to be transferred to the novel keypoints. Specifically, consider the \\( r^{\\mathrm{th}} \\) keypoint of \\( y \\), we compute its prototype \\( \\mathbf{w}_r \\) as maximum a posterior estimate of \\( \\zeta_r \\):" + }, + { + "type": "equation", + "bbox": [ + 0.556, + 0.531, + 0.892, + 0.561 + ], + "angle": 0, + "content": "\\[\n\\mathbf {w} _ {r} ^ {M A P} = \\underset {\\zeta_ {r}} {\\operatorname {a r g m a x}} \\log p \\left(\\zeta_ {r} \\mid \\left\\{\\mathbf {o} _ {r} ^ {i} \\right\\} _ {i = 1} ^ {K}, \\mathbf {M}\\right), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.573, + 0.892, + 0.62 + ], + "angle": 0, + "content": "where \\(\\mathbf{o}_r^i = f_\\theta (\\mathbf{I}^i)[\\mathbf{p}_r^i ]\\) is the feature of the \\(r^{\\mathrm{th}}\\) keypoint extracted from the \\(i^{\\mathrm{th}}\\) support sample in \\(\\mathcal{S}_{novel}^{y}\\), and \\(\\mathbf{M} = \\{\\mu_1,\\mu_2,\\dots ,\\mu_L\\}\\) is the set of learned super-keypoints." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.621, + 0.895, + 0.788 + ], + "angle": 0, + "content": "However, when adapting to a novel category, the superkeypoints of the novel keypoints are not available in general. We do not know which super-keypoint the \\( r^{\\mathrm{th}} \\) novel keypoint should belong to. To this extent, we adopt the expectation maximization (EM) algorithm [21], which first matches the \\( r^{\\mathrm{th}} \\) keypoint with the 'best' super-keypoint from M (the expectation step) and then optimizes for the prototype with the prior encoded in the matched super-keypoint (the maximization step). Below, we summarize the two steps; detailed derivation of the (EM) algorithm is provided in the Appendix." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.811, + 0.894, + 0.903 + ], + "angle": 0, + "content": "Expectation step. Let \\( z_{r} \\) be the super-keypoint assignment of the \\( r^{\\mathrm{th}} \\) keypoint of \\( y \\). We align the keypoint with the most likely super-keypoint \\( z_{r}^{\\star} \\), i.e., the \\( z_{r} \\) that maximizes \\( \\log p(\\zeta_{r}, z_{r}|\\{\\mathbf{o}_{r}^{i}\\}_{i=1}^{K}, \\mathbf{M}) \\). Note that keypoints of the same category cannot belong to the same super-keypoint. Therefore, we further adopt the Hungarian algorithm [15]" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "23494" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.14, + 0.087, + 0.837, + 0.403 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.414, + 0.895, + 0.485 + ], + "angle": 0, + "content": "Figure 4. Illustration of our adaptation procedure for a novel category. We leverage the EM algorithm to probabilistically optimize for the keypoint prototypes, considering both the support samples and the learned super-keypoints. We first align the novel keypoints with the different super-keypoints via a learnable matching network (expectation step). We then let the statistics encoded in the matched super-keypoints to be transferred directly to the keypoint prototypes (maximization step). The keypoint prototypes can be used to estimate poses for query samples." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.511, + 0.471, + 0.56 + ], + "angle": 0, + "content": "to optimally align the \\(T_{y}\\) keypoints of \\(y\\) with distinct superkeypoints \\(\\{\\mu_{z_r^*}\\}_{r = 1}^{T_y}\\subset \\mathbf{M}\\). We refer to this matching procedure as simple matching for the rest of the paper." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.578, + 0.47, + 0.625 + ], + "angle": 0, + "content": "Maximization step. Now, we know the 'best' superkeypoint for each keypoint of \\( y \\). We can compute their prototypes analytically as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.13, + 0.649, + 0.47, + 0.72 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathbf {w} _ {r} ^ {M A P} = \\underset {\\mathbf {w} _ {r}} {\\operatorname {a r g m a x}} \\log p \\left(\\mathbf {w} _ {r} | \\left\\{\\mathbf {o} _ {r} ^ {i} \\right\\} _ {i = 1} ^ {K}, \\mu_ {z _ {r} ^ {\\star}}\\right) (6) \\\\ = \\alpha \\frac {1}{K} \\sum_ {i = 1} ^ {K} \\mathbf {o} _ {r} ^ {i} + (1 - \\alpha) \\mu_ {z _ {r} ^ {\\star}}, (7) \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.732, + 0.471, + 0.824 + ], + "angle": 0, + "content": "where \\(\\alpha \\in (0,1)\\) depends on \\(\\phi\\) and \\(\\sigma\\). Since the distribution mean of a keypoint must distribute near its super-keypoint, the second term in Eq. 7 has an effect of pulling the estimated prototype closer to the true mean of the keypoint distribution, diminishing the dependence of the prototype on the support samples." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.832, + 0.338, + 0.849 + ], + "angle": 0, + "content": "5.2. Learnable matching network" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Matching novel keypoints with semantically related superkeypoints is crucial for the success of our method. A correct super-keypoint can pull the estimated prototype closer" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.511, + 0.892, + 0.648 + ], + "angle": 0, + "content": "to the distribution mean of the keypoint under consideration, while an incorrect super-keypoint pulls the prototype away. However, as later shown in the experiments, the simple matching in the expectation step fails to produce reliable super-keypoints assignments. We argue that this is because the simple matching treats the novel keypoints independently, discarding their structural relations. It ignores the keypoint coordinates in the support samples, and so the model cannot reason about poses of the object of interest." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.648, + 0.893, + 0.755 + ], + "angle": 0, + "content": "To overcome these limitations, we replace the simple matching in the expectation step with a learnable matching network to simultaneously produce matching scores between keypoints in an image and the learned super keypoints. The design is inspired by SuperGlue [30], a matching framework that has shown great performance in aligning two sequences of features." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.781, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Input features. Our matching network takes two sequences of features as input and produces the matching score matrix for their elements. The first sequence is augmented features of \\( T_{y} \\) keypoints extracted from a support image of \\( y \\). The second sequence is a set of features for the \\( L \\) super-keypoints. To encode features for the keypoints, we consider not only their visual appearance but also their coordinates within the support image. Specifically, given" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.958 + ], + "angle": 0, + "content": "23495" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.123, + 0.089, + 0.849, + 0.187 + ], + "angle": 0, + "content": "
MethodSplit 1Split 2Split 3Split 4Split 5Mean PCKSplit 1Split 2Split 3Split 4Split 5Mean PCK
ProtoNet [32]46.0540.8449.1343.3444.5444.7860.3153.5161.9258.4458.6158.56
MAML [10]68.1454.7264.1963.2457.2061.5070.0355.9863.2164.7958.4762.50
Finetune [22]70.6057.0466.0665.0059.2063.5871.6757.8466.7666.5360.2464.61
POMNet [42]84.2378.2578.1778.6879.1779.7084.7279.6178.0080.3880.8580.71
CapeFormer [31]84.1379.9079.5879.7880.3480.7590.4686.4785.8086.4786.7187.19
ESCAPE86.8982.5581.2581.7281.3282.7491.4187.4385.3387.2786.7687.63
" + }, + { + "type": "table_caption", + "bbox": [ + 0.098, + 0.199, + 0.87, + 0.213 + ], + "angle": 0, + "content": "Table 1. Comparison with state-of-the-art works on MP-100 dataset under 1-shot and 5-shot settings. \\( {}^{ \\dagger } \\) denotes keypoint identifiers." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.24, + 0.423, + 0.258 + ], + "angle": 0, + "content": "support image \\(\\mathbf{I}\\sim S_{novel}^{y}\\) , the keypoint features are:" + }, + { + "type": "equation", + "bbox": [ + 0.088, + 0.266, + 0.469, + 0.284 + ], + "angle": 0, + "content": "\\[\n\\mathbf {u} _ {r} = f _ {\\gamma} (\\mathbf {I}) \\left[ \\mathbf {p} _ {r} \\right] + P E _ {\\mathrm {g e o}} \\left(\\mathbf {p} _ {r}\\right), \\text {f o r} r = \\{1, \\dots , T _ {y} \\}, \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.293, + 0.469, + 0.354 + ], + "angle": 0, + "content": "where \\( f_{\\gamma} \\) is a CNN feature extractor, and \\( PE_{\\mathrm{geo}} \\) embeds a coordinate into the feature space. By considering the keypoint coordinates, our model can reason about the pose of the object in the support image." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.354, + 0.469, + 0.383 + ], + "angle": 0, + "content": "At the other end, the features for the super-keypoints are defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.124, + 0.395, + 0.469, + 0.412 + ], + "angle": 0, + "content": "\\[\n\\mathbf {v} _ {z} = \\omega_ {z} + P E _ {\\mathrm {a b s}} (z), \\quad \\text {f o r} z \\in \\{1, \\dots , L \\}, \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.423, + 0.469, + 0.484 + ], + "angle": 0, + "content": "where \\(\\omega_{z}\\) is a learnable feature vector for the \\(z^{\\mathrm{th}}\\) superkeypoint, and \\(PE_{\\mathrm{abs}}\\) embeds an index \\((1,\\dots ,L)\\) into the feature space. Note that the indexes of the super-keypoints are obtained from the clustering step described in Sec. 4.2." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.502, + 0.47, + 0.673 + ], + "angle": 0, + "content": "Feature aggregation with attention. The two sequences of features \\(\\{\\mathbf{u}_r\\}_{r=1}^{T_y}\\) and \\(\\{\\mathbf{v}_z\\}_{z=1}^L\\) are then passed to a neural network for information aggregation. The network comprises \\(U\\) blocks of self-attention and cross-attention to model interactions both within and between the elements of the sequences. Results are two sequences of refined features: \\(\\{\\mathbf{u}_r^\\star\\}_{r=1}^{T_y}\\) and \\(\\{\\mathbf{v}_z^\\star\\}_{z=1}^L\\). The predicted matching scores matrix is then \\(\\mathbf{S} \\in \\mathbb{R}^{T_y \\times L}\\), with each element \\(\\mathbf{S}_{r,z} = \\langle \\mathbf{u}_r^\\star, \\mathbf{v}_z^\\star \\rangle\\) measuring the matching score between the \\(r^{\\text{th}}\\) keypoint and the \\(z^{\\text{th}}\\) super-keypoint. An illustration of the matching network is provided at the bottom of Fig. 4." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.69, + 0.47, + 0.902 + ], + "angle": 0, + "content": "Matching as optimal transport problem. Since the keypoints within the same category must be matched to different super-keypoints, our goal is to find a maximum score bipartite matching between the two sequences. Following SuperGlue [30], we replace the Hungarian algorithm with the differentiable Sinkhorn algorithm [7] for optimal transport. It takes the matching score matrix \\(\\mathbf{S}\\) as input and produces a matrix \\(\\mathbf{Q} \\in \\mathbb{R}^{T_y \\times L}\\) where each element \\(\\mathbf{Q}_{r,z}\\) represents the probability that the \\(r^{\\text{th}}\\) keypoint is matched with the \\(z^{\\text{th}}\\) super-keypoint. We train the network end-to-end on the keypoints of base categories, whose super-keypoints assignments are available from the clustering step. Details of the Sinkhorn algorithm and the training procedure of the matching network are provided in the Appendix." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.24, + 0.892, + 0.36 + ], + "angle": 0, + "content": "Incorporating learnable matching network to novel category adaptation. We integrate the learned matching network into the novel category adaptation. Specifically, for every support image \\(\\mathbf{I}^i\\in S_{novel}^y\\), we compute the matching score matrix \\(\\mathbf{S}^i\\) from its keypoints and the learned super keypoints. The corresponding optimal assignment matrix is then denoted as \\(\\mathbf{Q}^i\\). The final prototype for the \\(r^{\\mathrm{th}}\\) keypoint are then estimated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.516, + 0.371, + 0.892, + 0.427 + ], + "angle": 0, + "content": "\\[\n\\mathbf {w} _ {r} ^ {\\text {E S C A P E}} = \\alpha \\frac {1}{K} \\sum_ {i = 1} ^ {K} \\mathbf {o} _ {r} ^ {i} + (1 - \\alpha) \\frac {L}{K} \\sum_ {i = 1} ^ {K} \\sum_ {z = 1} ^ {L} \\mathbf {Q} _ {r, z} ^ {i} \\mu_ {z}. \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.443, + 0.892, + 0.489 + ], + "angle": 0, + "content": "Here we keep the soft assignment form of \\(\\mathbf{Q}^i\\) instead of the hard assignment to model uncertainty. Our adaptation procedure is shown in Fig. 4." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.505, + 0.633, + 0.523 + ], + "angle": 0, + "content": "6. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.531, + 0.892, + 0.698 + ], + "angle": 0, + "content": "We perform experiments on Multi-category Pose (MP-100) proposed in [42]. The dataset consists of more than 20,000 data instances of 100 categories from 8 super-categories including human hand, human face, human body, animal body, animal face, clothes, furniture, and vehicle. We also follow the five splits from [42] for standard CAPE settings. Each split divides the dataset into 70 training categories, 10 validation categories, and 20 testing categories. The training set is used as base categories, while validation and testing sets are novel categories for validation and testing, respectively. Please refer to [42] for more details." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.71, + 0.696, + 0.726 + ], + "angle": 0, + "content": "6.1. Experimental details" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.734, + 0.892, + 0.824 + ], + "angle": 0, + "content": "Unless otherwise specified, for fair comparisons with previous works, we utilize ResNet-50 [12] pretrained on ImageNet [8] as the feature extractor. Also following [42], we crop the object of interest from the original image using the ground-truth bounding box, and resize it to a fixed size of \\( 256 \\times 256 \\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.825, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Probability of correct keypoint [44] with a threshold of 0.2 (PCK@0.2) is the evaluation metric for MP-100. During testing, we generate 60,000 episodes evenly distributed across the test categories and report the average PCK@0.2 over these episodes." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "23496" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.154, + 0.09, + 0.245, + 0.101 + ], + "angle": 0, + "content": "super-keypoint 1" + }, + { + "type": "image", + "bbox": [ + 0.1, + 0.102, + 0.3, + 0.164 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.154, + 0.166, + 0.246, + 0.177 + ], + "angle": 0, + "content": "super-keypoint 4" + }, + { + "type": "image", + "bbox": [ + 0.1, + 0.178, + 0.301, + 0.239 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.371, + 0.09, + 0.464, + 0.101 + ], + "angle": 0, + "content": "super-keypoint 2" + }, + { + "type": "image", + "bbox": [ + 0.318, + 0.102, + 0.383, + 0.153 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.321, + 0.154, + 0.382, + 0.163 + ], + "angle": 0, + "content": "arctic wolf" + }, + { + "type": "image", + "bbox": [ + 0.385, + 0.102, + 0.519, + 0.153 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.396, + 0.155, + 0.438, + 0.164 + ], + "angle": 0, + "content": "sealion" + }, + { + "type": "image_caption", + "bbox": [ + 0.468, + 0.155, + 0.5, + 0.164 + ], + "angle": 0, + "content": "ferret" + }, + { + "type": "image_caption", + "bbox": [ + 0.404, + 0.166, + 0.497, + 0.176 + ], + "angle": 0, + "content": "super-keypoint 5" + }, + { + "type": "image", + "bbox": [ + 0.318, + 0.177, + 0.384, + 0.228 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.336, + 0.23, + 0.369, + 0.24 + ], + "angle": 0, + "content": "grebe" + }, + { + "type": "image", + "bbox": [ + 0.387, + 0.177, + 0.452, + 0.228 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.391, + 0.23, + 0.446, + 0.24 + ], + "angle": 0, + "content": "kingfisher" + }, + { + "type": "image", + "bbox": [ + 0.453, + 0.177, + 0.585, + 0.23 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.468, + 0.23, + 0.498, + 0.24 + ], + "angle": 0, + "content": "wren" + }, + { + "type": "image_caption", + "bbox": [ + 0.508, + 0.23, + 0.56, + 0.24 + ], + "angle": 0, + "content": "gull" + }, + { + "type": "image_caption", + "bbox": [ + 0.657, + 0.09, + 0.748, + 0.101 + ], + "angle": 0, + "content": "super-keypoint 3" + }, + { + "type": "image", + "bbox": [ + 0.537, + 0.102, + 0.868, + 0.164 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.689, + 0.166, + 0.781, + 0.177 + ], + "angle": 0, + "content": "super-keypoint 6" + }, + { + "type": "image", + "bbox": [ + 0.601, + 0.177, + 0.67, + 0.229 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.602, + 0.23, + 0.671, + 0.251 + ], + "angle": 0, + "content": "long sleeved dress" + }, + { + "type": "image", + "bbox": [ + 0.673, + 0.177, + 0.868, + 0.229 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.69, + 0.23, + 0.713, + 0.24 + ], + "angle": 0, + "content": "vest" + }, + { + "type": "image_caption", + "bbox": [ + 0.74, + 0.23, + 0.868, + 0.251 + ], + "angle": 0, + "content": "vest dress long sleeved shirt" + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.262, + 0.892, + 0.277 + ], + "angle": 0, + "content": "Figure 5. Keypoints of four different super-keypoints discovered from the base categories. Below each keypoint is the name of its category." + }, + { + "type": "image", + "bbox": [ + 0.085, + 0.283, + 0.231, + 0.417 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.237, + 0.283, + 0.885, + 0.427 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.086, + 0.442, + 0.233, + 0.58 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.237, + 0.439, + 0.885, + 0.582 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.593, + 0.89, + 0.622 + ], + "angle": 0, + "content": "Figure 6. Qualitative comparison between simple matching and learnable matching network in aligning novel keypoints with superkeypoints. We show one base keypoint member of each matched super-keypoint." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.647, + 0.385, + 0.663 + ], + "angle": 0, + "content": "6.2. Comparison with previous methods" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.67, + 0.469, + 0.79 + ], + "angle": 0, + "content": "We compare ESCAPE against previous methods, including Prototypical Networks [32], MAML [10], Finetune [22], POMNet [32] and CapeFormer [31]. Note that we remove the support keypoint identifier from CapeFormer. The results are shown in Tab. 1. We can see that ESCAPE outperforms other methods by around \\(2.5\\%\\) and \\(0.5\\%\\) in 1-shot and 5-shot settings, respectively, establishing a new state-of-the-art on the MP-100 benchmark." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.802, + 0.227, + 0.818 + ], + "angle": 0, + "content": "6.3. Ablation study" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Super-keypoints discovering. We perform clustering on the keypoints of base categories to form super-keypoints. To verify that we can obtain meaningful super-keypoints from the clustering process, we show some keypoints members of a few super-keypoints in Fig. 5. We see that each super-" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.648, + 0.892, + 0.724 + ], + "angle": 0, + "content": "keypoint represents a meaningful concept. For example, super-keypoint 1 consists of 'left knee' of deer, antelope and zebra. super-keypoint 4 consists of 'right rear wheel' of SUV, bus, and car, while super-keypoint 6 is a landmark on the collars of different clothes categories." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Novel keypoints and super-keypoints matching. We qualitatively compare the performance of the simple matching and the learnable matching network in aligning novel keypoints with the learned super-keypoints. Fig. 6 shows the matching results for two novel categories, namely squirrel and woodpecker. The left side of the figure shows support images of the categories, while the right side shows the super-keypoints aligned by the simple matching and the learnable matching network. For each super-keypoint, we only show one base keypoints member. Compared with the" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "23497" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.08, + 0.091, + 0.245, + 0.157 + ], + "angle": 0, + "content": "
MethodPCK
MLE83.00 ± 0.16
MAP83.88 ± 0.15
ESCAPE86.89 ± 0.14
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.167, + 0.246, + 0.222 + ], + "angle": 0, + "content": "Table 2. Results of different prototype estimation methods, namely MLE, MAP and ESCAPE." + }, + { + "type": "table", + "bbox": [ + 0.26, + 0.089, + 0.472, + 0.159 + ], + "angle": 0, + "content": "
Feature extractorMethodPCK
DIFT [36]MLE60.61
DIFT [36]ESCAPE62.47
DINOv2 [27]MLE88.31
DINOv2 [27]ESCAPE89.14
" + }, + { + "type": "table_caption", + "bbox": [ + 0.258, + 0.17, + 0.47, + 0.224 + ], + "angle": 0, + "content": "Table 3. Results of integrating ESCAPE to feature extractors of different semantic correspondence models" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.26, + 0.47, + 0.381 + ], + "angle": 0, + "content": "simple matching, our learnable matching network provides more reliable super-keypoints for the novel keypoints. In particular, the simple matching aligns keypoints of squirrel with super-keypoints of keypoints of kingfisher, sling and suv, which are irrelevant to the squirrel. On the other hand, the learnable matching network aligns the novel keypoints with their actual concepts, e.g., the 'left elbow of squirrel' is matched with the super-keypoint of the 'left elbow of rat'." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.422, + 0.47, + 0.68 + ], + "angle": 0, + "content": "Ablation on novel category adaptation. In Tab. 2, we study the performance of different methods to estimate keypoint prototypes for a novel category, namely MLE, MAP and ESCAPE. MLE is our baseline, which estimate keypoint prototypes as means of support features as in Eq. 2. MAP and ESCAPE compute keypoint prototypes as maximum a posterior estimate with the EM algorithm. The difference between MAP and ESCAPE is that MAP (Eq. 7) uses the simple matching in the expectation step, whereas ESCAPE (Eq. 10) leverages the learnable matching network. We observe that incorporating super-keypoints into the estimation of keypoint prototypes helps improve the overall performance. Specifically, the MAP outperforms MLE by roughly \\(1\\%\\). Regardless ESCAPE, better novel keypoints and super-keypoints alignments from the learnable matching network immediately translate into \\(3\\%\\) improvements over the MAP counterpart." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.72, + 0.468, + 0.901 + ], + "angle": 0, + "content": "Ablation on different feature extractors. Feature extractors from semantic correspondence models can effectively solve CAPE since they can generate dense features for support and query images. In Tab. 3, we study the compatibility of ESCAPE with feature extractors from different semantic correspondence models, namely DIFT [36] and DINoV2 [27]. We take the pretrained stable diffusion 2.1 [28] as the backbone for DIFT, whereas we use ViT-ViT-S/14 for DINoV2. As seen, ESCAPE consistently boosts the performance of the considered semantic correspondence models. This shows that ESCAPE can benefit from future research for better representation." + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.089, + 0.89, + 0.322 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.5, + 0.334, + 0.892, + 0.363 + ], + "angle": 0, + "content": "Figure 7. We qualitatively compare pose estimation of Cape-Former and ESCAPE under 1-shot setting." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.393, + 0.677, + 0.409 + ], + "angle": 0, + "content": "6.4. Qualitative results" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.419, + 0.892, + 0.57 + ], + "angle": 0, + "content": "Figure 7 compares the pose estimation performance of CapeFormer and ESCAPE. Overall, ESCAPE gives more accurate pose predictions compared to CapeFormer. In particular, CapeFormer is overfitting to the support image of the sheep category. The predicted positions of the left front and back paws by CapeFormer are geometrically close to each other as they appear in the support image. In contrast, since ESCAPE allows learned knowledge about base animals to be transferred to the novel keypoints, it recovers the pose of the query sheep beautifully." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.589, + 0.618, + 0.605 + ], + "angle": 0, + "content": "7. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.617, + 0.892, + 0.798 + ], + "angle": 0, + "content": "In this paper, we address the task of category-agnostic pose estimation by introducing ESCAPE, a Bayesian framework that induces a prior over the keypoint features of different categories. This prior is a mixture of super-keypoints, with each super-keypoint encoding a distribution of a group of related keypoints. ESCAPE enables the transfer of knowledge from base categories to the novel ones, thereby preventing the query predictions from overfitting to sparse support samples. ESCAPE offers several advantages over previous works: it exhibits less overfitting to support samples, does not rely on support keypoint identifiers, and is model-agnostic." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.818, + 0.66, + 0.835 + ], + "angle": 0, + "content": "Acknowledgement" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.845, + 0.892, + 0.899 + ], + "angle": 0, + "content": "This research is supported in part by the National Research Foundation, Singapore under its Competitive Research Program Award NRF-CRP23-2019-0003." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "23498" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.197 + ], + "angle": 0, + "content": "[1] Yuanhao Cai, Zhicheng Wang, Zhengxiong Luo, Binyi Yin, Angang Du, Haoqian Wang, Xiangyu Zhang, Xinyu Zhou, Erjin Zhou, and Jian Sun. Learning delicate local representations for multi-person pose estimation. In European Conference on Computer Vision, pages 455-472. Springer, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.2, + 0.472, + 0.255 + ], + "angle": 0, + "content": "[2] Zhe Cao, Tomas Simon, Shih-En Wei, and Yaser Sheikh. Realtime multi-person 2d pose estimation using part affinity fields. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7291-7299, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.256, + 0.471, + 0.323 + ], + "angle": 0, + "content": "[3] Joao Carreira, Pulkit Agrawal, Katerina Fragkiadaki, and Jitendra Malik. Human pose estimation with iterative error feedback. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4733-4742, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.326, + 0.472, + 0.395 + ], + "angle": 0, + "content": "[4] Yilun Chen, Zhicheng Wang, Yuxiang Peng, Zhiqiang Zhang, Gang Yu, and Jian Sun. Cascaded pyramid network for multi-person pose estimation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7103-7112, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.397, + 0.47, + 0.478 + ], + "angle": 0, + "content": "[5] Bowen Cheng, Bin Xiao, Jingdong Wang, Honghui Shi, Thomas S Huang, and Lei Zhang. Higherhrnet: Scale-aware representation learning for bottom-up human pose estimation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5386-5395, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.48, + 0.47, + 0.534 + ], + "angle": 0, + "content": "[6] Christopher B Choy, JunYoung Gwak, Silvio Savarese, and Manmohan Chandraker. Universal correspondence network. Advances in neural information processing systems, 29, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.536, + 0.47, + 0.577 + ], + "angle": 0, + "content": "[7] Marco Cuturi. Sinkhorn distances: Lightspeed computation of optimal transport. Advances in neural information processing systems, 26, 2013. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.579, + 0.47, + 0.634 + ], + "angle": 0, + "content": "[8] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.635, + 0.47, + 0.689 + ], + "angle": 0, + "content": "[9] Mingjing Du, Shifei Ding, and Hongjie Jia. Study on density peaks clustering based on k-nearest neighbors and principal component analysis. Knowledge-Based Systems, 99: 135-145, 2016. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.691, + 0.47, + 0.745 + ], + "angle": 0, + "content": "[10] Chelsea Finn, Pieter Abbeel, and Sergey Levine. Model-agnostic meta-learning for fast adaptation of deep networks. In International conference on machine learning, pages 1126–1135. PMLR, 2017. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.747, + 0.47, + 0.802 + ], + "angle": 0, + "content": "[11] Jacob M Graving, Daniel Chae, Hemal Naik, Liang Li, Benjamin Koger, Blair R Costelloe, and Iain D Couzin. Deepposekit, a software toolkit for fast and robust animal pose estimation using deep learning. *Elife*, 8:e47994, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.804, + 0.47, + 0.858 + ], + "angle": 0, + "content": "[12] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.86, + 0.47, + 0.901 + ], + "angle": 0, + "content": "[13] Or Hirschorn and Shai Avidan. Pose anything: A graph-based approach for category-agnostic pose estimation. arXiv preprint arXiv:2311.17891, 2023. 3" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.147 + ], + "angle": 0, + "content": "[14] Timothy Hospedales, Antreas Antoniou, Paul Micaelli, and Amos Storkey. Meta-learning in neural networks: A survey. IEEE transactions on pattern analysis and machine intelligence, 44(9):5149-5169, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.149, + 0.892, + 0.188 + ], + "angle": 0, + "content": "[15] Harold W Kuhn. The hungarian method for the assignment problem. Naval research logistics quarterly, 2(1-2):83-97, 1955. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.19, + 0.892, + 0.258 + ], + "angle": 0, + "content": "[16] Duong Le, Khoi Duc Nguyen, Khoi Nguyen, Quoc-Huy Tran, Rang Nguyen, and Binh-Son Hua. Poodle: Improving few-shot learning via penalizing out-of-distribution samples. Advances in Neural Information Processing Systems, 34:23942-23955, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.261, + 0.892, + 0.302 + ], + "angle": 0, + "content": "[17] Shuyuan Li, Jianguo Li, Hanlin Tang, Rui Qian, and Weiyao Lin. Atrw: a benchmark for amur tiger re-identification in the wild. arXiv preprint arXiv:1906.05586, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.303, + 0.892, + 0.357 + ], + "angle": 0, + "content": "[18] Wenbo Li, Zhicheng Wang, Binyi Yin, Qixiang Peng, Yuming Du, Tianzi Xiao, Gang Yu, Hongtao Lu, Yichen Wei, and Jian Sun. Rethinking on multi-stage networks for human pose estimation. arXiv preprint arXiv:1901.00148, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.359, + 0.892, + 0.427 + ], + "angle": 0, + "content": "[19] Jiang-Jiang Liu, Qibin Hou, Ming-Ming Cheng, Changhu Wang, and Jiashi Feng. Improving convolutional networks with self-calibrated convolutions. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10096-10105, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.428, + 0.892, + 0.509 + ], + "angle": 0, + "content": "[20] Alexander Mathis, Thomas Biasi, Steffen Schneider, Mert Yuksekgonul, Byron Rogers, Matthias Bethge, and Mackenzie W Mathis. Pretraining boosts out-of-domain robustness for pose estimation. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1859-1868, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.511, + 0.892, + 0.538 + ], + "angle": 0, + "content": "[21] Kevin P. Murphy. Probabilistic Machine Learning: Advanced Topics. MIT Press, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.54, + 0.892, + 0.579 + ], + "angle": 0, + "content": "[22] Akihiro Nakamura and Tatsuya Harada. Revisiting fine-tuning for few-shot learning. arXiv preprint arXiv:1910.00216, 2019. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.581, + 0.892, + 0.65 + ], + "angle": 0, + "content": "[23] Alejandro Newell, Kaiyu Yang, and Jia Deng. Stacked hourglass networks for human pose estimation. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part VIII 14, pages 483-499. Springer, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.651, + 0.892, + 0.704 + ], + "angle": 0, + "content": "[24] Alejandro Newell, Zhiao Huang, and Jia Deng. Associative embedding: End-to-end learning for joint detection and grouping. Advances in neural information processing systems, 30, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.706, + 0.892, + 0.775 + ], + "angle": 0, + "content": "[25] Khoi D Nguyen, Quoc-Huy Tran, Khoi Nguyen, Binh-Son Hua, and Rang Nguyen. Inductive and transductive few-shot video classification via appearance and temporal alignments. In European Conference on Computer Vision, pages 471-487. Springer, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.777, + 0.892, + 0.831 + ], + "angle": 0, + "content": "[26] Xuecheng Nie, Jiashi Feng, Jianfeng Zhang, and Shuicheng Yan. Single-stage multi-person pose machines. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6951-6960, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.832, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[27] Maxime Oquab, Timothée Darcet, Théo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023. 2, 8" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "23499" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.161 + ], + "angle": 0, + "content": "[28] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.163, + 0.472, + 0.234 + ], + "angle": 0, + "content": "[29] Ruslan Salakhutdinov, Joshua Tenenbaum, and Antonio Torralba. One-shot learning with a hierarchical nonparametric bayesian model. In Proceedings of ICML Workshop on Unsupervised and Transfer Learning, pages 195-206. JMLR Workshop and Conference Proceedings, 2012. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.234, + 0.472, + 0.303 + ], + "angle": 0, + "content": "[30] Paul-Edouard Sarlin, Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. Superglue: Learning feature matching with graph neural networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4938–4947, 2020. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.305, + 0.472, + 0.375 + ], + "angle": 0, + "content": "[31] Min Shi, Zihao Huang, Xianzheng Ma, Xiaowei Hu, and Zhiguo Cao. Matching is not enough: A two-stage framework for category-agnostic pose estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7308-7317, 2023. 1, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.376, + 0.472, + 0.417 + ], + "angle": 0, + "content": "[32] Jake Snell, Kevin Swersky, and Richard Zemel. Prototypical networks for few-shot learning. Advances in neural information processing systems, 30, 2017. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.419, + 0.472, + 0.487 + ], + "angle": 0, + "content": "[33] Ke Sun, Bin Xiao, Dong Liu, and Jingdong Wang. Deep high-resolution representation learning for human pose estimation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5693-5703, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.489, + 0.472, + 0.545 + ], + "angle": 0, + "content": "[34] Xiao Sun, Jiaxiang Shang, Shuang Liang, and Yichen Wei. Compositional human pose regression. In Proceedings of the IEEE International Conference on Computer Vision, pages 2602-2611, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.547, + 0.472, + 0.602 + ], + "angle": 0, + "content": "[35] Xiao Sun, Bin Xiao, Fangyin Wei, Shuang Liang, and Yichen Wei. Integral human pose regression. In Proceedings of the European conference on computer vision (ECCV), pages 529-545, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.604, + 0.472, + 0.659 + ], + "angle": 0, + "content": "[36] Luming Tang, Menglin Jia, Qianqian Wang, Cheng Perng Phoo, and Bharath Hariharan. Emergent correspondence from image diffusion. Advances in Neural Information Processing Systems, 36:1363-1389, 2023. 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.661, + 0.472, + 0.702 + ], + "angle": 0, + "content": "[37] Zhi Tian, Hao Chen, and Chunhua Shen. Direct pose: Direct end-to-end multi-person pose estimation. arXiv preprint arXiv:1911.07451, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.704, + 0.472, + 0.759 + ], + "angle": 0, + "content": "[38] Jonathan J. Tompson, Arjun Jain, Yann LeCun, and Christoph Bregler. Joint training of a convolutional network and a graphical model for human pose estimation. Advances in neural information processing systems, 27, 2014. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.761, + 0.472, + 0.816 + ], + "angle": 0, + "content": "[39] Alexander Toshev and Christian Szegedy. Deeppose: Human pose estimation via deep neural networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1653-1660, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.818, + 0.472, + 0.9 + ], + "angle": 0, + "content": "[40] Jingdong Wang, Ke Sun, Tianheng Cheng, Borui Jiang, Chaorui Deng, Yang Zhao, Dong Liu, Yadong Mu, Mingkui Tan, Xinggang Wang, et al. Deep high-resolution representation learning for visual recognition. IEEE transactions on pattern analysis and machine intelligence, 43(10):3349-3364, 2020. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.092, + 0.894, + 0.147 + ], + "angle": 0, + "content": "[41] Bin Xiao, Haiping Wu, and Yichen Wei. Simple baselines for human pose estimation and tracking. In Proceedings of the European conference on computer vision (ECCV), pages 466-481, 2018. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.149, + 0.894, + 0.233 + ], + "angle": 0, + "content": "[42] Lumin Xu, Sheng Jin, Wang Zeng, Wentao Liu, Chen Qian, Wanli Ouyang, Ping Luo, and Xiaogang Wang. Pose for everything: Towards category-agnostic pose estimation. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23–27, 2022, Proceedings, Part VI, pages 398–416. Springer, 2022. 1, 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.234, + 0.894, + 0.275 + ], + "angle": 0, + "content": "[43] Shuo Yang, Lu Liu, and Min Xu. Free lunch for few-shot learning: Distribution calibration. arXiv preprint arXiv:2101.06395, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.277, + 0.894, + 0.33 + ], + "angle": 0, + "content": "[44] Yi Yang and Deva Ramanan. Articulated human detection with flexible mixtures of parts. IEEE transactions on pattern analysis and machine intelligence, 35(12):2878-2890, 2012. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.332, + 0.894, + 0.387 + ], + "angle": 0, + "content": "[45] Wang Yaqing, Yao Quanming, T Kwok James, and M Ni Lionel. Generalizing from a few examples: A survey on few-shot learning. ACM computing surveys, 53(3):1-34, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.389, + 0.894, + 0.444 + ], + "angle": 0, + "content": "[46] Hang Yu, Yufei Xu, Jing Zhang, Wei Zhao, Ziyu Guan, and Dacheng Tao. Ap-10k: A benchmark for animal pose estimation in the wild. arXiv preprint arXiv:2108.12617, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.446, + 0.894, + 0.529 + ], + "angle": 0, + "content": "[47] Wang Zeng, Sheng Jin, Wentao Liu, Chen Qian, Ping Luo, Wanli Ouyang, and Xiaogang Wang. Not all tokens are equal: Human-centric visual analysis via token clustering transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11101-11111, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.531, + 0.894, + 0.599 + ], + "angle": 0, + "content": "[48] Xueting Zhang, Debin Meng, Henry Gouk, and Timothy M Hospedales. Shallow bayesian meta learning for real-world few-shot recognition. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 651-660, 2021. 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.092, + 0.894, + 0.599 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "23500" + } + ] +] \ No newline at end of file diff --git a/2024/ESCAPE_ Encoding Super-keypoints for Category-Agnostic Pose Estimation/64b3bbb9-9580-410a-adfe-c8a2bd5b8bca_origin.pdf b/2024/ESCAPE_ Encoding Super-keypoints for Category-Agnostic Pose Estimation/64b3bbb9-9580-410a-adfe-c8a2bd5b8bca_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..d8fb9bc4b33c00e162d813669542ebf439213834 --- /dev/null +++ b/2024/ESCAPE_ Encoding Super-keypoints for Category-Agnostic Pose Estimation/64b3bbb9-9580-410a-adfe-c8a2bd5b8bca_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:cf6849ae4af2d4de3e63e3f6b1b528d71d144506420b7af39e1578bccdccb944 +size 7713097 diff --git a/2024/ESCAPE_ Encoding Super-keypoints for Category-Agnostic Pose Estimation/full.md b/2024/ESCAPE_ Encoding Super-keypoints for Category-Agnostic Pose Estimation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..da5e177b34ea22284b8a84990c0d2ad3fa9d0d86 --- /dev/null +++ b/2024/ESCAPE_ Encoding Super-keypoints for Category-Agnostic Pose Estimation/full.md @@ -0,0 +1,335 @@ +# ESCAPE: Encoding Super-keypoints for Category-Agnostic Pose Estimation + +Khoi Duc Nguyen1* +1University of Wisconsin-Madison +khoi.nguyen@wisc.edu lichen@u + +Chen Li $^{2}$ Gim Hee Lee $^{2}$ $^{2}$ National University of Singapore +.nus.edu gimhee.lee@comp.nus.edu. + +# Abstract + +In this paper, we tackle the task of category-agnostic pose estimation (CAPE), which aims to predict poses for objects of any category with few annotated samples. Previous works either rely on local matching between features of support and query samples or require support keypoint identifier. The former is prone to overfitting due to its sensitivity to sparse samples, while the latter is impractical for the open-world nature of the task. To overcome these limitations, we propose ESCAPE - a Bayesian framework that learns a prior over the features of keypoints. The prior can be expressed as a mixture of super-keypoints, each being a high-level abstract keypoint that captures the statistics of semantically related keypoints from different categories. We estimate the super-keypoints from base categories and use them in adaptation to novel categories. The adaptation to an unseen category involves two steps: first, we match each novel keypoint to a related super-keypoint; and second, we transfer the knowledge encoded in the matched super-keypoints to the novel keypoints. For the first step, we propose a learnable matching network to capture the relationship between the novel keypoints and the super-keypoints, resulting in a more reliable matching. ESCAPE mitigates overfitting by directly transferring learned knowledge to novel categories while it does not use keypoint identifiers. We achieve state-of-the-art performance on the standard MP-100 benchmark. Our code is available at https://github.com/khoiucd/escape-tgt. + +# 1. Introduction + +2D pose estimation aims to locate the semantic keypoints of an object of interest from an input image. This task is crucial for many visual understanding tasks, such as understanding human/animal behavior or interpreting the surrounding environment in autonomous driving. Existing works focus on detecting keypoints of only one specific + +![](images/c7d77b36b6441bbb24345f270f639eeb9aea1b2ddc508e68a47a3c9de239fe77.jpg) +Figure 1. We encode super-keypoint 'left knee' from the 'left knee' of base animals and use the super-keypoint in adaptation to the left knee of novel animal. + +(super-)category, such as human, animal, or vehicle. They are inefficient in the sense that they need to train a separate model for each application. Moreover, annotating a large number of samples for some categories, e.g., rare species, is too expensive, if not impossible, and thus hinders the usefulness of the deep learning approach. A unified model that can quickly adapt to keypoints of unseen categories is desirable. Addressing this, Xu et al. [42] propose the task of category-agnostic pose estimation (CAPE) in which a unified model is trained to detect the poses of objects from any category with limited supervision. + +Previous works attempts to compute keypoint prototypes from support and query samples and use them to locate keypoint positions on the query image via a matching step. Xu et al. [42] introduce POMNet, a transformer-based network, to capture the interactions between support and query images into the keypoint prototypes. POMNet is prone to overfitting since the keypoint prototypes heavily rely on very few, typically one or five, support samples. Recently, Shi et al. [31] propose a two-stage framework, called CapeFormer, to further calibrate the matching results. Although effective, CapeFormer uses support keypoint identifier to alleviate the ambiguity of keypoints with similar appearances. The keypoint identifier requires additional effort to consistently annotate the keypoints of different categories. This might prohibit their application in scenarios where keypoint identifiers are not available. + +In this work, we aim to improve the representation capacity of the keypoint prototypes by leveraging the knowl- + +edge learned from the base categories. We propose a Bayesian framework that models a prior over the keypoint prototypes. This allows us to analytically encode the knowledge from the base categories and explicitly transfer it to the novel ones. Specifically, we introduce the notion of super-keypoint, a high-level abstract keypoint that models the distribution of semantically related keypoints from different categories. This is based on the observation that keypoints from different categories can be closely related. For example, keypoints corresponding to the left knee of many animals are visually similar and can be grouped into a super-keypoint 'left knee'. Encoding the super-keypoint 'left knee' from many base animals can help generalize to the left knee of an unseen animal (see Fig. 1). The super-keypoints are discovered and estimated automatically from the keypoints of base categories, whose keypoint distributions are well estimated from the abundant number of samples. The adaptation to an unseen category involves two steps: first, we match each novel keypoint to a related super-keypoint; and second, we estimate the keypoint prototypes by considering both the features from the support samples and the prior knowledge encoded in the matched super-keypoints. The first step requires finding an optimal bipartite matching between the set of novel keypoints and the learned super-keypoints. To solve this, we leverage a learnable matching network to capture the structural relations between the novel keypoints and the super-keypoints before making predictions. + +We call our framework Encoding Super-keypoints for Category-Agnostic Pose Estimation, dubbed ESCAPE. ESCAPE prevents the estimated keypoints from overfitting to the sparse support samples by allowing knowledge from the abundant base categories to be explicitly transferred to the novel one during adaptation. Furthermore, ESCAPE does not use keypoint identifiers to alleviate keypoint ambiguity. Experiments show that our method outperform state-of-the-art methods on the standard MP-100 benchmark, assuming keypoint identifiers are not available. + +# 2. Related works + +2D pose estimation. 2D pose estimation aims to estimate the positions of a set of semantic keypoints of an object from its image. This task holds significant importance across various applications, such as human pose estimation for understanding human behaviors [19, 24, 40, 41], animal pose estimation for wildlife conservation [11, 17, 20, 46], and more. Existing methods can be roughly divided into regression-based methods [3, 26, 34, 35, 37, 39] and heatmap-based methods [1, 2, 4, 5, 18, 33]. The former directly predicts the coordinates of keypoints, while the latter infers heatmaps and takes the peak positions as it predictions. However, these methods are primarily designed for estimating poses for a single (super-)category with a fixed + +set of semantic keypoints, and they typically require access to large-scale datasets. + +Few-shot learning. Few-shot learning focuses on learning novel concepts from few examples [14, 45]. While few-shot learning has been extensively studied in domains such as image classification [16] and video classification [25], its application in 2D pose estimation remains relatively under-explored. Our notion of super-keypoints is related to the Bayesian approach in few-shot learning literature [29, 43, 48]. Salakhutdinov et al. [29] introduce super-categories that play role as higher-level abstract concepts for categories and encode distributions over category parameters. Zhang et al. [48] design a Bayesian framework to learn a single super-category that induces a prior over parameters for all categories; such an approach limits model flexibility and is not suitable for CAPE due to the diversity of keypoints. Yang et al. [43] observe that similar categories have nearly identical statistics and propose enhancing novel categories with similar base categories. However, it is not trivial to adopt this method for CAPE due to the lack of a structural way to measure similarity between categories with semantically and quantitatively different keypoints. + +Semantic correspondence. Semantic correspondence refers to the semantic matching between pixels or keypoints between images [6]. A semantic correspondence model typically includes a feature extractor that generates dense features from images and establishes the correspondence using a similarity function. Oquab et al. [27] show that self-supervised ViT models can produce expressive features for finding semantic correspondence. Recently, Tang et al. [36] observe that semantic correspondence naturally emerges as an ability of diffusion models. These models can effectively solve CAPE by matching query pixels to keypoints in support images. Since ESCAPE operates in the feature space, it complements semantic correspondence models. Integrating ESCAPE with an off-the-shelf feature extractor consistently enhances its performance in CAPE tasks. + +# 3. Background + +# 3.1. Category-agnostic pose estimation + +Category-agnostic pose estimation (CAPE) [42] seeks an unified model to estimate poses from any category given a few annotated samples. Each category $y$ has a set of $T_{y}$ keypoints of interest: $J_{y} = \{j_{r}\}_{r = 1}^{T_{y}}$ . Given an image $\mathbf{I}$ of $y$ , the model's objective is to accurately predict the coordinates $P = \{\mathbf{p}_r\}_{r = 1}^{T_y}$ of these keypoints. The setting of CAPE consists of two disjoint sets of categories that are $C_{base}$ for training the model and $C_{novel}$ for testing its generalization capabilities. While the base categories have abundant + +training data, the novel categories have very few training samples. Specifically, a training set $\mathcal{D}_{base}^{y} = \{(I^{i},P^{i})\}_{i = 1}^{n_{y}}$ ( $n_{y}$ is large) of each base category $y$ from $C_{base}$ is available for training the model. The testing stage includes a large number of episodes (or tasks) sampled from the novel categories. In each episode, a support set $S_{novel}^{y} = \{(\mathbf{I}^{i},P^{i})\}_{i = 1}^{K}$ of $K$ annotated samples from $y\in C_{novel}$ is given to the model for adaptation, and the model performance is evaluated on a set $\mathcal{Q}_{novel}^{y} = \{(\mathbf{I}^{i},P^{i})\}_{i = K + 1}^{K + Q}$ of $Q$ query samples. + +Support keypoint identifier. The MP-100 dataset is composed of existing widely used datasets, many of which impose an ordering over the keypoints of their categories. Shi et al [31] refer to this ordering information as support keypoint identifier and encode it into the keypoint features to alleviate the ambiguity of novel keypoints. However, incorporating keypoint identifiers may limit the applicability of CAPE to real-world scenarios where novel keypoints can be freely defined. In this paper, we focus on the challenging yet practical settings for CAPE that do not assume the order of keypoints to be available. We refer readers to [13] for detailed discussion of the support keypoint identifier. + +# 3.2. Simple baseline for CAPE + +In this section, we introduce a simple yet effective baseline for CAPE. The baseline includes a feature extractor $f_{\theta}$ that extracts dense features from input images. We denote the local feature extracted at a position $\mathbf{p}$ of an input image $\mathbf{I}$ as $f_{\theta}(\mathbf{I})[\mathbf{p}] \in \mathbb{R}^{d}$ . We aim to train $f_{\theta}$ so that it possesses strong semantic correspondence for the keypoints of interest. + +For each base category $y$ with keypoints of interest $J_{y}$ we define a set of prototypes $\mathbf{W}^y = \{\mathbf{w}_r\}_{r = 1}^{T_y}$ ; each prototype $\mathbf{w}_r\in \mathbb{R}^d$ is responsible for predicting a heatmap $\mathbf{H}_r$ for the $r^{\mathrm{th}}$ keypoint in $J_{y}$ as follows: + +$$ +\mathbf {H} _ {r} [ \mathbf {p} ] = \left\langle f _ {\theta} (\mathbf {I}) [ \mathbf {p} ], \mathbf {w} _ {r} \right\rangle , \tag {1} +$$ + +where $\langle \cdot, \cdot \rangle$ denotes the inner product, and $\mathbf{p}$ varies over the spatial dimensions. The peak position in the heatmap is then the prediction for that keypoint. + +We train the feature extractor $f_{\theta}$ and the base prototypes $\{\mathbf{W}^y\}_{y \in C_{base}}$ on the base training data $\bigcup_{y \in C_{base}} \{\mathcal{D}_{base}^y\}$ with the standard Mean Squared Error (MSE) loss [23, 38, 41] between the predicted heatmaps and the ground-truth heatmaps. + +To adapt to a novel category $y \in C_{\text{novel}}$ with support and query sets $(S_{\text{novel}}^y, Q_{\text{novel}}^y)$ , we freeze the feature extractor and compute the prototypes $\mathbf{W}^y = \{\mathbf{w}_r\}_{r=1}^{T_y}$ for novel keypoints as the means of keypoint features extracted from the support set: + +$$ +\mathbf {w} _ {r} = \frac {1}{K} \sum_ {\left(\mathbf {I} ^ {i}, P ^ {i}\right) \in \mathcal {S} _ {n o v e l} ^ {y}} f _ {\theta} \left(\mathbf {I} ^ {i}\right) \left[ \mathbf {p} _ {r} ^ {i} \right]. \tag {2} +$$ + +![](images/f8eae37c575b5467bca7dbb859695fd1c8f55767793bbe4f3cd04f09a43d6b62.jpg) +(a) zebra + +![](images/c6f86616d56748aa2372eaadbe603d88a655f0a4cc99e7d596c27f32b5dbe6e8.jpg) +(b) horse + +![](images/d658db9c6c42f13179b34da3b519c74e5545a9181ab05ec8cc45d3c4c53af1d8.jpg) +Figure 2. t-SNE visualization of distributions of 7 keypoints of zebra and horse. Different colors represent different keypoints. $\times$ denotes keypoint features extracted from images. (a) shows extracted features (denoted as $\times$ ) and distribution means (denoted as $\triangle$ ) for keypoints of zebra. (b) shows extracted features (denoted as $\times$ ) and distribution means (denoted as $\circ$ ) for keypoints of horse. +Figure 3. Super-keypoint modeling. Keypoint 'right back paw of zebra' and keypoint 'right back paw of horse' belong to the same super-keypoint 'right back paw'. The keypoints distribute closely around their super-keypoint. + +The computed prototypes are then used to predict keypoint heatmaps for query images as in Eq. 1. + +Our simple baseline provides results comparable to POMNet [42], which has more complicated matching network. The baseline, however, has the same problem of overfitting as POMNet and CapeFormer due to its heavy reliance on the sparse support samples. Building on our simple baseline, we propose a Bayesian framework that allows us to estimate better prototypes for the novel keypoints, while keeping the feature extractor the same. Specifically, we seek to equip the baseline with strong a prior to prevent the keypoint prototypes from overfitting to the support set. + +# 4. Prior as a mixture of super-keypoints + +We notice that keypoints of different categories can be visually similar, and hence their features should follow the same distribution. To verify our hypothesis, we use the feature extractor from our baseline model to extract features at the positions of seven keypoints in images of 'zebra' (a base category) and 'horse' (an unseen category). Fig. 2 (a) and (b) show keypoints features extracted from zebra and horse images respectively. For each keypoint, we also visualize the mean of its distribution (denoted as $\triangle$ for keypoint of zebra and $\bigcirc$ for keypoint of horse). As shown in the figure, although the model has not seen any horse sample during + +training, the keypoint features of horse still align well with those of zebra. This alignment suggests that capturing the distributions of zebra keypoints facilitates generalization to horse keypoints. From this observation, we introduce super-keypoints to model the connections between keypoints of horse and zebra. For instance, the keypoint 'right back paw of zebra' and the keypoint 'right back paw of horse' belong to the same super-keypoint 'right back paw'. + +# 4.1. Super-keypoints modeling + +We introduce the notion of super-keypoint to define a prior over a group of visually similar keypoints from different categories. Let us consider the $r^{\mathrm{th}}$ keypoint of a category $y$ . Given an image $\mathbf{I}$ of $y$ , we denote $\mathbf{o}_r \coloneqq f_\theta(\mathbf{I})[\mathbf{p}_r]$ as the local feature extracted at the position of the $r^{\mathrm{th}}$ keypoint in $\mathbf{I}$ . We assume that the keypoint feature $\mathbf{o}_r$ follows a Gaussian distribution with a mean $\zeta_r$ and covariance $\phi^2 I$ . We further assume that the mean of the keypoint distribution $\zeta_r$ obeys a Gaussian distribution with a mean $\mu_{z_r}$ and covariance $\sigma^2 I$ , where $\mu_{z_r}$ represent the super-keypoint of the $r^{\mathrm{th}}$ keypoint. Formally, + +$$ +\mathbf {o} _ {r} \sim \mathcal {N} (\mathbf {o} | \zeta_ {r}; \phi^ {2}), \quad \text {a n d} \quad \zeta_ {r} \sim \mathcal {N} (\zeta | \mu_ {z _ {r}}; \sigma^ {2}). \qquad (3) +$$ + +Fig. 3 illustrates our super-keypoint modeling. Intuitively, $\mu$ is the representation of a super-keypoint in the feature space, and its keypoint members distribute closely nearby. Our goal is to discover and estimate a set of superkeypoints from keypoints of base categories so that the super-keypoints can encode the statistics of groups of related keypoints across different categories. When adapting to a novel category, the statistics encoded in the superkeypoints are directly transferred to the prototypes of novel keypoints, preventing our model from overfitting to the sparse support set. + +# 4.2. Super-keypoints discovering and learning + +We obtain the set of super-keypoints by applying a clustering method to the keypoints of all base categories. Specifically, given the trained feature extractor $f_{\theta}$ from the baseline, we extract the features of keypoints from all base categories. We then estimate the distribution mean of each keypoint in the base categories as follows: + +$$ +\zeta_ {r} ^ {y} = \frac {1}{| \mathcal {D} _ {\text {b a s e}} ^ {y} |} \sum_ {(\mathbf {I} ^ {i}, P ^ {i}) \in \mathcal {D} _ {\text {b a s e}} ^ {y}} f _ {\theta} (\mathbf {I} ^ {i}) [ \mathbf {p} _ {r} ^ {i} ], \tag {4} +$$ + +where $\zeta_r^y$ represents the distribution mean of the $r^{\mathrm{th}}$ keypoint of category $y\in C_{base}$ . The set of keypoints from all base categories is then $\bigcup_{y\in C_{base}}\{\zeta_r^y\}_{r = 1}^{T_y}$ . In this work, we use the k-nearest neighbor based density peaks clustering [9, 47], which does not require the number of clusters to be known in advance. Note that keypoints of the same + +category are distinct, so they cannot belong to the same super-keypoint. A detailed description of the clustering is provided in the Appendix. Finally, we estimate each super-keypoint representation $\mu$ as the empirical mean of distribution means of keypoints within the cluster, resulting in a set of $L$ super-keypoints: $\mathbf{M} = \{\mu_1,\mu_2,\dots ,\mu_L\}$ . + +# 5. Novel category adaptation + +In this section, we describe how to adapt to a novel category $y$ whose support and query sets are $(\mathcal{S}_{novel}^{y},\mathcal{Q}_{novel}^{y})$ . Our goal is to estimate the prototypes that well represent the distributions of the novel keypoints from a few support samples. In particular, for the $r^{\mathrm{th}}$ keypoint of $y$ , we want the estimated prototype $\mathbf{w}_r$ to be close to the true distribution mean $\zeta_r$ as much as possible. With this interpretation, the estimated prototype for a novel keypoint in the baseline (Eq. 2) can be viewed as maximum likelihood estimate of the mean of the keypoint distribution. It is clear that the baseline is prone to overfitting since the number of support samples is extremely limited. + +# 5.1. Prototype estimate via maximum a posteriori + +We propose to alleviate the overfitting by letting the statistics captured in the super-keypoints to be transferred to the novel keypoints. Specifically, consider the $r^{\mathrm{th}}$ keypoint of $y$ , we compute its prototype $\mathbf{w}_r$ as maximum a posterior estimate of $\zeta_r$ : + +$$ +\mathbf {w} _ {r} ^ {M A P} = \underset {\zeta_ {r}} {\operatorname {a r g m a x}} \log p \left(\zeta_ {r} \mid \left\{\mathbf {o} _ {r} ^ {i} \right\} _ {i = 1} ^ {K}, \mathbf {M}\right), \tag {5} +$$ + +where $\mathbf{o}_r^i = f_\theta (\mathbf{I}^i)[\mathbf{p}_r^i ]$ is the feature of the $r^{\mathrm{th}}$ keypoint extracted from the $i^{\mathrm{th}}$ support sample in $\mathcal{S}_{novel}^{y}$ , and $\mathbf{M} = \{\mu_1,\mu_2,\dots ,\mu_L\}$ is the set of learned super-keypoints. + +However, when adapting to a novel category, the superkeypoints of the novel keypoints are not available in general. We do not know which super-keypoint the $r^{\mathrm{th}}$ novel keypoint should belong to. To this extent, we adopt the expectation maximization (EM) algorithm [21], which first matches the $r^{\mathrm{th}}$ keypoint with the 'best' super-keypoint from M (the expectation step) and then optimizes for the prototype with the prior encoded in the matched super-keypoint (the maximization step). Below, we summarize the two steps; detailed derivation of the (EM) algorithm is provided in the Appendix. + +Expectation step. Let $z_{r}$ be the super-keypoint assignment of the $r^{\mathrm{th}}$ keypoint of $y$ . We align the keypoint with the most likely super-keypoint $z_{r}^{\star}$ , i.e., the $z_{r}$ that maximizes $\log p(\zeta_{r}, z_{r}|\{\mathbf{o}_{r}^{i}\}_{i=1}^{K}, \mathbf{M})$ . Note that keypoints of the same category cannot belong to the same super-keypoint. Therefore, we further adopt the Hungarian algorithm [15] + +![](images/4ab82e4fd947cf03830e88ec00198e6ff4371112cabfe569843345bd3b004f0b.jpg) +Figure 4. Illustration of our adaptation procedure for a novel category. We leverage the EM algorithm to probabilistically optimize for the keypoint prototypes, considering both the support samples and the learned super-keypoints. We first align the novel keypoints with the different super-keypoints via a learnable matching network (expectation step). We then let the statistics encoded in the matched super-keypoints to be transferred directly to the keypoint prototypes (maximization step). The keypoint prototypes can be used to estimate poses for query samples. + +to optimally align the $T_{y}$ keypoints of $y$ with distinct superkeypoints $\{\mu_{z_r^*}\}_{r = 1}^{T_y}\subset \mathbf{M}$ . We refer to this matching procedure as simple matching for the rest of the paper. + +Maximization step. Now, we know the 'best' superkeypoint for each keypoint of $y$ . We can compute their prototypes analytically as follows: + +$$ +\begin{array}{l} \mathbf {w} _ {r} ^ {M A P} = \underset {\mathbf {w} _ {r}} {\operatorname {a r g m a x}} \log p \left(\mathbf {w} _ {r} | \left\{\mathbf {o} _ {r} ^ {i} \right\} _ {i = 1} ^ {K}, \mu_ {z _ {r} ^ {\star}}\right) (6) \\ = \alpha \frac {1}{K} \sum_ {i = 1} ^ {K} \mathbf {o} _ {r} ^ {i} + (1 - \alpha) \mu_ {z _ {r} ^ {\star}}, (7) \\ \end{array} +$$ + +where $\alpha \in (0,1)$ depends on $\phi$ and $\sigma$ . Since the distribution mean of a keypoint must distribute near its super-keypoint, the second term in Eq. 7 has an effect of pulling the estimated prototype closer to the true mean of the keypoint distribution, diminishing the dependence of the prototype on the support samples. + +# 5.2. Learnable matching network + +Matching novel keypoints with semantically related superkeypoints is crucial for the success of our method. A correct super-keypoint can pull the estimated prototype closer + +to the distribution mean of the keypoint under consideration, while an incorrect super-keypoint pulls the prototype away. However, as later shown in the experiments, the simple matching in the expectation step fails to produce reliable super-keypoints assignments. We argue that this is because the simple matching treats the novel keypoints independently, discarding their structural relations. It ignores the keypoint coordinates in the support samples, and so the model cannot reason about poses of the object of interest. + +To overcome these limitations, we replace the simple matching in the expectation step with a learnable matching network to simultaneously produce matching scores between keypoints in an image and the learned super keypoints. The design is inspired by SuperGlue [30], a matching framework that has shown great performance in aligning two sequences of features. + +Input features. Our matching network takes two sequences of features as input and produces the matching score matrix for their elements. The first sequence is augmented features of $T_{y}$ keypoints extracted from a support image of $y$ . The second sequence is a set of features for the $L$ super-keypoints. To encode features for the keypoints, we consider not only their visual appearance but also their coordinates within the support image. Specifically, given + +
MethodSplit 1Split 2Split 3Split 4Split 5Mean PCKSplit 1Split 2Split 3Split 4Split 5Mean PCK
ProtoNet [32]46.0540.8449.1343.3444.5444.7860.3153.5161.9258.4458.6158.56
MAML [10]68.1454.7264.1963.2457.2061.5070.0355.9863.2164.7958.4762.50
Finetune [22]70.6057.0466.0665.0059.2063.5871.6757.8466.7666.5360.2464.61
POMNet [42]84.2378.2578.1778.6879.1779.7084.7279.6178.0080.3880.8580.71
CapeFormer [31]84.1379.9079.5879.7880.3480.7590.4686.4785.8086.4786.7187.19
ESCAPE86.8982.5581.2581.7281.3282.7491.4187.4385.3387.2786.7687.63
+ +Table 1. Comparison with state-of-the-art works on MP-100 dataset under 1-shot and 5-shot settings. ${}^{ \dagger }$ denotes keypoint identifiers. + +support image $\mathbf{I}\sim S_{novel}^{y}$ , the keypoint features are: + +$$ +\mathbf {u} _ {r} = f _ {\gamma} (\mathbf {I}) \left[ \mathbf {p} _ {r} \right] + P E _ {\mathrm {g e o}} \left(\mathbf {p} _ {r}\right), \text {f o r} r = \{1, \dots , T _ {y} \}, \tag {8} +$$ + +where $f_{\gamma}$ is a CNN feature extractor, and $PE_{\mathrm{geo}}$ embeds a coordinate into the feature space. By considering the keypoint coordinates, our model can reason about the pose of the object in the support image. + +At the other end, the features for the super-keypoints are defined as: + +$$ +\mathbf {v} _ {z} = \omega_ {z} + P E _ {\mathrm {a b s}} (z), \quad \text {f o r} z \in \{1, \dots , L \}, \tag {9} +$$ + +where $\omega_{z}$ is a learnable feature vector for the $z^{\mathrm{th}}$ superkeypoint, and $PE_{\mathrm{abs}}$ embeds an index $(1,\dots ,L)$ into the feature space. Note that the indexes of the super-keypoints are obtained from the clustering step described in Sec. 4.2. + +Feature aggregation with attention. The two sequences of features $\{\mathbf{u}_r\}_{r=1}^{T_y}$ and $\{\mathbf{v}_z\}_{z=1}^L$ are then passed to a neural network for information aggregation. The network comprises $U$ blocks of self-attention and cross-attention to model interactions both within and between the elements of the sequences. Results are two sequences of refined features: $\{\mathbf{u}_r^\star\}_{r=1}^{T_y}$ and $\{\mathbf{v}_z^\star\}_{z=1}^L$ . The predicted matching scores matrix is then $\mathbf{S} \in \mathbb{R}^{T_y \times L}$ , with each element $\mathbf{S}_{r,z} = \langle \mathbf{u}_r^\star, \mathbf{v}_z^\star \rangle$ measuring the matching score between the $r^{\text{th}}$ keypoint and the $z^{\text{th}}$ super-keypoint. An illustration of the matching network is provided at the bottom of Fig. 4. + +Matching as optimal transport problem. Since the keypoints within the same category must be matched to different super-keypoints, our goal is to find a maximum score bipartite matching between the two sequences. Following SuperGlue [30], we replace the Hungarian algorithm with the differentiable Sinkhorn algorithm [7] for optimal transport. It takes the matching score matrix $\mathbf{S}$ as input and produces a matrix $\mathbf{Q} \in \mathbb{R}^{T_y \times L}$ where each element $\mathbf{Q}_{r,z}$ represents the probability that the $r^{\text{th}}$ keypoint is matched with the $z^{\text{th}}$ super-keypoint. We train the network end-to-end on the keypoints of base categories, whose super-keypoints assignments are available from the clustering step. Details of the Sinkhorn algorithm and the training procedure of the matching network are provided in the Appendix. + +Incorporating learnable matching network to novel category adaptation. We integrate the learned matching network into the novel category adaptation. Specifically, for every support image $\mathbf{I}^i\in S_{novel}^y$ , we compute the matching score matrix $\mathbf{S}^i$ from its keypoints and the learned super keypoints. The corresponding optimal assignment matrix is then denoted as $\mathbf{Q}^i$ . The final prototype for the $r^{\mathrm{th}}$ keypoint are then estimated as follows: + +$$ +\mathbf {w} _ {r} ^ {\text {E S C A P E}} = \alpha \frac {1}{K} \sum_ {i = 1} ^ {K} \mathbf {o} _ {r} ^ {i} + (1 - \alpha) \frac {L}{K} \sum_ {i = 1} ^ {K} \sum_ {z = 1} ^ {L} \mathbf {Q} _ {r, z} ^ {i} \mu_ {z}. \tag {10} +$$ + +Here we keep the soft assignment form of $\mathbf{Q}^i$ instead of the hard assignment to model uncertainty. Our adaptation procedure is shown in Fig. 4. + +# 6. Experiments + +We perform experiments on Multi-category Pose (MP-100) proposed in [42]. The dataset consists of more than 20,000 data instances of 100 categories from 8 super-categories including human hand, human face, human body, animal body, animal face, clothes, furniture, and vehicle. We also follow the five splits from [42] for standard CAPE settings. Each split divides the dataset into 70 training categories, 10 validation categories, and 20 testing categories. The training set is used as base categories, while validation and testing sets are novel categories for validation and testing, respectively. Please refer to [42] for more details. + +# 6.1. Experimental details + +Unless otherwise specified, for fair comparisons with previous works, we utilize ResNet-50 [12] pretrained on ImageNet [8] as the feature extractor. Also following [42], we crop the object of interest from the original image using the ground-truth bounding box, and resize it to a fixed size of $256 \times 256$ . + +Probability of correct keypoint [44] with a threshold of 0.2 (PCK@0.2) is the evaluation metric for MP-100. During testing, we generate 60,000 episodes evenly distributed across the test categories and report the average PCK@0.2 over these episodes. + +![](images/2dcb167bf1df6d2a3f0848cbf991a77c10385557f296ffba52d16b4497e2d3cc.jpg) +super-keypoint 1 + +![](images/dc6d2dd56b6ec5f1f181475c2fe1c16ab7c667692324d7245460aef242154200.jpg) +super-keypoint 4 + +![](images/7ab2abb81057dd1bd795501fd6a87507585c83a7247f6571d56827b2d5c89b82.jpg) +super-keypoint 2 +arctic wolf + +![](images/5becc95d4bbc53fd37ef419c7d35c1458503d22e41b4db10616afa3c288c4d7f.jpg) +sealion +ferret + +![](images/5821cad53eb20a3b71bf969b5d87fbd067c5a4483afdb5659f47b9b272ac65db.jpg) +grebe + +![](images/f936ba5b207f2eafb9bfbd02d131c92a99c29b07ac9f24e7ca7d93197999e3e9.jpg) +super-keypoint 5 +kingfisher + +![](images/86b51a155c2dfd9e02781812436417f8818af38079354f12e4f54474ad7c9ef9.jpg) +wren +gull + +![](images/745a79b6731c5e06dd435e94e3e370fba06a59ba037462611dc234c26267ccdd.jpg) +super-keypoint 3 + +![](images/529764b53507c5dd148173d39691175d465e94ba850e670a31e493d169546f71.jpg) +long sleeved dress + +![](images/864cc0f355e8c5fca8876a7a91a61925f3d5cb322aee6c66dfdbd2e2c0198f53.jpg) +super-keypoint 6 +vest +vest dress long sleeved shirt + +![](images/6959ff7436ec326b22dd0585880a33e7240f3a16460d31ddf18e3e5e0dc94065.jpg) +Figure 5. Keypoints of four different super-keypoints discovered from the base categories. Below each keypoint is the name of its category. + +![](images/d2ef5f652fa5ababd2292e8a42bea4f21d8b7ae7fae169876860d639a7579ddf.jpg) + +![](images/1ab32c0f7de342004916f7973eb8bc0093513bf88ac4d5cfbe4f91092111ae67.jpg) + +![](images/4855aa92067acae24b11129676c42da4b62c1843c6aebf796e3bb39043d4ddc9.jpg) +Figure 6. Qualitative comparison between simple matching and learnable matching network in aligning novel keypoints with superkeypoints. We show one base keypoint member of each matched super-keypoint. + +# 6.2. Comparison with previous methods + +We compare ESCAPE against previous methods, including Prototypical Networks [32], MAML [10], Finetune [22], POMNet [32] and CapeFormer [31]. Note that we remove the support keypoint identifier from CapeFormer. The results are shown in Tab. 1. We can see that ESCAPE outperforms other methods by around $2.5\%$ and $0.5\%$ in 1-shot and 5-shot settings, respectively, establishing a new state-of-the-art on the MP-100 benchmark. + +# 6.3. Ablation study + +Super-keypoints discovering. We perform clustering on the keypoints of base categories to form super-keypoints. To verify that we can obtain meaningful super-keypoints from the clustering process, we show some keypoints members of a few super-keypoints in Fig. 5. We see that each super- + +keypoint represents a meaningful concept. For example, super-keypoint 1 consists of 'left knee' of deer, antelope and zebra. super-keypoint 4 consists of 'right rear wheel' of SUV, bus, and car, while super-keypoint 6 is a landmark on the collars of different clothes categories. + +Novel keypoints and super-keypoints matching. We qualitatively compare the performance of the simple matching and the learnable matching network in aligning novel keypoints with the learned super-keypoints. Fig. 6 shows the matching results for two novel categories, namely squirrel and woodpecker. The left side of the figure shows support images of the categories, while the right side shows the super-keypoints aligned by the simple matching and the learnable matching network. For each super-keypoint, we only show one base keypoints member. Compared with the + +
MethodPCK
MLE83.00 ± 0.16
MAP83.88 ± 0.15
ESCAPE86.89 ± 0.14
+ +Table 2. Results of different prototype estimation methods, namely MLE, MAP and ESCAPE. + +
Feature extractorMethodPCK
DIFT [36]MLE60.61
DIFT [36]ESCAPE62.47
DINOv2 [27]MLE88.31
DINOv2 [27]ESCAPE89.14
+ +Table 3. Results of integrating ESCAPE to feature extractors of different semantic correspondence models + +simple matching, our learnable matching network provides more reliable super-keypoints for the novel keypoints. In particular, the simple matching aligns keypoints of squirrel with super-keypoints of keypoints of kingfisher, sling and suv, which are irrelevant to the squirrel. On the other hand, the learnable matching network aligns the novel keypoints with their actual concepts, e.g., the 'left elbow of squirrel' is matched with the super-keypoint of the 'left elbow of rat'. + +Ablation on novel category adaptation. In Tab. 2, we study the performance of different methods to estimate keypoint prototypes for a novel category, namely MLE, MAP and ESCAPE. MLE is our baseline, which estimate keypoint prototypes as means of support features as in Eq. 2. MAP and ESCAPE compute keypoint prototypes as maximum a posterior estimate with the EM algorithm. The difference between MAP and ESCAPE is that MAP (Eq. 7) uses the simple matching in the expectation step, whereas ESCAPE (Eq. 10) leverages the learnable matching network. We observe that incorporating super-keypoints into the estimation of keypoint prototypes helps improve the overall performance. Specifically, the MAP outperforms MLE by roughly $1\%$ . Regardless ESCAPE, better novel keypoints and super-keypoints alignments from the learnable matching network immediately translate into $3\%$ improvements over the MAP counterpart. + +Ablation on different feature extractors. Feature extractors from semantic correspondence models can effectively solve CAPE since they can generate dense features for support and query images. In Tab. 3, we study the compatibility of ESCAPE with feature extractors from different semantic correspondence models, namely DIFT [36] and DINoV2 [27]. We take the pretrained stable diffusion 2.1 [28] as the backbone for DIFT, whereas we use ViT-ViT-S/14 for DINoV2. As seen, ESCAPE consistently boosts the performance of the considered semantic correspondence models. This shows that ESCAPE can benefit from future research for better representation. + +![](images/cf502b8569da0778b0813a397dc07a58cc3c1dc4da2c3c3847660a35c76a943e.jpg) +Figure 7. We qualitatively compare pose estimation of Cape-Former and ESCAPE under 1-shot setting. + +# 6.4. Qualitative results + +Figure 7 compares the pose estimation performance of CapeFormer and ESCAPE. Overall, ESCAPE gives more accurate pose predictions compared to CapeFormer. In particular, CapeFormer is overfitting to the support image of the sheep category. The predicted positions of the left front and back paws by CapeFormer are geometrically close to each other as they appear in the support image. In contrast, since ESCAPE allows learned knowledge about base animals to be transferred to the novel keypoints, it recovers the pose of the query sheep beautifully. + +# 7. Conclusion + +In this paper, we address the task of category-agnostic pose estimation by introducing ESCAPE, a Bayesian framework that induces a prior over the keypoint features of different categories. This prior is a mixture of super-keypoints, with each super-keypoint encoding a distribution of a group of related keypoints. ESCAPE enables the transfer of knowledge from base categories to the novel ones, thereby preventing the query predictions from overfitting to sparse support samples. ESCAPE offers several advantages over previous works: it exhibits less overfitting to support samples, does not rely on support keypoint identifiers, and is model-agnostic. + +# Acknowledgement + +This research is supported in part by the National Research Foundation, Singapore under its Competitive Research Program Award NRF-CRP23-2019-0003. + +# References + +[1] Yuanhao Cai, Zhicheng Wang, Zhengxiong Luo, Binyi Yin, Angang Du, Haoqian Wang, Xiangyu Zhang, Xinyu Zhou, Erjin Zhou, and Jian Sun. Learning delicate local representations for multi-person pose estimation. In European Conference on Computer Vision, pages 455-472. Springer, 2020. 2 +[2] Zhe Cao, Tomas Simon, Shih-En Wei, and Yaser Sheikh. Realtime multi-person 2d pose estimation using part affinity fields. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7291-7299, 2017. 2 +[3] Joao Carreira, Pulkit Agrawal, Katerina Fragkiadaki, and Jitendra Malik. Human pose estimation with iterative error feedback. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4733-4742, 2016. 2 +[4] Yilun Chen, Zhicheng Wang, Yuxiang Peng, Zhiqiang Zhang, Gang Yu, and Jian Sun. Cascaded pyramid network for multi-person pose estimation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7103-7112, 2018. 2 +[5] Bowen Cheng, Bin Xiao, Jingdong Wang, Honghui Shi, Thomas S Huang, and Lei Zhang. Higherhrnet: Scale-aware representation learning for bottom-up human pose estimation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5386-5395, 2020. 2 +[6] Christopher B Choy, JunYoung Gwak, Silvio Savarese, and Manmohan Chandraker. Universal correspondence network. Advances in neural information processing systems, 29, 2016. 2 +[7] Marco Cuturi. Sinkhorn distances: Lightspeed computation of optimal transport. Advances in neural information processing systems, 26, 2013. 6 +[8] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 6 +[9] Mingjing Du, Shifei Ding, and Hongjie Jia. Study on density peaks clustering based on k-nearest neighbors and principal component analysis. Knowledge-Based Systems, 99: 135-145, 2016. 4 +[10] Chelsea Finn, Pieter Abbeel, and Sergey Levine. Model-agnostic meta-learning for fast adaptation of deep networks. In International conference on machine learning, pages 1126–1135. PMLR, 2017. 6, 7 +[11] Jacob M Graving, Daniel Chae, Hemal Naik, Liang Li, Benjamin Koger, Blair R Costelloe, and Iain D Couzin. Deepposekit, a software toolkit for fast and robust animal pose estimation using deep learning. *Elife*, 8:e47994, 2019. 2 +[12] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 6 +[13] Or Hirschorn and Shai Avidan. Pose anything: A graph-based approach for category-agnostic pose estimation. arXiv preprint arXiv:2311.17891, 2023. 3 + +[14] Timothy Hospedales, Antreas Antoniou, Paul Micaelli, and Amos Storkey. Meta-learning in neural networks: A survey. IEEE transactions on pattern analysis and machine intelligence, 44(9):5149-5169, 2021. 2 +[15] Harold W Kuhn. The hungarian method for the assignment problem. Naval research logistics quarterly, 2(1-2):83-97, 1955. 4 +[16] Duong Le, Khoi Duc Nguyen, Khoi Nguyen, Quoc-Huy Tran, Rang Nguyen, and Binh-Son Hua. Poodle: Improving few-shot learning via penalizing out-of-distribution samples. Advances in Neural Information Processing Systems, 34:23942-23955, 2021. 2 +[17] Shuyuan Li, Jianguo Li, Hanlin Tang, Rui Qian, and Weiyao Lin. Atrw: a benchmark for amur tiger re-identification in the wild. arXiv preprint arXiv:1906.05586, 2019. 2 +[18] Wenbo Li, Zhicheng Wang, Binyi Yin, Qixiang Peng, Yuming Du, Tianzi Xiao, Gang Yu, Hongtao Lu, Yichen Wei, and Jian Sun. Rethinking on multi-stage networks for human pose estimation. arXiv preprint arXiv:1901.00148, 2019. 2 +[19] Jiang-Jiang Liu, Qibin Hou, Ming-Ming Cheng, Changhu Wang, and Jiashi Feng. Improving convolutional networks with self-calibrated convolutions. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10096-10105, 2020. 2 +[20] Alexander Mathis, Thomas Biasi, Steffen Schneider, Mert Yuksekgonul, Byron Rogers, Matthias Bethge, and Mackenzie W Mathis. Pretraining boosts out-of-domain robustness for pose estimation. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1859-1868, 2021. 2 +[21] Kevin P. Murphy. Probabilistic Machine Learning: Advanced Topics. MIT Press, 2023. 4 +[22] Akihiro Nakamura and Tatsuya Harada. Revisiting fine-tuning for few-shot learning. arXiv preprint arXiv:1910.00216, 2019. 6, 7 +[23] Alejandro Newell, Kaiyu Yang, and Jia Deng. Stacked hourglass networks for human pose estimation. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part VIII 14, pages 483-499. Springer, 2016. 3 +[24] Alejandro Newell, Zhiao Huang, and Jia Deng. Associative embedding: End-to-end learning for joint detection and grouping. Advances in neural information processing systems, 30, 2017. 2 +[25] Khoi D Nguyen, Quoc-Huy Tran, Khoi Nguyen, Binh-Son Hua, and Rang Nguyen. Inductive and transductive few-shot video classification via appearance and temporal alignments. In European Conference on Computer Vision, pages 471-487. Springer, 2022. 2 +[26] Xuecheng Nie, Jiashi Feng, Jianfeng Zhang, and Shuicheng Yan. Single-stage multi-person pose machines. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6951-6960, 2019. 2 +[27] Maxime Oquab, Timothée Darcet, Théo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023. 2, 8 + +[28] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 8 +[29] Ruslan Salakhutdinov, Joshua Tenenbaum, and Antonio Torralba. One-shot learning with a hierarchical nonparametric bayesian model. In Proceedings of ICML Workshop on Unsupervised and Transfer Learning, pages 195-206. JMLR Workshop and Conference Proceedings, 2012. 2 +[30] Paul-Edouard Sarlin, Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. Superglue: Learning feature matching with graph neural networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4938–4947, 2020. 5, 6 +[31] Min Shi, Zihao Huang, Xianzheng Ma, Xiaowei Hu, and Zhiguo Cao. Matching is not enough: A two-stage framework for category-agnostic pose estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7308-7317, 2023. 1, 3, 6, 7 +[32] Jake Snell, Kevin Swersky, and Richard Zemel. Prototypical networks for few-shot learning. Advances in neural information processing systems, 30, 2017. 6, 7 +[33] Ke Sun, Bin Xiao, Dong Liu, and Jingdong Wang. Deep high-resolution representation learning for human pose estimation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5693-5703, 2019. 2 +[34] Xiao Sun, Jiaxiang Shang, Shuang Liang, and Yichen Wei. Compositional human pose regression. In Proceedings of the IEEE International Conference on Computer Vision, pages 2602-2611, 2017. 2 +[35] Xiao Sun, Bin Xiao, Fangyin Wei, Shuang Liang, and Yichen Wei. Integral human pose regression. In Proceedings of the European conference on computer vision (ECCV), pages 529-545, 2018. 2 +[36] Luming Tang, Menglin Jia, Qianqian Wang, Cheng Perng Phoo, and Bharath Hariharan. Emergent correspondence from image diffusion. Advances in Neural Information Processing Systems, 36:1363-1389, 2023. 2, 8 +[37] Zhi Tian, Hao Chen, and Chunhua Shen. Direct pose: Direct end-to-end multi-person pose estimation. arXiv preprint arXiv:1911.07451, 2019. 2 +[38] Jonathan J. Tompson, Arjun Jain, Yann LeCun, and Christoph Bregler. Joint training of a convolutional network and a graphical model for human pose estimation. Advances in neural information processing systems, 27, 2014. 3 +[39] Alexander Toshev and Christian Szegedy. Deeppose: Human pose estimation via deep neural networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1653-1660, 2014. 2 +[40] Jingdong Wang, Ke Sun, Tianheng Cheng, Borui Jiang, Chaorui Deng, Yang Zhao, Dong Liu, Yadong Mu, Mingkui Tan, Xinggang Wang, et al. Deep high-resolution representation learning for visual recognition. IEEE transactions on pattern analysis and machine intelligence, 43(10):3349-3364, 2020. 2 + +[41] Bin Xiao, Haiping Wu, and Yichen Wei. Simple baselines for human pose estimation and tracking. In Proceedings of the European conference on computer vision (ECCV), pages 466-481, 2018. 2, 3 +[42] Lumin Xu, Sheng Jin, Wang Zeng, Wentao Liu, Chen Qian, Wanli Ouyang, Ping Luo, and Xiaogang Wang. Pose for everything: Towards category-agnostic pose estimation. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23–27, 2022, Proceedings, Part VI, pages 398–416. Springer, 2022. 1, 2, 3, 6 +[43] Shuo Yang, Lu Liu, and Min Xu. Free lunch for few-shot learning: Distribution calibration. arXiv preprint arXiv:2101.06395, 2021. 2 +[44] Yi Yang and Deva Ramanan. Articulated human detection with flexible mixtures of parts. IEEE transactions on pattern analysis and machine intelligence, 35(12):2878-2890, 2012. 6 +[45] Wang Yaqing, Yao Quanming, T Kwok James, and M Ni Lionel. Generalizing from a few examples: A survey on few-shot learning. ACM computing surveys, 53(3):1-34, 2020. 2 +[46] Hang Yu, Yufei Xu, Jing Zhang, Wei Zhao, Ziyu Guan, and Dacheng Tao. Ap-10k: A benchmark for animal pose estimation in the wild. arXiv preprint arXiv:2108.12617, 2021. 2 +[47] Wang Zeng, Sheng Jin, Wentao Liu, Chen Qian, Ping Luo, Wanli Ouyang, and Xiaogang Wang. Not all tokens are equal: Human-centric visual analysis via token clustering transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11101-11111, 2022. 4 +[48] Xueting Zhang, Debin Meng, Henry Gouk, and Timothy M Hospedales. Shallow bayesian meta learning for real-world few-shot recognition. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 651-660, 2021. 2 \ No newline at end of file diff --git a/2024/ESCAPE_ Encoding Super-keypoints for Category-Agnostic Pose Estimation/images.zip b/2024/ESCAPE_ Encoding Super-keypoints for Category-Agnostic Pose Estimation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..3d47c6a7fa7dde5caa415fdc3bba941559c67122 --- /dev/null +++ b/2024/ESCAPE_ Encoding Super-keypoints for Category-Agnostic Pose Estimation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9a236f5635a7f889395b2fba22e4efea221b0e2cf03e7665bff17cdda3804d4e +size 657075 diff --git a/2024/ESCAPE_ Encoding Super-keypoints for Category-Agnostic Pose Estimation/layout.json b/2024/ESCAPE_ Encoding Super-keypoints for Category-Agnostic Pose Estimation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..a040adcc2f925ab6c76dd1ece3ef7ab87e1465e8 --- /dev/null +++ b/2024/ESCAPE_ Encoding Super-keypoints for Category-Agnostic Pose Estimation/layout.json @@ -0,0 +1,9935 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 61, + 103, + 533, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 103, + 533, + 121 + ], + "spans": [ + { + "bbox": [ + 61, + 103, + 533, + 121 + ], + "type": "text", + "content": "ESCAPE: Encoding Super-keypoints for Category-Agnostic Pose Estimation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 95, + 142, + 281, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 95, + 142, + 281, + 185 + ], + "spans": [ + { + "bbox": [ + 95, + 142, + 281, + 185 + ], + "type": "text", + "content": "Khoi Duc Nguyen1* \n1University of Wisconsin-Madison \nkhoi.nguyen@wisc.edu lichen@u" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 283, + 142, + 483, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 283, + 142, + 483, + 185 + ], + "spans": [ + { + "bbox": [ + 283, + 142, + 483, + 185 + ], + "type": "text", + "content": "Chen Li" + }, + { + "bbox": [ + 283, + 142, + 483, + 185 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 283, + 142, + 483, + 185 + ], + "type": "text", + "content": " Gim Hee Lee" + }, + { + "bbox": [ + 283, + 142, + 483, + 185 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 283, + 142, + 483, + 185 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 283, + 142, + 483, + 185 + ], + "type": "text", + "content": "National University of Singapore \n.nus.edu gimhee.lee@comp.nus.edu." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "spans": [ + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 239, + 290, + 563 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 239, + 290, + 563 + ], + "spans": [ + { + "bbox": [ + 46, + 239, + 290, + 563 + ], + "type": "text", + "content": "In this paper, we tackle the task of category-agnostic pose estimation (CAPE), which aims to predict poses for objects of any category with few annotated samples. Previous works either rely on local matching between features of support and query samples or require support keypoint identifier. The former is prone to overfitting due to its sensitivity to sparse samples, while the latter is impractical for the open-world nature of the task. To overcome these limitations, we propose ESCAPE - a Bayesian framework that learns a prior over the features of keypoints. The prior can be expressed as a mixture of super-keypoints, each being a high-level abstract keypoint that captures the statistics of semantically related keypoints from different categories. We estimate the super-keypoints from base categories and use them in adaptation to novel categories. The adaptation to an unseen category involves two steps: first, we match each novel keypoint to a related super-keypoint; and second, we transfer the knowledge encoded in the matched super-keypoints to the novel keypoints. For the first step, we propose a learnable matching network to capture the relationship between the novel keypoints and the super-keypoints, resulting in a more reliable matching. ESCAPE mitigates overfitting by directly transferring learned knowledge to novel categories while it does not use keypoint identifiers. We achieve state-of-the-art performance on the standard MP-100 benchmark. Our code is available at https://github.com/khoiucd/escape-tgt." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 599, + 128, + 611 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 599, + 128, + 611 + ], + "spans": [ + { + "bbox": [ + 47, + 599, + 128, + 611 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 620, + 287, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 620, + 287, + 693 + ], + "spans": [ + { + "bbox": [ + 46, + 620, + 287, + 693 + ], + "type": "text", + "content": "2D pose estimation aims to locate the semantic keypoints of an object of interest from an input image. This task is crucial for many visual understanding tasks, such as understanding human/animal behavior or interpreting the surrounding environment in autonomous driving. Existing works focus on detecting keypoints of only one specific" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 310, + 214, + 541, + 302 + ], + "blocks": [ + { + "bbox": [ + 310, + 214, + 541, + 302 + ], + "lines": [ + { + "bbox": [ + 310, + 214, + 541, + 302 + ], + "spans": [ + { + "bbox": [ + 310, + 214, + 541, + 302 + ], + "type": "image", + "image_path": "c7d77b36b6441bbb24345f270f639eeb9aea1b2ddc508e68a47a3c9de239fe77.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 312, + 546, + 345 + ], + "lines": [ + { + "bbox": [ + 305, + 312, + 546, + 345 + ], + "spans": [ + { + "bbox": [ + 305, + 312, + 546, + 345 + ], + "type": "text", + "content": "Figure 1. We encode super-keypoint 'left knee' from the 'left knee' of base animals and use the super-keypoint in adaptation to the left knee of novel animal." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 366, + 546, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 366, + 546, + 498 + ], + "spans": [ + { + "bbox": [ + 304, + 366, + 546, + 498 + ], + "type": "text", + "content": "(super-)category, such as human, animal, or vehicle. They are inefficient in the sense that they need to train a separate model for each application. Moreover, annotating a large number of samples for some categories, e.g., rare species, is too expensive, if not impossible, and thus hinders the usefulness of the deep learning approach. A unified model that can quickly adapt to keypoints of unseen categories is desirable. Addressing this, Xu et al. [42] propose the task of category-agnostic pose estimation (CAPE) in which a unified model is trained to detect the poses of objects from any category with limited supervision." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 498, + 547, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 498, + 547, + 689 + ], + "spans": [ + { + "bbox": [ + 304, + 498, + 547, + 689 + ], + "type": "text", + "content": "Previous works attempts to compute keypoint prototypes from support and query samples and use them to locate keypoint positions on the query image via a matching step. Xu et al. [42] introduce POMNet, a transformer-based network, to capture the interactions between support and query images into the keypoint prototypes. POMNet is prone to overfitting since the keypoint prototypes heavily rely on very few, typically one or five, support samples. Recently, Shi et al. [31] propose a two-stage framework, called CapeFormer, to further calibrate the matching results. Although effective, CapeFormer uses support keypoint identifier to alleviate the ambiguity of keypoints with similar appearances. The keypoint identifier requires additional effort to consistently annotate the keypoints of different categories. This might prohibit their application in scenarios where keypoint identifiers are not available." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 689, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 689, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 305, + 689, + 547, + 714 + ], + "type": "text", + "content": "In this work, we aim to improve the representation capacity of the keypoint prototypes by leveraging the knowl-" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 702, + 166, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 702, + 166, + 712 + ], + "spans": [ + { + "bbox": [ + 58, + 702, + 166, + 712 + ], + "type": "text", + "content": "*Work done as an intern at NUS." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "23491" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 289, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 289, + 407 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 289, + 407 + ], + "type": "text", + "content": "edge learned from the base categories. We propose a Bayesian framework that models a prior over the keypoint prototypes. This allows us to analytically encode the knowledge from the base categories and explicitly transfer it to the novel ones. Specifically, we introduce the notion of super-keypoint, a high-level abstract keypoint that models the distribution of semantically related keypoints from different categories. This is based on the observation that keypoints from different categories can be closely related. For example, keypoints corresponding to the left knee of many animals are visually similar and can be grouped into a super-keypoint 'left knee'. Encoding the super-keypoint 'left knee' from many base animals can help generalize to the left knee of an unseen animal (see Fig. 1). The super-keypoints are discovered and estimated automatically from the keypoints of base categories, whose keypoint distributions are well estimated from the abundant number of samples. The adaptation to an unseen category involves two steps: first, we match each novel keypoint to a related super-keypoint; and second, we estimate the keypoint prototypes by considering both the features from the support samples and the prior knowledge encoded in the matched super-keypoints. The first step requires finding an optimal bipartite matching between the set of novel keypoints and the learned super-keypoints. To solve this, we leverage a learnable matching network to capture the structural relations between the novel keypoints and the super-keypoints before making predictions." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 407, + 289, + 528 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 407, + 289, + 528 + ], + "spans": [ + { + "bbox": [ + 47, + 407, + 289, + 528 + ], + "type": "text", + "content": "We call our framework Encoding Super-keypoints for Category-Agnostic Pose Estimation, dubbed ESCAPE. ESCAPE prevents the estimated keypoints from overfitting to the sparse support samples by allowing knowledge from the abundant base categories to be explicitly transferred to the novel one during adaptation. Furthermore, ESCAPE does not use keypoint identifiers to alleviate keypoint ambiguity. Experiments show that our method outperform state-of-the-art methods on the standard MP-100 benchmark, assuming keypoint identifiers are not available." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 538, + 136, + 550 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 538, + 136, + 550 + ], + "spans": [ + { + "bbox": [ + 47, + 538, + 136, + 550 + ], + "type": "text", + "content": "2. Related works" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 558, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 558, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 47, + 558, + 289, + 715 + ], + "type": "text", + "content": "2D pose estimation. 2D pose estimation aims to estimate the positions of a set of semantic keypoints of an object from its image. This task holds significant importance across various applications, such as human pose estimation for understanding human behaviors [19, 24, 40, 41], animal pose estimation for wildlife conservation [11, 17, 20, 46], and more. Existing methods can be roughly divided into regression-based methods [3, 26, 34, 35, 37, 39] and heatmap-based methods [1, 2, 4, 5, 18, 33]. The former directly predicts the coordinates of keypoints, while the latter infers heatmaps and takes the peak positions as it predictions. However, these methods are primarily designed for estimating poses for a single (super-)category with a fixed" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "type": "text", + "content": "set of semantic keypoints, and they typically require access to large-scale datasets." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 110, + 547, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 110, + 547, + 350 + ], + "spans": [ + { + "bbox": [ + 304, + 110, + 547, + 350 + ], + "type": "text", + "content": "Few-shot learning. Few-shot learning focuses on learning novel concepts from few examples [14, 45]. While few-shot learning has been extensively studied in domains such as image classification [16] and video classification [25], its application in 2D pose estimation remains relatively under-explored. Our notion of super-keypoints is related to the Bayesian approach in few-shot learning literature [29, 43, 48]. Salakhutdinov et al. [29] introduce super-categories that play role as higher-level abstract concepts for categories and encode distributions over category parameters. Zhang et al. [48] design a Bayesian framework to learn a single super-category that induces a prior over parameters for all categories; such an approach limits model flexibility and is not suitable for CAPE due to the diversity of keypoints. Yang et al. [43] observe that similar categories have nearly identical statistics and propose enhancing novel categories with similar base categories. However, it is not trivial to adopt this method for CAPE due to the lack of a structural way to measure similarity between categories with semantically and quantitatively different keypoints." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 363, + 547, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 363, + 547, + 555 + ], + "spans": [ + { + "bbox": [ + 304, + 363, + 547, + 555 + ], + "type": "text", + "content": "Semantic correspondence. Semantic correspondence refers to the semantic matching between pixels or keypoints between images [6]. A semantic correspondence model typically includes a feature extractor that generates dense features from images and establishes the correspondence using a similarity function. Oquab et al. [27] show that self-supervised ViT models can produce expressive features for finding semantic correspondence. Recently, Tang et al. [36] observe that semantic correspondence naturally emerges as an ability of diffusion models. These models can effectively solve CAPE by matching query pixels to keypoints in support images. Since ESCAPE operates in the feature space, it complements semantic correspondence models. Integrating ESCAPE with an off-the-shelf feature extractor consistently enhances its performance in CAPE tasks." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 564, + 384, + 578 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 564, + 384, + 578 + ], + "spans": [ + { + "bbox": [ + 306, + 564, + 384, + 578 + ], + "type": "text", + "content": "3. Background" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 584, + 489, + 597 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 584, + 489, + 597 + ], + "spans": [ + { + "bbox": [ + 306, + 584, + 489, + 597 + ], + "type": "text", + "content": "3.1. Category-agnostic pose estimation" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 602, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 602, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 602, + 547, + 715 + ], + "type": "text", + "content": "Category-agnostic pose estimation (CAPE) [42] seeks an unified model to estimate poses from any category given a few annotated samples. Each category " + }, + { + "bbox": [ + 304, + 602, + 547, + 715 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 304, + 602, + 547, + 715 + ], + "type": "text", + "content": " has a set of " + }, + { + "bbox": [ + 304, + 602, + 547, + 715 + ], + "type": "inline_equation", + "content": "T_{y}" + }, + { + "bbox": [ + 304, + 602, + 547, + 715 + ], + "type": "text", + "content": " keypoints of interest: " + }, + { + "bbox": [ + 304, + 602, + 547, + 715 + ], + "type": "inline_equation", + "content": "J_{y} = \\{j_{r}\\}_{r = 1}^{T_{y}}" + }, + { + "bbox": [ + 304, + 602, + 547, + 715 + ], + "type": "text", + "content": ". Given an image " + }, + { + "bbox": [ + 304, + 602, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{I}" + }, + { + "bbox": [ + 304, + 602, + 547, + 715 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 304, + 602, + 547, + 715 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 304, + 602, + 547, + 715 + ], + "type": "text", + "content": ", the model's objective is to accurately predict the coordinates " + }, + { + "bbox": [ + 304, + 602, + 547, + 715 + ], + "type": "inline_equation", + "content": "P = \\{\\mathbf{p}_r\\}_{r = 1}^{T_y}" + }, + { + "bbox": [ + 304, + 602, + 547, + 715 + ], + "type": "text", + "content": " of these keypoints. The setting of CAPE consists of two disjoint sets of categories that are " + }, + { + "bbox": [ + 304, + 602, + 547, + 715 + ], + "type": "inline_equation", + "content": "C_{base}" + }, + { + "bbox": [ + 304, + 602, + 547, + 715 + ], + "type": "text", + "content": " for training the model and " + }, + { + "bbox": [ + 304, + 602, + 547, + 715 + ], + "type": "inline_equation", + "content": "C_{novel}" + }, + { + "bbox": [ + 304, + 602, + 547, + 715 + ], + "type": "text", + "content": " for testing its generalization capabilities. While the base categories have abundant" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "23492" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 289, + 193 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 289, + 193 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 289, + 193 + ], + "type": "text", + "content": "training data, the novel categories have very few training samples. Specifically, a training set " + }, + { + "bbox": [ + 47, + 72, + 289, + 193 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{base}^{y} = \\{(I^{i},P^{i})\\}_{i = 1}^{n_{y}}" + }, + { + "bbox": [ + 47, + 72, + 289, + 193 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 47, + 72, + 289, + 193 + ], + "type": "inline_equation", + "content": "n_{y}" + }, + { + "bbox": [ + 47, + 72, + 289, + 193 + ], + "type": "text", + "content": " is large) of each base category " + }, + { + "bbox": [ + 47, + 72, + 289, + 193 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 47, + 72, + 289, + 193 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 47, + 72, + 289, + 193 + ], + "type": "inline_equation", + "content": "C_{base}" + }, + { + "bbox": [ + 47, + 72, + 289, + 193 + ], + "type": "text", + "content": " is available for training the model. The testing stage includes a large number of episodes (or tasks) sampled from the novel categories. In each episode, a support set " + }, + { + "bbox": [ + 47, + 72, + 289, + 193 + ], + "type": "inline_equation", + "content": "S_{novel}^{y} = \\{(\\mathbf{I}^{i},P^{i})\\}_{i = 1}^{K}" + }, + { + "bbox": [ + 47, + 72, + 289, + 193 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 47, + 72, + 289, + 193 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 47, + 72, + 289, + 193 + ], + "type": "text", + "content": " annotated samples from " + }, + { + "bbox": [ + 47, + 72, + 289, + 193 + ], + "type": "inline_equation", + "content": "y\\in C_{novel}" + }, + { + "bbox": [ + 47, + 72, + 289, + 193 + ], + "type": "text", + "content": " is given to the model for adaptation, and the model performance is evaluated on a set " + }, + { + "bbox": [ + 47, + 72, + 289, + 193 + ], + "type": "inline_equation", + "content": "\\mathcal{Q}_{novel}^{y} = \\{(\\mathbf{I}^{i},P^{i})\\}_{i = K + 1}^{K + Q}" + }, + { + "bbox": [ + 47, + 72, + 289, + 193 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 47, + 72, + 289, + 193 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 47, + 72, + 289, + 193 + ], + "type": "text", + "content": " query samples." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 205, + 289, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 205, + 289, + 350 + ], + "spans": [ + { + "bbox": [ + 46, + 205, + 289, + 350 + ], + "type": "text", + "content": "Support keypoint identifier. The MP-100 dataset is composed of existing widely used datasets, many of which impose an ordering over the keypoints of their categories. Shi et al [31] refer to this ordering information as support keypoint identifier and encode it into the keypoint features to alleviate the ambiguity of novel keypoints. However, incorporating keypoint identifiers may limit the applicability of CAPE to real-world scenarios where novel keypoints can be freely defined. In this paper, we focus on the challenging yet practical settings for CAPE that do not assume the order of keypoints to be available. We refer readers to [13] for detailed discussion of the support keypoint identifier." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 354, + 192, + 368 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 354, + 192, + 368 + ], + "spans": [ + { + "bbox": [ + 47, + 354, + 192, + 368 + ], + "type": "text", + "content": "3.2. Simple baseline for CAPE" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 373, + 287, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 373, + 287, + 445 + ], + "spans": [ + { + "bbox": [ + 47, + 373, + 287, + 445 + ], + "type": "text", + "content": "In this section, we introduce a simple yet effective baseline for CAPE. The baseline includes a feature extractor " + }, + { + "bbox": [ + 47, + 373, + 287, + 445 + ], + "type": "inline_equation", + "content": "f_{\\theta}" + }, + { + "bbox": [ + 47, + 373, + 287, + 445 + ], + "type": "text", + "content": " that extracts dense features from input images. We denote the local feature extracted at a position " + }, + { + "bbox": [ + 47, + 373, + 287, + 445 + ], + "type": "inline_equation", + "content": "\\mathbf{p}" + }, + { + "bbox": [ + 47, + 373, + 287, + 445 + ], + "type": "text", + "content": " of an input image " + }, + { + "bbox": [ + 47, + 373, + 287, + 445 + ], + "type": "inline_equation", + "content": "\\mathbf{I}" + }, + { + "bbox": [ + 47, + 373, + 287, + 445 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 47, + 373, + 287, + 445 + ], + "type": "inline_equation", + "content": "f_{\\theta}(\\mathbf{I})[\\mathbf{p}] \\in \\mathbb{R}^{d}" + }, + { + "bbox": [ + 47, + 373, + 287, + 445 + ], + "type": "text", + "content": ". We aim to train " + }, + { + "bbox": [ + 47, + 373, + 287, + 445 + ], + "type": "inline_equation", + "content": "f_{\\theta}" + }, + { + "bbox": [ + 47, + 373, + 287, + 445 + ], + "type": "text", + "content": " so that it possesses strong semantic correspondence for the keypoints of interest." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 445, + 287, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 445, + 287, + 495 + ], + "spans": [ + { + "bbox": [ + 47, + 445, + 287, + 495 + ], + "type": "text", + "content": "For each base category " + }, + { + "bbox": [ + 47, + 445, + 287, + 495 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 47, + 445, + 287, + 495 + ], + "type": "text", + "content": " with keypoints of interest " + }, + { + "bbox": [ + 47, + 445, + 287, + 495 + ], + "type": "inline_equation", + "content": "J_{y}" + }, + { + "bbox": [ + 47, + 445, + 287, + 495 + ], + "type": "text", + "content": " we define a set of prototypes " + }, + { + "bbox": [ + 47, + 445, + 287, + 495 + ], + "type": "inline_equation", + "content": "\\mathbf{W}^y = \\{\\mathbf{w}_r\\}_{r = 1}^{T_y}" + }, + { + "bbox": [ + 47, + 445, + 287, + 495 + ], + "type": "text", + "content": "; each prototype " + }, + { + "bbox": [ + 47, + 445, + 287, + 495 + ], + "type": "inline_equation", + "content": "\\mathbf{w}_r\\in \\mathbb{R}^d" + }, + { + "bbox": [ + 47, + 445, + 287, + 495 + ], + "type": "text", + "content": " is responsible for predicting a heatmap " + }, + { + "bbox": [ + 47, + 445, + 287, + 495 + ], + "type": "inline_equation", + "content": "\\mathbf{H}_r" + }, + { + "bbox": [ + 47, + 445, + 287, + 495 + ], + "type": "text", + "content": " for the " + }, + { + "bbox": [ + 47, + 445, + 287, + 495 + ], + "type": "inline_equation", + "content": "r^{\\mathrm{th}}" + }, + { + "bbox": [ + 47, + 445, + 287, + 495 + ], + "type": "text", + "content": " keypoint in " + }, + { + "bbox": [ + 47, + 445, + 287, + 495 + ], + "type": "inline_equation", + "content": "J_{y}" + }, + { + "bbox": [ + 47, + 445, + 287, + 495 + ], + "type": "text", + "content": " as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 116, + 500, + 287, + 514 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 116, + 500, + 287, + 514 + ], + "spans": [ + { + "bbox": [ + 116, + 500, + 287, + 514 + ], + "type": "interline_equation", + "content": "\\mathbf {H} _ {r} [ \\mathbf {p} ] = \\left\\langle f _ {\\theta} (\\mathbf {I}) [ \\mathbf {p} ], \\mathbf {w} _ {r} \\right\\rangle , \\tag {1}", + "image_path": "86e96da7cd8ada1d0a5e0c0b61a3c8c0ed2297b9b92e5533676ad76e988047cc.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 517, + 287, + 553 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 517, + 287, + 553 + ], + "spans": [ + { + "bbox": [ + 47, + 517, + 287, + 553 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 517, + 287, + 553 + ], + "type": "inline_equation", + "content": "\\langle \\cdot, \\cdot \\rangle" + }, + { + "bbox": [ + 47, + 517, + 287, + 553 + ], + "type": "text", + "content": " denotes the inner product, and " + }, + { + "bbox": [ + 47, + 517, + 287, + 553 + ], + "type": "inline_equation", + "content": "\\mathbf{p}" + }, + { + "bbox": [ + 47, + 517, + 287, + 553 + ], + "type": "text", + "content": " varies over the spatial dimensions. The peak position in the heatmap is then the prediction for that keypoint." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 554, + 287, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 554, + 287, + 613 + ], + "spans": [ + { + "bbox": [ + 47, + 554, + 287, + 613 + ], + "type": "text", + "content": "We train the feature extractor " + }, + { + "bbox": [ + 47, + 554, + 287, + 613 + ], + "type": "inline_equation", + "content": "f_{\\theta}" + }, + { + "bbox": [ + 47, + 554, + 287, + 613 + ], + "type": "text", + "content": " and the base prototypes " + }, + { + "bbox": [ + 47, + 554, + 287, + 613 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{W}^y\\}_{y \\in C_{base}}" + }, + { + "bbox": [ + 47, + 554, + 287, + 613 + ], + "type": "text", + "content": " on the base training data " + }, + { + "bbox": [ + 47, + 554, + 287, + 613 + ], + "type": "inline_equation", + "content": "\\bigcup_{y \\in C_{base}} \\{\\mathcal{D}_{base}^y\\}" + }, + { + "bbox": [ + 47, + 554, + 287, + 613 + ], + "type": "text", + "content": " with the standard Mean Squared Error (MSE) loss [23, 38, 41] between the predicted heatmaps and the ground-truth heatmaps." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 613, + 287, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 613, + 287, + 676 + ], + "spans": [ + { + "bbox": [ + 47, + 613, + 287, + 676 + ], + "type": "text", + "content": "To adapt to a novel category " + }, + { + "bbox": [ + 47, + 613, + 287, + 676 + ], + "type": "inline_equation", + "content": "y \\in C_{\\text{novel}}" + }, + { + "bbox": [ + 47, + 613, + 287, + 676 + ], + "type": "text", + "content": " with support and query sets " + }, + { + "bbox": [ + 47, + 613, + 287, + 676 + ], + "type": "inline_equation", + "content": "(S_{\\text{novel}}^y, Q_{\\text{novel}}^y)" + }, + { + "bbox": [ + 47, + 613, + 287, + 676 + ], + "type": "text", + "content": ", we freeze the feature extractor and compute the prototypes " + }, + { + "bbox": [ + 47, + 613, + 287, + 676 + ], + "type": "inline_equation", + "content": "\\mathbf{W}^y = \\{\\mathbf{w}_r\\}_{r=1}^{T_y}" + }, + { + "bbox": [ + 47, + 613, + 287, + 676 + ], + "type": "text", + "content": " for novel keypoints as the means of keypoint features extracted from the support set:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 96, + 679, + 287, + 711 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 96, + 679, + 287, + 711 + ], + "spans": [ + { + "bbox": [ + 96, + 679, + 287, + 711 + ], + "type": "interline_equation", + "content": "\\mathbf {w} _ {r} = \\frac {1}{K} \\sum_ {\\left(\\mathbf {I} ^ {i}, P ^ {i}\\right) \\in \\mathcal {S} _ {n o v e l} ^ {y}} f _ {\\theta} \\left(\\mathbf {I} ^ {i}\\right) \\left[ \\mathbf {p} _ {r} ^ {i} \\right]. \\tag {2}", + "image_path": "e57698d9ad1ddf04e8943f3878292351de3a1b9a1b7eade43f2ab5cd9bcd5c7d.jpg" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 320, + 71, + 424, + 150 + ], + "blocks": [ + { + "bbox": [ + 320, + 71, + 424, + 150 + ], + "lines": [ + { + "bbox": [ + 320, + 71, + 424, + 150 + ], + "spans": [ + { + "bbox": [ + 320, + 71, + 424, + 150 + ], + "type": "image", + "image_path": "f8eae37c575b5467bca7dbb859695fd1c8f55767793bbe4f3cd04f09a43d6b62.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 350, + 152, + 384, + 161 + ], + "lines": [ + { + "bbox": [ + 350, + 152, + 384, + 161 + ], + "spans": [ + { + "bbox": [ + 350, + 152, + 384, + 161 + ], + "type": "text", + "content": "(a) zebra" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 429, + 71, + 533, + 150 + ], + "blocks": [ + { + "bbox": [ + 429, + 71, + 533, + 150 + ], + "lines": [ + { + "bbox": [ + 429, + 71, + 533, + 150 + ], + "spans": [ + { + "bbox": [ + 429, + 71, + 533, + 150 + ], + "type": "image", + "image_path": "c6f86616d56748aa2372eaadbe603d88a655f0a4cc99e7d596c27f32b5dbe6e8.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 463, + 152, + 495, + 161 + ], + "lines": [ + { + "bbox": [ + 463, + 152, + 495, + 161 + ], + "spans": [ + { + "bbox": [ + 463, + 152, + 495, + 161 + ], + "type": "text", + "content": "(b) horse" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 326, + 255, + 526, + 331 + ], + "blocks": [ + { + "bbox": [ + 304, + 171, + 547, + 247 + ], + "lines": [ + { + "bbox": [ + 304, + 171, + 547, + 247 + ], + "spans": [ + { + "bbox": [ + 304, + 171, + 547, + 247 + ], + "type": "text", + "content": "Figure 2. t-SNE visualization of distributions of 7 keypoints of zebra and horse. Different colors represent different keypoints. " + }, + { + "bbox": [ + 304, + 171, + 547, + 247 + ], + "type": "inline_equation", + "content": "\\times" + }, + { + "bbox": [ + 304, + 171, + 547, + 247 + ], + "type": "text", + "content": " denotes keypoint features extracted from images. (a) shows extracted features (denoted as " + }, + { + "bbox": [ + 304, + 171, + 547, + 247 + ], + "type": "inline_equation", + "content": "\\times" + }, + { + "bbox": [ + 304, + 171, + 547, + 247 + ], + "type": "text", + "content": ") and distribution means (denoted as " + }, + { + "bbox": [ + 304, + 171, + 547, + 247 + ], + "type": "inline_equation", + "content": "\\triangle" + }, + { + "bbox": [ + 304, + 171, + 547, + 247 + ], + "type": "text", + "content": ") for keypoints of zebra. (b) shows extracted features (denoted as " + }, + { + "bbox": [ + 304, + 171, + 547, + 247 + ], + "type": "inline_equation", + "content": "\\times" + }, + { + "bbox": [ + 304, + 171, + 547, + 247 + ], + "type": "text", + "content": ") and distribution means (denoted as " + }, + { + "bbox": [ + 304, + 171, + 547, + 247 + ], + "type": "inline_equation", + "content": "\\circ" + }, + { + "bbox": [ + 304, + 171, + 547, + 247 + ], + "type": "text", + "content": ") for keypoints of horse." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 326, + 255, + 526, + 331 + ], + "lines": [ + { + "bbox": [ + 326, + 255, + 526, + 331 + ], + "spans": [ + { + "bbox": [ + 326, + 255, + 526, + 331 + ], + "type": "image", + "image_path": "d658db9c6c42f13179b34da3b519c74e5545a9181ab05ec8cc45d3c4c53af1d8.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 338, + 547, + 384 + ], + "lines": [ + { + "bbox": [ + 304, + 338, + 547, + 384 + ], + "spans": [ + { + "bbox": [ + 304, + 338, + 547, + 384 + ], + "type": "text", + "content": "Figure 3. Super-keypoint modeling. Keypoint 'right back paw of zebra' and keypoint 'right back paw of horse' belong to the same super-keypoint 'right back paw'. The keypoints distribute closely around their super-keypoint." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 405, + 544, + 429 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 405, + 544, + 429 + ], + "spans": [ + { + "bbox": [ + 304, + 405, + 544, + 429 + ], + "type": "text", + "content": "The computed prototypes are then used to predict keypoint heatmaps for query images as in Eq. 1." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 430, + 545, + 549 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 430, + 545, + 549 + ], + "spans": [ + { + "bbox": [ + 304, + 430, + 545, + 549 + ], + "type": "text", + "content": "Our simple baseline provides results comparable to POMNet [42], which has more complicated matching network. The baseline, however, has the same problem of overfitting as POMNet and CapeFormer due to its heavy reliance on the sparse support samples. Building on our simple baseline, we propose a Bayesian framework that allows us to estimate better prototypes for the novel keypoints, while keeping the feature extractor the same. Specifically, we seek to equip the baseline with strong a prior to prevent the keypoint prototypes from overfitting to the support set." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 305, + 561, + 511, + 575 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 561, + 511, + 575 + ], + "spans": [ + { + "bbox": [ + 305, + 561, + 511, + 575 + ], + "type": "text", + "content": "4. Prior as a mixture of super-keypoints" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "content": "We notice that keypoints of different categories can be visually similar, and hence their features should follow the same distribution. To verify our hypothesis, we use the feature extractor from our baseline model to extract features at the positions of seven keypoints in images of 'zebra' (a base category) and 'horse' (an unseen category). Fig. 2 (a) and (b) show keypoints features extracted from zebra and horse images respectively. For each keypoint, we also visualize the mean of its distribution (denoted as " + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\triangle" + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "content": " for keypoint of zebra and " + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\bigcirc" + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "content": " for keypoint of horse). As shown in the figure, although the model has not seen any horse sample during" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "23493" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 168 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 168 + ], + "type": "text", + "content": "training, the keypoint features of horse still align well with those of zebra. This alignment suggests that capturing the distributions of zebra keypoints facilitates generalization to horse keypoints. From this observation, we introduce super-keypoints to model the connections between keypoints of horse and zebra. For instance, the keypoint 'right back paw of zebra' and the keypoint 'right back paw of horse' belong to the same super-keypoint 'right back paw'." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 176, + 194, + 190 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 176, + 194, + 190 + ], + "spans": [ + { + "bbox": [ + 47, + 176, + 194, + 190 + ], + "type": "text", + "content": "4.1. Super-keypoints modeling" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 194, + 289, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 194, + 289, + 327 + ], + "spans": [ + { + "bbox": [ + 46, + 194, + 289, + 327 + ], + "type": "text", + "content": "We introduce the notion of super-keypoint to define a prior over a group of visually similar keypoints from different categories. Let us consider the " + }, + { + "bbox": [ + 46, + 194, + 289, + 327 + ], + "type": "inline_equation", + "content": "r^{\\mathrm{th}}" + }, + { + "bbox": [ + 46, + 194, + 289, + 327 + ], + "type": "text", + "content": " keypoint of a category " + }, + { + "bbox": [ + 46, + 194, + 289, + 327 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 46, + 194, + 289, + 327 + ], + "type": "text", + "content": ". Given an image " + }, + { + "bbox": [ + 46, + 194, + 289, + 327 + ], + "type": "inline_equation", + "content": "\\mathbf{I}" + }, + { + "bbox": [ + 46, + 194, + 289, + 327 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 46, + 194, + 289, + 327 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 46, + 194, + 289, + 327 + ], + "type": "text", + "content": ", we denote " + }, + { + "bbox": [ + 46, + 194, + 289, + 327 + ], + "type": "inline_equation", + "content": "\\mathbf{o}_r \\coloneqq f_\\theta(\\mathbf{I})[\\mathbf{p}_r]" + }, + { + "bbox": [ + 46, + 194, + 289, + 327 + ], + "type": "text", + "content": " as the local feature extracted at the position of the " + }, + { + "bbox": [ + 46, + 194, + 289, + 327 + ], + "type": "inline_equation", + "content": "r^{\\mathrm{th}}" + }, + { + "bbox": [ + 46, + 194, + 289, + 327 + ], + "type": "text", + "content": " keypoint in " + }, + { + "bbox": [ + 46, + 194, + 289, + 327 + ], + "type": "inline_equation", + "content": "\\mathbf{I}" + }, + { + "bbox": [ + 46, + 194, + 289, + 327 + ], + "type": "text", + "content": ". We assume that the keypoint feature " + }, + { + "bbox": [ + 46, + 194, + 289, + 327 + ], + "type": "inline_equation", + "content": "\\mathbf{o}_r" + }, + { + "bbox": [ + 46, + 194, + 289, + 327 + ], + "type": "text", + "content": " follows a Gaussian distribution with a mean " + }, + { + "bbox": [ + 46, + 194, + 289, + 327 + ], + "type": "inline_equation", + "content": "\\zeta_r" + }, + { + "bbox": [ + 46, + 194, + 289, + 327 + ], + "type": "text", + "content": " and covariance " + }, + { + "bbox": [ + 46, + 194, + 289, + 327 + ], + "type": "inline_equation", + "content": "\\phi^2 I" + }, + { + "bbox": [ + 46, + 194, + 289, + 327 + ], + "type": "text", + "content": ". We further assume that the mean of the keypoint distribution " + }, + { + "bbox": [ + 46, + 194, + 289, + 327 + ], + "type": "inline_equation", + "content": "\\zeta_r" + }, + { + "bbox": [ + 46, + 194, + 289, + 327 + ], + "type": "text", + "content": " obeys a Gaussian distribution with a mean " + }, + { + "bbox": [ + 46, + 194, + 289, + 327 + ], + "type": "inline_equation", + "content": "\\mu_{z_r}" + }, + { + "bbox": [ + 46, + 194, + 289, + 327 + ], + "type": "text", + "content": " and covariance " + }, + { + "bbox": [ + 46, + 194, + 289, + 327 + ], + "type": "inline_equation", + "content": "\\sigma^2 I" + }, + { + "bbox": [ + 46, + 194, + 289, + 327 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 194, + 289, + 327 + ], + "type": "inline_equation", + "content": "\\mu_{z_r}" + }, + { + "bbox": [ + 46, + 194, + 289, + 327 + ], + "type": "text", + "content": " represent the super-keypoint of the " + }, + { + "bbox": [ + 46, + 194, + 289, + 327 + ], + "type": "inline_equation", + "content": "r^{\\mathrm{th}}" + }, + { + "bbox": [ + 46, + 194, + 289, + 327 + ], + "type": "text", + "content": " keypoint. Formally," + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 63, + 335, + 287, + 350 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 335, + 287, + 350 + ], + "spans": [ + { + "bbox": [ + 63, + 335, + 287, + 350 + ], + "type": "interline_equation", + "content": "\\mathbf {o} _ {r} \\sim \\mathcal {N} (\\mathbf {o} | \\zeta_ {r}; \\phi^ {2}), \\quad \\text {a n d} \\quad \\zeta_ {r} \\sim \\mathcal {N} (\\zeta | \\mu_ {z _ {r}}; \\sigma^ {2}). \\qquad (3)", + "image_path": "745415da43abddb9e26922e8abf40caf7494cc8eda619c780a210ed075b920b6.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 359, + 289, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 359, + 289, + 491 + ], + "spans": [ + { + "bbox": [ + 46, + 359, + 289, + 491 + ], + "type": "text", + "content": "Fig. 3 illustrates our super-keypoint modeling. Intuitively, " + }, + { + "bbox": [ + 46, + 359, + 289, + 491 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 46, + 359, + 289, + 491 + ], + "type": "text", + "content": " is the representation of a super-keypoint in the feature space, and its keypoint members distribute closely nearby. Our goal is to discover and estimate a set of superkeypoints from keypoints of base categories so that the super-keypoints can encode the statistics of groups of related keypoints across different categories. When adapting to a novel category, the statistics encoded in the superkeypoints are directly transferred to the prototypes of novel keypoints, preventing our model from overfitting to the sparse support set." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 499, + 264, + 513 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 499, + 264, + 513 + ], + "spans": [ + { + "bbox": [ + 47, + 499, + 264, + 513 + ], + "type": "text", + "content": "4.2. Super-keypoints discovering and learning" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 517, + 287, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 517, + 287, + 590 + ], + "spans": [ + { + "bbox": [ + 46, + 517, + 287, + 590 + ], + "type": "text", + "content": "We obtain the set of super-keypoints by applying a clustering method to the keypoints of all base categories. Specifically, given the trained feature extractor " + }, + { + "bbox": [ + 46, + 517, + 287, + 590 + ], + "type": "inline_equation", + "content": "f_{\\theta}" + }, + { + "bbox": [ + 46, + 517, + 287, + 590 + ], + "type": "text", + "content": " from the baseline, we extract the features of keypoints from all base categories. We then estimate the distribution mean of each keypoint in the base categories as follows:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 88, + 598, + 287, + 631 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 88, + 598, + 287, + 631 + ], + "spans": [ + { + "bbox": [ + 88, + 598, + 287, + 631 + ], + "type": "interline_equation", + "content": "\\zeta_ {r} ^ {y} = \\frac {1}{| \\mathcal {D} _ {\\text {b a s e}} ^ {y} |} \\sum_ {(\\mathbf {I} ^ {i}, P ^ {i}) \\in \\mathcal {D} _ {\\text {b a s e}} ^ {y}} f _ {\\theta} (\\mathbf {I} ^ {i}) [ \\mathbf {p} _ {r} ^ {i} ], \\tag {4}", + "image_path": "207a705fc3c8d48f1d0b0fa03ed3dd6f5dc69bb9c483d95a5bebfedb5df9d6dc.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 639, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 639, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 639, + 288, + 715 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 639, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\zeta_r^y" + }, + { + "bbox": [ + 46, + 639, + 288, + 715 + ], + "type": "text", + "content": " represents the distribution mean of the " + }, + { + "bbox": [ + 46, + 639, + 288, + 715 + ], + "type": "inline_equation", + "content": "r^{\\mathrm{th}}" + }, + { + "bbox": [ + 46, + 639, + 288, + 715 + ], + "type": "text", + "content": " keypoint of category " + }, + { + "bbox": [ + 46, + 639, + 288, + 715 + ], + "type": "inline_equation", + "content": "y\\in C_{base}" + }, + { + "bbox": [ + 46, + 639, + 288, + 715 + ], + "type": "text", + "content": ". The set of keypoints from all base categories is then " + }, + { + "bbox": [ + 46, + 639, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\bigcup_{y\\in C_{base}}\\{\\zeta_r^y\\}_{r = 1}^{T_y}" + }, + { + "bbox": [ + 46, + 639, + 288, + 715 + ], + "type": "text", + "content": ". In this work, we use the k-nearest neighbor based density peaks clustering [9, 47], which does not require the number of clusters to be known in advance. Note that keypoints of the same" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 72, + 546, + 145 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 546, + 145 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 546, + 145 + ], + "type": "text", + "content": "category are distinct, so they cannot belong to the same super-keypoint. A detailed description of the clustering is provided in the Appendix. Finally, we estimate each super-keypoint representation " + }, + { + "bbox": [ + 304, + 72, + 546, + 145 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 304, + 72, + 546, + 145 + ], + "type": "text", + "content": " as the empirical mean of distribution means of keypoints within the cluster, resulting in a set of " + }, + { + "bbox": [ + 304, + 72, + 546, + 145 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 304, + 72, + 546, + 145 + ], + "type": "text", + "content": " super-keypoints: " + }, + { + "bbox": [ + 304, + 72, + 546, + 145 + ], + "type": "inline_equation", + "content": "\\mathbf{M} = \\{\\mu_1,\\mu_2,\\dots ,\\mu_L\\}" + }, + { + "bbox": [ + 304, + 72, + 546, + 145 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 157, + 456, + 171 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 157, + 456, + 171 + ], + "spans": [ + { + "bbox": [ + 305, + 157, + 456, + 171 + ], + "type": "text", + "content": "5. Novel category adaptation" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 178, + 547, + 322 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 178, + 547, + 322 + ], + "spans": [ + { + "bbox": [ + 304, + 178, + 547, + 322 + ], + "type": "text", + "content": "In this section, we describe how to adapt to a novel category " + }, + { + "bbox": [ + 304, + 178, + 547, + 322 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 304, + 178, + 547, + 322 + ], + "type": "text", + "content": " whose support and query sets are " + }, + { + "bbox": [ + 304, + 178, + 547, + 322 + ], + "type": "inline_equation", + "content": "(\\mathcal{S}_{novel}^{y},\\mathcal{Q}_{novel}^{y})" + }, + { + "bbox": [ + 304, + 178, + 547, + 322 + ], + "type": "text", + "content": ". Our goal is to estimate the prototypes that well represent the distributions of the novel keypoints from a few support samples. In particular, for the " + }, + { + "bbox": [ + 304, + 178, + 547, + 322 + ], + "type": "inline_equation", + "content": "r^{\\mathrm{th}}" + }, + { + "bbox": [ + 304, + 178, + 547, + 322 + ], + "type": "text", + "content": " keypoint of " + }, + { + "bbox": [ + 304, + 178, + 547, + 322 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 304, + 178, + 547, + 322 + ], + "type": "text", + "content": ", we want the estimated prototype " + }, + { + "bbox": [ + 304, + 178, + 547, + 322 + ], + "type": "inline_equation", + "content": "\\mathbf{w}_r" + }, + { + "bbox": [ + 304, + 178, + 547, + 322 + ], + "type": "text", + "content": " to be close to the true distribution mean " + }, + { + "bbox": [ + 304, + 178, + 547, + 322 + ], + "type": "inline_equation", + "content": "\\zeta_r" + }, + { + "bbox": [ + 304, + 178, + 547, + 322 + ], + "type": "text", + "content": " as much as possible. With this interpretation, the estimated prototype for a novel keypoint in the baseline (Eq. 2) can be viewed as maximum likelihood estimate of the mean of the keypoint distribution. It is clear that the baseline is prone to overfitting since the number of support samples is extremely limited." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 331, + 541, + 345 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 331, + 541, + 345 + ], + "spans": [ + { + "bbox": [ + 305, + 331, + 541, + 345 + ], + "type": "text", + "content": "5.1. Prototype estimate via maximum a posteriori" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 351, + 547, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 351, + 547, + 411 + ], + "spans": [ + { + "bbox": [ + 304, + 351, + 547, + 411 + ], + "type": "text", + "content": "We propose to alleviate the overfitting by letting the statistics captured in the super-keypoints to be transferred to the novel keypoints. Specifically, consider the " + }, + { + "bbox": [ + 304, + 351, + 547, + 411 + ], + "type": "inline_equation", + "content": "r^{\\mathrm{th}}" + }, + { + "bbox": [ + 304, + 351, + 547, + 411 + ], + "type": "text", + "content": " keypoint of " + }, + { + "bbox": [ + 304, + 351, + 547, + 411 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 304, + 351, + 547, + 411 + ], + "type": "text", + "content": ", we compute its prototype " + }, + { + "bbox": [ + 304, + 351, + 547, + 411 + ], + "type": "inline_equation", + "content": "\\mathbf{w}_r" + }, + { + "bbox": [ + 304, + 351, + 547, + 411 + ], + "type": "text", + "content": " as maximum a posterior estimate of " + }, + { + "bbox": [ + 304, + 351, + 547, + 411 + ], + "type": "inline_equation", + "content": "\\zeta_r" + }, + { + "bbox": [ + 304, + 351, + 547, + 411 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 340, + 420, + 545, + 444 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 340, + 420, + 545, + 444 + ], + "spans": [ + { + "bbox": [ + 340, + 420, + 545, + 444 + ], + "type": "interline_equation", + "content": "\\mathbf {w} _ {r} ^ {M A P} = \\underset {\\zeta_ {r}} {\\operatorname {a r g m a x}} \\log p \\left(\\zeta_ {r} \\mid \\left\\{\\mathbf {o} _ {r} ^ {i} \\right\\} _ {i = 1} ^ {K}, \\mathbf {M}\\right), \\tag {5}", + "image_path": "4e89db4a9642f5ca2725d9e601ed7a7bae8dcf6db302ab8fc915c6b585d041a7.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 453, + 545, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 453, + 545, + 491 + ], + "spans": [ + { + "bbox": [ + 305, + 453, + 545, + 491 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 453, + 545, + 491 + ], + "type": "inline_equation", + "content": "\\mathbf{o}_r^i = f_\\theta (\\mathbf{I}^i)[\\mathbf{p}_r^i ]" + }, + { + "bbox": [ + 305, + 453, + 545, + 491 + ], + "type": "text", + "content": " is the feature of the " + }, + { + "bbox": [ + 305, + 453, + 545, + 491 + ], + "type": "inline_equation", + "content": "r^{\\mathrm{th}}" + }, + { + "bbox": [ + 305, + 453, + 545, + 491 + ], + "type": "text", + "content": " keypoint extracted from the " + }, + { + "bbox": [ + 305, + 453, + 545, + 491 + ], + "type": "inline_equation", + "content": "i^{\\mathrm{th}}" + }, + { + "bbox": [ + 305, + 453, + 545, + 491 + ], + "type": "text", + "content": " support sample in " + }, + { + "bbox": [ + 305, + 453, + 545, + 491 + ], + "type": "inline_equation", + "content": "\\mathcal{S}_{novel}^{y}" + }, + { + "bbox": [ + 305, + 453, + 545, + 491 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 305, + 453, + 545, + 491 + ], + "type": "inline_equation", + "content": "\\mathbf{M} = \\{\\mu_1,\\mu_2,\\dots ,\\mu_L\\}" + }, + { + "bbox": [ + 305, + 453, + 545, + 491 + ], + "type": "text", + "content": " is the set of learned super-keypoints." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 491, + 547, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 491, + 547, + 624 + ], + "spans": [ + { + "bbox": [ + 304, + 491, + 547, + 624 + ], + "type": "text", + "content": "However, when adapting to a novel category, the superkeypoints of the novel keypoints are not available in general. We do not know which super-keypoint the " + }, + { + "bbox": [ + 304, + 491, + 547, + 624 + ], + "type": "inline_equation", + "content": "r^{\\mathrm{th}}" + }, + { + "bbox": [ + 304, + 491, + 547, + 624 + ], + "type": "text", + "content": " novel keypoint should belong to. To this extent, we adopt the expectation maximization (EM) algorithm [21], which first matches the " + }, + { + "bbox": [ + 304, + 491, + 547, + 624 + ], + "type": "inline_equation", + "content": "r^{\\mathrm{th}}" + }, + { + "bbox": [ + 304, + 491, + 547, + 624 + ], + "type": "text", + "content": " keypoint with the 'best' super-keypoint from M (the expectation step) and then optimizes for the prototype with the prior encoded in the matched super-keypoint (the maximization step). Below, we summarize the two steps; detailed derivation of the (EM) algorithm is provided in the Appendix." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 642, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 642, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 642, + 547, + 715 + ], + "type": "text", + "content": "Expectation step. Let " + }, + { + "bbox": [ + 304, + 642, + 547, + 715 + ], + "type": "inline_equation", + "content": "z_{r}" + }, + { + "bbox": [ + 304, + 642, + 547, + 715 + ], + "type": "text", + "content": " be the super-keypoint assignment of the " + }, + { + "bbox": [ + 304, + 642, + 547, + 715 + ], + "type": "inline_equation", + "content": "r^{\\mathrm{th}}" + }, + { + "bbox": [ + 304, + 642, + 547, + 715 + ], + "type": "text", + "content": " keypoint of " + }, + { + "bbox": [ + 304, + 642, + 547, + 715 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 304, + 642, + 547, + 715 + ], + "type": "text", + "content": ". We align the keypoint with the most likely super-keypoint " + }, + { + "bbox": [ + 304, + 642, + 547, + 715 + ], + "type": "inline_equation", + "content": "z_{r}^{\\star}" + }, + { + "bbox": [ + 304, + 642, + 547, + 715 + ], + "type": "text", + "content": ", i.e., the " + }, + { + "bbox": [ + 304, + 642, + 547, + 715 + ], + "type": "inline_equation", + "content": "z_{r}" + }, + { + "bbox": [ + 304, + 642, + 547, + 715 + ], + "type": "text", + "content": " that maximizes " + }, + { + "bbox": [ + 304, + 642, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\log p(\\zeta_{r}, z_{r}|\\{\\mathbf{o}_{r}^{i}\\}_{i=1}^{K}, \\mathbf{M})" + }, + { + "bbox": [ + 304, + 642, + 547, + 715 + ], + "type": "text", + "content": ". Note that keypoints of the same category cannot belong to the same super-keypoint. Therefore, we further adopt the Hungarian algorithm [15]" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "23494" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 85, + 68, + 512, + 319 + ], + "blocks": [ + { + "bbox": [ + 85, + 68, + 512, + 319 + ], + "lines": [ + { + "bbox": [ + 85, + 68, + 512, + 319 + ], + "spans": [ + { + "bbox": [ + 85, + 68, + 512, + 319 + ], + "type": "image", + "image_path": "4ab82e4fd947cf03830e88ec00198e6ff4371112cabfe569843345bd3b004f0b.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 327, + 547, + 384 + ], + "lines": [ + { + "bbox": [ + 46, + 327, + 547, + 384 + ], + "spans": [ + { + "bbox": [ + 46, + 327, + 547, + 384 + ], + "type": "text", + "content": "Figure 4. Illustration of our adaptation procedure for a novel category. We leverage the EM algorithm to probabilistically optimize for the keypoint prototypes, considering both the support samples and the learned super-keypoints. We first align the novel keypoints with the different super-keypoints via a learnable matching network (expectation step). We then let the statistics encoded in the matched super-keypoints to be transferred directly to the keypoint prototypes (maximization step). The keypoint prototypes can be used to estimate poses for query samples." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 404, + 288, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 404, + 288, + 443 + ], + "spans": [ + { + "bbox": [ + 46, + 404, + 288, + 443 + ], + "type": "text", + "content": "to optimally align the " + }, + { + "bbox": [ + 46, + 404, + 288, + 443 + ], + "type": "inline_equation", + "content": "T_{y}" + }, + { + "bbox": [ + 46, + 404, + 288, + 443 + ], + "type": "text", + "content": " keypoints of " + }, + { + "bbox": [ + 46, + 404, + 288, + 443 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 46, + 404, + 288, + 443 + ], + "type": "text", + "content": " with distinct superkeypoints " + }, + { + "bbox": [ + 46, + 404, + 288, + 443 + ], + "type": "inline_equation", + "content": "\\{\\mu_{z_r^*}\\}_{r = 1}^{T_y}\\subset \\mathbf{M}" + }, + { + "bbox": [ + 46, + 404, + 288, + 443 + ], + "type": "text", + "content": ". We refer to this matching procedure as simple matching for the rest of the paper." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 457, + 287, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 457, + 287, + 495 + ], + "spans": [ + { + "bbox": [ + 46, + 457, + 287, + 495 + ], + "type": "text", + "content": "Maximization step. Now, we know the 'best' superkeypoint for each keypoint of " + }, + { + "bbox": [ + 46, + 457, + 287, + 495 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 46, + 457, + 287, + 495 + ], + "type": "text", + "content": ". We can compute their prototypes analytically as follows:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 79, + 514, + 287, + 570 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 79, + 514, + 287, + 570 + ], + "spans": [ + { + "bbox": [ + 79, + 514, + 287, + 570 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathbf {w} _ {r} ^ {M A P} = \\underset {\\mathbf {w} _ {r}} {\\operatorname {a r g m a x}} \\log p \\left(\\mathbf {w} _ {r} | \\left\\{\\mathbf {o} _ {r} ^ {i} \\right\\} _ {i = 1} ^ {K}, \\mu_ {z _ {r} ^ {\\star}}\\right) (6) \\\\ = \\alpha \\frac {1}{K} \\sum_ {i = 1} ^ {K} \\mathbf {o} _ {r} ^ {i} + (1 - \\alpha) \\mu_ {z _ {r} ^ {\\star}}, (7) \\\\ \\end{array}", + "image_path": "b9da502dafdce3aa6dc6eab51aa2f0c24014539409a62604f86f405a82144f80.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 579, + 288, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 579, + 288, + 652 + ], + "spans": [ + { + "bbox": [ + 46, + 579, + 288, + 652 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 579, + 288, + 652 + ], + "type": "inline_equation", + "content": "\\alpha \\in (0,1)" + }, + { + "bbox": [ + 46, + 579, + 288, + 652 + ], + "type": "text", + "content": " depends on " + }, + { + "bbox": [ + 46, + 579, + 288, + 652 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 46, + 579, + 288, + 652 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 579, + 288, + 652 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 46, + 579, + 288, + 652 + ], + "type": "text", + "content": ". Since the distribution mean of a keypoint must distribute near its super-keypoint, the second term in Eq. 7 has an effect of pulling the estimated prototype closer to the true mean of the keypoint distribution, diminishing the dependence of the prototype on the support samples." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 658, + 206, + 672 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 658, + 206, + 672 + ], + "spans": [ + { + "bbox": [ + 47, + 658, + 206, + 672 + ], + "type": "text", + "content": "5.2. Learnable matching network" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 677, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 288, + 715 + ], + "type": "text", + "content": "Matching novel keypoints with semantically related superkeypoints is crucial for the success of our method. A correct super-keypoint can pull the estimated prototype closer" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 404, + 545, + 513 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 404, + 545, + 513 + ], + "spans": [ + { + "bbox": [ + 304, + 404, + 545, + 513 + ], + "type": "text", + "content": "to the distribution mean of the keypoint under consideration, while an incorrect super-keypoint pulls the prototype away. However, as later shown in the experiments, the simple matching in the expectation step fails to produce reliable super-keypoints assignments. We argue that this is because the simple matching treats the novel keypoints independently, discarding their structural relations. It ignores the keypoint coordinates in the support samples, and so the model cannot reason about poses of the object of interest." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 513, + 546, + 597 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 513, + 546, + 597 + ], + "spans": [ + { + "bbox": [ + 304, + 513, + 546, + 597 + ], + "type": "text", + "content": "To overcome these limitations, we replace the simple matching in the expectation step with a learnable matching network to simultaneously produce matching scores between keypoints in an image and the learned super keypoints. The design is inspired by SuperGlue [30], a matching framework that has shown great performance in aligning two sequences of features." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 618, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 618, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 618, + 547, + 715 + ], + "type": "text", + "content": "Input features. Our matching network takes two sequences of features as input and produces the matching score matrix for their elements. The first sequence is augmented features of " + }, + { + "bbox": [ + 304, + 618, + 547, + 715 + ], + "type": "inline_equation", + "content": "T_{y}" + }, + { + "bbox": [ + 304, + 618, + 547, + 715 + ], + "type": "text", + "content": " keypoints extracted from a support image of " + }, + { + "bbox": [ + 304, + 618, + 547, + 715 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 304, + 618, + 547, + 715 + ], + "type": "text", + "content": ". The second sequence is a set of features for the " + }, + { + "bbox": [ + 304, + 618, + 547, + 715 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 304, + 618, + 547, + 715 + ], + "type": "text", + "content": " super-keypoints. To encode features for the keypoints, we consider not only their visual appearance but also their coordinates within the support image. Specifically, given" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "type": "text", + "content": "23495" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 75, + 70, + 519, + 148 + ], + "blocks": [ + { + "bbox": [ + 75, + 70, + 519, + 148 + ], + "lines": [ + { + "bbox": [ + 75, + 70, + 519, + 148 + ], + "spans": [ + { + "bbox": [ + 75, + 70, + 519, + 148 + ], + "type": "table", + "html": "
MethodSplit 1Split 2Split 3Split 4Split 5Mean PCKSplit 1Split 2Split 3Split 4Split 5Mean PCK
ProtoNet [32]46.0540.8449.1343.3444.5444.7860.3153.5161.9258.4458.6158.56
MAML [10]68.1454.7264.1963.2457.2061.5070.0355.9863.2164.7958.4762.50
Finetune [22]70.6057.0466.0665.0059.2063.5871.6757.8466.7666.5360.2464.61
POMNet [42]84.2378.2578.1778.6879.1779.7084.7279.6178.0080.3880.8580.71
CapeFormer [31]84.1379.9079.5879.7880.3480.7590.4686.4785.8086.4786.7187.19
ESCAPE86.8982.5581.2581.7281.3282.7491.4187.4385.3387.2786.7687.63
", + "image_path": "81785062c73240156340f985726695a0a06563cb629eeadba3a834f28fa4e7a1.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 59, + 157, + 532, + 168 + ], + "lines": [ + { + "bbox": [ + 59, + 157, + 532, + 168 + ], + "spans": [ + { + "bbox": [ + 59, + 157, + 532, + 168 + ], + "type": "text", + "content": "Table 1. Comparison with state-of-the-art works on MP-100 dataset under 1-shot and 5-shot settings. " + }, + { + "bbox": [ + 59, + 157, + 532, + 168 + ], + "type": "inline_equation", + "content": "{}^{ \\dagger }" + }, + { + "bbox": [ + 59, + 157, + 532, + 168 + ], + "type": "text", + "content": " denotes keypoint identifiers." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 47, + 190, + 258, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 190, + 258, + 204 + ], + "spans": [ + { + "bbox": [ + 47, + 190, + 258, + 204 + ], + "type": "text", + "content": "support image " + }, + { + "bbox": [ + 47, + 190, + 258, + 204 + ], + "type": "inline_equation", + "content": "\\mathbf{I}\\sim S_{novel}^{y}" + }, + { + "bbox": [ + 47, + 190, + 258, + 204 + ], + "type": "text", + "content": " , the keypoint features are:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 210, + 287, + 224 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 210, + 287, + 224 + ], + "spans": [ + { + "bbox": [ + 53, + 210, + 287, + 224 + ], + "type": "interline_equation", + "content": "\\mathbf {u} _ {r} = f _ {\\gamma} (\\mathbf {I}) \\left[ \\mathbf {p} _ {r} \\right] + P E _ {\\mathrm {g e o}} \\left(\\mathbf {p} _ {r}\\right), \\text {f o r} r = \\{1, \\dots , T _ {y} \\}, \\tag {8}", + "image_path": "1be8f0c2e2b4dc295a00535f404f27ee2a36b094971568ad040801c22dbf7fbc.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 232, + 287, + 280 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 232, + 287, + 280 + ], + "spans": [ + { + "bbox": [ + 47, + 232, + 287, + 280 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 232, + 287, + 280 + ], + "type": "inline_equation", + "content": "f_{\\gamma}" + }, + { + "bbox": [ + 47, + 232, + 287, + 280 + ], + "type": "text", + "content": " is a CNN feature extractor, and " + }, + { + "bbox": [ + 47, + 232, + 287, + 280 + ], + "type": "inline_equation", + "content": "PE_{\\mathrm{geo}}" + }, + { + "bbox": [ + 47, + 232, + 287, + 280 + ], + "type": "text", + "content": " embeds a coordinate into the feature space. By considering the keypoint coordinates, our model can reason about the pose of the object in the support image." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 280, + 287, + 303 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 280, + 287, + 303 + ], + "spans": [ + { + "bbox": [ + 47, + 280, + 287, + 303 + ], + "type": "text", + "content": "At the other end, the features for the super-keypoints are defined as:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 75, + 312, + 287, + 326 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 312, + 287, + 326 + ], + "spans": [ + { + "bbox": [ + 75, + 312, + 287, + 326 + ], + "type": "interline_equation", + "content": "\\mathbf {v} _ {z} = \\omega_ {z} + P E _ {\\mathrm {a b s}} (z), \\quad \\text {f o r} z \\in \\{1, \\dots , L \\}, \\tag {9}", + "image_path": "8ee627c12ddb0c1c0cf8127f0c928a00cdc94133db66185489dbdf3a7e0cdc0b.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 335, + 287, + 383 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 335, + 287, + 383 + ], + "spans": [ + { + "bbox": [ + 47, + 335, + 287, + 383 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 335, + 287, + 383 + ], + "type": "inline_equation", + "content": "\\omega_{z}" + }, + { + "bbox": [ + 47, + 335, + 287, + 383 + ], + "type": "text", + "content": " is a learnable feature vector for the " + }, + { + "bbox": [ + 47, + 335, + 287, + 383 + ], + "type": "inline_equation", + "content": "z^{\\mathrm{th}}" + }, + { + "bbox": [ + 47, + 335, + 287, + 383 + ], + "type": "text", + "content": " superkeypoint, and " + }, + { + "bbox": [ + 47, + 335, + 287, + 383 + ], + "type": "inline_equation", + "content": "PE_{\\mathrm{abs}}" + }, + { + "bbox": [ + 47, + 335, + 287, + 383 + ], + "type": "text", + "content": " embeds an index " + }, + { + "bbox": [ + 47, + 335, + 287, + 383 + ], + "type": "inline_equation", + "content": "(1,\\dots ,L)" + }, + { + "bbox": [ + 47, + 335, + 287, + 383 + ], + "type": "text", + "content": " into the feature space. Note that the indexes of the super-keypoints are obtained from the clustering step described in Sec. 4.2." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 397, + 287, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 397, + 287, + 533 + ], + "spans": [ + { + "bbox": [ + 47, + 397, + 287, + 533 + ], + "type": "text", + "content": "Feature aggregation with attention. The two sequences of features " + }, + { + "bbox": [ + 47, + 397, + 287, + 533 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{u}_r\\}_{r=1}^{T_y}" + }, + { + "bbox": [ + 47, + 397, + 287, + 533 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 397, + 287, + 533 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{v}_z\\}_{z=1}^L" + }, + { + "bbox": [ + 47, + 397, + 287, + 533 + ], + "type": "text", + "content": " are then passed to a neural network for information aggregation. The network comprises " + }, + { + "bbox": [ + 47, + 397, + 287, + 533 + ], + "type": "inline_equation", + "content": "U" + }, + { + "bbox": [ + 47, + 397, + 287, + 533 + ], + "type": "text", + "content": " blocks of self-attention and cross-attention to model interactions both within and between the elements of the sequences. Results are two sequences of refined features: " + }, + { + "bbox": [ + 47, + 397, + 287, + 533 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{u}_r^\\star\\}_{r=1}^{T_y}" + }, + { + "bbox": [ + 47, + 397, + 287, + 533 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 397, + 287, + 533 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{v}_z^\\star\\}_{z=1}^L" + }, + { + "bbox": [ + 47, + 397, + 287, + 533 + ], + "type": "text", + "content": ". The predicted matching scores matrix is then " + }, + { + "bbox": [ + 47, + 397, + 287, + 533 + ], + "type": "inline_equation", + "content": "\\mathbf{S} \\in \\mathbb{R}^{T_y \\times L}" + }, + { + "bbox": [ + 47, + 397, + 287, + 533 + ], + "type": "text", + "content": ", with each element " + }, + { + "bbox": [ + 47, + 397, + 287, + 533 + ], + "type": "inline_equation", + "content": "\\mathbf{S}_{r,z} = \\langle \\mathbf{u}_r^\\star, \\mathbf{v}_z^\\star \\rangle" + }, + { + "bbox": [ + 47, + 397, + 287, + 533 + ], + "type": "text", + "content": " measuring the matching score between the " + }, + { + "bbox": [ + 47, + 397, + 287, + 533 + ], + "type": "inline_equation", + "content": "r^{\\text{th}}" + }, + { + "bbox": [ + 47, + 397, + 287, + 533 + ], + "type": "text", + "content": " keypoint and the " + }, + { + "bbox": [ + 47, + 397, + 287, + 533 + ], + "type": "inline_equation", + "content": "z^{\\text{th}}" + }, + { + "bbox": [ + 47, + 397, + 287, + 533 + ], + "type": "text", + "content": " super-keypoint. An illustration of the matching network is provided at the bottom of Fig. 4." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 546, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 546, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 546, + 287, + 714 + ], + "type": "text", + "content": "Matching as optimal transport problem. Since the keypoints within the same category must be matched to different super-keypoints, our goal is to find a maximum score bipartite matching between the two sequences. Following SuperGlue [30], we replace the Hungarian algorithm with the differentiable Sinkhorn algorithm [7] for optimal transport. It takes the matching score matrix " + }, + { + "bbox": [ + 47, + 546, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{S}" + }, + { + "bbox": [ + 47, + 546, + 287, + 714 + ], + "type": "text", + "content": " as input and produces a matrix " + }, + { + "bbox": [ + 47, + 546, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{Q} \\in \\mathbb{R}^{T_y \\times L}" + }, + { + "bbox": [ + 47, + 546, + 287, + 714 + ], + "type": "text", + "content": " where each element " + }, + { + "bbox": [ + 47, + 546, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}_{r,z}" + }, + { + "bbox": [ + 47, + 546, + 287, + 714 + ], + "type": "text", + "content": " represents the probability that the " + }, + { + "bbox": [ + 47, + 546, + 287, + 714 + ], + "type": "inline_equation", + "content": "r^{\\text{th}}" + }, + { + "bbox": [ + 47, + 546, + 287, + 714 + ], + "type": "text", + "content": " keypoint is matched with the " + }, + { + "bbox": [ + 47, + 546, + 287, + 714 + ], + "type": "inline_equation", + "content": "z^{\\text{th}}" + }, + { + "bbox": [ + 47, + 546, + 287, + 714 + ], + "type": "text", + "content": " super-keypoint. We train the network end-to-end on the keypoints of base categories, whose super-keypoints assignments are available from the clustering step. Details of the Sinkhorn algorithm and the training procedure of the matching network are provided in the Appendix." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 190, + 545, + 285 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 190, + 545, + 285 + ], + "spans": [ + { + "bbox": [ + 304, + 190, + 545, + 285 + ], + "type": "text", + "content": "Incorporating learnable matching network to novel category adaptation. We integrate the learned matching network into the novel category adaptation. Specifically, for every support image " + }, + { + "bbox": [ + 304, + 190, + 545, + 285 + ], + "type": "inline_equation", + "content": "\\mathbf{I}^i\\in S_{novel}^y" + }, + { + "bbox": [ + 304, + 190, + 545, + 285 + ], + "type": "text", + "content": ", we compute the matching score matrix " + }, + { + "bbox": [ + 304, + 190, + 545, + 285 + ], + "type": "inline_equation", + "content": "\\mathbf{S}^i" + }, + { + "bbox": [ + 304, + 190, + 545, + 285 + ], + "type": "text", + "content": " from its keypoints and the learned super keypoints. The corresponding optimal assignment matrix is then denoted as " + }, + { + "bbox": [ + 304, + 190, + 545, + 285 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}^i" + }, + { + "bbox": [ + 304, + 190, + 545, + 285 + ], + "type": "text", + "content": ". The final prototype for the " + }, + { + "bbox": [ + 304, + 190, + 545, + 285 + ], + "type": "inline_equation", + "content": "r^{\\mathrm{th}}" + }, + { + "bbox": [ + 304, + 190, + 545, + 285 + ], + "type": "text", + "content": " keypoint are then estimated as follows:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 315, + 293, + 545, + 338 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 293, + 545, + 338 + ], + "spans": [ + { + "bbox": [ + 315, + 293, + 545, + 338 + ], + "type": "interline_equation", + "content": "\\mathbf {w} _ {r} ^ {\\text {E S C A P E}} = \\alpha \\frac {1}{K} \\sum_ {i = 1} ^ {K} \\mathbf {o} _ {r} ^ {i} + (1 - \\alpha) \\frac {L}{K} \\sum_ {i = 1} ^ {K} \\sum_ {z = 1} ^ {L} \\mathbf {Q} _ {r, z} ^ {i} \\mu_ {z}. \\tag {10}", + "image_path": "d87a6fa1e5f6b50caa07c408450bec4c8975088f7659665aa7a5c44096c87fc8.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 350, + 545, + 387 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 350, + 545, + 387 + ], + "spans": [ + { + "bbox": [ + 304, + 350, + 545, + 387 + ], + "type": "text", + "content": "Here we keep the soft assignment form of " + }, + { + "bbox": [ + 304, + 350, + 545, + 387 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}^i" + }, + { + "bbox": [ + 304, + 350, + 545, + 387 + ], + "type": "text", + "content": " instead of the hard assignment to model uncertainty. Our adaptation procedure is shown in Fig. 4." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 399, + 387, + 414 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 399, + 387, + 414 + ], + "spans": [ + { + "bbox": [ + 306, + 399, + 387, + 414 + ], + "type": "text", + "content": "6. Experiments" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 420, + 545, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 420, + 545, + 552 + ], + "spans": [ + { + "bbox": [ + 304, + 420, + 545, + 552 + ], + "type": "text", + "content": "We perform experiments on Multi-category Pose (MP-100) proposed in [42]. The dataset consists of more than 20,000 data instances of 100 categories from 8 super-categories including human hand, human face, human body, animal body, animal face, clothes, furniture, and vehicle. We also follow the five splits from [42] for standard CAPE settings. Each split divides the dataset into 70 training categories, 10 validation categories, and 20 testing categories. The training set is used as base categories, while validation and testing sets are novel categories for validation and testing, respectively. Please refer to [42] for more details." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 562, + 425, + 574 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 562, + 425, + 574 + ], + "spans": [ + { + "bbox": [ + 306, + 562, + 425, + 574 + ], + "type": "text", + "content": "6.1. Experimental details" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 581, + 545, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 581, + 545, + 652 + ], + "spans": [ + { + "bbox": [ + 304, + 581, + 545, + 652 + ], + "type": "text", + "content": "Unless otherwise specified, for fair comparisons with previous works, we utilize ResNet-50 [12] pretrained on ImageNet [8] as the feature extractor. Also following [42], we crop the object of interest from the original image using the ground-truth bounding box, and resize it to a fixed size of " + }, + { + "bbox": [ + 304, + 581, + 545, + 652 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 304, + 581, + 545, + 652 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "type": "text", + "content": "Probability of correct keypoint [44] with a threshold of 0.2 (PCK@0.2) is the evaluation metric for MP-100. During testing, we generate 60,000 episodes evenly distributed across the test categories and report the average PCK@0.2 over these episodes." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "23496" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 61, + 80, + 183, + 129 + ], + "blocks": [ + { + "bbox": [ + 94, + 71, + 149, + 79 + ], + "lines": [ + { + "bbox": [ + 94, + 71, + 149, + 79 + ], + "spans": [ + { + "bbox": [ + 94, + 71, + 149, + 79 + ], + "type": "text", + "content": "super-keypoint 1" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 61, + 80, + 183, + 129 + ], + "lines": [ + { + "bbox": [ + 61, + 80, + 183, + 129 + ], + "spans": [ + { + "bbox": [ + 61, + 80, + 183, + 129 + ], + "type": "image", + "image_path": "2dcb167bf1df6d2a3f0848cbf991a77c10385557f296ffba52d16b4497e2d3cc.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 61, + 140, + 184, + 189 + ], + "blocks": [ + { + "bbox": [ + 94, + 131, + 150, + 140 + ], + "lines": [ + { + "bbox": [ + 94, + 131, + 150, + 140 + ], + "spans": [ + { + "bbox": [ + 94, + 131, + 150, + 140 + ], + "type": "text", + "content": "super-keypoint 4" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 61, + 140, + 184, + 189 + ], + "lines": [ + { + "bbox": [ + 61, + 140, + 184, + 189 + ], + "spans": [ + { + "bbox": [ + 61, + 140, + 184, + 189 + ], + "type": "image", + "image_path": "dc6d2dd56b6ec5f1f181475c2fe1c16ab7c667692324d7245460aef242154200.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 194, + 80, + 234, + 121 + ], + "blocks": [ + { + "bbox": [ + 227, + 71, + 283, + 79 + ], + "lines": [ + { + "bbox": [ + 227, + 71, + 283, + 79 + ], + "spans": [ + { + "bbox": [ + 227, + 71, + 283, + 79 + ], + "type": "text", + "content": "super-keypoint 2" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 194, + 80, + 234, + 121 + ], + "lines": [ + { + "bbox": [ + 194, + 80, + 234, + 121 + ], + "spans": [ + { + "bbox": [ + 194, + 80, + 234, + 121 + ], + "type": "image", + "image_path": "7ab2abb81057dd1bd795501fd6a87507585c83a7247f6571d56827b2d5c89b82.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 196, + 121, + 233, + 129 + ], + "lines": [ + { + "bbox": [ + 196, + 121, + 233, + 129 + ], + "spans": [ + { + "bbox": [ + 196, + 121, + 233, + 129 + ], + "type": "text", + "content": "arctic wolf" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 235, + 80, + 317, + 121 + ], + "blocks": [ + { + "bbox": [ + 235, + 80, + 317, + 121 + ], + "lines": [ + { + "bbox": [ + 235, + 80, + 317, + 121 + ], + "spans": [ + { + "bbox": [ + 235, + 80, + 317, + 121 + ], + "type": "image", + "image_path": "5becc95d4bbc53fd37ef419c7d35c1458503d22e41b4db10616afa3c288c4d7f.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 242, + 122, + 268, + 129 + ], + "lines": [ + { + "bbox": [ + 242, + 122, + 268, + 129 + ], + "spans": [ + { + "bbox": [ + 242, + 122, + 268, + 129 + ], + "type": "text", + "content": "sealion" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 286, + 122, + 306, + 129 + ], + "lines": [ + { + "bbox": [ + 286, + 122, + 306, + 129 + ], + "spans": [ + { + "bbox": [ + 286, + 122, + 306, + 129 + ], + "type": "text", + "content": "ferret" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 194, + 140, + 235, + 180 + ], + "blocks": [ + { + "bbox": [ + 194, + 140, + 235, + 180 + ], + "lines": [ + { + "bbox": [ + 194, + 140, + 235, + 180 + ], + "spans": [ + { + "bbox": [ + 194, + 140, + 235, + 180 + ], + "type": "image", + "image_path": "5821cad53eb20a3b71bf969b5d87fbd067c5a4483afdb5659f47b9b272ac65db.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 205, + 182, + 225, + 190 + ], + "lines": [ + { + "bbox": [ + 205, + 182, + 225, + 190 + ], + "spans": [ + { + "bbox": [ + 205, + 182, + 225, + 190 + ], + "type": "text", + "content": "grebe" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 236, + 140, + 276, + 180 + ], + "blocks": [ + { + "bbox": [ + 247, + 131, + 304, + 139 + ], + "lines": [ + { + "bbox": [ + 247, + 131, + 304, + 139 + ], + "spans": [ + { + "bbox": [ + 247, + 131, + 304, + 139 + ], + "type": "text", + "content": "super-keypoint 5" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 236, + 140, + 276, + 180 + ], + "lines": [ + { + "bbox": [ + 236, + 140, + 276, + 180 + ], + "spans": [ + { + "bbox": [ + 236, + 140, + 276, + 180 + ], + "type": "image", + "image_path": "f936ba5b207f2eafb9bfbd02d131c92a99c29b07ac9f24e7ca7d93197999e3e9.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 239, + 182, + 272, + 190 + ], + "lines": [ + { + "bbox": [ + 239, + 182, + 272, + 190 + ], + "spans": [ + { + "bbox": [ + 239, + 182, + 272, + 190 + ], + "type": "text", + "content": "kingfisher" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 277, + 140, + 358, + 182 + ], + "blocks": [ + { + "bbox": [ + 277, + 140, + 358, + 182 + ], + "lines": [ + { + "bbox": [ + 277, + 140, + 358, + 182 + ], + "spans": [ + { + "bbox": [ + 277, + 140, + 358, + 182 + ], + "type": "image", + "image_path": "86b51a155c2dfd9e02781812436417f8818af38079354f12e4f54474ad7c9ef9.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 286, + 182, + 304, + 190 + ], + "lines": [ + { + "bbox": [ + 286, + 182, + 304, + 190 + ], + "spans": [ + { + "bbox": [ + 286, + 182, + 304, + 190 + ], + "type": "text", + "content": "wren" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 310, + 182, + 342, + 190 + ], + "lines": [ + { + "bbox": [ + 310, + 182, + 342, + 190 + ], + "spans": [ + { + "bbox": [ + 310, + 182, + 342, + 190 + ], + "type": "text", + "content": "gull" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 328, + 80, + 531, + 129 + ], + "blocks": [ + { + "bbox": [ + 402, + 71, + 457, + 79 + ], + "lines": [ + { + "bbox": [ + 402, + 71, + 457, + 79 + ], + "spans": [ + { + "bbox": [ + 402, + 71, + 457, + 79 + ], + "type": "text", + "content": "super-keypoint 3" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 328, + 80, + 531, + 129 + ], + "lines": [ + { + "bbox": [ + 328, + 80, + 531, + 129 + ], + "spans": [ + { + "bbox": [ + 328, + 80, + 531, + 129 + ], + "type": "image", + "image_path": "745a79b6731c5e06dd435e94e3e370fba06a59ba037462611dc234c26267ccdd.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 367, + 140, + 410, + 181 + ], + "blocks": [ + { + "bbox": [ + 367, + 140, + 410, + 181 + ], + "lines": [ + { + "bbox": [ + 367, + 140, + 410, + 181 + ], + "spans": [ + { + "bbox": [ + 367, + 140, + 410, + 181 + ], + "type": "image", + "image_path": "529764b53507c5dd148173d39691175d465e94ba850e670a31e493d169546f71.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 368, + 182, + 410, + 198 + ], + "lines": [ + { + "bbox": [ + 368, + 182, + 410, + 198 + ], + "spans": [ + { + "bbox": [ + 368, + 182, + 410, + 198 + ], + "type": "text", + "content": "long sleeved dress" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 411, + 140, + 531, + 181 + ], + "blocks": [ + { + "bbox": [ + 421, + 131, + 477, + 140 + ], + "lines": [ + { + "bbox": [ + 421, + 131, + 477, + 140 + ], + "spans": [ + { + "bbox": [ + 421, + 131, + 477, + 140 + ], + "type": "text", + "content": "super-keypoint 6" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 411, + 140, + 531, + 181 + ], + "lines": [ + { + "bbox": [ + 411, + 140, + 531, + 181 + ], + "spans": [ + { + "bbox": [ + 411, + 140, + 531, + 181 + ], + "type": "image", + "image_path": "864cc0f355e8c5fca8876a7a91a61925f3d5cb322aee6c66dfdbd2e2c0198f53.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 422, + 182, + 436, + 190 + ], + "lines": [ + { + "bbox": [ + 422, + 182, + 436, + 190 + ], + "spans": [ + { + "bbox": [ + 422, + 182, + 436, + 190 + ], + "type": "text", + "content": "vest" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 452, + 182, + 531, + 198 + ], + "lines": [ + { + "bbox": [ + 452, + 182, + 531, + 198 + ], + "spans": [ + { + "bbox": [ + 452, + 182, + 531, + 198 + ], + "type": "text", + "content": "vest dress long sleeved shirt" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 52, + 224, + 141, + 330 + ], + "blocks": [ + { + "bbox": [ + 47, + 207, + 545, + 219 + ], + "lines": [ + { + "bbox": [ + 47, + 207, + 545, + 219 + ], + "spans": [ + { + "bbox": [ + 47, + 207, + 545, + 219 + ], + "type": "text", + "content": "Figure 5. Keypoints of four different super-keypoints discovered from the base categories. Below each keypoint is the name of its category." + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 52, + 224, + 141, + 330 + ], + "lines": [ + { + "bbox": [ + 52, + 224, + 141, + 330 + ], + "spans": [ + { + "bbox": [ + 52, + 224, + 141, + 330 + ], + "type": "image", + "image_path": "6959ff7436ec326b22dd0585880a33e7240f3a16460d31ddf18e3e5e0dc94065.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_body" + } + ], + "index": 27 + }, + { + "type": "image", + "bbox": [ + 145, + 224, + 541, + 338 + ], + "blocks": [ + { + "bbox": [ + 145, + 224, + 541, + 338 + ], + "lines": [ + { + "bbox": [ + 145, + 224, + 541, + 338 + ], + "spans": [ + { + "bbox": [ + 145, + 224, + 541, + 338 + ], + "type": "image", + "image_path": "d2ef5f652fa5ababd2292e8a42bea4f21d8b7ae7fae169876860d639a7579ddf.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + } + ], + "index": 28 + }, + { + "type": "image", + "bbox": [ + 52, + 350, + 142, + 459 + ], + "blocks": [ + { + "bbox": [ + 52, + 350, + 142, + 459 + ], + "lines": [ + { + "bbox": [ + 52, + 350, + 142, + 459 + ], + "spans": [ + { + "bbox": [ + 52, + 350, + 142, + 459 + ], + "type": "image", + "image_path": "1ab32c0f7de342004916f7973eb8bc0093513bf88ac4d5cfbe4f91092111ae67.jpg" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_body" + } + ], + "index": 29 + }, + { + "type": "image", + "bbox": [ + 145, + 347, + 541, + 460 + ], + "blocks": [ + { + "bbox": [ + 145, + 347, + 541, + 460 + ], + "lines": [ + { + "bbox": [ + 145, + 347, + 541, + 460 + ], + "spans": [ + { + "bbox": [ + 145, + 347, + 541, + 460 + ], + "type": "image", + "image_path": "4855aa92067acae24b11129676c42da4b62c1843c6aebf796e3bb39043d4ddc9.jpg" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 469, + 544, + 492 + ], + "lines": [ + { + "bbox": [ + 46, + 469, + 544, + 492 + ], + "spans": [ + { + "bbox": [ + 46, + 469, + 544, + 492 + ], + "type": "text", + "content": "Figure 6. Qualitative comparison between simple matching and learnable matching network in aligning novel keypoints with superkeypoints. We show one base keypoint member of each matched super-keypoint." + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_caption" + } + ], + "index": 30 + }, + { + "bbox": [ + 47, + 512, + 235, + 525 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 512, + 235, + 525 + ], + "spans": [ + { + "bbox": [ + 47, + 512, + 235, + 525 + ], + "type": "text", + "content": "6.2. Comparison with previous methods" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 46, + 530, + 287, + 625 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 530, + 287, + 625 + ], + "spans": [ + { + "bbox": [ + 46, + 530, + 287, + 625 + ], + "type": "text", + "content": "We compare ESCAPE against previous methods, including Prototypical Networks [32], MAML [10], Finetune [22], POMNet [32] and CapeFormer [31]. Note that we remove the support keypoint identifier from CapeFormer. The results are shown in Tab. 1. We can see that ESCAPE outperforms other methods by around " + }, + { + "bbox": [ + 46, + 530, + 287, + 625 + ], + "type": "inline_equation", + "content": "2.5\\%" + }, + { + "bbox": [ + 46, + 530, + 287, + 625 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 530, + 287, + 625 + ], + "type": "inline_equation", + "content": "0.5\\%" + }, + { + "bbox": [ + 46, + 530, + 287, + 625 + ], + "type": "text", + "content": " in 1-shot and 5-shot settings, respectively, establishing a new state-of-the-art on the MP-100 benchmark." + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 47, + 635, + 138, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 635, + 138, + 647 + ], + "spans": [ + { + "bbox": [ + 47, + 635, + 138, + 647 + ], + "type": "text", + "content": "6.3. Ablation study" + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 46, + 653, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 287, + 713 + ], + "type": "text", + "content": "Super-keypoints discovering. We perform clustering on the keypoints of base categories to form super-keypoints. To verify that we can obtain meaningful super-keypoints from the clustering process, we show some keypoints members of a few super-keypoints in Fig. 5. We see that each super-" + } + ] + } + ], + "index": 35 + }, + { + "bbox": [ + 304, + 513, + 545, + 573 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 513, + 545, + 573 + ], + "spans": [ + { + "bbox": [ + 304, + 513, + 545, + 573 + ], + "type": "text", + "content": "keypoint represents a meaningful concept. For example, super-keypoint 1 consists of 'left knee' of deer, antelope and zebra. super-keypoint 4 consists of 'right rear wheel' of SUV, bus, and car, while super-keypoint 6 is a landmark on the collars of different clothes categories." + } + ] + } + ], + "index": 36 + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "text", + "content": "Novel keypoints and super-keypoints matching. We qualitatively compare the performance of the simple matching and the learnable matching network in aligning novel keypoints with the learned super-keypoints. Fig. 6 shows the matching results for two novel categories, namely squirrel and woodpecker. The left side of the figure shows support images of the categories, while the right side shows the super-keypoints aligned by the simple matching and the learnable matching network. For each super-keypoint, we only show one base keypoints member. Compared with the" + } + ] + } + ], + "index": 37 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "23497" + } + ] + } + ], + "index": 38 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 48, + 72, + 149, + 124 + ], + "blocks": [ + { + "bbox": [ + 48, + 72, + 149, + 124 + ], + "lines": [ + { + "bbox": [ + 48, + 72, + 149, + 124 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 149, + 124 + ], + "type": "table", + "html": "
MethodPCK
MLE83.00 ± 0.16
MAP83.88 ± 0.15
ESCAPE86.89 ± 0.14
", + "image_path": "aa60340e1a78f9c67d8affcdc0acba12cc20edca8eae4e52c31994de4afb1dd8.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 159, + 70, + 288, + 125 + ], + "blocks": [ + { + "bbox": [ + 47, + 132, + 150, + 175 + ], + "lines": [ + { + "bbox": [ + 47, + 132, + 150, + 175 + ], + "spans": [ + { + "bbox": [ + 47, + 132, + 150, + 175 + ], + "type": "text", + "content": "Table 2. Results of different prototype estimation methods, namely MLE, MAP and ESCAPE." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 159, + 70, + 288, + 125 + ], + "lines": [ + { + "bbox": [ + 159, + 70, + 288, + 125 + ], + "spans": [ + { + "bbox": [ + 159, + 70, + 288, + 125 + ], + "type": "table", + "html": "
Feature extractorMethodPCK
DIFT [36]MLE60.61
DIFT [36]ESCAPE62.47
DINOv2 [27]MLE88.31
DINOv2 [27]ESCAPE89.14
", + "image_path": "fdbe59bb7d561149dda8a5cb4dc0653b91e9b35ba954a7d249348f359ca5e96a.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 157, + 134, + 287, + 177 + ], + "lines": [ + { + "bbox": [ + 157, + 134, + 287, + 177 + ], + "spans": [ + { + "bbox": [ + 157, + 134, + 287, + 177 + ], + "type": "text", + "content": "Table 3. Results of integrating ESCAPE to feature extractors of different semantic correspondence models" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 205, + 287, + 301 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 205, + 287, + 301 + ], + "spans": [ + { + "bbox": [ + 46, + 205, + 287, + 301 + ], + "type": "text", + "content": "simple matching, our learnable matching network provides more reliable super-keypoints for the novel keypoints. In particular, the simple matching aligns keypoints of squirrel with super-keypoints of keypoints of kingfisher, sling and suv, which are irrelevant to the squirrel. On the other hand, the learnable matching network aligns the novel keypoints with their actual concepts, e.g., the 'left elbow of squirrel' is matched with the super-keypoint of the 'left elbow of rat'." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 334, + 287, + 538 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 334, + 287, + 538 + ], + "spans": [ + { + "bbox": [ + 46, + 334, + 287, + 538 + ], + "type": "text", + "content": "Ablation on novel category adaptation. In Tab. 2, we study the performance of different methods to estimate keypoint prototypes for a novel category, namely MLE, MAP and ESCAPE. MLE is our baseline, which estimate keypoint prototypes as means of support features as in Eq. 2. MAP and ESCAPE compute keypoint prototypes as maximum a posterior estimate with the EM algorithm. The difference between MAP and ESCAPE is that MAP (Eq. 7) uses the simple matching in the expectation step, whereas ESCAPE (Eq. 10) leverages the learnable matching network. We observe that incorporating super-keypoints into the estimation of keypoint prototypes helps improve the overall performance. Specifically, the MAP outperforms MLE by roughly " + }, + { + "bbox": [ + 46, + 334, + 287, + 538 + ], + "type": "inline_equation", + "content": "1\\%" + }, + { + "bbox": [ + 46, + 334, + 287, + 538 + ], + "type": "text", + "content": ". Regardless ESCAPE, better novel keypoints and super-keypoints alignments from the learnable matching network immediately translate into " + }, + { + "bbox": [ + 46, + 334, + 287, + 538 + ], + "type": "inline_equation", + "content": "3\\%" + }, + { + "bbox": [ + 46, + 334, + 287, + 538 + ], + "type": "text", + "content": " improvements over the MAP counterpart." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 570, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 570, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 570, + 286, + 713 + ], + "type": "text", + "content": "Ablation on different feature extractors. Feature extractors from semantic correspondence models can effectively solve CAPE since they can generate dense features for support and query images. In Tab. 3, we study the compatibility of ESCAPE with feature extractors from different semantic correspondence models, namely DIFT [36] and DINoV2 [27]. We take the pretrained stable diffusion 2.1 [28] as the backbone for DIFT, whereas we use ViT-ViT-S/14 for DINoV2. As seen, ESCAPE consistently boosts the performance of the considered semantic correspondence models. This shows that ESCAPE can benefit from future research for better representation." + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 309, + 70, + 544, + 255 + ], + "blocks": [ + { + "bbox": [ + 309, + 70, + 544, + 255 + ], + "lines": [ + { + "bbox": [ + 309, + 70, + 544, + 255 + ], + "spans": [ + { + "bbox": [ + 309, + 70, + 544, + 255 + ], + "type": "image", + "image_path": "cf502b8569da0778b0813a397dc07a58cc3c1dc4da2c3c3847660a35c76a943e.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 306, + 264, + 545, + 287 + ], + "lines": [ + { + "bbox": [ + 306, + 264, + 545, + 287 + ], + "spans": [ + { + "bbox": [ + 306, + 264, + 545, + 287 + ], + "type": "text", + "content": "Figure 7. We qualitatively compare pose estimation of Cape-Former and ESCAPE under 1-shot setting." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 311, + 414, + 323 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 311, + 414, + 323 + ], + "spans": [ + { + "bbox": [ + 306, + 311, + 414, + 323 + ], + "type": "text", + "content": "6.4. Qualitative results" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 331, + 545, + 451 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 331, + 545, + 451 + ], + "spans": [ + { + "bbox": [ + 304, + 331, + 545, + 451 + ], + "type": "text", + "content": "Figure 7 compares the pose estimation performance of CapeFormer and ESCAPE. Overall, ESCAPE gives more accurate pose predictions compared to CapeFormer. In particular, CapeFormer is overfitting to the support image of the sheep category. The predicted positions of the left front and back paws by CapeFormer are geometrically close to each other as they appear in the support image. In contrast, since ESCAPE allows learned knowledge about base animals to be transferred to the novel keypoints, it recovers the pose of the query sheep beautifully." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 466, + 378, + 479 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 466, + 378, + 479 + ], + "spans": [ + { + "bbox": [ + 306, + 466, + 378, + 479 + ], + "type": "text", + "content": "7. Conclusion" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 488, + 545, + 632 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 488, + 545, + 632 + ], + "spans": [ + { + "bbox": [ + 304, + 488, + 545, + 632 + ], + "type": "text", + "content": "In this paper, we address the task of category-agnostic pose estimation by introducing ESCAPE, a Bayesian framework that induces a prior over the keypoint features of different categories. This prior is a mixture of super-keypoints, with each super-keypoint encoding a distribution of a group of related keypoints. ESCAPE enables the transfer of knowledge from base categories to the novel ones, thereby preventing the query predictions from overfitting to sparse support samples. ESCAPE offers several advantages over previous works: it exhibits less overfitting to support samples, does not rely on support keypoint identifiers, and is model-agnostic." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 647, + 403, + 661 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 647, + 403, + 661 + ], + "spans": [ + { + "bbox": [ + 306, + 647, + 403, + 661 + ], + "type": "text", + "content": "Acknowledgement" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 669, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 669, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 669, + 545, + 712 + ], + "type": "text", + "content": "This research is supported in part by the National Research Foundation, Singapore under its Competitive Research Program Award NRF-CRP23-2019-0003." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "23498" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 156 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 156 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 156 + ], + "type": "text", + "content": "[1] Yuanhao Cai, Zhicheng Wang, Zhengxiong Luo, Binyi Yin, Angang Du, Haoqian Wang, Xiangyu Zhang, Xinyu Zhou, Erjin Zhou, and Jian Sun. Learning delicate local representations for multi-person pose estimation. In European Conference on Computer Vision, pages 455-472. Springer, 2020. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 158, + 288, + 201 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 158, + 288, + 201 + ], + "spans": [ + { + "bbox": [ + 53, + 158, + 288, + 201 + ], + "type": "text", + "content": "[2] Zhe Cao, Tomas Simon, Shih-En Wei, and Yaser Sheikh. Realtime multi-person 2d pose estimation using part affinity fields. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7291-7299, 2017. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 202, + 288, + 255 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 202, + 288, + 255 + ], + "spans": [ + { + "bbox": [ + 53, + 202, + 288, + 255 + ], + "type": "text", + "content": "[3] Joao Carreira, Pulkit Agrawal, Katerina Fragkiadaki, and Jitendra Malik. Human pose estimation with iterative error feedback. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4733-4742, 2016. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 258, + 288, + 312 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 258, + 288, + 312 + ], + "spans": [ + { + "bbox": [ + 53, + 258, + 288, + 312 + ], + "type": "text", + "content": "[4] Yilun Chen, Zhicheng Wang, Yuxiang Peng, Zhiqiang Zhang, Gang Yu, and Jian Sun. Cascaded pyramid network for multi-person pose estimation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7103-7112, 2018. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 314, + 287, + 378 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 314, + 287, + 378 + ], + "spans": [ + { + "bbox": [ + 53, + 314, + 287, + 378 + ], + "type": "text", + "content": "[5] Bowen Cheng, Bin Xiao, Jingdong Wang, Honghui Shi, Thomas S Huang, and Lei Zhang. Higherhrnet: Scale-aware representation learning for bottom-up human pose estimation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5386-5395, 2020. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 380, + 287, + 422 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 380, + 287, + 422 + ], + "spans": [ + { + "bbox": [ + 53, + 380, + 287, + 422 + ], + "type": "text", + "content": "[6] Christopher B Choy, JunYoung Gwak, Silvio Savarese, and Manmohan Chandraker. Universal correspondence network. Advances in neural information processing systems, 29, 2016. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 424, + 287, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 424, + 287, + 456 + ], + "spans": [ + { + "bbox": [ + 53, + 424, + 287, + 456 + ], + "type": "text", + "content": "[7] Marco Cuturi. Sinkhorn distances: Lightspeed computation of optimal transport. Advances in neural information processing systems, 26, 2013. 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 458, + 287, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 458, + 287, + 502 + ], + "spans": [ + { + "bbox": [ + 53, + 458, + 287, + 502 + ], + "type": "text", + "content": "[8] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 502, + 287, + 545 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 502, + 287, + 545 + ], + "spans": [ + { + "bbox": [ + 53, + 502, + 287, + 545 + ], + "type": "text", + "content": "[9] Mingjing Du, Shifei Ding, and Hongjie Jia. Study on density peaks clustering based on k-nearest neighbors and principal component analysis. Knowledge-Based Systems, 99: 135-145, 2016. 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 547, + 287, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 547, + 287, + 590 + ], + "spans": [ + { + "bbox": [ + 48, + 547, + 287, + 590 + ], + "type": "text", + "content": "[10] Chelsea Finn, Pieter Abbeel, and Sergey Levine. Model-agnostic meta-learning for fast adaptation of deep networks. In International conference on machine learning, pages 1126–1135. PMLR, 2017. 6, 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 591, + 287, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 591, + 287, + 635 + ], + "spans": [ + { + "bbox": [ + 48, + 591, + 287, + 635 + ], + "type": "text", + "content": "[11] Jacob M Graving, Daniel Chae, Hemal Naik, Liang Li, Benjamin Koger, Blair R Costelloe, and Iain D Couzin. Deepposekit, a software toolkit for fast and robust animal pose estimation using deep learning. *Elife*, 8:e47994, 2019. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 636, + 287, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 636, + 287, + 679 + ], + "spans": [ + { + "bbox": [ + 48, + 636, + 287, + 679 + ], + "type": "text", + "content": "[12] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 681, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 681, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 681, + 287, + 713 + ], + "type": "text", + "content": "[13] Or Hirschorn and Shai Avidan. Pose anything: A graph-based approach for category-agnostic pose estimation. arXiv preprint arXiv:2311.17891, 2023. 3" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "text", + "content": "[14] Timothy Hospedales, Antreas Antoniou, Paul Micaelli, and Amos Storkey. Meta-learning in neural networks: A survey. IEEE transactions on pattern analysis and machine intelligence, 44(9):5149-5169, 2021. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 118, + 545, + 148 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 545, + 148 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 545, + 148 + ], + "type": "text", + "content": "[15] Harold W Kuhn. The hungarian method for the assignment problem. Naval research logistics quarterly, 2(1-2):83-97, 1955. 4" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 150, + 545, + 204 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 150, + 545, + 204 + ], + "spans": [ + { + "bbox": [ + 308, + 150, + 545, + 204 + ], + "type": "text", + "content": "[16] Duong Le, Khoi Duc Nguyen, Khoi Nguyen, Quoc-Huy Tran, Rang Nguyen, and Binh-Son Hua. Poodle: Improving few-shot learning via penalizing out-of-distribution samples. Advances in Neural Information Processing Systems, 34:23942-23955, 2021. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 206, + 545, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 206, + 545, + 239 + ], + "spans": [ + { + "bbox": [ + 308, + 206, + 545, + 239 + ], + "type": "text", + "content": "[17] Shuyuan Li, Jianguo Li, Hanlin Tang, Rui Qian, and Weiyao Lin. Atrw: a benchmark for amur tiger re-identification in the wild. arXiv preprint arXiv:1906.05586, 2019. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 239, + 545, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 239, + 545, + 282 + ], + "spans": [ + { + "bbox": [ + 308, + 239, + 545, + 282 + ], + "type": "text", + "content": "[18] Wenbo Li, Zhicheng Wang, Binyi Yin, Qixiang Peng, Yuming Du, Tianzi Xiao, Gang Yu, Hongtao Lu, Yichen Wei, and Jian Sun. Rethinking on multi-stage networks for human pose estimation. arXiv preprint arXiv:1901.00148, 2019. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 284, + 545, + 338 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 284, + 545, + 338 + ], + "spans": [ + { + "bbox": [ + 308, + 284, + 545, + 338 + ], + "type": "text", + "content": "[19] Jiang-Jiang Liu, Qibin Hou, Ming-Ming Cheng, Changhu Wang, and Jiashi Feng. Improving convolutional networks with self-calibrated convolutions. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10096-10105, 2020. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 338, + 545, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 338, + 545, + 403 + ], + "spans": [ + { + "bbox": [ + 308, + 338, + 545, + 403 + ], + "type": "text", + "content": "[20] Alexander Mathis, Thomas Biasi, Steffen Schneider, Mert Yuksekgonul, Byron Rogers, Matthias Bethge, and Mackenzie W Mathis. Pretraining boosts out-of-domain robustness for pose estimation. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1859-1868, 2021. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 404, + 545, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 404, + 545, + 426 + ], + "spans": [ + { + "bbox": [ + 308, + 404, + 545, + 426 + ], + "type": "text", + "content": "[21] Kevin P. Murphy. Probabilistic Machine Learning: Advanced Topics. MIT Press, 2023. 4" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 427, + 545, + 458 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 427, + 545, + 458 + ], + "spans": [ + { + "bbox": [ + 308, + 427, + 545, + 458 + ], + "type": "text", + "content": "[22] Akihiro Nakamura and Tatsuya Harada. Revisiting fine-tuning for few-shot learning. arXiv preprint arXiv:1910.00216, 2019. 6, 7" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 460, + 545, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 460, + 545, + 514 + ], + "spans": [ + { + "bbox": [ + 308, + 460, + 545, + 514 + ], + "type": "text", + "content": "[23] Alejandro Newell, Kaiyu Yang, and Jia Deng. Stacked hourglass networks for human pose estimation. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part VIII 14, pages 483-499. Springer, 2016. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 515, + 545, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 515, + 545, + 557 + ], + "spans": [ + { + "bbox": [ + 308, + 515, + 545, + 557 + ], + "type": "text", + "content": "[24] Alejandro Newell, Zhiao Huang, and Jia Deng. Associative embedding: End-to-end learning for joint detection and grouping. Advances in neural information processing systems, 30, 2017. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 559, + 545, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 559, + 545, + 613 + ], + "spans": [ + { + "bbox": [ + 308, + 559, + 545, + 613 + ], + "type": "text", + "content": "[25] Khoi D Nguyen, Quoc-Huy Tran, Khoi Nguyen, Binh-Son Hua, and Rang Nguyen. Inductive and transductive few-shot video classification via appearance and temporal alignments. In European Conference on Computer Vision, pages 471-487. Springer, 2022. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 615, + 545, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 615, + 545, + 658 + ], + "spans": [ + { + "bbox": [ + 308, + 615, + 545, + 658 + ], + "type": "text", + "content": "[26] Xuecheng Nie, Jiashi Feng, Jianfeng Zhang, and Shuicheng Yan. Single-stage multi-person pose machines. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6951-6960, 2019. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 658, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 658, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 658, + 545, + 713 + ], + "type": "text", + "content": "[27] Maxime Oquab, Timothée Darcet, Théo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023. 2, 8" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "23499" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 288, + 127 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 288, + 127 + ], + "type": "text", + "content": "[28] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 129, + 288, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 129, + 288, + 185 + ], + "spans": [ + { + "bbox": [ + 48, + 129, + 288, + 185 + ], + "type": "text", + "content": "[29] Ruslan Salakhutdinov, Joshua Tenenbaum, and Antonio Torralba. One-shot learning with a hierarchical nonparametric bayesian model. In Proceedings of ICML Workshop on Unsupervised and Transfer Learning, pages 195-206. JMLR Workshop and Conference Proceedings, 2012. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 185, + 288, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 185, + 288, + 239 + ], + "spans": [ + { + "bbox": [ + 48, + 185, + 288, + 239 + ], + "type": "text", + "content": "[30] Paul-Edouard Sarlin, Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. Superglue: Learning feature matching with graph neural networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4938–4947, 2020. 5, 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 241, + 288, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 241, + 288, + 297 + ], + "spans": [ + { + "bbox": [ + 48, + 241, + 288, + 297 + ], + "type": "text", + "content": "[31] Min Shi, Zihao Huang, Xianzheng Ma, Xiaowei Hu, and Zhiguo Cao. Matching is not enough: A two-stage framework for category-agnostic pose estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7308-7317, 2023. 1, 3, 6, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 297, + 288, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 297, + 288, + 330 + ], + "spans": [ + { + "bbox": [ + 48, + 297, + 288, + 330 + ], + "type": "text", + "content": "[32] Jake Snell, Kevin Swersky, and Richard Zemel. Prototypical networks for few-shot learning. Advances in neural information processing systems, 30, 2017. 6, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 331, + 288, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 331, + 288, + 385 + ], + "spans": [ + { + "bbox": [ + 48, + 331, + 288, + 385 + ], + "type": "text", + "content": "[33] Ke Sun, Bin Xiao, Dong Liu, and Jingdong Wang. Deep high-resolution representation learning for human pose estimation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5693-5703, 2019. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 387, + 288, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 387, + 288, + 431 + ], + "spans": [ + { + "bbox": [ + 48, + 387, + 288, + 431 + ], + "type": "text", + "content": "[34] Xiao Sun, Jiaxiang Shang, Shuang Liang, and Yichen Wei. Compositional human pose regression. In Proceedings of the IEEE International Conference on Computer Vision, pages 2602-2611, 2017. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 433, + 288, + 476 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 433, + 288, + 476 + ], + "spans": [ + { + "bbox": [ + 48, + 433, + 288, + 476 + ], + "type": "text", + "content": "[35] Xiao Sun, Bin Xiao, Fangyin Wei, Shuang Liang, and Yichen Wei. Integral human pose regression. In Proceedings of the European conference on computer vision (ECCV), pages 529-545, 2018. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 478, + 288, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 478, + 288, + 521 + ], + "spans": [ + { + "bbox": [ + 48, + 478, + 288, + 521 + ], + "type": "text", + "content": "[36] Luming Tang, Menglin Jia, Qianqian Wang, Cheng Perng Phoo, and Bharath Hariharan. Emergent correspondence from image diffusion. Advances in Neural Information Processing Systems, 36:1363-1389, 2023. 2, 8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 523, + 288, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 523, + 288, + 555 + ], + "spans": [ + { + "bbox": [ + 48, + 523, + 288, + 555 + ], + "type": "text", + "content": "[37] Zhi Tian, Hao Chen, and Chunhua Shen. Direct pose: Direct end-to-end multi-person pose estimation. arXiv preprint arXiv:1911.07451, 2019. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 557, + 288, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 557, + 288, + 601 + ], + "spans": [ + { + "bbox": [ + 48, + 557, + 288, + 601 + ], + "type": "text", + "content": "[38] Jonathan J. Tompson, Arjun Jain, Yann LeCun, and Christoph Bregler. Joint training of a convolutional network and a graphical model for human pose estimation. Advances in neural information processing systems, 27, 2014. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 602, + 288, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 602, + 288, + 646 + ], + "spans": [ + { + "bbox": [ + 48, + 602, + 288, + 646 + ], + "type": "text", + "content": "[39] Alexander Toshev and Christian Szegedy. Deeppose: Human pose estimation via deep neural networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1653-1660, 2014. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 647, + 288, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 647, + 288, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 647, + 288, + 712 + ], + "type": "text", + "content": "[40] Jingdong Wang, Ke Sun, Tianheng Cheng, Borui Jiang, Chaorui Deng, Yang Zhao, Dong Liu, Yadong Mu, Mingkui Tan, Xinggang Wang, et al. Deep high-resolution representation learning for visual recognition. IEEE transactions on pattern analysis and machine intelligence, 43(10):3349-3364, 2020. 2" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 72, + 547, + 474 + ], + "type": "list", + "angle": 0, + "index": 22, + "blocks": [ + { + "bbox": [ + 307, + 72, + 547, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 72, + 547, + 116 + ], + "spans": [ + { + "bbox": [ + 307, + 72, + 547, + 116 + ], + "type": "text", + "content": "[41] Bin Xiao, Haiping Wu, and Yichen Wei. Simple baselines for human pose estimation and tracking. In Proceedings of the European conference on computer vision (ECCV), pages 466-481, 2018. 2, 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 118, + 547, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 547, + 184 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 547, + 184 + ], + "type": "text", + "content": "[42] Lumin Xu, Sheng Jin, Wang Zeng, Wentao Liu, Chen Qian, Wanli Ouyang, Ping Luo, and Xiaogang Wang. Pose for everything: Towards category-agnostic pose estimation. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23–27, 2022, Proceedings, Part VI, pages 398–416. Springer, 2022. 1, 2, 3, 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 185, + 547, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 185, + 547, + 217 + ], + "spans": [ + { + "bbox": [ + 307, + 185, + 547, + 217 + ], + "type": "text", + "content": "[43] Shuo Yang, Lu Liu, and Min Xu. Free lunch for few-shot learning: Distribution calibration. arXiv preprint arXiv:2101.06395, 2021. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 219, + 547, + 261 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 219, + 547, + 261 + ], + "spans": [ + { + "bbox": [ + 307, + 219, + 547, + 261 + ], + "type": "text", + "content": "[44] Yi Yang and Deva Ramanan. Articulated human detection with flexible mixtures of parts. IEEE transactions on pattern analysis and machine intelligence, 35(12):2878-2890, 2012. 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 262, + 547, + 306 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 262, + 547, + 306 + ], + "spans": [ + { + "bbox": [ + 307, + 262, + 547, + 306 + ], + "type": "text", + "content": "[45] Wang Yaqing, Yao Quanming, T Kwok James, and M Ni Lionel. Generalizing from a few examples: A survey on few-shot learning. ACM computing surveys, 53(3):1-34, 2020. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 308, + 547, + 351 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 308, + 547, + 351 + ], + "spans": [ + { + "bbox": [ + 307, + 308, + 547, + 351 + ], + "type": "text", + "content": "[46] Hang Yu, Yufei Xu, Jing Zhang, Wei Zhao, Ziyu Guan, and Dacheng Tao. Ap-10k: A benchmark for animal pose estimation in the wild. arXiv preprint arXiv:2108.12617, 2021. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 353, + 547, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 353, + 547, + 418 + ], + "spans": [ + { + "bbox": [ + 307, + 353, + 547, + 418 + ], + "type": "text", + "content": "[47] Wang Zeng, Sheng Jin, Wentao Liu, Chen Qian, Ping Luo, Wanli Ouyang, and Xiaogang Wang. Not all tokens are equal: Human-centric visual analysis via token clustering transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11101-11111, 2022. 4" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 420, + 547, + 474 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 420, + 547, + 474 + ], + "spans": [ + { + "bbox": [ + 307, + 420, + 547, + 474 + ], + "type": "text", + "content": "[48] Xueting Zhang, Debin Meng, Henry Gouk, and Timothy M Hospedales. Shallow bayesian meta learning for real-world few-shot recognition. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 651-660, 2021. 2" + } + ] + } + ], + "index": 21 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "23500" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/ESR-NeRF_ Emissive Source Reconstruction Using LDR Multi-view Images/2c6d481f-7a23-410c-a62a-93255bbc656e_content_list.json b/2024/ESR-NeRF_ Emissive Source Reconstruction Using LDR Multi-view Images/2c6d481f-7a23-410c-a62a-93255bbc656e_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..d5ef664ce329fa4248b293bbca5a7b84c1be830a --- /dev/null +++ b/2024/ESR-NeRF_ Emissive Source Reconstruction Using LDR Multi-view Images/2c6d481f-7a23-410c-a62a-93255bbc656e_content_list.json @@ -0,0 +1,2046 @@ +[ + { + "type": "text", + "text": "ESR-NeRF: Emissive Source Reconstruction Using LDR Multi-view Images", + "text_level": 1, + "bbox": [ + 102, + 130, + 866, + 151 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Jinseo Jeong $^{1}$ Junseo Koo $^{1}$ Qimeng Zhang $^{2}$ Gunhee Kim $^{1}$ Seoul National University Korea University", + "bbox": [ + 228, + 179, + 746, + 217 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "jinseo@vision.snu.ac.kr, junseo.koo@vision.snu.ac.kr, zoe1024@korea.ac.kr, gunhee@snu.ac.kr \nhttps://jinseo.kr/ESR-NeRF", + "bbox": [ + 86, + 219, + 892, + 251 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 286, + 313, + 301 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Existing NeRF-based inverse rendering methods suppose that scenes are exclusively illuminated by distant light sources, neglecting the potential influence of emissive sources within a scene. In this work, we confront this limitation using LDR multi-view images captured with emissive sources turned on and off. Two key issues must be addressed: 1) ambiguity arising from the limited dynamic range along with unknown lighting details, and 2) the expensive computational cost in volume rendering to backtrace the paths leading to final object colors. We present a novel approach, ESR-NeRF, leveraging neural networks as learnable functions to represent ray-traced fields. By training networks to satisfy light transport segments, we regulate outgoing radiances, progressively identifying emissive sources while being aware of reflection areas. The results on scenes encompassing emissive sources with various properties demonstrate the superiority of ESR-NeRF in qualitative and quantitative ways. Our approach also extends its applicability to the scenes devoid of emissive sources, achieving lower CD metrics on the DTU dataset.", + "bbox": [ + 75, + 318, + 473, + 621 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 647, + 209, + 662 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Extensive research has focused on reconstructing 3D object structures [16, 43, 47, 86], material properties [18, 29, 65], and lighting [15, 33, 34, 74, 79] from 2D images, applicable across domains including 3D graphics and augmented reality [62, 63, 69, 72]. This endeavor not only facilitates the creation of life-like virtual objects but also streamlines the process of scene manipulation [27, 58, 61, 73]. Recent advancements [24, 30, 36, 71] have built on Neural Radiance Fields (NeRF) [40] successes in novel view synthesis [3, 4, 45, 81, 91]. Significant progress in relighting [37, 38, 50] has facilitated scene editing via manipulating the reconstructed light sources. However, existing methods predominantly deal with the scenes lit by distant sources, like environment maps or collocated flashlights. Notably, NeRF-based inverse rendering has yet to", + "bbox": [ + 75, + 674, + 468, + 901 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/be789cdd189d4bd05cb65e9e4d54a65edb1901096f20ac60382ca27daa97eb57.jpg", + "image_caption": [ + "255 pixel values" + ], + "image_footnote": [], + "bbox": [ + 501, + 284, + 573, + 357 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/80db8dedaaea4ce3ac6829a80e9c359cab6dd83422264462eff09725a33280fb.jpg", + "image_caption": [ + "Information loss" + ], + "image_footnote": [], + "bbox": [ + 575, + 285, + 645, + 357 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/7843b6cfe8d393de564ff362e62133cbda7707f238fa4ecf3cfad14bd5dc81cc.jpg", + "image_caption": [ + "Threshold 0.65" + ], + "image_footnote": [], + "bbox": [ + 501, + 375, + 571, + 443 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/91a2824c953bde7a9f164294590355335cd4738daa64b0b63b095bbd1f12496f.jpg", + "image_caption": [ + "Threshold 0.75" + ], + "image_footnote": [], + "bbox": [ + 575, + 375, + 645, + 443 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/cb90e4c96d41e556988afaa3833d0fcca7b226e1c2055d7f3b263ab415c66fe8.jpg", + "image_caption": [ + "" + ], + "image_footnote": [], + "bbox": [ + 645, + 285, + 674, + 357 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/738b4216d6118c90695254c7dfc9ec18e1c04b9e7fa8ab91a620794a68fb5fc8.jpg", + "image_caption": [ + "Edit emission" + ], + "image_footnote": [], + "bbox": [ + 669, + 285, + 810, + 455 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/aa37142ac7b770d8a86342081c99d6da7914dbe9cc2bbeeed215360a535f0e20.jpg", + "image_caption": [ + "Desired" + ], + "image_footnote": [], + "bbox": [ + 818, + 299, + 887, + 364 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/c1f1c8ea760cb6b452e4444e5270a3c25bd6d3ebc5a978c6ad15c00985bc2a91.jpg", + "image_caption": [ + "Unrealistic", + "Figure 1. Challenges posed by emissive sources in LDR images. Green, red, and blue in thresholded images respectively show true positives, false negatives, and false positives of source identification. Thresholding values are scaled down divided by 255. The contrast between light on and off pixel values is more pronounced in surroundings than emissive sources. Inaccurate reconstruction of emissive sources disrupts scene editing, causing reflection areas to stay static while only the source colors change." + ], + "image_footnote": [], + "bbox": [ + 818, + 377, + 887, + 440 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "consider scenes with multiple emissive sources, a common real-world illumination condition.", + "bbox": [ + 496, + 593, + 890, + 622 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Emissive sources in a scene introduce critical challenges: (i) ambiguity in decomposing scene components and (ii) high computational costs for analyzing the causes of pixel colors. This ambiguity stems from difficulties in identifying emissive source regions, as illustrated in Fig. 1. Contrary to prior setups [6-8, 66, 85, 93], we allow the possibility of numerous emissive sources throughout the scene. In standard photographs with pixel values from 0 to 255, the distinction between emissive sources and nearby reflection areas is challenging. As shown in Fig. 1, relying solely on pixel value thresholding is insufficient for differentiating between emissive sources and their reflections. Naive inverse path tracing is impractical, due to the computational costs rising exponentially with the number of ray bounces in volume rendering. This can cause inaccuracy in emissive source reconstruction, yielding unrealistic illumination in reflective areas as users manipulate emissive sources.", + "bbox": [ + 496, + 626, + 892, + 881 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To address these challenges, we introduce ESR-NeRF", + "bbox": [ + 519, + 885, + 890, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 807, + 44 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "4598", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "(Emissive Sources Reconstructing NeRF), a novel approach capable of reconstructing any number of emissive sources by progressively discovering reflection areas. We assume that the scenes are observed in two lighting conditions: one with all emissive sources active and the other with them inactive. Our approach utilizes neural networks as learnable functions for representing ray-traced fields. By training networks to satisfy each light transport segment, we sidestep the computational overhead of ray tracing associated with ray bounces. In this work, we exclusively use low dynamic range (LDR) images, setting us apart from prior mesh-based methods that rely on high dynamic range (HDR) images [2, 19, 48, 76].", + "bbox": [ + 75, + 90, + 472, + 287 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our experiments encompass synthetic and real scenes, ranging from single to multiple lighting configurations with complex reflections. The scenes vary in light source counts, color, and intensity. Qualitative and quantitative evaluations show ESR-NeRF's superiority over state-of-the-art NeRF-based re-lighting methods. Furthermore, Chamfer Distance (CD) metrics on the DTU dataset [23] indicate ESR-NeRF's competitive performance in scene reconstruction, even without emissive sources.", + "bbox": [ + 75, + 287, + 472, + 421 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We summarize our contributions as follows.", + "bbox": [ + 96, + 422, + 387, + 436 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. Our work presents the first NeRF-based inverse rendering that can deal with the scenes with any number of emissive sources, challenging the distant light assumption of previous research.", + "2. Unlike existing mesh-based methods relying on HDR images, we use LDR images for the first time, overcoming the poor representation of emissive sources.", + "3. We provide a benchmark dataset designed to evaluate the performance of emissive source reconstruction.", + "4. Our method is applicable to the scenes with or without emissive sources, achieving superior mesh reconstruction results on the DTU dataset." + ], + "bbox": [ + 76, + 439, + 468, + 618 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related work", + "text_level": 1, + "bbox": [ + 76, + 633, + 215, + 648 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Neural Rendering. Advancements in implicit representations [51, 60] and volume rendering [39] have significantly enhanced neural rendering capabilities, enabling the reconstruction of scene components from 2D images. One of the key directions is mesh extraction [44, 70, 77, 78, 80, 97], with methods like NeuS [68] and VolSDF [87] utilizing signed distance function (SDF) values for volume rendering. Recently, the efficient computation of volume rendering has become a focal point due to the substantial computational cost associated with network inference for ray color calculation [41, 49, 88]. Several methods propose to directly predict ray color using the 4D light fields concept [1, 52, 55] or leveraging voxel grids for fast inference of spatial features [5, 11, 12, 14, 31, 56]. NeuralRadiosity [17] shares similarity with our method, as it predicts raytraced values instead of explicitly tracing individual rays.", + "bbox": [ + 75, + 657, + 472, + 902 + ], + "page_idx": 1 + }, + { + "type": "table", + "img_path": "images/f0fb17b5a2737860d7e30c85f898be2f8463abfbc7c165c6e07c05e8418eb9f8.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
VoxurfTensoIRPath TracingESR-NeRF
Big Onn · d(n · d)b+1n2 · d
Indirect illuminationx✓✓✓
BRDF decompositionx✓✓✓
Emissive source controlxx✓✓
", + "bbox": [ + 501, + 89, + 893, + 157 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Table 1. Computational cost comparison for inverse rendering methods. $n$ is the number of sampled points along a ray, $d$ is the number of scattering rays, and $b$ is the number of ray bounces.", + "bbox": [ + 498, + 167, + 890, + 209 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "However, they primarily focus on calculating the final object color when all scene information is available. In contrast, our inverse rendering approach aims to reconstruct emissive sources within a scene, addressing the ambiguities introduced by their presence in LDR images.", + "bbox": [ + 496, + 218, + 890, + 294 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Inverse Rendering. A growing emphasis revolves around the decomposition of materials represented by spatially varying bidirectional reflectance distribution functions (SVBRDF) [46, 83, 99]. To lessen the computational burden in inverse rendering [25, 54, 96, 98], several methods have adopted neural networks as lookup tables [9] or computational caches [54, 90, 95]. While NeRV [54] utilizes caching visibility and $\\mathrm{NeILF++}$ [90] adopts caching surface point radiance with the inter-reflection loss for incident radiance, our method diverges by focusing on tracing radiance origins. Specifically, we aim to identify emissive sources within a scene, moving beyond the simplification of incident radiance calculations. Several methods rely on diverse known lighting configurations to exploit variations in object appearances [59, 64, 84, 89]. Toggling emissive sources on and off resembles the common one-light-at-a-time (OLAT) technique, as seen in NLT [94] and ReN-eRF [82]. However, our setting does not need to know light source properties and to toggle lights individually. Instead, we allow for toggling all lights together. Recent works have also jointly reconstruct the mesh, materials, and lighting [20, 35, 42, 57]. They tackle with images captured under a single unknown lighting condition [92, 95], assuming that radiance already encodes global illumination [75, 96]. However, they confine to the scenes illuminated by far-distant lights, constrained to an 8-bit color spectrum. Our work considers the presence of multiple emissive sources within a scene captured in LDR images, questioning the prevailing notion that radiance fields trained with the image rendering loss faithfully represents global illumination. While some methods [2, 19, 32, 48, 76] deal with the scenes featuring emissive sources, they work outside the volume rendering framework and depend on HDR input images, assuming prior knowledge of scene geometry.", + "bbox": [ + 496, + 300, + 890, + 815 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3. Preliminaries", + "text_level": 1, + "bbox": [ + 500, + 829, + 638, + 845 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Surface Representation. Analogous to NeRF [40], neural network $f_{\\theta}$ predicts SDF values at arbitrary 3D spatial locations. NeuS [68] integrates surface representa", + "bbox": [ + 498, + 854, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "4599", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/114f18d634099b544fc56a064202769a983a7f934e4831d24cb881fa79374340.jpg", + "image_caption": [ + "On Images" + ], + "image_footnote": [], + "bbox": [ + 81, + 109, + 173, + 180 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/9a472dcd746a19a3e978d986bc74651228309b044ee968c37da34649cf9f85fb.jpg", + "image_caption": [ + "Off Images" + ], + "image_footnote": [], + "bbox": [ + 81, + 205, + 173, + 275 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/7c30826d9f9f5e59b8b5f0d05742991f50e3d6adc4853cb0fb80803759bb1545.jpg", + "image_caption": [ + "Multi-view Input", + "Training" + ], + "image_footnote": [], + "bbox": [ + 181, + 102, + 669, + 292 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/b6bae7cda4ddb5a80e5c74766f89f303c299dc9ed7c045fe30d0f9548596a16f.jpg", + "image_caption": [ + "Application", + "Figure 2. The pipeline of emissive source reconstruction. Given LDR images with emissive sources on and off, scene components are reconstructed by synthesizing training images and enforcing LTS requirements. Emissive sources are progressively refined via categorizing training rays into uncertain and certain groups. The scenes can be edited with new lighting conditions using reconstructed emissive sources." + ], + "image_footnote": [], + "bbox": [ + 679, + 103, + 890, + 292 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "tion into volume rendering using the SDF-based opacity $\\rho (x) = \\max (\\frac{-\\frac{d\\Phi_s}{dx}(f(x))}{\\Phi_s(f(x))},0)$ . Here $\\Phi_s(x) = (1 + e^{-sx})^{-1}$ is the sigmoid function where $s$ controls the sharpness of surfaces. The color of a ray can be calculated as", + "bbox": [ + 75, + 361, + 468, + 428 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {C} (r) = \\int_ {0} ^ {\\infty} T (r (t)) \\rho (r (t)) L _ {o} (r (t), \\omega_ {o}) d t, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 135, + 434, + 468, + 465 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\hat{C}(r)$ denotes the predicted ray color, $r(t; c, \\omega_o) = c - t \\cdot \\omega_o$ is the ray with camera center $c$ along direction $\\omega_o$ , $T(r(t)) = \\exp \\left( \\int_0^t -\\rho (r(u))du \\right)$ is the transmittance, and $L_{o}(r(t), \\omega_{o})$ is the outgoing radiance. Henceforth, we use $x$ to denote a point in $r(t; c, \\omega_o)$ for notational simplicity.", + "bbox": [ + 76, + 472, + 470, + 551 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Light Transport in Volume Rendering. Extracting light sources necessitates analyzing the causes affecting the final ray colors. Kajiya's rendering equation [26] factorizes the outgoing radiance $L_{o}(x,\\omega_{o})$ into emission and reflections:", + "bbox": [ + 75, + 556, + 468, + 617 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nL _ {o} (x, \\omega_ {o}) = E (x) + \\int_ {\\Omega} L _ {i} (x, \\omega_ {i}) R (x, \\omega_ {o}, \\omega_ {i}; b) d \\omega_ {i}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 99, + 622, + 468, + 651 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $E(x)$ is the emission, $R(x,\\omega_o,\\omega_i;b)$ represents the SVBRDF parametrized by parameters $b$ with Lambert cosine multiplied, and $L_{i}(x,\\omega_{i})$ is the incident radiance. In volume rendering, computing the incident radiance at point $x$ is akin to evaluating Eq. 1, with $x$ serving as the camera center. By iteratively factorizing the outgoing radiance in the incident radiance, the contribution of a path length $i$ for a pixel can be decomposed as in Eq. 3, where $\\mathcal{H}_i = \\Pi_{j=1}^{i-1}T(x_j)\\rho(x_j)R(x_j,\\omega_{j-1},\\omega_j)$ is the path throughput, $S(\\omega_i)$ is the environment map strength in direction $\\omega_i$ , and $V(x,\\omega_i) = \\exp\\left(\\int_0^\\infty -\\rho(r(u;x_i,-\\omega_i))du\\right)$ is the visibility of the environment map at point $x$ along direction $\\omega_i$ :", + "bbox": [ + 75, + 655, + 468, + 839 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} P _ {i} = \\int_ {l _ {1}} \\int_ {\\Omega} \\dots \\int_ {l _ {i - 1}} \\int_ {\\Omega} \\left(\\int_ {l _ {i}} T (x _ {i}) \\rho (x _ {i}) E (x _ {i}) d t _ {i} + \\right. \\tag {3} \\\\ S \\left(\\omega_ {i - 1}\\right) V \\left(x _ {i - 1}, \\omega_ {i - 1})\\right) \\mathcal {H} _ {i} d t _ {1} d \\omega_ {1} \\dots d t _ {i - 1} d \\omega_ {i - 1}. \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 99, + 853, + 468, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Extending the analysis to longer light paths, or equivalently, increasing the number of ray bounces, leads to exponential growth in computation complexity. This poses a challenge when attempting to decompose the influence of unknown emissive sources, as their ability to produce strong reflections makes ignoring indirect illumination infeasible.", + "bbox": [ + 498, + 361, + 892, + 453 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4. Methodology", + "text_level": 1, + "bbox": [ + 500, + 465, + 635, + 483 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "None of the previous works address the reconstruction of emissive sources from LDR multi-view images. Sec. § 4.1 through § 4.5 detail our method, ESR-NeRF, which reconstructs emissive sources without prior knowledge of scene geometry, materials, or lighting specifics (including their location, number, or colors). We also show how these reconstructed sources can be used for scene editing in § 4.5.", + "bbox": [ + 498, + 489, + 892, + 597 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4.1. Learnable Tone-mapper", + "text_level": 1, + "bbox": [ + 500, + 604, + 723, + 622 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Throughout the paper, we use $\\mathcal{R}$ to represent camera rays, $C$ for pixel values, and a binary flag $\\mathbb{I}$ to indicate whether an image is captured with emissive sources on or off.", + "bbox": [ + 498, + 628, + 890, + 672 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To extract HDR values from LDR images, we employ the softplus activation for outgoing radiance prediction and apply a clipping and gamma function $\\tau$ [21] for the rendering loss such that $\\hat{C}_{\\tau}(r) = \\tau (\\hat{C} (r))$ . Unlike previous NeRF-based works [25, 37, 54, 57] that limit radiance to the range of [0, 1], our approach allows for any positive radiance values. Yet, it creates difficulties in differentiating between the surface weight $T(x)\\rho (x)$ and the magnitude of radiance value $L_{o}(x,\\omega_{o})$ , since it allows for the possibility of assigning extreme radiance to the points with low surface weights to render same ray colors. Such ambiguity poses challenges, particularly in dark and high-contrast scenes, aggravating surface reconstruction (see Fig. 3). To address this, we introduce a learnable tone-mapper $m_{\\theta}:\\mathbb{R}_{+}^{3}\\to [0,1]^{3}$ , that takes positionally encoded HDR linear values as input:", + "bbox": [ + 496, + 674, + 892, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "4600", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/a3c9e215d5e892eb640ca4cdf9957db220e3027cb39b41a155a9983267ed6771.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 98, + 94, + 205, + 151 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/9e5cd54c95fe67b079c517c178a5061f7e5223ce6de42a3cbda1377eee353d56.jpg", + "image_caption": [ + "(a) Image" + ], + "image_footnote": [], + "bbox": [ + 89, + 157, + 212, + 229 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/31d8bf7ded52d96af3e1dc3ff1a5af6af66fa47b75a09972fe4ec68bfc51cd8a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 220, + 94, + 326, + 152 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/0614aa82aa0d17183c377cbc29216fbf0dcdda4bd9678bb34c6b2ffa50354f00.jpg", + "image_caption": [ + "(b) $\\mathrm{w / o}m_{\\theta}$" + ], + "image_footnote": [], + "bbox": [ + 223, + 157, + 316, + 224 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/cc73d8dc0bd607c86cea402aa62c9d82edeee5bf331f29c17f1c84ba3f8ad670.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 343, + 94, + 447, + 152 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/4367bb8019d50bddd926a6b6803b02803e6c00e46538ddaba63eda88a8c5eefe.jpg", + "image_caption": [ + "(c) ESR-NeRF", + "Figure 3. Reconstructed surfaces with the learnable tone-mapper." + ], + "image_footnote": [], + "bbox": [ + 344, + 157, + 437, + 224 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {C} _ {m _ {\\theta}} (r) = \\int_ {0} ^ {\\infty} T (x) \\rho (x) m _ {\\theta} \\left(L _ {o} \\left(x, \\omega_ {o}\\right)\\right) d t, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 135, + 284, + 468, + 314 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nL _ {o} (x, \\omega_ {o}) = L _ {o} ^ {S} (x, \\omega_ {o}) + L _ {o} ^ {E} (x, \\omega_ {o}) \\cdot \\mathbb {I}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 148, + 321, + 468, + 340 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $L_{o}^{S}(x,\\omega_{o})$ is radiance when emissive sources are turned off, while $L_{o}^{E}(x,\\omega_{o})$ stands for radiance added to the scene by emissive sources. Our rendering loss is then formulated as follows, with $\\lambda_{\\tau}$ as a hyper-parameter:", + "bbox": [ + 76, + 343, + 468, + 405 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {r e n d e r}} = \\sum_ {r \\in \\mathcal {R}} \\left(\\| C (r) - \\hat {C} _ {m _ {\\theta}} (r) \\| _ {2} ^ {2} + \\lambda_ {\\tau} \\| C (r) - \\hat {C} _ {\\tau} (r) \\| _ {2} ^ {2}\\right). \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 84, + 415, + 468, + 445 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.2. Learning of Light Transport Segments", + "text_level": 1, + "bbox": [ + 76, + 462, + 408, + 479 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The computational complexity of object appearance analysis in volume rendering is notably high, as shown in Eq. 3. We take an alternative approach by leveraging neural networks to represent ray-traced fields, rather than explicitly tracing every rays. Our distinct contribution to inverse rendering lies in precise adjustment of radiance. Specifically, we impose constraints on the predicted radiance to satisfy each light transport segments. The light transport segments (LTS) loss, $\\mathcal{L}_{lts}$ , plays a pivotal role in our method:", + "bbox": [ + 75, + 486, + 468, + 622 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {l t s} ^ {S} = \\sum_ {x, \\omega_ {o}} \\| L _ {o} ^ {S} (x, \\omega_ {o}) - \\hat {L} _ {o} ^ {S} (x, \\omega_ {o}) \\| _ {2} ^ {2}, \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 148, + 632, + 468, + 662 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {l t s} ^ {E} = \\sum_ {x, \\omega_ {o}} \\| L _ {o} ^ {E} (x, \\omega_ {o}) - \\hat {L} _ {o} ^ {E} (x, \\omega_ {o}) \\| _ {2} ^ {2}, \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 148, + 675, + 468, + 705 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {L} _ {o} ^ {S} (x, \\omega_ {o}) = \\int_ {\\Omega} \\underbrace {S (\\omega_ {i}) V (x , \\omega_ {i}) R (x , \\omega_ {o} , \\omega_ {i})} _ {\\text {d i r e c t i l l u m i n a t i o n b y a n e v i r o n m e n t m a p}} d \\omega_ {i} +\n$$\n", + "text_format": "latex", + "bbox": [ + 96, + 722, + 411, + 767 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\int_ {\\Omega} \\int_ {0} ^ {\\infty} \\underbrace {T (x ^ {\\prime}) \\rho (x ^ {\\prime}) L _ {o} ^ {S} (x ^ {\\prime} , - \\omega_ {i}) d t ^ {\\prime} R (x , \\omega_ {o} , \\omega_ {i})} _ {\\text {i n d r e c t i l l u m i n a t i o n b y a n e v i r o n m e n t m a p}} d \\omega_ {i}.\n$$\n", + "text_format": "latex", + "bbox": [ + 101, + 763, + 426, + 805 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {L} _ {o} ^ {E} (x, \\omega_ {o}) = \\underbrace {E (x)} _ {\\text {e m i s s i o n}} +\n$$\n", + "text_format": "latex", + "bbox": [ + 96, + 827, + 230, + 861 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\int_ {\\Omega} \\int_ {0} ^ {\\infty} \\underbrace {T \\left(x ^ {\\prime}\\right) \\rho \\left(x ^ {\\prime}\\right) L _ {o} ^ {E} \\left(x ^ {\\prime} , - \\omega_ {i}\\right) d t ^ {\\prime} R \\left(x , \\omega_ {o} , \\omega_ {i}\\right)} _ {\\text {d i r e c t ＆ i n d i r c t i l l u m i n a t i o n b y e m i s s i v e s o u r c e s}} d \\omega_ {i}. \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 96, + 859, + 468, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We ensure consistency between the radiance directly predicted by the network $L_{o}(x,\\omega_{o})$ and the radiance achievable based on the scene context $\\hat{L}_{o}(x,\\omega_{o})$ . Previous approaches have focused on matching $\\hat{L}_{o}(x,\\omega_{o})$ to training views, overlooking the relations to $L_{o}(x,\\omega_{o})$ . This hinders the restoration of HDR radiance by supervising scene components to LDR training views. In contrast, our LTS loss enables volumetric energy transfer of radiance, adjusting outgoing radiance based on their interrelations.", + "bbox": [ + 500, + 90, + 890, + 226 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To implement this concept, we train six dedicated networks for SDF $f(x)$ , SVBRDF parameters $b(x)$ , emission $E(x)$ , environment map $S(\\omega_{i})$ , outgoing radiances $L_{o}^{S}(x,\\omega_{o})$ and $L_{o}^{E}(x,\\omega_{o})$ , to adhere to these LTS requirements. For the environment map, we represent it using 48 Spherical Gaussians [67]: $\\sum_{k=1}^{M}\\mu_{k}e^{\\lambda_{k}(\\omega_{i}\\cdot\\xi_{k}-1)}$ , followed by the softplus activation. $\\mu\\in\\mathbb{R}^{3},\\lambda\\in\\mathbb{R}_{+}$ , and $\\xi\\in\\mathbb{S}^{2}$ respectively denote the lobe amplitude, sharpness, and axis.", + "bbox": [ + 498, + 227, + 890, + 348 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.3. Progressive Discovery of Reflection Areas", + "text_level": 1, + "bbox": [ + 500, + 356, + 852, + 372 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/e1575f4b7ab33fc445b5dd92db9dcbcce6315d95324ef4a2ac79713e6f3f9f93.jpg", + "image_caption": [ + "Figure 4. Left: Image with active emissive sources. Right: Identified emissive sources w/o progressive discovery of reflection areas." + ], + "image_footnote": [], + "bbox": [ + 501, + 382, + 594, + 459 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/93ca78abdc92500ef1608c9707574b5b883aec36de75958d071a0a5a87d6809b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 596, + 383, + 687, + 459 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Relying solely on LTS is insufficient for addressing ambiguity arising from low pixel values of emissive sources and intense reflections in adjacent regions, often leading to confusion between emission and reflection. The right image in Fig. 4 shows self-emitting objects restored with the", + "bbox": [ + 697, + 378, + 890, + 544 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "naive LTS loss. While emissive sources are small, large areas affected by them are also identified as emissive sources. We propose a reflection-aware progressive approach for precise identification of emissive sources. By leveraging LTS learning, we extend the regions that can be regarded as reflection areas. Fig. 5 illustrates our progressive algorithm.", + "bbox": [ + 496, + 545, + 890, + 636 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Reflection-Aware Emission Refinement. Since surface points are unknown and are updated during learning, we opt to utilize rays rather than surface points. This process involves categorizing training rays into two groups: uncertain $(\\mathcal{R}^U)$ and certain $(\\mathcal{R}^C)$ . The certain group contains the rays confidently identified as reflection, aiding the transfer of radiance energy to nearby points. For the points in the certain group, we use the Eq. 11 instead of Eq. 10 to exclusively attribute outgoing radiances to reflections. Satisfying the LTS loss on the certain group results in adjusting the outgoing radiances of influential points, as illustrated in Fig. 5(a):", + "bbox": [ + 496, + 642, + 890, + 808 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {L} _ {o} ^ {E} \\left(x, \\omega_ {o}\\right) = \\int_ {\\Omega} \\int_ {0} ^ {\\infty} T \\left(x ^ {\\prime}\\right) \\rho \\left(x ^ {\\prime}\\right) L _ {o} ^ {E} \\left(x ^ {\\prime}, - \\omega_ {i}\\right) d t ^ {\\prime} R \\left(x, \\omega_ {o}, \\omega_ {i}\\right) d \\omega_ {i}. \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 500, + 814, + 908, + 854 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The uncertain group includes the rays indicating the areas that are undetermined yet as reflection or emission. Using Eq. 12 to compute $\\hat{L}_o^E (x,\\omega_o)$ , this group adjusts emis", + "bbox": [ + 498, + 854, + 890, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4601", + "bbox": [ + 480, + 944, + 513, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "sions $E(x)$ based on the radiance updates by the certain group, where \"sg\" represents the stop-gradient:", + "bbox": [ + 76, + 90, + 468, + 121 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\hat {L} _ {o} ^ {E} (x, \\omega_ {o}) = E (x) + \\\\ \\left. \\operatorname {s g} \\left(\\int_ {\\Omega} \\int_ {0} ^ {\\infty} T \\left(x ^ {\\prime}\\right) \\rho \\left(x ^ {\\prime}\\right) L _ {o} ^ {E} \\left(x ^ {\\prime}, - \\omega_ {i}\\right) d t ^ {\\prime} R \\left(x, \\omega_ {o}, \\omega_ {i}\\right) d \\omega_ {i}\\right). \\right. \\tag {12} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 86, + 128, + 468, + 189 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As shown in Fig. 5(b), this leads to increased emissions for the regions whose radiances are adjusted to account for the reflections in the certain group. Conversely, emissions decrease for the regions where there is little change in outgoing radiance, but incident radiances are increased by surrounding influential points.", + "bbox": [ + 76, + 191, + 468, + 281 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Ray Group Management. As emissions and radiances are adjusted, the groups are dynamically updated at predefined training intervals through the following process. Within the uncertain group, we evaluate the expected emission strength of rays, retaining only those above a threshold $k_{i}$ . Rays below this threshold are then merged to the certain group:", + "bbox": [ + 76, + 287, + 468, + 378 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {R} _ {i} ^ {U} = \\left\\{r \\mid \\max _ {R G B} \\left(\\int_ {0} ^ {\\infty} T (x) \\rho (x) E (x) d t\\right) \\geq k _ {i}, r \\in \\mathcal {R} _ {i - 1} ^ {U} \\right\\}, \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 86, + 386, + 468, + 426 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {R} _ {i} ^ {C} = \\left(\\mathcal {R} _ {i - 1} ^ {U} - \\mathcal {R} _ {t} ^ {U}\\right) \\cup \\mathcal {R} _ {i - 1} ^ {C}. \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 174, + 438, + 468, + 460 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Subsequently, newly added rays to the certain group can be used to localize influential points and update their outgoing radiances. This iterative process progressively refines the separation between reflective and emissive regions, attaining more accurate identification of emissive sources.", + "bbox": [ + 76, + 464, + 468, + 539 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "LTS Loss Decomposition. The LTS loss, as detailed in Eq. 15, can be decomposed using a stop-gradient operation to refine the adjustment process.", + "bbox": [ + 76, + 545, + 468, + 590 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {l t s} ^ {E} = \\sum_ {x, \\omega_ {o}} \\left(\\lambda_ {l} \\| \\operatorname {s g} \\left(L _ {o} ^ {E} (x, \\omega_ {o})\\right) - \\hat {L} _ {o} ^ {E} (x, \\omega_ {o}) \\| _ {1} + \\right. \\tag {15}\n$$\n", + "text_format": "latex", + "bbox": [ + 109, + 598, + 468, + 630 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\lambda_ {r} \\| L _ {o} ^ {E} (x, \\omega_ {o}) - \\operatorname {s g} (\\hat {L} _ {o} ^ {E} (x, \\omega_ {o})) \\| _ {1}).\n$$\n", + "text_format": "latex", + "bbox": [ + 187, + 631, + 408, + 648 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We prioritize $\\lambda_{l}$ to enhance the update of scene context, affecting other points' radiance given the predicted $L_{o}(x,\\omega_{o})$ . $\\lambda_{r}$ prevents severe deviation of every $L_{o}(x,\\omega_{o})$ within the current scene context. This aligns with our focus on HDR source reconstruction from LDR images, addressing under-represented information in training data.", + "bbox": [ + 76, + 657, + 468, + 747 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.4. Training Details", + "text_level": 1, + "bbox": [ + 76, + 756, + 236, + 771 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We employ the Voxurf architecture [80] as backbone and adopt the simplified Disney BRDF model [10] for SVBRDF representation, with parameters including base color $\\in [0,1]^3$ , roughness $\\in [0,1]$ , and metallic $\\in [0,1]$ . The learnable tone-mapper, structured as a two-layer MLP, is utilized for the rendering loss only. Initially, we pre-train our networks using the rendering loss, subsequently integrating the basic LTS loss (Eq. 7 and Eq. 8) into our training regimen.", + "bbox": [ + 76, + 779, + 468, + 900 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/c0f1cde92af7bc78d655d0417f1087a8b316296c02b105f38ca8940638e37e8b.jpg", + "image_caption": [ + "(a) LTS on $R^C$" + ], + "image_footnote": [], + "bbox": [ + 501, + 88, + 622, + 148 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/12dab668070569d1475cefec26d1388d4bb0c1e5315387551bfbfa752f6ab277.jpg", + "image_caption": [ + "(b) LTS on $R^U$", + "Figure 5. Illustration of the progressive emissive source reconstruction with reflection awareness. Gray color represents the areas belonging to the certain group, while the red (emissive sources) and orange (their reflections) areas belong to the uncertain group." + ], + "image_footnote": [], + "bbox": [ + 633, + 88, + 754, + 148 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/2ea8217914c690e071b09f71ed17e37be0f69dcf599092c298054181d0ac1531.jpg", + "image_caption": [ + "(c) Ray Group Managing" + ], + "image_footnote": [], + "bbox": [ + 766, + 95, + 887, + 148 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "This phase transitions to the reflection-aware progressive training scheme, where we adopt the $\\ell_1$ loss due to its empirical stability in refining emissive source reconstruction. We use a smoothing regularization to promote local consistency in normals, BRDFs, and emissions. To ensure view-consistent labeling of 3D points as either reflective or emissive, we implement the emission suppression loss for points belonging to the certain group:", + "bbox": [ + 496, + 241, + 890, + 361 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {s u p p}} ^ {E} = \\sum_ {r \\in R _ {t} ^ {C}} \\| \\int_ {0} ^ {\\infty} T (x) \\rho (x) E (x) d t \\| _ {2} ^ {2}, \\tag {16}\n$$\n", + "text_format": "latex", + "bbox": [ + 566, + 367, + 890, + 404 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The threshold $k_{i}$ linearly increases with each time step $t$ , utilizing a grid search within a range of $[10^{-3}, 10^{-5}]$ to find the slope. We construct mini-batches via stratified sampling within each group. For a detailed description of our training procedure, please refer to Appendix.", + "bbox": [ + 496, + 411, + 890, + 487 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.5. Scene Editing", + "text_level": 1, + "bbox": [ + 500, + 494, + 643, + 511 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Reconstructed emissive sources enable scene editing; users select emissive sources using binary masks $M_{j=1\\dots N}$ and specify lighting conditions using colors $c_{j=1\\dots N}$ and intensities $i_{j=1\\dots N}$ within the HSV color space [53].", + "bbox": [ + 496, + 518, + 890, + 578 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We identify the rays in the uncertain group that match $M$ by projecting expected surface points $p$ of the rays onto the camera with the pose $\\mathbf{R}|\\mathbf{t}$ :", + "bbox": [ + 496, + 579, + 890, + 623 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\np = \\int_ {0} ^ {\\infty} T (x) \\rho (x) x d t, \\tag {17}\n$$\n", + "text_format": "latex", + "bbox": [ + 617, + 630, + 890, + 659 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbb {I} _ {j} ^ {\\text {h i t}} (x) = \\operatorname {i n t e r p} \\left(M _ {j}, p ^ {\\prime}\\right) > 0, \\text {w h e r e} p ^ {\\prime} = \\mathbf {K} [ \\mathbf {R} | \\mathbf {t} ] [ p | 1 ] ^ {T}. \\tag {18}\n$$\n", + "text_format": "latex", + "bbox": [ + 508, + 680, + 890, + 699 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For the rays satisfying $\\mathbb{I}_j^{hit}(x)$ , we apply the designated lighting conditions. The new emission values are computed by substituting the original hue (H) and saturation (S) of $E(x)$ with the user-specified color $c_{j}$ and adjusting the value (V) of $v(x)$ with the new intensity $i_j$ :", + "bbox": [ + 496, + 705, + 890, + 781 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nE (x) = \\operatorname {h s v} _ {-} \\operatorname {t o} _ {-} \\operatorname {r g b} ([ c _ {j} | (v (x) \\times i) ]) \\cdot \\mathbb {I} _ {j} ^ {h i t} + E (x) \\cdot \\neg \\mathbb {I} _ {j} ^ {h i t}. \\tag {19}\n$$\n", + "text_format": "latex", + "bbox": [ + 501, + 787, + 890, + 806 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "These modifications influence scene appearance by optimizing the loss in Eq. 20. During this process, all networks, except for $L_{o}^{E}(x,\\omega_{o})$ , are frozen:", + "bbox": [ + 496, + 813, + 890, + 859 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {e d i t}} = \\sum_ {x, \\omega_ {o}} \\| L _ {o} ^ {E} (x, \\omega_ {o}) - \\operatorname {s g} \\left(\\hat {L} _ {o} ^ {E} (x, \\omega_ {o})\\right) \\| _ {2} ^ {2}. \\tag {20}\n$$\n", + "text_format": "latex", + "bbox": [ + 555, + 864, + 890, + 895 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "4602", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/6596b2f39001e3c50406544f3244b3be821d46fc08e1b2cafde2b6713afc71e5.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
LegoGiftWhite colored BookCubeBillboardBallsVivid colored
IoUMSEIoUMSEIoUMSEIoUMSEIoUMSEIoUMSEIoUMSEIoUMSEIoUMSEIoUMSEIoUMSE
Twins0.2220.190.498.590.633.910.9531.830.691.120.900.060.256.960.246.090.552.630.9510.640.090.750.830.04
NeLF++0.4320.880.079.380.954.640.9332.670.011.950.910.800.307.650.096.860.953.360.9411.490.021.570.920.78
TensoIR0.7120.130.158.550.953.870.9531.730.761.110.950.050.336.930.156.050.952.590.9610.600.770.740.950.03
ESR-NeRF0.818.380.603.490.961.190.9717.870.840.460.950.040.515.480.592.500.960.510.977.940.880.260.940.03
", + "bbox": [ + 81, + 75, + 890, + 176 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 76, + 241, + 209, + 258 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We assess ESR-NeRF in reconstructing emissive sources by focusing on both identification and intensity restoration. To showcase its effectiveness, we conduct a range of experiments, including scene editing, ablation studies, illumination decomposition, and surface reconstruction, providing both quantitative and qualitative results.", + "bbox": [ + 75, + 267, + 470, + 358 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1. Experiment Settings", + "text_level": 1, + "bbox": [ + 76, + 369, + 269, + 386 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We curate 6 diverse synthetic scenes, each with 200 training images evenly distributed between on and off lighting conditions. To evaluate the robustness of our approach against light colors, we consider two distinct settings of white colored and vivid colored emissive sources, resulting in a total of 12 scenes. The vivid colors are selected with full saturation in the HSV color space. We measure source identification and radiance reconstruction using IoU and MSE metrics on novel view test images, comparing against ground truth data from Blender-rendered emission masks and EXR files. The emission strengths, the maximum EXR file values, range from 2 to 200. For quantitative scene editing evaluation, we alter the white-colored sources to various colors—red, green, blue, cyan, magenta, yellow—and adjust intensities to half or double their original values. Qualitative results include scene editing for vividly colored sources and real scenes captured with a Fuji 100s camera using Philips smart bulbs as emissive sources. Quantitative assessments are based on 50 test images from novel camera poses, except for MSE measured for 25 test", + "bbox": [ + 75, + 393, + 472, + 698 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/d007f314138098088d81eb12ba8af87dd4e8655fb8e78e7e43cb2e7a13fce647.jpg", + "image_caption": [ + "Figure 6. Comparison of identified emissive sources. ESR-NeRF excels through the reflection-aware progressive refinement." + ], + "image_footnote": [], + "bbox": [ + 81, + 702, + 465, + 861 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/703f7d45c49734b95212d29c670341090b96904f8005a10cd8e6a3d5da56c7c7.jpg", + "table_caption": [ + "Table 2. Results of emissive source identification. ESR-NeRF outperforms state-of-the-art re-lighting methods in reconstructing emissive sources, regardless of their color. The IoU measures the source area identification (a higher value is better), and the MSE quantifies the difference between reconstructed images and HDR ground truth images (a lower value is better)." + ], + "table_footnote": [], + "table_body": "
NVNV + INV+ CNV + I + C
PSNRLPIPSPSNRLPIPSPSNRLPIPSPSNRLPIPS
Twins36.520.014127.910.025231.020.025228.210.0310
NeRF-W36.440.014224.770.0417----
NeILF++24.400.055624.710.057924.060.075023.240.0770
TensoIR38.040.010327.280.041826.360.050525.180.0531
PaletteNeRF33.660.023323.270.048324.440.064622.580.0703
ESR-NeRF38.790.008329.990.019331.730.019631.630.0199
", + "bbox": [ + 500, + 238, + 898, + 349 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 3. Scene editing results. NV: novel view synthesis, I: intensity editing, and C: color editing. A higher PSNR or lower LPIPS value is better.", + "bbox": [ + 498, + 359, + 893, + 400 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "images. We denote the best performance with blue and the second-best with green. Additionally, we utilize the DTU dataset [23] to evaluate ESR-NeRF's performance in surface reconstruction tasks where emissive sources are absent.", + "bbox": [ + 498, + 407, + 890, + 468 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Baselines. We select two state-of-the-art re-lighting methods, TensoIR [25] and NeILF++ [90], that do not require prior lighting information. For thorough evaluation, we also implement a simple method, Twins, where separate models are trained under light on and off conditions. The Twins utilize the radiance discrepancies between the on and off models to distinguish and adjust emissive sources. For scene editing, we add NeRF-W [38] and PaletteNeRF [28] as baselines. Both NeRF-W and Twins adopt the Voxurf [80] architecture for fair comparison. For methods unable to individually control emissive sources, all sources are adjusted together to match the last lighting condition by a user. For the DTU dataset, we include state-of-the-art surface reconstruction methods that use object masks, such as NeuS [68] and Voxurf, as well as Neural-PBIR [57], that jointly reconstructs surfaces, materials, and environment maps.", + "bbox": [ + 496, + 474, + 892, + 717 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.2. Results", + "text_level": 1, + "bbox": [ + 500, + 726, + 591, + 739 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Emissive Source Recosnstruction. Tab. 2 shows that our approach excels in accurately identifying emissive source regions and restoring their intensity, regardless of the source color. While TensoIR and NeILF++ can restore emissions by modifying their physical rendering equations, they suffer from emissive source ambiguity, leading to near-zero IoU performance (see Appendix). For a comprehensive comparison, we report the best performance of the baseline methods using thresholding on the reconstructed emission strength at 0.01 intervals. ESR-NeRF consistently outper", + "bbox": [ + 496, + 750, + 890, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "4603", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/d36286269b953d85620873449b3bf2d8cdd8c1686bd39beb6787efc64e786250.jpg", + "image_caption": [ + "Figure 7. Comparison of scene editing. ESR-NeRF provides precise source control and faithfully represents reflection effects. For easy comparison in the Cube scene of low intensity, the bottom-right images are presented with a $40\\%$ increased brightness." + ], + "image_footnote": [], + "bbox": [ + 91, + 71, + 872, + 406 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "forms the baselines in identifying emissive source regions across all scenes. Our method also achieves significantly lower MSE values for restoring LDR to HDR images compared to the baselines, demonstrating its effectiveness of handling the ill-posed nature of the scenes with emissive sources. This is visually confirmed in Fig. 6, where ESR-NeRF surpasses the baselines in a complex scene with numerous small light bulbs.", + "bbox": [ + 75, + 445, + 472, + 568 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Scene Editing. Tab. 3 and Fig. 7 showcase the scene editing results under novel lighting conditions. Baseline methods struggle to adapt to lighting changes due to their inability to reconstruct emissive sources accurately. For example, in the Lego scene, TensoIR fails to adjust the illumination in surrounding regions when the color of emissive sources is changed, and in the Cube scene, both the hidden iPad screen and the cube surface covered by the user input mask change together. Twins introduces blue light onto yellow and red surfaces, leading to unintended white and purple appearances, even though there should be no reflection. PaletteNeRF, which manipulates scenes through re-colorization, lacks precise control over illumination, as seen in the synchronous color changes in the yellow ribbon and lighting. In contrast, ESR-NeRF demonstrates superior performance in", + "bbox": [ + 75, + 575, + 470, + 803 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/6dcd2109976ab572d9a3bbbb7d499046b79d2ce40f45aafb878806bc0c0c0012.jpg", + "image_caption": [ + "Figure 8. Reconstructed emitter and re-lighting at novel view." + ], + "image_footnote": [], + "bbox": [ + 81, + 811, + 470, + 883 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "scene editing outshining all baselines thanks to the accurate identification of emissive sources, as detailed in Table 3. ESR-NeRF effectively balances source reconstruction and novel view synthesis, ensuring high performance in both tasks. NeRF-W is excluded from color adjustments since it doesn't support direct color change through interpolating latent variables learned with light on and off conditions.", + "bbox": [ + 496, + 445, + 890, + 551 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Fig. 8 to 9 present additional examples of emissive source reconstruction and scene editing results. Fig. 10 shows results on real scenes, for which due to the impracticality of precise control over smart bulb colors, we offer emission reconstruction results with pseudo ground truth data. Our method effectively identifies emissive sources in real scenes, while it faces challenges in capturing complex", + "bbox": [ + 496, + 551, + 892, + 657 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/cacfadfd51a4853b5aea10c887268a2ed8cd3fc64bf02d6e4591d743f3084d9f.jpg", + "image_caption": [ + "Figure 9. Results of source reconstruction and scene editing." + ], + "image_footnote": [], + "bbox": [ + 500, + 659, + 890, + 885 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "4604", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b4b10d7590b1586aa937b771c07a96207d6996a5bc485653d2fe26cb0c572bc2.jpg", + "image_caption": [ + "Figure 10. Source reconstruction and scene editing on real scenes." + ], + "image_footnote": [], + "bbox": [ + 76, + 80, + 468, + 253 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "reflections within light bulbs, as evident in the bright spot at the center of the bulbs in the ground truth edit results.", + "bbox": [ + 75, + 286, + 468, + 316 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Ablation Analysis. Progressive refinement with the stop-gradient operation in Eq. 15 improves the identification of emissive sources and reduces MSE values. Without $m_{\\theta}$ , surface reconstructions become unreliable, complicating the accurate reconstruction of emissive sources. This issue is evident from the CD metrics and illustrated in Fig. 3. Further analyses are provided in Appendix.", + "bbox": [ + 75, + 323, + 468, + 429 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Illumination Decomposition. Fig. 11 demonstrates ESR-NeRF's decomposition of scene illumination into direct and indirect lighting from an environment map, as well as emissions and their reflections. The shadow behind the yellow ribbon in the direct figure and the illumination in the indirect figure showcase ESR-NeRF's ability to model both direct and indirect illumination. The reflection figure shows that our method accurately captures how emissive sources contribute to reflections on nearby regions.", + "bbox": [ + 75, + 435, + 468, + 571 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Surface Reconstruction. Interestingly, our approach can be applied to the scenes without emissive sources to enhance surface reconstruction, as evidenced by the lower CD values in Tab. 4 on the DTU dataset. For this experiment, we use Eq. 7 to 10 without our progressive refinement technique. ESR-NeRF's ability to adjust interrelated outgoing radiances helps prevent surface formations where radiances cannot be produced, considering the predicted scene context. Additional visualizations of the normals, BRDF, and environment maps are provided in Appendix.", + "bbox": [ + 75, + 578, + 468, + 729 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/6a30198e34f9cbaf226af7585a64625f99845035691aac99a7632a201aee4000.jpg", + "image_caption": [ + "Figure 11. An example of illumination decomposition." + ], + "image_footnote": [], + "bbox": [ + 78, + 738, + 468, + 882 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/fd10cc43e36a8259817e68e74a8789ca946a56f4684eac19aa394d68849b77f3.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ScanNeuSVoxurfNeural-PBIRESR-NeRF
240.830.650.570.58
370.980.740.750.71
400.560.390.380.38
550.370.350.360.33
631.130.961.040.93
650.590.640.730.57
690.600.850.650.78
831.451.581.281.18
970.951.010.970.95
1050.780.680.760.58
1060.520.600.530.54
1101.431.110.841.08
1140.360.370.380.33
1180.450.450.460.40
1220.450.470.490.44
mean0.770.720.680.65
", + "bbox": [ + 563, + 75, + 826, + 268 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/465a2083f3cc320eaf1015bec74cbf59919fb2eaa70527bb04f1fc03bc735466.jpg", + "table_caption": [ + "Table 4. Results of surface reconstruction via the Chamfer distance on the DTU dataset. A lower value is better." + ], + "table_footnote": [], + "table_body": "
WhiteVivid
IoU ↑MSE ↓IoU ↑MSE ↓
w/o progressive0.409.920.413.93
w/o sg0.716.450.603.47
ESR-NeRF0.865.240.812.79
", + "bbox": [ + 504, + 310, + 767, + 385 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/d47dabe4d6803dc5a8ff53273871aae0a0776e4b4c4c94310f24ff7f6635e82f.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
DTU\nCD ↓
w/o mθ0.93
w/o LTS0.71
ESR-NeRF0.65
", + "bbox": [ + 777, + 311, + 885, + 383 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 5. Ablation studies on the surface reconstruction (left) and the emissive source reconstruction (right).", + "bbox": [ + 498, + 388, + 890, + 417 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 429, + 617, + 444 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We present ESR-NeRF as the first NeRF-based inverse rendering method for the scenes with emissive sources. Our approach uses LDR images, eliminating the need of HDR images to reconstruct emissive sources. Furthermore, we demonstrate the application of reconstructed sources in scene editing, enabling color and intensity modifications.", + "bbox": [ + 496, + 454, + 890, + 544 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Limitations. Future work could explore using a single lighting condition to disentangle emissive sources, environmental lighting, and object texture. It is also promising to address the challenge of volume ray tracing in unbounded scenes to extend to indoor scenes. Additionally, LTS based re-lighting may be weak in representing new colors that traverse unobserved light paths during training. An alternative approach could be extracting emission texture maps and modifying it using the engines such as Blender [13] or Mitsuba [22]. More details on alternative re-lighting methods and radiance fine-tuning are provided in Appendix.", + "bbox": [ + 496, + 545, + 890, + 710 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "7. Acknowledgements", + "text_level": 1, + "bbox": [ + 500, + 724, + 687, + 741 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This work was supported by Samsung Electronics MX, Basic Science Research Program through the National Research Foundation of Korea(NRF) funded by the Ministry of Education(RS-2023-00274280), and Institute of Information & Communications Technology Planning & Evaluation (IITP) grant funded by the Korea government (MSIT) (No. 2019-0-01082, SW StarLab; No. 2022-0-00156, Fundamental research on continual meta-learning for quality enhancement of casual videos and their 3D metaverse transformation). Gunhee Kim is the corresponding author.", + "bbox": [ + 496, + 750, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "4605", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Benjamin Attal, Jia-Bin Huang, Michael Zollhöfer, Johannes Kopf, and Changil Kim. Learning neural light fields with ray-space embedding. In CVPR, 2022. 2", + "[2] Dejan Azinovic, Tzu-Mao Li, Anton Kaplanyan, and Matthias Nießner. Inverse path tracing for joint material and lighting estimation. In CVPR, 2019. 2", + "[3] Jonathan T Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P Srinivasan. Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In ICCV, 2021. 1", + "[4] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Mip-nerf 360: Unbounded anti-aliased neural radiance fields. In CVPR, 2022. 1", + "[5] Jonathan T. Barron, Ben Mildenhall, Dor Verbin, Pratul P. Srinivasan, and Peter Hedman. Zip-nerf: Anti-aliased grid-based neural radiance fields. In ICCV, 2023. 2", + "[6] Sai Bi, Zexiang Xu, Pratul Srinivasan, Ben Mildenhall, Kalyan Sunkavalli, Miloš Hasan, Yannick Hold-Geoffroy, David Kriegman, and Ravi Ramamoorthi. Neural reflectance fields for appearance acquisition. In arXiv, 2020. 1", + "[7] Mark Boss, Varun Jampani, Kihwan Kim, Hendrik Lensch, and Jan Kautz. Two-shot spatially-varying brdf and shape estimation. In CVPR, 2020.", + "[8] Mark Boss, Raphael Braun, Varun Jampani, Jonathan T. Barron, Ce Liu, and Hendrik P.A. Lensch. Nerd: Neural reflectance decomposition from image collections. In ICCV, 2021. 1", + "[9] Mark Boss, Varun Jampani, Raphael Braun, Ce Liu, Jonathan T. Barron, and Hendrik P.A. Lensch. Neural-pil: Neural pre-integrated lighting for reflectance decomposition. In NeurIPS, 2021. 2", + "[10] Brent Burley and Walt Disney Animation Studios. Physically-based shading at disney. In SIGGRAPH, 2012. 5", + "[11] Bowen Cai, Jinchi Huang, Rongfei Jia, Chengfei Lv, and Huan Fu. Neuda: Neural deformable anchor for high-fidelity implicit surface reconstruction. In CVPR, 2023. 2", + "[12] Anpei Chen, Zexiang Xu, Andreas Geiger, Jingyi Yu, and Hao Su. Tensorf: Tensorial radiance fields. In ECCV, 2022. 2", + "[13] Blender Online Community. Blender - a 3D modelling and rendering package. Blender Foundation, Stichting Blender Foundation, Amsterdam, 2018. 8", + "[14] Sara Fridovich-Keil, Alex Yu, Matthew Tancik, Qinhong Chen, Benjamin Recht, and Angjoo Kanazawa. Plenoxels: Radiance fields without neural networks. In CVPR, 2022. 2", + "[15] Mathieu Garon, Kalyan Sunkavalli, Sunil Hadap, Nathan Carr, and Jean-Francois Lalonde. Fast spatially-varying indoor lighting estimation. In CVPR, 2019. 1", + "[16] Wenhang Ge, Tao Hu, Haoyu Zhao, Shu Liu, and Ying-Cong Chen. Ref-neus: Ambiguity-reduced neural implicit surface learning for multi-view reconstruction with reflection. In ICCV, 2023. 1", + "[17] Saeed Hadadan, Shuhong Chen, and Matthias Zwicker. Neural radiosivity. In ACM TOG, 2021. 2" + ], + "bbox": [ + 78, + 114, + 467, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[18] Saeed Hadadan, Geng Lin, Jan Novák, Fabrice Rousselle, and Matthias Zwicker. Inverse global illumination using a neural radiometric prior. In SIGGRAPH Conference Proceedings, 2023. 1", + "[19] Bjoern Haefner, Simon Green, Alan Oursland, Daniel Andersen, Michael Goesele, Daniel Cremers, Richard Newcombe, and Thomas Whelan. Recovering real-world reflectance properties and shading from hdr imagery. In 3DV, 2021. 2", + "[20] Jon Hasselgren, Nikolai Hofmann, and Jacob Munkberg. Shape, light, and material decomposition from images using monte carlo rendering and denoising. In NeurIPS, 2022. 2", + "[21] IEC. IEC 61966-2-1:1999. Technical report, International Electrotechnical Commission, 1999. 3", + "[22] Wenzel Jakob, Sébastien Speierer, Nicolas Roussel, Merlin Nimier-David, Delio Vicini, Tizian Zeltner, Baptiste Nicolet, Miguel Crespo, Vincent Leroy, and Ziyi Zhang. Mitsuba 3 renderer, 2022. https://mitsuba-renderer.org. 8", + "[23] Rasmus Jensen, Anders Dahl, George Vogiatzis, Engil Tola, and Henrik Aanæs. Large scale multi-view stereopsis evaluation. In CVPR, 2014. 2, 6", + "[24] Chaonan Ji, Tao Yu, Kaiwen Guo, Jingxin Liu, and Yebin Liu. Geometry-aware single-image full-body human relighting. In ECCV, 2022. 1", + "[25] Haian Jin, Isabella Liu, Peijia Xu, Xiaoshuai Zhang, Songfang Han, Sai Bi, Xiaowei Zhou, Zexiang Xu, and Hao Su. Tensoir: Tensorial inverse rendering. In CVPR, 2023. 2, 3, 6", + "[26] James T. Kajiya. The rendering equation. In SIGGRAPH, 1986. 3", + "[27] Petr Kellnhofer, Lars C Jebe, Andrew Jones, Ryan Spicer, Kari Pulli, and Gordon Wetzstein. Neural lumigraph rendering. In CVPR, 2021. 1", + "[28] Zhengfei Kuang, Fujun Luan, Sai Bi, Zhixin Shu, Gordon Wetzstein, and Kalyan Sunkavalli. Palettenerf: Palette-based appearance editing of neural radiance fields. In CVPR, 2023. 6", + "[29] Zhengqin Li, Mohammad Shafiei, Ravi Ramamoorthi, Kalyan Sunkavalli, and Manmohan Chandraker. Inverse rendering for complex indoor scenes: Shape, spatially-varying lighting and svbrdf from a single image. In CVPR, 2020. 1", + "[30] Zhengqin Li, Jia Shi, Sai Bi, Rui Zhu, Kalyan Sunkavalli, Miloš Hašan, Zexiang Xu, Ravi Ramamoorthi, and Manmohan Chandraker. Physically-based editing of indoor scene lighting from a single image. In ECCV, 2022. 1", + "[31] Zhaoshuo Li, Thomas Müller, Alex Evans, Russell H Taylor, Mathias Unberath, Ming-Yu Liu, and Chen-Hsuan Lin. Neuralangelo: High-fidelity neural surface reconstruction. In CVPR, 2023. 2", + "[32] Zhen Li, Lingli Wang, Mofang Cheng, Cihui Pan, and Jiaqi. Yang. Multi-view inverse rendering for large-scale real-world indoor scenes. In CVPR, 2023. 2", + "[33] Ruofan Liang, Huiting Chen, Chunlin Li, Fan Chen, Selvakumar Panneer, and Nandita Vijaykumar. Envidr: Implicit differentiable renderer with neural environment lighting. In ICCV, 2023. 1" + ], + "bbox": [ + 501, + 92, + 890, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "4606", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[34] Fujun Luan, Shuang Zhao, Kavita Bala, and Zhao Dong. Unified shape and svbrdf recovery using differentiable monte carlo rendering. In EGSR, 2021. 1", + "[35] Linjie Lyu, Ayush Tewari, Thomas Leimkuehler, Marc Habermann, and Christian Theobalt. Neural radiance transfer fields for relightable novel-view synthesis with global illumination. In ECCV, 2022. 2", + "[36] Linjie Lyu, Ayush Tewari, Marc Habermann, Shunsuke Saito, Michael Zollhöfer, Thomas Leimkuehler, and Christian Theobalt. Diffusion posterior illumination for ambiguity-aware inverse rendering. In ACM TOG, 2023. 1", + "[37] Alexander Mai, Dor Verbin, Falko Kuester, and Sara Fridovich-Keil. Neural microfacet fields for inverse rendering. In ICCV, 2023. 1, 3", + "[38] Ricardo Martin-Brualla, Noha Radwan, Mehdi S. M. Sajjadi, Jonathan T. Barron, Alexey Dosovitskiy, and Daniel Duckworth. Nerf in the wild: Neural radiance fields for unconstrained photo collections. In CVPR, 2021. 1, 6", + "[39] Nelson Max. Optical models for direct volume rendering. IEEE Transactions on Visualization and Computer Graphics, 1(2):99-108, 1995. 2", + "[40] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In ECCV, 2020. 1, 2", + "[41] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. In ACM TOG, 2022. 2", + "[42] Jacob Munkberg, Jon Hasselgren, Tianchang Shen, Jun Gao, Wenzheng Chen, Alex Evans, Thomas Müller, and Sanja Fidler. Extracting triangular 3d models, materials, and lighting from images. In CVPR, 2022. 2", + "[43] Michael Niemeyer, Lars Mescheder, Michael Oechsle, and Andreas Geiger. Differentiable volumetric rendering: Learning implicit 3d representations without 3d supervision. In CVPR, 2020. 1", + "[44] Michael Oechsle, Songyou Peng, and Andreas Geiger. Unisurf: Unifying neural implicit surfaces and radiance fields for multi-view reconstruction. In ICCV, 2021. 2", + "[45] Julian Ost, Issam Laradji, Alejandro Newell, Yuval Bahat, and Felix Heide. Neural point light fields. In CVPR, 2022. 1", + "[46] Rohit Pandey, Sergio Orts-Escalano, Chloe LeGendre, Christian Haene, Sofien Bouaziz, Christoph Rhemann, Paul Debevec, and Seann Fanello. Total relighting: Learning to relight portraits for background replacement. In ACM TOG, 2021. 2", + "[47] Jeong Joon Park, Peter Florence, Julian Straub, Richard Newcombe, and Steven Lovegrove. Deepsdf: Learning continuous signed distance functions for shape representation. In CVPR, 2019. 1", + "[48] Julien Philip, Sébastien Morgenthaler, Michael Gharbi, and George Drettakis. Free-viewpoint indoor neural relighting from multi-view stereo. In ACM TOG, 2021. 2", + "[49] Martin Piala and Ronald Clark. Terminerf: Ray termination prediction for efficient neural rendering. In 3DV, 2021. 2", + "[50] Viktor Rudnev, Mohamed Elgharib, William Smith, Lingjie Liu, Vladislav Golyanik, and Christian Theobalt. Nerf for outdoor scene relighting. In ECCV, 2022. 1" + ], + "bbox": [ + 78, + 90, + 468, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[51] Vincent Sitzmann, Julien N.P. Martel, Alexander W. Bergman, David B. Lindell, and Gordon Wetzstein. Implicit neural representations with periodic activation functions. In NeurIPS, 2020. 2", + "[52] Vincent Sitzmann, Semon Rezchikov, Bill Freeman, Josh Tenenbaum, and Fredo Durand. Light field networks: Neural scene representations with single-evaluation rendering. In NeurIPS, 2021. 2", + "[53] Alvy Ray Smith. Color gamut transform pairs. In ACM TOG, 1978. 5", + "[54] Pratul P. Srinivasan, Boyang Deng, Xiuming Zhang, Matthew Tancik, Ben Mildenhall, and Jonathan T. Barron. Nerv: Neural reflectance and visibility fields for relighting and view synthesis. In CVPR, 2021. 2, 3", + "[55] Mohammed Suhail, Carlos Esteves, Leonid Sigal, and Ameesh Makadia. Light field neural rendering. In CVPR, 2022. 2", + "[56] Cheng Sun, Min Sun, and Hwann-Tzong Chen. Direct voxel grid optimization: Super-fast convergence for radiance fields reconstruction. In CVPR, 2022. 2", + "[57] Cheng Sun, Guangyan Cai, Zhengqin Li, Kai Yan, Cheng Zhang, Carl Marshall, Jia-Bin Huang, Shuang Zhao, and Zhao Dong. Neural-pbir reconstruction of shape, material, and illumination. In ICCV, 2023. 2, 3, 6", + "[58] Jiaming Sun, Xi Chen, Qianqian Wang, Zhengqi Li, Hadar Averbuch-Elor, Xiaowei Zhou, and Noah Snively. Neural 3D reconstruction in the wild. In SIGGRAPH Conference Proceedings, 2022. 1", + "[59] Tiancheng Sun, Kai-En Lin, Sai Bi, Zexiang Xu, and Ravi Ramamoorthi. Nelf: Neural light-transport field for portrait view synthesis and relighting. In EGSR, 2021. 2", + "[60] Matthew Tancik, Pratul P. Srinivasan, Ben Mildenhall, Sara Fridovich-Keil, Nithin Raghavan, Utkarsh Singhal, Ravi Ramamoorthi, Jonathan T. Barron, and Ren Ng. Fourier features let networks learn high frequency functions in low dimensional domains. In NeurIPS, 2020. 2", + "[61] Jiaxiang Tang, Hang Zhou, Xiaokang Chen, Tianshu Hu, Errui Ding, Jingdong Wang, and Gang Zeng. Delicate textured mesh recovery from nef via adaptive surface refinement. In ICCV, 2023. 1", + "[62] Ayush Tewari, Ohad Fried, Justus Thies, Vincent Sitzmann, Stephen Lombardi, Kalyan Sunkavalli, Ricardo Martin-Brualla, Tomas Simon, Jason Saragih, Matthias Nießner, et al. State of the art on neural rendering. In CGF, 2020. 1", + "[63] Ayush Tewari, Justus Thies, Ben Mildenhall, Pratul Srinivasan, Edgar Tretschk, Wang Yifan, Christoph Lassner, Vincent Sitzmann, Ricardo Martin-Brualla, Stephen Lombardi, et al. Advances in neural rendering. In CGF, 2022. 1", + "[64] Marco Toschi, Riccardo De Matteo, Riccardo Spezialetti, Daniele De Gregorio, Luigi Di Stefano, and Samuele Salti. Relight my nerf: A dataset for novel view synthesis and relighting of real world objects. In CVPR, 2023. 2", + "[65] Dor Verbin, Peter Hedman, Ben Mildenhall, Todd Zickler, Jonathan T Barron, and Pratul P Srinivasan. Ref-nerf: Structured view-dependent appearance for neural radiance fields. In CVPR, 2022. 1" + ], + "bbox": [ + 501, + 90, + 890, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "4607", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[66] Dongqing Wang, Tong Zhang, and Sabine Susstrunk. Nemto: Neural environment matting for novel view and relighting synthesis of transparent objects. In ICCV, 2023. 1", + "[67] Jiaping Wang, Peiran Ren, Minmin Gong, John Snyder, and Baining Guo. All-frequency rendering of dynamic, spatially-varying reflectance. ACM TOG, 2009. 4", + "[68] Peng Wang, Lingjie Liu, Yuan Liu, Christian Theobalt, Taku Komura, and Wenping Wang. Neus: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. In NeurIPS, 2021. 2, 6", + "[69] Shaofei Wang, Katja Schwarz, Andreas Geiger, and Siyu Tang. Arah: Animatable volume rendering of articulated human sdfs. In ECCV, 2022. 1", + "[70] Yiming Wang, Qin Han, Marc Habermann, Kostas Dani-ilidis, Christian Theobalt, and Lingjie Liu. Neus2: Fast learning of neural implicit surfaces for multi-view reconstruction. In ICCV, 2023. 2", + "[71] Zian Wang, Jonah Philion, Sanja Fidler, and Jan Kautz. Learning indoor inverse rendering with 3d spatially-varying lighting. In ICCV, 2021. 1", + "[72] Zian Wang, Wenzheng Chen, David Acuna, Jan Kautz, and Sanja Fidler. Neural light field estimation for street scenes with differentiable virtual object insertion. In ECCV, 2022. 1", + "[73] Zian Wang, Tianchang Shen, Jun Gao, Shengyu Huang, Jacob Munkberg, Jon Hasselgren, Zan Gojcic, Wenzheng Chen, and Sanja Fidler. Neural fields meet explicit geometric representations for inverse rendering of urban scenes. In CVPR, 2023. 1", + "[74] Henrique Weber, Mathieu Garon, and Jean-François Lalonde. Editable indoor lighting estimation. In ECCV, 2022. 1", + "[75] Haoqian Wu, Zhipeng Hu, Lincheng Li, Yongqiang Zhang, Changjie Fan, and Xin Yu. Nefii: Inverse rendering for reflectance decomposition with near-field indirect illumination. In CVPR, 2023. 2", + "[76] Liwen Wu, Rui Zhu, Mustafa B. Yaldiz, Yinhao Zhu, Hong Cai, Janarbek Matai, Fatih Porikli, Tzu-Mao Li, Manmohan Chandraker, and Ravi Ramamoorthi. Factorized inverse path tracing for efficient and accurate material-lighting estimation. In ICCV, 2023. 2", + "[77] Qianyi Wu, Xian Liu, Yuedong Chen, Kejie Li, Chuanxia Zheng, Jianfei Cai, and Jianmin Zheng. Object-compositional neural implicit surfaces. In ECCV, 2022. 2", + "[78] Qianyi Wu, Kaisiyuan Wang, Kejie Li, Jianmin Zheng, and Jianfei Cai. Objectsdf++: Improved object-compositional neural implicit surfaces. In ICCV, 2023. 2", + "[79] Tong Wu, Jia-Mu Sun, Yu-Kun Lai, and Lin Gao. De-nerf: Decoupled neural radiance fields for view-consistent appearance editing and high-frequency environmental relighting. In SIGGRAPH ASIA Conference Proceedings, 2023. 1", + "[80] Tong Wu, Jiaqi Wang, Xingang Pan, Xudong Xu, Christian Theobalt, Ziwei Liu, and Dahua Lin. Voxurf: Voxel-based efficient and accurate neural surface reconstruction. In ICLR, 2023. 2, 5, 6", + "[81] Yuanbo Xiangli, Linning Xu, Xingang Pan, Nanxuan Zhao, Anyi Rao, Christian Theobalt, Bo Dai, and Dahua Lin." + ], + "bbox": [ + 78, + 90, + 468, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Bungeenerf: Progressive neural radiance field for extreme multi-scale scene rendering. In ECCV, 2022. 1", + "[82] Yingyan Xu, Gaspard Zoss, Prashanth Chandran, Markus Gross, Derek Bradley, and Paulo Gotardo. Renerf: Relightable neural radiance fields with nearfield lighting. In ICCV, 2023. 2", + "[83] Wenqi Yang, Guanying Chen, Chaofeng Chen, Zhenfang Chen, and Kwan-Yee K. Wong. Ps-nerf: Neural inverse rendering for multi-view photometric stereo. In ECCV, 2022. 2", + "[84] Wenqi Yang, Guanying Chen, Chaofeng Chen, Zhenfang Chen, and Kwan-Yee K. Wong. S $^3$ -nerf: Neural reflectance field from shading and shadow under a single viewpoint. In NeurIPS, 2022. 2", + "[85] Yao Yao, Jingyang Zhang, Jingbo Liu, Yihang Qu, Tian Fang, David McKinnon, Yanghai Tsin, and Long Quan. Neilf: Neural incident light field for material and lighting estimation. In ECCV, 2022. 1", + "[86] Lior Yariv, Yoni Kasten, Dror Moran, Meirav Galun, Matan Atzmon, Basri Ronen, and Yaron Lipman. Multiview neural surface reconstruction by disentangling geometry and appearance. In NeurIPS, 2020. 1", + "[87] Lior Yariv, Jiatao Gu, Yoni Kasten, and Yaron Lipman. Volume rendering of neural implicit surfaces. In NeurIPS, 2021. 2", + "[88] Alex Yu, Ruilong Li, Matthew Tancik, Hao Li, Ren Ng, and Angjoo Kanazawa. PlenOctrees for real-time rendering of neural radiance fields. In ICCV, 2021. 2", + "[89] Chong Zeng, Guojun Chen, Yue Dong, Pieter Peers, Hongzhi Wu, and Xin Tong. Relighting neural radiance fields with shadow and highlight hints. In SIGGRAPH Conference Proceedings, 2023. 2", + "[90] Jingyang Zhang, Yao Yao, Shiwei Li, Jingbo Liu, Tian Fang, David McKinnon, Yanghai Tsin, and Long Quan. Neilf++: Inter-reflectable light fields for geometry and material estimation. In ICCV, 2023. 2, 6", + "[91] Kai Zhang, Gernot Riegler, Noah Snavely, and Vladlen Koltun. Nerf++: Analyzing and improving neural radiance fields. arXiv preprint arXiv:2010.07492, 2020. 1", + "[92] Kai Zhang, Fujun Luan, Qianqian Wang, Kavita Bala, and Noah Snavely. Physg: Inverse rendering with spherical gaussians for physics-based material editing and relighting. In CVPR, 2021. 2", + "[93] Kai Zhang, Fujun Luan, Zhengqi Li, and Noah Snavely. Iron: Inverse rendering by optimizing neural sdfs and materials from photometric images. In CVPR, 2022. 1", + "[94] Xiuming Zhang, Sean Fanello, Yun-Ta Tsai, Tiancheng Sun, Tianfan Xue, Rohit Pandey, Sergio Orts-Escalano, Philip Davidson, Christoph Rhemann, Paul Debevec, et al. Neural light transport for relighting and view synthesis. In ACM TOG, 2021. 2", + "[95] Xiuming Zhang, Pratul P. Srinivasan, Boyang Deng, Paul Debevec, William T. Freeman, and Jonathan T. Barron. Ner-factor: Neural factorization of shape and reflectance under an unknown illumination. In ACM TOG, 2021. 2", + "[96] Yuanqing Zhang, Jiaming Sun, Xingyi He, Huan Fu, Rongfei Jia, and Xiaowei Zhou. Modeling indirect illumination for inverse rendering. In CVPR, 2022. 2" + ], + "bbox": [ + 503, + 92, + 890, + 900 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "4608", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[97] Y. Zhang, Z. Hu, H. Wu, M. Zhao, L. Li, Z. Zou, and C. Fan. Towards unbiased volume rendering of neural implicit surfaces with geometry priors. In CVPR, 2023. 2", + "[98] Quan Zheng, Gurprit Singh, and Hans-Peter Seidel. Neural relightable participating media rendering. In NeurIPS, 2021. 2", + "[99] Jingsen Zhu, Fujun Luan, Yuchi Huo, Zihao Lin, Zhihua Zhong, Dianbing Xi, Rui Wang, Hujun Bao, Jiaxiang Zheng, and Rui Tang. Learning-based inverse rendering of complex indoor scenes with differentiable monte carlo raytracing. In SIGGRAPH ASIA Conference Proceedings, 2022. 2" + ], + "bbox": [ + 78, + 90, + 468, + 247 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "4609", + "bbox": [ + 482, + 945, + 516, + 955 + ], + "page_idx": 11 + } +] \ No newline at end of file diff --git a/2024/ESR-NeRF_ Emissive Source Reconstruction Using LDR Multi-view Images/2c6d481f-7a23-410c-a62a-93255bbc656e_model.json b/2024/ESR-NeRF_ Emissive Source Reconstruction Using LDR Multi-view Images/2c6d481f-7a23-410c-a62a-93255bbc656e_model.json new file mode 100644 index 0000000000000000000000000000000000000000..48352d86830c6e63dcd0c1733ac451a3798a1ecd --- /dev/null +++ b/2024/ESR-NeRF_ Emissive Source Reconstruction Using LDR Multi-view Images/2c6d481f-7a23-410c-a62a-93255bbc656e_model.json @@ -0,0 +1,3271 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.808, + 0.045 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.104, + 0.131, + 0.867, + 0.152 + ], + "angle": 0, + "content": "ESR-NeRF: Emissive Source Reconstruction Using LDR Multi-view Images" + }, + { + "type": "text", + "bbox": [ + 0.23, + 0.18, + 0.747, + 0.218 + ], + "angle": 0, + "content": "Jinseo Jeong\\(^{1}\\) Junseo Koo\\(^{1}\\) Qimeng Zhang\\(^{2}\\) Gunhee Kim\\(^{1}\\) Seoul National University Korea University" + }, + { + "type": "text", + "bbox": [ + 0.088, + 0.22, + 0.893, + 0.252 + ], + "angle": 0, + "content": "jinseo@vision.snu.ac.kr, junseo.koo@vision.snu.ac.kr, zoe1024@korea.ac.kr, gunhee@snu.ac.kr \nhttps://jinseo.kr/ESR-NeRF" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.287, + 0.314, + 0.303 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.319, + 0.474, + 0.622 + ], + "angle": 0, + "content": "Existing NeRF-based inverse rendering methods suppose that scenes are exclusively illuminated by distant light sources, neglecting the potential influence of emissive sources within a scene. In this work, we confront this limitation using LDR multi-view images captured with emissive sources turned on and off. Two key issues must be addressed: 1) ambiguity arising from the limited dynamic range along with unknown lighting details, and 2) the expensive computational cost in volume rendering to backtrace the paths leading to final object colors. We present a novel approach, ESR-NeRF, leveraging neural networks as learnable functions to represent ray-traced fields. By training networks to satisfy light transport segments, we regulate outgoing radiances, progressively identifying emissive sources while being aware of reflection areas. The results on scenes encompassing emissive sources with various properties demonstrate the superiority of ESR-NeRF in qualitative and quantitative ways. Our approach also extends its applicability to the scenes devoid of emissive sources, achieving lower CD metrics on the DTU dataset." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.648, + 0.21, + 0.664 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.675, + 0.47, + 0.902 + ], + "angle": 0, + "content": "Extensive research has focused on reconstructing 3D object structures [16, 43, 47, 86], material properties [18, 29, 65], and lighting [15, 33, 34, 74, 79] from 2D images, applicable across domains including 3D graphics and augmented reality [62, 63, 69, 72]. This endeavor not only facilitates the creation of life-like virtual objects but also streamlines the process of scene manipulation [27, 58, 61, 73]. Recent advancements [24, 30, 36, 71] have built on Neural Radiance Fields (NeRF) [40] successes in novel view synthesis [3, 4, 45, 81, 91]. Significant progress in relighting [37, 38, 50] has facilitated scene editing via manipulating the reconstructed light sources. However, existing methods predominantly deal with the scenes lit by distant sources, like environment maps or collocated flashlights. Notably, NeRF-based inverse rendering has yet to" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.285, + 0.574, + 0.358 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.504, + 0.36, + 0.572, + 0.369 + ], + "angle": 0, + "content": "255 pixel values" + }, + { + "type": "image", + "bbox": [ + 0.576, + 0.286, + 0.646, + 0.358 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.576, + 0.36, + 0.644, + 0.369 + ], + "angle": 0, + "content": "Information loss" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.375, + 0.573, + 0.444 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.507, + 0.447, + 0.57, + 0.457 + ], + "angle": 0, + "content": "Threshold 0.65" + }, + { + "type": "image", + "bbox": [ + 0.576, + 0.375, + 0.646, + 0.444 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.58, + 0.447, + 0.642, + 0.457 + ], + "angle": 0, + "content": "Threshold 0.75" + }, + { + "type": "image", + "bbox": [ + 0.647, + 0.286, + 0.675, + 0.358 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.647, + 0.36, + 0.649, + 0.369 + ], + "angle": 0, + "content": "" + }, + { + "type": "image", + "bbox": [ + 0.671, + 0.286, + 0.811, + 0.456 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.823, + 0.287, + 0.881, + 0.296 + ], + "angle": 0, + "content": "Edit emission" + }, + { + "type": "image", + "bbox": [ + 0.819, + 0.3, + 0.888, + 0.365 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.836, + 0.366, + 0.87, + 0.374 + ], + "angle": 0, + "content": "Desired" + }, + { + "type": "image", + "bbox": [ + 0.819, + 0.378, + 0.888, + 0.441 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.829, + 0.443, + 0.878, + 0.452 + ], + "angle": 0, + "content": "Unrealistic" + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.469, + 0.893, + 0.58 + ], + "angle": 0, + "content": "Figure 1. Challenges posed by emissive sources in LDR images. Green, red, and blue in thresholded images respectively show true positives, false negatives, and false positives of source identification. Thresholding values are scaled down divided by 255. The contrast between light on and off pixel values is more pronounced in surroundings than emissive sources. Inaccurate reconstruction of emissive sources disrupts scene editing, causing reflection areas to stay static while only the source colors change." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.594, + 0.892, + 0.623 + ], + "angle": 0, + "content": "consider scenes with multiple emissive sources, a common real-world illumination condition." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.627, + 0.893, + 0.882 + ], + "angle": 0, + "content": "Emissive sources in a scene introduce critical challenges: (i) ambiguity in decomposing scene components and (ii) high computational costs for analyzing the causes of pixel colors. This ambiguity stems from difficulties in identifying emissive source regions, as illustrated in Fig. 1. Contrary to prior setups [6-8, 66, 85, 93], we allow the possibility of numerous emissive sources throughout the scene. In standard photographs with pixel values from 0 to 255, the distinction between emissive sources and nearby reflection areas is challenging. As shown in Fig. 1, relying solely on pixel value thresholding is insufficient for differentiating between emissive sources and their reflections. Naive inverse path tracing is impractical, due to the computational costs rising exponentially with the number of ray bounces in volume rendering. This can cause inaccuracy in emissive source reconstruction, yielding unrealistic illumination in reflective areas as users manipulate emissive sources." + }, + { + "type": "text", + "bbox": [ + 0.52, + 0.886, + 0.892, + 0.901 + ], + "angle": 0, + "content": "To address these challenges, we introduce ESR-NeRF" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4598" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.288 + ], + "angle": 0, + "content": "(Emissive Sources Reconstructing NeRF), a novel approach capable of reconstructing any number of emissive sources by progressively discovering reflection areas. We assume that the scenes are observed in two lighting conditions: one with all emissive sources active and the other with them inactive. Our approach utilizes neural networks as learnable functions for representing ray-traced fields. By training networks to satisfy each light transport segment, we sidestep the computational overhead of ray tracing associated with ray bounces. In this work, we exclusively use low dynamic range (LDR) images, setting us apart from prior mesh-based methods that rely on high dynamic range (HDR) images [2, 19, 48, 76]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.289, + 0.473, + 0.422 + ], + "angle": 0, + "content": "Our experiments encompass synthetic and real scenes, ranging from single to multiple lighting configurations with complex reflections. The scenes vary in light source counts, color, and intensity. Qualitative and quantitative evaluations show ESR-NeRF's superiority over state-of-the-art NeRF-based re-lighting methods. Furthermore, Chamfer Distance (CD) metrics on the DTU dataset [23] indicate ESR-NeRF's competitive performance in scene reconstruction, even without emissive sources." + }, + { + "type": "text", + "bbox": [ + 0.098, + 0.424, + 0.388, + 0.438 + ], + "angle": 0, + "content": "We summarize our contributions as follows." + }, + { + "type": "text", + "bbox": [ + 0.079, + 0.44, + 0.47, + 0.498 + ], + "angle": 0, + "content": "1. Our work presents the first NeRF-based inverse rendering that can deal with the scenes with any number of emissive sources, challenging the distant light assumption of previous research." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.5, + 0.469, + 0.544 + ], + "angle": 0, + "content": "2. Unlike existing mesh-based methods relying on HDR images, we use LDR images for the first time, overcoming the poor representation of emissive sources." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.545, + 0.469, + 0.574 + ], + "angle": 0, + "content": "3. We provide a benchmark dataset designed to evaluate the performance of emissive source reconstruction." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.576, + 0.469, + 0.619 + ], + "angle": 0, + "content": "4. Our method is applicable to the scenes with or without emissive sources, achieving superior mesh reconstruction results on the DTU dataset." + }, + { + "type": "list", + "bbox": [ + 0.078, + 0.44, + 0.47, + 0.619 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.634, + 0.216, + 0.649 + ], + "angle": 0, + "content": "2. Related work" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.659, + 0.473, + 0.903 + ], + "angle": 0, + "content": "Neural Rendering. Advancements in implicit representations [51, 60] and volume rendering [39] have significantly enhanced neural rendering capabilities, enabling the reconstruction of scene components from 2D images. One of the key directions is mesh extraction [44, 70, 77, 78, 80, 97], with methods like NeuS [68] and VolSDF [87] utilizing signed distance function (SDF) values for volume rendering. Recently, the efficient computation of volume rendering has become a focal point due to the substantial computational cost associated with network inference for ray color calculation [41, 49, 88]. Several methods propose to directly predict ray color using the 4D light fields concept [1, 52, 55] or leveraging voxel grids for fast inference of spatial features [5, 11, 12, 14, 31, 56]. NeuralRadiosity [17] shares similarity with our method, as it predicts raytraced values instead of explicitly tracing individual rays." + }, + { + "type": "table", + "bbox": [ + 0.502, + 0.09, + 0.895, + 0.158 + ], + "angle": 0, + "content": "
VoxurfTensoIRPath TracingESR-NeRF
Big Onn · d(n · d)b+1n2 · d
Indirect illuminationx✓✓✓
BRDF decompositionx✓✓✓
Emissive source controlxx✓✓
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.168, + 0.892, + 0.21 + ], + "angle": 0, + "content": "Table 1. Computational cost comparison for inverse rendering methods. \\( n \\) is the number of sampled points along a ray, \\( d \\) is the number of scattering rays, and \\( b \\) is the number of ray bounces." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.219, + 0.892, + 0.295 + ], + "angle": 0, + "content": "However, they primarily focus on calculating the final object color when all scene information is available. In contrast, our inverse rendering approach aims to reconstruct emissive sources within a scene, addressing the ambiguities introduced by their presence in LDR images." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.301, + 0.892, + 0.816 + ], + "angle": 0, + "content": "Inverse Rendering. A growing emphasis revolves around the decomposition of materials represented by spatially varying bidirectional reflectance distribution functions (SVBRDF) [46, 83, 99]. To lessen the computational burden in inverse rendering [25, 54, 96, 98], several methods have adopted neural networks as lookup tables [9] or computational caches [54, 90, 95]. While NeRV [54] utilizes caching visibility and \\(\\mathrm{NeILF++}\\) [90] adopts caching surface point radiance with the inter-reflection loss for incident radiance, our method diverges by focusing on tracing radiance origins. Specifically, we aim to identify emissive sources within a scene, moving beyond the simplification of incident radiance calculations. Several methods rely on diverse known lighting configurations to exploit variations in object appearances [59, 64, 84, 89]. Toggling emissive sources on and off resembles the common one-light-at-a-time (OLAT) technique, as seen in NLT [94] and ReN-eRF [82]. However, our setting does not need to know light source properties and to toggle lights individually. Instead, we allow for toggling all lights together. Recent works have also jointly reconstruct the mesh, materials, and lighting [20, 35, 42, 57]. They tackle with images captured under a single unknown lighting condition [92, 95], assuming that radiance already encodes global illumination [75, 96]. However, they confine to the scenes illuminated by far-distant lights, constrained to an 8-bit color spectrum. Our work considers the presence of multiple emissive sources within a scene captured in LDR images, questioning the prevailing notion that radiance fields trained with the image rendering loss faithfully represents global illumination. While some methods [2, 19, 32, 48, 76] deal with the scenes featuring emissive sources, they work outside the volume rendering framework and depend on HDR input images, assuming prior knowledge of scene geometry." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.83, + 0.64, + 0.846 + ], + "angle": 0, + "content": "3. Preliminaries" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.856, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Surface Representation. Analogous to NeRF [40], neural network \\( f_{\\theta} \\) predicts SDF values at arbitrary 3D spatial locations. NeuS [68] integrates surface representa" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4599" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.081, + 0.089, + 0.175, + 0.102 + ], + "angle": 0, + "content": "Multi-view Input" + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.111, + 0.174, + 0.181 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.105, + 0.184, + 0.157, + 0.194 + ], + "angle": 0, + "content": "On Images" + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.207, + 0.174, + 0.276 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.105, + 0.279, + 0.158, + 0.289 + ], + "angle": 0, + "content": "Off Images" + }, + { + "type": "image_caption", + "bbox": [ + 0.405, + 0.089, + 0.454, + 0.102 + ], + "angle": 0, + "content": "Training" + }, + { + "type": "image", + "bbox": [ + 0.182, + 0.103, + 0.67, + 0.294 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.753, + 0.089, + 0.817, + 0.102 + ], + "angle": 0, + "content": "Application" + }, + { + "type": "image", + "bbox": [ + 0.681, + 0.104, + 0.892, + 0.293 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.306, + 0.894, + 0.35 + ], + "angle": 0, + "content": "Figure 2. The pipeline of emissive source reconstruction. Given LDR images with emissive sources on and off, scene components are reconstructed by synthesizing training images and enforcing LTS requirements. Emissive sources are progressively refined via categorizing training rays into uncertain and certain groups. The scenes can be edited with new lighting conditions using reconstructed emissive sources." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.362, + 0.47, + 0.429 + ], + "angle": 0, + "content": "tion into volume rendering using the SDF-based opacity \\(\\rho (x) = \\max (\\frac{-\\frac{d\\Phi_s}{dx}(f(x))}{\\Phi_s(f(x))},0)\\). Here \\(\\Phi_s(x) = (1 + e^{-sx})^{-1}\\) is the sigmoid function where \\(s\\) controls the sharpness of surfaces. The color of a ray can be calculated as" + }, + { + "type": "equation", + "bbox": [ + 0.136, + 0.435, + 0.47, + 0.466 + ], + "angle": 0, + "content": "\\[\n\\hat {C} (r) = \\int_ {0} ^ {\\infty} T (r (t)) \\rho (r (t)) L _ {o} (r (t), \\omega_ {o}) d t, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.473, + 0.472, + 0.553 + ], + "angle": 0, + "content": "where \\(\\hat{C}(r)\\) denotes the predicted ray color, \\(r(t; c, \\omega_o) = c - t \\cdot \\omega_o\\) is the ray with camera center \\(c\\) along direction \\(\\omega_o\\), \\(T(r(t)) = \\exp \\left( \\int_0^t -\\rho (r(u))du \\right)\\) is the transmittance, and \\(L_{o}(r(t), \\omega_{o})\\) is the outgoing radiance. Henceforth, we use \\(x\\) to denote a point in \\(r(t; c, \\omega_o)\\) for notational simplicity." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.557, + 0.47, + 0.618 + ], + "angle": 0, + "content": "Light Transport in Volume Rendering. Extracting light sources necessitates analyzing the causes affecting the final ray colors. Kajiya's rendering equation [26] factorizes the outgoing radiance \\( L_{o}(x,\\omega_{o}) \\) into emission and reflections:" + }, + { + "type": "equation", + "bbox": [ + 0.101, + 0.623, + 0.47, + 0.652 + ], + "angle": 0, + "content": "\\[\nL _ {o} (x, \\omega_ {o}) = E (x) + \\int_ {\\Omega} L _ {i} (x, \\omega_ {i}) R (x, \\omega_ {o}, \\omega_ {i}; b) d \\omega_ {i}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.656, + 0.47, + 0.84 + ], + "angle": 0, + "content": "where \\(E(x)\\) is the emission, \\(R(x,\\omega_o,\\omega_i;b)\\) represents the SVBRDF parametrized by parameters \\(b\\) with Lambert cosine multiplied, and \\(L_{i}(x,\\omega_{i})\\) is the incident radiance. In volume rendering, computing the incident radiance at point \\(x\\) is akin to evaluating Eq. 1, with \\(x\\) serving as the camera center. By iteratively factorizing the outgoing radiance in the incident radiance, the contribution of a path length \\(i\\) for a pixel can be decomposed as in Eq. 3, where \\(\\mathcal{H}_i = \\Pi_{j=1}^{i-1}T(x_j)\\rho(x_j)R(x_j,\\omega_{j-1},\\omega_j)\\) is the path throughput, \\(S(\\omega_i)\\) is the environment map strength in direction \\(\\omega_i\\), and \\(V(x,\\omega_i) = \\exp\\left(\\int_0^\\infty -\\rho(r(u;x_i,-\\omega_i))du\\right)\\) is the visibility of the environment map at point \\(x\\) along direction \\(\\omega_i\\):" + }, + { + "type": "equation", + "bbox": [ + 0.101, + 0.854, + 0.47, + 0.904 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} P _ {i} = \\int_ {l _ {1}} \\int_ {\\Omega} \\dots \\int_ {l _ {i - 1}} \\int_ {\\Omega} \\left(\\int_ {l _ {i}} T (x _ {i}) \\rho (x _ {i}) E (x _ {i}) d t _ {i} + \\right. \\tag {3} \\\\ S \\left(\\omega_ {i - 1}\\right) V \\left(x _ {i - 1}, \\omega_ {i - 1})\\right) \\mathcal {H} _ {i} d t _ {1} d \\omega_ {1} \\dots d t _ {i - 1} d \\omega_ {i - 1}. \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.362, + 0.893, + 0.454 + ], + "angle": 0, + "content": "Extending the analysis to longer light paths, or equivalently, increasing the number of ray bounces, leads to exponential growth in computation complexity. This poses a challenge when attempting to decompose the influence of unknown emissive sources, as their ability to produce strong reflections makes ignoring indirect illumination infeasible." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.466, + 0.636, + 0.484 + ], + "angle": 0, + "content": "4. Methodology" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.491, + 0.893, + 0.598 + ], + "angle": 0, + "content": "None of the previous works address the reconstruction of emissive sources from LDR multi-view images. Sec. § 4.1 through § 4.5 detail our method, ESR-NeRF, which reconstructs emissive sources without prior knowledge of scene geometry, materials, or lighting specifics (including their location, number, or colors). We also show how these reconstructed sources can be used for scene editing in § 4.5." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.606, + 0.724, + 0.623 + ], + "angle": 0, + "content": "4.1. Learnable Tone-mapper" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.629, + 0.892, + 0.674 + ], + "angle": 0, + "content": "Throughout the paper, we use \\(\\mathcal{R}\\) to represent camera rays, \\(C\\) for pixel values, and a binary flag \\(\\mathbb{I}\\) to indicate whether an image is captured with emissive sources on or off." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.675, + 0.893, + 0.901 + ], + "angle": 0, + "content": "To extract HDR values from LDR images, we employ the softplus activation for outgoing radiance prediction and apply a clipping and gamma function \\(\\tau\\) [21] for the rendering loss such that \\(\\hat{C}_{\\tau}(r) = \\tau (\\hat{C} (r))\\). Unlike previous NeRF-based works [25, 37, 54, 57] that limit radiance to the range of [0, 1], our approach allows for any positive radiance values. Yet, it creates difficulties in differentiating between the surface weight \\(T(x)\\rho (x)\\) and the magnitude of radiance value \\(L_{o}(x,\\omega_{o})\\), since it allows for the possibility of assigning extreme radiance to the points with low surface weights to render same ray colors. Such ambiguity poses challenges, particularly in dark and high-contrast scenes, aggravating surface reconstruction (see Fig. 3). To address this, we introduce a learnable tone-mapper \\(m_{\\theta}:\\mathbb{R}_{+}^{3}\\to [0,1]^{3}\\), that takes positionally encoded HDR linear values as input:" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4600" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.099, + 0.095, + 0.207, + 0.152 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.09, + 0.159, + 0.214, + 0.231 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.126, + 0.234, + 0.179, + 0.246 + ], + "angle": 0, + "content": "(a) Image" + }, + { + "type": "image", + "bbox": [ + 0.222, + 0.095, + 0.328, + 0.153 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.225, + 0.159, + 0.318, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.243, + 0.234, + 0.305, + 0.246 + ], + "angle": 0, + "content": "(b) \\(\\mathrm{w / o}m_{\\theta}\\)" + }, + { + "type": "image", + "bbox": [ + 0.344, + 0.095, + 0.449, + 0.153 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.346, + 0.159, + 0.438, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.356, + 0.234, + 0.436, + 0.245 + ], + "angle": 0, + "content": "(c) ESR-NeRF" + }, + { + "type": "image_caption", + "bbox": [ + 0.079, + 0.252, + 0.467, + 0.266 + ], + "angle": 0, + "content": "Figure 3. Reconstructed surfaces with the learnable tone-mapper." + }, + { + "type": "equation", + "bbox": [ + 0.136, + 0.285, + 0.47, + 0.315 + ], + "angle": 0, + "content": "\\[\n\\hat {C} _ {m _ {\\theta}} (r) = \\int_ {0} ^ {\\infty} T (x) \\rho (x) m _ {\\theta} \\left(L _ {o} \\left(x, \\omega_ {o}\\right)\\right) d t, \\tag {4}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.149, + 0.323, + 0.469, + 0.341 + ], + "angle": 0, + "content": "\\[\nL _ {o} (x, \\omega_ {o}) = L _ {o} ^ {S} (x, \\omega_ {o}) + L _ {o} ^ {E} (x, \\omega_ {o}) \\cdot \\mathbb {I}, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.344, + 0.469, + 0.406 + ], + "angle": 0, + "content": "where \\(L_{o}^{S}(x,\\omega_{o})\\) is radiance when emissive sources are turned off, while \\(L_{o}^{E}(x,\\omega_{o})\\) stands for radiance added to the scene by emissive sources. Our rendering loss is then formulated as follows, with \\(\\lambda_{\\tau}\\) as a hyper-parameter:" + }, + { + "type": "equation", + "bbox": [ + 0.085, + 0.416, + 0.469, + 0.446 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {r e n d e r}} = \\sum_ {r \\in \\mathcal {R}} \\left(\\| C (r) - \\hat {C} _ {m _ {\\theta}} (r) \\| _ {2} ^ {2} + \\lambda_ {\\tau} \\| C (r) - \\hat {C} _ {\\tau} (r) \\| _ {2} ^ {2}\\right). \\tag {6}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.463, + 0.41, + 0.48 + ], + "angle": 0, + "content": "4.2. Learning of Light Transport Segments" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.487, + 0.469, + 0.623 + ], + "angle": 0, + "content": "The computational complexity of object appearance analysis in volume rendering is notably high, as shown in Eq. 3. We take an alternative approach by leveraging neural networks to represent ray-traced fields, rather than explicitly tracing every rays. Our distinct contribution to inverse rendering lies in precise adjustment of radiance. Specifically, we impose constraints on the predicted radiance to satisfy each light transport segments. The light transport segments (LTS) loss, \\(\\mathcal{L}_{lts}\\), plays a pivotal role in our method:" + }, + { + "type": "equation", + "bbox": [ + 0.15, + 0.633, + 0.469, + 0.663 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {l t s} ^ {S} = \\sum_ {x, \\omega_ {o}} \\| L _ {o} ^ {S} (x, \\omega_ {o}) - \\hat {L} _ {o} ^ {S} (x, \\omega_ {o}) \\| _ {2} ^ {2}, \\tag {7}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.15, + 0.676, + 0.469, + 0.706 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {l t s} ^ {E} = \\sum_ {x, \\omega_ {o}} \\| L _ {o} ^ {E} (x, \\omega_ {o}) - \\hat {L} _ {o} ^ {E} (x, \\omega_ {o}) \\| _ {2} ^ {2}, \\tag {8}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.098, + 0.723, + 0.412, + 0.768 + ], + "angle": 0, + "content": "\\[\n\\hat {L} _ {o} ^ {S} (x, \\omega_ {o}) = \\int_ {\\Omega} \\underbrace {S (\\omega_ {i}) V (x , \\omega_ {i}) R (x , \\omega_ {o} , \\omega_ {i})} _ {\\text {d i r e c t i l l u m i n a t i o n b y a n e v i r o n m e n t m a p}} d \\omega_ {i} +\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.102, + 0.765, + 0.427, + 0.806 + ], + "angle": 0, + "content": "\\[\n\\int_ {\\Omega} \\int_ {0} ^ {\\infty} \\underbrace {T (x ^ {\\prime}) \\rho (x ^ {\\prime}) L _ {o} ^ {S} (x ^ {\\prime} , - \\omega_ {i}) d t ^ {\\prime} R (x , \\omega_ {o} , \\omega_ {i})} _ {\\text {i n d r e c t i l l u m i n a t i o n b y a n e v i r o n m e n t m a p}} d \\omega_ {i}.\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.098, + 0.828, + 0.232, + 0.862 + ], + "angle": 0, + "content": "\\[\n\\hat {L} _ {o} ^ {E} (x, \\omega_ {o}) = \\underbrace {E (x)} _ {\\text {e m i s s i o n}} +\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.098, + 0.861, + 0.469, + 0.904 + ], + "angle": 0, + "content": "\\[\n\\int_ {\\Omega} \\int_ {0} ^ {\\infty} \\underbrace {T \\left(x ^ {\\prime}\\right) \\rho \\left(x ^ {\\prime}\\right) L _ {o} ^ {E} \\left(x ^ {\\prime} , - \\omega_ {i}\\right) d t ^ {\\prime} R \\left(x , \\omega_ {o} , \\omega_ {i}\\right)} _ {\\text {d i r e c t ＆ i n d i r c t i l l u m i n a t i o n b y e m i s s i v e s o u r c e s}} d \\omega_ {i}. \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.092, + 0.892, + 0.227 + ], + "angle": 0, + "content": "We ensure consistency between the radiance directly predicted by the network \\( L_{o}(x,\\omega_{o}) \\) and the radiance achievable based on the scene context \\( \\hat{L}_{o}(x,\\omega_{o}) \\). Previous approaches have focused on matching \\( \\hat{L}_{o}(x,\\omega_{o}) \\) to training views, overlooking the relations to \\( L_{o}(x,\\omega_{o}) \\). This hinders the restoration of HDR radiance by supervising scene components to LDR training views. In contrast, our LTS loss enables volumetric energy transfer of radiance, adjusting outgoing radiance based on their interrelations." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.228, + 0.892, + 0.349 + ], + "angle": 0, + "content": "To implement this concept, we train six dedicated networks for SDF \\( f(x) \\), SVBRDF parameters \\( b(x) \\), emission \\( E(x) \\), environment map \\( S(\\omega_{i}) \\), outgoing radiances \\( L_{o}^{S}(x,\\omega_{o}) \\) and \\( L_{o}^{E}(x,\\omega_{o}) \\), to adhere to these LTS requirements. For the environment map, we represent it using 48 Spherical Gaussians [67]: \\( \\sum_{k=1}^{M}\\mu_{k}e^{\\lambda_{k}(\\omega_{i}\\cdot\\xi_{k}-1)} \\), followed by the softplus activation. \\( \\mu\\in\\mathbb{R}^{3},\\lambda\\in\\mathbb{R}_{+} \\), and \\( \\xi\\in\\mathbb{S}^{2} \\) respectively denote the lobe amplitude, sharpness, and axis." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.357, + 0.854, + 0.373 + ], + "angle": 0, + "content": "4.3. Progressive Discovery of Reflection Areas" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.383, + 0.596, + 0.46 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.598, + 0.384, + 0.689, + 0.46 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.47, + 0.692, + 0.539 + ], + "angle": 0, + "content": "Figure 4. Left: Image with active emissive sources. Right: Identified emissive sources w/o progressive discovery of reflection areas." + }, + { + "type": "text", + "bbox": [ + 0.698, + 0.379, + 0.892, + 0.545 + ], + "angle": 0, + "content": "Relying solely on LTS is insufficient for addressing ambiguity arising from low pixel values of emissive sources and intense reflections in adjacent regions, often leading to confusion between emission and reflection. The right image in Fig. 4 shows self-emitting objects restored with the" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.546, + 0.892, + 0.637 + ], + "angle": 0, + "content": "naive LTS loss. While emissive sources are small, large areas affected by them are also identified as emissive sources. We propose a reflection-aware progressive approach for precise identification of emissive sources. By leveraging LTS learning, we extend the regions that can be regarded as reflection areas. Fig. 5 illustrates our progressive algorithm." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.643, + 0.892, + 0.809 + ], + "angle": 0, + "content": "Reflection-Aware Emission Refinement. Since surface points are unknown and are updated during learning, we opt to utilize rays rather than surface points. This process involves categorizing training rays into two groups: uncertain \\((\\mathcal{R}^U)\\) and certain \\((\\mathcal{R}^C)\\). The certain group contains the rays confidently identified as reflection, aiding the transfer of radiance energy to nearby points. For the points in the certain group, we use the Eq. 11 instead of Eq. 10 to exclusively attribute outgoing radiances to reflections. Satisfying the LTS loss on the certain group results in adjusting the outgoing radiances of influential points, as illustrated in Fig. 5(a):" + }, + { + "type": "equation", + "bbox": [ + 0.5, + 0.815, + 0.91, + 0.855 + ], + "angle": 0, + "content": "\\[\n\\hat {L} _ {o} ^ {E} \\left(x, \\omega_ {o}\\right) = \\int_ {\\Omega} \\int_ {0} ^ {\\infty} T \\left(x ^ {\\prime}\\right) \\rho \\left(x ^ {\\prime}\\right) L _ {o} ^ {E} \\left(x ^ {\\prime}, - \\omega_ {i}\\right) d t ^ {\\prime} R \\left(x, \\omega_ {o}, \\omega_ {i}\\right) d \\omega_ {i}. \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.856, + 0.892, + 0.902 + ], + "angle": 0, + "content": "The uncertain group includes the rays indicating the areas that are undetermined yet as reflection or emission. Using Eq. 12 to compute \\(\\hat{L}_o^E (x,\\omega_o)\\), this group adjusts emis" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "4601" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.47, + 0.122 + ], + "angle": 0, + "content": "sions \\( E(x) \\) based on the radiance updates by the certain group, where \"sg\" represents the stop-gradient:" + }, + { + "type": "equation", + "bbox": [ + 0.088, + 0.129, + 0.469, + 0.19 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\hat {L} _ {o} ^ {E} (x, \\omega_ {o}) = E (x) + \\\\ \\left. \\operatorname {s g} \\left(\\int_ {\\Omega} \\int_ {0} ^ {\\infty} T \\left(x ^ {\\prime}\\right) \\rho \\left(x ^ {\\prime}\\right) L _ {o} ^ {E} \\left(x ^ {\\prime}, - \\omega_ {i}\\right) d t ^ {\\prime} R \\left(x, \\omega_ {o}, \\omega_ {i}\\right) d \\omega_ {i}\\right). \\right. \\tag {12} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.192, + 0.469, + 0.282 + ], + "angle": 0, + "content": "As shown in Fig. 5(b), this leads to increased emissions for the regions whose radiances are adjusted to account for the reflections in the certain group. Conversely, emissions decrease for the regions where there is little change in outgoing radiance, but incident radiances are increased by surrounding influential points." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.289, + 0.469, + 0.38 + ], + "angle": 0, + "content": "Ray Group Management. As emissions and radiances are adjusted, the groups are dynamically updated at predefined training intervals through the following process. Within the uncertain group, we evaluate the expected emission strength of rays, retaining only those above a threshold \\( k_{i} \\). Rays below this threshold are then merged to the certain group:" + }, + { + "type": "equation", + "bbox": [ + 0.088, + 0.387, + 0.469, + 0.428 + ], + "angle": 0, + "content": "\\[\n\\mathcal {R} _ {i} ^ {U} = \\left\\{r \\mid \\max _ {R G B} \\left(\\int_ {0} ^ {\\infty} T (x) \\rho (x) E (x) d t\\right) \\geq k _ {i}, r \\in \\mathcal {R} _ {i - 1} ^ {U} \\right\\}, \\tag {13}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.176, + 0.439, + 0.469, + 0.462 + ], + "angle": 0, + "content": "\\[\n\\mathcal {R} _ {i} ^ {C} = \\left(\\mathcal {R} _ {i - 1} ^ {U} - \\mathcal {R} _ {t} ^ {U}\\right) \\cup \\mathcal {R} _ {i - 1} ^ {C}. \\tag {14}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.465, + 0.469, + 0.54 + ], + "angle": 0, + "content": "Subsequently, newly added rays to the certain group can be used to localize influential points and update their outgoing radiances. This iterative process progressively refines the separation between reflective and emissive regions, attaining more accurate identification of emissive sources." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.546, + 0.469, + 0.591 + ], + "angle": 0, + "content": "LTS Loss Decomposition. The LTS loss, as detailed in Eq. 15, can be decomposed using a stop-gradient operation to refine the adjustment process." + }, + { + "type": "equation", + "bbox": [ + 0.11, + 0.599, + 0.469, + 0.631 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {l t s} ^ {E} = \\sum_ {x, \\omega_ {o}} \\left(\\lambda_ {l} \\| \\operatorname {s g} \\left(L _ {o} ^ {E} (x, \\omega_ {o})\\right) - \\hat {L} _ {o} ^ {E} (x, \\omega_ {o}) \\| _ {1} + \\right. \\tag {15}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.188, + 0.632, + 0.41, + 0.65 + ], + "angle": 0, + "content": "\\[\n\\lambda_ {r} \\| L _ {o} ^ {E} (x, \\omega_ {o}) - \\operatorname {s g} (\\hat {L} _ {o} ^ {E} (x, \\omega_ {o})) \\| _ {1}).\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.658, + 0.469, + 0.748 + ], + "angle": 0, + "content": "We prioritize \\(\\lambda_{l}\\) to enhance the update of scene context, affecting other points' radiance given the predicted \\(L_{o}(x,\\omega_{o})\\). \\(\\lambda_{r}\\) prevents severe deviation of every \\(L_{o}(x,\\omega_{o})\\) within the current scene context. This aligns with our focus on HDR source reconstruction from LDR images, addressing under-represented information in training data." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.757, + 0.238, + 0.772 + ], + "angle": 0, + "content": "4.4. Training Details" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.78, + 0.469, + 0.901 + ], + "angle": 0, + "content": "We employ the Voxurf architecture [80] as backbone and adopt the simplified Disney BRDF model [10] for SVBRDF representation, with parameters including base color \\(\\in [0,1]^3\\), roughness \\(\\in [0,1]\\), and metallic \\(\\in [0,1]\\). The learnable tone-mapper, structured as a two-layer MLP, is utilized for the rendering loss only. Initially, we pre-train our networks using the rendering loss, subsequently integrating the basic LTS loss (Eq. 7 and Eq. 8) into our training regimen." + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.089, + 0.624, + 0.15 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.521, + 0.154, + 0.591, + 0.166 + ], + "angle": 0, + "content": "(a) LTS on \\(R^C\\)" + }, + { + "type": "image", + "bbox": [ + 0.635, + 0.089, + 0.755, + 0.15 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.653, + 0.154, + 0.726, + 0.166 + ], + "angle": 0, + "content": "(b) LTS on \\(R^U\\)" + }, + { + "type": "image", + "bbox": [ + 0.767, + 0.097, + 0.888, + 0.15 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.765, + 0.154, + 0.889, + 0.167 + ], + "angle": 0, + "content": "(c) Ray Group Managing" + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.174, + 0.892, + 0.23 + ], + "angle": 0, + "content": "Figure 5. Illustration of the progressive emissive source reconstruction with reflection awareness. Gray color represents the areas belonging to the certain group, while the red (emissive sources) and orange (their reflections) areas belong to the uncertain group." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.242, + 0.892, + 0.362 + ], + "angle": 0, + "content": "This phase transitions to the reflection-aware progressive training scheme, where we adopt the \\(\\ell_1\\) loss due to its empirical stability in refining emissive source reconstruction. We use a smoothing regularization to promote local consistency in normals, BRDFs, and emissions. To ensure view-consistent labeling of 3D points as either reflective or emissive, we implement the emission suppression loss for points belonging to the certain group:" + }, + { + "type": "equation", + "bbox": [ + 0.567, + 0.368, + 0.891, + 0.405 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {s u p p}} ^ {E} = \\sum_ {r \\in R _ {t} ^ {C}} \\| \\int_ {0} ^ {\\infty} T (x) \\rho (x) E (x) d t \\| _ {2} ^ {2}, \\tag {16}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.412, + 0.892, + 0.488 + ], + "angle": 0, + "content": "The threshold \\( k_{i} \\) linearly increases with each time step \\( t \\), utilizing a grid search within a range of \\( [10^{-3}, 10^{-5}] \\) to find the slope. We construct mini-batches via stratified sampling within each group. For a detailed description of our training procedure, please refer to Appendix." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.496, + 0.644, + 0.512 + ], + "angle": 0, + "content": "4.5. Scene Editing" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.519, + 0.892, + 0.579 + ], + "angle": 0, + "content": "Reconstructed emissive sources enable scene editing; users select emissive sources using binary masks \\( M_{j=1\\dots N} \\) and specify lighting conditions using colors \\( c_{j=1\\dots N} \\) and intensities \\( i_{j=1\\dots N} \\) within the HSV color space [53]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.58, + 0.892, + 0.624 + ], + "angle": 0, + "content": "We identify the rays in the uncertain group that match \\(M\\) by projecting expected surface points \\(p\\) of the rays onto the camera with the pose \\(\\mathbf{R}|\\mathbf{t}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.619, + 0.631, + 0.891, + 0.66 + ], + "angle": 0, + "content": "\\[\np = \\int_ {0} ^ {\\infty} T (x) \\rho (x) x d t, \\tag {17}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.509, + 0.681, + 0.891, + 0.7 + ], + "angle": 0, + "content": "\\[\n\\mathbb {I} _ {j} ^ {\\text {h i t}} (x) = \\operatorname {i n t e r p} \\left(M _ {j}, p ^ {\\prime}\\right) > 0, \\text {w h e r e} p ^ {\\prime} = \\mathbf {K} [ \\mathbf {R} | \\mathbf {t} ] [ p | 1 ] ^ {T}. \\tag {18}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.706, + 0.892, + 0.782 + ], + "angle": 0, + "content": "For the rays satisfying \\(\\mathbb{I}_j^{hit}(x)\\), we apply the designated lighting conditions. The new emission values are computed by substituting the original hue (H) and saturation (S) of \\(E(x)\\) with the user-specified color \\(c_{j}\\) and adjusting the value (V) of \\(v(x)\\) with the new intensity \\(i_j\\):" + }, + { + "type": "equation", + "bbox": [ + 0.503, + 0.789, + 0.891, + 0.807 + ], + "angle": 0, + "content": "\\[\nE (x) = \\operatorname {h s v} _ {-} \\operatorname {t o} _ {-} \\operatorname {r g b} ([ c _ {j} | (v (x) \\times i) ]) \\cdot \\mathbb {I} _ {j} ^ {h i t} + E (x) \\cdot \\neg \\mathbb {I} _ {j} ^ {h i t}. \\tag {19}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.814, + 0.891, + 0.86 + ], + "angle": 0, + "content": "These modifications influence scene appearance by optimizing the loss in Eq. 20. During this process, all networks, except for \\( L_{o}^{E}(x,\\omega_{o}) \\), are frozen:" + }, + { + "type": "equation", + "bbox": [ + 0.556, + 0.866, + 0.891, + 0.896 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {e d i t}} = \\sum_ {x, \\omega_ {o}} \\| L _ {o} ^ {E} (x, \\omega_ {o}) - \\operatorname {s g} \\left(\\hat {L} _ {o} ^ {E} (x, \\omega_ {o})\\right) \\| _ {2} ^ {2}. \\tag {20}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4602" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.082, + 0.077, + 0.891, + 0.177 + ], + "angle": 0, + "content": "
LegoGiftWhite colored BookCubeBillboardBallsVivid colored
IoUMSEIoUMSEIoUMSEIoUMSEIoUMSEIoUMSEIoUMSEIoUMSEIoUMSEIoUMSEIoUMSE
Twins0.2220.190.498.590.633.910.9531.830.691.120.900.060.256.960.246.090.552.630.9510.640.090.750.830.04
NeLF++0.4320.880.079.380.954.640.9332.670.011.950.910.800.307.650.096.860.953.360.9411.490.021.570.920.78
TensoIR0.7120.130.158.550.953.870.9531.730.761.110.950.050.336.930.156.050.952.590.9610.600.770.740.950.03
ESR-NeRF0.818.380.603.490.961.190.9717.870.840.460.950.040.515.480.592.500.960.510.977.940.880.260.940.03
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.186, + 0.895, + 0.23 + ], + "angle": 0, + "content": "Table 2. Results of emissive source identification. ESR-NeRF outperforms state-of-the-art re-lighting methods in reconstructing emissive sources, regardless of their color. The IoU measures the source area identification (a higher value is better), and the MSE quantifies the difference between reconstructed images and HDR ground truth images (a lower value is better)." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.242, + 0.21, + 0.259 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.268, + 0.471, + 0.359 + ], + "angle": 0, + "content": "We assess ESR-NeRF in reconstructing emissive sources by focusing on both identification and intensity restoration. To showcase its effectiveness, we conduct a range of experiments, including scene editing, ablation studies, illumination decomposition, and surface reconstruction, providing both quantitative and qualitative results." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.371, + 0.271, + 0.387 + ], + "angle": 0, + "content": "5.1. Experiment Settings" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.395, + 0.473, + 0.699 + ], + "angle": 0, + "content": "We curate 6 diverse synthetic scenes, each with 200 training images evenly distributed between on and off lighting conditions. To evaluate the robustness of our approach against light colors, we consider two distinct settings of white colored and vivid colored emissive sources, resulting in a total of 12 scenes. The vivid colors are selected with full saturation in the HSV color space. We measure source identification and radiance reconstruction using IoU and MSE metrics on novel view test images, comparing against ground truth data from Blender-rendered emission masks and EXR files. The emission strengths, the maximum EXR file values, range from 2 to 200. For quantitative scene editing evaluation, we alter the white-colored sources to various colors—red, green, blue, cyan, magenta, yellow—and adjust intensities to half or double their original values. Qualitative results include scene editing for vividly colored sources and real scenes captured with a Fuji 100s camera using Philips smart bulbs as emissive sources. Quantitative assessments are based on 50 test images from novel camera poses, except for MSE measured for 25 test" + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.703, + 0.466, + 0.862 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.873, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Figure 6. Comparison of identified emissive sources. ESR-NeRF excels through the reflection-aware progressive refinement." + }, + { + "type": "table", + "bbox": [ + 0.501, + 0.239, + 0.899, + 0.35 + ], + "angle": 0, + "content": "
NVNV + INV+ CNV + I + C
PSNRLPIPSPSNRLPIPSPSNRLPIPSPSNRLPIPS
Twins36.520.014127.910.025231.020.025228.210.0310
NeRF-W36.440.014224.770.0417----
NeILF++24.400.055624.710.057924.060.075023.240.0770
TensoIR38.040.010327.280.041826.360.050525.180.0531
PaletteNeRF33.660.023323.270.048324.440.064622.580.0703
ESR-NeRF38.790.008329.990.019331.730.019631.630.0199
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.36, + 0.894, + 0.401 + ], + "angle": 0, + "content": "Table 3. Scene editing results. NV: novel view synthesis, I: intensity editing, and C: color editing. A higher PSNR or lower LPIPS value is better." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.409, + 0.892, + 0.469 + ], + "angle": 0, + "content": "images. We denote the best performance with blue and the second-best with green. Additionally, we utilize the DTU dataset [23] to evaluate ESR-NeRF's performance in surface reconstruction tasks where emissive sources are absent." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.476, + 0.893, + 0.718 + ], + "angle": 0, + "content": "Baselines. We select two state-of-the-art re-lighting methods, TensoIR [25] and NeILF++ [90], that do not require prior lighting information. For thorough evaluation, we also implement a simple method, Twins, where separate models are trained under light on and off conditions. The Twins utilize the radiance discrepancies between the on and off models to distinguish and adjust emissive sources. For scene editing, we add NeRF-W [38] and PaletteNeRF [28] as baselines. Both NeRF-W and Twins adopt the Voxurf [80] architecture for fair comparison. For methods unable to individually control emissive sources, all sources are adjusted together to match the last lighting condition by a user. For the DTU dataset, we include state-of-the-art surface reconstruction methods that use object masks, such as NeuS [68] and Voxurf, as well as Neural-PBIR [57], that jointly reconstructs surfaces, materials, and environment maps." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.727, + 0.593, + 0.74 + ], + "angle": 0, + "content": "5.2. Results" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.892, + 0.903 + ], + "angle": 0, + "content": "Emissive Source Recosnstruction. Tab. 2 shows that our approach excels in accurately identifying emissive source regions and restoring their intensity, regardless of the source color. While TensoIR and NeILF++ can restore emissions by modifying their physical rendering equations, they suffer from emissive source ambiguity, leading to near-zero IoU performance (see Appendix). For a comprehensive comparison, we report the best performance of the baseline methods using thresholding on the reconstructed emission strength at 0.01 intervals. ESR-NeRF consistently outper" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4603" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.092, + 0.072, + 0.874, + 0.407 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.411, + 0.893, + 0.44 + ], + "angle": 0, + "content": "Figure 7. Comparison of scene editing. ESR-NeRF provides precise source control and faithfully represents reflection effects. For easy comparison in the Cube scene of low intensity, the bottom-right images are presented with a \\(40\\%\\) increased brightness." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.446, + 0.473, + 0.569 + ], + "angle": 0, + "content": "forms the baselines in identifying emissive source regions across all scenes. Our method also achieves significantly lower MSE values for restoring LDR to HDR images compared to the baselines, demonstrating its effectiveness of handling the ill-posed nature of the scenes with emissive sources. This is visually confirmed in Fig. 6, where ESR-NeRF surpasses the baselines in a complex scene with numerous small light bulbs." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.577, + 0.472, + 0.804 + ], + "angle": 0, + "content": "Scene Editing. Tab. 3 and Fig. 7 showcase the scene editing results under novel lighting conditions. Baseline methods struggle to adapt to lighting changes due to their inability to reconstruct emissive sources accurately. For example, in the Lego scene, TensoIR fails to adjust the illumination in surrounding regions when the color of emissive sources is changed, and in the Cube scene, both the hidden iPad screen and the cube surface covered by the user input mask change together. Twins introduces blue light onto yellow and red surfaces, leading to unintended white and purple appearances, even though there should be no reflection. PaletteNeRF, which manipulates scenes through re-colorization, lacks precise control over illumination, as seen in the synchronous color changes in the yellow ribbon and lighting. In contrast, ESR-NeRF demonstrates superior performance in" + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.812, + 0.472, + 0.884 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.89, + 0.456, + 0.904 + ], + "angle": 0, + "content": "Figure 8. Reconstructed emitter and re-lighting at novel view." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.446, + 0.892, + 0.552 + ], + "angle": 0, + "content": "scene editing outshining all baselines thanks to the accurate identification of emissive sources, as detailed in Table 3. ESR-NeRF effectively balances source reconstruction and novel view synthesis, ensuring high performance in both tasks. NeRF-W is excluded from color adjustments since it doesn't support direct color change through interpolating latent variables learned with light on and off conditions." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.553, + 0.893, + 0.658 + ], + "angle": 0, + "content": "Fig. 8 to 9 present additional examples of emissive source reconstruction and scene editing results. Fig. 10 shows results on real scenes, for which due to the impracticality of precise control over smart bulb colors, we offer emission reconstruction results with pseudo ground truth data. Our method effectively identifies emissive sources in real scenes, while it faces challenges in capturing complex" + }, + { + "type": "image", + "bbox": [ + 0.501, + 0.66, + 0.892, + 0.886 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.515, + 0.89, + 0.875, + 0.905 + ], + "angle": 0, + "content": "Figure 9. Results of source reconstruction and scene editing." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "4604" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.078, + 0.082, + 0.47, + 0.254 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.263, + 0.468, + 0.278 + ], + "angle": 0, + "content": "Figure 10. Source reconstruction and scene editing on real scenes." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.287, + 0.469, + 0.318 + ], + "angle": 0, + "content": "reflections within light bulbs, as evident in the bright spot at the center of the bulbs in the ground truth edit results." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.324, + 0.469, + 0.43 + ], + "angle": 0, + "content": "Ablation Analysis. Progressive refinement with the stop-gradient operation in Eq. 15 improves the identification of emissive sources and reduces MSE values. Without \\( m_{\\theta} \\), surface reconstructions become unreliable, complicating the accurate reconstruction of emissive sources. This issue is evident from the CD metrics and illustrated in Fig. 3. Further analyses are provided in Appendix." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.436, + 0.469, + 0.572 + ], + "angle": 0, + "content": "Illumination Decomposition. Fig. 11 demonstrates ESR-NeRF's decomposition of scene illumination into direct and indirect lighting from an environment map, as well as emissions and their reflections. The shadow behind the yellow ribbon in the direct figure and the illumination in the indirect figure showcase ESR-NeRF's ability to model both direct and indirect illumination. The reflection figure shows that our method accurately captures how emissive sources contribute to reflections on nearby regions." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.579, + 0.469, + 0.73 + ], + "angle": 0, + "content": "Surface Reconstruction. Interestingly, our approach can be applied to the scenes without emissive sources to enhance surface reconstruction, as evidenced by the lower CD values in Tab. 4 on the DTU dataset. For this experiment, we use Eq. 7 to 10 without our progressive refinement technique. ESR-NeRF's ability to adjust interrelated outgoing radiances helps prevent surface formations where radiances cannot be produced, considering the predicted scene context. Additional visualizations of the normals, BRDF, and environment maps are provided in Appendix." + }, + { + "type": "image", + "bbox": [ + 0.08, + 0.739, + 0.47, + 0.883 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.11, + 0.888, + 0.435, + 0.902 + ], + "angle": 0, + "content": "Figure 11. An example of illumination decomposition." + }, + { + "type": "table", + "bbox": [ + 0.564, + 0.077, + 0.828, + 0.27 + ], + "angle": 0, + "content": "
ScanNeuSVoxurfNeural-PBIRESR-NeRF
240.830.650.570.58
370.980.740.750.71
400.560.390.380.38
550.370.350.360.33
631.130.961.040.93
650.590.640.730.57
690.600.850.650.78
831.451.581.281.18
970.951.010.970.95
1050.780.680.760.58
1060.520.600.530.54
1101.431.110.841.08
1140.360.370.380.33
1180.450.450.460.40
1220.450.470.490.44
mean0.770.720.680.65
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.276, + 0.892, + 0.303 + ], + "angle": 0, + "content": "Table 4. Results of surface reconstruction via the Chamfer distance on the DTU dataset. A lower value is better." + }, + { + "type": "table", + "bbox": [ + 0.505, + 0.311, + 0.768, + 0.386 + ], + "angle": 0, + "content": "
WhiteVivid
IoU ↑MSE ↓IoU ↑MSE ↓
w/o progressive0.409.920.413.93
w/o sg0.716.450.603.47
ESR-NeRF0.865.240.812.79
" + }, + { + "type": "table", + "bbox": [ + 0.779, + 0.312, + 0.887, + 0.385 + ], + "angle": 0, + "content": "
DTU\nCD ↓
w/o mθ0.93
w/o LTS0.71
ESR-NeRF0.65
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.39, + 0.892, + 0.418 + ], + "angle": 0, + "content": "Table 5. Ablation studies on the surface reconstruction (left) and the emissive source reconstruction (right)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.43, + 0.618, + 0.445 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.455, + 0.892, + 0.545 + ], + "angle": 0, + "content": "We present ESR-NeRF as the first NeRF-based inverse rendering method for the scenes with emissive sources. Our approach uses LDR images, eliminating the need of HDR images to reconstruct emissive sources. Furthermore, we demonstrate the application of reconstructed sources in scene editing, enabling color and intensity modifications." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.546, + 0.892, + 0.712 + ], + "angle": 0, + "content": "Limitations. Future work could explore using a single lighting condition to disentangle emissive sources, environmental lighting, and object texture. It is also promising to address the challenge of volume ray tracing in unbounded scenes to extend to indoor scenes. Additionally, LTS based re-lighting may be weak in representing new colors that traverse unobserved light paths during training. An alternative approach could be extracting emission texture maps and modifying it using the engines such as Blender [13] or Mitsuba [22]. More details on alternative re-lighting methods and radiance fine-tuning are provided in Appendix." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.725, + 0.688, + 0.742 + ], + "angle": 0, + "content": "7. Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.892, + 0.901 + ], + "angle": 0, + "content": "This work was supported by Samsung Electronics MX, Basic Science Research Program through the National Research Foundation of Korea(NRF) funded by the Ministry of Education(RS-2023-00274280), and Institute of Information & Communications Technology Planning & Evaluation (IITP) grant funded by the Korea government (MSIT) (No. 2019-0-01082, SW StarLab; No. 2022-0-00156, Fundamental research on continual meta-learning for quality enhancement of casual videos and their 3D metaverse transformation). Gunhee Kim is the corresponding author." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4605" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.468, + 0.156 + ], + "angle": 0, + "content": "[1] Benjamin Attal, Jia-Bin Huang, Michael Zollhöfer, Johannes Kopf, and Changil Kim. Learning neural light fields with ray-space embedding. In CVPR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.159, + 0.468, + 0.2 + ], + "angle": 0, + "content": "[2] Dejan Azinovic, Tzu-Mao Li, Anton Kaplanyan, and Matthias Nießner. Inverse path tracing for joint material and lighting estimation. In CVPR, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.089, + 0.202, + 0.468, + 0.256 + ], + "angle": 0, + "content": "[3] Jonathan T Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P Srinivasan. Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In ICCV, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.089, + 0.259, + 0.468, + 0.299 + ], + "angle": 0, + "content": "[4] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Mip-nerf 360: Unbounded anti-aliased neural radiance fields. In CVPR, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.302, + 0.468, + 0.342 + ], + "angle": 0, + "content": "[5] Jonathan T. Barron, Ben Mildenhall, Dor Verbin, Pratul P. Srinivasan, and Peter Hedman. Zip-nerf: Anti-aliased grid-based neural radiance fields. In ICCV, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.345, + 0.468, + 0.4 + ], + "angle": 0, + "content": "[6] Sai Bi, Zexiang Xu, Pratul Srinivasan, Ben Mildenhall, Kalyan Sunkavalli, Miloš Hasan, Yannick Hold-Geoffroy, David Kriegman, and Ravi Ramamoorthi. Neural reflectance fields for appearance acquisition. In arXiv, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.402, + 0.468, + 0.441 + ], + "angle": 0, + "content": "[7] Mark Boss, Varun Jampani, Kihwan Kim, Hendrik Lensch, and Jan Kautz. Two-shot spatially-varying brdf and shape estimation. In CVPR, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.445, + 0.468, + 0.498 + ], + "angle": 0, + "content": "[8] Mark Boss, Raphael Braun, Varun Jampani, Jonathan T. Barron, Ce Liu, and Hendrik P.A. Lensch. Nerd: Neural reflectance decomposition from image collections. In ICCV, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.502, + 0.468, + 0.556 + ], + "angle": 0, + "content": "[9] Mark Boss, Varun Jampani, Raphael Braun, Ce Liu, Jonathan T. Barron, and Hendrik P.A. Lensch. Neural-pil: Neural pre-integrated lighting for reflectance decomposition. In NeurIPS, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.558, + 0.468, + 0.598 + ], + "angle": 0, + "content": "[10] Brent Burley and Walt Disney Animation Studios. Physically-based shading at disney. In SIGGRAPH, 2012. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.602, + 0.468, + 0.642 + ], + "angle": 0, + "content": "[11] Bowen Cai, Jinchi Huang, Rongfei Jia, Chengfei Lv, and Huan Fu. Neuda: Neural deformable anchor for high-fidelity implicit surface reconstruction. In CVPR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.645, + 0.468, + 0.683 + ], + "angle": 0, + "content": "[12] Anpei Chen, Zexiang Xu, Andreas Geiger, Jingyi Yu, and Hao Su. Tensorf: Tensorial radiance fields. In ECCV, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.688, + 0.468, + 0.728 + ], + "angle": 0, + "content": "[13] Blender Online Community. Blender - a 3D modelling and rendering package. Blender Foundation, Stichting Blender Foundation, Amsterdam, 2018. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.731, + 0.468, + 0.771 + ], + "angle": 0, + "content": "[14] Sara Fridovich-Keil, Alex Yu, Matthew Tancik, Qinhong Chen, Benjamin Recht, and Angjoo Kanazawa. Plenoxels: Radiance fields without neural networks. In CVPR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.774, + 0.468, + 0.814 + ], + "angle": 0, + "content": "[15] Mathieu Garon, Kalyan Sunkavalli, Sunil Hadap, Nathan Carr, and Jean-Francois Lalonde. Fast spatially-varying indoor lighting estimation. In CVPR, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.817, + 0.468, + 0.87 + ], + "angle": 0, + "content": "[16] Wenhang Ge, Tao Hu, Haoyu Zhao, Shu Liu, and Ying-Cong Chen. Ref-neus: Ambiguity-reduced neural implicit surface learning for multi-view reconstruction with reflection. In ICCV, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.874, + 0.468, + 0.9 + ], + "angle": 0, + "content": "[17] Saeed Hadadan, Shuhong Chen, and Matthias Zwicker. Neural radiosivity. In ACM TOG, 2021. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.468, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.147 + ], + "angle": 0, + "content": "[18] Saeed Hadadan, Geng Lin, Jan Novák, Fabrice Rousselle, and Matthias Zwicker. Inverse global illumination using a neural radiometric prior. In SIGGRAPH Conference Proceedings, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.151, + 0.892, + 0.218 + ], + "angle": 0, + "content": "[19] Bjoern Haefner, Simon Green, Alan Oursland, Daniel Andersen, Michael Goesele, Daniel Cremers, Richard Newcombe, and Thomas Whelan. Recovering real-world reflectance properties and shading from hdr imagery. In 3DV, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.222, + 0.892, + 0.275 + ], + "angle": 0, + "content": "[20] Jon Hasselgren, Nikolai Hofmann, and Jacob Munkberg. Shape, light, and material decomposition from images using monte carlo rendering and denoising. In NeurIPS, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.279, + 0.892, + 0.306 + ], + "angle": 0, + "content": "[21] IEC. IEC 61966-2-1:1999. Technical report, International Electrotechnical Commission, 1999. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.31, + 0.892, + 0.364 + ], + "angle": 0, + "content": "[22] Wenzel Jakob, Sébastien Speierer, Nicolas Roussel, Merlin Nimier-David, Delio Vicini, Tizian Zeltner, Baptiste Nicolet, Miguel Crespo, Vincent Leroy, and Ziyi Zhang. Mitsuba 3 renderer, 2022. https://mitsuba-renderer.org. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.368, + 0.892, + 0.407 + ], + "angle": 0, + "content": "[23] Rasmus Jensen, Anders Dahl, George Vogiatzis, Engil Tola, and Henrik Aanæs. Large scale multi-view stereopsis evaluation. In CVPR, 2014. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.411, + 0.892, + 0.451 + ], + "angle": 0, + "content": "[24] Chaonan Ji, Tao Yu, Kaiwen Guo, Jingxin Liu, and Yebin Liu. Geometry-aware single-image full-body human relighting. In ECCV, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.455, + 0.892, + 0.495 + ], + "angle": 0, + "content": "[25] Haian Jin, Isabella Liu, Peijia Xu, Xiaoshuai Zhang, Songfang Han, Sai Bi, Xiaowei Zhou, Zexiang Xu, and Hao Su. Tensoir: Tensorial inverse rendering. In CVPR, 2023. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.499, + 0.892, + 0.524 + ], + "angle": 0, + "content": "[26] James T. Kajiya. The rendering equation. In SIGGRAPH, 1986. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.528, + 0.892, + 0.569 + ], + "angle": 0, + "content": "[27] Petr Kellnhofer, Lars C Jebe, Andrew Jones, Ryan Spicer, Kari Pulli, and Gordon Wetzstein. Neural lumigraph rendering. In CVPR, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.572, + 0.892, + 0.625 + ], + "angle": 0, + "content": "[28] Zhengfei Kuang, Fujun Luan, Sai Bi, Zhixin Shu, Gordon Wetzstein, and Kalyan Sunkavalli. Palettenerf: Palette-based appearance editing of neural radiance fields. In CVPR, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.63, + 0.892, + 0.684 + ], + "angle": 0, + "content": "[29] Zhengqin Li, Mohammad Shafiei, Ravi Ramamoorthi, Kalyan Sunkavalli, and Manmohan Chandraker. Inverse rendering for complex indoor scenes: Shape, spatially-varying lighting and svbrdf from a single image. In CVPR, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.687, + 0.892, + 0.742 + ], + "angle": 0, + "content": "[30] Zhengqin Li, Jia Shi, Sai Bi, Rui Zhu, Kalyan Sunkavalli, Miloš Hašan, Zexiang Xu, Ravi Ramamoorthi, and Manmohan Chandraker. Physically-based editing of indoor scene lighting from a single image. In ECCV, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.746, + 0.892, + 0.798 + ], + "angle": 0, + "content": "[31] Zhaoshuo Li, Thomas Müller, Alex Evans, Russell H Taylor, Mathias Unberath, Ming-Yu Liu, and Chen-Hsuan Lin. Neuralangelo: High-fidelity neural surface reconstruction. In CVPR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.802, + 0.892, + 0.842 + ], + "angle": 0, + "content": "[32] Zhen Li, Lingli Wang, Mofang Cheng, Cihui Pan, and Jiaqi. Yang. Multi-view inverse rendering for large-scale real-world indoor scenes. In CVPR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.846, + 0.892, + 0.899 + ], + "angle": 0, + "content": "[33] Ruofan Liang, Huiting Chen, Chunlin Li, Fan Chen, Selvakumar Panneer, and Nandita Vijaykumar. Envidr: Implicit differentiable renderer with neural environment lighting. In ICCV, 2023. 1" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.956 + ], + "angle": 0, + "content": "4606" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.469, + 0.134 + ], + "angle": 0, + "content": "[34] Fujun Luan, Shuang Zhao, Kavita Bala, and Zhao Dong. Unified shape and svbrdf recovery using differentiable monte carlo rendering. In EGSR, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.134, + 0.469, + 0.189 + ], + "angle": 0, + "content": "[35] Linjie Lyu, Ayush Tewari, Thomas Leimkuehler, Marc Habermann, and Christian Theobalt. Neural radiance transfer fields for relightable novel-view synthesis with global illumination. In ECCV, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.19, + 0.469, + 0.245 + ], + "angle": 0, + "content": "[36] Linjie Lyu, Ayush Tewari, Marc Habermann, Shunsuke Saito, Michael Zollhöfer, Thomas Leimkuehler, and Christian Theobalt. Diffusion posterior illumination for ambiguity-aware inverse rendering. In ACM TOG, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.246, + 0.469, + 0.287 + ], + "angle": 0, + "content": "[37] Alexander Mai, Dor Verbin, Falko Kuester, and Sara Fridovich-Keil. Neural microfacet fields for inverse rendering. In ICCV, 2023. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.288, + 0.469, + 0.343 + ], + "angle": 0, + "content": "[38] Ricardo Martin-Brualla, Noha Radwan, Mehdi S. M. Sajjadi, Jonathan T. Barron, Alexey Dosovitskiy, and Daniel Duckworth. Nerf in the wild: Neural radiance fields for unconstrained photo collections. In CVPR, 2021. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.343, + 0.469, + 0.384 + ], + "angle": 0, + "content": "[39] Nelson Max. Optical models for direct volume rendering. IEEE Transactions on Visualization and Computer Graphics, 1(2):99-108, 1995. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.385, + 0.469, + 0.44 + ], + "angle": 0, + "content": "[40] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In ECCV, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.441, + 0.469, + 0.482 + ], + "angle": 0, + "content": "[41] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. In ACM TOG, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.483, + 0.469, + 0.538 + ], + "angle": 0, + "content": "[42] Jacob Munkberg, Jon Hasselgren, Tianchang Shen, Jun Gao, Wenzheng Chen, Alex Evans, Thomas Müller, and Sanja Fidler. Extracting triangular 3d models, materials, and lighting from images. In CVPR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.539, + 0.469, + 0.593 + ], + "angle": 0, + "content": "[43] Michael Niemeyer, Lars Mescheder, Michael Oechsle, and Andreas Geiger. Differentiable volumetric rendering: Learning implicit 3d representations without 3d supervision. In CVPR, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.594, + 0.469, + 0.636 + ], + "angle": 0, + "content": "[44] Michael Oechsle, Songyou Peng, and Andreas Geiger. Unisurf: Unifying neural implicit surfaces and radiance fields for multi-view reconstruction. In ICCV, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.636, + 0.469, + 0.664 + ], + "angle": 0, + "content": "[45] Julian Ost, Issam Laradji, Alejandro Newell, Yuval Bahat, and Felix Heide. Neural point light fields. In CVPR, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.664, + 0.469, + 0.732 + ], + "angle": 0, + "content": "[46] Rohit Pandey, Sergio Orts-Escalano, Chloe LeGendre, Christian Haene, Sofien Bouaziz, Christoph Rhemann, Paul Debevec, and Seann Fanello. Total relighting: Learning to relight portraits for background replacement. In ACM TOG, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.733, + 0.469, + 0.788 + ], + "angle": 0, + "content": "[47] Jeong Joon Park, Peter Florence, Julian Straub, Richard Newcombe, and Steven Lovegrove. Deepsdf: Learning continuous signed distance functions for shape representation. In CVPR, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.789, + 0.469, + 0.83 + ], + "angle": 0, + "content": "[48] Julien Philip, Sébastien Morgenthaler, Michael Gharbi, and George Drettakis. Free-viewpoint indoor neural relighting from multi-view stereo. In ACM TOG, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.831, + 0.469, + 0.859 + ], + "angle": 0, + "content": "[49] Martin Piala and Ronald Clark. Terminerf: Ray termination prediction for efficient neural rendering. In 3DV, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.859, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[50] Viktor Rudnev, Mohamed Elgharib, William Smith, Lingjie Liu, Vladislav Golyanik, and Christian Theobalt. Nerf for outdoor scene relighting. In ECCV, 2022. 1" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.469, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.092, + 0.892, + 0.147 + ], + "angle": 0, + "content": "[51] Vincent Sitzmann, Julien N.P. Martel, Alexander W. Bergman, David B. Lindell, and Gordon Wetzstein. Implicit neural representations with periodic activation functions. In NeurIPS, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.149, + 0.892, + 0.204 + ], + "angle": 0, + "content": "[52] Vincent Sitzmann, Semon Rezchikov, Bill Freeman, Josh Tenenbaum, and Fredo Durand. Light field networks: Neural scene representations with single-evaluation rendering. In NeurIPS, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.205, + 0.892, + 0.233 + ], + "angle": 0, + "content": "[53] Alvy Ray Smith. Color gamut transform pairs. In ACM TOG, 1978. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.234, + 0.892, + 0.29 + ], + "angle": 0, + "content": "[54] Pratul P. Srinivasan, Boyang Deng, Xiuming Zhang, Matthew Tancik, Ben Mildenhall, and Jonathan T. Barron. Nerv: Neural reflectance and visibility fields for relighting and view synthesis. In CVPR, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.291, + 0.892, + 0.332 + ], + "angle": 0, + "content": "[55] Mohammed Suhail, Carlos Esteves, Leonid Sigal, and Ameesh Makadia. Light field neural rendering. In CVPR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.334, + 0.892, + 0.375 + ], + "angle": 0, + "content": "[56] Cheng Sun, Min Sun, and Hwann-Tzong Chen. Direct voxel grid optimization: Super-fast convergence for radiance fields reconstruction. In CVPR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.377, + 0.892, + 0.433 + ], + "angle": 0, + "content": "[57] Cheng Sun, Guangyan Cai, Zhengqin Li, Kai Yan, Cheng Zhang, Carl Marshall, Jia-Bin Huang, Shuang Zhao, and Zhao Dong. Neural-pbir reconstruction of shape, material, and illumination. In ICCV, 2023. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.434, + 0.892, + 0.49 + ], + "angle": 0, + "content": "[58] Jiaming Sun, Xi Chen, Qianqian Wang, Zhengqi Li, Hadar Averbuch-Elor, Xiaowei Zhou, and Noah Snively. Neural 3D reconstruction in the wild. In SIGGRAPH Conference Proceedings, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.491, + 0.892, + 0.533 + ], + "angle": 0, + "content": "[59] Tiancheng Sun, Kai-En Lin, Sai Bi, Zexiang Xu, and Ravi Ramamoorthi. Nelf: Neural light-transport field for portrait view synthesis and relighting. In EGSR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.534, + 0.892, + 0.603 + ], + "angle": 0, + "content": "[60] Matthew Tancik, Pratul P. Srinivasan, Ben Mildenhall, Sara Fridovich-Keil, Nithin Raghavan, Utkarsh Singhal, Ravi Ramamoorthi, Jonathan T. Barron, and Ren Ng. Fourier features let networks learn high frequency functions in low dimensional domains. In NeurIPS, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.604, + 0.892, + 0.659 + ], + "angle": 0, + "content": "[61] Jiaxiang Tang, Hang Zhou, Xiaokang Chen, Tianshu Hu, Errui Ding, Jingdong Wang, and Gang Zeng. Delicate textured mesh recovery from nef via adaptive surface refinement. In ICCV, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.661, + 0.892, + 0.73 + ], + "angle": 0, + "content": "[62] Ayush Tewari, Ohad Fried, Justus Thies, Vincent Sitzmann, Stephen Lombardi, Kalyan Sunkavalli, Ricardo Martin-Brualla, Tomas Simon, Jason Saragih, Matthias Nießner, et al. State of the art on neural rendering. In CGF, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.731, + 0.892, + 0.788 + ], + "angle": 0, + "content": "[63] Ayush Tewari, Justus Thies, Ben Mildenhall, Pratul Srinivasan, Edgar Tretschk, Wang Yifan, Christoph Lassner, Vincent Sitzmann, Ricardo Martin-Brualla, Stephen Lombardi, et al. Advances in neural rendering. In CGF, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.789, + 0.892, + 0.844 + ], + "angle": 0, + "content": "[64] Marco Toschi, Riccardo De Matteo, Riccardo Spezialetti, Daniele De Gregorio, Luigi Di Stefano, and Samuele Salti. Relight my nerf: A dataset for novel view synthesis and relighting of real world objects. In CVPR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.845, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[65] Dor Verbin, Peter Hedman, Ben Mildenhall, Todd Zickler, Jonathan T Barron, and Pratul P Srinivasan. Ref-nerf: Structured view-dependent appearance for neural radiance fields. In CVPR, 2022. 1" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.092, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "4607" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.134 + ], + "angle": 0, + "content": "[66] Dongqing Wang, Tong Zhang, and Sabine Susstrunk. Nemto: Neural environment matting for novel view and relighting synthesis of transparent objects. In ICCV, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.136, + 0.47, + 0.177 + ], + "angle": 0, + "content": "[67] Jiaping Wang, Peiran Ren, Minmin Gong, John Snyder, and Baining Guo. All-frequency rendering of dynamic, spatially-varying reflectance. ACM TOG, 2009. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.179, + 0.469, + 0.233 + ], + "angle": 0, + "content": "[68] Peng Wang, Lingjie Liu, Yuan Liu, Christian Theobalt, Taku Komura, and Wenping Wang. Neus: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. In NeurIPS, 2021. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.235, + 0.469, + 0.275 + ], + "angle": 0, + "content": "[69] Shaofei Wang, Katja Schwarz, Andreas Geiger, and Siyu Tang. Arah: Animatable volume rendering of articulated human sdfs. In ECCV, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.278, + 0.469, + 0.332 + ], + "angle": 0, + "content": "[70] Yiming Wang, Qin Han, Marc Habermann, Kostas Dani-ilidis, Christian Theobalt, and Lingjie Liu. Neus2: Fast learning of neural implicit surfaces for multi-view reconstruction. In ICCV, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.334, + 0.469, + 0.376 + ], + "angle": 0, + "content": "[71] Zian Wang, Jonah Philion, Sanja Fidler, and Jan Kautz. Learning indoor inverse rendering with 3d spatially-varying lighting. In ICCV, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.378, + 0.469, + 0.432 + ], + "angle": 0, + "content": "[72] Zian Wang, Wenzheng Chen, David Acuna, Jan Kautz, and Sanja Fidler. Neural light field estimation for street scenes with differentiable virtual object insertion. In ECCV, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.434, + 0.469, + 0.502 + ], + "angle": 0, + "content": "[73] Zian Wang, Tianchang Shen, Jun Gao, Shengyu Huang, Jacob Munkberg, Jon Hasselgren, Zan Gojcic, Wenzheng Chen, and Sanja Fidler. Neural fields meet explicit geometric representations for inverse rendering of urban scenes. In CVPR, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.504, + 0.469, + 0.544 + ], + "angle": 0, + "content": "[74] Henrique Weber, Mathieu Garon, and Jean-François Lalonde. Editable indoor lighting estimation. In ECCV, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.547, + 0.469, + 0.602 + ], + "angle": 0, + "content": "[75] Haoqian Wu, Zhipeng Hu, Lincheng Li, Yongqiang Zhang, Changjie Fan, and Xin Yu. Nefii: Inverse rendering for reflectance decomposition with near-field indirect illumination. In CVPR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.604, + 0.469, + 0.672 + ], + "angle": 0, + "content": "[76] Liwen Wu, Rui Zhu, Mustafa B. Yaldiz, Yinhao Zhu, Hong Cai, Janarbek Matai, Fatih Porikli, Tzu-Mao Li, Manmohan Chandraker, and Ravi Ramamoorthi. Factorized inverse path tracing for efficient and accurate material-lighting estimation. In ICCV, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.675, + 0.469, + 0.715 + ], + "angle": 0, + "content": "[77] Qianyi Wu, Xian Liu, Yuedong Chen, Kejie Li, Chuanxia Zheng, Jianfei Cai, and Jianmin Zheng. Object-compositional neural implicit surfaces. In ECCV, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.717, + 0.469, + 0.758 + ], + "angle": 0, + "content": "[78] Qianyi Wu, Kaisiyuan Wang, Kejie Li, Jianmin Zheng, and Jianfei Cai. Objectsdf++: Improved object-compositional neural implicit surfaces. In ICCV, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.76, + 0.469, + 0.815 + ], + "angle": 0, + "content": "[79] Tong Wu, Jia-Mu Sun, Yu-Kun Lai, and Lin Gao. De-nerf: Decoupled neural radiance fields for view-consistent appearance editing and high-frequency environmental relighting. In SIGGRAPH ASIA Conference Proceedings, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.817, + 0.469, + 0.87 + ], + "angle": 0, + "content": "[80] Tong Wu, Jiaqi Wang, Xingang Pan, Xudong Xu, Christian Theobalt, Ziwei Liu, and Dahua Lin. Voxurf: Voxel-based efficient and accurate neural surface reconstruction. In ICLR, 2023. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.873, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[81] Yuanbo Xiangli, Linning Xu, Xingang Pan, Nanxuan Zhao, Anyi Rao, Christian Theobalt, Bo Dai, and Dahua Lin." + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "Bungeenerf: Progressive neural radiance field for extreme multi-scale scene rendering. In ECCV, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.121, + 0.892, + 0.175 + ], + "angle": 0, + "content": "[82] Yingyan Xu, Gaspard Zoss, Prashanth Chandran, Markus Gross, Derek Bradley, and Paulo Gotardo. Renerf: Relightable neural radiance fields with nearfield lighting. In ICCV, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.177, + 0.892, + 0.23 + ], + "angle": 0, + "content": "[83] Wenqi Yang, Guanying Chen, Chaofeng Chen, Zhenfang Chen, and Kwan-Yee K. Wong. Ps-nerf: Neural inverse rendering for multi-view photometric stereo. In ECCV, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.233, + 0.892, + 0.287 + ], + "angle": 0, + "content": "[84] Wenqi Yang, Guanying Chen, Chaofeng Chen, Zhenfang Chen, and Kwan-Yee K. Wong. S\\(^3\\)-nerf: Neural reflectance field from shading and shadow under a single viewpoint. In NeurIPS, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.288, + 0.892, + 0.342 + ], + "angle": 0, + "content": "[85] Yao Yao, Jingyang Zhang, Jingbo Liu, Yihang Qu, Tian Fang, David McKinnon, Yanghai Tsin, and Long Quan. Neilf: Neural incident light field for material and lighting estimation. In ECCV, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.344, + 0.892, + 0.399 + ], + "angle": 0, + "content": "[86] Lior Yariv, Yoni Kasten, Dror Moran, Meirav Galun, Matan Atzmon, Basri Ronen, and Yaron Lipman. Multiview neural surface reconstruction by disentangling geometry and appearance. In NeurIPS, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.4, + 0.892, + 0.439 + ], + "angle": 0, + "content": "[87] Lior Yariv, Jiatao Gu, Yoni Kasten, and Yaron Lipman. Volume rendering of neural implicit surfaces. In NeurIPS, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.441, + 0.892, + 0.482 + ], + "angle": 0, + "content": "[88] Alex Yu, Ruilong Li, Matthew Tancik, Hao Li, Ren Ng, and Angjoo Kanazawa. PlenOctrees for real-time rendering of neural radiance fields. In ICCV, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.484, + 0.892, + 0.538 + ], + "angle": 0, + "content": "[89] Chong Zeng, Guojun Chen, Yue Dong, Pieter Peers, Hongzhi Wu, and Xin Tong. Relighting neural radiance fields with shadow and highlight hints. In SIGGRAPH Conference Proceedings, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.539, + 0.892, + 0.593 + ], + "angle": 0, + "content": "[90] Jingyang Zhang, Yao Yao, Shiwei Li, Jingbo Liu, Tian Fang, David McKinnon, Yanghai Tsin, and Long Quan. Neilf++: Inter-reflectable light fields for geometry and material estimation. In ICCV, 2023. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.595, + 0.892, + 0.636 + ], + "angle": 0, + "content": "[91] Kai Zhang, Gernot Riegler, Noah Snavely, and Vladlen Koltun. Nerf++: Analyzing and improving neural radiance fields. arXiv preprint arXiv:2010.07492, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.637, + 0.892, + 0.691 + ], + "angle": 0, + "content": "[92] Kai Zhang, Fujun Luan, Qianqian Wang, Kavita Bala, and Noah Snavely. Physg: Inverse rendering with spherical gaussians for physics-based material editing and relighting. In CVPR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.692, + 0.892, + 0.734 + ], + "angle": 0, + "content": "[93] Kai Zhang, Fujun Luan, Zhengqi Li, and Noah Snavely. Iron: Inverse rendering by optimizing neural sdfs and materials from photometric images. In CVPR, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.735, + 0.892, + 0.802 + ], + "angle": 0, + "content": "[94] Xiuming Zhang, Sean Fanello, Yun-Ta Tsai, Tiancheng Sun, Tianfan Xue, Rohit Pandey, Sergio Orts-Escalano, Philip Davidson, Christoph Rhemann, Paul Debevec, et al. Neural light transport for relighting and view synthesis. In ACM TOG, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.804, + 0.892, + 0.858 + ], + "angle": 0, + "content": "[95] Xiuming Zhang, Pratul P. Srinivasan, Boyang Deng, Paul Debevec, William T. Freeman, and Jonathan T. Barron. Ner-factor: Neural factorization of shape and reflectance under an unknown illumination. In ACM TOG, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.86, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[96] Yuanqing Zhang, Jiaming Sun, Xingyi He, Huan Fu, Rongfei Jia, and Xiaowei Zhou. Modeling indirect illumination for inverse rendering. In CVPR, 2022. 2" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.956 + ], + "angle": 0, + "content": "4608" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.134 + ], + "angle": 0, + "content": "[97] Y. Zhang, Z. Hu, H. Wu, M. Zhao, L. Li, Z. Zou, and C. Fan. Towards unbiased volume rendering of neural implicit surfaces with geometry priors. In CVPR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.136, + 0.468, + 0.175 + ], + "angle": 0, + "content": "[98] Quan Zheng, Gurprit Singh, and Hans-Peter Seidel. Neural relightable participating media rendering. In NeurIPS, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.178, + 0.468, + 0.248 + ], + "angle": 0, + "content": "[99] Jingsen Zhu, Fujun Luan, Yuchi Huo, Zihao Lin, Zhihua Zhong, Dianbing Xi, Rui Wang, Hujun Bao, Jiaxiang Zheng, and Rui Tang. Learning-based inverse rendering of complex indoor scenes with differentiable monte carlo raytracing. In SIGGRAPH ASIA Conference Proceedings, 2022. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.248 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.517, + 0.956 + ], + "angle": 0, + "content": "4609" + } + ] +] \ No newline at end of file diff --git a/2024/ESR-NeRF_ Emissive Source Reconstruction Using LDR Multi-view Images/2c6d481f-7a23-410c-a62a-93255bbc656e_origin.pdf b/2024/ESR-NeRF_ Emissive Source Reconstruction Using LDR Multi-view Images/2c6d481f-7a23-410c-a62a-93255bbc656e_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..d97e79ae73edb9cc12c42f9c97eec53eac892a7a --- /dev/null +++ b/2024/ESR-NeRF_ Emissive Source Reconstruction Using LDR Multi-view Images/2c6d481f-7a23-410c-a62a-93255bbc656e_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9d9976c61d074575173e248f257da09fc33dbd7b2c94faf7edcfee54eae18796 +size 4820301 diff --git a/2024/ESR-NeRF_ Emissive Source Reconstruction Using LDR Multi-view Images/full.md b/2024/ESR-NeRF_ Emissive Source Reconstruction Using LDR Multi-view Images/full.md new file mode 100644 index 0000000000000000000000000000000000000000..21259291106b91f99b9036f3a5170a897be27298 --- /dev/null +++ b/2024/ESR-NeRF_ Emissive Source Reconstruction Using LDR Multi-view Images/full.md @@ -0,0 +1,466 @@ +# ESR-NeRF: Emissive Source Reconstruction Using LDR Multi-view Images + +Jinseo Jeong $^{1}$ Junseo Koo $^{1}$ Qimeng Zhang $^{2}$ Gunhee Kim $^{1}$ Seoul National University Korea University + +jinseo@vision.snu.ac.kr, junseo.koo@vision.snu.ac.kr, zoe1024@korea.ac.kr, gunhee@snu.ac.kr +https://jinseo.kr/ESR-NeRF + +# Abstract + +Existing NeRF-based inverse rendering methods suppose that scenes are exclusively illuminated by distant light sources, neglecting the potential influence of emissive sources within a scene. In this work, we confront this limitation using LDR multi-view images captured with emissive sources turned on and off. Two key issues must be addressed: 1) ambiguity arising from the limited dynamic range along with unknown lighting details, and 2) the expensive computational cost in volume rendering to backtrace the paths leading to final object colors. We present a novel approach, ESR-NeRF, leveraging neural networks as learnable functions to represent ray-traced fields. By training networks to satisfy light transport segments, we regulate outgoing radiances, progressively identifying emissive sources while being aware of reflection areas. The results on scenes encompassing emissive sources with various properties demonstrate the superiority of ESR-NeRF in qualitative and quantitative ways. Our approach also extends its applicability to the scenes devoid of emissive sources, achieving lower CD metrics on the DTU dataset. + +# 1. Introduction + +Extensive research has focused on reconstructing 3D object structures [16, 43, 47, 86], material properties [18, 29, 65], and lighting [15, 33, 34, 74, 79] from 2D images, applicable across domains including 3D graphics and augmented reality [62, 63, 69, 72]. This endeavor not only facilitates the creation of life-like virtual objects but also streamlines the process of scene manipulation [27, 58, 61, 73]. Recent advancements [24, 30, 36, 71] have built on Neural Radiance Fields (NeRF) [40] successes in novel view synthesis [3, 4, 45, 81, 91]. Significant progress in relighting [37, 38, 50] has facilitated scene editing via manipulating the reconstructed light sources. However, existing methods predominantly deal with the scenes lit by distant sources, like environment maps or collocated flashlights. Notably, NeRF-based inverse rendering has yet to + +![](images/be789cdd189d4bd05cb65e9e4d54a65edb1901096f20ac60382ca27daa97eb57.jpg) +255 pixel values + +![](images/80db8dedaaea4ce3ac6829a80e9c359cab6dd83422264462eff09725a33280fb.jpg) +Information loss + +![](images/7843b6cfe8d393de564ff362e62133cbda7707f238fa4ecf3cfad14bd5dc81cc.jpg) +Threshold 0.65 + +![](images/91a2824c953bde7a9f164294590355335cd4738daa64b0b63b095bbd1f12496f.jpg) +Threshold 0.75 + +![](images/cb90e4c96d41e556988afaa3833d0fcca7b226e1c2055d7f3b263ab415c66fe8.jpg) + +![](images/738b4216d6118c90695254c7dfc9ec18e1c04b9e7fa8ab91a620794a68fb5fc8.jpg) +Edit emission + +![](images/aa37142ac7b770d8a86342081c99d6da7914dbe9cc2bbeeed215360a535f0e20.jpg) +Desired + +![](images/c1f1c8ea760cb6b452e4444e5270a3c25bd6d3ebc5a978c6ad15c00985bc2a91.jpg) +Unrealistic +Figure 1. Challenges posed by emissive sources in LDR images. Green, red, and blue in thresholded images respectively show true positives, false negatives, and false positives of source identification. Thresholding values are scaled down divided by 255. The contrast between light on and off pixel values is more pronounced in surroundings than emissive sources. Inaccurate reconstruction of emissive sources disrupts scene editing, causing reflection areas to stay static while only the source colors change. + +consider scenes with multiple emissive sources, a common real-world illumination condition. + +Emissive sources in a scene introduce critical challenges: (i) ambiguity in decomposing scene components and (ii) high computational costs for analyzing the causes of pixel colors. This ambiguity stems from difficulties in identifying emissive source regions, as illustrated in Fig. 1. Contrary to prior setups [6-8, 66, 85, 93], we allow the possibility of numerous emissive sources throughout the scene. In standard photographs with pixel values from 0 to 255, the distinction between emissive sources and nearby reflection areas is challenging. As shown in Fig. 1, relying solely on pixel value thresholding is insufficient for differentiating between emissive sources and their reflections. Naive inverse path tracing is impractical, due to the computational costs rising exponentially with the number of ray bounces in volume rendering. This can cause inaccuracy in emissive source reconstruction, yielding unrealistic illumination in reflective areas as users manipulate emissive sources. + +To address these challenges, we introduce ESR-NeRF + +(Emissive Sources Reconstructing NeRF), a novel approach capable of reconstructing any number of emissive sources by progressively discovering reflection areas. We assume that the scenes are observed in two lighting conditions: one with all emissive sources active and the other with them inactive. Our approach utilizes neural networks as learnable functions for representing ray-traced fields. By training networks to satisfy each light transport segment, we sidestep the computational overhead of ray tracing associated with ray bounces. In this work, we exclusively use low dynamic range (LDR) images, setting us apart from prior mesh-based methods that rely on high dynamic range (HDR) images [2, 19, 48, 76]. + +Our experiments encompass synthetic and real scenes, ranging from single to multiple lighting configurations with complex reflections. The scenes vary in light source counts, color, and intensity. Qualitative and quantitative evaluations show ESR-NeRF's superiority over state-of-the-art NeRF-based re-lighting methods. Furthermore, Chamfer Distance (CD) metrics on the DTU dataset [23] indicate ESR-NeRF's competitive performance in scene reconstruction, even without emissive sources. + +We summarize our contributions as follows. + +1. Our work presents the first NeRF-based inverse rendering that can deal with the scenes with any number of emissive sources, challenging the distant light assumption of previous research. +2. Unlike existing mesh-based methods relying on HDR images, we use LDR images for the first time, overcoming the poor representation of emissive sources. +3. We provide a benchmark dataset designed to evaluate the performance of emissive source reconstruction. +4. Our method is applicable to the scenes with or without emissive sources, achieving superior mesh reconstruction results on the DTU dataset. + +# 2. Related work + +Neural Rendering. Advancements in implicit representations [51, 60] and volume rendering [39] have significantly enhanced neural rendering capabilities, enabling the reconstruction of scene components from 2D images. One of the key directions is mesh extraction [44, 70, 77, 78, 80, 97], with methods like NeuS [68] and VolSDF [87] utilizing signed distance function (SDF) values for volume rendering. Recently, the efficient computation of volume rendering has become a focal point due to the substantial computational cost associated with network inference for ray color calculation [41, 49, 88]. Several methods propose to directly predict ray color using the 4D light fields concept [1, 52, 55] or leveraging voxel grids for fast inference of spatial features [5, 11, 12, 14, 31, 56]. NeuralRadiosity [17] shares similarity with our method, as it predicts raytraced values instead of explicitly tracing individual rays. + +
VoxurfTensoIRPath TracingESR-NeRF
Big Onn · d(n · d)b+1n2 · d
Indirect illuminationx✓✓✓
BRDF decompositionx✓✓✓
Emissive source controlxx✓✓
+ +Table 1. Computational cost comparison for inverse rendering methods. $n$ is the number of sampled points along a ray, $d$ is the number of scattering rays, and $b$ is the number of ray bounces. + +However, they primarily focus on calculating the final object color when all scene information is available. In contrast, our inverse rendering approach aims to reconstruct emissive sources within a scene, addressing the ambiguities introduced by their presence in LDR images. + +Inverse Rendering. A growing emphasis revolves around the decomposition of materials represented by spatially varying bidirectional reflectance distribution functions (SVBRDF) [46, 83, 99]. To lessen the computational burden in inverse rendering [25, 54, 96, 98], several methods have adopted neural networks as lookup tables [9] or computational caches [54, 90, 95]. While NeRV [54] utilizes caching visibility and $\mathrm{NeILF++}$ [90] adopts caching surface point radiance with the inter-reflection loss for incident radiance, our method diverges by focusing on tracing radiance origins. Specifically, we aim to identify emissive sources within a scene, moving beyond the simplification of incident radiance calculations. Several methods rely on diverse known lighting configurations to exploit variations in object appearances [59, 64, 84, 89]. Toggling emissive sources on and off resembles the common one-light-at-a-time (OLAT) technique, as seen in NLT [94] and ReN-eRF [82]. However, our setting does not need to know light source properties and to toggle lights individually. Instead, we allow for toggling all lights together. Recent works have also jointly reconstruct the mesh, materials, and lighting [20, 35, 42, 57]. They tackle with images captured under a single unknown lighting condition [92, 95], assuming that radiance already encodes global illumination [75, 96]. However, they confine to the scenes illuminated by far-distant lights, constrained to an 8-bit color spectrum. Our work considers the presence of multiple emissive sources within a scene captured in LDR images, questioning the prevailing notion that radiance fields trained with the image rendering loss faithfully represents global illumination. While some methods [2, 19, 32, 48, 76] deal with the scenes featuring emissive sources, they work outside the volume rendering framework and depend on HDR input images, assuming prior knowledge of scene geometry. + +# 3. Preliminaries + +Surface Representation. Analogous to NeRF [40], neural network $f_{\theta}$ predicts SDF values at arbitrary 3D spatial locations. NeuS [68] integrates surface representa + +![](images/114f18d634099b544fc56a064202769a983a7f934e4831d24cb881fa79374340.jpg) +On Images + +![](images/9a472dcd746a19a3e978d986bc74651228309b044ee968c37da34649cf9f85fb.jpg) +Off Images + +![](images/7c30826d9f9f5e59b8b5f0d05742991f50e3d6adc4853cb0fb80803759bb1545.jpg) +Multi-view Input +Training + +![](images/b6bae7cda4ddb5a80e5c74766f89f303c299dc9ed7c045fe30d0f9548596a16f.jpg) +Application +Figure 2. The pipeline of emissive source reconstruction. Given LDR images with emissive sources on and off, scene components are reconstructed by synthesizing training images and enforcing LTS requirements. Emissive sources are progressively refined via categorizing training rays into uncertain and certain groups. The scenes can be edited with new lighting conditions using reconstructed emissive sources. + +tion into volume rendering using the SDF-based opacity $\rho (x) = \max (\frac{-\frac{d\Phi_s}{dx}(f(x))}{\Phi_s(f(x))},0)$ . Here $\Phi_s(x) = (1 + e^{-sx})^{-1}$ is the sigmoid function where $s$ controls the sharpness of surfaces. The color of a ray can be calculated as + +$$ +\hat {C} (r) = \int_ {0} ^ {\infty} T (r (t)) \rho (r (t)) L _ {o} (r (t), \omega_ {o}) d t, \tag {1} +$$ + +where $\hat{C}(r)$ denotes the predicted ray color, $r(t; c, \omega_o) = c - t \cdot \omega_o$ is the ray with camera center $c$ along direction $\omega_o$ , $T(r(t)) = \exp \left( \int_0^t -\rho (r(u))du \right)$ is the transmittance, and $L_{o}(r(t), \omega_{o})$ is the outgoing radiance. Henceforth, we use $x$ to denote a point in $r(t; c, \omega_o)$ for notational simplicity. + +Light Transport in Volume Rendering. Extracting light sources necessitates analyzing the causes affecting the final ray colors. Kajiya's rendering equation [26] factorizes the outgoing radiance $L_{o}(x,\omega_{o})$ into emission and reflections: + +$$ +L _ {o} (x, \omega_ {o}) = E (x) + \int_ {\Omega} L _ {i} (x, \omega_ {i}) R (x, \omega_ {o}, \omega_ {i}; b) d \omega_ {i}, \tag {2} +$$ + +where $E(x)$ is the emission, $R(x,\omega_o,\omega_i;b)$ represents the SVBRDF parametrized by parameters $b$ with Lambert cosine multiplied, and $L_{i}(x,\omega_{i})$ is the incident radiance. In volume rendering, computing the incident radiance at point $x$ is akin to evaluating Eq. 1, with $x$ serving as the camera center. By iteratively factorizing the outgoing radiance in the incident radiance, the contribution of a path length $i$ for a pixel can be decomposed as in Eq. 3, where $\mathcal{H}_i = \Pi_{j=1}^{i-1}T(x_j)\rho(x_j)R(x_j,\omega_{j-1},\omega_j)$ is the path throughput, $S(\omega_i)$ is the environment map strength in direction $\omega_i$ , and $V(x,\omega_i) = \exp\left(\int_0^\infty -\rho(r(u;x_i,-\omega_i))du\right)$ is the visibility of the environment map at point $x$ along direction $\omega_i$ : + +$$ +\begin{array}{l} P _ {i} = \int_ {l _ {1}} \int_ {\Omega} \dots \int_ {l _ {i - 1}} \int_ {\Omega} \left(\int_ {l _ {i}} T (x _ {i}) \rho (x _ {i}) E (x _ {i}) d t _ {i} + \right. \tag {3} \\ S \left(\omega_ {i - 1}\right) V \left(x _ {i - 1}, \omega_ {i - 1})\right) \mathcal {H} _ {i} d t _ {1} d \omega_ {1} \dots d t _ {i - 1} d \omega_ {i - 1}. \\ \end{array} +$$ + +Extending the analysis to longer light paths, or equivalently, increasing the number of ray bounces, leads to exponential growth in computation complexity. This poses a challenge when attempting to decompose the influence of unknown emissive sources, as their ability to produce strong reflections makes ignoring indirect illumination infeasible. + +# 4. Methodology + +None of the previous works address the reconstruction of emissive sources from LDR multi-view images. Sec. § 4.1 through § 4.5 detail our method, ESR-NeRF, which reconstructs emissive sources without prior knowledge of scene geometry, materials, or lighting specifics (including their location, number, or colors). We also show how these reconstructed sources can be used for scene editing in § 4.5. + +# 4.1. Learnable Tone-mapper + +Throughout the paper, we use $\mathcal{R}$ to represent camera rays, $C$ for pixel values, and a binary flag $\mathbb{I}$ to indicate whether an image is captured with emissive sources on or off. + +To extract HDR values from LDR images, we employ the softplus activation for outgoing radiance prediction and apply a clipping and gamma function $\tau$ [21] for the rendering loss such that $\hat{C}_{\tau}(r) = \tau (\hat{C} (r))$ . Unlike previous NeRF-based works [25, 37, 54, 57] that limit radiance to the range of [0, 1], our approach allows for any positive radiance values. Yet, it creates difficulties in differentiating between the surface weight $T(x)\rho (x)$ and the magnitude of radiance value $L_{o}(x,\omega_{o})$ , since it allows for the possibility of assigning extreme radiance to the points with low surface weights to render same ray colors. Such ambiguity poses challenges, particularly in dark and high-contrast scenes, aggravating surface reconstruction (see Fig. 3). To address this, we introduce a learnable tone-mapper $m_{\theta}:\mathbb{R}_{+}^{3}\to [0,1]^{3}$ , that takes positionally encoded HDR linear values as input: + +![](images/a3c9e215d5e892eb640ca4cdf9957db220e3027cb39b41a155a9983267ed6771.jpg) + +![](images/9e5cd54c95fe67b079c517c178a5061f7e5223ce6de42a3cbda1377eee353d56.jpg) +(a) Image + +![](images/31d8bf7ded52d96af3e1dc3ff1a5af6af66fa47b75a09972fe4ec68bfc51cd8a.jpg) + +![](images/0614aa82aa0d17183c377cbc29216fbf0dcdda4bd9678bb34c6b2ffa50354f00.jpg) +(b) $\mathrm{w / o}m_{\theta}$ + +![](images/cc73d8dc0bd607c86cea402aa62c9d82edeee5bf331f29c17f1c84ba3f8ad670.jpg) + +![](images/4367bb8019d50bddd926a6b6803b02803e6c00e46538ddaba63eda88a8c5eefe.jpg) +(c) ESR-NeRF +Figure 3. Reconstructed surfaces with the learnable tone-mapper. + +$$ +\hat {C} _ {m _ {\theta}} (r) = \int_ {0} ^ {\infty} T (x) \rho (x) m _ {\theta} \left(L _ {o} \left(x, \omega_ {o}\right)\right) d t, \tag {4} +$$ + +$$ +L _ {o} (x, \omega_ {o}) = L _ {o} ^ {S} (x, \omega_ {o}) + L _ {o} ^ {E} (x, \omega_ {o}) \cdot \mathbb {I}, \tag {5} +$$ + +where $L_{o}^{S}(x,\omega_{o})$ is radiance when emissive sources are turned off, while $L_{o}^{E}(x,\omega_{o})$ stands for radiance added to the scene by emissive sources. Our rendering loss is then formulated as follows, with $\lambda_{\tau}$ as a hyper-parameter: + +$$ +\mathcal {L} _ {\text {r e n d e r}} = \sum_ {r \in \mathcal {R}} \left(\| C (r) - \hat {C} _ {m _ {\theta}} (r) \| _ {2} ^ {2} + \lambda_ {\tau} \| C (r) - \hat {C} _ {\tau} (r) \| _ {2} ^ {2}\right). \tag {6} +$$ + +# 4.2. Learning of Light Transport Segments + +The computational complexity of object appearance analysis in volume rendering is notably high, as shown in Eq. 3. We take an alternative approach by leveraging neural networks to represent ray-traced fields, rather than explicitly tracing every rays. Our distinct contribution to inverse rendering lies in precise adjustment of radiance. Specifically, we impose constraints on the predicted radiance to satisfy each light transport segments. The light transport segments (LTS) loss, $\mathcal{L}_{lts}$ , plays a pivotal role in our method: + +$$ +\mathcal {L} _ {l t s} ^ {S} = \sum_ {x, \omega_ {o}} \| L _ {o} ^ {S} (x, \omega_ {o}) - \hat {L} _ {o} ^ {S} (x, \omega_ {o}) \| _ {2} ^ {2}, \tag {7} +$$ + +$$ +\mathcal {L} _ {l t s} ^ {E} = \sum_ {x, \omega_ {o}} \| L _ {o} ^ {E} (x, \omega_ {o}) - \hat {L} _ {o} ^ {E} (x, \omega_ {o}) \| _ {2} ^ {2}, \tag {8} +$$ + +$$ +\hat {L} _ {o} ^ {S} (x, \omega_ {o}) = \int_ {\Omega} \underbrace {S (\omega_ {i}) V (x , \omega_ {i}) R (x , \omega_ {o} , \omega_ {i})} _ {\text {d i r e c t i l l u m i n a t i o n b y a n e v i r o n m e n t m a p}} d \omega_ {i} + +$$ + +$$ +\int_ {\Omega} \int_ {0} ^ {\infty} \underbrace {T (x ^ {\prime}) \rho (x ^ {\prime}) L _ {o} ^ {S} (x ^ {\prime} , - \omega_ {i}) d t ^ {\prime} R (x , \omega_ {o} , \omega_ {i})} _ {\text {i n d r e c t i l l u m i n a t i o n b y a n e v i r o n m e n t m a p}} d \omega_ {i}. +$$ + +$$ +\hat {L} _ {o} ^ {E} (x, \omega_ {o}) = \underbrace {E (x)} _ {\text {e m i s s i o n}} + +$$ + +$$ +\int_ {\Omega} \int_ {0} ^ {\infty} \underbrace {T \left(x ^ {\prime}\right) \rho \left(x ^ {\prime}\right) L _ {o} ^ {E} \left(x ^ {\prime} , - \omega_ {i}\right) d t ^ {\prime} R \left(x , \omega_ {o} , \omega_ {i}\right)} _ {\text {d i r e c t ＆ i n d i r c t i l l u m i n a t i o n b y e m i s s i v e s o u r c e s}} d \omega_ {i}. \tag {10} +$$ + +We ensure consistency between the radiance directly predicted by the network $L_{o}(x,\omega_{o})$ and the radiance achievable based on the scene context $\hat{L}_{o}(x,\omega_{o})$ . Previous approaches have focused on matching $\hat{L}_{o}(x,\omega_{o})$ to training views, overlooking the relations to $L_{o}(x,\omega_{o})$ . This hinders the restoration of HDR radiance by supervising scene components to LDR training views. In contrast, our LTS loss enables volumetric energy transfer of radiance, adjusting outgoing radiance based on their interrelations. + +To implement this concept, we train six dedicated networks for SDF $f(x)$ , SVBRDF parameters $b(x)$ , emission $E(x)$ , environment map $S(\omega_{i})$ , outgoing radiances $L_{o}^{S}(x,\omega_{o})$ and $L_{o}^{E}(x,\omega_{o})$ , to adhere to these LTS requirements. For the environment map, we represent it using 48 Spherical Gaussians [67]: $\sum_{k=1}^{M}\mu_{k}e^{\lambda_{k}(\omega_{i}\cdot\xi_{k}-1)}$ , followed by the softplus activation. $\mu\in\mathbb{R}^{3},\lambda\in\mathbb{R}_{+}$ , and $\xi\in\mathbb{S}^{2}$ respectively denote the lobe amplitude, sharpness, and axis. + +# 4.3. Progressive Discovery of Reflection Areas + +![](images/e1575f4b7ab33fc445b5dd92db9dcbcce6315d95324ef4a2ac79713e6f3f9f93.jpg) +Figure 4. Left: Image with active emissive sources. Right: Identified emissive sources w/o progressive discovery of reflection areas. + +![](images/93ca78abdc92500ef1608c9707574b5b883aec36de75958d071a0a5a87d6809b.jpg) + +Relying solely on LTS is insufficient for addressing ambiguity arising from low pixel values of emissive sources and intense reflections in adjacent regions, often leading to confusion between emission and reflection. The right image in Fig. 4 shows self-emitting objects restored with the + +naive LTS loss. While emissive sources are small, large areas affected by them are also identified as emissive sources. We propose a reflection-aware progressive approach for precise identification of emissive sources. By leveraging LTS learning, we extend the regions that can be regarded as reflection areas. Fig. 5 illustrates our progressive algorithm. + +Reflection-Aware Emission Refinement. Since surface points are unknown and are updated during learning, we opt to utilize rays rather than surface points. This process involves categorizing training rays into two groups: uncertain $(\mathcal{R}^U)$ and certain $(\mathcal{R}^C)$ . The certain group contains the rays confidently identified as reflection, aiding the transfer of radiance energy to nearby points. For the points in the certain group, we use the Eq. 11 instead of Eq. 10 to exclusively attribute outgoing radiances to reflections. Satisfying the LTS loss on the certain group results in adjusting the outgoing radiances of influential points, as illustrated in Fig. 5(a): + +$$ +\hat {L} _ {o} ^ {E} \left(x, \omega_ {o}\right) = \int_ {\Omega} \int_ {0} ^ {\infty} T \left(x ^ {\prime}\right) \rho \left(x ^ {\prime}\right) L _ {o} ^ {E} \left(x ^ {\prime}, - \omega_ {i}\right) d t ^ {\prime} R \left(x, \omega_ {o}, \omega_ {i}\right) d \omega_ {i}. \tag {11} +$$ + +The uncertain group includes the rays indicating the areas that are undetermined yet as reflection or emission. Using Eq. 12 to compute $\hat{L}_o^E (x,\omega_o)$ , this group adjusts emis + +sions $E(x)$ based on the radiance updates by the certain group, where "sg" represents the stop-gradient: + +$$ +\begin{array}{l} \hat {L} _ {o} ^ {E} (x, \omega_ {o}) = E (x) + \\ \left. \operatorname {s g} \left(\int_ {\Omega} \int_ {0} ^ {\infty} T \left(x ^ {\prime}\right) \rho \left(x ^ {\prime}\right) L _ {o} ^ {E} \left(x ^ {\prime}, - \omega_ {i}\right) d t ^ {\prime} R \left(x, \omega_ {o}, \omega_ {i}\right) d \omega_ {i}\right). \right. \tag {12} \\ \end{array} +$$ + +As shown in Fig. 5(b), this leads to increased emissions for the regions whose radiances are adjusted to account for the reflections in the certain group. Conversely, emissions decrease for the regions where there is little change in outgoing radiance, but incident radiances are increased by surrounding influential points. + +Ray Group Management. As emissions and radiances are adjusted, the groups are dynamically updated at predefined training intervals through the following process. Within the uncertain group, we evaluate the expected emission strength of rays, retaining only those above a threshold $k_{i}$ . Rays below this threshold are then merged to the certain group: + +$$ +\mathcal {R} _ {i} ^ {U} = \left\{r \mid \max _ {R G B} \left(\int_ {0} ^ {\infty} T (x) \rho (x) E (x) d t\right) \geq k _ {i}, r \in \mathcal {R} _ {i - 1} ^ {U} \right\}, \tag {13} +$$ + +$$ +\mathcal {R} _ {i} ^ {C} = \left(\mathcal {R} _ {i - 1} ^ {U} - \mathcal {R} _ {t} ^ {U}\right) \cup \mathcal {R} _ {i - 1} ^ {C}. \tag {14} +$$ + +Subsequently, newly added rays to the certain group can be used to localize influential points and update their outgoing radiances. This iterative process progressively refines the separation between reflective and emissive regions, attaining more accurate identification of emissive sources. + +LTS Loss Decomposition. The LTS loss, as detailed in Eq. 15, can be decomposed using a stop-gradient operation to refine the adjustment process. + +$$ +\mathcal {L} _ {l t s} ^ {E} = \sum_ {x, \omega_ {o}} \left(\lambda_ {l} \| \operatorname {s g} \left(L _ {o} ^ {E} (x, \omega_ {o})\right) - \hat {L} _ {o} ^ {E} (x, \omega_ {o}) \| _ {1} + \right. \tag {15} +$$ + +$$ +\lambda_ {r} \| L _ {o} ^ {E} (x, \omega_ {o}) - \operatorname {s g} (\hat {L} _ {o} ^ {E} (x, \omega_ {o})) \| _ {1}). +$$ + +We prioritize $\lambda_{l}$ to enhance the update of scene context, affecting other points' radiance given the predicted $L_{o}(x,\omega_{o})$ . $\lambda_{r}$ prevents severe deviation of every $L_{o}(x,\omega_{o})$ within the current scene context. This aligns with our focus on HDR source reconstruction from LDR images, addressing under-represented information in training data. + +# 4.4. Training Details + +We employ the Voxurf architecture [80] as backbone and adopt the simplified Disney BRDF model [10] for SVBRDF representation, with parameters including base color $\in [0,1]^3$ , roughness $\in [0,1]$ , and metallic $\in [0,1]$ . The learnable tone-mapper, structured as a two-layer MLP, is utilized for the rendering loss only. Initially, we pre-train our networks using the rendering loss, subsequently integrating the basic LTS loss (Eq. 7 and Eq. 8) into our training regimen. + +![](images/c0f1cde92af7bc78d655d0417f1087a8b316296c02b105f38ca8940638e37e8b.jpg) +(a) LTS on $R^C$ + +![](images/12dab668070569d1475cefec26d1388d4bb0c1e5315387551bfbfa752f6ab277.jpg) +(b) LTS on $R^U$ +Figure 5. Illustration of the progressive emissive source reconstruction with reflection awareness. Gray color represents the areas belonging to the certain group, while the red (emissive sources) and orange (their reflections) areas belong to the uncertain group. + +![](images/2ea8217914c690e071b09f71ed17e37be0f69dcf599092c298054181d0ac1531.jpg) +(c) Ray Group Managing + +This phase transitions to the reflection-aware progressive training scheme, where we adopt the $\ell_1$ loss due to its empirical stability in refining emissive source reconstruction. We use a smoothing regularization to promote local consistency in normals, BRDFs, and emissions. To ensure view-consistent labeling of 3D points as either reflective or emissive, we implement the emission suppression loss for points belonging to the certain group: + +$$ +\mathcal {L} _ {\text {s u p p}} ^ {E} = \sum_ {r \in R _ {t} ^ {C}} \| \int_ {0} ^ {\infty} T (x) \rho (x) E (x) d t \| _ {2} ^ {2}, \tag {16} +$$ + +The threshold $k_{i}$ linearly increases with each time step $t$ , utilizing a grid search within a range of $[10^{-3}, 10^{-5}]$ to find the slope. We construct mini-batches via stratified sampling within each group. For a detailed description of our training procedure, please refer to Appendix. + +# 4.5. Scene Editing + +Reconstructed emissive sources enable scene editing; users select emissive sources using binary masks $M_{j=1\dots N}$ and specify lighting conditions using colors $c_{j=1\dots N}$ and intensities $i_{j=1\dots N}$ within the HSV color space [53]. + +We identify the rays in the uncertain group that match $M$ by projecting expected surface points $p$ of the rays onto the camera with the pose $\mathbf{R}|\mathbf{t}$ : + +$$ +p = \int_ {0} ^ {\infty} T (x) \rho (x) x d t, \tag {17} +$$ + +$$ +\mathbb {I} _ {j} ^ {\text {h i t}} (x) = \operatorname {i n t e r p} \left(M _ {j}, p ^ {\prime}\right) > 0, \text {w h e r e} p ^ {\prime} = \mathbf {K} [ \mathbf {R} | \mathbf {t} ] [ p | 1 ] ^ {T}. \tag {18} +$$ + +For the rays satisfying $\mathbb{I}_j^{hit}(x)$ , we apply the designated lighting conditions. The new emission values are computed by substituting the original hue (H) and saturation (S) of $E(x)$ with the user-specified color $c_{j}$ and adjusting the value (V) of $v(x)$ with the new intensity $i_j$ : + +$$ +E (x) = \operatorname {h s v} _ {-} \operatorname {t o} _ {-} \operatorname {r g b} ([ c _ {j} | (v (x) \times i) ]) \cdot \mathbb {I} _ {j} ^ {h i t} + E (x) \cdot \neg \mathbb {I} _ {j} ^ {h i t}. \tag {19} +$$ + +These modifications influence scene appearance by optimizing the loss in Eq. 20. During this process, all networks, except for $L_{o}^{E}(x,\omega_{o})$ , are frozen: + +$$ +\mathcal {L} _ {\text {e d i t}} = \sum_ {x, \omega_ {o}} \| L _ {o} ^ {E} (x, \omega_ {o}) - \operatorname {s g} \left(\hat {L} _ {o} ^ {E} (x, \omega_ {o})\right) \| _ {2} ^ {2}. \tag {20} +$$ + +
LegoGiftWhite colored BookCubeBillboardBallsVivid colored
IoUMSEIoUMSEIoUMSEIoUMSEIoUMSEIoUMSEIoUMSEIoUMSEIoUMSEIoUMSEIoUMSE
Twins0.2220.190.498.590.633.910.9531.830.691.120.900.060.256.960.246.090.552.630.9510.640.090.750.830.04
NeLF++0.4320.880.079.380.954.640.9332.670.011.950.910.800.307.650.096.860.953.360.9411.490.021.570.920.78
TensoIR0.7120.130.158.550.953.870.9531.730.761.110.950.050.336.930.156.050.952.590.9610.600.770.740.950.03
ESR-NeRF0.818.380.603.490.961.190.9717.870.840.460.950.040.515.480.592.500.960.510.977.940.880.260.940.03
+ +# 5. Experiments + +We assess ESR-NeRF in reconstructing emissive sources by focusing on both identification and intensity restoration. To showcase its effectiveness, we conduct a range of experiments, including scene editing, ablation studies, illumination decomposition, and surface reconstruction, providing both quantitative and qualitative results. + +# 5.1. Experiment Settings + +We curate 6 diverse synthetic scenes, each with 200 training images evenly distributed between on and off lighting conditions. To evaluate the robustness of our approach against light colors, we consider two distinct settings of white colored and vivid colored emissive sources, resulting in a total of 12 scenes. The vivid colors are selected with full saturation in the HSV color space. We measure source identification and radiance reconstruction using IoU and MSE metrics on novel view test images, comparing against ground truth data from Blender-rendered emission masks and EXR files. The emission strengths, the maximum EXR file values, range from 2 to 200. For quantitative scene editing evaluation, we alter the white-colored sources to various colors—red, green, blue, cyan, magenta, yellow—and adjust intensities to half or double their original values. Qualitative results include scene editing for vividly colored sources and real scenes captured with a Fuji 100s camera using Philips smart bulbs as emissive sources. Quantitative assessments are based on 50 test images from novel camera poses, except for MSE measured for 25 test + +![](images/d007f314138098088d81eb12ba8af87dd4e8655fb8e78e7e43cb2e7a13fce647.jpg) +Figure 6. Comparison of identified emissive sources. ESR-NeRF excels through the reflection-aware progressive refinement. + +Table 2. Results of emissive source identification. ESR-NeRF outperforms state-of-the-art re-lighting methods in reconstructing emissive sources, regardless of their color. The IoU measures the source area identification (a higher value is better), and the MSE quantifies the difference between reconstructed images and HDR ground truth images (a lower value is better). + +
NVNV + INV+ CNV + I + C
PSNRLPIPSPSNRLPIPSPSNRLPIPSPSNRLPIPS
Twins36.520.014127.910.025231.020.025228.210.0310
NeRF-W36.440.014224.770.0417----
NeILF++24.400.055624.710.057924.060.075023.240.0770
TensoIR38.040.010327.280.041826.360.050525.180.0531
PaletteNeRF33.660.023323.270.048324.440.064622.580.0703
ESR-NeRF38.790.008329.990.019331.730.019631.630.0199
+ +Table 3. Scene editing results. NV: novel view synthesis, I: intensity editing, and C: color editing. A higher PSNR or lower LPIPS value is better. + +images. We denote the best performance with blue and the second-best with green. Additionally, we utilize the DTU dataset [23] to evaluate ESR-NeRF's performance in surface reconstruction tasks where emissive sources are absent. + +Baselines. We select two state-of-the-art re-lighting methods, TensoIR [25] and NeILF++ [90], that do not require prior lighting information. For thorough evaluation, we also implement a simple method, Twins, where separate models are trained under light on and off conditions. The Twins utilize the radiance discrepancies between the on and off models to distinguish and adjust emissive sources. For scene editing, we add NeRF-W [38] and PaletteNeRF [28] as baselines. Both NeRF-W and Twins adopt the Voxurf [80] architecture for fair comparison. For methods unable to individually control emissive sources, all sources are adjusted together to match the last lighting condition by a user. For the DTU dataset, we include state-of-the-art surface reconstruction methods that use object masks, such as NeuS [68] and Voxurf, as well as Neural-PBIR [57], that jointly reconstructs surfaces, materials, and environment maps. + +# 5.2. Results + +Emissive Source Recosnstruction. Tab. 2 shows that our approach excels in accurately identifying emissive source regions and restoring their intensity, regardless of the source color. While TensoIR and NeILF++ can restore emissions by modifying their physical rendering equations, they suffer from emissive source ambiguity, leading to near-zero IoU performance (see Appendix). For a comprehensive comparison, we report the best performance of the baseline methods using thresholding on the reconstructed emission strength at 0.01 intervals. ESR-NeRF consistently outper + +![](images/d36286269b953d85620873449b3bf2d8cdd8c1686bd39beb6787efc64e786250.jpg) +Figure 7. Comparison of scene editing. ESR-NeRF provides precise source control and faithfully represents reflection effects. For easy comparison in the Cube scene of low intensity, the bottom-right images are presented with a $40\%$ increased brightness. + +forms the baselines in identifying emissive source regions across all scenes. Our method also achieves significantly lower MSE values for restoring LDR to HDR images compared to the baselines, demonstrating its effectiveness of handling the ill-posed nature of the scenes with emissive sources. This is visually confirmed in Fig. 6, where ESR-NeRF surpasses the baselines in a complex scene with numerous small light bulbs. + +Scene Editing. Tab. 3 and Fig. 7 showcase the scene editing results under novel lighting conditions. Baseline methods struggle to adapt to lighting changes due to their inability to reconstruct emissive sources accurately. For example, in the Lego scene, TensoIR fails to adjust the illumination in surrounding regions when the color of emissive sources is changed, and in the Cube scene, both the hidden iPad screen and the cube surface covered by the user input mask change together. Twins introduces blue light onto yellow and red surfaces, leading to unintended white and purple appearances, even though there should be no reflection. PaletteNeRF, which manipulates scenes through re-colorization, lacks precise control over illumination, as seen in the synchronous color changes in the yellow ribbon and lighting. In contrast, ESR-NeRF demonstrates superior performance in + +![](images/6dcd2109976ab572d9a3bbbb7d499046b79d2ce40f45aafb878806bc0c0c0012.jpg) +Figure 8. Reconstructed emitter and re-lighting at novel view. + +scene editing outshining all baselines thanks to the accurate identification of emissive sources, as detailed in Table 3. ESR-NeRF effectively balances source reconstruction and novel view synthesis, ensuring high performance in both tasks. NeRF-W is excluded from color adjustments since it doesn't support direct color change through interpolating latent variables learned with light on and off conditions. + +Fig. 8 to 9 present additional examples of emissive source reconstruction and scene editing results. Fig. 10 shows results on real scenes, for which due to the impracticality of precise control over smart bulb colors, we offer emission reconstruction results with pseudo ground truth data. Our method effectively identifies emissive sources in real scenes, while it faces challenges in capturing complex + +![](images/cacfadfd51a4853b5aea10c887268a2ed8cd3fc64bf02d6e4591d743f3084d9f.jpg) +Figure 9. Results of source reconstruction and scene editing. + +![](images/b4b10d7590b1586aa937b771c07a96207d6996a5bc485653d2fe26cb0c572bc2.jpg) +Figure 10. Source reconstruction and scene editing on real scenes. + +reflections within light bulbs, as evident in the bright spot at the center of the bulbs in the ground truth edit results. + +Ablation Analysis. Progressive refinement with the stop-gradient operation in Eq. 15 improves the identification of emissive sources and reduces MSE values. Without $m_{\theta}$ , surface reconstructions become unreliable, complicating the accurate reconstruction of emissive sources. This issue is evident from the CD metrics and illustrated in Fig. 3. Further analyses are provided in Appendix. + +Illumination Decomposition. Fig. 11 demonstrates ESR-NeRF's decomposition of scene illumination into direct and indirect lighting from an environment map, as well as emissions and their reflections. The shadow behind the yellow ribbon in the direct figure and the illumination in the indirect figure showcase ESR-NeRF's ability to model both direct and indirect illumination. The reflection figure shows that our method accurately captures how emissive sources contribute to reflections on nearby regions. + +Surface Reconstruction. Interestingly, our approach can be applied to the scenes without emissive sources to enhance surface reconstruction, as evidenced by the lower CD values in Tab. 4 on the DTU dataset. For this experiment, we use Eq. 7 to 10 without our progressive refinement technique. ESR-NeRF's ability to adjust interrelated outgoing radiances helps prevent surface formations where radiances cannot be produced, considering the predicted scene context. Additional visualizations of the normals, BRDF, and environment maps are provided in Appendix. + +![](images/6a30198e34f9cbaf226af7585a64625f99845035691aac99a7632a201aee4000.jpg) +Figure 11. An example of illumination decomposition. + +
ScanNeuSVoxurfNeural-PBIRESR-NeRF
240.830.650.570.58
370.980.740.750.71
400.560.390.380.38
550.370.350.360.33
631.130.961.040.93
650.590.640.730.57
690.600.850.650.78
831.451.581.281.18
970.951.010.970.95
1050.780.680.760.58
1060.520.600.530.54
1101.431.110.841.08
1140.360.370.380.33
1180.450.450.460.40
1220.450.470.490.44
mean0.770.720.680.65
+ +Table 4. Results of surface reconstruction via the Chamfer distance on the DTU dataset. A lower value is better. + +
WhiteVivid
IoU ↑MSE ↓IoU ↑MSE ↓
w/o progressive0.409.920.413.93
w/o sg0.716.450.603.47
ESR-NeRF0.865.240.812.79
+ +
DTU +CD ↓
w/o mθ0.93
w/o LTS0.71
ESR-NeRF0.65
+ +Table 5. Ablation studies on the surface reconstruction (left) and the emissive source reconstruction (right). + +# 6. Conclusion + +We present ESR-NeRF as the first NeRF-based inverse rendering method for the scenes with emissive sources. Our approach uses LDR images, eliminating the need of HDR images to reconstruct emissive sources. Furthermore, we demonstrate the application of reconstructed sources in scene editing, enabling color and intensity modifications. + +Limitations. Future work could explore using a single lighting condition to disentangle emissive sources, environmental lighting, and object texture. It is also promising to address the challenge of volume ray tracing in unbounded scenes to extend to indoor scenes. Additionally, LTS based re-lighting may be weak in representing new colors that traverse unobserved light paths during training. An alternative approach could be extracting emission texture maps and modifying it using the engines such as Blender [13] or Mitsuba [22]. More details on alternative re-lighting methods and radiance fine-tuning are provided in Appendix. + +# 7. Acknowledgements + +This work was supported by Samsung Electronics MX, Basic Science Research Program through the National Research Foundation of Korea(NRF) funded by the Ministry of Education(RS-2023-00274280), and Institute of Information & Communications Technology Planning & Evaluation (IITP) grant funded by the Korea government (MSIT) (No. 2019-0-01082, SW StarLab; No. 2022-0-00156, Fundamental research on continual meta-learning for quality enhancement of casual videos and their 3D metaverse transformation). Gunhee Kim is the corresponding author. + +# References + +[1] Benjamin Attal, Jia-Bin Huang, Michael Zollhöfer, Johannes Kopf, and Changil Kim. Learning neural light fields with ray-space embedding. In CVPR, 2022. 2 +[2] Dejan Azinovic, Tzu-Mao Li, Anton Kaplanyan, and Matthias Nießner. Inverse path tracing for joint material and lighting estimation. In CVPR, 2019. 2 +[3] Jonathan T Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P Srinivasan. Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In ICCV, 2021. 1 +[4] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Mip-nerf 360: Unbounded anti-aliased neural radiance fields. In CVPR, 2022. 1 +[5] Jonathan T. Barron, Ben Mildenhall, Dor Verbin, Pratul P. Srinivasan, and Peter Hedman. Zip-nerf: Anti-aliased grid-based neural radiance fields. In ICCV, 2023. 2 +[6] Sai Bi, Zexiang Xu, Pratul Srinivasan, Ben Mildenhall, Kalyan Sunkavalli, Miloš Hasan, Yannick Hold-Geoffroy, David Kriegman, and Ravi Ramamoorthi. Neural reflectance fields for appearance acquisition. In arXiv, 2020. 1 +[7] Mark Boss, Varun Jampani, Kihwan Kim, Hendrik Lensch, and Jan Kautz. Two-shot spatially-varying brdf and shape estimation. In CVPR, 2020. +[8] Mark Boss, Raphael Braun, Varun Jampani, Jonathan T. Barron, Ce Liu, and Hendrik P.A. Lensch. Nerd: Neural reflectance decomposition from image collections. In ICCV, 2021. 1 +[9] Mark Boss, Varun Jampani, Raphael Braun, Ce Liu, Jonathan T. Barron, and Hendrik P.A. Lensch. Neural-pil: Neural pre-integrated lighting for reflectance decomposition. In NeurIPS, 2021. 2 +[10] Brent Burley and Walt Disney Animation Studios. Physically-based shading at disney. In SIGGRAPH, 2012. 5 +[11] Bowen Cai, Jinchi Huang, Rongfei Jia, Chengfei Lv, and Huan Fu. Neuda: Neural deformable anchor for high-fidelity implicit surface reconstruction. In CVPR, 2023. 2 +[12] Anpei Chen, Zexiang Xu, Andreas Geiger, Jingyi Yu, and Hao Su. Tensorf: Tensorial radiance fields. In ECCV, 2022. 2 +[13] Blender Online Community. Blender - a 3D modelling and rendering package. Blender Foundation, Stichting Blender Foundation, Amsterdam, 2018. 8 +[14] Sara Fridovich-Keil, Alex Yu, Matthew Tancik, Qinhong Chen, Benjamin Recht, and Angjoo Kanazawa. Plenoxels: Radiance fields without neural networks. In CVPR, 2022. 2 +[15] Mathieu Garon, Kalyan Sunkavalli, Sunil Hadap, Nathan Carr, and Jean-Francois Lalonde. Fast spatially-varying indoor lighting estimation. In CVPR, 2019. 1 +[16] Wenhang Ge, Tao Hu, Haoyu Zhao, Shu Liu, and Ying-Cong Chen. Ref-neus: Ambiguity-reduced neural implicit surface learning for multi-view reconstruction with reflection. In ICCV, 2023. 1 +[17] Saeed Hadadan, Shuhong Chen, and Matthias Zwicker. Neural radiosivity. In ACM TOG, 2021. 2 + +[18] Saeed Hadadan, Geng Lin, Jan Novák, Fabrice Rousselle, and Matthias Zwicker. Inverse global illumination using a neural radiometric prior. In SIGGRAPH Conference Proceedings, 2023. 1 +[19] Bjoern Haefner, Simon Green, Alan Oursland, Daniel Andersen, Michael Goesele, Daniel Cremers, Richard Newcombe, and Thomas Whelan. Recovering real-world reflectance properties and shading from hdr imagery. In 3DV, 2021. 2 +[20] Jon Hasselgren, Nikolai Hofmann, and Jacob Munkberg. Shape, light, and material decomposition from images using monte carlo rendering and denoising. In NeurIPS, 2022. 2 +[21] IEC. IEC 61966-2-1:1999. Technical report, International Electrotechnical Commission, 1999. 3 +[22] Wenzel Jakob, Sébastien Speierer, Nicolas Roussel, Merlin Nimier-David, Delio Vicini, Tizian Zeltner, Baptiste Nicolet, Miguel Crespo, Vincent Leroy, and Ziyi Zhang. Mitsuba 3 renderer, 2022. https://mitsuba-renderer.org. 8 +[23] Rasmus Jensen, Anders Dahl, George Vogiatzis, Engil Tola, and Henrik Aanæs. Large scale multi-view stereopsis evaluation. In CVPR, 2014. 2, 6 +[24] Chaonan Ji, Tao Yu, Kaiwen Guo, Jingxin Liu, and Yebin Liu. Geometry-aware single-image full-body human relighting. In ECCV, 2022. 1 +[25] Haian Jin, Isabella Liu, Peijia Xu, Xiaoshuai Zhang, Songfang Han, Sai Bi, Xiaowei Zhou, Zexiang Xu, and Hao Su. Tensoir: Tensorial inverse rendering. In CVPR, 2023. 2, 3, 6 +[26] James T. Kajiya. The rendering equation. In SIGGRAPH, 1986. 3 +[27] Petr Kellnhofer, Lars C Jebe, Andrew Jones, Ryan Spicer, Kari Pulli, and Gordon Wetzstein. Neural lumigraph rendering. In CVPR, 2021. 1 +[28] Zhengfei Kuang, Fujun Luan, Sai Bi, Zhixin Shu, Gordon Wetzstein, and Kalyan Sunkavalli. Palettenerf: Palette-based appearance editing of neural radiance fields. In CVPR, 2023. 6 +[29] Zhengqin Li, Mohammad Shafiei, Ravi Ramamoorthi, Kalyan Sunkavalli, and Manmohan Chandraker. Inverse rendering for complex indoor scenes: Shape, spatially-varying lighting and svbrdf from a single image. In CVPR, 2020. 1 +[30] Zhengqin Li, Jia Shi, Sai Bi, Rui Zhu, Kalyan Sunkavalli, Miloš Hašan, Zexiang Xu, Ravi Ramamoorthi, and Manmohan Chandraker. Physically-based editing of indoor scene lighting from a single image. In ECCV, 2022. 1 +[31] Zhaoshuo Li, Thomas Müller, Alex Evans, Russell H Taylor, Mathias Unberath, Ming-Yu Liu, and Chen-Hsuan Lin. Neuralangelo: High-fidelity neural surface reconstruction. In CVPR, 2023. 2 +[32] Zhen Li, Lingli Wang, Mofang Cheng, Cihui Pan, and Jiaqi. Yang. Multi-view inverse rendering for large-scale real-world indoor scenes. In CVPR, 2023. 2 +[33] Ruofan Liang, Huiting Chen, Chunlin Li, Fan Chen, Selvakumar Panneer, and Nandita Vijaykumar. Envidr: Implicit differentiable renderer with neural environment lighting. In ICCV, 2023. 1 + +[34] Fujun Luan, Shuang Zhao, Kavita Bala, and Zhao Dong. Unified shape and svbrdf recovery using differentiable monte carlo rendering. In EGSR, 2021. 1 +[35] Linjie Lyu, Ayush Tewari, Thomas Leimkuehler, Marc Habermann, and Christian Theobalt. Neural radiance transfer fields for relightable novel-view synthesis with global illumination. In ECCV, 2022. 2 +[36] Linjie Lyu, Ayush Tewari, Marc Habermann, Shunsuke Saito, Michael Zollhöfer, Thomas Leimkuehler, and Christian Theobalt. Diffusion posterior illumination for ambiguity-aware inverse rendering. In ACM TOG, 2023. 1 +[37] Alexander Mai, Dor Verbin, Falko Kuester, and Sara Fridovich-Keil. Neural microfacet fields for inverse rendering. In ICCV, 2023. 1, 3 +[38] Ricardo Martin-Brualla, Noha Radwan, Mehdi S. M. Sajjadi, Jonathan T. Barron, Alexey Dosovitskiy, and Daniel Duckworth. Nerf in the wild: Neural radiance fields for unconstrained photo collections. In CVPR, 2021. 1, 6 +[39] Nelson Max. Optical models for direct volume rendering. IEEE Transactions on Visualization and Computer Graphics, 1(2):99-108, 1995. 2 +[40] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In ECCV, 2020. 1, 2 +[41] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. In ACM TOG, 2022. 2 +[42] Jacob Munkberg, Jon Hasselgren, Tianchang Shen, Jun Gao, Wenzheng Chen, Alex Evans, Thomas Müller, and Sanja Fidler. Extracting triangular 3d models, materials, and lighting from images. In CVPR, 2022. 2 +[43] Michael Niemeyer, Lars Mescheder, Michael Oechsle, and Andreas Geiger. Differentiable volumetric rendering: Learning implicit 3d representations without 3d supervision. In CVPR, 2020. 1 +[44] Michael Oechsle, Songyou Peng, and Andreas Geiger. Unisurf: Unifying neural implicit surfaces and radiance fields for multi-view reconstruction. In ICCV, 2021. 2 +[45] Julian Ost, Issam Laradji, Alejandro Newell, Yuval Bahat, and Felix Heide. Neural point light fields. In CVPR, 2022. 1 +[46] Rohit Pandey, Sergio Orts-Escalano, Chloe LeGendre, Christian Haene, Sofien Bouaziz, Christoph Rhemann, Paul Debevec, and Seann Fanello. Total relighting: Learning to relight portraits for background replacement. In ACM TOG, 2021. 2 +[47] Jeong Joon Park, Peter Florence, Julian Straub, Richard Newcombe, and Steven Lovegrove. Deepsdf: Learning continuous signed distance functions for shape representation. In CVPR, 2019. 1 +[48] Julien Philip, Sébastien Morgenthaler, Michael Gharbi, and George Drettakis. Free-viewpoint indoor neural relighting from multi-view stereo. In ACM TOG, 2021. 2 +[49] Martin Piala and Ronald Clark. Terminerf: Ray termination prediction for efficient neural rendering. In 3DV, 2021. 2 +[50] Viktor Rudnev, Mohamed Elgharib, William Smith, Lingjie Liu, Vladislav Golyanik, and Christian Theobalt. Nerf for outdoor scene relighting. In ECCV, 2022. 1 + +[51] Vincent Sitzmann, Julien N.P. Martel, Alexander W. Bergman, David B. Lindell, and Gordon Wetzstein. Implicit neural representations with periodic activation functions. In NeurIPS, 2020. 2 +[52] Vincent Sitzmann, Semon Rezchikov, Bill Freeman, Josh Tenenbaum, and Fredo Durand. Light field networks: Neural scene representations with single-evaluation rendering. In NeurIPS, 2021. 2 +[53] Alvy Ray Smith. Color gamut transform pairs. In ACM TOG, 1978. 5 +[54] Pratul P. Srinivasan, Boyang Deng, Xiuming Zhang, Matthew Tancik, Ben Mildenhall, and Jonathan T. Barron. Nerv: Neural reflectance and visibility fields for relighting and view synthesis. In CVPR, 2021. 2, 3 +[55] Mohammed Suhail, Carlos Esteves, Leonid Sigal, and Ameesh Makadia. Light field neural rendering. In CVPR, 2022. 2 +[56] Cheng Sun, Min Sun, and Hwann-Tzong Chen. Direct voxel grid optimization: Super-fast convergence for radiance fields reconstruction. In CVPR, 2022. 2 +[57] Cheng Sun, Guangyan Cai, Zhengqin Li, Kai Yan, Cheng Zhang, Carl Marshall, Jia-Bin Huang, Shuang Zhao, and Zhao Dong. Neural-pbir reconstruction of shape, material, and illumination. In ICCV, 2023. 2, 3, 6 +[58] Jiaming Sun, Xi Chen, Qianqian Wang, Zhengqi Li, Hadar Averbuch-Elor, Xiaowei Zhou, and Noah Snively. Neural 3D reconstruction in the wild. In SIGGRAPH Conference Proceedings, 2022. 1 +[59] Tiancheng Sun, Kai-En Lin, Sai Bi, Zexiang Xu, and Ravi Ramamoorthi. Nelf: Neural light-transport field for portrait view synthesis and relighting. In EGSR, 2021. 2 +[60] Matthew Tancik, Pratul P. Srinivasan, Ben Mildenhall, Sara Fridovich-Keil, Nithin Raghavan, Utkarsh Singhal, Ravi Ramamoorthi, Jonathan T. Barron, and Ren Ng. Fourier features let networks learn high frequency functions in low dimensional domains. In NeurIPS, 2020. 2 +[61] Jiaxiang Tang, Hang Zhou, Xiaokang Chen, Tianshu Hu, Errui Ding, Jingdong Wang, and Gang Zeng. Delicate textured mesh recovery from nef via adaptive surface refinement. In ICCV, 2023. 1 +[62] Ayush Tewari, Ohad Fried, Justus Thies, Vincent Sitzmann, Stephen Lombardi, Kalyan Sunkavalli, Ricardo Martin-Brualla, Tomas Simon, Jason Saragih, Matthias Nießner, et al. State of the art on neural rendering. In CGF, 2020. 1 +[63] Ayush Tewari, Justus Thies, Ben Mildenhall, Pratul Srinivasan, Edgar Tretschk, Wang Yifan, Christoph Lassner, Vincent Sitzmann, Ricardo Martin-Brualla, Stephen Lombardi, et al. Advances in neural rendering. In CGF, 2022. 1 +[64] Marco Toschi, Riccardo De Matteo, Riccardo Spezialetti, Daniele De Gregorio, Luigi Di Stefano, and Samuele Salti. Relight my nerf: A dataset for novel view synthesis and relighting of real world objects. In CVPR, 2023. 2 +[65] Dor Verbin, Peter Hedman, Ben Mildenhall, Todd Zickler, Jonathan T Barron, and Pratul P Srinivasan. Ref-nerf: Structured view-dependent appearance for neural radiance fields. In CVPR, 2022. 1 + +[66] Dongqing Wang, Tong Zhang, and Sabine Susstrunk. Nemto: Neural environment matting for novel view and relighting synthesis of transparent objects. In ICCV, 2023. 1 +[67] Jiaping Wang, Peiran Ren, Minmin Gong, John Snyder, and Baining Guo. All-frequency rendering of dynamic, spatially-varying reflectance. ACM TOG, 2009. 4 +[68] Peng Wang, Lingjie Liu, Yuan Liu, Christian Theobalt, Taku Komura, and Wenping Wang. Neus: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. In NeurIPS, 2021. 2, 6 +[69] Shaofei Wang, Katja Schwarz, Andreas Geiger, and Siyu Tang. Arah: Animatable volume rendering of articulated human sdfs. In ECCV, 2022. 1 +[70] Yiming Wang, Qin Han, Marc Habermann, Kostas Dani-ilidis, Christian Theobalt, and Lingjie Liu. Neus2: Fast learning of neural implicit surfaces for multi-view reconstruction. In ICCV, 2023. 2 +[71] Zian Wang, Jonah Philion, Sanja Fidler, and Jan Kautz. Learning indoor inverse rendering with 3d spatially-varying lighting. In ICCV, 2021. 1 +[72] Zian Wang, Wenzheng Chen, David Acuna, Jan Kautz, and Sanja Fidler. Neural light field estimation for street scenes with differentiable virtual object insertion. In ECCV, 2022. 1 +[73] Zian Wang, Tianchang Shen, Jun Gao, Shengyu Huang, Jacob Munkberg, Jon Hasselgren, Zan Gojcic, Wenzheng Chen, and Sanja Fidler. Neural fields meet explicit geometric representations for inverse rendering of urban scenes. In CVPR, 2023. 1 +[74] Henrique Weber, Mathieu Garon, and Jean-François Lalonde. Editable indoor lighting estimation. In ECCV, 2022. 1 +[75] Haoqian Wu, Zhipeng Hu, Lincheng Li, Yongqiang Zhang, Changjie Fan, and Xin Yu. Nefii: Inverse rendering for reflectance decomposition with near-field indirect illumination. In CVPR, 2023. 2 +[76] Liwen Wu, Rui Zhu, Mustafa B. Yaldiz, Yinhao Zhu, Hong Cai, Janarbek Matai, Fatih Porikli, Tzu-Mao Li, Manmohan Chandraker, and Ravi Ramamoorthi. Factorized inverse path tracing for efficient and accurate material-lighting estimation. In ICCV, 2023. 2 +[77] Qianyi Wu, Xian Liu, Yuedong Chen, Kejie Li, Chuanxia Zheng, Jianfei Cai, and Jianmin Zheng. Object-compositional neural implicit surfaces. In ECCV, 2022. 2 +[78] Qianyi Wu, Kaisiyuan Wang, Kejie Li, Jianmin Zheng, and Jianfei Cai. Objectsdf++: Improved object-compositional neural implicit surfaces. In ICCV, 2023. 2 +[79] Tong Wu, Jia-Mu Sun, Yu-Kun Lai, and Lin Gao. De-nerf: Decoupled neural radiance fields for view-consistent appearance editing and high-frequency environmental relighting. In SIGGRAPH ASIA Conference Proceedings, 2023. 1 +[80] Tong Wu, Jiaqi Wang, Xingang Pan, Xudong Xu, Christian Theobalt, Ziwei Liu, and Dahua Lin. Voxurf: Voxel-based efficient and accurate neural surface reconstruction. In ICLR, 2023. 2, 5, 6 +[81] Yuanbo Xiangli, Linning Xu, Xingang Pan, Nanxuan Zhao, Anyi Rao, Christian Theobalt, Bo Dai, and Dahua Lin. + +Bungeenerf: Progressive neural radiance field for extreme multi-scale scene rendering. In ECCV, 2022. 1 +[82] Yingyan Xu, Gaspard Zoss, Prashanth Chandran, Markus Gross, Derek Bradley, and Paulo Gotardo. Renerf: Relightable neural radiance fields with nearfield lighting. In ICCV, 2023. 2 +[83] Wenqi Yang, Guanying Chen, Chaofeng Chen, Zhenfang Chen, and Kwan-Yee K. Wong. Ps-nerf: Neural inverse rendering for multi-view photometric stereo. In ECCV, 2022. 2 +[84] Wenqi Yang, Guanying Chen, Chaofeng Chen, Zhenfang Chen, and Kwan-Yee K. Wong. S $^3$ -nerf: Neural reflectance field from shading and shadow under a single viewpoint. In NeurIPS, 2022. 2 +[85] Yao Yao, Jingyang Zhang, Jingbo Liu, Yihang Qu, Tian Fang, David McKinnon, Yanghai Tsin, and Long Quan. Neilf: Neural incident light field for material and lighting estimation. In ECCV, 2022. 1 +[86] Lior Yariv, Yoni Kasten, Dror Moran, Meirav Galun, Matan Atzmon, Basri Ronen, and Yaron Lipman. Multiview neural surface reconstruction by disentangling geometry and appearance. In NeurIPS, 2020. 1 +[87] Lior Yariv, Jiatao Gu, Yoni Kasten, and Yaron Lipman. Volume rendering of neural implicit surfaces. In NeurIPS, 2021. 2 +[88] Alex Yu, Ruilong Li, Matthew Tancik, Hao Li, Ren Ng, and Angjoo Kanazawa. PlenOctrees for real-time rendering of neural radiance fields. In ICCV, 2021. 2 +[89] Chong Zeng, Guojun Chen, Yue Dong, Pieter Peers, Hongzhi Wu, and Xin Tong. Relighting neural radiance fields with shadow and highlight hints. In SIGGRAPH Conference Proceedings, 2023. 2 +[90] Jingyang Zhang, Yao Yao, Shiwei Li, Jingbo Liu, Tian Fang, David McKinnon, Yanghai Tsin, and Long Quan. Neilf++: Inter-reflectable light fields for geometry and material estimation. In ICCV, 2023. 2, 6 +[91] Kai Zhang, Gernot Riegler, Noah Snavely, and Vladlen Koltun. Nerf++: Analyzing and improving neural radiance fields. arXiv preprint arXiv:2010.07492, 2020. 1 +[92] Kai Zhang, Fujun Luan, Qianqian Wang, Kavita Bala, and Noah Snavely. Physg: Inverse rendering with spherical gaussians for physics-based material editing and relighting. In CVPR, 2021. 2 +[93] Kai Zhang, Fujun Luan, Zhengqi Li, and Noah Snavely. Iron: Inverse rendering by optimizing neural sdfs and materials from photometric images. In CVPR, 2022. 1 +[94] Xiuming Zhang, Sean Fanello, Yun-Ta Tsai, Tiancheng Sun, Tianfan Xue, Rohit Pandey, Sergio Orts-Escalano, Philip Davidson, Christoph Rhemann, Paul Debevec, et al. Neural light transport for relighting and view synthesis. In ACM TOG, 2021. 2 +[95] Xiuming Zhang, Pratul P. Srinivasan, Boyang Deng, Paul Debevec, William T. Freeman, and Jonathan T. Barron. Ner-factor: Neural factorization of shape and reflectance under an unknown illumination. In ACM TOG, 2021. 2 +[96] Yuanqing Zhang, Jiaming Sun, Xingyi He, Huan Fu, Rongfei Jia, and Xiaowei Zhou. Modeling indirect illumination for inverse rendering. In CVPR, 2022. 2 + +[97] Y. Zhang, Z. Hu, H. Wu, M. Zhao, L. Li, Z. Zou, and C. Fan. Towards unbiased volume rendering of neural implicit surfaces with geometry priors. In CVPR, 2023. 2 +[98] Quan Zheng, Gurprit Singh, and Hans-Peter Seidel. Neural relightable participating media rendering. In NeurIPS, 2021. 2 +[99] Jingsen Zhu, Fujun Luan, Yuchi Huo, Zihao Lin, Zhihua Zhong, Dianbing Xi, Rui Wang, Hujun Bao, Jiaxiang Zheng, and Rui Tang. Learning-based inverse rendering of complex indoor scenes with differentiable monte carlo raytracing. In SIGGRAPH ASIA Conference Proceedings, 2022. 2 \ No newline at end of file diff --git a/2024/ESR-NeRF_ Emissive Source Reconstruction Using LDR Multi-view Images/images.zip b/2024/ESR-NeRF_ Emissive Source Reconstruction Using LDR Multi-view Images/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..712364c54fe61443cfebe37102433b820dc92e3e --- /dev/null +++ b/2024/ESR-NeRF_ Emissive Source Reconstruction Using LDR Multi-view Images/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9319cf0e62c063db90d3861cb5cfd9262c164a66d6f3f560f265d427182c6edb +size 846598 diff --git a/2024/ESR-NeRF_ Emissive Source Reconstruction Using LDR Multi-view Images/layout.json b/2024/ESR-NeRF_ Emissive Source Reconstruction Using LDR Multi-view Images/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..17f9ebb0289c9d7309706383b5508c5e26642d45 --- /dev/null +++ b/2024/ESR-NeRF_ Emissive Source Reconstruction Using LDR Multi-view Images/layout.json @@ -0,0 +1,12000 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 63, + 103, + 530, + 120 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 103, + 530, + 120 + ], + "spans": [ + { + "bbox": [ + 63, + 103, + 530, + 120 + ], + "type": "text", + "content": "ESR-NeRF: Emissive Source Reconstruction Using LDR Multi-view Images" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 140, + 142, + 457, + 172 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 140, + 142, + 457, + 172 + ], + "spans": [ + { + "bbox": [ + 140, + 142, + 457, + 172 + ], + "type": "text", + "content": "Jinseo Jeong" + }, + { + "bbox": [ + 140, + 142, + 457, + 172 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 140, + 142, + 457, + 172 + ], + "type": "text", + "content": " Junseo Koo" + }, + { + "bbox": [ + 140, + 142, + 457, + 172 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 140, + 142, + 457, + 172 + ], + "type": "text", + "content": " Qimeng Zhang" + }, + { + "bbox": [ + 140, + 142, + 457, + 172 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 140, + 142, + 457, + 172 + ], + "type": "text", + "content": " Gunhee Kim" + }, + { + "bbox": [ + 140, + 142, + 457, + 172 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 140, + 142, + 457, + 172 + ], + "type": "text", + "content": " Seoul National University Korea University" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 174, + 546, + 199 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 174, + 546, + 199 + ], + "spans": [ + { + "bbox": [ + 53, + 174, + 546, + 199 + ], + "type": "text", + "content": "jinseo@vision.snu.ac.kr, junseo.koo@vision.snu.ac.kr, zoe1024@korea.ac.kr, gunhee@snu.ac.kr \nhttps://jinseo.kr/ESR-NeRF" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 143, + 227, + 192, + 239 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 227, + 192, + 239 + ], + "spans": [ + { + "bbox": [ + 143, + 227, + 192, + 239 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 252, + 290, + 492 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 252, + 290, + 492 + ], + "spans": [ + { + "bbox": [ + 46, + 252, + 290, + 492 + ], + "type": "text", + "content": "Existing NeRF-based inverse rendering methods suppose that scenes are exclusively illuminated by distant light sources, neglecting the potential influence of emissive sources within a scene. In this work, we confront this limitation using LDR multi-view images captured with emissive sources turned on and off. Two key issues must be addressed: 1) ambiguity arising from the limited dynamic range along with unknown lighting details, and 2) the expensive computational cost in volume rendering to backtrace the paths leading to final object colors. We present a novel approach, ESR-NeRF, leveraging neural networks as learnable functions to represent ray-traced fields. By training networks to satisfy light transport segments, we regulate outgoing radiances, progressively identifying emissive sources while being aware of reflection areas. The results on scenes encompassing emissive sources with various properties demonstrate the superiority of ESR-NeRF in qualitative and quantitative ways. Our approach also extends its applicability to the scenes devoid of emissive sources, achieving lower CD metrics on the DTU dataset." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 513, + 128, + 525 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 513, + 128, + 525 + ], + "spans": [ + { + "bbox": [ + 47, + 513, + 128, + 525 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 534, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 534, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 534, + 287, + 714 + ], + "type": "text", + "content": "Extensive research has focused on reconstructing 3D object structures [16, 43, 47, 86], material properties [18, 29, 65], and lighting [15, 33, 34, 74, 79] from 2D images, applicable across domains including 3D graphics and augmented reality [62, 63, 69, 72]. This endeavor not only facilitates the creation of life-like virtual objects but also streamlines the process of scene manipulation [27, 58, 61, 73]. Recent advancements [24, 30, 36, 71] have built on Neural Radiance Fields (NeRF) [40] successes in novel view synthesis [3, 4, 45, 81, 91]. Significant progress in relighting [37, 38, 50] has facilitated scene editing via manipulating the reconstructed light sources. However, existing methods predominantly deal with the scenes lit by distant sources, like environment maps or collocated flashlights. Notably, NeRF-based inverse rendering has yet to" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 307, + 225, + 351, + 283 + ], + "blocks": [ + { + "bbox": [ + 307, + 225, + 351, + 283 + ], + "lines": [ + { + "bbox": [ + 307, + 225, + 351, + 283 + ], + "spans": [ + { + "bbox": [ + 307, + 225, + 351, + 283 + ], + "type": "image", + "image_path": "be789cdd189d4bd05cb65e9e4d54a65edb1901096f20ac60382ca27daa97eb57.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 308, + 285, + 350, + 292 + ], + "lines": [ + { + "bbox": [ + 308, + 285, + 350, + 292 + ], + "spans": [ + { + "bbox": [ + 308, + 285, + 350, + 292 + ], + "type": "text", + "content": "255 pixel values" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 352, + 226, + 395, + 283 + ], + "blocks": [ + { + "bbox": [ + 352, + 226, + 395, + 283 + ], + "lines": [ + { + "bbox": [ + 352, + 226, + 395, + 283 + ], + "spans": [ + { + "bbox": [ + 352, + 226, + 395, + 283 + ], + "type": "image", + "image_path": "80db8dedaaea4ce3ac6829a80e9c359cab6dd83422264462eff09725a33280fb.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 352, + 285, + 394, + 292 + ], + "lines": [ + { + "bbox": [ + 352, + 285, + 394, + 292 + ], + "spans": [ + { + "bbox": [ + 352, + 285, + 394, + 292 + ], + "type": "text", + "content": "Information loss" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 307, + 297, + 350, + 351 + ], + "blocks": [ + { + "bbox": [ + 307, + 297, + 350, + 351 + ], + "lines": [ + { + "bbox": [ + 307, + 297, + 350, + 351 + ], + "spans": [ + { + "bbox": [ + 307, + 297, + 350, + 351 + ], + "type": "image", + "image_path": "7843b6cfe8d393de564ff362e62133cbda7707f238fa4ecf3cfad14bd5dc81cc.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 310, + 354, + 348, + 361 + ], + "lines": [ + { + "bbox": [ + 310, + 354, + 348, + 361 + ], + "spans": [ + { + "bbox": [ + 310, + 354, + 348, + 361 + ], + "type": "text", + "content": "Threshold 0.65" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 352, + 297, + 395, + 351 + ], + "blocks": [ + { + "bbox": [ + 352, + 297, + 395, + 351 + ], + "lines": [ + { + "bbox": [ + 352, + 297, + 395, + 351 + ], + "spans": [ + { + "bbox": [ + 352, + 297, + 395, + 351 + ], + "type": "image", + "image_path": "91a2824c953bde7a9f164294590355335cd4738daa64b0b63b095bbd1f12496f.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 354, + 354, + 392, + 361 + ], + "lines": [ + { + "bbox": [ + 354, + 354, + 392, + 361 + ], + "spans": [ + { + "bbox": [ + 354, + 354, + 392, + 361 + ], + "type": "text", + "content": "Threshold 0.75" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 395, + 226, + 413, + 283 + ], + "blocks": [ + { + "bbox": [ + 395, + 226, + 413, + 283 + ], + "lines": [ + { + "bbox": [ + 395, + 226, + 413, + 283 + ], + "spans": [ + { + "bbox": [ + 395, + 226, + 413, + 283 + ], + "type": "image", + "image_path": "cb90e4c96d41e556988afaa3833d0fcca7b226e1c2055d7f3b263ab415c66fe8.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 395, + 285, + 397, + 292 + ], + "lines": [ + { + "bbox": [ + 395, + 285, + 397, + 292 + ], + "spans": [ + { + "bbox": [ + 395, + 285, + 397, + 292 + ], + "type": "text", + "content": "" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 410, + 226, + 496, + 361 + ], + "blocks": [ + { + "bbox": [ + 410, + 226, + 496, + 361 + ], + "lines": [ + { + "bbox": [ + 410, + 226, + 496, + 361 + ], + "spans": [ + { + "bbox": [ + 410, + 226, + 496, + 361 + ], + "type": "image", + "image_path": "738b4216d6118c90695254c7dfc9ec18e1c04b9e7fa8ab91a620794a68fb5fc8.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 503, + 227, + 539, + 234 + ], + "lines": [ + { + "bbox": [ + 503, + 227, + 539, + 234 + ], + "spans": [ + { + "bbox": [ + 503, + 227, + 539, + 234 + ], + "type": "text", + "content": "Edit emission" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 501, + 237, + 543, + 289 + ], + "blocks": [ + { + "bbox": [ + 501, + 237, + 543, + 289 + ], + "lines": [ + { + "bbox": [ + 501, + 237, + 543, + 289 + ], + "spans": [ + { + "bbox": [ + 501, + 237, + 543, + 289 + ], + "type": "image", + "image_path": "aa37142ac7b770d8a86342081c99d6da7914dbe9cc2bbeeed215360a535f0e20.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 511, + 289, + 532, + 296 + ], + "lines": [ + { + "bbox": [ + 511, + 289, + 532, + 296 + ], + "spans": [ + { + "bbox": [ + 511, + 289, + 532, + 296 + ], + "type": "text", + "content": "Desired" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 501, + 299, + 543, + 349 + ], + "blocks": [ + { + "bbox": [ + 501, + 299, + 543, + 349 + ], + "lines": [ + { + "bbox": [ + 501, + 299, + 543, + 349 + ], + "spans": [ + { + "bbox": [ + 501, + 299, + 543, + 349 + ], + "type": "image", + "image_path": "c1f1c8ea760cb6b452e4444e5270a3c25bd6d3ebc5a978c6ad15c00985bc2a91.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 507, + 350, + 537, + 357 + ], + "lines": [ + { + "bbox": [ + 507, + 350, + 537, + 357 + ], + "spans": [ + { + "bbox": [ + 507, + 350, + 537, + 357 + ], + "type": "text", + "content": "Unrealistic" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 304, + 371, + 546, + 459 + ], + "lines": [ + { + "bbox": [ + 304, + 371, + 546, + 459 + ], + "spans": [ + { + "bbox": [ + 304, + 371, + 546, + 459 + ], + "type": "text", + "content": "Figure 1. Challenges posed by emissive sources in LDR images. Green, red, and blue in thresholded images respectively show true positives, false negatives, and false positives of source identification. Thresholding values are scaled down divided by 255. The contrast between light on and off pixel values is more pronounced in surroundings than emissive sources. Inaccurate reconstruction of emissive sources disrupts scene editing, causing reflection areas to stay static while only the source colors change." + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + } + ], + "index": 23 + }, + { + "bbox": [ + 304, + 470, + 545, + 493 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 470, + 545, + 493 + ], + "spans": [ + { + "bbox": [ + 304, + 470, + 545, + 493 + ], + "type": "text", + "content": "consider scenes with multiple emissive sources, a common real-world illumination condition." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 304, + 496, + 546, + 698 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 496, + 546, + 698 + ], + "spans": [ + { + "bbox": [ + 304, + 496, + 546, + 698 + ], + "type": "text", + "content": "Emissive sources in a scene introduce critical challenges: (i) ambiguity in decomposing scene components and (ii) high computational costs for analyzing the causes of pixel colors. This ambiguity stems from difficulties in identifying emissive source regions, as illustrated in Fig. 1. Contrary to prior setups [6-8, 66, 85, 93], we allow the possibility of numerous emissive sources throughout the scene. In standard photographs with pixel values from 0 to 255, the distinction between emissive sources and nearby reflection areas is challenging. As shown in Fig. 1, relying solely on pixel value thresholding is insufficient for differentiating between emissive sources and their reflections. Naive inverse path tracing is impractical, due to the computational costs rising exponentially with the number of ray bounces in volume rendering. This can cause inaccuracy in emissive source reconstruction, yielding unrealistic illumination in reflective areas as users manipulate emissive sources." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 318, + 701, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 701, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 318, + 701, + 545, + 713 + ], + "type": "text", + "content": "To address these challenges, we introduce ESR-NeRF" + } + ] + } + ], + "index": 28 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 494, + 35 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 494, + 35 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 494, + 35 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "4598" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 228 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 228 + ], + "type": "text", + "content": "(Emissive Sources Reconstructing NeRF), a novel approach capable of reconstructing any number of emissive sources by progressively discovering reflection areas. We assume that the scenes are observed in two lighting conditions: one with all emissive sources active and the other with them inactive. Our approach utilizes neural networks as learnable functions for representing ray-traced fields. By training networks to satisfy each light transport segment, we sidestep the computational overhead of ray tracing associated with ray bounces. In this work, we exclusively use low dynamic range (LDR) images, setting us apart from prior mesh-based methods that rely on high dynamic range (HDR) images [2, 19, 48, 76]." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 228, + 289, + 334 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 228, + 289, + 334 + ], + "spans": [ + { + "bbox": [ + 46, + 228, + 289, + 334 + ], + "type": "text", + "content": "Our experiments encompass synthetic and real scenes, ranging from single to multiple lighting configurations with complex reflections. The scenes vary in light source counts, color, and intensity. Qualitative and quantitative evaluations show ESR-NeRF's superiority over state-of-the-art NeRF-based re-lighting methods. Furthermore, Chamfer Distance (CD) metrics on the DTU dataset [23] indicate ESR-NeRF's competitive performance in scene reconstruction, even without emissive sources." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 59, + 335, + 237, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 335, + 237, + 346 + ], + "spans": [ + { + "bbox": [ + 59, + 335, + 237, + 346 + ], + "type": "text", + "content": "We summarize our contributions as follows." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 348, + 287, + 490 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 48, + 348, + 287, + 394 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 348, + 287, + 394 + ], + "spans": [ + { + "bbox": [ + 48, + 348, + 287, + 394 + ], + "type": "text", + "content": "1. Our work presents the first NeRF-based inverse rendering that can deal with the scenes with any number of emissive sources, challenging the distant light assumption of previous research." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 396, + 287, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 396, + 287, + 430 + ], + "spans": [ + { + "bbox": [ + 47, + 396, + 287, + 430 + ], + "type": "text", + "content": "2. Unlike existing mesh-based methods relying on HDR images, we use LDR images for the first time, overcoming the poor representation of emissive sources." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 431, + 287, + 454 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 431, + 287, + 454 + ], + "spans": [ + { + "bbox": [ + 47, + 431, + 287, + 454 + ], + "type": "text", + "content": "3. We provide a benchmark dataset designed to evaluate the performance of emissive source reconstruction." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 456, + 287, + 490 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 456, + 287, + 490 + ], + "spans": [ + { + "bbox": [ + 47, + 456, + 287, + 490 + ], + "type": "text", + "content": "4. Our method is applicable to the scenes with or without emissive sources, achieving superior mesh reconstruction results on the DTU dataset." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 47, + 502, + 132, + 514 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 502, + 132, + 514 + ], + "spans": [ + { + "bbox": [ + 47, + 502, + 132, + 514 + ], + "type": "text", + "content": "2. Related work" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 521, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 521, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 521, + 289, + 715 + ], + "type": "text", + "content": "Neural Rendering. Advancements in implicit representations [51, 60] and volume rendering [39] have significantly enhanced neural rendering capabilities, enabling the reconstruction of scene components from 2D images. One of the key directions is mesh extraction [44, 70, 77, 78, 80, 97], with methods like NeuS [68] and VolSDF [87] utilizing signed distance function (SDF) values for volume rendering. Recently, the efficient computation of volume rendering has become a focal point due to the substantial computational cost associated with network inference for ray color calculation [41, 49, 88]. Several methods propose to directly predict ray color using the 4D light fields concept [1, 52, 55] or leveraging voxel grids for fast inference of spatial features [5, 11, 12, 14, 31, 56]. NeuralRadiosity [17] shares similarity with our method, as it predicts raytraced values instead of explicitly tracing individual rays." + } + ] + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 307, + 71, + 547, + 125 + ], + "blocks": [ + { + "bbox": [ + 307, + 71, + 547, + 125 + ], + "lines": [ + { + "bbox": [ + 307, + 71, + 547, + 125 + ], + "spans": [ + { + "bbox": [ + 307, + 71, + 547, + 125 + ], + "type": "table", + "html": "
VoxurfTensoIRPath TracingESR-NeRF
Big Onn · d(n · d)b+1n2 · d
Indirect illuminationx✓✓✓
BRDF decompositionx✓✓✓
Emissive source controlxx✓✓
", + "image_path": "f0fb17b5a2737860d7e30c85f898be2f8463abfbc7c165c6e07c05e8418eb9f8.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 133, + 545, + 166 + ], + "lines": [ + { + "bbox": [ + 305, + 133, + 545, + 166 + ], + "spans": [ + { + "bbox": [ + 305, + 133, + 545, + 166 + ], + "type": "text", + "content": "Table 1. Computational cost comparison for inverse rendering methods. " + }, + { + "bbox": [ + 305, + 133, + 545, + 166 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 305, + 133, + 545, + 166 + ], + "type": "text", + "content": " is the number of sampled points along a ray, " + }, + { + "bbox": [ + 305, + 133, + 545, + 166 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 305, + 133, + 545, + 166 + ], + "type": "text", + "content": " is the number of scattering rays, and " + }, + { + "bbox": [ + 305, + 133, + 545, + 166 + ], + "type": "inline_equation", + "content": "b" + }, + { + "bbox": [ + 305, + 133, + 545, + 166 + ], + "type": "text", + "content": " is the number of ray bounces." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 173, + 545, + 233 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 173, + 545, + 233 + ], + "spans": [ + { + "bbox": [ + 304, + 173, + 545, + 233 + ], + "type": "text", + "content": "However, they primarily focus on calculating the final object color when all scene information is available. In contrast, our inverse rendering approach aims to reconstruct emissive sources within a scene, addressing the ambiguities introduced by their presence in LDR images." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 238, + 545, + 646 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 238, + 545, + 646 + ], + "spans": [ + { + "bbox": [ + 304, + 238, + 545, + 646 + ], + "type": "text", + "content": "Inverse Rendering. A growing emphasis revolves around the decomposition of materials represented by spatially varying bidirectional reflectance distribution functions (SVBRDF) [46, 83, 99]. To lessen the computational burden in inverse rendering [25, 54, 96, 98], several methods have adopted neural networks as lookup tables [9] or computational caches [54, 90, 95]. While NeRV [54] utilizes caching visibility and " + }, + { + "bbox": [ + 304, + 238, + 545, + 646 + ], + "type": "inline_equation", + "content": "\\mathrm{NeILF++}" + }, + { + "bbox": [ + 304, + 238, + 545, + 646 + ], + "type": "text", + "content": " [90] adopts caching surface point radiance with the inter-reflection loss for incident radiance, our method diverges by focusing on tracing radiance origins. Specifically, we aim to identify emissive sources within a scene, moving beyond the simplification of incident radiance calculations. Several methods rely on diverse known lighting configurations to exploit variations in object appearances [59, 64, 84, 89]. Toggling emissive sources on and off resembles the common one-light-at-a-time (OLAT) technique, as seen in NLT [94] and ReN-eRF [82]. However, our setting does not need to know light source properties and to toggle lights individually. Instead, we allow for toggling all lights together. Recent works have also jointly reconstruct the mesh, materials, and lighting [20, 35, 42, 57]. They tackle with images captured under a single unknown lighting condition [92, 95], assuming that radiance already encodes global illumination [75, 96]. However, they confine to the scenes illuminated by far-distant lights, constrained to an 8-bit color spectrum. Our work considers the presence of multiple emissive sources within a scene captured in LDR images, questioning the prevailing notion that radiance fields trained with the image rendering loss faithfully represents global illumination. While some methods [2, 19, 32, 48, 76] deal with the scenes featuring emissive sources, they work outside the volume rendering framework and depend on HDR input images, assuming prior knowledge of scene geometry." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 657, + 391, + 670 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 657, + 391, + 670 + ], + "spans": [ + { + "bbox": [ + 306, + 657, + 391, + 670 + ], + "type": "text", + "content": "3. Preliminaries" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 677, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 677, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 677, + 545, + 713 + ], + "type": "text", + "content": "Surface Representation. Analogous to NeRF [40], neural network " + }, + { + "bbox": [ + 305, + 677, + 545, + 713 + ], + "type": "inline_equation", + "content": "f_{\\theta}" + }, + { + "bbox": [ + 305, + 677, + 545, + 713 + ], + "type": "text", + "content": " predicts SDF values at arbitrary 3D spatial locations. NeuS [68] integrates surface representa" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4599" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 87, + 106, + 143 + ], + "blocks": [ + { + "bbox": [ + 50, + 87, + 106, + 143 + ], + "lines": [ + { + "bbox": [ + 50, + 87, + 106, + 143 + ], + "spans": [ + { + "bbox": [ + 50, + 87, + 106, + 143 + ], + "type": "image", + "image_path": "114f18d634099b544fc56a064202769a983a7f934e4831d24cb881fa79374340.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 64, + 145, + 96, + 153 + ], + "lines": [ + { + "bbox": [ + 64, + 145, + 96, + 153 + ], + "spans": [ + { + "bbox": [ + 64, + 145, + 96, + 153 + ], + "type": "text", + "content": "On Images" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 50, + 163, + 106, + 218 + ], + "blocks": [ + { + "bbox": [ + 50, + 163, + 106, + 218 + ], + "lines": [ + { + "bbox": [ + 50, + 163, + 106, + 218 + ], + "spans": [ + { + "bbox": [ + 50, + 163, + 106, + 218 + ], + "type": "image", + "image_path": "9a472dcd746a19a3e978d986bc74651228309b044ee968c37da34649cf9f85fb.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 64, + 220, + 96, + 228 + ], + "lines": [ + { + "bbox": [ + 64, + 220, + 96, + 228 + ], + "spans": [ + { + "bbox": [ + 64, + 220, + 96, + 228 + ], + "type": "text", + "content": "Off Images" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 111, + 81, + 410, + 232 + ], + "blocks": [ + { + "bbox": [ + 49, + 70, + 107, + 80 + ], + "lines": [ + { + "bbox": [ + 49, + 70, + 107, + 80 + ], + "spans": [ + { + "bbox": [ + 49, + 70, + 107, + 80 + ], + "type": "text", + "content": "Multi-view Input" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 247, + 70, + 277, + 80 + ], + "lines": [ + { + "bbox": [ + 247, + 70, + 277, + 80 + ], + "spans": [ + { + "bbox": [ + 247, + 70, + 277, + 80 + ], + "type": "text", + "content": "Training" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 111, + 81, + 410, + 232 + ], + "lines": [ + { + "bbox": [ + 111, + 81, + 410, + 232 + ], + "spans": [ + { + "bbox": [ + 111, + 81, + 410, + 232 + ], + "type": "image", + "image_path": "7c30826d9f9f5e59b8b5f0d05742991f50e3d6adc4853cb0fb80803759bb1545.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 416, + 82, + 545, + 232 + ], + "blocks": [ + { + "bbox": [ + 460, + 70, + 500, + 80 + ], + "lines": [ + { + "bbox": [ + 460, + 70, + 500, + 80 + ], + "spans": [ + { + "bbox": [ + 460, + 70, + 500, + 80 + ], + "type": "text", + "content": "Application" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 416, + 82, + 545, + 232 + ], + "lines": [ + { + "bbox": [ + 416, + 82, + 545, + 232 + ], + "spans": [ + { + "bbox": [ + 416, + 82, + 545, + 232 + ], + "type": "image", + "image_path": "b6bae7cda4ddb5a80e5c74766f89f303c299dc9ed7c045fe30d0f9548596a16f.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 242, + 547, + 277 + ], + "lines": [ + { + "bbox": [ + 46, + 242, + 547, + 277 + ], + "spans": [ + { + "bbox": [ + 46, + 242, + 547, + 277 + ], + "type": "text", + "content": "Figure 2. The pipeline of emissive source reconstruction. Given LDR images with emissive sources on and off, scene components are reconstructed by synthesizing training images and enforcing LTS requirements. Emissive sources are progressively refined via categorizing training rays into uncertain and certain groups. The scenes can be edited with new lighting conditions using reconstructed emissive sources." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 286, + 287, + 339 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 286, + 287, + 339 + ], + "spans": [ + { + "bbox": [ + 46, + 286, + 287, + 339 + ], + "type": "text", + "content": "tion into volume rendering using the SDF-based opacity " + }, + { + "bbox": [ + 46, + 286, + 287, + 339 + ], + "type": "inline_equation", + "content": "\\rho (x) = \\max (\\frac{-\\frac{d\\Phi_s}{dx}(f(x))}{\\Phi_s(f(x))},0)" + }, + { + "bbox": [ + 46, + 286, + 287, + 339 + ], + "type": "text", + "content": ". Here " + }, + { + "bbox": [ + 46, + 286, + 287, + 339 + ], + "type": "inline_equation", + "content": "\\Phi_s(x) = (1 + e^{-sx})^{-1}" + }, + { + "bbox": [ + 46, + 286, + 287, + 339 + ], + "type": "text", + "content": " is the sigmoid function where " + }, + { + "bbox": [ + 46, + 286, + 287, + 339 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 46, + 286, + 287, + 339 + ], + "type": "text", + "content": " controls the sharpness of surfaces. The color of a ray can be calculated as" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 83, + 344, + 287, + 369 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 83, + 344, + 287, + 369 + ], + "spans": [ + { + "bbox": [ + 83, + 344, + 287, + 369 + ], + "type": "interline_equation", + "content": "\\hat {C} (r) = \\int_ {0} ^ {\\infty} T (r (t)) \\rho (r (t)) L _ {o} (r (t), \\omega_ {o}) d t, \\tag {1}", + "image_path": "50635ec966c99c626df07dedf486cedbdabed36c5ae0456a45f4be535e620d61.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 47, + 374, + 288, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 374, + 288, + 437 + ], + "spans": [ + { + "bbox": [ + 47, + 374, + 288, + 437 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 374, + 288, + 437 + ], + "type": "inline_equation", + "content": "\\hat{C}(r)" + }, + { + "bbox": [ + 47, + 374, + 288, + 437 + ], + "type": "text", + "content": " denotes the predicted ray color, " + }, + { + "bbox": [ + 47, + 374, + 288, + 437 + ], + "type": "inline_equation", + "content": "r(t; c, \\omega_o) = c - t \\cdot \\omega_o" + }, + { + "bbox": [ + 47, + 374, + 288, + 437 + ], + "type": "text", + "content": " is the ray with camera center " + }, + { + "bbox": [ + 47, + 374, + 288, + 437 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 47, + 374, + 288, + 437 + ], + "type": "text", + "content": " along direction " + }, + { + "bbox": [ + 47, + 374, + 288, + 437 + ], + "type": "inline_equation", + "content": "\\omega_o" + }, + { + "bbox": [ + 47, + 374, + 288, + 437 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 47, + 374, + 288, + 437 + ], + "type": "inline_equation", + "content": "T(r(t)) = \\exp \\left( \\int_0^t -\\rho (r(u))du \\right)" + }, + { + "bbox": [ + 47, + 374, + 288, + 437 + ], + "type": "text", + "content": " is the transmittance, and " + }, + { + "bbox": [ + 47, + 374, + 288, + 437 + ], + "type": "inline_equation", + "content": "L_{o}(r(t), \\omega_{o})" + }, + { + "bbox": [ + 47, + 374, + 288, + 437 + ], + "type": "text", + "content": " is the outgoing radiance. Henceforth, we use " + }, + { + "bbox": [ + 47, + 374, + 288, + 437 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 47, + 374, + 288, + 437 + ], + "type": "text", + "content": " to denote a point in " + }, + { + "bbox": [ + 47, + 374, + 288, + 437 + ], + "type": "inline_equation", + "content": "r(t; c, \\omega_o)" + }, + { + "bbox": [ + 47, + 374, + 288, + 437 + ], + "type": "text", + "content": " for notational simplicity." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 46, + 441, + 287, + 489 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 441, + 287, + 489 + ], + "spans": [ + { + "bbox": [ + 46, + 441, + 287, + 489 + ], + "type": "text", + "content": "Light Transport in Volume Rendering. Extracting light sources necessitates analyzing the causes affecting the final ray colors. Kajiya's rendering equation [26] factorizes the outgoing radiance " + }, + { + "bbox": [ + 46, + 441, + 287, + 489 + ], + "type": "inline_equation", + "content": "L_{o}(x,\\omega_{o})" + }, + { + "bbox": [ + 46, + 441, + 287, + 489 + ], + "type": "text", + "content": " into emission and reflections:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 61, + 493, + 287, + 516 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 493, + 287, + 516 + ], + "spans": [ + { + "bbox": [ + 61, + 493, + 287, + 516 + ], + "type": "interline_equation", + "content": "L _ {o} (x, \\omega_ {o}) = E (x) + \\int_ {\\Omega} L _ {i} (x, \\omega_ {i}) R (x, \\omega_ {o}, \\omega_ {i}; b) d \\omega_ {i}, \\tag {2}", + "image_path": "4a7d9f28872fead83a5713fc50177cb7799147bfaf1b7e4e73c2eed801ba0cd2.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 46, + 519, + 287, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 519, + 287, + 665 + ], + "spans": [ + { + "bbox": [ + 46, + 519, + 287, + 665 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 519, + 287, + 665 + ], + "type": "inline_equation", + "content": "E(x)" + }, + { + "bbox": [ + 46, + 519, + 287, + 665 + ], + "type": "text", + "content": " is the emission, " + }, + { + "bbox": [ + 46, + 519, + 287, + 665 + ], + "type": "inline_equation", + "content": "R(x,\\omega_o,\\omega_i;b)" + }, + { + "bbox": [ + 46, + 519, + 287, + 665 + ], + "type": "text", + "content": " represents the SVBRDF parametrized by parameters " + }, + { + "bbox": [ + 46, + 519, + 287, + 665 + ], + "type": "inline_equation", + "content": "b" + }, + { + "bbox": [ + 46, + 519, + 287, + 665 + ], + "type": "text", + "content": " with Lambert cosine multiplied, and " + }, + { + "bbox": [ + 46, + 519, + 287, + 665 + ], + "type": "inline_equation", + "content": "L_{i}(x,\\omega_{i})" + }, + { + "bbox": [ + 46, + 519, + 287, + 665 + ], + "type": "text", + "content": " is the incident radiance. In volume rendering, computing the incident radiance at point " + }, + { + "bbox": [ + 46, + 519, + 287, + 665 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 46, + 519, + 287, + 665 + ], + "type": "text", + "content": " is akin to evaluating Eq. 1, with " + }, + { + "bbox": [ + 46, + 519, + 287, + 665 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 46, + 519, + 287, + 665 + ], + "type": "text", + "content": " serving as the camera center. By iteratively factorizing the outgoing radiance in the incident radiance, the contribution of a path length " + }, + { + "bbox": [ + 46, + 519, + 287, + 665 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 519, + 287, + 665 + ], + "type": "text", + "content": " for a pixel can be decomposed as in Eq. 3, where " + }, + { + "bbox": [ + 46, + 519, + 287, + 665 + ], + "type": "inline_equation", + "content": "\\mathcal{H}_i = \\Pi_{j=1}^{i-1}T(x_j)\\rho(x_j)R(x_j,\\omega_{j-1},\\omega_j)" + }, + { + "bbox": [ + 46, + 519, + 287, + 665 + ], + "type": "text", + "content": " is the path throughput, " + }, + { + "bbox": [ + 46, + 519, + 287, + 665 + ], + "type": "inline_equation", + "content": "S(\\omega_i)" + }, + { + "bbox": [ + 46, + 519, + 287, + 665 + ], + "type": "text", + "content": " is the environment map strength in direction " + }, + { + "bbox": [ + 46, + 519, + 287, + 665 + ], + "type": "inline_equation", + "content": "\\omega_i" + }, + { + "bbox": [ + 46, + 519, + 287, + 665 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 519, + 287, + 665 + ], + "type": "inline_equation", + "content": "V(x,\\omega_i) = \\exp\\left(\\int_0^\\infty -\\rho(r(u;x_i,-\\omega_i))du\\right)" + }, + { + "bbox": [ + 46, + 519, + 287, + 665 + ], + "type": "text", + "content": " is the visibility of the environment map at point " + }, + { + "bbox": [ + 46, + 519, + 287, + 665 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 46, + 519, + 287, + 665 + ], + "type": "text", + "content": " along direction " + }, + { + "bbox": [ + 46, + 519, + 287, + 665 + ], + "type": "inline_equation", + "content": "\\omega_i" + }, + { + "bbox": [ + 46, + 519, + 287, + 665 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 61, + 676, + 287, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 676, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 61, + 676, + 287, + 715 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} P _ {i} = \\int_ {l _ {1}} \\int_ {\\Omega} \\dots \\int_ {l _ {i - 1}} \\int_ {\\Omega} \\left(\\int_ {l _ {i}} T (x _ {i}) \\rho (x _ {i}) E (x _ {i}) d t _ {i} + \\right. \\tag {3} \\\\ S \\left(\\omega_ {i - 1}\\right) V \\left(x _ {i - 1}, \\omega_ {i - 1})\\right) \\mathcal {H} _ {i} d t _ {1} d \\omega_ {1} \\dots d t _ {i - 1} d \\omega_ {i - 1}. \\\\ \\end{array}", + "image_path": "ece9c8256e44f10f1412d125dcab896210ec1a5f3a4a34991fcb39fb31656d93.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 305, + 286, + 546, + 359 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 286, + 546, + 359 + ], + "spans": [ + { + "bbox": [ + 305, + 286, + 546, + 359 + ], + "type": "text", + "content": "Extending the analysis to longer light paths, or equivalently, increasing the number of ray bounces, leads to exponential growth in computation complexity. This poses a challenge when attempting to decompose the influence of unknown emissive sources, as their ability to produce strong reflections makes ignoring indirect illumination infeasible." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 306, + 369, + 389, + 383 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 369, + 389, + 383 + ], + "spans": [ + { + "bbox": [ + 306, + 369, + 389, + 383 + ], + "type": "text", + "content": "4. Methodology" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 305, + 388, + 546, + 473 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 388, + 546, + 473 + ], + "spans": [ + { + "bbox": [ + 305, + 388, + 546, + 473 + ], + "type": "text", + "content": "None of the previous works address the reconstruction of emissive sources from LDR multi-view images. Sec. § 4.1 through § 4.5 detail our method, ESR-NeRF, which reconstructs emissive sources without prior knowledge of scene geometry, materials, or lighting specifics (including their location, number, or colors). We also show how these reconstructed sources can be used for scene editing in § 4.5." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 306, + 479, + 443, + 493 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 479, + 443, + 493 + ], + "spans": [ + { + "bbox": [ + 306, + 479, + 443, + 493 + ], + "type": "text", + "content": "4.1. Learnable Tone-mapper" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 305, + 498, + 545, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 498, + 545, + 533 + ], + "spans": [ + { + "bbox": [ + 305, + 498, + 545, + 533 + ], + "type": "text", + "content": "Throughout the paper, we use " + }, + { + "bbox": [ + 305, + 498, + 545, + 533 + ], + "type": "inline_equation", + "content": "\\mathcal{R}" + }, + { + "bbox": [ + 305, + 498, + 545, + 533 + ], + "type": "text", + "content": " to represent camera rays, " + }, + { + "bbox": [ + 305, + 498, + 545, + 533 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 305, + 498, + 545, + 533 + ], + "type": "text", + "content": " for pixel values, and a binary flag " + }, + { + "bbox": [ + 305, + 498, + 545, + 533 + ], + "type": "inline_equation", + "content": "\\mathbb{I}" + }, + { + "bbox": [ + 305, + 498, + 545, + 533 + ], + "type": "text", + "content": " to indicate whether an image is captured with emissive sources on or off." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 304, + 534, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 534, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 534, + 546, + 713 + ], + "type": "text", + "content": "To extract HDR values from LDR images, we employ the softplus activation for outgoing radiance prediction and apply a clipping and gamma function " + }, + { + "bbox": [ + 304, + 534, + 546, + 713 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 304, + 534, + 546, + 713 + ], + "type": "text", + "content": " [21] for the rendering loss such that " + }, + { + "bbox": [ + 304, + 534, + 546, + 713 + ], + "type": "inline_equation", + "content": "\\hat{C}_{\\tau}(r) = \\tau (\\hat{C} (r))" + }, + { + "bbox": [ + 304, + 534, + 546, + 713 + ], + "type": "text", + "content": ". Unlike previous NeRF-based works [25, 37, 54, 57] that limit radiance to the range of [0, 1], our approach allows for any positive radiance values. Yet, it creates difficulties in differentiating between the surface weight " + }, + { + "bbox": [ + 304, + 534, + 546, + 713 + ], + "type": "inline_equation", + "content": "T(x)\\rho (x)" + }, + { + "bbox": [ + 304, + 534, + 546, + 713 + ], + "type": "text", + "content": " and the magnitude of radiance value " + }, + { + "bbox": [ + 304, + 534, + 546, + 713 + ], + "type": "inline_equation", + "content": "L_{o}(x,\\omega_{o})" + }, + { + "bbox": [ + 304, + 534, + 546, + 713 + ], + "type": "text", + "content": ", since it allows for the possibility of assigning extreme radiance to the points with low surface weights to render same ray colors. Such ambiguity poses challenges, particularly in dark and high-contrast scenes, aggravating surface reconstruction (see Fig. 3). To address this, we introduce a learnable tone-mapper " + }, + { + "bbox": [ + 304, + 534, + 546, + 713 + ], + "type": "inline_equation", + "content": "m_{\\theta}:\\mathbb{R}_{+}^{3}\\to [0,1]^{3}" + }, + { + "bbox": [ + 304, + 534, + 546, + 713 + ], + "type": "text", + "content": ", that takes positionally encoded HDR linear values as input:" + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4600" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 60, + 75, + 126, + 120 + ], + "blocks": [ + { + "bbox": [ + 60, + 75, + 126, + 120 + ], + "lines": [ + { + "bbox": [ + 60, + 75, + 126, + 120 + ], + "spans": [ + { + "bbox": [ + 60, + 75, + 126, + 120 + ], + "type": "image", + "image_path": "a3c9e215d5e892eb640ca4cdf9957db220e3027cb39b41a155a9983267ed6771.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 55, + 125, + 130, + 182 + ], + "blocks": [ + { + "bbox": [ + 55, + 125, + 130, + 182 + ], + "lines": [ + { + "bbox": [ + 55, + 125, + 130, + 182 + ], + "spans": [ + { + "bbox": [ + 55, + 125, + 130, + 182 + ], + "type": "image", + "image_path": "9e5cd54c95fe67b079c517c178a5061f7e5223ce6de42a3cbda1377eee353d56.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 77, + 185, + 109, + 194 + ], + "lines": [ + { + "bbox": [ + 77, + 185, + 109, + 194 + ], + "spans": [ + { + "bbox": [ + 77, + 185, + 109, + 194 + ], + "type": "text", + "content": "(a) Image" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 135, + 75, + 200, + 121 + ], + "blocks": [ + { + "bbox": [ + 135, + 75, + 200, + 121 + ], + "lines": [ + { + "bbox": [ + 135, + 75, + 200, + 121 + ], + "spans": [ + { + "bbox": [ + 135, + 75, + 200, + 121 + ], + "type": "image", + "image_path": "31d8bf7ded52d96af3e1dc3ff1a5af6af66fa47b75a09972fe4ec68bfc51cd8a.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 137, + 125, + 194, + 178 + ], + "blocks": [ + { + "bbox": [ + 137, + 125, + 194, + 178 + ], + "lines": [ + { + "bbox": [ + 137, + 125, + 194, + 178 + ], + "spans": [ + { + "bbox": [ + 137, + 125, + 194, + 178 + ], + "type": "image", + "image_path": "0614aa82aa0d17183c377cbc29216fbf0dcdda4bd9678bb34c6b2ffa50354f00.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 148, + 185, + 186, + 194 + ], + "lines": [ + { + "bbox": [ + 148, + 185, + 186, + 194 + ], + "spans": [ + { + "bbox": [ + 148, + 185, + 186, + 194 + ], + "type": "text", + "content": "(b) " + }, + { + "bbox": [ + 148, + 185, + 186, + 194 + ], + "type": "inline_equation", + "content": "\\mathrm{w / o}m_{\\theta}" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 210, + 75, + 274, + 121 + ], + "blocks": [ + { + "bbox": [ + 210, + 75, + 274, + 121 + ], + "lines": [ + { + "bbox": [ + 210, + 75, + 274, + 121 + ], + "spans": [ + { + "bbox": [ + 210, + 75, + 274, + 121 + ], + "type": "image", + "image_path": "cc73d8dc0bd607c86cea402aa62c9d82edeee5bf331f29c17f1c84ba3f8ad670.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 211, + 125, + 268, + 178 + ], + "blocks": [ + { + "bbox": [ + 211, + 125, + 268, + 178 + ], + "lines": [ + { + "bbox": [ + 211, + 125, + 268, + 178 + ], + "spans": [ + { + "bbox": [ + 211, + 125, + 268, + 178 + ], + "type": "image", + "image_path": "4367bb8019d50bddd926a6b6803b02803e6c00e46538ddaba63eda88a8c5eefe.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 217, + 185, + 266, + 194 + ], + "lines": [ + { + "bbox": [ + 217, + 185, + 266, + 194 + ], + "spans": [ + { + "bbox": [ + 217, + 185, + 266, + 194 + ], + "type": "text", + "content": "(c) ESR-NeRF" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 48, + 199, + 285, + 210 + ], + "lines": [ + { + "bbox": [ + 48, + 199, + 285, + 210 + ], + "spans": [ + { + "bbox": [ + 48, + 199, + 285, + 210 + ], + "type": "text", + "content": "Figure 3. Reconstructed surfaces with the learnable tone-mapper." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 83, + 225, + 287, + 249 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 83, + 225, + 287, + 249 + ], + "spans": [ + { + "bbox": [ + 83, + 225, + 287, + 249 + ], + "type": "interline_equation", + "content": "\\hat {C} _ {m _ {\\theta}} (r) = \\int_ {0} ^ {\\infty} T (x) \\rho (x) m _ {\\theta} \\left(L _ {o} \\left(x, \\omega_ {o}\\right)\\right) d t, \\tag {4}", + "image_path": "f4f5b2426ca0c4c2b41adb8ef74bb0dba14c0829f453033055b1f826e3201772.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 91, + 255, + 287, + 270 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 255, + 287, + 270 + ], + "spans": [ + { + "bbox": [ + 91, + 255, + 287, + 270 + ], + "type": "interline_equation", + "content": "L _ {o} (x, \\omega_ {o}) = L _ {o} ^ {S} (x, \\omega_ {o}) + L _ {o} ^ {E} (x, \\omega_ {o}) \\cdot \\mathbb {I}, \\tag {5}", + "image_path": "151b9488f8099bd49fc0553431ab00b09c184a97d8959cbd8eb0ccd64969e7dc.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 47, + 272, + 287, + 321 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 272, + 287, + 321 + ], + "spans": [ + { + "bbox": [ + 47, + 272, + 287, + 321 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 272, + 287, + 321 + ], + "type": "inline_equation", + "content": "L_{o}^{S}(x,\\omega_{o})" + }, + { + "bbox": [ + 47, + 272, + 287, + 321 + ], + "type": "text", + "content": " is radiance when emissive sources are turned off, while " + }, + { + "bbox": [ + 47, + 272, + 287, + 321 + ], + "type": "inline_equation", + "content": "L_{o}^{E}(x,\\omega_{o})" + }, + { + "bbox": [ + 47, + 272, + 287, + 321 + ], + "type": "text", + "content": " stands for radiance added to the scene by emissive sources. Our rendering loss is then formulated as follows, with " + }, + { + "bbox": [ + 47, + 272, + 287, + 321 + ], + "type": "inline_equation", + "content": "\\lambda_{\\tau}" + }, + { + "bbox": [ + 47, + 272, + 287, + 321 + ], + "type": "text", + "content": " as a hyper-parameter:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 52, + 329, + 287, + 353 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 52, + 329, + 287, + 353 + ], + "spans": [ + { + "bbox": [ + 52, + 329, + 287, + 353 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {r e n d e r}} = \\sum_ {r \\in \\mathcal {R}} \\left(\\| C (r) - \\hat {C} _ {m _ {\\theta}} (r) \\| _ {2} ^ {2} + \\lambda_ {\\tau} \\| C (r) - \\hat {C} _ {\\tau} (r) \\| _ {2} ^ {2}\\right). \\tag {6}", + "image_path": "e105b73acba8dc560ec2f4390a7815517c054cea4a61c96975e904bfe4312d31.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 47, + 366, + 250, + 380 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 366, + 250, + 380 + ], + "spans": [ + { + "bbox": [ + 47, + 366, + 250, + 380 + ], + "type": "text", + "content": "4.2. Learning of Light Transport Segments" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 46, + 385, + 287, + 493 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 385, + 287, + 493 + ], + "spans": [ + { + "bbox": [ + 46, + 385, + 287, + 493 + ], + "type": "text", + "content": "The computational complexity of object appearance analysis in volume rendering is notably high, as shown in Eq. 3. We take an alternative approach by leveraging neural networks to represent ray-traced fields, rather than explicitly tracing every rays. Our distinct contribution to inverse rendering lies in precise adjustment of radiance. Specifically, we impose constraints on the predicted radiance to satisfy each light transport segments. The light transport segments (LTS) loss, " + }, + { + "bbox": [ + 46, + 385, + 287, + 493 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{lts}" + }, + { + "bbox": [ + 46, + 385, + 287, + 493 + ], + "type": "text", + "content": ", plays a pivotal role in our method:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 91, + 501, + 287, + 525 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 501, + 287, + 525 + ], + "spans": [ + { + "bbox": [ + 91, + 501, + 287, + 525 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {l t s} ^ {S} = \\sum_ {x, \\omega_ {o}} \\| L _ {o} ^ {S} (x, \\omega_ {o}) - \\hat {L} _ {o} ^ {S} (x, \\omega_ {o}) \\| _ {2} ^ {2}, \\tag {7}", + "image_path": "425eaec825a3bc4c80569df95d3f74787a07e54b783c2966101e9ae4ed2d2116.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 91, + 535, + 287, + 559 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 535, + 287, + 559 + ], + "spans": [ + { + "bbox": [ + 91, + 535, + 287, + 559 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {l t s} ^ {E} = \\sum_ {x, \\omega_ {o}} \\| L _ {o} ^ {E} (x, \\omega_ {o}) - \\hat {L} _ {o} ^ {E} (x, \\omega_ {o}) \\| _ {2} ^ {2}, \\tag {8}", + "image_path": "b42fcdf9b2728eb314455a7fb204864a6c27c0074b59209c8a23ef4e2c925f48.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 59, + 572, + 252, + 608 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 572, + 252, + 608 + ], + "spans": [ + { + "bbox": [ + 59, + 572, + 252, + 608 + ], + "type": "interline_equation", + "content": "\\hat {L} _ {o} ^ {S} (x, \\omega_ {o}) = \\int_ {\\Omega} \\underbrace {S (\\omega_ {i}) V (x , \\omega_ {i}) R (x , \\omega_ {o} , \\omega_ {i})} _ {\\text {d i r e c t i l l u m i n a t i o n b y a n e v i r o n m e n t m a p}} d \\omega_ {i} +", + "image_path": "c6bf94c860c983b8525b47c966bfac45d8bd3baa6bbcce041fc038bd7fe8a18a.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 62, + 605, + 261, + 638 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 605, + 261, + 638 + ], + "spans": [ + { + "bbox": [ + 62, + 605, + 261, + 638 + ], + "type": "interline_equation", + "content": "\\int_ {\\Omega} \\int_ {0} ^ {\\infty} \\underbrace {T (x ^ {\\prime}) \\rho (x ^ {\\prime}) L _ {o} ^ {S} (x ^ {\\prime} , - \\omega_ {i}) d t ^ {\\prime} R (x , \\omega_ {o} , \\omega_ {i})} _ {\\text {i n d r e c t i l l u m i n a t i o n b y a n e v i r o n m e n t m a p}} d \\omega_ {i}.", + "image_path": "2b46d3bac3b45d8d51c8f415c0bd809fe2ab6afc604d81e8c5037bb73d9d51cf.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 59, + 655, + 141, + 682 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 655, + 141, + 682 + ], + "spans": [ + { + "bbox": [ + 59, + 655, + 141, + 682 + ], + "type": "interline_equation", + "content": "\\hat {L} _ {o} ^ {E} (x, \\omega_ {o}) = \\underbrace {E (x)} _ {\\text {e m i s s i o n}} +", + "image_path": "8b26f1240c0d5ceb52f0b7e130b6ab2702d3b7d5293917724152b3be125c495e.jpg" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 59, + 681, + 287, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 681, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 59, + 681, + 287, + 715 + ], + "type": "interline_equation", + "content": "\\int_ {\\Omega} \\int_ {0} ^ {\\infty} \\underbrace {T \\left(x ^ {\\prime}\\right) \\rho \\left(x ^ {\\prime}\\right) L _ {o} ^ {E} \\left(x ^ {\\prime} , - \\omega_ {i}\\right) d t ^ {\\prime} R \\left(x , \\omega_ {o} , \\omega_ {i}\\right)} _ {\\text {d i r e c t ＆ i n d i r c t i l l u m i n a t i o n b y e m i s s i v e s o u r c e s}} d \\omega_ {i}. \\tag {10}", + "image_path": "8bf4d745b44d8e4d9853262879b9f8471692f1a8643749b7a3735fd37e2aaaff.jpg" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 306, + 72, + 545, + 179 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 545, + 179 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 545, + 179 + ], + "type": "text", + "content": "We ensure consistency between the radiance directly predicted by the network " + }, + { + "bbox": [ + 306, + 72, + 545, + 179 + ], + "type": "inline_equation", + "content": "L_{o}(x,\\omega_{o})" + }, + { + "bbox": [ + 306, + 72, + 545, + 179 + ], + "type": "text", + "content": " and the radiance achievable based on the scene context " + }, + { + "bbox": [ + 306, + 72, + 545, + 179 + ], + "type": "inline_equation", + "content": "\\hat{L}_{o}(x,\\omega_{o})" + }, + { + "bbox": [ + 306, + 72, + 545, + 179 + ], + "type": "text", + "content": ". Previous approaches have focused on matching " + }, + { + "bbox": [ + 306, + 72, + 545, + 179 + ], + "type": "inline_equation", + "content": "\\hat{L}_{o}(x,\\omega_{o})" + }, + { + "bbox": [ + 306, + 72, + 545, + 179 + ], + "type": "text", + "content": " to training views, overlooking the relations to " + }, + { + "bbox": [ + 306, + 72, + 545, + 179 + ], + "type": "inline_equation", + "content": "L_{o}(x,\\omega_{o})" + }, + { + "bbox": [ + 306, + 72, + 545, + 179 + ], + "type": "text", + "content": ". This hinders the restoration of HDR radiance by supervising scene components to LDR training views. In contrast, our LTS loss enables volumetric energy transfer of radiance, adjusting outgoing radiance based on their interrelations." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 305, + 180, + 545, + 276 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 180, + 545, + 276 + ], + "spans": [ + { + "bbox": [ + 305, + 180, + 545, + 276 + ], + "type": "text", + "content": "To implement this concept, we train six dedicated networks for SDF " + }, + { + "bbox": [ + 305, + 180, + 545, + 276 + ], + "type": "inline_equation", + "content": "f(x)" + }, + { + "bbox": [ + 305, + 180, + 545, + 276 + ], + "type": "text", + "content": ", SVBRDF parameters " + }, + { + "bbox": [ + 305, + 180, + 545, + 276 + ], + "type": "inline_equation", + "content": "b(x)" + }, + { + "bbox": [ + 305, + 180, + 545, + 276 + ], + "type": "text", + "content": ", emission " + }, + { + "bbox": [ + 305, + 180, + 545, + 276 + ], + "type": "inline_equation", + "content": "E(x)" + }, + { + "bbox": [ + 305, + 180, + 545, + 276 + ], + "type": "text", + "content": ", environment map " + }, + { + "bbox": [ + 305, + 180, + 545, + 276 + ], + "type": "inline_equation", + "content": "S(\\omega_{i})" + }, + { + "bbox": [ + 305, + 180, + 545, + 276 + ], + "type": "text", + "content": ", outgoing radiances " + }, + { + "bbox": [ + 305, + 180, + 545, + 276 + ], + "type": "inline_equation", + "content": "L_{o}^{S}(x,\\omega_{o})" + }, + { + "bbox": [ + 305, + 180, + 545, + 276 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 180, + 545, + 276 + ], + "type": "inline_equation", + "content": "L_{o}^{E}(x,\\omega_{o})" + }, + { + "bbox": [ + 305, + 180, + 545, + 276 + ], + "type": "text", + "content": ", to adhere to these LTS requirements. For the environment map, we represent it using 48 Spherical Gaussians [67]: " + }, + { + "bbox": [ + 305, + 180, + 545, + 276 + ], + "type": "inline_equation", + "content": "\\sum_{k=1}^{M}\\mu_{k}e^{\\lambda_{k}(\\omega_{i}\\cdot\\xi_{k}-1)}" + }, + { + "bbox": [ + 305, + 180, + 545, + 276 + ], + "type": "text", + "content": ", followed by the softplus activation. " + }, + { + "bbox": [ + 305, + 180, + 545, + 276 + ], + "type": "inline_equation", + "content": "\\mu\\in\\mathbb{R}^{3},\\lambda\\in\\mathbb{R}_{+}" + }, + { + "bbox": [ + 305, + 180, + 545, + 276 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 305, + 180, + 545, + 276 + ], + "type": "inline_equation", + "content": "\\xi\\in\\mathbb{S}^{2}" + }, + { + "bbox": [ + 305, + 180, + 545, + 276 + ], + "type": "text", + "content": " respectively denote the lobe amplitude, sharpness, and axis." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 306, + 282, + 522, + 295 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 282, + 522, + 295 + ], + "spans": [ + { + "bbox": [ + 306, + 282, + 522, + 295 + ], + "type": "text", + "content": "4.3. Progressive Discovery of Reflection Areas" + } + ] + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 307, + 303, + 364, + 364 + ], + "blocks": [ + { + "bbox": [ + 307, + 303, + 364, + 364 + ], + "lines": [ + { + "bbox": [ + 307, + 303, + 364, + 364 + ], + "spans": [ + { + "bbox": [ + 307, + 303, + 364, + 364 + ], + "type": "image", + "image_path": "e1575f4b7ab33fc445b5dd92db9dcbcce6315d95324ef4a2ac79713e6f3f9f93.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 372, + 423, + 426 + ], + "lines": [ + { + "bbox": [ + 305, + 372, + 423, + 426 + ], + "spans": [ + { + "bbox": [ + 305, + 372, + 423, + 426 + ], + "type": "text", + "content": "Figure 4. Left: Image with active emissive sources. Right: Identified emissive sources w/o progressive discovery of reflection areas." + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_caption" + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 365, + 304, + 421, + 364 + ], + "blocks": [ + { + "bbox": [ + 365, + 304, + 421, + 364 + ], + "lines": [ + { + "bbox": [ + 365, + 304, + 421, + 364 + ], + "spans": [ + { + "bbox": [ + 365, + 304, + 421, + 364 + ], + "type": "image", + "image_path": "93ca78abdc92500ef1608c9707574b5b883aec36de75958d071a0a5a87d6809b.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + } + ], + "index": 26 + }, + { + "bbox": [ + 427, + 300, + 545, + 431 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 427, + 300, + 545, + 431 + ], + "spans": [ + { + "bbox": [ + 427, + 300, + 545, + 431 + ], + "type": "text", + "content": "Relying solely on LTS is insufficient for addressing ambiguity arising from low pixel values of emissive sources and intense reflections in adjacent regions, often leading to confusion between emission and reflection. The right image in Fig. 4 shows self-emitting objects restored with the" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 304, + 432, + 545, + 504 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 432, + 545, + 504 + ], + "spans": [ + { + "bbox": [ + 304, + 432, + 545, + 504 + ], + "type": "text", + "content": "naive LTS loss. While emissive sources are small, large areas affected by them are also identified as emissive sources. We propose a reflection-aware progressive approach for precise identification of emissive sources. By leveraging LTS learning, we extend the regions that can be regarded as reflection areas. Fig. 5 illustrates our progressive algorithm." + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 304, + 509, + 545, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 509, + 545, + 640 + ], + "spans": [ + { + "bbox": [ + 304, + 509, + 545, + 640 + ], + "type": "text", + "content": "Reflection-Aware Emission Refinement. Since surface points are unknown and are updated during learning, we opt to utilize rays rather than surface points. This process involves categorizing training rays into two groups: uncertain " + }, + { + "bbox": [ + 304, + 509, + 545, + 640 + ], + "type": "inline_equation", + "content": "(\\mathcal{R}^U)" + }, + { + "bbox": [ + 304, + 509, + 545, + 640 + ], + "type": "text", + "content": " and certain " + }, + { + "bbox": [ + 304, + 509, + 545, + 640 + ], + "type": "inline_equation", + "content": "(\\mathcal{R}^C)" + }, + { + "bbox": [ + 304, + 509, + 545, + 640 + ], + "type": "text", + "content": ". The certain group contains the rays confidently identified as reflection, aiding the transfer of radiance energy to nearby points. For the points in the certain group, we use the Eq. 11 instead of Eq. 10 to exclusively attribute outgoing radiances to reflections. Satisfying the LTS loss on the certain group results in adjusting the outgoing radiances of influential points, as illustrated in Fig. 5(a):" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 306, + 645, + 556, + 677 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 645, + 556, + 677 + ], + "spans": [ + { + "bbox": [ + 306, + 645, + 556, + 677 + ], + "type": "interline_equation", + "content": "\\hat {L} _ {o} ^ {E} \\left(x, \\omega_ {o}\\right) = \\int_ {\\Omega} \\int_ {0} ^ {\\infty} T \\left(x ^ {\\prime}\\right) \\rho \\left(x ^ {\\prime}\\right) L _ {o} ^ {E} \\left(x ^ {\\prime}, - \\omega_ {i}\\right) d t ^ {\\prime} R \\left(x, \\omega_ {o}, \\omega_ {i}\\right) d \\omega_ {i}. \\tag {11}", + "image_path": "f4efbdb0b6c7035f0dd510cf2e6afb0b0106d05a3df78f3cfa73ab4f29cf87cc.jpg" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 305, + 677, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 677, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 305, + 677, + 545, + 714 + ], + "type": "text", + "content": "The uncertain group includes the rays indicating the areas that are undetermined yet as reflection or emission. Using Eq. 12 to compute " + }, + { + "bbox": [ + 305, + 677, + 545, + 714 + ], + "type": "inline_equation", + "content": "\\hat{L}_o^E (x,\\omega_o)" + }, + { + "bbox": [ + 305, + 677, + 545, + 714 + ], + "type": "text", + "content": ", this group adjusts emis" + } + ] + } + ], + "index": 32 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 314, + 757 + ], + "type": "text", + "content": "4601" + } + ] + } + ], + "index": 33 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 287, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 287, + 96 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 287, + 96 + ], + "type": "text", + "content": "sions " + }, + { + "bbox": [ + 47, + 72, + 287, + 96 + ], + "type": "inline_equation", + "content": "E(x)" + }, + { + "bbox": [ + 47, + 72, + 287, + 96 + ], + "type": "text", + "content": " based on the radiance updates by the certain group, where \"sg\" represents the stop-gradient:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 53, + 102, + 287, + 150 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 102, + 287, + 150 + ], + "spans": [ + { + "bbox": [ + 53, + 102, + 287, + 150 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\hat {L} _ {o} ^ {E} (x, \\omega_ {o}) = E (x) + \\\\ \\left. \\operatorname {s g} \\left(\\int_ {\\Omega} \\int_ {0} ^ {\\infty} T \\left(x ^ {\\prime}\\right) \\rho \\left(x ^ {\\prime}\\right) L _ {o} ^ {E} \\left(x ^ {\\prime}, - \\omega_ {i}\\right) d t ^ {\\prime} R \\left(x, \\omega_ {o}, \\omega_ {i}\\right) d \\omega_ {i}\\right). \\right. \\tag {12} \\\\ \\end{array}", + "image_path": "1583571fed48d952a1d6eec3d4ca5e2a2d6076dc53bf8733e5d604b54ae1992f.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 152, + 287, + 223 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 152, + 287, + 223 + ], + "spans": [ + { + "bbox": [ + 47, + 152, + 287, + 223 + ], + "type": "text", + "content": "As shown in Fig. 5(b), this leads to increased emissions for the regions whose radiances are adjusted to account for the reflections in the certain group. Conversely, emissions decrease for the regions where there is little change in outgoing radiance, but incident radiances are increased by surrounding influential points." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 228, + 287, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 228, + 287, + 300 + ], + "spans": [ + { + "bbox": [ + 47, + 228, + 287, + 300 + ], + "type": "text", + "content": "Ray Group Management. As emissions and radiances are adjusted, the groups are dynamically updated at predefined training intervals through the following process. Within the uncertain group, we evaluate the expected emission strength of rays, retaining only those above a threshold " + }, + { + "bbox": [ + 47, + 228, + 287, + 300 + ], + "type": "inline_equation", + "content": "k_{i}" + }, + { + "bbox": [ + 47, + 228, + 287, + 300 + ], + "type": "text", + "content": ". Rays below this threshold are then merged to the certain group:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 306, + 287, + 338 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 306, + 287, + 338 + ], + "spans": [ + { + "bbox": [ + 53, + 306, + 287, + 338 + ], + "type": "interline_equation", + "content": "\\mathcal {R} _ {i} ^ {U} = \\left\\{r \\mid \\max _ {R G B} \\left(\\int_ {0} ^ {\\infty} T (x) \\rho (x) E (x) d t\\right) \\geq k _ {i}, r \\in \\mathcal {R} _ {i - 1} ^ {U} \\right\\}, \\tag {13}", + "image_path": "4aa1b111c30c243ca5b383cce1b476aa7d7d66c7383dd05dc1db3f85e2b090c7.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 107, + 347, + 287, + 365 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 107, + 347, + 287, + 365 + ], + "spans": [ + { + "bbox": [ + 107, + 347, + 287, + 365 + ], + "type": "interline_equation", + "content": "\\mathcal {R} _ {i} ^ {C} = \\left(\\mathcal {R} _ {i - 1} ^ {U} - \\mathcal {R} _ {t} ^ {U}\\right) \\cup \\mathcal {R} _ {i - 1} ^ {C}. \\tag {14}", + "image_path": "67e48d46a970097a7e9eb0aea8c5d83a5f787dddb2a986643d2d46bd7d480cf2.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 368, + 287, + 427 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 368, + 287, + 427 + ], + "spans": [ + { + "bbox": [ + 47, + 368, + 287, + 427 + ], + "type": "text", + "content": "Subsequently, newly added rays to the certain group can be used to localize influential points and update their outgoing radiances. This iterative process progressively refines the separation between reflective and emissive regions, attaining more accurate identification of emissive sources." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 432, + 287, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 432, + 287, + 468 + ], + "spans": [ + { + "bbox": [ + 47, + 432, + 287, + 468 + ], + "type": "text", + "content": "LTS Loss Decomposition. The LTS loss, as detailed in Eq. 15, can be decomposed using a stop-gradient operation to refine the adjustment process." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 67, + 474, + 287, + 499 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 474, + 287, + 499 + ], + "spans": [ + { + "bbox": [ + 67, + 474, + 287, + 499 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {l t s} ^ {E} = \\sum_ {x, \\omega_ {o}} \\left(\\lambda_ {l} \\| \\operatorname {s g} \\left(L _ {o} ^ {E} (x, \\omega_ {o})\\right) - \\hat {L} _ {o} ^ {E} (x, \\omega_ {o}) \\| _ {1} + \\right. \\tag {15}", + "image_path": "c9a6921f7e4e8457b12901d9fe02308a582166790d86d90b4732bc2233dfe890.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 115, + 500, + 250, + 514 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 115, + 500, + 250, + 514 + ], + "spans": [ + { + "bbox": [ + 115, + 500, + 250, + 514 + ], + "type": "interline_equation", + "content": "\\lambda_ {r} \\| L _ {o} ^ {E} (x, \\omega_ {o}) - \\operatorname {s g} (\\hat {L} _ {o} ^ {E} (x, \\omega_ {o})) \\| _ {1}).", + "image_path": "1461ab74ca791cd0b8a8de5c3cd19d7cf17fd83af41dd52bc6f5307d86fa6f58.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 521, + 287, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 521, + 287, + 592 + ], + "spans": [ + { + "bbox": [ + 47, + 521, + 287, + 592 + ], + "type": "text", + "content": "We prioritize " + }, + { + "bbox": [ + 47, + 521, + 287, + 592 + ], + "type": "inline_equation", + "content": "\\lambda_{l}" + }, + { + "bbox": [ + 47, + 521, + 287, + 592 + ], + "type": "text", + "content": " to enhance the update of scene context, affecting other points' radiance given the predicted " + }, + { + "bbox": [ + 47, + 521, + 287, + 592 + ], + "type": "inline_equation", + "content": "L_{o}(x,\\omega_{o})" + }, + { + "bbox": [ + 47, + 521, + 287, + 592 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 47, + 521, + 287, + 592 + ], + "type": "inline_equation", + "content": "\\lambda_{r}" + }, + { + "bbox": [ + 47, + 521, + 287, + 592 + ], + "type": "text", + "content": " prevents severe deviation of every " + }, + { + "bbox": [ + 47, + 521, + 287, + 592 + ], + "type": "inline_equation", + "content": "L_{o}(x,\\omega_{o})" + }, + { + "bbox": [ + 47, + 521, + 287, + 592 + ], + "type": "text", + "content": " within the current scene context. This aligns with our focus on HDR source reconstruction from LDR images, addressing under-represented information in training data." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 599, + 145, + 611 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 599, + 145, + 611 + ], + "spans": [ + { + "bbox": [ + 47, + 599, + 145, + 611 + ], + "type": "text", + "content": "4.4. Training Details" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 47, + 617, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 617, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 617, + 287, + 713 + ], + "type": "text", + "content": "We employ the Voxurf architecture [80] as backbone and adopt the simplified Disney BRDF model [10] for SVBRDF representation, with parameters including base color " + }, + { + "bbox": [ + 47, + 617, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\in [0,1]^3" + }, + { + "bbox": [ + 47, + 617, + 287, + 713 + ], + "type": "text", + "content": ", roughness " + }, + { + "bbox": [ + 47, + 617, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\in [0,1]" + }, + { + "bbox": [ + 47, + 617, + 287, + 713 + ], + "type": "text", + "content": ", and metallic " + }, + { + "bbox": [ + 47, + 617, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\in [0,1]" + }, + { + "bbox": [ + 47, + 617, + 287, + 713 + ], + "type": "text", + "content": ". The learnable tone-mapper, structured as a two-layer MLP, is utilized for the rendering loss only. Initially, we pre-train our networks using the rendering loss, subsequently integrating the basic LTS loss (Eq. 7 and Eq. 8) into our training regimen." + } + ] + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 307, + 70, + 381, + 118 + ], + "blocks": [ + { + "bbox": [ + 307, + 70, + 381, + 118 + ], + "lines": [ + { + "bbox": [ + 307, + 70, + 381, + 118 + ], + "spans": [ + { + "bbox": [ + 307, + 70, + 381, + 118 + ], + "type": "image", + "image_path": "c0f1cde92af7bc78d655d0417f1087a8b316296c02b105f38ca8940638e37e8b.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 318, + 121, + 361, + 131 + ], + "lines": [ + { + "bbox": [ + 318, + 121, + 361, + 131 + ], + "spans": [ + { + "bbox": [ + 318, + 121, + 361, + 131 + ], + "type": "text", + "content": "(a) LTS on " + }, + { + "bbox": [ + 318, + 121, + 361, + 131 + ], + "type": "inline_equation", + "content": "R^C" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 388, + 70, + 462, + 118 + ], + "blocks": [ + { + "bbox": [ + 388, + 70, + 462, + 118 + ], + "lines": [ + { + "bbox": [ + 388, + 70, + 462, + 118 + ], + "spans": [ + { + "bbox": [ + 388, + 70, + 462, + 118 + ], + "type": "image", + "image_path": "12dab668070569d1475cefec26d1388d4bb0c1e5315387551bfbfa752f6ab277.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 399, + 121, + 444, + 131 + ], + "lines": [ + { + "bbox": [ + 399, + 121, + 444, + 131 + ], + "spans": [ + { + "bbox": [ + 399, + 121, + 444, + 131 + ], + "type": "text", + "content": "(b) LTS on " + }, + { + "bbox": [ + 399, + 121, + 444, + 131 + ], + "type": "inline_equation", + "content": "R^U" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 304, + 137, + 545, + 182 + ], + "lines": [ + { + "bbox": [ + 304, + 137, + 545, + 182 + ], + "spans": [ + { + "bbox": [ + 304, + 137, + 545, + 182 + ], + "type": "text", + "content": "Figure 5. Illustration of the progressive emissive source reconstruction with reflection awareness. Gray color represents the areas belonging to the certain group, while the red (emissive sources) and orange (their reflections) areas belong to the uncertain group." + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 469, + 76, + 543, + 118 + ], + "blocks": [ + { + "bbox": [ + 469, + 76, + 543, + 118 + ], + "lines": [ + { + "bbox": [ + 469, + 76, + 543, + 118 + ], + "spans": [ + { + "bbox": [ + 469, + 76, + 543, + 118 + ], + "type": "image", + "image_path": "2ea8217914c690e071b09f71ed17e37be0f69dcf599092c298054181d0ac1531.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 468, + 121, + 544, + 132 + ], + "lines": [ + { + "bbox": [ + 468, + 121, + 544, + 132 + ], + "spans": [ + { + "bbox": [ + 468, + 121, + 544, + 132 + ], + "type": "text", + "content": "(c) Ray Group Managing" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 191, + 545, + 286 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 191, + 545, + 286 + ], + "spans": [ + { + "bbox": [ + 304, + 191, + 545, + 286 + ], + "type": "text", + "content": "This phase transitions to the reflection-aware progressive training scheme, where we adopt the " + }, + { + "bbox": [ + 304, + 191, + 545, + 286 + ], + "type": "inline_equation", + "content": "\\ell_1" + }, + { + "bbox": [ + 304, + 191, + 545, + 286 + ], + "type": "text", + "content": " loss due to its empirical stability in refining emissive source reconstruction. We use a smoothing regularization to promote local consistency in normals, BRDFs, and emissions. To ensure view-consistent labeling of 3D points as either reflective or emissive, we implement the emission suppression loss for points belonging to the certain group:" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 347, + 291, + 545, + 320 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 347, + 291, + 545, + 320 + ], + "spans": [ + { + "bbox": [ + 347, + 291, + 545, + 320 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {s u p p}} ^ {E} = \\sum_ {r \\in R _ {t} ^ {C}} \\| \\int_ {0} ^ {\\infty} T (x) \\rho (x) E (x) d t \\| _ {2} ^ {2}, \\tag {16}", + "image_path": "bb17ae20b18502e696062dc4bd6dadcd5af2bbf825308a100df24dd657d64ac8.jpg" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 304, + 326, + 545, + 386 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 326, + 545, + 386 + ], + "spans": [ + { + "bbox": [ + 304, + 326, + 545, + 386 + ], + "type": "text", + "content": "The threshold " + }, + { + "bbox": [ + 304, + 326, + 545, + 386 + ], + "type": "inline_equation", + "content": "k_{i}" + }, + { + "bbox": [ + 304, + 326, + 545, + 386 + ], + "type": "text", + "content": " linearly increases with each time step " + }, + { + "bbox": [ + 304, + 326, + 545, + 386 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 326, + 545, + 386 + ], + "type": "text", + "content": ", utilizing a grid search within a range of " + }, + { + "bbox": [ + 304, + 326, + 545, + 386 + ], + "type": "inline_equation", + "content": "[10^{-3}, 10^{-5}]" + }, + { + "bbox": [ + 304, + 326, + 545, + 386 + ], + "type": "text", + "content": " to find the slope. We construct mini-batches via stratified sampling within each group. For a detailed description of our training procedure, please refer to Appendix." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 306, + 392, + 394, + 405 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 392, + 394, + 405 + ], + "spans": [ + { + "bbox": [ + 306, + 392, + 394, + 405 + ], + "type": "text", + "content": "4.5. Scene Editing" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 304, + 411, + 545, + 458 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 411, + 545, + 458 + ], + "spans": [ + { + "bbox": [ + 304, + 411, + 545, + 458 + ], + "type": "text", + "content": "Reconstructed emissive sources enable scene editing; users select emissive sources using binary masks " + }, + { + "bbox": [ + 304, + 411, + 545, + 458 + ], + "type": "inline_equation", + "content": "M_{j=1\\dots N}" + }, + { + "bbox": [ + 304, + 411, + 545, + 458 + ], + "type": "text", + "content": " and specify lighting conditions using colors " + }, + { + "bbox": [ + 304, + 411, + 545, + 458 + ], + "type": "inline_equation", + "content": "c_{j=1\\dots N}" + }, + { + "bbox": [ + 304, + 411, + 545, + 458 + ], + "type": "text", + "content": " and intensities " + }, + { + "bbox": [ + 304, + 411, + 545, + 458 + ], + "type": "inline_equation", + "content": "i_{j=1\\dots N}" + }, + { + "bbox": [ + 304, + 411, + 545, + 458 + ], + "type": "text", + "content": " within the HSV color space [53]." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 304, + 459, + 545, + 494 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 459, + 545, + 494 + ], + "spans": [ + { + "bbox": [ + 304, + 459, + 545, + 494 + ], + "type": "text", + "content": "We identify the rays in the uncertain group that match " + }, + { + "bbox": [ + 304, + 459, + 545, + 494 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 304, + 459, + 545, + 494 + ], + "type": "text", + "content": " by projecting expected surface points " + }, + { + "bbox": [ + 304, + 459, + 545, + 494 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 304, + 459, + 545, + 494 + ], + "type": "text", + "content": " of the rays onto the camera with the pose " + }, + { + "bbox": [ + 304, + 459, + 545, + 494 + ], + "type": "inline_equation", + "content": "\\mathbf{R}|\\mathbf{t}" + }, + { + "bbox": [ + 304, + 459, + 545, + 494 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 378, + 499, + 545, + 522 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 378, + 499, + 545, + 522 + ], + "spans": [ + { + "bbox": [ + 378, + 499, + 545, + 522 + ], + "type": "interline_equation", + "content": "p = \\int_ {0} ^ {\\infty} T (x) \\rho (x) x d t, \\tag {17}", + "image_path": "8fe1d6feb5df35b091218624928dbdf3e4b9cde75fa1acf474342db46c63e891.jpg" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 311, + 539, + 545, + 554 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 539, + 545, + 554 + ], + "spans": [ + { + "bbox": [ + 311, + 539, + 545, + 554 + ], + "type": "interline_equation", + "content": "\\mathbb {I} _ {j} ^ {\\text {h i t}} (x) = \\operatorname {i n t e r p} \\left(M _ {j}, p ^ {\\prime}\\right) > 0, \\text {w h e r e} p ^ {\\prime} = \\mathbf {K} [ \\mathbf {R} | \\mathbf {t} ] [ p | 1 ] ^ {T}. \\tag {18}", + "image_path": "2864fa1c13b3e79b230bc688bd733a3f2c95c6937b77b994ede7d7b6c61c7a62.jpg" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 304, + 559, + 545, + 619 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 559, + 545, + 619 + ], + "spans": [ + { + "bbox": [ + 304, + 559, + 545, + 619 + ], + "type": "text", + "content": "For the rays satisfying " + }, + { + "bbox": [ + 304, + 559, + 545, + 619 + ], + "type": "inline_equation", + "content": "\\mathbb{I}_j^{hit}(x)" + }, + { + "bbox": [ + 304, + 559, + 545, + 619 + ], + "type": "text", + "content": ", we apply the designated lighting conditions. The new emission values are computed by substituting the original hue (H) and saturation (S) of " + }, + { + "bbox": [ + 304, + 559, + 545, + 619 + ], + "type": "inline_equation", + "content": "E(x)" + }, + { + "bbox": [ + 304, + 559, + 545, + 619 + ], + "type": "text", + "content": " with the user-specified color " + }, + { + "bbox": [ + 304, + 559, + 545, + 619 + ], + "type": "inline_equation", + "content": "c_{j}" + }, + { + "bbox": [ + 304, + 559, + 545, + 619 + ], + "type": "text", + "content": " and adjusting the value (V) of " + }, + { + "bbox": [ + 304, + 559, + 545, + 619 + ], + "type": "inline_equation", + "content": "v(x)" + }, + { + "bbox": [ + 304, + 559, + 545, + 619 + ], + "type": "text", + "content": " with the new intensity " + }, + { + "bbox": [ + 304, + 559, + 545, + 619 + ], + "type": "inline_equation", + "content": "i_j" + }, + { + "bbox": [ + 304, + 559, + 545, + 619 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 624, + 545, + 639 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 624, + 545, + 639 + ], + "spans": [ + { + "bbox": [ + 307, + 624, + 545, + 639 + ], + "type": "interline_equation", + "content": "E (x) = \\operatorname {h s v} _ {-} \\operatorname {t o} _ {-} \\operatorname {r g b} ([ c _ {j} | (v (x) \\times i) ]) \\cdot \\mathbb {I} _ {j} ^ {h i t} + E (x) \\cdot \\neg \\mathbb {I} _ {j} ^ {h i t}. \\tag {19}", + "image_path": "35a5492bd87f1de37fb15a2901eb1b23c1d14b5781bf720fe6ee14bd048f6470.jpg" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 304, + 644, + 545, + 681 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 644, + 545, + 681 + ], + "spans": [ + { + "bbox": [ + 304, + 644, + 545, + 681 + ], + "type": "text", + "content": "These modifications influence scene appearance by optimizing the loss in Eq. 20. During this process, all networks, except for " + }, + { + "bbox": [ + 304, + 644, + 545, + 681 + ], + "type": "inline_equation", + "content": "L_{o}^{E}(x,\\omega_{o})" + }, + { + "bbox": [ + 304, + 644, + 545, + 681 + ], + "type": "text", + "content": ", are frozen:" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 340, + 685, + 545, + 709 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 340, + 685, + 545, + 709 + ], + "spans": [ + { + "bbox": [ + 340, + 685, + 545, + 709 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {e d i t}} = \\sum_ {x, \\omega_ {o}} \\| L _ {o} ^ {E} (x, \\omega_ {o}) - \\operatorname {s g} \\left(\\hat {L} _ {o} ^ {E} (x, \\omega_ {o})\\right) \\| _ {2} ^ {2}. \\tag {20}", + "image_path": "20ecfbe241ed48fdcf3275ed8d4a4881bf42a77cb0640c27675b3d804b1d0e10.jpg" + } + ] + } + ], + "index": 31 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "4602" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 50, + 60, + 545, + 140 + ], + "blocks": [ + { + "bbox": [ + 50, + 60, + 545, + 140 + ], + "lines": [ + { + "bbox": [ + 50, + 60, + 545, + 140 + ], + "spans": [ + { + "bbox": [ + 50, + 60, + 545, + 140 + ], + "type": "table", + "html": "
LegoGiftWhite colored BookCubeBillboardBallsVivid colored
IoUMSEIoUMSEIoUMSEIoUMSEIoUMSEIoUMSEIoUMSEIoUMSEIoUMSEIoUMSEIoUMSE
Twins0.2220.190.498.590.633.910.9531.830.691.120.900.060.256.960.246.090.552.630.9510.640.090.750.830.04
NeLF++0.4320.880.079.380.954.640.9332.670.011.950.910.800.307.650.096.860.953.360.9411.490.021.570.920.78
TensoIR0.7120.130.158.550.953.870.9531.730.761.110.950.050.336.930.156.050.952.590.9610.600.770.740.950.03
ESR-NeRF0.818.380.603.490.961.190.9717.870.840.460.950.040.515.480.592.500.960.510.977.940.880.260.940.03
", + "image_path": "6596b2f39001e3c50406544f3244b3be821d46fc08e1b2cafde2b6713afc71e5.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 191, + 128, + 205 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 191, + 128, + 205 + ], + "spans": [ + { + "bbox": [ + 47, + 191, + 128, + 205 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 212, + 288, + 284 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 212, + 288, + 284 + ], + "spans": [ + { + "bbox": [ + 46, + 212, + 288, + 284 + ], + "type": "text", + "content": "We assess ESR-NeRF in reconstructing emissive sources by focusing on both identification and intensity restoration. To showcase its effectiveness, we conduct a range of experiments, including scene editing, ablation studies, illumination decomposition, and surface reconstruction, providing both quantitative and qualitative results." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 293, + 165, + 306 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 293, + 165, + 306 + ], + "spans": [ + { + "bbox": [ + 47, + 293, + 165, + 306 + ], + "type": "text", + "content": "5.1. Experiment Settings" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 312, + 289, + 553 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 312, + 289, + 553 + ], + "spans": [ + { + "bbox": [ + 46, + 312, + 289, + 553 + ], + "type": "text", + "content": "We curate 6 diverse synthetic scenes, each with 200 training images evenly distributed between on and off lighting conditions. To evaluate the robustness of our approach against light colors, we consider two distinct settings of white colored and vivid colored emissive sources, resulting in a total of 12 scenes. The vivid colors are selected with full saturation in the HSV color space. We measure source identification and radiance reconstruction using IoU and MSE metrics on novel view test images, comparing against ground truth data from Blender-rendered emission masks and EXR files. The emission strengths, the maximum EXR file values, range from 2 to 200. For quantitative scene editing evaluation, we alter the white-colored sources to various colors—red, green, blue, cyan, magenta, yellow—and adjust intensities to half or double their original values. Qualitative results include scene editing for vividly colored sources and real scenes captured with a Fuji 100s camera using Philips smart bulbs as emissive sources. Quantitative assessments are based on 50 test images from novel camera poses, except for MSE measured for 25 test" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 50, + 556, + 285, + 682 + ], + "blocks": [ + { + "bbox": [ + 50, + 556, + 285, + 682 + ], + "lines": [ + { + "bbox": [ + 50, + 556, + 285, + 682 + ], + "spans": [ + { + "bbox": [ + 50, + 556, + 285, + 682 + ], + "type": "image", + "image_path": "d007f314138098088d81eb12ba8af87dd4e8655fb8e78e7e43cb2e7a13fce647.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 691, + 288, + 715 + ], + "lines": [ + { + "bbox": [ + 46, + 691, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 691, + 288, + 715 + ], + "type": "text", + "content": "Figure 6. Comparison of identified emissive sources. ESR-NeRF excels through the reflection-aware progressive refinement." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 306, + 189, + 550, + 277 + ], + "blocks": [ + { + "bbox": [ + 46, + 147, + 547, + 182 + ], + "lines": [ + { + "bbox": [ + 46, + 147, + 547, + 182 + ], + "spans": [ + { + "bbox": [ + 46, + 147, + 547, + 182 + ], + "type": "text", + "content": "Table 2. Results of emissive source identification. ESR-NeRF outperforms state-of-the-art re-lighting methods in reconstructing emissive sources, regardless of their color. The IoU measures the source area identification (a higher value is better), and the MSE quantifies the difference between reconstructed images and HDR ground truth images (a lower value is better)." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 306, + 189, + 550, + 277 + ], + "lines": [ + { + "bbox": [ + 306, + 189, + 550, + 277 + ], + "spans": [ + { + "bbox": [ + 306, + 189, + 550, + 277 + ], + "type": "table", + "html": "
NVNV + INV+ CNV + I + C
PSNRLPIPSPSNRLPIPSPSNRLPIPSPSNRLPIPS
Twins36.520.014127.910.025231.020.025228.210.0310
NeRF-W36.440.014224.770.0417----
NeILF++24.400.055624.710.057924.060.075023.240.0770
TensoIR38.040.010327.280.041826.360.050525.180.0531
PaletteNeRF33.660.023323.270.048324.440.064622.580.0703
ESR-NeRF38.790.008329.990.019331.730.019631.630.0199
", + "image_path": "703f7d45c49734b95212d29c670341090b96904f8005a10cd8e6a3d5da56c7c7.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 285, + 547, + 317 + ], + "lines": [ + { + "bbox": [ + 305, + 285, + 547, + 317 + ], + "spans": [ + { + "bbox": [ + 305, + 285, + 547, + 317 + ], + "type": "text", + "content": "Table 3. Scene editing results. NV: novel view synthesis, I: intensity editing, and C: color editing. A higher PSNR or lower LPIPS value is better." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 305, + 323, + 545, + 371 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 323, + 545, + 371 + ], + "spans": [ + { + "bbox": [ + 305, + 323, + 545, + 371 + ], + "type": "text", + "content": "images. We denote the best performance with blue and the second-best with green. Additionally, we utilize the DTU dataset [23] to evaluate ESR-NeRF's performance in surface reconstruction tasks where emissive sources are absent." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 376, + 546, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 376, + 546, + 568 + ], + "spans": [ + { + "bbox": [ + 304, + 376, + 546, + 568 + ], + "type": "text", + "content": "Baselines. We select two state-of-the-art re-lighting methods, TensoIR [25] and NeILF++ [90], that do not require prior lighting information. For thorough evaluation, we also implement a simple method, Twins, where separate models are trained under light on and off conditions. The Twins utilize the radiance discrepancies between the on and off models to distinguish and adjust emissive sources. For scene editing, we add NeRF-W [38] and PaletteNeRF [28] as baselines. Both NeRF-W and Twins adopt the Voxurf [80] architecture for fair comparison. For methods unable to individually control emissive sources, all sources are adjusted together to match the last lighting condition by a user. For the DTU dataset, we include state-of-the-art surface reconstruction methods that use object masks, such as NeuS [68] and Voxurf, as well as Neural-PBIR [57], that jointly reconstructs surfaces, materials, and environment maps." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 575, + 362, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 575, + 362, + 586 + ], + "spans": [ + { + "bbox": [ + 306, + 575, + 362, + 586 + ], + "type": "text", + "content": "5.2. Results" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 594, + 545, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 545, + 715 + ], + "type": "text", + "content": "Emissive Source Recosnstruction. Tab. 2 shows that our approach excels in accurately identifying emissive source regions and restoring their intensity, regardless of the source color. While TensoIR and NeILF++ can restore emissions by modifying their physical rendering equations, they suffer from emissive source ambiguity, leading to near-zero IoU performance (see Appendix). For a comprehensive comparison, we report the best performance of the baseline methods using thresholding on the reconstructed emission strength at 0.01 intervals. ESR-NeRF consistently outper" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4603" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 57, + 534, + 322 + ], + "blocks": [ + { + "bbox": [ + 56, + 57, + 534, + 322 + ], + "lines": [ + { + "bbox": [ + 56, + 57, + 534, + 322 + ], + "spans": [ + { + "bbox": [ + 56, + 57, + 534, + 322 + ], + "type": "image", + "image_path": "d36286269b953d85620873449b3bf2d8cdd8c1686bd39beb6787efc64e786250.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 325, + 546, + 348 + ], + "lines": [ + { + "bbox": [ + 46, + 325, + 546, + 348 + ], + "spans": [ + { + "bbox": [ + 46, + 325, + 546, + 348 + ], + "type": "text", + "content": "Figure 7. Comparison of scene editing. ESR-NeRF provides precise source control and faithfully represents reflection effects. For easy comparison in the Cube scene of low intensity, the bottom-right images are presented with a " + }, + { + "bbox": [ + 46, + 325, + 546, + 348 + ], + "type": "inline_equation", + "content": "40\\%" + }, + { + "bbox": [ + 46, + 325, + 546, + 348 + ], + "type": "text", + "content": " increased brightness." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 353, + 289, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 353, + 289, + 450 + ], + "spans": [ + { + "bbox": [ + 46, + 353, + 289, + 450 + ], + "type": "text", + "content": "forms the baselines in identifying emissive source regions across all scenes. Our method also achieves significantly lower MSE values for restoring LDR to HDR images compared to the baselines, demonstrating its effectiveness of handling the ill-posed nature of the scenes with emissive sources. This is visually confirmed in Fig. 6, where ESR-NeRF surpasses the baselines in a complex scene with numerous small light bulbs." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 456, + 288, + 636 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 456, + 288, + 636 + ], + "spans": [ + { + "bbox": [ + 46, + 456, + 288, + 636 + ], + "type": "text", + "content": "Scene Editing. Tab. 3 and Fig. 7 showcase the scene editing results under novel lighting conditions. Baseline methods struggle to adapt to lighting changes due to their inability to reconstruct emissive sources accurately. For example, in the Lego scene, TensoIR fails to adjust the illumination in surrounding regions when the color of emissive sources is changed, and in the Cube scene, both the hidden iPad screen and the cube surface covered by the user input mask change together. Twins introduces blue light onto yellow and red surfaces, leading to unintended white and purple appearances, even though there should be no reflection. PaletteNeRF, which manipulates scenes through re-colorization, lacks precise control over illumination, as seen in the synchronous color changes in the yellow ribbon and lighting. In contrast, ESR-NeRF demonstrates superior performance in" + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 50, + 643, + 288, + 700 + ], + "blocks": [ + { + "bbox": [ + 50, + 643, + 288, + 700 + ], + "lines": [ + { + "bbox": [ + 50, + 643, + 288, + 700 + ], + "spans": [ + { + "bbox": [ + 50, + 643, + 288, + 700 + ], + "type": "image", + "image_path": "6dcd2109976ab572d9a3bbbb7d499046b79d2ce40f45aafb878806bc0c0c0012.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 704, + 279, + 715 + ], + "lines": [ + { + "bbox": [ + 54, + 704, + 279, + 715 + ], + "spans": [ + { + "bbox": [ + 54, + 704, + 279, + 715 + ], + "type": "text", + "content": "Figure 8. Reconstructed emitter and re-lighting at novel view." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 353, + 545, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 353, + 545, + 437 + ], + "spans": [ + { + "bbox": [ + 304, + 353, + 545, + 437 + ], + "type": "text", + "content": "scene editing outshining all baselines thanks to the accurate identification of emissive sources, as detailed in Table 3. ESR-NeRF effectively balances source reconstruction and novel view synthesis, ensuring high performance in both tasks. NeRF-W is excluded from color adjustments since it doesn't support direct color change through interpolating latent variables learned with light on and off conditions." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 437, + 546, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 437, + 546, + 521 + ], + "spans": [ + { + "bbox": [ + 304, + 437, + 546, + 521 + ], + "type": "text", + "content": "Fig. 8 to 9 present additional examples of emissive source reconstruction and scene editing results. Fig. 10 shows results on real scenes, for which due to the impracticality of precise control over smart bulb colors, we offer emission reconstruction results with pseudo ground truth data. Our method effectively identifies emissive sources in real scenes, while it faces challenges in capturing complex" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 306, + 522, + 545, + 701 + ], + "blocks": [ + { + "bbox": [ + 306, + 522, + 545, + 701 + ], + "lines": [ + { + "bbox": [ + 306, + 522, + 545, + 701 + ], + "spans": [ + { + "bbox": [ + 306, + 522, + 545, + 701 + ], + "type": "image", + "image_path": "cacfadfd51a4853b5aea10c887268a2ed8cd3fc64bf02d6e4591d743f3084d9f.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 315, + 704, + 535, + 716 + ], + "lines": [ + { + "bbox": [ + 315, + 704, + 535, + 716 + ], + "spans": [ + { + "bbox": [ + 315, + 704, + 535, + 716 + ], + "type": "text", + "content": "Figure 9. Results of source reconstruction and scene editing." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4604" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 47, + 64, + 287, + 201 + ], + "blocks": [ + { + "bbox": [ + 47, + 64, + 287, + 201 + ], + "lines": [ + { + "bbox": [ + 47, + 64, + 287, + 201 + ], + "spans": [ + { + "bbox": [ + 47, + 64, + 287, + 201 + ], + "type": "image", + "image_path": "b4b10d7590b1586aa937b771c07a96207d6996a5bc485653d2fe26cb0c572bc2.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 47, + 208, + 286, + 220 + ], + "lines": [ + { + "bbox": [ + 47, + 208, + 286, + 220 + ], + "spans": [ + { + "bbox": [ + 47, + 208, + 286, + 220 + ], + "type": "text", + "content": "Figure 10. Source reconstruction and scene editing on real scenes." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 227, + 287, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 227, + 287, + 251 + ], + "spans": [ + { + "bbox": [ + 46, + 227, + 287, + 251 + ], + "type": "text", + "content": "reflections within light bulbs, as evident in the bright spot at the center of the bulbs in the ground truth edit results." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 256, + 287, + 340 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 256, + 287, + 340 + ], + "spans": [ + { + "bbox": [ + 46, + 256, + 287, + 340 + ], + "type": "text", + "content": "Ablation Analysis. Progressive refinement with the stop-gradient operation in Eq. 15 improves the identification of emissive sources and reduces MSE values. Without " + }, + { + "bbox": [ + 46, + 256, + 287, + 340 + ], + "type": "inline_equation", + "content": "m_{\\theta}" + }, + { + "bbox": [ + 46, + 256, + 287, + 340 + ], + "type": "text", + "content": ", surface reconstructions become unreliable, complicating the accurate reconstruction of emissive sources. This issue is evident from the CD metrics and illustrated in Fig. 3. Further analyses are provided in Appendix." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 345, + 287, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 345, + 287, + 453 + ], + "spans": [ + { + "bbox": [ + 46, + 345, + 287, + 453 + ], + "type": "text", + "content": "Illumination Decomposition. Fig. 11 demonstrates ESR-NeRF's decomposition of scene illumination into direct and indirect lighting from an environment map, as well as emissions and their reflections. The shadow behind the yellow ribbon in the direct figure and the illumination in the indirect figure showcase ESR-NeRF's ability to model both direct and indirect illumination. The reflection figure shows that our method accurately captures how emissive sources contribute to reflections on nearby regions." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 458, + 287, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 458, + 287, + 578 + ], + "spans": [ + { + "bbox": [ + 46, + 458, + 287, + 578 + ], + "type": "text", + "content": "Surface Reconstruction. Interestingly, our approach can be applied to the scenes without emissive sources to enhance surface reconstruction, as evidenced by the lower CD values in Tab. 4 on the DTU dataset. For this experiment, we use Eq. 7 to 10 without our progressive refinement technique. ESR-NeRF's ability to adjust interrelated outgoing radiances helps prevent surface formations where radiances cannot be produced, considering the predicted scene context. Additional visualizations of the normals, BRDF, and environment maps are provided in Appendix." + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 48, + 585, + 287, + 699 + ], + "blocks": [ + { + "bbox": [ + 48, + 585, + 287, + 699 + ], + "lines": [ + { + "bbox": [ + 48, + 585, + 287, + 699 + ], + "spans": [ + { + "bbox": [ + 48, + 585, + 287, + 699 + ], + "type": "image", + "image_path": "6a30198e34f9cbaf226af7585a64625f99845035691aac99a7632a201aee4000.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 67, + 703, + 266, + 714 + ], + "lines": [ + { + "bbox": [ + 67, + 703, + 266, + 714 + ], + "spans": [ + { + "bbox": [ + 67, + 703, + 266, + 714 + ], + "type": "text", + "content": "Figure 11. An example of illumination decomposition." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 345, + 60, + 506, + 213 + ], + "blocks": [ + { + "bbox": [ + 345, + 60, + 506, + 213 + ], + "lines": [ + { + "bbox": [ + 345, + 60, + 506, + 213 + ], + "spans": [ + { + "bbox": [ + 345, + 60, + 506, + 213 + ], + "type": "table", + "html": "
ScanNeuSVoxurfNeural-PBIRESR-NeRF
240.830.650.570.58
370.980.740.750.71
400.560.390.380.38
550.370.350.360.33
631.130.961.040.93
650.590.640.730.57
690.600.850.650.78
831.451.581.281.18
970.951.010.970.95
1050.780.680.760.58
1060.520.600.530.54
1101.431.110.841.08
1140.360.370.380.33
1180.450.450.460.40
1220.450.470.490.44
mean0.770.720.680.65
", + "image_path": "fd10cc43e36a8259817e68e74a8789ca946a56f4684eac19aa394d68849b77f3.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 309, + 246, + 470, + 305 + ], + "blocks": [ + { + "bbox": [ + 305, + 218, + 545, + 239 + ], + "lines": [ + { + "bbox": [ + 305, + 218, + 545, + 239 + ], + "spans": [ + { + "bbox": [ + 305, + 218, + 545, + 239 + ], + "type": "text", + "content": "Table 4. Results of surface reconstruction via the Chamfer distance on the DTU dataset. A lower value is better." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 309, + 246, + 470, + 305 + ], + "lines": [ + { + "bbox": [ + 309, + 246, + 470, + 305 + ], + "spans": [ + { + "bbox": [ + 309, + 246, + 470, + 305 + ], + "type": "table", + "html": "
WhiteVivid
IoU ↑MSE ↓IoU ↑MSE ↓
w/o progressive0.409.920.413.93
w/o sg0.716.450.603.47
ESR-NeRF0.865.240.812.79
", + "image_path": "465a2083f3cc320eaf1015bec74cbf59919fb2eaa70527bb04f1fc03bc735466.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 476, + 247, + 542, + 304 + ], + "blocks": [ + { + "bbox": [ + 476, + 247, + 542, + 304 + ], + "lines": [ + { + "bbox": [ + 476, + 247, + 542, + 304 + ], + "spans": [ + { + "bbox": [ + 476, + 247, + 542, + 304 + ], + "type": "table", + "html": "
DTU\nCD ↓
w/o mθ0.93
w/o LTS0.71
ESR-NeRF0.65
", + "image_path": "d47dabe4d6803dc5a8ff53273871aae0a0776e4b4c4c94310f24ff7f6635e82f.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 308, + 545, + 331 + ], + "lines": [ + { + "bbox": [ + 305, + 308, + 545, + 331 + ], + "spans": [ + { + "bbox": [ + 305, + 308, + 545, + 331 + ], + "type": "text", + "content": "Table 5. Ablation studies on the surface reconstruction (left) and the emissive source reconstruction (right)." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 306, + 340, + 378, + 352 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 340, + 378, + 352 + ], + "spans": [ + { + "bbox": [ + 306, + 340, + 378, + 352 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 360, + 545, + 431 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 360, + 545, + 431 + ], + "spans": [ + { + "bbox": [ + 304, + 360, + 545, + 431 + ], + "type": "text", + "content": "We present ESR-NeRF as the first NeRF-based inverse rendering method for the scenes with emissive sources. Our approach uses LDR images, eliminating the need of HDR images to reconstruct emissive sources. Furthermore, we demonstrate the application of reconstructed sources in scene editing, enabling color and intensity modifications." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 432, + 545, + 563 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 432, + 545, + 563 + ], + "spans": [ + { + "bbox": [ + 304, + 432, + 545, + 563 + ], + "type": "text", + "content": "Limitations. Future work could explore using a single lighting condition to disentangle emissive sources, environmental lighting, and object texture. It is also promising to address the challenge of volume ray tracing in unbounded scenes to extend to indoor scenes. Additionally, LTS based re-lighting may be weak in representing new colors that traverse unobserved light paths during training. An alternative approach could be extracting emission texture maps and modifying it using the engines such as Blender [13] or Mitsuba [22]. More details on alternative re-lighting methods and radiance fine-tuning are provided in Appendix." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 574, + 421, + 587 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 574, + 421, + 587 + ], + "spans": [ + { + "bbox": [ + 306, + 574, + 421, + 587 + ], + "type": "text", + "content": "7. Acknowledgements" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "text", + "content": "This work was supported by Samsung Electronics MX, Basic Science Research Program through the National Research Foundation of Korea(NRF) funded by the Ministry of Education(RS-2023-00274280), and Institute of Information & Communications Technology Planning & Evaluation (IITP) grant funded by the Korea government (MSIT) (No. 2019-0-01082, SW StarLab; No. 2022-0-00156, Fundamental research on continual meta-learning for quality enhancement of casual videos and their 3D metaverse transformation). Gunhee Kim is the corresponding author." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4605" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 286, + 712 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 53, + 91, + 286, + 123 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 286, + 123 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 286, + 123 + ], + "type": "text", + "content": "[1] Benjamin Attal, Jia-Bin Huang, Michael Zollhöfer, Johannes Kopf, and Changil Kim. Learning neural light fields with ray-space embedding. In CVPR, 2022. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 125, + 286, + 158 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 125, + 286, + 158 + ], + "spans": [ + { + "bbox": [ + 53, + 125, + 286, + 158 + ], + "type": "text", + "content": "[2] Dejan Azinovic, Tzu-Mao Li, Anton Kaplanyan, and Matthias Nießner. Inverse path tracing for joint material and lighting estimation. In CVPR, 2019. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 159, + 286, + 202 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 159, + 286, + 202 + ], + "spans": [ + { + "bbox": [ + 54, + 159, + 286, + 202 + ], + "type": "text", + "content": "[3] Jonathan T Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P Srinivasan. Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In ICCV, 2021. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 205, + 286, + 236 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 205, + 286, + 236 + ], + "spans": [ + { + "bbox": [ + 54, + 205, + 286, + 236 + ], + "type": "text", + "content": "[4] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Mip-nerf 360: Unbounded anti-aliased neural radiance fields. In CVPR, 2022. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 239, + 286, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 239, + 286, + 270 + ], + "spans": [ + { + "bbox": [ + 53, + 239, + 286, + 270 + ], + "type": "text", + "content": "[5] Jonathan T. Barron, Ben Mildenhall, Dor Verbin, Pratul P. Srinivasan, and Peter Hedman. Zip-nerf: Anti-aliased grid-based neural radiance fields. In ICCV, 2023. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 273, + 286, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 273, + 286, + 316 + ], + "spans": [ + { + "bbox": [ + 53, + 273, + 286, + 316 + ], + "type": "text", + "content": "[6] Sai Bi, Zexiang Xu, Pratul Srinivasan, Ben Mildenhall, Kalyan Sunkavalli, Miloš Hasan, Yannick Hold-Geoffroy, David Kriegman, and Ravi Ramamoorthi. Neural reflectance fields for appearance acquisition. In arXiv, 2020. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 318, + 286, + 349 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 318, + 286, + 349 + ], + "spans": [ + { + "bbox": [ + 53, + 318, + 286, + 349 + ], + "type": "text", + "content": "[7] Mark Boss, Varun Jampani, Kihwan Kim, Hendrik Lensch, and Jan Kautz. Two-shot spatially-varying brdf and shape estimation. In CVPR, 2020." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 352, + 286, + 394 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 352, + 286, + 394 + ], + "spans": [ + { + "bbox": [ + 53, + 352, + 286, + 394 + ], + "type": "text", + "content": "[8] Mark Boss, Raphael Braun, Varun Jampani, Jonathan T. Barron, Ce Liu, and Hendrik P.A. Lensch. Nerd: Neural reflectance decomposition from image collections. In ICCV, 2021. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 397, + 286, + 440 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 397, + 286, + 440 + ], + "spans": [ + { + "bbox": [ + 53, + 397, + 286, + 440 + ], + "type": "text", + "content": "[9] Mark Boss, Varun Jampani, Raphael Braun, Ce Liu, Jonathan T. Barron, and Hendrik P.A. Lensch. Neural-pil: Neural pre-integrated lighting for reflectance decomposition. In NeurIPS, 2021. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 441, + 286, + 473 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 441, + 286, + 473 + ], + "spans": [ + { + "bbox": [ + 48, + 441, + 286, + 473 + ], + "type": "text", + "content": "[10] Brent Burley and Walt Disney Animation Studios. Physically-based shading at disney. In SIGGRAPH, 2012. 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 476, + 286, + 508 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 476, + 286, + 508 + ], + "spans": [ + { + "bbox": [ + 48, + 476, + 286, + 508 + ], + "type": "text", + "content": "[11] Bowen Cai, Jinchi Huang, Rongfei Jia, Chengfei Lv, and Huan Fu. Neuda: Neural deformable anchor for high-fidelity implicit surface reconstruction. In CVPR, 2023. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 510, + 286, + 540 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 510, + 286, + 540 + ], + "spans": [ + { + "bbox": [ + 48, + 510, + 286, + 540 + ], + "type": "text", + "content": "[12] Anpei Chen, Zexiang Xu, Andreas Geiger, Jingyi Yu, and Hao Su. Tensorf: Tensorial radiance fields. In ECCV, 2022. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 544, + 286, + 576 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 544, + 286, + 576 + ], + "spans": [ + { + "bbox": [ + 48, + 544, + 286, + 576 + ], + "type": "text", + "content": "[13] Blender Online Community. Blender - a 3D modelling and rendering package. Blender Foundation, Stichting Blender Foundation, Amsterdam, 2018. 8" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 578, + 286, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 578, + 286, + 610 + ], + "spans": [ + { + "bbox": [ + 48, + 578, + 286, + 610 + ], + "type": "text", + "content": "[14] Sara Fridovich-Keil, Alex Yu, Matthew Tancik, Qinhong Chen, Benjamin Recht, and Angjoo Kanazawa. Plenoxels: Radiance fields without neural networks. In CVPR, 2022. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 613, + 286, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 613, + 286, + 644 + ], + "spans": [ + { + "bbox": [ + 48, + 613, + 286, + 644 + ], + "type": "text", + "content": "[15] Mathieu Garon, Kalyan Sunkavalli, Sunil Hadap, Nathan Carr, and Jean-Francois Lalonde. Fast spatially-varying indoor lighting estimation. In CVPR, 2019. 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 48, + 647, + 286, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 647, + 286, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 647, + 286, + 689 + ], + "type": "text", + "content": "[16] Wenhang Ge, Tao Hu, Haoyu Zhao, Shu Liu, and Ying-Cong Chen. Ref-neus: Ambiguity-reduced neural implicit surface learning for multi-view reconstruction with reflection. In ICCV, 2023. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 48, + 692, + 286, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 692, + 286, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 692, + 286, + 712 + ], + "type": "text", + "content": "[17] Saeed Hadadan, Shuhong Chen, and Matthias Zwicker. Neural radiosivity. In ACM TOG, 2021. 2" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 35, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "text", + "content": "[18] Saeed Hadadan, Geng Lin, Jan Novák, Fabrice Rousselle, and Matthias Zwicker. Inverse global illumination using a neural radiometric prior. In SIGGRAPH Conference Proceedings, 2023. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 119, + 545, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 119, + 545, + 172 + ], + "spans": [ + { + "bbox": [ + 307, + 119, + 545, + 172 + ], + "type": "text", + "content": "[19] Bjoern Haefner, Simon Green, Alan Oursland, Daniel Andersen, Michael Goesele, Daniel Cremers, Richard Newcombe, and Thomas Whelan. Recovering real-world reflectance properties and shading from hdr imagery. In 3DV, 2021. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 175, + 545, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 175, + 545, + 217 + ], + "spans": [ + { + "bbox": [ + 307, + 175, + 545, + 217 + ], + "type": "text", + "content": "[20] Jon Hasselgren, Nikolai Hofmann, and Jacob Munkberg. Shape, light, and material decomposition from images using monte carlo rendering and denoising. In NeurIPS, 2022. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 220, + 545, + 242 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 220, + 545, + 242 + ], + "spans": [ + { + "bbox": [ + 307, + 220, + 545, + 242 + ], + "type": "text", + "content": "[21] IEC. IEC 61966-2-1:1999. Technical report, International Electrotechnical Commission, 1999. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 245, + 545, + 288 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 245, + 545, + 288 + ], + "spans": [ + { + "bbox": [ + 307, + 245, + 545, + 288 + ], + "type": "text", + "content": "[22] Wenzel Jakob, Sébastien Speierer, Nicolas Roussel, Merlin Nimier-David, Delio Vicini, Tizian Zeltner, Baptiste Nicolet, Miguel Crespo, Vincent Leroy, and Ziyi Zhang. Mitsuba 3 renderer, 2022. https://mitsuba-renderer.org. 8" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 291, + 545, + 322 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 291, + 545, + 322 + ], + "spans": [ + { + "bbox": [ + 307, + 291, + 545, + 322 + ], + "type": "text", + "content": "[23] Rasmus Jensen, Anders Dahl, George Vogiatzis, Engil Tola, and Henrik Aanæs. Large scale multi-view stereopsis evaluation. In CVPR, 2014. 2, 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 325, + 545, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 325, + 545, + 357 + ], + "spans": [ + { + "bbox": [ + 307, + 325, + 545, + 357 + ], + "type": "text", + "content": "[24] Chaonan Ji, Tao Yu, Kaiwen Guo, Jingxin Liu, and Yebin Liu. Geometry-aware single-image full-body human relighting. In ECCV, 2022. 1" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 360, + 545, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 360, + 545, + 392 + ], + "spans": [ + { + "bbox": [ + 307, + 360, + 545, + 392 + ], + "type": "text", + "content": "[25] Haian Jin, Isabella Liu, Peijia Xu, Xiaoshuai Zhang, Songfang Han, Sai Bi, Xiaowei Zhou, Zexiang Xu, and Hao Su. Tensoir: Tensorial inverse rendering. In CVPR, 2023. 2, 3, 6" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 395, + 545, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 395, + 545, + 415 + ], + "spans": [ + { + "bbox": [ + 307, + 395, + 545, + 415 + ], + "type": "text", + "content": "[26] James T. Kajiya. The rendering equation. In SIGGRAPH, 1986. 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 418, + 545, + 450 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 418, + 545, + 450 + ], + "spans": [ + { + "bbox": [ + 307, + 418, + 545, + 450 + ], + "type": "text", + "content": "[27] Petr Kellnhofer, Lars C Jebe, Andrew Jones, Ryan Spicer, Kari Pulli, and Gordon Wetzstein. Neural lumigraph rendering. In CVPR, 2021. 1" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 453, + 545, + 495 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 453, + 545, + 495 + ], + "spans": [ + { + "bbox": [ + 307, + 453, + 545, + 495 + ], + "type": "text", + "content": "[28] Zhengfei Kuang, Fujun Luan, Sai Bi, Zhixin Shu, Gordon Wetzstein, and Kalyan Sunkavalli. Palettenerf: Palette-based appearance editing of neural radiance fields. In CVPR, 2023. 6" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 498, + 545, + 541 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 498, + 545, + 541 + ], + "spans": [ + { + "bbox": [ + 307, + 498, + 545, + 541 + ], + "type": "text", + "content": "[29] Zhengqin Li, Mohammad Shafiei, Ravi Ramamoorthi, Kalyan Sunkavalli, and Manmohan Chandraker. Inverse rendering for complex indoor scenes: Shape, spatially-varying lighting and svbrdf from a single image. In CVPR, 2020. 1" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 307, + 544, + 545, + 587 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 544, + 545, + 587 + ], + "spans": [ + { + "bbox": [ + 307, + 544, + 545, + 587 + ], + "type": "text", + "content": "[30] Zhengqin Li, Jia Shi, Sai Bi, Rui Zhu, Kalyan Sunkavalli, Miloš Hašan, Zexiang Xu, Ravi Ramamoorthi, and Manmohan Chandraker. Physically-based editing of indoor scene lighting from a single image. In ECCV, 2022. 1" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 307, + 590, + 545, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 590, + 545, + 632 + ], + "spans": [ + { + "bbox": [ + 307, + 590, + 545, + 632 + ], + "type": "text", + "content": "[31] Zhaoshuo Li, Thomas Müller, Alex Evans, Russell H Taylor, Mathias Unberath, Ming-Yu Liu, and Chen-Hsuan Lin. Neuralangelo: High-fidelity neural surface reconstruction. In CVPR, 2023. 2" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 307, + 635, + 545, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 635, + 545, + 666 + ], + "spans": [ + { + "bbox": [ + 307, + 635, + 545, + 666 + ], + "type": "text", + "content": "[32] Zhen Li, Lingli Wang, Mofang Cheng, Cihui Pan, and Jiaqi. Yang. Multi-view inverse rendering for large-scale real-world indoor scenes. In CVPR, 2023. 2" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 307, + 670, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 670, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 307, + 670, + 545, + 712 + ], + "type": "text", + "content": "[33] Ruofan Liang, Huiting Chen, Chunlin Li, Fan Chen, Selvakumar Panneer, and Nandita Vijaykumar. Envidr: Implicit differentiable renderer with neural environment lighting. In ICCV, 2023. 1" + } + ] + } + ], + "index": 34 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "4606" + } + ] + } + ], + "index": 36 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "type": "text", + "content": "[34] Fujun Luan, Shuang Zhao, Kavita Bala, and Zhao Dong. Unified shape and svbrdf recovery using differentiable monte carlo rendering. In EGSR, 2021. 1" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 106, + 287, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 106, + 287, + 149 + ], + "spans": [ + { + "bbox": [ + 48, + 106, + 287, + 149 + ], + "type": "text", + "content": "[35] Linjie Lyu, Ayush Tewari, Thomas Leimkuehler, Marc Habermann, and Christian Theobalt. Neural radiance transfer fields for relightable novel-view synthesis with global illumination. In ECCV, 2022. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 150, + 287, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 150, + 287, + 194 + ], + "spans": [ + { + "bbox": [ + 48, + 150, + 287, + 194 + ], + "type": "text", + "content": "[36] Linjie Lyu, Ayush Tewari, Marc Habermann, Shunsuke Saito, Michael Zollhöfer, Thomas Leimkuehler, and Christian Theobalt. Diffusion posterior illumination for ambiguity-aware inverse rendering. In ACM TOG, 2023. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 194, + 287, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 194, + 287, + 227 + ], + "spans": [ + { + "bbox": [ + 48, + 194, + 287, + 227 + ], + "type": "text", + "content": "[37] Alexander Mai, Dor Verbin, Falko Kuester, and Sara Fridovich-Keil. Neural microfacet fields for inverse rendering. In ICCV, 2023. 1, 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 228, + 287, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 228, + 287, + 271 + ], + "spans": [ + { + "bbox": [ + 48, + 228, + 287, + 271 + ], + "type": "text", + "content": "[38] Ricardo Martin-Brualla, Noha Radwan, Mehdi S. M. Sajjadi, Jonathan T. Barron, Alexey Dosovitskiy, and Daniel Duckworth. Nerf in the wild: Neural radiance fields for unconstrained photo collections. In CVPR, 2021. 1, 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 271, + 287, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 271, + 287, + 304 + ], + "spans": [ + { + "bbox": [ + 48, + 271, + 287, + 304 + ], + "type": "text", + "content": "[39] Nelson Max. Optical models for direct volume rendering. IEEE Transactions on Visualization and Computer Graphics, 1(2):99-108, 1995. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 304, + 287, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 304, + 287, + 348 + ], + "spans": [ + { + "bbox": [ + 48, + 304, + 287, + 348 + ], + "type": "text", + "content": "[40] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In ECCV, 2020. 1, 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 349, + 287, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 349, + 287, + 381 + ], + "spans": [ + { + "bbox": [ + 48, + 349, + 287, + 381 + ], + "type": "text", + "content": "[41] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. In ACM TOG, 2022. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 382, + 287, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 382, + 287, + 426 + ], + "spans": [ + { + "bbox": [ + 48, + 382, + 287, + 426 + ], + "type": "text", + "content": "[42] Jacob Munkberg, Jon Hasselgren, Tianchang Shen, Jun Gao, Wenzheng Chen, Alex Evans, Thomas Müller, and Sanja Fidler. Extracting triangular 3d models, materials, and lighting from images. In CVPR, 2022. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 426, + 287, + 469 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 426, + 287, + 469 + ], + "spans": [ + { + "bbox": [ + 48, + 426, + 287, + 469 + ], + "type": "text", + "content": "[43] Michael Niemeyer, Lars Mescheder, Michael Oechsle, and Andreas Geiger. Differentiable volumetric rendering: Learning implicit 3d representations without 3d supervision. In CVPR, 2020. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 470, + 287, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 470, + 287, + 503 + ], + "spans": [ + { + "bbox": [ + 48, + 470, + 287, + 503 + ], + "type": "text", + "content": "[44] Michael Oechsle, Songyou Peng, and Andreas Geiger. Unisurf: Unifying neural implicit surfaces and radiance fields for multi-view reconstruction. In ICCV, 2021. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 503, + 287, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 503, + 287, + 525 + ], + "spans": [ + { + "bbox": [ + 48, + 503, + 287, + 525 + ], + "type": "text", + "content": "[45] Julian Ost, Issam Laradji, Alejandro Newell, Yuval Bahat, and Felix Heide. Neural point light fields. In CVPR, 2022. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 525, + 287, + 579 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 525, + 287, + 579 + ], + "spans": [ + { + "bbox": [ + 48, + 525, + 287, + 579 + ], + "type": "text", + "content": "[46] Rohit Pandey, Sergio Orts-Escalano, Chloe LeGendre, Christian Haene, Sofien Bouaziz, Christoph Rhemann, Paul Debevec, and Seann Fanello. Total relighting: Learning to relight portraits for background replacement. In ACM TOG, 2021. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 580, + 287, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 580, + 287, + 624 + ], + "spans": [ + { + "bbox": [ + 48, + 580, + 287, + 624 + ], + "type": "text", + "content": "[47] Jeong Joon Park, Peter Florence, Julian Straub, Richard Newcombe, and Steven Lovegrove. Deepsdf: Learning continuous signed distance functions for shape representation. In CVPR, 2019. 1" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 624, + 287, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 624, + 287, + 657 + ], + "spans": [ + { + "bbox": [ + 48, + 624, + 287, + 657 + ], + "type": "text", + "content": "[48] Julien Philip, Sébastien Morgenthaler, Michael Gharbi, and George Drettakis. Free-viewpoint indoor neural relighting from multi-view stereo. In ACM TOG, 2021. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 658, + 287, + 680 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 287, + 680 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 287, + 680 + ], + "type": "text", + "content": "[49] Martin Piala and Ronald Clark. Terminerf: Ray termination prediction for efficient neural rendering. In 3DV, 2021. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "type": "text", + "content": "[50] Viktor Rudnev, Mohamed Elgharib, William Smith, Lingjie Liu, Vladislav Golyanik, and Christian Theobalt. Nerf for outdoor scene relighting. In ECCV, 2022. 1" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 72, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 33, + "blocks": [ + { + "bbox": [ + 307, + 72, + 545, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 72, + 545, + 116 + ], + "spans": [ + { + "bbox": [ + 307, + 72, + 545, + 116 + ], + "type": "text", + "content": "[51] Vincent Sitzmann, Julien N.P. Martel, Alexander W. Bergman, David B. Lindell, and Gordon Wetzstein. Implicit neural representations with periodic activation functions. In NeurIPS, 2020. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 118, + 545, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 545, + 161 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 545, + 161 + ], + "type": "text", + "content": "[52] Vincent Sitzmann, Semon Rezchikov, Bill Freeman, Josh Tenenbaum, and Fredo Durand. Light field networks: Neural scene representations with single-evaluation rendering. In NeurIPS, 2021. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 162, + 545, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 162, + 545, + 184 + ], + "spans": [ + { + "bbox": [ + 307, + 162, + 545, + 184 + ], + "type": "text", + "content": "[53] Alvy Ray Smith. Color gamut transform pairs. In ACM TOG, 1978. 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 185, + 545, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 185, + 545, + 229 + ], + "spans": [ + { + "bbox": [ + 307, + 185, + 545, + 229 + ], + "type": "text", + "content": "[54] Pratul P. Srinivasan, Boyang Deng, Xiuming Zhang, Matthew Tancik, Ben Mildenhall, and Jonathan T. Barron. Nerv: Neural reflectance and visibility fields for relighting and view synthesis. In CVPR, 2021. 2, 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 230, + 545, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 230, + 545, + 262 + ], + "spans": [ + { + "bbox": [ + 307, + 230, + 545, + 262 + ], + "type": "text", + "content": "[55] Mohammed Suhail, Carlos Esteves, Leonid Sigal, and Ameesh Makadia. Light field neural rendering. In CVPR, 2022. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 264, + 545, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 264, + 545, + 297 + ], + "spans": [ + { + "bbox": [ + 307, + 264, + 545, + 297 + ], + "type": "text", + "content": "[56] Cheng Sun, Min Sun, and Hwann-Tzong Chen. Direct voxel grid optimization: Super-fast convergence for radiance fields reconstruction. In CVPR, 2022. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 298, + 545, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 298, + 545, + 342 + ], + "spans": [ + { + "bbox": [ + 307, + 298, + 545, + 342 + ], + "type": "text", + "content": "[57] Cheng Sun, Guangyan Cai, Zhengqin Li, Kai Yan, Cheng Zhang, Carl Marshall, Jia-Bin Huang, Shuang Zhao, and Zhao Dong. Neural-pbir reconstruction of shape, material, and illumination. In ICCV, 2023. 2, 3, 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 343, + 545, + 388 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 343, + 545, + 388 + ], + "spans": [ + { + "bbox": [ + 307, + 343, + 545, + 388 + ], + "type": "text", + "content": "[58] Jiaming Sun, Xi Chen, Qianqian Wang, Zhengqi Li, Hadar Averbuch-Elor, Xiaowei Zhou, and Noah Snively. Neural 3D reconstruction in the wild. In SIGGRAPH Conference Proceedings, 2022. 1" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 388, + 545, + 422 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 388, + 545, + 422 + ], + "spans": [ + { + "bbox": [ + 307, + 388, + 545, + 422 + ], + "type": "text", + "content": "[59] Tiancheng Sun, Kai-En Lin, Sai Bi, Zexiang Xu, and Ravi Ramamoorthi. Nelf: Neural light-transport field for portrait view synthesis and relighting. In EGSR, 2021. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 422, + 545, + 477 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 422, + 545, + 477 + ], + "spans": [ + { + "bbox": [ + 307, + 422, + 545, + 477 + ], + "type": "text", + "content": "[60] Matthew Tancik, Pratul P. Srinivasan, Ben Mildenhall, Sara Fridovich-Keil, Nithin Raghavan, Utkarsh Singhal, Ravi Ramamoorthi, Jonathan T. Barron, and Ren Ng. Fourier features let networks learn high frequency functions in low dimensional domains. In NeurIPS, 2020. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 478, + 545, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 478, + 545, + 521 + ], + "spans": [ + { + "bbox": [ + 307, + 478, + 545, + 521 + ], + "type": "text", + "content": "[61] Jiaxiang Tang, Hang Zhou, Xiaokang Chen, Tianshu Hu, Errui Ding, Jingdong Wang, and Gang Zeng. Delicate textured mesh recovery from nef via adaptive surface refinement. In ICCV, 2023. 1" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 523, + 545, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 523, + 545, + 578 + ], + "spans": [ + { + "bbox": [ + 307, + 523, + 545, + 578 + ], + "type": "text", + "content": "[62] Ayush Tewari, Ohad Fried, Justus Thies, Vincent Sitzmann, Stephen Lombardi, Kalyan Sunkavalli, Ricardo Martin-Brualla, Tomas Simon, Jason Saragih, Matthias Nießner, et al. State of the art on neural rendering. In CGF, 2020. 1" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 578, + 545, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 578, + 545, + 624 + ], + "spans": [ + { + "bbox": [ + 307, + 578, + 545, + 624 + ], + "type": "text", + "content": "[63] Ayush Tewari, Justus Thies, Ben Mildenhall, Pratul Srinivasan, Edgar Tretschk, Wang Yifan, Christoph Lassner, Vincent Sitzmann, Ricardo Martin-Brualla, Stephen Lombardi, et al. Advances in neural rendering. In CGF, 2022. 1" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 307, + 624, + 545, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 624, + 545, + 668 + ], + "spans": [ + { + "bbox": [ + 307, + 624, + 545, + 668 + ], + "type": "text", + "content": "[64] Marco Toschi, Riccardo De Matteo, Riccardo Spezialetti, Daniele De Gregorio, Luigi Di Stefano, and Samuele Salti. Relight my nerf: A dataset for novel view synthesis and relighting of real world objects. In CVPR, 2023. 2" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 307, + 669, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 669, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 307, + 669, + 545, + 712 + ], + "type": "text", + "content": "[65] Dor Verbin, Peter Hedman, Ben Mildenhall, Todd Zickler, Jonathan T Barron, and Pratul P Srinivasan. Ref-nerf: Structured view-dependent appearance for neural radiance fields. In CVPR, 2022. 1" + } + ] + } + ], + "index": 32 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4607" + } + ] + } + ], + "index": 34 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "type": "text", + "content": "[66] Dongqing Wang, Tong Zhang, and Sabine Susstrunk. Nemto: Neural environment matting for novel view and relighting synthesis of transparent objects. In ICCV, 2023. 1" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 107, + 287, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 107, + 287, + 140 + ], + "spans": [ + { + "bbox": [ + 49, + 107, + 287, + 140 + ], + "type": "text", + "content": "[67] Jiaping Wang, Peiran Ren, Minmin Gong, John Snyder, and Baining Guo. All-frequency rendering of dynamic, spatially-varying reflectance. ACM TOG, 2009. 4" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 141, + 287, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 141, + 287, + 184 + ], + "spans": [ + { + "bbox": [ + 49, + 141, + 287, + 184 + ], + "type": "text", + "content": "[68] Peng Wang, Lingjie Liu, Yuan Liu, Christian Theobalt, Taku Komura, and Wenping Wang. Neus: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. In NeurIPS, 2021. 2, 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 186, + 287, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 186, + 287, + 217 + ], + "spans": [ + { + "bbox": [ + 49, + 186, + 287, + 217 + ], + "type": "text", + "content": "[69] Shaofei Wang, Katja Schwarz, Andreas Geiger, and Siyu Tang. Arah: Animatable volume rendering of articulated human sdfs. In ECCV, 2022. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 49, + 220, + 287, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 220, + 287, + 262 + ], + "spans": [ + { + "bbox": [ + 49, + 220, + 287, + 262 + ], + "type": "text", + "content": "[70] Yiming Wang, Qin Han, Marc Habermann, Kostas Dani-ilidis, Christian Theobalt, and Lingjie Liu. Neus2: Fast learning of neural implicit surfaces for multi-view reconstruction. In ICCV, 2023. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 49, + 264, + 287, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 264, + 287, + 297 + ], + "spans": [ + { + "bbox": [ + 49, + 264, + 287, + 297 + ], + "type": "text", + "content": "[71] Zian Wang, Jonah Philion, Sanja Fidler, and Jan Kautz. Learning indoor inverse rendering with 3d spatially-varying lighting. In ICCV, 2021. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 49, + 299, + 287, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 299, + 287, + 342 + ], + "spans": [ + { + "bbox": [ + 49, + 299, + 287, + 342 + ], + "type": "text", + "content": "[72] Zian Wang, Wenzheng Chen, David Acuna, Jan Kautz, and Sanja Fidler. Neural light field estimation for street scenes with differentiable virtual object insertion. In ECCV, 2022. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 49, + 343, + 287, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 343, + 287, + 397 + ], + "spans": [ + { + "bbox": [ + 49, + 343, + 287, + 397 + ], + "type": "text", + "content": "[73] Zian Wang, Tianchang Shen, Jun Gao, Shengyu Huang, Jacob Munkberg, Jon Hasselgren, Zan Gojcic, Wenzheng Chen, and Sanja Fidler. Neural fields meet explicit geometric representations for inverse rendering of urban scenes. In CVPR, 2023. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 49, + 399, + 287, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 399, + 287, + 430 + ], + "spans": [ + { + "bbox": [ + 49, + 399, + 287, + 430 + ], + "type": "text", + "content": "[74] Henrique Weber, Mathieu Garon, and Jean-François Lalonde. Editable indoor lighting estimation. In ECCV, 2022. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 49, + 433, + 287, + 476 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 433, + 287, + 476 + ], + "spans": [ + { + "bbox": [ + 49, + 433, + 287, + 476 + ], + "type": "text", + "content": "[75] Haoqian Wu, Zhipeng Hu, Lincheng Li, Yongqiang Zhang, Changjie Fan, and Xin Yu. Nefii: Inverse rendering for reflectance decomposition with near-field indirect illumination. In CVPR, 2023. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 49, + 478, + 287, + 532 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 478, + 287, + 532 + ], + "spans": [ + { + "bbox": [ + 49, + 478, + 287, + 532 + ], + "type": "text", + "content": "[76] Liwen Wu, Rui Zhu, Mustafa B. Yaldiz, Yinhao Zhu, Hong Cai, Janarbek Matai, Fatih Porikli, Tzu-Mao Li, Manmohan Chandraker, and Ravi Ramamoorthi. Factorized inverse path tracing for efficient and accurate material-lighting estimation. In ICCV, 2023. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 49, + 534, + 287, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 534, + 287, + 566 + ], + "spans": [ + { + "bbox": [ + 49, + 534, + 287, + 566 + ], + "type": "text", + "content": "[77] Qianyi Wu, Xian Liu, Yuedong Chen, Kejie Li, Chuanxia Zheng, Jianfei Cai, and Jianmin Zheng. Object-compositional neural implicit surfaces. In ECCV, 2022. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 49, + 567, + 287, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 567, + 287, + 600 + ], + "spans": [ + { + "bbox": [ + 49, + 567, + 287, + 600 + ], + "type": "text", + "content": "[78] Qianyi Wu, Kaisiyuan Wang, Kejie Li, Jianmin Zheng, and Jianfei Cai. Objectsdf++: Improved object-compositional neural implicit surfaces. In ICCV, 2023. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 49, + 601, + 287, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 601, + 287, + 645 + ], + "spans": [ + { + "bbox": [ + 49, + 601, + 287, + 645 + ], + "type": "text", + "content": "[79] Tong Wu, Jia-Mu Sun, Yu-Kun Lai, and Lin Gao. De-nerf: Decoupled neural radiance fields for view-consistent appearance editing and high-frequency environmental relighting. In SIGGRAPH ASIA Conference Proceedings, 2023. 1" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 49, + 647, + 287, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 647, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 49, + 647, + 287, + 689 + ], + "type": "text", + "content": "[80] Tong Wu, Jiaqi Wang, Xingang Pan, Xudong Xu, Christian Theobalt, Ziwei Liu, and Dahua Lin. Voxurf: Voxel-based efficient and accurate neural surface reconstruction. In ICLR, 2023. 2, 5, 6" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 49, + 691, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 691, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 49, + 691, + 287, + 713 + ], + "type": "text", + "content": "[81] Yuanbo Xiangli, Linning Xu, Xingang Pan, Nanxuan Zhao, Anyi Rao, Christian Theobalt, Bo Dai, and Dahua Lin." + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 33, + "blocks": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "type": "text", + "content": "Bungeenerf: Progressive neural radiance field for extreme multi-scale scene rendering. In ECCV, 2022. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 95, + 545, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 95, + 545, + 138 + ], + "spans": [ + { + "bbox": [ + 308, + 95, + 545, + 138 + ], + "type": "text", + "content": "[82] Yingyan Xu, Gaspard Zoss, Prashanth Chandran, Markus Gross, Derek Bradley, and Paulo Gotardo. Renerf: Relightable neural radiance fields with nearfield lighting. In ICCV, 2023. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 140, + 545, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 140, + 545, + 182 + ], + "spans": [ + { + "bbox": [ + 308, + 140, + 545, + 182 + ], + "type": "text", + "content": "[83] Wenqi Yang, Guanying Chen, Chaofeng Chen, Zhenfang Chen, and Kwan-Yee K. Wong. Ps-nerf: Neural inverse rendering for multi-view photometric stereo. In ECCV, 2022. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 184, + 545, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 184, + 545, + 227 + ], + "spans": [ + { + "bbox": [ + 308, + 184, + 545, + 227 + ], + "type": "text", + "content": "[84] Wenqi Yang, Guanying Chen, Chaofeng Chen, Zhenfang Chen, and Kwan-Yee K. Wong. S" + }, + { + "bbox": [ + 308, + 184, + 545, + 227 + ], + "type": "inline_equation", + "content": "^3" + }, + { + "bbox": [ + 308, + 184, + 545, + 227 + ], + "type": "text", + "content": "-nerf: Neural reflectance field from shading and shadow under a single viewpoint. In NeurIPS, 2022. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 228, + 545, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 228, + 545, + 270 + ], + "spans": [ + { + "bbox": [ + 308, + 228, + 545, + 270 + ], + "type": "text", + "content": "[85] Yao Yao, Jingyang Zhang, Jingbo Liu, Yihang Qu, Tian Fang, David McKinnon, Yanghai Tsin, and Long Quan. Neilf: Neural incident light field for material and lighting estimation. In ECCV, 2022. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 272, + 545, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 272, + 545, + 316 + ], + "spans": [ + { + "bbox": [ + 308, + 272, + 545, + 316 + ], + "type": "text", + "content": "[86] Lior Yariv, Yoni Kasten, Dror Moran, Meirav Galun, Matan Atzmon, Basri Ronen, and Yaron Lipman. Multiview neural surface reconstruction by disentangling geometry and appearance. In NeurIPS, 2020. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 316, + 545, + 347 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 316, + 545, + 347 + ], + "spans": [ + { + "bbox": [ + 308, + 316, + 545, + 347 + ], + "type": "text", + "content": "[87] Lior Yariv, Jiatao Gu, Yoni Kasten, and Yaron Lipman. Volume rendering of neural implicit surfaces. In NeurIPS, 2021. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 349, + 545, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 349, + 545, + 381 + ], + "spans": [ + { + "bbox": [ + 308, + 349, + 545, + 381 + ], + "type": "text", + "content": "[88] Alex Yu, Ruilong Li, Matthew Tancik, Hao Li, Ren Ng, and Angjoo Kanazawa. PlenOctrees for real-time rendering of neural radiance fields. In ICCV, 2021. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 383, + 545, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 383, + 545, + 426 + ], + "spans": [ + { + "bbox": [ + 308, + 383, + 545, + 426 + ], + "type": "text", + "content": "[89] Chong Zeng, Guojun Chen, Yue Dong, Pieter Peers, Hongzhi Wu, and Xin Tong. Relighting neural radiance fields with shadow and highlight hints. In SIGGRAPH Conference Proceedings, 2023. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 426, + 545, + 469 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 426, + 545, + 469 + ], + "spans": [ + { + "bbox": [ + 308, + 426, + 545, + 469 + ], + "type": "text", + "content": "[90] Jingyang Zhang, Yao Yao, Shiwei Li, Jingbo Liu, Tian Fang, David McKinnon, Yanghai Tsin, and Long Quan. Neilf++: Inter-reflectable light fields for geometry and material estimation. In ICCV, 2023. 2, 6" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 471, + 545, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 471, + 545, + 503 + ], + "spans": [ + { + "bbox": [ + 308, + 471, + 545, + 503 + ], + "type": "text", + "content": "[91] Kai Zhang, Gernot Riegler, Noah Snavely, and Vladlen Koltun. Nerf++: Analyzing and improving neural radiance fields. arXiv preprint arXiv:2010.07492, 2020. 1" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 504, + 545, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 504, + 545, + 547 + ], + "spans": [ + { + "bbox": [ + 308, + 504, + 545, + 547 + ], + "type": "text", + "content": "[92] Kai Zhang, Fujun Luan, Qianqian Wang, Kavita Bala, and Noah Snavely. Physg: Inverse rendering with spherical gaussians for physics-based material editing and relighting. In CVPR, 2021. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 548, + 545, + 581 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 548, + 545, + 581 + ], + "spans": [ + { + "bbox": [ + 308, + 548, + 545, + 581 + ], + "type": "text", + "content": "[93] Kai Zhang, Fujun Luan, Zhengqi Li, and Noah Snavely. Iron: Inverse rendering by optimizing neural sdfs and materials from photometric images. In CVPR, 2022. 1" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 308, + 582, + 545, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 582, + 545, + 635 + ], + "spans": [ + { + "bbox": [ + 308, + 582, + 545, + 635 + ], + "type": "text", + "content": "[94] Xiuming Zhang, Sean Fanello, Yun-Ta Tsai, Tiancheng Sun, Tianfan Xue, Rohit Pandey, Sergio Orts-Escalano, Philip Davidson, Christoph Rhemann, Paul Debevec, et al. Neural light transport for relighting and view synthesis. In ACM TOG, 2021. 2" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 308, + 636, + 545, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 636, + 545, + 679 + ], + "spans": [ + { + "bbox": [ + 308, + 636, + 545, + 679 + ], + "type": "text", + "content": "[95] Xiuming Zhang, Pratul P. Srinivasan, Boyang Deng, Paul Debevec, William T. Freeman, and Jonathan T. Barron. Ner-factor: Neural factorization of shape and reflectance under an unknown illumination. In ACM TOG, 2021. 2" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 308, + 681, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 681, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 681, + 545, + 713 + ], + "type": "text", + "content": "[96] Yuanqing Zhang, Jiaming Sun, Xingyi He, Huan Fu, Rongfei Jia, and Xiaowei Zhou. Modeling indirect illumination for inverse rendering. In CVPR, 2022. 2" + } + ] + } + ], + "index": 32 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "4608" + } + ] + } + ], + "index": 34 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 196 + ], + "type": "list", + "angle": 0, + "index": 3, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "type": "text", + "content": "[97] Y. Zhang, Z. Hu, H. Wu, M. Zhao, L. Li, Z. Zou, and C. Fan. Towards unbiased volume rendering of neural implicit surfaces with geometry priors. In CVPR, 2023. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 107, + 286, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 107, + 286, + 138 + ], + "spans": [ + { + "bbox": [ + 48, + 107, + 286, + 138 + ], + "type": "text", + "content": "[98] Quan Zheng, Gurprit Singh, and Hans-Peter Seidel. Neural relightable participating media rendering. In NeurIPS, 2021. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 140, + 286, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 140, + 286, + 196 + ], + "spans": [ + { + "bbox": [ + 48, + 140, + 286, + 196 + ], + "type": "text", + "content": "[99] Jingsen Zhu, Fujun Luan, Yuchi Huo, Zihao Lin, Zhihua Zhong, Dianbing Xi, Rui Wang, Hujun Bao, Jiaxiang Zheng, and Rui Tang. Learning-based inverse rendering of complex indoor scenes with differentiable monte carlo raytracing. In SIGGRAPH ASIA Conference Proceedings, 2022. 2" + } + ] + } + ], + "index": 2 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 316, + 757 + ], + "type": "text", + "content": "4609" + } + ] + } + ], + "index": 4 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/EVCap_ Retrieval-Augmented Image Captioning with External Visual-Name Memory for Open-World Comprehension/003cc44f-e449-40f1-aec0-baf1c6235010_content_list.json b/2024/EVCap_ Retrieval-Augmented Image Captioning with External Visual-Name Memory for Open-World Comprehension/003cc44f-e449-40f1-aec0-baf1c6235010_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..d323e28a00f5d3fa6e4c03f5d75f7bf92e7e7bc7 --- /dev/null +++ b/2024/EVCap_ Retrieval-Augmented Image Captioning with External Visual-Name Memory for Open-World Comprehension/003cc44f-e449-40f1-aec0-baf1c6235010_content_list.json @@ -0,0 +1,1864 @@ +[ + { + "type": "text", + "text": "EVCAP: Retrieval-Augmented Image Captioning with External Visual-Name Memory for Open-World Comprehension", + "text_level": 1, + "bbox": [ + 133, + 130, + 836, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Jiaxuan $\\mathrm{Li}^{1*}$ , Duc Minh $\\mathrm{Vo}^{1*}$ , Akihiro Sugimoto $^{2}$ , Hideki Nakayama $^{1}$", + "bbox": [ + 197, + 202, + 776, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1The University of Tokyo, Japan 2National Institute of Informatics, Japan", + "bbox": [ + 194, + 220, + 774, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{li, vm Duc}@nlab.ci.i.u-tokyo.ac.jp sugimoto@nii.ac.jp nakayama@ci.i.u-tokyo.ac.jp", + "bbox": [ + 119, + 241, + 849, + 257 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 291, + 313, + 306 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Large language models (LLMs)-based image captioning has the capability of describing objects not explicitly observed in training data; yet novel objects occur frequently, necessitating the requirement of sustaining up-to-date object knowledge for open-world comprehension. Instead of relying on large amounts of data and/or scaling up network parameters, we introduce a highly effective retrieval-augmented image captioning method that prompts LLMs with object names retrieved from External Visual-name memory (EVCAP). We build ever-changing object knowledge memory using objects' visuals and names, enabling us to (i) update the memory at a minimal cost and (ii) effortlessly augment LLMs with retrieved object names by utilizing a lightweight and fast-to-train model. Our model, which was trained only on the COCO dataset, can adapt to out-of-domain without requiring additional fine-tuning or re-training. Our experiments conducted on benchmarks and synthetic commonsense-violating data show that EVCAP, with only 3.97M trainable parameters, exhibits superior performance compared to other methods based on frozen pre-trained LLMs. Its performance is also competitive to specialist SOTAs that require extensive training.", + "bbox": [ + 75, + 323, + 473, + 654 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 683, + 209, + 699 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Advanced image captioning based on large language models (LLMs) [3, 8, 9, 25] has focused on the approach using big-scale models trained on ever-increasingly large-scale datasets, which is no longer viable. This is because the computational cost to train the models increases exponentially and, more importantly, updating training data is almost impossible to keep pace with the growth of novel objects in our daily lives. Sustaining ever-changing object knowledge with a reasonable cost is a pressing concern in LLMs-based models to truly unlock open-world comprehension.", + "bbox": [ + 75, + 709, + 468, + 859 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Retrieval-augmented image captioning [20, 35] is", + "bbox": [ + 96, + 859, + 468, + 876 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/71d2fc87b5ba24b951155fde0165bac7dc20438920282706b7fa1b5866e3486c.jpg", + "image_caption": [ + "Figure 1. Overall comparison of our EVCAP and SOTAs. (Upper) Generated captions by SmallCap, BLIP-2, and our EVCAP for a commonsense-violating image from the WHOOPS dataset. $\\times$ and $\\checkmark$ indicate incorrect and correct predictions, respectively. Incorrect objects in captions are highlighted in red, while correct ones are in blue. SmallCap and BLIP-2 give incorrect predictions for \"ice skates\" and \"wood floor\", respectively, while our EVCAP utilizes an external visual-name memory to enhance attention to objects within the image, leading to superior performance for image captioning. (Lower) Comparison of the number of trainable parameters, CIDEr score on COCO and NoCaps datasets. The size of each circle reflects the log number of trainable parameters. EVCAP (3.97M) has less trainable parameters than others while achieving comparable results with SOTAs at scale." + ], + "image_footnote": [], + "bbox": [ + 504, + 291, + 890, + 542 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "emerging as an alternative since it considerably reduces training costs in both time and data while producing encouraging results. Nonetheless, with their huge datastore, it is obvious that LLMs would imitate the given texts, limiting their ability to describe open-world objects properly. For instance, SmallCap [35] considers the words \"skateboard\" and \"wooden floor\" to be a pair regardless of visual appearances containing a commonsense-violating pair", + "bbox": [ + 496, + 780, + 893, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Equal contributions. Code is available at https://jiaxuan-li.github.io/EVCap.", + "bbox": [ + 99, + 885, + 473, + 896 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "13733", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "of \"ice skates\" and \"wood floor\" (Fig. 1, upper). Additionally, prompting the LLMs given a lot of retrieved texts becomes cumbersome, requiring more trainable parameters. Fig. 1 (lower) shows that the CIDEr scores obtained by a lightweight SmallCap [35] with 43M trainable parameters are far away from those obtained by a heavy REVEAL [20] with 2.1B trainable parameters. Beyond that, due to the frequent occurrence of new objects, access to their sample texts is not always feasible, making the memory utilized in [20, 35] difficult to grow. We thus aim to streamline the external memory used in previous work [20, 35] by storing a sufficiently small amount of object information. And, of course, not only does the model not stereotype the example sentences, but the number of trainable parameters would be reduced drastically as a result of the causation (Fig. 1).", + "bbox": [ + 75, + 90, + 472, + 316 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We follow [13, 40] to construct a key-value memory where the key is represented by object's features, and the value corresponds to object's name. Unlike [13, 40], which rely on object definition as the key, our method leverages the visual appearance of the object as the key because of the abundance of object images readily available on the internet. We propose an external visual-name memory tailored for ease of expansion and cost-effectiveness in upholding up-to-date object information. We present a highly effective retrieval-augmented LLMs-based image captioning method, called EVCAP, that prompts frozen LLMs with object names retrieved from our proposed memory for open-world comprehension. EVCAP contains a frozen image encoder ViT [14] and Q-Former [25] with trainable image query tokens for object retrieval, an attentive fusion module, a trainable linear layer for mapping between vision and language latent spaces, and a frozen LLM decoder [10] for generating captions. Specifically, the attentive fusion module feeds retrieved object names and visual features into a customized frozen Q-Former using trainable object name query tokens to implicitly reduce the presence of superfluous object names. As a result, EVCAP amounts to only 3.97M trainable parameters. Once trained, the model can be adapted to new domains and large-scale data without further fine-tuning or re-training. Our contributions are as follows:", + "bbox": [ + 75, + 316, + 472, + 691 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We provide an extensible external visual-name memory with minimal but useful object information, which enables LLMs-based models to comprehend the open world.", + "- We present a highly efficacious retrieval-augmented image captioning EVCAP with 3.97M trainable parameters." + ], + "bbox": [ + 76, + 693, + 468, + 768 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "On in-/out-domain benchmarks and synthetic commonsense-violating dataset, EVCAP trained solely on COCO dataset competes with other lightweight methods by a margin while being on par with other specialist SOTAs.", + "bbox": [ + 75, + 771, + 468, + 832 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 76, + 845, + 218, + 861 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Image captioning aims to describe the contents of a given image. It can be roughly divided into two approaches: non-", + "bbox": [ + 75, + 869, + 468, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "LLMs-based methods and LLMs-based ones. The former approaches [4, 22, 42] typically employ a visual encoder and a language decoder in an end-to-end fashion to generate captions. However, they are incapable of describing open-world objects. The latter one leverages pre-trained large-scale vision models (CLIP [32], ViT [12]) and LLMs (GPTs [7, 31], T5 [33], LLaMA [37]) by bridging the gap between two modalities using either pre-training with largescale data or the learned mapper or prompt techniques. LLMs-based models [8, 9, 25, 29] demonstrate advancements in image captioning challenges, allowing the capacity to describe anything as long as pre-trained vision models can recognize it. Our method belongs to the LLMs-based approaches, but instead of relying fully on the pre-trained vision model, we use object names retrieved from the external memory to augment LLMs-based image captioning.", + "bbox": [ + 496, + 90, + 890, + 330 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Novel object captioning is a branch of image captioning that describes images containing objects that were not seen during training. Non-LLMs-based methods explore more objects by learning from unpaired image-sentence sources (DCC [19], NOC [39]) or rely on novel object detectors to recognize novel concepts (NBT [28], OSCAR [26] and VinVL [45]). LLMs-based methods such as ViECap [15] leverage the pre-trained CLIP [32] to obtain object entities. Nevertheless, the cut-off in training time of the pre-trained object detector or CLIP prevents it from detecting novel objects that arise quickly in reality. Unlike earlier work, we can readily update our recognition of novel concepts by adding them to external memory, ensuring that we keep any new objects from the past and even the future.", + "bbox": [ + 496, + 330, + 892, + 541 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Retrieval-augmented image captioning is a recently popular approach that augments the captioning model with retrieved information for better open-world understanding. AoANet [16] uses a memory bank of image-sentence pairs and target words. SmallCap [35] employs image-to-text retrieval to obtain sampled captions from a captions dataset. RA-CM3 [44] retrieves documents from an external memory of a mixture of text and image via a dense multimodal retriever. EXTRA [34] and Re-ViLM [43] exploit the similarity of the input image and vision candidates to retrieve captions. Different from the previous methods, our external memory contains visual-name pairs to avoid redundant information in the external captions/documents. In addition, we use an attentive fusion module to mitigate the effects of irrelevant retrieved object names on caption generation.", + "bbox": [ + 496, + 541, + 892, + 767 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3. Proposed EVCAP", + "text_level": 1, + "bbox": [ + 500, + 777, + 674, + 794 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3.1. Idea of EVCAP", + "text_level": 1, + "bbox": [ + 500, + 801, + 648, + 816 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We aim to build a retrieval-augmented LLMs-based image captioning model with a sufficiently small yet informative external memory. It involves two challenges: (1) constructing an expandable external memory and (2) building an effective LLMs-based model using retrieved object names.", + "bbox": [ + 496, + 825, + 890, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "13734", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/ba4975506959fdf05710a3b6655208a431a0dd4cb0e004e06a6ab95c19e0379c.jpg", + "image_caption": [ + "Figure 2. Schematic of our proposed EVCAP. It consists of an external visual-name memory with image embeddings and object names (upper), a frozen ViT and Q-Former equipped with trainable image query tokens, an attentive fusion module developed by a customized frozen Q-Former and trainable object name query tokens, and a frozen LLM with a trainable linear layer (lower). The ViT and Q-Former extract learned visual features from the input image, which are then used to retrieve object names from the external memory. These retrieved object names and learned visual features undergo cross-attention in the customized Q-Former, creating refined object name features. Finally, the object name features combined with visual features are fed into the LLM post a linear layer for generating captions." + ], + "image_footnote": [], + "bbox": [ + 117, + 85, + 851, + 390 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "As discussed above, challenge (1) can be resolved by utilizing the visual appearance of objects. However, if we restrict our memory to only a visual-name pair for each object, our memory will be lacking in diversity. Therefore, we gather several images for each target object. Additionally, we keep the synthetic images in our memory to avoid the harm that synthetic images might cause to our method, as pointed out in [18]. With the capability to collect images from the internet, EVCAP can be easily expanded to include novel objects from the real world effortlessly.", + "bbox": [ + 75, + 508, + 468, + 657 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We base our method on the frozen pre-trained vision model and LLM with several trainable layers (Fig. 2), giving in a model that is cheap to train. To guide the LLM, we adopt a recently popular approach called prompting as in [11, 25, 29, 35, 46]. We begin by matching the learned visual features from the input image with image embeddings stored in memory, retrieving object names. We also introduce an attentive fusion module designed to implicitly remove irrelevant retrieved names. Finally, following the attentive fusion, we combine the learned visual features and object name features to form a prompt for the LLM to generate a caption, thus addressing challenge (2).", + "bbox": [ + 75, + 659, + 470, + 840 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. External visual-name memory", + "text_level": 1, + "bbox": [ + 76, + 847, + 349, + 864 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To build the external visual-name memory, we first collect image-name pairs from the external data source. After", + "bbox": [ + 75, + 869, + 470, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "that, we encode these images into image embeddings, which serve as keys in memory, and use their names as values.", + "bbox": [ + 498, + 508, + 890, + 539 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "External data source. We utilize object images from LVIS dataset [17] to construct our external visual-name memory $\\mathcal{M}$ . Specifically, we use 1203 objects in LVIS, where we randomly select from one to ten images for each object, amounting to 8581 object images. Furthermore, as mentioned in Sec. 3.1, we also incorporate synthetic images in our memory construction. Using stable diffusion [36], we generate five additional images for each object, with a prompt of \"a photo of {object name}\", resulting in a total of $M = 14596$ ( $8581 + 5 \\times 1203$ ) images. Each object image $X^i$ is associated with an object name $v^i$ . Note that many object images may share the same object name. For the sake of simplicity, we may regard each image as corresponding to a single name. In summary, we have $M$ image-name pairs $\\{(X^i,v^i)\\}_{i=1}^M$ for external memory construction.", + "bbox": [ + 496, + 540, + 892, + 765 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "External memory construction. For each image $X^i$ , we use a frozen vision encoder $\\mathcal{E}(\\cdot)$ (see Sec. 3.3 for detail) to project it into 32 embeddings with the size of $1 \\times 768$ each: $\\{\\mathbf{k}_1^i, \\mathbf{k}_2^i, \\dots, \\mathbf{k}_{32}^i\\} = \\mathcal{E}(X^i)$ . We then average 32 embeddings to produce a single embedding $\\mathbf{k}^i$ ( $1 \\times 768$ ) that serves as the key (visual) in $\\mathcal{M}$ . The paired object name $v^i$ acts as its value (name). Consequently, we have the visual-name memory $\\mathcal{M} = \\{(\\mathbf{k}^i, v^i)\\}_{i=1}^M$ which is indexed using FAISS [21], facilitating rapid searches based on sim", + "bbox": [ + 496, + 766, + 893, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "13735", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "ilarity measures. Our memory can be expanded effortlessly by gathering additional visual-name pairs (see Sec. 5.3).", + "bbox": [ + 75, + 90, + 468, + 121 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Object names retrieval", + "text_level": 1, + "bbox": [ + 76, + 128, + 289, + 143 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Image encoding. We feed a frozen vision encoder $\\mathcal{E}$ image $X$ and image query tokens $\\mathbf{T}_{\\mathrm{img}}$ to produce visual features $\\mathcal{Q}$ . To enable the retrieval process controllable, we make image query tokens to be trainable. Thus, the image encoding process can be summarized as $\\mathcal{Q} = \\mathcal{E}(X,\\mathbf{T}_{\\mathrm{img}})$ . We use the BLIP-2 pre-trained vision encoder [25], which consists of a pre-trained vision transformer ViT-g [14] outputting image features $(257\\times 1408)$ , and a Q-Former receiving image features producing $|\\mathcal{Q}| = 32$ learned visual features $(1\\times 768$ each). We denote $\\mathcal{Q} = \\{\\mathbf{q}_1,\\mathbf{q}_2,\\dots,\\mathbf{q}_{32}\\}$ .", + "bbox": [ + 75, + 152, + 467, + 301 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Retrieval. Having obtained $\\mathcal{Q}$ , we calculate the cosine similarity between the query $\\mathbf{q}_j \\in \\mathcal{Q}$ and the key $\\mathbf{k}^i \\in \\mathcal{M}$ . The similarity calculation is given by $\\mathrm{SIM}(\\mathbf{q}_j, \\mathbf{k}^i) = \\frac{\\mathbf{q}_j^\\top \\mathbf{k}^i}{\\|\\mathbf{q}_j\\| \\|\\mathbf{k}^i\\|}$ , where $i \\in [1, M]$ , $j \\in [1, 32]$ . Given each $\\mathbf{q}_j$ , we select one key with the highest similarity score, resulting in 32 key-value candidates $\\{\\mathbf{k}_j^{\\mathrm{best}}, v_j^{\\mathrm{best}}\\}_{j=1}^{32}$ .", + "bbox": [ + 75, + 303, + 467, + 401 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "After that, we filter out candidates with repeated object names (values), and then select the top-K values. In particular, we determine the index $j$ from the key that has the highest SIM score. These selected values $v_{j}^{\\mathrm{best}}$ are redefined as the new notation $v_{l}$ in the retrieved top-K object names for the input image, which can be summarized as follows:", + "bbox": [ + 75, + 401, + 467, + 489 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\{\\mathbf {k} _ {j} ^ {\\text {b e s t}}, v _ {j} ^ {\\text {b e s t}} \\} = \\arg \\max _ {\\mathbf {k} ^ {i}} \\operatorname {S I M} \\left(\\mathbf {q} _ {j}, \\mathbf {k} ^ {i}\\right),\n$$\n", + "text_format": "latex", + "bbox": [ + 135, + 503, + 406, + 527 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nj = \\arg \\max _ {j} \\mathrm {S I M} (\\mathbf {q} _ {j}, \\mathbf {k} _ {j} ^ {\\mathrm {b e s t}}), v _ {l} \\gets v _ {j} ^ {\\mathrm {b e s t}},\n$$\n", + "text_format": "latex", + "bbox": [ + 133, + 539, + 411, + 564 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $l \\in [1, \\mathrm{K}]$ . As a result, the retrieved top- $\\mathbf{K}$ object names are $\\{v_l\\}_{l=1}^{\\mathrm{K}}$ .", + "bbox": [ + 75, + 568, + 467, + 601 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.4. Attentive fusion", + "text_level": 1, + "bbox": [ + 76, + 607, + 235, + 622 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Since the object names obtained from the retrieval process may be redundant, we develop an attentive fusion module to selectively distill object name features.", + "bbox": [ + 75, + 630, + 467, + 674 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The retrieved object names $\\{v_{l}\\}_{l = 1}^{\\mathrm{K}}$ are concatenated together into a sequence $S$ , each separated by a delimiter: $S = \\{v_{1},[\\mathrm{SEP}],v_{2},[\\mathrm{SEP}],\\dots ,[\\mathrm{SEP}],v_{\\mathrm{K}}\\}$ . The sequence $S$ and visual features $\\mathcal{Q}$ are fed into a customized Q-Former $\\mathcal{F}(\\cdot)$ , which is constructed from the frozen pretrained Q-Former as we used in vision encoder $\\mathcal{E}$ . Nonetheless, in order to enable object names to get attention from visual features, we switch the image embedding port and the text instruction port (see the supplement for architecture detail). Like in the image encoding process in Sec. 3.3, we make the object name query tokens $\\mathbf{T}_{\\mathrm{obj}}$ learnable during training to assist in learning object name features related to the caption. The size of $\\mathbf{T}_{\\mathrm{obj}}$ is $P\\times 768$ , where $P$ indicates the number of object name query tokens. We get the object name features $\\mathcal{V} = \\mathcal{F}(S,\\mathcal{Q},\\mathbf{T}_{\\mathrm{obj}})$ .", + "bbox": [ + 75, + 675, + 467, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.5. Caption generation", + "text_level": 1, + "bbox": [ + 500, + 90, + 684, + 106 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Before inputting the visual features $\\mathcal{Q}$ and object name features $\\nu$ into the LLM decoder, we concatenate $(\\oplus)$ them and use a linear layer $\\phi(\\cdot)$ to project them into the input latent space of the LLM as $\\phi(\\mathcal{Q} \\oplus \\mathcal{V})$ . The LLM used for caption generation in this work is the pre-trained Vicuna-13B [10], an open-source chatbot constructed from LLaMA [37]. During training and evaluation, we design a prompt in a conversational format, that is similar to [46]:", + "bbox": [ + 496, + 114, + 890, + 233 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Human: Describe this image in detail. \nAssistant:", + "bbox": [ + 524, + 244, + 857, + 290 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "in which, ProjFeature denotes the projected feature $\\phi (\\mathcal{Q}\\oplus \\mathcal{V})$ after the linear layer. In training phase, given input caption tokens $\\{c_i\\}_{i = 1}^L$ , the LLM decoder concatenates the embedded prompt $\\{\\mathbf{w}_i\\}_{i = 1}^N$ and the embedded caption tokens $\\{\\mathbf{c}_i\\}_{i = 1}^L$ as input, and predicts the caption tokens in an autoregressive fashion, while in the evaluation phase, we only need to input the embedded prompt. We train EV-CAP by minimizing the cross-entropy loss in an end-to-end way: $\\mathcal{L}_{\\theta} = -\\sum_{i = 1}^{L}\\log p_{\\theta}(c_i|\\mathbf{w}_1,\\dots \\mathbf{w}_N,\\mathbf{c}_1,\\dots ,\\mathbf{c}_{i - 1})$ , in which $\\theta$ indicates the trainable parameters.", + "bbox": [ + 496, + 300, + 890, + 450 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Experimental Settings", + "text_level": 1, + "bbox": [ + 500, + 467, + 709, + 484 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1. Training setup", + "text_level": 1, + "bbox": [ + 500, + 493, + 648, + 508 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Implementation. EVCAP uses the same image encoder as in BLIP-2 [25], consisting of a ViT-g [14] and their pretrained Q-Former. Since we intend to obtain object name features through cross-attention between retrieved object names and visual features, we develop a customized Q-Former, which consists of BERT [23] with cross-attention layers inserted at every other transformer block. We use a frozen Vicuna-13B [10] as the caption generator.", + "bbox": [ + 496, + 517, + 890, + 637 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Training dataset. For all experiments, we exclusively train EVCAP using the training set of COCO dataset [27], consisting of 82k images and 5 captions per images. The entire training process takes about 3 hours on 4 A6000 GPUs, using mixed precisions (more details in the supplementary).", + "bbox": [ + 496, + 638, + 890, + 714 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.2. Evaluation setup", + "text_level": 1, + "bbox": [ + 500, + 726, + 665, + 742 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Evaluation dataset. We evaluate EVCAP, trained using the COCO training set, across four datasets: its test set, two challenging benchmarks - NoCaps validation set and Flickr30k test set, and a synthetic commonsense-violating dataset - WHOOPS. We adhere to follow prior work [15, 41] to use the same images of Karpathy split [22] on COCO test set, NoCaps [2] validation set, and Karpathy split on Flickr30k [30] test set. In addition, WHOOPS [6] is a synthetic image captioning dataset comprising 500 synthetic commonsense-violating images and 2500 paired captions.", + "bbox": [ + 496, + 750, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "13736", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/82c1a4c9cda268d875841fd66754cb0bd89005508b338685a8f58c897073195a.jpg", + "table_caption": [ + "Table 1. Quantitative comparison against SOTA methods on three common image captioning benchmarks. * denotes using a memory bank. We report the size of training data and parameters; BLEU@4 (B@4), METEOR (M), CIDEr (C), and SPICE (S) scores on COCO test set; C and S scores on in-domain, near-domain, out-domain and overall data of NoCaps validation set; C and S scores on Flickr30k test set. Higher score is better. Bold indicates the best results among compared methods, normal indicates the second best results." + ], + "table_footnote": [], + "table_body": "
MethodTrainingCOCONoCaps valFlickr30k Test
DataPara.TestIn-domainNear-domainOut-domainOverallCS
B@4MCSCSCSCSCSCS
Heavyweight-training models
VinVL [45]8.9M110M38.230.3129.323.696.813.590.713.187.411.690.912.8--
AoANet+MA* [16]COCO-38.028.7121.021.8----------
NOC-REK* [40]COCO110M----104.714.8100.214.1100.713.0100.914.0--
RCA-NOC* [13]COCO110M37.429.6128.423.192.212.987.812.687.511.588.312.4--
ViEcap GPT2 [15]COCO124M27.224.892.918.261.110.464.39.965.08.666.29.547.913.6
InstructBLIP Vicuna-13B [11]129M188M----------121.9-82.8-
OSCAR [26]4.1M338M37.430.7127.823.583.412.081.612.077.610.681.111.7--
BLIP [24]129M446M40.4-136.7-114.915.2112.114.9115.314.4113.214.8--
BLIP-2 FlanT5-XL [25]129M1.2B42.4-144.5-123.716.3120.215.9124.815.1121.615.8--
REVEAL* T5 [20]1.3B2.1B--145.4-------123.0---
Lightweight-training models
MiniGPT4 Vicuna-13B [46]5M3.94M38.029.6129.623.499.014.8106.915.3110.814.9108.815.178.416.9
SmallCap* GPT2 [35]COCO7M37.027.9119.721.3--------60.6-
ClipCap GPT2 [29]COCO43M33.527.5113.121.184.912.166.810.949.19.665.810.9--
EVCAP* Vicuna-13BCOCO3.97M41.531.2140.124.7111.715.3119.515.6116.514.7119.315.384.418.0
Specialist SOTAs
Qwen-VL Qwen-7B [5]1.4B9.6B----------121.4-85.8-
CogVLM Vicuna-7B [41]1.5B6.5B--148.7-----132.6-128.3-94.9-
PaLI mT5-XXL [9]1.6B17B--149.1-------127.0---
PaLI-X_UL2-32B [8]2.2B55B--149.2-------126.3---
", + "bbox": [ + 80, + 151, + 890, + 426 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Compared methods. We compare EVCAP with several SOTAs. According to the trainable parameters size, they can be divided into 1) Heavyweight-training (between 100M to 5B): VinVL [45], AoANet [16], NOC-REK [40], RCA-NOC [13], ViECap [15], InstructBLIP [11], OSCAR [26], BLIP [24], BLIP-2 [25], REVEAL [20]; 2) Lightweight-training (less than 100M): MiniGPT4 [46], SmallCap [35], ClipCap [29]; and also 3) Specialist SOTAs with huge trainable parameters (larger than 5B): QwenVL [5], CogVLM [41], PaLI [9], PaLI-X [8]. Among these methods, AoANet, NOC-REK, RCA-NOC, REVEAL, and SmallCap are retrieval-augmented captioning methods.", + "bbox": [ + 75, + 450, + 472, + 633 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5. Experimental Results", + "text_level": 1, + "bbox": [ + 76, + 651, + 282, + 667 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.1. Results on in-/out-domain benchmarks", + "text_level": 1, + "bbox": [ + 76, + 678, + 410, + 693 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We assess EVCAP against SOTAs on both in-domain and out-domain benchmarks. The COCO test set can be considered as in-domain data as we only train our model on the COCO training set. Out-domain benchmarks are the NoCaps validation set and the Flickr30k test set.", + "bbox": [ + 75, + 703, + 468, + 777 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Quantitative results. Tab. 1 details our EVCAP's performance in comparison with SOTA methods. We first evaluate training costs in terms of training data sizes and parameters. Similar to various heavyweight-training models that exclude LLMs and the majority of lightweight-training models, EVCAP is trained solely on the COCO training set. It utilizes only 3.97M trainable parameters, positioning it as the second smallest, slightly larger than MiniGPT4", + "bbox": [ + 75, + 780, + 470, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "with 3.94M. Among lightweight-training models, our approach outperforms others, achieving the highest scores on all benchmarks. Despite using less training data and nearly identical trainable parameters as MiniGPT4, EVCAP significantly surpasses it, with a marked improvement of 10.5, 10.5, and 6.0 in CIDEr scores for each benchmark. When further compared with heavyweight-training models, the performance of EVCAP stands out among million-level models, nearly matching InstructBLIP, except in NoCaps. Note that since BLIP-2 does not include Vicuna checkpoints, InstructBLIP performs pre-training with Vicuna using the same procedure as BLIP-2, whereas EVCAP does not involve pre-training. Against REVEAL, which also uses external memory, our EVCAP utilizes about 1/3000 training data and 1/500 training parameters yet yields comparable results. Moreover, EVCAP's performance is on par with BLIP-2, the top-performing model with 1.2B trainable parameters. This highlights EVCAP's efficiency and effectiveness despite its significantly smaller training cost, thanks to our external visual-name memory. Regarding specialist SOTAs, they use billion-level training data and over 5B trainable parameters, so it is acceptable that they can achieve exceptionally strong performance, surpassing EVCAP by nearly 10 on all benchmarks in CIDEr scores.", + "bbox": [ + 496, + 450, + 890, + 811 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Qualitative results. Fig. 3 presents a comparison of captions generated by our EVCAP and three SOTA models across three benchmarks. The captions of SmallCap are generated by its publicly accessible demo [1]. We generate captions of MiniGPT4 and BLIP-2 using their respec", + "bbox": [ + 496, + 825, + 890, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "13737", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "COCO Test", + "text_level": 1, + "bbox": [ + 205, + 92, + 281, + 103 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/f0c7cd5fc601182ad9ef28928e582b8e31464589b2a50de68eebf324384de677.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 138, + 106, + 346, + 198 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "GT: A green bus driving through a rural area with trees in the background.", + "bbox": [ + 135, + 200, + 349, + 222 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "SmallCap: A bus driving down a street next to trees.", + "bbox": [ + 135, + 224, + 349, + 246 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "MiniGPT4: A green bus is driving down the street.", + "bbox": [ + 135, + 248, + 318, + 271 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "BLIP-2: A green bus driving down a road with trees in the background.", + "bbox": [ + 135, + 273, + 334, + 296 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "EVCap: A green bus driving down a road next to trees.", + "bbox": [ + 135, + 297, + 331, + 320 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/31f38804d88ed9718adb39319090b734d39a5158c437da97db65dc1d7b8286a4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 140, + 324, + 346, + 417 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "GT: A woman in a blue top with headphones and two cellphones.", + "bbox": [ + 135, + 419, + 316, + 441 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "SmallCap: A woman sitting in front of a laptop computer.", + "bbox": [ + 135, + 443, + 341, + 465 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "MiniGPT4: A woman sitting on a couch holding two phones.", + "bbox": [ + 135, + 468, + 315, + 491 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "BLIP-2: A woman sitting on a couch with two cell phones.", + "bbox": [ + 135, + 493, + 333, + 516 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "EVCap: A woman wearing headphones holding two cell phones.", + "bbox": [ + 135, + 517, + 338, + 540 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "NoCaps Val", + "text_level": 1, + "bbox": [ + 446, + 92, + 524, + 104 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/db8065873222d29d9fccd12b292d519b520e4922206aa27ef18f77eb871b7e91.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 382, + 106, + 589, + 198 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "GT: The two guinea pigs are getting dried off in a yellow towel.", + "bbox": [ + 377, + 200, + 573, + 220 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "SmallCap: A person holding a small animal in a towel.", + "bbox": [ + 377, + 223, + 576, + 246 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "MiniGPT4: Two small animals are wrapped in a towel.", + "bbox": [ + 377, + 248, + 565, + 271 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "BLIP-2: Two guinea pigs wrapped in a yellow towel.", + "bbox": [ + 377, + 273, + 588, + 296 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "EVCap: Two guinea pigs are wrapped in a yellow towel.", + "bbox": [ + 377, + 297, + 586, + 321 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/371ec2cfc3a252917fb63fa2cc8dfb97a18c771ee7895a3edfd0c8bd8b5755e9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 382, + 324, + 589, + 417 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "GT: A computer screen showing two men sitting at a table.", + "bbox": [ + 377, + 419, + 576, + 441 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "SmallCap: Two men sitting at a table with a laptop.", + "bbox": [ + 377, + 443, + 583, + 465 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "MiniGPT4: A laptop computer sitting on top of a table.", + "bbox": [ + 377, + 468, + 586, + 489 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "BLIP-2: A laptop computer with a picture of two men on it.", + "bbox": [ + 377, + 492, + 560, + 515 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "EVCap: A laptop computer with a picture of two men on the screen.", + "bbox": [ + 377, + 517, + 560, + 539 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Flickr30k Test", + "text_level": 1, + "bbox": [ + 679, + 92, + 772, + 103 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/26e81763047a240653926a483fd60d9f7d4d2836fe77c73f79e1d6146590b7b0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 622, + 106, + 830, + 198 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "GT: A very young child in a denim baseball cap eats a green apple.", + "bbox": [ + 617, + 200, + 803, + 222 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "SmallCap: A young boy holding an apple in his hand.", + "bbox": [ + 617, + 223, + 810, + 246 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "MiniGPT4: A baby sitting in a high chair eating an apple.", + "bbox": [ + 617, + 248, + 807, + 272 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "BLIP-2: A baby sitting in a white chair eating a green apple.", + "bbox": [ + 617, + 273, + 825, + 296 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "EVCap: A toddler eating a green apple while wearing a hat.", + "bbox": [ + 617, + 297, + 831, + 321 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/dd0c497f9a55af1019c1e154598255be1fcefe9d34c0a9b98cf0ca951e4b2dbb.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 622, + 325, + 830, + 417 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "GT: Two men are riding on a wooden vehicle pulled by two donkeys.", + "bbox": [ + 617, + 419, + 821, + 441 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "SmallCap: A donkey pulling a cart with a man in the background.", + "bbox": [ + 617, + 443, + 831, + 465 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "MiniGPT4: Two men riding on a donkey in the dirt.", + "bbox": [ + 617, + 467, + 795, + 491 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "BLIP-2: Two men riding a horse drawn cart through a field.", + "bbox": [ + 617, + 493, + 831, + 515 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "EVCap: Two men riding in a cart pulled by two donkeys.", + "bbox": [ + 617, + 517, + 797, + 540 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "tive pre-trained models. As a lightweight and retrieval-augmented captioning method, SmallCap struggles to produce accurate captions for given images, primarily because it relies on retrieved captions laden with extraneous information. MiniGPT4, though aligned with the primary content of images, sometimes misses certain objects like \"trees\" and \"headphones\". This oversight stems from its focus on the main objects in images, without integrating additional cues for other objects provided by the retrieved object names. In contrast, the captions generated by our EVCAP are comparable to those of BLIP-2.", + "bbox": [ + 75, + 625, + 468, + 789 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.2. Results on commonsense-violating data", + "text_level": 1, + "bbox": [ + 76, + 800, + 413, + 816 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To explore our EVCAP's capability in describing contents in open-word settings, we further evaluate it on WHOOPS dataset, which contains commonsense-violating images.", + "bbox": [ + 75, + 824, + 468, + 869 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Quantitative results. In Tab. 2, we compare the performance of EVCAP, MiniGPT4, BLIP, and BLIP-2 on", + "bbox": [ + 75, + 869, + 468, + 898 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/dfb894901fe60fd7e49981692efbf55e5abf7afaf7457c96a642077659221d9c.jpg", + "image_caption": [ + "Figure 3. Examples of captions generated by our EVCAP and three SOTA methods on COCO test set, NoCaps validation set, and Flickr30k test set. GT refers to the Ground Truth captions. Incorrect objects in captions are highlighted in red, while correct ones are in blue. Our EVCAP correctly generates captions across different datasets, showing performance comparable to BLIP-2.", + "Figure 4. Examples of captions generated by our EVCAP, EVCAP (w/ WHOOPS), and three SOTAs on WHOOPS dataset. Incorrect objects are highlighted in red, while correct ones are in blue." + ], + "image_footnote": [], + "bbox": [ + 503, + 626, + 588, + 691 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "GT: This is an image of a blue Pikachu with yellow accents.", + "bbox": [ + 589, + 623, + 852, + 633 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "SmallCap: A blue and white stuffed animal on a table.", + "bbox": [ + 589, + 633, + 831, + 643 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "MiniGPT4: A blue poker sitting on the floor with its eyes closed.", + "bbox": [ + 589, + 643, + 887, + 654 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "BLIP-2: A blue Pikachu sitting in a dark room.", + "bbox": [ + 589, + 655, + 792, + 664 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "EVCap: A blue and yellow cartoon character sitting on a dark background.", + "bbox": [ + 589, + 665, + 862, + 683 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "EVCap (w/ WHOOPS): A blue Pikachu is sitting on the floor.", + "bbox": [ + 589, + 684, + 857, + 694 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "WHOOPS dataset. This dataset is particularly challenging due to its inclusion of unusual objects [6]. Initially, as an end-to-end trained model, our EVCAP exhibits performance similar to MiniGPT4. However, there is a noticeable improvement in the CIDEr score, after the external memory is enriched with 2396 new objects from the WHOOPS dataset, each represented by 5 synthesized images generated using stable diffusion [36]. It highlights the effectiveness of", + "bbox": [ + 496, + 780, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "13738", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/b7da35384deaeb912f913760fdb15001de9389f1dca09d1f270a8551e7e1f6c5.jpg", + "table_caption": [ + "Table 2. Quantitative results on commonsense-violating data - WHOOPS dataset. EVCAP (w/ WHOOPS) denotes EVCAP using the memory expanded by WHOOPS objects. The results reveal the open-world comprehension ability and expandability of EVCAP." + ], + "table_footnote": [], + "table_body": "
MethodB@4MCS
Only pre-trained models
BLIP [24] (from [6])13-65-
BLIP-2 FlanT5-XXL [25] (from [6])31-120-
BLIP-2 FlanT5-XXL [25] (reproduced)2826.793.117.9
Finetuned models on COCO
MiniGPT4 [46]24.226.784.818.2
BLIP [24]22.925.079.317.1
BLIP-2 FlanT5-XL [25]25.827.089.118.3
End-to-end trained models on COCO
EVCAP24.126.185.317.7
EVCAP (w/ WHOOPS)24.426.186.317.8
", + "bbox": [ + 80, + 148, + 467, + 327 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/3c410699e381dc1838b3aff36ed69e3b2626a5ba083231bcb25a1eb04d23dfa2.jpg", + "table_caption": [ + "Table 3. Ablation study on components prior to the LLM decoder in EVCAP. The result of \"+\" Attentive fusion\" demonstrates the substantial impact of the external visual-name memory." + ], + "table_footnote": [], + "table_body": "
MethodCOCO testNoCaps valFlickr30k test
CSCSCS
ViT + Q-Former (Baseline)134.423.9108.814.276.817.3
+ Image query tokens (Baseline+)134.123.8109.014.377.317.2
+ Attentive fusion (EVCAP)140.124.7119.315.384.418.0
", + "bbox": [ + 80, + 386, + 467, + 452 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/1a7a4498382342c04d35bddc17714b4bcfea88d424b95000a66b935bcedfc598.jpg", + "image_caption": [ + "Figure 5. Visualization of the captions generated from ablation study on the NoCaps validation set. We also show the retrieved object names by EVCAP, presented in gray. Incorrect objects in captions are highlighted in red, while correct ones are in blue." + ], + "image_footnote": [], + "bbox": [ + 80, + 468, + 165, + 529 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "GT: A hamster on a blanket with a hand behind it.", + "bbox": [ + 166, + 468, + 385, + 478 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Baseline: A hamster sitting on a couch with a person's hand near it.", + "bbox": [ + 166, + 479, + 467, + 489 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Baseline+: A close up of a small hamster on a bed.", + "bbox": [ + 168, + 489, + 393, + 500 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "EVCap: A small hamster sitting on a blanket next to a hand.", + "bbox": [ + 168, + 500, + 431, + 510 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "(pillow, hamster, grizzly, gameboard, kitten, spectacles, cat, blanket, giraffe, pug-dog)", + "bbox": [ + 168, + 511, + 467, + 529 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "our idea of incorporating an expandable external memory into the captioning model for open-world comprehension.", + "bbox": [ + 75, + 626, + 467, + 656 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Qualitative results. Fig. 4 illustrates the captions generated by EVCAP, EVCAP (w/WHOOPS), and three SOTAs for one image from the WHOOPS dataset. Similar to other methods except for BLIP-2, EVCAP can not recognize \"blue cartoon character\" as \"Pikachu\", while EVCAP (w/WHOOPS) successfully predicts it because of the updated memory. SmallCap and MiniGPT4 tend to generate captions with hallucinatory objects, a result of commonsense-violating contents present in the images.", + "bbox": [ + 75, + 656, + 468, + 792 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.3. Detailed analysis", + "text_level": 1, + "bbox": [ + 76, + 801, + 243, + 816 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Ablation study. We assess the contribution of each component prior to the LLM decoder in EVCAP by incrementally integrating the image query tokens and the attentive fusion module into our baseline model. The baseline model comprises a ViT+Q-Former, a linear layer, and a LLM decoder.", + "bbox": [ + 75, + 825, + 467, + 900 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/954475c57db53357a2a3c352f997d05f6391c647ae4e380cc63d1da9b6448115.jpg", + "image_caption": [ + "Figure 6. Visualization of the visual features in external memory using t-SNE. For visual features in LVIS dataset's objects (blue), the related objects fall in the same cluster. After adding more visual features of synthesized images from WHOOPS' objects, new objects (red) are located at appropriate clusters (zoom-in view)." + ], + "image_footnote": [], + "bbox": [ + 504, + 88, + 885, + 261 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The quantitative results are shown in Tab. 3. When employing only the baseline model (Baseline), CIDEr scores drop notably by 5.7, 10.5, and 7.6 on COCO, NoCaps, and Flickr30k, respectively. The inclusion of trainable image query tokens (Baseline+) brings a marginal improvement on NoCaps and Flickr30k. However, the performance is significantly enhanced with the addition of attentive fusion (along with the introduction of external memory), indicating the pivotal role of the external visual-name memory in the overall effectiveness of EVCAP. This is further corroborated by the qualitative results in Fig. 5, where captions from Baseline and Baseline+ inaccurately include objects like \"couch\" and \"bed\", and Baseline+ overlooks \"hand\".", + "bbox": [ + 496, + 369, + 890, + 566 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Exploration for external memory expandability. To demonstrate the scalability of the external memory in EVCAP, we visualize the visual features stored in LVIS external memory, and newly synthesized data from objects appearing in the WHOOPS dataset. We employ t-SNE [38] to plot visual features after reducing their dimensions to 2-D (Fig. 6). For clear visualization, we only randomly display 3649 visual features in LVIS memory, and add 479 visual features from WHOOPS objects. Among them, 35 samples are randomly labeled. The result shows a clear clustering of LVIS objects (blue) in the external memory, as well as the successful integration and appropriate localization of new objects from WHOOPS (red) into these clusters. This pattern not only confirms the distinctiveness of visual features already present in the memory but also demonstrates the potential to accurately incorporate and differentiate new objects introduced from updated data. These findings highlight our external memory's ability to expand and maintain its effectiveness even as new data is incorporated.", + "bbox": [ + 496, + 568, + 892, + 852 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Impact of external memory size. We examine the impact of external memory size in Tab. 4. On the one hand, we randomly remove $30\\%$ , $60\\%$ , and $90\\%$ data in the external", + "bbox": [ + 496, + 854, + 893, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "13739", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/1d2c22461d754c669a628371b471ef48079dff4cbd4a7a07853f690db5f47028.jpg", + "table_caption": [ + "Table 4. Impact of the external memory size on the performance of EVCAP by evaluation under CIDEr scores. Changes in the size of external memory result in changes in performance." + ], + "table_footnote": [], + "table_body": "
MethodNoCaps valFlickr30k
InNearOutOverallTest
LVIS objects (EVCAP)111.7119.5116.5119.384.4
- 30% LVIS112.0119.2115.3118.885.0
- 60% LVIS111.4119.1116.2119.085.1
- 90% LVIS110.6118.2115.8118.383.6
+ WHOOPS110.7118.9116.7119.084.9
", + "bbox": [ + 81, + 133, + 467, + 236 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/c00825c28763cc8b55ca76a5d0aa5ee5fe752db79122330b7159a2860ab28ec3.jpg", + "image_caption": [ + "Figure 7. CIDEr scores after training EVCAP with the number of retrieved object names K from 0 to 20. The results indicate that the performance is relatively optimal when K is set to be 10." + ], + "image_footnote": [], + "bbox": [ + 81, + 250, + 465, + 367 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "memory constructed from LVIS objects. The results show the performance gradually degrades on NoCaps as reducing $30\\%$ and $90\\%$ LVIS. Despite some unexpected increases in certain results on NoCaps (5th row) and Flickr30k (4th - 5th rows), they do not alter the overall downward trend. Similar phenomena are also noted in SmallCap [35], we speculate it is due to data distribution. On the other hand, as we infuse WHOOPS knowledge into LVIS memory, there is a slight improvement on NoCaps (out) and Flickr30k. These observations validate the model's capability to effectively retrieve object names from an updated memory, enhancing its performance in generating captions.", + "bbox": [ + 75, + 448, + 468, + 628 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Impact of the number of retrieved object names. We investigate how the number of retrieved object names K (Sec. 3.3) affect EVCAP in Fig. 7. We train the model with K from 0 to 20 and evaluate the performance under CIDER on all three benchmarks. From the results, we can find that the model works worst on the out-domain dataset (NoCaps) with zero object names. It confirms that the performance boost from Baseline+ to EVCAP (Tab. 3) is primarily attributed to the retrieval-augmented mechanism, but not the customized Q-Former itself. With more object names, performance fluctuates but improves. Furthermore, we observe that setting K to 10 yields relatively optimal overall performance, validating the choice of $\\mathrm{K} = 10$ in EVCAP.", + "bbox": [ + 75, + 628, + 468, + 824 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Analysis with different decoders. To explore the influence of different LLMs decoders on our EVCAP, we experiment by substituting Vicuna-13B with GPT2 and Vicuna-7B, as detailed in Tab. 5. With GPT2 as the decoder, EVCAP still markedly surpasses other GPT2-based models, achieving", + "bbox": [ + 75, + 825, + 470, + 902 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/17e5f056a4540629f5df2aa9da5fea6697c41dc0b59d6e904d7d75e285d69700.jpg", + "table_caption": [ + "Table 5. Analysis with different LLM decoders including GPT2, Vicuna-7B, and Vicuna-13B. The results reveal EVCAP is effective when applying it in different LLM decoders." + ], + "table_footnote": [], + "table_body": "
MethodLLMCOCO testNoCaps valFlickr30k test
CSCSCS
SmallCap [35]GPT2119.721.3--60.6-
ViEcap [15]GPT292.918.266.29.547.913.6
EVCAPGPT2131.023.297.613.370.616.1
MiniGPT4 [46]Vicuna-7B119.423.5108.715.773.917.2
InstructBLIP [11]Vicuna-7B--123.1-82.4-
EVCAPVicuna-7B139.024.7116.815.382.718.0
MiniGPT4 [46]Vicuna-13B129.623.4108.815.178.416.9
InstructBLIP [11]Vicuna-13B--121.9-82.8-
EVCAPVicuna-13B140.124.7119.315.384.418.0
", + "bbox": [ + 503, + 133, + 890, + 275 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "impressive gains of 11.3 and 10.0 under CIDEr on COCO and Flickr30k, compared to SmallCap. When employing Vicuna-7B, the comparison of performance trends mirrors those observed with Vicuna-13B, further attesting to the robustness and adaptability of EVCAP across different LLM decoders. Notably, both SmallCap, which retrieves captions, and our GPT2-based EVCAP, which retrieves object names, use the same GPT2 decoder. Therefore, their comparison also underscores the effectiveness of our method's object name retrieval and attentive fusion strategy.", + "bbox": [ + 496, + 303, + 890, + 455 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Limitations. First, EVCAP cannot retrieve all objects that appear in the given image due to the memory coverage limits, leading to incomplete image descriptions (Fig. 4). We will investigate integrating object detection with image captioning to enhance completeness. Second, our focus on object representation restricts consideration of other crucial captioning elements, affecting overall performance. Similar to all models trained with COCO dataset, EVCAP has limitations in generating varied styles, which is reflected in our relatively modest performance improvements in Tab. 2, compared to MiniGPT4. We will overcome it by exploring methodologies that encourage style diversity in the future.", + "bbox": [ + 496, + 455, + 892, + 636 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 654, + 619, + 671 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We further advance image captioning in real-world scenarios by introducing EVCAP, a novel image captioning model with object names retrieved from an external visual-name memory. The external memory is easily expandable, allowing for effortless updates with new object visuals and names. We extensively compare EVCAP with SOTAs on various benchmarks and commonsense-violating data, demonstrating its significant superiority in performance.", + "bbox": [ + 496, + 681, + 890, + 801 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements. This work was supported by JSPS/MEXT KAKENHI Grant Numbers JP22H05015, JP23H03449, JP23KJ0404, and JP22K17947, and the commissioned research (No. 225) by the National Institute of Information and Communications Technology (NICT), Japan, and the Institute of AI and Beyond of the University of Tokyo, and ROIS NII Open Collaborative Research 2023-23FC01, 2024-24S1201.", + "bbox": [ + 496, + 801, + 893, + 901 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "13740", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 174, + 106 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] https : / / huggingface . co / spaces / RitaParadaRamos/SmallCapDemo.5", + "[2] Harsh Agrawal, Karan Desai, Yufei Wang, Xinlei Chen, Rishabh Jain, Mark Johnson, Dhruv Batra, Devi Parikh, Stefan Lee, and Peter Anderson. Nocaps: Novel object captioning at scale. In Proc. IEEE International Conference on Computer Vision (ICCV), 2019. 4", + "[3] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: A visual language model for few-shot learning. In Advances in Neural Information Processing Systems (NeurIPS), 2022. 1", + "[4] Peter Anderson, Xiaodong He, Chris Buehler, Damien Teney, Mark Johnson, Stephen Gould, and Lei Zhang. Bottom-up and top-down attention for image captioning and visual question answering. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 2", + "[5] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A frontier large vision-language model with versatile abilities. arXiv preprint arXiv:2308.12966, 2023. 5", + "[6] Nitzan Bitton-Guetta, Yonatan Bitton, Jack Hessel, Ludwig Schmidt, Yuval Elovici, Gabriel Stanovsky, and Roy Schwartz. Breaking common sense: Whoops! a vision-and-language benchmark of synthetic and compositional images. In Proc. IEEE International Conference on Computer Vision (ICCV), 2023. 4, 6, 7", + "[7] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. In Advances in Neural Information Processing Systems (NeurIPS), 2020. 2", + "[8] Xi Chen, Josip Djolonga, Piotr Padlewski, Basil Mustafa, Soravit Changpinyo, Jialin Wu, Carlos Riquelme Ruiz, Sebastian Goodman, Xiao Wang, Yi Tay, et al. PaLI-X: On scaling up a multilingual vision and language model. arXiv preprint arXiv:2305.18565, 2023. 1, 2, 5", + "[9] Xi Chen, Xiao Wang, Soravit Changpinyo, AJ Piergiovanni, Piotr Padlewski, Daniel Salz, Sebastian Goodman, Adam Grycner, Basil Mustafa, Lucas Beyer, Alexander Kolesnikov, Joan Puigcerver, Nan Ding, Keran Rong, Hassan Akbari, Gaurav Mishra, Linting Xue, Ashish V Thapliyal, James Bradbury, Weicheng Kuo, Mojtaba Seyedhosseini, Chao Jia, Burcu Karagol Ayan, Carlos Riquelme Ruiz, Andreas Peter Steiner, Anelia Angelova, Xiaohua Zhai, Neil Houlsby, and Radu Soricut. PaLI: A jointly-scaled multilingual language-image model. In International Conference on Learning Representations (ICLR), 2023. 1, 2, 5", + "[10] Wei-Lin Chiang, Zhuohan Li, Zi Lin, Ying Sheng, Zhanghao Wu, Hao Zhang, Lianmin Zheng, Siyuan Zhuang, Yonghao Zhuang, Joseph E Gonzalez, et al. Vicuna: An open-source chatbot impressing gpt-4 with $90\\%$ chatgpt quality. See https://vicuna.lmsys.org (accessed 14 April 2023), 2023. 2, 4", + "[11] Wenliang Dai, Junnan Li, Dongxu Li, Anthony Meng Huat Tiong, Junqi Zhao, Weisheng Wang, Boyang Li, Pascale" + ], + "bbox": [ + 78, + 114, + 470, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Fung, and Steven Hoi. Instructclip: Towards general-purpose vision-language models with instruction tuning. In Advances in Neural Information Processing Systems (NeurIPS), 2023. 3, 5, 8", + "[12] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations (ICLR), 2021. 2", + "[13] Jiashuo Fan, Yaoyuan Liang, Leyao Liu, Shaolun Huang, and Lei Zhang. Rca-noc: Relative contrastive alignment for novel object captioning. In Proc. IEEE International Conference on Computer Vision (ICCV), 2023. 2, 5", + "[14] Yuxin Fang, Wen Wang, Binhui Xie, Quan Sun, Ledell Wu, Xinggang Wang, Tiejun Huang, Xinlong Wang, and Yue Cao. Eva: Exploring the limits of masked visual representation learning at scale. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 2, 4", + "[15] Junjie Fei, Teng Wang, Jinrui Zhang, Zhenyu He, Chengjie Wang, and Feng Zheng. Transferable decoding with visual entities for zero-shot image captioning. In Proc. IEEE International Conference on Computer Vision (ICCV), 2023. 2, 4, 5, 8", + "[16] Zhengcong Fei. Memory-augmented image captioning. In Proc. AAAI Conference on Artificial Intelligence (AAAI), 2021. 2, 5", + "[17] Agrim Gupta, Piotr Dollar, and Ross Girshick. LVIS: A dataset for large vocabulary instance segmentation. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 3", + "[18] Ryuichiro Hataya, Han Bao, and Hiromi Arai. Will large-scale generative models corrupt future datasets? In Proc. IEEE International Conference on Computer Vision (ICCV), 2023. 3", + "[19] Lisa Anne Hendricks, Subhashini Venugopalan, Marcus Rohrbach, Raymond Mooney, Saenko Kate, and Trevor Darrell. Deep compositional captioning: Describing novel object categories without paired training data. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 2", + "[20] Ziniu Hu, Ahmet Iscen, Chen Sun, Zirui Wang, Kai-Wei Chang, Yizhou Sun, Cordelia Schmid, David A Ross, and Alireza Fathi. Reveal: Retrieval-augmented visual-language pre-training with multi-source multimodal knowledge memory. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 1, 2, 5", + "[21] Jeff Johnson, Matthijs Douze, and Hervé Jégou. Billionscale similarity search with gpus. IEEE Transactions on Big Data, 7(3):535-547, 2019. 3", + "[22] Andrej Karpathy and Li Fei-Fei. Deep visual-semantic alignments for generating image descriptions. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2015. 2, 4", + "[23] Jacob Devlin Ming-Wei Chang Kenton and Lee Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of" + ], + "bbox": [ + 501, + 92, + 893, + 901 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "13741", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "North American Chapter of the Association for Computational Linguistics: Human Language Technologies (NAACL-HLT), 2019. 4", + "[24] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In Proc. International conference on machine learning (ICML), 2022. 5, 7", + "[25] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In Proc. International conference on machine learning (ICML), 2023. 1, 2, 3, 4, 5, 7", + "[26] Xiujun Li, Xi Yin, Chunyuan Li, Pengchuan Zhang, Xiaowei Hu, Lei Zhang, Lijuan Wang, Houdong Hu, Li Dong, Furu Wei, et al. Oscar: Object-semantics aligned pre-training for vision-language tasks. In Proc. European Conference on Computer Vision (ECCV), 2020. 2, 5", + "[27] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dálár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Proc. European Conference on Computer Vision (ECCV), 2014. 4", + "[28] Jiasen Lu, Jianwei Yang, Dhruv Batra, and Devi Parikh. Neural baby talk. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 2", + "[29] Ron Mokady, Amir Hertz, and Amit H Bermano. Clipcap: Clip prefix for image captioning. arXiv preprint arXiv:2111.09734, 2021. 2, 3, 5", + "[30] Bryan A Plummer, Liwei Wang, Chris M Cervantes, Juan C Caicedo, Julia Hockenmaier, and Svetlana Lazebnik. Flickr30k entities: Collecting region-to-phrase correspondences for richer image-to-sentence models. In Proc. IEEE International Conference on Computer Vision (ICCV), 2015. 4", + "[31] Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, Ilya Sutskever, et al. Language models are unsupervised multitask learners. OpenAI blog, 1(8):9, 2019. 2", + "[32] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In Proc. International conference on machine learning (ICML), 2021. 2", + "[33] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J Liu. Exploring the limits of transfer learning with a unified text-to-text transformer. Journal of Machine Learning Research (JMLR), 21(1):5485-5551, 2020. 2", + "[34] Rita Ramos, Desmond Elliott, and Bruno Martins. Retrievalaugmented image captioning. In Proc. Conference of the European Chapter of the Association for Computational Linguistics (EACL), 2023. 2", + "[35] Rita Ramos, Bruno Martins, Desmond Elliott, and Yova Kementchedjhieva. Smallcap: Lightweight image captioning prompted with retrieval augmentation. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 1, 2, 3, 5, 8" + ], + "bbox": [ + 78, + 90, + 468, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[36] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 3, 6", + "[37] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 2, 4", + "[38] Laurens Van der Maaten and Geoffrey Hinton. Visualizing data using t-sne. Journal of machine learning research (JMLR), 9(11), 2008. 7", + "[39] Subhashini Venugopalan, Lisa Anne Hendricks, Marcus Rohrbach, Raymond Mooney, Trevor Darrell, and Kate Saenko. Captioning images with diverse objects. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2017. 2", + "[40] Duc Minh Vo, Hong Chen, Akihiro Sugimoto, and Hideki Nakayama. Noc-rek: Novel object captioning with retrieved vocabulary from external knowledge. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2, 5", + "[41] Weihan Wang, Qingsong Lv, Wenmeng Yu, Wenyi Hong, Ji Qi, Yan Wang, Junhui Ji, Zhuoyi Yang, Lei Zhao, Xixuan Song, Jiazheng Xu, Bin Xu, Juanzi Li, Yuxiao Dong, Ming Ding, and Jie Tang. Cogvlm: Visual expert for pretrained language models. arXiv preprint arXiv:2311.03079, 2023. 4, 5", + "[42] Kelvin Xu, Jimmy Ba, Ryan Kiros, Kyunghyun Cho, Aaron Courville, Ruslan Salakhudinov, Rich Zemel, and Yoshua Bengio. Show, attend and tell: Neural image caption generation with visual attention. In Proc. International conference on machine learning (ICML). 2", + "[43] Zhuolin Yang, Wei Ping, Zihan Liu, Vijay Korthikanti, Weili Nie, De-An Huang, Linxi Fan, Zhiding Yu, Shiyi Lan, Bo Li, et al. Re-ViLM: Retrieval-augmented visual language model for zero and few-shot image captioning. In Findings of the Association for Computational Linguistics: EMNLP, 2023. 2", + "[44] Michihiro Yasunaga, Armen Aghajanyan, Weijia Shi, Richard James, Jure Leskovec, Percy Liang, Mike Lewis, Luke Zettlemoyer, and Wen-tau Yih. Retrieval-augmented multimodal language modeling. In International Conference on Learning Representations (ICLR), 2023. 2", + "[45] Pengchuan Zhang, Xiujun Li, Xiaowei Hu, Jianwei Yang, Lei Zhang, Lijuan Wang, Yejin Choi, and Jianfeng Gao. Vinvl: Revisiting visual representations in vision-language models. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 2, 5", + "[46] Deyao Zhu, Jun Chen, Xiaoqian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models. arXiv preprint arXiv:2304.10592, 2023. 3, 4, 5, 7, 8" + ], + "bbox": [ + 501, + 92, + 890, + 852 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "13742", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2024/EVCap_ Retrieval-Augmented Image Captioning with External Visual-Name Memory for Open-World Comprehension/003cc44f-e449-40f1-aec0-baf1c6235010_model.json b/2024/EVCap_ Retrieval-Augmented Image Captioning with External Visual-Name Memory for Open-World Comprehension/003cc44f-e449-40f1-aec0-baf1c6235010_model.json new file mode 100644 index 0000000000000000000000000000000000000000..bea0f284d863f259d7c390526030e85ff0839c4a --- /dev/null +++ b/2024/EVCap_ Retrieval-Augmented Image Captioning with External Visual-Name Memory for Open-World Comprehension/003cc44f-e449-40f1-aec0-baf1c6235010_model.json @@ -0,0 +1,2420 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.134, + 0.131, + 0.838, + 0.177 + ], + "angle": 0, + "content": "EVCAP: Retrieval-Augmented Image Captioning with External Visual-Name Memory for Open-World Comprehension" + }, + { + "type": "text", + "bbox": [ + 0.199, + 0.203, + 0.777, + 0.221 + ], + "angle": 0, + "content": "Jiaxuan \\(\\mathrm{Li}^{1*}\\), Duc Minh \\(\\mathrm{Vo}^{1*}\\), Akihiro Sugimoto\\(^{2}\\), Hideki Nakayama\\(^{1}\\)" + }, + { + "type": "text", + "bbox": [ + 0.195, + 0.222, + 0.776, + 0.24 + ], + "angle": 0, + "content": "1The University of Tokyo, Japan 2National Institute of Informatics, Japan" + }, + { + "type": "text", + "bbox": [ + 0.12, + 0.242, + 0.851, + 0.258 + ], + "angle": 0, + "content": "{li, vm Duc}@nlab.ci.i.u-tokyo.ac.jp sugimoto@nii.ac.jp nakayama@ci.i.u-tokyo.ac.jp" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.292, + 0.314, + 0.308 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.324, + 0.474, + 0.655 + ], + "angle": 0, + "content": "Large language models (LLMs)-based image captioning has the capability of describing objects not explicitly observed in training data; yet novel objects occur frequently, necessitating the requirement of sustaining up-to-date object knowledge for open-world comprehension. Instead of relying on large amounts of data and/or scaling up network parameters, we introduce a highly effective retrieval-augmented image captioning method that prompts LLMs with object names retrieved from External Visual-name memory (EVCAP). We build ever-changing object knowledge memory using objects' visuals and names, enabling us to (i) update the memory at a minimal cost and (ii) effortlessly augment LLMs with retrieved object names by utilizing a lightweight and fast-to-train model. Our model, which was trained only on the COCO dataset, can adapt to out-of-domain without requiring additional fine-tuning or re-training. Our experiments conducted on benchmarks and synthetic commonsense-violating data show that EVCAP, with only 3.97M trainable parameters, exhibits superior performance compared to other methods based on frozen pre-trained LLMs. Its performance is also competitive to specialist SOTAs that require extensive training." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.684, + 0.21, + 0.7 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.71, + 0.47, + 0.86 + ], + "angle": 0, + "content": "Advanced image captioning based on large language models (LLMs) [3, 8, 9, 25] has focused on the approach using big-scale models trained on ever-increasingly large-scale datasets, which is no longer viable. This is because the computational cost to train the models increases exponentially and, more importantly, updating training data is almost impossible to keep pace with the growth of novel objects in our daily lives. Sustaining ever-changing object knowledge with a reasonable cost is a pressing concern in LLMs-based models to truly unlock open-world comprehension." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.861, + 0.47, + 0.877 + ], + "angle": 0, + "content": "Retrieval-augmented image captioning [20, 35] is" + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.292, + 0.891, + 0.543 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.554, + 0.895, + 0.753 + ], + "angle": 0, + "content": "Figure 1. Overall comparison of our EVCAP and SOTAs. (Upper) Generated captions by SmallCap, BLIP-2, and our EVCAP for a commonsense-violating image from the WHOOPS dataset. \\(\\times\\) and \\(\\checkmark\\) indicate incorrect and correct predictions, respectively. Incorrect objects in captions are highlighted in red, while correct ones are in blue. SmallCap and BLIP-2 give incorrect predictions for \"ice skates\" and \"wood floor\", respectively, while our EVCAP utilizes an external visual-name memory to enhance attention to objects within the image, leading to superior performance for image captioning. (Lower) Comparison of the number of trainable parameters, CIDEr score on COCO and NoCaps datasets. The size of each circle reflects the log number of trainable parameters. EVCAP (3.97M) has less trainable parameters than others while achieving comparable results with SOTAs at scale." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.781, + 0.895, + 0.903 + ], + "angle": 0, + "content": "emerging as an alternative since it considerably reduces training costs in both time and data while producing encouraging results. Nonetheless, with their huge datastore, it is obvious that LLMs would imitate the given texts, limiting their ability to describe open-world objects properly. For instance, SmallCap [35] considers the words \"skateboard\" and \"wooden floor\" to be a pair regardless of visual appearances containing a commonsense-violating pair" + }, + { + "type": "page_footnote", + "bbox": [ + 0.101, + 0.886, + 0.475, + 0.897 + ], + "angle": 0, + "content": "*Equal contributions. Code is available at https://jiaxuan-li.github.io/EVCap." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13733" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.317 + ], + "angle": 0, + "content": "of \"ice skates\" and \"wood floor\" (Fig. 1, upper). Additionally, prompting the LLMs given a lot of retrieved texts becomes cumbersome, requiring more trainable parameters. Fig. 1 (lower) shows that the CIDEr scores obtained by a lightweight SmallCap [35] with 43M trainable parameters are far away from those obtained by a heavy REVEAL [20] with 2.1B trainable parameters. Beyond that, due to the frequent occurrence of new objects, access to their sample texts is not always feasible, making the memory utilized in [20, 35] difficult to grow. We thus aim to streamline the external memory used in previous work [20, 35] by storing a sufficiently small amount of object information. And, of course, not only does the model not stereotype the example sentences, but the number of trainable parameters would be reduced drastically as a result of the causation (Fig. 1)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.318, + 0.473, + 0.692 + ], + "angle": 0, + "content": "We follow [13, 40] to construct a key-value memory where the key is represented by object's features, and the value corresponds to object's name. Unlike [13, 40], which rely on object definition as the key, our method leverages the visual appearance of the object as the key because of the abundance of object images readily available on the internet. We propose an external visual-name memory tailored for ease of expansion and cost-effectiveness in upholding up-to-date object information. We present a highly effective retrieval-augmented LLMs-based image captioning method, called EVCAP, that prompts frozen LLMs with object names retrieved from our proposed memory for open-world comprehension. EVCAP contains a frozen image encoder ViT [14] and Q-Former [25] with trainable image query tokens for object retrieval, an attentive fusion module, a trainable linear layer for mapping between vision and language latent spaces, and a frozen LLM decoder [10] for generating captions. Specifically, the attentive fusion module feeds retrieved object names and visual features into a customized frozen Q-Former using trainable object name query tokens to implicitly reduce the presence of superfluous object names. As a result, EVCAP amounts to only 3.97M trainable parameters. Once trained, the model can be adapted to new domains and large-scale data without further fine-tuning or re-training. Our contributions are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.694, + 0.47, + 0.739 + ], + "angle": 0, + "content": "- We provide an extensible external visual-name memory with minimal but useful object information, which enables LLMs-based models to comprehend the open world." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.74, + 0.469, + 0.77 + ], + "angle": 0, + "content": "- We present a highly efficacious retrieval-augmented image captioning EVCAP with 3.97M trainable parameters." + }, + { + "type": "list", + "bbox": [ + 0.077, + 0.694, + 0.47, + 0.77 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.772, + 0.469, + 0.833 + ], + "angle": 0, + "content": "On in-/out-domain benchmarks and synthetic commonsense-violating dataset, EVCAP trained solely on COCO dataset competes with other lightweight methods by a margin while being on par with other specialist SOTAs." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.846, + 0.22, + 0.862 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.871, + 0.47, + 0.903 + ], + "angle": 0, + "content": "Image captioning aims to describe the contents of a given image. It can be roughly divided into two approaches: non-" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.331 + ], + "angle": 0, + "content": "LLMs-based methods and LLMs-based ones. The former approaches [4, 22, 42] typically employ a visual encoder and a language decoder in an end-to-end fashion to generate captions. However, they are incapable of describing open-world objects. The latter one leverages pre-trained large-scale vision models (CLIP [32], ViT [12]) and LLMs (GPTs [7, 31], T5 [33], LLaMA [37]) by bridging the gap between two modalities using either pre-training with largescale data or the learned mapper or prompt techniques. LLMs-based models [8, 9, 25, 29] demonstrate advancements in image captioning challenges, allowing the capacity to describe anything as long as pre-trained vision models can recognize it. Our method belongs to the LLMs-based approaches, but instead of relying fully on the pre-trained vision model, we use object names retrieved from the external memory to augment LLMs-based image captioning." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.332, + 0.893, + 0.542 + ], + "angle": 0, + "content": "Novel object captioning is a branch of image captioning that describes images containing objects that were not seen during training. Non-LLMs-based methods explore more objects by learning from unpaired image-sentence sources (DCC [19], NOC [39]) or rely on novel object detectors to recognize novel concepts (NBT [28], OSCAR [26] and VinVL [45]). LLMs-based methods such as ViECap [15] leverage the pre-trained CLIP [32] to obtain object entities. Nevertheless, the cut-off in training time of the pre-trained object detector or CLIP prevents it from detecting novel objects that arise quickly in reality. Unlike earlier work, we can readily update our recognition of novel concepts by adding them to external memory, ensuring that we keep any new objects from the past and even the future." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.542, + 0.893, + 0.768 + ], + "angle": 0, + "content": "Retrieval-augmented image captioning is a recently popular approach that augments the captioning model with retrieved information for better open-world understanding. AoANet [16] uses a memory bank of image-sentence pairs and target words. SmallCap [35] employs image-to-text retrieval to obtain sampled captions from a captions dataset. RA-CM3 [44] retrieves documents from an external memory of a mixture of text and image via a dense multimodal retriever. EXTRA [34] and Re-ViLM [43] exploit the similarity of the input image and vision candidates to retrieve captions. Different from the previous methods, our external memory contains visual-name pairs to avoid redundant information in the external captions/documents. In addition, we use an attentive fusion module to mitigate the effects of irrelevant retrieved object names on caption generation." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.778, + 0.676, + 0.795 + ], + "angle": 0, + "content": "3. Proposed EVCAP" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.803, + 0.65, + 0.818 + ], + "angle": 0, + "content": "3.1. Idea of EVCAP" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.826, + 0.892, + 0.903 + ], + "angle": 0, + "content": "We aim to build a retrieval-augmented LLMs-based image captioning model with a sufficiently small yet informative external memory. It involves two challenges: (1) constructing an expandable external memory and (2) building an effective LLMs-based model using retrieved object names." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "13734" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.118, + 0.087, + 0.852, + 0.391 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.4, + 0.896, + 0.485 + ], + "angle": 0, + "content": "Figure 2. Schematic of our proposed EVCAP. It consists of an external visual-name memory with image embeddings and object names (upper), a frozen ViT and Q-Former equipped with trainable image query tokens, an attentive fusion module developed by a customized frozen Q-Former and trainable object name query tokens, and a frozen LLM with a trainable linear layer (lower). The ViT and Q-Former extract learned visual features from the input image, which are then used to retrieve object names from the external memory. These retrieved object names and learned visual features undergo cross-attention in the customized Q-Former, creating refined object name features. Finally, the object name features combined with visual features are fed into the LLM post a linear layer for generating captions." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.51, + 0.47, + 0.659 + ], + "angle": 0, + "content": "As discussed above, challenge (1) can be resolved by utilizing the visual appearance of objects. However, if we restrict our memory to only a visual-name pair for each object, our memory will be lacking in diversity. Therefore, we gather several images for each target object. Additionally, we keep the synthetic images in our memory to avoid the harm that synthetic images might cause to our method, as pointed out in [18]. With the capability to collect images from the internet, EVCAP can be easily expanded to include novel objects from the real world effortlessly." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.66, + 0.471, + 0.841 + ], + "angle": 0, + "content": "We base our method on the frozen pre-trained vision model and LLM with several trainable layers (Fig. 2), giving in a model that is cheap to train. To guide the LLM, we adopt a recently popular approach called prompting as in [11, 25, 29, 35, 46]. We begin by matching the learned visual features from the input image with image embeddings stored in memory, retrieving object names. We also introduce an attentive fusion module designed to implicitly remove irrelevant retrieved names. Finally, following the attentive fusion, we combine the learned visual features and object name features to form a prompt for the LLM to generate a caption, thus addressing challenge (2)." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.848, + 0.35, + 0.865 + ], + "angle": 0, + "content": "3.2. External visual-name memory" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.871, + 0.471, + 0.903 + ], + "angle": 0, + "content": "To build the external visual-name memory, we first collect image-name pairs from the external data source. After" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.51, + 0.892, + 0.54 + ], + "angle": 0, + "content": "that, we encode these images into image embeddings, which serve as keys in memory, and use their names as values." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.541, + 0.893, + 0.766 + ], + "angle": 0, + "content": "External data source. We utilize object images from LVIS dataset [17] to construct our external visual-name memory \\(\\mathcal{M}\\). Specifically, we use 1203 objects in LVIS, where we randomly select from one to ten images for each object, amounting to 8581 object images. Furthermore, as mentioned in Sec. 3.1, we also incorporate synthetic images in our memory construction. Using stable diffusion [36], we generate five additional images for each object, with a prompt of \"a photo of {object name}\", resulting in a total of \\(M = 14596\\) (\\(8581 + 5 \\times 1203\\)) images. Each object image \\(X^i\\) is associated with an object name \\(v^i\\). Note that many object images may share the same object name. For the sake of simplicity, we may regard each image as corresponding to a single name. In summary, we have \\(M\\) image-name pairs \\(\\{(X^i,v^i)\\}_{i=1}^M\\) for external memory construction." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.767, + 0.895, + 0.902 + ], + "angle": 0, + "content": "External memory construction. For each image \\(X^i\\), we use a frozen vision encoder \\(\\mathcal{E}(\\cdot)\\) (see Sec. 3.3 for detail) to project it into 32 embeddings with the size of \\(1 \\times 768\\) each: \\(\\{\\mathbf{k}_1^i, \\mathbf{k}_2^i, \\dots, \\mathbf{k}_{32}^i\\} = \\mathcal{E}(X^i)\\). We then average 32 embeddings to produce a single embedding \\(\\mathbf{k}^i\\) (\\(1 \\times 768\\)) that serves as the key (visual) in \\(\\mathcal{M}\\). The paired object name \\(v^i\\) acts as its value (name). Consequently, we have the visual-name memory \\(\\mathcal{M} = \\{(\\mathbf{k}^i, v^i)\\}_{i=1}^M\\) which is indexed using FAISS [21], facilitating rapid searches based on sim" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13735" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.122 + ], + "angle": 0, + "content": "ilarity measures. Our memory can be expanded effortlessly by gathering additional visual-name pairs (see Sec. 5.3)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.13, + 0.29, + 0.145 + ], + "angle": 0, + "content": "3.3. Object names retrieval" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.153, + 0.468, + 0.303 + ], + "angle": 0, + "content": "Image encoding. We feed a frozen vision encoder \\(\\mathcal{E}\\) image \\(X\\) and image query tokens \\(\\mathbf{T}_{\\mathrm{img}}\\) to produce visual features \\(\\mathcal{Q}\\). To enable the retrieval process controllable, we make image query tokens to be trainable. Thus, the image encoding process can be summarized as \\(\\mathcal{Q} = \\mathcal{E}(X,\\mathbf{T}_{\\mathrm{img}})\\). We use the BLIP-2 pre-trained vision encoder [25], which consists of a pre-trained vision transformer ViT-g [14] outputting image features \\((257\\times 1408)\\), and a Q-Former receiving image features producing \\(|\\mathcal{Q}| = 32\\) learned visual features \\((1\\times 768\\) each). We denote \\(\\mathcal{Q} = \\{\\mathbf{q}_1,\\mathbf{q}_2,\\dots,\\mathbf{q}_{32}\\}\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.304, + 0.468, + 0.402 + ], + "angle": 0, + "content": "Retrieval. Having obtained \\(\\mathcal{Q}\\), we calculate the cosine similarity between the query \\(\\mathbf{q}_j \\in \\mathcal{Q}\\) and the key \\(\\mathbf{k}^i \\in \\mathcal{M}\\). The similarity calculation is given by \\(\\mathrm{SIM}(\\mathbf{q}_j, \\mathbf{k}^i) = \\frac{\\mathbf{q}_j^\\top \\mathbf{k}^i}{\\|\\mathbf{q}_j\\| \\|\\mathbf{k}^i\\|}\\), where \\(i \\in [1, M]\\), \\(j \\in [1, 32]\\). Given each \\(\\mathbf{q}_j\\), we select one key with the highest similarity score, resulting in 32 key-value candidates \\(\\{\\mathbf{k}_j^{\\mathrm{best}}, v_j^{\\mathrm{best}}\\}_{j=1}^{32}\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.402, + 0.468, + 0.491 + ], + "angle": 0, + "content": "After that, we filter out candidates with repeated object names (values), and then select the top-K values. In particular, we determine the index \\( j \\) from the key that has the highest SIM score. These selected values \\( v_{j}^{\\mathrm{best}} \\) are redefined as the new notation \\( v_{l} \\) in the retrieved top-K object names for the input image, which can be summarized as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.136, + 0.504, + 0.407, + 0.528 + ], + "angle": 0, + "content": "\\[\n\\{\\mathbf {k} _ {j} ^ {\\text {b e s t}}, v _ {j} ^ {\\text {b e s t}} \\} = \\arg \\max _ {\\mathbf {k} ^ {i}} \\operatorname {S I M} \\left(\\mathbf {q} _ {j}, \\mathbf {k} ^ {i}\\right),\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.134, + 0.54, + 0.412, + 0.565 + ], + "angle": 0, + "content": "\\[\nj = \\arg \\max _ {j} \\mathrm {S I M} (\\mathbf {q} _ {j}, \\mathbf {k} _ {j} ^ {\\mathrm {b e s t}}), v _ {l} \\gets v _ {j} ^ {\\mathrm {b e s t}},\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.569, + 0.468, + 0.602 + ], + "angle": 0, + "content": "where \\(l \\in [1, \\mathrm{K}]\\). As a result, the retrieved top-\\(\\mathbf{K}\\) object names are \\(\\{v_l\\}_{l=1}^{\\mathrm{K}}\\)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.608, + 0.236, + 0.623 + ], + "angle": 0, + "content": "3.4. Attentive fusion" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.631, + 0.468, + 0.675 + ], + "angle": 0, + "content": "Since the object names obtained from the retrieval process may be redundant, we develop an attentive fusion module to selectively distill object name features." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.676, + 0.468, + 0.902 + ], + "angle": 0, + "content": "The retrieved object names \\(\\{v_{l}\\}_{l = 1}^{\\mathrm{K}}\\) are concatenated together into a sequence \\(S\\), each separated by a delimiter: \\(S = \\{v_{1},[\\mathrm{SEP}],v_{2},[\\mathrm{SEP}],\\dots ,[\\mathrm{SEP}],v_{\\mathrm{K}}\\}\\). The sequence \\(S\\) and visual features \\(\\mathcal{Q}\\) are fed into a customized Q-Former \\(\\mathcal{F}(\\cdot)\\), which is constructed from the frozen pretrained Q-Former as we used in vision encoder \\(\\mathcal{E}\\). Nonetheless, in order to enable object names to get attention from visual features, we switch the image embedding port and the text instruction port (see the supplement for architecture detail). Like in the image encoding process in Sec. 3.3, we make the object name query tokens \\(\\mathbf{T}_{\\mathrm{obj}}\\) learnable during training to assist in learning object name features related to the caption. The size of \\(\\mathbf{T}_{\\mathrm{obj}}\\) is \\(P\\times 768\\), where \\(P\\) indicates the number of object name query tokens. We get the object name features \\(\\mathcal{V} = \\mathcal{F}(S,\\mathcal{Q},\\mathbf{T}_{\\mathrm{obj}})\\)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.091, + 0.685, + 0.107 + ], + "angle": 0, + "content": "3.5. Caption generation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.115, + 0.892, + 0.234 + ], + "angle": 0, + "content": "Before inputting the visual features \\(\\mathcal{Q}\\) and object name features \\(\\nu\\) into the LLM decoder, we concatenate \\((\\oplus)\\) them and use a linear layer \\(\\phi(\\cdot)\\) to project them into the input latent space of the LLM as \\(\\phi(\\mathcal{Q} \\oplus \\mathcal{V})\\). The LLM used for caption generation in this work is the pre-trained Vicuna-13B [10], an open-source chatbot constructed from LLaMA [37]. During training and evaluation, we design a prompt in a conversational format, that is similar to [46]:" + }, + { + "type": "text", + "bbox": [ + 0.526, + 0.246, + 0.858, + 0.291 + ], + "angle": 0, + "content": "Human: Describe this image in detail. \nAssistant:" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.301, + 0.892, + 0.452 + ], + "angle": 0, + "content": "in which, ProjFeature denotes the projected feature \\(\\phi (\\mathcal{Q}\\oplus \\mathcal{V})\\) after the linear layer. In training phase, given input caption tokens \\(\\{c_i\\}_{i = 1}^L\\), the LLM decoder concatenates the embedded prompt \\(\\{\\mathbf{w}_i\\}_{i = 1}^N\\) and the embedded caption tokens \\(\\{\\mathbf{c}_i\\}_{i = 1}^L\\) as input, and predicts the caption tokens in an autoregressive fashion, while in the evaluation phase, we only need to input the embedded prompt. We train EV-CAP by minimizing the cross-entropy loss in an end-to-end way: \\(\\mathcal{L}_{\\theta} = -\\sum_{i = 1}^{L}\\log p_{\\theta}(c_i|\\mathbf{w}_1,\\dots \\mathbf{w}_N,\\mathbf{c}_1,\\dots ,\\mathbf{c}_{i - 1})\\), in which \\(\\theta\\) indicates the trainable parameters." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.468, + 0.71, + 0.485 + ], + "angle": 0, + "content": "4. Experimental Settings" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.494, + 0.649, + 0.51 + ], + "angle": 0, + "content": "4.1. Training setup" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.518, + 0.892, + 0.638 + ], + "angle": 0, + "content": "Implementation. EVCAP uses the same image encoder as in BLIP-2 [25], consisting of a ViT-g [14] and their pretrained Q-Former. Since we intend to obtain object name features through cross-attention between retrieved object names and visual features, we develop a customized Q-Former, which consists of BERT [23] with cross-attention layers inserted at every other transformer block. We use a frozen Vicuna-13B [10] as the caption generator." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.639, + 0.892, + 0.715 + ], + "angle": 0, + "content": "Training dataset. For all experiments, we exclusively train EVCAP using the training set of COCO dataset [27], consisting of 82k images and 5 captions per images. The entire training process takes about 3 hours on 4 A6000 GPUs, using mixed precisions (more details in the supplementary)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.727, + 0.666, + 0.743 + ], + "angle": 0, + "content": "4.2. Evaluation setup" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.751, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Evaluation dataset. We evaluate EVCAP, trained using the COCO training set, across four datasets: its test set, two challenging benchmarks - NoCaps validation set and Flickr30k test set, and a synthetic commonsense-violating dataset - WHOOPS. We adhere to follow prior work [15, 41] to use the same images of Karpathy split [22] on COCO test set, NoCaps [2] validation set, and Karpathy split on Flickr30k [30] test set. In addition, WHOOPS [6] is a synthetic image captioning dataset comprising 500 synthetic commonsense-violating images and 2500 paired captions." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13736" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.09, + 0.892, + 0.146 + ], + "angle": 0, + "content": "Table 1. Quantitative comparison against SOTA methods on three common image captioning benchmarks. * denotes using a memory bank. We report the size of training data and parameters; BLEU@4 (B@4), METEOR (M), CIDEr (C), and SPICE (S) scores on COCO test set; C and S scores on in-domain, near-domain, out-domain and overall data of NoCaps validation set; C and S scores on Flickr30k test set. Higher score is better. Bold indicates the best results among compared methods, normal indicates the second best results." + }, + { + "type": "table", + "bbox": [ + 0.081, + 0.152, + 0.892, + 0.428 + ], + "angle": 0, + "content": "
MethodTrainingCOCONoCaps valFlickr30k Test
DataPara.TestIn-domainNear-domainOut-domainOverallCS
B@4MCSCSCSCSCSCS
Heavyweight-training models
VinVL [45]8.9M110M38.230.3129.323.696.813.590.713.187.411.690.912.8--
AoANet+MA* [16]COCO-38.028.7121.021.8----------
NOC-REK* [40]COCO110M----104.714.8100.214.1100.713.0100.914.0--
RCA-NOC* [13]COCO110M37.429.6128.423.192.212.987.812.687.511.588.312.4--
ViEcap GPT2 [15]COCO124M27.224.892.918.261.110.464.39.965.08.666.29.547.913.6
InstructBLIP Vicuna-13B [11]129M188M----------121.9-82.8-
OSCAR [26]4.1M338M37.430.7127.823.583.412.081.612.077.610.681.111.7--
BLIP [24]129M446M40.4-136.7-114.915.2112.114.9115.314.4113.214.8--
BLIP-2 FlanT5-XL [25]129M1.2B42.4-144.5-123.716.3120.215.9124.815.1121.615.8--
REVEAL* T5 [20]1.3B2.1B--145.4-------123.0---
Lightweight-training models
MiniGPT4 Vicuna-13B [46]5M3.94M38.029.6129.623.499.014.8106.915.3110.814.9108.815.178.416.9
SmallCap* GPT2 [35]COCO7M37.027.9119.721.3--------60.6-
ClipCap GPT2 [29]COCO43M33.527.5113.121.184.912.166.810.949.19.665.810.9--
EVCAP* Vicuna-13BCOCO3.97M41.531.2140.124.7111.715.3119.515.6116.514.7119.315.384.418.0
Specialist SOTAs
Qwen-VL Qwen-7B [5]1.4B9.6B----------121.4-85.8-
CogVLM Vicuna-7B [41]1.5B6.5B--148.7-----132.6-128.3-94.9-
PaLI mT5-XXL [9]1.6B17B--149.1-------127.0---
PaLI-X_UL2-32B [8]2.2B55B--149.2-------126.3---
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.452, + 0.473, + 0.634 + ], + "angle": 0, + "content": "Compared methods. We compare EVCAP with several SOTAs. According to the trainable parameters size, they can be divided into 1) Heavyweight-training (between 100M to 5B): VinVL [45], AoANet [16], NOC-REK [40], RCA-NOC [13], ViECap [15], InstructBLIP [11], OSCAR [26], BLIP [24], BLIP-2 [25], REVEAL [20]; 2) Lightweight-training (less than 100M): MiniGPT4 [46], SmallCap [35], ClipCap [29]; and also 3) Specialist SOTAs with huge trainable parameters (larger than 5B): QwenVL [5], CogVLM [41], PaLI [9], PaLI-X [8]. Among these methods, AoANet, NOC-REK, RCA-NOC, REVEAL, and SmallCap are retrieval-augmented captioning methods." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.652, + 0.283, + 0.669 + ], + "angle": 0, + "content": "5. Experimental Results" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.679, + 0.411, + 0.694 + ], + "angle": 0, + "content": "5.1. Results on in-/out-domain benchmarks" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.704, + 0.47, + 0.779 + ], + "angle": 0, + "content": "We assess EVCAP against SOTAs on both in-domain and out-domain benchmarks. The COCO test set can be considered as in-domain data as we only train our model on the COCO training set. Out-domain benchmarks are the NoCaps validation set and the Flickr30k test set." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.781, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Quantitative results. Tab. 1 details our EVCAP's performance in comparison with SOTA methods. We first evaluate training costs in terms of training data sizes and parameters. Similar to various heavyweight-training models that exclude LLMs and the majority of lightweight-training models, EVCAP is trained solely on the COCO training set. It utilizes only 3.97M trainable parameters, positioning it as the second smallest, slightly larger than MiniGPT4" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.452, + 0.892, + 0.813 + ], + "angle": 0, + "content": "with 3.94M. Among lightweight-training models, our approach outperforms others, achieving the highest scores on all benchmarks. Despite using less training data and nearly identical trainable parameters as MiniGPT4, EVCAP significantly surpasses it, with a marked improvement of 10.5, 10.5, and 6.0 in CIDEr scores for each benchmark. When further compared with heavyweight-training models, the performance of EVCAP stands out among million-level models, nearly matching InstructBLIP, except in NoCaps. Note that since BLIP-2 does not include Vicuna checkpoints, InstructBLIP performs pre-training with Vicuna using the same procedure as BLIP-2, whereas EVCAP does not involve pre-training. Against REVEAL, which also uses external memory, our EVCAP utilizes about 1/3000 training data and 1/500 training parameters yet yields comparable results. Moreover, EVCAP's performance is on par with BLIP-2, the top-performing model with 1.2B trainable parameters. This highlights EVCAP's efficiency and effectiveness despite its significantly smaller training cost, thanks to our external visual-name memory. Regarding specialist SOTAs, they use billion-level training data and over 5B trainable parameters, so it is acceptable that they can achieve exceptionally strong performance, surpassing EVCAP by nearly 10 on all benchmarks in CIDEr scores." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.826, + 0.892, + 0.903 + ], + "angle": 0, + "content": "Qualitative results. Fig. 3 presents a comparison of captions generated by our EVCAP and three SOTA models across three benchmarks. The captions of SmallCap are generated by its publicly accessible demo [1]. We generate captions of MiniGPT4 and BLIP-2 using their respec" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13737" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.207, + 0.093, + 0.282, + 0.104 + ], + "angle": 0, + "content": "COCO Test" + }, + { + "type": "image", + "bbox": [ + 0.14, + 0.107, + 0.347, + 0.199 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.137, + 0.201, + 0.35, + 0.223 + ], + "angle": 0, + "content": "GT: A green bus driving through a rural area with trees in the background." + }, + { + "type": "text", + "bbox": [ + 0.137, + 0.225, + 0.35, + 0.247 + ], + "angle": 0, + "content": "SmallCap: A bus driving down a street next to trees." + }, + { + "type": "text", + "bbox": [ + 0.137, + 0.249, + 0.32, + 0.272 + ], + "angle": 0, + "content": "MiniGPT4: A green bus is driving down the street." + }, + { + "type": "text", + "bbox": [ + 0.137, + 0.274, + 0.336, + 0.297 + ], + "angle": 0, + "content": "BLIP-2: A green bus driving down a road with trees in the background." + }, + { + "type": "text", + "bbox": [ + 0.137, + 0.299, + 0.333, + 0.321 + ], + "angle": 0, + "content": "EVCap: A green bus driving down a road next to trees." + }, + { + "type": "image", + "bbox": [ + 0.141, + 0.325, + 0.348, + 0.418 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.137, + 0.42, + 0.317, + 0.442 + ], + "angle": 0, + "content": "GT: A woman in a blue top with headphones and two cellphones." + }, + { + "type": "text", + "bbox": [ + 0.137, + 0.444, + 0.343, + 0.467 + ], + "angle": 0, + "content": "SmallCap: A woman sitting in front of a laptop computer." + }, + { + "type": "text", + "bbox": [ + 0.137, + 0.469, + 0.316, + 0.492 + ], + "angle": 0, + "content": "MiniGPT4: A woman sitting on a couch holding two phones." + }, + { + "type": "text", + "bbox": [ + 0.137, + 0.494, + 0.334, + 0.517 + ], + "angle": 0, + "content": "BLIP-2: A woman sitting on a couch with two cell phones." + }, + { + "type": "text", + "bbox": [ + 0.137, + 0.518, + 0.339, + 0.541 + ], + "angle": 0, + "content": "EVCap: A woman wearing headphones holding two cell phones." + }, + { + "type": "title", + "bbox": [ + 0.447, + 0.093, + 0.526, + 0.105 + ], + "angle": 0, + "content": "NoCaps Val" + }, + { + "type": "image", + "bbox": [ + 0.383, + 0.107, + 0.591, + 0.199 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.379, + 0.201, + 0.575, + 0.222 + ], + "angle": 0, + "content": "GT: The two guinea pigs are getting dried off in a yellow towel." + }, + { + "type": "text", + "bbox": [ + 0.379, + 0.224, + 0.578, + 0.247 + ], + "angle": 0, + "content": "SmallCap: A person holding a small animal in a towel." + }, + { + "type": "text", + "bbox": [ + 0.379, + 0.249, + 0.566, + 0.272 + ], + "angle": 0, + "content": "MiniGPT4: Two small animals are wrapped in a towel." + }, + { + "type": "text", + "bbox": [ + 0.379, + 0.274, + 0.589, + 0.297 + ], + "angle": 0, + "content": "BLIP-2: Two guinea pigs wrapped in a yellow towel." + }, + { + "type": "text", + "bbox": [ + 0.379, + 0.299, + 0.588, + 0.322 + ], + "angle": 0, + "content": "EVCap: Two guinea pigs are wrapped in a yellow towel." + }, + { + "type": "image", + "bbox": [ + 0.383, + 0.325, + 0.591, + 0.418 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.378, + 0.42, + 0.578, + 0.442 + ], + "angle": 0, + "content": "GT: A computer screen showing two men sitting at a table." + }, + { + "type": "text", + "bbox": [ + 0.378, + 0.444, + 0.584, + 0.467 + ], + "angle": 0, + "content": "SmallCap: Two men sitting at a table with a laptop." + }, + { + "type": "text", + "bbox": [ + 0.378, + 0.469, + 0.588, + 0.491 + ], + "angle": 0, + "content": "MiniGPT4: A laptop computer sitting on top of a table." + }, + { + "type": "text", + "bbox": [ + 0.378, + 0.493, + 0.562, + 0.516 + ], + "angle": 0, + "content": "BLIP-2: A laptop computer with a picture of two men on it." + }, + { + "type": "text", + "bbox": [ + 0.378, + 0.518, + 0.562, + 0.54 + ], + "angle": 0, + "content": "EVCap: A laptop computer with a picture of two men on the screen." + }, + { + "type": "title", + "bbox": [ + 0.681, + 0.093, + 0.774, + 0.104 + ], + "angle": 0, + "content": "Flickr30k Test" + }, + { + "type": "image", + "bbox": [ + 0.623, + 0.107, + 0.831, + 0.199 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.619, + 0.201, + 0.805, + 0.223 + ], + "angle": 0, + "content": "GT: A very young child in a denim baseball cap eats a green apple." + }, + { + "type": "text", + "bbox": [ + 0.619, + 0.224, + 0.812, + 0.247 + ], + "angle": 0, + "content": "SmallCap: A young boy holding an apple in his hand." + }, + { + "type": "text", + "bbox": [ + 0.619, + 0.249, + 0.808, + 0.273 + ], + "angle": 0, + "content": "MiniGPT4: A baby sitting in a high chair eating an apple." + }, + { + "type": "text", + "bbox": [ + 0.619, + 0.274, + 0.826, + 0.297 + ], + "angle": 0, + "content": "BLIP-2: A baby sitting in a white chair eating a green apple." + }, + { + "type": "text", + "bbox": [ + 0.619, + 0.299, + 0.832, + 0.322 + ], + "angle": 0, + "content": "EVCap: A toddler eating a green apple while wearing a hat." + }, + { + "type": "image", + "bbox": [ + 0.623, + 0.326, + 0.831, + 0.419 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.619, + 0.42, + 0.823, + 0.442 + ], + "angle": 0, + "content": "GT: Two men are riding on a wooden vehicle pulled by two donkeys." + }, + { + "type": "text", + "bbox": [ + 0.619, + 0.444, + 0.833, + 0.466 + ], + "angle": 0, + "content": "SmallCap: A donkey pulling a cart with a man in the background." + }, + { + "type": "text", + "bbox": [ + 0.619, + 0.468, + 0.797, + 0.492 + ], + "angle": 0, + "content": "MiniGPT4: Two men riding on a donkey in the dirt." + }, + { + "type": "text", + "bbox": [ + 0.619, + 0.494, + 0.832, + 0.516 + ], + "angle": 0, + "content": "BLIP-2: Two men riding a horse drawn cart through a field." + }, + { + "type": "text", + "bbox": [ + 0.619, + 0.518, + 0.799, + 0.541 + ], + "angle": 0, + "content": "EVCap: Two men riding in a cart pulled by two donkeys." + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.555, + 0.894, + 0.598 + ], + "angle": 0, + "content": "Figure 3. Examples of captions generated by our EVCAP and three SOTA methods on COCO test set, NoCaps validation set, and Flickr30k test set. GT refers to the Ground Truth captions. Incorrect objects in captions are highlighted in red, while correct ones are in blue. Our EVCAP correctly generates captions across different datasets, showing performance comparable to BLIP-2." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.625, + 0.47, + 0.79 + ], + "angle": 0, + "content": "tive pre-trained models. As a lightweight and retrieval-augmented captioning method, SmallCap struggles to produce accurate captions for given images, primarily because it relies on retrieved captions laden with extraneous information. MiniGPT4, though aligned with the primary content of images, sometimes misses certain objects like \"trees\" and \"headphones\". This oversight stems from its focus on the main objects in images, without integrating additional cues for other objects provided by the retrieved object names. In contrast, the captions generated by our EVCAP are comparable to those of BLIP-2." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.801, + 0.414, + 0.817 + ], + "angle": 0, + "content": "5.2. Results on commonsense-violating data" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.469, + 0.87 + ], + "angle": 0, + "content": "To explore our EVCAP's capability in describing contents in open-word settings, we further evaluate it on WHOOPS dataset, which contains commonsense-violating images." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.871, + 0.469, + 0.9 + ], + "angle": 0, + "content": "Quantitative results. In Tab. 2, we compare the performance of EVCAP, MiniGPT4, BLIP, and BLIP-2 on" + }, + { + "type": "image", + "bbox": [ + 0.504, + 0.627, + 0.589, + 0.692 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.591, + 0.624, + 0.854, + 0.634 + ], + "angle": 0, + "content": "GT: This is an image of a blue Pikachu with yellow accents." + }, + { + "type": "text", + "bbox": [ + 0.591, + 0.635, + 0.832, + 0.644 + ], + "angle": 0, + "content": "SmallCap: A blue and white stuffed animal on a table." + }, + { + "type": "text", + "bbox": [ + 0.591, + 0.645, + 0.888, + 0.655 + ], + "angle": 0, + "content": "MiniGPT4: A blue poker sitting on the floor with its eyes closed." + }, + { + "type": "text", + "bbox": [ + 0.591, + 0.656, + 0.794, + 0.665 + ], + "angle": 0, + "content": "BLIP-2: A blue Pikachu sitting in a dark room." + }, + { + "type": "text", + "bbox": [ + 0.591, + 0.666, + 0.863, + 0.684 + ], + "angle": 0, + "content": "EVCap: A blue and yellow cartoon character sitting on a dark background." + }, + { + "type": "text", + "bbox": [ + 0.591, + 0.685, + 0.859, + 0.695 + ], + "angle": 0, + "content": "EVCap (w/ WHOOPS): A blue Pikachu is sitting on the floor." + }, + { + "type": "image_caption", + "bbox": [ + 0.5, + 0.708, + 0.892, + 0.75 + ], + "angle": 0, + "content": "Figure 4. Examples of captions generated by our EVCAP, EVCAP (w/ WHOOPS), and three SOTAs on WHOOPS dataset. Incorrect objects are highlighted in red, while correct ones are in blue." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.781, + 0.892, + 0.901 + ], + "angle": 0, + "content": "WHOOPS dataset. This dataset is particularly challenging due to its inclusion of unusual objects [6]. Initially, as an end-to-end trained model, our EVCAP exhibits performance similar to MiniGPT4. However, there is a noticeable improvement in the CIDEr score, after the external memory is enriched with 2396 new objects from the WHOOPS dataset, each represented by 5 synthesized images generated using stable diffusion [36]. It highlights the effectiveness of" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13738" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.09, + 0.47, + 0.146 + ], + "angle": 0, + "content": "Table 2. Quantitative results on commonsense-violating data - WHOOPS dataset. EVCAP (w/ WHOOPS) denotes EVCAP using the memory expanded by WHOOPS objects. The results reveal the open-world comprehension ability and expandability of EVCAP." + }, + { + "type": "table", + "bbox": [ + 0.081, + 0.15, + 0.468, + 0.328 + ], + "angle": 0, + "content": "
MethodB@4MCS
Only pre-trained models
BLIP [24] (from [6])13-65-
BLIP-2 FlanT5-XXL [25] (from [6])31-120-
BLIP-2 FlanT5-XXL [25] (reproduced)2826.793.117.9
Finetuned models on COCO
MiniGPT4 [46]24.226.784.818.2
BLIP [24]22.925.079.317.1
BLIP-2 FlanT5-XL [25]25.827.089.118.3
End-to-end trained models on COCO
EVCAP24.126.185.317.7
EVCAP (w/ WHOOPS)24.426.186.317.8
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.342, + 0.47, + 0.384 + ], + "angle": 0, + "content": "Table 3. Ablation study on components prior to the LLM decoder in EVCAP. The result of \"+\" Attentive fusion\" demonstrates the substantial impact of the external visual-name memory." + }, + { + "type": "table", + "bbox": [ + 0.081, + 0.387, + 0.468, + 0.453 + ], + "angle": 0, + "content": "
MethodCOCO testNoCaps valFlickr30k test
CSCSCS
ViT + Q-Former (Baseline)134.423.9108.814.276.817.3
+ Image query tokens (Baseline+)134.123.8109.014.377.317.2
+ Attentive fusion (EVCAP)140.124.7119.315.384.418.0
" + }, + { + "type": "image", + "bbox": [ + 0.081, + 0.469, + 0.166, + 0.53 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.168, + 0.469, + 0.387, + 0.479 + ], + "angle": 0, + "content": "GT: A hamster on a blanket with a hand behind it." + }, + { + "type": "text", + "bbox": [ + 0.168, + 0.48, + 0.468, + 0.49 + ], + "angle": 0, + "content": "Baseline: A hamster sitting on a couch with a person's hand near it." + }, + { + "type": "text", + "bbox": [ + 0.169, + 0.491, + 0.395, + 0.5 + ], + "angle": 0, + "content": "Baseline+: A close up of a small hamster on a bed." + }, + { + "type": "text", + "bbox": [ + 0.169, + 0.501, + 0.432, + 0.511 + ], + "angle": 0, + "content": "EVCap: A small hamster sitting on a blanket next to a hand." + }, + { + "type": "text", + "bbox": [ + 0.169, + 0.512, + 0.468, + 0.53 + ], + "angle": 0, + "content": "(pillow, hamster, grizzly, gameboard, kitten, spectacles, cat, blanket, giraffe, pug-dog)" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.542, + 0.47, + 0.601 + ], + "angle": 0, + "content": "Figure 5. Visualization of the captions generated from ablation study on the NoCaps validation set. We also show the retrieved object names by EVCAP, presented in gray. Incorrect objects in captions are highlighted in red, while correct ones are in blue." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.627, + 0.468, + 0.657 + ], + "angle": 0, + "content": "our idea of incorporating an expandable external memory into the captioning model for open-world comprehension." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.657, + 0.47, + 0.793 + ], + "angle": 0, + "content": "Qualitative results. Fig. 4 illustrates the captions generated by EVCAP, EVCAP (w/WHOOPS), and three SOTAs for one image from the WHOOPS dataset. Similar to other methods except for BLIP-2, EVCAP can not recognize \"blue cartoon character\" as \"Pikachu\", while EVCAP (w/WHOOPS) successfully predicts it because of the updated memory. SmallCap and MiniGPT4 tend to generate captions with hallucinatory objects, a result of commonsense-violating contents present in the images." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.802, + 0.245, + 0.818 + ], + "angle": 0, + "content": "5.3. Detailed analysis" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.826, + 0.468, + 0.901 + ], + "angle": 0, + "content": "Ablation study. We assess the contribution of each component prior to the LLM decoder in EVCAP by incrementally integrating the image query tokens and the attentive fusion module into our baseline model. The baseline model comprises a ViT+Q-Former, a linear layer, and a LLM decoder." + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.089, + 0.887, + 0.262 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.271, + 0.892, + 0.342 + ], + "angle": 0, + "content": "Figure 6. Visualization of the visual features in external memory using t-SNE. For visual features in LVIS dataset's objects (blue), the related objects fall in the same cluster. After adding more visual features of synthesized images from WHOOPS' objects, new objects (red) are located at appropriate clusters (zoom-in view)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.371, + 0.892, + 0.567 + ], + "angle": 0, + "content": "The quantitative results are shown in Tab. 3. When employing only the baseline model (Baseline), CIDEr scores drop notably by 5.7, 10.5, and 7.6 on COCO, NoCaps, and Flickr30k, respectively. The inclusion of trainable image query tokens (Baseline+) brings a marginal improvement on NoCaps and Flickr30k. However, the performance is significantly enhanced with the addition of attentive fusion (along with the introduction of external memory), indicating the pivotal role of the external visual-name memory in the overall effectiveness of EVCAP. This is further corroborated by the qualitative results in Fig. 5, where captions from Baseline and Baseline+ inaccurately include objects like \"couch\" and \"bed\", and Baseline+ overlooks \"hand\"." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.569, + 0.893, + 0.853 + ], + "angle": 0, + "content": "Exploration for external memory expandability. To demonstrate the scalability of the external memory in EVCAP, we visualize the visual features stored in LVIS external memory, and newly synthesized data from objects appearing in the WHOOPS dataset. We employ t-SNE [38] to plot visual features after reducing their dimensions to 2-D (Fig. 6). For clear visualization, we only randomly display 3649 visual features in LVIS memory, and add 479 visual features from WHOOPS objects. Among them, 35 samples are randomly labeled. The result shows a clear clustering of LVIS objects (blue) in the external memory, as well as the successful integration and appropriate localization of new objects from WHOOPS (red) into these clusters. This pattern not only confirms the distinctiveness of visual features already present in the memory but also demonstrates the potential to accurately incorporate and differentiate new objects introduced from updated data. These findings highlight our external memory's ability to expand and maintain its effectiveness even as new data is incorporated." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.894, + 0.901 + ], + "angle": 0, + "content": "Impact of external memory size. We examine the impact of external memory size in Tab. 4. On the one hand, we randomly remove \\(30\\%\\), \\(60\\%\\), and \\(90\\%\\) data in the external" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13739" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.09, + 0.472, + 0.132 + ], + "angle": 0, + "content": "Table 4. Impact of the external memory size on the performance of EVCAP by evaluation under CIDEr scores. Changes in the size of external memory result in changes in performance." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.135, + 0.468, + 0.237 + ], + "angle": 0, + "content": "
MethodNoCaps valFlickr30k
InNearOutOverallTest
LVIS objects (EVCAP)111.7119.5116.5119.384.4
- 30% LVIS112.0119.2115.3118.885.0
- 60% LVIS111.4119.1116.2119.085.1
- 90% LVIS110.6118.2115.8118.383.6
+ WHOOPS110.7118.9116.7119.084.9
" + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.25, + 0.466, + 0.368 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.378, + 0.471, + 0.421 + ], + "angle": 0, + "content": "Figure 7. CIDEr scores after training EVCAP with the number of retrieved object names K from 0 to 20. The results indicate that the performance is relatively optimal when K is set to be 10." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.449, + 0.47, + 0.629 + ], + "angle": 0, + "content": "memory constructed from LVIS objects. The results show the performance gradually degrades on NoCaps as reducing \\(30\\%\\) and \\(90\\%\\) LVIS. Despite some unexpected increases in certain results on NoCaps (5th row) and Flickr30k (4th - 5th rows), they do not alter the overall downward trend. Similar phenomena are also noted in SmallCap [35], we speculate it is due to data distribution. On the other hand, as we infuse WHOOPS knowledge into LVIS memory, there is a slight improvement on NoCaps (out) and Flickr30k. These observations validate the model's capability to effectively retrieve object names from an updated memory, enhancing its performance in generating captions." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.63, + 0.47, + 0.825 + ], + "angle": 0, + "content": "Impact of the number of retrieved object names. We investigate how the number of retrieved object names K (Sec. 3.3) affect EVCAP in Fig. 7. We train the model with K from 0 to 20 and evaluate the performance under CIDER on all three benchmarks. From the results, we can find that the model works worst on the out-domain dataset (NoCaps) with zero object names. It confirms that the performance boost from Baseline+ to EVCAP (Tab. 3) is primarily attributed to the retrieval-augmented mechanism, but not the customized Q-Former itself. With more object names, performance fluctuates but improves. Furthermore, we observe that setting K to 10 yields relatively optimal overall performance, validating the choice of \\( \\mathrm{K} = 10 \\) in EVCAP." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.826, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Analysis with different decoders. To explore the influence of different LLMs decoders on our EVCAP, we experiment by substituting Vicuna-13B with GPT2 and Vicuna-7B, as detailed in Tab. 5. With GPT2 as the decoder, EVCAP still markedly surpasses other GPT2-based models, achieving" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.09, + 0.892, + 0.132 + ], + "angle": 0, + "content": "Table 5. Analysis with different LLM decoders including GPT2, Vicuna-7B, and Vicuna-13B. The results reveal EVCAP is effective when applying it in different LLM decoders." + }, + { + "type": "table", + "bbox": [ + 0.504, + 0.135, + 0.892, + 0.276 + ], + "angle": 0, + "content": "
MethodLLMCOCO testNoCaps valFlickr30k test
CSCSCS
SmallCap [35]GPT2119.721.3--60.6-
ViEcap [15]GPT292.918.266.29.547.913.6
EVCAPGPT2131.023.297.613.370.616.1
MiniGPT4 [46]Vicuna-7B119.423.5108.715.773.917.2
InstructBLIP [11]Vicuna-7B--123.1-82.4-
EVCAPVicuna-7B139.024.7116.815.382.718.0
MiniGPT4 [46]Vicuna-13B129.623.4108.815.178.416.9
InstructBLIP [11]Vicuna-13B--121.9-82.8-
EVCAPVicuna-13B140.124.7119.315.384.418.0
" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.304, + 0.892, + 0.456 + ], + "angle": 0, + "content": "impressive gains of 11.3 and 10.0 under CIDEr on COCO and Flickr30k, compared to SmallCap. When employing Vicuna-7B, the comparison of performance trends mirrors those observed with Vicuna-13B, further attesting to the robustness and adaptability of EVCAP across different LLM decoders. Notably, both SmallCap, which retrieves captions, and our GPT2-based EVCAP, which retrieves object names, use the same GPT2 decoder. Therefore, their comparison also underscores the effectiveness of our method's object name retrieval and attentive fusion strategy." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.456, + 0.893, + 0.637 + ], + "angle": 0, + "content": "Limitations. First, EVCAP cannot retrieve all objects that appear in the given image due to the memory coverage limits, leading to incomplete image descriptions (Fig. 4). We will investigate integrating object detection with image captioning to enhance completeness. Second, our focus on object representation restricts consideration of other crucial captioning elements, affecting overall performance. Similar to all models trained with COCO dataset, EVCAP has limitations in generating varied styles, which is reflected in our relatively modest performance improvements in Tab. 2, compared to MiniGPT4. We will overcome it by exploring methodologies that encourage style diversity in the future." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.655, + 0.62, + 0.672 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.682, + 0.892, + 0.803 + ], + "angle": 0, + "content": "We further advance image captioning in real-world scenarios by introducing EVCAP, a novel image captioning model with object names retrieved from an external visual-name memory. The external memory is easily expandable, allowing for effortless updates with new object visuals and names. We extensively compare EVCAP with SOTAs on various benchmarks and commonsense-violating data, demonstrating its significant superiority in performance." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.803, + 0.895, + 0.902 + ], + "angle": 0, + "content": "Acknowledgements. This work was supported by JSPS/MEXT KAKENHI Grant Numbers JP22H05015, JP23H03449, JP23KJ0404, and JP22K17947, and the commissioned research (No. 225) by the National Institute of Information and Communications Technology (NICT), Japan, and the Institute of AI and Beyond of the University of Tokyo, and ROIS NII Open Collaborative Research 2023-23FC01, 2024-24S1201." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "13740" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.079, + 0.09, + 0.176, + 0.107 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.116, + 0.47, + 0.143 + ], + "angle": 0, + "content": "[1] https : / / huggingface . co / spaces / RitaParadaRamos/SmallCapDemo.5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.145, + 0.472, + 0.213 + ], + "angle": 0, + "content": "[2] Harsh Agrawal, Karan Desai, Yufei Wang, Xinlei Chen, Rishabh Jain, Mark Johnson, Dhruv Batra, Devi Parikh, Stefan Lee, and Peter Anderson. Nocaps: Novel object captioning at scale. In Proc. IEEE International Conference on Computer Vision (ICCV), 2019. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.215, + 0.47, + 0.284 + ], + "angle": 0, + "content": "[3] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: A visual language model for few-shot learning. In Advances in Neural Information Processing Systems (NeurIPS), 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.285, + 0.47, + 0.354 + ], + "angle": 0, + "content": "[4] Peter Anderson, Xiaodong He, Chris Buehler, Damien Teney, Mark Johnson, Stephen Gould, and Lei Zhang. Bottom-up and top-down attention for image captioning and visual question answering. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.355, + 0.47, + 0.411 + ], + "angle": 0, + "content": "[5] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A frontier large vision-language model with versatile abilities. arXiv preprint arXiv:2308.12966, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.412, + 0.47, + 0.495 + ], + "angle": 0, + "content": "[6] Nitzan Bitton-Guetta, Yonatan Bitton, Jack Hessel, Ludwig Schmidt, Yuval Elovici, Gabriel Stanovsky, and Roy Schwartz. Breaking common sense: Whoops! a vision-and-language benchmark of synthetic and compositional images. In Proc. IEEE International Conference on Computer Vision (ICCV), 2023. 4, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.496, + 0.47, + 0.564 + ], + "angle": 0, + "content": "[7] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. In Advances in Neural Information Processing Systems (NeurIPS), 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.566, + 0.47, + 0.635 + ], + "angle": 0, + "content": "[8] Xi Chen, Josip Djolonga, Piotr Padlewski, Basil Mustafa, Soravit Changpinyo, Jialin Wu, Carlos Riquelme Ruiz, Sebastian Goodman, Xiao Wang, Yi Tay, et al. PaLI-X: On scaling up a multilingual vision and language model. arXiv preprint arXiv:2305.18565, 2023. 1, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.636, + 0.47, + 0.788 + ], + "angle": 0, + "content": "[9] Xi Chen, Xiao Wang, Soravit Changpinyo, AJ Piergiovanni, Piotr Padlewski, Daniel Salz, Sebastian Goodman, Adam Grycner, Basil Mustafa, Lucas Beyer, Alexander Kolesnikov, Joan Puigcerver, Nan Ding, Keran Rong, Hassan Akbari, Gaurav Mishra, Linting Xue, Ashish V Thapliyal, James Bradbury, Weicheng Kuo, Mojtaba Seyedhosseini, Chao Jia, Burcu Karagol Ayan, Carlos Riquelme Ruiz, Andreas Peter Steiner, Anelia Angelova, Xiaohua Zhai, Neil Houlsby, and Radu Soricut. PaLI: A jointly-scaled multilingual language-image model. In International Conference on Learning Representations (ICLR), 2023. 1, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.79, + 0.47, + 0.871 + ], + "angle": 0, + "content": "[10] Wei-Lin Chiang, Zhuohan Li, Zi Lin, Ying Sheng, Zhanghao Wu, Hao Zhang, Lianmin Zheng, Siyuan Zhuang, Yonghao Zhuang, Joseph E Gonzalez, et al. Vicuna: An open-source chatbot impressing gpt-4 with \\(90\\%\\) chatgpt quality. See https://vicuna.lmsys.org (accessed 14 April 2023), 2023. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.873, + 0.47, + 0.902 + ], + "angle": 0, + "content": "[11] Wenliang Dai, Junnan Li, Dongxu Li, Anthony Meng Huat Tiong, Junqi Zhao, Weisheng Wang, Boyang Li, Pascale" + }, + { + "type": "list", + "bbox": [ + 0.079, + 0.116, + 0.472, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.147 + ], + "angle": 0, + "content": "Fung, and Steven Hoi. Instructclip: Towards general-purpose vision-language models with instruction tuning. In Advances in Neural Information Processing Systems (NeurIPS), 2023. 3, 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.149, + 0.892, + 0.245 + ], + "angle": 0, + "content": "[12] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations (ICLR), 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.247, + 0.892, + 0.302 + ], + "angle": 0, + "content": "[13] Jiashuo Fan, Yaoyuan Liang, Leyao Liu, Shaolun Huang, and Lei Zhang. Rca-noc: Relative contrastive alignment for novel object captioning. In Proc. IEEE International Conference on Computer Vision (ICCV), 2023. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.303, + 0.892, + 0.37 + ], + "angle": 0, + "content": "[14] Yuxin Fang, Wen Wang, Binhui Xie, Quan Sun, Ledell Wu, Xinggang Wang, Tiejun Huang, Xinlong Wang, and Yue Cao. Eva: Exploring the limits of masked visual representation learning at scale. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.372, + 0.892, + 0.439 + ], + "angle": 0, + "content": "[15] Junjie Fei, Teng Wang, Jinrui Zhang, Zhenyu He, Chengjie Wang, and Feng Zheng. Transferable decoding with visual entities for zero-shot image captioning. In Proc. IEEE International Conference on Computer Vision (ICCV), 2023. 2, 4, 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.442, + 0.892, + 0.482 + ], + "angle": 0, + "content": "[16] Zhengcong Fei. Memory-augmented image captioning. In Proc. AAAI Conference on Artificial Intelligence (AAAI), 2021. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.484, + 0.892, + 0.537 + ], + "angle": 0, + "content": "[17] Agrim Gupta, Piotr Dollar, and Ross Girshick. LVIS: A dataset for large vocabulary instance segmentation. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.539, + 0.892, + 0.593 + ], + "angle": 0, + "content": "[18] Ryuichiro Hataya, Han Bao, and Hiromi Arai. Will large-scale generative models corrupt future datasets? In Proc. IEEE International Conference on Computer Vision (ICCV), 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.596, + 0.892, + 0.677 + ], + "angle": 0, + "content": "[19] Lisa Anne Hendricks, Subhashini Venugopalan, Marcus Rohrbach, Raymond Mooney, Saenko Kate, and Trevor Darrell. Deep compositional captioning: Describing novel object categories without paired training data. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.679, + 0.892, + 0.761 + ], + "angle": 0, + "content": "[20] Ziniu Hu, Ahmet Iscen, Chen Sun, Zirui Wang, Kai-Wei Chang, Yizhou Sun, Cordelia Schmid, David A Ross, and Alireza Fathi. Reveal: Retrieval-augmented visual-language pre-training with multi-source multimodal knowledge memory. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 1, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.763, + 0.892, + 0.802 + ], + "angle": 0, + "content": "[21] Jeff Johnson, Matthijs Douze, and Hervé Jégou. Billionscale similarity search with gpus. IEEE Transactions on Big Data, 7(3):535-547, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.804, + 0.892, + 0.858 + ], + "angle": 0, + "content": "[22] Andrej Karpathy and Li Fei-Fei. Deep visual-semantic alignments for generating image descriptions. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2015. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.86, + 0.894, + 0.902 + ], + "angle": 0, + "content": "[23] Jacob Devlin Ming-Wei Chang Kenton and Lee Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "13741" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.092, + 0.47, + 0.134 + ], + "angle": 0, + "content": "North American Chapter of the Association for Computational Linguistics: Human Language Technologies (NAACL-HLT), 2019. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.135, + 0.47, + 0.203 + ], + "angle": 0, + "content": "[24] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In Proc. International conference on machine learning (ICML), 2022. 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.204, + 0.47, + 0.272 + ], + "angle": 0, + "content": "[25] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In Proc. International conference on machine learning (ICML), 2023. 1, 2, 3, 4, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.274, + 0.47, + 0.343 + ], + "angle": 0, + "content": "[26] Xiujun Li, Xi Yin, Chunyuan Li, Pengchuan Zhang, Xiaowei Hu, Lei Zhang, Lijuan Wang, Houdong Hu, Li Dong, Furu Wei, et al. Oscar: Object-semantics aligned pre-training for vision-language tasks. In Proc. European Conference on Computer Vision (ECCV), 2020. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.344, + 0.47, + 0.411 + ], + "angle": 0, + "content": "[27] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dálár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Proc. European Conference on Computer Vision (ECCV), 2014. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.413, + 0.47, + 0.455 + ], + "angle": 0, + "content": "[28] Jiasen Lu, Jianwei Yang, Dhruv Batra, and Devi Parikh. Neural baby talk. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.455, + 0.47, + 0.496 + ], + "angle": 0, + "content": "[29] Ron Mokady, Amir Hertz, and Amit H Bermano. Clipcap: Clip prefix for image captioning. arXiv preprint arXiv:2111.09734, 2021. 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.497, + 0.47, + 0.579 + ], + "angle": 0, + "content": "[30] Bryan A Plummer, Liwei Wang, Chris M Cervantes, Juan C Caicedo, Julia Hockenmaier, and Svetlana Lazebnik. Flickr30k entities: Collecting region-to-phrase correspondences for richer image-to-sentence models. In Proc. IEEE International Conference on Computer Vision (ICCV), 2015. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.581, + 0.47, + 0.622 + ], + "angle": 0, + "content": "[31] Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, Ilya Sutskever, et al. Language models are unsupervised multitask learners. OpenAI blog, 1(8):9, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.623, + 0.47, + 0.705 + ], + "angle": 0, + "content": "[32] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In Proc. International conference on machine learning (ICML), 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.707, + 0.47, + 0.774 + ], + "angle": 0, + "content": "[33] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J Liu. Exploring the limits of transfer learning with a unified text-to-text transformer. Journal of Machine Learning Research (JMLR), 21(1):5485-5551, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.776, + 0.47, + 0.831 + ], + "angle": 0, + "content": "[34] Rita Ramos, Desmond Elliott, and Bruno Martins. Retrievalaugmented image captioning. In Proc. Conference of the European Chapter of the Association for Computational Linguistics (EACL), 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.832, + 0.47, + 0.9 + ], + "angle": 0, + "content": "[35] Rita Ramos, Bruno Martins, Desmond Elliott, and Yova Kementchedjhieva. Smallcap: Lightweight image captioning prompted with retrieval augmentation. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 1, 2, 3, 5, 8" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.161 + ], + "angle": 0, + "content": "[36] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.163, + 0.892, + 0.232 + ], + "angle": 0, + "content": "[37] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.234, + 0.892, + 0.275 + ], + "angle": 0, + "content": "[38] Laurens Van der Maaten and Geoffrey Hinton. Visualizing data using t-sne. Journal of machine learning research (JMLR), 9(11), 2008. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.277, + 0.892, + 0.344 + ], + "angle": 0, + "content": "[39] Subhashini Venugopalan, Lisa Anne Hendricks, Marcus Rohrbach, Raymond Mooney, Trevor Darrell, and Kate Saenko. Captioning images with diverse objects. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.346, + 0.892, + 0.415 + ], + "angle": 0, + "content": "[40] Duc Minh Vo, Hong Chen, Akihiro Sugimoto, and Hideki Nakayama. Noc-rek: Novel object captioning with retrieved vocabulary from external knowledge. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.417, + 0.892, + 0.499 + ], + "angle": 0, + "content": "[41] Weihan Wang, Qingsong Lv, Wenmeng Yu, Wenyi Hong, Ji Qi, Yan Wang, Junhui Ji, Zhuoyi Yang, Lei Zhao, Xixuan Song, Jiazheng Xu, Bin Xu, Juanzi Li, Yuxiao Dong, Ming Ding, and Jie Tang. Cogvlm: Visual expert for pretrained language models. arXiv preprint arXiv:2311.03079, 2023. 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.501, + 0.892, + 0.571 + ], + "angle": 0, + "content": "[42] Kelvin Xu, Jimmy Ba, Ryan Kiros, Kyunghyun Cho, Aaron Courville, Ruslan Salakhudinov, Rich Zemel, and Yoshua Bengio. Show, attend and tell: Neural image caption generation with visual attention. In Proc. International conference on machine learning (ICML). 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.572, + 0.892, + 0.653 + ], + "angle": 0, + "content": "[43] Zhuolin Yang, Wei Ping, Zihan Liu, Vijay Korthikanti, Weili Nie, De-An Huang, Linxi Fan, Zhiding Yu, Shiyi Lan, Bo Li, et al. Re-ViLM: Retrieval-augmented visual language model for zero and few-shot image captioning. In Findings of the Association for Computational Linguistics: EMNLP, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.655, + 0.892, + 0.726 + ], + "angle": 0, + "content": "[44] Michihiro Yasunaga, Armen Aghajanyan, Weijia Shi, Richard James, Jure Leskovec, Percy Liang, Mike Lewis, Luke Zettlemoyer, and Wen-tau Yih. Retrieval-augmented multimodal language modeling. In International Conference on Learning Representations (ICLR), 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.727, + 0.892, + 0.796 + ], + "angle": 0, + "content": "[45] Pengchuan Zhang, Xiujun Li, Xiaowei Hu, Jianwei Yang, Lei Zhang, Lijuan Wang, Yejin Choi, and Jianfeng Gao. Vinvl: Revisiting visual representations in vision-language models. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.797, + 0.892, + 0.853 + ], + "angle": 0, + "content": "[46] Deyao Zhu, Jun Chen, Xiaoqian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models. arXiv preprint arXiv:2304.10592, 2023. 3, 4, 5, 7, 8" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.853 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "13742" + } + ] +] \ No newline at end of file diff --git a/2024/EVCap_ Retrieval-Augmented Image Captioning with External Visual-Name Memory for Open-World Comprehension/003cc44f-e449-40f1-aec0-baf1c6235010_origin.pdf b/2024/EVCap_ Retrieval-Augmented Image Captioning with External Visual-Name Memory for Open-World Comprehension/003cc44f-e449-40f1-aec0-baf1c6235010_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..a645fc7f3edc1de035f5ee180cfe543f06e8d46d --- /dev/null +++ b/2024/EVCap_ Retrieval-Augmented Image Captioning with External Visual-Name Memory for Open-World Comprehension/003cc44f-e449-40f1-aec0-baf1c6235010_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5d9719a7598499b5c28a62c0b7ab342ff7370324ef654fce8a15df3c8b8d019c +size 1678479 diff --git a/2024/EVCap_ Retrieval-Augmented Image Captioning with External Visual-Name Memory for Open-World Comprehension/full.md b/2024/EVCap_ Retrieval-Augmented Image Captioning with External Visual-Name Memory for Open-World Comprehension/full.md new file mode 100644 index 0000000000000000000000000000000000000000..de4fe746ec7779d072bf1c97f09cb9a2cd106883 --- /dev/null +++ b/2024/EVCap_ Retrieval-Augmented Image Captioning with External Visual-Name Memory for Open-World Comprehension/full.md @@ -0,0 +1,352 @@ +# EVCAP: Retrieval-Augmented Image Captioning with External Visual-Name Memory for Open-World Comprehension + +Jiaxuan $\mathrm{Li}^{1*}$ , Duc Minh $\mathrm{Vo}^{1*}$ , Akihiro Sugimoto $^{2}$ , Hideki Nakayama $^{1}$ + +1The University of Tokyo, Japan 2National Institute of Informatics, Japan + +{li, vm Duc}@nlab.ci.i.u-tokyo.ac.jp sugimoto@nii.ac.jp nakayama@ci.i.u-tokyo.ac.jp + +# Abstract + +Large language models (LLMs)-based image captioning has the capability of describing objects not explicitly observed in training data; yet novel objects occur frequently, necessitating the requirement of sustaining up-to-date object knowledge for open-world comprehension. Instead of relying on large amounts of data and/or scaling up network parameters, we introduce a highly effective retrieval-augmented image captioning method that prompts LLMs with object names retrieved from External Visual-name memory (EVCAP). We build ever-changing object knowledge memory using objects' visuals and names, enabling us to (i) update the memory at a minimal cost and (ii) effortlessly augment LLMs with retrieved object names by utilizing a lightweight and fast-to-train model. Our model, which was trained only on the COCO dataset, can adapt to out-of-domain without requiring additional fine-tuning or re-training. Our experiments conducted on benchmarks and synthetic commonsense-violating data show that EVCAP, with only 3.97M trainable parameters, exhibits superior performance compared to other methods based on frozen pre-trained LLMs. Its performance is also competitive to specialist SOTAs that require extensive training. + +# 1. Introduction + +Advanced image captioning based on large language models (LLMs) [3, 8, 9, 25] has focused on the approach using big-scale models trained on ever-increasingly large-scale datasets, which is no longer viable. This is because the computational cost to train the models increases exponentially and, more importantly, updating training data is almost impossible to keep pace with the growth of novel objects in our daily lives. Sustaining ever-changing object knowledge with a reasonable cost is a pressing concern in LLMs-based models to truly unlock open-world comprehension. + +Retrieval-augmented image captioning [20, 35] is + +![](images/71d2fc87b5ba24b951155fde0165bac7dc20438920282706b7fa1b5866e3486c.jpg) +Figure 1. Overall comparison of our EVCAP and SOTAs. (Upper) Generated captions by SmallCap, BLIP-2, and our EVCAP for a commonsense-violating image from the WHOOPS dataset. $\times$ and $\checkmark$ indicate incorrect and correct predictions, respectively. Incorrect objects in captions are highlighted in red, while correct ones are in blue. SmallCap and BLIP-2 give incorrect predictions for "ice skates" and "wood floor", respectively, while our EVCAP utilizes an external visual-name memory to enhance attention to objects within the image, leading to superior performance for image captioning. (Lower) Comparison of the number of trainable parameters, CIDEr score on COCO and NoCaps datasets. The size of each circle reflects the log number of trainable parameters. EVCAP (3.97M) has less trainable parameters than others while achieving comparable results with SOTAs at scale. + +emerging as an alternative since it considerably reduces training costs in both time and data while producing encouraging results. Nonetheless, with their huge datastore, it is obvious that LLMs would imitate the given texts, limiting their ability to describe open-world objects properly. For instance, SmallCap [35] considers the words "skateboard" and "wooden floor" to be a pair regardless of visual appearances containing a commonsense-violating pair + +of "ice skates" and "wood floor" (Fig. 1, upper). Additionally, prompting the LLMs given a lot of retrieved texts becomes cumbersome, requiring more trainable parameters. Fig. 1 (lower) shows that the CIDEr scores obtained by a lightweight SmallCap [35] with 43M trainable parameters are far away from those obtained by a heavy REVEAL [20] with 2.1B trainable parameters. Beyond that, due to the frequent occurrence of new objects, access to their sample texts is not always feasible, making the memory utilized in [20, 35] difficult to grow. We thus aim to streamline the external memory used in previous work [20, 35] by storing a sufficiently small amount of object information. And, of course, not only does the model not stereotype the example sentences, but the number of trainable parameters would be reduced drastically as a result of the causation (Fig. 1). + +We follow [13, 40] to construct a key-value memory where the key is represented by object's features, and the value corresponds to object's name. Unlike [13, 40], which rely on object definition as the key, our method leverages the visual appearance of the object as the key because of the abundance of object images readily available on the internet. We propose an external visual-name memory tailored for ease of expansion and cost-effectiveness in upholding up-to-date object information. We present a highly effective retrieval-augmented LLMs-based image captioning method, called EVCAP, that prompts frozen LLMs with object names retrieved from our proposed memory for open-world comprehension. EVCAP contains a frozen image encoder ViT [14] and Q-Former [25] with trainable image query tokens for object retrieval, an attentive fusion module, a trainable linear layer for mapping between vision and language latent spaces, and a frozen LLM decoder [10] for generating captions. Specifically, the attentive fusion module feeds retrieved object names and visual features into a customized frozen Q-Former using trainable object name query tokens to implicitly reduce the presence of superfluous object names. As a result, EVCAP amounts to only 3.97M trainable parameters. Once trained, the model can be adapted to new domains and large-scale data without further fine-tuning or re-training. Our contributions are as follows: + +- We provide an extensible external visual-name memory with minimal but useful object information, which enables LLMs-based models to comprehend the open world. +- We present a highly efficacious retrieval-augmented image captioning EVCAP with 3.97M trainable parameters. + +On in-/out-domain benchmarks and synthetic commonsense-violating dataset, EVCAP trained solely on COCO dataset competes with other lightweight methods by a margin while being on par with other specialist SOTAs. + +# 2. Related Work + +Image captioning aims to describe the contents of a given image. It can be roughly divided into two approaches: non- + +LLMs-based methods and LLMs-based ones. The former approaches [4, 22, 42] typically employ a visual encoder and a language decoder in an end-to-end fashion to generate captions. However, they are incapable of describing open-world objects. The latter one leverages pre-trained large-scale vision models (CLIP [32], ViT [12]) and LLMs (GPTs [7, 31], T5 [33], LLaMA [37]) by bridging the gap between two modalities using either pre-training with largescale data or the learned mapper or prompt techniques. LLMs-based models [8, 9, 25, 29] demonstrate advancements in image captioning challenges, allowing the capacity to describe anything as long as pre-trained vision models can recognize it. Our method belongs to the LLMs-based approaches, but instead of relying fully on the pre-trained vision model, we use object names retrieved from the external memory to augment LLMs-based image captioning. + +Novel object captioning is a branch of image captioning that describes images containing objects that were not seen during training. Non-LLMs-based methods explore more objects by learning from unpaired image-sentence sources (DCC [19], NOC [39]) or rely on novel object detectors to recognize novel concepts (NBT [28], OSCAR [26] and VinVL [45]). LLMs-based methods such as ViECap [15] leverage the pre-trained CLIP [32] to obtain object entities. Nevertheless, the cut-off in training time of the pre-trained object detector or CLIP prevents it from detecting novel objects that arise quickly in reality. Unlike earlier work, we can readily update our recognition of novel concepts by adding them to external memory, ensuring that we keep any new objects from the past and even the future. + +Retrieval-augmented image captioning is a recently popular approach that augments the captioning model with retrieved information for better open-world understanding. AoANet [16] uses a memory bank of image-sentence pairs and target words. SmallCap [35] employs image-to-text retrieval to obtain sampled captions from a captions dataset. RA-CM3 [44] retrieves documents from an external memory of a mixture of text and image via a dense multimodal retriever. EXTRA [34] and Re-ViLM [43] exploit the similarity of the input image and vision candidates to retrieve captions. Different from the previous methods, our external memory contains visual-name pairs to avoid redundant information in the external captions/documents. In addition, we use an attentive fusion module to mitigate the effects of irrelevant retrieved object names on caption generation. + +# 3. Proposed EVCAP + +# 3.1. Idea of EVCAP + +We aim to build a retrieval-augmented LLMs-based image captioning model with a sufficiently small yet informative external memory. It involves two challenges: (1) constructing an expandable external memory and (2) building an effective LLMs-based model using retrieved object names. + +![](images/ba4975506959fdf05710a3b6655208a431a0dd4cb0e004e06a6ab95c19e0379c.jpg) +Figure 2. Schematic of our proposed EVCAP. It consists of an external visual-name memory with image embeddings and object names (upper), a frozen ViT and Q-Former equipped with trainable image query tokens, an attentive fusion module developed by a customized frozen Q-Former and trainable object name query tokens, and a frozen LLM with a trainable linear layer (lower). The ViT and Q-Former extract learned visual features from the input image, which are then used to retrieve object names from the external memory. These retrieved object names and learned visual features undergo cross-attention in the customized Q-Former, creating refined object name features. Finally, the object name features combined with visual features are fed into the LLM post a linear layer for generating captions. + +As discussed above, challenge (1) can be resolved by utilizing the visual appearance of objects. However, if we restrict our memory to only a visual-name pair for each object, our memory will be lacking in diversity. Therefore, we gather several images for each target object. Additionally, we keep the synthetic images in our memory to avoid the harm that synthetic images might cause to our method, as pointed out in [18]. With the capability to collect images from the internet, EVCAP can be easily expanded to include novel objects from the real world effortlessly. + +We base our method on the frozen pre-trained vision model and LLM with several trainable layers (Fig. 2), giving in a model that is cheap to train. To guide the LLM, we adopt a recently popular approach called prompting as in [11, 25, 29, 35, 46]. We begin by matching the learned visual features from the input image with image embeddings stored in memory, retrieving object names. We also introduce an attentive fusion module designed to implicitly remove irrelevant retrieved names. Finally, following the attentive fusion, we combine the learned visual features and object name features to form a prompt for the LLM to generate a caption, thus addressing challenge (2). + +# 3.2. External visual-name memory + +To build the external visual-name memory, we first collect image-name pairs from the external data source. After + +that, we encode these images into image embeddings, which serve as keys in memory, and use their names as values. + +External data source. We utilize object images from LVIS dataset [17] to construct our external visual-name memory $\mathcal{M}$ . Specifically, we use 1203 objects in LVIS, where we randomly select from one to ten images for each object, amounting to 8581 object images. Furthermore, as mentioned in Sec. 3.1, we also incorporate synthetic images in our memory construction. Using stable diffusion [36], we generate five additional images for each object, with a prompt of "a photo of {object name}", resulting in a total of $M = 14596$ ( $8581 + 5 \times 1203$ ) images. Each object image $X^i$ is associated with an object name $v^i$ . Note that many object images may share the same object name. For the sake of simplicity, we may regard each image as corresponding to a single name. In summary, we have $M$ image-name pairs $\{(X^i,v^i)\}_{i=1}^M$ for external memory construction. + +External memory construction. For each image $X^i$ , we use a frozen vision encoder $\mathcal{E}(\cdot)$ (see Sec. 3.3 for detail) to project it into 32 embeddings with the size of $1 \times 768$ each: $\{\mathbf{k}_1^i, \mathbf{k}_2^i, \dots, \mathbf{k}_{32}^i\} = \mathcal{E}(X^i)$ . We then average 32 embeddings to produce a single embedding $\mathbf{k}^i$ ( $1 \times 768$ ) that serves as the key (visual) in $\mathcal{M}$ . The paired object name $v^i$ acts as its value (name). Consequently, we have the visual-name memory $\mathcal{M} = \{(\mathbf{k}^i, v^i)\}_{i=1}^M$ which is indexed using FAISS [21], facilitating rapid searches based on sim + +ilarity measures. Our memory can be expanded effortlessly by gathering additional visual-name pairs (see Sec. 5.3). + +# 3.3. Object names retrieval + +Image encoding. We feed a frozen vision encoder $\mathcal{E}$ image $X$ and image query tokens $\mathbf{T}_{\mathrm{img}}$ to produce visual features $\mathcal{Q}$ . To enable the retrieval process controllable, we make image query tokens to be trainable. Thus, the image encoding process can be summarized as $\mathcal{Q} = \mathcal{E}(X,\mathbf{T}_{\mathrm{img}})$ . We use the BLIP-2 pre-trained vision encoder [25], which consists of a pre-trained vision transformer ViT-g [14] outputting image features $(257\times 1408)$ , and a Q-Former receiving image features producing $|\mathcal{Q}| = 32$ learned visual features $(1\times 768$ each). We denote $\mathcal{Q} = \{\mathbf{q}_1,\mathbf{q}_2,\dots,\mathbf{q}_{32}\}$ . + +Retrieval. Having obtained $\mathcal{Q}$ , we calculate the cosine similarity between the query $\mathbf{q}_j \in \mathcal{Q}$ and the key $\mathbf{k}^i \in \mathcal{M}$ . The similarity calculation is given by $\mathrm{SIM}(\mathbf{q}_j, \mathbf{k}^i) = \frac{\mathbf{q}_j^\top \mathbf{k}^i}{\|\mathbf{q}_j\| \|\mathbf{k}^i\|}$ , where $i \in [1, M]$ , $j \in [1, 32]$ . Given each $\mathbf{q}_j$ , we select one key with the highest similarity score, resulting in 32 key-value candidates $\{\mathbf{k}_j^{\mathrm{best}}, v_j^{\mathrm{best}}\}_{j=1}^{32}$ . + +After that, we filter out candidates with repeated object names (values), and then select the top-K values. In particular, we determine the index $j$ from the key that has the highest SIM score. These selected values $v_{j}^{\mathrm{best}}$ are redefined as the new notation $v_{l}$ in the retrieved top-K object names for the input image, which can be summarized as follows: + +$$ +\{\mathbf {k} _ {j} ^ {\text {b e s t}}, v _ {j} ^ {\text {b e s t}} \} = \arg \max _ {\mathbf {k} ^ {i}} \operatorname {S I M} \left(\mathbf {q} _ {j}, \mathbf {k} ^ {i}\right), +$$ + +$$ +j = \arg \max _ {j} \mathrm {S I M} (\mathbf {q} _ {j}, \mathbf {k} _ {j} ^ {\mathrm {b e s t}}), v _ {l} \gets v _ {j} ^ {\mathrm {b e s t}}, +$$ + +where $l \in [1, \mathrm{K}]$ . As a result, the retrieved top- $\mathbf{K}$ object names are $\{v_l\}_{l=1}^{\mathrm{K}}$ . + +# 3.4. Attentive fusion + +Since the object names obtained from the retrieval process may be redundant, we develop an attentive fusion module to selectively distill object name features. + +The retrieved object names $\{v_{l}\}_{l = 1}^{\mathrm{K}}$ are concatenated together into a sequence $S$ , each separated by a delimiter: $S = \{v_{1},[\mathrm{SEP}],v_{2},[\mathrm{SEP}],\dots ,[\mathrm{SEP}],v_{\mathrm{K}}\}$ . The sequence $S$ and visual features $\mathcal{Q}$ are fed into a customized Q-Former $\mathcal{F}(\cdot)$ , which is constructed from the frozen pretrained Q-Former as we used in vision encoder $\mathcal{E}$ . Nonetheless, in order to enable object names to get attention from visual features, we switch the image embedding port and the text instruction port (see the supplement for architecture detail). Like in the image encoding process in Sec. 3.3, we make the object name query tokens $\mathbf{T}_{\mathrm{obj}}$ learnable during training to assist in learning object name features related to the caption. The size of $\mathbf{T}_{\mathrm{obj}}$ is $P\times 768$ , where $P$ indicates the number of object name query tokens. We get the object name features $\mathcal{V} = \mathcal{F}(S,\mathcal{Q},\mathbf{T}_{\mathrm{obj}})$ . + +# 3.5. Caption generation + +Before inputting the visual features $\mathcal{Q}$ and object name features $\nu$ into the LLM decoder, we concatenate $(\oplus)$ them and use a linear layer $\phi(\cdot)$ to project them into the input latent space of the LLM as $\phi(\mathcal{Q} \oplus \mathcal{V})$ . The LLM used for caption generation in this work is the pre-trained Vicuna-13B [10], an open-source chatbot constructed from LLaMA [37]. During training and evaluation, we design a prompt in a conversational format, that is similar to [46]: + +Human: Describe this image in detail. +Assistant: + +in which, ProjFeature denotes the projected feature $\phi (\mathcal{Q}\oplus \mathcal{V})$ after the linear layer. In training phase, given input caption tokens $\{c_i\}_{i = 1}^L$ , the LLM decoder concatenates the embedded prompt $\{\mathbf{w}_i\}_{i = 1}^N$ and the embedded caption tokens $\{\mathbf{c}_i\}_{i = 1}^L$ as input, and predicts the caption tokens in an autoregressive fashion, while in the evaluation phase, we only need to input the embedded prompt. We train EV-CAP by minimizing the cross-entropy loss in an end-to-end way: $\mathcal{L}_{\theta} = -\sum_{i = 1}^{L}\log p_{\theta}(c_i|\mathbf{w}_1,\dots \mathbf{w}_N,\mathbf{c}_1,\dots ,\mathbf{c}_{i - 1})$ , in which $\theta$ indicates the trainable parameters. + +# 4. Experimental Settings + +# 4.1. Training setup + +Implementation. EVCAP uses the same image encoder as in BLIP-2 [25], consisting of a ViT-g [14] and their pretrained Q-Former. Since we intend to obtain object name features through cross-attention between retrieved object names and visual features, we develop a customized Q-Former, which consists of BERT [23] with cross-attention layers inserted at every other transformer block. We use a frozen Vicuna-13B [10] as the caption generator. + +Training dataset. For all experiments, we exclusively train EVCAP using the training set of COCO dataset [27], consisting of 82k images and 5 captions per images. The entire training process takes about 3 hours on 4 A6000 GPUs, using mixed precisions (more details in the supplementary). + +# 4.2. Evaluation setup + +Evaluation dataset. We evaluate EVCAP, trained using the COCO training set, across four datasets: its test set, two challenging benchmarks - NoCaps validation set and Flickr30k test set, and a synthetic commonsense-violating dataset - WHOOPS. We adhere to follow prior work [15, 41] to use the same images of Karpathy split [22] on COCO test set, NoCaps [2] validation set, and Karpathy split on Flickr30k [30] test set. In addition, WHOOPS [6] is a synthetic image captioning dataset comprising 500 synthetic commonsense-violating images and 2500 paired captions. + +Table 1. Quantitative comparison against SOTA methods on three common image captioning benchmarks. * denotes using a memory bank. We report the size of training data and parameters; BLEU@4 (B@4), METEOR (M), CIDEr (C), and SPICE (S) scores on COCO test set; C and S scores on in-domain, near-domain, out-domain and overall data of NoCaps validation set; C and S scores on Flickr30k test set. Higher score is better. Bold indicates the best results among compared methods, normal indicates the second best results. + +
MethodTrainingCOCONoCaps valFlickr30k Test
DataPara.TestIn-domainNear-domainOut-domainOverallCS
B@4MCSCSCSCSCSCS
Heavyweight-training models
VinVL [45]8.9M110M38.230.3129.323.696.813.590.713.187.411.690.912.8--
AoANet+MA* [16]COCO-38.028.7121.021.8----------
NOC-REK* [40]COCO110M----104.714.8100.214.1100.713.0100.914.0--
RCA-NOC* [13]COCO110M37.429.6128.423.192.212.987.812.687.511.588.312.4--
ViEcap GPT2 [15]COCO124M27.224.892.918.261.110.464.39.965.08.666.29.547.913.6
InstructBLIP Vicuna-13B [11]129M188M----------121.9-82.8-
OSCAR [26]4.1M338M37.430.7127.823.583.412.081.612.077.610.681.111.7--
BLIP [24]129M446M40.4-136.7-114.915.2112.114.9115.314.4113.214.8--
BLIP-2 FlanT5-XL [25]129M1.2B42.4-144.5-123.716.3120.215.9124.815.1121.615.8--
REVEAL* T5 [20]1.3B2.1B--145.4-------123.0---
Lightweight-training models
MiniGPT4 Vicuna-13B [46]5M3.94M38.029.6129.623.499.014.8106.915.3110.814.9108.815.178.416.9
SmallCap* GPT2 [35]COCO7M37.027.9119.721.3--------60.6-
ClipCap GPT2 [29]COCO43M33.527.5113.121.184.912.166.810.949.19.665.810.9--
EVCAP* Vicuna-13BCOCO3.97M41.531.2140.124.7111.715.3119.515.6116.514.7119.315.384.418.0
Specialist SOTAs
Qwen-VL Qwen-7B [5]1.4B9.6B----------121.4-85.8-
CogVLM Vicuna-7B [41]1.5B6.5B--148.7-----132.6-128.3-94.9-
PaLI mT5-XXL [9]1.6B17B--149.1-------127.0---
PaLI-X_UL2-32B [8]2.2B55B--149.2-------126.3---
+ +Compared methods. We compare EVCAP with several SOTAs. According to the trainable parameters size, they can be divided into 1) Heavyweight-training (between 100M to 5B): VinVL [45], AoANet [16], NOC-REK [40], RCA-NOC [13], ViECap [15], InstructBLIP [11], OSCAR [26], BLIP [24], BLIP-2 [25], REVEAL [20]; 2) Lightweight-training (less than 100M): MiniGPT4 [46], SmallCap [35], ClipCap [29]; and also 3) Specialist SOTAs with huge trainable parameters (larger than 5B): QwenVL [5], CogVLM [41], PaLI [9], PaLI-X [8]. Among these methods, AoANet, NOC-REK, RCA-NOC, REVEAL, and SmallCap are retrieval-augmented captioning methods. + +# 5. Experimental Results + +# 5.1. Results on in-/out-domain benchmarks + +We assess EVCAP against SOTAs on both in-domain and out-domain benchmarks. The COCO test set can be considered as in-domain data as we only train our model on the COCO training set. Out-domain benchmarks are the NoCaps validation set and the Flickr30k test set. + +Quantitative results. Tab. 1 details our EVCAP's performance in comparison with SOTA methods. We first evaluate training costs in terms of training data sizes and parameters. Similar to various heavyweight-training models that exclude LLMs and the majority of lightweight-training models, EVCAP is trained solely on the COCO training set. It utilizes only 3.97M trainable parameters, positioning it as the second smallest, slightly larger than MiniGPT4 + +with 3.94M. Among lightweight-training models, our approach outperforms others, achieving the highest scores on all benchmarks. Despite using less training data and nearly identical trainable parameters as MiniGPT4, EVCAP significantly surpasses it, with a marked improvement of 10.5, 10.5, and 6.0 in CIDEr scores for each benchmark. When further compared with heavyweight-training models, the performance of EVCAP stands out among million-level models, nearly matching InstructBLIP, except in NoCaps. Note that since BLIP-2 does not include Vicuna checkpoints, InstructBLIP performs pre-training with Vicuna using the same procedure as BLIP-2, whereas EVCAP does not involve pre-training. Against REVEAL, which also uses external memory, our EVCAP utilizes about 1/3000 training data and 1/500 training parameters yet yields comparable results. Moreover, EVCAP's performance is on par with BLIP-2, the top-performing model with 1.2B trainable parameters. This highlights EVCAP's efficiency and effectiveness despite its significantly smaller training cost, thanks to our external visual-name memory. Regarding specialist SOTAs, they use billion-level training data and over 5B trainable parameters, so it is acceptable that they can achieve exceptionally strong performance, surpassing EVCAP by nearly 10 on all benchmarks in CIDEr scores. + +Qualitative results. Fig. 3 presents a comparison of captions generated by our EVCAP and three SOTA models across three benchmarks. The captions of SmallCap are generated by its publicly accessible demo [1]. We generate captions of MiniGPT4 and BLIP-2 using their respec + +# COCO Test + +![](images/f0c7cd5fc601182ad9ef28928e582b8e31464589b2a50de68eebf324384de677.jpg) + +GT: A green bus driving through a rural area with trees in the background. + +SmallCap: A bus driving down a street next to trees. + +MiniGPT4: A green bus is driving down the street. + +BLIP-2: A green bus driving down a road with trees in the background. + +EVCap: A green bus driving down a road next to trees. + +![](images/31f38804d88ed9718adb39319090b734d39a5158c437da97db65dc1d7b8286a4.jpg) + +GT: A woman in a blue top with headphones and two cellphones. + +SmallCap: A woman sitting in front of a laptop computer. + +MiniGPT4: A woman sitting on a couch holding two phones. + +BLIP-2: A woman sitting on a couch with two cell phones. + +EVCap: A woman wearing headphones holding two cell phones. + +# NoCaps Val + +![](images/db8065873222d29d9fccd12b292d519b520e4922206aa27ef18f77eb871b7e91.jpg) + +GT: The two guinea pigs are getting dried off in a yellow towel. + +SmallCap: A person holding a small animal in a towel. + +MiniGPT4: Two small animals are wrapped in a towel. + +BLIP-2: Two guinea pigs wrapped in a yellow towel. + +EVCap: Two guinea pigs are wrapped in a yellow towel. + +![](images/371ec2cfc3a252917fb63fa2cc8dfb97a18c771ee7895a3edfd0c8bd8b5755e9.jpg) + +GT: A computer screen showing two men sitting at a table. + +SmallCap: Two men sitting at a table with a laptop. + +MiniGPT4: A laptop computer sitting on top of a table. + +BLIP-2: A laptop computer with a picture of two men on it. + +EVCap: A laptop computer with a picture of two men on the screen. + +# Flickr30k Test + +![](images/26e81763047a240653926a483fd60d9f7d4d2836fe77c73f79e1d6146590b7b0.jpg) + +GT: A very young child in a denim baseball cap eats a green apple. + +SmallCap: A young boy holding an apple in his hand. + +MiniGPT4: A baby sitting in a high chair eating an apple. + +BLIP-2: A baby sitting in a white chair eating a green apple. + +EVCap: A toddler eating a green apple while wearing a hat. + +![](images/dd0c497f9a55af1019c1e154598255be1fcefe9d34c0a9b98cf0ca951e4b2dbb.jpg) + +GT: Two men are riding on a wooden vehicle pulled by two donkeys. + +SmallCap: A donkey pulling a cart with a man in the background. + +MiniGPT4: Two men riding on a donkey in the dirt. + +BLIP-2: Two men riding a horse drawn cart through a field. + +EVCap: Two men riding in a cart pulled by two donkeys. + +tive pre-trained models. As a lightweight and retrieval-augmented captioning method, SmallCap struggles to produce accurate captions for given images, primarily because it relies on retrieved captions laden with extraneous information. MiniGPT4, though aligned with the primary content of images, sometimes misses certain objects like "trees" and "headphones". This oversight stems from its focus on the main objects in images, without integrating additional cues for other objects provided by the retrieved object names. In contrast, the captions generated by our EVCAP are comparable to those of BLIP-2. + +# 5.2. Results on commonsense-violating data + +To explore our EVCAP's capability in describing contents in open-word settings, we further evaluate it on WHOOPS dataset, which contains commonsense-violating images. + +Quantitative results. In Tab. 2, we compare the performance of EVCAP, MiniGPT4, BLIP, and BLIP-2 on + +![](images/dfb894901fe60fd7e49981692efbf55e5abf7afaf7457c96a642077659221d9c.jpg) +Figure 3. Examples of captions generated by our EVCAP and three SOTA methods on COCO test set, NoCaps validation set, and Flickr30k test set. GT refers to the Ground Truth captions. Incorrect objects in captions are highlighted in red, while correct ones are in blue. Our EVCAP correctly generates captions across different datasets, showing performance comparable to BLIP-2. +Figure 4. Examples of captions generated by our EVCAP, EVCAP (w/ WHOOPS), and three SOTAs on WHOOPS dataset. Incorrect objects are highlighted in red, while correct ones are in blue. + +GT: This is an image of a blue Pikachu with yellow accents. + +SmallCap: A blue and white stuffed animal on a table. + +MiniGPT4: A blue poker sitting on the floor with its eyes closed. + +BLIP-2: A blue Pikachu sitting in a dark room. + +EVCap: A blue and yellow cartoon character sitting on a dark background. + +EVCap (w/ WHOOPS): A blue Pikachu is sitting on the floor. + +WHOOPS dataset. This dataset is particularly challenging due to its inclusion of unusual objects [6]. Initially, as an end-to-end trained model, our EVCAP exhibits performance similar to MiniGPT4. However, there is a noticeable improvement in the CIDEr score, after the external memory is enriched with 2396 new objects from the WHOOPS dataset, each represented by 5 synthesized images generated using stable diffusion [36]. It highlights the effectiveness of + +Table 2. Quantitative results on commonsense-violating data - WHOOPS dataset. EVCAP (w/ WHOOPS) denotes EVCAP using the memory expanded by WHOOPS objects. The results reveal the open-world comprehension ability and expandability of EVCAP. + +
MethodB@4MCS
Only pre-trained models
BLIP [24] (from [6])13-65-
BLIP-2 FlanT5-XXL [25] (from [6])31-120-
BLIP-2 FlanT5-XXL [25] (reproduced)2826.793.117.9
Finetuned models on COCO
MiniGPT4 [46]24.226.784.818.2
BLIP [24]22.925.079.317.1
BLIP-2 FlanT5-XL [25]25.827.089.118.3
End-to-end trained models on COCO
EVCAP24.126.185.317.7
EVCAP (w/ WHOOPS)24.426.186.317.8
+ +Table 3. Ablation study on components prior to the LLM decoder in EVCAP. The result of "+" Attentive fusion" demonstrates the substantial impact of the external visual-name memory. + +
MethodCOCO testNoCaps valFlickr30k test
CSCSCS
ViT + Q-Former (Baseline)134.423.9108.814.276.817.3
+ Image query tokens (Baseline+)134.123.8109.014.377.317.2
+ Attentive fusion (EVCAP)140.124.7119.315.384.418.0
+ +![](images/1a7a4498382342c04d35bddc17714b4bcfea88d424b95000a66b935bcedfc598.jpg) +Figure 5. Visualization of the captions generated from ablation study on the NoCaps validation set. We also show the retrieved object names by EVCAP, presented in gray. Incorrect objects in captions are highlighted in red, while correct ones are in blue. + +GT: A hamster on a blanket with a hand behind it. + +Baseline: A hamster sitting on a couch with a person's hand near it. + +Baseline+: A close up of a small hamster on a bed. + +EVCap: A small hamster sitting on a blanket next to a hand. + +(pillow, hamster, grizzly, gameboard, kitten, spectacles, cat, blanket, giraffe, pug-dog) + +our idea of incorporating an expandable external memory into the captioning model for open-world comprehension. + +Qualitative results. Fig. 4 illustrates the captions generated by EVCAP, EVCAP (w/WHOOPS), and three SOTAs for one image from the WHOOPS dataset. Similar to other methods except for BLIP-2, EVCAP can not recognize "blue cartoon character" as "Pikachu", while EVCAP (w/WHOOPS) successfully predicts it because of the updated memory. SmallCap and MiniGPT4 tend to generate captions with hallucinatory objects, a result of commonsense-violating contents present in the images. + +# 5.3. Detailed analysis + +Ablation study. We assess the contribution of each component prior to the LLM decoder in EVCAP by incrementally integrating the image query tokens and the attentive fusion module into our baseline model. The baseline model comprises a ViT+Q-Former, a linear layer, and a LLM decoder. + +![](images/954475c57db53357a2a3c352f997d05f6391c647ae4e380cc63d1da9b6448115.jpg) +Figure 6. Visualization of the visual features in external memory using t-SNE. For visual features in LVIS dataset's objects (blue), the related objects fall in the same cluster. After adding more visual features of synthesized images from WHOOPS' objects, new objects (red) are located at appropriate clusters (zoom-in view). + +The quantitative results are shown in Tab. 3. When employing only the baseline model (Baseline), CIDEr scores drop notably by 5.7, 10.5, and 7.6 on COCO, NoCaps, and Flickr30k, respectively. The inclusion of trainable image query tokens (Baseline+) brings a marginal improvement on NoCaps and Flickr30k. However, the performance is significantly enhanced with the addition of attentive fusion (along with the introduction of external memory), indicating the pivotal role of the external visual-name memory in the overall effectiveness of EVCAP. This is further corroborated by the qualitative results in Fig. 5, where captions from Baseline and Baseline+ inaccurately include objects like "couch" and "bed", and Baseline+ overlooks "hand". + +Exploration for external memory expandability. To demonstrate the scalability of the external memory in EVCAP, we visualize the visual features stored in LVIS external memory, and newly synthesized data from objects appearing in the WHOOPS dataset. We employ t-SNE [38] to plot visual features after reducing their dimensions to 2-D (Fig. 6). For clear visualization, we only randomly display 3649 visual features in LVIS memory, and add 479 visual features from WHOOPS objects. Among them, 35 samples are randomly labeled. The result shows a clear clustering of LVIS objects (blue) in the external memory, as well as the successful integration and appropriate localization of new objects from WHOOPS (red) into these clusters. This pattern not only confirms the distinctiveness of visual features already present in the memory but also demonstrates the potential to accurately incorporate and differentiate new objects introduced from updated data. These findings highlight our external memory's ability to expand and maintain its effectiveness even as new data is incorporated. + +Impact of external memory size. We examine the impact of external memory size in Tab. 4. On the one hand, we randomly remove $30\%$ , $60\%$ , and $90\%$ data in the external + +Table 4. Impact of the external memory size on the performance of EVCAP by evaluation under CIDEr scores. Changes in the size of external memory result in changes in performance. + +
MethodNoCaps valFlickr30k
InNearOutOverallTest
LVIS objects (EVCAP)111.7119.5116.5119.384.4
- 30% LVIS112.0119.2115.3118.885.0
- 60% LVIS111.4119.1116.2119.085.1
- 90% LVIS110.6118.2115.8118.383.6
+ WHOOPS110.7118.9116.7119.084.9
+ +![](images/c00825c28763cc8b55ca76a5d0aa5ee5fe752db79122330b7159a2860ab28ec3.jpg) +Figure 7. CIDEr scores after training EVCAP with the number of retrieved object names K from 0 to 20. The results indicate that the performance is relatively optimal when K is set to be 10. + +memory constructed from LVIS objects. The results show the performance gradually degrades on NoCaps as reducing $30\%$ and $90\%$ LVIS. Despite some unexpected increases in certain results on NoCaps (5th row) and Flickr30k (4th - 5th rows), they do not alter the overall downward trend. Similar phenomena are also noted in SmallCap [35], we speculate it is due to data distribution. On the other hand, as we infuse WHOOPS knowledge into LVIS memory, there is a slight improvement on NoCaps (out) and Flickr30k. These observations validate the model's capability to effectively retrieve object names from an updated memory, enhancing its performance in generating captions. + +Impact of the number of retrieved object names. We investigate how the number of retrieved object names K (Sec. 3.3) affect EVCAP in Fig. 7. We train the model with K from 0 to 20 and evaluate the performance under CIDER on all three benchmarks. From the results, we can find that the model works worst on the out-domain dataset (NoCaps) with zero object names. It confirms that the performance boost from Baseline+ to EVCAP (Tab. 3) is primarily attributed to the retrieval-augmented mechanism, but not the customized Q-Former itself. With more object names, performance fluctuates but improves. Furthermore, we observe that setting K to 10 yields relatively optimal overall performance, validating the choice of $\mathrm{K} = 10$ in EVCAP. + +Analysis with different decoders. To explore the influence of different LLMs decoders on our EVCAP, we experiment by substituting Vicuna-13B with GPT2 and Vicuna-7B, as detailed in Tab. 5. With GPT2 as the decoder, EVCAP still markedly surpasses other GPT2-based models, achieving + +Table 5. Analysis with different LLM decoders including GPT2, Vicuna-7B, and Vicuna-13B. The results reveal EVCAP is effective when applying it in different LLM decoders. + +
MethodLLMCOCO testNoCaps valFlickr30k test
CSCSCS
SmallCap [35]GPT2119.721.3--60.6-
ViEcap [15]GPT292.918.266.29.547.913.6
EVCAPGPT2131.023.297.613.370.616.1
MiniGPT4 [46]Vicuna-7B119.423.5108.715.773.917.2
InstructBLIP [11]Vicuna-7B--123.1-82.4-
EVCAPVicuna-7B139.024.7116.815.382.718.0
MiniGPT4 [46]Vicuna-13B129.623.4108.815.178.416.9
InstructBLIP [11]Vicuna-13B--121.9-82.8-
EVCAPVicuna-13B140.124.7119.315.384.418.0
+ +impressive gains of 11.3 and 10.0 under CIDEr on COCO and Flickr30k, compared to SmallCap. When employing Vicuna-7B, the comparison of performance trends mirrors those observed with Vicuna-13B, further attesting to the robustness and adaptability of EVCAP across different LLM decoders. Notably, both SmallCap, which retrieves captions, and our GPT2-based EVCAP, which retrieves object names, use the same GPT2 decoder. Therefore, their comparison also underscores the effectiveness of our method's object name retrieval and attentive fusion strategy. + +Limitations. First, EVCAP cannot retrieve all objects that appear in the given image due to the memory coverage limits, leading to incomplete image descriptions (Fig. 4). We will investigate integrating object detection with image captioning to enhance completeness. Second, our focus on object representation restricts consideration of other crucial captioning elements, affecting overall performance. Similar to all models trained with COCO dataset, EVCAP has limitations in generating varied styles, which is reflected in our relatively modest performance improvements in Tab. 2, compared to MiniGPT4. We will overcome it by exploring methodologies that encourage style diversity in the future. + +# 6. Conclusion + +We further advance image captioning in real-world scenarios by introducing EVCAP, a novel image captioning model with object names retrieved from an external visual-name memory. The external memory is easily expandable, allowing for effortless updates with new object visuals and names. We extensively compare EVCAP with SOTAs on various benchmarks and commonsense-violating data, demonstrating its significant superiority in performance. + +Acknowledgements. This work was supported by JSPS/MEXT KAKENHI Grant Numbers JP22H05015, JP23H03449, JP23KJ0404, and JP22K17947, and the commissioned research (No. 225) by the National Institute of Information and Communications Technology (NICT), Japan, and the Institute of AI and Beyond of the University of Tokyo, and ROIS NII Open Collaborative Research 2023-23FC01, 2024-24S1201. + +# References + +[1] https : / / huggingface . co / spaces / RitaParadaRamos/SmallCapDemo.5 +[2] Harsh Agrawal, Karan Desai, Yufei Wang, Xinlei Chen, Rishabh Jain, Mark Johnson, Dhruv Batra, Devi Parikh, Stefan Lee, and Peter Anderson. Nocaps: Novel object captioning at scale. In Proc. IEEE International Conference on Computer Vision (ICCV), 2019. 4 +[3] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: A visual language model for few-shot learning. In Advances in Neural Information Processing Systems (NeurIPS), 2022. 1 +[4] Peter Anderson, Xiaodong He, Chris Buehler, Damien Teney, Mark Johnson, Stephen Gould, and Lei Zhang. Bottom-up and top-down attention for image captioning and visual question answering. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 2 +[5] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A frontier large vision-language model with versatile abilities. arXiv preprint arXiv:2308.12966, 2023. 5 +[6] Nitzan Bitton-Guetta, Yonatan Bitton, Jack Hessel, Ludwig Schmidt, Yuval Elovici, Gabriel Stanovsky, and Roy Schwartz. Breaking common sense: Whoops! a vision-and-language benchmark of synthetic and compositional images. In Proc. IEEE International Conference on Computer Vision (ICCV), 2023. 4, 6, 7 +[7] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. In Advances in Neural Information Processing Systems (NeurIPS), 2020. 2 +[8] Xi Chen, Josip Djolonga, Piotr Padlewski, Basil Mustafa, Soravit Changpinyo, Jialin Wu, Carlos Riquelme Ruiz, Sebastian Goodman, Xiao Wang, Yi Tay, et al. PaLI-X: On scaling up a multilingual vision and language model. arXiv preprint arXiv:2305.18565, 2023. 1, 2, 5 +[9] Xi Chen, Xiao Wang, Soravit Changpinyo, AJ Piergiovanni, Piotr Padlewski, Daniel Salz, Sebastian Goodman, Adam Grycner, Basil Mustafa, Lucas Beyer, Alexander Kolesnikov, Joan Puigcerver, Nan Ding, Keran Rong, Hassan Akbari, Gaurav Mishra, Linting Xue, Ashish V Thapliyal, James Bradbury, Weicheng Kuo, Mojtaba Seyedhosseini, Chao Jia, Burcu Karagol Ayan, Carlos Riquelme Ruiz, Andreas Peter Steiner, Anelia Angelova, Xiaohua Zhai, Neil Houlsby, and Radu Soricut. PaLI: A jointly-scaled multilingual language-image model. In International Conference on Learning Representations (ICLR), 2023. 1, 2, 5 +[10] Wei-Lin Chiang, Zhuohan Li, Zi Lin, Ying Sheng, Zhanghao Wu, Hao Zhang, Lianmin Zheng, Siyuan Zhuang, Yonghao Zhuang, Joseph E Gonzalez, et al. Vicuna: An open-source chatbot impressing gpt-4 with $90\%$ chatgpt quality. See https://vicuna.lmsys.org (accessed 14 April 2023), 2023. 2, 4 +[11] Wenliang Dai, Junnan Li, Dongxu Li, Anthony Meng Huat Tiong, Junqi Zhao, Weisheng Wang, Boyang Li, Pascale + +Fung, and Steven Hoi. Instructclip: Towards general-purpose vision-language models with instruction tuning. In Advances in Neural Information Processing Systems (NeurIPS), 2023. 3, 5, 8 +[12] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations (ICLR), 2021. 2 +[13] Jiashuo Fan, Yaoyuan Liang, Leyao Liu, Shaolun Huang, and Lei Zhang. Rca-noc: Relative contrastive alignment for novel object captioning. In Proc. IEEE International Conference on Computer Vision (ICCV), 2023. 2, 5 +[14] Yuxin Fang, Wen Wang, Binhui Xie, Quan Sun, Ledell Wu, Xinggang Wang, Tiejun Huang, Xinlong Wang, and Yue Cao. Eva: Exploring the limits of masked visual representation learning at scale. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 2, 4 +[15] Junjie Fei, Teng Wang, Jinrui Zhang, Zhenyu He, Chengjie Wang, and Feng Zheng. Transferable decoding with visual entities for zero-shot image captioning. In Proc. IEEE International Conference on Computer Vision (ICCV), 2023. 2, 4, 5, 8 +[16] Zhengcong Fei. Memory-augmented image captioning. In Proc. AAAI Conference on Artificial Intelligence (AAAI), 2021. 2, 5 +[17] Agrim Gupta, Piotr Dollar, and Ross Girshick. LVIS: A dataset for large vocabulary instance segmentation. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 3 +[18] Ryuichiro Hataya, Han Bao, and Hiromi Arai. Will large-scale generative models corrupt future datasets? In Proc. IEEE International Conference on Computer Vision (ICCV), 2023. 3 +[19] Lisa Anne Hendricks, Subhashini Venugopalan, Marcus Rohrbach, Raymond Mooney, Saenko Kate, and Trevor Darrell. Deep compositional captioning: Describing novel object categories without paired training data. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 2 +[20] Ziniu Hu, Ahmet Iscen, Chen Sun, Zirui Wang, Kai-Wei Chang, Yizhou Sun, Cordelia Schmid, David A Ross, and Alireza Fathi. Reveal: Retrieval-augmented visual-language pre-training with multi-source multimodal knowledge memory. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 1, 2, 5 +[21] Jeff Johnson, Matthijs Douze, and Hervé Jégou. Billionscale similarity search with gpus. IEEE Transactions on Big Data, 7(3):535-547, 2019. 3 +[22] Andrej Karpathy and Li Fei-Fei. Deep visual-semantic alignments for generating image descriptions. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2015. 2, 4 +[23] Jacob Devlin Ming-Wei Chang Kenton and Lee Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of + +North American Chapter of the Association for Computational Linguistics: Human Language Technologies (NAACL-HLT), 2019. 4 +[24] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In Proc. International conference on machine learning (ICML), 2022. 5, 7 +[25] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In Proc. International conference on machine learning (ICML), 2023. 1, 2, 3, 4, 5, 7 +[26] Xiujun Li, Xi Yin, Chunyuan Li, Pengchuan Zhang, Xiaowei Hu, Lei Zhang, Lijuan Wang, Houdong Hu, Li Dong, Furu Wei, et al. Oscar: Object-semantics aligned pre-training for vision-language tasks. In Proc. European Conference on Computer Vision (ECCV), 2020. 2, 5 +[27] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dálár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Proc. European Conference on Computer Vision (ECCV), 2014. 4 +[28] Jiasen Lu, Jianwei Yang, Dhruv Batra, and Devi Parikh. Neural baby talk. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 2 +[29] Ron Mokady, Amir Hertz, and Amit H Bermano. Clipcap: Clip prefix for image captioning. arXiv preprint arXiv:2111.09734, 2021. 2, 3, 5 +[30] Bryan A Plummer, Liwei Wang, Chris M Cervantes, Juan C Caicedo, Julia Hockenmaier, and Svetlana Lazebnik. Flickr30k entities: Collecting region-to-phrase correspondences for richer image-to-sentence models. In Proc. IEEE International Conference on Computer Vision (ICCV), 2015. 4 +[31] Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, Ilya Sutskever, et al. Language models are unsupervised multitask learners. OpenAI blog, 1(8):9, 2019. 2 +[32] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In Proc. International conference on machine learning (ICML), 2021. 2 +[33] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J Liu. Exploring the limits of transfer learning with a unified text-to-text transformer. Journal of Machine Learning Research (JMLR), 21(1):5485-5551, 2020. 2 +[34] Rita Ramos, Desmond Elliott, and Bruno Martins. Retrievalaugmented image captioning. In Proc. Conference of the European Chapter of the Association for Computational Linguistics (EACL), 2023. 2 +[35] Rita Ramos, Bruno Martins, Desmond Elliott, and Yova Kementchedjhieva. Smallcap: Lightweight image captioning prompted with retrieval augmentation. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 1, 2, 3, 5, 8 + +[36] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 3, 6 +[37] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 2, 4 +[38] Laurens Van der Maaten and Geoffrey Hinton. Visualizing data using t-sne. Journal of machine learning research (JMLR), 9(11), 2008. 7 +[39] Subhashini Venugopalan, Lisa Anne Hendricks, Marcus Rohrbach, Raymond Mooney, Trevor Darrell, and Kate Saenko. Captioning images with diverse objects. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2017. 2 +[40] Duc Minh Vo, Hong Chen, Akihiro Sugimoto, and Hideki Nakayama. Noc-rek: Novel object captioning with retrieved vocabulary from external knowledge. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2, 5 +[41] Weihan Wang, Qingsong Lv, Wenmeng Yu, Wenyi Hong, Ji Qi, Yan Wang, Junhui Ji, Zhuoyi Yang, Lei Zhao, Xixuan Song, Jiazheng Xu, Bin Xu, Juanzi Li, Yuxiao Dong, Ming Ding, and Jie Tang. Cogvlm: Visual expert for pretrained language models. arXiv preprint arXiv:2311.03079, 2023. 4, 5 +[42] Kelvin Xu, Jimmy Ba, Ryan Kiros, Kyunghyun Cho, Aaron Courville, Ruslan Salakhudinov, Rich Zemel, and Yoshua Bengio. Show, attend and tell: Neural image caption generation with visual attention. In Proc. International conference on machine learning (ICML). 2 +[43] Zhuolin Yang, Wei Ping, Zihan Liu, Vijay Korthikanti, Weili Nie, De-An Huang, Linxi Fan, Zhiding Yu, Shiyi Lan, Bo Li, et al. Re-ViLM: Retrieval-augmented visual language model for zero and few-shot image captioning. In Findings of the Association for Computational Linguistics: EMNLP, 2023. 2 +[44] Michihiro Yasunaga, Armen Aghajanyan, Weijia Shi, Richard James, Jure Leskovec, Percy Liang, Mike Lewis, Luke Zettlemoyer, and Wen-tau Yih. Retrieval-augmented multimodal language modeling. In International Conference on Learning Representations (ICLR), 2023. 2 +[45] Pengchuan Zhang, Xiujun Li, Xiaowei Hu, Jianwei Yang, Lei Zhang, Lijuan Wang, Yejin Choi, and Jianfeng Gao. Vinvl: Revisiting visual representations in vision-language models. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 2, 5 +[46] Deyao Zhu, Jun Chen, Xiaoqian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models. arXiv preprint arXiv:2304.10592, 2023. 3, 4, 5, 7, 8 \ No newline at end of file diff --git a/2024/EVCap_ Retrieval-Augmented Image Captioning with External Visual-Name Memory for Open-World Comprehension/images.zip b/2024/EVCap_ Retrieval-Augmented Image Captioning with External Visual-Name Memory for Open-World Comprehension/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..e557ba1cb88db287619bf59455123ff9667a968a --- /dev/null +++ b/2024/EVCap_ Retrieval-Augmented Image Captioning with External Visual-Name Memory for Open-World Comprehension/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5fcf30f316157554f8409f2af859ad3e0cf453ca3a9933f8cc02f6d44600eb25 +size 559196 diff --git a/2024/EVCap_ Retrieval-Augmented Image Captioning with External Visual-Name Memory for Open-World Comprehension/layout.json b/2024/EVCap_ Retrieval-Augmented Image Captioning with External Visual-Name Memory for Open-World Comprehension/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..c50f4d38ce5f5543ba0f231cd340748426f9f174 --- /dev/null +++ b/2024/EVCap_ Retrieval-Augmented Image Captioning with External Visual-Name Memory for Open-World Comprehension/layout.json @@ -0,0 +1,8916 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 82, + 103, + 512, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 82, + 103, + 512, + 140 + ], + "spans": [ + { + "bbox": [ + 82, + 103, + 512, + 140 + ], + "type": "text", + "content": "EVCAP: Retrieval-Augmented Image Captioning with External Visual-Name Memory for Open-World Comprehension" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 121, + 160, + 475, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 121, + 160, + 475, + 175 + ], + "spans": [ + { + "bbox": [ + 121, + 160, + 475, + 175 + ], + "type": "text", + "content": "Jiaxuan " + }, + { + "bbox": [ + 121, + 160, + 475, + 175 + ], + "type": "inline_equation", + "content": "\\mathrm{Li}^{1*}" + }, + { + "bbox": [ + 121, + 160, + 475, + 175 + ], + "type": "text", + "content": ", Duc Minh " + }, + { + "bbox": [ + 121, + 160, + 475, + 175 + ], + "type": "inline_equation", + "content": "\\mathrm{Vo}^{1*}" + }, + { + "bbox": [ + 121, + 160, + 475, + 175 + ], + "type": "text", + "content": ", Akihiro Sugimoto" + }, + { + "bbox": [ + 121, + 160, + 475, + 175 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 121, + 160, + 475, + 175 + ], + "type": "text", + "content": ", Hideki Nakayama" + }, + { + "bbox": [ + 121, + 160, + 475, + 175 + ], + "type": "inline_equation", + "content": "^{1}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 119, + 175, + 474, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 119, + 175, + 474, + 190 + ], + "spans": [ + { + "bbox": [ + 119, + 175, + 474, + 190 + ], + "type": "text", + "content": "1The University of Tokyo, Japan 2National Institute of Informatics, Japan" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 73, + 191, + 520, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 73, + 191, + 520, + 204 + ], + "spans": [ + { + "bbox": [ + 73, + 191, + 520, + 204 + ], + "type": "text", + "content": "{li, vm Duc}@nlab.ci.i.u-tokyo.ac.jp sugimoto@nii.ac.jp nakayama@ci.i.u-tokyo.ac.jp" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 143, + 231, + 192, + 243 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 231, + 192, + 243 + ], + "spans": [ + { + "bbox": [ + 143, + 231, + 192, + 243 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 256, + 290, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 256, + 290, + 518 + ], + "spans": [ + { + "bbox": [ + 46, + 256, + 290, + 518 + ], + "type": "text", + "content": "Large language models (LLMs)-based image captioning has the capability of describing objects not explicitly observed in training data; yet novel objects occur frequently, necessitating the requirement of sustaining up-to-date object knowledge for open-world comprehension. Instead of relying on large amounts of data and/or scaling up network parameters, we introduce a highly effective retrieval-augmented image captioning method that prompts LLMs with object names retrieved from External Visual-name memory (EVCAP). We build ever-changing object knowledge memory using objects' visuals and names, enabling us to (i) update the memory at a minimal cost and (ii) effortlessly augment LLMs with retrieved object names by utilizing a lightweight and fast-to-train model. Our model, which was trained only on the COCO dataset, can adapt to out-of-domain without requiring additional fine-tuning or re-training. Our experiments conducted on benchmarks and synthetic commonsense-violating data show that EVCAP, with only 3.97M trainable parameters, exhibits superior performance compared to other methods based on frozen pre-trained LLMs. Its performance is also competitive to specialist SOTAs that require extensive training." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 541, + 128, + 554 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 541, + 128, + 554 + ], + "spans": [ + { + "bbox": [ + 47, + 541, + 128, + 554 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 562, + 287, + 681 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 562, + 287, + 681 + ], + "spans": [ + { + "bbox": [ + 46, + 562, + 287, + 681 + ], + "type": "text", + "content": "Advanced image captioning based on large language models (LLMs) [3, 8, 9, 25] has focused on the approach using big-scale models trained on ever-increasingly large-scale datasets, which is no longer viable. This is because the computational cost to train the models increases exponentially and, more importantly, updating training data is almost impossible to keep pace with the growth of novel objects in our daily lives. Sustaining ever-changing object knowledge with a reasonable cost is a pressing concern in LLMs-based models to truly unlock open-world comprehension." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 59, + 681, + 287, + 694 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 681, + 287, + 694 + ], + "spans": [ + { + "bbox": [ + 59, + 681, + 287, + 694 + ], + "type": "text", + "content": "Retrieval-augmented image captioning [20, 35] is" + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 309, + 231, + 545, + 430 + ], + "blocks": [ + { + "bbox": [ + 309, + 231, + 545, + 430 + ], + "lines": [ + { + "bbox": [ + 309, + 231, + 545, + 430 + ], + "spans": [ + { + "bbox": [ + 309, + 231, + 545, + 430 + ], + "type": "image", + "image_path": "71d2fc87b5ba24b951155fde0165bac7dc20438920282706b7fa1b5866e3486c.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 438, + 547, + 596 + ], + "lines": [ + { + "bbox": [ + 304, + 438, + 547, + 596 + ], + "spans": [ + { + "bbox": [ + 304, + 438, + 547, + 596 + ], + "type": "text", + "content": "Figure 1. Overall comparison of our EVCAP and SOTAs. (Upper) Generated captions by SmallCap, BLIP-2, and our EVCAP for a commonsense-violating image from the WHOOPS dataset. " + }, + { + "bbox": [ + 304, + 438, + 547, + 596 + ], + "type": "inline_equation", + "content": "\\times" + }, + { + "bbox": [ + 304, + 438, + 547, + 596 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 438, + 547, + 596 + ], + "type": "inline_equation", + "content": "\\checkmark" + }, + { + "bbox": [ + 304, + 438, + 547, + 596 + ], + "type": "text", + "content": " indicate incorrect and correct predictions, respectively. Incorrect objects in captions are highlighted in red, while correct ones are in blue. SmallCap and BLIP-2 give incorrect predictions for \"ice skates\" and \"wood floor\", respectively, while our EVCAP utilizes an external visual-name memory to enhance attention to objects within the image, leading to superior performance for image captioning. (Lower) Comparison of the number of trainable parameters, CIDEr score on COCO and NoCaps datasets. The size of each circle reflects the log number of trainable parameters. EVCAP (3.97M) has less trainable parameters than others while achieving comparable results with SOTAs at scale." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 618, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 618, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 618, + 547, + 715 + ], + "type": "text", + "content": "emerging as an alternative since it considerably reduces training costs in both time and data while producing encouraging results. Nonetheless, with their huge datastore, it is obvious that LLMs would imitate the given texts, limiting their ability to describe open-world objects properly. For instance, SmallCap [35] considers the words \"skateboard\" and \"wooden floor\" to be a pair regardless of visual appearances containing a commonsense-violating pair" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 701, + 290, + 710 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 701, + 290, + 710 + ], + "spans": [ + { + "bbox": [ + 61, + 701, + 290, + 710 + ], + "type": "text", + "content": "*Equal contributions. Code is available at https://jiaxuan-li.github.io/EVCap." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13733" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 251 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 251 + ], + "type": "text", + "content": "of \"ice skates\" and \"wood floor\" (Fig. 1, upper). Additionally, prompting the LLMs given a lot of retrieved texts becomes cumbersome, requiring more trainable parameters. Fig. 1 (lower) shows that the CIDEr scores obtained by a lightweight SmallCap [35] with 43M trainable parameters are far away from those obtained by a heavy REVEAL [20] with 2.1B trainable parameters. Beyond that, due to the frequent occurrence of new objects, access to their sample texts is not always feasible, making the memory utilized in [20, 35] difficult to grow. We thus aim to streamline the external memory used in previous work [20, 35] by storing a sufficiently small amount of object information. And, of course, not only does the model not stereotype the example sentences, but the number of trainable parameters would be reduced drastically as a result of the causation (Fig. 1)." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 251, + 289, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 251, + 289, + 548 + ], + "spans": [ + { + "bbox": [ + 46, + 251, + 289, + 548 + ], + "type": "text", + "content": "We follow [13, 40] to construct a key-value memory where the key is represented by object's features, and the value corresponds to object's name. Unlike [13, 40], which rely on object definition as the key, our method leverages the visual appearance of the object as the key because of the abundance of object images readily available on the internet. We propose an external visual-name memory tailored for ease of expansion and cost-effectiveness in upholding up-to-date object information. We present a highly effective retrieval-augmented LLMs-based image captioning method, called EVCAP, that prompts frozen LLMs with object names retrieved from our proposed memory for open-world comprehension. EVCAP contains a frozen image encoder ViT [14] and Q-Former [25] with trainable image query tokens for object retrieval, an attentive fusion module, a trainable linear layer for mapping between vision and language latent spaces, and a frozen LLM decoder [10] for generating captions. Specifically, the attentive fusion module feeds retrieved object names and visual features into a customized frozen Q-Former using trainable object name query tokens to implicitly reduce the presence of superfluous object names. As a result, EVCAP amounts to only 3.97M trainable parameters. Once trained, the model can be adapted to new domains and large-scale data without further fine-tuning or re-training. Our contributions are as follows:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 549, + 287, + 609 + ], + "type": "list", + "angle": 0, + "index": 4, + "blocks": [ + { + "bbox": [ + 47, + 549, + 287, + 585 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 549, + 287, + 585 + ], + "spans": [ + { + "bbox": [ + 47, + 549, + 287, + 585 + ], + "type": "text", + "content": "- We provide an extensible external visual-name memory with minimal but useful object information, which enables LLMs-based models to comprehend the open world." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 586, + 287, + 609 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 586, + 287, + 609 + ], + "spans": [ + { + "bbox": [ + 47, + 586, + 287, + 609 + ], + "type": "text", + "content": "- We present a highly efficacious retrieval-augmented image captioning EVCAP with 3.97M trainable parameters." + } + ] + } + ], + "index": 3 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 46, + 611, + 287, + 659 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 611, + 287, + 659 + ], + "spans": [ + { + "bbox": [ + 46, + 611, + 287, + 659 + ], + "type": "text", + "content": "On in-/out-domain benchmarks and synthetic commonsense-violating dataset, EVCAP trained solely on COCO dataset competes with other lightweight methods by a margin while being on par with other specialist SOTAs." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 670, + 134, + 682 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 670, + 134, + 682 + ], + "spans": [ + { + "bbox": [ + 47, + 670, + 134, + 682 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 689, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 689, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 689, + 287, + 715 + ], + "type": "text", + "content": "Image captioning aims to describe the contents of a given image. It can be roughly divided into two approaches: non-" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 72, + 545, + 262 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 262 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 262 + ], + "type": "text", + "content": "LLMs-based methods and LLMs-based ones. The former approaches [4, 22, 42] typically employ a visual encoder and a language decoder in an end-to-end fashion to generate captions. However, they are incapable of describing open-world objects. The latter one leverages pre-trained large-scale vision models (CLIP [32], ViT [12]) and LLMs (GPTs [7, 31], T5 [33], LLaMA [37]) by bridging the gap between two modalities using either pre-training with largescale data or the learned mapper or prompt techniques. LLMs-based models [8, 9, 25, 29] demonstrate advancements in image captioning challenges, allowing the capacity to describe anything as long as pre-trained vision models can recognize it. Our method belongs to the LLMs-based approaches, but instead of relying fully on the pre-trained vision model, we use object names retrieved from the external memory to augment LLMs-based image captioning." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 262, + 546, + 429 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 262, + 546, + 429 + ], + "spans": [ + { + "bbox": [ + 304, + 262, + 546, + 429 + ], + "type": "text", + "content": "Novel object captioning is a branch of image captioning that describes images containing objects that were not seen during training. Non-LLMs-based methods explore more objects by learning from unpaired image-sentence sources (DCC [19], NOC [39]) or rely on novel object detectors to recognize novel concepts (NBT [28], OSCAR [26] and VinVL [45]). LLMs-based methods such as ViECap [15] leverage the pre-trained CLIP [32] to obtain object entities. Nevertheless, the cut-off in training time of the pre-trained object detector or CLIP prevents it from detecting novel objects that arise quickly in reality. Unlike earlier work, we can readily update our recognition of novel concepts by adding them to external memory, ensuring that we keep any new objects from the past and even the future." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 429, + 546, + 608 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 429, + 546, + 608 + ], + "spans": [ + { + "bbox": [ + 304, + 429, + 546, + 608 + ], + "type": "text", + "content": "Retrieval-augmented image captioning is a recently popular approach that augments the captioning model with retrieved information for better open-world understanding. AoANet [16] uses a memory bank of image-sentence pairs and target words. SmallCap [35] employs image-to-text retrieval to obtain sampled captions from a captions dataset. RA-CM3 [44] retrieves documents from an external memory of a mixture of text and image via a dense multimodal retriever. EXTRA [34] and Re-ViLM [43] exploit the similarity of the input image and vision candidates to retrieve captions. Different from the previous methods, our external memory contains visual-name pairs to avoid redundant information in the external captions/documents. In addition, we use an attentive fusion module to mitigate the effects of irrelevant retrieved object names on caption generation." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 616, + 413, + 629 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 616, + 413, + 629 + ], + "spans": [ + { + "bbox": [ + 306, + 616, + 413, + 629 + ], + "type": "text", + "content": "3. Proposed EVCAP" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 635, + 397, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 635, + 397, + 647 + ], + "spans": [ + { + "bbox": [ + 306, + 635, + 397, + 647 + ], + "type": "text", + "content": "3.1. Idea of EVCAP" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 654, + 545, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 545, + 715 + ], + "type": "text", + "content": "We aim to build a retrieval-augmented LLMs-based image captioning model with a sufficiently small yet informative external memory. It involves two challenges: (1) constructing an expandable external memory and (2) building an effective LLMs-based model using retrieved object names." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "13734" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 72, + 68, + 521, + 309 + ], + "blocks": [ + { + "bbox": [ + 72, + 68, + 521, + 309 + ], + "lines": [ + { + "bbox": [ + 72, + 68, + 521, + 309 + ], + "spans": [ + { + "bbox": [ + 72, + 68, + 521, + 309 + ], + "type": "image", + "image_path": "ba4975506959fdf05710a3b6655208a431a0dd4cb0e004e06a6ab95c19e0379c.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 316, + 548, + 384 + ], + "lines": [ + { + "bbox": [ + 46, + 316, + 548, + 384 + ], + "spans": [ + { + "bbox": [ + 46, + 316, + 548, + 384 + ], + "type": "text", + "content": "Figure 2. Schematic of our proposed EVCAP. It consists of an external visual-name memory with image embeddings and object names (upper), a frozen ViT and Q-Former equipped with trainable image query tokens, an attentive fusion module developed by a customized frozen Q-Former and trainable object name query tokens, and a frozen LLM with a trainable linear layer (lower). The ViT and Q-Former extract learned visual features from the input image, which are then used to retrieve object names from the external memory. These retrieved object names and learned visual features undergo cross-attention in the customized Q-Former, creating refined object name features. Finally, the object name features combined with visual features are fed into the LLM post a linear layer for generating captions." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 403, + 287, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 403, + 287, + 521 + ], + "spans": [ + { + "bbox": [ + 46, + 403, + 287, + 521 + ], + "type": "text", + "content": "As discussed above, challenge (1) can be resolved by utilizing the visual appearance of objects. However, if we restrict our memory to only a visual-name pair for each object, our memory will be lacking in diversity. Therefore, we gather several images for each target object. Additionally, we keep the synthetic images in our memory to avoid the harm that synthetic images might cause to our method, as pointed out in [18]. With the capability to collect images from the internet, EVCAP can be easily expanded to include novel objects from the real world effortlessly." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 522, + 288, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 522, + 288, + 666 + ], + "spans": [ + { + "bbox": [ + 46, + 522, + 288, + 666 + ], + "type": "text", + "content": "We base our method on the frozen pre-trained vision model and LLM with several trainable layers (Fig. 2), giving in a model that is cheap to train. To guide the LLM, we adopt a recently popular approach called prompting as in [11, 25, 29, 35, 46]. We begin by matching the learned visual features from the input image with image embeddings stored in memory, retrieving object names. We also introduce an attentive fusion module designed to implicitly remove irrelevant retrieved names. Finally, following the attentive fusion, we combine the learned visual features and object name features to form a prompt for the LLM to generate a caption, thus addressing challenge (2)." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 671, + 214, + 685 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 671, + 214, + 685 + ], + "spans": [ + { + "bbox": [ + 47, + 671, + 214, + 685 + ], + "type": "text", + "content": "3.2. External visual-name memory" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 689, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 689, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 689, + 288, + 715 + ], + "type": "text", + "content": "To build the external visual-name memory, we first collect image-name pairs from the external data source. After" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 305, + 403, + 545, + 427 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 403, + 545, + 427 + ], + "spans": [ + { + "bbox": [ + 305, + 403, + 545, + 427 + ], + "type": "text", + "content": "that, we encode these images into image embeddings, which serve as keys in memory, and use their names as values." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 428, + 546, + 606 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 428, + 546, + 606 + ], + "spans": [ + { + "bbox": [ + 304, + 428, + 546, + 606 + ], + "type": "text", + "content": "External data source. We utilize object images from LVIS dataset [17] to construct our external visual-name memory " + }, + { + "bbox": [ + 304, + 428, + 546, + 606 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 304, + 428, + 546, + 606 + ], + "type": "text", + "content": ". Specifically, we use 1203 objects in LVIS, where we randomly select from one to ten images for each object, amounting to 8581 object images. Furthermore, as mentioned in Sec. 3.1, we also incorporate synthetic images in our memory construction. Using stable diffusion [36], we generate five additional images for each object, with a prompt of \"a photo of {object name}\", resulting in a total of " + }, + { + "bbox": [ + 304, + 428, + 546, + 606 + ], + "type": "inline_equation", + "content": "M = 14596" + }, + { + "bbox": [ + 304, + 428, + 546, + 606 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 304, + 428, + 546, + 606 + ], + "type": "inline_equation", + "content": "8581 + 5 \\times 1203" + }, + { + "bbox": [ + 304, + 428, + 546, + 606 + ], + "type": "text", + "content": ") images. Each object image " + }, + { + "bbox": [ + 304, + 428, + 546, + 606 + ], + "type": "inline_equation", + "content": "X^i" + }, + { + "bbox": [ + 304, + 428, + 546, + 606 + ], + "type": "text", + "content": " is associated with an object name " + }, + { + "bbox": [ + 304, + 428, + 546, + 606 + ], + "type": "inline_equation", + "content": "v^i" + }, + { + "bbox": [ + 304, + 428, + 546, + 606 + ], + "type": "text", + "content": ". Note that many object images may share the same object name. For the sake of simplicity, we may regard each image as corresponding to a single name. In summary, we have " + }, + { + "bbox": [ + 304, + 428, + 546, + 606 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 304, + 428, + 546, + 606 + ], + "type": "text", + "content": " image-name pairs " + }, + { + "bbox": [ + 304, + 428, + 546, + 606 + ], + "type": "inline_equation", + "content": "\\{(X^i,v^i)\\}_{i=1}^M" + }, + { + "bbox": [ + 304, + 428, + 546, + 606 + ], + "type": "text", + "content": " for external memory construction." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 607, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 607, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 607, + 547, + 714 + ], + "type": "text", + "content": "External memory construction. For each image " + }, + { + "bbox": [ + 304, + 607, + 547, + 714 + ], + "type": "inline_equation", + "content": "X^i" + }, + { + "bbox": [ + 304, + 607, + 547, + 714 + ], + "type": "text", + "content": ", we use a frozen vision encoder " + }, + { + "bbox": [ + 304, + 607, + 547, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{E}(\\cdot)" + }, + { + "bbox": [ + 304, + 607, + 547, + 714 + ], + "type": "text", + "content": " (see Sec. 3.3 for detail) to project it into 32 embeddings with the size of " + }, + { + "bbox": [ + 304, + 607, + 547, + 714 + ], + "type": "inline_equation", + "content": "1 \\times 768" + }, + { + "bbox": [ + 304, + 607, + 547, + 714 + ], + "type": "text", + "content": " each: " + }, + { + "bbox": [ + 304, + 607, + 547, + 714 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{k}_1^i, \\mathbf{k}_2^i, \\dots, \\mathbf{k}_{32}^i\\} = \\mathcal{E}(X^i)" + }, + { + "bbox": [ + 304, + 607, + 547, + 714 + ], + "type": "text", + "content": ". We then average 32 embeddings to produce a single embedding " + }, + { + "bbox": [ + 304, + 607, + 547, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{k}^i" + }, + { + "bbox": [ + 304, + 607, + 547, + 714 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 304, + 607, + 547, + 714 + ], + "type": "inline_equation", + "content": "1 \\times 768" + }, + { + "bbox": [ + 304, + 607, + 547, + 714 + ], + "type": "text", + "content": ") that serves as the key (visual) in " + }, + { + "bbox": [ + 304, + 607, + 547, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 304, + 607, + 547, + 714 + ], + "type": "text", + "content": ". The paired object name " + }, + { + "bbox": [ + 304, + 607, + 547, + 714 + ], + "type": "inline_equation", + "content": "v^i" + }, + { + "bbox": [ + 304, + 607, + 547, + 714 + ], + "type": "text", + "content": " acts as its value (name). Consequently, we have the visual-name memory " + }, + { + "bbox": [ + 304, + 607, + 547, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{M} = \\{(\\mathbf{k}^i, v^i)\\}_{i=1}^M" + }, + { + "bbox": [ + 304, + 607, + 547, + 714 + ], + "type": "text", + "content": " which is indexed using FAISS [21], facilitating rapid searches based on sim" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13735" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 96 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 96 + ], + "type": "text", + "content": "ilarity measures. Our memory can be expanded effortlessly by gathering additional visual-name pairs (see Sec. 5.3)." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 102, + 177, + 114 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 102, + 177, + 114 + ], + "spans": [ + { + "bbox": [ + 47, + 102, + 177, + 114 + ], + "type": "text", + "content": "3.3. Object names retrieval" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 121, + 286, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 121, + 286, + 239 + ], + "spans": [ + { + "bbox": [ + 46, + 121, + 286, + 239 + ], + "type": "text", + "content": "Image encoding. We feed a frozen vision encoder " + }, + { + "bbox": [ + 46, + 121, + 286, + 239 + ], + "type": "inline_equation", + "content": "\\mathcal{E}" + }, + { + "bbox": [ + 46, + 121, + 286, + 239 + ], + "type": "text", + "content": " image " + }, + { + "bbox": [ + 46, + 121, + 286, + 239 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 46, + 121, + 286, + 239 + ], + "type": "text", + "content": " and image query tokens " + }, + { + "bbox": [ + 46, + 121, + 286, + 239 + ], + "type": "inline_equation", + "content": "\\mathbf{T}_{\\mathrm{img}}" + }, + { + "bbox": [ + 46, + 121, + 286, + 239 + ], + "type": "text", + "content": " to produce visual features " + }, + { + "bbox": [ + 46, + 121, + 286, + 239 + ], + "type": "inline_equation", + "content": "\\mathcal{Q}" + }, + { + "bbox": [ + 46, + 121, + 286, + 239 + ], + "type": "text", + "content": ". To enable the retrieval process controllable, we make image query tokens to be trainable. Thus, the image encoding process can be summarized as " + }, + { + "bbox": [ + 46, + 121, + 286, + 239 + ], + "type": "inline_equation", + "content": "\\mathcal{Q} = \\mathcal{E}(X,\\mathbf{T}_{\\mathrm{img}})" + }, + { + "bbox": [ + 46, + 121, + 286, + 239 + ], + "type": "text", + "content": ". We use the BLIP-2 pre-trained vision encoder [25], which consists of a pre-trained vision transformer ViT-g [14] outputting image features " + }, + { + "bbox": [ + 46, + 121, + 286, + 239 + ], + "type": "inline_equation", + "content": "(257\\times 1408)" + }, + { + "bbox": [ + 46, + 121, + 286, + 239 + ], + "type": "text", + "content": ", and a Q-Former receiving image features producing " + }, + { + "bbox": [ + 46, + 121, + 286, + 239 + ], + "type": "inline_equation", + "content": "|\\mathcal{Q}| = 32" + }, + { + "bbox": [ + 46, + 121, + 286, + 239 + ], + "type": "text", + "content": " learned visual features " + }, + { + "bbox": [ + 46, + 121, + 286, + 239 + ], + "type": "inline_equation", + "content": "(1\\times 768" + }, + { + "bbox": [ + 46, + 121, + 286, + 239 + ], + "type": "text", + "content": " each). We denote " + }, + { + "bbox": [ + 46, + 121, + 286, + 239 + ], + "type": "inline_equation", + "content": "\\mathcal{Q} = \\{\\mathbf{q}_1,\\mathbf{q}_2,\\dots,\\mathbf{q}_{32}\\}" + }, + { + "bbox": [ + 46, + 121, + 286, + 239 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 240, + 286, + 318 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 240, + 286, + 318 + ], + "spans": [ + { + "bbox": [ + 46, + 240, + 286, + 318 + ], + "type": "text", + "content": "Retrieval. Having obtained " + }, + { + "bbox": [ + 46, + 240, + 286, + 318 + ], + "type": "inline_equation", + "content": "\\mathcal{Q}" + }, + { + "bbox": [ + 46, + 240, + 286, + 318 + ], + "type": "text", + "content": ", we calculate the cosine similarity between the query " + }, + { + "bbox": [ + 46, + 240, + 286, + 318 + ], + "type": "inline_equation", + "content": "\\mathbf{q}_j \\in \\mathcal{Q}" + }, + { + "bbox": [ + 46, + 240, + 286, + 318 + ], + "type": "text", + "content": " and the key " + }, + { + "bbox": [ + 46, + 240, + 286, + 318 + ], + "type": "inline_equation", + "content": "\\mathbf{k}^i \\in \\mathcal{M}" + }, + { + "bbox": [ + 46, + 240, + 286, + 318 + ], + "type": "text", + "content": ". The similarity calculation is given by " + }, + { + "bbox": [ + 46, + 240, + 286, + 318 + ], + "type": "inline_equation", + "content": "\\mathrm{SIM}(\\mathbf{q}_j, \\mathbf{k}^i) = \\frac{\\mathbf{q}_j^\\top \\mathbf{k}^i}{\\|\\mathbf{q}_j\\| \\|\\mathbf{k}^i\\|}" + }, + { + "bbox": [ + 46, + 240, + 286, + 318 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 240, + 286, + 318 + ], + "type": "inline_equation", + "content": "i \\in [1, M]" + }, + { + "bbox": [ + 46, + 240, + 286, + 318 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 240, + 286, + 318 + ], + "type": "inline_equation", + "content": "j \\in [1, 32]" + }, + { + "bbox": [ + 46, + 240, + 286, + 318 + ], + "type": "text", + "content": ". Given each " + }, + { + "bbox": [ + 46, + 240, + 286, + 318 + ], + "type": "inline_equation", + "content": "\\mathbf{q}_j" + }, + { + "bbox": [ + 46, + 240, + 286, + 318 + ], + "type": "text", + "content": ", we select one key with the highest similarity score, resulting in 32 key-value candidates " + }, + { + "bbox": [ + 46, + 240, + 286, + 318 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{k}_j^{\\mathrm{best}}, v_j^{\\mathrm{best}}\\}_{j=1}^{32}" + }, + { + "bbox": [ + 46, + 240, + 286, + 318 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 318, + 286, + 388 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 318, + 286, + 388 + ], + "spans": [ + { + "bbox": [ + 46, + 318, + 286, + 388 + ], + "type": "text", + "content": "After that, we filter out candidates with repeated object names (values), and then select the top-K values. In particular, we determine the index " + }, + { + "bbox": [ + 46, + 318, + 286, + 388 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 46, + 318, + 286, + 388 + ], + "type": "text", + "content": " from the key that has the highest SIM score. These selected values " + }, + { + "bbox": [ + 46, + 318, + 286, + 388 + ], + "type": "inline_equation", + "content": "v_{j}^{\\mathrm{best}}" + }, + { + "bbox": [ + 46, + 318, + 286, + 388 + ], + "type": "text", + "content": " are redefined as the new notation " + }, + { + "bbox": [ + 46, + 318, + 286, + 388 + ], + "type": "inline_equation", + "content": "v_{l}" + }, + { + "bbox": [ + 46, + 318, + 286, + 388 + ], + "type": "text", + "content": " in the retrieved top-K object names for the input image, which can be summarized as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 83, + 399, + 249, + 418 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 83, + 399, + 249, + 418 + ], + "spans": [ + { + "bbox": [ + 83, + 399, + 249, + 418 + ], + "type": "interline_equation", + "content": "\\{\\mathbf {k} _ {j} ^ {\\text {b e s t}}, v _ {j} ^ {\\text {b e s t}} \\} = \\arg \\max _ {\\mathbf {k} ^ {i}} \\operatorname {S I M} \\left(\\mathbf {q} _ {j}, \\mathbf {k} ^ {i}\\right),", + "image_path": "bce47cfee45f34852f9d53aa4aefb2c7a5357697fe825ec13c39d06ece685bb8.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 82, + 427, + 252, + 447 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 82, + 427, + 252, + 447 + ], + "spans": [ + { + "bbox": [ + 82, + 427, + 252, + 447 + ], + "type": "interline_equation", + "content": "j = \\arg \\max _ {j} \\mathrm {S I M} (\\mathbf {q} _ {j}, \\mathbf {k} _ {j} ^ {\\mathrm {b e s t}}), v _ {l} \\gets v _ {j} ^ {\\mathrm {b e s t}},", + "image_path": "e1058e350bc4cbda0a21aa7b66707dab039b628e2e13f1331c8e182dc449ce7a.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 450, + 286, + 476 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 450, + 286, + 476 + ], + "spans": [ + { + "bbox": [ + 46, + 450, + 286, + 476 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 450, + 286, + 476 + ], + "type": "inline_equation", + "content": "l \\in [1, \\mathrm{K}]" + }, + { + "bbox": [ + 46, + 450, + 286, + 476 + ], + "type": "text", + "content": ". As a result, the retrieved top-" + }, + { + "bbox": [ + 46, + 450, + 286, + 476 + ], + "type": "inline_equation", + "content": "\\mathbf{K}" + }, + { + "bbox": [ + 46, + 450, + 286, + 476 + ], + "type": "text", + "content": " object names are " + }, + { + "bbox": [ + 46, + 450, + 286, + 476 + ], + "type": "inline_equation", + "content": "\\{v_l\\}_{l=1}^{\\mathrm{K}}" + }, + { + "bbox": [ + 46, + 450, + 286, + 476 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 481, + 144, + 493 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 481, + 144, + 493 + ], + "spans": [ + { + "bbox": [ + 47, + 481, + 144, + 493 + ], + "type": "text", + "content": "3.4. Attentive fusion" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 499, + 286, + 534 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 499, + 286, + 534 + ], + "spans": [ + { + "bbox": [ + 46, + 499, + 286, + 534 + ], + "type": "text", + "content": "Since the object names obtained from the retrieval process may be redundant, we develop an attentive fusion module to selectively distill object name features." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 535, + 286, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 535, + 286, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 535, + 286, + 714 + ], + "type": "text", + "content": "The retrieved object names " + }, + { + "bbox": [ + 46, + 535, + 286, + 714 + ], + "type": "inline_equation", + "content": "\\{v_{l}\\}_{l = 1}^{\\mathrm{K}}" + }, + { + "bbox": [ + 46, + 535, + 286, + 714 + ], + "type": "text", + "content": " are concatenated together into a sequence " + }, + { + "bbox": [ + 46, + 535, + 286, + 714 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 46, + 535, + 286, + 714 + ], + "type": "text", + "content": ", each separated by a delimiter: " + }, + { + "bbox": [ + 46, + 535, + 286, + 714 + ], + "type": "inline_equation", + "content": "S = \\{v_{1},[\\mathrm{SEP}],v_{2},[\\mathrm{SEP}],\\dots ,[\\mathrm{SEP}],v_{\\mathrm{K}}\\}" + }, + { + "bbox": [ + 46, + 535, + 286, + 714 + ], + "type": "text", + "content": ". The sequence " + }, + { + "bbox": [ + 46, + 535, + 286, + 714 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 46, + 535, + 286, + 714 + ], + "type": "text", + "content": " and visual features " + }, + { + "bbox": [ + 46, + 535, + 286, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{Q}" + }, + { + "bbox": [ + 46, + 535, + 286, + 714 + ], + "type": "text", + "content": " are fed into a customized Q-Former " + }, + { + "bbox": [ + 46, + 535, + 286, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{F}(\\cdot)" + }, + { + "bbox": [ + 46, + 535, + 286, + 714 + ], + "type": "text", + "content": ", which is constructed from the frozen pretrained Q-Former as we used in vision encoder " + }, + { + "bbox": [ + 46, + 535, + 286, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{E}" + }, + { + "bbox": [ + 46, + 535, + 286, + 714 + ], + "type": "text", + "content": ". Nonetheless, in order to enable object names to get attention from visual features, we switch the image embedding port and the text instruction port (see the supplement for architecture detail). Like in the image encoding process in Sec. 3.3, we make the object name query tokens " + }, + { + "bbox": [ + 46, + 535, + 286, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{T}_{\\mathrm{obj}}" + }, + { + "bbox": [ + 46, + 535, + 286, + 714 + ], + "type": "text", + "content": " learnable during training to assist in learning object name features related to the caption. The size of " + }, + { + "bbox": [ + 46, + 535, + 286, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{T}_{\\mathrm{obj}}" + }, + { + "bbox": [ + 46, + 535, + 286, + 714 + ], + "type": "text", + "content": " is " + }, + { + "bbox": [ + 46, + 535, + 286, + 714 + ], + "type": "inline_equation", + "content": "P\\times 768" + }, + { + "bbox": [ + 46, + 535, + 286, + 714 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 535, + 286, + 714 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 46, + 535, + 286, + 714 + ], + "type": "text", + "content": " indicates the number of object name query tokens. We get the object name features " + }, + { + "bbox": [ + 46, + 535, + 286, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{V} = \\mathcal{F}(S,\\mathcal{Q},\\mathbf{T}_{\\mathrm{obj}})" + }, + { + "bbox": [ + 46, + 535, + 286, + 714 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 72, + 419, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 419, + 84 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 419, + 84 + ], + "type": "text", + "content": "3.5. Caption generation" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 91, + 545, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 91, + 545, + 185 + ], + "spans": [ + { + "bbox": [ + 304, + 91, + 545, + 185 + ], + "type": "text", + "content": "Before inputting the visual features " + }, + { + "bbox": [ + 304, + 91, + 545, + 185 + ], + "type": "inline_equation", + "content": "\\mathcal{Q}" + }, + { + "bbox": [ + 304, + 91, + 545, + 185 + ], + "type": "text", + "content": " and object name features " + }, + { + "bbox": [ + 304, + 91, + 545, + 185 + ], + "type": "inline_equation", + "content": "\\nu" + }, + { + "bbox": [ + 304, + 91, + 545, + 185 + ], + "type": "text", + "content": " into the LLM decoder, we concatenate " + }, + { + "bbox": [ + 304, + 91, + 545, + 185 + ], + "type": "inline_equation", + "content": "(\\oplus)" + }, + { + "bbox": [ + 304, + 91, + 545, + 185 + ], + "type": "text", + "content": " them and use a linear layer " + }, + { + "bbox": [ + 304, + 91, + 545, + 185 + ], + "type": "inline_equation", + "content": "\\phi(\\cdot)" + }, + { + "bbox": [ + 304, + 91, + 545, + 185 + ], + "type": "text", + "content": " to project them into the input latent space of the LLM as " + }, + { + "bbox": [ + 304, + 91, + 545, + 185 + ], + "type": "inline_equation", + "content": "\\phi(\\mathcal{Q} \\oplus \\mathcal{V})" + }, + { + "bbox": [ + 304, + 91, + 545, + 185 + ], + "type": "text", + "content": ". The LLM used for caption generation in this work is the pre-trained Vicuna-13B [10], an open-source chatbot constructed from LLaMA [37]. During training and evaluation, we design a prompt in a conversational format, that is similar to [46]:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 321, + 194, + 525, + 230 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 194, + 525, + 230 + ], + "spans": [ + { + "bbox": [ + 321, + 194, + 525, + 230 + ], + "type": "text", + "content": "Human: Describe this image in detail. \nAssistant:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 238, + 545, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 238, + 545, + 357 + ], + "spans": [ + { + "bbox": [ + 304, + 238, + 545, + 357 + ], + "type": "text", + "content": "in which, ProjFeature denotes the projected feature " + }, + { + "bbox": [ + 304, + 238, + 545, + 357 + ], + "type": "inline_equation", + "content": "\\phi (\\mathcal{Q}\\oplus \\mathcal{V})" + }, + { + "bbox": [ + 304, + 238, + 545, + 357 + ], + "type": "text", + "content": " after the linear layer. In training phase, given input caption tokens " + }, + { + "bbox": [ + 304, + 238, + 545, + 357 + ], + "type": "inline_equation", + "content": "\\{c_i\\}_{i = 1}^L" + }, + { + "bbox": [ + 304, + 238, + 545, + 357 + ], + "type": "text", + "content": ", the LLM decoder concatenates the embedded prompt " + }, + { + "bbox": [ + 304, + 238, + 545, + 357 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{w}_i\\}_{i = 1}^N" + }, + { + "bbox": [ + 304, + 238, + 545, + 357 + ], + "type": "text", + "content": " and the embedded caption tokens " + }, + { + "bbox": [ + 304, + 238, + 545, + 357 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{c}_i\\}_{i = 1}^L" + }, + { + "bbox": [ + 304, + 238, + 545, + 357 + ], + "type": "text", + "content": " as input, and predicts the caption tokens in an autoregressive fashion, while in the evaluation phase, we only need to input the embedded prompt. We train EV-CAP by minimizing the cross-entropy loss in an end-to-end way: " + }, + { + "bbox": [ + 304, + 238, + 545, + 357 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\theta} = -\\sum_{i = 1}^{L}\\log p_{\\theta}(c_i|\\mathbf{w}_1,\\dots \\mathbf{w}_N,\\mathbf{c}_1,\\dots ,\\mathbf{c}_{i - 1})" + }, + { + "bbox": [ + 304, + 238, + 545, + 357 + ], + "type": "text", + "content": ", in which " + }, + { + "bbox": [ + 304, + 238, + 545, + 357 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 304, + 238, + 545, + 357 + ], + "type": "text", + "content": " indicates the trainable parameters." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 370, + 434, + 384 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 370, + 434, + 384 + ], + "spans": [ + { + "bbox": [ + 306, + 370, + 434, + 384 + ], + "type": "text", + "content": "4. Experimental Settings" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 391, + 397, + 403 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 391, + 397, + 403 + ], + "spans": [ + { + "bbox": [ + 306, + 391, + 397, + 403 + ], + "type": "text", + "content": "4.1. Training setup" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 410, + 545, + 505 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 410, + 545, + 505 + ], + "spans": [ + { + "bbox": [ + 304, + 410, + 545, + 505 + ], + "type": "text", + "content": "Implementation. EVCAP uses the same image encoder as in BLIP-2 [25], consisting of a ViT-g [14] and their pretrained Q-Former. Since we intend to obtain object name features through cross-attention between retrieved object names and visual features, we develop a customized Q-Former, which consists of BERT [23] with cross-attention layers inserted at every other transformer block. We use a frozen Vicuna-13B [10] as the caption generator." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 506, + 545, + 566 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 506, + 545, + 566 + ], + "spans": [ + { + "bbox": [ + 304, + 506, + 545, + 566 + ], + "type": "text", + "content": "Training dataset. For all experiments, we exclusively train EVCAP using the training set of COCO dataset [27], consisting of 82k images and 5 captions per images. The entire training process takes about 3 hours on 4 A6000 GPUs, using mixed precisions (more details in the supplementary)." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 306, + 575, + 407, + 588 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 575, + 407, + 588 + ], + "spans": [ + { + "bbox": [ + 306, + 575, + 407, + 588 + ], + "type": "text", + "content": "4.2. Evaluation setup" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "text", + "content": "Evaluation dataset. We evaluate EVCAP, trained using the COCO training set, across four datasets: its test set, two challenging benchmarks - NoCaps validation set and Flickr30k test set, and a synthetic commonsense-violating dataset - WHOOPS. We adhere to follow prior work [15, 41] to use the same images of Karpathy split [22] on COCO test set, NoCaps [2] validation set, and Karpathy split on Flickr30k [30] test set. In addition, WHOOPS [6] is a synthetic image captioning dataset comprising 500 synthetic commonsense-violating images and 2500 paired captions." + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13736" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 49, + 120, + 545, + 338 + ], + "blocks": [ + { + "bbox": [ + 46, + 71, + 545, + 115 + ], + "lines": [ + { + "bbox": [ + 46, + 71, + 545, + 115 + ], + "spans": [ + { + "bbox": [ + 46, + 71, + 545, + 115 + ], + "type": "text", + "content": "Table 1. Quantitative comparison against SOTA methods on three common image captioning benchmarks. * denotes using a memory bank. We report the size of training data and parameters; BLEU@4 (B@4), METEOR (M), CIDEr (C), and SPICE (S) scores on COCO test set; C and S scores on in-domain, near-domain, out-domain and overall data of NoCaps validation set; C and S scores on Flickr30k test set. Higher score is better. Bold indicates the best results among compared methods, normal indicates the second best results." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 49, + 120, + 545, + 338 + ], + "lines": [ + { + "bbox": [ + 49, + 120, + 545, + 338 + ], + "spans": [ + { + "bbox": [ + 49, + 120, + 545, + 338 + ], + "type": "table", + "html": "
MethodTrainingCOCONoCaps valFlickr30k Test
DataPara.TestIn-domainNear-domainOut-domainOverallCS
B@4MCSCSCSCSCSCS
Heavyweight-training models
VinVL [45]8.9M110M38.230.3129.323.696.813.590.713.187.411.690.912.8--
AoANet+MA* [16]COCO-38.028.7121.021.8----------
NOC-REK* [40]COCO110M----104.714.8100.214.1100.713.0100.914.0--
RCA-NOC* [13]COCO110M37.429.6128.423.192.212.987.812.687.511.588.312.4--
ViEcap GPT2 [15]COCO124M27.224.892.918.261.110.464.39.965.08.666.29.547.913.6
InstructBLIP Vicuna-13B [11]129M188M----------121.9-82.8-
OSCAR [26]4.1M338M37.430.7127.823.583.412.081.612.077.610.681.111.7--
BLIP [24]129M446M40.4-136.7-114.915.2112.114.9115.314.4113.214.8--
BLIP-2 FlanT5-XL [25]129M1.2B42.4-144.5-123.716.3120.215.9124.815.1121.615.8--
REVEAL* T5 [20]1.3B2.1B--145.4-------123.0---
Lightweight-training models
MiniGPT4 Vicuna-13B [46]5M3.94M38.029.6129.623.499.014.8106.915.3110.814.9108.815.178.416.9
SmallCap* GPT2 [35]COCO7M37.027.9119.721.3--------60.6-
ClipCap GPT2 [29]COCO43M33.527.5113.121.184.912.166.810.949.19.665.810.9--
EVCAP* Vicuna-13BCOCO3.97M41.531.2140.124.7111.715.3119.515.6116.514.7119.315.384.418.0
Specialist SOTAs
Qwen-VL Qwen-7B [5]1.4B9.6B----------121.4-85.8-
CogVLM Vicuna-7B [41]1.5B6.5B--148.7-----132.6-128.3-94.9-
PaLI mT5-XXL [9]1.6B17B--149.1-------127.0---
PaLI-X_UL2-32B [8]2.2B55B--149.2-------126.3---
", + "image_path": "82c1a4c9cda268d875841fd66754cb0bd89005508b338685a8f58c897073195a.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 357, + 289, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 357, + 289, + 502 + ], + "spans": [ + { + "bbox": [ + 46, + 357, + 289, + 502 + ], + "type": "text", + "content": "Compared methods. We compare EVCAP with several SOTAs. According to the trainable parameters size, they can be divided into 1) Heavyweight-training (between 100M to 5B): VinVL [45], AoANet [16], NOC-REK [40], RCA-NOC [13], ViECap [15], InstructBLIP [11], OSCAR [26], BLIP [24], BLIP-2 [25], REVEAL [20]; 2) Lightweight-training (less than 100M): MiniGPT4 [46], SmallCap [35], ClipCap [29]; and also 3) Specialist SOTAs with huge trainable parameters (larger than 5B): QwenVL [5], CogVLM [41], PaLI [9], PaLI-X [8]. Among these methods, AoANet, NOC-REK, RCA-NOC, REVEAL, and SmallCap are retrieval-augmented captioning methods." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 516, + 173, + 529 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 516, + 173, + 529 + ], + "spans": [ + { + "bbox": [ + 47, + 516, + 173, + 529 + ], + "type": "text", + "content": "5. Experimental Results" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 537, + 251, + 549 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 537, + 251, + 549 + ], + "spans": [ + { + "bbox": [ + 47, + 537, + 251, + 549 + ], + "type": "text", + "content": "5.1. Results on in-/out-domain benchmarks" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 557, + 287, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 557, + 287, + 616 + ], + "spans": [ + { + "bbox": [ + 46, + 557, + 287, + 616 + ], + "type": "text", + "content": "We assess EVCAP against SOTAs on both in-domain and out-domain benchmarks. The COCO test set can be considered as in-domain data as we only train our model on the COCO training set. Out-domain benchmarks are the NoCaps validation set and the Flickr30k test set." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 618, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 618, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 618, + 288, + 714 + ], + "type": "text", + "content": "Quantitative results. Tab. 1 details our EVCAP's performance in comparison with SOTA methods. We first evaluate training costs in terms of training data sizes and parameters. Similar to various heavyweight-training models that exclude LLMs and the majority of lightweight-training models, EVCAP is trained solely on the COCO training set. It utilizes only 3.97M trainable parameters, positioning it as the second smallest, slightly larger than MiniGPT4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 357, + 545, + 643 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 357, + 545, + 643 + ], + "spans": [ + { + "bbox": [ + 304, + 357, + 545, + 643 + ], + "type": "text", + "content": "with 3.94M. Among lightweight-training models, our approach outperforms others, achieving the highest scores on all benchmarks. Despite using less training data and nearly identical trainable parameters as MiniGPT4, EVCAP significantly surpasses it, with a marked improvement of 10.5, 10.5, and 6.0 in CIDEr scores for each benchmark. When further compared with heavyweight-training models, the performance of EVCAP stands out among million-level models, nearly matching InstructBLIP, except in NoCaps. Note that since BLIP-2 does not include Vicuna checkpoints, InstructBLIP performs pre-training with Vicuna using the same procedure as BLIP-2, whereas EVCAP does not involve pre-training. Against REVEAL, which also uses external memory, our EVCAP utilizes about 1/3000 training data and 1/500 training parameters yet yields comparable results. Moreover, EVCAP's performance is on par with BLIP-2, the top-performing model with 1.2B trainable parameters. This highlights EVCAP's efficiency and effectiveness despite its significantly smaller training cost, thanks to our external visual-name memory. Regarding specialist SOTAs, they use billion-level training data and over 5B trainable parameters, so it is acceptable that they can achieve exceptionally strong performance, surpassing EVCAP by nearly 10 on all benchmarks in CIDEr scores." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 654, + 545, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 545, + 715 + ], + "type": "text", + "content": "Qualitative results. Fig. 3 presents a comparison of captions generated by our EVCAP and three SOTA models across three benchmarks. The captions of SmallCap are generated by its publicly accessible demo [1]. We generate captions of MiniGPT4 and BLIP-2 using their respec" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13737" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 126, + 73, + 172, + 82 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 126, + 73, + 172, + 82 + ], + "spans": [ + { + "bbox": [ + 126, + 73, + 172, + 82 + ], + "type": "text", + "content": "COCO Test" + } + ] + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 85, + 84, + 212, + 157 + ], + "blocks": [ + { + "bbox": [ + 85, + 84, + 212, + 157 + ], + "lines": [ + { + "bbox": [ + 85, + 84, + 212, + 157 + ], + "spans": [ + { + "bbox": [ + 85, + 84, + 212, + 157 + ], + "type": "image", + "image_path": "f0c7cd5fc601182ad9ef28928e582b8e31464589b2a50de68eebf324384de677.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 83, + 159, + 214, + 176 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 83, + 159, + 214, + 176 + ], + "spans": [ + { + "bbox": [ + 83, + 159, + 214, + 176 + ], + "type": "text", + "content": "GT: A green bus driving through a rural area with trees in the background." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 83, + 178, + 214, + 195 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 83, + 178, + 214, + 195 + ], + "spans": [ + { + "bbox": [ + 83, + 178, + 214, + 195 + ], + "type": "text", + "content": "SmallCap: A bus driving down a street next to trees." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 83, + 197, + 195, + 215 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 83, + 197, + 195, + 215 + ], + "spans": [ + { + "bbox": [ + 83, + 197, + 195, + 215 + ], + "type": "text", + "content": "MiniGPT4: A green bus is driving down the street." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 83, + 217, + 205, + 235 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 83, + 217, + 205, + 235 + ], + "spans": [ + { + "bbox": [ + 83, + 217, + 205, + 235 + ], + "type": "text", + "content": "BLIP-2: A green bus driving down a road with trees in the background." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 83, + 236, + 203, + 254 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 83, + 236, + 203, + 254 + ], + "spans": [ + { + "bbox": [ + 83, + 236, + 203, + 254 + ], + "type": "text", + "content": "EVCap: A green bus driving down a road next to trees." + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 86, + 257, + 212, + 331 + ], + "blocks": [ + { + "bbox": [ + 86, + 257, + 212, + 331 + ], + "lines": [ + { + "bbox": [ + 86, + 257, + 212, + 331 + ], + "spans": [ + { + "bbox": [ + 86, + 257, + 212, + 331 + ], + "type": "image", + "image_path": "31f38804d88ed9718adb39319090b734d39a5158c437da97db65dc1d7b8286a4.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 83, + 332, + 194, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 83, + 332, + 194, + 350 + ], + "spans": [ + { + "bbox": [ + 83, + 332, + 194, + 350 + ], + "type": "text", + "content": "GT: A woman in a blue top with headphones and two cellphones." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 83, + 351, + 209, + 369 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 83, + 351, + 209, + 369 + ], + "spans": [ + { + "bbox": [ + 83, + 351, + 209, + 369 + ], + "type": "text", + "content": "SmallCap: A woman sitting in front of a laptop computer." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 83, + 371, + 193, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 83, + 371, + 193, + 389 + ], + "spans": [ + { + "bbox": [ + 83, + 371, + 193, + 389 + ], + "type": "text", + "content": "MiniGPT4: A woman sitting on a couch holding two phones." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 83, + 391, + 204, + 409 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 83, + 391, + 204, + 409 + ], + "spans": [ + { + "bbox": [ + 83, + 391, + 204, + 409 + ], + "type": "text", + "content": "BLIP-2: A woman sitting on a couch with two cell phones." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 83, + 410, + 207, + 428 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 83, + 410, + 207, + 428 + ], + "spans": [ + { + "bbox": [ + 83, + 410, + 207, + 428 + ], + "type": "text", + "content": "EVCap: A woman wearing headphones holding two cell phones." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 273, + 73, + 321, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 273, + 73, + 321, + 83 + ], + "spans": [ + { + "bbox": [ + 273, + 73, + 321, + 83 + ], + "type": "text", + "content": "NoCaps Val" + } + ] + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 234, + 84, + 361, + 157 + ], + "blocks": [ + { + "bbox": [ + 234, + 84, + 361, + 157 + ], + "lines": [ + { + "bbox": [ + 234, + 84, + 361, + 157 + ], + "spans": [ + { + "bbox": [ + 234, + 84, + 361, + 157 + ], + "type": "image", + "image_path": "db8065873222d29d9fccd12b292d519b520e4922206aa27ef18f77eb871b7e91.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "bbox": [ + 231, + 159, + 351, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 231, + 159, + 351, + 175 + ], + "spans": [ + { + "bbox": [ + 231, + 159, + 351, + 175 + ], + "type": "text", + "content": "GT: The two guinea pigs are getting dried off in a yellow towel." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 231, + 177, + 353, + 195 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 231, + 177, + 353, + 195 + ], + "spans": [ + { + "bbox": [ + 231, + 177, + 353, + 195 + ], + "type": "text", + "content": "SmallCap: A person holding a small animal in a towel." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 231, + 197, + 346, + 215 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 231, + 197, + 346, + 215 + ], + "spans": [ + { + "bbox": [ + 231, + 197, + 346, + 215 + ], + "type": "text", + "content": "MiniGPT4: Two small animals are wrapped in a towel." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 231, + 217, + 360, + 235 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 231, + 217, + 360, + 235 + ], + "spans": [ + { + "bbox": [ + 231, + 217, + 360, + 235 + ], + "type": "text", + "content": "BLIP-2: Two guinea pigs wrapped in a yellow towel." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 231, + 236, + 359, + 255 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 231, + 236, + 359, + 255 + ], + "spans": [ + { + "bbox": [ + 231, + 236, + 359, + 255 + ], + "type": "text", + "content": "EVCap: Two guinea pigs are wrapped in a yellow towel." + } + ] + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 234, + 257, + 361, + 331 + ], + "blocks": [ + { + "bbox": [ + 234, + 257, + 361, + 331 + ], + "lines": [ + { + "bbox": [ + 234, + 257, + 361, + 331 + ], + "spans": [ + { + "bbox": [ + 234, + 257, + 361, + 331 + ], + "type": "image", + "image_path": "371ec2cfc3a252917fb63fa2cc8dfb97a18c771ee7895a3edfd0c8bd8b5755e9.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + } + ], + "index": 20 + }, + { + "bbox": [ + 231, + 332, + 353, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 231, + 332, + 353, + 350 + ], + "spans": [ + { + "bbox": [ + 231, + 332, + 353, + 350 + ], + "type": "text", + "content": "GT: A computer screen showing two men sitting at a table." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 231, + 351, + 357, + 369 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 231, + 351, + 357, + 369 + ], + "spans": [ + { + "bbox": [ + 231, + 351, + 357, + 369 + ], + "type": "text", + "content": "SmallCap: Two men sitting at a table with a laptop." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 231, + 371, + 359, + 388 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 231, + 371, + 359, + 388 + ], + "spans": [ + { + "bbox": [ + 231, + 371, + 359, + 388 + ], + "type": "text", + "content": "MiniGPT4: A laptop computer sitting on top of a table." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 231, + 390, + 343, + 408 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 231, + 390, + 343, + 408 + ], + "spans": [ + { + "bbox": [ + 231, + 390, + 343, + 408 + ], + "type": "text", + "content": "BLIP-2: A laptop computer with a picture of two men on it." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 231, + 410, + 343, + 427 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 231, + 410, + 343, + 427 + ], + "spans": [ + { + "bbox": [ + 231, + 410, + 343, + 427 + ], + "type": "text", + "content": "EVCap: A laptop computer with a picture of two men on the screen." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 416, + 73, + 473, + 82 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 416, + 73, + 473, + 82 + ], + "spans": [ + { + "bbox": [ + 416, + 73, + 473, + 82 + ], + "type": "text", + "content": "Flickr30k Test" + } + ] + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 381, + 84, + 508, + 157 + ], + "blocks": [ + { + "bbox": [ + 381, + 84, + 508, + 157 + ], + "lines": [ + { + "bbox": [ + 381, + 84, + 508, + 157 + ], + "spans": [ + { + "bbox": [ + 381, + 84, + 508, + 157 + ], + "type": "image", + "image_path": "26e81763047a240653926a483fd60d9f7d4d2836fe77c73f79e1d6146590b7b0.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_body" + } + ], + "index": 27 + }, + { + "bbox": [ + 378, + 159, + 492, + 176 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 378, + 159, + 492, + 176 + ], + "spans": [ + { + "bbox": [ + 378, + 159, + 492, + 176 + ], + "type": "text", + "content": "GT: A very young child in a denim baseball cap eats a green apple." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 378, + 177, + 496, + 195 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 378, + 177, + 496, + 195 + ], + "spans": [ + { + "bbox": [ + 378, + 177, + 496, + 195 + ], + "type": "text", + "content": "SmallCap: A young boy holding an apple in his hand." + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 378, + 197, + 494, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 378, + 197, + 494, + 216 + ], + "spans": [ + { + "bbox": [ + 378, + 197, + 494, + 216 + ], + "type": "text", + "content": "MiniGPT4: A baby sitting in a high chair eating an apple." + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 378, + 217, + 505, + 235 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 378, + 217, + 505, + 235 + ], + "spans": [ + { + "bbox": [ + 378, + 217, + 505, + 235 + ], + "type": "text", + "content": "BLIP-2: A baby sitting in a white chair eating a green apple." + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 378, + 236, + 509, + 255 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 378, + 236, + 509, + 255 + ], + "spans": [ + { + "bbox": [ + 378, + 236, + 509, + 255 + ], + "type": "text", + "content": "EVCap: A toddler eating a green apple while wearing a hat." + } + ] + } + ], + "index": 32 + }, + { + "type": "image", + "bbox": [ + 381, + 258, + 508, + 331 + ], + "blocks": [ + { + "bbox": [ + 381, + 258, + 508, + 331 + ], + "lines": [ + { + "bbox": [ + 381, + 258, + 508, + 331 + ], + "spans": [ + { + "bbox": [ + 381, + 258, + 508, + 331 + ], + "type": "image", + "image_path": "dd0c497f9a55af1019c1e154598255be1fcefe9d34c0a9b98cf0ca951e4b2dbb.jpg" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_body" + } + ], + "index": 33 + }, + { + "bbox": [ + 378, + 332, + 503, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 378, + 332, + 503, + 350 + ], + "spans": [ + { + "bbox": [ + 378, + 332, + 503, + 350 + ], + "type": "text", + "content": "GT: Two men are riding on a wooden vehicle pulled by two donkeys." + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 378, + 351, + 509, + 369 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 378, + 351, + 509, + 369 + ], + "spans": [ + { + "bbox": [ + 378, + 351, + 509, + 369 + ], + "type": "text", + "content": "SmallCap: A donkey pulling a cart with a man in the background." + } + ] + } + ], + "index": 35 + }, + { + "bbox": [ + 378, + 370, + 487, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 378, + 370, + 487, + 389 + ], + "spans": [ + { + "bbox": [ + 378, + 370, + 487, + 389 + ], + "type": "text", + "content": "MiniGPT4: Two men riding on a donkey in the dirt." + } + ] + } + ], + "index": 36 + }, + { + "bbox": [ + 378, + 391, + 509, + 408 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 378, + 391, + 509, + 408 + ], + "spans": [ + { + "bbox": [ + 378, + 391, + 509, + 408 + ], + "type": "text", + "content": "BLIP-2: Two men riding a horse drawn cart through a field." + } + ] + } + ], + "index": 37 + }, + { + "bbox": [ + 378, + 410, + 488, + 428 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 378, + 410, + 488, + 428 + ], + "spans": [ + { + "bbox": [ + 378, + 410, + 488, + 428 + ], + "type": "text", + "content": "EVCap: Two men riding in a cart pulled by two donkeys." + } + ] + } + ], + "index": 38 + }, + { + "bbox": [ + 46, + 495, + 287, + 625 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 495, + 287, + 625 + ], + "spans": [ + { + "bbox": [ + 46, + 495, + 287, + 625 + ], + "type": "text", + "content": "tive pre-trained models. As a lightweight and retrieval-augmented captioning method, SmallCap struggles to produce accurate captions for given images, primarily because it relies on retrieved captions laden with extraneous information. MiniGPT4, though aligned with the primary content of images, sometimes misses certain objects like \"trees\" and \"headphones\". This oversight stems from its focus on the main objects in images, without integrating additional cues for other objects provided by the retrieved object names. In contrast, the captions generated by our EVCAP are comparable to those of BLIP-2." + } + ] + } + ], + "index": 40 + }, + { + "bbox": [ + 47, + 634, + 253, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 634, + 253, + 647 + ], + "spans": [ + { + "bbox": [ + 47, + 634, + 253, + 647 + ], + "type": "text", + "content": "5.2. Results on commonsense-violating data" + } + ] + } + ], + "index": 41 + }, + { + "bbox": [ + 46, + 653, + 287, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 287, + 689 + ], + "type": "text", + "content": "To explore our EVCAP's capability in describing contents in open-word settings, we further evaluate it on WHOOPS dataset, which contains commonsense-violating images." + } + ] + } + ], + "index": 42 + }, + { + "bbox": [ + 46, + 689, + 287, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 689, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 46, + 689, + 287, + 712 + ], + "type": "text", + "content": "Quantitative results. In Tab. 2, we compare the performance of EVCAP, MiniGPT4, BLIP, and BLIP-2 on" + } + ] + } + ], + "index": 43 + }, + { + "type": "image", + "bbox": [ + 308, + 496, + 360, + 548 + ], + "blocks": [ + { + "bbox": [ + 46, + 439, + 547, + 473 + ], + "lines": [ + { + "bbox": [ + 46, + 439, + 547, + 473 + ], + "spans": [ + { + "bbox": [ + 46, + 439, + 547, + 473 + ], + "type": "text", + "content": "Figure 3. Examples of captions generated by our EVCAP and three SOTA methods on COCO test set, NoCaps validation set, and Flickr30k test set. GT refers to the Ground Truth captions. Incorrect objects in captions are highlighted in red, while correct ones are in blue. Our EVCAP correctly generates captions across different datasets, showing performance comparable to BLIP-2." + } + ] + } + ], + "index": 39, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 308, + 496, + 360, + 548 + ], + "lines": [ + { + "bbox": [ + 308, + 496, + 360, + 548 + ], + "spans": [ + { + "bbox": [ + 308, + 496, + 360, + 548 + ], + "type": "image", + "image_path": "dfb894901fe60fd7e49981692efbf55e5abf7afaf7457c96a642077659221d9c.jpg" + } + ] + } + ], + "index": 44, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 306, + 560, + 545, + 594 + ], + "lines": [ + { + "bbox": [ + 306, + 560, + 545, + 594 + ], + "spans": [ + { + "bbox": [ + 306, + 560, + 545, + 594 + ], + "type": "text", + "content": "Figure 4. Examples of captions generated by our EVCAP, EVCAP (w/ WHOOPS), and three SOTAs on WHOOPS dataset. Incorrect objects are highlighted in red, while correct ones are in blue." + } + ] + } + ], + "index": 51, + "angle": 0, + "type": "image_caption" + } + ], + "index": 44 + }, + { + "bbox": [ + 361, + 494, + 522, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 361, + 494, + 522, + 502 + ], + "spans": [ + { + "bbox": [ + 361, + 494, + 522, + 502 + ], + "type": "text", + "content": "GT: This is an image of a blue Pikachu with yellow accents." + } + ] + } + ], + "index": 45 + }, + { + "bbox": [ + 361, + 502, + 509, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 361, + 502, + 509, + 510 + ], + "spans": [ + { + "bbox": [ + 361, + 502, + 509, + 510 + ], + "type": "text", + "content": "SmallCap: A blue and white stuffed animal on a table." + } + ] + } + ], + "index": 46 + }, + { + "bbox": [ + 361, + 510, + 543, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 361, + 510, + 543, + 518 + ], + "spans": [ + { + "bbox": [ + 361, + 510, + 543, + 518 + ], + "type": "text", + "content": "MiniGPT4: A blue poker sitting on the floor with its eyes closed." + } + ] + } + ], + "index": 47 + }, + { + "bbox": [ + 361, + 519, + 485, + 526 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 361, + 519, + 485, + 526 + ], + "spans": [ + { + "bbox": [ + 361, + 519, + 485, + 526 + ], + "type": "text", + "content": "BLIP-2: A blue Pikachu sitting in a dark room." + } + ] + } + ], + "index": 48 + }, + { + "bbox": [ + 361, + 527, + 528, + 541 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 361, + 527, + 528, + 541 + ], + "spans": [ + { + "bbox": [ + 361, + 527, + 528, + 541 + ], + "type": "text", + "content": "EVCap: A blue and yellow cartoon character sitting on a dark background." + } + ] + } + ], + "index": 49 + }, + { + "bbox": [ + 361, + 542, + 525, + 550 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 361, + 542, + 525, + 550 + ], + "spans": [ + { + "bbox": [ + 361, + 542, + 525, + 550 + ], + "type": "text", + "content": "EVCap (w/ WHOOPS): A blue Pikachu is sitting on the floor." + } + ] + } + ], + "index": 50 + }, + { + "bbox": [ + 304, + 618, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 618, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 618, + 545, + 713 + ], + "type": "text", + "content": "WHOOPS dataset. This dataset is particularly challenging due to its inclusion of unusual objects [6]. Initially, as an end-to-end trained model, our EVCAP exhibits performance similar to MiniGPT4. However, there is a noticeable improvement in the CIDEr score, after the external memory is enriched with 2396 new objects from the WHOOPS dataset, each represented by 5 synthesized images generated using stable diffusion [36]. It highlights the effectiveness of" + } + ] + } + ], + "index": 52 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13738" + } + ] + } + ], + "index": 53 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 49, + 118, + 286, + 259 + ], + "blocks": [ + { + "bbox": [ + 47, + 71, + 287, + 115 + ], + "lines": [ + { + "bbox": [ + 47, + 71, + 287, + 115 + ], + "spans": [ + { + "bbox": [ + 47, + 71, + 287, + 115 + ], + "type": "text", + "content": "Table 2. Quantitative results on commonsense-violating data - WHOOPS dataset. EVCAP (w/ WHOOPS) denotes EVCAP using the memory expanded by WHOOPS objects. The results reveal the open-world comprehension ability and expandability of EVCAP." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 49, + 118, + 286, + 259 + ], + "lines": [ + { + "bbox": [ + 49, + 118, + 286, + 259 + ], + "spans": [ + { + "bbox": [ + 49, + 118, + 286, + 259 + ], + "type": "table", + "html": "
MethodB@4MCS
Only pre-trained models
BLIP [24] (from [6])13-65-
BLIP-2 FlanT5-XXL [25] (from [6])31-120-
BLIP-2 FlanT5-XXL [25] (reproduced)2826.793.117.9
Finetuned models on COCO
MiniGPT4 [46]24.226.784.818.2
BLIP [24]22.925.079.317.1
BLIP-2 FlanT5-XL [25]25.827.089.118.3
End-to-end trained models on COCO
EVCAP24.126.185.317.7
EVCAP (w/ WHOOPS)24.426.186.317.8
", + "image_path": "b7da35384deaeb912f913760fdb15001de9389f1dca09d1f270a8551e7e1f6c5.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 49, + 306, + 286, + 358 + ], + "blocks": [ + { + "bbox": [ + 47, + 270, + 287, + 304 + ], + "lines": [ + { + "bbox": [ + 47, + 270, + 287, + 304 + ], + "spans": [ + { + "bbox": [ + 47, + 270, + 287, + 304 + ], + "type": "text", + "content": "Table 3. Ablation study on components prior to the LLM decoder in EVCAP. The result of \"+\" Attentive fusion\" demonstrates the substantial impact of the external visual-name memory." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 49, + 306, + 286, + 358 + ], + "lines": [ + { + "bbox": [ + 49, + 306, + 286, + 358 + ], + "spans": [ + { + "bbox": [ + 49, + 306, + 286, + 358 + ], + "type": "table", + "html": "
MethodCOCO testNoCaps valFlickr30k test
CSCSCS
ViT + Q-Former (Baseline)134.423.9108.814.276.817.3
+ Image query tokens (Baseline+)134.123.8109.014.377.317.2
+ Attentive fusion (EVCAP)140.124.7119.315.384.418.0
", + "image_path": "3c410699e381dc1838b3aff36ed69e3b2626a5ba083231bcb25a1eb04d23dfa2.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 49, + 371, + 101, + 419 + ], + "blocks": [ + { + "bbox": [ + 49, + 371, + 101, + 419 + ], + "lines": [ + { + "bbox": [ + 49, + 371, + 101, + 419 + ], + "spans": [ + { + "bbox": [ + 49, + 371, + 101, + 419 + ], + "type": "image", + "image_path": "1a7a4498382342c04d35bddc17714b4bcfea88d424b95000a66b935bcedfc598.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 429, + 287, + 475 + ], + "lines": [ + { + "bbox": [ + 46, + 429, + 287, + 475 + ], + "spans": [ + { + "bbox": [ + 46, + 429, + 287, + 475 + ], + "type": "text", + "content": "Figure 5. Visualization of the captions generated from ablation study on the NoCaps validation set. We also show the retrieved object names by EVCAP, presented in gray. Incorrect objects in captions are highlighted in red, while correct ones are in blue." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 102, + 371, + 236, + 379 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 102, + 371, + 236, + 379 + ], + "spans": [ + { + "bbox": [ + 102, + 371, + 236, + 379 + ], + "type": "text", + "content": "GT: A hamster on a blanket with a hand behind it." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 102, + 380, + 286, + 388 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 102, + 380, + 286, + 388 + ], + "spans": [ + { + "bbox": [ + 102, + 380, + 286, + 388 + ], + "type": "text", + "content": "Baseline: A hamster sitting on a couch with a person's hand near it." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 103, + 388, + 241, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 103, + 388, + 241, + 396 + ], + "spans": [ + { + "bbox": [ + 103, + 388, + 241, + 396 + ], + "type": "text", + "content": "Baseline+: A close up of a small hamster on a bed." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 103, + 396, + 264, + 404 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 103, + 396, + 264, + 404 + ], + "spans": [ + { + "bbox": [ + 103, + 396, + 264, + 404 + ], + "type": "text", + "content": "EVCap: A small hamster sitting on a blanket next to a hand." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 103, + 405, + 286, + 419 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 103, + 405, + 286, + 419 + ], + "spans": [ + { + "bbox": [ + 103, + 405, + 286, + 419 + ], + "type": "text", + "content": "(pillow, hamster, grizzly, gameboard, kitten, spectacles, cat, blanket, giraffe, pug-dog)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 496, + 286, + 520 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 496, + 286, + 520 + ], + "spans": [ + { + "bbox": [ + 46, + 496, + 286, + 520 + ], + "type": "text", + "content": "our idea of incorporating an expandable external memory into the captioning model for open-world comprehension." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 520, + 287, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 520, + 287, + 628 + ], + "spans": [ + { + "bbox": [ + 46, + 520, + 287, + 628 + ], + "type": "text", + "content": "Qualitative results. Fig. 4 illustrates the captions generated by EVCAP, EVCAP (w/WHOOPS), and three SOTAs for one image from the WHOOPS dataset. Similar to other methods except for BLIP-2, EVCAP can not recognize \"blue cartoon character\" as \"Pikachu\", while EVCAP (w/WHOOPS) successfully predicts it because of the updated memory. SmallCap and MiniGPT4 tend to generate captions with hallucinatory objects, a result of commonsense-violating contents present in the images." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 47, + 635, + 149, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 635, + 149, + 647 + ], + "spans": [ + { + "bbox": [ + 47, + 635, + 149, + 647 + ], + "type": "text", + "content": "5.3. Detailed analysis" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 46, + 654, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 654, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 654, + 286, + 713 + ], + "type": "text", + "content": "Ablation study. We assess the contribution of each component prior to the LLM decoder in EVCAP by incrementally integrating the image query tokens and the attentive fusion module into our baseline model. The baseline model comprises a ViT+Q-Former, a linear layer, and a LLM decoder." + } + ] + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 309, + 70, + 542, + 207 + ], + "blocks": [ + { + "bbox": [ + 309, + 70, + 542, + 207 + ], + "lines": [ + { + "bbox": [ + 309, + 70, + 542, + 207 + ], + "spans": [ + { + "bbox": [ + 309, + 70, + 542, + 207 + ], + "type": "image", + "image_path": "954475c57db53357a2a3c352f997d05f6391c647ae4e380cc63d1da9b6448115.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 214, + 545, + 270 + ], + "lines": [ + { + "bbox": [ + 304, + 214, + 545, + 270 + ], + "spans": [ + { + "bbox": [ + 304, + 214, + 545, + 270 + ], + "type": "text", + "content": "Figure 6. Visualization of the visual features in external memory using t-SNE. For visual features in LVIS dataset's objects (blue), the related objects fall in the same cluster. After adding more visual features of synthesized images from WHOOPS' objects, new objects (red) are located at appropriate clusters (zoom-in view)." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 293, + 545, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 293, + 545, + 449 + ], + "spans": [ + { + "bbox": [ + 304, + 293, + 545, + 449 + ], + "type": "text", + "content": "The quantitative results are shown in Tab. 3. When employing only the baseline model (Baseline), CIDEr scores drop notably by 5.7, 10.5, and 7.6 on COCO, NoCaps, and Flickr30k, respectively. The inclusion of trainable image query tokens (Baseline+) brings a marginal improvement on NoCaps and Flickr30k. However, the performance is significantly enhanced with the addition of attentive fusion (along with the introduction of external memory), indicating the pivotal role of the external visual-name memory in the overall effectiveness of EVCAP. This is further corroborated by the qualitative results in Fig. 5, where captions from Baseline and Baseline+ inaccurately include objects like \"couch\" and \"bed\", and Baseline+ overlooks \"hand\"." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 450, + 546, + 675 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 450, + 546, + 675 + ], + "spans": [ + { + "bbox": [ + 304, + 450, + 546, + 675 + ], + "type": "text", + "content": "Exploration for external memory expandability. To demonstrate the scalability of the external memory in EVCAP, we visualize the visual features stored in LVIS external memory, and newly synthesized data from objects appearing in the WHOOPS dataset. We employ t-SNE [38] to plot visual features after reducing their dimensions to 2-D (Fig. 6). For clear visualization, we only randomly display 3649 visual features in LVIS memory, and add 479 visual features from WHOOPS objects. Among them, 35 samples are randomly labeled. The result shows a clear clustering of LVIS objects (blue) in the external memory, as well as the successful integration and appropriate localization of new objects from WHOOPS (red) into these clusters. This pattern not only confirms the distinctiveness of visual features already present in the memory but also demonstrates the potential to accurately incorporate and differentiate new objects introduced from updated data. These findings highlight our external memory's ability to expand and maintain its effectiveness even as new data is incorporated." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 677, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 547, + 713 + ], + "type": "text", + "content": "Impact of external memory size. We examine the impact of external memory size in Tab. 4. On the one hand, we randomly remove " + }, + { + "bbox": [ + 304, + 677, + 547, + 713 + ], + "type": "inline_equation", + "content": "30\\%" + }, + { + "bbox": [ + 304, + 677, + 547, + 713 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 677, + 547, + 713 + ], + "type": "inline_equation", + "content": "60\\%" + }, + { + "bbox": [ + 304, + 677, + 547, + 713 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 677, + 547, + 713 + ], + "type": "inline_equation", + "content": "90\\%" + }, + { + "bbox": [ + 304, + 677, + 547, + 713 + ], + "type": "text", + "content": " data in the external" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13739" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 50, + 106, + 286, + 187 + ], + "blocks": [ + { + "bbox": [ + 47, + 71, + 288, + 104 + ], + "lines": [ + { + "bbox": [ + 47, + 71, + 288, + 104 + ], + "spans": [ + { + "bbox": [ + 47, + 71, + 288, + 104 + ], + "type": "text", + "content": "Table 4. Impact of the external memory size on the performance of EVCAP by evaluation under CIDEr scores. Changes in the size of external memory result in changes in performance." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 50, + 106, + 286, + 187 + ], + "lines": [ + { + "bbox": [ + 50, + 106, + 286, + 187 + ], + "spans": [ + { + "bbox": [ + 50, + 106, + 286, + 187 + ], + "type": "table", + "html": "
MethodNoCaps valFlickr30k
InNearOutOverallTest
LVIS objects (EVCAP)111.7119.5116.5119.384.4
- 30% LVIS112.0119.2115.3118.885.0
- 60% LVIS111.4119.1116.2119.085.1
- 90% LVIS110.6118.2115.8118.383.6
+ WHOOPS110.7118.9116.7119.084.9
", + "image_path": "1d2c22461d754c669a628371b471ef48079dff4cbd4a7a07853f690db5f47028.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 50, + 198, + 285, + 291 + ], + "blocks": [ + { + "bbox": [ + 50, + 198, + 285, + 291 + ], + "lines": [ + { + "bbox": [ + 50, + 198, + 285, + 291 + ], + "spans": [ + { + "bbox": [ + 50, + 198, + 285, + 291 + ], + "type": "image", + "image_path": "c00825c28763cc8b55ca76a5d0aa5ee5fe752db79122330b7159a2860ab28ec3.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 299, + 288, + 333 + ], + "lines": [ + { + "bbox": [ + 46, + 299, + 288, + 333 + ], + "spans": [ + { + "bbox": [ + 46, + 299, + 288, + 333 + ], + "type": "text", + "content": "Figure 7. CIDEr scores after training EVCAP with the number of retrieved object names K from 0 to 20. The results indicate that the performance is relatively optimal when K is set to be 10." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 355, + 287, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 355, + 287, + 498 + ], + "spans": [ + { + "bbox": [ + 46, + 355, + 287, + 498 + ], + "type": "text", + "content": "memory constructed from LVIS objects. The results show the performance gradually degrades on NoCaps as reducing " + }, + { + "bbox": [ + 46, + 355, + 287, + 498 + ], + "type": "inline_equation", + "content": "30\\%" + }, + { + "bbox": [ + 46, + 355, + 287, + 498 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 355, + 287, + 498 + ], + "type": "inline_equation", + "content": "90\\%" + }, + { + "bbox": [ + 46, + 355, + 287, + 498 + ], + "type": "text", + "content": " LVIS. Despite some unexpected increases in certain results on NoCaps (5th row) and Flickr30k (4th - 5th rows), they do not alter the overall downward trend. Similar phenomena are also noted in SmallCap [35], we speculate it is due to data distribution. On the other hand, as we infuse WHOOPS knowledge into LVIS memory, there is a slight improvement on NoCaps (out) and Flickr30k. These observations validate the model's capability to effectively retrieve object names from an updated memory, enhancing its performance in generating captions." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 498, + 287, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 498, + 287, + 653 + ], + "spans": [ + { + "bbox": [ + 46, + 498, + 287, + 653 + ], + "type": "text", + "content": "Impact of the number of retrieved object names. We investigate how the number of retrieved object names K (Sec. 3.3) affect EVCAP in Fig. 7. We train the model with K from 0 to 20 and evaluate the performance under CIDER on all three benchmarks. From the results, we can find that the model works worst on the out-domain dataset (NoCaps) with zero object names. It confirms that the performance boost from Baseline+ to EVCAP (Tab. 3) is primarily attributed to the retrieval-augmented mechanism, but not the customized Q-Former itself. With more object names, performance fluctuates but improves. Furthermore, we observe that setting K to 10 yields relatively optimal overall performance, validating the choice of " + }, + { + "bbox": [ + 46, + 498, + 287, + 653 + ], + "type": "inline_equation", + "content": "\\mathrm{K} = 10" + }, + { + "bbox": [ + 46, + 498, + 287, + 653 + ], + "type": "text", + "content": " in EVCAP." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 654, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 654, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 654, + 288, + 715 + ], + "type": "text", + "content": "Analysis with different decoders. To explore the influence of different LLMs decoders on our EVCAP, we experiment by substituting Vicuna-13B with GPT2 and Vicuna-7B, as detailed in Tab. 5. With GPT2 as the decoder, EVCAP still markedly surpasses other GPT2-based models, achieving" + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 308, + 106, + 545, + 218 + ], + "blocks": [ + { + "bbox": [ + 305, + 71, + 545, + 104 + ], + "lines": [ + { + "bbox": [ + 305, + 71, + 545, + 104 + ], + "spans": [ + { + "bbox": [ + 305, + 71, + 545, + 104 + ], + "type": "text", + "content": "Table 5. Analysis with different LLM decoders including GPT2, Vicuna-7B, and Vicuna-13B. The results reveal EVCAP is effective when applying it in different LLM decoders." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 308, + 106, + 545, + 218 + ], + "lines": [ + { + "bbox": [ + 308, + 106, + 545, + 218 + ], + "spans": [ + { + "bbox": [ + 308, + 106, + 545, + 218 + ], + "type": "table", + "html": "
MethodLLMCOCO testNoCaps valFlickr30k test
CSCSCS
SmallCap [35]GPT2119.721.3--60.6-
ViEcap [15]GPT292.918.266.29.547.913.6
EVCAPGPT2131.023.297.613.370.616.1
MiniGPT4 [46]Vicuna-7B119.423.5108.715.773.917.2
InstructBLIP [11]Vicuna-7B--123.1-82.4-
EVCAPVicuna-7B139.024.7116.815.382.718.0
MiniGPT4 [46]Vicuna-13B129.623.4108.815.178.416.9
InstructBLIP [11]Vicuna-13B--121.9-82.8-
EVCAPVicuna-13B140.124.7119.315.384.418.0
", + "image_path": "17e5f056a4540629f5df2aa9da5fea6697c41dc0b59d6e904d7d75e285d69700.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 240, + 545, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 240, + 545, + 361 + ], + "spans": [ + { + "bbox": [ + 304, + 240, + 545, + 361 + ], + "type": "text", + "content": "impressive gains of 11.3 and 10.0 under CIDEr on COCO and Flickr30k, compared to SmallCap. When employing Vicuna-7B, the comparison of performance trends mirrors those observed with Vicuna-13B, further attesting to the robustness and adaptability of EVCAP across different LLM decoders. Notably, both SmallCap, which retrieves captions, and our GPT2-based EVCAP, which retrieves object names, use the same GPT2 decoder. Therefore, their comparison also underscores the effectiveness of our method's object name retrieval and attentive fusion strategy." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 361, + 546, + 504 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 361, + 546, + 504 + ], + "spans": [ + { + "bbox": [ + 304, + 361, + 546, + 504 + ], + "type": "text", + "content": "Limitations. First, EVCAP cannot retrieve all objects that appear in the given image due to the memory coverage limits, leading to incomplete image descriptions (Fig. 4). We will investigate integrating object detection with image captioning to enhance completeness. Second, our focus on object representation restricts consideration of other crucial captioning elements, affecting overall performance. Similar to all models trained with COCO dataset, EVCAP has limitations in generating varied styles, which is reflected in our relatively modest performance improvements in Tab. 2, compared to MiniGPT4. We will overcome it by exploring methodologies that encourage style diversity in the future." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 518, + 379, + 532 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 518, + 379, + 532 + ], + "spans": [ + { + "bbox": [ + 306, + 518, + 379, + 532 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 540, + 545, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 540, + 545, + 635 + ], + "spans": [ + { + "bbox": [ + 304, + 540, + 545, + 635 + ], + "type": "text", + "content": "We further advance image captioning in real-world scenarios by introducing EVCAP, a novel image captioning model with object names retrieved from an external visual-name memory. The external memory is easily expandable, allowing for effortless updates with new object visuals and names. We extensively compare EVCAP with SOTAs on various benchmarks and commonsense-violating data, demonstrating its significant superiority in performance." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 635, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 635, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 635, + 547, + 714 + ], + "type": "text", + "content": "Acknowledgements. This work was supported by JSPS/MEXT KAKENHI Grant Numbers JP22H05015, JP23H03449, JP23KJ0404, and JP22K17947, and the commissioned research (No. 225) by the National Institute of Information and Communications Technology (NICT), Japan, and the Institute of AI and Beyond of the University of Tokyo, and ROIS NII Open Collaborative Research 2023-23FC01, 2024-24S1201." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "13740" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 714 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 113 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 113 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 113 + ], + "type": "text", + "content": "[1] https : / / huggingface . co / spaces / RitaParadaRamos/SmallCapDemo.5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 114, + 288, + 168 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 114, + 288, + 168 + ], + "spans": [ + { + "bbox": [ + 53, + 114, + 288, + 168 + ], + "type": "text", + "content": "[2] Harsh Agrawal, Karan Desai, Yufei Wang, Xinlei Chen, Rishabh Jain, Mark Johnson, Dhruv Batra, Devi Parikh, Stefan Lee, and Peter Anderson. Nocaps: Novel object captioning at scale. In Proc. IEEE International Conference on Computer Vision (ICCV), 2019. 4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 170, + 287, + 224 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 170, + 287, + 224 + ], + "spans": [ + { + "bbox": [ + 53, + 170, + 287, + 224 + ], + "type": "text", + "content": "[3] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: A visual language model for few-shot learning. In Advances in Neural Information Processing Systems (NeurIPS), 2022. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 225, + 287, + 280 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 225, + 287, + 280 + ], + "spans": [ + { + "bbox": [ + 53, + 225, + 287, + 280 + ], + "type": "text", + "content": "[4] Peter Anderson, Xiaodong He, Chris Buehler, Damien Teney, Mark Johnson, Stephen Gould, and Lei Zhang. Bottom-up and top-down attention for image captioning and visual question answering. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 281, + 287, + 325 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 281, + 287, + 325 + ], + "spans": [ + { + "bbox": [ + 53, + 281, + 287, + 325 + ], + "type": "text", + "content": "[5] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A frontier large vision-language model with versatile abilities. arXiv preprint arXiv:2308.12966, 2023. 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 326, + 287, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 326, + 287, + 392 + ], + "spans": [ + { + "bbox": [ + 53, + 326, + 287, + 392 + ], + "type": "text", + "content": "[6] Nitzan Bitton-Guetta, Yonatan Bitton, Jack Hessel, Ludwig Schmidt, Yuval Elovici, Gabriel Stanovsky, and Roy Schwartz. Breaking common sense: Whoops! a vision-and-language benchmark of synthetic and compositional images. In Proc. IEEE International Conference on Computer Vision (ICCV), 2023. 4, 6, 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 392, + 287, + 446 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 392, + 287, + 446 + ], + "spans": [ + { + "bbox": [ + 53, + 392, + 287, + 446 + ], + "type": "text", + "content": "[7] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. In Advances in Neural Information Processing Systems (NeurIPS), 2020. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 448, + 287, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 448, + 287, + 502 + ], + "spans": [ + { + "bbox": [ + 53, + 448, + 287, + 502 + ], + "type": "text", + "content": "[8] Xi Chen, Josip Djolonga, Piotr Padlewski, Basil Mustafa, Soravit Changpinyo, Jialin Wu, Carlos Riquelme Ruiz, Sebastian Goodman, Xiao Wang, Yi Tay, et al. PaLI-X: On scaling up a multilingual vision and language model. arXiv preprint arXiv:2305.18565, 2023. 1, 2, 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 503, + 287, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 503, + 287, + 624 + ], + "spans": [ + { + "bbox": [ + 53, + 503, + 287, + 624 + ], + "type": "text", + "content": "[9] Xi Chen, Xiao Wang, Soravit Changpinyo, AJ Piergiovanni, Piotr Padlewski, Daniel Salz, Sebastian Goodman, Adam Grycner, Basil Mustafa, Lucas Beyer, Alexander Kolesnikov, Joan Puigcerver, Nan Ding, Keran Rong, Hassan Akbari, Gaurav Mishra, Linting Xue, Ashish V Thapliyal, James Bradbury, Weicheng Kuo, Mojtaba Seyedhosseini, Chao Jia, Burcu Karagol Ayan, Carlos Riquelme Ruiz, Andreas Peter Steiner, Anelia Angelova, Xiaohua Zhai, Neil Houlsby, and Radu Soricut. PaLI: A jointly-scaled multilingual language-image model. In International Conference on Learning Representations (ICLR), 2023. 1, 2, 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 625, + 287, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 625, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 625, + 287, + 689 + ], + "type": "text", + "content": "[10] Wei-Lin Chiang, Zhuohan Li, Zi Lin, Ying Sheng, Zhanghao Wu, Hao Zhang, Lianmin Zheng, Siyuan Zhuang, Yonghao Zhuang, Joseph E Gonzalez, et al. Vicuna: An open-source chatbot impressing gpt-4 with " + }, + { + "bbox": [ + 48, + 625, + 287, + 689 + ], + "type": "inline_equation", + "content": "90\\%" + }, + { + "bbox": [ + 48, + 625, + 287, + 689 + ], + "type": "text", + "content": " chatgpt quality. See https://vicuna.lmsys.org (accessed 14 April 2023), 2023. 2, 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 691, + 287, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 287, + 714 + ], + "type": "text", + "content": "[11] Wenliang Dai, Junnan Li, Dongxu Li, Anthony Meng Huat Tiong, Junqi Zhao, Weisheng Wang, Boyang Li, Pascale" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 714 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 116 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 116 + ], + "type": "text", + "content": "Fung, and Steven Hoi. Instructclip: Towards general-purpose vision-language models with instruction tuning. In Advances in Neural Information Processing Systems (NeurIPS), 2023. 3, 5, 8" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 308, + 118, + 545, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 118, + 545, + 194 + ], + "spans": [ + { + "bbox": [ + 308, + 118, + 545, + 194 + ], + "type": "text", + "content": "[12] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations (ICLR), 2021. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 195, + 545, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 195, + 545, + 239 + ], + "spans": [ + { + "bbox": [ + 307, + 195, + 545, + 239 + ], + "type": "text", + "content": "[13] Jiashuo Fan, Yaoyuan Liang, Leyao Liu, Shaolun Huang, and Lei Zhang. Rca-noc: Relative contrastive alignment for novel object captioning. In Proc. IEEE International Conference on Computer Vision (ICCV), 2023. 2, 5" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 239, + 545, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 239, + 545, + 293 + ], + "spans": [ + { + "bbox": [ + 307, + 239, + 545, + 293 + ], + "type": "text", + "content": "[14] Yuxin Fang, Wen Wang, Binhui Xie, Quan Sun, Ledell Wu, Xinggang Wang, Tiejun Huang, Xinlong Wang, and Yue Cao. Eva: Exploring the limits of masked visual representation learning at scale. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 2, 4" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 294, + 545, + 347 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 294, + 545, + 347 + ], + "spans": [ + { + "bbox": [ + 307, + 294, + 545, + 347 + ], + "type": "text", + "content": "[15] Junjie Fei, Teng Wang, Jinrui Zhang, Zhenyu He, Chengjie Wang, and Feng Zheng. Transferable decoding with visual entities for zero-shot image captioning. In Proc. IEEE International Conference on Computer Vision (ICCV), 2023. 2, 4, 5, 8" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 350, + 545, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 350, + 545, + 381 + ], + "spans": [ + { + "bbox": [ + 307, + 350, + 545, + 381 + ], + "type": "text", + "content": "[16] Zhengcong Fei. Memory-augmented image captioning. In Proc. AAAI Conference on Artificial Intelligence (AAAI), 2021. 2, 5" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 383, + 545, + 425 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 383, + 545, + 425 + ], + "spans": [ + { + "bbox": [ + 307, + 383, + 545, + 425 + ], + "type": "text", + "content": "[17] Agrim Gupta, Piotr Dollar, and Ross Girshick. LVIS: A dataset for large vocabulary instance segmentation. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 426, + 545, + 469 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 426, + 545, + 469 + ], + "spans": [ + { + "bbox": [ + 307, + 426, + 545, + 469 + ], + "type": "text", + "content": "[18] Ryuichiro Hataya, Han Bao, and Hiromi Arai. Will large-scale generative models corrupt future datasets? In Proc. IEEE International Conference on Computer Vision (ICCV), 2023. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 472, + 545, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 472, + 545, + 536 + ], + "spans": [ + { + "bbox": [ + 307, + 472, + 545, + 536 + ], + "type": "text", + "content": "[19] Lisa Anne Hendricks, Subhashini Venugopalan, Marcus Rohrbach, Raymond Mooney, Saenko Kate, and Trevor Darrell. Deep compositional captioning: Describing novel object categories without paired training data. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 537, + 545, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 537, + 545, + 602 + ], + "spans": [ + { + "bbox": [ + 307, + 537, + 545, + 602 + ], + "type": "text", + "content": "[20] Ziniu Hu, Ahmet Iscen, Chen Sun, Zirui Wang, Kai-Wei Chang, Yizhou Sun, Cordelia Schmid, David A Ross, and Alireza Fathi. Reveal: Retrieval-augmented visual-language pre-training with multi-source multimodal knowledge memory. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 1, 2, 5" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 604, + 545, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 604, + 545, + 635 + ], + "spans": [ + { + "bbox": [ + 307, + 604, + 545, + 635 + ], + "type": "text", + "content": "[21] Jeff Johnson, Matthijs Douze, and Hervé Jégou. Billionscale similarity search with gpus. IEEE Transactions on Big Data, 7(3):535-547, 2019. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 636, + 545, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 636, + 545, + 679 + ], + "spans": [ + { + "bbox": [ + 307, + 636, + 545, + 679 + ], + "type": "text", + "content": "[22] Andrej Karpathy and Li Fei-Fei. Deep visual-semantic alignments for generating image descriptions. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2015. 2, 4" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 681, + 547, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 681, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 307, + 681, + 547, + 714 + ], + "type": "text", + "content": "[23] Jacob Devlin Ming-Wei Chang Kenton and Lee Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13741" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 66, + 72, + 287, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 72, + 287, + 106 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 287, + 106 + ], + "type": "text", + "content": "North American Chapter of the Association for Computational Linguistics: Human Language Technologies (NAACL-HLT), 2019. 4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 106, + 287, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 106, + 287, + 160 + ], + "spans": [ + { + "bbox": [ + 48, + 106, + 287, + 160 + ], + "type": "text", + "content": "[24] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In Proc. International conference on machine learning (ICML), 2022. 5, 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 161, + 287, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 161, + 287, + 215 + ], + "spans": [ + { + "bbox": [ + 48, + 161, + 287, + 215 + ], + "type": "text", + "content": "[25] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In Proc. International conference on machine learning (ICML), 2023. 1, 2, 3, 4, 5, 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 217, + 287, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 217, + 287, + 271 + ], + "spans": [ + { + "bbox": [ + 48, + 217, + 287, + 271 + ], + "type": "text", + "content": "[26] Xiujun Li, Xi Yin, Chunyuan Li, Pengchuan Zhang, Xiaowei Hu, Lei Zhang, Lijuan Wang, Houdong Hu, Li Dong, Furu Wei, et al. Oscar: Object-semantics aligned pre-training for vision-language tasks. In Proc. European Conference on Computer Vision (ECCV), 2020. 2, 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 272, + 287, + 325 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 272, + 287, + 325 + ], + "spans": [ + { + "bbox": [ + 48, + 272, + 287, + 325 + ], + "type": "text", + "content": "[27] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dálár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Proc. European Conference on Computer Vision (ECCV), 2014. 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 327, + 287, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 327, + 287, + 360 + ], + "spans": [ + { + "bbox": [ + 48, + 327, + 287, + 360 + ], + "type": "text", + "content": "[28] Jiasen Lu, Jianwei Yang, Dhruv Batra, and Devi Parikh. Neural baby talk. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 360, + 287, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 360, + 287, + 392 + ], + "spans": [ + { + "bbox": [ + 48, + 360, + 287, + 392 + ], + "type": "text", + "content": "[29] Ron Mokady, Amir Hertz, and Amit H Bermano. Clipcap: Clip prefix for image captioning. arXiv preprint arXiv:2111.09734, 2021. 2, 3, 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 393, + 287, + 458 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 393, + 287, + 458 + ], + "spans": [ + { + "bbox": [ + 48, + 393, + 287, + 458 + ], + "type": "text", + "content": "[30] Bryan A Plummer, Liwei Wang, Chris M Cervantes, Juan C Caicedo, Julia Hockenmaier, and Svetlana Lazebnik. Flickr30k entities: Collecting region-to-phrase correspondences for richer image-to-sentence models. In Proc. IEEE International Conference on Computer Vision (ICCV), 2015. 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 460, + 287, + 492 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 460, + 287, + 492 + ], + "spans": [ + { + "bbox": [ + 48, + 460, + 287, + 492 + ], + "type": "text", + "content": "[31] Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, Ilya Sutskever, et al. Language models are unsupervised multitask learners. OpenAI blog, 1(8):9, 2019. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 493, + 287, + 558 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 493, + 287, + 558 + ], + "spans": [ + { + "bbox": [ + 48, + 493, + 287, + 558 + ], + "type": "text", + "content": "[32] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In Proc. International conference on machine learning (ICML), 2021. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 559, + 287, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 559, + 287, + 613 + ], + "spans": [ + { + "bbox": [ + 48, + 559, + 287, + 613 + ], + "type": "text", + "content": "[33] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J Liu. Exploring the limits of transfer learning with a unified text-to-text transformer. Journal of Machine Learning Research (JMLR), 21(1):5485-5551, 2020. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 614, + 287, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 614, + 287, + 658 + ], + "spans": [ + { + "bbox": [ + 48, + 614, + 287, + 658 + ], + "type": "text", + "content": "[34] Rita Ramos, Desmond Elliott, and Bruno Martins. Retrievalaugmented image captioning. In Proc. Conference of the European Chapter of the Association for Computational Linguistics (EACL), 2023. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 658, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 287, + 712 + ], + "type": "text", + "content": "[35] Rita Ramos, Bruno Martins, Desmond Elliott, and Yova Kementchedjhieva. Smallcap: Lightweight image captioning prompted with retrieval augmentation. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 1, 2, 3, 5, 8" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 675 + ], + "type": "list", + "angle": 0, + "index": 25, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 127 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 127 + ], + "type": "text", + "content": "[36] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 3, 6" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 129, + 545, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 129, + 545, + 183 + ], + "spans": [ + { + "bbox": [ + 307, + 129, + 545, + 183 + ], + "type": "text", + "content": "[37] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 2, 4" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 185, + 545, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 185, + 545, + 217 + ], + "spans": [ + { + "bbox": [ + 307, + 185, + 545, + 217 + ], + "type": "text", + "content": "[38] Laurens Van der Maaten and Geoffrey Hinton. Visualizing data using t-sne. Journal of machine learning research (JMLR), 9(11), 2008. 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 219, + 545, + 272 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 219, + 545, + 272 + ], + "spans": [ + { + "bbox": [ + 307, + 219, + 545, + 272 + ], + "type": "text", + "content": "[39] Subhashini Venugopalan, Lisa Anne Hendricks, Marcus Rohrbach, Raymond Mooney, Trevor Darrell, and Kate Saenko. Captioning images with diverse objects. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2017. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 274, + 545, + 328 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 274, + 545, + 328 + ], + "spans": [ + { + "bbox": [ + 307, + 274, + 545, + 328 + ], + "type": "text", + "content": "[40] Duc Minh Vo, Hong Chen, Akihiro Sugimoto, and Hideki Nakayama. Noc-rek: Novel object captioning with retrieved vocabulary from external knowledge. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2, 5" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 330, + 545, + 395 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 330, + 545, + 395 + ], + "spans": [ + { + "bbox": [ + 307, + 330, + 545, + 395 + ], + "type": "text", + "content": "[41] Weihan Wang, Qingsong Lv, Wenmeng Yu, Wenyi Hong, Ji Qi, Yan Wang, Junhui Ji, Zhuoyi Yang, Lei Zhao, Xixuan Song, Jiazheng Xu, Bin Xu, Juanzi Li, Yuxiao Dong, Ming Ding, and Jie Tang. Cogvlm: Visual expert for pretrained language models. arXiv preprint arXiv:2311.03079, 2023. 4, 5" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 396, + 545, + 452 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 396, + 545, + 452 + ], + "spans": [ + { + "bbox": [ + 307, + 396, + 545, + 452 + ], + "type": "text", + "content": "[42] Kelvin Xu, Jimmy Ba, Ryan Kiros, Kyunghyun Cho, Aaron Courville, Ruslan Salakhudinov, Rich Zemel, and Yoshua Bengio. Show, attend and tell: Neural image caption generation with visual attention. In Proc. International conference on machine learning (ICML). 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 453, + 545, + 517 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 453, + 545, + 517 + ], + "spans": [ + { + "bbox": [ + 307, + 453, + 545, + 517 + ], + "type": "text", + "content": "[43] Zhuolin Yang, Wei Ping, Zihan Liu, Vijay Korthikanti, Weili Nie, De-An Huang, Linxi Fan, Zhiding Yu, Shiyi Lan, Bo Li, et al. Re-ViLM: Retrieval-augmented visual language model for zero and few-shot image captioning. In Findings of the Association for Computational Linguistics: EMNLP, 2023. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 518, + 545, + 574 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 518, + 545, + 574 + ], + "spans": [ + { + "bbox": [ + 307, + 518, + 545, + 574 + ], + "type": "text", + "content": "[44] Michihiro Yasunaga, Armen Aghajanyan, Weijia Shi, Richard James, Jure Leskovec, Percy Liang, Mike Lewis, Luke Zettlemoyer, and Wen-tau Yih. Retrieval-augmented multimodal language modeling. In International Conference on Learning Representations (ICLR), 2023. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 575, + 545, + 630 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 575, + 545, + 630 + ], + "spans": [ + { + "bbox": [ + 307, + 575, + 545, + 630 + ], + "type": "text", + "content": "[45] Pengchuan Zhang, Xiujun Li, Xiaowei Hu, Jianwei Yang, Lei Zhang, Lijuan Wang, Yejin Choi, and Jianfeng Gao. Vinvl: Revisiting visual representations in vision-language models. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 2, 5" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 631, + 545, + 675 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 631, + 545, + 675 + ], + "spans": [ + { + "bbox": [ + 307, + 631, + 545, + 675 + ], + "type": "text", + "content": "[46] Deyao Zhu, Jun Chen, Xiaoqian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models. arXiv preprint arXiv:2304.10592, 2023. 3, 4, 5, 7, 8" + } + ] + } + ], + "index": 24 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "13742" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/EVS-assisted Joint Deblurring Rolling-Shutter Correction and Video Frame Interpolation through Sensor Inverse Modeling/2babbe27-3090-44dc-8f9d-2d16b50cc5ab_content_list.json b/2024/EVS-assisted Joint Deblurring Rolling-Shutter Correction and Video Frame Interpolation through Sensor Inverse Modeling/2babbe27-3090-44dc-8f9d-2d16b50cc5ab_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..3da693216480e67b7d652dcb3f94e77aa5509994 --- /dev/null +++ b/2024/EVS-assisted Joint Deblurring Rolling-Shutter Correction and Video Frame Interpolation through Sensor Inverse Modeling/2babbe27-3090-44dc-8f9d-2d16b50cc5ab_content_list.json @@ -0,0 +1,1629 @@ +[ + { + "type": "text", + "text": "EVS-assisted joint Deblurring, Rolling-Shutter Correction and Video Frame Interpolation through Sensor Inverse Modeling", + "text_level": 1, + "bbox": [ + 101, + 128, + 870, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Rui Jiang*, Fangwen Tu*, Yixuan Long, Aabhaas Vaish, Bowen Zhou, Qinyi Wang Wei Zhang, Yuntan Fang, Luis Eduardo Garcia Capel, Bo Mu, Tiejun Dai, Andreas Suess", + "bbox": [ + 138, + 203, + 844, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "OMNIVISION", + "bbox": [ + 433, + 244, + 535, + 258 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "*Equal contribution", + "bbox": [ + 424, + 263, + 544, + 277 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/0a6598df303e0632b86ab8d92416d59c00dbbc089ffe296a5c611f8025de048d.jpg", + "image_caption": [ + "Figure 1. Qualitative results comparing the proposed method with EDI [27], AKF [39], Time Lens [37], FILM [32], CBMNet [23], EvUnroll [44], and REFID [36] on simulation datasets. In the \"basketball\" scene, our method provides frames that are rolling shutter effect corrected and deblurred. In the \"slingshot egg\" scene, our method can reconstruct the shape of the pellet with the best quality, as indicated in the yellow box. The images have been magnified for optimal viewing quality." + ], + "image_footnote": [], + "bbox": [ + 81, + 313, + 890, + 407 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 487, + 313, + 502 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Event-based Vision Sensors (EVS) gain popularity in enhancing CMOS Image Sensor (CIS) video capture. Nonidealities of EVS such as pixel or readout latency can significantly influence the quality of the enhanced images and warrant dedicated consideration in the design of fusion algorithms. A novel approach for jointly computing deblurred, rolling-shutter artifact corrected high-speed videos with frame rates up to 10 000 FPS using inherently blurry rolling shutter CIS frames of 120 FPS to 150 FPS in conjunction with EVS data from a hybrid CIS-EVS sensor is presented. EVS pixel latency, readout latency and the sensor's refractory period are explicitly incorporated into the measurement model. This inverse function problem is solved on a per-pixel manner using an optimization-based framework. The interpolated images are subsequently processed by a novel refinement network. The proposed method is evaluated using simulated and measured datasets, under natural and controlled environments. Extensive experiments show reduced shadowing effect, a 4 dB increment in PSNR, and a $12\\%$ improvement in LPIPS score compared to state-of-the-art methods.", + "bbox": [ + 73, + 518, + 473, + 835 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 500, + 487, + 630, + 502 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Event-Based Vision Sensors (EVS) capture our world from a completely different perspective compared to CMOS Image Sensors (CIS) [9, 14, 24, 25]. In EVS, each pixel independently triggers an event if a relative illuminance change on the pixel reaches a certain triggering threshold (typically in the range of $10\\%$ to $30\\%$ ). Conversely to CIS, EVS operate time-continuously and asynchronously - there is no common exposure period or \"frame rate\". Thus, EVS pixels can be read out comparatively fast in a low-power manner enabling efficient high-speed data capture. Recent work on enhancement of high-quality images [23, 27, 34, 37, 39] focuses on fusing CIS and EVS information so that the high fidelity of CIS measurement and the fast response of EVS measurement can complement each other.", + "bbox": [ + 496, + 508, + 892, + 719 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In practice it is observed [3, 14, 21, 25] that the dynamic characteristics of EVS pixels are highly dependent on the illumininance in the sensor plane. Under low-light conditions such as indoor capture, a slower EVS response during video enhancement can result in ghosting or blurry frames and needs to be considered in algorithm design. Furthermore, the inevitable readout latency and refractory period of EVS pixels can significantly aggravate artifacts. Current methods are not specifically designed to address sensor nonidealities.", + "bbox": [ + 496, + 720, + 893, + 857 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "25172", + "bbox": [ + 478, + 922, + 519, + 934 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "This paper presents an inverse model capturing the EVS pixel latency (namely the $V_{\\mathrm{FE}}$ delay), the readout latency, and the refractory period. The inverse model is solved through nonlinear graph optimization to jointly address deblurring, rolling-shutter (RS) correction and video-frame interpolation while compensating for sensor nonidealities. In addition to the inverse model, a learning-based refinement module is proposed to enhance the image quality by further mitigating the noise and artifacts. The proposed method is compared with existing CIS-based [32] and EVS-assisted [8, 23, 27, 36, 37, 39] video enhancement methods under challenging indoor scenarios with fast motion. As illustrated in Figure 1, the proposed algorithm produces excellent image quality with much fewer artifacts.", + "bbox": [ + 75, + 90, + 472, + 301 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The contributions of the paper can be summarized as:", + "bbox": [ + 96, + 303, + 450, + 316 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- A formulation of a joint deblurring, rolling-shutter correction and video-frame interpolation problem based on CIS and EVS fusion, where EVS sensor nonidealities are explicitly modeled (Section 3),", + "- An optimization methodology jointly solving this coupled inverse problem (Section 4.1), as well as a post-processing network for artifact removal (Section 4.2),", + "- A comprehensive comparison and analysis of simulation and measured datasets in natural and controlled environments (Section 5)." + ], + "bbox": [ + 94, + 319, + 467, + 468 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 76, + 481, + 218, + 496 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Video Frame Interpolation Video interpolation aims to synthesize intermediate frames between two images to achieve a temporally coherent video sequence. Traditional CIS-based methods depend on optical flow for a smooth transition. In SuperSlomo [18], the bi-directional flows are combined with a linear combination followed by a refinement using U-Net [33]. Such optical flow-based methods, however, suffer from occlusion problems. In DAIN [1], researchers estimate occlusion areas using depth information and use contextual information from neighboring pixels to fill the occluded regions. In RIFE [17], the flow and fusion maps are generated simultaneously without an additional optical flow module for final result synthesis. FILM [32] presents a multi-scale feature extractor within a unified network to achieve an enlarged receptive field aimed at handling large object motion in the video.", + "bbox": [ + 75, + 503, + 472, + 744 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Event-assisted Frame Enhancement A key issue lowering the image quality of CIS-based frame interpolation methods is the lack of information between frames. EVS provides such information thanks to its asynchronous low latency capture not relying on a global exposure period. Some early studies [31, 34] show that intensity reconstruction purely based on EVS modality is feasible. However, as EVS captures events based on differential changes of illuminance, this approach struggles to estimate the absolute light level. In [2], the optical flow and intensity frames are", + "bbox": [ + 75, + 750, + 472, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "estimated simultaneously by solving an optimization problem that takes EVS measurements as data terms. The authors in [22] take a step further by formulating the event-based SLAM problem, which can be decomposed into camera ego-motion estimation and 3-D scene reconstruction. However, the quantization error makes it difficult to reconstruct high-quality frames from EVS alone. As EVS pixels trigger independently, researchers simplify the problem formulation to pixel-wise estimation by fusing CIS and EVS, without seeking explicit representation of the 3-D environment. The Event-based Double Integral (EDI) [27, 28] method is proposed as a straightforward and effective way to deblur CIS frames and generate high frame rate videos. EVS and CIS measurements are connected by integrating events twice, such that the continuous intensity curve for each pixel can be determined. [34] presents a complementary filter-based framework to combine a high-pass signal from EVS and a low-pass signal from CIS to compute an all-pass signal for pixel intensity. To make the complementary filter gain adaptive to measurement noise, a pixel-wise Kalman filter is used in [39] where a unifying EVS/CIS uncertainty model is proposed.", + "bbox": [ + 496, + 90, + 890, + 422 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Machine learning methods such as Time Lens [37] and its subsequent versions [10, 36, 38, 41] or others [29, 43] concatenate both CIS and EVS data using a voxel grid representation [12] as input to feed the network of two branches: an optical flow warping-based branch and a synthesis-based branch. The final result is blended with a trained attention module to avoid the disadvantages of any individual branch. CBMNet [23] is proposed to handle complex real-world motion using a novel cross-modal asymmetric bidirectional motion field estimation. Due to the high temporal resolution of EVS data, the rolling shutter problem which is commonly observed in fast-motion scenarios can be corrected using EVS data [6, 44].", + "bbox": [ + 496, + 425, + 890, + 621 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "EVS Latency Many studies model EVS latency [20, 25, 30, 35], but few of them have been applied to algorithm design such as frame enhancement. To the best of the authors' knowledge, this is the first work that considers an EVS latency model explicitly in generating high frame rate, deblurred, rolling-shutter artifact-free videos.", + "bbox": [ + 496, + 633, + 890, + 724 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3. Problem Formulation", + "text_level": 1, + "bbox": [ + 500, + 739, + 705, + 756 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Event Pixel Operation and Pixel Latency Model In this work, a hybrid CIS-EVS sensor [14, 15] is used. As shown in Figure 2(a), EVS and CIS share the same photodiode design on the hybrid sensor and it is assumed that the light level is identical for nearby CIS and EVS pixels. The photocurrent $i_{\\mathrm{pd}}$ is converted to front-end voltage $V_{\\mathrm{FE}}$ in module $\\mathrm{i}2\\mathrm{v}$ , which models dynamic characteristics as a 1st-order Low-Pass Filter (LPF) [21, 26]. The update rule for $V_{\\mathrm{FE}}$ between discrete event indices $k$ and $k + 1$ can be written as:", + "bbox": [ + 496, + 763, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "25173", + "bbox": [ + 478, + 922, + 519, + 934 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/d5b8847e1319d15d45a2c52e093e33b06a6599f473f2183a404a6577da4202ab.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 83, + 90, + 477, + 157 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/d400ca00c23e5f4cac2721238af5c6c712e24b8f18b992c3794696c18c2b4678.jpg", + "image_caption": [ + "Figure 2. (a) Hybrid CIS-EVS sensor principle. The upper and lower branches indicate physical quantity conversion in CIS pixels and EVS pixels, respectively. (b) Example of $V_{\\mathrm{DC}}$ and $V_{\\mathrm{FE}}$ curves and triggered events that illustrate the impact of $V_{\\mathrm{FE}}$ delay, readout latency, and refractory period." + ], + "image_footnote": [], + "bbox": [ + 83, + 169, + 468, + 305 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nV _ {\\mathrm {F E}} (k + 1) = V _ {\\mathrm {F E}} (k) + \\gamma (k) \\left(V _ {\\mathrm {D C}} (k + 1) - V _ {\\mathrm {F E}} (k)\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 96, + 429, + 468, + 445 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $V_{\\mathrm{DC}}(k + 1) = f_{\\mathrm{DC}}(i_{\\mathrm{pd}}(k + 1))$ represents the $V_{\\mathrm{FE}}$ under DC excitation (i.e., the $V_{\\mathrm{FE}}$ with infinite LPF bandwidth); $f_{\\mathrm{DC}}(i_{\\mathrm{pd}})$ is the logarithmic-like current-voltage conversion function and", + "bbox": [ + 76, + 448, + 468, + 506 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\gamma (k) = 1 - \\exp \\left(- \\frac {\\Delta t}{\\tau (k)}\\right) \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 179, + 508, + 468, + 541 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "is the time-varying coefficient of the LPF, where $\\Delta t$ is the time interval between two indices. The photocurrent-dependent time constant $\\tau (k)$ can be modeled as [16, 26]:", + "bbox": [ + 75, + 542, + 468, + 587 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\tau (k) = \\tau_ {0} + \\frac {\\alpha}{i _ {\\mathrm {p d}} (k)}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 184, + 590, + 468, + 619 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\tau_0$ and $\\alpha$ are pixel parameters. The pixel latency in module $\\pm 2\\mathrm{v}$ due to the LPF behavior is called \" $V_{\\mathrm{FE}}$ delay\". As illustrated in Figure 2(b), an event is triggered at time $t_\\mathrm{in}$ once the difference between $V_{\\mathrm{FE}}$ and reference voltage $V_{\\mathrm{ref}}$ reaches the triggering threshold voltage:", + "bbox": [ + 75, + 621, + 468, + 696 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\left\\{ \\begin{array}{l l} V _ {\\mathrm {F E}} \\left(t _ {\\mathrm {i n}}\\right) - V _ {\\mathrm {r e f}} \\left(t _ {\\mathrm {i n}}\\right) & \\geq c _ {\\mathrm {u p}} \\quad \\text {f o r p o s i t i v e e v e n t s} \\\\ V _ {\\mathrm {F E}} \\left(t _ {\\mathrm {i n}}\\right) - V _ {\\mathrm {r e f}} \\left(t _ {\\mathrm {i n}}\\right) & \\leq c _ {\\mathrm {d o w n}} \\quad \\text {f o r n e g a t i v e e v e n t s} \\end{array} , \\right. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 78, + 699, + 468, + 731 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $c_{\\mathrm{up / down}}$ denotes the contrast threshold voltage. The in-pixel timestamp $t_{\\mathrm{in}}$ is recorded when the event fires, after which the pixel experiences \"readout latency $t_{\\mathrm{rl}}$ \" and \"refractory period $t_{\\mathrm{rp}}$ \". The $V_{\\mathrm{ref}}$ is updated to follow $V_{\\mathrm{FE}}$ after the refractory period. Finally, we have a 5-dimensional output $\\{x,y,t_{\\mathrm{in}},t_{\\mathrm{off}},p\\}$ , which contains row and column indices, in-pixel and off-pixel timestamps, and polarity.", + "bbox": [ + 75, + 733, + 468, + 838 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Pixel-wise Photocurrent Estimation Suppose that the illuminance on the sensor plane results in photocurrent $i_{\\mathrm{pd}}$ for a single pixel. We aim to estimate $i_{\\mathrm{pd}}$ when an event is triggered to avoid the impact of pixel latency in the $\\pm 2\\mathrm{v}$", + "bbox": [ + 75, + 839, + 468, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "module. Based on Figure 2(a), we have the following CIS measurement equation:", + "bbox": [ + 498, + 90, + 890, + 119 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nD N = G \\int_ {t _ {\\mathrm {s}}} ^ {t _ {\\mathrm {e}}} i _ {\\mathrm {p d}} (t) \\mathrm {d} t, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 599, + 123, + 890, + 157 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $DN$ is the CIS intensity frame measured in digital numbers; $G$ indicates the total gain of charge-voltage conversion and ADC; $t_{\\mathrm{s}}$ and $t_{\\mathrm{e}}$ denote the CIS exposure start and end, respectively. As for EVS measurements, suppose we have an event sequence with $N$ events $\\{x(k),y(k),t_{\\mathrm{in}}(k),t_{\\mathrm{off}}(k),p(k)\\}$ . By assuming $V_{\\mathrm{FE}}(1) = V_{\\mathrm{DC}}(1)$ , with Eq. (1) we have the recurrence relation:", + "bbox": [ + 498, + 159, + 890, + 265 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nV _ {\\mathrm {F E}} (1) = f _ {\\mathrm {D C}} \\left(i _ {\\mathrm {p d}} (1)\\right) \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 526, + 268, + 890, + 285 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nV _ {\\mathrm {F E}} (k + 1) = f _ {\\gamma} \\left(i _ {\\mathrm {p d}} (k)\\right) \\left(f _ {\\mathrm {D C}} \\left(i _ {\\mathrm {p d}} (k + 1)\\right) - V _ {\\mathrm {F E}} (k)\\right), \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 498, + 287, + 890, + 304 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $k \\in \\{1, \\dots, N-1\\}$ and $f_{\\gamma}(\\cdot)$ denotes the function that maps $i_{\\mathrm{pd}}$ onto $\\gamma$ . By assuming $V_{\\mathrm{ref}}(k+1) \\approx V_{\\mathrm{FE}}(k)$ , the EVS measurements provide additional $N-1$ equations as:", + "bbox": [ + 498, + 306, + 890, + 364 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nc (k + 1) \\approx V _ {\\mathrm {F E}} (k + 1) - V _ {\\mathrm {F E}} (k), \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 565, + 371, + 890, + 386 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $c(k) \\in \\{c_{\\mathrm{up}}, c_{\\mathrm{down}}\\}$ are the known triggering thresholds that are configurable. These state transition equations and measurement equations (6)-(8) can be expressed as a system of $2N - 1$ nonlinear equations and $2N$ unknowns $-i_{\\mathrm{pd}}(k)$ and $V_{\\mathrm{FE}}(k)$ . By using a zero-order holder, the discrete states $i_{\\mathrm{pd}}(k)$ and $V_{\\mathrm{FE}}(k)$ can be approximated as continuous signals $i_{\\mathrm{pd}}(t)$ and $V_{\\mathrm{FE}}(t)$ . This makes it possible to utilize CIS measurements in Eq. (5) as additional constraints. In conclusion, the pixel latency compensation can be expressed as pixel-wise optimization problem, solving the nonlinear system with $2N$ unknown states and $2N + M - 1$ equations, where $M$ and $N$ denote the number of CIS and EVS measurements, respectively.", + "bbox": [ + 496, + 388, + 890, + 585 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Modeling Readout Latency and Refractory Period The EVS measurement model equation (8) is an approximation since the state variation during the readout latency and the refractory period is not considered. Using both in-pixel timestamps and off-pixel timestamps, it is possible to compensate for some of the \"missed\" $V_{\\mathrm{FE}}$ change by assuming that the slope of $V_{\\mathrm{FE}}$ is locally constant for a short period. As illustrated in Figure 2(b), it is noted that:", + "bbox": [ + 496, + 611, + 890, + 731 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nm (k + 1) = \\frac {c (k + 1)}{t _ {\\mathrm {i n}} (k + 1) - t _ {\\mathrm {o f f}} (k) - t _ {\\mathrm {r p}}}, \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 547, + 734, + 890, + 767 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $m$ describes the slope approximation of the $V_{\\mathrm{FE}}$ curve. For the readout latency, the missed $V_{\\mathrm{FE}}$ is represented as:", + "bbox": [ + 496, + 768, + 890, + 811 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\Delta V _ {\\mathrm {F E}} ^ {\\mathrm {r l}} (k) = m (k + 1) \\left[ t _ {\\text {o f f}} (k) - t _ {\\text {i n}} (k) \\right]. \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 545, + 816, + 890, + 834 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "For the refractory period, we have $\\Delta V_{\\mathrm{FE}}^{\\mathrm{rp}}(k) = m(k + 1)t_{\\mathrm{rp}}$ since $t_\\mathrm{rp}$ is a fixed parameter. Thus the reference voltage after compensation can be adjusted to:", + "bbox": [ + 496, + 835, + 890, + 881 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nV _ {\\text {r e f}} (k + 1) = V _ {\\mathrm {F E}} (k) + \\Delta V _ {\\mathrm {F E}} ^ {\\mathrm {r l}} (k) + \\Delta V _ {\\mathrm {F E}} ^ {\\mathrm {r p}} (k). \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 524, + 883, + 890, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "25174", + "bbox": [ + 478, + 922, + 517, + 934 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/8fd2a7b91b0979a19d54ffcd360ed1de72ee3a74f991f18b974e821b8c1bda04.jpg", + "image_caption": [ + "Figure 3. Block diagram of the proposed method. Note that after the frame reconstruction the blurry texture due to motion is enhanced and the RS effect is removed. The noise and artifacts are further reduced by the refinement network." + ], + "image_footnote": [], + "bbox": [ + 80, + 88, + 478, + 265 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "By substituting the compensation equation (11) into (4), the impact of readout latency and refractory period can be compensated:", + "bbox": [ + 75, + 362, + 468, + 407 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "$c(k + 1)\\approx V_{\\mathrm{FE}}(k + 1) - V_{\\mathrm{FE}}(k) - \\Delta V_{\\mathrm{FE}}^{\\mathrm{sl}}(k) - \\Delta V_{\\mathrm{FE}}^{\\mathrm{rp}}(k),$ (12) where $\\Delta V_{\\mathrm{FE}}^{\\mathrm{rl}}(k)$ and $\\Delta V_{\\mathrm{FE}}^{\\mathrm{rp}}(k)$ are voltage compensation for readout latency and refractory period, respectively. Note that this compensation is enabled by the availability of in-pixel and off-pixel timestamps.", + "bbox": [ + 76, + 410, + 468, + 492 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Method", + "text_level": 1, + "bbox": [ + 76, + 500, + 168, + 515 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The proposed frame enhancement framework is presented in Figure 3. The frame reconstruction module jointly deblurs images, removes the RS effect, and interpolates high frame rate images. The refinement network focuses on removing noise and improving image quality.", + "bbox": [ + 75, + 522, + 468, + 599 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1. Solving the Inverse Model via Optimization", + "text_level": 1, + "bbox": [ + 76, + 606, + 447, + 625 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Graph Modeling for a Single Pixel Due to the presence of pixel latency, the system is governed by nonlinear equations, rendering the models and solutions designed for linear systems inapplicable. Graph optimization [13] provides a straightforward way to model nonlinear systems, where the states and measurements are represented as nodes and edges in a graph as shown in Figure 4(a). Let $l \\in \\{1, \\dots, M\\}$ and $k \\in \\{1, \\dots, N-1\\}$ be the indices of CIS and EVS, respectively. By defining states $\\boldsymbol{x} = [i_{\\mathrm{pd}}(k)]^{\\top}$ and measurements $z_{\\mathrm{CIS}} = [DN(l)]^{\\top}$ , $z_{\\mathrm{EVS}} = [c(k+1)]^{\\top}$ , the pixel-wise graph optimization framework aims to determine the state vector $\\boldsymbol{x}$ :", + "bbox": [ + 75, + 630, + 468, + 810 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {x} ^ {*} = \\underset {\\boldsymbol {x}} {\\operatorname {a r g m i n}} \\left(\\boldsymbol {e} _ {\\mathrm {C I S}} ^ {\\top} \\boldsymbol {\\Omega} _ {\\mathrm {C I S}} \\boldsymbol {e} _ {\\mathrm {C I S}} + \\boldsymbol {e} _ {\\mathrm {E V S}} ^ {\\top} \\boldsymbol {\\Omega} _ {\\mathrm {E V S}} \\boldsymbol {e} _ {\\mathrm {E V S}}\\right) \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 99, + 811, + 468, + 840 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $e_{\\mathrm{CIS}} = z_{\\mathrm{CIS}} - \\tilde{z}_{\\mathrm{CIS}}$ , $e_{\\mathrm{EVS}} = z_{\\mathrm{EVS}} - \\tilde{z}_{\\mathrm{EVS}}$ denote measurement error vectors; $\\tilde{z}_{\\mathrm{CIS}}$ and $\\tilde{z}_{\\mathrm{EVS}}$ are predicted measurements which can be calculated from estimated states according to CIS measurement equation (5) and EVS mea", + "bbox": [ + 75, + 840, + 470, + 901 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/77e4328e7b85c58d782e23bb33abec072591107f7992dbc86e7c20e55805411f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 501, + 90, + 895, + 210 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/cb62d3b02e5fa2c6832a628e5522cd82486490e12ccbfcbc8ebc60eb1bd21f54.jpg", + "image_caption": [ + "Figure 4. (a) Graph modeling of the pixel-wise photocurrent estimation problem, where nodes and edges denote states and measurements. The arrows indicate the order of state calculation. \"zCIS\" and \"zEVS\" indicate CIS and EVS measurements, respectively. Eqn. (8) is replaced with Eqn. (12) when voltage compensation for readout latency and refractory period is enabled. (b) Row-based reconstruction scheme showing three CIS images as input. After row-wise computation, interpolated and deblurred frames without RS effect can be generated as shown in the blue box. The proposed scheme can be applied to video enhancement with any number of CIS images." + ], + "image_footnote": [], + "bbox": [ + 511, + 223, + 897, + 375 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "surement equation (8); $z_{\\mathrm{EVS}}$ is adjusted using register settings and calibrated using the \"S-curve method\" [7]. The statistical distribution of the threshold is not explicitly modeled in the current method. The diagonal weighing matrices $\\Omega_{\\mathrm{CIS}}$ and $\\Omega_{\\mathrm{EVS}}$ reflect the reliability of CIS and EVS, respectively. Using the cost function equation (13), the $V_{\\mathrm{FE}}$ delay is modeled explicitly through the EVS measurement equations. Solving the optimization problem leads to state estimation such that the $V_{\\mathrm{FE}}$ delay is compensated. Many approaches have been proposed to solve nonlinear optimization problems iteratively [19]. We parallelize pixelwise tasks as there is no data interaction or computational dependency between pixels.", + "bbox": [ + 496, + 565, + 892, + 762 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Rolling Shutter (RS) Effect Correction As CIS parameters such as row exposure time and RS scanning speed are known, we formulate the frame estimation problem as pixel-wise computation using adjusted CIS timing parameters for each row. After solving the photocurrent at each event time-step through optimization, a zero-order hold is used to generate a time-continuous photocurrent $i_{\\mathrm{pd}}(t)$ for all pixels. The CIS measurement equation (5) is used to generate a rolling shutter artifact-free target frame rate (such", + "bbox": [ + 496, + 763, + 893, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "25175", + "bbox": [ + 478, + 922, + 519, + 934 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/a941213ff7db73c7ecfd1ea401ec15d50529b2461cb01d3bfcfcbec8f943f606.jpg", + "image_caption": [ + "Figure 5. Modified NAF block in the refinement network." + ], + "image_footnote": [], + "bbox": [ + 114, + 92, + 436, + 258 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "as 10 000 FPS).", + "bbox": [ + 76, + 305, + 187, + 319 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Video Glitch Effect Deblurring and video frame interpolation performance can experience glitches in case these are computed on a frame-by-frame basis. To mitigate this issue, we solve the coupled deblurring, rolling-shutter correction and interpolation problem jointly for subsequent frames utilizing the row-specific rolling shutter timing characteristics as shown in Figure 4(b), where we reconstruct each row according to its specific exposure time by fusing two consecutive CIS frames and their corresponding events.", + "bbox": [ + 75, + 329, + 470, + 465 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Refinement Network", + "text_level": 1, + "bbox": [ + 76, + 479, + 272, + 494 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "A refinement network is developed as a post-processing module to further mitigate the noise and artifacts in the reconstructed frames. NAFNet [4] is designed for image restoration. It is structured as a U-Net, arranging NAF blocks to allow the network to progressively downsample the input for semantic feature extraction and then progressively upsample it to restore the output. NAFNet provides additional skip connections that preserve information and allow feature reusability from earlier layers. NAF blocks are structured as original ResNet blocks, with additional layers such as layer normalization and channel attention. We adopt the NAFNet basic concept, and further modify making two major additions: 1) spatial attention [40] as shown in Figure 5 and 2) perceptual loss [32]. The reason for choosing spatial attention is that most of the noise is located in motion areas instead of being ubiquitous across the image. Spatial attention allows the network to focus on these noisy regions of the image, which contributes to better denoising results. Moreover, the rationale for choosing the perceptual loss is that even though the proposed network does well in removing the pixel-level saturation noise, it also introduces blurriness or texture degradation. Perceptual loss helps to recover the image quality in a way that better aligns with the human vision system. The losses applied in this paper are summarized as:", + "bbox": [ + 75, + 503, + 470, + 881 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {s}} = w _ {l} \\mathcal {L} _ {1} + w _ {\\mathrm {V G G}} \\mathcal {L} _ {\\mathrm {V G G}} + w _ {\\text {G r a m}} \\mathcal {L} _ {\\text {G r a m}}, \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 135, + 885, + 468, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where the $\\mathcal{L}_1$ loss is the pixel-wise difference between the reconstructed RGB image and ground truth image; the $\\mathcal{L}_{\\mathrm{VGG}}$ loss is calculated as $\\mathcal{L}_{\\mathrm{VGG}} = \\frac{1}{L}\\sum_{l=1}^{L}\\alpha_l\\|\\Psi_l(\\boldsymbol{I}_{\\mathrm{Recon}}) - \\Psi_l(\\boldsymbol{I}_{\\mathrm{GT}})\\|_1;\\Psi_l(\\boldsymbol{I})$ is the feature tensor from the $l_{\\mathrm{th}}$ layer of an ImageNet pre-trained VGG-19 network generated by passing the image $\\boldsymbol{I}$ through the network. In this case, $L$ specifies the total number of layers in the VGG-19 network, while $\\alpha_l$ is the hyperparameter assigned to each layer. Finally, the Gram loss [11] is computed as the L2 difference between the autocorrelation of the VGG-19 features of the reconstructed image and ground truth image with the formulation: $\\mathcal{L}_{\\mathrm{Gram}} = \\frac{1}{L}\\sum_{l=1}^{L}\\alpha_l\\|\\boldsymbol{M}_l(\\boldsymbol{I}_{\\mathrm{Recon}}) - \\boldsymbol{M}_l(\\boldsymbol{I}_{\\mathrm{GT}})\\|_2$ , where given an image $I$ , the Gram matrix is computed for every layer $l$ as $\\boldsymbol{M}_l = (\\Psi_l(\\boldsymbol{I}))^\\top(\\Psi_l(\\boldsymbol{I}))$ .", + "bbox": [ + 498, + 90, + 893, + 319 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 500, + 329, + 633, + 345 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We compare our proposed method with state-of-the-art EVS+CIS video reconstruction methods: EDI [27], AKF [39], Time Lens [37], CBMNet [23], EvUnroll [8] and RE-FID [36]. Among them, EvUnroll proposes an RS correction based reconstruction and REFID combines image de-burring with EVS based frame interpolation. We also show the results of the CIS-only interpolation method FILM [32] as reference. Before optimization, $i_{\\mathrm{pd}}$ are initialized according to the first CIS image DN and the inverse photocurrent-DN relation. The elements in the weighing matrices for CIS, EVS terms are set to 1. The performance of the interpolation methods are evaluated both quantitatively and qualitatively, based on the metrics of Peak Signal-to-Noise Ratio (PSNR), LPIPS [42], and pixel-wise reconstruction error on DN. Furthermore, to study the performance degradation concerning ambient illuminance and motion speed, we measured a test set utilizing a rotating disk in a controlled environment. We analyze the Blurred Edge Width (BEW) [5] of the reconstructed frames - a larger BEW indicates a blurrier image and, therefore, a worse deblurring result.", + "bbox": [ + 498, + 352, + 893, + 667 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.1. Datasets", + "text_level": 1, + "bbox": [ + 500, + 680, + 601, + 694 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We evaluate the proposed method using the simulation data based on the CIS-EVS hybrid sensor simulator [26] which considers $V_{\\mathrm{FE}}$ delay, readout latency, and refractory period. Our measurement data was collected from the hybrid CIS-EVS sensor [14, 15]. Here, only off-pixel times- tamp data was captured. Two measured datasets were created for evaluation: The natural scene dataset focuses on indoor scenarios where EVS is significantly influenced by pixel latency. The other dataset captured a rotating disk with a Siemens star pattern under different ambient illumination levels and rotating speeds. In addition to the above datasets specifically captured for this work's evaluation, we also test our method using the publicly available HS-ERGB", + "bbox": [ + 498, + 704, + 893, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "25176", + "bbox": [ + 478, + 922, + 519, + 934 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/6a1cfcfbeaaec01dc9072d902d370e628e102cfe0ea3a14aca6489a36dd83142.jpg", + "table_caption": [ + "Table 1. Quantitative comparison of state-of-the-art video reconstruction methods on the proposed simulation dataset in terms of PSNR (in dB, higher is better) and LPIPS (a dimensionless quantity, lower is better). Each row shows results for a particular scene. Results from different methods are listed column-wise. The first and second places are highlighted with bold underline and bold, respectively." + ], + "table_footnote": [], + "table_body": "
Ours w/o refinementOursEDIAKFTime LensFILMCBMNetEvUnrollREFID
ScenePSNR ↑LPIPS ↓PSNR ↑LPIPS ↓PSNR ↑LPIPS ↓PSNR ↑LPIPS ↓PSNR ↑LPIPS ↓PSNR ↑LPIPS ↓PSNR ↑LPIPS ↓PSNR ↑LPIPS ↓PSNR ↑LPIPS ↓
basketball24.010.21727.230.15522.490.22015.630.28023.650.24023.860.22024.510.25223.130.21822.380.223
checkerboard24.230.20328.240.18720.480.21819.560.26122.600.23923.020.21622.760.21223.840.20421.230.235
slingshot egg25.230.10032.850.10624.950.16415.590.11025.190.24424.040.20322.760.23424.610.21023.920.201
running man24.600.16424.720.18922.290.13520.710.18925.010.11725.060.08925.160.06923.040.11723.520.163
fan22.400.18924.500.16318.170.25510.780.18420.640.23318.930.18421.280.20419.640.19018.300.198
Average24.090.17527.510.1621.680.19816.450.20523.420.21522.980.18223.290.19422.850.18821.870.204
", + "bbox": [ + 78, + 141, + 893, + 227 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/e8d383a756fe1bcf0a009e9fefde5c2ef96c829688b6f2dbc11ef1f15a803c21.jpg", + "table_caption": [ + "Table 2. Quantitative results of ablation study. The first and second places are highlighted with bold underline and bold, respectively." + ], + "table_footnote": [], + "table_body": "
without compensationwith VFE delay compensationwith RL+RP compensationwith VFE+RL+RP compensationwith VFE+RL+RP compensation and refinement
ScenePSNR ↑LPIPS ↓PSNR ↑LPIPS ↓PSNR ↑LPIPS ↓PSNR ↑LPIPS ↓PSNR ↑LPIPS ↓
basketball22.950.31623.470.30623.500.30723.560.30223.640.295
slingshot egg23.570.09924.000.09923.600.09923.710.10229.040.106
", + "bbox": [ + 114, + 265, + 854, + 335 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/29969f719a04c87b0e6c3220e145fad3cf4ee891447d76451681b3ebf152ada7.jpg", + "image_caption": [ + "Figure 6. Qualitative comparison of state-of-the-art video reconstruction methods on the proposed measured datasets in terms of image quality. The images have been magnified for optimal viewing quality." + ], + "image_footnote": [], + "bbox": [ + 81, + 347, + 890, + 554 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/8f9a77d30aee6659949c823f28d806ae217c9ca3960f79364c9fe5d1f26a5d6f.jpg", + "image_caption": [ + "Figure 7. Frame-wise, quantitative comparison of state-of-the-art video reconstruction methods in terms of PSNR for varying scenarios of the simulation dataset. Note that some algorithms exhibit pronounced discontinuities in the observed PSNR curves indicating the presence of video glitches." + ], + "image_footnote": [], + "bbox": [ + 78, + 609, + 230, + 707 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/40d56ba35ed8aa577db3383cf57a6cfa73d930d79242859c7d50e167ac2cb6b1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 235, + 613, + 372, + 707 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/918086ff77003e143e0ae0536c41b0aa403cd75bed8fcf0e47d481014d414e61.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 375, + 612, + 514, + 707 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/1ec2efe781dc867223e22b3a398dd3345f5e0f4b5f5fdb6dffb6aefcf94ffa80.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 517, + 612, + 656, + 707 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/7628439d5201acf790fe3536544619871746a1aa2c0c5fca2e5906fbc7bbce65.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 658, + 611, + 795, + 707 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/4cb9e1d501fb3607c001cf4c54594e197aa4e1596d7da03ebc296894e1f0ffa9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 805, + 609, + 885, + 709 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "dataset based on a EVS camera that is collocated next to a CIS camera. Since the paper focuses on mitigating artifacts due to pixel latency, two indoor scenarios (\"spinning plate\" and \"spinning umbrella\") were selected. More details of the dataset used are provided in the supplementary material.", + "bbox": [ + 75, + 796, + 470, + 875 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.2. Results", + "text_level": 1, + "bbox": [ + 500, + 796, + 594, + 811 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Image Quality Table 1 shows quantitative results for the reconstructed video. Results in PSNR and LPIPS show that the proposed method outperforms other state-of-the-art methods in most of the scenes, especially with fast motion (\"slingshot egg\") or complicated rotation motion (\"basket-", + "bbox": [ + 498, + 825, + 893, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "25177", + "bbox": [ + 478, + 922, + 519, + 934 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "ball\", \"fan\"). As expected the refinement network which aims to alleviate artifacts and improve image quality reports a higher score. Snippets of the reconstructed videos are presented in Figure 1 and Figure 6 for simulated and measured data, respectively. Comprehensive input images are provided in the supplementary material. Thanks to the explicit EVS latency modeling and compensation, our results avoid obvious shadowing/ghosting artifacts - especially if there is fast motion under challenging lighting conditions. It is also noted that the proposed method successfully corrects the RS effect, as shown in the \"basketball\" scene. Learning-based methods face difficulties, as latency and the RS effect have not been considered in model design and training. The CIS-only method FILM provides natural image quality with low noise but fails when the objects are too fast to be captured by CIS, as shown in the \"slingshot egg\" scene. For the \"basketball\" scene, our method shows an advantage in details and sharpness compared to FILM.", + "bbox": [ + 75, + 90, + 472, + 361 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In specific scenarios, such as the \"spinning umbrella\" scene within the HS-ERGB dataset and the \"running man\" scene in the simulation dataset, our method did not yield improved image quality. This observation can be attributed to several factors. Firstly, the HS-ERGB dataset employs a different EVS with different pixel and readout latency as well as refractory period. Secondly, this dataset is based on a dual camera instead of a hybrid sensor so the spatiotemporal alignment may differ. Thirdly, the scenes in question involve relatively slower object movement speeds resulting in reduced ghosting. In essence, the dynamic characteristics of the EVS, the input signal frequency range, and the reconstruction algorithm influence the quality of the reconstructed images.", + "bbox": [ + 75, + 363, + 472, + 575 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Video Glitch Effect To evaluate the glitch issue, we compute the PSNR for each reconstructed frame as shown in Figure 7. Overall, our method gives temporally more stable PSNR curves, which stem from the proposed row-based reconstruction scheme. It is also noted that the refinement network contributes significantly to higher PSNR.", + "bbox": [ + 75, + 579, + 468, + 670 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Pixel-wise Error Analysis We select a pixel within the region of interest (indicated as the yellow box in Figure 1(b)) and verify the pixel-wise reconstruction accuracy. As shown in Figure 8(a), with the help of pixel latency compensation the reconstructed $V_{\\mathrm{DC}}$ after optimization shows a sharper peak (with $250~\\mu \\mathrm{s}$ duration) compared to $V_{\\mathrm{FE}}$ (with over $1\\mathrm{ms}$ duration). The high-frequency fluctuations on ground truth $V_{\\mathrm{DC}}$ curve cannot be reconstructed due to the $V_{\\mathrm{FE}}$ bandwidth limit. An overshoot is observed after the peak resulting in a dark tail along the pellet's trajectory, as shown in Figure 1(b). This may stem from 1) the $V_{\\mathrm{FE}}$ bandwidth, 2) the approximation error in equation (12), 3) EVS measurement noise, and 4) choice of weighing matrices for CIS and EVS. Figure 8(b), indicates that the proposed method gives an excellent reconstruction of the pel", + "bbox": [ + 75, + 672, + 472, + 902 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e59a31c0c039cf129bfcd2d939a0cbca7716bc9bbbda3418fca6d19558c19005.jpg", + "image_caption": [ + "(a)" + ], + "image_footnote": [], + "bbox": [ + 511, + 94, + 890, + 186 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/0903f66bcfe91b0e67ca6280b8f589197a8c8e9f970927d46cdce5238e0c2d70.jpg", + "image_caption": [ + "(b)", + "time $(\\mu s)$" + ], + "image_footnote": [], + "bbox": [ + 506, + 205, + 888, + 296 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/5f28b27d9ca784fa765a3b8d41656f9cc8c4706ccc763c11eeca66738fd940fb.jpg", + "image_caption": [ + "Figure 8. Pixel-wise reconstruction results for the \"slingshot egg\" scene. The selected pixel is located at the path of the pellet, in the yellow box of Figure 1. The refinement network is not used here. (a) Comparison of the ground truth $V_{\\mathrm{DC}}$ , reconstructed $V_{\\mathrm{FE}}$ and $V_{\\mathrm{DC}}$ , where the ground truth $V_{\\mathrm{DC}}$ is computed from the ground truth DN of the pixel at 10000 FPS. (b) Comparison of reconstructed pixel-wise DN of state-of-the-art video reconstruction methods.", + "(a) with no compensation" + ], + "image_footnote": [], + "bbox": [ + 511, + 455, + 694, + 551 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/9623e737ab247512b849601955966bf3cd145117bc583e92d9b67a0abba37a04.jpg", + "image_caption": [ + "(b) with $V_{\\mathrm{FE}}$ delay compensation" + ], + "image_footnote": [], + "bbox": [ + 694, + 455, + 875, + 551 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/84c86a856507df835cd2d02721b030013398fa71167135634dc9ac8c1753bd57.jpg", + "image_caption": [ + "(c) with RL+RP compensation", + "Figure 9. Qualitative results of ablation study on the \"basketball\" scene. Refer to Figure 1(a) for \"with $V_{\\mathrm{FE}} + \\mathrm{RL} + \\mathrm{RP}$ compensation and refinement\" result." + ], + "image_footnote": [], + "bbox": [ + 511, + 565, + 694, + 660 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/1d06d331807319f189d35f3097b64f7ee960371b235c277a5d356c98d04ffa1c.jpg", + "image_caption": [ + "(d) with $V_{\\mathrm{FE}} + \\mathrm{RL} + \\mathrm{RP}$ compensation" + ], + "image_footnote": [], + "bbox": [ + 694, + 565, + 875, + 660 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "let's peak DN despite the overshoot.", + "bbox": [ + 498, + 746, + 740, + 761 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Ablation Study We ablate the refinement network, $V_{\\mathrm{FE}}$ delay compensation, readout latency (RL) and refractory period (RP) compensation of our method. The numerical evaluation are subject to selected ROIs on motion region. To ablate $V_{\\mathrm{FE}}$ delay compensation, we set the time constant $\\tau = 0$ such that $V_{\\mathrm{FE}} = V_{\\mathrm{DC}}$ in Eq. (7). To ablate readout latency and refractory period compensation, we use Eq. (8) instead of Eq. (12) in the EVS measurement model. Results are summarized in Table 2 and Figure 9. The con", + "bbox": [ + 496, + 763, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "25178", + "bbox": [ + 478, + 922, + 519, + 934 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Figuration using all modules gives significantly better visual and slightly better numerical results in PSNR compared to having latency compensation and refinement disabled. Disabling $V_{\\mathrm{FE}}$ or RL+RP compensation significantly worsens the image quality, as severe shadowing artifacts are observed. As for LPIPS in \"slingshot egg\" scene, the difference is not significant due to the small area of pellet ROI. It is also noted that adding RL+RP compensation leads to improvement in the \"basketball\" but degrades the performance in \"slingshot egg\" when comparing with $V_{\\mathrm{FE}}$ delay compensation only. This is because RL+RP compensation is based on the assumption that the slope of VFE is locally constant for a short period. In cases such as \"slingshot egg\", the assumption may not hold due to close positive and negative events. An ablation study against the use of NAFNet-only is included in supplementary material.", + "bbox": [ + 75, + 90, + 472, + 332 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Performance Regarding Illuminance and Speed The proposed method is further investigated under a controlled environment using varying illuminance and object motion speed. BEW [5] is used to evaluate the image quality by measuring edge sharpness. Figure 10(a)-(h) presents the comparison of the state-of-the-art methods on a rotating disk under $1000\\mathrm{lx}$ and $292\\mathrm{rpm}$ . It is observed that the proposed method reconstructs sharper edges compared to the others. Other EVS-based methods struggle resolving blurry edges in presence of fast motion and also the CIS-based FILM method fails to generate a sharp reconstruction. Figure 10(i) shows that the proposed method reports the lowest BEW score (0.7) and has the sharpest edge profile.", + "bbox": [ + 75, + 335, + 468, + 532 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The BEW scores under different speed and illumination conditions are summarized in Figure 11. Figure 11(a) shows that the BEW increases with faster rotation speeds. In Figure 11(b), the lower ambient light condition leads to higher $V_{\\mathrm{FE}}$ latency degrading image quality. Our method achieves the lowest BEW throughout all illuminance levels.", + "bbox": [ + 75, + 534, + 470, + 625 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 76, + 637, + 194, + 652 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "An EVS-assisted framework to solve motion blur, rolling-shutter artifacts and video-frame interpolation in a joint manner was presented. Existing methods use either a simple double integral model or a learning-based model, neither of which explicitly address EVS sensor imperfections such as EVS pixel latency, readout latency or refractory period. This work overcomes these limitations by explicitly modeling these characteristics. The compensation of readout latency and refractory period is enabled by having access to in-pixel and off-pixel time-stamps. The nonlinear inverse model was solved through joint graph optimization of subsequent frames utilizing row-by-row exposure information to overcome rolling-shutter artifacts and possible glitches between subsequent frames. A refinement network was proposed to further improve the image quality by using specifically designed spatial attention blocks. We showed that", + "bbox": [ + 75, + 657, + 472, + 900 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/ac5f187dbcf96d40eb3e6c02796a0804c55602294a83345f77bf9c0adebac150.jpg", + "image_caption": [ + "(a) Ours w/o refinement" + ], + "image_footnote": [], + "bbox": [ + 506, + 87, + 604, + 157 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/7c824604139f204e8ee97040bd04c00e5bf0fe194137654db705377c0490709d.jpg", + "image_caption": [ + "(b) EDI" + ], + "image_footnote": [], + "bbox": [ + 604, + 88, + 700, + 157 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/0bb38fd3e7a4c604a13814662e2932089c2f12959cde9fbdff3d87f149a3b387.jpg", + "image_caption": [ + "(c)AKF" + ], + "image_footnote": [], + "bbox": [ + 700, + 88, + 795, + 157 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/5fbd68f799e6321bfcf5b89c9e7134912537c696687e899a5325c2d6dcd14380.jpg", + "image_caption": [ + "(d) Time Lens" + ], + "image_footnote": [], + "bbox": [ + 795, + 88, + 893, + 157 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/181c11d8c4ef18733787bb7d61eff62d02de1f62ec20e8ef54cd9a125c775dbc.jpg", + "image_caption": [ + "(e) FILM" + ], + "image_footnote": [], + "bbox": [ + 508, + 169, + 604, + 239 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/8adac60a1e6e2c7288c4b69289e84a65e56b98a53137230ba2e0612898e4638c.jpg", + "image_caption": [ + "(f) CBMNet" + ], + "image_footnote": [], + "bbox": [ + 604, + 169, + 700, + 239 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/50d88950284ba515c9491230e2725d201856e92e35fdb6f1c99eb6254b195442.jpg", + "image_caption": [ + "(g) EvUnroll" + ], + "image_footnote": [], + "bbox": [ + 700, + 169, + 795, + 239 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/d58e086e10466bcdf13e7306801568e1f12c80e269db05ce17912a301ffd7e7c.jpg", + "image_caption": [ + "(h) REFID" + ], + "image_footnote": [], + "bbox": [ + 795, + 169, + 893, + 239 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/15725040120936c38a6ecf6d17f65b6bca40c7c44b7c4d7674a5e04693710be4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 509, + 253, + 887, + 356 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/122c30b05db8c9c5cb059e32e470ec0d4ddfa289390133eac089943b667e7cde.jpg", + "image_caption": [ + "Figure 10. (a)-(h) Comparison of state-of-the-art video reconstruction methods on the proposed rotating disk datasets with rotating speed $292\\mathrm{rpm}$ and environment illuminance $1000\\mathrm{lx}$ . The yellow box shows magnified edges. The visual \"misalignment\" is due to EVS latency. (i) The DN-edge width relation at a fixed segment of 40 pixels (red line in (a)) on reconstructed rotating disk edge. The refinement network is not used.", + "(a) speed (rpm)", + "Figure 11. Quantitative comparison of state-of-the-art video reconstruction methods on the proposed \"rotating disk\" scene in terms of BEW. The lower the BEW values, the sharper the edge. The refinement network is not used. (a) BEW under various rotation speeds with a fixed illuminance of $1000\\mathrm{lx}$ . (b) BEW under various illuminance levels with a fixed speed of $292\\mathrm{rpm}$ ." + ], + "image_footnote": [], + "bbox": [ + 506, + 488, + 702, + 638 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/b906a65537ee2753c868c9c208278975d5f4c076cc05dbed7596616761135e8c.jpg", + "image_caption": [ + "(b) illumination (lux)" + ], + "image_footnote": [], + "bbox": [ + 707, + 491, + 883, + 638 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "the proposed method outperforms state-of-the-art methods in reconstructed image quality with an up to $4\\mathrm{dB}$ improvement in PSNR and $12\\%$ improvement in LPIPS score.", + "bbox": [ + 496, + 777, + 890, + 823 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Limitation Since the pixel latency highly depends on EVS design parameters, object motion, and ambient illuminance, the image quality improvement after latency compensation may be subtle for slow objects or in bright scenes such as \"running man\" and \"spinning umbrella\".", + "bbox": [ + 496, + 824, + 890, + 898 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "25179", + "bbox": [ + 478, + 922, + 517, + 934 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Wenbo Bao, Wei-Sheng Lai, Chao Ma, Xiaoyun Zhang, Zhiyong Gao, and Ming-Hsuan Yang. Depth-aware video frame interpolation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3703-3712, 2019. 2", + "[2] Patrick Bardow, Andrew J Davison, and Stefan Leutenegger. Simultaneous optical flow and intensity estimation from an event camera. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 884-892, 2016. 2", + "[3] Christian Brandli, Raphael Berner, Minhao Yang, Shih-Chii Liu, and Tobi Delbruck. A $240 \\times 180$ 130 db $3\\mu s$ latency global shutter spatiotemporal vision sensor. IEEE Journal of Solid-State Circuits, 49(10):2333-2341, 2014. 1", + "[4] Xiaojie Chu, Liangyu Chen, and Wenqing Yu. Nafssr: Stereo image super-resolution using nafnet. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, pages 1239-1248, 2022. 5", + "[5] Hai Dinh, Qinyi Wang, Fangwen Tu, Brett Frymire, and Bo Mu. Evaluation of motion blur image quality in video frame interpolation. *Electronic Imaging*, 35:262-1, 2023. 5, 8", + "[6] Julius Erbach, Stepan Tulyakov, Patricia Vitoria, Alfredo Bochicchio, and Yuanyou Li. Evshutter: Transforming events for unconstrained rolling shutter correction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13904-13913, 2023. 2", + "[7] C.Posch et al. Sensitivity and uniformity of a $0.18\\mu \\mathrm{m}$ CMOS temporal contrast pixel array. In ISCAS, 2011. 4", + "[8] Xinyu Zhou et al. Evunroll: Neuromorphic events based rolling shutter image correction. In CVPR, 2022. 2, 5", + "[9] Guillermo Gallego, Tobi Delbrück, Garrick Orchard, Chiara Bartolozzi, Brian Taba, Andrea Censi, Stefan Leutenegger, Andrew J Davison, Jörg Conradt, Kostas Daniilidis, et al. Event-based vision: A survey. IEEE transactions on pattern analysis and machine intelligence, 44(1):154-180, 2020. 1", + "[10] Yue Gao, Siqi Li, Yipeng Li, Yandong Guo, and Qionghai Dai. Superfast: $200 \\times$ video frame interpolation via event camera. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2022. 2", + "[11] Leon A Gatys, Alexander S Ecker, and Matthias Bethge. Image style transfer using convolutional neural networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2414-2423, 2016. 5", + "[12] Daniel Gehrig, Antonio Loquercio, Konstantinos G Derpanis, and Davide Scaramuzza. End-to-end learning of representations for asynchronous event-based data. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5633-5643, 2019. 2", + "[13] Giorgio Grisetti, Rainer Kümmerle, Hauke Strasdat, and Kurt Konolige. g2o: A general framework for (hyper) graph optimization. In Proceedings of the IEEE International Conference on Robotics and Automation (ICRA), pages 9-13, 2011. 4", + "[14] Menghan Guo, Shoushun Chen, Zhe Gao, Wenlei Yang, Peter Bartkovjak, Qing Qin, Xiaqin Hu, Dahai Zhou, Qiping Huang, Masayuki Uchiyama, Yoshiharu Kudo, Shimpei" + ], + "bbox": [ + 78, + 114, + 470, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Fukuoka, Chengcheng Xu, Hiroaki Ebihara, Xueqing Wang, Peiwen Jiang, Bo Jiang, Bo Mu, Huan Chen, Jason Yang, T. J. Dai, and Andreas Suess. A three-wafer-stacked hybrid 15-mpixel cis + 1-mpixel evs with 4.6-GeV/s readout, inpixel tdc, and on-chip isp and esp function. IEEE Journal of Solid-State Circuits, pages 1–10, 2023. 1, 2, 5", + "[15] Menghan Guo, Shoushun Chen, Zhe Gao, Wenlei Yang, Peter Bartkovjak, Qing Qin, Xiaqin Hu, Dahei Zhou, Masayuki Uchiyama, Yoshiharu Kudo, Shimpei Fukuoka, Chengcheng Xu, Hiroaki Ebihara, Andy Wang, Peiwen Jiang, Bo Jiang, Bo Mu, Huan Chen, Jason Yang, TJ Dai, and Andreas Suess. A 3-wafer-stacked hybrid 15MPixel CIS + 1MPixel EVS with 4.6GEvent/s readout, in-pixel TDC and on-chip ISP and ESP function. In 2022 IEEE International Solid-State Circuits Conference (ISSCC), 2023. 2, 5", + "[16] Yuhuang Hu, Shih-Chii Liu, and Tobi Delbruck. v2e: From video frames to realistic dvs events. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1312-1321, 2021. 3", + "[17] Zhewei Huang, Tianyuan Zhang, Wen Heng, Boxin Shi, and Shuchang Zhou. Real-time intermediate flow estimation for video frame interpolation. In European Conference on Computer Vision, pages 624-642. Springer, 2022. 2", + "[18] Huaizu Jiang, Deqing Sun, Varun Jampani, Ming-Hsuan Yang, Erik Learned-Miller, and Jan Kautz. Super slomo: High quality estimation of multiple intermediate frames for video interpolation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 9000-9008, 2018. 2", + "[19] Nocedal Jorge and J Wright Stephen. Numerical optimization. Springer, 2006. 4", + "[20] Damien Joubert, Mathieu Hébert, Hubert Konik, and Christophe Lavergne. Characterization setup for event-based imagers applied to modulated light signal detection. Applied optics, 58(6):1305-1317, 2019. 2", + "[21] Damien Joubert, Alexandre Marcireau, Nic Ralph, Andrew Jolley, André van Schaik, and Gregory Cohen. Event camera simulator improvements via characterized parameters. Frontiers in Neuroscience, page 910, 2021. 1, 2", + "[22] Hanme Kim, Stefan Leutenegger, and Andrew J Davison. Real-time 3d reconstruction and 6-dof tracking with an event camera. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part VI 14, pages 349-364. Springer, 2016. 2", + "[23] Taewoo Kim, Yujeong Chae, Hyun-Kurl Jang, and Kuk-Jin Yoon. Event-based video frame interpolation with cross-modal asymmetric bidirectional motion fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 18032-18042, 2023. 1, 2, 5", + "[24] Kazutoshi Kodama, Yusuke Sato, Yuhi Yorikado, Raphael Berner, Kyoji Mizoguchi, Takahiro Miyazaki, Masahiro Tsukamoto, Yoshihisa Matoba, Hirotaka Shinozaki, Atsumi Niwa, et al. 1.22 $\\mu$ m 35.6 mpixel rgb hybrid event-based vision sensor with 4.88 $\\mu$ m-pitch event pixels and up to 10k event frame rate by adaptive control on event sparsity." + ], + "bbox": [ + 501, + 92, + 893, + 901 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "25180", + "bbox": [ + 478, + 922, + 519, + 934 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "In 2023 IEEE International Solid-State Circuits Conference (ISSCC), pages 92-94. IEEE, 2023. 1", + "[25] Patrick Lichtsteiner, Christoph Posch, and Tobi Delbruck. A $128 \\times 128$ 120 db $15\\mu s$ latency asynchronous temporal contrast vision sensor. IEEE Journal of Solid-State Circuits, 43 (2):566-576, 2008. 1, 2", + "[26] Xiaozheng Mou, Kaijun Feng, Alex Yi, Steve Wang, Huan Chen, Xiaoqin Hu, Menghan Guo, Shoushun Chen, and Andreas Suess. Accurate event simulation using high-speed videos. *Electronic Imaging*, 34(7):242-1, 2022. 2, 3, 5", + "[27] Liyuan Pan, Cedric Scheerlinck, Xin Yu, Richard Hartley, Miaomiao Liu, and Yuchao Dai. Bringing a blurry frame alive at high frame-rate with an event camera. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6820-6829, 2019. 1, 2, 5", + "[28] Liyuan Pan, Richard Hartley, Cedric Scheerlinck, Miaomiao Liu, Xin Yu, and Yuchao Dai. High frame rate video reconstruction based on an event camera. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(5):2519-2533, 2020. 2", + "[29] Federico Paredes-Valles and Guido CHE de Croon. Back to event basics: Self-supervised learning of image reconstruction for event cameras via photometric constancy. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3446–3455, 2021. 2", + "[30] Christoph Posch and Daniel Matolin. Sensitivity and uniformity of a $0.18\\mu \\mathrm{m}$ cmos temporal contrast pixel array. In 2011 IEEE International Symposium of Circuits and Systems (ISCAS), pages 1572-1575. IEEE, 2011. 2", + "[31] Henri Rebecq, René Ranftl, Vladlen Koltun, and Davide Scaramuzza. High speed and high dynamic range video with an event camera. IEEE transactions on pattern analysis and machine intelligence, 43(6):1964-1980, 2019. 2", + "[32] Fitsum Reda, Janne Kontkanen, Eric Tabellion, Deqing Sun, Caroline Pantofaru, and Brian Curless. Film: Frame interpolation for large motion. In European Conference on Computer Vision, pages 250-266. Springer, 2022. 1, 2, 5", + "[33] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical Image Computing and Computer-Assisted Intervention-MICCAI 2015: 18th International Conference, Munich, Germany, October 5-9, 2015, Proceedings, Part III 18, pages 234-241. Springer, 2015. 2", + "[34] Cedric Scheerlinck, Nick Barnes, and Robert Mahony. Continuous-time intensity estimation using event cameras. In Asian Conference on Computer Vision, pages 308-324. Springer, 2018. 1, 2", + "[35] Andreas Suess, Menghan Guo, Rui Jiang, Xiaozheng Mou, Qiping Huang, Wenlei Yang, and Shoushun Chen. Physical modeling and parameter extraction for event-based vision sensors. In Proc. IISS Int. Image Sensor Workshop (IISW), Edinburgh, UK, page R5, 2023. 2", + "[36] Lei Sun, Christos Sakaridis, Jingyun Liang, Peng Sun, Jiezhang Cao, Kai Zhang, Qi Jiang, Kaiwei Wang, and Luc Van Gool. Event-based frame interpolation with ad-hoc deblurring. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18043-18052, 2023. 1, 2, 5" + ], + "bbox": [ + 78, + 90, + 468, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[37] Stepan Tulyakov, Daniel Gehrig, Stamatios Georgoulis, Julius Erbach, Mathias Gehrig, Yuanyou Li, and Davide Scaramuzza. Time lens: Event-based video frame interpolation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16155-16164, 2021. 1, 2, 5", + "[38] Stepan Tulyakov, Alfredo Bochicchio, Daniel Gehrig, Stamatios Georgoulis, Yuanyou Li, and Davide Scaramuzza. Time lens: Event-based frame interpolation with parametric non-linear flow and multi-scale fusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17755-17764, 2022. 2", + "[39] Ziwei Wang, Yonhon Ng, Cedric Scheerlinck, and Robert Mahony. An asynchronous kalman filter for hybrid event cameras. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 448-457, 2021. 1, 2, 5", + "[40] Sanghyun Woo, Jongchan Park, Joon-Young Lee, and In So Kweon. Cbam: Convolutional block attention module. In Proceedings of the European conference on computer vision (ECCV), pages 3–19, 2018. 5", + "[41] Zeyu Xiao, Wenming Weng, Yueyi Zhang, and Zhiwei Xiong. Eva2: Event-assisted video frame interpolation via cross-modal alignment and aggregation. IEEE Transactions on Computational Imaging, 8:1145–1158, 2022. 2", + "[42] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 5", + "[43] Xiang Zhang and Lei Yu. Unifying motion deblurring and frame interpolation with events. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17765-17774, 2022. 2", + "[44] Xinyu Zhou, Peiqi Duan, Yi Ma, and Boxin Shi. Evunroll: Neuromorphic events based rolling shutter image correction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17775-17784, 2022. 1, 2" + ], + "bbox": [ + 501, + 92, + 890, + 638 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "25181", + "bbox": [ + 478, + 922, + 517, + 934 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2024/EVS-assisted Joint Deblurring Rolling-Shutter Correction and Video Frame Interpolation through Sensor Inverse Modeling/2babbe27-3090-44dc-8f9d-2d16b50cc5ab_model.json b/2024/EVS-assisted Joint Deblurring Rolling-Shutter Correction and Video Frame Interpolation through Sensor Inverse Modeling/2babbe27-3090-44dc-8f9d-2d16b50cc5ab_model.json new file mode 100644 index 0000000000000000000000000000000000000000..998da81397c50dc60fc2c63522b747f246001907 --- /dev/null +++ b/2024/EVS-assisted Joint Deblurring Rolling-Shutter Correction and Video Frame Interpolation through Sensor Inverse Modeling/2babbe27-3090-44dc-8f9d-2d16b50cc5ab_model.json @@ -0,0 +1,2310 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.0, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.102, + 0.13, + 0.872, + 0.177 + ], + "angle": 0, + "content": "EVS-assisted joint Deblurring, Rolling-Shutter Correction and Video Frame Interpolation through Sensor Inverse Modeling" + }, + { + "type": "text", + "bbox": [ + 0.139, + 0.204, + 0.846, + 0.24 + ], + "angle": 0, + "content": "Rui Jiang*, Fangwen Tu*, Yixuan Long, Aabhaas Vaish, Bowen Zhou, Qinyi Wang Wei Zhang, Yuntan Fang, Luis Eduardo Garcia Capel, Bo Mu, Tiejun Dai, Andreas Suess" + }, + { + "type": "text", + "bbox": [ + 0.434, + 0.246, + 0.536, + 0.26 + ], + "angle": 0, + "content": "OMNIVISION" + }, + { + "type": "text", + "bbox": [ + 0.425, + 0.264, + 0.545, + 0.278 + ], + "angle": 0, + "content": "*Equal contribution" + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.314, + 0.891, + 0.409 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.419, + 0.894, + 0.476 + ], + "angle": 0, + "content": "Figure 1. Qualitative results comparing the proposed method with EDI [27], AKF [39], Time Lens [37], FILM [32], CBMNet [23], EvUnroll [44], and REFID [36] on simulation datasets. In the \"basketball\" scene, our method provides frames that are rolling shutter effect corrected and deblurred. In the \"slingshot egg\" scene, our method can reconstruct the shape of the pellet with the best quality, as indicated in the yellow box. The images have been magnified for optimal viewing quality." + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.488, + 0.314, + 0.503 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.52, + 0.474, + 0.837 + ], + "angle": 0, + "content": "Event-based Vision Sensors (EVS) gain popularity in enhancing CMOS Image Sensor (CIS) video capture. Nonidealities of EVS such as pixel or readout latency can significantly influence the quality of the enhanced images and warrant dedicated consideration in the design of fusion algorithms. A novel approach for jointly computing deblurred, rolling-shutter artifact corrected high-speed videos with frame rates up to 10 000 FPS using inherently blurry rolling shutter CIS frames of 120 FPS to 150 FPS in conjunction with EVS data from a hybrid CIS-EVS sensor is presented. EVS pixel latency, readout latency and the sensor's refractory period are explicitly incorporated into the measurement model. This inverse function problem is solved on a per-pixel manner using an optimization-based framework. The interpolated images are subsequently processed by a novel refinement network. The proposed method is evaluated using simulated and measured datasets, under natural and controlled environments. Extensive experiments show reduced shadowing effect, a 4 dB increment in PSNR, and a \\(12\\%\\) improvement in LPIPS score compared to state-of-the-art methods." + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.488, + 0.631, + 0.503 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.509, + 0.893, + 0.72 + ], + "angle": 0, + "content": "Event-Based Vision Sensors (EVS) capture our world from a completely different perspective compared to CMOS Image Sensors (CIS) [9, 14, 24, 25]. In EVS, each pixel independently triggers an event if a relative illuminance change on the pixel reaches a certain triggering threshold (typically in the range of \\(10\\%\\) to \\(30\\%\\)). Conversely to CIS, EVS operate time-continuously and asynchronously - there is no common exposure period or \"frame rate\". Thus, EVS pixels can be read out comparatively fast in a low-power manner enabling efficient high-speed data capture. Recent work on enhancement of high-quality images [23, 27, 34, 37, 39] focuses on fusing CIS and EVS information so that the high fidelity of CIS measurement and the fast response of EVS measurement can complement each other." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.721, + 0.895, + 0.858 + ], + "angle": 0, + "content": "In practice it is observed [3, 14, 21, 25] that the dynamic characteristics of EVS pixels are highly dependent on the illumininance in the sensor plane. Under low-light conditions such as indoor capture, a slower EVS response during video enhancement can result in ghosting or blurry frames and needs to be considered in algorithm design. Furthermore, the inevitable readout latency and refractory period of EVS pixels can significantly aggravate artifacts. Current methods are not specifically designed to address sensor nonidealities." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.923, + 0.521, + 0.935 + ], + "angle": 0, + "content": "25172" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.302 + ], + "angle": 0, + "content": "This paper presents an inverse model capturing the EVS pixel latency (namely the \\( V_{\\mathrm{FE}} \\) delay), the readout latency, and the refractory period. The inverse model is solved through nonlinear graph optimization to jointly address deblurring, rolling-shutter (RS) correction and video-frame interpolation while compensating for sensor nonidealities. In addition to the inverse model, a learning-based refinement module is proposed to enhance the image quality by further mitigating the noise and artifacts. The proposed method is compared with existing CIS-based [32] and EVS-assisted [8, 23, 27, 36, 37, 39] video enhancement methods under challenging indoor scenarios with fast motion. As illustrated in Figure 1, the proposed algorithm produces excellent image quality with much fewer artifacts." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.304, + 0.451, + 0.318 + ], + "angle": 0, + "content": "The contributions of the paper can be summarized as:" + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.32, + 0.468, + 0.379 + ], + "angle": 0, + "content": "- A formulation of a joint deblurring, rolling-shutter correction and video-frame interpolation problem based on CIS and EVS fusion, where EVS sensor nonidealities are explicitly modeled (Section 3)," + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.38, + 0.468, + 0.425 + ], + "angle": 0, + "content": "- An optimization methodology jointly solving this coupled inverse problem (Section 4.1), as well as a post-processing network for artifact removal (Section 4.2)," + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.426, + 0.468, + 0.469 + ], + "angle": 0, + "content": "- A comprehensive comparison and analysis of simulation and measured datasets in natural and controlled environments (Section 5)." + }, + { + "type": "list", + "bbox": [ + 0.096, + 0.32, + 0.468, + 0.469 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.482, + 0.22, + 0.497 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.505, + 0.473, + 0.746 + ], + "angle": 0, + "content": "Video Frame Interpolation Video interpolation aims to synthesize intermediate frames between two images to achieve a temporally coherent video sequence. Traditional CIS-based methods depend on optical flow for a smooth transition. In SuperSlomo [18], the bi-directional flows are combined with a linear combination followed by a refinement using U-Net [33]. Such optical flow-based methods, however, suffer from occlusion problems. In DAIN [1], researchers estimate occlusion areas using depth information and use contextual information from neighboring pixels to fill the occluded regions. In RIFE [17], the flow and fusion maps are generated simultaneously without an additional optical flow module for final result synthesis. FILM [32] presents a multi-scale feature extractor within a unified network to achieve an enlarged receptive field aimed at handling large object motion in the video." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.473, + 0.902 + ], + "angle": 0, + "content": "Event-assisted Frame Enhancement A key issue lowering the image quality of CIS-based frame interpolation methods is the lack of information between frames. EVS provides such information thanks to its asynchronous low latency capture not relying on a global exposure period. Some early studies [31, 34] show that intensity reconstruction purely based on EVS modality is feasible. However, as EVS captures events based on differential changes of illuminance, this approach struggles to estimate the absolute light level. In [2], the optical flow and intensity frames are" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.424 + ], + "angle": 0, + "content": "estimated simultaneously by solving an optimization problem that takes EVS measurements as data terms. The authors in [22] take a step further by formulating the event-based SLAM problem, which can be decomposed into camera ego-motion estimation and 3-D scene reconstruction. However, the quantization error makes it difficult to reconstruct high-quality frames from EVS alone. As EVS pixels trigger independently, researchers simplify the problem formulation to pixel-wise estimation by fusing CIS and EVS, without seeking explicit representation of the 3-D environment. The Event-based Double Integral (EDI) [27, 28] method is proposed as a straightforward and effective way to deblur CIS frames and generate high frame rate videos. EVS and CIS measurements are connected by integrating events twice, such that the continuous intensity curve for each pixel can be determined. [34] presents a complementary filter-based framework to combine a high-pass signal from EVS and a low-pass signal from CIS to compute an all-pass signal for pixel intensity. To make the complementary filter gain adaptive to measurement noise, a pixel-wise Kalman filter is used in [39] where a unifying EVS/CIS uncertainty model is proposed." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.426, + 0.892, + 0.622 + ], + "angle": 0, + "content": "Machine learning methods such as Time Lens [37] and its subsequent versions [10, 36, 38, 41] or others [29, 43] concatenate both CIS and EVS data using a voxel grid representation [12] as input to feed the network of two branches: an optical flow warping-based branch and a synthesis-based branch. The final result is blended with a trained attention module to avoid the disadvantages of any individual branch. CBMNet [23] is proposed to handle complex real-world motion using a novel cross-modal asymmetric bidirectional motion field estimation. Due to the high temporal resolution of EVS data, the rolling shutter problem which is commonly observed in fast-motion scenarios can be corrected using EVS data [6, 44]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.634, + 0.892, + 0.725 + ], + "angle": 0, + "content": "EVS Latency Many studies model EVS latency [20, 25, 30, 35], but few of them have been applied to algorithm design such as frame enhancement. To the best of the authors' knowledge, this is the first work that considers an EVS latency model explicitly in generating high frame rate, deblurred, rolling-shutter artifact-free videos." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.741, + 0.707, + 0.757 + ], + "angle": 0, + "content": "3. Problem Formulation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Event Pixel Operation and Pixel Latency Model In this work, a hybrid CIS-EVS sensor [14, 15] is used. As shown in Figure 2(a), EVS and CIS share the same photodiode design on the hybrid sensor and it is assumed that the light level is identical for nearby CIS and EVS pixels. The photocurrent \\( i_{\\mathrm{pd}} \\) is converted to front-end voltage \\( V_{\\mathrm{FE}} \\) in module \\( \\mathrm{i}2\\mathrm{v} \\), which models dynamic characteristics as a 1st-order Low-Pass Filter (LPF) [21, 26]. The update rule for \\( V_{\\mathrm{FE}} \\) between discrete event indices \\( k \\) and \\( k + 1 \\) can be written as:" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.923, + 0.521, + 0.935 + ], + "angle": 0, + "content": "25173" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.084, + 0.091, + 0.478, + 0.158 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.084, + 0.17, + 0.47, + 0.306 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.322, + 0.47, + 0.392 + ], + "angle": 0, + "content": "Figure 2. (a) Hybrid CIS-EVS sensor principle. The upper and lower branches indicate physical quantity conversion in CIS pixels and EVS pixels, respectively. (b) Example of \\( V_{\\mathrm{DC}} \\) and \\( V_{\\mathrm{FE}} \\) curves and triggered events that illustrate the impact of \\( V_{\\mathrm{FE}} \\) delay, readout latency, and refractory period." + }, + { + "type": "equation", + "bbox": [ + 0.097, + 0.43, + 0.469, + 0.446 + ], + "angle": 0, + "content": "\\[\nV _ {\\mathrm {F E}} (k + 1) = V _ {\\mathrm {F E}} (k) + \\gamma (k) \\left(V _ {\\mathrm {D C}} (k + 1) - V _ {\\mathrm {F E}} (k)\\right), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.449, + 0.469, + 0.507 + ], + "angle": 0, + "content": "where \\( V_{\\mathrm{DC}}(k + 1) = f_{\\mathrm{DC}}(i_{\\mathrm{pd}}(k + 1)) \\) represents the \\( V_{\\mathrm{FE}} \\) under DC excitation (i.e., the \\( V_{\\mathrm{FE}} \\) with infinite LPF bandwidth); \\( f_{\\mathrm{DC}}(i_{\\mathrm{pd}}) \\) is the logarithmic-like current-voltage conversion function and" + }, + { + "type": "equation", + "bbox": [ + 0.18, + 0.51, + 0.469, + 0.542 + ], + "angle": 0, + "content": "\\[\n\\gamma (k) = 1 - \\exp \\left(- \\frac {\\Delta t}{\\tau (k)}\\right) \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.544, + 0.469, + 0.588 + ], + "angle": 0, + "content": "is the time-varying coefficient of the LPF, where \\(\\Delta t\\) is the time interval between two indices. The photocurrent-dependent time constant \\(\\tau (k)\\) can be modeled as [16, 26]:" + }, + { + "type": "equation", + "bbox": [ + 0.186, + 0.591, + 0.469, + 0.62 + ], + "angle": 0, + "content": "\\[\n\\tau (k) = \\tau_ {0} + \\frac {\\alpha}{i _ {\\mathrm {p d}} (k)}, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.622, + 0.469, + 0.697 + ], + "angle": 0, + "content": "where \\(\\tau_0\\) and \\(\\alpha\\) are pixel parameters. The pixel latency in module \\(\\pm 2\\mathrm{v}\\) due to the LPF behavior is called \" \\(V_{\\mathrm{FE}}\\) delay\". As illustrated in Figure 2(b), an event is triggered at time \\(t_\\mathrm{in}\\) once the difference between \\(V_{\\mathrm{FE}}\\) and reference voltage \\(V_{\\mathrm{ref}}\\) reaches the triggering threshold voltage:" + }, + { + "type": "equation", + "bbox": [ + 0.08, + 0.7, + 0.469, + 0.732 + ], + "angle": 0, + "content": "\\[\n\\left\\{ \\begin{array}{l l} V _ {\\mathrm {F E}} \\left(t _ {\\mathrm {i n}}\\right) - V _ {\\mathrm {r e f}} \\left(t _ {\\mathrm {i n}}\\right) & \\geq c _ {\\mathrm {u p}} \\quad \\text {f o r p o s i t i v e e v e n t s} \\\\ V _ {\\mathrm {F E}} \\left(t _ {\\mathrm {i n}}\\right) - V _ {\\mathrm {r e f}} \\left(t _ {\\mathrm {i n}}\\right) & \\leq c _ {\\mathrm {d o w n}} \\quad \\text {f o r n e g a t i v e e v e n t s} \\end{array} , \\right. \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.734, + 0.469, + 0.839 + ], + "angle": 0, + "content": "where \\( c_{\\mathrm{up / down}} \\) denotes the contrast threshold voltage. The in-pixel timestamp \\( t_{\\mathrm{in}} \\) is recorded when the event fires, after which the pixel experiences \"readout latency \\( t_{\\mathrm{rl}} \\)\" and \"refractory period \\( t_{\\mathrm{rp}} \\)\". The \\( V_{\\mathrm{ref}} \\) is updated to follow \\( V_{\\mathrm{FE}} \\) after the refractory period. Finally, we have a 5-dimensional output \\( \\{x,y,t_{\\mathrm{in}},t_{\\mathrm{off}},p\\} \\), which contains row and column indices, in-pixel and off-pixel timestamps, and polarity." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Pixel-wise Photocurrent Estimation Suppose that the illuminance on the sensor plane results in photocurrent \\( i_{\\mathrm{pd}} \\) for a single pixel. We aim to estimate \\( i_{\\mathrm{pd}} \\) when an event is triggered to avoid the impact of pixel latency in the \\( \\pm 2\\mathrm{v} \\)" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.892, + 0.121 + ], + "angle": 0, + "content": "module. Based on Figure 2(a), we have the following CIS measurement equation:" + }, + { + "type": "equation", + "bbox": [ + 0.601, + 0.124, + 0.892, + 0.159 + ], + "angle": 0, + "content": "\\[\nD N = G \\int_ {t _ {\\mathrm {s}}} ^ {t _ {\\mathrm {e}}} i _ {\\mathrm {p d}} (t) \\mathrm {d} t, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.16, + 0.892, + 0.266 + ], + "angle": 0, + "content": "where \\(DN\\) is the CIS intensity frame measured in digital numbers; \\(G\\) indicates the total gain of charge-voltage conversion and ADC; \\(t_{\\mathrm{s}}\\) and \\(t_{\\mathrm{e}}\\) denote the CIS exposure start and end, respectively. As for EVS measurements, suppose we have an event sequence with \\(N\\) events \\(\\{x(k),y(k),t_{\\mathrm{in}}(k),t_{\\mathrm{off}}(k),p(k)\\}\\). By assuming \\(V_{\\mathrm{FE}}(1) = V_{\\mathrm{DC}}(1)\\), with Eq. (1) we have the recurrence relation:" + }, + { + "type": "equation", + "bbox": [ + 0.527, + 0.27, + 0.891, + 0.286 + ], + "angle": 0, + "content": "\\[\nV _ {\\mathrm {F E}} (1) = f _ {\\mathrm {D C}} \\left(i _ {\\mathrm {p d}} (1)\\right) \\tag {6}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.499, + 0.288, + 0.891, + 0.305 + ], + "angle": 0, + "content": "\\[\nV _ {\\mathrm {F E}} (k + 1) = f _ {\\gamma} \\left(i _ {\\mathrm {p d}} (k)\\right) \\left(f _ {\\mathrm {D C}} \\left(i _ {\\mathrm {p d}} (k + 1)\\right) - V _ {\\mathrm {F E}} (k)\\right), \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.307, + 0.891, + 0.366 + ], + "angle": 0, + "content": "where \\(k \\in \\{1, \\dots, N-1\\}\\) and \\(f_{\\gamma}(\\cdot)\\) denotes the function that maps \\(i_{\\mathrm{pd}}\\) onto \\(\\gamma\\). By assuming \\(V_{\\mathrm{ref}}(k+1) \\approx V_{\\mathrm{FE}}(k)\\), the EVS measurements provide additional \\(N-1\\) equations as:" + }, + { + "type": "equation", + "bbox": [ + 0.566, + 0.372, + 0.891, + 0.387 + ], + "angle": 0, + "content": "\\[\nc (k + 1) \\approx V _ {\\mathrm {F E}} (k + 1) - V _ {\\mathrm {F E}} (k), \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.39, + 0.892, + 0.587 + ], + "angle": 0, + "content": "where \\( c(k) \\in \\{c_{\\mathrm{up}}, c_{\\mathrm{down}}\\} \\) are the known triggering thresholds that are configurable. These state transition equations and measurement equations (6)-(8) can be expressed as a system of \\( 2N - 1 \\) nonlinear equations and \\( 2N \\) unknowns \\( -i_{\\mathrm{pd}}(k) \\) and \\( V_{\\mathrm{FE}}(k) \\). By using a zero-order holder, the discrete states \\( i_{\\mathrm{pd}}(k) \\) and \\( V_{\\mathrm{FE}}(k) \\) can be approximated as continuous signals \\( i_{\\mathrm{pd}}(t) \\) and \\( V_{\\mathrm{FE}}(t) \\). This makes it possible to utilize CIS measurements in Eq. (5) as additional constraints. In conclusion, the pixel latency compensation can be expressed as pixel-wise optimization problem, solving the nonlinear system with \\( 2N \\) unknown states and \\( 2N + M - 1 \\) equations, where \\( M \\) and \\( N \\) denote the number of CIS and EVS measurements, respectively." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.612, + 0.892, + 0.732 + ], + "angle": 0, + "content": "Modeling Readout Latency and Refractory Period The EVS measurement model equation (8) is an approximation since the state variation during the readout latency and the refractory period is not considered. Using both in-pixel timestamps and off-pixel timestamps, it is possible to compensate for some of the \"missed\" \\( V_{\\mathrm{FE}} \\) change by assuming that the slope of \\( V_{\\mathrm{FE}} \\) is locally constant for a short period. As illustrated in Figure 2(b), it is noted that:" + }, + { + "type": "equation", + "bbox": [ + 0.548, + 0.735, + 0.891, + 0.768 + ], + "angle": 0, + "content": "\\[\nm (k + 1) = \\frac {c (k + 1)}{t _ {\\mathrm {i n}} (k + 1) - t _ {\\mathrm {o f f}} (k) - t _ {\\mathrm {r p}}}, \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.77, + 0.891, + 0.813 + ], + "angle": 0, + "content": "where \\(m\\) describes the slope approximation of the \\(V_{\\mathrm{FE}}\\) curve. For the readout latency, the missed \\(V_{\\mathrm{FE}}\\) is represented as:" + }, + { + "type": "equation", + "bbox": [ + 0.546, + 0.817, + 0.891, + 0.835 + ], + "angle": 0, + "content": "\\[\n\\Delta V _ {\\mathrm {F E}} ^ {\\mathrm {r l}} (k) = m (k + 1) \\left[ t _ {\\text {o f f}} (k) - t _ {\\text {i n}} (k) \\right]. \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.837, + 0.891, + 0.882 + ], + "angle": 0, + "content": "For the refractory period, we have \\(\\Delta V_{\\mathrm{FE}}^{\\mathrm{rp}}(k) = m(k + 1)t_{\\mathrm{rp}}\\) since \\(t_\\mathrm{rp}\\) is a fixed parameter. Thus the reference voltage after compensation can be adjusted to:" + }, + { + "type": "equation", + "bbox": [ + 0.526, + 0.885, + 0.891, + 0.903 + ], + "angle": 0, + "content": "\\[\nV _ {\\text {r e f}} (k + 1) = V _ {\\mathrm {F E}} (k) + \\Delta V _ {\\mathrm {F E}} ^ {\\mathrm {r l}} (k) + \\Delta V _ {\\mathrm {F E}} ^ {\\mathrm {r p}} (k). \\tag {11}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.923, + 0.519, + 0.935 + ], + "angle": 0, + "content": "25174" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.081, + 0.089, + 0.48, + 0.266 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.281, + 0.471, + 0.337 + ], + "angle": 0, + "content": "Figure 3. Block diagram of the proposed method. Note that after the frame reconstruction the blurry texture due to motion is enhanced and the RS effect is removed. The noise and artifacts are further reduced by the refinement network." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.363, + 0.47, + 0.408 + ], + "angle": 0, + "content": "By substituting the compensation equation (11) into (4), the impact of readout latency and refractory period can be compensated:" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.411, + 0.47, + 0.493 + ], + "angle": 0, + "content": "\\(c(k + 1)\\approx V_{\\mathrm{FE}}(k + 1) - V_{\\mathrm{FE}}(k) - \\Delta V_{\\mathrm{FE}}^{\\mathrm{sl}}(k) - \\Delta V_{\\mathrm{FE}}^{\\mathrm{rp}}(k),\\) (12) where \\(\\Delta V_{\\mathrm{FE}}^{\\mathrm{rl}}(k)\\) and \\(\\Delta V_{\\mathrm{FE}}^{\\mathrm{rp}}(k)\\) are voltage compensation for readout latency and refractory period, respectively. Note that this compensation is enabled by the availability of in-pixel and off-pixel timestamps." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.501, + 0.169, + 0.516 + ], + "angle": 0, + "content": "4. Method" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.523, + 0.47, + 0.6 + ], + "angle": 0, + "content": "The proposed frame enhancement framework is presented in Figure 3. The frame reconstruction module jointly deblurs images, removes the RS effect, and interpolates high frame rate images. The refinement network focuses on removing noise and improving image quality." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.607, + 0.449, + 0.625 + ], + "angle": 0, + "content": "4.1. Solving the Inverse Model via Optimization" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.631, + 0.47, + 0.811 + ], + "angle": 0, + "content": "Graph Modeling for a Single Pixel Due to the presence of pixel latency, the system is governed by nonlinear equations, rendering the models and solutions designed for linear systems inapplicable. Graph optimization [13] provides a straightforward way to model nonlinear systems, where the states and measurements are represented as nodes and edges in a graph as shown in Figure 4(a). Let \\( l \\in \\{1, \\dots, M\\} \\) and \\( k \\in \\{1, \\dots, N-1\\} \\) be the indices of CIS and EVS, respectively. By defining states \\( \\boldsymbol{x} = [i_{\\mathrm{pd}}(k)]^{\\top} \\) and measurements \\( z_{\\mathrm{CIS}} = [DN(l)]^{\\top} \\), \\( z_{\\mathrm{EVS}} = [c(k+1)]^{\\top} \\), the pixel-wise graph optimization framework aims to determine the state vector \\( \\boldsymbol{x} \\):" + }, + { + "type": "equation", + "bbox": [ + 0.1, + 0.813, + 0.47, + 0.841 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {x} ^ {*} = \\underset {\\boldsymbol {x}} {\\operatorname {a r g m i n}} \\left(\\boldsymbol {e} _ {\\mathrm {C I S}} ^ {\\top} \\boldsymbol {\\Omega} _ {\\mathrm {C I S}} \\boldsymbol {e} _ {\\mathrm {C I S}} + \\boldsymbol {e} _ {\\mathrm {E V S}} ^ {\\top} \\boldsymbol {\\Omega} _ {\\mathrm {E V S}} \\boldsymbol {e} _ {\\mathrm {E V S}}\\right) \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.841, + 0.471, + 0.902 + ], + "angle": 0, + "content": "where \\( e_{\\mathrm{CIS}} = z_{\\mathrm{CIS}} - \\tilde{z}_{\\mathrm{CIS}} \\), \\( e_{\\mathrm{EVS}} = z_{\\mathrm{EVS}} - \\tilde{z}_{\\mathrm{EVS}} \\) denote measurement error vectors; \\( \\tilde{z}_{\\mathrm{CIS}} \\) and \\( \\tilde{z}_{\\mathrm{EVS}} \\) are predicted measurements which can be calculated from estimated states according to CIS measurement equation (5) and EVS mea" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.091, + 0.897, + 0.212 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.513, + 0.224, + 0.898, + 0.375 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.392, + 0.895, + 0.546 + ], + "angle": 0, + "content": "Figure 4. (a) Graph modeling of the pixel-wise photocurrent estimation problem, where nodes and edges denote states and measurements. The arrows indicate the order of state calculation. \"zCIS\" and \"zEVS\" indicate CIS and EVS measurements, respectively. Eqn. (8) is replaced with Eqn. (12) when voltage compensation for readout latency and refractory period is enabled. (b) Row-based reconstruction scheme showing three CIS images as input. After row-wise computation, interpolated and deblurred frames without RS effect can be generated as shown in the blue box. The proposed scheme can be applied to video enhancement with any number of CIS images." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.566, + 0.893, + 0.763 + ], + "angle": 0, + "content": "surement equation (8); \\( z_{\\mathrm{EVS}} \\) is adjusted using register settings and calibrated using the \"S-curve method\" [7]. The statistical distribution of the threshold is not explicitly modeled in the current method. The diagonal weighing matrices \\( \\Omega_{\\mathrm{CIS}} \\) and \\( \\Omega_{\\mathrm{EVS}} \\) reflect the reliability of CIS and EVS, respectively. Using the cost function equation (13), the \\( V_{\\mathrm{FE}} \\) delay is modeled explicitly through the EVS measurement equations. Solving the optimization problem leads to state estimation such that the \\( V_{\\mathrm{FE}} \\) delay is compensated. Many approaches have been proposed to solve nonlinear optimization problems iteratively [19]. We parallelize pixelwise tasks as there is no data interaction or computational dependency between pixels." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Rolling Shutter (RS) Effect Correction As CIS parameters such as row exposure time and RS scanning speed are known, we formulate the frame estimation problem as pixel-wise computation using adjusted CIS timing parameters for each row. After solving the photocurrent at each event time-step through optimization, a zero-order hold is used to generate a time-continuous photocurrent \\( i_{\\mathrm{pd}}(t) \\) for all pixels. The CIS measurement equation (5) is used to generate a rolling shutter artifact-free target frame rate (such" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.923, + 0.521, + 0.935 + ], + "angle": 0, + "content": "25175" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.116, + 0.093, + 0.437, + 0.26 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.102, + 0.269, + 0.444, + 0.283 + ], + "angle": 0, + "content": "Figure 5. Modified NAF block in the refinement network." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.306, + 0.188, + 0.32 + ], + "angle": 0, + "content": "as 10 000 FPS)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.33, + 0.471, + 0.466 + ], + "angle": 0, + "content": "Video Glitch Effect Deblurring and video frame interpolation performance can experience glitches in case these are computed on a frame-by-frame basis. To mitigate this issue, we solve the coupled deblurring, rolling-shutter correction and interpolation problem jointly for subsequent frames utilizing the row-specific rolling shutter timing characteristics as shown in Figure 4(b), where we reconstruct each row according to its specific exposure time by fusing two consecutive CIS frames and their corresponding events." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.48, + 0.273, + 0.495 + ], + "angle": 0, + "content": "4.2. Refinement Network" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.505, + 0.471, + 0.882 + ], + "angle": 0, + "content": "A refinement network is developed as a post-processing module to further mitigate the noise and artifacts in the reconstructed frames. NAFNet [4] is designed for image restoration. It is structured as a U-Net, arranging NAF blocks to allow the network to progressively downsample the input for semantic feature extraction and then progressively upsample it to restore the output. NAFNet provides additional skip connections that preserve information and allow feature reusability from earlier layers. NAF blocks are structured as original ResNet blocks, with additional layers such as layer normalization and channel attention. We adopt the NAFNet basic concept, and further modify making two major additions: 1) spatial attention [40] as shown in Figure 5 and 2) perceptual loss [32]. The reason for choosing spatial attention is that most of the noise is located in motion areas instead of being ubiquitous across the image. Spatial attention allows the network to focus on these noisy regions of the image, which contributes to better denoising results. Moreover, the rationale for choosing the perceptual loss is that even though the proposed network does well in removing the pixel-level saturation noise, it also introduces blurriness or texture degradation. Perceptual loss helps to recover the image quality in a way that better aligns with the human vision system. The losses applied in this paper are summarized as:" + }, + { + "type": "equation", + "bbox": [ + 0.137, + 0.886, + 0.469, + 0.902 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {s}} = w _ {l} \\mathcal {L} _ {1} + w _ {\\mathrm {V G G}} \\mathcal {L} _ {\\mathrm {V G G}} + w _ {\\text {G r a m}} \\mathcal {L} _ {\\text {G r a m}}, \\tag {14}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.895, + 0.32 + ], + "angle": 0, + "content": "where the \\(\\mathcal{L}_1\\) loss is the pixel-wise difference between the reconstructed RGB image and ground truth image; the \\(\\mathcal{L}_{\\mathrm{VGG}}\\) loss is calculated as \\(\\mathcal{L}_{\\mathrm{VGG}} = \\frac{1}{L}\\sum_{l=1}^{L}\\alpha_l\\|\\Psi_l(\\boldsymbol{I}_{\\mathrm{Recon}}) - \\Psi_l(\\boldsymbol{I}_{\\mathrm{GT}})\\|_1;\\Psi_l(\\boldsymbol{I})\\) is the feature tensor from the \\(l_{\\mathrm{th}}\\) layer of an ImageNet pre-trained VGG-19 network generated by passing the image \\(\\boldsymbol{I}\\) through the network. In this case, \\(L\\) specifies the total number of layers in the VGG-19 network, while \\(\\alpha_l\\) is the hyperparameter assigned to each layer. Finally, the Gram loss [11] is computed as the L2 difference between the autocorrelation of the VGG-19 features of the reconstructed image and ground truth image with the formulation: \\(\\mathcal{L}_{\\mathrm{Gram}} = \\frac{1}{L}\\sum_{l=1}^{L}\\alpha_l\\|\\boldsymbol{M}_l(\\boldsymbol{I}_{\\mathrm{Recon}}) - \\boldsymbol{M}_l(\\boldsymbol{I}_{\\mathrm{GT}})\\|_2\\), where given an image \\(I\\), the Gram matrix is computed for every layer \\(l\\) as \\(\\boldsymbol{M}_l = (\\Psi_l(\\boldsymbol{I}))^\\top(\\Psi_l(\\boldsymbol{I}))\\)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.33, + 0.634, + 0.347 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.353, + 0.895, + 0.669 + ], + "angle": 0, + "content": "We compare our proposed method with state-of-the-art EVS+CIS video reconstruction methods: EDI [27], AKF [39], Time Lens [37], CBMNet [23], EvUnroll [8] and RE-FID [36]. Among them, EvUnroll proposes an RS correction based reconstruction and REFID combines image de-burring with EVS based frame interpolation. We also show the results of the CIS-only interpolation method FILM [32] as reference. Before optimization, \\( i_{\\mathrm{pd}} \\) are initialized according to the first CIS image DN and the inverse photocurrent-DN relation. The elements in the weighing matrices for CIS, EVS terms are set to 1. The performance of the interpolation methods are evaluated both quantitatively and qualitatively, based on the metrics of Peak Signal-to-Noise Ratio (PSNR), LPIPS [42], and pixel-wise reconstruction error on DN. Furthermore, to study the performance degradation concerning ambient illuminance and motion speed, we measured a test set utilizing a rotating disk in a controlled environment. We analyze the Blurred Edge Width (BEW) [5] of the reconstructed frames - a larger BEW indicates a blurrier image and, therefore, a worse deblurring result." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.681, + 0.602, + 0.695 + ], + "angle": 0, + "content": "5.1. Datasets" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.705, + 0.894, + 0.901 + ], + "angle": 0, + "content": "We evaluate the proposed method using the simulation data based on the CIS-EVS hybrid sensor simulator [26] which considers \\( V_{\\mathrm{FE}} \\) delay, readout latency, and refractory period. Our measurement data was collected from the hybrid CIS-EVS sensor [14, 15]. Here, only off-pixel times- tamp data was captured. Two measured datasets were created for evaluation: The natural scene dataset focuses on indoor scenarios where EVS is significantly influenced by pixel latency. The other dataset captured a rotating disk with a Siemens star pattern under different ambient illumination levels and rotating speeds. In addition to the above datasets specifically captured for this work's evaluation, we also test our method using the publicly available HS-ERGB" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.923, + 0.521, + 0.935 + ], + "angle": 0, + "content": "25176" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.09, + 0.895, + 0.134 + ], + "angle": 0, + "content": "Table 1. Quantitative comparison of state-of-the-art video reconstruction methods on the proposed simulation dataset in terms of PSNR (in dB, higher is better) and LPIPS (a dimensionless quantity, lower is better). Each row shows results for a particular scene. Results from different methods are listed column-wise. The first and second places are highlighted with bold underline and bold, respectively." + }, + { + "type": "table", + "bbox": [ + 0.079, + 0.142, + 0.894, + 0.228 + ], + "angle": 0, + "content": "
Ours w/o refinementOursEDIAKFTime LensFILMCBMNetEvUnrollREFID
ScenePSNR ↑LPIPS ↓PSNR ↑LPIPS ↓PSNR ↑LPIPS ↓PSNR ↑LPIPS ↓PSNR ↑LPIPS ↓PSNR ↑LPIPS ↓PSNR ↑LPIPS ↓PSNR ↑LPIPS ↓PSNR ↑LPIPS ↓
basketball24.010.21727.230.15522.490.22015.630.28023.650.24023.860.22024.510.25223.130.21822.380.223
checkerboard24.230.20328.240.18720.480.21819.560.26122.600.23923.020.21622.760.21223.840.20421.230.235
slingshot egg25.230.10032.850.10624.950.16415.590.11025.190.24424.040.20322.760.23424.610.21023.920.201
running man24.600.16424.720.18922.290.13520.710.18925.010.11725.060.08925.160.06923.040.11723.520.163
fan22.400.18924.500.16318.170.25510.780.18420.640.23318.930.18421.280.20419.640.19018.300.198
Average24.090.17527.510.1621.680.19816.450.20523.420.21522.980.18223.290.19422.850.18821.870.204
" + }, + { + "type": "table_caption", + "bbox": [ + 0.088, + 0.241, + 0.88, + 0.256 + ], + "angle": 0, + "content": "Table 2. Quantitative results of ablation study. The first and second places are highlighted with bold underline and bold, respectively." + }, + { + "type": "table", + "bbox": [ + 0.116, + 0.266, + 0.856, + 0.336 + ], + "angle": 0, + "content": "
without compensationwith VFE delay compensationwith RL+RP compensationwith VFE+RL+RP compensationwith VFE+RL+RP compensation and refinement
ScenePSNR ↑LPIPS ↓PSNR ↑LPIPS ↓PSNR ↑LPIPS ↓PSNR ↑LPIPS ↓PSNR ↑LPIPS ↓
basketball22.950.31623.470.30623.500.30723.560.30223.640.295
slingshot egg23.570.09924.000.09923.600.09923.710.10229.040.106
" + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.348, + 0.891, + 0.555 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.564, + 0.895, + 0.594 + ], + "angle": 0, + "content": "Figure 6. Qualitative comparison of state-of-the-art video reconstruction methods on the proposed measured datasets in terms of image quality. The images have been magnified for optimal viewing quality." + }, + { + "type": "image", + "bbox": [ + 0.08, + 0.611, + 0.231, + 0.708 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.236, + 0.614, + 0.374, + 0.708 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.377, + 0.613, + 0.516, + 0.708 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.613, + 0.657, + 0.708 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.66, + 0.612, + 0.797, + 0.708 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.806, + 0.61, + 0.887, + 0.71 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.729, + 0.897, + 0.772 + ], + "angle": 0, + "content": "Figure 7. Frame-wise, quantitative comparison of state-of-the-art video reconstruction methods in terms of PSNR for varying scenarios of the simulation dataset. Note that some algorithms exhibit pronounced discontinuities in the observed PSNR curves indicating the presence of video glitches." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.797, + 0.471, + 0.875 + ], + "angle": 0, + "content": "dataset based on a EVS camera that is collocated next to a CIS camera. Since the paper focuses on mitigating artifacts due to pixel latency, two indoor scenarios (\"spinning plate\" and \"spinning umbrella\") were selected. More details of the dataset used are provided in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.797, + 0.595, + 0.812 + ], + "angle": 0, + "content": "5.2. Results" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.826, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Image Quality Table 1 shows quantitative results for the reconstructed video. Results in PSNR and LPIPS show that the proposed method outperforms other state-of-the-art methods in most of the scenes, especially with fast motion (\"slingshot egg\") or complicated rotation motion (\"basket-" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.923, + 0.521, + 0.935 + ], + "angle": 0, + "content": "25177" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.091, + 0.473, + 0.362 + ], + "angle": 0, + "content": "ball\", \"fan\"). As expected the refinement network which aims to alleviate artifacts and improve image quality reports a higher score. Snippets of the reconstructed videos are presented in Figure 1 and Figure 6 for simulated and measured data, respectively. Comprehensive input images are provided in the supplementary material. Thanks to the explicit EVS latency modeling and compensation, our results avoid obvious shadowing/ghosting artifacts - especially if there is fast motion under challenging lighting conditions. It is also noted that the proposed method successfully corrects the RS effect, as shown in the \"basketball\" scene. Learning-based methods face difficulties, as latency and the RS effect have not been considered in model design and training. The CIS-only method FILM provides natural image quality with low noise but fails when the objects are too fast to be captured by CIS, as shown in the \"slingshot egg\" scene. For the \"basketball\" scene, our method shows an advantage in details and sharpness compared to FILM." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.364, + 0.473, + 0.576 + ], + "angle": 0, + "content": "In specific scenarios, such as the \"spinning umbrella\" scene within the HS-ERGB dataset and the \"running man\" scene in the simulation dataset, our method did not yield improved image quality. This observation can be attributed to several factors. Firstly, the HS-ERGB dataset employs a different EVS with different pixel and readout latency as well as refractory period. Secondly, this dataset is based on a dual camera instead of a hybrid sensor so the spatiotemporal alignment may differ. Thirdly, the scenes in question involve relatively slower object movement speeds resulting in reduced ghosting. In essence, the dynamic characteristics of the EVS, the input signal frequency range, and the reconstruction algorithm influence the quality of the reconstructed images." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.58, + 0.47, + 0.671 + ], + "angle": 0, + "content": "Video Glitch Effect To evaluate the glitch issue, we compute the PSNR for each reconstructed frame as shown in Figure 7. Overall, our method gives temporally more stable PSNR curves, which stem from the proposed row-based reconstruction scheme. It is also noted that the refinement network contributes significantly to higher PSNR." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.674, + 0.473, + 0.903 + ], + "angle": 0, + "content": "Pixel-wise Error Analysis We select a pixel within the region of interest (indicated as the yellow box in Figure 1(b)) and verify the pixel-wise reconstruction accuracy. As shown in Figure 8(a), with the help of pixel latency compensation the reconstructed \\( V_{\\mathrm{DC}} \\) after optimization shows a sharper peak (with \\( 250~\\mu \\mathrm{s} \\) duration) compared to \\( V_{\\mathrm{FE}} \\) (with over \\( 1\\mathrm{ms} \\) duration). The high-frequency fluctuations on ground truth \\( V_{\\mathrm{DC}} \\) curve cannot be reconstructed due to the \\( V_{\\mathrm{FE}} \\) bandwidth limit. An overshoot is observed after the peak resulting in a dark tail along the pellet's trajectory, as shown in Figure 1(b). This may stem from 1) the \\( V_{\\mathrm{FE}} \\) bandwidth, 2) the approximation error in equation (12), 3) EVS measurement noise, and 4) choice of weighing matrices for CIS and EVS. Figure 8(b), indicates that the proposed method gives an excellent reconstruction of the pel" + }, + { + "type": "image", + "bbox": [ + 0.513, + 0.095, + 0.891, + 0.187 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.521, + 0.187, + 0.541, + 0.2 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.206, + 0.89, + 0.297 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.521, + 0.297, + 0.54, + 0.31 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image_caption", + "bbox": [ + 0.688, + 0.298, + 0.748, + 0.31 + ], + "angle": 0, + "content": "time \\((\\mu s)\\)" + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.33, + 0.895, + 0.442 + ], + "angle": 0, + "content": "Figure 8. Pixel-wise reconstruction results for the \"slingshot egg\" scene. The selected pixel is located at the path of the pellet, in the yellow box of Figure 1. The refinement network is not used here. (a) Comparison of the ground truth \\( V_{\\mathrm{DC}} \\), reconstructed \\( V_{\\mathrm{FE}} \\) and \\( V_{\\mathrm{DC}} \\), where the ground truth \\( V_{\\mathrm{DC}} \\) is computed from the ground truth DN of the pixel at 10000 FPS. (b) Comparison of reconstructed pixel-wise DN of state-of-the-art video reconstruction methods." + }, + { + "type": "image", + "bbox": [ + 0.512, + 0.457, + 0.695, + 0.552 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.546, + 0.554, + 0.661, + 0.564 + ], + "angle": 0, + "content": "(a) with no compensation" + }, + { + "type": "image", + "bbox": [ + 0.695, + 0.457, + 0.877, + 0.552 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.71, + 0.554, + 0.858, + 0.564 + ], + "angle": 0, + "content": "(b) with \\(V_{\\mathrm{FE}}\\) delay compensation" + }, + { + "type": "image", + "bbox": [ + 0.512, + 0.566, + 0.695, + 0.661 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.535, + 0.663, + 0.671, + 0.674 + ], + "angle": 0, + "content": "(c) with RL+RP compensation" + }, + { + "type": "image", + "bbox": [ + 0.695, + 0.566, + 0.877, + 0.661 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.707, + 0.663, + 0.865, + 0.675 + ], + "angle": 0, + "content": "(d) with \\(V_{\\mathrm{FE}} + \\mathrm{RL} + \\mathrm{RP}\\) compensation" + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.677, + 0.894, + 0.718 + ], + "angle": 0, + "content": "Figure 9. Qualitative results of ablation study on the \"basketball\" scene. Refer to Figure 1(a) for \"with \\( V_{\\mathrm{FE}} + \\mathrm{RL} + \\mathrm{RP} \\) compensation and refinement\" result." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.747, + 0.741, + 0.762 + ], + "angle": 0, + "content": "let's peak DN despite the overshoot." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Ablation Study We ablate the refinement network, \\( V_{\\mathrm{FE}} \\) delay compensation, readout latency (RL) and refractory period (RP) compensation of our method. The numerical evaluation are subject to selected ROIs on motion region. To ablate \\( V_{\\mathrm{FE}} \\) delay compensation, we set the time constant \\( \\tau = 0 \\) such that \\( V_{\\mathrm{FE}} = V_{\\mathrm{DC}} \\) in Eq. (7). To ablate readout latency and refractory period compensation, we use Eq. (8) instead of Eq. (12) in the EVS measurement model. Results are summarized in Table 2 and Figure 9. The con" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.923, + 0.521, + 0.935 + ], + "angle": 0, + "content": "25178" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.333 + ], + "angle": 0, + "content": "Figuration using all modules gives significantly better visual and slightly better numerical results in PSNR compared to having latency compensation and refinement disabled. Disabling \\( V_{\\mathrm{FE}} \\) or RL+RP compensation significantly worsens the image quality, as severe shadowing artifacts are observed. As for LPIPS in \"slingshot egg\" scene, the difference is not significant due to the small area of pellet ROI. It is also noted that adding RL+RP compensation leads to improvement in the \"basketball\" but degrades the performance in \"slingshot egg\" when comparing with \\( V_{\\mathrm{FE}} \\) delay compensation only. This is because RL+RP compensation is based on the assumption that the slope of VFE is locally constant for a short period. In cases such as \"slingshot egg\", the assumption may not hold due to close positive and negative events. An ablation study against the use of NAFNet-only is included in supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.337, + 0.47, + 0.534 + ], + "angle": 0, + "content": "Performance Regarding Illuminance and Speed The proposed method is further investigated under a controlled environment using varying illuminance and object motion speed. BEW [5] is used to evaluate the image quality by measuring edge sharpness. Figure 10(a)-(h) presents the comparison of the state-of-the-art methods on a rotating disk under \\(1000\\mathrm{lx}\\) and \\(292\\mathrm{rpm}\\). It is observed that the proposed method reconstructs sharper edges compared to the others. Other EVS-based methods struggle resolving blurry edges in presence of fast motion and also the CIS-based FILM method fails to generate a sharp reconstruction. Figure 10(i) shows that the proposed method reports the lowest BEW score (0.7) and has the sharpest edge profile." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.535, + 0.471, + 0.626 + ], + "angle": 0, + "content": "The BEW scores under different speed and illumination conditions are summarized in Figure 11. Figure 11(a) shows that the BEW increases with faster rotation speeds. In Figure 11(b), the lower ambient light condition leads to higher \\( V_{\\mathrm{FE}} \\) latency degrading image quality. Our method achieves the lowest BEW throughout all illuminance levels." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.638, + 0.196, + 0.653 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.659, + 0.473, + 0.901 + ], + "angle": 0, + "content": "An EVS-assisted framework to solve motion blur, rolling-shutter artifacts and video-frame interpolation in a joint manner was presented. Existing methods use either a simple double integral model or a learning-based model, neither of which explicitly address EVS sensor imperfections such as EVS pixel latency, readout latency or refractory period. This work overcomes these limitations by explicitly modeling these characteristics. The compensation of readout latency and refractory period is enabled by having access to in-pixel and off-pixel time-stamps. The nonlinear inverse model was solved through joint graph optimization of subsequent frames utilizing row-by-row exposure information to overcome rolling-shutter artifacts and possible glitches between subsequent frames. A refinement network was proposed to further improve the image quality by using specifically designed spatial attention blocks. We showed that" + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.088, + 0.605, + 0.159 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.508, + 0.16, + 0.607, + 0.168 + ], + "angle": 0, + "content": "(a) Ours w/o refinement" + }, + { + "type": "image", + "bbox": [ + 0.606, + 0.089, + 0.701, + 0.159 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.636, + 0.16, + 0.667, + 0.168 + ], + "angle": 0, + "content": "(b) EDI" + }, + { + "type": "image", + "bbox": [ + 0.701, + 0.089, + 0.796, + 0.159 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.734, + 0.16, + 0.764, + 0.168 + ], + "angle": 0, + "content": "(c)AKF" + }, + { + "type": "image", + "bbox": [ + 0.797, + 0.089, + 0.894, + 0.159 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.815, + 0.16, + 0.873, + 0.168 + ], + "angle": 0, + "content": "(d) Time Lens" + }, + { + "type": "image", + "bbox": [ + 0.509, + 0.17, + 0.605, + 0.24 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.531, + 0.242, + 0.568, + 0.251 + ], + "angle": 0, + "content": "(e) FILM" + }, + { + "type": "image", + "bbox": [ + 0.606, + 0.17, + 0.701, + 0.24 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.627, + 0.242, + 0.674, + 0.251 + ], + "angle": 0, + "content": "(f) CBMNet" + }, + { + "type": "image", + "bbox": [ + 0.701, + 0.17, + 0.796, + 0.24 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.724, + 0.242, + 0.773, + 0.251 + ], + "angle": 0, + "content": "(g) EvUnroll" + }, + { + "type": "image", + "bbox": [ + 0.797, + 0.17, + 0.894, + 0.24 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.822, + 0.242, + 0.866, + 0.251 + ], + "angle": 0, + "content": "(h) REFID" + }, + { + "type": "image", + "bbox": [ + 0.51, + 0.255, + 0.888, + 0.357 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.373, + 0.893, + 0.469 + ], + "angle": 0, + "content": "Figure 10. (a)-(h) Comparison of state-of-the-art video reconstruction methods on the proposed rotating disk datasets with rotating speed \\(292\\mathrm{rpm}\\) and environment illuminance \\(1000\\mathrm{lx}\\). The yellow box shows magnified edges. The visual \"misalignment\" is due to EVS latency. (i) The DN-edge width relation at a fixed segment of 40 pixels (red line in (a)) on reconstructed rotating disk edge. The refinement network is not used." + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.489, + 0.703, + 0.64 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.584, + 0.64, + 0.669, + 0.651 + ], + "angle": 0, + "content": "(a) speed (rpm)" + }, + { + "type": "image", + "bbox": [ + 0.709, + 0.492, + 0.884, + 0.639 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.751, + 0.64, + 0.862, + 0.65 + ], + "angle": 0, + "content": "(b) illumination (lux)" + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.669, + 0.892, + 0.753 + ], + "angle": 0, + "content": "Figure 11. Quantitative comparison of state-of-the-art video reconstruction methods on the proposed \"rotating disk\" scene in terms of BEW. The lower the BEW values, the sharper the edge. The refinement network is not used. (a) BEW under various rotation speeds with a fixed illuminance of \\(1000\\mathrm{lx}\\). (b) BEW under various illuminance levels with a fixed speed of \\(292\\mathrm{rpm}\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.779, + 0.892, + 0.824 + ], + "angle": 0, + "content": "the proposed method outperforms state-of-the-art methods in reconstructed image quality with an up to \\(4\\mathrm{dB}\\) improvement in PSNR and \\(12\\%\\) improvement in LPIPS score." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.825, + 0.892, + 0.9 + ], + "angle": 0, + "content": "Limitation Since the pixel latency highly depends on EVS design parameters, object motion, and ambient illuminance, the image quality improvement after latency compensation may be subtle for slow objects or in bright scenes such as \"running man\" and \"spinning umbrella\"." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.923, + 0.519, + 0.935 + ], + "angle": 0, + "content": "25179" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.184 + ], + "angle": 0, + "content": "[1] Wenbo Bao, Wei-Sheng Lai, Chao Ma, Xiaoyun Zhang, Zhiyong Gao, and Ming-Hsuan Yang. Depth-aware video frame interpolation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3703-3712, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.185, + 0.472, + 0.253 + ], + "angle": 0, + "content": "[2] Patrick Bardow, Andrew J Davison, and Stefan Leutenegger. Simultaneous optical flow and intensity estimation from an event camera. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 884-892, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.255, + 0.472, + 0.31 + ], + "angle": 0, + "content": "[3] Christian Brandli, Raphael Berner, Minhao Yang, Shih-Chii Liu, and Tobi Delbruck. A \\(240 \\times 180\\) 130 db \\(3\\mu s\\) latency global shutter spatiotemporal vision sensor. IEEE Journal of Solid-State Circuits, 49(10):2333-2341, 2014. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.311, + 0.472, + 0.367 + ], + "angle": 0, + "content": "[4] Xiaojie Chu, Liangyu Chen, and Wenqing Yu. Nafssr: Stereo image super-resolution using nafnet. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, pages 1239-1248, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.368, + 0.471, + 0.409 + ], + "angle": 0, + "content": "[5] Hai Dinh, Qinyi Wang, Fangwen Tu, Brett Frymire, and Bo Mu. Evaluation of motion blur image quality in video frame interpolation. *Electronic Imaging*, 35:262-1, 2023. 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.41, + 0.471, + 0.479 + ], + "angle": 0, + "content": "[6] Julius Erbach, Stepan Tulyakov, Patricia Vitoria, Alfredo Bochicchio, and Yuanyou Li. Evshutter: Transforming events for unconstrained rolling shutter correction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13904-13913, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.48, + 0.471, + 0.508 + ], + "angle": 0, + "content": "[7] C.Posch et al. Sensitivity and uniformity of a \\(0.18\\mu \\mathrm{m}\\) CMOS temporal contrast pixel array. In ISCAS, 2011. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.509, + 0.471, + 0.536 + ], + "angle": 0, + "content": "[8] Xinyu Zhou et al. Evunroll: Neuromorphic events based rolling shutter image correction. In CVPR, 2022. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.537, + 0.471, + 0.606 + ], + "angle": 0, + "content": "[9] Guillermo Gallego, Tobi Delbrück, Garrick Orchard, Chiara Bartolozzi, Brian Taba, Andrea Censi, Stefan Leutenegger, Andrew J Davison, Jörg Conradt, Kostas Daniilidis, et al. Event-based vision: A survey. IEEE transactions on pattern analysis and machine intelligence, 44(1):154-180, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.607, + 0.471, + 0.662 + ], + "angle": 0, + "content": "[10] Yue Gao, Siqi Li, Yipeng Li, Yandong Guo, and Qionghai Dai. Superfast: \\(200 \\times\\) video frame interpolation via event camera. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.663, + 0.471, + 0.718 + ], + "angle": 0, + "content": "[11] Leon A Gatys, Alexander S Ecker, and Matthias Bethge. Image style transfer using convolutional neural networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2414-2423, 2016. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.719, + 0.471, + 0.789 + ], + "angle": 0, + "content": "[12] Daniel Gehrig, Antonio Loquercio, Konstantinos G Derpanis, and Davide Scaramuzza. End-to-end learning of representations for asynchronous event-based data. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5633-5643, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.79, + 0.471, + 0.857 + ], + "angle": 0, + "content": "[13] Giorgio Grisetti, Rainer Kümmerle, Hauke Strasdat, and Kurt Konolige. g2o: A general framework for (hyper) graph optimization. In Proceedings of the IEEE International Conference on Robotics and Automation (ICRA), pages 9-13, 2011. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.859, + 0.471, + 0.902 + ], + "angle": 0, + "content": "[14] Menghan Guo, Shoushun Chen, Zhe Gao, Wenlei Yang, Peter Bartkovjak, Qing Qin, Xiaqin Hu, Dahai Zhou, Qiping Huang, Masayuki Uchiyama, Yoshiharu Kudo, Shimpei" + }, + { + "type": "list", + "bbox": [ + 0.079, + 0.115, + 0.472, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.894, + 0.176 + ], + "angle": 0, + "content": "Fukuoka, Chengcheng Xu, Hiroaki Ebihara, Xueqing Wang, Peiwen Jiang, Bo Jiang, Bo Mu, Huan Chen, Jason Yang, T. J. Dai, and Andreas Suess. A three-wafer-stacked hybrid 15-mpixel cis + 1-mpixel evs with 4.6-GeV/s readout, inpixel tdc, and on-chip isp and esp function. IEEE Journal of Solid-State Circuits, pages 1–10, 2023. 1, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.178, + 0.894, + 0.302 + ], + "angle": 0, + "content": "[15] Menghan Guo, Shoushun Chen, Zhe Gao, Wenlei Yang, Peter Bartkovjak, Qing Qin, Xiaqin Hu, Dahei Zhou, Masayuki Uchiyama, Yoshiharu Kudo, Shimpei Fukuoka, Chengcheng Xu, Hiroaki Ebihara, Andy Wang, Peiwen Jiang, Bo Jiang, Bo Mu, Huan Chen, Jason Yang, TJ Dai, and Andreas Suess. A 3-wafer-stacked hybrid 15MPixel CIS + 1MPixel EVS with 4.6GEvent/s readout, in-pixel TDC and on-chip ISP and ESP function. In 2022 IEEE International Solid-State Circuits Conference (ISSCC), 2023. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.304, + 0.894, + 0.359 + ], + "angle": 0, + "content": "[16] Yuhuang Hu, Shih-Chii Liu, and Tobi Delbruck. v2e: From video frames to realistic dvs events. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1312-1321, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.361, + 0.893, + 0.417 + ], + "angle": 0, + "content": "[17] Zhewei Huang, Tianyuan Zhang, Wen Heng, Boxin Shi, and Shuchang Zhou. Real-time intermediate flow estimation for video frame interpolation. In European Conference on Computer Vision, pages 624-642. Springer, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.418, + 0.893, + 0.501 + ], + "angle": 0, + "content": "[18] Huaizu Jiang, Deqing Sun, Varun Jampani, Ming-Hsuan Yang, Erik Learned-Miller, and Jan Kautz. Super slomo: High quality estimation of multiple intermediate frames for video interpolation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 9000-9008, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.503, + 0.893, + 0.532 + ], + "angle": 0, + "content": "[19] Nocedal Jorge and J Wright Stephen. Numerical optimization. Springer, 2006. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.533, + 0.894, + 0.588 + ], + "angle": 0, + "content": "[20] Damien Joubert, Mathieu Hébert, Hubert Konik, and Christophe Lavergne. Characterization setup for event-based imagers applied to modulated light signal detection. Applied optics, 58(6):1305-1317, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.59, + 0.893, + 0.646 + ], + "angle": 0, + "content": "[21] Damien Joubert, Alexandre Marcireau, Nic Ralph, Andrew Jolley, André van Schaik, and Gregory Cohen. Event camera simulator improvements via characterized parameters. Frontiers in Neuroscience, page 910, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.648, + 0.893, + 0.73 + ], + "angle": 0, + "content": "[22] Hanme Kim, Stefan Leutenegger, and Andrew J Davison. Real-time 3d reconstruction and 6-dof tracking with an event camera. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part VI 14, pages 349-364. Springer, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.732, + 0.894, + 0.815 + ], + "angle": 0, + "content": "[23] Taewoo Kim, Yujeong Chae, Hyun-Kurl Jang, and Kuk-Jin Yoon. Event-based video frame interpolation with cross-modal asymmetric bidirectional motion fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 18032-18042, 2023. 1, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.817, + 0.894, + 0.902 + ], + "angle": 0, + "content": "[24] Kazutoshi Kodama, Yusuke Sato, Yuhi Yorikado, Raphael Berner, Kyoji Mizoguchi, Takahiro Miyazaki, Masahiro Tsukamoto, Yoshihisa Matoba, Hirotaka Shinozaki, Atsumi Niwa, et al. 1.22 \\(\\mu\\)m 35.6 mpixel rgb hybrid event-based vision sensor with 4.88 \\(\\mu\\)m-pitch event pixels and up to 10k event frame rate by adaptive control on event sparsity." + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.923, + 0.52, + 0.935 + ], + "angle": 0, + "content": "25180" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.11, + 0.092, + 0.469, + 0.12 + ], + "angle": 0, + "content": "In 2023 IEEE International Solid-State Circuits Conference (ISSCC), pages 92-94. IEEE, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.122, + 0.469, + 0.175 + ], + "angle": 0, + "content": "[25] Patrick Lichtsteiner, Christoph Posch, and Tobi Delbruck. A \\(128 \\times 128\\) 120 db \\(15\\mu s\\) latency asynchronous temporal contrast vision sensor. IEEE Journal of Solid-State Circuits, 43 (2):566-576, 2008. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.178, + 0.469, + 0.232 + ], + "angle": 0, + "content": "[26] Xiaozheng Mou, Kaijun Feng, Alex Yi, Steve Wang, Huan Chen, Xiaoqin Hu, Menghan Guo, Shoushun Chen, and Andreas Suess. Accurate event simulation using high-speed videos. *Electronic Imaging*, 34(7):242-1, 2022. 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.234, + 0.469, + 0.302 + ], + "angle": 0, + "content": "[27] Liyuan Pan, Cedric Scheerlinck, Xin Yu, Richard Hartley, Miaomiao Liu, and Yuchao Dai. Bringing a blurry frame alive at high frame-rate with an event camera. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6820-6829, 2019. 1, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.303, + 0.469, + 0.37 + ], + "angle": 0, + "content": "[28] Liyuan Pan, Richard Hartley, Cedric Scheerlinck, Miaomiao Liu, Xin Yu, and Yuchao Dai. High frame rate video reconstruction based on an event camera. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(5):2519-2533, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.373, + 0.469, + 0.44 + ], + "angle": 0, + "content": "[29] Federico Paredes-Valles and Guido CHE de Croon. Back to event basics: Self-supervised learning of image reconstruction for event cameras via photometric constancy. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3446–3455, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.443, + 0.469, + 0.496 + ], + "angle": 0, + "content": "[30] Christoph Posch and Daniel Matolin. Sensitivity and uniformity of a \\(0.18\\mu \\mathrm{m}\\) cmos temporal contrast pixel array. In 2011 IEEE International Symposium of Circuits and Systems (ISCAS), pages 1572-1575. IEEE, 2011. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.498, + 0.469, + 0.552 + ], + "angle": 0, + "content": "[31] Henri Rebecq, René Ranftl, Vladlen Koltun, and Davide Scaramuzza. High speed and high dynamic range video with an event camera. IEEE transactions on pattern analysis and machine intelligence, 43(6):1964-1980, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.554, + 0.469, + 0.609 + ], + "angle": 0, + "content": "[32] Fitsum Reda, Janne Kontkanen, Eric Tabellion, Deqing Sun, Caroline Pantofaru, and Brian Curless. Film: Frame interpolation for large motion. In European Conference on Computer Vision, pages 250-266. Springer, 2022. 1, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.61, + 0.469, + 0.692 + ], + "angle": 0, + "content": "[33] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical Image Computing and Computer-Assisted Intervention-MICCAI 2015: 18th International Conference, Munich, Germany, October 5-9, 2015, Proceedings, Part III 18, pages 234-241. Springer, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.693, + 0.469, + 0.748 + ], + "angle": 0, + "content": "[34] Cedric Scheerlinck, Nick Barnes, and Robert Mahony. Continuous-time intensity estimation using event cameras. In Asian Conference on Computer Vision, pages 308-324. Springer, 2018. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.749, + 0.469, + 0.817 + ], + "angle": 0, + "content": "[35] Andreas Suess, Menghan Guo, Rui Jiang, Xiaozheng Mou, Qiping Huang, Wenlei Yang, and Shoushun Chen. Physical modeling and parameter extraction for event-based vision sensors. In Proc. IISS Int. Image Sensor Workshop (IISW), Edinburgh, UK, page R5, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.818, + 0.469, + 0.899 + ], + "angle": 0, + "content": "[36] Lei Sun, Christos Sakaridis, Jingyun Liang, Peng Sun, Jiezhang Cao, Kai Zhang, Qi Jiang, Kaiwei Wang, and Luc Van Gool. Event-based frame interpolation with ad-hoc deblurring. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18043-18052, 2023. 1, 2, 5" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.469, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.175 + ], + "angle": 0, + "content": "[37] Stepan Tulyakov, Daniel Gehrig, Stamatios Georgoulis, Julius Erbach, Mathias Gehrig, Yuanyou Li, and Davide Scaramuzza. Time lens: Event-based video frame interpolation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16155-16164, 2021. 1, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.178, + 0.892, + 0.26 + ], + "angle": 0, + "content": "[38] Stepan Tulyakov, Alfredo Bochicchio, Daniel Gehrig, Stamatios Georgoulis, Yuanyou Li, and Davide Scaramuzza. Time lens: Event-based frame interpolation with parametric non-linear flow and multi-scale fusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17755-17764, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.262, + 0.892, + 0.329 + ], + "angle": 0, + "content": "[39] Ziwei Wang, Yonhon Ng, Cedric Scheerlinck, and Robert Mahony. An asynchronous kalman filter for hybrid event cameras. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 448-457, 2021. 1, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.332, + 0.892, + 0.387 + ], + "angle": 0, + "content": "[40] Sanghyun Woo, Jongchan Park, Joon-Young Lee, and In So Kweon. Cbam: Convolutional block attention module. In Proceedings of the European conference on computer vision (ECCV), pages 3–19, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.389, + 0.892, + 0.443 + ], + "angle": 0, + "content": "[41] Zeyu Xiao, Wenming Weng, Yueyi Zhang, and Zhiwei Xiong. Eva2: Event-assisted video frame interpolation via cross-modal alignment and aggregation. IEEE Transactions on Computational Imaging, 8:1145–1158, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.446, + 0.892, + 0.514 + ], + "angle": 0, + "content": "[42] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.516, + 0.892, + 0.571 + ], + "angle": 0, + "content": "[43] Xiang Zhang and Lei Yu. Unifying motion deblurring and frame interpolation with events. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17765-17774, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.573, + 0.892, + 0.64 + ], + "angle": 0, + "content": "[44] Xinyu Zhou, Peiqi Duan, Yi Ma, and Boxin Shi. Evunroll: Neuromorphic events based rolling shutter image correction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17775-17784, 2022. 1, 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.64 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.923, + 0.518, + 0.935 + ], + "angle": 0, + "content": "25181" + } + ] +] \ No newline at end of file diff --git a/2024/EVS-assisted Joint Deblurring Rolling-Shutter Correction and Video Frame Interpolation through Sensor Inverse Modeling/2babbe27-3090-44dc-8f9d-2d16b50cc5ab_origin.pdf b/2024/EVS-assisted Joint Deblurring Rolling-Shutter Correction and Video Frame Interpolation through Sensor Inverse Modeling/2babbe27-3090-44dc-8f9d-2d16b50cc5ab_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..abee85d0c8aa2cdb1982713f2c09478032e0f31e --- /dev/null +++ b/2024/EVS-assisted Joint Deblurring Rolling-Shutter Correction and Video Frame Interpolation through Sensor Inverse Modeling/2babbe27-3090-44dc-8f9d-2d16b50cc5ab_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:de1a66029fb858694943c03d198d89d879bfebb7e6592ea83a8717aa34b6311d +size 9002106 diff --git a/2024/EVS-assisted Joint Deblurring Rolling-Shutter Correction and Video Frame Interpolation through Sensor Inverse Modeling/full.md b/2024/EVS-assisted Joint Deblurring Rolling-Shutter Correction and Video Frame Interpolation through Sensor Inverse Modeling/full.md new file mode 100644 index 0000000000000000000000000000000000000000..fb7c876fd8ad780b90b9a179cfd73ae91086fbdb --- /dev/null +++ b/2024/EVS-assisted Joint Deblurring Rolling-Shutter Correction and Video Frame Interpolation through Sensor Inverse Modeling/full.md @@ -0,0 +1,335 @@ +# EVS-assisted joint Deblurring, Rolling-Shutter Correction and Video Frame Interpolation through Sensor Inverse Modeling + +Rui Jiang*, Fangwen Tu*, Yixuan Long, Aabhaas Vaish, Bowen Zhou, Qinyi Wang Wei Zhang, Yuntan Fang, Luis Eduardo Garcia Capel, Bo Mu, Tiejun Dai, Andreas Suess + +OMNIVISION + +*Equal contribution + +![](images/0a6598df303e0632b86ab8d92416d59c00dbbc089ffe296a5c611f8025de048d.jpg) +Figure 1. Qualitative results comparing the proposed method with EDI [27], AKF [39], Time Lens [37], FILM [32], CBMNet [23], EvUnroll [44], and REFID [36] on simulation datasets. In the "basketball" scene, our method provides frames that are rolling shutter effect corrected and deblurred. In the "slingshot egg" scene, our method can reconstruct the shape of the pellet with the best quality, as indicated in the yellow box. The images have been magnified for optimal viewing quality. + +# Abstract + +Event-based Vision Sensors (EVS) gain popularity in enhancing CMOS Image Sensor (CIS) video capture. Nonidealities of EVS such as pixel or readout latency can significantly influence the quality of the enhanced images and warrant dedicated consideration in the design of fusion algorithms. A novel approach for jointly computing deblurred, rolling-shutter artifact corrected high-speed videos with frame rates up to 10 000 FPS using inherently blurry rolling shutter CIS frames of 120 FPS to 150 FPS in conjunction with EVS data from a hybrid CIS-EVS sensor is presented. EVS pixel latency, readout latency and the sensor's refractory period are explicitly incorporated into the measurement model. This inverse function problem is solved on a per-pixel manner using an optimization-based framework. The interpolated images are subsequently processed by a novel refinement network. The proposed method is evaluated using simulated and measured datasets, under natural and controlled environments. Extensive experiments show reduced shadowing effect, a 4 dB increment in PSNR, and a $12\%$ improvement in LPIPS score compared to state-of-the-art methods. + +# 1. Introduction + +Event-Based Vision Sensors (EVS) capture our world from a completely different perspective compared to CMOS Image Sensors (CIS) [9, 14, 24, 25]. In EVS, each pixel independently triggers an event if a relative illuminance change on the pixel reaches a certain triggering threshold (typically in the range of $10\%$ to $30\%$ ). Conversely to CIS, EVS operate time-continuously and asynchronously - there is no common exposure period or "frame rate". Thus, EVS pixels can be read out comparatively fast in a low-power manner enabling efficient high-speed data capture. Recent work on enhancement of high-quality images [23, 27, 34, 37, 39] focuses on fusing CIS and EVS information so that the high fidelity of CIS measurement and the fast response of EVS measurement can complement each other. + +In practice it is observed [3, 14, 21, 25] that the dynamic characteristics of EVS pixels are highly dependent on the illumininance in the sensor plane. Under low-light conditions such as indoor capture, a slower EVS response during video enhancement can result in ghosting or blurry frames and needs to be considered in algorithm design. Furthermore, the inevitable readout latency and refractory period of EVS pixels can significantly aggravate artifacts. Current methods are not specifically designed to address sensor nonidealities. + +This paper presents an inverse model capturing the EVS pixel latency (namely the $V_{\mathrm{FE}}$ delay), the readout latency, and the refractory period. The inverse model is solved through nonlinear graph optimization to jointly address deblurring, rolling-shutter (RS) correction and video-frame interpolation while compensating for sensor nonidealities. In addition to the inverse model, a learning-based refinement module is proposed to enhance the image quality by further mitigating the noise and artifacts. The proposed method is compared with existing CIS-based [32] and EVS-assisted [8, 23, 27, 36, 37, 39] video enhancement methods under challenging indoor scenarios with fast motion. As illustrated in Figure 1, the proposed algorithm produces excellent image quality with much fewer artifacts. + +The contributions of the paper can be summarized as: + +- A formulation of a joint deblurring, rolling-shutter correction and video-frame interpolation problem based on CIS and EVS fusion, where EVS sensor nonidealities are explicitly modeled (Section 3), +- An optimization methodology jointly solving this coupled inverse problem (Section 4.1), as well as a post-processing network for artifact removal (Section 4.2), +- A comprehensive comparison and analysis of simulation and measured datasets in natural and controlled environments (Section 5). + +# 2. Related Work + +Video Frame Interpolation Video interpolation aims to synthesize intermediate frames between two images to achieve a temporally coherent video sequence. Traditional CIS-based methods depend on optical flow for a smooth transition. In SuperSlomo [18], the bi-directional flows are combined with a linear combination followed by a refinement using U-Net [33]. Such optical flow-based methods, however, suffer from occlusion problems. In DAIN [1], researchers estimate occlusion areas using depth information and use contextual information from neighboring pixels to fill the occluded regions. In RIFE [17], the flow and fusion maps are generated simultaneously without an additional optical flow module for final result synthesis. FILM [32] presents a multi-scale feature extractor within a unified network to achieve an enlarged receptive field aimed at handling large object motion in the video. + +Event-assisted Frame Enhancement A key issue lowering the image quality of CIS-based frame interpolation methods is the lack of information between frames. EVS provides such information thanks to its asynchronous low latency capture not relying on a global exposure period. Some early studies [31, 34] show that intensity reconstruction purely based on EVS modality is feasible. However, as EVS captures events based on differential changes of illuminance, this approach struggles to estimate the absolute light level. In [2], the optical flow and intensity frames are + +estimated simultaneously by solving an optimization problem that takes EVS measurements as data terms. The authors in [22] take a step further by formulating the event-based SLAM problem, which can be decomposed into camera ego-motion estimation and 3-D scene reconstruction. However, the quantization error makes it difficult to reconstruct high-quality frames from EVS alone. As EVS pixels trigger independently, researchers simplify the problem formulation to pixel-wise estimation by fusing CIS and EVS, without seeking explicit representation of the 3-D environment. The Event-based Double Integral (EDI) [27, 28] method is proposed as a straightforward and effective way to deblur CIS frames and generate high frame rate videos. EVS and CIS measurements are connected by integrating events twice, such that the continuous intensity curve for each pixel can be determined. [34] presents a complementary filter-based framework to combine a high-pass signal from EVS and a low-pass signal from CIS to compute an all-pass signal for pixel intensity. To make the complementary filter gain adaptive to measurement noise, a pixel-wise Kalman filter is used in [39] where a unifying EVS/CIS uncertainty model is proposed. + +Machine learning methods such as Time Lens [37] and its subsequent versions [10, 36, 38, 41] or others [29, 43] concatenate both CIS and EVS data using a voxel grid representation [12] as input to feed the network of two branches: an optical flow warping-based branch and a synthesis-based branch. The final result is blended with a trained attention module to avoid the disadvantages of any individual branch. CBMNet [23] is proposed to handle complex real-world motion using a novel cross-modal asymmetric bidirectional motion field estimation. Due to the high temporal resolution of EVS data, the rolling shutter problem which is commonly observed in fast-motion scenarios can be corrected using EVS data [6, 44]. + +EVS Latency Many studies model EVS latency [20, 25, 30, 35], but few of them have been applied to algorithm design such as frame enhancement. To the best of the authors' knowledge, this is the first work that considers an EVS latency model explicitly in generating high frame rate, deblurred, rolling-shutter artifact-free videos. + +# 3. Problem Formulation + +Event Pixel Operation and Pixel Latency Model In this work, a hybrid CIS-EVS sensor [14, 15] is used. As shown in Figure 2(a), EVS and CIS share the same photodiode design on the hybrid sensor and it is assumed that the light level is identical for nearby CIS and EVS pixels. The photocurrent $i_{\mathrm{pd}}$ is converted to front-end voltage $V_{\mathrm{FE}}$ in module $\mathrm{i}2\mathrm{v}$ , which models dynamic characteristics as a 1st-order Low-Pass Filter (LPF) [21, 26]. The update rule for $V_{\mathrm{FE}}$ between discrete event indices $k$ and $k + 1$ can be written as: + +![](images/d5b8847e1319d15d45a2c52e093e33b06a6599f473f2183a404a6577da4202ab.jpg) + +![](images/d400ca00c23e5f4cac2721238af5c6c712e24b8f18b992c3794696c18c2b4678.jpg) +Figure 2. (a) Hybrid CIS-EVS sensor principle. The upper and lower branches indicate physical quantity conversion in CIS pixels and EVS pixels, respectively. (b) Example of $V_{\mathrm{DC}}$ and $V_{\mathrm{FE}}$ curves and triggered events that illustrate the impact of $V_{\mathrm{FE}}$ delay, readout latency, and refractory period. + +$$ +V _ {\mathrm {F E}} (k + 1) = V _ {\mathrm {F E}} (k) + \gamma (k) \left(V _ {\mathrm {D C}} (k + 1) - V _ {\mathrm {F E}} (k)\right), \tag {1} +$$ + +where $V_{\mathrm{DC}}(k + 1) = f_{\mathrm{DC}}(i_{\mathrm{pd}}(k + 1))$ represents the $V_{\mathrm{FE}}$ under DC excitation (i.e., the $V_{\mathrm{FE}}$ with infinite LPF bandwidth); $f_{\mathrm{DC}}(i_{\mathrm{pd}})$ is the logarithmic-like current-voltage conversion function and + +$$ +\gamma (k) = 1 - \exp \left(- \frac {\Delta t}{\tau (k)}\right) \tag {2} +$$ + +is the time-varying coefficient of the LPF, where $\Delta t$ is the time interval between two indices. The photocurrent-dependent time constant $\tau (k)$ can be modeled as [16, 26]: + +$$ +\tau (k) = \tau_ {0} + \frac {\alpha}{i _ {\mathrm {p d}} (k)}, \tag {3} +$$ + +where $\tau_0$ and $\alpha$ are pixel parameters. The pixel latency in module $\pm 2\mathrm{v}$ due to the LPF behavior is called " $V_{\mathrm{FE}}$ delay". As illustrated in Figure 2(b), an event is triggered at time $t_\mathrm{in}$ once the difference between $V_{\mathrm{FE}}$ and reference voltage $V_{\mathrm{ref}}$ reaches the triggering threshold voltage: + +$$ +\left\{ \begin{array}{l l} V _ {\mathrm {F E}} \left(t _ {\mathrm {i n}}\right) - V _ {\mathrm {r e f}} \left(t _ {\mathrm {i n}}\right) & \geq c _ {\mathrm {u p}} \quad \text {f o r p o s i t i v e e v e n t s} \\ V _ {\mathrm {F E}} \left(t _ {\mathrm {i n}}\right) - V _ {\mathrm {r e f}} \left(t _ {\mathrm {i n}}\right) & \leq c _ {\mathrm {d o w n}} \quad \text {f o r n e g a t i v e e v e n t s} \end{array} , \right. \tag {4} +$$ + +where $c_{\mathrm{up / down}}$ denotes the contrast threshold voltage. The in-pixel timestamp $t_{\mathrm{in}}$ is recorded when the event fires, after which the pixel experiences "readout latency $t_{\mathrm{rl}}$ " and "refractory period $t_{\mathrm{rp}}$ ". The $V_{\mathrm{ref}}$ is updated to follow $V_{\mathrm{FE}}$ after the refractory period. Finally, we have a 5-dimensional output $\{x,y,t_{\mathrm{in}},t_{\mathrm{off}},p\}$ , which contains row and column indices, in-pixel and off-pixel timestamps, and polarity. + +Pixel-wise Photocurrent Estimation Suppose that the illuminance on the sensor plane results in photocurrent $i_{\mathrm{pd}}$ for a single pixel. We aim to estimate $i_{\mathrm{pd}}$ when an event is triggered to avoid the impact of pixel latency in the $\pm 2\mathrm{v}$ + +module. Based on Figure 2(a), we have the following CIS measurement equation: + +$$ +D N = G \int_ {t _ {\mathrm {s}}} ^ {t _ {\mathrm {e}}} i _ {\mathrm {p d}} (t) \mathrm {d} t, \tag {5} +$$ + +where $DN$ is the CIS intensity frame measured in digital numbers; $G$ indicates the total gain of charge-voltage conversion and ADC; $t_{\mathrm{s}}$ and $t_{\mathrm{e}}$ denote the CIS exposure start and end, respectively. As for EVS measurements, suppose we have an event sequence with $N$ events $\{x(k),y(k),t_{\mathrm{in}}(k),t_{\mathrm{off}}(k),p(k)\}$ . By assuming $V_{\mathrm{FE}}(1) = V_{\mathrm{DC}}(1)$ , with Eq. (1) we have the recurrence relation: + +$$ +V _ {\mathrm {F E}} (1) = f _ {\mathrm {D C}} \left(i _ {\mathrm {p d}} (1)\right) \tag {6} +$$ + +$$ +V _ {\mathrm {F E}} (k + 1) = f _ {\gamma} \left(i _ {\mathrm {p d}} (k)\right) \left(f _ {\mathrm {D C}} \left(i _ {\mathrm {p d}} (k + 1)\right) - V _ {\mathrm {F E}} (k)\right), \tag {7} +$$ + +where $k \in \{1, \dots, N-1\}$ and $f_{\gamma}(\cdot)$ denotes the function that maps $i_{\mathrm{pd}}$ onto $\gamma$ . By assuming $V_{\mathrm{ref}}(k+1) \approx V_{\mathrm{FE}}(k)$ , the EVS measurements provide additional $N-1$ equations as: + +$$ +c (k + 1) \approx V _ {\mathrm {F E}} (k + 1) - V _ {\mathrm {F E}} (k), \tag {8} +$$ + +where $c(k) \in \{c_{\mathrm{up}}, c_{\mathrm{down}}\}$ are the known triggering thresholds that are configurable. These state transition equations and measurement equations (6)-(8) can be expressed as a system of $2N - 1$ nonlinear equations and $2N$ unknowns $-i_{\mathrm{pd}}(k)$ and $V_{\mathrm{FE}}(k)$ . By using a zero-order holder, the discrete states $i_{\mathrm{pd}}(k)$ and $V_{\mathrm{FE}}(k)$ can be approximated as continuous signals $i_{\mathrm{pd}}(t)$ and $V_{\mathrm{FE}}(t)$ . This makes it possible to utilize CIS measurements in Eq. (5) as additional constraints. In conclusion, the pixel latency compensation can be expressed as pixel-wise optimization problem, solving the nonlinear system with $2N$ unknown states and $2N + M - 1$ equations, where $M$ and $N$ denote the number of CIS and EVS measurements, respectively. + +Modeling Readout Latency and Refractory Period The EVS measurement model equation (8) is an approximation since the state variation during the readout latency and the refractory period is not considered. Using both in-pixel timestamps and off-pixel timestamps, it is possible to compensate for some of the "missed" $V_{\mathrm{FE}}$ change by assuming that the slope of $V_{\mathrm{FE}}$ is locally constant for a short period. As illustrated in Figure 2(b), it is noted that: + +$$ +m (k + 1) = \frac {c (k + 1)}{t _ {\mathrm {i n}} (k + 1) - t _ {\mathrm {o f f}} (k) - t _ {\mathrm {r p}}}, \tag {9} +$$ + +where $m$ describes the slope approximation of the $V_{\mathrm{FE}}$ curve. For the readout latency, the missed $V_{\mathrm{FE}}$ is represented as: + +$$ +\Delta V _ {\mathrm {F E}} ^ {\mathrm {r l}} (k) = m (k + 1) \left[ t _ {\text {o f f}} (k) - t _ {\text {i n}} (k) \right]. \tag {10} +$$ + +For the refractory period, we have $\Delta V_{\mathrm{FE}}^{\mathrm{rp}}(k) = m(k + 1)t_{\mathrm{rp}}$ since $t_\mathrm{rp}$ is a fixed parameter. Thus the reference voltage after compensation can be adjusted to: + +$$ +V _ {\text {r e f}} (k + 1) = V _ {\mathrm {F E}} (k) + \Delta V _ {\mathrm {F E}} ^ {\mathrm {r l}} (k) + \Delta V _ {\mathrm {F E}} ^ {\mathrm {r p}} (k). \tag {11} +$$ + +![](images/8fd2a7b91b0979a19d54ffcd360ed1de72ee3a74f991f18b974e821b8c1bda04.jpg) +Figure 3. Block diagram of the proposed method. Note that after the frame reconstruction the blurry texture due to motion is enhanced and the RS effect is removed. The noise and artifacts are further reduced by the refinement network. + +By substituting the compensation equation (11) into (4), the impact of readout latency and refractory period can be compensated: + +$c(k + 1)\approx V_{\mathrm{FE}}(k + 1) - V_{\mathrm{FE}}(k) - \Delta V_{\mathrm{FE}}^{\mathrm{sl}}(k) - \Delta V_{\mathrm{FE}}^{\mathrm{rp}}(k),$ (12) where $\Delta V_{\mathrm{FE}}^{\mathrm{rl}}(k)$ and $\Delta V_{\mathrm{FE}}^{\mathrm{rp}}(k)$ are voltage compensation for readout latency and refractory period, respectively. Note that this compensation is enabled by the availability of in-pixel and off-pixel timestamps. + +# 4. Method + +The proposed frame enhancement framework is presented in Figure 3. The frame reconstruction module jointly deblurs images, removes the RS effect, and interpolates high frame rate images. The refinement network focuses on removing noise and improving image quality. + +# 4.1. Solving the Inverse Model via Optimization + +Graph Modeling for a Single Pixel Due to the presence of pixel latency, the system is governed by nonlinear equations, rendering the models and solutions designed for linear systems inapplicable. Graph optimization [13] provides a straightforward way to model nonlinear systems, where the states and measurements are represented as nodes and edges in a graph as shown in Figure 4(a). Let $l \in \{1, \dots, M\}$ and $k \in \{1, \dots, N-1\}$ be the indices of CIS and EVS, respectively. By defining states $\boldsymbol{x} = [i_{\mathrm{pd}}(k)]^{\top}$ and measurements $z_{\mathrm{CIS}} = [DN(l)]^{\top}$ , $z_{\mathrm{EVS}} = [c(k+1)]^{\top}$ , the pixel-wise graph optimization framework aims to determine the state vector $\boldsymbol{x}$ : + +$$ +\boldsymbol {x} ^ {*} = \underset {\boldsymbol {x}} {\operatorname {a r g m i n}} \left(\boldsymbol {e} _ {\mathrm {C I S}} ^ {\top} \boldsymbol {\Omega} _ {\mathrm {C I S}} \boldsymbol {e} _ {\mathrm {C I S}} + \boldsymbol {e} _ {\mathrm {E V S}} ^ {\top} \boldsymbol {\Omega} _ {\mathrm {E V S}} \boldsymbol {e} _ {\mathrm {E V S}}\right) \tag {13} +$$ + +where $e_{\mathrm{CIS}} = z_{\mathrm{CIS}} - \tilde{z}_{\mathrm{CIS}}$ , $e_{\mathrm{EVS}} = z_{\mathrm{EVS}} - \tilde{z}_{\mathrm{EVS}}$ denote measurement error vectors; $\tilde{z}_{\mathrm{CIS}}$ and $\tilde{z}_{\mathrm{EVS}}$ are predicted measurements which can be calculated from estimated states according to CIS measurement equation (5) and EVS mea + +![](images/77e4328e7b85c58d782e23bb33abec072591107f7992dbc86e7c20e55805411f.jpg) + +![](images/cb62d3b02e5fa2c6832a628e5522cd82486490e12ccbfcbc8ebc60eb1bd21f54.jpg) +Figure 4. (a) Graph modeling of the pixel-wise photocurrent estimation problem, where nodes and edges denote states and measurements. The arrows indicate the order of state calculation. "zCIS" and "zEVS" indicate CIS and EVS measurements, respectively. Eqn. (8) is replaced with Eqn. (12) when voltage compensation for readout latency and refractory period is enabled. (b) Row-based reconstruction scheme showing three CIS images as input. After row-wise computation, interpolated and deblurred frames without RS effect can be generated as shown in the blue box. The proposed scheme can be applied to video enhancement with any number of CIS images. + +surement equation (8); $z_{\mathrm{EVS}}$ is adjusted using register settings and calibrated using the "S-curve method" [7]. The statistical distribution of the threshold is not explicitly modeled in the current method. The diagonal weighing matrices $\Omega_{\mathrm{CIS}}$ and $\Omega_{\mathrm{EVS}}$ reflect the reliability of CIS and EVS, respectively. Using the cost function equation (13), the $V_{\mathrm{FE}}$ delay is modeled explicitly through the EVS measurement equations. Solving the optimization problem leads to state estimation such that the $V_{\mathrm{FE}}$ delay is compensated. Many approaches have been proposed to solve nonlinear optimization problems iteratively [19]. We parallelize pixelwise tasks as there is no data interaction or computational dependency between pixels. + +Rolling Shutter (RS) Effect Correction As CIS parameters such as row exposure time and RS scanning speed are known, we formulate the frame estimation problem as pixel-wise computation using adjusted CIS timing parameters for each row. After solving the photocurrent at each event time-step through optimization, a zero-order hold is used to generate a time-continuous photocurrent $i_{\mathrm{pd}}(t)$ for all pixels. The CIS measurement equation (5) is used to generate a rolling shutter artifact-free target frame rate (such + +![](images/a941213ff7db73c7ecfd1ea401ec15d50529b2461cb01d3bfcfcbec8f943f606.jpg) +Figure 5. Modified NAF block in the refinement network. + +as 10 000 FPS). + +Video Glitch Effect Deblurring and video frame interpolation performance can experience glitches in case these are computed on a frame-by-frame basis. To mitigate this issue, we solve the coupled deblurring, rolling-shutter correction and interpolation problem jointly for subsequent frames utilizing the row-specific rolling shutter timing characteristics as shown in Figure 4(b), where we reconstruct each row according to its specific exposure time by fusing two consecutive CIS frames and their corresponding events. + +# 4.2. Refinement Network + +A refinement network is developed as a post-processing module to further mitigate the noise and artifacts in the reconstructed frames. NAFNet [4] is designed for image restoration. It is structured as a U-Net, arranging NAF blocks to allow the network to progressively downsample the input for semantic feature extraction and then progressively upsample it to restore the output. NAFNet provides additional skip connections that preserve information and allow feature reusability from earlier layers. NAF blocks are structured as original ResNet blocks, with additional layers such as layer normalization and channel attention. We adopt the NAFNet basic concept, and further modify making two major additions: 1) spatial attention [40] as shown in Figure 5 and 2) perceptual loss [32]. The reason for choosing spatial attention is that most of the noise is located in motion areas instead of being ubiquitous across the image. Spatial attention allows the network to focus on these noisy regions of the image, which contributes to better denoising results. Moreover, the rationale for choosing the perceptual loss is that even though the proposed network does well in removing the pixel-level saturation noise, it also introduces blurriness or texture degradation. Perceptual loss helps to recover the image quality in a way that better aligns with the human vision system. The losses applied in this paper are summarized as: + +$$ +\mathcal {L} _ {\mathrm {s}} = w _ {l} \mathcal {L} _ {1} + w _ {\mathrm {V G G}} \mathcal {L} _ {\mathrm {V G G}} + w _ {\text {G r a m}} \mathcal {L} _ {\text {G r a m}}, \tag {14} +$$ + +where the $\mathcal{L}_1$ loss is the pixel-wise difference between the reconstructed RGB image and ground truth image; the $\mathcal{L}_{\mathrm{VGG}}$ loss is calculated as $\mathcal{L}_{\mathrm{VGG}} = \frac{1}{L}\sum_{l=1}^{L}\alpha_l\|\Psi_l(\boldsymbol{I}_{\mathrm{Recon}}) - \Psi_l(\boldsymbol{I}_{\mathrm{GT}})\|_1;\Psi_l(\boldsymbol{I})$ is the feature tensor from the $l_{\mathrm{th}}$ layer of an ImageNet pre-trained VGG-19 network generated by passing the image $\boldsymbol{I}$ through the network. In this case, $L$ specifies the total number of layers in the VGG-19 network, while $\alpha_l$ is the hyperparameter assigned to each layer. Finally, the Gram loss [11] is computed as the L2 difference between the autocorrelation of the VGG-19 features of the reconstructed image and ground truth image with the formulation: $\mathcal{L}_{\mathrm{Gram}} = \frac{1}{L}\sum_{l=1}^{L}\alpha_l\|\boldsymbol{M}_l(\boldsymbol{I}_{\mathrm{Recon}}) - \boldsymbol{M}_l(\boldsymbol{I}_{\mathrm{GT}})\|_2$ , where given an image $I$ , the Gram matrix is computed for every layer $l$ as $\boldsymbol{M}_l = (\Psi_l(\boldsymbol{I}))^\top(\Psi_l(\boldsymbol{I}))$ . + +# 5. Experiments + +We compare our proposed method with state-of-the-art EVS+CIS video reconstruction methods: EDI [27], AKF [39], Time Lens [37], CBMNet [23], EvUnroll [8] and RE-FID [36]. Among them, EvUnroll proposes an RS correction based reconstruction and REFID combines image de-burring with EVS based frame interpolation. We also show the results of the CIS-only interpolation method FILM [32] as reference. Before optimization, $i_{\mathrm{pd}}$ are initialized according to the first CIS image DN and the inverse photocurrent-DN relation. The elements in the weighing matrices for CIS, EVS terms are set to 1. The performance of the interpolation methods are evaluated both quantitatively and qualitatively, based on the metrics of Peak Signal-to-Noise Ratio (PSNR), LPIPS [42], and pixel-wise reconstruction error on DN. Furthermore, to study the performance degradation concerning ambient illuminance and motion speed, we measured a test set utilizing a rotating disk in a controlled environment. We analyze the Blurred Edge Width (BEW) [5] of the reconstructed frames - a larger BEW indicates a blurrier image and, therefore, a worse deblurring result. + +# 5.1. Datasets + +We evaluate the proposed method using the simulation data based on the CIS-EVS hybrid sensor simulator [26] which considers $V_{\mathrm{FE}}$ delay, readout latency, and refractory period. Our measurement data was collected from the hybrid CIS-EVS sensor [14, 15]. Here, only off-pixel times- tamp data was captured. Two measured datasets were created for evaluation: The natural scene dataset focuses on indoor scenarios where EVS is significantly influenced by pixel latency. The other dataset captured a rotating disk with a Siemens star pattern under different ambient illumination levels and rotating speeds. In addition to the above datasets specifically captured for this work's evaluation, we also test our method using the publicly available HS-ERGB + +Table 1. Quantitative comparison of state-of-the-art video reconstruction methods on the proposed simulation dataset in terms of PSNR (in dB, higher is better) and LPIPS (a dimensionless quantity, lower is better). Each row shows results for a particular scene. Results from different methods are listed column-wise. The first and second places are highlighted with bold underline and bold, respectively. + +
Ours w/o refinementOursEDIAKFTime LensFILMCBMNetEvUnrollREFID
ScenePSNR ↑LPIPS ↓PSNR ↑LPIPS ↓PSNR ↑LPIPS ↓PSNR ↑LPIPS ↓PSNR ↑LPIPS ↓PSNR ↑LPIPS ↓PSNR ↑LPIPS ↓PSNR ↑LPIPS ↓PSNR ↑LPIPS ↓
basketball24.010.21727.230.15522.490.22015.630.28023.650.24023.860.22024.510.25223.130.21822.380.223
checkerboard24.230.20328.240.18720.480.21819.560.26122.600.23923.020.21622.760.21223.840.20421.230.235
slingshot egg25.230.10032.850.10624.950.16415.590.11025.190.24424.040.20322.760.23424.610.21023.920.201
running man24.600.16424.720.18922.290.13520.710.18925.010.11725.060.08925.160.06923.040.11723.520.163
fan22.400.18924.500.16318.170.25510.780.18420.640.23318.930.18421.280.20419.640.19018.300.198
Average24.090.17527.510.1621.680.19816.450.20523.420.21522.980.18223.290.19422.850.18821.870.204
+ +Table 2. Quantitative results of ablation study. The first and second places are highlighted with bold underline and bold, respectively. + +
without compensationwith VFE delay compensationwith RL+RP compensationwith VFE+RL+RP compensationwith VFE+RL+RP compensation and refinement
ScenePSNR ↑LPIPS ↓PSNR ↑LPIPS ↓PSNR ↑LPIPS ↓PSNR ↑LPIPS ↓PSNR ↑LPIPS ↓
basketball22.950.31623.470.30623.500.30723.560.30223.640.295
slingshot egg23.570.09924.000.09923.600.09923.710.10229.040.106
+ +![](images/29969f719a04c87b0e6c3220e145fad3cf4ee891447d76451681b3ebf152ada7.jpg) +Figure 6. Qualitative comparison of state-of-the-art video reconstruction methods on the proposed measured datasets in terms of image quality. The images have been magnified for optimal viewing quality. + +![](images/8f9a77d30aee6659949c823f28d806ae217c9ca3960f79364c9fe5d1f26a5d6f.jpg) +Figure 7. Frame-wise, quantitative comparison of state-of-the-art video reconstruction methods in terms of PSNR for varying scenarios of the simulation dataset. Note that some algorithms exhibit pronounced discontinuities in the observed PSNR curves indicating the presence of video glitches. + +![](images/40d56ba35ed8aa577db3383cf57a6cfa73d930d79242859c7d50e167ac2cb6b1.jpg) + +![](images/918086ff77003e143e0ae0536c41b0aa403cd75bed8fcf0e47d481014d414e61.jpg) + +![](images/1ec2efe781dc867223e22b3a398dd3345f5e0f4b5f5fdb6dffb6aefcf94ffa80.jpg) + +![](images/7628439d5201acf790fe3536544619871746a1aa2c0c5fca2e5906fbc7bbce65.jpg) + +![](images/4cb9e1d501fb3607c001cf4c54594e197aa4e1596d7da03ebc296894e1f0ffa9.jpg) + +dataset based on a EVS camera that is collocated next to a CIS camera. Since the paper focuses on mitigating artifacts due to pixel latency, two indoor scenarios ("spinning plate" and "spinning umbrella") were selected. More details of the dataset used are provided in the supplementary material. + +# 5.2. Results + +Image Quality Table 1 shows quantitative results for the reconstructed video. Results in PSNR and LPIPS show that the proposed method outperforms other state-of-the-art methods in most of the scenes, especially with fast motion ("slingshot egg") or complicated rotation motion ("basket- + +ball", "fan"). As expected the refinement network which aims to alleviate artifacts and improve image quality reports a higher score. Snippets of the reconstructed videos are presented in Figure 1 and Figure 6 for simulated and measured data, respectively. Comprehensive input images are provided in the supplementary material. Thanks to the explicit EVS latency modeling and compensation, our results avoid obvious shadowing/ghosting artifacts - especially if there is fast motion under challenging lighting conditions. It is also noted that the proposed method successfully corrects the RS effect, as shown in the "basketball" scene. Learning-based methods face difficulties, as latency and the RS effect have not been considered in model design and training. The CIS-only method FILM provides natural image quality with low noise but fails when the objects are too fast to be captured by CIS, as shown in the "slingshot egg" scene. For the "basketball" scene, our method shows an advantage in details and sharpness compared to FILM. + +In specific scenarios, such as the "spinning umbrella" scene within the HS-ERGB dataset and the "running man" scene in the simulation dataset, our method did not yield improved image quality. This observation can be attributed to several factors. Firstly, the HS-ERGB dataset employs a different EVS with different pixel and readout latency as well as refractory period. Secondly, this dataset is based on a dual camera instead of a hybrid sensor so the spatiotemporal alignment may differ. Thirdly, the scenes in question involve relatively slower object movement speeds resulting in reduced ghosting. In essence, the dynamic characteristics of the EVS, the input signal frequency range, and the reconstruction algorithm influence the quality of the reconstructed images. + +Video Glitch Effect To evaluate the glitch issue, we compute the PSNR for each reconstructed frame as shown in Figure 7. Overall, our method gives temporally more stable PSNR curves, which stem from the proposed row-based reconstruction scheme. It is also noted that the refinement network contributes significantly to higher PSNR. + +Pixel-wise Error Analysis We select a pixel within the region of interest (indicated as the yellow box in Figure 1(b)) and verify the pixel-wise reconstruction accuracy. As shown in Figure 8(a), with the help of pixel latency compensation the reconstructed $V_{\mathrm{DC}}$ after optimization shows a sharper peak (with $250~\mu \mathrm{s}$ duration) compared to $V_{\mathrm{FE}}$ (with over $1\mathrm{ms}$ duration). The high-frequency fluctuations on ground truth $V_{\mathrm{DC}}$ curve cannot be reconstructed due to the $V_{\mathrm{FE}}$ bandwidth limit. An overshoot is observed after the peak resulting in a dark tail along the pellet's trajectory, as shown in Figure 1(b). This may stem from 1) the $V_{\mathrm{FE}}$ bandwidth, 2) the approximation error in equation (12), 3) EVS measurement noise, and 4) choice of weighing matrices for CIS and EVS. Figure 8(b), indicates that the proposed method gives an excellent reconstruction of the pel + +![](images/e59a31c0c039cf129bfcd2d939a0cbca7716bc9bbbda3418fca6d19558c19005.jpg) +(a) + +![](images/0903f66bcfe91b0e67ca6280b8f589197a8c8e9f970927d46cdce5238e0c2d70.jpg) +(b) +time $(\mu s)$ + +![](images/5f28b27d9ca784fa765a3b8d41656f9cc8c4706ccc763c11eeca66738fd940fb.jpg) +Figure 8. Pixel-wise reconstruction results for the "slingshot egg" scene. The selected pixel is located at the path of the pellet, in the yellow box of Figure 1. The refinement network is not used here. (a) Comparison of the ground truth $V_{\mathrm{DC}}$ , reconstructed $V_{\mathrm{FE}}$ and $V_{\mathrm{DC}}$ , where the ground truth $V_{\mathrm{DC}}$ is computed from the ground truth DN of the pixel at 10000 FPS. (b) Comparison of reconstructed pixel-wise DN of state-of-the-art video reconstruction methods. +(a) with no compensation + +![](images/9623e737ab247512b849601955966bf3cd145117bc583e92d9b67a0abba37a04.jpg) +(b) with $V_{\mathrm{FE}}$ delay compensation + +![](images/84c86a856507df835cd2d02721b030013398fa71167135634dc9ac8c1753bd57.jpg) +(c) with RL+RP compensation +Figure 9. Qualitative results of ablation study on the "basketball" scene. Refer to Figure 1(a) for "with $V_{\mathrm{FE}} + \mathrm{RL} + \mathrm{RP}$ compensation and refinement" result. + +![](images/1d06d331807319f189d35f3097b64f7ee960371b235c277a5d356c98d04ffa1c.jpg) +(d) with $V_{\mathrm{FE}} + \mathrm{RL} + \mathrm{RP}$ compensation + +let's peak DN despite the overshoot. + +Ablation Study We ablate the refinement network, $V_{\mathrm{FE}}$ delay compensation, readout latency (RL) and refractory period (RP) compensation of our method. The numerical evaluation are subject to selected ROIs on motion region. To ablate $V_{\mathrm{FE}}$ delay compensation, we set the time constant $\tau = 0$ such that $V_{\mathrm{FE}} = V_{\mathrm{DC}}$ in Eq. (7). To ablate readout latency and refractory period compensation, we use Eq. (8) instead of Eq. (12) in the EVS measurement model. Results are summarized in Table 2 and Figure 9. The con + +Figuration using all modules gives significantly better visual and slightly better numerical results in PSNR compared to having latency compensation and refinement disabled. Disabling $V_{\mathrm{FE}}$ or RL+RP compensation significantly worsens the image quality, as severe shadowing artifacts are observed. As for LPIPS in "slingshot egg" scene, the difference is not significant due to the small area of pellet ROI. It is also noted that adding RL+RP compensation leads to improvement in the "basketball" but degrades the performance in "slingshot egg" when comparing with $V_{\mathrm{FE}}$ delay compensation only. This is because RL+RP compensation is based on the assumption that the slope of VFE is locally constant for a short period. In cases such as "slingshot egg", the assumption may not hold due to close positive and negative events. An ablation study against the use of NAFNet-only is included in supplementary material. + +Performance Regarding Illuminance and Speed The proposed method is further investigated under a controlled environment using varying illuminance and object motion speed. BEW [5] is used to evaluate the image quality by measuring edge sharpness. Figure 10(a)-(h) presents the comparison of the state-of-the-art methods on a rotating disk under $1000\mathrm{lx}$ and $292\mathrm{rpm}$ . It is observed that the proposed method reconstructs sharper edges compared to the others. Other EVS-based methods struggle resolving blurry edges in presence of fast motion and also the CIS-based FILM method fails to generate a sharp reconstruction. Figure 10(i) shows that the proposed method reports the lowest BEW score (0.7) and has the sharpest edge profile. + +The BEW scores under different speed and illumination conditions are summarized in Figure 11. Figure 11(a) shows that the BEW increases with faster rotation speeds. In Figure 11(b), the lower ambient light condition leads to higher $V_{\mathrm{FE}}$ latency degrading image quality. Our method achieves the lowest BEW throughout all illuminance levels. + +# 6. Conclusion + +An EVS-assisted framework to solve motion blur, rolling-shutter artifacts and video-frame interpolation in a joint manner was presented. Existing methods use either a simple double integral model or a learning-based model, neither of which explicitly address EVS sensor imperfections such as EVS pixel latency, readout latency or refractory period. This work overcomes these limitations by explicitly modeling these characteristics. The compensation of readout latency and refractory period is enabled by having access to in-pixel and off-pixel time-stamps. The nonlinear inverse model was solved through joint graph optimization of subsequent frames utilizing row-by-row exposure information to overcome rolling-shutter artifacts and possible glitches between subsequent frames. A refinement network was proposed to further improve the image quality by using specifically designed spatial attention blocks. We showed that + +![](images/ac5f187dbcf96d40eb3e6c02796a0804c55602294a83345f77bf9c0adebac150.jpg) +(a) Ours w/o refinement + +![](images/7c824604139f204e8ee97040bd04c00e5bf0fe194137654db705377c0490709d.jpg) +(b) EDI + +![](images/0bb38fd3e7a4c604a13814662e2932089c2f12959cde9fbdff3d87f149a3b387.jpg) +(c)AKF + +![](images/5fbd68f799e6321bfcf5b89c9e7134912537c696687e899a5325c2d6dcd14380.jpg) +(d) Time Lens + +![](images/181c11d8c4ef18733787bb7d61eff62d02de1f62ec20e8ef54cd9a125c775dbc.jpg) +(e) FILM + +![](images/8adac60a1e6e2c7288c4b69289e84a65e56b98a53137230ba2e0612898e4638c.jpg) +(f) CBMNet + +![](images/50d88950284ba515c9491230e2725d201856e92e35fdb6f1c99eb6254b195442.jpg) +(g) EvUnroll + +![](images/d58e086e10466bcdf13e7306801568e1f12c80e269db05ce17912a301ffd7e7c.jpg) +(h) REFID + +![](images/15725040120936c38a6ecf6d17f65b6bca40c7c44b7c4d7674a5e04693710be4.jpg) + +![](images/122c30b05db8c9c5cb059e32e470ec0d4ddfa289390133eac089943b667e7cde.jpg) +Figure 10. (a)-(h) Comparison of state-of-the-art video reconstruction methods on the proposed rotating disk datasets with rotating speed $292\mathrm{rpm}$ and environment illuminance $1000\mathrm{lx}$ . The yellow box shows magnified edges. The visual "misalignment" is due to EVS latency. (i) The DN-edge width relation at a fixed segment of 40 pixels (red line in (a)) on reconstructed rotating disk edge. The refinement network is not used. +(a) speed (rpm) +Figure 11. Quantitative comparison of state-of-the-art video reconstruction methods on the proposed "rotating disk" scene in terms of BEW. The lower the BEW values, the sharper the edge. The refinement network is not used. (a) BEW under various rotation speeds with a fixed illuminance of $1000\mathrm{lx}$ . (b) BEW under various illuminance levels with a fixed speed of $292\mathrm{rpm}$ . + +![](images/b906a65537ee2753c868c9c208278975d5f4c076cc05dbed7596616761135e8c.jpg) +(b) illumination (lux) + +the proposed method outperforms state-of-the-art methods in reconstructed image quality with an up to $4\mathrm{dB}$ improvement in PSNR and $12\%$ improvement in LPIPS score. + +Limitation Since the pixel latency highly depends on EVS design parameters, object motion, and ambient illuminance, the image quality improvement after latency compensation may be subtle for slow objects or in bright scenes such as "running man" and "spinning umbrella". + +# References + +[1] Wenbo Bao, Wei-Sheng Lai, Chao Ma, Xiaoyun Zhang, Zhiyong Gao, and Ming-Hsuan Yang. Depth-aware video frame interpolation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3703-3712, 2019. 2 +[2] Patrick Bardow, Andrew J Davison, and Stefan Leutenegger. Simultaneous optical flow and intensity estimation from an event camera. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 884-892, 2016. 2 +[3] Christian Brandli, Raphael Berner, Minhao Yang, Shih-Chii Liu, and Tobi Delbruck. A $240 \times 180$ 130 db $3\mu s$ latency global shutter spatiotemporal vision sensor. IEEE Journal of Solid-State Circuits, 49(10):2333-2341, 2014. 1 +[4] Xiaojie Chu, Liangyu Chen, and Wenqing Yu. Nafssr: Stereo image super-resolution using nafnet. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, pages 1239-1248, 2022. 5 +[5] Hai Dinh, Qinyi Wang, Fangwen Tu, Brett Frymire, and Bo Mu. Evaluation of motion blur image quality in video frame interpolation. *Electronic Imaging*, 35:262-1, 2023. 5, 8 +[6] Julius Erbach, Stepan Tulyakov, Patricia Vitoria, Alfredo Bochicchio, and Yuanyou Li. Evshutter: Transforming events for unconstrained rolling shutter correction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13904-13913, 2023. 2 +[7] C.Posch et al. Sensitivity and uniformity of a $0.18\mu \mathrm{m}$ CMOS temporal contrast pixel array. In ISCAS, 2011. 4 +[8] Xinyu Zhou et al. Evunroll: Neuromorphic events based rolling shutter image correction. In CVPR, 2022. 2, 5 +[9] Guillermo Gallego, Tobi Delbrück, Garrick Orchard, Chiara Bartolozzi, Brian Taba, Andrea Censi, Stefan Leutenegger, Andrew J Davison, Jörg Conradt, Kostas Daniilidis, et al. Event-based vision: A survey. IEEE transactions on pattern analysis and machine intelligence, 44(1):154-180, 2020. 1 +[10] Yue Gao, Siqi Li, Yipeng Li, Yandong Guo, and Qionghai Dai. Superfast: $200 \times$ video frame interpolation via event camera. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2022. 2 +[11] Leon A Gatys, Alexander S Ecker, and Matthias Bethge. Image style transfer using convolutional neural networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2414-2423, 2016. 5 +[12] Daniel Gehrig, Antonio Loquercio, Konstantinos G Derpanis, and Davide Scaramuzza. End-to-end learning of representations for asynchronous event-based data. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5633-5643, 2019. 2 +[13] Giorgio Grisetti, Rainer Kümmerle, Hauke Strasdat, and Kurt Konolige. g2o: A general framework for (hyper) graph optimization. In Proceedings of the IEEE International Conference on Robotics and Automation (ICRA), pages 9-13, 2011. 4 +[14] Menghan Guo, Shoushun Chen, Zhe Gao, Wenlei Yang, Peter Bartkovjak, Qing Qin, Xiaqin Hu, Dahai Zhou, Qiping Huang, Masayuki Uchiyama, Yoshiharu Kudo, Shimpei + +Fukuoka, Chengcheng Xu, Hiroaki Ebihara, Xueqing Wang, Peiwen Jiang, Bo Jiang, Bo Mu, Huan Chen, Jason Yang, T. J. Dai, and Andreas Suess. A three-wafer-stacked hybrid 15-mpixel cis + 1-mpixel evs with 4.6-GeV/s readout, inpixel tdc, and on-chip isp and esp function. IEEE Journal of Solid-State Circuits, pages 1–10, 2023. 1, 2, 5 +[15] Menghan Guo, Shoushun Chen, Zhe Gao, Wenlei Yang, Peter Bartkovjak, Qing Qin, Xiaqin Hu, Dahei Zhou, Masayuki Uchiyama, Yoshiharu Kudo, Shimpei Fukuoka, Chengcheng Xu, Hiroaki Ebihara, Andy Wang, Peiwen Jiang, Bo Jiang, Bo Mu, Huan Chen, Jason Yang, TJ Dai, and Andreas Suess. A 3-wafer-stacked hybrid 15MPixel CIS + 1MPixel EVS with 4.6GEvent/s readout, in-pixel TDC and on-chip ISP and ESP function. In 2022 IEEE International Solid-State Circuits Conference (ISSCC), 2023. 2, 5 +[16] Yuhuang Hu, Shih-Chii Liu, and Tobi Delbruck. v2e: From video frames to realistic dvs events. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1312-1321, 2021. 3 +[17] Zhewei Huang, Tianyuan Zhang, Wen Heng, Boxin Shi, and Shuchang Zhou. Real-time intermediate flow estimation for video frame interpolation. In European Conference on Computer Vision, pages 624-642. Springer, 2022. 2 +[18] Huaizu Jiang, Deqing Sun, Varun Jampani, Ming-Hsuan Yang, Erik Learned-Miller, and Jan Kautz. Super slomo: High quality estimation of multiple intermediate frames for video interpolation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 9000-9008, 2018. 2 +[19] Nocedal Jorge and J Wright Stephen. Numerical optimization. Springer, 2006. 4 +[20] Damien Joubert, Mathieu Hébert, Hubert Konik, and Christophe Lavergne. Characterization setup for event-based imagers applied to modulated light signal detection. Applied optics, 58(6):1305-1317, 2019. 2 +[21] Damien Joubert, Alexandre Marcireau, Nic Ralph, Andrew Jolley, André van Schaik, and Gregory Cohen. Event camera simulator improvements via characterized parameters. Frontiers in Neuroscience, page 910, 2021. 1, 2 +[22] Hanme Kim, Stefan Leutenegger, and Andrew J Davison. Real-time 3d reconstruction and 6-dof tracking with an event camera. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part VI 14, pages 349-364. Springer, 2016. 2 +[23] Taewoo Kim, Yujeong Chae, Hyun-Kurl Jang, and Kuk-Jin Yoon. Event-based video frame interpolation with cross-modal asymmetric bidirectional motion fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 18032-18042, 2023. 1, 2, 5 +[24] Kazutoshi Kodama, Yusuke Sato, Yuhi Yorikado, Raphael Berner, Kyoji Mizoguchi, Takahiro Miyazaki, Masahiro Tsukamoto, Yoshihisa Matoba, Hirotaka Shinozaki, Atsumi Niwa, et al. 1.22 $\mu$ m 35.6 mpixel rgb hybrid event-based vision sensor with 4.88 $\mu$ m-pitch event pixels and up to 10k event frame rate by adaptive control on event sparsity. + +In 2023 IEEE International Solid-State Circuits Conference (ISSCC), pages 92-94. IEEE, 2023. 1 +[25] Patrick Lichtsteiner, Christoph Posch, and Tobi Delbruck. A $128 \times 128$ 120 db $15\mu s$ latency asynchronous temporal contrast vision sensor. IEEE Journal of Solid-State Circuits, 43 (2):566-576, 2008. 1, 2 +[26] Xiaozheng Mou, Kaijun Feng, Alex Yi, Steve Wang, Huan Chen, Xiaoqin Hu, Menghan Guo, Shoushun Chen, and Andreas Suess. Accurate event simulation using high-speed videos. *Electronic Imaging*, 34(7):242-1, 2022. 2, 3, 5 +[27] Liyuan Pan, Cedric Scheerlinck, Xin Yu, Richard Hartley, Miaomiao Liu, and Yuchao Dai. Bringing a blurry frame alive at high frame-rate with an event camera. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6820-6829, 2019. 1, 2, 5 +[28] Liyuan Pan, Richard Hartley, Cedric Scheerlinck, Miaomiao Liu, Xin Yu, and Yuchao Dai. High frame rate video reconstruction based on an event camera. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(5):2519-2533, 2020. 2 +[29] Federico Paredes-Valles and Guido CHE de Croon. Back to event basics: Self-supervised learning of image reconstruction for event cameras via photometric constancy. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3446–3455, 2021. 2 +[30] Christoph Posch and Daniel Matolin. Sensitivity and uniformity of a $0.18\mu \mathrm{m}$ cmos temporal contrast pixel array. In 2011 IEEE International Symposium of Circuits and Systems (ISCAS), pages 1572-1575. IEEE, 2011. 2 +[31] Henri Rebecq, René Ranftl, Vladlen Koltun, and Davide Scaramuzza. High speed and high dynamic range video with an event camera. IEEE transactions on pattern analysis and machine intelligence, 43(6):1964-1980, 2019. 2 +[32] Fitsum Reda, Janne Kontkanen, Eric Tabellion, Deqing Sun, Caroline Pantofaru, and Brian Curless. Film: Frame interpolation for large motion. In European Conference on Computer Vision, pages 250-266. Springer, 2022. 1, 2, 5 +[33] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical Image Computing and Computer-Assisted Intervention-MICCAI 2015: 18th International Conference, Munich, Germany, October 5-9, 2015, Proceedings, Part III 18, pages 234-241. Springer, 2015. 2 +[34] Cedric Scheerlinck, Nick Barnes, and Robert Mahony. Continuous-time intensity estimation using event cameras. In Asian Conference on Computer Vision, pages 308-324. Springer, 2018. 1, 2 +[35] Andreas Suess, Menghan Guo, Rui Jiang, Xiaozheng Mou, Qiping Huang, Wenlei Yang, and Shoushun Chen. Physical modeling and parameter extraction for event-based vision sensors. In Proc. IISS Int. Image Sensor Workshop (IISW), Edinburgh, UK, page R5, 2023. 2 +[36] Lei Sun, Christos Sakaridis, Jingyun Liang, Peng Sun, Jiezhang Cao, Kai Zhang, Qi Jiang, Kaiwei Wang, and Luc Van Gool. Event-based frame interpolation with ad-hoc deblurring. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18043-18052, 2023. 1, 2, 5 + +[37] Stepan Tulyakov, Daniel Gehrig, Stamatios Georgoulis, Julius Erbach, Mathias Gehrig, Yuanyou Li, and Davide Scaramuzza. Time lens: Event-based video frame interpolation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16155-16164, 2021. 1, 2, 5 +[38] Stepan Tulyakov, Alfredo Bochicchio, Daniel Gehrig, Stamatios Georgoulis, Yuanyou Li, and Davide Scaramuzza. Time lens: Event-based frame interpolation with parametric non-linear flow and multi-scale fusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17755-17764, 2022. 2 +[39] Ziwei Wang, Yonhon Ng, Cedric Scheerlinck, and Robert Mahony. An asynchronous kalman filter for hybrid event cameras. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 448-457, 2021. 1, 2, 5 +[40] Sanghyun Woo, Jongchan Park, Joon-Young Lee, and In So Kweon. Cbam: Convolutional block attention module. In Proceedings of the European conference on computer vision (ECCV), pages 3–19, 2018. 5 +[41] Zeyu Xiao, Wenming Weng, Yueyi Zhang, and Zhiwei Xiong. Eva2: Event-assisted video frame interpolation via cross-modal alignment and aggregation. IEEE Transactions on Computational Imaging, 8:1145–1158, 2022. 2 +[42] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 5 +[43] Xiang Zhang and Lei Yu. Unifying motion deblurring and frame interpolation with events. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17765-17774, 2022. 2 +[44] Xinyu Zhou, Peiqi Duan, Yi Ma, and Boxin Shi. Evunroll: Neuromorphic events based rolling shutter image correction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17775-17784, 2022. 1, 2 \ No newline at end of file diff --git a/2024/EVS-assisted Joint Deblurring Rolling-Shutter Correction and Video Frame Interpolation through Sensor Inverse Modeling/images.zip b/2024/EVS-assisted Joint Deblurring Rolling-Shutter Correction and Video Frame Interpolation through Sensor Inverse Modeling/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..4806cbb45ede2eafa3511347530024a0e2edfa97 --- /dev/null +++ b/2024/EVS-assisted Joint Deblurring Rolling-Shutter Correction and Video Frame Interpolation through Sensor Inverse Modeling/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:caec237cc5bc977d9ffa6a8dea75614925b001c2f7f51ee269b19ab6af2a7d36 +size 724667 diff --git a/2024/EVS-assisted Joint Deblurring Rolling-Shutter Correction and Video Frame Interpolation through Sensor Inverse Modeling/layout.json b/2024/EVS-assisted Joint Deblurring Rolling-Shutter Correction and Video Frame Interpolation through Sensor Inverse Modeling/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..3527e12945407d1b15e47fb48f6605eea1476676 --- /dev/null +++ b/2024/EVS-assisted Joint Deblurring Rolling-Shutter Correction and Video Frame Interpolation through Sensor Inverse Modeling/layout.json @@ -0,0 +1,10066 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 62, + 102, + 533, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 102, + 533, + 140 + ], + "spans": [ + { + "bbox": [ + 62, + 102, + 533, + 140 + ], + "type": "text", + "content": "EVS-assisted joint Deblurring, Rolling-Shutter Correction and Video Frame Interpolation through Sensor Inverse Modeling" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 85, + 161, + 517, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 85, + 161, + 517, + 190 + ], + "spans": [ + { + "bbox": [ + 85, + 161, + 517, + 190 + ], + "type": "text", + "content": "Rui Jiang*, Fangwen Tu*, Yixuan Long, Aabhaas Vaish, Bowen Zhou, Qinyi Wang Wei Zhang, Yuntan Fang, Luis Eduardo Garcia Capel, Bo Mu, Tiejun Dai, Andreas Suess" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 265, + 194, + 328, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 265, + 194, + 328, + 205 + ], + "spans": [ + { + "bbox": [ + 265, + 194, + 328, + 205 + ], + "type": "text", + "content": "OMNIVISION" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 260, + 209, + 333, + 220 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 260, + 209, + 333, + 220 + ], + "spans": [ + { + "bbox": [ + 260, + 209, + 333, + 220 + ], + "type": "text", + "content": "*Equal contribution" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 50, + 248, + 545, + 323 + ], + "blocks": [ + { + "bbox": [ + 50, + 248, + 545, + 323 + ], + "lines": [ + { + "bbox": [ + 50, + 248, + 545, + 323 + ], + "spans": [ + { + "bbox": [ + 50, + 248, + 545, + 323 + ], + "type": "image", + "image_path": "0a6598df303e0632b86ab8d92416d59c00dbbc089ffe296a5c611f8025de048d.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 331, + 547, + 376 + ], + "lines": [ + { + "bbox": [ + 46, + 331, + 547, + 376 + ], + "spans": [ + { + "bbox": [ + 46, + 331, + 547, + 376 + ], + "type": "text", + "content": "Figure 1. Qualitative results comparing the proposed method with EDI [27], AKF [39], Time Lens [37], FILM [32], CBMNet [23], EvUnroll [44], and REFID [36] on simulation datasets. In the \"basketball\" scene, our method provides frames that are rolling shutter effect corrected and deblurred. In the \"slingshot egg\" scene, our method can reconstruct the shape of the pellet with the best quality, as indicated in the yellow box. The images have been magnified for optimal viewing quality." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 143, + 386, + 192, + 398 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 386, + 192, + 398 + ], + "spans": [ + { + "bbox": [ + 143, + 386, + 192, + 398 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 45, + 411, + 290, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 411, + 290, + 662 + ], + "spans": [ + { + "bbox": [ + 45, + 411, + 290, + 662 + ], + "type": "text", + "content": "Event-based Vision Sensors (EVS) gain popularity in enhancing CMOS Image Sensor (CIS) video capture. Nonidealities of EVS such as pixel or readout latency can significantly influence the quality of the enhanced images and warrant dedicated consideration in the design of fusion algorithms. A novel approach for jointly computing deblurred, rolling-shutter artifact corrected high-speed videos with frame rates up to 10 000 FPS using inherently blurry rolling shutter CIS frames of 120 FPS to 150 FPS in conjunction with EVS data from a hybrid CIS-EVS sensor is presented. EVS pixel latency, readout latency and the sensor's refractory period are explicitly incorporated into the measurement model. This inverse function problem is solved on a per-pixel manner using an optimization-based framework. The interpolated images are subsequently processed by a novel refinement network. The proposed method is evaluated using simulated and measured datasets, under natural and controlled environments. Extensive experiments show reduced shadowing effect, a 4 dB increment in PSNR, and a " + }, + { + "bbox": [ + 45, + 411, + 290, + 662 + ], + "type": "inline_equation", + "content": "12\\%" + }, + { + "bbox": [ + 45, + 411, + 290, + 662 + ], + "type": "text", + "content": " improvement in LPIPS score compared to state-of-the-art methods." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 386, + 386, + 398 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 386, + 386, + 398 + ], + "spans": [ + { + "bbox": [ + 306, + 386, + 386, + 398 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 403, + 546, + 570 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 403, + 546, + 570 + ], + "spans": [ + { + "bbox": [ + 304, + 403, + 546, + 570 + ], + "type": "text", + "content": "Event-Based Vision Sensors (EVS) capture our world from a completely different perspective compared to CMOS Image Sensors (CIS) [9, 14, 24, 25]. In EVS, each pixel independently triggers an event if a relative illuminance change on the pixel reaches a certain triggering threshold (typically in the range of " + }, + { + "bbox": [ + 304, + 403, + 546, + 570 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 304, + 403, + 546, + 570 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 304, + 403, + 546, + 570 + ], + "type": "inline_equation", + "content": "30\\%" + }, + { + "bbox": [ + 304, + 403, + 546, + 570 + ], + "type": "text", + "content": "). Conversely to CIS, EVS operate time-continuously and asynchronously - there is no common exposure period or \"frame rate\". Thus, EVS pixels can be read out comparatively fast in a low-power manner enabling efficient high-speed data capture. Recent work on enhancement of high-quality images [23, 27, 34, 37, 39] focuses on fusing CIS and EVS information so that the high fidelity of CIS measurement and the fast response of EVS measurement can complement each other." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 571, + 547, + 679 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 571, + 547, + 679 + ], + "spans": [ + { + "bbox": [ + 304, + 571, + 547, + 679 + ], + "type": "text", + "content": "In practice it is observed [3, 14, 21, 25] that the dynamic characteristics of EVS pixels are highly dependent on the illumininance in the sensor plane. Under low-light conditions such as indoor capture, a slower EVS response during video enhancement can result in ghosting or blurry frames and needs to be considered in algorithm design. Furthermore, the inevitable readout latency and refractory period of EVS pixels can significantly aggravate artifacts. Current methods are not specifically designed to address sensor nonidealities." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 293, + 731, + 318, + 740 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 731, + 318, + 740 + ], + "spans": [ + { + "bbox": [ + 293, + 731, + 318, + 740 + ], + "type": "text", + "content": "25172" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 239 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 239 + ], + "type": "text", + "content": "This paper presents an inverse model capturing the EVS pixel latency (namely the " + }, + { + "bbox": [ + 46, + 72, + 289, + 239 + ], + "type": "inline_equation", + "content": "V_{\\mathrm{FE}}" + }, + { + "bbox": [ + 46, + 72, + 289, + 239 + ], + "type": "text", + "content": " delay), the readout latency, and the refractory period. The inverse model is solved through nonlinear graph optimization to jointly address deblurring, rolling-shutter (RS) correction and video-frame interpolation while compensating for sensor nonidealities. In addition to the inverse model, a learning-based refinement module is proposed to enhance the image quality by further mitigating the noise and artifacts. The proposed method is compared with existing CIS-based [32] and EVS-assisted [8, 23, 27, 36, 37, 39] video enhancement methods under challenging indoor scenarios with fast motion. As illustrated in Figure 1, the proposed algorithm produces excellent image quality with much fewer artifacts." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 59, + 240, + 276, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 240, + 276, + 251 + ], + "spans": [ + { + "bbox": [ + 59, + 240, + 276, + 251 + ], + "type": "text", + "content": "The contributions of the paper can be summarized as:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 253, + 286, + 371 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 58, + 253, + 286, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 253, + 286, + 300 + ], + "spans": [ + { + "bbox": [ + 58, + 253, + 286, + 300 + ], + "type": "text", + "content": "- A formulation of a joint deblurring, rolling-shutter correction and video-frame interpolation problem based on CIS and EVS fusion, where EVS sensor nonidealities are explicitly modeled (Section 3)," + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 58, + 300, + 286, + 336 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 300, + 286, + 336 + ], + "spans": [ + { + "bbox": [ + 58, + 300, + 286, + 336 + ], + "type": "text", + "content": "- An optimization methodology jointly solving this coupled inverse problem (Section 4.1), as well as a post-processing network for artifact removal (Section 4.2)," + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 58, + 337, + 286, + 371 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 337, + 286, + 371 + ], + "spans": [ + { + "bbox": [ + 58, + 337, + 286, + 371 + ], + "type": "text", + "content": "- A comprehensive comparison and analysis of simulation and measured datasets in natural and controlled environments (Section 5)." + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 47, + 381, + 134, + 393 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 381, + 134, + 393 + ], + "spans": [ + { + "bbox": [ + 47, + 381, + 134, + 393 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 399, + 289, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 399, + 289, + 590 + ], + "spans": [ + { + "bbox": [ + 46, + 399, + 289, + 590 + ], + "type": "text", + "content": "Video Frame Interpolation Video interpolation aims to synthesize intermediate frames between two images to achieve a temporally coherent video sequence. Traditional CIS-based methods depend on optical flow for a smooth transition. In SuperSlomo [18], the bi-directional flows are combined with a linear combination followed by a refinement using U-Net [33]. Such optical flow-based methods, however, suffer from occlusion problems. In DAIN [1], researchers estimate occlusion areas using depth information and use contextual information from neighboring pixels to fill the occluded regions. In RIFE [17], the flow and fusion maps are generated simultaneously without an additional optical flow module for final result synthesis. FILM [32] presents a multi-scale feature extractor within a unified network to achieve an enlarged receptive field aimed at handling large object motion in the video." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 594, + 289, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 289, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 289, + 714 + ], + "type": "text", + "content": "Event-assisted Frame Enhancement A key issue lowering the image quality of CIS-based frame interpolation methods is the lack of information between frames. EVS provides such information thanks to its asynchronous low latency capture not relying on a global exposure period. Some early studies [31, 34] show that intensity reconstruction purely based on EVS modality is feasible. However, as EVS captures events based on differential changes of illuminance, this approach struggles to estimate the absolute light level. In [2], the optical flow and intensity frames are" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 72, + 545, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 335 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 335 + ], + "type": "text", + "content": "estimated simultaneously by solving an optimization problem that takes EVS measurements as data terms. The authors in [22] take a step further by formulating the event-based SLAM problem, which can be decomposed into camera ego-motion estimation and 3-D scene reconstruction. However, the quantization error makes it difficult to reconstruct high-quality frames from EVS alone. As EVS pixels trigger independently, researchers simplify the problem formulation to pixel-wise estimation by fusing CIS and EVS, without seeking explicit representation of the 3-D environment. The Event-based Double Integral (EDI) [27, 28] method is proposed as a straightforward and effective way to deblur CIS frames and generate high frame rate videos. EVS and CIS measurements are connected by integrating events twice, such that the continuous intensity curve for each pixel can be determined. [34] presents a complementary filter-based framework to combine a high-pass signal from EVS and a low-pass signal from CIS to compute an all-pass signal for pixel intensity. To make the complementary filter gain adaptive to measurement noise, a pixel-wise Kalman filter is used in [39] where a unifying EVS/CIS uncertainty model is proposed." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 337, + 545, + 492 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 337, + 545, + 492 + ], + "spans": [ + { + "bbox": [ + 304, + 337, + 545, + 492 + ], + "type": "text", + "content": "Machine learning methods such as Time Lens [37] and its subsequent versions [10, 36, 38, 41] or others [29, 43] concatenate both CIS and EVS data using a voxel grid representation [12] as input to feed the network of two branches: an optical flow warping-based branch and a synthesis-based branch. The final result is blended with a trained attention module to avoid the disadvantages of any individual branch. CBMNet [23] is proposed to handle complex real-world motion using a novel cross-modal asymmetric bidirectional motion field estimation. Due to the high temporal resolution of EVS data, the rolling shutter problem which is commonly observed in fast-motion scenarios can be corrected using EVS data [6, 44]." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 502, + 545, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 502, + 545, + 574 + ], + "spans": [ + { + "bbox": [ + 304, + 502, + 545, + 574 + ], + "type": "text", + "content": "EVS Latency Many studies model EVS latency [20, 25, 30, 35], but few of them have been applied to algorithm design such as frame enhancement. To the best of the authors' knowledge, this is the first work that considers an EVS latency model explicitly in generating high frame rate, deblurred, rolling-shutter artifact-free videos." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 586, + 432, + 599 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 586, + 432, + 599 + ], + "spans": [ + { + "bbox": [ + 306, + 586, + 432, + 599 + ], + "type": "text", + "content": "3. Problem Formulation" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": "Event Pixel Operation and Pixel Latency Model In this work, a hybrid CIS-EVS sensor [14, 15] is used. As shown in Figure 2(a), EVS and CIS share the same photodiode design on the hybrid sensor and it is assumed that the light level is identical for nearby CIS and EVS pixels. The photocurrent " + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "inline_equation", + "content": "i_{\\mathrm{pd}}" + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": " is converted to front-end voltage " + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "inline_equation", + "content": "V_{\\mathrm{FE}}" + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": " in module " + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{i}2\\mathrm{v}" + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": ", which models dynamic characteristics as a 1st-order Low-Pass Filter (LPF) [21, 26]. The update rule for " + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "inline_equation", + "content": "V_{\\mathrm{FE}}" + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": " between discrete event indices " + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "inline_equation", + "content": "k + 1" + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": " can be written as:" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 731, + 318, + 740 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 731, + 318, + 740 + ], + "spans": [ + { + "bbox": [ + 293, + 731, + 318, + 740 + ], + "type": "text", + "content": "25173" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 51, + 72, + 292, + 125 + ], + "blocks": [ + { + "bbox": [ + 51, + 72, + 292, + 125 + ], + "lines": [ + { + "bbox": [ + 51, + 72, + 292, + 125 + ], + "spans": [ + { + "bbox": [ + 51, + 72, + 292, + 125 + ], + "type": "image", + "image_path": "d5b8847e1319d15d45a2c52e093e33b06a6599f473f2183a404a6577da4202ab.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 51, + 134, + 287, + 242 + ], + "blocks": [ + { + "bbox": [ + 51, + 134, + 287, + 242 + ], + "lines": [ + { + "bbox": [ + 51, + 134, + 287, + 242 + ], + "spans": [ + { + "bbox": [ + 51, + 134, + 287, + 242 + ], + "type": "image", + "image_path": "d400ca00c23e5f4cac2721238af5c6c712e24b8f18b992c3794696c18c2b4678.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 255, + 287, + 310 + ], + "lines": [ + { + "bbox": [ + 46, + 255, + 287, + 310 + ], + "spans": [ + { + "bbox": [ + 46, + 255, + 287, + 310 + ], + "type": "text", + "content": "Figure 2. (a) Hybrid CIS-EVS sensor principle. The upper and lower branches indicate physical quantity conversion in CIS pixels and EVS pixels, respectively. (b) Example of " + }, + { + "bbox": [ + 46, + 255, + 287, + 310 + ], + "type": "inline_equation", + "content": "V_{\\mathrm{DC}}" + }, + { + "bbox": [ + 46, + 255, + 287, + 310 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 255, + 287, + 310 + ], + "type": "inline_equation", + "content": "V_{\\mathrm{FE}}" + }, + { + "bbox": [ + 46, + 255, + 287, + 310 + ], + "type": "text", + "content": " curves and triggered events that illustrate the impact of " + }, + { + "bbox": [ + 46, + 255, + 287, + 310 + ], + "type": "inline_equation", + "content": "V_{\\mathrm{FE}}" + }, + { + "bbox": [ + 46, + 255, + 287, + 310 + ], + "type": "text", + "content": " delay, readout latency, and refractory period." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "bbox": [ + 59, + 340, + 287, + 353 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 340, + 287, + 353 + ], + "spans": [ + { + "bbox": [ + 59, + 340, + 287, + 353 + ], + "type": "interline_equation", + "content": "V _ {\\mathrm {F E}} (k + 1) = V _ {\\mathrm {F E}} (k) + \\gamma (k) \\left(V _ {\\mathrm {D C}} (k + 1) - V _ {\\mathrm {F E}} (k)\\right), \\tag {1}", + "image_path": "b3c9297e9cf803ab8a0bbbaa7b7d166e20cbcf779b5ee58c04009a1e26171002.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 355, + 287, + 401 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 355, + 287, + 401 + ], + "spans": [ + { + "bbox": [ + 47, + 355, + 287, + 401 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 355, + 287, + 401 + ], + "type": "inline_equation", + "content": "V_{\\mathrm{DC}}(k + 1) = f_{\\mathrm{DC}}(i_{\\mathrm{pd}}(k + 1))" + }, + { + "bbox": [ + 47, + 355, + 287, + 401 + ], + "type": "text", + "content": " represents the " + }, + { + "bbox": [ + 47, + 355, + 287, + 401 + ], + "type": "inline_equation", + "content": "V_{\\mathrm{FE}}" + }, + { + "bbox": [ + 47, + 355, + 287, + 401 + ], + "type": "text", + "content": " under DC excitation (i.e., the " + }, + { + "bbox": [ + 47, + 355, + 287, + 401 + ], + "type": "inline_equation", + "content": "V_{\\mathrm{FE}}" + }, + { + "bbox": [ + 47, + 355, + 287, + 401 + ], + "type": "text", + "content": " with infinite LPF bandwidth); " + }, + { + "bbox": [ + 47, + 355, + 287, + 401 + ], + "type": "inline_equation", + "content": "f_{\\mathrm{DC}}(i_{\\mathrm{pd}})" + }, + { + "bbox": [ + 47, + 355, + 287, + 401 + ], + "type": "text", + "content": " is the logarithmic-like current-voltage conversion function and" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 110, + 403, + 287, + 429 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 110, + 403, + 287, + 429 + ], + "spans": [ + { + "bbox": [ + 110, + 403, + 287, + 429 + ], + "type": "interline_equation", + "content": "\\gamma (k) = 1 - \\exp \\left(- \\frac {\\Delta t}{\\tau (k)}\\right) \\tag {2}", + "image_path": "3d987c7246488b840dd234bcb964513673c14fa6003ed3aab077a8835677225c.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 430, + 287, + 465 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 430, + 287, + 465 + ], + "spans": [ + { + "bbox": [ + 46, + 430, + 287, + 465 + ], + "type": "text", + "content": "is the time-varying coefficient of the LPF, where " + }, + { + "bbox": [ + 46, + 430, + 287, + 465 + ], + "type": "inline_equation", + "content": "\\Delta t" + }, + { + "bbox": [ + 46, + 430, + 287, + 465 + ], + "type": "text", + "content": " is the time interval between two indices. The photocurrent-dependent time constant " + }, + { + "bbox": [ + 46, + 430, + 287, + 465 + ], + "type": "inline_equation", + "content": "\\tau (k)" + }, + { + "bbox": [ + 46, + 430, + 287, + 465 + ], + "type": "text", + "content": " can be modeled as [16, 26]:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 113, + 468, + 287, + 491 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 113, + 468, + 287, + 491 + ], + "spans": [ + { + "bbox": [ + 113, + 468, + 287, + 491 + ], + "type": "interline_equation", + "content": "\\tau (k) = \\tau_ {0} + \\frac {\\alpha}{i _ {\\mathrm {p d}} (k)}, \\tag {3}", + "image_path": "7f3764d9b3d113ec93d8ad738a25c7a427a40aaad8cdbf21957826442646d674.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 492, + 287, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 492, + 287, + 552 + ], + "spans": [ + { + "bbox": [ + 46, + 492, + 287, + 552 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 492, + 287, + 552 + ], + "type": "inline_equation", + "content": "\\tau_0" + }, + { + "bbox": [ + 46, + 492, + 287, + 552 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 492, + 287, + 552 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 46, + 492, + 287, + 552 + ], + "type": "text", + "content": " are pixel parameters. The pixel latency in module " + }, + { + "bbox": [ + 46, + 492, + 287, + 552 + ], + "type": "inline_equation", + "content": "\\pm 2\\mathrm{v}" + }, + { + "bbox": [ + 46, + 492, + 287, + 552 + ], + "type": "text", + "content": " due to the LPF behavior is called \" " + }, + { + "bbox": [ + 46, + 492, + 287, + 552 + ], + "type": "inline_equation", + "content": "V_{\\mathrm{FE}}" + }, + { + "bbox": [ + 46, + 492, + 287, + 552 + ], + "type": "text", + "content": " delay\". As illustrated in Figure 2(b), an event is triggered at time " + }, + { + "bbox": [ + 46, + 492, + 287, + 552 + ], + "type": "inline_equation", + "content": "t_\\mathrm{in}" + }, + { + "bbox": [ + 46, + 492, + 287, + 552 + ], + "type": "text", + "content": " once the difference between " + }, + { + "bbox": [ + 46, + 492, + 287, + 552 + ], + "type": "inline_equation", + "content": "V_{\\mathrm{FE}}" + }, + { + "bbox": [ + 46, + 492, + 287, + 552 + ], + "type": "text", + "content": " and reference voltage " + }, + { + "bbox": [ + 46, + 492, + 287, + 552 + ], + "type": "inline_equation", + "content": "V_{\\mathrm{ref}}" + }, + { + "bbox": [ + 46, + 492, + 287, + 552 + ], + "type": "text", + "content": " reaches the triggering threshold voltage:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 554, + 287, + 579 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 554, + 287, + 579 + ], + "spans": [ + { + "bbox": [ + 48, + 554, + 287, + 579 + ], + "type": "interline_equation", + "content": "\\left\\{ \\begin{array}{l l} V _ {\\mathrm {F E}} \\left(t _ {\\mathrm {i n}}\\right) - V _ {\\mathrm {r e f}} \\left(t _ {\\mathrm {i n}}\\right) & \\geq c _ {\\mathrm {u p}} \\quad \\text {f o r p o s i t i v e e v e n t s} \\\\ V _ {\\mathrm {F E}} \\left(t _ {\\mathrm {i n}}\\right) - V _ {\\mathrm {r e f}} \\left(t _ {\\mathrm {i n}}\\right) & \\leq c _ {\\mathrm {d o w n}} \\quad \\text {f o r n e g a t i v e e v e n t s} \\end{array} , \\right. \\tag {4}", + "image_path": "192ea3e72299afea26abad538579b5e49dd58f806bfd9d85b0c5a136112a2f58.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 581, + 287, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 581, + 287, + 664 + ], + "spans": [ + { + "bbox": [ + 46, + 581, + 287, + 664 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 581, + 287, + 664 + ], + "type": "inline_equation", + "content": "c_{\\mathrm{up / down}}" + }, + { + "bbox": [ + 46, + 581, + 287, + 664 + ], + "type": "text", + "content": " denotes the contrast threshold voltage. The in-pixel timestamp " + }, + { + "bbox": [ + 46, + 581, + 287, + 664 + ], + "type": "inline_equation", + "content": "t_{\\mathrm{in}}" + }, + { + "bbox": [ + 46, + 581, + 287, + 664 + ], + "type": "text", + "content": " is recorded when the event fires, after which the pixel experiences \"readout latency " + }, + { + "bbox": [ + 46, + 581, + 287, + 664 + ], + "type": "inline_equation", + "content": "t_{\\mathrm{rl}}" + }, + { + "bbox": [ + 46, + 581, + 287, + 664 + ], + "type": "text", + "content": "\" and \"refractory period " + }, + { + "bbox": [ + 46, + 581, + 287, + 664 + ], + "type": "inline_equation", + "content": "t_{\\mathrm{rp}}" + }, + { + "bbox": [ + 46, + 581, + 287, + 664 + ], + "type": "text", + "content": "\". The " + }, + { + "bbox": [ + 46, + 581, + 287, + 664 + ], + "type": "inline_equation", + "content": "V_{\\mathrm{ref}}" + }, + { + "bbox": [ + 46, + 581, + 287, + 664 + ], + "type": "text", + "content": " is updated to follow " + }, + { + "bbox": [ + 46, + 581, + 287, + 664 + ], + "type": "inline_equation", + "content": "V_{\\mathrm{FE}}" + }, + { + "bbox": [ + 46, + 581, + 287, + 664 + ], + "type": "text", + "content": " after the refractory period. Finally, we have a 5-dimensional output " + }, + { + "bbox": [ + 46, + 581, + 287, + 664 + ], + "type": "inline_equation", + "content": "\\{x,y,t_{\\mathrm{in}},t_{\\mathrm{off}},p\\}" + }, + { + "bbox": [ + 46, + 581, + 287, + 664 + ], + "type": "text", + "content": ", which contains row and column indices, in-pixel and off-pixel timestamps, and polarity." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 665, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 287, + 713 + ], + "type": "text", + "content": "Pixel-wise Photocurrent Estimation Suppose that the illuminance on the sensor plane results in photocurrent " + }, + { + "bbox": [ + 46, + 665, + 287, + 713 + ], + "type": "inline_equation", + "content": "i_{\\mathrm{pd}}" + }, + { + "bbox": [ + 46, + 665, + 287, + 713 + ], + "type": "text", + "content": " for a single pixel. We aim to estimate " + }, + { + "bbox": [ + 46, + 665, + 287, + 713 + ], + "type": "inline_equation", + "content": "i_{\\mathrm{pd}}" + }, + { + "bbox": [ + 46, + 665, + 287, + 713 + ], + "type": "text", + "content": " when an event is triggered to avoid the impact of pixel latency in the " + }, + { + "bbox": [ + 46, + 665, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\pm 2\\mathrm{v}" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 72, + 545, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 95 + ], + "type": "text", + "content": "module. Based on Figure 2(a), we have the following CIS measurement equation:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 367, + 98, + 545, + 125 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 367, + 98, + 545, + 125 + ], + "spans": [ + { + "bbox": [ + 367, + 98, + 545, + 125 + ], + "type": "interline_equation", + "content": "D N = G \\int_ {t _ {\\mathrm {s}}} ^ {t _ {\\mathrm {e}}} i _ {\\mathrm {p d}} (t) \\mathrm {d} t, \\tag {5}", + "image_path": "2089a7f069d1548f6ad800c29e02bd394e023c665df55da3d6267d28bd83d4a6.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 126, + 545, + 210 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 126, + 545, + 210 + ], + "spans": [ + { + "bbox": [ + 305, + 126, + 545, + 210 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 126, + 545, + 210 + ], + "type": "inline_equation", + "content": "DN" + }, + { + "bbox": [ + 305, + 126, + 545, + 210 + ], + "type": "text", + "content": " is the CIS intensity frame measured in digital numbers; " + }, + { + "bbox": [ + 305, + 126, + 545, + 210 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 305, + 126, + 545, + 210 + ], + "type": "text", + "content": " indicates the total gain of charge-voltage conversion and ADC; " + }, + { + "bbox": [ + 305, + 126, + 545, + 210 + ], + "type": "inline_equation", + "content": "t_{\\mathrm{s}}" + }, + { + "bbox": [ + 305, + 126, + 545, + 210 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 126, + 545, + 210 + ], + "type": "inline_equation", + "content": "t_{\\mathrm{e}}" + }, + { + "bbox": [ + 305, + 126, + 545, + 210 + ], + "type": "text", + "content": " denote the CIS exposure start and end, respectively. As for EVS measurements, suppose we have an event sequence with " + }, + { + "bbox": [ + 305, + 126, + 545, + 210 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 305, + 126, + 545, + 210 + ], + "type": "text", + "content": " events " + }, + { + "bbox": [ + 305, + 126, + 545, + 210 + ], + "type": "inline_equation", + "content": "\\{x(k),y(k),t_{\\mathrm{in}}(k),t_{\\mathrm{off}}(k),p(k)\\}" + }, + { + "bbox": [ + 305, + 126, + 545, + 210 + ], + "type": "text", + "content": ". By assuming " + }, + { + "bbox": [ + 305, + 126, + 545, + 210 + ], + "type": "inline_equation", + "content": "V_{\\mathrm{FE}}(1) = V_{\\mathrm{DC}}(1)" + }, + { + "bbox": [ + 305, + 126, + 545, + 210 + ], + "type": "text", + "content": ", with Eq. (1) we have the recurrence relation:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 322, + 213, + 545, + 226 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 322, + 213, + 545, + 226 + ], + "spans": [ + { + "bbox": [ + 322, + 213, + 545, + 226 + ], + "type": "interline_equation", + "content": "V _ {\\mathrm {F E}} (1) = f _ {\\mathrm {D C}} \\left(i _ {\\mathrm {p d}} (1)\\right) \\tag {6}", + "image_path": "0d0da8eecfad8ce5ef72b807b9c4cc52c937360e6d4e0204c546b73b32c01887.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 305, + 228, + 545, + 241 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 228, + 545, + 241 + ], + "spans": [ + { + "bbox": [ + 305, + 228, + 545, + 241 + ], + "type": "interline_equation", + "content": "V _ {\\mathrm {F E}} (k + 1) = f _ {\\gamma} \\left(i _ {\\mathrm {p d}} (k)\\right) \\left(f _ {\\mathrm {D C}} \\left(i _ {\\mathrm {p d}} (k + 1)\\right) - V _ {\\mathrm {F E}} (k)\\right), \\tag {7}", + "image_path": "91a16f01042d5aa071a020650e1252f3a90d7b2c92c3ec6ce4d480c8e7e7842c.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 305, + 243, + 545, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 243, + 545, + 289 + ], + "spans": [ + { + "bbox": [ + 305, + 243, + 545, + 289 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 243, + 545, + 289 + ], + "type": "inline_equation", + "content": "k \\in \\{1, \\dots, N-1\\}" + }, + { + "bbox": [ + 305, + 243, + 545, + 289 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 243, + 545, + 289 + ], + "type": "inline_equation", + "content": "f_{\\gamma}(\\cdot)" + }, + { + "bbox": [ + 305, + 243, + 545, + 289 + ], + "type": "text", + "content": " denotes the function that maps " + }, + { + "bbox": [ + 305, + 243, + 545, + 289 + ], + "type": "inline_equation", + "content": "i_{\\mathrm{pd}}" + }, + { + "bbox": [ + 305, + 243, + 545, + 289 + ], + "type": "text", + "content": " onto " + }, + { + "bbox": [ + 305, + 243, + 545, + 289 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 305, + 243, + 545, + 289 + ], + "type": "text", + "content": ". By assuming " + }, + { + "bbox": [ + 305, + 243, + 545, + 289 + ], + "type": "inline_equation", + "content": "V_{\\mathrm{ref}}(k+1) \\approx V_{\\mathrm{FE}}(k)" + }, + { + "bbox": [ + 305, + 243, + 545, + 289 + ], + "type": "text", + "content": ", the EVS measurements provide additional " + }, + { + "bbox": [ + 305, + 243, + 545, + 289 + ], + "type": "inline_equation", + "content": "N-1" + }, + { + "bbox": [ + 305, + 243, + 545, + 289 + ], + "type": "text", + "content": " equations as:" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 346, + 294, + 545, + 306 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 346, + 294, + 545, + 306 + ], + "spans": [ + { + "bbox": [ + 346, + 294, + 545, + 306 + ], + "type": "interline_equation", + "content": "c (k + 1) \\approx V _ {\\mathrm {F E}} (k + 1) - V _ {\\mathrm {F E}} (k), \\tag {8}", + "image_path": "475f51e568feaa0976b8c2072eddf1a1fd9ffd138f9bf1b343288148f9e2642f.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 308, + 545, + 464 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 308, + 545, + 464 + ], + "spans": [ + { + "bbox": [ + 304, + 308, + 545, + 464 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 308, + 545, + 464 + ], + "type": "inline_equation", + "content": "c(k) \\in \\{c_{\\mathrm{up}}, c_{\\mathrm{down}}\\}" + }, + { + "bbox": [ + 304, + 308, + 545, + 464 + ], + "type": "text", + "content": " are the known triggering thresholds that are configurable. These state transition equations and measurement equations (6)-(8) can be expressed as a system of " + }, + { + "bbox": [ + 304, + 308, + 545, + 464 + ], + "type": "inline_equation", + "content": "2N - 1" + }, + { + "bbox": [ + 304, + 308, + 545, + 464 + ], + "type": "text", + "content": " nonlinear equations and " + }, + { + "bbox": [ + 304, + 308, + 545, + 464 + ], + "type": "inline_equation", + "content": "2N" + }, + { + "bbox": [ + 304, + 308, + 545, + 464 + ], + "type": "text", + "content": " unknowns " + }, + { + "bbox": [ + 304, + 308, + 545, + 464 + ], + "type": "inline_equation", + "content": "-i_{\\mathrm{pd}}(k)" + }, + { + "bbox": [ + 304, + 308, + 545, + 464 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 308, + 545, + 464 + ], + "type": "inline_equation", + "content": "V_{\\mathrm{FE}}(k)" + }, + { + "bbox": [ + 304, + 308, + 545, + 464 + ], + "type": "text", + "content": ". By using a zero-order holder, the discrete states " + }, + { + "bbox": [ + 304, + 308, + 545, + 464 + ], + "type": "inline_equation", + "content": "i_{\\mathrm{pd}}(k)" + }, + { + "bbox": [ + 304, + 308, + 545, + 464 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 308, + 545, + 464 + ], + "type": "inline_equation", + "content": "V_{\\mathrm{FE}}(k)" + }, + { + "bbox": [ + 304, + 308, + 545, + 464 + ], + "type": "text", + "content": " can be approximated as continuous signals " + }, + { + "bbox": [ + 304, + 308, + 545, + 464 + ], + "type": "inline_equation", + "content": "i_{\\mathrm{pd}}(t)" + }, + { + "bbox": [ + 304, + 308, + 545, + 464 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 308, + 545, + 464 + ], + "type": "inline_equation", + "content": "V_{\\mathrm{FE}}(t)" + }, + { + "bbox": [ + 304, + 308, + 545, + 464 + ], + "type": "text", + "content": ". This makes it possible to utilize CIS measurements in Eq. (5) as additional constraints. In conclusion, the pixel latency compensation can be expressed as pixel-wise optimization problem, solving the nonlinear system with " + }, + { + "bbox": [ + 304, + 308, + 545, + 464 + ], + "type": "inline_equation", + "content": "2N" + }, + { + "bbox": [ + 304, + 308, + 545, + 464 + ], + "type": "text", + "content": " unknown states and " + }, + { + "bbox": [ + 304, + 308, + 545, + 464 + ], + "type": "inline_equation", + "content": "2N + M - 1" + }, + { + "bbox": [ + 304, + 308, + 545, + 464 + ], + "type": "text", + "content": " equations, where " + }, + { + "bbox": [ + 304, + 308, + 545, + 464 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 304, + 308, + 545, + 464 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 308, + 545, + 464 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 308, + 545, + 464 + ], + "type": "text", + "content": " denote the number of CIS and EVS measurements, respectively." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 484, + 545, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 484, + 545, + 579 + ], + "spans": [ + { + "bbox": [ + 304, + 484, + 545, + 579 + ], + "type": "text", + "content": "Modeling Readout Latency and Refractory Period The EVS measurement model equation (8) is an approximation since the state variation during the readout latency and the refractory period is not considered. Using both in-pixel timestamps and off-pixel timestamps, it is possible to compensate for some of the \"missed\" " + }, + { + "bbox": [ + 304, + 484, + 545, + 579 + ], + "type": "inline_equation", + "content": "V_{\\mathrm{FE}}" + }, + { + "bbox": [ + 304, + 484, + 545, + 579 + ], + "type": "text", + "content": " change by assuming that the slope of " + }, + { + "bbox": [ + 304, + 484, + 545, + 579 + ], + "type": "inline_equation", + "content": "V_{\\mathrm{FE}}" + }, + { + "bbox": [ + 304, + 484, + 545, + 579 + ], + "type": "text", + "content": " is locally constant for a short period. As illustrated in Figure 2(b), it is noted that:" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 335, + 582, + 545, + 608 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 582, + 545, + 608 + ], + "spans": [ + { + "bbox": [ + 335, + 582, + 545, + 608 + ], + "type": "interline_equation", + "content": "m (k + 1) = \\frac {c (k + 1)}{t _ {\\mathrm {i n}} (k + 1) - t _ {\\mathrm {o f f}} (k) - t _ {\\mathrm {r p}}}, \\tag {9}", + "image_path": "634ca84b437c999c25be2d8a8313146c775e7a6d650240c14f67e68091e61274.jpg" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 304, + 609, + 545, + 643 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 609, + 545, + 643 + ], + "spans": [ + { + "bbox": [ + 304, + 609, + 545, + 643 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 609, + 545, + 643 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 304, + 609, + 545, + 643 + ], + "type": "text", + "content": " describes the slope approximation of the " + }, + { + "bbox": [ + 304, + 609, + 545, + 643 + ], + "type": "inline_equation", + "content": "V_{\\mathrm{FE}}" + }, + { + "bbox": [ + 304, + 609, + 545, + 643 + ], + "type": "text", + "content": " curve. For the readout latency, the missed " + }, + { + "bbox": [ + 304, + 609, + 545, + 643 + ], + "type": "inline_equation", + "content": "V_{\\mathrm{FE}}" + }, + { + "bbox": [ + 304, + 609, + 545, + 643 + ], + "type": "text", + "content": " is represented as:" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 334, + 647, + 545, + 661 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 334, + 647, + 545, + 661 + ], + "spans": [ + { + "bbox": [ + 334, + 647, + 545, + 661 + ], + "type": "interline_equation", + "content": "\\Delta V _ {\\mathrm {F E}} ^ {\\mathrm {r l}} (k) = m (k + 1) \\left[ t _ {\\text {o f f}} (k) - t _ {\\text {i n}} (k) \\right]. \\tag {10}", + "image_path": "73dc435a83f5dcef6228a42ce4d09dbd15d393b5791ece988e1919845387a95e.jpg" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 304, + 662, + 545, + 698 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 662, + 545, + 698 + ], + "spans": [ + { + "bbox": [ + 304, + 662, + 545, + 698 + ], + "type": "text", + "content": "For the refractory period, we have " + }, + { + "bbox": [ + 304, + 662, + 545, + 698 + ], + "type": "inline_equation", + "content": "\\Delta V_{\\mathrm{FE}}^{\\mathrm{rp}}(k) = m(k + 1)t_{\\mathrm{rp}}" + }, + { + "bbox": [ + 304, + 662, + 545, + 698 + ], + "type": "text", + "content": " since " + }, + { + "bbox": [ + 304, + 662, + 545, + 698 + ], + "type": "inline_equation", + "content": "t_\\mathrm{rp}" + }, + { + "bbox": [ + 304, + 662, + 545, + 698 + ], + "type": "text", + "content": " is a fixed parameter. Thus the reference voltage after compensation can be adjusted to:" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 321, + 700, + 545, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 700, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 321, + 700, + 545, + 715 + ], + "type": "interline_equation", + "content": "V _ {\\text {r e f}} (k + 1) = V _ {\\mathrm {F E}} (k) + \\Delta V _ {\\mathrm {F E}} ^ {\\mathrm {r l}} (k) + \\Delta V _ {\\mathrm {F E}} ^ {\\mathrm {r p}} (k). \\tag {11}", + "image_path": "cb10c72964d17e99973dfd3781e3a24000b09761f946fb54d29dd3195cb5208c.jpg" + } + ] + } + ], + "index": 25 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 731, + 317, + 740 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 731, + 317, + 740 + ], + "spans": [ + { + "bbox": [ + 293, + 731, + 317, + 740 + ], + "type": "text", + "content": "25174" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 49, + 70, + 293, + 210 + ], + "blocks": [ + { + "bbox": [ + 49, + 70, + 293, + 210 + ], + "lines": [ + { + "bbox": [ + 49, + 70, + 293, + 210 + ], + "spans": [ + { + "bbox": [ + 49, + 70, + 293, + 210 + ], + "type": "image", + "image_path": "8fd2a7b91b0979a19d54ffcd360ed1de72ee3a74f991f18b974e821b8c1bda04.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 222, + 288, + 266 + ], + "lines": [ + { + "bbox": [ + 46, + 222, + 288, + 266 + ], + "spans": [ + { + "bbox": [ + 46, + 222, + 288, + 266 + ], + "type": "text", + "content": "Figure 3. Block diagram of the proposed method. Note that after the frame reconstruction the blurry texture due to motion is enhanced and the RS effect is removed. The noise and artifacts are further reduced by the refinement network." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 287, + 287, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 287, + 287, + 323 + ], + "spans": [ + { + "bbox": [ + 46, + 287, + 287, + 323 + ], + "type": "text", + "content": "By substituting the compensation equation (11) into (4), the impact of readout latency and refractory period can be compensated:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 325, + 287, + 390 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 325, + 287, + 390 + ], + "spans": [ + { + "bbox": [ + 47, + 325, + 287, + 390 + ], + "type": "inline_equation", + "content": "c(k + 1)\\approx V_{\\mathrm{FE}}(k + 1) - V_{\\mathrm{FE}}(k) - \\Delta V_{\\mathrm{FE}}^{\\mathrm{sl}}(k) - \\Delta V_{\\mathrm{FE}}^{\\mathrm{rp}}(k)," + }, + { + "bbox": [ + 47, + 325, + 287, + 390 + ], + "type": "text", + "content": " (12) where " + }, + { + "bbox": [ + 47, + 325, + 287, + 390 + ], + "type": "inline_equation", + "content": "\\Delta V_{\\mathrm{FE}}^{\\mathrm{rl}}(k)" + }, + { + "bbox": [ + 47, + 325, + 287, + 390 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 325, + 287, + 390 + ], + "type": "inline_equation", + "content": "\\Delta V_{\\mathrm{FE}}^{\\mathrm{rp}}(k)" + }, + { + "bbox": [ + 47, + 325, + 287, + 390 + ], + "type": "text", + "content": " are voltage compensation for readout latency and refractory period, respectively. Note that this compensation is enabled by the availability of in-pixel and off-pixel timestamps." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 396, + 103, + 408 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 396, + 103, + 408 + ], + "spans": [ + { + "bbox": [ + 47, + 396, + 103, + 408 + ], + "type": "text", + "content": "4. Method" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 414, + 287, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 414, + 287, + 475 + ], + "spans": [ + { + "bbox": [ + 46, + 414, + 287, + 475 + ], + "type": "text", + "content": "The proposed frame enhancement framework is presented in Figure 3. The frame reconstruction module jointly deblurs images, removes the RS effect, and interpolates high frame rate images. The refinement network focuses on removing noise and improving image quality." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 480, + 274, + 495 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 480, + 274, + 495 + ], + "spans": [ + { + "bbox": [ + 47, + 480, + 274, + 495 + ], + "type": "text", + "content": "4.1. Solving the Inverse Model via Optimization" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 499, + 287, + 642 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 499, + 287, + 642 + ], + "spans": [ + { + "bbox": [ + 46, + 499, + 287, + 642 + ], + "type": "text", + "content": "Graph Modeling for a Single Pixel Due to the presence of pixel latency, the system is governed by nonlinear equations, rendering the models and solutions designed for linear systems inapplicable. Graph optimization [13] provides a straightforward way to model nonlinear systems, where the states and measurements are represented as nodes and edges in a graph as shown in Figure 4(a). Let " + }, + { + "bbox": [ + 46, + 499, + 287, + 642 + ], + "type": "inline_equation", + "content": "l \\in \\{1, \\dots, M\\}" + }, + { + "bbox": [ + 46, + 499, + 287, + 642 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 499, + 287, + 642 + ], + "type": "inline_equation", + "content": "k \\in \\{1, \\dots, N-1\\}" + }, + { + "bbox": [ + 46, + 499, + 287, + 642 + ], + "type": "text", + "content": " be the indices of CIS and EVS, respectively. By defining states " + }, + { + "bbox": [ + 46, + 499, + 287, + 642 + ], + "type": "inline_equation", + "content": "\\boldsymbol{x} = [i_{\\mathrm{pd}}(k)]^{\\top}" + }, + { + "bbox": [ + 46, + 499, + 287, + 642 + ], + "type": "text", + "content": " and measurements " + }, + { + "bbox": [ + 46, + 499, + 287, + 642 + ], + "type": "inline_equation", + "content": "z_{\\mathrm{CIS}} = [DN(l)]^{\\top}" + }, + { + "bbox": [ + 46, + 499, + 287, + 642 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 499, + 287, + 642 + ], + "type": "inline_equation", + "content": "z_{\\mathrm{EVS}} = [c(k+1)]^{\\top}" + }, + { + "bbox": [ + 46, + 499, + 287, + 642 + ], + "type": "text", + "content": ", the pixel-wise graph optimization framework aims to determine the state vector " + }, + { + "bbox": [ + 46, + 499, + 287, + 642 + ], + "type": "inline_equation", + "content": "\\boldsymbol{x}" + }, + { + "bbox": [ + 46, + 499, + 287, + 642 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 61, + 643, + 287, + 666 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 643, + 287, + 666 + ], + "spans": [ + { + "bbox": [ + 61, + 643, + 287, + 666 + ], + "type": "interline_equation", + "content": "\\boldsymbol {x} ^ {*} = \\underset {\\boldsymbol {x}} {\\operatorname {a r g m i n}} \\left(\\boldsymbol {e} _ {\\mathrm {C I S}} ^ {\\top} \\boldsymbol {\\Omega} _ {\\mathrm {C I S}} \\boldsymbol {e} _ {\\mathrm {C I S}} + \\boldsymbol {e} _ {\\mathrm {E V S}} ^ {\\top} \\boldsymbol {\\Omega} _ {\\mathrm {E V S}} \\boldsymbol {e} _ {\\mathrm {E V S}}\\right) \\tag {13}", + "image_path": "4ac01209fe3fcffb57c5513ca0ab515771d2df9eab88b4be353d0431064e994d.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 666, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 666, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 666, + 288, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 666, + 288, + 714 + ], + "type": "inline_equation", + "content": "e_{\\mathrm{CIS}} = z_{\\mathrm{CIS}} - \\tilde{z}_{\\mathrm{CIS}}" + }, + { + "bbox": [ + 46, + 666, + 288, + 714 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 666, + 288, + 714 + ], + "type": "inline_equation", + "content": "e_{\\mathrm{EVS}} = z_{\\mathrm{EVS}} - \\tilde{z}_{\\mathrm{EVS}}" + }, + { + "bbox": [ + 46, + 666, + 288, + 714 + ], + "type": "text", + "content": " denote measurement error vectors; " + }, + { + "bbox": [ + 46, + 666, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\tilde{z}_{\\mathrm{CIS}}" + }, + { + "bbox": [ + 46, + 666, + 288, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 666, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\tilde{z}_{\\mathrm{EVS}}" + }, + { + "bbox": [ + 46, + 666, + 288, + 714 + ], + "type": "text", + "content": " are predicted measurements which can be calculated from estimated states according to CIS measurement equation (5) and EVS mea" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 307, + 72, + 548, + 167 + ], + "blocks": [ + { + "bbox": [ + 307, + 72, + 548, + 167 + ], + "lines": [ + { + "bbox": [ + 307, + 72, + 548, + 167 + ], + "spans": [ + { + "bbox": [ + 307, + 72, + 548, + 167 + ], + "type": "image", + "image_path": "77e4328e7b85c58d782e23bb33abec072591107f7992dbc86e7c20e55805411f.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 313, + 177, + 549, + 297 + ], + "blocks": [ + { + "bbox": [ + 313, + 177, + 549, + 297 + ], + "lines": [ + { + "bbox": [ + 313, + 177, + 549, + 297 + ], + "spans": [ + { + "bbox": [ + 313, + 177, + 549, + 297 + ], + "type": "image", + "image_path": "cb62d3b02e5fa2c6832a628e5522cd82486490e12ccbfcbc8ebc60eb1bd21f54.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 310, + 547, + 432 + ], + "lines": [ + { + "bbox": [ + 304, + 310, + 547, + 432 + ], + "spans": [ + { + "bbox": [ + 304, + 310, + 547, + 432 + ], + "type": "text", + "content": "Figure 4. (a) Graph modeling of the pixel-wise photocurrent estimation problem, where nodes and edges denote states and measurements. The arrows indicate the order of state calculation. \"zCIS\" and \"zEVS\" indicate CIS and EVS measurements, respectively. Eqn. (8) is replaced with Eqn. (12) when voltage compensation for readout latency and refractory period is enabled. (b) Row-based reconstruction scheme showing three CIS images as input. After row-wise computation, interpolated and deblurred frames without RS effect can be generated as shown in the blue box. The proposed scheme can be applied to video enhancement with any number of CIS images." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 448, + 546, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 448, + 546, + 604 + ], + "spans": [ + { + "bbox": [ + 304, + 448, + 546, + 604 + ], + "type": "text", + "content": "surement equation (8); " + }, + { + "bbox": [ + 304, + 448, + 546, + 604 + ], + "type": "inline_equation", + "content": "z_{\\mathrm{EVS}}" + }, + { + "bbox": [ + 304, + 448, + 546, + 604 + ], + "type": "text", + "content": " is adjusted using register settings and calibrated using the \"S-curve method\" [7]. The statistical distribution of the threshold is not explicitly modeled in the current method. The diagonal weighing matrices " + }, + { + "bbox": [ + 304, + 448, + 546, + 604 + ], + "type": "inline_equation", + "content": "\\Omega_{\\mathrm{CIS}}" + }, + { + "bbox": [ + 304, + 448, + 546, + 604 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 448, + 546, + 604 + ], + "type": "inline_equation", + "content": "\\Omega_{\\mathrm{EVS}}" + }, + { + "bbox": [ + 304, + 448, + 546, + 604 + ], + "type": "text", + "content": " reflect the reliability of CIS and EVS, respectively. Using the cost function equation (13), the " + }, + { + "bbox": [ + 304, + 448, + 546, + 604 + ], + "type": "inline_equation", + "content": "V_{\\mathrm{FE}}" + }, + { + "bbox": [ + 304, + 448, + 546, + 604 + ], + "type": "text", + "content": " delay is modeled explicitly through the EVS measurement equations. Solving the optimization problem leads to state estimation such that the " + }, + { + "bbox": [ + 304, + 448, + 546, + 604 + ], + "type": "inline_equation", + "content": "V_{\\mathrm{FE}}" + }, + { + "bbox": [ + 304, + 448, + 546, + 604 + ], + "type": "text", + "content": " delay is compensated. Many approaches have been proposed to solve nonlinear optimization problems iteratively [19]. We parallelize pixelwise tasks as there is no data interaction or computational dependency between pixels." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "type": "text", + "content": "Rolling Shutter (RS) Effect Correction As CIS parameters such as row exposure time and RS scanning speed are known, we formulate the frame estimation problem as pixel-wise computation using adjusted CIS timing parameters for each row. After solving the photocurrent at each event time-step through optimization, a zero-order hold is used to generate a time-continuous photocurrent " + }, + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "type": "inline_equation", + "content": "i_{\\mathrm{pd}}(t)" + }, + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "type": "text", + "content": " for all pixels. The CIS measurement equation (5) is used to generate a rolling shutter artifact-free target frame rate (such" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 731, + 318, + 740 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 731, + 318, + 740 + ], + "spans": [ + { + "bbox": [ + 293, + 731, + 318, + 740 + ], + "type": "text", + "content": "25175" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 70, + 73, + 267, + 205 + ], + "blocks": [ + { + "bbox": [ + 70, + 73, + 267, + 205 + ], + "lines": [ + { + "bbox": [ + 70, + 73, + 267, + 205 + ], + "spans": [ + { + "bbox": [ + 70, + 73, + 267, + 205 + ], + "type": "image", + "image_path": "a941213ff7db73c7ecfd1ea401ec15d50529b2461cb01d3bfcfcbec8f943f606.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 62, + 213, + 271, + 224 + ], + "lines": [ + { + "bbox": [ + 62, + 213, + 271, + 224 + ], + "spans": [ + { + "bbox": [ + 62, + 213, + 271, + 224 + ], + "type": "text", + "content": "Figure 5. Modified NAF block in the refinement network." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 242, + 115, + 253 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 242, + 115, + 253 + ], + "spans": [ + { + "bbox": [ + 47, + 242, + 115, + 253 + ], + "type": "text", + "content": "as 10 000 FPS)." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 261, + 288, + 369 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 261, + 288, + 369 + ], + "spans": [ + { + "bbox": [ + 46, + 261, + 288, + 369 + ], + "type": "text", + "content": "Video Glitch Effect Deblurring and video frame interpolation performance can experience glitches in case these are computed on a frame-by-frame basis. To mitigate this issue, we solve the coupled deblurring, rolling-shutter correction and interpolation problem jointly for subsequent frames utilizing the row-specific rolling shutter timing characteristics as shown in Figure 4(b), where we reconstruct each row according to its specific exposure time by fusing two consecutive CIS frames and their corresponding events." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 380, + 167, + 392 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 380, + 167, + 392 + ], + "spans": [ + { + "bbox": [ + 47, + 380, + 167, + 392 + ], + "type": "text", + "content": "4.2. Refinement Network" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 399, + 288, + 698 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 399, + 288, + 698 + ], + "spans": [ + { + "bbox": [ + 46, + 399, + 288, + 698 + ], + "type": "text", + "content": "A refinement network is developed as a post-processing module to further mitigate the noise and artifacts in the reconstructed frames. NAFNet [4] is designed for image restoration. It is structured as a U-Net, arranging NAF blocks to allow the network to progressively downsample the input for semantic feature extraction and then progressively upsample it to restore the output. NAFNet provides additional skip connections that preserve information and allow feature reusability from earlier layers. NAF blocks are structured as original ResNet blocks, with additional layers such as layer normalization and channel attention. We adopt the NAFNet basic concept, and further modify making two major additions: 1) spatial attention [40] as shown in Figure 5 and 2) perceptual loss [32]. The reason for choosing spatial attention is that most of the noise is located in motion areas instead of being ubiquitous across the image. Spatial attention allows the network to focus on these noisy regions of the image, which contributes to better denoising results. Moreover, the rationale for choosing the perceptual loss is that even though the proposed network does well in removing the pixel-level saturation noise, it also introduces blurriness or texture degradation. Perceptual loss helps to recover the image quality in a way that better aligns with the human vision system. The losses applied in this paper are summarized as:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 83, + 701, + 287, + 714 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 83, + 701, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 83, + 701, + 287, + 714 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {s}} = w _ {l} \\mathcal {L} _ {1} + w _ {\\mathrm {V G G}} \\mathcal {L} _ {\\mathrm {V G G}} + w _ {\\text {G r a m}} \\mathcal {L} _ {\\text {G r a m}}, \\tag {14}", + "image_path": "4cd2e56cb186a9965fff1ded451648515374308eb2bb13f4db37926d75c31c14.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 72, + 547, + 253 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 547, + 253 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 547, + 253 + ], + "type": "text", + "content": "where the " + }, + { + "bbox": [ + 305, + 72, + 547, + 253 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_1" + }, + { + "bbox": [ + 305, + 72, + 547, + 253 + ], + "type": "text", + "content": " loss is the pixel-wise difference between the reconstructed RGB image and ground truth image; the " + }, + { + "bbox": [ + 305, + 72, + 547, + 253 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{VGG}}" + }, + { + "bbox": [ + 305, + 72, + 547, + 253 + ], + "type": "text", + "content": " loss is calculated as " + }, + { + "bbox": [ + 305, + 72, + 547, + 253 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{VGG}} = \\frac{1}{L}\\sum_{l=1}^{L}\\alpha_l\\|\\Psi_l(\\boldsymbol{I}_{\\mathrm{Recon}}) - \\Psi_l(\\boldsymbol{I}_{\\mathrm{GT}})\\|_1;\\Psi_l(\\boldsymbol{I})" + }, + { + "bbox": [ + 305, + 72, + 547, + 253 + ], + "type": "text", + "content": " is the feature tensor from the " + }, + { + "bbox": [ + 305, + 72, + 547, + 253 + ], + "type": "inline_equation", + "content": "l_{\\mathrm{th}}" + }, + { + "bbox": [ + 305, + 72, + 547, + 253 + ], + "type": "text", + "content": " layer of an ImageNet pre-trained VGG-19 network generated by passing the image " + }, + { + "bbox": [ + 305, + 72, + 547, + 253 + ], + "type": "inline_equation", + "content": "\\boldsymbol{I}" + }, + { + "bbox": [ + 305, + 72, + 547, + 253 + ], + "type": "text", + "content": " through the network. In this case, " + }, + { + "bbox": [ + 305, + 72, + 547, + 253 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 305, + 72, + 547, + 253 + ], + "type": "text", + "content": " specifies the total number of layers in the VGG-19 network, while " + }, + { + "bbox": [ + 305, + 72, + 547, + 253 + ], + "type": "inline_equation", + "content": "\\alpha_l" + }, + { + "bbox": [ + 305, + 72, + 547, + 253 + ], + "type": "text", + "content": " is the hyperparameter assigned to each layer. Finally, the Gram loss [11] is computed as the L2 difference between the autocorrelation of the VGG-19 features of the reconstructed image and ground truth image with the formulation: " + }, + { + "bbox": [ + 305, + 72, + 547, + 253 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{Gram}} = \\frac{1}{L}\\sum_{l=1}^{L}\\alpha_l\\|\\boldsymbol{M}_l(\\boldsymbol{I}_{\\mathrm{Recon}}) - \\boldsymbol{M}_l(\\boldsymbol{I}_{\\mathrm{GT}})\\|_2" + }, + { + "bbox": [ + 305, + 72, + 547, + 253 + ], + "type": "text", + "content": ", where given an image " + }, + { + "bbox": [ + 305, + 72, + 547, + 253 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 305, + 72, + 547, + 253 + ], + "type": "text", + "content": ", the Gram matrix is computed for every layer " + }, + { + "bbox": [ + 305, + 72, + 547, + 253 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 305, + 72, + 547, + 253 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 305, + 72, + 547, + 253 + ], + "type": "inline_equation", + "content": "\\boldsymbol{M}_l = (\\Psi_l(\\boldsymbol{I}))^\\top(\\Psi_l(\\boldsymbol{I}))" + }, + { + "bbox": [ + 305, + 72, + 547, + 253 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 261, + 388, + 274 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 261, + 388, + 274 + ], + "spans": [ + { + "bbox": [ + 306, + 261, + 388, + 274 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 279, + 547, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 279, + 547, + 529 + ], + "spans": [ + { + "bbox": [ + 305, + 279, + 547, + 529 + ], + "type": "text", + "content": "We compare our proposed method with state-of-the-art EVS+CIS video reconstruction methods: EDI [27], AKF [39], Time Lens [37], CBMNet [23], EvUnroll [8] and RE-FID [36]. Among them, EvUnroll proposes an RS correction based reconstruction and REFID combines image de-burring with EVS based frame interpolation. We also show the results of the CIS-only interpolation method FILM [32] as reference. Before optimization, " + }, + { + "bbox": [ + 305, + 279, + 547, + 529 + ], + "type": "inline_equation", + "content": "i_{\\mathrm{pd}}" + }, + { + "bbox": [ + 305, + 279, + 547, + 529 + ], + "type": "text", + "content": " are initialized according to the first CIS image DN and the inverse photocurrent-DN relation. The elements in the weighing matrices for CIS, EVS terms are set to 1. The performance of the interpolation methods are evaluated both quantitatively and qualitatively, based on the metrics of Peak Signal-to-Noise Ratio (PSNR), LPIPS [42], and pixel-wise reconstruction error on DN. Furthermore, to study the performance degradation concerning ambient illuminance and motion speed, we measured a test set utilizing a rotating disk in a controlled environment. We analyze the Blurred Edge Width (BEW) [5] of the reconstructed frames - a larger BEW indicates a blurrier image and, therefore, a worse deblurring result." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 539, + 368, + 550 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 539, + 368, + 550 + ], + "spans": [ + { + "bbox": [ + 306, + 539, + 368, + 550 + ], + "type": "text", + "content": "5.1. Datasets" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 558, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 558, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 558, + 547, + 713 + ], + "type": "text", + "content": "We evaluate the proposed method using the simulation data based on the CIS-EVS hybrid sensor simulator [26] which considers " + }, + { + "bbox": [ + 305, + 558, + 547, + 713 + ], + "type": "inline_equation", + "content": "V_{\\mathrm{FE}}" + }, + { + "bbox": [ + 305, + 558, + 547, + 713 + ], + "type": "text", + "content": " delay, readout latency, and refractory period. Our measurement data was collected from the hybrid CIS-EVS sensor [14, 15]. Here, only off-pixel times- tamp data was captured. Two measured datasets were created for evaluation: The natural scene dataset focuses on indoor scenarios where EVS is significantly influenced by pixel latency. The other dataset captured a rotating disk with a Siemens star pattern under different ambient illumination levels and rotating speeds. In addition to the above datasets specifically captured for this work's evaluation, we also test our method using the publicly available HS-ERGB" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 731, + 318, + 740 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 731, + 318, + 740 + ], + "spans": [ + { + "bbox": [ + 293, + 731, + 318, + 740 + ], + "type": "text", + "content": "25176" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 48, + 112, + 547, + 180 + ], + "blocks": [ + { + "bbox": [ + 46, + 71, + 547, + 106 + ], + "lines": [ + { + "bbox": [ + 46, + 71, + 547, + 106 + ], + "spans": [ + { + "bbox": [ + 46, + 71, + 547, + 106 + ], + "type": "text", + "content": "Table 1. Quantitative comparison of state-of-the-art video reconstruction methods on the proposed simulation dataset in terms of PSNR (in dB, higher is better) and LPIPS (a dimensionless quantity, lower is better). Each row shows results for a particular scene. Results from different methods are listed column-wise. The first and second places are highlighted with bold underline and bold, respectively." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 48, + 112, + 547, + 180 + ], + "lines": [ + { + "bbox": [ + 48, + 112, + 547, + 180 + ], + "spans": [ + { + "bbox": [ + 48, + 112, + 547, + 180 + ], + "type": "table", + "html": "
Ours w/o refinementOursEDIAKFTime LensFILMCBMNetEvUnrollREFID
ScenePSNR ↑LPIPS ↓PSNR ↑LPIPS ↓PSNR ↑LPIPS ↓PSNR ↑LPIPS ↓PSNR ↑LPIPS ↓PSNR ↑LPIPS ↓PSNR ↑LPIPS ↓PSNR ↑LPIPS ↓PSNR ↑LPIPS ↓
basketball24.010.21727.230.15522.490.22015.630.28023.650.24023.860.22024.510.25223.130.21822.380.223
checkerboard24.230.20328.240.18720.480.21819.560.26122.600.23923.020.21622.760.21223.840.20421.230.235
slingshot egg25.230.10032.850.10624.950.16415.590.11025.190.24424.040.20322.760.23424.610.21023.920.201
running man24.600.16424.720.18922.290.13520.710.18925.010.11725.060.08925.160.06923.040.11723.520.163
fan22.400.18924.500.16318.170.25510.780.18420.640.23318.930.18421.280.20419.640.19018.300.198
Average24.090.17527.510.1621.680.19816.450.20523.420.21522.980.18223.290.19422.850.18821.870.204
", + "image_path": "6a1cfcfbeaaec01dc9072d902d370e628e102cfe0ea3a14aca6489a36dd83142.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 70, + 210, + 523, + 266 + ], + "blocks": [ + { + "bbox": [ + 53, + 190, + 538, + 202 + ], + "lines": [ + { + "bbox": [ + 53, + 190, + 538, + 202 + ], + "spans": [ + { + "bbox": [ + 53, + 190, + 538, + 202 + ], + "type": "text", + "content": "Table 2. Quantitative results of ablation study. The first and second places are highlighted with bold underline and bold, respectively." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 70, + 210, + 523, + 266 + ], + "lines": [ + { + "bbox": [ + 70, + 210, + 523, + 266 + ], + "spans": [ + { + "bbox": [ + 70, + 210, + 523, + 266 + ], + "type": "table", + "html": "
without compensationwith VFE delay compensationwith RL+RP compensationwith VFE+RL+RP compensationwith VFE+RL+RP compensation and refinement
ScenePSNR ↑LPIPS ↓PSNR ↑LPIPS ↓PSNR ↑LPIPS ↓PSNR ↑LPIPS ↓PSNR ↑LPIPS ↓
basketball22.950.31623.470.30623.500.30723.560.30223.640.295
slingshot egg23.570.09924.000.09923.600.09923.710.10229.040.106
", + "image_path": "e8d383a756fe1bcf0a009e9fefde5c2ef96c829688b6f2dbc11ef1f15a803c21.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 50, + 275, + 545, + 439 + ], + "blocks": [ + { + "bbox": [ + 50, + 275, + 545, + 439 + ], + "lines": [ + { + "bbox": [ + 50, + 275, + 545, + 439 + ], + "spans": [ + { + "bbox": [ + 50, + 275, + 545, + 439 + ], + "type": "image", + "image_path": "29969f719a04c87b0e6c3220e145fad3cf4ee891447d76451681b3ebf152ada7.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 446, + 547, + 470 + ], + "lines": [ + { + "bbox": [ + 46, + 446, + 547, + 470 + ], + "spans": [ + { + "bbox": [ + 46, + 446, + 547, + 470 + ], + "type": "text", + "content": "Figure 6. Qualitative comparison of state-of-the-art video reconstruction methods on the proposed measured datasets in terms of image quality. The images have been magnified for optimal viewing quality." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 48, + 483, + 141, + 560 + ], + "blocks": [ + { + "bbox": [ + 48, + 483, + 141, + 560 + ], + "lines": [ + { + "bbox": [ + 48, + 483, + 141, + 560 + ], + "spans": [ + { + "bbox": [ + 48, + 483, + 141, + 560 + ], + "type": "image", + "image_path": "8f9a77d30aee6659949c823f28d806ae217c9ca3960f79364c9fe5d1f26a5d6f.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 577, + 548, + 611 + ], + "lines": [ + { + "bbox": [ + 46, + 577, + 548, + 611 + ], + "spans": [ + { + "bbox": [ + 46, + 577, + 548, + 611 + ], + "type": "text", + "content": "Figure 7. Frame-wise, quantitative comparison of state-of-the-art video reconstruction methods in terms of PSNR for varying scenarios of the simulation dataset. Note that some algorithms exhibit pronounced discontinuities in the observed PSNR curves indicating the presence of video glitches." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 144, + 486, + 228, + 560 + ], + "blocks": [ + { + "bbox": [ + 144, + 486, + 228, + 560 + ], + "lines": [ + { + "bbox": [ + 144, + 486, + 228, + 560 + ], + "spans": [ + { + "bbox": [ + 144, + 486, + 228, + 560 + ], + "type": "image", + "image_path": "40d56ba35ed8aa577db3383cf57a6cfa73d930d79242859c7d50e167ac2cb6b1.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 230, + 485, + 315, + 560 + ], + "blocks": [ + { + "bbox": [ + 230, + 485, + 315, + 560 + ], + "lines": [ + { + "bbox": [ + 230, + 485, + 315, + 560 + ], + "spans": [ + { + "bbox": [ + 230, + 485, + 315, + 560 + ], + "type": "image", + "image_path": "918086ff77003e143e0ae0536c41b0aa403cd75bed8fcf0e47d481014d414e61.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 317, + 485, + 402, + 560 + ], + "blocks": [ + { + "bbox": [ + 317, + 485, + 402, + 560 + ], + "lines": [ + { + "bbox": [ + 317, + 485, + 402, + 560 + ], + "spans": [ + { + "bbox": [ + 317, + 485, + 402, + 560 + ], + "type": "image", + "image_path": "1ec2efe781dc867223e22b3a398dd3345f5e0f4b5f5fdb6dffb6aefcf94ffa80.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 403, + 484, + 487, + 560 + ], + "blocks": [ + { + "bbox": [ + 403, + 484, + 487, + 560 + ], + "lines": [ + { + "bbox": [ + 403, + 484, + 487, + 560 + ], + "spans": [ + { + "bbox": [ + 403, + 484, + 487, + 560 + ], + "type": "image", + "image_path": "7628439d5201acf790fe3536544619871746a1aa2c0c5fca2e5906fbc7bbce65.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 493, + 483, + 542, + 562 + ], + "blocks": [ + { + "bbox": [ + 493, + 483, + 542, + 562 + ], + "lines": [ + { + "bbox": [ + 493, + 483, + 542, + 562 + ], + "spans": [ + { + "bbox": [ + 493, + 483, + 542, + 562 + ], + "type": "image", + "image_path": "4cb9e1d501fb3607c001cf4c54594e197aa4e1596d7da03ebc296894e1f0ffa9.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 631, + 288, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 631, + 288, + 693 + ], + "spans": [ + { + "bbox": [ + 46, + 631, + 288, + 693 + ], + "type": "text", + "content": "dataset based on a EVS camera that is collocated next to a CIS camera. Since the paper focuses on mitigating artifacts due to pixel latency, two indoor scenarios (\"spinning plate\" and \"spinning umbrella\") were selected. More details of the dataset used are provided in the supplementary material." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 631, + 364, + 643 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 631, + 364, + 643 + ], + "spans": [ + { + "bbox": [ + 306, + 631, + 364, + 643 + ], + "type": "text", + "content": "5.2. Results" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 654, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 654, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 305, + 654, + 547, + 715 + ], + "type": "text", + "content": "Image Quality Table 1 shows quantitative results for the reconstructed video. Results in PSNR and LPIPS show that the proposed method outperforms other state-of-the-art methods in most of the scenes, especially with fast motion (\"slingshot egg\") or complicated rotation motion (\"basket-" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 731, + 318, + 740 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 731, + 318, + 740 + ], + "spans": [ + { + "bbox": [ + 293, + 731, + 318, + 740 + ], + "type": "text", + "content": "25177" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 286 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 286 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 286 + ], + "type": "text", + "content": "ball\", \"fan\"). As expected the refinement network which aims to alleviate artifacts and improve image quality reports a higher score. Snippets of the reconstructed videos are presented in Figure 1 and Figure 6 for simulated and measured data, respectively. Comprehensive input images are provided in the supplementary material. Thanks to the explicit EVS latency modeling and compensation, our results avoid obvious shadowing/ghosting artifacts - especially if there is fast motion under challenging lighting conditions. It is also noted that the proposed method successfully corrects the RS effect, as shown in the \"basketball\" scene. Learning-based methods face difficulties, as latency and the RS effect have not been considered in model design and training. The CIS-only method FILM provides natural image quality with low noise but fails when the objects are too fast to be captured by CIS, as shown in the \"slingshot egg\" scene. For the \"basketball\" scene, our method shows an advantage in details and sharpness compared to FILM." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 288, + 289, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 288, + 289, + 456 + ], + "spans": [ + { + "bbox": [ + 46, + 288, + 289, + 456 + ], + "type": "text", + "content": "In specific scenarios, such as the \"spinning umbrella\" scene within the HS-ERGB dataset and the \"running man\" scene in the simulation dataset, our method did not yield improved image quality. This observation can be attributed to several factors. Firstly, the HS-ERGB dataset employs a different EVS with different pixel and readout latency as well as refractory period. Secondly, this dataset is based on a dual camera instead of a hybrid sensor so the spatiotemporal alignment may differ. Thirdly, the scenes in question involve relatively slower object movement speeds resulting in reduced ghosting. In essence, the dynamic characteristics of the EVS, the input signal frequency range, and the reconstruction algorithm influence the quality of the reconstructed images." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 459, + 287, + 531 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 459, + 287, + 531 + ], + "spans": [ + { + "bbox": [ + 46, + 459, + 287, + 531 + ], + "type": "text", + "content": "Video Glitch Effect To evaluate the glitch issue, we compute the PSNR for each reconstructed frame as shown in Figure 7. Overall, our method gives temporally more stable PSNR curves, which stem from the proposed row-based reconstruction scheme. It is also noted that the refinement network contributes significantly to higher PSNR." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 533, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 533, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 533, + 289, + 715 + ], + "type": "text", + "content": "Pixel-wise Error Analysis We select a pixel within the region of interest (indicated as the yellow box in Figure 1(b)) and verify the pixel-wise reconstruction accuracy. As shown in Figure 8(a), with the help of pixel latency compensation the reconstructed " + }, + { + "bbox": [ + 46, + 533, + 289, + 715 + ], + "type": "inline_equation", + "content": "V_{\\mathrm{DC}}" + }, + { + "bbox": [ + 46, + 533, + 289, + 715 + ], + "type": "text", + "content": " after optimization shows a sharper peak (with " + }, + { + "bbox": [ + 46, + 533, + 289, + 715 + ], + "type": "inline_equation", + "content": "250~\\mu \\mathrm{s}" + }, + { + "bbox": [ + 46, + 533, + 289, + 715 + ], + "type": "text", + "content": " duration) compared to " + }, + { + "bbox": [ + 46, + 533, + 289, + 715 + ], + "type": "inline_equation", + "content": "V_{\\mathrm{FE}}" + }, + { + "bbox": [ + 46, + 533, + 289, + 715 + ], + "type": "text", + "content": " (with over " + }, + { + "bbox": [ + 46, + 533, + 289, + 715 + ], + "type": "inline_equation", + "content": "1\\mathrm{ms}" + }, + { + "bbox": [ + 46, + 533, + 289, + 715 + ], + "type": "text", + "content": " duration). The high-frequency fluctuations on ground truth " + }, + { + "bbox": [ + 46, + 533, + 289, + 715 + ], + "type": "inline_equation", + "content": "V_{\\mathrm{DC}}" + }, + { + "bbox": [ + 46, + 533, + 289, + 715 + ], + "type": "text", + "content": " curve cannot be reconstructed due to the " + }, + { + "bbox": [ + 46, + 533, + 289, + 715 + ], + "type": "inline_equation", + "content": "V_{\\mathrm{FE}}" + }, + { + "bbox": [ + 46, + 533, + 289, + 715 + ], + "type": "text", + "content": " bandwidth limit. An overshoot is observed after the peak resulting in a dark tail along the pellet's trajectory, as shown in Figure 1(b). This may stem from 1) the " + }, + { + "bbox": [ + 46, + 533, + 289, + 715 + ], + "type": "inline_equation", + "content": "V_{\\mathrm{FE}}" + }, + { + "bbox": [ + 46, + 533, + 289, + 715 + ], + "type": "text", + "content": " bandwidth, 2) the approximation error in equation (12), 3) EVS measurement noise, and 4) choice of weighing matrices for CIS and EVS. Figure 8(b), indicates that the proposed method gives an excellent reconstruction of the pel" + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 313, + 75, + 545, + 148 + ], + "blocks": [ + { + "bbox": [ + 313, + 75, + 545, + 148 + ], + "lines": [ + { + "bbox": [ + 313, + 75, + 545, + 148 + ], + "spans": [ + { + "bbox": [ + 313, + 75, + 545, + 148 + ], + "type": "image", + "image_path": "e59a31c0c039cf129bfcd2d939a0cbca7716bc9bbbda3418fca6d19558c19005.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 318, + 148, + 331, + 158 + ], + "lines": [ + { + "bbox": [ + 318, + 148, + 331, + 158 + ], + "spans": [ + { + "bbox": [ + 318, + 148, + 331, + 158 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 310, + 163, + 544, + 235 + ], + "blocks": [ + { + "bbox": [ + 310, + 163, + 544, + 235 + ], + "lines": [ + { + "bbox": [ + 310, + 163, + 544, + 235 + ], + "spans": [ + { + "bbox": [ + 310, + 163, + 544, + 235 + ], + "type": "image", + "image_path": "0903f66bcfe91b0e67ca6280b8f589197a8c8e9f970927d46cdce5238e0c2d70.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 318, + 235, + 330, + 245 + ], + "lines": [ + { + "bbox": [ + 318, + 235, + 330, + 245 + ], + "spans": [ + { + "bbox": [ + 318, + 235, + 330, + 245 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 421, + 236, + 457, + 245 + ], + "lines": [ + { + "bbox": [ + 421, + 236, + 457, + 245 + ], + "spans": [ + { + "bbox": [ + 421, + 236, + 457, + 245 + ], + "type": "text", + "content": "time " + }, + { + "bbox": [ + 421, + 236, + 457, + 245 + ], + "type": "inline_equation", + "content": "(\\mu s)" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 313, + 361, + 425, + 437 + ], + "blocks": [ + { + "bbox": [ + 304, + 261, + 547, + 350 + ], + "lines": [ + { + "bbox": [ + 304, + 261, + 547, + 350 + ], + "spans": [ + { + "bbox": [ + 304, + 261, + 547, + 350 + ], + "type": "text", + "content": "Figure 8. Pixel-wise reconstruction results for the \"slingshot egg\" scene. The selected pixel is located at the path of the pellet, in the yellow box of Figure 1. The refinement network is not used here. (a) Comparison of the ground truth " + }, + { + "bbox": [ + 304, + 261, + 547, + 350 + ], + "type": "inline_equation", + "content": "V_{\\mathrm{DC}}" + }, + { + "bbox": [ + 304, + 261, + 547, + 350 + ], + "type": "text", + "content": ", reconstructed " + }, + { + "bbox": [ + 304, + 261, + 547, + 350 + ], + "type": "inline_equation", + "content": "V_{\\mathrm{FE}}" + }, + { + "bbox": [ + 304, + 261, + 547, + 350 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 261, + 547, + 350 + ], + "type": "inline_equation", + "content": "V_{\\mathrm{DC}}" + }, + { + "bbox": [ + 304, + 261, + 547, + 350 + ], + "type": "text", + "content": ", where the ground truth " + }, + { + "bbox": [ + 304, + 261, + 547, + 350 + ], + "type": "inline_equation", + "content": "V_{\\mathrm{DC}}" + }, + { + "bbox": [ + 304, + 261, + 547, + 350 + ], + "type": "text", + "content": " is computed from the ground truth DN of the pixel at 10000 FPS. (b) Comparison of reconstructed pixel-wise DN of state-of-the-art video reconstruction methods." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 361, + 425, + 437 + ], + "lines": [ + { + "bbox": [ + 313, + 361, + 425, + 437 + ], + "spans": [ + { + "bbox": [ + 313, + 361, + 425, + 437 + ], + "type": "image", + "image_path": "5f28b27d9ca784fa765a3b8d41656f9cc8c4706ccc763c11eeca66738fd940fb.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 334, + 438, + 404, + 446 + ], + "lines": [ + { + "bbox": [ + 334, + 438, + 404, + 446 + ], + "spans": [ + { + "bbox": [ + 334, + 438, + 404, + 446 + ], + "type": "text", + "content": "(a) with no compensation" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 425, + 361, + 536, + 437 + ], + "blocks": [ + { + "bbox": [ + 425, + 361, + 536, + 437 + ], + "lines": [ + { + "bbox": [ + 425, + 361, + 536, + 437 + ], + "spans": [ + { + "bbox": [ + 425, + 361, + 536, + 437 + ], + "type": "image", + "image_path": "9623e737ab247512b849601955966bf3cd145117bc583e92d9b67a0abba37a04.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 434, + 438, + 525, + 446 + ], + "lines": [ + { + "bbox": [ + 434, + 438, + 525, + 446 + ], + "spans": [ + { + "bbox": [ + 434, + 438, + 525, + 446 + ], + "type": "text", + "content": "(b) with " + }, + { + "bbox": [ + 434, + 438, + 525, + 446 + ], + "type": "inline_equation", + "content": "V_{\\mathrm{FE}}" + }, + { + "bbox": [ + 434, + 438, + 525, + 446 + ], + "type": "text", + "content": " delay compensation" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 313, + 448, + 425, + 523 + ], + "blocks": [ + { + "bbox": [ + 313, + 448, + 425, + 523 + ], + "lines": [ + { + "bbox": [ + 313, + 448, + 425, + 523 + ], + "spans": [ + { + "bbox": [ + 313, + 448, + 425, + 523 + ], + "type": "image", + "image_path": "84c86a856507df835cd2d02721b030013398fa71167135634dc9ac8c1753bd57.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 327, + 525, + 410, + 533 + ], + "lines": [ + { + "bbox": [ + 327, + 525, + 410, + 533 + ], + "spans": [ + { + "bbox": [ + 327, + 525, + 410, + 533 + ], + "type": "text", + "content": "(c) with RL+RP compensation" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 304, + 536, + 547, + 568 + ], + "lines": [ + { + "bbox": [ + 304, + 536, + 547, + 568 + ], + "spans": [ + { + "bbox": [ + 304, + 536, + 547, + 568 + ], + "type": "text", + "content": "Figure 9. Qualitative results of ablation study on the \"basketball\" scene. Refer to Figure 1(a) for \"with " + }, + { + "bbox": [ + 304, + 536, + 547, + 568 + ], + "type": "inline_equation", + "content": "V_{\\mathrm{FE}} + \\mathrm{RL} + \\mathrm{RP}" + }, + { + "bbox": [ + 304, + 536, + 547, + 568 + ], + "type": "text", + "content": " compensation and refinement\" result." + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 425, + 448, + 536, + 523 + ], + "blocks": [ + { + "bbox": [ + 425, + 448, + 536, + 523 + ], + "lines": [ + { + "bbox": [ + 425, + 448, + 536, + 523 + ], + "spans": [ + { + "bbox": [ + 425, + 448, + 536, + 523 + ], + "type": "image", + "image_path": "1d06d331807319f189d35f3097b64f7ee960371b235c277a5d356c98d04ffa1c.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 432, + 525, + 529, + 534 + ], + "lines": [ + { + "bbox": [ + 432, + 525, + 529, + 534 + ], + "spans": [ + { + "bbox": [ + 432, + 525, + 529, + 534 + ], + "type": "text", + "content": "(d) with " + }, + { + "bbox": [ + 432, + 525, + 529, + 534 + ], + "type": "inline_equation", + "content": "V_{\\mathrm{FE}} + \\mathrm{RL} + \\mathrm{RP}" + }, + { + "bbox": [ + 432, + 525, + 529, + 534 + ], + "type": "text", + "content": " compensation" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "bbox": [ + 305, + 591, + 453, + 603 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 591, + 453, + 603 + ], + "spans": [ + { + "bbox": [ + 305, + 591, + 453, + 603 + ], + "type": "text", + "content": "let's peak DN despite the overshoot." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": "Ablation Study We ablate the refinement network, " + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "inline_equation", + "content": "V_{\\mathrm{FE}}" + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": " delay compensation, readout latency (RL) and refractory period (RP) compensation of our method. The numerical evaluation are subject to selected ROIs on motion region. To ablate " + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "inline_equation", + "content": "V_{\\mathrm{FE}}" + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": " delay compensation, we set the time constant " + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\tau = 0" + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": " such that " + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "inline_equation", + "content": "V_{\\mathrm{FE}} = V_{\\mathrm{DC}}" + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": " in Eq. (7). To ablate readout latency and refractory period compensation, we use Eq. (8) instead of Eq. (12) in the EVS measurement model. Results are summarized in Table 2 and Figure 9. The con" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 731, + 318, + 740 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 731, + 318, + 740 + ], + "spans": [ + { + "bbox": [ + 293, + 731, + 318, + 740 + ], + "type": "text", + "content": "25178" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 263 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 263 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 263 + ], + "type": "text", + "content": "Figuration using all modules gives significantly better visual and slightly better numerical results in PSNR compared to having latency compensation and refinement disabled. Disabling " + }, + { + "bbox": [ + 46, + 72, + 289, + 263 + ], + "type": "inline_equation", + "content": "V_{\\mathrm{FE}}" + }, + { + "bbox": [ + 46, + 72, + 289, + 263 + ], + "type": "text", + "content": " or RL+RP compensation significantly worsens the image quality, as severe shadowing artifacts are observed. As for LPIPS in \"slingshot egg\" scene, the difference is not significant due to the small area of pellet ROI. It is also noted that adding RL+RP compensation leads to improvement in the \"basketball\" but degrades the performance in \"slingshot egg\" when comparing with " + }, + { + "bbox": [ + 46, + 72, + 289, + 263 + ], + "type": "inline_equation", + "content": "V_{\\mathrm{FE}}" + }, + { + "bbox": [ + 46, + 72, + 289, + 263 + ], + "type": "text", + "content": " delay compensation only. This is because RL+RP compensation is based on the assumption that the slope of VFE is locally constant for a short period. In cases such as \"slingshot egg\", the assumption may not hold due to close positive and negative events. An ablation study against the use of NAFNet-only is included in supplementary material." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 266, + 287, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 266, + 287, + 422 + ], + "spans": [ + { + "bbox": [ + 46, + 266, + 287, + 422 + ], + "type": "text", + "content": "Performance Regarding Illuminance and Speed The proposed method is further investigated under a controlled environment using varying illuminance and object motion speed. BEW [5] is used to evaluate the image quality by measuring edge sharpness. Figure 10(a)-(h) presents the comparison of the state-of-the-art methods on a rotating disk under " + }, + { + "bbox": [ + 46, + 266, + 287, + 422 + ], + "type": "inline_equation", + "content": "1000\\mathrm{lx}" + }, + { + "bbox": [ + 46, + 266, + 287, + 422 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 266, + 287, + 422 + ], + "type": "inline_equation", + "content": "292\\mathrm{rpm}" + }, + { + "bbox": [ + 46, + 266, + 287, + 422 + ], + "type": "text", + "content": ". It is observed that the proposed method reconstructs sharper edges compared to the others. Other EVS-based methods struggle resolving blurry edges in presence of fast motion and also the CIS-based FILM method fails to generate a sharp reconstruction. Figure 10(i) shows that the proposed method reports the lowest BEW score (0.7) and has the sharpest edge profile." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 423, + 288, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 423, + 288, + 495 + ], + "spans": [ + { + "bbox": [ + 46, + 423, + 288, + 495 + ], + "type": "text", + "content": "The BEW scores under different speed and illumination conditions are summarized in Figure 11. Figure 11(a) shows that the BEW increases with faster rotation speeds. In Figure 11(b), the lower ambient light condition leads to higher " + }, + { + "bbox": [ + 46, + 423, + 288, + 495 + ], + "type": "inline_equation", + "content": "V_{\\mathrm{FE}}" + }, + { + "bbox": [ + 46, + 423, + 288, + 495 + ], + "type": "text", + "content": " latency degrading image quality. Our method achieves the lowest BEW throughout all illuminance levels." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 505, + 119, + 517 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 505, + 119, + 517 + ], + "spans": [ + { + "bbox": [ + 47, + 505, + 119, + 517 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 521, + 289, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 521, + 289, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 521, + 289, + 713 + ], + "type": "text", + "content": "An EVS-assisted framework to solve motion blur, rolling-shutter artifacts and video-frame interpolation in a joint manner was presented. Existing methods use either a simple double integral model or a learning-based model, neither of which explicitly address EVS sensor imperfections such as EVS pixel latency, readout latency or refractory period. This work overcomes these limitations by explicitly modeling these characteristics. The compensation of readout latency and refractory period is enabled by having access to in-pixel and off-pixel time-stamps. The nonlinear inverse model was solved through joint graph optimization of subsequent frames utilizing row-by-row exposure information to overcome rolling-shutter artifacts and possible glitches between subsequent frames. A refinement network was proposed to further improve the image quality by using specifically designed spatial attention blocks. We showed that" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 310, + 69, + 370, + 125 + ], + "blocks": [ + { + "bbox": [ + 310, + 69, + 370, + 125 + ], + "lines": [ + { + "bbox": [ + 310, + 69, + 370, + 125 + ], + "spans": [ + { + "bbox": [ + 310, + 69, + 370, + 125 + ], + "type": "image", + "image_path": "ac5f187dbcf96d40eb3e6c02796a0804c55602294a83345f77bf9c0adebac150.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 310, + 126, + 371, + 133 + ], + "lines": [ + { + "bbox": [ + 310, + 126, + 371, + 133 + ], + "spans": [ + { + "bbox": [ + 310, + 126, + 371, + 133 + ], + "type": "text", + "content": "(a) Ours w/o refinement" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 370, + 70, + 429, + 125 + ], + "blocks": [ + { + "bbox": [ + 370, + 70, + 429, + 125 + ], + "lines": [ + { + "bbox": [ + 370, + 70, + 429, + 125 + ], + "spans": [ + { + "bbox": [ + 370, + 70, + 429, + 125 + ], + "type": "image", + "image_path": "7c824604139f204e8ee97040bd04c00e5bf0fe194137654db705377c0490709d.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 389, + 126, + 408, + 133 + ], + "lines": [ + { + "bbox": [ + 389, + 126, + 408, + 133 + ], + "spans": [ + { + "bbox": [ + 389, + 126, + 408, + 133 + ], + "type": "text", + "content": "(b) EDI" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 429, + 70, + 487, + 125 + ], + "blocks": [ + { + "bbox": [ + 429, + 70, + 487, + 125 + ], + "lines": [ + { + "bbox": [ + 429, + 70, + 487, + 125 + ], + "spans": [ + { + "bbox": [ + 429, + 70, + 487, + 125 + ], + "type": "image", + "image_path": "0bb38fd3e7a4c604a13814662e2932089c2f12959cde9fbdff3d87f149a3b387.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 449, + 126, + 467, + 133 + ], + "lines": [ + { + "bbox": [ + 449, + 126, + 467, + 133 + ], + "spans": [ + { + "bbox": [ + 449, + 126, + 467, + 133 + ], + "type": "text", + "content": "(c)AKF" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 487, + 70, + 547, + 125 + ], + "blocks": [ + { + "bbox": [ + 487, + 70, + 547, + 125 + ], + "lines": [ + { + "bbox": [ + 487, + 70, + 547, + 125 + ], + "spans": [ + { + "bbox": [ + 487, + 70, + 547, + 125 + ], + "type": "image", + "image_path": "5fbd68f799e6321bfcf5b89c9e7134912537c696687e899a5325c2d6dcd14380.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 498, + 126, + 534, + 133 + ], + "lines": [ + { + "bbox": [ + 498, + 126, + 534, + 133 + ], + "spans": [ + { + "bbox": [ + 498, + 126, + 534, + 133 + ], + "type": "text", + "content": "(d) Time Lens" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 311, + 134, + 370, + 190 + ], + "blocks": [ + { + "bbox": [ + 311, + 134, + 370, + 190 + ], + "lines": [ + { + "bbox": [ + 311, + 134, + 370, + 190 + ], + "spans": [ + { + "bbox": [ + 311, + 134, + 370, + 190 + ], + "type": "image", + "image_path": "181c11d8c4ef18733787bb7d61eff62d02de1f62ec20e8ef54cd9a125c775dbc.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 324, + 191, + 347, + 198 + ], + "lines": [ + { + "bbox": [ + 324, + 191, + 347, + 198 + ], + "spans": [ + { + "bbox": [ + 324, + 191, + 347, + 198 + ], + "type": "text", + "content": "(e) FILM" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 370, + 134, + 429, + 190 + ], + "blocks": [ + { + "bbox": [ + 370, + 134, + 429, + 190 + ], + "lines": [ + { + "bbox": [ + 370, + 134, + 429, + 190 + ], + "spans": [ + { + "bbox": [ + 370, + 134, + 429, + 190 + ], + "type": "image", + "image_path": "8adac60a1e6e2c7288c4b69289e84a65e56b98a53137230ba2e0612898e4638c.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 383, + 191, + 412, + 198 + ], + "lines": [ + { + "bbox": [ + 383, + 191, + 412, + 198 + ], + "spans": [ + { + "bbox": [ + 383, + 191, + 412, + 198 + ], + "type": "text", + "content": "(f) CBMNet" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 429, + 134, + 487, + 190 + ], + "blocks": [ + { + "bbox": [ + 429, + 134, + 487, + 190 + ], + "lines": [ + { + "bbox": [ + 429, + 134, + 487, + 190 + ], + "spans": [ + { + "bbox": [ + 429, + 134, + 487, + 190 + ], + "type": "image", + "image_path": "50d88950284ba515c9491230e2725d201856e92e35fdb6f1c99eb6254b195442.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 443, + 191, + 473, + 198 + ], + "lines": [ + { + "bbox": [ + 443, + 191, + 473, + 198 + ], + "spans": [ + { + "bbox": [ + 443, + 191, + 473, + 198 + ], + "type": "text", + "content": "(g) EvUnroll" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 487, + 134, + 547, + 190 + ], + "blocks": [ + { + "bbox": [ + 487, + 134, + 547, + 190 + ], + "lines": [ + { + "bbox": [ + 487, + 134, + 547, + 190 + ], + "spans": [ + { + "bbox": [ + 487, + 134, + 547, + 190 + ], + "type": "image", + "image_path": "d58e086e10466bcdf13e7306801568e1f12c80e269db05ce17912a301ffd7e7c.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 503, + 191, + 529, + 198 + ], + "lines": [ + { + "bbox": [ + 503, + 191, + 529, + 198 + ], + "spans": [ + { + "bbox": [ + 503, + 191, + 529, + 198 + ], + "type": "text", + "content": "(h) REFID" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 312, + 201, + 543, + 282 + ], + "blocks": [ + { + "bbox": [ + 312, + 201, + 543, + 282 + ], + "lines": [ + { + "bbox": [ + 312, + 201, + 543, + 282 + ], + "spans": [ + { + "bbox": [ + 312, + 201, + 543, + 282 + ], + "type": "image", + "image_path": "15725040120936c38a6ecf6d17f65b6bca40c7c44b7c4d7674a5e04693710be4.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 310, + 387, + 430, + 506 + ], + "blocks": [ + { + "bbox": [ + 304, + 295, + 546, + 371 + ], + "lines": [ + { + "bbox": [ + 304, + 295, + 546, + 371 + ], + "spans": [ + { + "bbox": [ + 304, + 295, + 546, + 371 + ], + "type": "text", + "content": "Figure 10. (a)-(h) Comparison of state-of-the-art video reconstruction methods on the proposed rotating disk datasets with rotating speed " + }, + { + "bbox": [ + 304, + 295, + 546, + 371 + ], + "type": "inline_equation", + "content": "292\\mathrm{rpm}" + }, + { + "bbox": [ + 304, + 295, + 546, + 371 + ], + "type": "text", + "content": " and environment illuminance " + }, + { + "bbox": [ + 304, + 295, + 546, + 371 + ], + "type": "inline_equation", + "content": "1000\\mathrm{lx}" + }, + { + "bbox": [ + 304, + 295, + 546, + 371 + ], + "type": "text", + "content": ". The yellow box shows magnified edges. The visual \"misalignment\" is due to EVS latency. (i) The DN-edge width relation at a fixed segment of 40 pixels (red line in (a)) on reconstructed rotating disk edge. The refinement network is not used." + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 310, + 387, + 430, + 506 + ], + "lines": [ + { + "bbox": [ + 310, + 387, + 430, + 506 + ], + "spans": [ + { + "bbox": [ + 310, + 387, + 430, + 506 + ], + "type": "image", + "image_path": "122c30b05db8c9c5cb059e32e470ec0d4ddfa289390133eac089943b667e7cde.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 357, + 506, + 409, + 515 + ], + "lines": [ + { + "bbox": [ + 357, + 506, + 409, + 515 + ], + "spans": [ + { + "bbox": [ + 357, + 506, + 409, + 515 + ], + "type": "text", + "content": "(a) speed (rpm)" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 304, + 529, + 545, + 596 + ], + "lines": [ + { + "bbox": [ + 304, + 529, + 545, + 596 + ], + "spans": [ + { + "bbox": [ + 304, + 529, + 545, + 596 + ], + "type": "text", + "content": "Figure 11. Quantitative comparison of state-of-the-art video reconstruction methods on the proposed \"rotating disk\" scene in terms of BEW. The lower the BEW values, the sharper the edge. The refinement network is not used. (a) BEW under various rotation speeds with a fixed illuminance of " + }, + { + "bbox": [ + 304, + 529, + 545, + 596 + ], + "type": "inline_equation", + "content": "1000\\mathrm{lx}" + }, + { + "bbox": [ + 304, + 529, + 545, + 596 + ], + "type": "text", + "content": ". (b) BEW under various illuminance levels with a fixed speed of " + }, + { + "bbox": [ + 304, + 529, + 545, + 596 + ], + "type": "inline_equation", + "content": "292\\mathrm{rpm}" + }, + { + "bbox": [ + 304, + 529, + 545, + 596 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_caption" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 433, + 389, + 541, + 506 + ], + "blocks": [ + { + "bbox": [ + 433, + 389, + 541, + 506 + ], + "lines": [ + { + "bbox": [ + 433, + 389, + 541, + 506 + ], + "spans": [ + { + "bbox": [ + 433, + 389, + 541, + 506 + ], + "type": "image", + "image_path": "b906a65537ee2753c868c9c208278975d5f4c076cc05dbed7596616761135e8c.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 459, + 506, + 527, + 514 + ], + "lines": [ + { + "bbox": [ + 459, + 506, + 527, + 514 + ], + "spans": [ + { + "bbox": [ + 459, + 506, + 527, + 514 + ], + "type": "text", + "content": "(b) illumination (lux)" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_caption" + } + ], + "index": 25 + }, + { + "bbox": [ + 304, + 616, + 545, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 616, + 545, + 652 + ], + "spans": [ + { + "bbox": [ + 304, + 616, + 545, + 652 + ], + "type": "text", + "content": "the proposed method outperforms state-of-the-art methods in reconstructed image quality with an up to " + }, + { + "bbox": [ + 304, + 616, + 545, + 652 + ], + "type": "inline_equation", + "content": "4\\mathrm{dB}" + }, + { + "bbox": [ + 304, + 616, + 545, + 652 + ], + "type": "text", + "content": " improvement in PSNR and " + }, + { + "bbox": [ + 304, + 616, + 545, + 652 + ], + "type": "inline_equation", + "content": "12\\%" + }, + { + "bbox": [ + 304, + 616, + 545, + 652 + ], + "type": "text", + "content": " improvement in LPIPS score." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 304, + 653, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 653, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 653, + 545, + 712 + ], + "type": "text", + "content": "Limitation Since the pixel latency highly depends on EVS design parameters, object motion, and ambient illuminance, the image quality improvement after latency compensation may be subtle for slow objects or in bright scenes such as \"running man\" and \"spinning umbrella\"." + } + ] + } + ], + "index": 29 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 731, + 317, + 740 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 731, + 317, + 740 + ], + "spans": [ + { + "bbox": [ + 293, + 731, + 317, + 740 + ], + "type": "text", + "content": "25179" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 714 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "type": "text", + "content": "[1] Wenbo Bao, Wei-Sheng Lai, Chao Ma, Xiaoyun Zhang, Zhiyong Gao, and Ming-Hsuan Yang. Depth-aware video frame interpolation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3703-3712, 2019. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 146, + 288, + 200 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 146, + 288, + 200 + ], + "spans": [ + { + "bbox": [ + 53, + 146, + 288, + 200 + ], + "type": "text", + "content": "[2] Patrick Bardow, Andrew J Davison, and Stefan Leutenegger. Simultaneous optical flow and intensity estimation from an event camera. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 884-892, 2016. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 201, + 288, + 245 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 201, + 288, + 245 + ], + "spans": [ + { + "bbox": [ + 53, + 201, + 288, + 245 + ], + "type": "text", + "content": "[3] Christian Brandli, Raphael Berner, Minhao Yang, Shih-Chii Liu, and Tobi Delbruck. A " + }, + { + "bbox": [ + 53, + 201, + 288, + 245 + ], + "type": "inline_equation", + "content": "240 \\times 180" + }, + { + "bbox": [ + 53, + 201, + 288, + 245 + ], + "type": "text", + "content": " 130 db " + }, + { + "bbox": [ + 53, + 201, + 288, + 245 + ], + "type": "inline_equation", + "content": "3\\mu s" + }, + { + "bbox": [ + 53, + 201, + 288, + 245 + ], + "type": "text", + "content": " latency global shutter spatiotemporal vision sensor. IEEE Journal of Solid-State Circuits, 49(10):2333-2341, 2014. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 246, + 288, + 290 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 246, + 288, + 290 + ], + "spans": [ + { + "bbox": [ + 53, + 246, + 288, + 290 + ], + "type": "text", + "content": "[4] Xiaojie Chu, Liangyu Chen, and Wenqing Yu. Nafssr: Stereo image super-resolution using nafnet. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, pages 1239-1248, 2022. 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 291, + 288, + 323 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 291, + 288, + 323 + ], + "spans": [ + { + "bbox": [ + 53, + 291, + 288, + 323 + ], + "type": "text", + "content": "[5] Hai Dinh, Qinyi Wang, Fangwen Tu, Brett Frymire, and Bo Mu. Evaluation of motion blur image quality in video frame interpolation. *Electronic Imaging*, 35:262-1, 2023. 5, 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 324, + 288, + 379 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 324, + 288, + 379 + ], + "spans": [ + { + "bbox": [ + 53, + 324, + 288, + 379 + ], + "type": "text", + "content": "[6] Julius Erbach, Stepan Tulyakov, Patricia Vitoria, Alfredo Bochicchio, and Yuanyou Li. Evshutter: Transforming events for unconstrained rolling shutter correction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13904-13913, 2023. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 380, + 288, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 380, + 288, + 402 + ], + "spans": [ + { + "bbox": [ + 53, + 380, + 288, + 402 + ], + "type": "text", + "content": "[7] C.Posch et al. Sensitivity and uniformity of a " + }, + { + "bbox": [ + 53, + 380, + 288, + 402 + ], + "type": "inline_equation", + "content": "0.18\\mu \\mathrm{m}" + }, + { + "bbox": [ + 53, + 380, + 288, + 402 + ], + "type": "text", + "content": " CMOS temporal contrast pixel array. In ISCAS, 2011. 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 403, + 288, + 424 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 403, + 288, + 424 + ], + "spans": [ + { + "bbox": [ + 53, + 403, + 288, + 424 + ], + "type": "text", + "content": "[8] Xinyu Zhou et al. Evunroll: Neuromorphic events based rolling shutter image correction. In CVPR, 2022. 2, 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 425, + 288, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 425, + 288, + 479 + ], + "spans": [ + { + "bbox": [ + 53, + 425, + 288, + 479 + ], + "type": "text", + "content": "[9] Guillermo Gallego, Tobi Delbrück, Garrick Orchard, Chiara Bartolozzi, Brian Taba, Andrea Censi, Stefan Leutenegger, Andrew J Davison, Jörg Conradt, Kostas Daniilidis, et al. Event-based vision: A survey. IEEE transactions on pattern analysis and machine intelligence, 44(1):154-180, 2020. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 480, + 288, + 524 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 480, + 288, + 524 + ], + "spans": [ + { + "bbox": [ + 48, + 480, + 288, + 524 + ], + "type": "text", + "content": "[10] Yue Gao, Siqi Li, Yipeng Li, Yandong Guo, and Qionghai Dai. Superfast: " + }, + { + "bbox": [ + 48, + 480, + 288, + 524 + ], + "type": "inline_equation", + "content": "200 \\times" + }, + { + "bbox": [ + 48, + 480, + 288, + 524 + ], + "type": "text", + "content": " video frame interpolation via event camera. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2022. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 525, + 288, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 525, + 288, + 568 + ], + "spans": [ + { + "bbox": [ + 48, + 525, + 288, + 568 + ], + "type": "text", + "content": "[11] Leon A Gatys, Alexander S Ecker, and Matthias Bethge. Image style transfer using convolutional neural networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2414-2423, 2016. 5" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 569, + 288, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 569, + 288, + 624 + ], + "spans": [ + { + "bbox": [ + 48, + 569, + 288, + 624 + ], + "type": "text", + "content": "[12] Daniel Gehrig, Antonio Loquercio, Konstantinos G Derpanis, and Davide Scaramuzza. End-to-end learning of representations for asynchronous event-based data. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5633-5643, 2019. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 625, + 288, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 625, + 288, + 678 + ], + "spans": [ + { + "bbox": [ + 48, + 625, + 288, + 678 + ], + "type": "text", + "content": "[13] Giorgio Grisetti, Rainer Kümmerle, Hauke Strasdat, and Kurt Konolige. g2o: A general framework for (hyper) graph optimization. In Proceedings of the IEEE International Conference on Robotics and Automation (ICRA), pages 9-13, 2011. 4" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 680, + 288, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 680, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 48, + 680, + 288, + 714 + ], + "type": "text", + "content": "[14] Menghan Guo, Shoushun Chen, Zhe Gao, Wenlei Yang, Peter Bartkovjak, Qing Qin, Xiaqin Hu, Dahai Zhou, Qiping Huang, Masayuki Uchiyama, Yoshiharu Kudo, Shimpei" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 714 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 326, + 73, + 547, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 547, + 139 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 547, + 139 + ], + "type": "text", + "content": "Fukuoka, Chengcheng Xu, Hiroaki Ebihara, Xueqing Wang, Peiwen Jiang, Bo Jiang, Bo Mu, Huan Chen, Jason Yang, T. J. Dai, and Andreas Suess. A three-wafer-stacked hybrid 15-mpixel cis + 1-mpixel evs with 4.6-GeV/s readout, inpixel tdc, and on-chip isp and esp function. IEEE Journal of Solid-State Circuits, pages 1–10, 2023. 1, 2, 5" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 140, + 547, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 140, + 547, + 239 + ], + "spans": [ + { + "bbox": [ + 307, + 140, + 547, + 239 + ], + "type": "text", + "content": "[15] Menghan Guo, Shoushun Chen, Zhe Gao, Wenlei Yang, Peter Bartkovjak, Qing Qin, Xiaqin Hu, Dahei Zhou, Masayuki Uchiyama, Yoshiharu Kudo, Shimpei Fukuoka, Chengcheng Xu, Hiroaki Ebihara, Andy Wang, Peiwen Jiang, Bo Jiang, Bo Mu, Huan Chen, Jason Yang, TJ Dai, and Andreas Suess. A 3-wafer-stacked hybrid 15MPixel CIS + 1MPixel EVS with 4.6GEvent/s readout, in-pixel TDC and on-chip ISP and ESP function. In 2022 IEEE International Solid-State Circuits Conference (ISSCC), 2023. 2, 5" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 240, + 547, + 284 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 240, + 547, + 284 + ], + "spans": [ + { + "bbox": [ + 307, + 240, + 547, + 284 + ], + "type": "text", + "content": "[16] Yuhuang Hu, Shih-Chii Liu, and Tobi Delbruck. v2e: From video frames to realistic dvs events. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1312-1321, 2021. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 285, + 546, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 285, + 546, + 330 + ], + "spans": [ + { + "bbox": [ + 307, + 285, + 546, + 330 + ], + "type": "text", + "content": "[17] Zhewei Huang, Tianyuan Zhang, Wen Heng, Boxin Shi, and Shuchang Zhou. Real-time intermediate flow estimation for video frame interpolation. In European Conference on Computer Vision, pages 624-642. Springer, 2022. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 331, + 546, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 331, + 546, + 396 + ], + "spans": [ + { + "bbox": [ + 307, + 331, + 546, + 396 + ], + "type": "text", + "content": "[18] Huaizu Jiang, Deqing Sun, Varun Jampani, Ming-Hsuan Yang, Erik Learned-Miller, and Jan Kautz. Super slomo: High quality estimation of multiple intermediate frames for video interpolation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 9000-9008, 2018. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 398, + 546, + 421 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 398, + 546, + 421 + ], + "spans": [ + { + "bbox": [ + 307, + 398, + 546, + 421 + ], + "type": "text", + "content": "[19] Nocedal Jorge and J Wright Stephen. Numerical optimization. Springer, 2006. 4" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 422, + 547, + 465 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 422, + 547, + 465 + ], + "spans": [ + { + "bbox": [ + 307, + 422, + 547, + 465 + ], + "type": "text", + "content": "[20] Damien Joubert, Mathieu Hébert, Hubert Konik, and Christophe Lavergne. Characterization setup for event-based imagers applied to modulated light signal detection. Applied optics, 58(6):1305-1317, 2019. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 467, + 546, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 467, + 546, + 511 + ], + "spans": [ + { + "bbox": [ + 307, + 467, + 546, + 511 + ], + "type": "text", + "content": "[21] Damien Joubert, Alexandre Marcireau, Nic Ralph, Andrew Jolley, André van Schaik, and Gregory Cohen. Event camera simulator improvements via characterized parameters. Frontiers in Neuroscience, page 910, 2021. 1, 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 513, + 546, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 513, + 546, + 578 + ], + "spans": [ + { + "bbox": [ + 307, + 513, + 546, + 578 + ], + "type": "text", + "content": "[22] Hanme Kim, Stefan Leutenegger, and Andrew J Davison. Real-time 3d reconstruction and 6-dof tracking with an event camera. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part VI 14, pages 349-364. Springer, 2016. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 579, + 547, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 579, + 547, + 645 + ], + "spans": [ + { + "bbox": [ + 307, + 579, + 547, + 645 + ], + "type": "text", + "content": "[23] Taewoo Kim, Yujeong Chae, Hyun-Kurl Jang, and Kuk-Jin Yoon. Event-based video frame interpolation with cross-modal asymmetric bidirectional motion fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 18032-18042, 2023. 1, 2, 5" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 647, + 547, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 647, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 307, + 647, + 547, + 714 + ], + "type": "text", + "content": "[24] Kazutoshi Kodama, Yusuke Sato, Yuhi Yorikado, Raphael Berner, Kyoji Mizoguchi, Takahiro Miyazaki, Masahiro Tsukamoto, Yoshihisa Matoba, Hirotaka Shinozaki, Atsumi Niwa, et al. 1.22 " + }, + { + "bbox": [ + 307, + 647, + 547, + 714 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 307, + 647, + 547, + 714 + ], + "type": "text", + "content": "m 35.6 mpixel rgb hybrid event-based vision sensor with 4.88 " + }, + { + "bbox": [ + 307, + 647, + 547, + 714 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 307, + 647, + 547, + 714 + ], + "type": "text", + "content": "m-pitch event pixels and up to 10k event frame rate by adaptive control on event sparsity." + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 731, + 318, + 740 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 731, + 318, + 740 + ], + "spans": [ + { + "bbox": [ + 293, + 731, + 318, + 740 + ], + "type": "text", + "content": "25180" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 67, + 72, + 287, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 72, + 287, + 95 + ], + "spans": [ + { + "bbox": [ + 67, + 72, + 287, + 95 + ], + "type": "text", + "content": "In 2023 IEEE International Solid-State Circuits Conference (ISSCC), pages 92-94. IEEE, 2023. 1" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 96, + 287, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 96, + 287, + 138 + ], + "spans": [ + { + "bbox": [ + 48, + 96, + 287, + 138 + ], + "type": "text", + "content": "[25] Patrick Lichtsteiner, Christoph Posch, and Tobi Delbruck. A " + }, + { + "bbox": [ + 48, + 96, + 287, + 138 + ], + "type": "inline_equation", + "content": "128 \\times 128" + }, + { + "bbox": [ + 48, + 96, + 287, + 138 + ], + "type": "text", + "content": " 120 db " + }, + { + "bbox": [ + 48, + 96, + 287, + 138 + ], + "type": "inline_equation", + "content": "15\\mu s" + }, + { + "bbox": [ + 48, + 96, + 287, + 138 + ], + "type": "text", + "content": " latency asynchronous temporal contrast vision sensor. IEEE Journal of Solid-State Circuits, 43 (2):566-576, 2008. 1, 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 140, + 287, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 140, + 287, + 183 + ], + "spans": [ + { + "bbox": [ + 48, + 140, + 287, + 183 + ], + "type": "text", + "content": "[26] Xiaozheng Mou, Kaijun Feng, Alex Yi, Steve Wang, Huan Chen, Xiaoqin Hu, Menghan Guo, Shoushun Chen, and Andreas Suess. Accurate event simulation using high-speed videos. *Electronic Imaging*, 34(7):242-1, 2022. 2, 3, 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 185, + 287, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 185, + 287, + 239 + ], + "spans": [ + { + "bbox": [ + 48, + 185, + 287, + 239 + ], + "type": "text", + "content": "[27] Liyuan Pan, Cedric Scheerlinck, Xin Yu, Richard Hartley, Miaomiao Liu, and Yuchao Dai. Bringing a blurry frame alive at high frame-rate with an event camera. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6820-6829, 2019. 1, 2, 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 239, + 287, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 239, + 287, + 293 + ], + "spans": [ + { + "bbox": [ + 48, + 239, + 287, + 293 + ], + "type": "text", + "content": "[28] Liyuan Pan, Richard Hartley, Cedric Scheerlinck, Miaomiao Liu, Xin Yu, and Yuchao Dai. High frame rate video reconstruction based on an event camera. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(5):2519-2533, 2020. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 295, + 287, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 295, + 287, + 348 + ], + "spans": [ + { + "bbox": [ + 48, + 295, + 287, + 348 + ], + "type": "text", + "content": "[29] Federico Paredes-Valles and Guido CHE de Croon. Back to event basics: Self-supervised learning of image reconstruction for event cameras via photometric constancy. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3446–3455, 2021. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 350, + 287, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 350, + 287, + 392 + ], + "spans": [ + { + "bbox": [ + 48, + 350, + 287, + 392 + ], + "type": "text", + "content": "[30] Christoph Posch and Daniel Matolin. Sensitivity and uniformity of a " + }, + { + "bbox": [ + 48, + 350, + 287, + 392 + ], + "type": "inline_equation", + "content": "0.18\\mu \\mathrm{m}" + }, + { + "bbox": [ + 48, + 350, + 287, + 392 + ], + "type": "text", + "content": " cmos temporal contrast pixel array. In 2011 IEEE International Symposium of Circuits and Systems (ISCAS), pages 1572-1575. IEEE, 2011. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 394, + 287, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 394, + 287, + 437 + ], + "spans": [ + { + "bbox": [ + 48, + 394, + 287, + 437 + ], + "type": "text", + "content": "[31] Henri Rebecq, René Ranftl, Vladlen Koltun, and Davide Scaramuzza. High speed and high dynamic range video with an event camera. IEEE transactions on pattern analysis and machine intelligence, 43(6):1964-1980, 2019. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 438, + 287, + 482 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 438, + 287, + 482 + ], + "spans": [ + { + "bbox": [ + 48, + 438, + 287, + 482 + ], + "type": "text", + "content": "[32] Fitsum Reda, Janne Kontkanen, Eric Tabellion, Deqing Sun, Caroline Pantofaru, and Brian Curless. Film: Frame interpolation for large motion. In European Conference on Computer Vision, pages 250-266. Springer, 2022. 1, 2, 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 483, + 287, + 548 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 483, + 287, + 548 + ], + "spans": [ + { + "bbox": [ + 48, + 483, + 287, + 548 + ], + "type": "text", + "content": "[33] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical Image Computing and Computer-Assisted Intervention-MICCAI 2015: 18th International Conference, Munich, Germany, October 5-9, 2015, Proceedings, Part III 18, pages 234-241. Springer, 2015. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 548, + 287, + 592 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 548, + 287, + 592 + ], + "spans": [ + { + "bbox": [ + 48, + 548, + 287, + 592 + ], + "type": "text", + "content": "[34] Cedric Scheerlinck, Nick Barnes, and Robert Mahony. Continuous-time intensity estimation using event cameras. In Asian Conference on Computer Vision, pages 308-324. Springer, 2018. 1, 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 593, + 287, + 647 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 593, + 287, + 647 + ], + "spans": [ + { + "bbox": [ + 48, + 593, + 287, + 647 + ], + "type": "text", + "content": "[35] Andreas Suess, Menghan Guo, Rui Jiang, Xiaozheng Mou, Qiping Huang, Wenlei Yang, and Shoushun Chen. Physical modeling and parameter extraction for event-based vision sensors. In Proc. IISS Int. Image Sensor Workshop (IISW), Edinburgh, UK, page R5, 2023. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 647, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 647, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 647, + 287, + 712 + ], + "type": "text", + "content": "[36] Lei Sun, Christos Sakaridis, Jingyun Liang, Peng Sun, Jiezhang Cao, Kai Zhang, Qi Jiang, Kaiwei Wang, and Luc Van Gool. Event-based frame interpolation with ad-hoc deblurring. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18043-18052, 2023. 1, 2, 5" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 506 + ], + "type": "list", + "angle": 0, + "index": 22, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 138 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 138 + ], + "type": "text", + "content": "[37] Stepan Tulyakov, Daniel Gehrig, Stamatios Georgoulis, Julius Erbach, Mathias Gehrig, Yuanyou Li, and Davide Scaramuzza. Time lens: Event-based video frame interpolation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16155-16164, 2021. 1, 2, 5" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 140, + 545, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 140, + 545, + 205 + ], + "spans": [ + { + "bbox": [ + 307, + 140, + 545, + 205 + ], + "type": "text", + "content": "[38] Stepan Tulyakov, Alfredo Bochicchio, Daniel Gehrig, Stamatios Georgoulis, Yuanyou Li, and Davide Scaramuzza. Time lens: Event-based frame interpolation with parametric non-linear flow and multi-scale fusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17755-17764, 2022. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 207, + 545, + 260 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 207, + 545, + 260 + ], + "spans": [ + { + "bbox": [ + 307, + 207, + 545, + 260 + ], + "type": "text", + "content": "[39] Ziwei Wang, Yonhon Ng, Cedric Scheerlinck, and Robert Mahony. An asynchronous kalman filter for hybrid event cameras. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 448-457, 2021. 1, 2, 5" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 262, + 545, + 306 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 262, + 545, + 306 + ], + "spans": [ + { + "bbox": [ + 307, + 262, + 545, + 306 + ], + "type": "text", + "content": "[40] Sanghyun Woo, Jongchan Park, Joon-Young Lee, and In So Kweon. Cbam: Convolutional block attention module. In Proceedings of the European conference on computer vision (ECCV), pages 3–19, 2018. 5" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 308, + 545, + 350 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 308, + 545, + 350 + ], + "spans": [ + { + "bbox": [ + 307, + 308, + 545, + 350 + ], + "type": "text", + "content": "[41] Zeyu Xiao, Wenming Weng, Yueyi Zhang, and Zhiwei Xiong. Eva2: Event-assisted video frame interpolation via cross-modal alignment and aggregation. IEEE Transactions on Computational Imaging, 8:1145–1158, 2022. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 353, + 545, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 353, + 545, + 407 + ], + "spans": [ + { + "bbox": [ + 307, + 353, + 545, + 407 + ], + "type": "text", + "content": "[42] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 5" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 408, + 545, + 452 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 408, + 545, + 452 + ], + "spans": [ + { + "bbox": [ + 307, + 408, + 545, + 452 + ], + "type": "text", + "content": "[43] Xiang Zhang and Lei Yu. Unifying motion deblurring and frame interpolation with events. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17765-17774, 2022. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 453, + 545, + 506 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 453, + 545, + 506 + ], + "spans": [ + { + "bbox": [ + 307, + 453, + 545, + 506 + ], + "type": "text", + "content": "[44] Xinyu Zhou, Peiqi Duan, Yi Ma, and Boxin Shi. Evunroll: Neuromorphic events based rolling shutter image correction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17775-17784, 2022. 1, 2" + } + ] + } + ], + "index": 21 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 731, + 317, + 740 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 731, + 317, + 740 + ], + "spans": [ + { + "bbox": [ + 293, + 731, + 317, + 740 + ], + "type": "text", + "content": "25181" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Each Test Image Deserves A Specific Prompt_ Continual Test-Time Adaptation for 2D Medical Image Segmentation/b547c627-f0bf-4b29-9f5d-10851e261fb1_content_list.json b/2024/Each Test Image Deserves A Specific Prompt_ Continual Test-Time Adaptation for 2D Medical Image Segmentation/b547c627-f0bf-4b29-9f5d-10851e261fb1_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..95dd7286f8bb16be7cc2953099213f2c34a50b8f --- /dev/null +++ b/2024/Each Test Image Deserves A Specific Prompt_ Continual Test-Time Adaptation for 2D Medical Image Segmentation/b547c627-f0bf-4b29-9f5d-10851e261fb1_content_list.json @@ -0,0 +1,1544 @@ +[ + { + "type": "text", + "text": "Each Test Image Deserves A Specific Prompt: Continual Test-Time Adaptation for 2D Medical Image Segmentation", + "text_level": 1, + "bbox": [ + 88, + 130, + 880, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Ziyang Chen $^{1}$ Yongsheng Pan $^{1\\dagger}$ Yiwen Ye $^{1}$ Mengkang Lu $^{1}$ Yong Xia $^{1,2,3\\dagger}$", + "bbox": [ + 171, + 202, + 792, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ School of Computer Science and Engineering, Northwestern Polytechnical University, China \n $^{2}$ Research & Development Institute of Northwestern Polytechnical University in Shenzhen, China \n $^{3}$ Ningbo Institute of Northwestern Polytechnical University, China", + "bbox": [ + 96, + 220, + 872, + 275 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "zychen@mail.nwpu.edu.cn, yspan@nwpu.edu.cn, ywye@mail.nwpu.edu.cn", + "bbox": [ + 196, + 277, + 772, + 292 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "lmk@mail.nwpu.edu.cn, yxia@nwpu.edu.cn", + "bbox": [ + 313, + 295, + 651, + 309 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/b7f9bd6d88491770ebe0d90b0c346003251b1c61ebacdc964fdb0880fc4578ee.jpg", + "image_caption": [ + "Source Model $f_{\\theta}$" + ], + "image_footnote": [], + "bbox": [ + 91, + 359, + 169, + 410 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/b8b46250f831afc9abdf05c7c6e2827ddb1168403b421ac04fd66a94dc21150c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 176, + 345, + 277, + 414 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/cd04bb644b92668adb948c54e6dd1e02cd01489ef8ab0b93524ac05af68b8489.jpg", + "image_caption": [ + "Domain 2", + "Domain 3" + ], + "image_footnote": [], + "bbox": [ + 285, + 358, + 359, + 414 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/fc535331f710fa7ac4fce2eee57cf00ca969e955b8339f281b41741b0244954e.jpg", + "image_caption": [ + "Domain 4", + "$f_{\\theta}^{\\prime}\\rightarrow f_{\\theta}$" + ], + "image_footnote": [], + "bbox": [ + 359, + 358, + 473, + 414 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/7319de94f101abffc1db56458e03dba8862b88e6626de9b750b8ffadb1077ef0.jpg", + "image_caption": [ + "" + ], + "image_footnote": [], + "bbox": [ + 473, + 358, + 535, + 414 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/e494199ddcc9269b92378aae9e5e5dd9177811b8a547070f8b40a5a81cf05d25.jpg", + "image_caption": [ + "Domain N" + ], + "image_footnote": [], + "bbox": [ + 535, + 358, + 571, + 412 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/7e3ad50669bb3bfaf2ad32d3263adc01c67d8f7dc067ec88311807d661551c85.jpg", + "image_caption": [ + "", + "$\\mathcal{L}_{self} \\to \\mathcal{L}_{prompt}$" + ], + "image_footnote": [], + "bbox": [ + 571, + 358, + 656, + 414 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/a05f686c0ccc852d2a5dab91a906236fcdcad7d871b369ef903edea534b0d07e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 656, + 358, + 722, + 414 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/fed712e153f9e5774cd0cc8f4a24a59cab2db2c8c82ea713d3e05bc69d46a2de.jpg", + "image_caption": [ + "Test" + ], + "image_footnote": [], + "bbox": [ + 532, + 438, + 563, + 489 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/f72b2d431c2fb13b0ab4e39ad1af38019d53f964edf9309d24df6ec683889042.jpg", + "image_caption": [ + "Adapted Output" + ], + "image_footnote": [], + "bbox": [ + 563, + 438, + 676, + 489 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/88efd7ca147b209c872a8aa52987eaa8196794d516936ba32fc8fc5f217a3219.jpg", + "image_caption": [ + "(c) Our VPTTA", + "EA:", + "CF:" + ], + "image_footnote": [], + "bbox": [ + 614, + 489, + 640, + 503 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/eac8d5775df914bcf760ff26ff202347d72471ca16344356ea573d4ddd730952.jpg", + "image_caption": [ + "Freeze" + ], + "image_footnote": [], + "bbox": [ + 754, + 359, + 782, + 378 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/d0488d37486fb777069da662dc3fdd126d023911f2b61598462fba29c36f68af.jpg", + "image_caption": [ + "Fine-tune" + ], + "image_footnote": [], + "bbox": [ + 756, + 385, + 781, + 407 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Freeze", + "bbox": [ + 805, + 364, + 848, + 376 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/c2450a86b24cea5720114978d05f0a713b50915b93014079c783837ff7f8bf24.jpg", + "image_caption": [ + "Prompt" + ], + "image_footnote": [], + "bbox": [ + 756, + 414, + 779, + 436 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/40332400e9aa9af88781e712236ab1c13d8f18a18d90c69267432e81c2cd9168.jpg", + "image_caption": [ + "Terrible" + ], + "image_footnote": [], + "bbox": [ + 754, + 446, + 782, + 463 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/5da86f88f4bdf7572c8ca547f9c4b4f6f98caf315180fda582195c5b5a87c355.jpg", + "image_caption": [ + "Passable" + ], + "image_footnote": [], + "bbox": [ + 754, + 474, + 782, + 494 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/7d936a17d3505ddcbca639d400faa5ff0772bef947668a13d6d47f9b19f07ace.jpg", + "image_caption": [ + "Good", + "Figure 1. Comparison between our VPTTA and existing solutions under the CTTA setup. Our VPTTA avoids both error accumulation (EA) and catastrophic forgetting (CF) by freezing model parameters and achieves adaptation by training a prompt for each test image. The commonly used self-supervised loss and the optimized one are denoted by $\\mathcal{L}_{self}$ and $\\mathcal{L}_{self}^{\\prime}$ , respectively." + ], + "image_footnote": [], + "bbox": [ + 754, + 503, + 782, + 523 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 604, + 313, + 619 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Distribution shift widely exists in medical images acquired from different medical centres and poses a significant obstacle to deploying the pre-trained semantic segmentation model in real-world applications. Test-time adaptation has proven its effectiveness in tackling the cross-domain distribution shift during inference. However, most existing methods achieve adaptation by updating the pre-trained models, rendering them susceptible to error accumulation and catastrophic forgetting when encountering a series of distribution shifts (i.e., under the continual test-time adaptation setup). To overcome these challenges caused by updating the models, in this paper, we freeze the pre-trained model and propose the Visual Prompt-based Test-Time Adaptation (VPTTA) method to train a specific prompt for each test image to align the statistics in the batch normalization layers. Specifically, we present the low-frequency prompt, which is lightweight with only a few parameters", + "bbox": [ + 75, + 643, + 472, + 902 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "and can be effectively trained in a single iteration. To enhance prompt initialization, we equip VPTTA with a memory bank to benefit the current prompt from previous ones. Additionally, we design a warm-up mechanism, which mixes source and target statistics to construct warm-up statistics, thereby facilitating the training process. Extensive experiments demonstrate the superiority of our VPTTA over other state-of-the-art methods on two medical image segmentation benchmark tasks. The code and weights of pre-trained source models are available at https://github.com/Chen-Ziyang/VPTTA.", + "bbox": [ + 496, + 604, + 893, + 772 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 808, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "$\\dagger$ Yong Xia and Yongsheng Pan are the corresponding authors. This work was supported in part by the National Natural Science Foundation of China under Grants 62171377, in part by STI2030-Major Project under Grants 2022ZD0213100, in part by Shenzhen Science and Technology Program under Grants JCYJ20220530161616036, and in part by the Ningbo Clinical Research Center for Medical Imaging under Grant 2021L003 (Open Project 2022LYKFZD06).", + "bbox": [ + 496, + 814, + 893, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "11184", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 78, + 89, + 207, + 104 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Semantic segmentation serves a pivotal role in medical image processing to outline specific anatomical structures. However, the segmentation model pre-trained on the source dataset may be hindered from deploying in real-world applications due to the commonly existing distribution shift, which is typically caused by the variations in imaging protocols, operators, and scanners [16, 34]. Test-time adaptation (TTA) has emerged as a promising paradigm to replace domain adaptation (DA) [5, 43, 46, 53], as TTA requires only test data during the inference phase, while DA necessitates a sufficiently large and representative target dataset [45]. The mainstream TTA methods [29-31, 37, 40, 41, 47, 48] leverage the self-supervised tasks to construct auxiliary losses, such as entropy minimization [41] and self-training [37], to update the parameters of pre-trained models. Although these TTA methods successfully alleviate the domain gap and performance degradation, they are based on the assumption that the target domains exhibit static distributions. Unfortunately, the target domains are continuously changing rather than static in most real-world scenarios.", + "bbox": [ + 76, + 114, + 472, + 417 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "This was first presented as the continual test-time adaptation (CTTA) setup [42] that addresses TTA on a series of distribution shifts. Under this setup, pre-trained models are required to adapt to several distinct domains and therefore are more prone to suffer error accumulation and catastrophic forgetting. As mentioned above, many TTA methods rely on self-supervised tasks and may be affected by the noisy losses due to unreliable supervision, leading to error accumulation [2, 23]. Catastrophic forgetting typically occurs when the model is continually trained on new domains, resulting in unexpected performance degradation [21, 30]. Some research has been studied in addressing these two challenges based on optimizing the loss functions for better supervision [6, 31, 42, 42, 45] (see Figure 1 (a)) or prevent the model parameters from excessive alterations by model resetting or introducing regularizers [30, 31, 39, 42] (see Figure 1 (b)). Despite their efforts, loss optimization cannot ensure a robust training process to avoid catastrophic forgetting during the long-term inference, insufficient or excessive updating of the model parameters may lead to inadequate adaptation or error accumulation. Since the pretrained model is updated by the self-supervised loss where the noise cannot be completely removed, these methods may still suffer from these two issues.", + "bbox": [ + 76, + 417, + 472, + 779 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Considering the above limitations in CTTA, updating the pre-trained model during inference seems to be inappropriate. To this end, we propose the Visual Prompt-based Test-Time Adaptation (VPTTA) method to avoid error accumulation and catastrophic forgetting associated with updating model by freezing the model parameters and conduct adaptation by learning a specific visual prompt for each test image (see Figure 1 (c)). The prompt can adapt each test im", + "bbox": [ + 76, + 780, + 470, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "age to the frozen pre-trained model to take into account both adaptation ability and knowledge retaining. We attempt to achieve this by answering the following three questions.", + "bbox": [ + 496, + 90, + 893, + 137 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "- How to design the prompt? Each prompt is restricted to a limited number of training iterations (sometimes just one) to satisfy the demands of online inference, requiring the prompt to be lightweight. Based on this principle, we introduce the prompt for only the low-frequency components of each sample, resulting in the low-frequency prompt. Since the low-frequency components are strongly associated with style textures and can serve as the primary source of distribution shift [46], our low-frequency prompt allows for maintaining lightweight and being able to reduce distribution shifts by modifying the low-frequency components.", + "bbox": [ + 500, + 138, + 893, + 320 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "- How to initialize the prompt? It is well known that model training can benefit from proper parameter initialization, resulting in not only fewer training iterations but also improved performance [3, 52]. We argue that this is also applicable to prompt training and therefore construct a memory bank to store the prompts and low-frequency components of previous test images. By comparing the similarity between the low-frequency components of the current and previous images, we select several historical prompts of the most similar images to initialize the current prompt.", + "bbox": [ + 500, + 320, + 893, + 486 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "- How to train the prompt? Since the statistics mismatch between the source and target domains in the batch normalization (BN) layers is a significant cause of distribution shift [33, 35, 44], we optimize the prompt with the objective of minimizing the distances between BN statistics and the statistics extracted from the features of test data for statistics alignment [19]. To further address the training difficulty posed by the large distribution shift at the beginning of the inference phase, we devise a statistics-fusion-based warm-up mechanism to assist in learning the prompt from easy to hard.", + "bbox": [ + 500, + 486, + 893, + 654 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To summarise, our contributions are three-fold: (1) We present the low-frequency prompt to adjust the test image to adapt the model instead of updating the model and therefore prevent the model from error accumulation and catastrophic forgetting. (2) We propose the novel VPTTA framework to alleviate the distribution shift by training a sample-specific visual prompt, which is lightweight, well-initialized, and follows a gradual learning process. (3) The powerful framework we provide can be effectively deployed, requiring only a single iteration and a few dozen parameters for adaptation.", + "bbox": [ + 496, + 655, + 893, + 808 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 821, + 640, + 837 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Test-time Adaptation (TTA)", + "text_level": 1, + "bbox": [ + 500, + 845, + 751, + 863 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "TTA aims to adapt the model pre-trained on the source domain to the test data during inference in a source-free", + "bbox": [ + 500, + 869, + 893, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "11185", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "and online manner [40, 41]. The mainstream TTA methods are based on model updating, which constructs self-supervised auxiliary tasks to guide the model to learn on the test data [4, 13, 26, 29, 37, 40, 41, 48, 49]. Sun et al. [40] utilized an auxiliary task to predict the rotation to assist the model learning on the test data. Wang et al. [41] presented a test-time entropy minimization scheme to reduce generalization error by reducing the entropy of model predictions on test data. Zhang et al. [49] introduced a supplementary network to fit the test data, which can work in a plug-and-play fashion, necessitating minimal hyperparameter tuning. Some other methods are based on BN statistics, which alleviate the distribution shifts by modifying the statistics within BN layers [25, 27, 44]. Mirza et al. [25] constantly adapted the statistics of the BN layers to modify the feature representations of the model. These TTA methods improve the adaptation performance on the static target domain but neglect the continually changing target domains in most real-world scenarios.", + "bbox": [ + 75, + 90, + 472, + 377 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Recently, the continual test-time adaptation (CTTA) [42] is proposed to tackle TTA under a series of continually changing target domains. Due to long-term inference, both error accumulation [2, 23] and catastrophic forgetting [21, 30] are more prone to occur in this setup. Many methods [6, 30, 31, 39, 42, 45, 47] optimize the self-supervised losses or prevent excessive alterations to the model parameters via model resetting or regularizer to improve this situation but are still impacted by these issues. Wang et al. [42] refined pseudo labels by the weight- and augmentation-averaged predictions and randomly restored a small part of model parameters. Niu et al. [31] removed partial noisy samples with large gradients and encouraged model weights toward a flat minimum. Yang et al. [45] designed a dynamic strategy to adjust the learning rate for each test image. Zhang et al. [47] presented to adaptively fuse source and test statistics and improve the entropy minimization loss to a generalized one. In contrast, our proposed VPTTA employs a learnable prompt for each test image to mitigate distribution shifts during inference, requiring no changes to the pre-trained model and therefore alleviating the error accumulation and catastrophic forgetting.", + "bbox": [ + 75, + 378, + 472, + 712 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. Prompt Learning", + "text_level": 1, + "bbox": [ + 76, + 724, + 246, + 742 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Prompt learning was initially employed to devise additional text instructions for input text to fine-tune the largescale Natural Language Processing (NLP) models on downstream tasks [24]. With remarkable success in NLP, it has also been involved in computer vision and greatly benefited visual models in recent years [12, 15, 19, 51]. Jia et al. [15] introduced a small number of trainable parameters in the input space while keeping the model backbone frozen to fine-tune the model with high quality and efficiency. Zhou et al. [51] presented the conditional context optimization", + "bbox": [ + 75, + 750, + 472, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "to generate an input-conditional token for each image to generalize the pre-trained vision-language model to unseen classes. Hu et al. [19] employed the visual prompt at the pixel level and added it to the input images to explicitly minimize the domain discrepancy. Gan et al. [12] learned a domain-specific and domain-agnostic visual prompt for adaptation by using the exponential-moving-average strategy under the guidance of self-training. Different from them, our VPTTA presents the low-frequency prompt to address a series of changing distribution shifts with a few dozen parameters. We further consider optimizing the initialization and warm-up mechanism, enabling our visual prompts to be effectively trained within a single iteration to satisfy the demands of CTTA.", + "bbox": [ + 496, + 90, + 893, + 303 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Methodology", + "text_level": 1, + "bbox": [ + 498, + 314, + 635, + 332 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Problem Definition and Method Overview", + "text_level": 1, + "bbox": [ + 498, + 339, + 857, + 356 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Let the labeled source domain dataset and unlabeled testing target dataset be $\\mathbb{D}^s = \\{X_i^s,Y_i^s\\}_{i = 1}^{N^s}$ and $\\mathbb{D}^t = \\{X_i^t\\}_{i = 1}^{N^t}$ , respectively, where $X_{i}^{*}\\in \\mathbb{R}^{H\\times W\\times C}$ is the $i$ -th image with C channels and size of $H\\times W$ , and $Y_{i}^{*}$ is its label. Our goal is to learn a low-frequency prompt $\\mathcal{P}_i$ for each test image $X_{i}^{t}$ to obtain the modified image $\\tilde{X}_i^t$ that can be adapted to the model $f_{\\theta}:X\\to Y$ pre-trained on $\\mathbb{D}^s$ .", + "bbox": [ + 496, + 362, + 890, + 470 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To accomplish this goal, our VPTTA considers three key facets: prompt design, prompt initialization, and prompt training. For prompt design, the prompt is designed to change the low-frequency component of the amplitude of the test image. For prompt initialization, we adopt a memory bank to initialize the current prompt with previous ones. For prompt training, we employ the absolute distance to train the prompt to align the statistics between features extracted from test images and the batch normalization (BN) layers and also present a statistics-fusion-based warm-up mechanism. An overview of our VPTTA is illustrated in Figure 2. We now delve into its details.", + "bbox": [ + 496, + 472, + 893, + 654 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Prompt Design: Low-frequency Prompt", + "text_level": 1, + "bbox": [ + 498, + 661, + 843, + 678 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Let $\\mathcal{F}(\\cdot)$ and $\\mathcal{F}^{-1}(\\cdot)$ be the Fast Fourier Transform (FFT) and Inverse Fast Fourier Transform (IFFT) operations [10], respectively. The amplitude component and phase component are denoted by $\\mathcal{F}^A (\\cdot)$ and $\\mathcal{F}^P (\\cdot)$ respectively. Note that the low-frequency component in $\\mathcal{F}^A (\\cdot)$ is shifted to the center. Utilizing the prompt $\\mathcal{P}_i\\in \\mathbb{R}^{(\\alpha \\times H)\\times (\\alpha \\times W)\\times C}$ to obtain the adapted image $\\bar{X}_i^t$ can be formalized as", + "bbox": [ + 496, + 684, + 890, + 791 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {X} _ {i} ^ {t} = \\mathcal {F} ^ {- 1} ([ \\text {O n e P a d} (\\mathcal {P} _ {i}) \\odot \\mathcal {F} ^ {A} (X _ {i} ^ {t}), \\mathcal {F} ^ {P} (X _ {i} ^ {t}) ]) \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 514, + 797, + 890, + 816 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\odot$ is the element-wise multiplication, and OnePad means padding one around $\\mathcal{P}_i$ to expand it to the size of $H\\times W$ . Under the control of the small coefficient $\\alpha \\in (0,1)$ , we can ensure that the prompt $\\mathcal{P}_i$ remains lightweight and focuses on the low-frequency component.", + "bbox": [ + 496, + 824, + 893, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "11186", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/62cb4d2a81e70ceeeda78d942341260a70bc8d077afc8e161d3cbac4508c6ad2.jpg", + "image_caption": [ + "Figure 2. Overview of our VPTTA. For each test image, (1) the Fast Fourier Transform (FFT) is first applied to transform it into the frequency domain, where the low-frequency component of amplitude is used to query in a memory bank to initialize the current prompt, and then the amplitude is multiplied with the prompt at the low-frequency component and transformed back to the spatial domain using the Inverse Fast Fourier Transform (IFFT). (2) The memory bank is built on the previous low-frequency components and their corresponding prompts and is updated using the First In First Out (FIFO) strategy. (3) We convert the source statistics stored in BN layers into the warm-up statistics and calculate the absolute distance between warm-up and target statistics as the loss to fine-tune the prompt. (4) Finally, we feed the image corrected by its fine-tuned prompt to the pre-trained model to produce the output. 'Stat': Abbreviation of 'Statistics'." + ], + "image_footnote": [], + "bbox": [ + 122, + 88, + 851, + 406 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Simultaneously, we recognize that the low-rank decomposition is widely employed in the field of data compression [22] and model fine-tuning [18]. Consequently, we also design a low-rank prompt $\\mathcal{P}_i^{LoRa} = \\mathcal{B}_i@A_i$ as a variant (denoted as VPTTA-LR), where $\\mathcal{B}_i\\in \\mathbb{R}^{H\\times r\\times C}(r\\ll W)$ $\\mathcal{A}_i\\in \\mathbb{R}^{r\\times W\\times C}(r\\ll H)$ , and @ denotes the matrix multiplication operation. In this way, the adapted image can be formalized as $\\tilde{X}_i^t = X_i^t +\\mathcal{P}_i^{LoRa}$ .", + "bbox": [ + 76, + 537, + 472, + 659 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Prompt Initialization: Memory Bank", + "text_level": 1, + "bbox": [ + 76, + 676, + 398, + 694 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To achieve efficient training, it is critical to produce a suitable initialization for $\\mathcal{P}_i$ . Therefore, we present a memory-bank-based initialization strategy to benefit the current prompt from previous ones. Specifically, the memory bank $M$ stores $S$ pairs of keys and values $\\{q_s,v_s\\}_{s = 1}^S$ and holds the First In First Out (FIFO) principle. As shown in Figure 2, the keys are the low-frequency components of previous test images, and the values are their corresponding prompts. Note that when using VPTTA-LR, the keys are the low-rank reconstruction matrix of previous test images. To initialize the current prompt $\\mathcal{P}_i$ , we first extract the low-frequency component of the current test image $\\mathcal{F}_{low}^{A}(X_{i}^{t})$ and calculate the cosine similarity between $\\mathcal{F}_{low}^{A}(X_{i}^{t})$ and", + "bbox": [ + 75, + 703, + 472, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "each key $q_{s}$ as follows:", + "bbox": [ + 500, + 537, + 656, + 553 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nC o s \\left(\\mathcal {F} _ {l o w} ^ {A} \\left(X _ {i} ^ {t}\\right), q _ {s}\\right) = \\frac {\\left\\langle \\mathcal {F} _ {l o w} ^ {A} \\left(X _ {i} ^ {t}\\right) , q _ {s} \\right\\rangle}{\\left\\| \\mathcal {F} _ {l o w} ^ {A} \\left(X _ {i} ^ {t}\\right) \\right\\| * \\left\\| q _ {s} \\right\\|}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 547, + 556, + 890, + 593 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\langle \\cdot ,\\cdot \\rangle$ denotes the dot product operation, and $\\| \\cdot \\|$ denotes the Euclidean norm. Since the low-frequency component can represent the style texture of an image, we rank the similarity scores and then retrieve the values of the $K$ most relevant keys from $M$ to construct a support set $R_{i} = \\{q_{k},v_{k}\\}_{k = 1}^{K}$ for the current test image. Finally, we assign a weight $w_{k}$ to $v_{k}$ to initialize the current prompt by $\\mathcal{P}_i = \\sum_{k = 1}^K w_k v_k$ where the weights $\\{w_{k}\\}$ are calculated based on the similarity scores [45] and $\\sum_{k = 1}^{K}w_{k} = 1$", + "bbox": [ + 498, + 597, + 890, + 737 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.4. Prompt Training: Statistics Alignment", + "text_level": 1, + "bbox": [ + 500, + 742, + 833, + 758 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Inspired by [19, 27, 47], the distribution shift largely exists in distinct BN statistics calculated on the test features (i.e., $\\mu_t$ and $\\sigma_t$ ) and stored in the source model $f_{\\theta}$ (i.e., $\\mu_s$ and $\\sigma_s$ ), where $\\mu$ and $\\sigma$ respectively denote the mean and standard deviation. To train the prompts to align the statistics, we adopt the absolute-distance-based loss $\\mathcal{L}_p$ [19] as follows:", + "bbox": [ + 496, + 765, + 890, + 869 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {p} = \\frac {1}{j} \\sum \\left| \\mu_ {s} ^ {j} - \\mu_ {t} ^ {j} \\right| + \\left| \\sigma_ {s} ^ {j} - \\sigma_ {t} ^ {j} \\right|, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 576, + 873, + 890, + 905 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "11187", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "code", + "sub_type": "algorithm", + "code_caption": [ + "Algorithm 1: VPTTA algorithm." + ], + "code_body": "Initialize: Frozen source model $f_{\\theta}$ , initial prompt $\\mathcal{P}$ filled with ones, memory bank $M$ , hyperparameters $S$ , $K$ and $\\tau$ .", + "bbox": [ + 98, + 112, + 415, + 152 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Input: For each time step $i$ , current test image $X_{i}^{t}$ .", + "bbox": [ + 98, + 154, + 400, + 166 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1: ▷ Initialize the prompt $\\mathcal{P}_i$ :", + "2: if $i < K$ then", + "3: $\\mathcal{P}_i = \\mathcal{P}$", + "4: else", + "5: Construct the support set $R_{i}$ from $M$", + "6: $\\mathcal{P}_i = \\sum_{k=1}^K w_k v_k, v_k \\in R_i$", + "7: end if", + "8: $\\triangleright$ Train the prompt $P_{i}$ :", + "9: Obtain $\\hat{X}_i^t$ using $\\mathcal{P}_i$ by Eq. (1)", + "10: Forward $f_{\\theta}(X_i^t)$ and calculate the warm-up statistics $\\mu_w$ and $\\sigma_w$ by Eq. (4) to replace the $\\mu_s$ and $\\sigma_s$ stored in $f_{\\theta}$", + "11: Backward and Update $\\mathcal{P}_i$ by Eq. (5)", + "12: Inference:", + "13: Obtain $\\hat{X}_i^t$ using updated $\\mathcal{P}_i$ by Eq. (1)", + "14: Forward $O_{i} = f_{\\theta}(X_{i}^{t})$", + "15: Update the memory bank $M$ :", + "16: M.enqueue $(q_{i},v_{i})$ $q_{i} = \\mathcal{F}^{A}\\left(X_{i}^{t}\\right),v_{i} = \\mathcal{P}_{i}$", + "17: if $\\text{len}(M) > S$ then", + "18: M.dequeue $(q_{\\text{first}}, v_{\\text{first}})$", + "19: end if" + ], + "bbox": [ + 101, + 167, + 483, + 441 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Output:Adapted prediction $O_{i}$", + "bbox": [ + 96, + 446, + 289, + 460 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $j$ indicates the $j$ -th BN layer in $f_{\\theta}$ .", + "bbox": [ + 76, + 494, + 352, + 510 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "However, aligning the source and target domains within a limited number of iterations (one in this study) provides challenges due to the empty memory bank and undiminished distribution shifts at the beginning of the inference phase. Therefore, we design a warm-up mechanism to address this issue by simulating the warm-up statistics (i.e., $\\mu_w$ and $\\sigma_w$ ) as follows:", + "bbox": [ + 75, + 511, + 468, + 617 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mu_ {w} = \\lambda \\mu_ {t} + (1 - \\lambda) \\mu_ {s}, \\sigma_ {w} = \\lambda \\sigma_ {t} + (1 - \\lambda) \\sigma_ {s}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 99, + 631, + 468, + 648 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\lambda = \\frac{1}{\\sqrt{i} / \\tau + 1}$ , $i$ indicates the index of the current test image and starts from 1, and $\\tau$ is a temperature coefficient to control the rate of transition from the warm-up statistics to the source statistics. Now, the Eq. (3) can be rewritten by:", + "bbox": [ + 75, + 660, + 468, + 738 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {p} = \\frac {1}{j} \\sum \\left| \\mu_ {w} ^ {j} - \\mu_ {t} ^ {j} \\right| + \\left| \\sigma_ {w} ^ {j} - \\sigma_ {t} ^ {j} \\right|, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 150, + 739, + 468, + 771 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "and we utilize $\\mu_w$ and $\\sigma_w$ for normalization instead of $\\mu_s$ and $\\sigma_s$ . As the inference progresses, the warm-up statistics will gradually shift towards the source statistics. By simulating the warm-up statistics between the source and target statistics, we can significantly reduce the training difficulty at the beginning of the inference phase, resulting in learning prompts from easy to difficult. The overall process of VPTTA is summarized in Algorithm 1.", + "bbox": [ + 75, + 780, + 468, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 500, + 89, + 632, + 107 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We evaluate our proposed VPTTA on two 2D segmentation benchmark tasks: the joint optic disc (OD) and cup (OC) segmentation task, and the polyp segmentation task.", + "bbox": [ + 498, + 114, + 890, + 160 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Datasets and Evaluation Metrics", + "text_level": 1, + "bbox": [ + 498, + 167, + 785, + 181 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The OD/OC segmentation dataset comprises five public datasets collected from different medical centres, denoted as domain A (RIM-ONE-r3 [11]), B (REFuge [32]), C (ORIGA [50]), D (REFuge-Validation/Test [32]), and E (Drishti-GS [38]). There are 159, 400, 650, 800, and 101 images from these datasets. We cropped a region of interest (ROI) centering at OD with size of $800 \\times 800$ for each image following [19], and each ROI is further resized to $512 \\times 512$ and normalized by min-max normalization. The Dice score metric $(DSC)$ is utilized for evaluation in this task.", + "bbox": [ + 496, + 190, + 890, + 340 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The polyp segmentation dataset consists of four public datasets collected from different medical centres, denoted as domain A (BKAI-IGH-NeoPolyp [28]), B (CVC-ClinicDB [1]), C (ETIS-LaribPolypDB [36]), and D (Kvasir-Seg [20]), which have 1000, 612, 196, and 1000 images, respectively. We followed [9] to resize each image to $352 \\times 352$ and normalize the resized image by the statistics computed on ImageNet. The $DSC$ , enhanced-alignment metric $(E_{\\phi}^{max})$ [8], and structural similarity metric $(S_{\\alpha})$ [7] are utilized for evaluation in this task.", + "bbox": [ + 496, + 340, + 890, + 491 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Implementation Details", + "text_level": 1, + "bbox": [ + 498, + 500, + 715, + 515 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For each task, we trained the source model on each single domain (source domain) and tested it on the left domains (target domains) to calculate the mean metrics to evaluate the methods on different scenarios. In the source-training phase, we trained a ResUNet-34 [17] backbone, following [19], as the baseline for OD/OC segmentation task and trained the PraNet [9] with a Res2Net-based [14] backbone for polyp segmentation task. Due to the complexity of PraNet's decoder, we only calculated the prompt loss $\\mathcal{L}_p$ for each BN layer in the encoder instead of the whole network. In the test-adaptation phase, We performed one-iteration adaptation for each batch of test data with a batch size of 1 on all experiments of our VPTTA and other competing methods [45]. As our VPTTA involves an iteration to update and an inference step, distinguishing it from other methods like [41] that do not require a dedicated inference step, we applied the same configuration as VPTTA to other competing methods to ensure consistency in the experimental setup. To deploy our VPTTA, we utilized the Adam optimizer with a learning rate of 0.05 and 0.01 for the OD/OC segmentation task and polyp segmentation task, respectively. The hyperparameters $\\alpha$ (size of prompt), $S$ (size of memory bank), $K$ (size of support set), and $\\tau$ (temperature coefficient in warm-up) are set to 0.01, 40, 16, and 5 for both segmentation tasks.", + "bbox": [ + 496, + 523, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "11188", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/8fd954f0893c4820081f1b86d15d86c66d9b4c9de30aba6b62969226756b1ac6.jpg", + "table_caption": [ + "Table 1. Performance of our VPTTA, 'Source Only' baseline, and six competing methods on the OD/OC segmentation task. The best and second-best results in each column are highlighted in bold and underline, respectively." + ], + "table_footnote": [], + "table_body": "
MethodsDomain ADomain BDomain CDomain DDomain EAverage
DSCDSCDSCDSCDSCDSC↑
Source Only (ResUNet-34)64.5376.0671.1852.6764.8765.86
TENT-continual (ICLR 2021) [41]73.0778.6671.9446.8170.2068.13
CoTTA (CVPR 2022) [42]75.3975.9869.1453.9970.4068.98
DLTTA (TMI 2022) [45]75.1178.8573.8951.6469.7169.84
DUA (CVPR 2022) [25]72.2876.5970.1356.1771.3869.31
SAR (ICLR 2023) [31]74.5577.7170.7855.4071.7270.03
DomainAdaptor (CVPR 2023) [47]74.5076.3971.8156.7870.5570.01
VPTTA (Ours)73.9179.3674.5156.5175.3571.93
", + "bbox": [ + 179, + 128, + 785, + 253 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/9634239044c431f8dadbc366191c2b1851142db7fa988c64c041e94d9ea1bb0a.jpg", + "table_caption": [ + "Table 2. Performance of our VPTTA, 'Source Only' baseline, and six competing methods on the polyp segmentation task. The best and second-best results in each column are highlighted in bold and underline, respectively." + ], + "table_footnote": [], + "table_body": "
MethodsDomain ADomain BDomain CDomain DAverage
\\(DSC\\)\\(E_{\\phi}^{max}\\)\\(S_{\\alpha}\\)\\(DSC\\)\\(E_{\\phi}^{max}\\)\\(S_{\\alpha}\\)\\(DSC\\)\\(E_{\\phi}^{max}\\)\\(S_{\\alpha}\\)\\(DSC\\)\\(E_{\\phi}^{max}\\)\\(S_{\\alpha}\\)\\(DSC \\uparrow\\)\\(E_{\\phi}^{max} \\uparrow\\)\\(S_{\\alpha} \\uparrow\\)
Source Only (PraNet)79.9087.9784.6666.3378.5176.7273.8984.6481.2882.9590.8488.0875.7785.4982.69
TENT-continual (ICLR 2021) [41]74.8684.5880.5267.5178.6678.0517.7940.0453.3073.5583.3882.4158.4371.6773.57
CoTTA (CVPR 2022) [42]76.4685.3782.5666.7776.7579.1771.3983.4280.1870.7179.8182.5471.3381.3481.11
DLTTA (TMI 2022) [45]76.2785.2382.4166.5877.0079.2463.7278.2375.5671.2081.3283.4769.4480.4580.17
DUA (CVPR 2022) [25]78.9387.3783.9666.8478.5277.5176.5386.4583.0586.2493.2389.8277.1386.3983.58
SAR (ICLR 2023) [31]76.4885.8981.4966.4577.3578.0571.4683.2379.4070.4180.1181.0771.2081.6580.00
DomainAdaptor (CVPR 2023) [47]77.4886.3182.4070.8281.7680.8871.9683.0679.9776.8985.8984.4574.2984.2681.93
VPTTA (Ours)81.0088.9184.9176.8787.3184.0877.5887.4883.6486.3993.4789.8780.4689.2985.62
", + "bbox": [ + 80, + 305, + 890, + 419 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/939cf4dca1e1de0085c672d9cddd7409743840ba63b8513c426ca955bd096682.jpg", + "table_caption": [ + "Table 3. Results of ablation study on the OD/OC segmentation task. The best and second-best results in each column are highlighted in bold and underline, respectively." + ], + "table_footnote": [], + "table_body": "
MethodsDomain ADomain BDomain CDomain DDomain EAverage
Low-frequency PromptMemory BankWarm-upDSCDSCDSCDSCDSCDSC↑
64.5376.0671.1852.6764.8765.86
✓65.9976.5871.7552.8667.3066.90
✓✓70.7079.3672.8255.5072.8470.24
✓✓74.4177.4774.1454.0367.8569.58
✓✓✓73.9179.3674.5156.5175.3571.93
", + "bbox": [ + 142, + 470, + 828, + 561 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Experimental Results", + "text_level": 1, + "bbox": [ + 76, + 584, + 279, + 601 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We compare our VPTTA with the 'Source Only' baseline (training on the source domain and testing without adaptation), and six competing methods, including a pseudo-label-based method (CoTTA [42]), two entropy-based methods (TENT-continual [41] and SAR [31]), a method dynamically adjusting the learning rate (DLTTA [45]), a method combining entropy-loss and BN statistics fusion (DomainAdaptor [47]), and a BN statistics-based method (DUA [25]) on two segmentation tasks. Specifically, we re-implemented all the competing methods using the same baseline as VPTTA and combined DLTTA with the entropy loss proposed in TENT [45]. The result of each domain is calculated by using the current domain as the source domain and applying the pre-trained model on the left domains.", + "bbox": [ + 75, + 609, + 470, + 821 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Comparison on the OD/OC segmentation task. The results of the proposed VPTTA, the \"Source Only\" baseline, and six competing methods on the OD/OC segmentation task are detailed in Table 1. Obviously, all competing methods outperform the \"Source Only\" baseline, indicating the", + "bbox": [ + 75, + 824, + 470, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "effectiveness of adaptation towards the distribution shifts. It can be found that our VPTTA achieves the best overall performance across all domains, underscoring its superior applicability and robustness. Meanwhile, our VPTTA also shows competitive performance in domains where most methods exhibit satisfactory performance (such as D) and also performs well in domains where other methods exhibit suboptimal results (such as C).", + "bbox": [ + 496, + 585, + 893, + 707 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Comparison on the polyp segmentation task. We conducted the comparison experiment on the polyp segmentation task under the same experimental setup, as shown in Table 2. In contrast to the results of the OD/OC segmentation task, the methods based on model updating (i.e., TENT-continual, CoTTA, DLTTA, SAR, and DomainAdaptor) exhibit significantly worse performance in this task and are even inferior to the 'Source Only' baseline. It can be attributed to the fact that the polyps are relatively hidden, and the models tend to segment almost nothing rather than poorly segment under the impact of distribution shifts, resulting in confident and low entropy predictions but wrong", + "bbox": [ + 496, + 719, + 895, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "11189", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/94d284a1840b9156326e8cf956d20c3c1507350ec98d944cf0da0f56a1f69e5a.jpg", + "image_caption": [ + "Figure 3. Visualization of the original images, estimated prompts, and adapted images on the OD/OC segmentation task. We normalize the prompts to [0, 1] for better visualization. The DSC of applying the frozen source model on the original and adapted images is displayed below each image. We also show an example of each source domain on the left side of this diagram. 'Ori': Abbreviation of 'Original'." + ], + "image_footnote": [], + "bbox": [ + 101, + 88, + 869, + 498 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "gradients for these methods. Since DUA is backward-free, it still performs well in this task. However, DUA also fails when using domain A as the source domain and has limited performance due to the absence of training for the current test image. Our VPTTA simultaneously benefits from training on the current test image and refraining from updating the model, thus surpassing all competing methods and demonstrating advanced performance.", + "bbox": [ + 75, + 575, + 468, + 696 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Visualization of prompts and adapted images. We visualized the estimated prompts and adapted images on the OD/OC segmentation task in Figure 3 and displayed the DSC values predicted on the original and adapted images below the corresponding images. It can be observed that: (1) Applying prompts to the original test images can achieve significant performance gain; (2) Images from different target domains have similar prompts, proving that involving the memory bank to initialize the current prompt using previous ones is reasonable; (3) The appearances of adapted images are close to the source domain, explicitly demonstrating the reduction of distribution shifts via our VPTTA.", + "bbox": [ + 75, + 700, + 468, + 882 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Comparison in a long-term continual test-time adap", + "bbox": [ + 76, + 885, + 468, + 901 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "tation. To evaluate our VPTTA in a long-term continual test-time adaptation [42], we conducted the experiments for the OD/OC segmentation task multiple rounds and reported the results in Table 4. Since DUA is a backward-free method and DomainAdaptor resets the pre-trained model before each adaptation, they are unsuitable to be evaluated in this experiment. We compared our VPTTA with TENT, CoTTA, DLTTA, and SAR, all of which involve continuous training of the pre-trained model. The performance across different rounds with the same source domain reflects the extent of catastrophic forgetting in the model, and the average performance across multiple rounds highlights the error accumulation of the model within a series of distinct domains. The results reveal that both optimizing loss functions (CoTTA, DLTTA, SAR) and resetting the model (CoTTA, SAR) are effective for mitigating error accumulation and catastrophic forgetting, while TENT with vanilla entropy-minimization loss suffers from severe performance degradation. Additionally, after multiple rounds of testing, our VPTTA maintains superior performance with minimal performance degradation, emphasizing the effectiveness of", + "bbox": [ + 496, + 575, + 890, + 893 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "11190", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/4ace3b55ebda693e4bc20c346b4298793c6c6ffe67fc0d9ad5747e9f4adc1e7d.jpg", + "table_caption": [ + "Table 4. Performance of our VPTTA and four competing methods on the OD/OC segmentation task in a long-term continual test-time adaptation. The red numbers indicate the performance gain relative to the 'Source Only' baseline. The performance degradation is calculated between the overall average DSC and the average DSC of round 1. 'Ave': Abbreviation of 'Average'." + ], + "table_footnote": [], + "table_body": "
Time→
Round123Average DSC ↑Perform. Degra. ↓
MethodsABCDEAveABCDEAveABCDEAve
Source Only64.5376.0671.1852.6764.8765.8664.5376.0671.1852.6764.8765.8664.5376.0671.1852.6764.8765.8665.86
TENT73.0778.6671.9446.8170.2068.1362.0969.3270.6739.0268.2261.8657.0562.4770.2039.0266.3759.0263.01 (-2.85)
CoTTA75.3975.9869.1453.9970.4068.9874.3175.0067.9951.0468.2867.3273.2274.3366.7250.2367.0866.3267.54 (+1.68)
DLTTA75.1178.8573.8951.6469.7169.8474.1479.6574.2545.0569.0468.4372.2878.9372.8742.3769.2667.1468.47 (+2.61)
SAR74.5577.7170.7855.4071.7270.0374.7478.0971.0052.1369.0269.0074.9078.2471.1850.1668.4468.5869.20 (+3.34)
VPTTA (Ours)73.9179.3674.5156.5175.3571.9373.5778.8473.6156.9174.8071.5573.1278.4572.6357.1174.0471.0771.51 (+5.65)
", + "bbox": [ + 80, + 142, + 890, + 229 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "training prompts over updating the pre-trained model.", + "bbox": [ + 75, + 255, + 433, + 270 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Ablation study. To evaluate the contributions of the low-frequency prompt, memory-bank-based initialization, and warm-up mechanism, we conducted ablation studies on the OD/OC segmentation task, as shown in Table 3. Note that Eq. (3)/(5) is used as the loss function when training the prompt without/with our warm-up mechanism. The results show that (1) using the low-frequency prompt only has limited performance gain due to the poorly trained prompt; (2) both the memory bank and warm-up mechanism are beneficial to training the prompt; (3) the best performance is achieved when the memory bank and warm-up mechanism are jointly used, which means they are complementary.", + "bbox": [ + 75, + 271, + 470, + 452 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/585cc210003974459a61397dbb3425e8ea254f78960f4a424dc0e4ba883f8fb7.jpg", + "image_caption": [ + "Average DSC", + "Figure 4. Performance of our VPTTA with various $\\alpha$ on the OD/OC segmentation task." + ], + "image_footnote": [], + "bbox": [ + 84, + 484, + 459, + 571 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/6a6f9dace7fe6ab79db53b357d3d6f2f284c705985fe75caf97aa6473a009c5a.jpg", + "image_caption": [ + "Average DSC", + "Figure 5. Performance of our VPTTA with various $S$ , $K$ , and $\\tau$ on the OD/OC segmentation task." + ], + "image_footnote": [], + "bbox": [ + 83, + 650, + 459, + 742 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Analysis of hyperparameter $\\alpha$ , $S$ , $K$ , and $\\tau$ . We conducted the experiments for the OD/OC segmentation task to discuss the hyper-parameters used in our VPTTA and displayed the results in Figure 4 and Figure 5. It can be seen that $\\alpha = 0.01$ is optimal, as an excessively small prompt makes adaptation challenging, while an overly large prompt hinders effective training. Similar trends are also reflected", + "bbox": [ + 75, + 794, + 470, + 900 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/a3b57ec2acd4b3b02424cd527f60a9e5fe7676ad8c81a05db550a5dfe215d916.jpg", + "table_caption": [ + "Table 5. Performance of our VPTTA and VPTTA-LR on two segmentation tasks." + ], + "table_footnote": [], + "table_body": "
MethodsOD/OCPolyp
DSC↑Param↓DSC↑Eφmax↑Sα↑Param↓
VPTTA-LR (r=3)69.019,21678.2588.2784.206,336
VPTTA71.937580.4689.9486.8427
", + "bbox": [ + 504, + 292, + 883, + 339 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "in $S$ and $K$ : as they increase, the DSC values improve, while the excessively large $S$ and $K$ may introduce unrelated samples, affecting the initialization. With the increase of $\\tau$ , performance gradually approaches the upper bound and tends to remain stable, suggesting that the warm-up has reached saturation. Therefore, we finally chose $\\alpha = 0.01$ , $S = 40$ , $K = 16$ , and $\\tau = 5$ as the best configuration.", + "bbox": [ + 496, + 363, + 890, + 468 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Low-frequency prompt vs. low-rank prompt. To compare the effectiveness of VPTTA and VPTTA-LR, we repeated the experiments on two segmentation tasks and showed results in Table 5. It can be seen that VPTTA significantly outperforms VPTTA-LR, attributed to the fewer learnable parameters and superior representation of style texture in the frequency domain. Benefiting from the low-frequency prompt, VPTTA can be effectively trained with over a hundred times fewer parameters than VPTTA-LR.", + "bbox": [ + 496, + 469, + 892, + 604 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 618, + 619, + 633 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we proposed the VPTTA, a prompt-based method to alleviate the potential risks (i.e., error accumulation, and catastrophic forgetting) by freezing the pre-trained model. Specially, we presented the low-frequency prompt that trains a lightweight prompt with only a small number of parameters in a single iteration to alleviate the distribution shifts in BN layers, constructed a memory bank to better initialize each prompt by utilizing the prompts of previous and the most similar test images, and designed the statistics-fusion-based warm-up mechanism to simulate the warm-up statistics to overcome the training difficulty at the beginning of the inference phase. The extensive experimental results on two medical image segmentation benchmark tasks with multiple domains verified the superiority of our VPTTA over other methods and the effectiveness of each special design. In this regard, we hope that our efforts can offer a different paradigm for continual test-time adaptation.", + "bbox": [ + 496, + 643, + 893, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "11191", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Jorge Bernal, F Javier Sánchez, Gloria Fernández-Esparrach, Debora Gil, Cristina Rodríguez, and Fernando Vilarino. Wm-dova maps for accurate polyp highlighting in colonoscopy: Validation vs. saliency maps from physicians. Comput. Med. Imaging and Graph., 43:99-111, 2015. 5", + "[2] Paola Cascante-Bonilla, Fuwen Tan, Yanjun Qi, and Vicente Ordonez. Curriculum labeling: Revisiting pseudo-labeling for semi-supervised learning. In AAAI, pages 6912-6920, 2021. 2, 3", + "[3] Oscar Chang, Lampros Flokas, and Hod Lipson. Principled weight initialization for hypernetworks. In Int. Conf. Learn. Represent., 2019. 2", + "[4] Dian Chen, Dequan Wang, Trevor Darrell, and Sayna Ebrahimi. Contrastive test-time adaptation. In IEEE Conf. Comput. Vis. Pattern Recog., pages 295-305, 2022. 3", + "[5] Lin Chen, Huaian Chen, Zhixiang Wei, Xin Jin, Xiao Tan, Yi Jin, and Enhong Chen. Reusing the task-specific classifier as a discriminator: Discriminator-free adversarial domain adaptation. In IEEE Conf. Comput. Vis. Pattern Recog., pages 7181-7190, 2022. 2", + "[6] Mario Döbler, Robert A Marsden, and Bin Yang. Robust mean teacher for continual and gradual test-time adaptation. In IEEE Conf. Comput. Vis. Pattern Recog., pages 7704-7714, 2023. 2, 3", + "[7] Deng-Ping Fan, Ming-Ming Cheng, Yun Liu, Tao Li, and Ali Borji. Structure-measure: A new way to evaluate foreground maps. In Int. Conf. Comput. Vis., pages 4548-4557, 2017. 5", + "[8] Deng-Ping Fan, Cheng Gong, Yang Cao, Bo Ren, Ming-Ming Cheng, and Ali Borji. Enhanced-alignment measure for binary foreground map evaluation. In IJCAI, pages 698-704, 2018. 5", + "[9] Deng-Ping Fan, Ge-Peng Ji, Tao Zhou, Geng Chen, Huazhu Fu, Jianbing Shen, and Ling Shao. Pranet: Parallel reverse attention network for polyp segmentation. In Int. Conf. Med. Image Comput. Comput.-Assist. Intervention., pages 263-273. Springer, 2020. 5", + "[10] Matteo Frigo and Steven G Johnson. FFTW: An adaptive software architecture for the fft. In ICASSP, pages 1381-1384. IEEE, 1998. 3", + "[11] Francisco Fumero, Silvia Alayón, José L Sanchez, Jose Sigut, and M Gonzalez-Hernandez. RIM-ONE: An open retinal image database for optic nerve evaluation. In Int. Symp. Comput.-based Med. Syst., pages 1-6. IEEE, 2011. 5", + "[12] Yulu Gan, Yan Bai, Yihang Lou, Xianzheng Ma, Renrui Zhang, Nian Shi, and Lin Luo. Decorate the newcomers: Visual domain prompt for continual test time adaptation. In AAAI, pages 7595-7603, 2023. 3", + "[13] Yossi Gandelsman, Yu Sun, Xinlei Chen, and Alexei Efros. Test-time training with masked autoencoders. Adv. Neural Inform. Process. Syst., 35:29374-29385, 2022. 3", + "[14] Shang-Hua Gao, Ming-Ming Cheng, Kai Zhao, Xin-Yu Zhang, Ming-Hsuan Yang, and Philip Torr. Res2net: A new multi-scale backbone architecture. IEEE Trans. Pattern Anal. Mach. Intell., 43(2):652-662, 2019. 5" + ], + "bbox": [ + 78, + 114, + 470, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[15] Yunhe Gao, Xingjian Shi, Yi Zhu, Hao Wang, Zhiqiang Tang, Xiong Zhou, Mu Li, and Dimitris N Metaxas. Visual prompt tuning for test-time domain adaptation. arXiv preprint arXiv:2210.04831, 2022. 3", + "[16] Mohsen Ghafoorian, Alireza Mehrtash, Tina Kapur, Nico Karssemeijer, Elena Marchiori, Mehran Pesteie, Charles RG Guttmann, Frank-Erik de Leeuw, Clare M Tempany, Bram Van Ginneken, et al. Transfer learning for domain adaptation in mri: Application in brain lesion segmentation. In Int. Conf. Med. Image Comput. Comput.-Assist. Intervent., pages 516-524. Springer, 2017. 2", + "[17] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In IEEE Conf. Comput. Vis. Pattern Recog., pages 770-778, 2016. 5", + "[18] Edward J Hu, Phillip Wallis, Zeyuan Allen-Zhu, Yuzhhi Li, Shean Wang, Lu Wang, Weizhu Chen, et al. Lora: Low-rank adaptation of large language models. In Int. Conf. Learn. Represent., 2022. 4", + "[19] Shishuai Hu, Zehui Liao, and Yong Xia. Prosfda: Prompt learning based source-free domain adaptation for medical image segmentation. arXiv preprint arXiv:2211.11514, 2022. 2, 3, 4, 5", + "[20] Debesh Jha, Pia H Smedsrud, Michael A Riegler, Pål Halvorsen, Thomas de Lange, Dag Johansen, and Håvard D Johansen. Kvasir-seg: A segmented polyp dataset. In Int. Conf. Multimedia Modeling, pages 451-462. Springer, 2020. 5", + "[21] Christian Lamers, René Vidal, Nabil Belbachir, Niki van Stein, Thomas Bäeck, and Paris Giampouras. Clustering-based domain-incremental learning. In Int. Conf. Comput. Vis., pages 3384-3392, 2023. 2, 3", + "[22] Joonseok Lee, Seungyeon Kim, Guy Lebanon, and Yoram Singer. Local low-rank matrix approximation. In Int. Conf. Mach. Learn., pages 82-90. PMLR, 2013. 4", + "[23] Jian Liang, Ran He, Zhenan Sun, and Tieniu Tan. Exploring uncertainty in pseudo-label guided unsupervised domain adaptation. Pattern Recognition, 96:106996, 2019. 2, 3", + "[24] Pengfei Liu, Weizhe Yuan, Jinlan Fu, Zhengbao Jiang, Hiroaki Hayashi, and Graham Neubig. Pre-train, prompt, and predict: A systematic survey of prompting methods in natural language processing. ACM Comput. Surveys, 55(9):1-35, 2023. 3", + "[25] M Jehanzeb Mirza, Jakub Micorek, Horst Possegger, and Horst Bischof. The norm must go on: Dynamic unsupervised domain adaptation by normalization. In IEEE Conf. Comput. Vis. Pattern Recog., pages 14765-14775, 2022. 3, 6", + "[26] Chaithanya Kumar Mummadi, Robin Hutmacher, Kilian Rambach, Evgeny Levinkov, Thomas Brox, and Jan Hendrik Metzen. Test-time adaptation to distribution shift by confidence maximization and input transformation. arXiv preprint arXiv:2106.14999, 2021. 3", + "[27] Zachary Nado, Shreyas Padhy, D Sculley, Alexander D'Amour, Balaji Lakshminarayanan, and Jasper Snoek. Evaluating prediction-time batch normalization for robustness under covariate shift. arXiv preprint arXiv:2006.10963, 2020. 3, 4" + ], + "bbox": [ + 501, + 92, + 890, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "11192", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[28] Phan Ngoc Lan, Nguyen Sy An, Dao Viet Hang, Dao Van Long, Tran Quang Trung, Nguyen Thi Thuy, and Dinh Viet Sang. Neounet: Towards accurate colon polyp segmentation and neoplasm detection. In Adv. Vis. Comput. Int. Symp., pages 15-28. Springer, 2021. 5", + "[29] A Tuan Nguyen, Thanh Nguyen-Tang, Ser-Nam Lim, and Philip HS Torr. Tipi: Test time adaptation with transformation invariance. In IEEE Conf. Comput. Vis. Pattern Recog., pages 24162-24171, 2023. 2, 3", + "[30] Shuaicheng Niu, Jiaxiang Wu, Yifan Zhang, Yafo Chen, Shijian Zheng, Peilin Zhao, and Mingkui Tan. Efficient test-time model adaptation without forgetting. In Int. Conf. Mach. Learn., pages 16888-16905. PMLR, 2022. 2, 3", + "[31] Shuaicheng Niu, Jiaxiang Wu, Yifan Zhang, Zhiquan Wen, Yaofo Chen, Peilin Zhao, and Mingkui Tan. Towards stable test-time adaptation in dynamic wild world. In Int. Conf. Learn. Represent., 2023. 2, 3, 6", + "[32] José Ignacio Orlando, Huazhu Fu, João Barbosa Breda, Karel van Keer, Deepti R Bathula, Andrés Diaz-Pinto, Ruogu Fang, Pheng-Ann Heng, Jyoung Kim, JoonHo Lee, et al. REFuge Challenge: A unified framework for evaluating automated methods for glaucoma assessment from fundus photographs. Med. Image Anal., 59:101570, 2020. 5", + "[33] Xingang Pan, Ping Luo, Jianping Shi, and Xiaou Tang. Two at once: Enhancing learning and generalization capacities via ibn-net. In *Eur. Conf. Comput. Vis.*, pages 464-479, 2018. 2", + "[34] Swami Sankaranarayanan, Yogesh Balaji, Arpit Jain, Ser Nam Lim, and Rama Chellappa. Learning from synthetic data: Addressing domain shift for semantic segmentation. In IEEE Conf. Comput. Vis. Pattern Recog., pages 3752-3761, 2018. 2", + "[35] Steffen Schneider, Evgenia Rusak, Luisa Eck, Oliver Bringmann, Wieland Brendel, and Matthias Bethge. Improving robustness against common corruptions by covariate shift adaptation. Adv. Neural Inform. Process. Syst., 33:11539-11551, 2020. 2", + "[36] Juan Silva, Aymeric Histace, Olivier Romain, Xavier Dray, and Bertrand Granado. Toward embedded detection of polyps in wce images for early diagnosis of colorectal cancer. Int. J. Comput.-Assist. Rad. Sur., 9(2):283-293, 2014. 5", + "[37] Samarth Sinha, Peter Gehler, Francesco Locatello, and Bernt Schiele. Test: Test-time self-training under distribution shift. In IEEE Wint. Conf. Appl. Comput. Vis., pages 2759-2769, 2023. 2, 3", + "[38] Jayanthi Sivaswamy, SR Krishnadas, Gopal Datt Joshi, Madhulika Jain, and A Ujjwaft Syed Tabish. Drishti-GS: Retinal image dataset for optic nerve head (onh) segmentation. In IEEE Int. Symp. on Bio. Imaging, pages 53-56. IEEE, 2014. 5", + "[39] Junha Song, Jungsoo Lee, In So Kweon, and Sungha Choi. Ecotta: Memory-efficient continual test-time adaptation via self-distilled regularization. In IEEE Conf. Comput. Vis. Pattern Recog., pages 11920-11929, 2023. 2, 3", + "[40] Yu Sun, Xiaolong Wang, Zhuang Liu, John Miller, Alexei Efros, and Moritz Hardt. Test-time training with self-supervision for generalization under distribution shifts. In" + ], + "bbox": [ + 78, + 90, + 468, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Int. Conf. Mach. Learn., pages 9229-9248. PMLR, 2020. 2, 3", + "[41] Dequan Wang, Evan Shelhamer, Shaoteng Liu, Bruno Olshausen, and Trevor Darrell. Tent: Fully test-time adaptation by entropy minimization. In Int. Conf. Learn. Represent., 2021. 2, 3, 5, 6", + "[42] Qin Wang, Olga Fink, Luc Van Gool, and Dengxin Dai. Continual test-time domain adaptation. In IEEE Conf. Comput. Vis. Pattern Recog., pages 7201-7211, 2022. 2, 3, 6, 7", + "[43] Shujun Wang, Lequan Yu, Kang Li, Xin Yang, Chi-Wing Fu, and Pheng-Ann Heng. Boundary and entropy-driven adversarial learning for fundus image segmentation. In Int. Conf. Med. Image Comput. Comput.-Assist. Intervention., pages 102-110. Springer, 2019. 2", + "[44] Wei Wang, Zhun Zhong, Weijie Wang, Xi Chen, Charles Ling, Boyu Wang, and Nicu Sebe. Dynamically instance-guided adaptation: A backward-free approach for test-time domain adaptive semantic segmentation. In IEEE Conf. Comput. Vis. Pattern Recog., pages 24090-24099, 2023. 2, 3", + "[45] Hongzheng Yang, Cheng Chen, Meirui Jiang, Quande Liu, Jianfeng Cao, Pheng Ann Heng, and Qi Dou. Dltta: Dynamic learning rate for test-time adaptation on cross-domain medical images. IEEE Trans. Med. Imaging, 41(12):3575-3586, 2022. 2, 3, 4, 5, 6", + "[46] Yanchao Yang and Stefano Soatto. FDA: Fourier domain adaptation for semantic segmentation. In IEEE Conf. Comput. Vis. Pattern Recog., pages 4085-4095, 2020. 2", + "[47] Jian Zhang, Lei Qi, Yinghuan Shi, and Yang Gao. Domain adaptor: A novel approach to test-time adaptation. In IEEE Conf. Comput. Vis. Pattern Recog., pages 18971-18981, 2023. 2, 3, 4, 6", + "[48] Marvin Zhang, Sergey Levine, and Chelsea Finn. Memo: Test time robustness via adaptation and augmentation. Adv. Neural Inform. Process. Syst., 35:38629-38642, 2022. 2, 3", + "[49] Yizhe Zhang, Tao Zhou, Yuhui Tao, Shuo Wang, Ye Wu, Benyuan Liu, Pengfei Gu, Qiang Chen, and Danny Z Chen. TestFit: A plug-and-play one-pass test time method for medical image segmentation. Medical Image Analysis, 92: 103069, 2024. 3", + "[50] Zhuo Zhang, Feng Shou Yin, Jiang Liu, Wing Kee Wong, Ngan Meng Tan, Beng Hai Lee, Jun Cheng, and Tien Yin Wong. ORIGA-light: An online retinal fundus image database for glaucoma analysis and research. In Int. Conf. IEEE Eng. in Med. and Bio., pages 3065-3068. IEEE, 2010. 5", + "[51] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Zwei Liu. Conditional prompt learning for vision-language models. In IEEE Conf. Comput. Vis. Pattern Recog., pages 16816-16825, 2022. 3", + "[52] Chen Zhu, Renkun Ni, Zheng Xu, Kezhi Kong, W Ronny Huang, and Tom Goldstein. Gradinit: Learning to initialize neural networks for stable and efficient training. Adv. Neural Inform. Process. Syst., 34:16410-16422, 2021. 2", + "[53] Yukun Zuo, Hantao Yao, and Changsheng Xu. Attention-based multi-source domain adaptation. IEEE Trans. Image Process., 30:3793-3803, 2021. 2" + ], + "bbox": [ + 501, + 92, + 890, + 896 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "11193", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2024/Each Test Image Deserves A Specific Prompt_ Continual Test-Time Adaptation for 2D Medical Image Segmentation/b547c627-f0bf-4b29-9f5d-10851e261fb1_model.json b/2024/Each Test Image Deserves A Specific Prompt_ Continual Test-Time Adaptation for 2D Medical Image Segmentation/b547c627-f0bf-4b29-9f5d-10851e261fb1_model.json new file mode 100644 index 0000000000000000000000000000000000000000..8f4ae5595f83450a4484ea38ccc78bc311a58118 --- /dev/null +++ b/2024/Each Test Image Deserves A Specific Prompt_ Continual Test-Time Adaptation for 2D Medical Image Segmentation/b547c627-f0bf-4b29-9f5d-10851e261fb1_model.json @@ -0,0 +1,2508 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.809, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.089, + 0.131, + 0.882, + 0.177 + ], + "angle": 0, + "content": "Each Test Image Deserves A Specific Prompt: Continual Test-Time Adaptation for 2D Medical Image Segmentation" + }, + { + "type": "text", + "bbox": [ + 0.173, + 0.203, + 0.794, + 0.222 + ], + "angle": 0, + "content": "Ziyang Chen\\(^{1}\\) Yongsheng Pan\\(^{1\\dagger}\\) Yiwen Ye\\(^{1}\\) Mengkang Lu\\(^{1}\\) Yong Xia\\(^{1,2,3\\dagger}\\)" + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.222, + 0.873, + 0.276 + ], + "angle": 0, + "content": "\\(^{1}\\) School of Computer Science and Engineering, Northwestern Polytechnical University, China \n\\(^{2}\\) Research & Development Institute of Northwestern Polytechnical University in Shenzhen, China \n\\(^{3}\\) Ningbo Institute of Northwestern Polytechnical University, China" + }, + { + "type": "text", + "bbox": [ + 0.197, + 0.278, + 0.773, + 0.293 + ], + "angle": 0, + "content": "zychen@mail.nwpu.edu.cn, yspan@nwpu.edu.cn, ywye@mail.nwpu.edu.cn" + }, + { + "type": "text", + "bbox": [ + 0.315, + 0.296, + 0.652, + 0.31 + ], + "angle": 0, + "content": "lmk@mail.nwpu.edu.cn, yxia@nwpu.edu.cn" + }, + { + "type": "image_caption", + "bbox": [ + 0.083, + 0.345, + 0.184, + 0.36 + ], + "angle": 0, + "content": "Source Model \\(f_{\\theta}\\)" + }, + { + "type": "image", + "bbox": [ + 0.092, + 0.361, + 0.171, + 0.411 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.178, + 0.347, + 0.279, + 0.415 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.286, + 0.347, + 0.361, + 0.359 + ], + "angle": 0, + "content": "Domain 2" + }, + { + "type": "image", + "bbox": [ + 0.287, + 0.359, + 0.36, + 0.415 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.36, + 0.347, + 0.444, + 0.359 + ], + "angle": 0, + "content": "Domain 3" + }, + { + "type": "image", + "bbox": [ + 0.36, + 0.359, + 0.474, + 0.415 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.474, + 0.347, + 0.534, + 0.359 + ], + "angle": 0, + "content": "Domain 4" + }, + { + "type": "image", + "bbox": [ + 0.474, + 0.359, + 0.537, + 0.415 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.537, + 0.347, + 0.59, + 0.359 + ], + "angle": 0, + "content": "" + }, + { + "type": "image", + "bbox": [ + 0.537, + 0.359, + 0.573, + 0.414 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.573, + 0.347, + 0.655, + 0.359 + ], + "angle": 0, + "content": "Domain N" + }, + { + "type": "image", + "bbox": [ + 0.573, + 0.359, + 0.657, + 0.415 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.657, + 0.347, + 0.724, + 0.359 + ], + "angle": 0, + "content": "" + }, + { + "type": "image", + "bbox": [ + 0.657, + 0.359, + 0.723, + 0.415 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.379, + 0.417, + 0.427, + 0.432 + ], + "angle": 0, + "content": "\\(f_{\\theta}^{\\prime}\\rightarrow f_{\\theta}\\)" + }, + { + "type": "image_caption", + "bbox": [ + 0.535, + 0.426, + 0.564, + 0.436 + ], + "angle": 0, + "content": "Test" + }, + { + "type": "image_caption", + "bbox": [ + 0.564, + 0.419, + 0.675, + 0.435 + ], + "angle": 0, + "content": "\\(\\mathcal{L}_{self} \\to \\mathcal{L}_{prompt}\\)" + }, + { + "type": "image", + "bbox": [ + 0.533, + 0.439, + 0.564, + 0.49 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.564, + 0.439, + 0.677, + 0.49 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.677, + 0.448, + 0.725, + 0.477 + ], + "angle": 0, + "content": "Adapted Output" + }, + { + "type": "image", + "bbox": [ + 0.615, + 0.49, + 0.641, + 0.505 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.581, + 0.508, + 0.68, + 0.519 + ], + "angle": 0, + "content": "(c) Our VPTTA" + }, + { + "type": "image_caption", + "bbox": [ + 0.567, + 0.521, + 0.614, + 0.537 + ], + "angle": 0, + "content": "EA:" + }, + { + "type": "image_caption", + "bbox": [ + 0.624, + 0.523, + 0.696, + 0.537 + ], + "angle": 0, + "content": "CF:" + }, + { + "type": "image", + "bbox": [ + 0.755, + 0.36, + 0.783, + 0.379 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.784, + 0.365, + 0.849, + 0.377 + ], + "angle": 0, + "content": "Freeze" + }, + { + "type": "image", + "bbox": [ + 0.757, + 0.386, + 0.782, + 0.408 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.807, + 0.365, + 0.849, + 0.377 + ], + "angle": 0, + "content": "Freeze" + }, + { + "type": "image_caption", + "bbox": [ + 0.807, + 0.391, + 0.865, + 0.402 + ], + "angle": 0, + "content": "Fine-tune" + }, + { + "type": "image", + "bbox": [ + 0.757, + 0.415, + 0.781, + 0.437 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.807, + 0.42, + 0.853, + 0.432 + ], + "angle": 0, + "content": "Prompt" + }, + { + "type": "image", + "bbox": [ + 0.755, + 0.447, + 0.784, + 0.464 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.807, + 0.448, + 0.856, + 0.459 + ], + "angle": 0, + "content": "Terrible" + }, + { + "type": "image", + "bbox": [ + 0.756, + 0.476, + 0.783, + 0.495 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.807, + 0.478, + 0.86, + 0.49 + ], + "angle": 0, + "content": "Passable" + }, + { + "type": "image", + "bbox": [ + 0.756, + 0.504, + 0.784, + 0.525 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.807, + 0.508, + 0.844, + 0.519 + ], + "angle": 0, + "content": "Good" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.548, + 0.894, + 0.593 + ], + "angle": 0, + "content": "Figure 1. Comparison between our VPTTA and existing solutions under the CTTA setup. Our VPTTA avoids both error accumulation (EA) and catastrophic forgetting (CF) by freezing model parameters and achieves adaptation by training a prompt for each test image. The commonly used self-supervised loss and the optimized one are denoted by \\(\\mathcal{L}_{self}\\) and \\(\\mathcal{L}_{self}^{\\prime}\\), respectively." + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.605, + 0.314, + 0.621 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.645, + 0.473, + 0.903 + ], + "angle": 0, + "content": "Distribution shift widely exists in medical images acquired from different medical centres and poses a significant obstacle to deploying the pre-trained semantic segmentation model in real-world applications. Test-time adaptation has proven its effectiveness in tackling the cross-domain distribution shift during inference. However, most existing methods achieve adaptation by updating the pre-trained models, rendering them susceptible to error accumulation and catastrophic forgetting when encountering a series of distribution shifts (i.e., under the continual test-time adaptation setup). To overcome these challenges caused by updating the models, in this paper, we freeze the pre-trained model and propose the Visual Prompt-based Test-Time Adaptation (VPTTA) method to train a specific prompt for each test image to align the statistics in the batch normalization layers. Specifically, we present the low-frequency prompt, which is lightweight with only a few parameters" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.606, + 0.895, + 0.773 + ], + "angle": 0, + "content": "and can be effectively trained in a single iteration. To enhance prompt initialization, we equip VPTTA with a memory bank to benefit the current prompt from previous ones. Additionally, we design a warm-up mechanism, which mixes source and target statistics to construct warm-up statistics, thereby facilitating the training process. Extensive experiments demonstrate the superiority of our VPTTA over other state-of-the-art methods on two medical image segmentation benchmark tasks. The code and weights of pre-trained source models are available at https://github.com/Chen-Ziyang/VPTTA." + }, + { + "type": "page_footnote", + "bbox": [ + 0.498, + 0.815, + 0.895, + 0.901 + ], + "angle": 0, + "content": "\\(\\dagger\\)Yong Xia and Yongsheng Pan are the corresponding authors. This work was supported in part by the National Natural Science Foundation of China under Grants 62171377, in part by STI2030-Major Project under Grants 2022ZD0213100, in part by Shenzhen Science and Technology Program under Grants JCYJ20220530161616036, and in part by the Ningbo Clinical Research Center for Medical Imaging under Grant 2021L003 (Open Project 2022LYKFZD06)." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "11184" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.079, + 0.09, + 0.208, + 0.106 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.115, + 0.473, + 0.418 + ], + "angle": 0, + "content": "Semantic segmentation serves a pivotal role in medical image processing to outline specific anatomical structures. However, the segmentation model pre-trained on the source dataset may be hindered from deploying in real-world applications due to the commonly existing distribution shift, which is typically caused by the variations in imaging protocols, operators, and scanners [16, 34]. Test-time adaptation (TTA) has emerged as a promising paradigm to replace domain adaptation (DA) [5, 43, 46, 53], as TTA requires only test data during the inference phase, while DA necessitates a sufficiently large and representative target dataset [45]. The mainstream TTA methods [29-31, 37, 40, 41, 47, 48] leverage the self-supervised tasks to construct auxiliary losses, such as entropy minimization [41] and self-training [37], to update the parameters of pre-trained models. Although these TTA methods successfully alleviate the domain gap and performance degradation, they are based on the assumption that the target domains exhibit static distributions. Unfortunately, the target domains are continuously changing rather than static in most real-world scenarios." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.418, + 0.473, + 0.78 + ], + "angle": 0, + "content": "This was first presented as the continual test-time adaptation (CTTA) setup [42] that addresses TTA on a series of distribution shifts. Under this setup, pre-trained models are required to adapt to several distinct domains and therefore are more prone to suffer error accumulation and catastrophic forgetting. As mentioned above, many TTA methods rely on self-supervised tasks and may be affected by the noisy losses due to unreliable supervision, leading to error accumulation [2, 23]. Catastrophic forgetting typically occurs when the model is continually trained on new domains, resulting in unexpected performance degradation [21, 30]. Some research has been studied in addressing these two challenges based on optimizing the loss functions for better supervision [6, 31, 42, 42, 45] (see Figure 1 (a)) or prevent the model parameters from excessive alterations by model resetting or introducing regularizers [30, 31, 39, 42] (see Figure 1 (b)). Despite their efforts, loss optimization cannot ensure a robust training process to avoid catastrophic forgetting during the long-term inference, insufficient or excessive updating of the model parameters may lead to inadequate adaptation or error accumulation. Since the pretrained model is updated by the self-supervised loss where the noise cannot be completely removed, these methods may still suffer from these two issues." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.781, + 0.472, + 0.904 + ], + "angle": 0, + "content": "Considering the above limitations in CTTA, updating the pre-trained model during inference seems to be inappropriate. To this end, we propose the Visual Prompt-based Test-Time Adaptation (VPTTA) method to avoid error accumulation and catastrophic forgetting associated with updating model by freezing the model parameters and conduct adaptation by learning a specific visual prompt for each test image (see Figure 1 (c)). The prompt can adapt each test im" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.894, + 0.138 + ], + "angle": 0, + "content": "age to the frozen pre-trained model to take into account both adaptation ability and knowledge retaining. We attempt to achieve this by answering the following three questions." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.14, + 0.895, + 0.321 + ], + "angle": 0, + "content": "- How to design the prompt? Each prompt is restricted to a limited number of training iterations (sometimes just one) to satisfy the demands of online inference, requiring the prompt to be lightweight. Based on this principle, we introduce the prompt for only the low-frequency components of each sample, resulting in the low-frequency prompt. Since the low-frequency components are strongly associated with style textures and can serve as the primary source of distribution shift [46], our low-frequency prompt allows for maintaining lightweight and being able to reduce distribution shifts by modifying the low-frequency components." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.321, + 0.895, + 0.487 + ], + "angle": 0, + "content": "- How to initialize the prompt? It is well known that model training can benefit from proper parameter initialization, resulting in not only fewer training iterations but also improved performance [3, 52]. We argue that this is also applicable to prompt training and therefore construct a memory bank to store the prompts and low-frequency components of previous test images. By comparing the similarity between the low-frequency components of the current and previous images, we select several historical prompts of the most similar images to initialize the current prompt." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.487, + 0.895, + 0.655 + ], + "angle": 0, + "content": "- How to train the prompt? Since the statistics mismatch between the source and target domains in the batch normalization (BN) layers is a significant cause of distribution shift [33, 35, 44], we optimize the prompt with the objective of minimizing the distances between BN statistics and the statistics extracted from the features of test data for statistics alignment [19]. To further address the training difficulty posed by the large distribution shift at the beginning of the inference phase, we devise a statistics-fusion-based warm-up mechanism to assist in learning the prompt from easy to hard." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.656, + 0.895, + 0.809 + ], + "angle": 0, + "content": "To summarise, our contributions are three-fold: (1) We present the low-frequency prompt to adjust the test image to adapt the model instead of updating the model and therefore prevent the model from error accumulation and catastrophic forgetting. (2) We propose the novel VPTTA framework to alleviate the distribution shift by training a sample-specific visual prompt, which is lightweight, well-initialized, and follows a gradual learning process. (3) The powerful framework we provide can be effectively deployed, requiring only a single iteration and a few dozen parameters for adaptation." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.822, + 0.642, + 0.838 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.847, + 0.753, + 0.864 + ], + "angle": 0, + "content": "2.1. Test-time Adaptation (TTA)" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.894, + 0.901 + ], + "angle": 0, + "content": "TTA aims to adapt the model pre-trained on the source domain to the test data during inference in a source-free" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "11185" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.091, + 0.473, + 0.378 + ], + "angle": 0, + "content": "and online manner [40, 41]. The mainstream TTA methods are based on model updating, which constructs self-supervised auxiliary tasks to guide the model to learn on the test data [4, 13, 26, 29, 37, 40, 41, 48, 49]. Sun et al. [40] utilized an auxiliary task to predict the rotation to assist the model learning on the test data. Wang et al. [41] presented a test-time entropy minimization scheme to reduce generalization error by reducing the entropy of model predictions on test data. Zhang et al. [49] introduced a supplementary network to fit the test data, which can work in a plug-and-play fashion, necessitating minimal hyperparameter tuning. Some other methods are based on BN statistics, which alleviate the distribution shifts by modifying the statistics within BN layers [25, 27, 44]. Mirza et al. [25] constantly adapted the statistics of the BN layers to modify the feature representations of the model. These TTA methods improve the adaptation performance on the static target domain but neglect the continually changing target domains in most real-world scenarios." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.38, + 0.473, + 0.713 + ], + "angle": 0, + "content": "Recently, the continual test-time adaptation (CTTA) [42] is proposed to tackle TTA under a series of continually changing target domains. Due to long-term inference, both error accumulation [2, 23] and catastrophic forgetting [21, 30] are more prone to occur in this setup. Many methods [6, 30, 31, 39, 42, 45, 47] optimize the self-supervised losses or prevent excessive alterations to the model parameters via model resetting or regularizer to improve this situation but are still impacted by these issues. Wang et al. [42] refined pseudo labels by the weight- and augmentation-averaged predictions and randomly restored a small part of model parameters. Niu et al. [31] removed partial noisy samples with large gradients and encouraged model weights toward a flat minimum. Yang et al. [45] designed a dynamic strategy to adjust the learning rate for each test image. Zhang et al. [47] presented to adaptively fuse source and test statistics and improve the entropy minimization loss to a generalized one. In contrast, our proposed VPTTA employs a learnable prompt for each test image to mitigate distribution shifts during inference, requiring no changes to the pre-trained model and therefore alleviating the error accumulation and catastrophic forgetting." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.725, + 0.248, + 0.743 + ], + "angle": 0, + "content": "2.2. Prompt Learning" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.473, + 0.903 + ], + "angle": 0, + "content": "Prompt learning was initially employed to devise additional text instructions for input text to fine-tune the largescale Natural Language Processing (NLP) models on downstream tasks [24]. With remarkable success in NLP, it has also been involved in computer vision and greatly benefited visual models in recent years [12, 15, 19, 51]. Jia et al. [15] introduced a small number of trainable parameters in the input space while keeping the model backbone frozen to fine-tune the model with high quality and efficiency. Zhou et al. [51] presented the conditional context optimization" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.304 + ], + "angle": 0, + "content": "to generate an input-conditional token for each image to generalize the pre-trained vision-language model to unseen classes. Hu et al. [19] employed the visual prompt at the pixel level and added it to the input images to explicitly minimize the domain discrepancy. Gan et al. [12] learned a domain-specific and domain-agnostic visual prompt for adaptation by using the exponential-moving-average strategy under the guidance of self-training. Different from them, our VPTTA presents the low-frequency prompt to address a series of changing distribution shifts with a few dozen parameters. We further consider optimizing the initialization and warm-up mechanism, enabling our visual prompts to be effectively trained within a single iteration to satisfy the demands of CTTA." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.315, + 0.637, + 0.333 + ], + "angle": 0, + "content": "3. Methodology" + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.34, + 0.859, + 0.357 + ], + "angle": 0, + "content": "3.1. Problem Definition and Method Overview" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.363, + 0.892, + 0.472 + ], + "angle": 0, + "content": "Let the labeled source domain dataset and unlabeled testing target dataset be \\(\\mathbb{D}^s = \\{X_i^s,Y_i^s\\}_{i = 1}^{N^s}\\) and \\(\\mathbb{D}^t = \\{X_i^t\\}_{i = 1}^{N^t}\\), respectively, where \\(X_{i}^{*}\\in \\mathbb{R}^{H\\times W\\times C}\\) is the \\(i\\)-th image with C channels and size of \\(H\\times W\\), and \\(Y_{i}^{*}\\) is its label. Our goal is to learn a low-frequency prompt \\(\\mathcal{P}_i\\) for each test image \\(X_{i}^{t}\\) to obtain the modified image \\(\\tilde{X}_i^t\\) that can be adapted to the model \\(f_{\\theta}:X\\to Y\\) pre-trained on \\(\\mathbb{D}^s\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.473, + 0.895, + 0.655 + ], + "angle": 0, + "content": "To accomplish this goal, our VPTTA considers three key facets: prompt design, prompt initialization, and prompt training. For prompt design, the prompt is designed to change the low-frequency component of the amplitude of the test image. For prompt initialization, we adopt a memory bank to initialize the current prompt with previous ones. For prompt training, we employ the absolute distance to train the prompt to align the statistics between features extracted from test images and the batch normalization (BN) layers and also present a statistics-fusion-based warm-up mechanism. An overview of our VPTTA is illustrated in Figure 2. We now delve into its details." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.662, + 0.844, + 0.679 + ], + "angle": 0, + "content": "3.2. Prompt Design: Low-frequency Prompt" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.685, + 0.892, + 0.792 + ], + "angle": 0, + "content": "Let \\(\\mathcal{F}(\\cdot)\\) and \\(\\mathcal{F}^{-1}(\\cdot)\\) be the Fast Fourier Transform (FFT) and Inverse Fast Fourier Transform (IFFT) operations [10], respectively. The amplitude component and phase component are denoted by \\(\\mathcal{F}^A (\\cdot)\\) and \\(\\mathcal{F}^P (\\cdot)\\) respectively. Note that the low-frequency component in \\(\\mathcal{F}^A (\\cdot)\\) is shifted to the center. Utilizing the prompt \\(\\mathcal{P}_i\\in \\mathbb{R}^{(\\alpha \\times H)\\times (\\alpha \\times W)\\times C}\\) to obtain the adapted image \\(\\bar{X}_i^t\\) can be formalized as" + }, + { + "type": "equation", + "bbox": [ + 0.515, + 0.799, + 0.892, + 0.818 + ], + "angle": 0, + "content": "\\[\n\\tilde {X} _ {i} ^ {t} = \\mathcal {F} ^ {- 1} ([ \\text {O n e P a d} (\\mathcal {P} _ {i}) \\odot \\mathcal {F} ^ {A} (X _ {i} ^ {t}), \\mathcal {F} ^ {P} (X _ {i} ^ {t}) ]) \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.825, + 0.895, + 0.902 + ], + "angle": 0, + "content": "where \\(\\odot\\) is the element-wise multiplication, and OnePad means padding one around \\(\\mathcal{P}_i\\) to expand it to the size of \\(H\\times W\\). Under the control of the small coefficient \\(\\alpha \\in (0,1)\\), we can ensure that the prompt \\(\\mathcal{P}_i\\) remains lightweight and focuses on the low-frequency component." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "11186" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.123, + 0.089, + 0.852, + 0.407 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.414, + 0.896, + 0.513 + ], + "angle": 0, + "content": "Figure 2. Overview of our VPTTA. For each test image, (1) the Fast Fourier Transform (FFT) is first applied to transform it into the frequency domain, where the low-frequency component of amplitude is used to query in a memory bank to initialize the current prompt, and then the amplitude is multiplied with the prompt at the low-frequency component and transformed back to the spatial domain using the Inverse Fast Fourier Transform (IFFT). (2) The memory bank is built on the previous low-frequency components and their corresponding prompts and is updated using the First In First Out (FIFO) strategy. (3) We convert the source statistics stored in BN layers into the warm-up statistics and calculate the absolute distance between warm-up and target statistics as the loss to fine-tune the prompt. (4) Finally, we feed the image corrected by its fine-tuned prompt to the pre-trained model to produce the output. 'Stat': Abbreviation of 'Statistics'." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.539, + 0.473, + 0.66 + ], + "angle": 0, + "content": "Simultaneously, we recognize that the low-rank decomposition is widely employed in the field of data compression [22] and model fine-tuning [18]. Consequently, we also design a low-rank prompt \\(\\mathcal{P}_i^{LoRa} = \\mathcal{B}_i@A_i\\) as a variant (denoted as VPTTA-LR), where \\(\\mathcal{B}_i\\in \\mathbb{R}^{H\\times r\\times C}(r\\ll W)\\) \\(\\mathcal{A}_i\\in \\mathbb{R}^{r\\times W\\times C}(r\\ll H)\\), and @ denotes the matrix multiplication operation. In this way, the adapted image can be formalized as \\(\\tilde{X}_i^t = X_i^t +\\mathcal{P}_i^{LoRa}\\)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.678, + 0.4, + 0.695 + ], + "angle": 0, + "content": "3.3. Prompt Initialization: Memory Bank" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.704, + 0.473, + 0.904 + ], + "angle": 0, + "content": "To achieve efficient training, it is critical to produce a suitable initialization for \\(\\mathcal{P}_i\\). Therefore, we present a memory-bank-based initialization strategy to benefit the current prompt from previous ones. Specifically, the memory bank \\(M\\) stores \\(S\\) pairs of keys and values \\(\\{q_s,v_s\\}_{s = 1}^S\\) and holds the First In First Out (FIFO) principle. As shown in Figure 2, the keys are the low-frequency components of previous test images, and the values are their corresponding prompts. Note that when using VPTTA-LR, the keys are the low-rank reconstruction matrix of previous test images. To initialize the current prompt \\(\\mathcal{P}_i\\), we first extract the low-frequency component of the current test image \\(\\mathcal{F}_{low}^{A}(X_{i}^{t})\\) and calculate the cosine similarity between \\(\\mathcal{F}_{low}^{A}(X_{i}^{t})\\) and" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.539, + 0.658, + 0.554 + ], + "angle": 0, + "content": "each key \\(q_{s}\\) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.549, + 0.558, + 0.892, + 0.594 + ], + "angle": 0, + "content": "\\[\nC o s \\left(\\mathcal {F} _ {l o w} ^ {A} \\left(X _ {i} ^ {t}\\right), q _ {s}\\right) = \\frac {\\left\\langle \\mathcal {F} _ {l o w} ^ {A} \\left(X _ {i} ^ {t}\\right) , q _ {s} \\right\\rangle}{\\left\\| \\mathcal {F} _ {l o w} ^ {A} \\left(X _ {i} ^ {t}\\right) \\right\\| * \\left\\| q _ {s} \\right\\|}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.598, + 0.892, + 0.738 + ], + "angle": 0, + "content": "where \\(\\langle \\cdot ,\\cdot \\rangle\\) denotes the dot product operation, and \\(\\| \\cdot \\|\\) denotes the Euclidean norm. Since the low-frequency component can represent the style texture of an image, we rank the similarity scores and then retrieve the values of the \\(K\\) most relevant keys from \\(M\\) to construct a support set \\(R_{i} = \\{q_{k},v_{k}\\}_{k = 1}^{K}\\) for the current test image. Finally, we assign a weight \\(w_{k}\\) to \\(v_{k}\\) to initialize the current prompt by \\(\\mathcal{P}_i = \\sum_{k = 1}^K w_k v_k\\) where the weights \\(\\{w_{k}\\}\\) are calculated based on the similarity scores [45] and \\(\\sum_{k = 1}^{K}w_{k} = 1\\)" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.743, + 0.834, + 0.76 + ], + "angle": 0, + "content": "3.4. Prompt Training: Statistics Alignment" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.766, + 0.892, + 0.87 + ], + "angle": 0, + "content": "Inspired by [19, 27, 47], the distribution shift largely exists in distinct BN statistics calculated on the test features (i.e., \\(\\mu_t\\) and \\(\\sigma_t\\)) and stored in the source model \\(f_{\\theta}\\) (i.e., \\(\\mu_s\\) and \\(\\sigma_s\\)), where \\(\\mu\\) and \\(\\sigma\\) respectively denote the mean and standard deviation. To train the prompts to align the statistics, we adopt the absolute-distance-based loss \\(\\mathcal{L}_p\\) [19] as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.578, + 0.874, + 0.892, + 0.906 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {p} = \\frac {1}{j} \\sum \\left| \\mu_ {s} ^ {j} - \\mu_ {t} ^ {j} \\right| + \\left| \\sigma_ {s} ^ {j} - \\sigma_ {t} ^ {j} \\right|, \\tag {3}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "11187" + } + ], + [ + { + "type": "code_caption", + "bbox": [ + 0.088, + 0.096, + 0.308, + 0.11 + ], + "angle": 0, + "content": "Algorithm 1: VPTTA algorithm." + }, + { + "type": "algorithm", + "bbox": [ + 0.099, + 0.113, + 0.416, + 0.154 + ], + "angle": 0, + "content": "Initialize: Frozen source model \\(f_{\\theta}\\), initial prompt \\(\\mathcal{P}\\) filled with ones, memory bank \\(M\\), hyperparameters \\(S\\), \\(K\\) and \\(\\tau\\)." + }, + { + "type": "text", + "bbox": [ + 0.099, + 0.155, + 0.401, + 0.167 + ], + "angle": 0, + "content": "Input: For each time step \\(i\\), current test image \\(X_{i}^{t}\\)." + }, + { + "type": "text", + "bbox": [ + 0.108, + 0.168, + 0.291, + 0.182 + ], + "angle": 0, + "content": "1: ▷ Initialize the prompt \\(\\mathcal{P}_i\\):" + }, + { + "type": "text", + "bbox": [ + 0.107, + 0.183, + 0.21, + 0.195 + ], + "angle": 0, + "content": "2: if \\( i < K \\) then" + }, + { + "type": "text", + "bbox": [ + 0.107, + 0.196, + 0.19, + 0.209 + ], + "angle": 0, + "content": "3: \\(\\mathcal{P}_i = \\mathcal{P}\\)" + }, + { + "type": "text", + "bbox": [ + 0.107, + 0.211, + 0.151, + 0.221 + ], + "angle": 0, + "content": "4: else" + }, + { + "type": "text", + "bbox": [ + 0.107, + 0.223, + 0.361, + 0.236 + ], + "angle": 0, + "content": "5: Construct the support set \\( R_{i} \\) from \\( M \\)" + }, + { + "type": "text", + "bbox": [ + 0.107, + 0.237, + 0.308, + 0.251 + ], + "angle": 0, + "content": "6: \\(\\mathcal{P}_i = \\sum_{k=1}^K w_k v_k, v_k \\in R_i\\)" + }, + { + "type": "text", + "bbox": [ + 0.107, + 0.253, + 0.163, + 0.263 + ], + "angle": 0, + "content": "7: end if" + }, + { + "type": "text", + "bbox": [ + 0.107, + 0.266, + 0.271, + 0.278 + ], + "angle": 0, + "content": "8: \\(\\triangleright\\) Train the prompt \\(P_{i}\\):" + }, + { + "type": "text", + "bbox": [ + 0.107, + 0.279, + 0.307, + 0.293 + ], + "angle": 0, + "content": "9: Obtain \\(\\hat{X}_i^t\\) using \\(\\mathcal{P}_i\\) by Eq. (1)" + }, + { + "type": "text", + "bbox": [ + 0.102, + 0.294, + 0.485, + 0.32 + ], + "angle": 0, + "content": "10: Forward \\( f_{\\theta}(X_i^t) \\) and calculate the warm-up statistics \\( \\mu_w \\) and \\( \\sigma_w \\) by Eq. (4) to replace the \\( \\mu_s \\) and \\( \\sigma_s \\) stored in \\( f_{\\theta} \\)" + }, + { + "type": "text", + "bbox": [ + 0.102, + 0.321, + 0.339, + 0.334 + ], + "angle": 0, + "content": "11: Backward and Update \\(\\mathcal{P}_i\\) by Eq. (5)" + }, + { + "type": "text", + "bbox": [ + 0.102, + 0.336, + 0.201, + 0.346 + ], + "angle": 0, + "content": "12: Inference:" + }, + { + "type": "text", + "bbox": [ + 0.102, + 0.348, + 0.357, + 0.362 + ], + "angle": 0, + "content": "13: Obtain \\(\\hat{X}_i^t\\) using updated \\(\\mathcal{P}_i\\) by Eq. (1)" + }, + { + "type": "text", + "bbox": [ + 0.102, + 0.363, + 0.261, + 0.376 + ], + "angle": 0, + "content": "14: Forward \\(O_{i} = f_{\\theta}(X_{i}^{t})\\)" + }, + { + "type": "text", + "bbox": [ + 0.102, + 0.377, + 0.323, + 0.389 + ], + "angle": 0, + "content": "15: Update the memory bank \\(M\\):" + }, + { + "type": "text", + "bbox": [ + 0.102, + 0.39, + 0.395, + 0.404 + ], + "angle": 0, + "content": "16: M.enqueue \\((q_{i},v_{i})\\) \\(q_{i} = \\mathcal{F}^{A}\\left(X_{i}^{t}\\right),v_{i} = \\mathcal{P}_{i}\\)" + }, + { + "type": "text", + "bbox": [ + 0.102, + 0.405, + 0.25, + 0.417 + ], + "angle": 0, + "content": "17: if \\( \\text{len}(M) > S \\) then" + }, + { + "type": "text", + "bbox": [ + 0.102, + 0.418, + 0.304, + 0.432 + ], + "angle": 0, + "content": "18: M.dequeue \\((q_{\\text{first}}, v_{\\text{first}})\\)" + }, + { + "type": "text", + "bbox": [ + 0.102, + 0.433, + 0.164, + 0.443 + ], + "angle": 0, + "content": "19: end if" + }, + { + "type": "list", + "bbox": [ + 0.102, + 0.168, + 0.485, + 0.443 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.098, + 0.447, + 0.29, + 0.461 + ], + "angle": 0, + "content": "Output:Adapted prediction \\(O_{i}\\)" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.495, + 0.353, + 0.511 + ], + "angle": 0, + "content": "where \\(j\\) indicates the \\(j\\)-th BN layer in \\(f_{\\theta}\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.512, + 0.469, + 0.618 + ], + "angle": 0, + "content": "However, aligning the source and target domains within a limited number of iterations (one in this study) provides challenges due to the empty memory bank and undiminished distribution shifts at the beginning of the inference phase. Therefore, we design a warm-up mechanism to address this issue by simulating the warm-up statistics (i.e., \\(\\mu_w\\) and \\(\\sigma_w\\)) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.1, + 0.632, + 0.469, + 0.649 + ], + "angle": 0, + "content": "\\[\n\\mu_ {w} = \\lambda \\mu_ {t} + (1 - \\lambda) \\mu_ {s}, \\sigma_ {w} = \\lambda \\sigma_ {t} + (1 - \\lambda) \\sigma_ {s}, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.661, + 0.469, + 0.739 + ], + "angle": 0, + "content": "where \\(\\lambda = \\frac{1}{\\sqrt{i} / \\tau + 1}\\), \\(i\\) indicates the index of the current test image and starts from 1, and \\(\\tau\\) is a temperature coefficient to control the rate of transition from the warm-up statistics to the source statistics. Now, the Eq. (3) can be rewritten by:" + }, + { + "type": "equation", + "bbox": [ + 0.151, + 0.741, + 0.469, + 0.772 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {p} = \\frac {1}{j} \\sum \\left| \\mu_ {w} ^ {j} - \\mu_ {t} ^ {j} \\right| + \\left| \\sigma_ {w} ^ {j} - \\sigma_ {t} ^ {j} \\right|, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.781, + 0.47, + 0.901 + ], + "angle": 0, + "content": "and we utilize \\(\\mu_w\\) and \\(\\sigma_w\\) for normalization instead of \\(\\mu_s\\) and \\(\\sigma_s\\). As the inference progresses, the warm-up statistics will gradually shift towards the source statistics. By simulating the warm-up statistics between the source and target statistics, we can significantly reduce the training difficulty at the beginning of the inference phase, resulting in learning prompts from easy to difficult. The overall process of VPTTA is summarized in Algorithm 1." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.09, + 0.633, + 0.108 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.115, + 0.892, + 0.161 + ], + "angle": 0, + "content": "We evaluate our proposed VPTTA on two 2D segmentation benchmark tasks: the joint optic disc (OD) and cup (OC) segmentation task, and the polyp segmentation task." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.168, + 0.787, + 0.183 + ], + "angle": 0, + "content": "4.1. Datasets and Evaluation Metrics" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.191, + 0.892, + 0.341 + ], + "angle": 0, + "content": "The OD/OC segmentation dataset comprises five public datasets collected from different medical centres, denoted as domain A (RIM-ONE-r3 [11]), B (REFuge [32]), C (ORIGA [50]), D (REFuge-Validation/Test [32]), and E (Drishti-GS [38]). There are 159, 400, 650, 800, and 101 images from these datasets. We cropped a region of interest (ROI) centering at OD with size of \\(800 \\times 800\\) for each image following [19], and each ROI is further resized to \\(512 \\times 512\\) and normalized by min-max normalization. The Dice score metric \\((DSC)\\) is utilized for evaluation in this task." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.342, + 0.892, + 0.492 + ], + "angle": 0, + "content": "The polyp segmentation dataset consists of four public datasets collected from different medical centres, denoted as domain A (BKAI-IGH-NeoPolyp [28]), B (CVC-ClinicDB [1]), C (ETIS-LaribPolypDB [36]), and D (Kvasir-Seg [20]), which have 1000, 612, 196, and 1000 images, respectively. We followed [9] to resize each image to \\(352 \\times 352\\) and normalize the resized image by the statistics computed on ImageNet. The \\(DSC\\), enhanced-alignment metric \\((E_{\\phi}^{max})\\) [8], and structural similarity metric \\((S_{\\alpha})\\) [7] are utilized for evaluation in this task." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.5, + 0.716, + 0.516 + ], + "angle": 0, + "content": "4.2. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.524, + 0.892, + 0.901 + ], + "angle": 0, + "content": "For each task, we trained the source model on each single domain (source domain) and tested it on the left domains (target domains) to calculate the mean metrics to evaluate the methods on different scenarios. In the source-training phase, we trained a ResUNet-34 [17] backbone, following [19], as the baseline for OD/OC segmentation task and trained the PraNet [9] with a Res2Net-based [14] backbone for polyp segmentation task. Due to the complexity of PraNet's decoder, we only calculated the prompt loss \\(\\mathcal{L}_p\\) for each BN layer in the encoder instead of the whole network. In the test-adaptation phase, We performed one-iteration adaptation for each batch of test data with a batch size of 1 on all experiments of our VPTTA and other competing methods [45]. As our VPTTA involves an iteration to update and an inference step, distinguishing it from other methods like [41] that do not require a dedicated inference step, we applied the same configuration as VPTTA to other competing methods to ensure consistency in the experimental setup. To deploy our VPTTA, we utilized the Adam optimizer with a learning rate of 0.05 and 0.01 for the OD/OC segmentation task and polyp segmentation task, respectively. The hyperparameters \\(\\alpha\\) (size of prompt), \\(S\\) (size of memory bank), \\(K\\) (size of support set), and \\(\\tau\\) (temperature coefficient in warm-up) are set to 0.01, 40, 16, and 5 for both segmentation tasks." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "11188" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.089, + 0.895, + 0.12 + ], + "angle": 0, + "content": "Table 1. Performance of our VPTTA, 'Source Only' baseline, and six competing methods on the OD/OC segmentation task. The best and second-best results in each column are highlighted in bold and underline, respectively." + }, + { + "type": "table", + "bbox": [ + 0.18, + 0.129, + 0.787, + 0.254 + ], + "angle": 0, + "content": "
MethodsDomain ADomain BDomain CDomain DDomain EAverage
DSCDSCDSCDSCDSCDSC↑
Source Only (ResUNet-34)64.5376.0671.1852.6764.8765.86
TENT-continual (ICLR 2021) [41]73.0778.6671.9446.8170.2068.13
CoTTA (CVPR 2022) [42]75.3975.9869.1453.9970.4068.98
DLTTA (TMI 2022) [45]75.1178.8573.8951.6469.7169.84
DUA (CVPR 2022) [25]72.2876.5970.1356.1771.3869.31
SAR (ICLR 2023) [31]74.5577.7170.7855.4071.7270.03
DomainAdaptor (CVPR 2023) [47]74.5076.3971.8156.7870.5570.01
VPTTA (Ours)73.9179.3674.5156.5175.3571.93
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.267, + 0.895, + 0.298 + ], + "angle": 0, + "content": "Table 2. Performance of our VPTTA, 'Source Only' baseline, and six competing methods on the polyp segmentation task. The best and second-best results in each column are highlighted in bold and underline, respectively." + }, + { + "type": "table", + "bbox": [ + 0.081, + 0.306, + 0.892, + 0.42 + ], + "angle": 0, + "content": "
MethodsDomain ADomain BDomain CDomain DAverage
\\(DSC\\)\\(E_{\\phi}^{max}\\)\\(S_{\\alpha}\\)\\(DSC\\)\\(E_{\\phi}^{max}\\)\\(S_{\\alpha}\\)\\(DSC\\)\\(E_{\\phi}^{max}\\)\\(S_{\\alpha}\\)\\(DSC\\)\\(E_{\\phi}^{max}\\)\\(S_{\\alpha}\\)\\(DSC \\uparrow\\)\\(E_{\\phi}^{max} \\uparrow\\)\\(S_{\\alpha} \\uparrow\\)
Source Only (PraNet)79.9087.9784.6666.3378.5176.7273.8984.6481.2882.9590.8488.0875.7785.4982.69
TENT-continual (ICLR 2021) [41]74.8684.5880.5267.5178.6678.0517.7940.0453.3073.5583.3882.4158.4371.6773.57
CoTTA (CVPR 2022) [42]76.4685.3782.5666.7776.7579.1771.3983.4280.1870.7179.8182.5471.3381.3481.11
DLTTA (TMI 2022) [45]76.2785.2382.4166.5877.0079.2463.7278.2375.5671.2081.3283.4769.4480.4580.17
DUA (CVPR 2022) [25]78.9387.3783.9666.8478.5277.5176.5386.4583.0586.2493.2389.8277.1386.3983.58
SAR (ICLR 2023) [31]76.4885.8981.4966.4577.3578.0571.4683.2379.4070.4180.1181.0771.2081.6580.00
DomainAdaptor (CVPR 2023) [47]77.4886.3182.4070.8281.7680.8871.9683.0679.9776.8985.8984.4574.2984.2681.93
VPTTA (Ours)81.0088.9184.9176.8787.3184.0877.5887.4883.6486.3993.4789.8780.4689.2985.62
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.433, + 0.895, + 0.464 + ], + "angle": 0, + "content": "Table 3. Results of ablation study on the OD/OC segmentation task. The best and second-best results in each column are highlighted in bold and underline, respectively." + }, + { + "type": "table", + "bbox": [ + 0.143, + 0.472, + 0.83, + 0.562 + ], + "angle": 0, + "content": "
MethodsDomain ADomain BDomain CDomain DDomain EAverage
Low-frequency PromptMemory BankWarm-upDSCDSCDSCDSCDSCDSC↑
64.5376.0671.1852.6764.8765.86
✓65.9976.5871.7552.8667.3066.90
✓✓70.7079.3672.8255.5072.8470.24
✓✓74.4177.4774.1454.0367.8569.58
✓✓✓73.9179.3674.5156.5175.3571.93
" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.585, + 0.281, + 0.602 + ], + "angle": 0, + "content": "4.3. Experimental Results" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.611, + 0.471, + 0.822 + ], + "angle": 0, + "content": "We compare our VPTTA with the 'Source Only' baseline (training on the source domain and testing without adaptation), and six competing methods, including a pseudo-label-based method (CoTTA [42]), two entropy-based methods (TENT-continual [41] and SAR [31]), a method dynamically adjusting the learning rate (DLTTA [45]), a method combining entropy-loss and BN statistics fusion (DomainAdaptor [47]), and a BN statistics-based method (DUA [25]) on two segmentation tasks. Specifically, we re-implemented all the competing methods using the same baseline as VPTTA and combined DLTTA with the entropy loss proposed in TENT [45]. The result of each domain is calculated by using the current domain as the source domain and applying the pre-trained model on the left domains." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.472, + 0.903 + ], + "angle": 0, + "content": "Comparison on the OD/OC segmentation task. The results of the proposed VPTTA, the \"Source Only\" baseline, and six competing methods on the OD/OC segmentation task are detailed in Table 1. Obviously, all competing methods outperform the \"Source Only\" baseline, indicating the" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.586, + 0.895, + 0.708 + ], + "angle": 0, + "content": "effectiveness of adaptation towards the distribution shifts. It can be found that our VPTTA achieves the best overall performance across all domains, underscoring its superior applicability and robustness. Meanwhile, our VPTTA also shows competitive performance in domains where most methods exhibit satisfactory performance (such as D) and also performs well in domains where other methods exhibit suboptimal results (such as C)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.897, + 0.903 + ], + "angle": 0, + "content": "Comparison on the polyp segmentation task. We conducted the comparison experiment on the polyp segmentation task under the same experimental setup, as shown in Table 2. In contrast to the results of the OD/OC segmentation task, the methods based on model updating (i.e., TENT-continual, CoTTA, DLTTA, SAR, and DomainAdaptor) exhibit significantly worse performance in this task and are even inferior to the 'Source Only' baseline. It can be attributed to the fact that the polyps are relatively hidden, and the models tend to segment almost nothing rather than poorly segment under the impact of distribution shifts, resulting in confident and low entropy predictions but wrong" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "11189" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.102, + 0.089, + 0.87, + 0.499 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.508, + 0.895, + 0.552 + ], + "angle": 0, + "content": "Figure 3. Visualization of the original images, estimated prompts, and adapted images on the OD/OC segmentation task. We normalize the prompts to [0, 1] for better visualization. The DSC of applying the frozen source model on the original and adapted images is displayed below each image. We also show an example of each source domain on the left side of this diagram. 'Ori': Abbreviation of 'Original'." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.577, + 0.47, + 0.698 + ], + "angle": 0, + "content": "gradients for these methods. Since DUA is backward-free, it still performs well in this task. However, DUA also fails when using domain A as the source domain and has limited performance due to the absence of training for the current test image. Our VPTTA simultaneously benefits from training on the current test image and refraining from updating the model, thus surpassing all competing methods and demonstrating advanced performance." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.701, + 0.47, + 0.883 + ], + "angle": 0, + "content": "Visualization of prompts and adapted images. We visualized the estimated prompts and adapted images on the OD/OC segmentation task in Figure 3 and displayed the DSC values predicted on the original and adapted images below the corresponding images. It can be observed that: (1) Applying prompts to the original test images can achieve significant performance gain; (2) Images from different target domains have similar prompts, proving that involving the memory bank to initialize the current prompt using previous ones is reasonable; (3) The appearances of adapted images are close to the source domain, explicitly demonstrating the reduction of distribution shifts via our VPTTA." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.886, + 0.47, + 0.902 + ], + "angle": 0, + "content": "Comparison in a long-term continual test-time adap" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.577, + 0.892, + 0.895 + ], + "angle": 0, + "content": "tation. To evaluate our VPTTA in a long-term continual test-time adaptation [42], we conducted the experiments for the OD/OC segmentation task multiple rounds and reported the results in Table 4. Since DUA is a backward-free method and DomainAdaptor resets the pre-trained model before each adaptation, they are unsuitable to be evaluated in this experiment. We compared our VPTTA with TENT, CoTTA, DLTTA, and SAR, all of which involve continuous training of the pre-trained model. The performance across different rounds with the same source domain reflects the extent of catastrophic forgetting in the model, and the average performance across multiple rounds highlights the error accumulation of the model within a series of distinct domains. The results reveal that both optimizing loss functions (CoTTA, DLTTA, SAR) and resetting the model (CoTTA, SAR) are effective for mitigating error accumulation and catastrophic forgetting, while TENT with vanilla entropy-minimization loss suffers from severe performance degradation. Additionally, after multiple rounds of testing, our VPTTA maintains superior performance with minimal performance degradation, emphasizing the effectiveness of" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "11190" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.089, + 0.895, + 0.133 + ], + "angle": 0, + "content": "Table 4. Performance of our VPTTA and four competing methods on the OD/OC segmentation task in a long-term continual test-time adaptation. The red numbers indicate the performance gain relative to the 'Source Only' baseline. The performance degradation is calculated between the overall average DSC and the average DSC of round 1. 'Ave': Abbreviation of 'Average'." + }, + { + "type": "table", + "bbox": [ + 0.081, + 0.143, + 0.891, + 0.23 + ], + "angle": 0, + "content": "
Time→
Round123Average DSC ↑Perform. Degra. ↓
MethodsABCDEAveABCDEAveABCDEAve
Source Only64.5376.0671.1852.6764.8765.8664.5376.0671.1852.6764.8765.8664.5376.0671.1852.6764.8765.8665.86
TENT73.0778.6671.9446.8170.2068.1362.0969.3270.6739.0268.2261.8657.0562.4770.2039.0266.3759.0263.01 (-2.85)
CoTTA75.3975.9869.1453.9970.4068.9874.3175.0067.9951.0468.2867.3273.2274.3366.7250.2367.0866.3267.54 (+1.68)
DLTTA75.1178.8573.8951.6469.7169.8474.1479.6574.2545.0569.0468.4372.2878.9372.8742.3769.2667.1468.47 (+2.61)
SAR74.5577.7170.7855.4071.7270.0374.7478.0971.0052.1369.0269.0074.9078.2471.1850.1668.4468.5869.20 (+3.34)
VPTTA (Ours)73.9179.3674.5156.5175.3571.9373.5778.8473.6156.9174.8071.5573.1278.4572.6357.1174.0471.0771.51 (+5.65)
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.256, + 0.434, + 0.271 + ], + "angle": 0, + "content": "training prompts over updating the pre-trained model." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.272, + 0.471, + 0.453 + ], + "angle": 0, + "content": "Ablation study. To evaluate the contributions of the low-frequency prompt, memory-bank-based initialization, and warm-up mechanism, we conducted ablation studies on the OD/OC segmentation task, as shown in Table 3. Note that Eq. (3)/(5) is used as the loss function when training the prompt without/with our warm-up mechanism. The results show that (1) using the low-frequency prompt only has limited performance gain due to the poorly trained prompt; (2) both the memory bank and warm-up mechanism are beneficial to training the prompt; (3) the best performance is achieved when the memory bank and warm-up mechanism are jointly used, which means they are complementary." + }, + { + "type": "image_caption", + "bbox": [ + 0.231, + 0.465, + 0.32, + 0.478 + ], + "angle": 0, + "content": "Average DSC" + }, + { + "type": "image", + "bbox": [ + 0.085, + 0.485, + 0.46, + 0.573 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.582, + 0.471, + 0.611 + ], + "angle": 0, + "content": "Figure 4. Performance of our VPTTA with various \\(\\alpha\\) on the OD/OC segmentation task." + }, + { + "type": "image_caption", + "bbox": [ + 0.231, + 0.637, + 0.317, + 0.651 + ], + "angle": 0, + "content": "Average DSC" + }, + { + "type": "image", + "bbox": [ + 0.084, + 0.651, + 0.46, + 0.743 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.752, + 0.47, + 0.78 + ], + "angle": 0, + "content": "Figure 5. Performance of our VPTTA with various \\( S \\), \\( K \\), and \\( \\tau \\) on the OD/OC segmentation task." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.795, + 0.471, + 0.901 + ], + "angle": 0, + "content": "Analysis of hyperparameter \\(\\alpha\\), \\(S\\), \\(K\\), and \\(\\tau\\). We conducted the experiments for the OD/OC segmentation task to discuss the hyper-parameters used in our VPTTA and displayed the results in Figure 4 and Figure 5. It can be seen that \\(\\alpha = 0.01\\) is optimal, as an excessively small prompt makes adaptation challenging, while an overly large prompt hinders effective training. Similar trends are also reflected" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.254, + 0.892, + 0.281 + ], + "angle": 0, + "content": "Table 5. Performance of our VPTTA and VPTTA-LR on two segmentation tasks." + }, + { + "type": "table", + "bbox": [ + 0.505, + 0.293, + 0.885, + 0.34 + ], + "angle": 0, + "content": "
MethodsOD/OCPolyp
DSC↑Param↓DSC↑Eφmax↑Sα↑Param↓
VPTTA-LR (r=3)69.019,21678.2588.2784.206,336
VPTTA71.937580.4689.9486.8427
" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.364, + 0.892, + 0.469 + ], + "angle": 0, + "content": "in \\( S \\) and \\( K \\): as they increase, the DSC values improve, while the excessively large \\( S \\) and \\( K \\) may introduce unrelated samples, affecting the initialization. With the increase of \\( \\tau \\), performance gradually approaches the upper bound and tends to remain stable, suggesting that the warm-up has reached saturation. Therefore, we finally chose \\( \\alpha = 0.01 \\), \\( S = 40 \\), \\( K = 16 \\), and \\( \\tau = 5 \\) as the best configuration." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.47, + 0.893, + 0.606 + ], + "angle": 0, + "content": "Low-frequency prompt vs. low-rank prompt. To compare the effectiveness of VPTTA and VPTTA-LR, we repeated the experiments on two segmentation tasks and showed results in Table 5. It can be seen that VPTTA significantly outperforms VPTTA-LR, attributed to the fewer learnable parameters and superior representation of style texture in the frequency domain. Benefiting from the low-frequency prompt, VPTTA can be effectively trained with over a hundred times fewer parameters than VPTTA-LR." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.619, + 0.62, + 0.635 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.644, + 0.895, + 0.901 + ], + "angle": 0, + "content": "In this paper, we proposed the VPTTA, a prompt-based method to alleviate the potential risks (i.e., error accumulation, and catastrophic forgetting) by freezing the pre-trained model. Specially, we presented the low-frequency prompt that trains a lightweight prompt with only a small number of parameters in a single iteration to alleviate the distribution shifts in BN layers, constructed a memory bank to better initialize each prompt by utilizing the prompts of previous and the most similar test images, and designed the statistics-fusion-based warm-up mechanism to simulate the warm-up statistics to overcome the training difficulty at the beginning of the inference phase. The extensive experimental results on two medical image segmentation benchmark tasks with multiple domains verified the superiority of our VPTTA over other methods and the effectiveness of each special design. In this regard, we hope that our efforts can offer a different paradigm for continual test-time adaptation." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "11191" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.185 + ], + "angle": 0, + "content": "[1] Jorge Bernal, F Javier Sánchez, Gloria Fernández-Esparrach, Debora Gil, Cristina Rodríguez, and Fernando Vilarino. Wm-dova maps for accurate polyp highlighting in colonoscopy: Validation vs. saliency maps from physicians. Comput. Med. Imaging and Graph., 43:99-111, 2015. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.187, + 0.472, + 0.242 + ], + "angle": 0, + "content": "[2] Paola Cascante-Bonilla, Fuwen Tan, Yanjun Qi, and Vicente Ordonez. Curriculum labeling: Revisiting pseudo-labeling for semi-supervised learning. In AAAI, pages 6912-6920, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.244, + 0.47, + 0.285 + ], + "angle": 0, + "content": "[3] Oscar Chang, Lampros Flokas, and Hod Lipson. Principled weight initialization for hypernetworks. In Int. Conf. Learn. Represent., 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.287, + 0.47, + 0.328 + ], + "angle": 0, + "content": "[4] Dian Chen, Dequan Wang, Trevor Darrell, and Sayna Ebrahimi. Contrastive test-time adaptation. In IEEE Conf. Comput. Vis. Pattern Recog., pages 295-305, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.33, + 0.47, + 0.399 + ], + "angle": 0, + "content": "[5] Lin Chen, Huaian Chen, Zhixiang Wei, Xin Jin, Xiao Tan, Yi Jin, and Enhong Chen. Reusing the task-specific classifier as a discriminator: Discriminator-free adversarial domain adaptation. In IEEE Conf. Comput. Vis. Pattern Recog., pages 7181-7190, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.402, + 0.47, + 0.457 + ], + "angle": 0, + "content": "[6] Mario Döbler, Robert A Marsden, and Bin Yang. Robust mean teacher for continual and gradual test-time adaptation. In IEEE Conf. Comput. Vis. Pattern Recog., pages 7704-7714, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.459, + 0.47, + 0.5 + ], + "angle": 0, + "content": "[7] Deng-Ping Fan, Ming-Ming Cheng, Yun Liu, Tao Li, and Ali Borji. Structure-measure: A new way to evaluate foreground maps. In Int. Conf. Comput. Vis., pages 4548-4557, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.502, + 0.47, + 0.557 + ], + "angle": 0, + "content": "[8] Deng-Ping Fan, Cheng Gong, Yang Cao, Bo Ren, Ming-Ming Cheng, and Ali Borji. Enhanced-alignment measure for binary foreground map evaluation. In IJCAI, pages 698-704, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.56, + 0.47, + 0.628 + ], + "angle": 0, + "content": "[9] Deng-Ping Fan, Ge-Peng Ji, Tao Zhou, Geng Chen, Huazhu Fu, Jianbing Shen, and Ling Shao. Pranet: Parallel reverse attention network for polyp segmentation. In Int. Conf. Med. Image Comput. Comput.-Assist. Intervention., pages 263-273. Springer, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.631, + 0.47, + 0.671 + ], + "angle": 0, + "content": "[10] Matteo Frigo and Steven G Johnson. FFTW: An adaptive software architecture for the fft. In ICASSP, pages 1381-1384. IEEE, 1998. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.674, + 0.47, + 0.741 + ], + "angle": 0, + "content": "[11] Francisco Fumero, Silvia Alayón, José L Sanchez, Jose Sigut, and M Gonzalez-Hernandez. RIM-ONE: An open retinal image database for optic nerve evaluation. In Int. Symp. Comput.-based Med. Syst., pages 1-6. IEEE, 2011. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.745, + 0.47, + 0.8 + ], + "angle": 0, + "content": "[12] Yulu Gan, Yan Bai, Yihang Lou, Xianzheng Ma, Renrui Zhang, Nian Shi, and Lin Luo. Decorate the newcomers: Visual domain prompt for continual test time adaptation. In AAAI, pages 7595-7603, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.802, + 0.47, + 0.843 + ], + "angle": 0, + "content": "[13] Yossi Gandelsman, Yu Sun, Xinlei Chen, and Alexei Efros. Test-time training with masked autoencoders. Adv. Neural Inform. Process. Syst., 35:29374-29385, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.846, + 0.47, + 0.9 + ], + "angle": 0, + "content": "[14] Shang-Hua Gao, Ming-Ming Cheng, Kai Zhao, Xin-Yu Zhang, Ming-Hsuan Yang, and Philip Torr. Res2net: A new multi-scale backbone architecture. IEEE Trans. Pattern Anal. Mach. Intell., 43(2):652-662, 2019. 5" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.472, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.148 + ], + "angle": 0, + "content": "[15] Yunhe Gao, Xingjian Shi, Yi Zhu, Hao Wang, Zhiqiang Tang, Xiong Zhou, Mu Li, and Dimitris N Metaxas. Visual prompt tuning for test-time domain adaptation. arXiv preprint arXiv:2210.04831, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.151, + 0.892, + 0.246 + ], + "angle": 0, + "content": "[16] Mohsen Ghafoorian, Alireza Mehrtash, Tina Kapur, Nico Karssemeijer, Elena Marchiori, Mehran Pesteie, Charles RG Guttmann, Frank-Erik de Leeuw, Clare M Tempany, Bram Van Ginneken, et al. Transfer learning for domain adaptation in mri: Application in brain lesion segmentation. In Int. Conf. Med. Image Comput. Comput.-Assist. Intervent., pages 516-524. Springer, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.249, + 0.892, + 0.289 + ], + "angle": 0, + "content": "[17] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In IEEE Conf. Comput. Vis. Pattern Recog., pages 770-778, 2016. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.291, + 0.892, + 0.346 + ], + "angle": 0, + "content": "[18] Edward J Hu, Phillip Wallis, Zeyuan Allen-Zhu, Yuzhhi Li, Shean Wang, Lu Wang, Weizhu Chen, et al. Lora: Low-rank adaptation of large language models. In Int. Conf. Learn. Represent., 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.349, + 0.892, + 0.403 + ], + "angle": 0, + "content": "[19] Shishuai Hu, Zehui Liao, and Yong Xia. Prosfda: Prompt learning based source-free domain adaptation for medical image segmentation. arXiv preprint arXiv:2211.11514, 2022. 2, 3, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.406, + 0.892, + 0.473 + ], + "angle": 0, + "content": "[20] Debesh Jha, Pia H Smedsrud, Michael A Riegler, Pål Halvorsen, Thomas de Lange, Dag Johansen, and Håvard D Johansen. Kvasir-seg: A segmented polyp dataset. In Int. Conf. Multimedia Modeling, pages 451-462. Springer, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.476, + 0.892, + 0.531 + ], + "angle": 0, + "content": "[21] Christian Lamers, René Vidal, Nabil Belbachir, Niki van Stein, Thomas Bäeck, and Paris Giampouras. Clustering-based domain-incremental learning. In Int. Conf. Comput. Vis., pages 3384-3392, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.533, + 0.892, + 0.574 + ], + "angle": 0, + "content": "[22] Joonseok Lee, Seungyeon Kim, Guy Lebanon, and Yoram Singer. Local low-rank matrix approximation. In Int. Conf. Mach. Learn., pages 82-90. PMLR, 2013. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.577, + 0.892, + 0.617 + ], + "angle": 0, + "content": "[23] Jian Liang, Ran He, Zhenan Sun, and Tieniu Tan. Exploring uncertainty in pseudo-label guided unsupervised domain adaptation. Pattern Recognition, 96:106996, 2019. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.62, + 0.892, + 0.687 + ], + "angle": 0, + "content": "[24] Pengfei Liu, Weizhe Yuan, Jinlan Fu, Zhengbao Jiang, Hiroaki Hayashi, and Graham Neubig. Pre-train, prompt, and predict: A systematic survey of prompting methods in natural language processing. ACM Comput. Surveys, 55(9):1-35, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.69, + 0.892, + 0.757 + ], + "angle": 0, + "content": "[25] M Jehanzeb Mirza, Jakub Micorek, Horst Possegger, and Horst Bischof. The norm must go on: Dynamic unsupervised domain adaptation by normalization. In IEEE Conf. Comput. Vis. Pattern Recog., pages 14765-14775, 2022. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.761, + 0.892, + 0.829 + ], + "angle": 0, + "content": "[26] Chaithanya Kumar Mummadi, Robin Hutmacher, Kilian Rambach, Evgeny Levinkov, Thomas Brox, and Jan Hendrik Metzen. Test-time adaptation to distribution shift by confidence maximization and input transformation. arXiv preprint arXiv:2106.14999, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.832, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[27] Zachary Nado, Shreyas Padhy, D Sculley, Alexander D'Amour, Balaji Lakshminarayanan, and Jasper Snoek. Evaluating prediction-time batch normalization for robustness under covariate shift. arXiv preprint arXiv:2006.10963, 2020. 3, 4" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "11192" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.161 + ], + "angle": 0, + "content": "[28] Phan Ngoc Lan, Nguyen Sy An, Dao Viet Hang, Dao Van Long, Tran Quang Trung, Nguyen Thi Thuy, and Dinh Viet Sang. Neounet: Towards accurate colon polyp segmentation and neoplasm detection. In Adv. Vis. Comput. Int. Symp., pages 15-28. Springer, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.162, + 0.47, + 0.218 + ], + "angle": 0, + "content": "[29] A Tuan Nguyen, Thanh Nguyen-Tang, Ser-Nam Lim, and Philip HS Torr. Tipi: Test time adaptation with transformation invariance. In IEEE Conf. Comput. Vis. Pattern Recog., pages 24162-24171, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.218, + 0.469, + 0.273 + ], + "angle": 0, + "content": "[30] Shuaicheng Niu, Jiaxiang Wu, Yifan Zhang, Yafo Chen, Shijian Zheng, Peilin Zhao, and Mingkui Tan. Efficient test-time model adaptation without forgetting. In Int. Conf. Mach. Learn., pages 16888-16905. PMLR, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.274, + 0.469, + 0.328 + ], + "angle": 0, + "content": "[31] Shuaicheng Niu, Jiaxiang Wu, Yifan Zhang, Zhiquan Wen, Yaofo Chen, Peilin Zhao, and Mingkui Tan. Towards stable test-time adaptation in dynamic wild world. In Int. Conf. Learn. Represent., 2023. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.33, + 0.469, + 0.412 + ], + "angle": 0, + "content": "[32] José Ignacio Orlando, Huazhu Fu, João Barbosa Breda, Karel van Keer, Deepti R Bathula, Andrés Diaz-Pinto, Ruogu Fang, Pheng-Ann Heng, Jyoung Kim, JoonHo Lee, et al. REFuge Challenge: A unified framework for evaluating automated methods for glaucoma assessment from fundus photographs. Med. Image Anal., 59:101570, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.413, + 0.469, + 0.467 + ], + "angle": 0, + "content": "[33] Xingang Pan, Ping Luo, Jianping Shi, and Xiaou Tang. Two at once: Enhancing learning and generalization capacities via ibn-net. In *Eur. Conf. Comput. Vis.*, pages 464-479, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.469, + 0.469, + 0.537 + ], + "angle": 0, + "content": "[34] Swami Sankaranarayanan, Yogesh Balaji, Arpit Jain, Ser Nam Lim, and Rama Chellappa. Learning from synthetic data: Addressing domain shift for semantic segmentation. In IEEE Conf. Comput. Vis. Pattern Recog., pages 3752-3761, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.538, + 0.469, + 0.607 + ], + "angle": 0, + "content": "[35] Steffen Schneider, Evgenia Rusak, Luisa Eck, Oliver Bringmann, Wieland Brendel, and Matthias Bethge. Improving robustness against common corruptions by covariate shift adaptation. Adv. Neural Inform. Process. Syst., 33:11539-11551, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.608, + 0.469, + 0.676 + ], + "angle": 0, + "content": "[36] Juan Silva, Aymeric Histace, Olivier Romain, Xavier Dray, and Bertrand Granado. Toward embedded detection of polyps in wce images for early diagnosis of colorectal cancer. Int. J. Comput.-Assist. Rad. Sur., 9(2):283-293, 2014. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.678, + 0.469, + 0.733 + ], + "angle": 0, + "content": "[37] Samarth Sinha, Peter Gehler, Francesco Locatello, and Bernt Schiele. Test: Test-time self-training under distribution shift. In IEEE Wint. Conf. Appl. Comput. Vis., pages 2759-2769, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.734, + 0.469, + 0.802 + ], + "angle": 0, + "content": "[38] Jayanthi Sivaswamy, SR Krishnadas, Gopal Datt Joshi, Madhulika Jain, and A Ujjwaft Syed Tabish. Drishti-GS: Retinal image dataset for optic nerve head (onh) segmentation. In IEEE Int. Symp. on Bio. Imaging, pages 53-56. IEEE, 2014. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.803, + 0.469, + 0.859 + ], + "angle": 0, + "content": "[39] Junha Song, Jungsoo Lee, In So Kweon, and Sungha Choi. Ecotta: Memory-efficient continual test-time adaptation via self-distilled regularization. In IEEE Conf. Comput. Vis. Pattern Recog., pages 11920-11929, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.859, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[40] Yu Sun, Xiaolong Wang, Zhuang Liu, John Miller, Alexei Efros, and Moritz Hardt. Test-time training with self-supervision for generalization under distribution shifts. In" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.892, + 0.119 + ], + "angle": 0, + "content": "Int. Conf. Mach. Learn., pages 9229-9248. PMLR, 2020. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.121, + 0.892, + 0.177 + ], + "angle": 0, + "content": "[41] Dequan Wang, Evan Shelhamer, Shaoteng Liu, Bruno Olshausen, and Trevor Darrell. Tent: Fully test-time adaptation by entropy minimization. In Int. Conf. Learn. Represent., 2021. 2, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.179, + 0.892, + 0.22 + ], + "angle": 0, + "content": "[42] Qin Wang, Olga Fink, Luc Van Gool, and Dengxin Dai. Continual test-time domain adaptation. In IEEE Conf. Comput. Vis. Pattern Recog., pages 7201-7211, 2022. 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.221, + 0.892, + 0.291 + ], + "angle": 0, + "content": "[43] Shujun Wang, Lequan Yu, Kang Li, Xin Yang, Chi-Wing Fu, and Pheng-Ann Heng. Boundary and entropy-driven adversarial learning for fundus image segmentation. In Int. Conf. Med. Image Comput. Comput.-Assist. Intervention., pages 102-110. Springer, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.291, + 0.892, + 0.373 + ], + "angle": 0, + "content": "[44] Wei Wang, Zhun Zhong, Weijie Wang, Xi Chen, Charles Ling, Boyu Wang, and Nicu Sebe. Dynamically instance-guided adaptation: A backward-free approach for test-time domain adaptive semantic segmentation. In IEEE Conf. Comput. Vis. Pattern Recog., pages 24090-24099, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.375, + 0.892, + 0.445 + ], + "angle": 0, + "content": "[45] Hongzheng Yang, Cheng Chen, Meirui Jiang, Quande Liu, Jianfeng Cao, Pheng Ann Heng, and Qi Dou. Dltta: Dynamic learning rate for test-time adaptation on cross-domain medical images. IEEE Trans. Med. Imaging, 41(12):3575-3586, 2022. 2, 3, 4, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.446, + 0.892, + 0.489 + ], + "angle": 0, + "content": "[46] Yanchao Yang and Stefano Soatto. FDA: Fourier domain adaptation for semantic segmentation. In IEEE Conf. Comput. Vis. Pattern Recog., pages 4085-4095, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.489, + 0.892, + 0.544 + ], + "angle": 0, + "content": "[47] Jian Zhang, Lei Qi, Yinghuan Shi, and Yang Gao. Domain adaptor: A novel approach to test-time adaptation. In IEEE Conf. Comput. Vis. Pattern Recog., pages 18971-18981, 2023. 2, 3, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.545, + 0.892, + 0.587 + ], + "angle": 0, + "content": "[48] Marvin Zhang, Sergey Levine, and Chelsea Finn. Memo: Test time robustness via adaptation and augmentation. Adv. Neural Inform. Process. Syst., 35:38629-38642, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.588, + 0.892, + 0.657 + ], + "angle": 0, + "content": "[49] Yizhe Zhang, Tao Zhou, Yuhui Tao, Shuo Wang, Ye Wu, Benyuan Liu, Pengfei Gu, Qiang Chen, and Danny Z Chen. TestFit: A plug-and-play one-pass test time method for medical image segmentation. Medical Image Analysis, 92: 103069, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.659, + 0.892, + 0.74 + ], + "angle": 0, + "content": "[50] Zhuo Zhang, Feng Shou Yin, Jiang Liu, Wing Kee Wong, Ngan Meng Tan, Beng Hai Lee, Jun Cheng, and Tien Yin Wong. ORIGA-light: An online retinal fundus image database for glaucoma analysis and research. In Int. Conf. IEEE Eng. in Med. and Bio., pages 3065-3068. IEEE, 2010. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.742, + 0.892, + 0.798 + ], + "angle": 0, + "content": "[51] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Zwei Liu. Conditional prompt learning for vision-language models. In IEEE Conf. Comput. Vis. Pattern Recog., pages 16816-16825, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.799, + 0.892, + 0.855 + ], + "angle": 0, + "content": "[52] Chen Zhu, Renkun Ni, Zheng Xu, Kezhi Kong, W Ronny Huang, and Tom Goldstein. Gradinit: Learning to initialize neural networks for stable and efficient training. Adv. Neural Inform. Process. Syst., 34:16410-16422, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.856, + 0.892, + 0.897 + ], + "angle": 0, + "content": "[53] Yukun Zuo, Hantao Yao, and Changsheng Xu. Attention-based multi-source domain adaptation. IEEE Trans. Image Process., 30:3793-3803, 2021. 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.897 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "11193" + } + ] +] \ No newline at end of file diff --git a/2024/Each Test Image Deserves A Specific Prompt_ Continual Test-Time Adaptation for 2D Medical Image Segmentation/b547c627-f0bf-4b29-9f5d-10851e261fb1_origin.pdf b/2024/Each Test Image Deserves A Specific Prompt_ Continual Test-Time Adaptation for 2D Medical Image Segmentation/b547c627-f0bf-4b29-9f5d-10851e261fb1_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..a26e30aa0a2f8235778d9fba9ed7dab16d8d7c71 --- /dev/null +++ b/2024/Each Test Image Deserves A Specific Prompt_ Continual Test-Time Adaptation for 2D Medical Image Segmentation/b547c627-f0bf-4b29-9f5d-10851e261fb1_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ff2f348265e221db5f7118851aa550fa57bd57dad3fae815452ccf527a40b545 +size 2303818 diff --git a/2024/Each Test Image Deserves A Specific Prompt_ Continual Test-Time Adaptation for 2D Medical Image Segmentation/full.md b/2024/Each Test Image Deserves A Specific Prompt_ Continual Test-Time Adaptation for 2D Medical Image Segmentation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..f46763e15f7c7e8443b1f948c2fd60d37fe9f643 --- /dev/null +++ b/2024/Each Test Image Deserves A Specific Prompt_ Continual Test-Time Adaptation for 2D Medical Image Segmentation/full.md @@ -0,0 +1,330 @@ +# Each Test Image Deserves A Specific Prompt: Continual Test-Time Adaptation for 2D Medical Image Segmentation + +Ziyang Chen $^{1}$ Yongsheng Pan $^{1\dagger}$ Yiwen Ye $^{1}$ Mengkang Lu $^{1}$ Yong Xia $^{1,2,3\dagger}$ + +$^{1}$ School of Computer Science and Engineering, Northwestern Polytechnical University, China + $^{2}$ Research & Development Institute of Northwestern Polytechnical University in Shenzhen, China + $^{3}$ Ningbo Institute of Northwestern Polytechnical University, China + +zychen@mail.nwpu.edu.cn, yspan@nwpu.edu.cn, ywye@mail.nwpu.edu.cn + +lmk@mail.nwpu.edu.cn, yxia@nwpu.edu.cn + +![](images/b7f9bd6d88491770ebe0d90b0c346003251b1c61ebacdc964fdb0880fc4578ee.jpg) +Source Model $f_{\theta}$ + +![](images/b8b46250f831afc9abdf05c7c6e2827ddb1168403b421ac04fd66a94dc21150c.jpg) + +![](images/cd04bb644b92668adb948c54e6dd1e02cd01489ef8ab0b93524ac05af68b8489.jpg) +Domain 2 +Domain 3 + +![](images/fc535331f710fa7ac4fce2eee57cf00ca969e955b8339f281b41741b0244954e.jpg) +Domain 4 +$f_{\theta}^{\prime}\rightarrow f_{\theta}$ + +![](images/7319de94f101abffc1db56458e03dba8862b88e6626de9b750b8ffadb1077ef0.jpg) + +![](images/e494199ddcc9269b92378aae9e5e5dd9177811b8a547070f8b40a5a81cf05d25.jpg) +Domain N + +![](images/7e3ad50669bb3bfaf2ad32d3263adc01c67d8f7dc067ec88311807d661551c85.jpg) + +$\mathcal{L}_{self} \to \mathcal{L}_{prompt}$ + +![](images/a05f686c0ccc852d2a5dab91a906236fcdcad7d871b369ef903edea534b0d07e.jpg) + +![](images/fed712e153f9e5774cd0cc8f4a24a59cab2db2c8c82ea713d3e05bc69d46a2de.jpg) +Test + +![](images/f72b2d431c2fb13b0ab4e39ad1af38019d53f964edf9309d24df6ec683889042.jpg) +Adapted Output + +![](images/88efd7ca147b209c872a8aa52987eaa8196794d516936ba32fc8fc5f217a3219.jpg) +(c) Our VPTTA +EA: +CF: + +![](images/eac8d5775df914bcf760ff26ff202347d72471ca16344356ea573d4ddd730952.jpg) +Freeze + +![](images/d0488d37486fb777069da662dc3fdd126d023911f2b61598462fba29c36f68af.jpg) +Fine-tune + +Freeze + +![](images/c2450a86b24cea5720114978d05f0a713b50915b93014079c783837ff7f8bf24.jpg) +Prompt + +![](images/40332400e9aa9af88781e712236ab1c13d8f18a18d90c69267432e81c2cd9168.jpg) +Terrible + +![](images/5da86f88f4bdf7572c8ca547f9c4b4f6f98caf315180fda582195c5b5a87c355.jpg) +Passable + +![](images/7d936a17d3505ddcbca639d400faa5ff0772bef947668a13d6d47f9b19f07ace.jpg) +Good +Figure 1. Comparison between our VPTTA and existing solutions under the CTTA setup. Our VPTTA avoids both error accumulation (EA) and catastrophic forgetting (CF) by freezing model parameters and achieves adaptation by training a prompt for each test image. The commonly used self-supervised loss and the optimized one are denoted by $\mathcal{L}_{self}$ and $\mathcal{L}_{self}^{\prime}$ , respectively. + +# Abstract + +Distribution shift widely exists in medical images acquired from different medical centres and poses a significant obstacle to deploying the pre-trained semantic segmentation model in real-world applications. Test-time adaptation has proven its effectiveness in tackling the cross-domain distribution shift during inference. However, most existing methods achieve adaptation by updating the pre-trained models, rendering them susceptible to error accumulation and catastrophic forgetting when encountering a series of distribution shifts (i.e., under the continual test-time adaptation setup). To overcome these challenges caused by updating the models, in this paper, we freeze the pre-trained model and propose the Visual Prompt-based Test-Time Adaptation (VPTTA) method to train a specific prompt for each test image to align the statistics in the batch normalization layers. Specifically, we present the low-frequency prompt, which is lightweight with only a few parameters + +and can be effectively trained in a single iteration. To enhance prompt initialization, we equip VPTTA with a memory bank to benefit the current prompt from previous ones. Additionally, we design a warm-up mechanism, which mixes source and target statistics to construct warm-up statistics, thereby facilitating the training process. Extensive experiments demonstrate the superiority of our VPTTA over other state-of-the-art methods on two medical image segmentation benchmark tasks. The code and weights of pre-trained source models are available at https://github.com/Chen-Ziyang/VPTTA. + +# 1. Introduction + +Semantic segmentation serves a pivotal role in medical image processing to outline specific anatomical structures. However, the segmentation model pre-trained on the source dataset may be hindered from deploying in real-world applications due to the commonly existing distribution shift, which is typically caused by the variations in imaging protocols, operators, and scanners [16, 34]. Test-time adaptation (TTA) has emerged as a promising paradigm to replace domain adaptation (DA) [5, 43, 46, 53], as TTA requires only test data during the inference phase, while DA necessitates a sufficiently large and representative target dataset [45]. The mainstream TTA methods [29-31, 37, 40, 41, 47, 48] leverage the self-supervised tasks to construct auxiliary losses, such as entropy minimization [41] and self-training [37], to update the parameters of pre-trained models. Although these TTA methods successfully alleviate the domain gap and performance degradation, they are based on the assumption that the target domains exhibit static distributions. Unfortunately, the target domains are continuously changing rather than static in most real-world scenarios. + +This was first presented as the continual test-time adaptation (CTTA) setup [42] that addresses TTA on a series of distribution shifts. Under this setup, pre-trained models are required to adapt to several distinct domains and therefore are more prone to suffer error accumulation and catastrophic forgetting. As mentioned above, many TTA methods rely on self-supervised tasks and may be affected by the noisy losses due to unreliable supervision, leading to error accumulation [2, 23]. Catastrophic forgetting typically occurs when the model is continually trained on new domains, resulting in unexpected performance degradation [21, 30]. Some research has been studied in addressing these two challenges based on optimizing the loss functions for better supervision [6, 31, 42, 42, 45] (see Figure 1 (a)) or prevent the model parameters from excessive alterations by model resetting or introducing regularizers [30, 31, 39, 42] (see Figure 1 (b)). Despite their efforts, loss optimization cannot ensure a robust training process to avoid catastrophic forgetting during the long-term inference, insufficient or excessive updating of the model parameters may lead to inadequate adaptation or error accumulation. Since the pretrained model is updated by the self-supervised loss where the noise cannot be completely removed, these methods may still suffer from these two issues. + +Considering the above limitations in CTTA, updating the pre-trained model during inference seems to be inappropriate. To this end, we propose the Visual Prompt-based Test-Time Adaptation (VPTTA) method to avoid error accumulation and catastrophic forgetting associated with updating model by freezing the model parameters and conduct adaptation by learning a specific visual prompt for each test image (see Figure 1 (c)). The prompt can adapt each test im + +age to the frozen pre-trained model to take into account both adaptation ability and knowledge retaining. We attempt to achieve this by answering the following three questions. + +- How to design the prompt? Each prompt is restricted to a limited number of training iterations (sometimes just one) to satisfy the demands of online inference, requiring the prompt to be lightweight. Based on this principle, we introduce the prompt for only the low-frequency components of each sample, resulting in the low-frequency prompt. Since the low-frequency components are strongly associated with style textures and can serve as the primary source of distribution shift [46], our low-frequency prompt allows for maintaining lightweight and being able to reduce distribution shifts by modifying the low-frequency components. + +- How to initialize the prompt? It is well known that model training can benefit from proper parameter initialization, resulting in not only fewer training iterations but also improved performance [3, 52]. We argue that this is also applicable to prompt training and therefore construct a memory bank to store the prompts and low-frequency components of previous test images. By comparing the similarity between the low-frequency components of the current and previous images, we select several historical prompts of the most similar images to initialize the current prompt. + +- How to train the prompt? Since the statistics mismatch between the source and target domains in the batch normalization (BN) layers is a significant cause of distribution shift [33, 35, 44], we optimize the prompt with the objective of minimizing the distances between BN statistics and the statistics extracted from the features of test data for statistics alignment [19]. To further address the training difficulty posed by the large distribution shift at the beginning of the inference phase, we devise a statistics-fusion-based warm-up mechanism to assist in learning the prompt from easy to hard. + +To summarise, our contributions are three-fold: (1) We present the low-frequency prompt to adjust the test image to adapt the model instead of updating the model and therefore prevent the model from error accumulation and catastrophic forgetting. (2) We propose the novel VPTTA framework to alleviate the distribution shift by training a sample-specific visual prompt, which is lightweight, well-initialized, and follows a gradual learning process. (3) The powerful framework we provide can be effectively deployed, requiring only a single iteration and a few dozen parameters for adaptation. + +# 2. Related Work + +# 2.1. Test-time Adaptation (TTA) + +TTA aims to adapt the model pre-trained on the source domain to the test data during inference in a source-free + +and online manner [40, 41]. The mainstream TTA methods are based on model updating, which constructs self-supervised auxiliary tasks to guide the model to learn on the test data [4, 13, 26, 29, 37, 40, 41, 48, 49]. Sun et al. [40] utilized an auxiliary task to predict the rotation to assist the model learning on the test data. Wang et al. [41] presented a test-time entropy minimization scheme to reduce generalization error by reducing the entropy of model predictions on test data. Zhang et al. [49] introduced a supplementary network to fit the test data, which can work in a plug-and-play fashion, necessitating minimal hyperparameter tuning. Some other methods are based on BN statistics, which alleviate the distribution shifts by modifying the statistics within BN layers [25, 27, 44]. Mirza et al. [25] constantly adapted the statistics of the BN layers to modify the feature representations of the model. These TTA methods improve the adaptation performance on the static target domain but neglect the continually changing target domains in most real-world scenarios. + +Recently, the continual test-time adaptation (CTTA) [42] is proposed to tackle TTA under a series of continually changing target domains. Due to long-term inference, both error accumulation [2, 23] and catastrophic forgetting [21, 30] are more prone to occur in this setup. Many methods [6, 30, 31, 39, 42, 45, 47] optimize the self-supervised losses or prevent excessive alterations to the model parameters via model resetting or regularizer to improve this situation but are still impacted by these issues. Wang et al. [42] refined pseudo labels by the weight- and augmentation-averaged predictions and randomly restored a small part of model parameters. Niu et al. [31] removed partial noisy samples with large gradients and encouraged model weights toward a flat minimum. Yang et al. [45] designed a dynamic strategy to adjust the learning rate for each test image. Zhang et al. [47] presented to adaptively fuse source and test statistics and improve the entropy minimization loss to a generalized one. In contrast, our proposed VPTTA employs a learnable prompt for each test image to mitigate distribution shifts during inference, requiring no changes to the pre-trained model and therefore alleviating the error accumulation and catastrophic forgetting. + +# 2.2. Prompt Learning + +Prompt learning was initially employed to devise additional text instructions for input text to fine-tune the largescale Natural Language Processing (NLP) models on downstream tasks [24]. With remarkable success in NLP, it has also been involved in computer vision and greatly benefited visual models in recent years [12, 15, 19, 51]. Jia et al. [15] introduced a small number of trainable parameters in the input space while keeping the model backbone frozen to fine-tune the model with high quality and efficiency. Zhou et al. [51] presented the conditional context optimization + +to generate an input-conditional token for each image to generalize the pre-trained vision-language model to unseen classes. Hu et al. [19] employed the visual prompt at the pixel level and added it to the input images to explicitly minimize the domain discrepancy. Gan et al. [12] learned a domain-specific and domain-agnostic visual prompt for adaptation by using the exponential-moving-average strategy under the guidance of self-training. Different from them, our VPTTA presents the low-frequency prompt to address a series of changing distribution shifts with a few dozen parameters. We further consider optimizing the initialization and warm-up mechanism, enabling our visual prompts to be effectively trained within a single iteration to satisfy the demands of CTTA. + +# 3. Methodology + +# 3.1. Problem Definition and Method Overview + +Let the labeled source domain dataset and unlabeled testing target dataset be $\mathbb{D}^s = \{X_i^s,Y_i^s\}_{i = 1}^{N^s}$ and $\mathbb{D}^t = \{X_i^t\}_{i = 1}^{N^t}$ , respectively, where $X_{i}^{*}\in \mathbb{R}^{H\times W\times C}$ is the $i$ -th image with C channels and size of $H\times W$ , and $Y_{i}^{*}$ is its label. Our goal is to learn a low-frequency prompt $\mathcal{P}_i$ for each test image $X_{i}^{t}$ to obtain the modified image $\tilde{X}_i^t$ that can be adapted to the model $f_{\theta}:X\to Y$ pre-trained on $\mathbb{D}^s$ . + +To accomplish this goal, our VPTTA considers three key facets: prompt design, prompt initialization, and prompt training. For prompt design, the prompt is designed to change the low-frequency component of the amplitude of the test image. For prompt initialization, we adopt a memory bank to initialize the current prompt with previous ones. For prompt training, we employ the absolute distance to train the prompt to align the statistics between features extracted from test images and the batch normalization (BN) layers and also present a statistics-fusion-based warm-up mechanism. An overview of our VPTTA is illustrated in Figure 2. We now delve into its details. + +# 3.2. Prompt Design: Low-frequency Prompt + +Let $\mathcal{F}(\cdot)$ and $\mathcal{F}^{-1}(\cdot)$ be the Fast Fourier Transform (FFT) and Inverse Fast Fourier Transform (IFFT) operations [10], respectively. The amplitude component and phase component are denoted by $\mathcal{F}^A (\cdot)$ and $\mathcal{F}^P (\cdot)$ respectively. Note that the low-frequency component in $\mathcal{F}^A (\cdot)$ is shifted to the center. Utilizing the prompt $\mathcal{P}_i\in \mathbb{R}^{(\alpha \times H)\times (\alpha \times W)\times C}$ to obtain the adapted image $\bar{X}_i^t$ can be formalized as + +$$ +\tilde {X} _ {i} ^ {t} = \mathcal {F} ^ {- 1} ([ \text {O n e P a d} (\mathcal {P} _ {i}) \odot \mathcal {F} ^ {A} (X _ {i} ^ {t}), \mathcal {F} ^ {P} (X _ {i} ^ {t}) ]) \tag {1} +$$ + +where $\odot$ is the element-wise multiplication, and OnePad means padding one around $\mathcal{P}_i$ to expand it to the size of $H\times W$ . Under the control of the small coefficient $\alpha \in (0,1)$ , we can ensure that the prompt $\mathcal{P}_i$ remains lightweight and focuses on the low-frequency component. + +![](images/62cb4d2a81e70ceeeda78d942341260a70bc8d077afc8e161d3cbac4508c6ad2.jpg) +Figure 2. Overview of our VPTTA. For each test image, (1) the Fast Fourier Transform (FFT) is first applied to transform it into the frequency domain, where the low-frequency component of amplitude is used to query in a memory bank to initialize the current prompt, and then the amplitude is multiplied with the prompt at the low-frequency component and transformed back to the spatial domain using the Inverse Fast Fourier Transform (IFFT). (2) The memory bank is built on the previous low-frequency components and their corresponding prompts and is updated using the First In First Out (FIFO) strategy. (3) We convert the source statistics stored in BN layers into the warm-up statistics and calculate the absolute distance between warm-up and target statistics as the loss to fine-tune the prompt. (4) Finally, we feed the image corrected by its fine-tuned prompt to the pre-trained model to produce the output. 'Stat': Abbreviation of 'Statistics'. + +Simultaneously, we recognize that the low-rank decomposition is widely employed in the field of data compression [22] and model fine-tuning [18]. Consequently, we also design a low-rank prompt $\mathcal{P}_i^{LoRa} = \mathcal{B}_i@A_i$ as a variant (denoted as VPTTA-LR), where $\mathcal{B}_i\in \mathbb{R}^{H\times r\times C}(r\ll W)$ $\mathcal{A}_i\in \mathbb{R}^{r\times W\times C}(r\ll H)$ , and @ denotes the matrix multiplication operation. In this way, the adapted image can be formalized as $\tilde{X}_i^t = X_i^t +\mathcal{P}_i^{LoRa}$ . + +# 3.3. Prompt Initialization: Memory Bank + +To achieve efficient training, it is critical to produce a suitable initialization for $\mathcal{P}_i$ . Therefore, we present a memory-bank-based initialization strategy to benefit the current prompt from previous ones. Specifically, the memory bank $M$ stores $S$ pairs of keys and values $\{q_s,v_s\}_{s = 1}^S$ and holds the First In First Out (FIFO) principle. As shown in Figure 2, the keys are the low-frequency components of previous test images, and the values are their corresponding prompts. Note that when using VPTTA-LR, the keys are the low-rank reconstruction matrix of previous test images. To initialize the current prompt $\mathcal{P}_i$ , we first extract the low-frequency component of the current test image $\mathcal{F}_{low}^{A}(X_{i}^{t})$ and calculate the cosine similarity between $\mathcal{F}_{low}^{A}(X_{i}^{t})$ and + +each key $q_{s}$ as follows: + +$$ +C o s \left(\mathcal {F} _ {l o w} ^ {A} \left(X _ {i} ^ {t}\right), q _ {s}\right) = \frac {\left\langle \mathcal {F} _ {l o w} ^ {A} \left(X _ {i} ^ {t}\right) , q _ {s} \right\rangle}{\left\| \mathcal {F} _ {l o w} ^ {A} \left(X _ {i} ^ {t}\right) \right\| * \left\| q _ {s} \right\|}, \tag {2} +$$ + +where $\langle \cdot ,\cdot \rangle$ denotes the dot product operation, and $\| \cdot \|$ denotes the Euclidean norm. Since the low-frequency component can represent the style texture of an image, we rank the similarity scores and then retrieve the values of the $K$ most relevant keys from $M$ to construct a support set $R_{i} = \{q_{k},v_{k}\}_{k = 1}^{K}$ for the current test image. Finally, we assign a weight $w_{k}$ to $v_{k}$ to initialize the current prompt by $\mathcal{P}_i = \sum_{k = 1}^K w_k v_k$ where the weights $\{w_{k}\}$ are calculated based on the similarity scores [45] and $\sum_{k = 1}^{K}w_{k} = 1$ + +# 3.4. Prompt Training: Statistics Alignment + +Inspired by [19, 27, 47], the distribution shift largely exists in distinct BN statistics calculated on the test features (i.e., $\mu_t$ and $\sigma_t$ ) and stored in the source model $f_{\theta}$ (i.e., $\mu_s$ and $\sigma_s$ ), where $\mu$ and $\sigma$ respectively denote the mean and standard deviation. To train the prompts to align the statistics, we adopt the absolute-distance-based loss $\mathcal{L}_p$ [19] as follows: + +$$ +\mathcal {L} _ {p} = \frac {1}{j} \sum \left| \mu_ {s} ^ {j} - \mu_ {t} ^ {j} \right| + \left| \sigma_ {s} ^ {j} - \sigma_ {t} ^ {j} \right|, \tag {3} +$$ + +Algorithm 1: VPTTA algorithm. +Initialize: Frozen source model $f_{\theta}$ , initial prompt $\mathcal{P}$ filled with ones, memory bank $M$ , hyperparameters $S$ , $K$ and $\tau$ . + +Input: For each time step $i$ , current test image $X_{i}^{t}$ . + +1: ▷ Initialize the prompt $\mathcal{P}_i$ : +2: if $i < K$ then +3: $\mathcal{P}_i = \mathcal{P}$ +4: else +5: Construct the support set $R_{i}$ from $M$ +6: $\mathcal{P}_i = \sum_{k=1}^K w_k v_k, v_k \in R_i$ +7: end if +8: $\triangleright$ Train the prompt $P_{i}$ : +9: Obtain $\hat{X}_i^t$ using $\mathcal{P}_i$ by Eq. (1) +10: Forward $f_{\theta}(X_i^t)$ and calculate the warm-up statistics $\mu_w$ and $\sigma_w$ by Eq. (4) to replace the $\mu_s$ and $\sigma_s$ stored in $f_{\theta}$ +11: Backward and Update $\mathcal{P}_i$ by Eq. (5) +12: Inference: +13: Obtain $\hat{X}_i^t$ using updated $\mathcal{P}_i$ by Eq. (1) +14: Forward $O_{i} = f_{\theta}(X_{i}^{t})$ +15: Update the memory bank $M$ : +16: M.enqueue $(q_{i},v_{i})$ $q_{i} = \mathcal{F}^{A}\left(X_{i}^{t}\right),v_{i} = \mathcal{P}_{i}$ +17: if $\text{len}(M) > S$ then +18: M.dequeue $(q_{\text{first}}, v_{\text{first}})$ +19: end if + +Output:Adapted prediction $O_{i}$ + +where $j$ indicates the $j$ -th BN layer in $f_{\theta}$ . + +However, aligning the source and target domains within a limited number of iterations (one in this study) provides challenges due to the empty memory bank and undiminished distribution shifts at the beginning of the inference phase. Therefore, we design a warm-up mechanism to address this issue by simulating the warm-up statistics (i.e., $\mu_w$ and $\sigma_w$ ) as follows: + +$$ +\mu_ {w} = \lambda \mu_ {t} + (1 - \lambda) \mu_ {s}, \sigma_ {w} = \lambda \sigma_ {t} + (1 - \lambda) \sigma_ {s}, \tag {4} +$$ + +where $\lambda = \frac{1}{\sqrt{i} / \tau + 1}$ , $i$ indicates the index of the current test image and starts from 1, and $\tau$ is a temperature coefficient to control the rate of transition from the warm-up statistics to the source statistics. Now, the Eq. (3) can be rewritten by: + +$$ +\mathcal {L} _ {p} = \frac {1}{j} \sum \left| \mu_ {w} ^ {j} - \mu_ {t} ^ {j} \right| + \left| \sigma_ {w} ^ {j} - \sigma_ {t} ^ {j} \right|, \tag {5} +$$ + +and we utilize $\mu_w$ and $\sigma_w$ for normalization instead of $\mu_s$ and $\sigma_s$ . As the inference progresses, the warm-up statistics will gradually shift towards the source statistics. By simulating the warm-up statistics between the source and target statistics, we can significantly reduce the training difficulty at the beginning of the inference phase, resulting in learning prompts from easy to difficult. The overall process of VPTTA is summarized in Algorithm 1. + +# 4. Experiments + +We evaluate our proposed VPTTA on two 2D segmentation benchmark tasks: the joint optic disc (OD) and cup (OC) segmentation task, and the polyp segmentation task. + +# 4.1. Datasets and Evaluation Metrics + +The OD/OC segmentation dataset comprises five public datasets collected from different medical centres, denoted as domain A (RIM-ONE-r3 [11]), B (REFuge [32]), C (ORIGA [50]), D (REFuge-Validation/Test [32]), and E (Drishti-GS [38]). There are 159, 400, 650, 800, and 101 images from these datasets. We cropped a region of interest (ROI) centering at OD with size of $800 \times 800$ for each image following [19], and each ROI is further resized to $512 \times 512$ and normalized by min-max normalization. The Dice score metric $(DSC)$ is utilized for evaluation in this task. + +The polyp segmentation dataset consists of four public datasets collected from different medical centres, denoted as domain A (BKAI-IGH-NeoPolyp [28]), B (CVC-ClinicDB [1]), C (ETIS-LaribPolypDB [36]), and D (Kvasir-Seg [20]), which have 1000, 612, 196, and 1000 images, respectively. We followed [9] to resize each image to $352 \times 352$ and normalize the resized image by the statistics computed on ImageNet. The $DSC$ , enhanced-alignment metric $(E_{\phi}^{max})$ [8], and structural similarity metric $(S_{\alpha})$ [7] are utilized for evaluation in this task. + +# 4.2. Implementation Details + +For each task, we trained the source model on each single domain (source domain) and tested it on the left domains (target domains) to calculate the mean metrics to evaluate the methods on different scenarios. In the source-training phase, we trained a ResUNet-34 [17] backbone, following [19], as the baseline for OD/OC segmentation task and trained the PraNet [9] with a Res2Net-based [14] backbone for polyp segmentation task. Due to the complexity of PraNet's decoder, we only calculated the prompt loss $\mathcal{L}_p$ for each BN layer in the encoder instead of the whole network. In the test-adaptation phase, We performed one-iteration adaptation for each batch of test data with a batch size of 1 on all experiments of our VPTTA and other competing methods [45]. As our VPTTA involves an iteration to update and an inference step, distinguishing it from other methods like [41] that do not require a dedicated inference step, we applied the same configuration as VPTTA to other competing methods to ensure consistency in the experimental setup. To deploy our VPTTA, we utilized the Adam optimizer with a learning rate of 0.05 and 0.01 for the OD/OC segmentation task and polyp segmentation task, respectively. The hyperparameters $\alpha$ (size of prompt), $S$ (size of memory bank), $K$ (size of support set), and $\tau$ (temperature coefficient in warm-up) are set to 0.01, 40, 16, and 5 for both segmentation tasks. + +Table 1. Performance of our VPTTA, 'Source Only' baseline, and six competing methods on the OD/OC segmentation task. The best and second-best results in each column are highlighted in bold and underline, respectively. + +
MethodsDomain ADomain BDomain CDomain DDomain EAverage
DSCDSCDSCDSCDSCDSC↑
Source Only (ResUNet-34)64.5376.0671.1852.6764.8765.86
TENT-continual (ICLR 2021) [41]73.0778.6671.9446.8170.2068.13
CoTTA (CVPR 2022) [42]75.3975.9869.1453.9970.4068.98
DLTTA (TMI 2022) [45]75.1178.8573.8951.6469.7169.84
DUA (CVPR 2022) [25]72.2876.5970.1356.1771.3869.31
SAR (ICLR 2023) [31]74.5577.7170.7855.4071.7270.03
DomainAdaptor (CVPR 2023) [47]74.5076.3971.8156.7870.5570.01
VPTTA (Ours)73.9179.3674.5156.5175.3571.93
+ +Table 2. Performance of our VPTTA, 'Source Only' baseline, and six competing methods on the polyp segmentation task. The best and second-best results in each column are highlighted in bold and underline, respectively. + +
MethodsDomain ADomain BDomain CDomain DAverage
\(DSC\)\(E_{\phi}^{max}\)\(S_{\alpha}\)\(DSC\)\(E_{\phi}^{max}\)\(S_{\alpha}\)\(DSC\)\(E_{\phi}^{max}\)\(S_{\alpha}\)\(DSC\)\(E_{\phi}^{max}\)\(S_{\alpha}\)\(DSC \uparrow\)\(E_{\phi}^{max} \uparrow\)\(S_{\alpha} \uparrow\)
Source Only (PraNet)79.9087.9784.6666.3378.5176.7273.8984.6481.2882.9590.8488.0875.7785.4982.69
TENT-continual (ICLR 2021) [41]74.8684.5880.5267.5178.6678.0517.7940.0453.3073.5583.3882.4158.4371.6773.57
CoTTA (CVPR 2022) [42]76.4685.3782.5666.7776.7579.1771.3983.4280.1870.7179.8182.5471.3381.3481.11
DLTTA (TMI 2022) [45]76.2785.2382.4166.5877.0079.2463.7278.2375.5671.2081.3283.4769.4480.4580.17
DUA (CVPR 2022) [25]78.9387.3783.9666.8478.5277.5176.5386.4583.0586.2493.2389.8277.1386.3983.58
SAR (ICLR 2023) [31]76.4885.8981.4966.4577.3578.0571.4683.2379.4070.4180.1181.0771.2081.6580.00
DomainAdaptor (CVPR 2023) [47]77.4886.3182.4070.8281.7680.8871.9683.0679.9776.8985.8984.4574.2984.2681.93
VPTTA (Ours)81.0088.9184.9176.8787.3184.0877.5887.4883.6486.3993.4789.8780.4689.2985.62
+ +Table 3. Results of ablation study on the OD/OC segmentation task. The best and second-best results in each column are highlighted in bold and underline, respectively. + +
MethodsDomain ADomain BDomain CDomain DDomain EAverage
Low-frequency PromptMemory BankWarm-upDSCDSCDSCDSCDSCDSC↑
64.5376.0671.1852.6764.8765.86
✓65.9976.5871.7552.8667.3066.90
✓✓70.7079.3672.8255.5072.8470.24
✓✓74.4177.4774.1454.0367.8569.58
✓✓✓73.9179.3674.5156.5175.3571.93
+ +# 4.3. Experimental Results + +We compare our VPTTA with the 'Source Only' baseline (training on the source domain and testing without adaptation), and six competing methods, including a pseudo-label-based method (CoTTA [42]), two entropy-based methods (TENT-continual [41] and SAR [31]), a method dynamically adjusting the learning rate (DLTTA [45]), a method combining entropy-loss and BN statistics fusion (DomainAdaptor [47]), and a BN statistics-based method (DUA [25]) on two segmentation tasks. Specifically, we re-implemented all the competing methods using the same baseline as VPTTA and combined DLTTA with the entropy loss proposed in TENT [45]. The result of each domain is calculated by using the current domain as the source domain and applying the pre-trained model on the left domains. + +Comparison on the OD/OC segmentation task. The results of the proposed VPTTA, the "Source Only" baseline, and six competing methods on the OD/OC segmentation task are detailed in Table 1. Obviously, all competing methods outperform the "Source Only" baseline, indicating the + +effectiveness of adaptation towards the distribution shifts. It can be found that our VPTTA achieves the best overall performance across all domains, underscoring its superior applicability and robustness. Meanwhile, our VPTTA also shows competitive performance in domains where most methods exhibit satisfactory performance (such as D) and also performs well in domains where other methods exhibit suboptimal results (such as C). + +Comparison on the polyp segmentation task. We conducted the comparison experiment on the polyp segmentation task under the same experimental setup, as shown in Table 2. In contrast to the results of the OD/OC segmentation task, the methods based on model updating (i.e., TENT-continual, CoTTA, DLTTA, SAR, and DomainAdaptor) exhibit significantly worse performance in this task and are even inferior to the 'Source Only' baseline. It can be attributed to the fact that the polyps are relatively hidden, and the models tend to segment almost nothing rather than poorly segment under the impact of distribution shifts, resulting in confident and low entropy predictions but wrong + +![](images/94d284a1840b9156326e8cf956d20c3c1507350ec98d944cf0da0f56a1f69e5a.jpg) +Figure 3. Visualization of the original images, estimated prompts, and adapted images on the OD/OC segmentation task. We normalize the prompts to [0, 1] for better visualization. The DSC of applying the frozen source model on the original and adapted images is displayed below each image. We also show an example of each source domain on the left side of this diagram. 'Ori': Abbreviation of 'Original'. + +gradients for these methods. Since DUA is backward-free, it still performs well in this task. However, DUA also fails when using domain A as the source domain and has limited performance due to the absence of training for the current test image. Our VPTTA simultaneously benefits from training on the current test image and refraining from updating the model, thus surpassing all competing methods and demonstrating advanced performance. + +Visualization of prompts and adapted images. We visualized the estimated prompts and adapted images on the OD/OC segmentation task in Figure 3 and displayed the DSC values predicted on the original and adapted images below the corresponding images. It can be observed that: (1) Applying prompts to the original test images can achieve significant performance gain; (2) Images from different target domains have similar prompts, proving that involving the memory bank to initialize the current prompt using previous ones is reasonable; (3) The appearances of adapted images are close to the source domain, explicitly demonstrating the reduction of distribution shifts via our VPTTA. + +Comparison in a long-term continual test-time adap + +tation. To evaluate our VPTTA in a long-term continual test-time adaptation [42], we conducted the experiments for the OD/OC segmentation task multiple rounds and reported the results in Table 4. Since DUA is a backward-free method and DomainAdaptor resets the pre-trained model before each adaptation, they are unsuitable to be evaluated in this experiment. We compared our VPTTA with TENT, CoTTA, DLTTA, and SAR, all of which involve continuous training of the pre-trained model. The performance across different rounds with the same source domain reflects the extent of catastrophic forgetting in the model, and the average performance across multiple rounds highlights the error accumulation of the model within a series of distinct domains. The results reveal that both optimizing loss functions (CoTTA, DLTTA, SAR) and resetting the model (CoTTA, SAR) are effective for mitigating error accumulation and catastrophic forgetting, while TENT with vanilla entropy-minimization loss suffers from severe performance degradation. Additionally, after multiple rounds of testing, our VPTTA maintains superior performance with minimal performance degradation, emphasizing the effectiveness of + +Table 4. Performance of our VPTTA and four competing methods on the OD/OC segmentation task in a long-term continual test-time adaptation. The red numbers indicate the performance gain relative to the 'Source Only' baseline. The performance degradation is calculated between the overall average DSC and the average DSC of round 1. 'Ave': Abbreviation of 'Average'. + +
Time→
Round123Average DSC ↑Perform. Degra. ↓
MethodsABCDEAveABCDEAveABCDEAve
Source Only64.5376.0671.1852.6764.8765.8664.5376.0671.1852.6764.8765.8664.5376.0671.1852.6764.8765.8665.86
TENT73.0778.6671.9446.8170.2068.1362.0969.3270.6739.0268.2261.8657.0562.4770.2039.0266.3759.0263.01 (-2.85)
CoTTA75.3975.9869.1453.9970.4068.9874.3175.0067.9951.0468.2867.3273.2274.3366.7250.2367.0866.3267.54 (+1.68)
DLTTA75.1178.8573.8951.6469.7169.8474.1479.6574.2545.0569.0468.4372.2878.9372.8742.3769.2667.1468.47 (+2.61)
SAR74.5577.7170.7855.4071.7270.0374.7478.0971.0052.1369.0269.0074.9078.2471.1850.1668.4468.5869.20 (+3.34)
VPTTA (Ours)73.9179.3674.5156.5175.3571.9373.5778.8473.6156.9174.8071.5573.1278.4572.6357.1174.0471.0771.51 (+5.65)
+ +training prompts over updating the pre-trained model. + +Ablation study. To evaluate the contributions of the low-frequency prompt, memory-bank-based initialization, and warm-up mechanism, we conducted ablation studies on the OD/OC segmentation task, as shown in Table 3. Note that Eq. (3)/(5) is used as the loss function when training the prompt without/with our warm-up mechanism. The results show that (1) using the low-frequency prompt only has limited performance gain due to the poorly trained prompt; (2) both the memory bank and warm-up mechanism are beneficial to training the prompt; (3) the best performance is achieved when the memory bank and warm-up mechanism are jointly used, which means they are complementary. + +![](images/585cc210003974459a61397dbb3425e8ea254f78960f4a424dc0e4ba883f8fb7.jpg) +Average DSC +Figure 4. Performance of our VPTTA with various $\alpha$ on the OD/OC segmentation task. + +![](images/6a6f9dace7fe6ab79db53b357d3d6f2f284c705985fe75caf97aa6473a009c5a.jpg) +Average DSC +Figure 5. Performance of our VPTTA with various $S$ , $K$ , and $\tau$ on the OD/OC segmentation task. + +Analysis of hyperparameter $\alpha$ , $S$ , $K$ , and $\tau$ . We conducted the experiments for the OD/OC segmentation task to discuss the hyper-parameters used in our VPTTA and displayed the results in Figure 4 and Figure 5. It can be seen that $\alpha = 0.01$ is optimal, as an excessively small prompt makes adaptation challenging, while an overly large prompt hinders effective training. Similar trends are also reflected + +Table 5. Performance of our VPTTA and VPTTA-LR on two segmentation tasks. + +
MethodsOD/OCPolyp
DSC↑Param↓DSC↑Eφmax↑Sα↑Param↓
VPTTA-LR (r=3)69.019,21678.2588.2784.206,336
VPTTA71.937580.4689.9486.8427
+ +in $S$ and $K$ : as they increase, the DSC values improve, while the excessively large $S$ and $K$ may introduce unrelated samples, affecting the initialization. With the increase of $\tau$ , performance gradually approaches the upper bound and tends to remain stable, suggesting that the warm-up has reached saturation. Therefore, we finally chose $\alpha = 0.01$ , $S = 40$ , $K = 16$ , and $\tau = 5$ as the best configuration. + +Low-frequency prompt vs. low-rank prompt. To compare the effectiveness of VPTTA and VPTTA-LR, we repeated the experiments on two segmentation tasks and showed results in Table 5. It can be seen that VPTTA significantly outperforms VPTTA-LR, attributed to the fewer learnable parameters and superior representation of style texture in the frequency domain. Benefiting from the low-frequency prompt, VPTTA can be effectively trained with over a hundred times fewer parameters than VPTTA-LR. + +# 5. Conclusion + +In this paper, we proposed the VPTTA, a prompt-based method to alleviate the potential risks (i.e., error accumulation, and catastrophic forgetting) by freezing the pre-trained model. Specially, we presented the low-frequency prompt that trains a lightweight prompt with only a small number of parameters in a single iteration to alleviate the distribution shifts in BN layers, constructed a memory bank to better initialize each prompt by utilizing the prompts of previous and the most similar test images, and designed the statistics-fusion-based warm-up mechanism to simulate the warm-up statistics to overcome the training difficulty at the beginning of the inference phase. The extensive experimental results on two medical image segmentation benchmark tasks with multiple domains verified the superiority of our VPTTA over other methods and the effectiveness of each special design. In this regard, we hope that our efforts can offer a different paradigm for continual test-time adaptation. + +# References + +[1] Jorge Bernal, F Javier Sánchez, Gloria Fernández-Esparrach, Debora Gil, Cristina Rodríguez, and Fernando Vilarino. Wm-dova maps for accurate polyp highlighting in colonoscopy: Validation vs. saliency maps from physicians. Comput. Med. Imaging and Graph., 43:99-111, 2015. 5 +[2] Paola Cascante-Bonilla, Fuwen Tan, Yanjun Qi, and Vicente Ordonez. Curriculum labeling: Revisiting pseudo-labeling for semi-supervised learning. In AAAI, pages 6912-6920, 2021. 2, 3 +[3] Oscar Chang, Lampros Flokas, and Hod Lipson. Principled weight initialization for hypernetworks. In Int. Conf. Learn. Represent., 2019. 2 +[4] Dian Chen, Dequan Wang, Trevor Darrell, and Sayna Ebrahimi. Contrastive test-time adaptation. In IEEE Conf. Comput. Vis. Pattern Recog., pages 295-305, 2022. 3 +[5] Lin Chen, Huaian Chen, Zhixiang Wei, Xin Jin, Xiao Tan, Yi Jin, and Enhong Chen. Reusing the task-specific classifier as a discriminator: Discriminator-free adversarial domain adaptation. In IEEE Conf. Comput. Vis. Pattern Recog., pages 7181-7190, 2022. 2 +[6] Mario Döbler, Robert A Marsden, and Bin Yang. Robust mean teacher for continual and gradual test-time adaptation. In IEEE Conf. Comput. Vis. Pattern Recog., pages 7704-7714, 2023. 2, 3 +[7] Deng-Ping Fan, Ming-Ming Cheng, Yun Liu, Tao Li, and Ali Borji. Structure-measure: A new way to evaluate foreground maps. In Int. Conf. Comput. Vis., pages 4548-4557, 2017. 5 +[8] Deng-Ping Fan, Cheng Gong, Yang Cao, Bo Ren, Ming-Ming Cheng, and Ali Borji. Enhanced-alignment measure for binary foreground map evaluation. In IJCAI, pages 698-704, 2018. 5 +[9] Deng-Ping Fan, Ge-Peng Ji, Tao Zhou, Geng Chen, Huazhu Fu, Jianbing Shen, and Ling Shao. Pranet: Parallel reverse attention network for polyp segmentation. In Int. Conf. Med. Image Comput. Comput.-Assist. Intervention., pages 263-273. Springer, 2020. 5 +[10] Matteo Frigo and Steven G Johnson. FFTW: An adaptive software architecture for the fft. In ICASSP, pages 1381-1384. IEEE, 1998. 3 +[11] Francisco Fumero, Silvia Alayón, José L Sanchez, Jose Sigut, and M Gonzalez-Hernandez. RIM-ONE: An open retinal image database for optic nerve evaluation. In Int. Symp. Comput.-based Med. Syst., pages 1-6. IEEE, 2011. 5 +[12] Yulu Gan, Yan Bai, Yihang Lou, Xianzheng Ma, Renrui Zhang, Nian Shi, and Lin Luo. Decorate the newcomers: Visual domain prompt for continual test time adaptation. In AAAI, pages 7595-7603, 2023. 3 +[13] Yossi Gandelsman, Yu Sun, Xinlei Chen, and Alexei Efros. Test-time training with masked autoencoders. Adv. Neural Inform. Process. Syst., 35:29374-29385, 2022. 3 +[14] Shang-Hua Gao, Ming-Ming Cheng, Kai Zhao, Xin-Yu Zhang, Ming-Hsuan Yang, and Philip Torr. Res2net: A new multi-scale backbone architecture. IEEE Trans. Pattern Anal. Mach. Intell., 43(2):652-662, 2019. 5 + +[15] Yunhe Gao, Xingjian Shi, Yi Zhu, Hao Wang, Zhiqiang Tang, Xiong Zhou, Mu Li, and Dimitris N Metaxas. Visual prompt tuning for test-time domain adaptation. arXiv preprint arXiv:2210.04831, 2022. 3 +[16] Mohsen Ghafoorian, Alireza Mehrtash, Tina Kapur, Nico Karssemeijer, Elena Marchiori, Mehran Pesteie, Charles RG Guttmann, Frank-Erik de Leeuw, Clare M Tempany, Bram Van Ginneken, et al. Transfer learning for domain adaptation in mri: Application in brain lesion segmentation. In Int. Conf. Med. Image Comput. Comput.-Assist. Intervent., pages 516-524. Springer, 2017. 2 +[17] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In IEEE Conf. Comput. Vis. Pattern Recog., pages 770-778, 2016. 5 +[18] Edward J Hu, Phillip Wallis, Zeyuan Allen-Zhu, Yuzhhi Li, Shean Wang, Lu Wang, Weizhu Chen, et al. Lora: Low-rank adaptation of large language models. In Int. Conf. Learn. Represent., 2022. 4 +[19] Shishuai Hu, Zehui Liao, and Yong Xia. Prosfda: Prompt learning based source-free domain adaptation for medical image segmentation. arXiv preprint arXiv:2211.11514, 2022. 2, 3, 4, 5 +[20] Debesh Jha, Pia H Smedsrud, Michael A Riegler, Pål Halvorsen, Thomas de Lange, Dag Johansen, and Håvard D Johansen. Kvasir-seg: A segmented polyp dataset. In Int. Conf. Multimedia Modeling, pages 451-462. Springer, 2020. 5 +[21] Christian Lamers, René Vidal, Nabil Belbachir, Niki van Stein, Thomas Bäeck, and Paris Giampouras. Clustering-based domain-incremental learning. In Int. Conf. Comput. Vis., pages 3384-3392, 2023. 2, 3 +[22] Joonseok Lee, Seungyeon Kim, Guy Lebanon, and Yoram Singer. Local low-rank matrix approximation. In Int. Conf. Mach. Learn., pages 82-90. PMLR, 2013. 4 +[23] Jian Liang, Ran He, Zhenan Sun, and Tieniu Tan. Exploring uncertainty in pseudo-label guided unsupervised domain adaptation. Pattern Recognition, 96:106996, 2019. 2, 3 +[24] Pengfei Liu, Weizhe Yuan, Jinlan Fu, Zhengbao Jiang, Hiroaki Hayashi, and Graham Neubig. Pre-train, prompt, and predict: A systematic survey of prompting methods in natural language processing. ACM Comput. Surveys, 55(9):1-35, 2023. 3 +[25] M Jehanzeb Mirza, Jakub Micorek, Horst Possegger, and Horst Bischof. The norm must go on: Dynamic unsupervised domain adaptation by normalization. In IEEE Conf. Comput. Vis. Pattern Recog., pages 14765-14775, 2022. 3, 6 +[26] Chaithanya Kumar Mummadi, Robin Hutmacher, Kilian Rambach, Evgeny Levinkov, Thomas Brox, and Jan Hendrik Metzen. Test-time adaptation to distribution shift by confidence maximization and input transformation. arXiv preprint arXiv:2106.14999, 2021. 3 +[27] Zachary Nado, Shreyas Padhy, D Sculley, Alexander D'Amour, Balaji Lakshminarayanan, and Jasper Snoek. Evaluating prediction-time batch normalization for robustness under covariate shift. arXiv preprint arXiv:2006.10963, 2020. 3, 4 + +[28] Phan Ngoc Lan, Nguyen Sy An, Dao Viet Hang, Dao Van Long, Tran Quang Trung, Nguyen Thi Thuy, and Dinh Viet Sang. Neounet: Towards accurate colon polyp segmentation and neoplasm detection. In Adv. Vis. Comput. Int. Symp., pages 15-28. Springer, 2021. 5 +[29] A Tuan Nguyen, Thanh Nguyen-Tang, Ser-Nam Lim, and Philip HS Torr. Tipi: Test time adaptation with transformation invariance. In IEEE Conf. Comput. Vis. Pattern Recog., pages 24162-24171, 2023. 2, 3 +[30] Shuaicheng Niu, Jiaxiang Wu, Yifan Zhang, Yafo Chen, Shijian Zheng, Peilin Zhao, and Mingkui Tan. Efficient test-time model adaptation without forgetting. In Int. Conf. Mach. Learn., pages 16888-16905. PMLR, 2022. 2, 3 +[31] Shuaicheng Niu, Jiaxiang Wu, Yifan Zhang, Zhiquan Wen, Yaofo Chen, Peilin Zhao, and Mingkui Tan. Towards stable test-time adaptation in dynamic wild world. In Int. Conf. Learn. Represent., 2023. 2, 3, 6 +[32] José Ignacio Orlando, Huazhu Fu, João Barbosa Breda, Karel van Keer, Deepti R Bathula, Andrés Diaz-Pinto, Ruogu Fang, Pheng-Ann Heng, Jyoung Kim, JoonHo Lee, et al. REFuge Challenge: A unified framework for evaluating automated methods for glaucoma assessment from fundus photographs. Med. Image Anal., 59:101570, 2020. 5 +[33] Xingang Pan, Ping Luo, Jianping Shi, and Xiaou Tang. Two at once: Enhancing learning and generalization capacities via ibn-net. In *Eur. Conf. Comput. Vis.*, pages 464-479, 2018. 2 +[34] Swami Sankaranarayanan, Yogesh Balaji, Arpit Jain, Ser Nam Lim, and Rama Chellappa. Learning from synthetic data: Addressing domain shift for semantic segmentation. In IEEE Conf. Comput. Vis. Pattern Recog., pages 3752-3761, 2018. 2 +[35] Steffen Schneider, Evgenia Rusak, Luisa Eck, Oliver Bringmann, Wieland Brendel, and Matthias Bethge. Improving robustness against common corruptions by covariate shift adaptation. Adv. Neural Inform. Process. Syst., 33:11539-11551, 2020. 2 +[36] Juan Silva, Aymeric Histace, Olivier Romain, Xavier Dray, and Bertrand Granado. Toward embedded detection of polyps in wce images for early diagnosis of colorectal cancer. Int. J. Comput.-Assist. Rad. Sur., 9(2):283-293, 2014. 5 +[37] Samarth Sinha, Peter Gehler, Francesco Locatello, and Bernt Schiele. Test: Test-time self-training under distribution shift. In IEEE Wint. Conf. Appl. Comput. Vis., pages 2759-2769, 2023. 2, 3 +[38] Jayanthi Sivaswamy, SR Krishnadas, Gopal Datt Joshi, Madhulika Jain, and A Ujjwaft Syed Tabish. Drishti-GS: Retinal image dataset for optic nerve head (onh) segmentation. In IEEE Int. Symp. on Bio. Imaging, pages 53-56. IEEE, 2014. 5 +[39] Junha Song, Jungsoo Lee, In So Kweon, and Sungha Choi. Ecotta: Memory-efficient continual test-time adaptation via self-distilled regularization. In IEEE Conf. Comput. Vis. Pattern Recog., pages 11920-11929, 2023. 2, 3 +[40] Yu Sun, Xiaolong Wang, Zhuang Liu, John Miller, Alexei Efros, and Moritz Hardt. Test-time training with self-supervision for generalization under distribution shifts. In + +Int. Conf. Mach. Learn., pages 9229-9248. PMLR, 2020. 2, 3 +[41] Dequan Wang, Evan Shelhamer, Shaoteng Liu, Bruno Olshausen, and Trevor Darrell. Tent: Fully test-time adaptation by entropy minimization. In Int. Conf. Learn. Represent., 2021. 2, 3, 5, 6 +[42] Qin Wang, Olga Fink, Luc Van Gool, and Dengxin Dai. Continual test-time domain adaptation. In IEEE Conf. Comput. Vis. Pattern Recog., pages 7201-7211, 2022. 2, 3, 6, 7 +[43] Shujun Wang, Lequan Yu, Kang Li, Xin Yang, Chi-Wing Fu, and Pheng-Ann Heng. Boundary and entropy-driven adversarial learning for fundus image segmentation. In Int. Conf. Med. Image Comput. Comput.-Assist. Intervention., pages 102-110. Springer, 2019. 2 +[44] Wei Wang, Zhun Zhong, Weijie Wang, Xi Chen, Charles Ling, Boyu Wang, and Nicu Sebe. Dynamically instance-guided adaptation: A backward-free approach for test-time domain adaptive semantic segmentation. In IEEE Conf. Comput. Vis. Pattern Recog., pages 24090-24099, 2023. 2, 3 +[45] Hongzheng Yang, Cheng Chen, Meirui Jiang, Quande Liu, Jianfeng Cao, Pheng Ann Heng, and Qi Dou. Dltta: Dynamic learning rate for test-time adaptation on cross-domain medical images. IEEE Trans. Med. Imaging, 41(12):3575-3586, 2022. 2, 3, 4, 5, 6 +[46] Yanchao Yang and Stefano Soatto. FDA: Fourier domain adaptation for semantic segmentation. In IEEE Conf. Comput. Vis. Pattern Recog., pages 4085-4095, 2020. 2 +[47] Jian Zhang, Lei Qi, Yinghuan Shi, and Yang Gao. Domain adaptor: A novel approach to test-time adaptation. In IEEE Conf. Comput. Vis. Pattern Recog., pages 18971-18981, 2023. 2, 3, 4, 6 +[48] Marvin Zhang, Sergey Levine, and Chelsea Finn. Memo: Test time robustness via adaptation and augmentation. Adv. Neural Inform. Process. Syst., 35:38629-38642, 2022. 2, 3 +[49] Yizhe Zhang, Tao Zhou, Yuhui Tao, Shuo Wang, Ye Wu, Benyuan Liu, Pengfei Gu, Qiang Chen, and Danny Z Chen. TestFit: A plug-and-play one-pass test time method for medical image segmentation. Medical Image Analysis, 92: 103069, 2024. 3 +[50] Zhuo Zhang, Feng Shou Yin, Jiang Liu, Wing Kee Wong, Ngan Meng Tan, Beng Hai Lee, Jun Cheng, and Tien Yin Wong. ORIGA-light: An online retinal fundus image database for glaucoma analysis and research. In Int. Conf. IEEE Eng. in Med. and Bio., pages 3065-3068. IEEE, 2010. 5 +[51] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Zwei Liu. Conditional prompt learning for vision-language models. In IEEE Conf. Comput. Vis. Pattern Recog., pages 16816-16825, 2022. 3 +[52] Chen Zhu, Renkun Ni, Zheng Xu, Kezhi Kong, W Ronny Huang, and Tom Goldstein. Gradinit: Learning to initialize neural networks for stable and efficient training. Adv. Neural Inform. Process. Syst., 34:16410-16422, 2021. 2 +[53] Yukun Zuo, Hantao Yao, and Changsheng Xu. Attention-based multi-source domain adaptation. IEEE Trans. Image Process., 30:3793-3803, 2021. 2 \ No newline at end of file diff --git a/2024/Each Test Image Deserves A Specific Prompt_ Continual Test-Time Adaptation for 2D Medical Image Segmentation/images.zip b/2024/Each Test Image Deserves A Specific Prompt_ Continual Test-Time Adaptation for 2D Medical Image Segmentation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..73d994cb4d42b96afc9d7ae7467e177c418d0d17 --- /dev/null +++ b/2024/Each Test Image Deserves A Specific Prompt_ Continual Test-Time Adaptation for 2D Medical Image Segmentation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5f155929d4bf903cf3f7f28664f1af9c0f5e7daed815e9e5bf785c6efe3573e7 +size 690634 diff --git a/2024/Each Test Image Deserves A Specific Prompt_ Continual Test-Time Adaptation for 2D Medical Image Segmentation/layout.json b/2024/Each Test Image Deserves A Specific Prompt_ Continual Test-Time Adaptation for 2D Medical Image Segmentation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..370357146987a4ac4bc098bbde68d92fe2dc28b3 --- /dev/null +++ b/2024/Each Test Image Deserves A Specific Prompt_ Continual Test-Time Adaptation for 2D Medical Image Segmentation/layout.json @@ -0,0 +1,10484 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 54, + 103, + 539, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 103, + 539, + 140 + ], + "spans": [ + { + "bbox": [ + 54, + 103, + 539, + 140 + ], + "type": "text", + "content": "Each Test Image Deserves A Specific Prompt: Continual Test-Time Adaptation for 2D Medical Image Segmentation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 160, + 485, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 105, + 160, + 485, + 175 + ], + "spans": [ + { + "bbox": [ + 105, + 160, + 485, + 175 + ], + "type": "text", + "content": "Ziyang Chen" + }, + { + "bbox": [ + 105, + 160, + 485, + 175 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 105, + 160, + 485, + 175 + ], + "type": "text", + "content": " Yongsheng Pan" + }, + { + "bbox": [ + 105, + 160, + 485, + 175 + ], + "type": "inline_equation", + "content": "^{1\\dagger}" + }, + { + "bbox": [ + 105, + 160, + 485, + 175 + ], + "type": "text", + "content": " Yiwen Ye" + }, + { + "bbox": [ + 105, + 160, + 485, + 175 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 105, + 160, + 485, + 175 + ], + "type": "text", + "content": " Mengkang Lu" + }, + { + "bbox": [ + 105, + 160, + 485, + 175 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 105, + 160, + 485, + 175 + ], + "type": "text", + "content": " Yong Xia" + }, + { + "bbox": [ + 105, + 160, + 485, + 175 + ], + "type": "inline_equation", + "content": "^{1,2,3\\dagger}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 59, + 175, + 534, + 218 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 175, + 534, + 218 + ], + "spans": [ + { + "bbox": [ + 59, + 175, + 534, + 218 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 59, + 175, + 534, + 218 + ], + "type": "text", + "content": " School of Computer Science and Engineering, Northwestern Polytechnical University, China \n" + }, + { + "bbox": [ + 59, + 175, + 534, + 218 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 59, + 175, + 534, + 218 + ], + "type": "text", + "content": " Research & Development Institute of Northwestern Polytechnical University in Shenzhen, China \n" + }, + { + "bbox": [ + 59, + 175, + 534, + 218 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 59, + 175, + 534, + 218 + ], + "type": "text", + "content": " Ningbo Institute of Northwestern Polytechnical University, China" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 120, + 220, + 473, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 120, + 220, + 473, + 232 + ], + "spans": [ + { + "bbox": [ + 120, + 220, + 473, + 232 + ], + "type": "text", + "content": "zychen@mail.nwpu.edu.cn, yspan@nwpu.edu.cn, ywye@mail.nwpu.edu.cn" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 192, + 234, + 399, + 245 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 192, + 234, + 399, + 245 + ], + "spans": [ + { + "bbox": [ + 192, + 234, + 399, + 245 + ], + "type": "text", + "content": "lmk@mail.nwpu.edu.cn, yxia@nwpu.edu.cn" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 56, + 285, + 104, + 325 + ], + "blocks": [ + { + "bbox": [ + 50, + 273, + 112, + 285 + ], + "lines": [ + { + "bbox": [ + 50, + 273, + 112, + 285 + ], + "spans": [ + { + "bbox": [ + 50, + 273, + 112, + 285 + ], + "type": "text", + "content": "Source Model " + }, + { + "bbox": [ + 50, + 273, + 112, + 285 + ], + "type": "inline_equation", + "content": "f_{\\theta}" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 56, + 285, + 104, + 325 + ], + "lines": [ + { + "bbox": [ + 56, + 285, + 104, + 325 + ], + "spans": [ + { + "bbox": [ + 56, + 285, + 104, + 325 + ], + "type": "image", + "image_path": "b7f9bd6d88491770ebe0d90b0c346003251b1c61ebacdc964fdb0880fc4578ee.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 108, + 274, + 170, + 328 + ], + "blocks": [ + { + "bbox": [ + 108, + 274, + 170, + 328 + ], + "lines": [ + { + "bbox": [ + 108, + 274, + 170, + 328 + ], + "spans": [ + { + "bbox": [ + 108, + 274, + 170, + 328 + ], + "type": "image", + "image_path": "b8b46250f831afc9abdf05c7c6e2827ddb1168403b421ac04fd66a94dc21150c.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 175, + 284, + 220, + 328 + ], + "blocks": [ + { + "bbox": [ + 175, + 274, + 220, + 284 + ], + "lines": [ + { + "bbox": [ + 175, + 274, + 220, + 284 + ], + "spans": [ + { + "bbox": [ + 175, + 274, + 220, + 284 + ], + "type": "text", + "content": "Domain 2" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 175, + 284, + 220, + 328 + ], + "lines": [ + { + "bbox": [ + 175, + 284, + 220, + 328 + ], + "spans": [ + { + "bbox": [ + 175, + 284, + 220, + 328 + ], + "type": "image", + "image_path": "cd04bb644b92668adb948c54e6dd1e02cd01489ef8ab0b93524ac05af68b8489.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 220, + 274, + 271, + 284 + ], + "lines": [ + { + "bbox": [ + 220, + 274, + 271, + 284 + ], + "spans": [ + { + "bbox": [ + 220, + 274, + 271, + 284 + ], + "type": "text", + "content": "Domain 3" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 220, + 284, + 290, + 328 + ], + "blocks": [ + { + "bbox": [ + 220, + 284, + 290, + 328 + ], + "lines": [ + { + "bbox": [ + 220, + 284, + 290, + 328 + ], + "spans": [ + { + "bbox": [ + 220, + 284, + 290, + 328 + ], + "type": "image", + "image_path": "fc535331f710fa7ac4fce2eee57cf00ca969e955b8339f281b41741b0244954e.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 290, + 274, + 326, + 284 + ], + "lines": [ + { + "bbox": [ + 290, + 274, + 326, + 284 + ], + "spans": [ + { + "bbox": [ + 290, + 274, + 326, + 284 + ], + "type": "text", + "content": "Domain 4" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 231, + 330, + 261, + 342 + ], + "lines": [ + { + "bbox": [ + 231, + 330, + 261, + 342 + ], + "spans": [ + { + "bbox": [ + 231, + 330, + 261, + 342 + ], + "type": "inline_equation", + "content": "f_{\\theta}^{\\prime}\\rightarrow f_{\\theta}" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 290, + 284, + 328, + 328 + ], + "blocks": [ + { + "bbox": [ + 290, + 284, + 328, + 328 + ], + "lines": [ + { + "bbox": [ + 290, + 284, + 328, + 328 + ], + "spans": [ + { + "bbox": [ + 290, + 284, + 328, + 328 + ], + "type": "image", + "image_path": "7319de94f101abffc1db56458e03dba8862b88e6626de9b750b8ffadb1077ef0.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 328, + 274, + 361, + 284 + ], + "lines": [ + { + "bbox": [ + 328, + 274, + 361, + 284 + ], + "spans": [ + { + "bbox": [ + 328, + 274, + 361, + 284 + ], + "type": "text", + "content": "" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 328, + 284, + 350, + 327 + ], + "blocks": [ + { + "bbox": [ + 328, + 284, + 350, + 327 + ], + "lines": [ + { + "bbox": [ + 328, + 284, + 350, + 327 + ], + "spans": [ + { + "bbox": [ + 328, + 284, + 350, + 327 + ], + "type": "image", + "image_path": "e494199ddcc9269b92378aae9e5e5dd9177811b8a547070f8b40a5a81cf05d25.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 350, + 274, + 400, + 284 + ], + "lines": [ + { + "bbox": [ + 350, + 274, + 400, + 284 + ], + "spans": [ + { + "bbox": [ + 350, + 274, + 400, + 284 + ], + "type": "text", + "content": "Domain N" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 350, + 284, + 402, + 328 + ], + "blocks": [ + { + "bbox": [ + 350, + 284, + 402, + 328 + ], + "lines": [ + { + "bbox": [ + 350, + 284, + 402, + 328 + ], + "spans": [ + { + "bbox": [ + 350, + 284, + 402, + 328 + ], + "type": "image", + "image_path": "7e3ad50669bb3bfaf2ad32d3263adc01c67d8f7dc067ec88311807d661551c85.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 402, + 274, + 443, + 284 + ], + "lines": [ + { + "bbox": [ + 402, + 274, + 443, + 284 + ], + "spans": [ + { + "bbox": [ + 402, + 274, + 443, + 284 + ], + "type": "text", + "content": "" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 345, + 331, + 413, + 344 + ], + "lines": [ + { + "bbox": [ + 345, + 331, + 413, + 344 + ], + "spans": [ + { + "bbox": [ + 345, + 331, + 413, + 344 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{self} \\to \\mathcal{L}_{prompt}" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 402, + 284, + 442, + 328 + ], + "blocks": [ + { + "bbox": [ + 402, + 284, + 442, + 328 + ], + "lines": [ + { + "bbox": [ + 402, + 284, + 442, + 328 + ], + "spans": [ + { + "bbox": [ + 402, + 284, + 442, + 328 + ], + "type": "image", + "image_path": "a05f686c0ccc852d2a5dab91a906236fcdcad7d871b369ef903edea534b0d07e.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 326, + 347, + 345, + 388 + ], + "blocks": [ + { + "bbox": [ + 327, + 337, + 345, + 345 + ], + "lines": [ + { + "bbox": [ + 327, + 337, + 345, + 345 + ], + "spans": [ + { + "bbox": [ + 327, + 337, + 345, + 345 + ], + "type": "text", + "content": "Test" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 326, + 347, + 345, + 388 + ], + "lines": [ + { + "bbox": [ + 326, + 347, + 345, + 388 + ], + "spans": [ + { + "bbox": [ + 326, + 347, + 345, + 388 + ], + "type": "image", + "image_path": "fed712e153f9e5774cd0cc8f4a24a59cab2db2c8c82ea713d3e05bc69d46a2de.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 345, + 347, + 414, + 388 + ], + "blocks": [ + { + "bbox": [ + 345, + 347, + 414, + 388 + ], + "lines": [ + { + "bbox": [ + 345, + 347, + 414, + 388 + ], + "spans": [ + { + "bbox": [ + 345, + 347, + 414, + 388 + ], + "type": "image", + "image_path": "f72b2d431c2fb13b0ab4e39ad1af38019d53f964edf9309d24df6ec683889042.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 414, + 354, + 443, + 377 + ], + "lines": [ + { + "bbox": [ + 414, + 354, + 443, + 377 + ], + "spans": [ + { + "bbox": [ + 414, + 354, + 443, + 377 + ], + "type": "text", + "content": "Adapted Output" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_caption" + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 376, + 388, + 392, + 399 + ], + "blocks": [ + { + "bbox": [ + 376, + 388, + 392, + 399 + ], + "lines": [ + { + "bbox": [ + 376, + 388, + 392, + 399 + ], + "spans": [ + { + "bbox": [ + 376, + 388, + 392, + 399 + ], + "type": "image", + "image_path": "88efd7ca147b209c872a8aa52987eaa8196794d516936ba32fc8fc5f217a3219.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 355, + 402, + 416, + 411 + ], + "lines": [ + { + "bbox": [ + 355, + 402, + 416, + 411 + ], + "spans": [ + { + "bbox": [ + 355, + 402, + 416, + 411 + ], + "type": "text", + "content": "(c) Our VPTTA" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 347, + 412, + 375, + 425 + ], + "lines": [ + { + "bbox": [ + 347, + 412, + 375, + 425 + ], + "spans": [ + { + "bbox": [ + 347, + 412, + 375, + 425 + ], + "type": "text", + "content": "EA:" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 381, + 414, + 425, + 425 + ], + "lines": [ + { + "bbox": [ + 381, + 414, + 425, + 425 + ], + "spans": [ + { + "bbox": [ + 381, + 414, + 425, + 425 + ], + "type": "text", + "content": "CF:" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_caption" + } + ], + "index": 28 + }, + { + "type": "image", + "bbox": [ + 462, + 285, + 479, + 300 + ], + "blocks": [ + { + "bbox": [ + 462, + 285, + 479, + 300 + ], + "lines": [ + { + "bbox": [ + 462, + 285, + 479, + 300 + ], + "spans": [ + { + "bbox": [ + 462, + 285, + 479, + 300 + ], + "type": "image", + "image_path": "eac8d5775df914bcf760ff26ff202347d72471ca16344356ea573d4ddd730952.jpg" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 479, + 289, + 519, + 298 + ], + "lines": [ + { + "bbox": [ + 479, + 289, + 519, + 298 + ], + "spans": [ + { + "bbox": [ + 479, + 289, + 519, + 298 + ], + "type": "text", + "content": "Freeze" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_caption" + } + ], + "index": 32 + }, + { + "type": "image", + "bbox": [ + 463, + 305, + 478, + 323 + ], + "blocks": [ + { + "bbox": [ + 463, + 305, + 478, + 323 + ], + "lines": [ + { + "bbox": [ + 463, + 305, + 478, + 323 + ], + "spans": [ + { + "bbox": [ + 463, + 305, + 478, + 323 + ], + "type": "image", + "image_path": "d0488d37486fb777069da662dc3fdd126d023911f2b61598462fba29c36f68af.jpg" + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 493, + 309, + 529, + 318 + ], + "lines": [ + { + "bbox": [ + 493, + 309, + 529, + 318 + ], + "spans": [ + { + "bbox": [ + 493, + 309, + 529, + 318 + ], + "type": "text", + "content": "Fine-tune" + } + ] + } + ], + "index": 36, + "angle": 0, + "type": "image_caption" + } + ], + "index": 34 + }, + { + "bbox": [ + 493, + 289, + 519, + 298 + ], + "angle": 0, + "lines": [ + { + "bbox": [ + 493, + 289, + 519, + 298 + ], + "spans": [ + { + "bbox": [ + 493, + 289, + 519, + 298 + ], + "type": "text", + "content": "Freeze" + } + ] + } + ], + "index": 35, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 463, + 328, + 477, + 346 + ], + "blocks": [ + { + "bbox": [ + 463, + 328, + 477, + 346 + ], + "lines": [ + { + "bbox": [ + 463, + 328, + 477, + 346 + ], + "spans": [ + { + "bbox": [ + 463, + 328, + 477, + 346 + ], + "type": "image", + "image_path": "c2450a86b24cea5720114978d05f0a713b50915b93014079c783837ff7f8bf24.jpg" + } + ] + } + ], + "index": 37, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 493, + 332, + 522, + 342 + ], + "lines": [ + { + "bbox": [ + 493, + 332, + 522, + 342 + ], + "spans": [ + { + "bbox": [ + 493, + 332, + 522, + 342 + ], + "type": "text", + "content": "Prompt" + } + ] + } + ], + "index": 38, + "angle": 0, + "type": "image_caption" + } + ], + "index": 37 + }, + { + "type": "image", + "bbox": [ + 462, + 354, + 479, + 367 + ], + "blocks": [ + { + "bbox": [ + 462, + 354, + 479, + 367 + ], + "lines": [ + { + "bbox": [ + 462, + 354, + 479, + 367 + ], + "spans": [ + { + "bbox": [ + 462, + 354, + 479, + 367 + ], + "type": "image", + "image_path": "40332400e9aa9af88781e712236ab1c13d8f18a18d90c69267432e81c2cd9168.jpg" + } + ] + } + ], + "index": 39, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 493, + 354, + 523, + 363 + ], + "lines": [ + { + "bbox": [ + 493, + 354, + 523, + 363 + ], + "spans": [ + { + "bbox": [ + 493, + 354, + 523, + 363 + ], + "type": "text", + "content": "Terrible" + } + ] + } + ], + "index": 40, + "angle": 0, + "type": "image_caption" + } + ], + "index": 39 + }, + { + "type": "image", + "bbox": [ + 462, + 376, + 479, + 392 + ], + "blocks": [ + { + "bbox": [ + 462, + 376, + 479, + 392 + ], + "lines": [ + { + "bbox": [ + 462, + 376, + 479, + 392 + ], + "spans": [ + { + "bbox": [ + 462, + 376, + 479, + 392 + ], + "type": "image", + "image_path": "5da86f88f4bdf7572c8ca547f9c4b4f6f98caf315180fda582195c5b5a87c355.jpg" + } + ] + } + ], + "index": 41, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 493, + 378, + 526, + 388 + ], + "lines": [ + { + "bbox": [ + 493, + 378, + 526, + 388 + ], + "spans": [ + { + "bbox": [ + 493, + 378, + 526, + 388 + ], + "type": "text", + "content": "Passable" + } + ] + } + ], + "index": 42, + "angle": 0, + "type": "image_caption" + } + ], + "index": 41 + }, + { + "type": "image", + "bbox": [ + 462, + 399, + 479, + 415 + ], + "blocks": [ + { + "bbox": [ + 462, + 399, + 479, + 415 + ], + "lines": [ + { + "bbox": [ + 462, + 399, + 479, + 415 + ], + "spans": [ + { + "bbox": [ + 462, + 399, + 479, + 415 + ], + "type": "image", + "image_path": "7d936a17d3505ddcbca639d400faa5ff0772bef947668a13d6d47f9b19f07ace.jpg" + } + ] + } + ], + "index": 43, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 493, + 402, + 516, + 411 + ], + "lines": [ + { + "bbox": [ + 493, + 402, + 516, + 411 + ], + "spans": [ + { + "bbox": [ + 493, + 402, + 516, + 411 + ], + "type": "text", + "content": "Good" + } + ] + } + ], + "index": 44, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 434, + 547, + 469 + ], + "lines": [ + { + "bbox": [ + 46, + 434, + 547, + 469 + ], + "spans": [ + { + "bbox": [ + 46, + 434, + 547, + 469 + ], + "type": "text", + "content": "Figure 1. Comparison between our VPTTA and existing solutions under the CTTA setup. Our VPTTA avoids both error accumulation (EA) and catastrophic forgetting (CF) by freezing model parameters and achieves adaptation by training a prompt for each test image. The commonly used self-supervised loss and the optimized one are denoted by " + }, + { + "bbox": [ + 46, + 434, + 547, + 469 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{self}" + }, + { + "bbox": [ + 46, + 434, + 547, + 469 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 434, + 547, + 469 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{self}^{\\prime}" + }, + { + "bbox": [ + 46, + 434, + 547, + 469 + ], + "type": "text", + "content": ", respectively." + } + ] + } + ], + "index": 45, + "angle": 0, + "type": "image_caption" + } + ], + "index": 43 + }, + { + "bbox": [ + 143, + 479, + 192, + 491 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 479, + 192, + 491 + ], + "spans": [ + { + "bbox": [ + 143, + 479, + 192, + 491 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 46 + }, + { + "bbox": [ + 46, + 510, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 510, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 510, + 289, + 715 + ], + "type": "text", + "content": "Distribution shift widely exists in medical images acquired from different medical centres and poses a significant obstacle to deploying the pre-trained semantic segmentation model in real-world applications. Test-time adaptation has proven its effectiveness in tackling the cross-domain distribution shift during inference. However, most existing methods achieve adaptation by updating the pre-trained models, rendering them susceptible to error accumulation and catastrophic forgetting when encountering a series of distribution shifts (i.e., under the continual test-time adaptation setup). To overcome these challenges caused by updating the models, in this paper, we freeze the pre-trained model and propose the Visual Prompt-based Test-Time Adaptation (VPTTA) method to train a specific prompt for each test image to align the statistics in the batch normalization layers. Specifically, we present the low-frequency prompt, which is lightweight with only a few parameters" + } + ] + } + ], + "index": 47 + }, + { + "bbox": [ + 304, + 479, + 547, + 612 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 479, + 547, + 612 + ], + "spans": [ + { + "bbox": [ + 304, + 479, + 547, + 612 + ], + "type": "text", + "content": "and can be effectively trained in a single iteration. To enhance prompt initialization, we equip VPTTA with a memory bank to benefit the current prompt from previous ones. Additionally, we design a warm-up mechanism, which mixes source and target statistics to construct warm-up statistics, thereby facilitating the training process. Extensive experiments demonstrate the superiority of our VPTTA over other state-of-the-art methods on two medical image segmentation benchmark tasks. The code and weights of pre-trained source models are available at https://github.com/Chen-Ziyang/VPTTA." + } + ] + } + ], + "index": 48 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 495, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 495, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 495, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 304, + 645, + 547, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 645, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 645, + 547, + 713 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 304, + 645, + 547, + 713 + ], + "type": "text", + "content": "Yong Xia and Yongsheng Pan are the corresponding authors. This work was supported in part by the National Natural Science Foundation of China under Grants 62171377, in part by STI2030-Major Project under Grants 2022ZD0213100, in part by Shenzhen Science and Technology Program under Grants JCYJ20220530161616036, and in part by the Ningbo Clinical Research Center for Medical Imaging under Grant 2021L003 (Open Project 2022LYKFZD06)." + } + ] + } + ], + "index": 49 + }, + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "11184" + } + ] + } + ], + "index": 50 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 127, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 127, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 127, + 83 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 91, + 289, + 331 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 91, + 289, + 331 + ], + "spans": [ + { + "bbox": [ + 47, + 91, + 289, + 331 + ], + "type": "text", + "content": "Semantic segmentation serves a pivotal role in medical image processing to outline specific anatomical structures. However, the segmentation model pre-trained on the source dataset may be hindered from deploying in real-world applications due to the commonly existing distribution shift, which is typically caused by the variations in imaging protocols, operators, and scanners [16, 34]. Test-time adaptation (TTA) has emerged as a promising paradigm to replace domain adaptation (DA) [5, 43, 46, 53], as TTA requires only test data during the inference phase, while DA necessitates a sufficiently large and representative target dataset [45]. The mainstream TTA methods [29-31, 37, 40, 41, 47, 48] leverage the self-supervised tasks to construct auxiliary losses, such as entropy minimization [41] and self-training [37], to update the parameters of pre-trained models. Although these TTA methods successfully alleviate the domain gap and performance degradation, they are based on the assumption that the target domains exhibit static distributions. Unfortunately, the target domains are continuously changing rather than static in most real-world scenarios." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 331, + 289, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 331, + 289, + 617 + ], + "spans": [ + { + "bbox": [ + 47, + 331, + 289, + 617 + ], + "type": "text", + "content": "This was first presented as the continual test-time adaptation (CTTA) setup [42] that addresses TTA on a series of distribution shifts. Under this setup, pre-trained models are required to adapt to several distinct domains and therefore are more prone to suffer error accumulation and catastrophic forgetting. As mentioned above, many TTA methods rely on self-supervised tasks and may be affected by the noisy losses due to unreliable supervision, leading to error accumulation [2, 23]. Catastrophic forgetting typically occurs when the model is continually trained on new domains, resulting in unexpected performance degradation [21, 30]. Some research has been studied in addressing these two challenges based on optimizing the loss functions for better supervision [6, 31, 42, 42, 45] (see Figure 1 (a)) or prevent the model parameters from excessive alterations by model resetting or introducing regularizers [30, 31, 39, 42] (see Figure 1 (b)). Despite their efforts, loss optimization cannot ensure a robust training process to avoid catastrophic forgetting during the long-term inference, insufficient or excessive updating of the model parameters may lead to inadequate adaptation or error accumulation. Since the pretrained model is updated by the self-supervised loss where the noise cannot be completely removed, these methods may still suffer from these two issues." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 618, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 618, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 47, + 618, + 288, + 715 + ], + "type": "text", + "content": "Considering the above limitations in CTTA, updating the pre-trained model during inference seems to be inappropriate. To this end, we propose the Visual Prompt-based Test-Time Adaptation (VPTTA) method to avoid error accumulation and catastrophic forgetting associated with updating model by freezing the model parameters and conduct adaptation by learning a specific visual prompt for each test image (see Figure 1 (c)). The prompt can adapt each test im" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 304, + 72, + 547, + 109 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 109 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 109 + ], + "type": "text", + "content": "age to the frozen pre-trained model to take into account both adaptation ability and knowledge retaining. We attempt to achieve this by answering the following three questions." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 306, + 110, + 547, + 254 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 110, + 547, + 254 + ], + "spans": [ + { + "bbox": [ + 306, + 110, + 547, + 254 + ], + "type": "text", + "content": "- How to design the prompt? Each prompt is restricted to a limited number of training iterations (sometimes just one) to satisfy the demands of online inference, requiring the prompt to be lightweight. Based on this principle, we introduce the prompt for only the low-frequency components of each sample, resulting in the low-frequency prompt. Since the low-frequency components are strongly associated with style textures and can serve as the primary source of distribution shift [46], our low-frequency prompt allows for maintaining lightweight and being able to reduce distribution shifts by modifying the low-frequency components." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 254, + 547, + 385 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 254, + 547, + 385 + ], + "spans": [ + { + "bbox": [ + 306, + 254, + 547, + 385 + ], + "type": "text", + "content": "- How to initialize the prompt? It is well known that model training can benefit from proper parameter initialization, resulting in not only fewer training iterations but also improved performance [3, 52]. We argue that this is also applicable to prompt training and therefore construct a memory bank to store the prompts and low-frequency components of previous test images. By comparing the similarity between the low-frequency components of the current and previous images, we select several historical prompts of the most similar images to initialize the current prompt." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 385, + 547, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 385, + 547, + 518 + ], + "spans": [ + { + "bbox": [ + 306, + 385, + 547, + 518 + ], + "type": "text", + "content": "- How to train the prompt? Since the statistics mismatch between the source and target domains in the batch normalization (BN) layers is a significant cause of distribution shift [33, 35, 44], we optimize the prompt with the objective of minimizing the distances between BN statistics and the statistics extracted from the features of test data for statistics alignment [19]. To further address the training difficulty posed by the large distribution shift at the beginning of the inference phase, we devise a statistics-fusion-based warm-up mechanism to assist in learning the prompt from easy to hard." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 519, + 547, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 519, + 547, + 640 + ], + "spans": [ + { + "bbox": [ + 304, + 519, + 547, + 640 + ], + "type": "text", + "content": "To summarise, our contributions are three-fold: (1) We present the low-frequency prompt to adjust the test image to adapt the model instead of updating the model and therefore prevent the model from error accumulation and catastrophic forgetting. (2) We propose the novel VPTTA framework to alleviate the distribution shift by training a sample-specific visual prompt, which is lightweight, well-initialized, and follows a gradual learning process. (3) The powerful framework we provide can be effectively deployed, requiring only a single iteration and a few dozen parameters for adaptation." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 651, + 392, + 663 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 651, + 392, + 663 + ], + "spans": [ + { + "bbox": [ + 306, + 651, + 392, + 663 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 670, + 460, + 684 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 670, + 460, + 684 + ], + "spans": [ + { + "bbox": [ + 306, + 670, + 460, + 684 + ], + "type": "text", + "content": "2.1. Test-time Adaptation (TTA)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 689, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 547, + 713 + ], + "type": "text", + "content": "TTA aims to adapt the model pre-trained on the source domain to the test data during inference in a source-free" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "11185" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 299 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 299 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 299 + ], + "type": "text", + "content": "and online manner [40, 41]. The mainstream TTA methods are based on model updating, which constructs self-supervised auxiliary tasks to guide the model to learn on the test data [4, 13, 26, 29, 37, 40, 41, 48, 49]. Sun et al. [40] utilized an auxiliary task to predict the rotation to assist the model learning on the test data. Wang et al. [41] presented a test-time entropy minimization scheme to reduce generalization error by reducing the entropy of model predictions on test data. Zhang et al. [49] introduced a supplementary network to fit the test data, which can work in a plug-and-play fashion, necessitating minimal hyperparameter tuning. Some other methods are based on BN statistics, which alleviate the distribution shifts by modifying the statistics within BN layers [25, 27, 44]. Mirza et al. [25] constantly adapted the statistics of the BN layers to modify the feature representations of the model. These TTA methods improve the adaptation performance on the static target domain but neglect the continually changing target domains in most real-world scenarios." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 300, + 289, + 564 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 300, + 289, + 564 + ], + "spans": [ + { + "bbox": [ + 46, + 300, + 289, + 564 + ], + "type": "text", + "content": "Recently, the continual test-time adaptation (CTTA) [42] is proposed to tackle TTA under a series of continually changing target domains. Due to long-term inference, both error accumulation [2, 23] and catastrophic forgetting [21, 30] are more prone to occur in this setup. Many methods [6, 30, 31, 39, 42, 45, 47] optimize the self-supervised losses or prevent excessive alterations to the model parameters via model resetting or regularizer to improve this situation but are still impacted by these issues. Wang et al. [42] refined pseudo labels by the weight- and augmentation-averaged predictions and randomly restored a small part of model parameters. Niu et al. [31] removed partial noisy samples with large gradients and encouraged model weights toward a flat minimum. Yang et al. [45] designed a dynamic strategy to adjust the learning rate for each test image. Zhang et al. [47] presented to adaptively fuse source and test statistics and improve the entropy minimization loss to a generalized one. In contrast, our proposed VPTTA employs a learnable prompt for each test image to mitigate distribution shifts during inference, requiring no changes to the pre-trained model and therefore alleviating the error accumulation and catastrophic forgetting." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 574, + 151, + 588 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 574, + 151, + 588 + ], + "spans": [ + { + "bbox": [ + 47, + 574, + 151, + 588 + ], + "type": "text", + "content": "2.2. Prompt Learning" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 594, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 289, + 715 + ], + "type": "text", + "content": "Prompt learning was initially employed to devise additional text instructions for input text to fine-tune the largescale Natural Language Processing (NLP) models on downstream tasks [24]. With remarkable success in NLP, it has also been involved in computer vision and greatly benefited visual models in recent years [12, 15, 19, 51]. Jia et al. [15] introduced a small number of trainable parameters in the input space while keeping the model backbone frozen to fine-tune the model with high quality and efficiency. Zhou et al. [51] presented the conditional context optimization" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 304, + 72, + 547, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 240 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 240 + ], + "type": "text", + "content": "to generate an input-conditional token for each image to generalize the pre-trained vision-language model to unseen classes. Hu et al. [19] employed the visual prompt at the pixel level and added it to the input images to explicitly minimize the domain discrepancy. Gan et al. [12] learned a domain-specific and domain-agnostic visual prompt for adaptation by using the exponential-moving-average strategy under the guidance of self-training. Different from them, our VPTTA presents the low-frequency prompt to address a series of changing distribution shifts with a few dozen parameters. We further consider optimizing the initialization and warm-up mechanism, enabling our visual prompts to be effectively trained within a single iteration to satisfy the demands of CTTA." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 305, + 249, + 389, + 263 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 249, + 389, + 263 + ], + "spans": [ + { + "bbox": [ + 305, + 249, + 389, + 263 + ], + "type": "text", + "content": "3. Methodology" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 305, + 269, + 525, + 282 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 269, + 525, + 282 + ], + "spans": [ + { + "bbox": [ + 305, + 269, + 525, + 282 + ], + "type": "text", + "content": "3.1. Problem Definition and Method Overview" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 287, + 545, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 287, + 545, + 373 + ], + "spans": [ + { + "bbox": [ + 304, + 287, + 545, + 373 + ], + "type": "text", + "content": "Let the labeled source domain dataset and unlabeled testing target dataset be " + }, + { + "bbox": [ + 304, + 287, + 545, + 373 + ], + "type": "inline_equation", + "content": "\\mathbb{D}^s = \\{X_i^s,Y_i^s\\}_{i = 1}^{N^s}" + }, + { + "bbox": [ + 304, + 287, + 545, + 373 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 287, + 545, + 373 + ], + "type": "inline_equation", + "content": "\\mathbb{D}^t = \\{X_i^t\\}_{i = 1}^{N^t}" + }, + { + "bbox": [ + 304, + 287, + 545, + 373 + ], + "type": "text", + "content": ", respectively, where " + }, + { + "bbox": [ + 304, + 287, + 545, + 373 + ], + "type": "inline_equation", + "content": "X_{i}^{*}\\in \\mathbb{R}^{H\\times W\\times C}" + }, + { + "bbox": [ + 304, + 287, + 545, + 373 + ], + "type": "text", + "content": " is the " + }, + { + "bbox": [ + 304, + 287, + 545, + 373 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 287, + 545, + 373 + ], + "type": "text", + "content": "-th image with C channels and size of " + }, + { + "bbox": [ + 304, + 287, + 545, + 373 + ], + "type": "inline_equation", + "content": "H\\times W" + }, + { + "bbox": [ + 304, + 287, + 545, + 373 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 287, + 545, + 373 + ], + "type": "inline_equation", + "content": "Y_{i}^{*}" + }, + { + "bbox": [ + 304, + 287, + 545, + 373 + ], + "type": "text", + "content": " is its label. Our goal is to learn a low-frequency prompt " + }, + { + "bbox": [ + 304, + 287, + 545, + 373 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_i" + }, + { + "bbox": [ + 304, + 287, + 545, + 373 + ], + "type": "text", + "content": " for each test image " + }, + { + "bbox": [ + 304, + 287, + 545, + 373 + ], + "type": "inline_equation", + "content": "X_{i}^{t}" + }, + { + "bbox": [ + 304, + 287, + 545, + 373 + ], + "type": "text", + "content": " to obtain the modified image " + }, + { + "bbox": [ + 304, + 287, + 545, + 373 + ], + "type": "inline_equation", + "content": "\\tilde{X}_i^t" + }, + { + "bbox": [ + 304, + 287, + 545, + 373 + ], + "type": "text", + "content": " that can be adapted to the model " + }, + { + "bbox": [ + 304, + 287, + 545, + 373 + ], + "type": "inline_equation", + "content": "f_{\\theta}:X\\to Y" + }, + { + "bbox": [ + 304, + 287, + 545, + 373 + ], + "type": "text", + "content": " pre-trained on " + }, + { + "bbox": [ + 304, + 287, + 545, + 373 + ], + "type": "inline_equation", + "content": "\\mathbb{D}^s" + }, + { + "bbox": [ + 304, + 287, + 545, + 373 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 374, + 547, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 374, + 547, + 518 + ], + "spans": [ + { + "bbox": [ + 304, + 374, + 547, + 518 + ], + "type": "text", + "content": "To accomplish this goal, our VPTTA considers three key facets: prompt design, prompt initialization, and prompt training. For prompt design, the prompt is designed to change the low-frequency component of the amplitude of the test image. For prompt initialization, we adopt a memory bank to initialize the current prompt with previous ones. For prompt training, we employ the absolute distance to train the prompt to align the statistics between features extracted from test images and the batch normalization (BN) layers and also present a statistics-fusion-based warm-up mechanism. An overview of our VPTTA is illustrated in Figure 2. We now delve into its details." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 524, + 516, + 537 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 524, + 516, + 537 + ], + "spans": [ + { + "bbox": [ + 305, + 524, + 516, + 537 + ], + "type": "text", + "content": "3.2. Prompt Design: Low-frequency Prompt" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 542, + 545, + 627 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 542, + 545, + 627 + ], + "spans": [ + { + "bbox": [ + 304, + 542, + 545, + 627 + ], + "type": "text", + "content": "Let " + }, + { + "bbox": [ + 304, + 542, + 545, + 627 + ], + "type": "inline_equation", + "content": "\\mathcal{F}(\\cdot)" + }, + { + "bbox": [ + 304, + 542, + 545, + 627 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 542, + 545, + 627 + ], + "type": "inline_equation", + "content": "\\mathcal{F}^{-1}(\\cdot)" + }, + { + "bbox": [ + 304, + 542, + 545, + 627 + ], + "type": "text", + "content": " be the Fast Fourier Transform (FFT) and Inverse Fast Fourier Transform (IFFT) operations [10], respectively. The amplitude component and phase component are denoted by " + }, + { + "bbox": [ + 304, + 542, + 545, + 627 + ], + "type": "inline_equation", + "content": "\\mathcal{F}^A (\\cdot)" + }, + { + "bbox": [ + 304, + 542, + 545, + 627 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 542, + 545, + 627 + ], + "type": "inline_equation", + "content": "\\mathcal{F}^P (\\cdot)" + }, + { + "bbox": [ + 304, + 542, + 545, + 627 + ], + "type": "text", + "content": " respectively. Note that the low-frequency component in " + }, + { + "bbox": [ + 304, + 542, + 545, + 627 + ], + "type": "inline_equation", + "content": "\\mathcal{F}^A (\\cdot)" + }, + { + "bbox": [ + 304, + 542, + 545, + 627 + ], + "type": "text", + "content": " is shifted to the center. Utilizing the prompt " + }, + { + "bbox": [ + 304, + 542, + 545, + 627 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_i\\in \\mathbb{R}^{(\\alpha \\times H)\\times (\\alpha \\times W)\\times C}" + }, + { + "bbox": [ + 304, + 542, + 545, + 627 + ], + "type": "text", + "content": " to obtain the adapted image " + }, + { + "bbox": [ + 304, + 542, + 545, + 627 + ], + "type": "inline_equation", + "content": "\\bar{X}_i^t" + }, + { + "bbox": [ + 304, + 542, + 545, + 627 + ], + "type": "text", + "content": " can be formalized as" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 315, + 632, + 545, + 647 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 632, + 545, + 647 + ], + "spans": [ + { + "bbox": [ + 315, + 632, + 545, + 647 + ], + "type": "interline_equation", + "content": "\\tilde {X} _ {i} ^ {t} = \\mathcal {F} ^ {- 1} ([ \\text {O n e P a d} (\\mathcal {P} _ {i}) \\odot \\mathcal {F} ^ {A} (X _ {i} ^ {t}), \\mathcal {F} ^ {P} (X _ {i} ^ {t}) ]) \\tag {1}", + "image_path": "8e61f343f709ab62dba94d0d93641edebc3481c76f2e0e42c360bda68db1de87.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 653, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 653, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 653, + 547, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 653, + 547, + 714 + ], + "type": "inline_equation", + "content": "\\odot" + }, + { + "bbox": [ + 304, + 653, + 547, + 714 + ], + "type": "text", + "content": " is the element-wise multiplication, and OnePad means padding one around " + }, + { + "bbox": [ + 304, + 653, + 547, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_i" + }, + { + "bbox": [ + 304, + 653, + 547, + 714 + ], + "type": "text", + "content": " to expand it to the size of " + }, + { + "bbox": [ + 304, + 653, + 547, + 714 + ], + "type": "inline_equation", + "content": "H\\times W" + }, + { + "bbox": [ + 304, + 653, + 547, + 714 + ], + "type": "text", + "content": ". Under the control of the small coefficient " + }, + { + "bbox": [ + 304, + 653, + 547, + 714 + ], + "type": "inline_equation", + "content": "\\alpha \\in (0,1)" + }, + { + "bbox": [ + 304, + 653, + 547, + 714 + ], + "type": "text", + "content": ", we can ensure that the prompt " + }, + { + "bbox": [ + 304, + 653, + 547, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_i" + }, + { + "bbox": [ + 304, + 653, + 547, + 714 + ], + "type": "text", + "content": " remains lightweight and focuses on the low-frequency component." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "11186" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 75, + 70, + 521, + 322 + ], + "blocks": [ + { + "bbox": [ + 75, + 70, + 521, + 322 + ], + "lines": [ + { + "bbox": [ + 75, + 70, + 521, + 322 + ], + "spans": [ + { + "bbox": [ + 75, + 70, + 521, + 322 + ], + "type": "image", + "image_path": "62cb4d2a81e70ceeeda78d942341260a70bc8d077afc8e161d3cbac4508c6ad2.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 327, + 548, + 406 + ], + "lines": [ + { + "bbox": [ + 46, + 327, + 548, + 406 + ], + "spans": [ + { + "bbox": [ + 46, + 327, + 548, + 406 + ], + "type": "text", + "content": "Figure 2. Overview of our VPTTA. For each test image, (1) the Fast Fourier Transform (FFT) is first applied to transform it into the frequency domain, where the low-frequency component of amplitude is used to query in a memory bank to initialize the current prompt, and then the amplitude is multiplied with the prompt at the low-frequency component and transformed back to the spatial domain using the Inverse Fast Fourier Transform (IFFT). (2) The memory bank is built on the previous low-frequency components and their corresponding prompts and is updated using the First In First Out (FIFO) strategy. (3) We convert the source statistics stored in BN layers into the warm-up statistics and calculate the absolute distance between warm-up and target statistics as the loss to fine-tune the prompt. (4) Finally, we feed the image corrected by its fine-tuned prompt to the pre-trained model to produce the output. 'Stat': Abbreviation of 'Statistics'." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 426, + 289, + 522 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 426, + 289, + 522 + ], + "spans": [ + { + "bbox": [ + 47, + 426, + 289, + 522 + ], + "type": "text", + "content": "Simultaneously, we recognize that the low-rank decomposition is widely employed in the field of data compression [22] and model fine-tuning [18]. Consequently, we also design a low-rank prompt " + }, + { + "bbox": [ + 47, + 426, + 289, + 522 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_i^{LoRa} = \\mathcal{B}_i@A_i" + }, + { + "bbox": [ + 47, + 426, + 289, + 522 + ], + "type": "text", + "content": " as a variant (denoted as VPTTA-LR), where " + }, + { + "bbox": [ + 47, + 426, + 289, + 522 + ], + "type": "inline_equation", + "content": "\\mathcal{B}_i\\in \\mathbb{R}^{H\\times r\\times C}(r\\ll W)" + }, + { + "bbox": [ + 47, + 426, + 289, + 522 + ], + "type": "inline_equation", + "content": "\\mathcal{A}_i\\in \\mathbb{R}^{r\\times W\\times C}(r\\ll H)" + }, + { + "bbox": [ + 47, + 426, + 289, + 522 + ], + "type": "text", + "content": ", and @ denotes the matrix multiplication operation. In this way, the adapted image can be formalized as " + }, + { + "bbox": [ + 47, + 426, + 289, + 522 + ], + "type": "inline_equation", + "content": "\\tilde{X}_i^t = X_i^t +\\mathcal{P}_i^{LoRa}" + }, + { + "bbox": [ + 47, + 426, + 289, + 522 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 536, + 244, + 550 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 536, + 244, + 550 + ], + "spans": [ + { + "bbox": [ + 47, + 536, + 244, + 550 + ], + "type": "text", + "content": "3.3. Prompt Initialization: Memory Bank" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 557, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 557, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 557, + 289, + 715 + ], + "type": "text", + "content": "To achieve efficient training, it is critical to produce a suitable initialization for " + }, + { + "bbox": [ + 46, + 557, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_i" + }, + { + "bbox": [ + 46, + 557, + 289, + 715 + ], + "type": "text", + "content": ". Therefore, we present a memory-bank-based initialization strategy to benefit the current prompt from previous ones. Specifically, the memory bank " + }, + { + "bbox": [ + 46, + 557, + 289, + 715 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 46, + 557, + 289, + 715 + ], + "type": "text", + "content": " stores " + }, + { + "bbox": [ + 46, + 557, + 289, + 715 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 46, + 557, + 289, + 715 + ], + "type": "text", + "content": " pairs of keys and values " + }, + { + "bbox": [ + 46, + 557, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\{q_s,v_s\\}_{s = 1}^S" + }, + { + "bbox": [ + 46, + 557, + 289, + 715 + ], + "type": "text", + "content": " and holds the First In First Out (FIFO) principle. As shown in Figure 2, the keys are the low-frequency components of previous test images, and the values are their corresponding prompts. Note that when using VPTTA-LR, the keys are the low-rank reconstruction matrix of previous test images. To initialize the current prompt " + }, + { + "bbox": [ + 46, + 557, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_i" + }, + { + "bbox": [ + 46, + 557, + 289, + 715 + ], + "type": "text", + "content": ", we first extract the low-frequency component of the current test image " + }, + { + "bbox": [ + 46, + 557, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{low}^{A}(X_{i}^{t})" + }, + { + "bbox": [ + 46, + 557, + 289, + 715 + ], + "type": "text", + "content": " and calculate the cosine similarity between " + }, + { + "bbox": [ + 46, + 557, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{low}^{A}(X_{i}^{t})" + }, + { + "bbox": [ + 46, + 557, + 289, + 715 + ], + "type": "text", + "content": " and" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 306, + 426, + 402, + 438 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 426, + 402, + 438 + ], + "spans": [ + { + "bbox": [ + 306, + 426, + 402, + 438 + ], + "type": "text", + "content": "each key " + }, + { + "bbox": [ + 306, + 426, + 402, + 438 + ], + "type": "inline_equation", + "content": "q_{s}" + }, + { + "bbox": [ + 306, + 426, + 402, + 438 + ], + "type": "text", + "content": " as follows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 335, + 441, + 545, + 470 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 441, + 545, + 470 + ], + "spans": [ + { + "bbox": [ + 335, + 441, + 545, + 470 + ], + "type": "interline_equation", + "content": "C o s \\left(\\mathcal {F} _ {l o w} ^ {A} \\left(X _ {i} ^ {t}\\right), q _ {s}\\right) = \\frac {\\left\\langle \\mathcal {F} _ {l o w} ^ {A} \\left(X _ {i} ^ {t}\\right) , q _ {s} \\right\\rangle}{\\left\\| \\mathcal {F} _ {l o w} ^ {A} \\left(X _ {i} ^ {t}\\right) \\right\\| * \\left\\| q _ {s} \\right\\|}, \\tag {2}", + "image_path": "31ed45b5c5b2fc372dc458a9613ed5e72dbf933192fabb09ccff6e1c923f5237.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 473, + 545, + 584 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 473, + 545, + 584 + ], + "spans": [ + { + "bbox": [ + 305, + 473, + 545, + 584 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 473, + 545, + 584 + ], + "type": "inline_equation", + "content": "\\langle \\cdot ,\\cdot \\rangle" + }, + { + "bbox": [ + 305, + 473, + 545, + 584 + ], + "type": "text", + "content": " denotes the dot product operation, and " + }, + { + "bbox": [ + 305, + 473, + 545, + 584 + ], + "type": "inline_equation", + "content": "\\| \\cdot \\|" + }, + { + "bbox": [ + 305, + 473, + 545, + 584 + ], + "type": "text", + "content": " denotes the Euclidean norm. Since the low-frequency component can represent the style texture of an image, we rank the similarity scores and then retrieve the values of the " + }, + { + "bbox": [ + 305, + 473, + 545, + 584 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 305, + 473, + 545, + 584 + ], + "type": "text", + "content": " most relevant keys from " + }, + { + "bbox": [ + 305, + 473, + 545, + 584 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 305, + 473, + 545, + 584 + ], + "type": "text", + "content": " to construct a support set " + }, + { + "bbox": [ + 305, + 473, + 545, + 584 + ], + "type": "inline_equation", + "content": "R_{i} = \\{q_{k},v_{k}\\}_{k = 1}^{K}" + }, + { + "bbox": [ + 305, + 473, + 545, + 584 + ], + "type": "text", + "content": " for the current test image. Finally, we assign a weight " + }, + { + "bbox": [ + 305, + 473, + 545, + 584 + ], + "type": "inline_equation", + "content": "w_{k}" + }, + { + "bbox": [ + 305, + 473, + 545, + 584 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 305, + 473, + 545, + 584 + ], + "type": "inline_equation", + "content": "v_{k}" + }, + { + "bbox": [ + 305, + 473, + 545, + 584 + ], + "type": "text", + "content": " to initialize the current prompt by " + }, + { + "bbox": [ + 305, + 473, + 545, + 584 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_i = \\sum_{k = 1}^K w_k v_k" + }, + { + "bbox": [ + 305, + 473, + 545, + 584 + ], + "type": "text", + "content": " where the weights " + }, + { + "bbox": [ + 305, + 473, + 545, + 584 + ], + "type": "inline_equation", + "content": "\\{w_{k}\\}" + }, + { + "bbox": [ + 305, + 473, + 545, + 584 + ], + "type": "text", + "content": " are calculated based on the similarity scores [45] and " + }, + { + "bbox": [ + 305, + 473, + 545, + 584 + ], + "type": "inline_equation", + "content": "\\sum_{k = 1}^{K}w_{k} = 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 588, + 510, + 601 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 588, + 510, + 601 + ], + "spans": [ + { + "bbox": [ + 306, + 588, + 510, + 601 + ], + "type": "text", + "content": "3.4. Prompt Training: Statistics Alignment" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 606, + 545, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 606, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 304, + 606, + 545, + 689 + ], + "type": "text", + "content": "Inspired by [19, 27, 47], the distribution shift largely exists in distinct BN statistics calculated on the test features (i.e., " + }, + { + "bbox": [ + 304, + 606, + 545, + 689 + ], + "type": "inline_equation", + "content": "\\mu_t" + }, + { + "bbox": [ + 304, + 606, + 545, + 689 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 606, + 545, + 689 + ], + "type": "inline_equation", + "content": "\\sigma_t" + }, + { + "bbox": [ + 304, + 606, + 545, + 689 + ], + "type": "text", + "content": ") and stored in the source model " + }, + { + "bbox": [ + 304, + 606, + 545, + 689 + ], + "type": "inline_equation", + "content": "f_{\\theta}" + }, + { + "bbox": [ + 304, + 606, + 545, + 689 + ], + "type": "text", + "content": " (i.e., " + }, + { + "bbox": [ + 304, + 606, + 545, + 689 + ], + "type": "inline_equation", + "content": "\\mu_s" + }, + { + "bbox": [ + 304, + 606, + 545, + 689 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 606, + 545, + 689 + ], + "type": "inline_equation", + "content": "\\sigma_s" + }, + { + "bbox": [ + 304, + 606, + 545, + 689 + ], + "type": "text", + "content": "), where " + }, + { + "bbox": [ + 304, + 606, + 545, + 689 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 304, + 606, + 545, + 689 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 606, + 545, + 689 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 304, + 606, + 545, + 689 + ], + "type": "text", + "content": " respectively denote the mean and standard deviation. To train the prompts to align the statistics, we adopt the absolute-distance-based loss " + }, + { + "bbox": [ + 304, + 606, + 545, + 689 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_p" + }, + { + "bbox": [ + 304, + 606, + 545, + 689 + ], + "type": "text", + "content": " [19] as follows:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 353, + 692, + 545, + 717 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 353, + 692, + 545, + 717 + ], + "spans": [ + { + "bbox": [ + 353, + 692, + 545, + 717 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {p} = \\frac {1}{j} \\sum \\left| \\mu_ {s} ^ {j} - \\mu_ {t} ^ {j} \\right| + \\left| \\sigma_ {s} ^ {j} - \\sigma_ {t} ^ {j} \\right|, \\tag {3}", + "image_path": "610214c1da3460427f19f1c70caca822335f64c569ce08836180c221449d6110.jpg" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "11187" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "code", + "bbox": [ + 60, + 89, + 254, + 121 + ], + "blocks": [ + { + "bbox": [ + 53, + 76, + 188, + 87 + ], + "lines": [ + { + "bbox": [ + 53, + 76, + 188, + 87 + ], + "spans": [ + { + "bbox": [ + 53, + 76, + 188, + 87 + ], + "type": "text", + "content": "Algorithm 1: VPTTA algorithm." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "code_caption" + }, + { + "bbox": [ + 60, + 89, + 254, + 121 + ], + "lines": [ + { + "bbox": [ + 60, + 89, + 254, + 121 + ], + "spans": [ + { + "bbox": [ + 60, + 89, + 254, + 121 + ], + "type": "text", + "content": "Initialize: Frozen source model " + }, + { + "bbox": [ + 60, + 89, + 254, + 121 + ], + "type": "inline_equation", + "content": "f_{\\theta}" + }, + { + "bbox": [ + 60, + 89, + 254, + 121 + ], + "type": "text", + "content": ", initial prompt " + }, + { + "bbox": [ + 60, + 89, + 254, + 121 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 60, + 89, + 254, + 121 + ], + "type": "text", + "content": " filled with ones, memory bank " + }, + { + "bbox": [ + 60, + 89, + 254, + 121 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 60, + 89, + 254, + 121 + ], + "type": "text", + "content": ", hyperparameters " + }, + { + "bbox": [ + 60, + 89, + 254, + 121 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 60, + 89, + 254, + 121 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 60, + 89, + 254, + 121 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 60, + 89, + 254, + 121 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 60, + 89, + 254, + 121 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 60, + 89, + 254, + 121 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "code_body" + } + ], + "index": 1, + "sub_type": "algorithm" + }, + { + "bbox": [ + 60, + 122, + 245, + 132 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 60, + 122, + 245, + 132 + ], + "spans": [ + { + "bbox": [ + 60, + 122, + 245, + 132 + ], + "type": "text", + "content": "Input: For each time step " + }, + { + "bbox": [ + 60, + 122, + 245, + 132 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 60, + 122, + 245, + 132 + ], + "type": "text", + "content": ", current test image " + }, + { + "bbox": [ + 60, + 122, + 245, + 132 + ], + "type": "inline_equation", + "content": "X_{i}^{t}" + }, + { + "bbox": [ + 60, + 122, + 245, + 132 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 133, + 296, + 350 + ], + "type": "list", + "angle": 0, + "index": 22, + "blocks": [ + { + "bbox": [ + 66, + 133, + 178, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 133, + 178, + 144 + ], + "spans": [ + { + "bbox": [ + 66, + 133, + 178, + 144 + ], + "type": "text", + "content": "1: ▷ Initialize the prompt " + }, + { + "bbox": [ + 66, + 133, + 178, + 144 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_i" + }, + { + "bbox": [ + 66, + 133, + 178, + 144 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 65, + 144, + 128, + 154 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 144, + 128, + 154 + ], + "spans": [ + { + "bbox": [ + 65, + 144, + 128, + 154 + ], + "type": "text", + "content": "2: if " + }, + { + "bbox": [ + 65, + 144, + 128, + 154 + ], + "type": "inline_equation", + "content": "i < K" + }, + { + "bbox": [ + 65, + 144, + 128, + 154 + ], + "type": "text", + "content": " then" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 65, + 155, + 116, + 165 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 155, + 116, + 165 + ], + "spans": [ + { + "bbox": [ + 65, + 155, + 116, + 165 + ], + "type": "text", + "content": "3: " + }, + { + "bbox": [ + 65, + 155, + 116, + 165 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_i = \\mathcal{P}" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 65, + 167, + 92, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 167, + 92, + 175 + ], + "spans": [ + { + "bbox": [ + 65, + 167, + 92, + 175 + ], + "type": "text", + "content": "4: else" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 65, + 176, + 220, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 176, + 220, + 186 + ], + "spans": [ + { + "bbox": [ + 65, + 176, + 220, + 186 + ], + "type": "text", + "content": "5: Construct the support set " + }, + { + "bbox": [ + 65, + 176, + 220, + 186 + ], + "type": "inline_equation", + "content": "R_{i}" + }, + { + "bbox": [ + 65, + 176, + 220, + 186 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 65, + 176, + 220, + 186 + ], + "type": "inline_equation", + "content": "M" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 65, + 187, + 188, + 198 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 187, + 188, + 198 + ], + "spans": [ + { + "bbox": [ + 65, + 187, + 188, + 198 + ], + "type": "text", + "content": "6: " + }, + { + "bbox": [ + 65, + 187, + 188, + 198 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_i = \\sum_{k=1}^K w_k v_k, v_k \\in R_i" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 65, + 200, + 99, + 208 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 200, + 99, + 208 + ], + "spans": [ + { + "bbox": [ + 65, + 200, + 99, + 208 + ], + "type": "text", + "content": "7: end if" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 65, + 210, + 165, + 220 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 210, + 165, + 220 + ], + "spans": [ + { + "bbox": [ + 65, + 210, + 165, + 220 + ], + "type": "text", + "content": "8: " + }, + { + "bbox": [ + 65, + 210, + 165, + 220 + ], + "type": "inline_equation", + "content": "\\triangleright" + }, + { + "bbox": [ + 65, + 210, + 165, + 220 + ], + "type": "text", + "content": " Train the prompt " + }, + { + "bbox": [ + 65, + 210, + 165, + 220 + ], + "type": "inline_equation", + "content": "P_{i}" + }, + { + "bbox": [ + 65, + 210, + 165, + 220 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 65, + 220, + 187, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 220, + 187, + 232 + ], + "spans": [ + { + "bbox": [ + 65, + 220, + 187, + 232 + ], + "type": "text", + "content": "9: Obtain " + }, + { + "bbox": [ + 65, + 220, + 187, + 232 + ], + "type": "inline_equation", + "content": "\\hat{X}_i^t" + }, + { + "bbox": [ + 65, + 220, + 187, + 232 + ], + "type": "text", + "content": " using " + }, + { + "bbox": [ + 65, + 220, + 187, + 232 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_i" + }, + { + "bbox": [ + 65, + 220, + 187, + 232 + ], + "type": "text", + "content": " by Eq. (1)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 62, + 232, + 296, + 253 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 232, + 296, + 253 + ], + "spans": [ + { + "bbox": [ + 62, + 232, + 296, + 253 + ], + "type": "text", + "content": "10: Forward " + }, + { + "bbox": [ + 62, + 232, + 296, + 253 + ], + "type": "inline_equation", + "content": "f_{\\theta}(X_i^t)" + }, + { + "bbox": [ + 62, + 232, + 296, + 253 + ], + "type": "text", + "content": " and calculate the warm-up statistics " + }, + { + "bbox": [ + 62, + 232, + 296, + 253 + ], + "type": "inline_equation", + "content": "\\mu_w" + }, + { + "bbox": [ + 62, + 232, + 296, + 253 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 62, + 232, + 296, + 253 + ], + "type": "inline_equation", + "content": "\\sigma_w" + }, + { + "bbox": [ + 62, + 232, + 296, + 253 + ], + "type": "text", + "content": " by Eq. (4) to replace the " + }, + { + "bbox": [ + 62, + 232, + 296, + 253 + ], + "type": "inline_equation", + "content": "\\mu_s" + }, + { + "bbox": [ + 62, + 232, + 296, + 253 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 62, + 232, + 296, + 253 + ], + "type": "inline_equation", + "content": "\\sigma_s" + }, + { + "bbox": [ + 62, + 232, + 296, + 253 + ], + "type": "text", + "content": " stored in " + }, + { + "bbox": [ + 62, + 232, + 296, + 253 + ], + "type": "inline_equation", + "content": "f_{\\theta}" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 62, + 254, + 207, + 264 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 254, + 207, + 264 + ], + "spans": [ + { + "bbox": [ + 62, + 254, + 207, + 264 + ], + "type": "text", + "content": "11: Backward and Update " + }, + { + "bbox": [ + 62, + 254, + 207, + 264 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_i" + }, + { + "bbox": [ + 62, + 254, + 207, + 264 + ], + "type": "text", + "content": " by Eq. (5)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 62, + 266, + 123, + 274 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 266, + 123, + 274 + ], + "spans": [ + { + "bbox": [ + 62, + 266, + 123, + 274 + ], + "type": "text", + "content": "12: Inference:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 62, + 275, + 218, + 286 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 275, + 218, + 286 + ], + "spans": [ + { + "bbox": [ + 62, + 275, + 218, + 286 + ], + "type": "text", + "content": "13: Obtain " + }, + { + "bbox": [ + 62, + 275, + 218, + 286 + ], + "type": "inline_equation", + "content": "\\hat{X}_i^t" + }, + { + "bbox": [ + 62, + 275, + 218, + 286 + ], + "type": "text", + "content": " using updated " + }, + { + "bbox": [ + 62, + 275, + 218, + 286 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_i" + }, + { + "bbox": [ + 62, + 275, + 218, + 286 + ], + "type": "text", + "content": " by Eq. (1)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 62, + 287, + 159, + 297 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 287, + 159, + 297 + ], + "spans": [ + { + "bbox": [ + 62, + 287, + 159, + 297 + ], + "type": "text", + "content": "14: Forward " + }, + { + "bbox": [ + 62, + 287, + 159, + 297 + ], + "type": "inline_equation", + "content": "O_{i} = f_{\\theta}(X_{i}^{t})" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 62, + 298, + 197, + 308 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 298, + 197, + 308 + ], + "spans": [ + { + "bbox": [ + 62, + 298, + 197, + 308 + ], + "type": "text", + "content": "15: Update the memory bank " + }, + { + "bbox": [ + 62, + 298, + 197, + 308 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 62, + 298, + 197, + 308 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 62, + 308, + 241, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 308, + 241, + 319 + ], + "spans": [ + { + "bbox": [ + 62, + 308, + 241, + 319 + ], + "type": "text", + "content": "16: M.enqueue " + }, + { + "bbox": [ + 62, + 308, + 241, + 319 + ], + "type": "inline_equation", + "content": "(q_{i},v_{i})" + }, + { + "bbox": [ + 62, + 308, + 241, + 319 + ], + "type": "inline_equation", + "content": "q_{i} = \\mathcal{F}^{A}\\left(X_{i}^{t}\\right),v_{i} = \\mathcal{P}_{i}" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 62, + 320, + 153, + 330 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 320, + 153, + 330 + ], + "spans": [ + { + "bbox": [ + 62, + 320, + 153, + 330 + ], + "type": "text", + "content": "17: if " + }, + { + "bbox": [ + 62, + 320, + 153, + 330 + ], + "type": "inline_equation", + "content": "\\text{len}(M) > S" + }, + { + "bbox": [ + 62, + 320, + 153, + 330 + ], + "type": "text", + "content": " then" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 62, + 331, + 186, + 342 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 331, + 186, + 342 + ], + "spans": [ + { + "bbox": [ + 62, + 331, + 186, + 342 + ], + "type": "text", + "content": "18: M.dequeue " + }, + { + "bbox": [ + 62, + 331, + 186, + 342 + ], + "type": "inline_equation", + "content": "(q_{\\text{first}}, v_{\\text{first}})" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 62, + 342, + 100, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 342, + 100, + 350 + ], + "spans": [ + { + "bbox": [ + 62, + 342, + 100, + 350 + ], + "type": "text", + "content": "19: end if" + } + ] + } + ], + "index": 21 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 59, + 354, + 177, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 354, + 177, + 365 + ], + "spans": [ + { + "bbox": [ + 59, + 354, + 177, + 365 + ], + "type": "text", + "content": "Output:Adapted prediction " + }, + { + "bbox": [ + 59, + 354, + 177, + 365 + ], + "type": "inline_equation", + "content": "O_{i}" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 47, + 392, + 216, + 404 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 392, + 216, + 404 + ], + "spans": [ + { + "bbox": [ + 47, + 392, + 216, + 404 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 392, + 216, + 404 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 47, + 392, + 216, + 404 + ], + "type": "text", + "content": " indicates the " + }, + { + "bbox": [ + 47, + 392, + 216, + 404 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 47, + 392, + 216, + 404 + ], + "type": "text", + "content": "-th BN layer in " + }, + { + "bbox": [ + 47, + 392, + 216, + 404 + ], + "type": "inline_equation", + "content": "f_{\\theta}" + }, + { + "bbox": [ + 47, + 392, + 216, + 404 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 46, + 405, + 287, + 489 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 405, + 287, + 489 + ], + "spans": [ + { + "bbox": [ + 46, + 405, + 287, + 489 + ], + "type": "text", + "content": "However, aligning the source and target domains within a limited number of iterations (one in this study) provides challenges due to the empty memory bank and undiminished distribution shifts at the beginning of the inference phase. Therefore, we design a warm-up mechanism to address this issue by simulating the warm-up statistics (i.e., " + }, + { + "bbox": [ + 46, + 405, + 287, + 489 + ], + "type": "inline_equation", + "content": "\\mu_w" + }, + { + "bbox": [ + 46, + 405, + 287, + 489 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 405, + 287, + 489 + ], + "type": "inline_equation", + "content": "\\sigma_w" + }, + { + "bbox": [ + 46, + 405, + 287, + 489 + ], + "type": "text", + "content": ") as follows:" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 61, + 500, + 287, + 514 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 500, + 287, + 514 + ], + "spans": [ + { + "bbox": [ + 61, + 500, + 287, + 514 + ], + "type": "interline_equation", + "content": "\\mu_ {w} = \\lambda \\mu_ {t} + (1 - \\lambda) \\mu_ {s}, \\sigma_ {w} = \\lambda \\sigma_ {t} + (1 - \\lambda) \\sigma_ {s}, \\tag {4}", + "image_path": "c776323d59cc17aadc2a80e5b8107d3684e14cb197e5c6bef4bc66499a7d46d9.jpg" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 46, + 523, + 287, + 585 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 523, + 287, + 585 + ], + "spans": [ + { + "bbox": [ + 46, + 523, + 287, + 585 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 523, + 287, + 585 + ], + "type": "inline_equation", + "content": "\\lambda = \\frac{1}{\\sqrt{i} / \\tau + 1}" + }, + { + "bbox": [ + 46, + 523, + 287, + 585 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 523, + 287, + 585 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 523, + 287, + 585 + ], + "type": "text", + "content": " indicates the index of the current test image and starts from 1, and " + }, + { + "bbox": [ + 46, + 523, + 287, + 585 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 46, + 523, + 287, + 585 + ], + "type": "text", + "content": " is a temperature coefficient to control the rate of transition from the warm-up statistics to the source statistics. Now, the Eq. (3) can be rewritten by:" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 92, + 586, + 287, + 611 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 92, + 586, + 287, + 611 + ], + "spans": [ + { + "bbox": [ + 92, + 586, + 287, + 611 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {p} = \\frac {1}{j} \\sum \\left| \\mu_ {w} ^ {j} - \\mu_ {t} ^ {j} \\right| + \\left| \\sigma_ {w} ^ {j} - \\sigma_ {t} ^ {j} \\right|, \\tag {5}", + "image_path": "c8ee097eeba7cf2cada56f862ef1d7287bbad7c82d513b17fa3fb6e757872886.jpg" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 46, + 618, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 618, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 618, + 287, + 713 + ], + "type": "text", + "content": "and we utilize " + }, + { + "bbox": [ + 46, + 618, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\mu_w" + }, + { + "bbox": [ + 46, + 618, + 287, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 618, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\sigma_w" + }, + { + "bbox": [ + 46, + 618, + 287, + 713 + ], + "type": "text", + "content": " for normalization instead of " + }, + { + "bbox": [ + 46, + 618, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\mu_s" + }, + { + "bbox": [ + 46, + 618, + 287, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 618, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\sigma_s" + }, + { + "bbox": [ + 46, + 618, + 287, + 713 + ], + "type": "text", + "content": ". As the inference progresses, the warm-up statistics will gradually shift towards the source statistics. By simulating the warm-up statistics between the source and target statistics, we can significantly reduce the training difficulty at the beginning of the inference phase, resulting in learning prompts from easy to difficult. The overall process of VPTTA is summarized in Algorithm 1." + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 306, + 71, + 387, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 71, + 387, + 85 + ], + "spans": [ + { + "bbox": [ + 306, + 71, + 387, + 85 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 305, + 91, + 545, + 127 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 91, + 545, + 127 + ], + "spans": [ + { + "bbox": [ + 305, + 91, + 545, + 127 + ], + "type": "text", + "content": "We evaluate our proposed VPTTA on two 2D segmentation benchmark tasks: the joint optic disc (OD) and cup (OC) segmentation task, and the polyp segmentation task." + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 305, + 133, + 481, + 144 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 133, + 481, + 144 + ], + "spans": [ + { + "bbox": [ + 305, + 133, + 481, + 144 + ], + "type": "text", + "content": "4.1. Datasets and Evaluation Metrics" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 304, + 151, + 545, + 270 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 151, + 545, + 270 + ], + "spans": [ + { + "bbox": [ + 304, + 151, + 545, + 270 + ], + "type": "text", + "content": "The OD/OC segmentation dataset comprises five public datasets collected from different medical centres, denoted as domain A (RIM-ONE-r3 [11]), B (REFuge [32]), C (ORIGA [50]), D (REFuge-Validation/Test [32]), and E (Drishti-GS [38]). There are 159, 400, 650, 800, and 101 images from these datasets. We cropped a region of interest (ROI) centering at OD with size of " + }, + { + "bbox": [ + 304, + 151, + 545, + 270 + ], + "type": "inline_equation", + "content": "800 \\times 800" + }, + { + "bbox": [ + 304, + 151, + 545, + 270 + ], + "type": "text", + "content": " for each image following [19], and each ROI is further resized to " + }, + { + "bbox": [ + 304, + 151, + 545, + 270 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 304, + 151, + 545, + 270 + ], + "type": "text", + "content": " and normalized by min-max normalization. The Dice score metric " + }, + { + "bbox": [ + 304, + 151, + 545, + 270 + ], + "type": "inline_equation", + "content": "(DSC)" + }, + { + "bbox": [ + 304, + 151, + 545, + 270 + ], + "type": "text", + "content": " is utilized for evaluation in this task." + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 304, + 270, + 545, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 270, + 545, + 389 + ], + "spans": [ + { + "bbox": [ + 304, + 270, + 545, + 389 + ], + "type": "text", + "content": "The polyp segmentation dataset consists of four public datasets collected from different medical centres, denoted as domain A (BKAI-IGH-NeoPolyp [28]), B (CVC-ClinicDB [1]), C (ETIS-LaribPolypDB [36]), and D (Kvasir-Seg [20]), which have 1000, 612, 196, and 1000 images, respectively. We followed [9] to resize each image to " + }, + { + "bbox": [ + 304, + 270, + 545, + 389 + ], + "type": "inline_equation", + "content": "352 \\times 352" + }, + { + "bbox": [ + 304, + 270, + 545, + 389 + ], + "type": "text", + "content": " and normalize the resized image by the statistics computed on ImageNet. The " + }, + { + "bbox": [ + 304, + 270, + 545, + 389 + ], + "type": "inline_equation", + "content": "DSC" + }, + { + "bbox": [ + 304, + 270, + 545, + 389 + ], + "type": "text", + "content": ", enhanced-alignment metric " + }, + { + "bbox": [ + 304, + 270, + 545, + 389 + ], + "type": "inline_equation", + "content": "(E_{\\phi}^{max})" + }, + { + "bbox": [ + 304, + 270, + 545, + 389 + ], + "type": "text", + "content": " [8], and structural similarity metric " + }, + { + "bbox": [ + 304, + 270, + 545, + 389 + ], + "type": "inline_equation", + "content": "(S_{\\alpha})" + }, + { + "bbox": [ + 304, + 270, + 545, + 389 + ], + "type": "text", + "content": " [7] are utilized for evaluation in this task." + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 305, + 396, + 438, + 408 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 396, + 438, + 408 + ], + "spans": [ + { + "bbox": [ + 305, + 396, + 438, + 408 + ], + "type": "text", + "content": "4.2. Implementation Details" + } + ] + } + ], + "index": 35 + }, + { + "bbox": [ + 304, + 415, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 415, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 415, + 545, + 713 + ], + "type": "text", + "content": "For each task, we trained the source model on each single domain (source domain) and tested it on the left domains (target domains) to calculate the mean metrics to evaluate the methods on different scenarios. In the source-training phase, we trained a ResUNet-34 [17] backbone, following [19], as the baseline for OD/OC segmentation task and trained the PraNet [9] with a Res2Net-based [14] backbone for polyp segmentation task. Due to the complexity of PraNet's decoder, we only calculated the prompt loss " + }, + { + "bbox": [ + 304, + 415, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_p" + }, + { + "bbox": [ + 304, + 415, + 545, + 713 + ], + "type": "text", + "content": " for each BN layer in the encoder instead of the whole network. In the test-adaptation phase, We performed one-iteration adaptation for each batch of test data with a batch size of 1 on all experiments of our VPTTA and other competing methods [45]. As our VPTTA involves an iteration to update and an inference step, distinguishing it from other methods like [41] that do not require a dedicated inference step, we applied the same configuration as VPTTA to other competing methods to ensure consistency in the experimental setup. To deploy our VPTTA, we utilized the Adam optimizer with a learning rate of 0.05 and 0.01 for the OD/OC segmentation task and polyp segmentation task, respectively. The hyperparameters " + }, + { + "bbox": [ + 304, + 415, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 304, + 415, + 545, + 713 + ], + "type": "text", + "content": " (size of prompt), " + }, + { + "bbox": [ + 304, + 415, + 545, + 713 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 304, + 415, + 545, + 713 + ], + "type": "text", + "content": " (size of memory bank), " + }, + { + "bbox": [ + 304, + 415, + 545, + 713 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 304, + 415, + 545, + 713 + ], + "type": "text", + "content": " (size of support set), and " + }, + { + "bbox": [ + 304, + 415, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 304, + 415, + 545, + 713 + ], + "type": "text", + "content": " (temperature coefficient in warm-up) are set to 0.01, 40, 16, and 5 for both segmentation tasks." + } + ] + } + ], + "index": 36 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "11188" + } + ] + } + ], + "index": 37 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 110, + 102, + 481, + 201 + ], + "blocks": [ + { + "bbox": [ + 46, + 70, + 547, + 95 + ], + "lines": [ + { + "bbox": [ + 46, + 70, + 547, + 95 + ], + "spans": [ + { + "bbox": [ + 46, + 70, + 547, + 95 + ], + "type": "text", + "content": "Table 1. Performance of our VPTTA, 'Source Only' baseline, and six competing methods on the OD/OC segmentation task. The best and second-best results in each column are highlighted in bold and underline, respectively." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 110, + 102, + 481, + 201 + ], + "lines": [ + { + "bbox": [ + 110, + 102, + 481, + 201 + ], + "spans": [ + { + "bbox": [ + 110, + 102, + 481, + 201 + ], + "type": "table", + "html": "
MethodsDomain ADomain BDomain CDomain DDomain EAverage
DSCDSCDSCDSCDSCDSC↑
Source Only (ResUNet-34)64.5376.0671.1852.6764.8765.86
TENT-continual (ICLR 2021) [41]73.0778.6671.9446.8170.2068.13
CoTTA (CVPR 2022) [42]75.3975.9869.1453.9970.4068.98
DLTTA (TMI 2022) [45]75.1178.8573.8951.6469.7169.84
DUA (CVPR 2022) [25]72.2876.5970.1356.1771.3869.31
SAR (ICLR 2023) [31]74.5577.7170.7855.4071.7270.03
DomainAdaptor (CVPR 2023) [47]74.5076.3971.8156.7870.5570.01
VPTTA (Ours)73.9179.3674.5156.5175.3571.93
", + "image_path": "8fd954f0893c4820081f1b86d15d86c66d9b4c9de30aba6b62969226756b1ac6.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 49, + 242, + 545, + 332 + ], + "blocks": [ + { + "bbox": [ + 46, + 211, + 547, + 236 + ], + "lines": [ + { + "bbox": [ + 46, + 211, + 547, + 236 + ], + "spans": [ + { + "bbox": [ + 46, + 211, + 547, + 236 + ], + "type": "text", + "content": "Table 2. Performance of our VPTTA, 'Source Only' baseline, and six competing methods on the polyp segmentation task. The best and second-best results in each column are highlighted in bold and underline, respectively." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 49, + 242, + 545, + 332 + ], + "lines": [ + { + "bbox": [ + 49, + 242, + 545, + 332 + ], + "spans": [ + { + "bbox": [ + 49, + 242, + 545, + 332 + ], + "type": "table", + "html": "
MethodsDomain ADomain BDomain CDomain DAverage
\\(DSC\\)\\(E_{\\phi}^{max}\\)\\(S_{\\alpha}\\)\\(DSC\\)\\(E_{\\phi}^{max}\\)\\(S_{\\alpha}\\)\\(DSC\\)\\(E_{\\phi}^{max}\\)\\(S_{\\alpha}\\)\\(DSC\\)\\(E_{\\phi}^{max}\\)\\(S_{\\alpha}\\)\\(DSC \\uparrow\\)\\(E_{\\phi}^{max} \\uparrow\\)\\(S_{\\alpha} \\uparrow\\)
Source Only (PraNet)79.9087.9784.6666.3378.5176.7273.8984.6481.2882.9590.8488.0875.7785.4982.69
TENT-continual (ICLR 2021) [41]74.8684.5880.5267.5178.6678.0517.7940.0453.3073.5583.3882.4158.4371.6773.57
CoTTA (CVPR 2022) [42]76.4685.3782.5666.7776.7579.1771.3983.4280.1870.7179.8182.5471.3381.3481.11
DLTTA (TMI 2022) [45]76.2785.2382.4166.5877.0079.2463.7278.2375.5671.2081.3283.4769.4480.4580.17
DUA (CVPR 2022) [25]78.9387.3783.9666.8478.5277.5176.5386.4583.0586.2493.2389.8277.1386.3983.58
SAR (ICLR 2023) [31]76.4885.8981.4966.4577.3578.0571.4683.2379.4070.4180.1181.0771.2081.6580.00
DomainAdaptor (CVPR 2023) [47]77.4886.3182.4070.8281.7680.8871.9683.0679.9776.8985.8984.4574.2984.2681.93
VPTTA (Ours)81.0088.9184.9176.8787.3184.0877.5887.4883.6486.3993.4789.8780.4689.2985.62
", + "image_path": "9634239044c431f8dadbc366191c2b1851142db7fa988c64c041e94d9ea1bb0a.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 87, + 373, + 507, + 445 + ], + "blocks": [ + { + "bbox": [ + 46, + 342, + 547, + 367 + ], + "lines": [ + { + "bbox": [ + 46, + 342, + 547, + 367 + ], + "spans": [ + { + "bbox": [ + 46, + 342, + 547, + 367 + ], + "type": "text", + "content": "Table 3. Results of ablation study on the OD/OC segmentation task. The best and second-best results in each column are highlighted in bold and underline, respectively." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 87, + 373, + 507, + 445 + ], + "lines": [ + { + "bbox": [ + 87, + 373, + 507, + 445 + ], + "spans": [ + { + "bbox": [ + 87, + 373, + 507, + 445 + ], + "type": "table", + "html": "
MethodsDomain ADomain BDomain CDomain DDomain EAverage
Low-frequency PromptMemory BankWarm-upDSCDSCDSCDSCDSCDSC↑
64.5376.0671.1852.6764.8765.86
✓65.9976.5871.7552.8667.3066.90
✓✓70.7079.3672.8255.5072.8470.24
✓✓74.4177.4774.1454.0367.8569.58
✓✓✓73.9179.3674.5156.5175.3571.93
", + "image_path": "939cf4dca1e1de0085c672d9cddd7409743840ba63b8513c426ca955bd096682.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 463, + 171, + 476 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 463, + 171, + 476 + ], + "spans": [ + { + "bbox": [ + 47, + 463, + 171, + 476 + ], + "type": "text", + "content": "4.3. Experimental Results" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 483, + 288, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 483, + 288, + 651 + ], + "spans": [ + { + "bbox": [ + 46, + 483, + 288, + 651 + ], + "type": "text", + "content": "We compare our VPTTA with the 'Source Only' baseline (training on the source domain and testing without adaptation), and six competing methods, including a pseudo-label-based method (CoTTA [42]), two entropy-based methods (TENT-continual [41] and SAR [31]), a method dynamically adjusting the learning rate (DLTTA [45]), a method combining entropy-loss and BN statistics fusion (DomainAdaptor [47]), and a BN statistics-based method (DUA [25]) on two segmentation tasks. Specifically, we re-implemented all the competing methods using the same baseline as VPTTA and combined DLTTA with the entropy loss proposed in TENT [45]. The result of each domain is calculated by using the current domain as the source domain and applying the pre-trained model on the left domains." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "type": "text", + "content": "Comparison on the OD/OC segmentation task. The results of the proposed VPTTA, the \"Source Only\" baseline, and six competing methods on the OD/OC segmentation task are detailed in Table 1. Obviously, all competing methods outperform the \"Source Only\" baseline, indicating the" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 464, + 547, + 560 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 464, + 547, + 560 + ], + "spans": [ + { + "bbox": [ + 304, + 464, + 547, + 560 + ], + "type": "text", + "content": "effectiveness of adaptation towards the distribution shifts. It can be found that our VPTTA achieves the best overall performance across all domains, underscoring its superior applicability and robustness. Meanwhile, our VPTTA also shows competitive performance in domains where most methods exhibit satisfactory performance (such as D) and also performs well in domains where other methods exhibit suboptimal results (such as C)." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 570, + 548, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 548, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 548, + 715 + ], + "type": "text", + "content": "Comparison on the polyp segmentation task. We conducted the comparison experiment on the polyp segmentation task under the same experimental setup, as shown in Table 2. In contrast to the results of the OD/OC segmentation task, the methods based on model updating (i.e., TENT-continual, CoTTA, DLTTA, SAR, and DomainAdaptor) exhibit significantly worse performance in this task and are even inferior to the 'Source Only' baseline. It can be attributed to the fact that the polyps are relatively hidden, and the models tend to segment almost nothing rather than poorly segment under the impact of distribution shifts, resulting in confident and low entropy predictions but wrong" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "11189" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 62, + 70, + 532, + 395 + ], + "blocks": [ + { + "bbox": [ + 62, + 70, + 532, + 395 + ], + "lines": [ + { + "bbox": [ + 62, + 70, + 532, + 395 + ], + "spans": [ + { + "bbox": [ + 62, + 70, + 532, + 395 + ], + "type": "image", + "image_path": "94d284a1840b9156326e8cf956d20c3c1507350ec98d944cf0da0f56a1f69e5a.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 402, + 547, + 437 + ], + "lines": [ + { + "bbox": [ + 46, + 402, + 547, + 437 + ], + "spans": [ + { + "bbox": [ + 46, + 402, + 547, + 437 + ], + "type": "text", + "content": "Figure 3. Visualization of the original images, estimated prompts, and adapted images on the OD/OC segmentation task. We normalize the prompts to [0, 1] for better visualization. The DSC of applying the frozen source model on the original and adapted images is displayed below each image. We also show an example of each source domain on the left side of this diagram. 'Ori': Abbreviation of 'Original'." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 456, + 287, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 456, + 287, + 552 + ], + "spans": [ + { + "bbox": [ + 46, + 456, + 287, + 552 + ], + "type": "text", + "content": "gradients for these methods. Since DUA is backward-free, it still performs well in this task. However, DUA also fails when using domain A as the source domain and has limited performance due to the absence of training for the current test image. Our VPTTA simultaneously benefits from training on the current test image and refraining from updating the model, thus surpassing all competing methods and demonstrating advanced performance." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 555, + 287, + 699 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 555, + 287, + 699 + ], + "spans": [ + { + "bbox": [ + 46, + 555, + 287, + 699 + ], + "type": "text", + "content": "Visualization of prompts and adapted images. We visualized the estimated prompts and adapted images on the OD/OC segmentation task in Figure 3 and displayed the DSC values predicted on the original and adapted images below the corresponding images. It can be observed that: (1) Applying prompts to the original test images can achieve significant performance gain; (2) Images from different target domains have similar prompts, proving that involving the memory bank to initialize the current prompt using previous ones is reasonable; (3) The appearances of adapted images are close to the source domain, explicitly demonstrating the reduction of distribution shifts via our VPTTA." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 701, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 701, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 701, + 287, + 714 + ], + "type": "text", + "content": "Comparison in a long-term continual test-time adap" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 456, + 545, + 708 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 456, + 545, + 708 + ], + "spans": [ + { + "bbox": [ + 304, + 456, + 545, + 708 + ], + "type": "text", + "content": "tation. To evaluate our VPTTA in a long-term continual test-time adaptation [42], we conducted the experiments for the OD/OC segmentation task multiple rounds and reported the results in Table 4. Since DUA is a backward-free method and DomainAdaptor resets the pre-trained model before each adaptation, they are unsuitable to be evaluated in this experiment. We compared our VPTTA with TENT, CoTTA, DLTTA, and SAR, all of which involve continuous training of the pre-trained model. The performance across different rounds with the same source domain reflects the extent of catastrophic forgetting in the model, and the average performance across multiple rounds highlights the error accumulation of the model within a series of distinct domains. The results reveal that both optimizing loss functions (CoTTA, DLTTA, SAR) and resetting the model (CoTTA, SAR) are effective for mitigating error accumulation and catastrophic forgetting, while TENT with vanilla entropy-minimization loss suffers from severe performance degradation. Additionally, after multiple rounds of testing, our VPTTA maintains superior performance with minimal performance degradation, emphasizing the effectiveness of" + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "11190" + } + ] + } + ], + "index": 6 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 49, + 113, + 545, + 182 + ], + "blocks": [ + { + "bbox": [ + 46, + 70, + 547, + 105 + ], + "lines": [ + { + "bbox": [ + 46, + 70, + 547, + 105 + ], + "spans": [ + { + "bbox": [ + 46, + 70, + 547, + 105 + ], + "type": "text", + "content": "Table 4. Performance of our VPTTA and four competing methods on the OD/OC segmentation task in a long-term continual test-time adaptation. The red numbers indicate the performance gain relative to the 'Source Only' baseline. The performance degradation is calculated between the overall average DSC and the average DSC of round 1. 'Ave': Abbreviation of 'Average'." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 49, + 113, + 545, + 182 + ], + "lines": [ + { + "bbox": [ + 49, + 113, + 545, + 182 + ], + "spans": [ + { + "bbox": [ + 49, + 113, + 545, + 182 + ], + "type": "table", + "html": "
Time→
Round123Average DSC ↑Perform. Degra. ↓
MethodsABCDEAveABCDEAveABCDEAve
Source Only64.5376.0671.1852.6764.8765.8664.5376.0671.1852.6764.8765.8664.5376.0671.1852.6764.8765.8665.86
TENT73.0778.6671.9446.8170.2068.1362.0969.3270.6739.0268.2261.8657.0562.4770.2039.0266.3759.0263.01 (-2.85)
CoTTA75.3975.9869.1453.9970.4068.9874.3175.0067.9951.0468.2867.3273.2274.3366.7250.2367.0866.3267.54 (+1.68)
DLTTA75.1178.8573.8951.6469.7169.8474.1479.6574.2545.0569.0468.4372.2878.9372.8742.3769.2667.1468.47 (+2.61)
SAR74.5577.7170.7855.4071.7270.0374.7478.0971.0052.1369.0269.0074.9078.2471.1850.1668.4468.5869.20 (+3.34)
VPTTA (Ours)73.9179.3674.5156.5175.3571.9373.5778.8473.6156.9174.8071.5573.1278.4572.6357.1174.0471.0771.51 (+5.65)
", + "image_path": "4ace3b55ebda693e4bc20c346b4298793c6c6ffe67fc0d9ad5747e9f4adc1e7d.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 202, + 265, + 214 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 202, + 265, + 214 + ], + "spans": [ + { + "bbox": [ + 46, + 202, + 265, + 214 + ], + "type": "text", + "content": "training prompts over updating the pre-trained model." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 215, + 288, + 358 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 215, + 288, + 358 + ], + "spans": [ + { + "bbox": [ + 46, + 215, + 288, + 358 + ], + "type": "text", + "content": "Ablation study. To evaluate the contributions of the low-frequency prompt, memory-bank-based initialization, and warm-up mechanism, we conducted ablation studies on the OD/OC segmentation task, as shown in Table 3. Note that Eq. (3)/(5) is used as the loss function when training the prompt without/with our warm-up mechanism. The results show that (1) using the low-frequency prompt only has limited performance gain due to the poorly trained prompt; (2) both the memory bank and warm-up mechanism are beneficial to training the prompt; (3) the best performance is achieved when the memory bank and warm-up mechanism are jointly used, which means they are complementary." + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 52, + 384, + 281, + 453 + ], + "blocks": [ + { + "bbox": [ + 141, + 368, + 195, + 378 + ], + "lines": [ + { + "bbox": [ + 141, + 368, + 195, + 378 + ], + "spans": [ + { + "bbox": [ + 141, + 368, + 195, + 378 + ], + "type": "text", + "content": "Average DSC" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 52, + 384, + 281, + 453 + ], + "lines": [ + { + "bbox": [ + 52, + 384, + 281, + 453 + ], + "spans": [ + { + "bbox": [ + 52, + 384, + 281, + 453 + ], + "type": "image", + "image_path": "585cc210003974459a61397dbb3425e8ea254f78960f4a424dc0e4ba883f8fb7.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 460, + 288, + 483 + ], + "lines": [ + { + "bbox": [ + 46, + 460, + 288, + 483 + ], + "spans": [ + { + "bbox": [ + 46, + 460, + 288, + 483 + ], + "type": "text", + "content": "Figure 4. Performance of our VPTTA with various " + }, + { + "bbox": [ + 46, + 460, + 288, + 483 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 46, + 460, + 288, + 483 + ], + "type": "text", + "content": " on the OD/OC segmentation task." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 51, + 515, + 281, + 588 + ], + "blocks": [ + { + "bbox": [ + 141, + 504, + 194, + 515 + ], + "lines": [ + { + "bbox": [ + 141, + 504, + 194, + 515 + ], + "spans": [ + { + "bbox": [ + 141, + 504, + 194, + 515 + ], + "type": "text", + "content": "Average DSC" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 51, + 515, + 281, + 588 + ], + "lines": [ + { + "bbox": [ + 51, + 515, + 281, + 588 + ], + "spans": [ + { + "bbox": [ + 51, + 515, + 281, + 588 + ], + "type": "image", + "image_path": "6a6f9dace7fe6ab79db53b357d3d6f2f284c705985fe75caf97aa6473a009c5a.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 595, + 287, + 617 + ], + "lines": [ + { + "bbox": [ + 46, + 595, + 287, + 617 + ], + "spans": [ + { + "bbox": [ + 46, + 595, + 287, + 617 + ], + "type": "text", + "content": "Figure 5. Performance of our VPTTA with various " + }, + { + "bbox": [ + 46, + 595, + 287, + 617 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 46, + 595, + 287, + 617 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 595, + 287, + 617 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 46, + 595, + 287, + 617 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 595, + 287, + 617 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 46, + 595, + 287, + 617 + ], + "type": "text", + "content": " on the OD/OC segmentation task." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 629, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 629, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 629, + 288, + 713 + ], + "type": "text", + "content": "Analysis of hyperparameter " + }, + { + "bbox": [ + 46, + 629, + 288, + 713 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 46, + 629, + 288, + 713 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 629, + 288, + 713 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 46, + 629, + 288, + 713 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 629, + 288, + 713 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 46, + 629, + 288, + 713 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 629, + 288, + 713 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 46, + 629, + 288, + 713 + ], + "type": "text", + "content": ". We conducted the experiments for the OD/OC segmentation task to discuss the hyper-parameters used in our VPTTA and displayed the results in Figure 4 and Figure 5. It can be seen that " + }, + { + "bbox": [ + 46, + 629, + 288, + 713 + ], + "type": "inline_equation", + "content": "\\alpha = 0.01" + }, + { + "bbox": [ + 46, + 629, + 288, + 713 + ], + "type": "text", + "content": " is optimal, as an excessively small prompt makes adaptation challenging, while an overly large prompt hinders effective training. Similar trends are also reflected" + } + ] + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 309, + 232, + 541, + 269 + ], + "blocks": [ + { + "bbox": [ + 305, + 201, + 545, + 222 + ], + "lines": [ + { + "bbox": [ + 305, + 201, + 545, + 222 + ], + "spans": [ + { + "bbox": [ + 305, + 201, + 545, + 222 + ], + "type": "text", + "content": "Table 5. Performance of our VPTTA and VPTTA-LR on two segmentation tasks." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 309, + 232, + 541, + 269 + ], + "lines": [ + { + "bbox": [ + 309, + 232, + 541, + 269 + ], + "spans": [ + { + "bbox": [ + 309, + 232, + 541, + 269 + ], + "type": "table", + "html": "
MethodsOD/OCPolyp
DSC↑Param↓DSC↑Eφmax↑Sα↑Param↓
VPTTA-LR (r=3)69.019,21678.2588.2784.206,336
VPTTA71.937580.4689.9486.8427
", + "image_path": "a3b57ec2acd4b3b02424cd527f60a9e5fe7676ad8c81a05db550a5dfe215d916.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_body" + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 288, + 545, + 371 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 288, + 545, + 371 + ], + "spans": [ + { + "bbox": [ + 304, + 288, + 545, + 371 + ], + "type": "text", + "content": "in " + }, + { + "bbox": [ + 304, + 288, + 545, + 371 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 304, + 288, + 545, + 371 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 288, + 545, + 371 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 304, + 288, + 545, + 371 + ], + "type": "text", + "content": ": as they increase, the DSC values improve, while the excessively large " + }, + { + "bbox": [ + 304, + 288, + 545, + 371 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 304, + 288, + 545, + 371 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 288, + 545, + 371 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 304, + 288, + 545, + 371 + ], + "type": "text", + "content": " may introduce unrelated samples, affecting the initialization. With the increase of " + }, + { + "bbox": [ + 304, + 288, + 545, + 371 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 304, + 288, + 545, + 371 + ], + "type": "text", + "content": ", performance gradually approaches the upper bound and tends to remain stable, suggesting that the warm-up has reached saturation. Therefore, we finally chose " + }, + { + "bbox": [ + 304, + 288, + 545, + 371 + ], + "type": "inline_equation", + "content": "\\alpha = 0.01" + }, + { + "bbox": [ + 304, + 288, + 545, + 371 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 288, + 545, + 371 + ], + "type": "inline_equation", + "content": "S = 40" + }, + { + "bbox": [ + 304, + 288, + 545, + 371 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 288, + 545, + 371 + ], + "type": "inline_equation", + "content": "K = 16" + }, + { + "bbox": [ + 304, + 288, + 545, + 371 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 288, + 545, + 371 + ], + "type": "inline_equation", + "content": "\\tau = 5" + }, + { + "bbox": [ + 304, + 288, + 545, + 371 + ], + "type": "text", + "content": " as the best configuration." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 372, + 546, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 372, + 546, + 479 + ], + "spans": [ + { + "bbox": [ + 304, + 372, + 546, + 479 + ], + "type": "text", + "content": "Low-frequency prompt vs. low-rank prompt. To compare the effectiveness of VPTTA and VPTTA-LR, we repeated the experiments on two segmentation tasks and showed results in Table 5. It can be seen that VPTTA significantly outperforms VPTTA-LR, attributed to the fewer learnable parameters and superior representation of style texture in the frequency domain. Benefiting from the low-frequency prompt, VPTTA can be effectively trained with over a hundred times fewer parameters than VPTTA-LR." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 490, + 379, + 502 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 490, + 379, + 502 + ], + "spans": [ + { + "bbox": [ + 306, + 490, + 379, + 502 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 510, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 510, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 510, + 547, + 713 + ], + "type": "text", + "content": "In this paper, we proposed the VPTTA, a prompt-based method to alleviate the potential risks (i.e., error accumulation, and catastrophic forgetting) by freezing the pre-trained model. Specially, we presented the low-frequency prompt that trains a lightweight prompt with only a small number of parameters in a single iteration to alleviate the distribution shifts in BN layers, constructed a memory bank to better initialize each prompt by utilizing the prompts of previous and the most similar test images, and designed the statistics-fusion-based warm-up mechanism to simulate the warm-up statistics to overcome the training difficulty at the beginning of the inference phase. The extensive experimental results on two medical image segmentation benchmark tasks with multiple domains verified the superiority of our VPTTA over other methods and the effectiveness of each special design. In this regard, we hope that our efforts can offer a different paradigm for continual test-time adaptation." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "11191" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 146 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 146 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 146 + ], + "type": "text", + "content": "[1] Jorge Bernal, F Javier Sánchez, Gloria Fernández-Esparrach, Debora Gil, Cristina Rodríguez, and Fernando Vilarino. Wm-dova maps for accurate polyp highlighting in colonoscopy: Validation vs. saliency maps from physicians. Comput. Med. Imaging and Graph., 43:99-111, 2015. 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 148, + 288, + 191 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 148, + 288, + 191 + ], + "spans": [ + { + "bbox": [ + 53, + 148, + 288, + 191 + ], + "type": "text", + "content": "[2] Paola Cascante-Bonilla, Fuwen Tan, Yanjun Qi, and Vicente Ordonez. Curriculum labeling: Revisiting pseudo-labeling for semi-supervised learning. In AAAI, pages 6912-6920, 2021. 2, 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 193, + 287, + 225 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 193, + 287, + 225 + ], + "spans": [ + { + "bbox": [ + 53, + 193, + 287, + 225 + ], + "type": "text", + "content": "[3] Oscar Chang, Lampros Flokas, and Hod Lipson. Principled weight initialization for hypernetworks. In Int. Conf. Learn. Represent., 2019. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 227, + 287, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 227, + 287, + 259 + ], + "spans": [ + { + "bbox": [ + 53, + 227, + 287, + 259 + ], + "type": "text", + "content": "[4] Dian Chen, Dequan Wang, Trevor Darrell, and Sayna Ebrahimi. Contrastive test-time adaptation. In IEEE Conf. Comput. Vis. Pattern Recog., pages 295-305, 2022. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 261, + 287, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 261, + 287, + 316 + ], + "spans": [ + { + "bbox": [ + 53, + 261, + 287, + 316 + ], + "type": "text", + "content": "[5] Lin Chen, Huaian Chen, Zhixiang Wei, Xin Jin, Xiao Tan, Yi Jin, and Enhong Chen. Reusing the task-specific classifier as a discriminator: Discriminator-free adversarial domain adaptation. In IEEE Conf. Comput. Vis. Pattern Recog., pages 7181-7190, 2022. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 318, + 287, + 361 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 318, + 287, + 361 + ], + "spans": [ + { + "bbox": [ + 53, + 318, + 287, + 361 + ], + "type": "text", + "content": "[6] Mario Döbler, Robert A Marsden, and Bin Yang. Robust mean teacher for continual and gradual test-time adaptation. In IEEE Conf. Comput. Vis. Pattern Recog., pages 7704-7714, 2023. 2, 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 363, + 287, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 363, + 287, + 396 + ], + "spans": [ + { + "bbox": [ + 53, + 363, + 287, + 396 + ], + "type": "text", + "content": "[7] Deng-Ping Fan, Ming-Ming Cheng, Yun Liu, Tao Li, and Ali Borji. Structure-measure: A new way to evaluate foreground maps. In Int. Conf. Comput. Vis., pages 4548-4557, 2017. 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 397, + 287, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 397, + 287, + 441 + ], + "spans": [ + { + "bbox": [ + 53, + 397, + 287, + 441 + ], + "type": "text", + "content": "[8] Deng-Ping Fan, Cheng Gong, Yang Cao, Bo Ren, Ming-Ming Cheng, and Ali Borji. Enhanced-alignment measure for binary foreground map evaluation. In IJCAI, pages 698-704, 2018. 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 443, + 287, + 497 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 443, + 287, + 497 + ], + "spans": [ + { + "bbox": [ + 53, + 443, + 287, + 497 + ], + "type": "text", + "content": "[9] Deng-Ping Fan, Ge-Peng Ji, Tao Zhou, Geng Chen, Huazhu Fu, Jianbing Shen, and Ling Shao. Pranet: Parallel reverse attention network for polyp segmentation. In Int. Conf. Med. Image Comput. Comput.-Assist. Intervention., pages 263-273. Springer, 2020. 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 499, + 287, + 531 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 499, + 287, + 531 + ], + "spans": [ + { + "bbox": [ + 48, + 499, + 287, + 531 + ], + "type": "text", + "content": "[10] Matteo Frigo and Steven G Johnson. FFTW: An adaptive software architecture for the fft. In ICASSP, pages 1381-1384. IEEE, 1998. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 533, + 287, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 533, + 287, + 586 + ], + "spans": [ + { + "bbox": [ + 48, + 533, + 287, + 586 + ], + "type": "text", + "content": "[11] Francisco Fumero, Silvia Alayón, José L Sanchez, Jose Sigut, and M Gonzalez-Hernandez. RIM-ONE: An open retinal image database for optic nerve evaluation. In Int. Symp. Comput.-based Med. Syst., pages 1-6. IEEE, 2011. 5" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 590, + 287, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 590, + 287, + 633 + ], + "spans": [ + { + "bbox": [ + 48, + 590, + 287, + 633 + ], + "type": "text", + "content": "[12] Yulu Gan, Yan Bai, Yihang Lou, Xianzheng Ma, Renrui Zhang, Nian Shi, and Lin Luo. Decorate the newcomers: Visual domain prompt for continual test time adaptation. In AAAI, pages 7595-7603, 2023. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 635, + 287, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 635, + 287, + 667 + ], + "spans": [ + { + "bbox": [ + 48, + 635, + 287, + 667 + ], + "type": "text", + "content": "[13] Yossi Gandelsman, Yu Sun, Xinlei Chen, and Alexei Efros. Test-time training with masked autoencoders. Adv. Neural Inform. Process. Syst., 35:29374-29385, 2022. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "type": "text", + "content": "[14] Shang-Hua Gao, Ming-Ming Cheng, Kai Zhao, Xin-Yu Zhang, Ming-Hsuan Yang, and Philip Torr. Res2net: A new multi-scale backbone architecture. IEEE Trans. Pattern Anal. Mach. Intell., 43(2):652-662, 2019. 5" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 117 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 117 + ], + "type": "text", + "content": "[15] Yunhe Gao, Xingjian Shi, Yi Zhu, Hao Wang, Zhiqiang Tang, Xiong Zhou, Mu Li, and Dimitris N Metaxas. Visual prompt tuning for test-time domain adaptation. arXiv preprint arXiv:2210.04831, 2022. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 119, + 545, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 119, + 545, + 194 + ], + "spans": [ + { + "bbox": [ + 307, + 119, + 545, + 194 + ], + "type": "text", + "content": "[16] Mohsen Ghafoorian, Alireza Mehrtash, Tina Kapur, Nico Karssemeijer, Elena Marchiori, Mehran Pesteie, Charles RG Guttmann, Frank-Erik de Leeuw, Clare M Tempany, Bram Van Ginneken, et al. Transfer learning for domain adaptation in mri: Application in brain lesion segmentation. In Int. Conf. Med. Image Comput. Comput.-Assist. Intervent., pages 516-524. Springer, 2017. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 197, + 545, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 197, + 545, + 228 + ], + "spans": [ + { + "bbox": [ + 307, + 197, + 545, + 228 + ], + "type": "text", + "content": "[17] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In IEEE Conf. Comput. Vis. Pattern Recog., pages 770-778, 2016. 5" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 230, + 545, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 230, + 545, + 274 + ], + "spans": [ + { + "bbox": [ + 308, + 230, + 545, + 274 + ], + "type": "text", + "content": "[18] Edward J Hu, Phillip Wallis, Zeyuan Allen-Zhu, Yuzhhi Li, Shean Wang, Lu Wang, Weizhu Chen, et al. Lora: Low-rank adaptation of large language models. In Int. Conf. Learn. Represent., 2022. 4" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 276, + 545, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 276, + 545, + 319 + ], + "spans": [ + { + "bbox": [ + 308, + 276, + 545, + 319 + ], + "type": "text", + "content": "[19] Shishuai Hu, Zehui Liao, and Yong Xia. Prosfda: Prompt learning based source-free domain adaptation for medical image segmentation. arXiv preprint arXiv:2211.11514, 2022. 2, 3, 4, 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 321, + 545, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 321, + 545, + 374 + ], + "spans": [ + { + "bbox": [ + 308, + 321, + 545, + 374 + ], + "type": "text", + "content": "[20] Debesh Jha, Pia H Smedsrud, Michael A Riegler, Pål Halvorsen, Thomas de Lange, Dag Johansen, and Håvard D Johansen. Kvasir-seg: A segmented polyp dataset. In Int. Conf. Multimedia Modeling, pages 451-462. Springer, 2020. 5" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 376, + 545, + 420 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 376, + 545, + 420 + ], + "spans": [ + { + "bbox": [ + 307, + 376, + 545, + 420 + ], + "type": "text", + "content": "[21] Christian Lamers, René Vidal, Nabil Belbachir, Niki van Stein, Thomas Bäeck, and Paris Giampouras. Clustering-based domain-incremental learning. In Int. Conf. Comput. Vis., pages 3384-3392, 2023. 2, 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 422, + 545, + 454 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 422, + 545, + 454 + ], + "spans": [ + { + "bbox": [ + 307, + 422, + 545, + 454 + ], + "type": "text", + "content": "[22] Joonseok Lee, Seungyeon Kim, Guy Lebanon, and Yoram Singer. Local low-rank matrix approximation. In Int. Conf. Mach. Learn., pages 82-90. PMLR, 2013. 4" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 456, + 545, + 488 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 456, + 545, + 488 + ], + "spans": [ + { + "bbox": [ + 307, + 456, + 545, + 488 + ], + "type": "text", + "content": "[23] Jian Liang, Ran He, Zhenan Sun, and Tieniu Tan. Exploring uncertainty in pseudo-label guided unsupervised domain adaptation. Pattern Recognition, 96:106996, 2019. 2, 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 491, + 545, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 491, + 545, + 544 + ], + "spans": [ + { + "bbox": [ + 307, + 491, + 545, + 544 + ], + "type": "text", + "content": "[24] Pengfei Liu, Weizhe Yuan, Jinlan Fu, Zhengbao Jiang, Hiroaki Hayashi, and Graham Neubig. Pre-train, prompt, and predict: A systematic survey of prompting methods in natural language processing. ACM Comput. Surveys, 55(9):1-35, 2023. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 546, + 545, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 546, + 545, + 599 + ], + "spans": [ + { + "bbox": [ + 307, + 546, + 545, + 599 + ], + "type": "text", + "content": "[25] M Jehanzeb Mirza, Jakub Micorek, Horst Possegger, and Horst Bischof. The norm must go on: Dynamic unsupervised domain adaptation by normalization. In IEEE Conf. Comput. Vis. Pattern Recog., pages 14765-14775, 2022. 3, 6" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 602, + 545, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 602, + 545, + 656 + ], + "spans": [ + { + "bbox": [ + 307, + 602, + 545, + 656 + ], + "type": "text", + "content": "[26] Chaithanya Kumar Mummadi, Robin Hutmacher, Kilian Rambach, Evgeny Levinkov, Thomas Brox, and Jan Hendrik Metzen. Test-time adaptation to distribution shift by confidence maximization and input transformation. arXiv preprint arXiv:2106.14999, 2021. 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 658, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 658, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 307, + 658, + 545, + 712 + ], + "type": "text", + "content": "[27] Zachary Nado, Shreyas Padhy, D Sculley, Alexander D'Amour, Balaji Lakshminarayanan, and Jasper Snoek. Evaluating prediction-time batch normalization for robustness under covariate shift. arXiv preprint arXiv:2006.10963, 2020. 3, 4" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "11192" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "text", + "content": "[28] Phan Ngoc Lan, Nguyen Sy An, Dao Viet Hang, Dao Van Long, Tran Quang Trung, Nguyen Thi Thuy, and Dinh Viet Sang. Neounet: Towards accurate colon polyp segmentation and neoplasm detection. In Adv. Vis. Comput. Int. Symp., pages 15-28. Springer, 2021. 5" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 128, + 287, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 128, + 287, + 172 + ], + "spans": [ + { + "bbox": [ + 48, + 128, + 287, + 172 + ], + "type": "text", + "content": "[29] A Tuan Nguyen, Thanh Nguyen-Tang, Ser-Nam Lim, and Philip HS Torr. Tipi: Test time adaptation with transformation invariance. In IEEE Conf. Comput. Vis. Pattern Recog., pages 24162-24171, 2023. 2, 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 172, + 287, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 172, + 287, + 216 + ], + "spans": [ + { + "bbox": [ + 48, + 172, + 287, + 216 + ], + "type": "text", + "content": "[30] Shuaicheng Niu, Jiaxiang Wu, Yifan Zhang, Yafo Chen, Shijian Zheng, Peilin Zhao, and Mingkui Tan. Efficient test-time model adaptation without forgetting. In Int. Conf. Mach. Learn., pages 16888-16905. PMLR, 2022. 2, 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 217, + 287, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 217, + 287, + 259 + ], + "spans": [ + { + "bbox": [ + 48, + 217, + 287, + 259 + ], + "type": "text", + "content": "[31] Shuaicheng Niu, Jiaxiang Wu, Yifan Zhang, Zhiquan Wen, Yaofo Chen, Peilin Zhao, and Mingkui Tan. Towards stable test-time adaptation in dynamic wild world. In Int. Conf. Learn. Represent., 2023. 2, 3, 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 261, + 287, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 261, + 287, + 326 + ], + "spans": [ + { + "bbox": [ + 48, + 261, + 287, + 326 + ], + "type": "text", + "content": "[32] José Ignacio Orlando, Huazhu Fu, João Barbosa Breda, Karel van Keer, Deepti R Bathula, Andrés Diaz-Pinto, Ruogu Fang, Pheng-Ann Heng, Jyoung Kim, JoonHo Lee, et al. REFuge Challenge: A unified framework for evaluating automated methods for glaucoma assessment from fundus photographs. Med. Image Anal., 59:101570, 2020. 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 327, + 287, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 327, + 287, + 369 + ], + "spans": [ + { + "bbox": [ + 48, + 327, + 287, + 369 + ], + "type": "text", + "content": "[33] Xingang Pan, Ping Luo, Jianping Shi, and Xiaou Tang. Two at once: Enhancing learning and generalization capacities via ibn-net. In *Eur. Conf. Comput. Vis.*, pages 464-479, 2018. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 371, + 287, + 425 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 371, + 287, + 425 + ], + "spans": [ + { + "bbox": [ + 48, + 371, + 287, + 425 + ], + "type": "text", + "content": "[34] Swami Sankaranarayanan, Yogesh Balaji, Arpit Jain, Ser Nam Lim, and Rama Chellappa. Learning from synthetic data: Addressing domain shift for semantic segmentation. In IEEE Conf. Comput. Vis. Pattern Recog., pages 3752-3761, 2018. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 426, + 287, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 426, + 287, + 480 + ], + "spans": [ + { + "bbox": [ + 48, + 426, + 287, + 480 + ], + "type": "text", + "content": "[35] Steffen Schneider, Evgenia Rusak, Luisa Eck, Oliver Bringmann, Wieland Brendel, and Matthias Bethge. Improving robustness against common corruptions by covariate shift adaptation. Adv. Neural Inform. Process. Syst., 33:11539-11551, 2020. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 481, + 287, + 535 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 481, + 287, + 535 + ], + "spans": [ + { + "bbox": [ + 48, + 481, + 287, + 535 + ], + "type": "text", + "content": "[36] Juan Silva, Aymeric Histace, Olivier Romain, Xavier Dray, and Bertrand Granado. Toward embedded detection of polyps in wce images for early diagnosis of colorectal cancer. Int. J. Comput.-Assist. Rad. Sur., 9(2):283-293, 2014. 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 536, + 287, + 580 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 536, + 287, + 580 + ], + "spans": [ + { + "bbox": [ + 48, + 536, + 287, + 580 + ], + "type": "text", + "content": "[37] Samarth Sinha, Peter Gehler, Francesco Locatello, and Bernt Schiele. Test: Test-time self-training under distribution shift. In IEEE Wint. Conf. Appl. Comput. Vis., pages 2759-2769, 2023. 2, 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 581, + 287, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 581, + 287, + 635 + ], + "spans": [ + { + "bbox": [ + 48, + 581, + 287, + 635 + ], + "type": "text", + "content": "[38] Jayanthi Sivaswamy, SR Krishnadas, Gopal Datt Joshi, Madhulika Jain, and A Ujjwaft Syed Tabish. Drishti-GS: Retinal image dataset for optic nerve head (onh) segmentation. In IEEE Int. Symp. on Bio. Imaging, pages 53-56. IEEE, 2014. 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 635, + 287, + 680 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 635, + 287, + 680 + ], + "spans": [ + { + "bbox": [ + 48, + 635, + 287, + 680 + ], + "type": "text", + "content": "[39] Junha Song, Jungsoo Lee, In So Kweon, and Sungha Choi. Ecotta: Memory-efficient continual test-time adaptation via self-distilled regularization. In IEEE Conf. Comput. Vis. Pattern Recog., pages 11920-11929, 2023. 2, 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "type": "text", + "content": "[40] Yu Sun, Xiaolong Wang, Zhuang Liu, John Miller, Alexei Efros, and Moritz Hardt. Test-time training with self-supervision for generalization under distribution shifts. In" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 710 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 327, + 73, + 545, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 545, + 94 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 545, + 94 + ], + "type": "text", + "content": "Int. Conf. Mach. Learn., pages 9229-9248. PMLR, 2020. 2, 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 95, + 545, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 95, + 545, + 140 + ], + "spans": [ + { + "bbox": [ + 307, + 95, + 545, + 140 + ], + "type": "text", + "content": "[41] Dequan Wang, Evan Shelhamer, Shaoteng Liu, Bruno Olshausen, and Trevor Darrell. Tent: Fully test-time adaptation by entropy minimization. In Int. Conf. Learn. Represent., 2021. 2, 3, 5, 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 141, + 545, + 174 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 141, + 545, + 174 + ], + "spans": [ + { + "bbox": [ + 307, + 141, + 545, + 174 + ], + "type": "text", + "content": "[42] Qin Wang, Olga Fink, Luc Van Gool, and Dengxin Dai. Continual test-time domain adaptation. In IEEE Conf. Comput. Vis. Pattern Recog., pages 7201-7211, 2022. 2, 3, 6, 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 175, + 545, + 230 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 175, + 545, + 230 + ], + "spans": [ + { + "bbox": [ + 307, + 175, + 545, + 230 + ], + "type": "text", + "content": "[43] Shujun Wang, Lequan Yu, Kang Li, Xin Yang, Chi-Wing Fu, and Pheng-Ann Heng. Boundary and entropy-driven adversarial learning for fundus image segmentation. In Int. Conf. Med. Image Comput. Comput.-Assist. Intervention., pages 102-110. Springer, 2019. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 230, + 545, + 295 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 230, + 545, + 295 + ], + "spans": [ + { + "bbox": [ + 308, + 230, + 545, + 295 + ], + "type": "text", + "content": "[44] Wei Wang, Zhun Zhong, Weijie Wang, Xi Chen, Charles Ling, Boyu Wang, and Nicu Sebe. Dynamically instance-guided adaptation: A backward-free approach for test-time domain adaptive semantic segmentation. In IEEE Conf. Comput. Vis. Pattern Recog., pages 24090-24099, 2023. 2, 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 297, + 545, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 297, + 545, + 352 + ], + "spans": [ + { + "bbox": [ + 308, + 297, + 545, + 352 + ], + "type": "text", + "content": "[45] Hongzheng Yang, Cheng Chen, Meirui Jiang, Quande Liu, Jianfeng Cao, Pheng Ann Heng, and Qi Dou. Dltta: Dynamic learning rate for test-time adaptation on cross-domain medical images. IEEE Trans. Med. Imaging, 41(12):3575-3586, 2022. 2, 3, 4, 5, 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 353, + 545, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 353, + 545, + 387 + ], + "spans": [ + { + "bbox": [ + 308, + 353, + 545, + 387 + ], + "type": "text", + "content": "[46] Yanchao Yang and Stefano Soatto. FDA: Fourier domain adaptation for semantic segmentation. In IEEE Conf. Comput. Vis. Pattern Recog., pages 4085-4095, 2020. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 387, + 545, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 387, + 545, + 430 + ], + "spans": [ + { + "bbox": [ + 308, + 387, + 545, + 430 + ], + "type": "text", + "content": "[47] Jian Zhang, Lei Qi, Yinghuan Shi, and Yang Gao. Domain adaptor: A novel approach to test-time adaptation. In IEEE Conf. Comput. Vis. Pattern Recog., pages 18971-18981, 2023. 2, 3, 4, 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 431, + 545, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 431, + 545, + 464 + ], + "spans": [ + { + "bbox": [ + 308, + 431, + 545, + 464 + ], + "type": "text", + "content": "[48] Marvin Zhang, Sergey Levine, and Chelsea Finn. Memo: Test time robustness via adaptation and augmentation. Adv. Neural Inform. Process. Syst., 35:38629-38642, 2022. 2, 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 465, + 545, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 465, + 545, + 520 + ], + "spans": [ + { + "bbox": [ + 308, + 465, + 545, + 520 + ], + "type": "text", + "content": "[49] Yizhe Zhang, Tao Zhou, Yuhui Tao, Shuo Wang, Ye Wu, Benyuan Liu, Pengfei Gu, Qiang Chen, and Danny Z Chen. TestFit: A plug-and-play one-pass test time method for medical image segmentation. Medical Image Analysis, 92: 103069, 2024. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 521, + 545, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 521, + 545, + 586 + ], + "spans": [ + { + "bbox": [ + 308, + 521, + 545, + 586 + ], + "type": "text", + "content": "[50] Zhuo Zhang, Feng Shou Yin, Jiang Liu, Wing Kee Wong, Ngan Meng Tan, Beng Hai Lee, Jun Cheng, and Tien Yin Wong. ORIGA-light: An online retinal fundus image database for glaucoma analysis and research. In Int. Conf. IEEE Eng. in Med. and Bio., pages 3065-3068. IEEE, 2010. 5" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 587, + 545, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 587, + 545, + 632 + ], + "spans": [ + { + "bbox": [ + 308, + 587, + 545, + 632 + ], + "type": "text", + "content": "[51] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Zwei Liu. Conditional prompt learning for vision-language models. In IEEE Conf. Comput. Vis. Pattern Recog., pages 16816-16825, 2022. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 632, + 545, + 677 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 632, + 545, + 677 + ], + "spans": [ + { + "bbox": [ + 308, + 632, + 545, + 677 + ], + "type": "text", + "content": "[52] Chen Zhu, Renkun Ni, Zheng Xu, Kezhi Kong, W Ronny Huang, and Tom Goldstein. Gradinit: Learning to initialize neural networks for stable and efficient training. Adv. Neural Inform. Process. Syst., 34:16410-16422, 2021. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 677, + 545, + 710 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 677, + 545, + 710 + ], + "spans": [ + { + "bbox": [ + 308, + 677, + 545, + 710 + ], + "type": "text", + "content": "[53] Yukun Zuo, Hantao Yao, and Changsheng Xu. Attention-based multi-source domain adaptation. IEEE Trans. Image Process., 30:3793-3803, 2021. 2" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "11193" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/EarthLoc_ Astronaut Photography Localization by Indexing Earth from Space/09848c3c-ff6b-4e53-9236-ab0f624d8495_content_list.json b/2024/EarthLoc_ Astronaut Photography Localization by Indexing Earth from Space/09848c3c-ff6b-4e53-9236-ab0f624d8495_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..c0b2664dfd3f8717671a29305d43528f83b3c002 --- /dev/null +++ b/2024/EarthLoc_ Astronaut Photography Localization by Indexing Earth from Space/09848c3c-ff6b-4e53-9236-ab0f624d8495_content_list.json @@ -0,0 +1,1845 @@ +[ + { + "type": "text", + "text": "EarthLoc: Astronaut Photography Localization by Indexing Earth from Space", + "text_level": 1, + "bbox": [ + 88, + 130, + 880, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Gabriele Berton1 Alex Stoken2 Barbara Caputo1 Carlo Masone1 \n1Politecnico di Torino 2Jacobs Technology, NASA Johnson Space Center \ngabriele.berton@polito.it", + "bbox": [ + 196, + 179, + 772, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 268, + 313, + 284 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Astronaut photography, spanning six decades of human spaceflight, presents a unique Earth observations dataset with immense value for both scientific research and disaster response. Despite their significance, accurately localizing the geographical extent of these images, which is crucial for effective utilization, poses substantial challenges. Current, manual localization efforts are time-consuming, motivating the need for automated solutions. We propose a novel approach - leveraging image retrieval - to address this challenge efficiently. We introduce innovative training techniques which contribute to the development of a high-performance model, EarthLoc. We develop six evaluation datasets and perform a comprehensive benchmark comparing EarthLoc to existing methods, showcasing its superior efficiency and accuracy. Our approach marks a significant advancement in automating the localization of astronaut photography, which will help bridge a critical gap in Earth observations data. Code and datasets are available at https://github.com/gmberton/EarthLoc.", + "bbox": [ + 75, + 300, + 473, + 589 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 618, + 209, + 633 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Astronaut photography of Earth is a unique remote sensing dataset that spans 60 years of human spaceflight, offering a rare perspective on our planet to the public and valuable data to Earth and atmospheric science researchers. This dataset contains over 4.5 million images and is growing by the tens of thousands per month, as astronauts are continually tasked with taking new photographs that enable scientific research as well as assist in natural disaster response efforts in the wake of events like floods and wildfires. To effectively use these images, the geographical area depicted in them needs to be identified. Unfortunately, this task - Astronaut Photography Localization (APL) - is very challenging. For each photo, only a coarse estimate of location is known, given by the point on Earth directly under the International Space Station (ISS) at the time the photo is taken. This point - called the nadir - can be easily computed using the image's timestamp and the ISS's orbit path. However, two", + "bbox": [ + 75, + 643, + 472, + 902 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/306c03fd150f2bcdbeb0615054bd2cd5358be753de30f69f07101315d886e142.jpg", + "image_caption": [ + "Figure 1. Overview of the astronaut photography localization task. Astronauts take hundreds of photos a day from the International Space Station (ISS) cupola (top-left) with hand-held cameras. For each image (example bottom right), the geographic location depicted is not known, and needs to be searched for across a huge area centered at the ISS's (known) nadir point at the moment that the photo is taken. A simulated view of the astronaut's perspective when the ISS is above Europe is shown. The goal of our paper is to automate the task of localizing these images, which could be anywhere within the view. In the figure's example, the photo the astronaut took is indicated by the green line and shown in inset - other possible photo extents are in red, illustrating the wide array of potential locations to search." + ], + "image_footnote": [], + "bbox": [ + 547, + 267, + 846, + 539 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "images taken with the same nadir can be thousands of kilometers apart, as even a slight inclination of the astronaut's hand-held camera can move the image's location hundreds of kilometers in any direction, as depicted in Fig. 1. Localization must thus be performed over a wide area, and is additionally complicated by (i) astronauts using hand-held cameras and a variety of zoom lenses, (ii) the large, 2500 kilometer (km) visibility range in all directions, (iii) most photographs being oblique, and (iv) the Earth's appearance", + "bbox": [ + 496, + 763, + 893, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "12754", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "changing over time, mostly due to weather or seasonal variations. Despite these challenges, the high value of properly localized imagery for Earth science research and disaster response has lead to more than 300,000 [21] astronaut photos being manually localized by identifying the geographic coordinates of the photo's center pixel. This process can take minutes to hours per photo for experts in NASA's Earth Science and Remote Sensing Unit (ESRS) and citizen scientists through ESRS's Image Detective program. The resulting geo-located images have lead to a multitude of peer reviewed papers, technical reports, and articles about light pollution and urban planning [15, 22, 47, 53], atmospheric phenomena [28, 69], changes in land usage and glacial extent [40], and other Earth science topics [27, 48, 68]. Most importantly, astronaut photography has a fast response time, crucial for disaster management - astronauts are alerted of a disaster, provide photos to a ground crew which localizes them, and sends them to first responders [56]. For example, in 2013 this protocol was activated in response to Cyclone Haiyan in the Philippines, wildfires in Australia, flooding in China, Russia, Pakistan, and the USA, and multiple other events. [56].", + "bbox": [ + 76, + 90, + 472, + 422 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Even with the abundance and importance of satellite imagery in modern applications, astronaut photography fills an unserved gap among other remotely sensed data. Unlike satellites that nominally take top-down imagery in similar illumination conditions at fixed temporal intervals, astronaut photographs can show topography via oblique views from multiple orientations, are taken in various lighting conditions (including nighttime), and vary in focal length to show detail at different resolutions. These qualities produce a complementary data product that would be difficult or impossible to gather from traditional satellites. Additionally, having a human in the loop for data collection allows for real-time response to natural disasters as well as a natural sense to avoid clouds and other obstructions.", + "bbox": [ + 76, + 422, + 472, + 633 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Given the uniqueness and importance of astronaut photography, and the large amount of time spent by human experts on geo-localizing them, researchers have been exploring solutions to automate the task through computer vision methods. Previous works have shown promising results using a pairwise matching setup that iteratively compares an astronaut photo to satellite imagery [58]. Yet, such methods suffer from high latency, searching multiple directions sequentially and using compute-heavy, dense correspondences to determine matching. The time required to find the location of all 4.5 million images using these methods is estimated to be over 20 compute years [58]. Furthermore, low latency is of paramount importance for pho", + "bbox": [ + 76, + 633, + 472, + 832 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "tography in areas that are affected by natural disasters (e.g., hurricanes or wildfires), for which speedy localization can help real-time, on-the-ground operations.", + "bbox": [ + 501, + 89, + 895, + 133 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To overcome the high latency of current methods, we propose to instead localize astronaut photography through image retrieval, by matching each astronaut photograph to a worldwide database containing satellite images of known position. In reformulating this problem as a retrieval task, we encounter a range of new challenges, from efficiently training a robust model to achieving low-latency inference, as well as defining a success metric and creating evaluation sets, both of which are crucial pieces when assessing how well a given model will perform when deployed to localize astronaut photography.", + "bbox": [ + 503, + 136, + 893, + 303 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The ultimate goal of this paper is twofold: (i) provide an efficient method to localize the archive of 4.5M queries as well as all new, incoming imagery, so that researchers studying a given geographical area can have access to a large amount of imagery spanning many years; and (ii) prove the viability of world-wide APL through image retrieval (which has never been attempted and was believed to be infeasible4) in order to spark a new line of research, with direct benefit to all space-based photography and its users across the globe.", + "bbox": [ + 503, + 303, + 893, + 453 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To achieve its goal, this paper makes the following contributions:", + "bbox": [ + 503, + 454, + 890, + 482 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- we propose approaching APL through image retrieval;", + "- we develop novel training techniques which show large quantitative improvement on the task;", + "- we provide six evaluation sets and a large benchmark of results with methods from a variety of relevant domains;", + "- we show that a model trained with these new techniques, EarthLoc, allows us to localize a large number of images, widely outperforming all other methods while being fast and efficient." + ], + "bbox": [ + 500, + 484, + 890, + 618 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Finally, we note how EarthLoc is being used to localize ISS astronaut photographs, which are publicly available and conveniently searchable at https://eol.jsc.nasa.gov/ExplorePhotos/.", + "bbox": [ + 500, + 619, + 890, + 681 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 504, + 691, + 645, + 705 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Image Retrieval Image retrieval involves searching a database for images similar to a query image. Traditionally, methodologies combined hand-crafted local features like SIFT [37], SURF [8], RootSIFT [3] into a global embedding through means like Bag of Words [13], Fisher Vectors [42] or VLAD [29], to allow for fast retrieval through kNN. With the rise of deep learning, local features have been replaced by CNN-derived features [6], significantly", + "bbox": [ + 500, + 720, + 893, + 842 + ], + "page_idx": 1 + }, + { + "type": "page_footnote", + "text": "4 in [59] (Sec. 2.1) scientists at NASA claim that retrieval for APL is infeasible because we cannot precompute a database of reference features due to their belief that the database must be query-specific. Results show that our pipeline can overcome this issue.", + "bbox": [ + 500, + 849, + 893, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_footnote", + "text": "1https://eol.jsc.nasa.gov/BeyondThePhotography/ ImageDetective/", + "bbox": [ + 73, + 835, + 465, + 859 + ], + "page_idx": 1 + }, + { + "type": "page_footnote", + "text": "2https://eol.jsc.nasa.gov/AboutCEO/PubList.htm", + "bbox": [ + 94, + 862, + 457, + 875 + ], + "page_idx": 1 + }, + { + "type": "page_footnote", + "text": "3https : / / storymaps . arcgis . com / stories / 947eb734e811465cb0425947b16b62b3", + "bbox": [ + 80, + 876, + 467, + 898 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "12755", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "improving retrieval performance. Since then, most of the attention of the retrieval community has been focused on how to best aggregate feature maps and efficiently train robust neural networks. For feature map aggregation, some of the most notable examples in literature include Max pooling [44], Regional Max pooling (R-MAC) [61] and Generalized Mean (GeM) [43]. To train these retrieval models, a number of losses have been proposed, which can be generally grouped into two categories: stateful losses, which have weights whose size depends on the number of classes, like the Large-Margin Softmax [35], SphereFace [36], CosFace [65] and ArcFace [16]; and their stateless counterparts, like the Contrastive and Triplet losses, and the more recent Lifted Structure Loss [54], NTXentLoss [64], Multi-Similarity Loss[67], FastAP [11], Supervised Contrastive [31] and the Circle Loss [60].", + "bbox": [ + 75, + 90, + 472, + 333 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Geo-localization and retrieval with Aerial and Remote Sensing Imagery To the best of our knowledge, only two previous works have specifically addressed the localization problem for astronaut photography. Both rely on local features and compute pairwise comparisons between astronaut photography and geo-located reference imagery. Find My Astronaut Photo [58] focuses on daytime imagery, presenting an evaluation of multiple methods as well as the note that pre-trained self-supervised models were not suitable for a retrieval-based approach to this task. Schwind and Storch [49] instead address nighttime imagery, using synthetically generated street light maps as reference. Both works discuss the importance of proper orientation for matching and the high cost of pairwise comparisons.", + "bbox": [ + 75, + 354, + 468, + 566 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "More commonly, studies of aerial imagery focus on data from unmanned aerial vehicles (UAVs) and satellite platforms. Aerial geo-localization is an important aspect of UAV navigation systems. The University-1652 dataset [71] and associated challenge encourage work in cross-view geo-localization for drone-satellite imagery, with strong approaches modifying different parts of the general pipeline, from backbone [73] to training setup [18], to feature partitioning [66]. Most use contrastive and self-supervised losses in training. Other works address cross-view geo-localization in more extreme view point differences, like that between aerial and street views [17, 20, 50-52, 74, 75].", + "bbox": [ + 75, + 566, + 472, + 748 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Remotely sensed data from satellites, on the other hand, often comes with reliable geo-location information from the sensor itself. Extracting features from satellite imagery is still a common task, with such features most often trained on a pretext objective before use on downstream tasks [5, 12, 39, 45]. Occasionally, features are used for retrieval itself [55], to identify areas of the Earth that are similar in appearance and thus might have similar properties to be studied together. In this paper we take a separate direction from previous literature, and aim at using image", + "bbox": [ + 75, + 750, + 472, + 902 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/2f0079bf20b434a2931132166f9337d4bee16a0d1eb527d4a2af3b53155c734b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 532, + 89, + 640, + 145 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/fbf8302a98adbf5a1b2ba44d37a91e9fb5f5d1fb519f46a2036e34c5e5761641.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 642, + 89, + 750, + 145 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/db1eaf7a3f1e24fedaeff0859841a2163133c83faa378cdba6457089d880f879.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 751, + 89, + 859, + 145 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/c8f695f11d7b80967238d9fc90f867b4115cc3343d2cd411d2b4c44a0014c0cb.jpg", + "image_caption": [ + "Figure 2. Astronaut photo query examples, showcasing the large variability in covered area and appearance." + ], + "image_footnote": [], + "bbox": [ + 532, + 146, + 640, + 200 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/4b5b11011707b7e985bdae857a14c2f35e6ee8128514cba300809aeabcdf2064.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 642, + 146, + 750, + 200 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/0c5720bcdb102fd8a201f9a8149d19b734e9a4371b9eb9217cfa5434dac4a455.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 751, + 146, + 859, + 200 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "retrieval techniques for APL.", + "bbox": [ + 500, + 246, + 694, + 261 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Other localization tasks Various other localization challenges are approached through image retrieval, including visual place recognition (VPR) and visual localization. The former aims at coarsely localizing a given query by matching it to a database of geo-tagged photos, and is mostly studied within urban environments using street-view imagery [7, 63]. Performance is improved either by the use of smart aggregation (e.g. NetVLAD [4]), attention layers (e.g. CRN [32]), or through large-scale training [1, 2, 9, 34]. More recently, the task of universal visual place recognition has been proposed by AnyLoc [30], whose authors provide a model that performs competitively on a large range of scenarios, including aerial imagery. The task of visual localization is focused on finding the precise camera pose of an image, and is commonly approached through feature matching techniques [7, 38, 46, 59]. It can also be used for visual place recognition [24, 76], although this leads to a noticeable increase in runtime when compared to pure retrieval.", + "bbox": [ + 496, + 279, + 893, + 551 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Dataset", + "text_level": 1, + "bbox": [ + 500, + 564, + 588, + 579 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section we discuss the datasets that we use in this work. Specifically, in Sec. 3.1 we introduce the images that we need to localize, called queries; in Sec. 3.2 we outline the collection of the database, made of satellite images (i.e., taken automatically, not with a hand-held camera) of known location; and in Sec. 3.3 we describe the creation of the evaluation datasets that we use to understand the capabilities of our models, combining queries and database in a way that will reflect the real-world use case of our method.", + "bbox": [ + 496, + 589, + 890, + 726 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Queries", + "text_level": 1, + "bbox": [ + 500, + 734, + 596, + 750 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Queries are taken from the Gateway to Astronaut Photography of Earth $^{5}$ , a collection of over 4.5 million photos of Earth taken by astronauts on the International Space Station (Fig. 2). This unique photography setting yields only coarse location information for each photograph, namely the position on Earth below the ISS (the nadir point) at the time the photo was taken. We seek to find the geographic area on the Earth that each photo encompasses (the photo's location),", + "bbox": [ + 496, + 757, + 890, + 878 + ], + "page_idx": 2 + }, + { + "type": "page_footnote", + "text": "5https://eol.jsc.nasa.gov/", + "bbox": [ + 514, + 886, + 723, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "12756", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/0261613d2655e91a2e02144006f754a2b00be1f2ff93651d9d66a20711547ebe.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 109, + 88, + 189, + 152 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/4c1e94bbae9a8fd5ea35601da20353a5b45b2b86c83a6365111f7468fdc03865.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 192, + 88, + 272, + 151 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/2a76d37bed00f53ccbd74b9b783cdc18a4566330491d3f21b287de192959200a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 274, + 88, + 356, + 151 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/84371a693e602f29fc8e8ff84d60f67964dee33653ca998e538f56a1083a61f5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 357, + 88, + 437, + 151 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/99e2ceef17f8a5c6c099c9214ca03d72799641480fc5f27dc4f678d47ae0317c.jpg", + "image_caption": [ + "Figure 3. Examples of database images. Top: four images from different regions, with $25\\%$ or $50\\%$ overlap between any pair. The red dot in each image represents the same geographic point. Bottom: 4 images from the same region across different years." + ], + "image_footnote": [], + "bbox": [ + 109, + 156, + 189, + 220 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/e105281a6895053b103e5a7466d139bb4db6b2e4fd30ea96d5a80ca737dcffef.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 192, + 156, + 272, + 220 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/42f578694e46a3b97f73fc8651a7343093ec202fe314dbebf8425bad1d8022dd.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 274, + 156, + 356, + 220 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/1a5387552fd0583aacc6454dfae4af460e6e66d7915cef119a9c6b37c9c1c881.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 357, + 156, + 437, + 220 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "which can be thousands of kilometers away from the ISS nadir point. This distance of maximum visibility (horizon) can be quickly computed as", + "bbox": [ + 75, + 303, + 468, + 349 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nd _ {\\text {v i s i b l e}} = \\sqrt {2 R h + h ^ {2}} = 2 4 3 6 \\mathrm {k m} \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 156, + 359, + 468, + 378 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $R = 6317$ is Earth's radius and $h = 450$ is the ISS maximum orbiting altitude. We follow common practice and round this number to 2500 kilometers [58].", + "bbox": [ + 75, + 388, + 468, + 436 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Challenges to APL Some astronaut photo queries, taken with high focal length lenses, can cover only a few kilometers within the almost 20 million sq. km area visible to the astronaut taking the photo, which makes localizing the images extremely challenging. Additional hurdles presented by the imagery acquisition process include the varying quality of the images themselves due to motion blur from the fast moving space station, and occlusion of unique, location-identifying features by cloud cover, shadow, or other portions of the ISS (Fig. 2). More queries are shown in Sec. 10 of the Supplementary.", + "bbox": [ + 75, + 453, + 468, + 619 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "From the 4.5 million photos available, we consider only the day-time images that have a ground truth obtained by FMAP [58], and keep only those which cover areas between 5000 and 900,000 sq. km, in order to match the areas of database images (see Sec. 3.2). This leads to 17,764 queries which we can use for validation and testing.", + "bbox": [ + 75, + 619, + 468, + 710 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Database", + "text_level": 1, + "bbox": [ + 76, + 719, + 184, + 733 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To estimate the location of the queries through image retrieval, we need a worldwide database of images, where each image is labeled with its position. We build such a database from an open-source, composited, cloudless collection of Sentinel-2 satellite imagery6. This collection contains imagery of worldwide landmasses at resolutions up to 15 meters per pixel. Some examples are shown in Fig. 3.", + "bbox": [ + 75, + 741, + 468, + 847 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For our database we include any land area between latitudes $60^{\\circ}$ and $-60^{\\circ}$ (i.e., the area traversed by the ISS). The", + "bbox": [ + 75, + 848, + 468, + 877 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/33b3b1bed9735d37e312500862e5cda6e8b4d980cb872faa7031a76982e95ee0.jpg", + "image_caption": [ + "Figure 4. To create evaluation sets we choose all images that could contain a Point of Interest (POI) - all photos with nadir point within $d_{visible}$ . To localize all photos within this range, even if they do not contain the POI, we create a database that contains all areas visible from the nadir points of all selected photos, yielding a database area of about $5000\\mathrm{km}^2$ per POI." + ], + "image_footnote": [], + "bbox": [ + 504, + 90, + 887, + 215 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/14c0ad698dc9285b07d6ddfd2f69e54cd9bdd8709f140135ec9d59f9c88fce7c.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Name(Lat, Lon)# queries# DBInterest
Texas(30, -95)614234kOften photoed (many queries)
Alps(45, 10)239453kSci. interest - glacial change
California(38, -122)356830kDisaster response - wildfires
Gobi Desert(40, 105)72654kSci. interest - desertification
Amazon(-3, -60)68219kSci. interest - deforestation
Toshka Lakes(23, 30)216463kSci. interest - flood monitoring
", + "bbox": [ + 501, + 324, + 888, + 411 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Table 1. Evaluation sets. The six sets cover diverse geographies and relevant areas for research on topics like climate change and disaster management. DB stands for database.", + "bbox": [ + 498, + 415, + 890, + 457 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Sentinel-2 imagery is available as map tiles, and we use tiles at zoom levels 9, 10, 11, with resolutions ranging from $\\sim 300$ to $\\sim 75$ meters per pixel7, to ensure that the database encompasses the large scale range of the query images.", + "bbox": [ + 498, + 467, + 890, + 527 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "These parameters lead us to define roughly 175k regions, where each region $\\mathcal{R}$ is the projection of a square with known corner coordinates onto the Earth's surface. Regions have areas ranging from 5000 sq. km to 900,000 sq. km. We take partially (up to $50\\%$ ) overlapping regions so that each query has at least one image in the database with considerable overlap (Fig. 3 top). From each region we collect four images, one per year from 2018, 2019, 2020 and 2021, covering temporal changes in a given area (Fig. 3, bottom). This produces a database of 700k images.", + "bbox": [ + 496, + 527, + 892, + 679 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Unlike our queries, these database images have undergone post-processing that includes atmospheric corrections as well as cloud pixel minimization via composition of images from multiple days. Additionally, the database images are always taken nadir facing (i.e. perpendicular to Earth's surface) to minimize obliquity effects. Each database image has resolution $1024 \\times 1024$ . Further images from the database are shown in Sec. 10 of the Supplementary.", + "bbox": [ + 496, + 679, + 890, + 800 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Evaluation sets", + "text_level": 1, + "bbox": [ + 500, + 810, + 653, + 824 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "With the end goal of fast and accurate localization of the astronaut photography archive, we propose a deployment", + "bbox": [ + 498, + 834, + 890, + 864 + ], + "page_idx": 3 + }, + { + "type": "page_footnote", + "text": "$^{6}$ https://s2maps.eu", + "bbox": [ + 94, + 886, + 238, + 898 + ], + "page_idx": 3 + }, + { + "type": "page_footnote", + "text": "7https://wiki.openstreetmap.org/wiki/Zoom levels", + "bbox": [ + 500, + 875, + 890, + 898 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "12757", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/7e61e7cb16e152fa920fc6f0f1a235a889bbc949d1e81cd9b07bffdec444231e.jpg", + "image_caption": [ + "Figure 5. Training strategy. Naive batching produces too easy a task for the model to learn robust representations from. We increase difficulty by clustering batches by similarity and adding year-wise augmentations, which move images of different regions, but from the same year, closer together in feature space. The model must then learn representations that discern similarity in this more difficult context." + ], + "image_footnote": [], + "bbox": [ + 140, + 87, + 831, + 277 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "environment that takes advantage of the known nadir points and visibility limits associated with each image. So while localization can be performed by matching any given query against the whole world-wide database, a more suitable approach is to follow a divide and conquer paradigm by grouping queries according to their nadir, which allows us to use only a subset of the database for each search, increasing speed and improving results.", + "bbox": [ + 75, + 340, + 470, + 462 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We seek to create evaluation sets that mirror the situation seen in deployment, so we evaluate on sets of queries that have a nadir within a $2500\\mathrm{km}$ radius from a chosen Point of Interest (POI). We then build a database of images such that the database fully encompasses the area that could have been photographed from each collected query, which is an area $2*d_{visible}$ from the original POI, as shown in Fig. 4.", + "bbox": [ + 75, + 463, + 468, + 568 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Since we need the queries' ground truth position for evaluation, we choose queries that have already been localized with exhaustive image matching techniques [58]. For these queries, the full photo extent is publicly available8.", + "bbox": [ + 75, + 568, + 468, + 628 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To properly evaluate different methods, we choose six POIs that represent various geographies but all have scientific relevance, and center an evaluation set around each. These are summarized in Tab. 1.", + "bbox": [ + 75, + 628, + 468, + 688 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Method", + "text_level": 1, + "bbox": [ + 76, + 703, + 168, + 718 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The goal of this work is to estimate the location of each astronaut photo query, which we tackle through image retrieval: for any given query, the objective is to find the most similar image(s) from the database, and use the corresponding location labels to confidently localize the query. With this in mind, we aim to train a retrieval model on satellite imagery such that the model is robust to changes in scale, perspective, and color, as well as temporal/seasonal variation. Specifically, the model's task is to extract features", + "bbox": [ + 75, + 729, + 468, + 864 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "from each image. Then, for each query, the most similar database image is retrieved via kNN in feature space.", + "bbox": [ + 500, + 340, + 890, + 371 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In the next subsections we describe how we train our retrieval model using the proposed database.", + "bbox": [ + 498, + 372, + 890, + 402 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Baseline Training", + "text_level": 1, + "bbox": [ + 500, + 409, + 671, + 425 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We want our model to extract relevant features of satellite imagery, so a straightforward approach would be to train via contrastive learning: for a given batch of images (samples), generate a number of augmented views (typically two), and apply a contrastive loss to maximize the distance in feature space between different samples and minimize distance between different views of the same sample. However, this solution would not allow the model to learn robustness to temporal changes, because the augmented views come from the same sample, with no inherent temporal variation.", + "bbox": [ + 496, + 431, + 890, + 582 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Therefore, instead of producing multiple views of the same image via synthetic augmentation, we train the model by randomly selecting a number of regions, and for each region we consider its four images (a quadruplet), one per year. Within a batch of quadruplets, the four images from each region are mutual positives, whereas any two images from different regions are considered negatives: we can therefore consider each region as a class. This provides a natural way to address temporal variation. We can then train the model through a multi-similarity loss [67]. This idea is shown in Fig. 5 (left).", + "bbox": [ + 496, + 583, + 890, + 750 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Clustered Batches", + "text_level": 1, + "bbox": [ + 500, + 756, + 676, + 771 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The baseline paradigm simply creates batches of random regions – a single batch of quadruplets could contain samples depicting shorelines, islands, deserts, and mountains. The loss operates batchwise, so large intra-batch variability can lead the model to learn too swiftly to distinguish between such diverse regions. In these situations, training converges quickly and the model stops short of learning to discriminate between similar-looking, but distinct, regions.", + "bbox": [ + 496, + 779, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_footnote", + "text": "https://eol.jsc.nasa.gov/SearchPhotos/PhotosDatabaseAPI/(mlcoordTable)", + "bbox": [ + 76, + 875, + 467, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "12758", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To avoid this issue, we want to train with batches of highly similar images, which are still from different regions. To this end, we compute clusters of images with similar representations. Every $N_{iter}$ training iterations, we extract the features from each region with the trained model, using only images from a single year, and group them into $C$ clusters.", + "bbox": [ + 75, + 90, + 468, + 181 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Then, we create training batches using samples from a given cluster, choosing a different cluster for each batch. This increases the difficulty of the task in each batch, ensuring that the model learns to discriminate between similar-looking images of different regions. Note that clusters are not geographically related, but related in feature space. For example, a cluster may contain images from the Gobi and the Sahara deserts, which both appear similar in feature space despite being geographically distant. This procedure, which can be seen as an offline mining approach, increases training time by less than $5\\%$ but improves recall by roughly $15\\%$ on average (see Sec. 5.3), a worthwhile trade off.", + "bbox": [ + 75, + 181, + 470, + 363 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "This concept is illustrated in Fig. 5 (middle), and qualitative examples of images that are clustered together are shown in Sec. 11.2 of the Supplementary.", + "bbox": [ + 75, + 364, + 468, + 410 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Year-Wise Data Augmentation", + "text_level": 1, + "bbox": [ + 76, + 421, + 346, + 436 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Query images span the ongoing $20+$ years of ISS operations, as well as multiple camera types, camera operators, and view points, so queries depicting the same location can have a number of natural \"transformations\". These range from blueish hues from lens filters, to skewed representations in highly oblique shots, to seasonal variation in ground cover (e.g. snow, autumn colors).", + "bbox": [ + 75, + 445, + 468, + 549 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To make the model robust to such transformations, we employ data augmentation. Typically, augmentations are applied per image or per batch, but we argue that to truly learn robustness to the augmentations in this setting, the same augmentation should be performed on images from different regions, but from the same year.", + "bbox": [ + 75, + 551, + 468, + 642 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Formally, at each training iteration we choose a set of 4 random augmentations, one for each year $A_{2018}, A_{2019}, A_{2020}, A_{2021}$ . We then apply the corresponding augmentation to all images from that year (i.e. $A_{2018}$ is applied to all images from 2018, across all regions). Such \"year-wise\" augmentation has the effect of moving images from a given year closer in feature space to each other, as shown in Fig. 5 (right). In the loss, images from the same region, but different years, are treated as positives, while images from different regions are treated as negatives. This has the effect that images that receive the same augmentation are assigned different classes – forcing the model to learn to ignore such augmentations – and instead pull together images from the same region, across different years.", + "bbox": [ + 75, + 643, + 468, + 854 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Furthermore, given that the same augmentation is applied to multiple images, augmentation can be batched and performed on the GPU, making training slightly faster (by", + "bbox": [ + 75, + 854, + 468, + 900 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/57ed52742026028b9e6109c4422fbc58eb7526e4700d658a56c4124fc9e8dae4.jpg", + "image_caption": [ + "Figure 6. Dataset collection from satellite imagery and handling in Neutral-Aware Multi-Similarity Loss. For a given image (yellow), images with no overlap are negatives (red), images from the same region are positives (green, overlapping yellow), and images with partial overlap are neutral (grey). Note that the database is made of images at 3 different zooms (each double in size than the next), which explains the regions of different size." + ], + "image_footnote": [], + "bbox": [ + 527, + 93, + 867, + 271 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "$5\\%$ on average, although this depends on the hardware).", + "bbox": [ + 500, + 382, + 867, + 398 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.4. Neutral-Aware Multi-Similarity Loss", + "text_level": 1, + "bbox": [ + 500, + 407, + 818, + 422 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The powerful Multi-Similarity loss is designed to take as input multiple images from a number of classes, along with the corresponding class indices, and compute a loss from the positives and negatives for each image.", + "bbox": [ + 496, + 430, + 890, + 491 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "However, it does not take into account that in some cases, pairs of images are neither positives nor negatives, and should be considered \"neutral\". While the concept of neutral images does not exist in standard image retrieval tasks, as an image either belongs to class A or class B, in APL it must be taken into consideration. For example, should two images, which have an intersection over union of $25\\%$ , be considered positives or negatives? This is precisely the case when two images of the same area, but from different zoom levels, are batched together, as well as overlapping images at the same level (see Fig. 6). We believe the answer is neither, so we add the \"neutral\" case to the loss.", + "bbox": [ + 496, + 491, + 890, + 671 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Formally, we model the neutral case with the indicator function $\\mathcal{I}_{NA}$ , which is 0 when the regions $\\mathcal{R}_i$ and $\\mathcal{R}_j$ intersect but are not equal, and 1 otherwise.", + "bbox": [ + 496, + 671, + 890, + 717 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {I} _ {N A} \\left(\\mathcal {R} _ {i}, \\mathcal {R} _ {j}\\right) = \\mathbb {1} _ {\\left(\\mathcal {R} _ {i} \\cap \\mathcal {R} _ {j}\\right) \\wedge \\left(\\mathcal {R} _ {i} \\neq \\mathcal {R} _ {j}\\right)} \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 571, + 728, + 890, + 747 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "With this, the neutral-aware multi-similarity loss becomes", + "bbox": [ + 500, + 756, + 880, + 771 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {N A M S} = \\frac {1}{B S} \\sum_ {i = 1} ^ {B S} \\left\\{\\frac {1}{\\alpha} \\log \\left[ 1 + \\sum_ {k \\in \\mathcal {R} _ {i}} e ^ {- \\alpha \\left(\\mathcal {S} _ {i k} - \\lambda\\right)} \\right] \\right. \\\\ \\left. + \\frac {1}{\\beta} \\log \\left[ 1 + \\sum_ {k \\notin \\mathcal {R} _ {i}} \\mathcal {I} _ {N A} \\left(\\mathcal {R} _ {i}, \\mathcal {R} _ {k}\\right) e ^ {- \\beta \\left(\\mathcal {S} _ {i k} - \\lambda\\right)} \\right] \\right\\} \\end{array} \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 509, + 781, + 890, + 877 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $BS$ is the number of images in a batch, $\\mathcal{R}_i$ is the", + "bbox": [ + 500, + 885, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "12759", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/33800021aeed8899d218f9b4d84da704d50213491d242e1ce954c3dfc9503d46.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodType of Training ImageryTexas (validation)AlpsCaliforniaGobi DesertAmazonToshka Lakes
R@1R@10R@100R@1R@10R@100R@1R@10R@100R@1R@10R@100R@1R@10R@100R@1R@10R@100
Nadir-2.4--1.2--2.4--1.8--3.1--1.4--
Random Choice-0.21.715.50.11.111.60.22.320.10.11.013.20.11.111.50.21.29.1
NetVLAD [4]Urban VPR3.613.434.85.215.236.24.314.439.02.17.724.56.619.946.39.622.545.6
SFRS [23]Urban VPR6.817.440.97.017.841.45.417.244.45.214.737.69.423.852.510.820.442.4
Conv-AP [1]Urban VPR3.713.537.72.08.825.03.815.441.23.79.024.27.021.147.75.012.332.6
CosPlace [9]Urban VPR4.915.242.05.919.345.24.815.243.37.917.539.58.523.954.49.825.453.0
MixVPR [2]Urban VPR4.515.340.13.18.622.64.417.543.24.010.124.59.828.758.75.915.436.5
EigenPlaces [10]Urban VPR6.421.652.48.721.450.36.022.854.38.120.040.910.926.557.914.330.960.4
AnyLoc [30] (DINOv2 [41] + NetVLAD)Universal VPR44.168.787.840.770.892.048.775.091.628.757.081.738.663.886.263.784.596.3
TorchGeo [57] (ResNet50 w MOCO [26])Satellite1.03.211.60.31.35.71.14.316.40.42.88.31.96.619.40.72.99.9
TorchGeo [57] (ResNet50 w SeCo [70])Satellite6.115.641.77.420.249.25.114.537.13.714.038.94.613.332.95.715.638.5
TorchGeo [57] (ResNet50 w GASSL [5])Satellite9.722.846.49.123.150.513.331.458.86.317.545.48.320.340.120.438.664.2
OGCL UAV-View [18] (ConvNeXt-XXLarge)UAV16.235.365.814.434.164.719.642.171.05.520.545.310.026.354.721.138.463.4
OGCL UAV-View [18] (ViT-L/14) [19]UAV17.633.255.914.633.263.922.848.174.97.622.850.120.439.162.531.851.874.4
MBEG [73] (ViT-L/14)UAV7.017.635.16.619.345.98.720.741.54.415.038.06.417.139.38.120.749.1
EarthLoc (Ours)Satellite55.973.088.358.476.889.558.076.091.451.167.586.547.267.984.672.285.093.3
", + "bbox": [ + 78, + 88, + 893, + 241 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 2. Results from different methods on our query sets. Methods are grouped into (1) naive, (2) VPR, (3) Remote Sensing (UAV+Satellite), (4) other baselines trained on our database with clustered batches (Sec. 4.2). For all methods we use $4x90TTA$ . VPR stands for Visual Place Recognition, UAV for Unmanned Aerial Vehicle. Best results overall in bold, best per-group underlined. EarthLoc outperforms other models, and performs competitively with AnyLoc, while having 50 times faster features extraction and 10 times smaller features.", + "bbox": [ + 75, + 247, + 893, + 316 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "region of image $i$ , $\\mathcal{S}_{ik}$ is the similarity between image $i$ and $j$ , $\\lambda$ is a margin, and $\\alpha, \\beta$ are hyperparameters.", + "bbox": [ + 75, + 325, + 468, + 356 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We believe that the neutral-aware multi-similarity loss is applicable to any task where it is not straightforward how to enforce full positivity or negativity between any pair of images, including localization tasks like VPR.", + "bbox": [ + 75, + 357, + 468, + 417 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 76, + 431, + 209, + 450 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.1. Experimental setting", + "text_level": 1, + "bbox": [ + 76, + 457, + 274, + 474 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.1.1 Training", + "text_level": 1, + "bbox": [ + 76, + 481, + 192, + 498 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In training, we use a batch size of 96 quadruplets (384 images), where one quadruplet corresponds to four images within a region. Our model has a MixVPR-style architecture [2], with a ResNet50 [25] backbone and an MLP-Mixer[62] with output dimension 4096. We train for 10k iterations with the Adam optimizer [33] and learning rate 0.0001. Clustering (see Sec. 4.2) is performed every 2k iterations with the number of clusters $C = 50$ . As augmentation, we use torchvision's color jittering, random perspective, and random rotation. Training takes between 5 and 7 hours on an A100 GPU for all methods described in Sec. 4.", + "bbox": [ + 75, + 506, + 470, + 672 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.1.2 Inference", + "text_level": 1, + "bbox": [ + 76, + 694, + 197, + 708 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Due to the imagery acquisition conditions, we cannot assume any canonical orientation for the photos (in space, there's no gravity vector, i.e., no \"up\" or \"down\").", + "bbox": [ + 75, + 718, + 468, + 763 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "This poses a challenge for any retrieval method – to address it, we generate four feature embeddings for each database image, one each after rotating the image for angles $r \\in [0,90,180,270]$ , in a process we call $4 \\times 90^{\\circ}$ test time augmentation (4x90TTA). For fairness, we do this for all methods in all experiments, given that the models are not trained to be robust to such large arbitrary rotations. This increases latency by $4\\mathrm{x}$ , but the method is still orders of magnitude faster than previous approaches (retrieval takes 0.05", + "bbox": [ + 75, + 765, + 468, + 901 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "seconds per query with $4x90TTA$ , while previous works required $\\sim 1$ minute [58] per query). Memory requirements also increase by four times, although this is far from being an issue for a standard workstation, since all database features for the whole Earth fit in less than 12GB of RAM. Experiments without $4x90TTA$ are shown in the Supplementary in Sec. 7. As an added bonus, since each feature from the database refers to an image with a certain orientation (w.r.t. North), retrieval predictions also give an estimate of the orientation of the query. We use features from the 2021 satellite imagery as our global database.", + "bbox": [ + 496, + 325, + 893, + 492 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Success Criteria With the final goal of localizing the 4.5M astronaut photograph queries in mind, we define a predicted image as correct if there is non-zero overlap between the prediction and the query. As a metric we use Recall@N, defined as the percentage of queries for which at least one of the top-N predictions is correct.", + "bbox": [ + 496, + 515, + 893, + 608 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.2. Results", + "text_level": 1, + "bbox": [ + 500, + 618, + 591, + 633 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In Tab. 2 we present results from a large number of models from multiple domains. We first show the recall when naively predicting the query to be directly nadir, proving that the vast majority of images are taken at an angle, and the poor performance of randomly choosing a prediction from the database - a symptom of the large-scale nature of the task (i.e., localizing a 5000 sq.km wide image within a 20M sq.km area).", + "bbox": [ + 496, + 643, + 893, + 763 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Results with methods trained for Urban Visual Place Recognition (VPR) provide, somewhat unsurprisingly, poor results - on the other hand, AnyLoc [30] shows remarkably strong results, correctly predicting the location of the query as the first prediction for roughly $50\\%$ of the queries across all evaluation sets, although at the cost of over 50x slower feature extraction and 10x bigger features than its VPR counterparts (AnyLoc uses 49152-D features vs 4096 of NetVLAD).", + "bbox": [ + 496, + 763, + 893, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "12760", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/838e8fba8270e7df4d564df08bc7499eca19f97f5fde4450f061e8403921ffaf.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Multi-similarity LossUsing regional quadrupletsClustered BatchesYear-Wise Data AugmentationNeutral-Aware Multi-SimilarityTexas (validation)AlpsCaliforniaGobi DesertAmazonToshka LakesAverage
R@1R@100R@1R@100R@1R@100R@1R@100R@1R@100R@1R@100R@1R@100
✓25.571.424.877.327.876.616.868.823.661.538.880.826.272.7
✓✓45.084.145.487.349.588.535.779.934.280.262.991.145.485.2
✓✓✓51.487.549.689.553.990.640.884.339.382.864.093.049.888.0
✓✓✓✓54.687.553.987.255.991.646.882.945.682.467.691.954.187.2
✓✓✓55.388.255.190.057.290.344.685.848.484.868.992.954.988.7
✓✓✓✓55.988.358.489.558.091.451.186.547.284.672.293.357.188.9
", + "bbox": [ + 81, + 88, + 890, + 159 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 3. Ablations over test sets. EarthLoc uses Clustered Batches (Sec. 4.2), Year-Wise Data Augmentation (Sec. 4.3) and Neutral-Aware Multi-Similarity Loss (Sec. 4.4). Best results in bold.", + "bbox": [ + 75, + 167, + 892, + 196 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/18d34865b1211f2e2c026250989968e605577ab93e31d82c1152e596d2fbd645.jpg", + "image_caption": [ + "Figure 7. Qualitative results. Each group of four images represents a query (left) and its top-3 predictions. Green and red indicate if the prediction is correct or incorrect." + ], + "image_footnote": [], + "bbox": [ + 78, + 208, + 468, + 289 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We then test a number of models trained with self-supervised techniques, like MOCO [26], SeCo [70], and Geography-aware self-supervised learning (GASSL) [5], conveniently provided by the TorchGeo library [57]. Surprisingly, these methods did not provide competitive results, and are on average about equal with Urban VPR methods.", + "bbox": [ + 75, + 362, + 468, + 452 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "From the Unmanned Aerial Vehicle (UAV) domain, we use the latest state of the art (SOTA) methods, i.e., the winners of the recent challenge in UAV localization at ACM MM (October 2023) [72], namely OGCL [18] and MBEG [73]. Despite being SOTA in the related task of UAV localization, we see that these methods are not competitive with AnyLoc, despite outperforming most other methods.", + "bbox": [ + 75, + 455, + 468, + 561 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Finally, EarthLoc outperforms all previous works, localizing $57.1\\%$ of queries within the top-1 prediction (Recall@1) averaged on the 6 evaluation sets.", + "bbox": [ + 75, + 566, + 468, + 612 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.3. Ablations", + "text_level": 1, + "bbox": [ + 76, + 631, + 189, + 645 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In Tab. 3 we ablate the components of our method to better understand performance. Results show that not only do Clustered Batches, Year-Wise Data Augmentation and Neutral-Aware Multi-Similarity Loss provide significant increases over the baseline, but they are also orthogonal improvements and can be easily combined.", + "bbox": [ + 75, + 657, + 468, + 748 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.4. Qualitative Analysis and Failure Cases", + "text_level": 1, + "bbox": [ + 76, + 767, + 410, + 782 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "A sample of qualitative results from EarthLoc are shown in Fig. 7. The samples show that EarthLoc picks locations that have similar characteristics to the query (top left), is robust to rotations and changes in scale (bottom right), and provides correct predictions that are difficult to correctly match even for the human eye (top left, bottom left). More qualitative results in Sec. 11 of the Supplementary.", + "bbox": [ + 75, + 794, + 468, + 902 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.5. Limitations and Future Work", + "text_level": 1, + "bbox": [ + 500, + 210, + 764, + 226 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "While this work takes a major step forward in the study of APL, EarthLoc focuses on daytime queries, and we did not investigate panoramic or nighttime imagery. The localization of nighttime astronaut photos in particular is important (e.g. to study light pollution [14]), but we believe this to be a separate challenge that should be addressed in future work. Another limitation is due to the zoom limits of the database: we chose database tiles at zoom levels 9, 10, 11 (see Sec. 3.2), which allows us to localize queries of area between 5000 and 900,000 square km.", + "bbox": [ + 496, + 236, + 893, + 386 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The originality of this work opens up a large number of possible future research directions, like training directly on (some of the) localized queries, using seasonal data, post-processing predictions, obtaining pixel-wise geolocation, applying domain adaptation, and using available metadata (like camera lens) during training, to name a few.", + "bbox": [ + 496, + 388, + 893, + 479 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusions", + "text_level": 1, + "bbox": [ + 500, + 497, + 627, + 512 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We introduce EarthLoc, a method for localizing astronaut photography of Earth. By reframing APL from an image matching task to an image retrieval task, and introducing a custom loss function and training scheme, our method efficiently determines the geographic location of astronaut photography given only the position of the International Space Station at the time the photo was taken. Our main contributions are the Neutral-Aware Multi-Similarity Loss, Year-Wise Data Augmentation technique, and new astronaut photography-oriented validation and test datasets to encourage future work on this problem.", + "bbox": [ + 496, + 523, + 890, + 690 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements. We acknowledge the CINECA award under the ISCRA initiative, for the availability of high performance computing resources. This work was supported by CINI. Project supported by ESA Network of Resources Initiative. This study was carried out within the project FAIR - Future Artificial Intelligence Research - and received funding from the European Union Next-GenerationEU (PIANO NAZIONALE DI RIPRESA E RESILLENZA (PNRR) - MISSIONE 4 COMPONENTE 2, INVESTIMENTO 1.3 - D.D. 1555 11/10/2022, PE00000013). This manuscript reflects only the authors' views and opinions, neither the European Union nor the European Commission can be considered responsible for them. European Lighthouse on Secure and Safe AI - ELSA, HORIZON EU Grant ID: 101070617", + "bbox": [ + 496, + 705, + 892, + 898 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "12761", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Amar Ali-bey, Brahim Chaib-draa, and Philippe Giguère. Gsv-cities: Toward appropriate supervised visual place recognition. Neurocomputing, 513:194-203, 2022. 3, 7", + "[2] Amar Ali-bey, Brahim Chaib-draa, and Philippe Giguère. Mixvpr: Feature mixing for visual place recognition. In WACV, pages 2998-3007, 2023. 3, 7", + "[3] R. Arandjelović and Andrew Zisserman. Three things everyone should know to improve object retrieval. In CVPR, pages 2911-2918, 2012. 2", + "[4] Relja Arandjelović, Petr Gronat, Akihiko Torii, Tomas Pajdla, and Josef Sivic. NetVLAD: CNN architecture for weakly supervised place recognition. IEEE TPAMI, 40(6): 1437-1451, 2018. 3, 7", + "[5] Kumar Ayush, Burak Uzkent, Chenlin Meng, Kumar Tanmay, Marshall Burke, David Lobell, and Stefano Ermon. Geography-aware self-supervised learning. ICCV, 2021. 3, 7, 8", + "[6] Artem Babenko, Anton Slesarev, A. Chigorin, and V. Lempitsky. Neural codes for image retrieval. ECCV, abs/1404.1777, 2014. 2", + "[7] Giovanni Barbarani, Mohamad Mostafa, Hajali Bayramov, Gabriele Trivigno, Gabriele Berton, Carlo Masone, and Barbara Caputo. Are local features all you need for cross-domain visual place recognition? In CVPRW, pages 6155-6165, 2023. 3", + "[8] Herbert Bay, Andreas Ess, Tinne Tuytelaars, and Luc Van Gool. Speeded-up robust features (surf). Computer Vision and Image Understanding, 110:346-359, 2008. 2", + "[9] Gabriele Berton, Carlo Masone, and Barbara Caputo. Rethinking visual geo-localization for large-scale applications. In CVPR, 2022. 3, 7", + "[10] Gabriele Berton, Gabriele Trivigno, Barbara Caputo, and Carlo Masone. Eigenplaces: Training viewpoint robust models for visual place recognition. In ICCV, pages 11080-11090, 2023. 7", + "[11] Fatih Cakir, Kun He, Xide Xia, Brian Kulis, and Stan Sclaroff. Deep metric learning to rank. In CVPR, pages 1861-1870, 2019. 3", + "[12] Yezhen Cong, Samar Khanna, Chenlin Meng, Patrick Liu, Erik Rozi, Yutong He, Marshall Burke, David B. Lobell, and Stefano Ermon. SatMAE: Pre-training transformers for temporal and multi-spectral satellite imagery. In NeurIPS, 2022. 3", + "[13] Gabriela Csurka, Christopher Dance, Lixin Fan, Jutta Willamowski, and Cedric Bray. Visual categorization with bags of keypoints. In ECCV, 2004. 2", + "[14] Alejandro Sánchez de Miguel, José Gómez Castaño, Jaime Zamorano, Sergio Pascual, M Ángeles, L Cayuela, Guillermo Martín Martínez, Peter Challupner, and Christopher C M Kyba. Atlas of astronaut photos of earth at night. *Astronomy and Geophysics*, 55(4):4.36-4.36, 2014. 8", + "[15] Alejandro Sánchez de Miguel, Jonathan Bennie, Emma Rosenfeld, Simon Dzurjak, and Kevin J. Gaston. Environmental risks from artificial nighttime lighting widespread and increasing across Europe. Science Advances, 8(37): eabl6891, 2022. 2" + ], + "bbox": [ + 78, + 114, + 468, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[16] Jiankang Deng, J. Guo, and S. Zafeiriou. ArcFace: Additive angular margin loss for deep face recognition. In CVPR, pages 4685-4694, 2019. 3", + "[17] Fabian Deuser, Konrad Habel, and Norbert Oswald. Sample4geo: Hard negative sampling for cross-view geolocation. In ICCV, pages 16847-16856, 2023. 3", + "[18] Fabian Deuser, Konrad Habel, Martin Werner, and Norbert Oswald. Orientation-guided contrastive learning for uav-view geo-localisation. In Proceedings of the 2023 Workshop on UAVs in Multimedia: Capturing the World from a New Perspective, page 7-11, New York, NY, USA, 2023. Association for Computing Machinery. 3, 7, 8", + "[19] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale. In ICLR, 2021. 7", + "[20] Florian Fervers, Sebastian Bullinger, Christoph Bodensteiner, Michael Arens, and Rainer Stiefelhagen. Uncertainty-aware vision-based metric cross-view geolocation. In CVPR, pages 21621-21631, 2023. 3", + "[21] Kenton Fisher, Sara Schmidt, and Alex Stoken. Crew earth observations: New tools to support your research. In 12th Annual International Space Station Research and Development Conference. Center for the Advancement of Science in Space, Inc., 2023. 2", + "[22] Kevin J. Gaston and Alejandro Sánchez de Miguel. Environmental impacts of artificial light at night. Annual Review of Environment and Resources, 47(1):373-398, 2022. 2", + "[23] Yixiao Ge, Haibo Wang, Feng Zhu, Rui Zhao, and Hongsheng Li. Self-supervising fine-grained region similarities for large-scale image localization. In ECCV, pages 369-386, Cham, 2020. Springer International Publishing. 7", + "[24] Stephen Hausler, Sourav Garg, Ming Xu, Michael Milford, and Tobias Fischer. Patch-netvlad: Multi-scale fusion of locally-global descriptors for place recognition. In CVPR, pages 14141-14152, 2021. 3", + "[25] K. He, X. Zhang, S. Ren, and J. Sun. Deep residual learning for image recognition. In CVPR, pages 770-778, 2016. 7", + "[26] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In CVPR, pages 9726-9735, 2020. 7, 8", + "[27] Christopher. R. Jackson and Werner Alpers. The role of the critical angle in brightness reversals on sunglint images of the sea surface. Journal of Geophysical Research: Oceans, 115(C9), 2010. 2", + "[28] Augustin Juhl, Thomas Farges, and Elisabeth Blanc. Color pictures of sprites from non-dedicated observation on board the international space station. Journal of Geophysical Research: Space Physics, 118(1):454-461, 2013. 2", + "[29] Hervé Jégou, Matthijs Douze, Jorge Sánchez, Patrick Perez, and Cordelia Schmid. Aggregating local image descriptors into compact codes. IEEE TPAMI, 34, 2011. 2", + "[30] Nikhil Keetha, Avneesh Mishra, Jay Karhade, Krishna Murthy Jatavallabhula, Sebastian Scherer, Madhava" + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "12762", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Krishna, and Sourav Garg. Anyloc: Towards universal visual place recognition. IEEE Robotics and Automation Letters, 2023. 3, 7", + "[31] Prannay Khosla, Piotr Teterwak, Chen Wang, Aaron Sarna, Yonglong Tian, Phillip Isola, Aaron Maschinot, Ce Liu, and Dilip Krishnan. Supervised contrastive learning. In NeurIPS, 2020. 3", + "[32] Hyo Jin Kim, Enrique Dunn, and Jan-Michael Frahm. Learned contextual feature reweighting for image geolocation. In CVPR, pages 3251-3260, 2017. 3", + "[33] Diederik Kingma and Jimmy Ba. Adam: A method for stochastic optimization. *ICLR*, 2014. 7", + "[34] María Leyva-Vallina, Nicola Strisciuglio, and Nicolai Petkov. Data-efficient large scale place recognition with graded similarity supervision. CVPR, 2023. 3", + "[35] Weiyang Liu, Yandong Wen, Zhiding Yu, and Meng Yang. Large-margin softmax loss for convolutional neural networks. In ICML, pages 507-516. PMLR, 2016. 3", + "[36] Weiyang Liu, Yandong Wen, Zhiding Yu, Ming Li, Bhiksha Raj, and Le Song. Spherface: Deep hypersphere embedding for face recognition. In CVPR, pages 212-220, 2017. 3", + "[37] David G. Lowe. Distinctive image features from scale-invariant keypoints. Int. J. Comput. Vision, 60(2):91-110, 2004. 2", + "[38] Carlo Masone and Barbara Caputo. A survey on deep visual place recognition. IEEE Access, 9:19516-19547, 2021. 3", + "[39] Oscar Manas, Alexandre Lacoste, Xavier Giroi-nieto, David Vazquez, and Pau Rodriguez. Seasonal contrast: Unsupervised pre-training from uncurated remote sensing data. In ICCV, pages 9394-9403, 2021. 3", + "[40] Aadil Nathani, Rishi Iyer, Annabelle Wang, Aarnav Chitari, Adele Wilson, and Hannah Norris. Observing Earth From Space: Using Astronaut Photography to Analyze Geographical Climate Patterns. In AGU Fall Meeting Abstracts, pages ED44C-06, 2022. 2", + "[41] Maxime Oquab, Timothée Darcet, Theo Moutakanni, Huy V. Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, Russell Howes, Po-Yao Huang, Hu Xu, Vasu Sharma, Shang-Wen Li, Wojciech Galuba, Mike Rabbat, Mido Assran, Nicolas Ballas, Gabriel Synnaeve, Ishan Misra, Herve Jegou, Julien Mairal, Patrick Labatut, Armand Joulin, and Piotr Bojanowski. Dinov2: Learning robust visual features without supervision, 2023. 7", + "[42] Florent Perronnin, Yan Liu, Jorge Sánchez, and Herve Poirier. Large-scale image retrieval with compressed fisher vectors. In CVPR, pages 3384-3391, 2010. 2", + "[43] F. Radenović, G. Tolias, and O. Chum. Fine-tuning CNN Image Retrieval with No Human Annotation. IEEE TPAMI, 2018. 3", + "[44] A. Razavian, J. Sullivan, A. Maki, and S. Carlsson. Visual Instance Retrieval with Deep Convolutional Networks. CoRR, abs/1412.6574, 2015. 3", + "[45] Esther Rolf, Jonathan Proctor, Tamma Carleton, Ian Bolliger, Vaishaal Shankar, Miyabi Ishihara, Benjamin Recht, and Solomon Hsiang. A generalizable and accessible approach to machine learning with global satellite imagery. Nature Communications, 12(1), 2021. 3" + ], + "bbox": [ + 78, + 90, + 468, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[46] Paul-Edouard Sarlin, Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. SuperGlue: Learning feature matching with graph neural networks. In CVPR, 2020. 3", + "[47] Aaron Schirmer, Caleb Gallemore, Ting Liu, Seth Magle, Elisabeth DiNello, Humerah Ahmed, and Thomas Gilday. Mapping behaviorally relevant light pollution levels to improve urban habitat planning. Scientific Reports, 9, 2019. 2", + "[48] Johannes A. Schultz, Maik Hartmann, Sascha Heinemann, Jens Janke, Carsten Jürgens, Dieter Oertel, Gernot Rücker, Frank Thonfeld, and Andreas Rienow. Diego: A multispectral thermal mission for earth observation on the international space station. European Journal of Remote Sensing, 53(sup2):28-38, 2020. 2", + "[49] Peter Schwind and Tobias Storch. Georeferencing urban nighttime lights imagery using street network maps. Remote Sensing, 14(11), 2022. 3", + "[50] Yujiao Shi and Hongdong Li. Beyond cross-view image retrieval: Highly accurate vehicle localization using satellite image. In CVPR, pages 17010-17020, 2022. 3", + "[51] Yujiao Shi, Xin Yu, Shan Wang, and Hongdong Li. Cvlnet: Cross-view feature correspondence learning for video-based camera localization. In ACCV, pages 652-669, 2022.", + "[52] Yujiao Shi, Fei Wu, Akhil Perincherry, Ankit Vora, and Hongdong Li. Boosting 3-dof ground-to-satellite camera localization accuracy via geometry-guided cross-view transformer. In ICCV, pages 21516-21526, 2023. 3", + "[53] Christopher Small. Spectrometry of the urban lightscape. Technologies, 10(4), 2022. 2", + "[54] H. Song, Y. Xiang, S. Jegelka, and S. Savarese. Deep metric learning via lifted structured feature embedding. In CVPR, 2016. Spotlight presentation. 3", + "[55] Weiwei Song, Zhi Gao, Renwei Dian, Pedram Ghamisi, Yongjun Zhang, and Jón Atli Benediktsson. Asymmetric hash code learning for remote sensing image retrieval. IEEE Transactions on Geoscience and Remote Sensing, 60:1-14, 2022. 3", + "[56] W. L. Stefanov and C. A. Evans. Data collection for disaster response from the international space station. The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, XL-7/W3:851–855, 2015. 2", + "[57] Adam J. Stewart, Caleb Robinson, Isaac A. Corley, Anthony Ortiz, Juan M. Lavista Ferres, and Arindam Banerjee. Torch-Geo: Deep learning with geospatial data. In Proceedings of the 30th International Conference on Advances in Geographic Information Systems, pages 1–12, Seattle, Washington, 2022. Association for Computing Machinery. 7, 8", + "[58] Alex Stoken and Kenton Fisher. Find my astronaut photo: Automated localization and georectification of astronaut photography. In CVPRW, pages 6196-6205, 2023. 2, 3, 4, 5, 7", + "[59] Jiaming Sun, Zehong Shen, Yuang Wang, Hujun Bao, and Xiaowei Zhou. LoFTR: Detector-free local feature matching with transformers. CVPR, 2021. 3", + "[60] Yifan Sun, Changmao Cheng, Yuhan Zhang, Chi Zhang, Liang Zheng, Zhongdao Wang, and Yichen Wei. Circle loss: A unified perspective of pair similarity optimization. CVPR, pages 6397-6406, 2020. 3" + ], + "bbox": [ + 501, + 92, + 890, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "12763", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[61] Giorgos Tolias, R. Sicre, and H. Jégou. Particular object retrieval with integral max-pooling of CNN activations. CoRR, abs/1511.05879, 2016. 3", + "[62] Ilya O Tolstikhin, Neil Houlsby, Alexander Kolesnikov, Lucas Beyer, Xiaohua Zhai, Thomas Unterthiner, Jessica Yung, Andreas Steiner, Daniel Keysers, Jakob Uszkoreit, Mario Lucic, and Alexey Dosovitskiy. Mlp-mixer: An all-mlx architecture for vision. In NeurIPS, pages 24261–24272. Curran Associates, Inc., 2021. 7", + "[63] Gabriele Trivigno, Gabriele Berton, Juan Aragon, Barbara Caputo, and Carlo Masone. Divide&classify: Fine-grained classification for city-wide visual geo-localization. In ICCV, pages 11142-11152, 2023. 3", + "[64] Aäron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding. CoRR, abs/1807.03748, 2018. 3", + "[65] Hao Wang, Yitong Wang, Zheng Zhou, Xing Ji, Dihong Gong, Jingchao Zhou, Zhifeng Li, and Wei Liu. Cosface: Large margin cosine loss for deep face recognition. In CVPR, pages 5265-5274. Computer Vision Foundation / IEEE Computer Society, 2018. 3", + "[66] Tingyu Wang, Zhedong Zheng, Chenggang Yan, jiyong Zhang, Yaoqi Sun, Bolun Zheng, and Yi Yang. Each part matters: Local patterns facilitate cross-view geo-localization. IEEE TCSVT, 2021. doi: 10.1109/TCSVT.2021.3061265.3", + "[67] Xun Wang, Xintong Han, Weilin Huang, Dengke Dong, and Matthew R Scott. Multi-similarity loss with general pair weighting for deep metric learning. In CVPR, pages 5022-5030, 2019. 3, 5", + "[68] M. Justin Wilkinson and Yanni Gunnell. Fluvial Megafans on Earth and Mars. Cambridge University Press, 2023. 2", + "[69] Yoav Yair, Melody Korman, Colin Price, and Eytan Stibbe. Observing lightning and transient luminous events from the international space station during ilan-es: An astronaut's perspective. Acta Astronautica, 211:592-599, 2023. 2", + "[70] Ting Yao, Yiheng Zhang, Zhaofan Qiu, Yingwei Pan, and Tao Mei. Seco: Exploring sequence supervision for unsupervised representation learning. In AAAI, 2021. 7, 8", + "[71] Zhedong Zheng, Yunchao Wei, and Yi Yang. University-1652: A multi-view multi-source benchmark for drone-based geo-localization. ACM MM, 2020. 3", + "[72] Zhedong Zheng, Yu jiao Shi, Tingyu Wang, Jun Liu, Jianwu Fang, Yunchao Wei, and Tat-seng Chua. Uavm '23: 2023 workshop on uavs in multimedia: Capturing the world from a new perspective. In ACM MM, page 9715-9717, New York, NY, USA, 2023. Association for Computing Machinery. 8", + "[73] Runzhe Zhu, Mingze Yang, Kaiyu Zhang, Fei Wu, Ling Yin, and Yujin Zhang. Modern backbone for efficient geolocation. In Proceedings of the 2023 Workshop on UAVs in Multimedia: Capturing the World from a New Perspective, page 31-37, New York, NY, USA, 2023. Association for Computing Machinery. 3, 7, 8", + "[74] Sijie Zhu, Taojiannan Yang, and Chen Chen. Vigor: Cross-view image geo-localization beyond one-to-one retrieval. In CVPR, pages 3640-3649, 2021. 3" + ], + "bbox": [ + 78, + 90, + 470, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[75] Sijie Zhu, Mubarak Shah, and Chen Chen. Transgeo: Transformer is all you need for cross-view image geo-localization. In CVPR, pages 1162-1171, 2022. 3", + "[76] Sijie Zhu, Linjie Yang, Chen Chen, Mubarak Shah, Xiaohui Shen, and Heng Wang. R2former: Unified retrieval and reranking transformer for place recognition. In CVPR, pages 19370-19380, 2023. 3" + ], + "bbox": [ + 501, + 90, + 890, + 189 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "12764", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2024/EarthLoc_ Astronaut Photography Localization by Indexing Earth from Space/09848c3c-ff6b-4e53-9236-ab0f624d8495_model.json b/2024/EarthLoc_ Astronaut Photography Localization by Indexing Earth from Space/09848c3c-ff6b-4e53-9236-ab0f624d8495_model.json new file mode 100644 index 0000000000000000000000000000000000000000..b16727db3a2ee99cc13751f44888cd883c6dda68 --- /dev/null +++ b/2024/EarthLoc_ Astronaut Photography Localization by Indexing Earth from Space/09848c3c-ff6b-4e53-9236-ab0f624d8495_model.json @@ -0,0 +1,2664 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.089, + 0.131, + 0.882, + 0.154 + ], + "angle": 0, + "content": "EarthLoc: Astronaut Photography Localization by Indexing Earth from Space" + }, + { + "type": "text", + "bbox": [ + 0.197, + 0.18, + 0.774, + 0.234 + ], + "angle": 0, + "content": "Gabriele Berton1 Alex Stoken2 Barbara Caputo1 Carlo Masone1 \n1Politecnico di Torino 2Jacobs Technology, NASA Johnson Space Center \ngabriele.berton@polito.it" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.269, + 0.314, + 0.285 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.301, + 0.474, + 0.59 + ], + "angle": 0, + "content": "Astronaut photography, spanning six decades of human spaceflight, presents a unique Earth observations dataset with immense value for both scientific research and disaster response. Despite their significance, accurately localizing the geographical extent of these images, which is crucial for effective utilization, poses substantial challenges. Current, manual localization efforts are time-consuming, motivating the need for automated solutions. We propose a novel approach - leveraging image retrieval - to address this challenge efficiently. We introduce innovative training techniques which contribute to the development of a high-performance model, EarthLoc. We develop six evaluation datasets and perform a comprehensive benchmark comparing EarthLoc to existing methods, showcasing its superior efficiency and accuracy. Our approach marks a significant advancement in automating the localization of astronaut photography, which will help bridge a critical gap in Earth observations data. Code and datasets are available at https://github.com/gmberton/EarthLoc." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.619, + 0.21, + 0.634 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.644, + 0.473, + 0.903 + ], + "angle": 0, + "content": "Astronaut photography of Earth is a unique remote sensing dataset that spans 60 years of human spaceflight, offering a rare perspective on our planet to the public and valuable data to Earth and atmospheric science researchers. This dataset contains over 4.5 million images and is growing by the tens of thousands per month, as astronauts are continually tasked with taking new photographs that enable scientific research as well as assist in natural disaster response efforts in the wake of events like floods and wildfires. To effectively use these images, the geographical area depicted in them needs to be identified. Unfortunately, this task - Astronaut Photography Localization (APL) - is very challenging. For each photo, only a coarse estimate of location is known, given by the point on Earth directly under the International Space Station (ISS) at the time the photo is taken. This point - called the nadir - can be easily computed using the image's timestamp and the ISS's orbit path. However, two" + }, + { + "type": "image", + "bbox": [ + 0.548, + 0.268, + 0.848, + 0.54 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.553, + 0.895, + 0.734 + ], + "angle": 0, + "content": "Figure 1. Overview of the astronaut photography localization task. Astronauts take hundreds of photos a day from the International Space Station (ISS) cupola (top-left) with hand-held cameras. For each image (example bottom right), the geographic location depicted is not known, and needs to be searched for across a huge area centered at the ISS's (known) nadir point at the moment that the photo is taken. A simulated view of the astronaut's perspective when the ISS is above Europe is shown. The goal of our paper is to automate the task of localizing these images, which could be anywhere within the view. In the figure's example, the photo the astronaut took is indicated by the green line and shown in inset - other possible photo extents are in red, illustrating the wide array of potential locations to search." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.895, + 0.903 + ], + "angle": 0, + "content": "images taken with the same nadir can be thousands of kilometers apart, as even a slight inclination of the astronaut's hand-held camera can move the image's location hundreds of kilometers in any direction, as depicted in Fig. 1. Localization must thus be performed over a wide area, and is additionally complicated by (i) astronauts using hand-held cameras and a variety of zoom lenses, (ii) the large, 2500 kilometer (km) visibility range in all directions, (iii) most photographs being oblique, and (iv) the Earth's appearance" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "12754" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.473, + 0.424 + ], + "angle": 0, + "content": "changing over time, mostly due to weather or seasonal variations. Despite these challenges, the high value of properly localized imagery for Earth science research and disaster response has lead to more than 300,000 [21] astronaut photos being manually localized by identifying the geographic coordinates of the photo's center pixel. This process can take minutes to hours per photo for experts in NASA's Earth Science and Remote Sensing Unit (ESRS) and citizen scientists through ESRS's Image Detective program. The resulting geo-located images have lead to a multitude of peer reviewed papers, technical reports, and articles about light pollution and urban planning [15, 22, 47, 53], atmospheric phenomena [28, 69], changes in land usage and glacial extent [40], and other Earth science topics [27, 48, 68]. Most importantly, astronaut photography has a fast response time, crucial for disaster management - astronauts are alerted of a disaster, provide photos to a ground crew which localizes them, and sends them to first responders [56]. For example, in 2013 this protocol was activated in response to Cyclone Haiyan in the Philippines, wildfires in Australia, flooding in China, Russia, Pakistan, and the USA, and multiple other events. [56]." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.424, + 0.473, + 0.635 + ], + "angle": 0, + "content": "Even with the abundance and importance of satellite imagery in modern applications, astronaut photography fills an unserved gap among other remotely sensed data. Unlike satellites that nominally take top-down imagery in similar illumination conditions at fixed temporal intervals, astronaut photographs can show topography via oblique views from multiple orientations, are taken in various lighting conditions (including nighttime), and vary in focal length to show detail at different resolutions. These qualities produce a complementary data product that would be difficult or impossible to gather from traditional satellites. Additionally, having a human in the loop for data collection allows for real-time response to natural disasters as well as a natural sense to avoid clouds and other obstructions." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.635, + 0.473, + 0.833 + ], + "angle": 0, + "content": "Given the uniqueness and importance of astronaut photography, and the large amount of time spent by human experts on geo-localizing them, researchers have been exploring solutions to automate the task through computer vision methods. Previous works have shown promising results using a pairwise matching setup that iteratively compares an astronaut photo to satellite imagery [58]. Yet, such methods suffer from high latency, searching multiple directions sequentially and using compute-heavy, dense correspondences to determine matching. The time required to find the location of all 4.5 million images using these methods is estimated to be over 20 compute years [58]. Furthermore, low latency is of paramount importance for pho" + }, + { + "type": "text", + "bbox": [ + 0.503, + 0.09, + 0.896, + 0.135 + ], + "angle": 0, + "content": "tography in areas that are affected by natural disasters (e.g., hurricanes or wildfires), for which speedy localization can help real-time, on-the-ground operations." + }, + { + "type": "text", + "bbox": [ + 0.504, + 0.137, + 0.895, + 0.304 + ], + "angle": 0, + "content": "To overcome the high latency of current methods, we propose to instead localize astronaut photography through image retrieval, by matching each astronaut photograph to a worldwide database containing satellite images of known position. In reformulating this problem as a retrieval task, we encounter a range of new challenges, from efficiently training a robust model to achieving low-latency inference, as well as defining a success metric and creating evaluation sets, both of which are crucial pieces when assessing how well a given model will perform when deployed to localize astronaut photography." + }, + { + "type": "text", + "bbox": [ + 0.504, + 0.304, + 0.895, + 0.454 + ], + "angle": 0, + "content": "The ultimate goal of this paper is twofold: (i) provide an efficient method to localize the archive of 4.5M queries as well as all new, incoming imagery, so that researchers studying a given geographical area can have access to a large amount of imagery spanning many years; and (ii) prove the viability of world-wide APL through image retrieval (which has never been attempted and was believed to be infeasible4) in order to spark a new line of research, with direct benefit to all space-based photography and its users across the globe." + }, + { + "type": "text", + "bbox": [ + 0.504, + 0.455, + 0.892, + 0.483 + ], + "angle": 0, + "content": "To achieve its goal, this paper makes the following contributions:" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.485, + 0.875, + 0.5 + ], + "angle": 0, + "content": "- we propose approaching APL through image retrieval;" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.5, + 0.891, + 0.529 + ], + "angle": 0, + "content": "- we develop novel training techniques which show large quantitative improvement on the task;" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.53, + 0.892, + 0.56 + ], + "angle": 0, + "content": "- we provide six evaluation sets and a large benchmark of results with methods from a variety of relevant domains;" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.561, + 0.892, + 0.619 + ], + "angle": 0, + "content": "- we show that a model trained with these new techniques, EarthLoc, allows us to localize a large number of images, widely outperforming all other methods while being fast and efficient." + }, + { + "type": "list", + "bbox": [ + 0.5, + 0.485, + 0.892, + 0.619 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.621, + 0.892, + 0.682 + ], + "angle": 0, + "content": "Finally, we note how EarthLoc is being used to localize ISS astronaut photographs, which are publicly available and conveniently searchable at https://eol.jsc.nasa.gov/ExplorePhotos/." + }, + { + "type": "title", + "bbox": [ + 0.505, + 0.692, + 0.646, + 0.707 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.721, + 0.894, + 0.843 + ], + "angle": 0, + "content": "Image Retrieval Image retrieval involves searching a database for images similar to a query image. Traditionally, methodologies combined hand-crafted local features like SIFT [37], SURF [8], RootSIFT [3] into a global embedding through means like Bag of Words [13], Fisher Vectors [42] or VLAD [29], to allow for fast retrieval through kNN. With the rise of deep learning, local features have been replaced by CNN-derived features [6], significantly" + }, + { + "type": "page_footnote", + "bbox": [ + 0.5, + 0.851, + 0.894, + 0.901 + ], + "angle": 0, + "content": "4 in [59] (Sec. 2.1) scientists at NASA claim that retrieval for APL is infeasible because we cannot precompute a database of reference features due to their belief that the database must be query-specific. Results show that our pipeline can overcome this issue." + }, + { + "type": "page_footnote", + "bbox": [ + 0.075, + 0.836, + 0.467, + 0.861 + ], + "angle": 0, + "content": "1https://eol.jsc.nasa.gov/BeyondThePhotography/ ImageDetective/" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.863, + 0.458, + 0.876 + ], + "angle": 0, + "content": "2https://eol.jsc.nasa.gov/AboutCEO/PubList.htm" + }, + { + "type": "page_footnote", + "bbox": [ + 0.081, + 0.877, + 0.468, + 0.899 + ], + "angle": 0, + "content": "3https : / / storymaps . arcgis . com / stories / 947eb734e811465cb0425947b16b62b3" + }, + { + "type": "list", + "bbox": [ + 0.081, + 0.863, + 0.468, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12755" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.334 + ], + "angle": 0, + "content": "improving retrieval performance. Since then, most of the attention of the retrieval community has been focused on how to best aggregate feature maps and efficiently train robust neural networks. For feature map aggregation, some of the most notable examples in literature include Max pooling [44], Regional Max pooling (R-MAC) [61] and Generalized Mean (GeM) [43]. To train these retrieval models, a number of losses have been proposed, which can be generally grouped into two categories: stateful losses, which have weights whose size depends on the number of classes, like the Large-Margin Softmax [35], SphereFace [36], CosFace [65] and ArcFace [16]; and their stateless counterparts, like the Contrastive and Triplet losses, and the more recent Lifted Structure Loss [54], NTXentLoss [64], Multi-Similarity Loss[67], FastAP [11], Supervised Contrastive [31] and the Circle Loss [60]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.356, + 0.47, + 0.568 + ], + "angle": 0, + "content": "Geo-localization and retrieval with Aerial and Remote Sensing Imagery To the best of our knowledge, only two previous works have specifically addressed the localization problem for astronaut photography. Both rely on local features and compute pairwise comparisons between astronaut photography and geo-located reference imagery. Find My Astronaut Photo [58] focuses on daytime imagery, presenting an evaluation of multiple methods as well as the note that pre-trained self-supervised models were not suitable for a retrieval-based approach to this task. Schwind and Storch [49] instead address nighttime imagery, using synthetically generated street light maps as reference. Both works discuss the importance of proper orientation for matching and the high cost of pairwise comparisons." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.568, + 0.473, + 0.749 + ], + "angle": 0, + "content": "More commonly, studies of aerial imagery focus on data from unmanned aerial vehicles (UAVs) and satellite platforms. Aerial geo-localization is an important aspect of UAV navigation systems. The University-1652 dataset [71] and associated challenge encourage work in cross-view geo-localization for drone-satellite imagery, with strong approaches modifying different parts of the general pipeline, from backbone [73] to training setup [18], to feature partitioning [66]. Most use contrastive and self-supervised losses in training. Other works address cross-view geo-localization in more extreme view point differences, like that between aerial and street views [17, 20, 50-52, 74, 75]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.473, + 0.904 + ], + "angle": 0, + "content": "Remotely sensed data from satellites, on the other hand, often comes with reliable geo-location information from the sensor itself. Extracting features from satellite imagery is still a common task, with such features most often trained on a pretext objective before use on downstream tasks [5, 12, 39, 45]. Occasionally, features are used for retrieval itself [55], to identify areas of the Earth that are similar in appearance and thus might have similar properties to be studied together. In this paper we take a separate direction from previous literature, and aim at using image" + }, + { + "type": "image", + "bbox": [ + 0.534, + 0.09, + 0.641, + 0.146 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.643, + 0.09, + 0.75, + 0.146 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.752, + 0.09, + 0.86, + 0.146 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.534, + 0.147, + 0.641, + 0.202 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.643, + 0.147, + 0.75, + 0.202 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.753, + 0.147, + 0.86, + 0.202 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.215, + 0.892, + 0.243 + ], + "angle": 0, + "content": "Figure 2. Astronaut photo query examples, showcasing the large variability in covered area and appearance." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.247, + 0.695, + 0.262 + ], + "angle": 0, + "content": "retrieval techniques for APL." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.28, + 0.895, + 0.553 + ], + "angle": 0, + "content": "Other localization tasks Various other localization challenges are approached through image retrieval, including visual place recognition (VPR) and visual localization. The former aims at coarsely localizing a given query by matching it to a database of geo-tagged photos, and is mostly studied within urban environments using street-view imagery [7, 63]. Performance is improved either by the use of smart aggregation (e.g. NetVLAD [4]), attention layers (e.g. CRN [32]), or through large-scale training [1, 2, 9, 34]. More recently, the task of universal visual place recognition has been proposed by AnyLoc [30], whose authors provide a model that performs competitively on a large range of scenarios, including aerial imagery. The task of visual localization is focused on finding the precise camera pose of an image, and is commonly approached through feature matching techniques [7, 38, 46, 59]. It can also be used for visual place recognition [24, 76], although this leads to a noticeable increase in runtime when compared to pure retrieval." + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.565, + 0.589, + 0.58 + ], + "angle": 0, + "content": "3. Dataset" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.59, + 0.892, + 0.727 + ], + "angle": 0, + "content": "In this section we discuss the datasets that we use in this work. Specifically, in Sec. 3.1 we introduce the images that we need to localize, called queries; in Sec. 3.2 we outline the collection of the database, made of satellite images (i.e., taken automatically, not with a hand-held camera) of known location; and in Sec. 3.3 we describe the creation of the evaluation datasets that we use to understand the capabilities of our models, combining queries and database in a way that will reflect the real-world use case of our method." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.735, + 0.597, + 0.751 + ], + "angle": 0, + "content": "3.1. Queries" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.758, + 0.892, + 0.88 + ], + "angle": 0, + "content": "Queries are taken from the Gateway to Astronaut Photography of Earth \\(^{5}\\), a collection of over 4.5 million photos of Earth taken by astronauts on the International Space Station (Fig. 2). This unique photography setting yields only coarse location information for each photograph, namely the position on Earth below the ISS (the nadir point) at the time the photo was taken. We seek to find the geographic area on the Earth that each photo encompasses (the photo's location)," + }, + { + "type": "page_footnote", + "bbox": [ + 0.516, + 0.887, + 0.724, + 0.901 + ], + "angle": 0, + "content": "5https://eol.jsc.nasa.gov/" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "12756" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.111, + 0.089, + 0.191, + 0.153 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.194, + 0.089, + 0.273, + 0.152 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.275, + 0.089, + 0.357, + 0.152 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.358, + 0.089, + 0.438, + 0.152 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.111, + 0.157, + 0.19, + 0.221 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.194, + 0.157, + 0.273, + 0.221 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.275, + 0.157, + 0.357, + 0.221 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.358, + 0.157, + 0.438, + 0.221 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.233, + 0.471, + 0.289 + ], + "angle": 0, + "content": "Figure 3. Examples of database images. Top: four images from different regions, with \\(25\\%\\) or \\(50\\%\\) overlap between any pair. The red dot in each image represents the same geographic point. Bottom: 4 images from the same region across different years." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.304, + 0.47, + 0.351 + ], + "angle": 0, + "content": "which can be thousands of kilometers away from the ISS nadir point. This distance of maximum visibility (horizon) can be quickly computed as" + }, + { + "type": "equation", + "bbox": [ + 0.158, + 0.361, + 0.469, + 0.38 + ], + "angle": 0, + "content": "\\[\nd _ {\\text {v i s i b l e}} = \\sqrt {2 R h + h ^ {2}} = 2 4 3 6 \\mathrm {k m} \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.39, + 0.47, + 0.437 + ], + "angle": 0, + "content": "where \\( R = 6317 \\) is Earth's radius and \\( h = 450 \\) is the ISS maximum orbiting altitude. We follow common practice and round this number to 2500 kilometers [58]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.454, + 0.47, + 0.62 + ], + "angle": 0, + "content": "Challenges to APL Some astronaut photo queries, taken with high focal length lenses, can cover only a few kilometers within the almost 20 million sq. km area visible to the astronaut taking the photo, which makes localizing the images extremely challenging. Additional hurdles presented by the imagery acquisition process include the varying quality of the images themselves due to motion blur from the fast moving space station, and occlusion of unique, location-identifying features by cloud cover, shadow, or other portions of the ISS (Fig. 2). More queries are shown in Sec. 10 of the Supplementary." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.62, + 0.47, + 0.711 + ], + "angle": 0, + "content": "From the 4.5 million photos available, we consider only the day-time images that have a ground truth obtained by FMAP [58], and keep only those which cover areas between 5000 and 900,000 sq. km, in order to match the areas of database images (see Sec. 3.2). This leads to 17,764 queries which we can use for validation and testing." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.72, + 0.185, + 0.734 + ], + "angle": 0, + "content": "3.2. Database" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.742, + 0.469, + 0.848 + ], + "angle": 0, + "content": "To estimate the location of the queries through image retrieval, we need a worldwide database of images, where each image is labeled with its position. We build such a database from an open-source, composited, cloudless collection of Sentinel-2 satellite imagery6. This collection contains imagery of worldwide landmasses at resolutions up to 15 meters per pixel. Some examples are shown in Fig. 3." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.849, + 0.469, + 0.878 + ], + "angle": 0, + "content": "For our database we include any land area between latitudes \\(60^{\\circ}\\) and \\(-60^{\\circ}\\) (i.e., the area traversed by the ISS). The" + }, + { + "type": "image", + "bbox": [ + 0.506, + 0.092, + 0.888, + 0.216 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.227, + 0.894, + 0.311 + ], + "angle": 0, + "content": "Figure 4. To create evaluation sets we choose all images that could contain a Point of Interest (POI) - all photos with nadir point within \\( d_{visible} \\). To localize all photos within this range, even if they do not contain the POI, we create a database that contains all areas visible from the nadir points of all selected photos, yielding a database area of about \\( 5000\\mathrm{km}^2 \\) per POI." + }, + { + "type": "table", + "bbox": [ + 0.503, + 0.325, + 0.89, + 0.412 + ], + "angle": 0, + "content": "
Name(Lat, Lon)# queries# DBInterest
Texas(30, -95)614234kOften photoed (many queries)
Alps(45, 10)239453kSci. interest - glacial change
California(38, -122)356830kDisaster response - wildfires
Gobi Desert(40, 105)72654kSci. interest - desertification
Amazon(-3, -60)68219kSci. interest - deforestation
Toshka Lakes(23, 30)216463kSci. interest - flood monitoring
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.416, + 0.892, + 0.458 + ], + "angle": 0, + "content": "Table 1. Evaluation sets. The six sets cover diverse geographies and relevant areas for research on topics like climate change and disaster management. DB stands for database." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.468, + 0.892, + 0.528 + ], + "angle": 0, + "content": "Sentinel-2 imagery is available as map tiles, and we use tiles at zoom levels 9, 10, 11, with resolutions ranging from \\(\\sim 300\\) to \\(\\sim 75\\) meters per pixel7, to ensure that the database encompasses the large scale range of the query images." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.529, + 0.893, + 0.68 + ], + "angle": 0, + "content": "These parameters lead us to define roughly 175k regions, where each region \\(\\mathcal{R}\\) is the projection of a square with known corner coordinates onto the Earth's surface. Regions have areas ranging from 5000 sq. km to 900,000 sq. km. We take partially (up to \\(50\\%\\)) overlapping regions so that each query has at least one image in the database with considerable overlap (Fig. 3 top). From each region we collect four images, one per year from 2018, 2019, 2020 and 2021, covering temporal changes in a given area (Fig. 3, bottom). This produces a database of 700k images." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.68, + 0.892, + 0.801 + ], + "angle": 0, + "content": "Unlike our queries, these database images have undergone post-processing that includes atmospheric corrections as well as cloud pixel minimization via composition of images from multiple days. Additionally, the database images are always taken nadir facing (i.e. perpendicular to Earth's surface) to minimize obliquity effects. Each database image has resolution \\(1024 \\times 1024\\). Further images from the database are shown in Sec. 10 of the Supplementary." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.811, + 0.654, + 0.825 + ], + "angle": 0, + "content": "3.3. Evaluation sets" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.835, + 0.892, + 0.866 + ], + "angle": 0, + "content": "With the end goal of fast and accurate localization of the astronaut photography archive, we propose a deployment" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.887, + 0.239, + 0.9 + ], + "angle": 0, + "content": "\\(^{6}\\)https://s2maps.eu" + }, + { + "type": "page_footnote", + "bbox": [ + 0.501, + 0.875, + 0.892, + 0.899 + ], + "angle": 0, + "content": "7https://wiki.openstreetmap.org/wiki/Zoom levels" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "12757" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.141, + 0.088, + 0.832, + 0.278 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.288, + 0.895, + 0.332 + ], + "angle": 0, + "content": "Figure 5. Training strategy. Naive batching produces too easy a task for the model to learn robust representations from. We increase difficulty by clustering batches by similarity and adding year-wise augmentations, which move images of different regions, but from the same year, closer together in feature space. The model must then learn representations that discern similarity in this more difficult context." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.342, + 0.471, + 0.463 + ], + "angle": 0, + "content": "environment that takes advantage of the known nadir points and visibility limits associated with each image. So while localization can be performed by matching any given query against the whole world-wide database, a more suitable approach is to follow a divide and conquer paradigm by grouping queries according to their nadir, which allows us to use only a subset of the database for each search, increasing speed and improving results." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.464, + 0.47, + 0.569 + ], + "angle": 0, + "content": "We seek to create evaluation sets that mirror the situation seen in deployment, so we evaluate on sets of queries that have a nadir within a \\(2500\\mathrm{km}\\) radius from a chosen Point of Interest (POI). We then build a database of images such that the database fully encompasses the area that could have been photographed from each collected query, which is an area \\(2*d_{visible}\\) from the original POI, as shown in Fig. 4." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.569, + 0.47, + 0.629 + ], + "angle": 0, + "content": "Since we need the queries' ground truth position for evaluation, we choose queries that have already been localized with exhaustive image matching techniques [58]. For these queries, the full photo extent is publicly available8." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.63, + 0.47, + 0.689 + ], + "angle": 0, + "content": "To properly evaluate different methods, we choose six POIs that represent various geographies but all have scientific relevance, and center an evaluation set around each. These are summarized in Tab. 1." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.704, + 0.169, + 0.719 + ], + "angle": 0, + "content": "4. Method" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.73, + 0.47, + 0.866 + ], + "angle": 0, + "content": "The goal of this work is to estimate the location of each astronaut photo query, which we tackle through image retrieval: for any given query, the objective is to find the most similar image(s) from the database, and use the corresponding location labels to confidently localize the query. With this in mind, we aim to train a retrieval model on satellite imagery such that the model is robust to changes in scale, perspective, and color, as well as temporal/seasonal variation. Specifically, the model's task is to extract features" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.342, + 0.892, + 0.372 + ], + "angle": 0, + "content": "from each image. Then, for each query, the most similar database image is retrieved via kNN in feature space." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.373, + 0.892, + 0.403 + ], + "angle": 0, + "content": "In the next subsections we describe how we train our retrieval model using the proposed database." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.41, + 0.673, + 0.426 + ], + "angle": 0, + "content": "4.1. Baseline Training" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.433, + 0.892, + 0.583 + ], + "angle": 0, + "content": "We want our model to extract relevant features of satellite imagery, so a straightforward approach would be to train via contrastive learning: for a given batch of images (samples), generate a number of augmented views (typically two), and apply a contrastive loss to maximize the distance in feature space between different samples and minimize distance between different views of the same sample. However, this solution would not allow the model to learn robustness to temporal changes, because the augmented views come from the same sample, with no inherent temporal variation." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.584, + 0.892, + 0.75 + ], + "angle": 0, + "content": "Therefore, instead of producing multiple views of the same image via synthetic augmentation, we train the model by randomly selecting a number of regions, and for each region we consider its four images (a quadruplet), one per year. Within a batch of quadruplets, the four images from each region are mutual positives, whereas any two images from different regions are considered negatives: we can therefore consider each region as a class. This provides a natural way to address temporal variation. We can then train the model through a multi-similarity loss [67]. This idea is shown in Fig. 5 (left)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.757, + 0.677, + 0.772 + ], + "angle": 0, + "content": "4.2. Clustered Batches" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.78, + 0.892, + 0.901 + ], + "angle": 0, + "content": "The baseline paradigm simply creates batches of random regions – a single batch of quadruplets could contain samples depicting shorelines, islands, deserts, and mountains. The loss operates batchwise, so large intra-batch variability can lead the model to learn too swiftly to distinguish between such diverse regions. In these situations, training converges quickly and the model stops short of learning to discriminate between similar-looking, but distinct, regions." + }, + { + "type": "page_footnote", + "bbox": [ + 0.078, + 0.875, + 0.468, + 0.901 + ], + "angle": 0, + "content": "https://eol.jsc.nasa.gov/SearchPhotos/PhotosDatabaseAPI/(mlcoordTable)" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12758" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.182 + ], + "angle": 0, + "content": "To avoid this issue, we want to train with batches of highly similar images, which are still from different regions. To this end, we compute clusters of images with similar representations. Every \\(N_{iter}\\) training iterations, we extract the features from each region with the trained model, using only images from a single year, and group them into \\(C\\) clusters." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.183, + 0.471, + 0.364 + ], + "angle": 0, + "content": "Then, we create training batches using samples from a given cluster, choosing a different cluster for each batch. This increases the difficulty of the task in each batch, ensuring that the model learns to discriminate between similar-looking images of different regions. Note that clusters are not geographically related, but related in feature space. For example, a cluster may contain images from the Gobi and the Sahara deserts, which both appear similar in feature space despite being geographically distant. This procedure, which can be seen as an offline mining approach, increases training time by less than \\(5\\%\\) but improves recall by roughly \\(15\\%\\) on average (see Sec. 5.3), a worthwhile trade off." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.365, + 0.469, + 0.411 + ], + "angle": 0, + "content": "This concept is illustrated in Fig. 5 (middle), and qualitative examples of images that are clustered together are shown in Sec. 11.2 of the Supplementary." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.422, + 0.347, + 0.437 + ], + "angle": 0, + "content": "4.3. Year-Wise Data Augmentation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.446, + 0.469, + 0.55 + ], + "angle": 0, + "content": "Query images span the ongoing \\(20+\\) years of ISS operations, as well as multiple camera types, camera operators, and view points, so queries depicting the same location can have a number of natural \"transformations\". These range from blueish hues from lens filters, to skewed representations in highly oblique shots, to seasonal variation in ground cover (e.g. snow, autumn colors)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.552, + 0.469, + 0.643 + ], + "angle": 0, + "content": "To make the model robust to such transformations, we employ data augmentation. Typically, augmentations are applied per image or per batch, but we argue that to truly learn robustness to the augmentations in this setting, the same augmentation should be performed on images from different regions, but from the same year." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.644, + 0.469, + 0.856 + ], + "angle": 0, + "content": "Formally, at each training iteration we choose a set of 4 random augmentations, one for each year \\( A_{2018}, A_{2019}, A_{2020}, A_{2021} \\). We then apply the corresponding augmentation to all images from that year (i.e. \\( A_{2018} \\) is applied to all images from 2018, across all regions). Such \"year-wise\" augmentation has the effect of moving images from a given year closer in feature space to each other, as shown in Fig. 5 (right). In the loss, images from the same region, but different years, are treated as positives, while images from different regions are treated as negatives. This has the effect that images that receive the same augmentation are assigned different classes – forcing the model to learn to ignore such augmentations – and instead pull together images from the same region, across different years." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Furthermore, given that the same augmentation is applied to multiple images, augmentation can be batched and performed on the GPU, making training slightly faster (by" + }, + { + "type": "image", + "bbox": [ + 0.529, + 0.094, + 0.868, + 0.272 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.281, + 0.892, + 0.378 + ], + "angle": 0, + "content": "Figure 6. Dataset collection from satellite imagery and handling in Neutral-Aware Multi-Similarity Loss. For a given image (yellow), images with no overlap are negatives (red), images from the same region are positives (green, overlapping yellow), and images with partial overlap are neutral (grey). Note that the database is made of images at 3 different zooms (each double in size than the next), which explains the regions of different size." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.383, + 0.868, + 0.399 + ], + "angle": 0, + "content": "\\(5\\%\\) on average, although this depends on the hardware)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.408, + 0.82, + 0.423 + ], + "angle": 0, + "content": "4.4. Neutral-Aware Multi-Similarity Loss" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.431, + 0.892, + 0.492 + ], + "angle": 0, + "content": "The powerful Multi-Similarity loss is designed to take as input multiple images from a number of classes, along with the corresponding class indices, and compute a loss from the positives and negatives for each image." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.492, + 0.892, + 0.672 + ], + "angle": 0, + "content": "However, it does not take into account that in some cases, pairs of images are neither positives nor negatives, and should be considered \"neutral\". While the concept of neutral images does not exist in standard image retrieval tasks, as an image either belongs to class A or class B, in APL it must be taken into consideration. For example, should two images, which have an intersection over union of \\(25\\%\\), be considered positives or negatives? This is precisely the case when two images of the same area, but from different zoom levels, are batched together, as well as overlapping images at the same level (see Fig. 6). We believe the answer is neither, so we add the \"neutral\" case to the loss." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.672, + 0.892, + 0.718 + ], + "angle": 0, + "content": "Formally, we model the neutral case with the indicator function \\(\\mathcal{I}_{NA}\\), which is 0 when the regions \\(\\mathcal{R}_i\\) and \\(\\mathcal{R}_j\\) intersect but are not equal, and 1 otherwise." + }, + { + "type": "equation", + "bbox": [ + 0.573, + 0.729, + 0.891, + 0.748 + ], + "angle": 0, + "content": "\\[\n\\mathcal {I} _ {N A} \\left(\\mathcal {R} _ {i}, \\mathcal {R} _ {j}\\right) = \\mathbb {1} _ {\\left(\\mathcal {R} _ {i} \\cap \\mathcal {R} _ {j}\\right) \\wedge \\left(\\mathcal {R} _ {i} \\neq \\mathcal {R} _ {j}\\right)} \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.757, + 0.882, + 0.772 + ], + "angle": 0, + "content": "With this, the neutral-aware multi-similarity loss becomes" + }, + { + "type": "equation", + "bbox": [ + 0.511, + 0.782, + 0.892, + 0.878 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} _ {N A M S} = \\frac {1}{B S} \\sum_ {i = 1} ^ {B S} \\left\\{\\frac {1}{\\alpha} \\log \\left[ 1 + \\sum_ {k \\in \\mathcal {R} _ {i}} e ^ {- \\alpha \\left(\\mathcal {S} _ {i k} - \\lambda\\right)} \\right] \\right. \\\\ \\left. + \\frac {1}{\\beta} \\log \\left[ 1 + \\sum_ {k \\notin \\mathcal {R} _ {i}} \\mathcal {I} _ {N A} \\left(\\mathcal {R} _ {i}, \\mathcal {R} _ {k}\\right) e ^ {- \\beta \\left(\\mathcal {S} _ {i k} - \\lambda\\right)} \\right] \\right\\} \\end{array} \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.886, + 0.892, + 0.901 + ], + "angle": 0, + "content": "where \\(BS\\) is the number of images in a batch, \\(\\mathcal{R}_i\\) is the" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12759" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.08, + 0.089, + 0.894, + 0.242 + ], + "angle": 0, + "content": "
MethodType of Training ImageryTexas (validation)AlpsCaliforniaGobi DesertAmazonToshka Lakes
R@1R@10R@100R@1R@10R@100R@1R@10R@100R@1R@10R@100R@1R@10R@100R@1R@10R@100
Nadir-2.4--1.2--2.4--1.8--3.1--1.4--
Random Choice-0.21.715.50.11.111.60.22.320.10.11.013.20.11.111.50.21.29.1
NetVLAD [4]Urban VPR3.613.434.85.215.236.24.314.439.02.17.724.56.619.946.39.622.545.6
SFRS [23]Urban VPR6.817.440.97.017.841.45.417.244.45.214.737.69.423.852.510.820.442.4
Conv-AP [1]Urban VPR3.713.537.72.08.825.03.815.441.23.79.024.27.021.147.75.012.332.6
CosPlace [9]Urban VPR4.915.242.05.919.345.24.815.243.37.917.539.58.523.954.49.825.453.0
MixVPR [2]Urban VPR4.515.340.13.18.622.64.417.543.24.010.124.59.828.758.75.915.436.5
EigenPlaces [10]Urban VPR6.421.652.48.721.450.36.022.854.38.120.040.910.926.557.914.330.960.4
AnyLoc [30] (DINOv2 [41] + NetVLAD)Universal VPR44.168.787.840.770.892.048.775.091.628.757.081.738.663.886.263.784.596.3
TorchGeo [57] (ResNet50 w MOCO [26])Satellite1.03.211.60.31.35.71.14.316.40.42.88.31.96.619.40.72.99.9
TorchGeo [57] (ResNet50 w SeCo [70])Satellite6.115.641.77.420.249.25.114.537.13.714.038.94.613.332.95.715.638.5
TorchGeo [57] (ResNet50 w GASSL [5])Satellite9.722.846.49.123.150.513.331.458.86.317.545.48.320.340.120.438.664.2
OGCL UAV-View [18] (ConvNeXt-XXLarge)UAV16.235.365.814.434.164.719.642.171.05.520.545.310.026.354.721.138.463.4
OGCL UAV-View [18] (ViT-L/14) [19]UAV17.633.255.914.633.263.922.848.174.97.622.850.120.439.162.531.851.874.4
MBEG [73] (ViT-L/14)UAV7.017.635.16.619.345.98.720.741.54.415.038.06.417.139.38.120.749.1
EarthLoc (Ours)Satellite55.973.088.358.476.889.558.076.091.451.167.586.547.267.984.672.285.093.3
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.248, + 0.895, + 0.318 + ], + "angle": 0, + "content": "Table 2. Results from different methods on our query sets. Methods are grouped into (1) naive, (2) VPR, (3) Remote Sensing (UAV+Satellite), (4) other baselines trained on our database with clustered batches (Sec. 4.2). For all methods we use \\(4x90TTA\\). VPR stands for Visual Place Recognition, UAV for Unmanned Aerial Vehicle. Best results overall in bold, best per-group underlined. EarthLoc outperforms other models, and performs competitively with AnyLoc, while having 50 times faster features extraction and 10 times smaller features." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.327, + 0.47, + 0.357 + ], + "angle": 0, + "content": "region of image \\(i\\), \\(\\mathcal{S}_{ik}\\) is the similarity between image \\(i\\) and \\(j\\), \\(\\lambda\\) is a margin, and \\(\\alpha, \\beta\\) are hyperparameters." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.358, + 0.47, + 0.418 + ], + "angle": 0, + "content": "We believe that the neutral-aware multi-similarity loss is applicable to any task where it is not straightforward how to enforce full positivity or negativity between any pair of images, including localization tasks like VPR." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.433, + 0.21, + 0.451 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.458, + 0.275, + 0.475 + ], + "angle": 0, + "content": "5.1. Experimental setting" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.482, + 0.194, + 0.499 + ], + "angle": 0, + "content": "5.1.1 Training" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.507, + 0.471, + 0.673 + ], + "angle": 0, + "content": "In training, we use a batch size of 96 quadruplets (384 images), where one quadruplet corresponds to four images within a region. Our model has a MixVPR-style architecture [2], with a ResNet50 [25] backbone and an MLP-Mixer[62] with output dimension 4096. We train for 10k iterations with the Adam optimizer [33] and learning rate 0.0001. Clustering (see Sec. 4.2) is performed every 2k iterations with the number of clusters \\( C = 50 \\). As augmentation, we use torchvision's color jittering, random perspective, and random rotation. Training takes between 5 and 7 hours on an A100 GPU for all methods described in Sec. 4." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.695, + 0.199, + 0.709 + ], + "angle": 0, + "content": "5.1.2 Inference" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.719, + 0.469, + 0.765 + ], + "angle": 0, + "content": "Due to the imagery acquisition conditions, we cannot assume any canonical orientation for the photos (in space, there's no gravity vector, i.e., no \"up\" or \"down\")." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.766, + 0.47, + 0.902 + ], + "angle": 0, + "content": "This poses a challenge for any retrieval method – to address it, we generate four feature embeddings for each database image, one each after rotating the image for angles \\( r \\in [0,90,180,270] \\), in a process we call \\( 4 \\times 90^{\\circ} \\) test time augmentation (4x90TTA). For fairness, we do this for all methods in all experiments, given that the models are not trained to be robust to such large arbitrary rotations. This increases latency by \\( 4\\mathrm{x} \\), but the method is still orders of magnitude faster than previous approaches (retrieval takes 0.05" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.327, + 0.895, + 0.493 + ], + "angle": 0, + "content": "seconds per query with \\(4x90TTA\\), while previous works required \\(\\sim 1\\) minute [58] per query). Memory requirements also increase by four times, although this is far from being an issue for a standard workstation, since all database features for the whole Earth fit in less than 12GB of RAM. Experiments without \\(4x90TTA\\) are shown in the Supplementary in Sec. 7. As an added bonus, since each feature from the database refers to an image with a certain orientation (w.r.t. North), retrieval predictions also give an estimate of the orientation of the query. We use features from the 2021 satellite imagery as our global database." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.516, + 0.894, + 0.609 + ], + "angle": 0, + "content": "Success Criteria With the final goal of localizing the 4.5M astronaut photograph queries in mind, we define a predicted image as correct if there is non-zero overlap between the prediction and the query. As a metric we use Recall@N, defined as the percentage of queries for which at least one of the top-N predictions is correct." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.619, + 0.593, + 0.634 + ], + "angle": 0, + "content": "5.2. Results" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.644, + 0.894, + 0.764 + ], + "angle": 0, + "content": "In Tab. 2 we present results from a large number of models from multiple domains. We first show the recall when naively predicting the query to be directly nadir, proving that the vast majority of images are taken at an angle, and the poor performance of randomly choosing a prediction from the database - a symptom of the large-scale nature of the task (i.e., localizing a 5000 sq.km wide image within a 20M sq.km area)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.894, + 0.901 + ], + "angle": 0, + "content": "Results with methods trained for Urban Visual Place Recognition (VPR) provide, somewhat unsurprisingly, poor results - on the other hand, AnyLoc [30] shows remarkably strong results, correctly predicting the location of the query as the first prediction for roughly \\(50\\%\\) of the queries across all evaluation sets, although at the cost of over 50x slower feature extraction and 10x bigger features than its VPR counterparts (AnyLoc uses 49152-D features vs 4096 of NetVLAD)." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "12760" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.082, + 0.089, + 0.891, + 0.16 + ], + "angle": 0, + "content": "
Multi-similarity LossUsing regional quadrupletsClustered BatchesYear-Wise Data AugmentationNeutral-Aware Multi-SimilarityTexas (validation)AlpsCaliforniaGobi DesertAmazonToshka LakesAverage
R@1R@100R@1R@100R@1R@100R@1R@100R@1R@100R@1R@100R@1R@100
✓25.571.424.877.327.876.616.868.823.661.538.880.826.272.7
✓✓45.084.145.487.349.588.535.779.934.280.262.991.145.485.2
✓✓✓51.487.549.689.553.990.640.884.339.382.864.093.049.888.0
✓✓✓✓54.687.553.987.255.991.646.882.945.682.467.691.954.187.2
✓✓✓55.388.255.190.057.290.344.685.848.484.868.992.954.988.7
✓✓✓✓55.988.358.489.558.091.451.186.547.284.672.293.357.188.9
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.169, + 0.893, + 0.197 + ], + "angle": 0, + "content": "Table 3. Ablations over test sets. EarthLoc uses Clustered Batches (Sec. 4.2), Year-Wise Data Augmentation (Sec. 4.3) and Neutral-Aware Multi-Similarity Loss (Sec. 4.4). Best results in bold." + }, + { + "type": "image", + "bbox": [ + 0.079, + 0.209, + 0.47, + 0.29 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.299, + 0.471, + 0.34 + ], + "angle": 0, + "content": "Figure 7. Qualitative results. Each group of four images represents a query (left) and its top-3 predictions. Green and red indicate if the prediction is correct or incorrect." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.363, + 0.47, + 0.453 + ], + "angle": 0, + "content": "We then test a number of models trained with self-supervised techniques, like MOCO [26], SeCo [70], and Geography-aware self-supervised learning (GASSL) [5], conveniently provided by the TorchGeo library [57]. Surprisingly, these methods did not provide competitive results, and are on average about equal with Urban VPR methods." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.457, + 0.47, + 0.563 + ], + "angle": 0, + "content": "From the Unmanned Aerial Vehicle (UAV) domain, we use the latest state of the art (SOTA) methods, i.e., the winners of the recent challenge in UAV localization at ACM MM (October 2023) [72], namely OGCL [18] and MBEG [73]. Despite being SOTA in the related task of UAV localization, we see that these methods are not competitive with AnyLoc, despite outperforming most other methods." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.567, + 0.47, + 0.613 + ], + "angle": 0, + "content": "Finally, EarthLoc outperforms all previous works, localizing \\(57.1\\%\\) of queries within the top-1 prediction (Recall@1) averaged on the 6 evaluation sets." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.632, + 0.19, + 0.646 + ], + "angle": 0, + "content": "5.3. Ablations" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.658, + 0.47, + 0.749 + ], + "angle": 0, + "content": "In Tab. 3 we ablate the components of our method to better understand performance. Results show that not only do Clustered Batches, Year-Wise Data Augmentation and Neutral-Aware Multi-Similarity Loss provide significant increases over the baseline, but they are also orthogonal improvements and can be easily combined." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.768, + 0.411, + 0.784 + ], + "angle": 0, + "content": "5.4. Qualitative Analysis and Failure Cases" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.795, + 0.47, + 0.903 + ], + "angle": 0, + "content": "A sample of qualitative results from EarthLoc are shown in Fig. 7. The samples show that EarthLoc picks locations that have similar characteristics to the query (top left), is robust to rotations and changes in scale (bottom right), and provides correct predictions that are difficult to correctly match even for the human eye (top left, bottom left). More qualitative results in Sec. 11 of the Supplementary." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.212, + 0.766, + 0.227 + ], + "angle": 0, + "content": "5.5. Limitations and Future Work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.237, + 0.894, + 0.387 + ], + "angle": 0, + "content": "While this work takes a major step forward in the study of APL, EarthLoc focuses on daytime queries, and we did not investigate panoramic or nighttime imagery. The localization of nighttime astronaut photos in particular is important (e.g. to study light pollution [14]), but we believe this to be a separate challenge that should be addressed in future work. Another limitation is due to the zoom limits of the database: we chose database tiles at zoom levels 9, 10, 11 (see Sec. 3.2), which allows us to localize queries of area between 5000 and 900,000 square km." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.389, + 0.894, + 0.48 + ], + "angle": 0, + "content": "The originality of this work opens up a large number of possible future research directions, like training directly on (some of the) localized queries, using seasonal data, post-processing predictions, obtaining pixel-wise geolocation, applying domain adaptation, and using available metadata (like camera lens) during training, to name a few." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.498, + 0.628, + 0.513 + ], + "angle": 0, + "content": "6. Conclusions" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.524, + 0.892, + 0.691 + ], + "angle": 0, + "content": "We introduce EarthLoc, a method for localizing astronaut photography of Earth. By reframing APL from an image matching task to an image retrieval task, and introducing a custom loss function and training scheme, our method efficiently determines the geographic location of astronaut photography given only the position of the International Space Station at the time the photo was taken. Our main contributions are the Neutral-Aware Multi-Similarity Loss, Year-Wise Data Augmentation technique, and new astronaut photography-oriented validation and test datasets to encourage future work on this problem." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.706, + 0.893, + 0.9 + ], + "angle": 0, + "content": "Acknowledgements. We acknowledge the CINECA award under the ISCRA initiative, for the availability of high performance computing resources. This work was supported by CINI. Project supported by ESA Network of Resources Initiative. This study was carried out within the project FAIR - Future Artificial Intelligence Research - and received funding from the European Union Next-GenerationEU (PIANO NAZIONALE DI RIPRESA E RESILLENZA (PNRR) - MISSIONE 4 COMPONENTE 2, INVESTIMENTO 1.3 - D.D. 1555 11/10/2022, PE00000013). This manuscript reflects only the authors' views and opinions, neither the European Union nor the European Commission can be considered responsible for them. European Lighthouse on Secure and Safe AI - ELSA, HORIZON EU Grant ID: 101070617" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "12761" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.469, + 0.156 + ], + "angle": 0, + "content": "[1] Amar Ali-bey, Brahim Chaib-draa, and Philippe Giguère. Gsv-cities: Toward appropriate supervised visual place recognition. Neurocomputing, 513:194-203, 2022. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.157, + 0.469, + 0.199 + ], + "angle": 0, + "content": "[2] Amar Ali-bey, Brahim Chaib-draa, and Philippe Giguère. Mixvpr: Feature mixing for visual place recognition. In WACV, pages 2998-3007, 2023. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.089, + 0.201, + 0.469, + 0.24 + ], + "angle": 0, + "content": "[3] R. Arandjelović and Andrew Zisserman. Three things everyone should know to improve object retrieval. In CVPR, pages 2911-2918, 2012. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.243, + 0.469, + 0.297 + ], + "angle": 0, + "content": "[4] Relja Arandjelović, Petr Gronat, Akihiko Torii, Tomas Pajdla, and Josef Sivic. NetVLAD: CNN architecture for weakly supervised place recognition. IEEE TPAMI, 40(6): 1437-1451, 2018. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.298, + 0.469, + 0.352 + ], + "angle": 0, + "content": "[5] Kumar Ayush, Burak Uzkent, Chenlin Meng, Kumar Tanmay, Marshall Burke, David Lobell, and Stefano Ermon. Geography-aware self-supervised learning. ICCV, 2021. 3, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.354, + 0.469, + 0.395 + ], + "angle": 0, + "content": "[6] Artem Babenko, Anton Slesarev, A. Chigorin, and V. Lempitsky. Neural codes for image retrieval. ECCV, abs/1404.1777, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.397, + 0.469, + 0.464 + ], + "angle": 0, + "content": "[7] Giovanni Barbarani, Mohamad Mostafa, Hajali Bayramov, Gabriele Trivigno, Gabriele Berton, Carlo Masone, and Barbara Caputo. Are local features all you need for cross-domain visual place recognition? In CVPRW, pages 6155-6165, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.467, + 0.469, + 0.508 + ], + "angle": 0, + "content": "[8] Herbert Bay, Andreas Ess, Tinne Tuytelaars, and Luc Van Gool. Speeded-up robust features (surf). Computer Vision and Image Understanding, 110:346-359, 2008. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.509, + 0.469, + 0.55 + ], + "angle": 0, + "content": "[9] Gabriele Berton, Carlo Masone, and Barbara Caputo. Rethinking visual geo-localization for large-scale applications. In CVPR, 2022. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.552, + 0.469, + 0.605 + ], + "angle": 0, + "content": "[10] Gabriele Berton, Gabriele Trivigno, Barbara Caputo, and Carlo Masone. Eigenplaces: Training viewpoint robust models for visual place recognition. In ICCV, pages 11080-11090, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.607, + 0.469, + 0.647 + ], + "angle": 0, + "content": "[11] Fatih Cakir, Kun He, Xide Xia, Brian Kulis, and Stan Sclaroff. Deep metric learning to rank. In CVPR, pages 1861-1870, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.649, + 0.469, + 0.717 + ], + "angle": 0, + "content": "[12] Yezhen Cong, Samar Khanna, Chenlin Meng, Patrick Liu, Erik Rozi, Yutong He, Marshall Burke, David B. Lobell, and Stefano Ermon. SatMAE: Pre-training transformers for temporal and multi-spectral satellite imagery. In NeurIPS, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.719, + 0.469, + 0.761 + ], + "angle": 0, + "content": "[13] Gabriela Csurka, Christopher Dance, Lixin Fan, Jutta Willamowski, and Cedric Bray. Visual categorization with bags of keypoints. In ECCV, 2004. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.762, + 0.469, + 0.83 + ], + "angle": 0, + "content": "[14] Alejandro Sánchez de Miguel, José Gómez Castaño, Jaime Zamorano, Sergio Pascual, M Ángeles, L Cayuela, Guillermo Martín Martínez, Peter Challupner, and Christopher C M Kyba. Atlas of astronaut photos of earth at night. *Astronomy and Geophysics*, 55(4):4.36-4.36, 2014. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.832, + 0.469, + 0.9 + ], + "angle": 0, + "content": "[15] Alejandro Sánchez de Miguel, Jonathan Bennie, Emma Rosenfeld, Simon Dzurjak, and Kevin J. Gaston. Environmental risks from artificial nighttime lighting widespread and increasing across Europe. Science Advances, 8(37): eabl6891, 2022. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.469, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.134 + ], + "angle": 0, + "content": "[16] Jiankang Deng, J. Guo, and S. Zafeiriou. ArcFace: Additive angular margin loss for deep face recognition. In CVPR, pages 4685-4694, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.136, + 0.892, + 0.177 + ], + "angle": 0, + "content": "[17] Fabian Deuser, Konrad Habel, and Norbert Oswald. Sample4geo: Hard negative sampling for cross-view geolocation. In ICCV, pages 16847-16856, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.179, + 0.892, + 0.261 + ], + "angle": 0, + "content": "[18] Fabian Deuser, Konrad Habel, Martin Werner, and Norbert Oswald. Orientation-guided contrastive learning for uav-view geo-localisation. In Proceedings of the 2023 Workshop on UAVs in Multimedia: Capturing the World from a New Perspective, page 7-11, New York, NY, USA, 2023. Association for Computing Machinery. 3, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.263, + 0.892, + 0.345 + ], + "angle": 0, + "content": "[19] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale. In ICLR, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.347, + 0.892, + 0.403 + ], + "angle": 0, + "content": "[20] Florian Fervers, Sebastian Bullinger, Christoph Bodensteiner, Michael Arens, and Rainer Stiefelhagen. Uncertainty-aware vision-based metric cross-view geolocation. In CVPR, pages 21621-21631, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.405, + 0.892, + 0.472 + ], + "angle": 0, + "content": "[21] Kenton Fisher, Sara Schmidt, and Alex Stoken. Crew earth observations: New tools to support your research. In 12th Annual International Space Station Research and Development Conference. Center for the Advancement of Science in Space, Inc., 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.475, + 0.892, + 0.516 + ], + "angle": 0, + "content": "[22] Kevin J. Gaston and Alejandro Sánchez de Miguel. Environmental impacts of artificial light at night. Annual Review of Environment and Resources, 47(1):373-398, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.518, + 0.892, + 0.572 + ], + "angle": 0, + "content": "[23] Yixiao Ge, Haibo Wang, Feng Zhu, Rui Zhao, and Hongsheng Li. Self-supervising fine-grained region similarities for large-scale image localization. In ECCV, pages 369-386, Cham, 2020. Springer International Publishing. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.575, + 0.892, + 0.629 + ], + "angle": 0, + "content": "[24] Stephen Hausler, Sourav Garg, Ming Xu, Michael Milford, and Tobias Fischer. Patch-netvlad: Multi-scale fusion of locally-global descriptors for place recognition. In CVPR, pages 14141-14152, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.631, + 0.892, + 0.658 + ], + "angle": 0, + "content": "[25] K. He, X. Zhang, S. Ren, and J. Sun. Deep residual learning for image recognition. In CVPR, pages 770-778, 2016. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.66, + 0.892, + 0.714 + ], + "angle": 0, + "content": "[26] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In CVPR, pages 9726-9735, 2020. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.717, + 0.892, + 0.771 + ], + "angle": 0, + "content": "[27] Christopher. R. Jackson and Werner Alpers. The role of the critical angle in brightness reversals on sunglint images of the sea surface. Journal of Geophysical Research: Oceans, 115(C9), 2010. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.773, + 0.892, + 0.829 + ], + "angle": 0, + "content": "[28] Augustin Juhl, Thomas Farges, and Elisabeth Blanc. Color pictures of sprites from non-dedicated observation on board the international space station. Journal of Geophysical Research: Space Physics, 118(1):454-461, 2013. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.831, + 0.892, + 0.871 + ], + "angle": 0, + "content": "[29] Hervé Jégou, Matthijs Douze, Jorge Sánchez, Patrick Perez, and Cordelia Schmid. Aggregating local image descriptors into compact codes. IEEE TPAMI, 34, 2011. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.873, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[30] Nikhil Keetha, Avneesh Mishra, Jay Karhade, Krishna Murthy Jatavallabhula, Sebastian Scherer, Madhava" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "12762" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.092, + 0.469, + 0.134 + ], + "angle": 0, + "content": "Krishna, and Sourav Garg. Anyloc: Towards universal visual place recognition. IEEE Robotics and Automation Letters, 2023. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.135, + 0.469, + 0.188 + ], + "angle": 0, + "content": "[31] Prannay Khosla, Piotr Teterwak, Chen Wang, Aaron Sarna, Yonglong Tian, Phillip Isola, Aaron Maschinot, Ce Liu, and Dilip Krishnan. Supervised contrastive learning. In NeurIPS, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.19, + 0.469, + 0.232 + ], + "angle": 0, + "content": "[32] Hyo Jin Kim, Enrique Dunn, and Jan-Michael Frahm. Learned contextual feature reweighting for image geolocation. In CVPR, pages 3251-3260, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.233, + 0.469, + 0.26 + ], + "angle": 0, + "content": "[33] Diederik Kingma and Jimmy Ba. Adam: A method for stochastic optimization. *ICLR*, 2014. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.261, + 0.469, + 0.302 + ], + "angle": 0, + "content": "[34] María Leyva-Vallina, Nicola Strisciuglio, and Nicolai Petkov. Data-efficient large scale place recognition with graded similarity supervision. CVPR, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.303, + 0.469, + 0.344 + ], + "angle": 0, + "content": "[35] Weiyang Liu, Yandong Wen, Zhiding Yu, and Meng Yang. Large-margin softmax loss for convolutional neural networks. In ICML, pages 507-516. PMLR, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.344, + 0.469, + 0.385 + ], + "angle": 0, + "content": "[36] Weiyang Liu, Yandong Wen, Zhiding Yu, Ming Li, Bhiksha Raj, and Le Song. Spherface: Deep hypersphere embedding for face recognition. In CVPR, pages 212-220, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.387, + 0.469, + 0.426 + ], + "angle": 0, + "content": "[37] David G. Lowe. Distinctive image features from scale-invariant keypoints. Int. J. Comput. Vision, 60(2):91-110, 2004. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.428, + 0.469, + 0.455 + ], + "angle": 0, + "content": "[38] Carlo Masone and Barbara Caputo. A survey on deep visual place recognition. IEEE Access, 9:19516-19547, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.457, + 0.469, + 0.511 + ], + "angle": 0, + "content": "[39] Oscar Manas, Alexandre Lacoste, Xavier Giroi-nieto, David Vazquez, and Pau Rodriguez. Seasonal contrast: Unsupervised pre-training from uncurated remote sensing data. In ICCV, pages 9394-9403, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.512, + 0.469, + 0.58 + ], + "angle": 0, + "content": "[40] Aadil Nathani, Rishi Iyer, Annabelle Wang, Aarnav Chitari, Adele Wilson, and Hannah Norris. Observing Earth From Space: Using Astronaut Photography to Analyze Geographical Climate Patterns. In AGU Fall Meeting Abstracts, pages ED44C-06, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.581, + 0.469, + 0.706 + ], + "angle": 0, + "content": "[41] Maxime Oquab, Timothée Darcet, Theo Moutakanni, Huy V. Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, Russell Howes, Po-Yao Huang, Hu Xu, Vasu Sharma, Shang-Wen Li, Wojciech Galuba, Mike Rabbat, Mido Assran, Nicolas Ballas, Gabriel Synnaeve, Ishan Misra, Herve Jegou, Julien Mairal, Patrick Labatut, Armand Joulin, and Piotr Bojanowski. Dinov2: Learning robust visual features without supervision, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.707, + 0.469, + 0.748 + ], + "angle": 0, + "content": "[42] Florent Perronnin, Yan Liu, Jorge Sánchez, and Herve Poirier. Large-scale image retrieval with compressed fisher vectors. In CVPR, pages 3384-3391, 2010. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.749, + 0.469, + 0.789 + ], + "angle": 0, + "content": "[43] F. Radenović, G. Tolias, and O. Chum. Fine-tuning CNN Image Retrieval with No Human Annotation. IEEE TPAMI, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.79, + 0.469, + 0.831 + ], + "angle": 0, + "content": "[44] A. Razavian, J. Sullivan, A. Maki, and S. Carlsson. Visual Instance Retrieval with Deep Convolutional Networks. CoRR, abs/1412.6574, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.832, + 0.469, + 0.9 + ], + "angle": 0, + "content": "[45] Esther Rolf, Jonathan Proctor, Tamma Carleton, Ian Bolliger, Vaishaal Shankar, Miyabi Ishihara, Benjamin Recht, and Solomon Hsiang. A generalizable and accessible approach to machine learning with global satellite imagery. Nature Communications, 12(1), 2021. 3" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.469, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.135 + ], + "angle": 0, + "content": "[46] Paul-Edouard Sarlin, Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. SuperGlue: Learning feature matching with graph neural networks. In CVPR, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.136, + 0.892, + 0.192 + ], + "angle": 0, + "content": "[47] Aaron Schirmer, Caleb Gallemore, Ting Liu, Seth Magle, Elisabeth DiNello, Humerah Ahmed, and Thomas Gilday. Mapping behaviorally relevant light pollution levels to improve urban habitat planning. Scientific Reports, 9, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.192, + 0.892, + 0.275 + ], + "angle": 0, + "content": "[48] Johannes A. Schultz, Maik Hartmann, Sascha Heinemann, Jens Janke, Carsten Jürgens, Dieter Oertel, Gernot Rücker, Frank Thonfeld, and Andreas Rienow. Diego: A multispectral thermal mission for earth observation on the international space station. European Journal of Remote Sensing, 53(sup2):28-38, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.277, + 0.892, + 0.318 + ], + "angle": 0, + "content": "[49] Peter Schwind and Tobias Storch. Georeferencing urban nighttime lights imagery using street network maps. Remote Sensing, 14(11), 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.32, + 0.892, + 0.361 + ], + "angle": 0, + "content": "[50] Yujiao Shi and Hongdong Li. Beyond cross-view image retrieval: Highly accurate vehicle localization using satellite image. In CVPR, pages 17010-17020, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.362, + 0.892, + 0.404 + ], + "angle": 0, + "content": "[51] Yujiao Shi, Xin Yu, Shan Wang, and Hongdong Li. Cvlnet: Cross-view feature correspondence learning for video-based camera localization. In ACCV, pages 652-669, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.406, + 0.892, + 0.461 + ], + "angle": 0, + "content": "[52] Yujiao Shi, Fei Wu, Akhil Perincherry, Ankit Vora, and Hongdong Li. Boosting 3-dof ground-to-satellite camera localization accuracy via geometry-guided cross-view transformer. In ICCV, pages 21516-21526, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.462, + 0.892, + 0.49 + ], + "angle": 0, + "content": "[53] Christopher Small. Spectrometry of the urban lightscape. Technologies, 10(4), 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.491, + 0.892, + 0.533 + ], + "angle": 0, + "content": "[54] H. Song, Y. Xiang, S. Jegelka, and S. Savarese. Deep metric learning via lifted structured feature embedding. In CVPR, 2016. Spotlight presentation. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.534, + 0.892, + 0.602 + ], + "angle": 0, + "content": "[55] Weiwei Song, Zhi Gao, Renwei Dian, Pedram Ghamisi, Yongjun Zhang, and Jón Atli Benediktsson. Asymmetric hash code learning for remote sensing image retrieval. IEEE Transactions on Geoscience and Remote Sensing, 60:1-14, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.604, + 0.892, + 0.66 + ], + "angle": 0, + "content": "[56] W. L. Stefanov and C. A. Evans. Data collection for disaster response from the international space station. The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, XL-7/W3:851–855, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.661, + 0.892, + 0.745 + ], + "angle": 0, + "content": "[57] Adam J. Stewart, Caleb Robinson, Isaac A. Corley, Anthony Ortiz, Juan M. Lavista Ferres, and Arindam Banerjee. Torch-Geo: Deep learning with geospatial data. In Proceedings of the 30th International Conference on Advances in Geographic Information Systems, pages 1–12, Seattle, Washington, 2022. Association for Computing Machinery. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.746, + 0.892, + 0.801 + ], + "angle": 0, + "content": "[58] Alex Stoken and Kenton Fisher. Find my astronaut photo: Automated localization and georectification of astronaut photography. In CVPRW, pages 6196-6205, 2023. 2, 3, 4, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.802, + 0.892, + 0.843 + ], + "angle": 0, + "content": "[59] Jiaming Sun, Zehong Shen, Yuang Wang, Hujun Bao, and Xiaowei Zhou. LoFTR: Detector-free local feature matching with transformers. CVPR, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.845, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[60] Yifan Sun, Changmao Cheng, Yuhan Zhang, Chi Zhang, Liang Zheng, Zhongdao Wang, and Yichen Wei. Circle loss: A unified perspective of pair similarity optimization. CVPR, pages 6397-6406, 2020. 3" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12763" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.134 + ], + "angle": 0, + "content": "[61] Giorgos Tolias, R. Sicre, and H. Jégou. Particular object retrieval with integral max-pooling of CNN activations. CoRR, abs/1511.05879, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.136, + 0.472, + 0.219 + ], + "angle": 0, + "content": "[62] Ilya O Tolstikhin, Neil Houlsby, Alexander Kolesnikov, Lucas Beyer, Xiaohua Zhai, Thomas Unterthiner, Jessica Yung, Andreas Steiner, Daniel Keysers, Jakob Uszkoreit, Mario Lucic, and Alexey Dosovitskiy. Mlp-mixer: An all-mlx architecture for vision. In NeurIPS, pages 24261–24272. Curran Associates, Inc., 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.222, + 0.47, + 0.277 + ], + "angle": 0, + "content": "[63] Gabriele Trivigno, Gabriele Berton, Juan Aragon, Barbara Caputo, and Carlo Masone. Divide&classify: Fine-grained classification for city-wide visual geo-localization. In ICCV, pages 11142-11152, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.28, + 0.469, + 0.321 + ], + "angle": 0, + "content": "[64] Aäron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding. CoRR, abs/1807.03748, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.324, + 0.469, + 0.393 + ], + "angle": 0, + "content": "[65] Hao Wang, Yitong Wang, Zheng Zhou, Xing Ji, Dihong Gong, Jingchao Zhou, Zhifeng Li, and Wei Liu. Cosface: Large margin cosine loss for deep face recognition. In CVPR, pages 5265-5274. Computer Vision Foundation / IEEE Computer Society, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.396, + 0.469, + 0.464 + ], + "angle": 0, + "content": "[66] Tingyu Wang, Zhedong Zheng, Chenggang Yan, jiyong Zhang, Yaoqi Sun, Bolun Zheng, and Yi Yang. Each part matters: Local patterns facilitate cross-view geo-localization. IEEE TCSVT, 2021. doi: 10.1109/TCSVT.2021.3061265.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.468, + 0.469, + 0.522 + ], + "angle": 0, + "content": "[67] Xun Wang, Xintong Han, Weilin Huang, Dengke Dong, and Matthew R Scott. Multi-similarity loss with general pair weighting for deep metric learning. In CVPR, pages 5022-5030, 2019. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.525, + 0.469, + 0.553 + ], + "angle": 0, + "content": "[68] M. Justin Wilkinson and Yanni Gunnell. Fluvial Megafans on Earth and Mars. Cambridge University Press, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.556, + 0.469, + 0.61 + ], + "angle": 0, + "content": "[69] Yoav Yair, Melody Korman, Colin Price, and Eytan Stibbe. Observing lightning and transient luminous events from the international space station during ilan-es: An astronaut's perspective. Acta Astronautica, 211:592-599, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.613, + 0.469, + 0.655 + ], + "angle": 0, + "content": "[70] Ting Yao, Yiheng Zhang, Zhaofan Qiu, Yingwei Pan, and Tao Mei. Seco: Exploring sequence supervision for unsupervised representation learning. In AAAI, 2021. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.657, + 0.469, + 0.699 + ], + "angle": 0, + "content": "[71] Zhedong Zheng, Yunchao Wei, and Yi Yang. University-1652: A multi-view multi-source benchmark for drone-based geo-localization. ACM MM, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.702, + 0.469, + 0.771 + ], + "angle": 0, + "content": "[72] Zhedong Zheng, Yu jiao Shi, Tingyu Wang, Jun Liu, Jianwu Fang, Yunchao Wei, and Tat-seng Chua. Uavm '23: 2023 workshop on uavs in multimedia: Capturing the world from a new perspective. In ACM MM, page 9715-9717, New York, NY, USA, 2023. Association for Computing Machinery. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.774, + 0.469, + 0.857 + ], + "angle": 0, + "content": "[73] Runzhe Zhu, Mingze Yang, Kaiyu Zhang, Fei Wu, Ling Yin, and Yujin Zhang. Modern backbone for efficient geolocation. In Proceedings of the 2023 Workshop on UAVs in Multimedia: Capturing the World from a New Perspective, page 31-37, New York, NY, USA, 2023. Association for Computing Machinery. 3, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.859, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[74] Sijie Zhu, Taojiannan Yang, and Chen Chen. Vigor: Cross-view image geo-localization beyond one-to-one retrieval. In CVPR, pages 3640-3649, 2021. 3" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.092, + 0.892, + 0.134 + ], + "angle": 0, + "content": "[75] Sijie Zhu, Mubarak Shah, and Chen Chen. Transgeo: Transformer is all you need for cross-view image geo-localization. In CVPR, pages 1162-1171, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.136, + 0.892, + 0.19 + ], + "angle": 0, + "content": "[76] Sijie Zhu, Linjie Yang, Chen Chen, Mubarak Shah, Xiaohui Shen, and Heng Wang. R2former: Unified retrieval and reranking transformer for place recognition. In CVPR, pages 19370-19380, 2023. 3" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.092, + 0.892, + 0.19 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12764" + } + ] +] \ No newline at end of file diff --git a/2024/EarthLoc_ Astronaut Photography Localization by Indexing Earth from Space/09848c3c-ff6b-4e53-9236-ab0f624d8495_origin.pdf b/2024/EarthLoc_ Astronaut Photography Localization by Indexing Earth from Space/09848c3c-ff6b-4e53-9236-ab0f624d8495_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..fc20521cb58e2c8c94ce742d7a0766f3258db3b2 --- /dev/null +++ b/2024/EarthLoc_ Astronaut Photography Localization by Indexing Earth from Space/09848c3c-ff6b-4e53-9236-ab0f624d8495_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5d1b35ced1ae9522c0ac05c7c1eb149be3f56cfca3205a01ad2a668135cb188b +size 3439157 diff --git a/2024/EarthLoc_ Astronaut Photography Localization by Indexing Earth from Space/full.md b/2024/EarthLoc_ Astronaut Photography Localization by Indexing Earth from Space/full.md new file mode 100644 index 0000000000000000000000000000000000000000..5a2766aeb3f50dad6fb9171b98dc307992fc7c37 --- /dev/null +++ b/2024/EarthLoc_ Astronaut Photography Localization by Indexing Earth from Space/full.md @@ -0,0 +1,348 @@ +# EarthLoc: Astronaut Photography Localization by Indexing Earth from Space + +Gabriele Berton1 Alex Stoken2 Barbara Caputo1 Carlo Masone1 +1Politecnico di Torino 2Jacobs Technology, NASA Johnson Space Center +gabriele.berton@polito.it + +# Abstract + +Astronaut photography, spanning six decades of human spaceflight, presents a unique Earth observations dataset with immense value for both scientific research and disaster response. Despite their significance, accurately localizing the geographical extent of these images, which is crucial for effective utilization, poses substantial challenges. Current, manual localization efforts are time-consuming, motivating the need for automated solutions. We propose a novel approach - leveraging image retrieval - to address this challenge efficiently. We introduce innovative training techniques which contribute to the development of a high-performance model, EarthLoc. We develop six evaluation datasets and perform a comprehensive benchmark comparing EarthLoc to existing methods, showcasing its superior efficiency and accuracy. Our approach marks a significant advancement in automating the localization of astronaut photography, which will help bridge a critical gap in Earth observations data. Code and datasets are available at https://github.com/gmberton/EarthLoc. + +# 1. Introduction + +Astronaut photography of Earth is a unique remote sensing dataset that spans 60 years of human spaceflight, offering a rare perspective on our planet to the public and valuable data to Earth and atmospheric science researchers. This dataset contains over 4.5 million images and is growing by the tens of thousands per month, as astronauts are continually tasked with taking new photographs that enable scientific research as well as assist in natural disaster response efforts in the wake of events like floods and wildfires. To effectively use these images, the geographical area depicted in them needs to be identified. Unfortunately, this task - Astronaut Photography Localization (APL) - is very challenging. For each photo, only a coarse estimate of location is known, given by the point on Earth directly under the International Space Station (ISS) at the time the photo is taken. This point - called the nadir - can be easily computed using the image's timestamp and the ISS's orbit path. However, two + +![](images/306c03fd150f2bcdbeb0615054bd2cd5358be753de30f69f07101315d886e142.jpg) +Figure 1. Overview of the astronaut photography localization task. Astronauts take hundreds of photos a day from the International Space Station (ISS) cupola (top-left) with hand-held cameras. For each image (example bottom right), the geographic location depicted is not known, and needs to be searched for across a huge area centered at the ISS's (known) nadir point at the moment that the photo is taken. A simulated view of the astronaut's perspective when the ISS is above Europe is shown. The goal of our paper is to automate the task of localizing these images, which could be anywhere within the view. In the figure's example, the photo the astronaut took is indicated by the green line and shown in inset - other possible photo extents are in red, illustrating the wide array of potential locations to search. + +images taken with the same nadir can be thousands of kilometers apart, as even a slight inclination of the astronaut's hand-held camera can move the image's location hundreds of kilometers in any direction, as depicted in Fig. 1. Localization must thus be performed over a wide area, and is additionally complicated by (i) astronauts using hand-held cameras and a variety of zoom lenses, (ii) the large, 2500 kilometer (km) visibility range in all directions, (iii) most photographs being oblique, and (iv) the Earth's appearance + +changing over time, mostly due to weather or seasonal variations. Despite these challenges, the high value of properly localized imagery for Earth science research and disaster response has lead to more than 300,000 [21] astronaut photos being manually localized by identifying the geographic coordinates of the photo's center pixel. This process can take minutes to hours per photo for experts in NASA's Earth Science and Remote Sensing Unit (ESRS) and citizen scientists through ESRS's Image Detective program. The resulting geo-located images have lead to a multitude of peer reviewed papers, technical reports, and articles about light pollution and urban planning [15, 22, 47, 53], atmospheric phenomena [28, 69], changes in land usage and glacial extent [40], and other Earth science topics [27, 48, 68]. Most importantly, astronaut photography has a fast response time, crucial for disaster management - astronauts are alerted of a disaster, provide photos to a ground crew which localizes them, and sends them to first responders [56]. For example, in 2013 this protocol was activated in response to Cyclone Haiyan in the Philippines, wildfires in Australia, flooding in China, Russia, Pakistan, and the USA, and multiple other events. [56]. + +Even with the abundance and importance of satellite imagery in modern applications, astronaut photography fills an unserved gap among other remotely sensed data. Unlike satellites that nominally take top-down imagery in similar illumination conditions at fixed temporal intervals, astronaut photographs can show topography via oblique views from multiple orientations, are taken in various lighting conditions (including nighttime), and vary in focal length to show detail at different resolutions. These qualities produce a complementary data product that would be difficult or impossible to gather from traditional satellites. Additionally, having a human in the loop for data collection allows for real-time response to natural disasters as well as a natural sense to avoid clouds and other obstructions. + +Given the uniqueness and importance of astronaut photography, and the large amount of time spent by human experts on geo-localizing them, researchers have been exploring solutions to automate the task through computer vision methods. Previous works have shown promising results using a pairwise matching setup that iteratively compares an astronaut photo to satellite imagery [58]. Yet, such methods suffer from high latency, searching multiple directions sequentially and using compute-heavy, dense correspondences to determine matching. The time required to find the location of all 4.5 million images using these methods is estimated to be over 20 compute years [58]. Furthermore, low latency is of paramount importance for pho + +tography in areas that are affected by natural disasters (e.g., hurricanes or wildfires), for which speedy localization can help real-time, on-the-ground operations. + +To overcome the high latency of current methods, we propose to instead localize astronaut photography through image retrieval, by matching each astronaut photograph to a worldwide database containing satellite images of known position. In reformulating this problem as a retrieval task, we encounter a range of new challenges, from efficiently training a robust model to achieving low-latency inference, as well as defining a success metric and creating evaluation sets, both of which are crucial pieces when assessing how well a given model will perform when deployed to localize astronaut photography. + +The ultimate goal of this paper is twofold: (i) provide an efficient method to localize the archive of 4.5M queries as well as all new, incoming imagery, so that researchers studying a given geographical area can have access to a large amount of imagery spanning many years; and (ii) prove the viability of world-wide APL through image retrieval (which has never been attempted and was believed to be infeasible4) in order to spark a new line of research, with direct benefit to all space-based photography and its users across the globe. + +To achieve its goal, this paper makes the following contributions: + +- we propose approaching APL through image retrieval; +- we develop novel training techniques which show large quantitative improvement on the task; +- we provide six evaluation sets and a large benchmark of results with methods from a variety of relevant domains; +- we show that a model trained with these new techniques, EarthLoc, allows us to localize a large number of images, widely outperforming all other methods while being fast and efficient. + +Finally, we note how EarthLoc is being used to localize ISS astronaut photographs, which are publicly available and conveniently searchable at https://eol.jsc.nasa.gov/ExplorePhotos/. + +# 2. Related Work + +Image Retrieval Image retrieval involves searching a database for images similar to a query image. Traditionally, methodologies combined hand-crafted local features like SIFT [37], SURF [8], RootSIFT [3] into a global embedding through means like Bag of Words [13], Fisher Vectors [42] or VLAD [29], to allow for fast retrieval through kNN. With the rise of deep learning, local features have been replaced by CNN-derived features [6], significantly + +improving retrieval performance. Since then, most of the attention of the retrieval community has been focused on how to best aggregate feature maps and efficiently train robust neural networks. For feature map aggregation, some of the most notable examples in literature include Max pooling [44], Regional Max pooling (R-MAC) [61] and Generalized Mean (GeM) [43]. To train these retrieval models, a number of losses have been proposed, which can be generally grouped into two categories: stateful losses, which have weights whose size depends on the number of classes, like the Large-Margin Softmax [35], SphereFace [36], CosFace [65] and ArcFace [16]; and their stateless counterparts, like the Contrastive and Triplet losses, and the more recent Lifted Structure Loss [54], NTXentLoss [64], Multi-Similarity Loss[67], FastAP [11], Supervised Contrastive [31] and the Circle Loss [60]. + +Geo-localization and retrieval with Aerial and Remote Sensing Imagery To the best of our knowledge, only two previous works have specifically addressed the localization problem for astronaut photography. Both rely on local features and compute pairwise comparisons between astronaut photography and geo-located reference imagery. Find My Astronaut Photo [58] focuses on daytime imagery, presenting an evaluation of multiple methods as well as the note that pre-trained self-supervised models were not suitable for a retrieval-based approach to this task. Schwind and Storch [49] instead address nighttime imagery, using synthetically generated street light maps as reference. Both works discuss the importance of proper orientation for matching and the high cost of pairwise comparisons. + +More commonly, studies of aerial imagery focus on data from unmanned aerial vehicles (UAVs) and satellite platforms. Aerial geo-localization is an important aspect of UAV navigation systems. The University-1652 dataset [71] and associated challenge encourage work in cross-view geo-localization for drone-satellite imagery, with strong approaches modifying different parts of the general pipeline, from backbone [73] to training setup [18], to feature partitioning [66]. Most use contrastive and self-supervised losses in training. Other works address cross-view geo-localization in more extreme view point differences, like that between aerial and street views [17, 20, 50-52, 74, 75]. + +Remotely sensed data from satellites, on the other hand, often comes with reliable geo-location information from the sensor itself. Extracting features from satellite imagery is still a common task, with such features most often trained on a pretext objective before use on downstream tasks [5, 12, 39, 45]. Occasionally, features are used for retrieval itself [55], to identify areas of the Earth that are similar in appearance and thus might have similar properties to be studied together. In this paper we take a separate direction from previous literature, and aim at using image + +![](images/2f0079bf20b434a2931132166f9337d4bee16a0d1eb527d4a2af3b53155c734b.jpg) + +![](images/fbf8302a98adbf5a1b2ba44d37a91e9fb5f5d1fb519f46a2036e34c5e5761641.jpg) + +![](images/db1eaf7a3f1e24fedaeff0859841a2163133c83faa378cdba6457089d880f879.jpg) + +![](images/c8f695f11d7b80967238d9fc90f867b4115cc3343d2cd411d2b4c44a0014c0cb.jpg) +Figure 2. Astronaut photo query examples, showcasing the large variability in covered area and appearance. + +![](images/4b5b11011707b7e985bdae857a14c2f35e6ee8128514cba300809aeabcdf2064.jpg) + +![](images/0c5720bcdb102fd8a201f9a8149d19b734e9a4371b9eb9217cfa5434dac4a455.jpg) + +retrieval techniques for APL. + +Other localization tasks Various other localization challenges are approached through image retrieval, including visual place recognition (VPR) and visual localization. The former aims at coarsely localizing a given query by matching it to a database of geo-tagged photos, and is mostly studied within urban environments using street-view imagery [7, 63]. Performance is improved either by the use of smart aggregation (e.g. NetVLAD [4]), attention layers (e.g. CRN [32]), or through large-scale training [1, 2, 9, 34]. More recently, the task of universal visual place recognition has been proposed by AnyLoc [30], whose authors provide a model that performs competitively on a large range of scenarios, including aerial imagery. The task of visual localization is focused on finding the precise camera pose of an image, and is commonly approached through feature matching techniques [7, 38, 46, 59]. It can also be used for visual place recognition [24, 76], although this leads to a noticeable increase in runtime when compared to pure retrieval. + +# 3. Dataset + +In this section we discuss the datasets that we use in this work. Specifically, in Sec. 3.1 we introduce the images that we need to localize, called queries; in Sec. 3.2 we outline the collection of the database, made of satellite images (i.e., taken automatically, not with a hand-held camera) of known location; and in Sec. 3.3 we describe the creation of the evaluation datasets that we use to understand the capabilities of our models, combining queries and database in a way that will reflect the real-world use case of our method. + +# 3.1. Queries + +Queries are taken from the Gateway to Astronaut Photography of Earth $^{5}$ , a collection of over 4.5 million photos of Earth taken by astronauts on the International Space Station (Fig. 2). This unique photography setting yields only coarse location information for each photograph, namely the position on Earth below the ISS (the nadir point) at the time the photo was taken. We seek to find the geographic area on the Earth that each photo encompasses (the photo's location), + +![](images/0261613d2655e91a2e02144006f754a2b00be1f2ff93651d9d66a20711547ebe.jpg) + +![](images/4c1e94bbae9a8fd5ea35601da20353a5b45b2b86c83a6365111f7468fdc03865.jpg) + +![](images/2a76d37bed00f53ccbd74b9b783cdc18a4566330491d3f21b287de192959200a.jpg) + +![](images/84371a693e602f29fc8e8ff84d60f67964dee33653ca998e538f56a1083a61f5.jpg) + +![](images/99e2ceef17f8a5c6c099c9214ca03d72799641480fc5f27dc4f678d47ae0317c.jpg) +Figure 3. Examples of database images. Top: four images from different regions, with $25\%$ or $50\%$ overlap between any pair. The red dot in each image represents the same geographic point. Bottom: 4 images from the same region across different years. + +![](images/e105281a6895053b103e5a7466d139bb4db6b2e4fd30ea96d5a80ca737dcffef.jpg) + +![](images/42f578694e46a3b97f73fc8651a7343093ec202fe314dbebf8425bad1d8022dd.jpg) + +![](images/1a5387552fd0583aacc6454dfae4af460e6e66d7915cef119a9c6b37c9c1c881.jpg) + +which can be thousands of kilometers away from the ISS nadir point. This distance of maximum visibility (horizon) can be quickly computed as + +$$ +d _ {\text {v i s i b l e}} = \sqrt {2 R h + h ^ {2}} = 2 4 3 6 \mathrm {k m} \tag {1} +$$ + +where $R = 6317$ is Earth's radius and $h = 450$ is the ISS maximum orbiting altitude. We follow common practice and round this number to 2500 kilometers [58]. + +Challenges to APL Some astronaut photo queries, taken with high focal length lenses, can cover only a few kilometers within the almost 20 million sq. km area visible to the astronaut taking the photo, which makes localizing the images extremely challenging. Additional hurdles presented by the imagery acquisition process include the varying quality of the images themselves due to motion blur from the fast moving space station, and occlusion of unique, location-identifying features by cloud cover, shadow, or other portions of the ISS (Fig. 2). More queries are shown in Sec. 10 of the Supplementary. + +From the 4.5 million photos available, we consider only the day-time images that have a ground truth obtained by FMAP [58], and keep only those which cover areas between 5000 and 900,000 sq. km, in order to match the areas of database images (see Sec. 3.2). This leads to 17,764 queries which we can use for validation and testing. + +# 3.2. Database + +To estimate the location of the queries through image retrieval, we need a worldwide database of images, where each image is labeled with its position. We build such a database from an open-source, composited, cloudless collection of Sentinel-2 satellite imagery6. This collection contains imagery of worldwide landmasses at resolutions up to 15 meters per pixel. Some examples are shown in Fig. 3. + +For our database we include any land area between latitudes $60^{\circ}$ and $-60^{\circ}$ (i.e., the area traversed by the ISS). The + +![](images/33b3b1bed9735d37e312500862e5cda6e8b4d980cb872faa7031a76982e95ee0.jpg) +Figure 4. To create evaluation sets we choose all images that could contain a Point of Interest (POI) - all photos with nadir point within $d_{visible}$ . To localize all photos within this range, even if they do not contain the POI, we create a database that contains all areas visible from the nadir points of all selected photos, yielding a database area of about $5000\mathrm{km}^2$ per POI. + +
Name(Lat, Lon)# queries# DBInterest
Texas(30, -95)614234kOften photoed (many queries)
Alps(45, 10)239453kSci. interest - glacial change
California(38, -122)356830kDisaster response - wildfires
Gobi Desert(40, 105)72654kSci. interest - desertification
Amazon(-3, -60)68219kSci. interest - deforestation
Toshka Lakes(23, 30)216463kSci. interest - flood monitoring
+ +Table 1. Evaluation sets. The six sets cover diverse geographies and relevant areas for research on topics like climate change and disaster management. DB stands for database. + +Sentinel-2 imagery is available as map tiles, and we use tiles at zoom levels 9, 10, 11, with resolutions ranging from $\sim 300$ to $\sim 75$ meters per pixel7, to ensure that the database encompasses the large scale range of the query images. + +These parameters lead us to define roughly 175k regions, where each region $\mathcal{R}$ is the projection of a square with known corner coordinates onto the Earth's surface. Regions have areas ranging from 5000 sq. km to 900,000 sq. km. We take partially (up to $50\%$ ) overlapping regions so that each query has at least one image in the database with considerable overlap (Fig. 3 top). From each region we collect four images, one per year from 2018, 2019, 2020 and 2021, covering temporal changes in a given area (Fig. 3, bottom). This produces a database of 700k images. + +Unlike our queries, these database images have undergone post-processing that includes atmospheric corrections as well as cloud pixel minimization via composition of images from multiple days. Additionally, the database images are always taken nadir facing (i.e. perpendicular to Earth's surface) to minimize obliquity effects. Each database image has resolution $1024 \times 1024$ . Further images from the database are shown in Sec. 10 of the Supplementary. + +# 3.3. Evaluation sets + +With the end goal of fast and accurate localization of the astronaut photography archive, we propose a deployment + +![](images/7e61e7cb16e152fa920fc6f0f1a235a889bbc949d1e81cd9b07bffdec444231e.jpg) +Figure 5. Training strategy. Naive batching produces too easy a task for the model to learn robust representations from. We increase difficulty by clustering batches by similarity and adding year-wise augmentations, which move images of different regions, but from the same year, closer together in feature space. The model must then learn representations that discern similarity in this more difficult context. + +environment that takes advantage of the known nadir points and visibility limits associated with each image. So while localization can be performed by matching any given query against the whole world-wide database, a more suitable approach is to follow a divide and conquer paradigm by grouping queries according to their nadir, which allows us to use only a subset of the database for each search, increasing speed and improving results. + +We seek to create evaluation sets that mirror the situation seen in deployment, so we evaluate on sets of queries that have a nadir within a $2500\mathrm{km}$ radius from a chosen Point of Interest (POI). We then build a database of images such that the database fully encompasses the area that could have been photographed from each collected query, which is an area $2*d_{visible}$ from the original POI, as shown in Fig. 4. + +Since we need the queries' ground truth position for evaluation, we choose queries that have already been localized with exhaustive image matching techniques [58]. For these queries, the full photo extent is publicly available8. + +To properly evaluate different methods, we choose six POIs that represent various geographies but all have scientific relevance, and center an evaluation set around each. These are summarized in Tab. 1. + +# 4. Method + +The goal of this work is to estimate the location of each astronaut photo query, which we tackle through image retrieval: for any given query, the objective is to find the most similar image(s) from the database, and use the corresponding location labels to confidently localize the query. With this in mind, we aim to train a retrieval model on satellite imagery such that the model is robust to changes in scale, perspective, and color, as well as temporal/seasonal variation. Specifically, the model's task is to extract features + +from each image. Then, for each query, the most similar database image is retrieved via kNN in feature space. + +In the next subsections we describe how we train our retrieval model using the proposed database. + +# 4.1. Baseline Training + +We want our model to extract relevant features of satellite imagery, so a straightforward approach would be to train via contrastive learning: for a given batch of images (samples), generate a number of augmented views (typically two), and apply a contrastive loss to maximize the distance in feature space between different samples and minimize distance between different views of the same sample. However, this solution would not allow the model to learn robustness to temporal changes, because the augmented views come from the same sample, with no inherent temporal variation. + +Therefore, instead of producing multiple views of the same image via synthetic augmentation, we train the model by randomly selecting a number of regions, and for each region we consider its four images (a quadruplet), one per year. Within a batch of quadruplets, the four images from each region are mutual positives, whereas any two images from different regions are considered negatives: we can therefore consider each region as a class. This provides a natural way to address temporal variation. We can then train the model through a multi-similarity loss [67]. This idea is shown in Fig. 5 (left). + +# 4.2. Clustered Batches + +The baseline paradigm simply creates batches of random regions – a single batch of quadruplets could contain samples depicting shorelines, islands, deserts, and mountains. The loss operates batchwise, so large intra-batch variability can lead the model to learn too swiftly to distinguish between such diverse regions. In these situations, training converges quickly and the model stops short of learning to discriminate between similar-looking, but distinct, regions. + +To avoid this issue, we want to train with batches of highly similar images, which are still from different regions. To this end, we compute clusters of images with similar representations. Every $N_{iter}$ training iterations, we extract the features from each region with the trained model, using only images from a single year, and group them into $C$ clusters. + +Then, we create training batches using samples from a given cluster, choosing a different cluster for each batch. This increases the difficulty of the task in each batch, ensuring that the model learns to discriminate between similar-looking images of different regions. Note that clusters are not geographically related, but related in feature space. For example, a cluster may contain images from the Gobi and the Sahara deserts, which both appear similar in feature space despite being geographically distant. This procedure, which can be seen as an offline mining approach, increases training time by less than $5\%$ but improves recall by roughly $15\%$ on average (see Sec. 5.3), a worthwhile trade off. + +This concept is illustrated in Fig. 5 (middle), and qualitative examples of images that are clustered together are shown in Sec. 11.2 of the Supplementary. + +# 4.3. Year-Wise Data Augmentation + +Query images span the ongoing $20+$ years of ISS operations, as well as multiple camera types, camera operators, and view points, so queries depicting the same location can have a number of natural "transformations". These range from blueish hues from lens filters, to skewed representations in highly oblique shots, to seasonal variation in ground cover (e.g. snow, autumn colors). + +To make the model robust to such transformations, we employ data augmentation. Typically, augmentations are applied per image or per batch, but we argue that to truly learn robustness to the augmentations in this setting, the same augmentation should be performed on images from different regions, but from the same year. + +Formally, at each training iteration we choose a set of 4 random augmentations, one for each year $A_{2018}, A_{2019}, A_{2020}, A_{2021}$ . We then apply the corresponding augmentation to all images from that year (i.e. $A_{2018}$ is applied to all images from 2018, across all regions). Such "year-wise" augmentation has the effect of moving images from a given year closer in feature space to each other, as shown in Fig. 5 (right). In the loss, images from the same region, but different years, are treated as positives, while images from different regions are treated as negatives. This has the effect that images that receive the same augmentation are assigned different classes – forcing the model to learn to ignore such augmentations – and instead pull together images from the same region, across different years. + +Furthermore, given that the same augmentation is applied to multiple images, augmentation can be batched and performed on the GPU, making training slightly faster (by + +![](images/57ed52742026028b9e6109c4422fbc58eb7526e4700d658a56c4124fc9e8dae4.jpg) +Figure 6. Dataset collection from satellite imagery and handling in Neutral-Aware Multi-Similarity Loss. For a given image (yellow), images with no overlap are negatives (red), images from the same region are positives (green, overlapping yellow), and images with partial overlap are neutral (grey). Note that the database is made of images at 3 different zooms (each double in size than the next), which explains the regions of different size. + +$5\%$ on average, although this depends on the hardware). + +# 4.4. Neutral-Aware Multi-Similarity Loss + +The powerful Multi-Similarity loss is designed to take as input multiple images from a number of classes, along with the corresponding class indices, and compute a loss from the positives and negatives for each image. + +However, it does not take into account that in some cases, pairs of images are neither positives nor negatives, and should be considered "neutral". While the concept of neutral images does not exist in standard image retrieval tasks, as an image either belongs to class A or class B, in APL it must be taken into consideration. For example, should two images, which have an intersection over union of $25\%$ , be considered positives or negatives? This is precisely the case when two images of the same area, but from different zoom levels, are batched together, as well as overlapping images at the same level (see Fig. 6). We believe the answer is neither, so we add the "neutral" case to the loss. + +Formally, we model the neutral case with the indicator function $\mathcal{I}_{NA}$ , which is 0 when the regions $\mathcal{R}_i$ and $\mathcal{R}_j$ intersect but are not equal, and 1 otherwise. + +$$ +\mathcal {I} _ {N A} \left(\mathcal {R} _ {i}, \mathcal {R} _ {j}\right) = \mathbb {1} _ {\left(\mathcal {R} _ {i} \cap \mathcal {R} _ {j}\right) \wedge \left(\mathcal {R} _ {i} \neq \mathcal {R} _ {j}\right)} \tag {2} +$$ + +With this, the neutral-aware multi-similarity loss becomes + +$$ +\begin{array}{l} \mathcal {L} _ {N A M S} = \frac {1}{B S} \sum_ {i = 1} ^ {B S} \left\{\frac {1}{\alpha} \log \left[ 1 + \sum_ {k \in \mathcal {R} _ {i}} e ^ {- \alpha \left(\mathcal {S} _ {i k} - \lambda\right)} \right] \right. \\ \left. + \frac {1}{\beta} \log \left[ 1 + \sum_ {k \notin \mathcal {R} _ {i}} \mathcal {I} _ {N A} \left(\mathcal {R} _ {i}, \mathcal {R} _ {k}\right) e ^ {- \beta \left(\mathcal {S} _ {i k} - \lambda\right)} \right] \right\} \end{array} \tag {3} +$$ + +where $BS$ is the number of images in a batch, $\mathcal{R}_i$ is the + +
MethodType of Training ImageryTexas (validation)AlpsCaliforniaGobi DesertAmazonToshka Lakes
R@1R@10R@100R@1R@10R@100R@1R@10R@100R@1R@10R@100R@1R@10R@100R@1R@10R@100
Nadir-2.4--1.2--2.4--1.8--3.1--1.4--
Random Choice-0.21.715.50.11.111.60.22.320.10.11.013.20.11.111.50.21.29.1
NetVLAD [4]Urban VPR3.613.434.85.215.236.24.314.439.02.17.724.56.619.946.39.622.545.6
SFRS [23]Urban VPR6.817.440.97.017.841.45.417.244.45.214.737.69.423.852.510.820.442.4
Conv-AP [1]Urban VPR3.713.537.72.08.825.03.815.441.23.79.024.27.021.147.75.012.332.6
CosPlace [9]Urban VPR4.915.242.05.919.345.24.815.243.37.917.539.58.523.954.49.825.453.0
MixVPR [2]Urban VPR4.515.340.13.18.622.64.417.543.24.010.124.59.828.758.75.915.436.5
EigenPlaces [10]Urban VPR6.421.652.48.721.450.36.022.854.38.120.040.910.926.557.914.330.960.4
AnyLoc [30] (DINOv2 [41] + NetVLAD)Universal VPR44.168.787.840.770.892.048.775.091.628.757.081.738.663.886.263.784.596.3
TorchGeo [57] (ResNet50 w MOCO [26])Satellite1.03.211.60.31.35.71.14.316.40.42.88.31.96.619.40.72.99.9
TorchGeo [57] (ResNet50 w SeCo [70])Satellite6.115.641.77.420.249.25.114.537.13.714.038.94.613.332.95.715.638.5
TorchGeo [57] (ResNet50 w GASSL [5])Satellite9.722.846.49.123.150.513.331.458.86.317.545.48.320.340.120.438.664.2
OGCL UAV-View [18] (ConvNeXt-XXLarge)UAV16.235.365.814.434.164.719.642.171.05.520.545.310.026.354.721.138.463.4
OGCL UAV-View [18] (ViT-L/14) [19]UAV17.633.255.914.633.263.922.848.174.97.622.850.120.439.162.531.851.874.4
MBEG [73] (ViT-L/14)UAV7.017.635.16.619.345.98.720.741.54.415.038.06.417.139.38.120.749.1
EarthLoc (Ours)Satellite55.973.088.358.476.889.558.076.091.451.167.586.547.267.984.672.285.093.3
+ +Table 2. Results from different methods on our query sets. Methods are grouped into (1) naive, (2) VPR, (3) Remote Sensing (UAV+Satellite), (4) other baselines trained on our database with clustered batches (Sec. 4.2). For all methods we use $4x90TTA$ . VPR stands for Visual Place Recognition, UAV for Unmanned Aerial Vehicle. Best results overall in bold, best per-group underlined. EarthLoc outperforms other models, and performs competitively with AnyLoc, while having 50 times faster features extraction and 10 times smaller features. + +region of image $i$ , $\mathcal{S}_{ik}$ is the similarity between image $i$ and $j$ , $\lambda$ is a margin, and $\alpha, \beta$ are hyperparameters. + +We believe that the neutral-aware multi-similarity loss is applicable to any task where it is not straightforward how to enforce full positivity or negativity between any pair of images, including localization tasks like VPR. + +# 5. Experiments + +# 5.1. Experimental setting + +# 5.1.1 Training + +In training, we use a batch size of 96 quadruplets (384 images), where one quadruplet corresponds to four images within a region. Our model has a MixVPR-style architecture [2], with a ResNet50 [25] backbone and an MLP-Mixer[62] with output dimension 4096. We train for 10k iterations with the Adam optimizer [33] and learning rate 0.0001. Clustering (see Sec. 4.2) is performed every 2k iterations with the number of clusters $C = 50$ . As augmentation, we use torchvision's color jittering, random perspective, and random rotation. Training takes between 5 and 7 hours on an A100 GPU for all methods described in Sec. 4. + +# 5.1.2 Inference + +Due to the imagery acquisition conditions, we cannot assume any canonical orientation for the photos (in space, there's no gravity vector, i.e., no "up" or "down"). + +This poses a challenge for any retrieval method – to address it, we generate four feature embeddings for each database image, one each after rotating the image for angles $r \in [0,90,180,270]$ , in a process we call $4 \times 90^{\circ}$ test time augmentation (4x90TTA). For fairness, we do this for all methods in all experiments, given that the models are not trained to be robust to such large arbitrary rotations. This increases latency by $4\mathrm{x}$ , but the method is still orders of magnitude faster than previous approaches (retrieval takes 0.05 + +seconds per query with $4x90TTA$ , while previous works required $\sim 1$ minute [58] per query). Memory requirements also increase by four times, although this is far from being an issue for a standard workstation, since all database features for the whole Earth fit in less than 12GB of RAM. Experiments without $4x90TTA$ are shown in the Supplementary in Sec. 7. As an added bonus, since each feature from the database refers to an image with a certain orientation (w.r.t. North), retrieval predictions also give an estimate of the orientation of the query. We use features from the 2021 satellite imagery as our global database. + +Success Criteria With the final goal of localizing the 4.5M astronaut photograph queries in mind, we define a predicted image as correct if there is non-zero overlap between the prediction and the query. As a metric we use Recall@N, defined as the percentage of queries for which at least one of the top-N predictions is correct. + +# 5.2. Results + +In Tab. 2 we present results from a large number of models from multiple domains. We first show the recall when naively predicting the query to be directly nadir, proving that the vast majority of images are taken at an angle, and the poor performance of randomly choosing a prediction from the database - a symptom of the large-scale nature of the task (i.e., localizing a 5000 sq.km wide image within a 20M sq.km area). + +Results with methods trained for Urban Visual Place Recognition (VPR) provide, somewhat unsurprisingly, poor results - on the other hand, AnyLoc [30] shows remarkably strong results, correctly predicting the location of the query as the first prediction for roughly $50\%$ of the queries across all evaluation sets, although at the cost of over 50x slower feature extraction and 10x bigger features than its VPR counterparts (AnyLoc uses 49152-D features vs 4096 of NetVLAD). + +
Multi-similarity LossUsing regional quadrupletsClustered BatchesYear-Wise Data AugmentationNeutral-Aware Multi-SimilarityTexas (validation)AlpsCaliforniaGobi DesertAmazonToshka LakesAverage
R@1R@100R@1R@100R@1R@100R@1R@100R@1R@100R@1R@100R@1R@100
✓25.571.424.877.327.876.616.868.823.661.538.880.826.272.7
✓✓45.084.145.487.349.588.535.779.934.280.262.991.145.485.2
✓✓✓51.487.549.689.553.990.640.884.339.382.864.093.049.888.0
✓✓✓✓54.687.553.987.255.991.646.882.945.682.467.691.954.187.2
✓✓✓55.388.255.190.057.290.344.685.848.484.868.992.954.988.7
✓✓✓✓55.988.358.489.558.091.451.186.547.284.672.293.357.188.9
+ +Table 3. Ablations over test sets. EarthLoc uses Clustered Batches (Sec. 4.2), Year-Wise Data Augmentation (Sec. 4.3) and Neutral-Aware Multi-Similarity Loss (Sec. 4.4). Best results in bold. + +![](images/18d34865b1211f2e2c026250989968e605577ab93e31d82c1152e596d2fbd645.jpg) +Figure 7. Qualitative results. Each group of four images represents a query (left) and its top-3 predictions. Green and red indicate if the prediction is correct or incorrect. + +We then test a number of models trained with self-supervised techniques, like MOCO [26], SeCo [70], and Geography-aware self-supervised learning (GASSL) [5], conveniently provided by the TorchGeo library [57]. Surprisingly, these methods did not provide competitive results, and are on average about equal with Urban VPR methods. + +From the Unmanned Aerial Vehicle (UAV) domain, we use the latest state of the art (SOTA) methods, i.e., the winners of the recent challenge in UAV localization at ACM MM (October 2023) [72], namely OGCL [18] and MBEG [73]. Despite being SOTA in the related task of UAV localization, we see that these methods are not competitive with AnyLoc, despite outperforming most other methods. + +Finally, EarthLoc outperforms all previous works, localizing $57.1\%$ of queries within the top-1 prediction (Recall@1) averaged on the 6 evaluation sets. + +# 5.3. Ablations + +In Tab. 3 we ablate the components of our method to better understand performance. Results show that not only do Clustered Batches, Year-Wise Data Augmentation and Neutral-Aware Multi-Similarity Loss provide significant increases over the baseline, but they are also orthogonal improvements and can be easily combined. + +# 5.4. Qualitative Analysis and Failure Cases + +A sample of qualitative results from EarthLoc are shown in Fig. 7. The samples show that EarthLoc picks locations that have similar characteristics to the query (top left), is robust to rotations and changes in scale (bottom right), and provides correct predictions that are difficult to correctly match even for the human eye (top left, bottom left). More qualitative results in Sec. 11 of the Supplementary. + +# 5.5. Limitations and Future Work + +While this work takes a major step forward in the study of APL, EarthLoc focuses on daytime queries, and we did not investigate panoramic or nighttime imagery. The localization of nighttime astronaut photos in particular is important (e.g. to study light pollution [14]), but we believe this to be a separate challenge that should be addressed in future work. Another limitation is due to the zoom limits of the database: we chose database tiles at zoom levels 9, 10, 11 (see Sec. 3.2), which allows us to localize queries of area between 5000 and 900,000 square km. + +The originality of this work opens up a large number of possible future research directions, like training directly on (some of the) localized queries, using seasonal data, post-processing predictions, obtaining pixel-wise geolocation, applying domain adaptation, and using available metadata (like camera lens) during training, to name a few. + +# 6. Conclusions + +We introduce EarthLoc, a method for localizing astronaut photography of Earth. By reframing APL from an image matching task to an image retrieval task, and introducing a custom loss function and training scheme, our method efficiently determines the geographic location of astronaut photography given only the position of the International Space Station at the time the photo was taken. Our main contributions are the Neutral-Aware Multi-Similarity Loss, Year-Wise Data Augmentation technique, and new astronaut photography-oriented validation and test datasets to encourage future work on this problem. + +Acknowledgements. We acknowledge the CINECA award under the ISCRA initiative, for the availability of high performance computing resources. This work was supported by CINI. Project supported by ESA Network of Resources Initiative. This study was carried out within the project FAIR - Future Artificial Intelligence Research - and received funding from the European Union Next-GenerationEU (PIANO NAZIONALE DI RIPRESA E RESILLENZA (PNRR) - MISSIONE 4 COMPONENTE 2, INVESTIMENTO 1.3 - D.D. 1555 11/10/2022, PE00000013). This manuscript reflects only the authors' views and opinions, neither the European Union nor the European Commission can be considered responsible for them. European Lighthouse on Secure and Safe AI - ELSA, HORIZON EU Grant ID: 101070617 + +# References + +[1] Amar Ali-bey, Brahim Chaib-draa, and Philippe Giguère. Gsv-cities: Toward appropriate supervised visual place recognition. Neurocomputing, 513:194-203, 2022. 3, 7 +[2] Amar Ali-bey, Brahim Chaib-draa, and Philippe Giguère. Mixvpr: Feature mixing for visual place recognition. In WACV, pages 2998-3007, 2023. 3, 7 +[3] R. Arandjelović and Andrew Zisserman. Three things everyone should know to improve object retrieval. In CVPR, pages 2911-2918, 2012. 2 +[4] Relja Arandjelović, Petr Gronat, Akihiko Torii, Tomas Pajdla, and Josef Sivic. NetVLAD: CNN architecture for weakly supervised place recognition. IEEE TPAMI, 40(6): 1437-1451, 2018. 3, 7 +[5] Kumar Ayush, Burak Uzkent, Chenlin Meng, Kumar Tanmay, Marshall Burke, David Lobell, and Stefano Ermon. Geography-aware self-supervised learning. ICCV, 2021. 3, 7, 8 +[6] Artem Babenko, Anton Slesarev, A. Chigorin, and V. Lempitsky. Neural codes for image retrieval. ECCV, abs/1404.1777, 2014. 2 +[7] Giovanni Barbarani, Mohamad Mostafa, Hajali Bayramov, Gabriele Trivigno, Gabriele Berton, Carlo Masone, and Barbara Caputo. Are local features all you need for cross-domain visual place recognition? In CVPRW, pages 6155-6165, 2023. 3 +[8] Herbert Bay, Andreas Ess, Tinne Tuytelaars, and Luc Van Gool. Speeded-up robust features (surf). Computer Vision and Image Understanding, 110:346-359, 2008. 2 +[9] Gabriele Berton, Carlo Masone, and Barbara Caputo. Rethinking visual geo-localization for large-scale applications. In CVPR, 2022. 3, 7 +[10] Gabriele Berton, Gabriele Trivigno, Barbara Caputo, and Carlo Masone. Eigenplaces: Training viewpoint robust models for visual place recognition. In ICCV, pages 11080-11090, 2023. 7 +[11] Fatih Cakir, Kun He, Xide Xia, Brian Kulis, and Stan Sclaroff. Deep metric learning to rank. In CVPR, pages 1861-1870, 2019. 3 +[12] Yezhen Cong, Samar Khanna, Chenlin Meng, Patrick Liu, Erik Rozi, Yutong He, Marshall Burke, David B. Lobell, and Stefano Ermon. SatMAE: Pre-training transformers for temporal and multi-spectral satellite imagery. In NeurIPS, 2022. 3 +[13] Gabriela Csurka, Christopher Dance, Lixin Fan, Jutta Willamowski, and Cedric Bray. Visual categorization with bags of keypoints. In ECCV, 2004. 2 +[14] Alejandro Sánchez de Miguel, José Gómez Castaño, Jaime Zamorano, Sergio Pascual, M Ángeles, L Cayuela, Guillermo Martín Martínez, Peter Challupner, and Christopher C M Kyba. Atlas of astronaut photos of earth at night. *Astronomy and Geophysics*, 55(4):4.36-4.36, 2014. 8 +[15] Alejandro Sánchez de Miguel, Jonathan Bennie, Emma Rosenfeld, Simon Dzurjak, and Kevin J. Gaston. Environmental risks from artificial nighttime lighting widespread and increasing across Europe. Science Advances, 8(37): eabl6891, 2022. 2 + +[16] Jiankang Deng, J. Guo, and S. Zafeiriou. ArcFace: Additive angular margin loss for deep face recognition. In CVPR, pages 4685-4694, 2019. 3 +[17] Fabian Deuser, Konrad Habel, and Norbert Oswald. Sample4geo: Hard negative sampling for cross-view geolocation. In ICCV, pages 16847-16856, 2023. 3 +[18] Fabian Deuser, Konrad Habel, Martin Werner, and Norbert Oswald. Orientation-guided contrastive learning for uav-view geo-localisation. In Proceedings of the 2023 Workshop on UAVs in Multimedia: Capturing the World from a New Perspective, page 7-11, New York, NY, USA, 2023. Association for Computing Machinery. 3, 7, 8 +[19] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale. In ICLR, 2021. 7 +[20] Florian Fervers, Sebastian Bullinger, Christoph Bodensteiner, Michael Arens, and Rainer Stiefelhagen. Uncertainty-aware vision-based metric cross-view geolocation. In CVPR, pages 21621-21631, 2023. 3 +[21] Kenton Fisher, Sara Schmidt, and Alex Stoken. Crew earth observations: New tools to support your research. In 12th Annual International Space Station Research and Development Conference. Center for the Advancement of Science in Space, Inc., 2023. 2 +[22] Kevin J. Gaston and Alejandro Sánchez de Miguel. Environmental impacts of artificial light at night. Annual Review of Environment and Resources, 47(1):373-398, 2022. 2 +[23] Yixiao Ge, Haibo Wang, Feng Zhu, Rui Zhao, and Hongsheng Li. Self-supervising fine-grained region similarities for large-scale image localization. In ECCV, pages 369-386, Cham, 2020. Springer International Publishing. 7 +[24] Stephen Hausler, Sourav Garg, Ming Xu, Michael Milford, and Tobias Fischer. Patch-netvlad: Multi-scale fusion of locally-global descriptors for place recognition. In CVPR, pages 14141-14152, 2021. 3 +[25] K. He, X. Zhang, S. Ren, and J. Sun. Deep residual learning for image recognition. In CVPR, pages 770-778, 2016. 7 +[26] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In CVPR, pages 9726-9735, 2020. 7, 8 +[27] Christopher. R. Jackson and Werner Alpers. The role of the critical angle in brightness reversals on sunglint images of the sea surface. Journal of Geophysical Research: Oceans, 115(C9), 2010. 2 +[28] Augustin Juhl, Thomas Farges, and Elisabeth Blanc. Color pictures of sprites from non-dedicated observation on board the international space station. Journal of Geophysical Research: Space Physics, 118(1):454-461, 2013. 2 +[29] Hervé Jégou, Matthijs Douze, Jorge Sánchez, Patrick Perez, and Cordelia Schmid. Aggregating local image descriptors into compact codes. IEEE TPAMI, 34, 2011. 2 +[30] Nikhil Keetha, Avneesh Mishra, Jay Karhade, Krishna Murthy Jatavallabhula, Sebastian Scherer, Madhava + +Krishna, and Sourav Garg. Anyloc: Towards universal visual place recognition. IEEE Robotics and Automation Letters, 2023. 3, 7 +[31] Prannay Khosla, Piotr Teterwak, Chen Wang, Aaron Sarna, Yonglong Tian, Phillip Isola, Aaron Maschinot, Ce Liu, and Dilip Krishnan. Supervised contrastive learning. In NeurIPS, 2020. 3 +[32] Hyo Jin Kim, Enrique Dunn, and Jan-Michael Frahm. Learned contextual feature reweighting for image geolocation. In CVPR, pages 3251-3260, 2017. 3 +[33] Diederik Kingma and Jimmy Ba. Adam: A method for stochastic optimization. *ICLR*, 2014. 7 +[34] María Leyva-Vallina, Nicola Strisciuglio, and Nicolai Petkov. Data-efficient large scale place recognition with graded similarity supervision. CVPR, 2023. 3 +[35] Weiyang Liu, Yandong Wen, Zhiding Yu, and Meng Yang. Large-margin softmax loss for convolutional neural networks. In ICML, pages 507-516. PMLR, 2016. 3 +[36] Weiyang Liu, Yandong Wen, Zhiding Yu, Ming Li, Bhiksha Raj, and Le Song. Spherface: Deep hypersphere embedding for face recognition. In CVPR, pages 212-220, 2017. 3 +[37] David G. Lowe. Distinctive image features from scale-invariant keypoints. Int. J. Comput. Vision, 60(2):91-110, 2004. 2 +[38] Carlo Masone and Barbara Caputo. A survey on deep visual place recognition. IEEE Access, 9:19516-19547, 2021. 3 +[39] Oscar Manas, Alexandre Lacoste, Xavier Giroi-nieto, David Vazquez, and Pau Rodriguez. Seasonal contrast: Unsupervised pre-training from uncurated remote sensing data. In ICCV, pages 9394-9403, 2021. 3 +[40] Aadil Nathani, Rishi Iyer, Annabelle Wang, Aarnav Chitari, Adele Wilson, and Hannah Norris. Observing Earth From Space: Using Astronaut Photography to Analyze Geographical Climate Patterns. In AGU Fall Meeting Abstracts, pages ED44C-06, 2022. 2 +[41] Maxime Oquab, Timothée Darcet, Theo Moutakanni, Huy V. Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, Russell Howes, Po-Yao Huang, Hu Xu, Vasu Sharma, Shang-Wen Li, Wojciech Galuba, Mike Rabbat, Mido Assran, Nicolas Ballas, Gabriel Synnaeve, Ishan Misra, Herve Jegou, Julien Mairal, Patrick Labatut, Armand Joulin, and Piotr Bojanowski. Dinov2: Learning robust visual features without supervision, 2023. 7 +[42] Florent Perronnin, Yan Liu, Jorge Sánchez, and Herve Poirier. Large-scale image retrieval with compressed fisher vectors. In CVPR, pages 3384-3391, 2010. 2 +[43] F. Radenović, G. Tolias, and O. Chum. Fine-tuning CNN Image Retrieval with No Human Annotation. IEEE TPAMI, 2018. 3 +[44] A. Razavian, J. Sullivan, A. Maki, and S. Carlsson. Visual Instance Retrieval with Deep Convolutional Networks. CoRR, abs/1412.6574, 2015. 3 +[45] Esther Rolf, Jonathan Proctor, Tamma Carleton, Ian Bolliger, Vaishaal Shankar, Miyabi Ishihara, Benjamin Recht, and Solomon Hsiang. A generalizable and accessible approach to machine learning with global satellite imagery. Nature Communications, 12(1), 2021. 3 + +[46] Paul-Edouard Sarlin, Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. SuperGlue: Learning feature matching with graph neural networks. In CVPR, 2020. 3 +[47] Aaron Schirmer, Caleb Gallemore, Ting Liu, Seth Magle, Elisabeth DiNello, Humerah Ahmed, and Thomas Gilday. Mapping behaviorally relevant light pollution levels to improve urban habitat planning. Scientific Reports, 9, 2019. 2 +[48] Johannes A. Schultz, Maik Hartmann, Sascha Heinemann, Jens Janke, Carsten Jürgens, Dieter Oertel, Gernot Rücker, Frank Thonfeld, and Andreas Rienow. Diego: A multispectral thermal mission for earth observation on the international space station. European Journal of Remote Sensing, 53(sup2):28-38, 2020. 2 +[49] Peter Schwind and Tobias Storch. Georeferencing urban nighttime lights imagery using street network maps. Remote Sensing, 14(11), 2022. 3 +[50] Yujiao Shi and Hongdong Li. Beyond cross-view image retrieval: Highly accurate vehicle localization using satellite image. In CVPR, pages 17010-17020, 2022. 3 +[51] Yujiao Shi, Xin Yu, Shan Wang, and Hongdong Li. Cvlnet: Cross-view feature correspondence learning for video-based camera localization. In ACCV, pages 652-669, 2022. +[52] Yujiao Shi, Fei Wu, Akhil Perincherry, Ankit Vora, and Hongdong Li. Boosting 3-dof ground-to-satellite camera localization accuracy via geometry-guided cross-view transformer. In ICCV, pages 21516-21526, 2023. 3 +[53] Christopher Small. Spectrometry of the urban lightscape. Technologies, 10(4), 2022. 2 +[54] H. Song, Y. Xiang, S. Jegelka, and S. Savarese. Deep metric learning via lifted structured feature embedding. In CVPR, 2016. Spotlight presentation. 3 +[55] Weiwei Song, Zhi Gao, Renwei Dian, Pedram Ghamisi, Yongjun Zhang, and Jón Atli Benediktsson. Asymmetric hash code learning for remote sensing image retrieval. IEEE Transactions on Geoscience and Remote Sensing, 60:1-14, 2022. 3 +[56] W. L. Stefanov and C. A. Evans. Data collection for disaster response from the international space station. The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, XL-7/W3:851–855, 2015. 2 +[57] Adam J. Stewart, Caleb Robinson, Isaac A. Corley, Anthony Ortiz, Juan M. Lavista Ferres, and Arindam Banerjee. Torch-Geo: Deep learning with geospatial data. In Proceedings of the 30th International Conference on Advances in Geographic Information Systems, pages 1–12, Seattle, Washington, 2022. Association for Computing Machinery. 7, 8 +[58] Alex Stoken and Kenton Fisher. Find my astronaut photo: Automated localization and georectification of astronaut photography. In CVPRW, pages 6196-6205, 2023. 2, 3, 4, 5, 7 +[59] Jiaming Sun, Zehong Shen, Yuang Wang, Hujun Bao, and Xiaowei Zhou. LoFTR: Detector-free local feature matching with transformers. CVPR, 2021. 3 +[60] Yifan Sun, Changmao Cheng, Yuhan Zhang, Chi Zhang, Liang Zheng, Zhongdao Wang, and Yichen Wei. Circle loss: A unified perspective of pair similarity optimization. CVPR, pages 6397-6406, 2020. 3 + +[61] Giorgos Tolias, R. Sicre, and H. Jégou. Particular object retrieval with integral max-pooling of CNN activations. CoRR, abs/1511.05879, 2016. 3 +[62] Ilya O Tolstikhin, Neil Houlsby, Alexander Kolesnikov, Lucas Beyer, Xiaohua Zhai, Thomas Unterthiner, Jessica Yung, Andreas Steiner, Daniel Keysers, Jakob Uszkoreit, Mario Lucic, and Alexey Dosovitskiy. Mlp-mixer: An all-mlx architecture for vision. In NeurIPS, pages 24261–24272. Curran Associates, Inc., 2021. 7 +[63] Gabriele Trivigno, Gabriele Berton, Juan Aragon, Barbara Caputo, and Carlo Masone. Divide&classify: Fine-grained classification for city-wide visual geo-localization. In ICCV, pages 11142-11152, 2023. 3 +[64] Aäron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding. CoRR, abs/1807.03748, 2018. 3 +[65] Hao Wang, Yitong Wang, Zheng Zhou, Xing Ji, Dihong Gong, Jingchao Zhou, Zhifeng Li, and Wei Liu. Cosface: Large margin cosine loss for deep face recognition. In CVPR, pages 5265-5274. Computer Vision Foundation / IEEE Computer Society, 2018. 3 +[66] Tingyu Wang, Zhedong Zheng, Chenggang Yan, jiyong Zhang, Yaoqi Sun, Bolun Zheng, and Yi Yang. Each part matters: Local patterns facilitate cross-view geo-localization. IEEE TCSVT, 2021. doi: 10.1109/TCSVT.2021.3061265.3 +[67] Xun Wang, Xintong Han, Weilin Huang, Dengke Dong, and Matthew R Scott. Multi-similarity loss with general pair weighting for deep metric learning. In CVPR, pages 5022-5030, 2019. 3, 5 +[68] M. Justin Wilkinson and Yanni Gunnell. Fluvial Megafans on Earth and Mars. Cambridge University Press, 2023. 2 +[69] Yoav Yair, Melody Korman, Colin Price, and Eytan Stibbe. Observing lightning and transient luminous events from the international space station during ilan-es: An astronaut's perspective. Acta Astronautica, 211:592-599, 2023. 2 +[70] Ting Yao, Yiheng Zhang, Zhaofan Qiu, Yingwei Pan, and Tao Mei. Seco: Exploring sequence supervision for unsupervised representation learning. In AAAI, 2021. 7, 8 +[71] Zhedong Zheng, Yunchao Wei, and Yi Yang. University-1652: A multi-view multi-source benchmark for drone-based geo-localization. ACM MM, 2020. 3 +[72] Zhedong Zheng, Yu jiao Shi, Tingyu Wang, Jun Liu, Jianwu Fang, Yunchao Wei, and Tat-seng Chua. Uavm '23: 2023 workshop on uavs in multimedia: Capturing the world from a new perspective. In ACM MM, page 9715-9717, New York, NY, USA, 2023. Association for Computing Machinery. 8 +[73] Runzhe Zhu, Mingze Yang, Kaiyu Zhang, Fei Wu, Ling Yin, and Yujin Zhang. Modern backbone for efficient geolocation. In Proceedings of the 2023 Workshop on UAVs in Multimedia: Capturing the World from a New Perspective, page 31-37, New York, NY, USA, 2023. Association for Computing Machinery. 3, 7, 8 +[74] Sijie Zhu, Taojiannan Yang, and Chen Chen. Vigor: Cross-view image geo-localization beyond one-to-one retrieval. In CVPR, pages 3640-3649, 2021. 3 + +[75] Sijie Zhu, Mubarak Shah, and Chen Chen. Transgeo: Transformer is all you need for cross-view image geo-localization. In CVPR, pages 1162-1171, 2022. 3 +[76] Sijie Zhu, Linjie Yang, Chen Chen, Mubarak Shah, Xiaohui Shen, and Heng Wang. R2former: Unified retrieval and reranking transformer for place recognition. In CVPR, pages 19370-19380, 2023. 3 \ No newline at end of file diff --git a/2024/EarthLoc_ Astronaut Photography Localization by Indexing Earth from Space/images.zip b/2024/EarthLoc_ Astronaut Photography Localization by Indexing Earth from Space/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..5902947afc5421e53c43342eb2ac77ea83338f58 --- /dev/null +++ b/2024/EarthLoc_ Astronaut Photography Localization by Indexing Earth from Space/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e1922a13463c4292f2b1fa40a31120c5693a79f64774863632535df4a3d3dfc8 +size 459311 diff --git a/2024/EarthLoc_ Astronaut Photography Localization by Indexing Earth from Space/layout.json b/2024/EarthLoc_ Astronaut Photography Localization by Indexing Earth from Space/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..96d77981a43ce80cb2176d3667b6b65c9d9a5182 --- /dev/null +++ b/2024/EarthLoc_ Astronaut Photography Localization by Indexing Earth from Space/layout.json @@ -0,0 +1,9071 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 54, + 103, + 539, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 103, + 539, + 121 + ], + "spans": [ + { + "bbox": [ + 54, + 103, + 539, + 121 + ], + "type": "text", + "content": "EarthLoc: Astronaut Photography Localization by Indexing Earth from Space" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 120, + 142, + 473, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 120, + 142, + 473, + 185 + ], + "spans": [ + { + "bbox": [ + 120, + 142, + 473, + 185 + ], + "type": "text", + "content": "Gabriele Berton1 Alex Stoken2 Barbara Caputo1 Carlo Masone1 \n1Politecnico di Torino 2Jacobs Technology, NASA Johnson Space Center \ngabriele.berton@polito.it" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "spans": [ + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 238, + 290, + 467 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 238, + 290, + 467 + ], + "spans": [ + { + "bbox": [ + 46, + 238, + 290, + 467 + ], + "type": "text", + "content": "Astronaut photography, spanning six decades of human spaceflight, presents a unique Earth observations dataset with immense value for both scientific research and disaster response. Despite their significance, accurately localizing the geographical extent of these images, which is crucial for effective utilization, poses substantial challenges. Current, manual localization efforts are time-consuming, motivating the need for automated solutions. We propose a novel approach - leveraging image retrieval - to address this challenge efficiently. We introduce innovative training techniques which contribute to the development of a high-performance model, EarthLoc. We develop six evaluation datasets and perform a comprehensive benchmark comparing EarthLoc to existing methods, showcasing its superior efficiency and accuracy. Our approach marks a significant advancement in automating the localization of astronaut photography, which will help bridge a critical gap in Earth observations data. Code and datasets are available at https://github.com/gmberton/EarthLoc." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 490, + 128, + 502 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 490, + 128, + 502 + ], + "spans": [ + { + "bbox": [ + 47, + 490, + 128, + 502 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 510, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 510, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 510, + 289, + 715 + ], + "type": "text", + "content": "Astronaut photography of Earth is a unique remote sensing dataset that spans 60 years of human spaceflight, offering a rare perspective on our planet to the public and valuable data to Earth and atmospheric science researchers. This dataset contains over 4.5 million images and is growing by the tens of thousands per month, as astronauts are continually tasked with taking new photographs that enable scientific research as well as assist in natural disaster response efforts in the wake of events like floods and wildfires. To effectively use these images, the geographical area depicted in them needs to be identified. Unfortunately, this task - Astronaut Photography Localization (APL) - is very challenging. For each photo, only a coarse estimate of location is known, given by the point on Earth directly under the International Space Station (ISS) at the time the photo is taken. This point - called the nadir - can be easily computed using the image's timestamp and the ISS's orbit path. However, two" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 335, + 212, + 518, + 427 + ], + "blocks": [ + { + "bbox": [ + 335, + 212, + 518, + 427 + ], + "lines": [ + { + "bbox": [ + 335, + 212, + 518, + 427 + ], + "spans": [ + { + "bbox": [ + 335, + 212, + 518, + 427 + ], + "type": "image", + "image_path": "306c03fd150f2bcdbeb0615054bd2cd5358be753de30f69f07101315d886e142.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 437, + 547, + 581 + ], + "lines": [ + { + "bbox": [ + 304, + 437, + 547, + 581 + ], + "spans": [ + { + "bbox": [ + 304, + 437, + 547, + 581 + ], + "type": "text", + "content": "Figure 1. Overview of the astronaut photography localization task. Astronauts take hundreds of photos a day from the International Space Station (ISS) cupola (top-left) with hand-held cameras. For each image (example bottom right), the geographic location depicted is not known, and needs to be searched for across a huge area centered at the ISS's (known) nadir point at the moment that the photo is taken. A simulated view of the astronaut's perspective when the ISS is above Europe is shown. The goal of our paper is to automate the task of localizing these images, which could be anywhere within the view. In the figure's example, the photo the astronaut took is indicated by the green line and shown in inset - other possible photo extents are in red, illustrating the wide array of potential locations to search." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "type": "text", + "content": "images taken with the same nadir can be thousands of kilometers apart, as even a slight inclination of the astronaut's hand-held camera can move the image's location hundreds of kilometers in any direction, as depicted in Fig. 1. Localization must thus be performed over a wide area, and is additionally complicated by (i) astronauts using hand-held cameras and a variety of zoom lenses, (ii) the large, 2500 kilometer (km) visibility range in all directions, (iii) most photographs being oblique, and (iv) the Earth's appearance" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "12754" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 289, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 289, + 335 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 289, + 335 + ], + "type": "text", + "content": "changing over time, mostly due to weather or seasonal variations. Despite these challenges, the high value of properly localized imagery for Earth science research and disaster response has lead to more than 300,000 [21] astronaut photos being manually localized by identifying the geographic coordinates of the photo's center pixel. This process can take minutes to hours per photo for experts in NASA's Earth Science and Remote Sensing Unit (ESRS) and citizen scientists through ESRS's Image Detective program. The resulting geo-located images have lead to a multitude of peer reviewed papers, technical reports, and articles about light pollution and urban planning [15, 22, 47, 53], atmospheric phenomena [28, 69], changes in land usage and glacial extent [40], and other Earth science topics [27, 48, 68]. Most importantly, astronaut photography has a fast response time, crucial for disaster management - astronauts are alerted of a disaster, provide photos to a ground crew which localizes them, and sends them to first responders [56]. For example, in 2013 this protocol was activated in response to Cyclone Haiyan in the Philippines, wildfires in Australia, flooding in China, Russia, Pakistan, and the USA, and multiple other events. [56]." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 335, + 289, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 335, + 289, + 502 + ], + "spans": [ + { + "bbox": [ + 47, + 335, + 289, + 502 + ], + "type": "text", + "content": "Even with the abundance and importance of satellite imagery in modern applications, astronaut photography fills an unserved gap among other remotely sensed data. Unlike satellites that nominally take top-down imagery in similar illumination conditions at fixed temporal intervals, astronaut photographs can show topography via oblique views from multiple orientations, are taken in various lighting conditions (including nighttime), and vary in focal length to show detail at different resolutions. These qualities produce a complementary data product that would be difficult or impossible to gather from traditional satellites. Additionally, having a human in the loop for data collection allows for real-time response to natural disasters as well as a natural sense to avoid clouds and other obstructions." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 502, + 289, + 659 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 502, + 289, + 659 + ], + "spans": [ + { + "bbox": [ + 47, + 502, + 289, + 659 + ], + "type": "text", + "content": "Given the uniqueness and importance of astronaut photography, and the large amount of time spent by human experts on geo-localizing them, researchers have been exploring solutions to automate the task through computer vision methods. Previous works have shown promising results using a pairwise matching setup that iteratively compares an astronaut photo to satellite imagery [58]. Yet, such methods suffer from high latency, searching multiple directions sequentially and using compute-heavy, dense correspondences to determine matching. The time required to find the location of all 4.5 million images using these methods is estimated to be over 20 compute years [58]. Furthermore, low latency is of paramount importance for pho" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 307, + 71, + 548, + 106 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 71, + 548, + 106 + ], + "spans": [ + { + "bbox": [ + 307, + 71, + 548, + 106 + ], + "type": "text", + "content": "tography in areas that are affected by natural disasters (e.g., hurricanes or wildfires), for which speedy localization can help real-time, on-the-ground operations." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 308, + 108, + 547, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 108, + 547, + 240 + ], + "spans": [ + { + "bbox": [ + 308, + 108, + 547, + 240 + ], + "type": "text", + "content": "To overcome the high latency of current methods, we propose to instead localize astronaut photography through image retrieval, by matching each astronaut photograph to a worldwide database containing satellite images of known position. In reformulating this problem as a retrieval task, we encounter a range of new challenges, from efficiently training a robust model to achieving low-latency inference, as well as defining a success metric and creating evaluation sets, both of which are crucial pieces when assessing how well a given model will perform when deployed to localize astronaut photography." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 308, + 240, + 547, + 359 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 240, + 547, + 359 + ], + "spans": [ + { + "bbox": [ + 308, + 240, + 547, + 359 + ], + "type": "text", + "content": "The ultimate goal of this paper is twofold: (i) provide an efficient method to localize the archive of 4.5M queries as well as all new, incoming imagery, so that researchers studying a given geographical area can have access to a large amount of imagery spanning many years; and (ii) prove the viability of world-wide APL through image retrieval (which has never been attempted and was believed to be infeasible4) in order to spark a new line of research, with direct benefit to all space-based photography and its users across the globe." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 308, + 360, + 545, + 382 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 360, + 545, + 382 + ], + "spans": [ + { + "bbox": [ + 308, + 360, + 545, + 382 + ], + "type": "text", + "content": "To achieve its goal, this paper makes the following contributions:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 384, + 545, + 490 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 306, + 384, + 535, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 384, + 535, + 396 + ], + "spans": [ + { + "bbox": [ + 306, + 384, + 535, + 396 + ], + "type": "text", + "content": "- we propose approaching APL through image retrieval;" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 396, + 545, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 396, + 545, + 418 + ], + "spans": [ + { + "bbox": [ + 306, + 396, + 545, + 418 + ], + "type": "text", + "content": "- we develop novel training techniques which show large quantitative improvement on the task;" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 419, + 545, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 419, + 545, + 443 + ], + "spans": [ + { + "bbox": [ + 306, + 419, + 545, + 443 + ], + "type": "text", + "content": "- we provide six evaluation sets and a large benchmark of results with methods from a variety of relevant domains;" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 444, + 545, + 490 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 444, + 545, + 490 + ], + "spans": [ + { + "bbox": [ + 306, + 444, + 545, + 490 + ], + "type": "text", + "content": "- we show that a model trained with these new techniques, EarthLoc, allows us to localize a large number of images, widely outperforming all other methods while being fast and efficient." + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 491, + 545, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 491, + 545, + 540 + ], + "spans": [ + { + "bbox": [ + 306, + 491, + 545, + 540 + ], + "type": "text", + "content": "Finally, we note how EarthLoc is being used to localize ISS astronaut photographs, which are publicly available and conveniently searchable at https://eol.jsc.nasa.gov/ExplorePhotos/." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 309, + 548, + 395, + 559 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 309, + 548, + 395, + 559 + ], + "spans": [ + { + "bbox": [ + 309, + 548, + 395, + 559 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 571, + 547, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 571, + 547, + 667 + ], + "spans": [ + { + "bbox": [ + 306, + 571, + 547, + 667 + ], + "type": "text", + "content": "Image Retrieval Image retrieval involves searching a database for images similar to a query image. Traditionally, methodologies combined hand-crafted local features like SIFT [37], SURF [8], RootSIFT [3] into a global embedding through means like Bag of Words [13], Fisher Vectors [42] or VLAD [29], to allow for fast retrieval through kNN. With the rise of deep learning, local features have been replaced by CNN-derived features [6], significantly" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 306, + 673, + 547, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 673, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 673, + 547, + 713 + ], + "type": "text", + "content": "4 in [59] (Sec. 2.1) scientists at NASA claim that retrieval for APL is infeasible because we cannot precompute a database of reference features due to their belief that the database must be query-specific. Results show that our pipeline can overcome this issue." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 45, + 662, + 285, + 681 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 662, + 285, + 681 + ], + "spans": [ + { + "bbox": [ + 45, + 662, + 285, + 681 + ], + "type": "text", + "content": "1https://eol.jsc.nasa.gov/BeyondThePhotography/ ImageDetective/" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 58, + 683, + 280, + 693 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 683, + 280, + 693 + ], + "spans": [ + { + "bbox": [ + 58, + 683, + 280, + 693 + ], + "type": "text", + "content": "2https://eol.jsc.nasa.gov/AboutCEO/PubList.htm" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 49, + 694, + 286, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 694, + 286, + 712 + ], + "spans": [ + { + "bbox": [ + 49, + 694, + 286, + 712 + ], + "type": "text", + "content": "3https : / / storymaps . arcgis . com / stories / 947eb734e811465cb0425947b16b62b3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "12755" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 264 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 264 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 264 + ], + "type": "text", + "content": "improving retrieval performance. Since then, most of the attention of the retrieval community has been focused on how to best aggregate feature maps and efficiently train robust neural networks. For feature map aggregation, some of the most notable examples in literature include Max pooling [44], Regional Max pooling (R-MAC) [61] and Generalized Mean (GeM) [43]. To train these retrieval models, a number of losses have been proposed, which can be generally grouped into two categories: stateful losses, which have weights whose size depends on the number of classes, like the Large-Margin Softmax [35], SphereFace [36], CosFace [65] and ArcFace [16]; and their stateless counterparts, like the Contrastive and Triplet losses, and the more recent Lifted Structure Loss [54], NTXentLoss [64], Multi-Similarity Loss[67], FastAP [11], Supervised Contrastive [31] and the Circle Loss [60]." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 281, + 287, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 281, + 287, + 449 + ], + "spans": [ + { + "bbox": [ + 46, + 281, + 287, + 449 + ], + "type": "text", + "content": "Geo-localization and retrieval with Aerial and Remote Sensing Imagery To the best of our knowledge, only two previous works have specifically addressed the localization problem for astronaut photography. Both rely on local features and compute pairwise comparisons between astronaut photography and geo-located reference imagery. Find My Astronaut Photo [58] focuses on daytime imagery, presenting an evaluation of multiple methods as well as the note that pre-trained self-supervised models were not suitable for a retrieval-based approach to this task. Schwind and Storch [49] instead address nighttime imagery, using synthetically generated street light maps as reference. Both works discuss the importance of proper orientation for matching and the high cost of pairwise comparisons." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 449, + 289, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 449, + 289, + 593 + ], + "spans": [ + { + "bbox": [ + 46, + 449, + 289, + 593 + ], + "type": "text", + "content": "More commonly, studies of aerial imagery focus on data from unmanned aerial vehicles (UAVs) and satellite platforms. Aerial geo-localization is an important aspect of UAV navigation systems. The University-1652 dataset [71] and associated challenge encourage work in cross-view geo-localization for drone-satellite imagery, with strong approaches modifying different parts of the general pipeline, from backbone [73] to training setup [18], to feature partitioning [66]. Most use contrastive and self-supervised losses in training. Other works address cross-view geo-localization in more extreme view point differences, like that between aerial and street views [17, 20, 50-52, 74, 75]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 594, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 289, + 715 + ], + "type": "text", + "content": "Remotely sensed data from satellites, on the other hand, often comes with reliable geo-location information from the sensor itself. Extracting features from satellite imagery is still a common task, with such features most often trained on a pretext objective before use on downstream tasks [5, 12, 39, 45]. Occasionally, features are used for retrieval itself [55], to identify areas of the Earth that are similar in appearance and thus might have similar properties to be studied together. In this paper we take a separate direction from previous literature, and aim at using image" + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 326, + 71, + 392, + 115 + ], + "blocks": [ + { + "bbox": [ + 326, + 71, + 392, + 115 + ], + "lines": [ + { + "bbox": [ + 326, + 71, + 392, + 115 + ], + "spans": [ + { + "bbox": [ + 326, + 71, + 392, + 115 + ], + "type": "image", + "image_path": "2f0079bf20b434a2931132166f9337d4bee16a0d1eb527d4a2af3b53155c734b.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 393, + 71, + 459, + 115 + ], + "blocks": [ + { + "bbox": [ + 393, + 71, + 459, + 115 + ], + "lines": [ + { + "bbox": [ + 393, + 71, + 459, + 115 + ], + "spans": [ + { + "bbox": [ + 393, + 71, + 459, + 115 + ], + "type": "image", + "image_path": "fbf8302a98adbf5a1b2ba44d37a91e9fb5f5d1fb519f46a2036e34c5e5761641.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 460, + 71, + 526, + 115 + ], + "blocks": [ + { + "bbox": [ + 460, + 71, + 526, + 115 + ], + "lines": [ + { + "bbox": [ + 460, + 71, + 526, + 115 + ], + "spans": [ + { + "bbox": [ + 460, + 71, + 526, + 115 + ], + "type": "image", + "image_path": "db1eaf7a3f1e24fedaeff0859841a2163133c83faa378cdba6457089d880f879.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 326, + 116, + 392, + 159 + ], + "blocks": [ + { + "bbox": [ + 326, + 116, + 392, + 159 + ], + "lines": [ + { + "bbox": [ + 326, + 116, + 392, + 159 + ], + "spans": [ + { + "bbox": [ + 326, + 116, + 392, + 159 + ], + "type": "image", + "image_path": "c8f695f11d7b80967238d9fc90f867b4115cc3343d2cd411d2b4c44a0014c0cb.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 170, + 545, + 192 + ], + "lines": [ + { + "bbox": [ + 305, + 170, + 545, + 192 + ], + "spans": [ + { + "bbox": [ + 305, + 170, + 545, + 192 + ], + "type": "text", + "content": "Figure 2. Astronaut photo query examples, showcasing the large variability in covered area and appearance." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 393, + 116, + 459, + 159 + ], + "blocks": [ + { + "bbox": [ + 393, + 116, + 459, + 159 + ], + "lines": [ + { + "bbox": [ + 393, + 116, + 459, + 159 + ], + "spans": [ + { + "bbox": [ + 393, + 116, + 459, + 159 + ], + "type": "image", + "image_path": "4b5b11011707b7e985bdae857a14c2f35e6ee8128514cba300809aeabcdf2064.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 460, + 116, + 526, + 159 + ], + "blocks": [ + { + "bbox": [ + 460, + 116, + 526, + 159 + ], + "lines": [ + { + "bbox": [ + 460, + 116, + 526, + 159 + ], + "spans": [ + { + "bbox": [ + 460, + 116, + 526, + 159 + ], + "type": "image", + "image_path": "0c5720bcdb102fd8a201f9a8149d19b734e9a4371b9eb9217cfa5434dac4a455.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 195, + 425, + 207 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 195, + 425, + 207 + ], + "spans": [ + { + "bbox": [ + 306, + 195, + 425, + 207 + ], + "type": "text", + "content": "retrieval techniques for APL." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 221, + 547, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 221, + 547, + 437 + ], + "spans": [ + { + "bbox": [ + 304, + 221, + 547, + 437 + ], + "type": "text", + "content": "Other localization tasks Various other localization challenges are approached through image retrieval, including visual place recognition (VPR) and visual localization. The former aims at coarsely localizing a given query by matching it to a database of geo-tagged photos, and is mostly studied within urban environments using street-view imagery [7, 63]. Performance is improved either by the use of smart aggregation (e.g. NetVLAD [4]), attention layers (e.g. CRN [32]), or through large-scale training [1, 2, 9, 34]. More recently, the task of universal visual place recognition has been proposed by AnyLoc [30], whose authors provide a model that performs competitively on a large range of scenarios, including aerial imagery. The task of visual localization is focused on finding the precise camera pose of an image, and is commonly approached through feature matching techniques [7, 38, 46, 59]. It can also be used for visual place recognition [24, 76], although this leads to a noticeable increase in runtime when compared to pure retrieval." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 447, + 360, + 459 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 447, + 360, + 459 + ], + "spans": [ + { + "bbox": [ + 306, + 447, + 360, + 459 + ], + "type": "text", + "content": "3. Dataset" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 467, + 545, + 575 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 467, + 545, + 575 + ], + "spans": [ + { + "bbox": [ + 304, + 467, + 545, + 575 + ], + "type": "text", + "content": "In this section we discuss the datasets that we use in this work. Specifically, in Sec. 3.1 we introduce the images that we need to localize, called queries; in Sec. 3.2 we outline the collection of the database, made of satellite images (i.e., taken automatically, not with a hand-held camera) of known location; and in Sec. 3.3 we describe the creation of the evaluation datasets that we use to understand the capabilities of our models, combining queries and database in a way that will reflect the real-world use case of our method." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 582, + 365, + 594 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 582, + 365, + 594 + ], + "spans": [ + { + "bbox": [ + 306, + 582, + 365, + 594 + ], + "type": "text", + "content": "3.1. Queries" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 600, + 545, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 600, + 545, + 696 + ], + "spans": [ + { + "bbox": [ + 304, + 600, + 545, + 696 + ], + "type": "text", + "content": "Queries are taken from the Gateway to Astronaut Photography of Earth " + }, + { + "bbox": [ + 304, + 600, + 545, + 696 + ], + "type": "inline_equation", + "content": "^{5}" + }, + { + "bbox": [ + 304, + 600, + 545, + 696 + ], + "type": "text", + "content": ", a collection of over 4.5 million photos of Earth taken by astronauts on the International Space Station (Fig. 2). This unique photography setting yields only coarse location information for each photograph, namely the position on Earth below the ISS (the nadir point) at the time the photo was taken. We seek to find the geographic area on the Earth that each photo encompasses (the photo's location)," + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 315, + 702, + 443, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 702, + 443, + 713 + ], + "spans": [ + { + "bbox": [ + 315, + 702, + 443, + 713 + ], + "type": "text", + "content": "5https://eol.jsc.nasa.gov/" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "12756" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 67, + 70, + 116, + 121 + ], + "blocks": [ + { + "bbox": [ + 67, + 70, + 116, + 121 + ], + "lines": [ + { + "bbox": [ + 67, + 70, + 116, + 121 + ], + "spans": [ + { + "bbox": [ + 67, + 70, + 116, + 121 + ], + "type": "image", + "image_path": "0261613d2655e91a2e02144006f754a2b00be1f2ff93651d9d66a20711547ebe.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 118, + 70, + 167, + 120 + ], + "blocks": [ + { + "bbox": [ + 118, + 70, + 167, + 120 + ], + "lines": [ + { + "bbox": [ + 118, + 70, + 167, + 120 + ], + "spans": [ + { + "bbox": [ + 118, + 70, + 167, + 120 + ], + "type": "image", + "image_path": "4c1e94bbae9a8fd5ea35601da20353a5b45b2b86c83a6365111f7468fdc03865.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 168, + 70, + 218, + 120 + ], + "blocks": [ + { + "bbox": [ + 168, + 70, + 218, + 120 + ], + "lines": [ + { + "bbox": [ + 168, + 70, + 218, + 120 + ], + "spans": [ + { + "bbox": [ + 168, + 70, + 218, + 120 + ], + "type": "image", + "image_path": "2a76d37bed00f53ccbd74b9b783cdc18a4566330491d3f21b287de192959200a.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 219, + 70, + 268, + 120 + ], + "blocks": [ + { + "bbox": [ + 219, + 70, + 268, + 120 + ], + "lines": [ + { + "bbox": [ + 219, + 70, + 268, + 120 + ], + "spans": [ + { + "bbox": [ + 219, + 70, + 268, + 120 + ], + "type": "image", + "image_path": "84371a693e602f29fc8e8ff84d60f67964dee33653ca998e538f56a1083a61f5.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 67, + 124, + 116, + 175 + ], + "blocks": [ + { + "bbox": [ + 67, + 124, + 116, + 175 + ], + "lines": [ + { + "bbox": [ + 67, + 124, + 116, + 175 + ], + "spans": [ + { + "bbox": [ + 67, + 124, + 116, + 175 + ], + "type": "image", + "image_path": "99e2ceef17f8a5c6c099c9214ca03d72799641480fc5f27dc4f678d47ae0317c.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 184, + 288, + 228 + ], + "lines": [ + { + "bbox": [ + 46, + 184, + 288, + 228 + ], + "spans": [ + { + "bbox": [ + 46, + 184, + 288, + 228 + ], + "type": "text", + "content": "Figure 3. Examples of database images. Top: four images from different regions, with " + }, + { + "bbox": [ + 46, + 184, + 288, + 228 + ], + "type": "inline_equation", + "content": "25\\%" + }, + { + "bbox": [ + 46, + 184, + 288, + 228 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 46, + 184, + 288, + 228 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 46, + 184, + 288, + 228 + ], + "type": "text", + "content": " overlap between any pair. The red dot in each image represents the same geographic point. Bottom: 4 images from the same region across different years." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 118, + 124, + 167, + 175 + ], + "blocks": [ + { + "bbox": [ + 118, + 124, + 167, + 175 + ], + "lines": [ + { + "bbox": [ + 118, + 124, + 167, + 175 + ], + "spans": [ + { + "bbox": [ + 118, + 124, + 167, + 175 + ], + "type": "image", + "image_path": "e105281a6895053b103e5a7466d139bb4db6b2e4fd30ea96d5a80ca737dcffef.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 168, + 124, + 218, + 175 + ], + "blocks": [ + { + "bbox": [ + 168, + 124, + 218, + 175 + ], + "lines": [ + { + "bbox": [ + 168, + 124, + 218, + 175 + ], + "spans": [ + { + "bbox": [ + 168, + 124, + 218, + 175 + ], + "type": "image", + "image_path": "42f578694e46a3b97f73fc8651a7343093ec202fe314dbebf8425bad1d8022dd.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 219, + 124, + 268, + 175 + ], + "blocks": [ + { + "bbox": [ + 219, + 124, + 268, + 175 + ], + "lines": [ + { + "bbox": [ + 219, + 124, + 268, + 175 + ], + "spans": [ + { + "bbox": [ + 219, + 124, + 268, + 175 + ], + "type": "image", + "image_path": "1a5387552fd0583aacc6454dfae4af460e6e66d7915cef119a9c6b37c9c1c881.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 240, + 287, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 240, + 287, + 277 + ], + "spans": [ + { + "bbox": [ + 46, + 240, + 287, + 277 + ], + "type": "text", + "content": "which can be thousands of kilometers away from the ISS nadir point. This distance of maximum visibility (horizon) can be quickly computed as" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 96, + 285, + 287, + 300 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 96, + 285, + 287, + 300 + ], + "spans": [ + { + "bbox": [ + 96, + 285, + 287, + 300 + ], + "type": "interline_equation", + "content": "d _ {\\text {v i s i b l e}} = \\sqrt {2 R h + h ^ {2}} = 2 4 3 6 \\mathrm {k m} \\tag {1}", + "image_path": "97587940d3c34b6c5d04b1b1c5ad4db9b90cc89bdcea35ea940363560c5dcbee.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 308, + 287, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 308, + 287, + 346 + ], + "spans": [ + { + "bbox": [ + 46, + 308, + 287, + 346 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 308, + 287, + 346 + ], + "type": "inline_equation", + "content": "R = 6317" + }, + { + "bbox": [ + 46, + 308, + 287, + 346 + ], + "type": "text", + "content": " is Earth's radius and " + }, + { + "bbox": [ + 46, + 308, + 287, + 346 + ], + "type": "inline_equation", + "content": "h = 450" + }, + { + "bbox": [ + 46, + 308, + 287, + 346 + ], + "type": "text", + "content": " is the ISS maximum orbiting altitude. We follow common practice and round this number to 2500 kilometers [58]." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 359, + 287, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 359, + 287, + 491 + ], + "spans": [ + { + "bbox": [ + 46, + 359, + 287, + 491 + ], + "type": "text", + "content": "Challenges to APL Some astronaut photo queries, taken with high focal length lenses, can cover only a few kilometers within the almost 20 million sq. km area visible to the astronaut taking the photo, which makes localizing the images extremely challenging. Additional hurdles presented by the imagery acquisition process include the varying quality of the images themselves due to motion blur from the fast moving space station, and occlusion of unique, location-identifying features by cloud cover, shadow, or other portions of the ISS (Fig. 2). More queries are shown in Sec. 10 of the Supplementary." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 46, + 491, + 287, + 563 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 491, + 287, + 563 + ], + "spans": [ + { + "bbox": [ + 46, + 491, + 287, + 563 + ], + "type": "text", + "content": "From the 4.5 million photos available, we consider only the day-time images that have a ground truth obtained by FMAP [58], and keep only those which cover areas between 5000 and 900,000 sq. km, in order to match the areas of database images (see Sec. 3.2). This leads to 17,764 queries which we can use for validation and testing." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 47, + 570, + 113, + 581 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 570, + 113, + 581 + ], + "spans": [ + { + "bbox": [ + 47, + 570, + 113, + 581 + ], + "type": "text", + "content": "3.2. Database" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 46, + 587, + 287, + 671 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 587, + 287, + 671 + ], + "spans": [ + { + "bbox": [ + 46, + 587, + 287, + 671 + ], + "type": "text", + "content": "To estimate the location of the queries through image retrieval, we need a worldwide database of images, where each image is labeled with its position. We build such a database from an open-source, composited, cloudless collection of Sentinel-2 satellite imagery6. This collection contains imagery of worldwide landmasses at resolutions up to 15 meters per pixel. Some examples are shown in Fig. 3." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 46, + 672, + 287, + 695 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 672, + 287, + 695 + ], + "spans": [ + { + "bbox": [ + 46, + 672, + 287, + 695 + ], + "type": "text", + "content": "For our database we include any land area between latitudes " + }, + { + "bbox": [ + 46, + 672, + 287, + 695 + ], + "type": "inline_equation", + "content": "60^{\\circ}" + }, + { + "bbox": [ + 46, + 672, + 287, + 695 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 672, + 287, + 695 + ], + "type": "inline_equation", + "content": "-60^{\\circ}" + }, + { + "bbox": [ + 46, + 672, + 287, + 695 + ], + "type": "text", + "content": " (i.e., the area traversed by the ISS). The" + } + ] + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 309, + 72, + 543, + 171 + ], + "blocks": [ + { + "bbox": [ + 309, + 72, + 543, + 171 + ], + "lines": [ + { + "bbox": [ + 309, + 72, + 543, + 171 + ], + "spans": [ + { + "bbox": [ + 309, + 72, + 543, + 171 + ], + "type": "image", + "image_path": "33b3b1bed9735d37e312500862e5cda6e8b4d980cb872faa7031a76982e95ee0.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 179, + 547, + 246 + ], + "lines": [ + { + "bbox": [ + 305, + 179, + 547, + 246 + ], + "spans": [ + { + "bbox": [ + 305, + 179, + 547, + 246 + ], + "type": "text", + "content": "Figure 4. To create evaluation sets we choose all images that could contain a Point of Interest (POI) - all photos with nadir point within " + }, + { + "bbox": [ + 305, + 179, + 547, + 246 + ], + "type": "inline_equation", + "content": "d_{visible}" + }, + { + "bbox": [ + 305, + 179, + 547, + 246 + ], + "type": "text", + "content": ". To localize all photos within this range, even if they do not contain the POI, we create a database that contains all areas visible from the nadir points of all selected photos, yielding a database area of about " + }, + { + "bbox": [ + 305, + 179, + 547, + 246 + ], + "type": "inline_equation", + "content": "5000\\mathrm{km}^2" + }, + { + "bbox": [ + 305, + 179, + 547, + 246 + ], + "type": "text", + "content": " per POI." + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "type": "table", + "bbox": [ + 307, + 257, + 544, + 326 + ], + "blocks": [ + { + "bbox": [ + 307, + 257, + 544, + 326 + ], + "lines": [ + { + "bbox": [ + 307, + 257, + 544, + 326 + ], + "spans": [ + { + "bbox": [ + 307, + 257, + 544, + 326 + ], + "type": "table", + "html": "
Name(Lat, Lon)# queries# DBInterest
Texas(30, -95)614234kOften photoed (many queries)
Alps(45, 10)239453kSci. interest - glacial change
California(38, -122)356830kDisaster response - wildfires
Gobi Desert(40, 105)72654kSci. interest - desertification
Amazon(-3, -60)68219kSci. interest - deforestation
Toshka Lakes(23, 30)216463kSci. interest - flood monitoring
", + "image_path": "14c0ad698dc9285b07d6ddfd2f69e54cd9bdd8709f140135ec9d59f9c88fce7c.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "table_body" + } + ], + "index": 19 + }, + { + "bbox": [ + 305, + 329, + 545, + 362 + ], + "lines": [ + { + "bbox": [ + 305, + 329, + 545, + 362 + ], + "spans": [ + { + "bbox": [ + 305, + 329, + 545, + 362 + ], + "type": "text", + "content": "Table 1. Evaluation sets. The six sets cover diverse geographies and relevant areas for research on topics like climate change and disaster management. DB stands for database." + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 305, + 370, + 545, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 370, + 545, + 418 + ], + "spans": [ + { + "bbox": [ + 305, + 370, + 545, + 418 + ], + "type": "text", + "content": "Sentinel-2 imagery is available as map tiles, and we use tiles at zoom levels 9, 10, 11, with resolutions ranging from " + }, + { + "bbox": [ + 305, + 370, + 545, + 418 + ], + "type": "inline_equation", + "content": "\\sim 300" + }, + { + "bbox": [ + 305, + 370, + 545, + 418 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 305, + 370, + 545, + 418 + ], + "type": "inline_equation", + "content": "\\sim 75" + }, + { + "bbox": [ + 305, + 370, + 545, + 418 + ], + "type": "text", + "content": " meters per pixel7, to ensure that the database encompasses the large scale range of the query images." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 304, + 418, + 546, + 538 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 418, + 546, + 538 + ], + "spans": [ + { + "bbox": [ + 304, + 418, + 546, + 538 + ], + "type": "text", + "content": "These parameters lead us to define roughly 175k regions, where each region " + }, + { + "bbox": [ + 304, + 418, + 546, + 538 + ], + "type": "inline_equation", + "content": "\\mathcal{R}" + }, + { + "bbox": [ + 304, + 418, + 546, + 538 + ], + "type": "text", + "content": " is the projection of a square with known corner coordinates onto the Earth's surface. Regions have areas ranging from 5000 sq. km to 900,000 sq. km. We take partially (up to " + }, + { + "bbox": [ + 304, + 418, + 546, + 538 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 304, + 418, + 546, + 538 + ], + "type": "text", + "content": ") overlapping regions so that each query has at least one image in the database with considerable overlap (Fig. 3 top). From each region we collect four images, one per year from 2018, 2019, 2020 and 2021, covering temporal changes in a given area (Fig. 3, bottom). This produces a database of 700k images." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 304, + 538, + 545, + 634 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 538, + 545, + 634 + ], + "spans": [ + { + "bbox": [ + 304, + 538, + 545, + 634 + ], + "type": "text", + "content": "Unlike our queries, these database images have undergone post-processing that includes atmospheric corrections as well as cloud pixel minimization via composition of images from multiple days. Additionally, the database images are always taken nadir facing (i.e. perpendicular to Earth's surface) to minimize obliquity effects. Each database image has resolution " + }, + { + "bbox": [ + 304, + 538, + 545, + 634 + ], + "type": "inline_equation", + "content": "1024 \\times 1024" + }, + { + "bbox": [ + 304, + 538, + 545, + 634 + ], + "type": "text", + "content": ". Further images from the database are shown in Sec. 10 of the Supplementary." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 306, + 642, + 400, + 653 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 642, + 400, + 653 + ], + "spans": [ + { + "bbox": [ + 306, + 642, + 400, + 653 + ], + "type": "text", + "content": "3.3. Evaluation sets" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 305, + 661, + 545, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 661, + 545, + 685 + ], + "spans": [ + { + "bbox": [ + 305, + 661, + 545, + 685 + ], + "type": "text", + "content": "With the end goal of fast and accurate localization of the astronaut photography archive, we propose a deployment" + } + ] + } + ], + "index": 25 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 58, + 702, + 146, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 702, + 146, + 712 + ], + "spans": [ + { + "bbox": [ + 58, + 702, + 146, + 712 + ], + "type": "inline_equation", + "content": "^{6}" + }, + { + "bbox": [ + 58, + 702, + 146, + 712 + ], + "type": "text", + "content": "https://s2maps.eu" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 306, + 693, + 545, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 693, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 306, + 693, + 545, + 712 + ], + "type": "text", + "content": "7https://wiki.openstreetmap.org/wiki/Zoom levels" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12757" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 86, + 69, + 509, + 220 + ], + "blocks": [ + { + "bbox": [ + 86, + 69, + 509, + 220 + ], + "lines": [ + { + "bbox": [ + 86, + 69, + 509, + 220 + ], + "spans": [ + { + "bbox": [ + 86, + 69, + 509, + 220 + ], + "type": "image", + "image_path": "7e61e7cb16e152fa920fc6f0f1a235a889bbc949d1e81cd9b07bffdec444231e.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 228, + 547, + 262 + ], + "lines": [ + { + "bbox": [ + 46, + 228, + 547, + 262 + ], + "spans": [ + { + "bbox": [ + 46, + 228, + 547, + 262 + ], + "type": "text", + "content": "Figure 5. Training strategy. Naive batching produces too easy a task for the model to learn robust representations from. We increase difficulty by clustering batches by similarity and adding year-wise augmentations, which move images of different regions, but from the same year, closer together in feature space. The model must then learn representations that discern similarity in this more difficult context." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 270, + 288, + 366 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 270, + 288, + 366 + ], + "spans": [ + { + "bbox": [ + 46, + 270, + 288, + 366 + ], + "type": "text", + "content": "environment that takes advantage of the known nadir points and visibility limits associated with each image. So while localization can be performed by matching any given query against the whole world-wide database, a more suitable approach is to follow a divide and conquer paradigm by grouping queries according to their nadir, which allows us to use only a subset of the database for each search, increasing speed and improving results." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 367, + 287, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 367, + 287, + 450 + ], + "spans": [ + { + "bbox": [ + 46, + 367, + 287, + 450 + ], + "type": "text", + "content": "We seek to create evaluation sets that mirror the situation seen in deployment, so we evaluate on sets of queries that have a nadir within a " + }, + { + "bbox": [ + 46, + 367, + 287, + 450 + ], + "type": "inline_equation", + "content": "2500\\mathrm{km}" + }, + { + "bbox": [ + 46, + 367, + 287, + 450 + ], + "type": "text", + "content": " radius from a chosen Point of Interest (POI). We then build a database of images such that the database fully encompasses the area that could have been photographed from each collected query, which is an area " + }, + { + "bbox": [ + 46, + 367, + 287, + 450 + ], + "type": "inline_equation", + "content": "2*d_{visible}" + }, + { + "bbox": [ + 46, + 367, + 287, + 450 + ], + "type": "text", + "content": " from the original POI, as shown in Fig. 4." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 450, + 287, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 450, + 287, + 498 + ], + "spans": [ + { + "bbox": [ + 46, + 450, + 287, + 498 + ], + "type": "text", + "content": "Since we need the queries' ground truth position for evaluation, we choose queries that have already been localized with exhaustive image matching techniques [58]. For these queries, the full photo extent is publicly available8." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 498, + 287, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 498, + 287, + 545 + ], + "spans": [ + { + "bbox": [ + 46, + 498, + 287, + 545 + ], + "type": "text", + "content": "To properly evaluate different methods, we choose six POIs that represent various geographies but all have scientific relevance, and center an evaluation set around each. These are summarized in Tab. 1." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 557, + 103, + 569 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 557, + 103, + 569 + ], + "spans": [ + { + "bbox": [ + 47, + 557, + 103, + 569 + ], + "type": "text", + "content": "4. Method" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 578, + 287, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 578, + 287, + 685 + ], + "spans": [ + { + "bbox": [ + 46, + 578, + 287, + 685 + ], + "type": "text", + "content": "The goal of this work is to estimate the location of each astronaut photo query, which we tackle through image retrieval: for any given query, the objective is to find the most similar image(s) from the database, and use the corresponding location labels to confidently localize the query. With this in mind, we aim to train a retrieval model on satellite imagery such that the model is robust to changes in scale, perspective, and color, as well as temporal/seasonal variation. Specifically, the model's task is to extract features" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 270, + 545, + 294 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 270, + 545, + 294 + ], + "spans": [ + { + "bbox": [ + 306, + 270, + 545, + 294 + ], + "type": "text", + "content": "from each image. Then, for each query, the most similar database image is retrieved via kNN in feature space." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 295, + 545, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 295, + 545, + 319 + ], + "spans": [ + { + "bbox": [ + 305, + 295, + 545, + 319 + ], + "type": "text", + "content": "In the next subsections we describe how we train our retrieval model using the proposed database." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 324, + 411, + 337 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 324, + 411, + 337 + ], + "spans": [ + { + "bbox": [ + 306, + 324, + 411, + 337 + ], + "type": "text", + "content": "4.1. Baseline Training" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 342, + 545, + 461 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 342, + 545, + 461 + ], + "spans": [ + { + "bbox": [ + 304, + 342, + 545, + 461 + ], + "type": "text", + "content": "We want our model to extract relevant features of satellite imagery, so a straightforward approach would be to train via contrastive learning: for a given batch of images (samples), generate a number of augmented views (typically two), and apply a contrastive loss to maximize the distance in feature space between different samples and minimize distance between different views of the same sample. However, this solution would not allow the model to learn robustness to temporal changes, because the augmented views come from the same sample, with no inherent temporal variation." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 462, + 545, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 462, + 545, + 594 + ], + "spans": [ + { + "bbox": [ + 304, + 462, + 545, + 594 + ], + "type": "text", + "content": "Therefore, instead of producing multiple views of the same image via synthetic augmentation, we train the model by randomly selecting a number of regions, and for each region we consider its four images (a quadruplet), one per year. Within a batch of quadruplets, the four images from each region are mutual positives, whereas any two images from different regions are considered negatives: we can therefore consider each region as a class. This provides a natural way to address temporal variation. We can then train the model through a multi-similarity loss [67]. This idea is shown in Fig. 5 (left)." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 599, + 414, + 611 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 599, + 414, + 611 + ], + "spans": [ + { + "bbox": [ + 306, + 599, + 414, + 611 + ], + "type": "text", + "content": "4.2. Clustered Batches" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "content": "The baseline paradigm simply creates batches of random regions – a single batch of quadruplets could contain samples depicting shorelines, islands, deserts, and mountains. The loss operates batchwise, so large intra-batch variability can lead the model to learn too swiftly to distinguish between such diverse regions. In these situations, training converges quickly and the model stops short of learning to discriminate between similar-looking, but distinct, regions." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 47, + 693, + 286, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 693, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 693, + 286, + 713 + ], + "type": "text", + "content": "https://eol.jsc.nasa.gov/SearchPhotos/PhotosDatabaseAPI/(mlcoordTable)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12758" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 144 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 144 + ], + "type": "text", + "content": "To avoid this issue, we want to train with batches of highly similar images, which are still from different regions. To this end, we compute clusters of images with similar representations. Every " + }, + { + "bbox": [ + 46, + 72, + 287, + 144 + ], + "type": "inline_equation", + "content": "N_{iter}" + }, + { + "bbox": [ + 46, + 72, + 287, + 144 + ], + "type": "text", + "content": " training iterations, we extract the features from each region with the trained model, using only images from a single year, and group them into " + }, + { + "bbox": [ + 46, + 72, + 287, + 144 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 46, + 72, + 287, + 144 + ], + "type": "text", + "content": " clusters." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 144, + 288, + 288 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 144, + 288, + 288 + ], + "spans": [ + { + "bbox": [ + 46, + 144, + 288, + 288 + ], + "type": "text", + "content": "Then, we create training batches using samples from a given cluster, choosing a different cluster for each batch. This increases the difficulty of the task in each batch, ensuring that the model learns to discriminate between similar-looking images of different regions. Note that clusters are not geographically related, but related in feature space. For example, a cluster may contain images from the Gobi and the Sahara deserts, which both appear similar in feature space despite being geographically distant. This procedure, which can be seen as an offline mining approach, increases training time by less than " + }, + { + "bbox": [ + 46, + 144, + 288, + 288 + ], + "type": "inline_equation", + "content": "5\\%" + }, + { + "bbox": [ + 46, + 144, + 288, + 288 + ], + "type": "text", + "content": " but improves recall by roughly " + }, + { + "bbox": [ + 46, + 144, + 288, + 288 + ], + "type": "inline_equation", + "content": "15\\%" + }, + { + "bbox": [ + 46, + 144, + 288, + 288 + ], + "type": "text", + "content": " on average (see Sec. 5.3), a worthwhile trade off." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 289, + 287, + 325 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 289, + 287, + 325 + ], + "spans": [ + { + "bbox": [ + 46, + 289, + 287, + 325 + ], + "type": "text", + "content": "This concept is illustrated in Fig. 5 (middle), and qualitative examples of images that are clustered together are shown in Sec. 11.2 of the Supplementary." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 334, + 212, + 346 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 334, + 212, + 346 + ], + "spans": [ + { + "bbox": [ + 47, + 334, + 212, + 346 + ], + "type": "text", + "content": "4.3. Year-Wise Data Augmentation" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 353, + 287, + 435 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 353, + 287, + 435 + ], + "spans": [ + { + "bbox": [ + 46, + 353, + 287, + 435 + ], + "type": "text", + "content": "Query images span the ongoing " + }, + { + "bbox": [ + 46, + 353, + 287, + 435 + ], + "type": "inline_equation", + "content": "20+" + }, + { + "bbox": [ + 46, + 353, + 287, + 435 + ], + "type": "text", + "content": " years of ISS operations, as well as multiple camera types, camera operators, and view points, so queries depicting the same location can have a number of natural \"transformations\". These range from blueish hues from lens filters, to skewed representations in highly oblique shots, to seasonal variation in ground cover (e.g. snow, autumn colors)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 437, + 287, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 437, + 287, + 509 + ], + "spans": [ + { + "bbox": [ + 46, + 437, + 287, + 509 + ], + "type": "text", + "content": "To make the model robust to such transformations, we employ data augmentation. Typically, augmentations are applied per image or per batch, but we argue that to truly learn robustness to the augmentations in this setting, the same augmentation should be performed on images from different regions, but from the same year." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 510, + 287, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 510, + 287, + 677 + ], + "spans": [ + { + "bbox": [ + 46, + 510, + 287, + 677 + ], + "type": "text", + "content": "Formally, at each training iteration we choose a set of 4 random augmentations, one for each year " + }, + { + "bbox": [ + 46, + 510, + 287, + 677 + ], + "type": "inline_equation", + "content": "A_{2018}, A_{2019}, A_{2020}, A_{2021}" + }, + { + "bbox": [ + 46, + 510, + 287, + 677 + ], + "type": "text", + "content": ". We then apply the corresponding augmentation to all images from that year (i.e. " + }, + { + "bbox": [ + 46, + 510, + 287, + 677 + ], + "type": "inline_equation", + "content": "A_{2018}" + }, + { + "bbox": [ + 46, + 510, + 287, + 677 + ], + "type": "text", + "content": " is applied to all images from 2018, across all regions). Such \"year-wise\" augmentation has the effect of moving images from a given year closer in feature space to each other, as shown in Fig. 5 (right). In the loss, images from the same region, but different years, are treated as positives, while images from different regions are treated as negatives. This has the effect that images that receive the same augmentation are assigned different classes – forcing the model to learn to ignore such augmentations – and instead pull together images from the same region, across different years." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "text", + "content": "Furthermore, given that the same augmentation is applied to multiple images, augmentation can be batched and performed on the GPU, making training slightly faster (by" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 323, + 74, + 531, + 215 + ], + "blocks": [ + { + "bbox": [ + 323, + 74, + 531, + 215 + ], + "lines": [ + { + "bbox": [ + 323, + 74, + 531, + 215 + ], + "spans": [ + { + "bbox": [ + 323, + 74, + 531, + 215 + ], + "type": "image", + "image_path": "57ed52742026028b9e6109c4422fbc58eb7526e4700d658a56c4124fc9e8dae4.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 222, + 545, + 299 + ], + "lines": [ + { + "bbox": [ + 304, + 222, + 545, + 299 + ], + "spans": [ + { + "bbox": [ + 304, + 222, + 545, + 299 + ], + "type": "text", + "content": "Figure 6. Dataset collection from satellite imagery and handling in Neutral-Aware Multi-Similarity Loss. For a given image (yellow), images with no overlap are negatives (red), images from the same region are positives (green, overlapping yellow), and images with partial overlap are neutral (grey). Note that the database is made of images at 3 different zooms (each double in size than the next), which explains the regions of different size." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 303, + 531, + 316 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 303, + 531, + 316 + ], + "spans": [ + { + "bbox": [ + 306, + 303, + 531, + 316 + ], + "type": "inline_equation", + "content": "5\\%" + }, + { + "bbox": [ + 306, + 303, + 531, + 316 + ], + "type": "text", + "content": " on average, although this depends on the hardware)." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 323, + 501, + 335 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 323, + 501, + 335 + ], + "spans": [ + { + "bbox": [ + 306, + 323, + 501, + 335 + ], + "type": "text", + "content": "4.4. Neutral-Aware Multi-Similarity Loss" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 341, + 545, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 341, + 545, + 389 + ], + "spans": [ + { + "bbox": [ + 304, + 341, + 545, + 389 + ], + "type": "text", + "content": "The powerful Multi-Similarity loss is designed to take as input multiple images from a number of classes, along with the corresponding class indices, and compute a loss from the positives and negatives for each image." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 389, + 545, + 532 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 389, + 545, + 532 + ], + "spans": [ + { + "bbox": [ + 304, + 389, + 545, + 532 + ], + "type": "text", + "content": "However, it does not take into account that in some cases, pairs of images are neither positives nor negatives, and should be considered \"neutral\". While the concept of neutral images does not exist in standard image retrieval tasks, as an image either belongs to class A or class B, in APL it must be taken into consideration. For example, should two images, which have an intersection over union of " + }, + { + "bbox": [ + 304, + 389, + 545, + 532 + ], + "type": "inline_equation", + "content": "25\\%" + }, + { + "bbox": [ + 304, + 389, + 545, + 532 + ], + "type": "text", + "content": ", be considered positives or negatives? This is precisely the case when two images of the same area, but from different zoom levels, are batched together, as well as overlapping images at the same level (see Fig. 6). We believe the answer is neither, so we add the \"neutral\" case to the loss." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 532, + 545, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 532, + 545, + 568 + ], + "spans": [ + { + "bbox": [ + 304, + 532, + 545, + 568 + ], + "type": "text", + "content": "Formally, we model the neutral case with the indicator function " + }, + { + "bbox": [ + 304, + 532, + 545, + 568 + ], + "type": "inline_equation", + "content": "\\mathcal{I}_{NA}" + }, + { + "bbox": [ + 304, + 532, + 545, + 568 + ], + "type": "text", + "content": ", which is 0 when the regions " + }, + { + "bbox": [ + 304, + 532, + 545, + 568 + ], + "type": "inline_equation", + "content": "\\mathcal{R}_i" + }, + { + "bbox": [ + 304, + 532, + 545, + 568 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 532, + 545, + 568 + ], + "type": "inline_equation", + "content": "\\mathcal{R}_j" + }, + { + "bbox": [ + 304, + 532, + 545, + 568 + ], + "type": "text", + "content": " intersect but are not equal, and 1 otherwise." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 350, + 577, + 545, + 592 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 350, + 577, + 545, + 592 + ], + "spans": [ + { + "bbox": [ + 350, + 577, + 545, + 592 + ], + "type": "interline_equation", + "content": "\\mathcal {I} _ {N A} \\left(\\mathcal {R} _ {i}, \\mathcal {R} _ {j}\\right) = \\mathbb {1} _ {\\left(\\mathcal {R} _ {i} \\cap \\mathcal {R} _ {j}\\right) \\wedge \\left(\\mathcal {R} _ {i} \\neq \\mathcal {R} _ {j}\\right)} \\tag {2}", + "image_path": "79c0f34d68fffd7dcde9fea404d7ed7a0233cfb611f3e8b9ba494871ef74b3ed.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 599, + 539, + 611 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 599, + 539, + 611 + ], + "spans": [ + { + "bbox": [ + 306, + 599, + 539, + 611 + ], + "type": "text", + "content": "With this, the neutral-aware multi-similarity loss becomes" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 312, + 619, + 545, + 695 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 619, + 545, + 695 + ], + "spans": [ + { + "bbox": [ + 312, + 619, + 545, + 695 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} _ {N A M S} = \\frac {1}{B S} \\sum_ {i = 1} ^ {B S} \\left\\{\\frac {1}{\\alpha} \\log \\left[ 1 + \\sum_ {k \\in \\mathcal {R} _ {i}} e ^ {- \\alpha \\left(\\mathcal {S} _ {i k} - \\lambda\\right)} \\right] \\right. \\\\ \\left. + \\frac {1}{\\beta} \\log \\left[ 1 + \\sum_ {k \\notin \\mathcal {R} _ {i}} \\mathcal {I} _ {N A} \\left(\\mathcal {R} _ {i}, \\mathcal {R} _ {k}\\right) e ^ {- \\beta \\left(\\mathcal {S} _ {i k} - \\lambda\\right)} \\right] \\right\\} \\end{array} \\tag {3}", + "image_path": "99099920164b4900556554e6829be64cc61b3ec113e28fcec84e1d5a8882534f.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 306, + 701, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 701, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 701, + 545, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 306, + 701, + 545, + 713 + ], + "type": "inline_equation", + "content": "BS" + }, + { + "bbox": [ + 306, + 701, + 545, + 713 + ], + "type": "text", + "content": " is the number of images in a batch, " + }, + { + "bbox": [ + 306, + 701, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{R}_i" + }, + { + "bbox": [ + 306, + 701, + 545, + 713 + ], + "type": "text", + "content": " is the" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12759" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 48, + 70, + 547, + 191 + ], + "blocks": [ + { + "bbox": [ + 48, + 70, + 547, + 191 + ], + "lines": [ + { + "bbox": [ + 48, + 70, + 547, + 191 + ], + "spans": [ + { + "bbox": [ + 48, + 70, + 547, + 191 + ], + "type": "table", + "html": "
MethodType of Training ImageryTexas (validation)AlpsCaliforniaGobi DesertAmazonToshka Lakes
R@1R@10R@100R@1R@10R@100R@1R@10R@100R@1R@10R@100R@1R@10R@100R@1R@10R@100
Nadir-2.4--1.2--2.4--1.8--3.1--1.4--
Random Choice-0.21.715.50.11.111.60.22.320.10.11.013.20.11.111.50.21.29.1
NetVLAD [4]Urban VPR3.613.434.85.215.236.24.314.439.02.17.724.56.619.946.39.622.545.6
SFRS [23]Urban VPR6.817.440.97.017.841.45.417.244.45.214.737.69.423.852.510.820.442.4
Conv-AP [1]Urban VPR3.713.537.72.08.825.03.815.441.23.79.024.27.021.147.75.012.332.6
CosPlace [9]Urban VPR4.915.242.05.919.345.24.815.243.37.917.539.58.523.954.49.825.453.0
MixVPR [2]Urban VPR4.515.340.13.18.622.64.417.543.24.010.124.59.828.758.75.915.436.5
EigenPlaces [10]Urban VPR6.421.652.48.721.450.36.022.854.38.120.040.910.926.557.914.330.960.4
AnyLoc [30] (DINOv2 [41] + NetVLAD)Universal VPR44.168.787.840.770.892.048.775.091.628.757.081.738.663.886.263.784.596.3
TorchGeo [57] (ResNet50 w MOCO [26])Satellite1.03.211.60.31.35.71.14.316.40.42.88.31.96.619.40.72.99.9
TorchGeo [57] (ResNet50 w SeCo [70])Satellite6.115.641.77.420.249.25.114.537.13.714.038.94.613.332.95.715.638.5
TorchGeo [57] (ResNet50 w GASSL [5])Satellite9.722.846.49.123.150.513.331.458.86.317.545.48.320.340.120.438.664.2
OGCL UAV-View [18] (ConvNeXt-XXLarge)UAV16.235.365.814.434.164.719.642.171.05.520.545.310.026.354.721.138.463.4
OGCL UAV-View [18] (ViT-L/14) [19]UAV17.633.255.914.633.263.922.848.174.97.622.850.120.439.162.531.851.874.4
MBEG [73] (ViT-L/14)UAV7.017.635.16.619.345.98.720.741.54.415.038.06.417.139.38.120.749.1
EarthLoc (Ours)Satellite55.973.088.358.476.889.558.076.091.451.167.586.547.267.984.672.285.093.3
", + "image_path": "33800021aeed8899d218f9b4d84da704d50213491d242e1ce954c3dfc9503d46.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 196, + 547, + 251 + ], + "lines": [ + { + "bbox": [ + 46, + 196, + 547, + 251 + ], + "spans": [ + { + "bbox": [ + 46, + 196, + 547, + 251 + ], + "type": "text", + "content": "Table 2. Results from different methods on our query sets. Methods are grouped into (1) naive, (2) VPR, (3) Remote Sensing (UAV+Satellite), (4) other baselines trained on our database with clustered batches (Sec. 4.2). For all methods we use " + }, + { + "bbox": [ + 46, + 196, + 547, + 251 + ], + "type": "inline_equation", + "content": "4x90TTA" + }, + { + "bbox": [ + 46, + 196, + 547, + 251 + ], + "type": "text", + "content": ". VPR stands for Visual Place Recognition, UAV for Unmanned Aerial Vehicle. Best results overall in bold, best per-group underlined. EarthLoc outperforms other models, and performs competitively with AnyLoc, while having 50 times faster features extraction and 10 times smaller features." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 258, + 287, + 282 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 258, + 287, + 282 + ], + "spans": [ + { + "bbox": [ + 46, + 258, + 287, + 282 + ], + "type": "text", + "content": "region of image " + }, + { + "bbox": [ + 46, + 258, + 287, + 282 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 258, + 287, + 282 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 258, + 287, + 282 + ], + "type": "inline_equation", + "content": "\\mathcal{S}_{ik}" + }, + { + "bbox": [ + 46, + 258, + 287, + 282 + ], + "type": "text", + "content": " is the similarity between image " + }, + { + "bbox": [ + 46, + 258, + 287, + 282 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 258, + 287, + 282 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 258, + 287, + 282 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 46, + 258, + 287, + 282 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 258, + 287, + 282 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 46, + 258, + 287, + 282 + ], + "type": "text", + "content": " is a margin, and " + }, + { + "bbox": [ + 46, + 258, + 287, + 282 + ], + "type": "inline_equation", + "content": "\\alpha, \\beta" + }, + { + "bbox": [ + 46, + 258, + 287, + 282 + ], + "type": "text", + "content": " are hyperparameters." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 283, + 287, + 331 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 283, + 287, + 331 + ], + "spans": [ + { + "bbox": [ + 46, + 283, + 287, + 331 + ], + "type": "text", + "content": "We believe that the neutral-aware multi-similarity loss is applicable to any task where it is not straightforward how to enforce full positivity or negativity between any pair of images, including localization tasks like VPR." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 342, + 128, + 357 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 342, + 128, + 357 + ], + "spans": [ + { + "bbox": [ + 47, + 342, + 128, + 357 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 362, + 168, + 376 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 362, + 168, + 376 + ], + "spans": [ + { + "bbox": [ + 47, + 362, + 168, + 376 + ], + "type": "text", + "content": "5.1. Experimental setting" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 381, + 118, + 395 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 381, + 118, + 395 + ], + "spans": [ + { + "bbox": [ + 47, + 381, + 118, + 395 + ], + "type": "text", + "content": "5.1.1 Training" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 401, + 288, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 401, + 288, + 533 + ], + "spans": [ + { + "bbox": [ + 46, + 401, + 288, + 533 + ], + "type": "text", + "content": "In training, we use a batch size of 96 quadruplets (384 images), where one quadruplet corresponds to four images within a region. Our model has a MixVPR-style architecture [2], with a ResNet50 [25] backbone and an MLP-Mixer[62] with output dimension 4096. We train for 10k iterations with the Adam optimizer [33] and learning rate 0.0001. Clustering (see Sec. 4.2) is performed every 2k iterations with the number of clusters " + }, + { + "bbox": [ + 46, + 401, + 288, + 533 + ], + "type": "inline_equation", + "content": "C = 50" + }, + { + "bbox": [ + 46, + 401, + 288, + 533 + ], + "type": "text", + "content": ". As augmentation, we use torchvision's color jittering, random perspective, and random rotation. Training takes between 5 and 7 hours on an A100 GPU for all methods described in Sec. 4." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 550, + 121, + 561 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 550, + 121, + 561 + ], + "spans": [ + { + "bbox": [ + 47, + 550, + 121, + 561 + ], + "type": "text", + "content": "5.1.2 Inference" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 569, + 287, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 569, + 287, + 605 + ], + "spans": [ + { + "bbox": [ + 46, + 569, + 287, + 605 + ], + "type": "text", + "content": "Due to the imagery acquisition conditions, we cannot assume any canonical orientation for the photos (in space, there's no gravity vector, i.e., no \"up\" or \"down\")." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 606, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 606, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 606, + 287, + 714 + ], + "type": "text", + "content": "This poses a challenge for any retrieval method – to address it, we generate four feature embeddings for each database image, one each after rotating the image for angles " + }, + { + "bbox": [ + 46, + 606, + 287, + 714 + ], + "type": "inline_equation", + "content": "r \\in [0,90,180,270]" + }, + { + "bbox": [ + 46, + 606, + 287, + 714 + ], + "type": "text", + "content": ", in a process we call " + }, + { + "bbox": [ + 46, + 606, + 287, + 714 + ], + "type": "inline_equation", + "content": "4 \\times 90^{\\circ}" + }, + { + "bbox": [ + 46, + 606, + 287, + 714 + ], + "type": "text", + "content": " test time augmentation (4x90TTA). For fairness, we do this for all methods in all experiments, given that the models are not trained to be robust to such large arbitrary rotations. This increases latency by " + }, + { + "bbox": [ + 46, + 606, + 287, + 714 + ], + "type": "inline_equation", + "content": "4\\mathrm{x}" + }, + { + "bbox": [ + 46, + 606, + 287, + 714 + ], + "type": "text", + "content": ", but the method is still orders of magnitude faster than previous approaches (retrieval takes 0.05" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 258, + 547, + 390 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 258, + 547, + 390 + ], + "spans": [ + { + "bbox": [ + 304, + 258, + 547, + 390 + ], + "type": "text", + "content": "seconds per query with " + }, + { + "bbox": [ + 304, + 258, + 547, + 390 + ], + "type": "inline_equation", + "content": "4x90TTA" + }, + { + "bbox": [ + 304, + 258, + 547, + 390 + ], + "type": "text", + "content": ", while previous works required " + }, + { + "bbox": [ + 304, + 258, + 547, + 390 + ], + "type": "inline_equation", + "content": "\\sim 1" + }, + { + "bbox": [ + 304, + 258, + 547, + 390 + ], + "type": "text", + "content": " minute [58] per query). Memory requirements also increase by four times, although this is far from being an issue for a standard workstation, since all database features for the whole Earth fit in less than 12GB of RAM. Experiments without " + }, + { + "bbox": [ + 304, + 258, + 547, + 390 + ], + "type": "inline_equation", + "content": "4x90TTA" + }, + { + "bbox": [ + 304, + 258, + 547, + 390 + ], + "type": "text", + "content": " are shown in the Supplementary in Sec. 7. As an added bonus, since each feature from the database refers to an image with a certain orientation (w.r.t. North), retrieval predictions also give an estimate of the orientation of the query. We use features from the 2021 satellite imagery as our global database." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 408, + 547, + 482 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 408, + 547, + 482 + ], + "spans": [ + { + "bbox": [ + 304, + 408, + 547, + 482 + ], + "type": "text", + "content": "Success Criteria With the final goal of localizing the 4.5M astronaut photograph queries in mind, we define a predicted image as correct if there is non-zero overlap between the prediction and the query. As a metric we use Recall@N, defined as the percentage of queries for which at least one of the top-N predictions is correct." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 490, + 362, + 502 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 490, + 362, + 502 + ], + "spans": [ + { + "bbox": [ + 306, + 490, + 362, + 502 + ], + "type": "text", + "content": "5.2. Results" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 510, + 547, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 510, + 547, + 605 + ], + "spans": [ + { + "bbox": [ + 304, + 510, + 547, + 605 + ], + "type": "text", + "content": "In Tab. 2 we present results from a large number of models from multiple domains. We first show the recall when naively predicting the query to be directly nadir, proving that the vast majority of images are taken at an angle, and the poor performance of randomly choosing a prediction from the database - a symptom of the large-scale nature of the task (i.e., localizing a 5000 sq.km wide image within a 20M sq.km area)." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 605, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 547, + 713 + ], + "type": "text", + "content": "Results with methods trained for Urban Visual Place Recognition (VPR) provide, somewhat unsurprisingly, poor results - on the other hand, AnyLoc [30] shows remarkably strong results, correctly predicting the location of the query as the first prediction for roughly " + }, + { + "bbox": [ + 304, + 605, + 547, + 713 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 304, + 605, + 547, + 713 + ], + "type": "text", + "content": " of the queries across all evaluation sets, although at the cost of over 50x slower feature extraction and 10x bigger features than its VPR counterparts (AnyLoc uses 49152-D features vs 4096 of NetVLAD)." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "12760" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 50, + 70, + 545, + 126 + ], + "blocks": [ + { + "bbox": [ + 50, + 70, + 545, + 126 + ], + "lines": [ + { + "bbox": [ + 50, + 70, + 545, + 126 + ], + "spans": [ + { + "bbox": [ + 50, + 70, + 545, + 126 + ], + "type": "table", + "html": "
Multi-similarity LossUsing regional quadrupletsClustered BatchesYear-Wise Data AugmentationNeutral-Aware Multi-SimilarityTexas (validation)AlpsCaliforniaGobi DesertAmazonToshka LakesAverage
R@1R@100R@1R@100R@1R@100R@1R@100R@1R@100R@1R@100R@1R@100
✓25.571.424.877.327.876.616.868.823.661.538.880.826.272.7
✓✓45.084.145.487.349.588.535.779.934.280.262.991.145.485.2
✓✓✓51.487.549.689.553.990.640.884.339.382.864.093.049.888.0
✓✓✓✓54.687.553.987.255.991.646.882.945.682.467.691.954.187.2
✓✓✓55.388.255.190.057.290.344.685.848.484.868.992.954.988.7
✓✓✓✓55.988.358.489.558.091.451.186.547.284.672.293.357.188.9
", + "image_path": "838e8fba8270e7df4d564df08bc7499eca19f97f5fde4450f061e8403921ffaf.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 133, + 546, + 156 + ], + "lines": [ + { + "bbox": [ + 46, + 133, + 546, + 156 + ], + "spans": [ + { + "bbox": [ + 46, + 133, + 546, + 156 + ], + "type": "text", + "content": "Table 3. Ablations over test sets. EarthLoc uses Clustered Batches (Sec. 4.2), Year-Wise Data Augmentation (Sec. 4.3) and Neutral-Aware Multi-Similarity Loss (Sec. 4.4). Best results in bold." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 48, + 165, + 287, + 229 + ], + "blocks": [ + { + "bbox": [ + 48, + 165, + 287, + 229 + ], + "lines": [ + { + "bbox": [ + 48, + 165, + 287, + 229 + ], + "spans": [ + { + "bbox": [ + 48, + 165, + 287, + 229 + ], + "type": "image", + "image_path": "18d34865b1211f2e2c026250989968e605577ab93e31d82c1152e596d2fbd645.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 236, + 288, + 269 + ], + "lines": [ + { + "bbox": [ + 46, + 236, + 288, + 269 + ], + "spans": [ + { + "bbox": [ + 46, + 236, + 288, + 269 + ], + "type": "text", + "content": "Figure 7. Qualitative results. Each group of four images represents a query (left) and its top-3 predictions. Green and red indicate if the prediction is correct or incorrect." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 287, + 287, + 358 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 287, + 287, + 358 + ], + "spans": [ + { + "bbox": [ + 46, + 287, + 287, + 358 + ], + "type": "text", + "content": "We then test a number of models trained with self-supervised techniques, like MOCO [26], SeCo [70], and Geography-aware self-supervised learning (GASSL) [5], conveniently provided by the TorchGeo library [57]. Surprisingly, these methods did not provide competitive results, and are on average about equal with Urban VPR methods." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 361, + 287, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 361, + 287, + 445 + ], + "spans": [ + { + "bbox": [ + 46, + 361, + 287, + 445 + ], + "type": "text", + "content": "From the Unmanned Aerial Vehicle (UAV) domain, we use the latest state of the art (SOTA) methods, i.e., the winners of the recent challenge in UAV localization at ACM MM (October 2023) [72], namely OGCL [18] and MBEG [73]. Despite being SOTA in the related task of UAV localization, we see that these methods are not competitive with AnyLoc, despite outperforming most other methods." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 449, + 287, + 485 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 449, + 287, + 485 + ], + "spans": [ + { + "bbox": [ + 46, + 449, + 287, + 485 + ], + "type": "text", + "content": "Finally, EarthLoc outperforms all previous works, localizing " + }, + { + "bbox": [ + 46, + 449, + 287, + 485 + ], + "type": "inline_equation", + "content": "57.1\\%" + }, + { + "bbox": [ + 46, + 449, + 287, + 485 + ], + "type": "text", + "content": " of queries within the top-1 prediction (Recall@1) averaged on the 6 evaluation sets." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 500, + 116, + 511 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 500, + 116, + 511 + ], + "spans": [ + { + "bbox": [ + 47, + 500, + 116, + 511 + ], + "type": "text", + "content": "5.3. Ablations" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 521, + 287, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 521, + 287, + 593 + ], + "spans": [ + { + "bbox": [ + 46, + 521, + 287, + 593 + ], + "type": "text", + "content": "In Tab. 3 we ablate the components of our method to better understand performance. Results show that not only do Clustered Batches, Year-Wise Data Augmentation and Neutral-Aware Multi-Similarity Loss provide significant increases over the baseline, but they are also orthogonal improvements and can be easily combined." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 608, + 251, + 620 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 608, + 251, + 620 + ], + "spans": [ + { + "bbox": [ + 47, + 608, + 251, + 620 + ], + "type": "text", + "content": "5.4. Qualitative Analysis and Failure Cases" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 629, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 629, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 629, + 287, + 715 + ], + "type": "text", + "content": "A sample of qualitative results from EarthLoc are shown in Fig. 7. The samples show that EarthLoc picks locations that have similar characteristics to the query (top left), is robust to rotations and changes in scale (bottom right), and provides correct predictions that are difficult to correctly match even for the human eye (top left, bottom left). More qualitative results in Sec. 11 of the Supplementary." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 167, + 468, + 179 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 167, + 468, + 179 + ], + "spans": [ + { + "bbox": [ + 306, + 167, + 468, + 179 + ], + "type": "text", + "content": "5.5. Limitations and Future Work" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 187, + 547, + 306 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 187, + 547, + 306 + ], + "spans": [ + { + "bbox": [ + 304, + 187, + 547, + 306 + ], + "type": "text", + "content": "While this work takes a major step forward in the study of APL, EarthLoc focuses on daytime queries, and we did not investigate panoramic or nighttime imagery. The localization of nighttime astronaut photos in particular is important (e.g. to study light pollution [14]), but we believe this to be a separate challenge that should be addressed in future work. Another limitation is due to the zoom limits of the database: we chose database tiles at zoom levels 9, 10, 11 (see Sec. 3.2), which allows us to localize queries of area between 5000 and 900,000 square km." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 308, + 547, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 308, + 547, + 380 + ], + "spans": [ + { + "bbox": [ + 304, + 308, + 547, + 380 + ], + "type": "text", + "content": "The originality of this work opens up a large number of possible future research directions, like training directly on (some of the) localized queries, using seasonal data, post-processing predictions, obtaining pixel-wise geolocation, applying domain adaptation, and using available metadata (like camera lens) during training, to name a few." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 394, + 384, + 406 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 394, + 384, + 406 + ], + "spans": [ + { + "bbox": [ + 306, + 394, + 384, + 406 + ], + "type": "text", + "content": "6. Conclusions" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 415, + 545, + 547 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 415, + 545, + 547 + ], + "spans": [ + { + "bbox": [ + 304, + 415, + 545, + 547 + ], + "type": "text", + "content": "We introduce EarthLoc, a method for localizing astronaut photography of Earth. By reframing APL from an image matching task to an image retrieval task, and introducing a custom loss function and training scheme, our method efficiently determines the geographic location of astronaut photography given only the position of the International Space Station at the time the photo was taken. Our main contributions are the Neutral-Aware Multi-Similarity Loss, Year-Wise Data Augmentation technique, and new astronaut photography-oriented validation and test datasets to encourage future work on this problem." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 559, + 546, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 559, + 546, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 559, + 546, + 712 + ], + "type": "text", + "content": "Acknowledgements. We acknowledge the CINECA award under the ISCRA initiative, for the availability of high performance computing resources. This work was supported by CINI. Project supported by ESA Network of Resources Initiative. This study was carried out within the project FAIR - Future Artificial Intelligence Research - and received funding from the European Union Next-GenerationEU (PIANO NAZIONALE DI RIPRESA E RESILLENZA (PNRR) - MISSIONE 4 COMPONENTE 2, INVESTIMENTO 1.3 - D.D. 1555 11/10/2022, PE00000013). This manuscript reflects only the authors' views and opinions, neither the European Union nor the European Commission can be considered responsible for them. European Lighthouse on Secure and Safe AI - ELSA, HORIZON EU Grant ID: 101070617" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12761" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 123 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 123 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 123 + ], + "type": "text", + "content": "[1] Amar Ali-bey, Brahim Chaib-draa, and Philippe Giguère. Gsv-cities: Toward appropriate supervised visual place recognition. Neurocomputing, 513:194-203, 2022. 3, 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 124, + 287, + 157 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 124, + 287, + 157 + ], + "spans": [ + { + "bbox": [ + 53, + 124, + 287, + 157 + ], + "type": "text", + "content": "[2] Amar Ali-bey, Brahim Chaib-draa, and Philippe Giguère. Mixvpr: Feature mixing for visual place recognition. In WACV, pages 2998-3007, 2023. 3, 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 159, + 287, + 190 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 159, + 287, + 190 + ], + "spans": [ + { + "bbox": [ + 54, + 159, + 287, + 190 + ], + "type": "text", + "content": "[3] R. Arandjelović and Andrew Zisserman. Three things everyone should know to improve object retrieval. In CVPR, pages 2911-2918, 2012. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 192, + 287, + 235 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 192, + 287, + 235 + ], + "spans": [ + { + "bbox": [ + 53, + 192, + 287, + 235 + ], + "type": "text", + "content": "[4] Relja Arandjelović, Petr Gronat, Akihiko Torii, Tomas Pajdla, and Josef Sivic. NetVLAD: CNN architecture for weakly supervised place recognition. IEEE TPAMI, 40(6): 1437-1451, 2018. 3, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 236, + 287, + 278 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 236, + 287, + 278 + ], + "spans": [ + { + "bbox": [ + 53, + 236, + 287, + 278 + ], + "type": "text", + "content": "[5] Kumar Ayush, Burak Uzkent, Chenlin Meng, Kumar Tanmay, Marshall Burke, David Lobell, and Stefano Ermon. Geography-aware self-supervised learning. ICCV, 2021. 3, 7, 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 280, + 287, + 312 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 280, + 287, + 312 + ], + "spans": [ + { + "bbox": [ + 53, + 280, + 287, + 312 + ], + "type": "text", + "content": "[6] Artem Babenko, Anton Slesarev, A. Chigorin, and V. Lempitsky. Neural codes for image retrieval. ECCV, abs/1404.1777, 2014. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 314, + 287, + 367 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 314, + 287, + 367 + ], + "spans": [ + { + "bbox": [ + 53, + 314, + 287, + 367 + ], + "type": "text", + "content": "[7] Giovanni Barbarani, Mohamad Mostafa, Hajali Bayramov, Gabriele Trivigno, Gabriele Berton, Carlo Masone, and Barbara Caputo. Are local features all you need for cross-domain visual place recognition? In CVPRW, pages 6155-6165, 2023. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 369, + 287, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 369, + 287, + 402 + ], + "spans": [ + { + "bbox": [ + 53, + 369, + 287, + 402 + ], + "type": "text", + "content": "[8] Herbert Bay, Andreas Ess, Tinne Tuytelaars, and Luc Van Gool. Speeded-up robust features (surf). Computer Vision and Image Understanding, 110:346-359, 2008. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 403, + 287, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 403, + 287, + 435 + ], + "spans": [ + { + "bbox": [ + 53, + 403, + 287, + 435 + ], + "type": "text", + "content": "[9] Gabriele Berton, Carlo Masone, and Barbara Caputo. Rethinking visual geo-localization for large-scale applications. In CVPR, 2022. 3, 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 437, + 287, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 437, + 287, + 479 + ], + "spans": [ + { + "bbox": [ + 48, + 437, + 287, + 479 + ], + "type": "text", + "content": "[10] Gabriele Berton, Gabriele Trivigno, Barbara Caputo, and Carlo Masone. Eigenplaces: Training viewpoint robust models for visual place recognition. In ICCV, pages 11080-11090, 2023. 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 480, + 287, + 512 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 480, + 287, + 512 + ], + "spans": [ + { + "bbox": [ + 48, + 480, + 287, + 512 + ], + "type": "text", + "content": "[11] Fatih Cakir, Kun He, Xide Xia, Brian Kulis, and Stan Sclaroff. Deep metric learning to rank. In CVPR, pages 1861-1870, 2019. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 514, + 287, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 514, + 287, + 567 + ], + "spans": [ + { + "bbox": [ + 48, + 514, + 287, + 567 + ], + "type": "text", + "content": "[12] Yezhen Cong, Samar Khanna, Chenlin Meng, Patrick Liu, Erik Rozi, Yutong He, Marshall Burke, David B. Lobell, and Stefano Ermon. SatMAE: Pre-training transformers for temporal and multi-spectral satellite imagery. In NeurIPS, 2022. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 569, + 287, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 569, + 287, + 602 + ], + "spans": [ + { + "bbox": [ + 48, + 569, + 287, + 602 + ], + "type": "text", + "content": "[13] Gabriela Csurka, Christopher Dance, Lixin Fan, Jutta Willamowski, and Cedric Bray. Visual categorization with bags of keypoints. In ECCV, 2004. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 603, + 287, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 603, + 287, + 657 + ], + "spans": [ + { + "bbox": [ + 48, + 603, + 287, + 657 + ], + "type": "text", + "content": "[14] Alejandro Sánchez de Miguel, José Gómez Castaño, Jaime Zamorano, Sergio Pascual, M Ángeles, L Cayuela, Guillermo Martín Martínez, Peter Challupner, and Christopher C M Kyba. Atlas of astronaut photos of earth at night. *Astronomy and Geophysics*, 55(4):4.36-4.36, 2014. 8" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 658, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 287, + 712 + ], + "type": "text", + "content": "[15] Alejandro Sánchez de Miguel, Jonathan Bennie, Emma Rosenfeld, Simon Dzurjak, and Kevin J. Gaston. Environmental risks from artificial nighttime lighting widespread and increasing across Europe. Science Advances, 8(37): eabl6891, 2022. 2" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 32, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "text", + "content": "[16] Jiankang Deng, J. Guo, and S. Zafeiriou. ArcFace: Additive angular margin loss for deep face recognition. In CVPR, pages 4685-4694, 2019. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 107, + 545, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 107, + 545, + 140 + ], + "spans": [ + { + "bbox": [ + 308, + 107, + 545, + 140 + ], + "type": "text", + "content": "[17] Fabian Deuser, Konrad Habel, and Norbert Oswald. Sample4geo: Hard negative sampling for cross-view geolocation. In ICCV, pages 16847-16856, 2023. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 141, + 545, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 141, + 545, + 206 + ], + "spans": [ + { + "bbox": [ + 308, + 141, + 545, + 206 + ], + "type": "text", + "content": "[18] Fabian Deuser, Konrad Habel, Martin Werner, and Norbert Oswald. Orientation-guided contrastive learning for uav-view geo-localisation. In Proceedings of the 2023 Workshop on UAVs in Multimedia: Capturing the World from a New Perspective, page 7-11, New York, NY, USA, 2023. Association for Computing Machinery. 3, 7, 8" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 208, + 545, + 273 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 208, + 545, + 273 + ], + "spans": [ + { + "bbox": [ + 308, + 208, + 545, + 273 + ], + "type": "text", + "content": "[19] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale. In ICLR, 2021. 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 274, + 545, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 274, + 545, + 319 + ], + "spans": [ + { + "bbox": [ + 308, + 274, + 545, + 319 + ], + "type": "text", + "content": "[20] Florian Fervers, Sebastian Bullinger, Christoph Bodensteiner, Michael Arens, and Rainer Stiefelhagen. Uncertainty-aware vision-based metric cross-view geolocation. In CVPR, pages 21621-21631, 2023. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 320, + 545, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 320, + 545, + 373 + ], + "spans": [ + { + "bbox": [ + 308, + 320, + 545, + 373 + ], + "type": "text", + "content": "[21] Kenton Fisher, Sara Schmidt, and Alex Stoken. Crew earth observations: New tools to support your research. In 12th Annual International Space Station Research and Development Conference. Center for the Advancement of Science in Space, Inc., 2023. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 376, + 545, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 376, + 545, + 408 + ], + "spans": [ + { + "bbox": [ + 308, + 376, + 545, + 408 + ], + "type": "text", + "content": "[22] Kevin J. Gaston and Alejandro Sánchez de Miguel. Environmental impacts of artificial light at night. Annual Review of Environment and Resources, 47(1):373-398, 2022. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 410, + 545, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 410, + 545, + 453 + ], + "spans": [ + { + "bbox": [ + 308, + 410, + 545, + 453 + ], + "type": "text", + "content": "[23] Yixiao Ge, Haibo Wang, Feng Zhu, Rui Zhao, and Hongsheng Li. Self-supervising fine-grained region similarities for large-scale image localization. In ECCV, pages 369-386, Cham, 2020. Springer International Publishing. 7" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 455, + 545, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 455, + 545, + 498 + ], + "spans": [ + { + "bbox": [ + 308, + 455, + 545, + 498 + ], + "type": "text", + "content": "[24] Stephen Hausler, Sourav Garg, Ming Xu, Michael Milford, and Tobias Fischer. Patch-netvlad: Multi-scale fusion of locally-global descriptors for place recognition. In CVPR, pages 14141-14152, 2021. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 499, + 545, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 499, + 545, + 521 + ], + "spans": [ + { + "bbox": [ + 308, + 499, + 545, + 521 + ], + "type": "text", + "content": "[25] K. He, X. Zhang, S. Ren, and J. Sun. Deep residual learning for image recognition. In CVPR, pages 770-778, 2016. 7" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 522, + 545, + 565 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 522, + 545, + 565 + ], + "spans": [ + { + "bbox": [ + 308, + 522, + 545, + 565 + ], + "type": "text", + "content": "[26] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In CVPR, pages 9726-9735, 2020. 7, 8" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 567, + 545, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 567, + 545, + 610 + ], + "spans": [ + { + "bbox": [ + 308, + 567, + 545, + 610 + ], + "type": "text", + "content": "[27] Christopher. R. Jackson and Werner Alpers. The role of the critical angle in brightness reversals on sunglint images of the sea surface. Journal of Geophysical Research: Oceans, 115(C9), 2010. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 612, + 545, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 612, + 545, + 656 + ], + "spans": [ + { + "bbox": [ + 308, + 612, + 545, + 656 + ], + "type": "text", + "content": "[28] Augustin Juhl, Thomas Farges, and Elisabeth Blanc. Color pictures of sprites from non-dedicated observation on board the international space station. Journal of Geophysical Research: Space Physics, 118(1):454-461, 2013. 2" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 308, + 658, + 545, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 658, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 308, + 658, + 545, + 689 + ], + "type": "text", + "content": "[29] Hervé Jégou, Matthijs Douze, Jorge Sánchez, Patrick Perez, and Cordelia Schmid. Aggregating local image descriptors into compact codes. IEEE TPAMI, 34, 2011. 2" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 308, + 691, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 691, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 691, + 545, + 713 + ], + "type": "text", + "content": "[30] Nikhil Keetha, Avneesh Mishra, Jay Karhade, Krishna Murthy Jatavallabhula, Sebastian Scherer, Madhava" + } + ] + } + ], + "index": 31 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "12762" + } + ] + } + ], + "index": 33 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 66, + 72, + 287, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 72, + 287, + 106 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 287, + 106 + ], + "type": "text", + "content": "Krishna, and Sourav Garg. Anyloc: Towards universal visual place recognition. IEEE Robotics and Automation Letters, 2023. 3, 7" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 106, + 287, + 148 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 106, + 287, + 148 + ], + "spans": [ + { + "bbox": [ + 48, + 106, + 287, + 148 + ], + "type": "text", + "content": "[31] Prannay Khosla, Piotr Teterwak, Chen Wang, Aaron Sarna, Yonglong Tian, Phillip Isola, Aaron Maschinot, Ce Liu, and Dilip Krishnan. Supervised contrastive learning. In NeurIPS, 2020. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 150, + 287, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 150, + 287, + 183 + ], + "spans": [ + { + "bbox": [ + 48, + 150, + 287, + 183 + ], + "type": "text", + "content": "[32] Hyo Jin Kim, Enrique Dunn, and Jan-Michael Frahm. Learned contextual feature reweighting for image geolocation. In CVPR, pages 3251-3260, 2017. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 184, + 287, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 184, + 287, + 205 + ], + "spans": [ + { + "bbox": [ + 48, + 184, + 287, + 205 + ], + "type": "text", + "content": "[33] Diederik Kingma and Jimmy Ba. Adam: A method for stochastic optimization. *ICLR*, 2014. 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 206, + 287, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 206, + 287, + 239 + ], + "spans": [ + { + "bbox": [ + 48, + 206, + 287, + 239 + ], + "type": "text", + "content": "[34] María Leyva-Vallina, Nicola Strisciuglio, and Nicolai Petkov. Data-efficient large scale place recognition with graded similarity supervision. CVPR, 2023. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 239, + 287, + 272 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 239, + 287, + 272 + ], + "spans": [ + { + "bbox": [ + 48, + 239, + 287, + 272 + ], + "type": "text", + "content": "[35] Weiyang Liu, Yandong Wen, Zhiding Yu, and Meng Yang. Large-margin softmax loss for convolutional neural networks. In ICML, pages 507-516. PMLR, 2016. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 272, + 287, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 272, + 287, + 304 + ], + "spans": [ + { + "bbox": [ + 48, + 272, + 287, + 304 + ], + "type": "text", + "content": "[36] Weiyang Liu, Yandong Wen, Zhiding Yu, Ming Li, Bhiksha Raj, and Le Song. Spherface: Deep hypersphere embedding for face recognition. In CVPR, pages 212-220, 2017. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 306, + 287, + 337 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 306, + 287, + 337 + ], + "spans": [ + { + "bbox": [ + 48, + 306, + 287, + 337 + ], + "type": "text", + "content": "[37] David G. Lowe. Distinctive image features from scale-invariant keypoints. Int. J. Comput. Vision, 60(2):91-110, 2004. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 338, + 287, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 338, + 287, + 360 + ], + "spans": [ + { + "bbox": [ + 48, + 338, + 287, + 360 + ], + "type": "text", + "content": "[38] Carlo Masone and Barbara Caputo. A survey on deep visual place recognition. IEEE Access, 9:19516-19547, 2021. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 361, + 287, + 404 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 361, + 287, + 404 + ], + "spans": [ + { + "bbox": [ + 48, + 361, + 287, + 404 + ], + "type": "text", + "content": "[39] Oscar Manas, Alexandre Lacoste, Xavier Giroi-nieto, David Vazquez, and Pau Rodriguez. Seasonal contrast: Unsupervised pre-training from uncurated remote sensing data. In ICCV, pages 9394-9403, 2021. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 405, + 287, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 405, + 287, + 459 + ], + "spans": [ + { + "bbox": [ + 48, + 405, + 287, + 459 + ], + "type": "text", + "content": "[40] Aadil Nathani, Rishi Iyer, Annabelle Wang, Aarnav Chitari, Adele Wilson, and Hannah Norris. Observing Earth From Space: Using Astronaut Photography to Analyze Geographical Climate Patterns. In AGU Fall Meeting Abstracts, pages ED44C-06, 2022. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 460, + 287, + 559 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 460, + 287, + 559 + ], + "spans": [ + { + "bbox": [ + 48, + 460, + 287, + 559 + ], + "type": "text", + "content": "[41] Maxime Oquab, Timothée Darcet, Theo Moutakanni, Huy V. Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, Russell Howes, Po-Yao Huang, Hu Xu, Vasu Sharma, Shang-Wen Li, Wojciech Galuba, Mike Rabbat, Mido Assran, Nicolas Ballas, Gabriel Synnaeve, Ishan Misra, Herve Jegou, Julien Mairal, Patrick Labatut, Armand Joulin, and Piotr Bojanowski. Dinov2: Learning robust visual features without supervision, 2023. 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 559, + 287, + 592 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 559, + 287, + 592 + ], + "spans": [ + { + "bbox": [ + 48, + 559, + 287, + 592 + ], + "type": "text", + "content": "[42] Florent Perronnin, Yan Liu, Jorge Sánchez, and Herve Poirier. Large-scale image retrieval with compressed fisher vectors. In CVPR, pages 3384-3391, 2010. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 593, + 287, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 593, + 287, + 624 + ], + "spans": [ + { + "bbox": [ + 48, + 593, + 287, + 624 + ], + "type": "text", + "content": "[43] F. Radenović, G. Tolias, and O. Chum. Fine-tuning CNN Image Retrieval with No Human Annotation. IEEE TPAMI, 2018. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 625, + 287, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 625, + 287, + 658 + ], + "spans": [ + { + "bbox": [ + 48, + 625, + 287, + 658 + ], + "type": "text", + "content": "[44] A. Razavian, J. Sullivan, A. Maki, and S. Carlsson. Visual Instance Retrieval with Deep Convolutional Networks. CoRR, abs/1412.6574, 2015. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 658, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 287, + 712 + ], + "type": "text", + "content": "[45] Esther Rolf, Jonathan Proctor, Tamma Carleton, Ian Bolliger, Vaishaal Shankar, Miyabi Ishihara, Benjamin Recht, and Solomon Hsiang. A generalizable and accessible approach to machine learning with global satellite imagery. Nature Communications, 12(1), 2021. 3" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 32, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "text", + "content": "[46] Paul-Edouard Sarlin, Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. SuperGlue: Learning feature matching with graph neural networks. In CVPR, 2020. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 107, + 545, + 152 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 107, + 545, + 152 + ], + "spans": [ + { + "bbox": [ + 307, + 107, + 545, + 152 + ], + "type": "text", + "content": "[47] Aaron Schirmer, Caleb Gallemore, Ting Liu, Seth Magle, Elisabeth DiNello, Humerah Ahmed, and Thomas Gilday. Mapping behaviorally relevant light pollution levels to improve urban habitat planning. Scientific Reports, 9, 2019. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 152, + 545, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 152, + 545, + 217 + ], + "spans": [ + { + "bbox": [ + 307, + 152, + 545, + 217 + ], + "type": "text", + "content": "[48] Johannes A. Schultz, Maik Hartmann, Sascha Heinemann, Jens Janke, Carsten Jürgens, Dieter Oertel, Gernot Rücker, Frank Thonfeld, and Andreas Rienow. Diego: A multispectral thermal mission for earth observation on the international space station. European Journal of Remote Sensing, 53(sup2):28-38, 2020. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 219, + 545, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 219, + 545, + 251 + ], + "spans": [ + { + "bbox": [ + 307, + 219, + 545, + 251 + ], + "type": "text", + "content": "[49] Peter Schwind and Tobias Storch. Georeferencing urban nighttime lights imagery using street network maps. Remote Sensing, 14(11), 2022. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 253, + 545, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 253, + 545, + 285 + ], + "spans": [ + { + "bbox": [ + 307, + 253, + 545, + 285 + ], + "type": "text", + "content": "[50] Yujiao Shi and Hongdong Li. Beyond cross-view image retrieval: Highly accurate vehicle localization using satellite image. In CVPR, pages 17010-17020, 2022. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 286, + 545, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 286, + 545, + 319 + ], + "spans": [ + { + "bbox": [ + 307, + 286, + 545, + 319 + ], + "type": "text", + "content": "[51] Yujiao Shi, Xin Yu, Shan Wang, and Hongdong Li. Cvlnet: Cross-view feature correspondence learning for video-based camera localization. In ACCV, pages 652-669, 2022." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 321, + 545, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 321, + 545, + 365 + ], + "spans": [ + { + "bbox": [ + 307, + 321, + 545, + 365 + ], + "type": "text", + "content": "[52] Yujiao Shi, Fei Wu, Akhil Perincherry, Ankit Vora, and Hongdong Li. Boosting 3-dof ground-to-satellite camera localization accuracy via geometry-guided cross-view transformer. In ICCV, pages 21516-21526, 2023. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 365, + 545, + 388 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 365, + 545, + 388 + ], + "spans": [ + { + "bbox": [ + 307, + 365, + 545, + 388 + ], + "type": "text", + "content": "[53] Christopher Small. Spectrometry of the urban lightscape. Technologies, 10(4), 2022. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 388, + 545, + 422 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 388, + 545, + 422 + ], + "spans": [ + { + "bbox": [ + 307, + 388, + 545, + 422 + ], + "type": "text", + "content": "[54] H. Song, Y. Xiang, S. Jegelka, and S. Savarese. Deep metric learning via lifted structured feature embedding. In CVPR, 2016. Spotlight presentation. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 422, + 545, + 476 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 422, + 545, + 476 + ], + "spans": [ + { + "bbox": [ + 307, + 422, + 545, + 476 + ], + "type": "text", + "content": "[55] Weiwei Song, Zhi Gao, Renwei Dian, Pedram Ghamisi, Yongjun Zhang, and Jón Atli Benediktsson. Asymmetric hash code learning for remote sensing image retrieval. IEEE Transactions on Geoscience and Remote Sensing, 60:1-14, 2022. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 478, + 545, + 522 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 478, + 545, + 522 + ], + "spans": [ + { + "bbox": [ + 307, + 478, + 545, + 522 + ], + "type": "text", + "content": "[56] W. L. Stefanov and C. A. Evans. Data collection for disaster response from the international space station. The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, XL-7/W3:851–855, 2015. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 523, + 545, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 523, + 545, + 590 + ], + "spans": [ + { + "bbox": [ + 307, + 523, + 545, + 590 + ], + "type": "text", + "content": "[57] Adam J. Stewart, Caleb Robinson, Isaac A. Corley, Anthony Ortiz, Juan M. Lavista Ferres, and Arindam Banerjee. Torch-Geo: Deep learning with geospatial data. In Proceedings of the 30th International Conference on Advances in Geographic Information Systems, pages 1–12, Seattle, Washington, 2022. Association for Computing Machinery. 7, 8" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 590, + 545, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 590, + 545, + 634 + ], + "spans": [ + { + "bbox": [ + 307, + 590, + 545, + 634 + ], + "type": "text", + "content": "[58] Alex Stoken and Kenton Fisher. Find my astronaut photo: Automated localization and georectification of astronaut photography. In CVPRW, pages 6196-6205, 2023. 2, 3, 4, 5, 7" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 635, + 545, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 635, + 545, + 667 + ], + "spans": [ + { + "bbox": [ + 307, + 635, + 545, + 667 + ], + "type": "text", + "content": "[59] Jiaming Sun, Zehong Shen, Yuang Wang, Hujun Bao, and Xiaowei Zhou. LoFTR: Detector-free local feature matching with transformers. CVPR, 2021. 3" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 307, + 669, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 669, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 307, + 669, + 545, + 712 + ], + "type": "text", + "content": "[60] Yifan Sun, Changmao Cheng, Yuhan Zhang, Chi Zhang, Liang Zheng, Zhongdao Wang, and Yichen Wei. Circle loss: A unified perspective of pair similarity optimization. CVPR, pages 6397-6406, 2020. 3" + } + ] + } + ], + "index": 31 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12763" + } + ] + } + ], + "index": 33 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "type": "text", + "content": "[61] Giorgos Tolias, R. Sicre, and H. Jégou. Particular object retrieval with integral max-pooling of CNN activations. CoRR, abs/1511.05879, 2016. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 107, + 288, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 107, + 288, + 173 + ], + "spans": [ + { + "bbox": [ + 48, + 107, + 288, + 173 + ], + "type": "text", + "content": "[62] Ilya O Tolstikhin, Neil Houlsby, Alexander Kolesnikov, Lucas Beyer, Xiaohua Zhai, Thomas Unterthiner, Jessica Yung, Andreas Steiner, Daniel Keysers, Jakob Uszkoreit, Mario Lucic, and Alexey Dosovitskiy. Mlp-mixer: An all-mlx architecture for vision. In NeurIPS, pages 24261–24272. Curran Associates, Inc., 2021. 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 175, + 287, + 219 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 175, + 287, + 219 + ], + "spans": [ + { + "bbox": [ + 48, + 175, + 287, + 219 + ], + "type": "text", + "content": "[63] Gabriele Trivigno, Gabriele Berton, Juan Aragon, Barbara Caputo, and Carlo Masone. Divide&classify: Fine-grained classification for city-wide visual geo-localization. In ICCV, pages 11142-11152, 2023. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 221, + 287, + 254 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 221, + 287, + 254 + ], + "spans": [ + { + "bbox": [ + 48, + 221, + 287, + 254 + ], + "type": "text", + "content": "[64] Aäron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding. CoRR, abs/1807.03748, 2018. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 256, + 287, + 311 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 256, + 287, + 311 + ], + "spans": [ + { + "bbox": [ + 48, + 256, + 287, + 311 + ], + "type": "text", + "content": "[65] Hao Wang, Yitong Wang, Zheng Zhou, Xing Ji, Dihong Gong, Jingchao Zhou, Zhifeng Li, and Wei Liu. Cosface: Large margin cosine loss for deep face recognition. In CVPR, pages 5265-5274. Computer Vision Foundation / IEEE Computer Society, 2018. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 313, + 287, + 367 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 313, + 287, + 367 + ], + "spans": [ + { + "bbox": [ + 48, + 313, + 287, + 367 + ], + "type": "text", + "content": "[66] Tingyu Wang, Zhedong Zheng, Chenggang Yan, jiyong Zhang, Yaoqi Sun, Bolun Zheng, and Yi Yang. Each part matters: Local patterns facilitate cross-view geo-localization. IEEE TCSVT, 2021. doi: 10.1109/TCSVT.2021.3061265.3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 370, + 287, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 370, + 287, + 413 + ], + "spans": [ + { + "bbox": [ + 48, + 370, + 287, + 413 + ], + "type": "text", + "content": "[67] Xun Wang, Xintong Han, Weilin Huang, Dengke Dong, and Matthew R Scott. Multi-similarity loss with general pair weighting for deep metric learning. In CVPR, pages 5022-5030, 2019. 3, 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 415, + 287, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 415, + 287, + 437 + ], + "spans": [ + { + "bbox": [ + 48, + 415, + 287, + 437 + ], + "type": "text", + "content": "[68] M. Justin Wilkinson and Yanni Gunnell. Fluvial Megafans on Earth and Mars. Cambridge University Press, 2023. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 440, + 287, + 483 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 440, + 287, + 483 + ], + "spans": [ + { + "bbox": [ + 48, + 440, + 287, + 483 + ], + "type": "text", + "content": "[69] Yoav Yair, Melody Korman, Colin Price, and Eytan Stibbe. Observing lightning and transient luminous events from the international space station during ilan-es: An astronaut's perspective. Acta Astronautica, 211:592-599, 2023. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 485, + 287, + 518 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 485, + 287, + 518 + ], + "spans": [ + { + "bbox": [ + 48, + 485, + 287, + 518 + ], + "type": "text", + "content": "[70] Ting Yao, Yiheng Zhang, Zhaofan Qiu, Yingwei Pan, and Tao Mei. Seco: Exploring sequence supervision for unsupervised representation learning. In AAAI, 2021. 7, 8" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 520, + 287, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 520, + 287, + 553 + ], + "spans": [ + { + "bbox": [ + 48, + 520, + 287, + 553 + ], + "type": "text", + "content": "[71] Zhedong Zheng, Yunchao Wei, and Yi Yang. University-1652: A multi-view multi-source benchmark for drone-based geo-localization. ACM MM, 2020. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 555, + 287, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 555, + 287, + 610 + ], + "spans": [ + { + "bbox": [ + 48, + 555, + 287, + 610 + ], + "type": "text", + "content": "[72] Zhedong Zheng, Yu jiao Shi, Tingyu Wang, Jun Liu, Jianwu Fang, Yunchao Wei, and Tat-seng Chua. Uavm '23: 2023 workshop on uavs in multimedia: Capturing the world from a new perspective. In ACM MM, page 9715-9717, New York, NY, USA, 2023. Association for Computing Machinery. 8" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 613, + 287, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 613, + 287, + 678 + ], + "spans": [ + { + "bbox": [ + 48, + 613, + 287, + 678 + ], + "type": "text", + "content": "[73] Runzhe Zhu, Mingze Yang, Kaiyu Zhang, Fei Wu, Ling Yin, and Yujin Zhang. Modern backbone for efficient geolocation. In Proceedings of the 2023 Workshop on UAVs in Multimedia: Capturing the World from a New Perspective, page 31-37, New York, NY, USA, 2023. Association for Computing Machinery. 3, 7, 8" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "type": "text", + "content": "[74] Sijie Zhu, Taojiannan Yang, and Chen Chen. Vigor: Cross-view image geo-localization beyond one-to-one retrieval. In CVPR, pages 3640-3649, 2021. 3" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 72, + 545, + 150 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 307, + 72, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 72, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 307, + 72, + 545, + 106 + ], + "type": "text", + "content": "[75] Sijie Zhu, Mubarak Shah, and Chen Chen. Transgeo: Transformer is all you need for cross-view image geo-localization. In CVPR, pages 1162-1171, 2022. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 107, + 545, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 107, + 545, + 150 + ], + "spans": [ + { + "bbox": [ + 307, + 107, + 545, + 150 + ], + "type": "text", + "content": "[76] Sijie Zhu, Linjie Yang, Chen Chen, Mubarak Shah, Xiaohui Shen, and Heng Wang. R2former: Unified retrieval and reranking transformer for place recognition. In CVPR, pages 19370-19380, 2023. 3" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "12764" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/EasyDrag_ Efficient Point-based Manipulation on Diffusion Models/3c97514b-8a67-4bdc-8795-3e7cfc02ceca_content_list.json b/2024/EasyDrag_ Efficient Point-based Manipulation on Diffusion Models/3c97514b-8a67-4bdc-8795-3e7cfc02ceca_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..1436a03fd8c6fd66034a6197e31914e50dc4de4f --- /dev/null +++ b/2024/EasyDrag_ Efficient Point-based Manipulation on Diffusion Models/3c97514b-8a67-4bdc-8795-3e7cfc02ceca_content_list.json @@ -0,0 +1,1895 @@ +[ + { + "type": "text", + "text": "EasyDrag: Efficient Point-based Manipulation on Diffusion Models", + "text_level": 1, + "bbox": [ + 145, + 130, + 825, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Xingzhong Hou1,2", + "bbox": [ + 86, + 178, + 233, + 198 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Boxiao Liu3", + "bbox": [ + 274, + 180, + 372, + 196 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yi Zhang3", + "bbox": [ + 415, + 180, + 498, + 198 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Jihao Liu3,4", + "bbox": [ + 537, + 180, + 633, + 196 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yu Liu3", + "bbox": [ + 673, + 180, + 736, + 196 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Haihang You1*", + "bbox": [ + 777, + 180, + 897, + 198 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ State Key Lab of Processors, Institute of Computing Technology, Chinese Academy of Sciences $^{2}$ School of Computer Science and Technology, University of Chinese Academy of Sciences $^{3}$ SenseTime Research $^{4}$ CUHK MMLab", + "bbox": [ + 109, + 199, + 872, + 250 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{houxingzhong, youhaihang}@ict.ac.cn {liuboxiao, zhangyi17}@sensetime.com jihaoliu@link.cuhk.edu.hk liuyuisanai@gmail.com", + "bbox": [ + 161, + 253, + 818, + 286 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/be7aa8bb33edfafc4f6d1777fbe28b052deb88c8d41da4902d58fbb7958ebe0c.jpg", + "image_caption": [ + "Figure 1. DragGAN fails in both cases due to limited model capacity. Similarly, DragDiffusion relies on LoRA fine-tuning and hand-drawn masks to achieve high-quality results, and SDE-Drag also fails to address these two cases effectively. In contrast, our approach ensures precise manipulation and detail preservation without fine-tuning or masks." + ], + "image_footnote": [], + "bbox": [ + 120, + 297, + 851, + 455 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 537, + 313, + 553 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Generative models are gaining increasing popularity, and the demand for precisely generating images is on the rise. However, generating an image that perfectly aligns with users' expectations is extremely challenging. The shapes of objects, the poses of animals, the structures of landscapes, and more may not match the user's desires, and this applies to real images as well. This is where point-based image editing becomes essential. An excellent image editing method needs to meet the following criteria: user-friendly interaction, high performance, and good generalization capability. Due to the limitations of StyleGAN, DragGAN exhibits limited robustness across diverse scenarios, while DragDiffusion lacks user-friendliness due to the necessity of LoRA fine-tuning and masks. In this paper, we introduce a novel interactive point-based image editing framework, called EasyDrag, that leverages pretrained diffusion models to achieve high-quality editing outcomes and user-friendship. Extensive experimentation demonstrates that our approach surpasses DragDiffusion in terms of both image quality and editing precision for point-based", + "bbox": [ + 75, + 570, + 472, + 875 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "image manipulation tasks. The code will be available on https://github.com/Ace-Pegasus/EasyDrag.", + "bbox": [ + 500, + 537, + 890, + 569 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 500, + 601, + 632, + 616 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In the past few years, diffusion models [9, 24, 27] represent a groundbreaking advancement in the domain of generative modeling. In particular, text-to-image diffusion models [24, 25, 31] can produce remarkable images based on textual prompts. Recently, such models have also been extended for image editing [3, 7, 13, 18, 23, 33]. Since diffusion models depend on textual input, image elements, including objects and styles, can be altered with a high degree of quality and diversity guided by edited prompts.", + "bbox": [ + 496, + 627, + 890, + 763 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "However, text-guided image editing cannot precisely direct changes in the image, and it is unpredictable for a specific point in the image that one might want to move. Therefore, point-guided image editing becomes essential. A recent study DragGAN [21] introduces \"drag\" operation for precise image editing, which has received widespread attention for its point-based interaction. Nonetheless, it encounters limitations tied to the capacity and adaptability of StyleGAN [11, 12] as shown in Fig. 1. In contrast, diffu", + "bbox": [ + 496, + 763, + 893, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author.", + "bbox": [ + 94, + 886, + 220, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "8404", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "sion models [9, 24, 27] exhibit better stability and superior generative quality. DragDiffusion [26] enables interactive point-based image editing on diffusion models with LoRA fine-tuning [10] but introduces high time consumption at the same time.", + "bbox": [ + 76, + 90, + 468, + 165 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "What capabilities should point-based image editing possess? User-friendly interaction, good performance and effective generalization. User-friendly interaction implies ease of use and short response times for users, while good performance refers to precise dragging with maintaining identity. Effective generalization requires the method to work with a single model for various images. Unfortunately, existing methods have not been able to simultaneously exhibit these capabilities. DragGAN lacks generalization of dragging general domain image, while DragDiffusion is less user-friendly due to its reliance on LoRA fine-tuning and masks. In this paper, we propose a novel point-based image editing framework built on diffusion models, which can concurrently fulfill these requirements.", + "bbox": [ + 75, + 167, + 468, + 378 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In contrast to DragDiffusion, our method does not require any fine-tuning of pre-trained models, which significantly reduces the response time for users. Since the generative domain of diffusion models are extremely huge and lack of LoRA fine-tuning presents a certain degree of challenge for diffusion-based dragging, we introduce a novel motion supervision method to better align with the feature space, which can get precise point tracking on diffusion models. Similarly, the edited latent often generates images that are inconsistent with the original due to the broad generation space. Built on this, we propose reference guidance during the denoising process, to preserve identity consistency with the input image. Furthermore, we design a method for auto mask generation and achieve simplified dragging interaction, which meets the need of user-friendliness. As shown in Fig. 1, the performance of DragDiffusion relies on LoRA fine-tuning and masks, while our method does not require these and achieves better performance.", + "bbox": [ + 75, + 380, + 468, + 651 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our contributions can be summarized as follows:", + "bbox": [ + 94, + 654, + 419, + 667 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- User-friendly interaction. We introduce a point-based image editing method with auto mask generation for user-friendship, simplifying the dragging process while maintaining the comparable results.", + "- High performance. We propose a stable motion supervision and point-tracking method that has better adaptability to the latent space of the diffusion model to achieve precise manipulation. We also introduce reference guidance during the denoising process after the optimization of the latent code for better identity preservation.", + "- Effective Generalization. Based on diffusion models, our approach can handle a wide range of natural images and various types of dragging, obviously outperforming GAN-based methods.", + "- Extensive qualitative and quantitative results demonstrate" + ], + "bbox": [ + 76, + 674, + 467, + 898 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "that our method produces superior and more precise point dragging outcomes with clear gaps when compared to existing point-guided image editing methods.", + "bbox": [ + 511, + 90, + 890, + 136 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related work", + "text_level": 1, + "bbox": [ + 500, + 152, + 638, + 167 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Text-guided image editing", + "text_level": 1, + "bbox": [ + 500, + 176, + 735, + 193 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "With the success of GANs in the field of image generation, many previous image editing methods [1, 6, 29, 34] are based on StyleGAN. These editing approaches map real images into the latent space of StyleGAN ( $\\mathcal{W}$ space or $\\mathcal{W}+$ space) and then achieve image editing by manipulating the latent vectors. However, StyleGAN generates images in a highly restricted domain, and the image quality is not always high. This makes it challenging to successfully invert many real images to a satisfactory latent code. Recently, diffusion models have enabled image synthesis at high quality. An increasing number of image editing tasks [7, 17, 18] have shifted towards being based on diffusion models. Previous image editing methods that rely on diffusion models, such as SDEdit [16] and blended diffusion [2], cannot fully leverage the capabilities of accurate diffusion inversion as they introduced random noise into the input image to create a noisy initial state. Prompt-to-Prompt [7] is the first to accomplish extensive text-guided image editing without the need for diffusion model refinement, including local editing even when a known mask is unavailable. Null-text inversion [18] successfully reconstructs real images by optimizing the null-text embedding at each prediction step, which decides the unconditional prediction. Most recently, pix2pix-zero [22] proposes to learn editing directions in the textual embedding space for specific image translation tasks. In contrast to these methods, our work enables users to conduct image control through point-guided editing.", + "bbox": [ + 496, + 200, + 890, + 609 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Point-guided image editing", + "text_level": 1, + "bbox": [ + 500, + 619, + 743, + 636 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In order to facilitate fine-grained editing, several approaches have been introduced for point-based editing, including those by [5, 21, 26, 30]. DragGAN [21] introduces an interactive method for seamless point-based image editing, incorporating two innovative components: the optimization of latent codes to progressively move multiple handle points to their desired positions and a precise point tracking mechanism to accurately follow the path of these handle points. However, the results of DragGAN are constrained by the generative capacity of StyleGAN. Recently, DragDiffusion [26] extends the editing framework of DragGAN to diffusion models with LoRA fine-tuning. Conversely, LoRA fine-tuning significantly extends the waiting time, which is unfriendly to the users. DragonDiffusion [19] utilizes an energy function to guide the editing and a memory bank for editing consistency. SDE-Drag [20] proposes an SDE-based approach to formulate the image", + "bbox": [ + 496, + 643, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "8405", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/dfa241d1bd2daeebaeba2ded18d81943e2b99102dee3163da5c08de54aeed2cc.jpg", + "image_caption": [ + "Figure 2. The pipeline of our framework. Our method generates masks automatically for precise supervision and convenience for users. To improve motion supervision over existing methods like DragGAN, we propose a stable motion supervision method based on diffusion models. The reference guidance where guidance latents are inherited from DDIM inversion preserves identity with the input image commendably." + ], + "image_footnote": [], + "bbox": [ + 83, + 90, + 890, + 262 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "editing. In contrast to previous works, we propose a novel point-based image editing framework, which can generate high-quality results without LoRA fine-tuning and masks.", + "bbox": [ + 75, + 356, + 468, + 402 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 76, + 414, + 168, + 429 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Preliminaries", + "text_level": 1, + "bbox": [ + 76, + 438, + 217, + 452 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Denoising diffusion probabilistic models (DDPM) [9] aim to map a pure noise $z_{T}$ to an output image $z_{0}$ , which is guided by the given conditioning prompt. During the training process, the network $\\epsilon_{\\theta}$ is updated by predicting the noise $\\epsilon$ from the latent variables $z_{t}$ :", + "bbox": [ + 75, + 460, + 468, + 537 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\theta} = \\mathbb {E} _ {z _ {0}, \\epsilon \\sim N (0, I), t \\sim U (1, T)} \\| \\epsilon - \\epsilon_ {\\theta} \\left(z _ {t}, t, \\mathcal {C}\\right) \\| _ {2} ^ {2}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 116, + 560, + 468, + 579 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where the noised sample $z_{t}$ is calculated by adding noise $\\epsilon$ to $z_{0}$ according to diffusion step $t$ and $\\mathcal{C}$ is the conditioning input of $\\epsilon_{\\theta}$ . During inference, we use DDIM [27] for sampling method:", + "bbox": [ + 75, + 585, + 468, + 646 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nz _ {t - 1} = \\sqrt {\\frac {\\alpha_ {t - 1}}{\\alpha_ {t}}} z _ {t} + \\sqrt {\\alpha_ {t - 1}} \\left(\\sqrt {\\frac {1}{\\alpha_ {t - 1}} - 1} - \\sqrt {\\frac {1}{\\alpha_ {t}} - 1}\\right) \\epsilon_ {\\theta} \\left(z _ {t}\\right), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 76, + 666, + 473, + 718 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\alpha_{t}(t = 0,1,2,\\dots,T)$ represent the noise scales at corresponding diffusion steps.", + "bbox": [ + 76, + 719, + 468, + 750 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "DDIM inversion. The forward process of DDIM can be expressed in terms of $\\epsilon_{\\theta}(z_t,t,\\mathcal{C})$ , based on the assumption that the ODE process can be inverted in the limit of small steps:", + "bbox": [ + 75, + 766, + 468, + 828 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nz _ {t + 1} = \\sqrt {\\frac {\\alpha_ {t + 1}}{\\alpha_ {t}}} z _ {t} + \\sqrt {\\alpha_ {t + 1}} \\left(\\sqrt {\\frac {1}{\\alpha_ {t + 1}} - 1} - \\sqrt {\\frac {1}{\\alpha_ {t}} - 1}\\right) \\epsilon_ {\\theta} \\left(z _ {t}\\right). \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 76, + 847, + 472, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Stable Diffusion. Stable Diffusion (SD) [24] presents the input image into a lower-dimensional latent space with a Variational Auto-Encoder (VAE) [15], where the input latent $z_0$ is defined as $E(x_0)$ and the output image $x_0$ is obtained by $D(z_0)$ . We implement our work based on Stable Diffusion model.", + "bbox": [ + 498, + 356, + 890, + 446 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Classifier-free guidance. The extent to which the prompt influences the reverse diffusion process directly affects the quality of image generation. For this purpose, Ho et al. [8] proposes the classifier-free guidance, which controls the degree of impact from text conditioning. Specifically, the null next-embedding $\\phi = \\psi(\"\")$ is employed as a anchor point for unconditional prediction to enhance the text conditioning with guidance scale $w$ :", + "bbox": [ + 498, + 465, + 892, + 585 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {\\epsilon} _ {\\theta} \\left(z _ {t}, t, \\mathcal {C}, \\phi\\right) = w * \\epsilon_ {\\theta} \\left(z _ {t}, t, \\mathcal {C}\\right) + (1 - w) * \\epsilon_ {\\theta} \\left(z _ {t}, t, \\phi\\right). \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 506, + 611, + 890, + 628 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Overview", + "text_level": 1, + "bbox": [ + 500, + 635, + 609, + 650 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given an input image $\\mathcal{I}$ , the user should input a number of handle points $\\{\\pmb{p}_i = (x_{p,i},y_{p,i})|i = 1,2,\\dots,N\\}$ and their corresponding target points $\\{\\pmb{q}_i = (x_{q,i},y_{q,i})|i = 1,2,\\dots,N\\}$ . The objective is to drag the content of the handle points to the target points based on image $\\mathcal{I}$ . The user can also input a mask $\\mathbf{M}$ to specify the modifiable area.", + "bbox": [ + 498, + 659, + 890, + 750 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We operate DDIM inversion with $\\mathcal{C} = \\psi(\\mathcal{P})$ on the input image, where $\\mathcal{P}$ represents the prompt, and get the intermediate results $z_{T}, \\ldots, z_{t}, \\ldots, z_{0}$ , where $z_{t}$ is a specific intermediate latent in DDIM inversion. Loop of optimizations is performed on $z_{t}$ to make features of the target points similar enough to that of the handle points, with novel stable motion supervision and automatic mask generation introduced in this paper. Then, the result of optimization $z_{t}^{*}$ is fed into the denoising process with reference guidance to achieve identity consistency with $\\mathcal{I}$ . We present the pipeline", + "bbox": [ + 496, + 750, + 892, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "8406", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/f3d668311aaa51f0aeb69f1ebc1cff5fbcff5f9f1b3d6e0e1078aa414667f69a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 104, + 116, + 197, + 184 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/d54a159f4c6ab1630366a150fa66ea2cf5a8e05e347dc5319436e1dd432abfd9.jpg", + "image_caption": [ + "Stable motion supervision" + ], + "image_footnote": [], + "bbox": [ + 215, + 93, + 305, + 185 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/639f2fb7e9195743f74fc8061af117ca9c0b4808c8d151b2610161b73c640d0c.jpg", + "image_caption": [ + "in Fig. 2, and elaborate three key components in the following subsections." + ], + "image_footnote": [], + "bbox": [ + 326, + 93, + 418, + 184 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/ed9925cab50f09aa9096fa4664bfcd4564a964acad549fadab34158ec8c1fe16.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 429, + 119, + 455, + 147 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/6a69fc3d968cdbec1eadcbf9175becf4f5bd554278de73de8b095f8986cc8d68.jpg", + "image_caption": [ + "DragGAN motion supervision" + ], + "image_footnote": [], + "bbox": [ + 96, + 199, + 194, + 266 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/e19e33f2a65b13d7178bec8d57fd56072e6d2beb6ae6c05d0f162d6608140ba8.jpg", + "image_caption": [ + "Figure 3. Stable motion supervision. When the next point $p^{**}$ is more similar with the handle point $p$ than $p^*$ , the handle point has moved to the position of $p^{**}$ . Our approach differs significantly from DragGAN in that our method always learns features from the original image, while DragGAN learns features that are continually changing." + ], + "image_footnote": [], + "bbox": [ + 217, + 199, + 305, + 267 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/905e240854c90ac9f5e223705c101966b6b722bc645151f810788c05afaf3516.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 328, + 199, + 418, + 267 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/fc42870db319f82e2d0272956996e6e0b1b894a3cd2c89fea997bcaac6a83168.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 336, + 273, + 411, + 286 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Stable motion supervision", + "text_level": 1, + "bbox": [ + 76, + 462, + 310, + 478 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "There exists misleading in DragGAN due to the proposed motion supervision. In contrast to DragGAN, we introduce an efficient drag method that is more suitable for diffusion models. Inspired by DIFT [28], we utilize the feature maps of U-Net denoiser as the semantic information of the input image. Specially, we consider the second and third upper layers of U-Net with $(z_{t}, t, \\mathcal{C})$ as inputs, and concatenate the two layers as our feature maps $\\mathbf{F}$ used for stable motion supervision. $\\mathbf{F}$ is resized to have the same resolution as the input image in order to get precise manipulation.", + "bbox": [ + 75, + 487, + 468, + 638 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As shown in Fig. 3, $p^*$ represents current position of the handle point, and $p^{**}$ is the next position that we want to drag to. The method for calculating $p^{**}$ is presented in Algorithm 1. We start from $p^*$ , move along the direction of $q$ by $K$ units, and then find the nearest integer coordinates to determine $p^{**}$ . In contrast to DragGAN, we employ $p^*$ to learn features of the original image at the $p$ position, ensuring that the learned features are consistently accurate. Inspired by DragonDiffusion [19], we define the similarity loss function of two feature maps as follows:", + "bbox": [ + 75, + 640, + 468, + 790 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {c o s}} \\left(\\mathbf {F}, \\mathbf {F} _ {0}\\right) = \\frac {1}{1 + \\frac {1 + \\cos \\left(\\mathbf {F} , s g \\left(\\mathbf {F} _ {0}\\right)\\right)}{2}}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 153, + 808, + 468, + 844 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathbf{F}_0$ represents the feature maps corresponding to the input image and $sg(\\cdot)$ is the stop gradient operator. Therefore, the stable motion supervision loss is defined as:", + "bbox": [ + 76, + 854, + 468, + 898 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Algorithm 1: Next point", + "text_level": 1, + "bbox": [ + 509, + 95, + 676, + 109 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Input: Current point $p^*$ , Target point $q$", + "bbox": [ + 519, + 112, + 779, + 126 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Output: Next point $p^{**}$", + "bbox": [ + 521, + 127, + 679, + 142 + ], + "page_idx": 3 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1 if $\\| \\pmb {q} - \\pmb{p}^{*}\\|_{2} < K$ then", + "2 $\\pmb{p}^{**}\\gets \\pmb {q};$", + "3 else", + "4 $\\pmb{d} \\gets \\frac{\\pmb{q} - \\pmb{p}^{*}}{\\|\\pmb{q} - \\pmb{p}^{*}\\|_{2}};$", + "5 $\\pmb{p}^{**}\\gets \\pmb{p}^* +round(\\pmb {d}*\\pmb {K})$", + "6 end", + "7 return $p^{**}$" + ], + "bbox": [ + 508, + 147, + 730, + 260 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} = \\sum_ {i = 1} ^ {N} \\mathcal {L} _ {\\text {c o s}} \\left(\\mathbf {F} \\left(\\boldsymbol {p} _ {i} ^ {* *}\\right), \\mathbf {F} _ {0} \\left(\\boldsymbol {p} _ {i}\\right)\\right) \\tag {6} \\\\ + \\lambda \\mathcal {L} _ {\\cos} (\\mathbf {F} \\odot (1 - \\mathbf {M}), \\mathbf {F} _ {0} \\odot (1 - \\mathbf {M})), \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 555, + 304, + 890, + 364 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathbf{F}(\\pmb{p}_i^{**})$ denotes the feature values of $\\mathbf{F}$ at pixel $\\pmb{p}_i^{**}$ .", + "bbox": [ + 498, + 369, + 877, + 386 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "After motion supervision, the updated latent code $z_{t}^{*}$ may change the position of handle points. Consequently, we need to determine whether the handle point has reached the next target point. To do this, we compare which one, $\\mathbf{F}(\\pmb{p}_i^{**})$ or $\\mathbf{F}(\\pmb{p}_i^*)$ , resembles $\\mathbf{F}_0(\\pmb{p}_i)$ more. If $\\cos(\\mathbf{F}(\\pmb{p}_i^{**}), \\mathbf{F}_0(\\pmb{p}_i))$ is higher, the handle point $\\pmb{p}_i^*$ is updated to $\\pmb{p}_i^{**}$ and $\\pmb{p}_i^{**}$ is further updated to $\\text{next}(\\pmb{p}_i^*, \\pmb{q}_i)$ according to Algorithm 1.", + "bbox": [ + 496, + 386, + 890, + 492 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.4. Auto mask generation", + "text_level": 1, + "bbox": [ + 500, + 500, + 704, + 515 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For user-friendly interaction, image editing should avoid burdening users with the task of drawing masks. Existing drag methods based on diffusion models [19, 26] always rely on mask inputs, which are provided by users. This requires users to have a deep understanding of the drag method because an incorrectly drawn mask can lead to unpredictable outcomes. Consequently, we propose a method that can automatically generate masks, which are able to change adaptively during the training process.", + "bbox": [ + 496, + 523, + 890, + 657 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In the process of motion supervision, the parts with larger gradients of $z_{t}^{*}$ have a greater impact on the results, while those with smaller gradients are generally the parts that are expected to remain stationary. However, if the regions with relatively small gradients keep changing, the final result becomes unpredictable. Due to the self-attention layer in U-Net, minor changes made at step $t$ are very likely to have an unacceptable impact on the final image. Therefore, we still need control over these low-impact areas. We achieve this by using regions with normalized gradients greater than a threshold $g$ as the drag mask $\\mathbf{M}$ .", + "bbox": [ + 496, + 659, + 890, + 825 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Every time $\\pmb{p}_i^*$ is updated, the hot-zone of $z_t$ also changes due to the position shifting of the drag point. We update our dynamic mask $\\mathbf{M}$ each time $\\pmb{p}_i^*$ changes, but there is no need to change the mask when $\\pmb{p}_i^*$ remains the same. This is because for the same $\\pmb{p}_i^*$ and $\\pmb{p}_i^{**}$ , the hot-zone is consistent.", + "bbox": [ + 496, + 825, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "8407", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/1076bdb83fba895d1d28cc783296bd895605188cbdede29e7e6e1f0d5649c4e0.jpg", + "image_caption": [ + "(a)", + "Figure 4. Qualitative comparison of our approach to DragDiffusion. Due to the misleading of motion supervision, DragDiffusion always cannot move the handle points to the target. Furthermore, DragDiffusion often exhibits significant background changes in the absence of the mask. In our method, the handle points reach the target positions precisely even without the mask." + ], + "image_footnote": [], + "bbox": [ + 127, + 109, + 450, + 386 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/fecf87bf8c74650247a2579d8e28cd4520b7544482a98fd6e4ace7b433ab62a4.jpg", + "image_caption": [ + "(b)" + ], + "image_footnote": [], + "bbox": [ + 460, + 109, + 849, + 385 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "This way, $\\mathbf{M}$ adapts to changes of $z_{t}$ , aligning better with the dragging requirements.", + "bbox": [ + 75, + 472, + 468, + 502 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.5. Reference guidance", + "text_level": 1, + "bbox": [ + 76, + 510, + 264, + 526 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "With optimized $z_{i}^{*}$ , the generated images with direct DDIM denoising often exhibit identity changes and a decrease in image quality. Especially in areas where the pretrained models have limited generation capabilities, naively generating with $z_{i}$ after DDIM Inversion may not even reproduce the original image. Drawing inspiration from [4, 19, 26], we employ mutual self-attention to ensure that images generated from $z_{i}^{*}$ maintain content consistency with the original image.", + "bbox": [ + 75, + 532, + 468, + 669 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As illustrated in Fig. 2, we replace the key and value of self-attention layers in backward process of $z_{i}^{*}$ with the corresponding key and value in the process of reconstructing the source image $\\mathcal{I}$ . Due to the feature correlations in lower layers being relatively weak, the mutual self-attention is only applied to the upper layers of U-Net $\\epsilon_{\\theta}$ . We then perform sampling using the linear combination of the reference guidance and normal prediction:", + "bbox": [ + 75, + 670, + 470, + 791 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\epsilon_ {i} ^ {*} \\leftarrow w \\epsilon_ {\\theta} \\left(z _ {i} ^ {*}, z _ {i}, i, \\mathcal {C}\\right) + (1 - w) \\epsilon_ {\\theta} \\left(z _ {i} ^ {*}, i, \\mathcal {C}\\right), \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 122, + 815, + 468, + 832 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $w$ controls the strength of the reference guidance. Specifically, $z_{i}$ is the corresponding latent during DDIM inversion, rather than that generated from $z_{t}$ . When $w = 1$ , our reference guidance is the same as mutual self-attention.", + "bbox": [ + 75, + 840, + 468, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "However, in the cases where the input image is infrequent for the model and $z_{t}$ undergoes significant changes, using mutual self-attention alone cannot guarantee identity consistency. When we increase the strength of reference guidance, the generated results can better ensure content and texture consistency with the input image.", + "bbox": [ + 496, + 472, + 892, + 564 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 500, + 577, + 633, + 595 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this section, we evaluate the proposed method qualitatively and quantitatively, on both synthetic and real images.", + "bbox": [ + 498, + 603, + 890, + 635 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Implementation details. In our experiments, except for generated image manipulations, all tasks are based on Stable Diffusion 1.5 (SD-1.5). Unless otherwise specified, in both DDIM inversion and sampling we set the number of the sampling steps to be 50 via using the stride of 20 over 1000 diffusion steps. We set the latent from the 35-th DDIM step, which contains rich semantic information corresponding to the input image, as $z_{t}$ . The gradient threshold $g$ is 0.4, the motion step $K$ is 12 and the reference guidance scale $w$ is 4 by default. We use a learning rate of 0.01 for $z_{t}$ without any decay schedule and train the latent with Adam [14] solver.", + "bbox": [ + 496, + 652, + 892, + 834 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Dataset. We employ DragBench introduced in DragDiffusion as our evaluation dataset as it encompasses various types of images, and we also follow its dragging instruc", + "bbox": [ + 498, + 854, + 893, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "8408", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/5600e1d001a83d73f3ccbbaf5344749d3b16e5a066b08eeab0be254b326ecff7.jpg", + "image_caption": [ + "(a)", + "Inputs" + ], + "image_footnote": [], + "bbox": [ + 156, + 113, + 269, + 200 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/94692cfe6867f37db3afe49edefa00e535c70d570a67d5591df9db012e5a34cc.jpg", + "image_caption": [ + "Results" + ], + "image_footnote": [], + "bbox": [ + 272, + 112, + 385, + 200 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/c0cc08965b72012715ec0e2c0b139069777c7d8089f74898e00ab1ecd10202c7.jpg", + "image_caption": [ + "Inputs" + ], + "image_footnote": [], + "bbox": [ + 387, + 112, + 501, + 200 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/f09f5170ed21fe8ef65d342bdc256034f5e9d4a4a100a8680e79d03b2d9d0616.jpg", + "image_caption": [ + "Results" + ], + "image_footnote": [], + "bbox": [ + 503, + 112, + 616, + 200 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/5074d6b041ecf4089806b68f0aefaa98eb48adf0105b68b3fad6f1f945705eff.jpg", + "image_caption": [ + "Inputs" + ], + "image_footnote": [], + "bbox": [ + 620, + 112, + 733, + 200 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/3446da27be0e0bbce2cb970163d8dc19becfa3be5f7d4689d829fde8f713939d.jpg", + "image_caption": [ + "Results" + ], + "image_footnote": [], + "bbox": [ + 736, + 112, + 849, + 200 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/e82cac6e9863c552c63f110feeb9f59fc34480a9f388efde9ff8fea656c7d9ab.jpg", + "image_caption": [ + "(b)" + ], + "image_footnote": [], + "bbox": [ + 156, + 205, + 269, + 292 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/afc20d99245a8cd0a328a8d41f8e7b520a61e499875512cf8fcbbf8905793434.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 272, + 205, + 383, + 292 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/5988b24870a3319ccb873c0a00c4c3542e0fb0be088ad618c0b81cdb429d0788.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 388, + 205, + 501, + 292 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/eabd01accd354e303531c596f04c79c0e24c936c36eee532807369438a159577.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 504, + 205, + 616, + 292 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/bc2a6496f40586416aa881955896a511e1c83c7d1fea9f1d1f8ac2d3865acee9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 620, + 204, + 733, + 292 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/208994933d984a2212dce4e266454ba0d09aa49c087e60b0652ecba0a90a1e77.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 736, + 205, + 849, + 292 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/377c7ab307cea6c20f1bf9095a19fa823139f0e0b6bd924797ee53eae72eaba1.jpg", + "image_caption": [ + "(c)" + ], + "image_footnote": [], + "bbox": [ + 156, + 295, + 269, + 383 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/127c7a07f9e408808d573cf6f08f03ce3ed21872e6aef5b9c6add66f02f3a71f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 272, + 295, + 383, + 383 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/759abe29efba09868916adc3c7f935ae4e339085e06d4d5cdfbfd1785723fbfc.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 388, + 295, + 501, + 383 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/0434fa00e25c242a2976c474351d1f06a5d5fcfd5a701762f7b754a460159ebf.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 504, + 295, + 616, + 383 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/62da21abc77b0febb11056941ab8203c407a84fa658876a306fc125c7cceb8d4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 620, + 295, + 733, + 383 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/757ef7d9241db58cc1cad54919191c5b90f577826ca6ca52bf55d6b867b42a93.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 736, + 295, + 849, + 383 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/f6be91b9b38b006e8d8e42bc5fc7aeca8739f1e8240f04fc0ca1ef35f8da9fb3.jpg", + "image_caption": [ + "(d)", + "Figure 5. Qualitative results of our approach on generated images with SD-1.5 and its fine-tuned variants. (a) SD-1.5. (b) Majicmix Realistic V7. (c) Interior Design Supermix. (d) Dvarch." + ], + "image_footnote": [], + "bbox": [ + 156, + 387, + 269, + 474 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/8646d15d7956bafba38430089a1f6ce4717d9221687f8fb2298de12274fe2f3d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 272, + 387, + 383, + 474 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/91b40614ca1168e551f12fdea0dd87e2c6a3f1ce524c2701bff173db947695c3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 388, + 387, + 501, + 474 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/cb0f7c8c84c5387800ec55ef6fa35fb245fc4927586a4eef2b516291155386fc.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 504, + 387, + 616, + 474 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/5829854f416da08cf0b2343ab41c3cb501926da7c78ff9d3409f2fe80a1802ad.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 620, + 387, + 733, + 474 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/6566911f56d82707fd59c1013b00d85363f7c22a5a0166bb0d5eecf8180ca9e3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 736, + 387, + 849, + 474 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "tions. DragBench contains 205 images, with well-annotated prompts, masks, and dragging points.", + "bbox": [ + 75, + 541, + 468, + 571 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation metrics. We evaluate the dragging results from two aspects: the quality of dragged image and the accuracy of position of handle points. We employ Image Fidelity (IF) [13] and Mean Distance (MD) [21] to assess these two aspects, respectively. IF evaluates the similarity between the input and output images, which is calculated by subtracting the mean LPIPS [32] over all pairs of original and edited images from 1. MD quantifies how well the handle points move to the target points. The same as DragDiffusion, we utilize DIFT [28] to identify the points in the output image that correspond to the handle points in the input image. These identified points are then treated as the definitive handle points after the dragging process. Subsequently, MD is calculated as the average Euclidean distance between the positions of all target points and their corresponding final handle points. A higher IF and a lower MD indicate a better quality of the dragging approach.", + "bbox": [ + 75, + 643, + 470, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Qualitative evaluation", + "text_level": 1, + "bbox": [ + 500, + 540, + 705, + 556 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We conduct extensive experiments for the qualitative comparison between our method and DragDiffusion. To provide a thorough evaluation, we adopt SD-1.5 [24] as our base model for all real images, and also employ other fine-tuned variants of SD-1.5 for generated image manipulation.", + "bbox": [ + 498, + 565, + 890, + 641 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We present real image editing results for various object categories and user inputs in Fig. 4. Our approach accurately moves the handle points to reach the target positions, resulting in a wide range of natural and diverse manipulation effects such as altering the poses of animals, rearranging landscapes, modifying human body parts, and reshaping objects, in the absence of the mask. Conversely, DragDiffusion is unable to precisely move the handle points to the target positions, which is mainly attributed to the misleading of point tracking that is employed in DragGAN and DragDiffusion. It is worth noting that DragDiffusion always changes the non-dragged regions when no mask is provided.", + "bbox": [ + 496, + 642, + 892, + 824 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We also conduct a series of point manipulations based on the generated images, as shown in Fig. 5. The classifier-free guidance scale is configured at 7.5 for various SD-1.5 variants. Notably, the reference guidance is unnecessary for generative editing, as the prompt of generated images pro", + "bbox": [ + 496, + 824, + 893, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "8409", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/f07b53341e22f2521951b6396f31e03c7b5aae1818e914cc080cdfa8dc9d40ac.jpg", + "image_caption": [ + "Figure 6. Out-of-distribution manipulations. In some cases, the input image cannot be reconstructed with DDIM Inversion. Utilizing reference guidance, edited image preserves its identity with the original image." + ], + "image_footnote": [], + "bbox": [ + 161, + 89, + 812, + 244 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/12a617a0c7a404eb78c076562a59c7e946c4e8a663cf57c3c3a27903f7c5f91f.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodIF(↑)MD(↓)Time(s)
DragGAN0.69557.359.1
DragDiffusion0.88236.7169.4
DragDiffusion w/o LoRA0.86353.6333.7
DragDiffusion w/o mask0.76647.0365.6
SDE-Drag0.87145.4450.1
Ours w/ mask0.88929.0727.3
Ours0.91032.5530.6
", + "bbox": [ + 81, + 306, + 467, + 435 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 1. Quantitative evaluation of different methods with Drag-Bench. Evaluation metrics include Image Fidelity (IF) and Mean Distance (MD).", + "bbox": [ + 75, + 444, + 470, + 486 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "vides substantial guidance during the denoising process. As shown in Fig. 5, our method achieves point-based editing with high quality and precision on generated images. It can be observed that in some cases, the details of generated images have changed. This is because the self-attention layer in diffusion models may introduce slight variation in other parts when a specific area of the image is edited. If users want to preserve certain areas without any changes, they can employ inpainting during image generation with masks to determine which regions should remain unchanged.", + "bbox": [ + 75, + 520, + 468, + 672 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. Quantitative evaluation", + "text_level": 1, + "bbox": [ + 76, + 691, + 294, + 708 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We conduct quantitative evaluation by comparing our method to the baseline dragging method on diffusion, DragDiffusion. The evaluation, which incorporates IF and MD metrics, makes use of the DragBench dataset provided by DragDiffusion. The results, as shown in Tab. 1, demonstrate that our approach significantly surpasses DragDiffusion and SDE-Drag on the DragBench dataset. Notably, our method achieves superior dragging results without requiring LoRA fine-tuning, offering substantial time savings during extensive experiments. The performance of DragDiffusion degrades significantly without LoRA nor mask. The execution time in Tab. 1 is calculated on a single A100 GPU.", + "bbox": [ + 75, + 719, + 470, + 900 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/fc6519bc1e51dec6c62cc45781aa2504cdd8c46afda07d63298c44221fff5efd.jpg", + "image_caption": [ + "Figure 7. When users do not provide an appropriate mask, the generated results are prone to issues." + ], + "image_footnote": [], + "bbox": [ + 501, + 311, + 890, + 474 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Discussions", + "text_level": 1, + "bbox": [ + 500, + 541, + 624, + 555 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Reference Guidance. With our reference guidance, we enable the diffusion models to perform out-of-distribution manipulations. As shown in Fig. 6, the original image cannot be faithfully reconstructed using DDIM Inversion. When we introduce reference guidance, $z_{t}$ can better reconstruct the input image. It is worth noting that $z_{i}(i = t,t - 1,\\dots,1)$ is obtained during the DDIM Inversion process, so it contains richer information compared to $z_{t}$ . After editing $z_{t}$ , the image generated from $z_{t}^{*}$ is very different from the input image in terms of identity. As we increase the reference guidance scale $w$ , the consistency between the generated image and original image gradually improves. However, if $w$ is too large, the generated image may become oversaturated. In our experiments, a value for $w$ in the range of 3 to 5 is considered a good choice.", + "bbox": [ + 496, + 564, + 890, + 790 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Auto mask generation. When users create imprecise masks, the results are unpredictable. As shown in Fig. 7, when the mask is too small, it will restrict the editing of the image. On the other hand, when the mask is too large, changes may occur in the background. Therefore, an appropriate mask is necessary for point-based image editing. Our", + "bbox": [ + 496, + 810, + 892, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "8410", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/67076ba86c73d6678f9264f69f5a21174c6b306f5836f284cd78a169a1750322.jpg", + "image_caption": [ + "Figure 8. The changes in the gradient and prediction of $z_{t}$ during optimization. The term of \"K-th step\" means the K-th update of handle points $p^{*}$ ." + ], + "image_footnote": [], + "bbox": [ + 83, + 89, + 468, + 247 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/7ca4e4504b8eb48648883887a7a59122d81e3212d6946843369e789a264b6911.jpg", + "image_caption": [ + "(a) Which upper layer to use.", + "Figure 9. Ablation study." + ], + "image_footnote": [], + "bbox": [ + 81, + 321, + 261, + 425 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/c8a4132df9a3745e3f02c9e7d0aa42f6cb2d6e0e82b1c865b92aa648ef6d3cc2.jpg", + "image_caption": [ + "(b) Effects of K." + ], + "image_footnote": [], + "bbox": [ + 279, + 318, + 464, + 425 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "auto mask generation method liberates users from the need to manually draw masks, which is user-friendly, especially for novice users.", + "bbox": [ + 75, + 498, + 468, + 542 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Figure 8 shows the update of gradient and prediction with the optimization of $z_{t}$ . It can be observed that when the handle points update, the gradient of $z_{t}$ will change simultaneously. In the K-th step, the whole arm of the robot needs to be lifted, so the gradients for the arm part are relatively large. However, by the time we reach the (K+1)-th step, the robot's arm has already been dragged into the right place, and only the palm part needs to extend outward. Therefore, only the gradients for the palm part are high. Our auto mask effectively captures these details and incorporates this variation into the updates of $z_{t}$ , thereby providing better control over the direction of $z_{t}$ . That is the reason why the IF metric of without mask is higher than that with mask in Tab. 1.", + "bbox": [ + 75, + 545, + 468, + 741 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Ablation study. We perform ablation studies to analyze the impact of the feature used in stable motion supervision, and we evaluate the performance on DragBench, as shown in Fig. 9a. Notably, the concatenated feature maps from upper layers 1&2 exhibit the best performance, demonstrating an optimal balance between resolution and discriminativeness. Additionally, we investigate the effects of different values of $K$ in Fig. 9b, and it is observed that $K = 12$ yields better results.", + "bbox": [ + 75, + 763, + 468, + 902 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/d8b7c323fd0a5192187cf97328708488c32ee7df2a1c891587e1500a33ceb083.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodsIF(↑)MD(↓)
DragGAN0.85651.32
DragonDiffusion0.89952.21
Stable motion supervision0.88929.07
", + "bbox": [ + 539, + 88, + 854, + 152 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 2. Quantitative comparison of motion supervision between our method and previous works. Our approach achieves more accurate dragging results than other methods.", + "bbox": [ + 498, + 162, + 890, + 205 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/20d93ef55d4de00d352344e33e002f8dce81dca5326e5fc65bcfc337468cc134.jpg", + "image_caption": [ + "Figure 10. When the manipulated portion in the image is very small or the dragging distance is extensive, it is challenging for the dragging to have a noticeable effect." + ], + "image_footnote": [], + "bbox": [ + 519, + 220, + 870, + 301 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "A comparison of \"drag\" method between our approach and previous methods is shown in Tab. 2. We only replace stable motion supervision with the \"drag\" method in other works. It can be found that stable motion supervision is more suitable for the feature space of diffusion models, although DragonDiffusion achieves slightly higher IF.", + "bbox": [ + 498, + 381, + 890, + 472 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Limitations. Despite the outstanding performance, our method has some limitations. Particularly, when the editing area is too small or the dragging distance is too long, the dragging result may be unsatisfactory. As depicted in Fig. 10, when the girl's face occupies only a small part of the image, and we intend to edit the mouth, the desired changes of mouth do not occur as expected.", + "bbox": [ + 498, + 492, + 890, + 599 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 612, + 617, + 628 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this work, we attribute an excellent point-based image editing method to three criteria: user-friendly interaction, good performance, and effective generalization. We abandon the LoRA fine-tuning and masks, which is convenient for users. Furthermore, we propose a stable motion supervision method and reference guidance to generate high-quality results based on diffusion models. Leveraging the powerful generative capabilities of the diffusion model, we achieve impressive dragging results across various domains. However, in our experiments, we observe that for very small-scale manipulations or long-distance dragging in images, the results are often not satisfactory. Diffusion models with stronger generative capabilities may be needed to achieve a broader range of dragging manipulations.", + "bbox": [ + 496, + 638, + 890, + 849 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements: This work is partially supported by Natural Science Foundation of China Grant 41930110.", + "bbox": [ + 500, + 869, + 890, + 898 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8411", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 174, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Rameen Abdal, Peihao Zhu, Niloy J Mitra, and Peter Wonka. Styleflow: Attribute-conditioned exploration of stylegan-generated images using conditional continuous normalizing flows. ACM Transactions on Graphics (ToG), 40(3):1-21, 2021. 2", + "[2] Omri Avrahami, Dani Lischinski, and Ohad Fried. Blended diffusion for text-driven editing of natural images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18208-18218, 2022. 2", + "[3] Tim Brooks, Aleksander Holynski, and Alexei A Efros. Instructpix2pix: Learning to follow image editing instructions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18392-18402, 2023. 1", + "[4] Mingdeng Cao, Xintao Wang, Zhongang Qi, Ying Shan, Xiaohu Qie, and Yinqiang Zheng. Masactrl: Tuning-free mutual self-attention control for consistent image synthesis and editing. arXiv preprint arXiv:2304.08465, 2023. 5", + "[5] Yuki Endo. User-controllable latent transformer for stylegan image layout editing. In Computer Graphics Forum, pages 395-406. Wiley Online Library, 2022. 2", + "[6] Dave Epstein, Taesung Park, Richard Zhang, Eli Shechtman, and Alexei A Efros. Blobgan: Spatially disentangled scene representations. In European Conference on Computer Vision, pages 616-635. Springer, 2022. 2", + "[7] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-prompt image editing with cross attention control. arXiv preprint arXiv:2208.01626, 2022. 1, 2", + "[8] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598, 2022. 3", + "[9] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 1, 2, 3", + "[10] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 2", + "[11] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4401-4410, 2019. 1", + "[12] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of stylegan. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8110-8119, 2020. 1", + "[13] Bahjat Kawar, Shiran Zada, Oran Lang, Omer Tov, Huiwen Chang, Tali Dekel, Inbar Mosseri, and Michal Irani. Imagic: Text-based real image editing with diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6007-6017, 2023. 1, 6", + "[14] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014.5" + ], + "bbox": [ + 78, + 114, + 470, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[15] Diederik P Kingma and Max Welling. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114, 2013. 3", + "[16] Chenlin Meng, Yutong He, Yang Song, Jiaming Song, Jia-jun Wu, Jun-Yan Zhu, and Stefano Ermon. Sdedit: Guided image synthesis and editing with stochastic differential equations. arXiv preprint arXiv:2108.01073, 2021. 2", + "[17] Daiki Miyake, Akihiro Iohara, Yu Saito, and Toshiyuki Tanaka. Negative-prompt inversion: Fast image inversion for editing with text-guided diffusion models. arXiv preprint arXiv:2305.16807, 2023. 2", + "[18] Ron Mokady, Amir Hertz, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Null-text inversion for editing real images using guided diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6038–6047, 2023. 1, 2", + "[19] Chong Mou, Xintao Wang, Jiechong Song, Ying Shan, and Jian Zhang. Dragondiffusion: Enabling drag-style manipulation on diffusion models. arXiv preprint arXiv:2307.02421, 2023. 2, 4, 5", + "[20] Shen Nie, Hanzhong Allan Guo, Cheng Lu, Yuhao Zhou, Chenyu Zheng, and Chongxuan Li. The blessing of randomness: Sde beats ode in general diffusion-based image editing. arXiv preprint arXiv:2311.01410, 2023. 2", + "[21] Xingang Pan, Ayush Tewari, Thomas Leimkuhler, Lingjie Liu, Abhimitra Meka, and Christian Theobalt. Drag your gan: Interactive point-based manipulation on the generative image manifold. In ACM SIGGRAPH 2023 Conference Proceedings, pages 1-11, 2023. 1, 2, 6", + "[22] Gaurav Parmar, Krishna Kumar Singh, Richard Zhang, Yijun Li, Jingwan Lu, and Jun-Yan Zhu. Zero-shot image-to-image translation. In ACM SIGGRAPH 2023 Conference Proceedings, pages 1-11, 2023. 2", + "[23] Daniel Roich, Ron Mokady, Amit H Bermano, and Daniel Cohen-Or. Pivotal tuning for latent-based editing of real images. ACM Transactions on graphics (TOG), 42(1):1-13, 2022. 1", + "[24] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10684-10695, 2022. 1, 2, 3, 6", + "[25] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in Neural Information Processing Systems, 35:36479-36494, 2022. 1", + "[26] Yujun Shi, Chuhui Xue, Jiachun Pan, Wenqing Zhang, Vincent YF Tan, and Song Bai. Dragdiffusion: Harnessing diffusion models for interactive point-based image editing. arXiv preprint arXiv:2306.14435, 2023. 2, 4, 5", + "[27] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 1, 2, 3", + "[28] Luming Tang, Menglin Jia, Qianqian Wang, Cheng Perng Phoo, and Bharath Hariharan. Emergent correspondence from image diffusion. arXiv preprint arXiv:2306.03881, 2023.4,6" + ], + "bbox": [ + 501, + 92, + 893, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "8412", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[29] Jianyuan Wang, Ceyuan Yang, Yinghao Xu, Yujun Shen, Hongdong Li, and Bolei Zhou. Improving gan equilibrium by raising spatial awareness. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11285-11293, 2022. 2", + "[30] Sheng-Yu Wang, David Bau, and Jun-Yan Zhu. Rewriting geometric rules of a gan. ACM Transactions on Graphics (TOG), 41(4):1-16, 2022. 2", + "[31] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3836-3847, 2023. 1", + "[32] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 6", + "[33] Zhixing Zhang, Ligong Han, Arnab Ghosh, Dimitris N Metaxas, and Jian Ren. Sine: Single image editing with text-to-image diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6027-6037, 2023. 1", + "[34] Jiapeng Zhu, Ceyuan Yang, Yujun Shen, Zifan Shi, Bo Dai, Deli Zhao, and Qifeng Chen. Linkgan: Linking gan latents to pixels for controllable image synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7656-7666, 2023. 2" + ], + "bbox": [ + 78, + 90, + 468, + 470 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "8413", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2024/EasyDrag_ Efficient Point-based Manipulation on Diffusion Models/3c97514b-8a67-4bdc-8795-3e7cfc02ceca_model.json b/2024/EasyDrag_ Efficient Point-based Manipulation on Diffusion Models/3c97514b-8a67-4bdc-8795-3e7cfc02ceca_model.json new file mode 100644 index 0000000000000000000000000000000000000000..9d088800535b04b87748b419fcf3b076135abbdb --- /dev/null +++ b/2024/EasyDrag_ Efficient Point-based Manipulation on Diffusion Models/3c97514b-8a67-4bdc-8795-3e7cfc02ceca_model.json @@ -0,0 +1,2486 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.146, + 0.131, + 0.826, + 0.154 + ], + "angle": 0, + "content": "EasyDrag: Efficient Point-based Manipulation on Diffusion Models" + }, + { + "type": "text", + "bbox": [ + 0.088, + 0.179, + 0.235, + 0.199 + ], + "angle": 0, + "content": "Xingzhong Hou1,2" + }, + { + "type": "text", + "bbox": [ + 0.275, + 0.181, + 0.374, + 0.198 + ], + "angle": 0, + "content": "Boxiao Liu3" + }, + { + "type": "text", + "bbox": [ + 0.416, + 0.181, + 0.499, + 0.199 + ], + "angle": 0, + "content": "Yi Zhang3" + }, + { + "type": "text", + "bbox": [ + 0.539, + 0.181, + 0.634, + 0.198 + ], + "angle": 0, + "content": "Jihao Liu3,4" + }, + { + "type": "text", + "bbox": [ + 0.674, + 0.181, + 0.738, + 0.198 + ], + "angle": 0, + "content": "Yu Liu3" + }, + { + "type": "text", + "bbox": [ + 0.779, + 0.181, + 0.898, + 0.199 + ], + "angle": 0, + "content": "Haihang You1*" + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.2, + 0.873, + 0.251 + ], + "angle": 0, + "content": "\\(^{1}\\)State Key Lab of Processors, Institute of Computing Technology, Chinese Academy of Sciences \\(^{2}\\)School of Computer Science and Technology, University of Chinese Academy of Sciences \\(^{3}\\)SenseTime Research \\(^{4}\\)CUHK MMLab" + }, + { + "type": "text", + "bbox": [ + 0.162, + 0.255, + 0.82, + 0.287 + ], + "angle": 0, + "content": "{houxingzhong, youhaihang}@ict.ac.cn {liuboxiao, zhangyi17}@sensetime.com jihaoliu@link.cuhk.edu.hk liuyuisanai@gmail.com" + }, + { + "type": "image", + "bbox": [ + 0.122, + 0.298, + 0.852, + 0.457 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.467, + 0.895, + 0.511 + ], + "angle": 0, + "content": "Figure 1. DragGAN fails in both cases due to limited model capacity. Similarly, DragDiffusion relies on LoRA fine-tuning and hand-drawn masks to achieve high-quality results, and SDE-Drag also fails to address these two cases effectively. In contrast, our approach ensures precise manipulation and detail preservation without fine-tuning or masks." + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.538, + 0.314, + 0.554 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.571, + 0.473, + 0.875 + ], + "angle": 0, + "content": "Generative models are gaining increasing popularity, and the demand for precisely generating images is on the rise. However, generating an image that perfectly aligns with users' expectations is extremely challenging. The shapes of objects, the poses of animals, the structures of landscapes, and more may not match the user's desires, and this applies to real images as well. This is where point-based image editing becomes essential. An excellent image editing method needs to meet the following criteria: user-friendly interaction, high performance, and good generalization capability. Due to the limitations of StyleGAN, DragGAN exhibits limited robustness across diverse scenarios, while DragDiffusion lacks user-friendliness due to the necessity of LoRA fine-tuning and masks. In this paper, we introduce a novel interactive point-based image editing framework, called EasyDrag, that leverages pretrained diffusion models to achieve high-quality editing outcomes and user-friendship. Extensive experimentation demonstrates that our approach surpasses DragDiffusion in terms of both image quality and editing precision for point-based" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.539, + 0.892, + 0.57 + ], + "angle": 0, + "content": "image manipulation tasks. The code will be available on https://github.com/Ace-Pegasus/EasyDrag." + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.602, + 0.633, + 0.617 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.628, + 0.892, + 0.765 + ], + "angle": 0, + "content": "In the past few years, diffusion models [9, 24, 27] represent a groundbreaking advancement in the domain of generative modeling. In particular, text-to-image diffusion models [24, 25, 31] can produce remarkable images based on textual prompts. Recently, such models have also been extended for image editing [3, 7, 13, 18, 23, 33]. Since diffusion models depend on textual input, image elements, including objects and styles, can be altered with a high degree of quality and diversity guided by edited prompts." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.895, + 0.903 + ], + "angle": 0, + "content": "However, text-guided image editing cannot precisely direct changes in the image, and it is unpredictable for a specific point in the image that one might want to move. Therefore, point-guided image editing becomes essential. A recent study DragGAN [21] introduces \"drag\" operation for precise image editing, which has received widespread attention for its point-based interaction. Nonetheless, it encounters limitations tied to the capacity and adaptability of StyleGAN [11, 12] as shown in Fig. 1. In contrast, diffu" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.887, + 0.222, + 0.9 + ], + "angle": 0, + "content": "*Corresponding author." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8404" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.47, + 0.166 + ], + "angle": 0, + "content": "sion models [9, 24, 27] exhibit better stability and superior generative quality. DragDiffusion [26] enables interactive point-based image editing on diffusion models with LoRA fine-tuning [10] but introduces high time consumption at the same time." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.169, + 0.47, + 0.379 + ], + "angle": 0, + "content": "What capabilities should point-based image editing possess? User-friendly interaction, good performance and effective generalization. User-friendly interaction implies ease of use and short response times for users, while good performance refers to precise dragging with maintaining identity. Effective generalization requires the method to work with a single model for various images. Unfortunately, existing methods have not been able to simultaneously exhibit these capabilities. DragGAN lacks generalization of dragging general domain image, while DragDiffusion is less user-friendly due to its reliance on LoRA fine-tuning and masks. In this paper, we propose a novel point-based image editing framework built on diffusion models, which can concurrently fulfill these requirements." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.381, + 0.47, + 0.652 + ], + "angle": 0, + "content": "In contrast to DragDiffusion, our method does not require any fine-tuning of pre-trained models, which significantly reduces the response time for users. Since the generative domain of diffusion models are extremely huge and lack of LoRA fine-tuning presents a certain degree of challenge for diffusion-based dragging, we introduce a novel motion supervision method to better align with the feature space, which can get precise point tracking on diffusion models. Similarly, the edited latent often generates images that are inconsistent with the original due to the broad generation space. Built on this, we propose reference guidance during the denoising process, to preserve identity consistency with the input image. Furthermore, we design a method for auto mask generation and achieve simplified dragging interaction, which meets the need of user-friendliness. As shown in Fig. 1, the performance of DragDiffusion relies on LoRA fine-tuning and masks, while our method does not require these and achieves better performance." + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.655, + 0.421, + 0.668 + ], + "angle": 0, + "content": "Our contributions can be summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.675, + 0.468, + 0.733 + ], + "angle": 0, + "content": "- User-friendly interaction. We introduce a point-based image editing method with auto mask generation for user-friendship, simplifying the dragging process while maintaining the comparable results." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.736, + 0.468, + 0.824 + ], + "angle": 0, + "content": "- High performance. We propose a stable motion supervision and point-tracking method that has better adaptability to the latent space of the diffusion model to achieve precise manipulation. We also introduce reference guidance during the denoising process after the optimization of the latent code for better identity preservation." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.826, + 0.468, + 0.883 + ], + "angle": 0, + "content": "- Effective Generalization. Based on diffusion models, our approach can handle a wide range of natural images and various types of dragging, obviously outperforming GAN-based methods." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.887, + 0.468, + 0.899 + ], + "angle": 0, + "content": "- Extensive qualitative and quantitative results demonstrate" + }, + { + "type": "list", + "bbox": [ + 0.078, + 0.675, + 0.468, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.892, + 0.137 + ], + "angle": 0, + "content": "that our method produces superior and more precise point dragging outcomes with clear gaps when compared to existing point-guided image editing methods." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.153, + 0.639, + 0.168 + ], + "angle": 0, + "content": "2. Related work" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.177, + 0.736, + 0.194 + ], + "angle": 0, + "content": "2.1. Text-guided image editing" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.201, + 0.892, + 0.61 + ], + "angle": 0, + "content": "With the success of GANs in the field of image generation, many previous image editing methods [1, 6, 29, 34] are based on StyleGAN. These editing approaches map real images into the latent space of StyleGAN (\\(\\mathcal{W}\\) space or \\(\\mathcal{W}+\\) space) and then achieve image editing by manipulating the latent vectors. However, StyleGAN generates images in a highly restricted domain, and the image quality is not always high. This makes it challenging to successfully invert many real images to a satisfactory latent code. Recently, diffusion models have enabled image synthesis at high quality. An increasing number of image editing tasks [7, 17, 18] have shifted towards being based on diffusion models. Previous image editing methods that rely on diffusion models, such as SDEdit [16] and blended diffusion [2], cannot fully leverage the capabilities of accurate diffusion inversion as they introduced random noise into the input image to create a noisy initial state. Prompt-to-Prompt [7] is the first to accomplish extensive text-guided image editing without the need for diffusion model refinement, including local editing even when a known mask is unavailable. Null-text inversion [18] successfully reconstructs real images by optimizing the null-text embedding at each prediction step, which decides the unconditional prediction. Most recently, pix2pix-zero [22] proposes to learn editing directions in the textual embedding space for specific image translation tasks. In contrast to these methods, our work enables users to conduct image control through point-guided editing." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.62, + 0.744, + 0.637 + ], + "angle": 0, + "content": "2.2. Point-guided image editing" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.645, + 0.892, + 0.901 + ], + "angle": 0, + "content": "In order to facilitate fine-grained editing, several approaches have been introduced for point-based editing, including those by [5, 21, 26, 30]. DragGAN [21] introduces an interactive method for seamless point-based image editing, incorporating two innovative components: the optimization of latent codes to progressively move multiple handle points to their desired positions and a precise point tracking mechanism to accurately follow the path of these handle points. However, the results of DragGAN are constrained by the generative capacity of StyleGAN. Recently, DragDiffusion [26] extends the editing framework of DragGAN to diffusion models with LoRA fine-tuning. Conversely, LoRA fine-tuning significantly extends the waiting time, which is unfriendly to the users. DragonDiffusion [19] utilizes an energy function to guide the editing and a memory bank for editing consistency. SDE-Drag [20] proposes an SDE-based approach to formulate the image" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8405" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.084, + 0.092, + 0.892, + 0.263 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.275, + 0.894, + 0.332 + ], + "angle": 0, + "content": "Figure 2. The pipeline of our framework. Our method generates masks automatically for precise supervision and convenience for users. To improve motion supervision over existing methods like DragGAN, we propose a stable motion supervision method based on diffusion models. The reference guidance where guidance latents are inherited from DDIM inversion preserves identity with the input image commendably." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.357, + 0.47, + 0.403 + ], + "angle": 0, + "content": "editing. In contrast to previous works, we propose a novel point-based image editing framework, which can generate high-quality results without LoRA fine-tuning and masks." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.415, + 0.169, + 0.43 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.439, + 0.218, + 0.453 + ], + "angle": 0, + "content": "3.1. Preliminaries" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.462, + 0.47, + 0.539 + ], + "angle": 0, + "content": "Denoising diffusion probabilistic models (DDPM) [9] aim to map a pure noise \\( z_{T} \\) to an output image \\( z_{0} \\), which is guided by the given conditioning prompt. During the training process, the network \\( \\epsilon_{\\theta} \\) is updated by predicting the noise \\( \\epsilon \\) from the latent variables \\( z_{t} \\):" + }, + { + "type": "equation", + "bbox": [ + 0.117, + 0.561, + 0.469, + 0.58 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\theta} = \\mathbb {E} _ {z _ {0}, \\epsilon \\sim N (0, I), t \\sim U (1, T)} \\| \\epsilon - \\epsilon_ {\\theta} \\left(z _ {t}, t, \\mathcal {C}\\right) \\| _ {2} ^ {2}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.586, + 0.47, + 0.647 + ], + "angle": 0, + "content": "where the noised sample \\( z_{t} \\) is calculated by adding noise \\( \\epsilon \\) to \\( z_{0} \\) according to diffusion step \\( t \\) and \\( \\mathcal{C} \\) is the conditioning input of \\( \\epsilon_{\\theta} \\). During inference, we use DDIM [27] for sampling method:" + }, + { + "type": "equation", + "bbox": [ + 0.077, + 0.667, + 0.475, + 0.719 + ], + "angle": 0, + "content": "\\[\nz _ {t - 1} = \\sqrt {\\frac {\\alpha_ {t - 1}}{\\alpha_ {t}}} z _ {t} + \\sqrt {\\alpha_ {t - 1}} \\left(\\sqrt {\\frac {1}{\\alpha_ {t - 1}} - 1} - \\sqrt {\\frac {1}{\\alpha_ {t}} - 1}\\right) \\epsilon_ {\\theta} \\left(z _ {t}\\right), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.72, + 0.47, + 0.75 + ], + "angle": 0, + "content": "where \\(\\alpha_{t}(t = 0,1,2,\\dots,T)\\) represent the noise scales at corresponding diffusion steps." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.767, + 0.47, + 0.829 + ], + "angle": 0, + "content": "DDIM inversion. The forward process of DDIM can be expressed in terms of \\(\\epsilon_{\\theta}(z_t,t,\\mathcal{C})\\), based on the assumption that the ODE process can be inverted in the limit of small steps:" + }, + { + "type": "equation", + "bbox": [ + 0.077, + 0.848, + 0.473, + 0.902 + ], + "angle": 0, + "content": "\\[\nz _ {t + 1} = \\sqrt {\\frac {\\alpha_ {t + 1}}{\\alpha_ {t}}} z _ {t} + \\sqrt {\\alpha_ {t + 1}} \\left(\\sqrt {\\frac {1}{\\alpha_ {t + 1}} - 1} - \\sqrt {\\frac {1}{\\alpha_ {t}} - 1}\\right) \\epsilon_ {\\theta} \\left(z _ {t}\\right). \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.357, + 0.892, + 0.448 + ], + "angle": 0, + "content": "Stable Diffusion. Stable Diffusion (SD) [24] presents the input image into a lower-dimensional latent space with a Variational Auto-Encoder (VAE) [15], where the input latent \\( z_0 \\) is defined as \\( E(x_0) \\) and the output image \\( x_0 \\) is obtained by \\( D(z_0) \\). We implement our work based on Stable Diffusion model." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.466, + 0.893, + 0.587 + ], + "angle": 0, + "content": "Classifier-free guidance. The extent to which the prompt influences the reverse diffusion process directly affects the quality of image generation. For this purpose, Ho et al. [8] proposes the classifier-free guidance, which controls the degree of impact from text conditioning. Specifically, the null next-embedding \\(\\phi = \\psi(\"\")\\) is employed as a anchor point for unconditional prediction to enhance the text conditioning with guidance scale \\(w\\):" + }, + { + "type": "equation", + "bbox": [ + 0.507, + 0.612, + 0.892, + 0.629 + ], + "angle": 0, + "content": "\\[\n\\tilde {\\epsilon} _ {\\theta} \\left(z _ {t}, t, \\mathcal {C}, \\phi\\right) = w * \\epsilon_ {\\theta} \\left(z _ {t}, t, \\mathcal {C}\\right) + (1 - w) * \\epsilon_ {\\theta} \\left(z _ {t}, t, \\phi\\right). \\tag {4}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.636, + 0.611, + 0.651 + ], + "angle": 0, + "content": "3.2. Overview" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.66, + 0.892, + 0.75 + ], + "angle": 0, + "content": "Given an input image \\(\\mathcal{I}\\), the user should input a number of handle points \\(\\{\\pmb{p}_i = (x_{p,i},y_{p,i})|i = 1,2,\\dots,N\\}\\) and their corresponding target points \\(\\{\\pmb{q}_i = (x_{q,i},y_{q,i})|i = 1,2,\\dots,N\\}\\). The objective is to drag the content of the handle points to the target points based on image \\(\\mathcal{I}\\). The user can also input a mask \\(\\mathbf{M}\\) to specify the modifiable area." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.893, + 0.902 + ], + "angle": 0, + "content": "We operate DDIM inversion with \\(\\mathcal{C} = \\psi(\\mathcal{P})\\) on the input image, where \\(\\mathcal{P}\\) represents the prompt, and get the intermediate results \\(z_{T}, \\ldots, z_{t}, \\ldots, z_{0}\\), where \\(z_{t}\\) is a specific intermediate latent in DDIM inversion. Loop of optimizations is performed on \\(z_{t}\\) to make features of the target points similar enough to that of the handle points, with novel stable motion supervision and automatic mask generation introduced in this paper. Then, the result of optimization \\(z_{t}^{*}\\) is fed into the denoising process with reference guidance to achieve identity consistency with \\(\\mathcal{I}\\). We present the pipeline" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8406" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.088, + 0.096, + 0.21, + 0.108 + ], + "angle": 0, + "content": "Stable motion supervision" + }, + { + "type": "image", + "bbox": [ + 0.105, + 0.117, + 0.199, + 0.185 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.217, + 0.094, + 0.307, + 0.186 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.327, + 0.094, + 0.419, + 0.185 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.43, + 0.121, + 0.457, + 0.148 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.098, + 0.2, + 0.195, + 0.267 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.096, + 0.275, + 0.228, + 0.285 + ], + "angle": 0, + "content": "DragGAN motion supervision" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.2, + 0.307, + 0.268 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.329, + 0.2, + 0.419, + 0.268 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.338, + 0.274, + 0.412, + 0.287 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.305, + 0.47, + 0.388 + ], + "angle": 0, + "content": "Figure 3. Stable motion supervision. When the next point \\( p^{**} \\) is more similar with the handle point \\( p \\) than \\( p^* \\), the handle point has moved to the position of \\( p^{**} \\). Our approach differs significantly from DragGAN in that our method always learns features from the original image, while DragGAN learns features that are continually changing." + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.419, + 0.469, + 0.449 + ], + "angle": 0, + "content": "in Fig. 2, and elaborate three key components in the following subsections." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.463, + 0.312, + 0.479 + ], + "angle": 0, + "content": "3.3. Stable motion supervision" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.488, + 0.469, + 0.639 + ], + "angle": 0, + "content": "There exists misleading in DragGAN due to the proposed motion supervision. In contrast to DragGAN, we introduce an efficient drag method that is more suitable for diffusion models. Inspired by DIFT [28], we utilize the feature maps of U-Net denoiser as the semantic information of the input image. Specially, we consider the second and third upper layers of U-Net with \\((z_{t}, t, \\mathcal{C})\\) as inputs, and concatenate the two layers as our feature maps \\(\\mathbf{F}\\) used for stable motion supervision. \\(\\mathbf{F}\\) is resized to have the same resolution as the input image in order to get precise manipulation." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.641, + 0.469, + 0.791 + ], + "angle": 0, + "content": "As shown in Fig. 3, \\( p^* \\) represents current position of the handle point, and \\( p^{**} \\) is the next position that we want to drag to. The method for calculating \\( p^{**} \\) is presented in Algorithm 1. We start from \\( p^* \\), move along the direction of \\( q \\) by \\( K \\) units, and then find the nearest integer coordinates to determine \\( p^{**} \\). In contrast to DragGAN, we employ \\( p^* \\) to learn features of the original image at the \\( p \\) position, ensuring that the learned features are consistently accurate. Inspired by DragonDiffusion [19], we define the similarity loss function of two feature maps as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.154, + 0.809, + 0.469, + 0.845 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {c o s}} \\left(\\mathbf {F}, \\mathbf {F} _ {0}\\right) = \\frac {1}{1 + \\frac {1 + \\cos \\left(\\mathbf {F} , s g \\left(\\mathbf {F} _ {0}\\right)\\right)}{2}}, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.856, + 0.469, + 0.9 + ], + "angle": 0, + "content": "where \\(\\mathbf{F}_0\\) represents the feature maps corresponding to the input image and \\(sg(\\cdot)\\) is the stop gradient operator. Therefore, the stable motion supervision loss is defined as:" + }, + { + "type": "title", + "bbox": [ + 0.511, + 0.096, + 0.678, + 0.11 + ], + "angle": 0, + "content": "Algorithm 1: Next point" + }, + { + "type": "text", + "bbox": [ + 0.521, + 0.113, + 0.78, + 0.127 + ], + "angle": 0, + "content": "Input: Current point \\( p^* \\), Target point \\( q \\)" + }, + { + "type": "text", + "bbox": [ + 0.522, + 0.128, + 0.681, + 0.143 + ], + "angle": 0, + "content": "Output: Next point \\( p^{**} \\)" + }, + { + "type": "text", + "bbox": [ + 0.509, + 0.148, + 0.678, + 0.163 + ], + "angle": 0, + "content": "1 if \\(\\| \\pmb {q} - \\pmb{p}^{*}\\|_{2} < K\\) then" + }, + { + "type": "text", + "bbox": [ + 0.509, + 0.163, + 0.61, + 0.179 + ], + "angle": 0, + "content": "2 \\(\\pmb{p}^{**}\\gets \\pmb {q};\\)" + }, + { + "type": "text", + "bbox": [ + 0.509, + 0.18, + 0.551, + 0.192 + ], + "angle": 0, + "content": "3 else" + }, + { + "type": "text", + "bbox": [ + 0.509, + 0.193, + 0.64, + 0.216 + ], + "angle": 0, + "content": "4 \\(\\pmb{d} \\gets \\frac{\\pmb{q} - \\pmb{p}^{*}}{\\|\\pmb{q} - \\pmb{p}^{*}\\|_{2}};\\)" + }, + { + "type": "text", + "bbox": [ + 0.509, + 0.215, + 0.732, + 0.23 + ], + "angle": 0, + "content": "5 \\(\\pmb{p}^{**}\\gets \\pmb{p}^* +round(\\pmb {d}*\\pmb {K})\\)" + }, + { + "type": "text", + "bbox": [ + 0.509, + 0.231, + 0.55, + 0.243 + ], + "angle": 0, + "content": "6 end" + }, + { + "type": "text", + "bbox": [ + 0.509, + 0.246, + 0.594, + 0.261 + ], + "angle": 0, + "content": "7 return \\(p^{**}\\)" + }, + { + "type": "list", + "bbox": [ + 0.509, + 0.148, + 0.732, + 0.261 + ], + "angle": 0, + "content": null + }, + { + "type": "equation", + "bbox": [ + 0.556, + 0.305, + 0.891, + 0.365 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} = \\sum_ {i = 1} ^ {N} \\mathcal {L} _ {\\text {c o s}} \\left(\\mathbf {F} \\left(\\boldsymbol {p} _ {i} ^ {* *}\\right), \\mathbf {F} _ {0} \\left(\\boldsymbol {p} _ {i}\\right)\\right) \\tag {6} \\\\ + \\lambda \\mathcal {L} _ {\\cos} (\\mathbf {F} \\odot (1 - \\mathbf {M}), \\mathbf {F} _ {0} \\odot (1 - \\mathbf {M})), \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.37, + 0.879, + 0.387 + ], + "angle": 0, + "content": "where \\(\\mathbf{F}(\\pmb{p}_i^{**})\\) denotes the feature values of \\(\\mathbf{F}\\) at pixel \\(\\pmb{p}_i^{**}\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.387, + 0.891, + 0.493 + ], + "angle": 0, + "content": "After motion supervision, the updated latent code \\( z_{t}^{*} \\) may change the position of handle points. Consequently, we need to determine whether the handle point has reached the next target point. To do this, we compare which one, \\( \\mathbf{F}(\\pmb{p}_i^{**}) \\) or \\( \\mathbf{F}(\\pmb{p}_i^*) \\), resembles \\( \\mathbf{F}_0(\\pmb{p}_i) \\) more. If \\( \\cos(\\mathbf{F}(\\pmb{p}_i^{**}), \\mathbf{F}_0(\\pmb{p}_i)) \\) is higher, the handle point \\( \\pmb{p}_i^* \\) is updated to \\( \\pmb{p}_i^{**} \\) and \\( \\pmb{p}_i^{**} \\) is further updated to \\( \\text{next}(\\pmb{p}_i^*, \\pmb{q}_i) \\) according to Algorithm 1." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.5, + 0.705, + 0.516 + ], + "angle": 0, + "content": "3.4. Auto mask generation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.524, + 0.892, + 0.659 + ], + "angle": 0, + "content": "For user-friendly interaction, image editing should avoid burdening users with the task of drawing masks. Existing drag methods based on diffusion models [19, 26] always rely on mask inputs, which are provided by users. This requires users to have a deep understanding of the drag method because an incorrectly drawn mask can lead to unpredictable outcomes. Consequently, we propose a method that can automatically generate masks, which are able to change adaptively during the training process." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.66, + 0.892, + 0.826 + ], + "angle": 0, + "content": "In the process of motion supervision, the parts with larger gradients of \\( z_{t}^{*} \\) have a greater impact on the results, while those with smaller gradients are generally the parts that are expected to remain stationary. However, if the regions with relatively small gradients keep changing, the final result becomes unpredictable. Due to the self-attention layer in U-Net, minor changes made at step \\( t \\) are very likely to have an unacceptable impact on the final image. Therefore, we still need control over these low-impact areas. We achieve this by using regions with normalized gradients greater than a threshold \\( g \\) as the drag mask \\( \\mathbf{M} \\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.826, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Every time \\( \\pmb{p}_i^* \\) is updated, the hot-zone of \\( z_t \\) also changes due to the position shifting of the drag point. We update our dynamic mask \\( \\mathbf{M} \\) each time \\( \\pmb{p}_i^* \\) changes, but there is no need to change the mask when \\( \\pmb{p}_i^* \\) remains the same. This is because for the same \\( \\pmb{p}_i^* \\) and \\( \\pmb{p}_i^{**} \\), the hot-zone is consistent." + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "8407" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.129, + 0.096, + 0.151, + 0.109 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.129, + 0.111, + 0.452, + 0.387 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.466, + 0.096, + 0.488, + 0.109 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image", + "bbox": [ + 0.461, + 0.111, + 0.85, + 0.386 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.404, + 0.894, + 0.448 + ], + "angle": 0, + "content": "Figure 4. Qualitative comparison of our approach to DragDiffusion. Due to the misleading of motion supervision, DragDiffusion always cannot move the handle points to the target. Furthermore, DragDiffusion often exhibits significant background changes in the absence of the mask. In our method, the handle points reach the target positions precisely even without the mask." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.473, + 0.47, + 0.503 + ], + "angle": 0, + "content": "This way, \\(\\mathbf{M}\\) adapts to changes of \\(z_{t}\\), aligning better with the dragging requirements." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.511, + 0.265, + 0.527 + ], + "angle": 0, + "content": "3.5. Reference guidance" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.534, + 0.47, + 0.67 + ], + "angle": 0, + "content": "With optimized \\( z_{i}^{*} \\), the generated images with direct DDIM denoising often exhibit identity changes and a decrease in image quality. Especially in areas where the pretrained models have limited generation capabilities, naively generating with \\( z_{i} \\) after DDIM Inversion may not even reproduce the original image. Drawing inspiration from [4, 19, 26], we employ mutual self-attention to ensure that images generated from \\( z_{i}^{*} \\) maintain content consistency with the original image." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.671, + 0.471, + 0.792 + ], + "angle": 0, + "content": "As illustrated in Fig. 2, we replace the key and value of self-attention layers in backward process of \\( z_{i}^{*} \\) with the corresponding key and value in the process of reconstructing the source image \\( \\mathcal{I} \\). Due to the feature correlations in lower layers being relatively weak, the mutual self-attention is only applied to the upper layers of U-Net \\( \\epsilon_{\\theta} \\). We then perform sampling using the linear combination of the reference guidance and normal prediction:" + }, + { + "type": "equation", + "bbox": [ + 0.124, + 0.816, + 0.47, + 0.833 + ], + "angle": 0, + "content": "\\[\n\\epsilon_ {i} ^ {*} \\leftarrow w \\epsilon_ {\\theta} \\left(z _ {i} ^ {*}, z _ {i}, i, \\mathcal {C}\\right) + (1 - w) \\epsilon_ {\\theta} \\left(z _ {i} ^ {*}, i, \\mathcal {C}\\right), \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.841, + 0.47, + 0.901 + ], + "angle": 0, + "content": "where \\( w \\) controls the strength of the reference guidance. Specifically, \\( z_{i} \\) is the corresponding latent during DDIM inversion, rather than that generated from \\( z_{t} \\). When \\( w = 1 \\), our reference guidance is the same as mutual self-attention." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.473, + 0.893, + 0.565 + ], + "angle": 0, + "content": "However, in the cases where the input image is infrequent for the model and \\( z_{t} \\) undergoes significant changes, using mutual self-attention alone cannot guarantee identity consistency. When we increase the strength of reference guidance, the generated results can better ensure content and texture consistency with the input image." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.578, + 0.634, + 0.596 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.604, + 0.892, + 0.636 + ], + "angle": 0, + "content": "In this section, we evaluate the proposed method qualitatively and quantitatively, on both synthetic and real images." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.654, + 0.893, + 0.835 + ], + "angle": 0, + "content": "Implementation details. In our experiments, except for generated image manipulations, all tasks are based on Stable Diffusion 1.5 (SD-1.5). Unless otherwise specified, in both DDIM inversion and sampling we set the number of the sampling steps to be 50 via using the stride of 20 over 1000 diffusion steps. We set the latent from the 35-th DDIM step, which contains rich semantic information corresponding to the input image, as \\( z_{t} \\). The gradient threshold \\( g \\) is 0.4, the motion step \\( K \\) is 12 and the reference guidance scale \\( w \\) is 4 by default. We use a learning rate of 0.01 for \\( z_{t} \\) without any decay schedule and train the latent with Adam [14] solver." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.856, + 0.894, + 0.903 + ], + "angle": 0, + "content": "Dataset. We employ DragBench introduced in DragDiffusion as our evaluation dataset as it encompasses various types of images, and we also follow its dragging instruc" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "8408" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.129, + 0.151, + 0.153, + 0.166 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image_caption", + "bbox": [ + 0.188, + 0.096, + 0.236, + 0.112 + ], + "angle": 0, + "content": "Inputs" + }, + { + "type": "image", + "bbox": [ + 0.157, + 0.114, + 0.27, + 0.201 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.301, + 0.096, + 0.354, + 0.11 + ], + "angle": 0, + "content": "Results" + }, + { + "type": "image", + "bbox": [ + 0.273, + 0.113, + 0.386, + 0.201 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.423, + 0.096, + 0.469, + 0.111 + ], + "angle": 0, + "content": "Inputs" + }, + { + "type": "image", + "bbox": [ + 0.388, + 0.113, + 0.503, + 0.201 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.532, + 0.096, + 0.586, + 0.11 + ], + "angle": 0, + "content": "Results" + }, + { + "type": "image", + "bbox": [ + 0.504, + 0.113, + 0.617, + 0.201 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.656, + 0.096, + 0.701, + 0.111 + ], + "angle": 0, + "content": "Inputs" + }, + { + "type": "image", + "bbox": [ + 0.622, + 0.113, + 0.735, + 0.201 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.765, + 0.096, + 0.818, + 0.11 + ], + "angle": 0, + "content": "Results" + }, + { + "type": "image", + "bbox": [ + 0.737, + 0.113, + 0.851, + 0.201 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.129, + 0.241, + 0.153, + 0.257 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image", + "bbox": [ + 0.157, + 0.206, + 0.271, + 0.293 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.274, + 0.206, + 0.385, + 0.293 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.389, + 0.206, + 0.502, + 0.293 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.206, + 0.617, + 0.293 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.621, + 0.205, + 0.734, + 0.293 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.737, + 0.206, + 0.85, + 0.293 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.129, + 0.334, + 0.153, + 0.349 + ], + "angle": 0, + "content": "(c)" + }, + { + "type": "image", + "bbox": [ + 0.157, + 0.296, + 0.271, + 0.384 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.274, + 0.296, + 0.385, + 0.384 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.389, + 0.296, + 0.502, + 0.384 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.296, + 0.617, + 0.384 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.621, + 0.296, + 0.734, + 0.384 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.737, + 0.296, + 0.85, + 0.384 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.129, + 0.425, + 0.153, + 0.44 + ], + "angle": 0, + "content": "(d)" + }, + { + "type": "image", + "bbox": [ + 0.157, + 0.388, + 0.271, + 0.476 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.274, + 0.388, + 0.385, + 0.476 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.389, + 0.388, + 0.502, + 0.476 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.388, + 0.617, + 0.476 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.621, + 0.388, + 0.734, + 0.476 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.737, + 0.388, + 0.85, + 0.476 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.487, + 0.892, + 0.515 + ], + "angle": 0, + "content": "Figure 5. Qualitative results of our approach on generated images with SD-1.5 and its fine-tuned variants. (a) SD-1.5. (b) Majicmix Realistic V7. (c) Interior Design Supermix. (d) Dvarch." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.542, + 0.47, + 0.573 + ], + "angle": 0, + "content": "tions. DragBench contains 205 images, with well-annotated prompts, masks, and dragging points." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.644, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Evaluation metrics. We evaluate the dragging results from two aspects: the quality of dragged image and the accuracy of position of handle points. We employ Image Fidelity (IF) [13] and Mean Distance (MD) [21] to assess these two aspects, respectively. IF evaluates the similarity between the input and output images, which is calculated by subtracting the mean LPIPS [32] over all pairs of original and edited images from 1. MD quantifies how well the handle points move to the target points. The same as DragDiffusion, we utilize DIFT [28] to identify the points in the output image that correspond to the handle points in the input image. These identified points are then treated as the definitive handle points after the dragging process. Subsequently, MD is calculated as the average Euclidean distance between the positions of all target points and their corresponding final handle points. A higher IF and a lower MD indicate a better quality of the dragging approach." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.541, + 0.706, + 0.557 + ], + "angle": 0, + "content": "4.1. Qualitative evaluation" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.566, + 0.892, + 0.642 + ], + "angle": 0, + "content": "We conduct extensive experiments for the qualitative comparison between our method and DragDiffusion. To provide a thorough evaluation, we adopt SD-1.5 [24] as our base model for all real images, and also employ other fine-tuned variants of SD-1.5 for generated image manipulation." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.643, + 0.893, + 0.825 + ], + "angle": 0, + "content": "We present real image editing results for various object categories and user inputs in Fig. 4. Our approach accurately moves the handle points to reach the target positions, resulting in a wide range of natural and diverse manipulation effects such as altering the poses of animals, rearranging landscapes, modifying human body parts, and reshaping objects, in the absence of the mask. Conversely, DragDiffusion is unable to precisely move the handle points to the target positions, which is mainly attributed to the misleading of point tracking that is employed in DragGAN and DragDiffusion. It is worth noting that DragDiffusion always changes the non-dragged regions when no mask is provided." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.825, + 0.894, + 0.903 + ], + "angle": 0, + "content": "We also conduct a series of point manipulations based on the generated images, as shown in Fig. 5. The classifier-free guidance scale is configured at 7.5 for various SD-1.5 variants. Notably, the reference guidance is unnecessary for generative editing, as the prompt of generated images pro" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8409" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.162, + 0.09, + 0.813, + 0.246 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.257, + 0.895, + 0.288 + ], + "angle": 0, + "content": "Figure 6. Out-of-distribution manipulations. In some cases, the input image cannot be reconstructed with DDIM Inversion. Utilizing reference guidance, edited image preserves its identity with the original image." + }, + { + "type": "table", + "bbox": [ + 0.083, + 0.308, + 0.468, + 0.436 + ], + "angle": 0, + "content": "
MethodIF(↑)MD(↓)Time(s)
DragGAN0.69557.359.1
DragDiffusion0.88236.7169.4
DragDiffusion w/o LoRA0.86353.6333.7
DragDiffusion w/o mask0.76647.0365.6
SDE-Drag0.87145.4450.1
Ours w/ mask0.88929.0727.3
Ours0.91032.5530.6
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.445, + 0.472, + 0.487 + ], + "angle": 0, + "content": "Table 1. Quantitative evaluation of different methods with Drag-Bench. Evaluation metrics include Image Fidelity (IF) and Mean Distance (MD)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.521, + 0.47, + 0.673 + ], + "angle": 0, + "content": "vides substantial guidance during the denoising process. As shown in Fig. 5, our method achieves point-based editing with high quality and precision on generated images. It can be observed that in some cases, the details of generated images have changed. This is because the self-attention layer in diffusion models may introduce slight variation in other parts when a specific area of the image is edited. If users want to preserve certain areas without any changes, they can employ inpainting during image generation with masks to determine which regions should remain unchanged." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.693, + 0.295, + 0.709 + ], + "angle": 0, + "content": "4.2. Quantitative evaluation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.72, + 0.471, + 0.901 + ], + "angle": 0, + "content": "We conduct quantitative evaluation by comparing our method to the baseline dragging method on diffusion, DragDiffusion. The evaluation, which incorporates IF and MD metrics, makes use of the DragBench dataset provided by DragDiffusion. The results, as shown in Tab. 1, demonstrate that our approach significantly surpasses DragDiffusion and SDE-Drag on the DragBench dataset. Notably, our method achieves superior dragging results without requiring LoRA fine-tuning, offering substantial time savings during extensive experiments. The performance of DragDiffusion degrades significantly without LoRA nor mask. The execution time in Tab. 1 is calculated on a single A100 GPU." + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.312, + 0.892, + 0.476 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.487, + 0.892, + 0.515 + ], + "angle": 0, + "content": "Figure 7. When users do not provide an appropriate mask, the generated results are prone to issues." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.542, + 0.625, + 0.556 + ], + "angle": 0, + "content": "4.3. Discussions" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.565, + 0.892, + 0.791 + ], + "angle": 0, + "content": "Reference Guidance. With our reference guidance, we enable the diffusion models to perform out-of-distribution manipulations. As shown in Fig. 6, the original image cannot be faithfully reconstructed using DDIM Inversion. When we introduce reference guidance, \\( z_{t} \\) can better reconstruct the input image. It is worth noting that \\( z_{i}(i = t,t - 1,\\dots,1) \\) is obtained during the DDIM Inversion process, so it contains richer information compared to \\( z_{t} \\). After editing \\( z_{t} \\), the image generated from \\( z_{t}^{*} \\) is very different from the input image in terms of identity. As we increase the reference guidance scale \\( w \\), the consistency between the generated image and original image gradually improves. However, if \\( w \\) is too large, the generated image may become oversaturated. In our experiments, a value for \\( w \\) in the range of 3 to 5 is considered a good choice." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.811, + 0.893, + 0.903 + ], + "angle": 0, + "content": "Auto mask generation. When users create imprecise masks, the results are unpredictable. As shown in Fig. 7, when the mask is too small, it will restrict the editing of the image. On the other hand, when the mask is too large, changes may occur in the background. Therefore, an appropriate mask is necessary for point-based image editing. Our" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8410" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.084, + 0.09, + 0.47, + 0.248 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.259, + 0.471, + 0.302 + ], + "angle": 0, + "content": "Figure 8. The changes in the gradient and prediction of \\( z_{t} \\) during optimization. The term of \"K-th step\" means the K-th update of handle points \\( p^{*} \\)." + }, + { + "type": "image", + "bbox": [ + 0.083, + 0.322, + 0.262, + 0.426 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.095, + 0.432, + 0.249, + 0.444 + ], + "angle": 0, + "content": "(a) Which upper layer to use." + }, + { + "type": "image", + "bbox": [ + 0.28, + 0.319, + 0.465, + 0.426 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.329, + 0.432, + 0.416, + 0.443 + ], + "angle": 0, + "content": "(b) Effects of K." + }, + { + "type": "image_caption", + "bbox": [ + 0.197, + 0.456, + 0.348, + 0.471 + ], + "angle": 0, + "content": "Figure 9. Ablation study." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.499, + 0.469, + 0.543 + ], + "angle": 0, + "content": "auto mask generation method liberates users from the need to manually draw masks, which is user-friendly, especially for novice users." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.546, + 0.47, + 0.742 + ], + "angle": 0, + "content": "Figure 8 shows the update of gradient and prediction with the optimization of \\( z_{t} \\). It can be observed that when the handle points update, the gradient of \\( z_{t} \\) will change simultaneously. In the K-th step, the whole arm of the robot needs to be lifted, so the gradients for the arm part are relatively large. However, by the time we reach the (K+1)-th step, the robot's arm has already been dragged into the right place, and only the palm part needs to extend outward. Therefore, only the gradients for the palm part are high. Our auto mask effectively captures these details and incorporates this variation into the updates of \\( z_{t} \\), thereby providing better control over the direction of \\( z_{t} \\). That is the reason why the IF metric of without mask is higher than that with mask in Tab. 1." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.47, + 0.903 + ], + "angle": 0, + "content": "Ablation study. We perform ablation studies to analyze the impact of the feature used in stable motion supervision, and we evaluate the performance on DragBench, as shown in Fig. 9a. Notably, the concatenated feature maps from upper layers 1&2 exhibit the best performance, demonstrating an optimal balance between resolution and discriminativeness. Additionally, we investigate the effects of different values of \\( K \\) in Fig. 9b, and it is observed that \\( K = 12 \\) yields better results." + }, + { + "type": "table", + "bbox": [ + 0.54, + 0.089, + 0.855, + 0.153 + ], + "angle": 0, + "content": "
MethodsIF(↑)MD(↓)
DragGAN0.85651.32
DragonDiffusion0.89952.21
Stable motion supervision0.88929.07
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.164, + 0.892, + 0.207 + ], + "angle": 0, + "content": "Table 2. Quantitative comparison of motion supervision between our method and previous works. Our approach achieves more accurate dragging results than other methods." + }, + { + "type": "image", + "bbox": [ + 0.521, + 0.221, + 0.872, + 0.302 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.312, + 0.892, + 0.355 + ], + "angle": 0, + "content": "Figure 10. When the manipulated portion in the image is very small or the dragging distance is extensive, it is challenging for the dragging to have a noticeable effect." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.382, + 0.892, + 0.473 + ], + "angle": 0, + "content": "A comparison of \"drag\" method between our approach and previous methods is shown in Tab. 2. We only replace stable motion supervision with the \"drag\" method in other works. It can be found that stable motion supervision is more suitable for the feature space of diffusion models, although DragonDiffusion achieves slightly higher IF." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.493, + 0.892, + 0.6 + ], + "angle": 0, + "content": "Limitations. Despite the outstanding performance, our method has some limitations. Particularly, when the editing area is too small or the dragging distance is too long, the dragging result may be unsatisfactory. As depicted in Fig. 10, when the girl's face occupies only a small part of the image, and we intend to edit the mouth, the desired changes of mouth do not occur as expected." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.613, + 0.619, + 0.629 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.639, + 0.892, + 0.851 + ], + "angle": 0, + "content": "In this work, we attribute an excellent point-based image editing method to three criteria: user-friendly interaction, good performance, and effective generalization. We abandon the LoRA fine-tuning and masks, which is convenient for users. Furthermore, we propose a stable motion supervision method and reference guidance to generate high-quality results based on diffusion models. Leveraging the powerful generative capabilities of the diffusion model, we achieve impressive dragging results across various domains. However, in our experiments, we observe that for very small-scale manipulations or long-distance dragging in images, the results are often not satisfactory. Diffusion models with stronger generative capabilities may be needed to achieve a broader range of dragging manipulations." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.891, + 0.9 + ], + "angle": 0, + "content": "Acknowledgements: This work is partially supported by Natural Science Foundation of China Grant 41930110." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "8411" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.175, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.184 + ], + "angle": 0, + "content": "[1] Rameen Abdal, Peihao Zhu, Niloy J Mitra, and Peter Wonka. Styleflow: Attribute-conditioned exploration of stylegan-generated images using conditional continuous normalizing flows. ACM Transactions on Graphics (ToG), 40(3):1-21, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.186, + 0.472, + 0.243 + ], + "angle": 0, + "content": "[2] Omri Avrahami, Dani Lischinski, and Ohad Fried. Blended diffusion for text-driven editing of natural images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18208-18218, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.245, + 0.472, + 0.312 + ], + "angle": 0, + "content": "[3] Tim Brooks, Aleksander Holynski, and Alexei A Efros. Instructpix2pix: Learning to follow image editing instructions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18392-18402, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.314, + 0.472, + 0.37 + ], + "angle": 0, + "content": "[4] Mingdeng Cao, Xintao Wang, Zhongang Qi, Ying Shan, Xiaohu Qie, and Yinqiang Zheng. Masactrl: Tuning-free mutual self-attention control for consistent image synthesis and editing. arXiv preprint arXiv:2304.08465, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.372, + 0.472, + 0.413 + ], + "angle": 0, + "content": "[5] Yuki Endo. User-controllable latent transformer for stylegan image layout editing. In Computer Graphics Forum, pages 395-406. Wiley Online Library, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.415, + 0.472, + 0.471 + ], + "angle": 0, + "content": "[6] Dave Epstein, Taesung Park, Richard Zhang, Eli Shechtman, and Alexei A Efros. Blobgan: Spatially disentangled scene representations. In European Conference on Computer Vision, pages 616-635. Springer, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.473, + 0.472, + 0.528 + ], + "angle": 0, + "content": "[7] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-prompt image editing with cross attention control. arXiv preprint arXiv:2208.01626, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.53, + 0.472, + 0.558 + ], + "angle": 0, + "content": "[8] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.56, + 0.472, + 0.601 + ], + "angle": 0, + "content": "[9] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.603, + 0.47, + 0.658 + ], + "angle": 0, + "content": "[10] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.66, + 0.47, + 0.716 + ], + "angle": 0, + "content": "[11] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4401-4410, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.717, + 0.472, + 0.786 + ], + "angle": 0, + "content": "[12] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of stylegan. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8110-8119, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.788, + 0.472, + 0.858 + ], + "angle": 0, + "content": "[13] Bahjat Kawar, Shiran Zada, Oran Lang, Omer Tov, Huiwen Chang, Tali Dekel, Inbar Mosseri, and Michal Irani. Imagic: Text-based real image editing with diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6007-6017, 2023. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.859, + 0.472, + 0.9 + ], + "angle": 0, + "content": "[14] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014.5" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.472, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "[15] Diederik P Kingma and Max Welling. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114, 2013. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.121, + 0.892, + 0.176 + ], + "angle": 0, + "content": "[16] Chenlin Meng, Yutong He, Yang Song, Jiaming Song, Jia-jun Wu, Jun-Yan Zhu, and Stefano Ermon. Sdedit: Guided image synthesis and editing with stochastic differential equations. arXiv preprint arXiv:2108.01073, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.177, + 0.892, + 0.232 + ], + "angle": 0, + "content": "[17] Daiki Miyake, Akihiro Iohara, Yu Saito, and Toshiyuki Tanaka. Negative-prompt inversion: Fast image inversion for editing with text-guided diffusion models. arXiv preprint arXiv:2305.16807, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.233, + 0.895, + 0.302 + ], + "angle": 0, + "content": "[18] Ron Mokady, Amir Hertz, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Null-text inversion for editing real images using guided diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6038–6047, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.303, + 0.892, + 0.357 + ], + "angle": 0, + "content": "[19] Chong Mou, Xintao Wang, Jiechong Song, Ying Shan, and Jian Zhang. Dragondiffusion: Enabling drag-style manipulation on diffusion models. arXiv preprint arXiv:2307.02421, 2023. 2, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.359, + 0.892, + 0.413 + ], + "angle": 0, + "content": "[20] Shen Nie, Hanzhong Allan Guo, Cheng Lu, Yuhao Zhou, Chenyu Zheng, and Chongxuan Li. The blessing of randomness: Sde beats ode in general diffusion-based image editing. arXiv preprint arXiv:2311.01410, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.414, + 0.892, + 0.483 + ], + "angle": 0, + "content": "[21] Xingang Pan, Ayush Tewari, Thomas Leimkuhler, Lingjie Liu, Abhimitra Meka, and Christian Theobalt. Drag your gan: Interactive point-based manipulation on the generative image manifold. In ACM SIGGRAPH 2023 Conference Proceedings, pages 1-11, 2023. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.484, + 0.892, + 0.538 + ], + "angle": 0, + "content": "[22] Gaurav Parmar, Krishna Kumar Singh, Richard Zhang, Yijun Li, Jingwan Lu, and Jun-Yan Zhu. Zero-shot image-to-image translation. In ACM SIGGRAPH 2023 Conference Proceedings, pages 1-11, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.539, + 0.892, + 0.593 + ], + "angle": 0, + "content": "[23] Daniel Roich, Ron Mokady, Amit H Bermano, and Daniel Cohen-Or. Pivotal tuning for latent-based editing of real images. ACM Transactions on graphics (TOG), 42(1):1-13, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.595, + 0.892, + 0.664 + ], + "angle": 0, + "content": "[24] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10684-10695, 2022. 1, 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.665, + 0.892, + 0.747 + ], + "angle": 0, + "content": "[25] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in Neural Information Processing Systems, 35:36479-36494, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.748, + 0.892, + 0.803 + ], + "angle": 0, + "content": "[26] Yujun Shi, Chuhui Xue, Jiachun Pan, Wenqing Zhang, Vincent YF Tan, and Song Bai. Dragdiffusion: Harnessing diffusion models for interactive point-based image editing. arXiv preprint arXiv:2306.14435, 2023. 2, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.804, + 0.892, + 0.845 + ], + "angle": 0, + "content": "[27] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.846, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[28] Luming Tang, Menglin Jia, Qianqian Wang, Cheng Perng Phoo, and Bharath Hariharan. Emergent correspondence from image diffusion. arXiv preprint arXiv:2306.03881, 2023.4,6" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.895, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8412" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.161 + ], + "angle": 0, + "content": "[29] Jianyuan Wang, Ceyuan Yang, Yinghao Xu, Yujun Shen, Hongdong Li, and Bolei Zhou. Improving gan equilibrium by raising spatial awareness. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11285-11293, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.163, + 0.47, + 0.204 + ], + "angle": 0, + "content": "[30] Sheng-Yu Wang, David Bau, and Jun-Yan Zhu. Rewriting geometric rules of a gan. ACM Transactions on Graphics (TOG), 41(4):1-16, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.206, + 0.47, + 0.261 + ], + "angle": 0, + "content": "[31] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3836-3847, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.263, + 0.47, + 0.331 + ], + "angle": 0, + "content": "[32] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.333, + 0.47, + 0.402 + ], + "angle": 0, + "content": "[33] Zhixing Zhang, Ligong Han, Arnab Ghosh, Dimitris N Metaxas, and Jian Ren. Sine: Single image editing with text-to-image diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6027-6037, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.403, + 0.47, + 0.472 + ], + "angle": 0, + "content": "[34] Jiapeng Zhu, Ceyuan Yang, Yujun Shen, Zifan Shi, Bo Dai, Deli Zhao, and Qifeng Chen. Linkgan: Linking gan latents to pixels for controllable image synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7656-7666, 2023. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.472 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.957 + ], + "angle": 0, + "content": "8413" + } + ] +] \ No newline at end of file diff --git a/2024/EasyDrag_ Efficient Point-based Manipulation on Diffusion Models/3c97514b-8a67-4bdc-8795-3e7cfc02ceca_origin.pdf b/2024/EasyDrag_ Efficient Point-based Manipulation on Diffusion Models/3c97514b-8a67-4bdc-8795-3e7cfc02ceca_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..b5f0ff58a69c3f4f5a6b6c2d23d79bf3abdd91c3 --- /dev/null +++ b/2024/EasyDrag_ Efficient Point-based Manipulation on Diffusion Models/3c97514b-8a67-4bdc-8795-3e7cfc02ceca_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:59029e9ae0af40c9248743ed849d39ed53281cecc22f17a8a72e0133397355e0 +size 6297504 diff --git a/2024/EasyDrag_ Efficient Point-based Manipulation on Diffusion Models/full.md b/2024/EasyDrag_ Efficient Point-based Manipulation on Diffusion Models/full.md new file mode 100644 index 0000000000000000000000000000000000000000..59e67ded0da629f2bd8e5febb423054a78a11278 --- /dev/null +++ b/2024/EasyDrag_ Efficient Point-based Manipulation on Diffusion Models/full.md @@ -0,0 +1,358 @@ +# EasyDrag: Efficient Point-based Manipulation on Diffusion Models + +Xingzhong Hou1,2 + +Boxiao Liu3 + +Yi Zhang3 + +Jihao Liu3,4 + +Yu Liu3 + +Haihang You1* + +$^{1}$ State Key Lab of Processors, Institute of Computing Technology, Chinese Academy of Sciences $^{2}$ School of Computer Science and Technology, University of Chinese Academy of Sciences $^{3}$ SenseTime Research $^{4}$ CUHK MMLab + +{houxingzhong, youhaihang}@ict.ac.cn {liuboxiao, zhangyi17}@sensetime.com jihaoliu@link.cuhk.edu.hk liuyuisanai@gmail.com + +![](images/be7aa8bb33edfafc4f6d1777fbe28b052deb88c8d41da4902d58fbb7958ebe0c.jpg) +Figure 1. DragGAN fails in both cases due to limited model capacity. Similarly, DragDiffusion relies on LoRA fine-tuning and hand-drawn masks to achieve high-quality results, and SDE-Drag also fails to address these two cases effectively. In contrast, our approach ensures precise manipulation and detail preservation without fine-tuning or masks. + +# Abstract + +Generative models are gaining increasing popularity, and the demand for precisely generating images is on the rise. However, generating an image that perfectly aligns with users' expectations is extremely challenging. The shapes of objects, the poses of animals, the structures of landscapes, and more may not match the user's desires, and this applies to real images as well. This is where point-based image editing becomes essential. An excellent image editing method needs to meet the following criteria: user-friendly interaction, high performance, and good generalization capability. Due to the limitations of StyleGAN, DragGAN exhibits limited robustness across diverse scenarios, while DragDiffusion lacks user-friendliness due to the necessity of LoRA fine-tuning and masks. In this paper, we introduce a novel interactive point-based image editing framework, called EasyDrag, that leverages pretrained diffusion models to achieve high-quality editing outcomes and user-friendship. Extensive experimentation demonstrates that our approach surpasses DragDiffusion in terms of both image quality and editing precision for point-based + +image manipulation tasks. The code will be available on https://github.com/Ace-Pegasus/EasyDrag. + +# 1. Introduction + +In the past few years, diffusion models [9, 24, 27] represent a groundbreaking advancement in the domain of generative modeling. In particular, text-to-image diffusion models [24, 25, 31] can produce remarkable images based on textual prompts. Recently, such models have also been extended for image editing [3, 7, 13, 18, 23, 33]. Since diffusion models depend on textual input, image elements, including objects and styles, can be altered with a high degree of quality and diversity guided by edited prompts. + +However, text-guided image editing cannot precisely direct changes in the image, and it is unpredictable for a specific point in the image that one might want to move. Therefore, point-guided image editing becomes essential. A recent study DragGAN [21] introduces "drag" operation for precise image editing, which has received widespread attention for its point-based interaction. Nonetheless, it encounters limitations tied to the capacity and adaptability of StyleGAN [11, 12] as shown in Fig. 1. In contrast, diffu + +sion models [9, 24, 27] exhibit better stability and superior generative quality. DragDiffusion [26] enables interactive point-based image editing on diffusion models with LoRA fine-tuning [10] but introduces high time consumption at the same time. + +What capabilities should point-based image editing possess? User-friendly interaction, good performance and effective generalization. User-friendly interaction implies ease of use and short response times for users, while good performance refers to precise dragging with maintaining identity. Effective generalization requires the method to work with a single model for various images. Unfortunately, existing methods have not been able to simultaneously exhibit these capabilities. DragGAN lacks generalization of dragging general domain image, while DragDiffusion is less user-friendly due to its reliance on LoRA fine-tuning and masks. In this paper, we propose a novel point-based image editing framework built on diffusion models, which can concurrently fulfill these requirements. + +In contrast to DragDiffusion, our method does not require any fine-tuning of pre-trained models, which significantly reduces the response time for users. Since the generative domain of diffusion models are extremely huge and lack of LoRA fine-tuning presents a certain degree of challenge for diffusion-based dragging, we introduce a novel motion supervision method to better align with the feature space, which can get precise point tracking on diffusion models. Similarly, the edited latent often generates images that are inconsistent with the original due to the broad generation space. Built on this, we propose reference guidance during the denoising process, to preserve identity consistency with the input image. Furthermore, we design a method for auto mask generation and achieve simplified dragging interaction, which meets the need of user-friendliness. As shown in Fig. 1, the performance of DragDiffusion relies on LoRA fine-tuning and masks, while our method does not require these and achieves better performance. + +Our contributions can be summarized as follows: + +- User-friendly interaction. We introduce a point-based image editing method with auto mask generation for user-friendship, simplifying the dragging process while maintaining the comparable results. +- High performance. We propose a stable motion supervision and point-tracking method that has better adaptability to the latent space of the diffusion model to achieve precise manipulation. We also introduce reference guidance during the denoising process after the optimization of the latent code for better identity preservation. +- Effective Generalization. Based on diffusion models, our approach can handle a wide range of natural images and various types of dragging, obviously outperforming GAN-based methods. +- Extensive qualitative and quantitative results demonstrate + +that our method produces superior and more precise point dragging outcomes with clear gaps when compared to existing point-guided image editing methods. + +# 2. Related work + +# 2.1. Text-guided image editing + +With the success of GANs in the field of image generation, many previous image editing methods [1, 6, 29, 34] are based on StyleGAN. These editing approaches map real images into the latent space of StyleGAN ( $\mathcal{W}$ space or $\mathcal{W}+$ space) and then achieve image editing by manipulating the latent vectors. However, StyleGAN generates images in a highly restricted domain, and the image quality is not always high. This makes it challenging to successfully invert many real images to a satisfactory latent code. Recently, diffusion models have enabled image synthesis at high quality. An increasing number of image editing tasks [7, 17, 18] have shifted towards being based on diffusion models. Previous image editing methods that rely on diffusion models, such as SDEdit [16] and blended diffusion [2], cannot fully leverage the capabilities of accurate diffusion inversion as they introduced random noise into the input image to create a noisy initial state. Prompt-to-Prompt [7] is the first to accomplish extensive text-guided image editing without the need for diffusion model refinement, including local editing even when a known mask is unavailable. Null-text inversion [18] successfully reconstructs real images by optimizing the null-text embedding at each prediction step, which decides the unconditional prediction. Most recently, pix2pix-zero [22] proposes to learn editing directions in the textual embedding space for specific image translation tasks. In contrast to these methods, our work enables users to conduct image control through point-guided editing. + +# 2.2. Point-guided image editing + +In order to facilitate fine-grained editing, several approaches have been introduced for point-based editing, including those by [5, 21, 26, 30]. DragGAN [21] introduces an interactive method for seamless point-based image editing, incorporating two innovative components: the optimization of latent codes to progressively move multiple handle points to their desired positions and a precise point tracking mechanism to accurately follow the path of these handle points. However, the results of DragGAN are constrained by the generative capacity of StyleGAN. Recently, DragDiffusion [26] extends the editing framework of DragGAN to diffusion models with LoRA fine-tuning. Conversely, LoRA fine-tuning significantly extends the waiting time, which is unfriendly to the users. DragonDiffusion [19] utilizes an energy function to guide the editing and a memory bank for editing consistency. SDE-Drag [20] proposes an SDE-based approach to formulate the image + +![](images/dfa241d1bd2daeebaeba2ded18d81943e2b99102dee3163da5c08de54aeed2cc.jpg) +Figure 2. The pipeline of our framework. Our method generates masks automatically for precise supervision and convenience for users. To improve motion supervision over existing methods like DragGAN, we propose a stable motion supervision method based on diffusion models. The reference guidance where guidance latents are inherited from DDIM inversion preserves identity with the input image commendably. + +editing. In contrast to previous works, we propose a novel point-based image editing framework, which can generate high-quality results without LoRA fine-tuning and masks. + +# 3. Method + +# 3.1. Preliminaries + +Denoising diffusion probabilistic models (DDPM) [9] aim to map a pure noise $z_{T}$ to an output image $z_{0}$ , which is guided by the given conditioning prompt. During the training process, the network $\epsilon_{\theta}$ is updated by predicting the noise $\epsilon$ from the latent variables $z_{t}$ : + +$$ +\mathcal {L} _ {\theta} = \mathbb {E} _ {z _ {0}, \epsilon \sim N (0, I), t \sim U (1, T)} \| \epsilon - \epsilon_ {\theta} \left(z _ {t}, t, \mathcal {C}\right) \| _ {2} ^ {2}, \tag {1} +$$ + +where the noised sample $z_{t}$ is calculated by adding noise $\epsilon$ to $z_{0}$ according to diffusion step $t$ and $\mathcal{C}$ is the conditioning input of $\epsilon_{\theta}$ . During inference, we use DDIM [27] for sampling method: + +$$ +z _ {t - 1} = \sqrt {\frac {\alpha_ {t - 1}}{\alpha_ {t}}} z _ {t} + \sqrt {\alpha_ {t - 1}} \left(\sqrt {\frac {1}{\alpha_ {t - 1}} - 1} - \sqrt {\frac {1}{\alpha_ {t}} - 1}\right) \epsilon_ {\theta} \left(z _ {t}\right), \tag {2} +$$ + +where $\alpha_{t}(t = 0,1,2,\dots,T)$ represent the noise scales at corresponding diffusion steps. + +DDIM inversion. The forward process of DDIM can be expressed in terms of $\epsilon_{\theta}(z_t,t,\mathcal{C})$ , based on the assumption that the ODE process can be inverted in the limit of small steps: + +$$ +z _ {t + 1} = \sqrt {\frac {\alpha_ {t + 1}}{\alpha_ {t}}} z _ {t} + \sqrt {\alpha_ {t + 1}} \left(\sqrt {\frac {1}{\alpha_ {t + 1}} - 1} - \sqrt {\frac {1}{\alpha_ {t}} - 1}\right) \epsilon_ {\theta} \left(z _ {t}\right). \tag {3} +$$ + +Stable Diffusion. Stable Diffusion (SD) [24] presents the input image into a lower-dimensional latent space with a Variational Auto-Encoder (VAE) [15], where the input latent $z_0$ is defined as $E(x_0)$ and the output image $x_0$ is obtained by $D(z_0)$ . We implement our work based on Stable Diffusion model. + +Classifier-free guidance. The extent to which the prompt influences the reverse diffusion process directly affects the quality of image generation. For this purpose, Ho et al. [8] proposes the classifier-free guidance, which controls the degree of impact from text conditioning. Specifically, the null next-embedding $\phi = \psi("")$ is employed as a anchor point for unconditional prediction to enhance the text conditioning with guidance scale $w$ : + +$$ +\tilde {\epsilon} _ {\theta} \left(z _ {t}, t, \mathcal {C}, \phi\right) = w * \epsilon_ {\theta} \left(z _ {t}, t, \mathcal {C}\right) + (1 - w) * \epsilon_ {\theta} \left(z _ {t}, t, \phi\right). \tag {4} +$$ + +# 3.2. Overview + +Given an input image $\mathcal{I}$ , the user should input a number of handle points $\{\pmb{p}_i = (x_{p,i},y_{p,i})|i = 1,2,\dots,N\}$ and their corresponding target points $\{\pmb{q}_i = (x_{q,i},y_{q,i})|i = 1,2,\dots,N\}$ . The objective is to drag the content of the handle points to the target points based on image $\mathcal{I}$ . The user can also input a mask $\mathbf{M}$ to specify the modifiable area. + +We operate DDIM inversion with $\mathcal{C} = \psi(\mathcal{P})$ on the input image, where $\mathcal{P}$ represents the prompt, and get the intermediate results $z_{T}, \ldots, z_{t}, \ldots, z_{0}$ , where $z_{t}$ is a specific intermediate latent in DDIM inversion. Loop of optimizations is performed on $z_{t}$ to make features of the target points similar enough to that of the handle points, with novel stable motion supervision and automatic mask generation introduced in this paper. Then, the result of optimization $z_{t}^{*}$ is fed into the denoising process with reference guidance to achieve identity consistency with $\mathcal{I}$ . We present the pipeline + +![](images/f3d668311aaa51f0aeb69f1ebc1cff5fbcff5f9f1b3d6e0e1078aa414667f69a.jpg) + +![](images/d54a159f4c6ab1630366a150fa66ea2cf5a8e05e347dc5319436e1dd432abfd9.jpg) +Stable motion supervision + +![](images/639f2fb7e9195743f74fc8061af117ca9c0b4808c8d151b2610161b73c640d0c.jpg) +in Fig. 2, and elaborate three key components in the following subsections. + +![](images/ed9925cab50f09aa9096fa4664bfcd4564a964acad549fadab34158ec8c1fe16.jpg) + +![](images/6a69fc3d968cdbec1eadcbf9175becf4f5bd554278de73de8b095f8986cc8d68.jpg) +DragGAN motion supervision + +![](images/e19e33f2a65b13d7178bec8d57fd56072e6d2beb6ae6c05d0f162d6608140ba8.jpg) +Figure 3. Stable motion supervision. When the next point $p^{**}$ is more similar with the handle point $p$ than $p^*$ , the handle point has moved to the position of $p^{**}$ . Our approach differs significantly from DragGAN in that our method always learns features from the original image, while DragGAN learns features that are continually changing. + +![](images/905e240854c90ac9f5e223705c101966b6b722bc645151f810788c05afaf3516.jpg) + +![](images/fc42870db319f82e2d0272956996e6e0b1b894a3cd2c89fea997bcaac6a83168.jpg) + +# 3.3. Stable motion supervision + +There exists misleading in DragGAN due to the proposed motion supervision. In contrast to DragGAN, we introduce an efficient drag method that is more suitable for diffusion models. Inspired by DIFT [28], we utilize the feature maps of U-Net denoiser as the semantic information of the input image. Specially, we consider the second and third upper layers of U-Net with $(z_{t}, t, \mathcal{C})$ as inputs, and concatenate the two layers as our feature maps $\mathbf{F}$ used for stable motion supervision. $\mathbf{F}$ is resized to have the same resolution as the input image in order to get precise manipulation. + +As shown in Fig. 3, $p^*$ represents current position of the handle point, and $p^{**}$ is the next position that we want to drag to. The method for calculating $p^{**}$ is presented in Algorithm 1. We start from $p^*$ , move along the direction of $q$ by $K$ units, and then find the nearest integer coordinates to determine $p^{**}$ . In contrast to DragGAN, we employ $p^*$ to learn features of the original image at the $p$ position, ensuring that the learned features are consistently accurate. Inspired by DragonDiffusion [19], we define the similarity loss function of two feature maps as follows: + +$$ +\mathcal {L} _ {\text {c o s}} \left(\mathbf {F}, \mathbf {F} _ {0}\right) = \frac {1}{1 + \frac {1 + \cos \left(\mathbf {F} , s g \left(\mathbf {F} _ {0}\right)\right)}{2}}, \tag {5} +$$ + +where $\mathbf{F}_0$ represents the feature maps corresponding to the input image and $sg(\cdot)$ is the stop gradient operator. Therefore, the stable motion supervision loss is defined as: + +# Algorithm 1: Next point + +Input: Current point $p^*$ , Target point $q$ + +Output: Next point $p^{**}$ + +1 if $\| \pmb {q} - \pmb{p}^{*}\|_{2} < K$ then +2 $\pmb{p}^{**}\gets \pmb {q};$ +3 else +4 $\pmb{d} \gets \frac{\pmb{q} - \pmb{p}^{*}}{\|\pmb{q} - \pmb{p}^{*}\|_{2}};$ +5 $\pmb{p}^{**}\gets \pmb{p}^* +round(\pmb {d}*\pmb {K})$ +6 end +7 return $p^{**}$ + +$$ +\begin{array}{l} \mathcal {L} = \sum_ {i = 1} ^ {N} \mathcal {L} _ {\text {c o s}} \left(\mathbf {F} \left(\boldsymbol {p} _ {i} ^ {* *}\right), \mathbf {F} _ {0} \left(\boldsymbol {p} _ {i}\right)\right) \tag {6} \\ + \lambda \mathcal {L} _ {\cos} (\mathbf {F} \odot (1 - \mathbf {M}), \mathbf {F} _ {0} \odot (1 - \mathbf {M})), \\ \end{array} +$$ + +where $\mathbf{F}(\pmb{p}_i^{**})$ denotes the feature values of $\mathbf{F}$ at pixel $\pmb{p}_i^{**}$ . + +After motion supervision, the updated latent code $z_{t}^{*}$ may change the position of handle points. Consequently, we need to determine whether the handle point has reached the next target point. To do this, we compare which one, $\mathbf{F}(\pmb{p}_i^{**})$ or $\mathbf{F}(\pmb{p}_i^*)$ , resembles $\mathbf{F}_0(\pmb{p}_i)$ more. If $\cos(\mathbf{F}(\pmb{p}_i^{**}), \mathbf{F}_0(\pmb{p}_i))$ is higher, the handle point $\pmb{p}_i^*$ is updated to $\pmb{p}_i^{**}$ and $\pmb{p}_i^{**}$ is further updated to $\text{next}(\pmb{p}_i^*, \pmb{q}_i)$ according to Algorithm 1. + +# 3.4. Auto mask generation + +For user-friendly interaction, image editing should avoid burdening users with the task of drawing masks. Existing drag methods based on diffusion models [19, 26] always rely on mask inputs, which are provided by users. This requires users to have a deep understanding of the drag method because an incorrectly drawn mask can lead to unpredictable outcomes. Consequently, we propose a method that can automatically generate masks, which are able to change adaptively during the training process. + +In the process of motion supervision, the parts with larger gradients of $z_{t}^{*}$ have a greater impact on the results, while those with smaller gradients are generally the parts that are expected to remain stationary. However, if the regions with relatively small gradients keep changing, the final result becomes unpredictable. Due to the self-attention layer in U-Net, minor changes made at step $t$ are very likely to have an unacceptable impact on the final image. Therefore, we still need control over these low-impact areas. We achieve this by using regions with normalized gradients greater than a threshold $g$ as the drag mask $\mathbf{M}$ . + +Every time $\pmb{p}_i^*$ is updated, the hot-zone of $z_t$ also changes due to the position shifting of the drag point. We update our dynamic mask $\mathbf{M}$ each time $\pmb{p}_i^*$ changes, but there is no need to change the mask when $\pmb{p}_i^*$ remains the same. This is because for the same $\pmb{p}_i^*$ and $\pmb{p}_i^{**}$ , the hot-zone is consistent. + +![](images/1076bdb83fba895d1d28cc783296bd895605188cbdede29e7e6e1f0d5649c4e0.jpg) +(a) +Figure 4. Qualitative comparison of our approach to DragDiffusion. Due to the misleading of motion supervision, DragDiffusion always cannot move the handle points to the target. Furthermore, DragDiffusion often exhibits significant background changes in the absence of the mask. In our method, the handle points reach the target positions precisely even without the mask. + +![](images/fecf87bf8c74650247a2579d8e28cd4520b7544482a98fd6e4ace7b433ab62a4.jpg) +(b) + +This way, $\mathbf{M}$ adapts to changes of $z_{t}$ , aligning better with the dragging requirements. + +# 3.5. Reference guidance + +With optimized $z_{i}^{*}$ , the generated images with direct DDIM denoising often exhibit identity changes and a decrease in image quality. Especially in areas where the pretrained models have limited generation capabilities, naively generating with $z_{i}$ after DDIM Inversion may not even reproduce the original image. Drawing inspiration from [4, 19, 26], we employ mutual self-attention to ensure that images generated from $z_{i}^{*}$ maintain content consistency with the original image. + +As illustrated in Fig. 2, we replace the key and value of self-attention layers in backward process of $z_{i}^{*}$ with the corresponding key and value in the process of reconstructing the source image $\mathcal{I}$ . Due to the feature correlations in lower layers being relatively weak, the mutual self-attention is only applied to the upper layers of U-Net $\epsilon_{\theta}$ . We then perform sampling using the linear combination of the reference guidance and normal prediction: + +$$ +\epsilon_ {i} ^ {*} \leftarrow w \epsilon_ {\theta} \left(z _ {i} ^ {*}, z _ {i}, i, \mathcal {C}\right) + (1 - w) \epsilon_ {\theta} \left(z _ {i} ^ {*}, i, \mathcal {C}\right), \tag {7} +$$ + +where $w$ controls the strength of the reference guidance. Specifically, $z_{i}$ is the corresponding latent during DDIM inversion, rather than that generated from $z_{t}$ . When $w = 1$ , our reference guidance is the same as mutual self-attention. + +However, in the cases where the input image is infrequent for the model and $z_{t}$ undergoes significant changes, using mutual self-attention alone cannot guarantee identity consistency. When we increase the strength of reference guidance, the generated results can better ensure content and texture consistency with the input image. + +# 4. Experiments + +In this section, we evaluate the proposed method qualitatively and quantitatively, on both synthetic and real images. + +Implementation details. In our experiments, except for generated image manipulations, all tasks are based on Stable Diffusion 1.5 (SD-1.5). Unless otherwise specified, in both DDIM inversion and sampling we set the number of the sampling steps to be 50 via using the stride of 20 over 1000 diffusion steps. We set the latent from the 35-th DDIM step, which contains rich semantic information corresponding to the input image, as $z_{t}$ . The gradient threshold $g$ is 0.4, the motion step $K$ is 12 and the reference guidance scale $w$ is 4 by default. We use a learning rate of 0.01 for $z_{t}$ without any decay schedule and train the latent with Adam [14] solver. + +Dataset. We employ DragBench introduced in DragDiffusion as our evaluation dataset as it encompasses various types of images, and we also follow its dragging instruc + +![](images/5600e1d001a83d73f3ccbbaf5344749d3b16e5a066b08eeab0be254b326ecff7.jpg) +(a) +Inputs + +![](images/94692cfe6867f37db3afe49edefa00e535c70d570a67d5591df9db012e5a34cc.jpg) +Results + +![](images/c0cc08965b72012715ec0e2c0b139069777c7d8089f74898e00ab1ecd10202c7.jpg) +Inputs + +![](images/f09f5170ed21fe8ef65d342bdc256034f5e9d4a4a100a8680e79d03b2d9d0616.jpg) +Results + +![](images/5074d6b041ecf4089806b68f0aefaa98eb48adf0105b68b3fad6f1f945705eff.jpg) +Inputs + +![](images/3446da27be0e0bbce2cb970163d8dc19becfa3be5f7d4689d829fde8f713939d.jpg) +Results + +![](images/e82cac6e9863c552c63f110feeb9f59fc34480a9f388efde9ff8fea656c7d9ab.jpg) +(b) + +![](images/afc20d99245a8cd0a328a8d41f8e7b520a61e499875512cf8fcbbf8905793434.jpg) + +![](images/5988b24870a3319ccb873c0a00c4c3542e0fb0be088ad618c0b81cdb429d0788.jpg) + +![](images/eabd01accd354e303531c596f04c79c0e24c936c36eee532807369438a159577.jpg) + +![](images/bc2a6496f40586416aa881955896a511e1c83c7d1fea9f1d1f8ac2d3865acee9.jpg) + +![](images/208994933d984a2212dce4e266454ba0d09aa49c087e60b0652ecba0a90a1e77.jpg) + +![](images/377c7ab307cea6c20f1bf9095a19fa823139f0e0b6bd924797ee53eae72eaba1.jpg) +(c) + +![](images/127c7a07f9e408808d573cf6f08f03ce3ed21872e6aef5b9c6add66f02f3a71f.jpg) + +![](images/759abe29efba09868916adc3c7f935ae4e339085e06d4d5cdfbfd1785723fbfc.jpg) + +![](images/0434fa00e25c242a2976c474351d1f06a5d5fcfd5a701762f7b754a460159ebf.jpg) + +![](images/62da21abc77b0febb11056941ab8203c407a84fa658876a306fc125c7cceb8d4.jpg) + +![](images/757ef7d9241db58cc1cad54919191c5b90f577826ca6ca52bf55d6b867b42a93.jpg) + +![](images/f6be91b9b38b006e8d8e42bc5fc7aeca8739f1e8240f04fc0ca1ef35f8da9fb3.jpg) +(d) +Figure 5. Qualitative results of our approach on generated images with SD-1.5 and its fine-tuned variants. (a) SD-1.5. (b) Majicmix Realistic V7. (c) Interior Design Supermix. (d) Dvarch. + +![](images/8646d15d7956bafba38430089a1f6ce4717d9221687f8fb2298de12274fe2f3d.jpg) + +![](images/91b40614ca1168e551f12fdea0dd87e2c6a3f1ce524c2701bff173db947695c3.jpg) + +![](images/cb0f7c8c84c5387800ec55ef6fa35fb245fc4927586a4eef2b516291155386fc.jpg) + +![](images/5829854f416da08cf0b2343ab41c3cb501926da7c78ff9d3409f2fe80a1802ad.jpg) + +![](images/6566911f56d82707fd59c1013b00d85363f7c22a5a0166bb0d5eecf8180ca9e3.jpg) + +tions. DragBench contains 205 images, with well-annotated prompts, masks, and dragging points. + +Evaluation metrics. We evaluate the dragging results from two aspects: the quality of dragged image and the accuracy of position of handle points. We employ Image Fidelity (IF) [13] and Mean Distance (MD) [21] to assess these two aspects, respectively. IF evaluates the similarity between the input and output images, which is calculated by subtracting the mean LPIPS [32] over all pairs of original and edited images from 1. MD quantifies how well the handle points move to the target points. The same as DragDiffusion, we utilize DIFT [28] to identify the points in the output image that correspond to the handle points in the input image. These identified points are then treated as the definitive handle points after the dragging process. Subsequently, MD is calculated as the average Euclidean distance between the positions of all target points and their corresponding final handle points. A higher IF and a lower MD indicate a better quality of the dragging approach. + +# 4.1. Qualitative evaluation + +We conduct extensive experiments for the qualitative comparison between our method and DragDiffusion. To provide a thorough evaluation, we adopt SD-1.5 [24] as our base model for all real images, and also employ other fine-tuned variants of SD-1.5 for generated image manipulation. + +We present real image editing results for various object categories and user inputs in Fig. 4. Our approach accurately moves the handle points to reach the target positions, resulting in a wide range of natural and diverse manipulation effects such as altering the poses of animals, rearranging landscapes, modifying human body parts, and reshaping objects, in the absence of the mask. Conversely, DragDiffusion is unable to precisely move the handle points to the target positions, which is mainly attributed to the misleading of point tracking that is employed in DragGAN and DragDiffusion. It is worth noting that DragDiffusion always changes the non-dragged regions when no mask is provided. + +We also conduct a series of point manipulations based on the generated images, as shown in Fig. 5. The classifier-free guidance scale is configured at 7.5 for various SD-1.5 variants. Notably, the reference guidance is unnecessary for generative editing, as the prompt of generated images pro + +![](images/f07b53341e22f2521951b6396f31e03c7b5aae1818e914cc080cdfa8dc9d40ac.jpg) +Figure 6. Out-of-distribution manipulations. In some cases, the input image cannot be reconstructed with DDIM Inversion. Utilizing reference guidance, edited image preserves its identity with the original image. + +
MethodIF(↑)MD(↓)Time(s)
DragGAN0.69557.359.1
DragDiffusion0.88236.7169.4
DragDiffusion w/o LoRA0.86353.6333.7
DragDiffusion w/o mask0.76647.0365.6
SDE-Drag0.87145.4450.1
Ours w/ mask0.88929.0727.3
Ours0.91032.5530.6
+ +Table 1. Quantitative evaluation of different methods with Drag-Bench. Evaluation metrics include Image Fidelity (IF) and Mean Distance (MD). + +vides substantial guidance during the denoising process. As shown in Fig. 5, our method achieves point-based editing with high quality and precision on generated images. It can be observed that in some cases, the details of generated images have changed. This is because the self-attention layer in diffusion models may introduce slight variation in other parts when a specific area of the image is edited. If users want to preserve certain areas without any changes, they can employ inpainting during image generation with masks to determine which regions should remain unchanged. + +# 4.2. Quantitative evaluation + +We conduct quantitative evaluation by comparing our method to the baseline dragging method on diffusion, DragDiffusion. The evaluation, which incorporates IF and MD metrics, makes use of the DragBench dataset provided by DragDiffusion. The results, as shown in Tab. 1, demonstrate that our approach significantly surpasses DragDiffusion and SDE-Drag on the DragBench dataset. Notably, our method achieves superior dragging results without requiring LoRA fine-tuning, offering substantial time savings during extensive experiments. The performance of DragDiffusion degrades significantly without LoRA nor mask. The execution time in Tab. 1 is calculated on a single A100 GPU. + +![](images/fc6519bc1e51dec6c62cc45781aa2504cdd8c46afda07d63298c44221fff5efd.jpg) +Figure 7. When users do not provide an appropriate mask, the generated results are prone to issues. + +# 4.3. Discussions + +Reference Guidance. With our reference guidance, we enable the diffusion models to perform out-of-distribution manipulations. As shown in Fig. 6, the original image cannot be faithfully reconstructed using DDIM Inversion. When we introduce reference guidance, $z_{t}$ can better reconstruct the input image. It is worth noting that $z_{i}(i = t,t - 1,\dots,1)$ is obtained during the DDIM Inversion process, so it contains richer information compared to $z_{t}$ . After editing $z_{t}$ , the image generated from $z_{t}^{*}$ is very different from the input image in terms of identity. As we increase the reference guidance scale $w$ , the consistency between the generated image and original image gradually improves. However, if $w$ is too large, the generated image may become oversaturated. In our experiments, a value for $w$ in the range of 3 to 5 is considered a good choice. + +Auto mask generation. When users create imprecise masks, the results are unpredictable. As shown in Fig. 7, when the mask is too small, it will restrict the editing of the image. On the other hand, when the mask is too large, changes may occur in the background. Therefore, an appropriate mask is necessary for point-based image editing. Our + +![](images/67076ba86c73d6678f9264f69f5a21174c6b306f5836f284cd78a169a1750322.jpg) +Figure 8. The changes in the gradient and prediction of $z_{t}$ during optimization. The term of "K-th step" means the K-th update of handle points $p^{*}$ . + +![](images/7ca4e4504b8eb48648883887a7a59122d81e3212d6946843369e789a264b6911.jpg) +(a) Which upper layer to use. +Figure 9. Ablation study. + +![](images/c8a4132df9a3745e3f02c9e7d0aa42f6cb2d6e0e82b1c865b92aa648ef6d3cc2.jpg) +(b) Effects of K. + +auto mask generation method liberates users from the need to manually draw masks, which is user-friendly, especially for novice users. + +Figure 8 shows the update of gradient and prediction with the optimization of $z_{t}$ . It can be observed that when the handle points update, the gradient of $z_{t}$ will change simultaneously. In the K-th step, the whole arm of the robot needs to be lifted, so the gradients for the arm part are relatively large. However, by the time we reach the (K+1)-th step, the robot's arm has already been dragged into the right place, and only the palm part needs to extend outward. Therefore, only the gradients for the palm part are high. Our auto mask effectively captures these details and incorporates this variation into the updates of $z_{t}$ , thereby providing better control over the direction of $z_{t}$ . That is the reason why the IF metric of without mask is higher than that with mask in Tab. 1. + +Ablation study. We perform ablation studies to analyze the impact of the feature used in stable motion supervision, and we evaluate the performance on DragBench, as shown in Fig. 9a. Notably, the concatenated feature maps from upper layers 1&2 exhibit the best performance, demonstrating an optimal balance between resolution and discriminativeness. Additionally, we investigate the effects of different values of $K$ in Fig. 9b, and it is observed that $K = 12$ yields better results. + +
MethodsIF(↑)MD(↓)
DragGAN0.85651.32
DragonDiffusion0.89952.21
Stable motion supervision0.88929.07
+ +Table 2. Quantitative comparison of motion supervision between our method and previous works. Our approach achieves more accurate dragging results than other methods. + +![](images/20d93ef55d4de00d352344e33e002f8dce81dca5326e5fc65bcfc337468cc134.jpg) +Figure 10. When the manipulated portion in the image is very small or the dragging distance is extensive, it is challenging for the dragging to have a noticeable effect. + +A comparison of "drag" method between our approach and previous methods is shown in Tab. 2. We only replace stable motion supervision with the "drag" method in other works. It can be found that stable motion supervision is more suitable for the feature space of diffusion models, although DragonDiffusion achieves slightly higher IF. + +Limitations. Despite the outstanding performance, our method has some limitations. Particularly, when the editing area is too small or the dragging distance is too long, the dragging result may be unsatisfactory. As depicted in Fig. 10, when the girl's face occupies only a small part of the image, and we intend to edit the mouth, the desired changes of mouth do not occur as expected. + +# 5. Conclusion + +In this work, we attribute an excellent point-based image editing method to three criteria: user-friendly interaction, good performance, and effective generalization. We abandon the LoRA fine-tuning and masks, which is convenient for users. Furthermore, we propose a stable motion supervision method and reference guidance to generate high-quality results based on diffusion models. Leveraging the powerful generative capabilities of the diffusion model, we achieve impressive dragging results across various domains. However, in our experiments, we observe that for very small-scale manipulations or long-distance dragging in images, the results are often not satisfactory. Diffusion models with stronger generative capabilities may be needed to achieve a broader range of dragging manipulations. + +Acknowledgements: This work is partially supported by Natural Science Foundation of China Grant 41930110. + +# References + +[1] Rameen Abdal, Peihao Zhu, Niloy J Mitra, and Peter Wonka. Styleflow: Attribute-conditioned exploration of stylegan-generated images using conditional continuous normalizing flows. ACM Transactions on Graphics (ToG), 40(3):1-21, 2021. 2 +[2] Omri Avrahami, Dani Lischinski, and Ohad Fried. Blended diffusion for text-driven editing of natural images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18208-18218, 2022. 2 +[3] Tim Brooks, Aleksander Holynski, and Alexei A Efros. Instructpix2pix: Learning to follow image editing instructions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18392-18402, 2023. 1 +[4] Mingdeng Cao, Xintao Wang, Zhongang Qi, Ying Shan, Xiaohu Qie, and Yinqiang Zheng. Masactrl: Tuning-free mutual self-attention control for consistent image synthesis and editing. arXiv preprint arXiv:2304.08465, 2023. 5 +[5] Yuki Endo. User-controllable latent transformer for stylegan image layout editing. In Computer Graphics Forum, pages 395-406. Wiley Online Library, 2022. 2 +[6] Dave Epstein, Taesung Park, Richard Zhang, Eli Shechtman, and Alexei A Efros. Blobgan: Spatially disentangled scene representations. In European Conference on Computer Vision, pages 616-635. Springer, 2022. 2 +[7] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-prompt image editing with cross attention control. arXiv preprint arXiv:2208.01626, 2022. 1, 2 +[8] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598, 2022. 3 +[9] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 1, 2, 3 +[10] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 2 +[11] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4401-4410, 2019. 1 +[12] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of stylegan. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8110-8119, 2020. 1 +[13] Bahjat Kawar, Shiran Zada, Oran Lang, Omer Tov, Huiwen Chang, Tali Dekel, Inbar Mosseri, and Michal Irani. Imagic: Text-based real image editing with diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6007-6017, 2023. 1, 6 +[14] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014.5 + +[15] Diederik P Kingma and Max Welling. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114, 2013. 3 +[16] Chenlin Meng, Yutong He, Yang Song, Jiaming Song, Jia-jun Wu, Jun-Yan Zhu, and Stefano Ermon. Sdedit: Guided image synthesis and editing with stochastic differential equations. arXiv preprint arXiv:2108.01073, 2021. 2 +[17] Daiki Miyake, Akihiro Iohara, Yu Saito, and Toshiyuki Tanaka. Negative-prompt inversion: Fast image inversion for editing with text-guided diffusion models. arXiv preprint arXiv:2305.16807, 2023. 2 +[18] Ron Mokady, Amir Hertz, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Null-text inversion for editing real images using guided diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6038–6047, 2023. 1, 2 +[19] Chong Mou, Xintao Wang, Jiechong Song, Ying Shan, and Jian Zhang. Dragondiffusion: Enabling drag-style manipulation on diffusion models. arXiv preprint arXiv:2307.02421, 2023. 2, 4, 5 +[20] Shen Nie, Hanzhong Allan Guo, Cheng Lu, Yuhao Zhou, Chenyu Zheng, and Chongxuan Li. The blessing of randomness: Sde beats ode in general diffusion-based image editing. arXiv preprint arXiv:2311.01410, 2023. 2 +[21] Xingang Pan, Ayush Tewari, Thomas Leimkuhler, Lingjie Liu, Abhimitra Meka, and Christian Theobalt. Drag your gan: Interactive point-based manipulation on the generative image manifold. In ACM SIGGRAPH 2023 Conference Proceedings, pages 1-11, 2023. 1, 2, 6 +[22] Gaurav Parmar, Krishna Kumar Singh, Richard Zhang, Yijun Li, Jingwan Lu, and Jun-Yan Zhu. Zero-shot image-to-image translation. In ACM SIGGRAPH 2023 Conference Proceedings, pages 1-11, 2023. 2 +[23] Daniel Roich, Ron Mokady, Amit H Bermano, and Daniel Cohen-Or. Pivotal tuning for latent-based editing of real images. ACM Transactions on graphics (TOG), 42(1):1-13, 2022. 1 +[24] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10684-10695, 2022. 1, 2, 3, 6 +[25] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in Neural Information Processing Systems, 35:36479-36494, 2022. 1 +[26] Yujun Shi, Chuhui Xue, Jiachun Pan, Wenqing Zhang, Vincent YF Tan, and Song Bai. Dragdiffusion: Harnessing diffusion models for interactive point-based image editing. arXiv preprint arXiv:2306.14435, 2023. 2, 4, 5 +[27] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 1, 2, 3 +[28] Luming Tang, Menglin Jia, Qianqian Wang, Cheng Perng Phoo, and Bharath Hariharan. Emergent correspondence from image diffusion. arXiv preprint arXiv:2306.03881, 2023.4,6 + +[29] Jianyuan Wang, Ceyuan Yang, Yinghao Xu, Yujun Shen, Hongdong Li, and Bolei Zhou. Improving gan equilibrium by raising spatial awareness. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11285-11293, 2022. 2 +[30] Sheng-Yu Wang, David Bau, and Jun-Yan Zhu. Rewriting geometric rules of a gan. ACM Transactions on Graphics (TOG), 41(4):1-16, 2022. 2 +[31] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3836-3847, 2023. 1 +[32] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 6 +[33] Zhixing Zhang, Ligong Han, Arnab Ghosh, Dimitris N Metaxas, and Jian Ren. Sine: Single image editing with text-to-image diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6027-6037, 2023. 1 +[34] Jiapeng Zhu, Ceyuan Yang, Yujun Shen, Zifan Shi, Bo Dai, Deli Zhao, and Qifeng Chen. Linkgan: Linking gan latents to pixels for controllable image synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7656-7666, 2023. 2 \ No newline at end of file diff --git a/2024/EasyDrag_ Efficient Point-based Manipulation on Diffusion Models/images.zip b/2024/EasyDrag_ Efficient Point-based Manipulation on Diffusion Models/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..e57afd074d038e9d8adc40f8ef8c6aea505cf7ff --- /dev/null +++ b/2024/EasyDrag_ Efficient Point-based Manipulation on Diffusion Models/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d8fc28d95ffbb7eda5c7d09601aa3d2bbfefbb7aff73974533cff5251650b64b +size 701947 diff --git a/2024/EasyDrag_ Efficient Point-based Manipulation on Diffusion Models/layout.json b/2024/EasyDrag_ Efficient Point-based Manipulation on Diffusion Models/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..09edbdd595f76c67d026f199dce31976cf999819 --- /dev/null +++ b/2024/EasyDrag_ Efficient Point-based Manipulation on Diffusion Models/layout.json @@ -0,0 +1,10280 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 89, + 103, + 505, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 103, + 505, + 121 + ], + "spans": [ + { + "bbox": [ + 89, + 103, + 505, + 121 + ], + "type": "text", + "content": "EasyDrag: Efficient Point-based Manipulation on Diffusion Models" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 141, + 143, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 141, + 143, + 157 + ], + "spans": [ + { + "bbox": [ + 53, + 141, + 143, + 157 + ], + "type": "text", + "content": "Xingzhong Hou1,2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 168, + 143, + 228, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 168, + 143, + 228, + 156 + ], + "spans": [ + { + "bbox": [ + 168, + 143, + 228, + 156 + ], + "type": "text", + "content": "Boxiao Liu3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 254, + 143, + 305, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 254, + 143, + 305, + 157 + ], + "spans": [ + { + "bbox": [ + 254, + 143, + 305, + 157 + ], + "type": "text", + "content": "Yi Zhang3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 329, + 143, + 388, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 329, + 143, + 388, + 156 + ], + "spans": [ + { + "bbox": [ + 329, + 143, + 388, + 156 + ], + "type": "text", + "content": "Jihao Liu3,4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 412, + 143, + 451, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 412, + 143, + 451, + 156 + ], + "spans": [ + { + "bbox": [ + 412, + 143, + 451, + 156 + ], + "type": "text", + "content": "Yu Liu3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 476, + 143, + 549, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 476, + 143, + 549, + 157 + ], + "spans": [ + { + "bbox": [ + 476, + 143, + 549, + 157 + ], + "type": "text", + "content": "Haihang You1*" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 67, + 158, + 534, + 198 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 158, + 534, + 198 + ], + "spans": [ + { + "bbox": [ + 67, + 158, + 534, + 198 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 67, + 158, + 534, + 198 + ], + "type": "text", + "content": "State Key Lab of Processors, Institute of Computing Technology, Chinese Academy of Sciences " + }, + { + "bbox": [ + 67, + 158, + 534, + 198 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 67, + 158, + 534, + 198 + ], + "type": "text", + "content": "School of Computer Science and Technology, University of Chinese Academy of Sciences " + }, + { + "bbox": [ + 67, + 158, + 534, + 198 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 67, + 158, + 534, + 198 + ], + "type": "text", + "content": "SenseTime Research " + }, + { + "bbox": [ + 67, + 158, + 534, + 198 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 67, + 158, + 534, + 198 + ], + "type": "text", + "content": "CUHK MMLab" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 99, + 201, + 501, + 227 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 99, + 201, + 501, + 227 + ], + "spans": [ + { + "bbox": [ + 99, + 201, + 501, + 227 + ], + "type": "text", + "content": "{houxingzhong, youhaihang}@ict.ac.cn {liuboxiao, zhangyi17}@sensetime.com jihaoliu@link.cuhk.edu.hk liuyuisanai@gmail.com" + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 74, + 236, + 521, + 361 + ], + "blocks": [ + { + "bbox": [ + 74, + 236, + 521, + 361 + ], + "lines": [ + { + "bbox": [ + 74, + 236, + 521, + 361 + ], + "spans": [ + { + "bbox": [ + 74, + 236, + 521, + 361 + ], + "type": "image", + "image_path": "be7aa8bb33edfafc4f6d1777fbe28b052deb88c8d41da4902d58fbb7958ebe0c.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 369, + 547, + 404 + ], + "lines": [ + { + "bbox": [ + 46, + 369, + 547, + 404 + ], + "spans": [ + { + "bbox": [ + 46, + 369, + 547, + 404 + ], + "type": "text", + "content": "Figure 1. DragGAN fails in both cases due to limited model capacity. Similarly, DragDiffusion relies on LoRA fine-tuning and hand-drawn masks to achieve high-quality results, and SDE-Drag also fails to address these two cases effectively. In contrast, our approach ensures precise manipulation and detail preservation without fine-tuning or masks." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 143, + 426, + 192, + 438 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 426, + 192, + 438 + ], + "spans": [ + { + "bbox": [ + 143, + 426, + 192, + 438 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 46, + 452, + 289, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 452, + 289, + 693 + ], + "spans": [ + { + "bbox": [ + 46, + 452, + 289, + 693 + ], + "type": "text", + "content": "Generative models are gaining increasing popularity, and the demand for precisely generating images is on the rise. However, generating an image that perfectly aligns with users' expectations is extremely challenging. The shapes of objects, the poses of animals, the structures of landscapes, and more may not match the user's desires, and this applies to real images as well. This is where point-based image editing becomes essential. An excellent image editing method needs to meet the following criteria: user-friendly interaction, high performance, and good generalization capability. Due to the limitations of StyleGAN, DragGAN exhibits limited robustness across diverse scenarios, while DragDiffusion lacks user-friendliness due to the necessity of LoRA fine-tuning and masks. In this paper, we introduce a novel interactive point-based image editing framework, called EasyDrag, that leverages pretrained diffusion models to achieve high-quality editing outcomes and user-friendship. Extensive experimentation demonstrates that our approach surpasses DragDiffusion in terms of both image quality and editing precision for point-based" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 426, + 545, + 451 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 426, + 545, + 451 + ], + "spans": [ + { + "bbox": [ + 306, + 426, + 545, + 451 + ], + "type": "text", + "content": "image manipulation tasks. The code will be available on https://github.com/Ace-Pegasus/EasyDrag." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 476, + 387, + 488 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 476, + 387, + 488 + ], + "spans": [ + { + "bbox": [ + 306, + 476, + 387, + 488 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 497, + 545, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 497, + 545, + 605 + ], + "spans": [ + { + "bbox": [ + 304, + 497, + 545, + 605 + ], + "type": "text", + "content": "In the past few years, diffusion models [9, 24, 27] represent a groundbreaking advancement in the domain of generative modeling. In particular, text-to-image diffusion models [24, 25, 31] can produce remarkable images based on textual prompts. Recently, such models have also been extended for image editing [3, 7, 13, 18, 23, 33]. Since diffusion models depend on textual input, image elements, including objects and styles, can be altered with a high degree of quality and diversity guided by edited prompts." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "type": "text", + "content": "However, text-guided image editing cannot precisely direct changes in the image, and it is unpredictable for a specific point in the image that one might want to move. Therefore, point-guided image editing becomes essential. A recent study DragGAN [21] introduces \"drag\" operation for precise image editing, which has received widespread attention for its point-based interaction. Nonetheless, it encounters limitations tied to the capacity and adaptability of StyleGAN [11, 12] as shown in Fig. 1. In contrast, diffu" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 702, + 135, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 702, + 135, + 712 + ], + "spans": [ + { + "bbox": [ + 58, + 702, + 135, + 712 + ], + "type": "text", + "content": "*Corresponding author." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8404" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 287, + 131 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 287, + 131 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 287, + 131 + ], + "type": "text", + "content": "sion models [9, 24, 27] exhibit better stability and superior generative quality. DragDiffusion [26] enables interactive point-based image editing on diffusion models with LoRA fine-tuning [10] but introduces high time consumption at the same time." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 133, + 287, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 133, + 287, + 300 + ], + "spans": [ + { + "bbox": [ + 46, + 133, + 287, + 300 + ], + "type": "text", + "content": "What capabilities should point-based image editing possess? User-friendly interaction, good performance and effective generalization. User-friendly interaction implies ease of use and short response times for users, while good performance refers to precise dragging with maintaining identity. Effective generalization requires the method to work with a single model for various images. Unfortunately, existing methods have not been able to simultaneously exhibit these capabilities. DragGAN lacks generalization of dragging general domain image, while DragDiffusion is less user-friendly due to its reliance on LoRA fine-tuning and masks. In this paper, we propose a novel point-based image editing framework built on diffusion models, which can concurrently fulfill these requirements." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 301, + 287, + 516 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 301, + 287, + 516 + ], + "spans": [ + { + "bbox": [ + 46, + 301, + 287, + 516 + ], + "type": "text", + "content": "In contrast to DragDiffusion, our method does not require any fine-tuning of pre-trained models, which significantly reduces the response time for users. Since the generative domain of diffusion models are extremely huge and lack of LoRA fine-tuning presents a certain degree of challenge for diffusion-based dragging, we introduce a novel motion supervision method to better align with the feature space, which can get precise point tracking on diffusion models. Similarly, the edited latent often generates images that are inconsistent with the original due to the broad generation space. Built on this, we propose reference guidance during the denoising process, to preserve identity consistency with the input image. Furthermore, we design a method for auto mask generation and achieve simplified dragging interaction, which meets the need of user-friendliness. As shown in Fig. 1, the performance of DragDiffusion relies on LoRA fine-tuning and masks, while our method does not require these and achieves better performance." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 58, + 518, + 257, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 518, + 257, + 529 + ], + "spans": [ + { + "bbox": [ + 58, + 518, + 257, + 529 + ], + "type": "text", + "content": "Our contributions can be summarized as follows:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 534, + 286, + 712 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 47, + 534, + 286, + 580 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 534, + 286, + 580 + ], + "spans": [ + { + "bbox": [ + 47, + 534, + 286, + 580 + ], + "type": "text", + "content": "- User-friendly interaction. We introduce a point-based image editing method with auto mask generation for user-friendship, simplifying the dragging process while maintaining the comparable results." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 582, + 286, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 582, + 286, + 652 + ], + "spans": [ + { + "bbox": [ + 47, + 582, + 286, + 652 + ], + "type": "text", + "content": "- High performance. We propose a stable motion supervision and point-tracking method that has better adaptability to the latent space of the diffusion model to achieve precise manipulation. We also introduce reference guidance during the denoising process after the optimization of the latent code for better identity preservation." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 654, + 286, + 699 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 654, + 286, + 699 + ], + "spans": [ + { + "bbox": [ + 47, + 654, + 286, + 699 + ], + "type": "text", + "content": "- Effective Generalization. Based on diffusion models, our approach can handle a wide range of natural images and various types of dragging, obviously outperforming GAN-based methods." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 702, + 286, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 702, + 286, + 712 + ], + "spans": [ + { + "bbox": [ + 47, + 702, + 286, + 712 + ], + "type": "text", + "content": "- Extensive qualitative and quantitative results demonstrate" + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 313, + 72, + 545, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 545, + 108 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 545, + 108 + ], + "type": "text", + "content": "that our method produces superior and more precise point dragging outcomes with clear gaps when compared to existing point-guided image editing methods." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 121, + 391, + 133 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 121, + 391, + 133 + ], + "spans": [ + { + "bbox": [ + 306, + 121, + 391, + 133 + ], + "type": "text", + "content": "2. Related work" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 140, + 450, + 153 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 140, + 450, + 153 + ], + "spans": [ + { + "bbox": [ + 306, + 140, + 450, + 153 + ], + "type": "text", + "content": "2.1. Text-guided image editing" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 159, + 545, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 159, + 545, + 483 + ], + "spans": [ + { + "bbox": [ + 304, + 159, + 545, + 483 + ], + "type": "text", + "content": "With the success of GANs in the field of image generation, many previous image editing methods [1, 6, 29, 34] are based on StyleGAN. These editing approaches map real images into the latent space of StyleGAN (" + }, + { + "bbox": [ + 304, + 159, + 545, + 483 + ], + "type": "inline_equation", + "content": "\\mathcal{W}" + }, + { + "bbox": [ + 304, + 159, + 545, + 483 + ], + "type": "text", + "content": " space or " + }, + { + "bbox": [ + 304, + 159, + 545, + 483 + ], + "type": "inline_equation", + "content": "\\mathcal{W}+" + }, + { + "bbox": [ + 304, + 159, + 545, + 483 + ], + "type": "text", + "content": " space) and then achieve image editing by manipulating the latent vectors. However, StyleGAN generates images in a highly restricted domain, and the image quality is not always high. This makes it challenging to successfully invert many real images to a satisfactory latent code. Recently, diffusion models have enabled image synthesis at high quality. An increasing number of image editing tasks [7, 17, 18] have shifted towards being based on diffusion models. Previous image editing methods that rely on diffusion models, such as SDEdit [16] and blended diffusion [2], cannot fully leverage the capabilities of accurate diffusion inversion as they introduced random noise into the input image to create a noisy initial state. Prompt-to-Prompt [7] is the first to accomplish extensive text-guided image editing without the need for diffusion model refinement, including local editing even when a known mask is unavailable. Null-text inversion [18] successfully reconstructs real images by optimizing the null-text embedding at each prediction step, which decides the unconditional prediction. Most recently, pix2pix-zero [22] proposes to learn editing directions in the textual embedding space for specific image translation tasks. In contrast to these methods, our work enables users to conduct image control through point-guided editing." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 491, + 455, + 504 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 491, + 455, + 504 + ], + "spans": [ + { + "bbox": [ + 306, + 491, + 455, + 504 + ], + "type": "text", + "content": "2.2. Point-guided image editing" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "text", + "content": "In order to facilitate fine-grained editing, several approaches have been introduced for point-based editing, including those by [5, 21, 26, 30]. DragGAN [21] introduces an interactive method for seamless point-based image editing, incorporating two innovative components: the optimization of latent codes to progressively move multiple handle points to their desired positions and a precise point tracking mechanism to accurately follow the path of these handle points. However, the results of DragGAN are constrained by the generative capacity of StyleGAN. Recently, DragDiffusion [26] extends the editing framework of DragGAN to diffusion models with LoRA fine-tuning. Conversely, LoRA fine-tuning significantly extends the waiting time, which is unfriendly to the users. DragonDiffusion [19] utilizes an energy function to guide the editing and a memory bank for editing consistency. SDE-Drag [20] proposes an SDE-based approach to formulate the image" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8405" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 51, + 72, + 545, + 208 + ], + "blocks": [ + { + "bbox": [ + 51, + 72, + 545, + 208 + ], + "lines": [ + { + "bbox": [ + 51, + 72, + 545, + 208 + ], + "spans": [ + { + "bbox": [ + 51, + 72, + 545, + 208 + ], + "type": "image", + "image_path": "dfa241d1bd2daeebaeba2ded18d81943e2b99102dee3163da5c08de54aeed2cc.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 217, + 547, + 262 + ], + "lines": [ + { + "bbox": [ + 46, + 217, + 547, + 262 + ], + "spans": [ + { + "bbox": [ + 46, + 217, + 547, + 262 + ], + "type": "text", + "content": "Figure 2. The pipeline of our framework. Our method generates masks automatically for precise supervision and convenience for users. To improve motion supervision over existing methods like DragGAN, we propose a stable motion supervision method based on diffusion models. The reference guidance where guidance latents are inherited from DDIM inversion preserves identity with the input image commendably." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 282, + 287, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 282, + 287, + 319 + ], + "spans": [ + { + "bbox": [ + 46, + 282, + 287, + 319 + ], + "type": "text", + "content": "editing. In contrast to previous works, we propose a novel point-based image editing framework, which can generate high-quality results without LoRA fine-tuning and masks." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 328, + 103, + 340 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 328, + 103, + 340 + ], + "spans": [ + { + "bbox": [ + 47, + 328, + 103, + 340 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 347, + 133, + 358 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 347, + 133, + 358 + ], + "spans": [ + { + "bbox": [ + 47, + 347, + 133, + 358 + ], + "type": "text", + "content": "3.1. Preliminaries" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 365, + 287, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 365, + 287, + 426 + ], + "spans": [ + { + "bbox": [ + 46, + 365, + 287, + 426 + ], + "type": "text", + "content": "Denoising diffusion probabilistic models (DDPM) [9] aim to map a pure noise " + }, + { + "bbox": [ + 46, + 365, + 287, + 426 + ], + "type": "inline_equation", + "content": "z_{T}" + }, + { + "bbox": [ + 46, + 365, + 287, + 426 + ], + "type": "text", + "content": " to an output image " + }, + { + "bbox": [ + 46, + 365, + 287, + 426 + ], + "type": "inline_equation", + "content": "z_{0}" + }, + { + "bbox": [ + 46, + 365, + 287, + 426 + ], + "type": "text", + "content": ", which is guided by the given conditioning prompt. During the training process, the network " + }, + { + "bbox": [ + 46, + 365, + 287, + 426 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}" + }, + { + "bbox": [ + 46, + 365, + 287, + 426 + ], + "type": "text", + "content": " is updated by predicting the noise " + }, + { + "bbox": [ + 46, + 365, + 287, + 426 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 46, + 365, + 287, + 426 + ], + "type": "text", + "content": " from the latent variables " + }, + { + "bbox": [ + 46, + 365, + 287, + 426 + ], + "type": "inline_equation", + "content": "z_{t}" + }, + { + "bbox": [ + 46, + 365, + 287, + 426 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 71, + 444, + 287, + 459 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 71, + 444, + 287, + 459 + ], + "spans": [ + { + "bbox": [ + 71, + 444, + 287, + 459 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\theta} = \\mathbb {E} _ {z _ {0}, \\epsilon \\sim N (0, I), t \\sim U (1, T)} \\| \\epsilon - \\epsilon_ {\\theta} \\left(z _ {t}, t, \\mathcal {C}\\right) \\| _ {2} ^ {2}, \\tag {1}", + "image_path": "f81e953d16a5aff58763623cc295896ba953ce5cce8ed1467e82a1425d6e7062.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 464, + 287, + 512 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 464, + 287, + 512 + ], + "spans": [ + { + "bbox": [ + 46, + 464, + 287, + 512 + ], + "type": "text", + "content": "where the noised sample " + }, + { + "bbox": [ + 46, + 464, + 287, + 512 + ], + "type": "inline_equation", + "content": "z_{t}" + }, + { + "bbox": [ + 46, + 464, + 287, + 512 + ], + "type": "text", + "content": " is calculated by adding noise " + }, + { + "bbox": [ + 46, + 464, + 287, + 512 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 46, + 464, + 287, + 512 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 46, + 464, + 287, + 512 + ], + "type": "inline_equation", + "content": "z_{0}" + }, + { + "bbox": [ + 46, + 464, + 287, + 512 + ], + "type": "text", + "content": " according to diffusion step " + }, + { + "bbox": [ + 46, + 464, + 287, + 512 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 464, + 287, + 512 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 464, + 287, + 512 + ], + "type": "inline_equation", + "content": "\\mathcal{C}" + }, + { + "bbox": [ + 46, + 464, + 287, + 512 + ], + "type": "text", + "content": " is the conditioning input of " + }, + { + "bbox": [ + 46, + 464, + 287, + 512 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}" + }, + { + "bbox": [ + 46, + 464, + 287, + 512 + ], + "type": "text", + "content": ". During inference, we use DDIM [27] for sampling method:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 528, + 290, + 569 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 528, + 290, + 569 + ], + "spans": [ + { + "bbox": [ + 47, + 528, + 290, + 569 + ], + "type": "interline_equation", + "content": "z _ {t - 1} = \\sqrt {\\frac {\\alpha_ {t - 1}}{\\alpha_ {t}}} z _ {t} + \\sqrt {\\alpha_ {t - 1}} \\left(\\sqrt {\\frac {1}{\\alpha_ {t - 1}} - 1} - \\sqrt {\\frac {1}{\\alpha_ {t}} - 1}\\right) \\epsilon_ {\\theta} \\left(z _ {t}\\right), \\tag {2}", + "image_path": "a6d96cc21bf65b0eecb6e6be849a6a326c4fac2c9102a1dfd9dc9ab311e353a4.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 570, + 287, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 570, + 287, + 594 + ], + "spans": [ + { + "bbox": [ + 47, + 570, + 287, + 594 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 570, + 287, + 594 + ], + "type": "inline_equation", + "content": "\\alpha_{t}(t = 0,1,2,\\dots,T)" + }, + { + "bbox": [ + 47, + 570, + 287, + 594 + ], + "type": "text", + "content": " represent the noise scales at corresponding diffusion steps." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 607, + 287, + 656 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 607, + 287, + 656 + ], + "spans": [ + { + "bbox": [ + 46, + 607, + 287, + 656 + ], + "type": "text", + "content": "DDIM inversion. The forward process of DDIM can be expressed in terms of " + }, + { + "bbox": [ + 46, + 607, + 287, + 656 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}(z_t,t,\\mathcal{C})" + }, + { + "bbox": [ + 46, + 607, + 287, + 656 + ], + "type": "text", + "content": ", based on the assumption that the ODE process can be inverted in the limit of small steps:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 671, + 289, + 714 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 671, + 289, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 671, + 289, + 714 + ], + "type": "interline_equation", + "content": "z _ {t + 1} = \\sqrt {\\frac {\\alpha_ {t + 1}}{\\alpha_ {t}}} z _ {t} + \\sqrt {\\alpha_ {t + 1}} \\left(\\sqrt {\\frac {1}{\\alpha_ {t + 1}} - 1} - \\sqrt {\\frac {1}{\\alpha_ {t}} - 1}\\right) \\epsilon_ {\\theta} \\left(z _ {t}\\right). \\tag {3}", + "image_path": "e746a8c901dca9e8e65e0b49ab192db0f6247e428857261d31a0d69156fe75d2.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 282, + 545, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 282, + 545, + 354 + ], + "spans": [ + { + "bbox": [ + 305, + 282, + 545, + 354 + ], + "type": "text", + "content": "Stable Diffusion. Stable Diffusion (SD) [24] presents the input image into a lower-dimensional latent space with a Variational Auto-Encoder (VAE) [15], where the input latent " + }, + { + "bbox": [ + 305, + 282, + 545, + 354 + ], + "type": "inline_equation", + "content": "z_0" + }, + { + "bbox": [ + 305, + 282, + 545, + 354 + ], + "type": "text", + "content": " is defined as " + }, + { + "bbox": [ + 305, + 282, + 545, + 354 + ], + "type": "inline_equation", + "content": "E(x_0)" + }, + { + "bbox": [ + 305, + 282, + 545, + 354 + ], + "type": "text", + "content": " and the output image " + }, + { + "bbox": [ + 305, + 282, + 545, + 354 + ], + "type": "inline_equation", + "content": "x_0" + }, + { + "bbox": [ + 305, + 282, + 545, + 354 + ], + "type": "text", + "content": " is obtained by " + }, + { + "bbox": [ + 305, + 282, + 545, + 354 + ], + "type": "inline_equation", + "content": "D(z_0)" + }, + { + "bbox": [ + 305, + 282, + 545, + 354 + ], + "type": "text", + "content": ". We implement our work based on Stable Diffusion model." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 369, + 546, + 464 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 369, + 546, + 464 + ], + "spans": [ + { + "bbox": [ + 305, + 369, + 546, + 464 + ], + "type": "text", + "content": "Classifier-free guidance. The extent to which the prompt influences the reverse diffusion process directly affects the quality of image generation. For this purpose, Ho et al. [8] proposes the classifier-free guidance, which controls the degree of impact from text conditioning. Specifically, the null next-embedding " + }, + { + "bbox": [ + 305, + 369, + 546, + 464 + ], + "type": "inline_equation", + "content": "\\phi = \\psi(\"\")" + }, + { + "bbox": [ + 305, + 369, + 546, + 464 + ], + "type": "text", + "content": " is employed as a anchor point for unconditional prediction to enhance the text conditioning with guidance scale " + }, + { + "bbox": [ + 305, + 369, + 546, + 464 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 305, + 369, + 546, + 464 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 310, + 484, + 545, + 498 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 310, + 484, + 545, + 498 + ], + "spans": [ + { + "bbox": [ + 310, + 484, + 545, + 498 + ], + "type": "interline_equation", + "content": "\\tilde {\\epsilon} _ {\\theta} \\left(z _ {t}, t, \\mathcal {C}, \\phi\\right) = w * \\epsilon_ {\\theta} \\left(z _ {t}, t, \\mathcal {C}\\right) + (1 - w) * \\epsilon_ {\\theta} \\left(z _ {t}, t, \\phi\\right). \\tag {4}", + "image_path": "2706d0e6e33807104186f45b7f82200d7e68571afbea057dca99f815320babd9.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 503, + 373, + 515 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 503, + 373, + 515 + ], + "spans": [ + { + "bbox": [ + 306, + 503, + 373, + 515 + ], + "type": "text", + "content": "3.2. Overview" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 305, + 522, + 545, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 522, + 545, + 594 + ], + "spans": [ + { + "bbox": [ + 305, + 522, + 545, + 594 + ], + "type": "text", + "content": "Given an input image " + }, + { + "bbox": [ + 305, + 522, + 545, + 594 + ], + "type": "inline_equation", + "content": "\\mathcal{I}" + }, + { + "bbox": [ + 305, + 522, + 545, + 594 + ], + "type": "text", + "content": ", the user should input a number of handle points " + }, + { + "bbox": [ + 305, + 522, + 545, + 594 + ], + "type": "inline_equation", + "content": "\\{\\pmb{p}_i = (x_{p,i},y_{p,i})|i = 1,2,\\dots,N\\}" + }, + { + "bbox": [ + 305, + 522, + 545, + 594 + ], + "type": "text", + "content": " and their corresponding target points " + }, + { + "bbox": [ + 305, + 522, + 545, + 594 + ], + "type": "inline_equation", + "content": "\\{\\pmb{q}_i = (x_{q,i},y_{q,i})|i = 1,2,\\dots,N\\}" + }, + { + "bbox": [ + 305, + 522, + 545, + 594 + ], + "type": "text", + "content": ". The objective is to drag the content of the handle points to the target points based on image " + }, + { + "bbox": [ + 305, + 522, + 545, + 594 + ], + "type": "inline_equation", + "content": "\\mathcal{I}" + }, + { + "bbox": [ + 305, + 522, + 545, + 594 + ], + "type": "text", + "content": ". The user can also input a mask " + }, + { + "bbox": [ + 305, + 522, + 545, + 594 + ], + "type": "inline_equation", + "content": "\\mathbf{M}" + }, + { + "bbox": [ + 305, + 522, + 545, + 594 + ], + "type": "text", + "content": " to specify the modifiable area." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 594, + 546, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 546, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 546, + 714 + ], + "type": "text", + "content": "We operate DDIM inversion with " + }, + { + "bbox": [ + 304, + 594, + 546, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{C} = \\psi(\\mathcal{P})" + }, + { + "bbox": [ + 304, + 594, + 546, + 714 + ], + "type": "text", + "content": " on the input image, where " + }, + { + "bbox": [ + 304, + 594, + 546, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 304, + 594, + 546, + 714 + ], + "type": "text", + "content": " represents the prompt, and get the intermediate results " + }, + { + "bbox": [ + 304, + 594, + 546, + 714 + ], + "type": "inline_equation", + "content": "z_{T}, \\ldots, z_{t}, \\ldots, z_{0}" + }, + { + "bbox": [ + 304, + 594, + 546, + 714 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 594, + 546, + 714 + ], + "type": "inline_equation", + "content": "z_{t}" + }, + { + "bbox": [ + 304, + 594, + 546, + 714 + ], + "type": "text", + "content": " is a specific intermediate latent in DDIM inversion. Loop of optimizations is performed on " + }, + { + "bbox": [ + 304, + 594, + 546, + 714 + ], + "type": "inline_equation", + "content": "z_{t}" + }, + { + "bbox": [ + 304, + 594, + 546, + 714 + ], + "type": "text", + "content": " to make features of the target points similar enough to that of the handle points, with novel stable motion supervision and automatic mask generation introduced in this paper. Then, the result of optimization " + }, + { + "bbox": [ + 304, + 594, + 546, + 714 + ], + "type": "inline_equation", + "content": "z_{t}^{*}" + }, + { + "bbox": [ + 304, + 594, + 546, + 714 + ], + "type": "text", + "content": " is fed into the denoising process with reference guidance to achieve identity consistency with " + }, + { + "bbox": [ + 304, + 594, + 546, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{I}" + }, + { + "bbox": [ + 304, + 594, + 546, + 714 + ], + "type": "text", + "content": ". We present the pipeline" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8406" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 64, + 92, + 121, + 146 + ], + "blocks": [ + { + "bbox": [ + 64, + 92, + 121, + 146 + ], + "lines": [ + { + "bbox": [ + 64, + 92, + 121, + 146 + ], + "spans": [ + { + "bbox": [ + 64, + 92, + 121, + 146 + ], + "type": "image", + "image_path": "f3d668311aaa51f0aeb69f1ebc1cff5fbcff5f9f1b3d6e0e1078aa414667f69a.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 132, + 74, + 187, + 147 + ], + "blocks": [ + { + "bbox": [ + 53, + 76, + 128, + 85 + ], + "lines": [ + { + "bbox": [ + 53, + 76, + 128, + 85 + ], + "spans": [ + { + "bbox": [ + 53, + 76, + 128, + 85 + ], + "type": "text", + "content": "Stable motion supervision" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 132, + 74, + 187, + 147 + ], + "lines": [ + { + "bbox": [ + 132, + 74, + 187, + 147 + ], + "spans": [ + { + "bbox": [ + 132, + 74, + 187, + 147 + ], + "type": "image", + "image_path": "d54a159f4c6ab1630366a150fa66ea2cf5a8e05e347dc5319436e1dd432abfd9.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 200, + 74, + 256, + 146 + ], + "blocks": [ + { + "bbox": [ + 200, + 74, + 256, + 146 + ], + "lines": [ + { + "bbox": [ + 200, + 74, + 256, + 146 + ], + "spans": [ + { + "bbox": [ + 200, + 74, + 256, + 146 + ], + "type": "image", + "image_path": "639f2fb7e9195743f74fc8061af117ca9c0b4808c8d151b2610161b73c640d0c.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 331, + 287, + 355 + ], + "lines": [ + { + "bbox": [ + 46, + 331, + 287, + 355 + ], + "spans": [ + { + "bbox": [ + 46, + 331, + 287, + 355 + ], + "type": "text", + "content": "in Fig. 2, and elaborate three key components in the following subsections." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 263, + 95, + 279, + 117 + ], + "blocks": [ + { + "bbox": [ + 263, + 95, + 279, + 117 + ], + "lines": [ + { + "bbox": [ + 263, + 95, + 279, + 117 + ], + "spans": [ + { + "bbox": [ + 263, + 95, + 279, + 117 + ], + "type": "image", + "image_path": "ed9925cab50f09aa9096fa4664bfcd4564a964acad549fadab34158ec8c1fe16.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 59, + 158, + 119, + 211 + ], + "blocks": [ + { + "bbox": [ + 59, + 158, + 119, + 211 + ], + "lines": [ + { + "bbox": [ + 59, + 158, + 119, + 211 + ], + "spans": [ + { + "bbox": [ + 59, + 158, + 119, + 211 + ], + "type": "image", + "image_path": "6a69fc3d968cdbec1eadcbf9175becf4f5bd554278de73de8b095f8986cc8d68.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 58, + 217, + 139, + 225 + ], + "lines": [ + { + "bbox": [ + 58, + 217, + 139, + 225 + ], + "spans": [ + { + "bbox": [ + 58, + 217, + 139, + 225 + ], + "type": "text", + "content": "DragGAN motion supervision" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 133, + 158, + 187, + 212 + ], + "blocks": [ + { + "bbox": [ + 133, + 158, + 187, + 212 + ], + "lines": [ + { + "bbox": [ + 133, + 158, + 187, + 212 + ], + "spans": [ + { + "bbox": [ + 133, + 158, + 187, + 212 + ], + "type": "image", + "image_path": "e19e33f2a65b13d7178bec8d57fd56072e6d2beb6ae6c05d0f162d6608140ba8.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 241, + 287, + 307 + ], + "lines": [ + { + "bbox": [ + 46, + 241, + 287, + 307 + ], + "spans": [ + { + "bbox": [ + 46, + 241, + 287, + 307 + ], + "type": "text", + "content": "Figure 3. Stable motion supervision. When the next point " + }, + { + "bbox": [ + 46, + 241, + 287, + 307 + ], + "type": "inline_equation", + "content": "p^{**}" + }, + { + "bbox": [ + 46, + 241, + 287, + 307 + ], + "type": "text", + "content": " is more similar with the handle point " + }, + { + "bbox": [ + 46, + 241, + 287, + 307 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 46, + 241, + 287, + 307 + ], + "type": "text", + "content": " than " + }, + { + "bbox": [ + 46, + 241, + 287, + 307 + ], + "type": "inline_equation", + "content": "p^*" + }, + { + "bbox": [ + 46, + 241, + 287, + 307 + ], + "type": "text", + "content": ", the handle point has moved to the position of " + }, + { + "bbox": [ + 46, + 241, + 287, + 307 + ], + "type": "inline_equation", + "content": "p^{**}" + }, + { + "bbox": [ + 46, + 241, + 287, + 307 + ], + "type": "text", + "content": ". Our approach differs significantly from DragGAN in that our method always learns features from the original image, while DragGAN learns features that are continually changing." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 201, + 158, + 256, + 212 + ], + "blocks": [ + { + "bbox": [ + 201, + 158, + 256, + 212 + ], + "lines": [ + { + "bbox": [ + 201, + 158, + 256, + 212 + ], + "spans": [ + { + "bbox": [ + 201, + 158, + 256, + 212 + ], + "type": "image", + "image_path": "905e240854c90ac9f5e223705c101966b6b722bc645151f810788c05afaf3516.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 206, + 217, + 252, + 227 + ], + "blocks": [ + { + "bbox": [ + 206, + 217, + 252, + 227 + ], + "lines": [ + { + "bbox": [ + 206, + 217, + 252, + 227 + ], + "spans": [ + { + "bbox": [ + 206, + 217, + 252, + 227 + ], + "type": "image", + "image_path": "fc42870db319f82e2d0272956996e6e0b1b894a3cd2c89fea997bcaac6a83168.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 366, + 190, + 379 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 366, + 190, + 379 + ], + "spans": [ + { + "bbox": [ + 47, + 366, + 190, + 379 + ], + "type": "text", + "content": "3.3. Stable motion supervision" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 46, + 386, + 287, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 386, + 287, + 506 + ], + "spans": [ + { + "bbox": [ + 46, + 386, + 287, + 506 + ], + "type": "text", + "content": "There exists misleading in DragGAN due to the proposed motion supervision. In contrast to DragGAN, we introduce an efficient drag method that is more suitable for diffusion models. Inspired by DIFT [28], we utilize the feature maps of U-Net denoiser as the semantic information of the input image. Specially, we consider the second and third upper layers of U-Net with " + }, + { + "bbox": [ + 46, + 386, + 287, + 506 + ], + "type": "inline_equation", + "content": "(z_{t}, t, \\mathcal{C})" + }, + { + "bbox": [ + 46, + 386, + 287, + 506 + ], + "type": "text", + "content": " as inputs, and concatenate the two layers as our feature maps " + }, + { + "bbox": [ + 46, + 386, + 287, + 506 + ], + "type": "inline_equation", + "content": "\\mathbf{F}" + }, + { + "bbox": [ + 46, + 386, + 287, + 506 + ], + "type": "text", + "content": " used for stable motion supervision. " + }, + { + "bbox": [ + 46, + 386, + 287, + 506 + ], + "type": "inline_equation", + "content": "\\mathbf{F}" + }, + { + "bbox": [ + 46, + 386, + 287, + 506 + ], + "type": "text", + "content": " is resized to have the same resolution as the input image in order to get precise manipulation." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 46, + 507, + 287, + 626 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 507, + 287, + 626 + ], + "spans": [ + { + "bbox": [ + 46, + 507, + 287, + 626 + ], + "type": "text", + "content": "As shown in Fig. 3, " + }, + { + "bbox": [ + 46, + 507, + 287, + 626 + ], + "type": "inline_equation", + "content": "p^*" + }, + { + "bbox": [ + 46, + 507, + 287, + 626 + ], + "type": "text", + "content": " represents current position of the handle point, and " + }, + { + "bbox": [ + 46, + 507, + 287, + 626 + ], + "type": "inline_equation", + "content": "p^{**}" + }, + { + "bbox": [ + 46, + 507, + 287, + 626 + ], + "type": "text", + "content": " is the next position that we want to drag to. The method for calculating " + }, + { + "bbox": [ + 46, + 507, + 287, + 626 + ], + "type": "inline_equation", + "content": "p^{**}" + }, + { + "bbox": [ + 46, + 507, + 287, + 626 + ], + "type": "text", + "content": " is presented in Algorithm 1. We start from " + }, + { + "bbox": [ + 46, + 507, + 287, + 626 + ], + "type": "inline_equation", + "content": "p^*" + }, + { + "bbox": [ + 46, + 507, + 287, + 626 + ], + "type": "text", + "content": ", move along the direction of " + }, + { + "bbox": [ + 46, + 507, + 287, + 626 + ], + "type": "inline_equation", + "content": "q" + }, + { + "bbox": [ + 46, + 507, + 287, + 626 + ], + "type": "text", + "content": " by " + }, + { + "bbox": [ + 46, + 507, + 287, + 626 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 46, + 507, + 287, + 626 + ], + "type": "text", + "content": " units, and then find the nearest integer coordinates to determine " + }, + { + "bbox": [ + 46, + 507, + 287, + 626 + ], + "type": "inline_equation", + "content": "p^{**}" + }, + { + "bbox": [ + 46, + 507, + 287, + 626 + ], + "type": "text", + "content": ". In contrast to DragGAN, we employ " + }, + { + "bbox": [ + 46, + 507, + 287, + 626 + ], + "type": "inline_equation", + "content": "p^*" + }, + { + "bbox": [ + 46, + 507, + 287, + 626 + ], + "type": "text", + "content": " to learn features of the original image at the " + }, + { + "bbox": [ + 46, + 507, + 287, + 626 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 46, + 507, + 287, + 626 + ], + "type": "text", + "content": " position, ensuring that the learned features are consistently accurate. Inspired by DragonDiffusion [19], we define the similarity loss function of two feature maps as follows:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 94, + 640, + 287, + 669 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 94, + 640, + 287, + 669 + ], + "spans": [ + { + "bbox": [ + 94, + 640, + 287, + 669 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {c o s}} \\left(\\mathbf {F}, \\mathbf {F} _ {0}\\right) = \\frac {1}{1 + \\frac {1 + \\cos \\left(\\mathbf {F} , s g \\left(\\mathbf {F} _ {0}\\right)\\right)}{2}}, \\tag {5}", + "image_path": "4ca8140621b26d1130941534555c1a35d21560762f15b63f889ce04d8e063260.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 47, + 677, + 287, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 677, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 47, + 677, + 287, + 712 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 677, + 287, + 712 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_0" + }, + { + "bbox": [ + 47, + 677, + 287, + 712 + ], + "type": "text", + "content": " represents the feature maps corresponding to the input image and " + }, + { + "bbox": [ + 47, + 677, + 287, + 712 + ], + "type": "inline_equation", + "content": "sg(\\cdot)" + }, + { + "bbox": [ + 47, + 677, + 287, + 712 + ], + "type": "text", + "content": " is the stop gradient operator. Therefore, the stable motion supervision loss is defined as:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 312, + 76, + 414, + 87 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 76, + 414, + 87 + ], + "spans": [ + { + "bbox": [ + 312, + 76, + 414, + 87 + ], + "type": "text", + "content": "Algorithm 1: Next point" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 318, + 89, + 477, + 100 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 89, + 477, + 100 + ], + "spans": [ + { + "bbox": [ + 318, + 89, + 477, + 100 + ], + "type": "text", + "content": "Input: Current point " + }, + { + "bbox": [ + 318, + 89, + 477, + 100 + ], + "type": "inline_equation", + "content": "p^*" + }, + { + "bbox": [ + 318, + 89, + 477, + 100 + ], + "type": "text", + "content": ", Target point " + }, + { + "bbox": [ + 318, + 89, + 477, + 100 + ], + "type": "inline_equation", + "content": "q" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 319, + 101, + 416, + 113 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 101, + 416, + 113 + ], + "spans": [ + { + "bbox": [ + 319, + 101, + 416, + 113 + ], + "type": "text", + "content": "Output: Next point " + }, + { + "bbox": [ + 319, + 101, + 416, + 113 + ], + "type": "inline_equation", + "content": "p^{**}" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 311, + 117, + 447, + 206 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 311, + 117, + 414, + 129 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 117, + 414, + 129 + ], + "spans": [ + { + "bbox": [ + 311, + 117, + 414, + 129 + ], + "type": "text", + "content": "1 if " + }, + { + "bbox": [ + 311, + 117, + 414, + 129 + ], + "type": "inline_equation", + "content": "\\| \\pmb {q} - \\pmb{p}^{*}\\|_{2} < K" + }, + { + "bbox": [ + 311, + 117, + 414, + 129 + ], + "type": "text", + "content": " then" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 311, + 129, + 373, + 141 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 129, + 373, + 141 + ], + "spans": [ + { + "bbox": [ + 311, + 129, + 373, + 141 + ], + "type": "text", + "content": "2 " + }, + { + "bbox": [ + 311, + 129, + 373, + 141 + ], + "type": "inline_equation", + "content": "\\pmb{p}^{**}\\gets \\pmb {q};" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 311, + 142, + 337, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 142, + 337, + 152 + ], + "spans": [ + { + "bbox": [ + 311, + 142, + 337, + 152 + ], + "type": "text", + "content": "3 else" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 311, + 152, + 391, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 152, + 391, + 171 + ], + "spans": [ + { + "bbox": [ + 311, + 152, + 391, + 171 + ], + "type": "text", + "content": "4 " + }, + { + "bbox": [ + 311, + 152, + 391, + 171 + ], + "type": "inline_equation", + "content": "\\pmb{d} \\gets \\frac{\\pmb{q} - \\pmb{p}^{*}}{\\|\\pmb{q} - \\pmb{p}^{*}\\|_{2}};" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 311, + 170, + 447, + 182 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 170, + 447, + 182 + ], + "spans": [ + { + "bbox": [ + 311, + 170, + 447, + 182 + ], + "type": "text", + "content": "5 " + }, + { + "bbox": [ + 311, + 170, + 447, + 182 + ], + "type": "inline_equation", + "content": "\\pmb{p}^{**}\\gets \\pmb{p}^* +round(\\pmb {d}*\\pmb {K})" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 311, + 182, + 336, + 192 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 182, + 336, + 192 + ], + "spans": [ + { + "bbox": [ + 311, + 182, + 336, + 192 + ], + "type": "text", + "content": "6 end" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 311, + 194, + 363, + 206 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 194, + 363, + 206 + ], + "spans": [ + { + "bbox": [ + 311, + 194, + 363, + 206 + ], + "type": "text", + "content": "7 return " + }, + { + "bbox": [ + 311, + 194, + 363, + 206 + ], + "type": "inline_equation", + "content": "p^{**}" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 340, + 241, + 545, + 289 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 340, + 241, + 545, + 289 + ], + "spans": [ + { + "bbox": [ + 340, + 241, + 545, + 289 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} = \\sum_ {i = 1} ^ {N} \\mathcal {L} _ {\\text {c o s}} \\left(\\mathbf {F} \\left(\\boldsymbol {p} _ {i} ^ {* *}\\right), \\mathbf {F} _ {0} \\left(\\boldsymbol {p} _ {i}\\right)\\right) \\tag {6} \\\\ + \\lambda \\mathcal {L} _ {\\cos} (\\mathbf {F} \\odot (1 - \\mathbf {M}), \\mathbf {F} _ {0} \\odot (1 - \\mathbf {M})), \\\\ \\end{array}", + "image_path": "c138d3a5295973c0711308d699a75b949d9486a65c70f7403d70e1565de6f2de.jpg" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 305, + 293, + 537, + 306 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 293, + 537, + 306 + ], + "spans": [ + { + "bbox": [ + 305, + 293, + 537, + 306 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 293, + 537, + 306 + ], + "type": "inline_equation", + "content": "\\mathbf{F}(\\pmb{p}_i^{**})" + }, + { + "bbox": [ + 305, + 293, + 537, + 306 + ], + "type": "text", + "content": " denotes the feature values of " + }, + { + "bbox": [ + 305, + 293, + 537, + 306 + ], + "type": "inline_equation", + "content": "\\mathbf{F}" + }, + { + "bbox": [ + 305, + 293, + 537, + 306 + ], + "type": "text", + "content": " at pixel " + }, + { + "bbox": [ + 305, + 293, + 537, + 306 + ], + "type": "inline_equation", + "content": "\\pmb{p}_i^{**}" + }, + { + "bbox": [ + 305, + 293, + 537, + 306 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 304, + 306, + 545, + 390 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 306, + 545, + 390 + ], + "spans": [ + { + "bbox": [ + 304, + 306, + 545, + 390 + ], + "type": "text", + "content": "After motion supervision, the updated latent code " + }, + { + "bbox": [ + 304, + 306, + 545, + 390 + ], + "type": "inline_equation", + "content": "z_{t}^{*}" + }, + { + "bbox": [ + 304, + 306, + 545, + 390 + ], + "type": "text", + "content": " may change the position of handle points. Consequently, we need to determine whether the handle point has reached the next target point. To do this, we compare which one, " + }, + { + "bbox": [ + 304, + 306, + 545, + 390 + ], + "type": "inline_equation", + "content": "\\mathbf{F}(\\pmb{p}_i^{**})" + }, + { + "bbox": [ + 304, + 306, + 545, + 390 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 304, + 306, + 545, + 390 + ], + "type": "inline_equation", + "content": "\\mathbf{F}(\\pmb{p}_i^*)" + }, + { + "bbox": [ + 304, + 306, + 545, + 390 + ], + "type": "text", + "content": ", resembles " + }, + { + "bbox": [ + 304, + 306, + 545, + 390 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_0(\\pmb{p}_i)" + }, + { + "bbox": [ + 304, + 306, + 545, + 390 + ], + "type": "text", + "content": " more. If " + }, + { + "bbox": [ + 304, + 306, + 545, + 390 + ], + "type": "inline_equation", + "content": "\\cos(\\mathbf{F}(\\pmb{p}_i^{**}), \\mathbf{F}_0(\\pmb{p}_i))" + }, + { + "bbox": [ + 304, + 306, + 545, + 390 + ], + "type": "text", + "content": " is higher, the handle point " + }, + { + "bbox": [ + 304, + 306, + 545, + 390 + ], + "type": "inline_equation", + "content": "\\pmb{p}_i^*" + }, + { + "bbox": [ + 304, + 306, + 545, + 390 + ], + "type": "text", + "content": " is updated to " + }, + { + "bbox": [ + 304, + 306, + 545, + 390 + ], + "type": "inline_equation", + "content": "\\pmb{p}_i^{**}" + }, + { + "bbox": [ + 304, + 306, + 545, + 390 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 306, + 545, + 390 + ], + "type": "inline_equation", + "content": "\\pmb{p}_i^{**}" + }, + { + "bbox": [ + 304, + 306, + 545, + 390 + ], + "type": "text", + "content": " is further updated to " + }, + { + "bbox": [ + 304, + 306, + 545, + 390 + ], + "type": "inline_equation", + "content": "\\text{next}(\\pmb{p}_i^*, \\pmb{q}_i)" + }, + { + "bbox": [ + 304, + 306, + 545, + 390 + ], + "type": "text", + "content": " according to Algorithm 1." + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 306, + 396, + 431, + 408 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 396, + 431, + 408 + ], + "spans": [ + { + "bbox": [ + 306, + 396, + 431, + 408 + ], + "type": "text", + "content": "3.4. Auto mask generation" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 304, + 415, + 545, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 415, + 545, + 521 + ], + "spans": [ + { + "bbox": [ + 304, + 415, + 545, + 521 + ], + "type": "text", + "content": "For user-friendly interaction, image editing should avoid burdening users with the task of drawing masks. Existing drag methods based on diffusion models [19, 26] always rely on mask inputs, which are provided by users. This requires users to have a deep understanding of the drag method because an incorrectly drawn mask can lead to unpredictable outcomes. Consequently, we propose a method that can automatically generate masks, which are able to change adaptively during the training process." + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 304, + 522, + 545, + 654 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 522, + 545, + 654 + ], + "spans": [ + { + "bbox": [ + 304, + 522, + 545, + 654 + ], + "type": "text", + "content": "In the process of motion supervision, the parts with larger gradients of " + }, + { + "bbox": [ + 304, + 522, + 545, + 654 + ], + "type": "inline_equation", + "content": "z_{t}^{*}" + }, + { + "bbox": [ + 304, + 522, + 545, + 654 + ], + "type": "text", + "content": " have a greater impact on the results, while those with smaller gradients are generally the parts that are expected to remain stationary. However, if the regions with relatively small gradients keep changing, the final result becomes unpredictable. Due to the self-attention layer in U-Net, minor changes made at step " + }, + { + "bbox": [ + 304, + 522, + 545, + 654 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 522, + 545, + 654 + ], + "type": "text", + "content": " are very likely to have an unacceptable impact on the final image. Therefore, we still need control over these low-impact areas. We achieve this by using regions with normalized gradients greater than a threshold " + }, + { + "bbox": [ + 304, + 522, + 545, + 654 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 304, + 522, + 545, + 654 + ], + "type": "text", + "content": " as the drag mask " + }, + { + "bbox": [ + 304, + 522, + 545, + 654 + ], + "type": "inline_equation", + "content": "\\mathbf{M}" + }, + { + "bbox": [ + 304, + 522, + 545, + 654 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "content": "Every time " + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\pmb{p}_i^*" + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "content": " is updated, the hot-zone of " + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "inline_equation", + "content": "z_t" + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "content": " also changes due to the position shifting of the drag point. We update our dynamic mask " + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{M}" + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "content": " each time " + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\pmb{p}_i^*" + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "content": " changes, but there is no need to change the mask when " + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\pmb{p}_i^*" + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "content": " remains the same. This is because for the same " + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\pmb{p}_i^*" + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\pmb{p}_i^{**}" + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "content": ", the hot-zone is consistent." + } + ] + } + ], + "index": 34 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "8407" + } + ] + } + ], + "index": 35 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 78, + 87, + 276, + 306 + ], + "blocks": [ + { + "bbox": [ + 78, + 76, + 92, + 86 + ], + "lines": [ + { + "bbox": [ + 78, + 76, + 92, + 86 + ], + "spans": [ + { + "bbox": [ + 78, + 76, + 92, + 86 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 78, + 87, + 276, + 306 + ], + "lines": [ + { + "bbox": [ + 78, + 87, + 276, + 306 + ], + "spans": [ + { + "bbox": [ + 78, + 87, + 276, + 306 + ], + "type": "image", + "image_path": "1076bdb83fba895d1d28cc783296bd895605188cbdede29e7e6e1f0d5649c4e0.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 319, + 547, + 354 + ], + "lines": [ + { + "bbox": [ + 46, + 319, + 547, + 354 + ], + "spans": [ + { + "bbox": [ + 46, + 319, + 547, + 354 + ], + "type": "text", + "content": "Figure 4. Qualitative comparison of our approach to DragDiffusion. Due to the misleading of motion supervision, DragDiffusion always cannot move the handle points to the target. Furthermore, DragDiffusion often exhibits significant background changes in the absence of the mask. In our method, the handle points reach the target positions precisely even without the mask." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 282, + 87, + 520, + 305 + ], + "blocks": [ + { + "bbox": [ + 285, + 76, + 298, + 86 + ], + "lines": [ + { + "bbox": [ + 285, + 76, + 298, + 86 + ], + "spans": [ + { + "bbox": [ + 285, + 76, + 298, + 86 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 282, + 87, + 520, + 305 + ], + "lines": [ + { + "bbox": [ + 282, + 87, + 520, + 305 + ], + "spans": [ + { + "bbox": [ + 282, + 87, + 520, + 305 + ], + "type": "image", + "image_path": "fecf87bf8c74650247a2579d8e28cd4520b7544482a98fd6e4ace7b433ab62a4.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 374, + 287, + 398 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 374, + 287, + 398 + ], + "spans": [ + { + "bbox": [ + 46, + 374, + 287, + 398 + ], + "type": "text", + "content": "This way, " + }, + { + "bbox": [ + 46, + 374, + 287, + 398 + ], + "type": "inline_equation", + "content": "\\mathbf{M}" + }, + { + "bbox": [ + 46, + 374, + 287, + 398 + ], + "type": "text", + "content": " adapts to changes of " + }, + { + "bbox": [ + 46, + 374, + 287, + 398 + ], + "type": "inline_equation", + "content": "z_{t}" + }, + { + "bbox": [ + 46, + 374, + 287, + 398 + ], + "type": "text", + "content": ", aligning better with the dragging requirements." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 404, + 162, + 417 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 404, + 162, + 417 + ], + "spans": [ + { + "bbox": [ + 47, + 404, + 162, + 417 + ], + "type": "text", + "content": "3.5. Reference guidance" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 422, + 287, + 530 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 422, + 287, + 530 + ], + "spans": [ + { + "bbox": [ + 46, + 422, + 287, + 530 + ], + "type": "text", + "content": "With optimized " + }, + { + "bbox": [ + 46, + 422, + 287, + 530 + ], + "type": "inline_equation", + "content": "z_{i}^{*}" + }, + { + "bbox": [ + 46, + 422, + 287, + 530 + ], + "type": "text", + "content": ", the generated images with direct DDIM denoising often exhibit identity changes and a decrease in image quality. Especially in areas where the pretrained models have limited generation capabilities, naively generating with " + }, + { + "bbox": [ + 46, + 422, + 287, + 530 + ], + "type": "inline_equation", + "content": "z_{i}" + }, + { + "bbox": [ + 46, + 422, + 287, + 530 + ], + "type": "text", + "content": " after DDIM Inversion may not even reproduce the original image. Drawing inspiration from [4, 19, 26], we employ mutual self-attention to ensure that images generated from " + }, + { + "bbox": [ + 46, + 422, + 287, + 530 + ], + "type": "inline_equation", + "content": "z_{i}^{*}" + }, + { + "bbox": [ + 46, + 422, + 287, + 530 + ], + "type": "text", + "content": " maintain content consistency with the original image." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 531, + 288, + 627 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 531, + 288, + 627 + ], + "spans": [ + { + "bbox": [ + 46, + 531, + 288, + 627 + ], + "type": "text", + "content": "As illustrated in Fig. 2, we replace the key and value of self-attention layers in backward process of " + }, + { + "bbox": [ + 46, + 531, + 288, + 627 + ], + "type": "inline_equation", + "content": "z_{i}^{*}" + }, + { + "bbox": [ + 46, + 531, + 288, + 627 + ], + "type": "text", + "content": " with the corresponding key and value in the process of reconstructing the source image " + }, + { + "bbox": [ + 46, + 531, + 288, + 627 + ], + "type": "inline_equation", + "content": "\\mathcal{I}" + }, + { + "bbox": [ + 46, + 531, + 288, + 627 + ], + "type": "text", + "content": ". Due to the feature correlations in lower layers being relatively weak, the mutual self-attention is only applied to the upper layers of U-Net " + }, + { + "bbox": [ + 46, + 531, + 288, + 627 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}" + }, + { + "bbox": [ + 46, + 531, + 288, + 627 + ], + "type": "text", + "content": ". We then perform sampling using the linear combination of the reference guidance and normal prediction:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 75, + 646, + 287, + 659 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 646, + 287, + 659 + ], + "spans": [ + { + "bbox": [ + 75, + 646, + 287, + 659 + ], + "type": "interline_equation", + "content": "\\epsilon_ {i} ^ {*} \\leftarrow w \\epsilon_ {\\theta} \\left(z _ {i} ^ {*}, z _ {i}, i, \\mathcal {C}\\right) + (1 - w) \\epsilon_ {\\theta} \\left(z _ {i} ^ {*}, i, \\mathcal {C}\\right), \\tag {7}", + "image_path": "5ebaf1031bac8969dc7b25f43f3d5100d47491109f0320ae62c86994b717c3ad.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 666, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 666, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 666, + 287, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 666, + 287, + 713 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 46, + 666, + 287, + 713 + ], + "type": "text", + "content": " controls the strength of the reference guidance. Specifically, " + }, + { + "bbox": [ + 46, + 666, + 287, + 713 + ], + "type": "inline_equation", + "content": "z_{i}" + }, + { + "bbox": [ + 46, + 666, + 287, + 713 + ], + "type": "text", + "content": " is the corresponding latent during DDIM inversion, rather than that generated from " + }, + { + "bbox": [ + 46, + 666, + 287, + 713 + ], + "type": "inline_equation", + "content": "z_{t}" + }, + { + "bbox": [ + 46, + 666, + 287, + 713 + ], + "type": "text", + "content": ". When " + }, + { + "bbox": [ + 46, + 666, + 287, + 713 + ], + "type": "inline_equation", + "content": "w = 1" + }, + { + "bbox": [ + 46, + 666, + 287, + 713 + ], + "type": "text", + "content": ", our reference guidance is the same as mutual self-attention." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 374, + 546, + 447 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 374, + 546, + 447 + ], + "spans": [ + { + "bbox": [ + 304, + 374, + 546, + 447 + ], + "type": "text", + "content": "However, in the cases where the input image is infrequent for the model and " + }, + { + "bbox": [ + 304, + 374, + 546, + 447 + ], + "type": "inline_equation", + "content": "z_{t}" + }, + { + "bbox": [ + 304, + 374, + 546, + 447 + ], + "type": "text", + "content": " undergoes significant changes, using mutual self-attention alone cannot guarantee identity consistency. When we increase the strength of reference guidance, the generated results can better ensure content and texture consistency with the input image." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 457, + 388, + 472 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 457, + 388, + 472 + ], + "spans": [ + { + "bbox": [ + 306, + 457, + 388, + 472 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 478, + 545, + 503 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 478, + 545, + 503 + ], + "spans": [ + { + "bbox": [ + 305, + 478, + 545, + 503 + ], + "type": "text", + "content": "In this section, we evaluate the proposed method qualitatively and quantitatively, on both synthetic and real images." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 517, + 546, + 661 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 517, + 546, + 661 + ], + "spans": [ + { + "bbox": [ + 304, + 517, + 546, + 661 + ], + "type": "text", + "content": "Implementation details. In our experiments, except for generated image manipulations, all tasks are based on Stable Diffusion 1.5 (SD-1.5). Unless otherwise specified, in both DDIM inversion and sampling we set the number of the sampling steps to be 50 via using the stride of 20 over 1000 diffusion steps. We set the latent from the 35-th DDIM step, which contains rich semantic information corresponding to the input image, as " + }, + { + "bbox": [ + 304, + 517, + 546, + 661 + ], + "type": "inline_equation", + "content": "z_{t}" + }, + { + "bbox": [ + 304, + 517, + 546, + 661 + ], + "type": "text", + "content": ". The gradient threshold " + }, + { + "bbox": [ + 304, + 517, + 546, + 661 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 304, + 517, + 546, + 661 + ], + "type": "text", + "content": " is 0.4, the motion step " + }, + { + "bbox": [ + 304, + 517, + 546, + 661 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 304, + 517, + 546, + 661 + ], + "type": "text", + "content": " is 12 and the reference guidance scale " + }, + { + "bbox": [ + 304, + 517, + 546, + 661 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 304, + 517, + 546, + 661 + ], + "type": "text", + "content": " is 4 by default. We use a learning rate of 0.01 for " + }, + { + "bbox": [ + 304, + 517, + 546, + 661 + ], + "type": "inline_equation", + "content": "z_{t}" + }, + { + "bbox": [ + 304, + 517, + 546, + 661 + ], + "type": "text", + "content": " without any decay schedule and train the latent with Adam [14] solver." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 677, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 677, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 305, + 677, + 547, + 715 + ], + "type": "text", + "content": "Dataset. We employ DragBench introduced in DragDiffusion as our evaluation dataset as it encompasses various types of images, and we also follow its dragging instruc" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8408" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 96, + 90, + 165, + 159 + ], + "blocks": [ + { + "bbox": [ + 78, + 119, + 93, + 131 + ], + "lines": [ + { + "bbox": [ + 78, + 119, + 93, + 131 + ], + "spans": [ + { + "bbox": [ + 78, + 119, + 93, + 131 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 115, + 76, + 144, + 88 + ], + "lines": [ + { + "bbox": [ + 115, + 76, + 144, + 88 + ], + "spans": [ + { + "bbox": [ + 115, + 76, + 144, + 88 + ], + "type": "text", + "content": "Inputs" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 96, + 90, + 165, + 159 + ], + "lines": [ + { + "bbox": [ + 96, + 90, + 165, + 159 + ], + "spans": [ + { + "bbox": [ + 96, + 90, + 165, + 159 + ], + "type": "image", + "image_path": "5600e1d001a83d73f3ccbbaf5344749d3b16e5a066b08eeab0be254b326ecff7.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 167, + 89, + 236, + 159 + ], + "blocks": [ + { + "bbox": [ + 184, + 76, + 216, + 87 + ], + "lines": [ + { + "bbox": [ + 184, + 76, + 216, + 87 + ], + "spans": [ + { + "bbox": [ + 184, + 76, + 216, + 87 + ], + "type": "text", + "content": "Results" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 167, + 89, + 236, + 159 + ], + "lines": [ + { + "bbox": [ + 167, + 89, + 236, + 159 + ], + "spans": [ + { + "bbox": [ + 167, + 89, + 236, + 159 + ], + "type": "image", + "image_path": "94692cfe6867f37db3afe49edefa00e535c70d570a67d5591df9db012e5a34cc.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 237, + 89, + 307, + 159 + ], + "blocks": [ + { + "bbox": [ + 258, + 76, + 287, + 87 + ], + "lines": [ + { + "bbox": [ + 258, + 76, + 287, + 87 + ], + "spans": [ + { + "bbox": [ + 258, + 76, + 287, + 87 + ], + "type": "text", + "content": "Inputs" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 237, + 89, + 307, + 159 + ], + "lines": [ + { + "bbox": [ + 237, + 89, + 307, + 159 + ], + "spans": [ + { + "bbox": [ + 237, + 89, + 307, + 159 + ], + "type": "image", + "image_path": "c0cc08965b72012715ec0e2c0b139069777c7d8089f74898e00ab1ecd10202c7.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 308, + 89, + 377, + 159 + ], + "blocks": [ + { + "bbox": [ + 325, + 76, + 358, + 87 + ], + "lines": [ + { + "bbox": [ + 325, + 76, + 358, + 87 + ], + "spans": [ + { + "bbox": [ + 325, + 76, + 358, + 87 + ], + "type": "text", + "content": "Results" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 308, + 89, + 377, + 159 + ], + "lines": [ + { + "bbox": [ + 308, + 89, + 377, + 159 + ], + "spans": [ + { + "bbox": [ + 308, + 89, + 377, + 159 + ], + "type": "image", + "image_path": "f09f5170ed21fe8ef65d342bdc256034f5e9d4a4a100a8680e79d03b2d9d0616.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 380, + 89, + 449, + 159 + ], + "blocks": [ + { + "bbox": [ + 401, + 76, + 429, + 87 + ], + "lines": [ + { + "bbox": [ + 401, + 76, + 429, + 87 + ], + "spans": [ + { + "bbox": [ + 401, + 76, + 429, + 87 + ], + "type": "text", + "content": "Inputs" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 380, + 89, + 449, + 159 + ], + "lines": [ + { + "bbox": [ + 380, + 89, + 449, + 159 + ], + "spans": [ + { + "bbox": [ + 380, + 89, + 449, + 159 + ], + "type": "image", + "image_path": "5074d6b041ecf4089806b68f0aefaa98eb48adf0105b68b3fad6f1f945705eff.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 451, + 89, + 520, + 159 + ], + "blocks": [ + { + "bbox": [ + 468, + 76, + 500, + 87 + ], + "lines": [ + { + "bbox": [ + 468, + 76, + 500, + 87 + ], + "spans": [ + { + "bbox": [ + 468, + 76, + 500, + 87 + ], + "type": "text", + "content": "Results" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 451, + 89, + 520, + 159 + ], + "lines": [ + { + "bbox": [ + 451, + 89, + 520, + 159 + ], + "spans": [ + { + "bbox": [ + 451, + 89, + 520, + 159 + ], + "type": "image", + "image_path": "3446da27be0e0bbce2cb970163d8dc19becfa3be5f7d4689d829fde8f713939d.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 96, + 163, + 165, + 232 + ], + "blocks": [ + { + "bbox": [ + 78, + 190, + 93, + 203 + ], + "lines": [ + { + "bbox": [ + 78, + 190, + 93, + 203 + ], + "spans": [ + { + "bbox": [ + 78, + 190, + 93, + 203 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 96, + 163, + 165, + 232 + ], + "lines": [ + { + "bbox": [ + 96, + 163, + 165, + 232 + ], + "spans": [ + { + "bbox": [ + 96, + 163, + 165, + 232 + ], + "type": "image", + "image_path": "e82cac6e9863c552c63f110feeb9f59fc34480a9f388efde9ff8fea656c7d9ab.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 167, + 163, + 235, + 232 + ], + "blocks": [ + { + "bbox": [ + 167, + 163, + 235, + 232 + ], + "lines": [ + { + "bbox": [ + 167, + 163, + 235, + 232 + ], + "spans": [ + { + "bbox": [ + 167, + 163, + 235, + 232 + ], + "type": "image", + "image_path": "afc20d99245a8cd0a328a8d41f8e7b520a61e499875512cf8fcbbf8905793434.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 238, + 163, + 307, + 232 + ], + "blocks": [ + { + "bbox": [ + 238, + 163, + 307, + 232 + ], + "lines": [ + { + "bbox": [ + 238, + 163, + 307, + 232 + ], + "spans": [ + { + "bbox": [ + 238, + 163, + 307, + 232 + ], + "type": "image", + "image_path": "5988b24870a3319ccb873c0a00c4c3542e0fb0be088ad618c0b81cdb429d0788.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 309, + 163, + 377, + 232 + ], + "blocks": [ + { + "bbox": [ + 309, + 163, + 377, + 232 + ], + "lines": [ + { + "bbox": [ + 309, + 163, + 377, + 232 + ], + "spans": [ + { + "bbox": [ + 309, + 163, + 377, + 232 + ], + "type": "image", + "image_path": "eabd01accd354e303531c596f04c79c0e24c936c36eee532807369438a159577.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 380, + 162, + 449, + 232 + ], + "blocks": [ + { + "bbox": [ + 380, + 162, + 449, + 232 + ], + "lines": [ + { + "bbox": [ + 380, + 162, + 449, + 232 + ], + "spans": [ + { + "bbox": [ + 380, + 162, + 449, + 232 + ], + "type": "image", + "image_path": "bc2a6496f40586416aa881955896a511e1c83c7d1fea9f1d1f8ac2d3865acee9.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 451, + 163, + 520, + 232 + ], + "blocks": [ + { + "bbox": [ + 451, + 163, + 520, + 232 + ], + "lines": [ + { + "bbox": [ + 451, + 163, + 520, + 232 + ], + "spans": [ + { + "bbox": [ + 451, + 163, + 520, + 232 + ], + "type": "image", + "image_path": "208994933d984a2212dce4e266454ba0d09aa49c087e60b0652ecba0a90a1e77.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 96, + 234, + 165, + 304 + ], + "blocks": [ + { + "bbox": [ + 78, + 264, + 93, + 276 + ], + "lines": [ + { + "bbox": [ + 78, + 264, + 93, + 276 + ], + "spans": [ + { + "bbox": [ + 78, + 264, + 93, + 276 + ], + "type": "text", + "content": "(c)" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 96, + 234, + 165, + 304 + ], + "lines": [ + { + "bbox": [ + 96, + 234, + 165, + 304 + ], + "spans": [ + { + "bbox": [ + 96, + 234, + 165, + 304 + ], + "type": "image", + "image_path": "377c7ab307cea6c20f1bf9095a19fa823139f0e0b6bd924797ee53eae72eaba1.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 167, + 234, + 235, + 304 + ], + "blocks": [ + { + "bbox": [ + 167, + 234, + 235, + 304 + ], + "lines": [ + { + "bbox": [ + 167, + 234, + 235, + 304 + ], + "spans": [ + { + "bbox": [ + 167, + 234, + 235, + 304 + ], + "type": "image", + "image_path": "127c7a07f9e408808d573cf6f08f03ce3ed21872e6aef5b9c6add66f02f3a71f.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 238, + 234, + 307, + 304 + ], + "blocks": [ + { + "bbox": [ + 238, + 234, + 307, + 304 + ], + "lines": [ + { + "bbox": [ + 238, + 234, + 307, + 304 + ], + "spans": [ + { + "bbox": [ + 238, + 234, + 307, + 304 + ], + "type": "image", + "image_path": "759abe29efba09868916adc3c7f935ae4e339085e06d4d5cdfbfd1785723fbfc.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 309, + 234, + 377, + 304 + ], + "blocks": [ + { + "bbox": [ + 309, + 234, + 377, + 304 + ], + "lines": [ + { + "bbox": [ + 309, + 234, + 377, + 304 + ], + "spans": [ + { + "bbox": [ + 309, + 234, + 377, + 304 + ], + "type": "image", + "image_path": "0434fa00e25c242a2976c474351d1f06a5d5fcfd5a701762f7b754a460159ebf.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 380, + 234, + 449, + 304 + ], + "blocks": [ + { + "bbox": [ + 380, + 234, + 449, + 304 + ], + "lines": [ + { + "bbox": [ + 380, + 234, + 449, + 304 + ], + "spans": [ + { + "bbox": [ + 380, + 234, + 449, + 304 + ], + "type": "image", + "image_path": "62da21abc77b0febb11056941ab8203c407a84fa658876a306fc125c7cceb8d4.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 451, + 234, + 520, + 304 + ], + "blocks": [ + { + "bbox": [ + 451, + 234, + 520, + 304 + ], + "lines": [ + { + "bbox": [ + 451, + 234, + 520, + 304 + ], + "spans": [ + { + "bbox": [ + 451, + 234, + 520, + 304 + ], + "type": "image", + "image_path": "757ef7d9241db58cc1cad54919191c5b90f577826ca6ca52bf55d6b867b42a93.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 96, + 307, + 165, + 376 + ], + "blocks": [ + { + "bbox": [ + 78, + 336, + 93, + 348 + ], + "lines": [ + { + "bbox": [ + 78, + 336, + 93, + 348 + ], + "spans": [ + { + "bbox": [ + 78, + 336, + 93, + 348 + ], + "type": "text", + "content": "(d)" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 96, + 307, + 165, + 376 + ], + "lines": [ + { + "bbox": [ + 96, + 307, + 165, + 376 + ], + "spans": [ + { + "bbox": [ + 96, + 307, + 165, + 376 + ], + "type": "image", + "image_path": "f6be91b9b38b006e8d8e42bc5fc7aeca8739f1e8240f04fc0ca1ef35f8da9fb3.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 385, + 545, + 407 + ], + "lines": [ + { + "bbox": [ + 46, + 385, + 545, + 407 + ], + "spans": [ + { + "bbox": [ + 46, + 385, + 545, + 407 + ], + "type": "text", + "content": "Figure 5. Qualitative results of our approach on generated images with SD-1.5 and its fine-tuned variants. (a) SD-1.5. (b) Majicmix Realistic V7. (c) Interior Design Supermix. (d) Dvarch." + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_caption" + } + ], + "index": 28 + }, + { + "type": "image", + "bbox": [ + 167, + 307, + 235, + 376 + ], + "blocks": [ + { + "bbox": [ + 167, + 307, + 235, + 376 + ], + "lines": [ + { + "bbox": [ + 167, + 307, + 235, + 376 + ], + "spans": [ + { + "bbox": [ + 167, + 307, + 235, + 376 + ], + "type": "image", + "image_path": "8646d15d7956bafba38430089a1f6ce4717d9221687f8fb2298de12274fe2f3d.jpg" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_body" + } + ], + "index": 29 + }, + { + "type": "image", + "bbox": [ + 238, + 307, + 307, + 376 + ], + "blocks": [ + { + "bbox": [ + 238, + 307, + 307, + 376 + ], + "lines": [ + { + "bbox": [ + 238, + 307, + 307, + 376 + ], + "spans": [ + { + "bbox": [ + 238, + 307, + 307, + 376 + ], + "type": "image", + "image_path": "91b40614ca1168e551f12fdea0dd87e2c6a3f1ce524c2701bff173db947695c3.jpg" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_body" + } + ], + "index": 30 + }, + { + "type": "image", + "bbox": [ + 309, + 307, + 377, + 376 + ], + "blocks": [ + { + "bbox": [ + 309, + 307, + 377, + 376 + ], + "lines": [ + { + "bbox": [ + 309, + 307, + 377, + 376 + ], + "spans": [ + { + "bbox": [ + 309, + 307, + 377, + 376 + ], + "type": "image", + "image_path": "cb0f7c8c84c5387800ec55ef6fa35fb245fc4927586a4eef2b516291155386fc.jpg" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_body" + } + ], + "index": 31 + }, + { + "type": "image", + "bbox": [ + 380, + 307, + 449, + 376 + ], + "blocks": [ + { + "bbox": [ + 380, + 307, + 449, + 376 + ], + "lines": [ + { + "bbox": [ + 380, + 307, + 449, + 376 + ], + "spans": [ + { + "bbox": [ + 380, + 307, + 449, + 376 + ], + "type": "image", + "image_path": "5829854f416da08cf0b2343ab41c3cb501926da7c78ff9d3409f2fe80a1802ad.jpg" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_body" + } + ], + "index": 32 + }, + { + "type": "image", + "bbox": [ + 451, + 307, + 520, + 376 + ], + "blocks": [ + { + "bbox": [ + 451, + 307, + 520, + 376 + ], + "lines": [ + { + "bbox": [ + 451, + 307, + 520, + 376 + ], + "spans": [ + { + "bbox": [ + 451, + 307, + 520, + 376 + ], + "type": "image", + "image_path": "6566911f56d82707fd59c1013b00d85363f7c22a5a0166bb0d5eecf8180ca9e3.jpg" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_body" + } + ], + "index": 33 + }, + { + "bbox": [ + 46, + 429, + 287, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 429, + 287, + 453 + ], + "spans": [ + { + "bbox": [ + 46, + 429, + 287, + 453 + ], + "type": "text", + "content": "tions. DragBench contains 205 images, with well-annotated prompts, masks, and dragging points." + } + ] + } + ], + "index": 35 + }, + { + "bbox": [ + 46, + 510, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 510, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 510, + 288, + 715 + ], + "type": "text", + "content": "Evaluation metrics. We evaluate the dragging results from two aspects: the quality of dragged image and the accuracy of position of handle points. We employ Image Fidelity (IF) [13] and Mean Distance (MD) [21] to assess these two aspects, respectively. IF evaluates the similarity between the input and output images, which is calculated by subtracting the mean LPIPS [32] over all pairs of original and edited images from 1. MD quantifies how well the handle points move to the target points. The same as DragDiffusion, we utilize DIFT [28] to identify the points in the output image that correspond to the handle points in the input image. These identified points are then treated as the definitive handle points after the dragging process. Subsequently, MD is calculated as the average Euclidean distance between the positions of all target points and their corresponding final handle points. A higher IF and a lower MD indicate a better quality of the dragging approach." + } + ] + } + ], + "index": 36 + }, + { + "bbox": [ + 306, + 428, + 432, + 441 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 428, + 432, + 441 + ], + "spans": [ + { + "bbox": [ + 306, + 428, + 432, + 441 + ], + "type": "text", + "content": "4.1. Qualitative evaluation" + } + ] + } + ], + "index": 37 + }, + { + "bbox": [ + 305, + 448, + 545, + 508 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 448, + 545, + 508 + ], + "spans": [ + { + "bbox": [ + 305, + 448, + 545, + 508 + ], + "type": "text", + "content": "We conduct extensive experiments for the qualitative comparison between our method and DragDiffusion. To provide a thorough evaluation, we adopt SD-1.5 [24] as our base model for all real images, and also employ other fine-tuned variants of SD-1.5 for generated image manipulation." + } + ] + } + ], + "index": 38 + }, + { + "bbox": [ + 304, + 509, + 546, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 509, + 546, + 653 + ], + "spans": [ + { + "bbox": [ + 304, + 509, + 546, + 653 + ], + "type": "text", + "content": "We present real image editing results for various object categories and user inputs in Fig. 4. Our approach accurately moves the handle points to reach the target positions, resulting in a wide range of natural and diverse manipulation effects such as altering the poses of animals, rearranging landscapes, modifying human body parts, and reshaping objects, in the absence of the mask. Conversely, DragDiffusion is unable to precisely move the handle points to the target positions, which is mainly attributed to the misleading of point tracking that is employed in DragGAN and DragDiffusion. It is worth noting that DragDiffusion always changes the non-dragged regions when no mask is provided." + } + ] + } + ], + "index": 39 + }, + { + "bbox": [ + 304, + 653, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 653, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 653, + 547, + 715 + ], + "type": "text", + "content": "We also conduct a series of point manipulations based on the generated images, as shown in Fig. 5. The classifier-free guidance scale is configured at 7.5 for various SD-1.5 variants. Notably, the reference guidance is unnecessary for generative editing, as the prompt of generated images pro" + } + ] + } + ], + "index": 40 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8409" + } + ] + } + ], + "index": 41 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 99, + 71, + 497, + 194 + ], + "blocks": [ + { + "bbox": [ + 99, + 71, + 497, + 194 + ], + "lines": [ + { + "bbox": [ + 99, + 71, + 497, + 194 + ], + "spans": [ + { + "bbox": [ + 99, + 71, + 497, + 194 + ], + "type": "image", + "image_path": "f07b53341e22f2521951b6396f31e03c7b5aae1818e914cc080cdfa8dc9d40ac.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 203, + 547, + 228 + ], + "lines": [ + { + "bbox": [ + 46, + 203, + 547, + 228 + ], + "spans": [ + { + "bbox": [ + 46, + 203, + 547, + 228 + ], + "type": "text", + "content": "Figure 6. Out-of-distribution manipulations. In some cases, the input image cannot be reconstructed with DDIM Inversion. Utilizing reference guidance, edited image preserves its identity with the original image." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 50, + 243, + 286, + 345 + ], + "blocks": [ + { + "bbox": [ + 50, + 243, + 286, + 345 + ], + "lines": [ + { + "bbox": [ + 50, + 243, + 286, + 345 + ], + "spans": [ + { + "bbox": [ + 50, + 243, + 286, + 345 + ], + "type": "table", + "html": "
MethodIF(↑)MD(↓)Time(s)
DragGAN0.69557.359.1
DragDiffusion0.88236.7169.4
DragDiffusion w/o LoRA0.86353.6333.7
DragDiffusion w/o mask0.76647.0365.6
SDE-Drag0.87145.4450.1
Ours w/ mask0.88929.0727.3
Ours0.91032.5530.6
", + "image_path": "12a617a0c7a404eb78c076562a59c7e946c4e8a663cf57c3c3a27903f7c5f91f.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 352, + 288, + 385 + ], + "lines": [ + { + "bbox": [ + 46, + 352, + 288, + 385 + ], + "spans": [ + { + "bbox": [ + 46, + 352, + 288, + 385 + ], + "type": "text", + "content": "Table 1. Quantitative evaluation of different methods with Drag-Bench. Evaluation metrics include Image Fidelity (IF) and Mean Distance (MD)." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 412, + 287, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 412, + 287, + 533 + ], + "spans": [ + { + "bbox": [ + 46, + 412, + 287, + 533 + ], + "type": "text", + "content": "vides substantial guidance during the denoising process. As shown in Fig. 5, our method achieves point-based editing with high quality and precision on generated images. It can be observed that in some cases, the details of generated images have changed. This is because the self-attention layer in diffusion models may introduce slight variation in other parts when a specific area of the image is edited. If users want to preserve certain areas without any changes, they can employ inpainting during image generation with masks to determine which regions should remain unchanged." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 548, + 180, + 561 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 548, + 180, + 561 + ], + "spans": [ + { + "bbox": [ + 47, + 548, + 180, + 561 + ], + "type": "text", + "content": "4.2. Quantitative evaluation" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 570, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 570, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 570, + 288, + 713 + ], + "type": "text", + "content": "We conduct quantitative evaluation by comparing our method to the baseline dragging method on diffusion, DragDiffusion. The evaluation, which incorporates IF and MD metrics, makes use of the DragBench dataset provided by DragDiffusion. The results, as shown in Tab. 1, demonstrate that our approach significantly surpasses DragDiffusion and SDE-Drag on the DragBench dataset. Notably, our method achieves superior dragging results without requiring LoRA fine-tuning, offering substantial time savings during extensive experiments. The performance of DragDiffusion degrades significantly without LoRA nor mask. The execution time in Tab. 1 is calculated on a single A100 GPU." + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 307, + 247, + 545, + 376 + ], + "blocks": [ + { + "bbox": [ + 307, + 247, + 545, + 376 + ], + "lines": [ + { + "bbox": [ + 307, + 247, + 545, + 376 + ], + "spans": [ + { + "bbox": [ + 307, + 247, + 545, + 376 + ], + "type": "image", + "image_path": "fc6519bc1e51dec6c62cc45781aa2504cdd8c46afda07d63298c44221fff5efd.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 385, + 545, + 407 + ], + "lines": [ + { + "bbox": [ + 305, + 385, + 545, + 407 + ], + "spans": [ + { + "bbox": [ + 305, + 385, + 545, + 407 + ], + "type": "text", + "content": "Figure 7. When users do not provide an appropriate mask, the generated results are prone to issues." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 429, + 382, + 440 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 429, + 382, + 440 + ], + "spans": [ + { + "bbox": [ + 306, + 429, + 382, + 440 + ], + "type": "text", + "content": "4.3. Discussions" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 447, + 545, + 626 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 447, + 545, + 626 + ], + "spans": [ + { + "bbox": [ + 304, + 447, + 545, + 626 + ], + "type": "text", + "content": "Reference Guidance. With our reference guidance, we enable the diffusion models to perform out-of-distribution manipulations. As shown in Fig. 6, the original image cannot be faithfully reconstructed using DDIM Inversion. When we introduce reference guidance, " + }, + { + "bbox": [ + 304, + 447, + 545, + 626 + ], + "type": "inline_equation", + "content": "z_{t}" + }, + { + "bbox": [ + 304, + 447, + 545, + 626 + ], + "type": "text", + "content": " can better reconstruct the input image. It is worth noting that " + }, + { + "bbox": [ + 304, + 447, + 545, + 626 + ], + "type": "inline_equation", + "content": "z_{i}(i = t,t - 1,\\dots,1)" + }, + { + "bbox": [ + 304, + 447, + 545, + 626 + ], + "type": "text", + "content": " is obtained during the DDIM Inversion process, so it contains richer information compared to " + }, + { + "bbox": [ + 304, + 447, + 545, + 626 + ], + "type": "inline_equation", + "content": "z_{t}" + }, + { + "bbox": [ + 304, + 447, + 545, + 626 + ], + "type": "text", + "content": ". After editing " + }, + { + "bbox": [ + 304, + 447, + 545, + 626 + ], + "type": "inline_equation", + "content": "z_{t}" + }, + { + "bbox": [ + 304, + 447, + 545, + 626 + ], + "type": "text", + "content": ", the image generated from " + }, + { + "bbox": [ + 304, + 447, + 545, + 626 + ], + "type": "inline_equation", + "content": "z_{t}^{*}" + }, + { + "bbox": [ + 304, + 447, + 545, + 626 + ], + "type": "text", + "content": " is very different from the input image in terms of identity. As we increase the reference guidance scale " + }, + { + "bbox": [ + 304, + 447, + 545, + 626 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 304, + 447, + 545, + 626 + ], + "type": "text", + "content": ", the consistency between the generated image and original image gradually improves. However, if " + }, + { + "bbox": [ + 304, + 447, + 545, + 626 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 304, + 447, + 545, + 626 + ], + "type": "text", + "content": " is too large, the generated image may become oversaturated. In our experiments, a value for " + }, + { + "bbox": [ + 304, + 447, + 545, + 626 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 304, + 447, + 545, + 626 + ], + "type": "text", + "content": " in the range of 3 to 5 is considered a good choice." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 642, + 546, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 642, + 546, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 642, + 546, + 715 + ], + "type": "text", + "content": "Auto mask generation. When users create imprecise masks, the results are unpredictable. As shown in Fig. 7, when the mask is too small, it will restrict the editing of the image. On the other hand, when the mask is too large, changes may occur in the background. Therefore, an appropriate mask is necessary for point-based image editing. Our" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8410" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 51, + 71, + 287, + 196 + ], + "blocks": [ + { + "bbox": [ + 51, + 71, + 287, + 196 + ], + "lines": [ + { + "bbox": [ + 51, + 71, + 287, + 196 + ], + "spans": [ + { + "bbox": [ + 51, + 71, + 287, + 196 + ], + "type": "image", + "image_path": "67076ba86c73d6678f9264f69f5a21174c6b306f5836f284cd78a169a1750322.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 205, + 288, + 239 + ], + "lines": [ + { + "bbox": [ + 46, + 205, + 288, + 239 + ], + "spans": [ + { + "bbox": [ + 46, + 205, + 288, + 239 + ], + "type": "text", + "content": "Figure 8. The changes in the gradient and prediction of " + }, + { + "bbox": [ + 46, + 205, + 288, + 239 + ], + "type": "inline_equation", + "content": "z_{t}" + }, + { + "bbox": [ + 46, + 205, + 288, + 239 + ], + "type": "text", + "content": " during optimization. The term of \"K-th step\" means the K-th update of handle points " + }, + { + "bbox": [ + 46, + 205, + 288, + 239 + ], + "type": "inline_equation", + "content": "p^{*}" + }, + { + "bbox": [ + 46, + 205, + 288, + 239 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 50, + 255, + 160, + 337 + ], + "blocks": [ + { + "bbox": [ + 50, + 255, + 160, + 337 + ], + "lines": [ + { + "bbox": [ + 50, + 255, + 160, + 337 + ], + "spans": [ + { + "bbox": [ + 50, + 255, + 160, + 337 + ], + "type": "image", + "image_path": "7ca4e4504b8eb48648883887a7a59122d81e3212d6946843369e789a264b6911.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 58, + 342, + 152, + 351 + ], + "lines": [ + { + "bbox": [ + 58, + 342, + 152, + 351 + ], + "spans": [ + { + "bbox": [ + 58, + 342, + 152, + 351 + ], + "type": "text", + "content": "(a) Which upper layer to use." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 120, + 361, + 212, + 373 + ], + "lines": [ + { + "bbox": [ + 120, + 361, + 212, + 373 + ], + "spans": [ + { + "bbox": [ + 120, + 361, + 212, + 373 + ], + "type": "text", + "content": "Figure 9. Ablation study." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 171, + 252, + 284, + 337 + ], + "blocks": [ + { + "bbox": [ + 171, + 252, + 284, + 337 + ], + "lines": [ + { + "bbox": [ + 171, + 252, + 284, + 337 + ], + "spans": [ + { + "bbox": [ + 171, + 252, + 284, + 337 + ], + "type": "image", + "image_path": "c8a4132df9a3745e3f02c9e7d0aa42f6cb2d6e0e82b1c865b92aa648ef6d3cc2.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 201, + 342, + 254, + 350 + ], + "lines": [ + { + "bbox": [ + 201, + 342, + 254, + 350 + ], + "spans": [ + { + "bbox": [ + 201, + 342, + 254, + 350 + ], + "type": "text", + "content": "(b) Effects of K." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 395, + 287, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 395, + 287, + 430 + ], + "spans": [ + { + "bbox": [ + 46, + 395, + 287, + 430 + ], + "type": "text", + "content": "auto mask generation method liberates users from the need to manually draw masks, which is user-friendly, especially for novice users." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 432, + 287, + 587 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 432, + 287, + 587 + ], + "spans": [ + { + "bbox": [ + 46, + 432, + 287, + 587 + ], + "type": "text", + "content": "Figure 8 shows the update of gradient and prediction with the optimization of " + }, + { + "bbox": [ + 46, + 432, + 287, + 587 + ], + "type": "inline_equation", + "content": "z_{t}" + }, + { + "bbox": [ + 46, + 432, + 287, + 587 + ], + "type": "text", + "content": ". It can be observed that when the handle points update, the gradient of " + }, + { + "bbox": [ + 46, + 432, + 287, + 587 + ], + "type": "inline_equation", + "content": "z_{t}" + }, + { + "bbox": [ + 46, + 432, + 287, + 587 + ], + "type": "text", + "content": " will change simultaneously. In the K-th step, the whole arm of the robot needs to be lifted, so the gradients for the arm part are relatively large. However, by the time we reach the (K+1)-th step, the robot's arm has already been dragged into the right place, and only the palm part needs to extend outward. Therefore, only the gradients for the palm part are high. Our auto mask effectively captures these details and incorporates this variation into the updates of " + }, + { + "bbox": [ + 46, + 432, + 287, + 587 + ], + "type": "inline_equation", + "content": "z_{t}" + }, + { + "bbox": [ + 46, + 432, + 287, + 587 + ], + "type": "text", + "content": ", thereby providing better control over the direction of " + }, + { + "bbox": [ + 46, + 432, + 287, + 587 + ], + "type": "inline_equation", + "content": "z_{t}" + }, + { + "bbox": [ + 46, + 432, + 287, + 587 + ], + "type": "text", + "content": ". That is the reason why the IF metric of without mask is higher than that with mask in Tab. 1." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 605, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 287, + 715 + ], + "type": "text", + "content": "Ablation study. We perform ablation studies to analyze the impact of the feature used in stable motion supervision, and we evaluate the performance on DragBench, as shown in Fig. 9a. Notably, the concatenated feature maps from upper layers 1&2 exhibit the best performance, demonstrating an optimal balance between resolution and discriminativeness. Additionally, we investigate the effects of different values of " + }, + { + "bbox": [ + 46, + 605, + 287, + 715 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 46, + 605, + 287, + 715 + ], + "type": "text", + "content": " in Fig. 9b, and it is observed that " + }, + { + "bbox": [ + 46, + 605, + 287, + 715 + ], + "type": "inline_equation", + "content": "K = 12" + }, + { + "bbox": [ + 46, + 605, + 287, + 715 + ], + "type": "text", + "content": " yields better results." + } + ] + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 330, + 70, + 523, + 121 + ], + "blocks": [ + { + "bbox": [ + 330, + 70, + 523, + 121 + ], + "lines": [ + { + "bbox": [ + 330, + 70, + 523, + 121 + ], + "spans": [ + { + "bbox": [ + 330, + 70, + 523, + 121 + ], + "type": "table", + "html": "
MethodsIF(↑)MD(↓)
DragGAN0.85651.32
DragonDiffusion0.89952.21
Stable motion supervision0.88929.07
", + "image_path": "d8b7c323fd0a5192187cf97328708488c32ee7df2a1c891587e1500a33ceb083.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 129, + 545, + 163 + ], + "lines": [ + { + "bbox": [ + 305, + 129, + 545, + 163 + ], + "spans": [ + { + "bbox": [ + 305, + 129, + 545, + 163 + ], + "type": "text", + "content": "Table 2. Quantitative comparison of motion supervision between our method and previous works. Our approach achieves more accurate dragging results than other methods." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 318, + 175, + 533, + 239 + ], + "blocks": [ + { + "bbox": [ + 318, + 175, + 533, + 239 + ], + "lines": [ + { + "bbox": [ + 318, + 175, + 533, + 239 + ], + "spans": [ + { + "bbox": [ + 318, + 175, + 533, + 239 + ], + "type": "image", + "image_path": "20d93ef55d4de00d352344e33e002f8dce81dca5326e5fc65bcfc337468cc134.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 247, + 545, + 281 + ], + "lines": [ + { + "bbox": [ + 305, + 247, + 545, + 281 + ], + "spans": [ + { + "bbox": [ + 305, + 247, + 545, + 281 + ], + "type": "text", + "content": "Figure 10. When the manipulated portion in the image is very small or the dragging distance is extensive, it is challenging for the dragging to have a noticeable effect." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 302, + 545, + 374 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 302, + 545, + 374 + ], + "spans": [ + { + "bbox": [ + 305, + 302, + 545, + 374 + ], + "type": "text", + "content": "A comparison of \"drag\" method between our approach and previous methods is shown in Tab. 2. We only replace stable motion supervision with the \"drag\" method in other works. It can be found that stable motion supervision is more suitable for the feature space of diffusion models, although DragonDiffusion achieves slightly higher IF." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 390, + 545, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 390, + 545, + 475 + ], + "spans": [ + { + "bbox": [ + 305, + 390, + 545, + 475 + ], + "type": "text", + "content": "Limitations. Despite the outstanding performance, our method has some limitations. Particularly, when the editing area is too small or the dragging distance is too long, the dragging result may be unsatisfactory. As depicted in Fig. 10, when the girl's face occupies only a small part of the image, and we intend to edit the mouth, the desired changes of mouth do not occur as expected." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 485, + 378, + 498 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 485, + 378, + 498 + ], + "spans": [ + { + "bbox": [ + 306, + 485, + 378, + 498 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 506, + 545, + 673 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 506, + 545, + 673 + ], + "spans": [ + { + "bbox": [ + 304, + 506, + 545, + 673 + ], + "type": "text", + "content": "In this work, we attribute an excellent point-based image editing method to three criteria: user-friendly interaction, good performance, and effective generalization. We abandon the LoRA fine-tuning and masks, which is convenient for users. Furthermore, we propose a stable motion supervision method and reference guidance to generate high-quality results based on diffusion models. Leveraging the powerful generative capabilities of the diffusion model, we achieve impressive dragging results across various domains. However, in our experiments, we observe that for very small-scale manipulations or long-distance dragging in images, the results are often not satisfactory. Diffusion models with stronger generative capabilities may be needed to achieve a broader range of dragging manipulations." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 306, + 689, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 545, + 712 + ], + "type": "text", + "content": "Acknowledgements: This work is partially supported by Natural Science Foundation of China Grant 41930110." + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "8411" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "type": "text", + "content": "[1] Rameen Abdal, Peihao Zhu, Niloy J Mitra, and Peter Wonka. Styleflow: Attribute-conditioned exploration of stylegan-generated images using conditional continuous normalizing flows. ACM Transactions on Graphics (ToG), 40(3):1-21, 2021. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 147, + 288, + 192 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 147, + 288, + 192 + ], + "spans": [ + { + "bbox": [ + 53, + 147, + 288, + 192 + ], + "type": "text", + "content": "[2] Omri Avrahami, Dani Lischinski, and Ohad Fried. Blended diffusion for text-driven editing of natural images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18208-18218, 2022. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 194, + 288, + 247 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 194, + 288, + 247 + ], + "spans": [ + { + "bbox": [ + 53, + 194, + 288, + 247 + ], + "type": "text", + "content": "[3] Tim Brooks, Aleksander Holynski, and Alexei A Efros. Instructpix2pix: Learning to follow image editing instructions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18392-18402, 2023. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 248, + 288, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 248, + 288, + 293 + ], + "spans": [ + { + "bbox": [ + 53, + 248, + 288, + 293 + ], + "type": "text", + "content": "[4] Mingdeng Cao, Xintao Wang, Zhongang Qi, Ying Shan, Xiaohu Qie, and Yinqiang Zheng. Masactrl: Tuning-free mutual self-attention control for consistent image synthesis and editing. arXiv preprint arXiv:2304.08465, 2023. 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 294, + 288, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 294, + 288, + 327 + ], + "spans": [ + { + "bbox": [ + 53, + 294, + 288, + 327 + ], + "type": "text", + "content": "[5] Yuki Endo. User-controllable latent transformer for stylegan image layout editing. In Computer Graphics Forum, pages 395-406. Wiley Online Library, 2022. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 328, + 288, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 328, + 288, + 373 + ], + "spans": [ + { + "bbox": [ + 53, + 328, + 288, + 373 + ], + "type": "text", + "content": "[6] Dave Epstein, Taesung Park, Richard Zhang, Eli Shechtman, and Alexei A Efros. Blobgan: Spatially disentangled scene representations. In European Conference on Computer Vision, pages 616-635. Springer, 2022. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 374, + 288, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 374, + 288, + 418 + ], + "spans": [ + { + "bbox": [ + 53, + 374, + 288, + 418 + ], + "type": "text", + "content": "[7] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-prompt image editing with cross attention control. arXiv preprint arXiv:2208.01626, 2022. 1, 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 419, + 288, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 419, + 288, + 441 + ], + "spans": [ + { + "bbox": [ + 53, + 419, + 288, + 441 + ], + "type": "text", + "content": "[8] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598, 2022. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 443, + 288, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 443, + 288, + 475 + ], + "spans": [ + { + "bbox": [ + 53, + 443, + 288, + 475 + ], + "type": "text", + "content": "[9] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 1, 2, 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 477, + 287, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 477, + 287, + 521 + ], + "spans": [ + { + "bbox": [ + 48, + 477, + 287, + 521 + ], + "type": "text", + "content": "[10] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 522, + 287, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 522, + 287, + 567 + ], + "spans": [ + { + "bbox": [ + 48, + 522, + 287, + 567 + ], + "type": "text", + "content": "[11] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4401-4410, 2019. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 567, + 288, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 567, + 288, + 622 + ], + "spans": [ + { + "bbox": [ + 48, + 567, + 288, + 622 + ], + "type": "text", + "content": "[12] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of stylegan. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8110-8119, 2020. 1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 624, + 288, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 624, + 288, + 679 + ], + "spans": [ + { + "bbox": [ + 48, + 624, + 288, + 679 + ], + "type": "text", + "content": "[13] Bahjat Kawar, Shiran Zada, Oran Lang, Omer Tov, Huiwen Chang, Tali Dekel, Inbar Mosseri, and Michal Irani. Imagic: Text-based real image editing with diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6007-6017, 2023. 1, 6" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 680, + 288, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 680, + 288, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 680, + 288, + 712 + ], + "type": "text", + "content": "[14] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014.5" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 712 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 95 + ], + "type": "text", + "content": "[15] Diederik P Kingma and Max Welling. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114, 2013. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 95, + 545, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 95, + 545, + 139 + ], + "spans": [ + { + "bbox": [ + 307, + 95, + 545, + 139 + ], + "type": "text", + "content": "[16] Chenlin Meng, Yutong He, Yang Song, Jiaming Song, Jia-jun Wu, Jun-Yan Zhu, and Stefano Ermon. Sdedit: Guided image synthesis and editing with stochastic differential equations. arXiv preprint arXiv:2108.01073, 2021. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 140, + 545, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 140, + 545, + 183 + ], + "spans": [ + { + "bbox": [ + 308, + 140, + 545, + 183 + ], + "type": "text", + "content": "[17] Daiki Miyake, Akihiro Iohara, Yu Saito, and Toshiyuki Tanaka. Negative-prompt inversion: Fast image inversion for editing with text-guided diffusion models. arXiv preprint arXiv:2305.16807, 2023. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 184, + 547, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 184, + 547, + 239 + ], + "spans": [ + { + "bbox": [ + 308, + 184, + 547, + 239 + ], + "type": "text", + "content": "[18] Ron Mokady, Amir Hertz, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Null-text inversion for editing real images using guided diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6038–6047, 2023. 1, 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 239, + 545, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 239, + 545, + 282 + ], + "spans": [ + { + "bbox": [ + 308, + 239, + 545, + 282 + ], + "type": "text", + "content": "[19] Chong Mou, Xintao Wang, Jiechong Song, Ying Shan, and Jian Zhang. Dragondiffusion: Enabling drag-style manipulation on diffusion models. arXiv preprint arXiv:2307.02421, 2023. 2, 4, 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 284, + 545, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 284, + 545, + 327 + ], + "spans": [ + { + "bbox": [ + 308, + 284, + 545, + 327 + ], + "type": "text", + "content": "[20] Shen Nie, Hanzhong Allan Guo, Cheng Lu, Yuhao Zhou, Chenyu Zheng, and Chongxuan Li. The blessing of randomness: Sde beats ode in general diffusion-based image editing. arXiv preprint arXiv:2311.01410, 2023. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 327, + 545, + 382 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 327, + 545, + 382 + ], + "spans": [ + { + "bbox": [ + 308, + 327, + 545, + 382 + ], + "type": "text", + "content": "[21] Xingang Pan, Ayush Tewari, Thomas Leimkuhler, Lingjie Liu, Abhimitra Meka, and Christian Theobalt. Drag your gan: Interactive point-based manipulation on the generative image manifold. In ACM SIGGRAPH 2023 Conference Proceedings, pages 1-11, 2023. 1, 2, 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 383, + 545, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 383, + 545, + 426 + ], + "spans": [ + { + "bbox": [ + 308, + 383, + 545, + 426 + ], + "type": "text", + "content": "[22] Gaurav Parmar, Krishna Kumar Singh, Richard Zhang, Yijun Li, Jingwan Lu, and Jun-Yan Zhu. Zero-shot image-to-image translation. In ACM SIGGRAPH 2023 Conference Proceedings, pages 1-11, 2023. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 426, + 545, + 469 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 426, + 545, + 469 + ], + "spans": [ + { + "bbox": [ + 308, + 426, + 545, + 469 + ], + "type": "text", + "content": "[23] Daniel Roich, Ron Mokady, Amit H Bermano, and Daniel Cohen-Or. Pivotal tuning for latent-based editing of real images. ACM Transactions on graphics (TOG), 42(1):1-13, 2022. 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 471, + 545, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 471, + 545, + 525 + ], + "spans": [ + { + "bbox": [ + 308, + 471, + 545, + 525 + ], + "type": "text", + "content": "[24] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10684-10695, 2022. 1, 2, 3, 6" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 526, + 545, + 591 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 526, + 545, + 591 + ], + "spans": [ + { + "bbox": [ + 308, + 526, + 545, + 591 + ], + "type": "text", + "content": "[25] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in Neural Information Processing Systems, 35:36479-36494, 2022. 1" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 592, + 545, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 592, + 545, + 635 + ], + "spans": [ + { + "bbox": [ + 308, + 592, + 545, + 635 + ], + "type": "text", + "content": "[26] Yujun Shi, Chuhui Xue, Jiachun Pan, Wenqing Zhang, Vincent YF Tan, and Song Bai. Dragdiffusion: Harnessing diffusion models for interactive point-based image editing. arXiv preprint arXiv:2306.14435, 2023. 2, 4, 5" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 636, + 545, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 636, + 545, + 669 + ], + "spans": [ + { + "bbox": [ + 308, + 636, + 545, + 669 + ], + "type": "text", + "content": "[27] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 1, 2, 3" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 670, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 670, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 308, + 670, + 545, + 712 + ], + "type": "text", + "content": "[28] Luming Tang, Menglin Jia, Qianqian Wang, Cheng Perng Phoo, and Bharath Hariharan. Emergent correspondence from image diffusion. arXiv preprint arXiv:2306.03881, 2023.4,6" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8412" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 373 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "text", + "content": "[29] Jianyuan Wang, Ceyuan Yang, Yinghao Xu, Yujun Shen, Hongdong Li, and Bolei Zhou. Improving gan equilibrium by raising spatial awareness. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11285-11293, 2022. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 129, + 287, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 129, + 287, + 161 + ], + "spans": [ + { + "bbox": [ + 48, + 129, + 287, + 161 + ], + "type": "text", + "content": "[30] Sheng-Yu Wang, David Bau, and Jun-Yan Zhu. Rewriting geometric rules of a gan. ACM Transactions on Graphics (TOG), 41(4):1-16, 2022. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 163, + 287, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 163, + 287, + 206 + ], + "spans": [ + { + "bbox": [ + 48, + 163, + 287, + 206 + ], + "type": "text", + "content": "[31] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3836-3847, 2023. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 208, + 287, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 208, + 287, + 262 + ], + "spans": [ + { + "bbox": [ + 48, + 208, + 287, + 262 + ], + "type": "text", + "content": "[32] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 263, + 287, + 318 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 263, + 287, + 318 + ], + "spans": [ + { + "bbox": [ + 48, + 263, + 287, + 318 + ], + "type": "text", + "content": "[33] Zhixing Zhang, Ligong Han, Arnab Ghosh, Dimitris N Metaxas, and Jian Ren. Sine: Single image editing with text-to-image diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6027-6037, 2023. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 319, + 287, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 319, + 287, + 373 + ], + "spans": [ + { + "bbox": [ + 48, + 319, + 287, + 373 + ], + "type": "text", + "content": "[34] Jiapeng Zhu, Ceyuan Yang, Yujun Shen, Zifan Shi, Bo Dai, Deli Zhao, and Qifeng Chen. Linkgan: Linking gan latents to pixels for controllable image synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7656-7666, 2023. 2" + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "8413" + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Eclipse_ Disambiguating Illumination and Materials using Unintended Shadows/a9a99e7b-5bee-4200-8036-13e74c55e36e_content_list.json b/2024/Eclipse_ Disambiguating Illumination and Materials using Unintended Shadows/a9a99e7b-5bee-4200-8036-13e74c55e36e_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..cd887e5ebc4a46d9de3f60894b7346c299e36b3d --- /dev/null +++ b/2024/Eclipse_ Disambiguating Illumination and Materials using Unintended Shadows/a9a99e7b-5bee-4200-8036-13e74c55e36e_content_list.json @@ -0,0 +1,2033 @@ +[ + { + "type": "text", + "text": "Eclipse: Disambiguating Illumination and Materials using Unintended Shadows", + "text_level": 1, + "bbox": [ + 83, + 130, + 887, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Dor Verbin", + "bbox": [ + 285, + 180, + 383, + 196 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Ben Mildenhall1", + "bbox": [ + 406, + 180, + 537, + 196 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Peter Hedman", + "bbox": [ + 560, + 180, + 681, + 196 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Jonathan T. Barron1", + "bbox": [ + 240, + 198, + 398, + 214 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Todd Zickler $^{1,2}$", + "bbox": [ + 421, + 198, + 545, + 214 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Pratul P. Srinivasan1", + "bbox": [ + 563, + 198, + 725, + 214 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1Google Research", + "bbox": [ + 320, + 215, + 462, + 234 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ Harvard University", + "bbox": [ + 485, + 215, + 645, + 234 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/fa3830b6da4b09f6b20011f65294f8112da2ea60f58383464472f4f750b6e842.jpg", + "image_caption": [ + "(a) Problem setup" + ], + "image_footnote": [], + "bbox": [ + 78, + 315, + 274, + 440 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/20bd31f14abca80306765369e4d607fc075573d7a45bbcc446da76ae2ed5df02.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 313, + 305, + 343, + 333 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/070bd0bd8dc3f3b994e92d68b7875826c86978889e196dd1a21eef26bac1e987.jpg", + "image_caption": [ + "Albedo", + "Illumination and occluders" + ], + "image_footnote": [], + "bbox": [ + 370, + 277, + 584, + 367 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/a5310ae4b2c78c42d0dc5e88c9801c49f93d0dad67194147c0171334cd72522b.jpg", + "image_caption": [ + "Albedo" + ], + "image_footnote": [], + "bbox": [ + 596, + 279, + 660, + 354 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/1e2c5fbe3d6b8eb10743d8fc365297a366652b3164fecd91f5ab0640a4c46f8b.jpg", + "image_caption": [ + "Illumination and occluders" + ], + "image_footnote": [], + "bbox": [ + 669, + 277, + 883, + 367 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/7e50a000c9c96cee488df38da1c6d47e2e25346c892d823d5eb7aacd2af78c93.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 295, + 378, + 359, + 452 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/39f08ffd84ba2f6ae864755c702a4163eb3829faeceeb6f04f312931c2c2dc15.jpg", + "image_caption": [ + "Figure 1. (a) We exploit unintended shadows cast by camera operators (or other unobserved moving occluders) to recover high-fidelity environment lighting and object materials from a set of images. Without modeling such unobserved occluders, prior methods can only: (b) recover convolved lighting without explicit material decomposition [31]; or (c) exploit lighting occlusions that occur internally among the observed objects [37]. (d) We show that additionally modeling and recovering external, unobserved occluders enables lighting and material reconstructions that are closer to ground truth (e)." + ], + "image_footnote": [], + "bbox": [ + 369, + 377, + 584, + 467 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/2ad95e1e242aebb2bdc0886673b5558f8b5e651de1490597bb48f0d0e5f3c046.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 596, + 377, + 661, + 452 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/692bf59ab3edbda586f02177204c85bfd6f3b277a773600180d0f64915325b43.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 669, + 376, + 883, + 467 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 558, + 312, + 571 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Decomposing an object's appearance into representations of its materials and the surrounding illumination is difficult, even when the object's 3D shape is known beforehand. This problem is especially challenging for diffuse objects: it is ill-conditioned because diffuse materials severely blur incoming light, and it is ill-posed because diffuse materials under high-frequency lighting can be indistinguishable from shiny materials under low-frequency lighting. We show that it is possible to recover precise materials and illumination—even from diffuse objects—by exploiting unintended shadows, like the ones cast onto an object by the photographer who moves around it. These shadows are a nuisance in most previous inverse rendering pipelines, but here we exploit them as signals that improve conditioning and help resolve material-lighting ambiguities. We present a method based on differentiable Monte Carlo ray tracing that uses images of an object to jointly recover its spatially-varying materials, the surrounding illumination environment, and the shapes of the unseen light occluders who inadvertently cast shadows upon it.", + "bbox": [ + 75, + 580, + 470, + 883 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 501, + 558, + 630, + 571 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this work, we show that the long-standing inverse rendering problem of recovering an object's material properties and the surrounding illumination environment from a set of images can greatly benefit from considering the effects of unobserved moving light occluders, such as the camera operator, which partially block incoming light and inadvertently cast shadows onto the object being imaged.", + "bbox": [ + 496, + 583, + 890, + 688 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Joint recovery of materials and lighting is a challenging task because the BRDF (bidirectional reflectance distribution function, which represents how a material at any point on an object's surface maps from incoming to outgoing light) acts as a directional filter on incoming light. BRDFs for specular (shiny) materials act as all-pass directional filters, while BRDFs for diffuse materials act as low-pass directional filters. As a result, inverse rendering is fundamentally ambiguous since a shiny object illuminated by blurry lighting can be indistinguishable from a diffuse object illuminated by sharp lighting (Figure 2). Furthermore, since diffuse materials act as low-pass filters and strongly blur incoming light, even the simpler problem of recovering lighting from images of a known diffuse material can be", + "bbox": [ + 496, + 689, + 892, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "77", + "bbox": [ + 490, + 945, + 506, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "severely ill-conditioned, precluding the recovery of high-frequency illumination [33].", + "bbox": [ + 76, + 90, + 467, + 121 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We observe that both of these issues can be ameliorated by exploiting an effect that naturally exists whenever an object is imaged from a sequence of viewpoints under static environment lighting. Namely, between one image and the next, the positions of the camera and its operator(s) must change, and in doing so they occlude different portions of the surrounding environment, casting distinct sets of shadows onto the object being imaged. This effect is very noticeable under strongly-directional lighting or for shiny objects because the cast shadows are sharp or the specular highlights are misshapen and hidden. But even when the effect is barely perceptible, such as for diffuse objects under well-distributed lighting, there exists a subtle signal that one can take advantage of.", + "bbox": [ + 75, + 122, + 467, + 333 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we make first steps toward a practical algorithm that exploits this cue by using gradient descent with differentiable Monte Carlo ray tracing to jointly recover explicit representations of (i) spatially-varying reflectance, (ii) environment illumination, and (iii) the per-image shapes of any unseen, light-blocking occluders. We evaluate our algorithm using a challenging variety of simulated scenes, including scenes with diffuse-only materials and with object geometry that is not known beforehand. We also apply our method to an existing off-the-shelf dataset designed for inverse rendering, where the (unseen) camera rig moving to capture the scene acts as an accidental occluder. Our results demonstrate that external shadowing effects provide a strong and useful cue, even when they are subtle, and even without relying on strong domain-specific priors for the materials, illumination, or external occluder shapes. This suggests that incorporating unintended shadows into inverse rendering pipelines for real, captured data is valuable for improving the quality of material and lighting assets.", + "bbox": [ + 75, + 335, + 467, + 623 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 76, + 640, + 217, + 655 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We build on recent developments in differentiable ray tracing and on a long history of inverse rendering work, including the joint estimation of reflectance and lighting, and the estimation of lighting alone. We are also inspired by a separate line of work on passive non-line-of-sight imaging, which exploits similar light-blocking effects.", + "bbox": [ + 76, + 667, + 467, + 758 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Physics-based differentiable ray tracing. Modern differentiable ray tracers [22, 28] enable the computation of gradients of rendered images with respect to scene parameters (i.e., geometry, materials, and lighting) by differentiating through light transport simulation. Recent works have focused on improving efficiency and performance [14, 29] and on accurately differentiating through visibility discontinuities—such as those caused by shadow and occlusion boundaries—with respect to shape and light-", + "bbox": [ + 76, + 763, + 467, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "ing [1, 42]. Our implementation of differentiable ray tracing leverages insights from these works as well as from Zeltner et al. [45], who provide design intuition for Monte Carlo differentiable ray tracers.", + "bbox": [ + 498, + 90, + 890, + 151 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Inverse rendering of materials and lighting. Decomposing an object's appearance into representations of material and lighting is a long-standing problem in computer vision and graphics. In their foundational work, Ramamoorthi and Hanrahan [33] developed a signal processing approach by describing the outgoing light at a surface point as the spherical convolution of the BRDF and the incoming lighting. This formulation elucidates why inverse rendering is ill-posed and often ill-conditioned: it is ill-posed because there are multiple illumination-material pairs that convolve into the same image, and it is ill-conditioned because diffuse BRDFs act as a low-pass filters on lighting, which causes the estimates of medium- and high-frequency illumination to be very sensitive to noise.", + "bbox": [ + 496, + 160, + 890, + 371 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Because of this, most approaches to inverse rendering rely on strong priors on materials and lighting. Single-view approaches have used hand-designed priors [3] or priors in the form of neural networks trained with supervision from large datasets of material and lighting labels [23, 24]. More relevant to us are multi-view approaches, which typically either assume known lighting [4, 5, 36] or rely on strong priors, such as assuming a single, highly-specular BRDF for the entire object [46] or lighting from a prior distribution that was pre-trained on a dataset of environment maps [6, 47]. The most closely related work is that for which lighting information is not provided as input [17, 27].", + "bbox": [ + 496, + 375, + 890, + 556 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "For the sake of generality, we do not use such strong priors in our experiments because by avoiding them we can more directly measure the benefits that are gained by modeling unintended shadows as an additional cue. Because our model uses a generic gradient-descent framework, stronger application-specific priors can be added to it to improve performance on a specific domain of interest.", + "bbox": [ + 496, + 559, + 890, + 664 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Estimating environment lighting. Another thread of inverse rendering research focuses on the task of recovering high-fidelity lighting environments, typically for augmented reality applications where virtual objects are rendered into photographs with consistent reflections and shadows. Prior work by Debevec [9] used images of a chrome sphere to measure environment lighting directly, and subsequent works have demonstrated that plausible environment lighting can be estimated without chrome spheres, using images of indoor [12, 35] and outdoor scenes [13, 19, 20], or images of a specific class of objects like faces [21, 30]. These subsequent techniques use strong priors in the form of deep neural network weights that are trained with supervision to map from images to lighting, and they do not enforce physical rendering consistency between the recovered", + "bbox": [ + 496, + 674, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "78", + "bbox": [ + 490, + 945, + 506, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "lighting and the observations. In contrast, we avoid such strong priors, and we explicitly enforce consistency.", + "bbox": [ + 75, + 90, + 468, + 121 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "More related to our approach is the work of Park et al. [31], which recovers physically-consistent environment maps from RGBD videos of shiny objects. However, their algorithm can only recover the convolution of the environment map with the BRDF of the observed object (e.g., Figure 1(b)), which precludes the recovery of high illumination frequencies unless the object is highly specular.", + "bbox": [ + 75, + 122, + 467, + 227 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Also related is the work of Swedish et al. [37], which places a known, diffuse object on a ground plane (a scenario first studied by Sato et al. [34]) and uses its cast shadows to recover high-quality lighting. Their formulation is linear and it improves lighting estimates by leveraging self-shadowing among diffuse objects that have known geometry and albedo (e.g., Figure 1(c)). We generalize this by replacing the linear formulation with differentiable Monte Carlo ray tracing, which allows exploiting the additional shadowing effects caused by moving external occluders and leads to substantially improved results (e.g., Figure 1(d)). Our formulation also handles objects with more general spatially-varying BRDFs that are not known beforehand.", + "bbox": [ + 75, + 229, + 467, + 425 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Passive non-line-of-sight imaging with occlusions. Passive non-line-of-sight techniques can also be seen as recovering environment illumination: They observe a reflective surface (which is typically diffuse and planar) and recover the appearance of a \"hidden scene\" from these reflections. Similar to us, prior works in this area have observed that the presence of an occluder between the observed surface and the hidden scene/environment aids recovery by introducing sharp angular variations into the rendering integral. This insight was first leveraged in settings with simple occluder-shapes like pinholes and pinspecks [39] or corners joining walls [7]. Subsequent work by Baradad et al. [2] generalizes this by considering light-occlusion effects caused by an arbitrary but known 3D shape (a task that is closely related to Swedish et al. above). Yedidia et al. [44] additionally recover the unobserved occluder's shape assuming it is a planar mask parallel to the observed planar surface.", + "bbox": [ + 75, + 431, + 467, + 688 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We generalize these prior works by using differentiable Monte Carlo rendering to replace their deconvolution-based algorithms, which are tailored to the specific case where the observed reflector is planar and diffuse. This allows using reflective objects that have arbitrary shapes, and arbitrary spatially-varying BRDFs which are not known beforehand. It also allows recovering the shapes of unobserved occluders that are arbitrary and time-varying.", + "bbox": [ + 75, + 690, + 467, + 811 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Motivation and Problem Setup", + "text_level": 1, + "bbox": [ + 76, + 828, + 359, + 845 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "A well-known ambiguity in computer vision and graphics occurs when decomposing an image into its component lighting and materials. The top of Figure 2 recreates", + "bbox": [ + 75, + 854, + 467, + 900 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/b046042743bcf125cc9ab4f2f0648209edda8e1e76468ba73c3b13f950e29bbe.jpg", + "image_caption": [ + "Figure 2. Decomposing appearance into lighting and materials can be inherently ambiguous. Top row: a diffuse sphere lit by high-frequency lighting (inset) is visually indistinguishable from a shiny, mirror-like sphere lit by low-frequency lighting. Bottom row: A second image captured in the presence of a dark external occluder resolves the ambiguity. The occluder's effect is clearly visible on the shiny sphere while on the diffuse sphere it produces shadows that are soft and very subtle." + ], + "image_footnote": [], + "bbox": [ + 500, + 87, + 890, + 309 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "a common depiction of this ambiguity, where an image of a known shape (a sphere) is explained equally well by a diffuse material in a complicated lighting environment or a shiny, mirror-like material under low-frequency lighting. Now, imagine we capture a second image from the same viewpoint, after some external object enters the scene. This second object remains beyond the field of view and so is not directly observed, but it acts as an external occluder that prevents some of the environment's light from reaching the sphere. This second image, shown at the bottom of the figure for each case, clearly reveals which of the two materiallighting explanations is the correct one.", + "bbox": [ + 496, + 460, + 890, + 641 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We aim to make use of this unintended shadowing that naturally occurs whenever a camera and its operators move around an object while capturing images from different viewpoints. They affect each image by blocking different portions of the surrounding light, and this provides a helpful signal. We will show that this signal is helpful even when the occluder shapes and locations are quite arbitrary and not known beforehand, and when their shadowing effects are very subtle, like between the top and bottom images on the left of Figure 2.", + "bbox": [ + 496, + 642, + 890, + 792 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In addition to helping resolve material-lighting ambiguities, shadowing from moving external occluders also improves the conditioning of our inverse rendering problem. This is analyzed in Figure 3, which considers the simplified hypothetical 1D case of recovering an environment illumination from images of a diffuse disk-shaped object, with known occluder positions. Convolution with a dif", + "bbox": [ + 496, + 795, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "79", + "bbox": [ + 490, + 945, + 506, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/6ed747df87e5cfd7f61229da52dbbbebf5eb52a1f438457e7547d63cd8e50879.jpg", + "image_caption": [ + "Figure 3. Here we plot eigenvalue magnitudes as a function of illumination angular frequency for a one-dimensional circular Lambertian object with known albedo, under the simplifying assumption of known occluder locations and shapes. For a single input image without occlusion, almost all lighting frequencies are strongly attenuated. In the single-occluder case (similar to the scenario considered by Baradad et al. [2]), some low frequencies become recoverable but intermediate and high frequencies do not. Observing additional images enables the recovery of more lighting information, especially at the highest frequencies." + ], + "image_footnote": [], + "bbox": [ + 81, + 90, + 464, + 292 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "fuse BRDF acts as a low-pass filter of the incoming light, which means that the linear inverse problem's eigenvalues are vanishingly low for higher frequencies. Moving external occluders introduce a sharp angular variation in the per-image illumination, which better-conditions these high frequency components, and allows their recovery. As shown in Figure 3, in this simplified scenario a single image containing an occluder is better than a single image without any occluders, but using observations of more occluders makes the problem even better-conditioned. We demonstrate this effect in practice in the supplement.", + "bbox": [ + 75, + 470, + 467, + 638 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1. Formulation", + "text_level": 1, + "bbox": [ + 76, + 650, + 210, + 666 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Let $\\{I_t\\}_{t=1}^T$ be a collection of $T$ images of a scene with known camera poses; for every $t$ , $I_t$ is an RGB image with spatial size $H \\times W$ . We assume the observed scene $\\mathcal{S} \\subset \\mathbb{R}^3$ is illuminated by a far-field environment map $L(\\hat{\\omega}_i)$ that does not change over time, so the only temporal changes in the incident light field are caused by unobserved light-blockers (called occluders hereafter) that move around the scene outside the field of view. We can then write the incident light at any observed position $\\mathbf{x} \\in \\mathcal{S}$ as:", + "bbox": [ + 75, + 674, + 467, + 811 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nL _ {t} (\\mathbf {x}, \\hat {\\omega} _ {i}) = L (\\hat {\\omega} _ {i}) M _ {t} (\\mathbf {x}, \\hat {\\omega} _ {i}) V (\\mathbf {x}, \\hat {\\omega} _ {i}), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 138, + 825, + 468, + 840 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where for any $t$ , $M_{t}(\\mathbf{x}, \\hat{\\boldsymbol{\\omega}}_{i})$ is a binary signal with value 0 for directions blocked by the occluder and 1 otherwise, and the binary signal $V(\\mathbf{x}, \\hat{\\boldsymbol{\\omega}}_{i})$ models visibility effects that are", + "bbox": [ + 76, + 854, + 468, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "internal to the scene, with value 0 for directions from $\\mathbf{x}$ that are blocked by other scene elements, and value 1 otherwise.", + "bbox": [ + 498, + 90, + 890, + 119 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Omitting global illumination effects, a pixel $\\mathbf{u}$ corresponding to a surface point $\\mathbf{x}$ with BRDF $f$ viewed from direction $\\hat{\\omega}_o$ at time $t$ has color:", + "bbox": [ + 498, + 122, + 890, + 167 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nI _ {t} (\\mathbf {u}) = \\int_ {\\mathbb {S} ^ {2}} L _ {t} (\\mathbf {x}, \\hat {\\omega} _ {i}) f (\\mathbf {x}, \\hat {\\omega} _ {i}, \\hat {\\omega} _ {o}) (\\hat {\\mathbf {n}} (\\mathbf {x}) \\cdot \\hat {\\boldsymbol {\\omega}} _ {i}) _ {+} d \\hat {\\boldsymbol {\\omega}} _ {i}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 506, + 179, + 890, + 210 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\hat{\\mathbf{n}} (\\mathbf{x})$ is the surface normal corresponding to $\\mathbf{x}$ , and $(\\cdot)_{+}$ clamps negative values to zero.", + "bbox": [ + 498, + 224, + 890, + 253 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Given the set of observed images, we would like to recover the environment map $L$ , a spatially-varying BRDF $f$ at every point on the surface of the object $S$ , as well as the set of unobserved occluder masks $\\{M_t\\}_{t=1}^T$ , one for each observation time $t$ . Our goal is therefore to solve the following optimization problem:", + "bbox": [ + 496, + 255, + 890, + 347 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\underset {\\phi^ {(o)}, \\phi^ {(m)}, \\phi^ {(\\ell)}} {\\arg \\min } \\sum_ {t, \\mathbf {u}} \\left\\| I _ {t} (\\mathbf {u}) - \\mathcal {R} _ {t} \\left(\\mathbf {u}; \\phi^ {(o)}, \\phi^ {(m)}, \\phi^ {(\\ell)}\\right) \\right\\| ^ {2}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 509, + 358, + 890, + 407 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathcal{R}_t(\\mathbf{u};\\cdot)$ renders the pixel location $\\mathbf{u}$ at time $t$ using the occluder parameters $\\phi^{(o)}$ , material parameters $\\phi^{(m)}$ , and illumination parameters $\\phi^{(\\ell)}$ .", + "bbox": [ + 496, + 409, + 890, + 455 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "At first glance, it may seem that solving for occluders in addition to materials and illumination should make the resulting inverse problem more difficult. However, as illustrated in Figure 3 and demonstrated by our experimental results, modeling and solving for occluders actually makes the full problem more well-conditioned and thus improves recovery of material and illumination.", + "bbox": [ + 496, + 457, + 890, + 561 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Method", + "text_level": 1, + "bbox": [ + 500, + 579, + 589, + 594 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We first describe our parameterizations of the unseen oc-clusters, illumination, and materials (Sections 4.1-4.3). Section 4.4 then relates these parameters to rendered pixel colors, and Section 4.5 describes how we optimize to solve the inverse problem in Equation 3.", + "bbox": [ + 496, + 604, + 890, + 680 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We emphasize that our method is designed to exploit external shadows across as many scene-types as possible, and so its only priors are those implicit to our parameterizations. In our experiments, we verify that these relatively weak priors are sufficient for recovering high-fidelity occluder shapes, illumination maps, and material maps.", + "bbox": [ + 496, + 681, + 890, + 773 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1.Occluders", + "text_level": 1, + "bbox": [ + 500, + 786, + 612, + 800 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Define a world coordinate system with origin at the scene's center, and let $\\{r_t\\}_{t=1}^T$ be the radial distances from the origin to each (known) camera's center. Then we model each of the $T$ occluders $M_t$ as an independent binary signal defined on the surface of a finite, radius- $r_t$ sphere that is centered at the origin. The value of the binary occluder signal", + "bbox": [ + 496, + 810, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "80", + "bbox": [ + 488, + 945, + 506, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/692f6c4c60ad0fde441093d2f689c0ca2f7930225543b1dc458cbf36616af74a.jpg", + "image_caption": [ + "PSNR = 31.2 dB" + ], + "image_footnote": [], + "bbox": [ + 99, + 108, + 197, + 164 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/bfb05bd52a53941c2916344e8532302f0842f14a221170725fc67aa512bfd39d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 218, + 109, + 305, + 162 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/013a22404f239d105b52f4540b3b8d4addf4f0d663cac0d11549c80dc352e0b0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 328, + 111, + 416, + 162 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/e20880b2ecb8cb8e5c828ef3f744b3d45a6fda19a408a9c1599b49cd3ce1f4b7.jpg", + "image_caption": [ + "PSNR = 26.4 dB" + ], + "image_footnote": [], + "bbox": [ + 101, + 200, + 202, + 270 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/8c849c23079cdbd9c7fca8520f96e9926475449810db2fce6a9e75b24164059a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 212, + 200, + 313, + 270 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/21e82359b9c42bf0e73dedbe36315326d37630eca6f14f4c6703fd6784fbf1bb.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 325, + 200, + 424, + 270 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/839e78ad849635b81b1ab2d428c1448386e2aafa8d5989a4bd62043becaf3698.jpg", + "image_caption": [ + "(a) Sample input image" + ], + "image_footnote": [], + "bbox": [ + 143, + 299, + 169, + 359 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/5f6df5ae024dcc3b2b02d44617ecda302035578f4a78db0b62e8929ee6eeb829.jpg", + "image_caption": [ + "PSNR = 29.6 dB" + ], + "image_footnote": [], + "bbox": [ + 254, + 299, + 281, + 359 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/722be76b5e4fd12bcd9170e47a0875fdf998f8a8cfadb15a7da83cf7a69c6681.jpg", + "image_caption": [ + "(c) True albedo" + ], + "image_footnote": [], + "bbox": [ + 364, + 300, + 392, + 359 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/4869fa775491944f014ae608cc50095f92748748a1489303987aeb8ab41b44da.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 436, + 88, + 648, + 172 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/9e91b7227707495da3ab7d67985eaf1e2d50ae51823e0d7b4b609dcb47f66a07.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 651, + 88, + 866, + 172 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/6a87389ff443254ee77238bac0be9b75c018ad39691f15fe93d1604e0c22c4cd.jpg", + "image_caption": [ + "$\\mathrm{RMSE} = 0.073$" + ], + "image_footnote": [], + "bbox": [ + 436, + 189, + 648, + 273 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/9605522f36447750b8164f9d293250441442b3f48c2461898c3bc392724d1c0b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 651, + 189, + 864, + 273 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/a8a7a34b84fbd30c914914c874079ebe93f8d0ec280a427470361edbe8f17b20.jpg", + "image_caption": [ + "RMSE = 0.081", + "$\\mathrm{RMSE} = 0.093$", + "(d) Recovered illumination" + ], + "image_footnote": [], + "bbox": [ + 436, + 290, + 648, + 375 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/44a2d5b874791ddb6c04886f520d02945a25e1405f676b69fdce90795f914751.jpg", + "image_caption": [ + "(e) True illumination" + ], + "image_footnote": [], + "bbox": [ + 651, + 290, + 864, + 375 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/7f99c9753ef13d7446e332cb81696c838627fc37380141451535b4f3411392bc.jpg", + "image_caption": [ + "Figure 4. The results of our method on three additional diffuse objects. We report the RMSE of each environment map in linear color space but plot the images after tonemapping for better evaluation of the full dynamic range. The albedo PSNR values are reported on object pixels only." + ], + "image_footnote": [], + "bbox": [ + 81, + 476, + 166, + 599 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/18e55259764e22af73ac545f6283f314a90f787e3698caaf833184ad1ff1ca68.jpg", + "image_caption": [ + "(b) Recovered albedo" + ], + "image_footnote": [], + "bbox": [ + 176, + 477, + 263, + 599 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/9ffa38b2fc7850f3bfd49bbc996dba852cdf282a83fbce33954c7d997d393786.jpg", + "image_caption": [ + "Figure 5. Our occluders extracted from the potatoes scene on the top of Figure 4 resemble the true ones (left). However, as expected, inaccuracies in the recovered masks tend to occur at relatively-dark regions of the environment map: there is a very strong inverse correlation (right) between the mean error in the recovered masks and the amount of illumination energy that they block." + ], + "image_footnote": [], + "bbox": [ + 269, + 484, + 467, + 625 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "for a shadow ray emitted from surface point $\\mathbf{x} \\in S$ in direction $\\hat{\\omega}_i$ at time $t$ can be computed by intersecting the shadow ray with the occluder's spherical shell:", + "bbox": [ + 75, + 744, + 468, + 790 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nM _ {t} (\\mathbf {x}, \\hat {\\omega} _ {i}) = \\tilde {M} _ {t} \\left(S (\\mathbf {x}, \\hat {\\omega} _ {i}, r _ {t})\\right) \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 165, + 800, + 468, + 818 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $S(\\mathbf{x},\\hat{\\omega}_i,r)$ is the intersection of the ray with a sphere of radius $r_t$ , normalized to have unit length:", + "bbox": [ + 76, + 829, + 468, + 859 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nS (\\mathbf {x}, \\hat {\\boldsymbol {\\omega}} _ {i}, r) = \\frac {\\mathbf {x} + \\left(\\sqrt {(\\mathbf {x} \\cdot \\hat {\\boldsymbol {\\omega}} _ {i}) ^ {2} + r ^ {2} - \\| \\mathbf {x} \\| ^ {2}} - \\mathbf {x} \\cdot \\hat {\\boldsymbol {\\omega}} _ {i}\\right) \\hat {\\boldsymbol {\\omega}} _ {i}}{r}. \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 84, + 869, + 468, + 905 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Note that there is a single intersection, since we assume that the surface points are always inside the occluder spheres.", + "bbox": [ + 498, + 479, + 890, + 508 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Although our formulation models the occluders as binary signals, a binary representation is not well-suited for gradient-based optimization. Instead, we represent the occluders as a continuously-valued function on the sphere in a spherical harmonic basis, mapped using a sigmoid function $\\sigma$ to lie in [0, 1]:", + "bbox": [ + 498, + 510, + 892, + 601 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {M} _ {t} (\\hat {\\omega}) = \\sigma \\left(\\sum_ {\\ell = 0} ^ {P} \\sum_ {m = - \\ell} ^ {\\ell} a _ {t \\ell m} Y _ {\\ell} ^ {m} (\\hat {\\omega})\\right), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 562, + 611, + 890, + 655 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\phi^{(o)}\\triangleq \\{a_{t\\ell m}\\}$ are estimizable coefficients that parameterize the occluder at time $t$ , and $P$ is the degree required to span the space of spherical images with the same resolution as our environment illumination (see below) [11].", + "bbox": [ + 496, + 667, + 890, + 731 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Note that for radially symmetric BRDFs, the rendering integral in Equation 2 can be written as a spherical convolution [11, 33]. This makes spherical harmonics a natural choice for representing the occluder signals, since spherical convolutions are diagonal in the basis of spherical harmonics. See our supplement for more details.", + "bbox": [ + 496, + 731, + 890, + 821 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Environment Illumination", + "text_level": 1, + "bbox": [ + 500, + 832, + 736, + 845 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We assume that the scene's illumination is distant and can be represented as an environment map. The environment map is parameterized as an image pyramid of size $H' \\times W'$", + "bbox": [ + 498, + 854, + 892, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "81", + "bbox": [ + 488, + 945, + 506, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/eb9121225b634bdf165cbf8c0046e239316bcd9c87059ee4f484eaeb5e08f6c2.jpg", + "image_caption": [ + "(a) Recovered illumination (no occluder)" + ], + "image_footnote": [], + "bbox": [ + 107, + 88, + 356, + 186 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/43aa90649c1995ebeff7b2492ffb2fb682441daad0fe10b35775f1f02d43b0fc.jpg", + "image_caption": [ + "(b) Recovered illumination", + "Figure 6. In the case of purely Lambertian material (with the potatoes geometry), (a) only using self occlusions is insufficient for the recovery of high-frequency content in the illumination. (b) Modeling and estimating occluders recovers significantly higher-quality illumination that closely resembles the ground truth." + ], + "image_footnote": [], + "bbox": [ + 357, + 87, + 607, + 186 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/44147fbc4e337644558b9d78179e4ec47fea4a755b1cb9d1d8d8574f736de55d.jpg", + "image_caption": [ + "(c) True illumination" + ], + "image_footnote": [], + "bbox": [ + 611, + 88, + 857, + 186 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "in equirectangular coordinates $(50 \\times 100$ in our experiments), with an exponential nonlinearity:", + "bbox": [ + 75, + 272, + 468, + 303 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nL \\left(\\hat {\\omega} _ {i}\\right) = \\exp \\left(\\sum_ {k = 0} ^ {K - 1} a ^ {k} L _ {k} \\left(\\hat {\\omega} _ {i}\\right)\\right), \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 158, + 313, + 468, + 356 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\phi^{(\\ell)}\\triangleq \\{L_k\\}_{k = 0}^{K - 1}$ comprise a coarse ( $k = 0$ ) to fine ( $k = K - 1$ ) representation of the illumination, and we set $a = 2$ . For each $k$ , $L_{k}(\\hat{\\omega}_{i})$ is computed by bilinearly interpolating into a grid whose size exponentially increases with $k$ . The exponential nonlinearity is helpful for obtaining high dynamic range values in the environment map, as noted, e.g. in [3].", + "bbox": [ + 76, + 368, + 468, + 477 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Unlike the occluder masks, we find that representing the environment map as a pyramid results in better performance than what is achieved with spherical harmonics. See appendix for additional details.", + "bbox": [ + 75, + 477, + 468, + 537 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Materials", + "text_level": 1, + "bbox": [ + 76, + 546, + 187, + 561 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We represent the scene's spatially-varying BRDF as the sum of a diffuse and a specular component. We use a Lambertian model for the diffuse term and a GGX microfacet model [43] for the specular term. The spatially-varying BRDF parameters are represented by a coordinate-based multi-layer perceptron (MLP) with a positional encoding [38] function $\\gamma$ and estimizable parameters $\\phi^{(m)}$ :", + "bbox": [ + 75, + 569, + 468, + 676 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {\\alpha} (\\mathbf {x}) = \\operatorname {M L P} \\left(\\gamma (\\mathbf {x}); \\phi^ {(m)}\\right). \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 173, + 685, + 468, + 712 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Here, $\\gamma (\\mathbf{x})$ is the positionally-encoded position on the object's surface, and $\\alpha (\\mathbf{x})\\in \\mathbb{R}^5$ are BRDF coefficients, consisting of an RGB diffuse albedo, a spatially-varying scalar microfacet roughness, and a spatially-varying scalar specular reflectance at normal incidence (equivalent to a reparameterization of the material's index of refraction). See the supplement for an exact specification of our BRDF model.", + "bbox": [ + 75, + 715, + 468, + 824 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.4. Rendering", + "text_level": 1, + "bbox": [ + 76, + 832, + 194, + 848 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We render the occluders, materials, and illumination, i.e. the rendering operator $\\mathcal{R}_t$ from Problem 3, by approximating the integral in Equation 2 using Monte Carlo techniques.", + "bbox": [ + 75, + 854, + 468, + 901 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/82b89b13cf92cd41db5da1bf57de2b83ac32155d5f733d975a452065600067c9.jpg", + "image_caption": [ + "(a) Sample input image" + ], + "image_footnote": [], + "bbox": [ + 504, + 268, + 624, + 348 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/f9fd746aa9328e892228beb2630c16b8593a0ca51e8c74a4900d9a4d58336467.jpg", + "image_caption": [ + "(b) Recovered albedo" + ], + "image_footnote": [], + "bbox": [ + 638, + 268, + 756, + 349 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/efdaf10d6b5ca1f563d6c15c6b58408a9ec3efab29a9025dfe1057727bd12dc8.jpg", + "image_caption": [ + "(c) True albedo" + ], + "image_footnote": [], + "bbox": [ + 772, + 268, + 890, + 349 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/404210c2a17e307793a4aaa1e101134d675cb92cca16ff3d35274bd0e985283b.jpg", + "image_caption": [ + "(d) Recovered illumination", + "Figure 7. We are able to recover illumination even when geometry is unknown by first optimizing a volumetric representation of geometry using a NeRF-based method. Despite this inaccurate proxy geometry, our method still recovers plausible (albeit blurry) illumination (d) and albedo (b)." + ], + "image_footnote": [], + "bbox": [ + 501, + 367, + 697, + 444 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/b2d26c2b212e96a3245e2858551847fb6472036fc0dd078701b0a34e89a5c8f7.jpg", + "image_caption": [ + "(e) True illumination" + ], + "image_footnote": [], + "bbox": [ + 700, + 367, + 895, + 444 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We use a standard multiple importance sampler [40] based on the illumination and material model. We model shadows cast by the occluders and self-occlusions by the object itself, but for efficiency and due to our focus on diffuse objects, we neglect lower-order effects and global illumination. See the supplemental material for a full description of our renderer.", + "bbox": [ + 498, + 558, + 892, + 650 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.5. Optimization", + "text_level": 1, + "bbox": [ + 500, + 657, + 638, + 675 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We optimize the objective in Equation 3 using the $L_{2}$ error between rendered values and ground truth ones. We use Adam [16] to optimize all three components of our model, with a learning rate of $3 \\cdot 10^{-3}$ for the environment map and materials, and a learning rate of 1 for the occluders.", + "bbox": [ + 498, + 681, + 890, + 756 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In each iteration, we compute the $L_{2}$ loss using a batch size of $2^{16}$ pixels. In order to avoid bias in the gradient updates to our model's parameters, we use the same approach as [10]: we render every pixel value twice using independent samples and multiply the per-channel deviations of both from the ground truth to get the expected loss value:", + "bbox": [ + 496, + 757, + 892, + 849 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\sum_ {\\mathbf {u}} \\left(\\tilde {I} _ {t} ^ {(1)} (\\mathbf {u}) - I _ {t} (\\mathbf {u})\\right) \\cdot \\left(\\tilde {I} _ {t} ^ {(2)} (\\mathbf {u}) - I _ {t} (\\mathbf {u})\\right), \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 517, + 869, + 890, + 904 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "82", + "bbox": [ + 488, + 945, + 506, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/8969b121a22bd2300a109fbea4bd520f17f6be2aa2c08364b4bfcc6f495f85d4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 127, + 92, + 222, + 220 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/1b2ce43d5f006ad02de392dd16365648815d779d1d83390ccd913aa8a7e30f17.jpg", + "image_caption": [ + "PSNR = 19.3 dB" + ], + "image_footnote": [], + "bbox": [ + 333, + 93, + 426, + 219 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/52fe5dc510adbffb435d476eab6520a392cff5adf340609fb56ec57c39a77e6c.jpg", + "image_caption": [ + "PSNR = 14.4 dB" + ], + "image_footnote": [], + "bbox": [ + 540, + 94, + 633, + 219 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d7a52d26f6a3899dee3a2a66af0108304fbc5da7c0f635b203d4fb97fc50f56e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 746, + 93, + 839, + 219 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f9e5985df6daf92fc8905fdf3fa38da529fd1ee6402097a312d405851642c25c.jpg", + "image_caption": [ + "(a) Sample input image", + "RMSE $= 0.035$", + "(e) Recovered illumination", + "Figure 8. Results on the captured salt004 scene from Stanford-ORB [18]. Note that the \"warped true illumination\" environment map (g) was captured by a light probe that was not co-located with the object, any may therefore be significantly warped. The recovered RMSE reported is computed with respect to the average of all environment maps provided with the dataset rotated the same coordinate frame." + ], + "image_footnote": [], + "bbox": [ + 80, + 266, + 344, + 371 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ed7c83d0bb66b90b9e6f018f25bea90ce53965ff93f2932e63c4d3a1266e8e38.jpg", + "image_caption": [ + "(b) Recovered albedo", + "(c) Recovered albedo (no occluder)", + "RMSE = 0.087", + "(f) Recovered illumination (no occluder)" + ], + "image_footnote": [], + "bbox": [ + 348, + 267, + 616, + 371 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/55c65d68bc598be062d0925de02489462bc13d9af434d32d5a6cb2b0786ce13f.jpg", + "image_caption": [ + "(d) True albedo", + "(g) Warped true illumination" + ], + "image_footnote": [], + "bbox": [ + 620, + 267, + 887, + 371 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $\\tilde{I}_t^{(1)}(\\mathbf{u})$ and $\\tilde{I}_t^{(2)}(\\mathbf{u})$ are the two independently-rendered pixel values.", + "bbox": [ + 75, + 469, + 468, + 502 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 76, + 518, + 209, + 537 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We implement our entire rendering and optimization pipeline in JAX [8] and run each of our experiments on 8 NVIDIA V100 GPUs. We evaluate our method on two sources of data: synthetic objects, and captured data from the Stanford-ORB dataset [18]. Our synthetic results were generated using 256 input images of size $400 \\times 400$ , although our method works similarly well with as few as 32 captures (see supplement for the effect of the number of images). Unless noted otherwise, the occluders for all experiments are spherical caps with a size of $0.1 \\cdot 4\\pi$ steradians, centered at the camera.", + "bbox": [ + 75, + 546, + 468, + 712 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Estimating illumination, materials, and occluders. Our method can effectively recover environment illumination, spatially-varying material parameters, and the shapes of unseen lighting occluders. We first investigate our algorithm's performance using rendered images of a variety of diffuse objects (roughness 0.6, see BRDF model in appendix) with known geometry. The illumination and albedos recovered by our model are visualized with their true values in Figures 1 and 4. Figure 6 shows our reconstruction for purely Lambertian materials, compared with the reconstruction from the same number of images without unobserved occluders (but with self-occlusions), similar to the scenario", + "bbox": [ + 75, + 719, + 470, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "investigated by Swedish et al. [37]. For Lambertian objects, reconstruction is severely ill-conditioned even when self-occlusions are present, which results in significantly blurrier reconstructions. Figure 5 validates that our recovered occluder shapes are accurate, especially for occluders that block more of the illumination energy.", + "bbox": [ + 496, + 470, + 892, + 564 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Captured data. We apply our method to scenes from the Stanford-ORB dataset [18]. Each object in the dataset is placed on a platform and captured from multiple directions. As the camera rig and photographer move around the scene, the photographer is hidden under white cloth below the camera. The images are provided along with a scanned mesh which we use for geometry.", + "bbox": [ + 496, + 571, + 890, + 676 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Despite not being designed for our task, Figure 8 shows our recovered environment map obtained using the 66 training views in the salt004 scene, which features a rough cylindrical object. Note that the dataset only provides an estimated illumination obtained by a light probe placed at a large distance from the object, meaning there is a significant unknown non-rigid warp between the true incident light and the image in Figure 8g labeled \"warped true illumination\". See supplement for additional results.", + "bbox": [ + 496, + 680, + 892, + 816 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Without known geometry. Figure 7 presents preliminary evidence that our method can be used to recover lighting and materials even when the object geometry is not known. In this experiment, we first recover an estimate of object geometry using a Neural Radiance Field (NeRF) [25]-based", + "bbox": [ + 496, + 824, + 893, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "83", + "bbox": [ + 488, + 945, + 506, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e4323f1caa819e4b8bf8fe94aacff0288c6e9c013c450b6c677c615ce52a2961.jpg", + "image_caption": [ + "Figure 9. The signal provided by unintended shadows cast by unseen occluders improves the quality of recovered environment illumination. Here, we plot the relative MSE for the recovered environment maps under two scenarios: (blue) using images rendered with unobserved occluders and jointly estimating materials, illumination and occluder shape and (orange) using images rendered without occluders and only estimating materials and illumination (this is the problem setting considered by Swedish et al. [37]). The cue of unintended shadows consistently improves the quality of estimated illumination across varying (a) illumination frequency content, (b) object material roughness, and (c) mask sizes. Additionally, we show that larger occluders that block more light improve the reconstructed illumination quality." + ], + "image_footnote": [], + "bbox": [ + 83, + 90, + 334, + 195 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/021b1307690b9c70ac09ecade42aa419ad190b22284784c0d8a81d651e62e339.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 357, + 90, + 609, + 195 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/95bf54043295c17c2abccbb0c9ef7880ca1adbfe368d4f47b0cb2a86239c10a5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 632, + 90, + 880, + 195 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/42d997c8f4d3eafc5395befc6c63eb65b6cfb185dee5e3c69b07552940e61518.jpg", + "image_caption": [ + "(a) Illumination (pyramid)", + "Figure 10. Results from the experiment in Figure 4, where the pyramid environment map has been replaced with a direct optimization of its (pre-exponentiated) values. Directly optimizing the mask values instead of using spherical harmonics also results in worse reconstructions (see Table 1)." + ], + "image_footnote": [], + "bbox": [ + 81, + 309, + 269, + 459 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/4a4d7c374adae22a074b2b3851f6cd29d8d19f06d2d9c83b39cfe9bcec541ba6.jpg", + "image_caption": [ + "(b) Illumination (direct)" + ], + "image_footnote": [], + "bbox": [ + 272, + 309, + 460, + 459 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "model augmented using the orientation loss from Verbin et al. [41], and then use this (potentially-imprecise) proxy geometry in our inverse rendering optimization pipeline. Please refer to the supplemental materials for a detailed description of our NeRF-based model.", + "bbox": [ + 75, + 580, + 468, + 656 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Object self-occlusion does not contain enough signal. Figure 9 quantitatively demonstrates that only using the signal provided by object self-occlusions (\"without occluders\" plot in orange) cannot recover illumination as accurately as our method, which leverages the cue of unintended shadows (\"with occluders\" plot in blue). In particular, exploiting unintended shadows is increasingly important as the illumination contains higher frequencies (Figure 9a) and as the object material becomes increasingly diffuse (Figure 9b).", + "bbox": [ + 75, + 664, + 468, + 801 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Larger occluders improve illumination recovery. Figure 9c shows that increasing the size of the unseen occluders improves the accuracy of the recovered environment maps. Increasing the occluder size effectively improves the problem's conditioning by creating a larger variation in the incident lighting across points on the object.", + "bbox": [ + 75, + 809, + 470, + 902 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/11fbac838b09598e683a6e5fcd4e413fa55a8cc3e366ec34743d0ac411a97bce.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
chairmannequinpotatoestoad
Direct spherical harmonics0.8310.6540.9970.943
Direct environment map0.0420.0710.0150.073
Ours0.0380.0350.0110.046
", + "bbox": [ + 501, + 309, + 888, + 364 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 1. Errors (RMSE) on all four of our scenes for our full model compared with our model without the spherical harmonic occluder parameterization, and without the pyramid environment parameterization. While the pyramid representation of the environment map improves smoothness (see Figure 10), we find that representing the masks using spherical harmonics is critical for our method.", + "bbox": [ + 498, + 369, + 892, + 455 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Parameterization ablation study. Table 1 and Figure 10 show that representing the environment map as an image pyramid and the image masks with spherical harmonics, both perform better than simply optimizing their respective element values directly. See supplement for more details.", + "bbox": [ + 498, + 479, + 890, + 555 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Discussion", + "text_level": 1, + "bbox": [ + 500, + 566, + 612, + 583 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Even though unintended shadows are common in most real-world capture scenarios, they are often treated as outlier data. In this paper, we showed how to explicitly leverage these shadows as a signal to improve the quality of recovered lighting and materials in particularly challenging scenarios, such as objects made of diffuse materials. Furthermore, we have shown that our algorithm can be used with approximate geometry reconstructed by view synthesis techniques, and that it can produce imperfect, yet promising results on real captures, despite not being designed for this type of data which does not satisfy many of our model's assumptions. We believe that this work makes important first steps towards a general inverse rendering algorithm that can recover geometry, materials, and illuminations from images captured under realistic conditions.", + "bbox": [ + 496, + 593, + 892, + 819 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements. Our sincere thanks to Delio Vicini, Rick Szeliski, Aleksander Holski, and Guy Satat for helpful discussions and suggestions, and to David Salesin, Janne Kontkanen, and Lior Yariv for their help in improving the manuscript before submission.", + "bbox": [ + 496, + 825, + 893, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "84", + "bbox": [ + 488, + 945, + 506, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Sai Bangaru, Michael Gharbi, Tzu-Mao Li, Fujun Luan, Kalyan Sunkavalli, Milos Hasan, Sai Bi, Zexiang Xu, Gilbert Bernstein, and Fredo Durand. Differentiable rendering of neural SDFs through reparameterization. SIGGRAPH Asia, 2022. 2", + "[2] Manel Baradad, Vickie Ye, Adam B. Yedidia, Frédo Durand, William T. Freeman, Gregory W. Wornell, and Antonio Torralba. Inferring light fields from shadows. CVPR, 2018. 3, 4", + "[3] Jonathan T. Barron and Jitendra Malik. Shape, illumination, and reflectance from shading. TPAMI, 2015. 2, 6", + "[4] Sai Bi, Zexiang Xu, Pratul P. Srinivasan, Ben Mildenhall, Kalyan Sunkavalli, Milos Hasan, Yannick Hold-Geoffroy, David Kriegman, and Ravi Ramamoorthi. Neural reflectance fields for appearance acquisition. arXiv, 2020. 2", + "[5] Sai Bi, Zexiang Xu, Kalyan Sunkavalli, Milovs Havsan, Yannick Hold-Geoffroy, David J. Kriegman, and Ravi Ramamoorthi. Deep Reflectance Volumes: Relightable Reconstructions from Multi-View Photometric Images. ECCV, 2020. 2", + "[6] Mark Boss, Raphael Braun, Varun Jampani, Jonathan T. Barron, Ce Liu, and Hendrik P. A. Lensch. NeRD: Neural reflectance decomposition from image collections. ICCV, 2021. 2", + "[7] Katherine L. Bouman, Vickie Ye, Adam B. Yedidia, Frédo Durand, Gregory W. Wornell, Antonio Torralba, and William T. Freeman. Turning corners into cameras: Principles and methods. ICCV, 2017. 3", + "[8] James Bradbury, Roy Frostig, Peter Hawkins, Matthew James Johnson, Chris Leary, Dougal Maclaurin, George Necula, Adam Paszke, Jake VanderPlas, Skye Wanderman-Milne, and Qiao Zhang. JAX: composable transformations of Python+NumPy programs, 2018. http://github.com/google/jax.7", + "[9] Paul Debevec. Rendering synthetic objects into real scenes: Bridging traditional and image-based graphics with global illumination and high dynamic range photography. SIGGRAPH, 1998. 2", + "[10] Xi Deng, Fujun Luan, Bruce Walter, Kavita Bala, and Steve Marschner. Reconstructing Translucent Objects using Differentiable Rendering. SIGGRAPH, 2022. 6", + "[11] James R Driscoll and Dennis M Healy. Computing Fourier Transforms and Convolutions on the 2-Sphere. Advances in applied mathematics, 1994. 5", + "[12] Marc-Andre Gardner, Kalyan Sunkavalli, Ersin Yumer, Xiaohui Shen, Emiliano Gambaretto, Christian Gagné, and Jean-François Lalonde. Learning to predict indoor illumination from a single image. ACM Trans. Graph., 2017. 2", + "[13] Yannick Hold-Geoffroy, Kalyan Sunkavalli, Sunil Hadap, Emiliano Gambaretto, and Jean-François Lalonde. Deep outdoor illumination estimation. CVPR, 2017. 2", + "[14] Wenzel Jakob, Sébastien Speierer, Nicolas Roussel, and Delio Vicini. Dr.Jit: A Just-In-Time Compiler for Differentiable Rendering. SIGGRAPH, 2022. 2", + "[15] Brian Karis and Epic Games. Real shading in unreal engine 4. Proc. Physically Based Shading Theory Practice, 2013. 1" + ], + "bbox": [ + 78, + 114, + 470, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[16] P. Diederik Kingma and Lei Jimmy Ba. Adam: A Method for Stochastic Optimization. *ICLR*, 2015. 6, 7", + "[17] Zhengfei Kuang, Kyle Olszewski, Menglei Chai, Zeng Huang, Panos Achlioptas, and Sergey Tulyakov. Neroic: Neural rendering of objects from online image collections. ACM Trans. Graph., 2022. 2", + "[18] Zhengfei Kuang, Yunzhi Zhang, Hong-Xing Yu, Samir Agarwala, Shangzhe Wu, and Jiajun Wu. Stanford-ORB: A realworld 3d object inverse rendering benchmark. NeurIPS, 2023. 7, 1, 2, 6", + "[19] Jean-François Lalonde, Alexei A Efros, and Srinivasa G Narasimhan. Estimating the natural illumination conditions from a single outdoor image. IJCV, 2012. 2", + "[20] Chloe LeGendre, Wan-Chun Ma, Graham Fyffe, John Flynn, Laurent Charbonnel, Jay Busch, and Paul Debevec. Deeplight: Learning illumination for unconstrained mobile mixed reality. CVPR, 2019. 2", + "[21] Chloe LeGendre, Wan-Chun Ma, Rohit Pandey, Sean Fanello, Christoph Rhemann, Jason Dourgarian, Jay Busch, and Paul Debevec. Learning illumination from diverse portraits. SIGGRAPH Asia 2020 Technical Communications, 2020. 2", + "[22] Tzu-Mao Li, Miika Aittala, Frédo Durand, and Jaakko Lehtinen. Differentiable monte carlo ray tracing through edge sampling. SIGGRAPH Asia, 2018. 2", + "[23] Zhengqin Li, Zexiang Xu, Ravi Ramamoorthi, Kalyan Sunkavalli, and Manmohan Chandraker. Learning to reconstruct shape and spatially-varying reflectance from a single image. SIGGRAPH Asia, 2018. 2", + "[24] Zhengqin Li, Mohammad Shafiei, Ravi Ramamoorthi, Kalyan Sunkavalli, and Manmohan Chandraker. Inverse rendering for complex indoor scenes: Shape, spatially-varying lighting and svbrdf from a single image. CVPR, 2020. 2", + "[25] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. NeRF: Representing Scenes as Neural Radiance Fields for View Synthesis. ECCV, 2020. 7, 8", + "[26] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. SIGGRAPH, 2022. 8", + "[27] Jacob Munkberg, Jon Hasselgren, Tianchang Shen, Jun Gao, Wenzheng Chen, Alex Evans, Thomas Mueller, and Sanja Fidler. Extracting Triangular 3D Models, Materials, and Lighting From Images. CVPR, 2022. 2", + "[28] Merlin Nimier-David, Delio Vicini, Tizian Zeltner, and Wenzel Jakob. Mitsuba 2: A retargetable forward and inverse renderer. SIGGRAPH Asia, 2019. 2", + "[29] Merlin Nimier-David, Sébastien Speierer, Benoit Ruiz, and Wenzel Jakob. Radiative Backpropagation: An Adjoint Method for Lightning-Fast Differentiable Rendering. SIGGRAPH, 2020. 2", + "[30] Ko Nishino and Shree K Nayar. Eyes for relighting. ACM Trans. Graph., 2004. 2", + "[31] Jeong Joon Park, Aleksander Holynski, and Steven M Seitz. Seeing the world in a bag of chips. CVPR, 2020. 1, 3", + "[32] Matt Pharr, Wenzel Jakob, and Greg Humphreys. Physically Based Rendering: From Theory to Implementation (3rd" + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "85", + "bbox": [ + 490, + 946, + 506, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "ed.). Morgan Kaufmann Publishers Inc., San Francisco, CA, USA, 3rd edition, 2016. 1", + "[33] Ravi Ramamoorthi and Pat Hanrahan. A signal-processing framework for inverse rendering. SIGGRAPH, 2001. 2, 5", + "[34] Imari Sato, Yoichi Sato, and Katsushi Ikeuchi. Illumination from shadows. TPAMI, 2003. 3", + "[35] Shuran Song and Thomas Funkhouser. Neural illumination: Lighting prediction for indoor environments. CVPR, 2019. 2", + "[36] Pratul P. Srinivasan, Boyang Deng, Xiuming Zhang, Matthew Tancik, Ben Mildenhall, and Jonathan T. Barron. NeRV: Neural reflectance and visibility fields for relighting and view synthesis. CVPR, 2021. 2", + "[37] Tristan Swedish, Connor Henley, and Ramesh Raskar. Objects as cameras: Estimating high-frequency illumination from shadows. ICCV, 2021. 1, 3, 7, 8", + "[38] Matthew Tancik, Pratul Srinivasan, Ben Mildenhall, Sara Fridovich-Keil, Nithin Raghavan, Utkarsh Singhal, Ravi Ramamoorthi, Jonathan Barron, and Ren Ng. Fourier features let networks learn high frequency functions in low dimensional domains. NeurIPS, 2020. 6", + "[39] Antonio Torralba and William T. Freeman. Accidental pin-hole and pinspeck cameras: Revealing the scene outside the picture. CVPR, 2012. 3", + "[40] Eric Veach and Leonidas J Guibas. Optimally Combining Sampling Techniques for Monte Carlo Rendering. SIGGRAPH, 1995. 6, 4", + "[41] Dor Verbin, Peter Hedman, Ben Mildenhall, Todd Zickler, Jonathan T. Barron, and Pratul P. Srinivasan. Ref-NeRF: Structured View-Dependent Appearance for Neural Radiance Fields. CVPR, 2022. 8", + "[42] Delio Vicini, Sébastien Speierer, and Wenzel Jakob. Differentiable Signed Distance Function Rendering. SIGGRAPH, 2022. 2", + "[43] Bruce Walter, Stephen R Marschner, Hongsong Li, and Kenneth E Torrance. Microfacet models for refraction through rough surfaces. Eurographics, 2007. 6, 1", + "[44] Adam B Yedidia, Manel Baradad, Christos Thrampoulidis, William T Freeman, and Gregory W Wornell. Using unknown occluders to recover hidden scenes. CVPR, 2019. 3", + "[45] Tizian Zeltner, Sébastien Speierer, Iliyan Georgiev, and Wenzel Jakob. Monte Carlo Estimators for Differential Light Transport. SIGGRAPH, 2021. 2, 4", + "[46] Kai Zhang, Fujun Luan, Qianqian Wang, Kavita Bala, and Noah Snively. PhySG: Inverse rendering with spherical gaussians for physics-based material editing and relighting. CVPR, 2021. 2", + "[47] Xiuming Zhang, Pratul P. Srinivasan, Boyang Deng, Paul Debevec, William T. Freeman, and Jonathan T. Barron. NeR-Factor: Neural factorization of shape and reflectance under an unknown illumination. SIGGRAPH Asia, 2021. 2" + ], + "bbox": [ + 78, + 90, + 467, + 801 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "86", + "bbox": [ + 490, + 946, + 506, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2024/Eclipse_ Disambiguating Illumination and Materials using Unintended Shadows/a9a99e7b-5bee-4200-8036-13e74c55e36e_model.json b/2024/Eclipse_ Disambiguating Illumination and Materials using Unintended Shadows/a9a99e7b-5bee-4200-8036-13e74c55e36e_model.json new file mode 100644 index 0000000000000000000000000000000000000000..01d5eb1f9d86eb9e22c01aaeeba8a32009895617 --- /dev/null +++ b/2024/Eclipse_ Disambiguating Illumination and Materials using Unintended Shadows/a9a99e7b-5bee-4200-8036-13e74c55e36e_model.json @@ -0,0 +1,2838 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.084, + 0.131, + 0.888, + 0.154 + ], + "angle": 0, + "content": "Eclipse: Disambiguating Illumination and Materials using Unintended Shadows" + }, + { + "type": "text", + "bbox": [ + 0.287, + 0.181, + 0.385, + 0.197 + ], + "angle": 0, + "content": "Dor Verbin" + }, + { + "type": "text", + "bbox": [ + 0.407, + 0.181, + 0.539, + 0.197 + ], + "angle": 0, + "content": "Ben Mildenhall1" + }, + { + "type": "text", + "bbox": [ + 0.562, + 0.181, + 0.682, + 0.197 + ], + "angle": 0, + "content": "Peter Hedman" + }, + { + "type": "text", + "bbox": [ + 0.241, + 0.199, + 0.4, + 0.215 + ], + "angle": 0, + "content": "Jonathan T. Barron1" + }, + { + "type": "text", + "bbox": [ + 0.422, + 0.199, + 0.546, + 0.215 + ], + "angle": 0, + "content": "Todd Zickler\\(^{1,2}\\)" + }, + { + "type": "text", + "bbox": [ + 0.565, + 0.199, + 0.726, + 0.215 + ], + "angle": 0, + "content": "Pratul P. Srinivasan1" + }, + { + "type": "text", + "bbox": [ + 0.321, + 0.216, + 0.464, + 0.235 + ], + "angle": 0, + "content": "1Google Research" + }, + { + "type": "text", + "bbox": [ + 0.486, + 0.216, + 0.646, + 0.235 + ], + "angle": 0, + "content": "\\(^{2}\\)Harvard University" + }, + { + "type": "image", + "bbox": [ + 0.08, + 0.316, + 0.276, + 0.441 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.131, + 0.445, + 0.235, + 0.459 + ], + "angle": 0, + "content": "(a) Problem setup" + }, + { + "type": "image_caption", + "bbox": [ + 0.306, + 0.262, + 0.35, + 0.273 + ], + "angle": 0, + "content": "Albedo" + }, + { + "type": "image", + "bbox": [ + 0.315, + 0.306, + 0.344, + 0.334 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.391, + 0.262, + 0.546, + 0.273 + ], + "angle": 0, + "content": "Illumination and occluders" + }, + { + "type": "image", + "bbox": [ + 0.371, + 0.278, + 0.585, + 0.368 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.61, + 0.262, + 0.653, + 0.273 + ], + "angle": 0, + "content": "Albedo" + }, + { + "type": "image", + "bbox": [ + 0.597, + 0.28, + 0.661, + 0.355 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.694, + 0.262, + 0.848, + 0.273 + ], + "angle": 0, + "content": "Illumination and occluders" + }, + { + "type": "image", + "bbox": [ + 0.67, + 0.278, + 0.885, + 0.368 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.296, + 0.38, + 0.361, + 0.453 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.37, + 0.378, + 0.585, + 0.468 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.598, + 0.378, + 0.663, + 0.453 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.671, + 0.377, + 0.885, + 0.468 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.075, + 0.476, + 0.893, + 0.545 + ], + "angle": 0, + "content": "Figure 1. (a) We exploit unintended shadows cast by camera operators (or other unobserved moving occluders) to recover high-fidelity environment lighting and object materials from a set of images. Without modeling such unobserved occluders, prior methods can only: (b) recover convolved lighting without explicit material decomposition [31]; or (c) exploit lighting occlusions that occur internally among the observed objects [37]. (d) We show that additionally modeling and recovering external, unobserved occluders enables lighting and material reconstructions that are closer to ground truth (e)." + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.559, + 0.313, + 0.573 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.582, + 0.471, + 0.884 + ], + "angle": 0, + "content": "Decomposing an object's appearance into representations of its materials and the surrounding illumination is difficult, even when the object's 3D shape is known beforehand. This problem is especially challenging for diffuse objects: it is ill-conditioned because diffuse materials severely blur incoming light, and it is ill-posed because diffuse materials under high-frequency lighting can be indistinguishable from shiny materials under low-frequency lighting. We show that it is possible to recover precise materials and illumination—even from diffuse objects—by exploiting unintended shadows, like the ones cast onto an object by the photographer who moves around it. These shadows are a nuisance in most previous inverse rendering pipelines, but here we exploit them as signals that improve conditioning and help resolve material-lighting ambiguities. We present a method based on differentiable Monte Carlo ray tracing that uses images of an object to jointly recover its spatially-varying materials, the surrounding illumination environment, and the shapes of the unseen light occluders who inadvertently cast shadows upon it." + }, + { + "type": "title", + "bbox": [ + 0.502, + 0.559, + 0.631, + 0.573 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.584, + 0.892, + 0.689 + ], + "angle": 0, + "content": "In this work, we show that the long-standing inverse rendering problem of recovering an object's material properties and the surrounding illumination environment from a set of images can greatly benefit from considering the effects of unobserved moving light occluders, such as the camera operator, which partially block incoming light and inadvertently cast shadows onto the object being imaged." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.69, + 0.893, + 0.901 + ], + "angle": 0, + "content": "Joint recovery of materials and lighting is a challenging task because the BRDF (bidirectional reflectance distribution function, which represents how a material at any point on an object's surface maps from incoming to outgoing light) acts as a directional filter on incoming light. BRDFs for specular (shiny) materials act as all-pass directional filters, while BRDFs for diffuse materials act as low-pass directional filters. As a result, inverse rendering is fundamentally ambiguous since a shiny object illuminated by blurry lighting can be indistinguishable from a diffuse object illuminated by sharp lighting (Figure 2). Furthermore, since diffuse materials act as low-pass filters and strongly blur incoming light, even the simpler problem of recovering lighting from images of a known diffuse material can be" + }, + { + "type": "page_number", + "bbox": [ + 0.491, + 0.946, + 0.507, + 0.956 + ], + "angle": 0, + "content": "77" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.468, + 0.122 + ], + "angle": 0, + "content": "severely ill-conditioned, precluding the recovery of high-frequency illumination [33]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.123, + 0.468, + 0.334 + ], + "angle": 0, + "content": "We observe that both of these issues can be ameliorated by exploiting an effect that naturally exists whenever an object is imaged from a sequence of viewpoints under static environment lighting. Namely, between one image and the next, the positions of the camera and its operator(s) must change, and in doing so they occlude different portions of the surrounding environment, casting distinct sets of shadows onto the object being imaged. This effect is very noticeable under strongly-directional lighting or for shiny objects because the cast shadows are sharp or the specular highlights are misshapen and hidden. But even when the effect is barely perceptible, such as for diffuse objects under well-distributed lighting, there exists a subtle signal that one can take advantage of." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.336, + 0.468, + 0.624 + ], + "angle": 0, + "content": "In this paper, we make first steps toward a practical algorithm that exploits this cue by using gradient descent with differentiable Monte Carlo ray tracing to jointly recover explicit representations of (i) spatially-varying reflectance, (ii) environment illumination, and (iii) the per-image shapes of any unseen, light-blocking occluders. We evaluate our algorithm using a challenging variety of simulated scenes, including scenes with diffuse-only materials and with object geometry that is not known beforehand. We also apply our method to an existing off-the-shelf dataset designed for inverse rendering, where the (unseen) camera rig moving to capture the scene acts as an accidental occluder. Our results demonstrate that external shadowing effects provide a strong and useful cue, even when they are subtle, and even without relying on strong domain-specific priors for the materials, illumination, or external occluder shapes. This suggests that incorporating unintended shadows into inverse rendering pipelines for real, captured data is valuable for improving the quality of material and lighting assets." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.641, + 0.218, + 0.656 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.668, + 0.468, + 0.759 + ], + "angle": 0, + "content": "We build on recent developments in differentiable ray tracing and on a long history of inverse rendering work, including the joint estimation of reflectance and lighting, and the estimation of lighting alone. We are also inspired by a separate line of work on passive non-line-of-sight imaging, which exploits similar light-blocking effects." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.765, + 0.468, + 0.901 + ], + "angle": 0, + "content": "Physics-based differentiable ray tracing. Modern differentiable ray tracers [22, 28] enable the computation of gradients of rendered images with respect to scene parameters (i.e., geometry, materials, and lighting) by differentiating through light transport simulation. Recent works have focused on improving efficiency and performance [14, 29] and on accurately differentiating through visibility discontinuities—such as those caused by shadow and occlusion boundaries—with respect to shape and light-" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.892, + 0.152 + ], + "angle": 0, + "content": "ing [1, 42]. Our implementation of differentiable ray tracing leverages insights from these works as well as from Zeltner et al. [45], who provide design intuition for Monte Carlo differentiable ray tracers." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.161, + 0.892, + 0.372 + ], + "angle": 0, + "content": "Inverse rendering of materials and lighting. Decomposing an object's appearance into representations of material and lighting is a long-standing problem in computer vision and graphics. In their foundational work, Ramamoorthi and Hanrahan [33] developed a signal processing approach by describing the outgoing light at a surface point as the spherical convolution of the BRDF and the incoming lighting. This formulation elucidates why inverse rendering is ill-posed and often ill-conditioned: it is ill-posed because there are multiple illumination-material pairs that convolve into the same image, and it is ill-conditioned because diffuse BRDFs act as a low-pass filters on lighting, which causes the estimates of medium- and high-frequency illumination to be very sensitive to noise." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.376, + 0.892, + 0.558 + ], + "angle": 0, + "content": "Because of this, most approaches to inverse rendering rely on strong priors on materials and lighting. Single-view approaches have used hand-designed priors [3] or priors in the form of neural networks trained with supervision from large datasets of material and lighting labels [23, 24]. More relevant to us are multi-view approaches, which typically either assume known lighting [4, 5, 36] or rely on strong priors, such as assuming a single, highly-specular BRDF for the entire object [46] or lighting from a prior distribution that was pre-trained on a dataset of environment maps [6, 47]. The most closely related work is that for which lighting information is not provided as input [17, 27]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.56, + 0.892, + 0.665 + ], + "angle": 0, + "content": "For the sake of generality, we do not use such strong priors in our experiments because by avoiding them we can more directly measure the benefits that are gained by modeling unintended shadows as an additional cue. Because our model uses a generic gradient-descent framework, stronger application-specific priors can be added to it to improve performance on a specific domain of interest." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.675, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Estimating environment lighting. Another thread of inverse rendering research focuses on the task of recovering high-fidelity lighting environments, typically for augmented reality applications where virtual objects are rendered into photographs with consistent reflections and shadows. Prior work by Debevec [9] used images of a chrome sphere to measure environment lighting directly, and subsequent works have demonstrated that plausible environment lighting can be estimated without chrome spheres, using images of indoor [12, 35] and outdoor scenes [13, 19, 20], or images of a specific class of objects like faces [21, 30]. These subsequent techniques use strong priors in the form of deep neural network weights that are trained with supervision to map from images to lighting, and they do not enforce physical rendering consistency between the recovered" + }, + { + "type": "page_number", + "bbox": [ + 0.491, + 0.946, + 0.507, + 0.956 + ], + "angle": 0, + "content": "78" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.122 + ], + "angle": 0, + "content": "lighting and the observations. In contrast, we avoid such strong priors, and we explicitly enforce consistency." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.123, + 0.468, + 0.228 + ], + "angle": 0, + "content": "More related to our approach is the work of Park et al. [31], which recovers physically-consistent environment maps from RGBD videos of shiny objects. However, their algorithm can only recover the convolution of the environment map with the BRDF of the observed object (e.g., Figure 1(b)), which precludes the recovery of high illumination frequencies unless the object is highly specular." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.23, + 0.468, + 0.426 + ], + "angle": 0, + "content": "Also related is the work of Swedish et al. [37], which places a known, diffuse object on a ground plane (a scenario first studied by Sato et al. [34]) and uses its cast shadows to recover high-quality lighting. Their formulation is linear and it improves lighting estimates by leveraging self-shadowing among diffuse objects that have known geometry and albedo (e.g., Figure 1(c)). We generalize this by replacing the linear formulation with differentiable Monte Carlo ray tracing, which allows exploiting the additional shadowing effects caused by moving external occluders and leads to substantially improved results (e.g., Figure 1(d)). Our formulation also handles objects with more general spatially-varying BRDFs that are not known beforehand." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.433, + 0.468, + 0.689 + ], + "angle": 0, + "content": "Passive non-line-of-sight imaging with occlusions. Passive non-line-of-sight techniques can also be seen as recovering environment illumination: They observe a reflective surface (which is typically diffuse and planar) and recover the appearance of a \"hidden scene\" from these reflections. Similar to us, prior works in this area have observed that the presence of an occluder between the observed surface and the hidden scene/environment aids recovery by introducing sharp angular variations into the rendering integral. This insight was first leveraged in settings with simple occluder-shapes like pinholes and pinspecks [39] or corners joining walls [7]. Subsequent work by Baradad et al. [2] generalizes this by considering light-occlusion effects caused by an arbitrary but known 3D shape (a task that is closely related to Swedish et al. above). Yedidia et al. [44] additionally recover the unobserved occluder's shape assuming it is a planar mask parallel to the observed planar surface." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.691, + 0.468, + 0.812 + ], + "angle": 0, + "content": "We generalize these prior works by using differentiable Monte Carlo rendering to replace their deconvolution-based algorithms, which are tailored to the specific case where the observed reflector is planar and diffuse. This allows using reflective objects that have arbitrary shapes, and arbitrary spatially-varying BRDFs which are not known beforehand. It also allows recovering the shapes of unobserved occluders that are arbitrary and time-varying." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.829, + 0.36, + 0.847 + ], + "angle": 0, + "content": "3. Motivation and Problem Setup" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.468, + 0.901 + ], + "angle": 0, + "content": "A well-known ambiguity in computer vision and graphics occurs when decomposing an image into its component lighting and materials. The top of Figure 2 recreates" + }, + { + "type": "image", + "bbox": [ + 0.501, + 0.088, + 0.892, + 0.31 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.321, + 0.892, + 0.432 + ], + "angle": 0, + "content": "Figure 2. Decomposing appearance into lighting and materials can be inherently ambiguous. Top row: a diffuse sphere lit by high-frequency lighting (inset) is visually indistinguishable from a shiny, mirror-like sphere lit by low-frequency lighting. Bottom row: A second image captured in the presence of a dark external occluder resolves the ambiguity. The occluder's effect is clearly visible on the shiny sphere while on the diffuse sphere it produces shadows that are soft and very subtle." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.461, + 0.892, + 0.642 + ], + "angle": 0, + "content": "a common depiction of this ambiguity, where an image of a known shape (a sphere) is explained equally well by a diffuse material in a complicated lighting environment or a shiny, mirror-like material under low-frequency lighting. Now, imagine we capture a second image from the same viewpoint, after some external object enters the scene. This second object remains beyond the field of view and so is not directly observed, but it acts as an external occluder that prevents some of the environment's light from reaching the sphere. This second image, shown at the bottom of the figure for each case, clearly reveals which of the two materiallighting explanations is the correct one." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.643, + 0.892, + 0.794 + ], + "angle": 0, + "content": "We aim to make use of this unintended shadowing that naturally occurs whenever a camera and its operators move around an object while capturing images from different viewpoints. They affect each image by blocking different portions of the surrounding light, and this provides a helpful signal. We will show that this signal is helpful even when the occluder shapes and locations are quite arbitrary and not known beforehand, and when their shadowing effects are very subtle, like between the top and bottom images on the left of Figure 2." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.796, + 0.892, + 0.901 + ], + "angle": 0, + "content": "In addition to helping resolve material-lighting ambiguities, shadowing from moving external occluders also improves the conditioning of our inverse rendering problem. This is analyzed in Figure 3, which considers the simplified hypothetical 1D case of recovering an environment illumination from images of a diffuse disk-shaped object, with known occluder positions. Convolution with a dif" + }, + { + "type": "page_number", + "bbox": [ + 0.491, + 0.946, + 0.507, + 0.956 + ], + "angle": 0, + "content": "79" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.091, + 0.465, + 0.294 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.305, + 0.47, + 0.445 + ], + "angle": 0, + "content": "Figure 3. Here we plot eigenvalue magnitudes as a function of illumination angular frequency for a one-dimensional circular Lambertian object with known albedo, under the simplifying assumption of known occluder locations and shapes. For a single input image without occlusion, almost all lighting frequencies are strongly attenuated. In the single-occluder case (similar to the scenario considered by Baradad et al. [2]), some low frequencies become recoverable but intermediate and high frequencies do not. Observing additional images enables the recovery of more lighting information, especially at the highest frequencies." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.472, + 0.468, + 0.639 + ], + "angle": 0, + "content": "fuse BRDF acts as a low-pass filter of the incoming light, which means that the linear inverse problem's eigenvalues are vanishingly low for higher frequencies. Moving external occluders introduce a sharp angular variation in the per-image illumination, which better-conditions these high frequency components, and allows their recovery. As shown in Figure 3, in this simplified scenario a single image containing an occluder is better than a single image without any occluders, but using observations of more occluders makes the problem even better-conditioned. We demonstrate this effect in practice in the supplement." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.651, + 0.212, + 0.667 + ], + "angle": 0, + "content": "3.1. Formulation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.675, + 0.468, + 0.812 + ], + "angle": 0, + "content": "Let \\(\\{I_t\\}_{t=1}^T\\) be a collection of \\(T\\) images of a scene with known camera poses; for every \\(t\\), \\(I_t\\) is an RGB image with spatial size \\(H \\times W\\). We assume the observed scene \\(\\mathcal{S} \\subset \\mathbb{R}^3\\) is illuminated by a far-field environment map \\(L(\\hat{\\omega}_i)\\) that does not change over time, so the only temporal changes in the incident light field are caused by unobserved light-blockers (called occluders hereafter) that move around the scene outside the field of view. We can then write the incident light at any observed position \\(\\mathbf{x} \\in \\mathcal{S}\\) as:" + }, + { + "type": "equation", + "bbox": [ + 0.139, + 0.826, + 0.469, + 0.842 + ], + "angle": 0, + "content": "\\[\nL _ {t} (\\mathbf {x}, \\hat {\\omega} _ {i}) = L (\\hat {\\omega} _ {i}) M _ {t} (\\mathbf {x}, \\hat {\\omega} _ {i}) V (\\mathbf {x}, \\hat {\\omega} _ {i}), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.856, + 0.469, + 0.901 + ], + "angle": 0, + "content": "where for any \\(t\\), \\(M_{t}(\\mathbf{x}, \\hat{\\boldsymbol{\\omega}}_{i})\\) is a binary signal with value 0 for directions blocked by the occluder and 1 otherwise, and the binary signal \\(V(\\mathbf{x}, \\hat{\\boldsymbol{\\omega}}_{i})\\) models visibility effects that are" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.892, + 0.121 + ], + "angle": 0, + "content": "internal to the scene, with value 0 for directions from \\( \\mathbf{x} \\) that are blocked by other scene elements, and value 1 otherwise." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.123, + 0.892, + 0.168 + ], + "angle": 0, + "content": "Omitting global illumination effects, a pixel \\(\\mathbf{u}\\) corresponding to a surface point \\(\\mathbf{x}\\) with BRDF \\(f\\) viewed from direction \\(\\hat{\\omega}_o\\) at time \\(t\\) has color:" + }, + { + "type": "equation", + "bbox": [ + 0.508, + 0.18, + 0.892, + 0.212 + ], + "angle": 0, + "content": "\\[\nI _ {t} (\\mathbf {u}) = \\int_ {\\mathbb {S} ^ {2}} L _ {t} (\\mathbf {x}, \\hat {\\omega} _ {i}) f (\\mathbf {x}, \\hat {\\omega} _ {i}, \\hat {\\omega} _ {o}) (\\hat {\\mathbf {n}} (\\mathbf {x}) \\cdot \\hat {\\boldsymbol {\\omega}} _ {i}) _ {+} d \\hat {\\boldsymbol {\\omega}} _ {i}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.225, + 0.892, + 0.255 + ], + "angle": 0, + "content": "where \\(\\hat{\\mathbf{n}} (\\mathbf{x})\\) is the surface normal corresponding to \\(\\mathbf{x}\\), and \\((\\cdot)_{+}\\) clamps negative values to zero." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.256, + 0.892, + 0.348 + ], + "angle": 0, + "content": "Given the set of observed images, we would like to recover the environment map \\( L \\), a spatially-varying BRDF \\( f \\) at every point on the surface of the object \\( S \\), as well as the set of unobserved occluder masks \\( \\{M_t\\}_{t=1}^T \\), one for each observation time \\( t \\). Our goal is therefore to solve the following optimization problem:" + }, + { + "type": "equation", + "bbox": [ + 0.51, + 0.359, + 0.892, + 0.409 + ], + "angle": 0, + "content": "\\[\n\\underset {\\phi^ {(o)}, \\phi^ {(m)}, \\phi^ {(\\ell)}} {\\arg \\min } \\sum_ {t, \\mathbf {u}} \\left\\| I _ {t} (\\mathbf {u}) - \\mathcal {R} _ {t} \\left(\\mathbf {u}; \\phi^ {(o)}, \\phi^ {(m)}, \\phi^ {(\\ell)}\\right) \\right\\| ^ {2}, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.41, + 0.892, + 0.457 + ], + "angle": 0, + "content": "where \\(\\mathcal{R}_t(\\mathbf{u};\\cdot)\\) renders the pixel location \\(\\mathbf{u}\\) at time \\(t\\) using the occluder parameters \\(\\phi^{(o)}\\), material parameters \\(\\phi^{(m)}\\), and illumination parameters \\(\\phi^{(\\ell)}\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.458, + 0.892, + 0.563 + ], + "angle": 0, + "content": "At first glance, it may seem that solving for occluders in addition to materials and illumination should make the resulting inverse problem more difficult. However, as illustrated in Figure 3 and demonstrated by our experimental results, modeling and solving for occluders actually makes the full problem more well-conditioned and thus improves recovery of material and illumination." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.58, + 0.59, + 0.595 + ], + "angle": 0, + "content": "4. Method" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.606, + 0.892, + 0.681 + ], + "angle": 0, + "content": "We first describe our parameterizations of the unseen oc-clusters, illumination, and materials (Sections 4.1-4.3). Section 4.4 then relates these parameters to rendered pixel colors, and Section 4.5 describes how we optimize to solve the inverse problem in Equation 3." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.683, + 0.892, + 0.774 + ], + "angle": 0, + "content": "We emphasize that our method is designed to exploit external shadows across as many scene-types as possible, and so its only priors are those implicit to our parameterizations. In our experiments, we verify that these relatively weak priors are sufficient for recovering high-fidelity occluder shapes, illumination maps, and material maps." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.787, + 0.614, + 0.801 + ], + "angle": 0, + "content": "4.1.Occluders" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.811, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Define a world coordinate system with origin at the scene's center, and let \\(\\{r_t\\}_{t=1}^T\\) be the radial distances from the origin to each (known) camera's center. Then we model each of the \\(T\\) occluders \\(M_t\\) as an independent binary signal defined on the surface of a finite, radius-\\(r_t\\) sphere that is centered at the origin. The value of the binary occluder signal" + }, + { + "type": "page_number", + "bbox": [ + 0.49, + 0.946, + 0.508, + 0.956 + ], + "angle": 0, + "content": "80" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.101, + 0.109, + 0.198, + 0.165 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.219, + 0.111, + 0.307, + 0.164 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.329, + 0.112, + 0.417, + 0.164 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.22, + 0.177, + 0.313, + 0.189 + ], + "angle": 0, + "content": "PSNR = 31.2 dB" + }, + { + "type": "image", + "bbox": [ + 0.102, + 0.201, + 0.204, + 0.271 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.214, + 0.202, + 0.315, + 0.271 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.326, + 0.202, + 0.426, + 0.271 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.221, + 0.278, + 0.312, + 0.29 + ], + "angle": 0, + "content": "PSNR = 26.4 dB" + }, + { + "type": "image", + "bbox": [ + 0.144, + 0.3, + 0.171, + 0.361 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.255, + 0.3, + 0.282, + 0.361 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.366, + 0.301, + 0.393, + 0.361 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.221, + 0.379, + 0.312, + 0.391 + ], + "angle": 0, + "content": "PSNR = 29.6 dB" + }, + { + "type": "image_caption", + "bbox": [ + 0.1, + 0.395, + 0.211, + 0.407 + ], + "angle": 0, + "content": "(a) Sample input image" + }, + { + "type": "image_caption", + "bbox": [ + 0.218, + 0.395, + 0.318, + 0.407 + ], + "angle": 0, + "content": "(b) Recovered albedo" + }, + { + "type": "image_caption", + "bbox": [ + 0.341, + 0.395, + 0.414, + 0.406 + ], + "angle": 0, + "content": "(c) True albedo" + }, + { + "type": "image", + "bbox": [ + 0.437, + 0.089, + 0.65, + 0.173 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.652, + 0.089, + 0.867, + 0.173 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.503, + 0.177, + 0.585, + 0.189 + ], + "angle": 0, + "content": "\\(\\mathrm{RMSE} = 0.073\\)" + }, + { + "type": "image", + "bbox": [ + 0.437, + 0.19, + 0.65, + 0.274 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.652, + 0.19, + 0.866, + 0.274 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.503, + 0.278, + 0.583, + 0.289 + ], + "angle": 0, + "content": "RMSE = 0.081" + }, + { + "type": "image", + "bbox": [ + 0.437, + 0.291, + 0.65, + 0.375 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.652, + 0.291, + 0.866, + 0.375 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.503, + 0.378, + 0.585, + 0.391 + ], + "angle": 0, + "content": "\\(\\mathrm{RMSE} = 0.093\\)" + }, + { + "type": "image_caption", + "bbox": [ + 0.481, + 0.395, + 0.607, + 0.406 + ], + "angle": 0, + "content": "(d) Recovered illumination" + }, + { + "type": "image_caption", + "bbox": [ + 0.71, + 0.395, + 0.808, + 0.406 + ], + "angle": 0, + "content": "(e) True illumination" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.411, + 0.894, + 0.455 + ], + "angle": 0, + "content": "Figure 4. The results of our method on three additional diffuse objects. We report the RMSE of each environment map in linear color space but plot the images after tonemapping for better evaluation of the full dynamic range. The albedo PSNR values are reported on object pixels only." + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.477, + 0.168, + 0.601 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.177, + 0.478, + 0.264, + 0.6 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.271, + 0.485, + 0.468, + 0.626 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.636, + 0.47, + 0.72 + ], + "angle": 0, + "content": "Figure 5. Our occluders extracted from the potatoes scene on the top of Figure 4 resemble the true ones (left). However, as expected, inaccuracies in the recovered masks tend to occur at relatively-dark regions of the environment map: there is a very strong inverse correlation (right) between the mean error in the recovered masks and the amount of illumination energy that they block." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.745, + 0.47, + 0.791 + ], + "angle": 0, + "content": "for a shadow ray emitted from surface point \\(\\mathbf{x} \\in S\\) in direction \\(\\hat{\\omega}_i\\) at time \\(t\\) can be computed by intersecting the shadow ray with the occluder's spherical shell:" + }, + { + "type": "equation", + "bbox": [ + 0.166, + 0.801, + 0.47, + 0.819 + ], + "angle": 0, + "content": "\\[\nM _ {t} (\\mathbf {x}, \\hat {\\omega} _ {i}) = \\tilde {M} _ {t} \\left(S (\\mathbf {x}, \\hat {\\omega} _ {i}, r _ {t})\\right) \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.83, + 0.47, + 0.86 + ], + "angle": 0, + "content": "where \\( S(\\mathbf{x},\\hat{\\omega}_i,r) \\) is the intersection of the ray with a sphere of radius \\( r_t \\), normalized to have unit length:" + }, + { + "type": "equation", + "bbox": [ + 0.086, + 0.87, + 0.47, + 0.906 + ], + "angle": 0, + "content": "\\[\nS (\\mathbf {x}, \\hat {\\boldsymbol {\\omega}} _ {i}, r) = \\frac {\\mathbf {x} + \\left(\\sqrt {(\\mathbf {x} \\cdot \\hat {\\boldsymbol {\\omega}} _ {i}) ^ {2} + r ^ {2} - \\| \\mathbf {x} \\| ^ {2}} - \\mathbf {x} \\cdot \\hat {\\boldsymbol {\\omega}} _ {i}\\right) \\hat {\\boldsymbol {\\omega}} _ {i}}{r}. \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.48, + 0.892, + 0.51 + ], + "angle": 0, + "content": "Note that there is a single intersection, since we assume that the surface points are always inside the occluder spheres." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.511, + 0.893, + 0.602 + ], + "angle": 0, + "content": "Although our formulation models the occluders as binary signals, a binary representation is not well-suited for gradient-based optimization. Instead, we represent the occluders as a continuously-valued function on the sphere in a spherical harmonic basis, mapped using a sigmoid function \\(\\sigma\\) to lie in [0, 1]:" + }, + { + "type": "equation", + "bbox": [ + 0.563, + 0.612, + 0.892, + 0.656 + ], + "angle": 0, + "content": "\\[\n\\tilde {M} _ {t} (\\hat {\\omega}) = \\sigma \\left(\\sum_ {\\ell = 0} ^ {P} \\sum_ {m = - \\ell} ^ {\\ell} a _ {t \\ell m} Y _ {\\ell} ^ {m} (\\hat {\\omega})\\right), \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.669, + 0.892, + 0.732 + ], + "angle": 0, + "content": "where \\(\\phi^{(o)}\\triangleq \\{a_{t\\ell m}\\}\\) are estimizable coefficients that parameterize the occluder at time \\(t\\), and \\(P\\) is the degree required to span the space of spherical images with the same resolution as our environment illumination (see below) [11]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.732, + 0.892, + 0.822 + ], + "angle": 0, + "content": "Note that for radially symmetric BRDFs, the rendering integral in Equation 2 can be written as a spherical convolution [11, 33]. This makes spherical harmonics a natural choice for representing the occluder signals, since spherical convolutions are diagonal in the basis of spherical harmonics. See our supplement for more details." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.833, + 0.738, + 0.847 + ], + "angle": 0, + "content": "4.2. Environment Illumination" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.856, + 0.893, + 0.902 + ], + "angle": 0, + "content": "We assume that the scene's illumination is distant and can be represented as an environment map. The environment map is parameterized as an image pyramid of size \\( H' \\times W' \\)" + }, + { + "type": "page_number", + "bbox": [ + 0.49, + 0.946, + 0.507, + 0.957 + ], + "angle": 0, + "content": "81" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.109, + 0.089, + 0.357, + 0.187 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.125, + 0.19, + 0.341, + 0.202 + ], + "angle": 0, + "content": "(a) Recovered illumination (no occluder)" + }, + { + "type": "image", + "bbox": [ + 0.358, + 0.088, + 0.609, + 0.187 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.411, + 0.19, + 0.555, + 0.201 + ], + "angle": 0, + "content": "(b) Recovered illumination" + }, + { + "type": "image", + "bbox": [ + 0.612, + 0.089, + 0.859, + 0.187 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.678, + 0.19, + 0.791, + 0.201 + ], + "angle": 0, + "content": "(c) True illumination" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.205, + 0.894, + 0.247 + ], + "angle": 0, + "content": "Figure 6. In the case of purely Lambertian material (with the potatoes geometry), (a) only using self occlusions is insufficient for the recovery of high-frequency content in the illumination. (b) Modeling and estimating occluders recovers significantly higher-quality illumination that closely resembles the ground truth." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.273, + 0.47, + 0.304 + ], + "angle": 0, + "content": "in equirectangular coordinates \\((50 \\times 100\\) in our experiments), with an exponential nonlinearity:" + }, + { + "type": "equation", + "bbox": [ + 0.16, + 0.314, + 0.47, + 0.357 + ], + "angle": 0, + "content": "\\[\nL \\left(\\hat {\\omega} _ {i}\\right) = \\exp \\left(\\sum_ {k = 0} ^ {K - 1} a ^ {k} L _ {k} \\left(\\hat {\\omega} _ {i}\\right)\\right), \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.369, + 0.47, + 0.478 + ], + "angle": 0, + "content": "where \\(\\phi^{(\\ell)}\\triangleq \\{L_k\\}_{k = 0}^{K - 1}\\) comprise a coarse (\\(k = 0\\)) to fine (\\(k = K - 1\\)) representation of the illumination, and we set \\(a = 2\\). For each \\(k\\), \\(L_{k}(\\hat{\\omega}_{i})\\) is computed by bilinearly interpolating into a grid whose size exponentially increases with \\(k\\). The exponential nonlinearity is helpful for obtaining high dynamic range values in the environment map, as noted, e.g. in [3]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.478, + 0.47, + 0.539 + ], + "angle": 0, + "content": "Unlike the occluder masks, we find that representing the environment map as a pyramid results in better performance than what is achieved with spherical harmonics. See appendix for additional details." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.547, + 0.188, + 0.563 + ], + "angle": 0, + "content": "4.3. Materials" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.57, + 0.47, + 0.678 + ], + "angle": 0, + "content": "We represent the scene's spatially-varying BRDF as the sum of a diffuse and a specular component. We use a Lambertian model for the diffuse term and a GGX microfacet model [43] for the specular term. The spatially-varying BRDF parameters are represented by a coordinate-based multi-layer perceptron (MLP) with a positional encoding [38] function \\(\\gamma\\) and estimizable parameters \\(\\phi^{(m)}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.174, + 0.686, + 0.47, + 0.713 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {\\alpha} (\\mathbf {x}) = \\operatorname {M L P} \\left(\\gamma (\\mathbf {x}); \\phi^ {(m)}\\right). \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.717, + 0.47, + 0.825 + ], + "angle": 0, + "content": "Here, \\(\\gamma (\\mathbf{x})\\) is the positionally-encoded position on the object's surface, and \\(\\alpha (\\mathbf{x})\\in \\mathbb{R}^5\\) are BRDF coefficients, consisting of an RGB diffuse albedo, a spatially-varying scalar microfacet roughness, and a spatially-varying scalar specular reflectance at normal incidence (equivalent to a reparameterization of the material's index of refraction). See the supplement for an exact specification of our BRDF model." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.833, + 0.196, + 0.849 + ], + "angle": 0, + "content": "4.4. Rendering" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.47, + 0.902 + ], + "angle": 0, + "content": "We render the occluders, materials, and illumination, i.e. the rendering operator \\(\\mathcal{R}_t\\) from Problem 3, by approximating the integral in Equation 2 using Monte Carlo techniques." + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.27, + 0.625, + 0.349 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.504, + 0.351, + 0.63, + 0.363 + ], + "angle": 0, + "content": "(a) Sample input image" + }, + { + "type": "image", + "bbox": [ + 0.64, + 0.27, + 0.758, + 0.35 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.644, + 0.351, + 0.758, + 0.361 + ], + "angle": 0, + "content": "(b) Recovered albedo" + }, + { + "type": "image", + "bbox": [ + 0.774, + 0.27, + 0.892, + 0.35 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.793, + 0.351, + 0.875, + 0.361 + ], + "angle": 0, + "content": "(c) True albedo" + }, + { + "type": "image", + "bbox": [ + 0.502, + 0.368, + 0.698, + 0.445 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.527, + 0.446, + 0.671, + 0.457 + ], + "angle": 0, + "content": "(d) Recovered illumination" + }, + { + "type": "image", + "bbox": [ + 0.701, + 0.368, + 0.897, + 0.445 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.743, + 0.446, + 0.854, + 0.457 + ], + "angle": 0, + "content": "(e) True illumination" + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.463, + 0.893, + 0.532 + ], + "angle": 0, + "content": "Figure 7. We are able to recover illumination even when geometry is unknown by first optimizing a volumetric representation of geometry using a NeRF-based method. Despite this inaccurate proxy geometry, our method still recovers plausible (albeit blurry) illumination (d) and albedo (b)." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.559, + 0.893, + 0.651 + ], + "angle": 0, + "content": "We use a standard multiple importance sampler [40] based on the illumination and material model. We model shadows cast by the occluders and self-occlusions by the object itself, but for efficiency and due to our focus on diffuse objects, we neglect lower-order effects and global illumination. See the supplemental material for a full description of our renderer." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.659, + 0.64, + 0.676 + ], + "angle": 0, + "content": "4.5. Optimization" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.683, + 0.892, + 0.757 + ], + "angle": 0, + "content": "We optimize the objective in Equation 3 using the \\( L_{2} \\) error between rendered values and ground truth ones. We use Adam [16] to optimize all three components of our model, with a learning rate of \\( 3 \\cdot 10^{-3} \\) for the environment map and materials, and a learning rate of 1 for the occluders." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.758, + 0.893, + 0.85 + ], + "angle": 0, + "content": "In each iteration, we compute the \\( L_{2} \\) loss using a batch size of \\( 2^{16} \\) pixels. In order to avoid bias in the gradient updates to our model's parameters, we use the same approach as [10]: we render every pixel value twice using independent samples and multiply the per-channel deviations of both from the ground truth to get the expected loss value:" + }, + { + "type": "equation", + "bbox": [ + 0.518, + 0.871, + 0.892, + 0.905 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\sum_ {\\mathbf {u}} \\left(\\tilde {I} _ {t} ^ {(1)} (\\mathbf {u}) - I _ {t} (\\mathbf {u})\\right) \\cdot \\left(\\tilde {I} _ {t} ^ {(2)} (\\mathbf {u}) - I _ {t} (\\mathbf {u})\\right), \\tag {9}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.49, + 0.946, + 0.508, + 0.957 + ], + "angle": 0, + "content": "82" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.129, + 0.093, + 0.223, + 0.221 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.117, + 0.249, + 0.243, + 0.262 + ], + "angle": 0, + "content": "(a) Sample input image" + }, + { + "type": "image", + "bbox": [ + 0.334, + 0.094, + 0.428, + 0.22 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.34, + 0.234, + 0.432, + 0.246 + ], + "angle": 0, + "content": "PSNR = 19.3 dB" + }, + { + "type": "image_caption", + "bbox": [ + 0.33, + 0.25, + 0.443, + 0.261 + ], + "angle": 0, + "content": "(b) Recovered albedo" + }, + { + "type": "image", + "bbox": [ + 0.541, + 0.095, + 0.634, + 0.22 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.547, + 0.234, + 0.638, + 0.246 + ], + "angle": 0, + "content": "PSNR = 14.4 dB" + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.25, + 0.686, + 0.261 + ], + "angle": 0, + "content": "(c) Recovered albedo (no occluder)" + }, + { + "type": "image", + "bbox": [ + 0.748, + 0.094, + 0.841, + 0.22 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.757, + 0.25, + 0.841, + 0.261 + ], + "angle": 0, + "content": "(d) True albedo" + }, + { + "type": "image", + "bbox": [ + 0.081, + 0.267, + 0.346, + 0.372 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.172, + 0.373, + 0.253, + 0.383 + ], + "angle": 0, + "content": "RMSE \\(= 0.035\\)" + }, + { + "type": "image_caption", + "bbox": [ + 0.141, + 0.388, + 0.284, + 0.399 + ], + "angle": 0, + "content": "(e) Recovered illumination" + }, + { + "type": "image", + "bbox": [ + 0.349, + 0.268, + 0.617, + 0.372 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.442, + 0.373, + 0.525, + 0.383 + ], + "angle": 0, + "content": "RMSE = 0.087" + }, + { + "type": "image_caption", + "bbox": [ + 0.376, + 0.388, + 0.591, + 0.399 + ], + "angle": 0, + "content": "(f) Recovered illumination (no occluder)" + }, + { + "type": "image", + "bbox": [ + 0.622, + 0.268, + 0.888, + 0.372 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.676, + 0.388, + 0.833, + 0.4 + ], + "angle": 0, + "content": "(g) Warped true illumination" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.404, + 0.893, + 0.446 + ], + "angle": 0, + "content": "Figure 8. Results on the captured salt004 scene from Stanford-ORB [18]. Note that the \"warped true illumination\" environment map (g) was captured by a light probe that was not co-located with the object, any may therefore be significantly warped. The recovered RMSE reported is computed with respect to the average of all environment maps provided with the dataset rotated the same coordinate frame." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.47, + 0.469, + 0.503 + ], + "angle": 0, + "content": "where \\(\\tilde{I}_t^{(1)}(\\mathbf{u})\\) and \\(\\tilde{I}_t^{(2)}(\\mathbf{u})\\) are the two independently-rendered pixel values." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.52, + 0.21, + 0.538 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.547, + 0.47, + 0.713 + ], + "angle": 0, + "content": "We implement our entire rendering and optimization pipeline in JAX [8] and run each of our experiments on 8 NVIDIA V100 GPUs. We evaluate our method on two sources of data: synthetic objects, and captured data from the Stanford-ORB dataset [18]. Our synthetic results were generated using 256 input images of size \\(400 \\times 400\\), although our method works similarly well with as few as 32 captures (see supplement for the effect of the number of images). Unless noted otherwise, the occluders for all experiments are spherical caps with a size of \\(0.1 \\cdot 4\\pi\\) steradians, centered at the camera." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.72, + 0.471, + 0.901 + ], + "angle": 0, + "content": "Estimating illumination, materials, and occluders. Our method can effectively recover environment illumination, spatially-varying material parameters, and the shapes of unseen lighting occluders. We first investigate our algorithm's performance using rendered images of a variety of diffuse objects (roughness 0.6, see BRDF model in appendix) with known geometry. The illumination and albedos recovered by our model are visualized with their true values in Figures 1 and 4. Figure 6 shows our reconstruction for purely Lambertian materials, compared with the reconstruction from the same number of images without unobserved occluders (but with self-occlusions), similar to the scenario" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.472, + 0.893, + 0.565 + ], + "angle": 0, + "content": "investigated by Swedish et al. [37]. For Lambertian objects, reconstruction is severely ill-conditioned even when self-occlusions are present, which results in significantly blurrier reconstructions. Figure 5 validates that our recovered occluder shapes are accurate, especially for occluders that block more of the illumination energy." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.572, + 0.892, + 0.678 + ], + "angle": 0, + "content": "Captured data. We apply our method to scenes from the Stanford-ORB dataset [18]. Each object in the dataset is placed on a platform and captured from multiple directions. As the camera rig and photographer move around the scene, the photographer is hidden under white cloth below the camera. The images are provided along with a scanned mesh which we use for geometry." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.681, + 0.893, + 0.817 + ], + "angle": 0, + "content": "Despite not being designed for our task, Figure 8 shows our recovered environment map obtained using the 66 training views in the salt004 scene, which features a rough cylindrical object. Note that the dataset only provides an estimated illumination obtained by a light probe placed at a large distance from the object, meaning there is a significant unknown non-rigid warp between the true incident light and the image in Figure 8g labeled \"warped true illumination\". See supplement for additional results." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.825, + 0.894, + 0.901 + ], + "angle": 0, + "content": "Without known geometry. Figure 7 presents preliminary evidence that our method can be used to recover lighting and materials even when the object geometry is not known. In this experiment, we first recover an estimate of object geometry using a Neural Radiance Field (NeRF) [25]-based" + }, + { + "type": "page_number", + "bbox": [ + 0.49, + 0.946, + 0.507, + 0.956 + ], + "angle": 0, + "content": "83" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.084, + 0.091, + 0.335, + 0.196 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.359, + 0.091, + 0.611, + 0.196 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.633, + 0.091, + 0.882, + 0.196 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.202, + 0.894, + 0.287 + ], + "angle": 0, + "content": "Figure 9. The signal provided by unintended shadows cast by unseen occluders improves the quality of recovered environment illumination. Here, we plot the relative MSE for the recovered environment maps under two scenarios: (blue) using images rendered with unobserved occluders and jointly estimating materials, illumination and occluder shape and (orange) using images rendered without occluders and only estimating materials and illumination (this is the problem setting considered by Swedish et al. [37]). The cue of unintended shadows consistently improves the quality of estimated illumination across varying (a) illumination frequency content, (b) object material roughness, and (c) mask sizes. Additionally, we show that larger occluders that block more light improve the reconstructed illumination quality." + }, + { + "type": "image", + "bbox": [ + 0.083, + 0.31, + 0.271, + 0.46 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.115, + 0.462, + 0.238, + 0.473 + ], + "angle": 0, + "content": "(a) Illumination (pyramid)" + }, + { + "type": "image", + "bbox": [ + 0.274, + 0.31, + 0.462, + 0.46 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.311, + 0.462, + 0.424, + 0.473 + ], + "angle": 0, + "content": "(b) Illumination (direct)" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.477, + 0.471, + 0.547 + ], + "angle": 0, + "content": "Figure 10. Results from the experiment in Figure 4, where the pyramid environment map has been replaced with a direct optimization of its (pre-exponentiated) values. Directly optimizing the mask values instead of using spherical harmonics also results in worse reconstructions (see Table 1)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.581, + 0.47, + 0.657 + ], + "angle": 0, + "content": "model augmented using the orientation loss from Verbin et al. [41], and then use this (potentially-imprecise) proxy geometry in our inverse rendering optimization pipeline. Please refer to the supplemental materials for a detailed description of our NeRF-based model." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.665, + 0.47, + 0.803 + ], + "angle": 0, + "content": "Object self-occlusion does not contain enough signal. Figure 9 quantitatively demonstrates that only using the signal provided by object self-occlusions (\"without occluders\" plot in orange) cannot recover illumination as accurately as our method, which leverages the cue of unintended shadows (\"with occluders\" plot in blue). In particular, exploiting unintended shadows is increasingly important as the illumination contains higher frequencies (Figure 9a) and as the object material becomes increasingly diffuse (Figure 9b)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.81, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Larger occluders improve illumination recovery. Figure 9c shows that increasing the size of the unseen occluders improves the accuracy of the recovered environment maps. Increasing the occluder size effectively improves the problem's conditioning by creating a larger variation in the incident lighting across points on the object." + }, + { + "type": "table", + "bbox": [ + 0.503, + 0.31, + 0.89, + 0.366 + ], + "angle": 0, + "content": "
chairmannequinpotatoestoad
Direct spherical harmonics0.8310.6540.9970.943
Direct environment map0.0420.0710.0150.073
Ours0.0380.0350.0110.046
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.37, + 0.893, + 0.456 + ], + "angle": 0, + "content": "Table 1. Errors (RMSE) on all four of our scenes for our full model compared with our model without the spherical harmonic occluder parameterization, and without the pyramid environment parameterization. While the pyramid representation of the environment map improves smoothness (see Figure 10), we find that representing the masks using spherical harmonics is critical for our method." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.48, + 0.892, + 0.556 + ], + "angle": 0, + "content": "Parameterization ablation study. Table 1 and Figure 10 show that representing the environment map as an image pyramid and the image masks with spherical harmonics, both perform better than simply optimizing their respective element values directly. See supplement for more details." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.568, + 0.614, + 0.584 + ], + "angle": 0, + "content": "6. Discussion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.594, + 0.893, + 0.82 + ], + "angle": 0, + "content": "Even though unintended shadows are common in most real-world capture scenarios, they are often treated as outlier data. In this paper, we showed how to explicitly leverage these shadows as a signal to improve the quality of recovered lighting and materials in particularly challenging scenarios, such as objects made of diffuse materials. Furthermore, we have shown that our algorithm can be used with approximate geometry reconstructed by view synthesis techniques, and that it can produce imperfect, yet promising results on real captures, despite not being designed for this type of data which does not satisfy many of our model's assumptions. We believe that this work makes important first steps towards a general inverse rendering algorithm that can recover geometry, materials, and illuminations from images captured under realistic conditions." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.826, + 0.894, + 0.901 + ], + "angle": 0, + "content": "Acknowledgements. Our sincere thanks to Delio Vicini, Rick Szeliski, Aleksander Holski, and Guy Satat for helpful discussions and suggestions, and to David Salesin, Janne Kontkanen, and Lior Yariv for their help in improving the manuscript before submission." + }, + { + "type": "page_number", + "bbox": [ + 0.49, + 0.946, + 0.508, + 0.956 + ], + "angle": 0, + "content": "84" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.183 + ], + "angle": 0, + "content": "[1] Sai Bangaru, Michael Gharbi, Tzu-Mao Li, Fujun Luan, Kalyan Sunkavalli, Milos Hasan, Sai Bi, Zexiang Xu, Gilbert Bernstein, and Fredo Durand. Differentiable rendering of neural SDFs through reparameterization. SIGGRAPH Asia, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.185, + 0.472, + 0.239 + ], + "angle": 0, + "content": "[2] Manel Baradad, Vickie Ye, Adam B. Yedidia, Frédo Durand, William T. Freeman, Gregory W. Wornell, and Antonio Torralba. Inferring light fields from shadows. CVPR, 2018. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.242, + 0.47, + 0.269 + ], + "angle": 0, + "content": "[3] Jonathan T. Barron and Jitendra Malik. Shape, illumination, and reflectance from shading. TPAMI, 2015. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.27, + 0.47, + 0.325 + ], + "angle": 0, + "content": "[4] Sai Bi, Zexiang Xu, Pratul P. Srinivasan, Ben Mildenhall, Kalyan Sunkavalli, Milos Hasan, Yannick Hold-Geoffroy, David Kriegman, and Ravi Ramamoorthi. Neural reflectance fields for appearance acquisition. arXiv, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.327, + 0.47, + 0.394 + ], + "angle": 0, + "content": "[5] Sai Bi, Zexiang Xu, Kalyan Sunkavalli, Milovs Havsan, Yannick Hold-Geoffroy, David J. Kriegman, and Ravi Ramamoorthi. Deep Reflectance Volumes: Relightable Reconstructions from Multi-View Photometric Images. ECCV, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.396, + 0.47, + 0.45 + ], + "angle": 0, + "content": "[6] Mark Boss, Raphael Braun, Varun Jampani, Jonathan T. Barron, Ce Liu, and Hendrik P. A. Lensch. NeRD: Neural reflectance decomposition from image collections. ICCV, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.452, + 0.47, + 0.507 + ], + "angle": 0, + "content": "[7] Katherine L. Bouman, Vickie Ye, Adam B. Yedidia, Frédo Durand, Gregory W. Wornell, Antonio Torralba, and William T. Freeman. Turning corners into cameras: Principles and methods. ICCV, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.509, + 0.47, + 0.591 + ], + "angle": 0, + "content": "[8] James Bradbury, Roy Frostig, Peter Hawkins, Matthew James Johnson, Chris Leary, Dougal Maclaurin, George Necula, Adam Paszke, Jake VanderPlas, Skye Wanderman-Milne, and Qiao Zhang. JAX: composable transformations of Python+NumPy programs, 2018. http://github.com/google/jax.7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.593, + 0.47, + 0.646 + ], + "angle": 0, + "content": "[9] Paul Debevec. Rendering synthetic objects into real scenes: Bridging traditional and image-based graphics with global illumination and high dynamic range photography. SIGGRAPH, 1998. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.649, + 0.469, + 0.689 + ], + "angle": 0, + "content": "[10] Xi Deng, Fujun Luan, Bruce Walter, Kavita Bala, and Steve Marschner. Reconstructing Translucent Objects using Differentiable Rendering. SIGGRAPH, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.691, + 0.469, + 0.732 + ], + "angle": 0, + "content": "[11] James R Driscoll and Dennis M Healy. Computing Fourier Transforms and Convolutions on the 2-Sphere. Advances in applied mathematics, 1994. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.733, + 0.469, + 0.788 + ], + "angle": 0, + "content": "[12] Marc-Andre Gardner, Kalyan Sunkavalli, Ersin Yumer, Xiaohui Shen, Emiliano Gambaretto, Christian Gagné, and Jean-François Lalonde. Learning to predict indoor illumination from a single image. ACM Trans. Graph., 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.789, + 0.469, + 0.829 + ], + "angle": 0, + "content": "[13] Yannick Hold-Geoffroy, Kalyan Sunkavalli, Sunil Hadap, Emiliano Gambaretto, and Jean-François Lalonde. Deep outdoor illumination estimation. CVPR, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.831, + 0.469, + 0.872 + ], + "angle": 0, + "content": "[14] Wenzel Jakob, Sébastien Speierer, Nicolas Roussel, and Delio Vicini. Dr.Jit: A Just-In-Time Compiler for Differentiable Rendering. SIGGRAPH, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.874, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[15] Brian Karis and Epic Games. Real shading in unreal engine 4. Proc. Physically Based Shading Theory Practice, 2013. 1" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "[16] P. Diederik Kingma and Lei Jimmy Ba. Adam: A Method for Stochastic Optimization. *ICLR*, 2015. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.122, + 0.892, + 0.177 + ], + "angle": 0, + "content": "[17] Zhengfei Kuang, Kyle Olszewski, Menglei Chai, Zeng Huang, Panos Achlioptas, and Sergey Tulyakov. Neroic: Neural rendering of objects from online image collections. ACM Trans. Graph., 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.179, + 0.892, + 0.233 + ], + "angle": 0, + "content": "[18] Zhengfei Kuang, Yunzhi Zhang, Hong-Xing Yu, Samir Agarwala, Shangzhe Wu, and Jiajun Wu. Stanford-ORB: A realworld 3d object inverse rendering benchmark. NeurIPS, 2023. 7, 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.235, + 0.892, + 0.276 + ], + "angle": 0, + "content": "[19] Jean-François Lalonde, Alexei A Efros, and Srinivasa G Narasimhan. Estimating the natural illumination conditions from a single outdoor image. IJCV, 2012. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.278, + 0.892, + 0.332 + ], + "angle": 0, + "content": "[20] Chloe LeGendre, Wan-Chun Ma, Graham Fyffe, John Flynn, Laurent Charbonnel, Jay Busch, and Paul Debevec. Deeplight: Learning illumination for unconstrained mobile mixed reality. CVPR, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.334, + 0.892, + 0.401 + ], + "angle": 0, + "content": "[21] Chloe LeGendre, Wan-Chun Ma, Rohit Pandey, Sean Fanello, Christoph Rhemann, Jason Dourgarian, Jay Busch, and Paul Debevec. Learning illumination from diverse portraits. SIGGRAPH Asia 2020 Technical Communications, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.404, + 0.892, + 0.445 + ], + "angle": 0, + "content": "[22] Tzu-Mao Li, Miika Aittala, Frédo Durand, and Jaakko Lehtinen. Differentiable monte carlo ray tracing through edge sampling. SIGGRAPH Asia, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.447, + 0.892, + 0.502 + ], + "angle": 0, + "content": "[23] Zhengqin Li, Zexiang Xu, Ravi Ramamoorthi, Kalyan Sunkavalli, and Manmohan Chandraker. Learning to reconstruct shape and spatially-varying reflectance from a single image. SIGGRAPH Asia, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.504, + 0.892, + 0.559 + ], + "angle": 0, + "content": "[24] Zhengqin Li, Mohammad Shafiei, Ravi Ramamoorthi, Kalyan Sunkavalli, and Manmohan Chandraker. Inverse rendering for complex indoor scenes: Shape, spatially-varying lighting and svbrdf from a single image. CVPR, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.561, + 0.892, + 0.615 + ], + "angle": 0, + "content": "[25] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. NeRF: Representing Scenes as Neural Radiance Fields for View Synthesis. ECCV, 2020. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.617, + 0.892, + 0.657 + ], + "angle": 0, + "content": "[26] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. SIGGRAPH, 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.66, + 0.892, + 0.715 + ], + "angle": 0, + "content": "[27] Jacob Munkberg, Jon Hasselgren, Tianchang Shen, Jun Gao, Wenzheng Chen, Alex Evans, Thomas Mueller, and Sanja Fidler. Extracting Triangular 3D Models, Materials, and Lighting From Images. CVPR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.717, + 0.892, + 0.757 + ], + "angle": 0, + "content": "[28] Merlin Nimier-David, Delio Vicini, Tizian Zeltner, and Wenzel Jakob. Mitsuba 2: A retargetable forward and inverse renderer. SIGGRAPH Asia, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.759, + 0.892, + 0.813 + ], + "angle": 0, + "content": "[29] Merlin Nimier-David, Sébastien Speierer, Benoit Ruiz, and Wenzel Jakob. Radiative Backpropagation: An Adjoint Method for Lightning-Fast Differentiable Rendering. SIGGRAPH, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.815, + 0.892, + 0.843 + ], + "angle": 0, + "content": "[30] Ko Nishino and Shree K Nayar. Eyes for relighting. ACM Trans. Graph., 2004. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.845, + 0.892, + 0.872 + ], + "angle": 0, + "content": "[31] Jeong Joon Park, Aleksander Holynski, and Steven M Seitz. Seeing the world in a bag of chips. CVPR, 2020. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.874, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[32] Matt Pharr, Wenzel Jakob, and Greg Humphreys. Physically Based Rendering: From Theory to Implementation (3rd" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.491, + 0.947, + 0.507, + 0.956 + ], + "angle": 0, + "content": "85" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.11, + 0.092, + 0.468, + 0.12 + ], + "angle": 0, + "content": "ed.). Morgan Kaufmann Publishers Inc., San Francisco, CA, USA, 3rd edition, 2016. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.122, + 0.468, + 0.149 + ], + "angle": 0, + "content": "[33] Ravi Ramamoorthi and Pat Hanrahan. A signal-processing framework for inverse rendering. SIGGRAPH, 2001. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.151, + 0.468, + 0.177 + ], + "angle": 0, + "content": "[34] Imari Sato, Yoichi Sato, and Katsushi Ikeuchi. Illumination from shadows. TPAMI, 2003. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.18, + 0.468, + 0.207 + ], + "angle": 0, + "content": "[35] Shuran Song and Thomas Funkhouser. Neural illumination: Lighting prediction for indoor environments. CVPR, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.209, + 0.468, + 0.263 + ], + "angle": 0, + "content": "[36] Pratul P. Srinivasan, Boyang Deng, Xiuming Zhang, Matthew Tancik, Ben Mildenhall, and Jonathan T. Barron. NeRV: Neural reflectance and visibility fields for relighting and view synthesis. CVPR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.266, + 0.468, + 0.306 + ], + "angle": 0, + "content": "[37] Tristan Swedish, Connor Henley, and Ramesh Raskar. Objects as cameras: Estimating high-frequency illumination from shadows. ICCV, 2021. 1, 3, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.308, + 0.468, + 0.376 + ], + "angle": 0, + "content": "[38] Matthew Tancik, Pratul Srinivasan, Ben Mildenhall, Sara Fridovich-Keil, Nithin Raghavan, Utkarsh Singhal, Ravi Ramamoorthi, Jonathan Barron, and Ren Ng. Fourier features let networks learn high frequency functions in low dimensional domains. NeurIPS, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.379, + 0.468, + 0.419 + ], + "angle": 0, + "content": "[39] Antonio Torralba and William T. Freeman. Accidental pin-hole and pinspeck cameras: Revealing the scene outside the picture. CVPR, 2012. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.421, + 0.468, + 0.461 + ], + "angle": 0, + "content": "[40] Eric Veach and Leonidas J Guibas. Optimally Combining Sampling Techniques for Monte Carlo Rendering. SIGGRAPH, 1995. 6, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.464, + 0.468, + 0.518 + ], + "angle": 0, + "content": "[41] Dor Verbin, Peter Hedman, Ben Mildenhall, Todd Zickler, Jonathan T. Barron, and Pratul P. Srinivasan. Ref-NeRF: Structured View-Dependent Appearance for Neural Radiance Fields. CVPR, 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.521, + 0.468, + 0.56 + ], + "angle": 0, + "content": "[42] Delio Vicini, Sébastien Speierer, and Wenzel Jakob. Differentiable Signed Distance Function Rendering. SIGGRAPH, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.564, + 0.468, + 0.605 + ], + "angle": 0, + "content": "[43] Bruce Walter, Stephen R Marschner, Hongsong Li, and Kenneth E Torrance. Microfacet models for refraction through rough surfaces. Eurographics, 2007. 6, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.607, + 0.468, + 0.647 + ], + "angle": 0, + "content": "[44] Adam B Yedidia, Manel Baradad, Christos Thrampoulidis, William T Freeman, and Gregory W Wornell. Using unknown occluders to recover hidden scenes. CVPR, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.649, + 0.468, + 0.69 + ], + "angle": 0, + "content": "[45] Tizian Zeltner, Sébastien Speierer, Iliyan Georgiev, and Wenzel Jakob. Monte Carlo Estimators for Differential Light Transport. SIGGRAPH, 2021. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.692, + 0.468, + 0.746 + ], + "angle": 0, + "content": "[46] Kai Zhang, Fujun Luan, Qianqian Wang, Kavita Bala, and Noah Snively. PhySG: Inverse rendering with spherical gaussians for physics-based material editing and relighting. CVPR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.748, + 0.468, + 0.803 + ], + "angle": 0, + "content": "[47] Xiuming Zhang, Pratul P. Srinivasan, Boyang Deng, Paul Debevec, William T. Freeman, and Jonathan T. Barron. NeR-Factor: Neural factorization of shape and reflectance under an unknown illumination. SIGGRAPH Asia, 2021. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.468, + 0.803 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.491, + 0.947, + 0.508, + 0.956 + ], + "angle": 0, + "content": "86" + } + ] +] \ No newline at end of file diff --git a/2024/Eclipse_ Disambiguating Illumination and Materials using Unintended Shadows/a9a99e7b-5bee-4200-8036-13e74c55e36e_origin.pdf b/2024/Eclipse_ Disambiguating Illumination and Materials using Unintended Shadows/a9a99e7b-5bee-4200-8036-13e74c55e36e_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..bd541986f4f9cffc6c5162105b51b93e22b8f027 --- /dev/null +++ b/2024/Eclipse_ Disambiguating Illumination and Materials using Unintended Shadows/a9a99e7b-5bee-4200-8036-13e74c55e36e_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6dcc84c725dab82b59a2197e217d750f105013cef8601e30a0f5894981aee648 +size 5653098 diff --git a/2024/Eclipse_ Disambiguating Illumination and Materials using Unintended Shadows/full.md b/2024/Eclipse_ Disambiguating Illumination and Materials using Unintended Shadows/full.md new file mode 100644 index 0000000000000000000000000000000000000000..9a8132d95a0b08a3c3a20f00b8d959ded24d17f3 --- /dev/null +++ b/2024/Eclipse_ Disambiguating Illumination and Materials using Unintended Shadows/full.md @@ -0,0 +1,407 @@ +# Eclipse: Disambiguating Illumination and Materials using Unintended Shadows + +Dor Verbin + +Ben Mildenhall1 + +Peter Hedman + +Jonathan T. Barron1 + +Todd Zickler $^{1,2}$ + +Pratul P. Srinivasan1 + +1Google Research + +$^{2}$ Harvard University + +![](images/fa3830b6da4b09f6b20011f65294f8112da2ea60f58383464472f4f750b6e842.jpg) +(a) Problem setup + +![](images/20bd31f14abca80306765369e4d607fc075573d7a45bbcc446da76ae2ed5df02.jpg) + +![](images/070bd0bd8dc3f3b994e92d68b7875826c86978889e196dd1a21eef26bac1e987.jpg) +Albedo +Illumination and occluders + +![](images/a5310ae4b2c78c42d0dc5e88c9801c49f93d0dad67194147c0171334cd72522b.jpg) +Albedo + +![](images/1e2c5fbe3d6b8eb10743d8fc365297a366652b3164fecd91f5ab0640a4c46f8b.jpg) +Illumination and occluders + +![](images/7e50a000c9c96cee488df38da1c6d47e2e25346c892d823d5eb7aacd2af78c93.jpg) + +![](images/39f08ffd84ba2f6ae864755c702a4163eb3829faeceeb6f04f312931c2c2dc15.jpg) +Figure 1. (a) We exploit unintended shadows cast by camera operators (or other unobserved moving occluders) to recover high-fidelity environment lighting and object materials from a set of images. Without modeling such unobserved occluders, prior methods can only: (b) recover convolved lighting without explicit material decomposition [31]; or (c) exploit lighting occlusions that occur internally among the observed objects [37]. (d) We show that additionally modeling and recovering external, unobserved occluders enables lighting and material reconstructions that are closer to ground truth (e). + +![](images/2ad95e1e242aebb2bdc0886673b5558f8b5e651de1490597bb48f0d0e5f3c046.jpg) + +![](images/692bf59ab3edbda586f02177204c85bfd6f3b277a773600180d0f64915325b43.jpg) + +# Abstract + +Decomposing an object's appearance into representations of its materials and the surrounding illumination is difficult, even when the object's 3D shape is known beforehand. This problem is especially challenging for diffuse objects: it is ill-conditioned because diffuse materials severely blur incoming light, and it is ill-posed because diffuse materials under high-frequency lighting can be indistinguishable from shiny materials under low-frequency lighting. We show that it is possible to recover precise materials and illumination—even from diffuse objects—by exploiting unintended shadows, like the ones cast onto an object by the photographer who moves around it. These shadows are a nuisance in most previous inverse rendering pipelines, but here we exploit them as signals that improve conditioning and help resolve material-lighting ambiguities. We present a method based on differentiable Monte Carlo ray tracing that uses images of an object to jointly recover its spatially-varying materials, the surrounding illumination environment, and the shapes of the unseen light occluders who inadvertently cast shadows upon it. + +# 1. Introduction + +In this work, we show that the long-standing inverse rendering problem of recovering an object's material properties and the surrounding illumination environment from a set of images can greatly benefit from considering the effects of unobserved moving light occluders, such as the camera operator, which partially block incoming light and inadvertently cast shadows onto the object being imaged. + +Joint recovery of materials and lighting is a challenging task because the BRDF (bidirectional reflectance distribution function, which represents how a material at any point on an object's surface maps from incoming to outgoing light) acts as a directional filter on incoming light. BRDFs for specular (shiny) materials act as all-pass directional filters, while BRDFs for diffuse materials act as low-pass directional filters. As a result, inverse rendering is fundamentally ambiguous since a shiny object illuminated by blurry lighting can be indistinguishable from a diffuse object illuminated by sharp lighting (Figure 2). Furthermore, since diffuse materials act as low-pass filters and strongly blur incoming light, even the simpler problem of recovering lighting from images of a known diffuse material can be + +severely ill-conditioned, precluding the recovery of high-frequency illumination [33]. + +We observe that both of these issues can be ameliorated by exploiting an effect that naturally exists whenever an object is imaged from a sequence of viewpoints under static environment lighting. Namely, between one image and the next, the positions of the camera and its operator(s) must change, and in doing so they occlude different portions of the surrounding environment, casting distinct sets of shadows onto the object being imaged. This effect is very noticeable under strongly-directional lighting or for shiny objects because the cast shadows are sharp or the specular highlights are misshapen and hidden. But even when the effect is barely perceptible, such as for diffuse objects under well-distributed lighting, there exists a subtle signal that one can take advantage of. + +In this paper, we make first steps toward a practical algorithm that exploits this cue by using gradient descent with differentiable Monte Carlo ray tracing to jointly recover explicit representations of (i) spatially-varying reflectance, (ii) environment illumination, and (iii) the per-image shapes of any unseen, light-blocking occluders. We evaluate our algorithm using a challenging variety of simulated scenes, including scenes with diffuse-only materials and with object geometry that is not known beforehand. We also apply our method to an existing off-the-shelf dataset designed for inverse rendering, where the (unseen) camera rig moving to capture the scene acts as an accidental occluder. Our results demonstrate that external shadowing effects provide a strong and useful cue, even when they are subtle, and even without relying on strong domain-specific priors for the materials, illumination, or external occluder shapes. This suggests that incorporating unintended shadows into inverse rendering pipelines for real, captured data is valuable for improving the quality of material and lighting assets. + +# 2. Related Work + +We build on recent developments in differentiable ray tracing and on a long history of inverse rendering work, including the joint estimation of reflectance and lighting, and the estimation of lighting alone. We are also inspired by a separate line of work on passive non-line-of-sight imaging, which exploits similar light-blocking effects. + +Physics-based differentiable ray tracing. Modern differentiable ray tracers [22, 28] enable the computation of gradients of rendered images with respect to scene parameters (i.e., geometry, materials, and lighting) by differentiating through light transport simulation. Recent works have focused on improving efficiency and performance [14, 29] and on accurately differentiating through visibility discontinuities—such as those caused by shadow and occlusion boundaries—with respect to shape and light- + +ing [1, 42]. Our implementation of differentiable ray tracing leverages insights from these works as well as from Zeltner et al. [45], who provide design intuition for Monte Carlo differentiable ray tracers. + +Inverse rendering of materials and lighting. Decomposing an object's appearance into representations of material and lighting is a long-standing problem in computer vision and graphics. In their foundational work, Ramamoorthi and Hanrahan [33] developed a signal processing approach by describing the outgoing light at a surface point as the spherical convolution of the BRDF and the incoming lighting. This formulation elucidates why inverse rendering is ill-posed and often ill-conditioned: it is ill-posed because there are multiple illumination-material pairs that convolve into the same image, and it is ill-conditioned because diffuse BRDFs act as a low-pass filters on lighting, which causes the estimates of medium- and high-frequency illumination to be very sensitive to noise. + +Because of this, most approaches to inverse rendering rely on strong priors on materials and lighting. Single-view approaches have used hand-designed priors [3] or priors in the form of neural networks trained with supervision from large datasets of material and lighting labels [23, 24]. More relevant to us are multi-view approaches, which typically either assume known lighting [4, 5, 36] or rely on strong priors, such as assuming a single, highly-specular BRDF for the entire object [46] or lighting from a prior distribution that was pre-trained on a dataset of environment maps [6, 47]. The most closely related work is that for which lighting information is not provided as input [17, 27]. + +For the sake of generality, we do not use such strong priors in our experiments because by avoiding them we can more directly measure the benefits that are gained by modeling unintended shadows as an additional cue. Because our model uses a generic gradient-descent framework, stronger application-specific priors can be added to it to improve performance on a specific domain of interest. + +Estimating environment lighting. Another thread of inverse rendering research focuses on the task of recovering high-fidelity lighting environments, typically for augmented reality applications where virtual objects are rendered into photographs with consistent reflections and shadows. Prior work by Debevec [9] used images of a chrome sphere to measure environment lighting directly, and subsequent works have demonstrated that plausible environment lighting can be estimated without chrome spheres, using images of indoor [12, 35] and outdoor scenes [13, 19, 20], or images of a specific class of objects like faces [21, 30]. These subsequent techniques use strong priors in the form of deep neural network weights that are trained with supervision to map from images to lighting, and they do not enforce physical rendering consistency between the recovered + +lighting and the observations. In contrast, we avoid such strong priors, and we explicitly enforce consistency. + +More related to our approach is the work of Park et al. [31], which recovers physically-consistent environment maps from RGBD videos of shiny objects. However, their algorithm can only recover the convolution of the environment map with the BRDF of the observed object (e.g., Figure 1(b)), which precludes the recovery of high illumination frequencies unless the object is highly specular. + +Also related is the work of Swedish et al. [37], which places a known, diffuse object on a ground plane (a scenario first studied by Sato et al. [34]) and uses its cast shadows to recover high-quality lighting. Their formulation is linear and it improves lighting estimates by leveraging self-shadowing among diffuse objects that have known geometry and albedo (e.g., Figure 1(c)). We generalize this by replacing the linear formulation with differentiable Monte Carlo ray tracing, which allows exploiting the additional shadowing effects caused by moving external occluders and leads to substantially improved results (e.g., Figure 1(d)). Our formulation also handles objects with more general spatially-varying BRDFs that are not known beforehand. + +Passive non-line-of-sight imaging with occlusions. Passive non-line-of-sight techniques can also be seen as recovering environment illumination: They observe a reflective surface (which is typically diffuse and planar) and recover the appearance of a "hidden scene" from these reflections. Similar to us, prior works in this area have observed that the presence of an occluder between the observed surface and the hidden scene/environment aids recovery by introducing sharp angular variations into the rendering integral. This insight was first leveraged in settings with simple occluder-shapes like pinholes and pinspecks [39] or corners joining walls [7]. Subsequent work by Baradad et al. [2] generalizes this by considering light-occlusion effects caused by an arbitrary but known 3D shape (a task that is closely related to Swedish et al. above). Yedidia et al. [44] additionally recover the unobserved occluder's shape assuming it is a planar mask parallel to the observed planar surface. + +We generalize these prior works by using differentiable Monte Carlo rendering to replace their deconvolution-based algorithms, which are tailored to the specific case where the observed reflector is planar and diffuse. This allows using reflective objects that have arbitrary shapes, and arbitrary spatially-varying BRDFs which are not known beforehand. It also allows recovering the shapes of unobserved occluders that are arbitrary and time-varying. + +# 3. Motivation and Problem Setup + +A well-known ambiguity in computer vision and graphics occurs when decomposing an image into its component lighting and materials. The top of Figure 2 recreates + +![](images/b046042743bcf125cc9ab4f2f0648209edda8e1e76468ba73c3b13f950e29bbe.jpg) +Figure 2. Decomposing appearance into lighting and materials can be inherently ambiguous. Top row: a diffuse sphere lit by high-frequency lighting (inset) is visually indistinguishable from a shiny, mirror-like sphere lit by low-frequency lighting. Bottom row: A second image captured in the presence of a dark external occluder resolves the ambiguity. The occluder's effect is clearly visible on the shiny sphere while on the diffuse sphere it produces shadows that are soft and very subtle. + +a common depiction of this ambiguity, where an image of a known shape (a sphere) is explained equally well by a diffuse material in a complicated lighting environment or a shiny, mirror-like material under low-frequency lighting. Now, imagine we capture a second image from the same viewpoint, after some external object enters the scene. This second object remains beyond the field of view and so is not directly observed, but it acts as an external occluder that prevents some of the environment's light from reaching the sphere. This second image, shown at the bottom of the figure for each case, clearly reveals which of the two materiallighting explanations is the correct one. + +We aim to make use of this unintended shadowing that naturally occurs whenever a camera and its operators move around an object while capturing images from different viewpoints. They affect each image by blocking different portions of the surrounding light, and this provides a helpful signal. We will show that this signal is helpful even when the occluder shapes and locations are quite arbitrary and not known beforehand, and when their shadowing effects are very subtle, like between the top and bottom images on the left of Figure 2. + +In addition to helping resolve material-lighting ambiguities, shadowing from moving external occluders also improves the conditioning of our inverse rendering problem. This is analyzed in Figure 3, which considers the simplified hypothetical 1D case of recovering an environment illumination from images of a diffuse disk-shaped object, with known occluder positions. Convolution with a dif + +![](images/6ed747df87e5cfd7f61229da52dbbbebf5eb52a1f438457e7547d63cd8e50879.jpg) +Figure 3. Here we plot eigenvalue magnitudes as a function of illumination angular frequency for a one-dimensional circular Lambertian object with known albedo, under the simplifying assumption of known occluder locations and shapes. For a single input image without occlusion, almost all lighting frequencies are strongly attenuated. In the single-occluder case (similar to the scenario considered by Baradad et al. [2]), some low frequencies become recoverable but intermediate and high frequencies do not. Observing additional images enables the recovery of more lighting information, especially at the highest frequencies. + +fuse BRDF acts as a low-pass filter of the incoming light, which means that the linear inverse problem's eigenvalues are vanishingly low for higher frequencies. Moving external occluders introduce a sharp angular variation in the per-image illumination, which better-conditions these high frequency components, and allows their recovery. As shown in Figure 3, in this simplified scenario a single image containing an occluder is better than a single image without any occluders, but using observations of more occluders makes the problem even better-conditioned. We demonstrate this effect in practice in the supplement. + +# 3.1. Formulation + +Let $\{I_t\}_{t=1}^T$ be a collection of $T$ images of a scene with known camera poses; for every $t$ , $I_t$ is an RGB image with spatial size $H \times W$ . We assume the observed scene $\mathcal{S} \subset \mathbb{R}^3$ is illuminated by a far-field environment map $L(\hat{\omega}_i)$ that does not change over time, so the only temporal changes in the incident light field are caused by unobserved light-blockers (called occluders hereafter) that move around the scene outside the field of view. We can then write the incident light at any observed position $\mathbf{x} \in \mathcal{S}$ as: + +$$ +L _ {t} (\mathbf {x}, \hat {\omega} _ {i}) = L (\hat {\omega} _ {i}) M _ {t} (\mathbf {x}, \hat {\omega} _ {i}) V (\mathbf {x}, \hat {\omega} _ {i}), \tag {1} +$$ + +where for any $t$ , $M_{t}(\mathbf{x}, \hat{\boldsymbol{\omega}}_{i})$ is a binary signal with value 0 for directions blocked by the occluder and 1 otherwise, and the binary signal $V(\mathbf{x}, \hat{\boldsymbol{\omega}}_{i})$ models visibility effects that are + +internal to the scene, with value 0 for directions from $\mathbf{x}$ that are blocked by other scene elements, and value 1 otherwise. + +Omitting global illumination effects, a pixel $\mathbf{u}$ corresponding to a surface point $\mathbf{x}$ with BRDF $f$ viewed from direction $\hat{\omega}_o$ at time $t$ has color: + +$$ +I _ {t} (\mathbf {u}) = \int_ {\mathbb {S} ^ {2}} L _ {t} (\mathbf {x}, \hat {\omega} _ {i}) f (\mathbf {x}, \hat {\omega} _ {i}, \hat {\omega} _ {o}) (\hat {\mathbf {n}} (\mathbf {x}) \cdot \hat {\boldsymbol {\omega}} _ {i}) _ {+} d \hat {\boldsymbol {\omega}} _ {i}, \tag {2} +$$ + +where $\hat{\mathbf{n}} (\mathbf{x})$ is the surface normal corresponding to $\mathbf{x}$ , and $(\cdot)_{+}$ clamps negative values to zero. + +Given the set of observed images, we would like to recover the environment map $L$ , a spatially-varying BRDF $f$ at every point on the surface of the object $S$ , as well as the set of unobserved occluder masks $\{M_t\}_{t=1}^T$ , one for each observation time $t$ . Our goal is therefore to solve the following optimization problem: + +$$ +\underset {\phi^ {(o)}, \phi^ {(m)}, \phi^ {(\ell)}} {\arg \min } \sum_ {t, \mathbf {u}} \left\| I _ {t} (\mathbf {u}) - \mathcal {R} _ {t} \left(\mathbf {u}; \phi^ {(o)}, \phi^ {(m)}, \phi^ {(\ell)}\right) \right\| ^ {2}, \tag {3} +$$ + +where $\mathcal{R}_t(\mathbf{u};\cdot)$ renders the pixel location $\mathbf{u}$ at time $t$ using the occluder parameters $\phi^{(o)}$ , material parameters $\phi^{(m)}$ , and illumination parameters $\phi^{(\ell)}$ . + +At first glance, it may seem that solving for occluders in addition to materials and illumination should make the resulting inverse problem more difficult. However, as illustrated in Figure 3 and demonstrated by our experimental results, modeling and solving for occluders actually makes the full problem more well-conditioned and thus improves recovery of material and illumination. + +# 4. Method + +We first describe our parameterizations of the unseen oc-clusters, illumination, and materials (Sections 4.1-4.3). Section 4.4 then relates these parameters to rendered pixel colors, and Section 4.5 describes how we optimize to solve the inverse problem in Equation 3. + +We emphasize that our method is designed to exploit external shadows across as many scene-types as possible, and so its only priors are those implicit to our parameterizations. In our experiments, we verify that these relatively weak priors are sufficient for recovering high-fidelity occluder shapes, illumination maps, and material maps. + +# 4.1.Occluders + +Define a world coordinate system with origin at the scene's center, and let $\{r_t\}_{t=1}^T$ be the radial distances from the origin to each (known) camera's center. Then we model each of the $T$ occluders $M_t$ as an independent binary signal defined on the surface of a finite, radius- $r_t$ sphere that is centered at the origin. The value of the binary occluder signal + +![](images/692f6c4c60ad0fde441093d2f689c0ca2f7930225543b1dc458cbf36616af74a.jpg) +PSNR = 31.2 dB + +![](images/bfb05bd52a53941c2916344e8532302f0842f14a221170725fc67aa512bfd39d.jpg) + +![](images/013a22404f239d105b52f4540b3b8d4addf4f0d663cac0d11549c80dc352e0b0.jpg) + +![](images/e20880b2ecb8cb8e5c828ef3f744b3d45a6fda19a408a9c1599b49cd3ce1f4b7.jpg) +PSNR = 26.4 dB + +![](images/8c849c23079cdbd9c7fca8520f96e9926475449810db2fce6a9e75b24164059a.jpg) + +![](images/21e82359b9c42bf0e73dedbe36315326d37630eca6f14f4c6703fd6784fbf1bb.jpg) + +![](images/839e78ad849635b81b1ab2d428c1448386e2aafa8d5989a4bd62043becaf3698.jpg) +(a) Sample input image + +![](images/5f6df5ae024dcc3b2b02d44617ecda302035578f4a78db0b62e8929ee6eeb829.jpg) +PSNR = 29.6 dB + +![](images/722be76b5e4fd12bcd9170e47a0875fdf998f8a8cfadb15a7da83cf7a69c6681.jpg) +(c) True albedo + +![](images/4869fa775491944f014ae608cc50095f92748748a1489303987aeb8ab41b44da.jpg) + +![](images/9e91b7227707495da3ab7d67985eaf1e2d50ae51823e0d7b4b609dcb47f66a07.jpg) + +![](images/6a87389ff443254ee77238bac0be9b75c018ad39691f15fe93d1604e0c22c4cd.jpg) +$\mathrm{RMSE} = 0.073$ + +![](images/9605522f36447750b8164f9d293250441442b3f48c2461898c3bc392724d1c0b.jpg) + +![](images/a8a7a34b84fbd30c914914c874079ebe93f8d0ec280a427470361edbe8f17b20.jpg) +RMSE = 0.081 +$\mathrm{RMSE} = 0.093$ +(d) Recovered illumination + +![](images/44a2d5b874791ddb6c04886f520d02945a25e1405f676b69fdce90795f914751.jpg) +(e) True illumination + +![](images/7f99c9753ef13d7446e332cb81696c838627fc37380141451535b4f3411392bc.jpg) +Figure 4. The results of our method on three additional diffuse objects. We report the RMSE of each environment map in linear color space but plot the images after tonemapping for better evaluation of the full dynamic range. The albedo PSNR values are reported on object pixels only. + +![](images/18e55259764e22af73ac545f6283f314a90f787e3698caaf833184ad1ff1ca68.jpg) +(b) Recovered albedo + +![](images/9ffa38b2fc7850f3bfd49bbc996dba852cdf282a83fbce33954c7d997d393786.jpg) +Figure 5. Our occluders extracted from the potatoes scene on the top of Figure 4 resemble the true ones (left). However, as expected, inaccuracies in the recovered masks tend to occur at relatively-dark regions of the environment map: there is a very strong inverse correlation (right) between the mean error in the recovered masks and the amount of illumination energy that they block. + +for a shadow ray emitted from surface point $\mathbf{x} \in S$ in direction $\hat{\omega}_i$ at time $t$ can be computed by intersecting the shadow ray with the occluder's spherical shell: + +$$ +M _ {t} (\mathbf {x}, \hat {\omega} _ {i}) = \tilde {M} _ {t} \left(S (\mathbf {x}, \hat {\omega} _ {i}, r _ {t})\right) \tag {4} +$$ + +where $S(\mathbf{x},\hat{\omega}_i,r)$ is the intersection of the ray with a sphere of radius $r_t$ , normalized to have unit length: + +$$ +S (\mathbf {x}, \hat {\boldsymbol {\omega}} _ {i}, r) = \frac {\mathbf {x} + \left(\sqrt {(\mathbf {x} \cdot \hat {\boldsymbol {\omega}} _ {i}) ^ {2} + r ^ {2} - \| \mathbf {x} \| ^ {2}} - \mathbf {x} \cdot \hat {\boldsymbol {\omega}} _ {i}\right) \hat {\boldsymbol {\omega}} _ {i}}{r}. \tag {5} +$$ + +Note that there is a single intersection, since we assume that the surface points are always inside the occluder spheres. + +Although our formulation models the occluders as binary signals, a binary representation is not well-suited for gradient-based optimization. Instead, we represent the occluders as a continuously-valued function on the sphere in a spherical harmonic basis, mapped using a sigmoid function $\sigma$ to lie in [0, 1]: + +$$ +\tilde {M} _ {t} (\hat {\omega}) = \sigma \left(\sum_ {\ell = 0} ^ {P} \sum_ {m = - \ell} ^ {\ell} a _ {t \ell m} Y _ {\ell} ^ {m} (\hat {\omega})\right), \tag {6} +$$ + +where $\phi^{(o)}\triangleq \{a_{t\ell m}\}$ are estimizable coefficients that parameterize the occluder at time $t$ , and $P$ is the degree required to span the space of spherical images with the same resolution as our environment illumination (see below) [11]. + +Note that for radially symmetric BRDFs, the rendering integral in Equation 2 can be written as a spherical convolution [11, 33]. This makes spherical harmonics a natural choice for representing the occluder signals, since spherical convolutions are diagonal in the basis of spherical harmonics. See our supplement for more details. + +# 4.2. Environment Illumination + +We assume that the scene's illumination is distant and can be represented as an environment map. The environment map is parameterized as an image pyramid of size $H' \times W'$ + +![](images/eb9121225b634bdf165cbf8c0046e239316bcd9c87059ee4f484eaeb5e08f6c2.jpg) +(a) Recovered illumination (no occluder) + +![](images/43aa90649c1995ebeff7b2492ffb2fb682441daad0fe10b35775f1f02d43b0fc.jpg) +(b) Recovered illumination +Figure 6. In the case of purely Lambertian material (with the potatoes geometry), (a) only using self occlusions is insufficient for the recovery of high-frequency content in the illumination. (b) Modeling and estimating occluders recovers significantly higher-quality illumination that closely resembles the ground truth. + +![](images/44147fbc4e337644558b9d78179e4ec47fea4a755b1cb9d1d8d8574f736de55d.jpg) +(c) True illumination + +in equirectangular coordinates $(50 \times 100$ in our experiments), with an exponential nonlinearity: + +$$ +L \left(\hat {\omega} _ {i}\right) = \exp \left(\sum_ {k = 0} ^ {K - 1} a ^ {k} L _ {k} \left(\hat {\omega} _ {i}\right)\right), \tag {7} +$$ + +where $\phi^{(\ell)}\triangleq \{L_k\}_{k = 0}^{K - 1}$ comprise a coarse ( $k = 0$ ) to fine ( $k = K - 1$ ) representation of the illumination, and we set $a = 2$ . For each $k$ , $L_{k}(\hat{\omega}_{i})$ is computed by bilinearly interpolating into a grid whose size exponentially increases with $k$ . The exponential nonlinearity is helpful for obtaining high dynamic range values in the environment map, as noted, e.g. in [3]. + +Unlike the occluder masks, we find that representing the environment map as a pyramid results in better performance than what is achieved with spherical harmonics. See appendix for additional details. + +# 4.3. Materials + +We represent the scene's spatially-varying BRDF as the sum of a diffuse and a specular component. We use a Lambertian model for the diffuse term and a GGX microfacet model [43] for the specular term. The spatially-varying BRDF parameters are represented by a coordinate-based multi-layer perceptron (MLP) with a positional encoding [38] function $\gamma$ and estimizable parameters $\phi^{(m)}$ : + +$$ +\boldsymbol {\alpha} (\mathbf {x}) = \operatorname {M L P} \left(\gamma (\mathbf {x}); \phi^ {(m)}\right). \tag {8} +$$ + +Here, $\gamma (\mathbf{x})$ is the positionally-encoded position on the object's surface, and $\alpha (\mathbf{x})\in \mathbb{R}^5$ are BRDF coefficients, consisting of an RGB diffuse albedo, a spatially-varying scalar microfacet roughness, and a spatially-varying scalar specular reflectance at normal incidence (equivalent to a reparameterization of the material's index of refraction). See the supplement for an exact specification of our BRDF model. + +# 4.4. Rendering + +We render the occluders, materials, and illumination, i.e. the rendering operator $\mathcal{R}_t$ from Problem 3, by approximating the integral in Equation 2 using Monte Carlo techniques. + +![](images/82b89b13cf92cd41db5da1bf57de2b83ac32155d5f733d975a452065600067c9.jpg) +(a) Sample input image + +![](images/f9fd746aa9328e892228beb2630c16b8593a0ca51e8c74a4900d9a4d58336467.jpg) +(b) Recovered albedo + +![](images/efdaf10d6b5ca1f563d6c15c6b58408a9ec3efab29a9025dfe1057727bd12dc8.jpg) +(c) True albedo + +![](images/404210c2a17e307793a4aaa1e101134d675cb92cca16ff3d35274bd0e985283b.jpg) +(d) Recovered illumination +Figure 7. We are able to recover illumination even when geometry is unknown by first optimizing a volumetric representation of geometry using a NeRF-based method. Despite this inaccurate proxy geometry, our method still recovers plausible (albeit blurry) illumination (d) and albedo (b). + +![](images/b2d26c2b212e96a3245e2858551847fb6472036fc0dd078701b0a34e89a5c8f7.jpg) +(e) True illumination + +We use a standard multiple importance sampler [40] based on the illumination and material model. We model shadows cast by the occluders and self-occlusions by the object itself, but for efficiency and due to our focus on diffuse objects, we neglect lower-order effects and global illumination. See the supplemental material for a full description of our renderer. + +# 4.5. Optimization + +We optimize the objective in Equation 3 using the $L_{2}$ error between rendered values and ground truth ones. We use Adam [16] to optimize all three components of our model, with a learning rate of $3 \cdot 10^{-3}$ for the environment map and materials, and a learning rate of 1 for the occluders. + +In each iteration, we compute the $L_{2}$ loss using a batch size of $2^{16}$ pixels. In order to avoid bias in the gradient updates to our model's parameters, we use the same approach as [10]: we render every pixel value twice using independent samples and multiply the per-channel deviations of both from the ground truth to get the expected loss value: + +$$ +\mathcal {L} = \sum_ {\mathbf {u}} \left(\tilde {I} _ {t} ^ {(1)} (\mathbf {u}) - I _ {t} (\mathbf {u})\right) \cdot \left(\tilde {I} _ {t} ^ {(2)} (\mathbf {u}) - I _ {t} (\mathbf {u})\right), \tag {9} +$$ + +![](images/8969b121a22bd2300a109fbea4bd520f17f6be2aa2c08364b4bfcc6f495f85d4.jpg) + +![](images/1b2ce43d5f006ad02de392dd16365648815d779d1d83390ccd913aa8a7e30f17.jpg) +PSNR = 19.3 dB + +![](images/52fe5dc510adbffb435d476eab6520a392cff5adf340609fb56ec57c39a77e6c.jpg) +PSNR = 14.4 dB + +![](images/d7a52d26f6a3899dee3a2a66af0108304fbc5da7c0f635b203d4fb97fc50f56e.jpg) + +![](images/f9e5985df6daf92fc8905fdf3fa38da529fd1ee6402097a312d405851642c25c.jpg) +(a) Sample input image +RMSE $= 0.035$ +(e) Recovered illumination +Figure 8. Results on the captured salt004 scene from Stanford-ORB [18]. Note that the "warped true illumination" environment map (g) was captured by a light probe that was not co-located with the object, any may therefore be significantly warped. The recovered RMSE reported is computed with respect to the average of all environment maps provided with the dataset rotated the same coordinate frame. + +![](images/ed7c83d0bb66b90b9e6f018f25bea90ce53965ff93f2932e63c4d3a1266e8e38.jpg) +(b) Recovered albedo +(c) Recovered albedo (no occluder) +RMSE = 0.087 +(f) Recovered illumination (no occluder) + +![](images/55c65d68bc598be062d0925de02489462bc13d9af434d32d5a6cb2b0786ce13f.jpg) +(d) True albedo +(g) Warped true illumination + +where $\tilde{I}_t^{(1)}(\mathbf{u})$ and $\tilde{I}_t^{(2)}(\mathbf{u})$ are the two independently-rendered pixel values. + +# 5. Experiments + +We implement our entire rendering and optimization pipeline in JAX [8] and run each of our experiments on 8 NVIDIA V100 GPUs. We evaluate our method on two sources of data: synthetic objects, and captured data from the Stanford-ORB dataset [18]. Our synthetic results were generated using 256 input images of size $400 \times 400$ , although our method works similarly well with as few as 32 captures (see supplement for the effect of the number of images). Unless noted otherwise, the occluders for all experiments are spherical caps with a size of $0.1 \cdot 4\pi$ steradians, centered at the camera. + +Estimating illumination, materials, and occluders. Our method can effectively recover environment illumination, spatially-varying material parameters, and the shapes of unseen lighting occluders. We first investigate our algorithm's performance using rendered images of a variety of diffuse objects (roughness 0.6, see BRDF model in appendix) with known geometry. The illumination and albedos recovered by our model are visualized with their true values in Figures 1 and 4. Figure 6 shows our reconstruction for purely Lambertian materials, compared with the reconstruction from the same number of images without unobserved occluders (but with self-occlusions), similar to the scenario + +investigated by Swedish et al. [37]. For Lambertian objects, reconstruction is severely ill-conditioned even when self-occlusions are present, which results in significantly blurrier reconstructions. Figure 5 validates that our recovered occluder shapes are accurate, especially for occluders that block more of the illumination energy. + +Captured data. We apply our method to scenes from the Stanford-ORB dataset [18]. Each object in the dataset is placed on a platform and captured from multiple directions. As the camera rig and photographer move around the scene, the photographer is hidden under white cloth below the camera. The images are provided along with a scanned mesh which we use for geometry. + +Despite not being designed for our task, Figure 8 shows our recovered environment map obtained using the 66 training views in the salt004 scene, which features a rough cylindrical object. Note that the dataset only provides an estimated illumination obtained by a light probe placed at a large distance from the object, meaning there is a significant unknown non-rigid warp between the true incident light and the image in Figure 8g labeled "warped true illumination". See supplement for additional results. + +Without known geometry. Figure 7 presents preliminary evidence that our method can be used to recover lighting and materials even when the object geometry is not known. In this experiment, we first recover an estimate of object geometry using a Neural Radiance Field (NeRF) [25]-based + +![](images/e4323f1caa819e4b8bf8fe94aacff0288c6e9c013c450b6c677c615ce52a2961.jpg) +Figure 9. The signal provided by unintended shadows cast by unseen occluders improves the quality of recovered environment illumination. Here, we plot the relative MSE for the recovered environment maps under two scenarios: (blue) using images rendered with unobserved occluders and jointly estimating materials, illumination and occluder shape and (orange) using images rendered without occluders and only estimating materials and illumination (this is the problem setting considered by Swedish et al. [37]). The cue of unintended shadows consistently improves the quality of estimated illumination across varying (a) illumination frequency content, (b) object material roughness, and (c) mask sizes. Additionally, we show that larger occluders that block more light improve the reconstructed illumination quality. + +![](images/021b1307690b9c70ac09ecade42aa419ad190b22284784c0d8a81d651e62e339.jpg) + +![](images/95bf54043295c17c2abccbb0c9ef7880ca1adbfe368d4f47b0cb2a86239c10a5.jpg) + +![](images/42d997c8f4d3eafc5395befc6c63eb65b6cfb185dee5e3c69b07552940e61518.jpg) +(a) Illumination (pyramid) +Figure 10. Results from the experiment in Figure 4, where the pyramid environment map has been replaced with a direct optimization of its (pre-exponentiated) values. Directly optimizing the mask values instead of using spherical harmonics also results in worse reconstructions (see Table 1). + +![](images/4a4d7c374adae22a074b2b3851f6cd29d8d19f06d2d9c83b39cfe9bcec541ba6.jpg) +(b) Illumination (direct) + +model augmented using the orientation loss from Verbin et al. [41], and then use this (potentially-imprecise) proxy geometry in our inverse rendering optimization pipeline. Please refer to the supplemental materials for a detailed description of our NeRF-based model. + +Object self-occlusion does not contain enough signal. Figure 9 quantitatively demonstrates that only using the signal provided by object self-occlusions ("without occluders" plot in orange) cannot recover illumination as accurately as our method, which leverages the cue of unintended shadows ("with occluders" plot in blue). In particular, exploiting unintended shadows is increasingly important as the illumination contains higher frequencies (Figure 9a) and as the object material becomes increasingly diffuse (Figure 9b). + +Larger occluders improve illumination recovery. Figure 9c shows that increasing the size of the unseen occluders improves the accuracy of the recovered environment maps. Increasing the occluder size effectively improves the problem's conditioning by creating a larger variation in the incident lighting across points on the object. + +
chairmannequinpotatoestoad
Direct spherical harmonics0.8310.6540.9970.943
Direct environment map0.0420.0710.0150.073
Ours0.0380.0350.0110.046
+ +Table 1. Errors (RMSE) on all four of our scenes for our full model compared with our model without the spherical harmonic occluder parameterization, and without the pyramid environment parameterization. While the pyramid representation of the environment map improves smoothness (see Figure 10), we find that representing the masks using spherical harmonics is critical for our method. + +Parameterization ablation study. Table 1 and Figure 10 show that representing the environment map as an image pyramid and the image masks with spherical harmonics, both perform better than simply optimizing their respective element values directly. See supplement for more details. + +# 6. Discussion + +Even though unintended shadows are common in most real-world capture scenarios, they are often treated as outlier data. In this paper, we showed how to explicitly leverage these shadows as a signal to improve the quality of recovered lighting and materials in particularly challenging scenarios, such as objects made of diffuse materials. Furthermore, we have shown that our algorithm can be used with approximate geometry reconstructed by view synthesis techniques, and that it can produce imperfect, yet promising results on real captures, despite not being designed for this type of data which does not satisfy many of our model's assumptions. We believe that this work makes important first steps towards a general inverse rendering algorithm that can recover geometry, materials, and illuminations from images captured under realistic conditions. + +Acknowledgements. Our sincere thanks to Delio Vicini, Rick Szeliski, Aleksander Holski, and Guy Satat for helpful discussions and suggestions, and to David Salesin, Janne Kontkanen, and Lior Yariv for their help in improving the manuscript before submission. + +# References + +[1] Sai Bangaru, Michael Gharbi, Tzu-Mao Li, Fujun Luan, Kalyan Sunkavalli, Milos Hasan, Sai Bi, Zexiang Xu, Gilbert Bernstein, and Fredo Durand. Differentiable rendering of neural SDFs through reparameterization. SIGGRAPH Asia, 2022. 2 +[2] Manel Baradad, Vickie Ye, Adam B. Yedidia, Frédo Durand, William T. Freeman, Gregory W. Wornell, and Antonio Torralba. Inferring light fields from shadows. CVPR, 2018. 3, 4 +[3] Jonathan T. Barron and Jitendra Malik. Shape, illumination, and reflectance from shading. TPAMI, 2015. 2, 6 +[4] Sai Bi, Zexiang Xu, Pratul P. Srinivasan, Ben Mildenhall, Kalyan Sunkavalli, Milos Hasan, Yannick Hold-Geoffroy, David Kriegman, and Ravi Ramamoorthi. Neural reflectance fields for appearance acquisition. arXiv, 2020. 2 +[5] Sai Bi, Zexiang Xu, Kalyan Sunkavalli, Milovs Havsan, Yannick Hold-Geoffroy, David J. Kriegman, and Ravi Ramamoorthi. Deep Reflectance Volumes: Relightable Reconstructions from Multi-View Photometric Images. ECCV, 2020. 2 +[6] Mark Boss, Raphael Braun, Varun Jampani, Jonathan T. Barron, Ce Liu, and Hendrik P. A. Lensch. NeRD: Neural reflectance decomposition from image collections. ICCV, 2021. 2 +[7] Katherine L. Bouman, Vickie Ye, Adam B. Yedidia, Frédo Durand, Gregory W. Wornell, Antonio Torralba, and William T. Freeman. Turning corners into cameras: Principles and methods. ICCV, 2017. 3 +[8] James Bradbury, Roy Frostig, Peter Hawkins, Matthew James Johnson, Chris Leary, Dougal Maclaurin, George Necula, Adam Paszke, Jake VanderPlas, Skye Wanderman-Milne, and Qiao Zhang. JAX: composable transformations of Python+NumPy programs, 2018. http://github.com/google/jax.7 +[9] Paul Debevec. Rendering synthetic objects into real scenes: Bridging traditional and image-based graphics with global illumination and high dynamic range photography. SIGGRAPH, 1998. 2 +[10] Xi Deng, Fujun Luan, Bruce Walter, Kavita Bala, and Steve Marschner. Reconstructing Translucent Objects using Differentiable Rendering. SIGGRAPH, 2022. 6 +[11] James R Driscoll and Dennis M Healy. Computing Fourier Transforms and Convolutions on the 2-Sphere. Advances in applied mathematics, 1994. 5 +[12] Marc-Andre Gardner, Kalyan Sunkavalli, Ersin Yumer, Xiaohui Shen, Emiliano Gambaretto, Christian Gagné, and Jean-François Lalonde. Learning to predict indoor illumination from a single image. ACM Trans. Graph., 2017. 2 +[13] Yannick Hold-Geoffroy, Kalyan Sunkavalli, Sunil Hadap, Emiliano Gambaretto, and Jean-François Lalonde. Deep outdoor illumination estimation. CVPR, 2017. 2 +[14] Wenzel Jakob, Sébastien Speierer, Nicolas Roussel, and Delio Vicini. Dr.Jit: A Just-In-Time Compiler for Differentiable Rendering. SIGGRAPH, 2022. 2 +[15] Brian Karis and Epic Games. Real shading in unreal engine 4. Proc. Physically Based Shading Theory Practice, 2013. 1 + +[16] P. Diederik Kingma and Lei Jimmy Ba. Adam: A Method for Stochastic Optimization. *ICLR*, 2015. 6, 7 +[17] Zhengfei Kuang, Kyle Olszewski, Menglei Chai, Zeng Huang, Panos Achlioptas, and Sergey Tulyakov. Neroic: Neural rendering of objects from online image collections. ACM Trans. Graph., 2022. 2 +[18] Zhengfei Kuang, Yunzhi Zhang, Hong-Xing Yu, Samir Agarwala, Shangzhe Wu, and Jiajun Wu. Stanford-ORB: A realworld 3d object inverse rendering benchmark. NeurIPS, 2023. 7, 1, 2, 6 +[19] Jean-François Lalonde, Alexei A Efros, and Srinivasa G Narasimhan. Estimating the natural illumination conditions from a single outdoor image. IJCV, 2012. 2 +[20] Chloe LeGendre, Wan-Chun Ma, Graham Fyffe, John Flynn, Laurent Charbonnel, Jay Busch, and Paul Debevec. Deeplight: Learning illumination for unconstrained mobile mixed reality. CVPR, 2019. 2 +[21] Chloe LeGendre, Wan-Chun Ma, Rohit Pandey, Sean Fanello, Christoph Rhemann, Jason Dourgarian, Jay Busch, and Paul Debevec. Learning illumination from diverse portraits. SIGGRAPH Asia 2020 Technical Communications, 2020. 2 +[22] Tzu-Mao Li, Miika Aittala, Frédo Durand, and Jaakko Lehtinen. Differentiable monte carlo ray tracing through edge sampling. SIGGRAPH Asia, 2018. 2 +[23] Zhengqin Li, Zexiang Xu, Ravi Ramamoorthi, Kalyan Sunkavalli, and Manmohan Chandraker. Learning to reconstruct shape and spatially-varying reflectance from a single image. SIGGRAPH Asia, 2018. 2 +[24] Zhengqin Li, Mohammad Shafiei, Ravi Ramamoorthi, Kalyan Sunkavalli, and Manmohan Chandraker. Inverse rendering for complex indoor scenes: Shape, spatially-varying lighting and svbrdf from a single image. CVPR, 2020. 2 +[25] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. NeRF: Representing Scenes as Neural Radiance Fields for View Synthesis. ECCV, 2020. 7, 8 +[26] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. SIGGRAPH, 2022. 8 +[27] Jacob Munkberg, Jon Hasselgren, Tianchang Shen, Jun Gao, Wenzheng Chen, Alex Evans, Thomas Mueller, and Sanja Fidler. Extracting Triangular 3D Models, Materials, and Lighting From Images. CVPR, 2022. 2 +[28] Merlin Nimier-David, Delio Vicini, Tizian Zeltner, and Wenzel Jakob. Mitsuba 2: A retargetable forward and inverse renderer. SIGGRAPH Asia, 2019. 2 +[29] Merlin Nimier-David, Sébastien Speierer, Benoit Ruiz, and Wenzel Jakob. Radiative Backpropagation: An Adjoint Method for Lightning-Fast Differentiable Rendering. SIGGRAPH, 2020. 2 +[30] Ko Nishino and Shree K Nayar. Eyes for relighting. ACM Trans. Graph., 2004. 2 +[31] Jeong Joon Park, Aleksander Holynski, and Steven M Seitz. Seeing the world in a bag of chips. CVPR, 2020. 1, 3 +[32] Matt Pharr, Wenzel Jakob, and Greg Humphreys. Physically Based Rendering: From Theory to Implementation (3rd + +ed.). Morgan Kaufmann Publishers Inc., San Francisco, CA, USA, 3rd edition, 2016. 1 +[33] Ravi Ramamoorthi and Pat Hanrahan. A signal-processing framework for inverse rendering. SIGGRAPH, 2001. 2, 5 +[34] Imari Sato, Yoichi Sato, and Katsushi Ikeuchi. Illumination from shadows. TPAMI, 2003. 3 +[35] Shuran Song and Thomas Funkhouser. Neural illumination: Lighting prediction for indoor environments. CVPR, 2019. 2 +[36] Pratul P. Srinivasan, Boyang Deng, Xiuming Zhang, Matthew Tancik, Ben Mildenhall, and Jonathan T. Barron. NeRV: Neural reflectance and visibility fields for relighting and view synthesis. CVPR, 2021. 2 +[37] Tristan Swedish, Connor Henley, and Ramesh Raskar. Objects as cameras: Estimating high-frequency illumination from shadows. ICCV, 2021. 1, 3, 7, 8 +[38] Matthew Tancik, Pratul Srinivasan, Ben Mildenhall, Sara Fridovich-Keil, Nithin Raghavan, Utkarsh Singhal, Ravi Ramamoorthi, Jonathan Barron, and Ren Ng. Fourier features let networks learn high frequency functions in low dimensional domains. NeurIPS, 2020. 6 +[39] Antonio Torralba and William T. Freeman. Accidental pin-hole and pinspeck cameras: Revealing the scene outside the picture. CVPR, 2012. 3 +[40] Eric Veach and Leonidas J Guibas. Optimally Combining Sampling Techniques for Monte Carlo Rendering. SIGGRAPH, 1995. 6, 4 +[41] Dor Verbin, Peter Hedman, Ben Mildenhall, Todd Zickler, Jonathan T. Barron, and Pratul P. Srinivasan. Ref-NeRF: Structured View-Dependent Appearance for Neural Radiance Fields. CVPR, 2022. 8 +[42] Delio Vicini, Sébastien Speierer, and Wenzel Jakob. Differentiable Signed Distance Function Rendering. SIGGRAPH, 2022. 2 +[43] Bruce Walter, Stephen R Marschner, Hongsong Li, and Kenneth E Torrance. Microfacet models for refraction through rough surfaces. Eurographics, 2007. 6, 1 +[44] Adam B Yedidia, Manel Baradad, Christos Thrampoulidis, William T Freeman, and Gregory W Wornell. Using unknown occluders to recover hidden scenes. CVPR, 2019. 3 +[45] Tizian Zeltner, Sébastien Speierer, Iliyan Georgiev, and Wenzel Jakob. Monte Carlo Estimators for Differential Light Transport. SIGGRAPH, 2021. 2, 4 +[46] Kai Zhang, Fujun Luan, Qianqian Wang, Kavita Bala, and Noah Snively. PhySG: Inverse rendering with spherical gaussians for physics-based material editing and relighting. CVPR, 2021. 2 +[47] Xiuming Zhang, Pratul P. Srinivasan, Boyang Deng, Paul Debevec, William T. Freeman, and Jonathan T. Barron. NeR-Factor: Neural factorization of shape and reflectance under an unknown illumination. SIGGRAPH Asia, 2021. 2 \ No newline at end of file diff --git a/2024/Eclipse_ Disambiguating Illumination and Materials using Unintended Shadows/images.zip b/2024/Eclipse_ Disambiguating Illumination and Materials using Unintended Shadows/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..fc9b9dc87f0df95478316bbafe5ae20e43cac472 --- /dev/null +++ b/2024/Eclipse_ Disambiguating Illumination and Materials using Unintended Shadows/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:782f824230e84b174c1aa03c8b3173df9c649d586ac19007c0fcb4d884d1b967 +size 464685 diff --git a/2024/Eclipse_ Disambiguating Illumination and Materials using Unintended Shadows/layout.json b/2024/Eclipse_ Disambiguating Illumination and Materials using Unintended Shadows/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..b43ae79bc8b95dcc4d4802e492dd83108fb9f95f --- /dev/null +++ b/2024/Eclipse_ Disambiguating Illumination and Materials using Unintended Shadows/layout.json @@ -0,0 +1,10457 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 51, + 103, + 543, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 51, + 103, + 543, + 121 + ], + "spans": [ + { + "bbox": [ + 51, + 103, + 543, + 121 + ], + "type": "text", + "content": "Eclipse: Disambiguating Illumination and Materials using Unintended Shadows" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 175, + 143, + 235, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 175, + 143, + 235, + 156 + ], + "spans": [ + { + "bbox": [ + 175, + 143, + 235, + 156 + ], + "type": "text", + "content": "Dor Verbin" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 249, + 143, + 329, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 249, + 143, + 329, + 156 + ], + "spans": [ + { + "bbox": [ + 249, + 143, + 329, + 156 + ], + "type": "text", + "content": "Ben Mildenhall1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 343, + 143, + 417, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 343, + 143, + 417, + 156 + ], + "spans": [ + { + "bbox": [ + 343, + 143, + 417, + 156 + ], + "type": "text", + "content": "Peter Hedman" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 147, + 157, + 244, + 170 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 147, + 157, + 244, + 170 + ], + "spans": [ + { + "bbox": [ + 147, + 157, + 244, + 170 + ], + "type": "text", + "content": "Jonathan T. Barron1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 258, + 157, + 334, + 170 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 258, + 157, + 334, + 170 + ], + "spans": [ + { + "bbox": [ + 258, + 157, + 334, + 170 + ], + "type": "text", + "content": "Todd Zickler" + }, + { + "bbox": [ + 258, + 157, + 334, + 170 + ], + "type": "inline_equation", + "content": "^{1,2}" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 345, + 157, + 444, + 170 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 345, + 157, + 444, + 170 + ], + "spans": [ + { + "bbox": [ + 345, + 157, + 444, + 170 + ], + "type": "text", + "content": "Pratul P. Srinivasan1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 196, + 171, + 283, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 196, + 171, + 283, + 186 + ], + "spans": [ + { + "bbox": [ + 196, + 171, + 283, + 186 + ], + "type": "text", + "content": "1Google Research" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 297, + 171, + 395, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 171, + 395, + 186 + ], + "spans": [ + { + "bbox": [ + 297, + 171, + 395, + 186 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 297, + 171, + 395, + 186 + ], + "type": "text", + "content": "Harvard University" + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 48, + 250, + 168, + 349 + ], + "blocks": [ + { + "bbox": [ + 48, + 250, + 168, + 349 + ], + "lines": [ + { + "bbox": [ + 48, + 250, + 168, + 349 + ], + "spans": [ + { + "bbox": [ + 48, + 250, + 168, + 349 + ], + "type": "image", + "image_path": "fa3830b6da4b09f6b20011f65294f8112da2ea60f58383464472f4f750b6e842.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 80, + 352, + 143, + 363 + ], + "lines": [ + { + "bbox": [ + 80, + 352, + 143, + 363 + ], + "spans": [ + { + "bbox": [ + 80, + 352, + 143, + 363 + ], + "type": "text", + "content": "(a) Problem setup" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 192, + 242, + 210, + 264 + ], + "blocks": [ + { + "bbox": [ + 192, + 242, + 210, + 264 + ], + "lines": [ + { + "bbox": [ + 192, + 242, + 210, + 264 + ], + "spans": [ + { + "bbox": [ + 192, + 242, + 210, + 264 + ], + "type": "image", + "image_path": "20bd31f14abca80306765369e4d607fc075573d7a45bbcc446da76ae2ed5df02.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 227, + 220, + 358, + 291 + ], + "blocks": [ + { + "bbox": [ + 187, + 207, + 214, + 216 + ], + "lines": [ + { + "bbox": [ + 187, + 207, + 214, + 216 + ], + "spans": [ + { + "bbox": [ + 187, + 207, + 214, + 216 + ], + "type": "text", + "content": "Albedo" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 239, + 207, + 334, + 216 + ], + "lines": [ + { + "bbox": [ + 239, + 207, + 334, + 216 + ], + "spans": [ + { + "bbox": [ + 239, + 207, + 334, + 216 + ], + "type": "text", + "content": "Illumination and occluders" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 227, + 220, + 358, + 291 + ], + "lines": [ + { + "bbox": [ + 227, + 220, + 358, + 291 + ], + "spans": [ + { + "bbox": [ + 227, + 220, + 358, + 291 + ], + "type": "image", + "image_path": "070bd0bd8dc3f3b994e92d68b7875826c86978889e196dd1a21eef26bac1e987.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 365, + 221, + 404, + 281 + ], + "blocks": [ + { + "bbox": [ + 373, + 207, + 399, + 216 + ], + "lines": [ + { + "bbox": [ + 373, + 207, + 399, + 216 + ], + "spans": [ + { + "bbox": [ + 373, + 207, + 399, + 216 + ], + "type": "text", + "content": "Albedo" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 365, + 221, + 404, + 281 + ], + "lines": [ + { + "bbox": [ + 365, + 221, + 404, + 281 + ], + "spans": [ + { + "bbox": [ + 365, + 221, + 404, + 281 + ], + "type": "image", + "image_path": "a5310ae4b2c78c42d0dc5e88c9801c49f93d0dad67194147c0171334cd72522b.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 410, + 220, + 541, + 291 + ], + "blocks": [ + { + "bbox": [ + 424, + 207, + 518, + 216 + ], + "lines": [ + { + "bbox": [ + 424, + 207, + 518, + 216 + ], + "spans": [ + { + "bbox": [ + 424, + 207, + 518, + 216 + ], + "type": "text", + "content": "Illumination and occluders" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 410, + 220, + 541, + 291 + ], + "lines": [ + { + "bbox": [ + 410, + 220, + 541, + 291 + ], + "spans": [ + { + "bbox": [ + 410, + 220, + 541, + 291 + ], + "type": "image", + "image_path": "1e2c5fbe3d6b8eb10743d8fc365297a366652b3164fecd91f5ab0640a4c46f8b.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 181, + 300, + 220, + 358 + ], + "blocks": [ + { + "bbox": [ + 181, + 300, + 220, + 358 + ], + "lines": [ + { + "bbox": [ + 181, + 300, + 220, + 358 + ], + "spans": [ + { + "bbox": [ + 181, + 300, + 220, + 358 + ], + "type": "image", + "image_path": "7e50a000c9c96cee488df38da1c6d47e2e25346c892d823d5eb7aacd2af78c93.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 226, + 299, + 358, + 370 + ], + "blocks": [ + { + "bbox": [ + 226, + 299, + 358, + 370 + ], + "lines": [ + { + "bbox": [ + 226, + 299, + 358, + 370 + ], + "spans": [ + { + "bbox": [ + 226, + 299, + 358, + 370 + ], + "type": "image", + "image_path": "39f08ffd84ba2f6ae864755c702a4163eb3829faeceeb6f04f312931c2c2dc15.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 45, + 376, + 546, + 431 + ], + "lines": [ + { + "bbox": [ + 45, + 376, + 546, + 431 + ], + "spans": [ + { + "bbox": [ + 45, + 376, + 546, + 431 + ], + "type": "text", + "content": "Figure 1. (a) We exploit unintended shadows cast by camera operators (or other unobserved moving occluders) to recover high-fidelity environment lighting and object materials from a set of images. Without modeling such unobserved occluders, prior methods can only: (b) recover convolved lighting without explicit material decomposition [31]; or (c) exploit lighting occlusions that occur internally among the observed objects [37]. (d) We show that additionally modeling and recovering external, unobserved occluders enables lighting and material reconstructions that are closer to ground truth (e)." + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 365, + 299, + 405, + 358 + ], + "blocks": [ + { + "bbox": [ + 365, + 299, + 405, + 358 + ], + "lines": [ + { + "bbox": [ + 365, + 299, + 405, + 358 + ], + "spans": [ + { + "bbox": [ + 365, + 299, + 405, + 358 + ], + "type": "image", + "image_path": "2ad95e1e242aebb2bdc0886673b5558f8b5e651de1490597bb48f0d0e5f3c046.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 410, + 298, + 541, + 370 + ], + "blocks": [ + { + "bbox": [ + 410, + 298, + 541, + 370 + ], + "lines": [ + { + "bbox": [ + 410, + 298, + 541, + 370 + ], + "spans": [ + { + "bbox": [ + 410, + 298, + 541, + 370 + ], + "type": "image", + "image_path": "692bf59ab3edbda586f02177204c85bfd6f3b277a773600180d0f64915325b43.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + } + ], + "index": 24 + }, + { + "bbox": [ + 143, + 442, + 191, + 453 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 442, + 191, + 453 + ], + "spans": [ + { + "bbox": [ + 143, + 442, + 191, + 453 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 46, + 460, + 288, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 460, + 288, + 700 + ], + "spans": [ + { + "bbox": [ + 46, + 460, + 288, + 700 + ], + "type": "text", + "content": "Decomposing an object's appearance into representations of its materials and the surrounding illumination is difficult, even when the object's 3D shape is known beforehand. This problem is especially challenging for diffuse objects: it is ill-conditioned because diffuse materials severely blur incoming light, and it is ill-posed because diffuse materials under high-frequency lighting can be indistinguishable from shiny materials under low-frequency lighting. We show that it is possible to recover precise materials and illumination—even from diffuse objects—by exploiting unintended shadows, like the ones cast onto an object by the photographer who moves around it. These shadows are a nuisance in most previous inverse rendering pipelines, but here we exploit them as signals that improve conditioning and help resolve material-lighting ambiguities. We present a method based on differentiable Monte Carlo ray tracing that uses images of an object to jointly recover its spatially-varying materials, the surrounding illumination environment, and the shapes of the unseen light occluders who inadvertently cast shadows upon it." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 442, + 386, + 453 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 442, + 386, + 453 + ], + "spans": [ + { + "bbox": [ + 307, + 442, + 386, + 453 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 304, + 462, + 545, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 462, + 545, + 545 + ], + "spans": [ + { + "bbox": [ + 304, + 462, + 545, + 545 + ], + "type": "text", + "content": "In this work, we show that the long-standing inverse rendering problem of recovering an object's material properties and the surrounding illumination environment from a set of images can greatly benefit from considering the effects of unobserved moving light occluders, such as the camera operator, which partially block incoming light and inadvertently cast shadows onto the object being imaged." + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 304, + 546, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 546, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 546, + 546, + 713 + ], + "type": "text", + "content": "Joint recovery of materials and lighting is a challenging task because the BRDF (bidirectional reflectance distribution function, which represents how a material at any point on an object's surface maps from incoming to outgoing light) acts as a directional filter on incoming light. BRDFs for specular (shiny) materials act as all-pass directional filters, while BRDFs for diffuse materials act as low-pass directional filters. As a result, inverse rendering is fundamentally ambiguous since a shiny object illuminated by blurry lighting can be indistinguishable from a diffuse object illuminated by sharp lighting (Figure 2). Furthermore, since diffuse materials act as low-pass filters and strongly blur incoming light, even the simpler problem of recovering lighting from images of a known diffuse material can be" + } + ] + } + ], + "index": 30 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 300, + 749, + 310, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 300, + 749, + 310, + 757 + ], + "spans": [ + { + "bbox": [ + 300, + 749, + 310, + 757 + ], + "type": "text", + "content": "77" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 286, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 286, + 96 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 286, + 96 + ], + "type": "text", + "content": "severely ill-conditioned, precluding the recovery of high-frequency illumination [33]." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 97, + 286, + 264 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 97, + 286, + 264 + ], + "spans": [ + { + "bbox": [ + 46, + 97, + 286, + 264 + ], + "type": "text", + "content": "We observe that both of these issues can be ameliorated by exploiting an effect that naturally exists whenever an object is imaged from a sequence of viewpoints under static environment lighting. Namely, between one image and the next, the positions of the camera and its operator(s) must change, and in doing so they occlude different portions of the surrounding environment, casting distinct sets of shadows onto the object being imaged. This effect is very noticeable under strongly-directional lighting or for shiny objects because the cast shadows are sharp or the specular highlights are misshapen and hidden. But even when the effect is barely perceptible, such as for diffuse objects under well-distributed lighting, there exists a subtle signal that one can take advantage of." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 266, + 286, + 494 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 266, + 286, + 494 + ], + "spans": [ + { + "bbox": [ + 46, + 266, + 286, + 494 + ], + "type": "text", + "content": "In this paper, we make first steps toward a practical algorithm that exploits this cue by using gradient descent with differentiable Monte Carlo ray tracing to jointly recover explicit representations of (i) spatially-varying reflectance, (ii) environment illumination, and (iii) the per-image shapes of any unseen, light-blocking occluders. We evaluate our algorithm using a challenging variety of simulated scenes, including scenes with diffuse-only materials and with object geometry that is not known beforehand. We also apply our method to an existing off-the-shelf dataset designed for inverse rendering, where the (unseen) camera rig moving to capture the scene acts as an accidental occluder. Our results demonstrate that external shadowing effects provide a strong and useful cue, even when they are subtle, and even without relying on strong domain-specific priors for the materials, illumination, or external occluder shapes. This suggests that incorporating unintended shadows into inverse rendering pipelines for real, captured data is valuable for improving the quality of material and lighting assets." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 507, + 133, + 519 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 507, + 133, + 519 + ], + "spans": [ + { + "bbox": [ + 47, + 507, + 133, + 519 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 529, + 286, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 529, + 286, + 601 + ], + "spans": [ + { + "bbox": [ + 47, + 529, + 286, + 601 + ], + "type": "text", + "content": "We build on recent developments in differentiable ray tracing and on a long history of inverse rendering work, including the joint estimation of reflectance and lighting, and the estimation of lighting alone. We are also inspired by a separate line of work on passive non-line-of-sight imaging, which exploits similar light-blocking effects." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 605, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 605, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 605, + 286, + 713 + ], + "type": "text", + "content": "Physics-based differentiable ray tracing. Modern differentiable ray tracers [22, 28] enable the computation of gradients of rendered images with respect to scene parameters (i.e., geometry, materials, and lighting) by differentiating through light transport simulation. Recent works have focused on improving efficiency and performance [14, 29] and on accurately differentiating through visibility discontinuities—such as those caused by shadow and occlusion boundaries—with respect to shape and light-" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 305, + 72, + 545, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 120 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 120 + ], + "type": "text", + "content": "ing [1, 42]. Our implementation of differentiable ray tracing leverages insights from these works as well as from Zeltner et al. [45], who provide design intuition for Monte Carlo differentiable ray tracers." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 127, + 545, + 294 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 127, + 545, + 294 + ], + "spans": [ + { + "bbox": [ + 304, + 127, + 545, + 294 + ], + "type": "text", + "content": "Inverse rendering of materials and lighting. Decomposing an object's appearance into representations of material and lighting is a long-standing problem in computer vision and graphics. In their foundational work, Ramamoorthi and Hanrahan [33] developed a signal processing approach by describing the outgoing light at a surface point as the spherical convolution of the BRDF and the incoming lighting. This formulation elucidates why inverse rendering is ill-posed and often ill-conditioned: it is ill-posed because there are multiple illumination-material pairs that convolve into the same image, and it is ill-conditioned because diffuse BRDFs act as a low-pass filters on lighting, which causes the estimates of medium- and high-frequency illumination to be very sensitive to noise." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 297, + 545, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 297, + 545, + 441 + ], + "spans": [ + { + "bbox": [ + 304, + 297, + 545, + 441 + ], + "type": "text", + "content": "Because of this, most approaches to inverse rendering rely on strong priors on materials and lighting. Single-view approaches have used hand-designed priors [3] or priors in the form of neural networks trained with supervision from large datasets of material and lighting labels [23, 24]. More relevant to us are multi-view approaches, which typically either assume known lighting [4, 5, 36] or rely on strong priors, such as assuming a single, highly-specular BRDF for the entire object [46] or lighting from a prior distribution that was pre-trained on a dataset of environment maps [6, 47]. The most closely related work is that for which lighting information is not provided as input [17, 27]." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 443, + 545, + 526 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 443, + 545, + 526 + ], + "spans": [ + { + "bbox": [ + 304, + 443, + 545, + 526 + ], + "type": "text", + "content": "For the sake of generality, we do not use such strong priors in our experiments because by avoiding them we can more directly measure the benefits that are gained by modeling unintended shadows as an additional cue. Because our model uses a generic gradient-descent framework, stronger application-specific priors can be added to it to improve performance on a specific domain of interest." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 534, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 534, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 534, + 545, + 713 + ], + "type": "text", + "content": "Estimating environment lighting. Another thread of inverse rendering research focuses on the task of recovering high-fidelity lighting environments, typically for augmented reality applications where virtual objects are rendered into photographs with consistent reflections and shadows. Prior work by Debevec [9] used images of a chrome sphere to measure environment lighting directly, and subsequent works have demonstrated that plausible environment lighting can be estimated without chrome spheres, using images of indoor [12, 35] and outdoor scenes [13, 19, 20], or images of a specific class of objects like faces [21, 30]. These subsequent techniques use strong priors in the form of deep neural network weights that are trained with supervision to map from images to lighting, and they do not enforce physical rendering consistency between the recovered" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 300, + 749, + 310, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 300, + 749, + 310, + 757 + ], + "spans": [ + { + "bbox": [ + 300, + 749, + 310, + 757 + ], + "type": "text", + "content": "78" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 96 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 96 + ], + "type": "text", + "content": "lighting and the observations. In contrast, we avoid such strong priors, and we explicitly enforce consistency." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 97, + 286, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 97, + 286, + 180 + ], + "spans": [ + { + "bbox": [ + 46, + 97, + 286, + 180 + ], + "type": "text", + "content": "More related to our approach is the work of Park et al. [31], which recovers physically-consistent environment maps from RGBD videos of shiny objects. However, their algorithm can only recover the convolution of the environment map with the BRDF of the observed object (e.g., Figure 1(b)), which precludes the recovery of high illumination frequencies unless the object is highly specular." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 182, + 286, + 337 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 182, + 286, + 337 + ], + "spans": [ + { + "bbox": [ + 46, + 182, + 286, + 337 + ], + "type": "text", + "content": "Also related is the work of Swedish et al. [37], which places a known, diffuse object on a ground plane (a scenario first studied by Sato et al. [34]) and uses its cast shadows to recover high-quality lighting. Their formulation is linear and it improves lighting estimates by leveraging self-shadowing among diffuse objects that have known geometry and albedo (e.g., Figure 1(c)). We generalize this by replacing the linear formulation with differentiable Monte Carlo ray tracing, which allows exploiting the additional shadowing effects caused by moving external occluders and leads to substantially improved results (e.g., Figure 1(d)). Our formulation also handles objects with more general spatially-varying BRDFs that are not known beforehand." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 342, + 286, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 342, + 286, + 545 + ], + "spans": [ + { + "bbox": [ + 46, + 342, + 286, + 545 + ], + "type": "text", + "content": "Passive non-line-of-sight imaging with occlusions. Passive non-line-of-sight techniques can also be seen as recovering environment illumination: They observe a reflective surface (which is typically diffuse and planar) and recover the appearance of a \"hidden scene\" from these reflections. Similar to us, prior works in this area have observed that the presence of an occluder between the observed surface and the hidden scene/environment aids recovery by introducing sharp angular variations into the rendering integral. This insight was first leveraged in settings with simple occluder-shapes like pinholes and pinspecks [39] or corners joining walls [7]. Subsequent work by Baradad et al. [2] generalizes this by considering light-occlusion effects caused by an arbitrary but known 3D shape (a task that is closely related to Swedish et al. above). Yedidia et al. [44] additionally recover the unobserved occluder's shape assuming it is a planar mask parallel to the observed planar surface." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 547, + 286, + 643 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 547, + 286, + 643 + ], + "spans": [ + { + "bbox": [ + 46, + 547, + 286, + 643 + ], + "type": "text", + "content": "We generalize these prior works by using differentiable Monte Carlo rendering to replace their deconvolution-based algorithms, which are tailored to the specific case where the observed reflector is planar and diffuse. This allows using reflective objects that have arbitrary shapes, and arbitrary spatially-varying BRDFs which are not known beforehand. It also allows recovering the shapes of unobserved occluders that are arbitrary and time-varying." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 656, + 220, + 670 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 656, + 220, + 670 + ], + "spans": [ + { + "bbox": [ + 47, + 656, + 220, + 670 + ], + "type": "text", + "content": "3. Motivation and Problem Setup" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 677, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 286, + 713 + ], + "type": "text", + "content": "A well-known ambiguity in computer vision and graphics occurs when decomposing an image into its component lighting and materials. The top of Figure 2 recreates" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 306, + 69, + 545, + 245 + ], + "blocks": [ + { + "bbox": [ + 306, + 69, + 545, + 245 + ], + "lines": [ + { + "bbox": [ + 306, + 69, + 545, + 245 + ], + "spans": [ + { + "bbox": [ + 306, + 69, + 545, + 245 + ], + "type": "image", + "image_path": "b046042743bcf125cc9ab4f2f0648209edda8e1e76468ba73c3b13f950e29bbe.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 254, + 545, + 342 + ], + "lines": [ + { + "bbox": [ + 304, + 254, + 545, + 342 + ], + "spans": [ + { + "bbox": [ + 304, + 254, + 545, + 342 + ], + "type": "text", + "content": "Figure 2. Decomposing appearance into lighting and materials can be inherently ambiguous. Top row: a diffuse sphere lit by high-frequency lighting (inset) is visually indistinguishable from a shiny, mirror-like sphere lit by low-frequency lighting. Bottom row: A second image captured in the presence of a dark external occluder resolves the ambiguity. The occluder's effect is clearly visible on the shiny sphere while on the diffuse sphere it produces shadows that are soft and very subtle." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 365, + 545, + 508 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 365, + 545, + 508 + ], + "spans": [ + { + "bbox": [ + 304, + 365, + 545, + 508 + ], + "type": "text", + "content": "a common depiction of this ambiguity, where an image of a known shape (a sphere) is explained equally well by a diffuse material in a complicated lighting environment or a shiny, mirror-like material under low-frequency lighting. Now, imagine we capture a second image from the same viewpoint, after some external object enters the scene. This second object remains beyond the field of view and so is not directly observed, but it acts as an external occluder that prevents some of the environment's light from reaching the sphere. This second image, shown at the bottom of the figure for each case, clearly reveals which of the two materiallighting explanations is the correct one." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 509, + 545, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 509, + 545, + 628 + ], + "spans": [ + { + "bbox": [ + 304, + 509, + 545, + 628 + ], + "type": "text", + "content": "We aim to make use of this unintended shadowing that naturally occurs whenever a camera and its operators move around an object while capturing images from different viewpoints. They affect each image by blocking different portions of the surrounding light, and this provides a helpful signal. We will show that this signal is helpful even when the occluder shapes and locations are quite arbitrary and not known beforehand, and when their shadowing effects are very subtle, like between the top and bottom images on the left of Figure 2." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "content": "In addition to helping resolve material-lighting ambiguities, shadowing from moving external occluders also improves the conditioning of our inverse rendering problem. This is analyzed in Figure 3, which considers the simplified hypothetical 1D case of recovering an environment illumination from images of a diffuse disk-shaped object, with known occluder positions. Convolution with a dif" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 300, + 749, + 310, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 300, + 749, + 310, + 757 + ], + "spans": [ + { + "bbox": [ + 300, + 749, + 310, + 757 + ], + "type": "text", + "content": "79" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 72, + 284, + 232 + ], + "blocks": [ + { + "bbox": [ + 50, + 72, + 284, + 232 + ], + "lines": [ + { + "bbox": [ + 50, + 72, + 284, + 232 + ], + "spans": [ + { + "bbox": [ + 50, + 72, + 284, + 232 + ], + "type": "image", + "image_path": "6ed747df87e5cfd7f61229da52dbbbebf5eb52a1f438457e7547d63cd8e50879.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 241, + 287, + 352 + ], + "lines": [ + { + "bbox": [ + 46, + 241, + 287, + 352 + ], + "spans": [ + { + "bbox": [ + 46, + 241, + 287, + 352 + ], + "type": "text", + "content": "Figure 3. Here we plot eigenvalue magnitudes as a function of illumination angular frequency for a one-dimensional circular Lambertian object with known albedo, under the simplifying assumption of known occluder locations and shapes. For a single input image without occlusion, almost all lighting frequencies are strongly attenuated. In the single-occluder case (similar to the scenario considered by Baradad et al. [2]), some low frequencies become recoverable but intermediate and high frequencies do not. Observing additional images enables the recovery of more lighting information, especially at the highest frequencies." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 373, + 286, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 373, + 286, + 506 + ], + "spans": [ + { + "bbox": [ + 46, + 373, + 286, + 506 + ], + "type": "text", + "content": "fuse BRDF acts as a low-pass filter of the incoming light, which means that the linear inverse problem's eigenvalues are vanishingly low for higher frequencies. Moving external occluders introduce a sharp angular variation in the per-image illumination, which better-conditions these high frequency components, and allows their recovery. As shown in Figure 3, in this simplified scenario a single image containing an occluder is better than a single image without any occluders, but using observations of more occluders makes the problem even better-conditioned. We demonstrate this effect in practice in the supplement." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 515, + 129, + 528 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 515, + 129, + 528 + ], + "spans": [ + { + "bbox": [ + 47, + 515, + 129, + 528 + ], + "type": "text", + "content": "3.1. Formulation" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 534, + 286, + 643 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 534, + 286, + 643 + ], + "spans": [ + { + "bbox": [ + 46, + 534, + 286, + 643 + ], + "type": "text", + "content": "Let " + }, + { + "bbox": [ + 46, + 534, + 286, + 643 + ], + "type": "inline_equation", + "content": "\\{I_t\\}_{t=1}^T" + }, + { + "bbox": [ + 46, + 534, + 286, + 643 + ], + "type": "text", + "content": " be a collection of " + }, + { + "bbox": [ + 46, + 534, + 286, + 643 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 46, + 534, + 286, + 643 + ], + "type": "text", + "content": " images of a scene with known camera poses; for every " + }, + { + "bbox": [ + 46, + 534, + 286, + 643 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 534, + 286, + 643 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 534, + 286, + 643 + ], + "type": "inline_equation", + "content": "I_t" + }, + { + "bbox": [ + 46, + 534, + 286, + 643 + ], + "type": "text", + "content": " is an RGB image with spatial size " + }, + { + "bbox": [ + 46, + 534, + 286, + 643 + ], + "type": "inline_equation", + "content": "H \\times W" + }, + { + "bbox": [ + 46, + 534, + 286, + 643 + ], + "type": "text", + "content": ". We assume the observed scene " + }, + { + "bbox": [ + 46, + 534, + 286, + 643 + ], + "type": "inline_equation", + "content": "\\mathcal{S} \\subset \\mathbb{R}^3" + }, + { + "bbox": [ + 46, + 534, + 286, + 643 + ], + "type": "text", + "content": " is illuminated by a far-field environment map " + }, + { + "bbox": [ + 46, + 534, + 286, + 643 + ], + "type": "inline_equation", + "content": "L(\\hat{\\omega}_i)" + }, + { + "bbox": [ + 46, + 534, + 286, + 643 + ], + "type": "text", + "content": " that does not change over time, so the only temporal changes in the incident light field are caused by unobserved light-blockers (called occluders hereafter) that move around the scene outside the field of view. We can then write the incident light at any observed position " + }, + { + "bbox": [ + 46, + 534, + 286, + 643 + ], + "type": "inline_equation", + "content": "\\mathbf{x} \\in \\mathcal{S}" + }, + { + "bbox": [ + 46, + 534, + 286, + 643 + ], + "type": "text", + "content": " as:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 85, + 654, + 287, + 666 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 85, + 654, + 287, + 666 + ], + "spans": [ + { + "bbox": [ + 85, + 654, + 287, + 666 + ], + "type": "interline_equation", + "content": "L _ {t} (\\mathbf {x}, \\hat {\\omega} _ {i}) = L (\\hat {\\omega} _ {i}) M _ {t} (\\mathbf {x}, \\hat {\\omega} _ {i}) V (\\mathbf {x}, \\hat {\\omega} _ {i}), \\tag {1}", + "image_path": "53b123837d5cd1be1babce9c4891ef8997e12d258162e7875e84ef24329aca01.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 677, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 677, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 677, + 287, + 713 + ], + "type": "text", + "content": "where for any " + }, + { + "bbox": [ + 47, + 677, + 287, + 713 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 47, + 677, + 287, + 713 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 47, + 677, + 287, + 713 + ], + "type": "inline_equation", + "content": "M_{t}(\\mathbf{x}, \\hat{\\boldsymbol{\\omega}}_{i})" + }, + { + "bbox": [ + 47, + 677, + 287, + 713 + ], + "type": "text", + "content": " is a binary signal with value 0 for directions blocked by the occluder and 1 otherwise, and the binary signal " + }, + { + "bbox": [ + 47, + 677, + 287, + 713 + ], + "type": "inline_equation", + "content": "V(\\mathbf{x}, \\hat{\\boldsymbol{\\omega}}_{i})" + }, + { + "bbox": [ + 47, + 677, + 287, + 713 + ], + "type": "text", + "content": " models visibility effects that are" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 72, + 545, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 95 + ], + "type": "text", + "content": "internal to the scene, with value 0 for directions from " + }, + { + "bbox": [ + 305, + 72, + 545, + 95 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 305, + 72, + 545, + 95 + ], + "type": "text", + "content": " that are blocked by other scene elements, and value 1 otherwise." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 97, + 545, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 97, + 545, + 133 + ], + "spans": [ + { + "bbox": [ + 305, + 97, + 545, + 133 + ], + "type": "text", + "content": "Omitting global illumination effects, a pixel " + }, + { + "bbox": [ + 305, + 97, + 545, + 133 + ], + "type": "inline_equation", + "content": "\\mathbf{u}" + }, + { + "bbox": [ + 305, + 97, + 545, + 133 + ], + "type": "text", + "content": " corresponding to a surface point " + }, + { + "bbox": [ + 305, + 97, + 545, + 133 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 305, + 97, + 545, + 133 + ], + "type": "text", + "content": " with BRDF " + }, + { + "bbox": [ + 305, + 97, + 545, + 133 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 305, + 97, + 545, + 133 + ], + "type": "text", + "content": " viewed from direction " + }, + { + "bbox": [ + 305, + 97, + 545, + 133 + ], + "type": "inline_equation", + "content": "\\hat{\\omega}_o" + }, + { + "bbox": [ + 305, + 97, + 545, + 133 + ], + "type": "text", + "content": " at time " + }, + { + "bbox": [ + 305, + 97, + 545, + 133 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 305, + 97, + 545, + 133 + ], + "type": "text", + "content": " has color:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 310, + 142, + 545, + 167 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 310, + 142, + 545, + 167 + ], + "spans": [ + { + "bbox": [ + 310, + 142, + 545, + 167 + ], + "type": "interline_equation", + "content": "I _ {t} (\\mathbf {u}) = \\int_ {\\mathbb {S} ^ {2}} L _ {t} (\\mathbf {x}, \\hat {\\omega} _ {i}) f (\\mathbf {x}, \\hat {\\omega} _ {i}, \\hat {\\omega} _ {o}) (\\hat {\\mathbf {n}} (\\mathbf {x}) \\cdot \\hat {\\boldsymbol {\\omega}} _ {i}) _ {+} d \\hat {\\boldsymbol {\\omega}} _ {i}, \\tag {2}", + "image_path": "73a1f02bce47b66c48c494d793240f54a9de8764b497e2399b2c9ea492a2b7cf.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 178, + 545, + 201 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 178, + 545, + 201 + ], + "spans": [ + { + "bbox": [ + 305, + 178, + 545, + 201 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 178, + 545, + 201 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{n}} (\\mathbf{x})" + }, + { + "bbox": [ + 305, + 178, + 545, + 201 + ], + "type": "text", + "content": " is the surface normal corresponding to " + }, + { + "bbox": [ + 305, + 178, + 545, + 201 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 305, + 178, + 545, + 201 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 305, + 178, + 545, + 201 + ], + "type": "inline_equation", + "content": "(\\cdot)_{+}" + }, + { + "bbox": [ + 305, + 178, + 545, + 201 + ], + "type": "text", + "content": " clamps negative values to zero." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 202, + 545, + 275 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 202, + 545, + 275 + ], + "spans": [ + { + "bbox": [ + 304, + 202, + 545, + 275 + ], + "type": "text", + "content": "Given the set of observed images, we would like to recover the environment map " + }, + { + "bbox": [ + 304, + 202, + 545, + 275 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 304, + 202, + 545, + 275 + ], + "type": "text", + "content": ", a spatially-varying BRDF " + }, + { + "bbox": [ + 304, + 202, + 545, + 275 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 304, + 202, + 545, + 275 + ], + "type": "text", + "content": " at every point on the surface of the object " + }, + { + "bbox": [ + 304, + 202, + 545, + 275 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 304, + 202, + 545, + 275 + ], + "type": "text", + "content": ", as well as the set of unobserved occluder masks " + }, + { + "bbox": [ + 304, + 202, + 545, + 275 + ], + "type": "inline_equation", + "content": "\\{M_t\\}_{t=1}^T" + }, + { + "bbox": [ + 304, + 202, + 545, + 275 + ], + "type": "text", + "content": ", one for each observation time " + }, + { + "bbox": [ + 304, + 202, + 545, + 275 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 202, + 545, + 275 + ], + "type": "text", + "content": ". Our goal is therefore to solve the following optimization problem:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 312, + 284, + 545, + 323 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 284, + 545, + 323 + ], + "spans": [ + { + "bbox": [ + 312, + 284, + 545, + 323 + ], + "type": "interline_equation", + "content": "\\underset {\\phi^ {(o)}, \\phi^ {(m)}, \\phi^ {(\\ell)}} {\\arg \\min } \\sum_ {t, \\mathbf {u}} \\left\\| I _ {t} (\\mathbf {u}) - \\mathcal {R} _ {t} \\left(\\mathbf {u}; \\phi^ {(o)}, \\phi^ {(m)}, \\phi^ {(\\ell)}\\right) \\right\\| ^ {2}, \\tag {3}", + "image_path": "f4a6291ba33e5fcdd3662a1bb757d9241bc23388e08463dd5ebb76b15ec9fdfd.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 324, + 545, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 324, + 545, + 361 + ], + "spans": [ + { + "bbox": [ + 304, + 324, + 545, + 361 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 324, + 545, + 361 + ], + "type": "inline_equation", + "content": "\\mathcal{R}_t(\\mathbf{u};\\cdot)" + }, + { + "bbox": [ + 304, + 324, + 545, + 361 + ], + "type": "text", + "content": " renders the pixel location " + }, + { + "bbox": [ + 304, + 324, + 545, + 361 + ], + "type": "inline_equation", + "content": "\\mathbf{u}" + }, + { + "bbox": [ + 304, + 324, + 545, + 361 + ], + "type": "text", + "content": " at time " + }, + { + "bbox": [ + 304, + 324, + 545, + 361 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 324, + 545, + 361 + ], + "type": "text", + "content": " using the occluder parameters " + }, + { + "bbox": [ + 304, + 324, + 545, + 361 + ], + "type": "inline_equation", + "content": "\\phi^{(o)}" + }, + { + "bbox": [ + 304, + 324, + 545, + 361 + ], + "type": "text", + "content": ", material parameters " + }, + { + "bbox": [ + 304, + 324, + 545, + 361 + ], + "type": "inline_equation", + "content": "\\phi^{(m)}" + }, + { + "bbox": [ + 304, + 324, + 545, + 361 + ], + "type": "text", + "content": ", and illumination parameters " + }, + { + "bbox": [ + 304, + 324, + 545, + 361 + ], + "type": "inline_equation", + "content": "\\phi^{(\\ell)}" + }, + { + "bbox": [ + 304, + 324, + 545, + 361 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 362, + 545, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 362, + 545, + 445 + ], + "spans": [ + { + "bbox": [ + 304, + 362, + 545, + 445 + ], + "type": "text", + "content": "At first glance, it may seem that solving for occluders in addition to materials and illumination should make the resulting inverse problem more difficult. However, as illustrated in Figure 3 and demonstrated by our experimental results, modeling and solving for occluders actually makes the full problem more well-conditioned and thus improves recovery of material and illumination." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 459, + 361, + 471 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 459, + 361, + 471 + ], + "spans": [ + { + "bbox": [ + 306, + 459, + 361, + 471 + ], + "type": "text", + "content": "4. Method" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 479, + 545, + 539 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 479, + 545, + 539 + ], + "spans": [ + { + "bbox": [ + 304, + 479, + 545, + 539 + ], + "type": "text", + "content": "We first describe our parameterizations of the unseen oc-clusters, illumination, and materials (Sections 4.1-4.3). Section 4.4 then relates these parameters to rendered pixel colors, and Section 4.5 describes how we optimize to solve the inverse problem in Equation 3." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 540, + 545, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 540, + 545, + 613 + ], + "spans": [ + { + "bbox": [ + 304, + 540, + 545, + 613 + ], + "type": "text", + "content": "We emphasize that our method is designed to exploit external shadows across as many scene-types as possible, and so its only priors are those implicit to our parameterizations. In our experiments, we verify that these relatively weak priors are sufficient for recovering high-fidelity occluder shapes, illumination maps, and material maps." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 306, + 623, + 375, + 634 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 623, + 375, + 634 + ], + "spans": [ + { + "bbox": [ + 306, + 623, + 375, + 634 + ], + "type": "text", + "content": "4.1.Occluders" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "type": "text", + "content": "Define a world coordinate system with origin at the scene's center, and let " + }, + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\{r_t\\}_{t=1}^T" + }, + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "type": "text", + "content": " be the radial distances from the origin to each (known) camera's center. Then we model each of the " + }, + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "type": "text", + "content": " occluders " + }, + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "type": "inline_equation", + "content": "M_t" + }, + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "type": "text", + "content": " as an independent binary signal defined on the surface of a finite, radius-" + }, + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "type": "inline_equation", + "content": "r_t" + }, + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "type": "text", + "content": " sphere that is centered at the origin. The value of the binary occluder signal" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 299, + 749, + 310, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 749, + 310, + 757 + ], + "spans": [ + { + "bbox": [ + 299, + 749, + 310, + 757 + ], + "type": "text", + "content": "80" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 61, + 86, + 121, + 130 + ], + "blocks": [ + { + "bbox": [ + 61, + 86, + 121, + 130 + ], + "lines": [ + { + "bbox": [ + 61, + 86, + 121, + 130 + ], + "spans": [ + { + "bbox": [ + 61, + 86, + 121, + 130 + ], + "type": "image", + "image_path": "692f6c4c60ad0fde441093d2f689c0ca2f7930225543b1dc458cbf36616af74a.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 134, + 140, + 191, + 149 + ], + "lines": [ + { + "bbox": [ + 134, + 140, + 191, + 149 + ], + "spans": [ + { + "bbox": [ + 134, + 140, + 191, + 149 + ], + "type": "text", + "content": "PSNR = 31.2 dB" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 134, + 87, + 187, + 129 + ], + "blocks": [ + { + "bbox": [ + 134, + 87, + 187, + 129 + ], + "lines": [ + { + "bbox": [ + 134, + 87, + 187, + 129 + ], + "spans": [ + { + "bbox": [ + 134, + 87, + 187, + 129 + ], + "type": "image", + "image_path": "bfb05bd52a53941c2916344e8532302f0842f14a221170725fc67aa512bfd39d.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 201, + 88, + 255, + 129 + ], + "blocks": [ + { + "bbox": [ + 201, + 88, + 255, + 129 + ], + "lines": [ + { + "bbox": [ + 201, + 88, + 255, + 129 + ], + "spans": [ + { + "bbox": [ + 201, + 88, + 255, + 129 + ], + "type": "image", + "image_path": "013a22404f239d105b52f4540b3b8d4addf4f0d663cac0d11549c80dc352e0b0.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 62, + 159, + 124, + 214 + ], + "blocks": [ + { + "bbox": [ + 62, + 159, + 124, + 214 + ], + "lines": [ + { + "bbox": [ + 62, + 159, + 124, + 214 + ], + "spans": [ + { + "bbox": [ + 62, + 159, + 124, + 214 + ], + "type": "image", + "image_path": "e20880b2ecb8cb8e5c828ef3f744b3d45a6fda19a408a9c1599b49cd3ce1f4b7.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 135, + 220, + 190, + 229 + ], + "lines": [ + { + "bbox": [ + 135, + 220, + 190, + 229 + ], + "spans": [ + { + "bbox": [ + 135, + 220, + 190, + 229 + ], + "type": "text", + "content": "PSNR = 26.4 dB" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 130, + 159, + 192, + 214 + ], + "blocks": [ + { + "bbox": [ + 130, + 159, + 192, + 214 + ], + "lines": [ + { + "bbox": [ + 130, + 159, + 192, + 214 + ], + "spans": [ + { + "bbox": [ + 130, + 159, + 192, + 214 + ], + "type": "image", + "image_path": "8c849c23079cdbd9c7fca8520f96e9926475449810db2fce6a9e75b24164059a.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 199, + 159, + 260, + 214 + ], + "blocks": [ + { + "bbox": [ + 199, + 159, + 260, + 214 + ], + "lines": [ + { + "bbox": [ + 199, + 159, + 260, + 214 + ], + "spans": [ + { + "bbox": [ + 199, + 159, + 260, + 214 + ], + "type": "image", + "image_path": "21e82359b9c42bf0e73dedbe36315326d37630eca6f14f4c6703fd6784fbf1bb.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 88, + 237, + 104, + 285 + ], + "blocks": [ + { + "bbox": [ + 88, + 237, + 104, + 285 + ], + "lines": [ + { + "bbox": [ + 88, + 237, + 104, + 285 + ], + "spans": [ + { + "bbox": [ + 88, + 237, + 104, + 285 + ], + "type": "image", + "image_path": "839e78ad849635b81b1ab2d428c1448386e2aafa8d5989a4bd62043becaf3698.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 61, + 312, + 129, + 322 + ], + "lines": [ + { + "bbox": [ + 61, + 312, + 129, + 322 + ], + "spans": [ + { + "bbox": [ + 61, + 312, + 129, + 322 + ], + "type": "text", + "content": "(a) Sample input image" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 156, + 237, + 172, + 285 + ], + "blocks": [ + { + "bbox": [ + 156, + 237, + 172, + 285 + ], + "lines": [ + { + "bbox": [ + 156, + 237, + 172, + 285 + ], + "spans": [ + { + "bbox": [ + 156, + 237, + 172, + 285 + ], + "type": "image", + "image_path": "5f6df5ae024dcc3b2b02d44617ecda302035578f4a78db0b62e8929ee6eeb829.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 135, + 300, + 190, + 309 + ], + "lines": [ + { + "bbox": [ + 135, + 300, + 190, + 309 + ], + "spans": [ + { + "bbox": [ + 135, + 300, + 190, + 309 + ], + "type": "text", + "content": "PSNR = 29.6 dB" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 223, + 238, + 240, + 285 + ], + "blocks": [ + { + "bbox": [ + 223, + 238, + 240, + 285 + ], + "lines": [ + { + "bbox": [ + 223, + 238, + 240, + 285 + ], + "spans": [ + { + "bbox": [ + 223, + 238, + 240, + 285 + ], + "type": "image", + "image_path": "722be76b5e4fd12bcd9170e47a0875fdf998f8a8cfadb15a7da83cf7a69c6681.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 208, + 312, + 253, + 321 + ], + "lines": [ + { + "bbox": [ + 208, + 312, + 253, + 321 + ], + "spans": [ + { + "bbox": [ + 208, + 312, + 253, + 321 + ], + "type": "text", + "content": "(c) True albedo" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 267, + 70, + 397, + 137 + ], + "blocks": [ + { + "bbox": [ + 267, + 70, + 397, + 137 + ], + "lines": [ + { + "bbox": [ + 267, + 70, + 397, + 137 + ], + "spans": [ + { + "bbox": [ + 267, + 70, + 397, + 137 + ], + "type": "image", + "image_path": "4869fa775491944f014ae608cc50095f92748748a1489303987aeb8ab41b44da.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 399, + 70, + 530, + 137 + ], + "blocks": [ + { + "bbox": [ + 399, + 70, + 530, + 137 + ], + "lines": [ + { + "bbox": [ + 399, + 70, + 530, + 137 + ], + "spans": [ + { + "bbox": [ + 399, + 70, + 530, + 137 + ], + "type": "image", + "image_path": "9e91b7227707495da3ab7d67985eaf1e2d50ae51823e0d7b4b609dcb47f66a07.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 267, + 150, + 397, + 217 + ], + "blocks": [ + { + "bbox": [ + 307, + 140, + 358, + 149 + ], + "lines": [ + { + "bbox": [ + 307, + 140, + 358, + 149 + ], + "spans": [ + { + "bbox": [ + 307, + 140, + 358, + 149 + ], + "type": "inline_equation", + "content": "\\mathrm{RMSE} = 0.073" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 267, + 150, + 397, + 217 + ], + "lines": [ + { + "bbox": [ + 267, + 150, + 397, + 217 + ], + "spans": [ + { + "bbox": [ + 267, + 150, + 397, + 217 + ], + "type": "image", + "image_path": "6a87389ff443254ee77238bac0be9b75c018ad39691f15fe93d1604e0c22c4cd.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 399, + 150, + 529, + 217 + ], + "blocks": [ + { + "bbox": [ + 399, + 150, + 529, + 217 + ], + "lines": [ + { + "bbox": [ + 399, + 150, + 529, + 217 + ], + "spans": [ + { + "bbox": [ + 399, + 150, + 529, + 217 + ], + "type": "image", + "image_path": "9605522f36447750b8164f9d293250441442b3f48c2461898c3bc392724d1c0b.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 267, + 230, + 397, + 297 + ], + "blocks": [ + { + "bbox": [ + 307, + 220, + 356, + 228 + ], + "lines": [ + { + "bbox": [ + 307, + 220, + 356, + 228 + ], + "spans": [ + { + "bbox": [ + 307, + 220, + 356, + 228 + ], + "type": "text", + "content": "RMSE = 0.081" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 267, + 230, + 397, + 297 + ], + "lines": [ + { + "bbox": [ + 267, + 230, + 397, + 297 + ], + "spans": [ + { + "bbox": [ + 267, + 230, + 397, + 297 + ], + "type": "image", + "image_path": "a8a7a34b84fbd30c914914c874079ebe93f8d0ec280a427470361edbe8f17b20.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 307, + 299, + 358, + 309 + ], + "lines": [ + { + "bbox": [ + 307, + 299, + 358, + 309 + ], + "spans": [ + { + "bbox": [ + 307, + 299, + 358, + 309 + ], + "type": "inline_equation", + "content": "\\mathrm{RMSE} = 0.093" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 294, + 312, + 371, + 321 + ], + "lines": [ + { + "bbox": [ + 294, + 312, + 371, + 321 + ], + "spans": [ + { + "bbox": [ + 294, + 312, + 371, + 321 + ], + "type": "text", + "content": "(d) Recovered illumination" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 399, + 230, + 529, + 297 + ], + "blocks": [ + { + "bbox": [ + 399, + 230, + 529, + 297 + ], + "lines": [ + { + "bbox": [ + 399, + 230, + 529, + 297 + ], + "spans": [ + { + "bbox": [ + 399, + 230, + 529, + 297 + ], + "type": "image", + "image_path": "44a2d5b874791ddb6c04886f520d02945a25e1405f676b69fdce90795f914751.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 434, + 312, + 494, + 321 + ], + "lines": [ + { + "bbox": [ + 434, + 312, + 494, + 321 + ], + "spans": [ + { + "bbox": [ + 434, + 312, + 494, + 321 + ], + "type": "text", + "content": "(e) True illumination" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 50, + 377, + 102, + 475 + ], + "blocks": [ + { + "bbox": [ + 46, + 325, + 547, + 360 + ], + "lines": [ + { + "bbox": [ + 46, + 325, + 547, + 360 + ], + "spans": [ + { + "bbox": [ + 46, + 325, + 547, + 360 + ], + "type": "text", + "content": "Figure 4. The results of our method on three additional diffuse objects. We report the RMSE of each environment map in linear color space but plot the images after tonemapping for better evaluation of the full dynamic range. The albedo PSNR values are reported on object pixels only." + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 50, + 377, + 102, + 475 + ], + "lines": [ + { + "bbox": [ + 50, + 377, + 102, + 475 + ], + "spans": [ + { + "bbox": [ + 50, + 377, + 102, + 475 + ], + "type": "image", + "image_path": "7f99c9753ef13d7446e332cb81696c838627fc37380141451535b4f3411392bc.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_body" + } + ], + "index": 27 + }, + { + "type": "image", + "bbox": [ + 108, + 378, + 161, + 475 + ], + "blocks": [ + { + "bbox": [ + 133, + 312, + 194, + 322 + ], + "lines": [ + { + "bbox": [ + 133, + 312, + 194, + 322 + ], + "spans": [ + { + "bbox": [ + 133, + 312, + 194, + 322 + ], + "type": "text", + "content": "(b) Recovered albedo" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 108, + 378, + 161, + 475 + ], + "lines": [ + { + "bbox": [ + 108, + 378, + 161, + 475 + ], + "spans": [ + { + "bbox": [ + 108, + 378, + 161, + 475 + ], + "type": "image", + "image_path": "18e55259764e22af73ac545f6283f314a90f787e3698caaf833184ad1ff1ca68.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + } + ], + "index": 28 + }, + { + "type": "image", + "bbox": [ + 165, + 384, + 286, + 495 + ], + "blocks": [ + { + "bbox": [ + 165, + 384, + 286, + 495 + ], + "lines": [ + { + "bbox": [ + 165, + 384, + 286, + 495 + ], + "spans": [ + { + "bbox": [ + 165, + 384, + 286, + 495 + ], + "type": "image", + "image_path": "9ffa38b2fc7850f3bfd49bbc996dba852cdf282a83fbce33954c7d997d393786.jpg" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 503, + 287, + 570 + ], + "lines": [ + { + "bbox": [ + 46, + 503, + 287, + 570 + ], + "spans": [ + { + "bbox": [ + 46, + 503, + 287, + 570 + ], + "type": "text", + "content": "Figure 5. Our occluders extracted from the potatoes scene on the top of Figure 4 resemble the true ones (left). However, as expected, inaccuracies in the recovered masks tend to occur at relatively-dark regions of the environment map: there is a very strong inverse correlation (right) between the mean error in the recovered masks and the amount of illumination energy that they block." + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_caption" + } + ], + "index": 29 + }, + { + "bbox": [ + 46, + 590, + 287, + 626 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 590, + 287, + 626 + ], + "spans": [ + { + "bbox": [ + 46, + 590, + 287, + 626 + ], + "type": "text", + "content": "for a shadow ray emitted from surface point " + }, + { + "bbox": [ + 46, + 590, + 287, + 626 + ], + "type": "inline_equation", + "content": "\\mathbf{x} \\in S" + }, + { + "bbox": [ + 46, + 590, + 287, + 626 + ], + "type": "text", + "content": " in direction " + }, + { + "bbox": [ + 46, + 590, + 287, + 626 + ], + "type": "inline_equation", + "content": "\\hat{\\omega}_i" + }, + { + "bbox": [ + 46, + 590, + 287, + 626 + ], + "type": "text", + "content": " at time " + }, + { + "bbox": [ + 46, + 590, + 287, + 626 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 590, + 287, + 626 + ], + "type": "text", + "content": " can be computed by intersecting the shadow ray with the occluder's spherical shell:" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 101, + 634, + 287, + 648 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 101, + 634, + 287, + 648 + ], + "spans": [ + { + "bbox": [ + 101, + 634, + 287, + 648 + ], + "type": "interline_equation", + "content": "M _ {t} (\\mathbf {x}, \\hat {\\omega} _ {i}) = \\tilde {M} _ {t} \\left(S (\\mathbf {x}, \\hat {\\omega} _ {i}, r _ {t})\\right) \\tag {4}", + "image_path": "ee8765f774d9a655a45165dbaaf8ba6582e2c0413bf8f5c25c0403ca9327d593.jpg" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 47, + 657, + 287, + 681 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 657, + 287, + 681 + ], + "spans": [ + { + "bbox": [ + 47, + 657, + 287, + 681 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 657, + 287, + 681 + ], + "type": "inline_equation", + "content": "S(\\mathbf{x},\\hat{\\omega}_i,r)" + }, + { + "bbox": [ + 47, + 657, + 287, + 681 + ], + "type": "text", + "content": " is the intersection of the ray with a sphere of radius " + }, + { + "bbox": [ + 47, + 657, + 287, + 681 + ], + "type": "inline_equation", + "content": "r_t" + }, + { + "bbox": [ + 47, + 657, + 287, + 681 + ], + "type": "text", + "content": ", normalized to have unit length:" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 52, + 689, + 287, + 717 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 52, + 689, + 287, + 717 + ], + "spans": [ + { + "bbox": [ + 52, + 689, + 287, + 717 + ], + "type": "interline_equation", + "content": "S (\\mathbf {x}, \\hat {\\boldsymbol {\\omega}} _ {i}, r) = \\frac {\\mathbf {x} + \\left(\\sqrt {(\\mathbf {x} \\cdot \\hat {\\boldsymbol {\\omega}} _ {i}) ^ {2} + r ^ {2} - \\| \\mathbf {x} \\| ^ {2}} - \\mathbf {x} \\cdot \\hat {\\boldsymbol {\\omega}} _ {i}\\right) \\hat {\\boldsymbol {\\omega}} _ {i}}{r}. \\tag {5}", + "image_path": "872c0dae3f01028ff7987e067f22a61e63ac04805d018b80722174844cf75132.jpg" + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 305, + 380, + 545, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 380, + 545, + 403 + ], + "spans": [ + { + "bbox": [ + 305, + 380, + 545, + 403 + ], + "type": "text", + "content": "Note that there is a single intersection, since we assume that the surface points are always inside the occluder spheres." + } + ] + } + ], + "index": 35 + }, + { + "bbox": [ + 305, + 404, + 546, + 476 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 404, + 546, + 476 + ], + "spans": [ + { + "bbox": [ + 305, + 404, + 546, + 476 + ], + "type": "text", + "content": "Although our formulation models the occluders as binary signals, a binary representation is not well-suited for gradient-based optimization. Instead, we represent the occluders as a continuously-valued function on the sphere in a spherical harmonic basis, mapped using a sigmoid function " + }, + { + "bbox": [ + 305, + 404, + 546, + 476 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 305, + 404, + 546, + 476 + ], + "type": "text", + "content": " to lie in [0, 1]:" + } + ] + } + ], + "index": 36 + }, + { + "bbox": [ + 344, + 484, + 545, + 519 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 344, + 484, + 545, + 519 + ], + "spans": [ + { + "bbox": [ + 344, + 484, + 545, + 519 + ], + "type": "interline_equation", + "content": "\\tilde {M} _ {t} (\\hat {\\omega}) = \\sigma \\left(\\sum_ {\\ell = 0} ^ {P} \\sum_ {m = - \\ell} ^ {\\ell} a _ {t \\ell m} Y _ {\\ell} ^ {m} (\\hat {\\omega})\\right), \\tag {6}", + "image_path": "bc0951d69d7a732368d36dadce5bbaa5e09292240257820b311ac2d57a82b0d9.jpg" + } + ] + } + ], + "index": 37 + }, + { + "bbox": [ + 304, + 529, + 545, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 529, + 545, + 579 + ], + "spans": [ + { + "bbox": [ + 304, + 529, + 545, + 579 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 529, + 545, + 579 + ], + "type": "inline_equation", + "content": "\\phi^{(o)}\\triangleq \\{a_{t\\ell m}\\}" + }, + { + "bbox": [ + 304, + 529, + 545, + 579 + ], + "type": "text", + "content": " are estimizable coefficients that parameterize the occluder at time " + }, + { + "bbox": [ + 304, + 529, + 545, + 579 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 529, + 545, + 579 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 529, + 545, + 579 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 304, + 529, + 545, + 579 + ], + "type": "text", + "content": " is the degree required to span the space of spherical images with the same resolution as our environment illumination (see below) [11]." + } + ] + } + ], + "index": 38 + }, + { + "bbox": [ + 304, + 579, + 545, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 579, + 545, + 651 + ], + "spans": [ + { + "bbox": [ + 304, + 579, + 545, + 651 + ], + "type": "text", + "content": "Note that for radially symmetric BRDFs, the rendering integral in Equation 2 can be written as a spherical convolution [11, 33]. This makes spherical harmonics a natural choice for representing the occluder signals, since spherical convolutions are diagonal in the basis of spherical harmonics. See our supplement for more details." + } + ] + } + ], + "index": 39 + }, + { + "bbox": [ + 306, + 659, + 451, + 670 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 659, + 451, + 670 + ], + "spans": [ + { + "bbox": [ + 306, + 659, + 451, + 670 + ], + "type": "text", + "content": "4.2. Environment Illumination" + } + ] + } + ], + "index": 40 + }, + { + "bbox": [ + 305, + 677, + 546, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 677, + 546, + 714 + ], + "spans": [ + { + "bbox": [ + 305, + 677, + 546, + 714 + ], + "type": "text", + "content": "We assume that the scene's illumination is distant and can be represented as an environment map. The environment map is parameterized as an image pyramid of size " + }, + { + "bbox": [ + 305, + 677, + 546, + 714 + ], + "type": "inline_equation", + "content": "H' \\times W'" + } + ] + } + ], + "index": 41 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 299, + 749, + 310, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 749, + 310, + 757 + ], + "spans": [ + { + "bbox": [ + 299, + 749, + 310, + 757 + ], + "type": "text", + "content": "81" + } + ] + } + ], + "index": 42 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 66, + 70, + 218, + 148 + ], + "blocks": [ + { + "bbox": [ + 66, + 70, + 218, + 148 + ], + "lines": [ + { + "bbox": [ + 66, + 70, + 218, + 148 + ], + "spans": [ + { + "bbox": [ + 66, + 70, + 218, + 148 + ], + "type": "image", + "image_path": "eb9121225b634bdf165cbf8c0046e239316bcd9c87059ee4f484eaeb5e08f6c2.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 76, + 150, + 208, + 159 + ], + "lines": [ + { + "bbox": [ + 76, + 150, + 208, + 159 + ], + "spans": [ + { + "bbox": [ + 76, + 150, + 208, + 159 + ], + "type": "text", + "content": "(a) Recovered illumination (no occluder)" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 219, + 69, + 372, + 148 + ], + "blocks": [ + { + "bbox": [ + 219, + 69, + 372, + 148 + ], + "lines": [ + { + "bbox": [ + 219, + 69, + 372, + 148 + ], + "spans": [ + { + "bbox": [ + 219, + 69, + 372, + 148 + ], + "type": "image", + "image_path": "43aa90649c1995ebeff7b2492ffb2fb682441daad0fe10b35775f1f02d43b0fc.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 251, + 150, + 339, + 159 + ], + "lines": [ + { + "bbox": [ + 251, + 150, + 339, + 159 + ], + "spans": [ + { + "bbox": [ + 251, + 150, + 339, + 159 + ], + "type": "text", + "content": "(b) Recovered illumination" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 162, + 547, + 195 + ], + "lines": [ + { + "bbox": [ + 46, + 162, + 547, + 195 + ], + "spans": [ + { + "bbox": [ + 46, + 162, + 547, + 195 + ], + "type": "text", + "content": "Figure 6. In the case of purely Lambertian material (with the potatoes geometry), (a) only using self occlusions is insufficient for the recovery of high-frequency content in the illumination. (b) Modeling and estimating occluders recovers significantly higher-quality illumination that closely resembles the ground truth." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 374, + 70, + 525, + 148 + ], + "blocks": [ + { + "bbox": [ + 374, + 70, + 525, + 148 + ], + "lines": [ + { + "bbox": [ + 374, + 70, + 525, + 148 + ], + "spans": [ + { + "bbox": [ + 374, + 70, + 525, + 148 + ], + "type": "image", + "image_path": "44147fbc4e337644558b9d78179e4ec47fea4a755b1cb9d1d8d8574f736de55d.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 414, + 150, + 484, + 159 + ], + "lines": [ + { + "bbox": [ + 414, + 150, + 484, + 159 + ], + "spans": [ + { + "bbox": [ + 414, + 150, + 484, + 159 + ], + "type": "text", + "content": "(c) True illumination" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 216, + 287, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 216, + 287, + 240 + ], + "spans": [ + { + "bbox": [ + 46, + 216, + 287, + 240 + ], + "type": "text", + "content": "in equirectangular coordinates " + }, + { + "bbox": [ + 46, + 216, + 287, + 240 + ], + "type": "inline_equation", + "content": "(50 \\times 100" + }, + { + "bbox": [ + 46, + 216, + 287, + 240 + ], + "type": "text", + "content": " in our experiments), with an exponential nonlinearity:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 97, + 248, + 287, + 282 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 97, + 248, + 287, + 282 + ], + "spans": [ + { + "bbox": [ + 97, + 248, + 287, + 282 + ], + "type": "interline_equation", + "content": "L \\left(\\hat {\\omega} _ {i}\\right) = \\exp \\left(\\sum_ {k = 0} ^ {K - 1} a ^ {k} L _ {k} \\left(\\hat {\\omega} _ {i}\\right)\\right), \\tag {7}", + "image_path": "acd055ad84689a27a76d2ccdca6afbc284750198bebcb8aa54d6b42bf1cfa8d8.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 292, + 287, + 378 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 292, + 287, + 378 + ], + "spans": [ + { + "bbox": [ + 47, + 292, + 287, + 378 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 292, + 287, + 378 + ], + "type": "inline_equation", + "content": "\\phi^{(\\ell)}\\triangleq \\{L_k\\}_{k = 0}^{K - 1}" + }, + { + "bbox": [ + 47, + 292, + 287, + 378 + ], + "type": "text", + "content": " comprise a coarse (" + }, + { + "bbox": [ + 47, + 292, + 287, + 378 + ], + "type": "inline_equation", + "content": "k = 0" + }, + { + "bbox": [ + 47, + 292, + 287, + 378 + ], + "type": "text", + "content": ") to fine (" + }, + { + "bbox": [ + 47, + 292, + 287, + 378 + ], + "type": "inline_equation", + "content": "k = K - 1" + }, + { + "bbox": [ + 47, + 292, + 287, + 378 + ], + "type": "text", + "content": ") representation of the illumination, and we set " + }, + { + "bbox": [ + 47, + 292, + 287, + 378 + ], + "type": "inline_equation", + "content": "a = 2" + }, + { + "bbox": [ + 47, + 292, + 287, + 378 + ], + "type": "text", + "content": ". For each " + }, + { + "bbox": [ + 47, + 292, + 287, + 378 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 47, + 292, + 287, + 378 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 47, + 292, + 287, + 378 + ], + "type": "inline_equation", + "content": "L_{k}(\\hat{\\omega}_{i})" + }, + { + "bbox": [ + 47, + 292, + 287, + 378 + ], + "type": "text", + "content": " is computed by bilinearly interpolating into a grid whose size exponentially increases with " + }, + { + "bbox": [ + 47, + 292, + 287, + 378 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 47, + 292, + 287, + 378 + ], + "type": "text", + "content": ". The exponential nonlinearity is helpful for obtaining high dynamic range values in the environment map, as noted, e.g. in [3]." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 378, + 287, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 378, + 287, + 426 + ], + "spans": [ + { + "bbox": [ + 46, + 378, + 287, + 426 + ], + "type": "text", + "content": "Unlike the occluder masks, we find that representing the environment map as a pyramid results in better performance than what is achieved with spherical harmonics. See appendix for additional details." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 433, + 115, + 445 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 433, + 115, + 445 + ], + "spans": [ + { + "bbox": [ + 47, + 433, + 115, + 445 + ], + "type": "text", + "content": "4.3. Materials" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 451, + 287, + 536 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 451, + 287, + 536 + ], + "spans": [ + { + "bbox": [ + 46, + 451, + 287, + 536 + ], + "type": "text", + "content": "We represent the scene's spatially-varying BRDF as the sum of a diffuse and a specular component. We use a Lambertian model for the diffuse term and a GGX microfacet model [43] for the specular term. The spatially-varying BRDF parameters are represented by a coordinate-based multi-layer perceptron (MLP) with a positional encoding [38] function " + }, + { + "bbox": [ + 46, + 451, + 287, + 536 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 46, + 451, + 287, + 536 + ], + "type": "text", + "content": " and estimizable parameters " + }, + { + "bbox": [ + 46, + 451, + 287, + 536 + ], + "type": "inline_equation", + "content": "\\phi^{(m)}" + }, + { + "bbox": [ + 46, + 451, + 287, + 536 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 543, + 287, + 564 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 106, + 543, + 287, + 564 + ], + "spans": [ + { + "bbox": [ + 106, + 543, + 287, + 564 + ], + "type": "interline_equation", + "content": "\\boldsymbol {\\alpha} (\\mathbf {x}) = \\operatorname {M L P} \\left(\\gamma (\\mathbf {x}); \\phi^ {(m)}\\right). \\tag {8}", + "image_path": "7f33dd2b8beb64d56069381624677c247b39fb19bf1630700357dc6c7327e5bc.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 46, + 567, + 287, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 567, + 287, + 653 + ], + "spans": [ + { + "bbox": [ + 46, + 567, + 287, + 653 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 46, + 567, + 287, + 653 + ], + "type": "inline_equation", + "content": "\\gamma (\\mathbf{x})" + }, + { + "bbox": [ + 46, + 567, + 287, + 653 + ], + "type": "text", + "content": " is the positionally-encoded position on the object's surface, and " + }, + { + "bbox": [ + 46, + 567, + 287, + 653 + ], + "type": "inline_equation", + "content": "\\alpha (\\mathbf{x})\\in \\mathbb{R}^5" + }, + { + "bbox": [ + 46, + 567, + 287, + 653 + ], + "type": "text", + "content": " are BRDF coefficients, consisting of an RGB diffuse albedo, a spatially-varying scalar microfacet roughness, and a spatially-varying scalar specular reflectance at normal incidence (equivalent to a reparameterization of the material's index of refraction). See the supplement for an exact specification of our BRDF model." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 47, + 659, + 119, + 672 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 659, + 119, + 672 + ], + "spans": [ + { + "bbox": [ + 47, + 659, + 119, + 672 + ], + "type": "text", + "content": "4.4. Rendering" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 46, + 677, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 287, + 714 + ], + "type": "text", + "content": "We render the occluders, materials, and illumination, i.e. the rendering operator " + }, + { + "bbox": [ + 46, + 677, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{R}_t" + }, + { + "bbox": [ + 46, + 677, + 287, + 714 + ], + "type": "text", + "content": " from Problem 3, by approximating the integral in Equation 2 using Monte Carlo techniques." + } + ] + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 309, + 213, + 382, + 276 + ], + "blocks": [ + { + "bbox": [ + 309, + 213, + 382, + 276 + ], + "lines": [ + { + "bbox": [ + 309, + 213, + 382, + 276 + ], + "spans": [ + { + "bbox": [ + 309, + 213, + 382, + 276 + ], + "type": "image", + "image_path": "82b89b13cf92cd41db5da1bf57de2b83ac32155d5f733d975a452065600067c9.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 308, + 277, + 385, + 287 + ], + "lines": [ + { + "bbox": [ + 308, + 277, + 385, + 287 + ], + "spans": [ + { + "bbox": [ + 308, + 277, + 385, + 287 + ], + "type": "text", + "content": "(a) Sample input image" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 391, + 213, + 463, + 277 + ], + "blocks": [ + { + "bbox": [ + 391, + 213, + 463, + 277 + ], + "lines": [ + { + "bbox": [ + 391, + 213, + 463, + 277 + ], + "spans": [ + { + "bbox": [ + 391, + 213, + 463, + 277 + ], + "type": "image", + "image_path": "f9fd746aa9328e892228beb2630c16b8593a0ca51e8c74a4900d9a4d58336467.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 394, + 277, + 463, + 285 + ], + "lines": [ + { + "bbox": [ + 394, + 277, + 463, + 285 + ], + "spans": [ + { + "bbox": [ + 394, + 277, + 463, + 285 + ], + "type": "text", + "content": "(b) Recovered albedo" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 473, + 213, + 545, + 277 + ], + "blocks": [ + { + "bbox": [ + 473, + 213, + 545, + 277 + ], + "lines": [ + { + "bbox": [ + 473, + 213, + 545, + 277 + ], + "spans": [ + { + "bbox": [ + 473, + 213, + 545, + 277 + ], + "type": "image", + "image_path": "efdaf10d6b5ca1f563d6c15c6b58408a9ec3efab29a9025dfe1057727bd12dc8.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 485, + 277, + 535, + 285 + ], + "lines": [ + { + "bbox": [ + 485, + 277, + 535, + 285 + ], + "spans": [ + { + "bbox": [ + 485, + 277, + 535, + 285 + ], + "type": "text", + "content": "(c) True albedo" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 307, + 291, + 427, + 352 + ], + "blocks": [ + { + "bbox": [ + 307, + 291, + 427, + 352 + ], + "lines": [ + { + "bbox": [ + 307, + 291, + 427, + 352 + ], + "spans": [ + { + "bbox": [ + 307, + 291, + 427, + 352 + ], + "type": "image", + "image_path": "404210c2a17e307793a4aaa1e101134d675cb92cca16ff3d35274bd0e985283b.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 322, + 353, + 410, + 361 + ], + "lines": [ + { + "bbox": [ + 322, + 353, + 410, + 361 + ], + "spans": [ + { + "bbox": [ + 322, + 353, + 410, + 361 + ], + "type": "text", + "content": "(d) Recovered illumination" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 305, + 366, + 546, + 421 + ], + "lines": [ + { + "bbox": [ + 305, + 366, + 546, + 421 + ], + "spans": [ + { + "bbox": [ + 305, + 366, + 546, + 421 + ], + "type": "text", + "content": "Figure 7. We are able to recover illumination even when geometry is unknown by first optimizing a volumetric representation of geometry using a NeRF-based method. Despite this inaccurate proxy geometry, our method still recovers plausible (albeit blurry) illumination (d) and albedo (b)." + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_caption" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 429, + 291, + 548, + 352 + ], + "blocks": [ + { + "bbox": [ + 429, + 291, + 548, + 352 + ], + "lines": [ + { + "bbox": [ + 429, + 291, + 548, + 352 + ], + "spans": [ + { + "bbox": [ + 429, + 291, + 548, + 352 + ], + "type": "image", + "image_path": "b2d26c2b212e96a3245e2858551847fb6472036fc0dd078701b0a34e89a5c8f7.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 454, + 353, + 522, + 361 + ], + "lines": [ + { + "bbox": [ + 454, + 353, + 522, + 361 + ], + "spans": [ + { + "bbox": [ + 454, + 353, + 522, + 361 + ], + "type": "text", + "content": "(e) True illumination" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_caption" + } + ], + "index": 25 + }, + { + "bbox": [ + 305, + 442, + 546, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 442, + 546, + 515 + ], + "spans": [ + { + "bbox": [ + 305, + 442, + 546, + 515 + ], + "type": "text", + "content": "We use a standard multiple importance sampler [40] based on the illumination and material model. We model shadows cast by the occluders and self-occlusions by the object itself, but for efficiency and due to our focus on diffuse objects, we neglect lower-order effects and global illumination. See the supplemental material for a full description of our renderer." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 306, + 521, + 391, + 535 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 521, + 391, + 535 + ], + "spans": [ + { + "bbox": [ + 306, + 521, + 391, + 535 + ], + "type": "text", + "content": "4.5. Optimization" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 305, + 540, + 545, + 599 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 540, + 545, + 599 + ], + "spans": [ + { + "bbox": [ + 305, + 540, + 545, + 599 + ], + "type": "text", + "content": "We optimize the objective in Equation 3 using the " + }, + { + "bbox": [ + 305, + 540, + 545, + 599 + ], + "type": "inline_equation", + "content": "L_{2}" + }, + { + "bbox": [ + 305, + 540, + 545, + 599 + ], + "type": "text", + "content": " error between rendered values and ground truth ones. We use Adam [16] to optimize all three components of our model, with a learning rate of " + }, + { + "bbox": [ + 305, + 540, + 545, + 599 + ], + "type": "inline_equation", + "content": "3 \\cdot 10^{-3}" + }, + { + "bbox": [ + 305, + 540, + 545, + 599 + ], + "type": "text", + "content": " for the environment map and materials, and a learning rate of 1 for the occluders." + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 304, + 600, + 546, + 673 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 600, + 546, + 673 + ], + "spans": [ + { + "bbox": [ + 304, + 600, + 546, + 673 + ], + "type": "text", + "content": "In each iteration, we compute the " + }, + { + "bbox": [ + 304, + 600, + 546, + 673 + ], + "type": "inline_equation", + "content": "L_{2}" + }, + { + "bbox": [ + 304, + 600, + 546, + 673 + ], + "type": "text", + "content": " loss using a batch size of " + }, + { + "bbox": [ + 304, + 600, + 546, + 673 + ], + "type": "inline_equation", + "content": "2^{16}" + }, + { + "bbox": [ + 304, + 600, + 546, + 673 + ], + "type": "text", + "content": " pixels. In order to avoid bias in the gradient updates to our model's parameters, we use the same approach as [10]: we render every pixel value twice using independent samples and multiply the per-channel deviations of both from the ground truth to get the expected loss value:" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 317, + 689, + 545, + 716 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 689, + 545, + 716 + ], + "spans": [ + { + "bbox": [ + 317, + 689, + 545, + 716 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\sum_ {\\mathbf {u}} \\left(\\tilde {I} _ {t} ^ {(1)} (\\mathbf {u}) - I _ {t} (\\mathbf {u})\\right) \\cdot \\left(\\tilde {I} _ {t} ^ {(2)} (\\mathbf {u}) - I _ {t} (\\mathbf {u})\\right), \\tag {9}", + "image_path": "b2681d188276aa25a75e7afc2011935adbdbc202b56fae11d70fb2850d46102c.jpg" + } + ] + } + ], + "index": 32 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 299, + 749, + 310, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 749, + 310, + 757 + ], + "spans": [ + { + "bbox": [ + 299, + 749, + 310, + 757 + ], + "type": "text", + "content": "82" + } + ] + } + ], + "index": 33 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 78, + 73, + 136, + 175 + ], + "blocks": [ + { + "bbox": [ + 78, + 73, + 136, + 175 + ], + "lines": [ + { + "bbox": [ + 78, + 73, + 136, + 175 + ], + "spans": [ + { + "bbox": [ + 78, + 73, + 136, + 175 + ], + "type": "image", + "image_path": "8969b121a22bd2300a109fbea4bd520f17f6be2aa2c08364b4bfcc6f495f85d4.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 204, + 74, + 261, + 174 + ], + "blocks": [ + { + "bbox": [ + 204, + 74, + 261, + 174 + ], + "lines": [ + { + "bbox": [ + 204, + 74, + 261, + 174 + ], + "spans": [ + { + "bbox": [ + 204, + 74, + 261, + 174 + ], + "type": "image", + "image_path": "1b2ce43d5f006ad02de392dd16365648815d779d1d83390ccd913aa8a7e30f17.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 208, + 185, + 264, + 194 + ], + "lines": [ + { + "bbox": [ + 208, + 185, + 264, + 194 + ], + "spans": [ + { + "bbox": [ + 208, + 185, + 264, + 194 + ], + "type": "text", + "content": "PSNR = 19.3 dB" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 331, + 75, + 388, + 174 + ], + "blocks": [ + { + "bbox": [ + 331, + 75, + 388, + 174 + ], + "lines": [ + { + "bbox": [ + 331, + 75, + 388, + 174 + ], + "spans": [ + { + "bbox": [ + 331, + 75, + 388, + 174 + ], + "type": "image", + "image_path": "52fe5dc510adbffb435d476eab6520a392cff5adf340609fb56ec57c39a77e6c.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 334, + 185, + 390, + 194 + ], + "lines": [ + { + "bbox": [ + 334, + 185, + 390, + 194 + ], + "spans": [ + { + "bbox": [ + 334, + 185, + 390, + 194 + ], + "type": "text", + "content": "PSNR = 14.4 dB" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 457, + 74, + 514, + 174 + ], + "blocks": [ + { + "bbox": [ + 457, + 74, + 514, + 174 + ], + "lines": [ + { + "bbox": [ + 457, + 74, + 514, + 174 + ], + "spans": [ + { + "bbox": [ + 457, + 74, + 514, + 174 + ], + "type": "image", + "image_path": "d7a52d26f6a3899dee3a2a66af0108304fbc5da7c0f635b203d4fb97fc50f56e.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 49, + 211, + 211, + 294 + ], + "blocks": [ + { + "bbox": [ + 71, + 197, + 148, + 207 + ], + "lines": [ + { + "bbox": [ + 71, + 197, + 148, + 207 + ], + "spans": [ + { + "bbox": [ + 71, + 197, + 148, + 207 + ], + "type": "text", + "content": "(a) Sample input image" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 49, + 211, + 211, + 294 + ], + "lines": [ + { + "bbox": [ + 49, + 211, + 211, + 294 + ], + "spans": [ + { + "bbox": [ + 49, + 211, + 211, + 294 + ], + "type": "image", + "image_path": "f9e5985df6daf92fc8905fdf3fa38da529fd1ee6402097a312d405851642c25c.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 105, + 295, + 154, + 303 + ], + "lines": [ + { + "bbox": [ + 105, + 295, + 154, + 303 + ], + "spans": [ + { + "bbox": [ + 105, + 295, + 154, + 303 + ], + "type": "text", + "content": "RMSE " + }, + { + "bbox": [ + 105, + 295, + 154, + 303 + ], + "type": "inline_equation", + "content": "= 0.035" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 86, + 307, + 173, + 316 + ], + "lines": [ + { + "bbox": [ + 86, + 307, + 173, + 316 + ], + "spans": [ + { + "bbox": [ + 86, + 307, + 173, + 316 + ], + "type": "text", + "content": "(e) Recovered illumination" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 319, + 546, + 353 + ], + "lines": [ + { + "bbox": [ + 46, + 319, + 546, + 353 + ], + "spans": [ + { + "bbox": [ + 46, + 319, + 546, + 353 + ], + "type": "text", + "content": "Figure 8. Results on the captured salt004 scene from Stanford-ORB [18]. Note that the \"warped true illumination\" environment map (g) was captured by a light probe that was not co-located with the object, any may therefore be significantly warped. The recovered RMSE reported is computed with respect to the average of all environment maps provided with the dataset rotated the same coordinate frame." + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 213, + 212, + 377, + 294 + ], + "blocks": [ + { + "bbox": [ + 201, + 198, + 271, + 206 + ], + "lines": [ + { + "bbox": [ + 201, + 198, + 271, + 206 + ], + "spans": [ + { + "bbox": [ + 201, + 198, + 271, + 206 + ], + "type": "text", + "content": "(b) Recovered albedo" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 305, + 198, + 419, + 206 + ], + "lines": [ + { + "bbox": [ + 305, + 198, + 419, + 206 + ], + "spans": [ + { + "bbox": [ + 305, + 198, + 419, + 206 + ], + "type": "text", + "content": "(c) Recovered albedo (no occluder)" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 213, + 212, + 377, + 294 + ], + "lines": [ + { + "bbox": [ + 213, + 212, + 377, + 294 + ], + "spans": [ + { + "bbox": [ + 213, + 212, + 377, + 294 + ], + "type": "image", + "image_path": "ed7c83d0bb66b90b9e6f018f25bea90ce53965ff93f2932e63c4d3a1266e8e38.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 270, + 295, + 321, + 303 + ], + "lines": [ + { + "bbox": [ + 270, + 295, + 321, + 303 + ], + "spans": [ + { + "bbox": [ + 270, + 295, + 321, + 303 + ], + "type": "text", + "content": "RMSE = 0.087" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 230, + 307, + 361, + 316 + ], + "lines": [ + { + "bbox": [ + 230, + 307, + 361, + 316 + ], + "spans": [ + { + "bbox": [ + 230, + 307, + 361, + 316 + ], + "type": "text", + "content": "(f) Recovered illumination (no occluder)" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 380, + 212, + 543, + 294 + ], + "blocks": [ + { + "bbox": [ + 463, + 198, + 514, + 206 + ], + "lines": [ + { + "bbox": [ + 463, + 198, + 514, + 206 + ], + "spans": [ + { + "bbox": [ + 463, + 198, + 514, + 206 + ], + "type": "text", + "content": "(d) True albedo" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 380, + 212, + 543, + 294 + ], + "lines": [ + { + "bbox": [ + 380, + 212, + 543, + 294 + ], + "spans": [ + { + "bbox": [ + 380, + 212, + 543, + 294 + ], + "type": "image", + "image_path": "55c65d68bc598be062d0925de02489462bc13d9af434d32d5a6cb2b0786ce13f.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 413, + 307, + 509, + 316 + ], + "lines": [ + { + "bbox": [ + 413, + 307, + 509, + 316 + ], + "spans": [ + { + "bbox": [ + 413, + 307, + 509, + 316 + ], + "type": "text", + "content": "(g) Warped true illumination" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "bbox": [ + 46, + 372, + 287, + 398 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 372, + 287, + 398 + ], + "spans": [ + { + "bbox": [ + 46, + 372, + 287, + 398 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 372, + 287, + 398 + ], + "type": "inline_equation", + "content": "\\tilde{I}_t^{(1)}(\\mathbf{u})" + }, + { + "bbox": [ + 46, + 372, + 287, + 398 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 372, + 287, + 398 + ], + "type": "inline_equation", + "content": "\\tilde{I}_t^{(2)}(\\mathbf{u})" + }, + { + "bbox": [ + 46, + 372, + 287, + 398 + ], + "type": "text", + "content": " are the two independently-rendered pixel values." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 47, + 411, + 128, + 426 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 411, + 128, + 426 + ], + "spans": [ + { + "bbox": [ + 47, + 411, + 128, + 426 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 46, + 433, + 287, + 564 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 433, + 287, + 564 + ], + "spans": [ + { + "bbox": [ + 46, + 433, + 287, + 564 + ], + "type": "text", + "content": "We implement our entire rendering and optimization pipeline in JAX [8] and run each of our experiments on 8 NVIDIA V100 GPUs. We evaluate our method on two sources of data: synthetic objects, and captured data from the Stanford-ORB dataset [18]. Our synthetic results were generated using 256 input images of size " + }, + { + "bbox": [ + 46, + 433, + 287, + 564 + ], + "type": "inline_equation", + "content": "400 \\times 400" + }, + { + "bbox": [ + 46, + 433, + 287, + 564 + ], + "type": "text", + "content": ", although our method works similarly well with as few as 32 captures (see supplement for the effect of the number of images). Unless noted otherwise, the occluders for all experiments are spherical caps with a size of " + }, + { + "bbox": [ + 46, + 433, + 287, + 564 + ], + "type": "inline_equation", + "content": "0.1 \\cdot 4\\pi" + }, + { + "bbox": [ + 46, + 433, + 287, + 564 + ], + "type": "text", + "content": " steradians, centered at the camera." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 46, + 570, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 570, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 570, + 288, + 713 + ], + "type": "text", + "content": "Estimating illumination, materials, and occluders. Our method can effectively recover environment illumination, spatially-varying material parameters, and the shapes of unseen lighting occluders. We first investigate our algorithm's performance using rendered images of a variety of diffuse objects (roughness 0.6, see BRDF model in appendix) with known geometry. The illumination and albedos recovered by our model are visualized with their true values in Figures 1 and 4. Figure 6 shows our reconstruction for purely Lambertian materials, compared with the reconstruction from the same number of images without unobserved occluders (but with self-occlusions), similar to the scenario" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 304, + 373, + 546, + 447 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 373, + 546, + 447 + ], + "spans": [ + { + "bbox": [ + 304, + 373, + 546, + 447 + ], + "type": "text", + "content": "investigated by Swedish et al. [37]. For Lambertian objects, reconstruction is severely ill-conditioned even when self-occlusions are present, which results in significantly blurrier reconstructions. Figure 5 validates that our recovered occluder shapes are accurate, especially for occluders that block more of the illumination energy." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 304, + 453, + 545, + 536 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 453, + 545, + 536 + ], + "spans": [ + { + "bbox": [ + 304, + 453, + 545, + 536 + ], + "type": "text", + "content": "Captured data. We apply our method to scenes from the Stanford-ORB dataset [18]. Each object in the dataset is placed on a platform and captured from multiple directions. As the camera rig and photographer move around the scene, the photographer is hidden under white cloth below the camera. The images are provided along with a scanned mesh which we use for geometry." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 304, + 539, + 546, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 539, + 546, + 647 + ], + "spans": [ + { + "bbox": [ + 304, + 539, + 546, + 647 + ], + "type": "text", + "content": "Despite not being designed for our task, Figure 8 shows our recovered environment map obtained using the 66 training views in the salt004 scene, which features a rough cylindrical object. Note that the dataset only provides an estimated illumination obtained by a light probe placed at a large distance from the object, meaning there is a significant unknown non-rigid warp between the true incident light and the image in Figure 8g labeled \"warped true illumination\". See supplement for additional results." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 304, + 653, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 653, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 653, + 547, + 713 + ], + "type": "text", + "content": "Without known geometry. Figure 7 presents preliminary evidence that our method can be used to recover lighting and materials even when the object geometry is not known. In this experiment, we first recover an estimate of object geometry using a Neural Radiance Field (NeRF) [25]-based" + } + ] + } + ], + "index": 26 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 299, + 749, + 310, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 749, + 310, + 757 + ], + "spans": [ + { + "bbox": [ + 299, + 749, + 310, + 757 + ], + "type": "text", + "content": "83" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 51, + 72, + 205, + 155 + ], + "blocks": [ + { + "bbox": [ + 51, + 72, + 205, + 155 + ], + "lines": [ + { + "bbox": [ + 51, + 72, + 205, + 155 + ], + "spans": [ + { + "bbox": [ + 51, + 72, + 205, + 155 + ], + "type": "image", + "image_path": "e4323f1caa819e4b8bf8fe94aacff0288c6e9c013c450b6c677c615ce52a2961.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 159, + 547, + 227 + ], + "lines": [ + { + "bbox": [ + 46, + 159, + 547, + 227 + ], + "spans": [ + { + "bbox": [ + 46, + 159, + 547, + 227 + ], + "type": "text", + "content": "Figure 9. The signal provided by unintended shadows cast by unseen occluders improves the quality of recovered environment illumination. Here, we plot the relative MSE for the recovered environment maps under two scenarios: (blue) using images rendered with unobserved occluders and jointly estimating materials, illumination and occluder shape and (orange) using images rendered without occluders and only estimating materials and illumination (this is the problem setting considered by Swedish et al. [37]). The cue of unintended shadows consistently improves the quality of estimated illumination across varying (a) illumination frequency content, (b) object material roughness, and (c) mask sizes. Additionally, we show that larger occluders that block more light improve the reconstructed illumination quality." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 219, + 72, + 373, + 155 + ], + "blocks": [ + { + "bbox": [ + 219, + 72, + 373, + 155 + ], + "lines": [ + { + "bbox": [ + 219, + 72, + 373, + 155 + ], + "spans": [ + { + "bbox": [ + 219, + 72, + 373, + 155 + ], + "type": "image", + "image_path": "021b1307690b9c70ac09ecade42aa419ad190b22284784c0d8a81d651e62e339.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 387, + 72, + 539, + 155 + ], + "blocks": [ + { + "bbox": [ + 387, + 72, + 539, + 155 + ], + "lines": [ + { + "bbox": [ + 387, + 72, + 539, + 155 + ], + "spans": [ + { + "bbox": [ + 387, + 72, + 539, + 155 + ], + "type": "image", + "image_path": "95bf54043295c17c2abccbb0c9ef7880ca1adbfe368d4f47b0cb2a86239c10a5.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 50, + 245, + 165, + 364 + ], + "blocks": [ + { + "bbox": [ + 50, + 245, + 165, + 364 + ], + "lines": [ + { + "bbox": [ + 50, + 245, + 165, + 364 + ], + "spans": [ + { + "bbox": [ + 50, + 245, + 165, + 364 + ], + "type": "image", + "image_path": "42d997c8f4d3eafc5395befc6c63eb65b6cfb185dee5e3c69b07552940e61518.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 70, + 365, + 145, + 374 + ], + "lines": [ + { + "bbox": [ + 70, + 365, + 145, + 374 + ], + "spans": [ + { + "bbox": [ + 70, + 365, + 145, + 374 + ], + "type": "text", + "content": "(a) Illumination (pyramid)" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 377, + 288, + 433 + ], + "lines": [ + { + "bbox": [ + 46, + 377, + 288, + 433 + ], + "spans": [ + { + "bbox": [ + 46, + 377, + 288, + 433 + ], + "type": "text", + "content": "Figure 10. Results from the experiment in Figure 4, where the pyramid environment map has been replaced with a direct optimization of its (pre-exponentiated) values. Directly optimizing the mask values instead of using spherical harmonics also results in worse reconstructions (see Table 1)." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 167, + 245, + 282, + 364 + ], + "blocks": [ + { + "bbox": [ + 167, + 245, + 282, + 364 + ], + "lines": [ + { + "bbox": [ + 167, + 245, + 282, + 364 + ], + "spans": [ + { + "bbox": [ + 167, + 245, + 282, + 364 + ], + "type": "image", + "image_path": "4a4d7c374adae22a074b2b3851f6cd29d8d19f06d2d9c83b39cfe9bcec541ba6.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 190, + 365, + 259, + 374 + ], + "lines": [ + { + "bbox": [ + 190, + 365, + 259, + 374 + ], + "spans": [ + { + "bbox": [ + 190, + 365, + 259, + 374 + ], + "type": "text", + "content": "(b) Illumination (direct)" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 460, + 287, + 520 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 460, + 287, + 520 + ], + "spans": [ + { + "bbox": [ + 46, + 460, + 287, + 520 + ], + "type": "text", + "content": "model augmented using the orientation loss from Verbin et al. [41], and then use this (potentially-imprecise) proxy geometry in our inverse rendering optimization pipeline. Please refer to the supplemental materials for a detailed description of our NeRF-based model." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 526, + 287, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 526, + 287, + 635 + ], + "spans": [ + { + "bbox": [ + 46, + 526, + 287, + 635 + ], + "type": "text", + "content": "Object self-occlusion does not contain enough signal. Figure 9 quantitatively demonstrates that only using the signal provided by object self-occlusions (\"without occluders\" plot in orange) cannot recover illumination as accurately as our method, which leverages the cue of unintended shadows (\"with occluders\" plot in blue). In particular, exploiting unintended shadows is increasingly important as the illumination contains higher frequencies (Figure 9a) and as the object material becomes increasingly diffuse (Figure 9b)." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 641, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 641, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 641, + 288, + 715 + ], + "type": "text", + "content": "Larger occluders improve illumination recovery. Figure 9c shows that increasing the size of the unseen occluders improves the accuracy of the recovered environment maps. Increasing the occluder size effectively improves the problem's conditioning by creating a larger variation in the incident lighting across points on the object." + } + ] + } + ], + "index": 11 + }, + { + "type": "table", + "bbox": [ + 307, + 245, + 544, + 289 + ], + "blocks": [ + { + "bbox": [ + 307, + 245, + 544, + 289 + ], + "lines": [ + { + "bbox": [ + 307, + 245, + 544, + 289 + ], + "spans": [ + { + "bbox": [ + 307, + 245, + 544, + 289 + ], + "type": "table", + "html": "
chairmannequinpotatoestoad
Direct spherical harmonics0.8310.6540.9970.943
Direct environment map0.0420.0710.0150.073
Ours0.0380.0350.0110.046
", + "image_path": "11fbac838b09598e683a6e5fcd4e413fa55a8cc3e366ec34743d0ac411a97bce.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_body" + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 293, + 546, + 361 + ], + "lines": [ + { + "bbox": [ + 305, + 293, + 546, + 361 + ], + "spans": [ + { + "bbox": [ + 305, + 293, + 546, + 361 + ], + "type": "text", + "content": "Table 1. Errors (RMSE) on all four of our scenes for our full model compared with our model without the spherical harmonic occluder parameterization, and without the pyramid environment parameterization. While the pyramid representation of the environment map improves smoothness (see Figure 10), we find that representing the masks using spherical harmonics is critical for our method." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 305, + 380, + 545, + 440 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 380, + 545, + 440 + ], + "spans": [ + { + "bbox": [ + 305, + 380, + 545, + 440 + ], + "type": "text", + "content": "Parameterization ablation study. Table 1 and Figure 10 show that representing the environment map as an image pyramid and the image masks with spherical harmonics, both perform better than simply optimizing their respective element values directly. See supplement for more details." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 449, + 375, + 462 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 449, + 375, + 462 + ], + "spans": [ + { + "bbox": [ + 306, + 449, + 375, + 462 + ], + "type": "text", + "content": "6. Discussion" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 470, + 546, + 649 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 470, + 546, + 649 + ], + "spans": [ + { + "bbox": [ + 304, + 470, + 546, + 649 + ], + "type": "text", + "content": "Even though unintended shadows are common in most real-world capture scenarios, they are often treated as outlier data. In this paper, we showed how to explicitly leverage these shadows as a signal to improve the quality of recovered lighting and materials in particularly challenging scenarios, such as objects made of diffuse materials. Furthermore, we have shown that our algorithm can be used with approximate geometry reconstructed by view synthesis techniques, and that it can produce imperfect, yet promising results on real captures, despite not being designed for this type of data which does not satisfy many of our model's assumptions. We believe that this work makes important first steps towards a general inverse rendering algorithm that can recover geometry, materials, and illuminations from images captured under realistic conditions." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 654, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 547, + 713 + ], + "type": "text", + "content": "Acknowledgements. Our sincere thanks to Delio Vicini, Rick Szeliski, Aleksander Holski, and Guy Satat for helpful discussions and suggestions, and to David Salesin, Janne Kontkanen, and Lior Yariv for their help in improving the manuscript before submission." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 299, + 749, + 310, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 749, + 310, + 757 + ], + "spans": [ + { + "bbox": [ + 299, + 749, + 310, + 757 + ], + "type": "text", + "content": "84" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 144 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 144 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 144 + ], + "type": "text", + "content": "[1] Sai Bangaru, Michael Gharbi, Tzu-Mao Li, Fujun Luan, Kalyan Sunkavalli, Milos Hasan, Sai Bi, Zexiang Xu, Gilbert Bernstein, and Fredo Durand. Differentiable rendering of neural SDFs through reparameterization. SIGGRAPH Asia, 2022. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 146, + 288, + 189 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 146, + 288, + 189 + ], + "spans": [ + { + "bbox": [ + 53, + 146, + 288, + 189 + ], + "type": "text", + "content": "[2] Manel Baradad, Vickie Ye, Adam B. Yedidia, Frédo Durand, William T. Freeman, Gregory W. Wornell, and Antonio Torralba. Inferring light fields from shadows. CVPR, 2018. 3, 4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 191, + 287, + 213 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 191, + 287, + 213 + ], + "spans": [ + { + "bbox": [ + 53, + 191, + 287, + 213 + ], + "type": "text", + "content": "[3] Jonathan T. Barron and Jitendra Malik. Shape, illumination, and reflectance from shading. TPAMI, 2015. 2, 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 213, + 287, + 257 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 213, + 287, + 257 + ], + "spans": [ + { + "bbox": [ + 53, + 213, + 287, + 257 + ], + "type": "text", + "content": "[4] Sai Bi, Zexiang Xu, Pratul P. Srinivasan, Ben Mildenhall, Kalyan Sunkavalli, Milos Hasan, Yannick Hold-Geoffroy, David Kriegman, and Ravi Ramamoorthi. Neural reflectance fields for appearance acquisition. arXiv, 2020. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 258, + 287, + 312 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 258, + 287, + 312 + ], + "spans": [ + { + "bbox": [ + 53, + 258, + 287, + 312 + ], + "type": "text", + "content": "[5] Sai Bi, Zexiang Xu, Kalyan Sunkavalli, Milovs Havsan, Yannick Hold-Geoffroy, David J. Kriegman, and Ravi Ramamoorthi. Deep Reflectance Volumes: Relightable Reconstructions from Multi-View Photometric Images. ECCV, 2020. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 313, + 287, + 356 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 313, + 287, + 356 + ], + "spans": [ + { + "bbox": [ + 53, + 313, + 287, + 356 + ], + "type": "text", + "content": "[6] Mark Boss, Raphael Braun, Varun Jampani, Jonathan T. Barron, Ce Liu, and Hendrik P. A. Lensch. NeRD: Neural reflectance decomposition from image collections. ICCV, 2021. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 357, + 287, + 401 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 357, + 287, + 401 + ], + "spans": [ + { + "bbox": [ + 53, + 357, + 287, + 401 + ], + "type": "text", + "content": "[7] Katherine L. Bouman, Vickie Ye, Adam B. Yedidia, Frédo Durand, Gregory W. Wornell, Antonio Torralba, and William T. Freeman. Turning corners into cameras: Principles and methods. ICCV, 2017. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 403, + 287, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 403, + 287, + 468 + ], + "spans": [ + { + "bbox": [ + 53, + 403, + 287, + 468 + ], + "type": "text", + "content": "[8] James Bradbury, Roy Frostig, Peter Hawkins, Matthew James Johnson, Chris Leary, Dougal Maclaurin, George Necula, Adam Paszke, Jake VanderPlas, Skye Wanderman-Milne, and Qiao Zhang. JAX: composable transformations of Python+NumPy programs, 2018. http://github.com/google/jax.7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 469, + 287, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 469, + 287, + 511 + ], + "spans": [ + { + "bbox": [ + 53, + 469, + 287, + 511 + ], + "type": "text", + "content": "[9] Paul Debevec. Rendering synthetic objects into real scenes: Bridging traditional and image-based graphics with global illumination and high dynamic range photography. SIGGRAPH, 1998. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 514, + 287, + 545 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 514, + 287, + 545 + ], + "spans": [ + { + "bbox": [ + 48, + 514, + 287, + 545 + ], + "type": "text", + "content": "[10] Xi Deng, Fujun Luan, Bruce Walter, Kavita Bala, and Steve Marschner. Reconstructing Translucent Objects using Differentiable Rendering. SIGGRAPH, 2022. 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 547, + 287, + 579 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 547, + 287, + 579 + ], + "spans": [ + { + "bbox": [ + 48, + 547, + 287, + 579 + ], + "type": "text", + "content": "[11] James R Driscoll and Dennis M Healy. Computing Fourier Transforms and Convolutions on the 2-Sphere. Advances in applied mathematics, 1994. 5" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 580, + 287, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 580, + 287, + 624 + ], + "spans": [ + { + "bbox": [ + 48, + 580, + 287, + 624 + ], + "type": "text", + "content": "[12] Marc-Andre Gardner, Kalyan Sunkavalli, Ersin Yumer, Xiaohui Shen, Emiliano Gambaretto, Christian Gagné, and Jean-François Lalonde. Learning to predict indoor illumination from a single image. ACM Trans. Graph., 2017. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 624, + 287, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 624, + 287, + 656 + ], + "spans": [ + { + "bbox": [ + 48, + 624, + 287, + 656 + ], + "type": "text", + "content": "[13] Yannick Hold-Geoffroy, Kalyan Sunkavalli, Sunil Hadap, Emiliano Gambaretto, and Jean-François Lalonde. Deep outdoor illumination estimation. CVPR, 2017. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 658, + 287, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 287, + 690 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 287, + 690 + ], + "type": "text", + "content": "[14] Wenzel Jakob, Sébastien Speierer, Nicolas Roussel, and Delio Vicini. Dr.Jit: A Just-In-Time Compiler for Differentiable Rendering. SIGGRAPH, 2022. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 692, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 692, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 692, + 287, + 713 + ], + "type": "text", + "content": "[15] Brian Karis and Epic Games. Real shading in unreal engine 4. Proc. Physically Based Shading Theory Practice, 2013. 1" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 34, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 95 + ], + "type": "text", + "content": "[16] P. Diederik Kingma and Lei Jimmy Ba. Adam: A Method for Stochastic Optimization. *ICLR*, 2015. 6, 7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 96, + 545, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 96, + 545, + 140 + ], + "spans": [ + { + "bbox": [ + 307, + 96, + 545, + 140 + ], + "type": "text", + "content": "[17] Zhengfei Kuang, Kyle Olszewski, Menglei Chai, Zeng Huang, Panos Achlioptas, and Sergey Tulyakov. Neroic: Neural rendering of objects from online image collections. ACM Trans. Graph., 2022. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 141, + 545, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 141, + 545, + 184 + ], + "spans": [ + { + "bbox": [ + 307, + 141, + 545, + 184 + ], + "type": "text", + "content": "[18] Zhengfei Kuang, Yunzhi Zhang, Hong-Xing Yu, Samir Agarwala, Shangzhe Wu, and Jiajun Wu. Stanford-ORB: A realworld 3d object inverse rendering benchmark. NeurIPS, 2023. 7, 1, 2, 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 186, + 545, + 218 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 186, + 545, + 218 + ], + "spans": [ + { + "bbox": [ + 307, + 186, + 545, + 218 + ], + "type": "text", + "content": "[19] Jean-François Lalonde, Alexei A Efros, and Srinivasa G Narasimhan. Estimating the natural illumination conditions from a single outdoor image. IJCV, 2012. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 220, + 545, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 220, + 545, + 262 + ], + "spans": [ + { + "bbox": [ + 307, + 220, + 545, + 262 + ], + "type": "text", + "content": "[20] Chloe LeGendre, Wan-Chun Ma, Graham Fyffe, John Flynn, Laurent Charbonnel, Jay Busch, and Paul Debevec. Deeplight: Learning illumination for unconstrained mobile mixed reality. CVPR, 2019. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 264, + 545, + 317 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 264, + 545, + 317 + ], + "spans": [ + { + "bbox": [ + 307, + 264, + 545, + 317 + ], + "type": "text", + "content": "[21] Chloe LeGendre, Wan-Chun Ma, Rohit Pandey, Sean Fanello, Christoph Rhemann, Jason Dourgarian, Jay Busch, and Paul Debevec. Learning illumination from diverse portraits. SIGGRAPH Asia 2020 Technical Communications, 2020. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 319, + 545, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 319, + 545, + 352 + ], + "spans": [ + { + "bbox": [ + 307, + 319, + 545, + 352 + ], + "type": "text", + "content": "[22] Tzu-Mao Li, Miika Aittala, Frédo Durand, and Jaakko Lehtinen. Differentiable monte carlo ray tracing through edge sampling. SIGGRAPH Asia, 2018. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 354, + 545, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 354, + 545, + 397 + ], + "spans": [ + { + "bbox": [ + 307, + 354, + 545, + 397 + ], + "type": "text", + "content": "[23] Zhengqin Li, Zexiang Xu, Ravi Ramamoorthi, Kalyan Sunkavalli, and Manmohan Chandraker. Learning to reconstruct shape and spatially-varying reflectance from a single image. SIGGRAPH Asia, 2018. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 399, + 545, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 399, + 545, + 442 + ], + "spans": [ + { + "bbox": [ + 307, + 399, + 545, + 442 + ], + "type": "text", + "content": "[24] Zhengqin Li, Mohammad Shafiei, Ravi Ramamoorthi, Kalyan Sunkavalli, and Manmohan Chandraker. Inverse rendering for complex indoor scenes: Shape, spatially-varying lighting and svbrdf from a single image. CVPR, 2020. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 444, + 545, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 444, + 545, + 487 + ], + "spans": [ + { + "bbox": [ + 307, + 444, + 545, + 487 + ], + "type": "text", + "content": "[25] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. NeRF: Representing Scenes as Neural Radiance Fields for View Synthesis. ECCV, 2020. 7, 8" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 488, + 545, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 488, + 545, + 520 + ], + "spans": [ + { + "bbox": [ + 307, + 488, + 545, + 520 + ], + "type": "text", + "content": "[26] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. SIGGRAPH, 2022. 8" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 522, + 545, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 522, + 545, + 566 + ], + "spans": [ + { + "bbox": [ + 307, + 522, + 545, + 566 + ], + "type": "text", + "content": "[27] Jacob Munkberg, Jon Hasselgren, Tianchang Shen, Jun Gao, Wenzheng Chen, Alex Evans, Thomas Mueller, and Sanja Fidler. Extracting Triangular 3D Models, Materials, and Lighting From Images. CVPR, 2022. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 567, + 545, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 567, + 545, + 599 + ], + "spans": [ + { + "bbox": [ + 307, + 567, + 545, + 599 + ], + "type": "text", + "content": "[28] Merlin Nimier-David, Delio Vicini, Tizian Zeltner, and Wenzel Jakob. Mitsuba 2: A retargetable forward and inverse renderer. SIGGRAPH Asia, 2019. 2" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 601, + 545, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 601, + 545, + 643 + ], + "spans": [ + { + "bbox": [ + 307, + 601, + 545, + 643 + ], + "type": "text", + "content": "[29] Merlin Nimier-David, Sébastien Speierer, Benoit Ruiz, and Wenzel Jakob. Radiative Backpropagation: An Adjoint Method for Lightning-Fast Differentiable Rendering. SIGGRAPH, 2020. 2" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 307, + 645, + 545, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 645, + 545, + 667 + ], + "spans": [ + { + "bbox": [ + 307, + 645, + 545, + 667 + ], + "type": "text", + "content": "[30] Ko Nishino and Shree K Nayar. Eyes for relighting. ACM Trans. Graph., 2004. 2" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 307, + 669, + 545, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 669, + 545, + 690 + ], + "spans": [ + { + "bbox": [ + 307, + 669, + 545, + 690 + ], + "type": "text", + "content": "[31] Jeong Joon Park, Aleksander Holynski, and Steven M Seitz. Seeing the world in a bag of chips. CVPR, 2020. 1, 3" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 307, + 692, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 692, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 692, + 545, + 713 + ], + "type": "text", + "content": "[32] Matt Pharr, Wenzel Jakob, and Greg Humphreys. Physically Based Rendering: From Theory to Implementation (3rd" + } + ] + } + ], + "index": 33 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 300, + 750, + 310, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 300, + 750, + 310, + 757 + ], + "spans": [ + { + "bbox": [ + 300, + 750, + 310, + 757 + ], + "type": "text", + "content": "85" + } + ] + } + ], + "index": 35 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 286, + 635 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 67, + 72, + 286, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 72, + 286, + 95 + ], + "spans": [ + { + "bbox": [ + 67, + 72, + 286, + 95 + ], + "type": "text", + "content": "ed.). Morgan Kaufmann Publishers Inc., San Francisco, CA, USA, 3rd edition, 2016. 1" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 96, + 286, + 118 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 96, + 286, + 118 + ], + "spans": [ + { + "bbox": [ + 48, + 96, + 286, + 118 + ], + "type": "text", + "content": "[33] Ravi Ramamoorthi and Pat Hanrahan. A signal-processing framework for inverse rendering. SIGGRAPH, 2001. 2, 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 119, + 286, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 119, + 286, + 140 + ], + "spans": [ + { + "bbox": [ + 48, + 119, + 286, + 140 + ], + "type": "text", + "content": "[34] Imari Sato, Yoichi Sato, and Katsushi Ikeuchi. Illumination from shadows. TPAMI, 2003. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 142, + 286, + 163 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 142, + 286, + 163 + ], + "spans": [ + { + "bbox": [ + 48, + 142, + 286, + 163 + ], + "type": "text", + "content": "[35] Shuran Song and Thomas Funkhouser. Neural illumination: Lighting prediction for indoor environments. CVPR, 2019. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 165, + 286, + 208 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 165, + 286, + 208 + ], + "spans": [ + { + "bbox": [ + 48, + 165, + 286, + 208 + ], + "type": "text", + "content": "[36] Pratul P. Srinivasan, Boyang Deng, Xiuming Zhang, Matthew Tancik, Ben Mildenhall, and Jonathan T. Barron. NeRV: Neural reflectance and visibility fields for relighting and view synthesis. CVPR, 2021. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 210, + 286, + 242 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 210, + 286, + 242 + ], + "spans": [ + { + "bbox": [ + 48, + 210, + 286, + 242 + ], + "type": "text", + "content": "[37] Tristan Swedish, Connor Henley, and Ramesh Raskar. Objects as cameras: Estimating high-frequency illumination from shadows. ICCV, 2021. 1, 3, 7, 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 243, + 286, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 243, + 286, + 297 + ], + "spans": [ + { + "bbox": [ + 48, + 243, + 286, + 297 + ], + "type": "text", + "content": "[38] Matthew Tancik, Pratul Srinivasan, Ben Mildenhall, Sara Fridovich-Keil, Nithin Raghavan, Utkarsh Singhal, Ravi Ramamoorthi, Jonathan Barron, and Ren Ng. Fourier features let networks learn high frequency functions in low dimensional domains. NeurIPS, 2020. 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 300, + 286, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 300, + 286, + 331 + ], + "spans": [ + { + "bbox": [ + 48, + 300, + 286, + 331 + ], + "type": "text", + "content": "[39] Antonio Torralba and William T. Freeman. Accidental pin-hole and pinspeck cameras: Revealing the scene outside the picture. CVPR, 2012. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 333, + 286, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 333, + 286, + 365 + ], + "spans": [ + { + "bbox": [ + 48, + 333, + 286, + 365 + ], + "type": "text", + "content": "[40] Eric Veach and Leonidas J Guibas. Optimally Combining Sampling Techniques for Monte Carlo Rendering. SIGGRAPH, 1995. 6, 4" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 367, + 286, + 410 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 367, + 286, + 410 + ], + "spans": [ + { + "bbox": [ + 48, + 367, + 286, + 410 + ], + "type": "text", + "content": "[41] Dor Verbin, Peter Hedman, Ben Mildenhall, Todd Zickler, Jonathan T. Barron, and Pratul P. Srinivasan. Ref-NeRF: Structured View-Dependent Appearance for Neural Radiance Fields. CVPR, 2022. 8" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 412, + 286, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 412, + 286, + 443 + ], + "spans": [ + { + "bbox": [ + 48, + 412, + 286, + 443 + ], + "type": "text", + "content": "[42] Delio Vicini, Sébastien Speierer, and Wenzel Jakob. Differentiable Signed Distance Function Rendering. SIGGRAPH, 2022. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 446, + 286, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 446, + 286, + 479 + ], + "spans": [ + { + "bbox": [ + 48, + 446, + 286, + 479 + ], + "type": "text", + "content": "[43] Bruce Walter, Stephen R Marschner, Hongsong Li, and Kenneth E Torrance. Microfacet models for refraction through rough surfaces. Eurographics, 2007. 6, 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 480, + 286, + 512 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 480, + 286, + 512 + ], + "spans": [ + { + "bbox": [ + 48, + 480, + 286, + 512 + ], + "type": "text", + "content": "[44] Adam B Yedidia, Manel Baradad, Christos Thrampoulidis, William T Freeman, and Gregory W Wornell. Using unknown occluders to recover hidden scenes. CVPR, 2019. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 514, + 286, + 546 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 514, + 286, + 546 + ], + "spans": [ + { + "bbox": [ + 48, + 514, + 286, + 546 + ], + "type": "text", + "content": "[45] Tizian Zeltner, Sébastien Speierer, Iliyan Georgiev, and Wenzel Jakob. Monte Carlo Estimators for Differential Light Transport. SIGGRAPH, 2021. 2, 4" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 548, + 286, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 548, + 286, + 590 + ], + "spans": [ + { + "bbox": [ + 48, + 548, + 286, + 590 + ], + "type": "text", + "content": "[46] Kai Zhang, Fujun Luan, Qianqian Wang, Kavita Bala, and Noah Snively. PhySG: Inverse rendering with spherical gaussians for physics-based material editing and relighting. CVPR, 2021. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 592, + 286, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 592, + 286, + 635 + ], + "spans": [ + { + "bbox": [ + 48, + 592, + 286, + 635 + ], + "type": "text", + "content": "[47] Xiuming Zhang, Pratul P. Srinivasan, Boyang Deng, Paul Debevec, William T. Freeman, and Jonathan T. Barron. NeR-Factor: Neural factorization of shape and reflectance under an unknown illumination. SIGGRAPH Asia, 2021. 2" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 300, + 750, + 310, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 300, + 750, + 310, + 757 + ], + "spans": [ + { + "bbox": [ + 300, + 750, + 310, + 757 + ], + "type": "text", + "content": "86" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Edge-Aware 3D Instance Segmentation Network with Intelligent Semantic Prior/ef85b90b-3f1b-4322-ac5a-c5d6ae60802c_content_list.json b/2024/Edge-Aware 3D Instance Segmentation Network with Intelligent Semantic Prior/ef85b90b-3f1b-4322-ac5a-c5d6ae60802c_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..d888fe25d19c8b7e08ab622e1314ba05c232d57d --- /dev/null +++ b/2024/Edge-Aware 3D Instance Segmentation Network with Intelligent Semantic Prior/ef85b90b-3f1b-4322-ac5a-c5d6ae60802c_content_list.json @@ -0,0 +1,1431 @@ +[ + { + "type": "text", + "text": "Edge-Aware 3D Instance Segmentation Network with Intelligent Semantic Prior", + "text_level": 1, + "bbox": [ + 83, + 130, + 887, + 151 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Wonseok Roh1 Hwanhee Jung1 Giljoo Nam2 Jinseop Yeom1 Hyunjep Park1 Sang Ho Yoon3 Sanggil Kim1* \n1Korea University 2Meta Reality Labs 3KAIST", + "bbox": [ + 217, + 179, + 750, + 244 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 273, + 312, + 287 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "While recent 3D instance segmentation approaches show promising results based on transformer architectures, they often fail to correctly identify instances with similar appearances. They also ambiguously determine edges, leading to multiple misclassifications of adjacent edge points. In this work, we introduce a novel framework, called EASE, to overcome these challenges and improve the perception of complex 3D instances. We first propose a semantic guidance network to leverage rich semantic knowledge from a language model as intelligent priors, enhancing the functional understanding of real-world instances beyond relying solely on geometrical information. We explicitly instruct the basic instance queries using text embeddings of each instance to learn deep semantic details. Further, we utilize the edge prediction module, encouraging the segmentation network to be edge-aware. We extract voxel-wise edge maps from point features and use them as auxiliary information for learning edge cues. In our extensive experiments on large-scale benchmarks, ScanNetV2, ScanNet200, S3DIS, and STPLS3D, our EASE outperforms existing state-of-the-art models, demonstrating its superior performance.", + "bbox": [ + 75, + 292, + 473, + 609 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 628, + 207, + 643 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Understanding 3D scenes is a fundamental task within 3D computer vision. Given 3D point cloud scenes, identifying and perceiving instances on sparse points, along with assigning semantic class labels, play a crucial role in a comprehensive understanding of the entire spatial environment.", + "bbox": [ + 75, + 652, + 468, + 729 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In real-world 3D scenarios, significant occlusion and truncation often occur, especially when objects are overlapped or hidden by others. To address these issues, conventional works [3, 7, 8, 16, 18, 40, 43, 44] in 3D instance segmentation (3DIS) primarily focus on accurately generating region proposals (top-down) [16, 43, 44] or effectively grouping points with clustering algorithms (bottom-up) [3, 7, 8, 18, 40]. Inspired by the remarkable sensation of Mask-RCNN [14], the former (proposal-based) meth", + "bbox": [ + 75, + 729, + 468, + 864 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/ef5c115d04f726dae5371777545edb2e15643339b4bcb8ae951cc343d6944cce.jpg", + "image_caption": [ + "(b) Instance misinterpretation cases", + "Figure 1. Examples of 3DIS for two challenging cases. (a) Semantic failure cases: the baseline [35] confuses seat-shaped objects as chair or toilet (Scene1) and also misclassifies large cuboid blocks as desk (Scene2). And (b) Instance misinterpretation cases: baseline erroneously segments a single object into multiple parts (Scene3) or merges multiple objects into one instance (Scene4)." + ], + "image_footnote": [], + "bbox": [ + 506, + 273, + 890, + 580 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ods initially detect instances as bounding boxes and predict masks within each proposed region. However, these strategies are susceptible to the quality of the detection results. On the other hand, the latter (grouping-based) methods use clustering algorithms to group closely related points and aggregate point-wise class labels and instance features. While they show great advancements in 3DIS, these methods require additional manually tuned processing, such as point grouping [18] or voting mechanisms [31], to determine specific geometric properties (e.g., centers, occupancy).", + "bbox": [ + 496, + 688, + 892, + 839 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recently, transformer-based 3DIS frameworks [22, 26, 35, 36] have tackled several limitations of traditional approaches by introducing a fully end-to-end pipeline. They directly predict masks from 3D points without resource-", + "bbox": [ + 496, + 839, + 890, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 807, + 44 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author. \nProject Page: https://kuai-lab.github.io/ease2024.", + "bbox": [ + 94, + 875, + 455, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "20644", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "exhaustive processing. These methods have achieved high performance and quickly become predominant in the 3DIS domain. However, despite these breakthroughs, they still face challenges in accurately masking everyday 3D instances relying solely on geometrical information. Specifically, they frequently suffer from cases where each instance appears similarly, but their semantic roles differ, as illustrated in Fig. 1 (a). Thus, gaining insight into the semantic properties of instances beyond mere appearance or position is crucial for effectively addressing complex real-world 3D scenarios. From this observation, we consider leveraging semantic knowledge along with the geometry information.", + "bbox": [ + 75, + 90, + 472, + 272 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Built upon the classic structure of the transformer-based framework [22, 35, 36], which trains queries that encode instance-specific knowledge for direct mask prediction, we introduce a carefully designed semantic guidance network. One of the most promising ways to utilize semantic information is through visual language models [17, 32]. They learn generous visual-linguistic knowledge from large-scale image-text pairs, which empowers them to achieve outstanding progress in various 2D or 3D vision tasks [9, 11, 23, 29]. In this work, motivated by the success of these works, we explicitly instruct the transformer-based network to learn contextual variations among instances using text embeddings. Specifically, we strengthen the basic instance queries with profound semantic clues of text embeddings representing each instance. Here, the context details can serve as discriminative semantic priors for perceiving complex 3D instances adequately. Through our extensive experiments, we found that our semantic-guided approach effectively overcomes the challenges of existing methods.", + "bbox": [ + 75, + 275, + 472, + 561 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Different from the human vision system, which easily distinguishes boundaries between instances, we empirically observe that most 3DIS models [21, 22, 35, 36, 42] struggle with precisely capturing the boundaries in the 3D space. These methods often misinterpret the spatial range of instances and ambiguously determine their edges, leading to numerous misclassifications of adjacent edge points, as shown in Fig. 1 (b). We consider that these issues arise due to the absence of detailed guidance on inter-object boundaries, resulting in fuzzy edges. To tackle these challenges, we advocate for leveraging the edge prediction module to pilot the whole network to exploit edge-advanced features. We predict edge points across point cloud scenes and operate them as auxiliary information for learning practical edge cues, which encourages accurate recognition of the 3D spatial scope. Note that we supervise this module using dynamically generated point-wise pseudo edge labels.", + "bbox": [ + 75, + 565, + 472, + 821 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Given landmark datasets for 3DIS, ScanNetV2 [4], ScanNet200 [33], S3DIS [1], and STPLS3D [2], we validate the effectiveness and robustness of our framework. Our method outperforms the existing state-of-the-art methods. To summarize, our main contributions are listed as follows:", + "bbox": [ + 75, + 824, + 472, + 902 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose EASE, a novel 3D instance segmentation framework that utilizes the rich semantic clues from the language model. Specifically, context details of text embeddings serve as smart semantic priors, effectively enhancing the functional comprehension of 3D instances.", + "- We introduce the edge prediction module guiding the entire network to take advantage of edge-aware features. Also, we utilize the module's output as auxiliary knowledge for learning edge cues, boosting 3DIS performance.", + "- We analyze the effectiveness of our proposed method on multiple challenging benchmarks, including ScanNetV2, ScanNet200, S3DIS, and STPLS3D. Extensive experiments on various 3D scenarios validate that our method achieves new State-of-the-Art performance on 3DIS." + ], + "bbox": [ + 500, + 90, + 893, + 303 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 316, + 640, + 333 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3D Instance Segmentation. The 3D Instance Segmentation (3DIS) task aims to identify individual instances and assign semantic classes to each point within a 3D point scene. Traditional approaches in 3DIS are primarily categorized into three groups: proposal-based [16, 43, 44], grouping-based [3, 7, 8, 18, 40], and transformer-based methods [22, 35, 36]. The proposal-based approaches initially detect object proposals, such as 3D bounding boxes, and then estimate corresponding per-point semantic classes and instance masks. One of the grouping-based approaches, PointGroup [18], conducts point-wise clustering to group points into instances using dual point sets, which comprise original and shifted coordinates. Recently, transformer-based approaches predict instance queries that represent individual objects and decode them into per-point categories and instance labels directly. Mask3D [35] refines queries using masked cross-attention with hierarchical point features, leading the queries to focus on particular instances. SPFormer [36] updates instance queries via transformer decoder utilizing pre-computed superpoints, which can reduce the whole computational cost of the network. MAFT [22], on the other hand, recognizes representational limitations in the initial mask and replaces the masking process with object-localizing position queries. Based on the core idea of transformer-based methods, our novel approach incorporates a semantic guidance network and an auxiliary edge network to improve instance mask segmentation.", + "bbox": [ + 496, + 338, + 893, + 746 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Semantic Guided 3D Scene Understanding. Recent research in 3D scene understanding focuses on comprehending large-scale, real-world 3D environments. Conventional studies in this field are broadly categorized into two strategies: closed-set 3D scene understanding, a prevalent approach in diverse tasks [19, 27, 33, 34], and open-vocabulary methods [12, 29, 37], which focus on recognizing unseen categories within datasets. More recently, there has been significant progress in integrating pre-trained visual-language models like CLIP [32] or ALIGN [17] into", + "bbox": [ + 496, + 750, + 895, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "20645", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/6a7ca8f42b59b14a63cb9d7bdcdb2e6781b2dcae5dba120aefe28498d2ca4a84.jpg", + "image_caption": [ + "Figure 2. An overview of our framework EASE. Built upon the classic architecture of the transformer-based 3DIS model, our model takes 3D scenes and directly infers instance masks. Our model consists of four main modules: (1) Sparse Convolutional Backbone, (2) Semantic Network, which strengthen the queries to learn contextual variations of instances with rich semantic embeddings from the pre-trained language model CLIP, (3) Mask Transformer Decoder with Mask Module and Query Refinement blocks, and (4) Edge Prediction Module, where our edge extraction head is trained to predict a voxel-wise edge map, encouraging the network to utilize edge-advanced features." + ], + "image_footnote": [], + "bbox": [ + 84, + 94, + 890, + 392 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "these methodologies to enhance semantic understanding. Building on this idea, for instance, [33] utilizes a closed-set framework and enhances 3D feature learning by aligning it with CLIP text embedding space for semantic classification. In this work, we employ contextual semantic priors to improve the comprehension of complex 3D instances, enhancing the overall 3D instance segmentation performance.", + "bbox": [ + 75, + 487, + 468, + 592 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 76, + 609, + 168, + 625 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we introduce a novel 3D Instance Segmentation (3DIS) framework EASE leveraging the rich semantic guidance from the language model. Furthermore, we propose the edge prediction module, which encourages the pipeline to use edge-aware features for effectively perceiving instance boundaries. We provide an overview of the entire pipeline (Sec. 3.1, Fig. 2, and Alg. 1) and then elaborate on the details of our method: (i) Intelligent Semantic Prior (Sec. 3.2) and (ii) Edge-Aware 3DIS Framework (Sec. 3.3).", + "bbox": [ + 75, + 636, + 468, + 772 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Overview", + "text_level": 1, + "bbox": [ + 76, + 785, + 187, + 800 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our end-to-end 3DIS framework, as illustrated in Fig. 2 and Alg. 1, aims to enhance the comprehension of various instances across 3D point cloud scenes. Based on the classic transformer-based architecture [22, 35], which operates standard transformer building blocks, we directly infer instance masks from 3D point clouds. Our model consists", + "bbox": [ + 75, + 809, + 470, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "of four main modules: (1) Sparse Convolutional Backbone, (2) Semantic Network, which provide the deep semantic cues of diverse real-world instances, (3) Mask Transformer Decoder with Mask Module and Query Refinement blocks, and (4) Edge Prediction Module, which benefits the whole network to take advantage of edge-aware features.", + "bbox": [ + 496, + 487, + 890, + 577 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Mask Transformer. First, the Sparse Convolutional U-Net Backbone takes a colored point cloud $P \\in \\mathbb{R}^{N_p \\times 6}$ as input and voxelizes $P$ into $X_0$ voxels $V \\in \\mathbb{R}^{X_0 \\times 3}$ to extract a multi-resolution hierarchical feature maps $F_i \\in \\mathbb{R}^{X_i \\times D}$ , where $i \\in \\{0,1,2,3,4\\}$ . Following [28, 35], we set zero-initialized non-parametric instance queries $Q \\in \\mathbb{R}^{N_q \\times D}$ , referring to point positions sampled with furthest point sampling (FPS) [30]. Given the $F_i$ and $Q$ , the mask transformer decoder layer iteratively enhances the queries using the Mask Module (MM) and Query Refinement (QR) blocks. In the MM, as shown in Fig. 3 (a), we classify the category $c_i$ for $i = \\{1,2,\\dots,N_c\\}$ of each query via linear classification head $f_{\\mathrm{class}}$ using following cross-entropy loss:", + "bbox": [ + 496, + 580, + 892, + 777 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {c l s} = - \\mathbb {E} _ {c, w _ {c} \\sim \\mathbb {D}} \\left[ \\sum_ {r \\in N _ {c}} w _ {c} [ r ] \\log f _ {\\text {c l a s s}} (Q) [ r ] \\right] \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 526, + 787, + 890, + 829 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $w_{c}$ denotes one-hot encoded category labels and $\\mathbb{D}$ represents (input) data distribution. Also, the $Q$ are fed through the query head $f_{\\mathrm{query}}$ to project them to the same feature space as $F_{0}$ for mask prediction. Afterwards, we", + "bbox": [ + 496, + 839, + 892, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "20646", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 2 + }, + { + "type": "code", + "sub_type": "algorithm", + "code_caption": [ + "Algorithm 1: Overview of our framework EASE." + ], + "code_body": "Input: Colored point cloud $P\\in \\mathbb{R}^{N_p\\times 6}$ Result: Binary 3DIS mask $\\hat{M}_f\\in \\{0,1\\}^{X\\times N_q}$ Procedure: # Sparse Convolutional Backbone. 1 Extract multi-resolution feature maps $F_{i}$ from $P$ # Edge Prediction Module (EPM). 2 Predict edge map $\\hat{E}$ using full-resolution feature $F_{0}$ 3 Supervise $\\hat{E}$ with pseudo edge label $E$ 4 Encourage the network to learn edge-aware features. #Mask Transformer Decoder. 5 for $l\\in N_l$ do 6 for $i\\in \\{1,2,3,4\\}$ do 7 Update queries $Q$ using $M_{attn}$ and $F_{i}$ # Semantic Network. 8 Match the set of classes with text labels. 9 Encode text into text embeds $T$ using CLIP. 10 Reinforce $Q$ to contain semantic clues of $T$ 11 end 12 end", + "bbox": [ + 78, + 111, + 468, + 411 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "compute the similarity between the projected query features $Q'$ and $F_0$ using the dot product operation, then calculate the probability of the instance mask employing the sigmoid function as follows.", + "bbox": [ + 76, + 426, + 468, + 486 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nM _ {a t t n} = \\left\\{m _ {i, j} = \\left[ \\sigma \\left(F _ {0} \\cdot f _ {\\text {q u e r y}} (Q) ^ {T}\\right) _ {i, j} > 0. 5 \\right] \\right\\} \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 99, + 497, + 468, + 516 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where the threshold value is 0.5 for binary mask. Further, to refine the query representation, we leverage the $QR$ block, including masked cross-attention and self-attention mechanisms. Here, we utilize $M_{att}$ as the foreground mask for the masked cross-attention layer, where $Q$ iteratively attend to each multi-scale feature $(F_1 - F_4, i \\geqslant 1)$ as follows:", + "bbox": [ + 76, + 526, + 468, + 617 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nQ = \\operatorname {s o f t m a x} \\left(\\mathbf {Q K} ^ {T} / \\sqrt {D} + M _ {\\text {a t t n}}\\right) \\mathbf {V} \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 143, + 627, + 468, + 646 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathbf{K}$ and $\\mathbf{V}$ are linearly projected from each point feature, and $\\mathbf{Q}$ are from $Q$ . Hence, the transformer focuses more on valuable instances instead of the unessential background. Subsequently, in the self-attention layer, the keys, values, and queries are all linear projections of $Q$ from the cross-attention layer. Finally, the transformer decoder layer is recurrently applied for multiple iterations $N_{l}$ and feature scales $i$ , ultimately producing the final set of refined queries.", + "bbox": [ + 76, + 656, + 468, + 777 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Intelligent Semantic Prior", + "text_level": 1, + "bbox": [ + 76, + 786, + 313, + 801 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "It is important to highlight that precise geometric properties (e.g., locations, distances and orientations) from 3D points significantly enhance 3DIS accuracy. While previous studies [22, 35, 36] benefit from the generous geometric potential of physical information, they are still limited in perceiving intricate real-world 3D instances. Hence, we pro", + "bbox": [ + 76, + 810, + 468, + 900 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/b48eadeeee65a1d6a26e024a61b87e555937b855f6be5210dd54b4b0144e0d77.jpg", + "image_caption": [ + "(a) Transformer Decoder Layer" + ], + "image_footnote": [], + "bbox": [ + 506, + 89, + 887, + 172 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/2a3295bc7d96b84dc109c8b3b80d0cc7cd2ddf67fc5a276e6ff747e0d144deee.jpg", + "image_caption": [ + "(b) Semantic Network", + "Figure 3. Detailed architecture of (a) Transformer Decoder Layer: consists of Mask Module and Query Refinement blocks. And (b) Semantic Network: we first map the categories of queries $Q$ with the corresponding text description for text embeddings $T$ from CLIP [32]. Then, we explicitly strengthen the $Q$ with $T$ using the Semantic Fusion network to learn semantic details." + ], + "image_footnote": [], + "bbox": [ + 506, + 189, + 887, + 256 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "pose to utilize semantic knowledge of individual instances as smart priors, effectively addressing the challenges.", + "bbox": [ + 498, + 368, + 890, + 398 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Deep Semantic Knowledge for 3D Instances. Recently, visual language models [17, 32] have highlighted the benefits of their capability to provide general embedded knowledge from large-scale image-text multimodal data. In this work, we explicitly instruct the transformer-based 3DIS network to learn contextual variations among instances by utilizing the power of the prevalent language model CLIP [32]. To achieve this goal, we utilize a carefully designed Semantic Network, as shown in Fig. 3 (b). Specifically, with the refined query $Q$ from each $QR$ block, we first categorize each query following the linear classification head $f_{\\mathrm{class}}$ in the MM. Given the set of category candidates, we map the category number $c_{i}$ for $i = \\{1,2,\\dots,N_c\\}$ with the corresponding text descriptions (e.g., \"Cabinet\", \"Bookshelf\"). Next, these text labels go through the large-scale pre-trained language model, which outputs text embedding vectors $T$ . We then concatenate $Q$ with $T$ and strengthen the basic instance queries with deep semantic clues from text embeddings using the Semantic Fusion network as follows:", + "bbox": [ + 498, + 401, + 890, + 686 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nQ = \\operatorname {S i g m o i d} \\left(\\mathbf {W} _ {2} \\cdot \\phi \\left(\\mathbf {W} _ {1} \\cdot (Q \\oplus T)\\right) ^ {T}\\right) \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 557, + 694, + 890, + 712 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\oplus$ indicates channel-wise concatenation, $\\mathbf{W}_i$ denotes the learnable parameters of the $i$ -th linear layer, and $\\phi$ is the non-linear activation function. We apply this process for each iteration of the $QR$ . Ultimately, we consider these explicitly guided queries with contextual variations as discriminative semantic priors for the practical perception of complex 3D instances, enhancing 3DIS accuracy.", + "bbox": [ + 496, + 717, + 890, + 823 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Edge-Aware 3DIS Framework", + "text_level": 1, + "bbox": [ + 500, + 832, + 767, + 848 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In real-world 3D scenarios, instances are commonly positioned in diverse arrangements without following standard rules or patterns. Especially when instances are closely ad", + "bbox": [ + 498, + 854, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "20647", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/a2ea6740bf0d611246be912d5ba9e3a51ed2ef290a51c7db716348dd7eeb2998.jpg", + "image_caption": [ + "Figure 4. Pseudo edge label calculation using KNN algorithm. We collect $k$ neighboring points of each point (red border) and compare their instance labels. If the count of different label points (blue and purple) among its surroundings surpasses a predefined threshold $\\tau$ , we annotate the central point as an edge point (green)." + ], + "image_footnote": [], + "bbox": [ + 83, + 93, + 462, + 195 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "jacent or overlapping, it becomes increasingly challenging for the model to define the edges precisely. Therefore, even with significant progress in technical strategies, most conventional methods [22, 35, 36] often misinterpret the spatial extent of each instance and ambiguously estimate their edges as illustrated in Fig. 1. To this end, we introduce an edge-aware 3DIS framework with an edge prediction module, enabling the network to utilize edge-advanced features.", + "bbox": [ + 75, + 291, + 468, + 412 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Pseudo Edge Label Calculation. To optimize the edge prediction module, we first compute point-wise pseudo edge labels for all 3D point scenes. We operate a traditional $k$ -Nearest Neighbors (KNN) classification algorithm based on the KD-tree data structure to explore the $k$ nearest points. Here, we utilize point-wise instance labels from datasets. For all points in a 3D scene, we set each point as a central point and compare their instance labels with those of $k$ neighbor points. As shown in Fig. 4, if the count of distinct label points among its surroundings exceeds a predefined threshold $\\tau$ , we identify the central point as a boundary point. We finally generate pseudo-binary edge maps $E' \\in \\{0,1\\}^{N_p \\times 1}$ , where points on the edges are 1 and internal points are 0 for the entire set of points $P$ . Then, we voxelize the point-wise edge labels into the voxel-wise edge labels $E \\in \\{0,1\\}^{X_0 \\times 1}$ to supervise the edge extraction head. Note that we precompute the pseudo edge labels for every 3D scene in the datasets before training.", + "bbox": [ + 75, + 414, + 468, + 686 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Learning Edge Cues. With dynamically generated pseudo edge labels $E$ , we aim to enrich regular features to recognize the accurate spatial scope of 3D instances with precise boundaries. Specifically, we introduce an Edge Prediction Module, which predicts a voxel-wise dense edge map $\\hat{E} \\in (0,1)^{X_0 \\times 1}$ to interactively guide the network in capturing edges of various instances. Given a full-resolution feature map $F_0$ from the backbone network, our edge extraction head, which consists of multiple shared MLPs, estimates the probabilities of edges for each voxel as follows:", + "bbox": [ + 76, + 688, + 468, + 840 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {E} = \\operatorname {M L P} _ {\\text {e d g e}} \\left(F _ {0}\\right) = \\operatorname {S i g m o i d} \\left(\\mathbf {W} _ {2} \\cdot \\phi \\left(\\mathbf {W} _ {1} \\cdot F _ {0}\\right) ^ {T}\\right) \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 89, + 845, + 468, + 864 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathbf{W}_i$ denotes the learnable parameters of the $i$ -th linear layer, and $\\phi$ indicates the non-linear activation function.", + "bbox": [ + 76, + 869, + 468, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Following [6, 10], as edge voxels constitute a small portion of the entire voxel set, we formulate weighted binary cross entropy loss to train the edge extraction head as follows:", + "bbox": [ + 498, + 90, + 890, + 136 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {e d g e} = - \\sum_ {i = 1} ^ {X _ {0}} \\left(w \\cdot E _ {i} \\log \\hat {E} _ {i} + \\left(1 - E _ {i}\\right) \\log \\left(1 - \\hat {E} _ {i}\\right)\\right) (6)\n$$\n", + "text_format": "latex", + "bbox": [ + 506, + 143, + 890, + 184 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $w$ represent the weight value used to balance the substantial difference between the numbers of edge points and others. Then, we propagate the edge cues of prediction as auxiliary information for the backbone network, improving 3D instance understanding with edge-aware features. Furthermore, we train our edge network jointly in an end-to-end manner with low computational costs.", + "bbox": [ + 498, + 188, + 890, + 294 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Loss Function We finally predict instance mask $\\hat{M}_f$ using the final set of queries in the $MM$ . For training, we compute the following loss $\\mathcal{L}_{mask}$ as a sum of the two losses:", + "bbox": [ + 498, + 297, + 890, + 342 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {m a s k}} = \\lambda_ {\\mathrm {B C E}} \\mathcal {L} _ {\\mathrm {B C E}} \\left(M _ {f}, \\hat {M} _ {f}\\right) + \\lambda_ {\\mathrm {d i c e}} \\mathcal {L} _ {\\mathrm {d i c e}} \\left(M _ {f}, \\hat {M} _ {f}\\right) \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 508, + 352, + 890, + 371 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $M_{f}$ denotes the ground-truth instance mask, and $\\mathcal{L}_{\\text{dice}}$ represents the Dice loss [5]. Ultimately, our model is trained end-to-end by minimizing the following loss $\\mathcal{L}_{\\text{total}}$ :", + "bbox": [ + 500, + 380, + 890, + 426 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {t o t a l}} = \\lambda_ {\\text {m a s k}} \\mathcal {L} _ {\\text {m a s k}} + \\lambda_ {\\text {c l s}} \\mathcal {L} _ {\\text {c l s}} + \\lambda_ {\\text {e d g e}} \\mathcal {L} _ {\\text {e d g e}} \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 524, + 440, + 890, + 458 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where each $\\lambda$ is a hyperparameter derived from grid searches to handle the strength of respective loss term.", + "bbox": [ + 498, + 467, + 890, + 497 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 500, + 510, + 632, + 527 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Experimental Setup", + "text_level": 1, + "bbox": [ + 500, + 535, + 689, + 551 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Datasets. In this study, we train and evaluate the overall performance using four landmark datasets for 3D instance segmentation: ScanNetV2 [4], ScanNet200 [33], S3DIS [1], and STPLS3D [2]. These four datasets provide 3D point cloud scan data collected in various scenarios. Note that detailed descriptions of each dataset and all implementation details are provided in the supplemental material.", + "bbox": [ + 496, + 561, + 890, + 667 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Evaluation Metrics. To evaluate the 3D instance segmentation accuracy, we operate Average Precision (AP), a typical metric for various computer vision tasks. We report mean average precision (mAP), which represents an average score with IoU thresholds ranging from $50\\%$ to $95\\%$ with a step size of $5\\%$ . Also, we provide $\\mathrm{mAP}_{50}$ and $\\mathrm{mAP}_{25}$ , indicating the scores with IoU thresholds of $50\\%$ and $25\\%$ , respectively. Further, we estimate mean precision (mPrec) and mean recall (mRec) on the S3DIS [1] dataset.", + "bbox": [ + 496, + 669, + 890, + 805 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Performance Comparison with SOTA Methods.", + "text_level": 1, + "bbox": [ + 498, + 814, + 890, + 830 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "ScanNetV2. We quantitatively compare our proposed network EASE with existing state-of-the-art methods. These methods demonstrate significant capabilities, but they still rely on geometric information from 3D points and often", + "bbox": [ + 496, + 839, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "20648", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/a02187e0001d23e7c2a90fdadb723b34920123c26a6ede3b2a14682fccfb8ca4.jpg", + "table_caption": [], + "table_footnote": [ + "Table 1. Comparison of 3D Instance Segmentation performance with state-of-the-art approaches on the ScanNetV2 [4] dataset. We evaluate mean average precision (mAP) with different IoU thresholds over 18 classes on the ScanNetV2 validation / hidden test set." + ], + "table_body": "
MethodScanNet ValScanNet Test
mAP\\( mAP_{50} \\)mAP\\( mAP_{50} \\)
GSPN [44]19.337.8-30.6
3D-SIS [16]-18.716.138.2
MASC [25]--25.444.7
3D-Bonet [43]--25.348.8
MTML [20]20.340.228.254.9
3D-MPA [8]35.559.135.561.1
DyCo3D [15]35.457.639.564.1
PointGroup [18]34.856.740.763.6
MaskGroup [45]42.063.343.466.4
OccuSeg [13]44.260.748.667.2
SSTNet [24]49.464.350.669.8
HAIS [3]43.564.145.769.9
SoftGroup [39]46.067.650.476.1
Mask3D [35]55.273.756.678.0
QueryFormer [26]56.574.258.378.7
MAFT [22]59.976.559.678.6
EASE (Ours)60.277.259.878.9
", + "bbox": [ + 81, + 88, + 472, + 388 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "face challenges in effectively comprehending 3D instances. To address these limitations, we present a novel edge-aware framework featuring a semantic guidance network. As reported in Tab. 1, EASE generally outperforms other methods, achieving new state-of-the-art accuracy in terms of $\\mathrm{mAP}$ and $\\mathrm{mAP}_{50}$ on both validation (60.2 / 77.2) and hidden test (59.8 / 78.9) sets. These results confirm that our edge cues and smart semantic prior benefits the entire network, leading to high precision in 3D instance segmentation.", + "bbox": [ + 75, + 462, + 468, + 599 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "S3DIS. In Tab. 2, we evaluate 3DIS performance on Area 5 and 6-fold cross-validation of the S3DIS [1] dataset. For Area 5 evaluation, we employ data from Area 5 for validation and the other areas for training. Also, for 6-fold cross-validation, we assess validation scores across six other areas and compute the average. In both evaluations, EASE demonstrates significant performance improvements up to $+2.4 / 2.5$ (Area 5) and $+0.8 / 0.9$ (6-fold) for $\\mathrm{mAP} / \\mathrm{mAP}_{50}$ .", + "bbox": [ + 75, + 602, + 468, + 726 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "ScanNet200. We also demonstrate considerable performance on the ScanNet200 validation set (see Tab. 3). Our method EASE precisely segments 3D instances for various categories compared to current state-of-the-art approaches.", + "bbox": [ + 75, + 729, + 468, + 792 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "STPLS3D. Remarkably, our EASE network also achieves higher scores on the STPLS3D [2] dataset, which consists of outdoor 3D point cloud scenes. EASE exceeds the second-best results with $+1.1 / 1.6 / 1.3$ margins in $\\mathrm{mAP} / \\mathrm{mAP}_{50} / \\mathrm{mAP}_{25}$ . These results highlight the effectiveness of our method in understanding diverse real-world 3D instances, both from indoor and outdoor environments.", + "bbox": [ + 75, + 794, + 470, + 902 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/85e311bd0fd4af619c41fba4c9a12321289aa18ae7cce901f07dfd848e2a70aa.jpg", + "table_caption": [], + "table_footnote": [ + "Table 2. Comparison of 3D Instance Segmentation performance with state-of-the-art approaches on the S3DIS [1] Area 5 and 6-fold cross validation set. We evaluate mAP across different IoU thresholds, along with mean precision (mPrec) and mean recall (mRec) at a $50\\%$ IoU threshold over 13 classes of the S3DIS." + ], + "table_body": "
MethodS3DIS Area 5S3DIS 6-fold CV
AP\\( AP_{50} \\)\\( Prec_{50} \\)\\( Rec_{50} \\)AP\\( AP_{50} \\)\\( Prec_{50} \\)\\( Rec_{50} \\)
ASIS [41]--55.342.4--63.647.5
3D-Bonet [43]--57.540.2--65.647.6
3D-MPA [8]--63.158.0--66.764.1
PointGroup [18]-57.861.962.1-64.069.669.2
DyCo3D [15]--64.364.2----
MaskGroup [45]-65.062.964.7-69.966.669.6
SSTNet [24]42.759.365.564.254.167.873.573.4
SoftGroup [39]51.666.173.666.654.468.975.369.8
Mask3D [35]56.668.468.766.364.575.572.874.5
MAFT [22]-69.1------
QueryFormer [26]57.769.970.572.262.073.372.773.4
EASE (Ours)59.071.669.468.765.376.473.674.6
", + "bbox": [ + 498, + 88, + 906, + 311 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/781e596597928f1eb5777e44a85425c20be20691f8491d3ae7a3c3d73c0f83d7.jpg", + "table_caption": [], + "table_footnote": [ + "Table 3. Comparison of 3D Instance Segmentation performance with state-of-the-art approaches on the ScanNet200 [33] val set." + ], + "table_body": "
MethodmAP\\( \\mathrm{mAP}_{50} \\)\\( \\mathrm{mAP}_{25} \\)
SPFormer [36]25.233.839.6
Mask3D [35]27.437.042.3
QueryFormer [26]28.137.143.4
MAFT [22]29.238.243.3
EASE (Ours)29.938.844.7
", + "bbox": [ + 500, + 398, + 903, + 507 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/f21c2fc81c8bb4dce875e3d5767b8d7fbf7dcc1745282fbd7eaac6bdcb474b6d.jpg", + "table_caption": [], + "table_footnote": [ + "Table 4. Comparison of 3D Instance Segmentation performance with state-of-the-art approaches on the STPLS3D [2] test dataset." + ], + "table_body": "
MethodmAP\\( \\mathrm{mAP}_{50} \\)\\( \\mathrm{mAP}_{25} \\)
PointGroup [18]23.338.548.6
HAIS [3]35.146.752.8
SoftGroup [39]47.363.171.4
Mask3D [35]63.479.285.6
EASE (Ours)64.580.886.9
", + "bbox": [ + 500, + 554, + 903, + 661 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Ablation Studies", + "text_level": 1, + "bbox": [ + 500, + 705, + 663, + 720 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Effect of Semantic and Edge Modules. In Tab. 5, we evaluate the variants of our method w/ and w/o the Semantic Network (SN) and Edge Prediction Module (EPM). The addition of two modules improves 3DIS accuracy across all experiments. The SN especially underscores the significance of incorporating rich semantic knowledge for effective segmentation, achieving $+1.3 / 0.7$ and $+1.5 / 1.7$ improvements on $\\mathrm{mAP} / \\mathrm{mAP}_{50}$ for ScanNetV2 and S3DIS, respectively. EPM further enhances performance via the auxiliary edge information for edge-aware features, resulting in gains of $+0.9 / 0.8$ and $+1.9 / 0.9$ . Notably, integrating", + "bbox": [ + 496, + 734, + 893, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "20649", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/86575ad122e42ead3f8b3f697102e1b7cafc8ecf87130a556755a9b673a99fdc.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
BaselineMethodScanNet ValS3DIS Area 5
SemanticEdgemAP / mAP50mAP / mAP50
✓--58.4 / 75.956.6 / 68.4
✓✓-59.7 / 76.658.1 / 70.1
✓-✓59.3 / 76.758.5 / 69.3
✓✓✓60.2 / 77.259.0 / 71.6
", + "bbox": [ + 78, + 83, + 472, + 224 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/d7b10d50217883d356a7da33a646e119547440b7f34c4f9a23fc6b46c84354bd.jpg", + "table_caption": [ + "Table 5. Ablation study to see the effect of our proposed two main modules: Semantic Network and Edge Prediction Module." + ], + "table_footnote": [], + "table_body": "
UsageMethodScanNet ValS3DIS Area 5
E-wise SumConcatmAP / mAP50mAP / mAP50
---58.4 / 75.956.6 / 68.4
fcls✓-58.9 / 76.057.8 / 68.5
fcls-✓59.3 / 76.558.3 / 70.1
S·F✓-59.6 / 76.958.6 / 69.6
S·F-✓60.2 / 77.259.0 / 71.6
", + "bbox": [ + 78, + 262, + 472, + 416 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "SN and EPM leads to considerable advances $(+1.8 / 1.3$ and $+2.4 / 3.2)$ , confirming the effectiveness of each module.", + "bbox": [ + 75, + 472, + 468, + 502 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Applications of Text Embeddings. In Tab. 6, we analyze various applications of text embeddings to determine the most effective strategy for the functional understanding of complex 3D instances. We utilize text embeddings for category classification from $f_{cls}$ of the Mask Module to boost classification accuracy. Also, we strengthen basic queries using the Semantic Fusion ( $S \\cdot F$ ) network as described in Sec. 3.2. For experiments with the E-wise Sum, we map the text embeddings to the same feature space as the basic queries through MLPs and sum together. While incorporating semantic cues via element-wise sum enhances 3DIS accuracy, it exhibits limited improvements compared to concatenation. By comparison, applying the semantic knowledge across the entire transformer decoder through the $S \\cdot F$ outperforms using them only for the classification head.", + "bbox": [ + 73, + 505, + 470, + 733 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Weight Value of Edge Prediction Loss. To encourage the network to learn practical edge cues, we utilize an edge prediction network. In particular, we compute the weighted binary cross-entropy loss to effectively train the edge extraction head from sparse edge points. Here, we explore the impact of the coefficient value $w$ of $\\mathcal{L}_{edge}$ on 3DIS accuracy. First, as shown in Fig. 5, we empirically find that $w$ influences the thickness of edges in a 3D scene. In Tab. 7, for the ScanNet [4] consisting of relatively small-scale scenes, a smaller value of $w$ (e.g. 2), resulting in thin edges, leads to higher performance. However, for the vast-scale scenes of", + "bbox": [ + 75, + 734, + 470, + 901 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/67150f2bb3f008f5f4080bc18427ccd6e3e230d0fac2c9e55341405fba91698a.jpg", + "image_caption": [ + "Figure 5. Visualization of edge prediction results (red) for different weight values $w$ in the weighted binary cross-entropy loss $\\mathcal{L}_{edge}$ , along with corresponding pseudo-edge label (green)." + ], + "image_footnote": [], + "bbox": [ + 504, + 69, + 630, + 157 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/eadb8bdd0196298d2b073c8324cf301ce7f4c1d51468330c29f648cb5dd1fd1e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 630, + 69, + 759, + 156 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/afe05085399c88d9dc529810348cf903a93784fd1484e1bb2bbf4ddc574de952.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 761, + 70, + 888, + 156 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/74ddd89c04e9e46c1a051ee644a69742b027e1fd0d6f85c09a3f9685270472bb.jpg", + "table_caption": [ + "Table 6. Ablation study of leveraging text embeddings to provide semantic details. $f_{cls}$ represents the classification head in the Mask Module, and $S \\cdot F$ stands for the Semantic Fusion network." + ], + "table_footnote": [], + "table_body": "
Feature mapw in LedgeScanNet ValS3DIS Area 5
mAP / mAP50mAP / mAP50
F0259.3 / 76.756.9 / 68.6
F4259.0 / 76.156.7 / 68.5
F0658.9 / 76.358.4 / 68.9
F4658.5 / 76.058.2 / 68.6
F01058.4 / 75.858.5 / 69.3
F41058.5 / 75.758.2 / 69.0
", + "bbox": [ + 500, + 212, + 893, + 411 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 7. Ablation study to investigate the impact of varying weight values $w$ in the weighted binary cross-entropy ${\\mathcal{L}}_{\\text{edge }}$ ,using high $\\left( {F}_{0}\\right)$ and low $\\left( {F}_{4}\\right)$ resolution feature maps for edge prediction.", + "bbox": [ + 500, + 415, + 892, + 459 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "the S3DIS [1], a higher value of $w$ (e.g. 10) leads to better performance with thicker edges. We finally observe the correlation between the scene scale and the weight value $w$ ; the optimal value for edge learning is proportional to the scale. Besides, methods using high-resolution feature maps (e.g. $F_0$ ) for edge prediction generally perform better than those using low-resolution features maps (e.g. $F_4$ ).", + "bbox": [ + 498, + 468, + 890, + 575 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Qualitative Analyses", + "text_level": 1, + "bbox": [ + 500, + 585, + 694, + 602 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Visual Comparison. In this section, we qualitatively confirm the usefulness of our novel framework EASE. We visualize the predicted semantic and instance masks of the baseline model [35] and ours on the ScanNetV2 [4] validation set, in Fig. 6. We emphasize the key differences using green and yellow boxes. As shown in the semantic results (Sem.), ours accurately classifies instances with similar appearances (e.g., desk and bed, door and cabinet) instead of the baseline, which incorrectly recognizes them. Also, as illustrated in the instance mask results (Inst.) of Scene1 and Scene2, ours segments a single object as one, unlike the baseline, which splits it into multiple parts. Further, in Scene 3 (Inst.), ours correctly distinguishes multiple objects (door and cabinet) in contrast to the baseline model. These qualitative results demonstrate that our method provides more detailed and robust segmentation results than the baseline method.", + "bbox": [ + 496, + 609, + 890, + 852 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Impact of Utilizing Semantic Priors. To verify the effectiveness of the Semantic Network, we present t-SNE [38] visualizations of query features for instances with similar", + "bbox": [ + 498, + 854, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "20650", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/50b6fd6cbc393982acb1f116a4e3f476ecbae0c108ec67106571f8c33707713f.jpg", + "image_caption": [ + "Figure 6. Qualitative comparison of 3D Instance Segmentation performance on the ScanNetV2 [4] validation set. We visualize semantic (Sem.) and instance (Inst.) masks of the baseline model and ours with Ground Truth (GT) masks. The critical differences are emphasized using green and yellow-colored boxes. Note that the color map (top right) represents semantic labels. Best viewed in color." + ], + "image_footnote": [], + "bbox": [ + 80, + 66, + 890, + 309 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/27c0198d788fc79971e1ea3d9b0e6de9c8ae2fbd834829a37b9ab2a8933846b7.jpg", + "image_caption": [ + "Figure 7. Category probability map (spatial distribution) for queries with high confidence. The x-axis denotes queries, the y-axis represents categories, and the z-axis (right) is probabilities." + ], + "image_footnote": [], + "bbox": [ + 81, + 375, + 457, + 513 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "appearances, which pose challenges for 3DIS models. As shown in Fig. 8, the baseline (MAFT [22]) model exhibits disorganized clusters, whereas ours shows relatively distinct feature space for challenging instances. Further, in Fig. 7, we visualize the spatial distribution of category probability for instance queries. The baseline tends to confuse perquery categories and predicts multiple categories for each query with low probability, causing semantic misclassifications. However, our model EASE overcomes this problem using semantic priors, providing clearer insight. These qualitative findings confirm that our semantic network effectively prompts the network to learn instance-specific semantic knowledge from intelligent text embedding priors.", + "bbox": [ + 75, + 577, + 468, + 773 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 76, + 784, + 194, + 799 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Current 3D instance segmentation approaches often face challenges in understanding real-world 3D instances relying solely on geometrical information. Specifically, (i) these methods suffer from identifying instances with similar appearances, and (ii) they often misinterpret the spatial extent of instances, resulting in unclear edges. To tackle these", + "bbox": [ + 75, + 810, + 470, + 900 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/35c883f6ce8d111bed3542168efd7a733723329ce7901c8de1571996464441b6.jpg", + "image_caption": [ + "Figure 8. t-SNE [38] visualization of query features representing each instance. Compared to the baseline [22], which produces disorganized clusters, ours creates more distinct clusters." + ], + "image_footnote": [], + "bbox": [ + 504, + 376, + 694, + 481 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/91836c3bd17e74dc6b126d22e7f4d844031d3a862a617c6aa1c0847345d5a6d2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 694, + 376, + 887, + 481 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "challenges, EASE, our novel framework, focuses on context details of text embeddings from the language model as smart priors, enhancing practical semantic understanding. Also, we employ the edge prediction module, guiding the network to reduce misclassifications near edges with edge-aware features. Our extensive experiments verify the effectiveness of EASE, achieving new state-of-the-art scores.", + "bbox": [ + 496, + 542, + 890, + 647 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments This work was supported by Culture, Sports and Tourism R&D Program through the Korea Creative Content Agency grant funded by the Ministry of Culture, Sports and Tourism in 2023. (RS-2023-00227409, Development of sketch-based semantic 3D modeling technology for creating user-centric Metaverse content spaces for indoor spaces, $80\\%$ ), (R2022020068, 4D Content Generation and Copyright Protection with Artificial Intelligence, $10\\%$ ), Electronics and Telecommunications Research Institute (ETRI) grant funded by the Korean government. (24ZC1200, Research on hyper-realistic interaction technology for five senses and emotional experience, $9\\%$ ), and the Institute of Information & communications Technology Planning & Evaluation (IITP) grant funded by the Korea government (MSIT) (2019-0-00079, Artificial Intelligence Graduate School Program (Korea University), $1\\%$ ).", + "bbox": [ + 496, + 650, + 892, + 891 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "20651", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 174, + 106 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Iro Armeni, Ozan Sener, Amir R Zamir, Helen Jiang, Ioannis Brilakis, Martin Fischer, and Silvio Savarese. 3d semantic parsing of large-scale indoor spaces. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1534-1543, 2016. 2, 5, 6, 7", + "[2] Meida Chen, Qingyong Hu, Zifan Yu, Hugues Thomas, Andrew Feng, Yu Hou, Kyle McCullough, Fengbo Ren, and Lucio Soibelman. Stpls3d: A large-scale synthetic and real aerial photogrammetry 3d point cloud dataset. arXiv preprint arXiv:2203.09065, 2022. 2, 5, 6", + "[3] Shaoyu Chen, Jiemin Fang, Qian Zhang, Wenyu Liu, and Xinggang Wang. Hierarchical aggregation for 3d instance segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15467-15476, 2021. 1, 2, 6", + "[4] Angela Dai, Angel X Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5828-5839, 2017. 2, 5, 6, 7, 8", + "[5] Ruoxi Deng, Chunhua Shen, Shengjun Liu, Huibing Wang, and Xinru Liu. Learning to predict crisp boundaries. In Proceedings of the European conference on computer vision (ECCV), pages 562-578, 2018. 5", + "[6] Shenglan Du, Nail Ibrahimli, Jantien Stoter, Julian Kooij, and Liangliang Nan. Push-the-boundary: Boundary-aware feature propagation for semantic segmentation of 3d point clouds. In 2022 International Conference on 3D Vision (3DV), pages 1-10. IEEE, 2022. 5", + "[7] Cathrin Elich, Francis Engelmann, Theodora Kontogianni, and Bastian Leibe. 3d bird's-eye-view instance segmentation. In German Conference on Pattern Recognition, pages 48-61. Springer, 2019. 1, 2", + "[8] Francis Engelmann, Martin Bokeloh, Alireza Fathi, Bastian Leibe, and Matthias Nießner. 3d-mpa: Multi-proposal aggregation for 3d semantic instance segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9031-9040, 2020. 1, 2, 6", + "[9] Golnaz Ghiasi, Xiuye Gu, Yin Cui, and Tsung-Yi Lin. Scaling open-vocabulary image segmentation with image-level labels. In European Conference on Computer Vision, pages 540-557. Springer, 2022. 2", + "[10] Jingyu Gong, Jiachen Xu, Xin Tan, Jie Zhou, Yanyun Qu, Yuan Xie, and Lizhuang Ma. Boundary-aware geometric encoding for semantic segmentation of point clouds. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1424-1432, 2021. 5", + "[11] Xiuye Gu, Tsung-Yi Lin, Weicheng Kuo, and Yin Cui. Open-vocabulary object detection via vision and language knowledge distillation. arXiv preprint arXiv:2104.13921, 2021. 2", + "[12] Huy Ha and Shuran Song. Semantic abstraction: Openworld 3d scene understanding from 2d vision-language models. In 6th Annual Conference on Robot Learning, 2022. 2", + "[13] Lei Han, Tian Zheng, Lan Xu, and Lu Fang. Occuseg: Occupancy-aware 3d instance segmentation. In Proceedings" + ], + "bbox": [ + 78, + 114, + 472, + 902 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "of the IEEE/CVF conference on computer vision and pattern recognition, pages 2940-2949, 2020. 6", + "[14] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In Proceedings of the IEEE international conference on computer vision, pages 2961-2969, 2017. 1", + "[15] Tong He, Chunhua Shen, and Anton Van Den Hengel. Dyco3d: Robust instance segmentation of 3d point clouds through dynamic convolution. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 354-363, 2021. 6", + "[16] Ji Hou, Angela Dai, and Matthias Nießner. 3d-sis: 3d semantic instance segmentation of rgb-d scans. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4421-4430, 2019. 1, 2, 6", + "[17] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In International conference on machine learning, pages 4904-4916. PMLR, 2021. 2, 4", + "[18] Li Jiang, Hengshuang Zhao, Shaoshuai Shi, Shu Liu, Chi-Wing Fu, and Jiaya Jia. Pointgroup: Dual-set point grouping for 3d instance segmentation. In Proceedings of the IEEE/CVF conference on computer vision and Pattern recognition, pages 4867-4876, 2020. 1, 2, 6", + "[19] Abhijit Kundu, Xiaoqi Yin, Alireza Fathi, David Ross, Brian Brewington, Thomas Funkhouser, and Caroline Pantofaru. Virtual multi-view fusion for 3d semantic segmentation. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXIV 16, pages 518-535. Springer, 2020. 2", + "[20] Jean Lahoud, Bernard Ghanem, Marc Pollefeys, and Martin R Oswald. 3d instance segmentation via multi-task metric learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9256-9266, 2019. 6", + "[21] Xin Lai, Jianhui Liu, Li Jiang, Liwei Wang, Hengshuang Zhao, Shu Liu, Xiaojuan Qi, and Jiaya Jia. Stratified transformer for 3d point cloud segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8500-8509, 2022. 2", + "[22] Xin Lai, Yuhui Yuan, Ruihang Chu, Yukang Chen, Han Hu, and Jiaya Jia. Mask-attention-free transformer for 3d instance segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3693–3703, 2023. 1, 2, 3, 4, 5, 6, 8", + "[23] Feng Liang, Bichen Wu, Xiaoliang Dai, Kunpeng Li, Yinan Zhao, Hang Zhang, Peizhao Zhang, Peter Vajda, and Diana Marculescu. Open-vocabulary semantic segmentation with mask-adapted clip. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7061-7070, 2023. 2", + "[24] Zhihao Liang, Zhihao Li, Songcen Xu, Mingkui Tan, and Kui Jia. Instance segmentation in 3d scenes using semantic superpoint tree networks. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2783-2792, 2021. 6", + "[25] Chen Liu and Yasutaka Furukawa. Masc: Multi-scale affinity" + ], + "bbox": [ + 501, + 92, + 893, + 902 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "20652", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "with sparse convolution for 3d instance segmentation. arXiv preprint arXiv:1902.04478, 2019. 6", + "[26] Jiahao Lu, Jiacheng Deng, Chuxin Wang, Jianfeng He, and Tianzhu Zhang. Query refinement transformer for 3d instance segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 18516-18526, 2023. 1, 6", + "[27] John McCormac, Ankur Handa, Andrew Davison, and Stefan Leutenegger. Semantic fusion: Dense 3d semantic mapping with convolutional neural networks. In 2017 IEEE International Conference on Robotics and automation (ICRA), pages 4628-4635. IEEE, 2017. 2", + "[28] Ishan Misra, Rohit Girdhar, and Armand Joulin. An end-to-end transformer model for 3d object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2906-2917, 2021. 3", + "[29] Songyou Peng, Kyle Genova, Chiyu Jiang, Andrea Tagliasacchi, Marc Pollefeys, Thomas Funkhouser, et al. Openscene: 3d scene understanding with open vocabularies. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 815-824, 2023. 2", + "[30] Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J Guibas. Pointnet++: Deep hierarchical feature learning on point sets in a metric space. Advances in neural information processing systems, 30, 2017. 3", + "[31] Charles R Qi, Or Litany, Kaiming He, and Leonidas J Guibas. Deep hough voting for 3d object detection in point clouds. In proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9277-9286, 2019. 1", + "[32] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 2, 4", + "[33] David Rozenberszki, Or Litany, and Angela Dai. Language-grounded indoor 3d semantic segmentation in the wild. In European Conference on Computer Vision, pages 125-141. Springer, 2022. 2, 3, 5, 6", + "[34] Corentin Sautier, Gilles Puy, Spyros Gidaris, Alexandre Boulch, Andrei Bursuc, and Renaud Marlet. Image-to-lidar self-supervised distillation for autonomous driving data. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9891-9901, 2022. 2", + "[35] Jonas Schult, Francis Engelmann, Alexander Hermans, Or Litany, Siyu Tang, and Bastian Leibe. Mask3d for 3d semantic instance segmentation. arXiv preprint arXiv:2210.03105, 2022. 1, 2, 3, 4, 5, 6, 7", + "[36] Jiahao Sun, Chunmei Qing, Junpeng Tan, and Xiangmin Xu. Superpoint transformer for 3d scene instance segmentation. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2393-2401, 2023. 1, 2, 4, 5, 6", + "[37] Ayca Takmaz, Elisabetta Fedele, Robert W Sumner, Marc Pollefeys, Federico Tombari, and Francis Engelmann. Openmask3d: Open-vocabulary 3d instance segmentation. arXiv preprint arXiv:2306.13631, 2023. 2" + ], + "bbox": [ + 78, + 90, + 470, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[38] Laurens Van der Maaten and Geoffrey Hinton. Visualizing data using t-sne. Journal of machine learning research, 9 (11), 2008. 7, 8", + "[39] Thang Vu, Kookhoi Kim, Tung M Luu, Thanh Nguyen, and Chang D Yoo. Softgroup for 3d instance segmentation on point clouds. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2708-2717, 2022. 6", + "[40] Weiyue Wang, Ronald Yu, Qiangui Huang, and Ulrich Neumann. Sgpn: Similarity group proposal network for 3d point cloud instance segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2569-2578, 2018. 1, 2", + "[41] Xinlong Wang, Shu Liu, Xiaoyong Shen, Chunhua Shen, and Jiaya Jia. Associatively segmenting instances and semantics in point clouds. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4096-4105, 2019. 6", + "[42] Yizheng Wu, Min Shi, Shuaiyuan Du, Hao Lu, Zhiguo Cao, and Weicai Zhong. 3d instances as 1d kernels. In European Conference on Computer Vision, pages 235-252. Springer, 2022. 2", + "[43] Bo Yang, Jianan Wang, Ronald Clark, Qingyong Hu, Sen Wang, Andrew Markham, and Niki Trigoni. Learning object bounding boxes for 3d instance segmentation on point clouds. Advances in neural information processing systems, 32, 2019. 1, 2, 6", + "[44] Li Yi, Wang Zhao, He Wang, Minhyuk Sung, and Leonidas J Guibas. Gspn: Generative shape proposal network for 3d instance segmentation in point cloud. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3947-3956, 2019. 1, 2, 6", + "[45] Min Zhong, Xinghao Chen, Xiaokang Chen, Gang Zeng, and Yunhe Wang. Maskgroup: Hierarchical point grouping and masking for 3d instance segmentation. In 2022 IEEE International Conference on Multimedia and Expo (ICME), pages 1-6. IEEE, 2022. 6" + ], + "bbox": [ + 501, + 92, + 892, + 612 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "20653", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2024/Edge-Aware 3D Instance Segmentation Network with Intelligent Semantic Prior/ef85b90b-3f1b-4322-ac5a-c5d6ae60802c_model.json b/2024/Edge-Aware 3D Instance Segmentation Network with Intelligent Semantic Prior/ef85b90b-3f1b-4322-ac5a-c5d6ae60802c_model.json new file mode 100644 index 0000000000000000000000000000000000000000..30fbf5f102c87e483c68698d63e218a837730c53 --- /dev/null +++ b/2024/Edge-Aware 3D Instance Segmentation Network with Intelligent Semantic Prior/ef85b90b-3f1b-4322-ac5a-c5d6ae60802c_model.json @@ -0,0 +1,2035 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.808, + 0.045 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.084, + 0.131, + 0.888, + 0.152 + ], + "angle": 0, + "content": "Edge-Aware 3D Instance Segmentation Network with Intelligent Semantic Prior" + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.18, + 0.751, + 0.245 + ], + "angle": 0, + "content": "Wonseok Roh1 Hwanhee Jung1 Giljoo Nam2 Jinseop Yeom1 Hyunjep Park1 Sang Ho Yoon3 Sanggil Kim1* \n1Korea University 2Meta Reality Labs 3KAIST" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.274, + 0.313, + 0.289 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.293, + 0.474, + 0.611 + ], + "angle": 0, + "content": "While recent 3D instance segmentation approaches show promising results based on transformer architectures, they often fail to correctly identify instances with similar appearances. They also ambiguously determine edges, leading to multiple misclassifications of adjacent edge points. In this work, we introduce a novel framework, called EASE, to overcome these challenges and improve the perception of complex 3D instances. We first propose a semantic guidance network to leverage rich semantic knowledge from a language model as intelligent priors, enhancing the functional understanding of real-world instances beyond relying solely on geometrical information. We explicitly instruct the basic instance queries using text embeddings of each instance to learn deep semantic details. Further, we utilize the edge prediction module, encouraging the segmentation network to be edge-aware. We extract voxel-wise edge maps from point features and use them as auxiliary information for learning edge cues. In our extensive experiments on large-scale benchmarks, ScanNetV2, ScanNet200, S3DIS, and STPLS3D, our EASE outperforms existing state-of-the-art models, demonstrating its superior performance." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.629, + 0.208, + 0.644 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.654, + 0.47, + 0.73 + ], + "angle": 0, + "content": "Understanding 3D scenes is a fundamental task within 3D computer vision. Given 3D point cloud scenes, identifying and perceiving instances on sparse points, along with assigning semantic class labels, play a crucial role in a comprehensive understanding of the entire spatial environment." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.73, + 0.47, + 0.866 + ], + "angle": 0, + "content": "In real-world 3D scenarios, significant occlusion and truncation often occur, especially when objects are overlapped or hidden by others. To address these issues, conventional works [3, 7, 8, 16, 18, 40, 43, 44] in 3D instance segmentation (3DIS) primarily focus on accurately generating region proposals (top-down) [16, 43, 44] or effectively grouping points with clustering algorithms (bottom-up) [3, 7, 8, 18, 40]. Inspired by the remarkable sensation of Mask-RCNN [14], the former (proposal-based) meth" + }, + { + "type": "image", + "bbox": [ + 0.507, + 0.275, + 0.891, + 0.581 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.622, + 0.582, + 0.783, + 0.593 + ], + "angle": 0, + "content": "(b) Instance misinterpretation cases" + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.599, + 0.892, + 0.682 + ], + "angle": 0, + "content": "Figure 1. Examples of 3DIS for two challenging cases. (a) Semantic failure cases: the baseline [35] confuses seat-shaped objects as chair or toilet (Scene1) and also misclassifies large cuboid blocks as desk (Scene2). And (b) Instance misinterpretation cases: baseline erroneously segments a single object into multiple parts (Scene3) or merges multiple objects into one instance (Scene4)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.689, + 0.893, + 0.84 + ], + "angle": 0, + "content": "ods initially detect instances as bounding boxes and predict masks within each proposed region. However, these strategies are susceptible to the quality of the detection results. On the other hand, the latter (grouping-based) methods use clustering algorithms to group closely related points and aggregate point-wise class labels and instance features. While they show great advancements in 3DIS, these methods require additional manually tuned processing, such as point grouping [18] or voting mechanisms [31], to determine specific geometric properties (e.g., centers, occupancy)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Recently, transformer-based 3DIS frameworks [22, 26, 35, 36] have tackled several limitations of traditional approaches by introducing a fully end-to-end pipeline. They directly predict masks from 3D points without resource-" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.876, + 0.456, + 0.901 + ], + "angle": 0, + "content": "*Corresponding author. \nProject Page: https://kuai-lab.github.io/ease2024." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20644" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.273 + ], + "angle": 0, + "content": "exhaustive processing. These methods have achieved high performance and quickly become predominant in the 3DIS domain. However, despite these breakthroughs, they still face challenges in accurately masking everyday 3D instances relying solely on geometrical information. Specifically, they frequently suffer from cases where each instance appears similarly, but their semantic roles differ, as illustrated in Fig. 1 (a). Thus, gaining insight into the semantic properties of instances beyond mere appearance or position is crucial for effectively addressing complex real-world 3D scenarios. From this observation, we consider leveraging semantic knowledge along with the geometry information." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.276, + 0.473, + 0.563 + ], + "angle": 0, + "content": "Built upon the classic structure of the transformer-based framework [22, 35, 36], which trains queries that encode instance-specific knowledge for direct mask prediction, we introduce a carefully designed semantic guidance network. One of the most promising ways to utilize semantic information is through visual language models [17, 32]. They learn generous visual-linguistic knowledge from large-scale image-text pairs, which empowers them to achieve outstanding progress in various 2D or 3D vision tasks [9, 11, 23, 29]. In this work, motivated by the success of these works, we explicitly instruct the transformer-based network to learn contextual variations among instances using text embeddings. Specifically, we strengthen the basic instance queries with profound semantic clues of text embeddings representing each instance. Here, the context details can serve as discriminative semantic priors for perceiving complex 3D instances adequately. Through our extensive experiments, we found that our semantic-guided approach effectively overcomes the challenges of existing methods." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.566, + 0.473, + 0.822 + ], + "angle": 0, + "content": "Different from the human vision system, which easily distinguishes boundaries between instances, we empirically observe that most 3DIS models [21, 22, 35, 36, 42] struggle with precisely capturing the boundaries in the 3D space. These methods often misinterpret the spatial range of instances and ambiguously determine their edges, leading to numerous misclassifications of adjacent edge points, as shown in Fig. 1 (b). We consider that these issues arise due to the absence of detailed guidance on inter-object boundaries, resulting in fuzzy edges. To tackle these challenges, we advocate for leveraging the edge prediction module to pilot the whole network to exploit edge-advanced features. We predict edge points across point cloud scenes and operate them as auxiliary information for learning practical edge cues, which encourages accurate recognition of the 3D spatial scope. Note that we supervise this module using dynamically generated point-wise pseudo edge labels." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.473, + 0.903 + ], + "angle": 0, + "content": "Given landmark datasets for 3DIS, ScanNetV2 [4], ScanNet200 [33], S3DIS [1], and STPLS3D [2], we validate the effectiveness and robustness of our framework. Our method outperforms the existing state-of-the-art methods. To summarize, our main contributions are listed as follows:" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.092, + 0.894, + 0.167 + ], + "angle": 0, + "content": "- We propose EASE, a novel 3D instance segmentation framework that utilizes the rich semantic clues from the language model. Specifically, context details of text embeddings serve as smart semantic priors, effectively enhancing the functional comprehension of 3D instances." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.168, + 0.894, + 0.228 + ], + "angle": 0, + "content": "- We introduce the edge prediction module guiding the entire network to take advantage of edge-aware features. Also, we utilize the module's output as auxiliary knowledge for learning edge cues, boosting 3DIS performance." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.228, + 0.895, + 0.304 + ], + "angle": 0, + "content": "- We analyze the effectiveness of our proposed method on multiple challenging benchmarks, including ScanNetV2, ScanNet200, S3DIS, and STPLS3D. Extensive experiments on various 3D scenarios validate that our method achieves new State-of-the-Art performance on 3DIS." + }, + { + "type": "list", + "bbox": [ + 0.5, + 0.092, + 0.895, + 0.304 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.318, + 0.642, + 0.334 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.339, + 0.895, + 0.747 + ], + "angle": 0, + "content": "3D Instance Segmentation. The 3D Instance Segmentation (3DIS) task aims to identify individual instances and assign semantic classes to each point within a 3D point scene. Traditional approaches in 3DIS are primarily categorized into three groups: proposal-based [16, 43, 44], grouping-based [3, 7, 8, 18, 40], and transformer-based methods [22, 35, 36]. The proposal-based approaches initially detect object proposals, such as 3D bounding boxes, and then estimate corresponding per-point semantic classes and instance masks. One of the grouping-based approaches, PointGroup [18], conducts point-wise clustering to group points into instances using dual point sets, which comprise original and shifted coordinates. Recently, transformer-based approaches predict instance queries that represent individual objects and decode them into per-point categories and instance labels directly. Mask3D [35] refines queries using masked cross-attention with hierarchical point features, leading the queries to focus on particular instances. SPFormer [36] updates instance queries via transformer decoder utilizing pre-computed superpoints, which can reduce the whole computational cost of the network. MAFT [22], on the other hand, recognizes representational limitations in the initial mask and replaces the masking process with object-localizing position queries. Based on the core idea of transformer-based methods, our novel approach incorporates a semantic guidance network and an auxiliary edge network to improve instance mask segmentation." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.897, + 0.903 + ], + "angle": 0, + "content": "Semantic Guided 3D Scene Understanding. Recent research in 3D scene understanding focuses on comprehending large-scale, real-world 3D environments. Conventional studies in this field are broadly categorized into two strategies: closed-set 3D scene understanding, a prevalent approach in diverse tasks [19, 27, 33, 34], and open-vocabulary methods [12, 29, 37], which focus on recognizing unseen categories within datasets. More recently, there has been significant progress in integrating pre-trained visual-language models like CLIP [32] or ALIGN [17] into" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "20645" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.086, + 0.095, + 0.891, + 0.393 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.406, + 0.893, + 0.477 + ], + "angle": 0, + "content": "Figure 2. An overview of our framework EASE. Built upon the classic architecture of the transformer-based 3DIS model, our model takes 3D scenes and directly infers instance masks. Our model consists of four main modules: (1) Sparse Convolutional Backbone, (2) Semantic Network, which strengthen the queries to learn contextual variations of instances with rich semantic embeddings from the pre-trained language model CLIP, (3) Mask Transformer Decoder with Mask Module and Query Refinement blocks, and (4) Edge Prediction Module, where our edge extraction head is trained to predict a voxel-wise edge map, encouraging the network to utilize edge-advanced features." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.488, + 0.47, + 0.593 + ], + "angle": 0, + "content": "these methodologies to enhance semantic understanding. Building on this idea, for instance, [33] utilizes a closed-set framework and enhances 3D feature learning by aligning it with CLIP text embedding space for semantic classification. In this work, we employ contextual semantic priors to improve the comprehension of complex 3D instances, enhancing the overall 3D instance segmentation performance." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.61, + 0.169, + 0.626 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.637, + 0.47, + 0.773 + ], + "angle": 0, + "content": "In this section, we introduce a novel 3D Instance Segmentation (3DIS) framework EASE leveraging the rich semantic guidance from the language model. Furthermore, we propose the edge prediction module, which encourages the pipeline to use edge-aware features for effectively perceiving instance boundaries. We provide an overview of the entire pipeline (Sec. 3.1, Fig. 2, and Alg. 1) and then elaborate on the details of our method: (i) Intelligent Semantic Prior (Sec. 3.2) and (ii) Edge-Aware 3DIS Framework (Sec. 3.3)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.786, + 0.188, + 0.801 + ], + "angle": 0, + "content": "3.1. Overview" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.81, + 0.471, + 0.901 + ], + "angle": 0, + "content": "Our end-to-end 3DIS framework, as illustrated in Fig. 2 and Alg. 1, aims to enhance the comprehension of various instances across 3D point cloud scenes. Based on the classic transformer-based architecture [22, 35], which operates standard transformer building blocks, we directly infer instance masks from 3D point clouds. Our model consists" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.488, + 0.892, + 0.578 + ], + "angle": 0, + "content": "of four main modules: (1) Sparse Convolutional Backbone, (2) Semantic Network, which provide the deep semantic cues of diverse real-world instances, (3) Mask Transformer Decoder with Mask Module and Query Refinement blocks, and (4) Edge Prediction Module, which benefits the whole network to take advantage of edge-aware features." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.581, + 0.893, + 0.779 + ], + "angle": 0, + "content": "Mask Transformer. First, the Sparse Convolutional U-Net Backbone takes a colored point cloud \\( P \\in \\mathbb{R}^{N_p \\times 6} \\) as input and voxelizes \\( P \\) into \\( X_0 \\) voxels \\( V \\in \\mathbb{R}^{X_0 \\times 3} \\) to extract a multi-resolution hierarchical feature maps \\( F_i \\in \\mathbb{R}^{X_i \\times D} \\), where \\( i \\in \\{0,1,2,3,4\\} \\). Following [28, 35], we set zero-initialized non-parametric instance queries \\( Q \\in \\mathbb{R}^{N_q \\times D} \\), referring to point positions sampled with furthest point sampling (FPS) [30]. Given the \\( F_i \\) and \\( Q \\), the mask transformer decoder layer iteratively enhances the queries using the Mask Module (MM) and Query Refinement (QR) blocks. In the MM, as shown in Fig. 3 (a), we classify the category \\( c_i \\) for \\( i = \\{1,2,\\dots,N_c\\} \\) of each query via linear classification head \\( f_{\\mathrm{class}} \\) using following cross-entropy loss:" + }, + { + "type": "equation", + "bbox": [ + 0.527, + 0.788, + 0.892, + 0.83 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {c l s} = - \\mathbb {E} _ {c, w _ {c} \\sim \\mathbb {D}} \\left[ \\sum_ {r \\in N _ {c}} w _ {c} [ r ] \\log f _ {\\text {c l a s s}} (Q) [ r ] \\right] \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.893, + 0.901 + ], + "angle": 0, + "content": "where \\(w_{c}\\) denotes one-hot encoded category labels and \\(\\mathbb{D}\\) represents (input) data distribution. Also, the \\(Q\\) are fed through the query head \\(f_{\\mathrm{query}}\\) to project them to the same feature space as \\(F_{0}\\) for mask prediction. Afterwards, we" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "20646" + } + ], + [ + { + "type": "code_caption", + "bbox": [ + 0.089, + 0.095, + 0.417, + 0.11 + ], + "angle": 0, + "content": "Algorithm 1: Overview of our framework EASE." + }, + { + "type": "algorithm", + "bbox": [ + 0.08, + 0.112, + 0.47, + 0.412 + ], + "angle": 0, + "content": "Input: Colored point cloud \\(P\\in \\mathbb{R}^{N_p\\times 6}\\) Result: Binary 3DIS mask \\(\\hat{M}_f\\in \\{0,1\\}^{X\\times N_q}\\) Procedure: # Sparse Convolutional Backbone. 1 Extract multi-resolution feature maps \\(F_{i}\\) from \\(P\\) # Edge Prediction Module (EPM). 2 Predict edge map \\(\\hat{E}\\) using full-resolution feature \\(F_{0}\\) 3 Supervise \\(\\hat{E}\\) with pseudo edge label \\(E\\) 4 Encourage the network to learn edge-aware features. #Mask Transformer Decoder. 5 for \\(l\\in N_l\\) do 6 for \\(i\\in \\{1,2,3,4\\}\\) do 7 Update queries \\(Q\\) using \\(M_{attn}\\) and \\(F_{i}\\) # Semantic Network. 8 Match the set of classes with text labels. 9 Encode text into text embeds \\(T\\) using CLIP. 10 Reinforce \\(Q\\) to contain semantic clues of \\(T\\) 11 end 12 end" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.428, + 0.47, + 0.487 + ], + "angle": 0, + "content": "compute the similarity between the projected query features \\( Q' \\) and \\( F_0 \\) using the dot product operation, then calculate the probability of the instance mask employing the sigmoid function as follows." + }, + { + "type": "equation", + "bbox": [ + 0.101, + 0.498, + 0.469, + 0.517 + ], + "angle": 0, + "content": "\\[\nM _ {a t t n} = \\left\\{m _ {i, j} = \\left[ \\sigma \\left(F _ {0} \\cdot f _ {\\text {q u e r y}} (Q) ^ {T}\\right) _ {i, j} > 0. 5 \\right] \\right\\} \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.527, + 0.469, + 0.618 + ], + "angle": 0, + "content": "where the threshold value is 0.5 for binary mask. Further, to refine the query representation, we leverage the \\(QR\\) block, including masked cross-attention and self-attention mechanisms. Here, we utilize \\(M_{att}\\) as the foreground mask for the masked cross-attention layer, where \\(Q\\) iteratively attend to each multi-scale feature \\((F_1 - F_4, i \\geqslant 1)\\) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.145, + 0.628, + 0.469, + 0.647 + ], + "angle": 0, + "content": "\\[\nQ = \\operatorname {s o f t m a x} \\left(\\mathbf {Q K} ^ {T} / \\sqrt {D} + M _ {\\text {a t t n}}\\right) \\mathbf {V} \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.657, + 0.469, + 0.779 + ], + "angle": 0, + "content": "where \\(\\mathbf{K}\\) and \\(\\mathbf{V}\\) are linearly projected from each point feature, and \\(\\mathbf{Q}\\) are from \\(Q\\). Hence, the transformer focuses more on valuable instances instead of the unessential background. Subsequently, in the self-attention layer, the keys, values, and queries are all linear projections of \\(Q\\) from the cross-attention layer. Finally, the transformer decoder layer is recurrently applied for multiple iterations \\(N_{l}\\) and feature scales \\(i\\), ultimately producing the final set of refined queries." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.787, + 0.314, + 0.803 + ], + "angle": 0, + "content": "3.2. Intelligent Semantic Prior" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.811, + 0.469, + 0.901 + ], + "angle": 0, + "content": "It is important to highlight that precise geometric properties (e.g., locations, distances and orientations) from 3D points significantly enhance 3DIS accuracy. While previous studies [22, 35, 36] benefit from the generous geometric potential of physical information, they are still limited in perceiving intricate real-world 3D instances. Hence, we pro" + }, + { + "type": "image", + "bbox": [ + 0.507, + 0.09, + 0.888, + 0.173 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.617, + 0.175, + 0.776, + 0.186 + ], + "angle": 0, + "content": "(a) Transformer Decoder Layer" + }, + { + "type": "image", + "bbox": [ + 0.507, + 0.19, + 0.888, + 0.257 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.64, + 0.259, + 0.753, + 0.27 + ], + "angle": 0, + "content": "(b) Semantic Network" + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.274, + 0.892, + 0.357 + ], + "angle": 0, + "content": "Figure 3. Detailed architecture of (a) Transformer Decoder Layer: consists of Mask Module and Query Refinement blocks. And (b) Semantic Network: we first map the categories of queries \\( Q \\) with the corresponding text description for text embeddings \\( T \\) from CLIP [32]. Then, we explicitly strengthen the \\( Q \\) with \\( T \\) using the Semantic Fusion network to learn semantic details." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.369, + 0.891, + 0.399 + ], + "angle": 0, + "content": "pose to utilize semantic knowledge of individual instances as smart priors, effectively addressing the challenges." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.402, + 0.892, + 0.688 + ], + "angle": 0, + "content": "Deep Semantic Knowledge for 3D Instances. Recently, visual language models [17, 32] have highlighted the benefits of their capability to provide general embedded knowledge from large-scale image-text multimodal data. In this work, we explicitly instruct the transformer-based 3DIS network to learn contextual variations among instances by utilizing the power of the prevalent language model CLIP [32]. To achieve this goal, we utilize a carefully designed Semantic Network, as shown in Fig. 3 (b). Specifically, with the refined query \\( Q \\) from each \\( QR \\) block, we first categorize each query following the linear classification head \\( f_{\\mathrm{class}} \\) in the MM. Given the set of category candidates, we map the category number \\( c_{i} \\) for \\( i = \\{1,2,\\dots,N_c\\} \\) with the corresponding text descriptions (e.g., \"Cabinet\", \"Bookshelf\"). Next, these text labels go through the large-scale pre-trained language model, which outputs text embedding vectors \\( T \\). We then concatenate \\( Q \\) with \\( T \\) and strengthen the basic instance queries with deep semantic clues from text embeddings using the Semantic Fusion network as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.558, + 0.695, + 0.891, + 0.713 + ], + "angle": 0, + "content": "\\[\nQ = \\operatorname {S i g m o i d} \\left(\\mathbf {W} _ {2} \\cdot \\phi \\left(\\mathbf {W} _ {1} \\cdot (Q \\oplus T)\\right) ^ {T}\\right) \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.718, + 0.892, + 0.824 + ], + "angle": 0, + "content": "where \\(\\oplus\\) indicates channel-wise concatenation, \\(\\mathbf{W}_i\\) denotes the learnable parameters of the \\(i\\)-th linear layer, and \\(\\phi\\) is the non-linear activation function. We apply this process for each iteration of the \\(QR\\). Ultimately, we consider these explicitly guided queries with contextual variations as discriminative semantic priors for the practical perception of complex 3D instances, enhancing 3DIS accuracy." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.833, + 0.769, + 0.849 + ], + "angle": 0, + "content": "3.3. Edge-Aware 3DIS Framework" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.856, + 0.892, + 0.901 + ], + "angle": 0, + "content": "In real-world 3D scenarios, instances are commonly positioned in diverse arrangements without following standard rules or patterns. Especially when instances are closely ad" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "20647" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.084, + 0.094, + 0.464, + 0.196 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.21, + 0.47, + 0.281 + ], + "angle": 0, + "content": "Figure 4. Pseudo edge label calculation using KNN algorithm. We collect \\( k \\) neighboring points of each point (red border) and compare their instance labels. If the count of different label points (blue and purple) among its surroundings surpasses a predefined threshold \\( \\tau \\), we annotate the central point as an edge point (green)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.292, + 0.469, + 0.413 + ], + "angle": 0, + "content": "jacent or overlapping, it becomes increasingly challenging for the model to define the edges precisely. Therefore, even with significant progress in technical strategies, most conventional methods [22, 35, 36] often misinterpret the spatial extent of each instance and ambiguously estimate their edges as illustrated in Fig. 1. To this end, we introduce an edge-aware 3DIS framework with an edge prediction module, enabling the network to utilize edge-advanced features." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.415, + 0.47, + 0.687 + ], + "angle": 0, + "content": "Pseudo Edge Label Calculation. To optimize the edge prediction module, we first compute point-wise pseudo edge labels for all 3D point scenes. We operate a traditional \\(k\\)-Nearest Neighbors (KNN) classification algorithm based on the KD-tree data structure to explore the \\(k\\) nearest points. Here, we utilize point-wise instance labels from datasets. For all points in a 3D scene, we set each point as a central point and compare their instance labels with those of \\(k\\) neighbor points. As shown in Fig. 4, if the count of distinct label points among its surroundings exceeds a predefined threshold \\(\\tau\\), we identify the central point as a boundary point. We finally generate pseudo-binary edge maps \\(E' \\in \\{0,1\\}^{N_p \\times 1}\\), where points on the edges are 1 and internal points are 0 for the entire set of points \\(P\\). Then, we voxelize the point-wise edge labels into the voxel-wise edge labels \\(E \\in \\{0,1\\}^{X_0 \\times 1}\\) to supervise the edge extraction head. Note that we precompute the pseudo edge labels for every 3D scene in the datasets before training." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.689, + 0.47, + 0.841 + ], + "angle": 0, + "content": "Learning Edge Cues. With dynamically generated pseudo edge labels \\( E \\), we aim to enrich regular features to recognize the accurate spatial scope of 3D instances with precise boundaries. Specifically, we introduce an Edge Prediction Module, which predicts a voxel-wise dense edge map \\( \\hat{E} \\in (0,1)^{X_0 \\times 1} \\) to interactively guide the network in capturing edges of various instances. Given a full-resolution feature map \\( F_0 \\) from the backbone network, our edge extraction head, which consists of multiple shared MLPs, estimates the probabilities of edges for each voxel as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.09, + 0.847, + 0.469, + 0.866 + ], + "angle": 0, + "content": "\\[\n\\hat {E} = \\operatorname {M L P} _ {\\text {e d g e}} \\left(F _ {0}\\right) = \\operatorname {S i g m o i d} \\left(\\mathbf {W} _ {2} \\cdot \\phi \\left(\\mathbf {W} _ {1} \\cdot F _ {0}\\right) ^ {T}\\right) \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.469, + 0.901 + ], + "angle": 0, + "content": "where \\(\\mathbf{W}_i\\) denotes the learnable parameters of the \\(i\\)-th linear layer, and \\(\\phi\\) indicates the non-linear activation function." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.892, + 0.137 + ], + "angle": 0, + "content": "Following [6, 10], as edge voxels constitute a small portion of the entire voxel set, we formulate weighted binary cross entropy loss to train the edge extraction head as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.508, + 0.144, + 0.892, + 0.185 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {e d g e} = - \\sum_ {i = 1} ^ {X _ {0}} \\left(w \\cdot E _ {i} \\log \\hat {E} _ {i} + \\left(1 - E _ {i}\\right) \\log \\left(1 - \\hat {E} _ {i}\\right)\\right) (6)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.189, + 0.892, + 0.295 + ], + "angle": 0, + "content": "where \\( w \\) represent the weight value used to balance the substantial difference between the numbers of edge points and others. Then, we propagate the edge cues of prediction as auxiliary information for the backbone network, improving 3D instance understanding with edge-aware features. Furthermore, we train our edge network jointly in an end-to-end manner with low computational costs." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.298, + 0.892, + 0.343 + ], + "angle": 0, + "content": "Loss Function We finally predict instance mask \\(\\hat{M}_f\\) using the final set of queries in the \\(MM\\). For training, we compute the following loss \\(\\mathcal{L}_{mask}\\) as a sum of the two losses:" + }, + { + "type": "equation", + "bbox": [ + 0.509, + 0.353, + 0.892, + 0.372 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {m a s k}} = \\lambda_ {\\mathrm {B C E}} \\mathcal {L} _ {\\mathrm {B C E}} \\left(M _ {f}, \\hat {M} _ {f}\\right) + \\lambda_ {\\mathrm {d i c e}} \\mathcal {L} _ {\\mathrm {d i c e}} \\left(M _ {f}, \\hat {M} _ {f}\\right) \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.381, + 0.892, + 0.428 + ], + "angle": 0, + "content": "where \\(M_{f}\\) denotes the ground-truth instance mask, and \\(\\mathcal{L}_{\\text{dice}}\\) represents the Dice loss [5]. Ultimately, our model is trained end-to-end by minimizing the following loss \\(\\mathcal{L}_{\\text{total}}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.525, + 0.441, + 0.892, + 0.459 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {t o t a l}} = \\lambda_ {\\text {m a s k}} \\mathcal {L} _ {\\text {m a s k}} + \\lambda_ {\\text {c l s}} \\mathcal {L} _ {\\text {c l s}} + \\lambda_ {\\text {e d g e}} \\mathcal {L} _ {\\text {e d g e}} \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.468, + 0.892, + 0.498 + ], + "angle": 0, + "content": "where each \\(\\lambda\\) is a hyperparameter derived from grid searches to handle the strength of respective loss term." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.511, + 0.633, + 0.529 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.536, + 0.691, + 0.553 + ], + "angle": 0, + "content": "4.1. Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.562, + 0.892, + 0.668 + ], + "angle": 0, + "content": "Datasets. In this study, we train and evaluate the overall performance using four landmark datasets for 3D instance segmentation: ScanNetV2 [4], ScanNet200 [33], S3DIS [1], and STPLS3D [2]. These four datasets provide 3D point cloud scan data collected in various scenarios. Note that detailed descriptions of each dataset and all implementation details are provided in the supplemental material." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.67, + 0.892, + 0.806 + ], + "angle": 0, + "content": "Evaluation Metrics. To evaluate the 3D instance segmentation accuracy, we operate Average Precision (AP), a typical metric for various computer vision tasks. We report mean average precision (mAP), which represents an average score with IoU thresholds ranging from \\(50\\%\\) to \\(95\\%\\) with a step size of \\(5\\%\\). Also, we provide \\(\\mathrm{mAP}_{50}\\) and \\(\\mathrm{mAP}_{25}\\), indicating the scores with IoU thresholds of \\(50\\%\\) and \\(25\\%\\), respectively. Further, we estimate mean precision (mPrec) and mean recall (mRec) on the S3DIS [1] dataset." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.815, + 0.891, + 0.831 + ], + "angle": 0, + "content": "4.2. Performance Comparison with SOTA Methods." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.892, + 0.901 + ], + "angle": 0, + "content": "ScanNetV2. We quantitatively compare our proposed network EASE with existing state-of-the-art methods. These methods demonstrate significant capabilities, but they still rely on geometric information from 3D points and often" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20648" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.082, + 0.089, + 0.473, + 0.39 + ], + "angle": 0, + "content": "
MethodScanNet ValScanNet Test
mAP\\( mAP_{50} \\)mAP\\( mAP_{50} \\)
GSPN [44]19.337.8-30.6
3D-SIS [16]-18.716.138.2
MASC [25]--25.444.7
3D-Bonet [43]--25.348.8
MTML [20]20.340.228.254.9
3D-MPA [8]35.559.135.561.1
DyCo3D [15]35.457.639.564.1
PointGroup [18]34.856.740.763.6
MaskGroup [45]42.063.343.466.4
OccuSeg [13]44.260.748.667.2
SSTNet [24]49.464.350.669.8
HAIS [3]43.564.145.769.9
SoftGroup [39]46.067.650.476.1
Mask3D [35]55.273.756.678.0
QueryFormer [26]56.574.258.378.7
MAFT [22]59.976.559.678.6
EASE (Ours)60.277.259.878.9
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.076, + 0.395, + 0.471, + 0.452 + ], + "angle": 0, + "content": "Table 1. Comparison of 3D Instance Segmentation performance with state-of-the-art approaches on the ScanNetV2 [4] dataset. We evaluate mean average precision (mAP) with different IoU thresholds over 18 classes on the ScanNetV2 validation / hidden test set." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.463, + 0.47, + 0.6 + ], + "angle": 0, + "content": "face challenges in effectively comprehending 3D instances. To address these limitations, we present a novel edge-aware framework featuring a semantic guidance network. As reported in Tab. 1, EASE generally outperforms other methods, achieving new state-of-the-art accuracy in terms of \\(\\mathrm{mAP}\\) and \\(\\mathrm{mAP}_{50}\\) on both validation (60.2 / 77.2) and hidden test (59.8 / 78.9) sets. These results confirm that our edge cues and smart semantic prior benefits the entire network, leading to high precision in 3D instance segmentation." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.603, + 0.47, + 0.727 + ], + "angle": 0, + "content": "S3DIS. In Tab. 2, we evaluate 3DIS performance on Area 5 and 6-fold cross-validation of the S3DIS [1] dataset. For Area 5 evaluation, we employ data from Area 5 for validation and the other areas for training. Also, for 6-fold cross-validation, we assess validation scores across six other areas and compute the average. In both evaluations, EASE demonstrates significant performance improvements up to \\(+2.4 / 2.5\\) (Area 5) and \\(+0.8 / 0.9\\) (6-fold) for \\(\\mathrm{mAP} / \\mathrm{mAP}_{50}\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.73, + 0.47, + 0.793 + ], + "angle": 0, + "content": "ScanNet200. We also demonstrate considerable performance on the ScanNet200 validation set (see Tab. 3). Our method EASE precisely segments 3D instances for various categories compared to current state-of-the-art approaches." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.795, + 0.471, + 0.903 + ], + "angle": 0, + "content": "STPLS3D. Remarkably, our EASE network also achieves higher scores on the STPLS3D [2] dataset, which consists of outdoor 3D point cloud scenes. EASE exceeds the second-best results with \\(+1.1 / 1.6 / 1.3\\) margins in \\(\\mathrm{mAP} / \\mathrm{mAP}_{50} / \\mathrm{mAP}_{25}\\). These results highlight the effectiveness of our method in understanding diverse real-world 3D instances, both from indoor and outdoor environments." + }, + { + "type": "table", + "bbox": [ + 0.499, + 0.089, + 0.908, + 0.312 + ], + "angle": 0, + "content": "
MethodS3DIS Area 5S3DIS 6-fold CV
AP\\( AP_{50} \\)\\( Prec_{50} \\)\\( Rec_{50} \\)AP\\( AP_{50} \\)\\( Prec_{50} \\)\\( Rec_{50} \\)
ASIS [41]--55.342.4--63.647.5
3D-Bonet [43]--57.540.2--65.647.6
3D-MPA [8]--63.158.0--66.764.1
PointGroup [18]-57.861.962.1-64.069.669.2
DyCo3D [15]--64.364.2----
MaskGroup [45]-65.062.964.7-69.966.669.6
SSTNet [24]42.759.365.564.254.167.873.573.4
SoftGroup [39]51.666.173.666.654.468.975.369.8
Mask3D [35]56.668.468.766.364.575.572.874.5
MAFT [22]-69.1------
QueryFormer [26]57.769.970.572.262.073.372.773.4
EASE (Ours)59.071.669.468.765.376.473.674.6
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.499, + 0.318, + 0.895, + 0.389 + ], + "angle": 0, + "content": "Table 2. Comparison of 3D Instance Segmentation performance with state-of-the-art approaches on the S3DIS [1] Area 5 and 6-fold cross validation set. We evaluate mAP across different IoU thresholds, along with mean precision (mPrec) and mean recall (mRec) at a \\(50\\%\\) IoU threshold over 13 classes of the S3DIS." + }, + { + "type": "table", + "bbox": [ + 0.501, + 0.4, + 0.905, + 0.508 + ], + "angle": 0, + "content": "
MethodmAP\\( \\mathrm{mAP}_{50} \\)\\( \\mathrm{mAP}_{25} \\)
SPFormer [36]25.233.839.6
Mask3D [35]27.437.042.3
QueryFormer [26]28.137.143.4
MAFT [22]29.238.243.3
EASE (Ours)29.938.844.7
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.499, + 0.513, + 0.892, + 0.543 + ], + "angle": 0, + "content": "Table 3. Comparison of 3D Instance Segmentation performance with state-of-the-art approaches on the ScanNet200 [33] val set." + }, + { + "type": "table", + "bbox": [ + 0.501, + 0.555, + 0.905, + 0.662 + ], + "angle": 0, + "content": "
MethodmAP\\( \\mathrm{mAP}_{50} \\)\\( \\mathrm{mAP}_{25} \\)
PointGroup [18]23.338.548.6
HAIS [3]35.146.752.8
SoftGroup [39]47.363.171.4
Mask3D [35]63.479.285.6
EASE (Ours)64.580.886.9
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.499, + 0.667, + 0.892, + 0.696 + ], + "angle": 0, + "content": "Table 4. Comparison of 3D Instance Segmentation performance with state-of-the-art approaches on the STPLS3D [2] test dataset." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.706, + 0.665, + 0.722 + ], + "angle": 0, + "content": "4.3. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.735, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Effect of Semantic and Edge Modules. In Tab. 5, we evaluate the variants of our method w/ and w/o the Semantic Network (SN) and Edge Prediction Module (EPM). The addition of two modules improves 3DIS accuracy across all experiments. The SN especially underscores the significance of incorporating rich semantic knowledge for effective segmentation, achieving \\(+1.3 / 0.7\\) and \\(+1.5 / 1.7\\) improvements on \\(\\mathrm{mAP} / \\mathrm{mAP}_{50}\\) for ScanNetV2 and S3DIS, respectively. EPM further enhances performance via the auxiliary edge information for edge-aware features, resulting in gains of \\(+0.9 / 0.8\\) and \\(+1.9 / 0.9\\). Notably, integrating" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "20649" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.079, + 0.084, + 0.473, + 0.226 + ], + "angle": 0, + "content": "
BaselineMethodScanNet ValS3DIS Area 5
SemanticEdgemAP / mAP50mAP / mAP50
✓--58.4 / 75.956.6 / 68.4
✓✓-59.7 / 76.658.1 / 70.1
✓-✓59.3 / 76.758.5 / 69.3
✓✓✓60.2 / 77.259.0 / 71.6
" + }, + { + "type": "table_caption", + "bbox": [ + 0.078, + 0.23, + 0.47, + 0.259 + ], + "angle": 0, + "content": "Table 5. Ablation study to see the effect of our proposed two main modules: Semantic Network and Edge Prediction Module." + }, + { + "type": "table", + "bbox": [ + 0.079, + 0.263, + 0.473, + 0.417 + ], + "angle": 0, + "content": "
UsageMethodScanNet ValS3DIS Area 5
E-wise SumConcatmAP / mAP50mAP / mAP50
---58.4 / 75.956.6 / 68.4
fcls✓-58.9 / 76.057.8 / 68.5
fcls-✓59.3 / 76.558.3 / 70.1
S·F✓-59.6 / 76.958.6 / 69.6
S·F-✓60.2 / 77.259.0 / 71.6
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.422, + 0.471, + 0.464 + ], + "angle": 0, + "content": "Table 6. Ablation study of leveraging text embeddings to provide semantic details. \\( f_{cls} \\) represents the classification head in the Mask Module, and \\( S \\cdot F \\) stands for the Semantic Fusion network." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.473, + 0.47, + 0.503 + ], + "angle": 0, + "content": "SN and EPM leads to considerable advances \\((+1.8 / 1.3\\) and \\(+2.4 / 3.2)\\), confirming the effectiveness of each module." + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.506, + 0.471, + 0.734 + ], + "angle": 0, + "content": "Applications of Text Embeddings. In Tab. 6, we analyze various applications of text embeddings to determine the most effective strategy for the functional understanding of complex 3D instances. We utilize text embeddings for category classification from \\( f_{cls} \\) of the Mask Module to boost classification accuracy. Also, we strengthen basic queries using the Semantic Fusion (\\( S \\cdot F \\)) network as described in Sec. 3.2. For experiments with the E-wise Sum, we map the text embeddings to the same feature space as the basic queries through MLPs and sum together. While incorporating semantic cues via element-wise sum enhances 3DIS accuracy, it exhibits limited improvements compared to concatenation. By comparison, applying the semantic knowledge across the entire transformer decoder through the \\( S \\cdot F \\) outperforms using them only for the classification head." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.735, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Weight Value of Edge Prediction Loss. To encourage the network to learn practical edge cues, we utilize an edge prediction network. In particular, we compute the weighted binary cross-entropy loss to effectively train the edge extraction head from sparse edge points. Here, we explore the impact of the coefficient value \\( w \\) of \\( \\mathcal{L}_{edge} \\) on 3DIS accuracy. First, as shown in Fig. 5, we empirically find that \\( w \\) influences the thickness of edges in a 3D scene. In Tab. 7, for the ScanNet [4] consisting of relatively small-scale scenes, a smaller value of \\( w \\) (e.g. 2), resulting in thin edges, leads to higher performance. However, for the vast-scale scenes of" + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.07, + 0.631, + 0.158 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.631, + 0.07, + 0.76, + 0.157 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.762, + 0.071, + 0.889, + 0.157 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.5, + 0.163, + 0.892, + 0.205 + ], + "angle": 0, + "content": "Figure 5. Visualization of edge prediction results (red) for different weight values \\( w \\) in the weighted binary cross-entropy loss \\( \\mathcal{L}_{edge} \\), along with corresponding pseudo-edge label (green)." + }, + { + "type": "table", + "bbox": [ + 0.501, + 0.213, + 0.894, + 0.412 + ], + "angle": 0, + "content": "
Feature mapw in LedgeScanNet ValS3DIS Area 5
mAP / mAP50mAP / mAP50
F0259.3 / 76.756.9 / 68.6
F4259.0 / 76.156.7 / 68.5
F0658.9 / 76.358.4 / 68.9
F4658.5 / 76.058.2 / 68.6
F01058.4 / 75.858.5 / 69.3
F41058.5 / 75.758.2 / 69.0
" + }, + { + "type": "table_caption", + "bbox": [ + 0.5, + 0.416, + 0.893, + 0.46 + ], + "angle": 0, + "content": "Table 7. Ablation study to investigate the impact of varying weight values \\( w \\) in the weighted binary cross-entropy \\( {\\mathcal{L}}_{\\text{edge }} \\) ,using high \\( \\left( {F}_{0}\\right) \\) and low \\( \\left( {F}_{4}\\right) \\) resolution feature maps for edge prediction." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.469, + 0.892, + 0.577 + ], + "angle": 0, + "content": "the S3DIS [1], a higher value of \\( w \\) (e.g. 10) leads to better performance with thicker edges. We finally observe the correlation between the scene scale and the weight value \\( w \\); the optimal value for edge learning is proportional to the scale. Besides, methods using high-resolution feature maps (e.g. \\( F_0 \\)) for edge prediction generally perform better than those using low-resolution features maps (e.g. \\( F_4 \\))." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.586, + 0.695, + 0.603 + ], + "angle": 0, + "content": "4.4. Qualitative Analyses" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.611, + 0.892, + 0.853 + ], + "angle": 0, + "content": "Visual Comparison. In this section, we qualitatively confirm the usefulness of our novel framework EASE. We visualize the predicted semantic and instance masks of the baseline model [35] and ours on the ScanNetV2 [4] validation set, in Fig. 6. We emphasize the key differences using green and yellow boxes. As shown in the semantic results (Sem.), ours accurately classifies instances with similar appearances (e.g., desk and bed, door and cabinet) instead of the baseline, which incorrectly recognizes them. Also, as illustrated in the instance mask results (Inst.) of Scene1 and Scene2, ours segments a single object as one, unlike the baseline, which splits it into multiple parts. Further, in Scene 3 (Inst.), ours correctly distinguishes multiple objects (door and cabinet) in contrast to the baseline model. These qualitative results demonstrate that our method provides more detailed and robust segmentation results than the baseline method." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.856, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Impact of Utilizing Semantic Priors. To verify the effectiveness of the Semantic Network, we present t-SNE [38] visualizations of query features for instances with similar" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "20650" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.081, + 0.068, + 0.892, + 0.31 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.321, + 0.895, + 0.365 + ], + "angle": 0, + "content": "Figure 6. Qualitative comparison of 3D Instance Segmentation performance on the ScanNetV2 [4] validation set. We visualize semantic (Sem.) and instance (Inst.) masks of the baseline model and ours with Ground Truth (GT) masks. The critical differences are emphasized using green and yellow-colored boxes. Note that the color map (top right) represents semantic labels. Best viewed in color." + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.375, + 0.459, + 0.514 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.524, + 0.471, + 0.566 + ], + "angle": 0, + "content": "Figure 7. Category probability map (spatial distribution) for queries with high confidence. The x-axis denotes queries, the y-axis represents categories, and the z-axis (right) is probabilities." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.578, + 0.47, + 0.775 + ], + "angle": 0, + "content": "appearances, which pose challenges for 3DIS models. As shown in Fig. 8, the baseline (MAFT [22]) model exhibits disorganized clusters, whereas ours shows relatively distinct feature space for challenging instances. Further, in Fig. 7, we visualize the spatial distribution of category probability for instance queries. The baseline tends to confuse perquery categories and predicts multiple categories for each query with low probability, causing semantic misclassifications. However, our model EASE overcomes this problem using semantic priors, providing clearer insight. These qualitative findings confirm that our semantic network effectively prompts the network to learn instance-specific semantic knowledge from intelligent text embedding priors." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.785, + 0.196, + 0.8 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.811, + 0.471, + 0.901 + ], + "angle": 0, + "content": "Current 3D instance segmentation approaches often face challenges in understanding real-world 3D instances relying solely on geometrical information. Specifically, (i) these methods suffer from identifying instances with similar appearances, and (ii) they often misinterpret the spatial extent of instances, resulting in unclear edges. To tackle these" + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.377, + 0.695, + 0.482 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.696, + 0.377, + 0.888, + 0.482 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.489, + 0.892, + 0.531 + ], + "angle": 0, + "content": "Figure 8. t-SNE [38] visualization of query features representing each instance. Compared to the baseline [22], which produces disorganized clusters, ours creates more distinct clusters." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.543, + 0.892, + 0.648 + ], + "angle": 0, + "content": "challenges, EASE, our novel framework, focuses on context details of text embeddings from the language model as smart priors, enhancing practical semantic understanding. Also, we employ the edge prediction module, guiding the network to reduce misclassifications near edges with edge-aware features. Our extensive experiments verify the effectiveness of EASE, achieving new state-of-the-art scores." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.651, + 0.893, + 0.892 + ], + "angle": 0, + "content": "Acknowledgments This work was supported by Culture, Sports and Tourism R&D Program through the Korea Creative Content Agency grant funded by the Ministry of Culture, Sports and Tourism in 2023. (RS-2023-00227409, Development of sketch-based semantic 3D modeling technology for creating user-centric Metaverse content spaces for indoor spaces, \\(80\\%\\)), (R2022020068, 4D Content Generation and Copyright Protection with Artificial Intelligence, \\(10\\%\\)), Electronics and Telecommunications Research Institute (ETRI) grant funded by the Korean government. (24ZC1200, Research on hyper-realistic interaction technology for five senses and emotional experience, \\(9\\%\\)), and the Institute of Information & communications Technology Planning & Evaluation (IITP) grant funded by the Korea government (MSIT) (2019-0-00079, Artificial Intelligence Graduate School Program (Korea University), \\(1\\%\\))." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "20651" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.079, + 0.09, + 0.176, + 0.107 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.473, + 0.185 + ], + "angle": 0, + "content": "[1] Iro Armeni, Ozan Sener, Amir R Zamir, Helen Jiang, Ioannis Brilakis, Martin Fischer, and Silvio Savarese. 3d semantic parsing of large-scale indoor spaces. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1534-1543, 2016. 2, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.186, + 0.472, + 0.254 + ], + "angle": 0, + "content": "[2] Meida Chen, Qingyong Hu, Zifan Yu, Hugues Thomas, Andrew Feng, Yu Hou, Kyle McCullough, Fengbo Ren, and Lucio Soibelman. Stpls3d: A large-scale synthetic and real aerial photogrammetry 3d point cloud dataset. arXiv preprint arXiv:2203.09065, 2022. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.256, + 0.472, + 0.324 + ], + "angle": 0, + "content": "[3] Shaoyu Chen, Jiemin Fang, Qian Zhang, Wenyu Liu, and Xinggang Wang. Hierarchical aggregation for 3d instance segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15467-15476, 2021. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.326, + 0.472, + 0.395 + ], + "angle": 0, + "content": "[4] Angela Dai, Angel X Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5828-5839, 2017. 2, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.396, + 0.472, + 0.45 + ], + "angle": 0, + "content": "[5] Ruoxi Deng, Chunhua Shen, Shengjun Liu, Huibing Wang, and Xinru Liu. Learning to predict crisp boundaries. In Proceedings of the European conference on computer vision (ECCV), pages 562-578, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.452, + 0.472, + 0.521 + ], + "angle": 0, + "content": "[6] Shenglan Du, Nail Ibrahimli, Jantien Stoter, Julian Kooij, and Liangliang Nan. Push-the-boundary: Boundary-aware feature propagation for semantic segmentation of 3d point clouds. In 2022 International Conference on 3D Vision (3DV), pages 1-10. IEEE, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.522, + 0.472, + 0.577 + ], + "angle": 0, + "content": "[7] Cathrin Elich, Francis Engelmann, Theodora Kontogianni, and Bastian Leibe. 3d bird's-eye-view instance segmentation. In German Conference on Pattern Recognition, pages 48-61. Springer, 2019. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.579, + 0.472, + 0.648 + ], + "angle": 0, + "content": "[8] Francis Engelmann, Martin Bokeloh, Alireza Fathi, Bastian Leibe, and Matthias Nießner. 3d-mpa: Multi-proposal aggregation for 3d semantic instance segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9031-9040, 2020. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.649, + 0.472, + 0.703 + ], + "angle": 0, + "content": "[9] Golnaz Ghiasi, Xiuye Gu, Yin Cui, and Tsung-Yi Lin. Scaling open-vocabulary image segmentation with image-level labels. In European Conference on Computer Vision, pages 540-557. Springer, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.704, + 0.472, + 0.773 + ], + "angle": 0, + "content": "[10] Jingyu Gong, Jiachen Xu, Xin Tan, Jie Zhou, Yanyun Qu, Yuan Xie, and Lizhuang Ma. Boundary-aware geometric encoding for semantic segmentation of point clouds. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1424-1432, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.775, + 0.472, + 0.829 + ], + "angle": 0, + "content": "[11] Xiuye Gu, Tsung-Yi Lin, Weicheng Kuo, and Yin Cui. Open-vocabulary object detection via vision and language knowledge distillation. arXiv preprint arXiv:2104.13921, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.831, + 0.472, + 0.872 + ], + "angle": 0, + "content": "[12] Huy Ha and Shuran Song. Semantic abstraction: Openworld 3d scene understanding from 2d vision-language models. In 6th Annual Conference on Robot Learning, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.873, + 0.472, + 0.903 + ], + "angle": 0, + "content": "[13] Lei Han, Tian Zheng, Lan Xu, and Lu Fang. Occuseg: Occupancy-aware 3d instance segmentation. In Proceedings" + }, + { + "type": "list", + "bbox": [ + 0.079, + 0.115, + 0.473, + 0.903 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "of the IEEE/CVF conference on computer vision and pattern recognition, pages 2940-2949, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.122, + 0.894, + 0.164 + ], + "angle": 0, + "content": "[14] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In Proceedings of the IEEE international conference on computer vision, pages 2961-2969, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.165, + 0.895, + 0.234 + ], + "angle": 0, + "content": "[15] Tong He, Chunhua Shen, and Anton Van Den Hengel. Dyco3d: Robust instance segmentation of 3d point clouds through dynamic convolution. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 354-363, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.235, + 0.895, + 0.291 + ], + "angle": 0, + "content": "[16] Ji Hou, Angela Dai, and Matthias Nießner. 3d-sis: 3d semantic instance segmentation of rgb-d scans. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4421-4430, 2019. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.293, + 0.895, + 0.375 + ], + "angle": 0, + "content": "[17] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In International conference on machine learning, pages 4904-4916. PMLR, 2021. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.377, + 0.895, + 0.447 + ], + "angle": 0, + "content": "[18] Li Jiang, Hengshuang Zhao, Shaoshuai Shi, Shu Liu, Chi-Wing Fu, and Jiaya Jia. Pointgroup: Dual-set point grouping for 3d instance segmentation. In Proceedings of the IEEE/CVF conference on computer vision and Pattern recognition, pages 4867-4876, 2020. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.448, + 0.894, + 0.532 + ], + "angle": 0, + "content": "[19] Abhijit Kundu, Xiaoqi Yin, Alireza Fathi, David Ross, Brian Brewington, Thomas Funkhouser, and Caroline Pantofaru. Virtual multi-view fusion for 3d semantic segmentation. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXIV 16, pages 518-535. Springer, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.533, + 0.894, + 0.589 + ], + "angle": 0, + "content": "[20] Jean Lahoud, Bernard Ghanem, Marc Pollefeys, and Martin R Oswald. 3d instance segmentation via multi-task metric learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9256-9266, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.59, + 0.895, + 0.659 + ], + "angle": 0, + "content": "[21] Xin Lai, Jianhui Liu, Li Jiang, Liwei Wang, Hengshuang Zhao, Shu Liu, Xiaojuan Qi, and Jiaya Jia. Stratified transformer for 3d point cloud segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8500-8509, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.66, + 0.894, + 0.73 + ], + "angle": 0, + "content": "[22] Xin Lai, Yuhui Yuan, Ruihang Chu, Yukang Chen, Han Hu, and Jiaya Jia. Mask-attention-free transformer for 3d instance segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3693–3703, 2023. 1, 2, 3, 4, 5, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.732, + 0.894, + 0.813 + ], + "angle": 0, + "content": "[23] Feng Liang, Bichen Wu, Xiaoliang Dai, Kunpeng Li, Yinan Zhao, Hang Zhang, Peizhao Zhang, Peter Vajda, and Diana Marculescu. Open-vocabulary semantic segmentation with mask-adapted clip. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7061-7070, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.814, + 0.894, + 0.884 + ], + "angle": 0, + "content": "[24] Zhihao Liang, Zhihao Li, Songcen Xu, Mingkui Tan, and Kui Jia. Instance segmentation in 3d scenes using semantic superpoint tree networks. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2783-2792, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.887, + 0.894, + 0.903 + ], + "angle": 0, + "content": "[25] Chen Liu and Yasutaka Furukawa. Masc: Multi-scale affinity" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.895, + 0.903 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "20652" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.092, + 0.47, + 0.121 + ], + "angle": 0, + "content": "with sparse convolution for 3d instance segmentation. arXiv preprint arXiv:1902.04478, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.123, + 0.472, + 0.192 + ], + "angle": 0, + "content": "[26] Jiahao Lu, Jiacheng Deng, Chuxin Wang, Jianfeng He, and Tianzhu Zhang. Query refinement transformer for 3d instance segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 18516-18526, 2023. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.195, + 0.472, + 0.267 + ], + "angle": 0, + "content": "[27] John McCormac, Ankur Handa, Andrew Davison, and Stefan Leutenegger. Semantic fusion: Dense 3d semantic mapping with convolutional neural networks. In 2017 IEEE International Conference on Robotics and automation (ICRA), pages 4628-4635. IEEE, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.268, + 0.472, + 0.323 + ], + "angle": 0, + "content": "[28] Ishan Misra, Rohit Girdhar, and Armand Joulin. An end-to-end transformer model for 3d object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2906-2917, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.325, + 0.472, + 0.395 + ], + "angle": 0, + "content": "[29] Songyou Peng, Kyle Genova, Chiyu Jiang, Andrea Tagliasacchi, Marc Pollefeys, Thomas Funkhouser, et al. Openscene: 3d scene understanding with open vocabularies. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 815-824, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.397, + 0.472, + 0.453 + ], + "angle": 0, + "content": "[30] Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J Guibas. Pointnet++: Deep hierarchical feature learning on point sets in a metric space. Advances in neural information processing systems, 30, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.455, + 0.472, + 0.512 + ], + "angle": 0, + "content": "[31] Charles R Qi, Or Litany, Kaiming He, and Leonidas J Guibas. Deep hough voting for 3d object detection in point clouds. In proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9277-9286, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.513, + 0.472, + 0.596 + ], + "angle": 0, + "content": "[32] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.599, + 0.472, + 0.655 + ], + "angle": 0, + "content": "[33] David Rozenberszki, Or Litany, and Angela Dai. Language-grounded indoor 3d semantic segmentation in the wild. In European Conference on Computer Vision, pages 125-141. Springer, 2022. 2, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.657, + 0.472, + 0.727 + ], + "angle": 0, + "content": "[34] Corentin Sautier, Gilles Puy, Spyros Gidaris, Alexandre Boulch, Andrei Bursuc, and Renaud Marlet. Image-to-lidar self-supervised distillation for autonomous driving data. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9891-9901, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.729, + 0.472, + 0.784 + ], + "angle": 0, + "content": "[35] Jonas Schult, Francis Engelmann, Alexander Hermans, Or Litany, Siyu Tang, and Bastian Leibe. Mask3d for 3d semantic instance segmentation. arXiv preprint arXiv:2210.03105, 2022. 1, 2, 3, 4, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.787, + 0.472, + 0.843 + ], + "angle": 0, + "content": "[36] Jiahao Sun, Chunmei Qing, Junpeng Tan, and Xiangmin Xu. Superpoint transformer for 3d scene instance segmentation. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2393-2401, 2023. 1, 2, 4, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.845, + 0.472, + 0.901 + ], + "angle": 0, + "content": "[37] Ayca Takmaz, Elisabetta Fedele, Robert W Sumner, Marc Pollefeys, Federico Tombari, and Francis Engelmann. Openmask3d: Open-vocabulary 3d instance segmentation. arXiv preprint arXiv:2306.13631, 2023. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.893, + 0.134 + ], + "angle": 0, + "content": "[38] Laurens Van der Maaten and Geoffrey Hinton. Visualizing data using t-sne. Journal of machine learning research, 9 (11), 2008. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.136, + 0.893, + 0.204 + ], + "angle": 0, + "content": "[39] Thang Vu, Kookhoi Kim, Tung M Luu, Thanh Nguyen, and Chang D Yoo. Softgroup for 3d instance segmentation on point clouds. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2708-2717, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.206, + 0.893, + 0.275 + ], + "angle": 0, + "content": "[40] Weiyue Wang, Ronald Yu, Qiangui Huang, and Ulrich Neumann. Sgpn: Similarity group proposal network for 3d point cloud instance segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2569-2578, 2018. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.277, + 0.893, + 0.344 + ], + "angle": 0, + "content": "[41] Xinlong Wang, Shu Liu, Xiaoyong Shen, Chunhua Shen, and Jiaya Jia. Associatively segmenting instances and semantics in point clouds. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4096-4105, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.346, + 0.893, + 0.401 + ], + "angle": 0, + "content": "[42] Yizheng Wu, Min Shi, Shuaiyuan Du, Hao Lu, Zhiguo Cao, and Weicai Zhong. 3d instances as 1d kernels. In European Conference on Computer Vision, pages 235-252. Springer, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.403, + 0.893, + 0.471 + ], + "angle": 0, + "content": "[43] Bo Yang, Jianan Wang, Ronald Clark, Qingyong Hu, Sen Wang, Andrew Markham, and Niki Trigoni. Learning object bounding boxes for 3d instance segmentation on point clouds. Advances in neural information processing systems, 32, 2019. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.473, + 0.893, + 0.543 + ], + "angle": 0, + "content": "[44] Li Yi, Wang Zhao, He Wang, Minhyuk Sung, and Leonidas J Guibas. Gspn: Generative shape proposal network for 3d instance segmentation in point cloud. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3947-3956, 2019. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.545, + 0.893, + 0.613 + ], + "angle": 0, + "content": "[45] Min Zhong, Xinghao Chen, Xiaokang Chen, Gang Zeng, and Yunhe Wang. Maskgroup: Hierarchical point grouping and masking for 3d instance segmentation. In 2022 IEEE International Conference on Multimedia and Expo (ICME), pages 1-6. IEEE, 2022. 6" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.893, + 0.613 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20653" + } + ] +] \ No newline at end of file diff --git a/2024/Edge-Aware 3D Instance Segmentation Network with Intelligent Semantic Prior/ef85b90b-3f1b-4322-ac5a-c5d6ae60802c_origin.pdf b/2024/Edge-Aware 3D Instance Segmentation Network with Intelligent Semantic Prior/ef85b90b-3f1b-4322-ac5a-c5d6ae60802c_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..4898c97519ea2c241c55879bf1a3d906eee5234f --- /dev/null +++ b/2024/Edge-Aware 3D Instance Segmentation Network with Intelligent Semantic Prior/ef85b90b-3f1b-4322-ac5a-c5d6ae60802c_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:cfd3f58e0ac3b3b4265ff5b3cfac13403f6a1011edebc1bd0bbbbfb3f2bab114 +size 5116646 diff --git a/2024/Edge-Aware 3D Instance Segmentation Network with Intelligent Semantic Prior/full.md b/2024/Edge-Aware 3D Instance Segmentation Network with Intelligent Semantic Prior/full.md new file mode 100644 index 0000000000000000000000000000000000000000..9bf75c951d910fa110835c02d1753c463d71a9ff --- /dev/null +++ b/2024/Edge-Aware 3D Instance Segmentation Network with Intelligent Semantic Prior/full.md @@ -0,0 +1,289 @@ +# Edge-Aware 3D Instance Segmentation Network with Intelligent Semantic Prior + +Wonseok Roh1 Hwanhee Jung1 Giljoo Nam2 Jinseop Yeom1 Hyunjep Park1 Sang Ho Yoon3 Sanggil Kim1* +1Korea University 2Meta Reality Labs 3KAIST + +# Abstract + +While recent 3D instance segmentation approaches show promising results based on transformer architectures, they often fail to correctly identify instances with similar appearances. They also ambiguously determine edges, leading to multiple misclassifications of adjacent edge points. In this work, we introduce a novel framework, called EASE, to overcome these challenges and improve the perception of complex 3D instances. We first propose a semantic guidance network to leverage rich semantic knowledge from a language model as intelligent priors, enhancing the functional understanding of real-world instances beyond relying solely on geometrical information. We explicitly instruct the basic instance queries using text embeddings of each instance to learn deep semantic details. Further, we utilize the edge prediction module, encouraging the segmentation network to be edge-aware. We extract voxel-wise edge maps from point features and use them as auxiliary information for learning edge cues. In our extensive experiments on large-scale benchmarks, ScanNetV2, ScanNet200, S3DIS, and STPLS3D, our EASE outperforms existing state-of-the-art models, demonstrating its superior performance. + +# 1. Introduction + +Understanding 3D scenes is a fundamental task within 3D computer vision. Given 3D point cloud scenes, identifying and perceiving instances on sparse points, along with assigning semantic class labels, play a crucial role in a comprehensive understanding of the entire spatial environment. + +In real-world 3D scenarios, significant occlusion and truncation often occur, especially when objects are overlapped or hidden by others. To address these issues, conventional works [3, 7, 8, 16, 18, 40, 43, 44] in 3D instance segmentation (3DIS) primarily focus on accurately generating region proposals (top-down) [16, 43, 44] or effectively grouping points with clustering algorithms (bottom-up) [3, 7, 8, 18, 40]. Inspired by the remarkable sensation of Mask-RCNN [14], the former (proposal-based) meth + +![](images/ef5c115d04f726dae5371777545edb2e15643339b4bcb8ae951cc343d6944cce.jpg) +(b) Instance misinterpretation cases +Figure 1. Examples of 3DIS for two challenging cases. (a) Semantic failure cases: the baseline [35] confuses seat-shaped objects as chair or toilet (Scene1) and also misclassifies large cuboid blocks as desk (Scene2). And (b) Instance misinterpretation cases: baseline erroneously segments a single object into multiple parts (Scene3) or merges multiple objects into one instance (Scene4). + +ods initially detect instances as bounding boxes and predict masks within each proposed region. However, these strategies are susceptible to the quality of the detection results. On the other hand, the latter (grouping-based) methods use clustering algorithms to group closely related points and aggregate point-wise class labels and instance features. While they show great advancements in 3DIS, these methods require additional manually tuned processing, such as point grouping [18] or voting mechanisms [31], to determine specific geometric properties (e.g., centers, occupancy). + +Recently, transformer-based 3DIS frameworks [22, 26, 35, 36] have tackled several limitations of traditional approaches by introducing a fully end-to-end pipeline. They directly predict masks from 3D points without resource- + +exhaustive processing. These methods have achieved high performance and quickly become predominant in the 3DIS domain. However, despite these breakthroughs, they still face challenges in accurately masking everyday 3D instances relying solely on geometrical information. Specifically, they frequently suffer from cases where each instance appears similarly, but their semantic roles differ, as illustrated in Fig. 1 (a). Thus, gaining insight into the semantic properties of instances beyond mere appearance or position is crucial for effectively addressing complex real-world 3D scenarios. From this observation, we consider leveraging semantic knowledge along with the geometry information. + +Built upon the classic structure of the transformer-based framework [22, 35, 36], which trains queries that encode instance-specific knowledge for direct mask prediction, we introduce a carefully designed semantic guidance network. One of the most promising ways to utilize semantic information is through visual language models [17, 32]. They learn generous visual-linguistic knowledge from large-scale image-text pairs, which empowers them to achieve outstanding progress in various 2D or 3D vision tasks [9, 11, 23, 29]. In this work, motivated by the success of these works, we explicitly instruct the transformer-based network to learn contextual variations among instances using text embeddings. Specifically, we strengthen the basic instance queries with profound semantic clues of text embeddings representing each instance. Here, the context details can serve as discriminative semantic priors for perceiving complex 3D instances adequately. Through our extensive experiments, we found that our semantic-guided approach effectively overcomes the challenges of existing methods. + +Different from the human vision system, which easily distinguishes boundaries between instances, we empirically observe that most 3DIS models [21, 22, 35, 36, 42] struggle with precisely capturing the boundaries in the 3D space. These methods often misinterpret the spatial range of instances and ambiguously determine their edges, leading to numerous misclassifications of adjacent edge points, as shown in Fig. 1 (b). We consider that these issues arise due to the absence of detailed guidance on inter-object boundaries, resulting in fuzzy edges. To tackle these challenges, we advocate for leveraging the edge prediction module to pilot the whole network to exploit edge-advanced features. We predict edge points across point cloud scenes and operate them as auxiliary information for learning practical edge cues, which encourages accurate recognition of the 3D spatial scope. Note that we supervise this module using dynamically generated point-wise pseudo edge labels. + +Given landmark datasets for 3DIS, ScanNetV2 [4], ScanNet200 [33], S3DIS [1], and STPLS3D [2], we validate the effectiveness and robustness of our framework. Our method outperforms the existing state-of-the-art methods. To summarize, our main contributions are listed as follows: + +- We propose EASE, a novel 3D instance segmentation framework that utilizes the rich semantic clues from the language model. Specifically, context details of text embeddings serve as smart semantic priors, effectively enhancing the functional comprehension of 3D instances. +- We introduce the edge prediction module guiding the entire network to take advantage of edge-aware features. Also, we utilize the module's output as auxiliary knowledge for learning edge cues, boosting 3DIS performance. +- We analyze the effectiveness of our proposed method on multiple challenging benchmarks, including ScanNetV2, ScanNet200, S3DIS, and STPLS3D. Extensive experiments on various 3D scenarios validate that our method achieves new State-of-the-Art performance on 3DIS. + +# 2. Related Work + +3D Instance Segmentation. The 3D Instance Segmentation (3DIS) task aims to identify individual instances and assign semantic classes to each point within a 3D point scene. Traditional approaches in 3DIS are primarily categorized into three groups: proposal-based [16, 43, 44], grouping-based [3, 7, 8, 18, 40], and transformer-based methods [22, 35, 36]. The proposal-based approaches initially detect object proposals, such as 3D bounding boxes, and then estimate corresponding per-point semantic classes and instance masks. One of the grouping-based approaches, PointGroup [18], conducts point-wise clustering to group points into instances using dual point sets, which comprise original and shifted coordinates. Recently, transformer-based approaches predict instance queries that represent individual objects and decode them into per-point categories and instance labels directly. Mask3D [35] refines queries using masked cross-attention with hierarchical point features, leading the queries to focus on particular instances. SPFormer [36] updates instance queries via transformer decoder utilizing pre-computed superpoints, which can reduce the whole computational cost of the network. MAFT [22], on the other hand, recognizes representational limitations in the initial mask and replaces the masking process with object-localizing position queries. Based on the core idea of transformer-based methods, our novel approach incorporates a semantic guidance network and an auxiliary edge network to improve instance mask segmentation. + +Semantic Guided 3D Scene Understanding. Recent research in 3D scene understanding focuses on comprehending large-scale, real-world 3D environments. Conventional studies in this field are broadly categorized into two strategies: closed-set 3D scene understanding, a prevalent approach in diverse tasks [19, 27, 33, 34], and open-vocabulary methods [12, 29, 37], which focus on recognizing unseen categories within datasets. More recently, there has been significant progress in integrating pre-trained visual-language models like CLIP [32] or ALIGN [17] into + +![](images/6a7ca8f42b59b14a63cb9d7bdcdb2e6781b2dcae5dba120aefe28498d2ca4a84.jpg) +Figure 2. An overview of our framework EASE. Built upon the classic architecture of the transformer-based 3DIS model, our model takes 3D scenes and directly infers instance masks. Our model consists of four main modules: (1) Sparse Convolutional Backbone, (2) Semantic Network, which strengthen the queries to learn contextual variations of instances with rich semantic embeddings from the pre-trained language model CLIP, (3) Mask Transformer Decoder with Mask Module and Query Refinement blocks, and (4) Edge Prediction Module, where our edge extraction head is trained to predict a voxel-wise edge map, encouraging the network to utilize edge-advanced features. + +these methodologies to enhance semantic understanding. Building on this idea, for instance, [33] utilizes a closed-set framework and enhances 3D feature learning by aligning it with CLIP text embedding space for semantic classification. In this work, we employ contextual semantic priors to improve the comprehension of complex 3D instances, enhancing the overall 3D instance segmentation performance. + +# 3. Method + +In this section, we introduce a novel 3D Instance Segmentation (3DIS) framework EASE leveraging the rich semantic guidance from the language model. Furthermore, we propose the edge prediction module, which encourages the pipeline to use edge-aware features for effectively perceiving instance boundaries. We provide an overview of the entire pipeline (Sec. 3.1, Fig. 2, and Alg. 1) and then elaborate on the details of our method: (i) Intelligent Semantic Prior (Sec. 3.2) and (ii) Edge-Aware 3DIS Framework (Sec. 3.3). + +# 3.1. Overview + +Our end-to-end 3DIS framework, as illustrated in Fig. 2 and Alg. 1, aims to enhance the comprehension of various instances across 3D point cloud scenes. Based on the classic transformer-based architecture [22, 35], which operates standard transformer building blocks, we directly infer instance masks from 3D point clouds. Our model consists + +of four main modules: (1) Sparse Convolutional Backbone, (2) Semantic Network, which provide the deep semantic cues of diverse real-world instances, (3) Mask Transformer Decoder with Mask Module and Query Refinement blocks, and (4) Edge Prediction Module, which benefits the whole network to take advantage of edge-aware features. + +Mask Transformer. First, the Sparse Convolutional U-Net Backbone takes a colored point cloud $P \in \mathbb{R}^{N_p \times 6}$ as input and voxelizes $P$ into $X_0$ voxels $V \in \mathbb{R}^{X_0 \times 3}$ to extract a multi-resolution hierarchical feature maps $F_i \in \mathbb{R}^{X_i \times D}$ , where $i \in \{0,1,2,3,4\}$ . Following [28, 35], we set zero-initialized non-parametric instance queries $Q \in \mathbb{R}^{N_q \times D}$ , referring to point positions sampled with furthest point sampling (FPS) [30]. Given the $F_i$ and $Q$ , the mask transformer decoder layer iteratively enhances the queries using the Mask Module (MM) and Query Refinement (QR) blocks. In the MM, as shown in Fig. 3 (a), we classify the category $c_i$ for $i = \{1,2,\dots,N_c\}$ of each query via linear classification head $f_{\mathrm{class}}$ using following cross-entropy loss: + +$$ +\mathcal {L} _ {c l s} = - \mathbb {E} _ {c, w _ {c} \sim \mathbb {D}} \left[ \sum_ {r \in N _ {c}} w _ {c} [ r ] \log f _ {\text {c l a s s}} (Q) [ r ] \right] \tag {1} +$$ + +where $w_{c}$ denotes one-hot encoded category labels and $\mathbb{D}$ represents (input) data distribution. Also, the $Q$ are fed through the query head $f_{\mathrm{query}}$ to project them to the same feature space as $F_{0}$ for mask prediction. Afterwards, we + +Algorithm 1: Overview of our framework EASE. +Input: Colored point cloud $P\in \mathbb{R}^{N_p\times 6}$ Result: Binary 3DIS mask $\hat{M}_f\in \{0,1\}^{X\times N_q}$ Procedure: # Sparse Convolutional Backbone. 1 Extract multi-resolution feature maps $F_{i}$ from $P$ # Edge Prediction Module (EPM). 2 Predict edge map $\hat{E}$ using full-resolution feature $F_{0}$ 3 Supervise $\hat{E}$ with pseudo edge label $E$ 4 Encourage the network to learn edge-aware features. #Mask Transformer Decoder. 5 for $l\in N_l$ do 6 for $i\in \{1,2,3,4\}$ do 7 Update queries $Q$ using $M_{attn}$ and $F_{i}$ # Semantic Network. 8 Match the set of classes with text labels. 9 Encode text into text embeds $T$ using CLIP. 10 Reinforce $Q$ to contain semantic clues of $T$ 11 end 12 end + +compute the similarity between the projected query features $Q'$ and $F_0$ using the dot product operation, then calculate the probability of the instance mask employing the sigmoid function as follows. + +$$ +M _ {a t t n} = \left\{m _ {i, j} = \left[ \sigma \left(F _ {0} \cdot f _ {\text {q u e r y}} (Q) ^ {T}\right) _ {i, j} > 0. 5 \right] \right\} \tag {2} +$$ + +where the threshold value is 0.5 for binary mask. Further, to refine the query representation, we leverage the $QR$ block, including masked cross-attention and self-attention mechanisms. Here, we utilize $M_{att}$ as the foreground mask for the masked cross-attention layer, where $Q$ iteratively attend to each multi-scale feature $(F_1 - F_4, i \geqslant 1)$ as follows: + +$$ +Q = \operatorname {s o f t m a x} \left(\mathbf {Q K} ^ {T} / \sqrt {D} + M _ {\text {a t t n}}\right) \mathbf {V} \tag {3} +$$ + +where $\mathbf{K}$ and $\mathbf{V}$ are linearly projected from each point feature, and $\mathbf{Q}$ are from $Q$ . Hence, the transformer focuses more on valuable instances instead of the unessential background. Subsequently, in the self-attention layer, the keys, values, and queries are all linear projections of $Q$ from the cross-attention layer. Finally, the transformer decoder layer is recurrently applied for multiple iterations $N_{l}$ and feature scales $i$ , ultimately producing the final set of refined queries. + +# 3.2. Intelligent Semantic Prior + +It is important to highlight that precise geometric properties (e.g., locations, distances and orientations) from 3D points significantly enhance 3DIS accuracy. While previous studies [22, 35, 36] benefit from the generous geometric potential of physical information, they are still limited in perceiving intricate real-world 3D instances. Hence, we pro + +![](images/b48eadeeee65a1d6a26e024a61b87e555937b855f6be5210dd54b4b0144e0d77.jpg) +(a) Transformer Decoder Layer + +![](images/2a3295bc7d96b84dc109c8b3b80d0cc7cd2ddf67fc5a276e6ff747e0d144deee.jpg) +(b) Semantic Network +Figure 3. Detailed architecture of (a) Transformer Decoder Layer: consists of Mask Module and Query Refinement blocks. And (b) Semantic Network: we first map the categories of queries $Q$ with the corresponding text description for text embeddings $T$ from CLIP [32]. Then, we explicitly strengthen the $Q$ with $T$ using the Semantic Fusion network to learn semantic details. + +pose to utilize semantic knowledge of individual instances as smart priors, effectively addressing the challenges. + +Deep Semantic Knowledge for 3D Instances. Recently, visual language models [17, 32] have highlighted the benefits of their capability to provide general embedded knowledge from large-scale image-text multimodal data. In this work, we explicitly instruct the transformer-based 3DIS network to learn contextual variations among instances by utilizing the power of the prevalent language model CLIP [32]. To achieve this goal, we utilize a carefully designed Semantic Network, as shown in Fig. 3 (b). Specifically, with the refined query $Q$ from each $QR$ block, we first categorize each query following the linear classification head $f_{\mathrm{class}}$ in the MM. Given the set of category candidates, we map the category number $c_{i}$ for $i = \{1,2,\dots,N_c\}$ with the corresponding text descriptions (e.g., "Cabinet", "Bookshelf"). Next, these text labels go through the large-scale pre-trained language model, which outputs text embedding vectors $T$ . We then concatenate $Q$ with $T$ and strengthen the basic instance queries with deep semantic clues from text embeddings using the Semantic Fusion network as follows: + +$$ +Q = \operatorname {S i g m o i d} \left(\mathbf {W} _ {2} \cdot \phi \left(\mathbf {W} _ {1} \cdot (Q \oplus T)\right) ^ {T}\right) \tag {4} +$$ + +where $\oplus$ indicates channel-wise concatenation, $\mathbf{W}_i$ denotes the learnable parameters of the $i$ -th linear layer, and $\phi$ is the non-linear activation function. We apply this process for each iteration of the $QR$ . Ultimately, we consider these explicitly guided queries with contextual variations as discriminative semantic priors for the practical perception of complex 3D instances, enhancing 3DIS accuracy. + +# 3.3. Edge-Aware 3DIS Framework + +In real-world 3D scenarios, instances are commonly positioned in diverse arrangements without following standard rules or patterns. Especially when instances are closely ad + +![](images/a2ea6740bf0d611246be912d5ba9e3a51ed2ef290a51c7db716348dd7eeb2998.jpg) +Figure 4. Pseudo edge label calculation using KNN algorithm. We collect $k$ neighboring points of each point (red border) and compare their instance labels. If the count of different label points (blue and purple) among its surroundings surpasses a predefined threshold $\tau$ , we annotate the central point as an edge point (green). + +jacent or overlapping, it becomes increasingly challenging for the model to define the edges precisely. Therefore, even with significant progress in technical strategies, most conventional methods [22, 35, 36] often misinterpret the spatial extent of each instance and ambiguously estimate their edges as illustrated in Fig. 1. To this end, we introduce an edge-aware 3DIS framework with an edge prediction module, enabling the network to utilize edge-advanced features. + +Pseudo Edge Label Calculation. To optimize the edge prediction module, we first compute point-wise pseudo edge labels for all 3D point scenes. We operate a traditional $k$ -Nearest Neighbors (KNN) classification algorithm based on the KD-tree data structure to explore the $k$ nearest points. Here, we utilize point-wise instance labels from datasets. For all points in a 3D scene, we set each point as a central point and compare their instance labels with those of $k$ neighbor points. As shown in Fig. 4, if the count of distinct label points among its surroundings exceeds a predefined threshold $\tau$ , we identify the central point as a boundary point. We finally generate pseudo-binary edge maps $E' \in \{0,1\}^{N_p \times 1}$ , where points on the edges are 1 and internal points are 0 for the entire set of points $P$ . Then, we voxelize the point-wise edge labels into the voxel-wise edge labels $E \in \{0,1\}^{X_0 \times 1}$ to supervise the edge extraction head. Note that we precompute the pseudo edge labels for every 3D scene in the datasets before training. + +Learning Edge Cues. With dynamically generated pseudo edge labels $E$ , we aim to enrich regular features to recognize the accurate spatial scope of 3D instances with precise boundaries. Specifically, we introduce an Edge Prediction Module, which predicts a voxel-wise dense edge map $\hat{E} \in (0,1)^{X_0 \times 1}$ to interactively guide the network in capturing edges of various instances. Given a full-resolution feature map $F_0$ from the backbone network, our edge extraction head, which consists of multiple shared MLPs, estimates the probabilities of edges for each voxel as follows: + +$$ +\hat {E} = \operatorname {M L P} _ {\text {e d g e}} \left(F _ {0}\right) = \operatorname {S i g m o i d} \left(\mathbf {W} _ {2} \cdot \phi \left(\mathbf {W} _ {1} \cdot F _ {0}\right) ^ {T}\right) \tag {5} +$$ + +where $\mathbf{W}_i$ denotes the learnable parameters of the $i$ -th linear layer, and $\phi$ indicates the non-linear activation function. + +Following [6, 10], as edge voxels constitute a small portion of the entire voxel set, we formulate weighted binary cross entropy loss to train the edge extraction head as follows: + +$$ +\mathcal {L} _ {e d g e} = - \sum_ {i = 1} ^ {X _ {0}} \left(w \cdot E _ {i} \log \hat {E} _ {i} + \left(1 - E _ {i}\right) \log \left(1 - \hat {E} _ {i}\right)\right) (6) +$$ + +where $w$ represent the weight value used to balance the substantial difference between the numbers of edge points and others. Then, we propagate the edge cues of prediction as auxiliary information for the backbone network, improving 3D instance understanding with edge-aware features. Furthermore, we train our edge network jointly in an end-to-end manner with low computational costs. + +Loss Function We finally predict instance mask $\hat{M}_f$ using the final set of queries in the $MM$ . For training, we compute the following loss $\mathcal{L}_{mask}$ as a sum of the two losses: + +$$ +\mathcal {L} _ {\text {m a s k}} = \lambda_ {\mathrm {B C E}} \mathcal {L} _ {\mathrm {B C E}} \left(M _ {f}, \hat {M} _ {f}\right) + \lambda_ {\mathrm {d i c e}} \mathcal {L} _ {\mathrm {d i c e}} \left(M _ {f}, \hat {M} _ {f}\right) \tag {7} +$$ + +where $M_{f}$ denotes the ground-truth instance mask, and $\mathcal{L}_{\text{dice}}$ represents the Dice loss [5]. Ultimately, our model is trained end-to-end by minimizing the following loss $\mathcal{L}_{\text{total}}$ : + +$$ +\mathcal {L} _ {\text {t o t a l}} = \lambda_ {\text {m a s k}} \mathcal {L} _ {\text {m a s k}} + \lambda_ {\text {c l s}} \mathcal {L} _ {\text {c l s}} + \lambda_ {\text {e d g e}} \mathcal {L} _ {\text {e d g e}} \tag {8} +$$ + +where each $\lambda$ is a hyperparameter derived from grid searches to handle the strength of respective loss term. + +# 4. Experiments + +# 4.1. Experimental Setup + +Datasets. In this study, we train and evaluate the overall performance using four landmark datasets for 3D instance segmentation: ScanNetV2 [4], ScanNet200 [33], S3DIS [1], and STPLS3D [2]. These four datasets provide 3D point cloud scan data collected in various scenarios. Note that detailed descriptions of each dataset and all implementation details are provided in the supplemental material. + +Evaluation Metrics. To evaluate the 3D instance segmentation accuracy, we operate Average Precision (AP), a typical metric for various computer vision tasks. We report mean average precision (mAP), which represents an average score with IoU thresholds ranging from $50\%$ to $95\%$ with a step size of $5\%$ . Also, we provide $\mathrm{mAP}_{50}$ and $\mathrm{mAP}_{25}$ , indicating the scores with IoU thresholds of $50\%$ and $25\%$ , respectively. Further, we estimate mean precision (mPrec) and mean recall (mRec) on the S3DIS [1] dataset. + +# 4.2. Performance Comparison with SOTA Methods. + +ScanNetV2. We quantitatively compare our proposed network EASE with existing state-of-the-art methods. These methods demonstrate significant capabilities, but they still rely on geometric information from 3D points and often + +
MethodScanNet ValScanNet Test
mAP\( mAP_{50} \)mAP\( mAP_{50} \)
GSPN [44]19.337.8-30.6
3D-SIS [16]-18.716.138.2
MASC [25]--25.444.7
3D-Bonet [43]--25.348.8
MTML [20]20.340.228.254.9
3D-MPA [8]35.559.135.561.1
DyCo3D [15]35.457.639.564.1
PointGroup [18]34.856.740.763.6
MaskGroup [45]42.063.343.466.4
OccuSeg [13]44.260.748.667.2
SSTNet [24]49.464.350.669.8
HAIS [3]43.564.145.769.9
SoftGroup [39]46.067.650.476.1
Mask3D [35]55.273.756.678.0
QueryFormer [26]56.574.258.378.7
MAFT [22]59.976.559.678.6
EASE (Ours)60.277.259.878.9
+ +Table 1. Comparison of 3D Instance Segmentation performance with state-of-the-art approaches on the ScanNetV2 [4] dataset. We evaluate mean average precision (mAP) with different IoU thresholds over 18 classes on the ScanNetV2 validation / hidden test set. + +face challenges in effectively comprehending 3D instances. To address these limitations, we present a novel edge-aware framework featuring a semantic guidance network. As reported in Tab. 1, EASE generally outperforms other methods, achieving new state-of-the-art accuracy in terms of $\mathrm{mAP}$ and $\mathrm{mAP}_{50}$ on both validation (60.2 / 77.2) and hidden test (59.8 / 78.9) sets. These results confirm that our edge cues and smart semantic prior benefits the entire network, leading to high precision in 3D instance segmentation. + +S3DIS. In Tab. 2, we evaluate 3DIS performance on Area 5 and 6-fold cross-validation of the S3DIS [1] dataset. For Area 5 evaluation, we employ data from Area 5 for validation and the other areas for training. Also, for 6-fold cross-validation, we assess validation scores across six other areas and compute the average. In both evaluations, EASE demonstrates significant performance improvements up to $+2.4 / 2.5$ (Area 5) and $+0.8 / 0.9$ (6-fold) for $\mathrm{mAP} / \mathrm{mAP}_{50}$ . + +ScanNet200. We also demonstrate considerable performance on the ScanNet200 validation set (see Tab. 3). Our method EASE precisely segments 3D instances for various categories compared to current state-of-the-art approaches. + +STPLS3D. Remarkably, our EASE network also achieves higher scores on the STPLS3D [2] dataset, which consists of outdoor 3D point cloud scenes. EASE exceeds the second-best results with $+1.1 / 1.6 / 1.3$ margins in $\mathrm{mAP} / \mathrm{mAP}_{50} / \mathrm{mAP}_{25}$ . These results highlight the effectiveness of our method in understanding diverse real-world 3D instances, both from indoor and outdoor environments. + +
MethodS3DIS Area 5S3DIS 6-fold CV
AP\( AP_{50} \)\( Prec_{50} \)\( Rec_{50} \)AP\( AP_{50} \)\( Prec_{50} \)\( Rec_{50} \)
ASIS [41]--55.342.4--63.647.5
3D-Bonet [43]--57.540.2--65.647.6
3D-MPA [8]--63.158.0--66.764.1
PointGroup [18]-57.861.962.1-64.069.669.2
DyCo3D [15]--64.364.2----
MaskGroup [45]-65.062.964.7-69.966.669.6
SSTNet [24]42.759.365.564.254.167.873.573.4
SoftGroup [39]51.666.173.666.654.468.975.369.8
Mask3D [35]56.668.468.766.364.575.572.874.5
MAFT [22]-69.1------
QueryFormer [26]57.769.970.572.262.073.372.773.4
EASE (Ours)59.071.669.468.765.376.473.674.6
+ +Table 2. Comparison of 3D Instance Segmentation performance with state-of-the-art approaches on the S3DIS [1] Area 5 and 6-fold cross validation set. We evaluate mAP across different IoU thresholds, along with mean precision (mPrec) and mean recall (mRec) at a $50\%$ IoU threshold over 13 classes of the S3DIS. + +
MethodmAP\( \mathrm{mAP}_{50} \)\( \mathrm{mAP}_{25} \)
SPFormer [36]25.233.839.6
Mask3D [35]27.437.042.3
QueryFormer [26]28.137.143.4
MAFT [22]29.238.243.3
EASE (Ours)29.938.844.7
+ +Table 3. Comparison of 3D Instance Segmentation performance with state-of-the-art approaches on the ScanNet200 [33] val set. + +
MethodmAP\( \mathrm{mAP}_{50} \)\( \mathrm{mAP}_{25} \)
PointGroup [18]23.338.548.6
HAIS [3]35.146.752.8
SoftGroup [39]47.363.171.4
Mask3D [35]63.479.285.6
EASE (Ours)64.580.886.9
+ +Table 4. Comparison of 3D Instance Segmentation performance with state-of-the-art approaches on the STPLS3D [2] test dataset. + +# 4.3. Ablation Studies + +Effect of Semantic and Edge Modules. In Tab. 5, we evaluate the variants of our method w/ and w/o the Semantic Network (SN) and Edge Prediction Module (EPM). The addition of two modules improves 3DIS accuracy across all experiments. The SN especially underscores the significance of incorporating rich semantic knowledge for effective segmentation, achieving $+1.3 / 0.7$ and $+1.5 / 1.7$ improvements on $\mathrm{mAP} / \mathrm{mAP}_{50}$ for ScanNetV2 and S3DIS, respectively. EPM further enhances performance via the auxiliary edge information for edge-aware features, resulting in gains of $+0.9 / 0.8$ and $+1.9 / 0.9$ . Notably, integrating + +
BaselineMethodScanNet ValS3DIS Area 5
SemanticEdgemAP / mAP50mAP / mAP50
✓--58.4 / 75.956.6 / 68.4
✓✓-59.7 / 76.658.1 / 70.1
✓-✓59.3 / 76.758.5 / 69.3
✓✓✓60.2 / 77.259.0 / 71.6
+ +Table 5. Ablation study to see the effect of our proposed two main modules: Semantic Network and Edge Prediction Module. + +
UsageMethodScanNet ValS3DIS Area 5
E-wise SumConcatmAP / mAP50mAP / mAP50
---58.4 / 75.956.6 / 68.4
fcls✓-58.9 / 76.057.8 / 68.5
fcls-✓59.3 / 76.558.3 / 70.1
S·F✓-59.6 / 76.958.6 / 69.6
S·F-✓60.2 / 77.259.0 / 71.6
+ +SN and EPM leads to considerable advances $(+1.8 / 1.3$ and $+2.4 / 3.2)$ , confirming the effectiveness of each module. + +Applications of Text Embeddings. In Tab. 6, we analyze various applications of text embeddings to determine the most effective strategy for the functional understanding of complex 3D instances. We utilize text embeddings for category classification from $f_{cls}$ of the Mask Module to boost classification accuracy. Also, we strengthen basic queries using the Semantic Fusion ( $S \cdot F$ ) network as described in Sec. 3.2. For experiments with the E-wise Sum, we map the text embeddings to the same feature space as the basic queries through MLPs and sum together. While incorporating semantic cues via element-wise sum enhances 3DIS accuracy, it exhibits limited improvements compared to concatenation. By comparison, applying the semantic knowledge across the entire transformer decoder through the $S \cdot F$ outperforms using them only for the classification head. + +Weight Value of Edge Prediction Loss. To encourage the network to learn practical edge cues, we utilize an edge prediction network. In particular, we compute the weighted binary cross-entropy loss to effectively train the edge extraction head from sparse edge points. Here, we explore the impact of the coefficient value $w$ of $\mathcal{L}_{edge}$ on 3DIS accuracy. First, as shown in Fig. 5, we empirically find that $w$ influences the thickness of edges in a 3D scene. In Tab. 7, for the ScanNet [4] consisting of relatively small-scale scenes, a smaller value of $w$ (e.g. 2), resulting in thin edges, leads to higher performance. However, for the vast-scale scenes of + +![](images/67150f2bb3f008f5f4080bc18427ccd6e3e230d0fac2c9e55341405fba91698a.jpg) +Figure 5. Visualization of edge prediction results (red) for different weight values $w$ in the weighted binary cross-entropy loss $\mathcal{L}_{edge}$ , along with corresponding pseudo-edge label (green). + +![](images/eadb8bdd0196298d2b073c8324cf301ce7f4c1d51468330c29f648cb5dd1fd1e.jpg) + +![](images/afe05085399c88d9dc529810348cf903a93784fd1484e1bb2bbf4ddc574de952.jpg) + +Table 6. Ablation study of leveraging text embeddings to provide semantic details. $f_{cls}$ represents the classification head in the Mask Module, and $S \cdot F$ stands for the Semantic Fusion network. + +
Feature mapw in LedgeScanNet ValS3DIS Area 5
mAP / mAP50mAP / mAP50
F0259.3 / 76.756.9 / 68.6
F4259.0 / 76.156.7 / 68.5
F0658.9 / 76.358.4 / 68.9
F4658.5 / 76.058.2 / 68.6
F01058.4 / 75.858.5 / 69.3
F41058.5 / 75.758.2 / 69.0
+ +Table 7. Ablation study to investigate the impact of varying weight values $w$ in the weighted binary cross-entropy ${\mathcal{L}}_{\text{edge }}$ ,using high $\left( {F}_{0}\right)$ and low $\left( {F}_{4}\right)$ resolution feature maps for edge prediction. + +the S3DIS [1], a higher value of $w$ (e.g. 10) leads to better performance with thicker edges. We finally observe the correlation between the scene scale and the weight value $w$ ; the optimal value for edge learning is proportional to the scale. Besides, methods using high-resolution feature maps (e.g. $F_0$ ) for edge prediction generally perform better than those using low-resolution features maps (e.g. $F_4$ ). + +# 4.4. Qualitative Analyses + +Visual Comparison. In this section, we qualitatively confirm the usefulness of our novel framework EASE. We visualize the predicted semantic and instance masks of the baseline model [35] and ours on the ScanNetV2 [4] validation set, in Fig. 6. We emphasize the key differences using green and yellow boxes. As shown in the semantic results (Sem.), ours accurately classifies instances with similar appearances (e.g., desk and bed, door and cabinet) instead of the baseline, which incorrectly recognizes them. Also, as illustrated in the instance mask results (Inst.) of Scene1 and Scene2, ours segments a single object as one, unlike the baseline, which splits it into multiple parts. Further, in Scene 3 (Inst.), ours correctly distinguishes multiple objects (door and cabinet) in contrast to the baseline model. These qualitative results demonstrate that our method provides more detailed and robust segmentation results than the baseline method. + +Impact of Utilizing Semantic Priors. To verify the effectiveness of the Semantic Network, we present t-SNE [38] visualizations of query features for instances with similar + +![](images/50b6fd6cbc393982acb1f116a4e3f476ecbae0c108ec67106571f8c33707713f.jpg) +Figure 6. Qualitative comparison of 3D Instance Segmentation performance on the ScanNetV2 [4] validation set. We visualize semantic (Sem.) and instance (Inst.) masks of the baseline model and ours with Ground Truth (GT) masks. The critical differences are emphasized using green and yellow-colored boxes. Note that the color map (top right) represents semantic labels. Best viewed in color. + +![](images/27c0198d788fc79971e1ea3d9b0e6de9c8ae2fbd834829a37b9ab2a8933846b7.jpg) +Figure 7. Category probability map (spatial distribution) for queries with high confidence. The x-axis denotes queries, the y-axis represents categories, and the z-axis (right) is probabilities. + +appearances, which pose challenges for 3DIS models. As shown in Fig. 8, the baseline (MAFT [22]) model exhibits disorganized clusters, whereas ours shows relatively distinct feature space for challenging instances. Further, in Fig. 7, we visualize the spatial distribution of category probability for instance queries. The baseline tends to confuse perquery categories and predicts multiple categories for each query with low probability, causing semantic misclassifications. However, our model EASE overcomes this problem using semantic priors, providing clearer insight. These qualitative findings confirm that our semantic network effectively prompts the network to learn instance-specific semantic knowledge from intelligent text embedding priors. + +# 5. Conclusion + +Current 3D instance segmentation approaches often face challenges in understanding real-world 3D instances relying solely on geometrical information. Specifically, (i) these methods suffer from identifying instances with similar appearances, and (ii) they often misinterpret the spatial extent of instances, resulting in unclear edges. To tackle these + +![](images/35c883f6ce8d111bed3542168efd7a733723329ce7901c8de1571996464441b6.jpg) +Figure 8. t-SNE [38] visualization of query features representing each instance. Compared to the baseline [22], which produces disorganized clusters, ours creates more distinct clusters. + +![](images/91836c3bd17e74dc6b126d22e7f4d844031d3a862a617c6aa1c0847345d5a6d2.jpg) + +challenges, EASE, our novel framework, focuses on context details of text embeddings from the language model as smart priors, enhancing practical semantic understanding. Also, we employ the edge prediction module, guiding the network to reduce misclassifications near edges with edge-aware features. Our extensive experiments verify the effectiveness of EASE, achieving new state-of-the-art scores. + +Acknowledgments This work was supported by Culture, Sports and Tourism R&D Program through the Korea Creative Content Agency grant funded by the Ministry of Culture, Sports and Tourism in 2023. (RS-2023-00227409, Development of sketch-based semantic 3D modeling technology for creating user-centric Metaverse content spaces for indoor spaces, $80\%$ ), (R2022020068, 4D Content Generation and Copyright Protection with Artificial Intelligence, $10\%$ ), Electronics and Telecommunications Research Institute (ETRI) grant funded by the Korean government. (24ZC1200, Research on hyper-realistic interaction technology for five senses and emotional experience, $9\%$ ), and the Institute of Information & communications Technology Planning & Evaluation (IITP) grant funded by the Korea government (MSIT) (2019-0-00079, Artificial Intelligence Graduate School Program (Korea University), $1\%$ ). + +# References + +[1] Iro Armeni, Ozan Sener, Amir R Zamir, Helen Jiang, Ioannis Brilakis, Martin Fischer, and Silvio Savarese. 3d semantic parsing of large-scale indoor spaces. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1534-1543, 2016. 2, 5, 6, 7 +[2] Meida Chen, Qingyong Hu, Zifan Yu, Hugues Thomas, Andrew Feng, Yu Hou, Kyle McCullough, Fengbo Ren, and Lucio Soibelman. Stpls3d: A large-scale synthetic and real aerial photogrammetry 3d point cloud dataset. arXiv preprint arXiv:2203.09065, 2022. 2, 5, 6 +[3] Shaoyu Chen, Jiemin Fang, Qian Zhang, Wenyu Liu, and Xinggang Wang. Hierarchical aggregation for 3d instance segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15467-15476, 2021. 1, 2, 6 +[4] Angela Dai, Angel X Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5828-5839, 2017. 2, 5, 6, 7, 8 +[5] Ruoxi Deng, Chunhua Shen, Shengjun Liu, Huibing Wang, and Xinru Liu. Learning to predict crisp boundaries. In Proceedings of the European conference on computer vision (ECCV), pages 562-578, 2018. 5 +[6] Shenglan Du, Nail Ibrahimli, Jantien Stoter, Julian Kooij, and Liangliang Nan. Push-the-boundary: Boundary-aware feature propagation for semantic segmentation of 3d point clouds. In 2022 International Conference on 3D Vision (3DV), pages 1-10. IEEE, 2022. 5 +[7] Cathrin Elich, Francis Engelmann, Theodora Kontogianni, and Bastian Leibe. 3d bird's-eye-view instance segmentation. In German Conference on Pattern Recognition, pages 48-61. Springer, 2019. 1, 2 +[8] Francis Engelmann, Martin Bokeloh, Alireza Fathi, Bastian Leibe, and Matthias Nießner. 3d-mpa: Multi-proposal aggregation for 3d semantic instance segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9031-9040, 2020. 1, 2, 6 +[9] Golnaz Ghiasi, Xiuye Gu, Yin Cui, and Tsung-Yi Lin. Scaling open-vocabulary image segmentation with image-level labels. In European Conference on Computer Vision, pages 540-557. Springer, 2022. 2 +[10] Jingyu Gong, Jiachen Xu, Xin Tan, Jie Zhou, Yanyun Qu, Yuan Xie, and Lizhuang Ma. Boundary-aware geometric encoding for semantic segmentation of point clouds. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1424-1432, 2021. 5 +[11] Xiuye Gu, Tsung-Yi Lin, Weicheng Kuo, and Yin Cui. Open-vocabulary object detection via vision and language knowledge distillation. arXiv preprint arXiv:2104.13921, 2021. 2 +[12] Huy Ha and Shuran Song. Semantic abstraction: Openworld 3d scene understanding from 2d vision-language models. In 6th Annual Conference on Robot Learning, 2022. 2 +[13] Lei Han, Tian Zheng, Lan Xu, and Lu Fang. Occuseg: Occupancy-aware 3d instance segmentation. In Proceedings + +of the IEEE/CVF conference on computer vision and pattern recognition, pages 2940-2949, 2020. 6 +[14] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In Proceedings of the IEEE international conference on computer vision, pages 2961-2969, 2017. 1 +[15] Tong He, Chunhua Shen, and Anton Van Den Hengel. Dyco3d: Robust instance segmentation of 3d point clouds through dynamic convolution. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 354-363, 2021. 6 +[16] Ji Hou, Angela Dai, and Matthias Nießner. 3d-sis: 3d semantic instance segmentation of rgb-d scans. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4421-4430, 2019. 1, 2, 6 +[17] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In International conference on machine learning, pages 4904-4916. PMLR, 2021. 2, 4 +[18] Li Jiang, Hengshuang Zhao, Shaoshuai Shi, Shu Liu, Chi-Wing Fu, and Jiaya Jia. Pointgroup: Dual-set point grouping for 3d instance segmentation. In Proceedings of the IEEE/CVF conference on computer vision and Pattern recognition, pages 4867-4876, 2020. 1, 2, 6 +[19] Abhijit Kundu, Xiaoqi Yin, Alireza Fathi, David Ross, Brian Brewington, Thomas Funkhouser, and Caroline Pantofaru. Virtual multi-view fusion for 3d semantic segmentation. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXIV 16, pages 518-535. Springer, 2020. 2 +[20] Jean Lahoud, Bernard Ghanem, Marc Pollefeys, and Martin R Oswald. 3d instance segmentation via multi-task metric learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9256-9266, 2019. 6 +[21] Xin Lai, Jianhui Liu, Li Jiang, Liwei Wang, Hengshuang Zhao, Shu Liu, Xiaojuan Qi, and Jiaya Jia. Stratified transformer for 3d point cloud segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8500-8509, 2022. 2 +[22] Xin Lai, Yuhui Yuan, Ruihang Chu, Yukang Chen, Han Hu, and Jiaya Jia. Mask-attention-free transformer for 3d instance segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3693–3703, 2023. 1, 2, 3, 4, 5, 6, 8 +[23] Feng Liang, Bichen Wu, Xiaoliang Dai, Kunpeng Li, Yinan Zhao, Hang Zhang, Peizhao Zhang, Peter Vajda, and Diana Marculescu. Open-vocabulary semantic segmentation with mask-adapted clip. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7061-7070, 2023. 2 +[24] Zhihao Liang, Zhihao Li, Songcen Xu, Mingkui Tan, and Kui Jia. Instance segmentation in 3d scenes using semantic superpoint tree networks. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2783-2792, 2021. 6 +[25] Chen Liu and Yasutaka Furukawa. Masc: Multi-scale affinity + +with sparse convolution for 3d instance segmentation. arXiv preprint arXiv:1902.04478, 2019. 6 +[26] Jiahao Lu, Jiacheng Deng, Chuxin Wang, Jianfeng He, and Tianzhu Zhang. Query refinement transformer for 3d instance segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 18516-18526, 2023. 1, 6 +[27] John McCormac, Ankur Handa, Andrew Davison, and Stefan Leutenegger. Semantic fusion: Dense 3d semantic mapping with convolutional neural networks. In 2017 IEEE International Conference on Robotics and automation (ICRA), pages 4628-4635. IEEE, 2017. 2 +[28] Ishan Misra, Rohit Girdhar, and Armand Joulin. An end-to-end transformer model for 3d object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2906-2917, 2021. 3 +[29] Songyou Peng, Kyle Genova, Chiyu Jiang, Andrea Tagliasacchi, Marc Pollefeys, Thomas Funkhouser, et al. Openscene: 3d scene understanding with open vocabularies. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 815-824, 2023. 2 +[30] Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J Guibas. Pointnet++: Deep hierarchical feature learning on point sets in a metric space. Advances in neural information processing systems, 30, 2017. 3 +[31] Charles R Qi, Or Litany, Kaiming He, and Leonidas J Guibas. Deep hough voting for 3d object detection in point clouds. In proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9277-9286, 2019. 1 +[32] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 2, 4 +[33] David Rozenberszki, Or Litany, and Angela Dai. Language-grounded indoor 3d semantic segmentation in the wild. In European Conference on Computer Vision, pages 125-141. Springer, 2022. 2, 3, 5, 6 +[34] Corentin Sautier, Gilles Puy, Spyros Gidaris, Alexandre Boulch, Andrei Bursuc, and Renaud Marlet. Image-to-lidar self-supervised distillation for autonomous driving data. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9891-9901, 2022. 2 +[35] Jonas Schult, Francis Engelmann, Alexander Hermans, Or Litany, Siyu Tang, and Bastian Leibe. Mask3d for 3d semantic instance segmentation. arXiv preprint arXiv:2210.03105, 2022. 1, 2, 3, 4, 5, 6, 7 +[36] Jiahao Sun, Chunmei Qing, Junpeng Tan, and Xiangmin Xu. Superpoint transformer for 3d scene instance segmentation. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2393-2401, 2023. 1, 2, 4, 5, 6 +[37] Ayca Takmaz, Elisabetta Fedele, Robert W Sumner, Marc Pollefeys, Federico Tombari, and Francis Engelmann. Openmask3d: Open-vocabulary 3d instance segmentation. arXiv preprint arXiv:2306.13631, 2023. 2 + +[38] Laurens Van der Maaten and Geoffrey Hinton. Visualizing data using t-sne. Journal of machine learning research, 9 (11), 2008. 7, 8 +[39] Thang Vu, Kookhoi Kim, Tung M Luu, Thanh Nguyen, and Chang D Yoo. Softgroup for 3d instance segmentation on point clouds. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2708-2717, 2022. 6 +[40] Weiyue Wang, Ronald Yu, Qiangui Huang, and Ulrich Neumann. Sgpn: Similarity group proposal network for 3d point cloud instance segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2569-2578, 2018. 1, 2 +[41] Xinlong Wang, Shu Liu, Xiaoyong Shen, Chunhua Shen, and Jiaya Jia. Associatively segmenting instances and semantics in point clouds. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4096-4105, 2019. 6 +[42] Yizheng Wu, Min Shi, Shuaiyuan Du, Hao Lu, Zhiguo Cao, and Weicai Zhong. 3d instances as 1d kernels. In European Conference on Computer Vision, pages 235-252. Springer, 2022. 2 +[43] Bo Yang, Jianan Wang, Ronald Clark, Qingyong Hu, Sen Wang, Andrew Markham, and Niki Trigoni. Learning object bounding boxes for 3d instance segmentation on point clouds. Advances in neural information processing systems, 32, 2019. 1, 2, 6 +[44] Li Yi, Wang Zhao, He Wang, Minhyuk Sung, and Leonidas J Guibas. Gspn: Generative shape proposal network for 3d instance segmentation in point cloud. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3947-3956, 2019. 1, 2, 6 +[45] Min Zhong, Xinghao Chen, Xiaokang Chen, Gang Zeng, and Yunhe Wang. Maskgroup: Hierarchical point grouping and masking for 3d instance segmentation. In 2022 IEEE International Conference on Multimedia and Expo (ICME), pages 1-6. IEEE, 2022. 6 \ No newline at end of file diff --git a/2024/Edge-Aware 3D Instance Segmentation Network with Intelligent Semantic Prior/images.zip b/2024/Edge-Aware 3D Instance Segmentation Network with Intelligent Semantic Prior/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..ed6e2a7789286d2cfcfbd86c04e27d4526d82f88 --- /dev/null +++ b/2024/Edge-Aware 3D Instance Segmentation Network with Intelligent Semantic Prior/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c5725625b659bac95b5ea4f62c8907e030ca88a35b29e8a2915ab9f073e6d3d9 +size 857356 diff --git a/2024/Edge-Aware 3D Instance Segmentation Network with Intelligent Semantic Prior/layout.json b/2024/Edge-Aware 3D Instance Segmentation Network with Intelligent Semantic Prior/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..176318dcce80dbc81cc97552a09dd499b3d96270 --- /dev/null +++ b/2024/Edge-Aware 3D Instance Segmentation Network with Intelligent Semantic Prior/layout.json @@ -0,0 +1,8936 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 51, + 103, + 543, + 120 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 51, + 103, + 543, + 120 + ], + "spans": [ + { + "bbox": [ + 51, + 103, + 543, + 120 + ], + "type": "text", + "content": "Edge-Aware 3D Instance Segmentation Network with Intelligent Semantic Prior" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 142, + 459, + 194 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 142, + 459, + 194 + ], + "spans": [ + { + "bbox": [ + 133, + 142, + 459, + 194 + ], + "type": "text", + "content": "Wonseok Roh1 Hwanhee Jung1 Giljoo Nam2 Jinseop Yeom1 Hyunjep Park1 Sang Ho Yoon3 Sanggil Kim1* \n1Korea University 2Meta Reality Labs 3KAIST" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 143, + 217, + 191, + 228 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 217, + 191, + 228 + ], + "spans": [ + { + "bbox": [ + 143, + 217, + 191, + 228 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 232, + 290, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 232, + 290, + 483 + ], + "spans": [ + { + "bbox": [ + 46, + 232, + 290, + 483 + ], + "type": "text", + "content": "While recent 3D instance segmentation approaches show promising results based on transformer architectures, they often fail to correctly identify instances with similar appearances. They also ambiguously determine edges, leading to multiple misclassifications of adjacent edge points. In this work, we introduce a novel framework, called EASE, to overcome these challenges and improve the perception of complex 3D instances. We first propose a semantic guidance network to leverage rich semantic knowledge from a language model as intelligent priors, enhancing the functional understanding of real-world instances beyond relying solely on geometrical information. We explicitly instruct the basic instance queries using text embeddings of each instance to learn deep semantic details. Further, we utilize the edge prediction module, encouraging the segmentation network to be edge-aware. We extract voxel-wise edge maps from point features and use them as auxiliary information for learning edge cues. In our extensive experiments on large-scale benchmarks, ScanNetV2, ScanNet200, S3DIS, and STPLS3D, our EASE outperforms existing state-of-the-art models, demonstrating its superior performance." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 498, + 127, + 510 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 498, + 127, + 510 + ], + "spans": [ + { + "bbox": [ + 47, + 498, + 127, + 510 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 517, + 287, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 517, + 287, + 578 + ], + "spans": [ + { + "bbox": [ + 46, + 517, + 287, + 578 + ], + "type": "text", + "content": "Understanding 3D scenes is a fundamental task within 3D computer vision. Given 3D point cloud scenes, identifying and perceiving instances on sparse points, along with assigning semantic class labels, play a crucial role in a comprehensive understanding of the entire spatial environment." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 578, + 287, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 578, + 287, + 685 + ], + "spans": [ + { + "bbox": [ + 46, + 578, + 287, + 685 + ], + "type": "text", + "content": "In real-world 3D scenarios, significant occlusion and truncation often occur, especially when objects are overlapped or hidden by others. To address these issues, conventional works [3, 7, 8, 16, 18, 40, 43, 44] in 3D instance segmentation (3DIS) primarily focus on accurately generating region proposals (top-down) [16, 43, 44] or effectively grouping points with clustering algorithms (bottom-up) [3, 7, 8, 18, 40]. Inspired by the remarkable sensation of Mask-RCNN [14], the former (proposal-based) meth" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 310, + 217, + 545, + 460 + ], + "blocks": [ + { + "bbox": [ + 310, + 217, + 545, + 460 + ], + "lines": [ + { + "bbox": [ + 310, + 217, + 545, + 460 + ], + "spans": [ + { + "bbox": [ + 310, + 217, + 545, + 460 + ], + "type": "image", + "image_path": "ef5c115d04f726dae5371777545edb2e15643339b4bcb8ae951cc343d6944cce.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 380, + 460, + 479, + 469 + ], + "lines": [ + { + "bbox": [ + 380, + 460, + 479, + 469 + ], + "spans": [ + { + "bbox": [ + 380, + 460, + 479, + 469 + ], + "type": "text", + "content": "(b) Instance misinterpretation cases" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 304, + 474, + 545, + 540 + ], + "lines": [ + { + "bbox": [ + 304, + 474, + 545, + 540 + ], + "spans": [ + { + "bbox": [ + 304, + 474, + 545, + 540 + ], + "type": "text", + "content": "Figure 1. Examples of 3DIS for two challenging cases. (a) Semantic failure cases: the baseline [35] confuses seat-shaped objects as chair or toilet (Scene1) and also misclassifies large cuboid blocks as desk (Scene2). And (b) Instance misinterpretation cases: baseline erroneously segments a single object into multiple parts (Scene3) or merges multiple objects into one instance (Scene4)." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 545, + 546, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 545, + 546, + 665 + ], + "spans": [ + { + "bbox": [ + 304, + 545, + 546, + 665 + ], + "type": "text", + "content": "ods initially detect instances as bounding boxes and predict masks within each proposed region. However, these strategies are susceptible to the quality of the detection results. On the other hand, the latter (grouping-based) methods use clustering algorithms to group closely related points and aggregate point-wise class labels and instance features. While they show great advancements in 3DIS, these methods require additional manually tuned processing, such as point grouping [18] or voting mechanisms [31], to determine specific geometric properties (e.g., centers, occupancy)." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "content": "Recently, transformer-based 3DIS frameworks [22, 26, 35, 36] have tackled several limitations of traditional approaches by introducing a fully end-to-end pipeline. They directly predict masks from 3D points without resource-" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 494, + 35 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 494, + 35 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 494, + 35 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 693, + 279, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 693, + 279, + 713 + ], + "spans": [ + { + "bbox": [ + 58, + 693, + 279, + 713 + ], + "type": "text", + "content": "*Corresponding author. \nProject Page: https://kuai-lab.github.io/ease2024." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20644" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 216 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 216 + ], + "type": "text", + "content": "exhaustive processing. These methods have achieved high performance and quickly become predominant in the 3DIS domain. However, despite these breakthroughs, they still face challenges in accurately masking everyday 3D instances relying solely on geometrical information. Specifically, they frequently suffer from cases where each instance appears similarly, but their semantic roles differ, as illustrated in Fig. 1 (a). Thus, gaining insight into the semantic properties of instances beyond mere appearance or position is crucial for effectively addressing complex real-world 3D scenarios. From this observation, we consider leveraging semantic knowledge along with the geometry information." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 218, + 289, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 218, + 289, + 445 + ], + "spans": [ + { + "bbox": [ + 46, + 218, + 289, + 445 + ], + "type": "text", + "content": "Built upon the classic structure of the transformer-based framework [22, 35, 36], which trains queries that encode instance-specific knowledge for direct mask prediction, we introduce a carefully designed semantic guidance network. One of the most promising ways to utilize semantic information is through visual language models [17, 32]. They learn generous visual-linguistic knowledge from large-scale image-text pairs, which empowers them to achieve outstanding progress in various 2D or 3D vision tasks [9, 11, 23, 29]. In this work, motivated by the success of these works, we explicitly instruct the transformer-based network to learn contextual variations among instances using text embeddings. Specifically, we strengthen the basic instance queries with profound semantic clues of text embeddings representing each instance. Here, the context details can serve as discriminative semantic priors for perceiving complex 3D instances adequately. Through our extensive experiments, we found that our semantic-guided approach effectively overcomes the challenges of existing methods." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 448, + 289, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 448, + 289, + 651 + ], + "spans": [ + { + "bbox": [ + 46, + 448, + 289, + 651 + ], + "type": "text", + "content": "Different from the human vision system, which easily distinguishes boundaries between instances, we empirically observe that most 3DIS models [21, 22, 35, 36, 42] struggle with precisely capturing the boundaries in the 3D space. These methods often misinterpret the spatial range of instances and ambiguously determine their edges, leading to numerous misclassifications of adjacent edge points, as shown in Fig. 1 (b). We consider that these issues arise due to the absence of detailed guidance on inter-object boundaries, resulting in fuzzy edges. To tackle these challenges, we advocate for leveraging the edge prediction module to pilot the whole network to exploit edge-advanced features. We predict edge points across point cloud scenes and operate them as auxiliary information for learning practical edge cues, which encourages accurate recognition of the 3D spatial scope. Note that we supervise this module using dynamically generated point-wise pseudo edge labels." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 653, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 289, + 715 + ], + "type": "text", + "content": "Given landmark datasets for 3DIS, ScanNetV2 [4], ScanNet200 [33], S3DIS [1], and STPLS3D [2], we validate the effectiveness and robustness of our framework. Our method outperforms the existing state-of-the-art methods. To summarize, our main contributions are listed as follows:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 306, + 72, + 547, + 240 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 306, + 72, + 547, + 132 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 547, + 132 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 547, + 132 + ], + "type": "text", + "content": "- We propose EASE, a novel 3D instance segmentation framework that utilizes the rich semantic clues from the language model. Specifically, context details of text embeddings serve as smart semantic priors, effectively enhancing the functional comprehension of 3D instances." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 306, + 133, + 547, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 133, + 547, + 180 + ], + "spans": [ + { + "bbox": [ + 306, + 133, + 547, + 180 + ], + "type": "text", + "content": "- We introduce the edge prediction module guiding the entire network to take advantage of edge-aware features. Also, we utilize the module's output as auxiliary knowledge for learning edge cues, boosting 3DIS performance." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 180, + 547, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 180, + 547, + 240 + ], + "spans": [ + { + "bbox": [ + 306, + 180, + 547, + 240 + ], + "type": "text", + "content": "- We analyze the effectiveness of our proposed method on multiple challenging benchmarks, including ScanNetV2, ScanNet200, S3DIS, and STPLS3D. Extensive experiments on various 3D scenarios validate that our method achieves new State-of-the-Art performance on 3DIS." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 251, + 392, + 264 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 251, + 392, + 264 + ], + "spans": [ + { + "bbox": [ + 306, + 251, + 392, + 264 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 268, + 547, + 591 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 268, + 547, + 591 + ], + "spans": [ + { + "bbox": [ + 304, + 268, + 547, + 591 + ], + "type": "text", + "content": "3D Instance Segmentation. The 3D Instance Segmentation (3DIS) task aims to identify individual instances and assign semantic classes to each point within a 3D point scene. Traditional approaches in 3DIS are primarily categorized into three groups: proposal-based [16, 43, 44], grouping-based [3, 7, 8, 18, 40], and transformer-based methods [22, 35, 36]. The proposal-based approaches initially detect object proposals, such as 3D bounding boxes, and then estimate corresponding per-point semantic classes and instance masks. One of the grouping-based approaches, PointGroup [18], conducts point-wise clustering to group points into instances using dual point sets, which comprise original and shifted coordinates. Recently, transformer-based approaches predict instance queries that represent individual objects and decode them into per-point categories and instance labels directly. Mask3D [35] refines queries using masked cross-attention with hierarchical point features, leading the queries to focus on particular instances. SPFormer [36] updates instance queries via transformer decoder utilizing pre-computed superpoints, which can reduce the whole computational cost of the network. MAFT [22], on the other hand, recognizes representational limitations in the initial mask and replaces the masking process with object-localizing position queries. Based on the core idea of transformer-based methods, our novel approach incorporates a semantic guidance network and an auxiliary edge network to improve instance mask segmentation." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 594, + 548, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 548, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 548, + 715 + ], + "type": "text", + "content": "Semantic Guided 3D Scene Understanding. Recent research in 3D scene understanding focuses on comprehending large-scale, real-world 3D environments. Conventional studies in this field are broadly categorized into two strategies: closed-set 3D scene understanding, a prevalent approach in diverse tasks [19, 27, 33, 34], and open-vocabulary methods [12, 29, 37], which focus on recognizing unseen categories within datasets. More recently, there has been significant progress in integrating pre-trained visual-language models like CLIP [32] or ALIGN [17] into" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "20645" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 52, + 75, + 545, + 311 + ], + "blocks": [ + { + "bbox": [ + 52, + 75, + 545, + 311 + ], + "lines": [ + { + "bbox": [ + 52, + 75, + 545, + 311 + ], + "spans": [ + { + "bbox": [ + 52, + 75, + 545, + 311 + ], + "type": "image", + "image_path": "6a7ca8f42b59b14a63cb9d7bdcdb2e6781b2dcae5dba120aefe28498d2ca4a84.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 321, + 546, + 377 + ], + "lines": [ + { + "bbox": [ + 46, + 321, + 546, + 377 + ], + "spans": [ + { + "bbox": [ + 46, + 321, + 546, + 377 + ], + "type": "text", + "content": "Figure 2. An overview of our framework EASE. Built upon the classic architecture of the transformer-based 3DIS model, our model takes 3D scenes and directly infers instance masks. Our model consists of four main modules: (1) Sparse Convolutional Backbone, (2) Semantic Network, which strengthen the queries to learn contextual variations of instances with rich semantic embeddings from the pre-trained language model CLIP, (3) Mask Transformer Decoder with Mask Module and Query Refinement blocks, and (4) Edge Prediction Module, where our edge extraction head is trained to predict a voxel-wise edge map, encouraging the network to utilize edge-advanced features." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 386, + 287, + 469 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 386, + 287, + 469 + ], + "spans": [ + { + "bbox": [ + 46, + 386, + 287, + 469 + ], + "type": "text", + "content": "these methodologies to enhance semantic understanding. Building on this idea, for instance, [33] utilizes a closed-set framework and enhances 3D feature learning by aligning it with CLIP text embedding space for semantic classification. In this work, we employ contextual semantic priors to improve the comprehension of complex 3D instances, enhancing the overall 3D instance segmentation performance." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 483, + 103, + 495 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 483, + 103, + 495 + ], + "spans": [ + { + "bbox": [ + 47, + 483, + 103, + 495 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 504, + 287, + 612 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 504, + 287, + 612 + ], + "spans": [ + { + "bbox": [ + 46, + 504, + 287, + 612 + ], + "type": "text", + "content": "In this section, we introduce a novel 3D Instance Segmentation (3DIS) framework EASE leveraging the rich semantic guidance from the language model. Furthermore, we propose the edge prediction module, which encourages the pipeline to use edge-aware features for effectively perceiving instance boundaries. We provide an overview of the entire pipeline (Sec. 3.1, Fig. 2, and Alg. 1) and then elaborate on the details of our method: (i) Intelligent Semantic Prior (Sec. 3.2) and (ii) Edge-Aware 3DIS Framework (Sec. 3.3)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 622, + 115, + 634 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 622, + 115, + 634 + ], + "spans": [ + { + "bbox": [ + 47, + 622, + 115, + 634 + ], + "type": "text", + "content": "3.1. Overview" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 641, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 641, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 641, + 288, + 713 + ], + "type": "text", + "content": "Our end-to-end 3DIS framework, as illustrated in Fig. 2 and Alg. 1, aims to enhance the comprehension of various instances across 3D point cloud scenes. Based on the classic transformer-based architecture [22, 35], which operates standard transformer building blocks, we directly infer instance masks from 3D point clouds. Our model consists" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 386, + 545, + 457 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 386, + 545, + 457 + ], + "spans": [ + { + "bbox": [ + 304, + 386, + 545, + 457 + ], + "type": "text", + "content": "of four main modules: (1) Sparse Convolutional Backbone, (2) Semantic Network, which provide the deep semantic cues of diverse real-world instances, (3) Mask Transformer Decoder with Mask Module and Query Refinement blocks, and (4) Edge Prediction Module, which benefits the whole network to take advantage of edge-aware features." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 460, + 546, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 460, + 546, + 616 + ], + "spans": [ + { + "bbox": [ + 304, + 460, + 546, + 616 + ], + "type": "text", + "content": "Mask Transformer. First, the Sparse Convolutional U-Net Backbone takes a colored point cloud " + }, + { + "bbox": [ + 304, + 460, + 546, + 616 + ], + "type": "inline_equation", + "content": "P \\in \\mathbb{R}^{N_p \\times 6}" + }, + { + "bbox": [ + 304, + 460, + 546, + 616 + ], + "type": "text", + "content": " as input and voxelizes " + }, + { + "bbox": [ + 304, + 460, + 546, + 616 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 304, + 460, + 546, + 616 + ], + "type": "text", + "content": " into " + }, + { + "bbox": [ + 304, + 460, + 546, + 616 + ], + "type": "inline_equation", + "content": "X_0" + }, + { + "bbox": [ + 304, + 460, + 546, + 616 + ], + "type": "text", + "content": " voxels " + }, + { + "bbox": [ + 304, + 460, + 546, + 616 + ], + "type": "inline_equation", + "content": "V \\in \\mathbb{R}^{X_0 \\times 3}" + }, + { + "bbox": [ + 304, + 460, + 546, + 616 + ], + "type": "text", + "content": " to extract a multi-resolution hierarchical feature maps " + }, + { + "bbox": [ + 304, + 460, + 546, + 616 + ], + "type": "inline_equation", + "content": "F_i \\in \\mathbb{R}^{X_i \\times D}" + }, + { + "bbox": [ + 304, + 460, + 546, + 616 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 460, + 546, + 616 + ], + "type": "inline_equation", + "content": "i \\in \\{0,1,2,3,4\\}" + }, + { + "bbox": [ + 304, + 460, + 546, + 616 + ], + "type": "text", + "content": ". Following [28, 35], we set zero-initialized non-parametric instance queries " + }, + { + "bbox": [ + 304, + 460, + 546, + 616 + ], + "type": "inline_equation", + "content": "Q \\in \\mathbb{R}^{N_q \\times D}" + }, + { + "bbox": [ + 304, + 460, + 546, + 616 + ], + "type": "text", + "content": ", referring to point positions sampled with furthest point sampling (FPS) [30]. Given the " + }, + { + "bbox": [ + 304, + 460, + 546, + 616 + ], + "type": "inline_equation", + "content": "F_i" + }, + { + "bbox": [ + 304, + 460, + 546, + 616 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 460, + 546, + 616 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 304, + 460, + 546, + 616 + ], + "type": "text", + "content": ", the mask transformer decoder layer iteratively enhances the queries using the Mask Module (MM) and Query Refinement (QR) blocks. In the MM, as shown in Fig. 3 (a), we classify the category " + }, + { + "bbox": [ + 304, + 460, + 546, + 616 + ], + "type": "inline_equation", + "content": "c_i" + }, + { + "bbox": [ + 304, + 460, + 546, + 616 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 304, + 460, + 546, + 616 + ], + "type": "inline_equation", + "content": "i = \\{1,2,\\dots,N_c\\}" + }, + { + "bbox": [ + 304, + 460, + 546, + 616 + ], + "type": "text", + "content": " of each query via linear classification head " + }, + { + "bbox": [ + 304, + 460, + 546, + 616 + ], + "type": "inline_equation", + "content": "f_{\\mathrm{class}}" + }, + { + "bbox": [ + 304, + 460, + 546, + 616 + ], + "type": "text", + "content": " using following cross-entropy loss:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 322, + 624, + 545, + 657 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 322, + 624, + 545, + 657 + ], + "spans": [ + { + "bbox": [ + 322, + 624, + 545, + 657 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {c l s} = - \\mathbb {E} _ {c, w _ {c} \\sim \\mathbb {D}} \\left[ \\sum_ {r \\in N _ {c}} w _ {c} [ r ] \\log f _ {\\text {c l a s s}} (Q) [ r ] \\right] \\tag {1}", + "image_path": "97bf08ea6e0a5f458d00a9611d6223be90eab7f73afc6ea6a526688f74f78617.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 665, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 546, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 665, + 546, + 713 + ], + "type": "inline_equation", + "content": "w_{c}" + }, + { + "bbox": [ + 304, + 665, + 546, + 713 + ], + "type": "text", + "content": " denotes one-hot encoded category labels and " + }, + { + "bbox": [ + 304, + 665, + 546, + 713 + ], + "type": "inline_equation", + "content": "\\mathbb{D}" + }, + { + "bbox": [ + 304, + 665, + 546, + 713 + ], + "type": "text", + "content": " represents (input) data distribution. Also, the " + }, + { + "bbox": [ + 304, + 665, + 546, + 713 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 304, + 665, + 546, + 713 + ], + "type": "text", + "content": " are fed through the query head " + }, + { + "bbox": [ + 304, + 665, + 546, + 713 + ], + "type": "inline_equation", + "content": "f_{\\mathrm{query}}" + }, + { + "bbox": [ + 304, + 665, + 546, + 713 + ], + "type": "text", + "content": " to project them to the same feature space as " + }, + { + "bbox": [ + 304, + 665, + 546, + 713 + ], + "type": "inline_equation", + "content": "F_{0}" + }, + { + "bbox": [ + 304, + 665, + 546, + 713 + ], + "type": "text", + "content": " for mask prediction. Afterwards, we" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "20646" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "code", + "bbox": [ + 48, + 88, + 287, + 326 + ], + "blocks": [ + { + "bbox": [ + 54, + 75, + 255, + 87 + ], + "lines": [ + { + "bbox": [ + 54, + 75, + 255, + 87 + ], + "spans": [ + { + "bbox": [ + 54, + 75, + 255, + 87 + ], + "type": "text", + "content": "Algorithm 1: Overview of our framework EASE." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "code_caption" + }, + { + "bbox": [ + 48, + 88, + 287, + 326 + ], + "lines": [ + { + "bbox": [ + 48, + 88, + 287, + 326 + ], + "spans": [ + { + "bbox": [ + 48, + 88, + 287, + 326 + ], + "type": "text", + "content": "Input: Colored point cloud " + }, + { + "bbox": [ + 48, + 88, + 287, + 326 + ], + "type": "inline_equation", + "content": "P\\in \\mathbb{R}^{N_p\\times 6}" + }, + { + "bbox": [ + 48, + 88, + 287, + 326 + ], + "type": "text", + "content": " Result: Binary 3DIS mask " + }, + { + "bbox": [ + 48, + 88, + 287, + 326 + ], + "type": "inline_equation", + "content": "\\hat{M}_f\\in \\{0,1\\}^{X\\times N_q}" + }, + { + "bbox": [ + 48, + 88, + 287, + 326 + ], + "type": "text", + "content": " Procedure: # Sparse Convolutional Backbone. 1 Extract multi-resolution feature maps " + }, + { + "bbox": [ + 48, + 88, + 287, + 326 + ], + "type": "inline_equation", + "content": "F_{i}" + }, + { + "bbox": [ + 48, + 88, + 287, + 326 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 48, + 88, + 287, + 326 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 48, + 88, + 287, + 326 + ], + "type": "text", + "content": " # Edge Prediction Module (EPM). 2 Predict edge map " + }, + { + "bbox": [ + 48, + 88, + 287, + 326 + ], + "type": "inline_equation", + "content": "\\hat{E}" + }, + { + "bbox": [ + 48, + 88, + 287, + 326 + ], + "type": "text", + "content": " using full-resolution feature " + }, + { + "bbox": [ + 48, + 88, + 287, + 326 + ], + "type": "inline_equation", + "content": "F_{0}" + }, + { + "bbox": [ + 48, + 88, + 287, + 326 + ], + "type": "text", + "content": " 3 Supervise " + }, + { + "bbox": [ + 48, + 88, + 287, + 326 + ], + "type": "inline_equation", + "content": "\\hat{E}" + }, + { + "bbox": [ + 48, + 88, + 287, + 326 + ], + "type": "text", + "content": " with pseudo edge label " + }, + { + "bbox": [ + 48, + 88, + 287, + 326 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 48, + 88, + 287, + 326 + ], + "type": "text", + "content": " 4 Encourage the network to learn edge-aware features. #Mask Transformer Decoder. 5 for " + }, + { + "bbox": [ + 48, + 88, + 287, + 326 + ], + "type": "inline_equation", + "content": "l\\in N_l" + }, + { + "bbox": [ + 48, + 88, + 287, + 326 + ], + "type": "text", + "content": " do 6 for " + }, + { + "bbox": [ + 48, + 88, + 287, + 326 + ], + "type": "inline_equation", + "content": "i\\in \\{1,2,3,4\\}" + }, + { + "bbox": [ + 48, + 88, + 287, + 326 + ], + "type": "text", + "content": " do 7 Update queries " + }, + { + "bbox": [ + 48, + 88, + 287, + 326 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 48, + 88, + 287, + 326 + ], + "type": "text", + "content": " using " + }, + { + "bbox": [ + 48, + 88, + 287, + 326 + ], + "type": "inline_equation", + "content": "M_{attn}" + }, + { + "bbox": [ + 48, + 88, + 287, + 326 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 48, + 88, + 287, + 326 + ], + "type": "inline_equation", + "content": "F_{i}" + }, + { + "bbox": [ + 48, + 88, + 287, + 326 + ], + "type": "text", + "content": " # Semantic Network. 8 Match the set of classes with text labels. 9 Encode text into text embeds " + }, + { + "bbox": [ + 48, + 88, + 287, + 326 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 48, + 88, + 287, + 326 + ], + "type": "text", + "content": " using CLIP. 10 Reinforce " + }, + { + "bbox": [ + 48, + 88, + 287, + 326 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 48, + 88, + 287, + 326 + ], + "type": "text", + "content": " to contain semantic clues of " + }, + { + "bbox": [ + 48, + 88, + 287, + 326 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 48, + 88, + 287, + 326 + ], + "type": "text", + "content": " 11 end 12 end" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "code_body" + } + ], + "index": 1, + "sub_type": "algorithm" + }, + { + "bbox": [ + 47, + 338, + 287, + 385 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 338, + 287, + 385 + ], + "spans": [ + { + "bbox": [ + 47, + 338, + 287, + 385 + ], + "type": "text", + "content": "compute the similarity between the projected query features " + }, + { + "bbox": [ + 47, + 338, + 287, + 385 + ], + "type": "inline_equation", + "content": "Q'" + }, + { + "bbox": [ + 47, + 338, + 287, + 385 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 338, + 287, + 385 + ], + "type": "inline_equation", + "content": "F_0" + }, + { + "bbox": [ + 47, + 338, + 287, + 385 + ], + "type": "text", + "content": " using the dot product operation, then calculate the probability of the instance mask employing the sigmoid function as follows." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 394, + 287, + 409 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 394, + 287, + 409 + ], + "spans": [ + { + "bbox": [ + 61, + 394, + 287, + 409 + ], + "type": "interline_equation", + "content": "M _ {a t t n} = \\left\\{m _ {i, j} = \\left[ \\sigma \\left(F _ {0} \\cdot f _ {\\text {q u e r y}} (Q) ^ {T}\\right) _ {i, j} > 0. 5 \\right] \\right\\} \\tag {2}", + "image_path": "86ebc75328c2e4c14927973ab0e65ab2f011de0bf96397bef34f0d790b6e7402.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 417, + 287, + 489 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 417, + 287, + 489 + ], + "spans": [ + { + "bbox": [ + 47, + 417, + 287, + 489 + ], + "type": "text", + "content": "where the threshold value is 0.5 for binary mask. Further, to refine the query representation, we leverage the " + }, + { + "bbox": [ + 47, + 417, + 287, + 489 + ], + "type": "inline_equation", + "content": "QR" + }, + { + "bbox": [ + 47, + 417, + 287, + 489 + ], + "type": "text", + "content": " block, including masked cross-attention and self-attention mechanisms. Here, we utilize " + }, + { + "bbox": [ + 47, + 417, + 287, + 489 + ], + "type": "inline_equation", + "content": "M_{att}" + }, + { + "bbox": [ + 47, + 417, + 287, + 489 + ], + "type": "text", + "content": " as the foreground mask for the masked cross-attention layer, where " + }, + { + "bbox": [ + 47, + 417, + 287, + 489 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 47, + 417, + 287, + 489 + ], + "type": "text", + "content": " iteratively attend to each multi-scale feature " + }, + { + "bbox": [ + 47, + 417, + 287, + 489 + ], + "type": "inline_equation", + "content": "(F_1 - F_4, i \\geqslant 1)" + }, + { + "bbox": [ + 47, + 417, + 287, + 489 + ], + "type": "text", + "content": " as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 88, + 497, + 287, + 512 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 88, + 497, + 287, + 512 + ], + "spans": [ + { + "bbox": [ + 88, + 497, + 287, + 512 + ], + "type": "interline_equation", + "content": "Q = \\operatorname {s o f t m a x} \\left(\\mathbf {Q K} ^ {T} / \\sqrt {D} + M _ {\\text {a t t n}}\\right) \\mathbf {V} \\tag {3}", + "image_path": "aec218cd72354abab6ef6a92224765a1db9e16f2506b55d1d889792f37ef457b.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 520, + 287, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 520, + 287, + 616 + ], + "spans": [ + { + "bbox": [ + 47, + 520, + 287, + 616 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 520, + 287, + 616 + ], + "type": "inline_equation", + "content": "\\mathbf{K}" + }, + { + "bbox": [ + 47, + 520, + 287, + 616 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 520, + 287, + 616 + ], + "type": "inline_equation", + "content": "\\mathbf{V}" + }, + { + "bbox": [ + 47, + 520, + 287, + 616 + ], + "type": "text", + "content": " are linearly projected from each point feature, and " + }, + { + "bbox": [ + 47, + 520, + 287, + 616 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}" + }, + { + "bbox": [ + 47, + 520, + 287, + 616 + ], + "type": "text", + "content": " are from " + }, + { + "bbox": [ + 47, + 520, + 287, + 616 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 47, + 520, + 287, + 616 + ], + "type": "text", + "content": ". Hence, the transformer focuses more on valuable instances instead of the unessential background. Subsequently, in the self-attention layer, the keys, values, and queries are all linear projections of " + }, + { + "bbox": [ + 47, + 520, + 287, + 616 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 47, + 520, + 287, + 616 + ], + "type": "text", + "content": " from the cross-attention layer. Finally, the transformer decoder layer is recurrently applied for multiple iterations " + }, + { + "bbox": [ + 47, + 520, + 287, + 616 + ], + "type": "inline_equation", + "content": "N_{l}" + }, + { + "bbox": [ + 47, + 520, + 287, + 616 + ], + "type": "text", + "content": " and feature scales " + }, + { + "bbox": [ + 47, + 520, + 287, + 616 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 47, + 520, + 287, + 616 + ], + "type": "text", + "content": ", ultimately producing the final set of refined queries." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 623, + 192, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 623, + 192, + 635 + ], + "spans": [ + { + "bbox": [ + 47, + 623, + 192, + 635 + ], + "type": "text", + "content": "3.2. Intelligent Semantic Prior" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 642, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 642, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 642, + 287, + 713 + ], + "type": "text", + "content": "It is important to highlight that precise geometric properties (e.g., locations, distances and orientations) from 3D points significantly enhance 3DIS accuracy. While previous studies [22, 35, 36] benefit from the generous geometric potential of physical information, they are still limited in perceiving intricate real-world 3D instances. Hence, we pro" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 310, + 71, + 543, + 137 + ], + "blocks": [ + { + "bbox": [ + 310, + 71, + 543, + 137 + ], + "lines": [ + { + "bbox": [ + 310, + 71, + 543, + 137 + ], + "spans": [ + { + "bbox": [ + 310, + 71, + 543, + 137 + ], + "type": "image", + "image_path": "b48eadeeee65a1d6a26e024a61b87e555937b855f6be5210dd54b4b0144e0d77.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 377, + 138, + 474, + 147 + ], + "lines": [ + { + "bbox": [ + 377, + 138, + 474, + 147 + ], + "spans": [ + { + "bbox": [ + 377, + 138, + 474, + 147 + ], + "type": "text", + "content": "(a) Transformer Decoder Layer" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 310, + 150, + 543, + 203 + ], + "blocks": [ + { + "bbox": [ + 310, + 150, + 543, + 203 + ], + "lines": [ + { + "bbox": [ + 310, + 150, + 543, + 203 + ], + "spans": [ + { + "bbox": [ + 310, + 150, + 543, + 203 + ], + "type": "image", + "image_path": "2a3295bc7d96b84dc109c8b3b80d0cc7cd2ddf67fc5a276e6ff747e0d144deee.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 391, + 205, + 460, + 213 + ], + "lines": [ + { + "bbox": [ + 391, + 205, + 460, + 213 + ], + "spans": [ + { + "bbox": [ + 391, + 205, + 460, + 213 + ], + "type": "text", + "content": "(b) Semantic Network" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 305, + 217, + 545, + 282 + ], + "lines": [ + { + "bbox": [ + 305, + 217, + 545, + 282 + ], + "spans": [ + { + "bbox": [ + 305, + 217, + 545, + 282 + ], + "type": "text", + "content": "Figure 3. Detailed architecture of (a) Transformer Decoder Layer: consists of Mask Module and Query Refinement blocks. And (b) Semantic Network: we first map the categories of queries " + }, + { + "bbox": [ + 305, + 217, + 545, + 282 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 305, + 217, + 545, + 282 + ], + "type": "text", + "content": " with the corresponding text description for text embeddings " + }, + { + "bbox": [ + 305, + 217, + 545, + 282 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 305, + 217, + 545, + 282 + ], + "type": "text", + "content": " from CLIP [32]. Then, we explicitly strengthen the " + }, + { + "bbox": [ + 305, + 217, + 545, + 282 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 305, + 217, + 545, + 282 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 305, + 217, + 545, + 282 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 305, + 217, + 545, + 282 + ], + "type": "text", + "content": " using the Semantic Fusion network to learn semantic details." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 292, + 545, + 316 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 292, + 545, + 316 + ], + "spans": [ + { + "bbox": [ + 305, + 292, + 545, + 316 + ], + "type": "text", + "content": "pose to utilize semantic knowledge of individual instances as smart priors, effectively addressing the challenges." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 318, + 545, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 318, + 545, + 544 + ], + "spans": [ + { + "bbox": [ + 305, + 318, + 545, + 544 + ], + "type": "text", + "content": "Deep Semantic Knowledge for 3D Instances. Recently, visual language models [17, 32] have highlighted the benefits of their capability to provide general embedded knowledge from large-scale image-text multimodal data. In this work, we explicitly instruct the transformer-based 3DIS network to learn contextual variations among instances by utilizing the power of the prevalent language model CLIP [32]. To achieve this goal, we utilize a carefully designed Semantic Network, as shown in Fig. 3 (b). Specifically, with the refined query " + }, + { + "bbox": [ + 305, + 318, + 545, + 544 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 305, + 318, + 545, + 544 + ], + "type": "text", + "content": " from each " + }, + { + "bbox": [ + 305, + 318, + 545, + 544 + ], + "type": "inline_equation", + "content": "QR" + }, + { + "bbox": [ + 305, + 318, + 545, + 544 + ], + "type": "text", + "content": " block, we first categorize each query following the linear classification head " + }, + { + "bbox": [ + 305, + 318, + 545, + 544 + ], + "type": "inline_equation", + "content": "f_{\\mathrm{class}}" + }, + { + "bbox": [ + 305, + 318, + 545, + 544 + ], + "type": "text", + "content": " in the MM. Given the set of category candidates, we map the category number " + }, + { + "bbox": [ + 305, + 318, + 545, + 544 + ], + "type": "inline_equation", + "content": "c_{i}" + }, + { + "bbox": [ + 305, + 318, + 545, + 544 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 305, + 318, + 545, + 544 + ], + "type": "inline_equation", + "content": "i = \\{1,2,\\dots,N_c\\}" + }, + { + "bbox": [ + 305, + 318, + 545, + 544 + ], + "type": "text", + "content": " with the corresponding text descriptions (e.g., \"Cabinet\", \"Bookshelf\"). Next, these text labels go through the large-scale pre-trained language model, which outputs text embedding vectors " + }, + { + "bbox": [ + 305, + 318, + 545, + 544 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 305, + 318, + 545, + 544 + ], + "type": "text", + "content": ". We then concatenate " + }, + { + "bbox": [ + 305, + 318, + 545, + 544 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 305, + 318, + 545, + 544 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 305, + 318, + 545, + 544 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 305, + 318, + 545, + 544 + ], + "type": "text", + "content": " and strengthen the basic instance queries with deep semantic clues from text embeddings using the Semantic Fusion network as follows:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 341, + 550, + 545, + 564 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 341, + 550, + 545, + 564 + ], + "spans": [ + { + "bbox": [ + 341, + 550, + 545, + 564 + ], + "type": "interline_equation", + "content": "Q = \\operatorname {S i g m o i d} \\left(\\mathbf {W} _ {2} \\cdot \\phi \\left(\\mathbf {W} _ {1} \\cdot (Q \\oplus T)\\right) ^ {T}\\right) \\tag {4}", + "image_path": "fece36c88489692a88f0cdd2584128bdb6d22fc83b44b482fbc1fd4e87a2d879.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 568, + 545, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 568, + 545, + 652 + ], + "spans": [ + { + "bbox": [ + 304, + 568, + 545, + 652 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 568, + 545, + 652 + ], + "type": "inline_equation", + "content": "\\oplus" + }, + { + "bbox": [ + 304, + 568, + 545, + 652 + ], + "type": "text", + "content": " indicates channel-wise concatenation, " + }, + { + "bbox": [ + 304, + 568, + 545, + 652 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_i" + }, + { + "bbox": [ + 304, + 568, + 545, + 652 + ], + "type": "text", + "content": " denotes the learnable parameters of the " + }, + { + "bbox": [ + 304, + 568, + 545, + 652 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 568, + 545, + 652 + ], + "type": "text", + "content": "-th linear layer, and " + }, + { + "bbox": [ + 304, + 568, + 545, + 652 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 304, + 568, + 545, + 652 + ], + "type": "text", + "content": " is the non-linear activation function. We apply this process for each iteration of the " + }, + { + "bbox": [ + 304, + 568, + 545, + 652 + ], + "type": "inline_equation", + "content": "QR" + }, + { + "bbox": [ + 304, + 568, + 545, + 652 + ], + "type": "text", + "content": ". Ultimately, we consider these explicitly guided queries with contextual variations as discriminative semantic priors for the practical perception of complex 3D instances, enhancing 3DIS accuracy." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 306, + 659, + 470, + 672 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 659, + 470, + 672 + ], + "spans": [ + { + "bbox": [ + 306, + 659, + 470, + 672 + ], + "type": "text", + "content": "3.3. Edge-Aware 3DIS Framework" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 305, + 677, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 677, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 677, + 545, + 713 + ], + "type": "text", + "content": "In real-world 3D scenarios, instances are commonly positioned in diverse arrangements without following standard rules or patterns. Especially when instances are closely ad" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20647" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 51, + 74, + 283, + 155 + ], + "blocks": [ + { + "bbox": [ + 51, + 74, + 283, + 155 + ], + "lines": [ + { + "bbox": [ + 51, + 74, + 283, + 155 + ], + "spans": [ + { + "bbox": [ + 51, + 74, + 283, + 155 + ], + "type": "image", + "image_path": "a2ea6740bf0d611246be912d5ba9e3a51ed2ef290a51c7db716348dd7eeb2998.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 166, + 287, + 222 + ], + "lines": [ + { + "bbox": [ + 46, + 166, + 287, + 222 + ], + "spans": [ + { + "bbox": [ + 46, + 166, + 287, + 222 + ], + "type": "text", + "content": "Figure 4. Pseudo edge label calculation using KNN algorithm. We collect " + }, + { + "bbox": [ + 46, + 166, + 287, + 222 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 166, + 287, + 222 + ], + "type": "text", + "content": " neighboring points of each point (red border) and compare their instance labels. If the count of different label points (blue and purple) among its surroundings surpasses a predefined threshold " + }, + { + "bbox": [ + 46, + 166, + 287, + 222 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 46, + 166, + 287, + 222 + ], + "type": "text", + "content": ", we annotate the central point as an edge point (green)." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 231, + 287, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 231, + 287, + 327 + ], + "spans": [ + { + "bbox": [ + 46, + 231, + 287, + 327 + ], + "type": "text", + "content": "jacent or overlapping, it becomes increasingly challenging for the model to define the edges precisely. Therefore, even with significant progress in technical strategies, most conventional methods [22, 35, 36] often misinterpret the spatial extent of each instance and ambiguously estimate their edges as illustrated in Fig. 1. To this end, we introduce an edge-aware 3DIS framework with an edge prediction module, enabling the network to utilize edge-advanced features." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 328, + 287, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 328, + 287, + 544 + ], + "spans": [ + { + "bbox": [ + 46, + 328, + 287, + 544 + ], + "type": "text", + "content": "Pseudo Edge Label Calculation. To optimize the edge prediction module, we first compute point-wise pseudo edge labels for all 3D point scenes. We operate a traditional " + }, + { + "bbox": [ + 46, + 328, + 287, + 544 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 328, + 287, + 544 + ], + "type": "text", + "content": "-Nearest Neighbors (KNN) classification algorithm based on the KD-tree data structure to explore the " + }, + { + "bbox": [ + 46, + 328, + 287, + 544 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 328, + 287, + 544 + ], + "type": "text", + "content": " nearest points. Here, we utilize point-wise instance labels from datasets. For all points in a 3D scene, we set each point as a central point and compare their instance labels with those of " + }, + { + "bbox": [ + 46, + 328, + 287, + 544 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 328, + 287, + 544 + ], + "type": "text", + "content": " neighbor points. As shown in Fig. 4, if the count of distinct label points among its surroundings exceeds a predefined threshold " + }, + { + "bbox": [ + 46, + 328, + 287, + 544 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 46, + 328, + 287, + 544 + ], + "type": "text", + "content": ", we identify the central point as a boundary point. We finally generate pseudo-binary edge maps " + }, + { + "bbox": [ + 46, + 328, + 287, + 544 + ], + "type": "inline_equation", + "content": "E' \\in \\{0,1\\}^{N_p \\times 1}" + }, + { + "bbox": [ + 46, + 328, + 287, + 544 + ], + "type": "text", + "content": ", where points on the edges are 1 and internal points are 0 for the entire set of points " + }, + { + "bbox": [ + 46, + 328, + 287, + 544 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 46, + 328, + 287, + 544 + ], + "type": "text", + "content": ". Then, we voxelize the point-wise edge labels into the voxel-wise edge labels " + }, + { + "bbox": [ + 46, + 328, + 287, + 544 + ], + "type": "inline_equation", + "content": "E \\in \\{0,1\\}^{X_0 \\times 1}" + }, + { + "bbox": [ + 46, + 328, + 287, + 544 + ], + "type": "text", + "content": " to supervise the edge extraction head. Note that we precompute the pseudo edge labels for every 3D scene in the datasets before training." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 545, + 287, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 545, + 287, + 666 + ], + "spans": [ + { + "bbox": [ + 47, + 545, + 287, + 666 + ], + "type": "text", + "content": "Learning Edge Cues. With dynamically generated pseudo edge labels " + }, + { + "bbox": [ + 47, + 545, + 287, + 666 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 47, + 545, + 287, + 666 + ], + "type": "text", + "content": ", we aim to enrich regular features to recognize the accurate spatial scope of 3D instances with precise boundaries. Specifically, we introduce an Edge Prediction Module, which predicts a voxel-wise dense edge map " + }, + { + "bbox": [ + 47, + 545, + 287, + 666 + ], + "type": "inline_equation", + "content": "\\hat{E} \\in (0,1)^{X_0 \\times 1}" + }, + { + "bbox": [ + 47, + 545, + 287, + 666 + ], + "type": "text", + "content": " to interactively guide the network in capturing edges of various instances. Given a full-resolution feature map " + }, + { + "bbox": [ + 47, + 545, + 287, + 666 + ], + "type": "inline_equation", + "content": "F_0" + }, + { + "bbox": [ + 47, + 545, + 287, + 666 + ], + "type": "text", + "content": " from the backbone network, our edge extraction head, which consists of multiple shared MLPs, estimates the probabilities of edges for each voxel as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 670, + 287, + 685 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 670, + 287, + 685 + ], + "spans": [ + { + "bbox": [ + 55, + 670, + 287, + 685 + ], + "type": "interline_equation", + "content": "\\hat {E} = \\operatorname {M L P} _ {\\text {e d g e}} \\left(F _ {0}\\right) = \\operatorname {S i g m o i d} \\left(\\mathbf {W} _ {2} \\cdot \\phi \\left(\\mathbf {W} _ {1} \\cdot F _ {0}\\right) ^ {T}\\right) \\tag {5}", + "image_path": "08de7b0ff867e24ff49a4502c630aa038fd376fd43bff8fb78aec927b54cb51e.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_i" + }, + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "content": " denotes the learnable parameters of the " + }, + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "content": "-th linear layer, and " + }, + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "content": " indicates the non-linear activation function." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 72, + 545, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 108 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 108 + ], + "type": "text", + "content": "Following [6, 10], as edge voxels constitute a small portion of the entire voxel set, we formulate weighted binary cross entropy loss to train the edge extraction head as follows:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 310, + 114, + 545, + 146 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 310, + 114, + 545, + 146 + ], + "spans": [ + { + "bbox": [ + 310, + 114, + 545, + 146 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {e d g e} = - \\sum_ {i = 1} ^ {X _ {0}} \\left(w \\cdot E _ {i} \\log \\hat {E} _ {i} + \\left(1 - E _ {i}\\right) \\log \\left(1 - \\hat {E} _ {i}\\right)\\right) (6)", + "image_path": "c00dbea97f083aaeb7cd6489f50e81186b415471828aed75101672cad0184324.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 149, + 545, + 233 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 149, + 545, + 233 + ], + "spans": [ + { + "bbox": [ + 305, + 149, + 545, + 233 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 149, + 545, + 233 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 305, + 149, + 545, + 233 + ], + "type": "text", + "content": " represent the weight value used to balance the substantial difference between the numbers of edge points and others. Then, we propagate the edge cues of prediction as auxiliary information for the backbone network, improving 3D instance understanding with edge-aware features. Furthermore, we train our edge network jointly in an end-to-end manner with low computational costs." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 236, + 545, + 271 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 236, + 545, + 271 + ], + "spans": [ + { + "bbox": [ + 305, + 236, + 545, + 271 + ], + "type": "text", + "content": "Loss Function We finally predict instance mask " + }, + { + "bbox": [ + 305, + 236, + 545, + 271 + ], + "type": "inline_equation", + "content": "\\hat{M}_f" + }, + { + "bbox": [ + 305, + 236, + 545, + 271 + ], + "type": "text", + "content": " using the final set of queries in the " + }, + { + "bbox": [ + 305, + 236, + 545, + 271 + ], + "type": "inline_equation", + "content": "MM" + }, + { + "bbox": [ + 305, + 236, + 545, + 271 + ], + "type": "text", + "content": ". For training, we compute the following loss " + }, + { + "bbox": [ + 305, + 236, + 545, + 271 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{mask}" + }, + { + "bbox": [ + 305, + 236, + 545, + 271 + ], + "type": "text", + "content": " as a sum of the two losses:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 311, + 279, + 545, + 294 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 279, + 545, + 294 + ], + "spans": [ + { + "bbox": [ + 311, + 279, + 545, + 294 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {m a s k}} = \\lambda_ {\\mathrm {B C E}} \\mathcal {L} _ {\\mathrm {B C E}} \\left(M _ {f}, \\hat {M} _ {f}\\right) + \\lambda_ {\\mathrm {d i c e}} \\mathcal {L} _ {\\mathrm {d i c e}} \\left(M _ {f}, \\hat {M} _ {f}\\right) \\tag {7}", + "image_path": "f26bc3d7c6c45174dd79405538397a7b1f8a43cd0f5c36d2348e64fe9bcdccd9.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 301, + 545, + 338 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 301, + 545, + 338 + ], + "spans": [ + { + "bbox": [ + 306, + 301, + 545, + 338 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 306, + 301, + 545, + 338 + ], + "type": "inline_equation", + "content": "M_{f}" + }, + { + "bbox": [ + 306, + 301, + 545, + 338 + ], + "type": "text", + "content": " denotes the ground-truth instance mask, and " + }, + { + "bbox": [ + 306, + 301, + 545, + 338 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\text{dice}}" + }, + { + "bbox": [ + 306, + 301, + 545, + 338 + ], + "type": "text", + "content": " represents the Dice loss [5]. Ultimately, our model is trained end-to-end by minimizing the following loss " + }, + { + "bbox": [ + 306, + 301, + 545, + 338 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\text{total}}" + }, + { + "bbox": [ + 306, + 301, + 545, + 338 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 321, + 349, + 545, + 363 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 349, + 545, + 363 + ], + "spans": [ + { + "bbox": [ + 321, + 349, + 545, + 363 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {t o t a l}} = \\lambda_ {\\text {m a s k}} \\mathcal {L} _ {\\text {m a s k}} + \\lambda_ {\\text {c l s}} \\mathcal {L} _ {\\text {c l s}} + \\lambda_ {\\text {e d g e}} \\mathcal {L} _ {\\text {e d g e}} \\tag {8}", + "image_path": "6b7e7dde3457b322d54b9ff2c2816f7327c161831a56860cc88898140cc3ec6e.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 370, + 545, + 394 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 370, + 545, + 394 + ], + "spans": [ + { + "bbox": [ + 305, + 370, + 545, + 394 + ], + "type": "text", + "content": "where each " + }, + { + "bbox": [ + 305, + 370, + 545, + 394 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 305, + 370, + 545, + 394 + ], + "type": "text", + "content": " is a hyperparameter derived from grid searches to handle the strength of respective loss term." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 404, + 387, + 418 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 404, + 387, + 418 + ], + "spans": [ + { + "bbox": [ + 306, + 404, + 387, + 418 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 424, + 422, + 437 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 424, + 422, + 437 + ], + "spans": [ + { + "bbox": [ + 306, + 424, + 422, + 437 + ], + "type": "text", + "content": "4.1. Experimental Setup" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 445, + 545, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 445, + 545, + 529 + ], + "spans": [ + { + "bbox": [ + 304, + 445, + 545, + 529 + ], + "type": "text", + "content": "Datasets. In this study, we train and evaluate the overall performance using four landmark datasets for 3D instance segmentation: ScanNetV2 [4], ScanNet200 [33], S3DIS [1], and STPLS3D [2]. These four datasets provide 3D point cloud scan data collected in various scenarios. Note that detailed descriptions of each dataset and all implementation details are provided in the supplemental material." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 530, + 545, + 638 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 530, + 545, + 638 + ], + "spans": [ + { + "bbox": [ + 304, + 530, + 545, + 638 + ], + "type": "text", + "content": "Evaluation Metrics. To evaluate the 3D instance segmentation accuracy, we operate Average Precision (AP), a typical metric for various computer vision tasks. We report mean average precision (mAP), which represents an average score with IoU thresholds ranging from " + }, + { + "bbox": [ + 304, + 530, + 545, + 638 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 304, + 530, + 545, + 638 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 304, + 530, + 545, + 638 + ], + "type": "inline_equation", + "content": "95\\%" + }, + { + "bbox": [ + 304, + 530, + 545, + 638 + ], + "type": "text", + "content": " with a step size of " + }, + { + "bbox": [ + 304, + 530, + 545, + 638 + ], + "type": "inline_equation", + "content": "5\\%" + }, + { + "bbox": [ + 304, + 530, + 545, + 638 + ], + "type": "text", + "content": ". Also, we provide " + }, + { + "bbox": [ + 304, + 530, + 545, + 638 + ], + "type": "inline_equation", + "content": "\\mathrm{mAP}_{50}" + }, + { + "bbox": [ + 304, + 530, + 545, + 638 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 530, + 545, + 638 + ], + "type": "inline_equation", + "content": "\\mathrm{mAP}_{25}" + }, + { + "bbox": [ + 304, + 530, + 545, + 638 + ], + "type": "text", + "content": ", indicating the scores with IoU thresholds of " + }, + { + "bbox": [ + 304, + 530, + 545, + 638 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 304, + 530, + 545, + 638 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 530, + 545, + 638 + ], + "type": "inline_equation", + "content": "25\\%" + }, + { + "bbox": [ + 304, + 530, + 545, + 638 + ], + "type": "text", + "content": ", respectively. Further, we estimate mean precision (mPrec) and mean recall (mRec) on the S3DIS [1] dataset." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 305, + 645, + 545, + 658 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 645, + 545, + 658 + ], + "spans": [ + { + "bbox": [ + 305, + 645, + 545, + 658 + ], + "type": "text", + "content": "4.2. Performance Comparison with SOTA Methods." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "content": "ScanNetV2. We quantitatively compare our proposed network EASE with existing state-of-the-art methods. These methods demonstrate significant capabilities, but they still rely on geometric information from 3D points and often" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20648" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 50, + 70, + 289, + 308 + ], + "blocks": [ + { + "bbox": [ + 50, + 70, + 289, + 308 + ], + "lines": [ + { + "bbox": [ + 50, + 70, + 289, + 308 + ], + "spans": [ + { + "bbox": [ + 50, + 70, + 289, + 308 + ], + "type": "table", + "html": "
MethodScanNet ValScanNet Test
mAP\\( mAP_{50} \\)mAP\\( mAP_{50} \\)
GSPN [44]19.337.8-30.6
3D-SIS [16]-18.716.138.2
MASC [25]--25.444.7
3D-Bonet [43]--25.348.8
MTML [20]20.340.228.254.9
3D-MPA [8]35.559.135.561.1
DyCo3D [15]35.457.639.564.1
PointGroup [18]34.856.740.763.6
MaskGroup [45]42.063.343.466.4
OccuSeg [13]44.260.748.667.2
SSTNet [24]49.464.350.669.8
HAIS [3]43.564.145.769.9
SoftGroup [39]46.067.650.476.1
Mask3D [35]55.273.756.678.0
QueryFormer [26]56.574.258.378.7
MAFT [22]59.976.559.678.6
EASE (Ours)60.277.259.878.9
", + "image_path": "a02187e0001d23e7c2a90fdadb723b34920123c26a6ede3b2a14682fccfb8ca4.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 46, + 312, + 288, + 357 + ], + "lines": [ + { + "bbox": [ + 46, + 312, + 288, + 357 + ], + "spans": [ + { + "bbox": [ + 46, + 312, + 288, + 357 + ], + "type": "text", + "content": "Table 1. Comparison of 3D Instance Segmentation performance with state-of-the-art approaches on the ScanNetV2 [4] dataset. We evaluate mean average precision (mAP) with different IoU thresholds over 18 classes on the ScanNetV2 validation / hidden test set." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 366, + 287, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 366, + 287, + 475 + ], + "spans": [ + { + "bbox": [ + 46, + 366, + 287, + 475 + ], + "type": "text", + "content": "face challenges in effectively comprehending 3D instances. To address these limitations, we present a novel edge-aware framework featuring a semantic guidance network. As reported in Tab. 1, EASE generally outperforms other methods, achieving new state-of-the-art accuracy in terms of " + }, + { + "bbox": [ + 46, + 366, + 287, + 475 + ], + "type": "inline_equation", + "content": "\\mathrm{mAP}" + }, + { + "bbox": [ + 46, + 366, + 287, + 475 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 366, + 287, + 475 + ], + "type": "inline_equation", + "content": "\\mathrm{mAP}_{50}" + }, + { + "bbox": [ + 46, + 366, + 287, + 475 + ], + "type": "text", + "content": " on both validation (60.2 / 77.2) and hidden test (59.8 / 78.9) sets. These results confirm that our edge cues and smart semantic prior benefits the entire network, leading to high precision in 3D instance segmentation." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 477, + 287, + 575 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 477, + 287, + 575 + ], + "spans": [ + { + "bbox": [ + 46, + 477, + 287, + 575 + ], + "type": "text", + "content": "S3DIS. In Tab. 2, we evaluate 3DIS performance on Area 5 and 6-fold cross-validation of the S3DIS [1] dataset. For Area 5 evaluation, we employ data from Area 5 for validation and the other areas for training. Also, for 6-fold cross-validation, we assess validation scores across six other areas and compute the average. In both evaluations, EASE demonstrates significant performance improvements up to " + }, + { + "bbox": [ + 46, + 477, + 287, + 575 + ], + "type": "inline_equation", + "content": "+2.4 / 2.5" + }, + { + "bbox": [ + 46, + 477, + 287, + 575 + ], + "type": "text", + "content": " (Area 5) and " + }, + { + "bbox": [ + 46, + 477, + 287, + 575 + ], + "type": "inline_equation", + "content": "+0.8 / 0.9" + }, + { + "bbox": [ + 46, + 477, + 287, + 575 + ], + "type": "text", + "content": " (6-fold) for " + }, + { + "bbox": [ + 46, + 477, + 287, + 575 + ], + "type": "inline_equation", + "content": "\\mathrm{mAP} / \\mathrm{mAP}_{50}" + }, + { + "bbox": [ + 46, + 477, + 287, + 575 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 578, + 287, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 578, + 287, + 628 + ], + "spans": [ + { + "bbox": [ + 46, + 578, + 287, + 628 + ], + "type": "text", + "content": "ScanNet200. We also demonstrate considerable performance on the ScanNet200 validation set (see Tab. 3). Our method EASE precisely segments 3D instances for various categories compared to current state-of-the-art approaches." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "type": "text", + "content": "STPLS3D. Remarkably, our EASE network also achieves higher scores on the STPLS3D [2] dataset, which consists of outdoor 3D point cloud scenes. EASE exceeds the second-best results with " + }, + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "type": "inline_equation", + "content": "+1.1 / 1.6 / 1.3" + }, + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "type": "text", + "content": " margins in " + }, + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\mathrm{mAP} / \\mathrm{mAP}_{50} / \\mathrm{mAP}_{25}" + }, + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "type": "text", + "content": ". These results highlight the effectiveness of our method in understanding diverse real-world 3D instances, both from indoor and outdoor environments." + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 305, + 70, + 555, + 247 + ], + "blocks": [ + { + "bbox": [ + 305, + 70, + 555, + 247 + ], + "lines": [ + { + "bbox": [ + 305, + 70, + 555, + 247 + ], + "spans": [ + { + "bbox": [ + 305, + 70, + 555, + 247 + ], + "type": "table", + "html": "
MethodS3DIS Area 5S3DIS 6-fold CV
AP\\( AP_{50} \\)\\( Prec_{50} \\)\\( Rec_{50} \\)AP\\( AP_{50} \\)\\( Prec_{50} \\)\\( Rec_{50} \\)
ASIS [41]--55.342.4--63.647.5
3D-Bonet [43]--57.540.2--65.647.6
3D-MPA [8]--63.158.0--66.764.1
PointGroup [18]-57.861.962.1-64.069.669.2
DyCo3D [15]--64.364.2----
MaskGroup [45]-65.062.964.7-69.966.669.6
SSTNet [24]42.759.365.564.254.167.873.573.4
SoftGroup [39]51.666.173.666.654.468.975.369.8
Mask3D [35]56.668.468.766.364.575.572.874.5
MAFT [22]-69.1------
QueryFormer [26]57.769.970.572.262.073.372.773.4
EASE (Ours)59.071.669.468.765.376.473.674.6
", + "image_path": "85e311bd0fd4af619c41fba4c9a12321289aa18ae7cce901f07dfd848e2a70aa.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 305, + 251, + 547, + 308 + ], + "lines": [ + { + "bbox": [ + 305, + 251, + 547, + 308 + ], + "spans": [ + { + "bbox": [ + 305, + 251, + 547, + 308 + ], + "type": "text", + "content": "Table 2. Comparison of 3D Instance Segmentation performance with state-of-the-art approaches on the S3DIS [1] Area 5 and 6-fold cross validation set. We evaluate mAP across different IoU thresholds, along with mean precision (mPrec) and mean recall (mRec) at a " + }, + { + "bbox": [ + 305, + 251, + 547, + 308 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 305, + 251, + 547, + 308 + ], + "type": "text", + "content": " IoU threshold over 13 classes of the S3DIS." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 306, + 316, + 553, + 402 + ], + "blocks": [ + { + "bbox": [ + 306, + 316, + 553, + 402 + ], + "lines": [ + { + "bbox": [ + 306, + 316, + 553, + 402 + ], + "spans": [ + { + "bbox": [ + 306, + 316, + 553, + 402 + ], + "type": "table", + "html": "
MethodmAP\\( \\mathrm{mAP}_{50} \\)\\( \\mathrm{mAP}_{25} \\)
SPFormer [36]25.233.839.6
Mask3D [35]27.437.042.3
QueryFormer [26]28.137.143.4
MAFT [22]29.238.243.3
EASE (Ours)29.938.844.7
", + "image_path": "781e596597928f1eb5777e44a85425c20be20691f8491d3ae7a3c3d73c0f83d7.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 305, + 406, + 545, + 430 + ], + "lines": [ + { + "bbox": [ + 305, + 406, + 545, + 430 + ], + "spans": [ + { + "bbox": [ + 305, + 406, + 545, + 430 + ], + "type": "text", + "content": "Table 3. Comparison of 3D Instance Segmentation performance with state-of-the-art approaches on the ScanNet200 [33] val set." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 306, + 439, + 553, + 524 + ], + "blocks": [ + { + "bbox": [ + 306, + 439, + 553, + 524 + ], + "lines": [ + { + "bbox": [ + 306, + 439, + 553, + 524 + ], + "spans": [ + { + "bbox": [ + 306, + 439, + 553, + 524 + ], + "type": "table", + "html": "
MethodmAP\\( \\mathrm{mAP}_{50} \\)\\( \\mathrm{mAP}_{25} \\)
PointGroup [18]23.338.548.6
HAIS [3]35.146.752.8
SoftGroup [39]47.363.171.4
Mask3D [35]63.479.285.6
EASE (Ours)64.580.886.9
", + "image_path": "f21c2fc81c8bb4dce875e3d5767b8d7fbf7dcc1745282fbd7eaac6bdcb474b6d.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 305, + 528, + 545, + 551 + ], + "lines": [ + { + "bbox": [ + 305, + 528, + 545, + 551 + ], + "spans": [ + { + "bbox": [ + 305, + 528, + 545, + 551 + ], + "type": "text", + "content": "Table 4. Comparison of 3D Instance Segmentation performance with state-of-the-art approaches on the STPLS3D [2] test dataset." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 559, + 406, + 571 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 559, + 406, + 571 + ], + "spans": [ + { + "bbox": [ + 306, + 559, + 406, + 571 + ], + "type": "text", + "content": "4.3. Ablation Studies" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 582, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 582, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 582, + 547, + 715 + ], + "type": "text", + "content": "Effect of Semantic and Edge Modules. In Tab. 5, we evaluate the variants of our method w/ and w/o the Semantic Network (SN) and Edge Prediction Module (EPM). The addition of two modules improves 3DIS accuracy across all experiments. The SN especially underscores the significance of incorporating rich semantic knowledge for effective segmentation, achieving " + }, + { + "bbox": [ + 304, + 582, + 547, + 715 + ], + "type": "inline_equation", + "content": "+1.3 / 0.7" + }, + { + "bbox": [ + 304, + 582, + 547, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 582, + 547, + 715 + ], + "type": "inline_equation", + "content": "+1.5 / 1.7" + }, + { + "bbox": [ + 304, + 582, + 547, + 715 + ], + "type": "text", + "content": " improvements on " + }, + { + "bbox": [ + 304, + 582, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\mathrm{mAP} / \\mathrm{mAP}_{50}" + }, + { + "bbox": [ + 304, + 582, + 547, + 715 + ], + "type": "text", + "content": " for ScanNetV2 and S3DIS, respectively. EPM further enhances performance via the auxiliary edge information for edge-aware features, resulting in gains of " + }, + { + "bbox": [ + 304, + 582, + 547, + 715 + ], + "type": "inline_equation", + "content": "+0.9 / 0.8" + }, + { + "bbox": [ + 304, + 582, + 547, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 582, + 547, + 715 + ], + "type": "inline_equation", + "content": "+1.9 / 0.9" + }, + { + "bbox": [ + 304, + 582, + 547, + 715 + ], + "type": "text", + "content": ". Notably, integrating" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "20649" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 48, + 66, + 289, + 178 + ], + "blocks": [ + { + "bbox": [ + 48, + 66, + 289, + 178 + ], + "lines": [ + { + "bbox": [ + 48, + 66, + 289, + 178 + ], + "spans": [ + { + "bbox": [ + 48, + 66, + 289, + 178 + ], + "type": "table", + "html": "
BaselineMethodScanNet ValS3DIS Area 5
SemanticEdgemAP / mAP50mAP / mAP50
✓--58.4 / 75.956.6 / 68.4
✓✓-59.7 / 76.658.1 / 70.1
✓-✓59.3 / 76.758.5 / 69.3
✓✓✓60.2 / 77.259.0 / 71.6
", + "image_path": "86575ad122e42ead3f8b3f697102e1b7cafc8ecf87130a556755a9b673a99fdc.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 48, + 208, + 289, + 330 + ], + "blocks": [ + { + "bbox": [ + 47, + 182, + 287, + 205 + ], + "lines": [ + { + "bbox": [ + 47, + 182, + 287, + 205 + ], + "spans": [ + { + "bbox": [ + 47, + 182, + 287, + 205 + ], + "type": "text", + "content": "Table 5. Ablation study to see the effect of our proposed two main modules: Semantic Network and Edge Prediction Module." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 48, + 208, + 289, + 330 + ], + "lines": [ + { + "bbox": [ + 48, + 208, + 289, + 330 + ], + "spans": [ + { + "bbox": [ + 48, + 208, + 289, + 330 + ], + "type": "table", + "html": "
UsageMethodScanNet ValS3DIS Area 5
E-wise SumConcatmAP / mAP50mAP / mAP50
---58.4 / 75.956.6 / 68.4
fcls✓-58.9 / 76.057.8 / 68.5
fcls-✓59.3 / 76.558.3 / 70.1
S·F✓-59.6 / 76.958.6 / 69.6
S·F-✓60.2 / 77.259.0 / 71.6
", + "image_path": "d7b10d50217883d356a7da33a646e119547440b7f34c4f9a23fc6b46c84354bd.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 374, + 287, + 398 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 374, + 287, + 398 + ], + "spans": [ + { + "bbox": [ + 46, + 374, + 287, + 398 + ], + "type": "text", + "content": "SN and EPM leads to considerable advances " + }, + { + "bbox": [ + 46, + 374, + 287, + 398 + ], + "type": "inline_equation", + "content": "(+1.8 / 1.3" + }, + { + "bbox": [ + 46, + 374, + 287, + 398 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 374, + 287, + 398 + ], + "type": "inline_equation", + "content": "+2.4 / 3.2)" + }, + { + "bbox": [ + 46, + 374, + 287, + 398 + ], + "type": "text", + "content": ", confirming the effectiveness of each module." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 45, + 400, + 288, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 400, + 288, + 581 + ], + "spans": [ + { + "bbox": [ + 45, + 400, + 288, + 581 + ], + "type": "text", + "content": "Applications of Text Embeddings. In Tab. 6, we analyze various applications of text embeddings to determine the most effective strategy for the functional understanding of complex 3D instances. We utilize text embeddings for category classification from " + }, + { + "bbox": [ + 45, + 400, + 288, + 581 + ], + "type": "inline_equation", + "content": "f_{cls}" + }, + { + "bbox": [ + 45, + 400, + 288, + 581 + ], + "type": "text", + "content": " of the Mask Module to boost classification accuracy. Also, we strengthen basic queries using the Semantic Fusion (" + }, + { + "bbox": [ + 45, + 400, + 288, + 581 + ], + "type": "inline_equation", + "content": "S \\cdot F" + }, + { + "bbox": [ + 45, + 400, + 288, + 581 + ], + "type": "text", + "content": ") network as described in Sec. 3.2. For experiments with the E-wise Sum, we map the text embeddings to the same feature space as the basic queries through MLPs and sum together. While incorporating semantic cues via element-wise sum enhances 3DIS accuracy, it exhibits limited improvements compared to concatenation. By comparison, applying the semantic knowledge across the entire transformer decoder through the " + }, + { + "bbox": [ + 45, + 400, + 288, + 581 + ], + "type": "inline_equation", + "content": "S \\cdot F" + }, + { + "bbox": [ + 45, + 400, + 288, + 581 + ], + "type": "text", + "content": " outperforms using them only for the classification head." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 582, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 582, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 582, + 288, + 714 + ], + "type": "text", + "content": "Weight Value of Edge Prediction Loss. To encourage the network to learn practical edge cues, we utilize an edge prediction network. In particular, we compute the weighted binary cross-entropy loss to effectively train the edge extraction head from sparse edge points. Here, we explore the impact of the coefficient value " + }, + { + "bbox": [ + 46, + 582, + 288, + 714 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 46, + 582, + 288, + 714 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 46, + 582, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{edge}" + }, + { + "bbox": [ + 46, + 582, + 288, + 714 + ], + "type": "text", + "content": " on 3DIS accuracy. First, as shown in Fig. 5, we empirically find that " + }, + { + "bbox": [ + 46, + 582, + 288, + 714 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 46, + 582, + 288, + 714 + ], + "type": "text", + "content": " influences the thickness of edges in a 3D scene. In Tab. 7, for the ScanNet [4] consisting of relatively small-scale scenes, a smaller value of " + }, + { + "bbox": [ + 46, + 582, + 288, + 714 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 46, + 582, + 288, + 714 + ], + "type": "text", + "content": " (e.g. 2), resulting in thin edges, leads to higher performance. However, for the vast-scale scenes of" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 309, + 55, + 386, + 125 + ], + "blocks": [ + { + "bbox": [ + 309, + 55, + 386, + 125 + ], + "lines": [ + { + "bbox": [ + 309, + 55, + 386, + 125 + ], + "spans": [ + { + "bbox": [ + 309, + 55, + 386, + 125 + ], + "type": "image", + "image_path": "67150f2bb3f008f5f4080bc18427ccd6e3e230d0fac2c9e55341405fba91698a.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 306, + 129, + 545, + 162 + ], + "lines": [ + { + "bbox": [ + 306, + 129, + 545, + 162 + ], + "spans": [ + { + "bbox": [ + 306, + 129, + 545, + 162 + ], + "type": "text", + "content": "Figure 5. Visualization of edge prediction results (red) for different weight values " + }, + { + "bbox": [ + 306, + 129, + 545, + 162 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 306, + 129, + 545, + 162 + ], + "type": "text", + "content": " in the weighted binary cross-entropy loss " + }, + { + "bbox": [ + 306, + 129, + 545, + 162 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{edge}" + }, + { + "bbox": [ + 306, + 129, + 545, + 162 + ], + "type": "text", + "content": ", along with corresponding pseudo-edge label (green)." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 386, + 55, + 465, + 124 + ], + "blocks": [ + { + "bbox": [ + 386, + 55, + 465, + 124 + ], + "lines": [ + { + "bbox": [ + 386, + 55, + 465, + 124 + ], + "spans": [ + { + "bbox": [ + 386, + 55, + 465, + 124 + ], + "type": "image", + "image_path": "eadb8bdd0196298d2b073c8324cf301ce7f4c1d51468330c29f648cb5dd1fd1e.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 466, + 56, + 544, + 124 + ], + "blocks": [ + { + "bbox": [ + 466, + 56, + 544, + 124 + ], + "lines": [ + { + "bbox": [ + 466, + 56, + 544, + 124 + ], + "spans": [ + { + "bbox": [ + 466, + 56, + 544, + 124 + ], + "type": "image", + "image_path": "afe05085399c88d9dc529810348cf903a93784fd1484e1bb2bbf4ddc574de952.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 306, + 168, + 547, + 326 + ], + "blocks": [ + { + "bbox": [ + 46, + 334, + 288, + 367 + ], + "lines": [ + { + "bbox": [ + 46, + 334, + 288, + 367 + ], + "spans": [ + { + "bbox": [ + 46, + 334, + 288, + 367 + ], + "type": "text", + "content": "Table 6. Ablation study of leveraging text embeddings to provide semantic details. " + }, + { + "bbox": [ + 46, + 334, + 288, + 367 + ], + "type": "inline_equation", + "content": "f_{cls}" + }, + { + "bbox": [ + 46, + 334, + 288, + 367 + ], + "type": "text", + "content": " represents the classification head in the Mask Module, and " + }, + { + "bbox": [ + 46, + 334, + 288, + 367 + ], + "type": "inline_equation", + "content": "S \\cdot F" + }, + { + "bbox": [ + 46, + 334, + 288, + 367 + ], + "type": "text", + "content": " stands for the Semantic Fusion network." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 306, + 168, + 547, + 326 + ], + "lines": [ + { + "bbox": [ + 306, + 168, + 547, + 326 + ], + "spans": [ + { + "bbox": [ + 306, + 168, + 547, + 326 + ], + "type": "table", + "html": "
Feature mapw in LedgeScanNet ValS3DIS Area 5
mAP / mAP50mAP / mAP50
F0259.3 / 76.756.9 / 68.6
F4259.0 / 76.156.7 / 68.5
F0658.9 / 76.358.4 / 68.9
F4658.5 / 76.058.2 / 68.6
F01058.4 / 75.858.5 / 69.3
F41058.5 / 75.758.2 / 69.0
", + "image_path": "74ddd89c04e9e46c1a051ee644a69742b027e1fd0d6f85c09a3f9685270472bb.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 329, + 546, + 364 + ], + "lines": [ + { + "bbox": [ + 306, + 329, + 546, + 364 + ], + "spans": [ + { + "bbox": [ + 306, + 329, + 546, + 364 + ], + "type": "text", + "content": "Table 7. Ablation study to investigate the impact of varying weight values " + }, + { + "bbox": [ + 306, + 329, + 546, + 364 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 306, + 329, + 546, + 364 + ], + "type": "text", + "content": " in the weighted binary cross-entropy " + }, + { + "bbox": [ + 306, + 329, + 546, + 364 + ], + "type": "inline_equation", + "content": "{\\mathcal{L}}_{\\text{edge }}" + }, + { + "bbox": [ + 306, + 329, + 546, + 364 + ], + "type": "text", + "content": " ,using high " + }, + { + "bbox": [ + 306, + 329, + 546, + 364 + ], + "type": "inline_equation", + "content": "\\left( {F}_{0}\\right)" + }, + { + "bbox": [ + 306, + 329, + 546, + 364 + ], + "type": "text", + "content": " and low " + }, + { + "bbox": [ + 306, + 329, + 546, + 364 + ], + "type": "inline_equation", + "content": "\\left( {F}_{4}\\right)" + }, + { + "bbox": [ + 306, + 329, + 546, + 364 + ], + "type": "text", + "content": " resolution feature maps for edge prediction." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 305, + 371, + 545, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 371, + 545, + 456 + ], + "spans": [ + { + "bbox": [ + 305, + 371, + 545, + 456 + ], + "type": "text", + "content": "the S3DIS [1], a higher value of " + }, + { + "bbox": [ + 305, + 371, + 545, + 456 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 305, + 371, + 545, + 456 + ], + "type": "text", + "content": " (e.g. 10) leads to better performance with thicker edges. We finally observe the correlation between the scene scale and the weight value " + }, + { + "bbox": [ + 305, + 371, + 545, + 456 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 305, + 371, + 545, + 456 + ], + "type": "text", + "content": "; the optimal value for edge learning is proportional to the scale. Besides, methods using high-resolution feature maps (e.g. " + }, + { + "bbox": [ + 305, + 371, + 545, + 456 + ], + "type": "inline_equation", + "content": "F_0" + }, + { + "bbox": [ + 305, + 371, + 545, + 456 + ], + "type": "text", + "content": ") for edge prediction generally perform better than those using low-resolution features maps (e.g. " + }, + { + "bbox": [ + 305, + 371, + 545, + 456 + ], + "type": "inline_equation", + "content": "F_4" + }, + { + "bbox": [ + 305, + 371, + 545, + 456 + ], + "type": "text", + "content": ")." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 464, + 425, + 477 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 464, + 425, + 477 + ], + "spans": [ + { + "bbox": [ + 306, + 464, + 425, + 477 + ], + "type": "text", + "content": "4.4. Qualitative Analyses" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 483, + 545, + 675 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 483, + 545, + 675 + ], + "spans": [ + { + "bbox": [ + 304, + 483, + 545, + 675 + ], + "type": "text", + "content": "Visual Comparison. In this section, we qualitatively confirm the usefulness of our novel framework EASE. We visualize the predicted semantic and instance masks of the baseline model [35] and ours on the ScanNetV2 [4] validation set, in Fig. 6. We emphasize the key differences using green and yellow boxes. As shown in the semantic results (Sem.), ours accurately classifies instances with similar appearances (e.g., desk and bed, door and cabinet) instead of the baseline, which incorrectly recognizes them. Also, as illustrated in the instance mask results (Inst.) of Scene1 and Scene2, ours segments a single object as one, unlike the baseline, which splits it into multiple parts. Further, in Scene 3 (Inst.), ours correctly distinguishes multiple objects (door and cabinet) in contrast to the baseline model. These qualitative results demonstrate that our method provides more detailed and robust segmentation results than the baseline method." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 305, + 677, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 677, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 677, + 545, + 713 + ], + "type": "text", + "content": "Impact of Utilizing Semantic Priors. To verify the effectiveness of the Semantic Network, we present t-SNE [38] visualizations of query features for instances with similar" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "20650" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 49, + 53, + 545, + 245 + ], + "blocks": [ + { + "bbox": [ + 49, + 53, + 545, + 245 + ], + "lines": [ + { + "bbox": [ + 49, + 53, + 545, + 245 + ], + "spans": [ + { + "bbox": [ + 49, + 53, + 545, + 245 + ], + "type": "image", + "image_path": "50b6fd6cbc393982acb1f116a4e3f476ecbae0c108ec67106571f8c33707713f.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 254, + 547, + 289 + ], + "lines": [ + { + "bbox": [ + 46, + 254, + 547, + 289 + ], + "spans": [ + { + "bbox": [ + 46, + 254, + 547, + 289 + ], + "type": "text", + "content": "Figure 6. Qualitative comparison of 3D Instance Segmentation performance on the ScanNetV2 [4] validation set. We visualize semantic (Sem.) and instance (Inst.) masks of the baseline model and ours with Ground Truth (GT) masks. The critical differences are emphasized using green and yellow-colored boxes. Note that the color map (top right) represents semantic labels. Best viewed in color." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 50, + 297, + 280, + 407 + ], + "blocks": [ + { + "bbox": [ + 50, + 297, + 280, + 407 + ], + "lines": [ + { + "bbox": [ + 50, + 297, + 280, + 407 + ], + "spans": [ + { + "bbox": [ + 50, + 297, + 280, + 407 + ], + "type": "image", + "image_path": "27c0198d788fc79971e1ea3d9b0e6de9c8ae2fbd834829a37b9ab2a8933846b7.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 415, + 288, + 448 + ], + "lines": [ + { + "bbox": [ + 46, + 415, + 288, + 448 + ], + "spans": [ + { + "bbox": [ + 46, + 415, + 288, + 448 + ], + "type": "text", + "content": "Figure 7. Category probability map (spatial distribution) for queries with high confidence. The x-axis denotes queries, the y-axis represents categories, and the z-axis (right) is probabilities." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 457, + 287, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 457, + 287, + 613 + ], + "spans": [ + { + "bbox": [ + 46, + 457, + 287, + 613 + ], + "type": "text", + "content": "appearances, which pose challenges for 3DIS models. As shown in Fig. 8, the baseline (MAFT [22]) model exhibits disorganized clusters, whereas ours shows relatively distinct feature space for challenging instances. Further, in Fig. 7, we visualize the spatial distribution of category probability for instance queries. The baseline tends to confuse perquery categories and predicts multiple categories for each query with low probability, causing semantic misclassifications. However, our model EASE overcomes this problem using semantic priors, providing clearer insight. These qualitative findings confirm that our semantic network effectively prompts the network to learn instance-specific semantic knowledge from intelligent text embedding priors." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 621, + 119, + 633 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 621, + 119, + 633 + ], + "spans": [ + { + "bbox": [ + 47, + 621, + 119, + 633 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 642, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 642, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 642, + 288, + 713 + ], + "type": "text", + "content": "Current 3D instance segmentation approaches often face challenges in understanding real-world 3D instances relying solely on geometrical information. Specifically, (i) these methods suffer from identifying instances with similar appearances, and (ii) they often misinterpret the spatial extent of instances, resulting in unclear edges. To tackle these" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 309, + 298, + 425, + 381 + ], + "blocks": [ + { + "bbox": [ + 309, + 298, + 425, + 381 + ], + "lines": [ + { + "bbox": [ + 309, + 298, + 425, + 381 + ], + "spans": [ + { + "bbox": [ + 309, + 298, + 425, + 381 + ], + "type": "image", + "image_path": "35c883f6ce8d111bed3542168efd7a733723329ce7901c8de1571996464441b6.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 387, + 545, + 420 + ], + "lines": [ + { + "bbox": [ + 305, + 387, + 545, + 420 + ], + "spans": [ + { + "bbox": [ + 305, + 387, + 545, + 420 + ], + "type": "text", + "content": "Figure 8. t-SNE [38] visualization of query features representing each instance. Compared to the baseline [22], which produces disorganized clusters, ours creates more distinct clusters." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 425, + 298, + 543, + 381 + ], + "blocks": [ + { + "bbox": [ + 425, + 298, + 543, + 381 + ], + "lines": [ + { + "bbox": [ + 425, + 298, + 543, + 381 + ], + "spans": [ + { + "bbox": [ + 425, + 298, + 543, + 381 + ], + "type": "image", + "image_path": "91836c3bd17e74dc6b126d22e7f4d844031d3a862a617c6aa1c0847345d5a6d2.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 430, + 545, + 513 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 430, + 545, + 513 + ], + "spans": [ + { + "bbox": [ + 304, + 430, + 545, + 513 + ], + "type": "text", + "content": "challenges, EASE, our novel framework, focuses on context details of text embeddings from the language model as smart priors, enhancing practical semantic understanding. Also, we employ the edge prediction module, guiding the network to reduce misclassifications near edges with edge-aware features. Our extensive experiments verify the effectiveness of EASE, achieving new state-of-the-art scores." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 515, + 546, + 706 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 515, + 546, + 706 + ], + "spans": [ + { + "bbox": [ + 304, + 515, + 546, + 706 + ], + "type": "text", + "content": "Acknowledgments This work was supported by Culture, Sports and Tourism R&D Program through the Korea Creative Content Agency grant funded by the Ministry of Culture, Sports and Tourism in 2023. (RS-2023-00227409, Development of sketch-based semantic 3D modeling technology for creating user-centric Metaverse content spaces for indoor spaces, " + }, + { + "bbox": [ + 304, + 515, + 546, + 706 + ], + "type": "inline_equation", + "content": "80\\%" + }, + { + "bbox": [ + 304, + 515, + 546, + 706 + ], + "type": "text", + "content": "), (R2022020068, 4D Content Generation and Copyright Protection with Artificial Intelligence, " + }, + { + "bbox": [ + 304, + 515, + 546, + 706 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 304, + 515, + 546, + 706 + ], + "type": "text", + "content": "), Electronics and Telecommunications Research Institute (ETRI) grant funded by the Korean government. (24ZC1200, Research on hyper-realistic interaction technology for five senses and emotional experience, " + }, + { + "bbox": [ + 304, + 515, + 546, + 706 + ], + "type": "inline_equation", + "content": "9\\%" + }, + { + "bbox": [ + 304, + 515, + 546, + 706 + ], + "type": "text", + "content": "), and the Institute of Information & communications Technology Planning & Evaluation (IITP) grant funded by the Korea government (MSIT) (2019-0-00079, Artificial Intelligence Graduate School Program (Korea University), " + }, + { + "bbox": [ + 304, + 515, + 546, + 706 + ], + "type": "inline_equation", + "content": "1\\%" + }, + { + "bbox": [ + 304, + 515, + 546, + 706 + ], + "type": "text", + "content": ")." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20651" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 289, + 715 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 53, + 91, + 289, + 146 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 289, + 146 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 289, + 146 + ], + "type": "text", + "content": "[1] Iro Armeni, Ozan Sener, Amir R Zamir, Helen Jiang, Ioannis Brilakis, Martin Fischer, and Silvio Savarese. 3d semantic parsing of large-scale indoor spaces. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1534-1543, 2016. 2, 5, 6, 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 147, + 288, + 201 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 147, + 288, + 201 + ], + "spans": [ + { + "bbox": [ + 53, + 147, + 288, + 201 + ], + "type": "text", + "content": "[2] Meida Chen, Qingyong Hu, Zifan Yu, Hugues Thomas, Andrew Feng, Yu Hou, Kyle McCullough, Fengbo Ren, and Lucio Soibelman. Stpls3d: A large-scale synthetic and real aerial photogrammetry 3d point cloud dataset. arXiv preprint arXiv:2203.09065, 2022. 2, 5, 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 202, + 288, + 256 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 202, + 288, + 256 + ], + "spans": [ + { + "bbox": [ + 53, + 202, + 288, + 256 + ], + "type": "text", + "content": "[3] Shaoyu Chen, Jiemin Fang, Qian Zhang, Wenyu Liu, and Xinggang Wang. Hierarchical aggregation for 3d instance segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15467-15476, 2021. 1, 2, 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 258, + 288, + 312 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 258, + 288, + 312 + ], + "spans": [ + { + "bbox": [ + 53, + 258, + 288, + 312 + ], + "type": "text", + "content": "[4] Angela Dai, Angel X Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5828-5839, 2017. 2, 5, 6, 7, 8" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 313, + 288, + 356 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 313, + 288, + 356 + ], + "spans": [ + { + "bbox": [ + 53, + 313, + 288, + 356 + ], + "type": "text", + "content": "[5] Ruoxi Deng, Chunhua Shen, Shengjun Liu, Huibing Wang, and Xinru Liu. Learning to predict crisp boundaries. In Proceedings of the European conference on computer vision (ECCV), pages 562-578, 2018. 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 357, + 288, + 412 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 357, + 288, + 412 + ], + "spans": [ + { + "bbox": [ + 53, + 357, + 288, + 412 + ], + "type": "text", + "content": "[6] Shenglan Du, Nail Ibrahimli, Jantien Stoter, Julian Kooij, and Liangliang Nan. Push-the-boundary: Boundary-aware feature propagation for semantic segmentation of 3d point clouds. In 2022 International Conference on 3D Vision (3DV), pages 1-10. IEEE, 2022. 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 413, + 288, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 413, + 288, + 456 + ], + "spans": [ + { + "bbox": [ + 53, + 413, + 288, + 456 + ], + "type": "text", + "content": "[7] Cathrin Elich, Francis Engelmann, Theodora Kontogianni, and Bastian Leibe. 3d bird's-eye-view instance segmentation. In German Conference on Pattern Recognition, pages 48-61. Springer, 2019. 1, 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 458, + 288, + 513 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 458, + 288, + 513 + ], + "spans": [ + { + "bbox": [ + 53, + 458, + 288, + 513 + ], + "type": "text", + "content": "[8] Francis Engelmann, Martin Bokeloh, Alireza Fathi, Bastian Leibe, and Matthias Nießner. 3d-mpa: Multi-proposal aggregation for 3d semantic instance segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9031-9040, 2020. 1, 2, 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 514, + 288, + 556 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 514, + 288, + 556 + ], + "spans": [ + { + "bbox": [ + 53, + 514, + 288, + 556 + ], + "type": "text", + "content": "[9] Golnaz Ghiasi, Xiuye Gu, Yin Cui, and Tsung-Yi Lin. Scaling open-vocabulary image segmentation with image-level labels. In European Conference on Computer Vision, pages 540-557. Springer, 2022. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 557, + 288, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 557, + 288, + 612 + ], + "spans": [ + { + "bbox": [ + 48, + 557, + 288, + 612 + ], + "type": "text", + "content": "[10] Jingyu Gong, Jiachen Xu, Xin Tan, Jie Zhou, Yanyun Qu, Yuan Xie, and Lizhuang Ma. Boundary-aware geometric encoding for semantic segmentation of point clouds. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1424-1432, 2021. 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 613, + 288, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 613, + 288, + 656 + ], + "spans": [ + { + "bbox": [ + 48, + 613, + 288, + 656 + ], + "type": "text", + "content": "[11] Xiuye Gu, Tsung-Yi Lin, Weicheng Kuo, and Yin Cui. Open-vocabulary object detection via vision and language knowledge distillation. arXiv preprint arXiv:2104.13921, 2021. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 658, + 288, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 288, + 690 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 288, + 690 + ], + "type": "text", + "content": "[12] Huy Ha and Shuran Song. Semantic abstraction: Openworld 3d scene understanding from 2d vision-language models. In 6th Annual Conference on Robot Learning, 2022. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 691, + 288, + 715 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 288, + 715 + ], + "type": "text", + "content": "[13] Lei Han, Tian Zheng, Lan Xu, and Lu Fang. Occuseg: Occupancy-aware 3d instance segmentation. In Proceedings" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 715 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "text", + "content": "of the IEEE/CVF conference on computer vision and pattern recognition, pages 2940-2949, 2020. 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 96, + 547, + 129 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 96, + 547, + 129 + ], + "spans": [ + { + "bbox": [ + 307, + 96, + 547, + 129 + ], + "type": "text", + "content": "[14] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In Proceedings of the IEEE international conference on computer vision, pages 2961-2969, 2017. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 130, + 547, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 130, + 547, + 185 + ], + "spans": [ + { + "bbox": [ + 307, + 130, + 547, + 185 + ], + "type": "text", + "content": "[15] Tong He, Chunhua Shen, and Anton Van Den Hengel. Dyco3d: Robust instance segmentation of 3d point clouds through dynamic convolution. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 354-363, 2021. 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 186, + 547, + 230 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 186, + 547, + 230 + ], + "spans": [ + { + "bbox": [ + 307, + 186, + 547, + 230 + ], + "type": "text", + "content": "[16] Ji Hou, Angela Dai, and Matthias Nießner. 3d-sis: 3d semantic instance segmentation of rgb-d scans. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4421-4430, 2019. 1, 2, 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 232, + 547, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 232, + 547, + 297 + ], + "spans": [ + { + "bbox": [ + 307, + 232, + 547, + 297 + ], + "type": "text", + "content": "[17] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In International conference on machine learning, pages 4904-4916. PMLR, 2021. 2, 4" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 298, + 547, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 298, + 547, + 354 + ], + "spans": [ + { + "bbox": [ + 307, + 298, + 547, + 354 + ], + "type": "text", + "content": "[18] Li Jiang, Hengshuang Zhao, Shaoshuai Shi, Shu Liu, Chi-Wing Fu, and Jiaya Jia. Pointgroup: Dual-set point grouping for 3d instance segmentation. In Proceedings of the IEEE/CVF conference on computer vision and Pattern recognition, pages 4867-4876, 2020. 1, 2, 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 354, + 547, + 421 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 354, + 547, + 421 + ], + "spans": [ + { + "bbox": [ + 307, + 354, + 547, + 421 + ], + "type": "text", + "content": "[19] Abhijit Kundu, Xiaoqi Yin, Alireza Fathi, David Ross, Brian Brewington, Thomas Funkhouser, and Caroline Pantofaru. Virtual multi-view fusion for 3d semantic segmentation. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXIV 16, pages 518-535. Springer, 2020. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 422, + 547, + 466 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 422, + 547, + 466 + ], + "spans": [ + { + "bbox": [ + 307, + 422, + 547, + 466 + ], + "type": "text", + "content": "[20] Jean Lahoud, Bernard Ghanem, Marc Pollefeys, and Martin R Oswald. 3d instance segmentation via multi-task metric learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9256-9266, 2019. 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 467, + 547, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 467, + 547, + 521 + ], + "spans": [ + { + "bbox": [ + 307, + 467, + 547, + 521 + ], + "type": "text", + "content": "[21] Xin Lai, Jianhui Liu, Li Jiang, Liwei Wang, Hengshuang Zhao, Shu Liu, Xiaojuan Qi, and Jiaya Jia. Stratified transformer for 3d point cloud segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8500-8509, 2022. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 522, + 547, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 522, + 547, + 578 + ], + "spans": [ + { + "bbox": [ + 307, + 522, + 547, + 578 + ], + "type": "text", + "content": "[22] Xin Lai, Yuhui Yuan, Ruihang Chu, Yukang Chen, Han Hu, and Jiaya Jia. Mask-attention-free transformer for 3d instance segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3693–3703, 2023. 1, 2, 3, 4, 5, 6, 8" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 579, + 547, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 579, + 547, + 643 + ], + "spans": [ + { + "bbox": [ + 307, + 579, + 547, + 643 + ], + "type": "text", + "content": "[23] Feng Liang, Bichen Wu, Xiaoliang Dai, Kunpeng Li, Yinan Zhao, Hang Zhang, Peizhao Zhang, Peter Vajda, and Diana Marculescu. Open-vocabulary semantic segmentation with mask-adapted clip. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7061-7070, 2023. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 644, + 547, + 700 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 644, + 547, + 700 + ], + "spans": [ + { + "bbox": [ + 307, + 644, + 547, + 700 + ], + "type": "text", + "content": "[24] Zhihao Liang, Zhihao Li, Songcen Xu, Mingkui Tan, and Kui Jia. Instance segmentation in 3d scenes using semantic superpoint tree networks. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2783-2792, 2021. 6" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 702, + 547, + 715 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 702, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 307, + 702, + 547, + 715 + ], + "type": "text", + "content": "[25] Chen Liu and Yasutaka Furukawa. Masc: Multi-scale affinity" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "20652" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 66, + 72, + 287, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 72, + 287, + 95 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 287, + 95 + ], + "type": "text", + "content": "with sparse convolution for 3d instance segmentation. arXiv preprint arXiv:1902.04478, 2019. 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 97, + 288, + 152 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 97, + 288, + 152 + ], + "spans": [ + { + "bbox": [ + 48, + 97, + 288, + 152 + ], + "type": "text", + "content": "[26] Jiahao Lu, Jiacheng Deng, Chuxin Wang, Jianfeng He, and Tianzhu Zhang. Query refinement transformer for 3d instance segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 18516-18526, 2023. 1, 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 154, + 288, + 211 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 154, + 288, + 211 + ], + "spans": [ + { + "bbox": [ + 48, + 154, + 288, + 211 + ], + "type": "text", + "content": "[27] John McCormac, Ankur Handa, Andrew Davison, and Stefan Leutenegger. Semantic fusion: Dense 3d semantic mapping with convolutional neural networks. In 2017 IEEE International Conference on Robotics and automation (ICRA), pages 4628-4635. IEEE, 2017. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 212, + 288, + 255 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 212, + 288, + 255 + ], + "spans": [ + { + "bbox": [ + 48, + 212, + 288, + 255 + ], + "type": "text", + "content": "[28] Ishan Misra, Rohit Girdhar, and Armand Joulin. An end-to-end transformer model for 3d object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2906-2917, 2021. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 257, + 288, + 312 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 257, + 288, + 312 + ], + "spans": [ + { + "bbox": [ + 48, + 257, + 288, + 312 + ], + "type": "text", + "content": "[29] Songyou Peng, Kyle Genova, Chiyu Jiang, Andrea Tagliasacchi, Marc Pollefeys, Thomas Funkhouser, et al. Openscene: 3d scene understanding with open vocabularies. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 815-824, 2023. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 314, + 288, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 314, + 288, + 358 + ], + "spans": [ + { + "bbox": [ + 48, + 314, + 288, + 358 + ], + "type": "text", + "content": "[30] Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J Guibas. Pointnet++: Deep hierarchical feature learning on point sets in a metric space. Advances in neural information processing systems, 30, 2017. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 360, + 288, + 405 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 360, + 288, + 405 + ], + "spans": [ + { + "bbox": [ + 48, + 360, + 288, + 405 + ], + "type": "text", + "content": "[31] Charles R Qi, Or Litany, Kaiming He, and Leonidas J Guibas. Deep hough voting for 3d object detection in point clouds. In proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9277-9286, 2019. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 406, + 288, + 472 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 406, + 288, + 472 + ], + "spans": [ + { + "bbox": [ + 48, + 406, + 288, + 472 + ], + "type": "text", + "content": "[32] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 2, 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 474, + 288, + 518 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 474, + 288, + 518 + ], + "spans": [ + { + "bbox": [ + 48, + 474, + 288, + 518 + ], + "type": "text", + "content": "[33] David Rozenberszki, Or Litany, and Angela Dai. Language-grounded indoor 3d semantic segmentation in the wild. In European Conference on Computer Vision, pages 125-141. Springer, 2022. 2, 3, 5, 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 520, + 288, + 575 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 520, + 288, + 575 + ], + "spans": [ + { + "bbox": [ + 48, + 520, + 288, + 575 + ], + "type": "text", + "content": "[34] Corentin Sautier, Gilles Puy, Spyros Gidaris, Alexandre Boulch, Andrei Bursuc, and Renaud Marlet. Image-to-lidar self-supervised distillation for autonomous driving data. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9891-9901, 2022. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 577, + 288, + 620 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 577, + 288, + 620 + ], + "spans": [ + { + "bbox": [ + 48, + 577, + 288, + 620 + ], + "type": "text", + "content": "[35] Jonas Schult, Francis Engelmann, Alexander Hermans, Or Litany, Siyu Tang, and Bastian Leibe. Mask3d for 3d semantic instance segmentation. arXiv preprint arXiv:2210.03105, 2022. 1, 2, 3, 4, 5, 6, 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 623, + 288, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 623, + 288, + 667 + ], + "spans": [ + { + "bbox": [ + 48, + 623, + 288, + 667 + ], + "type": "text", + "content": "[36] Jiahao Sun, Chunmei Qing, Junpeng Tan, and Xiangmin Xu. Superpoint transformer for 3d scene instance segmentation. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2393-2401, 2023. 1, 2, 4, 5, 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 669, + 288, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 669, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 669, + 288, + 713 + ], + "type": "text", + "content": "[37] Ayca Takmaz, Elisabetta Fedele, Robert W Sumner, Marc Pollefeys, Federico Tombari, and Francis Engelmann. Openmask3d: Open-vocabulary 3d instance segmentation. arXiv preprint arXiv:2306.13631, 2023. 2" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 546, + 485 + ], + "type": "list", + "angle": 0, + "index": 22, + "blocks": [ + { + "bbox": [ + 307, + 73, + 546, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 546, + 106 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 546, + 106 + ], + "type": "text", + "content": "[38] Laurens Van der Maaten and Geoffrey Hinton. Visualizing data using t-sne. Journal of machine learning research, 9 (11), 2008. 7, 8" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 107, + 546, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 107, + 546, + 161 + ], + "spans": [ + { + "bbox": [ + 307, + 107, + 546, + 161 + ], + "type": "text", + "content": "[39] Thang Vu, Kookhoi Kim, Tung M Luu, Thanh Nguyen, and Chang D Yoo. Softgroup for 3d instance segmentation on point clouds. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2708-2717, 2022. 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 163, + 546, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 163, + 546, + 217 + ], + "spans": [ + { + "bbox": [ + 307, + 163, + 546, + 217 + ], + "type": "text", + "content": "[40] Weiyue Wang, Ronald Yu, Qiangui Huang, and Ulrich Neumann. Sgpn: Similarity group proposal network for 3d point cloud instance segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2569-2578, 2018. 1, 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 219, + 546, + 272 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 219, + 546, + 272 + ], + "spans": [ + { + "bbox": [ + 307, + 219, + 546, + 272 + ], + "type": "text", + "content": "[41] Xinlong Wang, Shu Liu, Xiaoyong Shen, Chunhua Shen, and Jiaya Jia. Associatively segmenting instances and semantics in point clouds. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4096-4105, 2019. 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 274, + 546, + 317 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 274, + 546, + 317 + ], + "spans": [ + { + "bbox": [ + 307, + 274, + 546, + 317 + ], + "type": "text", + "content": "[42] Yizheng Wu, Min Shi, Shuaiyuan Du, Hao Lu, Zhiguo Cao, and Weicai Zhong. 3d instances as 1d kernels. In European Conference on Computer Vision, pages 235-252. Springer, 2022. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 319, + 546, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 319, + 546, + 373 + ], + "spans": [ + { + "bbox": [ + 307, + 319, + 546, + 373 + ], + "type": "text", + "content": "[43] Bo Yang, Jianan Wang, Ronald Clark, Qingyong Hu, Sen Wang, Andrew Markham, and Niki Trigoni. Learning object bounding boxes for 3d instance segmentation on point clouds. Advances in neural information processing systems, 32, 2019. 1, 2, 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 374, + 546, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 374, + 546, + 430 + ], + "spans": [ + { + "bbox": [ + 307, + 374, + 546, + 430 + ], + "type": "text", + "content": "[44] Li Yi, Wang Zhao, He Wang, Minhyuk Sung, and Leonidas J Guibas. Gspn: Generative shape proposal network for 3d instance segmentation in point cloud. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3947-3956, 2019. 1, 2, 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 431, + 546, + 485 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 431, + 546, + 485 + ], + "spans": [ + { + "bbox": [ + 307, + 431, + 546, + 485 + ], + "type": "text", + "content": "[45] Min Zhong, Xinghao Chen, Xiaokang Chen, Gang Zeng, and Yunhe Wang. Maskgroup: Hierarchical point grouping and masking for 3d instance segmentation. In 2022 IEEE International Conference on Multimedia and Expo (ICME), pages 1-6. IEEE, 2022. 6" + } + ] + } + ], + "index": 21 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20653" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Edit One for All_ Interactive Batch Image Editing/d0962876-52e1-4fef-ac92-9ddbf4fa5230_content_list.json b/2024/Edit One for All_ Interactive Batch Image Editing/d0962876-52e1-4fef-ac92-9ddbf4fa5230_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..5cf418a5468de2e80ce22c852beb1c13537e0d5a --- /dev/null +++ b/2024/Edit One for All_ Interactive Batch Image Editing/d0962876-52e1-4fef-ac92-9ddbf4fa5230_content_list.json @@ -0,0 +1,1435 @@ +[ + { + "type": "text", + "text": "Edit One for All: Interactive Batch Image Editing", + "text_level": 1, + "bbox": [ + 233, + 130, + 736, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Thao Nguyen Utkarsh Ojha Yuheng Li Haotian Liu Yong Jae Lee University of Wisconsin-Madison", + "bbox": [ + 187, + 180, + 781, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "https://thaoshibe.github.io/edit-one-for-all", + "bbox": [ + 282, + 219, + 676, + 232 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 268, + 313, + 284 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In recent years, image editing has advanced remarkably. With increased human control, it is now possible to edit an image in a plethora of ways; from specifying in text what we want to change, to straight up dragging the contents of the image in an interactive point-based manner. However, most of the focus has remained on editing single images at a time. Whether and how we can simultaneously edit large batches of images has remained understudied. With the goal of minimizing human supervision in the editing process, this paper presents a novel method for interactive batch image editing using StyleGAN as the medium. Given an edit specified by users in an example image (e.g., make the face frontal), our method can automatically transfer that edit to other test images, so that regardless of their initial state (pose), they all arrive at the same final state (e.g., all facing front). Extensive experiments demonstrate that edits performed using our method have similar visual quality to existing single-image-editing methods, while having more visual consistency and saving significant time and human effort.", + "bbox": [ + 76, + 294, + 473, + 582 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 603, + 209, + 619 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Image editing has undergone a transformation in recent years with the aid of modern generative models. The process has been greatly democratized, where many of the sophisticated edits, which previously might have required hours and niche expertise, can now be completed with relative ease in a matter of minutes. For example, different types of learning based algorithms can be used for image correction/adjustment [1-12] and for manipulating the semantic contents of real images [13-22]. Moreover, there are many different ways to edit an image. For instance, a user can specify in text what changes they want - e.g., \"make the hair darker\" [14, 15, 23, 24], or they can drag the contents of the image in an interactive manner to shrink, enlarge, or move a part of an object [12, 16, 17, 19, 25].", + "bbox": [ + 75, + 628, + 468, + 839 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "A common theme across many of these works is that the edits are designed to work for a single image at a time. For example, given an image of a cat with open eyes, DragGAN [17] allows the user to drag the contents of that par", + "bbox": [ + 75, + 839, + 468, + 901 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/0eb91e7d6b9d8e795b992b5d273fc19a0ad33d6f0a5c2b84cecc4e7dd125563f.jpg", + "image_caption": [ + "Figure 1. Interactive Batch Image Editing. Given a single user edited image, the goal is to automatically transfer that edit to new unseen images, so that all edited images end up with the same final state as the user edited image." + ], + "image_footnote": [], + "bbox": [ + 506, + 267, + 883, + 561 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ticular image so that both the eyes can be closed in the resulting cat image. But what if we wish to apply the same edit to many different types of cats so that we can close all of their eyes? While one could perform dragging on each cat separately, the whole process would be quite cumbersome (requiring lots of human annotation) and time intensive (requiring optimization for each image).", + "bbox": [ + 496, + 641, + 890, + 747 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this paper, we introduce the new problem of Interactive Batch Image Editing. Given a user-specified edit $I \\rightarrow I'$ , the goal is to automatically transfer that edit to new unseen images, so that all edited images end up with the same final state as $I'$ (e.g., all cats with eyes closed) regardless of their original starting states (e.g., varying degrees of original eye openness); see Figure 1. Hence, a solution to this problem requires two components: (i) modeling the user edit in the example pair $(I, I')$ so that it can be transferred to new images; and (ii) controlling the degree of the", + "bbox": [ + 496, + 750, + 892, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 807, + 44 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "8271", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/a2fbce35151b420551fee3a48bde474afebeb2ffd3af5f9c5decb05340526b8d.jpg", + "image_caption": [ + "edit one image at a time", + "Figure 2. Single Image Editing vs. Batch Image Editing. (a) Prior work (e.g., [17, 19, 25]) focuses on single image editing. (b) We focus on batch image editing, where the user's edit on a single image is automatically transferred to new images, so that they all arrive at the same final state regardless of their initial starting state. In this way, we can achieve time speed up and reduce human effort in editing." + ], + "image_footnote": [], + "bbox": [ + 91, + 104, + 447, + 258 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/2ddb6c5d6e47094d0816beaae5e12608cb13fb44ebde5f8a66a4d9a97f06ab26.jpg", + "image_caption": [ + "edit one image, transfer to others at the same time!" + ], + "image_footnote": [], + "bbox": [ + 465, + 103, + 874, + 258 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "edit so that all edited images have the same final state.", + "bbox": [ + 75, + 318, + 434, + 333 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To model the user's edit, we build upon existing literature on Generative Adversarial Networks (GANs) [23, 26-29], and in particular StyleGAN [30, 31], which shows that their learned latent space emits globally linear editing directions where the edited attribute (e.g., eye closeness) varies linearly in magnitude along such a direction for all semantically related instances (e.g., all closed eyes cat). To try to discover the global direction corresponding to a user edit, we perform optimization in the latent space of StyleGAN2 [31] so that (i) the edited image with the discovered direction is visually similar to the original edited image $I'$ , and (ii) the linear correlation between the distance along the direction and the strength of the visual attribute is highest (e.g., going twice the amount in that direction closes a cat's eyes by twice the amount).", + "bbox": [ + 75, + 335, + 468, + 561 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To control the degree of the edit so that all edited images end up with the same final state, we derive a closed-form solution. Taking motivation from [29], we model the attribute strength (e.g., how much the cat's eyes are open) as the distance along the normal vector from a hyperplane. By setting the example edited image as lying on that hyperplane (i.e., its distance is 0), the objective is to move all other images along that same direction so that their distances to the hyperplane also become 0. In this way, regardless of the start state of any image (e.g., varying degree of eye openness), their edited end states all become the same (e.g., closed eyes). We show that we can analytically compute the exact amount of traversal along the direction for any image with a closed-form solution. Editing images in this way helps improve the results visually, as the final state of any new image matches that of the edited example given by the user.", + "bbox": [ + 75, + 565, + 468, + 806 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our method works with various edits given by an editing framework, e.g., it can be a change given by a dragging operation [17], or it can be a change given by a text-based edit [14, 23] (e.g., \"make eyes bigger\"), as long as it is a direction compatible with a StyleGAN2 [31] model. We present qualitative and", + "bbox": [ + 75, + 810, + 468, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "quantitative results in transferring edits for a variety of objects (cats/dogs/faces/humans/limons/arts/etc.), parts (mouths/ears/legs/etc.), and corresponding attributes (big eyes/short faces/pale skins/etc.). Importantly, we show that the final state of the edited images is comparable to the scenario in which a user performs the edit for each image separately. But since our method does this automatically (see Fig. 2), we show that it takes much less time (e.g., only 0.05s per image, compared to DragGAN [17] about 2s per image), and does not require laborious human annotation (e.g., only need 1 annotated image, compared to 4 required for DragGAN [17]). Finally, we show practical applications of batch image editing; e.g., changing wheel size in every photo of a car collection by editing just one.", + "bbox": [ + 496, + 318, + 890, + 529 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In sum, our contributions are: (1) We introduce the novel problem of interactive batch image editing, wherein a user-driven edit is automatically transferred to other similar test images. (2) We study what having the same final state means in a geometrical sense, and propose a principled approach to achieving that in StyleGAN2 [31]'s latent space. (3) We show that our method works on a wide variety of domains e.g., cats, dogs, humans etc., taking significantly less time than editing each image individually while having similar visual quality and more visual consistency.", + "bbox": [ + 496, + 530, + 892, + 681 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 694, + 640, + 709 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Latent space of GANs. The capabilities of generative adversarial networks (GANs) have transformed drastically, from the first GAN [32] designed for simple image datasets like MNIST [33] to big and powerful models like BigGAN [34], Progressive GAN [35] and StyleGANs [30, 31] designed for much more complex datasets [36, 37]. Parallel to the efforts to enhance their capabilities, there has also been work done to better understand and make use of their learned latent space. Specifically, there is a line of work which tries to find interpretable directions in that space, so that moving in such a direction majorly changes only one discernible attribute in the image. Some methods [38, 39]", + "bbox": [ + 496, + 719, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "8272", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/ed94eafe60804ef8f989e160c3fc16fc755bfb18cce8c08b180a96a1fa5d073d.jpg", + "image_caption": [ + "(a) Setting" + ], + "image_footnote": [], + "bbox": [ + 80, + 88, + 254, + 268 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/8dd88c9fcd19d6b656d62f993e17a4355cd8639d14d75990bae421b3283904c9.jpg", + "image_caption": [ + "(b) Naïve Approach" + ], + "image_footnote": [], + "bbox": [ + 269, + 87, + 405, + 268 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/cf3b505c6e9ae51a1836b378c628907361c396818e3a871ec748313ca0e39026.jpg", + "image_caption": [ + "(c) Optimizing Editing Direction", + "Figure 3. Different editing strategies. (a) Setting. (b) Naive Approach: The editing direction effective for an example may not generalize well to test images. (c) Optimizing Editing Direction: We optimize for a globally consistent direction that is effective for both example and test images. (d) Adjusting Editing Strength: Ensuring consistent final states requires adjusting the editing strength for each test image." + ], + "image_footnote": [], + "bbox": [ + 421, + 88, + 619, + 268 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/a4a06c1b516757ff2a327010b4edc31624f4671b2108a1a73a8da292a17eea67.jpg", + "image_caption": [ + "(d) Adjusting Edit Strength" + ], + "image_footnote": [], + "bbox": [ + 625, + 87, + 890, + 268 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "try to find such directions or particular activations using (self) supervised methods. Others [26-28] try getting rid of the need for supervision: discovering the directions representing the most important factors of variation, whatever they may be. Going beyond the $\\mathcal{W}$ space, the authors in [40] explore the $\\mathcal{W}+$ space, and show that it has even better disentanglement useful for spatial image editing. All of these hidden capabilities have made StyleGANs a useful tool for editing purposes. However, prior work mainly focuses on the effectiveness of the editing direction (e.g., whether the discovered direction can change yaw pose). The question of whether these directions can be applied to diverse images to yield consistent results (e.g., all faces facing frontal) and, if so, how to achieving such consistency, remains unanswered.", + "bbox": [ + 75, + 345, + 472, + 556 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Image editing with generative models. Researchers have used StyleGAN for real image editing by designing an encoder to invert a real image into StyleGAN's latent space [41-43]. StyleCLIP [23] presents a way to perform image editing through a text-based interface by making use of the CLIP encoder [44]. Very recently, StyleGAN has also been used to perform point-based editing [17, 25] so as to move any start point to reach a target location in the image; thereby elongating, rotating, shifting the objects. Most work on GANs has been for single object category datasets, but with the rise of text-to-image Diffusion Models which can generate complex images [13, 45, 46], editing such images is now possible. Complex scenes can be edited either using text [14, 15, 21, 24, 47, 48], or using the same idea of point-based manipulation [16, 19]. There have been some works which discuss the possibility to transfer the edit of one image to another. EditGAN [49] uses segmentation mask manipulations to edit an image, but to successfully transfer them to an unseen image, it needs to do post hoc manipulation of the editing scale. RewriteGAN [50] involves editing one generated sample; e.g., adding a patch of a tree onto a church tower. Following this, the rules of the", + "bbox": [ + 75, + 569, + 472, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "GAN are manipulated so that all churches have some tree on the top. Within diffusion models, visual prompting/ image analogies tries something similar, where users can define a triplet {before, after, test} to learn the edit and transfer it to a test image [51-55]. Our task is similar, in that we wish to transfer user-edits from the training example to new images, but it differs in one crucial aspect: along with transferring the edit, we wish to automatically learn its strength so that the edit produces the same final state for a new image.", + "bbox": [ + 496, + 345, + 893, + 482 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Approach", + "text_level": 1, + "bbox": [ + 500, + 494, + 607, + 513 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our focus is on the setting in which an image editing process edits an attribute of an image so that the attribute's value reaches a desired state; e.g., rotating the pose of a face looking sideways so that it faces front (attribute = face's pose, value = front).", + "bbox": [ + 496, + 521, + 893, + 595 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "With this view, we explain our framework for batch image editing, which can be broken down into two stages: (i) A user edits an image $I_0$ (e.g., using DragGAN [17]) to obtain an edited image $I_0'$ . We describe in Sec. 3.1 how we capture this user edit $I_0 \\rightarrow I_0'$ , so that the same edit can be applied to new images. (ii) Next, we describe in Sec. 3.2, how for any new image, we apply the modeled edit by automatically adjusting its strength so that the attribute's value in this new image matches that of $I_0'$ (e.g., any face, regardless of its initial pose, now faces front after the edit).", + "bbox": [ + 496, + 595, + 890, + 748 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Modeling the User Edit", + "text_level": 1, + "bbox": [ + 500, + 756, + 717, + 772 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We start with an image $I_0$ . This image could be a real image or a generated one. Either way, we get its latent representation in the $\\mathcal{W}$ space of a StyleGAN2 [31] model $G$ , so that $I_0 = G(w_0)$ . (For real images we can use GAN-inversion techniques [56].) The user, with the help of an image editing framework (e.g., DragGAN [17] or InstructPix2Pix [14]), edits $I_0$ to manipulate one or more of its attributes. The edit maps to the $\\mathcal{W}$ space as $w_0 \\to w_0'$ . The", + "bbox": [ + 496, + 779, + 890, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "8273", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "resulting edited image can thus be recovered as follows: $I_0' = G(w_0')$ . Fig. 3(a) shows an example of the original and edited image pair, $(I_0, I_0')$ , where the user intended to turn the face forward.", + "bbox": [ + 75, + 90, + 468, + 150 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Now, given the user edit $I_0 \\to I_0'$ , we wish to capture it in a way that can be applied to new images in a generalizable manner; i.e., the application of the edit changes the same property in a new image. This is where a nice property of GANs, and in particular StyleGANs [30, 31] comes in useful. It has been shown in prior works [26-28] that it is possible to discover directions with such properties (using supervised as well as unsupervised methods) in the learned $\\mathcal{W}$ space. In particular, it is possible to find directions $(\\Delta_w)$ that are globally consistent. Taking motivation from [29], we define a globally consistent direction $\\Delta_w$ as the following: for any arbitrary $w$ , moving along $\\Delta_w$ , $w \\to w + \\Delta_w$ (i) changes the same attribute, and (ii) by the same amount.", + "bbox": [ + 75, + 151, + 468, + 348 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To make the precise user edit $I_0 \\to I_0'$ applicable to other images, it needs to be captured as a globally consistent direction. The naive way to represent that edit will be through a simple difference in the $\\mathcal{W}$ space: $\\Delta_w = w_0' - w_0$ . However, empirically, and as we show in Fig. 3(b), applying this $\\Delta_w$ to the latent code $w_1$ corresponding to a new image $I_1$ does not always result in the same change; while $I_0 \\to I_0'$ results in a pose change of $\\sim 30^\\circ$ degrees in yaw, $I_1 \\to I_1'$ does not seem to change the same attribute or at least not by the same amount.", + "bbox": [ + 75, + 348, + 468, + 498 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Hence, our goal is to represent the $I_0 \\to I_0'$ edit through a $\\mathcal{W}$ space direction that can better satisfy both the properties of a globally consistent direction. For this task, we first take motivation from LARGE [29] to introduce a mathematical view of what those directions mean.", + "bbox": [ + 75, + 500, + 468, + 574 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Globally consistent direction. Let's say that $\\Delta_g$ is one such direction. Along with a bias term $b$ , we can define a hyperplane as follows:", + "bbox": [ + 75, + 583, + 468, + 628 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nw \\cdot \\Delta_ {g} + b = 0 \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 218, + 633, + 468, + 650 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "That is, any point $w$ which lies on the hyperplane will satisfy this equation. The authors in [29] argued that for such a hyperplane, whose normal vector is a globally consistent direction $(\\Delta_g)$ , the distance of an arbitrary point $w$ from that hyperplane $(w \\cdot \\Delta_g)$ will be linearly correlated to the actual attribute $(y)$ that results in the generated image:", + "bbox": [ + 75, + 662, + 468, + 753 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\ny = a \\times (w \\cdot \\Delta_ {g}) + b \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 197, + 766, + 468, + 782 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Here, $a$ and $b$ are unknown linear coefficients. For example, if $\\Delta_g$ corresponds to change in pose, then all the front facing people will lie at the same distance $d$ from the above hyperplane. Because of this, any hyperplane defined with respect to a globally consistent direction can be viewed as a semantic hyperplane. For simplicity, we can set $d = 0$ for the edited image $I_0'$ .", + "bbox": [ + 75, + 794, + 468, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Given that the original $\\Delta_w$ may not always be globally consistent, we aim to discover, through optimization, a different direction $\\Delta_w^*$ which produces a similar editing effect, but is more likely to be a globally consistent one. We initialize $\\Delta_w^*$ with 0's, and design two objective functions to optimize it. First, to make sure that it produces a similar effect as the original direction, we use an image reconstruction loss so that the edited image produced by the new edit matches the original edit given by the user:", + "bbox": [ + 496, + 90, + 890, + 227 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {i m g} = \\left\\| G \\left(w _ {0} + \\Delta_ {w}\\right) - G \\left(w _ {0} + \\Delta_ {w} ^ {*}\\right) \\right\\| _ {2} \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 553, + 234, + 890, + 252 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Additionally, user can provide the real value of distance $d$ , otherwise we set the distance of the edited image from the hypothetical hyperplane to be 0 (similar to [29]). We constrain the learned $\\Delta_w^*$ to follow this property for better interpretability. That is, the new direction should be such that when we traverse the original latent code in that direction, $w_0 \\rightarrow w_0 + \\Delta_w^*$ , the resulting image should lie at the hyperplane defined by that direction. Setting $b = 0$ in Eq. 1, we minimize the edited point's distance from the hyperplane:", + "bbox": [ + 496, + 261, + 890, + 412 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {a t t} = \\left| \\left(w _ {0} + \\Delta_ {w} ^ {*}\\right) \\cdot \\Delta_ {w} ^ {*} \\right| \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 606, + 417, + 890, + 433 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The overall loss function for optimizing the new direction $\\Delta_w^*$ is a weighted sum of the two objectives: $\\mathcal{L}_{\\Delta} = \\mathcal{L}_{img} + \\lambda \\mathcal{L}_{att}$ . We set $\\lambda$ to 0.02 in all experiments to balance the magnitude of $\\mathcal{L}_{img}$ and $\\mathcal{L}_{att}$ .", + "bbox": [ + 496, + 444, + 890, + 503 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Now, in Fig. 3 (b) and (c), we compare the difference in image editing results using two different directions when applied to the same new image $(I_1)$ : $w_1 \\to w_1 + \\Delta_w$ vs. $w_1 \\to w_1 + \\Delta_w^*$ . As discussed before, in this case, the editing using the naively computed direction $(\\Delta_w)$ is not able to accurately capture the pose rotation the way it did for the user edited example $(I_0 \\to I_0')$ . On the other hand, with $\\Delta_w^*$ , we see that the pose of the woman changes by a similar amount as the edited example. However, since the original pose of the woman $(I_1)$ was not the same as the pose of the man $(I_0)$ , the new edited image $(I_1')$ still does not arrive at the same final state as $I_0'$ . Therefore, our next goal is to figure out how to scale the learned direction so that $I_1'$ does arrive at the same final state as $I_0'$ .", + "bbox": [ + 496, + 505, + 890, + 717 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Adjusting Editing Strength for New Images", + "text_level": 1, + "bbox": [ + 498, + 724, + 867, + 741 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Given the optimized $\\Delta_w^*$ and some new image $I_{i} = G(w_{i})$ we wish to edit it in the following way:", + "bbox": [ + 496, + 748, + 888, + 777 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nw _ {i} ^ {\\prime} = w _ {i} + \\alpha_ {i} n \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 642, + 782, + 890, + 797 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\alpha_{i}$ is the editing strength computed separately for each image and $n = \\Delta_w^* /||\\Delta_w^* ||$ is a unit vector in the direction of $\\Delta_w^*$ . Fig. 3 (d) illustrates a geometric perspective that we will use to compute $\\alpha_{i}$ . First, we see $\\Delta_w^*$ represented as a vector and its corresponding hyperplane that is normal to it. Next, we depict the new latent point $w_{i}$ . The", + "bbox": [ + 496, + 810, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "8274", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/c8a7088e79d3a3ec3755d382e2c6186bf27b6e736d51de3ff8d7bbc1f2b622ea.jpg", + "image_caption": [ + "Original", + "UserID", + "DragGAN", + "DragDiff.", + "Ours", + "Figure 4. Qualitative comparisons between dragging baselines. For ours, green bounding box indicates automatic transfer from the red bounding box example in the first row (i.e., no point annotation needed!)." + ], + "image_footnote": [], + "bbox": [ + 81, + 88, + 478, + 335 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "", + "bbox": [ + 395, + 88, + 475, + 335 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "goal is to move it along the $\\Delta_w^*$ direction so that it arrives onto the hyperplane. Through this depiction, $\\alpha_{i}$ can be understood as the distance of $w_{1}$ from the hyperplane. We can get the distance by projecting $(w_0' - w_i)$ in the direction of $n$ . Therefore, $\\alpha_{i} = (w_{0}^{\\prime} - w_{i})\\cdot n$ .", + "bbox": [ + 75, + 396, + 468, + 470 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Since for each image, the only unique computation that needs to be done is the calculation of $\\alpha_{i}$ , we can see why our method will be much faster than, for example, annotating every image and running the DragGAN [17] optimization each of those times. We will show in our experiments the significant difference in time taken by our method compared to single image editing baselines.", + "bbox": [ + 75, + 473, + 468, + 577 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Importantly, there is an added benefit to computing $\\alpha$ 's in this way. Let's say the goal of the editing process is to rotate $n$ faces and make them frontal. After completion, all of them do become frontal, each with their own editing strengths $\\{\\alpha_{1},\\alpha_{2},\\dots,\\alpha_{n}\\}$ computed using the above formula. Suppose the user now wants the same faces facing a bit left instead. To do this, the user does not need to re-associate the original training example and run the optimization one more time. In an interactive manner, they can simply scale the $\\alpha$ for the training example to match the desired edit, and all other $\\alpha$ 's can be automatically recomputed.", + "bbox": [ + 75, + 579, + 467, + 744 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 76, + 760, + 207, + 777 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We study how well our method models and transfers the edits from an example, and how efficient it is compared to single image editing baselines using their official code.", + "bbox": [ + 75, + 785, + 468, + 830 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Categories. We evaluate on a variety of domains: Human faces (FFHQ) [30], AFHQ Cats, Dogs [57], MetFaces [58], Human bodies [59]. For each domain, we use the corresponding pretrained StyleGAN model to perform editing.", + "bbox": [ + 75, + 839, + 468, + 901 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/ea1041f98c84f017aca711fc32ed3636d05645398c67c2f22e408c5baca9685c.jpg", + "image_caption": [ + "Original", + "UserID", + "DragGAN", + "DragDiff.", + "Ours" + ], + "image_footnote": [], + "bbox": [ + 491, + 88, + 887, + 335 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Qualitative results", + "text_level": 1, + "bbox": [ + 500, + 396, + 676, + 411 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We start with qualitative comparisons with (i) interactive point-dragging, and (ii) text-based image editing baselines.", + "bbox": [ + 496, + 420, + 890, + 450 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Interactive point based editing. We compare against DragGAN [17], UserControllableLT [25], and DragDiffusion [16]. For each baseline, we, as a user, desire to edit a bunch of images to have the same final state for an attribute. Fig. 4 shows the results of editing two kinds of edits to four images (leftmost column). In the left case, the goal is to make everyone smile by the same degree. In the right, it is to vertically compress everyone's face by the same degree. For all three baselines, we manually annotate points for each test image. For our method, we only annotate points and perform DragGAN-based edit on one image (top row), then automatically transfer the edit to the three other test images.", + "bbox": [ + 496, + 460, + 890, + 657 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For the smiling case (left), we notice that when DragGAN drags the upper/lower lips up/down respectively, it can either make people smile (1st, 2nd and 4th rows), or it can make people look shocked (blonde girl; 3rd row). Both edits are technically correct based on lips movement, but it won't match the user expectation of making everyone smile by the same degree. Since all four images are edited independently, this is not unexpected. The other two baselines sometimes have issues introducing the same edit across different images; e.g., UserControllableLT can make the blonde girl smile a little bit, but not the other images. In general, we notice in our experiments that to introduce the same exact edit that we want, we often need to play around with some hyperparameters (e.g., # of iterations). In comparison, our method produces edited images that are smiling more equally (left) and have been compressed by a similar", + "bbox": [ + 496, + 659, + 892, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "8275", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/df55c72c3c40102246cc15bc76fd8a26172d87785963f53ca1907f922ed2d67f.jpg", + "image_caption": [ + "Figure 5. Qualitative comparisons to text-guided baselines. Ours transfers the edit from example (red), to other test images (green)." + ], + "image_footnote": [], + "bbox": [ + 80, + 89, + 480, + 358 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/eae8fe8079eafdcd4acca4e58f1046c8620c9b2f042b98a7a7d7afeb5da25fd6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 491, + 90, + 890, + 358 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "amount (right), and with much less human effort.", + "bbox": [ + 76, + 391, + 403, + 405 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Text-driven image editing. We consider two baselines: StyleCLIP [23] and InstructPix2Pix [14]. Each takes an image and text as input to produce an edited image: $I \\rightarrow I'$ . Fig. 5 shows the results of Ours vs. StyleCLIP (left) and Ours vs. InstructPix2Pix (right). (For the latter, we invert the output of InstructPix2Pix into StyleGAN's latent space using [42] for our method.) In each case, for the baselines, we use the same text prompt, e.g., \"A photo of a person with pale skin\" for each edited image (in four rows). For ours, we capture the edit from the first editing result of a baseline (first row), and then automatically transfer the edit to the remaining three images.", + "bbox": [ + 75, + 415, + 468, + 595 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Our goal with this particular setup is to test (i) how consistently the baselines introduce the edit denoted by text to different images (e.g., do all people become equally pale?), and (ii) how consistently our method captures and transfers the particular edit of the first example to the rest of the images, irrespective of how good that first edit was. For (i), we find that StyleCLIP sometimes has issues; e.g., the paleness of the edited face in first example is different from the second (first vs second rows). For InstructPix2Pix, sometimes it can give consistent results; the childness of different faces seems similar. But, it can have consistency issues in other examples; e.g., the type/strength of maleness introduced is different in each image. And this is where, we believe, the utility of our method lies: if one desires to edit every face to be male in a particular way (i.e., to the same degree) as the first example, e.g., thick eyebrows & light beard, our method has an advantage.", + "bbox": [ + 75, + 597, + 468, + 853 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Finally, we show our method's results for non-facial domains in Fig. 6. For the lion examples (top left), we see that it can preserve not just the type of edit, e.g., dragging", + "bbox": [ + 76, + 854, + 468, + 901 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/58f1e5f00f5735cf88163ff1e50a0075a74642cb43b36a3db508b2c3307ec400.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodPoint DraggingTimeAnno. #imgs
Dist.FID1 img1k imgs
UserControl. [25]13.6425.320.03s30s1000
DragGAN [17]4.1659.282s33.33m1000
DragDiffusion [16]26.5636.5560s16.67h1000
Ours9.4679.352s82s1
", + "bbox": [ + 500, + 387, + 888, + 477 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1. Time is estimated for 1 point drag, without human annotation time. (82s includes 2s to perform edit on the example, 30s to optimize the editing direction, and 50s to transfer the edit to 1000 test images.) Our method requires only one image annotation in total, while the baselines need one annotation per test image.", + "bbox": [ + 496, + 487, + 890, + 556 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "to make the lion roar, but also the strength of the roar. The strength being preserved can be observed more easily for the human body poses (bottom left), where we see that the extent of legs split, hand movements, is consistent enough to almost align the edited test images with the user edited one. Overall, results on these diverse set of domains highlight an observation that many kinds of edits can be thought of as a combination of {type, strength}, both of which can indeed be captured and transferred according to our needs.", + "bbox": [ + 496, + 561, + 892, + 699 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Quantitative results", + "text_level": 1, + "bbox": [ + 500, + 710, + 687, + 724 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Next, perform quantitative experiments following the setup proposed in [16, 17, 19]. We first randomly sample 10 facial test images and pair them with a randomly sampled target face image. We use dlib-ml [60] to detect keypoints in the test and target images. Our goal is to see if the keypoints in the test images can be moved to the target locations specified by keypoints in the target image. For the baselines [16, 17, 25], we perform dragging for each test image. For our method, we perform dragging on one test image, and then transfer the edit to the remaining nine images. This is repeated 100 times; i.e., each time a random target image is", + "bbox": [ + 496, + 734, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "8276", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/44f2934bfb3f2ad87d0e208d632ee7cc74a7d5cce1f23c1c0b29ef27845b63e6.jpg", + "image_caption": [ + "Figure 6. Additional qualitative results on various domains." + ], + "image_footnote": [], + "bbox": [ + 140, + 85, + 836, + 601 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "paired with 10 random test images.", + "bbox": [ + 76, + 626, + 313, + 642 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We report the Euclidean distance between the keypoints of edited and target images in Table 1. Even without requiring annotations for every test image, our method can move the points very close to the target; almost as close as DragGAN, which requires annotation each time. We also compute FID [61] between the original test images and their edited versions to ensure that our method does not distort the image; and our quality is comparable to DragGAN.", + "bbox": [ + 75, + 643, + 468, + 763 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5. Deeper Analysis", + "text_level": 1, + "bbox": [ + 76, + 781, + 236, + 799 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We perform deeper analysis on the usefulness of our batch image editing framework by focusing on a specific edit: face pose rotation. In particular, if we wish to rotate many faces to front, how do we visualize the improvements brought by different components of our method?", + "bbox": [ + 76, + 808, + 468, + 883 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Fig. 7 depicts the setup. We use DragGAN to perform", + "bbox": [ + 94, + 885, + 468, + 901 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "dragging on one image (top; $w_0 \\rightarrow w_0'$ ). We then transfer the edit to 1000 other test images (two shown in rows 2-3). We see that naively applying $\\Delta_w = w_0' - w_0$ from DragGAN to test images cannot make them frontal ( $0^\\circ$ ). Using our optimized $\\Delta_w^*$ (without dynamic scaling) does help it bring closer to facing front. Does this mean that $\\Delta_w^*$ might be a more global direction? We study this in Fig. 8 (a), where we visualize the degree to which the correlation property of global directions (Eq. 2) holds for $\\Delta_w$ and $\\Delta_w^*$ individually. We see that there is indeed a better correlation ( $R^2$ ) between the distances of latents from hyperplanes and the Yaw degree of resulting images. (We predict Yaw using 6DRepNet [62].) However, as we discussed before, $\\Delta_w^*$ in itself is insufficient: to completely bring the facial pose to front, we need to scale it with corresponding $\\alpha$ 's to bring them much closer to front (Fig. 7, 4th column). To study this in a more systematic way, we visualize the effect on all 1000 test images. Similar to Fig. 8 (a), we visualize", + "bbox": [ + 496, + 626, + 893, + 898 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "8277", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f60e5e943108b02f35a126b09dc2235f141c9eec4980e764b72229e506fa1bd1.jpg", + "image_caption": [ + "Figure 7. Effect of adjusted and optimized editing for test images. Yaw degree of each image is provided on top of each images." + ], + "image_footnote": [], + "bbox": [ + 83, + 88, + 462, + 334 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/ae8ac06b7f4964e0935dddaf672482e9c56b3867bdbc39f61e567831c351ae83.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodMAETime complexity
PrepareInference
Random11.295 ± 8.972--
DragGAN [17] + GANgealing [63]8.141 ± 7.22130s2000s
Ours2.120 ± 1.81832s50s
", + "bbox": [ + 78, + 375, + 467, + 444 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 2. Ablation Study: DragGAN + GANgealing.", + "bbox": [ + 117, + 445, + 426, + 458 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "distance-to-hyperplane vs. Yaw degree for the original test images (blue) and edited images (red) in Fig. 8 (b). We can see that the variation in Yaw for the edited images collapses at around $0^{\\circ}$ ; i.e., they mostly face front as we would like.", + "bbox": [ + 75, + 468, + 468, + 527 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We also compute the time it takes to perform editing in Table 1, last three columns. Our method only requires annotating one image, while the baselines require annotating each image (e.g., 1000 total in this case). Note the times do not include human annotation time.", + "bbox": [ + 75, + 529, + 468, + 603 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Other ways to automate batch image editing. One baseline for batch image editing (i.e., so that all edited images achieve the same final state) is shown in Fig. 9. After the user annotates the source/target points in the example image, we use GANgealing [63] to transfer the points to corresponding locations in each new test image. The edited images are then obtained using DragGAN (DragGAN + GANgealing). We compare our method to this baseline when the user edits one face image to become front $(\\mathrm{Yaw} = 0^{\\circ})$ and transferring it to 1000 other test images.", + "bbox": [ + 75, + 613, + 468, + 763 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Results are shown in Table 2. We report mean absolute error (MAE) between the Yaw degree of the edited image and its ideal frontal image $(\\mathrm{Yaw} = 0^{\\circ})$ . 'Random' shows the variations in Yaw for the original images (before editing). While 'DragGAN + GANgealing' does help in reducing the variation $(8.14 < 11.29)$ , our edited images are much more frontal, with an MAE of 2.12. This is because the baseline has no way to bring all edited images to the same final state. On top of that, since our method does not rely on", + "bbox": [ + 73, + 765, + 468, + 900 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/2344e8afeb8e2a2009e2d8ace1e9bc1b02dd78946e357c60d7b839bb4c43ccb5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 506, + 90, + 883, + 196 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/73454f2fb7eaedc5a334eca6e89fb4fe69459100f142d30f76d4839735cb0786.jpg", + "image_caption": [ + "Figure 8. (a) After optimization, the editing direction is more linearly correlated with the yaw attribute. (b) With automatically adjusted editing scale for each test image." + ], + "image_footnote": [], + "bbox": [ + 506, + 200, + 883, + 306 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/870b11d8902c50cfe052a24c06cf8e5498bccb50ab164a08a9254cf03e894bd9.jpg", + "image_caption": [ + "Figure 9. Ablation study set up. We use GANgealing [63] to transfer the annotated points from example to test image." + ], + "image_footnote": [], + "bbox": [ + 504, + 359, + 888, + 455 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "additional information about keypoints (which 'DragGAN + GANgealing' does), we believe that it is better suited for batch image editing for many kinds of domains (e.g., Fig. 6), where we might not have such information.", + "bbox": [ + 498, + 496, + 890, + 556 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 574, + 617, + 589 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We introduced the problem of interactive batch image editing. Given a user edit in an example image, our approach automatically transfers that edit to other test images, maintaining a consistent final state of the edit across images. Extensive experiments demonstrated that our method produces comparable quality to state-of-the-art single-image-editing methods while saving significant time and human effort. We are currently limited to StyleGAN models. Extending this problem and solution to diffusion-based models for more edits types would be an exciting future direction.", + "bbox": [ + 496, + 599, + 890, + 751 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgement", + "text_level": 1, + "bbox": [ + 500, + 768, + 658, + 785 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This work was supported in part by NSF CAREER IIS2150012, Adobe Data Science award, Sony Focused Research award, and Institute of Information & communications Technology Planning & Evaluation (IITP) grant funded by the Korea government (MSIT) (No. RS-2022-00187238, Development of Large Korean Language Model Technology for Efficient Pretraining).", + "bbox": [ + 496, + 794, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8278", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Man M. Ho and Jinjia Zhou. Deep preset: Blending and retouching photos with color style transfer. In WACV, 2021. 1", + "[2] Mahmoud Afifi, Konstantinos G. Derpanis, Bjorn Ommer, and Michael S. Brown. Learning multi-scale photo exposure correction. In CVPR, 2021.", + "[3] Jong-Hyeon Baek, DaeHyun Kim, Su-Min Choi, Hyo-jun Lee, Hanul Kim, and Yeong Jun Koh. Luminance-aware color transform for multiple exposure correction. In ICCV, 2023.", + "[4] Wenyan Cong, Jianfu Zhang, Li Niu, Liu Liu, Zhixin Ling, Weiyuan Li, and Liqing Zhang. Dovenet: Deep image harmonization via domain verification. In CVPR, 2020.", + "[5] Yi-Hsuan Tsai, Xiaohui Shen, Zhe Lin, Kalyan Sunkavalli, Xin Lu, and Ming-Hsuan Yang. Deep image harmonization, 2017.", + "[6] Zonghui Guo, Haiyong Zheng, Yufeng Jiang, Zhaorui Gu, and Bing Zheng. Intrinsic image harmonization. In CVPR, 2021.", + "[7] Jun Ling, Han Xue, Li Song, Rong Xie, and Xiao Gu. Region-aware adaptive instance normalization for image harmonization. In CVPR, 2021.", + "[8] Yifan Jiang, He Zhang, Jianming Zhang, Yilin Wang, Zhe Lin, Kalyan Sunkavalli, Simon Chen, Sohrab Amirghodsi, Sarah Kong, and Zhangyang Wang. Ssh: A self-supervised framework for image harmonization. In ICCV, 2021.", + "[9] Wenyan Cong, Xinhao Tao, Li Niu, Jing Liang, Xuesong Gao, Qihao Sun, and Liqing Zhang. High-resolution image harmonization via collaborative dual transformations. 2022.", + "[10] Jeya Maria Jose Valanarasu, He Zhang, Jianming Zhang, Yilin Wang, Zhe Lin, Jose Echevarria, Yinglan Ma, Zijun Wei, Kalyan Sunkavalli, and Vishal M Patel. Interactive portrait harmonization. 2022.", + "[11] Lingzhi Zhang, Tarmily Wen, and Jianbo Shi. Deep image blending. In The IEEE Winter Conference on Applications of Computer Vision, pages 231-240, 2020.", + "[12] Dave Epstein, Allan Jabri, Ben Poole, Alexei A. Efros, and Aleksander Holynski. Diffusion self-guidance for controllable image generation. arXiv preprint arXiv:2306.00986, 2023. 1", + "[13] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. 2021. 1, 3", + "[14] Tim Brooks, Aleksander Holynski, and Alexei A Efros. Instructpix2pix: Learning to follow image editing instructions. 2022. 1, 2, 3, 6", + "[15] Chenlin Meng, Yutong He, Yang Song, Jiaming Song, Jia-jun Wu, Jun-Yan Zhu, and Stefano Ermon. SDEdit: Guided image synthesis and editing with stochastic differential equations. In ICLR, 2022. 1, 3", + "[16] Yujun Shi, Chuhui Xue, Jiachun Pan, Wenqing Zhang, Vincent YF Tan, and Song Bai. Dragdiffusion: Harnessing diffusion models for interactive point-based image editing. 2023. 1, 3, 5, 6", + "[17] Xingang Pan, Ayush Tewari, Thomas Leimkuhler, Lingjie Liu, Abhimitra Meka, and Christian Theobalt. Drag your" + ], + "bbox": [ + 78, + 114, + 468, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "gan: Interactive point-based manipulation on the generative image manifold. In SIGGRAPH, 2023. 1, 2, 3, 5, 6, 8", + "[18] Guillaume Couairon, Jakob Verbeek, Holger Schwenk, and Matthieu Cord. Diffedit: Diffusion-based semantic image editing with mask guidance, 2022.", + "[19] Chong Mou, Xintao Wang, Jiechong Song, Ying Shan, and Jian Zhang. Dragondiffusion: Enabling drag-style manipulation on diffusion models. 2023. 1, 2, 3, 6", + "[20] Binxin Yang, Shuyang Gu, Bo Zhang, Ting Zhang, Xuejin Chen, Xiaoyan Sun, Dong Chen, and Fang Wen. Paint by example: Exemplar-based image editing with diffusion models. arXiv preprint arXiv:2211.13227, 2022.", + "[21] Adding Conditional Control to Text-to-Image Diffusion Models, 2023. 3", + "[22] GLIGEN: Open-Set Grounded Text-to-Image Generation, 2023. arXiv:2301.07093. 1", + "[23] Or Patashnik, Zongze Wu, Eli Shechtman, Daniel Cohen-Or, and Dani Lischinski. Styleclip: Text-driven manipulation of stylegan imagery. In ICCV, 2021. 1, 2, 3, 6", + "[24] Bahjat Kawar, Shiran Zada, Oran Lang, Omer Tov, Huiwen Chang, Tali Dekel, Inbar Mosseri, and Michal Irani. Imagic: Text-based real image editing with diffusion models. In CVPR, 2023. 1, 3", + "[25] Yuki Endo. User-controllable latent transformer for stylegan image layout editing. Computer Graphics Forum, 2022. 1, 2, 3, 5, 6", + "[26] Erik Härkönen, Aaron Hertzmann, Jaakko Lehtinen, and Sylvain Paris. Ganspace: Discovering interpretable gan controls. In NeurIPS, 2020. 2, 3, 4", + "[27] Yujun Shen, Jinjin Gu, Xiaou Tang, and Bolei Zhou. Interpreting the latent space of gans for semantic face editing. In CVPR, 2020.", + "[28] Yujun Shen and Bolei Zhou. Closed-form factorization of latent semantics in gans. In CVPR, 2021. 3, 4", + "[29] Yotam Nitzan, Rinon Gal, Ofir Brenner, and Daniel Cohen-Or. Large: Latent-based regression through gan semantics. In CVPR, 2022. 2, 4", + "[30] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. 2019. 2, 4, 5", + "[31] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of StyleGAN. In CVPR, 2020. 2, 3, 4", + "[32] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. In NeurIPS, 2014. 2", + "[33] Yann LeCun. The mnist database of handwritten digits. http://yann.lecun.com/exdb/mnist/, 1998. 2", + "[34] Andrew Brock, Jeff Donahue, and Karen Simonyan. Large scale gan training for high fidelity natural image synthesis. arXiv, 2018. 2", + "[35] Terro Karras, Timo Aila, Samuli Laine, and Jaakko Lehtinen. Progressive growing of gans for improved quality, stability, and variation. In ICLR, 2018. 2", + "[36] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR, 2009. 2" + ], + "bbox": [ + 503, + 92, + 890, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "8279", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[37] Fisher Yu, Yinda Zhang, Shuran Song, Ari Seff, and Jianxiong Xiao. Lsun: Construction of a large-scale image dataset using deep learning with humans in the loop. 2015. 2", + "[38] Ali Jahanian, Lucy Chai, and Phillip Isola. On the \"steerability\" of generative adversarial networks. In ICLR, 2020. 2", + "[39] David Bau, Jun-Yan Zhu, Hendrik Strobelt, Agata Lapedriza, Bolei Zhou, and Antonio Torralba. Understanding the role of individual units in a deep neural network. Proceedings of the National Academy of Sciences, 2020. 2", + "[40] Zongze Wu, Dani Lischinski, and Eli Shechtman. Stylespace analysis: Disentangled controls for stylegan image generation. In ICLR, 2020. 3", + "[41] Tengfei Wang, Yong Zhang, Yanbo Fan, Jue Wang, and Qifeng Chen. High-fidelity gan inversion for image attribute editing. In CVPR, 2022. 3", + "[42] Omer Tov, Yuval Alaluf, Yotam Nitzan, Or Patashnik, and Daniel Cohen-Or. Designing an encoder for stylegan image manipulation. In ACM Transactions on Graphics, 2021. 6", + "[43] Daniel Roich, Ron Mokady, Amit H Bermano, and Daniel Cohen-Or. Pivotal tuning for latent-based editing of real images. ACM Trans. Graph., 2021. 3", + "[44] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. 2021. 3", + "[45] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. 2021. 3", + "[46] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In H. Larochelle, M. Ranzato, R. Hadsell, M.F. Balcan, and H. Lin, editors, Advances in Neural Information Processing Systems. Curran Associates, Inc., 2020. 3", + "[47] Ron Mokady, Amir Hertz, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Null-text inversion for editing real images using guided diffusion models. arXiv preprint arXiv:2211.09794, 2022. 3", + "[48] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-prompt image editing with cross attention control. arXiv preprint arXiv:2208.01626, 2022. 3", + "[49] Huan Ling, Karsten Kreis, Daiqing Li, Seung Wook Kim, Antonio Torralba, and Sanja Fidler. Editgan: High-precision semantic image editing. In Advances in Neural Information Processing Systems (NeurIPS), 2021. 3", + "[50] David Bau, Steven Liu, Tongzhou Wang, Jun-Yan Zhu, and Antonio Torralba. Rewriting a deep generative model. In ECCV, 2020. 3", + "[51] Amir Bar, Yossi Gandelsman, Trevor Darrell, Amir Globerson, and Alexei A. Efros. Visual prompting via image inpainting. 2022. 3", + "[52] Thao Nguyen, Yuheng Li, Utkarsh Ojha, and Yong Jae Lee. Visual instruction inversion: Image editing via visual prompting. In NeurIPS, 2023." + ], + "bbox": [ + 78, + 90, + 470, + 901 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[53] Yasheng Sun, Yifan Yang, Houwen Peng, Yifei Shen, Yuqing Yang, Han Hu, Lili Qiu, and Hideki Koike. Imagebrush: Learning visual in-context instructions for exemplar-based image manipulation. 2023.", + "[54] Aaron Hertzmann, Charles E. Jacobs, Nuria Oliver, Brian Curless, and David H. Salesin. Image analogies. In Seminal Graphics Papers: Pushing the Boundaries, Volume 2, 2023.", + "[55] Jing Liao, Yuan Yao, Lu Yuan, Gang Hua, and Sing Bing Kang. Visual attribute transfer through deep image analogy. 2017. 3", + "[56] Weihao Xia, Yulun Zhang, Yujiu Yang, Jing-Hao Xue, Bolei Zhou, and Ming-Hsuan Yang. Gan inversion: A survey. TPAMI, 2022. 3", + "[57] Yunjey Choi, Youngjung Uh, Jaejun Yoo, and Jung-Woo Ha. Starganv2: Diverse image synthesis for multiple domains. In CVPR, 2020. 5", + "[58] Tero Karras, Miika Aittala, Janne Hellsten, Samuli Laine, Jaakko Lehtinen, and Timo Aila. Training generative adversarial networks with limited data, 2020. 5", + "[59] Jianglin Fu, Shikai Li, Yuming Jiang, Kwan-Yee Lin, Chen Qian, Chen-Change Loy, Wayne Wu, and Ziwei Liu. Stylegan-human: A data-centric odyssey of human generation. arXiv preprint, arXiv:2204.11823, 2022. 5", + "[60] Davis E. King. Dlib-ml: A machine learning toolkit. J. Mach. Learn. Res., 2009. 6", + "[61] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, Gunter Klambauer, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. In NeurIPS, 2017. 7", + "[62] Thorsten Hempel, Ahmed A. Abdelrahman, and Ayoub Al-Hamadi. 6d rotation representation for unconstrained head pose estimation. In ICIP, 2022. 7", + "[63] William Peebles, Jun-Yan Zhu, Richard Zhang, Antonio Torralba, Alexei Efros, and Eli Shechtman. Gan-supervised dense visual alignment. In CVPR, 2022. 8" + ], + "bbox": [ + 501, + 90, + 892, + 588 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "8280", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2024/Edit One for All_ Interactive Batch Image Editing/d0962876-52e1-4fef-ac92-9ddbf4fa5230_model.json b/2024/Edit One for All_ Interactive Batch Image Editing/d0962876-52e1-4fef-ac92-9ddbf4fa5230_model.json new file mode 100644 index 0000000000000000000000000000000000000000..a8c5a96b83f552c0b7cb02fe87c2d04ac4ce7b4a --- /dev/null +++ b/2024/Edit One for All_ Interactive Batch Image Editing/d0962876-52e1-4fef-ac92-9ddbf4fa5230_model.json @@ -0,0 +1,2266 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.808, + 0.045 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.234, + 0.131, + 0.738, + 0.154 + ], + "angle": 0, + "content": "Edit One for All: Interactive Batch Image Editing" + }, + { + "type": "text", + "bbox": [ + 0.188, + 0.181, + 0.782, + 0.217 + ], + "angle": 0, + "content": "Thao Nguyen Utkarsh Ojha Yuheng Li Haotian Liu Yong Jae Lee University of Wisconsin-Madison" + }, + { + "type": "text", + "bbox": [ + 0.283, + 0.22, + 0.678, + 0.233 + ], + "angle": 0, + "content": "https://thaoshibe.github.io/edit-one-for-all" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.269, + 0.314, + 0.285 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.295, + 0.474, + 0.583 + ], + "angle": 0, + "content": "In recent years, image editing has advanced remarkably. With increased human control, it is now possible to edit an image in a plethora of ways; from specifying in text what we want to change, to straight up dragging the contents of the image in an interactive point-based manner. However, most of the focus has remained on editing single images at a time. Whether and how we can simultaneously edit large batches of images has remained understudied. With the goal of minimizing human supervision in the editing process, this paper presents a novel method for interactive batch image editing using StyleGAN as the medium. Given an edit specified by users in an example image (e.g., make the face frontal), our method can automatically transfer that edit to other test images, so that regardless of their initial state (pose), they all arrive at the same final state (e.g., all facing front). Extensive experiments demonstrate that edits performed using our method have similar visual quality to existing single-image-editing methods, while having more visual consistency and saving significant time and human effort." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.604, + 0.21, + 0.62 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.629, + 0.47, + 0.84 + ], + "angle": 0, + "content": "Image editing has undergone a transformation in recent years with the aid of modern generative models. The process has been greatly democratized, where many of the sophisticated edits, which previously might have required hours and niche expertise, can now be completed with relative ease in a matter of minutes. For example, different types of learning based algorithms can be used for image correction/adjustment [1-12] and for manipulating the semantic contents of real images [13-22]. Moreover, there are many different ways to edit an image. For instance, a user can specify in text what changes they want - e.g., \"make the hair darker\" [14, 15, 23, 24], or they can drag the contents of the image in an interactive manner to shrink, enlarge, or move a part of an object [12, 16, 17, 19, 25]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.47, + 0.902 + ], + "angle": 0, + "content": "A common theme across many of these works is that the edits are designed to work for a single image at a time. For example, given an image of a cat with open eyes, DragGAN [17] allows the user to drag the contents of that par" + }, + { + "type": "image", + "bbox": [ + 0.507, + 0.268, + 0.885, + 0.562 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.566, + 0.892, + 0.622 + ], + "angle": 0, + "content": "Figure 1. Interactive Batch Image Editing. Given a single user edited image, the goal is to automatically transfer that edit to new unseen images, so that all edited images end up with the same final state as the user edited image." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.642, + 0.892, + 0.748 + ], + "angle": 0, + "content": "ticular image so that both the eyes can be closed in the resulting cat image. But what if we wish to apply the same edit to many different types of cats so that we can close all of their eyes? While one could perform dragging on each cat separately, the whole process would be quite cumbersome (requiring lots of human annotation) and time intensive (requiring optimization for each image)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.893, + 0.901 + ], + "angle": 0, + "content": "In this paper, we introduce the new problem of Interactive Batch Image Editing. Given a user-specified edit \\( I \\rightarrow I' \\), the goal is to automatically transfer that edit to new unseen images, so that all edited images end up with the same final state as \\( I' \\) (e.g., all cats with eyes closed) regardless of their original starting states (e.g., varying degrees of original eye openness); see Figure 1. Hence, a solution to this problem requires two components: (i) modeling the user edit in the example pair \\( (I, I') \\) so that it can be transferred to new images; and (ii) controlling the degree of the" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "8271" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.197, + 0.089, + 0.352, + 0.101 + ], + "angle": 0, + "content": "edit one image at a time" + }, + { + "type": "image", + "bbox": [ + 0.093, + 0.106, + 0.449, + 0.26 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.514, + 0.089, + 0.835, + 0.102 + ], + "angle": 0, + "content": "edit one image, transfer to others at the same time!" + }, + { + "type": "image", + "bbox": [ + 0.467, + 0.104, + 0.875, + 0.259 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.261, + 0.894, + 0.304 + ], + "angle": 0, + "content": "Figure 2. Single Image Editing vs. Batch Image Editing. (a) Prior work (e.g., [17, 19, 25]) focuses on single image editing. (b) We focus on batch image editing, where the user's edit on a single image is automatically transferred to new images, so that they all arrive at the same final state regardless of their initial starting state. In this way, we can achieve time speed up and reduce human effort in editing." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.319, + 0.436, + 0.334 + ], + "angle": 0, + "content": "edit so that all edited images have the same final state." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.337, + 0.47, + 0.563 + ], + "angle": 0, + "content": "To model the user's edit, we build upon existing literature on Generative Adversarial Networks (GANs) [23, 26-29], and in particular StyleGAN [30, 31], which shows that their learned latent space emits globally linear editing directions where the edited attribute (e.g., eye closeness) varies linearly in magnitude along such a direction for all semantically related instances (e.g., all closed eyes cat). To try to discover the global direction corresponding to a user edit, we perform optimization in the latent space of StyleGAN2 [31] so that (i) the edited image with the discovered direction is visually similar to the original edited image \\( I' \\), and (ii) the linear correlation between the distance along the direction and the strength of the visual attribute is highest (e.g., going twice the amount in that direction closes a cat's eyes by twice the amount)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.566, + 0.47, + 0.808 + ], + "angle": 0, + "content": "To control the degree of the edit so that all edited images end up with the same final state, we derive a closed-form solution. Taking motivation from [29], we model the attribute strength (e.g., how much the cat's eyes are open) as the distance along the normal vector from a hyperplane. By setting the example edited image as lying on that hyperplane (i.e., its distance is 0), the objective is to move all other images along that same direction so that their distances to the hyperplane also become 0. In this way, regardless of the start state of any image (e.g., varying degree of eye openness), their edited end states all become the same (e.g., closed eyes). We show that we can analytically compute the exact amount of traversal along the direction for any image with a closed-form solution. Editing images in this way helps improve the results visually, as the final state of any new image matches that of the edited example given by the user." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.811, + 0.47, + 0.903 + ], + "angle": 0, + "content": "Our method works with various edits given by an editing framework, e.g., it can be a change given by a dragging operation [17], or it can be a change given by a text-based edit [14, 23] (e.g., \"make eyes bigger\"), as long as it is a direction compatible with a StyleGAN2 [31] model. We present qualitative and" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.319, + 0.892, + 0.53 + ], + "angle": 0, + "content": "quantitative results in transferring edits for a variety of objects (cats/dogs/faces/humans/limons/arts/etc.), parts (mouths/ears/legs/etc.), and corresponding attributes (big eyes/short faces/pale skins/etc.). Importantly, we show that the final state of the edited images is comparable to the scenario in which a user performs the edit for each image separately. But since our method does this automatically (see Fig. 2), we show that it takes much less time (e.g., only 0.05s per image, compared to DragGAN [17] about 2s per image), and does not require laborious human annotation (e.g., only need 1 annotated image, compared to 4 required for DragGAN [17]). Finally, we show practical applications of batch image editing; e.g., changing wheel size in every photo of a car collection by editing just one." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.531, + 0.893, + 0.682 + ], + "angle": 0, + "content": "In sum, our contributions are: (1) We introduce the novel problem of interactive batch image editing, wherein a user-driven edit is automatically transferred to other similar test images. (2) We study what having the same final state means in a geometrical sense, and propose a principled approach to achieving that in StyleGAN2 [31]'s latent space. (3) We show that our method works on a wide variety of domains e.g., cats, dogs, humans etc., taking significantly less time than editing each image individually while having similar visual quality and more visual consistency." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.695, + 0.642, + 0.71 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Latent space of GANs. The capabilities of generative adversarial networks (GANs) have transformed drastically, from the first GAN [32] designed for simple image datasets like MNIST [33] to big and powerful models like BigGAN [34], Progressive GAN [35] and StyleGANs [30, 31] designed for much more complex datasets [36, 37]. Parallel to the efforts to enhance their capabilities, there has also been work done to better understand and make use of their learned latent space. Specifically, there is a line of work which tries to find interpretable directions in that space, so that moving in such a direction majorly changes only one discernible attribute in the image. Some methods [38, 39]" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8272" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.081, + 0.089, + 0.255, + 0.27 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.134, + 0.272, + 0.201, + 0.284 + ], + "angle": 0, + "content": "(a) Setting" + }, + { + "type": "image", + "bbox": [ + 0.271, + 0.088, + 0.406, + 0.27 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.272, + 0.272, + 0.395, + 0.284 + ], + "angle": 0, + "content": "(b) Naïve Approach" + }, + { + "type": "image", + "bbox": [ + 0.422, + 0.089, + 0.62, + 0.27 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.42, + 0.272, + 0.615, + 0.285 + ], + "angle": 0, + "content": "(c) Optimizing Editing Direction" + }, + { + "type": "image", + "bbox": [ + 0.627, + 0.088, + 0.892, + 0.27 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.674, + 0.272, + 0.838, + 0.285 + ], + "angle": 0, + "content": "(d) Adjusting Edit Strength" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.289, + 0.894, + 0.332 + ], + "angle": 0, + "content": "Figure 3. Different editing strategies. (a) Setting. (b) Naive Approach: The editing direction effective for an example may not generalize well to test images. (c) Optimizing Editing Direction: We optimize for a globally consistent direction that is effective for both example and test images. (d) Adjusting Editing Strength: Ensuring consistent final states requires adjusting the editing strength for each test image." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.347, + 0.473, + 0.558 + ], + "angle": 0, + "content": "try to find such directions or particular activations using (self) supervised methods. Others [26-28] try getting rid of the need for supervision: discovering the directions representing the most important factors of variation, whatever they may be. Going beyond the \\(\\mathcal{W}\\) space, the authors in [40] explore the \\(\\mathcal{W}+\\) space, and show that it has even better disentanglement useful for spatial image editing. All of these hidden capabilities have made StyleGANs a useful tool for editing purposes. However, prior work mainly focuses on the effectiveness of the editing direction (e.g., whether the discovered direction can change yaw pose). The question of whether these directions can be applied to diverse images to yield consistent results (e.g., all faces facing frontal) and, if so, how to achieving such consistency, remains unanswered." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.57, + 0.473, + 0.902 + ], + "angle": 0, + "content": "Image editing with generative models. Researchers have used StyleGAN for real image editing by designing an encoder to invert a real image into StyleGAN's latent space [41-43]. StyleCLIP [23] presents a way to perform image editing through a text-based interface by making use of the CLIP encoder [44]. Very recently, StyleGAN has also been used to perform point-based editing [17, 25] so as to move any start point to reach a target location in the image; thereby elongating, rotating, shifting the objects. Most work on GANs has been for single object category datasets, but with the rise of text-to-image Diffusion Models which can generate complex images [13, 45, 46], editing such images is now possible. Complex scenes can be edited either using text [14, 15, 21, 24, 47, 48], or using the same idea of point-based manipulation [16, 19]. There have been some works which discuss the possibility to transfer the edit of one image to another. EditGAN [49] uses segmentation mask manipulations to edit an image, but to successfully transfer them to an unseen image, it needs to do post hoc manipulation of the editing scale. RewriteGAN [50] involves editing one generated sample; e.g., adding a patch of a tree onto a church tower. Following this, the rules of the" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.347, + 0.895, + 0.483 + ], + "angle": 0, + "content": "GAN are manipulated so that all churches have some tree on the top. Within diffusion models, visual prompting/ image analogies tries something similar, where users can define a triplet {before, after, test} to learn the edit and transfer it to a test image [51-55]. Our task is similar, in that we wish to transfer user-edits from the training example to new images, but it differs in one crucial aspect: along with transferring the edit, we wish to automatically learn its strength so that the edit produces the same final state for a new image." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.496, + 0.609, + 0.514 + ], + "angle": 0, + "content": "3. Approach" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.522, + 0.894, + 0.596 + ], + "angle": 0, + "content": "Our focus is on the setting in which an image editing process edits an attribute of an image so that the attribute's value reaches a desired state; e.g., rotating the pose of a face looking sideways so that it faces front (attribute = face's pose, value = front)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.597, + 0.892, + 0.749 + ], + "angle": 0, + "content": "With this view, we explain our framework for batch image editing, which can be broken down into two stages: (i) A user edits an image \\( I_0 \\) (e.g., using DragGAN [17]) to obtain an edited image \\( I_0' \\). We describe in Sec. 3.1 how we capture this user edit \\( I_0 \\rightarrow I_0' \\), so that the same edit can be applied to new images. (ii) Next, we describe in Sec. 3.2, how for any new image, we apply the modeled edit by automatically adjusting its strength so that the attribute's value in this new image matches that of \\( I_0' \\) (e.g., any face, regardless of its initial pose, now faces front after the edit)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.757, + 0.718, + 0.773 + ], + "angle": 0, + "content": "3.1. Modeling the User Edit" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.78, + 0.892, + 0.903 + ], + "angle": 0, + "content": "We start with an image \\( I_0 \\). This image could be a real image or a generated one. Either way, we get its latent representation in the \\( \\mathcal{W} \\) space of a StyleGAN2 [31] model \\( G \\), so that \\( I_0 = G(w_0) \\). (For real images we can use GAN-inversion techniques [56].) The user, with the help of an image editing framework (e.g., DragGAN [17] or InstructPix2Pix [14]), edits \\( I_0 \\) to manipulate one or more of its attributes. The edit maps to the \\( \\mathcal{W} \\) space as \\( w_0 \\to w_0' \\). The" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8273" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.151 + ], + "angle": 0, + "content": "resulting edited image can thus be recovered as follows: \\( I_0' = G(w_0') \\). Fig. 3(a) shows an example of the original and edited image pair, \\( (I_0, I_0') \\), where the user intended to turn the face forward." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.152, + 0.47, + 0.349 + ], + "angle": 0, + "content": "Now, given the user edit \\( I_0 \\to I_0' \\), we wish to capture it in a way that can be applied to new images in a generalizable manner; i.e., the application of the edit changes the same property in a new image. This is where a nice property of GANs, and in particular StyleGANs [30, 31] comes in useful. It has been shown in prior works [26-28] that it is possible to discover directions with such properties (using supervised as well as unsupervised methods) in the learned \\( \\mathcal{W} \\) space. In particular, it is possible to find directions \\( (\\Delta_w) \\) that are globally consistent. Taking motivation from [29], we define a globally consistent direction \\( \\Delta_w \\) as the following: for any arbitrary \\( w \\), moving along \\( \\Delta_w \\), \\( w \\to w + \\Delta_w \\) (i) changes the same attribute, and (ii) by the same amount." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.349, + 0.469, + 0.499 + ], + "angle": 0, + "content": "To make the precise user edit \\( I_0 \\to I_0' \\) applicable to other images, it needs to be captured as a globally consistent direction. The naive way to represent that edit will be through a simple difference in the \\( \\mathcal{W} \\) space: \\( \\Delta_w = w_0' - w_0 \\). However, empirically, and as we show in Fig. 3(b), applying this \\( \\Delta_w \\) to the latent code \\( w_1 \\) corresponding to a new image \\( I_1 \\) does not always result in the same change; while \\( I_0 \\to I_0' \\) results in a pose change of \\( \\sim 30^\\circ \\) degrees in yaw, \\( I_1 \\to I_1' \\) does not seem to change the same attribute or at least not by the same amount." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.5, + 0.469, + 0.575 + ], + "angle": 0, + "content": "Hence, our goal is to represent the \\( I_0 \\to I_0' \\) edit through a \\( \\mathcal{W} \\) space direction that can better satisfy both the properties of a globally consistent direction. For this task, we first take motivation from LARGE [29] to introduce a mathematical view of what those directions mean." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.584, + 0.469, + 0.629 + ], + "angle": 0, + "content": "Globally consistent direction. Let's say that \\(\\Delta_g\\) is one such direction. Along with a bias term \\(b\\), we can define a hyperplane as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.22, + 0.635, + 0.469, + 0.651 + ], + "angle": 0, + "content": "\\[\nw \\cdot \\Delta_ {g} + b = 0 \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.663, + 0.469, + 0.754 + ], + "angle": 0, + "content": "That is, any point \\( w \\) which lies on the hyperplane will satisfy this equation. The authors in [29] argued that for such a hyperplane, whose normal vector is a globally consistent direction \\( (\\Delta_g) \\), the distance of an arbitrary point \\( w \\) from that hyperplane \\( (w \\cdot \\Delta_g) \\) will be linearly correlated to the actual attribute \\( (y) \\) that results in the generated image:" + }, + { + "type": "equation", + "bbox": [ + 0.198, + 0.767, + 0.469, + 0.783 + ], + "angle": 0, + "content": "\\[\ny = a \\times (w \\cdot \\Delta_ {g}) + b \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.795, + 0.469, + 0.902 + ], + "angle": 0, + "content": "Here, \\( a \\) and \\( b \\) are unknown linear coefficients. For example, if \\( \\Delta_g \\) corresponds to change in pose, then all the front facing people will lie at the same distance \\( d \\) from the above hyperplane. Because of this, any hyperplane defined with respect to a globally consistent direction can be viewed as a semantic hyperplane. For simplicity, we can set \\( d = 0 \\) for the edited image \\( I_0' \\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.228 + ], + "angle": 0, + "content": "Given that the original \\(\\Delta_w\\) may not always be globally consistent, we aim to discover, through optimization, a different direction \\(\\Delta_w^*\\) which produces a similar editing effect, but is more likely to be a globally consistent one. We initialize \\(\\Delta_w^*\\) with 0's, and design two objective functions to optimize it. First, to make sure that it produces a similar effect as the original direction, we use an image reconstruction loss so that the edited image produced by the new edit matches the original edit given by the user:" + }, + { + "type": "equation", + "bbox": [ + 0.555, + 0.235, + 0.892, + 0.253 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {i m g} = \\left\\| G \\left(w _ {0} + \\Delta_ {w}\\right) - G \\left(w _ {0} + \\Delta_ {w} ^ {*}\\right) \\right\\| _ {2} \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.262, + 0.892, + 0.413 + ], + "angle": 0, + "content": "Additionally, user can provide the real value of distance \\( d \\), otherwise we set the distance of the edited image from the hypothetical hyperplane to be 0 (similar to [29]). We constrain the learned \\( \\Delta_w^* \\) to follow this property for better interpretability. That is, the new direction should be such that when we traverse the original latent code in that direction, \\( w_0 \\rightarrow w_0 + \\Delta_w^* \\), the resulting image should lie at the hyperplane defined by that direction. Setting \\( b = 0 \\) in Eq. 1, we minimize the edited point's distance from the hyperplane:" + }, + { + "type": "equation", + "bbox": [ + 0.607, + 0.418, + 0.892, + 0.434 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {a t t} = \\left| \\left(w _ {0} + \\Delta_ {w} ^ {*}\\right) \\cdot \\Delta_ {w} ^ {*} \\right| \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.445, + 0.892, + 0.505 + ], + "angle": 0, + "content": "The overall loss function for optimizing the new direction \\(\\Delta_w^*\\) is a weighted sum of the two objectives: \\(\\mathcal{L}_{\\Delta} = \\mathcal{L}_{img} + \\lambda \\mathcal{L}_{att}\\). We set \\(\\lambda\\) to 0.02 in all experiments to balance the magnitude of \\(\\mathcal{L}_{img}\\) and \\(\\mathcal{L}_{att}\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.506, + 0.892, + 0.718 + ], + "angle": 0, + "content": "Now, in Fig. 3 (b) and (c), we compare the difference in image editing results using two different directions when applied to the same new image \\((I_1)\\): \\(w_1 \\to w_1 + \\Delta_w\\) vs. \\(w_1 \\to w_1 + \\Delta_w^*\\). As discussed before, in this case, the editing using the naively computed direction \\((\\Delta_w)\\) is not able to accurately capture the pose rotation the way it did for the user edited example \\((I_0 \\to I_0')\\). On the other hand, with \\(\\Delta_w^*\\), we see that the pose of the woman changes by a similar amount as the edited example. However, since the original pose of the woman \\((I_1)\\) was not the same as the pose of the man \\((I_0)\\), the new edited image \\((I_1')\\) still does not arrive at the same final state as \\(I_0'\\). Therefore, our next goal is to figure out how to scale the learned direction so that \\(I_1'\\) does arrive at the same final state as \\(I_0'\\)." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.726, + 0.869, + 0.742 + ], + "angle": 0, + "content": "3.2. Adjusting Editing Strength for New Images" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.749, + 0.89, + 0.779 + ], + "angle": 0, + "content": "Given the optimized \\(\\Delta_w^*\\) and some new image \\(I_{i} = G(w_{i})\\) we wish to edit it in the following way:" + }, + { + "type": "equation", + "bbox": [ + 0.643, + 0.784, + 0.891, + 0.799 + ], + "angle": 0, + "content": "\\[\nw _ {i} ^ {\\prime} = w _ {i} + \\alpha_ {i} n \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.811, + 0.892, + 0.901 + ], + "angle": 0, + "content": "where \\(\\alpha_{i}\\) is the editing strength computed separately for each image and \\(n = \\Delta_w^* /||\\Delta_w^* ||\\) is a unit vector in the direction of \\(\\Delta_w^*\\). Fig. 3 (d) illustrates a geometric perspective that we will use to compute \\(\\alpha_{i}\\). First, we see \\(\\Delta_w^*\\) represented as a vector and its corresponding hyperplane that is normal to it. Next, we depict the new latent point \\(w_{i}\\). The" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8274" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.089, + 0.48, + 0.336 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.095, + 0.337, + 0.147, + 0.349 + ], + "angle": 0, + "content": "Original" + }, + { + "type": "image_caption", + "bbox": [ + 0.164, + 0.337, + 0.235, + 0.349 + ], + "angle": 0, + "content": "UserID" + }, + { + "type": "image_caption", + "bbox": [ + 0.242, + 0.337, + 0.311, + 0.349 + ], + "angle": 0, + "content": "DragGAN" + }, + { + "type": "image_caption", + "bbox": [ + 0.318, + 0.337, + 0.386, + 0.349 + ], + "angle": 0, + "content": "DragDiff." + }, + { + "type": "image", + "bbox": [ + 0.397, + 0.089, + 0.477, + 0.337 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.418, + 0.338, + 0.451, + 0.348 + ], + "angle": 0, + "content": "Ours" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.353, + 0.892, + 0.381 + ], + "angle": 0, + "content": "Figure 4. Qualitative comparisons between dragging baselines. For ours, green bounding box indicates automatic transfer from the red bounding box example in the first row (i.e., no point annotation needed!)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.397, + 0.47, + 0.472 + ], + "angle": 0, + "content": "goal is to move it along the \\(\\Delta_w^*\\) direction so that it arrives onto the hyperplane. Through this depiction, \\(\\alpha_{i}\\) can be understood as the distance of \\(w_{1}\\) from the hyperplane. We can get the distance by projecting \\((w_0' - w_i)\\) in the direction of \\(n\\). Therefore, \\(\\alpha_{i} = (w_{0}^{\\prime} - w_{i})\\cdot n\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.474, + 0.47, + 0.578 + ], + "angle": 0, + "content": "Since for each image, the only unique computation that needs to be done is the calculation of \\(\\alpha_{i}\\), we can see why our method will be much faster than, for example, annotating every image and running the DragGAN [17] optimization each of those times. We will show in our experiments the significant difference in time taken by our method compared to single image editing baselines." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.58, + 0.468, + 0.745 + ], + "angle": 0, + "content": "Importantly, there is an added benefit to computing \\(\\alpha\\)'s in this way. Let's say the goal of the editing process is to rotate \\(n\\) faces and make them frontal. After completion, all of them do become frontal, each with their own editing strengths \\(\\{\\alpha_{1},\\alpha_{2},\\dots,\\alpha_{n}\\}\\) computed using the above formula. Suppose the user now wants the same faces facing a bit left instead. To do this, the user does not need to re-associate the original training example and run the optimization one more time. In an interactive manner, they can simply scale the \\(\\alpha\\) for the training example to match the desired edit, and all other \\(\\alpha\\)'s can be automatically recomputed." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.761, + 0.208, + 0.779 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.786, + 0.469, + 0.832 + ], + "angle": 0, + "content": "We study how well our method models and transfers the edits from an example, and how efficient it is compared to single image editing baselines using their official code." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.469, + 0.902 + ], + "angle": 0, + "content": "Categories. We evaluate on a variety of domains: Human faces (FFHQ) [30], AFHQ Cats, Dogs [57], MetFaces [58], Human bodies [59]. For each domain, we use the corresponding pretrained StyleGAN model to perform editing." + }, + { + "type": "image", + "bbox": [ + 0.493, + 0.089, + 0.888, + 0.336 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.507, + 0.337, + 0.558, + 0.349 + ], + "angle": 0, + "content": "Original" + }, + { + "type": "image_caption", + "bbox": [ + 0.576, + 0.337, + 0.647, + 0.349 + ], + "angle": 0, + "content": "UserID" + }, + { + "type": "image_caption", + "bbox": [ + 0.655, + 0.337, + 0.723, + 0.349 + ], + "angle": 0, + "content": "DragGAN" + }, + { + "type": "image_caption", + "bbox": [ + 0.738, + 0.337, + 0.796, + 0.349 + ], + "angle": 0, + "content": "DragDiff." + }, + { + "type": "image_caption", + "bbox": [ + 0.805, + 0.337, + 0.864, + 0.348 + ], + "angle": 0, + "content": "Ours" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.397, + 0.677, + 0.412 + ], + "angle": 0, + "content": "4.1. Qualitative results" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.421, + 0.891, + 0.452 + ], + "angle": 0, + "content": "We start with qualitative comparisons with (i) interactive point-dragging, and (ii) text-based image editing baselines." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.462, + 0.892, + 0.658 + ], + "angle": 0, + "content": "Interactive point based editing. We compare against DragGAN [17], UserControllableLT [25], and DragDiffusion [16]. For each baseline, we, as a user, desire to edit a bunch of images to have the same final state for an attribute. Fig. 4 shows the results of editing two kinds of edits to four images (leftmost column). In the left case, the goal is to make everyone smile by the same degree. In the right, it is to vertically compress everyone's face by the same degree. For all three baselines, we manually annotate points for each test image. For our method, we only annotate points and perform DragGAN-based edit on one image (top row), then automatically transfer the edit to the three other test images." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.66, + 0.893, + 0.901 + ], + "angle": 0, + "content": "For the smiling case (left), we notice that when DragGAN drags the upper/lower lips up/down respectively, it can either make people smile (1st, 2nd and 4th rows), or it can make people look shocked (blonde girl; 3rd row). Both edits are technically correct based on lips movement, but it won't match the user expectation of making everyone smile by the same degree. Since all four images are edited independently, this is not unexpected. The other two baselines sometimes have issues introducing the same edit across different images; e.g., UserControllableLT can make the blonde girl smile a little bit, but not the other images. In general, we notice in our experiments that to introduce the same exact edit that we want, we often need to play around with some hyperparameters (e.g., # of iterations). In comparison, our method produces edited images that are smiling more equally (left) and have been compressed by a similar" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.515, + 0.957 + ], + "angle": 0, + "content": "8275" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.081, + 0.09, + 0.482, + 0.359 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.493, + 0.091, + 0.892, + 0.359 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.096, + 0.361, + 0.871, + 0.375 + ], + "angle": 0, + "content": "Figure 5. Qualitative comparisons to text-guided baselines. Ours transfers the edit from example (red), to other test images (green)." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.392, + 0.405, + 0.406 + ], + "angle": 0, + "content": "amount (right), and with much less human effort." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.416, + 0.469, + 0.597 + ], + "angle": 0, + "content": "Text-driven image editing. We consider two baselines: StyleCLIP [23] and InstructPix2Pix [14]. Each takes an image and text as input to produce an edited image: \\( I \\rightarrow I' \\). Fig. 5 shows the results of Ours vs. StyleCLIP (left) and Ours vs. InstructPix2Pix (right). (For the latter, we invert the output of InstructPix2Pix into StyleGAN's latent space using [42] for our method.) In each case, for the baselines, we use the same text prompt, e.g., \"A photo of a person with pale skin\" for each edited image (in four rows). For ours, we capture the edit from the first editing result of a baseline (first row), and then automatically transfer the edit to the remaining three images." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.598, + 0.47, + 0.854 + ], + "angle": 0, + "content": "Our goal with this particular setup is to test (i) how consistently the baselines introduce the edit denoted by text to different images (e.g., do all people become equally pale?), and (ii) how consistently our method captures and transfers the particular edit of the first example to the rest of the images, irrespective of how good that first edit was. For (i), we find that StyleCLIP sometimes has issues; e.g., the paleness of the edited face in first example is different from the second (first vs second rows). For InstructPix2Pix, sometimes it can give consistent results; the childness of different faces seems similar. But, it can have consistency issues in other examples; e.g., the type/strength of maleness introduced is different in each image. And this is where, we believe, the utility of our method lies: if one desires to edit every face to be male in a particular way (i.e., to the same degree) as the first example, e.g., thick eyebrows & light beard, our method has an advantage." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.856, + 0.47, + 0.902 + ], + "angle": 0, + "content": "Finally, we show our method's results for non-facial domains in Fig. 6. For the lion examples (top left), we see that it can preserve not just the type of edit, e.g., dragging" + }, + { + "type": "table", + "bbox": [ + 0.5, + 0.388, + 0.89, + 0.478 + ], + "angle": 0, + "content": "
MethodPoint DraggingTimeAnno. #imgs
Dist.FID1 img1k imgs
UserControl. [25]13.6425.320.03s30s1000
DragGAN [17]4.1659.282s33.33m1000
DragDiffusion [16]26.5636.5560s16.67h1000
Ours9.4679.352s82s1
" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.488, + 0.892, + 0.558 + ], + "angle": 0, + "content": "Table 1. Time is estimated for 1 point drag, without human annotation time. (82s includes 2s to perform edit on the example, 30s to optimize the editing direction, and 50s to transfer the edit to 1000 test images.) Our method requires only one image annotation in total, while the baselines need one annotation per test image." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.563, + 0.893, + 0.7 + ], + "angle": 0, + "content": "to make the lion roar, but also the strength of the roar. The strength being preserved can be observed more easily for the human body poses (bottom left), where we see that the extent of legs split, hand movements, is consistent enough to almost align the edited test images with the user edited one. Overall, results on these diverse set of domains highlight an observation that many kinds of edits can be thought of as a combination of {type, strength}, both of which can indeed be captured and transferred according to our needs." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.711, + 0.688, + 0.726 + ], + "angle": 0, + "content": "4.2. Quantitative results" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.735, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Next, perform quantitative experiments following the setup proposed in [16, 17, 19]. We first randomly sample 10 facial test images and pair them with a randomly sampled target face image. We use dlib-ml [60] to detect keypoints in the test and target images. Our goal is to see if the keypoints in the test images can be moved to the target locations specified by keypoints in the target image. For the baselines [16, 17, 25], we perform dragging for each test image. For our method, we perform dragging on one test image, and then transfer the edit to the remaining nine images. This is repeated 100 times; i.e., each time a random target image is" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "8276" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.141, + 0.086, + 0.837, + 0.602 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.307, + 0.604, + 0.663, + 0.619 + ], + "angle": 0, + "content": "Figure 6. Additional qualitative results on various domains." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.627, + 0.314, + 0.643 + ], + "angle": 0, + "content": "paired with 10 random test images." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.644, + 0.47, + 0.765 + ], + "angle": 0, + "content": "We report the Euclidean distance between the keypoints of edited and target images in Table 1. Even without requiring annotations for every test image, our method can move the points very close to the target; almost as close as DragGAN, which requires annotation each time. We also compute FID [61] between the original test images and their edited versions to ensure that our method does not distort the image; and our quality is comparable to DragGAN." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.782, + 0.238, + 0.8 + ], + "angle": 0, + "content": "5. Deeper Analysis" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.809, + 0.47, + 0.884 + ], + "angle": 0, + "content": "We perform deeper analysis on the usefulness of our batch image editing framework by focusing on a specific edit: face pose rotation. In particular, if we wish to rotate many faces to front, how do we visualize the improvements brought by different components of our method?" + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.886, + 0.47, + 0.902 + ], + "angle": 0, + "content": "Fig. 7 depicts the setup. We use DragGAN to perform" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.627, + 0.895, + 0.9 + ], + "angle": 0, + "content": "dragging on one image (top; \\( w_0 \\rightarrow w_0' \\)). We then transfer the edit to 1000 other test images (two shown in rows 2-3). We see that naively applying \\( \\Delta_w = w_0' - w_0 \\) from DragGAN to test images cannot make them frontal (\\( 0^\\circ \\)). Using our optimized \\( \\Delta_w^* \\) (without dynamic scaling) does help it bring closer to facing front. Does this mean that \\( \\Delta_w^* \\) might be a more global direction? We study this in Fig. 8 (a), where we visualize the degree to which the correlation property of global directions (Eq. 2) holds for \\( \\Delta_w \\) and \\( \\Delta_w^* \\) individually. We see that there is indeed a better correlation (\\( R^2 \\)) between the distances of latents from hyperplanes and the Yaw degree of resulting images. (We predict Yaw using 6DRepNet [62].) However, as we discussed before, \\( \\Delta_w^* \\) in itself is insufficient: to completely bring the facial pose to front, we need to scale it with corresponding \\( \\alpha \\)'s to bring them much closer to front (Fig. 7, 4th column). To study this in a more systematic way, we visualize the effect on all 1000 test images. Similar to Fig. 8 (a), we visualize" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8277" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.084, + 0.089, + 0.464, + 0.335 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.078, + 0.337, + 0.468, + 0.365 + ], + "angle": 0, + "content": "Figure 7. Effect of adjusted and optimized editing for test images. Yaw degree of each image is provided on top of each images." + }, + { + "type": "table", + "bbox": [ + 0.08, + 0.375, + 0.468, + 0.445 + ], + "angle": 0, + "content": "
MethodMAETime complexity
PrepareInference
Random11.295 ± 8.972--
DragGAN [17] + GANgealing [63]8.141 ± 7.22130s2000s
Ours2.120 ± 1.81832s50s
" + }, + { + "type": "table_caption", + "bbox": [ + 0.118, + 0.446, + 0.428, + 0.459 + ], + "angle": 0, + "content": "Table 2. Ablation Study: DragGAN + GANgealing." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.469, + 0.469, + 0.529 + ], + "angle": 0, + "content": "distance-to-hyperplane vs. Yaw degree for the original test images (blue) and edited images (red) in Fig. 8 (b). We can see that the variation in Yaw for the edited images collapses at around \\(0^{\\circ}\\); i.e., they mostly face front as we would like." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.53, + 0.469, + 0.604 + ], + "angle": 0, + "content": "We also compute the time it takes to perform editing in Table 1, last three columns. Our method only requires annotating one image, while the baselines require annotating each image (e.g., 1000 total in this case). Note the times do not include human annotation time." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.614, + 0.469, + 0.765 + ], + "angle": 0, + "content": "Other ways to automate batch image editing. One baseline for batch image editing (i.e., so that all edited images achieve the same final state) is shown in Fig. 9. After the user annotates the source/target points in the example image, we use GANgealing [63] to transfer the points to corresponding locations in each new test image. The edited images are then obtained using DragGAN (DragGAN + GANgealing). We compare our method to this baseline when the user edits one face image to become front \\((\\mathrm{Yaw} = 0^{\\circ})\\) and transferring it to 1000 other test images." + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.766, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Results are shown in Table 2. We report mean absolute error (MAE) between the Yaw degree of the edited image and its ideal frontal image \\((\\mathrm{Yaw} = 0^{\\circ})\\). 'Random' shows the variations in Yaw for the original images (before editing). While 'DragGAN + GANgealing' does help in reducing the variation \\((8.14 < 11.29)\\), our edited images are much more frontal, with an MAE of 2.12. This is because the baseline has no way to bring all edited images to the same final state. On top of that, since our method does not rely on" + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.091, + 0.885, + 0.198 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.201, + 0.885, + 0.307 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.312, + 0.892, + 0.355 + ], + "angle": 0, + "content": "Figure 8. (a) After optimization, the editing direction is more linearly correlated with the yaw attribute. (b) With automatically adjusted editing scale for each test image." + }, + { + "type": "image", + "bbox": [ + 0.506, + 0.36, + 0.89, + 0.456 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.458, + 0.891, + 0.486 + ], + "angle": 0, + "content": "Figure 9. Ablation study set up. We use GANgealing [63] to transfer the annotated points from example to test image." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.497, + 0.892, + 0.557 + ], + "angle": 0, + "content": "additional information about keypoints (which 'DragGAN + GANgealing' does), we believe that it is better suited for batch image editing for many kinds of domains (e.g., Fig. 6), where we might not have such information." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.575, + 0.619, + 0.59 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.601, + 0.892, + 0.752 + ], + "angle": 0, + "content": "We introduced the problem of interactive batch image editing. Given a user edit in an example image, our approach automatically transfers that edit to other test images, maintaining a consistent final state of the edit across images. Extensive experiments demonstrated that our method produces comparable quality to state-of-the-art single-image-editing methods while saving significant time and human effort. We are currently limited to StyleGAN models. Extending this problem and solution to diffusion-based models for more edits types would be an exciting future direction." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.769, + 0.66, + 0.786 + ], + "angle": 0, + "content": "Acknowledgement" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.795, + 0.892, + 0.901 + ], + "angle": 0, + "content": "This work was supported in part by NSF CAREER IIS2150012, Adobe Data Science award, Sony Focused Research award, and Institute of Information & communications Technology Planning & Evaluation (IITP) grant funded by the Korea government (MSIT) (No. RS-2022-00187238, Development of Large Korean Language Model Technology for Efficient Pretraining)." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8278" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.155 + ], + "angle": 0, + "content": "[1] Man M. Ho and Jinjia Zhou. Deep preset: Blending and retouching photos with color style transfer. In WACV, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.157, + 0.47, + 0.198 + ], + "angle": 0, + "content": "[2] Mahmoud Afifi, Konstantinos G. Derpanis, Bjorn Ommer, and Michael S. Brown. Learning multi-scale photo exposure correction. In CVPR, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.2, + 0.468, + 0.253 + ], + "angle": 0, + "content": "[3] Jong-Hyeon Baek, DaeHyun Kim, Su-Min Choi, Hyo-jun Lee, Hanul Kim, and Yeong Jun Koh. Luminance-aware color transform for multiple exposure correction. In ICCV, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.256, + 0.468, + 0.297 + ], + "angle": 0, + "content": "[4] Wenyan Cong, Jianfu Zhang, Li Niu, Liu Liu, Zhixin Ling, Weiyuan Li, and Liqing Zhang. Dovenet: Deep image harmonization via domain verification. In CVPR, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.298, + 0.468, + 0.338 + ], + "angle": 0, + "content": "[5] Yi-Hsuan Tsai, Xiaohui Shen, Zhe Lin, Kalyan Sunkavalli, Xin Lu, and Ming-Hsuan Yang. Deep image harmonization, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.34, + 0.468, + 0.38 + ], + "angle": 0, + "content": "[6] Zonghui Guo, Haiyong Zheng, Yufeng Jiang, Zhaorui Gu, and Bing Zheng. Intrinsic image harmonization. In CVPR, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.382, + 0.468, + 0.422 + ], + "angle": 0, + "content": "[7] Jun Ling, Han Xue, Li Song, Rong Xie, and Xiao Gu. Region-aware adaptive instance normalization for image harmonization. In CVPR, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.424, + 0.468, + 0.479 + ], + "angle": 0, + "content": "[8] Yifan Jiang, He Zhang, Jianming Zhang, Yilin Wang, Zhe Lin, Kalyan Sunkavalli, Simon Chen, Sohrab Amirghodsi, Sarah Kong, and Zhangyang Wang. Ssh: A self-supervised framework for image harmonization. In ICCV, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.481, + 0.468, + 0.521 + ], + "angle": 0, + "content": "[9] Wenyan Cong, Xinhao Tao, Li Niu, Jing Liang, Xuesong Gao, Qihao Sun, and Liqing Zhang. High-resolution image harmonization via collaborative dual transformations. 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.523, + 0.468, + 0.577 + ], + "angle": 0, + "content": "[10] Jeya Maria Jose Valanarasu, He Zhang, Jianming Zhang, Yilin Wang, Zhe Lin, Jose Echevarria, Yinglan Ma, Zijun Wei, Kalyan Sunkavalli, and Vishal M Patel. Interactive portrait harmonization. 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.579, + 0.468, + 0.62 + ], + "angle": 0, + "content": "[11] Lingzhi Zhang, Tarmily Wen, and Jianbo Shi. Deep image blending. In The IEEE Winter Conference on Applications of Computer Vision, pages 231-240, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.621, + 0.468, + 0.675 + ], + "angle": 0, + "content": "[12] Dave Epstein, Allan Jabri, Ben Poole, Alexei A. Efros, and Aleksander Holynski. Diffusion self-guidance for controllable image generation. arXiv preprint arXiv:2306.00986, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.677, + 0.468, + 0.717 + ], + "angle": 0, + "content": "[13] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. 2021. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.719, + 0.468, + 0.759 + ], + "angle": 0, + "content": "[14] Tim Brooks, Aleksander Holynski, and Alexei A Efros. Instructpix2pix: Learning to follow image editing instructions. 2022. 1, 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.761, + 0.468, + 0.815 + ], + "angle": 0, + "content": "[15] Chenlin Meng, Yutong He, Yang Song, Jiaming Song, Jia-jun Wu, Jun-Yan Zhu, and Stefano Ermon. SDEdit: Guided image synthesis and editing with stochastic differential equations. In ICLR, 2022. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.817, + 0.468, + 0.871 + ], + "angle": 0, + "content": "[16] Yujun Shi, Chuhui Xue, Jiachun Pan, Wenqing Zhang, Vincent YF Tan, and Song Bai. Dragdiffusion: Harnessing diffusion models for interactive point-based image editing. 2023. 1, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.468, + 0.902 + ], + "angle": 0, + "content": "[17] Xingang Pan, Ayush Tewari, Thomas Leimkuhler, Lingjie Liu, Abhimitra Meka, and Christian Theobalt. Drag your" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.47, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "gan: Interactive point-based manipulation on the generative image manifold. In SIGGRAPH, 2023. 1, 2, 3, 5, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.121, + 0.892, + 0.162 + ], + "angle": 0, + "content": "[18] Guillaume Couairon, Jakob Verbeek, Holger Schwenk, and Matthieu Cord. Diffedit: Diffusion-based semantic image editing with mask guidance, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.163, + 0.892, + 0.203 + ], + "angle": 0, + "content": "[19] Chong Mou, Xintao Wang, Jiechong Song, Ying Shan, and Jian Zhang. Dragondiffusion: Enabling drag-style manipulation on diffusion models. 2023. 1, 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.205, + 0.892, + 0.259 + ], + "angle": 0, + "content": "[20] Binxin Yang, Shuyang Gu, Bo Zhang, Ting Zhang, Xuejin Chen, Xiaoyan Sun, Dong Chen, and Fang Wen. Paint by example: Exemplar-based image editing with diffusion models. arXiv preprint arXiv:2211.13227, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.261, + 0.892, + 0.286 + ], + "angle": 0, + "content": "[21] Adding Conditional Control to Text-to-Image Diffusion Models, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.288, + 0.892, + 0.314 + ], + "angle": 0, + "content": "[22] GLIGEN: Open-Set Grounded Text-to-Image Generation, 2023. arXiv:2301.07093. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.316, + 0.892, + 0.357 + ], + "angle": 0, + "content": "[23] Or Patashnik, Zongze Wu, Eli Shechtman, Daniel Cohen-Or, and Dani Lischinski. Styleclip: Text-driven manipulation of stylegan imagery. In ICCV, 2021. 1, 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.359, + 0.892, + 0.412 + ], + "angle": 0, + "content": "[24] Bahjat Kawar, Shiran Zada, Oran Lang, Omer Tov, Huiwen Chang, Tali Dekel, Inbar Mosseri, and Michal Irani. Imagic: Text-based real image editing with diffusion models. In CVPR, 2023. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.414, + 0.892, + 0.453 + ], + "angle": 0, + "content": "[25] Yuki Endo. User-controllable latent transformer for stylegan image layout editing. Computer Graphics Forum, 2022. 1, 2, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.455, + 0.892, + 0.496 + ], + "angle": 0, + "content": "[26] Erik Härkönen, Aaron Hertzmann, Jaakko Lehtinen, and Sylvain Paris. Ganspace: Discovering interpretable gan controls. In NeurIPS, 2020. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.498, + 0.892, + 0.537 + ], + "angle": 0, + "content": "[27] Yujun Shen, Jinjin Gu, Xiaou Tang, and Bolei Zhou. Interpreting the latent space of gans for semantic face editing. In CVPR, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.539, + 0.892, + 0.566 + ], + "angle": 0, + "content": "[28] Yujun Shen and Bolei Zhou. Closed-form factorization of latent semantics in gans. In CVPR, 2021. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.567, + 0.892, + 0.607 + ], + "angle": 0, + "content": "[29] Yotam Nitzan, Rinon Gal, Ofir Brenner, and Daniel Cohen-Or. Large: Latent-based regression through gan semantics. In CVPR, 2022. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.609, + 0.892, + 0.649 + ], + "angle": 0, + "content": "[30] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. 2019. 2, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.651, + 0.892, + 0.692 + ], + "angle": 0, + "content": "[31] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of StyleGAN. In CVPR, 2020. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.693, + 0.892, + 0.746 + ], + "angle": 0, + "content": "[32] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. In NeurIPS, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.748, + 0.892, + 0.774 + ], + "angle": 0, + "content": "[33] Yann LeCun. The mnist database of handwritten digits. http://yann.lecun.com/exdb/mnist/, 1998. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.776, + 0.892, + 0.815 + ], + "angle": 0, + "content": "[34] Andrew Brock, Jeff Donahue, and Karen Simonyan. Large scale gan training for high fidelity natural image synthesis. arXiv, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.817, + 0.892, + 0.858 + ], + "angle": 0, + "content": "[35] Terro Karras, Timo Aila, Samuli Laine, and Jaakko Lehtinen. Progressive growing of gans for improved quality, stability, and variation. In ICLR, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.86, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[36] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR, 2009. 2" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "8279" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.135 + ], + "angle": 0, + "content": "[37] Fisher Yu, Yinda Zhang, Shuran Song, Ari Seff, and Jianxiong Xiao. Lsun: Construction of a large-scale image dataset using deep learning with humans in the loop. 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.136, + 0.47, + 0.178 + ], + "angle": 0, + "content": "[38] Ali Jahanian, Lucy Chai, and Phillip Isola. On the \"steerability\" of generative adversarial networks. In ICLR, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.18, + 0.471, + 0.235 + ], + "angle": 0, + "content": "[39] David Bau, Jun-Yan Zhu, Hendrik Strobelt, Agata Lapedriza, Bolei Zhou, and Antonio Torralba. Understanding the role of individual units in a deep neural network. Proceedings of the National Academy of Sciences, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.238, + 0.47, + 0.279 + ], + "angle": 0, + "content": "[40] Zongze Wu, Dani Lischinski, and Eli Shechtman. Stylespace analysis: Disentangled controls for stylegan image generation. In ICLR, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.281, + 0.47, + 0.323 + ], + "angle": 0, + "content": "[41] Tengfei Wang, Yong Zhang, Yanbo Fan, Jue Wang, and Qifeng Chen. High-fidelity gan inversion for image attribute editing. In CVPR, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.325, + 0.47, + 0.367 + ], + "angle": 0, + "content": "[42] Omer Tov, Yuval Alaluf, Yotam Nitzan, Or Patashnik, and Daniel Cohen-Or. Designing an encoder for stylegan image manipulation. In ACM Transactions on Graphics, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.368, + 0.47, + 0.41 + ], + "angle": 0, + "content": "[43] Daniel Roich, Ron Mokady, Amit H Bermano, and Daniel Cohen-Or. Pivotal tuning for latent-based editing of real images. ACM Trans. Graph., 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.412, + 0.47, + 0.482 + ], + "angle": 0, + "content": "[44] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.484, + 0.469, + 0.525 + ], + "angle": 0, + "content": "[45] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.527, + 0.47, + 0.596 + ], + "angle": 0, + "content": "[46] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In H. Larochelle, M. Ranzato, R. Hadsell, M.F. Balcan, and H. Lin, editors, Advances in Neural Information Processing Systems. Curran Associates, Inc., 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.599, + 0.47, + 0.655 + ], + "angle": 0, + "content": "[47] Ron Mokady, Amir Hertz, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Null-text inversion for editing real images using guided diffusion models. arXiv preprint arXiv:2211.09794, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.656, + 0.47, + 0.712 + ], + "angle": 0, + "content": "[48] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-prompt image editing with cross attention control. arXiv preprint arXiv:2208.01626, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.714, + 0.47, + 0.769 + ], + "angle": 0, + "content": "[49] Huan Ling, Karsten Kreis, Daiqing Li, Seung Wook Kim, Antonio Torralba, and Sanja Fidler. Editgan: High-precision semantic image editing. In Advances in Neural Information Processing Systems (NeurIPS), 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.771, + 0.47, + 0.813 + ], + "angle": 0, + "content": "[50] David Bau, Steven Liu, Tongzhou Wang, Jun-Yan Zhu, and Antonio Torralba. Rewriting a deep generative model. In ECCV, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.814, + 0.47, + 0.858 + ], + "angle": 0, + "content": "[51] Amir Bar, Yossi Gandelsman, Trevor Darrell, Amir Globerson, and Alexei A. Efros. Visual prompting via image inpainting. 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.859, + 0.47, + 0.902 + ], + "angle": 0, + "content": "[52] Thao Nguyen, Yuheng Li, Utkarsh Ojha, and Yong Jae Lee. Visual instruction inversion: Image editing via visual prompting. In NeurIPS, 2023." + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.471, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.092, + 0.893, + 0.147 + ], + "angle": 0, + "content": "[53] Yasheng Sun, Yifan Yang, Houwen Peng, Yifei Shen, Yuqing Yang, Han Hu, Lili Qiu, and Hideki Koike. Imagebrush: Learning visual in-context instructions for exemplar-based image manipulation. 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.149, + 0.893, + 0.19 + ], + "angle": 0, + "content": "[54] Aaron Hertzmann, Charles E. Jacobs, Nuria Oliver, Brian Curless, and David H. Salesin. Image analogies. In Seminal Graphics Papers: Pushing the Boundaries, Volume 2, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.192, + 0.893, + 0.233 + ], + "angle": 0, + "content": "[55] Jing Liao, Yuan Yao, Lu Yuan, Gang Hua, and Sing Bing Kang. Visual attribute transfer through deep image analogy. 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.234, + 0.893, + 0.275 + ], + "angle": 0, + "content": "[56] Weihao Xia, Yulun Zhang, Yujiu Yang, Jing-Hao Xue, Bolei Zhou, and Ming-Hsuan Yang. Gan inversion: A survey. TPAMI, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.277, + 0.893, + 0.318 + ], + "angle": 0, + "content": "[57] Yunjey Choi, Youngjung Uh, Jaejun Yoo, and Jung-Woo Ha. Starganv2: Diverse image synthesis for multiple domains. In CVPR, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.32, + 0.893, + 0.361 + ], + "angle": 0, + "content": "[58] Tero Karras, Miika Aittala, Janne Hellsten, Samuli Laine, Jaakko Lehtinen, and Timo Aila. Training generative adversarial networks with limited data, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.362, + 0.893, + 0.418 + ], + "angle": 0, + "content": "[59] Jianglin Fu, Shikai Li, Yuming Jiang, Kwan-Yee Lin, Chen Qian, Chen-Change Loy, Wayne Wu, and Ziwei Liu. Stylegan-human: A data-centric odyssey of human generation. arXiv preprint, arXiv:2204.11823, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.42, + 0.893, + 0.447 + ], + "angle": 0, + "content": "[60] Davis E. King. Dlib-ml: A machine learning toolkit. J. Mach. Learn. Res., 2009. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.449, + 0.893, + 0.504 + ], + "angle": 0, + "content": "[61] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, Gunter Klambauer, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. In NeurIPS, 2017. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.505, + 0.893, + 0.546 + ], + "angle": 0, + "content": "[62] Thorsten Hempel, Ahmed A. Abdelrahman, and Ayoub Al-Hamadi. 6d rotation representation for unconstrained head pose estimation. In ICIP, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.548, + 0.893, + 0.589 + ], + "angle": 0, + "content": "[63] William Peebles, Jun-Yan Zhu, Richard Zhang, Antonio Torralba, Alexei Efros, and Eli Shechtman. Gan-supervised dense visual alignment. In CVPR, 2022. 8" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.092, + 0.893, + 0.589 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.956 + ], + "angle": 0, + "content": "8280" + } + ] +] \ No newline at end of file diff --git a/2024/Edit One for All_ Interactive Batch Image Editing/d0962876-52e1-4fef-ac92-9ddbf4fa5230_origin.pdf b/2024/Edit One for All_ Interactive Batch Image Editing/d0962876-52e1-4fef-ac92-9ddbf4fa5230_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..c00230e7590910ddd1cbc884b158cb789eb6a248 --- /dev/null +++ b/2024/Edit One for All_ Interactive Batch Image Editing/d0962876-52e1-4fef-ac92-9ddbf4fa5230_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d079d6aa7126a87a403c2a3bd17ecf7cead86224ea0b2826fb91d3a677fd2da7 +size 8625629 diff --git a/2024/Edit One for All_ Interactive Batch Image Editing/full.md b/2024/Edit One for All_ Interactive Batch Image Editing/full.md new file mode 100644 index 0000000000000000000000000000000000000000..28374163b40a73337b5e4fd034f569a69ab02eb6 --- /dev/null +++ b/2024/Edit One for All_ Interactive Batch Image Editing/full.md @@ -0,0 +1,298 @@ +# Edit One for All: Interactive Batch Image Editing + +Thao Nguyen Utkarsh Ojha Yuheng Li Haotian Liu Yong Jae Lee University of Wisconsin-Madison + +https://thaoshibe.github.io/edit-one-for-all + +# Abstract + +In recent years, image editing has advanced remarkably. With increased human control, it is now possible to edit an image in a plethora of ways; from specifying in text what we want to change, to straight up dragging the contents of the image in an interactive point-based manner. However, most of the focus has remained on editing single images at a time. Whether and how we can simultaneously edit large batches of images has remained understudied. With the goal of minimizing human supervision in the editing process, this paper presents a novel method for interactive batch image editing using StyleGAN as the medium. Given an edit specified by users in an example image (e.g., make the face frontal), our method can automatically transfer that edit to other test images, so that regardless of their initial state (pose), they all arrive at the same final state (e.g., all facing front). Extensive experiments demonstrate that edits performed using our method have similar visual quality to existing single-image-editing methods, while having more visual consistency and saving significant time and human effort. + +# 1. Introduction + +Image editing has undergone a transformation in recent years with the aid of modern generative models. The process has been greatly democratized, where many of the sophisticated edits, which previously might have required hours and niche expertise, can now be completed with relative ease in a matter of minutes. For example, different types of learning based algorithms can be used for image correction/adjustment [1-12] and for manipulating the semantic contents of real images [13-22]. Moreover, there are many different ways to edit an image. For instance, a user can specify in text what changes they want - e.g., "make the hair darker" [14, 15, 23, 24], or they can drag the contents of the image in an interactive manner to shrink, enlarge, or move a part of an object [12, 16, 17, 19, 25]. + +A common theme across many of these works is that the edits are designed to work for a single image at a time. For example, given an image of a cat with open eyes, DragGAN [17] allows the user to drag the contents of that par + +![](images/0eb91e7d6b9d8e795b992b5d273fc19a0ad33d6f0a5c2b84cecc4e7dd125563f.jpg) +Figure 1. Interactive Batch Image Editing. Given a single user edited image, the goal is to automatically transfer that edit to new unseen images, so that all edited images end up with the same final state as the user edited image. + +ticular image so that both the eyes can be closed in the resulting cat image. But what if we wish to apply the same edit to many different types of cats so that we can close all of their eyes? While one could perform dragging on each cat separately, the whole process would be quite cumbersome (requiring lots of human annotation) and time intensive (requiring optimization for each image). + +In this paper, we introduce the new problem of Interactive Batch Image Editing. Given a user-specified edit $I \rightarrow I'$ , the goal is to automatically transfer that edit to new unseen images, so that all edited images end up with the same final state as $I'$ (e.g., all cats with eyes closed) regardless of their original starting states (e.g., varying degrees of original eye openness); see Figure 1. Hence, a solution to this problem requires two components: (i) modeling the user edit in the example pair $(I, I')$ so that it can be transferred to new images; and (ii) controlling the degree of the + +![](images/a2fbce35151b420551fee3a48bde474afebeb2ffd3af5f9c5decb05340526b8d.jpg) +edit one image at a time +Figure 2. Single Image Editing vs. Batch Image Editing. (a) Prior work (e.g., [17, 19, 25]) focuses on single image editing. (b) We focus on batch image editing, where the user's edit on a single image is automatically transferred to new images, so that they all arrive at the same final state regardless of their initial starting state. In this way, we can achieve time speed up and reduce human effort in editing. + +![](images/2ddb6c5d6e47094d0816beaae5e12608cb13fb44ebde5f8a66a4d9a97f06ab26.jpg) +edit one image, transfer to others at the same time! + +edit so that all edited images have the same final state. + +To model the user's edit, we build upon existing literature on Generative Adversarial Networks (GANs) [23, 26-29], and in particular StyleGAN [30, 31], which shows that their learned latent space emits globally linear editing directions where the edited attribute (e.g., eye closeness) varies linearly in magnitude along such a direction for all semantically related instances (e.g., all closed eyes cat). To try to discover the global direction corresponding to a user edit, we perform optimization in the latent space of StyleGAN2 [31] so that (i) the edited image with the discovered direction is visually similar to the original edited image $I'$ , and (ii) the linear correlation between the distance along the direction and the strength of the visual attribute is highest (e.g., going twice the amount in that direction closes a cat's eyes by twice the amount). + +To control the degree of the edit so that all edited images end up with the same final state, we derive a closed-form solution. Taking motivation from [29], we model the attribute strength (e.g., how much the cat's eyes are open) as the distance along the normal vector from a hyperplane. By setting the example edited image as lying on that hyperplane (i.e., its distance is 0), the objective is to move all other images along that same direction so that their distances to the hyperplane also become 0. In this way, regardless of the start state of any image (e.g., varying degree of eye openness), their edited end states all become the same (e.g., closed eyes). We show that we can analytically compute the exact amount of traversal along the direction for any image with a closed-form solution. Editing images in this way helps improve the results visually, as the final state of any new image matches that of the edited example given by the user. + +Our method works with various edits given by an editing framework, e.g., it can be a change given by a dragging operation [17], or it can be a change given by a text-based edit [14, 23] (e.g., "make eyes bigger"), as long as it is a direction compatible with a StyleGAN2 [31] model. We present qualitative and + +quantitative results in transferring edits for a variety of objects (cats/dogs/faces/humans/limons/arts/etc.), parts (mouths/ears/legs/etc.), and corresponding attributes (big eyes/short faces/pale skins/etc.). Importantly, we show that the final state of the edited images is comparable to the scenario in which a user performs the edit for each image separately. But since our method does this automatically (see Fig. 2), we show that it takes much less time (e.g., only 0.05s per image, compared to DragGAN [17] about 2s per image), and does not require laborious human annotation (e.g., only need 1 annotated image, compared to 4 required for DragGAN [17]). Finally, we show practical applications of batch image editing; e.g., changing wheel size in every photo of a car collection by editing just one. + +In sum, our contributions are: (1) We introduce the novel problem of interactive batch image editing, wherein a user-driven edit is automatically transferred to other similar test images. (2) We study what having the same final state means in a geometrical sense, and propose a principled approach to achieving that in StyleGAN2 [31]'s latent space. (3) We show that our method works on a wide variety of domains e.g., cats, dogs, humans etc., taking significantly less time than editing each image individually while having similar visual quality and more visual consistency. + +# 2. Related Work + +Latent space of GANs. The capabilities of generative adversarial networks (GANs) have transformed drastically, from the first GAN [32] designed for simple image datasets like MNIST [33] to big and powerful models like BigGAN [34], Progressive GAN [35] and StyleGANs [30, 31] designed for much more complex datasets [36, 37]. Parallel to the efforts to enhance their capabilities, there has also been work done to better understand and make use of their learned latent space. Specifically, there is a line of work which tries to find interpretable directions in that space, so that moving in such a direction majorly changes only one discernible attribute in the image. Some methods [38, 39] + +![](images/ed94eafe60804ef8f989e160c3fc16fc755bfb18cce8c08b180a96a1fa5d073d.jpg) +(a) Setting + +![](images/8dd88c9fcd19d6b656d62f993e17a4355cd8639d14d75990bae421b3283904c9.jpg) +(b) Naïve Approach + +![](images/cf3b505c6e9ae51a1836b378c628907361c396818e3a871ec748313ca0e39026.jpg) +(c) Optimizing Editing Direction +Figure 3. Different editing strategies. (a) Setting. (b) Naive Approach: The editing direction effective for an example may not generalize well to test images. (c) Optimizing Editing Direction: We optimize for a globally consistent direction that is effective for both example and test images. (d) Adjusting Editing Strength: Ensuring consistent final states requires adjusting the editing strength for each test image. + +![](images/a4a06c1b516757ff2a327010b4edc31624f4671b2108a1a73a8da292a17eea67.jpg) +(d) Adjusting Edit Strength + +try to find such directions or particular activations using (self) supervised methods. Others [26-28] try getting rid of the need for supervision: discovering the directions representing the most important factors of variation, whatever they may be. Going beyond the $\mathcal{W}$ space, the authors in [40] explore the $\mathcal{W}+$ space, and show that it has even better disentanglement useful for spatial image editing. All of these hidden capabilities have made StyleGANs a useful tool for editing purposes. However, prior work mainly focuses on the effectiveness of the editing direction (e.g., whether the discovered direction can change yaw pose). The question of whether these directions can be applied to diverse images to yield consistent results (e.g., all faces facing frontal) and, if so, how to achieving such consistency, remains unanswered. + +Image editing with generative models. Researchers have used StyleGAN for real image editing by designing an encoder to invert a real image into StyleGAN's latent space [41-43]. StyleCLIP [23] presents a way to perform image editing through a text-based interface by making use of the CLIP encoder [44]. Very recently, StyleGAN has also been used to perform point-based editing [17, 25] so as to move any start point to reach a target location in the image; thereby elongating, rotating, shifting the objects. Most work on GANs has been for single object category datasets, but with the rise of text-to-image Diffusion Models which can generate complex images [13, 45, 46], editing such images is now possible. Complex scenes can be edited either using text [14, 15, 21, 24, 47, 48], or using the same idea of point-based manipulation [16, 19]. There have been some works which discuss the possibility to transfer the edit of one image to another. EditGAN [49] uses segmentation mask manipulations to edit an image, but to successfully transfer them to an unseen image, it needs to do post hoc manipulation of the editing scale. RewriteGAN [50] involves editing one generated sample; e.g., adding a patch of a tree onto a church tower. Following this, the rules of the + +GAN are manipulated so that all churches have some tree on the top. Within diffusion models, visual prompting/ image analogies tries something similar, where users can define a triplet {before, after, test} to learn the edit and transfer it to a test image [51-55]. Our task is similar, in that we wish to transfer user-edits from the training example to new images, but it differs in one crucial aspect: along with transferring the edit, we wish to automatically learn its strength so that the edit produces the same final state for a new image. + +# 3. Approach + +Our focus is on the setting in which an image editing process edits an attribute of an image so that the attribute's value reaches a desired state; e.g., rotating the pose of a face looking sideways so that it faces front (attribute = face's pose, value = front). + +With this view, we explain our framework for batch image editing, which can be broken down into two stages: (i) A user edits an image $I_0$ (e.g., using DragGAN [17]) to obtain an edited image $I_0'$ . We describe in Sec. 3.1 how we capture this user edit $I_0 \rightarrow I_0'$ , so that the same edit can be applied to new images. (ii) Next, we describe in Sec. 3.2, how for any new image, we apply the modeled edit by automatically adjusting its strength so that the attribute's value in this new image matches that of $I_0'$ (e.g., any face, regardless of its initial pose, now faces front after the edit). + +# 3.1. Modeling the User Edit + +We start with an image $I_0$ . This image could be a real image or a generated one. Either way, we get its latent representation in the $\mathcal{W}$ space of a StyleGAN2 [31] model $G$ , so that $I_0 = G(w_0)$ . (For real images we can use GAN-inversion techniques [56].) The user, with the help of an image editing framework (e.g., DragGAN [17] or InstructPix2Pix [14]), edits $I_0$ to manipulate one or more of its attributes. The edit maps to the $\mathcal{W}$ space as $w_0 \to w_0'$ . The + +resulting edited image can thus be recovered as follows: $I_0' = G(w_0')$ . Fig. 3(a) shows an example of the original and edited image pair, $(I_0, I_0')$ , where the user intended to turn the face forward. + +Now, given the user edit $I_0 \to I_0'$ , we wish to capture it in a way that can be applied to new images in a generalizable manner; i.e., the application of the edit changes the same property in a new image. This is where a nice property of GANs, and in particular StyleGANs [30, 31] comes in useful. It has been shown in prior works [26-28] that it is possible to discover directions with such properties (using supervised as well as unsupervised methods) in the learned $\mathcal{W}$ space. In particular, it is possible to find directions $(\Delta_w)$ that are globally consistent. Taking motivation from [29], we define a globally consistent direction $\Delta_w$ as the following: for any arbitrary $w$ , moving along $\Delta_w$ , $w \to w + \Delta_w$ (i) changes the same attribute, and (ii) by the same amount. + +To make the precise user edit $I_0 \to I_0'$ applicable to other images, it needs to be captured as a globally consistent direction. The naive way to represent that edit will be through a simple difference in the $\mathcal{W}$ space: $\Delta_w = w_0' - w_0$ . However, empirically, and as we show in Fig. 3(b), applying this $\Delta_w$ to the latent code $w_1$ corresponding to a new image $I_1$ does not always result in the same change; while $I_0 \to I_0'$ results in a pose change of $\sim 30^\circ$ degrees in yaw, $I_1 \to I_1'$ does not seem to change the same attribute or at least not by the same amount. + +Hence, our goal is to represent the $I_0 \to I_0'$ edit through a $\mathcal{W}$ space direction that can better satisfy both the properties of a globally consistent direction. For this task, we first take motivation from LARGE [29] to introduce a mathematical view of what those directions mean. + +Globally consistent direction. Let's say that $\Delta_g$ is one such direction. Along with a bias term $b$ , we can define a hyperplane as follows: + +$$ +w \cdot \Delta_ {g} + b = 0 \tag {1} +$$ + +That is, any point $w$ which lies on the hyperplane will satisfy this equation. The authors in [29] argued that for such a hyperplane, whose normal vector is a globally consistent direction $(\Delta_g)$ , the distance of an arbitrary point $w$ from that hyperplane $(w \cdot \Delta_g)$ will be linearly correlated to the actual attribute $(y)$ that results in the generated image: + +$$ +y = a \times (w \cdot \Delta_ {g}) + b \tag {2} +$$ + +Here, $a$ and $b$ are unknown linear coefficients. For example, if $\Delta_g$ corresponds to change in pose, then all the front facing people will lie at the same distance $d$ from the above hyperplane. Because of this, any hyperplane defined with respect to a globally consistent direction can be viewed as a semantic hyperplane. For simplicity, we can set $d = 0$ for the edited image $I_0'$ . + +Given that the original $\Delta_w$ may not always be globally consistent, we aim to discover, through optimization, a different direction $\Delta_w^*$ which produces a similar editing effect, but is more likely to be a globally consistent one. We initialize $\Delta_w^*$ with 0's, and design two objective functions to optimize it. First, to make sure that it produces a similar effect as the original direction, we use an image reconstruction loss so that the edited image produced by the new edit matches the original edit given by the user: + +$$ +\mathcal {L} _ {i m g} = \left\| G \left(w _ {0} + \Delta_ {w}\right) - G \left(w _ {0} + \Delta_ {w} ^ {*}\right) \right\| _ {2} \tag {3} +$$ + +Additionally, user can provide the real value of distance $d$ , otherwise we set the distance of the edited image from the hypothetical hyperplane to be 0 (similar to [29]). We constrain the learned $\Delta_w^*$ to follow this property for better interpretability. That is, the new direction should be such that when we traverse the original latent code in that direction, $w_0 \rightarrow w_0 + \Delta_w^*$ , the resulting image should lie at the hyperplane defined by that direction. Setting $b = 0$ in Eq. 1, we minimize the edited point's distance from the hyperplane: + +$$ +\mathcal {L} _ {a t t} = \left| \left(w _ {0} + \Delta_ {w} ^ {*}\right) \cdot \Delta_ {w} ^ {*} \right| \tag {4} +$$ + +The overall loss function for optimizing the new direction $\Delta_w^*$ is a weighted sum of the two objectives: $\mathcal{L}_{\Delta} = \mathcal{L}_{img} + \lambda \mathcal{L}_{att}$ . We set $\lambda$ to 0.02 in all experiments to balance the magnitude of $\mathcal{L}_{img}$ and $\mathcal{L}_{att}$ . + +Now, in Fig. 3 (b) and (c), we compare the difference in image editing results using two different directions when applied to the same new image $(I_1)$ : $w_1 \to w_1 + \Delta_w$ vs. $w_1 \to w_1 + \Delta_w^*$ . As discussed before, in this case, the editing using the naively computed direction $(\Delta_w)$ is not able to accurately capture the pose rotation the way it did for the user edited example $(I_0 \to I_0')$ . On the other hand, with $\Delta_w^*$ , we see that the pose of the woman changes by a similar amount as the edited example. However, since the original pose of the woman $(I_1)$ was not the same as the pose of the man $(I_0)$ , the new edited image $(I_1')$ still does not arrive at the same final state as $I_0'$ . Therefore, our next goal is to figure out how to scale the learned direction so that $I_1'$ does arrive at the same final state as $I_0'$ . + +# 3.2. Adjusting Editing Strength for New Images + +Given the optimized $\Delta_w^*$ and some new image $I_{i} = G(w_{i})$ we wish to edit it in the following way: + +$$ +w _ {i} ^ {\prime} = w _ {i} + \alpha_ {i} n \tag {5} +$$ + +where $\alpha_{i}$ is the editing strength computed separately for each image and $n = \Delta_w^* /||\Delta_w^* ||$ is a unit vector in the direction of $\Delta_w^*$ . Fig. 3 (d) illustrates a geometric perspective that we will use to compute $\alpha_{i}$ . First, we see $\Delta_w^*$ represented as a vector and its corresponding hyperplane that is normal to it. Next, we depict the new latent point $w_{i}$ . The + +![](images/c8a7088e79d3a3ec3755d382e2c6186bf27b6e736d51de3ff8d7bbc1f2b622ea.jpg) +Original +UserID +DragGAN +DragDiff. +Ours +Figure 4. Qualitative comparisons between dragging baselines. For ours, green bounding box indicates automatic transfer from the red bounding box example in the first row (i.e., no point annotation needed!). + +goal is to move it along the $\Delta_w^*$ direction so that it arrives onto the hyperplane. Through this depiction, $\alpha_{i}$ can be understood as the distance of $w_{1}$ from the hyperplane. We can get the distance by projecting $(w_0' - w_i)$ in the direction of $n$ . Therefore, $\alpha_{i} = (w_{0}^{\prime} - w_{i})\cdot n$ . + +Since for each image, the only unique computation that needs to be done is the calculation of $\alpha_{i}$ , we can see why our method will be much faster than, for example, annotating every image and running the DragGAN [17] optimization each of those times. We will show in our experiments the significant difference in time taken by our method compared to single image editing baselines. + +Importantly, there is an added benefit to computing $\alpha$ 's in this way. Let's say the goal of the editing process is to rotate $n$ faces and make them frontal. After completion, all of them do become frontal, each with their own editing strengths $\{\alpha_{1},\alpha_{2},\dots,\alpha_{n}\}$ computed using the above formula. Suppose the user now wants the same faces facing a bit left instead. To do this, the user does not need to re-associate the original training example and run the optimization one more time. In an interactive manner, they can simply scale the $\alpha$ for the training example to match the desired edit, and all other $\alpha$ 's can be automatically recomputed. + +# 4. Experiments + +We study how well our method models and transfers the edits from an example, and how efficient it is compared to single image editing baselines using their official code. + +Categories. We evaluate on a variety of domains: Human faces (FFHQ) [30], AFHQ Cats, Dogs [57], MetFaces [58], Human bodies [59]. For each domain, we use the corresponding pretrained StyleGAN model to perform editing. + +![](images/ea1041f98c84f017aca711fc32ed3636d05645398c67c2f22e408c5baca9685c.jpg) +Original +UserID +DragGAN +DragDiff. +Ours + +# 4.1. Qualitative results + +We start with qualitative comparisons with (i) interactive point-dragging, and (ii) text-based image editing baselines. + +Interactive point based editing. We compare against DragGAN [17], UserControllableLT [25], and DragDiffusion [16]. For each baseline, we, as a user, desire to edit a bunch of images to have the same final state for an attribute. Fig. 4 shows the results of editing two kinds of edits to four images (leftmost column). In the left case, the goal is to make everyone smile by the same degree. In the right, it is to vertically compress everyone's face by the same degree. For all three baselines, we manually annotate points for each test image. For our method, we only annotate points and perform DragGAN-based edit on one image (top row), then automatically transfer the edit to the three other test images. + +For the smiling case (left), we notice that when DragGAN drags the upper/lower lips up/down respectively, it can either make people smile (1st, 2nd and 4th rows), or it can make people look shocked (blonde girl; 3rd row). Both edits are technically correct based on lips movement, but it won't match the user expectation of making everyone smile by the same degree. Since all four images are edited independently, this is not unexpected. The other two baselines sometimes have issues introducing the same edit across different images; e.g., UserControllableLT can make the blonde girl smile a little bit, but not the other images. In general, we notice in our experiments that to introduce the same exact edit that we want, we often need to play around with some hyperparameters (e.g., # of iterations). In comparison, our method produces edited images that are smiling more equally (left) and have been compressed by a similar + +![](images/df55c72c3c40102246cc15bc76fd8a26172d87785963f53ca1907f922ed2d67f.jpg) +Figure 5. Qualitative comparisons to text-guided baselines. Ours transfers the edit from example (red), to other test images (green). + +![](images/eae8fe8079eafdcd4acca4e58f1046c8620c9b2f042b98a7a7d7afeb5da25fd6.jpg) + +amount (right), and with much less human effort. + +Text-driven image editing. We consider two baselines: StyleCLIP [23] and InstructPix2Pix [14]. Each takes an image and text as input to produce an edited image: $I \rightarrow I'$ . Fig. 5 shows the results of Ours vs. StyleCLIP (left) and Ours vs. InstructPix2Pix (right). (For the latter, we invert the output of InstructPix2Pix into StyleGAN's latent space using [42] for our method.) In each case, for the baselines, we use the same text prompt, e.g., "A photo of a person with pale skin" for each edited image (in four rows). For ours, we capture the edit from the first editing result of a baseline (first row), and then automatically transfer the edit to the remaining three images. + +Our goal with this particular setup is to test (i) how consistently the baselines introduce the edit denoted by text to different images (e.g., do all people become equally pale?), and (ii) how consistently our method captures and transfers the particular edit of the first example to the rest of the images, irrespective of how good that first edit was. For (i), we find that StyleCLIP sometimes has issues; e.g., the paleness of the edited face in first example is different from the second (first vs second rows). For InstructPix2Pix, sometimes it can give consistent results; the childness of different faces seems similar. But, it can have consistency issues in other examples; e.g., the type/strength of maleness introduced is different in each image. And this is where, we believe, the utility of our method lies: if one desires to edit every face to be male in a particular way (i.e., to the same degree) as the first example, e.g., thick eyebrows & light beard, our method has an advantage. + +Finally, we show our method's results for non-facial domains in Fig. 6. For the lion examples (top left), we see that it can preserve not just the type of edit, e.g., dragging + +
MethodPoint DraggingTimeAnno. #imgs
Dist.FID1 img1k imgs
UserControl. [25]13.6425.320.03s30s1000
DragGAN [17]4.1659.282s33.33m1000
DragDiffusion [16]26.5636.5560s16.67h1000
Ours9.4679.352s82s1
+ +Table 1. Time is estimated for 1 point drag, without human annotation time. (82s includes 2s to perform edit on the example, 30s to optimize the editing direction, and 50s to transfer the edit to 1000 test images.) Our method requires only one image annotation in total, while the baselines need one annotation per test image. + +to make the lion roar, but also the strength of the roar. The strength being preserved can be observed more easily for the human body poses (bottom left), where we see that the extent of legs split, hand movements, is consistent enough to almost align the edited test images with the user edited one. Overall, results on these diverse set of domains highlight an observation that many kinds of edits can be thought of as a combination of {type, strength}, both of which can indeed be captured and transferred according to our needs. + +# 4.2. Quantitative results + +Next, perform quantitative experiments following the setup proposed in [16, 17, 19]. We first randomly sample 10 facial test images and pair them with a randomly sampled target face image. We use dlib-ml [60] to detect keypoints in the test and target images. Our goal is to see if the keypoints in the test images can be moved to the target locations specified by keypoints in the target image. For the baselines [16, 17, 25], we perform dragging for each test image. For our method, we perform dragging on one test image, and then transfer the edit to the remaining nine images. This is repeated 100 times; i.e., each time a random target image is + +![](images/44f2934bfb3f2ad87d0e208d632ee7cc74a7d5cce1f23c1c0b29ef27845b63e6.jpg) +Figure 6. Additional qualitative results on various domains. + +paired with 10 random test images. + +We report the Euclidean distance between the keypoints of edited and target images in Table 1. Even without requiring annotations for every test image, our method can move the points very close to the target; almost as close as DragGAN, which requires annotation each time. We also compute FID [61] between the original test images and their edited versions to ensure that our method does not distort the image; and our quality is comparable to DragGAN. + +# 5. Deeper Analysis + +We perform deeper analysis on the usefulness of our batch image editing framework by focusing on a specific edit: face pose rotation. In particular, if we wish to rotate many faces to front, how do we visualize the improvements brought by different components of our method? + +Fig. 7 depicts the setup. We use DragGAN to perform + +dragging on one image (top; $w_0 \rightarrow w_0'$ ). We then transfer the edit to 1000 other test images (two shown in rows 2-3). We see that naively applying $\Delta_w = w_0' - w_0$ from DragGAN to test images cannot make them frontal ( $0^\circ$ ). Using our optimized $\Delta_w^*$ (without dynamic scaling) does help it bring closer to facing front. Does this mean that $\Delta_w^*$ might be a more global direction? We study this in Fig. 8 (a), where we visualize the degree to which the correlation property of global directions (Eq. 2) holds for $\Delta_w$ and $\Delta_w^*$ individually. We see that there is indeed a better correlation ( $R^2$ ) between the distances of latents from hyperplanes and the Yaw degree of resulting images. (We predict Yaw using 6DRepNet [62].) However, as we discussed before, $\Delta_w^*$ in itself is insufficient: to completely bring the facial pose to front, we need to scale it with corresponding $\alpha$ 's to bring them much closer to front (Fig. 7, 4th column). To study this in a more systematic way, we visualize the effect on all 1000 test images. Similar to Fig. 8 (a), we visualize + +![](images/f60e5e943108b02f35a126b09dc2235f141c9eec4980e764b72229e506fa1bd1.jpg) +Figure 7. Effect of adjusted and optimized editing for test images. Yaw degree of each image is provided on top of each images. + +
MethodMAETime complexity
PrepareInference
Random11.295 ± 8.972--
DragGAN [17] + GANgealing [63]8.141 ± 7.22130s2000s
Ours2.120 ± 1.81832s50s
+ +Table 2. Ablation Study: DragGAN + GANgealing. + +distance-to-hyperplane vs. Yaw degree for the original test images (blue) and edited images (red) in Fig. 8 (b). We can see that the variation in Yaw for the edited images collapses at around $0^{\circ}$ ; i.e., they mostly face front as we would like. + +We also compute the time it takes to perform editing in Table 1, last three columns. Our method only requires annotating one image, while the baselines require annotating each image (e.g., 1000 total in this case). Note the times do not include human annotation time. + +Other ways to automate batch image editing. One baseline for batch image editing (i.e., so that all edited images achieve the same final state) is shown in Fig. 9. After the user annotates the source/target points in the example image, we use GANgealing [63] to transfer the points to corresponding locations in each new test image. The edited images are then obtained using DragGAN (DragGAN + GANgealing). We compare our method to this baseline when the user edits one face image to become front $(\mathrm{Yaw} = 0^{\circ})$ and transferring it to 1000 other test images. + +Results are shown in Table 2. We report mean absolute error (MAE) between the Yaw degree of the edited image and its ideal frontal image $(\mathrm{Yaw} = 0^{\circ})$ . 'Random' shows the variations in Yaw for the original images (before editing). While 'DragGAN + GANgealing' does help in reducing the variation $(8.14 < 11.29)$ , our edited images are much more frontal, with an MAE of 2.12. This is because the baseline has no way to bring all edited images to the same final state. On top of that, since our method does not rely on + +![](images/2344e8afeb8e2a2009e2d8ace1e9bc1b02dd78946e357c60d7b839bb4c43ccb5.jpg) + +![](images/73454f2fb7eaedc5a334eca6e89fb4fe69459100f142d30f76d4839735cb0786.jpg) +Figure 8. (a) After optimization, the editing direction is more linearly correlated with the yaw attribute. (b) With automatically adjusted editing scale for each test image. + +![](images/870b11d8902c50cfe052a24c06cf8e5498bccb50ab164a08a9254cf03e894bd9.jpg) +Figure 9. Ablation study set up. We use GANgealing [63] to transfer the annotated points from example to test image. + +additional information about keypoints (which 'DragGAN + GANgealing' does), we believe that it is better suited for batch image editing for many kinds of domains (e.g., Fig. 6), where we might not have such information. + +# 6. Conclusion + +We introduced the problem of interactive batch image editing. Given a user edit in an example image, our approach automatically transfers that edit to other test images, maintaining a consistent final state of the edit across images. Extensive experiments demonstrated that our method produces comparable quality to state-of-the-art single-image-editing methods while saving significant time and human effort. We are currently limited to StyleGAN models. Extending this problem and solution to diffusion-based models for more edits types would be an exciting future direction. + +# Acknowledgement + +This work was supported in part by NSF CAREER IIS2150012, Adobe Data Science award, Sony Focused Research award, and Institute of Information & communications Technology Planning & Evaluation (IITP) grant funded by the Korea government (MSIT) (No. RS-2022-00187238, Development of Large Korean Language Model Technology for Efficient Pretraining). + +# References + +[1] Man M. Ho and Jinjia Zhou. Deep preset: Blending and retouching photos with color style transfer. In WACV, 2021. 1 +[2] Mahmoud Afifi, Konstantinos G. Derpanis, Bjorn Ommer, and Michael S. Brown. Learning multi-scale photo exposure correction. In CVPR, 2021. +[3] Jong-Hyeon Baek, DaeHyun Kim, Su-Min Choi, Hyo-jun Lee, Hanul Kim, and Yeong Jun Koh. Luminance-aware color transform for multiple exposure correction. In ICCV, 2023. +[4] Wenyan Cong, Jianfu Zhang, Li Niu, Liu Liu, Zhixin Ling, Weiyuan Li, and Liqing Zhang. Dovenet: Deep image harmonization via domain verification. In CVPR, 2020. +[5] Yi-Hsuan Tsai, Xiaohui Shen, Zhe Lin, Kalyan Sunkavalli, Xin Lu, and Ming-Hsuan Yang. Deep image harmonization, 2017. +[6] Zonghui Guo, Haiyong Zheng, Yufeng Jiang, Zhaorui Gu, and Bing Zheng. Intrinsic image harmonization. In CVPR, 2021. +[7] Jun Ling, Han Xue, Li Song, Rong Xie, and Xiao Gu. Region-aware adaptive instance normalization for image harmonization. In CVPR, 2021. +[8] Yifan Jiang, He Zhang, Jianming Zhang, Yilin Wang, Zhe Lin, Kalyan Sunkavalli, Simon Chen, Sohrab Amirghodsi, Sarah Kong, and Zhangyang Wang. Ssh: A self-supervised framework for image harmonization. In ICCV, 2021. +[9] Wenyan Cong, Xinhao Tao, Li Niu, Jing Liang, Xuesong Gao, Qihao Sun, and Liqing Zhang. High-resolution image harmonization via collaborative dual transformations. 2022. +[10] Jeya Maria Jose Valanarasu, He Zhang, Jianming Zhang, Yilin Wang, Zhe Lin, Jose Echevarria, Yinglan Ma, Zijun Wei, Kalyan Sunkavalli, and Vishal M Patel. Interactive portrait harmonization. 2022. +[11] Lingzhi Zhang, Tarmily Wen, and Jianbo Shi. Deep image blending. In The IEEE Winter Conference on Applications of Computer Vision, pages 231-240, 2020. +[12] Dave Epstein, Allan Jabri, Ben Poole, Alexei A. Efros, and Aleksander Holynski. Diffusion self-guidance for controllable image generation. arXiv preprint arXiv:2306.00986, 2023. 1 +[13] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. 2021. 1, 3 +[14] Tim Brooks, Aleksander Holynski, and Alexei A Efros. Instructpix2pix: Learning to follow image editing instructions. 2022. 1, 2, 3, 6 +[15] Chenlin Meng, Yutong He, Yang Song, Jiaming Song, Jia-jun Wu, Jun-Yan Zhu, and Stefano Ermon. SDEdit: Guided image synthesis and editing with stochastic differential equations. In ICLR, 2022. 1, 3 +[16] Yujun Shi, Chuhui Xue, Jiachun Pan, Wenqing Zhang, Vincent YF Tan, and Song Bai. Dragdiffusion: Harnessing diffusion models for interactive point-based image editing. 2023. 1, 3, 5, 6 +[17] Xingang Pan, Ayush Tewari, Thomas Leimkuhler, Lingjie Liu, Abhimitra Meka, and Christian Theobalt. Drag your + +gan: Interactive point-based manipulation on the generative image manifold. In SIGGRAPH, 2023. 1, 2, 3, 5, 6, 8 +[18] Guillaume Couairon, Jakob Verbeek, Holger Schwenk, and Matthieu Cord. Diffedit: Diffusion-based semantic image editing with mask guidance, 2022. +[19] Chong Mou, Xintao Wang, Jiechong Song, Ying Shan, and Jian Zhang. Dragondiffusion: Enabling drag-style manipulation on diffusion models. 2023. 1, 2, 3, 6 +[20] Binxin Yang, Shuyang Gu, Bo Zhang, Ting Zhang, Xuejin Chen, Xiaoyan Sun, Dong Chen, and Fang Wen. Paint by example: Exemplar-based image editing with diffusion models. arXiv preprint arXiv:2211.13227, 2022. +[21] Adding Conditional Control to Text-to-Image Diffusion Models, 2023. 3 +[22] GLIGEN: Open-Set Grounded Text-to-Image Generation, 2023. arXiv:2301.07093. 1 +[23] Or Patashnik, Zongze Wu, Eli Shechtman, Daniel Cohen-Or, and Dani Lischinski. Styleclip: Text-driven manipulation of stylegan imagery. In ICCV, 2021. 1, 2, 3, 6 +[24] Bahjat Kawar, Shiran Zada, Oran Lang, Omer Tov, Huiwen Chang, Tali Dekel, Inbar Mosseri, and Michal Irani. Imagic: Text-based real image editing with diffusion models. In CVPR, 2023. 1, 3 +[25] Yuki Endo. User-controllable latent transformer for stylegan image layout editing. Computer Graphics Forum, 2022. 1, 2, 3, 5, 6 +[26] Erik Härkönen, Aaron Hertzmann, Jaakko Lehtinen, and Sylvain Paris. Ganspace: Discovering interpretable gan controls. In NeurIPS, 2020. 2, 3, 4 +[27] Yujun Shen, Jinjin Gu, Xiaou Tang, and Bolei Zhou. Interpreting the latent space of gans for semantic face editing. In CVPR, 2020. +[28] Yujun Shen and Bolei Zhou. Closed-form factorization of latent semantics in gans. In CVPR, 2021. 3, 4 +[29] Yotam Nitzan, Rinon Gal, Ofir Brenner, and Daniel Cohen-Or. Large: Latent-based regression through gan semantics. In CVPR, 2022. 2, 4 +[30] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. 2019. 2, 4, 5 +[31] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of StyleGAN. In CVPR, 2020. 2, 3, 4 +[32] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. In NeurIPS, 2014. 2 +[33] Yann LeCun. The mnist database of handwritten digits. http://yann.lecun.com/exdb/mnist/, 1998. 2 +[34] Andrew Brock, Jeff Donahue, and Karen Simonyan. Large scale gan training for high fidelity natural image synthesis. arXiv, 2018. 2 +[35] Terro Karras, Timo Aila, Samuli Laine, and Jaakko Lehtinen. Progressive growing of gans for improved quality, stability, and variation. In ICLR, 2018. 2 +[36] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR, 2009. 2 + +[37] Fisher Yu, Yinda Zhang, Shuran Song, Ari Seff, and Jianxiong Xiao. Lsun: Construction of a large-scale image dataset using deep learning with humans in the loop. 2015. 2 +[38] Ali Jahanian, Lucy Chai, and Phillip Isola. On the "steerability" of generative adversarial networks. In ICLR, 2020. 2 +[39] David Bau, Jun-Yan Zhu, Hendrik Strobelt, Agata Lapedriza, Bolei Zhou, and Antonio Torralba. Understanding the role of individual units in a deep neural network. Proceedings of the National Academy of Sciences, 2020. 2 +[40] Zongze Wu, Dani Lischinski, and Eli Shechtman. Stylespace analysis: Disentangled controls for stylegan image generation. In ICLR, 2020. 3 +[41] Tengfei Wang, Yong Zhang, Yanbo Fan, Jue Wang, and Qifeng Chen. High-fidelity gan inversion for image attribute editing. In CVPR, 2022. 3 +[42] Omer Tov, Yuval Alaluf, Yotam Nitzan, Or Patashnik, and Daniel Cohen-Or. Designing an encoder for stylegan image manipulation. In ACM Transactions on Graphics, 2021. 6 +[43] Daniel Roich, Ron Mokady, Amit H Bermano, and Daniel Cohen-Or. Pivotal tuning for latent-based editing of real images. ACM Trans. Graph., 2021. 3 +[44] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. 2021. 3 +[45] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. 2021. 3 +[46] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In H. Larochelle, M. Ranzato, R. Hadsell, M.F. Balcan, and H. Lin, editors, Advances in Neural Information Processing Systems. Curran Associates, Inc., 2020. 3 +[47] Ron Mokady, Amir Hertz, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Null-text inversion for editing real images using guided diffusion models. arXiv preprint arXiv:2211.09794, 2022. 3 +[48] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-prompt image editing with cross attention control. arXiv preprint arXiv:2208.01626, 2022. 3 +[49] Huan Ling, Karsten Kreis, Daiqing Li, Seung Wook Kim, Antonio Torralba, and Sanja Fidler. Editgan: High-precision semantic image editing. In Advances in Neural Information Processing Systems (NeurIPS), 2021. 3 +[50] David Bau, Steven Liu, Tongzhou Wang, Jun-Yan Zhu, and Antonio Torralba. Rewriting a deep generative model. In ECCV, 2020. 3 +[51] Amir Bar, Yossi Gandelsman, Trevor Darrell, Amir Globerson, and Alexei A. Efros. Visual prompting via image inpainting. 2022. 3 +[52] Thao Nguyen, Yuheng Li, Utkarsh Ojha, and Yong Jae Lee. Visual instruction inversion: Image editing via visual prompting. In NeurIPS, 2023. + +[53] Yasheng Sun, Yifan Yang, Houwen Peng, Yifei Shen, Yuqing Yang, Han Hu, Lili Qiu, and Hideki Koike. Imagebrush: Learning visual in-context instructions for exemplar-based image manipulation. 2023. +[54] Aaron Hertzmann, Charles E. Jacobs, Nuria Oliver, Brian Curless, and David H. Salesin. Image analogies. In Seminal Graphics Papers: Pushing the Boundaries, Volume 2, 2023. +[55] Jing Liao, Yuan Yao, Lu Yuan, Gang Hua, and Sing Bing Kang. Visual attribute transfer through deep image analogy. 2017. 3 +[56] Weihao Xia, Yulun Zhang, Yujiu Yang, Jing-Hao Xue, Bolei Zhou, and Ming-Hsuan Yang. Gan inversion: A survey. TPAMI, 2022. 3 +[57] Yunjey Choi, Youngjung Uh, Jaejun Yoo, and Jung-Woo Ha. Starganv2: Diverse image synthesis for multiple domains. In CVPR, 2020. 5 +[58] Tero Karras, Miika Aittala, Janne Hellsten, Samuli Laine, Jaakko Lehtinen, and Timo Aila. Training generative adversarial networks with limited data, 2020. 5 +[59] Jianglin Fu, Shikai Li, Yuming Jiang, Kwan-Yee Lin, Chen Qian, Chen-Change Loy, Wayne Wu, and Ziwei Liu. Stylegan-human: A data-centric odyssey of human generation. arXiv preprint, arXiv:2204.11823, 2022. 5 +[60] Davis E. King. Dlib-ml: A machine learning toolkit. J. Mach. Learn. Res., 2009. 6 +[61] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, Gunter Klambauer, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. In NeurIPS, 2017. 7 +[62] Thorsten Hempel, Ahmed A. Abdelrahman, and Ayoub Al-Hamadi. 6d rotation representation for unconstrained head pose estimation. In ICIP, 2022. 7 +[63] William Peebles, Jun-Yan Zhu, Richard Zhang, Antonio Torralba, Alexei Efros, and Eli Shechtman. Gan-supervised dense visual alignment. In CVPR, 2022. 8 \ No newline at end of file diff --git a/2024/Edit One for All_ Interactive Batch Image Editing/images.zip b/2024/Edit One for All_ Interactive Batch Image Editing/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..7b1fa6fb8162ec63f7c4cf9e99944b8759cbc4bd --- /dev/null +++ b/2024/Edit One for All_ Interactive Batch Image Editing/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6b617419e3b12d74a2e70672cddc108d375f3fd35f7c61eda79664647b020359 +size 918408 diff --git a/2024/Edit One for All_ Interactive Batch Image Editing/layout.json b/2024/Edit One for All_ Interactive Batch Image Editing/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..d9bfe65cec8c7fb5f3ae2da1bfb375efbe1c7d20 --- /dev/null +++ b/2024/Edit One for All_ Interactive Batch Image Editing/layout.json @@ -0,0 +1,9175 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 143, + 103, + 451, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 103, + 451, + 121 + ], + "spans": [ + { + "bbox": [ + 143, + 103, + 451, + 121 + ], + "type": "text", + "content": "Edit One for All: Interactive Batch Image Editing" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 115, + 143, + 478, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 115, + 143, + 478, + 171 + ], + "spans": [ + { + "bbox": [ + 115, + 143, + 478, + 171 + ], + "type": "text", + "content": "Thao Nguyen Utkarsh Ojha Yuheng Li Haotian Liu Yong Jae Lee University of Wisconsin-Madison" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 173, + 174, + 414, + 184 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 173, + 174, + 414, + 184 + ], + "spans": [ + { + "bbox": [ + 173, + 174, + 414, + 184 + ], + "type": "text", + "content": "https://thaoshibe.github.io/edit-one-for-all" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "spans": [ + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 233, + 290, + 461 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 233, + 290, + 461 + ], + "spans": [ + { + "bbox": [ + 47, + 233, + 290, + 461 + ], + "type": "text", + "content": "In recent years, image editing has advanced remarkably. With increased human control, it is now possible to edit an image in a plethora of ways; from specifying in text what we want to change, to straight up dragging the contents of the image in an interactive point-based manner. However, most of the focus has remained on editing single images at a time. Whether and how we can simultaneously edit large batches of images has remained understudied. With the goal of minimizing human supervision in the editing process, this paper presents a novel method for interactive batch image editing using StyleGAN as the medium. Given an edit specified by users in an example image (e.g., make the face frontal), our method can automatically transfer that edit to other test images, so that regardless of their initial state (pose), they all arrive at the same final state (e.g., all facing front). Extensive experiments demonstrate that edits performed using our method have similar visual quality to existing single-image-editing methods, while having more visual consistency and saving significant time and human effort." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 478, + 128, + 491 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 478, + 128, + 491 + ], + "spans": [ + { + "bbox": [ + 47, + 478, + 128, + 491 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 498, + 287, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 498, + 287, + 665 + ], + "spans": [ + { + "bbox": [ + 46, + 498, + 287, + 665 + ], + "type": "text", + "content": "Image editing has undergone a transformation in recent years with the aid of modern generative models. The process has been greatly democratized, where many of the sophisticated edits, which previously might have required hours and niche expertise, can now be completed with relative ease in a matter of minutes. For example, different types of learning based algorithms can be used for image correction/adjustment [1-12] and for manipulating the semantic contents of real images [13-22]. Moreover, there are many different ways to edit an image. For instance, a user can specify in text what changes they want - e.g., \"make the hair darker\" [14, 15, 23, 24], or they can drag the contents of the image in an interactive manner to shrink, enlarge, or move a part of an object [12, 16, 17, 19, 25]." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "text", + "content": "A common theme across many of these works is that the edits are designed to work for a single image at a time. For example, given an image of a cat with open eyes, DragGAN [17] allows the user to drag the contents of that par" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 310, + 212, + 541, + 445 + ], + "blocks": [ + { + "bbox": [ + 310, + 212, + 541, + 445 + ], + "lines": [ + { + "bbox": [ + 310, + 212, + 541, + 445 + ], + "spans": [ + { + "bbox": [ + 310, + 212, + 541, + 445 + ], + "type": "image", + "image_path": "0eb91e7d6b9d8e795b992b5d273fc19a0ad33d6f0a5c2b84cecc4e7dd125563f.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 448, + 545, + 492 + ], + "lines": [ + { + "bbox": [ + 305, + 448, + 545, + 492 + ], + "spans": [ + { + "bbox": [ + 305, + 448, + 545, + 492 + ], + "type": "text", + "content": "Figure 1. Interactive Batch Image Editing. Given a single user edited image, the goal is to automatically transfer that edit to new unseen images, so that all edited images end up with the same final state as the user edited image." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 508, + 545, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 508, + 545, + 592 + ], + "spans": [ + { + "bbox": [ + 304, + 508, + 545, + 592 + ], + "type": "text", + "content": "ticular image so that both the eyes can be closed in the resulting cat image. But what if we wish to apply the same edit to many different types of cats so that we can close all of their eyes? While one could perform dragging on each cat separately, the whole process would be quite cumbersome (requiring lots of human annotation) and time intensive (requiring optimization for each image)." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 594, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 546, + 713 + ], + "type": "text", + "content": "In this paper, we introduce the new problem of Interactive Batch Image Editing. Given a user-specified edit " + }, + { + "bbox": [ + 304, + 594, + 546, + 713 + ], + "type": "inline_equation", + "content": "I \\rightarrow I'" + }, + { + "bbox": [ + 304, + 594, + 546, + 713 + ], + "type": "text", + "content": ", the goal is to automatically transfer that edit to new unseen images, so that all edited images end up with the same final state as " + }, + { + "bbox": [ + 304, + 594, + 546, + 713 + ], + "type": "inline_equation", + "content": "I'" + }, + { + "bbox": [ + 304, + 594, + 546, + 713 + ], + "type": "text", + "content": " (e.g., all cats with eyes closed) regardless of their original starting states (e.g., varying degrees of original eye openness); see Figure 1. Hence, a solution to this problem requires two components: (i) modeling the user edit in the example pair " + }, + { + "bbox": [ + 304, + 594, + 546, + 713 + ], + "type": "inline_equation", + "content": "(I, I')" + }, + { + "bbox": [ + 304, + 594, + 546, + 713 + ], + "type": "text", + "content": " so that it can be transferred to new images; and (ii) controlling the degree of the" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 494, + 35 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 494, + 35 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 494, + 35 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "8271" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 83, + 274, + 205 + ], + "blocks": [ + { + "bbox": [ + 120, + 70, + 215, + 79 + ], + "lines": [ + { + "bbox": [ + 120, + 70, + 215, + 79 + ], + "spans": [ + { + "bbox": [ + 120, + 70, + 215, + 79 + ], + "type": "text", + "content": "edit one image at a time" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 56, + 83, + 274, + 205 + ], + "lines": [ + { + "bbox": [ + 56, + 83, + 274, + 205 + ], + "spans": [ + { + "bbox": [ + 56, + 83, + 274, + 205 + ], + "type": "image", + "image_path": "a2fbce35151b420551fee3a48bde474afebeb2ffd3af5f9c5decb05340526b8d.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 206, + 547, + 240 + ], + "lines": [ + { + "bbox": [ + 46, + 206, + 547, + 240 + ], + "spans": [ + { + "bbox": [ + 46, + 206, + 547, + 240 + ], + "type": "text", + "content": "Figure 2. Single Image Editing vs. Batch Image Editing. (a) Prior work (e.g., [17, 19, 25]) focuses on single image editing. (b) We focus on batch image editing, where the user's edit on a single image is automatically transferred to new images, so that they all arrive at the same final state regardless of their initial starting state. In this way, we can achieve time speed up and reduce human effort in editing." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 285, + 82, + 535, + 205 + ], + "blocks": [ + { + "bbox": [ + 314, + 70, + 511, + 80 + ], + "lines": [ + { + "bbox": [ + 314, + 70, + 511, + 80 + ], + "spans": [ + { + "bbox": [ + 314, + 70, + 511, + 80 + ], + "type": "text", + "content": "edit one image, transfer to others at the same time!" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 285, + 82, + 535, + 205 + ], + "lines": [ + { + "bbox": [ + 285, + 82, + 535, + 205 + ], + "spans": [ + { + "bbox": [ + 285, + 82, + 535, + 205 + ], + "type": "image", + "image_path": "2ddb6c5d6e47094d0816beaae5e12608cb13fb44ebde5f8a66a4d9a97f06ab26.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 252, + 266, + 264 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 252, + 266, + 264 + ], + "spans": [ + { + "bbox": [ + 46, + 252, + 266, + 264 + ], + "type": "text", + "content": "edit so that all edited images have the same final state." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 266, + 287, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 266, + 287, + 445 + ], + "spans": [ + { + "bbox": [ + 46, + 266, + 287, + 445 + ], + "type": "text", + "content": "To model the user's edit, we build upon existing literature on Generative Adversarial Networks (GANs) [23, 26-29], and in particular StyleGAN [30, 31], which shows that their learned latent space emits globally linear editing directions where the edited attribute (e.g., eye closeness) varies linearly in magnitude along such a direction for all semantically related instances (e.g., all closed eyes cat). To try to discover the global direction corresponding to a user edit, we perform optimization in the latent space of StyleGAN2 [31] so that (i) the edited image with the discovered direction is visually similar to the original edited image " + }, + { + "bbox": [ + 46, + 266, + 287, + 445 + ], + "type": "inline_equation", + "content": "I'" + }, + { + "bbox": [ + 46, + 266, + 287, + 445 + ], + "type": "text", + "content": ", and (ii) the linear correlation between the distance along the direction and the strength of the visual attribute is highest (e.g., going twice the amount in that direction closes a cat's eyes by twice the amount)." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 448, + 287, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 448, + 287, + 639 + ], + "spans": [ + { + "bbox": [ + 46, + 448, + 287, + 639 + ], + "type": "text", + "content": "To control the degree of the edit so that all edited images end up with the same final state, we derive a closed-form solution. Taking motivation from [29], we model the attribute strength (e.g., how much the cat's eyes are open) as the distance along the normal vector from a hyperplane. By setting the example edited image as lying on that hyperplane (i.e., its distance is 0), the objective is to move all other images along that same direction so that their distances to the hyperplane also become 0. In this way, regardless of the start state of any image (e.g., varying degree of eye openness), their edited end states all become the same (e.g., closed eyes). We show that we can analytically compute the exact amount of traversal along the direction for any image with a closed-form solution. Editing images in this way helps improve the results visually, as the final state of any new image matches that of the edited example given by the user." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 642, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 642, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 642, + 287, + 715 + ], + "type": "text", + "content": "Our method works with various edits given by an editing framework, e.g., it can be a change given by a dragging operation [17], or it can be a change given by a text-based edit [14, 23] (e.g., \"make eyes bigger\"), as long as it is a direction compatible with a StyleGAN2 [31] model. We present qualitative and" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 252, + 545, + 419 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 252, + 545, + 419 + ], + "spans": [ + { + "bbox": [ + 304, + 252, + 545, + 419 + ], + "type": "text", + "content": "quantitative results in transferring edits for a variety of objects (cats/dogs/faces/humans/limons/arts/etc.), parts (mouths/ears/legs/etc.), and corresponding attributes (big eyes/short faces/pale skins/etc.). Importantly, we show that the final state of the edited images is comparable to the scenario in which a user performs the edit for each image separately. But since our method does this automatically (see Fig. 2), we show that it takes much less time (e.g., only 0.05s per image, compared to DragGAN [17] about 2s per image), and does not require laborious human annotation (e.g., only need 1 annotated image, compared to 4 required for DragGAN [17]). Finally, we show practical applications of batch image editing; e.g., changing wheel size in every photo of a car collection by editing just one." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 420, + 546, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 420, + 546, + 540 + ], + "spans": [ + { + "bbox": [ + 304, + 420, + 546, + 540 + ], + "type": "text", + "content": "In sum, our contributions are: (1) We introduce the novel problem of interactive batch image editing, wherein a user-driven edit is automatically transferred to other similar test images. (2) We study what having the same final state means in a geometrical sense, and propose a principled approach to achieving that in StyleGAN2 [31]'s latent space. (3) We show that our method works on a wide variety of domains e.g., cats, dogs, humans etc., taking significantly less time than editing each image individually while having similar visual quality and more visual consistency." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 550, + 392, + 562 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 550, + 392, + 562 + ], + "spans": [ + { + "bbox": [ + 306, + 550, + 392, + 562 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "text", + "content": "Latent space of GANs. The capabilities of generative adversarial networks (GANs) have transformed drastically, from the first GAN [32] designed for simple image datasets like MNIST [33] to big and powerful models like BigGAN [34], Progressive GAN [35] and StyleGANs [30, 31] designed for much more complex datasets [36, 37]. Parallel to the efforts to enhance their capabilities, there has also been work done to better understand and make use of their learned latent space. Specifically, there is a line of work which tries to find interpretable directions in that space, so that moving in such a direction majorly changes only one discernible attribute in the image. Some methods [38, 39]" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "8272" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 49, + 70, + 156, + 213 + ], + "blocks": [ + { + "bbox": [ + 49, + 70, + 156, + 213 + ], + "lines": [ + { + "bbox": [ + 49, + 70, + 156, + 213 + ], + "spans": [ + { + "bbox": [ + 49, + 70, + 156, + 213 + ], + "type": "image", + "image_path": "ed94eafe60804ef8f989e160c3fc16fc755bfb18cce8c08b180a96a1fa5d073d.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 82, + 215, + 123, + 224 + ], + "lines": [ + { + "bbox": [ + 82, + 215, + 123, + 224 + ], + "spans": [ + { + "bbox": [ + 82, + 215, + 123, + 224 + ], + "type": "text", + "content": "(a) Setting" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 165, + 69, + 248, + 213 + ], + "blocks": [ + { + "bbox": [ + 165, + 69, + 248, + 213 + ], + "lines": [ + { + "bbox": [ + 165, + 69, + 248, + 213 + ], + "spans": [ + { + "bbox": [ + 165, + 69, + 248, + 213 + ], + "type": "image", + "image_path": "8dd88c9fcd19d6b656d62f993e17a4355cd8639d14d75990bae421b3283904c9.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 166, + 215, + 241, + 224 + ], + "lines": [ + { + "bbox": [ + 166, + 215, + 241, + 224 + ], + "spans": [ + { + "bbox": [ + 166, + 215, + 241, + 224 + ], + "type": "text", + "content": "(b) Naïve Approach" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 258, + 70, + 379, + 213 + ], + "blocks": [ + { + "bbox": [ + 258, + 70, + 379, + 213 + ], + "lines": [ + { + "bbox": [ + 258, + 70, + 379, + 213 + ], + "spans": [ + { + "bbox": [ + 258, + 70, + 379, + 213 + ], + "type": "image", + "image_path": "cf3b505c6e9ae51a1836b378c628907361c396818e3a871ec748313ca0e39026.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 257, + 215, + 376, + 225 + ], + "lines": [ + { + "bbox": [ + 257, + 215, + 376, + 225 + ], + "spans": [ + { + "bbox": [ + 257, + 215, + 376, + 225 + ], + "type": "text", + "content": "(c) Optimizing Editing Direction" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 228, + 547, + 262 + ], + "lines": [ + { + "bbox": [ + 46, + 228, + 547, + 262 + ], + "spans": [ + { + "bbox": [ + 46, + 228, + 547, + 262 + ], + "type": "text", + "content": "Figure 3. Different editing strategies. (a) Setting. (b) Naive Approach: The editing direction effective for an example may not generalize well to test images. (c) Optimizing Editing Direction: We optimize for a globally consistent direction that is effective for both example and test images. (d) Adjusting Editing Strength: Ensuring consistent final states requires adjusting the editing strength for each test image." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 383, + 69, + 545, + 213 + ], + "blocks": [ + { + "bbox": [ + 383, + 69, + 545, + 213 + ], + "lines": [ + { + "bbox": [ + 383, + 69, + 545, + 213 + ], + "spans": [ + { + "bbox": [ + 383, + 69, + 545, + 213 + ], + "type": "image", + "image_path": "a4a06c1b516757ff2a327010b4edc31624f4671b2108a1a73a8da292a17eea67.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 412, + 215, + 512, + 225 + ], + "lines": [ + { + "bbox": [ + 412, + 215, + 512, + 225 + ], + "spans": [ + { + "bbox": [ + 412, + 215, + 512, + 225 + ], + "type": "text", + "content": "(d) Adjusting Edit Strength" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 274, + 289, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 274, + 289, + 441 + ], + "spans": [ + { + "bbox": [ + 46, + 274, + 289, + 441 + ], + "type": "text", + "content": "try to find such directions or particular activations using (self) supervised methods. Others [26-28] try getting rid of the need for supervision: discovering the directions representing the most important factors of variation, whatever they may be. Going beyond the " + }, + { + "bbox": [ + 46, + 274, + 289, + 441 + ], + "type": "inline_equation", + "content": "\\mathcal{W}" + }, + { + "bbox": [ + 46, + 274, + 289, + 441 + ], + "type": "text", + "content": " space, the authors in [40] explore the " + }, + { + "bbox": [ + 46, + 274, + 289, + 441 + ], + "type": "inline_equation", + "content": "\\mathcal{W}+" + }, + { + "bbox": [ + 46, + 274, + 289, + 441 + ], + "type": "text", + "content": " space, and show that it has even better disentanglement useful for spatial image editing. All of these hidden capabilities have made StyleGANs a useful tool for editing purposes. However, prior work mainly focuses on the effectiveness of the editing direction (e.g., whether the discovered direction can change yaw pose). The question of whether these directions can be applied to diverse images to yield consistent results (e.g., all faces facing frontal) and, if so, how to achieving such consistency, remains unanswered." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 451, + 289, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 451, + 289, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 451, + 289, + 714 + ], + "type": "text", + "content": "Image editing with generative models. Researchers have used StyleGAN for real image editing by designing an encoder to invert a real image into StyleGAN's latent space [41-43]. StyleCLIP [23] presents a way to perform image editing through a text-based interface by making use of the CLIP encoder [44]. Very recently, StyleGAN has also been used to perform point-based editing [17, 25] so as to move any start point to reach a target location in the image; thereby elongating, rotating, shifting the objects. Most work on GANs has been for single object category datasets, but with the rise of text-to-image Diffusion Models which can generate complex images [13, 45, 46], editing such images is now possible. Complex scenes can be edited either using text [14, 15, 21, 24, 47, 48], or using the same idea of point-based manipulation [16, 19]. There have been some works which discuss the possibility to transfer the edit of one image to another. EditGAN [49] uses segmentation mask manipulations to edit an image, but to successfully transfer them to an unseen image, it needs to do post hoc manipulation of the editing scale. RewriteGAN [50] involves editing one generated sample; e.g., adding a patch of a tree onto a church tower. Following this, the rules of the" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 274, + 547, + 382 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 274, + 547, + 382 + ], + "spans": [ + { + "bbox": [ + 304, + 274, + 547, + 382 + ], + "type": "text", + "content": "GAN are manipulated so that all churches have some tree on the top. Within diffusion models, visual prompting/ image analogies tries something similar, where users can define a triplet {before, after, test} to learn the edit and transfer it to a test image [51-55]. Our task is similar, in that we wish to transfer user-edits from the training example to new images, but it differs in one crucial aspect: along with transferring the edit, we wish to automatically learn its strength so that the edit produces the same final state for a new image." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 392, + 372, + 407 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 392, + 372, + 407 + ], + "spans": [ + { + "bbox": [ + 306, + 392, + 372, + 407 + ], + "type": "text", + "content": "3. Approach" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 413, + 547, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 413, + 547, + 472 + ], + "spans": [ + { + "bbox": [ + 304, + 413, + 547, + 472 + ], + "type": "text", + "content": "Our focus is on the setting in which an image editing process edits an attribute of an image so that the attribute's value reaches a desired state; e.g., rotating the pose of a face looking sideways so that it faces front (attribute = face's pose, value = front)." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 472, + 545, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 472, + 545, + 593 + ], + "spans": [ + { + "bbox": [ + 304, + 472, + 545, + 593 + ], + "type": "text", + "content": "With this view, we explain our framework for batch image editing, which can be broken down into two stages: (i) A user edits an image " + }, + { + "bbox": [ + 304, + 472, + 545, + 593 + ], + "type": "inline_equation", + "content": "I_0" + }, + { + "bbox": [ + 304, + 472, + 545, + 593 + ], + "type": "text", + "content": " (e.g., using DragGAN [17]) to obtain an edited image " + }, + { + "bbox": [ + 304, + 472, + 545, + 593 + ], + "type": "inline_equation", + "content": "I_0'" + }, + { + "bbox": [ + 304, + 472, + 545, + 593 + ], + "type": "text", + "content": ". We describe in Sec. 3.1 how we capture this user edit " + }, + { + "bbox": [ + 304, + 472, + 545, + 593 + ], + "type": "inline_equation", + "content": "I_0 \\rightarrow I_0'" + }, + { + "bbox": [ + 304, + 472, + 545, + 593 + ], + "type": "text", + "content": ", so that the same edit can be applied to new images. (ii) Next, we describe in Sec. 3.2, how for any new image, we apply the modeled edit by automatically adjusting its strength so that the attribute's value in this new image matches that of " + }, + { + "bbox": [ + 304, + 472, + 545, + 593 + ], + "type": "inline_equation", + "content": "I_0'" + }, + { + "bbox": [ + 304, + 472, + 545, + 593 + ], + "type": "text", + "content": " (e.g., any face, regardless of its initial pose, now faces front after the edit)." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 599, + 439, + 612 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 599, + 439, + 612 + ], + "spans": [ + { + "bbox": [ + 306, + 599, + 439, + 612 + ], + "type": "text", + "content": "3.1. Modeling the User Edit" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 617, + 545, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 617, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 617, + 545, + 715 + ], + "type": "text", + "content": "We start with an image " + }, + { + "bbox": [ + 304, + 617, + 545, + 715 + ], + "type": "inline_equation", + "content": "I_0" + }, + { + "bbox": [ + 304, + 617, + 545, + 715 + ], + "type": "text", + "content": ". This image could be a real image or a generated one. Either way, we get its latent representation in the " + }, + { + "bbox": [ + 304, + 617, + 545, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{W}" + }, + { + "bbox": [ + 304, + 617, + 545, + 715 + ], + "type": "text", + "content": " space of a StyleGAN2 [31] model " + }, + { + "bbox": [ + 304, + 617, + 545, + 715 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 304, + 617, + 545, + 715 + ], + "type": "text", + "content": ", so that " + }, + { + "bbox": [ + 304, + 617, + 545, + 715 + ], + "type": "inline_equation", + "content": "I_0 = G(w_0)" + }, + { + "bbox": [ + 304, + 617, + 545, + 715 + ], + "type": "text", + "content": ". (For real images we can use GAN-inversion techniques [56].) The user, with the help of an image editing framework (e.g., DragGAN [17] or InstructPix2Pix [14]), edits " + }, + { + "bbox": [ + 304, + 617, + 545, + 715 + ], + "type": "inline_equation", + "content": "I_0" + }, + { + "bbox": [ + 304, + 617, + 545, + 715 + ], + "type": "text", + "content": " to manipulate one or more of its attributes. The edit maps to the " + }, + { + "bbox": [ + 304, + 617, + 545, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{W}" + }, + { + "bbox": [ + 304, + 617, + 545, + 715 + ], + "type": "text", + "content": " space as " + }, + { + "bbox": [ + 304, + 617, + 545, + 715 + ], + "type": "inline_equation", + "content": "w_0 \\to w_0'" + }, + { + "bbox": [ + 304, + 617, + 545, + 715 + ], + "type": "text", + "content": ". The" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "8273" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 119 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 119 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 119 + ], + "type": "text", + "content": "resulting edited image can thus be recovered as follows: " + }, + { + "bbox": [ + 46, + 72, + 287, + 119 + ], + "type": "inline_equation", + "content": "I_0' = G(w_0')" + }, + { + "bbox": [ + 46, + 72, + 287, + 119 + ], + "type": "text", + "content": ". Fig. 3(a) shows an example of the original and edited image pair, " + }, + { + "bbox": [ + 46, + 72, + 287, + 119 + ], + "type": "inline_equation", + "content": "(I_0, I_0')" + }, + { + "bbox": [ + 46, + 72, + 287, + 119 + ], + "type": "text", + "content": ", where the user intended to turn the face forward." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 120, + 287, + 276 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 120, + 287, + 276 + ], + "spans": [ + { + "bbox": [ + 46, + 120, + 287, + 276 + ], + "type": "text", + "content": "Now, given the user edit " + }, + { + "bbox": [ + 46, + 120, + 287, + 276 + ], + "type": "inline_equation", + "content": "I_0 \\to I_0'" + }, + { + "bbox": [ + 46, + 120, + 287, + 276 + ], + "type": "text", + "content": ", we wish to capture it in a way that can be applied to new images in a generalizable manner; i.e., the application of the edit changes the same property in a new image. This is where a nice property of GANs, and in particular StyleGANs [30, 31] comes in useful. It has been shown in prior works [26-28] that it is possible to discover directions with such properties (using supervised as well as unsupervised methods) in the learned " + }, + { + "bbox": [ + 46, + 120, + 287, + 276 + ], + "type": "inline_equation", + "content": "\\mathcal{W}" + }, + { + "bbox": [ + 46, + 120, + 287, + 276 + ], + "type": "text", + "content": " space. In particular, it is possible to find directions " + }, + { + "bbox": [ + 46, + 120, + 287, + 276 + ], + "type": "inline_equation", + "content": "(\\Delta_w)" + }, + { + "bbox": [ + 46, + 120, + 287, + 276 + ], + "type": "text", + "content": " that are globally consistent. Taking motivation from [29], we define a globally consistent direction " + }, + { + "bbox": [ + 46, + 120, + 287, + 276 + ], + "type": "inline_equation", + "content": "\\Delta_w" + }, + { + "bbox": [ + 46, + 120, + 287, + 276 + ], + "type": "text", + "content": " as the following: for any arbitrary " + }, + { + "bbox": [ + 46, + 120, + 287, + 276 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 46, + 120, + 287, + 276 + ], + "type": "text", + "content": ", moving along " + }, + { + "bbox": [ + 46, + 120, + 287, + 276 + ], + "type": "inline_equation", + "content": "\\Delta_w" + }, + { + "bbox": [ + 46, + 120, + 287, + 276 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 120, + 287, + 276 + ], + "type": "inline_equation", + "content": "w \\to w + \\Delta_w" + }, + { + "bbox": [ + 46, + 120, + 287, + 276 + ], + "type": "text", + "content": " (i) changes the same attribute, and (ii) by the same amount." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 276, + 287, + 395 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 276, + 287, + 395 + ], + "spans": [ + { + "bbox": [ + 46, + 276, + 287, + 395 + ], + "type": "text", + "content": "To make the precise user edit " + }, + { + "bbox": [ + 46, + 276, + 287, + 395 + ], + "type": "inline_equation", + "content": "I_0 \\to I_0'" + }, + { + "bbox": [ + 46, + 276, + 287, + 395 + ], + "type": "text", + "content": " applicable to other images, it needs to be captured as a globally consistent direction. The naive way to represent that edit will be through a simple difference in the " + }, + { + "bbox": [ + 46, + 276, + 287, + 395 + ], + "type": "inline_equation", + "content": "\\mathcal{W}" + }, + { + "bbox": [ + 46, + 276, + 287, + 395 + ], + "type": "text", + "content": " space: " + }, + { + "bbox": [ + 46, + 276, + 287, + 395 + ], + "type": "inline_equation", + "content": "\\Delta_w = w_0' - w_0" + }, + { + "bbox": [ + 46, + 276, + 287, + 395 + ], + "type": "text", + "content": ". However, empirically, and as we show in Fig. 3(b), applying this " + }, + { + "bbox": [ + 46, + 276, + 287, + 395 + ], + "type": "inline_equation", + "content": "\\Delta_w" + }, + { + "bbox": [ + 46, + 276, + 287, + 395 + ], + "type": "text", + "content": " to the latent code " + }, + { + "bbox": [ + 46, + 276, + 287, + 395 + ], + "type": "inline_equation", + "content": "w_1" + }, + { + "bbox": [ + 46, + 276, + 287, + 395 + ], + "type": "text", + "content": " corresponding to a new image " + }, + { + "bbox": [ + 46, + 276, + 287, + 395 + ], + "type": "inline_equation", + "content": "I_1" + }, + { + "bbox": [ + 46, + 276, + 287, + 395 + ], + "type": "text", + "content": " does not always result in the same change; while " + }, + { + "bbox": [ + 46, + 276, + 287, + 395 + ], + "type": "inline_equation", + "content": "I_0 \\to I_0'" + }, + { + "bbox": [ + 46, + 276, + 287, + 395 + ], + "type": "text", + "content": " results in a pose change of " + }, + { + "bbox": [ + 46, + 276, + 287, + 395 + ], + "type": "inline_equation", + "content": "\\sim 30^\\circ" + }, + { + "bbox": [ + 46, + 276, + 287, + 395 + ], + "type": "text", + "content": " degrees in yaw, " + }, + { + "bbox": [ + 46, + 276, + 287, + 395 + ], + "type": "inline_equation", + "content": "I_1 \\to I_1'" + }, + { + "bbox": [ + 46, + 276, + 287, + 395 + ], + "type": "text", + "content": " does not seem to change the same attribute or at least not by the same amount." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 396, + 287, + 455 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 396, + 287, + 455 + ], + "spans": [ + { + "bbox": [ + 46, + 396, + 287, + 455 + ], + "type": "text", + "content": "Hence, our goal is to represent the " + }, + { + "bbox": [ + 46, + 396, + 287, + 455 + ], + "type": "inline_equation", + "content": "I_0 \\to I_0'" + }, + { + "bbox": [ + 46, + 396, + 287, + 455 + ], + "type": "text", + "content": " edit through a " + }, + { + "bbox": [ + 46, + 396, + 287, + 455 + ], + "type": "inline_equation", + "content": "\\mathcal{W}" + }, + { + "bbox": [ + 46, + 396, + 287, + 455 + ], + "type": "text", + "content": " space direction that can better satisfy both the properties of a globally consistent direction. For this task, we first take motivation from LARGE [29] to introduce a mathematical view of what those directions mean." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 462, + 287, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 462, + 287, + 498 + ], + "spans": [ + { + "bbox": [ + 46, + 462, + 287, + 498 + ], + "type": "text", + "content": "Globally consistent direction. Let's say that " + }, + { + "bbox": [ + 46, + 462, + 287, + 498 + ], + "type": "inline_equation", + "content": "\\Delta_g" + }, + { + "bbox": [ + 46, + 462, + 287, + 498 + ], + "type": "text", + "content": " is one such direction. Along with a bias term " + }, + { + "bbox": [ + 46, + 462, + 287, + 498 + ], + "type": "inline_equation", + "content": "b" + }, + { + "bbox": [ + 46, + 462, + 287, + 498 + ], + "type": "text", + "content": ", we can define a hyperplane as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 134, + 502, + 287, + 515 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 502, + 287, + 515 + ], + "spans": [ + { + "bbox": [ + 134, + 502, + 287, + 515 + ], + "type": "interline_equation", + "content": "w \\cdot \\Delta_ {g} + b = 0 \\tag {1}", + "image_path": "104a3d8aa40f3281fb1009fec66285464535af1a6dce49ee425de0266bca8c51.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 525, + 287, + 597 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 525, + 287, + 597 + ], + "spans": [ + { + "bbox": [ + 46, + 525, + 287, + 597 + ], + "type": "text", + "content": "That is, any point " + }, + { + "bbox": [ + 46, + 525, + 287, + 597 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 46, + 525, + 287, + 597 + ], + "type": "text", + "content": " which lies on the hyperplane will satisfy this equation. The authors in [29] argued that for such a hyperplane, whose normal vector is a globally consistent direction " + }, + { + "bbox": [ + 46, + 525, + 287, + 597 + ], + "type": "inline_equation", + "content": "(\\Delta_g)" + }, + { + "bbox": [ + 46, + 525, + 287, + 597 + ], + "type": "text", + "content": ", the distance of an arbitrary point " + }, + { + "bbox": [ + 46, + 525, + 287, + 597 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 46, + 525, + 287, + 597 + ], + "type": "text", + "content": " from that hyperplane " + }, + { + "bbox": [ + 46, + 525, + 287, + 597 + ], + "type": "inline_equation", + "content": "(w \\cdot \\Delta_g)" + }, + { + "bbox": [ + 46, + 525, + 287, + 597 + ], + "type": "text", + "content": " will be linearly correlated to the actual attribute " + }, + { + "bbox": [ + 46, + 525, + 287, + 597 + ], + "type": "inline_equation", + "content": "(y)" + }, + { + "bbox": [ + 46, + 525, + 287, + 597 + ], + "type": "text", + "content": " that results in the generated image:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 121, + 607, + 287, + 620 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 121, + 607, + 287, + 620 + ], + "spans": [ + { + "bbox": [ + 121, + 607, + 287, + 620 + ], + "type": "interline_equation", + "content": "y = a \\times (w \\cdot \\Delta_ {g}) + b \\tag {2}", + "image_path": "51e525cde19c6127f2bbbe9e03c7c38b73410b5118e28aff6f54c4df49376afb.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 629, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 629, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 629, + 287, + 714 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 46, + 629, + 287, + 714 + ], + "type": "inline_equation", + "content": "a" + }, + { + "bbox": [ + 46, + 629, + 287, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 629, + 287, + 714 + ], + "type": "inline_equation", + "content": "b" + }, + { + "bbox": [ + 46, + 629, + 287, + 714 + ], + "type": "text", + "content": " are unknown linear coefficients. For example, if " + }, + { + "bbox": [ + 46, + 629, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\Delta_g" + }, + { + "bbox": [ + 46, + 629, + 287, + 714 + ], + "type": "text", + "content": " corresponds to change in pose, then all the front facing people will lie at the same distance " + }, + { + "bbox": [ + 46, + 629, + 287, + 714 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 46, + 629, + 287, + 714 + ], + "type": "text", + "content": " from the above hyperplane. Because of this, any hyperplane defined with respect to a globally consistent direction can be viewed as a semantic hyperplane. For simplicity, we can set " + }, + { + "bbox": [ + 46, + 629, + 287, + 714 + ], + "type": "inline_equation", + "content": "d = 0" + }, + { + "bbox": [ + 46, + 629, + 287, + 714 + ], + "type": "text", + "content": " for the edited image " + }, + { + "bbox": [ + 46, + 629, + 287, + 714 + ], + "type": "inline_equation", + "content": "I_0'" + }, + { + "bbox": [ + 46, + 629, + 287, + 714 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 72, + 545, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 180 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 180 + ], + "type": "text", + "content": "Given that the original " + }, + { + "bbox": [ + 304, + 72, + 545, + 180 + ], + "type": "inline_equation", + "content": "\\Delta_w" + }, + { + "bbox": [ + 304, + 72, + 545, + 180 + ], + "type": "text", + "content": " may not always be globally consistent, we aim to discover, through optimization, a different direction " + }, + { + "bbox": [ + 304, + 72, + 545, + 180 + ], + "type": "inline_equation", + "content": "\\Delta_w^*" + }, + { + "bbox": [ + 304, + 72, + 545, + 180 + ], + "type": "text", + "content": " which produces a similar editing effect, but is more likely to be a globally consistent one. We initialize " + }, + { + "bbox": [ + 304, + 72, + 545, + 180 + ], + "type": "inline_equation", + "content": "\\Delta_w^*" + }, + { + "bbox": [ + 304, + 72, + 545, + 180 + ], + "type": "text", + "content": " with 0's, and design two objective functions to optimize it. First, to make sure that it produces a similar effect as the original direction, we use an image reconstruction loss so that the edited image produced by the new edit matches the original edit given by the user:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 339, + 186, + 545, + 200 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 339, + 186, + 545, + 200 + ], + "spans": [ + { + "bbox": [ + 339, + 186, + 545, + 200 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {i m g} = \\left\\| G \\left(w _ {0} + \\Delta_ {w}\\right) - G \\left(w _ {0} + \\Delta_ {w} ^ {*}\\right) \\right\\| _ {2} \\tag {3}", + "image_path": "26003e743709bb5cf3b1274d048c63b6c9f824cc221a739f25afafd439afb9ab.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 207, + 545, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 207, + 545, + 327 + ], + "spans": [ + { + "bbox": [ + 304, + 207, + 545, + 327 + ], + "type": "text", + "content": "Additionally, user can provide the real value of distance " + }, + { + "bbox": [ + 304, + 207, + 545, + 327 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 304, + 207, + 545, + 327 + ], + "type": "text", + "content": ", otherwise we set the distance of the edited image from the hypothetical hyperplane to be 0 (similar to [29]). We constrain the learned " + }, + { + "bbox": [ + 304, + 207, + 545, + 327 + ], + "type": "inline_equation", + "content": "\\Delta_w^*" + }, + { + "bbox": [ + 304, + 207, + 545, + 327 + ], + "type": "text", + "content": " to follow this property for better interpretability. That is, the new direction should be such that when we traverse the original latent code in that direction, " + }, + { + "bbox": [ + 304, + 207, + 545, + 327 + ], + "type": "inline_equation", + "content": "w_0 \\rightarrow w_0 + \\Delta_w^*" + }, + { + "bbox": [ + 304, + 207, + 545, + 327 + ], + "type": "text", + "content": ", the resulting image should lie at the hyperplane defined by that direction. Setting " + }, + { + "bbox": [ + 304, + 207, + 545, + 327 + ], + "type": "inline_equation", + "content": "b = 0" + }, + { + "bbox": [ + 304, + 207, + 545, + 327 + ], + "type": "text", + "content": " in Eq. 1, we minimize the edited point's distance from the hyperplane:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 371, + 331, + 545, + 343 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 371, + 331, + 545, + 343 + ], + "spans": [ + { + "bbox": [ + 371, + 331, + 545, + 343 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {a t t} = \\left| \\left(w _ {0} + \\Delta_ {w} ^ {*}\\right) \\cdot \\Delta_ {w} ^ {*} \\right| \\tag {4}", + "image_path": "68e2b2d597bce0c1a45f96da7bfdf429f00537af232bfbf537773fb192dd63d2.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 352, + 545, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 352, + 545, + 399 + ], + "spans": [ + { + "bbox": [ + 304, + 352, + 545, + 399 + ], + "type": "text", + "content": "The overall loss function for optimizing the new direction " + }, + { + "bbox": [ + 304, + 352, + 545, + 399 + ], + "type": "inline_equation", + "content": "\\Delta_w^*" + }, + { + "bbox": [ + 304, + 352, + 545, + 399 + ], + "type": "text", + "content": " is a weighted sum of the two objectives: " + }, + { + "bbox": [ + 304, + 352, + 545, + 399 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\Delta} = \\mathcal{L}_{img} + \\lambda \\mathcal{L}_{att}" + }, + { + "bbox": [ + 304, + 352, + 545, + 399 + ], + "type": "text", + "content": ". We set " + }, + { + "bbox": [ + 304, + 352, + 545, + 399 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 304, + 352, + 545, + 399 + ], + "type": "text", + "content": " to 0.02 in all experiments to balance the magnitude of " + }, + { + "bbox": [ + 304, + 352, + 545, + 399 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{img}" + }, + { + "bbox": [ + 304, + 352, + 545, + 399 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 352, + 545, + 399 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{att}" + }, + { + "bbox": [ + 304, + 352, + 545, + 399 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 400, + 545, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 400, + 545, + 568 + ], + "spans": [ + { + "bbox": [ + 304, + 400, + 545, + 568 + ], + "type": "text", + "content": "Now, in Fig. 3 (b) and (c), we compare the difference in image editing results using two different directions when applied to the same new image " + }, + { + "bbox": [ + 304, + 400, + 545, + 568 + ], + "type": "inline_equation", + "content": "(I_1)" + }, + { + "bbox": [ + 304, + 400, + 545, + 568 + ], + "type": "text", + "content": ": " + }, + { + "bbox": [ + 304, + 400, + 545, + 568 + ], + "type": "inline_equation", + "content": "w_1 \\to w_1 + \\Delta_w" + }, + { + "bbox": [ + 304, + 400, + 545, + 568 + ], + "type": "text", + "content": " vs. " + }, + { + "bbox": [ + 304, + 400, + 545, + 568 + ], + "type": "inline_equation", + "content": "w_1 \\to w_1 + \\Delta_w^*" + }, + { + "bbox": [ + 304, + 400, + 545, + 568 + ], + "type": "text", + "content": ". As discussed before, in this case, the editing using the naively computed direction " + }, + { + "bbox": [ + 304, + 400, + 545, + 568 + ], + "type": "inline_equation", + "content": "(\\Delta_w)" + }, + { + "bbox": [ + 304, + 400, + 545, + 568 + ], + "type": "text", + "content": " is not able to accurately capture the pose rotation the way it did for the user edited example " + }, + { + "bbox": [ + 304, + 400, + 545, + 568 + ], + "type": "inline_equation", + "content": "(I_0 \\to I_0')" + }, + { + "bbox": [ + 304, + 400, + 545, + 568 + ], + "type": "text", + "content": ". On the other hand, with " + }, + { + "bbox": [ + 304, + 400, + 545, + 568 + ], + "type": "inline_equation", + "content": "\\Delta_w^*" + }, + { + "bbox": [ + 304, + 400, + 545, + 568 + ], + "type": "text", + "content": ", we see that the pose of the woman changes by a similar amount as the edited example. However, since the original pose of the woman " + }, + { + "bbox": [ + 304, + 400, + 545, + 568 + ], + "type": "inline_equation", + "content": "(I_1)" + }, + { + "bbox": [ + 304, + 400, + 545, + 568 + ], + "type": "text", + "content": " was not the same as the pose of the man " + }, + { + "bbox": [ + 304, + 400, + 545, + 568 + ], + "type": "inline_equation", + "content": "(I_0)" + }, + { + "bbox": [ + 304, + 400, + 545, + 568 + ], + "type": "text", + "content": ", the new edited image " + }, + { + "bbox": [ + 304, + 400, + 545, + 568 + ], + "type": "inline_equation", + "content": "(I_1')" + }, + { + "bbox": [ + 304, + 400, + 545, + 568 + ], + "type": "text", + "content": " still does not arrive at the same final state as " + }, + { + "bbox": [ + 304, + 400, + 545, + 568 + ], + "type": "inline_equation", + "content": "I_0'" + }, + { + "bbox": [ + 304, + 400, + 545, + 568 + ], + "type": "text", + "content": ". Therefore, our next goal is to figure out how to scale the learned direction so that " + }, + { + "bbox": [ + 304, + 400, + 545, + 568 + ], + "type": "inline_equation", + "content": "I_1'" + }, + { + "bbox": [ + 304, + 400, + 545, + 568 + ], + "type": "text", + "content": " does arrive at the same final state as " + }, + { + "bbox": [ + 304, + 400, + 545, + 568 + ], + "type": "inline_equation", + "content": "I_0'" + }, + { + "bbox": [ + 304, + 400, + 545, + 568 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 574, + 531, + 587 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 574, + 531, + 587 + ], + "spans": [ + { + "bbox": [ + 305, + 574, + 531, + 587 + ], + "type": "text", + "content": "3.2. Adjusting Editing Strength for New Images" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 593, + 544, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 593, + 544, + 616 + ], + "spans": [ + { + "bbox": [ + 304, + 593, + 544, + 616 + ], + "type": "text", + "content": "Given the optimized " + }, + { + "bbox": [ + 304, + 593, + 544, + 616 + ], + "type": "inline_equation", + "content": "\\Delta_w^*" + }, + { + "bbox": [ + 304, + 593, + 544, + 616 + ], + "type": "text", + "content": " and some new image " + }, + { + "bbox": [ + 304, + 593, + 544, + 616 + ], + "type": "inline_equation", + "content": "I_{i} = G(w_{i})" + }, + { + "bbox": [ + 304, + 593, + 544, + 616 + ], + "type": "text", + "content": " we wish to edit it in the following way:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 393, + 620, + 545, + 632 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 393, + 620, + 545, + 632 + ], + "spans": [ + { + "bbox": [ + 393, + 620, + 545, + 632 + ], + "type": "interline_equation", + "content": "w _ {i} ^ {\\prime} = w _ {i} + \\alpha_ {i} n \\tag {5}", + "image_path": "a28a27fc69d23f1be5c552c0a5fe366f52843dfec31c991725f4d94b3b447b8c.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\alpha_{i}" + }, + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "type": "text", + "content": " is the editing strength computed separately for each image and " + }, + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "type": "inline_equation", + "content": "n = \\Delta_w^* /||\\Delta_w^* ||" + }, + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "type": "text", + "content": " is a unit vector in the direction of " + }, + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\Delta_w^*" + }, + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "type": "text", + "content": ". Fig. 3 (d) illustrates a geometric perspective that we will use to compute " + }, + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\alpha_{i}" + }, + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "type": "text", + "content": ". First, we see " + }, + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\Delta_w^*" + }, + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "type": "text", + "content": " represented as a vector and its corresponding hyperplane that is normal to it. Next, we depict the new latent point " + }, + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "type": "inline_equation", + "content": "w_{i}" + }, + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "type": "text", + "content": ". The" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "8274" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 70, + 293, + 266 + ], + "blocks": [ + { + "bbox": [ + 50, + 70, + 293, + 266 + ], + "lines": [ + { + "bbox": [ + 50, + 70, + 293, + 266 + ], + "spans": [ + { + "bbox": [ + 50, + 70, + 293, + 266 + ], + "type": "image", + "image_path": "c8a7088e79d3a3ec3755d382e2c6186bf27b6e736d51de3ff8d7bbc1f2b622ea.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 58, + 266, + 89, + 276 + ], + "lines": [ + { + "bbox": [ + 58, + 266, + 89, + 276 + ], + "spans": [ + { + "bbox": [ + 58, + 266, + 89, + 276 + ], + "type": "text", + "content": "Original" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 100, + 266, + 143, + 276 + ], + "lines": [ + { + "bbox": [ + 100, + 266, + 143, + 276 + ], + "spans": [ + { + "bbox": [ + 100, + 266, + 143, + 276 + ], + "type": "text", + "content": "UserID" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 148, + 266, + 190, + 276 + ], + "lines": [ + { + "bbox": [ + 148, + 266, + 190, + 276 + ], + "spans": [ + { + "bbox": [ + 148, + 266, + 190, + 276 + ], + "type": "text", + "content": "DragGAN" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 194, + 266, + 236, + 276 + ], + "lines": [ + { + "bbox": [ + 194, + 266, + 236, + 276 + ], + "spans": [ + { + "bbox": [ + 194, + 266, + 236, + 276 + ], + "type": "text", + "content": "DragDiff." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 255, + 267, + 276, + 275 + ], + "lines": [ + { + "bbox": [ + 255, + 267, + 276, + 275 + ], + "spans": [ + { + "bbox": [ + 255, + 267, + 276, + 275 + ], + "type": "text", + "content": "Ours" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 279, + 545, + 301 + ], + "lines": [ + { + "bbox": [ + 46, + 279, + 545, + 301 + ], + "spans": [ + { + "bbox": [ + 46, + 279, + 545, + 301 + ], + "type": "text", + "content": "Figure 4. Qualitative comparisons between dragging baselines. For ours, green bounding box indicates automatic transfer from the red bounding box example in the first row (i.e., no point annotation needed!)." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 242, + 70, + 291, + 266 + ], + "angle": 0, + "lines": [ + { + "bbox": [ + 242, + 70, + 291, + 266 + ], + "spans": [ + { + "bbox": [ + 242, + 70, + 291, + 266 + ], + "type": "image", + "image_path": "070513cba64059f1ea2862a73c5e444a02bdb4a440a1066d1cebbc9ee6194503.jpg" + } + ] + } + ], + "index": 5, + "type": "text" + }, + { + "bbox": [ + 46, + 314, + 287, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 314, + 287, + 373 + ], + "spans": [ + { + "bbox": [ + 46, + 314, + 287, + 373 + ], + "type": "text", + "content": "goal is to move it along the " + }, + { + "bbox": [ + 46, + 314, + 287, + 373 + ], + "type": "inline_equation", + "content": "\\Delta_w^*" + }, + { + "bbox": [ + 46, + 314, + 287, + 373 + ], + "type": "text", + "content": " direction so that it arrives onto the hyperplane. Through this depiction, " + }, + { + "bbox": [ + 46, + 314, + 287, + 373 + ], + "type": "inline_equation", + "content": "\\alpha_{i}" + }, + { + "bbox": [ + 46, + 314, + 287, + 373 + ], + "type": "text", + "content": " can be understood as the distance of " + }, + { + "bbox": [ + 46, + 314, + 287, + 373 + ], + "type": "inline_equation", + "content": "w_{1}" + }, + { + "bbox": [ + 46, + 314, + 287, + 373 + ], + "type": "text", + "content": " from the hyperplane. We can get the distance by projecting " + }, + { + "bbox": [ + 46, + 314, + 287, + 373 + ], + "type": "inline_equation", + "content": "(w_0' - w_i)" + }, + { + "bbox": [ + 46, + 314, + 287, + 373 + ], + "type": "text", + "content": " in the direction of " + }, + { + "bbox": [ + 46, + 314, + 287, + 373 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 46, + 314, + 287, + 373 + ], + "type": "text", + "content": ". Therefore, " + }, + { + "bbox": [ + 46, + 314, + 287, + 373 + ], + "type": "inline_equation", + "content": "\\alpha_{i} = (w_{0}^{\\prime} - w_{i})\\cdot n" + }, + { + "bbox": [ + 46, + 314, + 287, + 373 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 375, + 287, + 457 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 375, + 287, + 457 + ], + "spans": [ + { + "bbox": [ + 46, + 375, + 287, + 457 + ], + "type": "text", + "content": "Since for each image, the only unique computation that needs to be done is the calculation of " + }, + { + "bbox": [ + 46, + 375, + 287, + 457 + ], + "type": "inline_equation", + "content": "\\alpha_{i}" + }, + { + "bbox": [ + 46, + 375, + 287, + 457 + ], + "type": "text", + "content": ", we can see why our method will be much faster than, for example, annotating every image and running the DragGAN [17] optimization each of those times. We will show in our experiments the significant difference in time taken by our method compared to single image editing baselines." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 459, + 286, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 459, + 286, + 590 + ], + "spans": [ + { + "bbox": [ + 46, + 459, + 286, + 590 + ], + "type": "text", + "content": "Importantly, there is an added benefit to computing " + }, + { + "bbox": [ + 46, + 459, + 286, + 590 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 46, + 459, + 286, + 590 + ], + "type": "text", + "content": "'s in this way. Let's say the goal of the editing process is to rotate " + }, + { + "bbox": [ + 46, + 459, + 286, + 590 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 46, + 459, + 286, + 590 + ], + "type": "text", + "content": " faces and make them frontal. After completion, all of them do become frontal, each with their own editing strengths " + }, + { + "bbox": [ + 46, + 459, + 286, + 590 + ], + "type": "inline_equation", + "content": "\\{\\alpha_{1},\\alpha_{2},\\dots,\\alpha_{n}\\}" + }, + { + "bbox": [ + 46, + 459, + 286, + 590 + ], + "type": "text", + "content": " computed using the above formula. Suppose the user now wants the same faces facing a bit left instead. To do this, the user does not need to re-associate the original training example and run the optimization one more time. In an interactive manner, they can simply scale the " + }, + { + "bbox": [ + 46, + 459, + 286, + 590 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 46, + 459, + 286, + 590 + ], + "type": "text", + "content": " for the training example to match the desired edit, and all other " + }, + { + "bbox": [ + 46, + 459, + 286, + 590 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 46, + 459, + 286, + 590 + ], + "type": "text", + "content": "'s can be automatically recomputed." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 602, + 127, + 616 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 602, + 127, + 616 + ], + "spans": [ + { + "bbox": [ + 47, + 602, + 127, + 616 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 622, + 287, + 658 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 622, + 287, + 658 + ], + "spans": [ + { + "bbox": [ + 46, + 622, + 287, + 658 + ], + "type": "text", + "content": "We study how well our method models and transfers the edits from an example, and how efficient it is compared to single image editing baselines using their official code." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "text", + "content": "Categories. We evaluate on a variety of domains: Human faces (FFHQ) [30], AFHQ Cats, Dogs [57], MetFaces [58], Human bodies [59]. For each domain, we use the corresponding pretrained StyleGAN model to perform editing." + } + ] + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 301, + 70, + 543, + 266 + ], + "blocks": [ + { + "bbox": [ + 301, + 70, + 543, + 266 + ], + "lines": [ + { + "bbox": [ + 301, + 70, + 543, + 266 + ], + "spans": [ + { + "bbox": [ + 301, + 70, + 543, + 266 + ], + "type": "image", + "image_path": "ea1041f98c84f017aca711fc32ed3636d05645398c67c2f22e408c5baca9685c.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 310, + 266, + 341, + 276 + ], + "lines": [ + { + "bbox": [ + 310, + 266, + 341, + 276 + ], + "spans": [ + { + "bbox": [ + 310, + 266, + 341, + 276 + ], + "type": "text", + "content": "Original" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 352, + 266, + 395, + 276 + ], + "lines": [ + { + "bbox": [ + 352, + 266, + 395, + 276 + ], + "spans": [ + { + "bbox": [ + 352, + 266, + 395, + 276 + ], + "type": "text", + "content": "UserID" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 400, + 266, + 442, + 276 + ], + "lines": [ + { + "bbox": [ + 400, + 266, + 442, + 276 + ], + "spans": [ + { + "bbox": [ + 400, + 266, + 442, + 276 + ], + "type": "text", + "content": "DragGAN" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 451, + 266, + 487, + 276 + ], + "lines": [ + { + "bbox": [ + 451, + 266, + 487, + 276 + ], + "spans": [ + { + "bbox": [ + 451, + 266, + 487, + 276 + ], + "type": "text", + "content": "DragDiff." + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 492, + 266, + 528, + 275 + ], + "lines": [ + { + "bbox": [ + 492, + 266, + 528, + 275 + ], + "spans": [ + { + "bbox": [ + 492, + 266, + 528, + 275 + ], + "type": "text", + "content": "Ours" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 314, + 414, + 326 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 314, + 414, + 326 + ], + "spans": [ + { + "bbox": [ + 306, + 314, + 414, + 326 + ], + "type": "text", + "content": "4.1. Qualitative results" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 333, + 545, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 333, + 545, + 357 + ], + "spans": [ + { + "bbox": [ + 304, + 333, + 545, + 357 + ], + "type": "text", + "content": "We start with qualitative comparisons with (i) interactive point-dragging, and (ii) text-based image editing baselines." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 304, + 365, + 545, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 365, + 545, + 521 + ], + "spans": [ + { + "bbox": [ + 304, + 365, + 545, + 521 + ], + "type": "text", + "content": "Interactive point based editing. We compare against DragGAN [17], UserControllableLT [25], and DragDiffusion [16]. For each baseline, we, as a user, desire to edit a bunch of images to have the same final state for an attribute. Fig. 4 shows the results of editing two kinds of edits to four images (leftmost column). In the left case, the goal is to make everyone smile by the same degree. In the right, it is to vertically compress everyone's face by the same degree. For all three baselines, we manually annotate points for each test image. For our method, we only annotate points and perform DragGAN-based edit on one image (top row), then automatically transfer the edit to the three other test images." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 304, + 522, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 522, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 522, + 546, + 713 + ], + "type": "text", + "content": "For the smiling case (left), we notice that when DragGAN drags the upper/lower lips up/down respectively, it can either make people smile (1st, 2nd and 4th rows), or it can make people look shocked (blonde girl; 3rd row). Both edits are technically correct based on lips movement, but it won't match the user expectation of making everyone smile by the same degree. Since all four images are edited independently, this is not unexpected. The other two baselines sometimes have issues introducing the same edit across different images; e.g., UserControllableLT can make the blonde girl smile a little bit, but not the other images. In general, we notice in our experiments that to introduce the same exact edit that we want, we often need to play around with some hyperparameters (e.g., # of iterations). In comparison, our method produces edited images that are smiling more equally (left) and have been compressed by a similar" + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8275" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 49, + 71, + 294, + 284 + ], + "blocks": [ + { + "bbox": [ + 49, + 71, + 294, + 284 + ], + "lines": [ + { + "bbox": [ + 49, + 71, + 294, + 284 + ], + "spans": [ + { + "bbox": [ + 49, + 71, + 294, + 284 + ], + "type": "image", + "image_path": "df55c72c3c40102246cc15bc76fd8a26172d87785963f53ca1907f922ed2d67f.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 58, + 285, + 533, + 297 + ], + "lines": [ + { + "bbox": [ + 58, + 285, + 533, + 297 + ], + "spans": [ + { + "bbox": [ + 58, + 285, + 533, + 297 + ], + "type": "text", + "content": "Figure 5. Qualitative comparisons to text-guided baselines. Ours transfers the edit from example (red), to other test images (green)." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 301, + 72, + 545, + 284 + ], + "blocks": [ + { + "bbox": [ + 301, + 72, + 545, + 284 + ], + "lines": [ + { + "bbox": [ + 301, + 72, + 545, + 284 + ], + "spans": [ + { + "bbox": [ + 301, + 72, + 545, + 284 + ], + "type": "image", + "image_path": "eae8fe8079eafdcd4acca4e58f1046c8620c9b2f042b98a7a7d7afeb5da25fd6.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 310, + 247, + 321 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 310, + 247, + 321 + ], + "spans": [ + { + "bbox": [ + 47, + 310, + 247, + 321 + ], + "type": "text", + "content": "amount (right), and with much less human effort." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 329, + 287, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 329, + 287, + 472 + ], + "spans": [ + { + "bbox": [ + 46, + 329, + 287, + 472 + ], + "type": "text", + "content": "Text-driven image editing. We consider two baselines: StyleCLIP [23] and InstructPix2Pix [14]. Each takes an image and text as input to produce an edited image: " + }, + { + "bbox": [ + 46, + 329, + 287, + 472 + ], + "type": "inline_equation", + "content": "I \\rightarrow I'" + }, + { + "bbox": [ + 46, + 329, + 287, + 472 + ], + "type": "text", + "content": ". Fig. 5 shows the results of Ours vs. StyleCLIP (left) and Ours vs. InstructPix2Pix (right). (For the latter, we invert the output of InstructPix2Pix into StyleGAN's latent space using [42] for our method.) In each case, for the baselines, we use the same text prompt, e.g., \"A photo of a person with pale skin\" for each edited image (in four rows). For ours, we capture the edit from the first editing result of a baseline (first row), and then automatically transfer the edit to the remaining three images." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 473, + 287, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 473, + 287, + 676 + ], + "spans": [ + { + "bbox": [ + 46, + 473, + 287, + 676 + ], + "type": "text", + "content": "Our goal with this particular setup is to test (i) how consistently the baselines introduce the edit denoted by text to different images (e.g., do all people become equally pale?), and (ii) how consistently our method captures and transfers the particular edit of the first example to the rest of the images, irrespective of how good that first edit was. For (i), we find that StyleCLIP sometimes has issues; e.g., the paleness of the edited face in first example is different from the second (first vs second rows). For InstructPix2Pix, sometimes it can give consistent results; the childness of different faces seems similar. But, it can have consistency issues in other examples; e.g., the type/strength of maleness introduced is different in each image. And this is where, we believe, the utility of our method lies: if one desires to edit every face to be male in a particular way (i.e., to the same degree) as the first example, e.g., thick eyebrows & light beard, our method has an advantage." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 677, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 677, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 677, + 287, + 714 + ], + "type": "text", + "content": "Finally, we show our method's results for non-facial domains in Fig. 6. For the lion examples (top left), we see that it can preserve not just the type of edit, e.g., dragging" + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 306, + 307, + 544, + 378 + ], + "blocks": [ + { + "bbox": [ + 306, + 307, + 544, + 378 + ], + "lines": [ + { + "bbox": [ + 306, + 307, + 544, + 378 + ], + "spans": [ + { + "bbox": [ + 306, + 307, + 544, + 378 + ], + "type": "table", + "html": "
MethodPoint DraggingTimeAnno. #imgs
Dist.FID1 img1k imgs
UserControl. [25]13.6425.320.03s30s1000
DragGAN [17]4.1659.282s33.33m1000
DragDiffusion [16]26.5636.5560s16.67h1000
Ours9.4679.352s82s1
", + "image_path": "58f1e5f00f5735cf88163ff1e50a0075a74642cb43b36a3db508b2c3307ec400.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 386, + 545, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 386, + 545, + 441 + ], + "spans": [ + { + "bbox": [ + 304, + 386, + 545, + 441 + ], + "type": "text", + "content": "Table 1. Time is estimated for 1 point drag, without human annotation time. (82s includes 2s to perform edit on the example, 30s to optimize the editing direction, and 50s to transfer the edit to 1000 test images.) Our method requires only one image annotation in total, while the baselines need one annotation per test image." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 445, + 546, + 554 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 445, + 546, + 554 + ], + "spans": [ + { + "bbox": [ + 304, + 445, + 546, + 554 + ], + "type": "text", + "content": "to make the lion roar, but also the strength of the roar. The strength being preserved can be observed more easily for the human body poses (bottom left), where we see that the extent of legs split, hand movements, is consistent enough to almost align the edited test images with the user edited one. Overall, results on these diverse set of domains highlight an observation that many kinds of edits can be thought of as a combination of {type, strength}, both of which can indeed be captured and transferred according to our needs." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 563, + 421, + 574 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 563, + 421, + 574 + ], + "spans": [ + { + "bbox": [ + 306, + 563, + 421, + 574 + ], + "type": "text", + "content": "4.2. Quantitative results" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "content": "Next, perform quantitative experiments following the setup proposed in [16, 17, 19]. We first randomly sample 10 facial test images and pair them with a randomly sampled target face image. We use dlib-ml [60] to detect keypoints in the test and target images. Our goal is to see if the keypoints in the test images can be moved to the target locations specified by keypoints in the target image. For the baselines [16, 17, 25], we perform dragging for each test image. For our method, we perform dragging on one test image, and then transfer the edit to the remaining nine images. This is repeated 100 times; i.e., each time a random target image is" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8276" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 86, + 68, + 512, + 476 + ], + "blocks": [ + { + "bbox": [ + 86, + 68, + 512, + 476 + ], + "lines": [ + { + "bbox": [ + 86, + 68, + 512, + 476 + ], + "spans": [ + { + "bbox": [ + 86, + 68, + 512, + 476 + ], + "type": "image", + "image_path": "44f2934bfb3f2ad87d0e208d632ee7cc74a7d5cce1f23c1c0b29ef27845b63e6.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 187, + 478, + 405, + 490 + ], + "lines": [ + { + "bbox": [ + 187, + 478, + 405, + 490 + ], + "spans": [ + { + "bbox": [ + 187, + 478, + 405, + 490 + ], + "type": "text", + "content": "Figure 6. Additional qualitative results on various domains." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 496, + 192, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 496, + 192, + 509 + ], + "spans": [ + { + "bbox": [ + 47, + 496, + 192, + 509 + ], + "type": "text", + "content": "paired with 10 random test images." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 510, + 287, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 510, + 287, + 605 + ], + "spans": [ + { + "bbox": [ + 46, + 510, + 287, + 605 + ], + "type": "text", + "content": "We report the Euclidean distance between the keypoints of edited and target images in Table 1. Even without requiring annotations for every test image, our method can move the points very close to the target; almost as close as DragGAN, which requires annotation each time. We also compute FID [61] between the original test images and their edited versions to ensure that our method does not distort the image; and our quality is comparable to DragGAN." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 619, + 145, + 633 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 619, + 145, + 633 + ], + "spans": [ + { + "bbox": [ + 47, + 619, + 145, + 633 + ], + "type": "text", + "content": "5. Deeper Analysis" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 640, + 287, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 640, + 287, + 700 + ], + "spans": [ + { + "bbox": [ + 47, + 640, + 287, + 700 + ], + "type": "text", + "content": "We perform deeper analysis on the usefulness of our batch image editing framework by focusing on a specific edit: face pose rotation. In particular, if we wish to rotate many faces to front, how do we visualize the improvements brought by different components of our method?" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 58, + 701, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 701, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 58, + 701, + 287, + 714 + ], + "type": "text", + "content": "Fig. 7 depicts the setup. We use DragGAN to perform" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 496, + 547, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 496, + 547, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 496, + 547, + 712 + ], + "type": "text", + "content": "dragging on one image (top; " + }, + { + "bbox": [ + 304, + 496, + 547, + 712 + ], + "type": "inline_equation", + "content": "w_0 \\rightarrow w_0'" + }, + { + "bbox": [ + 304, + 496, + 547, + 712 + ], + "type": "text", + "content": "). We then transfer the edit to 1000 other test images (two shown in rows 2-3). We see that naively applying " + }, + { + "bbox": [ + 304, + 496, + 547, + 712 + ], + "type": "inline_equation", + "content": "\\Delta_w = w_0' - w_0" + }, + { + "bbox": [ + 304, + 496, + 547, + 712 + ], + "type": "text", + "content": " from DragGAN to test images cannot make them frontal (" + }, + { + "bbox": [ + 304, + 496, + 547, + 712 + ], + "type": "inline_equation", + "content": "0^\\circ" + }, + { + "bbox": [ + 304, + 496, + 547, + 712 + ], + "type": "text", + "content": "). Using our optimized " + }, + { + "bbox": [ + 304, + 496, + 547, + 712 + ], + "type": "inline_equation", + "content": "\\Delta_w^*" + }, + { + "bbox": [ + 304, + 496, + 547, + 712 + ], + "type": "text", + "content": " (without dynamic scaling) does help it bring closer to facing front. Does this mean that " + }, + { + "bbox": [ + 304, + 496, + 547, + 712 + ], + "type": "inline_equation", + "content": "\\Delta_w^*" + }, + { + "bbox": [ + 304, + 496, + 547, + 712 + ], + "type": "text", + "content": " might be a more global direction? We study this in Fig. 8 (a), where we visualize the degree to which the correlation property of global directions (Eq. 2) holds for " + }, + { + "bbox": [ + 304, + 496, + 547, + 712 + ], + "type": "inline_equation", + "content": "\\Delta_w" + }, + { + "bbox": [ + 304, + 496, + 547, + 712 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 496, + 547, + 712 + ], + "type": "inline_equation", + "content": "\\Delta_w^*" + }, + { + "bbox": [ + 304, + 496, + 547, + 712 + ], + "type": "text", + "content": " individually. We see that there is indeed a better correlation (" + }, + { + "bbox": [ + 304, + 496, + 547, + 712 + ], + "type": "inline_equation", + "content": "R^2" + }, + { + "bbox": [ + 304, + 496, + 547, + 712 + ], + "type": "text", + "content": ") between the distances of latents from hyperplanes and the Yaw degree of resulting images. (We predict Yaw using 6DRepNet [62].) However, as we discussed before, " + }, + { + "bbox": [ + 304, + 496, + 547, + 712 + ], + "type": "inline_equation", + "content": "\\Delta_w^*" + }, + { + "bbox": [ + 304, + 496, + 547, + 712 + ], + "type": "text", + "content": " in itself is insufficient: to completely bring the facial pose to front, we need to scale it with corresponding " + }, + { + "bbox": [ + 304, + 496, + 547, + 712 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 304, + 496, + 547, + 712 + ], + "type": "text", + "content": "'s to bring them much closer to front (Fig. 7, 4th column). To study this in a more systematic way, we visualize the effect on all 1000 test images. Similar to Fig. 8 (a), we visualize" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8277" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 51, + 70, + 283, + 265 + ], + "blocks": [ + { + "bbox": [ + 51, + 70, + 283, + 265 + ], + "lines": [ + { + "bbox": [ + 51, + 70, + 283, + 265 + ], + "spans": [ + { + "bbox": [ + 51, + 70, + 283, + 265 + ], + "type": "image", + "image_path": "f60e5e943108b02f35a126b09dc2235f141c9eec4980e764b72229e506fa1bd1.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 47, + 266, + 286, + 289 + ], + "lines": [ + { + "bbox": [ + 47, + 266, + 286, + 289 + ], + "spans": [ + { + "bbox": [ + 47, + 266, + 286, + 289 + ], + "type": "text", + "content": "Figure 7. Effect of adjusted and optimized editing for test images. Yaw degree of each image is provided on top of each images." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 48, + 297, + 286, + 352 + ], + "blocks": [ + { + "bbox": [ + 48, + 297, + 286, + 352 + ], + "lines": [ + { + "bbox": [ + 48, + 297, + 286, + 352 + ], + "spans": [ + { + "bbox": [ + 48, + 297, + 286, + 352 + ], + "type": "table", + "html": "
MethodMAETime complexity
PrepareInference
Random11.295 ± 8.972--
DragGAN [17] + GANgealing [63]8.141 ± 7.22130s2000s
Ours2.120 ± 1.81832s50s
", + "image_path": "ae8ac06b7f4964e0935dddaf672482e9c56b3867bdbc39f61e567831c351ae83.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 72, + 353, + 261, + 363 + ], + "lines": [ + { + "bbox": [ + 72, + 353, + 261, + 363 + ], + "spans": [ + { + "bbox": [ + 72, + 353, + 261, + 363 + ], + "type": "text", + "content": "Table 2. Ablation Study: DragGAN + GANgealing." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 371, + 287, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 371, + 287, + 418 + ], + "spans": [ + { + "bbox": [ + 46, + 371, + 287, + 418 + ], + "type": "text", + "content": "distance-to-hyperplane vs. Yaw degree for the original test images (blue) and edited images (red) in Fig. 8 (b). We can see that the variation in Yaw for the edited images collapses at around " + }, + { + "bbox": [ + 46, + 371, + 287, + 418 + ], + "type": "inline_equation", + "content": "0^{\\circ}" + }, + { + "bbox": [ + 46, + 371, + 287, + 418 + ], + "type": "text", + "content": "; i.e., they mostly face front as we would like." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 419, + 287, + 478 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 419, + 287, + 478 + ], + "spans": [ + { + "bbox": [ + 46, + 419, + 287, + 478 + ], + "type": "text", + "content": "We also compute the time it takes to perform editing in Table 1, last three columns. Our method only requires annotating one image, while the baselines require annotating each image (e.g., 1000 total in this case). Note the times do not include human annotation time." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 486, + 287, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 486, + 287, + 605 + ], + "spans": [ + { + "bbox": [ + 46, + 486, + 287, + 605 + ], + "type": "text", + "content": "Other ways to automate batch image editing. One baseline for batch image editing (i.e., so that all edited images achieve the same final state) is shown in Fig. 9. After the user annotates the source/target points in the example image, we use GANgealing [63] to transfer the points to corresponding locations in each new test image. The edited images are then obtained using DragGAN (DragGAN + GANgealing). We compare our method to this baseline when the user edits one face image to become front " + }, + { + "bbox": [ + 46, + 486, + 287, + 605 + ], + "type": "inline_equation", + "content": "(\\mathrm{Yaw} = 0^{\\circ})" + }, + { + "bbox": [ + 46, + 486, + 287, + 605 + ], + "type": "text", + "content": " and transferring it to 1000 other test images." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 45, + 606, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 606, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 45, + 606, + 287, + 713 + ], + "type": "text", + "content": "Results are shown in Table 2. We report mean absolute error (MAE) between the Yaw degree of the edited image and its ideal frontal image " + }, + { + "bbox": [ + 45, + 606, + 287, + 713 + ], + "type": "inline_equation", + "content": "(\\mathrm{Yaw} = 0^{\\circ})" + }, + { + "bbox": [ + 45, + 606, + 287, + 713 + ], + "type": "text", + "content": ". 'Random' shows the variations in Yaw for the original images (before editing). While 'DragGAN + GANgealing' does help in reducing the variation " + }, + { + "bbox": [ + 45, + 606, + 287, + 713 + ], + "type": "inline_equation", + "content": "(8.14 < 11.29)" + }, + { + "bbox": [ + 45, + 606, + 287, + 713 + ], + "type": "text", + "content": ", our edited images are much more frontal, with an MAE of 2.12. This is because the baseline has no way to bring all edited images to the same final state. On top of that, since our method does not rely on" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 310, + 72, + 541, + 156 + ], + "blocks": [ + { + "bbox": [ + 310, + 72, + 541, + 156 + ], + "lines": [ + { + "bbox": [ + 310, + 72, + 541, + 156 + ], + "spans": [ + { + "bbox": [ + 310, + 72, + 541, + 156 + ], + "type": "image", + "image_path": "2344e8afeb8e2a2009e2d8ace1e9bc1b02dd78946e357c60d7b839bb4c43ccb5.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 310, + 159, + 541, + 243 + ], + "blocks": [ + { + "bbox": [ + 310, + 159, + 541, + 243 + ], + "lines": [ + { + "bbox": [ + 310, + 159, + 541, + 243 + ], + "spans": [ + { + "bbox": [ + 310, + 159, + 541, + 243 + ], + "type": "image", + "image_path": "73454f2fb7eaedc5a334eca6e89fb4fe69459100f142d30f76d4839735cb0786.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 247, + 545, + 281 + ], + "lines": [ + { + "bbox": [ + 305, + 247, + 545, + 281 + ], + "spans": [ + { + "bbox": [ + 305, + 247, + 545, + 281 + ], + "type": "text", + "content": "Figure 8. (a) After optimization, the editing direction is more linearly correlated with the yaw attribute. (b) With automatically adjusted editing scale for each test image." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 309, + 285, + 544, + 361 + ], + "blocks": [ + { + "bbox": [ + 309, + 285, + 544, + 361 + ], + "lines": [ + { + "bbox": [ + 309, + 285, + 544, + 361 + ], + "spans": [ + { + "bbox": [ + 309, + 285, + 544, + 361 + ], + "type": "image", + "image_path": "870b11d8902c50cfe052a24c06cf8e5498bccb50ab164a08a9254cf03e894bd9.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 362, + 545, + 384 + ], + "lines": [ + { + "bbox": [ + 305, + 362, + 545, + 384 + ], + "spans": [ + { + "bbox": [ + 305, + 362, + 545, + 384 + ], + "type": "text", + "content": "Figure 9. Ablation study set up. We use GANgealing [63] to transfer the annotated points from example to test image." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 393, + 545, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 393, + 545, + 441 + ], + "spans": [ + { + "bbox": [ + 305, + 393, + 545, + 441 + ], + "type": "text", + "content": "additional information about keypoints (which 'DragGAN + GANgealing' does), we believe that it is better suited for batch image editing for many kinds of domains (e.g., Fig. 6), where we might not have such information." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 455, + 378, + 467 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 455, + 378, + 467 + ], + "spans": [ + { + "bbox": [ + 306, + 455, + 378, + 467 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 475, + 545, + 595 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 475, + 545, + 595 + ], + "spans": [ + { + "bbox": [ + 304, + 475, + 545, + 595 + ], + "type": "text", + "content": "We introduced the problem of interactive batch image editing. Given a user edit in an example image, our approach automatically transfers that edit to other test images, maintaining a consistent final state of the edit across images. Extensive experiments demonstrated that our method produces comparable quality to state-of-the-art single-image-editing methods while saving significant time and human effort. We are currently limited to StyleGAN models. Extending this problem and solution to diffusion-based models for more edits types would be an exciting future direction." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 609, + 403, + 622 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 609, + 403, + 622 + ], + "spans": [ + { + "bbox": [ + 306, + 609, + 403, + 622 + ], + "type": "text", + "content": "Acknowledgement" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "text", + "content": "This work was supported in part by NSF CAREER IIS2150012, Adobe Data Science award, Sony Focused Research award, and Institute of Information & communications Technology Planning & Evaluation (IITP) grant funded by the Korea government (MSIT) (No. RS-2022-00187238, Development of Large Korean Language Model Technology for Efficient Pretraining)." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8278" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 287, + 714 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 122 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 122 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 122 + ], + "type": "text", + "content": "[1] Man M. Ho and Jinjia Zhou. Deep preset: Blending and retouching photos with color style transfer. In WACV, 2021. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 124, + 287, + 156 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 124, + 287, + 156 + ], + "spans": [ + { + "bbox": [ + 53, + 124, + 287, + 156 + ], + "type": "text", + "content": "[2] Mahmoud Afifi, Konstantinos G. Derpanis, Bjorn Ommer, and Michael S. Brown. Learning multi-scale photo exposure correction. In CVPR, 2021." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 158, + 286, + 200 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 158, + 286, + 200 + ], + "spans": [ + { + "bbox": [ + 53, + 158, + 286, + 200 + ], + "type": "text", + "content": "[3] Jong-Hyeon Baek, DaeHyun Kim, Su-Min Choi, Hyo-jun Lee, Hanul Kim, and Yeong Jun Koh. Luminance-aware color transform for multiple exposure correction. In ICCV, 2023." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 202, + 286, + 235 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 202, + 286, + 235 + ], + "spans": [ + { + "bbox": [ + 53, + 202, + 286, + 235 + ], + "type": "text", + "content": "[4] Wenyan Cong, Jianfu Zhang, Li Niu, Liu Liu, Zhixin Ling, Weiyuan Li, and Liqing Zhang. Dovenet: Deep image harmonization via domain verification. In CVPR, 2020." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 236, + 286, + 267 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 236, + 286, + 267 + ], + "spans": [ + { + "bbox": [ + 53, + 236, + 286, + 267 + ], + "type": "text", + "content": "[5] Yi-Hsuan Tsai, Xiaohui Shen, Zhe Lin, Kalyan Sunkavalli, Xin Lu, and Ming-Hsuan Yang. Deep image harmonization, 2017." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 269, + 286, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 269, + 286, + 300 + ], + "spans": [ + { + "bbox": [ + 53, + 269, + 286, + 300 + ], + "type": "text", + "content": "[6] Zonghui Guo, Haiyong Zheng, Yufeng Jiang, Zhaorui Gu, and Bing Zheng. Intrinsic image harmonization. In CVPR, 2021." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 302, + 286, + 334 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 302, + 286, + 334 + ], + "spans": [ + { + "bbox": [ + 53, + 302, + 286, + 334 + ], + "type": "text", + "content": "[7] Jun Ling, Han Xue, Li Song, Rong Xie, and Xiao Gu. Region-aware adaptive instance normalization for image harmonization. In CVPR, 2021." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 335, + 286, + 379 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 335, + 286, + 379 + ], + "spans": [ + { + "bbox": [ + 53, + 335, + 286, + 379 + ], + "type": "text", + "content": "[8] Yifan Jiang, He Zhang, Jianming Zhang, Yilin Wang, Zhe Lin, Kalyan Sunkavalli, Simon Chen, Sohrab Amirghodsi, Sarah Kong, and Zhangyang Wang. Ssh: A self-supervised framework for image harmonization. In ICCV, 2021." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 380, + 286, + 412 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 380, + 286, + 412 + ], + "spans": [ + { + "bbox": [ + 53, + 380, + 286, + 412 + ], + "type": "text", + "content": "[9] Wenyan Cong, Xinhao Tao, Li Niu, Jing Liang, Xuesong Gao, Qihao Sun, and Liqing Zhang. High-resolution image harmonization via collaborative dual transformations. 2022." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 414, + 286, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 414, + 286, + 456 + ], + "spans": [ + { + "bbox": [ + 48, + 414, + 286, + 456 + ], + "type": "text", + "content": "[10] Jeya Maria Jose Valanarasu, He Zhang, Jianming Zhang, Yilin Wang, Zhe Lin, Jose Echevarria, Yinglan Ma, Zijun Wei, Kalyan Sunkavalli, and Vishal M Patel. Interactive portrait harmonization. 2022." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 458, + 286, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 458, + 286, + 491 + ], + "spans": [ + { + "bbox": [ + 48, + 458, + 286, + 491 + ], + "type": "text", + "content": "[11] Lingzhi Zhang, Tarmily Wen, and Jianbo Shi. Deep image blending. In The IEEE Winter Conference on Applications of Computer Vision, pages 231-240, 2020." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 491, + 286, + 534 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 491, + 286, + 534 + ], + "spans": [ + { + "bbox": [ + 48, + 491, + 286, + 534 + ], + "type": "text", + "content": "[12] Dave Epstein, Allan Jabri, Ben Poole, Alexei A. Efros, and Aleksander Holynski. Diffusion self-guidance for controllable image generation. arXiv preprint arXiv:2306.00986, 2023. 1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 536, + 286, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 536, + 286, + 567 + ], + "spans": [ + { + "bbox": [ + 48, + 536, + 286, + 567 + ], + "type": "text", + "content": "[13] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. 2021. 1, 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 569, + 286, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 569, + 286, + 601 + ], + "spans": [ + { + "bbox": [ + 48, + 569, + 286, + 601 + ], + "type": "text", + "content": "[14] Tim Brooks, Aleksander Holynski, and Alexei A Efros. Instructpix2pix: Learning to follow image editing instructions. 2022. 1, 2, 3, 6" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 602, + 286, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 602, + 286, + 645 + ], + "spans": [ + { + "bbox": [ + 48, + 602, + 286, + 645 + ], + "type": "text", + "content": "[15] Chenlin Meng, Yutong He, Yang Song, Jiaming Song, Jia-jun Wu, Jun-Yan Zhu, and Stefano Ermon. SDEdit: Guided image synthesis and editing with stochastic differential equations. In ICLR, 2022. 1, 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 48, + 647, + 286, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 647, + 286, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 647, + 286, + 689 + ], + "type": "text", + "content": "[16] Yujun Shi, Chuhui Xue, Jiachun Pan, Wenqing Zhang, Vincent YF Tan, and Song Bai. Dragdiffusion: Harnessing diffusion models for interactive point-based image editing. 2023. 1, 3, 5, 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 48, + 691, + 286, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 286, + 714 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 286, + 714 + ], + "type": "text", + "content": "[17] Xingang Pan, Ayush Tewari, Thomas Leimkuhler, Lingjie Liu, Abhimitra Meka, and Christian Theobalt. Drag your" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 39, + "blocks": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "type": "text", + "content": "gan: Interactive point-based manipulation on the generative image manifold. In SIGGRAPH, 2023. 1, 2, 3, 5, 6, 8" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 95, + 545, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 95, + 545, + 128 + ], + "spans": [ + { + "bbox": [ + 308, + 95, + 545, + 128 + ], + "type": "text", + "content": "[18] Guillaume Couairon, Jakob Verbeek, Holger Schwenk, and Matthieu Cord. Diffedit: Diffusion-based semantic image editing with mask guidance, 2022." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 129, + 545, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 129, + 545, + 160 + ], + "spans": [ + { + "bbox": [ + 308, + 129, + 545, + 160 + ], + "type": "text", + "content": "[19] Chong Mou, Xintao Wang, Jiechong Song, Ying Shan, and Jian Zhang. Dragondiffusion: Enabling drag-style manipulation on diffusion models. 2023. 1, 2, 3, 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 162, + 545, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 162, + 545, + 205 + ], + "spans": [ + { + "bbox": [ + 308, + 162, + 545, + 205 + ], + "type": "text", + "content": "[20] Binxin Yang, Shuyang Gu, Bo Zhang, Ting Zhang, Xuejin Chen, Xiaoyan Sun, Dong Chen, and Fang Wen. Paint by example: Exemplar-based image editing with diffusion models. arXiv preprint arXiv:2211.13227, 2022." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 206, + 545, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 206, + 545, + 226 + ], + "spans": [ + { + "bbox": [ + 308, + 206, + 545, + 226 + ], + "type": "text", + "content": "[21] Adding Conditional Control to Text-to-Image Diffusion Models, 2023. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 228, + 545, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 228, + 545, + 248 + ], + "spans": [ + { + "bbox": [ + 308, + 228, + 545, + 248 + ], + "type": "text", + "content": "[22] GLIGEN: Open-Set Grounded Text-to-Image Generation, 2023. arXiv:2301.07093. 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 250, + 545, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 250, + 545, + 282 + ], + "spans": [ + { + "bbox": [ + 308, + 250, + 545, + 282 + ], + "type": "text", + "content": "[23] Or Patashnik, Zongze Wu, Eli Shechtman, Daniel Cohen-Or, and Dani Lischinski. Styleclip: Text-driven manipulation of stylegan imagery. In ICCV, 2021. 1, 2, 3, 6" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 284, + 545, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 284, + 545, + 326 + ], + "spans": [ + { + "bbox": [ + 308, + 284, + 545, + 326 + ], + "type": "text", + "content": "[24] Bahjat Kawar, Shiran Zada, Oran Lang, Omer Tov, Huiwen Chang, Tali Dekel, Inbar Mosseri, and Michal Irani. Imagic: Text-based real image editing with diffusion models. In CVPR, 2023. 1, 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 327, + 545, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 327, + 545, + 358 + ], + "spans": [ + { + "bbox": [ + 308, + 327, + 545, + 358 + ], + "type": "text", + "content": "[25] Yuki Endo. User-controllable latent transformer for stylegan image layout editing. Computer Graphics Forum, 2022. 1, 2, 3, 5, 6" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 360, + 545, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 360, + 545, + 392 + ], + "spans": [ + { + "bbox": [ + 308, + 360, + 545, + 392 + ], + "type": "text", + "content": "[26] Erik Härkönen, Aaron Hertzmann, Jaakko Lehtinen, and Sylvain Paris. Ganspace: Discovering interpretable gan controls. In NeurIPS, 2020. 2, 3, 4" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 394, + 545, + 425 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 394, + 545, + 425 + ], + "spans": [ + { + "bbox": [ + 308, + 394, + 545, + 425 + ], + "type": "text", + "content": "[27] Yujun Shen, Jinjin Gu, Xiaou Tang, and Bolei Zhou. Interpreting the latent space of gans for semantic face editing. In CVPR, 2020." + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 308, + 426, + 545, + 448 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 426, + 545, + 448 + ], + "spans": [ + { + "bbox": [ + 308, + 426, + 545, + 448 + ], + "type": "text", + "content": "[28] Yujun Shen and Bolei Zhou. Closed-form factorization of latent semantics in gans. In CVPR, 2021. 3, 4" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 308, + 449, + 545, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 449, + 545, + 480 + ], + "spans": [ + { + "bbox": [ + 308, + 449, + 545, + 480 + ], + "type": "text", + "content": "[29] Yotam Nitzan, Rinon Gal, Ofir Brenner, and Daniel Cohen-Or. Large: Latent-based regression through gan semantics. In CVPR, 2022. 2, 4" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 308, + 482, + 545, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 482, + 545, + 514 + ], + "spans": [ + { + "bbox": [ + 308, + 482, + 545, + 514 + ], + "type": "text", + "content": "[30] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. 2019. 2, 4, 5" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 308, + 515, + 545, + 548 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 515, + 545, + 548 + ], + "spans": [ + { + "bbox": [ + 308, + 515, + 545, + 548 + ], + "type": "text", + "content": "[31] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of StyleGAN. In CVPR, 2020. 2, 3, 4" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 308, + 548, + 545, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 548, + 545, + 590 + ], + "spans": [ + { + "bbox": [ + 308, + 548, + 545, + 590 + ], + "type": "text", + "content": "[32] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. In NeurIPS, 2014. 2" + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 308, + 592, + 545, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 592, + 545, + 613 + ], + "spans": [ + { + "bbox": [ + 308, + 592, + 545, + 613 + ], + "type": "text", + "content": "[33] Yann LeCun. The mnist database of handwritten digits. http://yann.lecun.com/exdb/mnist/, 1998. 2" + } + ] + } + ], + "index": 35 + }, + { + "bbox": [ + 308, + 614, + 545, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 614, + 545, + 645 + ], + "spans": [ + { + "bbox": [ + 308, + 614, + 545, + 645 + ], + "type": "text", + "content": "[34] Andrew Brock, Jeff Donahue, and Karen Simonyan. Large scale gan training for high fidelity natural image synthesis. arXiv, 2018. 2" + } + ] + } + ], + "index": 36 + }, + { + "bbox": [ + 308, + 647, + 545, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 647, + 545, + 679 + ], + "spans": [ + { + "bbox": [ + 308, + 647, + 545, + 679 + ], + "type": "text", + "content": "[35] Terro Karras, Timo Aila, Samuli Laine, and Jaakko Lehtinen. Progressive growing of gans for improved quality, stability, and variation. In ICLR, 2018. 2" + } + ] + } + ], + "index": 37 + }, + { + "bbox": [ + 308, + 681, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 681, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 308, + 681, + 545, + 712 + ], + "type": "text", + "content": "[36] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR, 2009. 2" + } + ] + } + ], + "index": 38 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8279" + } + ] + } + ], + "index": 40 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 714 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "type": "text", + "content": "[37] Fisher Yu, Yinda Zhang, Shuran Song, Ari Seff, and Jianxiong Xiao. Lsun: Construction of a large-scale image dataset using deep learning with humans in the loop. 2015. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 107, + 287, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 107, + 287, + 140 + ], + "spans": [ + { + "bbox": [ + 48, + 107, + 287, + 140 + ], + "type": "text", + "content": "[38] Ali Jahanian, Lucy Chai, and Phillip Isola. On the \"steerability\" of generative adversarial networks. In ICLR, 2020. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 142, + 288, + 186 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 142, + 288, + 186 + ], + "spans": [ + { + "bbox": [ + 48, + 142, + 288, + 186 + ], + "type": "text", + "content": "[39] David Bau, Jun-Yan Zhu, Hendrik Strobelt, Agata Lapedriza, Bolei Zhou, and Antonio Torralba. Understanding the role of individual units in a deep neural network. Proceedings of the National Academy of Sciences, 2020. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 188, + 287, + 220 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 188, + 287, + 220 + ], + "spans": [ + { + "bbox": [ + 48, + 188, + 287, + 220 + ], + "type": "text", + "content": "[40] Zongze Wu, Dani Lischinski, and Eli Shechtman. Stylespace analysis: Disentangled controls for stylegan image generation. In ICLR, 2020. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 222, + 287, + 255 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 222, + 287, + 255 + ], + "spans": [ + { + "bbox": [ + 48, + 222, + 287, + 255 + ], + "type": "text", + "content": "[41] Tengfei Wang, Yong Zhang, Yanbo Fan, Jue Wang, and Qifeng Chen. High-fidelity gan inversion for image attribute editing. In CVPR, 2022. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 257, + 287, + 290 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 257, + 287, + 290 + ], + "spans": [ + { + "bbox": [ + 48, + 257, + 287, + 290 + ], + "type": "text", + "content": "[42] Omer Tov, Yuval Alaluf, Yotam Nitzan, Or Patashnik, and Daniel Cohen-Or. Designing an encoder for stylegan image manipulation. In ACM Transactions on Graphics, 2021. 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 291, + 287, + 324 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 291, + 287, + 324 + ], + "spans": [ + { + "bbox": [ + 48, + 291, + 287, + 324 + ], + "type": "text", + "content": "[43] Daniel Roich, Ron Mokady, Amit H Bermano, and Daniel Cohen-Or. Pivotal tuning for latent-based editing of real images. ACM Trans. Graph., 2021. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 326, + 287, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 326, + 287, + 381 + ], + "spans": [ + { + "bbox": [ + 48, + 326, + 287, + 381 + ], + "type": "text", + "content": "[44] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. 2021. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 383, + 287, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 383, + 287, + 415 + ], + "spans": [ + { + "bbox": [ + 48, + 383, + 287, + 415 + ], + "type": "text", + "content": "[45] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. 2021. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 417, + 287, + 472 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 417, + 287, + 472 + ], + "spans": [ + { + "bbox": [ + 48, + 417, + 287, + 472 + ], + "type": "text", + "content": "[46] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In H. Larochelle, M. Ranzato, R. Hadsell, M.F. Balcan, and H. Lin, editors, Advances in Neural Information Processing Systems. Curran Associates, Inc., 2020. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 474, + 287, + 518 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 474, + 287, + 518 + ], + "spans": [ + { + "bbox": [ + 48, + 474, + 287, + 518 + ], + "type": "text", + "content": "[47] Ron Mokady, Amir Hertz, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Null-text inversion for editing real images using guided diffusion models. arXiv preprint arXiv:2211.09794, 2022. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 519, + 287, + 563 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 519, + 287, + 563 + ], + "spans": [ + { + "bbox": [ + 48, + 519, + 287, + 563 + ], + "type": "text", + "content": "[48] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-prompt image editing with cross attention control. arXiv preprint arXiv:2208.01626, 2022. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 565, + 287, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 565, + 287, + 609 + ], + "spans": [ + { + "bbox": [ + 48, + 565, + 287, + 609 + ], + "type": "text", + "content": "[49] Huan Ling, Karsten Kreis, Daiqing Li, Seung Wook Kim, Antonio Torralba, and Sanja Fidler. Editgan: High-precision semantic image editing. In Advances in Neural Information Processing Systems (NeurIPS), 2021. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 610, + 287, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 610, + 287, + 643 + ], + "spans": [ + { + "bbox": [ + 48, + 610, + 287, + 643 + ], + "type": "text", + "content": "[50] David Bau, Steven Liu, Tongzhou Wang, Jun-Yan Zhu, and Antonio Torralba. Rewriting a deep generative model. In ECCV, 2020. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 644, + 287, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 644, + 287, + 679 + ], + "spans": [ + { + "bbox": [ + 48, + 644, + 287, + 679 + ], + "type": "text", + "content": "[51] Amir Bar, Yossi Gandelsman, Trevor Darrell, Amir Globerson, and Alexei A. Efros. Visual prompting via image inpainting. 2022. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 680, + 287, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 680, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 48, + 680, + 287, + 714 + ], + "type": "text", + "content": "[52] Thao Nguyen, Yuheng Li, Utkarsh Ojha, and Yong Jae Lee. Visual instruction inversion: Image editing via visual prompting. In NeurIPS, 2023." + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 72, + 546, + 466 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 307, + 72, + 546, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 72, + 546, + 116 + ], + "spans": [ + { + "bbox": [ + 307, + 72, + 546, + 116 + ], + "type": "text", + "content": "[53] Yasheng Sun, Yifan Yang, Houwen Peng, Yifei Shen, Yuqing Yang, Han Hu, Lili Qiu, and Hideki Koike. Imagebrush: Learning visual in-context instructions for exemplar-based image manipulation. 2023." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 118, + 546, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 546, + 150 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 546, + 150 + ], + "type": "text", + "content": "[54] Aaron Hertzmann, Charles E. Jacobs, Nuria Oliver, Brian Curless, and David H. Salesin. Image analogies. In Seminal Graphics Papers: Pushing the Boundaries, Volume 2, 2023." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 152, + 546, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 152, + 546, + 184 + ], + "spans": [ + { + "bbox": [ + 307, + 152, + 546, + 184 + ], + "type": "text", + "content": "[55] Jing Liao, Yuan Yao, Lu Yuan, Gang Hua, and Sing Bing Kang. Visual attribute transfer through deep image analogy. 2017. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 185, + 546, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 185, + 546, + 217 + ], + "spans": [ + { + "bbox": [ + 307, + 185, + 546, + 217 + ], + "type": "text", + "content": "[56] Weihao Xia, Yulun Zhang, Yujiu Yang, Jing-Hao Xue, Bolei Zhou, and Ming-Hsuan Yang. Gan inversion: A survey. TPAMI, 2022. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 219, + 546, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 219, + 546, + 251 + ], + "spans": [ + { + "bbox": [ + 307, + 219, + 546, + 251 + ], + "type": "text", + "content": "[57] Yunjey Choi, Youngjung Uh, Jaejun Yoo, and Jung-Woo Ha. Starganv2: Diverse image synthesis for multiple domains. In CVPR, 2020. 5" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 253, + 546, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 253, + 546, + 285 + ], + "spans": [ + { + "bbox": [ + 307, + 253, + 546, + 285 + ], + "type": "text", + "content": "[58] Tero Karras, Miika Aittala, Janne Hellsten, Samuli Laine, Jaakko Lehtinen, and Timo Aila. Training generative adversarial networks with limited data, 2020. 5" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 286, + 546, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 286, + 546, + 331 + ], + "spans": [ + { + "bbox": [ + 307, + 286, + 546, + 331 + ], + "type": "text", + "content": "[59] Jianglin Fu, Shikai Li, Yuming Jiang, Kwan-Yee Lin, Chen Qian, Chen-Change Loy, Wayne Wu, and Ziwei Liu. Stylegan-human: A data-centric odyssey of human generation. arXiv preprint, arXiv:2204.11823, 2022. 5" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 332, + 546, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 332, + 546, + 354 + ], + "spans": [ + { + "bbox": [ + 307, + 332, + 546, + 354 + ], + "type": "text", + "content": "[60] Davis E. King. Dlib-ml: A machine learning toolkit. J. Mach. Learn. Res., 2009. 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 355, + 546, + 399 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 355, + 546, + 399 + ], + "spans": [ + { + "bbox": [ + 307, + 355, + 546, + 399 + ], + "type": "text", + "content": "[61] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, Gunter Klambauer, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. In NeurIPS, 2017. 7" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 399, + 546, + 432 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 399, + 546, + 432 + ], + "spans": [ + { + "bbox": [ + 307, + 399, + 546, + 432 + ], + "type": "text", + "content": "[62] Thorsten Hempel, Ahmed A. Abdelrahman, and Ayoub Al-Hamadi. 6d rotation representation for unconstrained head pose estimation. In ICIP, 2022. 7" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 434, + 546, + 466 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 434, + 546, + 466 + ], + "spans": [ + { + "bbox": [ + 307, + 434, + 546, + 466 + ], + "type": "text", + "content": "[63] William Peebles, Jun-Yan Zhu, Richard Zhang, Antonio Torralba, Alexei Efros, and Eli Shechtman. Gan-supervised dense visual alignment. In CVPR, 2022. 8" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "8280" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/EditGuard_ Versatile Image Watermarking for Tamper Localization and Copyright Protection/43cb9c2b-dd1a-4670-9150-d4345483687e_content_list.json b/2024/EditGuard_ Versatile Image Watermarking for Tamper Localization and Copyright Protection/43cb9c2b-dd1a-4670-9150-d4345483687e_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..a7954247c54aacf891724f5ed5e5b63ef4db3b7b --- /dev/null +++ b/2024/EditGuard_ Versatile Image Watermarking for Tamper Localization and Copyright Protection/43cb9c2b-dd1a-4670-9150-d4345483687e_content_list.json @@ -0,0 +1,2440 @@ +[ + { + "type": "text", + "text": "EditGuard: Versatile Image Watermarking for Tamper Localization and Copyright Protection", + "text_level": 1, + "bbox": [ + 119, + 130, + 851, + 175 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Xuanyu Zhang $^{1,2}$ , Runyi Li $^{1}$ , Jiwen Yu $^{1}$ , Youmin Xu $^{1}$ , Weiqi Li $^{1}$ , Jian Zhang $^{1,2 \\boxtimes}$ $^{1}$ School of Electronic and Computer Engineering, Peking University", + "bbox": [ + 165, + 189, + 802, + 226 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ Peking University Shenzhen Graduate School-Rabbitpre AIGC Joint Research Laboratory", + "bbox": [ + 120, + 226, + 841, + 244 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/d8378e347524d202d48a803c7b961ea40fa20288455a143e19ee3a00096d0467.jpg", + "image_caption": [ + "Users" + ], + "image_footnote": [], + "bbox": [ + 86, + 257, + 130, + 292 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Please help protect my image against intentional tampering. My exclusive copyright information is 00111010110...", + "bbox": [ + 148, + 262, + 290, + 295 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/5607f94e273b084db6289341ec892ef5497d508e43c3a0c17f984efc08e968a9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 143, + 308, + 189, + 345 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "I've added the exclusive EditGuard watermark to it. You can post the picture on the left and I will be responsible for its protection.", + "bbox": [ + 196, + 310, + 341, + 344 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/14ebc880a3f4007db11fd8461f3ab11469a88a059d0a69dcb638c998a78d2614.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 292, + 258, + 341, + 297 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/a2473491b6f8037f8db2e32ad5296ae7fe40b42159dbfece582cd3295334ad29.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 362, + 305, + 413, + 343 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/ee9111f2a171ed5d97957c269b21f285bd9bb5ad367af501064b52780a751e81.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 81, + 352, + 138, + 396 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/416b2131a6a646bcd4b4ea856976f1a7d02bcc4351d5efc5192058194f05ad2a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 145, + 354, + 344, + 401 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/8ff376315174fa9cc4dcbf7c7f8fae287bd2ae36592540a47f66c922389dc984.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 88, + 414, + 130, + 458 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "I find a picture online that looks similar to mine. Can you provide evidence to prove my copyright and pinpoint tampered areas?", + "bbox": [ + 148, + 416, + 290, + 450 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/d39af36db5bf85a178dc5dfc20cf0b681c0da9f4f224560a93bf926d2bd719b1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 292, + 414, + 341, + 452 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/b24586fd356b7e063d993524975730611605f0ae08c4a4601554d75233276b06.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 362, + 460, + 413, + 496 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/560a5bbb01b919df6b94888b46d2f23b43d556222e6129e9f57072eba95a8914.jpg", + "image_caption": [ + "Original" + ], + "image_footnote": [], + "bbox": [ + 429, + 263, + 498, + 316 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/dad32f94dbfa1692713fc314f50d78e9f6ae1d2ff385d81bef20198e81a1f49d.jpg", + "image_caption": [ + "Tampered", + "Remove the wom" + ], + "image_footnote": [], + "bbox": [ + 504, + 263, + 573, + 318 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/24cb7d14e3c5b6e9bda8be532b5ae26396a8f43439fc87fb50a93158f37d47fe.jpg", + "image_caption": [ + "Our Predicted mask" + ], + "image_footnote": [], + "bbox": [ + 581, + 263, + 651, + 318 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/fce68ffe6804dba46fa7e6a3f4ca39cb656916e0619f82e81f1c2852be05af92.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 429, + 329, + 498, + 383 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/42a65b5e23bbd3bfacf3a62b020c01fa03d6a03fcda8805c156fd237dce7fe20.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 500, + 329, + 573, + 383 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/d76bc2f2c1f9c4f6a9ffff0b7f1ef5942951015effd3e05060def371d4432375.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 580, + 329, + 651, + 383 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/46bc5d2965fafcc79eef9aeb40a90a3555cf92ae32dd46f5139ed56306d6eebe.jpg", + "image_caption": [ + "Change the cloth" + ], + "image_footnote": [], + "bbox": [ + 429, + 393, + 498, + 448 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/66b03e5e20470d096f9996835dc598c65dd463d3bb1e22790cb9337a650226d3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 500, + 393, + 573, + 449 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/a7a50709d76a5eefaaf2fb72378462e3f5161ccac2dd89fb63d669e4b75e2c53.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 580, + 393, + 651, + 449 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/ef46a82a571e641ddad2420ae8d83655029caf679edbd3284c710e5a17c7323a.jpg", + "image_caption": [ + "Tamper with AI-generated images" + ], + "image_footnote": [], + "bbox": [ + 429, + 459, + 498, + 491 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/971d94980f5a423b41c9e7dc2fbc6de91c45d50ad53429365bfe38dc41235a4e.jpg", + "image_caption": [ + "Text Forgery", + "Figure 1. We propose a versatile proactive forensics framework EditGuard. The application scenario is shown on the left, wherein users embed invisible watermarks to their images via EditGuard in advance. If suffering tampering, users can defend their rights via the tampered areas and copyright information provided by EditGuard. Some supported tampering methods (marked in blue) and localization results of EditGuard are placed on the right. Our EditGuard can achieve over $95\\%$ localization precision and nearly $100\\%$ copyright accuracy." + ], + "image_footnote": [], + "bbox": [ + 500, + 459, + 573, + 513 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/d90b5b7215fe769f777d32aa4267f0e7d64858149ce81f84aae3ff33bf217968.jpg", + "image_caption": [ + "es, \"A volcano'" + ], + "image_footnote": [], + "bbox": [ + 580, + 459, + 651, + 513 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/1d62c40a83bc886c779407c8a053fdec1369912eb3c5ce8d96fafff48efd4310.jpg", + "image_caption": [ + "Original", + "AI Fill" + ], + "image_footnote": [], + "bbox": [ + 663, + 263, + 733, + 318 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/0bbb5738cc0dbc73621ea30a9504e05c302e3468022b90984db1ae12173ed25e.jpg", + "image_caption": [ + "Tampered", + "Picasso painting" + ], + "image_footnote": [], + "bbox": [ + 736, + 263, + 810, + 318 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/ef5e8fa582c6aeb0ef7f3babbcd898c34d8bc2cbd5ffe9345f140dbb7834ad80.jpg", + "image_caption": [ + "Our Predicted mask", + "on the wall" + ], + "image_footnote": [], + "bbox": [ + 815, + 263, + 885, + 318 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/b55ae49daacc5bf74a9f498b5e1af76d43b516d77223e6a6e30fbaea5bbc93b4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 663, + 329, + 733, + 383 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/4239f38f67c4141fd0b9043c6a100f4452549f3da6935110b6d78f4c63ea9fad.jpg", + "image_caption": [ + "Replace background" + ], + "image_footnote": [], + "bbox": [ + 663, + 393, + 733, + 449 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/ba7f575d98c5ab4ce391c376226cc462b2a4cf5a332b3ffc42477fc7c3c6bf7a.jpg", + "image_caption": [ + "ground, \"A man" + ], + "image_footnote": [], + "bbox": [ + 736, + 329, + 810, + 383 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/2a5384c2a631ee76fd571960084ccd4e9df042886f5ae6724e33eb9f29d3814c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 815, + 329, + 885, + 383 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/d7a5ce36ace3683d775141aa1d2560a26c3f3c151c5effc0dfe4b62da163ec0c.jpg", + "image_caption": [ + "in an office", + "\"A straw hat\"" + ], + "image_footnote": [], + "bbox": [ + 815, + 393, + 885, + 449 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/c852226a6f6691fc2c0bbb154d0f9f2addaeb99d34c00ed03a87681a891a5cea.jpg", + "image_caption": [ + "Secondary creation of artworks" + ], + "image_footnote": [], + "bbox": [ + 663, + 459, + 733, + 513 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/0a54d13aa873acdbc59699f323546ea56bee7fde1c33c7aa9140f6072c2bc195.jpg", + "image_caption": [ + "Swap Face" + ], + "image_footnote": [], + "bbox": [ + 738, + 459, + 810, + 513 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/8863579b2f6447b29ef99f80bc578a2ed5b9550e69081aeddd5b7240ffc19c45.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 815, + 459, + 885, + 513 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 599, + 312, + 616 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In the era of AI-generated content (AIGC), malicious tampering poses imminent threats to copyright integrity and information security. Current deep image watermarking, while widely accepted for safeguarding visual content, can only protect copyright and ensure traceability. They fall short in localizing increasingly realistic image tampering, potentially leading to trust crises, privacy violations, and legal disputes. To solve this challenge, we propose an innovative proactive forensics framework EditGuard, to unify copyright protection and tamper-agnostic localization, especially for AIGC-based editing methods. It can offer a meticulous embedding of imperceptible watermarks and precise decoding of tampered areas and copyright information. Leveraging our observed fragility and locality of image-into-image steganography, the realization of Edit", + "bbox": [ + 73, + 633, + 470, + 861 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Guard can be converted into a united image-bit steganography issue, thus completely decoupling the training process from the tampering types. Extensive experiments verify that our EditGuard balances the tamper localization accuracy, copyright recovery precision, and generalizability to various AIGC-based tampering methods, especially for image forgery that is difficult for the naked eye to detect.", + "bbox": [ + 496, + 601, + 892, + 708 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 500, + 722, + 632, + 738 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Owing to the advantageous properties of diffusion models and the bolstering of extensive datasets, AI-generated content (AIGC) models like DALL-E 3 [14], Imagen [52], and Stable Diffusion [51], can produce lifelike and wondrous images, which brings great convenience to photography enthusiasts and image editors. Nonetheless, the remarkable capabilities of these models come with a double-edged sword, presenting new challenges in copyright protection and information security. The efficiency of image manipulation [43, 47, 49, 51, 55, 68, 72] has blurred the", + "bbox": [ + 496, + 750, + 892, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 44 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "This work was supported by National Natural Science Foundation of China under Grant 62372016. (Corresponding author: Jian Zhang.)", + "bbox": [ + 75, + 875, + 470, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "11964", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "line between fact and forgery, ushering in myriad security and legal concerns. For instance, artistic works are vulnerable to malicious tampering or unauthorized AI-facilitated recreations, making it challenging to protect their original creations [13]. Meanwhile, forged images may be spread online or used as court evidence, causing adverse effects on public opinion, ethical issues, and social stability.", + "bbox": [ + 75, + 90, + 468, + 196 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Given the challenges of preventing image tampering from the source, image watermarking has become a consensus for proactive forensics [54]. However, most prevalent forensic image watermarking [12, 15, 44, 62] still focus on detecting image authenticity or protecting image copyrights, but fall short when it comes to advanced demands, such as localizing tampered areas. Tamper localization facilitates an evaluation of the severity of the image manipulation, and provides an understanding of the intent of attackers, potentially allowing for the partial reuse of the tampered images. However, most passive forensics methods such as previous black-box localization networks [8, 56, 61] tend to seek anomalies like artifacts or flickers in images but struggle to detect more realistic textures and more advanced AIGC models. Moreover, they inevitably need to introduce tampered data during the training and focus solely on specific \"CheapFake\" tampering like slicing and copy-and-paste [8, 45, 56], or on \"DeepFake\" targeting human faces [2, 50], restricted in generalizability. Thus, it is vital to develop an integrated watermarking framework that unites tamper-agnostic localization and copyright protection.", + "bbox": [ + 75, + 198, + 470, + 516 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To clarify our task scope, we re-emphasize the definition of dual forensics tasks as illustrated in Fig. 1: (1) Copyright protection: \"Who does this image belong to?\" We aspire to accurately retrieve the original copyright of an image, even suffering various tampering and degradation. (2) Tamper localization: \"Where was this image manipulated?\" We aim to precisely pinpoint the tampered areas, unrestrained by specific tampering types. To the best of our knowledge, no existing method accomplishes these two tasks simultaneously, while maintaining a balance of high precision and extensive generalizability.", + "bbox": [ + 75, + 517, + 468, + 684 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address this urgent demand, we propose a novel proactive forensics framework, dubbed EditGuard, to protect copyrights and localize tamper areas for AIGC-based editing methods. Specifically, drawing inspiration from our observed locality and fragility of image-into-image (I2I) steganography and inherent robustness of bit-into-image steganography, we can transform the realization of EditGuard into a joint image-bit steganography issue, which allows the training of EditGuard to be entirely decoupled from tampering types, thereby endowing it with exceptional generalizability and locate tampering in a zero-shot manner. In a nutshell, our contributions are as follows:", + "bbox": [ + 75, + 685, + 468, + 864 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "(1) We present the first attempt to design a deep versatile proactive forensics framework EditGuard for univer", + "bbox": [ + 76, + 869, + 468, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "sal tamper localization and copyright protection. It embeds dual invisible watermarks into original images and accurately decodes tampered areas and copyright information.", + "bbox": [ + 496, + 90, + 890, + 136 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "(2) We have observed the fragility and locality of I2I steganography and innovatively convert the solution of this dual forensics task into training a united Image-Bit Steganography Network (IBSN), and utilize the core components of IBSN to construct EditGuard.", + "(3) We introduce a prompt-based posterior estimation module to enhance the localization accuracy and degradation robustness of the proposed framework.", + "(4) The effectiveness of our method has been verified on our constructed dataset and classical benchmarks. Compared to other competitive methods, our approach has notable merits in localization precision, generalization abilities, and copyright accuracy without any labeled data or additional training required for specific tampering types." + ], + "bbox": [ + 496, + 138, + 890, + 354 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related works", + "text_level": 1, + "bbox": [ + 500, + 368, + 643, + 383 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Tamper Localization", + "text_level": 1, + "bbox": [ + 500, + 392, + 694, + 407 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Prevalent passive image forensic techniques have focused on localizing specific types of manipulations [25, 33, 34, 53, 61, 75]. Meanwhile, some universal tamper localization methods [5, 23, 31, 35, 63, 65-67] also tend to explore artifacts and anomalies in tampered images. For instance, MVSS-Net [8] employed multi-view feature learning and multi-scale supervision to jointly exploit boundary artifacts and the noise view of images. OSN [61] proposed a novel robust training scheme to address the challenges posed by lossy operations. Trufor [17] used a learned noise-sensitive fingerprint and extracted both high-level and low-level traces via transformer-based fusion. HiFi-Net [18] utilized multi-branch feature extractor and localization modules for both CNN-synthesized and edited images. SAFL-Net [56] constrained a feature extractor to learn semantic-agnostic features with specific modules and auxiliary tasks. However, the above-mentioned passive localization methods are often limited in terms of generalization and localization accuracy, which usually work on known tampering types that have been trained. Although MaLP [2] used template matching for proactive tamper localization, it still requires a large number of forgery images and cannot fully decouple the network training from the tamper types.", + "bbox": [ + 496, + 415, + 890, + 763 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Image Watermarking", + "text_level": 1, + "bbox": [ + 500, + 771, + 702, + 787 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Image watermarking can be broadly used for the verification, authenticity, and traceability of images. Although traditional fragile watermarking [6, 24, 29, 36, 37, 48] can also achieve block-wise tamper localization, their localization accuracy and flexibility are unsatisfactory. How to realize joint pixel-level tamper localization and copyright protection remains largely unexplored. Owing to the development", + "bbox": [ + 496, + 794, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "11965", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/f39961573342ac5698699c406fd9e570e66dc33b26fe8a12add7aac800fee9d6.jpg", + "image_caption": [ + "Figure 2. Fragility and locality of I2I steganography. The first line shows that when $\\mathbf{X}_{con}^{\\prime}$ is changed, $\\hat{\\mathbf{X}}_{sec}$ will also be fragility demaged. The second line plots the attribution maps $\\frac{1}{|\\mathbb{S}|}\\sum_{(i,j)\\in \\mathbb{S}}\\frac{\\partial\\hat{\\mathbf{X}}_{sec}[i,j]}{\\partial\\mathbf{X}_{con}^{\\prime}}$ of five point sets $\\mathbb{S}$ (marked by $①-⑤$ ) in $\\hat{\\mathbf{X}}_{sec}$ . We observed that $\\hat{\\mathbf{X}}_{sec}$ almost only has a strong response at the corresponding positions of $\\mathbf{X}_{con}^{\\prime}$ and its neighborhoods." + ], + "image_footnote": [], + "bbox": [ + 80, + 89, + 468, + 205 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "of deep learning, deep image watermarking has attracted increased attention. For instance, HiDDeN [74] firstly introduced a deep encoder-decoder network to hide and recover bitstream. Moreover, many distortion layers such as differentiable JPEG and screen-shooting [1, 10, 40, 62] were designed to enhance its robustness. Flow-based models [11, 44] were also utilized to further improve the fidelity of container images. Recently, researchers [7, 12, 27, 59, 73] have designed specialized watermarking mechanisms for large-scale image generation models [51], to merge watermarking into the generation process. However, these deep watermarking methods have a singular function and cannot accurately localize the tampered areas.", + "bbox": [ + 75, + 311, + 470, + 508 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Overall Framework of EditGuard", + "text_level": 1, + "bbox": [ + 76, + 525, + 385, + 542 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Motivation", + "text_level": 1, + "bbox": [ + 76, + 551, + 197, + 566 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Challenges of existing methods: (1) How to equip existing watermark methods, which are solely for copyright protection, with the ability to localize tampering is the crux of EditGuard. We will solve it via the framework design in Sec. 3.2. (2) Most previous tamper localization methods inevitably introduce specific tampering data during network training but tend to raise generalization concerns in unknown tampering types, which will be addressed in Sec. 3.3.", + "bbox": [ + 75, + 575, + 468, + 696 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our observations: Fortunately, we observed that imageinto-image (I2I) steganography exhibits distinct fragility and locality, possessing great potential to address these issues. Concretly, I2I steganography [3, 28, 42, 46, 64, 69] aims to hide a secret image $\\mathbf{X}_{\\text{sec}}$ into a cover image $\\mathbf{X}_{\\text{cov}}$ to produce a container image $\\mathbf{X}_{\\text{con}}$ , and reveal $\\hat{\\mathbf{X}}_{\\text{sec}}$ and $\\hat{\\mathbf{X}}_{\\text{cov}}$ with minimal distortion from the received image $\\mathbf{X}_{\\text{con}}^{\\prime}$ . We discover that when $\\mathbf{X}_{\\text{con}}^{\\prime}$ undergoes significant alteration compared to $\\mathbf{X}_{\\text{con}}$ , $\\hat{\\mathbf{X}}_{\\text{sec}}$ will also be damaged and generate artifacts (the first row of Fig. 2), which is called fragility. Furthermore, we notice that the artifacts in $\\hat{\\mathbf{X}}_{\\text{sec}}$ are almost pixel-level corresponding to the changes in $\\mathbf{X}_{\\text{con}}^{\\prime}$ relative to $\\mathbf{X}_{\\text{con}}$ , which is called locality. To demonstrate this locality,", + "bbox": [ + 75, + 704, + 470, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "we select five $7 \\times 7$ point sets on $\\hat{\\mathbf{X}}_{\\text {sec }}$ and calculated their attribution maps with respect to $\\mathbf{X}_{\\text {con }}^{\\prime}$ . As plotted in the second row of Fig. 2, $\\hat{\\mathbf{X}}_{\\text {sec }}$ only exhibits a strong response at the corresponding locations of $\\mathbf{X}_{\\text {con }}^{\\prime}$ and their immediate vicinity, almost irrelevant to other pixels. These properties inspire us to treat $\\mathbf{X}_{\\text {sec }}$ as a special localization watermark and embed it within existing watermarking frameworks.", + "bbox": [ + 496, + 90, + 890, + 196 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Framework Design and Forensics Process", + "text_level": 1, + "bbox": [ + 498, + 209, + 854, + 224 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To realize united tamper localization and copyright protection, EditGuard is envisioned to embed both a 2D localization watermark and a 1D copyright traceability watermark into the original image in an imperceptible manner, which allows the decoding end to obtain the copyright of the images and a binary mask reflecting tampered areas. However, designing such a framework needs to solve the compatibility issue of two types of watermarks.", + "bbox": [ + 496, + 232, + 890, + 352 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "(1) Local vs. Global: The localization watermark is required to be hidden in the corresponding pixel positions of the original image, while the copyright watermark should be unrelated to spatial location and embedded in the global area redundantly. (2) Semi-fragile vs. Robust: The desired attribute of the localization watermark is semi-fragile, which means it is fragile to tampering but robust against some common degradations (such as Gaussian noise, JPEG compression, and Poisson noise) during network transmission. However, the copyright should be extracted nearly losslessly, irrespective of tampering or degradation.", + "bbox": [ + 496, + 354, + 890, + 520 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To address the two pivotal conflicts, EditGuard employs a \"sequential encoding and parallel decoding\" structure, which comprises a dual-watermark encoder, a tamper locator, and a copyright extractor. As shown in Fig. 3, the dual-watermark encoder will sequentially add pre-defined localization watermark and global copyright watermark $\\mathbf{w}_{cop}$ provided by users to the original image $\\mathbf{I}_{ori}$ , forming the container image $\\mathbf{I}_{con}$ . Our experiments have proved that parallel encoding cannot effectively add dual watermarks into images (in supplementary materials (S.M.). In contrast, sequential embedding effectively prevents cross-interference by hiding these two watermarks. Furthermore, we model the network transmission process in which the received (tampered) image $\\mathbf{I}_{rec}$ is transformed from $\\mathbf{I}_{con}$ as:", + "bbox": [ + 496, + 521, + 890, + 733 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {I} _ {r e c} = \\mathcal {D} \\left(\\mathbf {I} _ {c o n} \\odot (\\mathbf {1} - \\mathbf {M}) + \\mathcal {T} \\left(\\mathbf {I} _ {c o n}\\right) \\odot \\mathbf {M}\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 542, + 739, + 890, + 755 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mathcal{T}(\\cdot),\\mathcal{D}(\\cdot)$ and $\\mathbf{M}$ respectively denote the tamper function, degradation operation, and tempered mask. Moreover, the parallel decoding processes allow us to flexibly train each branch under different levels of robustness and obtain the predicted mask $\\hat{\\mathbf{M}}$ via the tamper locator and traceability watermark $\\hat{\\mathbf{w}}_{cop}$ via the copyright extractor. We can categorize the dual forensic process of EditGuard into the following scenarios.", + "bbox": [ + 496, + 761, + 890, + 882 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Case 1: If $\\hat{\\mathbf{w}}_{\\text{cop}} \\not\\simeq \\mathbf{w}_{\\text{cop}}$ , suspicious $\\mathbf{I}_{\\text{rec}}$ is either not", + "bbox": [ + 500, + 885, + 890, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "11966", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/86330ab304ab8ee076a9aed3f95af6473f94c3b7e495a949fe4060b129a8b924.jpg", + "image_caption": [ + "Figure 3. Illustration of the proposed proactive forensics framework EditGuard. The dual-watermark encoder sequentially embeds the pre-defined localization watermark and copyright watermark $\\mathbf{w}_{\\text{cop}}$ into the original image $\\mathbf{I}_{\\text{ori}}$ , generating the container image $\\mathbf{I}_{\\text{con}}$ . After encountering potential malicious tampering and degradation during network transmission, tampered mask $\\hat{\\mathbf{M}}$ and copyright information $\\hat{\\mathbf{W}}_{\\text{cop}}$ are respectively extracted via the tamper locator and copyright extractor from the received image $\\mathbf{I}_{\\text{rec}}$ ." + ], + "image_footnote": [], + "bbox": [ + 81, + 88, + 893, + 253 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "registered in our EditGuard or underwent extremely severe global tampering, rendering it unreliable as evidence.", + "bbox": [ + 75, + 324, + 468, + 354 + ], + "page_idx": 3 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Case 2: If $\\hat{\\mathbf{w}}_{\\text {cop }} \\approx \\mathbf{w}_{\\text {cop }}$ and $\\hat{\\mathbf{M}} \\not\\approx \\mathbf{0}$ , suspicious $\\mathbf{I}_{\\text {rec }}$ has undergone tampering, disqualifying it as valid evidence. Users may infer the intention of tamperers based on $\\hat{\\mathbf{M}}$ and decide whether to reuse parts of the image.", + "$\\square$ Case 3: If $\\hat{\\mathbf{w}}_{cop} \\approx \\mathbf{w}_{cop}$ and $\\hat{\\mathbf{M}} \\approx \\mathbf{0}$ , $\\mathbf{I}_{rec}$ remains untampered and trustworthy under the shield of EditGuard." + ], + "bbox": [ + 75, + 354, + 468, + 446 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Transform Dual Forensics into Steganography", + "text_level": 1, + "bbox": [ + 76, + 455, + 467, + 470 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To realize universal and tamper-agnostic localization, we resort to our observed locality and fragility of I2I steganography. As described in Sec. 3.1, localization watermarking and tamper locator in Fig. 3 can be effectively realized via image hiding and revealing. Meanwhile, combining with the robustness of current bit-into-image steganography, copyright watermarking and extractor in Fig. 3 are achieved via bit encryption and recovery. Thus, we can convert the realization of the dual forensics framework EditGuard into a united image-bit steganography network. Our training objective is just a self-recovery mechanism, meaning it only needs to ensure the input and output of the steganography network maintain high fidelity under various robustness levels, with no need to introduce any labeled data or tampered samples. During inference, it can naturally locate tampering via simple comparisons in a zero-shot manner and extract copyright exactly.", + "bbox": [ + 75, + 478, + 472, + 734 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. United Image-bit Steganography Network", + "text_level": 1, + "bbox": [ + 76, + 746, + 452, + 763 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1. Network Architecture", + "text_level": 1, + "bbox": [ + 76, + 771, + 282, + 786 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As plotted in Fig. 4, the proposed IBSN includes an image hiding module (IHM), a bit encryption module (BEM), a bit recovery module (BRM), and an image revealing module (IRM). First, the IHM aims to hide a localization watermark $\\mathbf{W}_{loc} \\in \\mathbb{R}^{H \\times W \\times 3}$ into the original image $\\mathbf{I}_{ori} \\in \\mathbb{R}^{H \\times W \\times 3}$ , resulting in an intermediate output $\\mathbf{I}_{med} \\in \\mathbb{R}^{H \\times W \\times 3}$ . Subsequently, $\\mathbf{I}_{med}$ is fed to the BEM for feature refinement,", + "bbox": [ + 75, + 794, + 468, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "while the copyright watermark $\\mathbf{w}_{cop} \\in \\{0,1\\}^L$ is modulated into the BEM, forming the final container image $\\mathbf{I}_{con} \\in \\mathbb{R}^{H \\times W \\times 3}$ . After network transmission, the BRM will faithfully reconstruct the copyright watermark $\\hat{\\mathbf{w}}_{cop}$ from the received container image $\\mathbf{I}_{rec}$ . Meanwhile, $\\mathbf{I}_{rec}$ predicts the missing information $\\tilde{\\mathbf{Z}}$ via the prompt-based posterior estimation and uses it as the initialization for the invertible blocks, producing $\\hat{\\mathbf{I}}_{ori}$ and semi-fragile watermark $\\hat{\\mathbf{W}}_{loc}$ .", + "bbox": [ + 496, + 324, + 890, + 446 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.2. Invertible Blocks in IHM and IRM", + "text_level": 1, + "bbox": [ + 498, + 452, + 803, + 467 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Given the inherent capacity of flow-based models to precisely recover multimedia information, we harness stacked invertible blocks to construct image hiding and revealing modules. The original image $\\mathbf{I}_{ori} \\in \\mathbb{R}^{H \\times W \\times 3}$ and localization watermark $\\mathbf{W}_{loc} \\in \\mathbb{R}^{H \\times W \\times 3}$ will undergo discrete wavelet transformations (DWT) to yield frequency-decoupled image features. We then employ enhanced additive affine coupling layers to project the original image and its corresponding localization watermark branches. The transformation parameters are generated from each other. The enhanced affine coupling layer is composed of a five-layer dense convolution block [46] and a lightweight feature interaction module (LFIM) [4]. The LFIM can enhance the non-linearity of transformations and capture the long-range dependencies with low computational cost. More details are presented in S.M. Finally, the revealed features are then transformed to the image domain via the inverse DWT.", + "bbox": [ + 496, + 476, + 890, + 733 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.3. Prompt-based Posterior Estimation", + "text_level": 1, + "bbox": [ + 498, + 741, + 807, + 757 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To bolster the fidelity and robustness of the image hiding and revealing module, we introduce a degradation prompt-based posteriori estimation module (PPEM). Since the encoding network tends to compress $\\left[\\mathbf{I}_{ori};\\mathbf{W}_{loc}\\right] \\in \\mathbb{R}^{H \\times W \\times 6}$ into the container image $\\mathbf{I}_{con} \\in \\mathbb{R}^{H \\times W \\times 3}$ , previous methods [42, 64] typically utilized a random Gaussian initialization or an all-zero map at the decoding end to compensate for the lost high-frequency channels. Yet, our observations suggest that discarded information lurks within the", + "bbox": [ + 496, + 763, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "11967", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/01b5c21e4e64812cbbbcd0b7eb5a9a74b9248afa151d2966ce8537cfa107bdb0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 83, + 89, + 890, + 271 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/8d9ee22d3b133e6ac1c8906b3431073c5a2484c614038f51aa7f79bf152441c1.jpg", + "image_caption": [ + "Figure 4. Illustration of the united Image-bit Steganography Network (IBSN). In the training process, we randomly sample original image $\\mathbf{I}_{ori}$ , localization watermark $\\mathbf{W}_{loc}$ (a natural RGB image) and copyright watermark $\\mathbf{w}_{cop}$ and expect the IBSN to recover $\\hat{\\mathbf{I}}_{ori}$ , $\\hat{\\mathbf{W}}_{loc}$ and $\\hat{\\mathbf{w}}_{cop}$ with high fidelity. In the inference process, we use core components of the pre-trained IBSN with a mask extractor (ME) to construct our EditGuard, and pre-define a simple solid color image as a localization watermark." + ], + "image_footnote": [], + "bbox": [ + 83, + 275, + 887, + 431 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/9b662d95305022ca97a793ae622e6d31d1bc933c050499a507cebfed03184167.jpg", + "image_caption": [ + "Figure 5. Illustration of the proposed prompt-based posterior estimation. It will dynamically fuse degraded representations and extracted features to obtain posterior mean $\\hat{\\mathbf{Z}} = \\mathbb{E}[\\mathbf{Z}|\\mathbf{I}_{rec}]$ ." + ], + "image_footnote": [], + "bbox": [ + 76, + 498, + 468, + 647 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "edges and textures of the container image. Thus, deploying a dedicated network proves to be a more potent strategy in predicting the posterior mean of the vanished localization watermark information $\\hat{\\mathbf{Z}} = \\mathbb{E}[\\mathbf{Z}|\\mathbf{I}_{rec}]$ . Specifically, as shown in Fig. 5, we stack $M$ residual blocks $\\mathrm{Res}(\\cdot)$ [19] and $M$ channel-wise transformer blocks $\\mathrm{Trans}(\\cdot)$ [70] to extract the local and non-local features $\\mathbf{F}_c$ .", + "bbox": [ + 75, + 696, + 468, + 800 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {F} _ {c} = \\operatorname {T r a n s} (\\operatorname {R e s} (\\mathrm {D W T} (\\mathbf {I} _ {r e c}))) + \\operatorname {R e s} (\\mathrm {D W T} (\\mathbf {I} _ {r e c})). \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 84, + 804, + 468, + 821 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Considering that the container image is prone to various degradations during network transmission, we pre-define $N$ learnable embedding tensors as degradation prompts $\\mathbf{P} = [\\mathbf{P}_1,\\mathbf{P}_2,\\dots ,\\mathbf{P}_N]$ , where $N$ denotes the number of degradation types and is set to 3. These learned prompts $\\mathbf{P}$ can", + "bbox": [ + 76, + 825, + 468, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "adaptively learn a diverse range of degradation representations and are integrated with the intrinsic features extracted from $\\mathbf{I}_{rec}$ , enabling the proposed IBSN to handle multiple types of degradations using a single set of parameters. To better foster the interaction between the input features $\\mathbf{F}_c$ and the degradation prompt $\\mathbf{P}$ , the features $\\mathbf{F}_c$ are passed to a global average pooling (GAP) layer, a $1\\times 1$ convolution, and a softmax operator to produce a set of dynamic weight coefficients. Each degradation prompt $\\mathbf{P}_i$ is combined using these dynamic coefficients $\\mathbf{w}_{p\\otimes i}$ and subsequently integrated via an upsampling operator $\\uparrow$ and $3\\times 3$ convolution to obtain the enhanced representation $\\mathbf{P}_c$ .", + "bbox": [ + 496, + 500, + 890, + 681 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {P} _ {c} = \\operatorname {C o n v} _ {3 \\times 3} \\left(\\left(\\sum_ {i = 1} ^ {N} \\mathbf {w} _ {p * i} \\mathbf {P} _ {i}\\right) _ {\\uparrow}\\right), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 562, + 685, + 890, + 714 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathbf{w}_p = \\mathrm{Softmax}\\left(\\mathrm{Conv}_{1\\times 1}\\left(\\mathrm{GAP}\\left(\\mathbf{F}_c\\right)\\right)\\right)$", + "bbox": [ + 519, + 715, + 846, + 731 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Finally, we utilize a $3 \\times 3$ convolution to fuse the learned degradation representation conditioned on the extracted features $\\mathbf{F}_c$ to enrich the degradation-specific context, obtaining $\\hat{\\mathbf{Z}}$ . This process can be formulated as:", + "bbox": [ + 496, + 736, + 890, + 796 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathbf {Z}} = \\operatorname {C o n v} _ {3 \\times 3} \\left(\\left[ \\mathbf {P} _ {c}; \\mathbf {F} _ {c} \\right]\\right) \\in \\mathbb {R} ^ {\\frac {H}{2} \\times \\frac {W}{2} \\times 1 2}. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 560, + 800, + 890, + 819 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.4. Bit Encryption and Recovery Modules", + "text_level": 1, + "bbox": [ + 500, + 830, + 828, + 847 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As shown in Fig. 4, to encode the copyright watermark $\\mathbf{w}_{\\text{cop}}$ into $\\mathbf{I}_{\\text{med}}$ , we firstly expand $\\mathbf{w}_{\\text{cop}} \\in \\{0,1\\}^L$ via stacked MLPs and reshape it into several $L \\times L$ message feature", + "bbox": [ + 500, + 854, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "11968", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/d444b17703fece8231bf655fa0b91478d28d1ef369d55100bf1e0b8b801db008.jpg", + "table_caption": [ + "Table 1. Localization precision (F1-Score) and bit accuracy (BA) comparison with other competitive methods on [9, 16, 20, 58]." + ], + "table_footnote": [], + "table_body": "
MethodCAISAv1 [9]Coverage [58]NIST16 [16]Columbia [20]
F1BA(%)F1BA(%)F1BA(%)F1BA(%)
ManTraNet [63]0.320-0.486-0.225-0.650-
SPAN [22]0.169-0.428-0.363-0.873-
CAT-Net v2 [32]0.852-0.582-0.417-0.923-
OSN [61]0.676-0.472-0.449-0.836-
MVSS-Net [8]0.650-0.659-0.372-0.781-
PSCC-Net [39]0.670-0.615-0.210-0.760-
TruFor [17]0.822-0.735-0.470-0.914-
EditGuard (Ours)0.95499.910.9551000.91199.880.98899.93
", + "bbox": [ + 80, + 122, + 468, + 222 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/ac93b3cda087d4ec26260b5d1f14f6c77c235ad6139be81c5a78ef10a189739f.jpg", + "image_caption": [ + "Figure 6. Localization precision comparisons of our EditGuard and competitive methods [8, 39, 61] on four classical benchmarks." + ], + "image_footnote": [], + "bbox": [ + 81, + 227, + 467, + 388 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "maps. Meanwhile, $\\mathbf{I}_{med}$ is fed to a U-style feature enhancement network to extract features of each downsampling and upsampling layer. Finally, the message features will be upscaled and integrated with multi-level image features via the fusion mechanism [21, 62], achieving the modulation of bit-image information. In the decoding end, $\\mathbf{I}_{rec}$ is fed into a U-shaped sub-network and downsampled to a size of $L\\times L$ . The recovered copyright watermark $\\hat{\\mathbf{w}}_{cop}$ is then extracted via an MLP. More details are presented in S.M..", + "bbox": [ + 75, + 425, + 472, + 561 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.5. Construct EditGuard via the IBSN", + "text_level": 1, + "bbox": [ + 76, + 569, + 382, + 584 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To stabilize the optimization of the proposed IBSN, we propose a bi-level optimization strategy. Given an arbitrary original image $\\mathbf{I}_{med}$ and watermark $\\mathbf{w}_{cop}$ , we first train the bit encryption and recovery module via the $\\ell_{2}$ loss.", + "bbox": [ + 75, + 592, + 468, + 654 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\ell_ {c o p} = \\left\\| \\mathbf {I} _ {c o n} - \\mathbf {I} _ {m e d} \\right\\| _ {2} ^ {2} + \\lambda \\left\\| \\hat {\\mathbf {w}} _ {c o p} - \\mathbf {w} _ {c o p} \\right\\| _ {2} ^ {2}, \\quad (5)\n$$\n", + "text_format": "latex", + "bbox": [ + 119, + 656, + 468, + 676 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\lambda$ is set to 10. Furthermore, we freeze the weights of BEM and BRM and jointly train the IHM and IRM. Given a random original image $\\mathbf{I}_{ori}$ , localization watermark $\\mathbf{W}_{loc}$ and copyright watermark $\\mathbf{w}_{cop}$ , the loss function is:", + "bbox": [ + 75, + 678, + 468, + 739 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\ell_ {l o c} = \\left\\| \\hat {\\mathbf {I}} _ {o r i} - \\mathbf {I} _ {o r i} \\right\\| _ {1} + \\alpha \\left\\| \\mathbf {I} _ {c o n} - \\mathbf {I} _ {o r i} \\right\\| _ {2} ^ {2} + \\beta \\left\\| \\hat {\\mathbf {W}} _ {l o c} - \\mathbf {W} _ {l o c} \\right\\| _ {1}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 83, + 744, + 468, + 760 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\alpha$ and $\\beta$ are respectively set to 100 and 1. During training, we only introduced degradations to $\\mathbf{I}_{con}$ without being exposed to any tampering. After acquiring a pretrained IBSN, we can construct the proposed EditGuard via the components of IBSN. As plotted in Fig. 4, the dual-watermark encoder of EditGuard is composed of IHM and BEM, which correspond to the localization and copyright watermarking in Fig. 3 respectively. The copyright extractor strictly corresponds to BRM. The tamper locator", + "bbox": [ + 75, + 763, + 470, + 901 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/533fae2f42952dbc6a7275fde1683a63874f847939ae1901f747d96464322e1b.jpg", + "table_caption": [ + "Table 2. Visual quality of the container image $\\mathbf{I}_{con}$ and bit accuracy comparison with other pure watermarking methods." + ], + "table_footnote": [], + "table_body": "
MethodImage SizeM. L.PSNR (dB)SSIMNIQE(↓)BA(%)
MBRS [26]128×1283026.570.8867.219100
CIN [44]128×1283041.350.9817.17199.99
PIMoG [10]128×1283036.220.9417.11399.99
SepMark [62]128×1283035.420.9317.09599.86
EditGuard128×1283036.930.9445.56799.89
EditGuard512×5126437.770.9494.25799.95
", + "bbox": [ + 503, + 121, + 890, + 203 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "includes IRM and a mask extractor (ME). Note that we need to pre-define a localization watermark $\\mathbf{W}_{loc}$ , which is shared between the encoding and decoding ends. The choice of $\\mathbf{W}_{loc}$ is very general to our method. It can be any natural image or even a solid color image. Finally, by comparing the pre-defined watermark $\\mathbf{W}_{loc}$ with the decoded one $\\hat{\\mathbf{W}}_{loc}$ , we can obtain a binary mask $\\hat{\\mathbf{M}}\\in \\mathbb{R}^{H\\times W}$ :", + "bbox": [ + 496, + 207, + 890, + 313 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathbf {M}} [ i, j ] = \\theta_ {\\tau} \\left(\\max \\left(\\left| \\hat {\\mathbf {W}} _ {l o c} [ i, j,: ] - \\mathbf {W} _ {l o c} [ i, j,: ] \\right|\\right)\\right). \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 517, + 316, + 890, + 334 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $i\\in [0,H)$ and $j\\in [0,W)$ . $\\theta_{\\tau}(z) = 1$ ( $z\\geq \\tau$ ). $\\tau$ is set to 0.2. $|\\cdot |$ is an absolute value operation.", + "bbox": [ + 496, + 337, + 890, + 369 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 500, + 382, + 633, + 398 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1. Implementation Details", + "text_level": 1, + "bbox": [ + 500, + 407, + 718, + 422 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We trained our EditGuard via the training set of COCO [38] without any tampered data. Thus, for tamper localization, our method is actually zero-shot. The Adam [30] is used for training $250K$ iterations with $\\beta_{1} = 0.9$ and $\\beta_{2} = 0.5$ . The learning rate is initialized to $1\\times 10^{-4}$ and decreases by half for every $30K$ iterations, with the batch size set to 4. We embed a 64-bit copyright watermark and a simple localization watermark such as a pure blue image ([R, G, B] = [0, 0, 255]) to original images. Following [8, 17, 39], F1-score, AUC, IoU, and bit accuracy are used to evaluate localization and copyright protection performance. Since no prior methods can simultaneously achieve this dual forensics, we conducted separate comparisons with tamper localization and image watermarking methods.", + "bbox": [ + 496, + 430, + 890, + 642 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.2. Comparison with Localization Methods", + "text_level": 1, + "bbox": [ + 500, + 650, + 841, + 666 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For a fair comparison with tamper localization methods, we conducted extensive evaluations on four classical benchmarks [9, 16, 20, 58], as reported on Tab. 1. Since EditGuard is a proactive approach, we initially embed watermarks into authentic images and then paste the tampered areas into the container images. Remarkably, even for tamper types that existing methods specialize in, the localization accuracy of EditGuard consistently outperforms the SOTA method [17] across four datasets by margins of 0.102, 0.116, 0.441, and 0.065 in F1-score without any labeled data or tampered samples required, which verifies the superiority of our proactive localization mechanism. As shown in Fig. 6, our EditGuard can precisely pinpoint pixel-level tampered areas but other methods can only produce a rough outline or are only effective in some cases. Mean-", + "bbox": [ + 496, + 674, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "11969", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/ce030315e7bf53ca422742f2f12f38fb43f40f250dbab4c1d47c813512e4ca66.jpg", + "table_caption": [ + "Table 3. Comparison with other competitive tamper localization methods under different AIGC-based editing methods. Note that $\\dagger$ denotes the network finetuned in our constructed AGE-Set-C." + ], + "table_footnote": [], + "table_body": "
MethodStable Diffusion Inpaint [51]Controlnet [72]SDXL [49]RePaint [43]Lama [57]FaceSwap [60]
F1AUCIoUBA(%)F1AUCIoUBA(%)F1AUCIoUBA(%)F1AUCIoUBA(%)F1AUCIoUBA(%)F1AUCIoUBA(%)
MVSS-Net [8]0.1780.4880.103-0.1780.4920.103-0.0370.5030.028-0.1040.5460.082-0.0240.5050.022-0.2850.6120.192-
OSN [61]0.1740.4860.101-0.1910.6440.110-0.2000.7550.118-0.1830.6440.105-0.1700.4300.099-0.3080.7910.171-
PSCC-Net [39]0.1660.5010.112-0.1770.5650.116-0.1890.7040.115-0.1400.4690.109-0.1320.3290.104-0.1570.3460.180-
IML-VIT [45]0.2130.5960.135-0.2000.5760.128-0.2210.6030.145-0.1030.4970.059-0.1050.4650.064-0.1050.4650.064-
HiFi-Net [18]0.5470.7340.128-0.5420.7350.123-0.6330.8280.261-0.6810.8960.339-0.4830.7210.029-0.7810.8900.478-
MVSS-Net†[8]0.6940.9390.575-0.6780.9250.558-0.4820.8840.359-0.1850.5290.111-0.3930.8290.275-0.4590.7390.333-
EditGuard (Ours)0.9660.9710.93699.950.9680.9870.94099.960.9650.9890.93699.960.9670.9770.93899.950.9650.9690.93499.950.8960.9430.87699.86
", + "bbox": [ + 80, + 116, + 890, + 194 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/95875c5849a1cdbc6277286d2e769a7f1bd18d6ddaca9783cff1641b55880ea9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 81, + 200, + 161, + 256 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e42e68e144df1134327fbfcd2269b66657f377933b8c3fc5b2785224624d2386.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 163, + 200, + 236, + 257 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/fa30e311be2c9f3547a7b8c2161d1dfc5bf76dbf7f6ebb1f737811bf611765dc.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 241, + 200, + 315, + 257 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/83cb8d90245a0c0fdedf521e2b149960657dc5bae5a30eee4f340c70cf4e5e8e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 316, + 200, + 390, + 257 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/4bd55197d7c7b580de6b9d301db508e36d01df31e4ac53d4df4977dd2649048d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 393, + 200, + 467, + 257 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/1485a6728911b0188fb0bf86050c79bdae5b51c24a940aeb06082d5c4eb82843.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 80, + 260, + 161, + 316 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/dc9ae763a7b3ccf0c9aaae7c79651e92be21477f8d2253256501671c9512a36f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 165, + 260, + 236, + 316 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/41c00b7712cda2d2a90ef7fd3b6fc93bb44371f993c44847f11653440f36d9d4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 241, + 260, + 315, + 316 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/9b23195e6a3bd779476e07957bf8724e8326a4245ca6915e44f1728542337b0f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 316, + 260, + 390, + 316 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/efb52bc2d62e35da7636a0034a91a419b7900d3bb22e746ef6b36778ea57a0f2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 393, + 260, + 467, + 316 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/5ac68cb476e3c8b7111c8ea60d546a3b431b987f0854fb19fbea659d4803e0dd.jpg", + "image_caption": [ + "Figure 7. Visual results of the container image $\\mathbf{I}_{con}$ and the error map of the proposed EditGuard. Here, localization and copyright watermarks are randomly selected from the dataset." + ], + "image_footnote": [], + "bbox": [ + 80, + 319, + 161, + 375 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e1e1c8519147b61a444e57c089b3f515b3409cfed275c56496039ec91406b4d5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 165, + 319, + 236, + 375 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/450cda36558f90764ebfd9de06e5bbc54b1e1324e4813945fa25c2f9cb66e9ef.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 241, + 320, + 315, + 375 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ef1ebd48de8354d61bbaa44b137bf3f916d54f689186b9b4e920e1288961fd98.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 316, + 320, + 390, + 375 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a70af4c7c227f03986671dd12dd5b3528f9fc16cc7ba5282eb3468b33973902c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 393, + 320, + 467, + 375 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "while, our bit accuracy remains over $99.8\\%$ while all other methods can not realize effective copyright protection.", + "bbox": [ + 76, + 424, + 468, + 455 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.3. Comparison with Watermarking Methods", + "text_level": 1, + "bbox": [ + 76, + 465, + 434, + 481 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To evaluate the visual quality of $\\mathbf{I}_{con}$ , we compared EditGuard with other watermarking methods on $1K$ testing images from COCO [38] under the tampering of stable diffusion inpaint [51]. For a fair comparison, we also retrained our EditGuard on $128\\times 128$ original images and 30 bits. Tab. 2 reports that the fidelity of our container image far surpasses that of SepMark [62], PIMoG [10], and MBRS [26] but is slightly inferior to CIN [44]. Meanwhile, our method exhibits the best performance in perceptual quality measures like NIQE. As shown in Fig. 7, dual-watermarked images do not have noticeable artifacts and noise, making them imperceptible to the human eyes. When suffer malicious tampering, our method outperforms SepMark and is very close to PIMoG and CIN in bit accuracy. Note that other competitive methods only hide 30 bits, with a capacity of $30 / (128\\times 128)$ . In contrast, our EditGuard hides both an RGB localization watermark and a 1D copyright watermark, with a capacity far greater than $30 / (128\\times 128)$ . Here, we do not claim to achieve the best visual quality and bit accuracy, but just to demonstrate that our method is comparable to the current image watermarking methods.", + "bbox": [ + 75, + 488, + 468, + 806 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.4. Extension to AIGC-based Editing Methods", + "text_level": 1, + "bbox": [ + 76, + 816, + 442, + 832 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Dataset Preparation: We constructed a dataset tailored for AIGC Editing methods, dubbed AGE-Set, comprising two sub-datasets. The first AGE-Set-C is a batch-processed coarse tamper dataset. Its original images are sourced from", + "bbox": [ + 75, + 839, + 468, + 900 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a3d698dd3b7233fb982193bec98a5b9b6dc00da38e36f29d78b369f7739469f6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 501, + 200, + 604, + 242 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/2aa366912735e6902300754de3efbc2c7f8356075cae02d1740b9f8804c9efc4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 607, + 200, + 699, + 242 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/3fdedada16bcdab9164cd71199ca49fac996c152395eccc4c710bbba918fd5a0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 700, + 200, + 769, + 242 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/816a0d20ac6825bea0b8fee43275fcdc5903d35dac32e6b5462be513437c2261.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 771, + 200, + 813, + 242 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/fc4edcaef943973475910003308f05107398737bad26fd2de14f34d2c22e4594.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 816, + 200, + 888, + 242 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/4738da0e0397918d0663e84d28ad244b867a07e02fa7d91b40baaf698622d151.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 503, + 252, + 604, + 279 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d6bc9a0a9ad2fc1b7d3b3810c9af56a31f59f9c54be1dd9981caaefe5315c94e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 606, + 244, + 699, + 279 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a79116708e6ace07e34d9cd9b405a64ae2cde5d10c4220e84c866048561cda8a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 700, + 244, + 769, + 279 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/97136b7504c476d3f733252490dc8f28d6d237f3d713b80182826d44d2576899.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 771, + 244, + 813, + 279 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/61921a5bab7ef5eaabb3e6a77529d830d0a262dd2ff06037c2b592da94a9a45d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 816, + 244, + 888, + 279 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/3de38109b2947ebb0cfec34afd0a8030d377a8e777ee835ed30dd6d6153abb94.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 504, + 280, + 604, + 315 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/978c7704ed9dbc9ad97cef580bcb0f5ba4e75af3c2acc01a247fc8f5a54156fe.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 606, + 280, + 699, + 315 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c1b99e11191fbba0ca9ba1cc538d606061a288535ce85654b7b5054d7db1d2e4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 700, + 280, + 769, + 315 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d0ab8577cabc01bcd57112c4e17f47b8c1868b5c4403a245f1ce02858322b3d4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 771, + 280, + 813, + 315 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/fb07c2ec7d2fad52ba53093c3e6d0e7fc668b4719e15a082e3ace33a0ffe1804.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 816, + 280, + 888, + 315 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/43aebf93db511ac8ad02cd3c7a18779ac604d36214cb16e54407fd3242cf9acf.jpg", + "image_caption": [ + "Figure 8. Localization performance comparisons of our EditGuard and other methods [8, 18, 39, 61] on our constructed AGE-Set-C." + ], + "image_footnote": [], + "bbox": [ + 503, + 316, + 604, + 352 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/64aebf924695532abcc878f1d7ff14b5c84a341fdd4e1781705592337fc5bb07.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 606, + 316, + 699, + 352 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/510f586af59a8dd1331cfadacf30af9acb24d8412991393df83d34fa7cbb50bd.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 700, + 316, + 769, + 352 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/1961807dc7f2143728f3e9bd78759d2a60e337895b39be63ca524b3e73376d08.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 771, + 316, + 813, + 352 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/45c47d0401b8e85bf542c59842412290bef15467881628beb41afebe3bf1039d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 816, + 316, + 888, + 352 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "COCO 2017 [38] and CelebA [41], containing $30K$ training images and $1.2K$ testing images. We used some SOTA editing methods such as Stable Diffusion Inpaint [51], ControlNet [72], SDXL [49] to manipulate images with the prompt to be \"None\", and employed some unconditional methods like Repaint [43], Lama [57], and Faceswap [60]. Note that we only use the tampered data to train other methods, not our EditGuard. The second sub-dataset AGE-Set-F includes 100 finely edited images. It is edited manually via some sophisticated software such as SD-Web-UI, Photoshop, and Adobe Firefly. These AIGC-based editing methods can achieve a good fusion of the tampered and unchanged areas, making it hard for the naked eye to catch artifacts. More details are presented in S.M..", + "bbox": [ + 496, + 386, + 890, + 598 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "AGE-Set-C: Tab. 3 presents the comparison of our EditGuard and some SOTA tamper localization methods [8, 18, 39, 45, 61]. We observe that the F1-scores of other passive forensic methods are generally lower than 0.7 when applied to AGE-Set-C. Meanwhile, even when we try our best to finetune MVSS-Net using AGE-Set-C, the accuracy of MVSS-Net remains unsatisfactory, and they exhibit catastrophic forgetting across various tamper methods. In contrast, our method can guarantee an F1-score and AUC of over $95\\%$ , maintaining around $90\\%$ IOU, regardless of tampering types. As shown in Fig. 8, our EditGuard can accurately capture these imperceptible tampering traces produced by AIGC-based editing methods, but other methods are almost ineffective. Moreover, our EditGuard can effectively recover copyright information with a bit accuracy exceeding $99.8\\%$ . Noting that none of the comparison methods offer copyright protection capabilities.", + "bbox": [ + 496, + 598, + 892, + 854 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "AGE-Set-F: To further highlight the practicality of our EditGuard, we conducted subjective comparisons with other methods on the meticulously tampered AGE-Set-F.", + "bbox": [ + 496, + 854, + 892, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "11970", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/bce070c0cef5d5ad87cc6a09e540d2e125882f389da237eb7590e82583cf7c0a.jpg", + "image_caption": [ + "Figure 9. Localization precision comparisons of our EditGuard and other competitive methods on the meticulously tampered AGE-Set-F." + ], + "image_footnote": [], + "bbox": [ + 80, + 89, + 888, + 340 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/ab61ec84023430adfc76ec15cc0af35310b4847355245841e38a3b2b129c7708.jpg", + "table_caption": [ + "Table 4. Localization and bit recovery performance of our Edit-Guard and MVSS-Net† [8] under different levels of degradations." + ], + "table_footnote": [], + "table_body": "
MethodsMetricsCleanGaussian NoiseJPEGPoisson
σ=1σ=5Q=70Q=80Q=90
MVSS-Net†[8]F10.6940.6440.6190.4580.5070.5580.652
BA(%)-------
EditGuard (Ours)F10.9660.9370.9320.9200.9200.9250.943
BA(%)99.9599.9499.3797.6998.1698.2399.91
", + "bbox": [ + 80, + 390, + 467, + 452 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The tamper types in this subset did not appear in the training set. As shown in Fig. 9, when faced with real-world tampering, even the most powerful tamper localization methods almost entirely fail. This is due to their mechanisms to look for image artifacts and explore instance-wise semantic information. However, our EditGuard, which locates tampered masks via the natural fragility and locality of I2I steganography, can still clearly annotate the tampered area.", + "bbox": [ + 75, + 457, + 467, + 578 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.5. Robustness Analysis", + "text_level": 1, + "bbox": [ + 76, + 587, + 269, + 603 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "As shown in Tab. 4, we conducted robustness analysis on the tampering of \"Stable Diffusion Inpaint\" [51] under Gaussian noise with $\\sigma = 1$ and 5, JPEG compression with $Q = 70$ , 80 and 90, and Poisson noise with $\\alpha = 4$ [71]. We observed that our method still maintains a high localization accuracy (F1-score $>0.9$ ) and bit accuracy with a very slight performance decrease under various levels of degradations, while MVSS-Net† [8] exhibits a noticeable performance degradation compared to its results in clean conditions. It is attributed to our prompt-based estimation that can effectively learn the degradation representation.", + "bbox": [ + 75, + 611, + 467, + 776 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.6. Ablation Study", + "text_level": 1, + "bbox": [ + 76, + 786, + 228, + 801 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To verify the effectiveness of each component of the Edit-Guard, we conducted ablation studies on bi-level optimization (BO), lightweight feature interaction module (LFIM), transformer block (TB), and prompt-based fusion (PF) under the tampering of \"Stable Diffusion Inpaint\". As listed in Tab. 5, without BO, the joint training of all components", + "bbox": [ + 75, + 809, + 467, + 898 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/cabce7a34e141a4bb1209a7fc25977c64e8bfb656c98cc12391f57bfa9a40835.jpg", + "table_caption": [ + "Table 5. Abalation studies on the core components of EditGuard." + ], + "table_footnote": [], + "table_body": "
CaseDegradation Type D(·)PFTBLFIMBOF1AUCIoUBA(%)
(a)Clean✓✓✓---49.17
(b)Clean✓✓✓0.9500.9600.90499.73
(c)Clean✓✓✓0.9570.9660.92799.51
(d)Random Degradations✓✓✓0.9030.9330.84199.12
OursClean✓✓✓✓0.9660.9710.93699.95
Random Degradations✓✓✓✓0.9380.9640.88799.36
", + "bbox": [ + 503, + 375, + 888, + 446 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "cannot converge effectively, resulting in bit accuracy that is close to random guessing. Without LFIM and TB, the IoU of EditGuard will suffer 0.032 and 0.009 declines since these two modules can better perform feature fusion. Without PF, the robustness of the EditGuard will significantly decline. We observed that the F1/AUC/IoU of our method far surpasses that of case (d) by 0.035/0.031/0.046 under \"Random Degradations\", which indicates that the PF effectively enables a single network to support watermark recovery under various degradations. \"Random Degradations\" denotes that we randomly set the $\\mathcal{D}(\\cdot)$ to various levels of Gaussian noise, Poisson noise, and JPEG compression.", + "bbox": [ + 496, + 453, + 890, + 633 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 648, + 617, + 662 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We present the first attempt to design a deep versatile watermarking mechanism EditGuard. It enhances the credibility of images by embedding imperceptible localization and copyright watermarks, and decoding accurate copyright information and tampered areas, making it a reliable tool for artistic creation and legal forensic analysis. In the future, we will focus on improving the robustness of EditGuard and strive not only to offer pixel-wise localization results but also to provide semantic-wise outcomes. Additionally, we plan to further expand EditGuard to a broader range of modalities and applications, including video, audio, and 3D scenes. Our efforts at information authenticity serve not only the AIGC industry, but the trust in our digital world, ensuring that every pixel tells the truth and the rights of each individual are safeguarded.", + "bbox": [ + 496, + 672, + 890, + 898 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "11971", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 174, + 106 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Mahdi Ahmadi, Alireza Norouzi, Nader Karimi, Shadrokh Samavi, and Ali Emami. Redmark: Framework for residual diffusion watermarking based on deep networks. Expert Systems with Applications, 146:113157, 2020. 3", + "[2] Vishal Asnani, Xi Yin, Tal Hassner, and Xiaoming Liu. Malp: Manipulation localization using a proactive scheme. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 2", + "[3] Shumeet Baluja. Hiding images in plain sight: Deep steganography. In Proceedings of the International Conference on Neural Information Processing Systems (NeurIPS), 2017. 3", + "[4] Liangyu Chen, Xiaojie Chu, Xiangyu Zhang, and Jian Sun. Simple baselines for image restoration. In Proceedings of the European Conference on Computer Vision (ECCV), 2022. 4", + "[5] Xinru Chen, Chengbo Dong, Jiaqi Ji, Juan Cao, and Xirong Li. Image manipulation detection by multi-view multi-scale supervision. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 2", + "[6] Baotian Cheng, Rongrong Ni, and Yao Zhao. A refining localization watermarking for image tamper detection and recovery. In 2012 IEEE 11th International Conference on Signal Processing, pages 984-988, 2012. 2", + "[7] Yingqian Cui, Jie Ren, Han Xu, Pengfei He, Hui Liu, Lichao Sun, and Jiliang Tang. Diffusionshield: A watermark for copyright protection against generative diffusion models. arXiv preprint arXiv:2306.04642, 2023. 3", + "[8] Chengbo Dong, Xinru Chen, Ruohan Hu, Juan Cao, and Xirong Li. Mvss-net: Multi-view multi-scale supervised networks for image manipulation detection. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(3):3539-3553, 2022. 2, 6, 7, 8", + "[9] Jing Dong, Wei Wang, and Tieniu Tan. Casia image tampering detection evaluation database. In Proceedings of the IEEE China Summit and International Conference on Signal and Information Processing (ChinaSIP), 2013. 6", + "[10] Han Fang, Zhaoyang Jia, Zehua Ma, Ee-Chien Chang, and Weiming Zhang. Pimog: An effective screen-shooting noise-layer simulation for deep-learning-based watermarking network. In Proceedings of the 30th ACM International Conference on Multimedia (MM), 2022. 3, 6, 7", + "[11] Han Fang, Yupeng Qiu, Kejiang Chen, Jiyi Zhang, Weiming Zhang, and Ee-Chien Chang. Flow-based robust watermarking with invertible noise layer for black-box distortions. In Proceedings of the AAAI Conference on Artificial Intelligence (AAAI), 2023. 3", + "[12] Pierre Fernandez, Guillaume Couairon, Hervé Jégou, Matthijs Douze, and Teddy Furon. The stable signature: Rooting watermarks in latent diffusion models. Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2023. 2, 3", + "[13] Anirban Ghoshal. Artists lose first copyright battle in the fight against ai-generated images. https://www.computerworld.com/article/3709691/artists-lose-first-copyright-battle-in" + ], + "bbox": [ + 78, + 114, + 470, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "the-fight-against-ai-generated-images. \nhtml,2023.2", + "[14] Gabriel Goh, James Betker, Li Jing, Aditya Ramesh, et al. Improving image generation with better captions. https://cdn.openai.com/papers/dall-e-3.pdf, 2023.1", + "[15] Google DeepMind. Synthid. https://deepmind.google/technologies/synthid/, 2023.2", + "[16] Haiying Guan, Mark Kozak, Eric Robertson, Yooyoung Lee, Amy N Yates, Andrew Delgado, Daniel Zhou, Timothee Kheyrkhah, Jeff Smith, and Jonathan Fiscus. Mfc datasets: Large-scale benchmark datasets for media forensic challenge evaluation. In Proceedings of the IEEE Winter Applications of Computer Vision Workshops (WACVW), 2019. 6", + "[17] Fabrizio Guillaro, Davide Cozzolino, Avneesh Sud, Nicholas Dufour, and Luisa Verdoliva. Trufor: Leveraging all-round clues for trustworthy image forgery detection and localization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 2, 6", + "[18] Xiao Guo, Xiaohong Liu, Zhiyuan Ren, Steven Grosz, Iacopo Masi, and Xiaoming Liu. Hierarchical fine-grained image forgery detection and localization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 2, 7", + "[19] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 5", + "[20] Yu-Feng Hsu and Shih-Fu Chang. Detecting image splicing using geometry invariants and camera characteristics consistency. In Proceedings of the IEEE International Conference on Multimedia and Expo (ICME), 2006. 6", + "[21] Jie Hu, Li Shen, and Gang Sun. Squeeze-and-excitation networks. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 6", + "[22] Xuefeng Hu, Zhihan Zhang, Zhenye Jiang, Syomantak Chaudhuri, Zhenheng Yang, and Ram Nevatia. Span: Spatial pyramid attention network for image manipulation localization. In Proceedings of the European Conference on Computer Vision (ECCV), 2020. 6", + "[23] Xiaoxiao Hu, Qichao Ying, Zhenxing Qian, Sheng Li, and Xinpeng Zhang. Draw: Defending camera-shooted raw against image manipulation. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2023. 2", + "[24] Nasir N Hurrah, Shabir A Parah, Nazir A Loan, Javaid A Sheikh, Mohammad Elhoseny, and Khan Muhammad. Dual watermarking framework for privacy protection and content authentication of multimedia. Future generation computer Systems, 94:654-673, 2019. 2", + "[25] Ashraful Islam, Chengjiang Long, Arslan Basharat, and Anthony Hoogs. Doa-gan: Dual-order attentive generative adversarial network for image copy-move forgery detection and localization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 2" + ], + "bbox": [ + 501, + 92, + 893, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "11972", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[26] Zhaoyang Jia, Han Fang, and Weiming Zhang. Mbrs: Enhancing robustness of dnn-based watermarking by minibatch of real and simulatedJPEG compression. In Proceedings of the 29th ACM International Conference on Multimedia (MM), 2021. 6, 7", + "[27] Zhengyuan Jiang, Jinghuai Zhang, and Neil Zhenqiang Gong. Evading watermark based detection of ai-generated content. arXiv preprint arXiv:2305.03807, 2023. 3", + "[28] Junpeng Jing, Xin Deng, Mai Xu, Jianyi Wang, and Zhenyu Guan. Hinet: Deep image hiding by invertible network. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 3", + "[29] Asra Kamili, Nasir N Hurrah, Shabir A Parah, Ghulam Mohiuddin Bhat, and Khan Muhammad. Dwfcat: Dual watermarking framework for industrial image authentication and tamper localization. IEEE Transactions on Industrial Informatics, 17(7):5108-5117, 2020. 2", + "[30] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6", + "[31] Myung-Joon Kwon, In-Jae Yu, Seung-Hun Nam, and Heung-Kyu Lee. Cat-net: Compression artifact tracing network for detection and localization of image splicing. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), 2021. 2", + "[32] Myung-Joon Kwon, Seung-Hun Nam, In-Jae Yu, Heung-Kyu Lee, and Changick Kim. LearningJPEG compression artifacts for image manipulation detection and localization. International Journal of Computer Vision, 130(8):1875-1895, 2022.6", + "[33] Haodong Li and Jiwu Huang. Localization of deep inpainting using high-pass fully convolutional network. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2019. 2", + "[34] Yuanman Li and Jiantao Zhou. Fast and effective image copy-move forgery detection via hierarchical feature point matching. IEEE Transactions on Information Forensics and Security, 14(5):1307-1322, 2018. 2", + "[35] Yue Li, Dong Liu, Houqiang Li, Li Li, Zhu Li, and Feng Wu. Learning a convolutional neural network for image compact-resolution. IEEE Transactions on Image Processing, 28(3): 1092-1107, 2018. 2", + "[36] Siau-Chuin Liew, Siau-Way Liew, and Jasni Mohd Zain. Tamper localization and lossless recovery watermarking scheme with roi segmentation and multilevel authentication. Journal of digital imaging, 26:316-325, 2013. 2", + "[37] Chia-Chen Lin, Ting-Lin Lee, Ya-Fen Chang, Pei-Feng Shiu, and Bohan Zhang. Fragile watermarking for tamper localization and self-recovery based on ambtc and vq. Electronics, 12(2):415, 2023. 2", + "[38] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Proceedings of the European Conference on Computer Vision (ECCV), 2014. 6, 7", + "[39] Xiaohong Liu, Yaojie Liu, Jun Chen, and Xiaoming Liu. Pssc-net: Progressive spatio-channel correlation network for" + ], + "bbox": [ + 78, + 90, + 468, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "image manipulation detection and localization. IEEE Transactions on Circuits and Systems for Video Technology, 32 (11):7505-7517, 2022. 6, 7", + "[40] Yang Liu, Mengxi Guo, Jian Zhang, Yuesheng Zhu, and Xiaodong Xie. A novel two-stage separable deep learning framework for practical blind watermarking. In Proceedings of the ACM International Conference on Multimedia (MM), 2019. 3", + "[41] Ziwei Liu, Ping Luo, Xiaogang Wang, and Xiaou Tang. Deep learning face attributes in the wild. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), 2015. 7", + "[42] Shao-Ping Lu, Rong Wang, Tao Zhong, and Paul L Rosin. Large-capacity image steganography based on invertible neural networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 3, 4", + "[43] Andreas Lugmayr, Martin Danelljan, Andres Romero, Fisher Yu, Radu Timofte, and Luc Van Gool. Repaint: Inpainting using denoising diffusion probabilistic models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 1, 7", + "[44] Rui Ma, Mengxi Guo, Yi Hou, Fan Yang, Yuan Li, Huizhu Jia, and Xiaodong Xie. Towards blind watermarking: Combining invertible and non-invertible mechanisms. In Proceedings of the ACM International Conference on Multimedia (MM), 2022. 2, 3, 6, 7", + "[45] Xiaochen Ma, Bo Du, Xianggen Liu, Ahmed Y Al Hammadi, and Jizhe Zhou. Iml-vit: Image manipulation localization by vision transformer. arXiv preprint arXiv:2307.14863, 2023. 2, 7", + "[46] Chong Mou, Youmin Xu, Jiechong Song, Chen Zhao, Bernard Ghanem, and Jian Zhang. Large-capacity and flexible video steganography via invertible neural network. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 3, 4", + "[47] Chong Mou, Xintao Wang, Liangbin Xie, Yanze Wu, Jian Zhang, Zhongang Qi, Ying Shan, and Xiaohu Qie. T2i-adapter: Learning adapters to dig out more controllable ability for text-to-image diffusion models. In AAAI Conference on Artificial Intelligence (AAAI), 2024. 1", + "[48] Neena Raj NR and R Shreelekshmi. Fragile watermarking scheme for tamper localization in images using logistic map and singular value decomposition. Journal of Visual Communication and Image Representation, 85:103500, 2022. 2", + "[49] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952, 2023. 1, 7", + "[50] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022. 2", + "[51] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of" + ], + "bbox": [ + 503, + 92, + 890, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "11973", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 1, 3, 7, 8", + "[52] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, Seyed Kamyar Seyed Ghasemipour, Raphael Gontijo-Lopes, Burcu Karagol Ayan, Tim Salimans, Jonathan Ho, David J. Fleet, and Mohammad Norouzi. Photorealistic text-to-image diffusion models with deep language understanding. In Proceedings of the International conference on Neural Information Processing Systems (NeurIPS), 2022. 1", + "[53] Ronald Salloum, Yuzhuo Ren, and C-C Jay Kuo. Image splicing localization using a multi-task fully convolutional network (mfcn). Journal of Visual Communication and Image Representation, 51:201-209, 2018. 2", + "[54] Deepa Shivaram. The white house and big tech companies release commitments on managing ai. https://www.nprillinois.org/2023-07-21/the-white-house-and-big-tech-companies-release-commitments-on-managing-ai, 2023.2", + "[55] Xiaopeng Sun, Weiqi Li, Zhenyu Zhang, Qiufang Ma, Xuhan Sheng, Ming Cheng, Haoyu Ma, Shijie Zhao, Jian Zhang, Junlin Li, et al. Opdn: Omnidirectional position-aware deformable network for omnidirectional image superresolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshop, pages 1293-1301, 2023. 1", + "[56] Zhihao Sun, Haoran Jiang, Danding Wang, Xirong Li, and Juan Cao. Safl-net: Semantic-agnostic feature learning network with auxiliary plugins for image manipulation detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2023. 2", + "[57] Roman Suvorov, Elizaveta Logacheva, Anton Mashikhin, Anastasia Remizova, Armenii Ashukha, Aleksei Silvestrov, Naejin Kong, Harshith Goka, Kiwoong Park, and Victor Lempitsky. Resolution-robust large mask inpainting with fourier convolutions. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), 2022. 7", + "[58] Bihan Wen, Ye Zhu, Ramanathan Subramanian, Tian-Tsong Ng, Xuanjing Shen, and Stefan Winkler. Coverage—a novel database for copy-move forgery detection. In Proceedings of the IEEE International Conference on Image Processing (ICIP), 2016. 6", + "[59] Yuxin Wen, John Kirchenbauer, Jonas Geiping, and Tom Goldstein. Tree-rings watermarks: Invisible fingerprints for diffusion images. Advances in Neural Information Processing Systems, 36, 2024. 3", + "[60] Huikai Wu, Shaocheng Xiang, Gabriben, and Niczem. Faceswap. https://github.com/wuhuikai/FaceSwap, 2020.7", + "[61] Haiwei Wu, Jiantao Zhou, Jinyu Tian, and Jun Liu. Robust image forgery detection over online social network shared images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2, 6, 7", + "[62] Xiaoshuai Wu, Xin Liao, and Bo Ou. Sepmark: Deep separable watermarking for unified source tracing and deepfake" + ], + "bbox": [ + 78, + 92, + 468, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "detection. In Proceedings of the ACM international conference on Multimedia (MM), 2023. 2, 3, 6, 7", + "[63] Yue Wu, Wael AbdAlmageed, and Premkumar Natarajan. *Mantra-net: Manipulation tracing network for detection and localization of image forgeries with anomalous features*. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 2, 6", + "[64] Youmin Xu, Chong Mou, Yujie Hu, Jingfen Xie, and Jian Zhang. Robust invertible image steganography. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 3, 4", + "[65] Qichao Ying, Zhenxing Qian, Hang Zhou, Haisheng Xu, Xinpeng Zhang, and Siyi Li. From image to image: Immunized image generation. In Proceedings of the ACM international conference on Multimedia (MM), 2021. 2", + "[66] Qichao Ying, Xiaoxiao Hu, Xiangyu Zhang, Zhenxing Qian, Sheng Li, and Xinpeng Zhang. Rwn: Robust watermarking network for image cropping localization. In Proceedings of the IEEE International Conference on Image Processing (ICIP), 2022.", + "[67] Qichao Ying, Hang Zhou, Zhenxing Qian, Sheng Li, and Xinpeng Zhang. Learning to immunize images for tamper localization and self-recovery. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023. 2", + "[68] Jiwen Yu, Yinhuai Wang, Chen Zhao, Bernard Ghanem, and Jian Zhang. Freedom: Training-free energy-guided conditional diffusion model. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2023. 1", + "[69] Jiwen Yu, Xuanyu Zhang, Youmin Xu, and Jian Zhang. Cross: Diffusion model makes controllable, robust and secure image steganography. Advances in Neural Information Processing Systems, 36, 2024. 3", + "[70] Syed Waqas Zamir, Aditya Arora, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, and Ming-Hsuan Yang. Restormer: Efficient transformer for high-resolution image restoration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 5", + "[71] Kai Zhang, Yawei Li, Jingyun Liang, Jiezhang Cao, Yulun Zhang, Hao Tang, Radu Timofte, and Luc Van Gool. Practical blind denoising via swim-conv-unet and data synthesis. arXiv preprint arXiv:2203.13278, 2022. 8", + "[72] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2023. 1, 7", + "[73] Yunqing Zhao, Tianyu Pang, Chao Du, Xiao Yang, Ngai-Man Cheung, and Min Lin. A recipe for watermarking diffusion models. arXiv preprint arXiv:2303.10137, 2023. 3", + "[74] Jiren Zhu, Russell Kaplan, Justin Johnson, and Li Fei-Fei. Hidden: Hiding data with deep networks. In European Conference on Computer Vision (ECCV), 2018. 3", + "[75] Xinshan Zhu, Yongjun Qian, Xianfeng Zhao, Biao Sun, and Ya Sun. A deep learning approach to patch-based image inpainting forensics. Signal Processing: Image Communication, 67:90-99, 2018. 2" + ], + "bbox": [ + 501, + 92, + 890, + 882 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11974", + "bbox": [ + 480, + 945, + 519, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2024/EditGuard_ Versatile Image Watermarking for Tamper Localization and Copyright Protection/43cb9c2b-dd1a-4670-9150-d4345483687e_model.json b/2024/EditGuard_ Versatile Image Watermarking for Tamper Localization and Copyright Protection/43cb9c2b-dd1a-4670-9150-d4345483687e_model.json new file mode 100644 index 0000000000000000000000000000000000000000..ae23d368b886f19faf1dd851ad52d0dea5e0ee92 --- /dev/null +++ b/2024/EditGuard_ Versatile Image Watermarking for Tamper Localization and Copyright Protection/43cb9c2b-dd1a-4670-9150-d4345483687e_model.json @@ -0,0 +1,3412 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.0, + 0.812, + 0.045 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.12, + 0.131, + 0.852, + 0.176 + ], + "angle": 0, + "content": "EditGuard: Versatile Image Watermarking for Tamper Localization and Copyright Protection" + }, + { + "type": "text", + "bbox": [ + 0.166, + 0.19, + 0.803, + 0.227 + ], + "angle": 0, + "content": "Xuanyu Zhang\\(^{1,2}\\), Runyi Li\\(^{1}\\), Jiwen Yu\\(^{1}\\), Youmin Xu\\(^{1}\\), Weiqi Li\\(^{1}\\), Jian Zhang\\(^{1,2 \\boxtimes}\\) \n\\(^{1}\\) School of Electronic and Computer Engineering, Peking University" + }, + { + "type": "text", + "bbox": [ + 0.121, + 0.227, + 0.843, + 0.246 + ], + "angle": 0, + "content": "\\(^{2}\\) Peking University Shenzhen Graduate School-Rabbitpre AIGC Joint Research Laboratory" + }, + { + "type": "image", + "bbox": [ + 0.088, + 0.258, + 0.131, + 0.293 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.093, + 0.298, + 0.129, + 0.308 + ], + "angle": 0, + "content": "Users" + }, + { + "type": "text", + "bbox": [ + 0.149, + 0.263, + 0.291, + 0.296 + ], + "angle": 0, + "content": "Please help protect my image against intentional tampering. My exclusive copyright information is 00111010110..." + }, + { + "type": "image", + "bbox": [ + 0.144, + 0.309, + 0.191, + 0.347 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.197, + 0.311, + 0.343, + 0.345 + ], + "angle": 0, + "content": "I've added the exclusive EditGuard watermark to it. You can post the picture on the left and I will be responsible for its protection." + }, + { + "type": "image", + "bbox": [ + 0.293, + 0.259, + 0.343, + 0.298 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.363, + 0.306, + 0.414, + 0.344 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.083, + 0.353, + 0.139, + 0.397 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.147, + 0.356, + 0.346, + 0.402 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.089, + 0.415, + 0.131, + 0.459 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.149, + 0.417, + 0.292, + 0.451 + ], + "angle": 0, + "content": "I find a picture online that looks similar to mine. Can you provide evidence to prove my copyright and pinpoint tampered areas?" + }, + { + "type": "image", + "bbox": [ + 0.293, + 0.415, + 0.343, + 0.453 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.363, + 0.461, + 0.414, + 0.497 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.45, + 0.257, + 0.481, + 0.265 + ], + "angle": 0, + "content": "Original" + }, + { + "type": "image", + "bbox": [ + 0.43, + 0.265, + 0.499, + 0.318 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.522, + 0.257, + 0.56, + 0.265 + ], + "angle": 0, + "content": "Tampered" + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.265, + 0.575, + 0.319 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.582, + 0.257, + 0.654, + 0.265 + ], + "angle": 0, + "content": "Our Predicted mask" + }, + { + "type": "image", + "bbox": [ + 0.582, + 0.265, + 0.653, + 0.319 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.321, + 0.578, + 0.327 + ], + "angle": 0, + "content": "Remove the wom" + }, + { + "type": "image", + "bbox": [ + 0.43, + 0.33, + 0.499, + 0.385 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.501, + 0.33, + 0.574, + 0.385 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.581, + 0.33, + 0.652, + 0.385 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.435, + 0.387, + 0.512, + 0.395 + ], + "angle": 0, + "content": "Change the cloth" + }, + { + "type": "image", + "bbox": [ + 0.43, + 0.395, + 0.499, + 0.449 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.501, + 0.395, + 0.574, + 0.45 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.581, + 0.395, + 0.652, + 0.45 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.43, + 0.451, + 0.591, + 0.46 + ], + "angle": 0, + "content": "Tamper with AI-generated images" + }, + { + "type": "image_caption", + "bbox": [ + 0.581, + 0.451, + 0.65, + 0.46 + ], + "angle": 0, + "content": "es, \"A volcano'" + }, + { + "type": "image", + "bbox": [ + 0.43, + 0.46, + 0.499, + 0.492 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.501, + 0.46, + 0.574, + 0.514 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.581, + 0.46, + 0.652, + 0.515 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.508, + 0.515, + 0.573, + 0.526 + ], + "angle": 0, + "content": "Text Forgery" + }, + { + "type": "image_caption", + "bbox": [ + 0.686, + 0.257, + 0.715, + 0.264 + ], + "angle": 0, + "content": "Original" + }, + { + "type": "image", + "bbox": [ + 0.664, + 0.265, + 0.735, + 0.319 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.685, + 0.32, + 0.714, + 0.327 + ], + "angle": 0, + "content": "AI Fill" + }, + { + "type": "image_caption", + "bbox": [ + 0.757, + 0.257, + 0.795, + 0.265 + ], + "angle": 0, + "content": "Tampered" + }, + { + "type": "image", + "bbox": [ + 0.738, + 0.265, + 0.812, + 0.319 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.737, + 0.32, + 0.812, + 0.327 + ], + "angle": 0, + "content": "Picasso painting" + }, + { + "type": "image_caption", + "bbox": [ + 0.814, + 0.257, + 0.886, + 0.264 + ], + "angle": 0, + "content": "Our Predicted mask" + }, + { + "type": "image", + "bbox": [ + 0.816, + 0.265, + 0.887, + 0.319 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.816, + 0.32, + 0.865, + 0.327 + ], + "angle": 0, + "content": "on the wall" + }, + { + "type": "image", + "bbox": [ + 0.664, + 0.33, + 0.735, + 0.385 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.676, + 0.387, + 0.768, + 0.395 + ], + "angle": 0, + "content": "Replace background" + }, + { + "type": "image", + "bbox": [ + 0.664, + 0.395, + 0.735, + 0.45 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.664, + 0.451, + 0.811, + 0.46 + ], + "angle": 0, + "content": "Secondary creation of artworks" + }, + { + "type": "image", + "bbox": [ + 0.738, + 0.33, + 0.811, + 0.385 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.739, + 0.387, + 0.811, + 0.395 + ], + "angle": 0, + "content": "ground, \"A man" + }, + { + "type": "image", + "bbox": [ + 0.816, + 0.33, + 0.887, + 0.385 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.816, + 0.387, + 0.865, + 0.395 + ], + "angle": 0, + "content": "in an office" + }, + { + "type": "image", + "bbox": [ + 0.816, + 0.395, + 0.887, + 0.45 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.816, + 0.451, + 0.882, + 0.459 + ], + "angle": 0, + "content": "\"A straw hat\"" + }, + { + "type": "image", + "bbox": [ + 0.664, + 0.46, + 0.735, + 0.515 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.74, + 0.46, + 0.811, + 0.515 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.749, + 0.516, + 0.8, + 0.525 + ], + "angle": 0, + "content": "Swap Face" + }, + { + "type": "image", + "bbox": [ + 0.816, + 0.46, + 0.887, + 0.515 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.075, + 0.53, + 0.894, + 0.587 + ], + "angle": 0, + "content": "Figure 1. We propose a versatile proactive forensics framework EditGuard. The application scenario is shown on the left, wherein users embed invisible watermarks to their images via EditGuard in advance. If suffering tampering, users can defend their rights via the tampered areas and copyright information provided by EditGuard. Some supported tampering methods (marked in blue) and localization results of EditGuard are placed on the right. Our EditGuard can achieve over \\(95\\%\\) localization precision and nearly \\(100\\%\\) copyright accuracy." + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.601, + 0.313, + 0.617 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.634, + 0.471, + 0.862 + ], + "angle": 0, + "content": "In the era of AI-generated content (AIGC), malicious tampering poses imminent threats to copyright integrity and information security. Current deep image watermarking, while widely accepted for safeguarding visual content, can only protect copyright and ensure traceability. They fall short in localizing increasingly realistic image tampering, potentially leading to trust crises, privacy violations, and legal disputes. To solve this challenge, we propose an innovative proactive forensics framework EditGuard, to unify copyright protection and tamper-agnostic localization, especially for AIGC-based editing methods. It can offer a meticulous embedding of imperceptible watermarks and precise decoding of tampered areas and copyright information. Leveraging our observed fragility and locality of image-into-image steganography, the realization of Edit" + }, + { + "type": "page_footnote", + "bbox": [ + 0.076, + 0.875, + 0.471, + 0.901 + ], + "angle": 0, + "content": "This work was supported by National Natural Science Foundation of China under Grant 62372016. (Corresponding author: Jian Zhang.)" + }, + { + "type": "text", + "bbox": [ + 0.497, + 0.602, + 0.893, + 0.709 + ], + "angle": 0, + "content": "Guard can be converted into a united image-bit steganography issue, thus completely decoupling the training process from the tampering types. Extensive experiments verify that our EditGuard balances the tamper localization accuracy, copyright recovery precision, and generalizability to various AIGC-based tampering methods, especially for image forgery that is difficult for the naked eye to detect." + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.723, + 0.633, + 0.739 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.497, + 0.75, + 0.893, + 0.901 + ], + "angle": 0, + "content": "Owing to the advantageous properties of diffusion models and the bolstering of extensive datasets, AI-generated content (AIGC) models like DALL-E 3 [14], Imagen [52], and Stable Diffusion [51], can produce lifelike and wondrous images, which brings great convenience to photography enthusiasts and image editors. Nonetheless, the remarkable capabilities of these models come with a double-edged sword, presenting new challenges in copyright protection and information security. The efficiency of image manipulation [43, 47, 49, 51, 55, 68, 72] has blurred the" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "11964" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.198 + ], + "angle": 0, + "content": "line between fact and forgery, ushering in myriad security and legal concerns. For instance, artistic works are vulnerable to malicious tampering or unauthorized AI-facilitated recreations, making it challenging to protect their original creations [13]. Meanwhile, forged images may be spread online or used as court evidence, causing adverse effects on public opinion, ethical issues, and social stability." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.199, + 0.471, + 0.517 + ], + "angle": 0, + "content": "Given the challenges of preventing image tampering from the source, image watermarking has become a consensus for proactive forensics [54]. However, most prevalent forensic image watermarking [12, 15, 44, 62] still focus on detecting image authenticity or protecting image copyrights, but fall short when it comes to advanced demands, such as localizing tampered areas. Tamper localization facilitates an evaluation of the severity of the image manipulation, and provides an understanding of the intent of attackers, potentially allowing for the partial reuse of the tampered images. However, most passive forensics methods such as previous black-box localization networks [8, 56, 61] tend to seek anomalies like artifacts or flickers in images but struggle to detect more realistic textures and more advanced AIGC models. Moreover, they inevitably need to introduce tampered data during the training and focus solely on specific \"CheapFake\" tampering like slicing and copy-and-paste [8, 45, 56], or on \"DeepFake\" targeting human faces [2, 50], restricted in generalizability. Thus, it is vital to develop an integrated watermarking framework that unites tamper-agnostic localization and copyright protection." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.518, + 0.47, + 0.685 + ], + "angle": 0, + "content": "To clarify our task scope, we re-emphasize the definition of dual forensics tasks as illustrated in Fig. 1: (1) Copyright protection: \"Who does this image belong to?\" We aspire to accurately retrieve the original copyright of an image, even suffering various tampering and degradation. (2) Tamper localization: \"Where was this image manipulated?\" We aim to precisely pinpoint the tampered areas, unrestrained by specific tampering types. To the best of our knowledge, no existing method accomplishes these two tasks simultaneously, while maintaining a balance of high precision and extensive generalizability." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.686, + 0.47, + 0.866 + ], + "angle": 0, + "content": "To address this urgent demand, we propose a novel proactive forensics framework, dubbed EditGuard, to protect copyrights and localize tamper areas for AIGC-based editing methods. Specifically, drawing inspiration from our observed locality and fragility of image-into-image (I2I) steganography and inherent robustness of bit-into-image steganography, we can transform the realization of EditGuard into a joint image-bit steganography issue, which allows the training of EditGuard to be entirely decoupled from tampering types, thereby endowing it with exceptional generalizability and locate tampering in a zero-shot manner. In a nutshell, our contributions are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.47, + 0.901 + ], + "angle": 0, + "content": "(1) We present the first attempt to design a deep versatile proactive forensics framework EditGuard for univer" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.137 + ], + "angle": 0, + "content": "sal tamper localization and copyright protection. It embeds dual invisible watermarks into original images and accurately decodes tampered areas and copyright information." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.139, + 0.892, + 0.215 + ], + "angle": 0, + "content": "(2) We have observed the fragility and locality of I2I steganography and innovatively convert the solution of this dual forensics task into training a united Image-Bit Steganography Network (IBSN), and utilize the core components of IBSN to construct EditGuard." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.217, + 0.892, + 0.263 + ], + "angle": 0, + "content": "(3) We introduce a prompt-based posterior estimation module to enhance the localization accuracy and degradation robustness of the proposed framework." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.265, + 0.892, + 0.356 + ], + "angle": 0, + "content": "(4) The effectiveness of our method has been verified on our constructed dataset and classical benchmarks. Compared to other competitive methods, our approach has notable merits in localization precision, generalization abilities, and copyright accuracy without any labeled data or additional training required for specific tampering types." + }, + { + "type": "list", + "bbox": [ + 0.498, + 0.139, + 0.892, + 0.356 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.369, + 0.645, + 0.384 + ], + "angle": 0, + "content": "2. Related works" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.393, + 0.696, + 0.409 + ], + "angle": 0, + "content": "2.1. Tamper Localization" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.416, + 0.892, + 0.765 + ], + "angle": 0, + "content": "Prevalent passive image forensic techniques have focused on localizing specific types of manipulations [25, 33, 34, 53, 61, 75]. Meanwhile, some universal tamper localization methods [5, 23, 31, 35, 63, 65-67] also tend to explore artifacts and anomalies in tampered images. For instance, MVSS-Net [8] employed multi-view feature learning and multi-scale supervision to jointly exploit boundary artifacts and the noise view of images. OSN [61] proposed a novel robust training scheme to address the challenges posed by lossy operations. Trufor [17] used a learned noise-sensitive fingerprint and extracted both high-level and low-level traces via transformer-based fusion. HiFi-Net [18] utilized multi-branch feature extractor and localization modules for both CNN-synthesized and edited images. SAFL-Net [56] constrained a feature extractor to learn semantic-agnostic features with specific modules and auxiliary tasks. However, the above-mentioned passive localization methods are often limited in terms of generalization and localization accuracy, which usually work on known tampering types that have been trained. Although MaLP [2] used template matching for proactive tamper localization, it still requires a large number of forgery images and cannot fully decouple the network training from the tamper types." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.772, + 0.704, + 0.789 + ], + "angle": 0, + "content": "2.2. Image Watermarking" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.795, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Image watermarking can be broadly used for the verification, authenticity, and traceability of images. Although traditional fragile watermarking [6, 24, 29, 36, 37, 48] can also achieve block-wise tamper localization, their localization accuracy and flexibility are unsatisfactory. How to realize joint pixel-level tamper localization and copyright protection remains largely unexplored. Owing to the development" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "11965" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.081, + 0.09, + 0.47, + 0.207 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.213, + 0.47, + 0.304 + ], + "angle": 0, + "content": "Figure 2. Fragility and locality of I2I steganography. The first line shows that when \\(\\mathbf{X}_{con}^{\\prime}\\) is changed, \\(\\hat{\\mathbf{X}}_{sec}\\) will also be fragility demaged. The second line plots the attribution maps \\(\\frac{1}{|\\mathbb{S}|}\\sum_{(i,j)\\in \\mathbb{S}}\\frac{\\partial\\hat{\\mathbf{X}}_{sec}[i,j]}{\\partial\\mathbf{X}_{con}^{\\prime}}\\) of five point sets \\(\\mathbb{S}\\) (marked by \\(①-⑤\\)) in \\(\\hat{\\mathbf{X}}_{sec}\\). We observed that \\(\\hat{\\mathbf{X}}_{sec}\\) almost only has a strong response at the corresponding positions of \\(\\mathbf{X}_{con}^{\\prime}\\) and its neighborhoods." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.313, + 0.471, + 0.51 + ], + "angle": 0, + "content": "of deep learning, deep image watermarking has attracted increased attention. For instance, HiDDeN [74] firstly introduced a deep encoder-decoder network to hide and recover bitstream. Moreover, many distortion layers such as differentiable JPEG and screen-shooting [1, 10, 40, 62] were designed to enhance its robustness. Flow-based models [11, 44] were also utilized to further improve the fidelity of container images. Recently, researchers [7, 12, 27, 59, 73] have designed specialized watermarking mechanisms for large-scale image generation models [51], to merge watermarking into the generation process. However, these deep watermarking methods have a singular function and cannot accurately localize the tampered areas." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.526, + 0.386, + 0.543 + ], + "angle": 0, + "content": "3. Overall Framework of EditGuard" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.552, + 0.199, + 0.567 + ], + "angle": 0, + "content": "3.1. Motivation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.577, + 0.47, + 0.698 + ], + "angle": 0, + "content": "Challenges of existing methods: (1) How to equip existing watermark methods, which are solely for copyright protection, with the ability to localize tampering is the crux of EditGuard. We will solve it via the framework design in Sec. 3.2. (2) Most previous tamper localization methods inevitably introduce specific tampering data during network training but tend to raise generalization concerns in unknown tampering types, which will be addressed in Sec. 3.3." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.705, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Our observations: Fortunately, we observed that imageinto-image (I2I) steganography exhibits distinct fragility and locality, possessing great potential to address these issues. Concretly, I2I steganography [3, 28, 42, 46, 64, 69] aims to hide a secret image \\(\\mathbf{X}_{\\text{sec}}\\) into a cover image \\(\\mathbf{X}_{\\text{cov}}\\) to produce a container image \\(\\mathbf{X}_{\\text{con}}\\), and reveal \\(\\hat{\\mathbf{X}}_{\\text{sec}}\\) and \\(\\hat{\\mathbf{X}}_{\\text{cov}}\\) with minimal distortion from the received image \\(\\mathbf{X}_{\\text{con}}^{\\prime}\\). We discover that when \\(\\mathbf{X}_{\\text{con}}^{\\prime}\\) undergoes significant alteration compared to \\(\\mathbf{X}_{\\text{con}}\\), \\(\\hat{\\mathbf{X}}_{\\text{sec}}\\) will also be damaged and generate artifacts (the first row of Fig. 2), which is called fragility. Furthermore, we notice that the artifacts in \\(\\hat{\\mathbf{X}}_{\\text{sec}}\\) are almost pixel-level corresponding to the changes in \\(\\mathbf{X}_{\\text{con}}^{\\prime}\\) relative to \\(\\mathbf{X}_{\\text{con}}\\), which is called locality. To demonstrate this locality," + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.091, + 0.892, + 0.198 + ], + "angle": 0, + "content": "we select five \\(7 \\times 7\\) point sets on \\(\\hat{\\mathbf{X}}_{\\text {sec }}\\) and calculated their attribution maps with respect to \\(\\mathbf{X}_{\\text {con }}^{\\prime}\\). As plotted in the second row of Fig. 2, \\(\\hat{\\mathbf{X}}_{\\text {sec }}\\) only exhibits a strong response at the corresponding locations of \\(\\mathbf{X}_{\\text {con }}^{\\prime}\\) and their immediate vicinity, almost irrelevant to other pixels. These properties inspire us to treat \\(\\mathbf{X}_{\\text {sec }}\\) as a special localization watermark and embed it within existing watermarking frameworks." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.21, + 0.855, + 0.225 + ], + "angle": 0, + "content": "3.2. Framework Design and Forensics Process" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.233, + 0.892, + 0.353 + ], + "angle": 0, + "content": "To realize united tamper localization and copyright protection, EditGuard is envisioned to embed both a 2D localization watermark and a 1D copyright traceability watermark into the original image in an imperceptible manner, which allows the decoding end to obtain the copyright of the images and a binary mask reflecting tampered areas. However, designing such a framework needs to solve the compatibility issue of two types of watermarks." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.355, + 0.892, + 0.521 + ], + "angle": 0, + "content": "(1) Local vs. Global: The localization watermark is required to be hidden in the corresponding pixel positions of the original image, while the copyright watermark should be unrelated to spatial location and embedded in the global area redundantly. (2) Semi-fragile vs. Robust: The desired attribute of the localization watermark is semi-fragile, which means it is fragile to tampering but robust against some common degradations (such as Gaussian noise, JPEG compression, and Poisson noise) during network transmission. However, the copyright should be extracted nearly losslessly, irrespective of tampering or degradation." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.522, + 0.892, + 0.734 + ], + "angle": 0, + "content": "To address the two pivotal conflicts, EditGuard employs a \"sequential encoding and parallel decoding\" structure, which comprises a dual-watermark encoder, a tamper locator, and a copyright extractor. As shown in Fig. 3, the dual-watermark encoder will sequentially add pre-defined localization watermark and global copyright watermark \\(\\mathbf{w}_{cop}\\) provided by users to the original image \\(\\mathbf{I}_{ori}\\), forming the container image \\(\\mathbf{I}_{con}\\). Our experiments have proved that parallel encoding cannot effectively add dual watermarks into images (in supplementary materials (S.M.). In contrast, sequential embedding effectively prevents cross-interference by hiding these two watermarks. Furthermore, we model the network transmission process in which the received (tampered) image \\(\\mathbf{I}_{rec}\\) is transformed from \\(\\mathbf{I}_{con}\\) as:" + }, + { + "type": "equation", + "bbox": [ + 0.543, + 0.74, + 0.891, + 0.756 + ], + "angle": 0, + "content": "\\[\n\\mathbf {I} _ {r e c} = \\mathcal {D} \\left(\\mathbf {I} _ {c o n} \\odot (\\mathbf {1} - \\mathbf {M}) + \\mathcal {T} \\left(\\mathbf {I} _ {c o n}\\right) \\odot \\mathbf {M}\\right), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.762, + 0.892, + 0.883 + ], + "angle": 0, + "content": "where \\(\\mathcal{T}(\\cdot),\\mathcal{D}(\\cdot)\\) and \\(\\mathbf{M}\\) respectively denote the tamper function, degradation operation, and tempered mask. Moreover, the parallel decoding processes allow us to flexibly train each branch under different levels of robustness and obtain the predicted mask \\(\\hat{\\mathbf{M}}\\) via the tamper locator and traceability watermark \\(\\hat{\\mathbf{w}}_{cop}\\) via the copyright extractor. We can categorize the dual forensic process of EditGuard into the following scenarios." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.886, + 0.892, + 0.902 + ], + "angle": 0, + "content": "Case 1: If \\(\\hat{\\mathbf{w}}_{\\text{cop}} \\not\\simeq \\mathbf{w}_{\\text{cop}}\\), suspicious \\(\\mathbf{I}_{\\text{rec}}\\) is either not" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "11966" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.089, + 0.895, + 0.255 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.261, + 0.895, + 0.318 + ], + "angle": 0, + "content": "Figure 3. Illustration of the proposed proactive forensics framework EditGuard. The dual-watermark encoder sequentially embeds the pre-defined localization watermark and copyright watermark \\(\\mathbf{w}_{\\text{cop}}\\) into the original image \\(\\mathbf{I}_{\\text{ori}}\\), generating the container image \\(\\mathbf{I}_{\\text{con}}\\). After encountering potential malicious tampering and degradation during network transmission, tampered mask \\(\\hat{\\mathbf{M}}\\) and copyright information \\(\\hat{\\mathbf{W}}_{\\text{cop}}\\) are respectively extracted via the tamper locator and copyright extractor from the received image \\(\\mathbf{I}_{\\text{rec}}\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.325, + 0.469, + 0.355 + ], + "angle": 0, + "content": "registered in our EditGuard or underwent extremely severe global tampering, rendering it unreliable as evidence." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.356, + 0.469, + 0.417 + ], + "angle": 0, + "content": "- Case 2: If \\(\\hat{\\mathbf{w}}_{\\text {cop }} \\approx \\mathbf{w}_{\\text {cop }}\\) and \\(\\hat{\\mathbf{M}} \\not\\approx \\mathbf{0}\\), suspicious \\(\\mathbf{I}_{\\text {rec }}\\) has undergone tampering, disqualifying it as valid evidence. Users may infer the intention of tamperers based on \\(\\hat{\\mathbf{M}}\\) and decide whether to reuse parts of the image." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.418, + 0.468, + 0.448 + ], + "angle": 0, + "content": "\\(\\square\\) Case 3: If \\(\\hat{\\mathbf{w}}_{cop} \\approx \\mathbf{w}_{cop}\\) and \\(\\hat{\\mathbf{M}} \\approx \\mathbf{0}\\), \\(\\mathbf{I}_{rec}\\) remains untampered and trustworthy under the shield of EditGuard." + }, + { + "type": "list", + "bbox": [ + 0.076, + 0.356, + 0.469, + 0.448 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.456, + 0.468, + 0.472 + ], + "angle": 0, + "content": "3.3. Transform Dual Forensics into Steganography" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.479, + 0.473, + 0.736 + ], + "angle": 0, + "content": "To realize universal and tamper-agnostic localization, we resort to our observed locality and fragility of I2I steganography. As described in Sec. 3.1, localization watermarking and tamper locator in Fig. 3 can be effectively realized via image hiding and revealing. Meanwhile, combining with the robustness of current bit-into-image steganography, copyright watermarking and extractor in Fig. 3 are achieved via bit encryption and recovery. Thus, we can convert the realization of the dual forensics framework EditGuard into a united image-bit steganography network. Our training objective is just a self-recovery mechanism, meaning it only needs to ensure the input and output of the steganography network maintain high fidelity under various robustness levels, with no need to introduce any labeled data or tampered samples. During inference, it can naturally locate tampering via simple comparisons in a zero-shot manner and extract copyright exactly." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.747, + 0.454, + 0.765 + ], + "angle": 0, + "content": "4. United Image-bit Steganography Network" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.772, + 0.283, + 0.787 + ], + "angle": 0, + "content": "4.1. Network Architecture" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.795, + 0.47, + 0.901 + ], + "angle": 0, + "content": "As plotted in Fig. 4, the proposed IBSN includes an image hiding module (IHM), a bit encryption module (BEM), a bit recovery module (BRM), and an image revealing module (IRM). First, the IHM aims to hide a localization watermark \\(\\mathbf{W}_{loc} \\in \\mathbb{R}^{H \\times W \\times 3}\\) into the original image \\(\\mathbf{I}_{ori} \\in \\mathbb{R}^{H \\times W \\times 3}\\), resulting in an intermediate output \\(\\mathbf{I}_{med} \\in \\mathbb{R}^{H \\times W \\times 3}\\). Subsequently, \\(\\mathbf{I}_{med}\\) is fed to the BEM for feature refinement," + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.325, + 0.892, + 0.447 + ], + "angle": 0, + "content": "while the copyright watermark \\(\\mathbf{w}_{cop} \\in \\{0,1\\}^L\\) is modulated into the BEM, forming the final container image \\(\\mathbf{I}_{con} \\in \\mathbb{R}^{H \\times W \\times 3}\\). After network transmission, the BRM will faithfully reconstruct the copyright watermark \\(\\hat{\\mathbf{w}}_{cop}\\) from the received container image \\(\\mathbf{I}_{rec}\\). Meanwhile, \\(\\mathbf{I}_{rec}\\) predicts the missing information \\(\\tilde{\\mathbf{Z}}\\) via the prompt-based posterior estimation and uses it as the initialization for the invertible blocks, producing \\(\\hat{\\mathbf{I}}_{ori}\\) and semi-fragile watermark \\(\\hat{\\mathbf{W}}_{loc}\\)." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.453, + 0.804, + 0.468 + ], + "angle": 0, + "content": "4.2. Invertible Blocks in IHM and IRM" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.477, + 0.892, + 0.734 + ], + "angle": 0, + "content": "Given the inherent capacity of flow-based models to precisely recover multimedia information, we harness stacked invertible blocks to construct image hiding and revealing modules. The original image \\(\\mathbf{I}_{ori} \\in \\mathbb{R}^{H \\times W \\times 3}\\) and localization watermark \\(\\mathbf{W}_{loc} \\in \\mathbb{R}^{H \\times W \\times 3}\\) will undergo discrete wavelet transformations (DWT) to yield frequency-decoupled image features. We then employ enhanced additive affine coupling layers to project the original image and its corresponding localization watermark branches. The transformation parameters are generated from each other. The enhanced affine coupling layer is composed of a five-layer dense convolution block [46] and a lightweight feature interaction module (LFIM) [4]. The LFIM can enhance the non-linearity of transformations and capture the long-range dependencies with low computational cost. More details are presented in S.M. Finally, the revealed features are then transformed to the image domain via the inverse DWT." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.742, + 0.808, + 0.758 + ], + "angle": 0, + "content": "4.3. Prompt-based Posterior Estimation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.892, + 0.901 + ], + "angle": 0, + "content": "To bolster the fidelity and robustness of the image hiding and revealing module, we introduce a degradation prompt-based posteriori estimation module (PPEM). Since the encoding network tends to compress \\(\\left[\\mathbf{I}_{ori};\\mathbf{W}_{loc}\\right] \\in \\mathbb{R}^{H \\times W \\times 6}\\) into the container image \\(\\mathbf{I}_{con} \\in \\mathbb{R}^{H \\times W \\times 3}\\), previous methods [42, 64] typically utilized a random Gaussian initialization or an all-zero map at the decoding end to compensate for the lost high-frequency channels. Yet, our observations suggest that discarded information lurks within the" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "11967" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.084, + 0.09, + 0.891, + 0.272 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.084, + 0.276, + 0.888, + 0.433 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.438, + 0.893, + 0.493 + ], + "angle": 0, + "content": "Figure 4. Illustration of the united Image-bit Steganography Network (IBSN). In the training process, we randomly sample original image \\(\\mathbf{I}_{ori}\\), localization watermark \\(\\mathbf{W}_{loc}\\) (a natural RGB image) and copyright watermark \\(\\mathbf{w}_{cop}\\) and expect the IBSN to recover \\(\\hat{\\mathbf{I}}_{ori}\\), \\(\\hat{\\mathbf{W}}_{loc}\\) and \\(\\hat{\\mathbf{w}}_{cop}\\) with high fidelity. In the inference process, we use core components of the pre-trained IBSN with a mask extractor (ME) to construct our EditGuard, and pre-define a simple solid color image as a localization watermark." + }, + { + "type": "image", + "bbox": [ + 0.078, + 0.499, + 0.47, + 0.648 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.652, + 0.47, + 0.694 + ], + "angle": 0, + "content": "Figure 5. Illustration of the proposed prompt-based posterior estimation. It will dynamically fuse degraded representations and extracted features to obtain posterior mean \\(\\hat{\\mathbf{Z}} = \\mathbb{E}[\\mathbf{Z}|\\mathbf{I}_{rec}]\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.697, + 0.469, + 0.801 + ], + "angle": 0, + "content": "edges and textures of the container image. Thus, deploying a dedicated network proves to be a more potent strategy in predicting the posterior mean of the vanished localization watermark information \\(\\hat{\\mathbf{Z}} = \\mathbb{E}[\\mathbf{Z}|\\mathbf{I}_{rec}]\\). Specifically, as shown in Fig. 5, we stack \\(M\\) residual blocks \\(\\mathrm{Res}(\\cdot)\\) [19] and \\(M\\) channel-wise transformer blocks \\(\\mathrm{Trans}(\\cdot)\\) [70] to extract the local and non-local features \\(\\mathbf{F}_c\\)." + }, + { + "type": "equation", + "bbox": [ + 0.085, + 0.805, + 0.469, + 0.822 + ], + "angle": 0, + "content": "\\[\n\\mathbf {F} _ {c} = \\operatorname {T r a n s} (\\operatorname {R e s} (\\mathrm {D W T} (\\mathbf {I} _ {r e c}))) + \\operatorname {R e s} (\\mathrm {D W T} (\\mathbf {I} _ {r e c})). \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.826, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Considering that the container image is prone to various degradations during network transmission, we pre-define \\(N\\) learnable embedding tensors as degradation prompts \\(\\mathbf{P} = [\\mathbf{P}_1,\\mathbf{P}_2,\\dots ,\\mathbf{P}_N]\\), where \\(N\\) denotes the number of degradation types and is set to 3. These learned prompts \\(\\mathbf{P}\\) can" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.501, + 0.892, + 0.682 + ], + "angle": 0, + "content": "adaptively learn a diverse range of degradation representations and are integrated with the intrinsic features extracted from \\(\\mathbf{I}_{rec}\\), enabling the proposed IBSN to handle multiple types of degradations using a single set of parameters. To better foster the interaction between the input features \\(\\mathbf{F}_c\\) and the degradation prompt \\(\\mathbf{P}\\), the features \\(\\mathbf{F}_c\\) are passed to a global average pooling (GAP) layer, a \\(1\\times 1\\) convolution, and a softmax operator to produce a set of dynamic weight coefficients. Each degradation prompt \\(\\mathbf{P}_i\\) is combined using these dynamic coefficients \\(\\mathbf{w}_{p\\otimes i}\\) and subsequently integrated via an upsampling operator \\(\\uparrow\\) and \\(3\\times 3\\) convolution to obtain the enhanced representation \\(\\mathbf{P}_c\\)." + }, + { + "type": "equation", + "bbox": [ + 0.563, + 0.686, + 0.891, + 0.715 + ], + "angle": 0, + "content": "\\[\n\\mathbf {P} _ {c} = \\operatorname {C o n v} _ {3 \\times 3} \\left(\\left(\\sum_ {i = 1} ^ {N} \\mathbf {w} _ {p * i} \\mathbf {P} _ {i}\\right) _ {\\uparrow}\\right), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.521, + 0.716, + 0.848, + 0.732 + ], + "angle": 0, + "content": "where \\(\\mathbf{w}_p = \\mathrm{Softmax}\\left(\\mathrm{Conv}_{1\\times 1}\\left(\\mathrm{GAP}\\left(\\mathbf{F}_c\\right)\\right)\\right)\\)" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.737, + 0.892, + 0.797 + ], + "angle": 0, + "content": "Finally, we utilize a \\(3 \\times 3\\) convolution to fuse the learned degradation representation conditioned on the extracted features \\(\\mathbf{F}_c\\) to enrich the degradation-specific context, obtaining \\(\\hat{\\mathbf{Z}}\\). This process can be formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.561, + 0.801, + 0.891, + 0.82 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathbf {Z}} = \\operatorname {C o n v} _ {3 \\times 3} \\left(\\left[ \\mathbf {P} _ {c}; \\mathbf {F} _ {c} \\right]\\right) \\in \\mathbb {R} ^ {\\frac {H}{2} \\times \\frac {W}{2} \\times 1 2}. \\tag {4}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.832, + 0.83, + 0.848 + ], + "angle": 0, + "content": "4.4. Bit Encryption and Recovery Modules" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.856, + 0.892, + 0.901 + ], + "angle": 0, + "content": "As shown in Fig. 4, to encode the copyright watermark \\(\\mathbf{w}_{\\text{cop}}\\) into \\(\\mathbf{I}_{\\text{med}}\\), we firstly expand \\(\\mathbf{w}_{\\text{cop}} \\in \\{0,1\\}^L\\) via stacked MLPs and reshape it into several \\(L \\times L\\) message feature" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.956 + ], + "angle": 0, + "content": "11968" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.089, + 0.47, + 0.117 + ], + "angle": 0, + "content": "Table 1. Localization precision (F1-Score) and bit accuracy (BA) comparison with other competitive methods on [9, 16, 20, 58]." + }, + { + "type": "table", + "bbox": [ + 0.081, + 0.123, + 0.469, + 0.223 + ], + "angle": 0, + "content": "
MethodCAISAv1 [9]Coverage [58]NIST16 [16]Columbia [20]
F1BA(%)F1BA(%)F1BA(%)F1BA(%)
ManTraNet [63]0.320-0.486-0.225-0.650-
SPAN [22]0.169-0.428-0.363-0.873-
CAT-Net v2 [32]0.852-0.582-0.417-0.923-
OSN [61]0.676-0.472-0.449-0.836-
MVSS-Net [8]0.650-0.659-0.372-0.781-
PSCC-Net [39]0.670-0.615-0.210-0.760-
TruFor [17]0.822-0.735-0.470-0.914-
EditGuard (Ours)0.95499.910.9551000.91199.880.98899.93
" + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.228, + 0.468, + 0.389 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.393, + 0.47, + 0.422 + ], + "angle": 0, + "content": "Figure 6. Localization precision comparisons of our EditGuard and competitive methods [8, 39, 61] on four classical benchmarks." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.426, + 0.473, + 0.562 + ], + "angle": 0, + "content": "maps. Meanwhile, \\(\\mathbf{I}_{med}\\) is fed to a U-style feature enhancement network to extract features of each downsampling and upsampling layer. Finally, the message features will be upscaled and integrated with multi-level image features via the fusion mechanism [21, 62], achieving the modulation of bit-image information. In the decoding end, \\(\\mathbf{I}_{rec}\\) is fed into a U-shaped sub-network and downsampled to a size of \\(L\\times L\\). The recovered copyright watermark \\(\\hat{\\mathbf{w}}_{cop}\\) is then extracted via an MLP. More details are presented in S.M.." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.57, + 0.383, + 0.585 + ], + "angle": 0, + "content": "4.5. Construct EditGuard via the IBSN" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.593, + 0.47, + 0.655 + ], + "angle": 0, + "content": "To stabilize the optimization of the proposed IBSN, we propose a bi-level optimization strategy. Given an arbitrary original image \\(\\mathbf{I}_{med}\\) and watermark \\(\\mathbf{w}_{cop}\\), we first train the bit encryption and recovery module via the \\(\\ell_{2}\\) loss." + }, + { + "type": "equation", + "bbox": [ + 0.12, + 0.657, + 0.469, + 0.677 + ], + "angle": 0, + "content": "\\[\n\\ell_ {c o p} = \\left\\| \\mathbf {I} _ {c o n} - \\mathbf {I} _ {m e d} \\right\\| _ {2} ^ {2} + \\lambda \\left\\| \\hat {\\mathbf {w}} _ {c o p} - \\mathbf {w} _ {c o p} \\right\\| _ {2} ^ {2}, \\quad (5)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.679, + 0.47, + 0.741 + ], + "angle": 0, + "content": "where \\(\\lambda\\) is set to 10. Furthermore, we freeze the weights of BEM and BRM and jointly train the IHM and IRM. Given a random original image \\(\\mathbf{I}_{ori}\\), localization watermark \\(\\mathbf{W}_{loc}\\) and copyright watermark \\(\\mathbf{w}_{cop}\\), the loss function is:" + }, + { + "type": "equation", + "bbox": [ + 0.084, + 0.745, + 0.469, + 0.761 + ], + "angle": 0, + "content": "\\[\n\\ell_ {l o c} = \\left\\| \\hat {\\mathbf {I}} _ {o r i} - \\mathbf {I} _ {o r i} \\right\\| _ {1} + \\alpha \\left\\| \\mathbf {I} _ {c o n} - \\mathbf {I} _ {o r i} \\right\\| _ {2} ^ {2} + \\beta \\left\\| \\hat {\\mathbf {W}} _ {l o c} - \\mathbf {W} _ {l o c} \\right\\| _ {1}, \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.471, + 0.902 + ], + "angle": 0, + "content": "where \\(\\alpha\\) and \\(\\beta\\) are respectively set to 100 and 1. During training, we only introduced degradations to \\(\\mathbf{I}_{con}\\) without being exposed to any tampering. After acquiring a pretrained IBSN, we can construct the proposed EditGuard via the components of IBSN. As plotted in Fig. 4, the dual-watermark encoder of EditGuard is composed of IHM and BEM, which correspond to the localization and copyright watermarking in Fig. 3 respectively. The copyright extractor strictly corresponds to BRM. The tamper locator" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.089, + 0.892, + 0.118 + ], + "angle": 0, + "content": "Table 2. Visual quality of the container image \\( \\mathbf{I}_{con} \\) and bit accuracy comparison with other pure watermarking methods." + }, + { + "type": "table", + "bbox": [ + 0.504, + 0.122, + 0.892, + 0.204 + ], + "angle": 0, + "content": "
MethodImage SizeM. L.PSNR (dB)SSIMNIQE(↓)BA(%)
MBRS [26]128×1283026.570.8867.219100
CIN [44]128×1283041.350.9817.17199.99
PIMoG [10]128×1283036.220.9417.11399.99
SepMark [62]128×1283035.420.9317.09599.86
EditGuard128×1283036.930.9445.56799.89
EditGuard512×5126437.770.9494.25799.95
" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.208, + 0.892, + 0.314 + ], + "angle": 0, + "content": "includes IRM and a mask extractor (ME). Note that we need to pre-define a localization watermark \\(\\mathbf{W}_{loc}\\), which is shared between the encoding and decoding ends. The choice of \\(\\mathbf{W}_{loc}\\) is very general to our method. It can be any natural image or even a solid color image. Finally, by comparing the pre-defined watermark \\(\\mathbf{W}_{loc}\\) with the decoded one \\(\\hat{\\mathbf{W}}_{loc}\\), we can obtain a binary mask \\(\\hat{\\mathbf{M}}\\in \\mathbb{R}^{H\\times W}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.519, + 0.317, + 0.892, + 0.335 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathbf {M}} [ i, j ] = \\theta_ {\\tau} \\left(\\max \\left(\\left| \\hat {\\mathbf {W}} _ {l o c} [ i, j,: ] - \\mathbf {W} _ {l o c} [ i, j,: ] \\right|\\right)\\right). \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.338, + 0.892, + 0.37 + ], + "angle": 0, + "content": "where \\(i\\in [0,H)\\) and \\(j\\in [0,W)\\). \\(\\theta_{\\tau}(z) = 1\\) (\\(z\\geq \\tau\\)). \\(\\tau\\) is set to 0.2. \\(|\\cdot |\\) is an absolute value operation." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.383, + 0.634, + 0.4 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.408, + 0.719, + 0.423 + ], + "angle": 0, + "content": "5.1. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.431, + 0.892, + 0.643 + ], + "angle": 0, + "content": "We trained our EditGuard via the training set of COCO [38] without any tampered data. Thus, for tamper localization, our method is actually zero-shot. The Adam [30] is used for training \\(250K\\) iterations with \\(\\beta_{1} = 0.9\\) and \\(\\beta_{2} = 0.5\\). The learning rate is initialized to \\(1\\times 10^{-4}\\) and decreases by half for every \\(30K\\) iterations, with the batch size set to 4. We embed a 64-bit copyright watermark and a simple localization watermark such as a pure blue image ([R, G, B] = [0, 0, 255]) to original images. Following [8, 17, 39], F1-score, AUC, IoU, and bit accuracy are used to evaluate localization and copyright protection performance. Since no prior methods can simultaneously achieve this dual forensics, we conducted separate comparisons with tamper localization and image watermarking methods." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.651, + 0.842, + 0.667 + ], + "angle": 0, + "content": "5.2. Comparison with Localization Methods" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.675, + 0.892, + 0.901 + ], + "angle": 0, + "content": "For a fair comparison with tamper localization methods, we conducted extensive evaluations on four classical benchmarks [9, 16, 20, 58], as reported on Tab. 1. Since EditGuard is a proactive approach, we initially embed watermarks into authentic images and then paste the tampered areas into the container images. Remarkably, even for tamper types that existing methods specialize in, the localization accuracy of EditGuard consistently outperforms the SOTA method [17] across four datasets by margins of 0.102, 0.116, 0.441, and 0.065 in F1-score without any labeled data or tampered samples required, which verifies the superiority of our proactive localization mechanism. As shown in Fig. 6, our EditGuard can precisely pinpoint pixel-level tampered areas but other methods can only produce a rough outline or are only effective in some cases. Mean-" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "11969" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.09, + 0.892, + 0.116 + ], + "angle": 0, + "content": "Table 3. Comparison with other competitive tamper localization methods under different AIGC-based editing methods. Note that \\( \\dagger \\) denotes the network finetuned in our constructed AGE-Set-C." + }, + { + "type": "table", + "bbox": [ + 0.081, + 0.117, + 0.892, + 0.195 + ], + "angle": 0, + "content": "
MethodStable Diffusion Inpaint [51]Controlnet [72]SDXL [49]RePaint [43]Lama [57]FaceSwap [60]
F1AUCIoUBA(%)F1AUCIoUBA(%)F1AUCIoUBA(%)F1AUCIoUBA(%)F1AUCIoUBA(%)F1AUCIoUBA(%)
MVSS-Net [8]0.1780.4880.103-0.1780.4920.103-0.0370.5030.028-0.1040.5460.082-0.0240.5050.022-0.2850.6120.192-
OSN [61]0.1740.4860.101-0.1910.6440.110-0.2000.7550.118-0.1830.6440.105-0.1700.4300.099-0.3080.7910.171-
PSCC-Net [39]0.1660.5010.112-0.1770.5650.116-0.1890.7040.115-0.1400.4690.109-0.1320.3290.104-0.1570.3460.180-
IML-VIT [45]0.2130.5960.135-0.2000.5760.128-0.2210.6030.145-0.1030.4970.059-0.1050.4650.064-0.1050.4650.064-
HiFi-Net [18]0.5470.7340.128-0.5420.7350.123-0.6330.8280.261-0.6810.8960.339-0.4830.7210.029-0.7810.8900.478-
MVSS-Net†[8]0.6940.9390.575-0.6780.9250.558-0.4820.8840.359-0.1850.5290.111-0.3930.8290.275-0.4590.7390.333-
EditGuard (Ours)0.9660.9710.93699.950.9680.9870.94099.960.9650.9890.93699.960.9670.9770.93899.950.9650.9690.93499.950.8960.9430.87699.86
" + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.202, + 0.163, + 0.257 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.165, + 0.202, + 0.238, + 0.258 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.242, + 0.202, + 0.316, + 0.258 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.318, + 0.202, + 0.392, + 0.258 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.395, + 0.202, + 0.468, + 0.258 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.081, + 0.261, + 0.162, + 0.317 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.166, + 0.261, + 0.238, + 0.317 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.242, + 0.261, + 0.316, + 0.317 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.318, + 0.261, + 0.392, + 0.317 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.395, + 0.261, + 0.468, + 0.317 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.081, + 0.32, + 0.162, + 0.375 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.166, + 0.32, + 0.238, + 0.376 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.242, + 0.321, + 0.316, + 0.376 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.318, + 0.321, + 0.392, + 0.376 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.395, + 0.321, + 0.468, + 0.376 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.38, + 0.469, + 0.422 + ], + "angle": 0, + "content": "Figure 7. Visual results of the container image \\(\\mathbf{I}_{con}\\) and the error map of the proposed EditGuard. Here, localization and copyright watermarks are randomly selected from the dataset." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.425, + 0.469, + 0.456 + ], + "angle": 0, + "content": "while, our bit accuracy remains over \\(99.8\\%\\) while all other methods can not realize effective copyright protection." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.466, + 0.436, + 0.482 + ], + "angle": 0, + "content": "5.3. Comparison with Watermarking Methods" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.489, + 0.47, + 0.807 + ], + "angle": 0, + "content": "To evaluate the visual quality of \\(\\mathbf{I}_{con}\\), we compared EditGuard with other watermarking methods on \\(1K\\) testing images from COCO [38] under the tampering of stable diffusion inpaint [51]. For a fair comparison, we also retrained our EditGuard on \\(128\\times 128\\) original images and 30 bits. Tab. 2 reports that the fidelity of our container image far surpasses that of SepMark [62], PIMoG [10], and MBRS [26] but is slightly inferior to CIN [44]. Meanwhile, our method exhibits the best performance in perceptual quality measures like NIQE. As shown in Fig. 7, dual-watermarked images do not have noticeable artifacts and noise, making them imperceptible to the human eyes. When suffer malicious tampering, our method outperforms SepMark and is very close to PIMoG and CIN in bit accuracy. Note that other competitive methods only hide 30 bits, with a capacity of \\(30 / (128\\times 128)\\). In contrast, our EditGuard hides both an RGB localization watermark and a 1D copyright watermark, with a capacity far greater than \\(30 / (128\\times 128)\\). Here, we do not claim to achieve the best visual quality and bit accuracy, but just to demonstrate that our method is comparable to the current image watermarking methods." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.817, + 0.443, + 0.833 + ], + "angle": 0, + "content": "5.4. Extension to AIGC-based Editing Methods" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.47, + 0.901 + ], + "angle": 0, + "content": "Dataset Preparation: We constructed a dataset tailored for AIGC Editing methods, dubbed AGE-Set, comprising two sub-datasets. The first AGE-Set-C is a batch-processed coarse tamper dataset. Its original images are sourced from" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.202, + 0.606, + 0.243 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.608, + 0.202, + 0.7, + 0.243 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.702, + 0.202, + 0.77, + 0.243 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.772, + 0.202, + 0.815, + 0.243 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.817, + 0.202, + 0.89, + 0.243 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.504, + 0.253, + 0.605, + 0.28 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.607, + 0.245, + 0.7, + 0.28 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.702, + 0.245, + 0.77, + 0.28 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.772, + 0.245, + 0.815, + 0.28 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.817, + 0.245, + 0.89, + 0.28 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.281, + 0.605, + 0.316 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.607, + 0.281, + 0.7, + 0.316 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.702, + 0.281, + 0.77, + 0.316 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.772, + 0.281, + 0.815, + 0.316 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.817, + 0.281, + 0.89, + 0.316 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.504, + 0.317, + 0.605, + 0.353 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.607, + 0.317, + 0.7, + 0.353 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.702, + 0.317, + 0.77, + 0.353 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.772, + 0.317, + 0.815, + 0.353 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.817, + 0.317, + 0.89, + 0.353 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.5, + 0.355, + 0.892, + 0.383 + ], + "angle": 0, + "content": "Figure 8. Localization performance comparisons of our EditGuard and other methods [8, 18, 39, 61] on our constructed AGE-Set-C." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.387, + 0.892, + 0.599 + ], + "angle": 0, + "content": "COCO 2017 [38] and CelebA [41], containing \\(30K\\) training images and \\(1.2K\\) testing images. We used some SOTA editing methods such as Stable Diffusion Inpaint [51], ControlNet [72], SDXL [49] to manipulate images with the prompt to be \"None\", and employed some unconditional methods like Repaint [43], Lama [57], and Faceswap [60]. Note that we only use the tampered data to train other methods, not our EditGuard. The second sub-dataset AGE-Set-F includes 100 finely edited images. It is edited manually via some sophisticated software such as SD-Web-UI, Photoshop, and Adobe Firefly. These AIGC-based editing methods can achieve a good fusion of the tampered and unchanged areas, making it hard for the naked eye to catch artifacts. More details are presented in S.M.." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.599, + 0.893, + 0.856 + ], + "angle": 0, + "content": "AGE-Set-C: Tab. 3 presents the comparison of our EditGuard and some SOTA tamper localization methods [8, 18, 39, 45, 61]. We observe that the F1-scores of other passive forensic methods are generally lower than 0.7 when applied to AGE-Set-C. Meanwhile, even when we try our best to finetune MVSS-Net using AGE-Set-C, the accuracy of MVSS-Net remains unsatisfactory, and they exhibit catastrophic forgetting across various tamper methods. In contrast, our method can guarantee an F1-score and AUC of over \\(95\\%\\), maintaining around \\(90\\%\\) IOU, regardless of tampering types. As shown in Fig. 8, our EditGuard can accurately capture these imperceptible tampering traces produced by AIGC-based editing methods, but other methods are almost ineffective. Moreover, our EditGuard can effectively recover copyright information with a bit accuracy exceeding \\(99.8\\%\\). Noting that none of the comparison methods offer copyright protection capabilities." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.893, + 0.901 + ], + "angle": 0, + "content": "AGE-Set-F: To further highlight the practicality of our EditGuard, we conducted subjective comparisons with other methods on the meticulously tampered AGE-Set-F." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "11970" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.081, + 0.09, + 0.89, + 0.342 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.08, + 0.343, + 0.889, + 0.356 + ], + "angle": 0, + "content": "Figure 9. Localization precision comparisons of our EditGuard and other competitive methods on the meticulously tampered AGE-Set-F." + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.36, + 0.468, + 0.389 + ], + "angle": 0, + "content": "Table 4. Localization and bit recovery performance of our Edit-Guard and MVSS-Net† [8] under different levels of degradations." + }, + { + "type": "table", + "bbox": [ + 0.081, + 0.391, + 0.468, + 0.453 + ], + "angle": 0, + "content": "
MethodsMetricsCleanGaussian NoiseJPEGPoisson
σ=1σ=5Q=70Q=80Q=90
MVSS-Net†[8]F10.6940.6440.6190.4580.5070.5580.652
BA(%)-------
EditGuard (Ours)F10.9660.9370.9320.9200.9200.9250.943
BA(%)99.9599.9499.3797.6998.1698.2399.91
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.458, + 0.468, + 0.579 + ], + "angle": 0, + "content": "The tamper types in this subset did not appear in the training set. As shown in Fig. 9, when faced with real-world tampering, even the most powerful tamper localization methods almost entirely fail. This is due to their mechanisms to look for image artifacts and explore instance-wise semantic information. However, our EditGuard, which locates tampered masks via the natural fragility and locality of I2I steganography, can still clearly annotate the tampered area." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.588, + 0.27, + 0.604 + ], + "angle": 0, + "content": "5.5. Robustness Analysis" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.612, + 0.468, + 0.777 + ], + "angle": 0, + "content": "As shown in Tab. 4, we conducted robustness analysis on the tampering of \"Stable Diffusion Inpaint\" [51] under Gaussian noise with \\(\\sigma = 1\\) and 5, JPEG compression with \\(Q = 70\\), 80 and 90, and Poisson noise with \\(\\alpha = 4\\) [71]. We observed that our method still maintains a high localization accuracy (F1-score \\(>0.9\\)) and bit accuracy with a very slight performance decrease under various levels of degradations, while MVSS-Net† [8] exhibits a noticeable performance degradation compared to its results in clean conditions. It is attributed to our prompt-based estimation that can effectively learn the degradation representation." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.787, + 0.229, + 0.803 + ], + "angle": 0, + "content": "5.6. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.81, + 0.468, + 0.9 + ], + "angle": 0, + "content": "To verify the effectiveness of each component of the Edit-Guard, we conducted ablation studies on bi-level optimization (BO), lightweight feature interaction module (LFIM), transformer block (TB), and prompt-based fusion (PF) under the tampering of \"Stable Diffusion Inpaint\". As listed in Tab. 5, without BO, the joint training of all components" + }, + { + "type": "table_caption", + "bbox": [ + 0.503, + 0.36, + 0.889, + 0.374 + ], + "angle": 0, + "content": "Table 5. Abalation studies on the core components of EditGuard." + }, + { + "type": "table", + "bbox": [ + 0.504, + 0.376, + 0.89, + 0.448 + ], + "angle": 0, + "content": "
CaseDegradation Type D(·)PFTBLFIMBOF1AUCIoUBA(%)
(a)Clean✓✓✓---49.17
(b)Clean✓✓✓0.9500.9600.90499.73
(c)Clean✓✓✓0.9570.9660.92799.51
(d)Random Degradations✓✓✓0.9030.9330.84199.12
OursClean✓✓✓✓0.9660.9710.93699.95
Random Degradations✓✓✓✓0.9380.9640.88799.36
" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.454, + 0.892, + 0.635 + ], + "angle": 0, + "content": "cannot converge effectively, resulting in bit accuracy that is close to random guessing. Without LFIM and TB, the IoU of EditGuard will suffer 0.032 and 0.009 declines since these two modules can better perform feature fusion. Without PF, the robustness of the EditGuard will significantly decline. We observed that the F1/AUC/IoU of our method far surpasses that of case (d) by 0.035/0.031/0.046 under \"Random Degradations\", which indicates that the PF effectively enables a single network to support watermark recovery under various degradations. \"Random Degradations\" denotes that we randomly set the \\(\\mathcal{D}(\\cdot)\\) to various levels of Gaussian noise, Poisson noise, and JPEG compression." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.649, + 0.618, + 0.664 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.674, + 0.892, + 0.9 + ], + "angle": 0, + "content": "We present the first attempt to design a deep versatile watermarking mechanism EditGuard. It enhances the credibility of images by embedding imperceptible localization and copyright watermarks, and decoding accurate copyright information and tampered areas, making it a reliable tool for artistic creation and legal forensic analysis. In the future, we will focus on improving the robustness of EditGuard and strive not only to offer pixel-wise localization results but also to provide semantic-wise outcomes. Additionally, we plan to further expand EditGuard to a broader range of modalities and applications, including video, audio, and 3D scenes. Our efforts at information authenticity serve not only the AIGC industry, but the trust in our digital world, ensuring that every pixel tells the truth and the rights of each individual are safeguarded." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "11971" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.079, + 0.09, + 0.176, + 0.107 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.171 + ], + "angle": 0, + "content": "[1] Mahdi Ahmadi, Alireza Norouzi, Nader Karimi, Shadrokh Samavi, and Ali Emami. Redmark: Framework for residual diffusion watermarking based on deep networks. Expert Systems with Applications, 146:113157, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.173, + 0.472, + 0.228 + ], + "angle": 0, + "content": "[2] Vishal Asnani, Xi Yin, Tal Hassner, and Xiaoming Liu. Malp: Manipulation localization using a proactive scheme. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.23, + 0.47, + 0.285 + ], + "angle": 0, + "content": "[3] Shumeet Baluja. Hiding images in plain sight: Deep steganography. In Proceedings of the International Conference on Neural Information Processing Systems (NeurIPS), 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.287, + 0.47, + 0.329 + ], + "angle": 0, + "content": "[4] Liangyu Chen, Xiaojie Chu, Xiangyu Zhang, and Jian Sun. Simple baselines for image restoration. In Proceedings of the European Conference on Computer Vision (ECCV), 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.331, + 0.47, + 0.387 + ], + "angle": 0, + "content": "[5] Xinru Chen, Chengbo Dong, Jiaqi Ji, Juan Cao, and Xirong Li. Image manipulation detection by multi-view multi-scale supervision. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.388, + 0.47, + 0.445 + ], + "angle": 0, + "content": "[6] Baotian Cheng, Rongrong Ni, and Yao Zhao. A refining localization watermarking for image tamper detection and recovery. In 2012 IEEE 11th International Conference on Signal Processing, pages 984-988, 2012. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.446, + 0.47, + 0.502 + ], + "angle": 0, + "content": "[7] Yingqian Cui, Jie Ren, Han Xu, Pengfei He, Hui Liu, Lichao Sun, and Jiliang Tang. Diffusionshield: A watermark for copyright protection against generative diffusion models. arXiv preprint arXiv:2306.04642, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.503, + 0.47, + 0.572 + ], + "angle": 0, + "content": "[8] Chengbo Dong, Xinru Chen, Ruohan Hu, Juan Cao, and Xirong Li. Mvss-net: Multi-view multi-scale supervised networks for image manipulation detection. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(3):3539-3553, 2022. 2, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.574, + 0.47, + 0.631 + ], + "angle": 0, + "content": "[9] Jing Dong, Wei Wang, and Tieniu Tan. Casia image tampering detection evaluation database. In Proceedings of the IEEE China Summit and International Conference on Signal and Information Processing (ChinaSIP), 2013. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.632, + 0.47, + 0.702 + ], + "angle": 0, + "content": "[10] Han Fang, Zhaoyang Jia, Zehua Ma, Ee-Chien Chang, and Weiming Zhang. Pimog: An effective screen-shooting noise-layer simulation for deep-learning-based watermarking network. In Proceedings of the 30th ACM International Conference on Multimedia (MM), 2022. 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.703, + 0.47, + 0.773 + ], + "angle": 0, + "content": "[11] Han Fang, Yupeng Qiu, Kejiang Chen, Jiyi Zhang, Weiming Zhang, and Ee-Chien Chang. Flow-based robust watermarking with invertible noise layer for black-box distortions. In Proceedings of the AAAI Conference on Artificial Intelligence (AAAI), 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.774, + 0.47, + 0.844 + ], + "angle": 0, + "content": "[12] Pierre Fernandez, Guillaume Couairon, Hervé Jégou, Matthijs Douze, and Teddy Furon. The stable signature: Rooting watermarks in latent diffusion models. Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.845, + 0.47, + 0.902 + ], + "angle": 0, + "content": "[13] Anirban Ghoshal. Artists lose first copyright battle in the fight against ai-generated images. https://www.computerworld.com/article/3709691/artists-lose-first-copyright-battle-in" + }, + { + "type": "list", + "bbox": [ + 0.079, + 0.115, + 0.472, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.89, + 0.12 + ], + "angle": 0, + "content": "the-fight-against-ai-generated-images. \nhtml,2023.2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.123, + 0.893, + 0.178 + ], + "angle": 0, + "content": "[14] Gabriel Goh, James Betker, Li Jing, Aditya Ramesh, et al. Improving image generation with better captions. https://cdn.openai.com/papers/dall-e-3.pdf, 2023.1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.181, + 0.89, + 0.21 + ], + "angle": 0, + "content": "[15] Google DeepMind. Synthid. https://deepmind.google/technologies/synthid/, 2023.2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.212, + 0.893, + 0.295 + ], + "angle": 0, + "content": "[16] Haiying Guan, Mark Kozak, Eric Robertson, Yooyoung Lee, Amy N Yates, Andrew Delgado, Daniel Zhou, Timothee Kheyrkhah, Jeff Smith, and Jonathan Fiscus. Mfc datasets: Large-scale benchmark datasets for media forensic challenge evaluation. In Proceedings of the IEEE Winter Applications of Computer Vision Workshops (WACVW), 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.297, + 0.893, + 0.368 + ], + "angle": 0, + "content": "[17] Fabrizio Guillaro, Davide Cozzolino, Avneesh Sud, Nicholas Dufour, and Luisa Verdoliva. Trufor: Leveraging all-round clues for trustworthy image forgery detection and localization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.369, + 0.894, + 0.439 + ], + "angle": 0, + "content": "[18] Xiao Guo, Xiaohong Liu, Zhiyuan Ren, Steven Grosz, Iacopo Masi, and Xiaoming Liu. Hierarchical fine-grained image forgery detection and localization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.441, + 0.893, + 0.497 + ], + "angle": 0, + "content": "[19] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.499, + 0.893, + 0.556 + ], + "angle": 0, + "content": "[20] Yu-Feng Hsu and Shih-Fu Chang. Detecting image splicing using geometry invariants and camera characteristics consistency. In Proceedings of the IEEE International Conference on Multimedia and Expo (ICME), 2006. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.557, + 0.893, + 0.6 + ], + "angle": 0, + "content": "[21] Jie Hu, Li Shen, and Gang Sun. Squeeze-and-excitation networks. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.601, + 0.893, + 0.672 + ], + "angle": 0, + "content": "[22] Xuefeng Hu, Zhihan Zhang, Zhenye Jiang, Syomantak Chaudhuri, Zhenheng Yang, and Ram Nevatia. Span: Spatial pyramid attention network for image manipulation localization. In Proceedings of the European Conference on Computer Vision (ECCV), 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.673, + 0.893, + 0.743 + ], + "angle": 0, + "content": "[23] Xiaoxiao Hu, Qichao Ying, Zhenxing Qian, Sheng Li, and Xinpeng Zhang. Draw: Defending camera-shooted raw against image manipulation. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.745, + 0.894, + 0.815 + ], + "angle": 0, + "content": "[24] Nasir N Hurrah, Shabir A Parah, Nazir A Loan, Javaid A Sheikh, Mohammad Elhoseny, and Khan Muhammad. Dual watermarking framework for privacy protection and content authentication of multimedia. Future generation computer Systems, 94:654-673, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.817, + 0.894, + 0.9 + ], + "angle": 0, + "content": "[25] Ashraful Islam, Chengjiang Long, Arslan Basharat, and Anthony Hoogs. Doa-gan: Dual-order attentive generative adversarial network for image copy-move forgery detection and localization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "11972" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.161 + ], + "angle": 0, + "content": "[26] Zhaoyang Jia, Han Fang, and Weiming Zhang. Mbrs: Enhancing robustness of dnn-based watermarking by minibatch of real and simulatedJPEG compression. In Proceedings of the 29th ACM International Conference on Multimedia (MM), 2021. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.163, + 0.47, + 0.205 + ], + "angle": 0, + "content": "[27] Zhengyuan Jiang, Jinghuai Zhang, and Neil Zhenqiang Gong. Evading watermark based detection of ai-generated content. arXiv preprint arXiv:2305.03807, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.207, + 0.47, + 0.262 + ], + "angle": 0, + "content": "[28] Junpeng Jing, Xin Deng, Mai Xu, Jianyi Wang, and Zhenyu Guan. Hinet: Deep image hiding by invertible network. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.264, + 0.47, + 0.332 + ], + "angle": 0, + "content": "[29] Asra Kamili, Nasir N Hurrah, Shabir A Parah, Ghulam Mohiuddin Bhat, and Khan Muhammad. Dwfcat: Dual watermarking framework for industrial image authentication and tamper localization. IEEE Transactions on Industrial Informatics, 17(7):5108-5117, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.334, + 0.47, + 0.375 + ], + "angle": 0, + "content": "[30] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.377, + 0.47, + 0.446 + ], + "angle": 0, + "content": "[31] Myung-Joon Kwon, In-Jae Yu, Seung-Hun Nam, and Heung-Kyu Lee. Cat-net: Compression artifact tracing network for detection and localization of image splicing. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.448, + 0.469, + 0.516 + ], + "angle": 0, + "content": "[32] Myung-Joon Kwon, Seung-Hun Nam, In-Jae Yu, Heung-Kyu Lee, and Changick Kim. LearningJPEG compression artifacts for image manipulation detection and localization. International Journal of Computer Vision, 130(8):1875-1895, 2022.6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.518, + 0.469, + 0.573 + ], + "angle": 0, + "content": "[33] Haodong Li and Jiwu Huang. Localization of deep inpainting using high-pass fully convolutional network. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.575, + 0.469, + 0.63 + ], + "angle": 0, + "content": "[34] Yuanman Li and Jiantao Zhou. Fast and effective image copy-move forgery detection via hierarchical feature point matching. IEEE Transactions on Information Forensics and Security, 14(5):1307-1322, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.632, + 0.469, + 0.687 + ], + "angle": 0, + "content": "[35] Yue Li, Dong Liu, Houqiang Li, Li Li, Zhu Li, and Feng Wu. Learning a convolutional neural network for image compact-resolution. IEEE Transactions on Image Processing, 28(3): 1092-1107, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.689, + 0.469, + 0.744 + ], + "angle": 0, + "content": "[36] Siau-Chuin Liew, Siau-Way Liew, and Jasni Mohd Zain. Tamper localization and lossless recovery watermarking scheme with roi segmentation and multilevel authentication. Journal of digital imaging, 26:316-325, 2013. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.746, + 0.469, + 0.8 + ], + "angle": 0, + "content": "[37] Chia-Chen Lin, Ting-Lin Lee, Ya-Fen Chang, Pei-Feng Shiu, and Bohan Zhang. Fragile watermarking for tamper localization and self-recovery based on ambtc and vq. Electronics, 12(2):415, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.803, + 0.469, + 0.871 + ], + "angle": 0, + "content": "[38] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Proceedings of the European Conference on Computer Vision (ECCV), 2014. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[39] Xiaohong Liu, Yaojie Liu, Jun Chen, and Xiaoming Liu. Pssc-net: Progressive spatio-channel correlation network for" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.892, + 0.134 + ], + "angle": 0, + "content": "image manipulation detection and localization. IEEE Transactions on Circuits and Systems for Video Technology, 32 (11):7505-7517, 2022. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.136, + 0.892, + 0.204 + ], + "angle": 0, + "content": "[40] Yang Liu, Mengxi Guo, Jian Zhang, Yuesheng Zhu, and Xiaodong Xie. A novel two-stage separable deep learning framework for practical blind watermarking. In Proceedings of the ACM International Conference on Multimedia (MM), 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.207, + 0.892, + 0.262 + ], + "angle": 0, + "content": "[41] Ziwei Liu, Ping Luo, Xiaogang Wang, and Xiaou Tang. Deep learning face attributes in the wild. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), 2015. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.264, + 0.892, + 0.331 + ], + "angle": 0, + "content": "[42] Shao-Ping Lu, Rong Wang, Tao Zhong, and Paul L Rosin. Large-capacity image steganography based on invertible neural networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.334, + 0.892, + 0.403 + ], + "angle": 0, + "content": "[43] Andreas Lugmayr, Martin Danelljan, Andres Romero, Fisher Yu, Radu Timofte, and Luc Van Gool. Repaint: Inpainting using denoising diffusion probabilistic models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 1, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.405, + 0.892, + 0.473 + ], + "angle": 0, + "content": "[44] Rui Ma, Mengxi Guo, Yi Hou, Fan Yang, Yuan Li, Huizhu Jia, and Xiaodong Xie. Towards blind watermarking: Combining invertible and non-invertible mechanisms. In Proceedings of the ACM International Conference on Multimedia (MM), 2022. 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.476, + 0.892, + 0.53 + ], + "angle": 0, + "content": "[45] Xiaochen Ma, Bo Du, Xianggen Liu, Ahmed Y Al Hammadi, and Jizhe Zhou. Iml-vit: Image manipulation localization by vision transformer. arXiv preprint arXiv:2307.14863, 2023. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.533, + 0.892, + 0.602 + ], + "angle": 0, + "content": "[46] Chong Mou, Youmin Xu, Jiechong Song, Chen Zhao, Bernard Ghanem, and Jian Zhang. Large-capacity and flexible video steganography via invertible neural network. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.604, + 0.892, + 0.673 + ], + "angle": 0, + "content": "[47] Chong Mou, Xintao Wang, Liangbin Xie, Yanze Wu, Jian Zhang, Zhongang Qi, Ying Shan, and Xiaohu Qie. T2i-adapter: Learning adapters to dig out more controllable ability for text-to-image diffusion models. In AAAI Conference on Artificial Intelligence (AAAI), 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.675, + 0.892, + 0.731 + ], + "angle": 0, + "content": "[48] Neena Raj NR and R Shreelekshmi. Fragile watermarking scheme for tamper localization in images using logistic map and singular value decomposition. Journal of Visual Communication and Image Representation, 85:103500, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.732, + 0.892, + 0.8 + ], + "angle": 0, + "content": "[49] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952, 2023. 1, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.802, + 0.892, + 0.857 + ], + "angle": 0, + "content": "[50] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.859, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[51] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "11973" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.11, + 0.093, + 0.469, + 0.12 + ], + "angle": 0, + "content": "the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 1, 3, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.122, + 0.47, + 0.233 + ], + "angle": 0, + "content": "[52] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, Seyed Kamyar Seyed Ghasemipour, Raphael Gontijo-Lopes, Burcu Karagol Ayan, Tim Salimans, Jonathan Ho, David J. Fleet, and Mohammad Norouzi. Photorealistic text-to-image diffusion models with deep language understanding. In Proceedings of the International conference on Neural Information Processing Systems (NeurIPS), 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.235, + 0.47, + 0.29 + ], + "angle": 0, + "content": "[53] Ronald Salloum, Yuzhuo Ren, and C-C Jay Kuo. Image splicing localization using a multi-task fully convolutional network (mfcn). Journal of Visual Communication and Image Representation, 51:201-209, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.292, + 0.469, + 0.361 + ], + "angle": 0, + "content": "[54] Deepa Shivaram. The white house and big tech companies release commitments on managing ai. https://www.nprillinois.org/2023-07-21/the-white-house-and-big-tech-companies-release-commitments-on-managing-ai, 2023.2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.363, + 0.469, + 0.459 + ], + "angle": 0, + "content": "[55] Xiaopeng Sun, Weiqi Li, Zhenyu Zhang, Qiufang Ma, Xuhan Sheng, Ming Cheng, Haoyu Ma, Shijie Zhao, Jian Zhang, Junlin Li, et al. Opdn: Omnidirectional position-aware deformable network for omnidirectional image superresolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshop, pages 1293-1301, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.461, + 0.469, + 0.53 + ], + "angle": 0, + "content": "[56] Zhihao Sun, Haoran Jiang, Danding Wang, Xirong Li, and Juan Cao. Safl-net: Semantic-agnostic feature learning network with auxiliary plugins for image manipulation detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.532, + 0.469, + 0.628 + ], + "angle": 0, + "content": "[57] Roman Suvorov, Elizaveta Logacheva, Anton Mashikhin, Anastasia Remizova, Armenii Ashukha, Aleksei Silvestrov, Naejin Kong, Harshith Goka, Kiwoong Park, and Victor Lempitsky. Resolution-robust large mask inpainting with fourier convolutions. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.631, + 0.469, + 0.699 + ], + "angle": 0, + "content": "[58] Bihan Wen, Ye Zhu, Ramanathan Subramanian, Tian-Tsong Ng, Xuanjing Shen, and Stefan Winkler. Coverage—a novel database for copy-move forgery detection. In Proceedings of the IEEE International Conference on Image Processing (ICIP), 2016. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.702, + 0.469, + 0.757 + ], + "angle": 0, + "content": "[59] Yuxin Wen, John Kirchenbauer, Jonas Geiping, and Tom Goldstein. Tree-rings watermarks: Invisible fingerprints for diffusion images. Advances in Neural Information Processing Systems, 36, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.759, + 0.469, + 0.8 + ], + "angle": 0, + "content": "[60] Huikai Wu, Shaocheng Xiang, Gabriben, and Niczem. Faceswap. https://github.com/wuhuikai/FaceSwap, 2020.7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.802, + 0.469, + 0.87 + ], + "angle": 0, + "content": "[61] Haiwei Wu, Jiantao Zhou, Jinyu Tian, and Jun Liu. Robust image forgery detection over online social network shared images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[62] Xiaoshuai Wu, Xin Liao, and Bo Ou. Sepmark: Deep separable watermarking for unified source tracing and deepfake" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.093, + 0.47, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "detection. In Proceedings of the ACM international conference on Multimedia (MM), 2023. 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.122, + 0.892, + 0.19 + ], + "angle": 0, + "content": "[63] Yue Wu, Wael AbdAlmageed, and Premkumar Natarajan. *Mantra-net: Manipulation tracing network for detection and localization of image forgeries with anomalous features*. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.192, + 0.892, + 0.247 + ], + "angle": 0, + "content": "[64] Youmin Xu, Chong Mou, Yujie Hu, Jingfen Xie, and Jian Zhang. Robust invertible image steganography. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.249, + 0.892, + 0.304 + ], + "angle": 0, + "content": "[65] Qichao Ying, Zhenxing Qian, Hang Zhou, Haisheng Xu, Xinpeng Zhang, and Siyi Li. From image to image: Immunized image generation. In Proceedings of the ACM international conference on Multimedia (MM), 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.305, + 0.892, + 0.374 + ], + "angle": 0, + "content": "[66] Qichao Ying, Xiaoxiao Hu, Xiangyu Zhang, Zhenxing Qian, Sheng Li, and Xinpeng Zhang. Rwn: Robust watermarking network for image cropping localization. In Proceedings of the IEEE International Conference on Image Processing (ICIP), 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.375, + 0.892, + 0.432 + ], + "angle": 0, + "content": "[67] Qichao Ying, Hang Zhou, Zhenxing Qian, Sheng Li, and Xinpeng Zhang. Learning to immunize images for tamper localization and self-recovery. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.433, + 0.892, + 0.5 + ], + "angle": 0, + "content": "[68] Jiwen Yu, Yinhuai Wang, Chen Zhao, Bernard Ghanem, and Jian Zhang. Freedom: Training-free energy-guided conditional diffusion model. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.502, + 0.892, + 0.558 + ], + "angle": 0, + "content": "[69] Jiwen Yu, Xuanyu Zhang, Youmin Xu, and Jian Zhang. Cross: Diffusion model makes controllable, robust and secure image steganography. Advances in Neural Information Processing Systems, 36, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.559, + 0.892, + 0.628 + ], + "angle": 0, + "content": "[70] Syed Waqas Zamir, Aditya Arora, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, and Ming-Hsuan Yang. Restormer: Efficient transformer for high-resolution image restoration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.63, + 0.892, + 0.685 + ], + "angle": 0, + "content": "[71] Kai Zhang, Yawei Li, Jingyun Liang, Jiezhang Cao, Yulun Zhang, Hao Tang, Radu Timofte, and Luc Van Gool. Practical blind denoising via swim-conv-unet and data synthesis. arXiv preprint arXiv:2203.13278, 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.687, + 0.892, + 0.742 + ], + "angle": 0, + "content": "[72] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2023. 1, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.743, + 0.892, + 0.785 + ], + "angle": 0, + "content": "[73] Yunqing Zhao, Tianyu Pang, Chao Du, Xiao Yang, Ngai-Man Cheung, and Min Lin. A recipe for watermarking diffusion models. arXiv preprint arXiv:2303.10137, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.786, + 0.892, + 0.828 + ], + "angle": 0, + "content": "[74] Jiren Zhu, Russell Kaplan, Justin Johnson, and Li Fei-Fei. Hidden: Hiding data with deep networks. In European Conference on Computer Vision (ECCV), 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.829, + 0.892, + 0.883 + ], + "angle": 0, + "content": "[75] Xinshan Zhu, Yongjun Qian, Xianfeng Zhao, Biao Sun, and Ya Sun. A deep learning approach to patch-based image inpainting forensics. Signal Processing: Image Communication, 67:90-99, 2018. 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.883 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "11974" + } + ] +] \ No newline at end of file diff --git a/2024/EditGuard_ Versatile Image Watermarking for Tamper Localization and Copyright Protection/43cb9c2b-dd1a-4670-9150-d4345483687e_origin.pdf b/2024/EditGuard_ Versatile Image Watermarking for Tamper Localization and Copyright Protection/43cb9c2b-dd1a-4670-9150-d4345483687e_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..5edf102d7f1940f33f6a4a1453aef73eca229702 --- /dev/null +++ b/2024/EditGuard_ Versatile Image Watermarking for Tamper Localization and Copyright Protection/43cb9c2b-dd1a-4670-9150-d4345483687e_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:21e2f9102e7fa82a3f883e74215a266614f4183a40e03d3cdef67ae9cd0221c4 +size 7036280 diff --git a/2024/EditGuard_ Versatile Image Watermarking for Tamper Localization and Copyright Protection/full.md b/2024/EditGuard_ Versatile Image Watermarking for Tamper Localization and Copyright Protection/full.md new file mode 100644 index 0000000000000000000000000000000000000000..ab30520115368a6be11ab05bef413c0cb66d9f5b --- /dev/null +++ b/2024/EditGuard_ Versatile Image Watermarking for Tamper Localization and Copyright Protection/full.md @@ -0,0 +1,480 @@ +# EditGuard: Versatile Image Watermarking for Tamper Localization and Copyright Protection + +Xuanyu Zhang $^{1,2}$ , Runyi Li $^{1}$ , Jiwen Yu $^{1}$ , Youmin Xu $^{1}$ , Weiqi Li $^{1}$ , Jian Zhang $^{1,2 \boxtimes}$ $^{1}$ School of Electronic and Computer Engineering, Peking University + +$^{2}$ Peking University Shenzhen Graduate School-Rabbitpre AIGC Joint Research Laboratory + +![](images/d8378e347524d202d48a803c7b961ea40fa20288455a143e19ee3a00096d0467.jpg) +Users + +Please help protect my image against intentional tampering. My exclusive copyright information is 00111010110... + +![](images/5607f94e273b084db6289341ec892ef5497d508e43c3a0c17f984efc08e968a9.jpg) + +I've added the exclusive EditGuard watermark to it. You can post the picture on the left and I will be responsible for its protection. + +![](images/14ebc880a3f4007db11fd8461f3ab11469a88a059d0a69dcb638c998a78d2614.jpg) + +![](images/a2473491b6f8037f8db2e32ad5296ae7fe40b42159dbfece582cd3295334ad29.jpg) + +![](images/ee9111f2a171ed5d97957c269b21f285bd9bb5ad367af501064b52780a751e81.jpg) + +![](images/416b2131a6a646bcd4b4ea856976f1a7d02bcc4351d5efc5192058194f05ad2a.jpg) + +![](images/8ff376315174fa9cc4dcbf7c7f8fae287bd2ae36592540a47f66c922389dc984.jpg) + +I find a picture online that looks similar to mine. Can you provide evidence to prove my copyright and pinpoint tampered areas? + +![](images/d39af36db5bf85a178dc5dfc20cf0b681c0da9f4f224560a93bf926d2bd719b1.jpg) + +![](images/b24586fd356b7e063d993524975730611605f0ae08c4a4601554d75233276b06.jpg) + +![](images/560a5bbb01b919df6b94888b46d2f23b43d556222e6129e9f57072eba95a8914.jpg) +Original + +![](images/dad32f94dbfa1692713fc314f50d78e9f6ae1d2ff385d81bef20198e81a1f49d.jpg) +Tampered +Remove the wom + +![](images/24cb7d14e3c5b6e9bda8be532b5ae26396a8f43439fc87fb50a93158f37d47fe.jpg) +Our Predicted mask + +![](images/fce68ffe6804dba46fa7e6a3f4ca39cb656916e0619f82e81f1c2852be05af92.jpg) + +![](images/42a65b5e23bbd3bfacf3a62b020c01fa03d6a03fcda8805c156fd237dce7fe20.jpg) + +![](images/d76bc2f2c1f9c4f6a9ffff0b7f1ef5942951015effd3e05060def371d4432375.jpg) + +![](images/46bc5d2965fafcc79eef9aeb40a90a3555cf92ae32dd46f5139ed56306d6eebe.jpg) +Change the cloth + +![](images/66b03e5e20470d096f9996835dc598c65dd463d3bb1e22790cb9337a650226d3.jpg) + +![](images/a7a50709d76a5eefaaf2fb72378462e3f5161ccac2dd89fb63d669e4b75e2c53.jpg) + +![](images/ef46a82a571e641ddad2420ae8d83655029caf679edbd3284c710e5a17c7323a.jpg) +Tamper with AI-generated images + +![](images/971d94980f5a423b41c9e7dc2fbc6de91c45d50ad53429365bfe38dc41235a4e.jpg) +Text Forgery +Figure 1. We propose a versatile proactive forensics framework EditGuard. The application scenario is shown on the left, wherein users embed invisible watermarks to their images via EditGuard in advance. If suffering tampering, users can defend their rights via the tampered areas and copyright information provided by EditGuard. Some supported tampering methods (marked in blue) and localization results of EditGuard are placed on the right. Our EditGuard can achieve over $95\%$ localization precision and nearly $100\%$ copyright accuracy. + +![](images/d90b5b7215fe769f777d32aa4267f0e7d64858149ce81f84aae3ff33bf217968.jpg) +es, "A volcano' + +![](images/1d62c40a83bc886c779407c8a053fdec1369912eb3c5ce8d96fafff48efd4310.jpg) +Original +AI Fill + +![](images/0bbb5738cc0dbc73621ea30a9504e05c302e3468022b90984db1ae12173ed25e.jpg) +Tampered +Picasso painting + +![](images/ef5e8fa582c6aeb0ef7f3babbcd898c34d8bc2cbd5ffe9345f140dbb7834ad80.jpg) +Our Predicted mask +on the wall + +![](images/b55ae49daacc5bf74a9f498b5e1af76d43b516d77223e6a6e30fbaea5bbc93b4.jpg) + +![](images/4239f38f67c4141fd0b9043c6a100f4452549f3da6935110b6d78f4c63ea9fad.jpg) +Replace background + +![](images/ba7f575d98c5ab4ce391c376226cc462b2a4cf5a332b3ffc42477fc7c3c6bf7a.jpg) +ground, "A man + +![](images/2a5384c2a631ee76fd571960084ccd4e9df042886f5ae6724e33eb9f29d3814c.jpg) + +![](images/d7a5ce36ace3683d775141aa1d2560a26c3f3c151c5effc0dfe4b62da163ec0c.jpg) +in an office +"A straw hat" + +![](images/c852226a6f6691fc2c0bbb154d0f9f2addaeb99d34c00ed03a87681a891a5cea.jpg) +Secondary creation of artworks + +![](images/0a54d13aa873acdbc59699f323546ea56bee7fde1c33c7aa9140f6072c2bc195.jpg) +Swap Face + +![](images/8863579b2f6447b29ef99f80bc578a2ed5b9550e69081aeddd5b7240ffc19c45.jpg) + +# Abstract + +In the era of AI-generated content (AIGC), malicious tampering poses imminent threats to copyright integrity and information security. Current deep image watermarking, while widely accepted for safeguarding visual content, can only protect copyright and ensure traceability. They fall short in localizing increasingly realistic image tampering, potentially leading to trust crises, privacy violations, and legal disputes. To solve this challenge, we propose an innovative proactive forensics framework EditGuard, to unify copyright protection and tamper-agnostic localization, especially for AIGC-based editing methods. It can offer a meticulous embedding of imperceptible watermarks and precise decoding of tampered areas and copyright information. Leveraging our observed fragility and locality of image-into-image steganography, the realization of Edit + +Guard can be converted into a united image-bit steganography issue, thus completely decoupling the training process from the tampering types. Extensive experiments verify that our EditGuard balances the tamper localization accuracy, copyright recovery precision, and generalizability to various AIGC-based tampering methods, especially for image forgery that is difficult for the naked eye to detect. + +# 1. Introduction + +Owing to the advantageous properties of diffusion models and the bolstering of extensive datasets, AI-generated content (AIGC) models like DALL-E 3 [14], Imagen [52], and Stable Diffusion [51], can produce lifelike and wondrous images, which brings great convenience to photography enthusiasts and image editors. Nonetheless, the remarkable capabilities of these models come with a double-edged sword, presenting new challenges in copyright protection and information security. The efficiency of image manipulation [43, 47, 49, 51, 55, 68, 72] has blurred the + +line between fact and forgery, ushering in myriad security and legal concerns. For instance, artistic works are vulnerable to malicious tampering or unauthorized AI-facilitated recreations, making it challenging to protect their original creations [13]. Meanwhile, forged images may be spread online or used as court evidence, causing adverse effects on public opinion, ethical issues, and social stability. + +Given the challenges of preventing image tampering from the source, image watermarking has become a consensus for proactive forensics [54]. However, most prevalent forensic image watermarking [12, 15, 44, 62] still focus on detecting image authenticity or protecting image copyrights, but fall short when it comes to advanced demands, such as localizing tampered areas. Tamper localization facilitates an evaluation of the severity of the image manipulation, and provides an understanding of the intent of attackers, potentially allowing for the partial reuse of the tampered images. However, most passive forensics methods such as previous black-box localization networks [8, 56, 61] tend to seek anomalies like artifacts or flickers in images but struggle to detect more realistic textures and more advanced AIGC models. Moreover, they inevitably need to introduce tampered data during the training and focus solely on specific "CheapFake" tampering like slicing and copy-and-paste [8, 45, 56], or on "DeepFake" targeting human faces [2, 50], restricted in generalizability. Thus, it is vital to develop an integrated watermarking framework that unites tamper-agnostic localization and copyright protection. + +To clarify our task scope, we re-emphasize the definition of dual forensics tasks as illustrated in Fig. 1: (1) Copyright protection: "Who does this image belong to?" We aspire to accurately retrieve the original copyright of an image, even suffering various tampering and degradation. (2) Tamper localization: "Where was this image manipulated?" We aim to precisely pinpoint the tampered areas, unrestrained by specific tampering types. To the best of our knowledge, no existing method accomplishes these two tasks simultaneously, while maintaining a balance of high precision and extensive generalizability. + +To address this urgent demand, we propose a novel proactive forensics framework, dubbed EditGuard, to protect copyrights and localize tamper areas for AIGC-based editing methods. Specifically, drawing inspiration from our observed locality and fragility of image-into-image (I2I) steganography and inherent robustness of bit-into-image steganography, we can transform the realization of EditGuard into a joint image-bit steganography issue, which allows the training of EditGuard to be entirely decoupled from tampering types, thereby endowing it with exceptional generalizability and locate tampering in a zero-shot manner. In a nutshell, our contributions are as follows: + +(1) We present the first attempt to design a deep versatile proactive forensics framework EditGuard for univer + +sal tamper localization and copyright protection. It embeds dual invisible watermarks into original images and accurately decodes tampered areas and copyright information. + +(2) We have observed the fragility and locality of I2I steganography and innovatively convert the solution of this dual forensics task into training a united Image-Bit Steganography Network (IBSN), and utilize the core components of IBSN to construct EditGuard. +(3) We introduce a prompt-based posterior estimation module to enhance the localization accuracy and degradation robustness of the proposed framework. +(4) The effectiveness of our method has been verified on our constructed dataset and classical benchmarks. Compared to other competitive methods, our approach has notable merits in localization precision, generalization abilities, and copyright accuracy without any labeled data or additional training required for specific tampering types. + +# 2. Related works + +# 2.1. Tamper Localization + +Prevalent passive image forensic techniques have focused on localizing specific types of manipulations [25, 33, 34, 53, 61, 75]. Meanwhile, some universal tamper localization methods [5, 23, 31, 35, 63, 65-67] also tend to explore artifacts and anomalies in tampered images. For instance, MVSS-Net [8] employed multi-view feature learning and multi-scale supervision to jointly exploit boundary artifacts and the noise view of images. OSN [61] proposed a novel robust training scheme to address the challenges posed by lossy operations. Trufor [17] used a learned noise-sensitive fingerprint and extracted both high-level and low-level traces via transformer-based fusion. HiFi-Net [18] utilized multi-branch feature extractor and localization modules for both CNN-synthesized and edited images. SAFL-Net [56] constrained a feature extractor to learn semantic-agnostic features with specific modules and auxiliary tasks. However, the above-mentioned passive localization methods are often limited in terms of generalization and localization accuracy, which usually work on known tampering types that have been trained. Although MaLP [2] used template matching for proactive tamper localization, it still requires a large number of forgery images and cannot fully decouple the network training from the tamper types. + +# 2.2. Image Watermarking + +Image watermarking can be broadly used for the verification, authenticity, and traceability of images. Although traditional fragile watermarking [6, 24, 29, 36, 37, 48] can also achieve block-wise tamper localization, their localization accuracy and flexibility are unsatisfactory. How to realize joint pixel-level tamper localization and copyright protection remains largely unexplored. Owing to the development + +![](images/f39961573342ac5698699c406fd9e570e66dc33b26fe8a12add7aac800fee9d6.jpg) +Figure 2. Fragility and locality of I2I steganography. The first line shows that when $\mathbf{X}_{con}^{\prime}$ is changed, $\hat{\mathbf{X}}_{sec}$ will also be fragility demaged. The second line plots the attribution maps $\frac{1}{|\mathbb{S}|}\sum_{(i,j)\in \mathbb{S}}\frac{\partial\hat{\mathbf{X}}_{sec}[i,j]}{\partial\mathbf{X}_{con}^{\prime}}$ of five point sets $\mathbb{S}$ (marked by $①-⑤$ ) in $\hat{\mathbf{X}}_{sec}$ . We observed that $\hat{\mathbf{X}}_{sec}$ almost only has a strong response at the corresponding positions of $\mathbf{X}_{con}^{\prime}$ and its neighborhoods. + +of deep learning, deep image watermarking has attracted increased attention. For instance, HiDDeN [74] firstly introduced a deep encoder-decoder network to hide and recover bitstream. Moreover, many distortion layers such as differentiable JPEG and screen-shooting [1, 10, 40, 62] were designed to enhance its robustness. Flow-based models [11, 44] were also utilized to further improve the fidelity of container images. Recently, researchers [7, 12, 27, 59, 73] have designed specialized watermarking mechanisms for large-scale image generation models [51], to merge watermarking into the generation process. However, these deep watermarking methods have a singular function and cannot accurately localize the tampered areas. + +# 3. Overall Framework of EditGuard + +# 3.1. Motivation + +Challenges of existing methods: (1) How to equip existing watermark methods, which are solely for copyright protection, with the ability to localize tampering is the crux of EditGuard. We will solve it via the framework design in Sec. 3.2. (2) Most previous tamper localization methods inevitably introduce specific tampering data during network training but tend to raise generalization concerns in unknown tampering types, which will be addressed in Sec. 3.3. + +Our observations: Fortunately, we observed that imageinto-image (I2I) steganography exhibits distinct fragility and locality, possessing great potential to address these issues. Concretly, I2I steganography [3, 28, 42, 46, 64, 69] aims to hide a secret image $\mathbf{X}_{\text{sec}}$ into a cover image $\mathbf{X}_{\text{cov}}$ to produce a container image $\mathbf{X}_{\text{con}}$ , and reveal $\hat{\mathbf{X}}_{\text{sec}}$ and $\hat{\mathbf{X}}_{\text{cov}}$ with minimal distortion from the received image $\mathbf{X}_{\text{con}}^{\prime}$ . We discover that when $\mathbf{X}_{\text{con}}^{\prime}$ undergoes significant alteration compared to $\mathbf{X}_{\text{con}}$ , $\hat{\mathbf{X}}_{\text{sec}}$ will also be damaged and generate artifacts (the first row of Fig. 2), which is called fragility. Furthermore, we notice that the artifacts in $\hat{\mathbf{X}}_{\text{sec}}$ are almost pixel-level corresponding to the changes in $\mathbf{X}_{\text{con}}^{\prime}$ relative to $\mathbf{X}_{\text{con}}$ , which is called locality. To demonstrate this locality, + +we select five $7 \times 7$ point sets on $\hat{\mathbf{X}}_{\text {sec }}$ and calculated their attribution maps with respect to $\mathbf{X}_{\text {con }}^{\prime}$ . As plotted in the second row of Fig. 2, $\hat{\mathbf{X}}_{\text {sec }}$ only exhibits a strong response at the corresponding locations of $\mathbf{X}_{\text {con }}^{\prime}$ and their immediate vicinity, almost irrelevant to other pixels. These properties inspire us to treat $\mathbf{X}_{\text {sec }}$ as a special localization watermark and embed it within existing watermarking frameworks. + +# 3.2. Framework Design and Forensics Process + +To realize united tamper localization and copyright protection, EditGuard is envisioned to embed both a 2D localization watermark and a 1D copyright traceability watermark into the original image in an imperceptible manner, which allows the decoding end to obtain the copyright of the images and a binary mask reflecting tampered areas. However, designing such a framework needs to solve the compatibility issue of two types of watermarks. + +(1) Local vs. Global: The localization watermark is required to be hidden in the corresponding pixel positions of the original image, while the copyright watermark should be unrelated to spatial location and embedded in the global area redundantly. (2) Semi-fragile vs. Robust: The desired attribute of the localization watermark is semi-fragile, which means it is fragile to tampering but robust against some common degradations (such as Gaussian noise, JPEG compression, and Poisson noise) during network transmission. However, the copyright should be extracted nearly losslessly, irrespective of tampering or degradation. + +To address the two pivotal conflicts, EditGuard employs a "sequential encoding and parallel decoding" structure, which comprises a dual-watermark encoder, a tamper locator, and a copyright extractor. As shown in Fig. 3, the dual-watermark encoder will sequentially add pre-defined localization watermark and global copyright watermark $\mathbf{w}_{cop}$ provided by users to the original image $\mathbf{I}_{ori}$ , forming the container image $\mathbf{I}_{con}$ . Our experiments have proved that parallel encoding cannot effectively add dual watermarks into images (in supplementary materials (S.M.). In contrast, sequential embedding effectively prevents cross-interference by hiding these two watermarks. Furthermore, we model the network transmission process in which the received (tampered) image $\mathbf{I}_{rec}$ is transformed from $\mathbf{I}_{con}$ as: + +$$ +\mathbf {I} _ {r e c} = \mathcal {D} \left(\mathbf {I} _ {c o n} \odot (\mathbf {1} - \mathbf {M}) + \mathcal {T} \left(\mathbf {I} _ {c o n}\right) \odot \mathbf {M}\right), \tag {1} +$$ + +where $\mathcal{T}(\cdot),\mathcal{D}(\cdot)$ and $\mathbf{M}$ respectively denote the tamper function, degradation operation, and tempered mask. Moreover, the parallel decoding processes allow us to flexibly train each branch under different levels of robustness and obtain the predicted mask $\hat{\mathbf{M}}$ via the tamper locator and traceability watermark $\hat{\mathbf{w}}_{cop}$ via the copyright extractor. We can categorize the dual forensic process of EditGuard into the following scenarios. + +Case 1: If $\hat{\mathbf{w}}_{\text{cop}} \not\simeq \mathbf{w}_{\text{cop}}$ , suspicious $\mathbf{I}_{\text{rec}}$ is either not + +![](images/86330ab304ab8ee076a9aed3f95af6473f94c3b7e495a949fe4060b129a8b924.jpg) +Figure 3. Illustration of the proposed proactive forensics framework EditGuard. The dual-watermark encoder sequentially embeds the pre-defined localization watermark and copyright watermark $\mathbf{w}_{\text{cop}}$ into the original image $\mathbf{I}_{\text{ori}}$ , generating the container image $\mathbf{I}_{\text{con}}$ . After encountering potential malicious tampering and degradation during network transmission, tampered mask $\hat{\mathbf{M}}$ and copyright information $\hat{\mathbf{W}}_{\text{cop}}$ are respectively extracted via the tamper locator and copyright extractor from the received image $\mathbf{I}_{\text{rec}}$ . + +registered in our EditGuard or underwent extremely severe global tampering, rendering it unreliable as evidence. + +- Case 2: If $\hat{\mathbf{w}}_{\text {cop }} \approx \mathbf{w}_{\text {cop }}$ and $\hat{\mathbf{M}} \not\approx \mathbf{0}$ , suspicious $\mathbf{I}_{\text {rec }}$ has undergone tampering, disqualifying it as valid evidence. Users may infer the intention of tamperers based on $\hat{\mathbf{M}}$ and decide whether to reuse parts of the image. +$\square$ Case 3: If $\hat{\mathbf{w}}_{cop} \approx \mathbf{w}_{cop}$ and $\hat{\mathbf{M}} \approx \mathbf{0}$ , $\mathbf{I}_{rec}$ remains untampered and trustworthy under the shield of EditGuard. + +# 3.3. Transform Dual Forensics into Steganography + +To realize universal and tamper-agnostic localization, we resort to our observed locality and fragility of I2I steganography. As described in Sec. 3.1, localization watermarking and tamper locator in Fig. 3 can be effectively realized via image hiding and revealing. Meanwhile, combining with the robustness of current bit-into-image steganography, copyright watermarking and extractor in Fig. 3 are achieved via bit encryption and recovery. Thus, we can convert the realization of the dual forensics framework EditGuard into a united image-bit steganography network. Our training objective is just a self-recovery mechanism, meaning it only needs to ensure the input and output of the steganography network maintain high fidelity under various robustness levels, with no need to introduce any labeled data or tampered samples. During inference, it can naturally locate tampering via simple comparisons in a zero-shot manner and extract copyright exactly. + +# 4. United Image-bit Steganography Network + +# 4.1. Network Architecture + +As plotted in Fig. 4, the proposed IBSN includes an image hiding module (IHM), a bit encryption module (BEM), a bit recovery module (BRM), and an image revealing module (IRM). First, the IHM aims to hide a localization watermark $\mathbf{W}_{loc} \in \mathbb{R}^{H \times W \times 3}$ into the original image $\mathbf{I}_{ori} \in \mathbb{R}^{H \times W \times 3}$ , resulting in an intermediate output $\mathbf{I}_{med} \in \mathbb{R}^{H \times W \times 3}$ . Subsequently, $\mathbf{I}_{med}$ is fed to the BEM for feature refinement, + +while the copyright watermark $\mathbf{w}_{cop} \in \{0,1\}^L$ is modulated into the BEM, forming the final container image $\mathbf{I}_{con} \in \mathbb{R}^{H \times W \times 3}$ . After network transmission, the BRM will faithfully reconstruct the copyright watermark $\hat{\mathbf{w}}_{cop}$ from the received container image $\mathbf{I}_{rec}$ . Meanwhile, $\mathbf{I}_{rec}$ predicts the missing information $\tilde{\mathbf{Z}}$ via the prompt-based posterior estimation and uses it as the initialization for the invertible blocks, producing $\hat{\mathbf{I}}_{ori}$ and semi-fragile watermark $\hat{\mathbf{W}}_{loc}$ . + +# 4.2. Invertible Blocks in IHM and IRM + +Given the inherent capacity of flow-based models to precisely recover multimedia information, we harness stacked invertible blocks to construct image hiding and revealing modules. The original image $\mathbf{I}_{ori} \in \mathbb{R}^{H \times W \times 3}$ and localization watermark $\mathbf{W}_{loc} \in \mathbb{R}^{H \times W \times 3}$ will undergo discrete wavelet transformations (DWT) to yield frequency-decoupled image features. We then employ enhanced additive affine coupling layers to project the original image and its corresponding localization watermark branches. The transformation parameters are generated from each other. The enhanced affine coupling layer is composed of a five-layer dense convolution block [46] and a lightweight feature interaction module (LFIM) [4]. The LFIM can enhance the non-linearity of transformations and capture the long-range dependencies with low computational cost. More details are presented in S.M. Finally, the revealed features are then transformed to the image domain via the inverse DWT. + +# 4.3. Prompt-based Posterior Estimation + +To bolster the fidelity and robustness of the image hiding and revealing module, we introduce a degradation prompt-based posteriori estimation module (PPEM). Since the encoding network tends to compress $\left[\mathbf{I}_{ori};\mathbf{W}_{loc}\right] \in \mathbb{R}^{H \times W \times 6}$ into the container image $\mathbf{I}_{con} \in \mathbb{R}^{H \times W \times 3}$ , previous methods [42, 64] typically utilized a random Gaussian initialization or an all-zero map at the decoding end to compensate for the lost high-frequency channels. Yet, our observations suggest that discarded information lurks within the + +![](images/01b5c21e4e64812cbbbcd0b7eb5a9a74b9248afa151d2966ce8537cfa107bdb0.jpg) + +![](images/8d9ee22d3b133e6ac1c8906b3431073c5a2484c614038f51aa7f79bf152441c1.jpg) +Figure 4. Illustration of the united Image-bit Steganography Network (IBSN). In the training process, we randomly sample original image $\mathbf{I}_{ori}$ , localization watermark $\mathbf{W}_{loc}$ (a natural RGB image) and copyright watermark $\mathbf{w}_{cop}$ and expect the IBSN to recover $\hat{\mathbf{I}}_{ori}$ , $\hat{\mathbf{W}}_{loc}$ and $\hat{\mathbf{w}}_{cop}$ with high fidelity. In the inference process, we use core components of the pre-trained IBSN with a mask extractor (ME) to construct our EditGuard, and pre-define a simple solid color image as a localization watermark. + +![](images/9b662d95305022ca97a793ae622e6d31d1bc933c050499a507cebfed03184167.jpg) +Figure 5. Illustration of the proposed prompt-based posterior estimation. It will dynamically fuse degraded representations and extracted features to obtain posterior mean $\hat{\mathbf{Z}} = \mathbb{E}[\mathbf{Z}|\mathbf{I}_{rec}]$ . + +edges and textures of the container image. Thus, deploying a dedicated network proves to be a more potent strategy in predicting the posterior mean of the vanished localization watermark information $\hat{\mathbf{Z}} = \mathbb{E}[\mathbf{Z}|\mathbf{I}_{rec}]$ . Specifically, as shown in Fig. 5, we stack $M$ residual blocks $\mathrm{Res}(\cdot)$ [19] and $M$ channel-wise transformer blocks $\mathrm{Trans}(\cdot)$ [70] to extract the local and non-local features $\mathbf{F}_c$ . + +$$ +\mathbf {F} _ {c} = \operatorname {T r a n s} (\operatorname {R e s} (\mathrm {D W T} (\mathbf {I} _ {r e c}))) + \operatorname {R e s} (\mathrm {D W T} (\mathbf {I} _ {r e c})). \tag {2} +$$ + +Considering that the container image is prone to various degradations during network transmission, we pre-define $N$ learnable embedding tensors as degradation prompts $\mathbf{P} = [\mathbf{P}_1,\mathbf{P}_2,\dots ,\mathbf{P}_N]$ , where $N$ denotes the number of degradation types and is set to 3. These learned prompts $\mathbf{P}$ can + +adaptively learn a diverse range of degradation representations and are integrated with the intrinsic features extracted from $\mathbf{I}_{rec}$ , enabling the proposed IBSN to handle multiple types of degradations using a single set of parameters. To better foster the interaction between the input features $\mathbf{F}_c$ and the degradation prompt $\mathbf{P}$ , the features $\mathbf{F}_c$ are passed to a global average pooling (GAP) layer, a $1\times 1$ convolution, and a softmax operator to produce a set of dynamic weight coefficients. Each degradation prompt $\mathbf{P}_i$ is combined using these dynamic coefficients $\mathbf{w}_{p\otimes i}$ and subsequently integrated via an upsampling operator $\uparrow$ and $3\times 3$ convolution to obtain the enhanced representation $\mathbf{P}_c$ . + +$$ +\mathbf {P} _ {c} = \operatorname {C o n v} _ {3 \times 3} \left(\left(\sum_ {i = 1} ^ {N} \mathbf {w} _ {p * i} \mathbf {P} _ {i}\right) _ {\uparrow}\right), \tag {3} +$$ + +where $\mathbf{w}_p = \mathrm{Softmax}\left(\mathrm{Conv}_{1\times 1}\left(\mathrm{GAP}\left(\mathbf{F}_c\right)\right)\right)$ + +Finally, we utilize a $3 \times 3$ convolution to fuse the learned degradation representation conditioned on the extracted features $\mathbf{F}_c$ to enrich the degradation-specific context, obtaining $\hat{\mathbf{Z}}$ . This process can be formulated as: + +$$ +\hat {\mathbf {Z}} = \operatorname {C o n v} _ {3 \times 3} \left(\left[ \mathbf {P} _ {c}; \mathbf {F} _ {c} \right]\right) \in \mathbb {R} ^ {\frac {H}{2} \times \frac {W}{2} \times 1 2}. \tag {4} +$$ + +# 4.4. Bit Encryption and Recovery Modules + +As shown in Fig. 4, to encode the copyright watermark $\mathbf{w}_{\text{cop}}$ into $\mathbf{I}_{\text{med}}$ , we firstly expand $\mathbf{w}_{\text{cop}} \in \{0,1\}^L$ via stacked MLPs and reshape it into several $L \times L$ message feature + +Table 1. Localization precision (F1-Score) and bit accuracy (BA) comparison with other competitive methods on [9, 16, 20, 58]. + +
MethodCAISAv1 [9]Coverage [58]NIST16 [16]Columbia [20]
F1BA(%)F1BA(%)F1BA(%)F1BA(%)
ManTraNet [63]0.320-0.486-0.225-0.650-
SPAN [22]0.169-0.428-0.363-0.873-
CAT-Net v2 [32]0.852-0.582-0.417-0.923-
OSN [61]0.676-0.472-0.449-0.836-
MVSS-Net [8]0.650-0.659-0.372-0.781-
PSCC-Net [39]0.670-0.615-0.210-0.760-
TruFor [17]0.822-0.735-0.470-0.914-
EditGuard (Ours)0.95499.910.9551000.91199.880.98899.93
+ +![](images/ac93b3cda087d4ec26260b5d1f14f6c77c235ad6139be81c5a78ef10a189739f.jpg) +Figure 6. Localization precision comparisons of our EditGuard and competitive methods [8, 39, 61] on four classical benchmarks. + +maps. Meanwhile, $\mathbf{I}_{med}$ is fed to a U-style feature enhancement network to extract features of each downsampling and upsampling layer. Finally, the message features will be upscaled and integrated with multi-level image features via the fusion mechanism [21, 62], achieving the modulation of bit-image information. In the decoding end, $\mathbf{I}_{rec}$ is fed into a U-shaped sub-network and downsampled to a size of $L\times L$ . The recovered copyright watermark $\hat{\mathbf{w}}_{cop}$ is then extracted via an MLP. More details are presented in S.M.. + +# 4.5. Construct EditGuard via the IBSN + +To stabilize the optimization of the proposed IBSN, we propose a bi-level optimization strategy. Given an arbitrary original image $\mathbf{I}_{med}$ and watermark $\mathbf{w}_{cop}$ , we first train the bit encryption and recovery module via the $\ell_{2}$ loss. + +$$ +\ell_ {c o p} = \left\| \mathbf {I} _ {c o n} - \mathbf {I} _ {m e d} \right\| _ {2} ^ {2} + \lambda \left\| \hat {\mathbf {w}} _ {c o p} - \mathbf {w} _ {c o p} \right\| _ {2} ^ {2}, \quad (5) +$$ + +where $\lambda$ is set to 10. Furthermore, we freeze the weights of BEM and BRM and jointly train the IHM and IRM. Given a random original image $\mathbf{I}_{ori}$ , localization watermark $\mathbf{W}_{loc}$ and copyright watermark $\mathbf{w}_{cop}$ , the loss function is: + +$$ +\ell_ {l o c} = \left\| \hat {\mathbf {I}} _ {o r i} - \mathbf {I} _ {o r i} \right\| _ {1} + \alpha \left\| \mathbf {I} _ {c o n} - \mathbf {I} _ {o r i} \right\| _ {2} ^ {2} + \beta \left\| \hat {\mathbf {W}} _ {l o c} - \mathbf {W} _ {l o c} \right\| _ {1}, \tag {6} +$$ + +where $\alpha$ and $\beta$ are respectively set to 100 and 1. During training, we only introduced degradations to $\mathbf{I}_{con}$ without being exposed to any tampering. After acquiring a pretrained IBSN, we can construct the proposed EditGuard via the components of IBSN. As plotted in Fig. 4, the dual-watermark encoder of EditGuard is composed of IHM and BEM, which correspond to the localization and copyright watermarking in Fig. 3 respectively. The copyright extractor strictly corresponds to BRM. The tamper locator + +Table 2. Visual quality of the container image $\mathbf{I}_{con}$ and bit accuracy comparison with other pure watermarking methods. + +
MethodImage SizeM. L.PSNR (dB)SSIMNIQE(↓)BA(%)
MBRS [26]128×1283026.570.8867.219100
CIN [44]128×1283041.350.9817.17199.99
PIMoG [10]128×1283036.220.9417.11399.99
SepMark [62]128×1283035.420.9317.09599.86
EditGuard128×1283036.930.9445.56799.89
EditGuard512×5126437.770.9494.25799.95
+ +includes IRM and a mask extractor (ME). Note that we need to pre-define a localization watermark $\mathbf{W}_{loc}$ , which is shared between the encoding and decoding ends. The choice of $\mathbf{W}_{loc}$ is very general to our method. It can be any natural image or even a solid color image. Finally, by comparing the pre-defined watermark $\mathbf{W}_{loc}$ with the decoded one $\hat{\mathbf{W}}_{loc}$ , we can obtain a binary mask $\hat{\mathbf{M}}\in \mathbb{R}^{H\times W}$ : + +$$ +\hat {\mathbf {M}} [ i, j ] = \theta_ {\tau} \left(\max \left(\left| \hat {\mathbf {W}} _ {l o c} [ i, j,: ] - \mathbf {W} _ {l o c} [ i, j,: ] \right|\right)\right). \tag {7} +$$ + +where $i\in [0,H)$ and $j\in [0,W)$ . $\theta_{\tau}(z) = 1$ ( $z\geq \tau$ ). $\tau$ is set to 0.2. $|\cdot |$ is an absolute value operation. + +# 5. Experiments + +# 5.1. Implementation Details + +We trained our EditGuard via the training set of COCO [38] without any tampered data. Thus, for tamper localization, our method is actually zero-shot. The Adam [30] is used for training $250K$ iterations with $\beta_{1} = 0.9$ and $\beta_{2} = 0.5$ . The learning rate is initialized to $1\times 10^{-4}$ and decreases by half for every $30K$ iterations, with the batch size set to 4. We embed a 64-bit copyright watermark and a simple localization watermark such as a pure blue image ([R, G, B] = [0, 0, 255]) to original images. Following [8, 17, 39], F1-score, AUC, IoU, and bit accuracy are used to evaluate localization and copyright protection performance. Since no prior methods can simultaneously achieve this dual forensics, we conducted separate comparisons with tamper localization and image watermarking methods. + +# 5.2. Comparison with Localization Methods + +For a fair comparison with tamper localization methods, we conducted extensive evaluations on four classical benchmarks [9, 16, 20, 58], as reported on Tab. 1. Since EditGuard is a proactive approach, we initially embed watermarks into authentic images and then paste the tampered areas into the container images. Remarkably, even for tamper types that existing methods specialize in, the localization accuracy of EditGuard consistently outperforms the SOTA method [17] across four datasets by margins of 0.102, 0.116, 0.441, and 0.065 in F1-score without any labeled data or tampered samples required, which verifies the superiority of our proactive localization mechanism. As shown in Fig. 6, our EditGuard can precisely pinpoint pixel-level tampered areas but other methods can only produce a rough outline or are only effective in some cases. Mean- + +Table 3. Comparison with other competitive tamper localization methods under different AIGC-based editing methods. Note that $\dagger$ denotes the network finetuned in our constructed AGE-Set-C. + +
MethodStable Diffusion Inpaint [51]Controlnet [72]SDXL [49]RePaint [43]Lama [57]FaceSwap [60]
F1AUCIoUBA(%)F1AUCIoUBA(%)F1AUCIoUBA(%)F1AUCIoUBA(%)F1AUCIoUBA(%)F1AUCIoUBA(%)
MVSS-Net [8]0.1780.4880.103-0.1780.4920.103-0.0370.5030.028-0.1040.5460.082-0.0240.5050.022-0.2850.6120.192-
OSN [61]0.1740.4860.101-0.1910.6440.110-0.2000.7550.118-0.1830.6440.105-0.1700.4300.099-0.3080.7910.171-
PSCC-Net [39]0.1660.5010.112-0.1770.5650.116-0.1890.7040.115-0.1400.4690.109-0.1320.3290.104-0.1570.3460.180-
IML-VIT [45]0.2130.5960.135-0.2000.5760.128-0.2210.6030.145-0.1030.4970.059-0.1050.4650.064-0.1050.4650.064-
HiFi-Net [18]0.5470.7340.128-0.5420.7350.123-0.6330.8280.261-0.6810.8960.339-0.4830.7210.029-0.7810.8900.478-
MVSS-Net†[8]0.6940.9390.575-0.6780.9250.558-0.4820.8840.359-0.1850.5290.111-0.3930.8290.275-0.4590.7390.333-
EditGuard (Ours)0.9660.9710.93699.950.9680.9870.94099.960.9650.9890.93699.960.9670.9770.93899.950.9650.9690.93499.950.8960.9430.87699.86
+ +![](images/95875c5849a1cdbc6277286d2e769a7f1bd18d6ddaca9783cff1641b55880ea9.jpg) + +![](images/e42e68e144df1134327fbfcd2269b66657f377933b8c3fc5b2785224624d2386.jpg) + +![](images/fa30e311be2c9f3547a7b8c2161d1dfc5bf76dbf7f6ebb1f737811bf611765dc.jpg) + +![](images/83cb8d90245a0c0fdedf521e2b149960657dc5bae5a30eee4f340c70cf4e5e8e.jpg) + +![](images/4bd55197d7c7b580de6b9d301db508e36d01df31e4ac53d4df4977dd2649048d.jpg) + +![](images/1485a6728911b0188fb0bf86050c79bdae5b51c24a940aeb06082d5c4eb82843.jpg) + +![](images/dc9ae763a7b3ccf0c9aaae7c79651e92be21477f8d2253256501671c9512a36f.jpg) + +![](images/41c00b7712cda2d2a90ef7fd3b6fc93bb44371f993c44847f11653440f36d9d4.jpg) + +![](images/9b23195e6a3bd779476e07957bf8724e8326a4245ca6915e44f1728542337b0f.jpg) + +![](images/efb52bc2d62e35da7636a0034a91a419b7900d3bb22e746ef6b36778ea57a0f2.jpg) + +![](images/5ac68cb476e3c8b7111c8ea60d546a3b431b987f0854fb19fbea659d4803e0dd.jpg) +Figure 7. Visual results of the container image $\mathbf{I}_{con}$ and the error map of the proposed EditGuard. Here, localization and copyright watermarks are randomly selected from the dataset. + +![](images/e1e1c8519147b61a444e57c089b3f515b3409cfed275c56496039ec91406b4d5.jpg) + +![](images/450cda36558f90764ebfd9de06e5bbc54b1e1324e4813945fa25c2f9cb66e9ef.jpg) + +![](images/ef1ebd48de8354d61bbaa44b137bf3f916d54f689186b9b4e920e1288961fd98.jpg) + +![](images/a70af4c7c227f03986671dd12dd5b3528f9fc16cc7ba5282eb3468b33973902c.jpg) + +while, our bit accuracy remains over $99.8\%$ while all other methods can not realize effective copyright protection. + +# 5.3. Comparison with Watermarking Methods + +To evaluate the visual quality of $\mathbf{I}_{con}$ , we compared EditGuard with other watermarking methods on $1K$ testing images from COCO [38] under the tampering of stable diffusion inpaint [51]. For a fair comparison, we also retrained our EditGuard on $128\times 128$ original images and 30 bits. Tab. 2 reports that the fidelity of our container image far surpasses that of SepMark [62], PIMoG [10], and MBRS [26] but is slightly inferior to CIN [44]. Meanwhile, our method exhibits the best performance in perceptual quality measures like NIQE. As shown in Fig. 7, dual-watermarked images do not have noticeable artifacts and noise, making them imperceptible to the human eyes. When suffer malicious tampering, our method outperforms SepMark and is very close to PIMoG and CIN in bit accuracy. Note that other competitive methods only hide 30 bits, with a capacity of $30 / (128\times 128)$ . In contrast, our EditGuard hides both an RGB localization watermark and a 1D copyright watermark, with a capacity far greater than $30 / (128\times 128)$ . Here, we do not claim to achieve the best visual quality and bit accuracy, but just to demonstrate that our method is comparable to the current image watermarking methods. + +# 5.4. Extension to AIGC-based Editing Methods + +Dataset Preparation: We constructed a dataset tailored for AIGC Editing methods, dubbed AGE-Set, comprising two sub-datasets. The first AGE-Set-C is a batch-processed coarse tamper dataset. Its original images are sourced from + +![](images/a3d698dd3b7233fb982193bec98a5b9b6dc00da38e36f29d78b369f7739469f6.jpg) + +![](images/2aa366912735e6902300754de3efbc2c7f8356075cae02d1740b9f8804c9efc4.jpg) + +![](images/3fdedada16bcdab9164cd71199ca49fac996c152395eccc4c710bbba918fd5a0.jpg) + +![](images/816a0d20ac6825bea0b8fee43275fcdc5903d35dac32e6b5462be513437c2261.jpg) + +![](images/fc4edcaef943973475910003308f05107398737bad26fd2de14f34d2c22e4594.jpg) + +![](images/4738da0e0397918d0663e84d28ad244b867a07e02fa7d91b40baaf698622d151.jpg) + +![](images/d6bc9a0a9ad2fc1b7d3b3810c9af56a31f59f9c54be1dd9981caaefe5315c94e.jpg) + +![](images/a79116708e6ace07e34d9cd9b405a64ae2cde5d10c4220e84c866048561cda8a.jpg) + +![](images/97136b7504c476d3f733252490dc8f28d6d237f3d713b80182826d44d2576899.jpg) + +![](images/61921a5bab7ef5eaabb3e6a77529d830d0a262dd2ff06037c2b592da94a9a45d.jpg) + +![](images/3de38109b2947ebb0cfec34afd0a8030d377a8e777ee835ed30dd6d6153abb94.jpg) + +![](images/978c7704ed9dbc9ad97cef580bcb0f5ba4e75af3c2acc01a247fc8f5a54156fe.jpg) + +![](images/c1b99e11191fbba0ca9ba1cc538d606061a288535ce85654b7b5054d7db1d2e4.jpg) + +![](images/d0ab8577cabc01bcd57112c4e17f47b8c1868b5c4403a245f1ce02858322b3d4.jpg) + +![](images/fb07c2ec7d2fad52ba53093c3e6d0e7fc668b4719e15a082e3ace33a0ffe1804.jpg) + +![](images/43aebf93db511ac8ad02cd3c7a18779ac604d36214cb16e54407fd3242cf9acf.jpg) +Figure 8. Localization performance comparisons of our EditGuard and other methods [8, 18, 39, 61] on our constructed AGE-Set-C. + +![](images/64aebf924695532abcc878f1d7ff14b5c84a341fdd4e1781705592337fc5bb07.jpg) + +![](images/510f586af59a8dd1331cfadacf30af9acb24d8412991393df83d34fa7cbb50bd.jpg) + +![](images/1961807dc7f2143728f3e9bd78759d2a60e337895b39be63ca524b3e73376d08.jpg) + +![](images/45c47d0401b8e85bf542c59842412290bef15467881628beb41afebe3bf1039d.jpg) + +COCO 2017 [38] and CelebA [41], containing $30K$ training images and $1.2K$ testing images. We used some SOTA editing methods such as Stable Diffusion Inpaint [51], ControlNet [72], SDXL [49] to manipulate images with the prompt to be "None", and employed some unconditional methods like Repaint [43], Lama [57], and Faceswap [60]. Note that we only use the tampered data to train other methods, not our EditGuard. The second sub-dataset AGE-Set-F includes 100 finely edited images. It is edited manually via some sophisticated software such as SD-Web-UI, Photoshop, and Adobe Firefly. These AIGC-based editing methods can achieve a good fusion of the tampered and unchanged areas, making it hard for the naked eye to catch artifacts. More details are presented in S.M.. + +AGE-Set-C: Tab. 3 presents the comparison of our EditGuard and some SOTA tamper localization methods [8, 18, 39, 45, 61]. We observe that the F1-scores of other passive forensic methods are generally lower than 0.7 when applied to AGE-Set-C. Meanwhile, even when we try our best to finetune MVSS-Net using AGE-Set-C, the accuracy of MVSS-Net remains unsatisfactory, and they exhibit catastrophic forgetting across various tamper methods. In contrast, our method can guarantee an F1-score and AUC of over $95\%$ , maintaining around $90\%$ IOU, regardless of tampering types. As shown in Fig. 8, our EditGuard can accurately capture these imperceptible tampering traces produced by AIGC-based editing methods, but other methods are almost ineffective. Moreover, our EditGuard can effectively recover copyright information with a bit accuracy exceeding $99.8\%$ . Noting that none of the comparison methods offer copyright protection capabilities. + +AGE-Set-F: To further highlight the practicality of our EditGuard, we conducted subjective comparisons with other methods on the meticulously tampered AGE-Set-F. + +![](images/bce070c0cef5d5ad87cc6a09e540d2e125882f389da237eb7590e82583cf7c0a.jpg) +Figure 9. Localization precision comparisons of our EditGuard and other competitive methods on the meticulously tampered AGE-Set-F. + +Table 4. Localization and bit recovery performance of our Edit-Guard and MVSS-Net† [8] under different levels of degradations. + +
MethodsMetricsCleanGaussian NoiseJPEGPoisson
σ=1σ=5Q=70Q=80Q=90
MVSS-Net†[8]F10.6940.6440.6190.4580.5070.5580.652
BA(%)-------
EditGuard (Ours)F10.9660.9370.9320.9200.9200.9250.943
BA(%)99.9599.9499.3797.6998.1698.2399.91
+ +The tamper types in this subset did not appear in the training set. As shown in Fig. 9, when faced with real-world tampering, even the most powerful tamper localization methods almost entirely fail. This is due to their mechanisms to look for image artifacts and explore instance-wise semantic information. However, our EditGuard, which locates tampered masks via the natural fragility and locality of I2I steganography, can still clearly annotate the tampered area. + +# 5.5. Robustness Analysis + +As shown in Tab. 4, we conducted robustness analysis on the tampering of "Stable Diffusion Inpaint" [51] under Gaussian noise with $\sigma = 1$ and 5, JPEG compression with $Q = 70$ , 80 and 90, and Poisson noise with $\alpha = 4$ [71]. We observed that our method still maintains a high localization accuracy (F1-score $>0.9$ ) and bit accuracy with a very slight performance decrease under various levels of degradations, while MVSS-Net† [8] exhibits a noticeable performance degradation compared to its results in clean conditions. It is attributed to our prompt-based estimation that can effectively learn the degradation representation. + +# 5.6. Ablation Study + +To verify the effectiveness of each component of the Edit-Guard, we conducted ablation studies on bi-level optimization (BO), lightweight feature interaction module (LFIM), transformer block (TB), and prompt-based fusion (PF) under the tampering of "Stable Diffusion Inpaint". As listed in Tab. 5, without BO, the joint training of all components + +Table 5. Abalation studies on the core components of EditGuard. + +
CaseDegradation Type D(·)PFTBLFIMBOF1AUCIoUBA(%)
(a)Clean✓✓✓---49.17
(b)Clean✓✓✓0.9500.9600.90499.73
(c)Clean✓✓✓0.9570.9660.92799.51
(d)Random Degradations✓✓✓0.9030.9330.84199.12
OursClean✓✓✓✓0.9660.9710.93699.95
Random Degradations✓✓✓✓0.9380.9640.88799.36
+ +cannot converge effectively, resulting in bit accuracy that is close to random guessing. Without LFIM and TB, the IoU of EditGuard will suffer 0.032 and 0.009 declines since these two modules can better perform feature fusion. Without PF, the robustness of the EditGuard will significantly decline. We observed that the F1/AUC/IoU of our method far surpasses that of case (d) by 0.035/0.031/0.046 under "Random Degradations", which indicates that the PF effectively enables a single network to support watermark recovery under various degradations. "Random Degradations" denotes that we randomly set the $\mathcal{D}(\cdot)$ to various levels of Gaussian noise, Poisson noise, and JPEG compression. + +# 6. Conclusion + +We present the first attempt to design a deep versatile watermarking mechanism EditGuard. It enhances the credibility of images by embedding imperceptible localization and copyright watermarks, and decoding accurate copyright information and tampered areas, making it a reliable tool for artistic creation and legal forensic analysis. In the future, we will focus on improving the robustness of EditGuard and strive not only to offer pixel-wise localization results but also to provide semantic-wise outcomes. Additionally, we plan to further expand EditGuard to a broader range of modalities and applications, including video, audio, and 3D scenes. Our efforts at information authenticity serve not only the AIGC industry, but the trust in our digital world, ensuring that every pixel tells the truth and the rights of each individual are safeguarded. + +# References + +[1] Mahdi Ahmadi, Alireza Norouzi, Nader Karimi, Shadrokh Samavi, and Ali Emami. Redmark: Framework for residual diffusion watermarking based on deep networks. Expert Systems with Applications, 146:113157, 2020. 3 +[2] Vishal Asnani, Xi Yin, Tal Hassner, and Xiaoming Liu. Malp: Manipulation localization using a proactive scheme. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 2 +[3] Shumeet Baluja. Hiding images in plain sight: Deep steganography. In Proceedings of the International Conference on Neural Information Processing Systems (NeurIPS), 2017. 3 +[4] Liangyu Chen, Xiaojie Chu, Xiangyu Zhang, and Jian Sun. Simple baselines for image restoration. In Proceedings of the European Conference on Computer Vision (ECCV), 2022. 4 +[5] Xinru Chen, Chengbo Dong, Jiaqi Ji, Juan Cao, and Xirong Li. Image manipulation detection by multi-view multi-scale supervision. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 2 +[6] Baotian Cheng, Rongrong Ni, and Yao Zhao. A refining localization watermarking for image tamper detection and recovery. In 2012 IEEE 11th International Conference on Signal Processing, pages 984-988, 2012. 2 +[7] Yingqian Cui, Jie Ren, Han Xu, Pengfei He, Hui Liu, Lichao Sun, and Jiliang Tang. Diffusionshield: A watermark for copyright protection against generative diffusion models. arXiv preprint arXiv:2306.04642, 2023. 3 +[8] Chengbo Dong, Xinru Chen, Ruohan Hu, Juan Cao, and Xirong Li. Mvss-net: Multi-view multi-scale supervised networks for image manipulation detection. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(3):3539-3553, 2022. 2, 6, 7, 8 +[9] Jing Dong, Wei Wang, and Tieniu Tan. Casia image tampering detection evaluation database. In Proceedings of the IEEE China Summit and International Conference on Signal and Information Processing (ChinaSIP), 2013. 6 +[10] Han Fang, Zhaoyang Jia, Zehua Ma, Ee-Chien Chang, and Weiming Zhang. Pimog: An effective screen-shooting noise-layer simulation for deep-learning-based watermarking network. In Proceedings of the 30th ACM International Conference on Multimedia (MM), 2022. 3, 6, 7 +[11] Han Fang, Yupeng Qiu, Kejiang Chen, Jiyi Zhang, Weiming Zhang, and Ee-Chien Chang. Flow-based robust watermarking with invertible noise layer for black-box distortions. In Proceedings of the AAAI Conference on Artificial Intelligence (AAAI), 2023. 3 +[12] Pierre Fernandez, Guillaume Couairon, Hervé Jégou, Matthijs Douze, and Teddy Furon. The stable signature: Rooting watermarks in latent diffusion models. Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2023. 2, 3 +[13] Anirban Ghoshal. Artists lose first copyright battle in the fight against ai-generated images. https://www.computerworld.com/article/3709691/artists-lose-first-copyright-battle-in + +the-fight-against-ai-generated-images. +html,2023.2 +[14] Gabriel Goh, James Betker, Li Jing, Aditya Ramesh, et al. Improving image generation with better captions. https://cdn.openai.com/papers/dall-e-3.pdf, 2023.1 +[15] Google DeepMind. Synthid. https://deepmind.google/technologies/synthid/, 2023.2 +[16] Haiying Guan, Mark Kozak, Eric Robertson, Yooyoung Lee, Amy N Yates, Andrew Delgado, Daniel Zhou, Timothee Kheyrkhah, Jeff Smith, and Jonathan Fiscus. Mfc datasets: Large-scale benchmark datasets for media forensic challenge evaluation. In Proceedings of the IEEE Winter Applications of Computer Vision Workshops (WACVW), 2019. 6 +[17] Fabrizio Guillaro, Davide Cozzolino, Avneesh Sud, Nicholas Dufour, and Luisa Verdoliva. Trufor: Leveraging all-round clues for trustworthy image forgery detection and localization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 2, 6 +[18] Xiao Guo, Xiaohong Liu, Zhiyuan Ren, Steven Grosz, Iacopo Masi, and Xiaoming Liu. Hierarchical fine-grained image forgery detection and localization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 2, 7 +[19] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 5 +[20] Yu-Feng Hsu and Shih-Fu Chang. Detecting image splicing using geometry invariants and camera characteristics consistency. In Proceedings of the IEEE International Conference on Multimedia and Expo (ICME), 2006. 6 +[21] Jie Hu, Li Shen, and Gang Sun. Squeeze-and-excitation networks. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 6 +[22] Xuefeng Hu, Zhihan Zhang, Zhenye Jiang, Syomantak Chaudhuri, Zhenheng Yang, and Ram Nevatia. Span: Spatial pyramid attention network for image manipulation localization. In Proceedings of the European Conference on Computer Vision (ECCV), 2020. 6 +[23] Xiaoxiao Hu, Qichao Ying, Zhenxing Qian, Sheng Li, and Xinpeng Zhang. Draw: Defending camera-shooted raw against image manipulation. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2023. 2 +[24] Nasir N Hurrah, Shabir A Parah, Nazir A Loan, Javaid A Sheikh, Mohammad Elhoseny, and Khan Muhammad. Dual watermarking framework for privacy protection and content authentication of multimedia. Future generation computer Systems, 94:654-673, 2019. 2 +[25] Ashraful Islam, Chengjiang Long, Arslan Basharat, and Anthony Hoogs. Doa-gan: Dual-order attentive generative adversarial network for image copy-move forgery detection and localization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 2 + +[26] Zhaoyang Jia, Han Fang, and Weiming Zhang. Mbrs: Enhancing robustness of dnn-based watermarking by minibatch of real and simulatedJPEG compression. In Proceedings of the 29th ACM International Conference on Multimedia (MM), 2021. 6, 7 +[27] Zhengyuan Jiang, Jinghuai Zhang, and Neil Zhenqiang Gong. Evading watermark based detection of ai-generated content. arXiv preprint arXiv:2305.03807, 2023. 3 +[28] Junpeng Jing, Xin Deng, Mai Xu, Jianyi Wang, and Zhenyu Guan. Hinet: Deep image hiding by invertible network. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 3 +[29] Asra Kamili, Nasir N Hurrah, Shabir A Parah, Ghulam Mohiuddin Bhat, and Khan Muhammad. Dwfcat: Dual watermarking framework for industrial image authentication and tamper localization. IEEE Transactions on Industrial Informatics, 17(7):5108-5117, 2020. 2 +[30] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6 +[31] Myung-Joon Kwon, In-Jae Yu, Seung-Hun Nam, and Heung-Kyu Lee. Cat-net: Compression artifact tracing network for detection and localization of image splicing. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), 2021. 2 +[32] Myung-Joon Kwon, Seung-Hun Nam, In-Jae Yu, Heung-Kyu Lee, and Changick Kim. LearningJPEG compression artifacts for image manipulation detection and localization. International Journal of Computer Vision, 130(8):1875-1895, 2022.6 +[33] Haodong Li and Jiwu Huang. Localization of deep inpainting using high-pass fully convolutional network. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2019. 2 +[34] Yuanman Li and Jiantao Zhou. Fast and effective image copy-move forgery detection via hierarchical feature point matching. IEEE Transactions on Information Forensics and Security, 14(5):1307-1322, 2018. 2 +[35] Yue Li, Dong Liu, Houqiang Li, Li Li, Zhu Li, and Feng Wu. Learning a convolutional neural network for image compact-resolution. IEEE Transactions on Image Processing, 28(3): 1092-1107, 2018. 2 +[36] Siau-Chuin Liew, Siau-Way Liew, and Jasni Mohd Zain. Tamper localization and lossless recovery watermarking scheme with roi segmentation and multilevel authentication. Journal of digital imaging, 26:316-325, 2013. 2 +[37] Chia-Chen Lin, Ting-Lin Lee, Ya-Fen Chang, Pei-Feng Shiu, and Bohan Zhang. Fragile watermarking for tamper localization and self-recovery based on ambtc and vq. Electronics, 12(2):415, 2023. 2 +[38] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Proceedings of the European Conference on Computer Vision (ECCV), 2014. 6, 7 +[39] Xiaohong Liu, Yaojie Liu, Jun Chen, and Xiaoming Liu. Pssc-net: Progressive spatio-channel correlation network for + +image manipulation detection and localization. IEEE Transactions on Circuits and Systems for Video Technology, 32 (11):7505-7517, 2022. 6, 7 +[40] Yang Liu, Mengxi Guo, Jian Zhang, Yuesheng Zhu, and Xiaodong Xie. A novel two-stage separable deep learning framework for practical blind watermarking. In Proceedings of the ACM International Conference on Multimedia (MM), 2019. 3 +[41] Ziwei Liu, Ping Luo, Xiaogang Wang, and Xiaou Tang. Deep learning face attributes in the wild. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), 2015. 7 +[42] Shao-Ping Lu, Rong Wang, Tao Zhong, and Paul L Rosin. Large-capacity image steganography based on invertible neural networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 3, 4 +[43] Andreas Lugmayr, Martin Danelljan, Andres Romero, Fisher Yu, Radu Timofte, and Luc Van Gool. Repaint: Inpainting using denoising diffusion probabilistic models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 1, 7 +[44] Rui Ma, Mengxi Guo, Yi Hou, Fan Yang, Yuan Li, Huizhu Jia, and Xiaodong Xie. Towards blind watermarking: Combining invertible and non-invertible mechanisms. In Proceedings of the ACM International Conference on Multimedia (MM), 2022. 2, 3, 6, 7 +[45] Xiaochen Ma, Bo Du, Xianggen Liu, Ahmed Y Al Hammadi, and Jizhe Zhou. Iml-vit: Image manipulation localization by vision transformer. arXiv preprint arXiv:2307.14863, 2023. 2, 7 +[46] Chong Mou, Youmin Xu, Jiechong Song, Chen Zhao, Bernard Ghanem, and Jian Zhang. Large-capacity and flexible video steganography via invertible neural network. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 3, 4 +[47] Chong Mou, Xintao Wang, Liangbin Xie, Yanze Wu, Jian Zhang, Zhongang Qi, Ying Shan, and Xiaohu Qie. T2i-adapter: Learning adapters to dig out more controllable ability for text-to-image diffusion models. In AAAI Conference on Artificial Intelligence (AAAI), 2024. 1 +[48] Neena Raj NR and R Shreelekshmi. Fragile watermarking scheme for tamper localization in images using logistic map and singular value decomposition. Journal of Visual Communication and Image Representation, 85:103500, 2022. 2 +[49] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952, 2023. 1, 7 +[50] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022. 2 +[51] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of + +the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 1, 3, 7, 8 +[52] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, Seyed Kamyar Seyed Ghasemipour, Raphael Gontijo-Lopes, Burcu Karagol Ayan, Tim Salimans, Jonathan Ho, David J. Fleet, and Mohammad Norouzi. Photorealistic text-to-image diffusion models with deep language understanding. In Proceedings of the International conference on Neural Information Processing Systems (NeurIPS), 2022. 1 +[53] Ronald Salloum, Yuzhuo Ren, and C-C Jay Kuo. Image splicing localization using a multi-task fully convolutional network (mfcn). Journal of Visual Communication and Image Representation, 51:201-209, 2018. 2 +[54] Deepa Shivaram. The white house and big tech companies release commitments on managing ai. https://www.nprillinois.org/2023-07-21/the-white-house-and-big-tech-companies-release-commitments-on-managing-ai, 2023.2 +[55] Xiaopeng Sun, Weiqi Li, Zhenyu Zhang, Qiufang Ma, Xuhan Sheng, Ming Cheng, Haoyu Ma, Shijie Zhao, Jian Zhang, Junlin Li, et al. Opdn: Omnidirectional position-aware deformable network for omnidirectional image superresolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshop, pages 1293-1301, 2023. 1 +[56] Zhihao Sun, Haoran Jiang, Danding Wang, Xirong Li, and Juan Cao. Safl-net: Semantic-agnostic feature learning network with auxiliary plugins for image manipulation detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2023. 2 +[57] Roman Suvorov, Elizaveta Logacheva, Anton Mashikhin, Anastasia Remizova, Armenii Ashukha, Aleksei Silvestrov, Naejin Kong, Harshith Goka, Kiwoong Park, and Victor Lempitsky. Resolution-robust large mask inpainting with fourier convolutions. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), 2022. 7 +[58] Bihan Wen, Ye Zhu, Ramanathan Subramanian, Tian-Tsong Ng, Xuanjing Shen, and Stefan Winkler. Coverage—a novel database for copy-move forgery detection. In Proceedings of the IEEE International Conference on Image Processing (ICIP), 2016. 6 +[59] Yuxin Wen, John Kirchenbauer, Jonas Geiping, and Tom Goldstein. Tree-rings watermarks: Invisible fingerprints for diffusion images. Advances in Neural Information Processing Systems, 36, 2024. 3 +[60] Huikai Wu, Shaocheng Xiang, Gabriben, and Niczem. Faceswap. https://github.com/wuhuikai/FaceSwap, 2020.7 +[61] Haiwei Wu, Jiantao Zhou, Jinyu Tian, and Jun Liu. Robust image forgery detection over online social network shared images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2, 6, 7 +[62] Xiaoshuai Wu, Xin Liao, and Bo Ou. Sepmark: Deep separable watermarking for unified source tracing and deepfake + +detection. In Proceedings of the ACM international conference on Multimedia (MM), 2023. 2, 3, 6, 7 +[63] Yue Wu, Wael AbdAlmageed, and Premkumar Natarajan. *Mantra-net: Manipulation tracing network for detection and localization of image forgeries with anomalous features*. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 2, 6 +[64] Youmin Xu, Chong Mou, Yujie Hu, Jingfen Xie, and Jian Zhang. Robust invertible image steganography. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 3, 4 +[65] Qichao Ying, Zhenxing Qian, Hang Zhou, Haisheng Xu, Xinpeng Zhang, and Siyi Li. From image to image: Immunized image generation. In Proceedings of the ACM international conference on Multimedia (MM), 2021. 2 +[66] Qichao Ying, Xiaoxiao Hu, Xiangyu Zhang, Zhenxing Qian, Sheng Li, and Xinpeng Zhang. Rwn: Robust watermarking network for image cropping localization. In Proceedings of the IEEE International Conference on Image Processing (ICIP), 2022. +[67] Qichao Ying, Hang Zhou, Zhenxing Qian, Sheng Li, and Xinpeng Zhang. Learning to immunize images for tamper localization and self-recovery. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023. 2 +[68] Jiwen Yu, Yinhuai Wang, Chen Zhao, Bernard Ghanem, and Jian Zhang. Freedom: Training-free energy-guided conditional diffusion model. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2023. 1 +[69] Jiwen Yu, Xuanyu Zhang, Youmin Xu, and Jian Zhang. Cross: Diffusion model makes controllable, robust and secure image steganography. Advances in Neural Information Processing Systems, 36, 2024. 3 +[70] Syed Waqas Zamir, Aditya Arora, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, and Ming-Hsuan Yang. Restormer: Efficient transformer for high-resolution image restoration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 5 +[71] Kai Zhang, Yawei Li, Jingyun Liang, Jiezhang Cao, Yulun Zhang, Hao Tang, Radu Timofte, and Luc Van Gool. Practical blind denoising via swim-conv-unet and data synthesis. arXiv preprint arXiv:2203.13278, 2022. 8 +[72] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2023. 1, 7 +[73] Yunqing Zhao, Tianyu Pang, Chao Du, Xiao Yang, Ngai-Man Cheung, and Min Lin. A recipe for watermarking diffusion models. arXiv preprint arXiv:2303.10137, 2023. 3 +[74] Jiren Zhu, Russell Kaplan, Justin Johnson, and Li Fei-Fei. Hidden: Hiding data with deep networks. In European Conference on Computer Vision (ECCV), 2018. 3 +[75] Xinshan Zhu, Yongjun Qian, Xianfeng Zhao, Biao Sun, and Ya Sun. A deep learning approach to patch-based image inpainting forensics. Signal Processing: Image Communication, 67:90-99, 2018. 2 \ No newline at end of file diff --git a/2024/EditGuard_ Versatile Image Watermarking for Tamper Localization and Copyright Protection/images.zip b/2024/EditGuard_ Versatile Image Watermarking for Tamper Localization and Copyright Protection/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..b4ad2f1d5b37d41b1a07de3df88b31da57d54b04 --- /dev/null +++ b/2024/EditGuard_ Versatile Image Watermarking for Tamper Localization and Copyright Protection/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:02a0f626e6c3c1c71e848dad2afc17bbb76567805e89233dd982a08302a26ae8 +size 844726 diff --git a/2024/EditGuard_ Versatile Image Watermarking for Tamper Localization and Copyright Protection/layout.json b/2024/EditGuard_ Versatile Image Watermarking for Tamper Localization and Copyright Protection/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..7fa821465f680f9bd446a762719335ba86a9c35e --- /dev/null +++ b/2024/EditGuard_ Versatile Image Watermarking for Tamper Localization and Copyright Protection/layout.json @@ -0,0 +1,14279 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 73, + 103, + 521, + 139 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 73, + 103, + 521, + 139 + ], + "spans": [ + { + "bbox": [ + 73, + 103, + 521, + 139 + ], + "type": "text", + "content": "EditGuard: Versatile Image Watermarking for Tamper Localization and Copyright Protection" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 101, + 150, + 491, + 179 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 101, + 150, + 491, + 179 + ], + "spans": [ + { + "bbox": [ + 101, + 150, + 491, + 179 + ], + "type": "text", + "content": "Xuanyu Zhang" + }, + { + "bbox": [ + 101, + 150, + 491, + 179 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 101, + 150, + 491, + 179 + ], + "type": "text", + "content": ", Runyi Li" + }, + { + "bbox": [ + 101, + 150, + 491, + 179 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 101, + 150, + 491, + 179 + ], + "type": "text", + "content": ", Jiwen Yu" + }, + { + "bbox": [ + 101, + 150, + 491, + 179 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 101, + 150, + 491, + 179 + ], + "type": "text", + "content": ", Youmin Xu" + }, + { + "bbox": [ + 101, + 150, + 491, + 179 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 101, + 150, + 491, + 179 + ], + "type": "text", + "content": ", Weiqi Li" + }, + { + "bbox": [ + 101, + 150, + 491, + 179 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 101, + 150, + 491, + 179 + ], + "type": "text", + "content": ", Jian Zhang" + }, + { + "bbox": [ + 101, + 150, + 491, + 179 + ], + "type": "inline_equation", + "content": "^{1,2 \\boxtimes}" + }, + { + "bbox": [ + 101, + 150, + 491, + 179 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 101, + 150, + 491, + 179 + ], + "type": "text", + "content": " School of Electronic and Computer Engineering, Peking University" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 74, + 179, + 515, + 194 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 74, + 179, + 515, + 194 + ], + "spans": [ + { + "bbox": [ + 74, + 179, + 515, + 194 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 74, + 179, + 515, + 194 + ], + "type": "text", + "content": " Peking University Shenzhen Graduate School-Rabbitpre AIGC Joint Research Laboratory" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 53, + 204, + 80, + 232 + ], + "blocks": [ + { + "bbox": [ + 53, + 204, + 80, + 232 + ], + "lines": [ + { + "bbox": [ + 53, + 204, + 80, + 232 + ], + "spans": [ + { + "bbox": [ + 53, + 204, + 80, + 232 + ], + "type": "image", + "image_path": "d8378e347524d202d48a803c7b961ea40fa20288455a143e19ee3a00096d0467.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 56, + 236, + 78, + 243 + ], + "lines": [ + { + "bbox": [ + 56, + 236, + 78, + 243 + ], + "spans": [ + { + "bbox": [ + 56, + 236, + 78, + 243 + ], + "type": "text", + "content": "Users" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 91, + 208, + 178, + 234 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 208, + 178, + 234 + ], + "spans": [ + { + "bbox": [ + 91, + 208, + 178, + 234 + ], + "type": "text", + "content": "Please help protect my image against intentional tampering. My exclusive copyright information is 00111010110..." + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 88, + 244, + 116, + 274 + ], + "blocks": [ + { + "bbox": [ + 88, + 244, + 116, + 274 + ], + "lines": [ + { + "bbox": [ + 88, + 244, + 116, + 274 + ], + "spans": [ + { + "bbox": [ + 88, + 244, + 116, + 274 + ], + "type": "image", + "image_path": "5607f94e273b084db6289341ec892ef5497d508e43c3a0c17f984efc08e968a9.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 120, + 246, + 209, + 273 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 120, + 246, + 209, + 273 + ], + "spans": [ + { + "bbox": [ + 120, + 246, + 209, + 273 + ], + "type": "text", + "content": "I've added the exclusive EditGuard watermark to it. You can post the picture on the left and I will be responsible for its protection." + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 179, + 205, + 209, + 236 + ], + "blocks": [ + { + "bbox": [ + 179, + 205, + 209, + 236 + ], + "lines": [ + { + "bbox": [ + 179, + 205, + 209, + 236 + ], + "spans": [ + { + "bbox": [ + 179, + 205, + 209, + 236 + ], + "type": "image", + "image_path": "14ebc880a3f4007db11fd8461f3ab11469a88a059d0a69dcb638c998a78d2614.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 222, + 242, + 253, + 272 + ], + "blocks": [ + { + "bbox": [ + 222, + 242, + 253, + 272 + ], + "lines": [ + { + "bbox": [ + 222, + 242, + 253, + 272 + ], + "spans": [ + { + "bbox": [ + 222, + 242, + 253, + 272 + ], + "type": "image", + "image_path": "a2473491b6f8037f8db2e32ad5296ae7fe40b42159dbfece582cd3295334ad29.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 50, + 279, + 85, + 314 + ], + "blocks": [ + { + "bbox": [ + 50, + 279, + 85, + 314 + ], + "lines": [ + { + "bbox": [ + 50, + 279, + 85, + 314 + ], + "spans": [ + { + "bbox": [ + 50, + 279, + 85, + 314 + ], + "type": "image", + "image_path": "ee9111f2a171ed5d97957c269b21f285bd9bb5ad367af501064b52780a751e81.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 89, + 281, + 211, + 318 + ], + "blocks": [ + { + "bbox": [ + 89, + 281, + 211, + 318 + ], + "lines": [ + { + "bbox": [ + 89, + 281, + 211, + 318 + ], + "spans": [ + { + "bbox": [ + 89, + 281, + 211, + 318 + ], + "type": "image", + "image_path": "416b2131a6a646bcd4b4ea856976f1a7d02bcc4351d5efc5192058194f05ad2a.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 54, + 328, + 80, + 363 + ], + "blocks": [ + { + "bbox": [ + 54, + 328, + 80, + 363 + ], + "lines": [ + { + "bbox": [ + 54, + 328, + 80, + 363 + ], + "spans": [ + { + "bbox": [ + 54, + 328, + 80, + 363 + ], + "type": "image", + "image_path": "8ff376315174fa9cc4dcbf7c7f8fae287bd2ae36592540a47f66c922389dc984.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "bbox": [ + 91, + 330, + 178, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 330, + 178, + 357 + ], + "spans": [ + { + "bbox": [ + 91, + 330, + 178, + 357 + ], + "type": "text", + "content": "I find a picture online that looks similar to mine. Can you provide evidence to prove my copyright and pinpoint tampered areas?" + } + ] + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 179, + 328, + 209, + 358 + ], + "blocks": [ + { + "bbox": [ + 179, + 328, + 209, + 358 + ], + "lines": [ + { + "bbox": [ + 179, + 328, + 209, + 358 + ], + "spans": [ + { + "bbox": [ + 179, + 328, + 209, + 358 + ], + "type": "image", + "image_path": "d39af36db5bf85a178dc5dfc20cf0b681c0da9f4f224560a93bf926d2bd719b1.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 222, + 365, + 253, + 393 + ], + "blocks": [ + { + "bbox": [ + 222, + 365, + 253, + 393 + ], + "lines": [ + { + "bbox": [ + 222, + 365, + 253, + 393 + ], + "spans": [ + { + "bbox": [ + 222, + 365, + 253, + 393 + ], + "type": "image", + "image_path": "b24586fd356b7e063d993524975730611605f0ae08c4a4601554d75233276b06.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 263, + 209, + 305, + 251 + ], + "blocks": [ + { + "bbox": [ + 275, + 203, + 294, + 209 + ], + "lines": [ + { + "bbox": [ + 275, + 203, + 294, + 209 + ], + "spans": [ + { + "bbox": [ + 275, + 203, + 294, + 209 + ], + "type": "text", + "content": "Original" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 263, + 209, + 305, + 251 + ], + "lines": [ + { + "bbox": [ + 263, + 209, + 305, + 251 + ], + "spans": [ + { + "bbox": [ + 263, + 209, + 305, + 251 + ], + "type": "image", + "image_path": "560a5bbb01b919df6b94888b46d2f23b43d556222e6129e9f57072eba95a8914.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 309, + 209, + 351, + 252 + ], + "blocks": [ + { + "bbox": [ + 319, + 203, + 342, + 209 + ], + "lines": [ + { + "bbox": [ + 319, + 203, + 342, + 209 + ], + "spans": [ + { + "bbox": [ + 319, + 203, + 342, + 209 + ], + "type": "text", + "content": "Tampered" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 309, + 209, + 351, + 252 + ], + "lines": [ + { + "bbox": [ + 309, + 209, + 351, + 252 + ], + "spans": [ + { + "bbox": [ + 309, + 209, + 351, + 252 + ], + "type": "image", + "image_path": "dad32f94dbfa1692713fc314f50d78e9f6ae1d2ff385d81bef20198e81a1f49d.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 254, + 353, + 258 + ], + "lines": [ + { + "bbox": [ + 304, + 254, + 353, + 258 + ], + "spans": [ + { + "bbox": [ + 304, + 254, + 353, + 258 + ], + "type": "text", + "content": "Remove the wom" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 356, + 209, + 399, + 252 + ], + "blocks": [ + { + "bbox": [ + 356, + 203, + 400, + 209 + ], + "lines": [ + { + "bbox": [ + 356, + 203, + 400, + 209 + ], + "spans": [ + { + "bbox": [ + 356, + 203, + 400, + 209 + ], + "type": "text", + "content": "Our Predicted mask" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 356, + 209, + 399, + 252 + ], + "lines": [ + { + "bbox": [ + 356, + 209, + 399, + 252 + ], + "spans": [ + { + "bbox": [ + 356, + 209, + 399, + 252 + ], + "type": "image", + "image_path": "24cb7d14e3c5b6e9bda8be532b5ae26396a8f43439fc87fb50a93158f37d47fe.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 263, + 261, + 305, + 304 + ], + "blocks": [ + { + "bbox": [ + 263, + 261, + 305, + 304 + ], + "lines": [ + { + "bbox": [ + 263, + 261, + 305, + 304 + ], + "spans": [ + { + "bbox": [ + 263, + 261, + 305, + 304 + ], + "type": "image", + "image_path": "fce68ffe6804dba46fa7e6a3f4ca39cb656916e0619f82e81f1c2852be05af92.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 306, + 261, + 351, + 304 + ], + "blocks": [ + { + "bbox": [ + 306, + 261, + 351, + 304 + ], + "lines": [ + { + "bbox": [ + 306, + 261, + 351, + 304 + ], + "spans": [ + { + "bbox": [ + 306, + 261, + 351, + 304 + ], + "type": "image", + "image_path": "42a65b5e23bbd3bfacf3a62b020c01fa03d6a03fcda8805c156fd237dce7fe20.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 355, + 261, + 399, + 304 + ], + "blocks": [ + { + "bbox": [ + 355, + 261, + 399, + 304 + ], + "lines": [ + { + "bbox": [ + 355, + 261, + 399, + 304 + ], + "spans": [ + { + "bbox": [ + 355, + 261, + 399, + 304 + ], + "type": "image", + "image_path": "d76bc2f2c1f9c4f6a9ffff0b7f1ef5942951015effd3e05060def371d4432375.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_body" + } + ], + "index": 27 + }, + { + "type": "image", + "bbox": [ + 263, + 312, + 305, + 355 + ], + "blocks": [ + { + "bbox": [ + 266, + 306, + 313, + 312 + ], + "lines": [ + { + "bbox": [ + 266, + 306, + 313, + 312 + ], + "spans": [ + { + "bbox": [ + 266, + 306, + 313, + 312 + ], + "type": "text", + "content": "Change the cloth" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 263, + 312, + 305, + 355 + ], + "lines": [ + { + "bbox": [ + 263, + 312, + 305, + 355 + ], + "spans": [ + { + "bbox": [ + 263, + 312, + 305, + 355 + ], + "type": "image", + "image_path": "46bc5d2965fafcc79eef9aeb40a90a3555cf92ae32dd46f5139ed56306d6eebe.jpg" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_body" + } + ], + "index": 29 + }, + { + "type": "image", + "bbox": [ + 306, + 312, + 351, + 356 + ], + "blocks": [ + { + "bbox": [ + 306, + 312, + 351, + 356 + ], + "lines": [ + { + "bbox": [ + 306, + 312, + 351, + 356 + ], + "spans": [ + { + "bbox": [ + 306, + 312, + 351, + 356 + ], + "type": "image", + "image_path": "66b03e5e20470d096f9996835dc598c65dd463d3bb1e22790cb9337a650226d3.jpg" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_body" + } + ], + "index": 30 + }, + { + "type": "image", + "bbox": [ + 355, + 312, + 399, + 356 + ], + "blocks": [ + { + "bbox": [ + 355, + 312, + 399, + 356 + ], + "lines": [ + { + "bbox": [ + 355, + 312, + 399, + 356 + ], + "spans": [ + { + "bbox": [ + 355, + 312, + 399, + 356 + ], + "type": "image", + "image_path": "a7a50709d76a5eefaaf2fb72378462e3f5161ccac2dd89fb63d669e4b75e2c53.jpg" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_body" + } + ], + "index": 31 + }, + { + "type": "image", + "bbox": [ + 263, + 364, + 305, + 389 + ], + "blocks": [ + { + "bbox": [ + 263, + 357, + 361, + 364 + ], + "lines": [ + { + "bbox": [ + 263, + 357, + 361, + 364 + ], + "spans": [ + { + "bbox": [ + 263, + 357, + 361, + 364 + ], + "type": "text", + "content": "Tamper with AI-generated images" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 263, + 364, + 305, + 389 + ], + "lines": [ + { + "bbox": [ + 263, + 364, + 305, + 389 + ], + "spans": [ + { + "bbox": [ + 263, + 364, + 305, + 389 + ], + "type": "image", + "image_path": "ef46a82a571e641ddad2420ae8d83655029caf679edbd3284c710e5a17c7323a.jpg" + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_body" + } + ], + "index": 34 + }, + { + "type": "image", + "bbox": [ + 306, + 364, + 351, + 407 + ], + "blocks": [ + { + "bbox": [ + 306, + 364, + 351, + 407 + ], + "lines": [ + { + "bbox": [ + 306, + 364, + 351, + 407 + ], + "spans": [ + { + "bbox": [ + 306, + 364, + 351, + 407 + ], + "type": "image", + "image_path": "971d94980f5a423b41c9e7dc2fbc6de91c45d50ad53429365bfe38dc41235a4e.jpg" + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 310, + 407, + 350, + 416 + ], + "lines": [ + { + "bbox": [ + 310, + 407, + 350, + 416 + ], + "spans": [ + { + "bbox": [ + 310, + 407, + 350, + 416 + ], + "type": "text", + "content": "Text Forgery" + } + ] + } + ], + "index": 37, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 45, + 419, + 547, + 464 + ], + "lines": [ + { + "bbox": [ + 45, + 419, + 547, + 464 + ], + "spans": [ + { + "bbox": [ + 45, + 419, + 547, + 464 + ], + "type": "text", + "content": "Figure 1. We propose a versatile proactive forensics framework EditGuard. The application scenario is shown on the left, wherein users embed invisible watermarks to their images via EditGuard in advance. If suffering tampering, users can defend their rights via the tampered areas and copyright information provided by EditGuard. Some supported tampering methods (marked in blue) and localization results of EditGuard are placed on the right. Our EditGuard can achieve over " + }, + { + "bbox": [ + 45, + 419, + 547, + 464 + ], + "type": "inline_equation", + "content": "95\\%" + }, + { + "bbox": [ + 45, + 419, + 547, + 464 + ], + "type": "text", + "content": " localization precision and nearly " + }, + { + "bbox": [ + 45, + 419, + 547, + 464 + ], + "type": "inline_equation", + "content": "100\\%" + }, + { + "bbox": [ + 45, + 419, + 547, + 464 + ], + "type": "text", + "content": " copyright accuracy." + } + ] + } + ], + "index": 61, + "angle": 0, + "type": "image_caption" + } + ], + "index": 35 + }, + { + "type": "image", + "bbox": [ + 355, + 364, + 399, + 407 + ], + "blocks": [ + { + "bbox": [ + 355, + 357, + 397, + 364 + ], + "lines": [ + { + "bbox": [ + 355, + 357, + 397, + 364 + ], + "spans": [ + { + "bbox": [ + 355, + 357, + 397, + 364 + ], + "type": "text", + "content": "es, \"A volcano'" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 355, + 364, + 399, + 407 + ], + "lines": [ + { + "bbox": [ + 355, + 364, + 399, + 407 + ], + "spans": [ + { + "bbox": [ + 355, + 364, + 399, + 407 + ], + "type": "image", + "image_path": "d90b5b7215fe769f777d32aa4267f0e7d64858149ce81f84aae3ff33bf217968.jpg" + } + ] + } + ], + "index": 36, + "angle": 0, + "type": "image_body" + } + ], + "index": 36 + }, + { + "type": "image", + "bbox": [ + 406, + 209, + 449, + 252 + ], + "blocks": [ + { + "bbox": [ + 419, + 203, + 437, + 209 + ], + "lines": [ + { + "bbox": [ + 419, + 203, + 437, + 209 + ], + "spans": [ + { + "bbox": [ + 419, + 203, + 437, + 209 + ], + "type": "text", + "content": "Original" + } + ] + } + ], + "index": 38, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 406, + 209, + 449, + 252 + ], + "lines": [ + { + "bbox": [ + 406, + 209, + 449, + 252 + ], + "spans": [ + { + "bbox": [ + 406, + 209, + 449, + 252 + ], + "type": "image", + "image_path": "1d62c40a83bc886c779407c8a053fdec1369912eb3c5ce8d96fafff48efd4310.jpg" + } + ] + } + ], + "index": 39, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 419, + 253, + 436, + 258 + ], + "lines": [ + { + "bbox": [ + 419, + 253, + 436, + 258 + ], + "spans": [ + { + "bbox": [ + 419, + 253, + 436, + 258 + ], + "type": "text", + "content": "AI Fill" + } + ] + } + ], + "index": 40, + "angle": 0, + "type": "image_caption" + } + ], + "index": 39 + }, + { + "type": "image", + "bbox": [ + 451, + 209, + 496, + 252 + ], + "blocks": [ + { + "bbox": [ + 463, + 203, + 486, + 209 + ], + "lines": [ + { + "bbox": [ + 463, + 203, + 486, + 209 + ], + "spans": [ + { + "bbox": [ + 463, + 203, + 486, + 209 + ], + "type": "text", + "content": "Tampered" + } + ] + } + ], + "index": 41, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 451, + 209, + 496, + 252 + ], + "lines": [ + { + "bbox": [ + 451, + 209, + 496, + 252 + ], + "spans": [ + { + "bbox": [ + 451, + 209, + 496, + 252 + ], + "type": "image", + "image_path": "0bbb5738cc0dbc73621ea30a9504e05c302e3468022b90984db1ae12173ed25e.jpg" + } + ] + } + ], + "index": 42, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 451, + 253, + 496, + 258 + ], + "lines": [ + { + "bbox": [ + 451, + 253, + 496, + 258 + ], + "spans": [ + { + "bbox": [ + 451, + 253, + 496, + 258 + ], + "type": "text", + "content": "Picasso painting" + } + ] + } + ], + "index": 43, + "angle": 0, + "type": "image_caption" + } + ], + "index": 42 + }, + { + "type": "image", + "bbox": [ + 499, + 209, + 542, + 252 + ], + "blocks": [ + { + "bbox": [ + 498, + 203, + 542, + 209 + ], + "lines": [ + { + "bbox": [ + 498, + 203, + 542, + 209 + ], + "spans": [ + { + "bbox": [ + 498, + 203, + 542, + 209 + ], + "type": "text", + "content": "Our Predicted mask" + } + ] + } + ], + "index": 44, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 499, + 209, + 542, + 252 + ], + "lines": [ + { + "bbox": [ + 499, + 209, + 542, + 252 + ], + "spans": [ + { + "bbox": [ + 499, + 209, + 542, + 252 + ], + "type": "image", + "image_path": "ef5e8fa582c6aeb0ef7f3babbcd898c34d8bc2cbd5ffe9345f140dbb7834ad80.jpg" + } + ] + } + ], + "index": 45, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 499, + 253, + 529, + 258 + ], + "lines": [ + { + "bbox": [ + 499, + 253, + 529, + 258 + ], + "spans": [ + { + "bbox": [ + 499, + 253, + 529, + 258 + ], + "type": "text", + "content": "on the wall" + } + ] + } + ], + "index": 46, + "angle": 0, + "type": "image_caption" + } + ], + "index": 45 + }, + { + "type": "image", + "bbox": [ + 406, + 261, + 449, + 304 + ], + "blocks": [ + { + "bbox": [ + 406, + 261, + 449, + 304 + ], + "lines": [ + { + "bbox": [ + 406, + 261, + 449, + 304 + ], + "spans": [ + { + "bbox": [ + 406, + 261, + 449, + 304 + ], + "type": "image", + "image_path": "b55ae49daacc5bf74a9f498b5e1af76d43b516d77223e6a6e30fbaea5bbc93b4.jpg" + } + ] + } + ], + "index": 47, + "angle": 0, + "type": "image_body" + } + ], + "index": 47 + }, + { + "type": "image", + "bbox": [ + 406, + 312, + 449, + 356 + ], + "blocks": [ + { + "bbox": [ + 413, + 306, + 470, + 312 + ], + "lines": [ + { + "bbox": [ + 413, + 306, + 470, + 312 + ], + "spans": [ + { + "bbox": [ + 413, + 306, + 470, + 312 + ], + "type": "text", + "content": "Replace background" + } + ] + } + ], + "index": 48, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 406, + 312, + 449, + 356 + ], + "lines": [ + { + "bbox": [ + 406, + 312, + 449, + 356 + ], + "spans": [ + { + "bbox": [ + 406, + 312, + 449, + 356 + ], + "type": "image", + "image_path": "4239f38f67c4141fd0b9043c6a100f4452549f3da6935110b6d78f4c63ea9fad.jpg" + } + ] + } + ], + "index": 49, + "angle": 0, + "type": "image_body" + } + ], + "index": 49 + }, + { + "type": "image", + "bbox": [ + 451, + 261, + 496, + 304 + ], + "blocks": [ + { + "bbox": [ + 451, + 261, + 496, + 304 + ], + "lines": [ + { + "bbox": [ + 451, + 261, + 496, + 304 + ], + "spans": [ + { + "bbox": [ + 451, + 261, + 496, + 304 + ], + "type": "image", + "image_path": "ba7f575d98c5ab4ce391c376226cc462b2a4cf5a332b3ffc42477fc7c3c6bf7a.jpg" + } + ] + } + ], + "index": 51, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 452, + 306, + 496, + 312 + ], + "lines": [ + { + "bbox": [ + 452, + 306, + 496, + 312 + ], + "spans": [ + { + "bbox": [ + 452, + 306, + 496, + 312 + ], + "type": "text", + "content": "ground, \"A man" + } + ] + } + ], + "index": 52, + "angle": 0, + "type": "image_caption" + } + ], + "index": 51 + }, + { + "type": "image", + "bbox": [ + 499, + 261, + 542, + 304 + ], + "blocks": [ + { + "bbox": [ + 499, + 261, + 542, + 304 + ], + "lines": [ + { + "bbox": [ + 499, + 261, + 542, + 304 + ], + "spans": [ + { + "bbox": [ + 499, + 261, + 542, + 304 + ], + "type": "image", + "image_path": "2a5384c2a631ee76fd571960084ccd4e9df042886f5ae6724e33eb9f29d3814c.jpg" + } + ] + } + ], + "index": 53, + "angle": 0, + "type": "image_body" + } + ], + "index": 53 + }, + { + "type": "image", + "bbox": [ + 499, + 312, + 542, + 356 + ], + "blocks": [ + { + "bbox": [ + 499, + 306, + 529, + 312 + ], + "lines": [ + { + "bbox": [ + 499, + 306, + 529, + 312 + ], + "spans": [ + { + "bbox": [ + 499, + 306, + 529, + 312 + ], + "type": "text", + "content": "in an office" + } + ] + } + ], + "index": 54, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 499, + 312, + 542, + 356 + ], + "lines": [ + { + "bbox": [ + 499, + 312, + 542, + 356 + ], + "spans": [ + { + "bbox": [ + 499, + 312, + 542, + 356 + ], + "type": "image", + "image_path": "d7a5ce36ace3683d775141aa1d2560a26c3f3c151c5effc0dfe4b62da163ec0c.jpg" + } + ] + } + ], + "index": 55, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 499, + 357, + 539, + 363 + ], + "lines": [ + { + "bbox": [ + 499, + 357, + 539, + 363 + ], + "spans": [ + { + "bbox": [ + 499, + 357, + 539, + 363 + ], + "type": "text", + "content": "\"A straw hat\"" + } + ] + } + ], + "index": 56, + "angle": 0, + "type": "image_caption" + } + ], + "index": 55 + }, + { + "type": "image", + "bbox": [ + 406, + 364, + 449, + 407 + ], + "blocks": [ + { + "bbox": [ + 406, + 357, + 496, + 364 + ], + "lines": [ + { + "bbox": [ + 406, + 357, + 496, + 364 + ], + "spans": [ + { + "bbox": [ + 406, + 357, + 496, + 364 + ], + "type": "text", + "content": "Secondary creation of artworks" + } + ] + } + ], + "index": 50, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 406, + 364, + 449, + 407 + ], + "lines": [ + { + "bbox": [ + 406, + 364, + 449, + 407 + ], + "spans": [ + { + "bbox": [ + 406, + 364, + 449, + 407 + ], + "type": "image", + "image_path": "c852226a6f6691fc2c0bbb154d0f9f2addaeb99d34c00ed03a87681a891a5cea.jpg" + } + ] + } + ], + "index": 57, + "angle": 0, + "type": "image_body" + } + ], + "index": 57 + }, + { + "type": "image", + "bbox": [ + 452, + 364, + 496, + 407 + ], + "blocks": [ + { + "bbox": [ + 452, + 364, + 496, + 407 + ], + "lines": [ + { + "bbox": [ + 452, + 364, + 496, + 407 + ], + "spans": [ + { + "bbox": [ + 452, + 364, + 496, + 407 + ], + "type": "image", + "image_path": "0a54d13aa873acdbc59699f323546ea56bee7fde1c33c7aa9140f6072c2bc195.jpg" + } + ] + } + ], + "index": 58, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 458, + 408, + 489, + 415 + ], + "lines": [ + { + "bbox": [ + 458, + 408, + 489, + 415 + ], + "spans": [ + { + "bbox": [ + 458, + 408, + 489, + 415 + ], + "type": "text", + "content": "Swap Face" + } + ] + } + ], + "index": 59, + "angle": 0, + "type": "image_caption" + } + ], + "index": 58 + }, + { + "type": "image", + "bbox": [ + 499, + 364, + 542, + 407 + ], + "blocks": [ + { + "bbox": [ + 499, + 364, + 542, + 407 + ], + "lines": [ + { + "bbox": [ + 499, + 364, + 542, + 407 + ], + "spans": [ + { + "bbox": [ + 499, + 364, + 542, + 407 + ], + "type": "image", + "image_path": "8863579b2f6447b29ef99f80bc578a2ed5b9550e69081aeddd5b7240ffc19c45.jpg" + } + ] + } + ], + "index": 60, + "angle": 0, + "type": "image_body" + } + ], + "index": 60 + }, + { + "bbox": [ + 143, + 475, + 191, + 488 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 475, + 191, + 488 + ], + "spans": [ + { + "bbox": [ + 143, + 475, + 191, + 488 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 62 + }, + { + "bbox": [ + 45, + 502, + 288, + 682 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 502, + 288, + 682 + ], + "spans": [ + { + "bbox": [ + 45, + 502, + 288, + 682 + ], + "type": "text", + "content": "In the era of AI-generated content (AIGC), malicious tampering poses imminent threats to copyright integrity and information security. Current deep image watermarking, while widely accepted for safeguarding visual content, can only protect copyright and ensure traceability. They fall short in localizing increasingly realistic image tampering, potentially leading to trust crises, privacy violations, and legal disputes. To solve this challenge, we propose an innovative proactive forensics framework EditGuard, to unify copyright protection and tamper-agnostic localization, especially for AIGC-based editing methods. It can offer a meticulous embedding of imperceptible watermarks and precise decoding of tampered areas and copyright information. Leveraging our observed fragility and locality of image-into-image steganography, the realization of Edit" + } + ] + } + ], + "index": 63 + }, + { + "bbox": [ + 304, + 476, + 546, + 561 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 476, + 546, + 561 + ], + "spans": [ + { + "bbox": [ + 304, + 476, + 546, + 561 + ], + "type": "text", + "content": "Guard can be converted into a united image-bit steganography issue, thus completely decoupling the training process from the tampering types. Extensive experiments verify that our EditGuard balances the tamper localization accuracy, copyright recovery precision, and generalizability to various AIGC-based tampering methods, especially for image forgery that is difficult for the naked eye to detect." + } + ] + } + ], + "index": 65 + }, + { + "bbox": [ + 306, + 572, + 387, + 585 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 572, + 387, + 585 + ], + "spans": [ + { + "bbox": [ + 306, + 572, + 387, + 585 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 66 + }, + { + "bbox": [ + 304, + 594, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 546, + 713 + ], + "type": "text", + "content": "Owing to the advantageous properties of diffusion models and the bolstering of extensive datasets, AI-generated content (AIGC) models like DALL-E 3 [14], Imagen [52], and Stable Diffusion [51], can produce lifelike and wondrous images, which brings great convenience to photography enthusiasts and image editors. Nonetheless, the remarkable capabilities of these models come with a double-edged sword, presenting new challenges in copyright protection and information security. The efficiency of image manipulation [43, 47, 49, 51, 55, 68, 72] has blurred the" + } + ] + } + ], + "index": 67 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 693, + 288, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 693, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 693, + 288, + 713 + ], + "type": "text", + "content": "This work was supported by National Natural Science Foundation of China under Grant 62372016. (Corresponding author: Jian Zhang.)" + } + ] + } + ], + "index": 64 + }, + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "11964" + } + ] + } + ], + "index": 68 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 156 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 156 + ], + "type": "text", + "content": "line between fact and forgery, ushering in myriad security and legal concerns. For instance, artistic works are vulnerable to malicious tampering or unauthorized AI-facilitated recreations, making it challenging to protect their original creations [13]. Meanwhile, forged images may be spread online or used as court evidence, causing adverse effects on public opinion, ethical issues, and social stability." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 157, + 288, + 409 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 157, + 288, + 409 + ], + "spans": [ + { + "bbox": [ + 46, + 157, + 288, + 409 + ], + "type": "text", + "content": "Given the challenges of preventing image tampering from the source, image watermarking has become a consensus for proactive forensics [54]. However, most prevalent forensic image watermarking [12, 15, 44, 62] still focus on detecting image authenticity or protecting image copyrights, but fall short when it comes to advanced demands, such as localizing tampered areas. Tamper localization facilitates an evaluation of the severity of the image manipulation, and provides an understanding of the intent of attackers, potentially allowing for the partial reuse of the tampered images. However, most passive forensics methods such as previous black-box localization networks [8, 56, 61] tend to seek anomalies like artifacts or flickers in images but struggle to detect more realistic textures and more advanced AIGC models. Moreover, they inevitably need to introduce tampered data during the training and focus solely on specific \"CheapFake\" tampering like slicing and copy-and-paste [8, 45, 56], or on \"DeepFake\" targeting human faces [2, 50], restricted in generalizability. Thus, it is vital to develop an integrated watermarking framework that unites tamper-agnostic localization and copyright protection." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 410, + 287, + 542 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 410, + 287, + 542 + ], + "spans": [ + { + "bbox": [ + 46, + 410, + 287, + 542 + ], + "type": "text", + "content": "To clarify our task scope, we re-emphasize the definition of dual forensics tasks as illustrated in Fig. 1: (1) Copyright protection: \"Who does this image belong to?\" We aspire to accurately retrieve the original copyright of an image, even suffering various tampering and degradation. (2) Tamper localization: \"Where was this image manipulated?\" We aim to precisely pinpoint the tampered areas, unrestrained by specific tampering types. To the best of our knowledge, no existing method accomplishes these two tasks simultaneously, while maintaining a balance of high precision and extensive generalizability." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 543, + 287, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 543, + 287, + 685 + ], + "spans": [ + { + "bbox": [ + 46, + 543, + 287, + 685 + ], + "type": "text", + "content": "To address this urgent demand, we propose a novel proactive forensics framework, dubbed EditGuard, to protect copyrights and localize tamper areas for AIGC-based editing methods. Specifically, drawing inspiration from our observed locality and fragility of image-into-image (I2I) steganography and inherent robustness of bit-into-image steganography, we can transform the realization of EditGuard into a joint image-bit steganography issue, which allows the training of EditGuard to be entirely decoupled from tampering types, thereby endowing it with exceptional generalizability and locate tampering in a zero-shot manner. In a nutshell, our contributions are as follows:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "content": "(1) We present the first attempt to design a deep versatile proactive forensics framework EditGuard for univer" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 72, + 545, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 108 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 108 + ], + "type": "text", + "content": "sal tamper localization and copyright protection. It embeds dual invisible watermarks into original images and accurately decodes tampered areas and copyright information." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 110, + 545, + 281 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 304, + 110, + 545, + 170 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 110, + 545, + 170 + ], + "spans": [ + { + "bbox": [ + 304, + 110, + 545, + 170 + ], + "type": "text", + "content": "(2) We have observed the fragility and locality of I2I steganography and innovatively convert the solution of this dual forensics task into training a united Image-Bit Steganography Network (IBSN), and utilize the core components of IBSN to construct EditGuard." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 171, + 545, + 208 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 171, + 545, + 208 + ], + "spans": [ + { + "bbox": [ + 304, + 171, + 545, + 208 + ], + "type": "text", + "content": "(3) We introduce a prompt-based posterior estimation module to enhance the localization accuracy and degradation robustness of the proposed framework." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 209, + 545, + 281 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 209, + 545, + 281 + ], + "spans": [ + { + "bbox": [ + 304, + 209, + 545, + 281 + ], + "type": "text", + "content": "(4) The effectiveness of our method has been verified on our constructed dataset and classical benchmarks. Compared to other competitive methods, our approach has notable merits in localization precision, generalization abilities, and copyright accuracy without any labeled data or additional training required for specific tampering types." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 292, + 394, + 304 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 292, + 394, + 304 + ], + "spans": [ + { + "bbox": [ + 306, + 292, + 394, + 304 + ], + "type": "text", + "content": "2. Related works" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 311, + 425, + 323 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 311, + 425, + 323 + ], + "spans": [ + { + "bbox": [ + 306, + 311, + 425, + 323 + ], + "type": "text", + "content": "2.1. Tamper Localization" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 329, + 545, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 329, + 545, + 605 + ], + "spans": [ + { + "bbox": [ + 304, + 329, + 545, + 605 + ], + "type": "text", + "content": "Prevalent passive image forensic techniques have focused on localizing specific types of manipulations [25, 33, 34, 53, 61, 75]. Meanwhile, some universal tamper localization methods [5, 23, 31, 35, 63, 65-67] also tend to explore artifacts and anomalies in tampered images. For instance, MVSS-Net [8] employed multi-view feature learning and multi-scale supervision to jointly exploit boundary artifacts and the noise view of images. OSN [61] proposed a novel robust training scheme to address the challenges posed by lossy operations. Trufor [17] used a learned noise-sensitive fingerprint and extracted both high-level and low-level traces via transformer-based fusion. HiFi-Net [18] utilized multi-branch feature extractor and localization modules for both CNN-synthesized and edited images. SAFL-Net [56] constrained a feature extractor to learn semantic-agnostic features with specific modules and auxiliary tasks. However, the above-mentioned passive localization methods are often limited in terms of generalization and localization accuracy, which usually work on known tampering types that have been trained. Although MaLP [2] used template matching for proactive tamper localization, it still requires a large number of forgery images and cannot fully decouple the network training from the tamper types." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 611, + 430, + 624 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 611, + 430, + 624 + ], + "spans": [ + { + "bbox": [ + 306, + 611, + 430, + 624 + ], + "type": "text", + "content": "2.2. Image Watermarking" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "text", + "content": "Image watermarking can be broadly used for the verification, authenticity, and traceability of images. Although traditional fragile watermarking [6, 24, 29, 36, 37, 48] can also achieve block-wise tamper localization, their localization accuracy and flexibility are unsatisfactory. How to realize joint pixel-level tamper localization and copyright protection remains largely unexplored. Owing to the development" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "11965" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 49, + 71, + 287, + 163 + ], + "blocks": [ + { + "bbox": [ + 49, + 71, + 287, + 163 + ], + "lines": [ + { + "bbox": [ + 49, + 71, + 287, + 163 + ], + "spans": [ + { + "bbox": [ + 49, + 71, + 287, + 163 + ], + "type": "image", + "image_path": "f39961573342ac5698699c406fd9e570e66dc33b26fe8a12add7aac800fee9d6.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 47, + 168, + 287, + 240 + ], + "lines": [ + { + "bbox": [ + 47, + 168, + 287, + 240 + ], + "spans": [ + { + "bbox": [ + 47, + 168, + 287, + 240 + ], + "type": "text", + "content": "Figure 2. Fragility and locality of I2I steganography. The first line shows that when " + }, + { + "bbox": [ + 47, + 168, + 287, + 240 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_{con}^{\\prime}" + }, + { + "bbox": [ + 47, + 168, + 287, + 240 + ], + "type": "text", + "content": " is changed, " + }, + { + "bbox": [ + 47, + 168, + 287, + 240 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{X}}_{sec}" + }, + { + "bbox": [ + 47, + 168, + 287, + 240 + ], + "type": "text", + "content": " will also be fragility demaged. The second line plots the attribution maps " + }, + { + "bbox": [ + 47, + 168, + 287, + 240 + ], + "type": "inline_equation", + "content": "\\frac{1}{|\\mathbb{S}|}\\sum_{(i,j)\\in \\mathbb{S}}\\frac{\\partial\\hat{\\mathbf{X}}_{sec}[i,j]}{\\partial\\mathbf{X}_{con}^{\\prime}}" + }, + { + "bbox": [ + 47, + 168, + 287, + 240 + ], + "type": "text", + "content": " of five point sets " + }, + { + "bbox": [ + 47, + 168, + 287, + 240 + ], + "type": "inline_equation", + "content": "\\mathbb{S}" + }, + { + "bbox": [ + 47, + 168, + 287, + 240 + ], + "type": "text", + "content": " (marked by " + }, + { + "bbox": [ + 47, + 168, + 287, + 240 + ], + "type": "inline_equation", + "content": "①-⑤" + }, + { + "bbox": [ + 47, + 168, + 287, + 240 + ], + "type": "text", + "content": ") in " + }, + { + "bbox": [ + 47, + 168, + 287, + 240 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{X}}_{sec}" + }, + { + "bbox": [ + 47, + 168, + 287, + 240 + ], + "type": "text", + "content": ". We observed that " + }, + { + "bbox": [ + 47, + 168, + 287, + 240 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{X}}_{sec}" + }, + { + "bbox": [ + 47, + 168, + 287, + 240 + ], + "type": "text", + "content": " almost only has a strong response at the corresponding positions of " + }, + { + "bbox": [ + 47, + 168, + 287, + 240 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_{con}^{\\prime}" + }, + { + "bbox": [ + 47, + 168, + 287, + 240 + ], + "type": "text", + "content": " and its neighborhoods." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 247, + 288, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 247, + 288, + 403 + ], + "spans": [ + { + "bbox": [ + 46, + 247, + 288, + 403 + ], + "type": "text", + "content": "of deep learning, deep image watermarking has attracted increased attention. For instance, HiDDeN [74] firstly introduced a deep encoder-decoder network to hide and recover bitstream. Moreover, many distortion layers such as differentiable JPEG and screen-shooting [1, 10, 40, 62] were designed to enhance its robustness. Flow-based models [11, 44] were also utilized to further improve the fidelity of container images. Recently, researchers [7, 12, 27, 59, 73] have designed specialized watermarking mechanisms for large-scale image generation models [51], to merge watermarking into the generation process. However, these deep watermarking methods have a singular function and cannot accurately localize the tampered areas." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 416, + 236, + 430 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 416, + 236, + 430 + ], + "spans": [ + { + "bbox": [ + 47, + 416, + 236, + 430 + ], + "type": "text", + "content": "3. Overall Framework of EditGuard" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 437, + 121, + 449 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 437, + 121, + 449 + ], + "spans": [ + { + "bbox": [ + 47, + 437, + 121, + 449 + ], + "type": "text", + "content": "3.1. Motivation" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 456, + 287, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 456, + 287, + 552 + ], + "spans": [ + { + "bbox": [ + 46, + 456, + 287, + 552 + ], + "type": "text", + "content": "Challenges of existing methods: (1) How to equip existing watermark methods, which are solely for copyright protection, with the ability to localize tampering is the crux of EditGuard. We will solve it via the framework design in Sec. 3.2. (2) Most previous tamper localization methods inevitably introduce specific tampering data during network training but tend to raise generalization concerns in unknown tampering types, which will be addressed in Sec. 3.3." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 558, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 558, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 558, + 288, + 714 + ], + "type": "text", + "content": "Our observations: Fortunately, we observed that imageinto-image (I2I) steganography exhibits distinct fragility and locality, possessing great potential to address these issues. Concretly, I2I steganography [3, 28, 42, 46, 64, 69] aims to hide a secret image " + }, + { + "bbox": [ + 46, + 558, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_{\\text{sec}}" + }, + { + "bbox": [ + 46, + 558, + 288, + 714 + ], + "type": "text", + "content": " into a cover image " + }, + { + "bbox": [ + 46, + 558, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_{\\text{cov}}" + }, + { + "bbox": [ + 46, + 558, + 288, + 714 + ], + "type": "text", + "content": " to produce a container image " + }, + { + "bbox": [ + 46, + 558, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_{\\text{con}}" + }, + { + "bbox": [ + 46, + 558, + 288, + 714 + ], + "type": "text", + "content": ", and reveal " + }, + { + "bbox": [ + 46, + 558, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{X}}_{\\text{sec}}" + }, + { + "bbox": [ + 46, + 558, + 288, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 558, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{X}}_{\\text{cov}}" + }, + { + "bbox": [ + 46, + 558, + 288, + 714 + ], + "type": "text", + "content": " with minimal distortion from the received image " + }, + { + "bbox": [ + 46, + 558, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_{\\text{con}}^{\\prime}" + }, + { + "bbox": [ + 46, + 558, + 288, + 714 + ], + "type": "text", + "content": ". We discover that when " + }, + { + "bbox": [ + 46, + 558, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_{\\text{con}}^{\\prime}" + }, + { + "bbox": [ + 46, + 558, + 288, + 714 + ], + "type": "text", + "content": " undergoes significant alteration compared to " + }, + { + "bbox": [ + 46, + 558, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_{\\text{con}}" + }, + { + "bbox": [ + 46, + 558, + 288, + 714 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 558, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{X}}_{\\text{sec}}" + }, + { + "bbox": [ + 46, + 558, + 288, + 714 + ], + "type": "text", + "content": " will also be damaged and generate artifacts (the first row of Fig. 2), which is called fragility. Furthermore, we notice that the artifacts in " + }, + { + "bbox": [ + 46, + 558, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{X}}_{\\text{sec}}" + }, + { + "bbox": [ + 46, + 558, + 288, + 714 + ], + "type": "text", + "content": " are almost pixel-level corresponding to the changes in " + }, + { + "bbox": [ + 46, + 558, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_{\\text{con}}^{\\prime}" + }, + { + "bbox": [ + 46, + 558, + 288, + 714 + ], + "type": "text", + "content": " relative to " + }, + { + "bbox": [ + 46, + 558, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_{\\text{con}}" + }, + { + "bbox": [ + 46, + 558, + 288, + 714 + ], + "type": "text", + "content": ", which is called locality. To demonstrate this locality," + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "text", + "content": "we select five " + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "inline_equation", + "content": "7 \\times 7" + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "text", + "content": " point sets on " + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{X}}_{\\text {sec }}" + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "text", + "content": " and calculated their attribution maps with respect to " + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_{\\text {con }}^{\\prime}" + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "text", + "content": ". As plotted in the second row of Fig. 2, " + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{X}}_{\\text {sec }}" + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "text", + "content": " only exhibits a strong response at the corresponding locations of " + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_{\\text {con }}^{\\prime}" + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "text", + "content": " and their immediate vicinity, almost irrelevant to other pixels. These properties inspire us to treat " + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_{\\text {sec }}" + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "text", + "content": " as a special localization watermark and embed it within existing watermarking frameworks." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 166, + 523, + 178 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 166, + 523, + 178 + ], + "spans": [ + { + "bbox": [ + 305, + 166, + 523, + 178 + ], + "type": "text", + "content": "3.2. Framework Design and Forensics Process" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 184, + 545, + 279 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 184, + 545, + 279 + ], + "spans": [ + { + "bbox": [ + 304, + 184, + 545, + 279 + ], + "type": "text", + "content": "To realize united tamper localization and copyright protection, EditGuard is envisioned to embed both a 2D localization watermark and a 1D copyright traceability watermark into the original image in an imperceptible manner, which allows the decoding end to obtain the copyright of the images and a binary mask reflecting tampered areas. However, designing such a framework needs to solve the compatibility issue of two types of watermarks." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 281, + 545, + 412 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 281, + 545, + 412 + ], + "spans": [ + { + "bbox": [ + 304, + 281, + 545, + 412 + ], + "type": "text", + "content": "(1) Local vs. Global: The localization watermark is required to be hidden in the corresponding pixel positions of the original image, while the copyright watermark should be unrelated to spatial location and embedded in the global area redundantly. (2) Semi-fragile vs. Robust: The desired attribute of the localization watermark is semi-fragile, which means it is fragile to tampering but robust against some common degradations (such as Gaussian noise, JPEG compression, and Poisson noise) during network transmission. However, the copyright should be extracted nearly losslessly, irrespective of tampering or degradation." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 413, + 545, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 413, + 545, + 581 + ], + "spans": [ + { + "bbox": [ + 304, + 413, + 545, + 581 + ], + "type": "text", + "content": "To address the two pivotal conflicts, EditGuard employs a \"sequential encoding and parallel decoding\" structure, which comprises a dual-watermark encoder, a tamper locator, and a copyright extractor. As shown in Fig. 3, the dual-watermark encoder will sequentially add pre-defined localization watermark and global copyright watermark " + }, + { + "bbox": [ + 304, + 413, + 545, + 581 + ], + "type": "inline_equation", + "content": "\\mathbf{w}_{cop}" + }, + { + "bbox": [ + 304, + 413, + 545, + 581 + ], + "type": "text", + "content": " provided by users to the original image " + }, + { + "bbox": [ + 304, + 413, + 545, + 581 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_{ori}" + }, + { + "bbox": [ + 304, + 413, + 545, + 581 + ], + "type": "text", + "content": ", forming the container image " + }, + { + "bbox": [ + 304, + 413, + 545, + 581 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_{con}" + }, + { + "bbox": [ + 304, + 413, + 545, + 581 + ], + "type": "text", + "content": ". Our experiments have proved that parallel encoding cannot effectively add dual watermarks into images (in supplementary materials (S.M.). In contrast, sequential embedding effectively prevents cross-interference by hiding these two watermarks. Furthermore, we model the network transmission process in which the received (tampered) image " + }, + { + "bbox": [ + 304, + 413, + 545, + 581 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_{rec}" + }, + { + "bbox": [ + 304, + 413, + 545, + 581 + ], + "type": "text", + "content": " is transformed from " + }, + { + "bbox": [ + 304, + 413, + 545, + 581 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_{con}" + }, + { + "bbox": [ + 304, + 413, + 545, + 581 + ], + "type": "text", + "content": " as:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 332, + 586, + 545, + 598 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 332, + 586, + 545, + 598 + ], + "spans": [ + { + "bbox": [ + 332, + 586, + 545, + 598 + ], + "type": "interline_equation", + "content": "\\mathbf {I} _ {r e c} = \\mathcal {D} \\left(\\mathbf {I} _ {c o n} \\odot (\\mathbf {1} - \\mathbf {M}) + \\mathcal {T} \\left(\\mathbf {I} _ {c o n}\\right) \\odot \\mathbf {M}\\right), \\tag {1}", + "image_path": "62bcf3c6b2650dc7027a3aefc5a71c53e2cd24b2bca6de78887cbc0117195024.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 603, + 545, + 699 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 603, + 545, + 699 + ], + "spans": [ + { + "bbox": [ + 304, + 603, + 545, + 699 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 603, + 545, + 699 + ], + "type": "inline_equation", + "content": "\\mathcal{T}(\\cdot),\\mathcal{D}(\\cdot)" + }, + { + "bbox": [ + 304, + 603, + 545, + 699 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 603, + 545, + 699 + ], + "type": "inline_equation", + "content": "\\mathbf{M}" + }, + { + "bbox": [ + 304, + 603, + 545, + 699 + ], + "type": "text", + "content": " respectively denote the tamper function, degradation operation, and tempered mask. Moreover, the parallel decoding processes allow us to flexibly train each branch under different levels of robustness and obtain the predicted mask " + }, + { + "bbox": [ + 304, + 603, + 545, + 699 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{M}}" + }, + { + "bbox": [ + 304, + 603, + 545, + 699 + ], + "type": "text", + "content": " via the tamper locator and traceability watermark " + }, + { + "bbox": [ + 304, + 603, + 545, + 699 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{w}}_{cop}" + }, + { + "bbox": [ + 304, + 603, + 545, + 699 + ], + "type": "text", + "content": " via the copyright extractor. We can categorize the dual forensic process of EditGuard into the following scenarios." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 701, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 701, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 306, + 701, + 545, + 714 + ], + "type": "text", + "content": "Case 1: If " + }, + { + "bbox": [ + 306, + 701, + 545, + 714 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{w}}_{\\text{cop}} \\not\\simeq \\mathbf{w}_{\\text{cop}}" + }, + { + "bbox": [ + 306, + 701, + 545, + 714 + ], + "type": "text", + "content": ", suspicious " + }, + { + "bbox": [ + 306, + 701, + 545, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_{\\text{rec}}" + }, + { + "bbox": [ + 306, + 701, + 545, + 714 + ], + "type": "text", + "content": " is either not" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "11966" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 70, + 547, + 201 + ], + "blocks": [ + { + "bbox": [ + 50, + 70, + 547, + 201 + ], + "lines": [ + { + "bbox": [ + 50, + 70, + 547, + 201 + ], + "spans": [ + { + "bbox": [ + 50, + 70, + 547, + 201 + ], + "type": "image", + "image_path": "86330ab304ab8ee076a9aed3f95af6473f94c3b7e495a949fe4060b129a8b924.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 206, + 547, + 251 + ], + "lines": [ + { + "bbox": [ + 46, + 206, + 547, + 251 + ], + "spans": [ + { + "bbox": [ + 46, + 206, + 547, + 251 + ], + "type": "text", + "content": "Figure 3. Illustration of the proposed proactive forensics framework EditGuard. The dual-watermark encoder sequentially embeds the pre-defined localization watermark and copyright watermark " + }, + { + "bbox": [ + 46, + 206, + 547, + 251 + ], + "type": "inline_equation", + "content": "\\mathbf{w}_{\\text{cop}}" + }, + { + "bbox": [ + 46, + 206, + 547, + 251 + ], + "type": "text", + "content": " into the original image " + }, + { + "bbox": [ + 46, + 206, + 547, + 251 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_{\\text{ori}}" + }, + { + "bbox": [ + 46, + 206, + 547, + 251 + ], + "type": "text", + "content": ", generating the container image " + }, + { + "bbox": [ + 46, + 206, + 547, + 251 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_{\\text{con}}" + }, + { + "bbox": [ + 46, + 206, + 547, + 251 + ], + "type": "text", + "content": ". After encountering potential malicious tampering and degradation during network transmission, tampered mask " + }, + { + "bbox": [ + 46, + 206, + 547, + 251 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{M}}" + }, + { + "bbox": [ + 46, + 206, + 547, + 251 + ], + "type": "text", + "content": " and copyright information " + }, + { + "bbox": [ + 46, + 206, + 547, + 251 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{W}}_{\\text{cop}}" + }, + { + "bbox": [ + 46, + 206, + 547, + 251 + ], + "type": "text", + "content": " are respectively extracted via the tamper locator and copyright extractor from the received image " + }, + { + "bbox": [ + 46, + 206, + 547, + 251 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_{\\text{rec}}" + }, + { + "bbox": [ + 46, + 206, + 547, + 251 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 257, + 287, + 281 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 257, + 287, + 281 + ], + "spans": [ + { + "bbox": [ + 46, + 257, + 287, + 281 + ], + "type": "text", + "content": "registered in our EditGuard or underwent extremely severe global tampering, rendering it unreliable as evidence." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 281, + 287, + 354 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 46, + 281, + 287, + 330 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 281, + 287, + 330 + ], + "spans": [ + { + "bbox": [ + 46, + 281, + 287, + 330 + ], + "type": "text", + "content": "- Case 2: If " + }, + { + "bbox": [ + 46, + 281, + 287, + 330 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{w}}_{\\text {cop }} \\approx \\mathbf{w}_{\\text {cop }}" + }, + { + "bbox": [ + 46, + 281, + 287, + 330 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 281, + 287, + 330 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{M}} \\not\\approx \\mathbf{0}" + }, + { + "bbox": [ + 46, + 281, + 287, + 330 + ], + "type": "text", + "content": ", suspicious " + }, + { + "bbox": [ + 46, + 281, + 287, + 330 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_{\\text {rec }}" + }, + { + "bbox": [ + 46, + 281, + 287, + 330 + ], + "type": "text", + "content": " has undergone tampering, disqualifying it as valid evidence. Users may infer the intention of tamperers based on " + }, + { + "bbox": [ + 46, + 281, + 287, + 330 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{M}}" + }, + { + "bbox": [ + 46, + 281, + 287, + 330 + ], + "type": "text", + "content": " and decide whether to reuse parts of the image." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 331, + 286, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 331, + 286, + 354 + ], + "spans": [ + { + "bbox": [ + 47, + 331, + 286, + 354 + ], + "type": "inline_equation", + "content": "\\square" + }, + { + "bbox": [ + 47, + 331, + 286, + 354 + ], + "type": "text", + "content": " Case 3: If " + }, + { + "bbox": [ + 47, + 331, + 286, + 354 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{w}}_{cop} \\approx \\mathbf{w}_{cop}" + }, + { + "bbox": [ + 47, + 331, + 286, + 354 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 331, + 286, + 354 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{M}} \\approx \\mathbf{0}" + }, + { + "bbox": [ + 47, + 331, + 286, + 354 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 47, + 331, + 286, + 354 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_{rec}" + }, + { + "bbox": [ + 47, + 331, + 286, + 354 + ], + "type": "text", + "content": " remains untampered and trustworthy under the shield of EditGuard." + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 47, + 361, + 286, + 373 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 361, + 286, + 373 + ], + "spans": [ + { + "bbox": [ + 47, + 361, + 286, + 373 + ], + "type": "text", + "content": "3.3. Transform Dual Forensics into Steganography" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 379, + 289, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 379, + 289, + 582 + ], + "spans": [ + { + "bbox": [ + 46, + 379, + 289, + 582 + ], + "type": "text", + "content": "To realize universal and tamper-agnostic localization, we resort to our observed locality and fragility of I2I steganography. As described in Sec. 3.1, localization watermarking and tamper locator in Fig. 3 can be effectively realized via image hiding and revealing. Meanwhile, combining with the robustness of current bit-into-image steganography, copyright watermarking and extractor in Fig. 3 are achieved via bit encryption and recovery. Thus, we can convert the realization of the dual forensics framework EditGuard into a united image-bit steganography network. Our training objective is just a self-recovery mechanism, meaning it only needs to ensure the input and output of the steganography network maintain high fidelity under various robustness levels, with no need to introduce any labeled data or tampered samples. During inference, it can naturally locate tampering via simple comparisons in a zero-shot manner and extract copyright exactly." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 591, + 277, + 605 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 591, + 277, + 605 + ], + "spans": [ + { + "bbox": [ + 47, + 591, + 277, + 605 + ], + "type": "text", + "content": "4. United Image-bit Steganography Network" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 611, + 173, + 623 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 611, + 173, + 623 + ], + "spans": [ + { + "bbox": [ + 47, + 611, + 173, + 623 + ], + "type": "text", + "content": "4.1. Network Architecture" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 629, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 629, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 629, + 287, + 713 + ], + "type": "text", + "content": "As plotted in Fig. 4, the proposed IBSN includes an image hiding module (IHM), a bit encryption module (BEM), a bit recovery module (BRM), and an image revealing module (IRM). First, the IHM aims to hide a localization watermark " + }, + { + "bbox": [ + 46, + 629, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_{loc} \\in \\mathbb{R}^{H \\times W \\times 3}" + }, + { + "bbox": [ + 46, + 629, + 287, + 713 + ], + "type": "text", + "content": " into the original image " + }, + { + "bbox": [ + 46, + 629, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_{ori} \\in \\mathbb{R}^{H \\times W \\times 3}" + }, + { + "bbox": [ + 46, + 629, + 287, + 713 + ], + "type": "text", + "content": ", resulting in an intermediate output " + }, + { + "bbox": [ + 46, + 629, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_{med} \\in \\mathbb{R}^{H \\times W \\times 3}" + }, + { + "bbox": [ + 46, + 629, + 287, + 713 + ], + "type": "text", + "content": ". Subsequently, " + }, + { + "bbox": [ + 46, + 629, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_{med}" + }, + { + "bbox": [ + 46, + 629, + 287, + 713 + ], + "type": "text", + "content": " is fed to the BEM for feature refinement," + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 257, + 545, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 257, + 545, + 354 + ], + "spans": [ + { + "bbox": [ + 304, + 257, + 545, + 354 + ], + "type": "text", + "content": "while the copyright watermark " + }, + { + "bbox": [ + 304, + 257, + 545, + 354 + ], + "type": "inline_equation", + "content": "\\mathbf{w}_{cop} \\in \\{0,1\\}^L" + }, + { + "bbox": [ + 304, + 257, + 545, + 354 + ], + "type": "text", + "content": " is modulated into the BEM, forming the final container image " + }, + { + "bbox": [ + 304, + 257, + 545, + 354 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_{con} \\in \\mathbb{R}^{H \\times W \\times 3}" + }, + { + "bbox": [ + 304, + 257, + 545, + 354 + ], + "type": "text", + "content": ". After network transmission, the BRM will faithfully reconstruct the copyright watermark " + }, + { + "bbox": [ + 304, + 257, + 545, + 354 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{w}}_{cop}" + }, + { + "bbox": [ + 304, + 257, + 545, + 354 + ], + "type": "text", + "content": " from the received container image " + }, + { + "bbox": [ + 304, + 257, + 545, + 354 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_{rec}" + }, + { + "bbox": [ + 304, + 257, + 545, + 354 + ], + "type": "text", + "content": ". Meanwhile, " + }, + { + "bbox": [ + 304, + 257, + 545, + 354 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_{rec}" + }, + { + "bbox": [ + 304, + 257, + 545, + 354 + ], + "type": "text", + "content": " predicts the missing information " + }, + { + "bbox": [ + 304, + 257, + 545, + 354 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{Z}}" + }, + { + "bbox": [ + 304, + 257, + 545, + 354 + ], + "type": "text", + "content": " via the prompt-based posterior estimation and uses it as the initialization for the invertible blocks, producing " + }, + { + "bbox": [ + 304, + 257, + 545, + 354 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{I}}_{ori}" + }, + { + "bbox": [ + 304, + 257, + 545, + 354 + ], + "type": "text", + "content": " and semi-fragile watermark " + }, + { + "bbox": [ + 304, + 257, + 545, + 354 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{W}}_{loc}" + }, + { + "bbox": [ + 304, + 257, + 545, + 354 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 358, + 492, + 370 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 358, + 492, + 370 + ], + "spans": [ + { + "bbox": [ + 305, + 358, + 492, + 370 + ], + "type": "text", + "content": "4.2. Invertible Blocks in IHM and IRM" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 377, + 545, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 377, + 545, + 581 + ], + "spans": [ + { + "bbox": [ + 304, + 377, + 545, + 581 + ], + "type": "text", + "content": "Given the inherent capacity of flow-based models to precisely recover multimedia information, we harness stacked invertible blocks to construct image hiding and revealing modules. The original image " + }, + { + "bbox": [ + 304, + 377, + 545, + 581 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_{ori} \\in \\mathbb{R}^{H \\times W \\times 3}" + }, + { + "bbox": [ + 304, + 377, + 545, + 581 + ], + "type": "text", + "content": " and localization watermark " + }, + { + "bbox": [ + 304, + 377, + 545, + 581 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_{loc} \\in \\mathbb{R}^{H \\times W \\times 3}" + }, + { + "bbox": [ + 304, + 377, + 545, + 581 + ], + "type": "text", + "content": " will undergo discrete wavelet transformations (DWT) to yield frequency-decoupled image features. We then employ enhanced additive affine coupling layers to project the original image and its corresponding localization watermark branches. The transformation parameters are generated from each other. The enhanced affine coupling layer is composed of a five-layer dense convolution block [46] and a lightweight feature interaction module (LFIM) [4]. The LFIM can enhance the non-linearity of transformations and capture the long-range dependencies with low computational cost. More details are presented in S.M. Finally, the revealed features are then transformed to the image domain via the inverse DWT." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 587, + 494, + 600 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 587, + 494, + 600 + ], + "spans": [ + { + "bbox": [ + 305, + 587, + 494, + 600 + ], + "type": "text", + "content": "4.3. Prompt-based Posterior Estimation" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": "To bolster the fidelity and robustness of the image hiding and revealing module, we introduce a degradation prompt-based posteriori estimation module (PPEM). Since the encoding network tends to compress " + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\left[\\mathbf{I}_{ori};\\mathbf{W}_{loc}\\right] \\in \\mathbb{R}^{H \\times W \\times 6}" + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": " into the container image " + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_{con} \\in \\mathbb{R}^{H \\times W \\times 3}" + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": ", previous methods [42, 64] typically utilized a random Gaussian initialization or an all-zero map at the decoding end to compensate for the lost high-frequency channels. Yet, our observations suggest that discarded information lurks within the" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "11967" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 51, + 71, + 545, + 215 + ], + "blocks": [ + { + "bbox": [ + 51, + 71, + 545, + 215 + ], + "lines": [ + { + "bbox": [ + 51, + 71, + 545, + 215 + ], + "spans": [ + { + "bbox": [ + 51, + 71, + 545, + 215 + ], + "type": "image", + "image_path": "01b5c21e4e64812cbbbcd0b7eb5a9a74b9248afa151d2966ce8537cfa107bdb0.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 51, + 218, + 543, + 342 + ], + "blocks": [ + { + "bbox": [ + 51, + 218, + 543, + 342 + ], + "lines": [ + { + "bbox": [ + 51, + 218, + 543, + 342 + ], + "spans": [ + { + "bbox": [ + 51, + 218, + 543, + 342 + ], + "type": "image", + "image_path": "8d9ee22d3b133e6ac1c8906b3431073c5a2484c614038f51aa7f79bf152441c1.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 346, + 546, + 390 + ], + "lines": [ + { + "bbox": [ + 46, + 346, + 546, + 390 + ], + "spans": [ + { + "bbox": [ + 46, + 346, + 546, + 390 + ], + "type": "text", + "content": "Figure 4. Illustration of the united Image-bit Steganography Network (IBSN). In the training process, we randomly sample original image " + }, + { + "bbox": [ + 46, + 346, + 546, + 390 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_{ori}" + }, + { + "bbox": [ + 46, + 346, + 546, + 390 + ], + "type": "text", + "content": ", localization watermark " + }, + { + "bbox": [ + 46, + 346, + 546, + 390 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_{loc}" + }, + { + "bbox": [ + 46, + 346, + 546, + 390 + ], + "type": "text", + "content": " (a natural RGB image) and copyright watermark " + }, + { + "bbox": [ + 46, + 346, + 546, + 390 + ], + "type": "inline_equation", + "content": "\\mathbf{w}_{cop}" + }, + { + "bbox": [ + 46, + 346, + 546, + 390 + ], + "type": "text", + "content": " and expect the IBSN to recover " + }, + { + "bbox": [ + 46, + 346, + 546, + 390 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{I}}_{ori}" + }, + { + "bbox": [ + 46, + 346, + 546, + 390 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 346, + 546, + 390 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{W}}_{loc}" + }, + { + "bbox": [ + 46, + 346, + 546, + 390 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 346, + 546, + 390 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{w}}_{cop}" + }, + { + "bbox": [ + 46, + 346, + 546, + 390 + ], + "type": "text", + "content": " with high fidelity. In the inference process, we use core components of the pre-trained IBSN with a mask extractor (ME) to construct our EditGuard, and pre-define a simple solid color image as a localization watermark." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 47, + 395, + 287, + 513 + ], + "blocks": [ + { + "bbox": [ + 47, + 395, + 287, + 513 + ], + "lines": [ + { + "bbox": [ + 47, + 395, + 287, + 513 + ], + "spans": [ + { + "bbox": [ + 47, + 395, + 287, + 513 + ], + "type": "image", + "image_path": "9b662d95305022ca97a793ae622e6d31d1bc933c050499a507cebfed03184167.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 516, + 287, + 549 + ], + "lines": [ + { + "bbox": [ + 46, + 516, + 287, + 549 + ], + "spans": [ + { + "bbox": [ + 46, + 516, + 287, + 549 + ], + "type": "text", + "content": "Figure 5. Illustration of the proposed prompt-based posterior estimation. It will dynamically fuse degraded representations and extracted features to obtain posterior mean " + }, + { + "bbox": [ + 46, + 516, + 287, + 549 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{Z}} = \\mathbb{E}[\\mathbf{Z}|\\mathbf{I}_{rec}]" + }, + { + "bbox": [ + 46, + 516, + 287, + 549 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 552, + 287, + 634 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 552, + 287, + 634 + ], + "spans": [ + { + "bbox": [ + 46, + 552, + 287, + 634 + ], + "type": "text", + "content": "edges and textures of the container image. Thus, deploying a dedicated network proves to be a more potent strategy in predicting the posterior mean of the vanished localization watermark information " + }, + { + "bbox": [ + 46, + 552, + 287, + 634 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{Z}} = \\mathbb{E}[\\mathbf{Z}|\\mathbf{I}_{rec}]" + }, + { + "bbox": [ + 46, + 552, + 287, + 634 + ], + "type": "text", + "content": ". Specifically, as shown in Fig. 5, we stack " + }, + { + "bbox": [ + 46, + 552, + 287, + 634 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 46, + 552, + 287, + 634 + ], + "type": "text", + "content": " residual blocks " + }, + { + "bbox": [ + 46, + 552, + 287, + 634 + ], + "type": "inline_equation", + "content": "\\mathrm{Res}(\\cdot)" + }, + { + "bbox": [ + 46, + 552, + 287, + 634 + ], + "type": "text", + "content": " [19] and " + }, + { + "bbox": [ + 46, + 552, + 287, + 634 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 46, + 552, + 287, + 634 + ], + "type": "text", + "content": " channel-wise transformer blocks " + }, + { + "bbox": [ + 46, + 552, + 287, + 634 + ], + "type": "inline_equation", + "content": "\\mathrm{Trans}(\\cdot)" + }, + { + "bbox": [ + 46, + 552, + 287, + 634 + ], + "type": "text", + "content": " [70] to extract the local and non-local features " + }, + { + "bbox": [ + 46, + 552, + 287, + 634 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_c" + }, + { + "bbox": [ + 46, + 552, + 287, + 634 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 52, + 637, + 287, + 651 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 52, + 637, + 287, + 651 + ], + "spans": [ + { + "bbox": [ + 52, + 637, + 287, + 651 + ], + "type": "interline_equation", + "content": "\\mathbf {F} _ {c} = \\operatorname {T r a n s} (\\operatorname {R e s} (\\mathrm {D W T} (\\mathbf {I} _ {r e c}))) + \\operatorname {R e s} (\\mathrm {D W T} (\\mathbf {I} _ {r e c})). \\tag {2}", + "image_path": "b3cf7ed26bc228049b220cdd0a24fa913ddccd3e4b5f0c1f6a3cf05f770ae9f4.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 654, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 654, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 654, + 287, + 713 + ], + "type": "text", + "content": "Considering that the container image is prone to various degradations during network transmission, we pre-define " + }, + { + "bbox": [ + 47, + 654, + 287, + 713 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 47, + 654, + 287, + 713 + ], + "type": "text", + "content": " learnable embedding tensors as degradation prompts " + }, + { + "bbox": [ + 47, + 654, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{P} = [\\mathbf{P}_1,\\mathbf{P}_2,\\dots ,\\mathbf{P}_N]" + }, + { + "bbox": [ + 47, + 654, + 287, + 713 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 47, + 654, + 287, + 713 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 47, + 654, + 287, + 713 + ], + "type": "text", + "content": " denotes the number of degradation types and is set to 3. These learned prompts " + }, + { + "bbox": [ + 47, + 654, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{P}" + }, + { + "bbox": [ + 47, + 654, + 287, + 713 + ], + "type": "text", + "content": " can" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 396, + 545, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 396, + 545, + 540 + ], + "spans": [ + { + "bbox": [ + 304, + 396, + 545, + 540 + ], + "type": "text", + "content": "adaptively learn a diverse range of degradation representations and are integrated with the intrinsic features extracted from " + }, + { + "bbox": [ + 304, + 396, + 545, + 540 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_{rec}" + }, + { + "bbox": [ + 304, + 396, + 545, + 540 + ], + "type": "text", + "content": ", enabling the proposed IBSN to handle multiple types of degradations using a single set of parameters. To better foster the interaction between the input features " + }, + { + "bbox": [ + 304, + 396, + 545, + 540 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_c" + }, + { + "bbox": [ + 304, + 396, + 545, + 540 + ], + "type": "text", + "content": " and the degradation prompt " + }, + { + "bbox": [ + 304, + 396, + 545, + 540 + ], + "type": "inline_equation", + "content": "\\mathbf{P}" + }, + { + "bbox": [ + 304, + 396, + 545, + 540 + ], + "type": "text", + "content": ", the features " + }, + { + "bbox": [ + 304, + 396, + 545, + 540 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_c" + }, + { + "bbox": [ + 304, + 396, + 545, + 540 + ], + "type": "text", + "content": " are passed to a global average pooling (GAP) layer, a " + }, + { + "bbox": [ + 304, + 396, + 545, + 540 + ], + "type": "inline_equation", + "content": "1\\times 1" + }, + { + "bbox": [ + 304, + 396, + 545, + 540 + ], + "type": "text", + "content": " convolution, and a softmax operator to produce a set of dynamic weight coefficients. Each degradation prompt " + }, + { + "bbox": [ + 304, + 396, + 545, + 540 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_i" + }, + { + "bbox": [ + 304, + 396, + 545, + 540 + ], + "type": "text", + "content": " is combined using these dynamic coefficients " + }, + { + "bbox": [ + 304, + 396, + 545, + 540 + ], + "type": "inline_equation", + "content": "\\mathbf{w}_{p\\otimes i}" + }, + { + "bbox": [ + 304, + 396, + 545, + 540 + ], + "type": "text", + "content": " and subsequently integrated via an upsampling operator " + }, + { + "bbox": [ + 304, + 396, + 545, + 540 + ], + "type": "inline_equation", + "content": "\\uparrow" + }, + { + "bbox": [ + 304, + 396, + 545, + 540 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 396, + 545, + 540 + ], + "type": "inline_equation", + "content": "3\\times 3" + }, + { + "bbox": [ + 304, + 396, + 545, + 540 + ], + "type": "text", + "content": " convolution to obtain the enhanced representation " + }, + { + "bbox": [ + 304, + 396, + 545, + 540 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_c" + }, + { + "bbox": [ + 304, + 396, + 545, + 540 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 344, + 543, + 545, + 566 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 344, + 543, + 545, + 566 + ], + "spans": [ + { + "bbox": [ + 344, + 543, + 545, + 566 + ], + "type": "interline_equation", + "content": "\\mathbf {P} _ {c} = \\operatorname {C o n v} _ {3 \\times 3} \\left(\\left(\\sum_ {i = 1} ^ {N} \\mathbf {w} _ {p * i} \\mathbf {P} _ {i}\\right) _ {\\uparrow}\\right), \\tag {3}", + "image_path": "e93a154bdd5bac01ef0fa350fb25cfd4f68aa9250240ecd0ec5b11e5994005af.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 318, + 567, + 518, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 567, + 518, + 579 + ], + "spans": [ + { + "bbox": [ + 318, + 567, + 518, + 579 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 318, + 567, + 518, + 579 + ], + "type": "inline_equation", + "content": "\\mathbf{w}_p = \\mathrm{Softmax}\\left(\\mathrm{Conv}_{1\\times 1}\\left(\\mathrm{GAP}\\left(\\mathbf{F}_c\\right)\\right)\\right)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 583, + 545, + 631 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 583, + 545, + 631 + ], + "spans": [ + { + "bbox": [ + 304, + 583, + 545, + 631 + ], + "type": "text", + "content": "Finally, we utilize a " + }, + { + "bbox": [ + 304, + 583, + 545, + 631 + ], + "type": "inline_equation", + "content": "3 \\times 3" + }, + { + "bbox": [ + 304, + 583, + 545, + 631 + ], + "type": "text", + "content": " convolution to fuse the learned degradation representation conditioned on the extracted features " + }, + { + "bbox": [ + 304, + 583, + 545, + 631 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_c" + }, + { + "bbox": [ + 304, + 583, + 545, + 631 + ], + "type": "text", + "content": " to enrich the degradation-specific context, obtaining " + }, + { + "bbox": [ + 304, + 583, + 545, + 631 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{Z}}" + }, + { + "bbox": [ + 304, + 583, + 545, + 631 + ], + "type": "text", + "content": ". This process can be formulated as:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 343, + 634, + 545, + 649 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 343, + 634, + 545, + 649 + ], + "spans": [ + { + "bbox": [ + 343, + 634, + 545, + 649 + ], + "type": "interline_equation", + "content": "\\hat {\\mathbf {Z}} = \\operatorname {C o n v} _ {3 \\times 3} \\left(\\left[ \\mathbf {P} _ {c}; \\mathbf {F} _ {c} \\right]\\right) \\in \\mathbb {R} ^ {\\frac {H}{2} \\times \\frac {W}{2} \\times 1 2}. \\tag {4}", + "image_path": "18a19711c0825bdfd33652ba0a6fff710894fc3d63a3dc0acad5c97968244db8.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 658, + 507, + 671 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 658, + 507, + 671 + ], + "spans": [ + { + "bbox": [ + 306, + 658, + 507, + 671 + ], + "type": "text", + "content": "4.4. Bit Encryption and Recovery Modules" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 677, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 677, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 677, + 545, + 713 + ], + "type": "text", + "content": "As shown in Fig. 4, to encode the copyright watermark " + }, + { + "bbox": [ + 306, + 677, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{w}_{\\text{cop}}" + }, + { + "bbox": [ + 306, + 677, + 545, + 713 + ], + "type": "text", + "content": " into " + }, + { + "bbox": [ + 306, + 677, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_{\\text{med}}" + }, + { + "bbox": [ + 306, + 677, + 545, + 713 + ], + "type": "text", + "content": ", we firstly expand " + }, + { + "bbox": [ + 306, + 677, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{w}_{\\text{cop}} \\in \\{0,1\\}^L" + }, + { + "bbox": [ + 306, + 677, + 545, + 713 + ], + "type": "text", + "content": " via stacked MLPs and reshape it into several " + }, + { + "bbox": [ + 306, + 677, + 545, + 713 + ], + "type": "inline_equation", + "content": "L \\times L" + }, + { + "bbox": [ + 306, + 677, + 545, + 713 + ], + "type": "text", + "content": " message feature" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "11968" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 49, + 97, + 287, + 176 + ], + "blocks": [ + { + "bbox": [ + 47, + 70, + 287, + 92 + ], + "lines": [ + { + "bbox": [ + 47, + 70, + 287, + 92 + ], + "spans": [ + { + "bbox": [ + 47, + 70, + 287, + 92 + ], + "type": "text", + "content": "Table 1. Localization precision (F1-Score) and bit accuracy (BA) comparison with other competitive methods on [9, 16, 20, 58]." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 49, + 97, + 287, + 176 + ], + "lines": [ + { + "bbox": [ + 49, + 97, + 287, + 176 + ], + "spans": [ + { + "bbox": [ + 49, + 97, + 287, + 176 + ], + "type": "table", + "html": "
MethodCAISAv1 [9]Coverage [58]NIST16 [16]Columbia [20]
F1BA(%)F1BA(%)F1BA(%)F1BA(%)
ManTraNet [63]0.320-0.486-0.225-0.650-
SPAN [22]0.169-0.428-0.363-0.873-
CAT-Net v2 [32]0.852-0.582-0.417-0.923-
OSN [61]0.676-0.472-0.449-0.836-
MVSS-Net [8]0.650-0.659-0.372-0.781-
PSCC-Net [39]0.670-0.615-0.210-0.760-
TruFor [17]0.822-0.735-0.470-0.914-
EditGuard (Ours)0.95499.910.9551000.91199.880.98899.93
", + "image_path": "d444b17703fece8231bf655fa0b91478d28d1ef369d55100bf1e0b8b801db008.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 50, + 180, + 286, + 308 + ], + "blocks": [ + { + "bbox": [ + 50, + 180, + 286, + 308 + ], + "lines": [ + { + "bbox": [ + 50, + 180, + 286, + 308 + ], + "spans": [ + { + "bbox": [ + 50, + 180, + 286, + 308 + ], + "type": "image", + "image_path": "ac93b3cda087d4ec26260b5d1f14f6c77c235ad6139be81c5a78ef10a189739f.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 311, + 287, + 334 + ], + "lines": [ + { + "bbox": [ + 46, + 311, + 287, + 334 + ], + "spans": [ + { + "bbox": [ + 46, + 311, + 287, + 334 + ], + "type": "text", + "content": "Figure 6. Localization precision comparisons of our EditGuard and competitive methods [8, 39, 61] on four classical benchmarks." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 337, + 289, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 337, + 289, + 445 + ], + "spans": [ + { + "bbox": [ + 46, + 337, + 289, + 445 + ], + "type": "text", + "content": "maps. Meanwhile, " + }, + { + "bbox": [ + 46, + 337, + 289, + 445 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_{med}" + }, + { + "bbox": [ + 46, + 337, + 289, + 445 + ], + "type": "text", + "content": " is fed to a U-style feature enhancement network to extract features of each downsampling and upsampling layer. Finally, the message features will be upscaled and integrated with multi-level image features via the fusion mechanism [21, 62], achieving the modulation of bit-image information. In the decoding end, " + }, + { + "bbox": [ + 46, + 337, + 289, + 445 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_{rec}" + }, + { + "bbox": [ + 46, + 337, + 289, + 445 + ], + "type": "text", + "content": " is fed into a U-shaped sub-network and downsampled to a size of " + }, + { + "bbox": [ + 46, + 337, + 289, + 445 + ], + "type": "inline_equation", + "content": "L\\times L" + }, + { + "bbox": [ + 46, + 337, + 289, + 445 + ], + "type": "text", + "content": ". The recovered copyright watermark " + }, + { + "bbox": [ + 46, + 337, + 289, + 445 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{w}}_{cop}" + }, + { + "bbox": [ + 46, + 337, + 289, + 445 + ], + "type": "text", + "content": " is then extracted via an MLP. More details are presented in S.M.." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 451, + 234, + 463 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 451, + 234, + 463 + ], + "spans": [ + { + "bbox": [ + 47, + 451, + 234, + 463 + ], + "type": "text", + "content": "4.5. Construct EditGuard via the IBSN" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 469, + 287, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 469, + 287, + 518 + ], + "spans": [ + { + "bbox": [ + 46, + 469, + 287, + 518 + ], + "type": "text", + "content": "To stabilize the optimization of the proposed IBSN, we propose a bi-level optimization strategy. Given an arbitrary original image " + }, + { + "bbox": [ + 46, + 469, + 287, + 518 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_{med}" + }, + { + "bbox": [ + 46, + 469, + 287, + 518 + ], + "type": "text", + "content": " and watermark " + }, + { + "bbox": [ + 46, + 469, + 287, + 518 + ], + "type": "inline_equation", + "content": "\\mathbf{w}_{cop}" + }, + { + "bbox": [ + 46, + 469, + 287, + 518 + ], + "type": "text", + "content": ", we first train the bit encryption and recovery module via the " + }, + { + "bbox": [ + 46, + 469, + 287, + 518 + ], + "type": "inline_equation", + "content": "\\ell_{2}" + }, + { + "bbox": [ + 46, + 469, + 287, + 518 + ], + "type": "text", + "content": " loss." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 73, + 520, + 287, + 536 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 73, + 520, + 287, + 536 + ], + "spans": [ + { + "bbox": [ + 73, + 520, + 287, + 536 + ], + "type": "interline_equation", + "content": "\\ell_ {c o p} = \\left\\| \\mathbf {I} _ {c o n} - \\mathbf {I} _ {m e d} \\right\\| _ {2} ^ {2} + \\lambda \\left\\| \\hat {\\mathbf {w}} _ {c o p} - \\mathbf {w} _ {c o p} \\right\\| _ {2} ^ {2}, \\quad (5)", + "image_path": "fdb3677dbe334335eb819932e151681e7e946f888a82bbe6f17cfe9bf4a5b02f.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 537, + 287, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 537, + 287, + 586 + ], + "spans": [ + { + "bbox": [ + 46, + 537, + 287, + 586 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 537, + 287, + 586 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 46, + 537, + 287, + 586 + ], + "type": "text", + "content": " is set to 10. Furthermore, we freeze the weights of BEM and BRM and jointly train the IHM and IRM. Given a random original image " + }, + { + "bbox": [ + 46, + 537, + 287, + 586 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_{ori}" + }, + { + "bbox": [ + 46, + 537, + 287, + 586 + ], + "type": "text", + "content": ", localization watermark " + }, + { + "bbox": [ + 46, + 537, + 287, + 586 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_{loc}" + }, + { + "bbox": [ + 46, + 537, + 287, + 586 + ], + "type": "text", + "content": " and copyright watermark " + }, + { + "bbox": [ + 46, + 537, + 287, + 586 + ], + "type": "inline_equation", + "content": "\\mathbf{w}_{cop}" + }, + { + "bbox": [ + 46, + 537, + 287, + 586 + ], + "type": "text", + "content": ", the loss function is:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 51, + 590, + 287, + 602 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 51, + 590, + 287, + 602 + ], + "spans": [ + { + "bbox": [ + 51, + 590, + 287, + 602 + ], + "type": "interline_equation", + "content": "\\ell_ {l o c} = \\left\\| \\hat {\\mathbf {I}} _ {o r i} - \\mathbf {I} _ {o r i} \\right\\| _ {1} + \\alpha \\left\\| \\mathbf {I} _ {c o n} - \\mathbf {I} _ {o r i} \\right\\| _ {2} ^ {2} + \\beta \\left\\| \\hat {\\mathbf {W}} _ {l o c} - \\mathbf {W} _ {l o c} \\right\\| _ {1}, \\tag {6}", + "image_path": "165fa326fbf5f0ca2e9d3cb28a4db48e4c4033acd694a6cf04c692433102abf5.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 605, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 288, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 605, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 46, + 605, + 288, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 605, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 46, + 605, + 288, + 714 + ], + "type": "text", + "content": " are respectively set to 100 and 1. During training, we only introduced degradations to " + }, + { + "bbox": [ + 46, + 605, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_{con}" + }, + { + "bbox": [ + 46, + 605, + 288, + 714 + ], + "type": "text", + "content": " without being exposed to any tampering. After acquiring a pretrained IBSN, we can construct the proposed EditGuard via the components of IBSN. As plotted in Fig. 4, the dual-watermark encoder of EditGuard is composed of IHM and BEM, which correspond to the localization and copyright watermarking in Fig. 3 respectively. The copyright extractor strictly corresponds to BRM. The tamper locator" + } + ] + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 308, + 96, + 545, + 161 + ], + "blocks": [ + { + "bbox": [ + 305, + 70, + 545, + 93 + ], + "lines": [ + { + "bbox": [ + 305, + 70, + 545, + 93 + ], + "spans": [ + { + "bbox": [ + 305, + 70, + 545, + 93 + ], + "type": "text", + "content": "Table 2. Visual quality of the container image " + }, + { + "bbox": [ + 305, + 70, + 545, + 93 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_{con}" + }, + { + "bbox": [ + 305, + 70, + 545, + 93 + ], + "type": "text", + "content": " and bit accuracy comparison with other pure watermarking methods." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 308, + 96, + 545, + 161 + ], + "lines": [ + { + "bbox": [ + 308, + 96, + 545, + 161 + ], + "spans": [ + { + "bbox": [ + 308, + 96, + 545, + 161 + ], + "type": "table", + "html": "
MethodImage SizeM. L.PSNR (dB)SSIMNIQE(↓)BA(%)
MBRS [26]128×1283026.570.8867.219100
CIN [44]128×1283041.350.9817.17199.99
PIMoG [10]128×1283036.220.9417.11399.99
SepMark [62]128×1283035.420.9317.09599.86
EditGuard128×1283036.930.9445.56799.89
EditGuard512×5126437.770.9494.25799.95
", + "image_path": "533fae2f42952dbc6a7275fde1683a63874f847939ae1901f747d96464322e1b.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_body" + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 164, + 545, + 248 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 164, + 545, + 248 + ], + "spans": [ + { + "bbox": [ + 304, + 164, + 545, + 248 + ], + "type": "text", + "content": "includes IRM and a mask extractor (ME). Note that we need to pre-define a localization watermark " + }, + { + "bbox": [ + 304, + 164, + 545, + 248 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_{loc}" + }, + { + "bbox": [ + 304, + 164, + 545, + 248 + ], + "type": "text", + "content": ", which is shared between the encoding and decoding ends. The choice of " + }, + { + "bbox": [ + 304, + 164, + 545, + 248 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_{loc}" + }, + { + "bbox": [ + 304, + 164, + 545, + 248 + ], + "type": "text", + "content": " is very general to our method. It can be any natural image or even a solid color image. Finally, by comparing the pre-defined watermark " + }, + { + "bbox": [ + 304, + 164, + 545, + 248 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_{loc}" + }, + { + "bbox": [ + 304, + 164, + 545, + 248 + ], + "type": "text", + "content": " with the decoded one " + }, + { + "bbox": [ + 304, + 164, + 545, + 248 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{W}}_{loc}" + }, + { + "bbox": [ + 304, + 164, + 545, + 248 + ], + "type": "text", + "content": ", we can obtain a binary mask " + }, + { + "bbox": [ + 304, + 164, + 545, + 248 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{M}}\\in \\mathbb{R}^{H\\times W}" + }, + { + "bbox": [ + 304, + 164, + 545, + 248 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 317, + 251, + 545, + 265 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 251, + 545, + 265 + ], + "spans": [ + { + "bbox": [ + 317, + 251, + 545, + 265 + ], + "type": "interline_equation", + "content": "\\hat {\\mathbf {M}} [ i, j ] = \\theta_ {\\tau} \\left(\\max \\left(\\left| \\hat {\\mathbf {W}} _ {l o c} [ i, j,: ] - \\mathbf {W} _ {l o c} [ i, j,: ] \\right|\\right)\\right). \\tag {7}", + "image_path": "0157f56b871540f11920a0a487c83779cab02052ef15ac19b3fc20a7cfb0d44d.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 267, + 545, + 293 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 267, + 545, + 293 + ], + "spans": [ + { + "bbox": [ + 304, + 267, + 545, + 293 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 267, + 545, + 293 + ], + "type": "inline_equation", + "content": "i\\in [0,H)" + }, + { + "bbox": [ + 304, + 267, + 545, + 293 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 267, + 545, + 293 + ], + "type": "inline_equation", + "content": "j\\in [0,W)" + }, + { + "bbox": [ + 304, + 267, + 545, + 293 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 304, + 267, + 545, + 293 + ], + "type": "inline_equation", + "content": "\\theta_{\\tau}(z) = 1" + }, + { + "bbox": [ + 304, + 267, + 545, + 293 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 304, + 267, + 545, + 293 + ], + "type": "inline_equation", + "content": "z\\geq \\tau" + }, + { + "bbox": [ + 304, + 267, + 545, + 293 + ], + "type": "text", + "content": "). " + }, + { + "bbox": [ + 304, + 267, + 545, + 293 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 304, + 267, + 545, + 293 + ], + "type": "text", + "content": " is set to 0.2. " + }, + { + "bbox": [ + 304, + 267, + 545, + 293 + ], + "type": "inline_equation", + "content": "|\\cdot |" + }, + { + "bbox": [ + 304, + 267, + 545, + 293 + ], + "type": "text", + "content": " is an absolute value operation." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 303, + 388, + 316 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 303, + 388, + 316 + ], + "spans": [ + { + "bbox": [ + 306, + 303, + 388, + 316 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 323, + 440, + 335 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 323, + 440, + 335 + ], + "spans": [ + { + "bbox": [ + 306, + 323, + 440, + 335 + ], + "type": "text", + "content": "5.1. Implementation Details" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 341, + 545, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 341, + 545, + 509 + ], + "spans": [ + { + "bbox": [ + 304, + 341, + 545, + 509 + ], + "type": "text", + "content": "We trained our EditGuard via the training set of COCO [38] without any tampered data. Thus, for tamper localization, our method is actually zero-shot. The Adam [30] is used for training " + }, + { + "bbox": [ + 304, + 341, + 545, + 509 + ], + "type": "inline_equation", + "content": "250K" + }, + { + "bbox": [ + 304, + 341, + 545, + 509 + ], + "type": "text", + "content": " iterations with " + }, + { + "bbox": [ + 304, + 341, + 545, + 509 + ], + "type": "inline_equation", + "content": "\\beta_{1} = 0.9" + }, + { + "bbox": [ + 304, + 341, + 545, + 509 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 341, + 545, + 509 + ], + "type": "inline_equation", + "content": "\\beta_{2} = 0.5" + }, + { + "bbox": [ + 304, + 341, + 545, + 509 + ], + "type": "text", + "content": ". The learning rate is initialized to " + }, + { + "bbox": [ + 304, + 341, + 545, + 509 + ], + "type": "inline_equation", + "content": "1\\times 10^{-4}" + }, + { + "bbox": [ + 304, + 341, + 545, + 509 + ], + "type": "text", + "content": " and decreases by half for every " + }, + { + "bbox": [ + 304, + 341, + 545, + 509 + ], + "type": "inline_equation", + "content": "30K" + }, + { + "bbox": [ + 304, + 341, + 545, + 509 + ], + "type": "text", + "content": " iterations, with the batch size set to 4. We embed a 64-bit copyright watermark and a simple localization watermark such as a pure blue image ([R, G, B] = [0, 0, 255]) to original images. Following [8, 17, 39], F1-score, AUC, IoU, and bit accuracy are used to evaluate localization and copyright protection performance. Since no prior methods can simultaneously achieve this dual forensics, we conducted separate comparisons with tamper localization and image watermarking methods." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 306, + 515, + 515, + 528 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 515, + 515, + 528 + ], + "spans": [ + { + "bbox": [ + 306, + 515, + 515, + 528 + ], + "type": "text", + "content": "5.2. Comparison with Localization Methods" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 534, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 534, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 534, + 545, + 713 + ], + "type": "text", + "content": "For a fair comparison with tamper localization methods, we conducted extensive evaluations on four classical benchmarks [9, 16, 20, 58], as reported on Tab. 1. Since EditGuard is a proactive approach, we initially embed watermarks into authentic images and then paste the tampered areas into the container images. Remarkably, even for tamper types that existing methods specialize in, the localization accuracy of EditGuard consistently outperforms the SOTA method [17] across four datasets by margins of 0.102, 0.116, 0.441, and 0.065 in F1-score without any labeled data or tampered samples required, which verifies the superiority of our proactive localization mechanism. As shown in Fig. 6, our EditGuard can precisely pinpoint pixel-level tampered areas but other methods can only produce a rough outline or are only effective in some cases. Mean-" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "11969" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 49, + 92, + 545, + 154 + ], + "blocks": [ + { + "bbox": [ + 47, + 71, + 545, + 91 + ], + "lines": [ + { + "bbox": [ + 47, + 71, + 545, + 91 + ], + "spans": [ + { + "bbox": [ + 47, + 71, + 545, + 91 + ], + "type": "text", + "content": "Table 3. Comparison with other competitive tamper localization methods under different AIGC-based editing methods. Note that " + }, + { + "bbox": [ + 47, + 71, + 545, + 91 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 47, + 71, + 545, + 91 + ], + "type": "text", + "content": " denotes the network finetuned in our constructed AGE-Set-C." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 49, + 92, + 545, + 154 + ], + "lines": [ + { + "bbox": [ + 49, + 92, + 545, + 154 + ], + "spans": [ + { + "bbox": [ + 49, + 92, + 545, + 154 + ], + "type": "table", + "html": "
MethodStable Diffusion Inpaint [51]Controlnet [72]SDXL [49]RePaint [43]Lama [57]FaceSwap [60]
F1AUCIoUBA(%)F1AUCIoUBA(%)F1AUCIoUBA(%)F1AUCIoUBA(%)F1AUCIoUBA(%)F1AUCIoUBA(%)
MVSS-Net [8]0.1780.4880.103-0.1780.4920.103-0.0370.5030.028-0.1040.5460.082-0.0240.5050.022-0.2850.6120.192-
OSN [61]0.1740.4860.101-0.1910.6440.110-0.2000.7550.118-0.1830.6440.105-0.1700.4300.099-0.3080.7910.171-
PSCC-Net [39]0.1660.5010.112-0.1770.5650.116-0.1890.7040.115-0.1400.4690.109-0.1320.3290.104-0.1570.3460.180-
IML-VIT [45]0.2130.5960.135-0.2000.5760.128-0.2210.6030.145-0.1030.4970.059-0.1050.4650.064-0.1050.4650.064-
HiFi-Net [18]0.5470.7340.128-0.5420.7350.123-0.6330.8280.261-0.6810.8960.339-0.4830.7210.029-0.7810.8900.478-
MVSS-Net†[8]0.6940.9390.575-0.6780.9250.558-0.4820.8840.359-0.1850.5290.111-0.3930.8290.275-0.4590.7390.333-
EditGuard (Ours)0.9660.9710.93699.950.9680.9870.94099.960.9650.9890.93699.960.9670.9770.93899.950.9650.9690.93499.950.8960.9430.87699.86
", + "image_path": "ce030315e7bf53ca422742f2f12f38fb43f40f250dbab4c1d47c813512e4ca66.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 50, + 159, + 99, + 203 + ], + "blocks": [ + { + "bbox": [ + 50, + 159, + 99, + 203 + ], + "lines": [ + { + "bbox": [ + 50, + 159, + 99, + 203 + ], + "spans": [ + { + "bbox": [ + 50, + 159, + 99, + 203 + ], + "type": "image", + "image_path": "95875c5849a1cdbc6277286d2e769a7f1bd18d6ddaca9783cff1641b55880ea9.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 100, + 159, + 145, + 204 + ], + "blocks": [ + { + "bbox": [ + 100, + 159, + 145, + 204 + ], + "lines": [ + { + "bbox": [ + 100, + 159, + 145, + 204 + ], + "spans": [ + { + "bbox": [ + 100, + 159, + 145, + 204 + ], + "type": "image", + "image_path": "e42e68e144df1134327fbfcd2269b66657f377933b8c3fc5b2785224624d2386.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 148, + 159, + 193, + 204 + ], + "blocks": [ + { + "bbox": [ + 148, + 159, + 193, + 204 + ], + "lines": [ + { + "bbox": [ + 148, + 159, + 193, + 204 + ], + "spans": [ + { + "bbox": [ + 148, + 159, + 193, + 204 + ], + "type": "image", + "image_path": "fa30e311be2c9f3547a7b8c2161d1dfc5bf76dbf7f6ebb1f737811bf611765dc.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 194, + 159, + 239, + 204 + ], + "blocks": [ + { + "bbox": [ + 194, + 159, + 239, + 204 + ], + "lines": [ + { + "bbox": [ + 194, + 159, + 239, + 204 + ], + "spans": [ + { + "bbox": [ + 194, + 159, + 239, + 204 + ], + "type": "image", + "image_path": "83cb8d90245a0c0fdedf521e2b149960657dc5bae5a30eee4f340c70cf4e5e8e.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 241, + 159, + 286, + 204 + ], + "blocks": [ + { + "bbox": [ + 241, + 159, + 286, + 204 + ], + "lines": [ + { + "bbox": [ + 241, + 159, + 286, + 204 + ], + "spans": [ + { + "bbox": [ + 241, + 159, + 286, + 204 + ], + "type": "image", + "image_path": "4bd55197d7c7b580de6b9d301db508e36d01df31e4ac53d4df4977dd2649048d.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 49, + 206, + 99, + 251 + ], + "blocks": [ + { + "bbox": [ + 49, + 206, + 99, + 251 + ], + "lines": [ + { + "bbox": [ + 49, + 206, + 99, + 251 + ], + "spans": [ + { + "bbox": [ + 49, + 206, + 99, + 251 + ], + "type": "image", + "image_path": "1485a6728911b0188fb0bf86050c79bdae5b51c24a940aeb06082d5c4eb82843.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 101, + 206, + 145, + 251 + ], + "blocks": [ + { + "bbox": [ + 101, + 206, + 145, + 251 + ], + "lines": [ + { + "bbox": [ + 101, + 206, + 145, + 251 + ], + "spans": [ + { + "bbox": [ + 101, + 206, + 145, + 251 + ], + "type": "image", + "image_path": "dc9ae763a7b3ccf0c9aaae7c79651e92be21477f8d2253256501671c9512a36f.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 148, + 206, + 193, + 251 + ], + "blocks": [ + { + "bbox": [ + 148, + 206, + 193, + 251 + ], + "lines": [ + { + "bbox": [ + 148, + 206, + 193, + 251 + ], + "spans": [ + { + "bbox": [ + 148, + 206, + 193, + 251 + ], + "type": "image", + "image_path": "41c00b7712cda2d2a90ef7fd3b6fc93bb44371f993c44847f11653440f36d9d4.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 194, + 206, + 239, + 251 + ], + "blocks": [ + { + "bbox": [ + 194, + 206, + 239, + 251 + ], + "lines": [ + { + "bbox": [ + 194, + 206, + 239, + 251 + ], + "spans": [ + { + "bbox": [ + 194, + 206, + 239, + 251 + ], + "type": "image", + "image_path": "9b23195e6a3bd779476e07957bf8724e8326a4245ca6915e44f1728542337b0f.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 241, + 206, + 286, + 251 + ], + "blocks": [ + { + "bbox": [ + 241, + 206, + 286, + 251 + ], + "lines": [ + { + "bbox": [ + 241, + 206, + 286, + 251 + ], + "spans": [ + { + "bbox": [ + 241, + 206, + 286, + 251 + ], + "type": "image", + "image_path": "efb52bc2d62e35da7636a0034a91a419b7900d3bb22e746ef6b36778ea57a0f2.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 49, + 253, + 99, + 297 + ], + "blocks": [ + { + "bbox": [ + 49, + 253, + 99, + 297 + ], + "lines": [ + { + "bbox": [ + 49, + 253, + 99, + 297 + ], + "spans": [ + { + "bbox": [ + 49, + 253, + 99, + 297 + ], + "type": "image", + "image_path": "5ac68cb476e3c8b7111c8ea60d546a3b431b987f0854fb19fbea659d4803e0dd.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 47, + 300, + 287, + 334 + ], + "lines": [ + { + "bbox": [ + 47, + 300, + 287, + 334 + ], + "spans": [ + { + "bbox": [ + 47, + 300, + 287, + 334 + ], + "type": "text", + "content": "Figure 7. Visual results of the container image " + }, + { + "bbox": [ + 47, + 300, + 287, + 334 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_{con}" + }, + { + "bbox": [ + 47, + 300, + 287, + 334 + ], + "type": "text", + "content": " and the error map of the proposed EditGuard. Here, localization and copyright watermarks are randomly selected from the dataset." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 101, + 253, + 145, + 297 + ], + "blocks": [ + { + "bbox": [ + 101, + 253, + 145, + 297 + ], + "lines": [ + { + "bbox": [ + 101, + 253, + 145, + 297 + ], + "spans": [ + { + "bbox": [ + 101, + 253, + 145, + 297 + ], + "type": "image", + "image_path": "e1e1c8519147b61a444e57c089b3f515b3409cfed275c56496039ec91406b4d5.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 148, + 254, + 193, + 297 + ], + "blocks": [ + { + "bbox": [ + 148, + 254, + 193, + 297 + ], + "lines": [ + { + "bbox": [ + 148, + 254, + 193, + 297 + ], + "spans": [ + { + "bbox": [ + 148, + 254, + 193, + 297 + ], + "type": "image", + "image_path": "450cda36558f90764ebfd9de06e5bbc54b1e1324e4813945fa25c2f9cb66e9ef.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 194, + 254, + 239, + 297 + ], + "blocks": [ + { + "bbox": [ + 194, + 254, + 239, + 297 + ], + "lines": [ + { + "bbox": [ + 194, + 254, + 239, + 297 + ], + "spans": [ + { + "bbox": [ + 194, + 254, + 239, + 297 + ], + "type": "image", + "image_path": "ef1ebd48de8354d61bbaa44b137bf3f916d54f689186b9b4e920e1288961fd98.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 241, + 254, + 286, + 297 + ], + "blocks": [ + { + "bbox": [ + 241, + 254, + 286, + 297 + ], + "lines": [ + { + "bbox": [ + 241, + 254, + 286, + 297 + ], + "spans": [ + { + "bbox": [ + 241, + 254, + 286, + 297 + ], + "type": "image", + "image_path": "a70af4c7c227f03986671dd12dd5b3528f9fc16cc7ba5282eb3468b33973902c.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "bbox": [ + 47, + 336, + 287, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 336, + 287, + 361 + ], + "spans": [ + { + "bbox": [ + 47, + 336, + 287, + 361 + ], + "type": "text", + "content": "while, our bit accuracy remains over " + }, + { + "bbox": [ + 47, + 336, + 287, + 361 + ], + "type": "inline_equation", + "content": "99.8\\%" + }, + { + "bbox": [ + 47, + 336, + 287, + 361 + ], + "type": "text", + "content": " while all other methods can not realize effective copyright protection." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 47, + 369, + 266, + 381 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 369, + 266, + 381 + ], + "spans": [ + { + "bbox": [ + 47, + 369, + 266, + 381 + ], + "type": "text", + "content": "5.3. Comparison with Watermarking Methods" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 46, + 387, + 287, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 387, + 287, + 639 + ], + "spans": [ + { + "bbox": [ + 46, + 387, + 287, + 639 + ], + "type": "text", + "content": "To evaluate the visual quality of " + }, + { + "bbox": [ + 46, + 387, + 287, + 639 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_{con}" + }, + { + "bbox": [ + 46, + 387, + 287, + 639 + ], + "type": "text", + "content": ", we compared EditGuard with other watermarking methods on " + }, + { + "bbox": [ + 46, + 387, + 287, + 639 + ], + "type": "inline_equation", + "content": "1K" + }, + { + "bbox": [ + 46, + 387, + 287, + 639 + ], + "type": "text", + "content": " testing images from COCO [38] under the tampering of stable diffusion inpaint [51]. For a fair comparison, we also retrained our EditGuard on " + }, + { + "bbox": [ + 46, + 387, + 287, + 639 + ], + "type": "inline_equation", + "content": "128\\times 128" + }, + { + "bbox": [ + 46, + 387, + 287, + 639 + ], + "type": "text", + "content": " original images and 30 bits. Tab. 2 reports that the fidelity of our container image far surpasses that of SepMark [62], PIMoG [10], and MBRS [26] but is slightly inferior to CIN [44]. Meanwhile, our method exhibits the best performance in perceptual quality measures like NIQE. As shown in Fig. 7, dual-watermarked images do not have noticeable artifacts and noise, making them imperceptible to the human eyes. When suffer malicious tampering, our method outperforms SepMark and is very close to PIMoG and CIN in bit accuracy. Note that other competitive methods only hide 30 bits, with a capacity of " + }, + { + "bbox": [ + 46, + 387, + 287, + 639 + ], + "type": "inline_equation", + "content": "30 / (128\\times 128)" + }, + { + "bbox": [ + 46, + 387, + 287, + 639 + ], + "type": "text", + "content": ". In contrast, our EditGuard hides both an RGB localization watermark and a 1D copyright watermark, with a capacity far greater than " + }, + { + "bbox": [ + 46, + 387, + 287, + 639 + ], + "type": "inline_equation", + "content": "30 / (128\\times 128)" + }, + { + "bbox": [ + 46, + 387, + 287, + 639 + ], + "type": "text", + "content": ". Here, we do not claim to achieve the best visual quality and bit accuracy, but just to demonstrate that our method is comparable to the current image watermarking methods." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 47, + 647, + 271, + 659 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 647, + 271, + 659 + ], + "spans": [ + { + "bbox": [ + 47, + 647, + 271, + 659 + ], + "type": "text", + "content": "5.4. Extension to AIGC-based Editing Methods" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 46, + 665, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 287, + 713 + ], + "type": "text", + "content": "Dataset Preparation: We constructed a dataset tailored for AIGC Editing methods, dubbed AGE-Set, comprising two sub-datasets. The first AGE-Set-C is a batch-processed coarse tamper dataset. Its original images are sourced from" + } + ] + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 307, + 159, + 370, + 192 + ], + "blocks": [ + { + "bbox": [ + 307, + 159, + 370, + 192 + ], + "lines": [ + { + "bbox": [ + 307, + 159, + 370, + 192 + ], + "spans": [ + { + "bbox": [ + 307, + 159, + 370, + 192 + ], + "type": "image", + "image_path": "a3d698dd3b7233fb982193bec98a5b9b6dc00da38e36f29d78b369f7739469f6.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 372, + 159, + 428, + 192 + ], + "blocks": [ + { + "bbox": [ + 372, + 159, + 428, + 192 + ], + "lines": [ + { + "bbox": [ + 372, + 159, + 428, + 192 + ], + "spans": [ + { + "bbox": [ + 372, + 159, + 428, + 192 + ], + "type": "image", + "image_path": "2aa366912735e6902300754de3efbc2c7f8356075cae02d1740b9f8804c9efc4.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 429, + 159, + 471, + 192 + ], + "blocks": [ + { + "bbox": [ + 429, + 159, + 471, + 192 + ], + "lines": [ + { + "bbox": [ + 429, + 159, + 471, + 192 + ], + "spans": [ + { + "bbox": [ + 429, + 159, + 471, + 192 + ], + "type": "image", + "image_path": "3fdedada16bcdab9164cd71199ca49fac996c152395eccc4c710bbba918fd5a0.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 472, + 159, + 498, + 192 + ], + "blocks": [ + { + "bbox": [ + 472, + 159, + 498, + 192 + ], + "lines": [ + { + "bbox": [ + 472, + 159, + 498, + 192 + ], + "spans": [ + { + "bbox": [ + 472, + 159, + 498, + 192 + ], + "type": "image", + "image_path": "816a0d20ac6825bea0b8fee43275fcdc5903d35dac32e6b5462be513437c2261.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 500, + 159, + 544, + 192 + ], + "blocks": [ + { + "bbox": [ + 500, + 159, + 544, + 192 + ], + "lines": [ + { + "bbox": [ + 500, + 159, + 544, + 192 + ], + "spans": [ + { + "bbox": [ + 500, + 159, + 544, + 192 + ], + "type": "image", + "image_path": "fc4edcaef943973475910003308f05107398737bad26fd2de14f34d2c22e4594.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_body" + } + ], + "index": 27 + }, + { + "type": "image", + "bbox": [ + 308, + 200, + 370, + 221 + ], + "blocks": [ + { + "bbox": [ + 308, + 200, + 370, + 221 + ], + "lines": [ + { + "bbox": [ + 308, + 200, + 370, + 221 + ], + "spans": [ + { + "bbox": [ + 308, + 200, + 370, + 221 + ], + "type": "image", + "image_path": "4738da0e0397918d0663e84d28ad244b867a07e02fa7d91b40baaf698622d151.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + } + ], + "index": 28 + }, + { + "type": "image", + "bbox": [ + 371, + 194, + 428, + 221 + ], + "blocks": [ + { + "bbox": [ + 371, + 194, + 428, + 221 + ], + "lines": [ + { + "bbox": [ + 371, + 194, + 428, + 221 + ], + "spans": [ + { + "bbox": [ + 371, + 194, + 428, + 221 + ], + "type": "image", + "image_path": "d6bc9a0a9ad2fc1b7d3b3810c9af56a31f59f9c54be1dd9981caaefe5315c94e.jpg" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_body" + } + ], + "index": 29 + }, + { + "type": "image", + "bbox": [ + 429, + 194, + 471, + 221 + ], + "blocks": [ + { + "bbox": [ + 429, + 194, + 471, + 221 + ], + "lines": [ + { + "bbox": [ + 429, + 194, + 471, + 221 + ], + "spans": [ + { + "bbox": [ + 429, + 194, + 471, + 221 + ], + "type": "image", + "image_path": "a79116708e6ace07e34d9cd9b405a64ae2cde5d10c4220e84c866048561cda8a.jpg" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_body" + } + ], + "index": 30 + }, + { + "type": "image", + "bbox": [ + 472, + 194, + 498, + 221 + ], + "blocks": [ + { + "bbox": [ + 472, + 194, + 498, + 221 + ], + "lines": [ + { + "bbox": [ + 472, + 194, + 498, + 221 + ], + "spans": [ + { + "bbox": [ + 472, + 194, + 498, + 221 + ], + "type": "image", + "image_path": "97136b7504c476d3f733252490dc8f28d6d237f3d713b80182826d44d2576899.jpg" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_body" + } + ], + "index": 31 + }, + { + "type": "image", + "bbox": [ + 500, + 194, + 544, + 221 + ], + "blocks": [ + { + "bbox": [ + 500, + 194, + 544, + 221 + ], + "lines": [ + { + "bbox": [ + 500, + 194, + 544, + 221 + ], + "spans": [ + { + "bbox": [ + 500, + 194, + 544, + 221 + ], + "type": "image", + "image_path": "61921a5bab7ef5eaabb3e6a77529d830d0a262dd2ff06037c2b592da94a9a45d.jpg" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_body" + } + ], + "index": 32 + }, + { + "type": "image", + "bbox": [ + 309, + 222, + 370, + 250 + ], + "blocks": [ + { + "bbox": [ + 309, + 222, + 370, + 250 + ], + "lines": [ + { + "bbox": [ + 309, + 222, + 370, + 250 + ], + "spans": [ + { + "bbox": [ + 309, + 222, + 370, + 250 + ], + "type": "image", + "image_path": "3de38109b2947ebb0cfec34afd0a8030d377a8e777ee835ed30dd6d6153abb94.jpg" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_body" + } + ], + "index": 33 + }, + { + "type": "image", + "bbox": [ + 371, + 222, + 428, + 250 + ], + "blocks": [ + { + "bbox": [ + 371, + 222, + 428, + 250 + ], + "lines": [ + { + "bbox": [ + 371, + 222, + 428, + 250 + ], + "spans": [ + { + "bbox": [ + 371, + 222, + 428, + 250 + ], + "type": "image", + "image_path": "978c7704ed9dbc9ad97cef580bcb0f5ba4e75af3c2acc01a247fc8f5a54156fe.jpg" + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_body" + } + ], + "index": 34 + }, + { + "type": "image", + "bbox": [ + 429, + 222, + 471, + 250 + ], + "blocks": [ + { + "bbox": [ + 429, + 222, + 471, + 250 + ], + "lines": [ + { + "bbox": [ + 429, + 222, + 471, + 250 + ], + "spans": [ + { + "bbox": [ + 429, + 222, + 471, + 250 + ], + "type": "image", + "image_path": "c1b99e11191fbba0ca9ba1cc538d606061a288535ce85654b7b5054d7db1d2e4.jpg" + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "image_body" + } + ], + "index": 35 + }, + { + "type": "image", + "bbox": [ + 472, + 222, + 498, + 250 + ], + "blocks": [ + { + "bbox": [ + 472, + 222, + 498, + 250 + ], + "lines": [ + { + "bbox": [ + 472, + 222, + 498, + 250 + ], + "spans": [ + { + "bbox": [ + 472, + 222, + 498, + 250 + ], + "type": "image", + "image_path": "d0ab8577cabc01bcd57112c4e17f47b8c1868b5c4403a245f1ce02858322b3d4.jpg" + } + ] + } + ], + "index": 36, + "angle": 0, + "type": "image_body" + } + ], + "index": 36 + }, + { + "type": "image", + "bbox": [ + 500, + 222, + 544, + 250 + ], + "blocks": [ + { + "bbox": [ + 500, + 222, + 544, + 250 + ], + "lines": [ + { + "bbox": [ + 500, + 222, + 544, + 250 + ], + "spans": [ + { + "bbox": [ + 500, + 222, + 544, + 250 + ], + "type": "image", + "image_path": "fb07c2ec7d2fad52ba53093c3e6d0e7fc668b4719e15a082e3ace33a0ffe1804.jpg" + } + ] + } + ], + "index": 37, + "angle": 0, + "type": "image_body" + } + ], + "index": 37 + }, + { + "type": "image", + "bbox": [ + 308, + 251, + 370, + 279 + ], + "blocks": [ + { + "bbox": [ + 308, + 251, + 370, + 279 + ], + "lines": [ + { + "bbox": [ + 308, + 251, + 370, + 279 + ], + "spans": [ + { + "bbox": [ + 308, + 251, + 370, + 279 + ], + "type": "image", + "image_path": "43aebf93db511ac8ad02cd3c7a18779ac604d36214cb16e54407fd3242cf9acf.jpg" + } + ] + } + ], + "index": 38, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 306, + 281, + 545, + 303 + ], + "lines": [ + { + "bbox": [ + 306, + 281, + 545, + 303 + ], + "spans": [ + { + "bbox": [ + 306, + 281, + 545, + 303 + ], + "type": "text", + "content": "Figure 8. Localization performance comparisons of our EditGuard and other methods [8, 18, 39, 61] on our constructed AGE-Set-C." + } + ] + } + ], + "index": 43, + "angle": 0, + "type": "image_caption" + } + ], + "index": 38 + }, + { + "type": "image", + "bbox": [ + 371, + 251, + 428, + 279 + ], + "blocks": [ + { + "bbox": [ + 371, + 251, + 428, + 279 + ], + "lines": [ + { + "bbox": [ + 371, + 251, + 428, + 279 + ], + "spans": [ + { + "bbox": [ + 371, + 251, + 428, + 279 + ], + "type": "image", + "image_path": "64aebf924695532abcc878f1d7ff14b5c84a341fdd4e1781705592337fc5bb07.jpg" + } + ] + } + ], + "index": 39, + "angle": 0, + "type": "image_body" + } + ], + "index": 39 + }, + { + "type": "image", + "bbox": [ + 429, + 251, + 471, + 279 + ], + "blocks": [ + { + "bbox": [ + 429, + 251, + 471, + 279 + ], + "lines": [ + { + "bbox": [ + 429, + 251, + 471, + 279 + ], + "spans": [ + { + "bbox": [ + 429, + 251, + 471, + 279 + ], + "type": "image", + "image_path": "510f586af59a8dd1331cfadacf30af9acb24d8412991393df83d34fa7cbb50bd.jpg" + } + ] + } + ], + "index": 40, + "angle": 0, + "type": "image_body" + } + ], + "index": 40 + }, + { + "type": "image", + "bbox": [ + 472, + 251, + 498, + 279 + ], + "blocks": [ + { + "bbox": [ + 472, + 251, + 498, + 279 + ], + "lines": [ + { + "bbox": [ + 472, + 251, + 498, + 279 + ], + "spans": [ + { + "bbox": [ + 472, + 251, + 498, + 279 + ], + "type": "image", + "image_path": "1961807dc7f2143728f3e9bd78759d2a60e337895b39be63ca524b3e73376d08.jpg" + } + ] + } + ], + "index": 41, + "angle": 0, + "type": "image_body" + } + ], + "index": 41 + }, + { + "type": "image", + "bbox": [ + 500, + 251, + 544, + 279 + ], + "blocks": [ + { + "bbox": [ + 500, + 251, + 544, + 279 + ], + "lines": [ + { + "bbox": [ + 500, + 251, + 544, + 279 + ], + "spans": [ + { + "bbox": [ + 500, + 251, + 544, + 279 + ], + "type": "image", + "image_path": "45c47d0401b8e85bf542c59842412290bef15467881628beb41afebe3bf1039d.jpg" + } + ] + } + ], + "index": 42, + "angle": 0, + "type": "image_body" + } + ], + "index": 42 + }, + { + "bbox": [ + 304, + 306, + 545, + 474 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 306, + 545, + 474 + ], + "spans": [ + { + "bbox": [ + 304, + 306, + 545, + 474 + ], + "type": "text", + "content": "COCO 2017 [38] and CelebA [41], containing " + }, + { + "bbox": [ + 304, + 306, + 545, + 474 + ], + "type": "inline_equation", + "content": "30K" + }, + { + "bbox": [ + 304, + 306, + 545, + 474 + ], + "type": "text", + "content": " training images and " + }, + { + "bbox": [ + 304, + 306, + 545, + 474 + ], + "type": "inline_equation", + "content": "1.2K" + }, + { + "bbox": [ + 304, + 306, + 545, + 474 + ], + "type": "text", + "content": " testing images. We used some SOTA editing methods such as Stable Diffusion Inpaint [51], ControlNet [72], SDXL [49] to manipulate images with the prompt to be \"None\", and employed some unconditional methods like Repaint [43], Lama [57], and Faceswap [60]. Note that we only use the tampered data to train other methods, not our EditGuard. The second sub-dataset AGE-Set-F includes 100 finely edited images. It is edited manually via some sophisticated software such as SD-Web-UI, Photoshop, and Adobe Firefly. These AIGC-based editing methods can achieve a good fusion of the tampered and unchanged areas, making it hard for the naked eye to catch artifacts. More details are presented in S.M.." + } + ] + } + ], + "index": 44 + }, + { + "bbox": [ + 304, + 474, + 546, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 474, + 546, + 677 + ], + "spans": [ + { + "bbox": [ + 304, + 474, + 546, + 677 + ], + "type": "text", + "content": "AGE-Set-C: Tab. 3 presents the comparison of our EditGuard and some SOTA tamper localization methods [8, 18, 39, 45, 61]. We observe that the F1-scores of other passive forensic methods are generally lower than 0.7 when applied to AGE-Set-C. Meanwhile, even when we try our best to finetune MVSS-Net using AGE-Set-C, the accuracy of MVSS-Net remains unsatisfactory, and they exhibit catastrophic forgetting across various tamper methods. In contrast, our method can guarantee an F1-score and AUC of over " + }, + { + "bbox": [ + 304, + 474, + 546, + 677 + ], + "type": "inline_equation", + "content": "95\\%" + }, + { + "bbox": [ + 304, + 474, + 546, + 677 + ], + "type": "text", + "content": ", maintaining around " + }, + { + "bbox": [ + 304, + 474, + 546, + 677 + ], + "type": "inline_equation", + "content": "90\\%" + }, + { + "bbox": [ + 304, + 474, + 546, + 677 + ], + "type": "text", + "content": " IOU, regardless of tampering types. As shown in Fig. 8, our EditGuard can accurately capture these imperceptible tampering traces produced by AIGC-based editing methods, but other methods are almost ineffective. Moreover, our EditGuard can effectively recover copyright information with a bit accuracy exceeding " + }, + { + "bbox": [ + 304, + 474, + 546, + 677 + ], + "type": "inline_equation", + "content": "99.8\\%" + }, + { + "bbox": [ + 304, + 474, + 546, + 677 + ], + "type": "text", + "content": ". Noting that none of the comparison methods offer copyright protection capabilities." + } + ] + } + ], + "index": 45 + }, + { + "bbox": [ + 304, + 677, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 546, + 713 + ], + "type": "text", + "content": "AGE-Set-F: To further highlight the practicality of our EditGuard, we conducted subjective comparisons with other methods on the meticulously tampered AGE-Set-F." + } + ] + } + ], + "index": 46 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "11970" + } + ] + } + ], + "index": 47 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 49, + 71, + 544, + 270 + ], + "blocks": [ + { + "bbox": [ + 49, + 71, + 544, + 270 + ], + "lines": [ + { + "bbox": [ + 49, + 71, + 544, + 270 + ], + "spans": [ + { + "bbox": [ + 49, + 71, + 544, + 270 + ], + "type": "image", + "image_path": "bce070c0cef5d5ad87cc6a09e540d2e125882f389da237eb7590e82583cf7c0a.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 48, + 271, + 544, + 281 + ], + "lines": [ + { + "bbox": [ + 48, + 271, + 544, + 281 + ], + "spans": [ + { + "bbox": [ + 48, + 271, + 544, + 281 + ], + "type": "text", + "content": "Figure 9. Localization precision comparisons of our EditGuard and other competitive methods on the meticulously tampered AGE-Set-F." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 49, + 309, + 286, + 358 + ], + "blocks": [ + { + "bbox": [ + 47, + 285, + 286, + 308 + ], + "lines": [ + { + "bbox": [ + 47, + 285, + 286, + 308 + ], + "spans": [ + { + "bbox": [ + 47, + 285, + 286, + 308 + ], + "type": "text", + "content": "Table 4. Localization and bit recovery performance of our Edit-Guard and MVSS-Net† [8] under different levels of degradations." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 49, + 309, + 286, + 358 + ], + "lines": [ + { + "bbox": [ + 49, + 309, + 286, + 358 + ], + "spans": [ + { + "bbox": [ + 49, + 309, + 286, + 358 + ], + "type": "table", + "html": "
MethodsMetricsCleanGaussian NoiseJPEGPoisson
σ=1σ=5Q=70Q=80Q=90
MVSS-Net†[8]F10.6940.6440.6190.4580.5070.5580.652
BA(%)-------
EditGuard (Ours)F10.9660.9370.9320.9200.9200.9250.943
BA(%)99.9599.9499.3797.6998.1698.2399.91
", + "image_path": "ab61ec84023430adfc76ec15cc0af35310b4847355245841e38a3b2b129c7708.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 362, + 286, + 458 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 362, + 286, + 458 + ], + "spans": [ + { + "bbox": [ + 46, + 362, + 286, + 458 + ], + "type": "text", + "content": "The tamper types in this subset did not appear in the training set. As shown in Fig. 9, when faced with real-world tampering, even the most powerful tamper localization methods almost entirely fail. This is due to their mechanisms to look for image artifacts and explore instance-wise semantic information. However, our EditGuard, which locates tampered masks via the natural fragility and locality of I2I steganography, can still clearly annotate the tampered area." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 465, + 165, + 478 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 465, + 165, + 478 + ], + "spans": [ + { + "bbox": [ + 47, + 465, + 165, + 478 + ], + "type": "text", + "content": "5.5. Robustness Analysis" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 484, + 286, + 615 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 484, + 286, + 615 + ], + "spans": [ + { + "bbox": [ + 46, + 484, + 286, + 615 + ], + "type": "text", + "content": "As shown in Tab. 4, we conducted robustness analysis on the tampering of \"Stable Diffusion Inpaint\" [51] under Gaussian noise with " + }, + { + "bbox": [ + 46, + 484, + 286, + 615 + ], + "type": "inline_equation", + "content": "\\sigma = 1" + }, + { + "bbox": [ + 46, + 484, + 286, + 615 + ], + "type": "text", + "content": " and 5, JPEG compression with " + }, + { + "bbox": [ + 46, + 484, + 286, + 615 + ], + "type": "inline_equation", + "content": "Q = 70" + }, + { + "bbox": [ + 46, + 484, + 286, + 615 + ], + "type": "text", + "content": ", 80 and 90, and Poisson noise with " + }, + { + "bbox": [ + 46, + 484, + 286, + 615 + ], + "type": "inline_equation", + "content": "\\alpha = 4" + }, + { + "bbox": [ + 46, + 484, + 286, + 615 + ], + "type": "text", + "content": " [71]. We observed that our method still maintains a high localization accuracy (F1-score " + }, + { + "bbox": [ + 46, + 484, + 286, + 615 + ], + "type": "inline_equation", + "content": ">0.9" + }, + { + "bbox": [ + 46, + 484, + 286, + 615 + ], + "type": "text", + "content": ") and bit accuracy with a very slight performance decrease under various levels of degradations, while MVSS-Net† [8] exhibits a noticeable performance degradation compared to its results in clean conditions. It is attributed to our prompt-based estimation that can effectively learn the degradation representation." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 623, + 140, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 623, + 140, + 635 + ], + "spans": [ + { + "bbox": [ + 47, + 623, + 140, + 635 + ], + "type": "text", + "content": "5.6. Ablation Study" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 641, + 286, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 641, + 286, + 712 + ], + "spans": [ + { + "bbox": [ + 46, + 641, + 286, + 712 + ], + "type": "text", + "content": "To verify the effectiveness of each component of the Edit-Guard, we conducted ablation studies on bi-level optimization (BO), lightweight feature interaction module (LFIM), transformer block (TB), and prompt-based fusion (PF) under the tampering of \"Stable Diffusion Inpaint\". As listed in Tab. 5, without BO, the joint training of all components" + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 308, + 297, + 544, + 354 + ], + "blocks": [ + { + "bbox": [ + 307, + 285, + 544, + 296 + ], + "lines": [ + { + "bbox": [ + 307, + 285, + 544, + 296 + ], + "spans": [ + { + "bbox": [ + 307, + 285, + 544, + 296 + ], + "type": "text", + "content": "Table 5. Abalation studies on the core components of EditGuard." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 308, + 297, + 544, + 354 + ], + "lines": [ + { + "bbox": [ + 308, + 297, + 544, + 354 + ], + "spans": [ + { + "bbox": [ + 308, + 297, + 544, + 354 + ], + "type": "table", + "html": "
CaseDegradation Type D(·)PFTBLFIMBOF1AUCIoUBA(%)
(a)Clean✓✓✓---49.17
(b)Clean✓✓✓0.9500.9600.90499.73
(c)Clean✓✓✓0.9570.9660.92799.51
(d)Random Degradations✓✓✓0.9030.9330.84199.12
OursClean✓✓✓✓0.9660.9710.93699.95
Random Degradations✓✓✓✓0.9380.9640.88799.36
", + "image_path": "cabce7a34e141a4bb1209a7fc25977c64e8bfb656c98cc12391f57bfa9a40835.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 359, + 545, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 359, + 545, + 502 + ], + "spans": [ + { + "bbox": [ + 304, + 359, + 545, + 502 + ], + "type": "text", + "content": "cannot converge effectively, resulting in bit accuracy that is close to random guessing. Without LFIM and TB, the IoU of EditGuard will suffer 0.032 and 0.009 declines since these two modules can better perform feature fusion. Without PF, the robustness of the EditGuard will significantly decline. We observed that the F1/AUC/IoU of our method far surpasses that of case (d) by 0.035/0.031/0.046 under \"Random Degradations\", which indicates that the PF effectively enables a single network to support watermark recovery under various degradations. \"Random Degradations\" denotes that we randomly set the " + }, + { + "bbox": [ + 304, + 359, + 545, + 502 + ], + "type": "inline_equation", + "content": "\\mathcal{D}(\\cdot)" + }, + { + "bbox": [ + 304, + 359, + 545, + 502 + ], + "type": "text", + "content": " to various levels of Gaussian noise, Poisson noise, and JPEG compression." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 514, + 378, + 525 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 514, + 378, + 525 + ], + "spans": [ + { + "bbox": [ + 306, + 514, + 378, + 525 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 533, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 533, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 533, + 545, + 712 + ], + "type": "text", + "content": "We present the first attempt to design a deep versatile watermarking mechanism EditGuard. It enhances the credibility of images by embedding imperceptible localization and copyright watermarks, and decoding accurate copyright information and tampered areas, making it a reliable tool for artistic creation and legal forensic analysis. In the future, we will focus on improving the robustness of EditGuard and strive not only to offer pixel-wise localization results but also to provide semantic-wise outcomes. Additionally, we plan to further expand EditGuard to a broader range of modalities and applications, including video, audio, and 3D scenes. Our efforts at information authenticity serve not only the AIGC industry, but the trust in our digital world, ensuring that every pixel tells the truth and the rights of each individual are safeguarded." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "11971" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 714 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 135 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 135 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 135 + ], + "type": "text", + "content": "[1] Mahdi Ahmadi, Alireza Norouzi, Nader Karimi, Shadrokh Samavi, and Ali Emami. Redmark: Framework for residual diffusion watermarking based on deep networks. Expert Systems with Applications, 146:113157, 2020. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 137, + 288, + 180 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 137, + 288, + 180 + ], + "spans": [ + { + "bbox": [ + 53, + 137, + 288, + 180 + ], + "type": "text", + "content": "[2] Vishal Asnani, Xi Yin, Tal Hassner, and Xiaoming Liu. Malp: Manipulation localization using a proactive scheme. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 182, + 287, + 225 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 182, + 287, + 225 + ], + "spans": [ + { + "bbox": [ + 53, + 182, + 287, + 225 + ], + "type": "text", + "content": "[3] Shumeet Baluja. Hiding images in plain sight: Deep steganography. In Proceedings of the International Conference on Neural Information Processing Systems (NeurIPS), 2017. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 227, + 287, + 260 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 227, + 287, + 260 + ], + "spans": [ + { + "bbox": [ + 53, + 227, + 287, + 260 + ], + "type": "text", + "content": "[4] Liangyu Chen, Xiaojie Chu, Xiangyu Zhang, and Jian Sun. Simple baselines for image restoration. In Proceedings of the European Conference on Computer Vision (ECCV), 2022. 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 262, + 287, + 306 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 262, + 287, + 306 + ], + "spans": [ + { + "bbox": [ + 53, + 262, + 287, + 306 + ], + "type": "text", + "content": "[5] Xinru Chen, Chengbo Dong, Jiaqi Ji, Juan Cao, and Xirong Li. Image manipulation detection by multi-view multi-scale supervision. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 307, + 287, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 307, + 287, + 352 + ], + "spans": [ + { + "bbox": [ + 53, + 307, + 287, + 352 + ], + "type": "text", + "content": "[6] Baotian Cheng, Rongrong Ni, and Yao Zhao. A refining localization watermarking for image tamper detection and recovery. In 2012 IEEE 11th International Conference on Signal Processing, pages 984-988, 2012. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 353, + 287, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 353, + 287, + 397 + ], + "spans": [ + { + "bbox": [ + 53, + 353, + 287, + 397 + ], + "type": "text", + "content": "[7] Yingqian Cui, Jie Ren, Han Xu, Pengfei He, Hui Liu, Lichao Sun, and Jiliang Tang. Diffusionshield: A watermark for copyright protection against generative diffusion models. arXiv preprint arXiv:2306.04642, 2023. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 398, + 287, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 398, + 287, + 453 + ], + "spans": [ + { + "bbox": [ + 53, + 398, + 287, + 453 + ], + "type": "text", + "content": "[8] Chengbo Dong, Xinru Chen, Ruohan Hu, Juan Cao, and Xirong Li. Mvss-net: Multi-view multi-scale supervised networks for image manipulation detection. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(3):3539-3553, 2022. 2, 6, 7, 8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 454, + 287, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 454, + 287, + 499 + ], + "spans": [ + { + "bbox": [ + 53, + 454, + 287, + 499 + ], + "type": "text", + "content": "[9] Jing Dong, Wei Wang, and Tieniu Tan. Casia image tampering detection evaluation database. In Proceedings of the IEEE China Summit and International Conference on Signal and Information Processing (ChinaSIP), 2013. 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 500, + 287, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 500, + 287, + 555 + ], + "spans": [ + { + "bbox": [ + 48, + 500, + 287, + 555 + ], + "type": "text", + "content": "[10] Han Fang, Zhaoyang Jia, Zehua Ma, Ee-Chien Chang, and Weiming Zhang. Pimog: An effective screen-shooting noise-layer simulation for deep-learning-based watermarking network. In Proceedings of the 30th ACM International Conference on Multimedia (MM), 2022. 3, 6, 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 556, + 287, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 556, + 287, + 612 + ], + "spans": [ + { + "bbox": [ + 48, + 556, + 287, + 612 + ], + "type": "text", + "content": "[11] Han Fang, Yupeng Qiu, Kejiang Chen, Jiyi Zhang, Weiming Zhang, and Ee-Chien Chang. Flow-based robust watermarking with invertible noise layer for black-box distortions. In Proceedings of the AAAI Conference on Artificial Intelligence (AAAI), 2023. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 613, + 287, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 613, + 287, + 668 + ], + "spans": [ + { + "bbox": [ + 48, + 613, + 287, + 668 + ], + "type": "text", + "content": "[12] Pierre Fernandez, Guillaume Couairon, Hervé Jégou, Matthijs Douze, and Teddy Furon. The stable signature: Rooting watermarks in latent diffusion models. Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2023. 2, 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 669, + 287, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 669, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 48, + 669, + 287, + 714 + ], + "type": "text", + "content": "[13] Anirban Ghoshal. Artists lose first copyright battle in the fight against ai-generated images. https://www.computerworld.com/article/3709691/artists-lose-first-copyright-battle-in" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 712 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 327, + 73, + 544, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 544, + 95 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 544, + 95 + ], + "type": "text", + "content": "the-fight-against-ai-generated-images. \nhtml,2023.2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 97, + 546, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 97, + 546, + 140 + ], + "spans": [ + { + "bbox": [ + 307, + 97, + 546, + 140 + ], + "type": "text", + "content": "[14] Gabriel Goh, James Betker, Li Jing, Aditya Ramesh, et al. Improving image generation with better captions. https://cdn.openai.com/papers/dall-e-3.pdf, 2023.1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 143, + 544, + 166 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 143, + 544, + 166 + ], + "spans": [ + { + "bbox": [ + 307, + 143, + 544, + 166 + ], + "type": "text", + "content": "[15] Google DeepMind. Synthid. https://deepmind.google/technologies/synthid/, 2023.2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 167, + 546, + 233 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 167, + 546, + 233 + ], + "spans": [ + { + "bbox": [ + 308, + 167, + 546, + 233 + ], + "type": "text", + "content": "[16] Haiying Guan, Mark Kozak, Eric Robertson, Yooyoung Lee, Amy N Yates, Andrew Delgado, Daniel Zhou, Timothee Kheyrkhah, Jeff Smith, and Jonathan Fiscus. Mfc datasets: Large-scale benchmark datasets for media forensic challenge evaluation. In Proceedings of the IEEE Winter Applications of Computer Vision Workshops (WACVW), 2019. 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 235, + 546, + 291 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 235, + 546, + 291 + ], + "spans": [ + { + "bbox": [ + 308, + 235, + 546, + 291 + ], + "type": "text", + "content": "[17] Fabrizio Guillaro, Davide Cozzolino, Avneesh Sud, Nicholas Dufour, and Luisa Verdoliva. Trufor: Leveraging all-round clues for trustworthy image forgery detection and localization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 2, 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 292, + 547, + 347 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 292, + 547, + 347 + ], + "spans": [ + { + "bbox": [ + 308, + 292, + 547, + 347 + ], + "type": "text", + "content": "[18] Xiao Guo, Xiaohong Liu, Zhiyuan Ren, Steven Grosz, Iacopo Masi, and Xiaoming Liu. Hierarchical fine-grained image forgery detection and localization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 2, 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 349, + 546, + 393 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 349, + 546, + 393 + ], + "spans": [ + { + "bbox": [ + 308, + 349, + 546, + 393 + ], + "type": "text", + "content": "[19] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 5" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 395, + 546, + 440 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 395, + 546, + 440 + ], + "spans": [ + { + "bbox": [ + 308, + 395, + 546, + 440 + ], + "type": "text", + "content": "[20] Yu-Feng Hsu and Shih-Fu Chang. Detecting image splicing using geometry invariants and camera characteristics consistency. In Proceedings of the IEEE International Conference on Multimedia and Expo (ICME), 2006. 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 441, + 546, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 441, + 546, + 475 + ], + "spans": [ + { + "bbox": [ + 308, + 441, + 546, + 475 + ], + "type": "text", + "content": "[21] Jie Hu, Li Shen, and Gang Sun. Squeeze-and-excitation networks. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 475, + 546, + 532 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 475, + 546, + 532 + ], + "spans": [ + { + "bbox": [ + 308, + 475, + 546, + 532 + ], + "type": "text", + "content": "[22] Xuefeng Hu, Zhihan Zhang, Zhenye Jiang, Syomantak Chaudhuri, Zhenheng Yang, and Ram Nevatia. Span: Spatial pyramid attention network for image manipulation localization. In Proceedings of the European Conference on Computer Vision (ECCV), 2020. 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 533, + 546, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 533, + 546, + 588 + ], + "spans": [ + { + "bbox": [ + 308, + 533, + 546, + 588 + ], + "type": "text", + "content": "[23] Xiaoxiao Hu, Qichao Ying, Zhenxing Qian, Sheng Li, and Xinpeng Zhang. Draw: Defending camera-shooted raw against image manipulation. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2023. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 590, + 547, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 590, + 547, + 645 + ], + "spans": [ + { + "bbox": [ + 308, + 590, + 547, + 645 + ], + "type": "text", + "content": "[24] Nasir N Hurrah, Shabir A Parah, Nazir A Loan, Javaid A Sheikh, Mohammad Elhoseny, and Khan Muhammad. Dual watermarking framework for privacy protection and content authentication of multimedia. Future generation computer Systems, 94:654-673, 2019. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 647, + 547, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 647, + 547, + 712 + ], + "spans": [ + { + "bbox": [ + 308, + 647, + 547, + 712 + ], + "type": "text", + "content": "[25] Ashraful Islam, Chengjiang Long, Arslan Basharat, and Anthony Hoogs. Doa-gan: Dual-order attentive generative adversarial network for image copy-move forgery detection and localization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 2" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "11972" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "text", + "content": "[26] Zhaoyang Jia, Han Fang, and Weiming Zhang. Mbrs: Enhancing robustness of dnn-based watermarking by minibatch of real and simulatedJPEG compression. In Proceedings of the 29th ACM International Conference on Multimedia (MM), 2021. 6, 7" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 129, + 287, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 129, + 287, + 162 + ], + "spans": [ + { + "bbox": [ + 49, + 129, + 287, + 162 + ], + "type": "text", + "content": "[27] Zhengyuan Jiang, Jinghuai Zhang, and Neil Zhenqiang Gong. Evading watermark based detection of ai-generated content. arXiv preprint arXiv:2305.03807, 2023. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 163, + 287, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 163, + 287, + 207 + ], + "spans": [ + { + "bbox": [ + 49, + 163, + 287, + 207 + ], + "type": "text", + "content": "[28] Junpeng Jing, Xin Deng, Mai Xu, Jianyi Wang, and Zhenyu Guan. Hinet: Deep image hiding by invertible network. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 209, + 287, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 209, + 287, + 262 + ], + "spans": [ + { + "bbox": [ + 49, + 209, + 287, + 262 + ], + "type": "text", + "content": "[29] Asra Kamili, Nasir N Hurrah, Shabir A Parah, Ghulam Mohiuddin Bhat, and Khan Muhammad. Dwfcat: Dual watermarking framework for industrial image authentication and tamper localization. IEEE Transactions on Industrial Informatics, 17(7):5108-5117, 2020. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 49, + 264, + 287, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 264, + 287, + 297 + ], + "spans": [ + { + "bbox": [ + 49, + 264, + 287, + 297 + ], + "type": "text", + "content": "[30] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 49, + 298, + 287, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 298, + 287, + 353 + ], + "spans": [ + { + "bbox": [ + 49, + 298, + 287, + 353 + ], + "type": "text", + "content": "[31] Myung-Joon Kwon, In-Jae Yu, Seung-Hun Nam, and Heung-Kyu Lee. Cat-net: Compression artifact tracing network for detection and localization of image splicing. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), 2021. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 354, + 287, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 354, + 287, + 408 + ], + "spans": [ + { + "bbox": [ + 48, + 354, + 287, + 408 + ], + "type": "text", + "content": "[32] Myung-Joon Kwon, Seung-Hun Nam, In-Jae Yu, Heung-Kyu Lee, and Changick Kim. LearningJPEG compression artifacts for image manipulation detection and localization. International Journal of Computer Vision, 130(8):1875-1895, 2022.6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 410, + 287, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 410, + 287, + 453 + ], + "spans": [ + { + "bbox": [ + 48, + 410, + 287, + 453 + ], + "type": "text", + "content": "[33] Haodong Li and Jiwu Huang. Localization of deep inpainting using high-pass fully convolutional network. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2019. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 455, + 287, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 455, + 287, + 498 + ], + "spans": [ + { + "bbox": [ + 48, + 455, + 287, + 498 + ], + "type": "text", + "content": "[34] Yuanman Li and Jiantao Zhou. Fast and effective image copy-move forgery detection via hierarchical feature point matching. IEEE Transactions on Information Forensics and Security, 14(5):1307-1322, 2018. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 500, + 287, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 500, + 287, + 544 + ], + "spans": [ + { + "bbox": [ + 48, + 500, + 287, + 544 + ], + "type": "text", + "content": "[35] Yue Li, Dong Liu, Houqiang Li, Li Li, Zhu Li, and Feng Wu. Learning a convolutional neural network for image compact-resolution. IEEE Transactions on Image Processing, 28(3): 1092-1107, 2018. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 545, + 287, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 545, + 287, + 589 + ], + "spans": [ + { + "bbox": [ + 48, + 545, + 287, + 589 + ], + "type": "text", + "content": "[36] Siau-Chuin Liew, Siau-Way Liew, and Jasni Mohd Zain. Tamper localization and lossless recovery watermarking scheme with roi segmentation and multilevel authentication. Journal of digital imaging, 26:316-325, 2013. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 590, + 287, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 590, + 287, + 633 + ], + "spans": [ + { + "bbox": [ + 48, + 590, + 287, + 633 + ], + "type": "text", + "content": "[37] Chia-Chen Lin, Ting-Lin Lee, Ya-Fen Chang, Pei-Feng Shiu, and Bohan Zhang. Fragile watermarking for tamper localization and self-recovery based on ambtc and vq. Electronics, 12(2):415, 2023. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 635, + 287, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 635, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 635, + 287, + 689 + ], + "type": "text", + "content": "[38] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Proceedings of the European Conference on Computer Vision (ECCV), 2014. 6, 7" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "type": "text", + "content": "[39] Xiaohong Liu, Yaojie Liu, Jun Chen, and Xiaoming Liu. Pssc-net: Progressive spatio-channel correlation network for" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 327, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 545, + 106 + ], + "type": "text", + "content": "image manipulation detection and localization. IEEE Transactions on Circuits and Systems for Video Technology, 32 (11):7505-7517, 2022. 6, 7" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 107, + 545, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 107, + 545, + 161 + ], + "spans": [ + { + "bbox": [ + 308, + 107, + 545, + 161 + ], + "type": "text", + "content": "[40] Yang Liu, Mengxi Guo, Jian Zhang, Yuesheng Zhu, and Xiaodong Xie. A novel two-stage separable deep learning framework for practical blind watermarking. In Proceedings of the ACM International Conference on Multimedia (MM), 2019. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 163, + 545, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 163, + 545, + 207 + ], + "spans": [ + { + "bbox": [ + 308, + 163, + 545, + 207 + ], + "type": "text", + "content": "[41] Ziwei Liu, Ping Luo, Xiaogang Wang, and Xiaou Tang. Deep learning face attributes in the wild. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), 2015. 7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 209, + 545, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 209, + 545, + 262 + ], + "spans": [ + { + "bbox": [ + 308, + 209, + 545, + 262 + ], + "type": "text", + "content": "[42] Shao-Ping Lu, Rong Wang, Tao Zhong, and Paul L Rosin. Large-capacity image steganography based on invertible neural networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 3, 4" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 264, + 545, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 264, + 545, + 319 + ], + "spans": [ + { + "bbox": [ + 308, + 264, + 545, + 319 + ], + "type": "text", + "content": "[43] Andreas Lugmayr, Martin Danelljan, Andres Romero, Fisher Yu, Radu Timofte, and Luc Van Gool. Repaint: Inpainting using denoising diffusion probabilistic models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 1, 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 320, + 545, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 320, + 545, + 374 + ], + "spans": [ + { + "bbox": [ + 308, + 320, + 545, + 374 + ], + "type": "text", + "content": "[44] Rui Ma, Mengxi Guo, Yi Hou, Fan Yang, Yuan Li, Huizhu Jia, and Xiaodong Xie. Towards blind watermarking: Combining invertible and non-invertible mechanisms. In Proceedings of the ACM International Conference on Multimedia (MM), 2022. 2, 3, 6, 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 376, + 545, + 419 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 376, + 545, + 419 + ], + "spans": [ + { + "bbox": [ + 308, + 376, + 545, + 419 + ], + "type": "text", + "content": "[45] Xiaochen Ma, Bo Du, Xianggen Liu, Ahmed Y Al Hammadi, and Jizhe Zhou. Iml-vit: Image manipulation localization by vision transformer. arXiv preprint arXiv:2307.14863, 2023. 2, 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 422, + 545, + 476 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 422, + 545, + 476 + ], + "spans": [ + { + "bbox": [ + 308, + 422, + 545, + 476 + ], + "type": "text", + "content": "[46] Chong Mou, Youmin Xu, Jiechong Song, Chen Zhao, Bernard Ghanem, and Jian Zhang. Large-capacity and flexible video steganography via invertible neural network. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 3, 4" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 478, + 545, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 478, + 545, + 533 + ], + "spans": [ + { + "bbox": [ + 308, + 478, + 545, + 533 + ], + "type": "text", + "content": "[47] Chong Mou, Xintao Wang, Liangbin Xie, Yanze Wu, Jian Zhang, Zhongang Qi, Ying Shan, and Xiaohu Qie. T2i-adapter: Learning adapters to dig out more controllable ability for text-to-image diffusion models. In AAAI Conference on Artificial Intelligence (AAAI), 2024. 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 534, + 545, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 534, + 545, + 578 + ], + "spans": [ + { + "bbox": [ + 308, + 534, + 545, + 578 + ], + "type": "text", + "content": "[48] Neena Raj NR and R Shreelekshmi. Fragile watermarking scheme for tamper localization in images using logistic map and singular value decomposition. Journal of Visual Communication and Image Representation, 85:103500, 2022. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 579, + 545, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 579, + 545, + 633 + ], + "spans": [ + { + "bbox": [ + 308, + 579, + 545, + 633 + ], + "type": "text", + "content": "[49] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952, 2023. 1, 7" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 635, + 545, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 635, + 545, + 678 + ], + "spans": [ + { + "bbox": [ + 308, + 635, + 545, + 678 + ], + "type": "text", + "content": "[50] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 680, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 680, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 680, + 545, + 713 + ], + "type": "text", + "content": "[51] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "11973" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 73, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 67, + 73, + 287, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 73, + 287, + 95 + ], + "spans": [ + { + "bbox": [ + 67, + 73, + 287, + 95 + ], + "type": "text", + "content": "the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 1, 3, 7, 8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 96, + 287, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 96, + 287, + 184 + ], + "spans": [ + { + "bbox": [ + 48, + 96, + 287, + 184 + ], + "type": "text", + "content": "[52] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, Seyed Kamyar Seyed Ghasemipour, Raphael Gontijo-Lopes, Burcu Karagol Ayan, Tim Salimans, Jonathan Ho, David J. Fleet, and Mohammad Norouzi. Photorealistic text-to-image diffusion models with deep language understanding. In Proceedings of the International conference on Neural Information Processing Systems (NeurIPS), 2022. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 186, + 287, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 186, + 287, + 229 + ], + "spans": [ + { + "bbox": [ + 48, + 186, + 287, + 229 + ], + "type": "text", + "content": "[53] Ronald Salloum, Yuzhuo Ren, and C-C Jay Kuo. Image splicing localization using a multi-task fully convolutional network (mfcn). Journal of Visual Communication and Image Representation, 51:201-209, 2018. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 231, + 287, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 231, + 287, + 285 + ], + "spans": [ + { + "bbox": [ + 48, + 231, + 287, + 285 + ], + "type": "text", + "content": "[54] Deepa Shivaram. The white house and big tech companies release commitments on managing ai. https://www.nprillinois.org/2023-07-21/the-white-house-and-big-tech-companies-release-commitments-on-managing-ai, 2023.2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 287, + 287, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 287, + 287, + 363 + ], + "spans": [ + { + "bbox": [ + 48, + 287, + 287, + 363 + ], + "type": "text", + "content": "[55] Xiaopeng Sun, Weiqi Li, Zhenyu Zhang, Qiufang Ma, Xuhan Sheng, Ming Cheng, Haoyu Ma, Shijie Zhao, Jian Zhang, Junlin Li, et al. Opdn: Omnidirectional position-aware deformable network for omnidirectional image superresolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshop, pages 1293-1301, 2023. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 365, + 287, + 419 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 365, + 287, + 419 + ], + "spans": [ + { + "bbox": [ + 48, + 365, + 287, + 419 + ], + "type": "text", + "content": "[56] Zhihao Sun, Haoran Jiang, Danding Wang, Xirong Li, and Juan Cao. Safl-net: Semantic-agnostic feature learning network with auxiliary plugins for image manipulation detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2023. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 421, + 287, + 497 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 421, + 287, + 497 + ], + "spans": [ + { + "bbox": [ + 48, + 421, + 287, + 497 + ], + "type": "text", + "content": "[57] Roman Suvorov, Elizaveta Logacheva, Anton Mashikhin, Anastasia Remizova, Armenii Ashukha, Aleksei Silvestrov, Naejin Kong, Harshith Goka, Kiwoong Park, and Victor Lempitsky. Resolution-robust large mask inpainting with fourier convolutions. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), 2022. 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 499, + 287, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 499, + 287, + 553 + ], + "spans": [ + { + "bbox": [ + 48, + 499, + 287, + 553 + ], + "type": "text", + "content": "[58] Bihan Wen, Ye Zhu, Ramanathan Subramanian, Tian-Tsong Ng, Xuanjing Shen, and Stefan Winkler. Coverage—a novel database for copy-move forgery detection. In Proceedings of the IEEE International Conference on Image Processing (ICIP), 2016. 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 555, + 287, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 555, + 287, + 599 + ], + "spans": [ + { + "bbox": [ + 48, + 555, + 287, + 599 + ], + "type": "text", + "content": "[59] Yuxin Wen, John Kirchenbauer, Jonas Geiping, and Tom Goldstein. Tree-rings watermarks: Invisible fingerprints for diffusion images. Advances in Neural Information Processing Systems, 36, 2024. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 601, + 287, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 601, + 287, + 633 + ], + "spans": [ + { + "bbox": [ + 48, + 601, + 287, + 633 + ], + "type": "text", + "content": "[60] Huikai Wu, Shaocheng Xiang, Gabriben, and Niczem. Faceswap. https://github.com/wuhuikai/FaceSwap, 2020.7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 635, + 287, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 635, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 635, + 287, + 689 + ], + "type": "text", + "content": "[61] Haiwei Wu, Jiantao Zhou, Jinyu Tian, and Jun Liu. Robust image forgery detection over online social network shared images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2, 6, 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "type": "text", + "content": "[62] Xiaoshuai Wu, Xin Liao, and Bo Ou. Sepmark: Deep separable watermarking for unified source tracing and deepfake" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 699 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "text", + "content": "detection. In Proceedings of the ACM international conference on Multimedia (MM), 2023. 2, 3, 6, 7" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 307, + 96, + 545, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 96, + 545, + 150 + ], + "spans": [ + { + "bbox": [ + 307, + 96, + 545, + 150 + ], + "type": "text", + "content": "[63] Yue Wu, Wael AbdAlmageed, and Premkumar Natarajan. *Mantra-net: Manipulation tracing network for detection and localization of image forgeries with anomalous features*. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 2, 6" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 152, + 545, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 152, + 545, + 195 + ], + "spans": [ + { + "bbox": [ + 307, + 152, + 545, + 195 + ], + "type": "text", + "content": "[64] Youmin Xu, Chong Mou, Yujie Hu, Jingfen Xie, and Jian Zhang. Robust invertible image steganography. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 3, 4" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 197, + 545, + 240 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 197, + 545, + 240 + ], + "spans": [ + { + "bbox": [ + 307, + 197, + 545, + 240 + ], + "type": "text", + "content": "[65] Qichao Ying, Zhenxing Qian, Hang Zhou, Haisheng Xu, Xinpeng Zhang, and Siyi Li. From image to image: Immunized image generation. In Proceedings of the ACM international conference on Multimedia (MM), 2021. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 241, + 545, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 241, + 545, + 296 + ], + "spans": [ + { + "bbox": [ + 308, + 241, + 545, + 296 + ], + "type": "text", + "content": "[66] Qichao Ying, Xiaoxiao Hu, Xiangyu Zhang, Zhenxing Qian, Sheng Li, and Xinpeng Zhang. Rwn: Robust watermarking network for image cropping localization. In Proceedings of the IEEE International Conference on Image Processing (ICIP), 2022." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 297, + 545, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 297, + 545, + 342 + ], + "spans": [ + { + "bbox": [ + 308, + 297, + 545, + 342 + ], + "type": "text", + "content": "[67] Qichao Ying, Hang Zhou, Zhenxing Qian, Sheng Li, and Xinpeng Zhang. Learning to immunize images for tamper localization and self-recovery. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 342, + 545, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 342, + 545, + 396 + ], + "spans": [ + { + "bbox": [ + 308, + 342, + 545, + 396 + ], + "type": "text", + "content": "[68] Jiwen Yu, Yinhuai Wang, Chen Zhao, Bernard Ghanem, and Jian Zhang. Freedom: Training-free energy-guided conditional diffusion model. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2023. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 397, + 545, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 397, + 545, + 441 + ], + "spans": [ + { + "bbox": [ + 308, + 397, + 545, + 441 + ], + "type": "text", + "content": "[69] Jiwen Yu, Xuanyu Zhang, Youmin Xu, and Jian Zhang. Cross: Diffusion model makes controllable, robust and secure image steganography. Advances in Neural Information Processing Systems, 36, 2024. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 442, + 545, + 497 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 442, + 545, + 497 + ], + "spans": [ + { + "bbox": [ + 308, + 442, + 545, + 497 + ], + "type": "text", + "content": "[70] Syed Waqas Zamir, Aditya Arora, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, and Ming-Hsuan Yang. Restormer: Efficient transformer for high-resolution image restoration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 5" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 498, + 545, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 498, + 545, + 542 + ], + "spans": [ + { + "bbox": [ + 308, + 498, + 545, + 542 + ], + "type": "text", + "content": "[71] Kai Zhang, Yawei Li, Jingyun Liang, Jiezhang Cao, Yulun Zhang, Hao Tang, Radu Timofte, and Luc Van Gool. Practical blind denoising via swim-conv-unet and data synthesis. arXiv preprint arXiv:2203.13278, 2022. 8" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 544, + 545, + 587 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 544, + 545, + 587 + ], + "spans": [ + { + "bbox": [ + 308, + 544, + 545, + 587 + ], + "type": "text", + "content": "[72] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2023. 1, 7" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 588, + 545, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 588, + 545, + 621 + ], + "spans": [ + { + "bbox": [ + 308, + 588, + 545, + 621 + ], + "type": "text", + "content": "[73] Yunqing Zhao, Tianyu Pang, Chao Du, Xiao Yang, Ngai-Man Cheung, and Min Lin. A recipe for watermarking diffusion models. arXiv preprint arXiv:2303.10137, 2023. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 622, + 545, + 655 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 622, + 545, + 655 + ], + "spans": [ + { + "bbox": [ + 308, + 622, + 545, + 655 + ], + "type": "text", + "content": "[74] Jiren Zhu, Russell Kaplan, Justin Johnson, and Li Fei-Fei. Hidden: Hiding data with deep networks. In European Conference on Computer Vision (ECCV), 2018. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 656, + 545, + 699 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 656, + 545, + 699 + ], + "spans": [ + { + "bbox": [ + 308, + 656, + 545, + 699 + ], + "type": "text", + "content": "[75] Xinshan Zhu, Yongjun Qian, Xianfeng Zhao, Biao Sun, and Ya Sun. A deep learning approach to patch-based image inpainting forensics. Signal Processing: Image Communication, 67:90-99, 2018. 2" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "text", + "content": "11974" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Editable Scene Simulation for Autonomous Driving via Collaborative LLM-Agents/49eb43d9-92f1-4849-a7c5-3da61380ffc6_content_list.json b/2024/Editable Scene Simulation for Autonomous Driving via Collaborative LLM-Agents/49eb43d9-92f1-4849-a7c5-3da61380ffc6_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..805795be3b839143cc4c83b75576ac5fee39a921 --- /dev/null +++ b/2024/Editable Scene Simulation for Autonomous Driving via Collaborative LLM-Agents/49eb43d9-92f1-4849-a7c5-3da61380ffc6_content_list.json @@ -0,0 +1,1906 @@ +[ + { + "type": "text", + "text": "Editable Scene Simulation for Autonomous Driving via Collaborative LLM-Agents", + "text_level": 1, + "bbox": [ + 78, + 130, + 890, + 151 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yuxi Wei $^{1,5*}$ Zi Wang $^{3,5*}$ Yifan Lu $^{1,5*}$ Chenxin Xu $^{1,5*}$ Changxing Liu $^{1,5}$ Hao Zhao $^{4}$ Siheng Chen $^{1,2,5\\dagger}$ Yanfeng Wang $^{1,2}$ Shanghai Jiao Tong University", + "bbox": [ + 202, + 178, + 767, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{3}$ Carnegie Mellon University $^{4}$ Tsinghua University $^{5}$ Multi-Agent Governance & Intelligence Crew (MAGIC) {wyx3590236732, yifan_lu, xcxwakaka, cx-liu}@sjtu.edu.cn {sihengc, wangyanfeng}@sjtu.edu.cn ziwang2@andrew.cmu.edu zhaohao@air.tsinghua.edu.cn", + "bbox": [ + 138, + 234, + 826, + 304 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 339, + 313, + 354 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Scene simulation in autonomous driving has gained significant attention because of its huge potential for generating customized data. However, existing editable scene simulation approaches face limitations in terms of user interaction efficiency, multi-camera photo-realistic rendering and external digital assets integration. To address these challenges, this paper introduces ChatSim, the first system that enables editable photo-realistic 3D driving scene simulations via natural language commands with external digital assets. To enable editing with high command flexibility, ChatSim leverages a large language model (LLM) agent collaboration framework. To generate photo-realistic outcomes, ChatSim employs a novel multi-camera neural radiance field method. Furthermore, to unleash the potential of extensive high-quality digital assets, ChatSim employs a novel multi-camera lighting estimation method to achieve scene-consistent assets' rendering. Our experiments on Waymo Open Dataset demonstrate that ChatSim can handle complex language commands and generate corresponding photo-realistic scene videos. Code can be accessed at: https://github.com/yifanlu0227/ChatSim.", + "bbox": [ + 76, + 369, + 473, + 688 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 78, + 717, + 209, + 732 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Perception [12, 15-17, 38, 66] is the window of an autonomous vehicle into the external environment. To ensure the robustness of the vehicle's perceptual capabilities during both training and testing phases, it necessitates the collection of high-quality perception data in substantial volumes [13, 58]. However, the operation of a fleet for the acquisition of real-world data often incurs prohibitive expenses, particularly for specialized or customized requirements. For instance, in the aftermath of an accident or intervention in", + "bbox": [ + 76, + 741, + 470, + 878 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/369a5d248f75a13a44d4444719f784db88d394a33e5dda862aad944811a3c72c.jpg", + "image_caption": [ + "Figure 1. ChatSim enables the editing of photo-realistic 3D driving scene simulations via language commands." + ], + "image_footnote": [], + "bbox": [ + 517, + 339, + 874, + 526 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "volving an autonomous vehicle, it is imperative to test the vehicle's perception system across a spectrum of similar scenarios. While replicating such scenario data from real-world instances is nearly impossible due to the uncontrollability of actual scenes [9, 52], customized scene simulation emerges as a vital and feasible alternative. It enables the precise modeling of specific conditions without high costs and logistical complexities of real-world data collection [4, 72].", + "bbox": [ + 496, + 564, + 893, + 685 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To effectively simulate customized driving scenes, we identify three key properties as fundamental. First, the simulation should be capable of following sophisticated or abstract demands, thereby facilitating the production. Second, the simulation should generate photo-realistic, view-consistent outcomes, which allow for the closest approximation to vehicle observations in real-world scenarios. Third, it should allow for the integration of external digital assets [6, 44] with their photo-realistic textures and materials while fitting the lighting conditions. This capability would unlock the potential for data expansion by incorporating a wide array of external digital assets, satisfying customized needs.", + "bbox": [ + 496, + 686, + 893, + 868 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "A vast array of significant works have been proposed for scene simulation, yet they fail to meet all three of", + "bbox": [ + 500, + 869, + 893, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Equal contribution.", + "bbox": [ + 94, + 886, + 205, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "† Corresponding author.", + "bbox": [ + 233, + 887, + 357, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "15077", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "these requirements. Traditional graphics engines, such as CARLA [22] and UE [25], offer editable virtual environments with external digital assets, but the data realism is restricted by asset modeling and rendering qualities. Image generation based methods, such as BEVControl [71], DriveDreamer [63], MagicDrive [26], can generate realistic scene images based on various control signals, including BEV maps, bounding boxes and camera poses. However, they struggle to maintain view consistency and face challenges in importing external digital assets due to the absence of 3D spatial modeling. Rendering-based methods have been proposed to obtain photo-realistic and view-consistent scene simulation. Notable examples like UniSim [73] and MARS [68] come equipped with a suite of scene-editing tools. However, these systems require extensive user involvement in every trivial editing step via code implementation, which is ineffective when performing the editing. Furthermore, while they handle vehicles in observed scenarios effectively, their inability to support external digital assets restricts opportunities for data expansion and customization.", + "bbox": [ + 76, + 90, + 472, + 393 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To fulfill the identified requirements, we introduce ChatSim, the first system that enables editable photo-realistic 3D driving scene simulations via natural language commands with external digital assets. To use ChatSim, users simply engage in a conversation with the system, issuing commands through natural language without any involvement in intermediate simulation steps; see Figure 1 for illustration.", + "bbox": [ + 76, + 396, + 472, + 501 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address complex or abstract user commands effectively, ChatSim adopts a large language model (LLM)-based multi-agent collaboration framework. The key idea is to exploit multiple LLM agents, each with a specialized role, to decouple an overall simulation demand into specific editing tasks, thereby mirroring the task division and execution typically founded in the workflow of a human-operated company. This workflow offers two key advantages for scene simulation. First, LLM agents' ability to process human language commands allows for intuitive and dynamic editing of complex driving scenes, enabling precise adjustments and feedback. Second, the collaboration framework enhances simulation efficiency and accuracy by distributing specific editing tasks among specialized agents, ensuring detailed and realistic simulations with improved task completion rates.", + "bbox": [ + 76, + 505, + 472, + 733 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To generate photo-realistic outcomes, we propose McNeRF in ChatSim, a novel neural radiance field method that incorporates multi-camera inputs, offering a broader scene rendering. This integration fully exploits camera setups on vehicles but raises two significant challenges: camera pose misalignment due to asynchronizated trigger times and brightness inconsistency due to different camera exposure times. To address camera pose misalignment, McNeRF uses a multi-camera alignment to reduce extrinsic parameter noises, ensuring rendering quality. To address brightness inconsistency, McNeRF integrates the critical exposure times", + "bbox": [ + 76, + 734, + 472, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "to recover scene radiance in HDR, markedly mitigating the issue of color discrepancies at the intersections of two camera images with different exposure times.", + "bbox": [ + 496, + 90, + 893, + 136 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To import external digital assets with their realistic textures and materials, we propose McLight, a novel multicamera lighting estimation that blends skydome and surrounding lighting. Our skydome estimation restores accurate sun behavior with peak intensity residual connection, enabling the rendering of prominent shadows. For surrounding lighting, McLight queries McNeRF to achieve complex location-specific illumination effects, like those in the tree shade with sunlight being blocked. This significantly improves the rendering realism of the integrated 3D assets.", + "bbox": [ + 496, + 137, + 895, + 287 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We conduct extensive experiments on the Waymo autonomous driving dataset and show that ChatSim generates photo-realistic customized perception data including dangerous corner cases according to various human language commands. Our method is compatible with mixed, highly abstract and multi-round commands. Our method achieves SoTA performance with an improvement of $4.5\\%$ in photorealism with a wide-angle rendering. Moreover, we demonstrate our lighting estimation outperforms the SoTA methods both qualitatively and quantitatively, reducing the intensity error and angular error by $57.0\\%$ and $9.9\\%$ .", + "bbox": [ + 496, + 289, + 895, + 455 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 470, + 640, + 486 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Scene simulation for autonomous driving. Current scene simulation methods can be generally divided into three categories: graphics engines, image generation, and scene rendering. Graphics engines, such as CARLA [22], AirSim [54], OpenScenario Editor [24], 51Sim-One [1] and RoadRunner [21], create a virtual world for simulating a wide range of driving scenarios. However, there exists a significant domain gap between the virtual world and reality. Image generation methods can generate realistic scene images based on different control signals, such as HD maps [26, 39, 59], sketch layout [71], bounding boxes [26, 39, 63], text [26, 33, 39, 63] and driving actions [33, 63]. However, these approaches can hardly maintain scene consistency. To obtain a coherent driving scene, methods based on scene rendering target to reconstruct the 3D scene. READ [41] employs point clouds and uses a U-Net to render images. With the rapid development of Neural Radiance Field (NeRF) [7, 8, 43, 45, 56, 62], several works [28, 34, 47, 48, 61, 68, 69, 73] also exploit NeRFs to model cars and static street backgrounds in outdoor environments. Moreover, notable examples like UniSim [73] and MARS [68] come equipped with a suite of scene-editing tools. However, these methods require extensive user involvement in intermediate editing steps and they fail to support external digital assets for data expansion. In this work, we propose ChatSim that achieves automatic simulation editing via language commands and integrates external digital assets to enhance realism and flexibility. In ChatSim,", + "bbox": [ + 496, + 493, + 895, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "15078", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 1 + }, + { + "type": "table", + "img_path": "images/639f5bd6b6538d953ac0ba6d42130ee9ad15ee7a8a14ec3dae28a98f0020d135.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodPhoto-realisticDim.Multi-cameraEditableExternal assetsLanguageOpen-source
CARLA [22]×3D✓✓✓×✓
AirSim [54]×3D✓✓✓×✓
OpenScenario [24]×3D✓✓✓×✓
51Sim-One [1]×3D✓✓✓××
RoadRunner [21]×3D✓✓✓×✓
BEVGen [59]✓2D✓✓××✓
BEVCentral [71]✓2D✓✓×××
DriveDreamer [63]✓2D✓✓×✓×
DrivingDiffusion [39]✓2D✓✓×✓×
GAIA-1 [33]✓2D×✓×✓×
MagicDrive [26]✓2D✓✓×××
READ [41]✓3D××××✓
Neural SG [47]✓3D×✓××✓
Neural PLF [48]✓3D××××✓
S-NeRF [69]✓3D✓×××✓
UniSim [73]✓3D×✓×××
MARS [68]✓3D×✓××✓
ChatSim (Ours)✓3D✓✓✓✓✓
", + "bbox": [ + 80, + 88, + 465, + 292 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Table 1. Comparison of existing and proposed methods for autonomous driving simulation.", + "bbox": [ + 76, + 295, + 470, + 324 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "we integrate McNeRF, a novel neural radiance field designed to leverage multi-camera inputs for high-fidelity rendering.", + "bbox": [ + 75, + 330, + 468, + 359 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Lighting estimation. Lighting estimation focuses on assessing the illumination conditions of a real-world environment to seamlessly integrate digital objects. Early methods [35, 36] for outdoor environments use explicit cues like detected shadows on the ground. Recent works usually adopt learning-based approaches [27, 30, 31, 37, 40, 74] by predicting different lighting representations like spherical lobes [10, 40], light probes [37], environment map [53, 55], HDR sky model [31, 64, 74] and lighting volume [64]. However, few of them consider multi-camera input, which is common for driving scenarios. In this paper, we propose a novel multi-camera lighting estimation method, McLight, combining with our McNeRF, to estimate a wider range of lighting and obtain the spatially-varying lighting effects of assets.", + "bbox": [ + 75, + 361, + 470, + 573 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Large Language model and collaborative framework. Large Language Models (LLMs) are AI systems trained on extensive data to understand, generate, and respond to human language. GPT [11] is a pioneering work to generate human-like content. The following updated versions GPT-3.5 [14] and GPT-4 [46], provide more intelligent capabilities like chatting, browsing and coding. Notable other large language models include InstructGPT [49], LLaMA [60] and PaLM [5, 18]. Based on LLM, many works [3, 23, 29, 65, 76] improve the problem-solving abilities by integrating communication among multiple agents. [32] and [67] define a group of well-organized agents to form operating procedures with conversation and code programming. In this paper, we exploit the power of collaborative LLM agents in simulation for autonomous driving, enabling the various editing of 3D scenes via language commands.", + "bbox": [ + 75, + 574, + 470, + 816 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Collaborative LLM-Agents for Editing", + "text_level": 1, + "bbox": [ + 76, + 832, + 426, + 849 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The ChatSim system analyzes specific user commands and returns a video that meets customized needs; see Figure 2. Since user commands could be abstract and sophisticated,", + "bbox": [ + 76, + 854, + 470, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "it requires the system to have flexible task-handling ability. Directly applying a single LLM agent struggles with multi-step reasoning and cross-referencing. To address this, we design a series of collaborative LLM agents, where each agent is responsible for a unique aspect of the editing task.", + "bbox": [ + 496, + 90, + 893, + 167 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Specific Agent's Functionality", + "text_level": 1, + "bbox": [ + 498, + 179, + 764, + 196 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Agents in ChatSim comprise two key components: a Large Language Model (LLM) and the corresponding role functions. The LLM is responsible for understanding the received commands while the role functions process the received data. Each agent is equipped with unique LLM prompts and role functions tailored to their specific duties within the system. To accomplish their tasks, agents first convert the received commands to a structured configuration using LLM with the assistance of prompts. Then the role functions utilize the structured configuration as parameters to process the received data and produce the desired outcomes; see an agent example on the right side of Figure 2. This workflow endows agents with both language interpretation capabilities and precise execution capabilities.", + "bbox": [ + 496, + 200, + 893, + 412 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Project Manager Agent. The project manager agent decomposes direct commands into clear natural language instructions dispatched to other editing agents. To equip the project manager agent with the capability of command decomposition, we design a series of prompts for its LLM. The core idea of the prompts is to describe the action set, give the overall goal, and define the output form with examples; The role functions send the decomposed instructions to other agents for editing. The presence of the project manager agent enhances the system's robustness in interpreting various inputs and streamlines operations for clarity and fine granularity.", + "bbox": [ + 496, + 412, + 893, + 580 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Tech agent for view adjustment. The view adjustment agent generates suitable extrinsic camera parameters. The LLM in the agent translates the natural language instructions for viewpoint adjustment into movement parameters to the target viewpoint's position and angle. In role functions, these movement parameters are turned into transformation matrices required by the extrinsic, which are then multiplied by the original parameters to yield a new viewpoint.", + "bbox": [ + 496, + 580, + 893, + 702 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Tech agent for background rendering. The background rendering agent renders the scene background based on multi-camera images. The LLM receives the rendering command and then operates the role functions for rendering. Notably, in role functions, we specifically integrate a novel neural radiance field method (McNeRF) taking multi-camera inputs and considering exposure time, solving the problem of blurring and brightness inconsistency in multi-camera rendering, see more details in Section 4.1.", + "bbox": [ + 496, + 703, + 893, + 838 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Tech agent for vehicle deleting. The vehicle deleting agent removes specified vehicles from the background. It first identifies current vehicle attributes like 3D bounding boxes and colors from given scene information or results from a scene", + "bbox": [ + 496, + 839, + 893, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "15079", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/05f9bf54c389c3fc12db037c655fa35ff99434483de4cc028ef5f24bef591de3.jpg", + "image_caption": [ + "Figure 2. ChatSim system overview. The system exploit multiple collaborative LLM agents with specialized roles to decouple an overall demand into specific editing tasks. Each agent equips an LLM and corresponding role functions to interpret and execute its specific tasks." + ], + "image_footnote": [], + "bbox": [ + 99, + 90, + 867, + 323 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/3332b5392e5399f26ee735fa931df462f0f7f160fd44cf61038e0dcca3811bb1.jpg", + "image_caption": [ + "Figure 3. Prompt example of view adjustment agent." + ], + "image_footnote": [], + "bbox": [ + 86, + 364, + 455, + 484 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "perception model like [42]. The LLM gathers attributes of the vehicles and performs matching with user requests. Upon confirming the targeted vehicles, it employs a per-frame inpainting model as the role functions, such as latent diffusion methods [51], to effectively delete them from the scene.", + "bbox": [ + 75, + 503, + 470, + 579 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Tech agent for 3D asset management. The 3D asset management agent selects and modifies 3D digital assets according to user specifications. It constructs and maintains a 3D digital asset bank; see our bank details in the Appendix. To facilitate the addition of various objects, the agent first uses LLM to select the most suitable asset by key attributes matching with the requirements, such as color and type. If the matching is not perfect, the agent could modify the asset through its role functions like changing the color.", + "bbox": [ + 75, + 582, + 470, + 717 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Tech agent for vehicle motion. The vehicle motion agent creates the initial places and subsequent motions of vehicles following the requests. Existing vehicle motion generation methods cannot directly generate motion purely from text and the scene map. To solve the problem, here we propose a novel text-to-motion method. The key idea is linking a placement and planning module as role functions with LLMs to extract and turn motion attributes into coordinates. Motion attributes include position attributes (e.g., distance, direction) and movement attributes (e.g., speed, action). For the placement module, we endow each lane node in the lane map with its attributes to match with the position attributes. The", + "bbox": [ + 75, + 719, + 472, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "planning module plans the vehicle's approximate destination lane node and then plans the intermediate trajectory by fitting the Bezier curves. We also add trajectory tracking [70] to fit vehicle dynamics; see more details in the Appendix.", + "bbox": [ + 496, + 366, + 893, + 426 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Tech agent for foreground rendering. The foreground rendering agent integrates camera extrinsic information, 3D assets, and motion information to render foreground objects in the scene. Notably, to seamlessly integrate the external assets with the current scene, we design a multi-camera lighting estimation method (McLight) into the role functions, coupling with McNeRF. The estimated illumination is then utilized by Blender API to generate foreground images. The detailed technical aspects will be elaborated in Section 4.2.", + "bbox": [ + 496, + 426, + 892, + 561 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Agent Collaboration Workflow", + "text_level": 1, + "bbox": [ + 500, + 571, + 772, + 588 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Agents with tailored functions collaboratively work together to edit based on user commands. The project manager orchestrates and dispatches instructions to editing agents. The editing agents form two teams: background generation and foreground generation. For background generation, the background rendering agent generates rendered images using the extrinsic parameters from the view adjustment agent, followed by inpainting by the vehicle deleting agent. For foreground generation, the foreground rendering agent renders the images using the extrinsic parameters from the view adjustment agent, selected 3D assets from 3D asset management agent, and generated motions from vehicle motion agent. Finally, the foreground and background images are composed to create and deliver a video to the user. The editing information in each agent's configuration is recorded by the project manager agent for possible multi-round editorings.", + "bbox": [ + 496, + 593, + 893, + 834 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Novel Rendering Methods", + "text_level": 1, + "bbox": [ + 500, + 848, + 743, + 864 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Based on the collaborative LLM agents framework introduced in Section 3, this section presents two novel rendering", + "bbox": [ + 500, + 869, + 893, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "15080", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "techniques to enhance photo-realism in simulations. To tackle the rendering challenges caused by multiple cameras, we propose multi-camera neural radiance field (McNeRF), a novel NeRF model considering the varied camera exposure times for visual consistency. To render realistic external digital assets with location-specific lighting and accurate shadows, we propose McLight, a novel hybrid lighting estimation method combined with our McNeRF. Note that McNeRF and McLight are leveraged by the background rendering agent and the foreground rendering agent, respectively.", + "bbox": [ + 75, + 90, + 472, + 243 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. McNeRF for Background Rendering", + "text_level": 1, + "bbox": [ + 76, + 260, + 395, + 277 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "An autonomous vehicle typically equips multiple cameras to achieve a comprehensive perception view. However, this poses challenges for NeRF training due to the misaligned multi-camera poses from asynchronized camera trigger times and the brightness inconsistency originating from different exposure times. To address these challenges, the proposed McNeRF uses two techniques: multi-camera alignment and brightness-consistent rendering.", + "bbox": [ + 75, + 287, + 468, + 407 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Multi-camera alignment. Autonomous vehicles, despite having a localization module for accurate camera poses, face challenges with asynchronous trigger times across multiple cameras. To align camera extrinsics for NeRF training, our core idea is to leverage a consistent spatial coordinate system provided by Agisoft Metashape [2] to align the images captured by multiple cameras at different timestamps.", + "bbox": [ + 75, + 411, + 468, + 517 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Specifically, let $\\mathcal{I}^{(i,k)}$ and $\\xi^{(i,k)}$ be the image captured by the $i$ th camera at the $k$ th trigger and the corresponding camera pose in the vehicle's global coordinate space, respectively. We first input all images into Metashape for recalibration. The aligned camera pose is then obtained as:", + "bbox": [ + 75, + 518, + 470, + 595 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\widehat {\\xi} ^ {(i, k)} = T _ {M \\to G} \\cdot \\xi_ {M} ^ {(i, k)},\n$$\n", + "text_format": "latex", + "bbox": [ + 191, + 595, + 351, + 616 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\xi_{M}^{(i,k)}$ denotes the recalibrated camera pose in the Metashape's unified spatial coordinate space, and $T_{M\\rightarrow G}$ is the transformation from the Metashape's coordinate space to the vehicle's global coordinate space. After alignment, the pose noise can be significantly reduced. Then, the aligned camera pose $\\widehat{\\xi}^{(i,t)}$ can be used to generate the origins and directions of rays for McNeRF, enabling high-fidelity rendering. The aligned pose can also facilitate the foreground rendering agent's operations.", + "bbox": [ + 75, + 619, + 468, + 758 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Brightness-consistent rendering. The exposure times of cameras can differ substantially, causing significant brightness differences across images, hindering the NeRF training. As shown in Figure 4, McNeRF, addresses this by incorporating exposure times into HDR radiance fields, prompting brightness consistency.", + "bbox": [ + 75, + 761, + 470, + 851 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We adopt F2-NeRF [62] as our backbone model to handle the unbounded scene, sampling $K$ points along the ray $\\mathbf{r}$ and estimating each point's HDR radiance $\\mathbf{e}_k$ and density $\\sigma_k$ .", + "bbox": [ + 75, + 854, + 470, + 901 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/024a3e27b985c79138121fbda6995993109f49d829e053842b6a1b358f7f57b6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 504, + 90, + 883, + 181 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/7f9a739e4c4e43f357b42161a700e97bab79b2aa9ba7680c886f357178d0a0a1.jpg", + "image_caption": [ + "Figure 4. Rendering framework. The main components include McNeRF and McLight. Background rendering uses McNeRF to predict HDR pixel value and convert it to LDR with sRGB OETF. McLight includes a skydome lighting estimation network and adopts McNeRF to generate surrounding lighting." + ], + "image_footnote": [], + "bbox": [ + 504, + 181, + 883, + 281 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The HDR light intensity is then calculated as:", + "bbox": [ + 498, + 359, + 802, + 373 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\widehat {\\mathcal {I}} _ {\\mathrm {H D R}} (\\mathbf {r}) = f (\\Delta t) \\cdot \\sum_ {k = 1} ^ {K} T _ {k} \\alpha_ {k} \\mathbf {e} _ {k}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 581, + 375, + 890, + 416 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\alpha_{k} = 1 - \\exp (-\\sigma_{k}\\delta_{i})$ is the opacity, $\\delta_{i}$ is the point sampling interval, $T_{k} = \\prod_{i = 0}^{k - 1}(1 - \\alpha_{i})$ is the accumulated transmittance and $\\Delta t$ is the exposure time. The normalization function $f(\\Delta t) = 1 + \\epsilon (\\Delta t - \\mu) / \\sigma$ is designed to stabilize training, where $\\epsilon$ is a hyperparameter for scaling, $\\mu$ and $\\sigma$ are the mean and standard deviation of the exposure times of all images, respectively.", + "bbox": [ + 496, + 417, + 892, + 525 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "By predicting scene radiance in HDR and multiplying it by the exposure time, we recover the light intensity received by the sensor and tackle the inconsistent color supervision at the intersections of two camera images with distinct exposure times. Moreover, the HDR light intensity outputted by McNeRF can provide scene-level illumination for foreground object rendering, a topic further discussed in Section 4.2.", + "bbox": [ + 496, + 526, + 893, + 631 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To train the rendering network, we enforce the consistency of radiance between the rendered image (prediction) and the captured image (ground-truth). Given the ground-truth image $\\mathcal{I}$ , the loss function is then:", + "bbox": [ + 496, + 631, + 893, + 691 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\frac {1}{| R |} \\sum_ {\\mathbf {r} \\in R} \\left(\\mathrm {O E T F} \\left(\\widehat {\\mathcal {I}} _ {\\mathrm {H D R}} (\\mathbf {r})\\right) - \\mathcal {I} (\\mathbf {r})\\right) ^ {2},\n$$\n", + "text_format": "latex", + "bbox": [ + 540, + 693, + 846, + 729 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $R$ represents the ray set and $\\mathrm{OETF}(\\cdot)$ is the sRGB opto-electronic transfer function (gamma correction) [19] that converts HDR light intensity to LDR colors.", + "bbox": [ + 496, + 729, + 892, + 776 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. McLight for Foreground Rendering", + "text_level": 1, + "bbox": [ + 498, + 786, + 810, + 803 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To enrich the scene's content with substantial digital 3D assets, we employ Blender [20] foreground virtual objects' rendering. A seamless insertion critically depends on accurately estimating the scene's illumination conditions. Thus, as shown in Figure 4, we propose McLight, a novel hybrid lighting estimation consisting of skydome lighting and sur", + "bbox": [ + 496, + 809, + 893, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "15081", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "rounding lighting.", + "bbox": [ + 76, + 90, + 199, + 106 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Skydome lighting estimation. Estimating skydome lighting from images is challenging for restoring accurate sun behavior. To achieve this, we propose a novel residual connection from the estimated peak intensity to the HDR reconstruction to address over-smoothing output. Further, we adopt a self-attention mechanism to fuse multi-camera inputs, capturing complementary visual cues.", + "bbox": [ + 75, + 106, + 470, + 210 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Here we employ a two-stage process. In the first stage, we train an autoencoder to reconstruct the corresponding HDR panorama from an LDR panorama. Following [64], the encoder transforms the LDR skydome panorama into three intermediate vectors, including the peak direction vector $\\mathbf{f}_{\\mathrm{dir}} \\in \\mathbb{R}^3$ , the intensity vector $\\mathbf{f}_{\\mathrm{int}} \\in \\mathbb{R}_+^3$ , and the sky content vector $\\mathbf{f}_{\\mathrm{content}} \\in \\mathbb{R}^{64}$ . However, as HDR intensity behaves like an impulse response at its peak position, with pixel values thousands of times higher than its neighbors, it is difficult for the decoder to recover such patterns. To tackle this, we design a residual connection that injects $\\mathbf{f}_{\\mathrm{int}}$ into the decoded HDR panorama with a spherical Gaussian lobe attenuation. This explicitly restores the peak intensity of the sun in the reconstructed HDR panorama, allowing us to render strong shadows for virtual objects.", + "bbox": [ + 75, + 212, + 468, + 438 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In the second stage, we train an image encoder and a multi-camera fusion module built upon the pretrained decoder from the first stage. Specifically, for images from each camera, a shared image encoder predicts the peak direction vector $\\mathbf{f}_{\\mathrm{dir}}^{(i)}$ , the intensity vector $\\mathbf{f}_{\\mathrm{int}}^{(i)}$ , and the sky content vector $\\mathbf{f}_{\\mathrm{content}}^{(i)}$ for each image $\\mathcal{I}^{(i)}$ , where $i$ is the camera index. We design the latent vector fusion across the multiple camera views as follows: all $\\mathbf{f}_{\\mathrm{dir}}^{(i)}$ are aligned to the front-facing view using their extrinsic parameters and averaged to form $\\bar{\\mathbf{f}}_{\\mathrm{dir}}$ ; all $\\mathbf{f}_{\\mathrm{int}}^{(i)}$ are averaged to yield $\\bar{\\mathbf{f}}_{\\mathrm{int}}$ ; all $\\mathbf{f}_{\\mathrm{content}}^{(i)}$ are integrated into $\\bar{\\mathbf{f}}_{\\mathrm{content}}$ through a self-attention module. Finally, the pretrained decoder reconstructs the HDR skydome image $\\mathcal{I}_{\\mathrm{skydome}}$ from $\\bar{\\mathbf{f}}_{\\mathrm{dir}}$ , $\\bar{\\mathbf{f}}_{\\mathrm{int}}$ and $\\bar{\\mathbf{f}}_{\\mathrm{content}}$ .", + "bbox": [ + 75, + 439, + 470, + 643 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Compared to alternative approaches [31, 64], our multicamera sky dome estimation technique accurately reproduces the sun's intensity response behavior at its peak with our residual designs, significantly improving the accuracy and fidelity of the skydome reconstruction.", + "bbox": [ + 75, + 643, + 470, + 718 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Surrounding lighting estimation. Merely modeling the sky-dome cannot replicate the complex location-specific lighting effects, like those in the shade with sunlight blocked by trees or buildings. Our McNeRF is capable of storing precise 3D scene information, enabling us to capture the surrounding scene's impact on lighting. This approach facilitates the achievement of spatially-varying lighting estimation. Specifically, we sample the hemisphere rays at the virtual object's position $\\mathbf{o}$ . The rays' directions, $\\mathbf{d}_i$ , $i = 0,1,\\dots,h\\times w$ , are aligned with pixel coordinates on a unit sphere using equirectangular projection from an environment map, where $h$ and $w$ are map's height and width. With the ray", + "bbox": [ + 75, + 719, + 470, + 901 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/22006f29ef30b26d3218bf91ba48a7fddcd5eba050b07813c70aee86e212003d.jpg", + "image_caption": [ + "Front left" + ], + "image_footnote": [], + "bbox": [ + 516, + 99, + 630, + 157 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/911ac196eb6cba8fdf316d43be5a36f0d9a14ef7a2bec755b9cb0886e9ea1c90.jpg", + "image_caption": [ + "Front" + ], + "image_footnote": [], + "bbox": [ + 637, + 98, + 751, + 157 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/e19903211766eff71f549d3d33153a6bfdca3335d06d8ebd131424b6936e5b46.jpg", + "image_caption": [ + "Front right" + ], + "image_footnote": [], + "bbox": [ + 759, + 98, + 872, + 157 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/b091b38369abab6351eb8e2f4c0ef52abc36ffaf6f7c01c133713e6b629af4c6.jpg", + "image_caption": [ + "Command: \"Remove all cars in the scene and add a Porsche driving the wrong way toward me fast. Additionally, add a police car also driving the wrong way and chasing behind the Porsche. The view should be moved 5 meters ahead and 0.5 meters above.\"", + "Figure 5. Editing result under a complex and mixed command." + ], + "image_footnote": [], + "bbox": [ + 516, + 159, + 872, + 343 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "$\\mathbf{r} = \\mathbf{o} + t\\mathbf{d}_i$ , we query our McNeRF as Equation 1 to obtain HDR surrounding lighting $\\mathcal{I}_{\\mathrm{surround}}(\\mathbf{o},\\mathbf{d}_i)$ . The surrounding lighting estimation reconstructs complex environmental lighting, achieving a spatially varying effect and high consistency with the background.", + "bbox": [ + 498, + 412, + 893, + 487 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Blending. We blend the HDR intensity value from the sky-dome and surrounding lighting by transmittance of the final sampling point from McNeRF. The idea is that the rays emitted outside the radiance fields will definitely hit the sky-dome. Given the direction $\\mathbf{d}_i$ , we retrieve the skydome's intensity $\\mathcal{I}_{\\mathrm{skydome}}(\\mathbf{d}_i)$ with equirectangular projection. The final HDR light intensity $\\mathcal{I}_{\\mathrm{env}}(\\mathbf{o},\\mathbf{d}_i)$ is a combination of scene and skydome:", + "bbox": [ + 498, + 488, + 893, + 609 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {I} _ {\\text {e n v}} (\\mathbf {o}, \\mathbf {d} _ {i}) = \\mathcal {I} _ {\\text {s u r r o u n d}} (\\mathbf {o}, \\mathbf {d} _ {i}) + T _ {K} \\mathcal {I} _ {\\text {s k y d o m e}} (\\mathbf {d} _ {i}),\n$$\n", + "text_format": "latex", + "bbox": [ + 527, + 611, + 861, + 628 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $T_{K}$ is the last sampling point's transmittance.", + "bbox": [ + 500, + 628, + 843, + 643 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "McLight offers two main advantages: i) it explicitly recovers the illuminance behavior at the peak and use complementary information from multiple cameras to restore accurate skydome; and ii) it enables location-specific lighting with consideration of complex scene structures.", + "bbox": [ + 498, + 645, + 893, + 720 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Experimental Results", + "text_level": 1, + "bbox": [ + 500, + 734, + 705, + 751 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1. Datasets and implementation details", + "text_level": 1, + "bbox": [ + 500, + 758, + 813, + 775 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We demonstrate a variety of results mainly on the Waymo Open Dataset [57], which contains high-quality multicamera images and the corresponding calibrations. For McLight skydome estimation, we collect 449 HDRIs from online HDRI databases for the autoencoder training and use HoliCity [75], a street view panorama dataset for the second stage; see more dataset details in the Appendix.", + "bbox": [ + 498, + 779, + 893, + 883 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In our experiment, we use front, left front, and right front", + "bbox": [ + 517, + 885, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "15082", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/5bf4eea0fd9debc5a80d87e52d41fe9652ea193fc3f60c63d06cedfeb3463bad.jpg", + "image_caption": [ + "Front left" + ], + "image_footnote": [], + "bbox": [ + 93, + 97, + 207, + 157 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/88b45533493fe22c5aebd9727124043b947a6e88e0bf1c6d7c46ed0766439a47.jpg", + "image_caption": [ + "Front" + ], + "image_footnote": [], + "bbox": [ + 212, + 97, + 328, + 157 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f2352fabec77c98e64a729f724c314460506a2f9a7df1db37b598d2cfcf797d3.jpg", + "image_caption": [ + "Front right" + ], + "image_footnote": [], + "bbox": [ + 334, + 97, + 450, + 157 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/2f3d4818138188034491308613a4e64292983139cbd718eea3df731de29a8591.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 93, + 157, + 450, + 215 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/8750f1a59336fa996b2a6810a2fbeca36c4caad4b20c74a52315b933319234e8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 93, + 217, + 450, + 277 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/24f44ef39b8a2fe213825262032ceda185cf8b6c6479fbd001923d42aea24510.jpg", + "image_caption": [ + "Command: \"Create a traffic jam.\"", + "Figure 6. Editing result under a highly abstract command." + ], + "image_footnote": [], + "bbox": [ + 93, + 280, + 450, + 340 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "cameras in each frame. During the rendering process, we choose 40 frames per scene at a $10\\mathrm{Hz}$ sampling rate, totaling 120 images. We evenly select $1/8$ of these as the test set, with the remainder used for training. The input images are used at the dataset's initial resolution of $1920 \\times 1280$ ; we employ GPT-4 as the LLMs in all of our experiments; see more implementation details in the Appendix.", + "bbox": [ + 75, + 381, + 468, + 488 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.2. System results", + "text_level": 1, + "bbox": [ + 76, + 498, + 223, + 513 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Editing via language commands. We select three representative commands to demonstrate the editing results. All of the results demonstrate we achieve photo-realistic wide angle results, thanks to McNeRF and McLight.", + "bbox": [ + 75, + 521, + 468, + 580 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Mixed and complex command. We send the system with a mixed and complex command, implying that a police car is chasing a wrong-way racer. The target scene, command and the result are shown in Figure 5. We see that i) every requirement in the complex command is accurately executed thanks to our multi-agent collaboration design; ii) this command successfully simulates one rare but dangerous driving condition, which is significant in accident testing.", + "bbox": [ + 75, + 582, + 468, + 703 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Highly abstract command. The second type is a highly abstract command. The inputs and results are presented in Figure 6. We see that i) this highly abstract command is hard to decompose by sentence division but still can be correctly executed by our method, and ii) our 3D asset bank offers a large variety of objects for addition.", + "bbox": [ + 75, + 704, + 468, + 794 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Multi-round command. We also perform a multi-round chat with our system, and the commands in different rounds exist context dependencies. The final results are shown in Figure 7. We see that i) our system is well-equipped to handle multi-round commands and execute the commands in each round precisely; ii) our system can handle the context dependencies across different rounds thanks to the recording", + "bbox": [ + 75, + 795, + 470, + 901 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e4c5778bba8772af279f5d2bea99dba1dec6a0777e6ce27cd9ba94d467452580.jpg", + "image_caption": [ + "Front left" + ], + "image_footnote": [], + "bbox": [ + 503, + 99, + 609, + 155 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/bb3473fdec38bcc1dabfc85a69d7790b4777373c0e913c07731f38a4fa8030da.jpg", + "image_caption": [ + "Front" + ], + "image_footnote": [], + "bbox": [ + 640, + 99, + 746, + 155 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d608debd70fea97742d077397ad49aca403f5c604ed008564ec6b404a73257f5.jpg", + "image_caption": [ + "Front right" + ], + "image_footnote": [], + "bbox": [ + 776, + 99, + 882, + 155 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/5f523b0e48638bdba808421b135cb871cbb83d880f9e559cacc082a4923b763e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 503, + 157, + 691, + 222 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/3842b951399e78dfbcee77369c218814ea73c4e67cdcbd06488fba96dd74e7cf.jpg", + "image_caption": [ + "Command 1: \"Ego vehicle drives ahead slowly. Add a car to the close front that is moving ahead.\"" + ], + "image_footnote": [], + "bbox": [ + 503, + 222, + 691, + 319 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e02f9db9cea4e58a76eb8915bd9f894b2cea9d69eb6a1c0c1ff9462f90dd83c8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 692, + 157, + 882, + 222 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/06021c08aa05d33110e195aa01dcdcd49f9bc171a5bbf749b81dc5f46fe7d801.jpg", + "image_caption": [ + "Command 2: \"Modify the added car to turn left. Add a Chevrolet to the front of the added car. Add another vehicle to the left of the added Mini driving toward me.\"" + ], + "image_footnote": [], + "bbox": [ + 692, + 222, + 882, + 318 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/4cfd2cc41757614acabd5e77c875f05d89aeb79398df87090e2c313eed16bea4.jpg", + "image_caption": [ + "Figure 7. Editing result under multi-round commands.", + "Figure 8. Comparison of detection performance w/o and with our simulated data under different amounts of real data during training." + ], + "image_footnote": [], + "bbox": [ + 508, + 387, + 625, + 472 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/905eb43f45871bb8d325fee54b0b81cd9d099a391d5ba7dd6e269cfc0a09549c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 633, + 387, + 751, + 472 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/7ac17b933135773c0bebdac17ec94c71dc4f72a0852e53208171a60d3f2d3368.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 751, + 387, + 875, + 472 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "ability of the project manager agent.", + "bbox": [ + 500, + 508, + 740, + 523 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3D detection with simulation data. We validate the benefits of our simulation as data augmentation for a downstream 3D object detection task on Waymo Open Dataset [58]. We simulate 1960 frames, derived from scenes in the training dataset. In the simulation, cars with various types, locations, and orientations are incorporated. The detection model adopts Lift-Splat [50]. Figure 8 shows detection performances with and without fixed augmentation under various amounts of real data. We see that i) a significant and consistent improvement across different data sizes is achieved; ii) when real data is limited, our simulation notably aids in rough detection (AP30); iii) when the amount of real data increases, our simulation further significantly improves fine-grained detection (AP70), reflecting the high-quality of our simulation.", + "bbox": [ + 496, + 523, + 893, + 736 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.3. Component results", + "text_level": 1, + "bbox": [ + 500, + 744, + 681, + 760 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Multi-agent collaboration. We evaluated the effectiveness of the multi-agent collaboration by checking whether the command is successfully executed in Table 2. In scenarios without multi-agent collaboration, all operations are executed by a single LLM agent. We see that a single LLM agent leads to notably lower execution accuracy across all categories due to process limitations. In contrast, the collaborative multi-agent approach can manage most commands, attributed to its task division and agent role specificity.", + "bbox": [ + 496, + 763, + 893, + 901 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "15083", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/4caa2587df8a2367762f47bc8ac637b0973c255f5542aa592d3e64e7bd23423f.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Multi-agent collaborationLanguage command category
DeletionAdditionView changeRevisionAbstract
✓0.6170.3830.7170.3670.216
0.9830.8670.9670.9170.883
", + "bbox": [ + 98, + 89, + 450, + 143 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/6988eff828e2eb6c2cd4eef57bd61d71e36647ae8f52ace0105c019e29da90c1.jpg", + "table_caption": [ + "Table 2. The accuracy $(\\%)$ of task completion by LLM without and with multi-agent collaboration." + ], + "table_footnote": [], + "table_body": "
MethodsPSNR↑SSIM↑LPIPS↓Inf. time (s)↓
DVGO [56]23.570.7700.5087.7
Mip-NeRF360 [8]24.400.7540.528101.8
S-NeRF [69]24.710.7590.519114.5
F2NeRF [62]23.260.7730.4392.4
Ours w/o alignment23.320.7760.4372.5
Ours w/o exposure25.180.8190.3812.4
McNeRF (Ours)25.820.8220.3782.5
", + "bbox": [ + 99, + 176, + 444, + 279 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/0b85171da357d5b1be05d0a7ff7e0e46dd219956fc9bb7cc47b073fc4b23473a.jpg", + "image_caption": [ + "Figure 9. Comparisons of wide-angle images generation. (a) S-NeRF.(b) F2NeRF. (c) McNeRF (Ours). Last row: target images." + ], + "image_footnote": [], + "bbox": [ + 84, + 300, + 457, + 558 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Background rendering. We compare our McNeRF with several other state-of-the-art methods on the background novel view synthesis task. We perform reconstruction and rendering on 32 selected scenes. Table 3 shows the quantitative results comparison on three metrics: PSNR, SSIM, and LPIPS. We see that i) McNeRF achieves SoTA performance on all three metrics, significantly outperforming other baselines; ii) McNeRF has a fast inference speed, enabling quick responses to user requests for image rendering.", + "bbox": [ + 75, + 598, + 468, + 733 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Figure 9 demonstrates qualitative comparisons between other methods and ours. We see that existing NeRF methods do not consider the exposure time, leading to noticeable changes in brightness at the junctions of different cameras in the image, as well as an overall inconsistency in exposure across the wide-angle view. Our method can make the brightness of the entire image more consistent.", + "bbox": [ + 75, + 734, + 470, + 839 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Foreground rendering. We compare our McLight with the other two SoTA methods [31, 64]. Table 4 shows the comparison of relative intensity(log 10) error on our HDRI dataset, angular error on HoliCity [75], and user study. We see that", + "bbox": [ + 75, + 839, + 470, + 901 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/a62ba856d2e7105826c19c9872d9e137b225b7a3979e893d40c02f7d3ffb9a20.jpg", + "table_caption": [ + "Table 3. Background novel view rendering performance evaluation." + ], + "table_footnote": [], + "table_body": "
MethodPeak Intensity(log10) ErrorPeak Angular Error (deg)User study(%) ↑
Mean ↓Median ↓Mean ↓Median ↓
Hold-Geoffroy et al. [31]0.8990.97548.451.619.5
Wang et al. [64]0.5900.62833.529.437.3
McLight (Ours)0.4490.27032.326.543.1
", + "bbox": [ + 503, + 89, + 883, + 140 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 4. Comparison with previous methods on lighting estimation.", + "bbox": [ + 500, + 140, + 888, + 154 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/040a647041766cdad1cfb156f7c86f9ff6980b2de7daa1aae433af54102c829c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 514, + 157, + 602, + 255 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/b7ceb900ce547806bd527b7d1f9d3769b75bce40061dda59456290851f57fd86.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 604, + 157, + 687, + 255 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/ac2c7f5dc676175ca7c8baea74a45e917bfd0d38ffe62903fd152edac5040bb8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 691, + 157, + 779, + 255 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/2b4e959904e821cd521964733264a315c2842aea89eda6db7528caed6bf9c7d2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 785, + 157, + 872, + 255 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/24a02b02c09057bc5a763463a261d2d990bfd4256e193ec081b99c45a4316e27.jpg", + "table_caption": [ + "Figure 10. Comparison with different lighting estimation methods." + ], + "table_footnote": [], + "table_body": "
MethodsStraightLeft TurnRight TurnSpeedWithin-road
GPT2Code0.7380.5590.5360.8930.214
GPT2Motion0.5950.1190.1670.3450.277
Ours0.9880.9400.9760.9521.000
", + "bbox": [ + 511, + 273, + 879, + 321 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 5. Comparison with motion generation from text methods.", + "bbox": [ + 501, + 321, + 885, + 335 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "McLight achieves more accurate peak behavior and receives noticeably higher user preferences. Figure 10 shows the visualizations of vehicle insertion. The vehicles added through McLight feature significantly more realistic reflections and strong shadows consistent with the scene.", + "bbox": [ + 498, + 340, + 893, + 415 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Vehicle motion. As shown in Table 5, we compare the motion generation method from user commands with two of our designed baselines: 1. GPT2Motion, which directly uses LLM to return the motion coordinates; 2. GPT2Code, which first generates code using LLM and executes it to obtain the vehicle motion. We validate multiple actions in multiple scenarios and report the user study result. The user study is to determine if the generated motions matched the command intents and fitted with the lane map. We see that our method demonstrated a significant advantage in generating motions from language commands. Additionally, it maintained a high rate of keeping the trajectories within the lane boundaries.", + "bbox": [ + 498, + 416, + 892, + 598 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusions and Limitations", + "text_level": 1, + "bbox": [ + 500, + 603, + 764, + 619 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This paper introduces ChatSim, the first system for editing 3D driving scene simulations via language commands and realistic rendering with import of external digital assets. To effectively execute user commands, ChatSim adopts an LLM-agent collaboration workflow. To promote photorealistic simulation, we propose McNeRF and McLight for background and foreground rendering, respectively, accommodating multi-camera inputs. Experiments show that ChatSim successfully simulates customized data via language commands, achieving high-quality, photo-realistic outcomes. In future, we plan to integrate more background editing functionalities to ChatSim, such as weather changes.", + "bbox": [ + 496, + 625, + 893, + 806 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgement", + "text_level": 1, + "bbox": [ + 500, + 820, + 660, + 837 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This research is supported by NSFC under Grant 62171276 and the Science and Technology Commission of Shanghai Municipal under Grant 2151100900 and 22DZ2229005.", + "bbox": [ + 498, + 844, + 893, + 887 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "15084", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] 51Sim-One. https://wdp.51aes.com/news/27.2,3", + "[2] LLC Agisoft. Metashape-photogrammetric processing of digital images and 3d spatial data generation, 2019. 5", + "[3] Elif Akata, Lion Schulz, Julian Coda-Forno, Seong Joon Oh, Matthias Bethge, and Eric Schulz. Playing repeated games with large language models. arXiv preprint arXiv:2305.16867, 2023. 3", + "[4] Alexander Amini, Igor Gilitschenski, Jacob Phillips, Julia Moseyko, Rohan Banerjee, Sertac Karaman, and Daniela Rus. Learning robust control policies for end-to-end autonomous driving from data-driven simulation. IEEE Robotics and Automation Letters, 5(2):1143-1150, 2020. 1", + "[5] Rohan Anil, Andrew M Dai, Orhan First, Melvin Johnson, Dmitry Lepikhin, Alexandre Passos, Siamak Shakeri, Emanuel Taropa, Paige Bailey, Zhifeng Chen, et al. Palm 2 technical report. arXiv preprint arXiv:2305.10403, 2023. 3", + "[6] Natalie M Banta. Property interests in digital assets: The rise of digital feudalism. Cardozo L. Rev., 38:1099, 2016. 1", + "[7] Jonathan T Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P Srinivasan. Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5855–5864, 2021. 2", + "[8] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Mip-nerf 360: Unbounded anti-aliased neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5470–5479, 2022. 2, 8", + "[9] Luca Bergamini, Yawei Ye, Oliver Scheel, Long Chen, Chih Hu, Luca Del Pero, Băjej Osiński, Hugo Grimmett, and Peter Ondruska. Simnet: Learning reactive self-driving simulations from real-world observations. In 2021 IEEE International Conference on Robotics and Automation (ICRA), pages 5119-5125. IEEE, 2021. 1", + "[10] Mark Boss, Varun Jampani, Kihwan Kim, Hendrik Lensch, and Jan Kautz. Two-shot spatially-varying brdf and shape estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3982-3991, 2020. 3", + "[11] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. Advances in neural information processing systems, 33:1877-1901, 2020. 3", + "[12] Holger Caesar, Varun Bankiti, Alex H Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuscenes: A multimodal dataset for autonomous driving. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11621-11631, 2020. 1", + "[13] Ming-Fang Chang, John Lambert, Patsorn Sangkloy, Jagjeet Singh, Slawomir Bak, Andrew Hartnett, De Wang, Peter Carr, Simon Lucey, Deva Ramanan, et al. Argoverse: 3d tracking and forecasting with rich maps. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8748-8757, 2019. 1" + ], + "bbox": [ + 78, + 114, + 470, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[14] ChatGPT. https://openai.com/blog/chatgpt.3", + "[15] Chenyi Chen, Ari Seff, Alain Kornhauser, and Jianxiong Xiao. Deepdriving: Learning affordance for direct perception in autonomous driving. In Proceedings of the IEEE international conference on computer vision, pages 2722-2730, 2015. 1", + "[16] Xiaozhi Chen, Kaustav Kundu, Ziyu Zhang, Huimin Ma, Sanja Fidler, and Raquel Urtasun. Monocular 3d object detection for autonomous driving. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2147-2156, 2016.", + "[17] Xiaozhi Chen, Huimin Ma, Ji Wan, Bo Li, and Tian Xia. Multi-view 3d object detection network for autonomous driving. In Proceedings of the IEEE conference on Computer Vision and Pattern Recognition, pages 1907-1915, 2017. 1", + "[18] Aakanksha Chowdhery, Sharan Narang, Jacob Devlin, Maarten Bosma, Gaurav Mishra, Adam Roberts, Paul Barham, Hyung Won Chung, Charles Sutton, Sebastian Gehrmann, et al. Palm: Scaling language modeling with pathways. arXiv preprint arXiv:2204.02311, 2022. 3", + "[19] International Electrotechnical Commission et al. Iec 61966-2-1: 1999 multimedia systems and equipment- colour measurement and management- part 2-1: Colour management-default rgb colour space-srgb, 1999. 5", + "[20] Blender Online Community. Blender - a 3D modelling and rendering package. Blender Foundation, Stichting Blender Foundation, Amsterdam, 2018. 5", + "[21] Valter Crescenzi, Giansalvatore Mecca, Paolo Merialdo, et al. Roadrunner: Towards automatic data extraction from large web sites. In VLDB, pages 109-118, 2001. 2, 3", + "[22] Alexey Dosovitskiy, German Ros, Felipe Codevilla, Antonio Lopez, and Vladlen Koltun. Carla: An open urban driving simulator. In Conference on robot learning, pages 1-16. PMLR, 2017. 2, 3", + "[23] Yilun Du, Shuang Li, Antonio Torralba, Joshua B Tenenbaum, and Igor Mordatch. Improving factuality and reasoning in language models through multiagent debate. arXiv preprint arXiv:2305.14325, 2023. 3", + "[24] Openscanerio Editor. https://github.com/ebadi/openscenarioeditor. 2,3", + "[25] Unreal Engine. https://www.unrealengine.com/. 2", + "[26] Ruiyuan Gao, Kai Chen, Enze Xie, Lanqing Hong, Zhenguo Li, Dit-Yan Yeung, and Qiang Xu. Magicdrive: Street view generation with diverse 3d geometry control. arXiv preprint arXiv:2310.02601, 2023. 2, 3", + "[27] Mathieu Garon, Kalyan Sunkavalli, Sunil Hadap, Nathan Carr, and Jean-François Lalonde. Fast spatially-varying indoor lighting estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6908-6917, 2019. 3", + "[28] Jianfei Guo, Nianchen Deng, Xinyang Li, Yeqi Bai, Botian Shi, Chiyu Wang, Chenjing Ding, Dongliang Wang, and Yikang Li. Streetsurf: Extending multi-view implicit surface reconstruction to street views. arXiv preprint arXiv:2306.04988, 2023. 2", + "[29] Rui Hao, Linmei Hu, Weijian Qi, Qingliu Wu, Yirui Zhang, and Liqiang Nie. Chatllm network: More brains, more intelligence. arXiv preprint arXiv:2304.12998, 2023. 3" + ], + "bbox": [ + 501, + 92, + 915, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "15085", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[30] Yannick Hold-Geoffroy, Kalyan Sunkavalli, Sunil Hadap, Emiliano Gambaretto, and Jean-François Lalonde. Deep outdoor illumination estimation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7312–7321, 2017. 3", + "[31] Yannick Hold-Geoffroy, Akshaya Athawale, and Jean-François Lalonde. Deep sky modeling for single image outdoor lighting estimation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6927–6935, 2019. 3, 6, 8", + "[32] Sirui Hong, Xiawu Zheng, Jonathan Chen, Yuheng Cheng, Ceyao Zhang, Zili Wang, Steven Ka Shing Yau, Zijuan Lin, Liyang Zhou, Chenyu Ran, et al. Metagpt: Meta programming for multi-agent collaborative framework. arXiv preprint arXiv:2308.00352, 2023. 3", + "[33] Anthony Hu, Lloyd Russell, Hudson Yeo, Zak Murez, George Fedoseev, Alex Kendall, Jamie Shotton, and Gianluca Corrado. Gaia-1: A generative world model for autonomous driving. arXiv preprint arXiv:2309.17080, 2023. 2, 3", + "[34] Abhijit Kundu, Kyle Genova, Xiaoqi Yin, Alireza Fathi, Caroline Pantofaru, Leonidas J Guibas, Andrea Tagliasacchi, Frank Dellaert, and Thomas Funkhouser. Panoptic neural fields: A semantic object-aware neural scene representation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12871-12881, 2022. 2", + "[35] Jean-François Lalonde and Iain Matthews. Lighting estimation in outdoor image collections. In 2014 2nd international conference on 3D vision, pages 131-138. IEEE, 2014. 3", + "[36] Jean-François Lalonde, Srinivasa G Narasimhan, and Alexei A Efros. What do the sun and the sky tell us about the camera? International Journal of Computer Vision, 88: 24-51, 2010. 3", + "[37] Chloe LeGendre, Wan-Chun Ma, Graham Fyffe, John Flynn, Laurent Charbonnel, Jay Busch, and Paul Debevec. Deeplight: Learning illumination for unconstrained mobile mixed reality. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5918-5928, 2019. 3", + "[38] Peixuan Li, Huaici Zhao, Pengfei Liu, and Feidao Cao. Rtm3d: Real-time monocular 3d detection from object keypoints for autonomous driving. In European Conference on Computer Vision, pages 644-660. Springer, 2020. 1", + "[39] Xiaofan Li, Yifu Zhang, and Xiaoqing Ye. Drivingdiffusion: Layout-guided multi-view driving scene video generation with latent diffusion model. arXiv preprint arXiv:2310.07771, 2023. 2, 3", + "[40] Zhengqin Li, Zexiang Xu, Ravi Ramamoorthi, Kalyan Sunkavalli, and Manmohan Chandraker. Learning to reconstruct shape and spatially-varying reflectance from a single image. ACM Transactions on Graphics (TOG), 37(6):1-11, 2018. 3", + "[41] Zhuopeng Li, Lu Li, and Jianke Zhu. Read: Large-scale neural scene rendering for autonomous driving. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1522-1529, 2023. 2, 3", + "[42] Zhijian Liu, Haotian Tang, Alexander Amini, Xinyu Yang, Huizi Mao, Daniela L Rus, and Song Han. Bevfusion: Multi-task multi-sensor fusion with unified bird's-eye view represent" + ], + "bbox": [ + 78, + 90, + 470, + 901 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "tation. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 2774-2781. IEEE, 2023. 4", + "[43] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 2", + "[44] Milan Miric, Kevin J Boudreau, and Lars Bo Jeppesen. Protecting their digital assets: The use of formal & informal appropriability strategies by app developers. Research Policy, 48(8):103738, 2019. 1", + "[45] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM Transactions on Graphics (ToG), 41(4):1-15, 2022. 2", + "[46] OpenAI. Gpt-4 technical report. 2023. 3", + "[47] Julian Ost, Fahim Mannan, Nils Thuerey, Julian Knodt, and Felix Heide. Neural scene graphs for dynamic scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2856-2865, 2021. 2, 3", + "[48] Julian Ost, Issam Laradji, Alejandro Newell, Yuval Bahat, and Felix Heide. Neural point light fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18419-18429, 2022. 2, 3", + "[49] Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, et al. Training language models to follow instructions with human feedback. Advances in Neural Information Processing Systems, 35:27730-27744, 2022. 3", + "[50] Jonah Philion and Sanja Fidler. Lift, splat, shoot: Encoding images from arbitrary camera rigs by implicitly unprojecting to 3d. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XIV 16, pages 194-210. Springer, 2020. 7", + "[51] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models, 2021. 4", + "[52] Guodong Rong, Byung Hyun Shin, Hadi Tabatabaee, Qiang Lu, Steve Lemke, Martins Mozeiko, Eric Boise, Geehoon Uhm, Mark Gerow, Shalin Mehta, et al. Lgsvl simulator: A high fidelity simulator for autonomous driving. In 2020 IEEE 23rd International conference on intelligent transportation systems (ITSC), pages 1-6. IEEE, 2020. 1", + "[53] Soumyadip Sengupta, Jinwei Gu, Kihwan Kim, Guilin Liu, David W Jacobs, and Jan Kautz. Neural inverse rendering of an indoor scene from a single image. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8598-8607, 2019. 3", + "[54] Shital Shah, Debadeepta Dey, Chris Lovett, and Ashish Kapoor. Airsim: High-fidelity visual and physical simulation for autonomous vehicles. In Field and Service Robotics: Results of the 11th International Conference, pages 621-635. Springer, 2018. 2, 3", + "[55] Gowri Somanath and Daniel Kurz. Hdr environment map estimation for real-time augmented reality. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11298-11306, 2021. 3" + ], + "bbox": [ + 501, + 92, + 893, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "15086", + "bbox": [ + 480, + 945, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[56] Cheng Sun, Min Sun, and Hwann-Tzong Chen. Direct voxel grid optimization: Super-fast convergence for radiance fields reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5459-5469, 2022. 2, 8", + "[57] Pei Sun, Henrik Kretzschmar, Xerxes Dotiwalla, Aurelien Chouard, Vijaysai Patnaik, Paul Tsui, James Guo, Yin Zhou, Yuning Chai, Benjamin Caine, Vijay Vasudevan, Wei Han, Jiquan Ngiam, Hang Zhao, Aleksei Timofeev, Scott Ettinger, Maxim Krivokon, Amy Gao, Aditya Joshi, Yu Zhang, Jonathon Shlens, Zhifeng Chen, and Dragomir Anguelov. Scalability in perception for autonomous driving: Waymo open dataset. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 6", + "[58] Pei Sun, Henrik Kretzschmar, Xerxes Dotiwalla, Aurelien Chouard, Vijaysai Patnaik, Paul Tsui, James Guo, Yin Zhou, Yuning Chai, Benjamin Caine, et al. Scalability in perception for autonomous driving: Waymo open dataset. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2446-2454, 2020. 1, 7", + "[59] Alexander Swerdlow, Runsheng Xu, and Bolei Zhou. Street-view image generation from a bird's-eye view layout. arXiv preprint arXiv:2301.04634, 2023. 2, 3", + "[60] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288, 2023. 3", + "[61] Haithem Turki, Jason Y Zhang, Francesco Ferroni, and Deva Ramanan. Suds: Scalable urban dynamic scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12375-12385, 2023. 2", + "[62] Peng Wang, Yuan Liu, Zhaoxi Chen, Lingjie Liu, Ziwei Liu, Taku Komura, Christian Theobalt, and Wenping Wang. F2-nerf: Fast neural radiance field training with free camera trajectories. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4150–4159, 2023. 2, 5, 8", + "[63] Xiaofeng Wang, Zheng Zhu, Guan Huang, Xinze Chen, and Jiwen Lu. Drivedreamer: Towards real-world-driven world models for autonomous driving. arXiv preprint arXiv:2309.09777, 2023. 2, 3", + "[64] Zian Wang, Wenzheng Chen, David Acuna, Jan Kautz, and Sanja Fidler. Neural light field estimation for street scenes with differentiable virtual object insertion. In European Conference on Computer Vision, pages 380-397. Springer, 2022. 3, 6, 8", + "[65] Zhenhailong Wang, Shaoguang Mao, Wenshan Wu, Tao Ge, Furu Wei, and Heng Ji. Unleashing cognitive synergy in large language models: A task-solving agent through multi-personal self-collaboration. arXiv preprint arXiv:2307.05300, 2023. 3", + "[66] Bichen Wu, Forrest Iandola, Peter H Jin, and Kurt Keutzer. Squeezedet: Unified, small, low power fully convolutional neural networks for real-time object detection for autonomous driving. In Proceedings of the IEEE conference on computer vision and pattern recognition workshops, pages 129-137, 2017. 1" + ], + "bbox": [ + 78, + 90, + 470, + 898 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[67] Qingyun Wu, Gagan Bansal, Jieyu Zhang, Yiran Wu, Shaokun Zhang, Erkang Zhu, Beibin Li, Li Jiang, Xiaoyun Zhang, and Chi Wang. Autogen: Enabling next-gen llm applications via multi-agent conversation framework. arXiv preprint arXiv:2308.08155, 2023. 3", + "[68] Zirui Wu, Tianyu Liu, Liyi Luo, Zhide Zhong, Jianteng Chen, Hongmin Xiao, Chao Hou, Haozhe Lou, Yuantao Chen, Runyi Yang, Yuxin Huang, Xiaoyu Ye, Zike Yan, Yongliang Shi, Yiyi Liao, and Hao Zhao. Mars: An instance-aware, modular and realistic simulator for autonomous driving. CICAI, 2023. 2, 3", + "[69] Ziyang Xie, Junge Zhang, Wenye Li, Feihu Zhang, and Li Zhang. S-nerf: Neural radiance fields for street views. arXiv preprint arXiv:2303.00749, 2023. 2, 3, 8", + "[70] Yinda Xu and Lidong Yu. Drl-based trajectory tracking for motion-related modules in autonomous driving. arXiv preprint arXiv:2308.15991, 2023. 4", + "[71] Kairui Yang, Enhui Ma, Jibin Peng, Qing Guo, Di Lin, and Kaicheng Yu. Bevcontrol: Accurately controlling street-view elements with multi-perspective consistency via bev sketch layout. arXiv preprint arXiv:2308.01661, 2023. 2, 3", + "[72] Zhenpei Yang, Yuning Chai, Dragomir Anguelov, Yin Zhou, Pei Sun, Dumitru Erhan, Sean Rafferty, and Henrik Kretzschmar. Surfelgan: Synthesizing realistic sensor data for autonomous driving. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11118-11127, 2020. 1", + "[73] Ze Yang, Yun Chen, Jingkang Wang, Sivabalan Manivasagam, Wei-Chiu Ma, Anqi Joyce Yang, and Raquel Urtasun. Unisim: A neural closed-loop sensor simulator. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1389-1399, 2023. 2, 3", + "[74] Jinsong Zhang, Kalyan Sunkavalli, Yannick Hold-Geoffroy, Sunil Hadap, Jonathan Eisenman, and Jean-François Lalonde. All-weather deep outdoor lighting estimation. In Proceedings of the IEEE/CVF conference on Computer Vision and Pattern Recognition, pages 10158–10166, 2019. 3", + "[75] Yichao Zhou, Jingwei Huang, Xili Dai, Linjie Luo, Zhili Chen, and Yi Ma. HoliCity: A city-scale data platform for learning holistic 3D structures. 2020. arXiv:2008.03286 [cs.CV]. 6, 8", + "[76] Mingchen Zhuge, Haozhe Liu, Francesco Faccio, Dylan R Ashley, Robert Csordás, Anand Gopalakrishnan, Abdullah Hamdi, Hasan Abed Al Kader Hammoud, Vincent Herrmann, Kazuki Irie, et al. Mindstorms in natural language-based societies of mind. arXiv preprint arXiv:2305.17066, 2023. 3" + ], + "bbox": [ + 501, + 92, + 893, + 739 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "15087", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2024/Editable Scene Simulation for Autonomous Driving via Collaborative LLM-Agents/49eb43d9-92f1-4849-a7c5-3da61380ffc6_model.json b/2024/Editable Scene Simulation for Autonomous Driving via Collaborative LLM-Agents/49eb43d9-92f1-4849-a7c5-3da61380ffc6_model.json new file mode 100644 index 0000000000000000000000000000000000000000..f41495e4085ef8163ac32dccb1e1d4a6f1aae464 --- /dev/null +++ b/2024/Editable Scene Simulation for Autonomous Driving via Collaborative LLM-Agents/49eb43d9-92f1-4849-a7c5-3da61380ffc6_model.json @@ -0,0 +1,2818 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.079, + 0.131, + 0.892, + 0.152 + ], + "angle": 0, + "content": "Editable Scene Simulation for Autonomous Driving via Collaborative LLM-Agents" + }, + { + "type": "text", + "bbox": [ + 0.203, + 0.179, + 0.768, + 0.234 + ], + "angle": 0, + "content": "Yuxi Wei\\(^{1,5*}\\) Zi Wang\\(^{3,5*}\\) Yifan Lu\\(^{1,5*}\\) Chenxin Xu\\(^{1,5*}\\) Changxing Liu\\(^{1,5}\\) Hao Zhao\\(^{4}\\) Siheng Chen\\(^{1,2,5\\dagger}\\) Yanfeng Wang\\(^{1,2}\\) Shanghai Jiao Tong University" + }, + { + "type": "text", + "bbox": [ + 0.14, + 0.236, + 0.828, + 0.305 + ], + "angle": 0, + "content": "\\(^{3}\\) Carnegie Mellon University \\(^{4}\\) Tsinghua University \\(^{5}\\) Multi-Agent Governance & Intelligence Crew (MAGIC) {wyx3590236732, yifan_lu, xcxwakaka, cx-liu}@sjtu.edu.cn {sihengc, wangyanfeng}@sjtu.edu.cn ziwang2@andrew.cmu.edu zhaohao@air.tsinghua.edu.cn" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.34, + 0.314, + 0.356 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.371, + 0.474, + 0.689 + ], + "angle": 0, + "content": "Scene simulation in autonomous driving has gained significant attention because of its huge potential for generating customized data. However, existing editable scene simulation approaches face limitations in terms of user interaction efficiency, multi-camera photo-realistic rendering and external digital assets integration. To address these challenges, this paper introduces ChatSim, the first system that enables editable photo-realistic 3D driving scene simulations via natural language commands with external digital assets. To enable editing with high command flexibility, ChatSim leverages a large language model (LLM) agent collaboration framework. To generate photo-realistic outcomes, ChatSim employs a novel multi-camera neural radiance field method. Furthermore, to unleash the potential of extensive high-quality digital assets, ChatSim employs a novel multi-camera lighting estimation method to achieve scene-consistent assets' rendering. Our experiments on Waymo Open Dataset demonstrate that ChatSim can handle complex language commands and generate corresponding photo-realistic scene videos. Code can be accessed at: https://github.com/yifanlu0227/ChatSim." + }, + { + "type": "title", + "bbox": [ + 0.079, + 0.718, + 0.21, + 0.733 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.742, + 0.472, + 0.879 + ], + "angle": 0, + "content": "Perception [12, 15-17, 38, 66] is the window of an autonomous vehicle into the external environment. To ensure the robustness of the vehicle's perceptual capabilities during both training and testing phases, it necessitates the collection of high-quality perception data in substantial volumes [13, 58]. However, the operation of a fleet for the acquisition of real-world data often incurs prohibitive expenses, particularly for specialized or customized requirements. For instance, in the aftermath of an accident or intervention in" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.34, + 0.875, + 0.527 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.5, + 0.528, + 0.892, + 0.556 + ], + "angle": 0, + "content": "Figure 1. ChatSim enables the editing of photo-realistic 3D driving scene simulations via language commands." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.565, + 0.895, + 0.686 + ], + "angle": 0, + "content": "volving an autonomous vehicle, it is imperative to test the vehicle's perception system across a spectrum of similar scenarios. While replicating such scenario data from real-world instances is nearly impossible due to the uncontrollability of actual scenes [9, 52], customized scene simulation emerges as a vital and feasible alternative. It enables the precise modeling of specific conditions without high costs and logistical complexities of real-world data collection [4, 72]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.688, + 0.895, + 0.869 + ], + "angle": 0, + "content": "To effectively simulate customized driving scenes, we identify three key properties as fundamental. First, the simulation should be capable of following sophisticated or abstract demands, thereby facilitating the production. Second, the simulation should generate photo-realistic, view-consistent outcomes, which allow for the closest approximation to vehicle observations in real-world scenarios. Third, it should allow for the integration of external digital assets [6, 44] with their photo-realistic textures and materials while fitting the lighting conditions. This capability would unlock the potential for data expansion by incorporating a wide array of external digital assets, satisfying customized needs." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.894, + 0.901 + ], + "angle": 0, + "content": "A vast array of significant works have been proposed for scene simulation, yet they fail to meet all three of" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.887, + 0.207, + 0.901 + ], + "angle": 0, + "content": "*Equal contribution." + }, + { + "type": "page_footnote", + "bbox": [ + 0.234, + 0.888, + 0.358, + 0.901 + ], + "angle": 0, + "content": "† Corresponding author." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15077" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.473, + 0.394 + ], + "angle": 0, + "content": "these requirements. Traditional graphics engines, such as CARLA [22] and UE [25], offer editable virtual environments with external digital assets, but the data realism is restricted by asset modeling and rendering qualities. Image generation based methods, such as BEVControl [71], DriveDreamer [63], MagicDrive [26], can generate realistic scene images based on various control signals, including BEV maps, bounding boxes and camera poses. However, they struggle to maintain view consistency and face challenges in importing external digital assets due to the absence of 3D spatial modeling. Rendering-based methods have been proposed to obtain photo-realistic and view-consistent scene simulation. Notable examples like UniSim [73] and MARS [68] come equipped with a suite of scene-editing tools. However, these systems require extensive user involvement in every trivial editing step via code implementation, which is ineffective when performing the editing. Furthermore, while they handle vehicles in observed scenarios effectively, their inability to support external digital assets restricts opportunities for data expansion and customization." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.397, + 0.473, + 0.502 + ], + "angle": 0, + "content": "To fulfill the identified requirements, we introduce ChatSim, the first system that enables editable photo-realistic 3D driving scene simulations via natural language commands with external digital assets. To use ChatSim, users simply engage in a conversation with the system, issuing commands through natural language without any involvement in intermediate simulation steps; see Figure 1 for illustration." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.506, + 0.473, + 0.734 + ], + "angle": 0, + "content": "To address complex or abstract user commands effectively, ChatSim adopts a large language model (LLM)-based multi-agent collaboration framework. The key idea is to exploit multiple LLM agents, each with a specialized role, to decouple an overall simulation demand into specific editing tasks, thereby mirroring the task division and execution typically founded in the workflow of a human-operated company. This workflow offers two key advantages for scene simulation. First, LLM agents' ability to process human language commands allows for intuitive and dynamic editing of complex driving scenes, enabling precise adjustments and feedback. Second, the collaboration framework enhances simulation efficiency and accuracy by distributing specific editing tasks among specialized agents, ensuring detailed and realistic simulations with improved task completion rates." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.735, + 0.473, + 0.903 + ], + "angle": 0, + "content": "To generate photo-realistic outcomes, we propose McNeRF in ChatSim, a novel neural radiance field method that incorporates multi-camera inputs, offering a broader scene rendering. This integration fully exploits camera setups on vehicles but raises two significant challenges: camera pose misalignment due to asynchronizated trigger times and brightness inconsistency due to different camera exposure times. To address camera pose misalignment, McNeRF uses a multi-camera alignment to reduce extrinsic parameter noises, ensuring rendering quality. To address brightness inconsistency, McNeRF integrates the critical exposure times" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.137 + ], + "angle": 0, + "content": "to recover scene radiance in HDR, markedly mitigating the issue of color discrepancies at the intersections of two camera images with different exposure times." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.138, + 0.897, + 0.289 + ], + "angle": 0, + "content": "To import external digital assets with their realistic textures and materials, we propose McLight, a novel multicamera lighting estimation that blends skydome and surrounding lighting. Our skydome estimation restores accurate sun behavior with peak intensity residual connection, enabling the rendering of prominent shadows. For surrounding lighting, McLight queries McNeRF to achieve complex location-specific illumination effects, like those in the tree shade with sunlight being blocked. This significantly improves the rendering realism of the integrated 3D assets." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.29, + 0.897, + 0.456 + ], + "angle": 0, + "content": "We conduct extensive experiments on the Waymo autonomous driving dataset and show that ChatSim generates photo-realistic customized perception data including dangerous corner cases according to various human language commands. Our method is compatible with mixed, highly abstract and multi-round commands. Our method achieves SoTA performance with an improvement of \\(4.5\\%\\) in photorealism with a wide-angle rendering. Moreover, we demonstrate our lighting estimation outperforms the SoTA methods both qualitatively and quantitatively, reducing the intensity error and angular error by \\(57.0\\%\\) and \\(9.9\\%\\)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.471, + 0.642, + 0.487 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.494, + 0.897, + 0.903 + ], + "angle": 0, + "content": "Scene simulation for autonomous driving. Current scene simulation methods can be generally divided into three categories: graphics engines, image generation, and scene rendering. Graphics engines, such as CARLA [22], AirSim [54], OpenScenario Editor [24], 51Sim-One [1] and RoadRunner [21], create a virtual world for simulating a wide range of driving scenarios. However, there exists a significant domain gap between the virtual world and reality. Image generation methods can generate realistic scene images based on different control signals, such as HD maps [26, 39, 59], sketch layout [71], bounding boxes [26, 39, 63], text [26, 33, 39, 63] and driving actions [33, 63]. However, these approaches can hardly maintain scene consistency. To obtain a coherent driving scene, methods based on scene rendering target to reconstruct the 3D scene. READ [41] employs point clouds and uses a U-Net to render images. With the rapid development of Neural Radiance Field (NeRF) [7, 8, 43, 45, 56, 62], several works [28, 34, 47, 48, 61, 68, 69, 73] also exploit NeRFs to model cars and static street backgrounds in outdoor environments. Moreover, notable examples like UniSim [73] and MARS [68] come equipped with a suite of scene-editing tools. However, these methods require extensive user involvement in intermediate editing steps and they fail to support external digital assets for data expansion. In this work, we propose ChatSim that achieves automatic simulation editing via language commands and integrates external digital assets to enhance realism and flexibility. In ChatSim," + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "15078" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.081, + 0.089, + 0.467, + 0.294 + ], + "angle": 0, + "content": "
MethodPhoto-realisticDim.Multi-cameraEditableExternal assetsLanguageOpen-source
CARLA [22]×3D✓✓✓×✓
AirSim [54]×3D✓✓✓×✓
OpenScenario [24]×3D✓✓✓×✓
51Sim-One [1]×3D✓✓✓××
RoadRunner [21]×3D✓✓✓×✓
BEVGen [59]✓2D✓✓××✓
BEVCentral [71]✓2D✓✓×××
DriveDreamer [63]✓2D✓✓×✓×
DrivingDiffusion [39]✓2D✓✓×✓×
GAIA-1 [33]✓2D×✓×✓×
MagicDrive [26]✓2D✓✓×××
READ [41]✓3D××××✓
Neural SG [47]✓3D×✓××✓
Neural PLF [48]✓3D××××✓
S-NeRF [69]✓3D✓×××✓
UniSim [73]✓3D×✓×××
MARS [68]✓3D×✓××✓
ChatSim (Ours)✓3D✓✓✓✓✓
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.296, + 0.472, + 0.325 + ], + "angle": 0, + "content": "Table 1. Comparison of existing and proposed methods for autonomous driving simulation." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.331, + 0.47, + 0.361 + ], + "angle": 0, + "content": "we integrate McNeRF, a novel neural radiance field designed to leverage multi-camera inputs for high-fidelity rendering." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.362, + 0.472, + 0.574 + ], + "angle": 0, + "content": "Lighting estimation. Lighting estimation focuses on assessing the illumination conditions of a real-world environment to seamlessly integrate digital objects. Early methods [35, 36] for outdoor environments use explicit cues like detected shadows on the ground. Recent works usually adopt learning-based approaches [27, 30, 31, 37, 40, 74] by predicting different lighting representations like spherical lobes [10, 40], light probes [37], environment map [53, 55], HDR sky model [31, 64, 74] and lighting volume [64]. However, few of them consider multi-camera input, which is common for driving scenarios. In this paper, we propose a novel multi-camera lighting estimation method, McLight, combining with our McNeRF, to estimate a wider range of lighting and obtain the spatially-varying lighting effects of assets." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.575, + 0.472, + 0.817 + ], + "angle": 0, + "content": "Large Language model and collaborative framework. Large Language Models (LLMs) are AI systems trained on extensive data to understand, generate, and respond to human language. GPT [11] is a pioneering work to generate human-like content. The following updated versions GPT-3.5 [14] and GPT-4 [46], provide more intelligent capabilities like chatting, browsing and coding. Notable other large language models include InstructGPT [49], LLaMA [60] and PaLM [5, 18]. Based on LLM, many works [3, 23, 29, 65, 76] improve the problem-solving abilities by integrating communication among multiple agents. [32] and [67] define a group of well-organized agents to form operating procedures with conversation and code programming. In this paper, we exploit the power of collaborative LLM agents in simulation for autonomous driving, enabling the various editing of 3D scenes via language commands." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.833, + 0.427, + 0.851 + ], + "angle": 0, + "content": "3. Collaborative LLM-Agents for Editing" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.856, + 0.472, + 0.903 + ], + "angle": 0, + "content": "The ChatSim system analyzes specific user commands and returns a video that meets customized needs; see Figure 2. Since user commands could be abstract and sophisticated," + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.168 + ], + "angle": 0, + "content": "it requires the system to have flexible task-handling ability. Directly applying a single LLM agent struggles with multi-step reasoning and cross-referencing. To address this, we design a series of collaborative LLM agents, where each agent is responsible for a unique aspect of the editing task." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.18, + 0.766, + 0.197 + ], + "angle": 0, + "content": "3.1. Specific Agent's Functionality" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.201, + 0.895, + 0.413 + ], + "angle": 0, + "content": "Agents in ChatSim comprise two key components: a Large Language Model (LLM) and the corresponding role functions. The LLM is responsible for understanding the received commands while the role functions process the received data. Each agent is equipped with unique LLM prompts and role functions tailored to their specific duties within the system. To accomplish their tasks, agents first convert the received commands to a structured configuration using LLM with the assistance of prompts. Then the role functions utilize the structured configuration as parameters to process the received data and produce the desired outcomes; see an agent example on the right side of Figure 2. This workflow endows agents with both language interpretation capabilities and precise execution capabilities." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.414, + 0.895, + 0.581 + ], + "angle": 0, + "content": "Project Manager Agent. The project manager agent decomposes direct commands into clear natural language instructions dispatched to other editing agents. To equip the project manager agent with the capability of command decomposition, we design a series of prompts for its LLM. The core idea of the prompts is to describe the action set, give the overall goal, and define the output form with examples; The role functions send the decomposed instructions to other agents for editing. The presence of the project manager agent enhances the system's robustness in interpreting various inputs and streamlines operations for clarity and fine granularity." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.582, + 0.895, + 0.703 + ], + "angle": 0, + "content": "Tech agent for view adjustment. The view adjustment agent generates suitable extrinsic camera parameters. The LLM in the agent translates the natural language instructions for viewpoint adjustment into movement parameters to the target viewpoint's position and angle. In role functions, these movement parameters are turned into transformation matrices required by the extrinsic, which are then multiplied by the original parameters to yield a new viewpoint." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.704, + 0.895, + 0.839 + ], + "angle": 0, + "content": "Tech agent for background rendering. The background rendering agent renders the scene background based on multi-camera images. The LLM receives the rendering command and then operates the role functions for rendering. Notably, in role functions, we specifically integrate a novel neural radiance field method (McNeRF) taking multi-camera inputs and considering exposure time, solving the problem of blurring and brightness inconsistency in multi-camera rendering, see more details in Section 4.1." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.895, + 0.901 + ], + "angle": 0, + "content": "Tech agent for vehicle deleting. The vehicle deleting agent removes specified vehicles from the background. It first identifies current vehicle attributes like 3D bounding boxes and colors from given scene information or results from a scene" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "15079" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.1, + 0.092, + 0.868, + 0.324 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.327, + 0.892, + 0.355 + ], + "angle": 0, + "content": "Figure 2. ChatSim system overview. The system exploit multiple collaborative LLM agents with specialized roles to decouple an overall demand into specific editing tasks. Each agent equips an LLM and corresponding role functions to interpret and execute its specific tasks." + }, + { + "type": "image", + "bbox": [ + 0.088, + 0.365, + 0.456, + 0.485 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.116, + 0.486, + 0.429, + 0.5 + ], + "angle": 0, + "content": "Figure 3. Prompt example of view adjustment agent." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.505, + 0.471, + 0.58 + ], + "angle": 0, + "content": "perception model like [42]. The LLM gathers attributes of the vehicles and performs matching with user requests. Upon confirming the targeted vehicles, it employs a per-frame inpainting model as the role functions, such as latent diffusion methods [51], to effectively delete them from the scene." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.583, + 0.471, + 0.718 + ], + "angle": 0, + "content": "Tech agent for 3D asset management. The 3D asset management agent selects and modifies 3D digital assets according to user specifications. It constructs and maintains a 3D digital asset bank; see our bank details in the Appendix. To facilitate the addition of various objects, the agent first uses LLM to select the most suitable asset by key attributes matching with the requirements, such as color and type. If the matching is not perfect, the agent could modify the asset through its role functions like changing the color." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.72, + 0.473, + 0.902 + ], + "angle": 0, + "content": "Tech agent for vehicle motion. The vehicle motion agent creates the initial places and subsequent motions of vehicles following the requests. Existing vehicle motion generation methods cannot directly generate motion purely from text and the scene map. To solve the problem, here we propose a novel text-to-motion method. The key idea is linking a placement and planning module as role functions with LLMs to extract and turn motion attributes into coordinates. Motion attributes include position attributes (e.g., distance, direction) and movement attributes (e.g., speed, action). For the placement module, we endow each lane node in the lane map with its attributes to match with the position attributes. The" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.367, + 0.895, + 0.427 + ], + "angle": 0, + "content": "planning module plans the vehicle's approximate destination lane node and then plans the intermediate trajectory by fitting the Bezier curves. We also add trajectory tracking [70] to fit vehicle dynamics; see more details in the Appendix." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.428, + 0.893, + 0.563 + ], + "angle": 0, + "content": "Tech agent for foreground rendering. The foreground rendering agent integrates camera extrinsic information, 3D assets, and motion information to render foreground objects in the scene. Notably, to seamlessly integrate the external assets with the current scene, we design a multi-camera lighting estimation method (McLight) into the role functions, coupling with McNeRF. The estimated illumination is then utilized by Blender API to generate foreground images. The detailed technical aspects will be elaborated in Section 4.2." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.573, + 0.774, + 0.589 + ], + "angle": 0, + "content": "3.2. Agent Collaboration Workflow" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.594, + 0.895, + 0.835 + ], + "angle": 0, + "content": "Agents with tailored functions collaboratively work together to edit based on user commands. The project manager orchestrates and dispatches instructions to editing agents. The editing agents form two teams: background generation and foreground generation. For background generation, the background rendering agent generates rendered images using the extrinsic parameters from the view adjustment agent, followed by inpainting by the vehicle deleting agent. For foreground generation, the foreground rendering agent renders the images using the extrinsic parameters from the view adjustment agent, selected 3D assets from 3D asset management agent, and generated motions from vehicle motion agent. Finally, the foreground and background images are composed to create and deliver a video to the user. The editing information in each agent's configuration is recorded by the project manager agent for possible multi-round editorings." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.849, + 0.744, + 0.866 + ], + "angle": 0, + "content": "4. Novel Rendering Methods" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.895, + 0.902 + ], + "angle": 0, + "content": "Based on the collaborative LLM agents framework introduced in Section 3, this section presents two novel rendering" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15080" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.244 + ], + "angle": 0, + "content": "techniques to enhance photo-realism in simulations. To tackle the rendering challenges caused by multiple cameras, we propose multi-camera neural radiance field (McNeRF), a novel NeRF model considering the varied camera exposure times for visual consistency. To render realistic external digital assets with location-specific lighting and accurate shadows, we propose McLight, a novel hybrid lighting estimation method combined with our McNeRF. Note that McNeRF and McLight are leveraged by the background rendering agent and the foreground rendering agent, respectively." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.261, + 0.396, + 0.278 + ], + "angle": 0, + "content": "4.1. McNeRF for Background Rendering" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.288, + 0.47, + 0.409 + ], + "angle": 0, + "content": "An autonomous vehicle typically equips multiple cameras to achieve a comprehensive perception view. However, this poses challenges for NeRF training due to the misaligned multi-camera poses from asynchronized camera trigger times and the brightness inconsistency originating from different exposure times. To address these challenges, the proposed McNeRF uses two techniques: multi-camera alignment and brightness-consistent rendering." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.412, + 0.47, + 0.518 + ], + "angle": 0, + "content": "Multi-camera alignment. Autonomous vehicles, despite having a localization module for accurate camera poses, face challenges with asynchronous trigger times across multiple cameras. To align camera extrinsics for NeRF training, our core idea is to leverage a consistent spatial coordinate system provided by Agisoft Metashape [2] to align the images captured by multiple cameras at different timestamps." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.52, + 0.472, + 0.596 + ], + "angle": 0, + "content": "Specifically, let \\(\\mathcal{I}^{(i,k)}\\) and \\(\\xi^{(i,k)}\\) be the image captured by the \\(i\\)th camera at the \\(k\\)th trigger and the corresponding camera pose in the vehicle's global coordinate space, respectively. We first input all images into Metashape for recalibration. The aligned camera pose is then obtained as:" + }, + { + "type": "equation", + "bbox": [ + 0.192, + 0.597, + 0.352, + 0.617 + ], + "angle": 0, + "content": "\\[\n\\widehat {\\xi} ^ {(i, k)} = T _ {M \\to G} \\cdot \\xi_ {M} ^ {(i, k)},\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.62, + 0.47, + 0.759 + ], + "angle": 0, + "content": "where \\(\\xi_{M}^{(i,k)}\\) denotes the recalibrated camera pose in the Metashape's unified spatial coordinate space, and \\(T_{M\\rightarrow G}\\) is the transformation from the Metashape's coordinate space to the vehicle's global coordinate space. After alignment, the pose noise can be significantly reduced. Then, the aligned camera pose \\(\\widehat{\\xi}^{(i,t)}\\) can be used to generate the origins and directions of rays for McNeRF, enabling high-fidelity rendering. The aligned pose can also facilitate the foreground rendering agent's operations." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.762, + 0.471, + 0.852 + ], + "angle": 0, + "content": "Brightness-consistent rendering. The exposure times of cameras can differ substantially, causing significant brightness differences across images, hindering the NeRF training. As shown in Figure 4, McNeRF, addresses this by incorporating exposure times into HDR radiance fields, prompting brightness consistency." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.855, + 0.471, + 0.902 + ], + "angle": 0, + "content": "We adopt F2-NeRF [62] as our backbone model to handle the unbounded scene, sampling \\(K\\) points along the ray \\(\\mathbf{r}\\) and estimating each point's HDR radiance \\(\\mathbf{e}_k\\) and density \\(\\sigma_k\\)." + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.091, + 0.885, + 0.182 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.506, + 0.183, + 0.885, + 0.282 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.284, + 0.893, + 0.354 + ], + "angle": 0, + "content": "Figure 4. Rendering framework. The main components include McNeRF and McLight. Background rendering uses McNeRF to predict HDR pixel value and convert it to LDR with sRGB OETF. McLight includes a skydome lighting estimation network and adopts McNeRF to generate surrounding lighting." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.36, + 0.803, + 0.374 + ], + "angle": 0, + "content": "The HDR light intensity is then calculated as:" + }, + { + "type": "equation", + "bbox": [ + 0.583, + 0.376, + 0.892, + 0.417 + ], + "angle": 0, + "content": "\\[\n\\widehat {\\mathcal {I}} _ {\\mathrm {H D R}} (\\mathbf {r}) = f (\\Delta t) \\cdot \\sum_ {k = 1} ^ {K} T _ {k} \\alpha_ {k} \\mathbf {e} _ {k}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.419, + 0.893, + 0.526 + ], + "angle": 0, + "content": "where \\(\\alpha_{k} = 1 - \\exp (-\\sigma_{k}\\delta_{i})\\) is the opacity, \\(\\delta_{i}\\) is the point sampling interval, \\(T_{k} = \\prod_{i = 0}^{k - 1}(1 - \\alpha_{i})\\) is the accumulated transmittance and \\(\\Delta t\\) is the exposure time. The normalization function \\(f(\\Delta t) = 1 + \\epsilon (\\Delta t - \\mu) / \\sigma\\) is designed to stabilize training, where \\(\\epsilon\\) is a hyperparameter for scaling, \\(\\mu\\) and \\(\\sigma\\) are the mean and standard deviation of the exposure times of all images, respectively." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.527, + 0.894, + 0.632 + ], + "angle": 0, + "content": "By predicting scene radiance in HDR and multiplying it by the exposure time, we recover the light intensity received by the sensor and tackle the inconsistent color supervision at the intersections of two camera images with distinct exposure times. Moreover, the HDR light intensity outputted by McNeRF can provide scene-level illumination for foreground object rendering, a topic further discussed in Section 4.2." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.632, + 0.894, + 0.692 + ], + "angle": 0, + "content": "To train the rendering network, we enforce the consistency of radiance between the rendered image (prediction) and the captured image (ground-truth). Given the ground-truth image \\(\\mathcal{I}\\), the loss function is then:" + }, + { + "type": "equation", + "bbox": [ + 0.542, + 0.694, + 0.848, + 0.73 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\frac {1}{| R |} \\sum_ {\\mathbf {r} \\in R} \\left(\\mathrm {O E T F} \\left(\\widehat {\\mathcal {I}} _ {\\mathrm {H D R}} (\\mathbf {r})\\right) - \\mathcal {I} (\\mathbf {r})\\right) ^ {2},\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.731, + 0.893, + 0.777 + ], + "angle": 0, + "content": "where \\(R\\) represents the ray set and \\(\\mathrm{OETF}(\\cdot)\\) is the sRGB opto-electronic transfer function (gamma correction) [19] that converts HDR light intensity to LDR colors." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.787, + 0.811, + 0.804 + ], + "angle": 0, + "content": "4.2. McLight for Foreground Rendering" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.81, + 0.894, + 0.901 + ], + "angle": 0, + "content": "To enrich the scene's content with substantial digital 3D assets, we employ Blender [20] foreground virtual objects' rendering. A seamless insertion critically depends on accurately estimating the scene's illumination conditions. Thus, as shown in Figure 4, we propose McLight, a novel hybrid lighting estimation consisting of skydome lighting and sur" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "15081" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.2, + 0.107 + ], + "angle": 0, + "content": "rounding lighting." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.107, + 0.471, + 0.212 + ], + "angle": 0, + "content": "Skydome lighting estimation. Estimating skydome lighting from images is challenging for restoring accurate sun behavior. To achieve this, we propose a novel residual connection from the estimated peak intensity to the HDR reconstruction to address over-smoothing output. Further, we adopt a self-attention mechanism to fuse multi-camera inputs, capturing complementary visual cues." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.213, + 0.47, + 0.439 + ], + "angle": 0, + "content": "Here we employ a two-stage process. In the first stage, we train an autoencoder to reconstruct the corresponding HDR panorama from an LDR panorama. Following [64], the encoder transforms the LDR skydome panorama into three intermediate vectors, including the peak direction vector \\(\\mathbf{f}_{\\mathrm{dir}} \\in \\mathbb{R}^3\\), the intensity vector \\(\\mathbf{f}_{\\mathrm{int}} \\in \\mathbb{R}_+^3\\), and the sky content vector \\(\\mathbf{f}_{\\mathrm{content}} \\in \\mathbb{R}^{64}\\). However, as HDR intensity behaves like an impulse response at its peak position, with pixel values thousands of times higher than its neighbors, it is difficult for the decoder to recover such patterns. To tackle this, we design a residual connection that injects \\(\\mathbf{f}_{\\mathrm{int}}\\) into the decoded HDR panorama with a spherical Gaussian lobe attenuation. This explicitly restores the peak intensity of the sun in the reconstructed HDR panorama, allowing us to render strong shadows for virtual objects." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.44, + 0.471, + 0.645 + ], + "angle": 0, + "content": "In the second stage, we train an image encoder and a multi-camera fusion module built upon the pretrained decoder from the first stage. Specifically, for images from each camera, a shared image encoder predicts the peak direction vector \\(\\mathbf{f}_{\\mathrm{dir}}^{(i)}\\), the intensity vector \\(\\mathbf{f}_{\\mathrm{int}}^{(i)}\\), and the sky content vector \\(\\mathbf{f}_{\\mathrm{content}}^{(i)}\\) for each image \\(\\mathcal{I}^{(i)}\\), where \\(i\\) is the camera index. We design the latent vector fusion across the multiple camera views as follows: all \\(\\mathbf{f}_{\\mathrm{dir}}^{(i)}\\) are aligned to the front-facing view using their extrinsic parameters and averaged to form \\(\\bar{\\mathbf{f}}_{\\mathrm{dir}}\\); all \\(\\mathbf{f}_{\\mathrm{int}}^{(i)}\\) are averaged to yield \\(\\bar{\\mathbf{f}}_{\\mathrm{int}}\\); all \\(\\mathbf{f}_{\\mathrm{content}}^{(i)}\\) are integrated into \\(\\bar{\\mathbf{f}}_{\\mathrm{content}}\\) through a self-attention module. Finally, the pretrained decoder reconstructs the HDR skydome image \\(\\mathcal{I}_{\\mathrm{skydome}}\\) from \\(\\bar{\\mathbf{f}}_{\\mathrm{dir}}\\), \\(\\bar{\\mathbf{f}}_{\\mathrm{int}}\\) and \\(\\bar{\\mathbf{f}}_{\\mathrm{content}}\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.645, + 0.471, + 0.719 + ], + "angle": 0, + "content": "Compared to alternative approaches [31, 64], our multicamera sky dome estimation technique accurately reproduces the sun's intensity response behavior at its peak with our residual designs, significantly improving the accuracy and fidelity of the skydome reconstruction." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.72, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Surrounding lighting estimation. Merely modeling the sky-dome cannot replicate the complex location-specific lighting effects, like those in the shade with sunlight blocked by trees or buildings. Our McNeRF is capable of storing precise 3D scene information, enabling us to capture the surrounding scene's impact on lighting. This approach facilitates the achievement of spatially-varying lighting estimation. Specifically, we sample the hemisphere rays at the virtual object's position \\(\\mathbf{o}\\). The rays' directions, \\(\\mathbf{d}_i\\), \\(i = 0,1,\\dots,h\\times w\\), are aligned with pixel coordinates on a unit sphere using equirectangular projection from an environment map, where \\(h\\) and \\(w\\) are map's height and width. With the ray" + }, + { + "type": "image_caption", + "bbox": [ + 0.548, + 0.089, + 0.601, + 0.099 + ], + "angle": 0, + "content": "Front left" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.1, + 0.631, + 0.158 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.68, + 0.089, + 0.711, + 0.099 + ], + "angle": 0, + "content": "Front" + }, + { + "type": "image", + "bbox": [ + 0.638, + 0.099, + 0.753, + 0.158 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.786, + 0.089, + 0.845, + 0.099 + ], + "angle": 0, + "content": "Front right" + }, + { + "type": "image", + "bbox": [ + 0.76, + 0.099, + 0.874, + 0.158 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.16, + 0.874, + 0.344 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.518, + 0.344, + 0.865, + 0.389 + ], + "angle": 0, + "content": "Command: \"Remove all cars in the scene and add a Porsche driving the wrong way toward me fast. Additionally, add a police car also driving the wrong way and chasing behind the Porsche. The view should be moved 5 meters ahead and 0.5 meters above.\"" + }, + { + "type": "image_caption", + "bbox": [ + 0.51, + 0.39, + 0.882, + 0.404 + ], + "angle": 0, + "content": "Figure 5. Editing result under a complex and mixed command." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.413, + 0.895, + 0.488 + ], + "angle": 0, + "content": "\\(\\mathbf{r} = \\mathbf{o} + t\\mathbf{d}_i\\), we query our McNeRF as Equation 1 to obtain HDR surrounding lighting \\(\\mathcal{I}_{\\mathrm{surround}}(\\mathbf{o},\\mathbf{d}_i)\\). The surrounding lighting estimation reconstructs complex environmental lighting, achieving a spatially varying effect and high consistency with the background." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.489, + 0.895, + 0.61 + ], + "angle": 0, + "content": "Blending. We blend the HDR intensity value from the sky-dome and surrounding lighting by transmittance of the final sampling point from McNeRF. The idea is that the rays emitted outside the radiance fields will definitely hit the sky-dome. Given the direction \\(\\mathbf{d}_i\\), we retrieve the skydome's intensity \\(\\mathcal{I}_{\\mathrm{skydome}}(\\mathbf{d}_i)\\) with equirectangular projection. The final HDR light intensity \\(\\mathcal{I}_{\\mathrm{env}}(\\mathbf{o},\\mathbf{d}_i)\\) is a combination of scene and skydome:" + }, + { + "type": "equation", + "bbox": [ + 0.528, + 0.612, + 0.862, + 0.629 + ], + "angle": 0, + "content": "\\[\n\\mathcal {I} _ {\\text {e n v}} (\\mathbf {o}, \\mathbf {d} _ {i}) = \\mathcal {I} _ {\\text {s u r r o u n d}} (\\mathbf {o}, \\mathbf {d} _ {i}) + T _ {K} \\mathcal {I} _ {\\text {s k y d o m e}} (\\mathbf {d} _ {i}),\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.63, + 0.844, + 0.645 + ], + "angle": 0, + "content": "where \\( T_{K} \\) is the last sampling point's transmittance." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.646, + 0.894, + 0.721 + ], + "angle": 0, + "content": "McLight offers two main advantages: i) it explicitly recovers the illuminance behavior at the peak and use complementary information from multiple cameras to restore accurate skydome; and ii) it enables location-specific lighting with consideration of complex scene structures." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.735, + 0.706, + 0.752 + ], + "angle": 0, + "content": "5. Experimental Results" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.76, + 0.815, + 0.776 + ], + "angle": 0, + "content": "5.1. Datasets and implementation details" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.78, + 0.894, + 0.885 + ], + "angle": 0, + "content": "We demonstrate a variety of results mainly on the Waymo Open Dataset [57], which contains high-quality multicamera images and the corresponding calibrations. For McLight skydome estimation, we collect 449 HDRIs from online HDRI databases for the autoencoder training and use HoliCity [75], a street view panorama dataset for the second stage; see more dataset details in the Appendix." + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.886, + 0.892, + 0.901 + ], + "angle": 0, + "content": "In our experiment, we use front, left front, and right front" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15082" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.126, + 0.089, + 0.176, + 0.098 + ], + "angle": 0, + "content": "Front left" + }, + { + "type": "image", + "bbox": [ + 0.094, + 0.098, + 0.209, + 0.158 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.258, + 0.089, + 0.288, + 0.098 + ], + "angle": 0, + "content": "Front" + }, + { + "type": "image", + "bbox": [ + 0.214, + 0.098, + 0.33, + 0.158 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.364, + 0.089, + 0.422, + 0.098 + ], + "angle": 0, + "content": "Front right" + }, + { + "type": "image", + "bbox": [ + 0.335, + 0.098, + 0.451, + 0.158 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.094, + 0.158, + 0.451, + 0.217 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.094, + 0.218, + 0.451, + 0.279 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.094, + 0.281, + 0.451, + 0.342 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.098, + 0.342, + 0.284, + 0.354 + ], + "angle": 0, + "content": "Command: \"Create a traffic jam.\"" + }, + { + "type": "image_caption", + "bbox": [ + 0.1, + 0.355, + 0.445, + 0.369 + ], + "angle": 0, + "content": "Figure 6. Editing result under a highly abstract command." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.382, + 0.47, + 0.489 + ], + "angle": 0, + "content": "cameras in each frame. During the rendering process, we choose 40 frames per scene at a \\(10\\mathrm{Hz}\\) sampling rate, totaling 120 images. We evenly select \\(1/8\\) of these as the test set, with the remainder used for training. The input images are used at the dataset's initial resolution of \\(1920 \\times 1280\\); we employ GPT-4 as the LLMs in all of our experiments; see more implementation details in the Appendix." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.499, + 0.225, + 0.515 + ], + "angle": 0, + "content": "5.2. System results" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.522, + 0.47, + 0.582 + ], + "angle": 0, + "content": "Editing via language commands. We select three representative commands to demonstrate the editing results. All of the results demonstrate we achieve photo-realistic wide angle results, thanks to McNeRF and McLight." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.583, + 0.47, + 0.704 + ], + "angle": 0, + "content": "Mixed and complex command. We send the system with a mixed and complex command, implying that a police car is chasing a wrong-way racer. The target scene, command and the result are shown in Figure 5. We see that i) every requirement in the complex command is accurately executed thanks to our multi-agent collaboration design; ii) this command successfully simulates one rare but dangerous driving condition, which is significant in accident testing." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.705, + 0.47, + 0.795 + ], + "angle": 0, + "content": "Highly abstract command. The second type is a highly abstract command. The inputs and results are presented in Figure 6. We see that i) this highly abstract command is hard to decompose by sentence division but still can be correctly executed by our method, and ii) our 3D asset bank offers a large variety of objects for addition." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.796, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Multi-round command. We also perform a multi-round chat with our system, and the commands in different rounds exist context dependencies. The final results are shown in Figure 7. We see that i) our system is well-equipped to handle multi-round commands and execute the commands in each round precisely; ii) our system can handle the context dependencies across different rounds thanks to the recording" + }, + { + "type": "image_caption", + "bbox": [ + 0.533, + 0.09, + 0.578, + 0.1 + ], + "angle": 0, + "content": "Front left" + }, + { + "type": "image", + "bbox": [ + 0.504, + 0.101, + 0.61, + 0.156 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.679, + 0.09, + 0.706, + 0.1 + ], + "angle": 0, + "content": "Front" + }, + { + "type": "image", + "bbox": [ + 0.641, + 0.101, + 0.747, + 0.156 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.804, + 0.09, + 0.854, + 0.1 + ], + "angle": 0, + "content": "Front right" + }, + { + "type": "image", + "bbox": [ + 0.777, + 0.101, + 0.883, + 0.156 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.504, + 0.158, + 0.692, + 0.223 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.504, + 0.223, + 0.692, + 0.32 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.507, + 0.32, + 0.67, + 0.351 + ], + "angle": 0, + "content": "Command 1: \"Ego vehicle drives ahead slowly. Add a car to the close front that is moving ahead.\"" + }, + { + "type": "image", + "bbox": [ + 0.693, + 0.158, + 0.883, + 0.223 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.693, + 0.223, + 0.883, + 0.319 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.697, + 0.319, + 0.887, + 0.36 + ], + "angle": 0, + "content": "Command 2: \"Modify the added car to turn left. Add a Chevrolet to the front of the added car. Add another vehicle to the left of the added Mini driving toward me.\"" + }, + { + "type": "image_caption", + "bbox": [ + 0.534, + 0.361, + 0.857, + 0.375 + ], + "angle": 0, + "content": "Figure 7. Editing result under multi-round commands." + }, + { + "type": "image", + "bbox": [ + 0.509, + 0.388, + 0.627, + 0.473 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.635, + 0.388, + 0.752, + 0.473 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.753, + 0.388, + 0.877, + 0.473 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.5, + 0.473, + 0.892, + 0.501 + ], + "angle": 0, + "content": "Figure 8. Comparison of detection performance w/o and with our simulated data under different amounts of real data during training." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.51, + 0.741, + 0.525 + ], + "angle": 0, + "content": "ability of the project manager agent." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.525, + 0.895, + 0.737 + ], + "angle": 0, + "content": "3D detection with simulation data. We validate the benefits of our simulation as data augmentation for a downstream 3D object detection task on Waymo Open Dataset [58]. We simulate 1960 frames, derived from scenes in the training dataset. In the simulation, cars with various types, locations, and orientations are incorporated. The detection model adopts Lift-Splat [50]. Figure 8 shows detection performances with and without fixed augmentation under various amounts of real data. We see that i) a significant and consistent improvement across different data sizes is achieved; ii) when real data is limited, our simulation notably aids in rough detection (AP30); iii) when the amount of real data increases, our simulation further significantly improves fine-grained detection (AP70), reflecting the high-quality of our simulation." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.746, + 0.682, + 0.761 + ], + "angle": 0, + "content": "5.3. Component results" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.895, + 0.902 + ], + "angle": 0, + "content": "Multi-agent collaboration. We evaluated the effectiveness of the multi-agent collaboration by checking whether the command is successfully executed in Table 2. In scenarios without multi-agent collaboration, all operations are executed by a single LLM agent. We see that a single LLM agent leads to notably lower execution accuracy across all categories due to process limitations. In contrast, the collaborative multi-agent approach can manage most commands, attributed to its task division and agent role specificity." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15083" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.099, + 0.09, + 0.451, + 0.144 + ], + "angle": 0, + "content": "
Multi-agent collaborationLanguage command category
DeletionAdditionView changeRevisionAbstract
✓0.6170.3830.7170.3670.216
0.9830.8670.9670.9170.883
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.145, + 0.47, + 0.173 + ], + "angle": 0, + "content": "Table 2. The accuracy \\((\\%)\\) of task completion by LLM without and with multi-agent collaboration." + }, + { + "type": "table", + "bbox": [ + 0.1, + 0.177, + 0.445, + 0.28 + ], + "angle": 0, + "content": "
MethodsPSNR↑SSIM↑LPIPS↓Inf. time (s)↓
DVGO [56]23.570.7700.5087.7
Mip-NeRF360 [8]24.400.7540.528101.8
S-NeRF [69]24.710.7590.519114.5
F2NeRF [62]23.260.7730.4392.4
Ours w/o alignment23.320.7760.4372.5
Ours w/o exposure25.180.8190.3812.4
McNeRF (Ours)25.820.8220.3782.5
" + }, + { + "type": "table_caption", + "bbox": [ + 0.078, + 0.281, + 0.468, + 0.295 + ], + "angle": 0, + "content": "Table 3. Background novel view rendering performance evaluation." + }, + { + "type": "image", + "bbox": [ + 0.086, + 0.301, + 0.458, + 0.559 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.561, + 0.472, + 0.59 + ], + "angle": 0, + "content": "Figure 9. Comparisons of wide-angle images generation. (a) S-NeRF.(b) F2NeRF. (c) McNeRF (Ours). Last row: target images." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.599, + 0.47, + 0.734 + ], + "angle": 0, + "content": "Background rendering. We compare our McNeRF with several other state-of-the-art methods on the background novel view synthesis task. We perform reconstruction and rendering on 32 selected scenes. Table 3 shows the quantitative results comparison on three metrics: PSNR, SSIM, and LPIPS. We see that i) McNeRF achieves SoTA performance on all three metrics, significantly outperforming other baselines; ii) McNeRF has a fast inference speed, enabling quick responses to user requests for image rendering." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.735, + 0.471, + 0.84 + ], + "angle": 0, + "content": "Figure 9 demonstrates qualitative comparisons between other methods and ours. We see that existing NeRF methods do not consider the exposure time, leading to noticeable changes in brightness at the junctions of different cameras in the image, as well as an overall inconsistency in exposure across the wide-angle view. Our method can make the brightness of the entire image more consistent." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.472, + 0.902 + ], + "angle": 0, + "content": "Foreground rendering. We compare our McLight with the other two SoTA methods [31, 64]. Table 4 shows the comparison of relative intensity(log 10) error on our HDRI dataset, angular error on HoliCity [75], and user study. We see that" + }, + { + "type": "table", + "bbox": [ + 0.504, + 0.09, + 0.885, + 0.141 + ], + "angle": 0, + "content": "
MethodPeak Intensity(log10) ErrorPeak Angular Error (deg)User study(%) ↑
Mean ↓Median ↓Mean ↓Median ↓
Hold-Geoffroy et al. [31]0.8990.97548.451.619.5
Wang et al. [64]0.5900.62833.529.437.3
McLight (Ours)0.4490.27032.326.543.1
" + }, + { + "type": "table_caption", + "bbox": [ + 0.501, + 0.141, + 0.89, + 0.155 + ], + "angle": 0, + "content": "Table 4. Comparison with previous methods on lighting estimation." + }, + { + "type": "image", + "bbox": [ + 0.516, + 0.159, + 0.603, + 0.256 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.605, + 0.159, + 0.689, + 0.256 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.692, + 0.159, + 0.78, + 0.256 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.787, + 0.159, + 0.874, + 0.256 + ], + "angle": 0, + "content": null + }, + { + "type": "table_caption", + "bbox": [ + 0.501, + 0.257, + 0.89, + 0.271 + ], + "angle": 0, + "content": "Figure 10. Comparison with different lighting estimation methods." + }, + { + "type": "table", + "bbox": [ + 0.512, + 0.275, + 0.88, + 0.322 + ], + "angle": 0, + "content": "
MethodsStraightLeft TurnRight TurnSpeedWithin-road
GPT2Code0.7380.5590.5360.8930.214
GPT2Motion0.5950.1190.1670.3450.277
Ours0.9880.9400.9760.9521.000
" + }, + { + "type": "table_caption", + "bbox": [ + 0.503, + 0.322, + 0.887, + 0.336 + ], + "angle": 0, + "content": "Table 5. Comparison with motion generation from text methods." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.341, + 0.894, + 0.416 + ], + "angle": 0, + "content": "McLight achieves more accurate peak behavior and receives noticeably higher user preferences. Figure 10 shows the visualizations of vehicle insertion. The vehicles added through McLight feature significantly more realistic reflections and strong shadows consistent with the scene." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.417, + 0.893, + 0.599 + ], + "angle": 0, + "content": "Vehicle motion. As shown in Table 5, we compare the motion generation method from user commands with two of our designed baselines: 1. GPT2Motion, which directly uses LLM to return the motion coordinates; 2. GPT2Code, which first generates code using LLM and executes it to obtain the vehicle motion. We validate multiple actions in multiple scenarios and report the user study result. The user study is to determine if the generated motions matched the command intents and fitted with the lane map. We see that our method demonstrated a significant advantage in generating motions from language commands. Additionally, it maintained a high rate of keeping the trajectories within the lane boundaries." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.604, + 0.766, + 0.62 + ], + "angle": 0, + "content": "6. Conclusions and Limitations" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.626, + 0.895, + 0.807 + ], + "angle": 0, + "content": "This paper introduces ChatSim, the first system for editing 3D driving scene simulations via language commands and realistic rendering with import of external digital assets. To effectively execute user commands, ChatSim adopts an LLM-agent collaboration workflow. To promote photorealistic simulation, we propose McNeRF and McLight for background and foreground rendering, respectively, accommodating multi-camera inputs. Experiments show that ChatSim successfully simulates customized data via language commands, achieving high-quality, photo-realistic outcomes. In future, we plan to integrate more background editing functionalities to ChatSim, such as weather changes." + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.821, + 0.661, + 0.838 + ], + "angle": 0, + "content": "Acknowledgement" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.845, + 0.895, + 0.888 + ], + "angle": 0, + "content": "This research is supported by NSFC under Grant 62171276 and the Science and Technology Commission of Shanghai Municipal under Grant 2151100900 and 22DZ2229005." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "15084" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.105 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.408, + 0.128 + ], + "angle": 0, + "content": "[1] 51Sim-One. https://wdp.51aes.com/news/27.2,3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.131, + 0.47, + 0.157 + ], + "angle": 0, + "content": "[2] LLC Agisoft. Metashape-photogrammetric processing of digital images and 3d spatial data generation, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.16, + 0.472, + 0.212 + ], + "angle": 0, + "content": "[3] Elif Akata, Lion Schulz, Julian Coda-Forno, Seong Joon Oh, Matthias Bethge, and Eric Schulz. Playing repeated games with large language models. arXiv preprint arXiv:2305.16867, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.215, + 0.471, + 0.284 + ], + "angle": 0, + "content": "[4] Alexander Amini, Igor Gilitschenski, Jacob Phillips, Julia Moseyko, Rohan Banerjee, Sertac Karaman, and Daniela Rus. Learning robust control policies for end-to-end autonomous driving from data-driven simulation. IEEE Robotics and Automation Letters, 5(2):1143-1150, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.286, + 0.471, + 0.342 + ], + "angle": 0, + "content": "[5] Rohan Anil, Andrew M Dai, Orhan First, Melvin Johnson, Dmitry Lepikhin, Alexandre Passos, Siamak Shakeri, Emanuel Taropa, Paige Bailey, Zhifeng Chen, et al. Palm 2 technical report. arXiv preprint arXiv:2305.10403, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.343, + 0.469, + 0.369 + ], + "angle": 0, + "content": "[6] Natalie M Banta. Property interests in digital assets: The rise of digital feudalism. Cardozo L. Rev., 38:1099, 2016. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.371, + 0.47, + 0.439 + ], + "angle": 0, + "content": "[7] Jonathan T Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P Srinivasan. Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5855–5864, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.441, + 0.47, + 0.509 + ], + "angle": 0, + "content": "[8] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Mip-nerf 360: Unbounded anti-aliased neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5470–5479, 2022. 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.511, + 0.471, + 0.592 + ], + "angle": 0, + "content": "[9] Luca Bergamini, Yawei Ye, Oliver Scheel, Long Chen, Chih Hu, Luca Del Pero, Băjej Osiński, Hugo Grimmett, and Peter Ondruska. Simnet: Learning reactive self-driving simulations from real-world observations. In 2021 IEEE International Conference on Robotics and Automation (ICRA), pages 5119-5125. IEEE, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.595, + 0.47, + 0.661 + ], + "angle": 0, + "content": "[10] Mark Boss, Varun Jampani, Kihwan Kim, Hendrik Lensch, and Jan Kautz. Two-shot spatially-varying brdf and shape estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3982-3991, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.664, + 0.47, + 0.734 + ], + "angle": 0, + "content": "[11] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. Advances in neural information processing systems, 33:1877-1901, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.735, + 0.471, + 0.817 + ], + "angle": 0, + "content": "[12] Holger Caesar, Varun Bankiti, Alex H Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuscenes: A multimodal dataset for autonomous driving. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11621-11631, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.818, + 0.471, + 0.901 + ], + "angle": 0, + "content": "[13] Ming-Fang Chang, John Lambert, Patsorn Sangkloy, Jagjeet Singh, Slawomir Bak, Andrew Hartnett, De Wang, Peter Carr, Simon Lucey, Deva Ramanan, et al. Argoverse: 3d tracking and forecasting with rich maps. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8748-8757, 2019. 1" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.804, + 0.107 + ], + "angle": 0, + "content": "[14] ChatGPT. https://openai.com/blog/chatgpt.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.109, + 0.894, + 0.164 + ], + "angle": 0, + "content": "[15] Chenyi Chen, Ari Seff, Alain Kornhauser, and Jianxiong Xiao. Deepdriving: Learning affordance for direct perception in autonomous driving. In Proceedings of the IEEE international conference on computer vision, pages 2722-2730, 2015. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.166, + 0.894, + 0.233 + ], + "angle": 0, + "content": "[16] Xiaozhi Chen, Kaustav Kundu, Ziyu Zhang, Huimin Ma, Sanja Fidler, and Raquel Urtasun. Monocular 3d object detection for autonomous driving. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2147-2156, 2016." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.235, + 0.894, + 0.291 + ], + "angle": 0, + "content": "[17] Xiaozhi Chen, Huimin Ma, Ji Wan, Bo Li, and Tian Xia. Multi-view 3d object detection network for autonomous driving. In Proceedings of the IEEE conference on Computer Vision and Pattern Recognition, pages 1907-1915, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.293, + 0.893, + 0.36 + ], + "angle": 0, + "content": "[18] Aakanksha Chowdhery, Sharan Narang, Jacob Devlin, Maarten Bosma, Gaurav Mishra, Adam Roberts, Paul Barham, Hyung Won Chung, Charles Sutton, Sebastian Gehrmann, et al. Palm: Scaling language modeling with pathways. arXiv preprint arXiv:2204.02311, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.363, + 0.894, + 0.417 + ], + "angle": 0, + "content": "[19] International Electrotechnical Commission et al. Iec 61966-2-1: 1999 multimedia systems and equipment- colour measurement and management- part 2-1: Colour management-default rgb colour space-srgb, 1999. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.42, + 0.893, + 0.46 + ], + "angle": 0, + "content": "[20] Blender Online Community. Blender - a 3D modelling and rendering package. Blender Foundation, Stichting Blender Foundation, Amsterdam, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.462, + 0.893, + 0.503 + ], + "angle": 0, + "content": "[21] Valter Crescenzi, Giansalvatore Mecca, Paolo Merialdo, et al. Roadrunner: Towards automatic data extraction from large web sites. In VLDB, pages 109-118, 2001. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.505, + 0.894, + 0.559 + ], + "angle": 0, + "content": "[22] Alexey Dosovitskiy, German Ros, Felipe Codevilla, Antonio Lopez, and Vladlen Koltun. Carla: An open urban driving simulator. In Conference on robot learning, pages 1-16. PMLR, 2017. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.562, + 0.893, + 0.615 + ], + "angle": 0, + "content": "[23] Yilun Du, Shuang Li, Antonio Torralba, Joshua B Tenenbaum, and Igor Mordatch. Improving factuality and reasoning in language models through multiagent debate. arXiv preprint arXiv:2305.14325, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.618, + 0.916, + 0.644 + ], + "angle": 0, + "content": "[24] Openscanerio Editor. https://github.com/ebadi/openscenarioeditor. 2,3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.647, + 0.833, + 0.661 + ], + "angle": 0, + "content": "[25] Unreal Engine. https://www.unrealengine.com/. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.663, + 0.893, + 0.716 + ], + "angle": 0, + "content": "[26] Ruiyuan Gao, Kai Chen, Enze Xie, Lanqing Hong, Zhenguo Li, Dit-Yan Yeung, and Qiang Xu. Magicdrive: Street view generation with diverse 3d geometry control. arXiv preprint arXiv:2310.02601, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.719, + 0.894, + 0.787 + ], + "angle": 0, + "content": "[27] Mathieu Garon, Kalyan Sunkavalli, Sunil Hadap, Nathan Carr, and Jean-François Lalonde. Fast spatially-varying indoor lighting estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6908-6917, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.79, + 0.894, + 0.857 + ], + "angle": 0, + "content": "[28] Jianfei Guo, Nianchen Deng, Xinyang Li, Yeqi Bai, Botian Shi, Chiyu Wang, Chenjing Ding, Dongliang Wang, and Yikang Li. Streetsurf: Extending multi-view implicit surface reconstruction to street views. arXiv preprint arXiv:2306.04988, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.86, + 0.894, + 0.901 + ], + "angle": 0, + "content": "[29] Rui Hao, Linmei Hu, Weijian Qi, Qingliu Wu, Yirui Zhang, and Liqiang Nie. Chatllm network: More brains, more intelligence. arXiv preprint arXiv:2304.12998, 2023. 3" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.916, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15085" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.161 + ], + "angle": 0, + "content": "[30] Yannick Hold-Geoffroy, Kalyan Sunkavalli, Sunil Hadap, Emiliano Gambaretto, and Jean-François Lalonde. Deep outdoor illumination estimation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7312–7321, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.163, + 0.472, + 0.233 + ], + "angle": 0, + "content": "[31] Yannick Hold-Geoffroy, Akshaya Athawale, and Jean-François Lalonde. Deep sky modeling for single image outdoor lighting estimation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6927–6935, 2019. 3, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.234, + 0.472, + 0.303 + ], + "angle": 0, + "content": "[32] Sirui Hong, Xiawu Zheng, Jonathan Chen, Yuheng Cheng, Ceyao Zhang, Zili Wang, Steven Ka Shing Yau, Zijuan Lin, Liyang Zhou, Chenyu Ran, et al. Metagpt: Meta programming for multi-agent collaborative framework. arXiv preprint arXiv:2308.00352, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.305, + 0.472, + 0.36 + ], + "angle": 0, + "content": "[33] Anthony Hu, Lloyd Russell, Hudson Yeo, Zak Murez, George Fedoseev, Alex Kendall, Jamie Shotton, and Gianluca Corrado. Gaia-1: A generative world model for autonomous driving. arXiv preprint arXiv:2309.17080, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.362, + 0.472, + 0.445 + ], + "angle": 0, + "content": "[34] Abhijit Kundu, Kyle Genova, Xiaoqi Yin, Alireza Fathi, Caroline Pantofaru, Leonidas J Guibas, Andrea Tagliasacchi, Frank Dellaert, and Thomas Funkhouser. Panoptic neural fields: A semantic object-aware neural scene representation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12871-12881, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.447, + 0.472, + 0.489 + ], + "angle": 0, + "content": "[35] Jean-François Lalonde and Iain Matthews. Lighting estimation in outdoor image collections. In 2014 2nd international conference on 3D vision, pages 131-138. IEEE, 2014. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.49, + 0.472, + 0.545 + ], + "angle": 0, + "content": "[36] Jean-François Lalonde, Srinivasa G Narasimhan, and Alexei A Efros. What do the sun and the sky tell us about the camera? International Journal of Computer Vision, 88: 24-51, 2010. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.547, + 0.472, + 0.616 + ], + "angle": 0, + "content": "[37] Chloe LeGendre, Wan-Chun Ma, Graham Fyffe, John Flynn, Laurent Charbonnel, Jay Busch, and Paul Debevec. Deeplight: Learning illumination for unconstrained mobile mixed reality. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5918-5928, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.618, + 0.472, + 0.673 + ], + "angle": 0, + "content": "[38] Peixuan Li, Huaici Zhao, Pengfei Liu, and Feidao Cao. Rtm3d: Real-time monocular 3d detection from object keypoints for autonomous driving. In European Conference on Computer Vision, pages 644-660. Springer, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.675, + 0.472, + 0.73 + ], + "angle": 0, + "content": "[39] Xiaofan Li, Yifu Zhang, and Xiaoqing Ye. Drivingdiffusion: Layout-guided multi-view driving scene video generation with latent diffusion model. arXiv preprint arXiv:2310.07771, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.732, + 0.472, + 0.8 + ], + "angle": 0, + "content": "[40] Zhengqin Li, Zexiang Xu, Ravi Ramamoorthi, Kalyan Sunkavalli, and Manmohan Chandraker. Learning to reconstruct shape and spatially-varying reflectance from a single image. ACM Transactions on Graphics (TOG), 37(6):1-11, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.802, + 0.472, + 0.857 + ], + "angle": 0, + "content": "[41] Zhuopeng Li, Lu Li, and Jianke Zhu. Read: Large-scale neural scene rendering for autonomous driving. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1522-1529, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.859, + 0.472, + 0.902 + ], + "angle": 0, + "content": "[42] Zhijian Liu, Haotian Tang, Alexander Amini, Xinyu Yang, Huizi Mao, Daniela L Rus, and Song Han. Bevfusion: Multi-task multi-sensor fusion with unified bird's-eye view represent" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "tation. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 2774-2781. IEEE, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.121, + 0.895, + 0.188 + ], + "angle": 0, + "content": "[43] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.19, + 0.895, + 0.245 + ], + "angle": 0, + "content": "[44] Milan Miric, Kevin J Boudreau, and Lars Bo Jeppesen. Protecting their digital assets: The use of formal & informal appropriability strategies by app developers. Research Policy, 48(8):103738, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.247, + 0.895, + 0.301 + ], + "angle": 0, + "content": "[45] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM Transactions on Graphics (ToG), 41(4):1-15, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.302, + 0.779, + 0.316 + ], + "angle": 0, + "content": "[46] OpenAI. Gpt-4 technical report. 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.317, + 0.895, + 0.372 + ], + "angle": 0, + "content": "[47] Julian Ost, Fahim Mannan, Nils Thuerey, Julian Knodt, and Felix Heide. Neural scene graphs for dynamic scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2856-2865, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.373, + 0.895, + 0.427 + ], + "angle": 0, + "content": "[48] Julian Ost, Issam Laradji, Alejandro Newell, Yuval Bahat, and Felix Heide. Neural point light fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18419-18429, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.428, + 0.895, + 0.509 + ], + "angle": 0, + "content": "[49] Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, et al. Training language models to follow instructions with human feedback. Advances in Neural Information Processing Systems, 35:27730-27744, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.511, + 0.892, + 0.58 + ], + "angle": 0, + "content": "[50] Jonah Philion and Sanja Fidler. Lift, splat, shoot: Encoding images from arbitrary camera rigs by implicitly unprojecting to 3d. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XIV 16, pages 194-210. Springer, 2020. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.581, + 0.892, + 0.622 + ], + "angle": 0, + "content": "[51] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.623, + 0.892, + 0.705 + ], + "angle": 0, + "content": "[52] Guodong Rong, Byung Hyun Shin, Hadi Tabatabaee, Qiang Lu, Steve Lemke, Martins Mozeiko, Eric Boise, Geehoon Uhm, Mark Gerow, Shalin Mehta, et al. Lgsvl simulator: A high fidelity simulator for autonomous driving. In 2020 IEEE 23rd International conference on intelligent transportation systems (ITSC), pages 1-6. IEEE, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.707, + 0.895, + 0.775 + ], + "angle": 0, + "content": "[53] Soumyadip Sengupta, Jinwei Gu, Kihwan Kim, Guilin Liu, David W Jacobs, and Jan Kautz. Neural inverse rendering of an indoor scene from a single image. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8598-8607, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.776, + 0.895, + 0.845 + ], + "angle": 0, + "content": "[54] Shital Shah, Debadeepta Dey, Chris Lovett, and Ashish Kapoor. Airsim: High-fidelity visual and physical simulation for autonomous vehicles. In Field and Service Robotics: Results of the 11th International Conference, pages 621-635. Springer, 2018. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.846, + 0.895, + 0.901 + ], + "angle": 0, + "content": "[55] Gowri Somanath and Daniel Kurz. Hdr environment map estimation for real-time augmented reality. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11298-11306, 2021. 3" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.895, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "15086" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.161 + ], + "angle": 0, + "content": "[56] Cheng Sun, Min Sun, and Hwann-Tzong Chen. Direct voxel grid optimization: Super-fast convergence for radiance fields reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5459-5469, 2022. 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.164, + 0.472, + 0.289 + ], + "angle": 0, + "content": "[57] Pei Sun, Henrik Kretzschmar, Xerxes Dotiwalla, Aurelien Chouard, Vijaysai Patnaik, Paul Tsui, James Guo, Yin Zhou, Yuning Chai, Benjamin Caine, Vijay Vasudevan, Wei Han, Jiquan Ngiam, Hang Zhao, Aleksei Timofeev, Scott Ettinger, Maxim Krivokon, Amy Gao, Aditya Joshi, Yu Zhang, Jonathon Shlens, Zhifeng Chen, and Dragomir Anguelov. Scalability in perception for autonomous driving: Waymo open dataset. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.291, + 0.471, + 0.374 + ], + "angle": 0, + "content": "[58] Pei Sun, Henrik Kretzschmar, Xerxes Dotiwalla, Aurelien Chouard, Vijaysai Patnaik, Paul Tsui, James Guo, Yin Zhou, Yuning Chai, Benjamin Caine, et al. Scalability in perception for autonomous driving: Waymo open dataset. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2446-2454, 2020. 1, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.375, + 0.471, + 0.417 + ], + "angle": 0, + "content": "[59] Alexander Swerdlow, Runsheng Xu, and Bolei Zhou. Street-view image generation from a bird's-eye view layout. arXiv preprint arXiv:2301.04634, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.419, + 0.472, + 0.488 + ], + "angle": 0, + "content": "[60] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.49, + 0.472, + 0.546 + ], + "angle": 0, + "content": "[61] Haithem Turki, Jason Y Zhang, Francesco Ferroni, and Deva Ramanan. Suds: Scalable urban dynamic scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12375-12385, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.548, + 0.472, + 0.63 + ], + "angle": 0, + "content": "[62] Peng Wang, Yuan Liu, Zhaoxi Chen, Lingjie Liu, Ziwei Liu, Taku Komura, Christian Theobalt, and Wenping Wang. F2-nerf: Fast neural radiance field training with free camera trajectories. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4150–4159, 2023. 2, 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.632, + 0.472, + 0.688 + ], + "angle": 0, + "content": "[63] Xiaofeng Wang, Zheng Zhu, Guan Huang, Xinze Chen, and Jiwen Lu. Drivedreamer: Towards real-world-driven world models for autonomous driving. arXiv preprint arXiv:2309.09777, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.689, + 0.472, + 0.758 + ], + "angle": 0, + "content": "[64] Zian Wang, Wenzheng Chen, David Acuna, Jan Kautz, and Sanja Fidler. Neural light field estimation for street scenes with differentiable virtual object insertion. In European Conference on Computer Vision, pages 380-397. Springer, 2022. 3, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.761, + 0.472, + 0.816 + ], + "angle": 0, + "content": "[65] Zhenhailong Wang, Shaoguang Mao, Wenshan Wu, Tao Ge, Furu Wei, and Heng Ji. Unleashing cognitive synergy in large language models: A task-solving agent through multi-personal self-collaboration. arXiv preprint arXiv:2307.05300, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.818, + 0.472, + 0.9 + ], + "angle": 0, + "content": "[66] Bichen Wu, Forrest Iandola, Peter H Jin, and Kurt Keutzer. Squeezedet: Unified, small, low power fully convolutional neural networks for real-time object detection for autonomous driving. In Proceedings of the IEEE conference on computer vision and pattern recognition workshops, pages 129-137, 2017. 1" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.161 + ], + "angle": 0, + "content": "[67] Qingyun Wu, Gagan Bansal, Jieyu Zhang, Yiran Wu, Shaokun Zhang, Erkang Zhu, Beibin Li, Li Jiang, Xiaoyun Zhang, and Chi Wang. Autogen: Enabling next-gen llm applications via multi-agent conversation framework. arXiv preprint arXiv:2308.08155, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.164, + 0.894, + 0.245 + ], + "angle": 0, + "content": "[68] Zirui Wu, Tianyu Liu, Liyi Luo, Zhide Zhong, Jianteng Chen, Hongmin Xiao, Chao Hou, Haozhe Lou, Yuantao Chen, Runyi Yang, Yuxin Huang, Xiaoyu Ye, Zike Yan, Yongliang Shi, Yiyi Liao, and Hao Zhao. Mars: An instance-aware, modular and realistic simulator for autonomous driving. CICAI, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.248, + 0.893, + 0.289 + ], + "angle": 0, + "content": "[69] Ziyang Xie, Junge Zhang, Wenye Li, Feihu Zhang, and Li Zhang. S-nerf: Neural radiance fields for street views. arXiv preprint arXiv:2303.00749, 2023. 2, 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.291, + 0.893, + 0.331 + ], + "angle": 0, + "content": "[70] Yinda Xu and Lidong Yu. Drl-based trajectory tracking for motion-related modules in autonomous driving. arXiv preprint arXiv:2308.15991, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.334, + 0.893, + 0.389 + ], + "angle": 0, + "content": "[71] Kairui Yang, Enhui Ma, Jibin Peng, Qing Guo, Di Lin, and Kaicheng Yu. Bevcontrol: Accurately controlling street-view elements with multi-perspective consistency via bev sketch layout. arXiv preprint arXiv:2308.01661, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.391, + 0.894, + 0.471 + ], + "angle": 0, + "content": "[72] Zhenpei Yang, Yuning Chai, Dragomir Anguelov, Yin Zhou, Pei Sun, Dumitru Erhan, Sean Rafferty, and Henrik Kretzschmar. Surfelgan: Synthesizing realistic sensor data for autonomous driving. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11118-11127, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.474, + 0.894, + 0.543 + ], + "angle": 0, + "content": "[73] Ze Yang, Yun Chen, Jingkang Wang, Sivabalan Manivasagam, Wei-Chiu Ma, Anqi Joyce Yang, and Raquel Urtasun. Unisim: A neural closed-loop sensor simulator. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1389-1399, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.545, + 0.894, + 0.613 + ], + "angle": 0, + "content": "[74] Jinsong Zhang, Kalyan Sunkavalli, Yannick Hold-Geoffroy, Sunil Hadap, Jonathan Eisenman, and Jean-François Lalonde. All-weather deep outdoor lighting estimation. In Proceedings of the IEEE/CVF conference on Computer Vision and Pattern Recognition, pages 10158–10166, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.615, + 0.893, + 0.669 + ], + "angle": 0, + "content": "[75] Yichao Zhou, Jingwei Huang, Xili Dai, Linjie Luo, Zhili Chen, and Yi Ma. HoliCity: A city-scale data platform for learning holistic 3D structures. 2020. arXiv:2008.03286 [cs.CV]. 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.672, + 0.894, + 0.74 + ], + "angle": 0, + "content": "[76] Mingchen Zhuge, Haozhe Liu, Francesco Faccio, Dylan R Ashley, Robert Csordás, Anand Gopalakrishnan, Abdullah Hamdi, Hasan Abed Al Kader Hammoud, Vincent Herrmann, Kazuki Irie, et al. Mindstorms in natural language-based societies of mind. arXiv preprint arXiv:2305.17066, 2023. 3" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.74 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "15087" + } + ] +] \ No newline at end of file diff --git a/2024/Editable Scene Simulation for Autonomous Driving via Collaborative LLM-Agents/49eb43d9-92f1-4849-a7c5-3da61380ffc6_origin.pdf b/2024/Editable Scene Simulation for Autonomous Driving via Collaborative LLM-Agents/49eb43d9-92f1-4849-a7c5-3da61380ffc6_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..0c9a515357f8851836f74367a82c006d378ea8cf --- /dev/null +++ b/2024/Editable Scene Simulation for Autonomous Driving via Collaborative LLM-Agents/49eb43d9-92f1-4849-a7c5-3da61380ffc6_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:353474dbedc0f6362e1d00f211724c0fd8166f3e22f9b8bdcdfbb16e00f30858 +size 5002125 diff --git a/2024/Editable Scene Simulation for Autonomous Driving via Collaborative LLM-Agents/full.md b/2024/Editable Scene Simulation for Autonomous Driving via Collaborative LLM-Agents/full.md new file mode 100644 index 0000000000000000000000000000000000000000..7a809101fb4ff282f00852b11b7c9b6d0e052974 --- /dev/null +++ b/2024/Editable Scene Simulation for Autonomous Driving via Collaborative LLM-Agents/full.md @@ -0,0 +1,380 @@ +# Editable Scene Simulation for Autonomous Driving via Collaborative LLM-Agents + +Yuxi Wei $^{1,5*}$ Zi Wang $^{3,5*}$ Yifan Lu $^{1,5*}$ Chenxin Xu $^{1,5*}$ Changxing Liu $^{1,5}$ Hao Zhao $^{4}$ Siheng Chen $^{1,2,5\dagger}$ Yanfeng Wang $^{1,2}$ Shanghai Jiao Tong University + +$^{3}$ Carnegie Mellon University $^{4}$ Tsinghua University $^{5}$ Multi-Agent Governance & Intelligence Crew (MAGIC) {wyx3590236732, yifan_lu, xcxwakaka, cx-liu}@sjtu.edu.cn {sihengc, wangyanfeng}@sjtu.edu.cn ziwang2@andrew.cmu.edu zhaohao@air.tsinghua.edu.cn + +# Abstract + +Scene simulation in autonomous driving has gained significant attention because of its huge potential for generating customized data. However, existing editable scene simulation approaches face limitations in terms of user interaction efficiency, multi-camera photo-realistic rendering and external digital assets integration. To address these challenges, this paper introduces ChatSim, the first system that enables editable photo-realistic 3D driving scene simulations via natural language commands with external digital assets. To enable editing with high command flexibility, ChatSim leverages a large language model (LLM) agent collaboration framework. To generate photo-realistic outcomes, ChatSim employs a novel multi-camera neural radiance field method. Furthermore, to unleash the potential of extensive high-quality digital assets, ChatSim employs a novel multi-camera lighting estimation method to achieve scene-consistent assets' rendering. Our experiments on Waymo Open Dataset demonstrate that ChatSim can handle complex language commands and generate corresponding photo-realistic scene videos. Code can be accessed at: https://github.com/yifanlu0227/ChatSim. + +# 1. Introduction + +Perception [12, 15-17, 38, 66] is the window of an autonomous vehicle into the external environment. To ensure the robustness of the vehicle's perceptual capabilities during both training and testing phases, it necessitates the collection of high-quality perception data in substantial volumes [13, 58]. However, the operation of a fleet for the acquisition of real-world data often incurs prohibitive expenses, particularly for specialized or customized requirements. For instance, in the aftermath of an accident or intervention in + +![](images/369a5d248f75a13a44d4444719f784db88d394a33e5dda862aad944811a3c72c.jpg) +Figure 1. ChatSim enables the editing of photo-realistic 3D driving scene simulations via language commands. + +volving an autonomous vehicle, it is imperative to test the vehicle's perception system across a spectrum of similar scenarios. While replicating such scenario data from real-world instances is nearly impossible due to the uncontrollability of actual scenes [9, 52], customized scene simulation emerges as a vital and feasible alternative. It enables the precise modeling of specific conditions without high costs and logistical complexities of real-world data collection [4, 72]. + +To effectively simulate customized driving scenes, we identify three key properties as fundamental. First, the simulation should be capable of following sophisticated or abstract demands, thereby facilitating the production. Second, the simulation should generate photo-realistic, view-consistent outcomes, which allow for the closest approximation to vehicle observations in real-world scenarios. Third, it should allow for the integration of external digital assets [6, 44] with their photo-realistic textures and materials while fitting the lighting conditions. This capability would unlock the potential for data expansion by incorporating a wide array of external digital assets, satisfying customized needs. + +A vast array of significant works have been proposed for scene simulation, yet they fail to meet all three of + +these requirements. Traditional graphics engines, such as CARLA [22] and UE [25], offer editable virtual environments with external digital assets, but the data realism is restricted by asset modeling and rendering qualities. Image generation based methods, such as BEVControl [71], DriveDreamer [63], MagicDrive [26], can generate realistic scene images based on various control signals, including BEV maps, bounding boxes and camera poses. However, they struggle to maintain view consistency and face challenges in importing external digital assets due to the absence of 3D spatial modeling. Rendering-based methods have been proposed to obtain photo-realistic and view-consistent scene simulation. Notable examples like UniSim [73] and MARS [68] come equipped with a suite of scene-editing tools. However, these systems require extensive user involvement in every trivial editing step via code implementation, which is ineffective when performing the editing. Furthermore, while they handle vehicles in observed scenarios effectively, their inability to support external digital assets restricts opportunities for data expansion and customization. + +To fulfill the identified requirements, we introduce ChatSim, the first system that enables editable photo-realistic 3D driving scene simulations via natural language commands with external digital assets. To use ChatSim, users simply engage in a conversation with the system, issuing commands through natural language without any involvement in intermediate simulation steps; see Figure 1 for illustration. + +To address complex or abstract user commands effectively, ChatSim adopts a large language model (LLM)-based multi-agent collaboration framework. The key idea is to exploit multiple LLM agents, each with a specialized role, to decouple an overall simulation demand into specific editing tasks, thereby mirroring the task division and execution typically founded in the workflow of a human-operated company. This workflow offers two key advantages for scene simulation. First, LLM agents' ability to process human language commands allows for intuitive and dynamic editing of complex driving scenes, enabling precise adjustments and feedback. Second, the collaboration framework enhances simulation efficiency and accuracy by distributing specific editing tasks among specialized agents, ensuring detailed and realistic simulations with improved task completion rates. + +To generate photo-realistic outcomes, we propose McNeRF in ChatSim, a novel neural radiance field method that incorporates multi-camera inputs, offering a broader scene rendering. This integration fully exploits camera setups on vehicles but raises two significant challenges: camera pose misalignment due to asynchronizated trigger times and brightness inconsistency due to different camera exposure times. To address camera pose misalignment, McNeRF uses a multi-camera alignment to reduce extrinsic parameter noises, ensuring rendering quality. To address brightness inconsistency, McNeRF integrates the critical exposure times + +to recover scene radiance in HDR, markedly mitigating the issue of color discrepancies at the intersections of two camera images with different exposure times. + +To import external digital assets with their realistic textures and materials, we propose McLight, a novel multicamera lighting estimation that blends skydome and surrounding lighting. Our skydome estimation restores accurate sun behavior with peak intensity residual connection, enabling the rendering of prominent shadows. For surrounding lighting, McLight queries McNeRF to achieve complex location-specific illumination effects, like those in the tree shade with sunlight being blocked. This significantly improves the rendering realism of the integrated 3D assets. + +We conduct extensive experiments on the Waymo autonomous driving dataset and show that ChatSim generates photo-realistic customized perception data including dangerous corner cases according to various human language commands. Our method is compatible with mixed, highly abstract and multi-round commands. Our method achieves SoTA performance with an improvement of $4.5\%$ in photorealism with a wide-angle rendering. Moreover, we demonstrate our lighting estimation outperforms the SoTA methods both qualitatively and quantitatively, reducing the intensity error and angular error by $57.0\%$ and $9.9\%$ . + +# 2. Related Work + +Scene simulation for autonomous driving. Current scene simulation methods can be generally divided into three categories: graphics engines, image generation, and scene rendering. Graphics engines, such as CARLA [22], AirSim [54], OpenScenario Editor [24], 51Sim-One [1] and RoadRunner [21], create a virtual world for simulating a wide range of driving scenarios. However, there exists a significant domain gap between the virtual world and reality. Image generation methods can generate realistic scene images based on different control signals, such as HD maps [26, 39, 59], sketch layout [71], bounding boxes [26, 39, 63], text [26, 33, 39, 63] and driving actions [33, 63]. However, these approaches can hardly maintain scene consistency. To obtain a coherent driving scene, methods based on scene rendering target to reconstruct the 3D scene. READ [41] employs point clouds and uses a U-Net to render images. With the rapid development of Neural Radiance Field (NeRF) [7, 8, 43, 45, 56, 62], several works [28, 34, 47, 48, 61, 68, 69, 73] also exploit NeRFs to model cars and static street backgrounds in outdoor environments. Moreover, notable examples like UniSim [73] and MARS [68] come equipped with a suite of scene-editing tools. However, these methods require extensive user involvement in intermediate editing steps and they fail to support external digital assets for data expansion. In this work, we propose ChatSim that achieves automatic simulation editing via language commands and integrates external digital assets to enhance realism and flexibility. In ChatSim, + +
MethodPhoto-realisticDim.Multi-cameraEditableExternal assetsLanguageOpen-source
CARLA [22]×3D✓✓✓×✓
AirSim [54]×3D✓✓✓×✓
OpenScenario [24]×3D✓✓✓×✓
51Sim-One [1]×3D✓✓✓××
RoadRunner [21]×3D✓✓✓×✓
BEVGen [59]✓2D✓✓××✓
BEVCentral [71]✓2D✓✓×××
DriveDreamer [63]✓2D✓✓×✓×
DrivingDiffusion [39]✓2D✓✓×✓×
GAIA-1 [33]✓2D×✓×✓×
MagicDrive [26]✓2D✓✓×××
READ [41]✓3D××××✓
Neural SG [47]✓3D×✓××✓
Neural PLF [48]✓3D××××✓
S-NeRF [69]✓3D✓×××✓
UniSim [73]✓3D×✓×××
MARS [68]✓3D×✓××✓
ChatSim (Ours)✓3D✓✓✓✓✓
+ +Table 1. Comparison of existing and proposed methods for autonomous driving simulation. + +we integrate McNeRF, a novel neural radiance field designed to leverage multi-camera inputs for high-fidelity rendering. + +Lighting estimation. Lighting estimation focuses on assessing the illumination conditions of a real-world environment to seamlessly integrate digital objects. Early methods [35, 36] for outdoor environments use explicit cues like detected shadows on the ground. Recent works usually adopt learning-based approaches [27, 30, 31, 37, 40, 74] by predicting different lighting representations like spherical lobes [10, 40], light probes [37], environment map [53, 55], HDR sky model [31, 64, 74] and lighting volume [64]. However, few of them consider multi-camera input, which is common for driving scenarios. In this paper, we propose a novel multi-camera lighting estimation method, McLight, combining with our McNeRF, to estimate a wider range of lighting and obtain the spatially-varying lighting effects of assets. + +Large Language model and collaborative framework. Large Language Models (LLMs) are AI systems trained on extensive data to understand, generate, and respond to human language. GPT [11] is a pioneering work to generate human-like content. The following updated versions GPT-3.5 [14] and GPT-4 [46], provide more intelligent capabilities like chatting, browsing and coding. Notable other large language models include InstructGPT [49], LLaMA [60] and PaLM [5, 18]. Based on LLM, many works [3, 23, 29, 65, 76] improve the problem-solving abilities by integrating communication among multiple agents. [32] and [67] define a group of well-organized agents to form operating procedures with conversation and code programming. In this paper, we exploit the power of collaborative LLM agents in simulation for autonomous driving, enabling the various editing of 3D scenes via language commands. + +# 3. Collaborative LLM-Agents for Editing + +The ChatSim system analyzes specific user commands and returns a video that meets customized needs; see Figure 2. Since user commands could be abstract and sophisticated, + +it requires the system to have flexible task-handling ability. Directly applying a single LLM agent struggles with multi-step reasoning and cross-referencing. To address this, we design a series of collaborative LLM agents, where each agent is responsible for a unique aspect of the editing task. + +# 3.1. Specific Agent's Functionality + +Agents in ChatSim comprise two key components: a Large Language Model (LLM) and the corresponding role functions. The LLM is responsible for understanding the received commands while the role functions process the received data. Each agent is equipped with unique LLM prompts and role functions tailored to their specific duties within the system. To accomplish their tasks, agents first convert the received commands to a structured configuration using LLM with the assistance of prompts. Then the role functions utilize the structured configuration as parameters to process the received data and produce the desired outcomes; see an agent example on the right side of Figure 2. This workflow endows agents with both language interpretation capabilities and precise execution capabilities. + +Project Manager Agent. The project manager agent decomposes direct commands into clear natural language instructions dispatched to other editing agents. To equip the project manager agent with the capability of command decomposition, we design a series of prompts for its LLM. The core idea of the prompts is to describe the action set, give the overall goal, and define the output form with examples; The role functions send the decomposed instructions to other agents for editing. The presence of the project manager agent enhances the system's robustness in interpreting various inputs and streamlines operations for clarity and fine granularity. + +Tech agent for view adjustment. The view adjustment agent generates suitable extrinsic camera parameters. The LLM in the agent translates the natural language instructions for viewpoint adjustment into movement parameters to the target viewpoint's position and angle. In role functions, these movement parameters are turned into transformation matrices required by the extrinsic, which are then multiplied by the original parameters to yield a new viewpoint. + +Tech agent for background rendering. The background rendering agent renders the scene background based on multi-camera images. The LLM receives the rendering command and then operates the role functions for rendering. Notably, in role functions, we specifically integrate a novel neural radiance field method (McNeRF) taking multi-camera inputs and considering exposure time, solving the problem of blurring and brightness inconsistency in multi-camera rendering, see more details in Section 4.1. + +Tech agent for vehicle deleting. The vehicle deleting agent removes specified vehicles from the background. It first identifies current vehicle attributes like 3D bounding boxes and colors from given scene information or results from a scene + +![](images/05f9bf54c389c3fc12db037c655fa35ff99434483de4cc028ef5f24bef591de3.jpg) +Figure 2. ChatSim system overview. The system exploit multiple collaborative LLM agents with specialized roles to decouple an overall demand into specific editing tasks. Each agent equips an LLM and corresponding role functions to interpret and execute its specific tasks. + +![](images/3332b5392e5399f26ee735fa931df462f0f7f160fd44cf61038e0dcca3811bb1.jpg) +Figure 3. Prompt example of view adjustment agent. + +perception model like [42]. The LLM gathers attributes of the vehicles and performs matching with user requests. Upon confirming the targeted vehicles, it employs a per-frame inpainting model as the role functions, such as latent diffusion methods [51], to effectively delete them from the scene. + +Tech agent for 3D asset management. The 3D asset management agent selects and modifies 3D digital assets according to user specifications. It constructs and maintains a 3D digital asset bank; see our bank details in the Appendix. To facilitate the addition of various objects, the agent first uses LLM to select the most suitable asset by key attributes matching with the requirements, such as color and type. If the matching is not perfect, the agent could modify the asset through its role functions like changing the color. + +Tech agent for vehicle motion. The vehicle motion agent creates the initial places and subsequent motions of vehicles following the requests. Existing vehicle motion generation methods cannot directly generate motion purely from text and the scene map. To solve the problem, here we propose a novel text-to-motion method. The key idea is linking a placement and planning module as role functions with LLMs to extract and turn motion attributes into coordinates. Motion attributes include position attributes (e.g., distance, direction) and movement attributes (e.g., speed, action). For the placement module, we endow each lane node in the lane map with its attributes to match with the position attributes. The + +planning module plans the vehicle's approximate destination lane node and then plans the intermediate trajectory by fitting the Bezier curves. We also add trajectory tracking [70] to fit vehicle dynamics; see more details in the Appendix. + +Tech agent for foreground rendering. The foreground rendering agent integrates camera extrinsic information, 3D assets, and motion information to render foreground objects in the scene. Notably, to seamlessly integrate the external assets with the current scene, we design a multi-camera lighting estimation method (McLight) into the role functions, coupling with McNeRF. The estimated illumination is then utilized by Blender API to generate foreground images. The detailed technical aspects will be elaborated in Section 4.2. + +# 3.2. Agent Collaboration Workflow + +Agents with tailored functions collaboratively work together to edit based on user commands. The project manager orchestrates and dispatches instructions to editing agents. The editing agents form two teams: background generation and foreground generation. For background generation, the background rendering agent generates rendered images using the extrinsic parameters from the view adjustment agent, followed by inpainting by the vehicle deleting agent. For foreground generation, the foreground rendering agent renders the images using the extrinsic parameters from the view adjustment agent, selected 3D assets from 3D asset management agent, and generated motions from vehicle motion agent. Finally, the foreground and background images are composed to create and deliver a video to the user. The editing information in each agent's configuration is recorded by the project manager agent for possible multi-round editorings. + +# 4. Novel Rendering Methods + +Based on the collaborative LLM agents framework introduced in Section 3, this section presents two novel rendering + +techniques to enhance photo-realism in simulations. To tackle the rendering challenges caused by multiple cameras, we propose multi-camera neural radiance field (McNeRF), a novel NeRF model considering the varied camera exposure times for visual consistency. To render realistic external digital assets with location-specific lighting and accurate shadows, we propose McLight, a novel hybrid lighting estimation method combined with our McNeRF. Note that McNeRF and McLight are leveraged by the background rendering agent and the foreground rendering agent, respectively. + +# 4.1. McNeRF for Background Rendering + +An autonomous vehicle typically equips multiple cameras to achieve a comprehensive perception view. However, this poses challenges for NeRF training due to the misaligned multi-camera poses from asynchronized camera trigger times and the brightness inconsistency originating from different exposure times. To address these challenges, the proposed McNeRF uses two techniques: multi-camera alignment and brightness-consistent rendering. + +Multi-camera alignment. Autonomous vehicles, despite having a localization module for accurate camera poses, face challenges with asynchronous trigger times across multiple cameras. To align camera extrinsics for NeRF training, our core idea is to leverage a consistent spatial coordinate system provided by Agisoft Metashape [2] to align the images captured by multiple cameras at different timestamps. + +Specifically, let $\mathcal{I}^{(i,k)}$ and $\xi^{(i,k)}$ be the image captured by the $i$ th camera at the $k$ th trigger and the corresponding camera pose in the vehicle's global coordinate space, respectively. We first input all images into Metashape for recalibration. The aligned camera pose is then obtained as: + +$$ +\widehat {\xi} ^ {(i, k)} = T _ {M \to G} \cdot \xi_ {M} ^ {(i, k)}, +$$ + +where $\xi_{M}^{(i,k)}$ denotes the recalibrated camera pose in the Metashape's unified spatial coordinate space, and $T_{M\rightarrow G}$ is the transformation from the Metashape's coordinate space to the vehicle's global coordinate space. After alignment, the pose noise can be significantly reduced. Then, the aligned camera pose $\widehat{\xi}^{(i,t)}$ can be used to generate the origins and directions of rays for McNeRF, enabling high-fidelity rendering. The aligned pose can also facilitate the foreground rendering agent's operations. + +Brightness-consistent rendering. The exposure times of cameras can differ substantially, causing significant brightness differences across images, hindering the NeRF training. As shown in Figure 4, McNeRF, addresses this by incorporating exposure times into HDR radiance fields, prompting brightness consistency. + +We adopt F2-NeRF [62] as our backbone model to handle the unbounded scene, sampling $K$ points along the ray $\mathbf{r}$ and estimating each point's HDR radiance $\mathbf{e}_k$ and density $\sigma_k$ . + +![](images/024a3e27b985c79138121fbda6995993109f49d829e053842b6a1b358f7f57b6.jpg) + +![](images/7f9a739e4c4e43f357b42161a700e97bab79b2aa9ba7680c886f357178d0a0a1.jpg) +Figure 4. Rendering framework. The main components include McNeRF and McLight. Background rendering uses McNeRF to predict HDR pixel value and convert it to LDR with sRGB OETF. McLight includes a skydome lighting estimation network and adopts McNeRF to generate surrounding lighting. + +The HDR light intensity is then calculated as: + +$$ +\widehat {\mathcal {I}} _ {\mathrm {H D R}} (\mathbf {r}) = f (\Delta t) \cdot \sum_ {k = 1} ^ {K} T _ {k} \alpha_ {k} \mathbf {e} _ {k}, \tag {1} +$$ + +where $\alpha_{k} = 1 - \exp (-\sigma_{k}\delta_{i})$ is the opacity, $\delta_{i}$ is the point sampling interval, $T_{k} = \prod_{i = 0}^{k - 1}(1 - \alpha_{i})$ is the accumulated transmittance and $\Delta t$ is the exposure time. The normalization function $f(\Delta t) = 1 + \epsilon (\Delta t - \mu) / \sigma$ is designed to stabilize training, where $\epsilon$ is a hyperparameter for scaling, $\mu$ and $\sigma$ are the mean and standard deviation of the exposure times of all images, respectively. + +By predicting scene radiance in HDR and multiplying it by the exposure time, we recover the light intensity received by the sensor and tackle the inconsistent color supervision at the intersections of two camera images with distinct exposure times. Moreover, the HDR light intensity outputted by McNeRF can provide scene-level illumination for foreground object rendering, a topic further discussed in Section 4.2. + +To train the rendering network, we enforce the consistency of radiance between the rendered image (prediction) and the captured image (ground-truth). Given the ground-truth image $\mathcal{I}$ , the loss function is then: + +$$ +\mathcal {L} = \frac {1}{| R |} \sum_ {\mathbf {r} \in R} \left(\mathrm {O E T F} \left(\widehat {\mathcal {I}} _ {\mathrm {H D R}} (\mathbf {r})\right) - \mathcal {I} (\mathbf {r})\right) ^ {2}, +$$ + +where $R$ represents the ray set and $\mathrm{OETF}(\cdot)$ is the sRGB opto-electronic transfer function (gamma correction) [19] that converts HDR light intensity to LDR colors. + +# 4.2. McLight for Foreground Rendering + +To enrich the scene's content with substantial digital 3D assets, we employ Blender [20] foreground virtual objects' rendering. A seamless insertion critically depends on accurately estimating the scene's illumination conditions. Thus, as shown in Figure 4, we propose McLight, a novel hybrid lighting estimation consisting of skydome lighting and sur + +rounding lighting. + +Skydome lighting estimation. Estimating skydome lighting from images is challenging for restoring accurate sun behavior. To achieve this, we propose a novel residual connection from the estimated peak intensity to the HDR reconstruction to address over-smoothing output. Further, we adopt a self-attention mechanism to fuse multi-camera inputs, capturing complementary visual cues. + +Here we employ a two-stage process. In the first stage, we train an autoencoder to reconstruct the corresponding HDR panorama from an LDR panorama. Following [64], the encoder transforms the LDR skydome panorama into three intermediate vectors, including the peak direction vector $\mathbf{f}_{\mathrm{dir}} \in \mathbb{R}^3$ , the intensity vector $\mathbf{f}_{\mathrm{int}} \in \mathbb{R}_+^3$ , and the sky content vector $\mathbf{f}_{\mathrm{content}} \in \mathbb{R}^{64}$ . However, as HDR intensity behaves like an impulse response at its peak position, with pixel values thousands of times higher than its neighbors, it is difficult for the decoder to recover such patterns. To tackle this, we design a residual connection that injects $\mathbf{f}_{\mathrm{int}}$ into the decoded HDR panorama with a spherical Gaussian lobe attenuation. This explicitly restores the peak intensity of the sun in the reconstructed HDR panorama, allowing us to render strong shadows for virtual objects. + +In the second stage, we train an image encoder and a multi-camera fusion module built upon the pretrained decoder from the first stage. Specifically, for images from each camera, a shared image encoder predicts the peak direction vector $\mathbf{f}_{\mathrm{dir}}^{(i)}$ , the intensity vector $\mathbf{f}_{\mathrm{int}}^{(i)}$ , and the sky content vector $\mathbf{f}_{\mathrm{content}}^{(i)}$ for each image $\mathcal{I}^{(i)}$ , where $i$ is the camera index. We design the latent vector fusion across the multiple camera views as follows: all $\mathbf{f}_{\mathrm{dir}}^{(i)}$ are aligned to the front-facing view using their extrinsic parameters and averaged to form $\bar{\mathbf{f}}_{\mathrm{dir}}$ ; all $\mathbf{f}_{\mathrm{int}}^{(i)}$ are averaged to yield $\bar{\mathbf{f}}_{\mathrm{int}}$ ; all $\mathbf{f}_{\mathrm{content}}^{(i)}$ are integrated into $\bar{\mathbf{f}}_{\mathrm{content}}$ through a self-attention module. Finally, the pretrained decoder reconstructs the HDR skydome image $\mathcal{I}_{\mathrm{skydome}}$ from $\bar{\mathbf{f}}_{\mathrm{dir}}$ , $\bar{\mathbf{f}}_{\mathrm{int}}$ and $\bar{\mathbf{f}}_{\mathrm{content}}$ . + +Compared to alternative approaches [31, 64], our multicamera sky dome estimation technique accurately reproduces the sun's intensity response behavior at its peak with our residual designs, significantly improving the accuracy and fidelity of the skydome reconstruction. + +Surrounding lighting estimation. Merely modeling the sky-dome cannot replicate the complex location-specific lighting effects, like those in the shade with sunlight blocked by trees or buildings. Our McNeRF is capable of storing precise 3D scene information, enabling us to capture the surrounding scene's impact on lighting. This approach facilitates the achievement of spatially-varying lighting estimation. Specifically, we sample the hemisphere rays at the virtual object's position $\mathbf{o}$ . The rays' directions, $\mathbf{d}_i$ , $i = 0,1,\dots,h\times w$ , are aligned with pixel coordinates on a unit sphere using equirectangular projection from an environment map, where $h$ and $w$ are map's height and width. With the ray + +![](images/22006f29ef30b26d3218bf91ba48a7fddcd5eba050b07813c70aee86e212003d.jpg) +Front left + +![](images/911ac196eb6cba8fdf316d43be5a36f0d9a14ef7a2bec755b9cb0886e9ea1c90.jpg) +Front + +![](images/e19903211766eff71f549d3d33153a6bfdca3335d06d8ebd131424b6936e5b46.jpg) +Front right + +![](images/b091b38369abab6351eb8e2f4c0ef52abc36ffaf6f7c01c133713e6b629af4c6.jpg) +Command: "Remove all cars in the scene and add a Porsche driving the wrong way toward me fast. Additionally, add a police car also driving the wrong way and chasing behind the Porsche. The view should be moved 5 meters ahead and 0.5 meters above." +Figure 5. Editing result under a complex and mixed command. + +$\mathbf{r} = \mathbf{o} + t\mathbf{d}_i$ , we query our McNeRF as Equation 1 to obtain HDR surrounding lighting $\mathcal{I}_{\mathrm{surround}}(\mathbf{o},\mathbf{d}_i)$ . The surrounding lighting estimation reconstructs complex environmental lighting, achieving a spatially varying effect and high consistency with the background. + +Blending. We blend the HDR intensity value from the sky-dome and surrounding lighting by transmittance of the final sampling point from McNeRF. The idea is that the rays emitted outside the radiance fields will definitely hit the sky-dome. Given the direction $\mathbf{d}_i$ , we retrieve the skydome's intensity $\mathcal{I}_{\mathrm{skydome}}(\mathbf{d}_i)$ with equirectangular projection. The final HDR light intensity $\mathcal{I}_{\mathrm{env}}(\mathbf{o},\mathbf{d}_i)$ is a combination of scene and skydome: + +$$ +\mathcal {I} _ {\text {e n v}} (\mathbf {o}, \mathbf {d} _ {i}) = \mathcal {I} _ {\text {s u r r o u n d}} (\mathbf {o}, \mathbf {d} _ {i}) + T _ {K} \mathcal {I} _ {\text {s k y d o m e}} (\mathbf {d} _ {i}), +$$ + +where $T_{K}$ is the last sampling point's transmittance. + +McLight offers two main advantages: i) it explicitly recovers the illuminance behavior at the peak and use complementary information from multiple cameras to restore accurate skydome; and ii) it enables location-specific lighting with consideration of complex scene structures. + +# 5. Experimental Results + +# 5.1. Datasets and implementation details + +We demonstrate a variety of results mainly on the Waymo Open Dataset [57], which contains high-quality multicamera images and the corresponding calibrations. For McLight skydome estimation, we collect 449 HDRIs from online HDRI databases for the autoencoder training and use HoliCity [75], a street view panorama dataset for the second stage; see more dataset details in the Appendix. + +In our experiment, we use front, left front, and right front + +![](images/5bf4eea0fd9debc5a80d87e52d41fe9652ea193fc3f60c63d06cedfeb3463bad.jpg) +Front left + +![](images/88b45533493fe22c5aebd9727124043b947a6e88e0bf1c6d7c46ed0766439a47.jpg) +Front + +![](images/f2352fabec77c98e64a729f724c314460506a2f9a7df1db37b598d2cfcf797d3.jpg) +Front right + +![](images/2f3d4818138188034491308613a4e64292983139cbd718eea3df731de29a8591.jpg) + +![](images/8750f1a59336fa996b2a6810a2fbeca36c4caad4b20c74a52315b933319234e8.jpg) + +![](images/24f44ef39b8a2fe213825262032ceda185cf8b6c6479fbd001923d42aea24510.jpg) +Command: "Create a traffic jam." +Figure 6. Editing result under a highly abstract command. + +cameras in each frame. During the rendering process, we choose 40 frames per scene at a $10\mathrm{Hz}$ sampling rate, totaling 120 images. We evenly select $1/8$ of these as the test set, with the remainder used for training. The input images are used at the dataset's initial resolution of $1920 \times 1280$ ; we employ GPT-4 as the LLMs in all of our experiments; see more implementation details in the Appendix. + +# 5.2. System results + +Editing via language commands. We select three representative commands to demonstrate the editing results. All of the results demonstrate we achieve photo-realistic wide angle results, thanks to McNeRF and McLight. + +Mixed and complex command. We send the system with a mixed and complex command, implying that a police car is chasing a wrong-way racer. The target scene, command and the result are shown in Figure 5. We see that i) every requirement in the complex command is accurately executed thanks to our multi-agent collaboration design; ii) this command successfully simulates one rare but dangerous driving condition, which is significant in accident testing. + +Highly abstract command. The second type is a highly abstract command. The inputs and results are presented in Figure 6. We see that i) this highly abstract command is hard to decompose by sentence division but still can be correctly executed by our method, and ii) our 3D asset bank offers a large variety of objects for addition. + +Multi-round command. We also perform a multi-round chat with our system, and the commands in different rounds exist context dependencies. The final results are shown in Figure 7. We see that i) our system is well-equipped to handle multi-round commands and execute the commands in each round precisely; ii) our system can handle the context dependencies across different rounds thanks to the recording + +![](images/e4c5778bba8772af279f5d2bea99dba1dec6a0777e6ce27cd9ba94d467452580.jpg) +Front left + +![](images/bb3473fdec38bcc1dabfc85a69d7790b4777373c0e913c07731f38a4fa8030da.jpg) +Front + +![](images/d608debd70fea97742d077397ad49aca403f5c604ed008564ec6b404a73257f5.jpg) +Front right + +![](images/5f523b0e48638bdba808421b135cb871cbb83d880f9e559cacc082a4923b763e.jpg) + +![](images/3842b951399e78dfbcee77369c218814ea73c4e67cdcbd06488fba96dd74e7cf.jpg) +Command 1: "Ego vehicle drives ahead slowly. Add a car to the close front that is moving ahead." + +![](images/e02f9db9cea4e58a76eb8915bd9f894b2cea9d69eb6a1c0c1ff9462f90dd83c8.jpg) + +![](images/06021c08aa05d33110e195aa01dcdcd49f9bc171a5bbf749b81dc5f46fe7d801.jpg) +Command 2: "Modify the added car to turn left. Add a Chevrolet to the front of the added car. Add another vehicle to the left of the added Mini driving toward me." + +![](images/4cfd2cc41757614acabd5e77c875f05d89aeb79398df87090e2c313eed16bea4.jpg) +Figure 7. Editing result under multi-round commands. +Figure 8. Comparison of detection performance w/o and with our simulated data under different amounts of real data during training. + +![](images/905eb43f45871bb8d325fee54b0b81cd9d099a391d5ba7dd6e269cfc0a09549c.jpg) + +![](images/7ac17b933135773c0bebdac17ec94c71dc4f72a0852e53208171a60d3f2d3368.jpg) + +ability of the project manager agent. + +3D detection with simulation data. We validate the benefits of our simulation as data augmentation for a downstream 3D object detection task on Waymo Open Dataset [58]. We simulate 1960 frames, derived from scenes in the training dataset. In the simulation, cars with various types, locations, and orientations are incorporated. The detection model adopts Lift-Splat [50]. Figure 8 shows detection performances with and without fixed augmentation under various amounts of real data. We see that i) a significant and consistent improvement across different data sizes is achieved; ii) when real data is limited, our simulation notably aids in rough detection (AP30); iii) when the amount of real data increases, our simulation further significantly improves fine-grained detection (AP70), reflecting the high-quality of our simulation. + +# 5.3. Component results + +Multi-agent collaboration. We evaluated the effectiveness of the multi-agent collaboration by checking whether the command is successfully executed in Table 2. In scenarios without multi-agent collaboration, all operations are executed by a single LLM agent. We see that a single LLM agent leads to notably lower execution accuracy across all categories due to process limitations. In contrast, the collaborative multi-agent approach can manage most commands, attributed to its task division and agent role specificity. + +
Multi-agent collaborationLanguage command category
DeletionAdditionView changeRevisionAbstract
✓0.6170.3830.7170.3670.216
0.9830.8670.9670.9170.883
+ +Table 2. The accuracy $(\%)$ of task completion by LLM without and with multi-agent collaboration. + +
MethodsPSNR↑SSIM↑LPIPS↓Inf. time (s)↓
DVGO [56]23.570.7700.5087.7
Mip-NeRF360 [8]24.400.7540.528101.8
S-NeRF [69]24.710.7590.519114.5
F2NeRF [62]23.260.7730.4392.4
Ours w/o alignment23.320.7760.4372.5
Ours w/o exposure25.180.8190.3812.4
McNeRF (Ours)25.820.8220.3782.5
+ +![](images/0b85171da357d5b1be05d0a7ff7e0e46dd219956fc9bb7cc47b073fc4b23473a.jpg) +Figure 9. Comparisons of wide-angle images generation. (a) S-NeRF.(b) F2NeRF. (c) McNeRF (Ours). Last row: target images. + +Background rendering. We compare our McNeRF with several other state-of-the-art methods on the background novel view synthesis task. We perform reconstruction and rendering on 32 selected scenes. Table 3 shows the quantitative results comparison on three metrics: PSNR, SSIM, and LPIPS. We see that i) McNeRF achieves SoTA performance on all three metrics, significantly outperforming other baselines; ii) McNeRF has a fast inference speed, enabling quick responses to user requests for image rendering. + +Figure 9 demonstrates qualitative comparisons between other methods and ours. We see that existing NeRF methods do not consider the exposure time, leading to noticeable changes in brightness at the junctions of different cameras in the image, as well as an overall inconsistency in exposure across the wide-angle view. Our method can make the brightness of the entire image more consistent. + +Foreground rendering. We compare our McLight with the other two SoTA methods [31, 64]. Table 4 shows the comparison of relative intensity(log 10) error on our HDRI dataset, angular error on HoliCity [75], and user study. We see that + +Table 3. Background novel view rendering performance evaluation. + +
MethodPeak Intensity(log10) ErrorPeak Angular Error (deg)User study(%) ↑
Mean ↓Median ↓Mean ↓Median ↓
Hold-Geoffroy et al. [31]0.8990.97548.451.619.5
Wang et al. [64]0.5900.62833.529.437.3
McLight (Ours)0.4490.27032.326.543.1
+ +Table 4. Comparison with previous methods on lighting estimation. + +![](images/040a647041766cdad1cfb156f7c86f9ff6980b2de7daa1aae433af54102c829c.jpg) + +![](images/b7ceb900ce547806bd527b7d1f9d3769b75bce40061dda59456290851f57fd86.jpg) + +![](images/ac2c7f5dc676175ca7c8baea74a45e917bfd0d38ffe62903fd152edac5040bb8.jpg) + +![](images/2b4e959904e821cd521964733264a315c2842aea89eda6db7528caed6bf9c7d2.jpg) + +Figure 10. Comparison with different lighting estimation methods. + +
MethodsStraightLeft TurnRight TurnSpeedWithin-road
GPT2Code0.7380.5590.5360.8930.214
GPT2Motion0.5950.1190.1670.3450.277
Ours0.9880.9400.9760.9521.000
+ +Table 5. Comparison with motion generation from text methods. + +McLight achieves more accurate peak behavior and receives noticeably higher user preferences. Figure 10 shows the visualizations of vehicle insertion. The vehicles added through McLight feature significantly more realistic reflections and strong shadows consistent with the scene. + +Vehicle motion. As shown in Table 5, we compare the motion generation method from user commands with two of our designed baselines: 1. GPT2Motion, which directly uses LLM to return the motion coordinates; 2. GPT2Code, which first generates code using LLM and executes it to obtain the vehicle motion. We validate multiple actions in multiple scenarios and report the user study result. The user study is to determine if the generated motions matched the command intents and fitted with the lane map. We see that our method demonstrated a significant advantage in generating motions from language commands. Additionally, it maintained a high rate of keeping the trajectories within the lane boundaries. + +# 6. Conclusions and Limitations + +This paper introduces ChatSim, the first system for editing 3D driving scene simulations via language commands and realistic rendering with import of external digital assets. To effectively execute user commands, ChatSim adopts an LLM-agent collaboration workflow. To promote photorealistic simulation, we propose McNeRF and McLight for background and foreground rendering, respectively, accommodating multi-camera inputs. Experiments show that ChatSim successfully simulates customized data via language commands, achieving high-quality, photo-realistic outcomes. In future, we plan to integrate more background editing functionalities to ChatSim, such as weather changes. + +# Acknowledgement + +This research is supported by NSFC under Grant 62171276 and the Science and Technology Commission of Shanghai Municipal under Grant 2151100900 and 22DZ2229005. + +# References + +[1] 51Sim-One. https://wdp.51aes.com/news/27.2,3 +[2] LLC Agisoft. Metashape-photogrammetric processing of digital images and 3d spatial data generation, 2019. 5 +[3] Elif Akata, Lion Schulz, Julian Coda-Forno, Seong Joon Oh, Matthias Bethge, and Eric Schulz. Playing repeated games with large language models. arXiv preprint arXiv:2305.16867, 2023. 3 +[4] Alexander Amini, Igor Gilitschenski, Jacob Phillips, Julia Moseyko, Rohan Banerjee, Sertac Karaman, and Daniela Rus. Learning robust control policies for end-to-end autonomous driving from data-driven simulation. IEEE Robotics and Automation Letters, 5(2):1143-1150, 2020. 1 +[5] Rohan Anil, Andrew M Dai, Orhan First, Melvin Johnson, Dmitry Lepikhin, Alexandre Passos, Siamak Shakeri, Emanuel Taropa, Paige Bailey, Zhifeng Chen, et al. Palm 2 technical report. arXiv preprint arXiv:2305.10403, 2023. 3 +[6] Natalie M Banta. Property interests in digital assets: The rise of digital feudalism. Cardozo L. Rev., 38:1099, 2016. 1 +[7] Jonathan T Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P Srinivasan. Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5855–5864, 2021. 2 +[8] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Mip-nerf 360: Unbounded anti-aliased neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5470–5479, 2022. 2, 8 +[9] Luca Bergamini, Yawei Ye, Oliver Scheel, Long Chen, Chih Hu, Luca Del Pero, Băjej Osiński, Hugo Grimmett, and Peter Ondruska. Simnet: Learning reactive self-driving simulations from real-world observations. In 2021 IEEE International Conference on Robotics and Automation (ICRA), pages 5119-5125. IEEE, 2021. 1 +[10] Mark Boss, Varun Jampani, Kihwan Kim, Hendrik Lensch, and Jan Kautz. Two-shot spatially-varying brdf and shape estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3982-3991, 2020. 3 +[11] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. Advances in neural information processing systems, 33:1877-1901, 2020. 3 +[12] Holger Caesar, Varun Bankiti, Alex H Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuscenes: A multimodal dataset for autonomous driving. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11621-11631, 2020. 1 +[13] Ming-Fang Chang, John Lambert, Patsorn Sangkloy, Jagjeet Singh, Slawomir Bak, Andrew Hartnett, De Wang, Peter Carr, Simon Lucey, Deva Ramanan, et al. Argoverse: 3d tracking and forecasting with rich maps. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8748-8757, 2019. 1 + +[14] ChatGPT. https://openai.com/blog/chatgpt.3 +[15] Chenyi Chen, Ari Seff, Alain Kornhauser, and Jianxiong Xiao. Deepdriving: Learning affordance for direct perception in autonomous driving. In Proceedings of the IEEE international conference on computer vision, pages 2722-2730, 2015. 1 +[16] Xiaozhi Chen, Kaustav Kundu, Ziyu Zhang, Huimin Ma, Sanja Fidler, and Raquel Urtasun. Monocular 3d object detection for autonomous driving. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2147-2156, 2016. +[17] Xiaozhi Chen, Huimin Ma, Ji Wan, Bo Li, and Tian Xia. Multi-view 3d object detection network for autonomous driving. In Proceedings of the IEEE conference on Computer Vision and Pattern Recognition, pages 1907-1915, 2017. 1 +[18] Aakanksha Chowdhery, Sharan Narang, Jacob Devlin, Maarten Bosma, Gaurav Mishra, Adam Roberts, Paul Barham, Hyung Won Chung, Charles Sutton, Sebastian Gehrmann, et al. Palm: Scaling language modeling with pathways. arXiv preprint arXiv:2204.02311, 2022. 3 +[19] International Electrotechnical Commission et al. Iec 61966-2-1: 1999 multimedia systems and equipment- colour measurement and management- part 2-1: Colour management-default rgb colour space-srgb, 1999. 5 +[20] Blender Online Community. Blender - a 3D modelling and rendering package. Blender Foundation, Stichting Blender Foundation, Amsterdam, 2018. 5 +[21] Valter Crescenzi, Giansalvatore Mecca, Paolo Merialdo, et al. Roadrunner: Towards automatic data extraction from large web sites. In VLDB, pages 109-118, 2001. 2, 3 +[22] Alexey Dosovitskiy, German Ros, Felipe Codevilla, Antonio Lopez, and Vladlen Koltun. Carla: An open urban driving simulator. In Conference on robot learning, pages 1-16. PMLR, 2017. 2, 3 +[23] Yilun Du, Shuang Li, Antonio Torralba, Joshua B Tenenbaum, and Igor Mordatch. Improving factuality and reasoning in language models through multiagent debate. arXiv preprint arXiv:2305.14325, 2023. 3 +[24] Openscanerio Editor. https://github.com/ebadi/openscenarioeditor. 2,3 +[25] Unreal Engine. https://www.unrealengine.com/. 2 +[26] Ruiyuan Gao, Kai Chen, Enze Xie, Lanqing Hong, Zhenguo Li, Dit-Yan Yeung, and Qiang Xu. Magicdrive: Street view generation with diverse 3d geometry control. arXiv preprint arXiv:2310.02601, 2023. 2, 3 +[27] Mathieu Garon, Kalyan Sunkavalli, Sunil Hadap, Nathan Carr, and Jean-François Lalonde. Fast spatially-varying indoor lighting estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6908-6917, 2019. 3 +[28] Jianfei Guo, Nianchen Deng, Xinyang Li, Yeqi Bai, Botian Shi, Chiyu Wang, Chenjing Ding, Dongliang Wang, and Yikang Li. Streetsurf: Extending multi-view implicit surface reconstruction to street views. arXiv preprint arXiv:2306.04988, 2023. 2 +[29] Rui Hao, Linmei Hu, Weijian Qi, Qingliu Wu, Yirui Zhang, and Liqiang Nie. Chatllm network: More brains, more intelligence. arXiv preprint arXiv:2304.12998, 2023. 3 + +[30] Yannick Hold-Geoffroy, Kalyan Sunkavalli, Sunil Hadap, Emiliano Gambaretto, and Jean-François Lalonde. Deep outdoor illumination estimation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7312–7321, 2017. 3 +[31] Yannick Hold-Geoffroy, Akshaya Athawale, and Jean-François Lalonde. Deep sky modeling for single image outdoor lighting estimation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6927–6935, 2019. 3, 6, 8 +[32] Sirui Hong, Xiawu Zheng, Jonathan Chen, Yuheng Cheng, Ceyao Zhang, Zili Wang, Steven Ka Shing Yau, Zijuan Lin, Liyang Zhou, Chenyu Ran, et al. Metagpt: Meta programming for multi-agent collaborative framework. arXiv preprint arXiv:2308.00352, 2023. 3 +[33] Anthony Hu, Lloyd Russell, Hudson Yeo, Zak Murez, George Fedoseev, Alex Kendall, Jamie Shotton, and Gianluca Corrado. Gaia-1: A generative world model for autonomous driving. arXiv preprint arXiv:2309.17080, 2023. 2, 3 +[34] Abhijit Kundu, Kyle Genova, Xiaoqi Yin, Alireza Fathi, Caroline Pantofaru, Leonidas J Guibas, Andrea Tagliasacchi, Frank Dellaert, and Thomas Funkhouser. Panoptic neural fields: A semantic object-aware neural scene representation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12871-12881, 2022. 2 +[35] Jean-François Lalonde and Iain Matthews. Lighting estimation in outdoor image collections. In 2014 2nd international conference on 3D vision, pages 131-138. IEEE, 2014. 3 +[36] Jean-François Lalonde, Srinivasa G Narasimhan, and Alexei A Efros. What do the sun and the sky tell us about the camera? International Journal of Computer Vision, 88: 24-51, 2010. 3 +[37] Chloe LeGendre, Wan-Chun Ma, Graham Fyffe, John Flynn, Laurent Charbonnel, Jay Busch, and Paul Debevec. Deeplight: Learning illumination for unconstrained mobile mixed reality. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5918-5928, 2019. 3 +[38] Peixuan Li, Huaici Zhao, Pengfei Liu, and Feidao Cao. Rtm3d: Real-time monocular 3d detection from object keypoints for autonomous driving. In European Conference on Computer Vision, pages 644-660. Springer, 2020. 1 +[39] Xiaofan Li, Yifu Zhang, and Xiaoqing Ye. Drivingdiffusion: Layout-guided multi-view driving scene video generation with latent diffusion model. arXiv preprint arXiv:2310.07771, 2023. 2, 3 +[40] Zhengqin Li, Zexiang Xu, Ravi Ramamoorthi, Kalyan Sunkavalli, and Manmohan Chandraker. Learning to reconstruct shape and spatially-varying reflectance from a single image. ACM Transactions on Graphics (TOG), 37(6):1-11, 2018. 3 +[41] Zhuopeng Li, Lu Li, and Jianke Zhu. Read: Large-scale neural scene rendering for autonomous driving. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1522-1529, 2023. 2, 3 +[42] Zhijian Liu, Haotian Tang, Alexander Amini, Xinyu Yang, Huizi Mao, Daniela L Rus, and Song Han. Bevfusion: Multi-task multi-sensor fusion with unified bird's-eye view represent + +tation. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 2774-2781. IEEE, 2023. 4 +[43] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 2 +[44] Milan Miric, Kevin J Boudreau, and Lars Bo Jeppesen. Protecting their digital assets: The use of formal & informal appropriability strategies by app developers. Research Policy, 48(8):103738, 2019. 1 +[45] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM Transactions on Graphics (ToG), 41(4):1-15, 2022. 2 +[46] OpenAI. Gpt-4 technical report. 2023. 3 +[47] Julian Ost, Fahim Mannan, Nils Thuerey, Julian Knodt, and Felix Heide. Neural scene graphs for dynamic scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2856-2865, 2021. 2, 3 +[48] Julian Ost, Issam Laradji, Alejandro Newell, Yuval Bahat, and Felix Heide. Neural point light fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18419-18429, 2022. 2, 3 +[49] Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, et al. Training language models to follow instructions with human feedback. Advances in Neural Information Processing Systems, 35:27730-27744, 2022. 3 +[50] Jonah Philion and Sanja Fidler. Lift, splat, shoot: Encoding images from arbitrary camera rigs by implicitly unprojecting to 3d. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XIV 16, pages 194-210. Springer, 2020. 7 +[51] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models, 2021. 4 +[52] Guodong Rong, Byung Hyun Shin, Hadi Tabatabaee, Qiang Lu, Steve Lemke, Martins Mozeiko, Eric Boise, Geehoon Uhm, Mark Gerow, Shalin Mehta, et al. Lgsvl simulator: A high fidelity simulator for autonomous driving. In 2020 IEEE 23rd International conference on intelligent transportation systems (ITSC), pages 1-6. IEEE, 2020. 1 +[53] Soumyadip Sengupta, Jinwei Gu, Kihwan Kim, Guilin Liu, David W Jacobs, and Jan Kautz. Neural inverse rendering of an indoor scene from a single image. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8598-8607, 2019. 3 +[54] Shital Shah, Debadeepta Dey, Chris Lovett, and Ashish Kapoor. Airsim: High-fidelity visual and physical simulation for autonomous vehicles. In Field and Service Robotics: Results of the 11th International Conference, pages 621-635. Springer, 2018. 2, 3 +[55] Gowri Somanath and Daniel Kurz. Hdr environment map estimation for real-time augmented reality. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11298-11306, 2021. 3 + +[56] Cheng Sun, Min Sun, and Hwann-Tzong Chen. Direct voxel grid optimization: Super-fast convergence for radiance fields reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5459-5469, 2022. 2, 8 +[57] Pei Sun, Henrik Kretzschmar, Xerxes Dotiwalla, Aurelien Chouard, Vijaysai Patnaik, Paul Tsui, James Guo, Yin Zhou, Yuning Chai, Benjamin Caine, Vijay Vasudevan, Wei Han, Jiquan Ngiam, Hang Zhao, Aleksei Timofeev, Scott Ettinger, Maxim Krivokon, Amy Gao, Aditya Joshi, Yu Zhang, Jonathon Shlens, Zhifeng Chen, and Dragomir Anguelov. Scalability in perception for autonomous driving: Waymo open dataset. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 6 +[58] Pei Sun, Henrik Kretzschmar, Xerxes Dotiwalla, Aurelien Chouard, Vijaysai Patnaik, Paul Tsui, James Guo, Yin Zhou, Yuning Chai, Benjamin Caine, et al. Scalability in perception for autonomous driving: Waymo open dataset. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2446-2454, 2020. 1, 7 +[59] Alexander Swerdlow, Runsheng Xu, and Bolei Zhou. Street-view image generation from a bird's-eye view layout. arXiv preprint arXiv:2301.04634, 2023. 2, 3 +[60] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288, 2023. 3 +[61] Haithem Turki, Jason Y Zhang, Francesco Ferroni, and Deva Ramanan. Suds: Scalable urban dynamic scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12375-12385, 2023. 2 +[62] Peng Wang, Yuan Liu, Zhaoxi Chen, Lingjie Liu, Ziwei Liu, Taku Komura, Christian Theobalt, and Wenping Wang. F2-nerf: Fast neural radiance field training with free camera trajectories. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4150–4159, 2023. 2, 5, 8 +[63] Xiaofeng Wang, Zheng Zhu, Guan Huang, Xinze Chen, and Jiwen Lu. Drivedreamer: Towards real-world-driven world models for autonomous driving. arXiv preprint arXiv:2309.09777, 2023. 2, 3 +[64] Zian Wang, Wenzheng Chen, David Acuna, Jan Kautz, and Sanja Fidler. Neural light field estimation for street scenes with differentiable virtual object insertion. In European Conference on Computer Vision, pages 380-397. Springer, 2022. 3, 6, 8 +[65] Zhenhailong Wang, Shaoguang Mao, Wenshan Wu, Tao Ge, Furu Wei, and Heng Ji. Unleashing cognitive synergy in large language models: A task-solving agent through multi-personal self-collaboration. arXiv preprint arXiv:2307.05300, 2023. 3 +[66] Bichen Wu, Forrest Iandola, Peter H Jin, and Kurt Keutzer. Squeezedet: Unified, small, low power fully convolutional neural networks for real-time object detection for autonomous driving. In Proceedings of the IEEE conference on computer vision and pattern recognition workshops, pages 129-137, 2017. 1 + +[67] Qingyun Wu, Gagan Bansal, Jieyu Zhang, Yiran Wu, Shaokun Zhang, Erkang Zhu, Beibin Li, Li Jiang, Xiaoyun Zhang, and Chi Wang. Autogen: Enabling next-gen llm applications via multi-agent conversation framework. arXiv preprint arXiv:2308.08155, 2023. 3 +[68] Zirui Wu, Tianyu Liu, Liyi Luo, Zhide Zhong, Jianteng Chen, Hongmin Xiao, Chao Hou, Haozhe Lou, Yuantao Chen, Runyi Yang, Yuxin Huang, Xiaoyu Ye, Zike Yan, Yongliang Shi, Yiyi Liao, and Hao Zhao. Mars: An instance-aware, modular and realistic simulator for autonomous driving. CICAI, 2023. 2, 3 +[69] Ziyang Xie, Junge Zhang, Wenye Li, Feihu Zhang, and Li Zhang. S-nerf: Neural radiance fields for street views. arXiv preprint arXiv:2303.00749, 2023. 2, 3, 8 +[70] Yinda Xu and Lidong Yu. Drl-based trajectory tracking for motion-related modules in autonomous driving. arXiv preprint arXiv:2308.15991, 2023. 4 +[71] Kairui Yang, Enhui Ma, Jibin Peng, Qing Guo, Di Lin, and Kaicheng Yu. Bevcontrol: Accurately controlling street-view elements with multi-perspective consistency via bev sketch layout. arXiv preprint arXiv:2308.01661, 2023. 2, 3 +[72] Zhenpei Yang, Yuning Chai, Dragomir Anguelov, Yin Zhou, Pei Sun, Dumitru Erhan, Sean Rafferty, and Henrik Kretzschmar. Surfelgan: Synthesizing realistic sensor data for autonomous driving. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11118-11127, 2020. 1 +[73] Ze Yang, Yun Chen, Jingkang Wang, Sivabalan Manivasagam, Wei-Chiu Ma, Anqi Joyce Yang, and Raquel Urtasun. Unisim: A neural closed-loop sensor simulator. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1389-1399, 2023. 2, 3 +[74] Jinsong Zhang, Kalyan Sunkavalli, Yannick Hold-Geoffroy, Sunil Hadap, Jonathan Eisenman, and Jean-François Lalonde. All-weather deep outdoor lighting estimation. In Proceedings of the IEEE/CVF conference on Computer Vision and Pattern Recognition, pages 10158–10166, 2019. 3 +[75] Yichao Zhou, Jingwei Huang, Xili Dai, Linjie Luo, Zhili Chen, and Yi Ma. HoliCity: A city-scale data platform for learning holistic 3D structures. 2020. arXiv:2008.03286 [cs.CV]. 6, 8 +[76] Mingchen Zhuge, Haozhe Liu, Francesco Faccio, Dylan R Ashley, Robert Csordás, Anand Gopalakrishnan, Abdullah Hamdi, Hasan Abed Al Kader Hammoud, Vincent Herrmann, Kazuki Irie, et al. Mindstorms in natural language-based societies of mind. arXiv preprint arXiv:2305.17066, 2023. 3 \ No newline at end of file diff --git a/2024/Editable Scene Simulation for Autonomous Driving via Collaborative LLM-Agents/images.zip b/2024/Editable Scene Simulation for Autonomous Driving via Collaborative LLM-Agents/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..397303ad51a60b0e8592be26886dacc093abe60b --- /dev/null +++ b/2024/Editable Scene Simulation for Autonomous Driving via Collaborative LLM-Agents/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:cd3483d97b195d619c6d8b5860ccd99f93092a76f09c123d0ee102d975655873 +size 679971 diff --git a/2024/Editable Scene Simulation for Autonomous Driving via Collaborative LLM-Agents/layout.json b/2024/Editable Scene Simulation for Autonomous Driving via Collaborative LLM-Agents/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..ad481e654bf1accc0bd225601e5799bcfdd9327e --- /dev/null +++ b/2024/Editable Scene Simulation for Autonomous Driving via Collaborative LLM-Agents/layout.json @@ -0,0 +1,10204 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 48, + 103, + 545, + 120 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 103, + 545, + 120 + ], + "spans": [ + { + "bbox": [ + 48, + 103, + 545, + 120 + ], + "type": "text", + "content": "Editable Scene Simulation for Autonomous Driving via Collaborative LLM-Agents" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 124, + 141, + 470, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 124, + 141, + 470, + 185 + ], + "spans": [ + { + "bbox": [ + 124, + 141, + 470, + 185 + ], + "type": "text", + "content": "Yuxi Wei" + }, + { + "bbox": [ + 124, + 141, + 470, + 185 + ], + "type": "inline_equation", + "content": "^{1,5*}" + }, + { + "bbox": [ + 124, + 141, + 470, + 185 + ], + "type": "text", + "content": " Zi Wang" + }, + { + "bbox": [ + 124, + 141, + 470, + 185 + ], + "type": "inline_equation", + "content": "^{3,5*}" + }, + { + "bbox": [ + 124, + 141, + 470, + 185 + ], + "type": "text", + "content": " Yifan Lu" + }, + { + "bbox": [ + 124, + 141, + 470, + 185 + ], + "type": "inline_equation", + "content": "^{1,5*}" + }, + { + "bbox": [ + 124, + 141, + 470, + 185 + ], + "type": "text", + "content": " Chenxin Xu" + }, + { + "bbox": [ + 124, + 141, + 470, + 185 + ], + "type": "inline_equation", + "content": "^{1,5*}" + }, + { + "bbox": [ + 124, + 141, + 470, + 185 + ], + "type": "text", + "content": " Changxing Liu" + }, + { + "bbox": [ + 124, + 141, + 470, + 185 + ], + "type": "inline_equation", + "content": "^{1,5}" + }, + { + "bbox": [ + 124, + 141, + 470, + 185 + ], + "type": "text", + "content": " Hao Zhao" + }, + { + "bbox": [ + 124, + 141, + 470, + 185 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 124, + 141, + 470, + 185 + ], + "type": "text", + "content": " Siheng Chen" + }, + { + "bbox": [ + 124, + 141, + 470, + 185 + ], + "type": "inline_equation", + "content": "^{1,2,5\\dagger}" + }, + { + "bbox": [ + 124, + 141, + 470, + 185 + ], + "type": "text", + "content": " Yanfeng Wang" + }, + { + "bbox": [ + 124, + 141, + 470, + 185 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 124, + 141, + 470, + 185 + ], + "type": "text", + "content": " Shanghai Jiao Tong University" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 85, + 186, + 506, + 241 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 85, + 186, + 506, + 241 + ], + "spans": [ + { + "bbox": [ + 85, + 186, + 506, + 241 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 85, + 186, + 506, + 241 + ], + "type": "text", + "content": " Carnegie Mellon University " + }, + { + "bbox": [ + 85, + 186, + 506, + 241 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 85, + 186, + 506, + 241 + ], + "type": "text", + "content": " Tsinghua University " + }, + { + "bbox": [ + 85, + 186, + 506, + 241 + ], + "type": "inline_equation", + "content": "^{5}" + }, + { + "bbox": [ + 85, + 186, + 506, + 241 + ], + "type": "text", + "content": " Multi-Agent Governance & Intelligence Crew (MAGIC) {wyx3590236732, yifan_lu, xcxwakaka, cx-liu}@sjtu.edu.cn {sihengc, wangyanfeng}@sjtu.edu.cn ziwang2@andrew.cmu.edu zhaohao@air.tsinghua.edu.cn" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 143, + 269, + 192, + 281 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 269, + 192, + 281 + ], + "spans": [ + { + "bbox": [ + 143, + 269, + 192, + 281 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 293, + 290, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 293, + 290, + 545 + ], + "spans": [ + { + "bbox": [ + 47, + 293, + 290, + 545 + ], + "type": "text", + "content": "Scene simulation in autonomous driving has gained significant attention because of its huge potential for generating customized data. However, existing editable scene simulation approaches face limitations in terms of user interaction efficiency, multi-camera photo-realistic rendering and external digital assets integration. To address these challenges, this paper introduces ChatSim, the first system that enables editable photo-realistic 3D driving scene simulations via natural language commands with external digital assets. To enable editing with high command flexibility, ChatSim leverages a large language model (LLM) agent collaboration framework. To generate photo-realistic outcomes, ChatSim employs a novel multi-camera neural radiance field method. Furthermore, to unleash the potential of extensive high-quality digital assets, ChatSim employs a novel multi-camera lighting estimation method to achieve scene-consistent assets' rendering. Our experiments on Waymo Open Dataset demonstrate that ChatSim can handle complex language commands and generate corresponding photo-realistic scene videos. Code can be accessed at: https://github.com/yifanlu0227/ChatSim." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 568, + 128, + 580 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 568, + 128, + 580 + ], + "spans": [ + { + "bbox": [ + 48, + 568, + 128, + 580 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 587, + 288, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 587, + 288, + 696 + ], + "spans": [ + { + "bbox": [ + 47, + 587, + 288, + 696 + ], + "type": "text", + "content": "Perception [12, 15-17, 38, 66] is the window of an autonomous vehicle into the external environment. To ensure the robustness of the vehicle's perceptual capabilities during both training and testing phases, it necessitates the collection of high-quality perception data in substantial volumes [13, 58]. However, the operation of a fleet for the acquisition of real-world data often incurs prohibitive expenses, particularly for specialized or customized requirements. For instance, in the aftermath of an accident or intervention in" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 317, + 269, + 535, + 417 + ], + "blocks": [ + { + "bbox": [ + 317, + 269, + 535, + 417 + ], + "lines": [ + { + "bbox": [ + 317, + 269, + 535, + 417 + ], + "spans": [ + { + "bbox": [ + 317, + 269, + 535, + 417 + ], + "type": "image", + "image_path": "369a5d248f75a13a44d4444719f784db88d394a33e5dda862aad944811a3c72c.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 306, + 418, + 545, + 440 + ], + "lines": [ + { + "bbox": [ + 306, + 418, + 545, + 440 + ], + "spans": [ + { + "bbox": [ + 306, + 418, + 545, + 440 + ], + "type": "text", + "content": "Figure 1. ChatSim enables the editing of photo-realistic 3D driving scene simulations via language commands." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 447, + 547, + 543 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 447, + 547, + 543 + ], + "spans": [ + { + "bbox": [ + 304, + 447, + 547, + 543 + ], + "type": "text", + "content": "volving an autonomous vehicle, it is imperative to test the vehicle's perception system across a spectrum of similar scenarios. While replicating such scenario data from real-world instances is nearly impossible due to the uncontrollability of actual scenes [9, 52], customized scene simulation emerges as a vital and feasible alternative. It enables the precise modeling of specific conditions without high costs and logistical complexities of real-world data collection [4, 72]." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 544, + 547, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 544, + 547, + 688 + ], + "spans": [ + { + "bbox": [ + 304, + 544, + 547, + 688 + ], + "type": "text", + "content": "To effectively simulate customized driving scenes, we identify three key properties as fundamental. First, the simulation should be capable of following sophisticated or abstract demands, thereby facilitating the production. Second, the simulation should generate photo-realistic, view-consistent outcomes, which allow for the closest approximation to vehicle observations in real-world scenarios. Third, it should allow for the integration of external digital assets [6, 44] with their photo-realistic textures and materials while fitting the lighting conditions. This capability would unlock the potential for data expansion by incorporating a wide array of external digital assets, satisfying customized needs." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 689, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 547, + 713 + ], + "type": "text", + "content": "A vast array of significant works have been proposed for scene simulation, yet they fail to meet all three of" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 702, + 126, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 702, + 126, + 713 + ], + "spans": [ + { + "bbox": [ + 58, + 702, + 126, + 713 + ], + "type": "text", + "content": "*Equal contribution." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 143, + 703, + 219, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 703, + 219, + 713 + ], + "spans": [ + { + "bbox": [ + 143, + 703, + 219, + 713 + ], + "type": "text", + "content": "† Corresponding author." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15077" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 289, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 289, + 312 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 289, + 312 + ], + "type": "text", + "content": "these requirements. Traditional graphics engines, such as CARLA [22] and UE [25], offer editable virtual environments with external digital assets, but the data realism is restricted by asset modeling and rendering qualities. Image generation based methods, such as BEVControl [71], DriveDreamer [63], MagicDrive [26], can generate realistic scene images based on various control signals, including BEV maps, bounding boxes and camera poses. However, they struggle to maintain view consistency and face challenges in importing external digital assets due to the absence of 3D spatial modeling. Rendering-based methods have been proposed to obtain photo-realistic and view-consistent scene simulation. Notable examples like UniSim [73] and MARS [68] come equipped with a suite of scene-editing tools. However, these systems require extensive user involvement in every trivial editing step via code implementation, which is ineffective when performing the editing. Furthermore, while they handle vehicles in observed scenarios effectively, their inability to support external digital assets restricts opportunities for data expansion and customization." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 314, + 289, + 397 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 314, + 289, + 397 + ], + "spans": [ + { + "bbox": [ + 47, + 314, + 289, + 397 + ], + "type": "text", + "content": "To fulfill the identified requirements, we introduce ChatSim, the first system that enables editable photo-realistic 3D driving scene simulations via natural language commands with external digital assets. To use ChatSim, users simply engage in a conversation with the system, issuing commands through natural language without any involvement in intermediate simulation steps; see Figure 1 for illustration." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 400, + 289, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 400, + 289, + 581 + ], + "spans": [ + { + "bbox": [ + 47, + 400, + 289, + 581 + ], + "type": "text", + "content": "To address complex or abstract user commands effectively, ChatSim adopts a large language model (LLM)-based multi-agent collaboration framework. The key idea is to exploit multiple LLM agents, each with a specialized role, to decouple an overall simulation demand into specific editing tasks, thereby mirroring the task division and execution typically founded in the workflow of a human-operated company. This workflow offers two key advantages for scene simulation. First, LLM agents' ability to process human language commands allows for intuitive and dynamic editing of complex driving scenes, enabling precise adjustments and feedback. Second, the collaboration framework enhances simulation efficiency and accuracy by distributing specific editing tasks among specialized agents, ensuring detailed and realistic simulations with improved task completion rates." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 582, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 582, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 47, + 582, + 289, + 715 + ], + "type": "text", + "content": "To generate photo-realistic outcomes, we propose McNeRF in ChatSim, a novel neural radiance field method that incorporates multi-camera inputs, offering a broader scene rendering. This integration fully exploits camera setups on vehicles but raises two significant challenges: camera pose misalignment due to asynchronizated trigger times and brightness inconsistency due to different camera exposure times. To address camera pose misalignment, McNeRF uses a multi-camera alignment to reduce extrinsic parameter noises, ensuring rendering quality. To address brightness inconsistency, McNeRF integrates the critical exposure times" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 304, + 72, + 547, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 108 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 108 + ], + "type": "text", + "content": "to recover scene radiance in HDR, markedly mitigating the issue of color discrepancies at the intersections of two camera images with different exposure times." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 109, + 548, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 109, + 548, + 228 + ], + "spans": [ + { + "bbox": [ + 304, + 109, + 548, + 228 + ], + "type": "text", + "content": "To import external digital assets with their realistic textures and materials, we propose McLight, a novel multicamera lighting estimation that blends skydome and surrounding lighting. Our skydome estimation restores accurate sun behavior with peak intensity residual connection, enabling the rendering of prominent shadows. For surrounding lighting, McLight queries McNeRF to achieve complex location-specific illumination effects, like those in the tree shade with sunlight being blocked. This significantly improves the rendering realism of the integrated 3D assets." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 229, + 548, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 229, + 548, + 361 + ], + "spans": [ + { + "bbox": [ + 304, + 229, + 548, + 361 + ], + "type": "text", + "content": "We conduct extensive experiments on the Waymo autonomous driving dataset and show that ChatSim generates photo-realistic customized perception data including dangerous corner cases according to various human language commands. Our method is compatible with mixed, highly abstract and multi-round commands. Our method achieves SoTA performance with an improvement of " + }, + { + "bbox": [ + 304, + 229, + 548, + 361 + ], + "type": "inline_equation", + "content": "4.5\\%" + }, + { + "bbox": [ + 304, + 229, + 548, + 361 + ], + "type": "text", + "content": " in photorealism with a wide-angle rendering. Moreover, we demonstrate our lighting estimation outperforms the SoTA methods both qualitatively and quantitatively, reducing the intensity error and angular error by " + }, + { + "bbox": [ + 304, + 229, + 548, + 361 + ], + "type": "inline_equation", + "content": "57.0\\%" + }, + { + "bbox": [ + 304, + 229, + 548, + 361 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 229, + 548, + 361 + ], + "type": "inline_equation", + "content": "9.9\\%" + }, + { + "bbox": [ + 304, + 229, + 548, + 361 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 373, + 392, + 385 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 373, + 392, + 385 + ], + "spans": [ + { + "bbox": [ + 306, + 373, + 392, + 385 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 391, + 548, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 391, + 548, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 391, + 548, + 715 + ], + "type": "text", + "content": "Scene simulation for autonomous driving. Current scene simulation methods can be generally divided into three categories: graphics engines, image generation, and scene rendering. Graphics engines, such as CARLA [22], AirSim [54], OpenScenario Editor [24], 51Sim-One [1] and RoadRunner [21], create a virtual world for simulating a wide range of driving scenarios. However, there exists a significant domain gap between the virtual world and reality. Image generation methods can generate realistic scene images based on different control signals, such as HD maps [26, 39, 59], sketch layout [71], bounding boxes [26, 39, 63], text [26, 33, 39, 63] and driving actions [33, 63]. However, these approaches can hardly maintain scene consistency. To obtain a coherent driving scene, methods based on scene rendering target to reconstruct the 3D scene. READ [41] employs point clouds and uses a U-Net to render images. With the rapid development of Neural Radiance Field (NeRF) [7, 8, 43, 45, 56, 62], several works [28, 34, 47, 48, 61, 68, 69, 73] also exploit NeRFs to model cars and static street backgrounds in outdoor environments. Moreover, notable examples like UniSim [73] and MARS [68] come equipped with a suite of scene-editing tools. However, these methods require extensive user involvement in intermediate editing steps and they fail to support external digital assets for data expansion. In this work, we propose ChatSim that achieves automatic simulation editing via language commands and integrates external digital assets to enhance realism and flexibility. In ChatSim," + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "15078" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 49, + 70, + 285, + 232 + ], + "blocks": [ + { + "bbox": [ + 49, + 70, + 285, + 232 + ], + "lines": [ + { + "bbox": [ + 49, + 70, + 285, + 232 + ], + "spans": [ + { + "bbox": [ + 49, + 70, + 285, + 232 + ], + "type": "table", + "html": "
MethodPhoto-realisticDim.Multi-cameraEditableExternal assetsLanguageOpen-source
CARLA [22]×3D✓✓✓×✓
AirSim [54]×3D✓✓✓×✓
OpenScenario [24]×3D✓✓✓×✓
51Sim-One [1]×3D✓✓✓××
RoadRunner [21]×3D✓✓✓×✓
BEVGen [59]✓2D✓✓××✓
BEVCentral [71]✓2D✓✓×××
DriveDreamer [63]✓2D✓✓×✓×
DrivingDiffusion [39]✓2D✓✓×✓×
GAIA-1 [33]✓2D×✓×✓×
MagicDrive [26]✓2D✓✓×××
READ [41]✓3D××××✓
Neural SG [47]✓3D×✓××✓
Neural PLF [48]✓3D××××✓
S-NeRF [69]✓3D✓×××✓
UniSim [73]✓3D×✓×××
MARS [68]✓3D×✓××✓
ChatSim (Ours)✓3D✓✓✓✓✓
", + "image_path": "639f5bd6b6538d953ac0ba6d42130ee9ad15ee7a8a14ec3dae28a98f0020d135.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 234, + 288, + 257 + ], + "lines": [ + { + "bbox": [ + 47, + 234, + 288, + 257 + ], + "spans": [ + { + "bbox": [ + 47, + 234, + 288, + 257 + ], + "type": "text", + "content": "Table 1. Comparison of existing and proposed methods for autonomous driving simulation." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 262, + 287, + 285 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 262, + 287, + 285 + ], + "spans": [ + { + "bbox": [ + 46, + 262, + 287, + 285 + ], + "type": "text", + "content": "we integrate McNeRF, a novel neural radiance field designed to leverage multi-camera inputs for high-fidelity rendering." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 286, + 288, + 454 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 286, + 288, + 454 + ], + "spans": [ + { + "bbox": [ + 46, + 286, + 288, + 454 + ], + "type": "text", + "content": "Lighting estimation. Lighting estimation focuses on assessing the illumination conditions of a real-world environment to seamlessly integrate digital objects. Early methods [35, 36] for outdoor environments use explicit cues like detected shadows on the ground. Recent works usually adopt learning-based approaches [27, 30, 31, 37, 40, 74] by predicting different lighting representations like spherical lobes [10, 40], light probes [37], environment map [53, 55], HDR sky model [31, 64, 74] and lighting volume [64]. However, few of them consider multi-camera input, which is common for driving scenarios. In this paper, we propose a novel multi-camera lighting estimation method, McLight, combining with our McNeRF, to estimate a wider range of lighting and obtain the spatially-varying lighting effects of assets." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 455, + 288, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 455, + 288, + 647 + ], + "spans": [ + { + "bbox": [ + 46, + 455, + 288, + 647 + ], + "type": "text", + "content": "Large Language model and collaborative framework. Large Language Models (LLMs) are AI systems trained on extensive data to understand, generate, and respond to human language. GPT [11] is a pioneering work to generate human-like content. The following updated versions GPT-3.5 [14] and GPT-4 [46], provide more intelligent capabilities like chatting, browsing and coding. Notable other large language models include InstructGPT [49], LLaMA [60] and PaLM [5, 18]. Based on LLM, many works [3, 23, 29, 65, 76] improve the problem-solving abilities by integrating communication among multiple agents. [32] and [67] define a group of well-organized agents to form operating procedures with conversation and code programming. In this paper, we exploit the power of collaborative LLM agents in simulation for autonomous driving, enabling the various editing of 3D scenes via language commands." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 659, + 261, + 673 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 659, + 261, + 673 + ], + "spans": [ + { + "bbox": [ + 47, + 659, + 261, + 673 + ], + "type": "text", + "content": "3. Collaborative LLM-Agents for Editing" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 677, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 677, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 47, + 677, + 288, + 715 + ], + "type": "text", + "content": "The ChatSim system analyzes specific user commands and returns a video that meets customized needs; see Figure 2. Since user commands could be abstract and sophisticated," + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 72, + 547, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 133 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 133 + ], + "type": "text", + "content": "it requires the system to have flexible task-handling ability. Directly applying a single LLM agent struggles with multi-step reasoning and cross-referencing. To address this, we design a series of collaborative LLM agents, where each agent is responsible for a unique aspect of the editing task." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 142, + 468, + 156 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 142, + 468, + 156 + ], + "spans": [ + { + "bbox": [ + 305, + 142, + 468, + 156 + ], + "type": "text", + "content": "3.1. Specific Agent's Functionality" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 159, + 547, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 159, + 547, + 327 + ], + "spans": [ + { + "bbox": [ + 304, + 159, + 547, + 327 + ], + "type": "text", + "content": "Agents in ChatSim comprise two key components: a Large Language Model (LLM) and the corresponding role functions. The LLM is responsible for understanding the received commands while the role functions process the received data. Each agent is equipped with unique LLM prompts and role functions tailored to their specific duties within the system. To accomplish their tasks, agents first convert the received commands to a structured configuration using LLM with the assistance of prompts. Then the role functions utilize the structured configuration as parameters to process the received data and produce the desired outcomes; see an agent example on the right side of Figure 2. This workflow endows agents with both language interpretation capabilities and precise execution capabilities." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 327, + 547, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 327, + 547, + 460 + ], + "spans": [ + { + "bbox": [ + 304, + 327, + 547, + 460 + ], + "type": "text", + "content": "Project Manager Agent. The project manager agent decomposes direct commands into clear natural language instructions dispatched to other editing agents. To equip the project manager agent with the capability of command decomposition, we design a series of prompts for its LLM. The core idea of the prompts is to describe the action set, give the overall goal, and define the output form with examples; The role functions send the decomposed instructions to other agents for editing. The presence of the project manager agent enhances the system's robustness in interpreting various inputs and streamlines operations for clarity and fine granularity." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 460, + 547, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 460, + 547, + 556 + ], + "spans": [ + { + "bbox": [ + 304, + 460, + 547, + 556 + ], + "type": "text", + "content": "Tech agent for view adjustment. The view adjustment agent generates suitable extrinsic camera parameters. The LLM in the agent translates the natural language instructions for viewpoint adjustment into movement parameters to the target viewpoint's position and angle. In role functions, these movement parameters are turned into transformation matrices required by the extrinsic, which are then multiplied by the original parameters to yield a new viewpoint." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 557, + 547, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 557, + 547, + 664 + ], + "spans": [ + { + "bbox": [ + 304, + 557, + 547, + 664 + ], + "type": "text", + "content": "Tech agent for background rendering. The background rendering agent renders the scene background based on multi-camera images. The LLM receives the rendering command and then operates the role functions for rendering. Notably, in role functions, we specifically integrate a novel neural radiance field method (McNeRF) taking multi-camera inputs and considering exposure time, solving the problem of blurring and brightness inconsistency in multi-camera rendering, see more details in Section 4.1." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "type": "text", + "content": "Tech agent for vehicle deleting. The vehicle deleting agent removes specified vehicles from the background. It first identifies current vehicle attributes like 3D bounding boxes and colors from given scene information or results from a scene" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "15079" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 61, + 72, + 531, + 256 + ], + "blocks": [ + { + "bbox": [ + 61, + 72, + 531, + 256 + ], + "lines": [ + { + "bbox": [ + 61, + 72, + 531, + 256 + ], + "spans": [ + { + "bbox": [ + 61, + 72, + 531, + 256 + ], + "type": "image", + "image_path": "05f9bf54c389c3fc12db037c655fa35ff99434483de4cc028ef5f24bef591de3.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 258, + 545, + 281 + ], + "lines": [ + { + "bbox": [ + 46, + 258, + 545, + 281 + ], + "spans": [ + { + "bbox": [ + 46, + 258, + 545, + 281 + ], + "type": "text", + "content": "Figure 2. ChatSim system overview. The system exploit multiple collaborative LLM agents with specialized roles to decouple an overall demand into specific editing tasks. Each agent equips an LLM and corresponding role functions to interpret and execute its specific tasks." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 53, + 289, + 279, + 384 + ], + "blocks": [ + { + "bbox": [ + 53, + 289, + 279, + 384 + ], + "lines": [ + { + "bbox": [ + 53, + 289, + 279, + 384 + ], + "spans": [ + { + "bbox": [ + 53, + 289, + 279, + 384 + ], + "type": "image", + "image_path": "3332b5392e5399f26ee735fa931df462f0f7f160fd44cf61038e0dcca3811bb1.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 70, + 384, + 262, + 396 + ], + "lines": [ + { + "bbox": [ + 70, + 384, + 262, + 396 + ], + "spans": [ + { + "bbox": [ + 70, + 384, + 262, + 396 + ], + "type": "text", + "content": "Figure 3. Prompt example of view adjustment agent." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 399, + 288, + 459 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 399, + 288, + 459 + ], + "spans": [ + { + "bbox": [ + 46, + 399, + 288, + 459 + ], + "type": "text", + "content": "perception model like [42]. The LLM gathers attributes of the vehicles and performs matching with user requests. Upon confirming the targeted vehicles, it employs a per-frame inpainting model as the role functions, such as latent diffusion methods [51], to effectively delete them from the scene." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 461, + 288, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 461, + 288, + 568 + ], + "spans": [ + { + "bbox": [ + 46, + 461, + 288, + 568 + ], + "type": "text", + "content": "Tech agent for 3D asset management. The 3D asset management agent selects and modifies 3D digital assets according to user specifications. It constructs and maintains a 3D digital asset bank; see our bank details in the Appendix. To facilitate the addition of various objects, the agent first uses LLM to select the most suitable asset by key attributes matching with the requirements, such as color and type. If the matching is not perfect, the agent could modify the asset through its role functions like changing the color." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 570, + 289, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 570, + 289, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 570, + 289, + 714 + ], + "type": "text", + "content": "Tech agent for vehicle motion. The vehicle motion agent creates the initial places and subsequent motions of vehicles following the requests. Existing vehicle motion generation methods cannot directly generate motion purely from text and the scene map. To solve the problem, here we propose a novel text-to-motion method. The key idea is linking a placement and planning module as role functions with LLMs to extract and turn motion attributes into coordinates. Motion attributes include position attributes (e.g., distance, direction) and movement attributes (e.g., speed, action). For the placement module, we endow each lane node in the lane map with its attributes to match with the position attributes. The" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 290, + 547, + 338 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 290, + 547, + 338 + ], + "spans": [ + { + "bbox": [ + 304, + 290, + 547, + 338 + ], + "type": "text", + "content": "planning module plans the vehicle's approximate destination lane node and then plans the intermediate trajectory by fitting the Bezier curves. We also add trajectory tracking [70] to fit vehicle dynamics; see more details in the Appendix." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 338, + 546, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 338, + 546, + 445 + ], + "spans": [ + { + "bbox": [ + 304, + 338, + 546, + 445 + ], + "type": "text", + "content": "Tech agent for foreground rendering. The foreground rendering agent integrates camera extrinsic information, 3D assets, and motion information to render foreground objects in the scene. Notably, to seamlessly integrate the external assets with the current scene, we design a multi-camera lighting estimation method (McLight) into the role functions, coupling with McNeRF. The estimated illumination is then utilized by Blender API to generate foreground images. The detailed technical aspects will be elaborated in Section 4.2." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 453, + 473, + 466 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 453, + 473, + 466 + ], + "spans": [ + { + "bbox": [ + 306, + 453, + 473, + 466 + ], + "type": "text", + "content": "3.2. Agent Collaboration Workflow" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 470, + 547, + 661 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 470, + 547, + 661 + ], + "spans": [ + { + "bbox": [ + 304, + 470, + 547, + 661 + ], + "type": "text", + "content": "Agents with tailored functions collaboratively work together to edit based on user commands. The project manager orchestrates and dispatches instructions to editing agents. The editing agents form two teams: background generation and foreground generation. For background generation, the background rendering agent generates rendered images using the extrinsic parameters from the view adjustment agent, followed by inpainting by the vehicle deleting agent. For foreground generation, the foreground rendering agent renders the images using the extrinsic parameters from the view adjustment agent, selected 3D assets from 3D asset management agent, and generated motions from vehicle motion agent. Finally, the foreground and background images are composed to create and deliver a video to the user. The editing information in each agent's configuration is recorded by the project manager agent for possible multi-round editorings." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 672, + 455, + 685 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 672, + 455, + 685 + ], + "spans": [ + { + "bbox": [ + 306, + 672, + 455, + 685 + ], + "type": "text", + "content": "4. Novel Rendering Methods" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 689, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 547, + 714 + ], + "type": "text", + "content": "Based on the collaborative LLM agents framework introduced in Section 3, this section presents two novel rendering" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15080" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 193 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 193 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 193 + ], + "type": "text", + "content": "techniques to enhance photo-realism in simulations. To tackle the rendering challenges caused by multiple cameras, we propose multi-camera neural radiance field (McNeRF), a novel NeRF model considering the varied camera exposure times for visual consistency. To render realistic external digital assets with location-specific lighting and accurate shadows, we propose McLight, a novel hybrid lighting estimation method combined with our McNeRF. Note that McNeRF and McLight are leveraged by the background rendering agent and the foreground rendering agent, respectively." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 206, + 242, + 220 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 206, + 242, + 220 + ], + "spans": [ + { + "bbox": [ + 47, + 206, + 242, + 220 + ], + "type": "text", + "content": "4.1. McNeRF for Background Rendering" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 228, + 287, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 228, + 287, + 323 + ], + "spans": [ + { + "bbox": [ + 46, + 228, + 287, + 323 + ], + "type": "text", + "content": "An autonomous vehicle typically equips multiple cameras to achieve a comprehensive perception view. However, this poses challenges for NeRF training due to the misaligned multi-camera poses from asynchronized camera trigger times and the brightness inconsistency originating from different exposure times. To address these challenges, the proposed McNeRF uses two techniques: multi-camera alignment and brightness-consistent rendering." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 326, + 287, + 410 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 326, + 287, + 410 + ], + "spans": [ + { + "bbox": [ + 46, + 326, + 287, + 410 + ], + "type": "text", + "content": "Multi-camera alignment. Autonomous vehicles, despite having a localization module for accurate camera poses, face challenges with asynchronous trigger times across multiple cameras. To align camera extrinsics for NeRF training, our core idea is to leverage a consistent spatial coordinate system provided by Agisoft Metashape [2] to align the images captured by multiple cameras at different timestamps." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 411, + 288, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 411, + 288, + 472 + ], + "spans": [ + { + "bbox": [ + 46, + 411, + 288, + 472 + ], + "type": "text", + "content": "Specifically, let " + }, + { + "bbox": [ + 46, + 411, + 288, + 472 + ], + "type": "inline_equation", + "content": "\\mathcal{I}^{(i,k)}" + }, + { + "bbox": [ + 46, + 411, + 288, + 472 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 411, + 288, + 472 + ], + "type": "inline_equation", + "content": "\\xi^{(i,k)}" + }, + { + "bbox": [ + 46, + 411, + 288, + 472 + ], + "type": "text", + "content": " be the image captured by the " + }, + { + "bbox": [ + 46, + 411, + 288, + 472 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 411, + 288, + 472 + ], + "type": "text", + "content": "th camera at the " + }, + { + "bbox": [ + 46, + 411, + 288, + 472 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 411, + 288, + 472 + ], + "type": "text", + "content": "th trigger and the corresponding camera pose in the vehicle's global coordinate space, respectively. We first input all images into Metashape for recalibration. The aligned camera pose is then obtained as:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 117, + 472, + 215, + 488 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 117, + 472, + 215, + 488 + ], + "spans": [ + { + "bbox": [ + 117, + 472, + 215, + 488 + ], + "type": "interline_equation", + "content": "\\widehat {\\xi} ^ {(i, k)} = T _ {M \\to G} \\cdot \\xi_ {M} ^ {(i, k)},", + "image_path": "a714ee26cf25d20bafaa63d804d740fecc750081e8142eaf600bf75058fa59aa.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 491, + 287, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 491, + 287, + 601 + ], + "spans": [ + { + "bbox": [ + 46, + 491, + 287, + 601 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 491, + 287, + 601 + ], + "type": "inline_equation", + "content": "\\xi_{M}^{(i,k)}" + }, + { + "bbox": [ + 46, + 491, + 287, + 601 + ], + "type": "text", + "content": " denotes the recalibrated camera pose in the Metashape's unified spatial coordinate space, and " + }, + { + "bbox": [ + 46, + 491, + 287, + 601 + ], + "type": "inline_equation", + "content": "T_{M\\rightarrow G}" + }, + { + "bbox": [ + 46, + 491, + 287, + 601 + ], + "type": "text", + "content": " is the transformation from the Metashape's coordinate space to the vehicle's global coordinate space. After alignment, the pose noise can be significantly reduced. Then, the aligned camera pose " + }, + { + "bbox": [ + 46, + 491, + 287, + 601 + ], + "type": "inline_equation", + "content": "\\widehat{\\xi}^{(i,t)}" + }, + { + "bbox": [ + 46, + 491, + 287, + 601 + ], + "type": "text", + "content": " can be used to generate the origins and directions of rays for McNeRF, enabling high-fidelity rendering. The aligned pose can also facilitate the foreground rendering agent's operations." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 603, + 288, + 674 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 603, + 288, + 674 + ], + "spans": [ + { + "bbox": [ + 46, + 603, + 288, + 674 + ], + "type": "text", + "content": "Brightness-consistent rendering. The exposure times of cameras can differ substantially, causing significant brightness differences across images, hindering the NeRF training. As shown in Figure 4, McNeRF, addresses this by incorporating exposure times into HDR radiance fields, prompting brightness consistency." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 677, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 288, + 714 + ], + "type": "text", + "content": "We adopt F2-NeRF [62] as our backbone model to handle the unbounded scene, sampling " + }, + { + "bbox": [ + 46, + 677, + 288, + 714 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 46, + 677, + 288, + 714 + ], + "type": "text", + "content": " points along the ray " + }, + { + "bbox": [ + 46, + 677, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{r}" + }, + { + "bbox": [ + 46, + 677, + 288, + 714 + ], + "type": "text", + "content": " and estimating each point's HDR radiance " + }, + { + "bbox": [ + 46, + 677, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{e}_k" + }, + { + "bbox": [ + 46, + 677, + 288, + 714 + ], + "type": "text", + "content": " and density " + }, + { + "bbox": [ + 46, + 677, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\sigma_k" + }, + { + "bbox": [ + 46, + 677, + 288, + 714 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 309, + 72, + 541, + 144 + ], + "blocks": [ + { + "bbox": [ + 309, + 72, + 541, + 144 + ], + "lines": [ + { + "bbox": [ + 309, + 72, + 541, + 144 + ], + "spans": [ + { + "bbox": [ + 309, + 72, + 541, + 144 + ], + "type": "image", + "image_path": "024a3e27b985c79138121fbda6995993109f49d829e053842b6a1b358f7f57b6.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 309, + 144, + 541, + 223 + ], + "blocks": [ + { + "bbox": [ + 309, + 144, + 541, + 223 + ], + "lines": [ + { + "bbox": [ + 309, + 144, + 541, + 223 + ], + "spans": [ + { + "bbox": [ + 309, + 144, + 541, + 223 + ], + "type": "image", + "image_path": "7f9a739e4c4e43f357b42161a700e97bab79b2aa9ba7680c886f357178d0a0a1.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 224, + 546, + 280 + ], + "lines": [ + { + "bbox": [ + 304, + 224, + 546, + 280 + ], + "spans": [ + { + "bbox": [ + 304, + 224, + 546, + 280 + ], + "type": "text", + "content": "Figure 4. Rendering framework. The main components include McNeRF and McLight. Background rendering uses McNeRF to predict HDR pixel value and convert it to LDR with sRGB OETF. McLight includes a skydome lighting estimation network and adopts McNeRF to generate surrounding lighting." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 285, + 491, + 296 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 285, + 491, + 296 + ], + "spans": [ + { + "bbox": [ + 305, + 285, + 491, + 296 + ], + "type": "text", + "content": "The HDR light intensity is then calculated as:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 356, + 297, + 545, + 330 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 356, + 297, + 545, + 330 + ], + "spans": [ + { + "bbox": [ + 356, + 297, + 545, + 330 + ], + "type": "interline_equation", + "content": "\\widehat {\\mathcal {I}} _ {\\mathrm {H D R}} (\\mathbf {r}) = f (\\Delta t) \\cdot \\sum_ {k = 1} ^ {K} T _ {k} \\alpha_ {k} \\mathbf {e} _ {k}, \\tag {1}", + "image_path": "3f9663e5f20d7892ee73fc2830ac8ff04d165941cb56ca5b043eccc6f93a8165.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 331, + 546, + 416 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 331, + 546, + 416 + ], + "spans": [ + { + "bbox": [ + 304, + 331, + 546, + 416 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 331, + 546, + 416 + ], + "type": "inline_equation", + "content": "\\alpha_{k} = 1 - \\exp (-\\sigma_{k}\\delta_{i})" + }, + { + "bbox": [ + 304, + 331, + 546, + 416 + ], + "type": "text", + "content": " is the opacity, " + }, + { + "bbox": [ + 304, + 331, + 546, + 416 + ], + "type": "inline_equation", + "content": "\\delta_{i}" + }, + { + "bbox": [ + 304, + 331, + 546, + 416 + ], + "type": "text", + "content": " is the point sampling interval, " + }, + { + "bbox": [ + 304, + 331, + 546, + 416 + ], + "type": "inline_equation", + "content": "T_{k} = \\prod_{i = 0}^{k - 1}(1 - \\alpha_{i})" + }, + { + "bbox": [ + 304, + 331, + 546, + 416 + ], + "type": "text", + "content": " is the accumulated transmittance and " + }, + { + "bbox": [ + 304, + 331, + 546, + 416 + ], + "type": "inline_equation", + "content": "\\Delta t" + }, + { + "bbox": [ + 304, + 331, + 546, + 416 + ], + "type": "text", + "content": " is the exposure time. The normalization function " + }, + { + "bbox": [ + 304, + 331, + 546, + 416 + ], + "type": "inline_equation", + "content": "f(\\Delta t) = 1 + \\epsilon (\\Delta t - \\mu) / \\sigma" + }, + { + "bbox": [ + 304, + 331, + 546, + 416 + ], + "type": "text", + "content": " is designed to stabilize training, where " + }, + { + "bbox": [ + 304, + 331, + 546, + 416 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 304, + 331, + 546, + 416 + ], + "type": "text", + "content": " is a hyperparameter for scaling, " + }, + { + "bbox": [ + 304, + 331, + 546, + 416 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 304, + 331, + 546, + 416 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 331, + 546, + 416 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 304, + 331, + 546, + 416 + ], + "type": "text", + "content": " are the mean and standard deviation of the exposure times of all images, respectively." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 417, + 547, + 500 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 417, + 547, + 500 + ], + "spans": [ + { + "bbox": [ + 304, + 417, + 547, + 500 + ], + "type": "text", + "content": "By predicting scene radiance in HDR and multiplying it by the exposure time, we recover the light intensity received by the sensor and tackle the inconsistent color supervision at the intersections of two camera images with distinct exposure times. Moreover, the HDR light intensity outputted by McNeRF can provide scene-level illumination for foreground object rendering, a topic further discussed in Section 4.2." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 500, + 547, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 500, + 547, + 548 + ], + "spans": [ + { + "bbox": [ + 304, + 500, + 547, + 548 + ], + "type": "text", + "content": "To train the rendering network, we enforce the consistency of radiance between the rendered image (prediction) and the captured image (ground-truth). Given the ground-truth image " + }, + { + "bbox": [ + 304, + 500, + 547, + 548 + ], + "type": "inline_equation", + "content": "\\mathcal{I}" + }, + { + "bbox": [ + 304, + 500, + 547, + 548 + ], + "type": "text", + "content": ", the loss function is then:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 331, + 549, + 518, + 578 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 331, + 549, + 518, + 578 + ], + "spans": [ + { + "bbox": [ + 331, + 549, + 518, + 578 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\frac {1}{| R |} \\sum_ {\\mathbf {r} \\in R} \\left(\\mathrm {O E T F} \\left(\\widehat {\\mathcal {I}} _ {\\mathrm {H D R}} (\\mathbf {r})\\right) - \\mathcal {I} (\\mathbf {r})\\right) ^ {2},", + "image_path": "d07a5104b37118351e38317d3bc347cafa98a5ddce134920c285df0db66f8b03.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 578, + 546, + 615 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 578, + 546, + 615 + ], + "spans": [ + { + "bbox": [ + 304, + 578, + 546, + 615 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 578, + 546, + 615 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 304, + 578, + 546, + 615 + ], + "type": "text", + "content": " represents the ray set and " + }, + { + "bbox": [ + 304, + 578, + 546, + 615 + ], + "type": "inline_equation", + "content": "\\mathrm{OETF}(\\cdot)" + }, + { + "bbox": [ + 304, + 578, + 546, + 615 + ], + "type": "text", + "content": " is the sRGB opto-electronic transfer function (gamma correction) [19] that converts HDR light intensity to LDR colors." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 305, + 623, + 496, + 636 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 623, + 496, + 636 + ], + "spans": [ + { + "bbox": [ + 305, + 623, + 496, + 636 + ], + "type": "text", + "content": "4.2. McLight for Foreground Rendering" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 641, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 641, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 641, + 547, + 713 + ], + "type": "text", + "content": "To enrich the scene's content with substantial digital 3D assets, we employ Blender [20] foreground virtual objects' rendering. A seamless insertion critically depends on accurately estimating the scene's illumination conditions. Thus, as shown in Figure 4, we propose McLight, a novel hybrid lighting estimation consisting of skydome lighting and sur" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15081" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 122, + 84 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 122, + 84 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 122, + 84 + ], + "type": "text", + "content": "rounding lighting." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 84, + 288, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 84, + 288, + 167 + ], + "spans": [ + { + "bbox": [ + 46, + 84, + 288, + 167 + ], + "type": "text", + "content": "Skydome lighting estimation. Estimating skydome lighting from images is challenging for restoring accurate sun behavior. To achieve this, we propose a novel residual connection from the estimated peak intensity to the HDR reconstruction to address over-smoothing output. Further, we adopt a self-attention mechanism to fuse multi-camera inputs, capturing complementary visual cues." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 168, + 287, + 347 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 168, + 287, + 347 + ], + "spans": [ + { + "bbox": [ + 46, + 168, + 287, + 347 + ], + "type": "text", + "content": "Here we employ a two-stage process. In the first stage, we train an autoencoder to reconstruct the corresponding HDR panorama from an LDR panorama. Following [64], the encoder transforms the LDR skydome panorama into three intermediate vectors, including the peak direction vector " + }, + { + "bbox": [ + 46, + 168, + 287, + 347 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_{\\mathrm{dir}} \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 46, + 168, + 287, + 347 + ], + "type": "text", + "content": ", the intensity vector " + }, + { + "bbox": [ + 46, + 168, + 287, + 347 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_{\\mathrm{int}} \\in \\mathbb{R}_+^3" + }, + { + "bbox": [ + 46, + 168, + 287, + 347 + ], + "type": "text", + "content": ", and the sky content vector " + }, + { + "bbox": [ + 46, + 168, + 287, + 347 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_{\\mathrm{content}} \\in \\mathbb{R}^{64}" + }, + { + "bbox": [ + 46, + 168, + 287, + 347 + ], + "type": "text", + "content": ". However, as HDR intensity behaves like an impulse response at its peak position, with pixel values thousands of times higher than its neighbors, it is difficult for the decoder to recover such patterns. To tackle this, we design a residual connection that injects " + }, + { + "bbox": [ + 46, + 168, + 287, + 347 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_{\\mathrm{int}}" + }, + { + "bbox": [ + 46, + 168, + 287, + 347 + ], + "type": "text", + "content": " into the decoded HDR panorama with a spherical Gaussian lobe attenuation. This explicitly restores the peak intensity of the sun in the reconstructed HDR panorama, allowing us to render strong shadows for virtual objects." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 348, + 288, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 348, + 288, + 510 + ], + "spans": [ + { + "bbox": [ + 46, + 348, + 288, + 510 + ], + "type": "text", + "content": "In the second stage, we train an image encoder and a multi-camera fusion module built upon the pretrained decoder from the first stage. Specifically, for images from each camera, a shared image encoder predicts the peak direction vector " + }, + { + "bbox": [ + 46, + 348, + 288, + 510 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_{\\mathrm{dir}}^{(i)}" + }, + { + "bbox": [ + 46, + 348, + 288, + 510 + ], + "type": "text", + "content": ", the intensity vector " + }, + { + "bbox": [ + 46, + 348, + 288, + 510 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_{\\mathrm{int}}^{(i)}" + }, + { + "bbox": [ + 46, + 348, + 288, + 510 + ], + "type": "text", + "content": ", and the sky content vector " + }, + { + "bbox": [ + 46, + 348, + 288, + 510 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_{\\mathrm{content}}^{(i)}" + }, + { + "bbox": [ + 46, + 348, + 288, + 510 + ], + "type": "text", + "content": " for each image " + }, + { + "bbox": [ + 46, + 348, + 288, + 510 + ], + "type": "inline_equation", + "content": "\\mathcal{I}^{(i)}" + }, + { + "bbox": [ + 46, + 348, + 288, + 510 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 348, + 288, + 510 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 348, + 288, + 510 + ], + "type": "text", + "content": " is the camera index. We design the latent vector fusion across the multiple camera views as follows: all " + }, + { + "bbox": [ + 46, + 348, + 288, + 510 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_{\\mathrm{dir}}^{(i)}" + }, + { + "bbox": [ + 46, + 348, + 288, + 510 + ], + "type": "text", + "content": " are aligned to the front-facing view using their extrinsic parameters and averaged to form " + }, + { + "bbox": [ + 46, + 348, + 288, + 510 + ], + "type": "inline_equation", + "content": "\\bar{\\mathbf{f}}_{\\mathrm{dir}}" + }, + { + "bbox": [ + 46, + 348, + 288, + 510 + ], + "type": "text", + "content": "; all " + }, + { + "bbox": [ + 46, + 348, + 288, + 510 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_{\\mathrm{int}}^{(i)}" + }, + { + "bbox": [ + 46, + 348, + 288, + 510 + ], + "type": "text", + "content": " are averaged to yield " + }, + { + "bbox": [ + 46, + 348, + 288, + 510 + ], + "type": "inline_equation", + "content": "\\bar{\\mathbf{f}}_{\\mathrm{int}}" + }, + { + "bbox": [ + 46, + 348, + 288, + 510 + ], + "type": "text", + "content": "; all " + }, + { + "bbox": [ + 46, + 348, + 288, + 510 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_{\\mathrm{content}}^{(i)}" + }, + { + "bbox": [ + 46, + 348, + 288, + 510 + ], + "type": "text", + "content": " are integrated into " + }, + { + "bbox": [ + 46, + 348, + 288, + 510 + ], + "type": "inline_equation", + "content": "\\bar{\\mathbf{f}}_{\\mathrm{content}}" + }, + { + "bbox": [ + 46, + 348, + 288, + 510 + ], + "type": "text", + "content": " through a self-attention module. Finally, the pretrained decoder reconstructs the HDR skydome image " + }, + { + "bbox": [ + 46, + 348, + 288, + 510 + ], + "type": "inline_equation", + "content": "\\mathcal{I}_{\\mathrm{skydome}}" + }, + { + "bbox": [ + 46, + 348, + 288, + 510 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 46, + 348, + 288, + 510 + ], + "type": "inline_equation", + "content": "\\bar{\\mathbf{f}}_{\\mathrm{dir}}" + }, + { + "bbox": [ + 46, + 348, + 288, + 510 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 348, + 288, + 510 + ], + "type": "inline_equation", + "content": "\\bar{\\mathbf{f}}_{\\mathrm{int}}" + }, + { + "bbox": [ + 46, + 348, + 288, + 510 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 348, + 288, + 510 + ], + "type": "inline_equation", + "content": "\\bar{\\mathbf{f}}_{\\mathrm{content}}" + }, + { + "bbox": [ + 46, + 348, + 288, + 510 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 510, + 288, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 510, + 288, + 569 + ], + "spans": [ + { + "bbox": [ + 46, + 510, + 288, + 569 + ], + "type": "text", + "content": "Compared to alternative approaches [31, 64], our multicamera sky dome estimation technique accurately reproduces the sun's intensity response behavior at its peak with our residual designs, significantly improving the accuracy and fidelity of the skydome reconstruction." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "type": "text", + "content": "Surrounding lighting estimation. Merely modeling the sky-dome cannot replicate the complex location-specific lighting effects, like those in the shade with sunlight blocked by trees or buildings. Our McNeRF is capable of storing precise 3D scene information, enabling us to capture the surrounding scene's impact on lighting. This approach facilitates the achievement of spatially-varying lighting estimation. Specifically, we sample the hemisphere rays at the virtual object's position " + }, + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{o}" + }, + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "type": "text", + "content": ". The rays' directions, " + }, + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{d}_i" + }, + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "type": "inline_equation", + "content": "i = 0,1,\\dots,h\\times w" + }, + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "type": "text", + "content": ", are aligned with pixel coordinates on a unit sphere using equirectangular projection from an environment map, where " + }, + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "type": "text", + "content": " are map's height and width. With the ray" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 316, + 79, + 386, + 125 + ], + "blocks": [ + { + "bbox": [ + 335, + 70, + 367, + 78 + ], + "lines": [ + { + "bbox": [ + 335, + 70, + 367, + 78 + ], + "spans": [ + { + "bbox": [ + 335, + 70, + 367, + 78 + ], + "type": "text", + "content": "Front left" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 316, + 79, + 386, + 125 + ], + "lines": [ + { + "bbox": [ + 316, + 79, + 386, + 125 + ], + "spans": [ + { + "bbox": [ + 316, + 79, + 386, + 125 + ], + "type": "image", + "image_path": "22006f29ef30b26d3218bf91ba48a7fddcd5eba050b07813c70aee86e212003d.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 390, + 78, + 460, + 125 + ], + "blocks": [ + { + "bbox": [ + 416, + 70, + 435, + 78 + ], + "lines": [ + { + "bbox": [ + 416, + 70, + 435, + 78 + ], + "spans": [ + { + "bbox": [ + 416, + 70, + 435, + 78 + ], + "type": "text", + "content": "Front" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 390, + 78, + 460, + 125 + ], + "lines": [ + { + "bbox": [ + 390, + 78, + 460, + 125 + ], + "spans": [ + { + "bbox": [ + 390, + 78, + 460, + 125 + ], + "type": "image", + "image_path": "911ac196eb6cba8fdf316d43be5a36f0d9a14ef7a2bec755b9cb0886e9ea1c90.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 465, + 78, + 534, + 125 + ], + "blocks": [ + { + "bbox": [ + 481, + 70, + 517, + 78 + ], + "lines": [ + { + "bbox": [ + 481, + 70, + 517, + 78 + ], + "spans": [ + { + "bbox": [ + 481, + 70, + 517, + 78 + ], + "type": "text", + "content": "Front right" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 465, + 78, + 534, + 125 + ], + "lines": [ + { + "bbox": [ + 465, + 78, + 534, + 125 + ], + "spans": [ + { + "bbox": [ + 465, + 78, + 534, + 125 + ], + "type": "image", + "image_path": "e19903211766eff71f549d3d33153a6bfdca3335d06d8ebd131424b6936e5b46.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 316, + 126, + 534, + 272 + ], + "blocks": [ + { + "bbox": [ + 316, + 126, + 534, + 272 + ], + "lines": [ + { + "bbox": [ + 316, + 126, + 534, + 272 + ], + "spans": [ + { + "bbox": [ + 316, + 126, + 534, + 272 + ], + "type": "image", + "image_path": "b091b38369abab6351eb8e2f4c0ef52abc36ffaf6f7c01c133713e6b629af4c6.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 317, + 272, + 529, + 308 + ], + "lines": [ + { + "bbox": [ + 317, + 272, + 529, + 308 + ], + "spans": [ + { + "bbox": [ + 317, + 272, + 529, + 308 + ], + "type": "text", + "content": "Command: \"Remove all cars in the scene and add a Porsche driving the wrong way toward me fast. Additionally, add a police car also driving the wrong way and chasing behind the Porsche. The view should be moved 5 meters ahead and 0.5 meters above.\"" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 312, + 308, + 539, + 319 + ], + "lines": [ + { + "bbox": [ + 312, + 308, + 539, + 319 + ], + "spans": [ + { + "bbox": [ + 312, + 308, + 539, + 319 + ], + "type": "text", + "content": "Figure 5. Editing result under a complex and mixed command." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 327, + 547, + 386 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 327, + 547, + 386 + ], + "spans": [ + { + "bbox": [ + 305, + 327, + 547, + 386 + ], + "type": "inline_equation", + "content": "\\mathbf{r} = \\mathbf{o} + t\\mathbf{d}_i" + }, + { + "bbox": [ + 305, + 327, + 547, + 386 + ], + "type": "text", + "content": ", we query our McNeRF as Equation 1 to obtain HDR surrounding lighting " + }, + { + "bbox": [ + 305, + 327, + 547, + 386 + ], + "type": "inline_equation", + "content": "\\mathcal{I}_{\\mathrm{surround}}(\\mathbf{o},\\mathbf{d}_i)" + }, + { + "bbox": [ + 305, + 327, + 547, + 386 + ], + "type": "text", + "content": ". The surrounding lighting estimation reconstructs complex environmental lighting, achieving a spatially varying effect and high consistency with the background." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 305, + 387, + 547, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 387, + 547, + 483 + ], + "spans": [ + { + "bbox": [ + 305, + 387, + 547, + 483 + ], + "type": "text", + "content": "Blending. We blend the HDR intensity value from the sky-dome and surrounding lighting by transmittance of the final sampling point from McNeRF. The idea is that the rays emitted outside the radiance fields will definitely hit the sky-dome. Given the direction " + }, + { + "bbox": [ + 305, + 387, + 547, + 483 + ], + "type": "inline_equation", + "content": "\\mathbf{d}_i" + }, + { + "bbox": [ + 305, + 387, + 547, + 483 + ], + "type": "text", + "content": ", we retrieve the skydome's intensity " + }, + { + "bbox": [ + 305, + 387, + 547, + 483 + ], + "type": "inline_equation", + "content": "\\mathcal{I}_{\\mathrm{skydome}}(\\mathbf{d}_i)" + }, + { + "bbox": [ + 305, + 387, + 547, + 483 + ], + "type": "text", + "content": " with equirectangular projection. The final HDR light intensity " + }, + { + "bbox": [ + 305, + 387, + 547, + 483 + ], + "type": "inline_equation", + "content": "\\mathcal{I}_{\\mathrm{env}}(\\mathbf{o},\\mathbf{d}_i)" + }, + { + "bbox": [ + 305, + 387, + 547, + 483 + ], + "type": "text", + "content": " is a combination of scene and skydome:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 323, + 484, + 527, + 498 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 323, + 484, + 527, + 498 + ], + "spans": [ + { + "bbox": [ + 323, + 484, + 527, + 498 + ], + "type": "interline_equation", + "content": "\\mathcal {I} _ {\\text {e n v}} (\\mathbf {o}, \\mathbf {d} _ {i}) = \\mathcal {I} _ {\\text {s u r r o u n d}} (\\mathbf {o}, \\mathbf {d} _ {i}) + T _ {K} \\mathcal {I} _ {\\text {s k y d o m e}} (\\mathbf {d} _ {i}),", + "image_path": "beae71591b5a98d6cdcd0bbf5d295db63ef7c433d1c148ddda1a67891a742596.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 306, + 498, + 516, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 498, + 516, + 510 + ], + "spans": [ + { + "bbox": [ + 306, + 498, + 516, + 510 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 306, + 498, + 516, + 510 + ], + "type": "inline_equation", + "content": "T_{K}" + }, + { + "bbox": [ + 306, + 498, + 516, + 510 + ], + "type": "text", + "content": " is the last sampling point's transmittance." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 305, + 511, + 547, + 571 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 511, + 547, + 571 + ], + "spans": [ + { + "bbox": [ + 305, + 511, + 547, + 571 + ], + "type": "text", + "content": "McLight offers two main advantages: i) it explicitly recovers the illuminance behavior at the peak and use complementary information from multiple cameras to restore accurate skydome; and ii) it enables location-specific lighting with consideration of complex scene structures." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 306, + 582, + 432, + 595 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 582, + 432, + 595 + ], + "spans": [ + { + "bbox": [ + 306, + 582, + 432, + 595 + ], + "type": "text", + "content": "5. Experimental Results" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 306, + 601, + 498, + 614 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 601, + 498, + 614 + ], + "spans": [ + { + "bbox": [ + 306, + 601, + 498, + 614 + ], + "type": "text", + "content": "5.1. Datasets and implementation details" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 305, + 617, + 547, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 617, + 547, + 700 + ], + "spans": [ + { + "bbox": [ + 305, + 617, + 547, + 700 + ], + "type": "text", + "content": "We demonstrate a variety of results mainly on the Waymo Open Dataset [57], which contains high-quality multicamera images and the corresponding calibrations. For McLight skydome estimation, we collect 449 HDRIs from online HDRI databases for the autoencoder training and use HoliCity [75], a street view panorama dataset for the second stage; see more dataset details in the Appendix." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 701, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 701, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 317, + 701, + 545, + 713 + ], + "type": "text", + "content": "In our experiment, we use front, left front, and right front" + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15082" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 77, + 127, + 125 + ], + "blocks": [ + { + "bbox": [ + 77, + 70, + 107, + 77 + ], + "lines": [ + { + "bbox": [ + 77, + 70, + 107, + 77 + ], + "spans": [ + { + "bbox": [ + 77, + 70, + 107, + 77 + ], + "type": "text", + "content": "Front left" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 57, + 77, + 127, + 125 + ], + "lines": [ + { + "bbox": [ + 57, + 77, + 127, + 125 + ], + "spans": [ + { + "bbox": [ + 57, + 77, + 127, + 125 + ], + "type": "image", + "image_path": "5bf4eea0fd9debc5a80d87e52d41fe9652ea193fc3f60c63d06cedfeb3463bad.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 130, + 77, + 201, + 125 + ], + "blocks": [ + { + "bbox": [ + 157, + 70, + 176, + 77 + ], + "lines": [ + { + "bbox": [ + 157, + 70, + 176, + 77 + ], + "spans": [ + { + "bbox": [ + 157, + 70, + 176, + 77 + ], + "type": "text", + "content": "Front" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 77, + 201, + 125 + ], + "lines": [ + { + "bbox": [ + 130, + 77, + 201, + 125 + ], + "spans": [ + { + "bbox": [ + 130, + 77, + 201, + 125 + ], + "type": "image", + "image_path": "88b45533493fe22c5aebd9727124043b947a6e88e0bf1c6d7c46ed0766439a47.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 205, + 77, + 276, + 125 + ], + "blocks": [ + { + "bbox": [ + 222, + 70, + 258, + 77 + ], + "lines": [ + { + "bbox": [ + 222, + 70, + 258, + 77 + ], + "spans": [ + { + "bbox": [ + 222, + 70, + 258, + 77 + ], + "type": "text", + "content": "Front right" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 205, + 77, + 276, + 125 + ], + "lines": [ + { + "bbox": [ + 205, + 77, + 276, + 125 + ], + "spans": [ + { + "bbox": [ + 205, + 77, + 276, + 125 + ], + "type": "image", + "image_path": "f2352fabec77c98e64a729f724c314460506a2f9a7df1db37b598d2cfcf797d3.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 57, + 125, + 276, + 171 + ], + "blocks": [ + { + "bbox": [ + 57, + 125, + 276, + 171 + ], + "lines": [ + { + "bbox": [ + 57, + 125, + 276, + 171 + ], + "spans": [ + { + "bbox": [ + 57, + 125, + 276, + 171 + ], + "type": "image", + "image_path": "2f3d4818138188034491308613a4e64292983139cbd718eea3df731de29a8591.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 57, + 172, + 276, + 220 + ], + "blocks": [ + { + "bbox": [ + 57, + 172, + 276, + 220 + ], + "lines": [ + { + "bbox": [ + 57, + 172, + 276, + 220 + ], + "spans": [ + { + "bbox": [ + 57, + 172, + 276, + 220 + ], + "type": "image", + "image_path": "8750f1a59336fa996b2a6810a2fbeca36c4caad4b20c74a52315b933319234e8.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 57, + 222, + 276, + 270 + ], + "blocks": [ + { + "bbox": [ + 57, + 222, + 276, + 270 + ], + "lines": [ + { + "bbox": [ + 57, + 222, + 276, + 270 + ], + "spans": [ + { + "bbox": [ + 57, + 222, + 276, + 270 + ], + "type": "image", + "image_path": "24f44ef39b8a2fe213825262032ceda185cf8b6c6479fbd001923d42aea24510.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 59, + 270, + 173, + 280 + ], + "lines": [ + { + "bbox": [ + 59, + 270, + 173, + 280 + ], + "spans": [ + { + "bbox": [ + 59, + 270, + 173, + 280 + ], + "type": "text", + "content": "Command: \"Create a traffic jam.\"" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 61, + 281, + 272, + 292 + ], + "lines": [ + { + "bbox": [ + 61, + 281, + 272, + 292 + ], + "spans": [ + { + "bbox": [ + 61, + 281, + 272, + 292 + ], + "type": "text", + "content": "Figure 6. Editing result under a highly abstract command." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 302, + 287, + 387 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 302, + 287, + 387 + ], + "spans": [ + { + "bbox": [ + 46, + 302, + 287, + 387 + ], + "type": "text", + "content": "cameras in each frame. During the rendering process, we choose 40 frames per scene at a " + }, + { + "bbox": [ + 46, + 302, + 287, + 387 + ], + "type": "inline_equation", + "content": "10\\mathrm{Hz}" + }, + { + "bbox": [ + 46, + 302, + 287, + 387 + ], + "type": "text", + "content": " sampling rate, totaling 120 images. We evenly select " + }, + { + "bbox": [ + 46, + 302, + 287, + 387 + ], + "type": "inline_equation", + "content": "1/8" + }, + { + "bbox": [ + 46, + 302, + 287, + 387 + ], + "type": "text", + "content": " of these as the test set, with the remainder used for training. The input images are used at the dataset's initial resolution of " + }, + { + "bbox": [ + 46, + 302, + 287, + 387 + ], + "type": "inline_equation", + "content": "1920 \\times 1280" + }, + { + "bbox": [ + 46, + 302, + 287, + 387 + ], + "type": "text", + "content": "; we employ GPT-4 as the LLMs in all of our experiments; see more implementation details in the Appendix." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 47, + 395, + 137, + 407 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 395, + 137, + 407 + ], + "spans": [ + { + "bbox": [ + 47, + 395, + 137, + 407 + ], + "type": "text", + "content": "5.2. System results" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 46, + 413, + 287, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 413, + 287, + 460 + ], + "spans": [ + { + "bbox": [ + 46, + 413, + 287, + 460 + ], + "type": "text", + "content": "Editing via language commands. We select three representative commands to demonstrate the editing results. All of the results demonstrate we achieve photo-realistic wide angle results, thanks to McNeRF and McLight." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 46, + 461, + 287, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 461, + 287, + 557 + ], + "spans": [ + { + "bbox": [ + 46, + 461, + 287, + 557 + ], + "type": "text", + "content": "Mixed and complex command. We send the system with a mixed and complex command, implying that a police car is chasing a wrong-way racer. The target scene, command and the result are shown in Figure 5. We see that i) every requirement in the complex command is accurately executed thanks to our multi-agent collaboration design; ii) this command successfully simulates one rare but dangerous driving condition, which is significant in accident testing." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 46, + 558, + 287, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 558, + 287, + 629 + ], + "spans": [ + { + "bbox": [ + 46, + 558, + 287, + 629 + ], + "type": "text", + "content": "Highly abstract command. The second type is a highly abstract command. The inputs and results are presented in Figure 6. We see that i) this highly abstract command is hard to decompose by sentence division but still can be correctly executed by our method, and ii) our 3D asset bank offers a large variety of objects for addition." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 46, + 630, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 630, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 630, + 288, + 714 + ], + "type": "text", + "content": "Multi-round command. We also perform a multi-round chat with our system, and the commands in different rounds exist context dependencies. The final results are shown in Figure 7. We see that i) our system is well-equipped to handle multi-round commands and execute the commands in each round precisely; ii) our system can handle the context dependencies across different rounds thanks to the recording" + } + ] + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 308, + 79, + 373, + 123 + ], + "blocks": [ + { + "bbox": [ + 326, + 71, + 353, + 79 + ], + "lines": [ + { + "bbox": [ + 326, + 71, + 353, + 79 + ], + "spans": [ + { + "bbox": [ + 326, + 71, + 353, + 79 + ], + "type": "text", + "content": "Front left" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 308, + 79, + 373, + 123 + ], + "lines": [ + { + "bbox": [ + 308, + 79, + 373, + 123 + ], + "spans": [ + { + "bbox": [ + 308, + 79, + 373, + 123 + ], + "type": "image", + "image_path": "e4c5778bba8772af279f5d2bea99dba1dec6a0777e6ce27cd9ba94d467452580.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 392, + 79, + 457, + 123 + ], + "blocks": [ + { + "bbox": [ + 415, + 71, + 432, + 79 + ], + "lines": [ + { + "bbox": [ + 415, + 71, + 432, + 79 + ], + "spans": [ + { + "bbox": [ + 415, + 71, + 432, + 79 + ], + "type": "text", + "content": "Front" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 392, + 79, + 457, + 123 + ], + "lines": [ + { + "bbox": [ + 392, + 79, + 457, + 123 + ], + "spans": [ + { + "bbox": [ + 392, + 79, + 457, + 123 + ], + "type": "image", + "image_path": "bb3473fdec38bcc1dabfc85a69d7790b4777373c0e913c07731f38a4fa8030da.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 475, + 79, + 540, + 123 + ], + "blocks": [ + { + "bbox": [ + 492, + 71, + 522, + 79 + ], + "lines": [ + { + "bbox": [ + 492, + 71, + 522, + 79 + ], + "spans": [ + { + "bbox": [ + 492, + 71, + 522, + 79 + ], + "type": "text", + "content": "Front right" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 475, + 79, + 540, + 123 + ], + "lines": [ + { + "bbox": [ + 475, + 79, + 540, + 123 + ], + "spans": [ + { + "bbox": [ + 475, + 79, + 540, + 123 + ], + "type": "image", + "image_path": "d608debd70fea97742d077397ad49aca403f5c604ed008564ec6b404a73257f5.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 308, + 125, + 423, + 176 + ], + "blocks": [ + { + "bbox": [ + 308, + 125, + 423, + 176 + ], + "lines": [ + { + "bbox": [ + 308, + 125, + 423, + 176 + ], + "spans": [ + { + "bbox": [ + 308, + 125, + 423, + 176 + ], + "type": "image", + "image_path": "5f523b0e48638bdba808421b135cb871cbb83d880f9e559cacc082a4923b763e.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 308, + 176, + 423, + 253 + ], + "blocks": [ + { + "bbox": [ + 308, + 176, + 423, + 253 + ], + "lines": [ + { + "bbox": [ + 308, + 176, + 423, + 253 + ], + "spans": [ + { + "bbox": [ + 308, + 176, + 423, + 253 + ], + "type": "image", + "image_path": "3842b951399e78dfbcee77369c218814ea73c4e67cdcbd06488fba96dd74e7cf.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 310, + 253, + 410, + 277 + ], + "lines": [ + { + "bbox": [ + 310, + 253, + 410, + 277 + ], + "spans": [ + { + "bbox": [ + 310, + 253, + 410, + 277 + ], + "type": "text", + "content": "Command 1: \"Ego vehicle drives ahead slowly. Add a car to the close front that is moving ahead.\"" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 424, + 125, + 540, + 176 + ], + "blocks": [ + { + "bbox": [ + 424, + 125, + 540, + 176 + ], + "lines": [ + { + "bbox": [ + 424, + 125, + 540, + 176 + ], + "spans": [ + { + "bbox": [ + 424, + 125, + 540, + 176 + ], + "type": "image", + "image_path": "e02f9db9cea4e58a76eb8915bd9f894b2cea9d69eb6a1c0c1ff9462f90dd83c8.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 424, + 176, + 540, + 252 + ], + "blocks": [ + { + "bbox": [ + 424, + 176, + 540, + 252 + ], + "lines": [ + { + "bbox": [ + 424, + 176, + 540, + 252 + ], + "spans": [ + { + "bbox": [ + 424, + 176, + 540, + 252 + ], + "type": "image", + "image_path": "06021c08aa05d33110e195aa01dcdcd49f9bc171a5bbf749b81dc5f46fe7d801.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 426, + 252, + 542, + 285 + ], + "lines": [ + { + "bbox": [ + 426, + 252, + 542, + 285 + ], + "spans": [ + { + "bbox": [ + 426, + 252, + 542, + 285 + ], + "type": "text", + "content": "Command 2: \"Modify the added car to turn left. Add a Chevrolet to the front of the added car. Add another vehicle to the left of the added Mini driving toward me.\"" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_caption" + } + ], + "index": 27 + }, + { + "type": "image", + "bbox": [ + 311, + 307, + 383, + 374 + ], + "blocks": [ + { + "bbox": [ + 326, + 285, + 524, + 297 + ], + "lines": [ + { + "bbox": [ + 326, + 285, + 524, + 297 + ], + "spans": [ + { + "bbox": [ + 326, + 285, + 524, + 297 + ], + "type": "text", + "content": "Figure 7. Editing result under multi-round commands." + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 311, + 307, + 383, + 374 + ], + "lines": [ + { + "bbox": [ + 311, + 307, + 383, + 374 + ], + "spans": [ + { + "bbox": [ + 311, + 307, + 383, + 374 + ], + "type": "image", + "image_path": "4cfd2cc41757614acabd5e77c875f05d89aeb79398df87090e2c313eed16bea4.jpg" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 306, + 374, + 545, + 396 + ], + "lines": [ + { + "bbox": [ + 306, + 374, + 545, + 396 + ], + "spans": [ + { + "bbox": [ + 306, + 374, + 545, + 396 + ], + "type": "text", + "content": "Figure 8. Comparison of detection performance w/o and with our simulated data under different amounts of real data during training." + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_caption" + } + ], + "index": 30 + }, + { + "type": "image", + "bbox": [ + 388, + 307, + 460, + 374 + ], + "blocks": [ + { + "bbox": [ + 388, + 307, + 460, + 374 + ], + "lines": [ + { + "bbox": [ + 388, + 307, + 460, + 374 + ], + "spans": [ + { + "bbox": [ + 388, + 307, + 460, + 374 + ], + "type": "image", + "image_path": "905eb43f45871bb8d325fee54b0b81cd9d099a391d5ba7dd6e269cfc0a09549c.jpg" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_body" + } + ], + "index": 31 + }, + { + "type": "image", + "bbox": [ + 460, + 307, + 536, + 374 + ], + "blocks": [ + { + "bbox": [ + 460, + 307, + 536, + 374 + ], + "lines": [ + { + "bbox": [ + 460, + 307, + 536, + 374 + ], + "spans": [ + { + "bbox": [ + 460, + 307, + 536, + 374 + ], + "type": "image", + "image_path": "7ac17b933135773c0bebdac17ec94c71dc4f72a0852e53208171a60d3f2d3368.jpg" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_body" + } + ], + "index": 32 + }, + { + "bbox": [ + 306, + 403, + 453, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 403, + 453, + 415 + ], + "spans": [ + { + "bbox": [ + 306, + 403, + 453, + 415 + ], + "type": "text", + "content": "ability of the project manager agent." + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 304, + 415, + 547, + 583 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 415, + 547, + 583 + ], + "spans": [ + { + "bbox": [ + 304, + 415, + 547, + 583 + ], + "type": "text", + "content": "3D detection with simulation data. We validate the benefits of our simulation as data augmentation for a downstream 3D object detection task on Waymo Open Dataset [58]. We simulate 1960 frames, derived from scenes in the training dataset. In the simulation, cars with various types, locations, and orientations are incorporated. The detection model adopts Lift-Splat [50]. Figure 8 shows detection performances with and without fixed augmentation under various amounts of real data. We see that i) a significant and consistent improvement across different data sizes is achieved; ii) when real data is limited, our simulation notably aids in rough detection (AP30); iii) when the amount of real data increases, our simulation further significantly improves fine-grained detection (AP70), reflecting the high-quality of our simulation." + } + ] + } + ], + "index": 35 + }, + { + "bbox": [ + 306, + 590, + 417, + 602 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 590, + 417, + 602 + ], + "spans": [ + { + "bbox": [ + 306, + 590, + 417, + 602 + ], + "type": "text", + "content": "5.3. Component results" + } + ] + } + ], + "index": 36 + }, + { + "bbox": [ + 304, + 605, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 547, + 714 + ], + "type": "text", + "content": "Multi-agent collaboration. We evaluated the effectiveness of the multi-agent collaboration by checking whether the command is successfully executed in Table 2. In scenarios without multi-agent collaboration, all operations are executed by a single LLM agent. We see that a single LLM agent leads to notably lower execution accuracy across all categories due to process limitations. In contrast, the collaborative multi-agent approach can manage most commands, attributed to its task division and agent role specificity." + } + ] + } + ], + "index": 37 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15083" + } + ] + } + ], + "index": 38 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 60, + 71, + 276, + 114 + ], + "blocks": [ + { + "bbox": [ + 60, + 71, + 276, + 114 + ], + "lines": [ + { + "bbox": [ + 60, + 71, + 276, + 114 + ], + "spans": [ + { + "bbox": [ + 60, + 71, + 276, + 114 + ], + "type": "table", + "html": "
Multi-agent collaborationLanguage command category
DeletionAdditionView changeRevisionAbstract
✓0.6170.3830.7170.3670.216
0.9830.8670.9670.9170.883
", + "image_path": "4caa2587df8a2367762f47bc8ac637b0973c255f5542aa592d3e64e7bd23423f.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 61, + 140, + 272, + 221 + ], + "blocks": [ + { + "bbox": [ + 47, + 114, + 287, + 137 + ], + "lines": [ + { + "bbox": [ + 47, + 114, + 287, + 137 + ], + "spans": [ + { + "bbox": [ + 47, + 114, + 287, + 137 + ], + "type": "text", + "content": "Table 2. The accuracy " + }, + { + "bbox": [ + 47, + 114, + 287, + 137 + ], + "type": "inline_equation", + "content": "(\\%)" + }, + { + "bbox": [ + 47, + 114, + 287, + 137 + ], + "type": "text", + "content": " of task completion by LLM without and with multi-agent collaboration." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 61, + 140, + 272, + 221 + ], + "lines": [ + { + "bbox": [ + 61, + 140, + 272, + 221 + ], + "spans": [ + { + "bbox": [ + 61, + 140, + 272, + 221 + ], + "type": "table", + "html": "
MethodsPSNR↑SSIM↑LPIPS↓Inf. time (s)↓
DVGO [56]23.570.7700.5087.7
Mip-NeRF360 [8]24.400.7540.528101.8
S-NeRF [69]24.710.7590.519114.5
F2NeRF [62]23.260.7730.4392.4
Ours w/o alignment23.320.7760.4372.5
Ours w/o exposure25.180.8190.3812.4
McNeRF (Ours)25.820.8220.3782.5
", + "image_path": "6988eff828e2eb6c2cd4eef57bd61d71e36647ae8f52ace0105c019e29da90c1.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 52, + 238, + 280, + 442 + ], + "blocks": [ + { + "bbox": [ + 52, + 238, + 280, + 442 + ], + "lines": [ + { + "bbox": [ + 52, + 238, + 280, + 442 + ], + "spans": [ + { + "bbox": [ + 52, + 238, + 280, + 442 + ], + "type": "image", + "image_path": "0b85171da357d5b1be05d0a7ff7e0e46dd219956fc9bb7cc47b073fc4b23473a.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 47, + 444, + 288, + 467 + ], + "lines": [ + { + "bbox": [ + 47, + 444, + 288, + 467 + ], + "spans": [ + { + "bbox": [ + 47, + 444, + 288, + 467 + ], + "type": "text", + "content": "Figure 9. Comparisons of wide-angle images generation. (a) S-NeRF.(b) F2NeRF. (c) McNeRF (Ours). Last row: target images." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 474, + 287, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 474, + 287, + 581 + ], + "spans": [ + { + "bbox": [ + 46, + 474, + 287, + 581 + ], + "type": "text", + "content": "Background rendering. We compare our McNeRF with several other state-of-the-art methods on the background novel view synthesis task. We perform reconstruction and rendering on 32 selected scenes. Table 3 shows the quantitative results comparison on three metrics: PSNR, SSIM, and LPIPS. We see that i) McNeRF achieves SoTA performance on all three metrics, significantly outperforming other baselines; ii) McNeRF has a fast inference speed, enabling quick responses to user requests for image rendering." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 582, + 288, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 582, + 288, + 665 + ], + "spans": [ + { + "bbox": [ + 46, + 582, + 288, + 665 + ], + "type": "text", + "content": "Figure 9 demonstrates qualitative comparisons between other methods and ours. We see that existing NeRF methods do not consider the exposure time, leading to noticeable changes in brightness at the junctions of different cameras in the image, as well as an overall inconsistency in exposure across the wide-angle view. Our method can make the brightness of the entire image more consistent." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 665, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 288, + 714 + ], + "type": "text", + "content": "Foreground rendering. We compare our McLight with the other two SoTA methods [31, 64]. Table 4 shows the comparison of relative intensity(log 10) error on our HDRI dataset, angular error on HoliCity [75], and user study. We see that" + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 308, + 71, + 541, + 111 + ], + "blocks": [ + { + "bbox": [ + 47, + 222, + 286, + 233 + ], + "lines": [ + { + "bbox": [ + 47, + 222, + 286, + 233 + ], + "spans": [ + { + "bbox": [ + 47, + 222, + 286, + 233 + ], + "type": "text", + "content": "Table 3. Background novel view rendering performance evaluation." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 308, + 71, + 541, + 111 + ], + "lines": [ + { + "bbox": [ + 308, + 71, + 541, + 111 + ], + "spans": [ + { + "bbox": [ + 308, + 71, + 541, + 111 + ], + "type": "table", + "html": "
MethodPeak Intensity(log10) ErrorPeak Angular Error (deg)User study(%) ↑
Mean ↓Median ↓Mean ↓Median ↓
Hold-Geoffroy et al. [31]0.8990.97548.451.619.5
Wang et al. [64]0.5900.62833.529.437.3
McLight (Ours)0.4490.27032.326.543.1
", + "image_path": "a62ba856d2e7105826c19c9872d9e137b225b7a3979e893d40c02f7d3ffb9a20.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 111, + 544, + 122 + ], + "lines": [ + { + "bbox": [ + 306, + 111, + 544, + 122 + ], + "spans": [ + { + "bbox": [ + 306, + 111, + 544, + 122 + ], + "type": "text", + "content": "Table 4. Comparison with previous methods on lighting estimation." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 315, + 125, + 369, + 202 + ], + "blocks": [ + { + "bbox": [ + 315, + 125, + 369, + 202 + ], + "lines": [ + { + "bbox": [ + 315, + 125, + 369, + 202 + ], + "spans": [ + { + "bbox": [ + 315, + 125, + 369, + 202 + ], + "type": "image", + "image_path": "040a647041766cdad1cfb156f7c86f9ff6980b2de7daa1aae433af54102c829c.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 370, + 125, + 421, + 202 + ], + "blocks": [ + { + "bbox": [ + 370, + 125, + 421, + 202 + ], + "lines": [ + { + "bbox": [ + 370, + 125, + 421, + 202 + ], + "spans": [ + { + "bbox": [ + 370, + 125, + 421, + 202 + ], + "type": "image", + "image_path": "b7ceb900ce547806bd527b7d1f9d3769b75bce40061dda59456290851f57fd86.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 423, + 125, + 477, + 202 + ], + "blocks": [ + { + "bbox": [ + 423, + 125, + 477, + 202 + ], + "lines": [ + { + "bbox": [ + 423, + 125, + 477, + 202 + ], + "spans": [ + { + "bbox": [ + 423, + 125, + 477, + 202 + ], + "type": "image", + "image_path": "ac2c7f5dc676175ca7c8baea74a45e917bfd0d38ffe62903fd152edac5040bb8.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 481, + 125, + 534, + 202 + ], + "blocks": [ + { + "bbox": [ + 481, + 125, + 534, + 202 + ], + "lines": [ + { + "bbox": [ + 481, + 125, + 534, + 202 + ], + "spans": [ + { + "bbox": [ + 481, + 125, + 534, + 202 + ], + "type": "image", + "image_path": "2b4e959904e821cd521964733264a315c2842aea89eda6db7528caed6bf9c7d2.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "table", + "bbox": [ + 313, + 217, + 538, + 255 + ], + "blocks": [ + { + "bbox": [ + 306, + 203, + 544, + 214 + ], + "lines": [ + { + "bbox": [ + 306, + 203, + 544, + 214 + ], + "spans": [ + { + "bbox": [ + 306, + 203, + 544, + 214 + ], + "type": "text", + "content": "Figure 10. Comparison with different lighting estimation methods." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 313, + 217, + 538, + 255 + ], + "lines": [ + { + "bbox": [ + 313, + 217, + 538, + 255 + ], + "spans": [ + { + "bbox": [ + 313, + 217, + 538, + 255 + ], + "type": "table", + "html": "
MethodsStraightLeft TurnRight TurnSpeedWithin-road
GPT2Code0.7380.5590.5360.8930.214
GPT2Motion0.5950.1190.1670.3450.277
Ours0.9880.9400.9760.9521.000
", + "image_path": "24a02b02c09057bc5a763463a261d2d990bfd4256e193ec081b99c45a4316e27.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "table_body" + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 255, + 542, + 266 + ], + "lines": [ + { + "bbox": [ + 307, + 255, + 542, + 266 + ], + "spans": [ + { + "bbox": [ + 307, + 255, + 542, + 266 + ], + "type": "text", + "content": "Table 5. Comparison with motion generation from text methods." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 305, + 270, + 547, + 329 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 270, + 547, + 329 + ], + "spans": [ + { + "bbox": [ + 305, + 270, + 547, + 329 + ], + "type": "text", + "content": "McLight achieves more accurate peak behavior and receives noticeably higher user preferences. Figure 10 shows the visualizations of vehicle insertion. The vehicles added through McLight feature significantly more realistic reflections and strong shadows consistent with the scene." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 305, + 330, + 546, + 474 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 330, + 546, + 474 + ], + "spans": [ + { + "bbox": [ + 305, + 330, + 546, + 474 + ], + "type": "text", + "content": "Vehicle motion. As shown in Table 5, we compare the motion generation method from user commands with two of our designed baselines: 1. GPT2Motion, which directly uses LLM to return the motion coordinates; 2. GPT2Code, which first generates code using LLM and executes it to obtain the vehicle motion. We validate multiple actions in multiple scenarios and report the user study result. The user study is to determine if the generated motions matched the command intents and fitted with the lane map. We see that our method demonstrated a significant advantage in generating motions from language commands. Additionally, it maintained a high rate of keeping the trajectories within the lane boundaries." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 306, + 478, + 468, + 491 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 478, + 468, + 491 + ], + "spans": [ + { + "bbox": [ + 306, + 478, + 468, + 491 + ], + "type": "text", + "content": "6. Conclusions and Limitations" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 495, + 547, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 495, + 547, + 639 + ], + "spans": [ + { + "bbox": [ + 304, + 495, + 547, + 639 + ], + "type": "text", + "content": "This paper introduces ChatSim, the first system for editing 3D driving scene simulations via language commands and realistic rendering with import of external digital assets. To effectively execute user commands, ChatSim adopts an LLM-agent collaboration workflow. To promote photorealistic simulation, we propose McNeRF and McLight for background and foreground rendering, respectively, accommodating multi-camera inputs. Experiments show that ChatSim successfully simulates customized data via language commands, achieving high-quality, photo-realistic outcomes. In future, we plan to integrate more background editing functionalities to ChatSim, such as weather changes." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 306, + 650, + 404, + 663 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 650, + 404, + 663 + ], + "spans": [ + { + "bbox": [ + 306, + 650, + 404, + 663 + ], + "type": "text", + "content": "Acknowledgement" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 305, + 669, + 547, + 703 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 669, + 547, + 703 + ], + "spans": [ + { + "bbox": [ + 305, + 669, + 547, + 703 + ], + "type": "text", + "content": "This research is supported by NSFC under Grant 62171276 and the Science and Technology Commission of Shanghai Municipal under Grant 2151100900 and 22DZ2229005." + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "15084" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 53, + 91, + 249, + 101 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 249, + 101 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 249, + 101 + ], + "type": "text", + "content": "[1] 51Sim-One. https://wdp.51aes.com/news/27.2,3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 103, + 287, + 124 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 103, + 287, + 124 + ], + "spans": [ + { + "bbox": [ + 53, + 103, + 287, + 124 + ], + "type": "text", + "content": "[2] LLC Agisoft. Metashape-photogrammetric processing of digital images and 3d spatial data generation, 2019. 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 126, + 288, + 167 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 126, + 288, + 167 + ], + "spans": [ + { + "bbox": [ + 53, + 126, + 288, + 167 + ], + "type": "text", + "content": "[3] Elif Akata, Lion Schulz, Julian Coda-Forno, Seong Joon Oh, Matthias Bethge, and Eric Schulz. Playing repeated games with large language models. arXiv preprint arXiv:2305.16867, 2023. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 170, + 288, + 224 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 170, + 288, + 224 + ], + "spans": [ + { + "bbox": [ + 53, + 170, + 288, + 224 + ], + "type": "text", + "content": "[4] Alexander Amini, Igor Gilitschenski, Jacob Phillips, Julia Moseyko, Rohan Banerjee, Sertac Karaman, and Daniela Rus. Learning robust control policies for end-to-end autonomous driving from data-driven simulation. IEEE Robotics and Automation Letters, 5(2):1143-1150, 2020. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 226, + 288, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 226, + 288, + 270 + ], + "spans": [ + { + "bbox": [ + 53, + 226, + 288, + 270 + ], + "type": "text", + "content": "[5] Rohan Anil, Andrew M Dai, Orhan First, Melvin Johnson, Dmitry Lepikhin, Alexandre Passos, Siamak Shakeri, Emanuel Taropa, Paige Bailey, Zhifeng Chen, et al. Palm 2 technical report. arXiv preprint arXiv:2305.10403, 2023. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 271, + 287, + 292 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 271, + 287, + 292 + ], + "spans": [ + { + "bbox": [ + 53, + 271, + 287, + 292 + ], + "type": "text", + "content": "[6] Natalie M Banta. Property interests in digital assets: The rise of digital feudalism. Cardozo L. Rev., 38:1099, 2016. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 293, + 287, + 347 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 293, + 287, + 347 + ], + "spans": [ + { + "bbox": [ + 53, + 293, + 287, + 347 + ], + "type": "text", + "content": "[7] Jonathan T Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P Srinivasan. Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5855–5864, 2021. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 349, + 287, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 349, + 287, + 403 + ], + "spans": [ + { + "bbox": [ + 53, + 349, + 287, + 403 + ], + "type": "text", + "content": "[8] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Mip-nerf 360: Unbounded anti-aliased neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5470–5479, 2022. 2, 8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 404, + 288, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 404, + 288, + 468 + ], + "spans": [ + { + "bbox": [ + 53, + 404, + 288, + 468 + ], + "type": "text", + "content": "[9] Luca Bergamini, Yawei Ye, Oliver Scheel, Long Chen, Chih Hu, Luca Del Pero, Băjej Osiński, Hugo Grimmett, and Peter Ondruska. Simnet: Learning reactive self-driving simulations from real-world observations. In 2021 IEEE International Conference on Robotics and Automation (ICRA), pages 5119-5125. IEEE, 2021. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 471, + 287, + 523 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 471, + 287, + 523 + ], + "spans": [ + { + "bbox": [ + 48, + 471, + 287, + 523 + ], + "type": "text", + "content": "[10] Mark Boss, Varun Jampani, Kihwan Kim, Hendrik Lensch, and Jan Kautz. Two-shot spatially-varying brdf and shape estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3982-3991, 2020. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 525, + 287, + 581 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 525, + 287, + 581 + ], + "spans": [ + { + "bbox": [ + 48, + 525, + 287, + 581 + ], + "type": "text", + "content": "[11] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. Advances in neural information processing systems, 33:1877-1901, 2020. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 582, + 288, + 647 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 582, + 288, + 647 + ], + "spans": [ + { + "bbox": [ + 48, + 582, + 288, + 647 + ], + "type": "text", + "content": "[12] Holger Caesar, Varun Bankiti, Alex H Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuscenes: A multimodal dataset for autonomous driving. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11621-11631, 2020. 1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 647, + 288, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 647, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 647, + 288, + 713 + ], + "type": "text", + "content": "[13] Ming-Fang Chang, John Lambert, Patsorn Sangkloy, Jagjeet Singh, Slawomir Bak, Andrew Hartnett, De Wang, Peter Carr, Simon Lucey, Deva Ramanan, et al. Argoverse: 3d tracking and forecasting with rich maps. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8748-8757, 2019. 1" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 560, + 713 + ], + "type": "list", + "angle": 0, + "index": 31, + "blocks": [ + { + "bbox": [ + 307, + 73, + 492, + 84 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 492, + 84 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 492, + 84 + ], + "type": "text", + "content": "[14] ChatGPT. https://openai.com/blog/chatgpt.3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 86, + 547, + 129 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 86, + 547, + 129 + ], + "spans": [ + { + "bbox": [ + 308, + 86, + 547, + 129 + ], + "type": "text", + "content": "[15] Chenyi Chen, Ari Seff, Alain Kornhauser, and Jianxiong Xiao. Deepdriving: Learning affordance for direct perception in autonomous driving. In Proceedings of the IEEE international conference on computer vision, pages 2722-2730, 2015. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 131, + 547, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 131, + 547, + 184 + ], + "spans": [ + { + "bbox": [ + 308, + 131, + 547, + 184 + ], + "type": "text", + "content": "[16] Xiaozhi Chen, Kaustav Kundu, Ziyu Zhang, Huimin Ma, Sanja Fidler, and Raquel Urtasun. Monocular 3d object detection for autonomous driving. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2147-2156, 2016." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 186, + 547, + 230 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 186, + 547, + 230 + ], + "spans": [ + { + "bbox": [ + 308, + 186, + 547, + 230 + ], + "type": "text", + "content": "[17] Xiaozhi Chen, Huimin Ma, Ji Wan, Bo Li, and Tian Xia. Multi-view 3d object detection network for autonomous driving. In Proceedings of the IEEE conference on Computer Vision and Pattern Recognition, pages 1907-1915, 2017. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 232, + 546, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 232, + 546, + 285 + ], + "spans": [ + { + "bbox": [ + 308, + 232, + 546, + 285 + ], + "type": "text", + "content": "[18] Aakanksha Chowdhery, Sharan Narang, Jacob Devlin, Maarten Bosma, Gaurav Mishra, Adam Roberts, Paul Barham, Hyung Won Chung, Charles Sutton, Sebastian Gehrmann, et al. Palm: Scaling language modeling with pathways. arXiv preprint arXiv:2204.02311, 2022. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 287, + 547, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 287, + 547, + 330 + ], + "spans": [ + { + "bbox": [ + 308, + 287, + 547, + 330 + ], + "type": "text", + "content": "[19] International Electrotechnical Commission et al. Iec 61966-2-1: 1999 multimedia systems and equipment- colour measurement and management- part 2-1: Colour management-default rgb colour space-srgb, 1999. 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 332, + 546, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 332, + 546, + 364 + ], + "spans": [ + { + "bbox": [ + 308, + 332, + 546, + 364 + ], + "type": "text", + "content": "[20] Blender Online Community. Blender - a 3D modelling and rendering package. Blender Foundation, Stichting Blender Foundation, Amsterdam, 2018. 5" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 365, + 546, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 365, + 546, + 398 + ], + "spans": [ + { + "bbox": [ + 308, + 365, + 546, + 398 + ], + "type": "text", + "content": "[21] Valter Crescenzi, Giansalvatore Mecca, Paolo Merialdo, et al. Roadrunner: Towards automatic data extraction from large web sites. In VLDB, pages 109-118, 2001. 2, 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 399, + 547, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 399, + 547, + 442 + ], + "spans": [ + { + "bbox": [ + 308, + 399, + 547, + 442 + ], + "type": "text", + "content": "[22] Alexey Dosovitskiy, German Ros, Felipe Codevilla, Antonio Lopez, and Vladlen Koltun. Carla: An open urban driving simulator. In Conference on robot learning, pages 1-16. PMLR, 2017. 2, 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 445, + 546, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 445, + 546, + 487 + ], + "spans": [ + { + "bbox": [ + 308, + 445, + 546, + 487 + ], + "type": "text", + "content": "[23] Yilun Du, Shuang Li, Antonio Torralba, Joshua B Tenenbaum, and Igor Mordatch. Improving factuality and reasoning in language models through multiagent debate. arXiv preprint arXiv:2305.14325, 2023. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 489, + 560, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 489, + 560, + 510 + ], + "spans": [ + { + "bbox": [ + 308, + 489, + 560, + 510 + ], + "type": "text", + "content": "[24] Openscanerio Editor. https://github.com/ebadi/openscenarioeditor. 2,3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 512, + 509, + 523 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 512, + 509, + 523 + ], + "spans": [ + { + "bbox": [ + 308, + 512, + 509, + 523 + ], + "type": "text", + "content": "[25] Unreal Engine. https://www.unrealengine.com/. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 525, + 546, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 525, + 546, + 567 + ], + "spans": [ + { + "bbox": [ + 308, + 525, + 546, + 567 + ], + "type": "text", + "content": "[26] Ruiyuan Gao, Kai Chen, Enze Xie, Lanqing Hong, Zhenguo Li, Dit-Yan Yeung, and Qiang Xu. Magicdrive: Street view generation with diverse 3d geometry control. arXiv preprint arXiv:2310.02601, 2023. 2, 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 569, + 547, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 569, + 547, + 623 + ], + "spans": [ + { + "bbox": [ + 308, + 569, + 547, + 623 + ], + "type": "text", + "content": "[27] Mathieu Garon, Kalyan Sunkavalli, Sunil Hadap, Nathan Carr, and Jean-François Lalonde. Fast spatially-varying indoor lighting estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6908-6917, 2019. 3" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 625, + 547, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 625, + 547, + 678 + ], + "spans": [ + { + "bbox": [ + 308, + 625, + 547, + 678 + ], + "type": "text", + "content": "[28] Jianfei Guo, Nianchen Deng, Xinyang Li, Yeqi Bai, Botian Shi, Chiyu Wang, Chenjing Ding, Dongliang Wang, and Yikang Li. Streetsurf: Extending multi-view implicit surface reconstruction to street views. arXiv preprint arXiv:2306.04988, 2023. 2" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 308, + 681, + 547, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 681, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 681, + 547, + 713 + ], + "type": "text", + "content": "[29] Rui Hao, Linmei Hu, Weijian Qi, Qingliu Wu, Yirui Zhang, and Liqiang Nie. Chatllm network: More brains, more intelligence. arXiv preprint arXiv:2304.12998, 2023. 3" + } + ] + } + ], + "index": 30 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15085" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 714 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "text", + "content": "[30] Yannick Hold-Geoffroy, Kalyan Sunkavalli, Sunil Hadap, Emiliano Gambaretto, and Jean-François Lalonde. Deep outdoor illumination estimation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7312–7321, 2017. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 129, + 288, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 129, + 288, + 184 + ], + "spans": [ + { + "bbox": [ + 48, + 129, + 288, + 184 + ], + "type": "text", + "content": "[31] Yannick Hold-Geoffroy, Akshaya Athawale, and Jean-François Lalonde. Deep sky modeling for single image outdoor lighting estimation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6927–6935, 2019. 3, 6, 8" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 185, + 288, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 185, + 288, + 239 + ], + "spans": [ + { + "bbox": [ + 48, + 185, + 288, + 239 + ], + "type": "text", + "content": "[32] Sirui Hong, Xiawu Zheng, Jonathan Chen, Yuheng Cheng, Ceyao Zhang, Zili Wang, Steven Ka Shing Yau, Zijuan Lin, Liyang Zhou, Chenyu Ran, et al. Metagpt: Meta programming for multi-agent collaborative framework. arXiv preprint arXiv:2308.00352, 2023. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 241, + 288, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 241, + 288, + 285 + ], + "spans": [ + { + "bbox": [ + 48, + 241, + 288, + 285 + ], + "type": "text", + "content": "[33] Anthony Hu, Lloyd Russell, Hudson Yeo, Zak Murez, George Fedoseev, Alex Kendall, Jamie Shotton, and Gianluca Corrado. Gaia-1: A generative world model for autonomous driving. arXiv preprint arXiv:2309.17080, 2023. 2, 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 286, + 288, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 286, + 288, + 352 + ], + "spans": [ + { + "bbox": [ + 48, + 286, + 288, + 352 + ], + "type": "text", + "content": "[34] Abhijit Kundu, Kyle Genova, Xiaoqi Yin, Alireza Fathi, Caroline Pantofaru, Leonidas J Guibas, Andrea Tagliasacchi, Frank Dellaert, and Thomas Funkhouser. Panoptic neural fields: A semantic object-aware neural scene representation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12871-12881, 2022. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 354, + 288, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 354, + 288, + 387 + ], + "spans": [ + { + "bbox": [ + 48, + 354, + 288, + 387 + ], + "type": "text", + "content": "[35] Jean-François Lalonde and Iain Matthews. Lighting estimation in outdoor image collections. In 2014 2nd international conference on 3D vision, pages 131-138. IEEE, 2014. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 388, + 288, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 388, + 288, + 431 + ], + "spans": [ + { + "bbox": [ + 48, + 388, + 288, + 431 + ], + "type": "text", + "content": "[36] Jean-François Lalonde, Srinivasa G Narasimhan, and Alexei A Efros. What do the sun and the sky tell us about the camera? International Journal of Computer Vision, 88: 24-51, 2010. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 433, + 288, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 433, + 288, + 487 + ], + "spans": [ + { + "bbox": [ + 48, + 433, + 288, + 487 + ], + "type": "text", + "content": "[37] Chloe LeGendre, Wan-Chun Ma, Graham Fyffe, John Flynn, Laurent Charbonnel, Jay Busch, and Paul Debevec. Deeplight: Learning illumination for unconstrained mobile mixed reality. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5918-5928, 2019. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 489, + 288, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 489, + 288, + 533 + ], + "spans": [ + { + "bbox": [ + 48, + 489, + 288, + 533 + ], + "type": "text", + "content": "[38] Peixuan Li, Huaici Zhao, Pengfei Liu, and Feidao Cao. Rtm3d: Real-time monocular 3d detection from object keypoints for autonomous driving. In European Conference on Computer Vision, pages 644-660. Springer, 2020. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 534, + 288, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 534, + 288, + 578 + ], + "spans": [ + { + "bbox": [ + 48, + 534, + 288, + 578 + ], + "type": "text", + "content": "[39] Xiaofan Li, Yifu Zhang, and Xiaoqing Ye. Drivingdiffusion: Layout-guided multi-view driving scene video generation with latent diffusion model. arXiv preprint arXiv:2310.07771, 2023. 2, 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 579, + 288, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 579, + 288, + 633 + ], + "spans": [ + { + "bbox": [ + 48, + 579, + 288, + 633 + ], + "type": "text", + "content": "[40] Zhengqin Li, Zexiang Xu, Ravi Ramamoorthi, Kalyan Sunkavalli, and Manmohan Chandraker. Learning to reconstruct shape and spatially-varying reflectance from a single image. ACM Transactions on Graphics (TOG), 37(6):1-11, 2018. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 635, + 288, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 635, + 288, + 678 + ], + "spans": [ + { + "bbox": [ + 48, + 635, + 288, + 678 + ], + "type": "text", + "content": "[41] Zhuopeng Li, Lu Li, and Jianke Zhu. Read: Large-scale neural scene rendering for autonomous driving. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1522-1529, 2023. 2, 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 680, + 288, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 680, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 48, + 680, + 288, + 714 + ], + "type": "text", + "content": "[42] Zhijian Liu, Haotian Tang, Alexander Amini, Xinyu Yang, Huizi Mao, Daniela L Rus, and Song Han. Bevfusion: Multi-task multi-sensor fusion with unified bird's-eye view represent" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "text", + "content": "tation. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 2774-2781. IEEE, 2023. 4" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 95, + 547, + 148 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 95, + 547, + 148 + ], + "spans": [ + { + "bbox": [ + 307, + 95, + 547, + 148 + ], + "type": "text", + "content": "[43] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 150, + 547, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 150, + 547, + 194 + ], + "spans": [ + { + "bbox": [ + 307, + 150, + 547, + 194 + ], + "type": "text", + "content": "[44] Milan Miric, Kevin J Boudreau, and Lars Bo Jeppesen. Protecting their digital assets: The use of formal & informal appropriability strategies by app developers. Research Policy, 48(8):103738, 2019. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 195, + 547, + 238 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 195, + 547, + 238 + ], + "spans": [ + { + "bbox": [ + 307, + 195, + 547, + 238 + ], + "type": "text", + "content": "[45] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM Transactions on Graphics (ToG), 41(4):1-15, 2022. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 239, + 476, + 250 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 239, + 476, + 250 + ], + "spans": [ + { + "bbox": [ + 307, + 239, + 476, + 250 + ], + "type": "text", + "content": "[46] OpenAI. Gpt-4 technical report. 2023. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 251, + 547, + 294 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 251, + 547, + 294 + ], + "spans": [ + { + "bbox": [ + 307, + 251, + 547, + 294 + ], + "type": "text", + "content": "[47] Julian Ost, Fahim Mannan, Nils Thuerey, Julian Knodt, and Felix Heide. Neural scene graphs for dynamic scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2856-2865, 2021. 2, 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 295, + 547, + 338 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 295, + 547, + 338 + ], + "spans": [ + { + "bbox": [ + 307, + 295, + 547, + 338 + ], + "type": "text", + "content": "[48] Julian Ost, Issam Laradji, Alejandro Newell, Yuval Bahat, and Felix Heide. Neural point light fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18419-18429, 2022. 2, 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 338, + 547, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 338, + 547, + 403 + ], + "spans": [ + { + "bbox": [ + 307, + 338, + 547, + 403 + ], + "type": "text", + "content": "[49] Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, et al. Training language models to follow instructions with human feedback. Advances in Neural Information Processing Systems, 35:27730-27744, 2022. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 404, + 545, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 404, + 545, + 459 + ], + "spans": [ + { + "bbox": [ + 307, + 404, + 545, + 459 + ], + "type": "text", + "content": "[50] Jonah Philion and Sanja Fidler. Lift, splat, shoot: Encoding images from arbitrary camera rigs by implicitly unprojecting to 3d. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XIV 16, pages 194-210. Springer, 2020. 7" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 460, + 545, + 492 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 460, + 545, + 492 + ], + "spans": [ + { + "bbox": [ + 307, + 460, + 545, + 492 + ], + "type": "text", + "content": "[51] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models, 2021. 4" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 493, + 545, + 558 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 493, + 545, + 558 + ], + "spans": [ + { + "bbox": [ + 307, + 493, + 545, + 558 + ], + "type": "text", + "content": "[52] Guodong Rong, Byung Hyun Shin, Hadi Tabatabaee, Qiang Lu, Steve Lemke, Martins Mozeiko, Eric Boise, Geehoon Uhm, Mark Gerow, Shalin Mehta, et al. Lgsvl simulator: A high fidelity simulator for autonomous driving. In 2020 IEEE 23rd International conference on intelligent transportation systems (ITSC), pages 1-6. IEEE, 2020. 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 559, + 547, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 559, + 547, + 613 + ], + "spans": [ + { + "bbox": [ + 307, + 559, + 547, + 613 + ], + "type": "text", + "content": "[53] Soumyadip Sengupta, Jinwei Gu, Kihwan Kim, Guilin Liu, David W Jacobs, and Jan Kautz. Neural inverse rendering of an indoor scene from a single image. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8598-8607, 2019. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 614, + 547, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 614, + 547, + 669 + ], + "spans": [ + { + "bbox": [ + 307, + 614, + 547, + 669 + ], + "type": "text", + "content": "[54] Shital Shah, Debadeepta Dey, Chris Lovett, and Ashish Kapoor. Airsim: High-fidelity visual and physical simulation for autonomous vehicles. In Field and Service Robotics: Results of the 11th International Conference, pages 621-635. Springer, 2018. 2, 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 670, + 547, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 670, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 670, + 547, + 713 + ], + "type": "text", + "content": "[55] Gowri Somanath and Daniel Kurz. Hdr environment map estimation for real-time augmented reality. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11298-11306, 2021. 3" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "text", + "content": "15086" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 288, + 127 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 288, + 127 + ], + "type": "text", + "content": "[56] Cheng Sun, Min Sun, and Hwann-Tzong Chen. Direct voxel grid optimization: Super-fast convergence for radiance fields reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5459-5469, 2022. 2, 8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 129, + 288, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 129, + 288, + 228 + ], + "spans": [ + { + "bbox": [ + 48, + 129, + 288, + 228 + ], + "type": "text", + "content": "[57] Pei Sun, Henrik Kretzschmar, Xerxes Dotiwalla, Aurelien Chouard, Vijaysai Patnaik, Paul Tsui, James Guo, Yin Zhou, Yuning Chai, Benjamin Caine, Vijay Vasudevan, Wei Han, Jiquan Ngiam, Hang Zhao, Aleksei Timofeev, Scott Ettinger, Maxim Krivokon, Amy Gao, Aditya Joshi, Yu Zhang, Jonathon Shlens, Zhifeng Chen, and Dragomir Anguelov. Scalability in perception for autonomous driving: Waymo open dataset. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 230, + 288, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 230, + 288, + 296 + ], + "spans": [ + { + "bbox": [ + 48, + 230, + 288, + 296 + ], + "type": "text", + "content": "[58] Pei Sun, Henrik Kretzschmar, Xerxes Dotiwalla, Aurelien Chouard, Vijaysai Patnaik, Paul Tsui, James Guo, Yin Zhou, Yuning Chai, Benjamin Caine, et al. Scalability in perception for autonomous driving: Waymo open dataset. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2446-2454, 2020. 1, 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 297, + 288, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 297, + 288, + 330 + ], + "spans": [ + { + "bbox": [ + 48, + 297, + 288, + 330 + ], + "type": "text", + "content": "[59] Alexander Swerdlow, Runsheng Xu, and Bolei Zhou. Street-view image generation from a bird's-eye view layout. arXiv preprint arXiv:2301.04634, 2023. 2, 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 331, + 288, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 331, + 288, + 386 + ], + "spans": [ + { + "bbox": [ + 48, + 331, + 288, + 386 + ], + "type": "text", + "content": "[60] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288, 2023. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 388, + 288, + 432 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 388, + 288, + 432 + ], + "spans": [ + { + "bbox": [ + 48, + 388, + 288, + 432 + ], + "type": "text", + "content": "[61] Haithem Turki, Jason Y Zhang, Francesco Ferroni, and Deva Ramanan. Suds: Scalable urban dynamic scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12375-12385, 2023. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 434, + 288, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 434, + 288, + 498 + ], + "spans": [ + { + "bbox": [ + 48, + 434, + 288, + 498 + ], + "type": "text", + "content": "[62] Peng Wang, Yuan Liu, Zhaoxi Chen, Lingjie Liu, Ziwei Liu, Taku Komura, Christian Theobalt, and Wenping Wang. F2-nerf: Fast neural radiance field training with free camera trajectories. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4150–4159, 2023. 2, 5, 8" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 500, + 288, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 500, + 288, + 544 + ], + "spans": [ + { + "bbox": [ + 48, + 500, + 288, + 544 + ], + "type": "text", + "content": "[63] Xiaofeng Wang, Zheng Zhu, Guan Huang, Xinze Chen, and Jiwen Lu. Drivedreamer: Towards real-world-driven world models for autonomous driving. arXiv preprint arXiv:2309.09777, 2023. 2, 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 545, + 288, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 545, + 288, + 600 + ], + "spans": [ + { + "bbox": [ + 48, + 545, + 288, + 600 + ], + "type": "text", + "content": "[64] Zian Wang, Wenzheng Chen, David Acuna, Jan Kautz, and Sanja Fidler. Neural light field estimation for street scenes with differentiable virtual object insertion. In European Conference on Computer Vision, pages 380-397. Springer, 2022. 3, 6, 8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 602, + 288, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 602, + 288, + 646 + ], + "spans": [ + { + "bbox": [ + 48, + 602, + 288, + 646 + ], + "type": "text", + "content": "[65] Zhenhailong Wang, Shaoguang Mao, Wenshan Wu, Tao Ge, Furu Wei, and Heng Ji. Unleashing cognitive synergy in large language models: A task-solving agent through multi-personal self-collaboration. arXiv preprint arXiv:2307.05300, 2023. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 647, + 288, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 647, + 288, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 647, + 288, + 712 + ], + "type": "text", + "content": "[66] Bichen Wu, Forrest Iandola, Peter H Jin, and Kurt Keutzer. Squeezedet: Unified, small, low power fully convolutional neural networks for real-time object detection for autonomous driving. In Proceedings of the IEEE conference on computer vision and pattern recognition workshops, pages 129-137, 2017. 1" + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 586 + ], + "type": "list", + "angle": 0, + "index": 22, + "blocks": [ + { + "bbox": [ + 307, + 73, + 547, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 547, + 127 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 547, + 127 + ], + "type": "text", + "content": "[67] Qingyun Wu, Gagan Bansal, Jieyu Zhang, Yiran Wu, Shaokun Zhang, Erkang Zhu, Beibin Li, Li Jiang, Xiaoyun Zhang, and Chi Wang. Autogen: Enabling next-gen llm applications via multi-agent conversation framework. arXiv preprint arXiv:2308.08155, 2023. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 307, + 129, + 547, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 129, + 547, + 194 + ], + "spans": [ + { + "bbox": [ + 307, + 129, + 547, + 194 + ], + "type": "text", + "content": "[68] Zirui Wu, Tianyu Liu, Liyi Luo, Zhide Zhong, Jianteng Chen, Hongmin Xiao, Chao Hou, Haozhe Lou, Yuantao Chen, Runyi Yang, Yuxin Huang, Xiaoyu Ye, Zike Yan, Yongliang Shi, Yiyi Liao, and Hao Zhao. Mars: An instance-aware, modular and realistic simulator for autonomous driving. CICAI, 2023. 2, 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 307, + 196, + 546, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 196, + 546, + 228 + ], + "spans": [ + { + "bbox": [ + 307, + 196, + 546, + 228 + ], + "type": "text", + "content": "[69] Ziyang Xie, Junge Zhang, Wenye Li, Feihu Zhang, and Li Zhang. S-nerf: Neural radiance fields for street views. arXiv preprint arXiv:2303.00749, 2023. 2, 3, 8" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 230, + 546, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 230, + 546, + 262 + ], + "spans": [ + { + "bbox": [ + 307, + 230, + 546, + 262 + ], + "type": "text", + "content": "[70] Yinda Xu and Lidong Yu. Drl-based trajectory tracking for motion-related modules in autonomous driving. arXiv preprint arXiv:2308.15991, 2023. 4" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 264, + 546, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 264, + 546, + 308 + ], + "spans": [ + { + "bbox": [ + 307, + 264, + 546, + 308 + ], + "type": "text", + "content": "[71] Kairui Yang, Enhui Ma, Jibin Peng, Qing Guo, Di Lin, and Kaicheng Yu. Bevcontrol: Accurately controlling street-view elements with multi-perspective consistency via bev sketch layout. arXiv preprint arXiv:2308.01661, 2023. 2, 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 309, + 547, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 309, + 547, + 373 + ], + "spans": [ + { + "bbox": [ + 308, + 309, + 547, + 373 + ], + "type": "text", + "content": "[72] Zhenpei Yang, Yuning Chai, Dragomir Anguelov, Yin Zhou, Pei Sun, Dumitru Erhan, Sean Rafferty, and Henrik Kretzschmar. Surfelgan: Synthesizing realistic sensor data for autonomous driving. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11118-11127, 2020. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 375, + 547, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 375, + 547, + 430 + ], + "spans": [ + { + "bbox": [ + 308, + 375, + 547, + 430 + ], + "type": "text", + "content": "[73] Ze Yang, Yun Chen, Jingkang Wang, Sivabalan Manivasagam, Wei-Chiu Ma, Anqi Joyce Yang, and Raquel Urtasun. Unisim: A neural closed-loop sensor simulator. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1389-1399, 2023. 2, 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 431, + 547, + 485 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 431, + 547, + 485 + ], + "spans": [ + { + "bbox": [ + 308, + 431, + 547, + 485 + ], + "type": "text", + "content": "[74] Jinsong Zhang, Kalyan Sunkavalli, Yannick Hold-Geoffroy, Sunil Hadap, Jonathan Eisenman, and Jean-François Lalonde. All-weather deep outdoor lighting estimation. In Proceedings of the IEEE/CVF conference on Computer Vision and Pattern Recognition, pages 10158–10166, 2019. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 487, + 546, + 529 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 487, + 546, + 529 + ], + "spans": [ + { + "bbox": [ + 308, + 487, + 546, + 529 + ], + "type": "text", + "content": "[75] Yichao Zhou, Jingwei Huang, Xili Dai, Linjie Luo, Zhili Chen, and Yi Ma. HoliCity: A city-scale data platform for learning holistic 3D structures. 2020. arXiv:2008.03286 [cs.CV]. 6, 8" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 532, + 547, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 532, + 547, + 586 + ], + "spans": [ + { + "bbox": [ + 308, + 532, + 547, + 586 + ], + "type": "text", + "content": "[76] Mingchen Zhuge, Haozhe Liu, Francesco Faccio, Dylan R Ashley, Robert Csordás, Anand Gopalakrishnan, Abdullah Hamdi, Hasan Abed Al Kader Hammoud, Vincent Herrmann, Kazuki Irie, et al. Mindstorms in natural language-based societies of mind. arXiv preprint arXiv:2305.17066, 2023. 3" + } + ] + } + ], + "index": 21 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "15087" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Effective Video Mirror Detection with Inconsistent Motion Cues/423d0156-aefb-4c1d-9839-9825d419ff81_content_list.json b/2024/Effective Video Mirror Detection with Inconsistent Motion Cues/423d0156-aefb-4c1d-9839-9825d419ff81_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..9159c9760f4592b9263c06716514e8ec0eeb7640 --- /dev/null +++ b/2024/Effective Video Mirror Detection with Inconsistent Motion Cues/423d0156-aefb-4c1d-9839-9825d419ff81_content_list.json @@ -0,0 +1,1482 @@ +[ + { + "type": "text", + "text": "Effective Video Mirror Detection with Inconsistent Motion Cues", + "text_level": 1, + "bbox": [ + 161, + 130, + 808, + 150 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Alex Warren1, Ke Xu2, Jiaying Lin2, Gary K.L. Tam1, Rynson W.H. Lau1,2", + "bbox": [ + 169, + 179, + 799, + 199 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Department of Computer Science, Swansea University1 and City University of Hong Kong2", + "bbox": [ + 171, + 200, + 795, + 217 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "alex.warren@swansea.ac.uk, kkangwing@gmail.com, jiayinlin5-c@my.cityu.edu.hk", + "bbox": [ + 148, + 219, + 820, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "k.l.tam@swansea.ac.uk, Rynson.Lau@cityu.edu.hk", + "bbox": [ + 277, + 237, + 687, + 250 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 286, + 313, + 301 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Image-based mirror detection has recently undergone rapid research due to its significance in applications such as robotic navigation, semantic segmentation and scene reconstruction. Recently, VMD-Net was proposed as the first video mirror detection technique, by modeling dual correspondences between the inside and outside of the mirror both spatially and temporally. However, this approach is not reliable, as correspondences can occur completely inside or outside of the mirrors. In addition, the proposed dataset VMD-D contains many small mirrors, limiting its applicability to real-world scenarios. To address these problems, we developed a more challenging dataset that includes mirrors of various shapes and sizes at different locations of the frames, providing a better reflection of real-world scenarios. Next, we observed that the motions between the inside and outside of the mirror are often inconsistent. For instance, when moving in front of a mirror, the motion inside the mirror is often much smaller than the motion outside due to increased depth perception. With these observations, we propose modeling inconsistent motion cues to detect mirrors, and a new network with two novel modules. The Motion Attention Module (MAM) explicitly models inconsistent motions around mirrors via optical flow, and the Motion-Guided Edge Detection Module (MEDM) uses motions to guide mirror edge feature learning. Experimental results on our proposed dataset show that our method outperforms state-of-the-arts. The code and dataset are available at https://github.com/ AlexAnthonyWarren/MG-VMD.", + "bbox": [ + 76, + 316, + 473, + 758 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 784, + 209, + 800 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Mirrors and reflective surfaces are abundant in our surroundings. Successful detection of mirrors underpins many vision applications such as autonomous drone navigation and vehicle vision systems [16]. When a scene contains mirrored regions, mirror identification is critical to reducing errors in applications such as salient object de", + "bbox": [ + 75, + 810, + 468, + 902 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/0dfd4ee24860de30e9d715bf6232ccd9c165459ff99a7c121c670413d046e0af.jpg", + "image_caption": [ + "Figure 1. We propose to model motion inconsistency as a cue for mirror detection. The $1^{st}$ column shows the inconsistent motions (depicted with arrows). The locations of these cues align well with the mirror regions. The $2^{nd}$ column shows the optical flow computed between frame N ( $3^{rd}$ column), and frame N-1 from respective videos in our new dataset. Our method ( $5^{th}$ column) predicts more reliable and consistent mirror regions, outperforming VMD-Net [12] ( $4^{th}$ column). The $6^{th}$ column is the ground truth." + ], + "image_footnote": [], + "bbox": [ + 508, + 287, + 883, + 489 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "tection [25, 26], semantic segmentation [16, 21], scene reconstruction[3, 4] and NeRF modeling [6, 31].", + "bbox": [ + 496, + 625, + 892, + 655 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Mirror detection involves binary classification of individual pixels in RGB images to determine if they are inside or outside of the mirror. It is a challenging research as mirrors typically do not have their own visual appearance. They instead reflect objects within their surroundings. This visual attribute makes mirrors difficult to detect.", + "bbox": [ + 496, + 657, + 890, + 747 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "There are a few image-based and one video-based mirror detection methods in the literature. Yang et al. [30] first introduced a mirror detection method using contrast and semantics. Lin et al. [11] introduced a PMD-Net that utilizes edge detection and leverages contextual contrasting objects inside and outside of mirrored regions to detect mirrored regions. Tan et al. [22] detect mirrors using Visual Chirality Cue [13] that encodes mirror-like symmetrical properties at the pixel level. Mei et al. [15] factor depth estimation into their mirror detection process and show good results", + "bbox": [ + 496, + 750, + 893, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "17244", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "in RGB and RGB-D mirror datasets. Recently, Lin et al. [12] introduced the first video mirror detection method using inter-frame and intra-frame dual correspondences.", + "bbox": [ + 75, + 90, + 468, + 136 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Despite these exciting results, we observe research gaps and limitations in these works, especially in video mirror detection. The state-of-the-art VMD-Net [12] models dual correspondences inside and outside of mirrors to aid in the detection of mirror regions in videos. However, the approach is not reliable as correspondences can appear completely inside or outside of mirrors as discussed in [22]. In addition, the proposed VMD-D dataset contains primarily small mirrors, with small inter-frame motion, mirror coverage and content variation. This limits its applicability to mirror detection in real-world scenarios.", + "bbox": [ + 75, + 138, + 468, + 303 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address these issues, we first developed a new, more challenging dataset, Mirror Motion Dataset (MMD). The dataset contains 37 ( $\\approx$ 9-second) videos with substantial motion, each accompanied by manual mirror annotations. To address the small mirror issue in VMD-D, we capture a high variance of mirror sizes and shapes at different spatial locations, generalizing to real-world scenarios. We further include challenging examples, including low-contrast environments, multi-/large/full-mirror scenes and under different lighting conditions. Next, we observe that the motions between the inside and outside of mirrors are often inconsistent. For instance, when moving in front of a mirror, the motion inside is often smaller than outside due to the increased depth perception. The inconsistent motion cues (depicted by arrows in the $1^{st}$ column in Figure 1) also apply in challenging scenarios like low-contrast and multi-mirror scenes. These observations motivate us to ask the question: Can we apply the motion inconsistency cue to improve mirror detection in challenging complex video scenes?", + "bbox": [ + 75, + 306, + 468, + 592 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Neuroscience studies suggest that humans often rely on dynamic perceptual cues to identify mirrored regions in daily life. Tamura et al. [18] [19] show that humans are capable of distinguishing mirrors in dynamic scenes involving rotational movement. [20] highlights that rotational, parallax, forward and backward motions are exploitable motion inconsistencies for mirror detection. To our knowledge, such motion inconsistency cues have not been utilized before for the video mirror detection task.", + "bbox": [ + 75, + 595, + 468, + 729 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To this end, we propose a novel network to model motion inconsistency for mirror detection, with two novel modules. First, the Motion Attention Module (MAM) predicts mirror regions by exploiting inconsistent motion cues obtained from optical flow fields. Second, the Motion-guided Edge Detection Module (MEDM) further uses motion to guide mirror edge detection. Experimental results show that the proposed network outperforms relevant state-of-the-art methods. To summarise, our main contributions include:", + "bbox": [ + 75, + 732, + 468, + 867 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "- We propose a novel deep-learning video mirror detection method based on the motion inconsistency cue.", + "bbox": [ + 76, + 869, + 468, + 900 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Our network includes two novel modules. Motion Attention Module (MAM) attentively predicts mirror regions using motion inconsistency in the optical flow fields. Motion-guided Edge Detection Module (MEDM) uses the motions to guide mirror edge feature extraction.", + "- We further present a new video dataset (MMD). Compared to VMD-D, our dataset contains mirrors of various shapes and size at different locations of the images. It also covers various scenes, e.g., low-light scenes, and multi-large/full mirror scenes. It is a more challenging dataset, more suitable for real-life scenarios.", + "- Extensive evaluations show that the proposed model outperforms existing methods on video mirror detection." + ], + "bbox": [ + 500, + 90, + 890, + 286 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 500, + 301, + 648, + 316 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Video Mirror Detection. VMD-Net [12] introduced the first video mirror detection method. It employs dual correspondences across both inter-frame and intra-frame to predict mirrored regions. They further presented a video mirror dataset for this purpose. However, despite its demonstrated success, we observe two limitations. First, the idea of dual correspondences is not reliable as correspondences can occur completely inside or outside of the mirrors [22]. Second, their dataset contains an excess of small mirrors, which does not generalize well to real-world scenarios.", + "bbox": [ + 496, + 325, + 890, + 476 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "This paper aims to facilitate video mirror detection in the wild from two perspectives. First, we propose to model the inconsistent motion cues to learn more robust mirror representations. Second, we introduce a new benchmark dataset comprising videos with more complex scenarios.", + "bbox": [ + 496, + 478, + 890, + 551 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Image-Based Mirror Detection. [30] pioneered mirror detection by utilizing contextual contrasting features and introducing a mirror detection dataset. [15] refined detection using RGB-D data and presented the first RGB-D mirror dataset. [11] introduced a method leveraging contextual contrasts to relate mirror interior and exterior contents, along with an expanded image-based mirror dataset. [22] incorporated chirality cues for mirror detection. Although these work achieved good results, they mainly focus on single-image mirror detection and lack the capacity to identify mirrors within temporal videos.", + "bbox": [ + 496, + 553, + 890, + 718 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Video Salient Object Detection. Video Salient Object Detection (VSOD) seeks to identify the most visually prominent object in videos. Early VSOD studies [27] [28] utilized manually crafted features to detect these salient objects. In more recent research [9] [8] [9] [14] [23], the focus has shifted to deep learning models, which have demonstrated promising results in VSOD. These techniques however are not optimized for the video mirror detection task. This is because mirrored regions may not always represent the most visually significant objects within a scene. Our technique aims to identify mirrors in real-life scenarios.", + "bbox": [ + 496, + 719, + 890, + 883 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Optical Flow Estimation. Optical flow estimation fo", + "bbox": [ + 500, + 885, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "17245", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/f336c6d5b1be67764b040e70639b9b997823bba7bad736a5844ac93c660c1564.jpg", + "image_caption": [ + "Figure 2. Snapshots of our proposed Motion Mirror Dataset (MMD), with pixel-level mirror and edge annotations." + ], + "image_footnote": [], + "bbox": [ + 86, + 88, + 883, + 275 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/e2803215de5f929742c0998604c203196ac6331bdcb616cfd48c907e9dd1a766.jpg", + "image_caption": [ + "Figure 3. Cumulative Distributions. Figure 5. Color Contrast." + ], + "image_footnote": [], + "bbox": [ + 81, + 305, + 465, + 474 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "cuses on extracting per-pixel motion from video sequences. Early efforts [5] used hand-crafted features of image brightness between consecutive frames. In recent studies, convolutional deep learning and transformer-based approaches have made significant improvements [24] [7]. In our context, optical flow estimation is not inherently applicable to video mirror detection and cannot serve as a direct detection method. This study aims to leverage deep-learning optical flow as a way to model inconsistent motion cues, which enables a robust end-to-end video mirror detection framework.", + "bbox": [ + 75, + 510, + 468, + 660 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Proposed Dataset", + "text_level": 1, + "bbox": [ + 76, + 678, + 246, + 695 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To create our Motion Mirror Dataset (MMD), we first recruited six volunteering participants (final year university students) to capture 37 nine-second smartphone videos from their surroundings (e.g., living rooms, bathrooms). Participants were instructed to maintain consistent motion around mirror regions while recording. We recorded videos under various lighting conditions, including daily light and low-light settings. The dataset was subsequently divided into non-overlapping training (4,653 samples) and testing (5,074 samples) sets. To accomplish this, we randomly split the 37 videos into the two sets (training and testing). Specifically 18 videos (4,653 samples) are put into the training set and 19 videos (5,074 samples) are put into the testing", + "bbox": [ + 75, + 704, + 470, + 902 + ], + "page_idx": 2 + }, + { + "type": "table", + "img_path": "images/1d998f791eaa10f88eb6ebb4214696080fd2da362136af1cd862e87eb19c1030.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
DatasetSSIM ↑
VMD-D0.7950544833184201
Ours (MotionMirrorDataset)0.9033545507233643
", + "bbox": [ + 500, + 301, + 893, + 349 + ], + "page_idx": 2 + }, + { + "type": "table", + "img_path": "images/9e2b8104320d7e1bd0d2f1fdea6f1ed2cd106f0ef4860dc4d27e98631c0d65d7.jpg", + "table_caption": [ + "Table 1. SSIM of MMD and VMD-D. A higher SSIM indicates a more reasonable split of dataset." + ], + "table_footnote": [], + "table_body": "
DatasetAreas not covered by mirrors ↓
(Ours) MMD Training0.000061
(Ours) MMD Testing0.000015
VMD-D Training0.00038147
VMD-D Testing0.00024414
", + "bbox": [ + 511, + 398, + 870, + 474 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Table 2. Mirror Coverage. A smaller non-coverage rate indicates a more even spatial distribution of mirrors.", + "bbox": [ + 500, + 477, + 890, + 505 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "set. Each frame was meticulously annotated with pixel-level mirror masks. We applied a depth-first search algorithm to the pixel-level mirror masks to extract edge masks, which we then manually verified for ground truth accuracy. All videos in the dataset have a frame rate of 30 fps.", + "bbox": [ + 496, + 522, + 890, + 598 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Statistic Analysis. We compare our proposed MMD with VMD-D [12] in terms of Cumulative Distribution and Structural Similarity between training and testing sets, Mirror Size Distribution, and Area Coverage.", + "bbox": [ + 496, + 598, + 890, + 657 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Figure 3 shows the cumulative distribution of all mirrors in both training and testing sets. Compared to VMD-D, our MMD dataset contains more intricate scenarios. VMD-D [12] however is highly biased with majority of mirror regions fall in the center in the train/test sets whilst there are nearly none along the edge (darker color on left/right). The mirror regions of our proposed dataset (MMD) are spread out across the entire heatmap where mirrors are also found at the image edges. We further compute the Structural Similarity Index (SSIM) on the cumulative area distributions (Figure 3) directly to compare the structure between the whole training and testing sets. The higher SSIM of our MMD in Table 1 shows that the training and testing sets are similar in structure and fairer for evaluation purpose.", + "bbox": [ + 496, + 657, + 892, + 869 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Figure 4 shows the size distribution of mirrors in both MMD and VMD-D datasets. VMD-D contains a lot of", + "bbox": [ + 500, + 869, + 893, + 898 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "17246", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/228c0a33d6db9084e592569962c0eeff8d0207a5c8ea770bec02354033962aa8.jpg", + "image_caption": [ + "Figure 6. Method overview. Our MG-VMD network takes three adjacent frames $(\\mathrm{I}_{\\mathrm{N - 2}},\\mathrm{I}_{\\mathrm{N - 1}},\\mathrm{I}_{\\mathrm{N}})$ as inputs and extracts their multi-scale features via a shared ResNext-101 [29] backbone and DeepLabV3[2]. We apply the optical flow coherence block to process $\\mathrm{I}_{\\mathrm{N - 2}}$ and $\\mathrm{I}_{\\mathrm{N - 1}}$ to compute an optical flow map for $\\mathrm{I}_{\\mathrm{N - 1}}$ , followed by an optical flow vector field extrapolation to estimate the optical flow map for $\\mathrm{I}_{\\mathrm{N}}$ . These two optical flow maps are fed into our proposed MAM and MEDM modules to guide the mirror localization and mirror-edge detection processes, respectively. The proposed Motion Attention Module (MAM) learns to locate mirrors from optical flow fields, exploiting motion cues, attentively modeling motion inconsistency. The proposed Motion-guided Edge Detection Module (MEDM) learns to extract and refine mirror boundary edge features. Finally, a fusion refinement module is used to predict the mirror maps by fusing the predicted mirror boundary maps of MEDM, the mirror localization maps of MAM, and the extracted multi-scale image backbone features." + ], + "image_footnote": [], + "bbox": [ + 98, + 101, + 883, + 388 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "small mirrors with few/no large mirrors, which does not represent real-world scenarios. In contrast, our MMD dataset has a more reasonable distribution, containing a variety of small, medium and large sized mirrors, and therefore is more challenging than VMD-D.", + "bbox": [ + 75, + 523, + 470, + 599 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Table 2 shows the areas (ratio to the whole image) where there is no mirror coverage within the two datasets. The smaller ratios show that MMD contains mirrors that are more evenly spatially distributed across the whole image. MMD is more challenging and will introduce less spatial bias in the model learning.", + "bbox": [ + 75, + 601, + 468, + 691 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We further compare the color contrast between mirror and non-mirror regions in all images between MMD and VMD-D. The histogram in Figure 5 shows that MMD contains more images with lower color contrast between mirror and non-mirror regions. Lower color contrast implies a decrease in the amount of contextual information available for a model to leverage, making it more challenging to detect mirrors. The $4^{\\text{th}}$ column of Figure 2 shows one example. These show that MMD is more challenging than VMD-D.", + "bbox": [ + 75, + 691, + 470, + 829 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Proposed Method", + "text_level": 1, + "bbox": [ + 76, + 844, + 250, + 861 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our proposed method, Motion-guided Video Mirror Detection (MG-VMD), aims to enhance mirror detection by", + "bbox": [ + 75, + 869, + 470, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "modelling motion inconsistency cues. Drawing inspiration from our own observations and neuroscience studies, we hypothesise that modeling motion inconsistency cues can improve the accuracy of mirror detection, especially in dynamic video scenarios.", + "bbox": [ + 496, + 523, + 890, + 599 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The core idea behind MG-VMD is to utilize optical flow to model motion inconsistency and use it to guide the extraction of multi-scale image and edge features, allowing for the temporal detection of mirrored regions in videos. To achieve this, we introduce two key modules: the Motion Attention Module (MAM) and the Motion-guided Edge Detection Module (MEDM).", + "bbox": [ + 496, + 604, + 890, + 710 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Specifically, the Motion Attention Module (MAM) is designed to explicitly predict mirror locations by modelling motion inconsistency cues from the underlying optical flow fields. This module enhances video mirror prediction by providing spatial contextual information through the multiscale image features and motion cues. It informs the subsequent stages of the model about the location of mirror regions in the video.", + "bbox": [ + 496, + 714, + 890, + 835 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The Motion-guided Edge Detection Module (MEDM) is dedicated to predicting mirror boundaries, contributing to the enhancement of overall mirror region predictions. This module takes advantage of image features and employs mo", + "bbox": [ + 496, + 839, + 893, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "17247", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/4214f93536c8abbf728bdc63c5971369f1817a668c77eb59c7ad31826e05faa4.jpg", + "image_caption": [ + "Figure 7. The proposed MAM (Motion Attention Module) aims to model the inconsistent motion cues to localize the mirrors." + ], + "image_footnote": [], + "bbox": [ + 84, + 90, + 467, + 181 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "tion inconsistency cues to guide the detection of features at the edges, specifically tailored for mirror detection.", + "bbox": [ + 75, + 233, + 468, + 262 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The Fusion Refinement Module, inspired from PMD-Net [11], further takes the MAM predicted mirror locations and the MEDM refined mirror boundaries to predict mirrors. The integration plays a pivotal role in improving mirror prediction by alleviating errors arising from the limited capability of edge detection or inconsistent motion cues alone. It overcomes the limitations of individual modules and provides spatial and temporal context for accurate mirror prediction.", + "bbox": [ + 75, + 263, + 468, + 398 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In Figure 6 (Architecture Overview), the MG-VMD processes adjacent images $(\\mathrm{I}_{\\mathrm{N - 2}},\\mathrm{I}_{\\mathrm{N - 1}},\\mathrm{I}_{\\mathrm{N}})$ as inputs. Optical flow field extrapolation is employed to enhance coherency and address temporal errors between adjacent branches (N-1 and N). The model utilizes a shared ResNext-101 [29] backbone and DeepLabV3 [2] to extract multi-scale image features. The outputs include predicted mirror binary maps $(\\mathbf{M}_{\\mathrm{N - 1}}$ and $\\mathbf{M}_{\\mathrm{N}}$ ), predicted mirror-boundary edge maps $(\\mathbf{E}_{\\mathrm{N - 1}}$ and $\\mathbf{E}_{\\mathrm{N}}$ ), and mirror motion predictions $(\\mathbf{M}\\mathbf{p}_{\\mathrm{N - 1}}$ and $\\mathbf{M}\\mathbf{p}_{\\mathrm{N}}$ ). The collaboration of MAM, MEDM, and the Fusion Refinement Module ensures accurate and robust video mirror detection by considering both spatial and temporal aspects of mirror cues.", + "bbox": [ + 75, + 398, + 468, + 595 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Optical Flow Coherence Block", + "text_level": 1, + "bbox": [ + 76, + 604, + 346, + 619 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our architecture leverages a pre-trained optical flow model. Similar to previous work [10], our optical flow block utilizes frozen weights. However, optical flow may become unreliable in scenarios with low contrast in image details or when images become saturated. To achieve a simpler design while obtaining more stable optical flow, we introduce our Optical Flow Coherence Block, inspired by the work in [24] for motion estimation. This block extracts motion and extends it using linear extrapolation. The extrapolated optical flow motion, denoted as $flo$ , is then utilized in our subsequent motion-guided modules, MAM and MEDM. We empirically find that this simple design outperforms alternatives (refer to Section 6 of the Supplemental) whilst enhancing coherency.", + "bbox": [ + 75, + 628, + 467, + 839 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Specifically, the block calculates motion vectors $f(N - 2,N - 1)$ between frames $\\mathrm{I}_{\\mathrm{N - 2}}$ and $\\mathrm{I}_{\\mathrm{N - 1}}$ . These vectors are then extrapolated by one frame to estimate motion $f^{\\prime}(N - 1,N)$ between $\\mathrm{I}_{\\mathrm{N - 1}}$ and $\\mathrm{I}_{\\mathrm{N}}$ . By using extrapolation, we can calculate the motion vectors of the frame $\\mathrm{I}_{\\mathrm{N - 1}}$ and $\\mathrm{I}_{\\mathrm{N}}$ .", + "bbox": [ + 75, + 839, + 467, + 901 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/ef55c47ca706413ecfab9687706bb2aaf8664b6bcf40210a82c483e9d4481867.jpg", + "image_caption": [ + "Figure 8. Our proposed MEDM (Motion-Guided Edge Detection Module) aims to detect the mirror boundaries based on the captured motion cues in the optical flow vector fields." + ], + "image_footnote": [], + "bbox": [ + 504, + 89, + 890, + 191 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "tion, we connect the motion information from the frame at N-2 to N-1, and then to N. This connection helps maintain a smooth and consistent flow of motion information, reducing errors from neighboring frames and sudden changes in motion, especially around mirror regions.", + "bbox": [ + 498, + 252, + 890, + 327 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Motion Attention Module", + "text_level": 1, + "bbox": [ + 500, + 339, + 732, + 354 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Figure 7 shows the architecture of the Motion Attention Module (MAM). This module focuses on direct mirror prediction from the optical flow fields. It leverages self-attention to capture and refine better features, to model motion inconsistency cue and to inform mirror locations. The output is then fed to the Fusion Refinement Module for the final mirror refinement.", + "bbox": [ + 498, + 363, + 890, + 468 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Specifically, we initiate the MAM by embedding patches with positional information to create $flo_{\\text{embed}}$ . Subsequently, we guide $flo_{\\text{embed}}$ through six transformer blocks. Each block incorporates self-attention, treating $flo_{\\text{embed}}$ as Q, K, and V in the self-attention mechanism, with output denoted as $flo_{\\text{attend}}$ . To enhance its representation, $flo_{\\text{attend}}$ undergoes batch normalization, followed by processing through a feed-forward design.", + "bbox": [ + 498, + 470, + 890, + 590 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The output of the self-attention mechanism within one Transformer BlockLayer N is then passed to the subsequent Transformer BlockLayer N+1. Following this sequence, the output from Transformer BlockLayer 6 is channeled through a convolutional layer. The intentional use of six sequential transformer blocks serves a dual purpose: it equips the model to comprehend long-range dependencies, crucial for temporal relationships, and enables the capture of intricate spatial dependencies. This enhanced capability is particularly valuable for predicting spatial locations of the attended mirrors, denoted as $\\mathrm{Mp}_{\\mathrm{N - 1}}$ and $\\mathrm{Mp_N}$", + "bbox": [ + 498, + 592, + 890, + 758 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.3. Motion-guided Edge Detection Module", + "text_level": 1, + "bbox": [ + 500, + 768, + 831, + 786 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Figure 8 shows the architecture of the Motion-guided Edge Detection Module (MEDM). Unlike previous mirror edge detection in [11, 22] that only exploits image features, we also leverage the inconsistent motion cues captured in the optical flow field to guide the mirror edge detection process.", + "bbox": [ + 498, + 792, + 890, + 869 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "MEDM processes three inputs: $\\mathrm{I}_{\\mathrm{low}}$ (a 2nd-level multiscale image feature), $\\mathrm{I}_{\\mathrm{high}}$ (a 5th-layer multi-scale image", + "bbox": [ + 500, + 869, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "17248", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "feature), and $flo$ (the optical flow field for motion guidance). The process begins with convolutional layers that extract and predict edge features—referred to as $E_N$ or $E_{N-1}$ —from $\\mathrm{I}_{\\mathrm{low}}$ , $\\mathrm{I}_{\\mathrm{high}}$ , and $flo$ .", + "bbox": [ + 75, + 90, + 468, + 151 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The cross-attention mechanism involves matrix multiplication between the transformed $flo$ and $E_N$ or $E_{N - 1}$ . A Softmax activation determines the cross-attended beta value between $flo$ and $E_N$ or $E_{N - 1}$ . This beta value, representing the attention weight, is then used to focus on $flo$ , determining the contribution of each element in the optical flow field to the final cross-attended motion-guided edge feature. We specifically employ a cross-attention mechanism in this step to guide the extracted edge feature with the optical flow field. The motion-guided edge feature undergoes a convolutional pass, producing the output: the motion-guided edge feature map, $E_N$ or $E_{N - 1}$ , respectively.", + "bbox": [ + 75, + 152, + 472, + 333 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.4. Fusion Refinement Module", + "text_level": 1, + "bbox": [ + 76, + 340, + 321, + 354 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We refine the motion-guided edge detection features from MEDM and the motion-based mirror predictions from MAM. This refining step is crucial to address limitations in each module (MEDM, MAM) and ensure accurate mirror prediction with both spatial and temporal context.", + "bbox": [ + 75, + 363, + 468, + 439 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The process involves fusion-refinement, where model information is combined, refined, and fed into a convolutional binary classification head. This produces per-pixel binary classification maps, named $\\mathbf{M}_{\\mathrm{N - 1}}$ and $\\mathbf{M}_{\\mathrm{N}}$ , using two instances of the Fusion Refinement Module. Our approach is inspired by PMD-Net's [11] refinement strategy, adapted here to improve multiple modalities and guide mirror prediction for $\\mathbf{M}_{\\mathrm{N - 1}}$ and $\\mathbf{M}_{\\mathrm{N}}$ .", + "bbox": [ + 75, + 439, + 468, + 560 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.5. Loss Function", + "text_level": 1, + "bbox": [ + 76, + 569, + 222, + 583 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To train the model, we use the loss function $(\\mathcal{L}_{\\mathrm{Final}})$ :", + "bbox": [ + 76, + 592, + 421, + 607 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {A} = \\left(\\mathcal {L} _ {\\text {M i r r o r A}} \\cdot \\alpha\\right) + \\left(\\mathcal {L} _ {\\text {M o t i o n A}} \\cdot \\zeta\\right) + \\left(\\mathcal {L} _ {\\text {E d g e A}} \\cdot \\gamma\\right)\n$$\n", + "text_format": "latex", + "bbox": [ + 125, + 617, + 459, + 633 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {B} = \\left(\\mathcal {L} _ {\\text {M i r r o r B}} \\cdot \\alpha\\right) + \\left(\\mathcal {L} _ {\\text {M o t i o n B}} \\cdot \\zeta\\right) + \\left(\\mathcal {L} _ {\\text {E d g e B}} \\cdot \\gamma\\right)\n$$\n", + "text_format": "latex", + "bbox": [ + 125, + 636, + 457, + 651 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {\\text {C o h e r e n c e}} = (\\text {M i r r o r} _ {\\text {C o h e r e n c e}} \\\\ + \\operatorname {E d g e} _ {\\text {C o h e r e n c e}} + \\operatorname {M o t i o n} _ {\\text {C o h e r e n c e}}) \\cdot \\delta \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 86, + 656, + 411, + 690 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {F i n a l}} = \\mathcal {L} _ {\\text {C o h e r e n c e}} + \\mathcal {L} _ {A} + \\mathcal {L} _ {B}\n$$\n", + "text_format": "latex", + "bbox": [ + 112, + 695, + 315, + 709 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We discuss the loss component below:", + "bbox": [ + 96, + 719, + 351, + 734 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Temporal Loss $(\\mathcal{L}_{\\mathrm{Coherence}})$ : To make predictions across consecutive frames more consistent over time, we use BCELossWithLogits (binary cross-entropy) between $\\mathbf{M}_{\\mathrm{N - 1}}$ and $\\mathbf{M_N}$ , $\\mathbf{M}\\mathfrak{p}_{\\mathrm{N - 1}}$ and $\\mathbf{M}\\mathfrak{p}_{\\mathrm{N}}$ , as well as $\\mathbf{E}_{\\mathrm{N - 1}}$ and $\\mathbf{E}_{\\mathrm{N}}$ . We set $\\delta$ to 1, calling this loss $\\mathcal{L}_{\\mathrm{Coherence}}$ . This loss is aimed at reducing temporal inaccuracies in edge prediction, motion mirror location, and mirror prediction between frames in a video. It ensures that predictions from our model's combined features remain consistent over time, addressing a current limitation in video mirror detection.", + "bbox": [ + 75, + 750, + 468, + 898 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Mirror Binary Losses $(\\mathcal{L}_{\\mathrm{MirrorA}}$ and $\\mathcal{L}_{\\mathrm{MirrorB}})$ : We use two BCELossWithLogits loss functions to predict mirrored regions in $\\mathrm{M_{N - 1}}$ $(\\mathcal{L}_{\\mathrm{MirrorA}})$ and $\\mathrm{M_N}$ $(\\mathcal{L}_{\\mathrm{MirrorB}})$ . These losses are then compared with the actual ground truth mirror maps. We set $\\alpha$ to 3 for these losses, emphasizing the importance of accurately detecting mirrors in the model.", + "bbox": [ + 496, + 90, + 890, + 181 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Motion Mirror Losses $(\\mathcal{L}_{\\mathrm{MotionA}}$ and $\\mathcal{L}_{\\mathrm{MotionB}})$ : Two additional BCELossWithLogits loss functions are used to predict mirror regions from motion-guided image features $\\mathrm{Mp_{N - 1}}$ $(\\mathcal{L}_{\\mathrm{MotionA}})$ and $\\mathrm{Mp_N}$ $(\\mathcal{L}_{\\mathrm{MotionB}})$ . These losses are then compared with the ground-truth mirror maps. We set $\\zeta$ to 1 for both of these losses.", + "bbox": [ + 496, + 181, + 890, + 271 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Mirror Edge Losses $(\\mathcal{L}_{\\mathrm{EdgeA}}$ and $\\mathcal{L}_{\\mathrm{EdgeB}})$ : Two BCELossWithLogits loss functions are utilized for predicting mirror boundaries in $\\mathrm{E_{N - 1}}$ $(\\mathcal{L}_{\\mathrm{EdgeA}})$ and $\\mathrm{E_N}$ $(\\mathcal{L}_{\\mathrm{EdgeB}})$ . These losses are compared against the ground-truth mirror edge maps. We set $\\gamma = 2$ for each of these losses, emphasizing the precision of mirror edge prediction.", + "bbox": [ + 496, + 272, + 890, + 363 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 500, + 375, + 630, + 392 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1. Setups and Metrics", + "text_level": 1, + "bbox": [ + 500, + 400, + 684, + 416 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We implemented MG-VMD using PyTorch, trained and validated it against comparative models on a NVIDIA RTX 3090 GPU. In the data preprocessing phase, all images, ground-truth mirror maps, and edge maps were resized to a resolution of $224 \\times 224$ . Our model's ResNext-101 backbone and DeepLabV3 parameters were initialized using the pretrained ResNext-101 backbone and pretrained DeepLabV3 weights from VMD-Net [12].", + "bbox": [ + 496, + 422, + 890, + 542 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We utilized a Stochastic Gradient Descent Optimizer with a starting learning rate of 9e-3, a momentum of 1.9, and a weight decay of 5e-4. Then, we applied an adaptive learning rate to the optimizer, interpolating the learning rate from 9e-3 to 3e-3 across epochs 1 to 15. The model underwent 15 epochs of training with a batch size of 8.", + "bbox": [ + 496, + 544, + 890, + 635 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For validation, we employed four quantitative metrics: Mean Absolute Error (MAE↓), F-measure $(F_{\\beta}\\uparrow)$ , Accuracy↑, and Intersection over Union (IoU↑). F-measure $(F_{\\beta}\\uparrow)$ is computed as follows:", + "bbox": [ + 496, + 635, + 890, + 696 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nF _ {\\beta} = \\frac {1 + \\beta^ {2} (p r e c i s i o n * r e c a l l)}{\\beta^ {2} + r e c a l l}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 576, + 705, + 890, + 738 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\beta$ is set to 0.3, as suggested by previous research [11] [1]. The F-measure $(\\mathrm{F}_{\\beta}\\uparrow)$ assesses the harmonic mean between precision and recall.", + "bbox": [ + 496, + 747, + 890, + 792 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.2. Experimental Results", + "text_level": 1, + "bbox": [ + 500, + 801, + 700, + 816 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In Table 3 and Figure 9, we evaluate our proposed technique against 8 state-of-the-art methods in Video Salient Object Detection, and Image-Based/Video Mirror Detection. Using their respective pre-trained weights, we fine-tune and validate these results on our proposed MMD dataset..", + "bbox": [ + 496, + 824, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "17249", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/69c2078ac8b2b3b5ea2c04258db100265172934d33da06687524b04d03907fe2.jpg", + "image_caption": [ + "Figure 9. Qualitative results table comparing our proposed method against state-of-the-art Video Salient Object Detection, and Image-Based/Video Mirror Detection FS-Net [8], MGA [10], ALGRF [23], F3Net [9], UFO[17], MirrorNet[30], PMD-Net [11], VMD-Net[12] trained and validated on our proposed MMD dataset." + ], + "image_footnote": [], + "bbox": [ + 76, + 87, + 916, + 566 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.3. Comparison Against State-of-the-art", + "text_level": 1, + "bbox": [ + 76, + 626, + 395, + 643 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Our MG-VMD achieves better performance in Mean Absolute Error (MAE↓), F-measure ( $F_{\\beta} \\uparrow$ ), Accuracy↑, and Intersection over Union (IoU↑) when compared to state-of-the-art Video Salient Object Detection, Image-Based Mirror Detection and Video Mirror Detection methods fine-tuned and validated on our proposed Motion Mirror Dataset. We attribute the better performance to the exploiting of motion inconsistency cues across adjacent frames through our novel modules: Motion-Guided Edge Detection Module (MEDM) and Motion Attention Module (MAM). Figure 9 shows that our method is able to detect mirrors well in complex (e.g., low contrast) scenarios, outperforming state-of-the-arts.", + "bbox": [ + 75, + 654, + 468, + 851 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We also evaluate the generalizability of our model. In the Supplemental, we show results comparing MG-VMD and VMD-Net trained on our MMD dataset and tested on", + "bbox": [ + 75, + 854, + 470, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "VMD-D [12]. MG-VMD outperforms VMD-Net, showing that our model has slightly better/close generalization performances.", + "bbox": [ + 498, + 627, + 890, + 672 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.4. Ablation", + "text_level": 1, + "bbox": [ + 500, + 681, + 604, + 695 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We conducted an ablation study to assess the effectiveness of the proposed modules in MG-VMD. Initially, we removed both the Motion Attention Module and the Motion-guided Edge Detection Module, resulting in the baseline with only the Refinement module, which incorporates convolutional layers. This configuration is referred to as the \"baseline\" for comparative analysis. The rationale behind evaluating the baseline's performance without the Optical Flow Coherence Block, Motion-Guided Edge Detection Module, and Motion Attention Module was to quantitatively gauge the efficacy of our deep-learning approach without leveraging inconsistent motion cues.", + "bbox": [ + 496, + 703, + 890, + 885 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 4 shows the performance of the ablated models:", + "bbox": [ + 519, + 885, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "17250", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/598b38fe10e1eafb9eda1568ce42afdb47fe6255f138d1c07f67862787156b75.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelsAccuracy↑MAE↓Fβ↑IoU↑
MGA0.5660370.4339630.4891760.281285
UFO0.7967970.2032040.7974470.598641
FSNet0.8532870.1467130.8377320.707848
ALGRF0.7237720.2762290.6928290.502814
F3Net0.8388630.1611370.8473210.670618
PMD-Net0.7409090.2590910.8472110.424085
MirrorNet0.8352270.1647720.8389870.666242
VMDNet0.8544150.1455850.8123470.722634
Ours0.8725320.1274680.8694190.725130
", + "bbox": [ + 81, + 88, + 467, + 233 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/18322604f24696b3b464cd9d1bb229b3f6e63b0ef66dfc8a1ef1b4c375ecbf88.jpg", + "table_caption": [ + "Table 3. Quantitative results table comparing our proposed method against state-of-the-art Video Salient Object Detection, and Image-Based/Video Mirror Detection FS-Net [8], MGA [10], AL-GRF [23], F3Net [9], UFO [17], PMD-Net [11], MirrorNet[30], VMD-Net[12] methods trained and validated on our proposed MMD dataset. Red, Blue and Green indicate the best, second and third best performances, respectively." + ], + "table_footnote": [], + "table_body": "
ModelsAccuracy↑MAE↓Fβ↑IoU↑
baseline0.8575190.1424810.8375780.700855
b + MEDM0.8581760.1418240.8478530.707350
b + MAM0.8662180.1337820.8416140.713005
Ours0.8725320.1274680.8694190.725130
", + "bbox": [ + 81, + 349, + 465, + 422 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 4. Ablation Results: Qualitative results table comparing our proposed method (MG-VMD (Ours)) against ablated versions of our model: b + MAM, b + MEDM, and baseline (b), trained and validated on our proposed MMD dataset. Red, Blue and Green indicate the best, second and third-best performances, respectively.", + "bbox": [ + 75, + 426, + 470, + 498 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "\"baseline (b)\", \"b + MEDM\", \"b + MAM\", and \"MG-VMD (Ours)\". Both MAM and MEDM exploit motion inconsistency cues, demonstrating better performance compared to the baseline. The full model combines the complementary outputs from MAM and MEDM, yielding the best results.", + "bbox": [ + 75, + 508, + 468, + 585 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Our baseline outperforms state-of-the-art VMD-Net [12] for two reasons. First, our model is fine-tuned on VMD-Net's [12] ResNext-101 backbone and DeepLabV3 weights (Section 5). Second, VMD-Net [12] assumes dual correspondence, but in real-life datasets, dual correspondence does not always occur. This shows the usefulness of our MMD dataset.", + "bbox": [ + 75, + 585, + 468, + 691 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 4 shows that MEDM and MAM each contributes to the model individually. Without MEDM, mirror edges become noisy (e.g., Figure 10 Column 4 Rows 2 and 4). Without MAM, MEDM may mistakenly detect object edges inside mirrors (e.g., cabinet edges in Figure 10 Column 3 Row 2), leading to inaccurate mirror detection. (Best viewed by zooming in.) Both modules use motion features, complementing each other and boost performance (Figure 10 Column 5, and Table 4).", + "bbox": [ + 75, + 691, + 470, + 829 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 76, + 844, + 194, + 859 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This paper introduces a novel deep-learning approach to tackle video mirror detection by leveraging inconsistent", + "bbox": [ + 75, + 869, + 470, + 900 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/15d4f9447aaa4fcb42a89d36b5392d61dfa4b40d163684e32178f2a405cc3c06.jpg", + "image_caption": [ + "Figure 10. Qualitative results of the ablated models." + ], + "image_footnote": [], + "bbox": [ + 501, + 88, + 893, + 303 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/0004acba8a3a80f61163a5505f1f5de033e94223b9cfba8b3d73e5eea875dea7.jpg", + "image_caption": [ + "Figure 11. Failure cases. Our Method may not be good in scenarios containing both mirror regions and windows, due to windows irregular motion compared to a scene." + ], + "image_footnote": [], + "bbox": [ + 501, + 338, + 890, + 453 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "motion cues within and outside mirrored regions. We also present a benchmark video mirror dataset featuring consistent inter-frame motion across long videos. Our experimental results demonstrate that our proposed MG-VMD deep-learning method outperforms state-of-the-arts in Video Mirror Detection, Image-Based Mirror Detection, and Video Salient Object Detection. Furthermore, we conclude that our proposed Motion Mirror Dataset (MMD) is challenging compared to state-of-the-art Video Mirror Datasets.", + "bbox": [ + 496, + 508, + 890, + 645 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Our method is not without limitation. Figure 11 shows a scenario with both mirror and window in a scene. We found that similar to mirror regions, windows can have different motions from their surroundings. Thus, our method tends to over-predict window regions as mirrors. We plan to improve our design to mitigate this in the future.", + "bbox": [ + 496, + 652, + 893, + 743 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements: Alex is supported by a Swansea GTA Research Scholarship. This project is in part supported by a GRF grant from the Research Grants Council of Hong Kong (Ref.: 11211223). We gratefully acknowledge the support of the HEFCW HERC fund (W21/21HE) for the provision of GPU equipment used in this research. For the purpose of Open Access, the author has applied a Creative Commons Attribution (CC BY) license to any Author Accepted Manuscript (AAM) version arising from this submission.", + "bbox": [ + 496, + 763, + 893, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "17251", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Radhakrishna Achanta, Sheila Hemami, Francisco Estrada, and Sabine Susstrunk. Frequency-tuned salient region detection. In CVPR, 2009. 6", + "[2] Liang-Chieh Chen, George Papandreou, Florian Schroff, and Hartwig Adam. Rethinking atrous convolution for semantic image segmentation, 2017. 4, 5", + "[3] Zheng Dong, Ke Xu, Ziheng Duan, Hujun Bao, Weiwei Xu, and Rynson W. H. Lau. Geometry-aware two-scale pifu representation for human reconstruction. In NeurIPS, 2022. 1", + "[4] Zheng Dong, Ke Xu, Yaoan Gao, Qilin Sun, Hujun Bao, Weiwei Xu, and Rynson W. H. Lau. Sailor: Synergizing radiance and occupancy fields for live human performance capture. ACM Trans. Graph., 2023. 1", + "[5] Gunnar Farnebäck. Two-frame motion estimation based on polynomial expansion. In Proceedings of the 13th Scandinavian Conference on Image Analysis, 2003. 3", + "[6] Yuan-Chen Guo, Di Kang, Linchao Bao, Yu He, and Song-Hai Zhang. Nerfren: Neural radiance fields with reflections. In CVPR, 2022. 1", + "[7] Eddy Ig, Nikolaus Mayer, Tonmoy Saikia, Margret Keuper, Alexey Dosovitskiy, and Thomas Brox. Flownet 2.0: Evolution of optical flow estimation with deep networks. In CVPR, 2017. 3", + "[8] Ge-Peng Ji, Keren Fu, Zhe Wu, Deng-Ping Fan, Jianbing Shen, and Ling Shao. Full-duplex strategy for video object segmentation. In ICCV, 2021. 2, 7, 8", + "[9] Qingming Huang Jun Wei, Shuhui Wang. F3net: Fusion, feedback and focus for salient object detection. In AAAI, 2020. 2, 7, 8", + "[10] Haofeng Li, Guanqi Chen, Guanbin Li, and Yizhou Yu. Motion guided attention for video salient object detection. In ICCV, 2019. 5, 7, 8", + "[11] Jiaying Lin, Guodong Wang, and Rynson W.H. Lau. Progressive mirror detection. In CVPR, 2020. 1, 2, 5, 6, 7, 8", + "[12] Jiaying Lin, Xin Tan, and Rynson W.H. Lau. Learning to detect mirrors from videos via dual correspondences. In CVPR, 2023. 1, 2, 3, 6, 7, 8", + "[13] Zhiqiu Lin, Jin Sun, Abe Davis, and Noah Snavely. Visual chirality. In CVPR, 2020. 1", + "[14] Jing Liu, Jiaxiang Wang, Weikang Wang, and Yuting Su. Ds-net: Dynamic spatiotemporal network for video salient object detection. arXiv:2012.04886, 2022. 2", + "[15] Haiyang Mei, Bo Dong, Wen Dong, Pieter Peers, Xin Yang, Qiang Zhang, and Xiaopeng Wei. Depth-aware mirror segmentation. In CVPR, 2021. 1, 2", + "[16] Anwesan Pal, Sayan Mondal, and Henrik I Christensen. \"looking at the right stuff\"-guided semantic-gaze for autonomous driving. In CVPR, 2020. 1", + "[17] Yukun Su, Jingliang Deng, Ruizhou Sun, Guosheng Lin, and Qingyao Wu. A unified transformer framework for group-based segmentation: Co-segmentation, co-saliency detection and video salient object detection. IEEE Trans. Multimedia, 2023. 7, 8", + "[18] Hideki Tamura, Maki Tsukuda, Hiroshi Higashi, and Shigeki Nakauchi. Perceptual segregation between mirror and glass" + ], + "bbox": [ + 78, + 114, + 467, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "material under natural and unnatural illumination. Journal of Vision, 2016. 2", + "[19] Hideki Tamura, Hiroshi Higashi, and Shigeki Nakauchi. Multiple cues for visual perception of mirror and glass materials. Journal of Vision, 2017. 2", + "[20] Hideki Tamura, Hiroshi Higashi, and Shigeki Nakauchi. Dynamic visual cues for differentiating mirror and glass. Scientific Reports, 2018. 2", + "[21] Xin Tan, Ke Xu, Ying Cao, Yiheng Zhang, Lizhuang Ma, and Rynson W. H. Lau. Night-time scene parsing with a large real dataset. IEEE Transactions on Image Processing, 2021. 1", + "[22] Xin Tan, Jiaying Lin, Ke Xu, Pan Chen, Lizhuang Ma, and Rynson W.H. Lau. Mirror detection with the visual chirality cue. IEEE Trans. Pattern Anal. Mach. Intell., 2022. 1, 2, 5", + "[23] Yi Tang, Yuanman Li, and Guoliang Xing. Video salient object detection via adaptive local-global refinement. arXiv:2104.14360, 2021. 2, 7, 8", + "[24] Zachary Teed and Jia Deng. Raft: Recurrent all-pairs field transforms for optical flow. In ECCV, 2020. 3, 5", + "[25] Xin Tian, Ke Xu, Xin Yang, Lin Du, Baocai Yin, and Rynson WH Lau. Bi-directional object-context prioritization learning for saliency ranking. In CVPR, 2022. 1", + "[26] Xin Tian, Ke Xu, Xin Yang, Baocai Yin, and Rynson Lau. Learning to detect instance-level salient objects using complementary image labels. Int. J. Comput. Vis., 2022. 1", + "[27] Pavel Tokmakov, Karteek Alahari, and Cordelia Schmid. Learning video object segmentation with visual memory. arXiv:1704.05737, 2017. 2", + "[28] Wenguan Wang, Xiankai Lu, Jianbing Shen, David Crandall, and Ling Shao. Zero-shot video object segmentation via attentive graph neural networks. In ICCV, 2019. 2", + "[29] Saining Xie, Ross Girshick, Piotr Dólár, Zhuowen Tu, and Kaiming He. Aggregated residual transformations for deep neural networks. In CVPR, 2017. 4, 5", + "[30] Xin Yang, Haiyang Mei, Ke Xu, Xiaopeng Wei, Baocai Yin, and Rynson W. H. Lau. Where is my mirror? In ICCV, 2019. 1, 2, 7, 8", + "[31] Junyi Zeng, Chong Bao, Rui Chen, Zilong Dong, Guofeng Zhang, Hujun Bao, and Zhaopeng Cui. Mirror-nerf: Learning neural radiance fields for mirrors with whitted-style ray tracing. In ACM MM, 2023. 1" + ], + "bbox": [ + 503, + 93, + 890, + 689 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "17252", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 8 + } +] \ No newline at end of file diff --git a/2024/Effective Video Mirror Detection with Inconsistent Motion Cues/423d0156-aefb-4c1d-9839-9825d419ff81_model.json b/2024/Effective Video Mirror Detection with Inconsistent Motion Cues/423d0156-aefb-4c1d-9839-9825d419ff81_model.json new file mode 100644 index 0000000000000000000000000000000000000000..a4ad7af1f974d248b4a6cc8a633c0f5b231abc38 --- /dev/null +++ b/2024/Effective Video Mirror Detection with Inconsistent Motion Cues/423d0156-aefb-4c1d-9839-9825d419ff81_model.json @@ -0,0 +1,1890 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.163, + 0.131, + 0.809, + 0.151 + ], + "angle": 0, + "content": "Effective Video Mirror Detection with Inconsistent Motion Cues" + }, + { + "type": "text", + "bbox": [ + 0.17, + 0.18, + 0.8, + 0.2 + ], + "angle": 0, + "content": "Alex Warren1, Ke Xu2, Jiaying Lin2, Gary K.L. Tam1, Rynson W.H. Lau1,2" + }, + { + "type": "text", + "bbox": [ + 0.173, + 0.201, + 0.796, + 0.218 + ], + "angle": 0, + "content": "Department of Computer Science, Swansea University1 and City University of Hong Kong2" + }, + { + "type": "text", + "bbox": [ + 0.149, + 0.22, + 0.821, + 0.234 + ], + "angle": 0, + "content": "alex.warren@swansea.ac.uk, kkangwing@gmail.com, jiayinlin5-c@my.cityu.edu.hk" + }, + { + "type": "text", + "bbox": [ + 0.278, + 0.238, + 0.688, + 0.251 + ], + "angle": 0, + "content": "k.l.tam@swansea.ac.uk, Rynson.Lau@cityu.edu.hk" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.287, + 0.314, + 0.303 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.318, + 0.474, + 0.759 + ], + "angle": 0, + "content": "Image-based mirror detection has recently undergone rapid research due to its significance in applications such as robotic navigation, semantic segmentation and scene reconstruction. Recently, VMD-Net was proposed as the first video mirror detection technique, by modeling dual correspondences between the inside and outside of the mirror both spatially and temporally. However, this approach is not reliable, as correspondences can occur completely inside or outside of the mirrors. In addition, the proposed dataset VMD-D contains many small mirrors, limiting its applicability to real-world scenarios. To address these problems, we developed a more challenging dataset that includes mirrors of various shapes and sizes at different locations of the frames, providing a better reflection of real-world scenarios. Next, we observed that the motions between the inside and outside of the mirror are often inconsistent. For instance, when moving in front of a mirror, the motion inside the mirror is often much smaller than the motion outside due to increased depth perception. With these observations, we propose modeling inconsistent motion cues to detect mirrors, and a new network with two novel modules. The Motion Attention Module (MAM) explicitly models inconsistent motions around mirrors via optical flow, and the Motion-Guided Edge Detection Module (MEDM) uses motions to guide mirror edge feature learning. Experimental results on our proposed dataset show that our method outperforms state-of-the-arts. The code and dataset are available at https://github.com/ AlexAnthonyWarren/MG-VMD." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.785, + 0.21, + 0.801 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.811, + 0.47, + 0.903 + ], + "angle": 0, + "content": "Mirrors and reflective surfaces are abundant in our surroundings. Successful detection of mirrors underpins many vision applications such as autonomous drone navigation and vehicle vision systems [16]. When a scene contains mirrored regions, mirror identification is critical to reducing errors in applications such as salient object de" + }, + { + "type": "image", + "bbox": [ + 0.509, + 0.288, + 0.885, + 0.49 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.495, + 0.895, + 0.607 + ], + "angle": 0, + "content": "Figure 1. We propose to model motion inconsistency as a cue for mirror detection. The \\(1^{st}\\) column shows the inconsistent motions (depicted with arrows). The locations of these cues align well with the mirror regions. The \\(2^{nd}\\) column shows the optical flow computed between frame N (\\(3^{rd}\\) column), and frame N-1 from respective videos in our new dataset. Our method (\\(5^{th}\\) column) predicts more reliable and consistent mirror regions, outperforming VMD-Net [12] (\\(4^{th}\\) column). The \\(6^{th}\\) column is the ground truth." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.626, + 0.893, + 0.656 + ], + "angle": 0, + "content": "tection [25, 26], semantic segmentation [16, 21], scene reconstruction[3, 4] and NeRF modeling [6, 31]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.658, + 0.892, + 0.748 + ], + "angle": 0, + "content": "Mirror detection involves binary classification of individual pixels in RGB images to determine if they are inside or outside of the mirror. It is a challenging research as mirrors typically do not have their own visual appearance. They instead reflect objects within their surroundings. This visual attribute makes mirrors difficult to detect." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.895, + 0.902 + ], + "angle": 0, + "content": "There are a few image-based and one video-based mirror detection methods in the literature. Yang et al. [30] first introduced a mirror detection method using contrast and semantics. Lin et al. [11] introduced a PMD-Net that utilizes edge detection and leverages contextual contrasting objects inside and outside of mirrored regions to detect mirrored regions. Tan et al. [22] detect mirrors using Visual Chirality Cue [13] that encodes mirror-like symmetrical properties at the pixel level. Mei et al. [15] factor depth estimation into their mirror detection process and show good results" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "17244" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.137 + ], + "angle": 0, + "content": "in RGB and RGB-D mirror datasets. Recently, Lin et al. [12] introduced the first video mirror detection method using inter-frame and intra-frame dual correspondences." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.139, + 0.47, + 0.304 + ], + "angle": 0, + "content": "Despite these exciting results, we observe research gaps and limitations in these works, especially in video mirror detection. The state-of-the-art VMD-Net [12] models dual correspondences inside and outside of mirrors to aid in the detection of mirror regions in videos. However, the approach is not reliable as correspondences can appear completely inside or outside of mirrors as discussed in [22]. In addition, the proposed VMD-D dataset contains primarily small mirrors, with small inter-frame motion, mirror coverage and content variation. This limits its applicability to mirror detection in real-world scenarios." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.307, + 0.47, + 0.593 + ], + "angle": 0, + "content": "To address these issues, we first developed a new, more challenging dataset, Mirror Motion Dataset (MMD). The dataset contains 37 (\\(\\approx\\)9-second) videos with substantial motion, each accompanied by manual mirror annotations. To address the small mirror issue in VMD-D, we capture a high variance of mirror sizes and shapes at different spatial locations, generalizing to real-world scenarios. We further include challenging examples, including low-contrast environments, multi-/large/full-mirror scenes and under different lighting conditions. Next, we observe that the motions between the inside and outside of mirrors are often inconsistent. For instance, when moving in front of a mirror, the motion inside is often smaller than outside due to the increased depth perception. The inconsistent motion cues (depicted by arrows in the \\(1^{st}\\) column in Figure 1) also apply in challenging scenarios like low-contrast and multi-mirror scenes. These observations motivate us to ask the question: Can we apply the motion inconsistency cue to improve mirror detection in challenging complex video scenes?" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.596, + 0.47, + 0.731 + ], + "angle": 0, + "content": "Neuroscience studies suggest that humans often rely on dynamic perceptual cues to identify mirrored regions in daily life. Tamura et al. [18] [19] show that humans are capable of distinguishing mirrors in dynamic scenes involving rotational movement. [20] highlights that rotational, parallax, forward and backward motions are exploitable motion inconsistencies for mirror detection. To our knowledge, such motion inconsistency cues have not been utilized before for the video mirror detection task." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.733, + 0.47, + 0.868 + ], + "angle": 0, + "content": "To this end, we propose a novel network to model motion inconsistency for mirror detection, with two novel modules. First, the Motion Attention Module (MAM) predicts mirror regions by exploiting inconsistent motion cues obtained from optical flow fields. Second, the Motion-guided Edge Detection Module (MEDM) further uses motion to guide mirror edge detection. Experimental results show that the proposed network outperforms relevant state-of-the-art methods. To summarise, our main contributions include:" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.47, + 0.901 + ], + "angle": 0, + "content": "- We propose a novel deep-learning video mirror detection method based on the motion inconsistency cue." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.092, + 0.891, + 0.166 + ], + "angle": 0, + "content": "- Our network includes two novel modules. Motion Attention Module (MAM) attentively predicts mirror regions using motion inconsistency in the optical flow fields. Motion-guided Edge Detection Module (MEDM) uses the motions to guide mirror edge feature extraction." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.168, + 0.891, + 0.256 + ], + "angle": 0, + "content": "- We further present a new video dataset (MMD). Compared to VMD-D, our dataset contains mirrors of various shapes and size at different locations of the images. It also covers various scenes, e.g., low-light scenes, and multi-large/full mirror scenes. It is a more challenging dataset, more suitable for real-life scenarios." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.258, + 0.891, + 0.287 + ], + "angle": 0, + "content": "- Extensive evaluations show that the proposed model outperforms existing methods on video mirror detection." + }, + { + "type": "list", + "bbox": [ + 0.5, + 0.092, + 0.891, + 0.287 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.302, + 0.65, + 0.318 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.327, + 0.892, + 0.477 + ], + "angle": 0, + "content": "Video Mirror Detection. VMD-Net [12] introduced the first video mirror detection method. It employs dual correspondences across both inter-frame and intra-frame to predict mirrored regions. They further presented a video mirror dataset for this purpose. However, despite its demonstrated success, we observe two limitations. First, the idea of dual correspondences is not reliable as correspondences can occur completely inside or outside of the mirrors [22]. Second, their dataset contains an excess of small mirrors, which does not generalize well to real-world scenarios." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.479, + 0.892, + 0.553 + ], + "angle": 0, + "content": "This paper aims to facilitate video mirror detection in the wild from two perspectives. First, we propose to model the inconsistent motion cues to learn more robust mirror representations. Second, we introduce a new benchmark dataset comprising videos with more complex scenarios." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.554, + 0.892, + 0.719 + ], + "angle": 0, + "content": "Image-Based Mirror Detection. [30] pioneered mirror detection by utilizing contextual contrasting features and introducing a mirror detection dataset. [15] refined detection using RGB-D data and presented the first RGB-D mirror dataset. [11] introduced a method leveraging contextual contrasts to relate mirror interior and exterior contents, along with an expanded image-based mirror dataset. [22] incorporated chirality cues for mirror detection. Although these work achieved good results, they mainly focus on single-image mirror detection and lack the capacity to identify mirrors within temporal videos." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.892, + 0.884 + ], + "angle": 0, + "content": "Video Salient Object Detection. Video Salient Object Detection (VSOD) seeks to identify the most visually prominent object in videos. Early VSOD studies [27] [28] utilized manually crafted features to detect these salient objects. In more recent research [9] [8] [9] [14] [23], the focus has shifted to deep learning models, which have demonstrated promising results in VSOD. These techniques however are not optimized for the video mirror detection task. This is because mirrored regions may not always represent the most visually significant objects within a scene. Our technique aims to identify mirrors in real-life scenarios." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.886, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Optical Flow Estimation. Optical flow estimation fo" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "17245" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.088, + 0.089, + 0.885, + 0.276 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.145, + 0.278, + 0.823, + 0.293 + ], + "angle": 0, + "content": "Figure 2. Snapshots of our proposed Motion Mirror Dataset (MMD), with pixel-level mirror and edge annotations." + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.306, + 0.466, + 0.476 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.479, + 0.46, + 0.493 + ], + "angle": 0, + "content": "Figure 3. Cumulative Distributions. Figure 5. Color Contrast." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.511, + 0.47, + 0.661 + ], + "angle": 0, + "content": "cuses on extracting per-pixel motion from video sequences. Early efforts [5] used hand-crafted features of image brightness between consecutive frames. In recent studies, convolutional deep learning and transformer-based approaches have made significant improvements [24] [7]. In our context, optical flow estimation is not inherently applicable to video mirror detection and cannot serve as a direct detection method. This study aims to leverage deep-learning optical flow as a way to model inconsistent motion cues, which enables a robust end-to-end video mirror detection framework." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.679, + 0.248, + 0.696 + ], + "angle": 0, + "content": "3. Proposed Dataset" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.705, + 0.471, + 0.903 + ], + "angle": 0, + "content": "To create our Motion Mirror Dataset (MMD), we first recruited six volunteering participants (final year university students) to capture 37 nine-second smartphone videos from their surroundings (e.g., living rooms, bathrooms). Participants were instructed to maintain consistent motion around mirror regions while recording. We recorded videos under various lighting conditions, including daily light and low-light settings. The dataset was subsequently divided into non-overlapping training (4,653 samples) and testing (5,074 samples) sets. To accomplish this, we randomly split the 37 videos into the two sets (training and testing). Specifically 18 videos (4,653 samples) are put into the training set and 19 videos (5,074 samples) are put into the testing" + }, + { + "type": "table", + "bbox": [ + 0.501, + 0.302, + 0.895, + 0.35 + ], + "angle": 0, + "content": "
DatasetSSIM ↑
VMD-D0.7950544833184201
Ours (MotionMirrorDataset)0.9033545507233643
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.357, + 0.892, + 0.385 + ], + "angle": 0, + "content": "Table 1. SSIM of MMD and VMD-D. A higher SSIM indicates a more reasonable split of dataset." + }, + { + "type": "table", + "bbox": [ + 0.513, + 0.399, + 0.872, + 0.475 + ], + "angle": 0, + "content": "
DatasetAreas not covered by mirrors ↓
(Ours) MMD Training0.000061
(Ours) MMD Testing0.000015
VMD-D Training0.00038147
VMD-D Testing0.00024414
" + }, + { + "type": "table_caption", + "bbox": [ + 0.5, + 0.478, + 0.891, + 0.506 + ], + "angle": 0, + "content": "Table 2. Mirror Coverage. A smaller non-coverage rate indicates a more even spatial distribution of mirrors." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.523, + 0.892, + 0.599 + ], + "angle": 0, + "content": "set. Each frame was meticulously annotated with pixel-level mirror masks. We applied a depth-first search algorithm to the pixel-level mirror masks to extract edge masks, which we then manually verified for ground truth accuracy. All videos in the dataset have a frame rate of 30 fps." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.599, + 0.892, + 0.659 + ], + "angle": 0, + "content": "Statistic Analysis. We compare our proposed MMD with VMD-D [12] in terms of Cumulative Distribution and Structural Similarity between training and testing sets, Mirror Size Distribution, and Area Coverage." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.659, + 0.893, + 0.87 + ], + "angle": 0, + "content": "Figure 3 shows the cumulative distribution of all mirrors in both training and testing sets. Compared to VMD-D, our MMD dataset contains more intricate scenarios. VMD-D [12] however is highly biased with majority of mirror regions fall in the center in the train/test sets whilst there are nearly none along the edge (darker color on left/right). The mirror regions of our proposed dataset (MMD) are spread out across the entire heatmap where mirrors are also found at the image edges. We further compute the Structural Similarity Index (SSIM) on the cumulative area distributions (Figure 3) directly to compare the structure between the whole training and testing sets. The higher SSIM of our MMD in Table 1 shows that the training and testing sets are similar in structure and fairer for evaluation purpose." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.894, + 0.9 + ], + "angle": 0, + "content": "Figure 4 shows the size distribution of mirrors in both MMD and VMD-D datasets. VMD-D contains a lot of" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "17246" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.099, + 0.102, + 0.885, + 0.389 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.398, + 0.896, + 0.511 + ], + "angle": 0, + "content": "Figure 6. Method overview. Our MG-VMD network takes three adjacent frames \\((\\mathrm{I}_{\\mathrm{N - 2}},\\mathrm{I}_{\\mathrm{N - 1}},\\mathrm{I}_{\\mathrm{N}})\\) as inputs and extracts their multi-scale features via a shared ResNext-101 [29] backbone and DeepLabV3[2]. We apply the optical flow coherence block to process \\(\\mathrm{I}_{\\mathrm{N - 2}}\\) and \\(\\mathrm{I}_{\\mathrm{N - 1}}\\) to compute an optical flow map for \\(\\mathrm{I}_{\\mathrm{N - 1}}\\), followed by an optical flow vector field extrapolation to estimate the optical flow map for \\(\\mathrm{I}_{\\mathrm{N}}\\). These two optical flow maps are fed into our proposed MAM and MEDM modules to guide the mirror localization and mirror-edge detection processes, respectively. The proposed Motion Attention Module (MAM) learns to locate mirrors from optical flow fields, exploiting motion cues, attentively modeling motion inconsistency. The proposed Motion-guided Edge Detection Module (MEDM) learns to extract and refine mirror boundary edge features. Finally, a fusion refinement module is used to predict the mirror maps by fusing the predicted mirror boundary maps of MEDM, the mirror localization maps of MAM, and the extracted multi-scale image backbone features." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.525, + 0.471, + 0.601 + ], + "angle": 0, + "content": "small mirrors with few/no large mirrors, which does not represent real-world scenarios. In contrast, our MMD dataset has a more reasonable distribution, containing a variety of small, medium and large sized mirrors, and therefore is more challenging than VMD-D." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.602, + 0.47, + 0.692 + ], + "angle": 0, + "content": "Table 2 shows the areas (ratio to the whole image) where there is no mirror coverage within the two datasets. The smaller ratios show that MMD contains mirrors that are more evenly spatially distributed across the whole image. MMD is more challenging and will introduce less spatial bias in the model learning." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.693, + 0.471, + 0.83 + ], + "angle": 0, + "content": "We further compare the color contrast between mirror and non-mirror regions in all images between MMD and VMD-D. The histogram in Figure 5 shows that MMD contains more images with lower color contrast between mirror and non-mirror regions. Lower color contrast implies a decrease in the amount of contextual information available for a model to leverage, making it more challenging to detect mirrors. The \\(4^{\\text{th}}\\) column of Figure 2 shows one example. These show that MMD is more challenging than VMD-D." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.845, + 0.251, + 0.862 + ], + "angle": 0, + "content": "4. Proposed Method" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.871, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Our proposed method, Motion-guided Video Mirror Detection (MG-VMD), aims to enhance mirror detection by" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.525, + 0.892, + 0.6 + ], + "angle": 0, + "content": "modelling motion inconsistency cues. Drawing inspiration from our own observations and neuroscience studies, we hypothesise that modeling motion inconsistency cues can improve the accuracy of mirror detection, especially in dynamic video scenarios." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.605, + 0.892, + 0.711 + ], + "angle": 0, + "content": "The core idea behind MG-VMD is to utilize optical flow to model motion inconsistency and use it to guide the extraction of multi-scale image and edge features, allowing for the temporal detection of mirrored regions in videos. To achieve this, we introduce two key modules: the Motion Attention Module (MAM) and the Motion-guided Edge Detection Module (MEDM)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.715, + 0.892, + 0.837 + ], + "angle": 0, + "content": "Specifically, the Motion Attention Module (MAM) is designed to explicitly predict mirror locations by modelling motion inconsistency cues from the underlying optical flow fields. This module enhances video mirror prediction by providing spatial contextual information through the multiscale image features and motion cues. It informs the subsequent stages of the model about the location of mirror regions in the video." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.895, + 0.903 + ], + "angle": 0, + "content": "The Motion-guided Edge Detection Module (MEDM) is dedicated to predicting mirror boundaries, contributing to the enhancement of overall mirror region predictions. This module takes advantage of image features and employs mo" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "17247" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.086, + 0.092, + 0.468, + 0.183 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.197, + 0.47, + 0.224 + ], + "angle": 0, + "content": "Figure 7. The proposed MAM (Motion Attention Module) aims to model the inconsistent motion cues to localize the mirrors." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.234, + 0.47, + 0.263 + ], + "angle": 0, + "content": "tion inconsistency cues to guide the detection of features at the edges, specifically tailored for mirror detection." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.264, + 0.47, + 0.399 + ], + "angle": 0, + "content": "The Fusion Refinement Module, inspired from PMD-Net [11], further takes the MAM predicted mirror locations and the MEDM refined mirror boundaries to predict mirrors. The integration plays a pivotal role in improving mirror prediction by alleviating errors arising from the limited capability of edge detection or inconsistent motion cues alone. It overcomes the limitations of individual modules and provides spatial and temporal context for accurate mirror prediction." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.4, + 0.47, + 0.596 + ], + "angle": 0, + "content": "In Figure 6 (Architecture Overview), the MG-VMD processes adjacent images \\((\\mathrm{I}_{\\mathrm{N - 2}},\\mathrm{I}_{\\mathrm{N - 1}},\\mathrm{I}_{\\mathrm{N}})\\) as inputs. Optical flow field extrapolation is employed to enhance coherency and address temporal errors between adjacent branches (N-1 and N). The model utilizes a shared ResNext-101 [29] backbone and DeepLabV3 [2] to extract multi-scale image features. The outputs include predicted mirror binary maps \\((\\mathbf{M}_{\\mathrm{N - 1}}\\) and \\(\\mathbf{M}_{\\mathrm{N}}\\) ), predicted mirror-boundary edge maps \\((\\mathbf{E}_{\\mathrm{N - 1}}\\) and \\(\\mathbf{E}_{\\mathrm{N}}\\) ), and mirror motion predictions \\((\\mathbf{M}\\mathbf{p}_{\\mathrm{N - 1}}\\) and \\(\\mathbf{M}\\mathbf{p}_{\\mathrm{N}}\\) ). The collaboration of MAM, MEDM, and the Fusion Refinement Module ensures accurate and robust video mirror detection by considering both spatial and temporal aspects of mirror cues." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.606, + 0.347, + 0.621 + ], + "angle": 0, + "content": "4.1. Optical Flow Coherence Block" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.629, + 0.468, + 0.84 + ], + "angle": 0, + "content": "Our architecture leverages a pre-trained optical flow model. Similar to previous work [10], our optical flow block utilizes frozen weights. However, optical flow may become unreliable in scenarios with low contrast in image details or when images become saturated. To achieve a simpler design while obtaining more stable optical flow, we introduce our Optical Flow Coherence Block, inspired by the work in [24] for motion estimation. This block extracts motion and extends it using linear extrapolation. The extrapolated optical flow motion, denoted as \\( flo \\), is then utilized in our subsequent motion-guided modules, MAM and MEDM. We empirically find that this simple design outperforms alternatives (refer to Section 6 of the Supplemental) whilst enhancing coherency." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.468, + 0.902 + ], + "angle": 0, + "content": "Specifically, the block calculates motion vectors \\( f(N - 2,N - 1) \\) between frames \\( \\mathrm{I}_{\\mathrm{N - 2}} \\) and \\( \\mathrm{I}_{\\mathrm{N - 1}} \\). These vectors are then extrapolated by one frame to estimate motion \\( f^{\\prime}(N - 1,N) \\) between \\( \\mathrm{I}_{\\mathrm{N - 1}} \\) and \\( \\mathrm{I}_{\\mathrm{N}} \\). By using extrapolation, we can calculate the motion vectors of the frame \\( \\mathrm{I}_{\\mathrm{N - 1}} \\) and \\( \\mathrm{I}_{\\mathrm{N}} \\)." + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.09, + 0.891, + 0.192 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.204, + 0.892, + 0.245 + ], + "angle": 0, + "content": "Figure 8. Our proposed MEDM (Motion-Guided Edge Detection Module) aims to detect the mirror boundaries based on the captured motion cues in the optical flow vector fields." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.253, + 0.892, + 0.328 + ], + "angle": 0, + "content": "tion, we connect the motion information from the frame at N-2 to N-1, and then to N. This connection helps maintain a smooth and consistent flow of motion information, reducing errors from neighboring frames and sudden changes in motion, especially around mirror regions." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.34, + 0.733, + 0.355 + ], + "angle": 0, + "content": "4.2. Motion Attention Module" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.364, + 0.892, + 0.469 + ], + "angle": 0, + "content": "Figure 7 shows the architecture of the Motion Attention Module (MAM). This module focuses on direct mirror prediction from the optical flow fields. It leverages self-attention to capture and refine better features, to model motion inconsistency cue and to inform mirror locations. The output is then fed to the Fusion Refinement Module for the final mirror refinement." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.471, + 0.892, + 0.591 + ], + "angle": 0, + "content": "Specifically, we initiate the MAM by embedding patches with positional information to create \\( flo_{\\text{embed}} \\). Subsequently, we guide \\( flo_{\\text{embed}} \\) through six transformer blocks. Each block incorporates self-attention, treating \\( flo_{\\text{embed}} \\) as Q, K, and V in the self-attention mechanism, with output denoted as \\( flo_{\\text{attend}} \\). To enhance its representation, \\( flo_{\\text{attend}} \\) undergoes batch normalization, followed by processing through a feed-forward design." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.593, + 0.892, + 0.759 + ], + "angle": 0, + "content": "The output of the self-attention mechanism within one Transformer BlockLayer N is then passed to the subsequent Transformer BlockLayer N+1. Following this sequence, the output from Transformer BlockLayer 6 is channeled through a convolutional layer. The intentional use of six sequential transformer blocks serves a dual purpose: it equips the model to comprehend long-range dependencies, crucial for temporal relationships, and enables the capture of intricate spatial dependencies. This enhanced capability is particularly valuable for predicting spatial locations of the attended mirrors, denoted as \\(\\mathrm{Mp}_{\\mathrm{N - 1}}\\) and \\(\\mathrm{Mp_N}\\)" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.77, + 0.833, + 0.787 + ], + "angle": 0, + "content": "4.3. Motion-guided Edge Detection Module" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.794, + 0.891, + 0.87 + ], + "angle": 0, + "content": "Figure 8 shows the architecture of the Motion-guided Edge Detection Module (MEDM). Unlike previous mirror edge detection in [11, 22] that only exploits image features, we also leverage the inconsistent motion cues captured in the optical flow field to guide the mirror edge detection process." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.891, + 0.901 + ], + "angle": 0, + "content": "MEDM processes three inputs: \\( \\mathrm{I}_{\\mathrm{low}} \\) (a 2nd-level multiscale image feature), \\( \\mathrm{I}_{\\mathrm{high}} \\) (a 5th-layer multi-scale image" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.956 + ], + "angle": 0, + "content": "17248" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.152 + ], + "angle": 0, + "content": "feature), and \\(flo\\) (the optical flow field for motion guidance). The process begins with convolutional layers that extract and predict edge features—referred to as \\(E_N\\) or \\(E_{N-1}\\)—from \\(\\mathrm{I}_{\\mathrm{low}}\\), \\(\\mathrm{I}_{\\mathrm{high}}\\), and \\(flo\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.153, + 0.473, + 0.334 + ], + "angle": 0, + "content": "The cross-attention mechanism involves matrix multiplication between the transformed \\(flo\\) and \\(E_N\\) or \\(E_{N - 1}\\). A Softmax activation determines the cross-attended beta value between \\(flo\\) and \\(E_N\\) or \\(E_{N - 1}\\). This beta value, representing the attention weight, is then used to focus on \\(flo\\), determining the contribution of each element in the optical flow field to the final cross-attended motion-guided edge feature. We specifically employ a cross-attention mechanism in this step to guide the extracted edge feature with the optical flow field. The motion-guided edge feature undergoes a convolutional pass, producing the output: the motion-guided edge feature map, \\(E_N\\) or \\(E_{N - 1}\\), respectively." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.342, + 0.322, + 0.356 + ], + "angle": 0, + "content": "4.4. Fusion Refinement Module" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.364, + 0.47, + 0.44 + ], + "angle": 0, + "content": "We refine the motion-guided edge detection features from MEDM and the motion-based mirror predictions from MAM. This refining step is crucial to address limitations in each module (MEDM, MAM) and ensure accurate mirror prediction with both spatial and temporal context." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.44, + 0.47, + 0.561 + ], + "angle": 0, + "content": "The process involves fusion-refinement, where model information is combined, refined, and fed into a convolutional binary classification head. This produces per-pixel binary classification maps, named \\( \\mathbf{M}_{\\mathrm{N - 1}} \\) and \\( \\mathbf{M}_{\\mathrm{N}} \\), using two instances of the Fusion Refinement Module. Our approach is inspired by PMD-Net's [11] refinement strategy, adapted here to improve multiple modalities and guide mirror prediction for \\( \\mathbf{M}_{\\mathrm{N - 1}} \\) and \\( \\mathbf{M}_{\\mathrm{N}} \\)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.57, + 0.223, + 0.584 + ], + "angle": 0, + "content": "4.5. Loss Function" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.593, + 0.422, + 0.608 + ], + "angle": 0, + "content": "To train the model, we use the loss function \\((\\mathcal{L}_{\\mathrm{Final}})\\):" + }, + { + "type": "equation", + "bbox": [ + 0.127, + 0.618, + 0.46, + 0.634 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {A} = \\left(\\mathcal {L} _ {\\text {M i r r o r A}} \\cdot \\alpha\\right) + \\left(\\mathcal {L} _ {\\text {M o t i o n A}} \\cdot \\zeta\\right) + \\left(\\mathcal {L} _ {\\text {E d g e A}} \\cdot \\gamma\\right)\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.127, + 0.637, + 0.459, + 0.652 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {B} = \\left(\\mathcal {L} _ {\\text {M i r r o r B}} \\cdot \\alpha\\right) + \\left(\\mathcal {L} _ {\\text {M o t i o n B}} \\cdot \\zeta\\right) + \\left(\\mathcal {L} _ {\\text {E d g e B}} \\cdot \\gamma\\right)\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.088, + 0.657, + 0.412, + 0.691 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} _ {\\text {C o h e r e n c e}} = (\\text {M i r r o r} _ {\\text {C o h e r e n c e}} \\\\ + \\operatorname {E d g e} _ {\\text {C o h e r e n c e}} + \\operatorname {M o t i o n} _ {\\text {C o h e r e n c e}}) \\cdot \\delta \\\\ \\end{array}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.113, + 0.696, + 0.316, + 0.71 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {F i n a l}} = \\mathcal {L} _ {\\text {C o h e r e n c e}} + \\mathcal {L} _ {A} + \\mathcal {L} _ {B}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.72, + 0.352, + 0.735 + ], + "angle": 0, + "content": "We discuss the loss component below:" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.47, + 0.9 + ], + "angle": 0, + "content": "Temporal Loss \\((\\mathcal{L}_{\\mathrm{Coherence}})\\): To make predictions across consecutive frames more consistent over time, we use BCELossWithLogits (binary cross-entropy) between \\(\\mathbf{M}_{\\mathrm{N - 1}}\\) and \\(\\mathbf{M_N}\\), \\(\\mathbf{M}\\mathfrak{p}_{\\mathrm{N - 1}}\\) and \\(\\mathbf{M}\\mathfrak{p}_{\\mathrm{N}}\\), as well as \\(\\mathbf{E}_{\\mathrm{N - 1}}\\) and \\(\\mathbf{E}_{\\mathrm{N}}\\). We set \\(\\delta\\) to 1, calling this loss \\(\\mathcal{L}_{\\mathrm{Coherence}}\\). This loss is aimed at reducing temporal inaccuracies in edge prediction, motion mirror location, and mirror prediction between frames in a video. It ensures that predictions from our model's combined features remain consistent over time, addressing a current limitation in video mirror detection." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.182 + ], + "angle": 0, + "content": "Mirror Binary Losses \\((\\mathcal{L}_{\\mathrm{MirrorA}}\\) and \\(\\mathcal{L}_{\\mathrm{MirrorB}})\\): We use two BCELossWithLogits loss functions to predict mirrored regions in \\(\\mathrm{M_{N - 1}}\\) \\((\\mathcal{L}_{\\mathrm{MirrorA}})\\) and \\(\\mathrm{M_N}\\) \\((\\mathcal{L}_{\\mathrm{MirrorB}})\\). These losses are then compared with the actual ground truth mirror maps. We set \\(\\alpha\\) to 3 for these losses, emphasizing the importance of accurately detecting mirrors in the model." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.182, + 0.892, + 0.272 + ], + "angle": 0, + "content": "Motion Mirror Losses \\((\\mathcal{L}_{\\mathrm{MotionA}}\\) and \\(\\mathcal{L}_{\\mathrm{MotionB}})\\): Two additional BCELossWithLogits loss functions are used to predict mirror regions from motion-guided image features \\(\\mathrm{Mp_{N - 1}}\\) \\((\\mathcal{L}_{\\mathrm{MotionA}})\\) and \\(\\mathrm{Mp_N}\\) \\((\\mathcal{L}_{\\mathrm{MotionB}})\\). These losses are then compared with the ground-truth mirror maps. We set \\(\\zeta\\) to 1 for both of these losses." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.273, + 0.892, + 0.364 + ], + "angle": 0, + "content": "Mirror Edge Losses \\((\\mathcal{L}_{\\mathrm{EdgeA}}\\) and \\(\\mathcal{L}_{\\mathrm{EdgeB}})\\): Two BCELossWithLogits loss functions are utilized for predicting mirror boundaries in \\(\\mathrm{E_{N - 1}}\\) \\((\\mathcal{L}_{\\mathrm{EdgeA}})\\) and \\(\\mathrm{E_N}\\) \\((\\mathcal{L}_{\\mathrm{EdgeB}})\\). These losses are compared against the ground-truth mirror edge maps. We set \\(\\gamma = 2\\) for each of these losses, emphasizing the precision of mirror edge prediction." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.376, + 0.632, + 0.393 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.401, + 0.685, + 0.417 + ], + "angle": 0, + "content": "5.1. Setups and Metrics" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.424, + 0.892, + 0.544 + ], + "angle": 0, + "content": "We implemented MG-VMD using PyTorch, trained and validated it against comparative models on a NVIDIA RTX 3090 GPU. In the data preprocessing phase, all images, ground-truth mirror maps, and edge maps were resized to a resolution of \\(224 \\times 224\\). Our model's ResNext-101 backbone and DeepLabV3 parameters were initialized using the pretrained ResNext-101 backbone and pretrained DeepLabV3 weights from VMD-Net [12]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.545, + 0.892, + 0.636 + ], + "angle": 0, + "content": "We utilized a Stochastic Gradient Descent Optimizer with a starting learning rate of 9e-3, a momentum of 1.9, and a weight decay of 5e-4. Then, we applied an adaptive learning rate to the optimizer, interpolating the learning rate from 9e-3 to 3e-3 across epochs 1 to 15. The model underwent 15 epochs of training with a batch size of 8." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.636, + 0.892, + 0.697 + ], + "angle": 0, + "content": "For validation, we employed four quantitative metrics: Mean Absolute Error (MAE↓), F-measure \\((F_{\\beta}\\uparrow)\\), Accuracy↑, and Intersection over Union (IoU↑). F-measure \\((F_{\\beta}\\uparrow)\\) is computed as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.578, + 0.707, + 0.892, + 0.739 + ], + "angle": 0, + "content": "\\[\nF _ {\\beta} = \\frac {1 + \\beta^ {2} (p r e c i s i o n * r e c a l l)}{\\beta^ {2} + r e c a l l}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.748, + 0.892, + 0.793 + ], + "angle": 0, + "content": "where \\(\\beta\\) is set to 0.3, as suggested by previous research [11] [1]. The F-measure \\((\\mathrm{F}_{\\beta}\\uparrow)\\) assesses the harmonic mean between precision and recall." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.802, + 0.702, + 0.818 + ], + "angle": 0, + "content": "5.2. Experimental Results" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.825, + 0.892, + 0.901 + ], + "angle": 0, + "content": "In Table 3 and Figure 9, we evaluate our proposed technique against 8 state-of-the-art methods in Video Salient Object Detection, and Image-Based/Video Mirror Detection. Using their respective pre-trained weights, we fine-tune and validate these results on our proposed MMD dataset.." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "17249" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.078, + 0.088, + 0.917, + 0.568 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.571, + 0.895, + 0.613 + ], + "angle": 0, + "content": "Figure 9. Qualitative results table comparing our proposed method against state-of-the-art Video Salient Object Detection, and Image-Based/Video Mirror Detection FS-Net [8], MGA [10], ALGRF [23], F3Net [9], UFO[17], MirrorNet[30], PMD-Net [11], VMD-Net[12] trained and validated on our proposed MMD dataset." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.627, + 0.396, + 0.644 + ], + "angle": 0, + "content": "5.3. Comparison Against State-of-the-art" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.655, + 0.47, + 0.852 + ], + "angle": 0, + "content": "Our MG-VMD achieves better performance in Mean Absolute Error (MAE↓), F-measure (\\(F_{\\beta} \\uparrow\\)), Accuracy↑, and Intersection over Union (IoU↑) when compared to state-of-the-art Video Salient Object Detection, Image-Based Mirror Detection and Video Mirror Detection methods fine-tuned and validated on our proposed Motion Mirror Dataset. We attribute the better performance to the exploiting of motion inconsistency cues across adjacent frames through our novel modules: Motion-Guided Edge Detection Module (MEDM) and Motion Attention Module (MAM). Figure 9 shows that our method is able to detect mirrors well in complex (e.g., low contrast) scenarios, outperforming state-of-the-arts." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.471, + 0.901 + ], + "angle": 0, + "content": "We also evaluate the generalizability of our model. In the Supplemental, we show results comparing MG-VMD and VMD-Net trained on our MMD dataset and tested on" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.628, + 0.892, + 0.673 + ], + "angle": 0, + "content": "VMD-D [12]. MG-VMD outperforms VMD-Net, showing that our model has slightly better/close generalization performances." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.682, + 0.605, + 0.696 + ], + "angle": 0, + "content": "5.4. Ablation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.704, + 0.892, + 0.886 + ], + "angle": 0, + "content": "We conducted an ablation study to assess the effectiveness of the proposed modules in MG-VMD. Initially, we removed both the Motion Attention Module and the Motion-guided Edge Detection Module, resulting in the baseline with only the Refinement module, which incorporates convolutional layers. This configuration is referred to as the \"baseline\" for comparative analysis. The rationale behind evaluating the baseline's performance without the Optical Flow Coherence Block, Motion-Guided Edge Detection Module, and Motion Attention Module was to quantitatively gauge the efficacy of our deep-learning approach without leveraging inconsistent motion cues." + }, + { + "type": "text", + "bbox": [ + 0.52, + 0.886, + 0.891, + 0.901 + ], + "angle": 0, + "content": "Table 4 shows the performance of the ablated models:" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "17250" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.082, + 0.089, + 0.468, + 0.234 + ], + "angle": 0, + "content": "
ModelsAccuracy↑MAE↓Fβ↑IoU↑
MGA0.5660370.4339630.4891760.281285
UFO0.7967970.2032040.7974470.598641
FSNet0.8532870.1467130.8377320.707848
ALGRF0.7237720.2762290.6928290.502814
F3Net0.8388630.1611370.8473210.670618
PMD-Net0.7409090.2590910.8472110.424085
MirrorNet0.8352270.1647720.8389870.666242
VMDNet0.8544150.1455850.8123470.722634
Ours0.8725320.1274680.8694190.725130
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.238, + 0.473, + 0.336 + ], + "angle": 0, + "content": "Table 3. Quantitative results table comparing our proposed method against state-of-the-art Video Salient Object Detection, and Image-Based/Video Mirror Detection FS-Net [8], MGA [10], AL-GRF [23], F3Net [9], UFO [17], PMD-Net [11], MirrorNet[30], VMD-Net[12] methods trained and validated on our proposed MMD dataset. Red, Blue and Green indicate the best, second and third best performances, respectively." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.35, + 0.466, + 0.424 + ], + "angle": 0, + "content": "
ModelsAccuracy↑MAE↓Fβ↑IoU↑
baseline0.8575190.1424810.8375780.700855
b + MEDM0.8581760.1418240.8478530.707350
b + MAM0.8662180.1337820.8416140.713005
Ours0.8725320.1274680.8694190.725130
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.428, + 0.472, + 0.499 + ], + "angle": 0, + "content": "Table 4. Ablation Results: Qualitative results table comparing our proposed method (MG-VMD (Ours)) against ablated versions of our model: b + MAM, b + MEDM, and baseline (b), trained and validated on our proposed MMD dataset. Red, Blue and Green indicate the best, second and third-best performances, respectively." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.51, + 0.47, + 0.587 + ], + "angle": 0, + "content": "\"baseline (b)\", \"b + MEDM\", \"b + MAM\", and \"MG-VMD (Ours)\". Both MAM and MEDM exploit motion inconsistency cues, demonstrating better performance compared to the baseline. The full model combines the complementary outputs from MAM and MEDM, yielding the best results." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.587, + 0.47, + 0.692 + ], + "angle": 0, + "content": "Our baseline outperforms state-of-the-art VMD-Net [12] for two reasons. First, our model is fine-tuned on VMD-Net's [12] ResNext-101 backbone and DeepLabV3 weights (Section 5). Second, VMD-Net [12] assumes dual correspondence, but in real-life datasets, dual correspondence does not always occur. This shows the usefulness of our MMD dataset." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.693, + 0.471, + 0.83 + ], + "angle": 0, + "content": "Table 4 shows that MEDM and MAM each contributes to the model individually. Without MEDM, mirror edges become noisy (e.g., Figure 10 Column 4 Rows 2 and 4). Without MAM, MEDM may mistakenly detect object edges inside mirrors (e.g., cabinet edges in Figure 10 Column 3 Row 2), leading to inaccurate mirror detection. (Best viewed by zooming in.) Both modules use motion features, complementing each other and boost performance (Figure 10 Column 5, and Table 4)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.845, + 0.196, + 0.86 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.871, + 0.471, + 0.901 + ], + "angle": 0, + "content": "This paper introduces a novel deep-learning approach to tackle video mirror detection by leveraging inconsistent" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.089, + 0.895, + 0.304 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.541, + 0.304, + 0.851, + 0.318 + ], + "angle": 0, + "content": "Figure 10. Qualitative results of the ablated models." + }, + { + "type": "image", + "bbox": [ + 0.502, + 0.339, + 0.891, + 0.454 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.457, + 0.892, + 0.499 + ], + "angle": 0, + "content": "Figure 11. Failure cases. Our Method may not be good in scenarios containing both mirror regions and windows, due to windows irregular motion compared to a scene." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.51, + 0.892, + 0.646 + ], + "angle": 0, + "content": "motion cues within and outside mirrored regions. We also present a benchmark video mirror dataset featuring consistent inter-frame motion across long videos. Our experimental results demonstrate that our proposed MG-VMD deep-learning method outperforms state-of-the-arts in Video Mirror Detection, Image-Based Mirror Detection, and Video Salient Object Detection. Furthermore, we conclude that our proposed Motion Mirror Dataset (MMD) is challenging compared to state-of-the-art Video Mirror Datasets." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.653, + 0.894, + 0.744 + ], + "angle": 0, + "content": "Our method is not without limitation. Figure 11 shows a scenario with both mirror and window in a scene. We found that similar to mirror regions, windows can have different motions from their surroundings. Thus, our method tends to over-predict window regions as mirrors. We plan to improve our design to mitigate this in the future." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Acknowledgements: Alex is supported by a Swansea GTA Research Scholarship. This project is in part supported by a GRF grant from the Research Grants Council of Hong Kong (Ref.: 11211223). We gratefully acknowledge the support of the HEFCW HERC fund (W21/21HE) for the provision of GPU equipment used in this research. For the purpose of Open Access, the author has applied a Creative Commons Attribution (CC BY) license to any Author Accepted Manuscript (AAM) version arising from this submission." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "17251" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.468, + 0.156 + ], + "angle": 0, + "content": "[1] Radhakrishna Achanta, Sheila Hemami, Francisco Estrada, and Sabine Susstrunk. Frequency-tuned salient region detection. In CVPR, 2009. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.159, + 0.468, + 0.2 + ], + "angle": 0, + "content": "[2] Liang-Chieh Chen, George Papandreou, Florian Schroff, and Hartwig Adam. Rethinking atrous convolution for semantic image segmentation, 2017. 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.089, + 0.202, + 0.468, + 0.242 + ], + "angle": 0, + "content": "[3] Zheng Dong, Ke Xu, Ziheng Duan, Hujun Bao, Weiwei Xu, and Rynson W. H. Lau. Geometry-aware two-scale pifu representation for human reconstruction. In NeurIPS, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.245, + 0.468, + 0.3 + ], + "angle": 0, + "content": "[4] Zheng Dong, Ke Xu, Yaoan Gao, Qilin Sun, Hujun Bao, Weiwei Xu, and Rynson W. H. Lau. Sailor: Synergizing radiance and occupancy fields for live human performance capture. ACM Trans. Graph., 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.302, + 0.468, + 0.342 + ], + "angle": 0, + "content": "[5] Gunnar Farnebäck. Two-frame motion estimation based on polynomial expansion. In Proceedings of the 13th Scandinavian Conference on Image Analysis, 2003. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.345, + 0.468, + 0.384 + ], + "angle": 0, + "content": "[6] Yuan-Chen Guo, Di Kang, Linchao Bao, Yu He, and Song-Hai Zhang. Nerfren: Neural radiance fields with reflections. In CVPR, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.387, + 0.468, + 0.44 + ], + "angle": 0, + "content": "[7] Eddy Ig, Nikolaus Mayer, Tonmoy Saikia, Margret Keuper, Alexey Dosovitskiy, and Thomas Brox. Flownet 2.0: Evolution of optical flow estimation with deep networks. In CVPR, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.444, + 0.468, + 0.485 + ], + "angle": 0, + "content": "[8] Ge-Peng Ji, Keren Fu, Zhe Wu, Deng-Ping Fan, Jianbing Shen, and Ling Shao. Full-duplex strategy for video object segmentation. In ICCV, 2021. 2, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.488, + 0.468, + 0.527 + ], + "angle": 0, + "content": "[9] Qingming Huang Jun Wei, Shuhui Wang. F3net: Fusion, feedback and focus for salient object detection. In AAAI, 2020. 2, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.531, + 0.468, + 0.571 + ], + "angle": 0, + "content": "[10] Haofeng Li, Guanqi Chen, Guanbin Li, and Yizhou Yu. Motion guided attention for video salient object detection. In ICCV, 2019. 5, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.573, + 0.468, + 0.6 + ], + "angle": 0, + "content": "[11] Jiaying Lin, Guodong Wang, and Rynson W.H. Lau. Progressive mirror detection. In CVPR, 2020. 1, 2, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.602, + 0.468, + 0.642 + ], + "angle": 0, + "content": "[12] Jiaying Lin, Xin Tan, and Rynson W.H. Lau. Learning to detect mirrors from videos via dual correspondences. In CVPR, 2023. 1, 2, 3, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.645, + 0.468, + 0.672 + ], + "angle": 0, + "content": "[13] Zhiqiu Lin, Jin Sun, Abe Davis, and Noah Snavely. Visual chirality. In CVPR, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.675, + 0.468, + 0.715 + ], + "angle": 0, + "content": "[14] Jing Liu, Jiaxiang Wang, Weikang Wang, and Yuting Su. Ds-net: Dynamic spatiotemporal network for video salient object detection. arXiv:2012.04886, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.717, + 0.468, + 0.757 + ], + "angle": 0, + "content": "[15] Haiyang Mei, Bo Dong, Wen Dong, Pieter Peers, Xin Yang, Qiang Zhang, and Xiaopeng Wei. Depth-aware mirror segmentation. In CVPR, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.76, + 0.468, + 0.801 + ], + "angle": 0, + "content": "[16] Anwesan Pal, Sayan Mondal, and Henrik I Christensen. \"looking at the right stuff\"-guided semantic-gaze for autonomous driving. In CVPR, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.804, + 0.468, + 0.87 + ], + "angle": 0, + "content": "[17] Yukun Su, Jingliang Deng, Ruizhou Sun, Guosheng Lin, and Qingyao Wu. A unified transformer framework for group-based segmentation: Co-segmentation, co-saliency detection and video salient object detection. IEEE Trans. Multimedia, 2023. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.468, + 0.901 + ], + "angle": 0, + "content": "[18] Hideki Tamura, Maki Tsukuda, Hiroshi Higashi, and Shigeki Nakauchi. Perceptual segregation between mirror and glass" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.468, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.094, + 0.892, + 0.119 + ], + "angle": 0, + "content": "material under natural and unnatural illumination. Journal of Vision, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.122, + 0.892, + 0.163 + ], + "angle": 0, + "content": "[19] Hideki Tamura, Hiroshi Higashi, and Shigeki Nakauchi. Multiple cues for visual perception of mirror and glass materials. Journal of Vision, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.165, + 0.892, + 0.206 + ], + "angle": 0, + "content": "[20] Hideki Tamura, Hiroshi Higashi, and Shigeki Nakauchi. Dynamic visual cues for differentiating mirror and glass. Scientific Reports, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.208, + 0.892, + 0.261 + ], + "angle": 0, + "content": "[21] Xin Tan, Ke Xu, Ying Cao, Yiheng Zhang, Lizhuang Ma, and Rynson W. H. Lau. Night-time scene parsing with a large real dataset. IEEE Transactions on Image Processing, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.264, + 0.892, + 0.305 + ], + "angle": 0, + "content": "[22] Xin Tan, Jiaying Lin, Ke Xu, Pan Chen, Lizhuang Ma, and Rynson W.H. Lau. Mirror detection with the visual chirality cue. IEEE Trans. Pattern Anal. Mach. Intell., 2022. 1, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.307, + 0.892, + 0.347 + ], + "angle": 0, + "content": "[23] Yi Tang, Yuanman Li, and Guoliang Xing. Video salient object detection via adaptive local-global refinement. arXiv:2104.14360, 2021. 2, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.35, + 0.892, + 0.377 + ], + "angle": 0, + "content": "[24] Zachary Teed and Jia Deng. Raft: Recurrent all-pairs field transforms for optical flow. In ECCV, 2020. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.379, + 0.892, + 0.42 + ], + "angle": 0, + "content": "[25] Xin Tian, Ke Xu, Xin Yang, Lin Du, Baocai Yin, and Rynson WH Lau. Bi-directional object-context prioritization learning for saliency ranking. In CVPR, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.422, + 0.892, + 0.463 + ], + "angle": 0, + "content": "[26] Xin Tian, Ke Xu, Xin Yang, Baocai Yin, and Rynson Lau. Learning to detect instance-level salient objects using complementary image labels. Int. J. Comput. Vis., 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.465, + 0.892, + 0.504 + ], + "angle": 0, + "content": "[27] Pavel Tokmakov, Karteek Alahari, and Cordelia Schmid. Learning video object segmentation with visual memory. arXiv:1704.05737, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.507, + 0.892, + 0.548 + ], + "angle": 0, + "content": "[28] Wenguan Wang, Xiankai Lu, Jianbing Shen, David Crandall, and Ling Shao. Zero-shot video object segmentation via attentive graph neural networks. In ICCV, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.55, + 0.892, + 0.59 + ], + "angle": 0, + "content": "[29] Saining Xie, Ross Girshick, Piotr Dólár, Zhuowen Tu, and Kaiming He. Aggregated residual transformations for deep neural networks. In CVPR, 2017. 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.592, + 0.892, + 0.632 + ], + "angle": 0, + "content": "[30] Xin Yang, Haiyang Mei, Ke Xu, Xiaopeng Wei, Baocai Yin, and Rynson W. H. Lau. Where is my mirror? In ICCV, 2019. 1, 2, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.635, + 0.892, + 0.69 + ], + "angle": 0, + "content": "[31] Junyi Zeng, Chong Bao, Rui Chen, Zilong Dong, Guofeng Zhang, Hujun Bao, and Zhaopeng Cui. Mirror-nerf: Learning neural radiance fields for mirrors with whitted-style ray tracing. In ACM MM, 2023. 1" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.094, + 0.892, + 0.69 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "17252" + } + ] +] \ No newline at end of file diff --git a/2024/Effective Video Mirror Detection with Inconsistent Motion Cues/423d0156-aefb-4c1d-9839-9825d419ff81_origin.pdf b/2024/Effective Video Mirror Detection with Inconsistent Motion Cues/423d0156-aefb-4c1d-9839-9825d419ff81_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..4676cdb80ee203e972d0c6045c19d91dc1d5faa6 --- /dev/null +++ b/2024/Effective Video Mirror Detection with Inconsistent Motion Cues/423d0156-aefb-4c1d-9839-9825d419ff81_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:87c12e8025e994ecf34da857d1378fb051ed2cf919216ea4e774a52eca53510e +size 4787302 diff --git a/2024/Effective Video Mirror Detection with Inconsistent Motion Cues/full.md b/2024/Effective Video Mirror Detection with Inconsistent Motion Cues/full.md new file mode 100644 index 0000000000000000000000000000000000000000..61ebf39789c440b261d37ad923afc905400465c6 --- /dev/null +++ b/2024/Effective Video Mirror Detection with Inconsistent Motion Cues/full.md @@ -0,0 +1,280 @@ +# Effective Video Mirror Detection with Inconsistent Motion Cues + +Alex Warren1, Ke Xu2, Jiaying Lin2, Gary K.L. Tam1, Rynson W.H. Lau1,2 + +Department of Computer Science, Swansea University1 and City University of Hong Kong2 + +alex.warren@swansea.ac.uk, kkangwing@gmail.com, jiayinlin5-c@my.cityu.edu.hk + +k.l.tam@swansea.ac.uk, Rynson.Lau@cityu.edu.hk + +# Abstract + +Image-based mirror detection has recently undergone rapid research due to its significance in applications such as robotic navigation, semantic segmentation and scene reconstruction. Recently, VMD-Net was proposed as the first video mirror detection technique, by modeling dual correspondences between the inside and outside of the mirror both spatially and temporally. However, this approach is not reliable, as correspondences can occur completely inside or outside of the mirrors. In addition, the proposed dataset VMD-D contains many small mirrors, limiting its applicability to real-world scenarios. To address these problems, we developed a more challenging dataset that includes mirrors of various shapes and sizes at different locations of the frames, providing a better reflection of real-world scenarios. Next, we observed that the motions between the inside and outside of the mirror are often inconsistent. For instance, when moving in front of a mirror, the motion inside the mirror is often much smaller than the motion outside due to increased depth perception. With these observations, we propose modeling inconsistent motion cues to detect mirrors, and a new network with two novel modules. The Motion Attention Module (MAM) explicitly models inconsistent motions around mirrors via optical flow, and the Motion-Guided Edge Detection Module (MEDM) uses motions to guide mirror edge feature learning. Experimental results on our proposed dataset show that our method outperforms state-of-the-arts. The code and dataset are available at https://github.com/ AlexAnthonyWarren/MG-VMD. + +# 1. Introduction + +Mirrors and reflective surfaces are abundant in our surroundings. Successful detection of mirrors underpins many vision applications such as autonomous drone navigation and vehicle vision systems [16]. When a scene contains mirrored regions, mirror identification is critical to reducing errors in applications such as salient object de + +![](images/0dfd4ee24860de30e9d715bf6232ccd9c165459ff99a7c121c670413d046e0af.jpg) +Figure 1. We propose to model motion inconsistency as a cue for mirror detection. The $1^{st}$ column shows the inconsistent motions (depicted with arrows). The locations of these cues align well with the mirror regions. The $2^{nd}$ column shows the optical flow computed between frame N ( $3^{rd}$ column), and frame N-1 from respective videos in our new dataset. Our method ( $5^{th}$ column) predicts more reliable and consistent mirror regions, outperforming VMD-Net [12] ( $4^{th}$ column). The $6^{th}$ column is the ground truth. + +tection [25, 26], semantic segmentation [16, 21], scene reconstruction[3, 4] and NeRF modeling [6, 31]. + +Mirror detection involves binary classification of individual pixels in RGB images to determine if they are inside or outside of the mirror. It is a challenging research as mirrors typically do not have their own visual appearance. They instead reflect objects within their surroundings. This visual attribute makes mirrors difficult to detect. + +There are a few image-based and one video-based mirror detection methods in the literature. Yang et al. [30] first introduced a mirror detection method using contrast and semantics. Lin et al. [11] introduced a PMD-Net that utilizes edge detection and leverages contextual contrasting objects inside and outside of mirrored regions to detect mirrored regions. Tan et al. [22] detect mirrors using Visual Chirality Cue [13] that encodes mirror-like symmetrical properties at the pixel level. Mei et al. [15] factor depth estimation into their mirror detection process and show good results + +in RGB and RGB-D mirror datasets. Recently, Lin et al. [12] introduced the first video mirror detection method using inter-frame and intra-frame dual correspondences. + +Despite these exciting results, we observe research gaps and limitations in these works, especially in video mirror detection. The state-of-the-art VMD-Net [12] models dual correspondences inside and outside of mirrors to aid in the detection of mirror regions in videos. However, the approach is not reliable as correspondences can appear completely inside or outside of mirrors as discussed in [22]. In addition, the proposed VMD-D dataset contains primarily small mirrors, with small inter-frame motion, mirror coverage and content variation. This limits its applicability to mirror detection in real-world scenarios. + +To address these issues, we first developed a new, more challenging dataset, Mirror Motion Dataset (MMD). The dataset contains 37 ( $\approx$ 9-second) videos with substantial motion, each accompanied by manual mirror annotations. To address the small mirror issue in VMD-D, we capture a high variance of mirror sizes and shapes at different spatial locations, generalizing to real-world scenarios. We further include challenging examples, including low-contrast environments, multi-/large/full-mirror scenes and under different lighting conditions. Next, we observe that the motions between the inside and outside of mirrors are often inconsistent. For instance, when moving in front of a mirror, the motion inside is often smaller than outside due to the increased depth perception. The inconsistent motion cues (depicted by arrows in the $1^{st}$ column in Figure 1) also apply in challenging scenarios like low-contrast and multi-mirror scenes. These observations motivate us to ask the question: Can we apply the motion inconsistency cue to improve mirror detection in challenging complex video scenes? + +Neuroscience studies suggest that humans often rely on dynamic perceptual cues to identify mirrored regions in daily life. Tamura et al. [18] [19] show that humans are capable of distinguishing mirrors in dynamic scenes involving rotational movement. [20] highlights that rotational, parallax, forward and backward motions are exploitable motion inconsistencies for mirror detection. To our knowledge, such motion inconsistency cues have not been utilized before for the video mirror detection task. + +To this end, we propose a novel network to model motion inconsistency for mirror detection, with two novel modules. First, the Motion Attention Module (MAM) predicts mirror regions by exploiting inconsistent motion cues obtained from optical flow fields. Second, the Motion-guided Edge Detection Module (MEDM) further uses motion to guide mirror edge detection. Experimental results show that the proposed network outperforms relevant state-of-the-art methods. To summarise, our main contributions include: + +- We propose a novel deep-learning video mirror detection method based on the motion inconsistency cue. + +- Our network includes two novel modules. Motion Attention Module (MAM) attentively predicts mirror regions using motion inconsistency in the optical flow fields. Motion-guided Edge Detection Module (MEDM) uses the motions to guide mirror edge feature extraction. +- We further present a new video dataset (MMD). Compared to VMD-D, our dataset contains mirrors of various shapes and size at different locations of the images. It also covers various scenes, e.g., low-light scenes, and multi-large/full mirror scenes. It is a more challenging dataset, more suitable for real-life scenarios. +- Extensive evaluations show that the proposed model outperforms existing methods on video mirror detection. + +# 2. Related Works + +Video Mirror Detection. VMD-Net [12] introduced the first video mirror detection method. It employs dual correspondences across both inter-frame and intra-frame to predict mirrored regions. They further presented a video mirror dataset for this purpose. However, despite its demonstrated success, we observe two limitations. First, the idea of dual correspondences is not reliable as correspondences can occur completely inside or outside of the mirrors [22]. Second, their dataset contains an excess of small mirrors, which does not generalize well to real-world scenarios. + +This paper aims to facilitate video mirror detection in the wild from two perspectives. First, we propose to model the inconsistent motion cues to learn more robust mirror representations. Second, we introduce a new benchmark dataset comprising videos with more complex scenarios. + +Image-Based Mirror Detection. [30] pioneered mirror detection by utilizing contextual contrasting features and introducing a mirror detection dataset. [15] refined detection using RGB-D data and presented the first RGB-D mirror dataset. [11] introduced a method leveraging contextual contrasts to relate mirror interior and exterior contents, along with an expanded image-based mirror dataset. [22] incorporated chirality cues for mirror detection. Although these work achieved good results, they mainly focus on single-image mirror detection and lack the capacity to identify mirrors within temporal videos. + +Video Salient Object Detection. Video Salient Object Detection (VSOD) seeks to identify the most visually prominent object in videos. Early VSOD studies [27] [28] utilized manually crafted features to detect these salient objects. In more recent research [9] [8] [9] [14] [23], the focus has shifted to deep learning models, which have demonstrated promising results in VSOD. These techniques however are not optimized for the video mirror detection task. This is because mirrored regions may not always represent the most visually significant objects within a scene. Our technique aims to identify mirrors in real-life scenarios. + +Optical Flow Estimation. Optical flow estimation fo + +![](images/f336c6d5b1be67764b040e70639b9b997823bba7bad736a5844ac93c660c1564.jpg) +Figure 2. Snapshots of our proposed Motion Mirror Dataset (MMD), with pixel-level mirror and edge annotations. + +![](images/e2803215de5f929742c0998604c203196ac6331bdcb616cfd48c907e9dd1a766.jpg) +Figure 3. Cumulative Distributions. Figure 5. Color Contrast. + +cuses on extracting per-pixel motion from video sequences. Early efforts [5] used hand-crafted features of image brightness between consecutive frames. In recent studies, convolutional deep learning and transformer-based approaches have made significant improvements [24] [7]. In our context, optical flow estimation is not inherently applicable to video mirror detection and cannot serve as a direct detection method. This study aims to leverage deep-learning optical flow as a way to model inconsistent motion cues, which enables a robust end-to-end video mirror detection framework. + +# 3. Proposed Dataset + +To create our Motion Mirror Dataset (MMD), we first recruited six volunteering participants (final year university students) to capture 37 nine-second smartphone videos from their surroundings (e.g., living rooms, bathrooms). Participants were instructed to maintain consistent motion around mirror regions while recording. We recorded videos under various lighting conditions, including daily light and low-light settings. The dataset was subsequently divided into non-overlapping training (4,653 samples) and testing (5,074 samples) sets. To accomplish this, we randomly split the 37 videos into the two sets (training and testing). Specifically 18 videos (4,653 samples) are put into the training set and 19 videos (5,074 samples) are put into the testing + +
DatasetSSIM ↑
VMD-D0.7950544833184201
Ours (MotionMirrorDataset)0.9033545507233643
+ +Table 1. SSIM of MMD and VMD-D. A higher SSIM indicates a more reasonable split of dataset. + +
DatasetAreas not covered by mirrors ↓
(Ours) MMD Training0.000061
(Ours) MMD Testing0.000015
VMD-D Training0.00038147
VMD-D Testing0.00024414
+ +Table 2. Mirror Coverage. A smaller non-coverage rate indicates a more even spatial distribution of mirrors. + +set. Each frame was meticulously annotated with pixel-level mirror masks. We applied a depth-first search algorithm to the pixel-level mirror masks to extract edge masks, which we then manually verified for ground truth accuracy. All videos in the dataset have a frame rate of 30 fps. + +Statistic Analysis. We compare our proposed MMD with VMD-D [12] in terms of Cumulative Distribution and Structural Similarity between training and testing sets, Mirror Size Distribution, and Area Coverage. + +Figure 3 shows the cumulative distribution of all mirrors in both training and testing sets. Compared to VMD-D, our MMD dataset contains more intricate scenarios. VMD-D [12] however is highly biased with majority of mirror regions fall in the center in the train/test sets whilst there are nearly none along the edge (darker color on left/right). The mirror regions of our proposed dataset (MMD) are spread out across the entire heatmap where mirrors are also found at the image edges. We further compute the Structural Similarity Index (SSIM) on the cumulative area distributions (Figure 3) directly to compare the structure between the whole training and testing sets. The higher SSIM of our MMD in Table 1 shows that the training and testing sets are similar in structure and fairer for evaluation purpose. + +Figure 4 shows the size distribution of mirrors in both MMD and VMD-D datasets. VMD-D contains a lot of + +![](images/228c0a33d6db9084e592569962c0eeff8d0207a5c8ea770bec02354033962aa8.jpg) +Figure 6. Method overview. Our MG-VMD network takes three adjacent frames $(\mathrm{I}_{\mathrm{N - 2}},\mathrm{I}_{\mathrm{N - 1}},\mathrm{I}_{\mathrm{N}})$ as inputs and extracts their multi-scale features via a shared ResNext-101 [29] backbone and DeepLabV3[2]. We apply the optical flow coherence block to process $\mathrm{I}_{\mathrm{N - 2}}$ and $\mathrm{I}_{\mathrm{N - 1}}$ to compute an optical flow map for $\mathrm{I}_{\mathrm{N - 1}}$ , followed by an optical flow vector field extrapolation to estimate the optical flow map for $\mathrm{I}_{\mathrm{N}}$ . These two optical flow maps are fed into our proposed MAM and MEDM modules to guide the mirror localization and mirror-edge detection processes, respectively. The proposed Motion Attention Module (MAM) learns to locate mirrors from optical flow fields, exploiting motion cues, attentively modeling motion inconsistency. The proposed Motion-guided Edge Detection Module (MEDM) learns to extract and refine mirror boundary edge features. Finally, a fusion refinement module is used to predict the mirror maps by fusing the predicted mirror boundary maps of MEDM, the mirror localization maps of MAM, and the extracted multi-scale image backbone features. + +small mirrors with few/no large mirrors, which does not represent real-world scenarios. In contrast, our MMD dataset has a more reasonable distribution, containing a variety of small, medium and large sized mirrors, and therefore is more challenging than VMD-D. + +Table 2 shows the areas (ratio to the whole image) where there is no mirror coverage within the two datasets. The smaller ratios show that MMD contains mirrors that are more evenly spatially distributed across the whole image. MMD is more challenging and will introduce less spatial bias in the model learning. + +We further compare the color contrast between mirror and non-mirror regions in all images between MMD and VMD-D. The histogram in Figure 5 shows that MMD contains more images with lower color contrast between mirror and non-mirror regions. Lower color contrast implies a decrease in the amount of contextual information available for a model to leverage, making it more challenging to detect mirrors. The $4^{\text{th}}$ column of Figure 2 shows one example. These show that MMD is more challenging than VMD-D. + +# 4. Proposed Method + +Our proposed method, Motion-guided Video Mirror Detection (MG-VMD), aims to enhance mirror detection by + +modelling motion inconsistency cues. Drawing inspiration from our own observations and neuroscience studies, we hypothesise that modeling motion inconsistency cues can improve the accuracy of mirror detection, especially in dynamic video scenarios. + +The core idea behind MG-VMD is to utilize optical flow to model motion inconsistency and use it to guide the extraction of multi-scale image and edge features, allowing for the temporal detection of mirrored regions in videos. To achieve this, we introduce two key modules: the Motion Attention Module (MAM) and the Motion-guided Edge Detection Module (MEDM). + +Specifically, the Motion Attention Module (MAM) is designed to explicitly predict mirror locations by modelling motion inconsistency cues from the underlying optical flow fields. This module enhances video mirror prediction by providing spatial contextual information through the multiscale image features and motion cues. It informs the subsequent stages of the model about the location of mirror regions in the video. + +The Motion-guided Edge Detection Module (MEDM) is dedicated to predicting mirror boundaries, contributing to the enhancement of overall mirror region predictions. This module takes advantage of image features and employs mo + +![](images/4214f93536c8abbf728bdc63c5971369f1817a668c77eb59c7ad31826e05faa4.jpg) +Figure 7. The proposed MAM (Motion Attention Module) aims to model the inconsistent motion cues to localize the mirrors. + +tion inconsistency cues to guide the detection of features at the edges, specifically tailored for mirror detection. + +The Fusion Refinement Module, inspired from PMD-Net [11], further takes the MAM predicted mirror locations and the MEDM refined mirror boundaries to predict mirrors. The integration plays a pivotal role in improving mirror prediction by alleviating errors arising from the limited capability of edge detection or inconsistent motion cues alone. It overcomes the limitations of individual modules and provides spatial and temporal context for accurate mirror prediction. + +In Figure 6 (Architecture Overview), the MG-VMD processes adjacent images $(\mathrm{I}_{\mathrm{N - 2}},\mathrm{I}_{\mathrm{N - 1}},\mathrm{I}_{\mathrm{N}})$ as inputs. Optical flow field extrapolation is employed to enhance coherency and address temporal errors between adjacent branches (N-1 and N). The model utilizes a shared ResNext-101 [29] backbone and DeepLabV3 [2] to extract multi-scale image features. The outputs include predicted mirror binary maps $(\mathbf{M}_{\mathrm{N - 1}}$ and $\mathbf{M}_{\mathrm{N}}$ ), predicted mirror-boundary edge maps $(\mathbf{E}_{\mathrm{N - 1}}$ and $\mathbf{E}_{\mathrm{N}}$ ), and mirror motion predictions $(\mathbf{M}\mathbf{p}_{\mathrm{N - 1}}$ and $\mathbf{M}\mathbf{p}_{\mathrm{N}}$ ). The collaboration of MAM, MEDM, and the Fusion Refinement Module ensures accurate and robust video mirror detection by considering both spatial and temporal aspects of mirror cues. + +# 4.1. Optical Flow Coherence Block + +Our architecture leverages a pre-trained optical flow model. Similar to previous work [10], our optical flow block utilizes frozen weights. However, optical flow may become unreliable in scenarios with low contrast in image details or when images become saturated. To achieve a simpler design while obtaining more stable optical flow, we introduce our Optical Flow Coherence Block, inspired by the work in [24] for motion estimation. This block extracts motion and extends it using linear extrapolation. The extrapolated optical flow motion, denoted as $flo$ , is then utilized in our subsequent motion-guided modules, MAM and MEDM. We empirically find that this simple design outperforms alternatives (refer to Section 6 of the Supplemental) whilst enhancing coherency. + +Specifically, the block calculates motion vectors $f(N - 2,N - 1)$ between frames $\mathrm{I}_{\mathrm{N - 2}}$ and $\mathrm{I}_{\mathrm{N - 1}}$ . These vectors are then extrapolated by one frame to estimate motion $f^{\prime}(N - 1,N)$ between $\mathrm{I}_{\mathrm{N - 1}}$ and $\mathrm{I}_{\mathrm{N}}$ . By using extrapolation, we can calculate the motion vectors of the frame $\mathrm{I}_{\mathrm{N - 1}}$ and $\mathrm{I}_{\mathrm{N}}$ . + +![](images/ef55c47ca706413ecfab9687706bb2aaf8664b6bcf40210a82c483e9d4481867.jpg) +Figure 8. Our proposed MEDM (Motion-Guided Edge Detection Module) aims to detect the mirror boundaries based on the captured motion cues in the optical flow vector fields. + +tion, we connect the motion information from the frame at N-2 to N-1, and then to N. This connection helps maintain a smooth and consistent flow of motion information, reducing errors from neighboring frames and sudden changes in motion, especially around mirror regions. + +# 4.2. Motion Attention Module + +Figure 7 shows the architecture of the Motion Attention Module (MAM). This module focuses on direct mirror prediction from the optical flow fields. It leverages self-attention to capture and refine better features, to model motion inconsistency cue and to inform mirror locations. The output is then fed to the Fusion Refinement Module for the final mirror refinement. + +Specifically, we initiate the MAM by embedding patches with positional information to create $flo_{\text{embed}}$ . Subsequently, we guide $flo_{\text{embed}}$ through six transformer blocks. Each block incorporates self-attention, treating $flo_{\text{embed}}$ as Q, K, and V in the self-attention mechanism, with output denoted as $flo_{\text{attend}}$ . To enhance its representation, $flo_{\text{attend}}$ undergoes batch normalization, followed by processing through a feed-forward design. + +The output of the self-attention mechanism within one Transformer BlockLayer N is then passed to the subsequent Transformer BlockLayer N+1. Following this sequence, the output from Transformer BlockLayer 6 is channeled through a convolutional layer. The intentional use of six sequential transformer blocks serves a dual purpose: it equips the model to comprehend long-range dependencies, crucial for temporal relationships, and enables the capture of intricate spatial dependencies. This enhanced capability is particularly valuable for predicting spatial locations of the attended mirrors, denoted as $\mathrm{Mp}_{\mathrm{N - 1}}$ and $\mathrm{Mp_N}$ + +# 4.3. Motion-guided Edge Detection Module + +Figure 8 shows the architecture of the Motion-guided Edge Detection Module (MEDM). Unlike previous mirror edge detection in [11, 22] that only exploits image features, we also leverage the inconsistent motion cues captured in the optical flow field to guide the mirror edge detection process. + +MEDM processes three inputs: $\mathrm{I}_{\mathrm{low}}$ (a 2nd-level multiscale image feature), $\mathrm{I}_{\mathrm{high}}$ (a 5th-layer multi-scale image + +feature), and $flo$ (the optical flow field for motion guidance). The process begins with convolutional layers that extract and predict edge features—referred to as $E_N$ or $E_{N-1}$ —from $\mathrm{I}_{\mathrm{low}}$ , $\mathrm{I}_{\mathrm{high}}$ , and $flo$ . + +The cross-attention mechanism involves matrix multiplication between the transformed $flo$ and $E_N$ or $E_{N - 1}$ . A Softmax activation determines the cross-attended beta value between $flo$ and $E_N$ or $E_{N - 1}$ . This beta value, representing the attention weight, is then used to focus on $flo$ , determining the contribution of each element in the optical flow field to the final cross-attended motion-guided edge feature. We specifically employ a cross-attention mechanism in this step to guide the extracted edge feature with the optical flow field. The motion-guided edge feature undergoes a convolutional pass, producing the output: the motion-guided edge feature map, $E_N$ or $E_{N - 1}$ , respectively. + +# 4.4. Fusion Refinement Module + +We refine the motion-guided edge detection features from MEDM and the motion-based mirror predictions from MAM. This refining step is crucial to address limitations in each module (MEDM, MAM) and ensure accurate mirror prediction with both spatial and temporal context. + +The process involves fusion-refinement, where model information is combined, refined, and fed into a convolutional binary classification head. This produces per-pixel binary classification maps, named $\mathbf{M}_{\mathrm{N - 1}}$ and $\mathbf{M}_{\mathrm{N}}$ , using two instances of the Fusion Refinement Module. Our approach is inspired by PMD-Net's [11] refinement strategy, adapted here to improve multiple modalities and guide mirror prediction for $\mathbf{M}_{\mathrm{N - 1}}$ and $\mathbf{M}_{\mathrm{N}}$ . + +# 4.5. Loss Function + +To train the model, we use the loss function $(\mathcal{L}_{\mathrm{Final}})$ : + +$$ +\mathcal {L} _ {A} = \left(\mathcal {L} _ {\text {M i r r o r A}} \cdot \alpha\right) + \left(\mathcal {L} _ {\text {M o t i o n A}} \cdot \zeta\right) + \left(\mathcal {L} _ {\text {E d g e A}} \cdot \gamma\right) +$$ + +$$ +\mathcal {L} _ {B} = \left(\mathcal {L} _ {\text {M i r r o r B}} \cdot \alpha\right) + \left(\mathcal {L} _ {\text {M o t i o n B}} \cdot \zeta\right) + \left(\mathcal {L} _ {\text {E d g e B}} \cdot \gamma\right) +$$ + +$$ +\begin{array}{l} \mathcal {L} _ {\text {C o h e r e n c e}} = (\text {M i r r o r} _ {\text {C o h e r e n c e}} \\ + \operatorname {E d g e} _ {\text {C o h e r e n c e}} + \operatorname {M o t i o n} _ {\text {C o h e r e n c e}}) \cdot \delta \\ \end{array} +$$ + +$$ +\mathcal {L} _ {\text {F i n a l}} = \mathcal {L} _ {\text {C o h e r e n c e}} + \mathcal {L} _ {A} + \mathcal {L} _ {B} +$$ + +We discuss the loss component below: + +Temporal Loss $(\mathcal{L}_{\mathrm{Coherence}})$ : To make predictions across consecutive frames more consistent over time, we use BCELossWithLogits (binary cross-entropy) between $\mathbf{M}_{\mathrm{N - 1}}$ and $\mathbf{M_N}$ , $\mathbf{M}\mathfrak{p}_{\mathrm{N - 1}}$ and $\mathbf{M}\mathfrak{p}_{\mathrm{N}}$ , as well as $\mathbf{E}_{\mathrm{N - 1}}$ and $\mathbf{E}_{\mathrm{N}}$ . We set $\delta$ to 1, calling this loss $\mathcal{L}_{\mathrm{Coherence}}$ . This loss is aimed at reducing temporal inaccuracies in edge prediction, motion mirror location, and mirror prediction between frames in a video. It ensures that predictions from our model's combined features remain consistent over time, addressing a current limitation in video mirror detection. + +Mirror Binary Losses $(\mathcal{L}_{\mathrm{MirrorA}}$ and $\mathcal{L}_{\mathrm{MirrorB}})$ : We use two BCELossWithLogits loss functions to predict mirrored regions in $\mathrm{M_{N - 1}}$ $(\mathcal{L}_{\mathrm{MirrorA}})$ and $\mathrm{M_N}$ $(\mathcal{L}_{\mathrm{MirrorB}})$ . These losses are then compared with the actual ground truth mirror maps. We set $\alpha$ to 3 for these losses, emphasizing the importance of accurately detecting mirrors in the model. + +Motion Mirror Losses $(\mathcal{L}_{\mathrm{MotionA}}$ and $\mathcal{L}_{\mathrm{MotionB}})$ : Two additional BCELossWithLogits loss functions are used to predict mirror regions from motion-guided image features $\mathrm{Mp_{N - 1}}$ $(\mathcal{L}_{\mathrm{MotionA}})$ and $\mathrm{Mp_N}$ $(\mathcal{L}_{\mathrm{MotionB}})$ . These losses are then compared with the ground-truth mirror maps. We set $\zeta$ to 1 for both of these losses. + +Mirror Edge Losses $(\mathcal{L}_{\mathrm{EdgeA}}$ and $\mathcal{L}_{\mathrm{EdgeB}})$ : Two BCELossWithLogits loss functions are utilized for predicting mirror boundaries in $\mathrm{E_{N - 1}}$ $(\mathcal{L}_{\mathrm{EdgeA}})$ and $\mathrm{E_N}$ $(\mathcal{L}_{\mathrm{EdgeB}})$ . These losses are compared against the ground-truth mirror edge maps. We set $\gamma = 2$ for each of these losses, emphasizing the precision of mirror edge prediction. + +# 5. Experiments + +# 5.1. Setups and Metrics + +We implemented MG-VMD using PyTorch, trained and validated it against comparative models on a NVIDIA RTX 3090 GPU. In the data preprocessing phase, all images, ground-truth mirror maps, and edge maps were resized to a resolution of $224 \times 224$ . Our model's ResNext-101 backbone and DeepLabV3 parameters were initialized using the pretrained ResNext-101 backbone and pretrained DeepLabV3 weights from VMD-Net [12]. + +We utilized a Stochastic Gradient Descent Optimizer with a starting learning rate of 9e-3, a momentum of 1.9, and a weight decay of 5e-4. Then, we applied an adaptive learning rate to the optimizer, interpolating the learning rate from 9e-3 to 3e-3 across epochs 1 to 15. The model underwent 15 epochs of training with a batch size of 8. + +For validation, we employed four quantitative metrics: Mean Absolute Error (MAE↓), F-measure $(F_{\beta}\uparrow)$ , Accuracy↑, and Intersection over Union (IoU↑). F-measure $(F_{\beta}\uparrow)$ is computed as follows: + +$$ +F _ {\beta} = \frac {1 + \beta^ {2} (p r e c i s i o n * r e c a l l)}{\beta^ {2} + r e c a l l}, \tag {1} +$$ + +where $\beta$ is set to 0.3, as suggested by previous research [11] [1]. The F-measure $(\mathrm{F}_{\beta}\uparrow)$ assesses the harmonic mean between precision and recall. + +# 5.2. Experimental Results + +In Table 3 and Figure 9, we evaluate our proposed technique against 8 state-of-the-art methods in Video Salient Object Detection, and Image-Based/Video Mirror Detection. Using their respective pre-trained weights, we fine-tune and validate these results on our proposed MMD dataset.. + +![](images/69c2078ac8b2b3b5ea2c04258db100265172934d33da06687524b04d03907fe2.jpg) +Figure 9. Qualitative results table comparing our proposed method against state-of-the-art Video Salient Object Detection, and Image-Based/Video Mirror Detection FS-Net [8], MGA [10], ALGRF [23], F3Net [9], UFO[17], MirrorNet[30], PMD-Net [11], VMD-Net[12] trained and validated on our proposed MMD dataset. + +# 5.3. Comparison Against State-of-the-art + +Our MG-VMD achieves better performance in Mean Absolute Error (MAE↓), F-measure ( $F_{\beta} \uparrow$ ), Accuracy↑, and Intersection over Union (IoU↑) when compared to state-of-the-art Video Salient Object Detection, Image-Based Mirror Detection and Video Mirror Detection methods fine-tuned and validated on our proposed Motion Mirror Dataset. We attribute the better performance to the exploiting of motion inconsistency cues across adjacent frames through our novel modules: Motion-Guided Edge Detection Module (MEDM) and Motion Attention Module (MAM). Figure 9 shows that our method is able to detect mirrors well in complex (e.g., low contrast) scenarios, outperforming state-of-the-arts. + +We also evaluate the generalizability of our model. In the Supplemental, we show results comparing MG-VMD and VMD-Net trained on our MMD dataset and tested on + +VMD-D [12]. MG-VMD outperforms VMD-Net, showing that our model has slightly better/close generalization performances. + +# 5.4. Ablation + +We conducted an ablation study to assess the effectiveness of the proposed modules in MG-VMD. Initially, we removed both the Motion Attention Module and the Motion-guided Edge Detection Module, resulting in the baseline with only the Refinement module, which incorporates convolutional layers. This configuration is referred to as the "baseline" for comparative analysis. The rationale behind evaluating the baseline's performance without the Optical Flow Coherence Block, Motion-Guided Edge Detection Module, and Motion Attention Module was to quantitatively gauge the efficacy of our deep-learning approach without leveraging inconsistent motion cues. + +Table 4 shows the performance of the ablated models: + +
ModelsAccuracy↑MAE↓Fβ↑IoU↑
MGA0.5660370.4339630.4891760.281285
UFO0.7967970.2032040.7974470.598641
FSNet0.8532870.1467130.8377320.707848
ALGRF0.7237720.2762290.6928290.502814
F3Net0.8388630.1611370.8473210.670618
PMD-Net0.7409090.2590910.8472110.424085
MirrorNet0.8352270.1647720.8389870.666242
VMDNet0.8544150.1455850.8123470.722634
Ours0.8725320.1274680.8694190.725130
+ +Table 3. Quantitative results table comparing our proposed method against state-of-the-art Video Salient Object Detection, and Image-Based/Video Mirror Detection FS-Net [8], MGA [10], AL-GRF [23], F3Net [9], UFO [17], PMD-Net [11], MirrorNet[30], VMD-Net[12] methods trained and validated on our proposed MMD dataset. Red, Blue and Green indicate the best, second and third best performances, respectively. + +
ModelsAccuracy↑MAE↓Fβ↑IoU↑
baseline0.8575190.1424810.8375780.700855
b + MEDM0.8581760.1418240.8478530.707350
b + MAM0.8662180.1337820.8416140.713005
Ours0.8725320.1274680.8694190.725130
+ +Table 4. Ablation Results: Qualitative results table comparing our proposed method (MG-VMD (Ours)) against ablated versions of our model: b + MAM, b + MEDM, and baseline (b), trained and validated on our proposed MMD dataset. Red, Blue and Green indicate the best, second and third-best performances, respectively. + +"baseline (b)", "b + MEDM", "b + MAM", and "MG-VMD (Ours)". Both MAM and MEDM exploit motion inconsistency cues, demonstrating better performance compared to the baseline. The full model combines the complementary outputs from MAM and MEDM, yielding the best results. + +Our baseline outperforms state-of-the-art VMD-Net [12] for two reasons. First, our model is fine-tuned on VMD-Net's [12] ResNext-101 backbone and DeepLabV3 weights (Section 5). Second, VMD-Net [12] assumes dual correspondence, but in real-life datasets, dual correspondence does not always occur. This shows the usefulness of our MMD dataset. + +Table 4 shows that MEDM and MAM each contributes to the model individually. Without MEDM, mirror edges become noisy (e.g., Figure 10 Column 4 Rows 2 and 4). Without MAM, MEDM may mistakenly detect object edges inside mirrors (e.g., cabinet edges in Figure 10 Column 3 Row 2), leading to inaccurate mirror detection. (Best viewed by zooming in.) Both modules use motion features, complementing each other and boost performance (Figure 10 Column 5, and Table 4). + +# 6. Conclusion + +This paper introduces a novel deep-learning approach to tackle video mirror detection by leveraging inconsistent + +![](images/15d4f9447aaa4fcb42a89d36b5392d61dfa4b40d163684e32178f2a405cc3c06.jpg) +Figure 10. Qualitative results of the ablated models. + +![](images/0004acba8a3a80f61163a5505f1f5de033e94223b9cfba8b3d73e5eea875dea7.jpg) +Figure 11. Failure cases. Our Method may not be good in scenarios containing both mirror regions and windows, due to windows irregular motion compared to a scene. + +motion cues within and outside mirrored regions. We also present a benchmark video mirror dataset featuring consistent inter-frame motion across long videos. Our experimental results demonstrate that our proposed MG-VMD deep-learning method outperforms state-of-the-arts in Video Mirror Detection, Image-Based Mirror Detection, and Video Salient Object Detection. Furthermore, we conclude that our proposed Motion Mirror Dataset (MMD) is challenging compared to state-of-the-art Video Mirror Datasets. + +Our method is not without limitation. Figure 11 shows a scenario with both mirror and window in a scene. We found that similar to mirror regions, windows can have different motions from their surroundings. Thus, our method tends to over-predict window regions as mirrors. We plan to improve our design to mitigate this in the future. + +Acknowledgements: Alex is supported by a Swansea GTA Research Scholarship. This project is in part supported by a GRF grant from the Research Grants Council of Hong Kong (Ref.: 11211223). We gratefully acknowledge the support of the HEFCW HERC fund (W21/21HE) for the provision of GPU equipment used in this research. For the purpose of Open Access, the author has applied a Creative Commons Attribution (CC BY) license to any Author Accepted Manuscript (AAM) version arising from this submission. + +# References + +[1] Radhakrishna Achanta, Sheila Hemami, Francisco Estrada, and Sabine Susstrunk. Frequency-tuned salient region detection. In CVPR, 2009. 6 +[2] Liang-Chieh Chen, George Papandreou, Florian Schroff, and Hartwig Adam. Rethinking atrous convolution for semantic image segmentation, 2017. 4, 5 +[3] Zheng Dong, Ke Xu, Ziheng Duan, Hujun Bao, Weiwei Xu, and Rynson W. H. Lau. Geometry-aware two-scale pifu representation for human reconstruction. In NeurIPS, 2022. 1 +[4] Zheng Dong, Ke Xu, Yaoan Gao, Qilin Sun, Hujun Bao, Weiwei Xu, and Rynson W. H. Lau. Sailor: Synergizing radiance and occupancy fields for live human performance capture. ACM Trans. Graph., 2023. 1 +[5] Gunnar Farnebäck. Two-frame motion estimation based on polynomial expansion. In Proceedings of the 13th Scandinavian Conference on Image Analysis, 2003. 3 +[6] Yuan-Chen Guo, Di Kang, Linchao Bao, Yu He, and Song-Hai Zhang. Nerfren: Neural radiance fields with reflections. In CVPR, 2022. 1 +[7] Eddy Ig, Nikolaus Mayer, Tonmoy Saikia, Margret Keuper, Alexey Dosovitskiy, and Thomas Brox. Flownet 2.0: Evolution of optical flow estimation with deep networks. In CVPR, 2017. 3 +[8] Ge-Peng Ji, Keren Fu, Zhe Wu, Deng-Ping Fan, Jianbing Shen, and Ling Shao. Full-duplex strategy for video object segmentation. In ICCV, 2021. 2, 7, 8 +[9] Qingming Huang Jun Wei, Shuhui Wang. F3net: Fusion, feedback and focus for salient object detection. In AAAI, 2020. 2, 7, 8 +[10] Haofeng Li, Guanqi Chen, Guanbin Li, and Yizhou Yu. Motion guided attention for video salient object detection. In ICCV, 2019. 5, 7, 8 +[11] Jiaying Lin, Guodong Wang, and Rynson W.H. Lau. Progressive mirror detection. In CVPR, 2020. 1, 2, 5, 6, 7, 8 +[12] Jiaying Lin, Xin Tan, and Rynson W.H. Lau. Learning to detect mirrors from videos via dual correspondences. In CVPR, 2023. 1, 2, 3, 6, 7, 8 +[13] Zhiqiu Lin, Jin Sun, Abe Davis, and Noah Snavely. Visual chirality. In CVPR, 2020. 1 +[14] Jing Liu, Jiaxiang Wang, Weikang Wang, and Yuting Su. Ds-net: Dynamic spatiotemporal network for video salient object detection. arXiv:2012.04886, 2022. 2 +[15] Haiyang Mei, Bo Dong, Wen Dong, Pieter Peers, Xin Yang, Qiang Zhang, and Xiaopeng Wei. Depth-aware mirror segmentation. In CVPR, 2021. 1, 2 +[16] Anwesan Pal, Sayan Mondal, and Henrik I Christensen. "looking at the right stuff"-guided semantic-gaze for autonomous driving. In CVPR, 2020. 1 +[17] Yukun Su, Jingliang Deng, Ruizhou Sun, Guosheng Lin, and Qingyao Wu. A unified transformer framework for group-based segmentation: Co-segmentation, co-saliency detection and video salient object detection. IEEE Trans. Multimedia, 2023. 7, 8 +[18] Hideki Tamura, Maki Tsukuda, Hiroshi Higashi, and Shigeki Nakauchi. Perceptual segregation between mirror and glass + +material under natural and unnatural illumination. Journal of Vision, 2016. 2 +[19] Hideki Tamura, Hiroshi Higashi, and Shigeki Nakauchi. Multiple cues for visual perception of mirror and glass materials. Journal of Vision, 2017. 2 +[20] Hideki Tamura, Hiroshi Higashi, and Shigeki Nakauchi. Dynamic visual cues for differentiating mirror and glass. Scientific Reports, 2018. 2 +[21] Xin Tan, Ke Xu, Ying Cao, Yiheng Zhang, Lizhuang Ma, and Rynson W. H. Lau. Night-time scene parsing with a large real dataset. IEEE Transactions on Image Processing, 2021. 1 +[22] Xin Tan, Jiaying Lin, Ke Xu, Pan Chen, Lizhuang Ma, and Rynson W.H. Lau. Mirror detection with the visual chirality cue. IEEE Trans. Pattern Anal. Mach. Intell., 2022. 1, 2, 5 +[23] Yi Tang, Yuanman Li, and Guoliang Xing. Video salient object detection via adaptive local-global refinement. arXiv:2104.14360, 2021. 2, 7, 8 +[24] Zachary Teed and Jia Deng. Raft: Recurrent all-pairs field transforms for optical flow. In ECCV, 2020. 3, 5 +[25] Xin Tian, Ke Xu, Xin Yang, Lin Du, Baocai Yin, and Rynson WH Lau. Bi-directional object-context prioritization learning for saliency ranking. In CVPR, 2022. 1 +[26] Xin Tian, Ke Xu, Xin Yang, Baocai Yin, and Rynson Lau. Learning to detect instance-level salient objects using complementary image labels. Int. J. Comput. Vis., 2022. 1 +[27] Pavel Tokmakov, Karteek Alahari, and Cordelia Schmid. Learning video object segmentation with visual memory. arXiv:1704.05737, 2017. 2 +[28] Wenguan Wang, Xiankai Lu, Jianbing Shen, David Crandall, and Ling Shao. Zero-shot video object segmentation via attentive graph neural networks. In ICCV, 2019. 2 +[29] Saining Xie, Ross Girshick, Piotr Dólár, Zhuowen Tu, and Kaiming He. Aggregated residual transformations for deep neural networks. In CVPR, 2017. 4, 5 +[30] Xin Yang, Haiyang Mei, Ke Xu, Xiaopeng Wei, Baocai Yin, and Rynson W. H. Lau. Where is my mirror? In ICCV, 2019. 1, 2, 7, 8 +[31] Junyi Zeng, Chong Bao, Rui Chen, Zilong Dong, Guofeng Zhang, Hujun Bao, and Zhaopeng Cui. Mirror-nerf: Learning neural radiance fields for mirrors with whitted-style ray tracing. In ACM MM, 2023. 1 \ No newline at end of file diff --git a/2024/Effective Video Mirror Detection with Inconsistent Motion Cues/images.zip b/2024/Effective Video Mirror Detection with Inconsistent Motion Cues/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..18c3fe725410d44fa0d6a9be7ee2bf2b746ea399 --- /dev/null +++ b/2024/Effective Video Mirror Detection with Inconsistent Motion Cues/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:11a407496349cdfeed6b9badb5c9109431f5f9e7d33c08975900ce88709fb813 +size 545163 diff --git a/2024/Effective Video Mirror Detection with Inconsistent Motion Cues/layout.json b/2024/Effective Video Mirror Detection with Inconsistent Motion Cues/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..851a9ea6444260daee105e23d15106ba47db5f0d --- /dev/null +++ b/2024/Effective Video Mirror Detection with Inconsistent Motion Cues/layout.json @@ -0,0 +1,7653 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 99, + 103, + 495, + 119 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 99, + 103, + 495, + 119 + ], + "spans": [ + { + "bbox": [ + 99, + 103, + 495, + 119 + ], + "type": "text", + "content": "Effective Video Mirror Detection with Inconsistent Motion Cues" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 104, + 142, + 489, + 158 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 104, + 142, + 489, + 158 + ], + "spans": [ + { + "bbox": [ + 104, + 142, + 489, + 158 + ], + "type": "text", + "content": "Alex Warren1, Ke Xu2, Jiaying Lin2, Gary K.L. Tam1, Rynson W.H. Lau1,2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 159, + 487, + 172 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 105, + 159, + 487, + 172 + ], + "spans": [ + { + "bbox": [ + 105, + 159, + 487, + 172 + ], + "type": "text", + "content": "Department of Computer Science, Swansea University1 and City University of Hong Kong2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 91, + 174, + 502, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 174, + 502, + 185 + ], + "spans": [ + { + "bbox": [ + 91, + 174, + 502, + 185 + ], + "type": "text", + "content": "alex.warren@swansea.ac.uk, kkangwing@gmail.com, jiayinlin5-c@my.cityu.edu.hk" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 170, + 188, + 421, + 198 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 170, + 188, + 421, + 198 + ], + "spans": [ + { + "bbox": [ + 170, + 188, + 421, + 198 + ], + "type": "text", + "content": "k.l.tam@swansea.ac.uk, Rynson.Lau@cityu.edu.hk" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 143, + 227, + 192, + 239 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 227, + 192, + 239 + ], + "spans": [ + { + "bbox": [ + 143, + 227, + 192, + 239 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 251, + 290, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 251, + 290, + 601 + ], + "spans": [ + { + "bbox": [ + 47, + 251, + 290, + 601 + ], + "type": "text", + "content": "Image-based mirror detection has recently undergone rapid research due to its significance in applications such as robotic navigation, semantic segmentation and scene reconstruction. Recently, VMD-Net was proposed as the first video mirror detection technique, by modeling dual correspondences between the inside and outside of the mirror both spatially and temporally. However, this approach is not reliable, as correspondences can occur completely inside or outside of the mirrors. In addition, the proposed dataset VMD-D contains many small mirrors, limiting its applicability to real-world scenarios. To address these problems, we developed a more challenging dataset that includes mirrors of various shapes and sizes at different locations of the frames, providing a better reflection of real-world scenarios. Next, we observed that the motions between the inside and outside of the mirror are often inconsistent. For instance, when moving in front of a mirror, the motion inside the mirror is often much smaller than the motion outside due to increased depth perception. With these observations, we propose modeling inconsistent motion cues to detect mirrors, and a new network with two novel modules. The Motion Attention Module (MAM) explicitly models inconsistent motions around mirrors via optical flow, and the Motion-Guided Edge Detection Module (MEDM) uses motions to guide mirror edge feature learning. Experimental results on our proposed dataset show that our method outperforms state-of-the-arts. The code and dataset are available at https://github.com/ AlexAnthonyWarren/MG-VMD." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 621, + 128, + 634 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 621, + 128, + 634 + ], + "spans": [ + { + "bbox": [ + 47, + 621, + 128, + 634 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 642, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 642, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 642, + 287, + 715 + ], + "type": "text", + "content": "Mirrors and reflective surfaces are abundant in our surroundings. Successful detection of mirrors underpins many vision applications such as autonomous drone navigation and vehicle vision systems [16]. When a scene contains mirrored regions, mirror identification is critical to reducing errors in applications such as salient object de" + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 311, + 228, + 541, + 388 + ], + "blocks": [ + { + "bbox": [ + 311, + 228, + 541, + 388 + ], + "lines": [ + { + "bbox": [ + 311, + 228, + 541, + 388 + ], + "spans": [ + { + "bbox": [ + 311, + 228, + 541, + 388 + ], + "type": "image", + "image_path": "0dfd4ee24860de30e9d715bf6232ccd9c165459ff99a7c121c670413d046e0af.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 392, + 547, + 480 + ], + "lines": [ + { + "bbox": [ + 304, + 392, + 547, + 480 + ], + "spans": [ + { + "bbox": [ + 304, + 392, + 547, + 480 + ], + "type": "text", + "content": "Figure 1. We propose to model motion inconsistency as a cue for mirror detection. The " + }, + { + "bbox": [ + 304, + 392, + 547, + 480 + ], + "type": "inline_equation", + "content": "1^{st}" + }, + { + "bbox": [ + 304, + 392, + 547, + 480 + ], + "type": "text", + "content": " column shows the inconsistent motions (depicted with arrows). The locations of these cues align well with the mirror regions. The " + }, + { + "bbox": [ + 304, + 392, + 547, + 480 + ], + "type": "inline_equation", + "content": "2^{nd}" + }, + { + "bbox": [ + 304, + 392, + 547, + 480 + ], + "type": "text", + "content": " column shows the optical flow computed between frame N (" + }, + { + "bbox": [ + 304, + 392, + 547, + 480 + ], + "type": "inline_equation", + "content": "3^{rd}" + }, + { + "bbox": [ + 304, + 392, + 547, + 480 + ], + "type": "text", + "content": " column), and frame N-1 from respective videos in our new dataset. Our method (" + }, + { + "bbox": [ + 304, + 392, + 547, + 480 + ], + "type": "inline_equation", + "content": "5^{th}" + }, + { + "bbox": [ + 304, + 392, + 547, + 480 + ], + "type": "text", + "content": " column) predicts more reliable and consistent mirror regions, outperforming VMD-Net [12] (" + }, + { + "bbox": [ + 304, + 392, + 547, + 480 + ], + "type": "inline_equation", + "content": "4^{th}" + }, + { + "bbox": [ + 304, + 392, + 547, + 480 + ], + "type": "text", + "content": " column). The " + }, + { + "bbox": [ + 304, + 392, + 547, + 480 + ], + "type": "inline_equation", + "content": "6^{th}" + }, + { + "bbox": [ + 304, + 392, + 547, + 480 + ], + "type": "text", + "content": " column is the ground truth." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 495, + 546, + 519 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 495, + 546, + 519 + ], + "spans": [ + { + "bbox": [ + 304, + 495, + 546, + 519 + ], + "type": "text", + "content": "tection [25, 26], semantic segmentation [16, 21], scene reconstruction[3, 4] and NeRF modeling [6, 31]." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 521, + 545, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 521, + 545, + 592 + ], + "spans": [ + { + "bbox": [ + 304, + 521, + 545, + 592 + ], + "type": "text", + "content": "Mirror detection involves binary classification of individual pixels in RGB images to determine if they are inside or outside of the mirror. It is a challenging research as mirrors typically do not have their own visual appearance. They instead reflect objects within their surroundings. This visual attribute makes mirrors difficult to detect." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 594, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 547, + 714 + ], + "type": "text", + "content": "There are a few image-based and one video-based mirror detection methods in the literature. Yang et al. [30] first introduced a mirror detection method using contrast and semantics. Lin et al. [11] introduced a PMD-Net that utilizes edge detection and leverages contextual contrasting objects inside and outside of mirrored regions to detect mirrored regions. Tan et al. [22] detect mirrors using Visual Chirality Cue [13] that encodes mirror-like symmetrical properties at the pixel level. Mei et al. [15] factor depth estimation into their mirror detection process and show good results" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "17244" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 108 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 108 + ], + "type": "text", + "content": "in RGB and RGB-D mirror datasets. Recently, Lin et al. [12] introduced the first video mirror detection method using inter-frame and intra-frame dual correspondences." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 110, + 287, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 110, + 287, + 240 + ], + "spans": [ + { + "bbox": [ + 46, + 110, + 287, + 240 + ], + "type": "text", + "content": "Despite these exciting results, we observe research gaps and limitations in these works, especially in video mirror detection. The state-of-the-art VMD-Net [12] models dual correspondences inside and outside of mirrors to aid in the detection of mirror regions in videos. However, the approach is not reliable as correspondences can appear completely inside or outside of mirrors as discussed in [22]. In addition, the proposed VMD-D dataset contains primarily small mirrors, with small inter-frame motion, mirror coverage and content variation. This limits its applicability to mirror detection in real-world scenarios." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 243, + 287, + 469 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 243, + 287, + 469 + ], + "spans": [ + { + "bbox": [ + 46, + 243, + 287, + 469 + ], + "type": "text", + "content": "To address these issues, we first developed a new, more challenging dataset, Mirror Motion Dataset (MMD). The dataset contains 37 (" + }, + { + "bbox": [ + 46, + 243, + 287, + 469 + ], + "type": "inline_equation", + "content": "\\approx" + }, + { + "bbox": [ + 46, + 243, + 287, + 469 + ], + "type": "text", + "content": "9-second) videos with substantial motion, each accompanied by manual mirror annotations. To address the small mirror issue in VMD-D, we capture a high variance of mirror sizes and shapes at different spatial locations, generalizing to real-world scenarios. We further include challenging examples, including low-contrast environments, multi-/large/full-mirror scenes and under different lighting conditions. Next, we observe that the motions between the inside and outside of mirrors are often inconsistent. For instance, when moving in front of a mirror, the motion inside is often smaller than outside due to the increased depth perception. The inconsistent motion cues (depicted by arrows in the " + }, + { + "bbox": [ + 46, + 243, + 287, + 469 + ], + "type": "inline_equation", + "content": "1^{st}" + }, + { + "bbox": [ + 46, + 243, + 287, + 469 + ], + "type": "text", + "content": " column in Figure 1) also apply in challenging scenarios like low-contrast and multi-mirror scenes. These observations motivate us to ask the question: Can we apply the motion inconsistency cue to improve mirror detection in challenging complex video scenes?" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 472, + 287, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 472, + 287, + 578 + ], + "spans": [ + { + "bbox": [ + 46, + 472, + 287, + 578 + ], + "type": "text", + "content": "Neuroscience studies suggest that humans often rely on dynamic perceptual cues to identify mirrored regions in daily life. Tamura et al. [18] [19] show that humans are capable of distinguishing mirrors in dynamic scenes involving rotational movement. [20] highlights that rotational, parallax, forward and backward motions are exploitable motion inconsistencies for mirror detection. To our knowledge, such motion inconsistency cues have not been utilized before for the video mirror detection task." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 580, + 287, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 580, + 287, + 687 + ], + "spans": [ + { + "bbox": [ + 46, + 580, + 287, + 687 + ], + "type": "text", + "content": "To this end, we propose a novel network to model motion inconsistency for mirror detection, with two novel modules. First, the Motion Attention Module (MAM) predicts mirror regions by exploiting inconsistent motion cues obtained from optical flow fields. Second, the Motion-guided Edge Detection Module (MEDM) further uses motion to guide mirror edge detection. Experimental results show that the proposed network outperforms relevant state-of-the-art methods. To summarise, our main contributions include:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "content": "- We propose a novel deep-learning video mirror detection method based on the motion inconsistency cue." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 72, + 545, + 227 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 306, + 72, + 545, + 131 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 545, + 131 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 545, + 131 + ], + "type": "text", + "content": "- Our network includes two novel modules. Motion Attention Module (MAM) attentively predicts mirror regions using motion inconsistency in the optical flow fields. Motion-guided Edge Detection Module (MEDM) uses the motions to guide mirror edge feature extraction." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 133, + 545, + 202 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 133, + 545, + 202 + ], + "spans": [ + { + "bbox": [ + 306, + 133, + 545, + 202 + ], + "type": "text", + "content": "- We further present a new video dataset (MMD). Compared to VMD-D, our dataset contains mirrors of various shapes and size at different locations of the images. It also covers various scenes, e.g., low-light scenes, and multi-large/full mirror scenes. It is a more challenging dataset, more suitable for real-life scenarios." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 204, + 545, + 227 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 204, + 545, + 227 + ], + "spans": [ + { + "bbox": [ + 306, + 204, + 545, + 227 + ], + "type": "text", + "content": "- Extensive evaluations show that the proposed model outperforms existing methods on video mirror detection." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 239, + 397, + 251 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 239, + 397, + 251 + ], + "spans": [ + { + "bbox": [ + 306, + 239, + 397, + 251 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 258, + 545, + 377 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 258, + 545, + 377 + ], + "spans": [ + { + "bbox": [ + 304, + 258, + 545, + 377 + ], + "type": "text", + "content": "Video Mirror Detection. VMD-Net [12] introduced the first video mirror detection method. It employs dual correspondences across both inter-frame and intra-frame to predict mirrored regions. They further presented a video mirror dataset for this purpose. However, despite its demonstrated success, we observe two limitations. First, the idea of dual correspondences is not reliable as correspondences can occur completely inside or outside of the mirrors [22]. Second, their dataset contains an excess of small mirrors, which does not generalize well to real-world scenarios." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 379, + 545, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 379, + 545, + 437 + ], + "spans": [ + { + "bbox": [ + 304, + 379, + 545, + 437 + ], + "type": "text", + "content": "This paper aims to facilitate video mirror detection in the wild from two perspectives. First, we propose to model the inconsistent motion cues to learn more robust mirror representations. Second, we introduce a new benchmark dataset comprising videos with more complex scenarios." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 438, + 545, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 438, + 545, + 569 + ], + "spans": [ + { + "bbox": [ + 304, + 438, + 545, + 569 + ], + "type": "text", + "content": "Image-Based Mirror Detection. [30] pioneered mirror detection by utilizing contextual contrasting features and introducing a mirror detection dataset. [15] refined detection using RGB-D data and presented the first RGB-D mirror dataset. [11] introduced a method leveraging contextual contrasts to relate mirror interior and exterior contents, along with an expanded image-based mirror dataset. [22] incorporated chirality cues for mirror detection. Although these work achieved good results, they mainly focus on single-image mirror detection and lack the capacity to identify mirrors within temporal videos." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 570, + 545, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 545, + 700 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 545, + 700 + ], + "type": "text", + "content": "Video Salient Object Detection. Video Salient Object Detection (VSOD) seeks to identify the most visually prominent object in videos. Early VSOD studies [27] [28] utilized manually crafted features to detect these salient objects. In more recent research [9] [8] [9] [14] [23], the focus has shifted to deep learning models, which have demonstrated promising results in VSOD. These techniques however are not optimized for the video mirror detection task. This is because mirrored regions may not always represent the most visually significant objects within a scene. Our technique aims to identify mirrors in real-life scenarios." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 701, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 701, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 701, + 545, + 713 + ], + "type": "text", + "content": "Optical Flow Estimation. Optical flow estimation fo" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17245" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 53, + 70, + 541, + 218 + ], + "blocks": [ + { + "bbox": [ + 53, + 70, + 541, + 218 + ], + "lines": [ + { + "bbox": [ + 53, + 70, + 541, + 218 + ], + "spans": [ + { + "bbox": [ + 53, + 70, + 541, + 218 + ], + "type": "image", + "image_path": "f336c6d5b1be67764b040e70639b9b997823bba7bad736a5844ac93c660c1564.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 88, + 220, + 503, + 232 + ], + "lines": [ + { + "bbox": [ + 88, + 220, + 503, + 232 + ], + "spans": [ + { + "bbox": [ + 88, + 220, + 503, + 232 + ], + "type": "text", + "content": "Figure 2. Snapshots of our proposed Motion Mirror Dataset (MMD), with pixel-level mirror and edge annotations." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 50, + 242, + 285, + 376 + ], + "blocks": [ + { + "bbox": [ + 50, + 242, + 285, + 376 + ], + "lines": [ + { + "bbox": [ + 50, + 242, + 285, + 376 + ], + "spans": [ + { + "bbox": [ + 50, + 242, + 285, + 376 + ], + "type": "image", + "image_path": "e2803215de5f929742c0998604c203196ac6331bdcb616cfd48c907e9dd1a766.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 379, + 281, + 390 + ], + "lines": [ + { + "bbox": [ + 54, + 379, + 281, + 390 + ], + "spans": [ + { + "bbox": [ + 54, + 379, + 281, + 390 + ], + "type": "text", + "content": "Figure 3. Cumulative Distributions. Figure 5. Color Contrast." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 404, + 287, + 523 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 404, + 287, + 523 + ], + "spans": [ + { + "bbox": [ + 46, + 404, + 287, + 523 + ], + "type": "text", + "content": "cuses on extracting per-pixel motion from video sequences. Early efforts [5] used hand-crafted features of image brightness between consecutive frames. In recent studies, convolutional deep learning and transformer-based approaches have made significant improvements [24] [7]. In our context, optical flow estimation is not inherently applicable to video mirror detection and cannot serve as a direct detection method. This study aims to leverage deep-learning optical flow as a way to model inconsistent motion cues, which enables a robust end-to-end video mirror detection framework." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 537, + 151, + 551 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 537, + 151, + 551 + ], + "spans": [ + { + "bbox": [ + 47, + 537, + 151, + 551 + ], + "type": "text", + "content": "3. Proposed Dataset" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 558, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 558, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 558, + 288, + 715 + ], + "type": "text", + "content": "To create our Motion Mirror Dataset (MMD), we first recruited six volunteering participants (final year university students) to capture 37 nine-second smartphone videos from their surroundings (e.g., living rooms, bathrooms). Participants were instructed to maintain consistent motion around mirror regions while recording. We recorded videos under various lighting conditions, including daily light and low-light settings. The dataset was subsequently divided into non-overlapping training (4,653 samples) and testing (5,074 samples) sets. To accomplish this, we randomly split the 37 videos into the two sets (training and testing). Specifically 18 videos (4,653 samples) are put into the training set and 19 videos (5,074 samples) are put into the testing" + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 306, + 239, + 547, + 277 + ], + "blocks": [ + { + "bbox": [ + 306, + 239, + 547, + 277 + ], + "lines": [ + { + "bbox": [ + 306, + 239, + 547, + 277 + ], + "spans": [ + { + "bbox": [ + 306, + 239, + 547, + 277 + ], + "type": "table", + "html": "
DatasetSSIM ↑
VMD-D0.7950544833184201
Ours (MotionMirrorDataset)0.9033545507233643
", + "image_path": "1d998f791eaa10f88eb6ebb4214696080fd2da362136af1cd862e87eb19c1030.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 313, + 316, + 533, + 376 + ], + "blocks": [ + { + "bbox": [ + 305, + 282, + 545, + 304 + ], + "lines": [ + { + "bbox": [ + 305, + 282, + 545, + 304 + ], + "spans": [ + { + "bbox": [ + 305, + 282, + 545, + 304 + ], + "type": "text", + "content": "Table 1. SSIM of MMD and VMD-D. A higher SSIM indicates a more reasonable split of dataset." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 313, + 316, + 533, + 376 + ], + "lines": [ + { + "bbox": [ + 313, + 316, + 533, + 376 + ], + "spans": [ + { + "bbox": [ + 313, + 316, + 533, + 376 + ], + "type": "table", + "html": "
DatasetAreas not covered by mirrors ↓
(Ours) MMD Training0.000061
(Ours) MMD Testing0.000015
VMD-D Training0.00038147
VMD-D Testing0.00024414
", + "image_path": "9e2b8104320d7e1bd0d2f1fdea6f1ed2cd106f0ef4860dc4d27e98631c0d65d7.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 378, + 545, + 400 + ], + "lines": [ + { + "bbox": [ + 306, + 378, + 545, + 400 + ], + "spans": [ + { + "bbox": [ + 306, + 378, + 545, + 400 + ], + "type": "text", + "content": "Table 2. Mirror Coverage. A smaller non-coverage rate indicates a more even spatial distribution of mirrors." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 414, + 545, + 474 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 414, + 545, + 474 + ], + "spans": [ + { + "bbox": [ + 304, + 414, + 545, + 474 + ], + "type": "text", + "content": "set. Each frame was meticulously annotated with pixel-level mirror masks. We applied a depth-first search algorithm to the pixel-level mirror masks to extract edge masks, which we then manually verified for ground truth accuracy. All videos in the dataset have a frame rate of 30 fps." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 474, + 545, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 474, + 545, + 521 + ], + "spans": [ + { + "bbox": [ + 304, + 474, + 545, + 521 + ], + "type": "text", + "content": "Statistic Analysis. We compare our proposed MMD with VMD-D [12] in terms of Cumulative Distribution and Structural Similarity between training and testing sets, Mirror Size Distribution, and Area Coverage." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 521, + 546, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 521, + 546, + 689 + ], + "spans": [ + { + "bbox": [ + 304, + 521, + 546, + 689 + ], + "type": "text", + "content": "Figure 3 shows the cumulative distribution of all mirrors in both training and testing sets. Compared to VMD-D, our MMD dataset contains more intricate scenarios. VMD-D [12] however is highly biased with majority of mirror regions fall in the center in the train/test sets whilst there are nearly none along the edge (darker color on left/right). The mirror regions of our proposed dataset (MMD) are spread out across the entire heatmap where mirrors are also found at the image edges. We further compute the Structural Similarity Index (SSIM) on the cumulative area distributions (Figure 3) directly to compare the structure between the whole training and testing sets. The higher SSIM of our MMD in Table 1 shows that the training and testing sets are similar in structure and fairer for evaluation purpose." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 689, + 547, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 547, + 712 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 547, + 712 + ], + "type": "text", + "content": "Figure 4 shows the size distribution of mirrors in both MMD and VMD-D datasets. VMD-D contains a lot of" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17246" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 60, + 80, + 541, + 308 + ], + "blocks": [ + { + "bbox": [ + 60, + 80, + 541, + 308 + ], + "lines": [ + { + "bbox": [ + 60, + 80, + 541, + 308 + ], + "spans": [ + { + "bbox": [ + 60, + 80, + 541, + 308 + ], + "type": "image", + "image_path": "228c0a33d6db9084e592569962c0eeff8d0207a5c8ea770bec02354033962aa8.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 315, + 548, + 404 + ], + "lines": [ + { + "bbox": [ + 46, + 315, + 548, + 404 + ], + "spans": [ + { + "bbox": [ + 46, + 315, + 548, + 404 + ], + "type": "text", + "content": "Figure 6. Method overview. Our MG-VMD network takes three adjacent frames " + }, + { + "bbox": [ + 46, + 315, + 548, + 404 + ], + "type": "inline_equation", + "content": "(\\mathrm{I}_{\\mathrm{N - 2}},\\mathrm{I}_{\\mathrm{N - 1}},\\mathrm{I}_{\\mathrm{N}})" + }, + { + "bbox": [ + 46, + 315, + 548, + 404 + ], + "type": "text", + "content": " as inputs and extracts their multi-scale features via a shared ResNext-101 [29] backbone and DeepLabV3[2]. We apply the optical flow coherence block to process " + }, + { + "bbox": [ + 46, + 315, + 548, + 404 + ], + "type": "inline_equation", + "content": "\\mathrm{I}_{\\mathrm{N - 2}}" + }, + { + "bbox": [ + 46, + 315, + 548, + 404 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 315, + 548, + 404 + ], + "type": "inline_equation", + "content": "\\mathrm{I}_{\\mathrm{N - 1}}" + }, + { + "bbox": [ + 46, + 315, + 548, + 404 + ], + "type": "text", + "content": " to compute an optical flow map for " + }, + { + "bbox": [ + 46, + 315, + 548, + 404 + ], + "type": "inline_equation", + "content": "\\mathrm{I}_{\\mathrm{N - 1}}" + }, + { + "bbox": [ + 46, + 315, + 548, + 404 + ], + "type": "text", + "content": ", followed by an optical flow vector field extrapolation to estimate the optical flow map for " + }, + { + "bbox": [ + 46, + 315, + 548, + 404 + ], + "type": "inline_equation", + "content": "\\mathrm{I}_{\\mathrm{N}}" + }, + { + "bbox": [ + 46, + 315, + 548, + 404 + ], + "type": "text", + "content": ". These two optical flow maps are fed into our proposed MAM and MEDM modules to guide the mirror localization and mirror-edge detection processes, respectively. The proposed Motion Attention Module (MAM) learns to locate mirrors from optical flow fields, exploiting motion cues, attentively modeling motion inconsistency. The proposed Motion-guided Edge Detection Module (MEDM) learns to extract and refine mirror boundary edge features. Finally, a fusion refinement module is used to predict the mirror maps by fusing the predicted mirror boundary maps of MEDM, the mirror localization maps of MAM, and the extracted multi-scale image backbone features." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 415, + 288, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 415, + 288, + 475 + ], + "spans": [ + { + "bbox": [ + 46, + 415, + 288, + 475 + ], + "type": "text", + "content": "small mirrors with few/no large mirrors, which does not represent real-world scenarios. In contrast, our MMD dataset has a more reasonable distribution, containing a variety of small, medium and large sized mirrors, and therefore is more challenging than VMD-D." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 476, + 287, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 476, + 287, + 548 + ], + "spans": [ + { + "bbox": [ + 46, + 476, + 287, + 548 + ], + "type": "text", + "content": "Table 2 shows the areas (ratio to the whole image) where there is no mirror coverage within the two datasets. The smaller ratios show that MMD contains mirrors that are more evenly spatially distributed across the whole image. MMD is more challenging and will introduce less spatial bias in the model learning." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 548, + 288, + 657 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 548, + 288, + 657 + ], + "spans": [ + { + "bbox": [ + 46, + 548, + 288, + 657 + ], + "type": "text", + "content": "We further compare the color contrast between mirror and non-mirror regions in all images between MMD and VMD-D. The histogram in Figure 5 shows that MMD contains more images with lower color contrast between mirror and non-mirror regions. Lower color contrast implies a decrease in the amount of contextual information available for a model to leverage, making it more challenging to detect mirrors. The " + }, + { + "bbox": [ + 46, + 548, + 288, + 657 + ], + "type": "inline_equation", + "content": "4^{\\text{th}}" + }, + { + "bbox": [ + 46, + 548, + 288, + 657 + ], + "type": "text", + "content": " column of Figure 2 shows one example. These show that MMD is more challenging than VMD-D." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 669, + 153, + 682 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 669, + 153, + 682 + ], + "spans": [ + { + "bbox": [ + 47, + 669, + 153, + 682 + ], + "type": "text", + "content": "4. Proposed Method" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 689, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 689, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 689, + 288, + 715 + ], + "type": "text", + "content": "Our proposed method, Motion-guided Video Mirror Detection (MG-VMD), aims to enhance mirror detection by" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 415, + 545, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 415, + 545, + 475 + ], + "spans": [ + { + "bbox": [ + 304, + 415, + 545, + 475 + ], + "type": "text", + "content": "modelling motion inconsistency cues. Drawing inspiration from our own observations and neuroscience studies, we hypothesise that modeling motion inconsistency cues can improve the accuracy of mirror detection, especially in dynamic video scenarios." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 479, + 545, + 563 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 479, + 545, + 563 + ], + "spans": [ + { + "bbox": [ + 304, + 479, + 545, + 563 + ], + "type": "text", + "content": "The core idea behind MG-VMD is to utilize optical flow to model motion inconsistency and use it to guide the extraction of multi-scale image and edge features, allowing for the temporal detection of mirrored regions in videos. To achieve this, we introduce two key modules: the Motion Attention Module (MAM) and the Motion-guided Edge Detection Module (MEDM)." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 566, + 545, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 566, + 545, + 662 + ], + "spans": [ + { + "bbox": [ + 304, + 566, + 545, + 662 + ], + "type": "text", + "content": "Specifically, the Motion Attention Module (MAM) is designed to explicitly predict mirror locations by modelling motion inconsistency cues from the underlying optical flow fields. This module enhances video mirror prediction by providing spatial contextual information through the multiscale image features and motion cues. It informs the subsequent stages of the model about the location of mirror regions in the video." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 665, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 547, + 715 + ], + "type": "text", + "content": "The Motion-guided Edge Detection Module (MEDM) is dedicated to predicting mirror boundaries, contributing to the enhancement of overall mirror region predictions. This module takes advantage of image features and employs mo" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17247" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 52, + 72, + 286, + 144 + ], + "blocks": [ + { + "bbox": [ + 52, + 72, + 286, + 144 + ], + "lines": [ + { + "bbox": [ + 52, + 72, + 286, + 144 + ], + "spans": [ + { + "bbox": [ + 52, + 72, + 286, + 144 + ], + "type": "image", + "image_path": "4214f93536c8abbf728bdc63c5971369f1817a668c77eb59c7ad31826e05faa4.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 156, + 287, + 177 + ], + "lines": [ + { + "bbox": [ + 46, + 156, + 287, + 177 + ], + "spans": [ + { + "bbox": [ + 46, + 156, + 287, + 177 + ], + "type": "text", + "content": "Figure 7. The proposed MAM (Motion Attention Module) aims to model the inconsistent motion cues to localize the mirrors." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 185, + 287, + 208 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 185, + 287, + 208 + ], + "spans": [ + { + "bbox": [ + 46, + 185, + 287, + 208 + ], + "type": "text", + "content": "tion inconsistency cues to guide the detection of features at the edges, specifically tailored for mirror detection." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 209, + 287, + 316 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 209, + 287, + 316 + ], + "spans": [ + { + "bbox": [ + 46, + 209, + 287, + 316 + ], + "type": "text", + "content": "The Fusion Refinement Module, inspired from PMD-Net [11], further takes the MAM predicted mirror locations and the MEDM refined mirror boundaries to predict mirrors. The integration plays a pivotal role in improving mirror prediction by alleviating errors arising from the limited capability of edge detection or inconsistent motion cues alone. It overcomes the limitations of individual modules and provides spatial and temporal context for accurate mirror prediction." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 316, + 287, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 316, + 287, + 472 + ], + "spans": [ + { + "bbox": [ + 46, + 316, + 287, + 472 + ], + "type": "text", + "content": "In Figure 6 (Architecture Overview), the MG-VMD processes adjacent images " + }, + { + "bbox": [ + 46, + 316, + 287, + 472 + ], + "type": "inline_equation", + "content": "(\\mathrm{I}_{\\mathrm{N - 2}},\\mathrm{I}_{\\mathrm{N - 1}},\\mathrm{I}_{\\mathrm{N}})" + }, + { + "bbox": [ + 46, + 316, + 287, + 472 + ], + "type": "text", + "content": " as inputs. Optical flow field extrapolation is employed to enhance coherency and address temporal errors between adjacent branches (N-1 and N). The model utilizes a shared ResNext-101 [29] backbone and DeepLabV3 [2] to extract multi-scale image features. The outputs include predicted mirror binary maps " + }, + { + "bbox": [ + 46, + 316, + 287, + 472 + ], + "type": "inline_equation", + "content": "(\\mathbf{M}_{\\mathrm{N - 1}}" + }, + { + "bbox": [ + 46, + 316, + 287, + 472 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 316, + 287, + 472 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_{\\mathrm{N}}" + }, + { + "bbox": [ + 46, + 316, + 287, + 472 + ], + "type": "text", + "content": " ), predicted mirror-boundary edge maps " + }, + { + "bbox": [ + 46, + 316, + 287, + 472 + ], + "type": "inline_equation", + "content": "(\\mathbf{E}_{\\mathrm{N - 1}}" + }, + { + "bbox": [ + 46, + 316, + 287, + 472 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 316, + 287, + 472 + ], + "type": "inline_equation", + "content": "\\mathbf{E}_{\\mathrm{N}}" + }, + { + "bbox": [ + 46, + 316, + 287, + 472 + ], + "type": "text", + "content": " ), and mirror motion predictions " + }, + { + "bbox": [ + 46, + 316, + 287, + 472 + ], + "type": "inline_equation", + "content": "(\\mathbf{M}\\mathbf{p}_{\\mathrm{N - 1}}" + }, + { + "bbox": [ + 46, + 316, + 287, + 472 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 316, + 287, + 472 + ], + "type": "inline_equation", + "content": "\\mathbf{M}\\mathbf{p}_{\\mathrm{N}}" + }, + { + "bbox": [ + 46, + 316, + 287, + 472 + ], + "type": "text", + "content": " ). The collaboration of MAM, MEDM, and the Fusion Refinement Module ensures accurate and robust video mirror detection by considering both spatial and temporal aspects of mirror cues." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 479, + 212, + 491 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 479, + 212, + 491 + ], + "spans": [ + { + "bbox": [ + 47, + 479, + 212, + 491 + ], + "type": "text", + "content": "4.1. Optical Flow Coherence Block" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 498, + 286, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 498, + 286, + 665 + ], + "spans": [ + { + "bbox": [ + 46, + 498, + 286, + 665 + ], + "type": "text", + "content": "Our architecture leverages a pre-trained optical flow model. Similar to previous work [10], our optical flow block utilizes frozen weights. However, optical flow may become unreliable in scenarios with low contrast in image details or when images become saturated. To achieve a simpler design while obtaining more stable optical flow, we introduce our Optical Flow Coherence Block, inspired by the work in [24] for motion estimation. This block extracts motion and extends it using linear extrapolation. The extrapolated optical flow motion, denoted as " + }, + { + "bbox": [ + 46, + 498, + 286, + 665 + ], + "type": "inline_equation", + "content": "flo" + }, + { + "bbox": [ + 46, + 498, + 286, + 665 + ], + "type": "text", + "content": ", is then utilized in our subsequent motion-guided modules, MAM and MEDM. We empirically find that this simple design outperforms alternatives (refer to Section 6 of the Supplemental) whilst enhancing coherency." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 665, + 286, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 286, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 286, + 714 + ], + "type": "text", + "content": "Specifically, the block calculates motion vectors " + }, + { + "bbox": [ + 46, + 665, + 286, + 714 + ], + "type": "inline_equation", + "content": "f(N - 2,N - 1)" + }, + { + "bbox": [ + 46, + 665, + 286, + 714 + ], + "type": "text", + "content": " between frames " + }, + { + "bbox": [ + 46, + 665, + 286, + 714 + ], + "type": "inline_equation", + "content": "\\mathrm{I}_{\\mathrm{N - 2}}" + }, + { + "bbox": [ + 46, + 665, + 286, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 665, + 286, + 714 + ], + "type": "inline_equation", + "content": "\\mathrm{I}_{\\mathrm{N - 1}}" + }, + { + "bbox": [ + 46, + 665, + 286, + 714 + ], + "type": "text", + "content": ". These vectors are then extrapolated by one frame to estimate motion " + }, + { + "bbox": [ + 46, + 665, + 286, + 714 + ], + "type": "inline_equation", + "content": "f^{\\prime}(N - 1,N)" + }, + { + "bbox": [ + 46, + 665, + 286, + 714 + ], + "type": "text", + "content": " between " + }, + { + "bbox": [ + 46, + 665, + 286, + 714 + ], + "type": "inline_equation", + "content": "\\mathrm{I}_{\\mathrm{N - 1}}" + }, + { + "bbox": [ + 46, + 665, + 286, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 665, + 286, + 714 + ], + "type": "inline_equation", + "content": "\\mathrm{I}_{\\mathrm{N}}" + }, + { + "bbox": [ + 46, + 665, + 286, + 714 + ], + "type": "text", + "content": ". By using extrapolation, we can calculate the motion vectors of the frame " + }, + { + "bbox": [ + 46, + 665, + 286, + 714 + ], + "type": "inline_equation", + "content": "\\mathrm{I}_{\\mathrm{N - 1}}" + }, + { + "bbox": [ + 46, + 665, + 286, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 665, + 286, + 714 + ], + "type": "inline_equation", + "content": "\\mathrm{I}_{\\mathrm{N}}" + }, + { + "bbox": [ + 46, + 665, + 286, + 714 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 309, + 71, + 545, + 152 + ], + "blocks": [ + { + "bbox": [ + 309, + 71, + 545, + 152 + ], + "lines": [ + { + "bbox": [ + 309, + 71, + 545, + 152 + ], + "spans": [ + { + "bbox": [ + 309, + 71, + 545, + 152 + ], + "type": "image", + "image_path": "ef55c47ca706413ecfab9687706bb2aaf8664b6bcf40210a82c483e9d4481867.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 161, + 545, + 194 + ], + "lines": [ + { + "bbox": [ + 305, + 161, + 545, + 194 + ], + "spans": [ + { + "bbox": [ + 305, + 161, + 545, + 194 + ], + "type": "text", + "content": "Figure 8. Our proposed MEDM (Motion-Guided Edge Detection Module) aims to detect the mirror boundaries based on the captured motion cues in the optical flow vector fields." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 200, + 545, + 259 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 200, + 545, + 259 + ], + "spans": [ + { + "bbox": [ + 305, + 200, + 545, + 259 + ], + "type": "text", + "content": "tion, we connect the motion information from the frame at N-2 to N-1, and then to N. This connection helps maintain a smooth and consistent flow of motion information, reducing errors from neighboring frames and sudden changes in motion, especially around mirror regions." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 269, + 448, + 281 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 269, + 448, + 281 + ], + "spans": [ + { + "bbox": [ + 306, + 269, + 448, + 281 + ], + "type": "text", + "content": "4.2. Motion Attention Module" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 288, + 545, + 371 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 288, + 545, + 371 + ], + "spans": [ + { + "bbox": [ + 305, + 288, + 545, + 371 + ], + "type": "text", + "content": "Figure 7 shows the architecture of the Motion Attention Module (MAM). This module focuses on direct mirror prediction from the optical flow fields. It leverages self-attention to capture and refine better features, to model motion inconsistency cue and to inform mirror locations. The output is then fed to the Fusion Refinement Module for the final mirror refinement." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 373, + 545, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 373, + 545, + 468 + ], + "spans": [ + { + "bbox": [ + 305, + 373, + 545, + 468 + ], + "type": "text", + "content": "Specifically, we initiate the MAM by embedding patches with positional information to create " + }, + { + "bbox": [ + 305, + 373, + 545, + 468 + ], + "type": "inline_equation", + "content": "flo_{\\text{embed}}" + }, + { + "bbox": [ + 305, + 373, + 545, + 468 + ], + "type": "text", + "content": ". Subsequently, we guide " + }, + { + "bbox": [ + 305, + 373, + 545, + 468 + ], + "type": "inline_equation", + "content": "flo_{\\text{embed}}" + }, + { + "bbox": [ + 305, + 373, + 545, + 468 + ], + "type": "text", + "content": " through six transformer blocks. Each block incorporates self-attention, treating " + }, + { + "bbox": [ + 305, + 373, + 545, + 468 + ], + "type": "inline_equation", + "content": "flo_{\\text{embed}}" + }, + { + "bbox": [ + 305, + 373, + 545, + 468 + ], + "type": "text", + "content": " as Q, K, and V in the self-attention mechanism, with output denoted as " + }, + { + "bbox": [ + 305, + 373, + 545, + 468 + ], + "type": "inline_equation", + "content": "flo_{\\text{attend}}" + }, + { + "bbox": [ + 305, + 373, + 545, + 468 + ], + "type": "text", + "content": ". To enhance its representation, " + }, + { + "bbox": [ + 305, + 373, + 545, + 468 + ], + "type": "inline_equation", + "content": "flo_{\\text{attend}}" + }, + { + "bbox": [ + 305, + 373, + 545, + 468 + ], + "type": "text", + "content": " undergoes batch normalization, followed by processing through a feed-forward design." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 469, + 545, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 469, + 545, + 601 + ], + "spans": [ + { + "bbox": [ + 305, + 469, + 545, + 601 + ], + "type": "text", + "content": "The output of the self-attention mechanism within one Transformer BlockLayer N is then passed to the subsequent Transformer BlockLayer N+1. Following this sequence, the output from Transformer BlockLayer 6 is channeled through a convolutional layer. The intentional use of six sequential transformer blocks serves a dual purpose: it equips the model to comprehend long-range dependencies, crucial for temporal relationships, and enables the capture of intricate spatial dependencies. This enhanced capability is particularly valuable for predicting spatial locations of the attended mirrors, denoted as " + }, + { + "bbox": [ + 305, + 469, + 545, + 601 + ], + "type": "inline_equation", + "content": "\\mathrm{Mp}_{\\mathrm{N - 1}}" + }, + { + "bbox": [ + 305, + 469, + 545, + 601 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 469, + 545, + 601 + ], + "type": "inline_equation", + "content": "\\mathrm{Mp_N}" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 609, + 509, + 623 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 609, + 509, + 623 + ], + "spans": [ + { + "bbox": [ + 306, + 609, + 509, + 623 + ], + "type": "text", + "content": "4.3. Motion-guided Edge Detection Module" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 305, + 628, + 545, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 628, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 305, + 628, + 545, + 689 + ], + "type": "text", + "content": "Figure 8 shows the architecture of the Motion-guided Edge Detection Module (MEDM). Unlike previous mirror edge detection in [11, 22] that only exploits image features, we also leverage the inconsistent motion cues captured in the optical flow field to guide the mirror edge detection process." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "content": "MEDM processes three inputs: " + }, + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{I}_{\\mathrm{low}}" + }, + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "content": " (a 2nd-level multiscale image feature), " + }, + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{I}_{\\mathrm{high}}" + }, + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "content": " (a 5th-layer multi-scale image" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17248" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "text", + "content": "feature), and " + }, + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "inline_equation", + "content": "flo" + }, + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "text", + "content": " (the optical flow field for motion guidance). The process begins with convolutional layers that extract and predict edge features—referred to as " + }, + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "inline_equation", + "content": "E_N" + }, + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "inline_equation", + "content": "E_{N-1}" + }, + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "text", + "content": "—from " + }, + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "inline_equation", + "content": "\\mathrm{I}_{\\mathrm{low}}" + }, + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "inline_equation", + "content": "\\mathrm{I}_{\\mathrm{high}}" + }, + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "inline_equation", + "content": "flo" + }, + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 121, + 289, + 264 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 121, + 289, + 264 + ], + "spans": [ + { + "bbox": [ + 46, + 121, + 289, + 264 + ], + "type": "text", + "content": "The cross-attention mechanism involves matrix multiplication between the transformed " + }, + { + "bbox": [ + 46, + 121, + 289, + 264 + ], + "type": "inline_equation", + "content": "flo" + }, + { + "bbox": [ + 46, + 121, + 289, + 264 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 121, + 289, + 264 + ], + "type": "inline_equation", + "content": "E_N" + }, + { + "bbox": [ + 46, + 121, + 289, + 264 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 46, + 121, + 289, + 264 + ], + "type": "inline_equation", + "content": "E_{N - 1}" + }, + { + "bbox": [ + 46, + 121, + 289, + 264 + ], + "type": "text", + "content": ". A Softmax activation determines the cross-attended beta value between " + }, + { + "bbox": [ + 46, + 121, + 289, + 264 + ], + "type": "inline_equation", + "content": "flo" + }, + { + "bbox": [ + 46, + 121, + 289, + 264 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 121, + 289, + 264 + ], + "type": "inline_equation", + "content": "E_N" + }, + { + "bbox": [ + 46, + 121, + 289, + 264 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 46, + 121, + 289, + 264 + ], + "type": "inline_equation", + "content": "E_{N - 1}" + }, + { + "bbox": [ + 46, + 121, + 289, + 264 + ], + "type": "text", + "content": ". This beta value, representing the attention weight, is then used to focus on " + }, + { + "bbox": [ + 46, + 121, + 289, + 264 + ], + "type": "inline_equation", + "content": "flo" + }, + { + "bbox": [ + 46, + 121, + 289, + 264 + ], + "type": "text", + "content": ", determining the contribution of each element in the optical flow field to the final cross-attended motion-guided edge feature. We specifically employ a cross-attention mechanism in this step to guide the extracted edge feature with the optical flow field. The motion-guided edge feature undergoes a convolutional pass, producing the output: the motion-guided edge feature map, " + }, + { + "bbox": [ + 46, + 121, + 289, + 264 + ], + "type": "inline_equation", + "content": "E_N" + }, + { + "bbox": [ + 46, + 121, + 289, + 264 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 46, + 121, + 289, + 264 + ], + "type": "inline_equation", + "content": "E_{N - 1}" + }, + { + "bbox": [ + 46, + 121, + 289, + 264 + ], + "type": "text", + "content": ", respectively." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 270, + 197, + 281 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 270, + 197, + 281 + ], + "spans": [ + { + "bbox": [ + 47, + 270, + 197, + 281 + ], + "type": "text", + "content": "4.4. Fusion Refinement Module" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 288, + 287, + 348 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 288, + 287, + 348 + ], + "spans": [ + { + "bbox": [ + 46, + 288, + 287, + 348 + ], + "type": "text", + "content": "We refine the motion-guided edge detection features from MEDM and the motion-based mirror predictions from MAM. This refining step is crucial to address limitations in each module (MEDM, MAM) and ensure accurate mirror prediction with both spatial and temporal context." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 348, + 287, + 444 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 348, + 287, + 444 + ], + "spans": [ + { + "bbox": [ + 46, + 348, + 287, + 444 + ], + "type": "text", + "content": "The process involves fusion-refinement, where model information is combined, refined, and fed into a convolutional binary classification head. This produces per-pixel binary classification maps, named " + }, + { + "bbox": [ + 46, + 348, + 287, + 444 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_{\\mathrm{N - 1}}" + }, + { + "bbox": [ + 46, + 348, + 287, + 444 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 348, + 287, + 444 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_{\\mathrm{N}}" + }, + { + "bbox": [ + 46, + 348, + 287, + 444 + ], + "type": "text", + "content": ", using two instances of the Fusion Refinement Module. Our approach is inspired by PMD-Net's [11] refinement strategy, adapted here to improve multiple modalities and guide mirror prediction for " + }, + { + "bbox": [ + 46, + 348, + 287, + 444 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_{\\mathrm{N - 1}}" + }, + { + "bbox": [ + 46, + 348, + 287, + 444 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 348, + 287, + 444 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_{\\mathrm{N}}" + }, + { + "bbox": [ + 46, + 348, + 287, + 444 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 451, + 136, + 462 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 451, + 136, + 462 + ], + "spans": [ + { + "bbox": [ + 47, + 451, + 136, + 462 + ], + "type": "text", + "content": "4.5. Loss Function" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 469, + 258, + 481 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 469, + 258, + 481 + ], + "spans": [ + { + "bbox": [ + 47, + 469, + 258, + 481 + ], + "type": "text", + "content": "To train the model, we use the loss function " + }, + { + "bbox": [ + 47, + 469, + 258, + 481 + ], + "type": "inline_equation", + "content": "(\\mathcal{L}_{\\mathrm{Final}})" + }, + { + "bbox": [ + 47, + 469, + 258, + 481 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 77, + 489, + 281, + 502 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 489, + 281, + 502 + ], + "spans": [ + { + "bbox": [ + 77, + 489, + 281, + 502 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {A} = \\left(\\mathcal {L} _ {\\text {M i r r o r A}} \\cdot \\alpha\\right) + \\left(\\mathcal {L} _ {\\text {M o t i o n A}} \\cdot \\zeta\\right) + \\left(\\mathcal {L} _ {\\text {E d g e A}} \\cdot \\gamma\\right)", + "image_path": "5a76f3ea27970aa9b4099b3d70523a6e30065e838d93a9fb7dbf08891f8c3fa8.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 77, + 504, + 280, + 516 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 504, + 280, + 516 + ], + "spans": [ + { + "bbox": [ + 77, + 504, + 280, + 516 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {B} = \\left(\\mathcal {L} _ {\\text {M i r r o r B}} \\cdot \\alpha\\right) + \\left(\\mathcal {L} _ {\\text {M o t i o n B}} \\cdot \\zeta\\right) + \\left(\\mathcal {L} _ {\\text {E d g e B}} \\cdot \\gamma\\right)", + "image_path": "ee4ab400dbed543c392be34e93e59e74e5a3fd57854a87c2a2ac141ea078e4cc.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 520, + 252, + 547 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 520, + 252, + 547 + ], + "spans": [ + { + "bbox": [ + 53, + 520, + 252, + 547 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} _ {\\text {C o h e r e n c e}} = (\\text {M i r r o r} _ {\\text {C o h e r e n c e}} \\\\ + \\operatorname {E d g e} _ {\\text {C o h e r e n c e}} + \\operatorname {M o t i o n} _ {\\text {C o h e r e n c e}}) \\cdot \\delta \\\\ \\end{array}", + "image_path": "00af1c6cb98de7251ff65cbc12a73ccfdc93fe65967944d06025acbdcc907598.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 69, + 551, + 193, + 562 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 551, + 193, + 562 + ], + "spans": [ + { + "bbox": [ + 69, + 551, + 193, + 562 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {F i n a l}} = \\mathcal {L} _ {\\text {C o h e r e n c e}} + \\mathcal {L} _ {A} + \\mathcal {L} _ {B}", + "image_path": "fba1107c358b4ab88a240a1158dab34596e4ff77528acda2f5b798f2f41901a4.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 59, + 570, + 215, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 570, + 215, + 582 + ], + "spans": [ + { + "bbox": [ + 59, + 570, + 215, + 582 + ], + "type": "text", + "content": "We discuss the loss component below:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 594, + 287, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 287, + 712 + ], + "type": "text", + "content": "Temporal Loss " + }, + { + "bbox": [ + 46, + 594, + 287, + 712 + ], + "type": "inline_equation", + "content": "(\\mathcal{L}_{\\mathrm{Coherence}})" + }, + { + "bbox": [ + 46, + 594, + 287, + 712 + ], + "type": "text", + "content": ": To make predictions across consecutive frames more consistent over time, we use BCELossWithLogits (binary cross-entropy) between " + }, + { + "bbox": [ + 46, + 594, + 287, + 712 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_{\\mathrm{N - 1}}" + }, + { + "bbox": [ + 46, + 594, + 287, + 712 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 594, + 287, + 712 + ], + "type": "inline_equation", + "content": "\\mathbf{M_N}" + }, + { + "bbox": [ + 46, + 594, + 287, + 712 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 594, + 287, + 712 + ], + "type": "inline_equation", + "content": "\\mathbf{M}\\mathfrak{p}_{\\mathrm{N - 1}}" + }, + { + "bbox": [ + 46, + 594, + 287, + 712 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 594, + 287, + 712 + ], + "type": "inline_equation", + "content": "\\mathbf{M}\\mathfrak{p}_{\\mathrm{N}}" + }, + { + "bbox": [ + 46, + 594, + 287, + 712 + ], + "type": "text", + "content": ", as well as " + }, + { + "bbox": [ + 46, + 594, + 287, + 712 + ], + "type": "inline_equation", + "content": "\\mathbf{E}_{\\mathrm{N - 1}}" + }, + { + "bbox": [ + 46, + 594, + 287, + 712 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 594, + 287, + 712 + ], + "type": "inline_equation", + "content": "\\mathbf{E}_{\\mathrm{N}}" + }, + { + "bbox": [ + 46, + 594, + 287, + 712 + ], + "type": "text", + "content": ". We set " + }, + { + "bbox": [ + 46, + 594, + 287, + 712 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 46, + 594, + 287, + 712 + ], + "type": "text", + "content": " to 1, calling this loss " + }, + { + "bbox": [ + 46, + 594, + 287, + 712 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{Coherence}}" + }, + { + "bbox": [ + 46, + 594, + 287, + 712 + ], + "type": "text", + "content": ". This loss is aimed at reducing temporal inaccuracies in edge prediction, motion mirror location, and mirror prediction between frames in a video. It ensures that predictions from our model's combined features remain consistent over time, addressing a current limitation in video mirror detection." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "text", + "content": "Mirror Binary Losses " + }, + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "inline_equation", + "content": "(\\mathcal{L}_{\\mathrm{MirrorA}}" + }, + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{MirrorB}})" + }, + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "text", + "content": ": We use two BCELossWithLogits loss functions to predict mirrored regions in " + }, + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "inline_equation", + "content": "\\mathrm{M_{N - 1}}" + }, + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "inline_equation", + "content": "(\\mathcal{L}_{\\mathrm{MirrorA}})" + }, + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "inline_equation", + "content": "\\mathrm{M_N}" + }, + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "inline_equation", + "content": "(\\mathcal{L}_{\\mathrm{MirrorB}})" + }, + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "text", + "content": ". These losses are then compared with the actual ground truth mirror maps. We set " + }, + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "text", + "content": " to 3 for these losses, emphasizing the importance of accurately detecting mirrors in the model." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 144, + 545, + 215 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 144, + 545, + 215 + ], + "spans": [ + { + "bbox": [ + 304, + 144, + 545, + 215 + ], + "type": "text", + "content": "Motion Mirror Losses " + }, + { + "bbox": [ + 304, + 144, + 545, + 215 + ], + "type": "inline_equation", + "content": "(\\mathcal{L}_{\\mathrm{MotionA}}" + }, + { + "bbox": [ + 304, + 144, + 545, + 215 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 144, + 545, + 215 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{MotionB}})" + }, + { + "bbox": [ + 304, + 144, + 545, + 215 + ], + "type": "text", + "content": ": Two additional BCELossWithLogits loss functions are used to predict mirror regions from motion-guided image features " + }, + { + "bbox": [ + 304, + 144, + 545, + 215 + ], + "type": "inline_equation", + "content": "\\mathrm{Mp_{N - 1}}" + }, + { + "bbox": [ + 304, + 144, + 545, + 215 + ], + "type": "inline_equation", + "content": "(\\mathcal{L}_{\\mathrm{MotionA}})" + }, + { + "bbox": [ + 304, + 144, + 545, + 215 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 144, + 545, + 215 + ], + "type": "inline_equation", + "content": "\\mathrm{Mp_N}" + }, + { + "bbox": [ + 304, + 144, + 545, + 215 + ], + "type": "inline_equation", + "content": "(\\mathcal{L}_{\\mathrm{MotionB}})" + }, + { + "bbox": [ + 304, + 144, + 545, + 215 + ], + "type": "text", + "content": ". These losses are then compared with the ground-truth mirror maps. We set " + }, + { + "bbox": [ + 304, + 144, + 545, + 215 + ], + "type": "inline_equation", + "content": "\\zeta" + }, + { + "bbox": [ + 304, + 144, + 545, + 215 + ], + "type": "text", + "content": " to 1 for both of these losses." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 216, + 545, + 288 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 216, + 545, + 288 + ], + "spans": [ + { + "bbox": [ + 304, + 216, + 545, + 288 + ], + "type": "text", + "content": "Mirror Edge Losses " + }, + { + "bbox": [ + 304, + 216, + 545, + 288 + ], + "type": "inline_equation", + "content": "(\\mathcal{L}_{\\mathrm{EdgeA}}" + }, + { + "bbox": [ + 304, + 216, + 545, + 288 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 216, + 545, + 288 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{EdgeB}})" + }, + { + "bbox": [ + 304, + 216, + 545, + 288 + ], + "type": "text", + "content": ": Two BCELossWithLogits loss functions are utilized for predicting mirror boundaries in " + }, + { + "bbox": [ + 304, + 216, + 545, + 288 + ], + "type": "inline_equation", + "content": "\\mathrm{E_{N - 1}}" + }, + { + "bbox": [ + 304, + 216, + 545, + 288 + ], + "type": "inline_equation", + "content": "(\\mathcal{L}_{\\mathrm{EdgeA}})" + }, + { + "bbox": [ + 304, + 216, + 545, + 288 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 216, + 545, + 288 + ], + "type": "inline_equation", + "content": "\\mathrm{E_N}" + }, + { + "bbox": [ + 304, + 216, + 545, + 288 + ], + "type": "inline_equation", + "content": "(\\mathcal{L}_{\\mathrm{EdgeB}})" + }, + { + "bbox": [ + 304, + 216, + 545, + 288 + ], + "type": "text", + "content": ". These losses are compared against the ground-truth mirror edge maps. We set " + }, + { + "bbox": [ + 304, + 216, + 545, + 288 + ], + "type": "inline_equation", + "content": "\\gamma = 2" + }, + { + "bbox": [ + 304, + 216, + 545, + 288 + ], + "type": "text", + "content": " for each of these losses, emphasizing the precision of mirror edge prediction." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 297, + 386, + 311 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 297, + 386, + 311 + ], + "spans": [ + { + "bbox": [ + 306, + 297, + 386, + 311 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 317, + 419, + 330 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 317, + 419, + 330 + ], + "spans": [ + { + "bbox": [ + 306, + 317, + 419, + 330 + ], + "type": "text", + "content": "5.1. Setups and Metrics" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 335, + 545, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 335, + 545, + 430 + ], + "spans": [ + { + "bbox": [ + 304, + 335, + 545, + 430 + ], + "type": "text", + "content": "We implemented MG-VMD using PyTorch, trained and validated it against comparative models on a NVIDIA RTX 3090 GPU. In the data preprocessing phase, all images, ground-truth mirror maps, and edge maps were resized to a resolution of " + }, + { + "bbox": [ + 304, + 335, + 545, + 430 + ], + "type": "inline_equation", + "content": "224 \\times 224" + }, + { + "bbox": [ + 304, + 335, + 545, + 430 + ], + "type": "text", + "content": ". Our model's ResNext-101 backbone and DeepLabV3 parameters were initialized using the pretrained ResNext-101 backbone and pretrained DeepLabV3 weights from VMD-Net [12]." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 431, + 545, + 503 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 431, + 545, + 503 + ], + "spans": [ + { + "bbox": [ + 304, + 431, + 545, + 503 + ], + "type": "text", + "content": "We utilized a Stochastic Gradient Descent Optimizer with a starting learning rate of 9e-3, a momentum of 1.9, and a weight decay of 5e-4. Then, we applied an adaptive learning rate to the optimizer, interpolating the learning rate from 9e-3 to 3e-3 across epochs 1 to 15. The model underwent 15 epochs of training with a batch size of 8." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 503, + 545, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 503, + 545, + 552 + ], + "spans": [ + { + "bbox": [ + 304, + 503, + 545, + 552 + ], + "type": "text", + "content": "For validation, we employed four quantitative metrics: Mean Absolute Error (MAE↓), F-measure " + }, + { + "bbox": [ + 304, + 503, + 545, + 552 + ], + "type": "inline_equation", + "content": "(F_{\\beta}\\uparrow)" + }, + { + "bbox": [ + 304, + 503, + 545, + 552 + ], + "type": "text", + "content": ", Accuracy↑, and Intersection over Union (IoU↑). F-measure " + }, + { + "bbox": [ + 304, + 503, + 545, + 552 + ], + "type": "inline_equation", + "content": "(F_{\\beta}\\uparrow)" + }, + { + "bbox": [ + 304, + 503, + 545, + 552 + ], + "type": "text", + "content": " is computed as follows:" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 353, + 559, + 545, + 585 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 353, + 559, + 545, + 585 + ], + "spans": [ + { + "bbox": [ + 353, + 559, + 545, + 585 + ], + "type": "interline_equation", + "content": "F _ {\\beta} = \\frac {1 + \\beta^ {2} (p r e c i s i o n * r e c a l l)}{\\beta^ {2} + r e c a l l}, \\tag {1}", + "image_path": "103934d6e86ba13633d4abc0705507e042bdb6969a65eb2394e6e2d5a670cdb9.jpg" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 304, + 592, + 545, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 592, + 545, + 628 + ], + "spans": [ + { + "bbox": [ + 304, + 592, + 545, + 628 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 592, + 545, + 628 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 304, + 592, + 545, + 628 + ], + "type": "text", + "content": " is set to 0.3, as suggested by previous research [11] [1]. The F-measure " + }, + { + "bbox": [ + 304, + 592, + 545, + 628 + ], + "type": "inline_equation", + "content": "(\\mathrm{F}_{\\beta}\\uparrow)" + }, + { + "bbox": [ + 304, + 592, + 545, + 628 + ], + "type": "text", + "content": " assesses the harmonic mean between precision and recall." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 306, + 635, + 429, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 635, + 429, + 647 + ], + "spans": [ + { + "bbox": [ + 306, + 635, + 429, + 647 + ], + "type": "text", + "content": "5.2. Experimental Results" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "type": "text", + "content": "In Table 3 and Figure 9, we evaluate our proposed technique against 8 state-of-the-art methods in Video Salient Object Detection, and Image-Based/Video Mirror Detection. Using their respective pre-trained weights, we fine-tune and validate these results on our proposed MMD dataset.." + } + ] + } + ], + "index": 24 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17249" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 47, + 69, + 561, + 449 + ], + "blocks": [ + { + "bbox": [ + 47, + 69, + 561, + 449 + ], + "lines": [ + { + "bbox": [ + 47, + 69, + 561, + 449 + ], + "spans": [ + { + "bbox": [ + 47, + 69, + 561, + 449 + ], + "type": "image", + "image_path": "69c2078ac8b2b3b5ea2c04258db100265172934d33da06687524b04d03907fe2.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 452, + 547, + 485 + ], + "lines": [ + { + "bbox": [ + 46, + 452, + 547, + 485 + ], + "spans": [ + { + "bbox": [ + 46, + 452, + 547, + 485 + ], + "type": "text", + "content": "Figure 9. Qualitative results table comparing our proposed method against state-of-the-art Video Salient Object Detection, and Image-Based/Video Mirror Detection FS-Net [8], MGA [10], ALGRF [23], F3Net [9], UFO[17], MirrorNet[30], PMD-Net [11], VMD-Net[12] trained and validated on our proposed MMD dataset." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 496, + 242, + 510 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 496, + 242, + 510 + ], + "spans": [ + { + "bbox": [ + 47, + 496, + 242, + 510 + ], + "type": "text", + "content": "5.3. Comparison Against State-of-the-art" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 518, + 287, + 674 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 518, + 287, + 674 + ], + "spans": [ + { + "bbox": [ + 46, + 518, + 287, + 674 + ], + "type": "text", + "content": "Our MG-VMD achieves better performance in Mean Absolute Error (MAE↓), F-measure (" + }, + { + "bbox": [ + 46, + 518, + 287, + 674 + ], + "type": "inline_equation", + "content": "F_{\\beta} \\uparrow" + }, + { + "bbox": [ + 46, + 518, + 287, + 674 + ], + "type": "text", + "content": "), Accuracy↑, and Intersection over Union (IoU↑) when compared to state-of-the-art Video Salient Object Detection, Image-Based Mirror Detection and Video Mirror Detection methods fine-tuned and validated on our proposed Motion Mirror Dataset. We attribute the better performance to the exploiting of motion inconsistency cues across adjacent frames through our novel modules: Motion-Guided Edge Detection Module (MEDM) and Motion Attention Module (MAM). Figure 9 shows that our method is able to detect mirrors well in complex (e.g., low contrast) scenarios, outperforming state-of-the-arts." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 677, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 288, + 713 + ], + "type": "text", + "content": "We also evaluate the generalizability of our model. In the Supplemental, we show results comparing MG-VMD and VMD-Net trained on our MMD dataset and tested on" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 305, + 497, + 545, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 497, + 545, + 533 + ], + "spans": [ + { + "bbox": [ + 305, + 497, + 545, + 533 + ], + "type": "text", + "content": "VMD-D [12]. MG-VMD outperforms VMD-Net, showing that our model has slightly better/close generalization performances." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 540, + 370, + 551 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 540, + 370, + 551 + ], + "spans": [ + { + "bbox": [ + 306, + 540, + 370, + 551 + ], + "type": "text", + "content": "5.4. Ablation" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 557, + 545, + 701 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 557, + 545, + 701 + ], + "spans": [ + { + "bbox": [ + 304, + 557, + 545, + 701 + ], + "type": "text", + "content": "We conducted an ablation study to assess the effectiveness of the proposed modules in MG-VMD. Initially, we removed both the Motion Attention Module and the Motion-guided Edge Detection Module, resulting in the baseline with only the Refinement module, which incorporates convolutional layers. This configuration is referred to as the \"baseline\" for comparative analysis. The rationale behind evaluating the baseline's performance without the Optical Flow Coherence Block, Motion-Guided Edge Detection Module, and Motion Attention Module was to quantitatively gauge the efficacy of our deep-learning approach without leveraging inconsistent motion cues." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 318, + 701, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 701, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 318, + 701, + 545, + 713 + ], + "type": "text", + "content": "Table 4 shows the performance of the ablated models:" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17250" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 50, + 70, + 286, + 185 + ], + "blocks": [ + { + "bbox": [ + 50, + 70, + 286, + 185 + ], + "lines": [ + { + "bbox": [ + 50, + 70, + 286, + 185 + ], + "spans": [ + { + "bbox": [ + 50, + 70, + 286, + 185 + ], + "type": "table", + "html": "
ModelsAccuracy↑MAE↓Fβ↑IoU↑
MGA0.5660370.4339630.4891760.281285
UFO0.7967970.2032040.7974470.598641
FSNet0.8532870.1467130.8377320.707848
ALGRF0.7237720.2762290.6928290.502814
F3Net0.8388630.1611370.8473210.670618
PMD-Net0.7409090.2590910.8472110.424085
MirrorNet0.8352270.1647720.8389870.666242
VMDNet0.8544150.1455850.8123470.722634
Ours0.8725320.1274680.8694190.725130
", + "image_path": "598b38fe10e1eafb9eda1568ce42afdb47fe6255f138d1c07f67862787156b75.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 50, + 277, + 285, + 335 + ], + "blocks": [ + { + "bbox": [ + 46, + 188, + 289, + 266 + ], + "lines": [ + { + "bbox": [ + 46, + 188, + 289, + 266 + ], + "spans": [ + { + "bbox": [ + 46, + 188, + 289, + 266 + ], + "type": "text", + "content": "Table 3. Quantitative results table comparing our proposed method against state-of-the-art Video Salient Object Detection, and Image-Based/Video Mirror Detection FS-Net [8], MGA [10], AL-GRF [23], F3Net [9], UFO [17], PMD-Net [11], MirrorNet[30], VMD-Net[12] methods trained and validated on our proposed MMD dataset. Red, Blue and Green indicate the best, second and third best performances, respectively." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 50, + 277, + 285, + 335 + ], + "lines": [ + { + "bbox": [ + 50, + 277, + 285, + 335 + ], + "spans": [ + { + "bbox": [ + 50, + 277, + 285, + 335 + ], + "type": "table", + "html": "
ModelsAccuracy↑MAE↓Fβ↑IoU↑
baseline0.8575190.1424810.8375780.700855
b + MEDM0.8581760.1418240.8478530.707350
b + MAM0.8662180.1337820.8416140.713005
Ours0.8725320.1274680.8694190.725130
", + "image_path": "18322604f24696b3b464cd9d1bb229b3f6e63b0ef66dfc8a1ef1b4c375ecbf88.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 338, + 288, + 395 + ], + "lines": [ + { + "bbox": [ + 46, + 338, + 288, + 395 + ], + "spans": [ + { + "bbox": [ + 46, + 338, + 288, + 395 + ], + "type": "text", + "content": "Table 4. Ablation Results: Qualitative results table comparing our proposed method (MG-VMD (Ours)) against ablated versions of our model: b + MAM, b + MEDM, and baseline (b), trained and validated on our proposed MMD dataset. Red, Blue and Green indicate the best, second and third-best performances, respectively." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 403, + 287, + 464 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 403, + 287, + 464 + ], + "spans": [ + { + "bbox": [ + 46, + 403, + 287, + 464 + ], + "type": "text", + "content": "\"baseline (b)\", \"b + MEDM\", \"b + MAM\", and \"MG-VMD (Ours)\". Both MAM and MEDM exploit motion inconsistency cues, demonstrating better performance compared to the baseline. The full model combines the complementary outputs from MAM and MEDM, yielding the best results." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 464, + 287, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 464, + 287, + 548 + ], + "spans": [ + { + "bbox": [ + 46, + 464, + 287, + 548 + ], + "type": "text", + "content": "Our baseline outperforms state-of-the-art VMD-Net [12] for two reasons. First, our model is fine-tuned on VMD-Net's [12] ResNext-101 backbone and DeepLabV3 weights (Section 5). Second, VMD-Net [12] assumes dual correspondence, but in real-life datasets, dual correspondence does not always occur. This shows the usefulness of our MMD dataset." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 548, + 288, + 657 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 548, + 288, + 657 + ], + "spans": [ + { + "bbox": [ + 46, + 548, + 288, + 657 + ], + "type": "text", + "content": "Table 4 shows that MEDM and MAM each contributes to the model individually. Without MEDM, mirror edges become noisy (e.g., Figure 10 Column 4 Rows 2 and 4). Without MAM, MEDM may mistakenly detect object edges inside mirrors (e.g., cabinet edges in Figure 10 Column 3 Row 2), leading to inaccurate mirror detection. (Best viewed by zooming in.) Both modules use motion features, complementing each other and boost performance (Figure 10 Column 5, and Table 4)." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 669, + 119, + 681 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 669, + 119, + 681 + ], + "spans": [ + { + "bbox": [ + 47, + 669, + 119, + 681 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 689, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 689, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 689, + 288, + 713 + ], + "type": "text", + "content": "This paper introduces a novel deep-learning approach to tackle video mirror detection by leveraging inconsistent" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 307, + 70, + 547, + 240 + ], + "blocks": [ + { + "bbox": [ + 307, + 70, + 547, + 240 + ], + "lines": [ + { + "bbox": [ + 307, + 70, + 547, + 240 + ], + "spans": [ + { + "bbox": [ + 307, + 70, + 547, + 240 + ], + "type": "image", + "image_path": "15d4f9447aaa4fcb42a89d36b5392d61dfa4b40d163684e32178f2a405cc3c06.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 331, + 240, + 520, + 251 + ], + "lines": [ + { + "bbox": [ + 331, + 240, + 520, + 251 + ], + "spans": [ + { + "bbox": [ + 331, + 240, + 520, + 251 + ], + "type": "text", + "content": "Figure 10. Qualitative results of the ablated models." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 307, + 268, + 545, + 359 + ], + "blocks": [ + { + "bbox": [ + 307, + 268, + 545, + 359 + ], + "lines": [ + { + "bbox": [ + 307, + 268, + 545, + 359 + ], + "spans": [ + { + "bbox": [ + 307, + 268, + 545, + 359 + ], + "type": "image", + "image_path": "0004acba8a3a80f61163a5505f1f5de033e94223b9cfba8b3d73e5eea875dea7.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 361, + 545, + 395 + ], + "lines": [ + { + "bbox": [ + 305, + 361, + 545, + 395 + ], + "spans": [ + { + "bbox": [ + 305, + 361, + 545, + 395 + ], + "type": "text", + "content": "Figure 11. Failure cases. Our Method may not be good in scenarios containing both mirror regions and windows, due to windows irregular motion compared to a scene." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 403, + 545, + 511 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 403, + 545, + 511 + ], + "spans": [ + { + "bbox": [ + 304, + 403, + 545, + 511 + ], + "type": "text", + "content": "motion cues within and outside mirrored regions. We also present a benchmark video mirror dataset featuring consistent inter-frame motion across long videos. Our experimental results demonstrate that our proposed MG-VMD deep-learning method outperforms state-of-the-arts in Video Mirror Detection, Image-Based Mirror Detection, and Video Salient Object Detection. Furthermore, we conclude that our proposed Motion Mirror Dataset (MMD) is challenging compared to state-of-the-art Video Mirror Datasets." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 517, + 547, + 589 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 517, + 547, + 589 + ], + "spans": [ + { + "bbox": [ + 304, + 517, + 547, + 589 + ], + "type": "text", + "content": "Our method is not without limitation. Figure 11 shows a scenario with both mirror and window in a scene. We found that similar to mirror regions, windows can have different motions from their surroundings. Thus, our method tends to over-predict window regions as mirrors. We plan to improve our design to mitigate this in the future." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "type": "text", + "content": "Acknowledgements: Alex is supported by a Swansea GTA Research Scholarship. This project is in part supported by a GRF grant from the Research Grants Council of Hong Kong (Ref.: 11211223). We gratefully acknowledge the support of the HEFCW HERC fund (W21/21HE) for the provision of GPU equipment used in this research. For the purpose of Open Access, the author has applied a Creative Commons Attribution (CC BY) license to any Author Accepted Manuscript (AAM) version arising from this submission." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17251" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 286, + 713 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 53, + 91, + 286, + 123 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 286, + 123 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 286, + 123 + ], + "type": "text", + "content": "[1] Radhakrishna Achanta, Sheila Hemami, Francisco Estrada, and Sabine Susstrunk. Frequency-tuned salient region detection. In CVPR, 2009. 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 125, + 286, + 158 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 125, + 286, + 158 + ], + "spans": [ + { + "bbox": [ + 53, + 125, + 286, + 158 + ], + "type": "text", + "content": "[2] Liang-Chieh Chen, George Papandreou, Florian Schroff, and Hartwig Adam. Rethinking atrous convolution for semantic image segmentation, 2017. 4, 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 159, + 286, + 191 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 159, + 286, + 191 + ], + "spans": [ + { + "bbox": [ + 54, + 159, + 286, + 191 + ], + "type": "text", + "content": "[3] Zheng Dong, Ke Xu, Ziheng Duan, Hujun Bao, Weiwei Xu, and Rynson W. H. Lau. Geometry-aware two-scale pifu representation for human reconstruction. In NeurIPS, 2022. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 194, + 286, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 194, + 286, + 237 + ], + "spans": [ + { + "bbox": [ + 53, + 194, + 286, + 237 + ], + "type": "text", + "content": "[4] Zheng Dong, Ke Xu, Yaoan Gao, Qilin Sun, Hujun Bao, Weiwei Xu, and Rynson W. H. Lau. Sailor: Synergizing radiance and occupancy fields for live human performance capture. ACM Trans. Graph., 2023. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 239, + 286, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 239, + 286, + 270 + ], + "spans": [ + { + "bbox": [ + 53, + 239, + 286, + 270 + ], + "type": "text", + "content": "[5] Gunnar Farnebäck. Two-frame motion estimation based on polynomial expansion. In Proceedings of the 13th Scandinavian Conference on Image Analysis, 2003. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 273, + 286, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 273, + 286, + 304 + ], + "spans": [ + { + "bbox": [ + 53, + 273, + 286, + 304 + ], + "type": "text", + "content": "[6] Yuan-Chen Guo, Di Kang, Linchao Bao, Yu He, and Song-Hai Zhang. Nerfren: Neural radiance fields with reflections. In CVPR, 2022. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 306, + 286, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 306, + 286, + 348 + ], + "spans": [ + { + "bbox": [ + 53, + 306, + 286, + 348 + ], + "type": "text", + "content": "[7] Eddy Ig, Nikolaus Mayer, Tonmoy Saikia, Margret Keuper, Alexey Dosovitskiy, and Thomas Brox. Flownet 2.0: Evolution of optical flow estimation with deep networks. In CVPR, 2017. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 351, + 286, + 384 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 351, + 286, + 384 + ], + "spans": [ + { + "bbox": [ + 53, + 351, + 286, + 384 + ], + "type": "text", + "content": "[8] Ge-Peng Ji, Keren Fu, Zhe Wu, Deng-Ping Fan, Jianbing Shen, and Ling Shao. Full-duplex strategy for video object segmentation. In ICCV, 2021. 2, 7, 8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 386, + 286, + 417 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 386, + 286, + 417 + ], + "spans": [ + { + "bbox": [ + 53, + 386, + 286, + 417 + ], + "type": "text", + "content": "[9] Qingming Huang Jun Wei, Shuhui Wang. F3net: Fusion, feedback and focus for salient object detection. In AAAI, 2020. 2, 7, 8" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 420, + 286, + 452 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 420, + 286, + 452 + ], + "spans": [ + { + "bbox": [ + 48, + 420, + 286, + 452 + ], + "type": "text", + "content": "[10] Haofeng Li, Guanqi Chen, Guanbin Li, and Yizhou Yu. Motion guided attention for video salient object detection. In ICCV, 2019. 5, 7, 8" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 453, + 286, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 453, + 286, + 475 + ], + "spans": [ + { + "bbox": [ + 48, + 453, + 286, + 475 + ], + "type": "text", + "content": "[11] Jiaying Lin, Guodong Wang, and Rynson W.H. Lau. Progressive mirror detection. In CVPR, 2020. 1, 2, 5, 6, 7, 8" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 476, + 286, + 508 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 476, + 286, + 508 + ], + "spans": [ + { + "bbox": [ + 48, + 476, + 286, + 508 + ], + "type": "text", + "content": "[12] Jiaying Lin, Xin Tan, and Rynson W.H. Lau. Learning to detect mirrors from videos via dual correspondences. In CVPR, 2023. 1, 2, 3, 6, 7, 8" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 510, + 286, + 532 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 510, + 286, + 532 + ], + "spans": [ + { + "bbox": [ + 48, + 510, + 286, + 532 + ], + "type": "text", + "content": "[13] Zhiqiu Lin, Jin Sun, Abe Davis, and Noah Snavely. Visual chirality. In CVPR, 2020. 1" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 534, + 286, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 534, + 286, + 566 + ], + "spans": [ + { + "bbox": [ + 48, + 534, + 286, + 566 + ], + "type": "text", + "content": "[14] Jing Liu, Jiaxiang Wang, Weikang Wang, and Yuting Su. Ds-net: Dynamic spatiotemporal network for video salient object detection. arXiv:2012.04886, 2022. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 567, + 286, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 567, + 286, + 599 + ], + "spans": [ + { + "bbox": [ + 48, + 567, + 286, + 599 + ], + "type": "text", + "content": "[15] Haiyang Mei, Bo Dong, Wen Dong, Pieter Peers, Xin Yang, Qiang Zhang, and Xiaopeng Wei. Depth-aware mirror segmentation. In CVPR, 2021. 1, 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 48, + 601, + 286, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 601, + 286, + 634 + ], + "spans": [ + { + "bbox": [ + 48, + 601, + 286, + 634 + ], + "type": "text", + "content": "[16] Anwesan Pal, Sayan Mondal, and Henrik I Christensen. \"looking at the right stuff\"-guided semantic-gaze for autonomous driving. In CVPR, 2020. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 48, + 636, + 286, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 636, + 286, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 636, + 286, + 689 + ], + "type": "text", + "content": "[17] Yukun Su, Jingliang Deng, Ruizhou Sun, Guosheng Lin, and Qingyao Wu. A unified transformer framework for group-based segmentation: Co-segmentation, co-saliency detection and video salient object detection. IEEE Trans. Multimedia, 2023. 7, 8" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 48, + 691, + 286, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 286, + 713 + ], + "type": "text", + "content": "[18] Hideki Tamura, Maki Tsukuda, Hiroshi Higashi, and Shigeki Nakauchi. Perceptual segregation between mirror and glass" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 74, + 545, + 546 + ], + "type": "list", + "angle": 0, + "index": 34, + "blocks": [ + { + "bbox": [ + 327, + 74, + 545, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 74, + 545, + 94 + ], + "spans": [ + { + "bbox": [ + 327, + 74, + 545, + 94 + ], + "type": "text", + "content": "material under natural and unnatural illumination. Journal of Vision, 2016. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 96, + 545, + 129 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 96, + 545, + 129 + ], + "spans": [ + { + "bbox": [ + 308, + 96, + 545, + 129 + ], + "type": "text", + "content": "[19] Hideki Tamura, Hiroshi Higashi, and Shigeki Nakauchi. Multiple cues for visual perception of mirror and glass materials. Journal of Vision, 2017. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 130, + 545, + 163 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 130, + 545, + 163 + ], + "spans": [ + { + "bbox": [ + 308, + 130, + 545, + 163 + ], + "type": "text", + "content": "[20] Hideki Tamura, Hiroshi Higashi, and Shigeki Nakauchi. Dynamic visual cues for differentiating mirror and glass. Scientific Reports, 2018. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 164, + 545, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 164, + 545, + 206 + ], + "spans": [ + { + "bbox": [ + 308, + 164, + 545, + 206 + ], + "type": "text", + "content": "[21] Xin Tan, Ke Xu, Ying Cao, Yiheng Zhang, Lizhuang Ma, and Rynson W. H. Lau. Night-time scene parsing with a large real dataset. IEEE Transactions on Image Processing, 2021. 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 209, + 545, + 241 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 209, + 545, + 241 + ], + "spans": [ + { + "bbox": [ + 308, + 209, + 545, + 241 + ], + "type": "text", + "content": "[22] Xin Tan, Jiaying Lin, Ke Xu, Pan Chen, Lizhuang Ma, and Rynson W.H. Lau. Mirror detection with the visual chirality cue. IEEE Trans. Pattern Anal. Mach. Intell., 2022. 1, 2, 5" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 243, + 545, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 243, + 545, + 274 + ], + "spans": [ + { + "bbox": [ + 308, + 243, + 545, + 274 + ], + "type": "text", + "content": "[23] Yi Tang, Yuanman Li, and Guoliang Xing. Video salient object detection via adaptive local-global refinement. arXiv:2104.14360, 2021. 2, 7, 8" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 277, + 545, + 298 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 277, + 545, + 298 + ], + "spans": [ + { + "bbox": [ + 308, + 277, + 545, + 298 + ], + "type": "text", + "content": "[24] Zachary Teed and Jia Deng. Raft: Recurrent all-pairs field transforms for optical flow. In ECCV, 2020. 3, 5" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 300, + 545, + 332 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 300, + 545, + 332 + ], + "spans": [ + { + "bbox": [ + 308, + 300, + 545, + 332 + ], + "type": "text", + "content": "[25] Xin Tian, Ke Xu, Xin Yang, Lin Du, Baocai Yin, and Rynson WH Lau. Bi-directional object-context prioritization learning for saliency ranking. In CVPR, 2022. 1" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 334, + 545, + 366 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 334, + 545, + 366 + ], + "spans": [ + { + "bbox": [ + 308, + 334, + 545, + 366 + ], + "type": "text", + "content": "[26] Xin Tian, Ke Xu, Xin Yang, Baocai Yin, and Rynson Lau. Learning to detect instance-level salient objects using complementary image labels. Int. J. Comput. Vis., 2022. 1" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 368, + 545, + 399 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 368, + 545, + 399 + ], + "spans": [ + { + "bbox": [ + 308, + 368, + 545, + 399 + ], + "type": "text", + "content": "[27] Pavel Tokmakov, Karteek Alahari, and Cordelia Schmid. Learning video object segmentation with visual memory. arXiv:1704.05737, 2017. 2" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 308, + 401, + 545, + 434 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 401, + 545, + 434 + ], + "spans": [ + { + "bbox": [ + 308, + 401, + 545, + 434 + ], + "type": "text", + "content": "[28] Wenguan Wang, Xiankai Lu, Jianbing Shen, David Crandall, and Ling Shao. Zero-shot video object segmentation via attentive graph neural networks. In ICCV, 2019. 2" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 308, + 435, + 545, + 467 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 435, + 545, + 467 + ], + "spans": [ + { + "bbox": [ + 308, + 435, + 545, + 467 + ], + "type": "text", + "content": "[29] Saining Xie, Ross Girshick, Piotr Dólár, Zhuowen Tu, and Kaiming He. Aggregated residual transformations for deep neural networks. In CVPR, 2017. 4, 5" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 308, + 468, + 545, + 500 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 468, + 545, + 500 + ], + "spans": [ + { + "bbox": [ + 308, + 468, + 545, + 500 + ], + "type": "text", + "content": "[30] Xin Yang, Haiyang Mei, Ke Xu, Xiaopeng Wei, Baocai Yin, and Rynson W. H. Lau. Where is my mirror? In ICCV, 2019. 1, 2, 7, 8" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 308, + 502, + 545, + 546 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 502, + 545, + 546 + ], + "spans": [ + { + "bbox": [ + 308, + 502, + 545, + 546 + ], + "type": "text", + "content": "[31] Junyi Zeng, Chong Bao, Rui Chen, Zilong Dong, Guofeng Zhang, Hujun Bao, and Zhaopeng Cui. Mirror-nerf: Learning neural radiance fields for mirrors with whitted-style ray tracing. In ACM MM, 2023. 1" + } + ] + } + ], + "index": 33 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17252" + } + ] + } + ], + "index": 35 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Efficient 3D Implicit Head Avatar with Mesh-anchored Hash Table Blendshapes/afaaa8c5-698e-4fcb-9bde-21aa2469de9a_content_list.json b/2024/Efficient 3D Implicit Head Avatar with Mesh-anchored Hash Table Blendshapes/afaaa8c5-698e-4fcb-9bde-21aa2469de9a_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..35669e66de3c41cd718297feec8383482e2d092f --- /dev/null +++ b/2024/Efficient 3D Implicit Head Avatar with Mesh-anchored Hash Table Blendshapes/afaaa8c5-698e-4fcb-9bde-21aa2469de9a_content_list.json @@ -0,0 +1,1285 @@ +[ + { + "type": "text", + "text": "Efficient 3D Implicit Head Avatar with Mesh-anchored Hash Table Blendshapes", + "text_level": 1, + "bbox": [ + 83, + 130, + 887, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Ziqian Bai $^{1,2*}$ Feitong Tan $^{1}$ Sean Fanello $^{1}$ Rohit Pandey $^{1}$ Mingsong Dou $^{1}$ Shichen Liu $^{1}$ Ping Tan $^{3}$ Yinda Zhang $^{1}$", + "bbox": [ + 227, + 178, + 740, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Google", + "bbox": [ + 143, + 217, + 207, + 234 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ Simon Fraser University", + "bbox": [ + 238, + 217, + 415, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3 The Hong Kong University of Science and Technology", + "bbox": [ + 447, + 217, + 820, + 234 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/5711f29f7cf67a1111456bdda84716b1dcc8c7f8a5f2a528900c44276fb04533.jpg", + "image_caption": [ + "Figure 1. Visual Quality vs. Rendering Speed comparisons between our monocular 3D head avatars and prior state-of-the-art methods including NeRFBlendshape [14], INSTA [42], PointAvatar [41] and MonoAvatar [5]. Our approach achieves real-time rendering (i.e. $>30$ mean FPS) with a $512 \\times 512$ resolution, while produces comparable visual quality comparing to prior SOTAs [5], and gives significantly better results on challenging expressions than prior efficient avatars [14, 41, 42]. Webpage augmentedperception.github.io/monoavatar-plus" + ], + "image_footnote": [], + "bbox": [ + 78, + 271, + 274, + 419 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/e6c022d54e15cc4bdf00f6a0bc6e4efafe906c34aa5218377f999d8304c8f86d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 282, + 277, + 401, + 412 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/a377184159282ef49a05f86238bbba261f699c7ecce41884e41d486ee32b5422.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 405, + 277, + 522, + 412 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/fc3e527d29141d1109789cdb8834ff2002fe41ca2761b9caa4d9a5e003d0f9ca.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 524, + 277, + 645, + 412 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/35ac6f899365e031ab2e1b353b6adc15bc15f472d6724242609bb85f11ee6651.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 648, + 277, + 767, + 412 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/2f4e777735712c9a94daba4203b3452828327dd0ef6d2e3ea2d471af69b60c91.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 769, + 279, + 890, + 412 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 488, + 313, + 503 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3D head avatars built with neural implicit volumetric representations have achieved unprecedented levels of photorealism. However, the computational cost of these methods remains a significant barrier to their widespread adoption, particularly in real-time applications such as virtual reality and teleconferencing. While attempts have been made to develop fast neural rendering approaches for static scenes, these methods cannot be simply employed to support realistic facial expressions, such as in the case of a dynamic facial performance. To address these challenges, we propose a novel fast 3D neural implicit head avatar model that achieves real-time rendering while maintaining fine-grained controllability and high rendering quality. Our key idea lies in the introduction of local hash table blendshapes, which are learned and attached to the vertices of an underlying face parametric model. These per-vertex hash-tables are linearly merged with weights predicted via a CNN, resulting in expression dependent embeddings. Our novel representation enables efficient density and color predictions using a lightweight MLP, which is further accelerated by a hierarchical nearest neighbor search method. Extensive experiments show that our approach runs in real-time while achieving comparable rendering quality to state-of-the-arts and decent results on challenging expressions.", + "bbox": [ + 75, + 520, + 472, + 883 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 501, + 488, + 630, + 503 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The demand of high performing photo-realistic human avatars has dramatically increased with emerging VR/AR applications, e.g. VR gaming [2, 33], virtual assistant [3], tele-presence [28], and 3D videos [15, 26]. How to build efficient high quality avatars from monocular RGB videos becomes a promising direction due to the convenience of monocular data acquisition. While early works mostly adopt surface-based models for convenient controllability, recent methods (e.g. MonoAvatar [5]) leverage a sophisticated pipeline to build human avatars on neural radiance fields, which delivers vivid animations as well as significantly better rendering quality, especially over challenging parts such as hairs and glasses. On the downside, these approaches tend to be prohibitively slow, and most of the computation is consumed by the neural radiance field inference with large Multilayer Perceptrons (MLPs).", + "bbox": [ + 496, + 529, + 890, + 770 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recently, fast approaches for neural radiance fields (e.g. hash encoding in Instant NGPs [27]) have been proposed, which are designed mostly for static scenes or pre-recorded temporal sequences. Despite their great success, it is not straightforward to extend these approaches for human avatars, which requires real-time rendering of dynamic facial performances when controlling the avatar. NeRFblend", + "bbox": [ + 496, + 771, + 890, + 876 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Work was conducted while Ziqian Bai was an intern at Google.", + "bbox": [ + 522, + 886, + 864, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "1975", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "shape [14] address these issues by learning multiple feature hash tables, one for each face blendshape. These hash tables are linearly combined with blending weights to render the target facial expression via hash encoding. However, the expressiveness of the avatar is compromised by the global nature of the blending shapes, which cannot accurately capture vertex-level local deformations. On the other hand, INSTA [42] proposes to build the appearance model in a canonical space using a vanilla Instant NGPs [27] with expression codes as the additional MLP input to capture dynamic details, which is then transformed into target expression via a face parametric model (i.e. 3D morphable model (3DMM)). However, the lightweight MLP used in the vanilla Instant NGPs limits their model capacity, resulting in inferior animation quality, especially on extreme expressions.", + "bbox": [ + 75, + 90, + 472, + 332 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this work, we propose a novel 3D neural avatar system that achieves efficient inference while maintaining fine-grained controllability and high fidelity quality. To achieve this, we introduce mesh-anchored hash table blendshapes, where we attach multiple, small hash tables to each of the 3DMM mesh vertices. These hash tables act as per-vertex \"local blendshapes\" (i.e., each \"blendshape\" is controlled by one local hash table) and influence only a local region. The mesh-anchored blendshapes are linearly merged with per-vertex weights predicted by a convolutional neural network in UV space from avatar driving signals, such as expression and head rotation. This results in expression-dependent hash table embeddings which offer several advantages over a global linear combination of blendshapes. Indeed by associating hash tables with individual vertices, we enhance the expressiveness of the model, allowing for more localized and nuanced facial expressions. This contrasts with global blendshapes, which apply uniform transformations across the entire face, limiting expressiveness.", + "bbox": [ + 75, + 335, + 472, + 625 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In more detail, our model starts from 3D query points, uses hash encoding [27] to gather the merged hash table embeddings from $k$ -nearest-neighbor vertices around the query point, and predicts the density and color via a small MLP. The hash encoding [27] allows us to use a very lightweight MLP to significantly reduce computation, leading to efficient inference. Additionally, the vertex-attached hash table blendshapes represent a 3DMM-anchored neural radiance field (NeRF), which can be easily controlled by the underlying 3DMM and produce high fidelity renderings as demonstrated by MonoAvatar [5]. To further accelerate our rendering speed, we propose a hierarchical $k$ -nearest-neighbor search method.", + "bbox": [ + 75, + 626, + 472, + 821 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our contributions are summarized as follows. We propose a novel approach for high quality and efficient 3D neural implicit head avatars. At the core of our model, vertex-attached local hash table blendshapes are proposed to support efficient rendering, controllability, and capturing fine", + "bbox": [ + 75, + 824, + 472, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "grained rendering details in dynamic facial performances. We also design a hierarchical querying solution to speed up the $k$ -nearest-neighbor search when pulling hash table embeddings from neighbor vertices. Extensive experiments on multiple datasets verify that we are able to speed up avatar rendering to real-time (i.e., average $\\geq 30$ FPS to render a $512 \\times 512$ video) while maintaining comparable rendering quality with the state-of-the-art high quality 3D avatar [5] and being largely superior on challenging expressions than existing efficient 3D avatars [14, 41, 42].", + "bbox": [ + 496, + 90, + 893, + 244 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 498, + 256, + 640, + 272 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Constructing photorealistic digital humans has been a extensively researched topic. Here, we focus on discussing prior work on implicit monocular head avatars and efficient rendering. We refer readers to state-of-the-art surveys [12, 32, 43] for a comprehensive literature review.", + "bbox": [ + 496, + 281, + 893, + 358 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "High Quality Head Avatar. Traditionally, high-quality head avatars have been achieved under expensive equipment configurations, such as camera arrays [6, 23, 28], depth sensors [7], and light stages [15, 25], or require laborious manual intervention [1]. Recent research efforts have focused on constructing high quality avatars from monocular RGB videos. One typical class of approaches [4, 13, 35, 40] use implicit 3D representations (i.e., neural radiance fields (NeRFs), implicit occupancy fields) to build the head avatar, which are parameterized by Multilayer perceptrons (MLPs). Although reasonable results are obtained, their rendering quality is still unsatisfactory especially for more challenging expressions. More recently, Bai et al. [5] proposed a head avatar based on 3DMM-anchored NeRFs with expression-dependent features produced by a convolution neural network in UV space. Chen et al. [9] designed local deformation fields to capture expression-dependent deformations applied on canonical NeRFs. Despite the impressive results, their methods didn't demonstrate real-time rendering capability due to expensive inference using large MLPs. In contrast, with our proposed mesh-anchored hash table blendshapes (Sec. 3), we achieve much faster rendering speed while maintaining high fidelity results.", + "bbox": [ + 496, + 364, + 893, + 712 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Efficient Neural Radiance Fields. There has been a plethora of work in recent years attempting to accelerate rendering with neural implicit representations for static objects and scenes. SNeRG [22], DVGO [31] and Plenoxels [37] propose to directly optimize voxel grids of (neural or SH) features for faster performance. However, their approach still requires a large memory footprint to store pervoxel features in 3D space. KiloNeRF [30] dramatically accelerates the original NeRF by representing the scene with thousands of tiny MLPs, however, this approach requires a complex training strategy. TensoRF [8] factorizes the feature grid into compact components, resulting in sig", + "bbox": [ + 496, + 719, + 895, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "1976", + "bbox": [ + 483, + 944, + 517, + 957 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/44e43a35c15cca6871b74a7c81e903c099dfc1401b4ff664a9d59249d41a42a1.jpg", + "image_caption": [ + "Figure 2. Overview of our pipeline. Our core avatar representation is Mesh-anchored Hash Table Blendshapes (Sec. 3.1), where multiple small hash tables are attached to each 3DMM vertex. During inference, our method starts from a displacement map encoding the facial expression, which is then fed into a U-Net to predict hash table weights and per-vertex features. The predicted weights are used to linearly combine the multiple hash tables attached on each 3DMM vertex (Sec. 3.2). During volumetric rendering (Sec. 3.3 and Fig. 3), for each query point, we search its $k$ -nearest-neighbor vertices, then pull embeddings from the merged hash tables and concatenate with the per-vertex feature to decode local density and color via a tiny MLP with two hidden layers." + ], + "image_footnote": [], + "bbox": [ + 76, + 85, + 890, + 272 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "nificantly higher memory efficiency. Concurrently, Instant NGPs [27] utilizes multi-resolution hashing for efficient encoding, resulting in high compactness. MobileNeRF [10] propose to represent NeRF based on polygons, which allows leveraging a traditional polygon pipeline to enable their method to run in real-time on mobile devices. 3D Gaussian Splatting [17] represents a radiance field with 3D Gaussian point clouds, and leverages a point rasterization pipeline to enable fast rendering. These methods, however, cannot be easily extended to controllable dynamic contents. More recently, NeRFBlendshape [14] was proposed to handle controllable expressions by learning multiple hash tables for different global blendshapes and linearly combine them with expression codes. INSTA [42] transforms all expressions into a shared 3D canonical space, then adopts the vanilla Instant NGPs [27] conditioned on expression codes to model the head avatar. Despite the fast rendering speed achieved, their methods suffer from unsatisfactory rendering quality. PointAvatar [41] utilizes point clouds to represent the head avatar and uses large MLPs to predict the colors and motions of each point, leading to slow inference. Another type of works [11, 18, 19, 34, 36, 39] use 2D convolution neural networks to directly synthesize images (i.e. 2D neural rendering) from rasterized 3DMM meshes or low resolution feature maps generated by volumetric rendering. Despite their fast speed, the 2D CNNs may break the 3D consistency, leading to temporally unstable results especially for high frequency details. In contrast, our method simultaneously achieves controllability, high quality, and efficient rendering with a fully 3D representation.", + "bbox": [ + 75, + 380, + 472, + 834 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 76, + 845, + 168, + 859 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given a monocular RGB video, our method learns a neural radiance field (NeRF) based head avatar, which can be ren-", + "bbox": [ + 75, + 869, + 468, + 898 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "dered under any specified cameras, articulated poses (i.e., neck, jaw, and eyes) and facial expressions defined by a face parametric model (i.e. 3DMM). We use FLAME [20] as the parametric model in this work, but our method can be generalized to any other mesh-based parametric models. Fig. 2 shows the overview of our method.", + "bbox": [ + 498, + 380, + 892, + 470 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our goal is to design a 3D neural implicit head avatar architecture that can simultaneously achieve high image quality, controllability, and computationally efficient rendering. To achieve this, we propose mesh-anchored hash table blendshapes (Sec. 3.1) as a novel avatar representation that can leverage both advantages from recent high-quality (i.e., 3DMM-anchored NeRF [5]) and efficient (i.e., hash encoding [27]) frameworks. More specifically, we propose to attach multiple small hash tables on each 3DMM vertex. These vertex-attached hash tables form a set of local \"blendshapes\", which will be linearly merged with predicted blending weights (Sec. 3.2), and decoded into a 3DMM-anchored NeRF to support fine-grained control and high fidelity rendering. During NeRF decoding on a query point (Sec. 3.3), we pull the embeddings from the linearly merged hash tables attached on $k$ -nearest-neighbor ( $k$ -NN) vertices from the 3DMM mesh. Using hash encoding allows us to use a very light weight MLP (only 2 hidden layers) to predict the final densities and colors, which is the key for efficient rendering. To further accelerate our approach to real-time, we leverage the fact that close query points likely share similar $k$ -NN vertices, and thus propose to group the query points into voxels and hierarchically search for $k$ -NN vertices (Sec. 3.4). Finally, our proposed avatar representation can be trained with only monocular RGB videos without any 3D scans or multi-view data (Sec. 3.5).", + "bbox": [ + 496, + 508, + 892, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "1977", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Mesh-anchored Hash Table Blendshapes", + "text_level": 1, + "bbox": [ + 76, + 90, + 424, + 107 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The core of our model is an avatar representation that can represent a 3DMM-anchored neural radiance field (NeRF) while allowing us to adopt the hash encoding [27] technique for acceleration. Recent approaches manage to adopt hash encoding into head avatars with different avatar representations (i.e., global blendshapes [14] and canonical NeRF [42]). In contrast, our solution is built upon the most recent 3DMM-anchored NeRF [5], which is superior for high quality renderings as demonstrated in our experiments (Sec. 4.2).", + "bbox": [ + 75, + 113, + 468, + 263 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We propose mesh-anchored hash table blendshapes as the new avatar representation. Given a target expression $i$ with 3DMM pose code $\\theta_{i}$ and expression code $\\psi_{i}$ , we first get the deformed 3DMM mesh $\\mathbf{V}_{i} = \\mathcal{F}_{3\\mathrm{DMM}}(\\psi_{i},\\theta_{i})$ with $J$ vertices. For the $j$ -th 3DMM mesh vertex $\\mathbf{v}_{ij}$ , we attach $M$ small hash tables $\\{\\mathbf{H}_j^{(m)}\\}_{M}$ on it, where each hash table has multiple resolutions following instant NGPs [27]. Intuitively, these hash tables form a set of vertex-level \"blendshapes\" anchored on the mesh, where each \"blendshape\" is a hash table, whose embeddings encode the information of a local radiance field around vertex $\\mathbf{v}_{ij}$ . Given a target expression to render, these hash table blendshapes are linearly summed via expression-dependent weights (Sec. 3.2), such that the merged embeddings encode the fine details specific to the target expression. Simultaneously, the coarse motion of the target expression is captured by the 3DMM vertex movement, which moves the attached hash tables accordingly, and hence the corresponding local radiance field.", + "bbox": [ + 75, + 265, + 470, + 540 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Merge Mesh-anchored Blendshapes", + "text_level": 1, + "bbox": [ + 76, + 547, + 387, + 564 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We obtain per-vertex blending weights by running a convolution neural network (CNN) on the 3DMM deformation represented in UV atlas space. Specifically, we calculate the vertex displacements with respect to the neutral face $\\mathbf{D}_i = \\mathcal{F}_{\\mathrm{3DMM}}(\\psi_i,\\pmb {\\theta}_i) - \\mathcal{F}_{\\mathrm{3DMM}}(\\mathbf{0},\\mathbf{0})$ . The displacements are then warped into the UV space and fed into a U-Net to predict a weights map in $\\mathbb{R}^{\\mathrm{H_t}\\times \\mathrm{W_t}\\times \\mathrm{M}}$ , where $\\mathrm{H_t}\\times \\mathrm{W_t}$ is the UV resolution, and $M$ is the number of hash table blendshapes on each vertex (pre-defined as 5 in our experiments). The weights map is then sampled back to 3DMM vertices, serving as the expression-dependent weights $\\{w_{ij}^{(m)}\\}_{M}$ to take a weighted sum of the embeddings in the hash tables on each vertex, which produces the merged hash tables", + "bbox": [ + 75, + 570, + 468, + 770 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathbf {H}} _ {i j} = \\sum_ {m = 1} ^ {M} w _ {i j} ^ {(m)} \\mathbf {H} _ {j} ^ {(m)}. \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 191, + 777, + 468, + 818 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The U-Net also produces a UV feature map. We sample a per-vertex feature $\\mathbf{f}_{ij}$ from it, similar to MonoAvatar [5]. We empirically found this benefits the geometry quality. The mesh-anchored hash tables $\\hat{\\mathbf{H}}_{ij}$ and features $\\mathbf{f}_{ij}$ are decoded into a neural radiance field as described in Sec. 3.3.", + "bbox": [ + 75, + 824, + 468, + 898 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/72db2b65c8180081479f4da2a586cfa6c43a25b8ea455a2da735e2d2c62df568.jpg", + "image_caption": [ + "Figure 3. Pipeline of Neural Radiance Field Decoding (Sec. 3.3). Our method only needs a very lightweight MLP for NeRF inference, where a two-hidden-layer MLP is used to predict the density and the color of each query point." + ], + "image_footnote": [], + "bbox": [ + 511, + 88, + 883, + 316 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Neural Radiance Field Decoding", + "text_level": 1, + "bbox": [ + 500, + 405, + 785, + 422 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Given the mesh-anchored hash tables $\\hat{\\mathbf{H}}_{ij}$ and features $\\mathbf{f}_{ij}$ described in Sec. 3.2, the final step is to decode them into a Neural Radiance Field (NeRF) to render the output image as shown in Fig. 3. The key idea is to associate a query point to neighbor vertices, and pull the embeddings from the attached hash tables via hash encoding [27]. Finally, we decode these pulled embeddings and the nearest per-vertex feature into the color and the density of this query point, followed by volumetric rendering to obtain the output image.", + "bbox": [ + 496, + 428, + 890, + 564 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For a 3D query point $\\mathbf{q}$ when rendering a particular facial expression $i$ , we first obtain its $k$ -nearest-neighbors, denoted as $\\{\\mathbf{v}_{ik}\\}_{k\\in \\mathcal{N}_{\\mathbf{q}}^{K}}$ , from the 3DMM vertices, with $k^*$ denoting the nearest vertex index. For each neighbor vertex $\\mathbf{v}_{ik}$ with an attached hash table $\\hat{\\mathbf{H}}_{ik}$ , we denote $\\mathbf{q}_{ik}$ as the coordinates of $\\mathbf{q}$ in the tangent space of $\\mathbf{v}_{ik}$ . We then use $\\mathbf{q}_{ik}$ to query the hash table $\\hat{\\mathbf{H}}_{ik}$ using a hash encoding function $\\mathcal{H}(\\cdot)$ and obtain the embedding $\\mathbf{h}_{ik} = \\mathcal{H}(\\mathbf{q}_{ik};\\hat{\\mathbf{H}}_{ik})$ . To interpolate the embeddings from all $k$ -nearest-neighbors, we use the weighted sum of the inverse distances $z_{k} = 1 / \\| \\mathbf{q}_{ik}\\|_{2}$ . Next, the summed embedding, together with the nearest per-vertex feature $\\mathbf{f}_{ik^*}$ and the query point tangent coordinate $\\mathbf{q}_{ik^*}$ of the nearest vertex, are fed into a two-hidden-layer MLP to predict the density and color as", + "bbox": [ + 496, + 565, + 892, + 777 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\overline {{\\mathbf {h}}} _ {i} = \\sum_ {k \\in \\mathcal {N} _ {\\mathbf {q}} ^ {K}} \\bar {w} _ {k} \\mathbf {h} _ {i k}, \\text {w h e r e} \\bar {w} _ {k} = \\frac {z _ {k}}{\\sum_ {k ^ {\\prime} \\in \\mathcal {N} _ {\\mathbf {q}} ^ {K}} z _ {k ^ {\\prime}}}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 535, + 782, + 890, + 819 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\left[ \\mathbf {c} _ {i} (\\mathbf {q}, \\mathbf {d}), \\sigma_ {i} (\\mathbf {q}) \\right] = \\mathcal {F} _ {\\mathrm {M L P}} \\left(\\overline {{\\mathbf {h}}} _ {i}, \\mathbf {f} _ {i k ^ {*}}, \\operatorname {P E} \\left(\\mathbf {q} _ {i k ^ {*}}\\right), \\operatorname {P E} (\\mathbf {d})\\right),\n$$\n", + "text_format": "latex", + "bbox": [ + 506, + 823, + 883, + 848 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\overline{w}_k$ is the normalized inverse-distance based weight, $\\mathbf{d}$ denotes the camera view direction, $\\mathrm{PE}(\\cdot)$ denotes positional encoding, $\\mathbf{c}_i$ denotes color, and $\\sigma_{i}$ denotes density.", + "bbox": [ + 496, + 854, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "1978", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Finally, we render the output pixel with the given camera ray $\\mathbf{r}$ by volumetric rendering, where we reparameterize the query point with samples on the ray $\\mathbf{q} = \\mathbf{r}(t) = \\mathbf{o} + t\\mathbf{d}$ :", + "bbox": [ + 75, + 90, + 468, + 137 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {C} _ {i} (\\mathbf {r}) = \\int_ {t _ {n}} ^ {t _ {f}} T (t) \\sigma_ {i} (\\mathbf {r} (t)) \\mathbf {c} _ {i} (\\mathbf {r} (t), \\mathbf {d}) d t,\n$$\n", + "text_format": "latex", + "bbox": [ + 153, + 143, + 428, + 180 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {w h e r e} T (t) = \\exp \\left(- \\int_ {t _ {n}} ^ {t} \\sigma_ {i} (\\mathbf {r} (s)) d s\\right). \\qquad (3)\n$$\n", + "text_format": "latex", + "bbox": [ + 119, + 181, + 468, + 217 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Following prior works [5, 16], we also introduce a per-frame error-correction warping field during training to reduce misalignments due to the noise in 3DMM tracking and unmodeled per-frame contents such as hair movements. We feed the query point $\\mathbf{q}$ , together with a per-frame latent code $\\mathbf{e}_i$ , into an MLP $\\mathcal{F}_{\\mathcal{E}}(\\cdot)$ to obtain a rigid transformation applied on the original query point, denoted as $\\mathbf{q}' = \\mathcal{T}_i(\\mathbf{q}) = \\mathcal{F}_{\\mathcal{E}}(\\mathbf{q},\\mathbf{e}_i)$ . The warped query point $\\mathbf{q}'$ is then used to compute the density and color for volumetric rendering. Since the warping fields are overfit to corresponding training frames, we disable the warping field during testing similar to previous works [5, 16], and hence $\\mathcal{F}_{\\mathcal{E}}$ does not affect rendering efficiency.", + "bbox": [ + 75, + 220, + 470, + 417 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Hierarchical $k$ -nearest-neighbor Search", + "text_level": 1, + "bbox": [ + 76, + 424, + 416, + 439 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As described in Sec. 3.3, our method involves a $k$ -nearest-neighbor ( $k$ -NN) search, which is computationally expensive and cannot be naively accelerated with pre-calculated structures (e.g., KD-Tree) due to the dynamically changing search pool (i.e., the 3DMM vertices driven by poses and expressions). To speed up the process, we propose a hierarchical $k$ -NN search algorithm following a coarse-to-fine strategy. The key idea is to group nearby query points into a cluster as they likely share similar nearest neighbors. Specifically, we use a 3D grid with resolution 64 and treat all query points that fall in each voxel as a cluster. For each cluster, we first search $K'$ (where $K' > K$ ) nearest neighbors of the voxel center from all 3DMM vertices. Then, for each query point, we search $K$ nearest neighbors from the $K'$ nearest neighbors of the corresponding cluster. In practice, for a 3DMM with the vertices number of $J = 1772$ , we set $K' = 12$ and $K = 3$ . Our experiments empirically show that, with a proper grid resolution, this design significantly improves the nearest neighbor search speed, and does not introduce noticeable rendering artifacts, even though the $k$ -NNs may not be accurate on some of the query points.", + "bbox": [ + 75, + 446, + 468, + 765 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.5. Training", + "text_level": 1, + "bbox": [ + 76, + 771, + 181, + 786 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Only monocular RGB videos are required to train our model. Three losses are used during the training process: (1) a photometric loss that minimizes the $l_{2}$ -norm distance between the rendered and ground truth pixel colors over all camera rays $\\mathbf{r}$ from all training frames $i$ . Formally, we have $\\mathcal{L}_{\\mathrm{rgb}} = \\sum_{i} \\sum_{\\mathbf{r}} \\| \\mathbf{C}_i(\\mathbf{r}) - \\mathbf{I}_i(\\mathbf{r}) \\|_2$ ; (2) a elastic regularization loss $\\mathcal{L}_{\\mathrm{elastic}}$ applied on the learned error-correction", + "bbox": [ + 75, + 794, + 467, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "warping field $\\mathcal{T}(\\mathbf{q})$ , which is introduced in Nerfies [29]; (3) a magnitude regularization loss to encourage small warping fields, which is defined as $\\mathcal{L}_{\\mathrm{mag}} = \\sum_{\\mathbf{q}}\\| \\mathbf{q} - \\mathcal{T}(\\mathbf{q})\\| _2^2$ . Finally, we combine all three loss terms:", + "bbox": [ + 496, + 90, + 890, + 151 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\mathcal {L} _ {\\mathrm {r g b}} + \\lambda_ {\\text {e l a s t i c}} \\mathcal {L} _ {\\text {e l a s t i c}} + \\lambda_ {\\text {m a g}} \\mathcal {L} _ {\\text {m a g}}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 557, + 164, + 890, + 180 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where we set $\\lambda_{\\mathrm{elastic}} = 10^{-4}$ at the beginning of the training and decay it to $10^{-5}$ after 150k iterations, and set $\\lambda_{\\mathrm{mag}} = 10^{-2}$ . To warm start training, we replace the $l_{2}$ -norm distance in the photometric loss $\\mathcal{L}_{\\mathrm{rgb}}$ with the $l_{2}$ distance for the first 10k iterations. Please refer to the supplementary for more details.", + "bbox": [ + 496, + 191, + 890, + 282 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 498, + 295, + 632, + 313 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this section, we first introduce the data and metrics used for training and evaluation (Sec. 4.1). Then, we show that our avatar model achieves real-time rendering speed, while producing superior rendering quality on challenging expressions than recent efficient avatars [14, 41, 42] and being comparable to previous high-quality approaches [5] (Sec. 4.2). Finally, we provide ablation studies to justify the design choices and hyper-parameters of our avatar representation, and demonstrate the rendering speed improvements contributed from each of our newly proposed algorithmic components.", + "bbox": [ + 496, + 321, + 890, + 487 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Data and Metrics", + "text_level": 1, + "bbox": [ + 498, + 496, + 669, + 511 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Data. We use monocular RGB videos of multiple subjects (i.e., one video for one subject) to train and evaluate our method, and compare to prior state-of-the-art (SOTA) approaches. Our dataset consists of 10 videos in total, which are a mix of videos captured by us, as well as videos from prior works including PointAvatar [41], INSTA [42], and MonoAvatar [5]. We filter out the background of the videos with off-the-shelf segmentation [24] and matting [21] methods, then crop and resize the videos into a VGA resolution that preserves the original aspect ratio. We compute the camera and 3DMM parameters from the videos following the 3DMM fitting optimization used in INSTA [42]. We reserve a short clip from the end of each video as the testing frames, and use the rest frames for training.", + "bbox": [ + 496, + 518, + 890, + 731 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Metrics. Following prior arts [13], we use PSNR, SSIM (higher is preferable), and LPIPS (lower is preferable) to measure the image quality. As observed by Zhang et al. [38], LPIPS is a more effective metric in judging the perceptual quality compared to PSNR and SSIM. When computing PSNR and SSIM, we weigh the mean squared error map and the SSIM map with a foreground mask (eroded and smoothed), in order to focus on non-empty areas and avoid the inaccurate foreground segmentation from dominating the metrics.", + "bbox": [ + 496, + 750, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "1979", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To evaluate the computational cost, we measure the rendering speed in frames-per-second (FPS) on a RTX3090Ti and compare across different approaches with their available implementations. We also estimate the number of FLOPs (floating-point operations) of all methods as the theoretical measurement for the rendering speed. When estimating FLOPs, we fix the contribution of the ray-marching part to 16 points sampled along each camera ray. This simplifies the estimate, since the ray-marching varies the number of FLOPs needed across cameras and scenes, and it applies to all the considered methods.", + "bbox": [ + 75, + 90, + 472, + 257 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Comparison to State-of-the-art", + "text_level": 1, + "bbox": [ + 76, + 270, + 352, + 286 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We compare our method with several prior works, including: NeRFBlendshape [13], INSTA [42], PointAvatar [41], and MonoAvatar [5]. NeRFBlendshape [13] and INSTA [42] adopt hash encoding [27] into head avatars, leading to efficient renderings. PointAvatar [41] leverages point clouds to represent the head avatar. MonoAvatar [5] is based on a 3DMM-anchored NeRF, and produces high-quality renderings but is slow in speed. We use the same camera and 3DMM parameters to train and test all methods.", + "bbox": [ + 75, + 294, + 468, + 430 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "From Tab. 1, ours and INSTA [42] are the only 2 methods that can achieve real-time rendering (i.e., $\\geq 30$ mean FPS). However, INSTA is quantitatively inferior than our method by a large margin, and gives obvious artifacts in Fig. 4, especially for challenging expressions. PointAvatar [41] has the potential to run in real-time with an optimized implementation thanks to its point cloud representation, but their renderings are overall blurrier than ours, leading to worse quantitative results. Although NeRFBlendshape [14] gives relatively good numbers in Tab. 1, it produces severe artifacts in dynamically changing regions (e.g., mouth and eyebrows in Fig. 4) for several median and large expressions and also gives more floaters, resulting in implausible animations. We highly suggest readers to see the supplementary videos for more comparisons. MonoAvatar [5] gives good rendering qualities and animations, but is one order of magnitudes slower and slightly blurrier on high frequency details such as forehead wrinkles, presumably because that the hash encoding in our model can better capture high frequency contents. Among these compared approaches, our method is the only one that achieves real-time rendering while being one of the best on image qualities.", + "bbox": [ + 75, + 431, + 472, + 763 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We also compare the theoretical FLOPs of all methods in Tab. 1, where our method requires the least computation mostly because of the smaller MLPs we use (e.g., 2 hidden layers of ours vs. $\\geq 5$ hidden layers of others). Note that INSTA [42] is implemented with a highly optimized pure $\\mathrm{C}++$ and CUDA codebase, while other methods use python (tensorflow/pytorch) with customized CUDA kernels. This implementation advantage of INSTA makes it running in a high FPS even with a relatively larger FLOPs.", + "bbox": [ + 75, + 763, + 470, + 902 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/21032d3b9ac2a6e87ce9fd6ec3c97548436808291297b9d9f706be99286ae919.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
LPIPSSSIMPSNRMean FPSGFLOPs
PointAvatar [41]0.1170.72821.125.0933
INSTA [42]0.1490.75822.1246.2266
NeRFBlendshape [14]0.1100.79322.7711.2223
MonoAvatar [5]0.1140.79822.740.52385
Ours0.1000.79522.7735.9113
", + "bbox": [ + 501, + 89, + 893, + 186 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1. Quantitative comparisons with state-of-the-art approaches. Our method achieves rendering quality among the best, while supports real-time rendering with a $512 \\times 512$ resolution.", + "bbox": [ + 498, + 188, + 890, + 231 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Ablation Study", + "text_level": 1, + "bbox": [ + 500, + 244, + 653, + 262 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this section we show the impact of the proposed design choices, in particular proving the importance of mesh-anchored hash table blendshapes and the proposed hierarchical $k$ -NN search.", + "bbox": [ + 498, + 268, + 890, + 329 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3.1 Mesh-anchored Hash Table Blendshapes", + "text_level": 1, + "bbox": [ + 498, + 347, + 836, + 364 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We hereby investigate alternative design choices and different hyper-parameters to justify the necessity of our mesh-anchored hash table blendshapes.", + "bbox": [ + 496, + 371, + 890, + 416 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Static Hash + 3DMM Param. We first build a naive alternative approach to incorporate hash encoding into 3DMM-anchored NeRF, by attaching a single hash table to each vertex. For a query point, we concatenate the 3DMM pose and expression codes $(\\theta_{i},\\psi_{i})$ with the embedding pulled from the hash tables, and send them into the MLP. Note that there is no convolution running in UV space. As shown in Tab. 2 and Fig. 5, obvious rendering artifacts show up, and the rendering quality metrics drop significantly compared to our full model. This is presumably because the lightweight MLP, which is crucial for good efficiency, does not have enough capacity to process detailed expression-dependent information from compact 3DMM codes.", + "bbox": [ + 496, + 417, + 892, + 613 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Static Hash + UV CNN. We then increase the model capacity by adding back the UV CNN branch, but use only one hash table per-vertex, i.e., single hash table without blend-shape formulation. As shown in Tab. 2 and Fig. 5, the rendered images show less artifacts over the previous case, but still contain blurry textures and floaters compared to our full model. This demonstrates that the blendshape formulation of mesh-anchored hash tables are necessary in order to obtain good expression dependent local embeddings, leading to a superior rendering quality.", + "bbox": [ + 496, + 613, + 893, + 763 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Number of Hash Table Blendshapes. Here we investigate how the number of hash table blendshapes per vertex influences the final rendering quality. As shown in Tab. 2, we can see that more blendshapes per vertex leads to higher rendering qualities, which saturates as the number of tables increases. In Fig. 5, increasing the number of blendshapes also gives better details especially for eyelids and ears. Although further adding more blendshapes may produce a better quality, we choose to use 5 blendshapes per-vertex as our", + "bbox": [ + 496, + 763, + 893, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "1980", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/7ccb91c29e5b751006d3df318f43de24172b32ed48cc011e3641c1362821c9fe.jpg", + "image_caption": [ + "Figure 4. Comparisons on the rendering quality to previous state-of-the-art methods. From left to right, each column contains the images of: 1) PointAvatar [41], 2) INSTA [42], 3) NeRFBlendshape [14], 4) MonoAvatar [5], 5) Ours, 6) Ground Truth. Our method faithfully reconstructs the personalized expressions and high-frequency details, achieving one of the best rendering quality with real-time rendering speed." + ], + "image_footnote": [], + "bbox": [ + 84, + 87, + 890, + 806 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "1981", + "bbox": [ + 483, + 945, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/75c55b298d7d131cb7a338c066782833fa72f3a48110563b9632b328914526b0.jpg", + "image_caption": [ + "Figure 5. Qualitative results of our full model and various alternative design choices to demonstrate the necessity of our mesh-anchored hash table blendshapes. Please refer to Sec. 4.3.1 for details of each alternative design choice. \"Static Hash + 3DMM Param\" gives unfaithful expressions and obvious artifacts around eyes and ears. \"Static Hash + UV CNN\" produces more faithful expressions, but still suffers from floaters. By utilizing the blendshape formulation of hash tables, \"3 Blendshapes\" eliminates most of the artifacts, but produces blurriness on the details of eyes and ears. \"Ours\" gives the best rendering quality by increasing the number of blendshapes to 5, leading to cleaner details of the eyelids and ear boundary." + ], + "image_footnote": [], + "bbox": [ + 83, + 88, + 890, + 212 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/b74324f3e847ea6e233f71207889aff1afaf2aa8018d81e238365f14eab65d8e.jpg", + "image_caption": [ + "Figure 6. Qualitative results of using different 3D grid resolutions during hierarchical $k$ -NN search and without hierarchical $k$ -NN search for ablation studies. \"16 grid resolution\" results in blocky artifacts around mouth and nose. \"32 grid resolution\" removes blocky artifacts, but produces floaters inside mouth. \"64 grid resolution\" produces a similar level of visual quality as \"w/o hierarchical $k$ -NN search\"." + ], + "image_footnote": [], + "bbox": [ + 81, + 306, + 890, + 411 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "final setting to maintain a relatively small model size and computation cost.", + "bbox": [ + 75, + 469, + 468, + 500 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.3.2 Hierarchical $k$ -NN Search", + "text_level": 1, + "bbox": [ + 76, + 518, + 313, + 534 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We evaluate our model with and without hierarchical $k$ -NN search in terms of speed and quality. From the comparison over rendering speeds (w/: 35.9 FPS; w/o: 26.4 FPS), we can see that hierarchical $k$ -NN search gives around $36\\%$ improvements on the frame rate, which is crucial for achieving real-time rendering. From Fig. 6, we empirically find that enabling hierarchical $k$ -NN search will not lead to observable drops on the rendering quality, as long as a proper grid resolution is used (i.e., 64 in our case).", + "bbox": [ + 75, + 542, + 468, + 678 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We also investigate the affect of using different 3D grid resolutions during hierarchical $k$ -NN search. As shown in Fig. 6, we observe more artifacts around the mouth region when using smaller 3D grid resolutions (i.e., 32 and 16). Therefore, we choose to use 64 resolution in our final setting, which is a good trade-off between quality and speed.", + "bbox": [ + 75, + 679, + 468, + 772 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Discussion", + "text_level": 1, + "bbox": [ + 76, + 784, + 191, + 800 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We present a high quality 3D neural volumetric head avatar that can be rendered efficiently, while only requires monocular RGB videos for construction. We propose the mesh-anchored hash table blendshapes as our avatar representation, which enable a significantly faster rendering speed by utilizing hash encoding and lightweight MLPs, while still", + "bbox": [ + 75, + 809, + 470, + 902 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/414a31889138248f198c203ddfad31931b450d380e336321e8a9ef91c417dfce.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
LPIPSSSIMPSNR
Static Hash + 3DMM Param0.1250.76321.99
1 Blendshape (Static Hash + UV CNN)0.1150.78522.52
3 Blendshapes0.1040.79122.70
5 Blendshapes (Ours)0.1000.79522.77
", + "bbox": [ + 501, + 467, + 890, + 551 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 2. Quantitative results for the ablation study. Our full model (i.e., mesh-anchored hash table blendshapes) consistently outperforms alternative design choices discussed in Sec. 4.3.1.", + "bbox": [ + 498, + 556, + 890, + 598 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "maintaining superior controllability to support realistic facial animations, and producing vivid expression-dependent details thank to the local blendshape formulation of hash tables. The experiments indicate that our approach runs in real-time at a $512 \\times 512$ resolution, while giving a rendering quality comparable to state-of-the-art, with better challenging expressions than prior efficient approaches.", + "bbox": [ + 496, + 613, + 890, + 719 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "As a limitation, we observe floaters under camera viewpoints and expressions that are far from the training distribution, which is a common issue in instant NGPs [27] based approaches. We also notice that performance around the mouth interior regions tends to be less stable because of the relatively poor tracking in these areas on the training data. Fortunately, the fast rendering could enable the possibility to adopt more expensive training strategies, such as regularization terms, adversarial loss, or joint face fitting refinement during the training, which could potentially mitigate these issues and further improve the rendering expressiveness and quality.", + "bbox": [ + 496, + 719, + 892, + 888 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "1982", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] MetaHuman - Unreal Engine. https://www.unrealengine.com/en-US/metahuman. Accessed: 2022-10-17. 2", + "[2] VRChat. https://hello.vrchat.com. Accessed: 2023-02-05.1", + "[3] XHolo Virtual Assistant. https://www.digalix.com/en/virtual-assistant-augmented-reality-xholo. Accessed: 2023-03-01. 1", + "[4] ShahRukh Athar, Zexiang Xu, Kalyan Sunkavalli, Eli Shechtman, and Zhixin Shu. RigNeRF: Fully Controllable Neural 3D Portraits. pages 20364-20373, 2022. 2", + "[5] Ziqian Bai, Feitong Tan, Zeng Huang, Kripasindhu Sarkar, Danhang Tang, Di Qiu, Abhimitra Meka, Ruofei Du, Mingsong Dou, Sergio Orts-Escalano, et al. Learning personalized high quality volumetric head avatars from monocular rgb videos. In CVPR, pages 16890-16900, 2023. 1, 2, 3, 4, 5, 6, 7", + "[6] Thabo Beeler, Fabian Hahn, Derek Bradley, Bernd Bickel, Paul Beardsley, Craig Gotsman, Robert W Sumner, and Markus Gross. High-Quality Passive Facial Performance Capture Using Anchor Frames. In ACM SIGGRAPH 2011 Papers, pages 1-10. 2011. 2", + "[7] Chen Cao, Tomas Simon, Jin Kyu Kim, Gabe Schwartz, Michael Zollhoefer, Shun-Suke Saito, Stephen Lombardi, Shih-En Wei, Danielle Belko, Shouu-I Yu, Yaser Sheikh, and Jason Saragih. Authentic volumetric avatars from a phone scan. ACM TOG, 2022. 2", + "[8] Anpei Chen, Zexiang Xu, Andreas Geiger, Jingyi Yu, and Hao Su. Tensorf: Tensorial radiance fields. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXXII, pages 333-350. Springer, 2022. 2", + "[9] Chuhan Chen, Matthew O'Toole, Gaurav Bharaj, and Pablo Garrido. Implicit neural head synthesis via controllable local deformation fields. In CVPR, pages 416-426, 2023. 2", + "[10] Zhiqin Chen, Thomas Funkhouser, Peter Hedman, and Andrea Tagliasacchi. Mobilenerf: Exploiting the polygon rasterization pipeline for efficient neural field rendering on mobile architectures. arXiv preprint arXiv:2208.00277, 2022.3", + "[11] Michail Christos Doukas, Mohammad Rami Koujan, Viktoriaa Sharmanska, Anastasios Roussos, and Stefanos Zafeiriou. Head2head++: Deep Facial Attributes Re-Targeting. IEEE Transactions on Biometrics, Behavior, and Identity Science, 3(1):31-43, 2021. 3", + "[12] Bernhard Egger, William AP Smith, Ayush Tewari, Stefanie Wuhrer, Michael Zollhoefer, Thabo Beeler, Florian Bernard, Timo Bolkart, Adam Kortylewski, Sami Romdhani, et al. 3D Morphable Face Models—Past, Present, and Future. 39(5): 1-38, 2020. 2", + "[13] Guy Gafni, Justus Thies, Michael Zollhofer, and Matthias Nießner. Dynamic neural radiance fields for monocular 4d facial avatar reconstruction. In CVPR, pages 8649-8658, 2021. 2, 5, 6", + "[14] Xuan Gao, Chenglai Zhong, Jun Xiang, Yang Hong, Yudong Guo, and Juyong Zhang. Reconstructing person" + ], + "bbox": [ + 78, + 114, + 467, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "alized semantic facial nerf models from monocular video. ACM Transactions on Graphics (Proceedings of SIGGRAPH Asia), 41(6), 2022. 1, 2, 3, 4, 5, 6, 7", + "[15] Kaiwen Guo, Peter Lincoln, Philip Davidson, Jay Busch, Xueming Yu, Matt Whalen, Geoff Harvey, Sergio Orts-Escolano, Rohit Pandey, Jason Dourgarian, Danhang Tang, Anastasia Tkach, Adarsh Kowdle, Emily Cooper, Ming-song Dou, Sean Fanello, Graham Fyffe, Christoph Rhemann, Jonathan Taylor, Paul Debevec, and Shahram Izadi. The Relightables: Volumetric Performance Capture of Humans With Realistic Relighting. ACM TOG, 2019. 1, 2", + "[16] Wei Jiang, Kwang Moo Yi, Golnoosh Samei, Oncel Tuzel, and Anurag Ranjan. NeuMan: Neural Human Radiance Field From a Single Video. 2022. 5", + "[17] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM TOG, 42(4):1-14, 2023. 3", + "[18] Hyeongwoo Kim, Pablo Garrido, Ayush Tewari, Weipeng Xu, Justus Thies, Matthias Niessner, Patrick Pérez, Christian Richardt, Michael Zollhöfer, and Christian Theobalt. Deep Video Portraits. ACM TOG, 37(4):1-14, 2018. 3", + "[19] Mohammad Rami Koujan, Michail Christos Doukas, Anastasios Roussos, and Stefanos Zafeiriou. Head2head: Video-Based Neural Head Synthesis. In 2020 15th IEEE International Conference on Automatic Face and Gesture Recognition (FG 2020), pages 16-23. IEEE, IEEE, 2020. 3", + "[20] Tianye Li, Timo Bolkart, Michael. J. Black, Hao Li, and Javier Romero. Learning a Model of Facial Shape and Expression From 4D Scans. ACM TOG, 36(6):194:1-194:17, 2017. 3", + "[21] Shanchuan Lin, Linjie Yang, Imran Saleemi, and Soumyadip Sengupta. Robust high-resolution video matting with temporal guidance. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 238-247, 2022. 5", + "[22] Lingjie Liu, Jiatao Gu, Kyaw Zaw Lin, Tat-Seng Chua, and Christian Theobalt. Neural sparse voxel fields. Advances in Neural Information Processing Systems, 33:15651-15663, 2020. 2", + "[23] Stephen Lombardi, Tomas Simon, Gabriel Schwartz, Michael Zollhoefer, Yaser Sheikh, and Jason Saragih. Mixture of volumetric primitives for efficient neural rendering. ACM TOG, 40(4):1-13, 2021. 2", + "[24] Camillo Lugaresi, Jiuqiang Tang, Hadon Nash, Chris McClanahan, Esha Uboweja, Michael Hays, Fan Zhang, Chuoling Chang, Ming Guang Yong, Juhyun Lee, et al. Mediapipe: A framework for building perception pipelines. arXiv preprint arXiv:1906.08172, 2019. 5", + "[25] Abhimitra Meka, Rohit Pandey, Christian Haene, Sergio Orts-Escalano, Peter Barnum, Philip David-Son, Daniel Erickson, Yinda Zhang, Jonathan Taylor, Sofien Bouaziz, et al. Deep Relightable Textures: Volumetric Performance Capture With Neural Rendering. ACM TOG, 39(6):1-21, 2020. 2", + "[26] Abhinitra Meka, Rohit Pandey, Christian Haene, Sergio Orts-Escolano, Peter Barnum, Philip Davidson, Daniel Erickson, Yinda Zhang, Jonathan Taylor, Sofien Bouaziz," + ], + "bbox": [ + 503, + 93, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "1983", + "bbox": [ + 483, + 945, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Chloe Legendre, Wan-Chun Ma, Ryan Overbeck, Thabo Beeler, Paul Debevec, Shahram Izadi, Christian Theobalt, Christoph Rhemann, and Sean Fanello. Deep relightable textures - volumetric performance capture with neural rendering. 2020. 1", + "[27] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM Trans. Graph., 41(4):102:1-102:15, 2022. 1, 2, 3, 4, 6, 8", + "[28] Sergio Orts-Escolano, Christoph Rhemann, Sean Fanello, Wayne Chang, Adarsh Kowdle, Yury Degtyarev, David Kim, Philip Davidson, Sameh Khamis, Mingsong Dou, Vladimir Tankovich, Charles Loop, Qin Cai, Philip Chou, Sarah Mennicken, Julien Valentin, Vivek Pradeep, Shenlong Wang, Sing Bing Kang, Pushmeet Kohli, Yuliya Lutchyn, Cem Keskin, and Shahram Izadi. Holoportation: Virtual 3D Teleportation in Real-Time. In Proceedings of the 29th Annual Symposium on User Interface Software and Technology. ACM, 2016. 1, 2", + "[29] Keunhong Park, Utkarsh Sinha, Jonathan T Barron, Sofien Bouaziz, Dan B Goldman, Steven M Seitz, and Ricardo Martin-Brualla. Nerfies: Deformable Neural Radiance Fields. pages 5865-5874, 2021. 5", + "[30] Christian Reiser, Songyou Peng, Yiyi Liao, and Andreas Geiger. Kilonerf: Speeding up neural radiance fields with thousands of tiny mlps. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14335-14345, 2021. 2", + "[31] Cheng Sun, Min Sun, and Hwann-Tzong Chen. Direct voxel grid optimization: Super-fast convergence for radiance fields reconstruction. In CVPR, 2022. 2", + "[32] Ayush Tewari, Justus Thies, Ben Mildenhall, Pratul Srinivasan, Edgar Tretschk, Wang Yifan, Christoph Lassner, Vincent Sitzmann, Ricardo Martin-Brualla, Stephen Lombardi, et al. Advances in neural rendering. In Computer Graphics Forum, pages 703-735. Wiley Online Library, 2022. 2", + "[33] Zach Waggoner. My Avatar, My Self: Identity in Video Role-Playing Games. McFarland, 2009. 1", + "[34] Lizhen Wang, Xiaochen Zhao, Jingxiang Sun, Yuxiang Zhang, Hongwen Zhang, Tao Yu, and Yebin Liu. Stylea-vatar: Real-time photo-realistic portrait avatar from a single video. In ACM SIGGRAPH 2023 Conference Proceedings, pages 1-10, 2023. 3", + "[35] Yuelang Xu, Lizhen Wang, Xiaochen Zhao, Hongwen Zhang, and Yebin Liu. Avatarmav: Fast 3d head avatar reconstruction using motion-aware neural voxels. In ACM SIGGRAPH 2023 Conference Proceedings, pages 1-10, 2023. 2", + "[36] Yuelang Xu, Hongwen Zhang, Lizhen Wang, Xiaochen Zhao, Han Huang, Guojun Qi, and Yebin Liu. Latentavatar: Learning latent expression code for expressive neural head avatar. In ACM SIGGRAPH 2023 Conference Proceedings, pages 1-10, 2023. 3", + "[37] Alex Yu, Ruilong Li, Matthew Tancik, Hao Li, Ren Ng, and Angjoo Kanazawa. Plenoctrees for real-time rendering of neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5752-5761, 2021. 2" + ], + "bbox": [ + 78, + 90, + 470, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[38] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The Unreasonable Effectiveness of Deep Features As a Perceptual Metric. pages 586-595, 2018. 5", + "[39] Xiaochen Zhao, Lizhen Wang, Jingxiang Sun, Hongwen Zhang, Jinli Suo, and Yebin Liu. Havatar: High-fidelity head avatar via facial model conditioned neural radiance field. ACM TOG, 43(1):1-16, 2023. 3", + "[40] Yufeng Zheng, Victoria Fernández Abrevaya, Marcel C Bühler, Xu Chen, Michael J Black, and Otmar Hilliges. Im avatar: Implicit morphable head avatars from videos. In CVPR, pages 13545-13555, 2022. 2", + "[41] Yufeng Zheng, Wang Yifan, Gordon Wetzstein, Michael J Black, and Otmar Hilliges. Pointavatar: Deformable point-based head avatars from videos. In CVPR, pages 21057-21067, 2023. 1, 2, 3, 5, 6, 7", + "[42] Wojciech Zielonka, Timo Bolkart, and Justus Thies. Instant volumetric head avatars. In CVPR, pages 4574-4584, 2023. 1, 2, 3, 4, 5, 6, 7", + "[43] Michael Zollhöfer, Justus Thies, Pablo Garrido, Derek Bradley, Thabo Beeler, Patrick Pérez, Marc Stamminger, Matthias Nießner, and Christian Theobalt. State of the Art on Monocular 3D Face Reconstruction, Tracking, and Applications. pages 523-550, 2018. 2" + ], + "bbox": [ + 501, + 92, + 890, + 416 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "1984", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2024/Efficient 3D Implicit Head Avatar with Mesh-anchored Hash Table Blendshapes/afaaa8c5-698e-4fcb-9bde-21aa2469de9a_model.json b/2024/Efficient 3D Implicit Head Avatar with Mesh-anchored Hash Table Blendshapes/afaaa8c5-698e-4fcb-9bde-21aa2469de9a_model.json new file mode 100644 index 0000000000000000000000000000000000000000..32e902c6bcd904d1bd630dde41545023d7857c08 --- /dev/null +++ b/2024/Efficient 3D Implicit Head Avatar with Mesh-anchored Hash Table Blendshapes/afaaa8c5-698e-4fcb-9bde-21aa2469de9a_model.json @@ -0,0 +1,1749 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.084, + 0.131, + 0.888, + 0.154 + ], + "angle": 0, + "content": "Efficient 3D Implicit Head Avatar with Mesh-anchored Hash Table Blendshapes" + }, + { + "type": "text", + "bbox": [ + 0.228, + 0.179, + 0.741, + 0.217 + ], + "angle": 0, + "content": "Ziqian Bai\\(^{1,2*}\\) Feitong Tan\\(^{1}\\) Sean Fanello\\(^{1}\\) Rohit Pandey\\(^{1}\\) Mingsong Dou\\(^{1}\\) Shichen Liu\\(^{1}\\) Ping Tan\\(^{3}\\) Yinda Zhang\\(^{1}\\)" + }, + { + "type": "text", + "bbox": [ + 0.145, + 0.218, + 0.208, + 0.235 + ], + "angle": 0, + "content": "Google" + }, + { + "type": "text", + "bbox": [ + 0.24, + 0.218, + 0.416, + 0.234 + ], + "angle": 0, + "content": "\\(^{2}\\) Simon Fraser University" + }, + { + "type": "text", + "bbox": [ + 0.448, + 0.218, + 0.821, + 0.235 + ], + "angle": 0, + "content": "3 The Hong Kong University of Science and Technology" + }, + { + "type": "image", + "bbox": [ + 0.08, + 0.272, + 0.276, + 0.42 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.283, + 0.279, + 0.403, + 0.414 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.406, + 0.279, + 0.524, + 0.414 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.526, + 0.279, + 0.647, + 0.413 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.649, + 0.279, + 0.768, + 0.413 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.77, + 0.28, + 0.892, + 0.413 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.422, + 0.894, + 0.478 + ], + "angle": 0, + "content": "Figure 1. Visual Quality vs. Rendering Speed comparisons between our monocular 3D head avatars and prior state-of-the-art methods including NeRFBlendshape [14], INSTA [42], PointAvatar [41] and MonoAvatar [5]. Our approach achieves real-time rendering (i.e. \\(>30\\) mean FPS) with a \\(512 \\times 512\\) resolution, while produces comparable visual quality comparing to prior SOTAs [5], and gives significantly better results on challenging expressions than prior efficient avatars [14, 41, 42]. Webpage augmentedperception.github.io/monoavatar-plus" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.489, + 0.314, + 0.505 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.521, + 0.473, + 0.884 + ], + "angle": 0, + "content": "3D head avatars built with neural implicit volumetric representations have achieved unprecedented levels of photorealism. However, the computational cost of these methods remains a significant barrier to their widespread adoption, particularly in real-time applications such as virtual reality and teleconferencing. While attempts have been made to develop fast neural rendering approaches for static scenes, these methods cannot be simply employed to support realistic facial expressions, such as in the case of a dynamic facial performance. To address these challenges, we propose a novel fast 3D neural implicit head avatar model that achieves real-time rendering while maintaining fine-grained controllability and high rendering quality. Our key idea lies in the introduction of local hash table blendshapes, which are learned and attached to the vertices of an underlying face parametric model. These per-vertex hash-tables are linearly merged with weights predicted via a CNN, resulting in expression dependent embeddings. Our novel representation enables efficient density and color predictions using a lightweight MLP, which is further accelerated by a hierarchical nearest neighbor search method. Extensive experiments show that our approach runs in real-time while achieving comparable rendering quality to state-of-the-arts and decent results on challenging expressions." + }, + { + "type": "title", + "bbox": [ + 0.502, + 0.489, + 0.631, + 0.505 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.53, + 0.892, + 0.771 + ], + "angle": 0, + "content": "The demand of high performing photo-realistic human avatars has dramatically increased with emerging VR/AR applications, e.g. VR gaming [2, 33], virtual assistant [3], tele-presence [28], and 3D videos [15, 26]. How to build efficient high quality avatars from monocular RGB videos becomes a promising direction due to the convenience of monocular data acquisition. While early works mostly adopt surface-based models for convenient controllability, recent methods (e.g. MonoAvatar [5]) leverage a sophisticated pipeline to build human avatars on neural radiance fields, which delivers vivid animations as well as significantly better rendering quality, especially over challenging parts such as hairs and glasses. On the downside, these approaches tend to be prohibitively slow, and most of the computation is consumed by the neural radiance field inference with large Multilayer Perceptrons (MLPs)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.772, + 0.892, + 0.877 + ], + "angle": 0, + "content": "Recently, fast approaches for neural radiance fields (e.g. hash encoding in Instant NGPs [27]) have been proposed, which are designed mostly for static scenes or pre-recorded temporal sequences. Despite their great success, it is not straightforward to extend these approaches for human avatars, which requires real-time rendering of dynamic facial performances when controlling the avatar. NeRFblend" + }, + { + "type": "page_footnote", + "bbox": [ + 0.524, + 0.887, + 0.866, + 0.901 + ], + "angle": 0, + "content": "*Work was conducted while Ziqian Bai was an intern at Google." + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1975" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.333 + ], + "angle": 0, + "content": "shape [14] address these issues by learning multiple feature hash tables, one for each face blendshape. These hash tables are linearly combined with blending weights to render the target facial expression via hash encoding. However, the expressiveness of the avatar is compromised by the global nature of the blending shapes, which cannot accurately capture vertex-level local deformations. On the other hand, INSTA [42] proposes to build the appearance model in a canonical space using a vanilla Instant NGPs [27] with expression codes as the additional MLP input to capture dynamic details, which is then transformed into target expression via a face parametric model (i.e. 3D morphable model (3DMM)). However, the lightweight MLP used in the vanilla Instant NGPs limits their model capacity, resulting in inferior animation quality, especially on extreme expressions." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.336, + 0.473, + 0.625 + ], + "angle": 0, + "content": "In this work, we propose a novel 3D neural avatar system that achieves efficient inference while maintaining fine-grained controllability and high fidelity quality. To achieve this, we introduce mesh-anchored hash table blendshapes, where we attach multiple, small hash tables to each of the 3DMM mesh vertices. These hash tables act as per-vertex \"local blendshapes\" (i.e., each \"blendshape\" is controlled by one local hash table) and influence only a local region. The mesh-anchored blendshapes are linearly merged with per-vertex weights predicted by a convolutional neural network in UV space from avatar driving signals, such as expression and head rotation. This results in expression-dependent hash table embeddings which offer several advantages over a global linear combination of blendshapes. Indeed by associating hash tables with individual vertices, we enhance the expressiveness of the model, allowing for more localized and nuanced facial expressions. This contrasts with global blendshapes, which apply uniform transformations across the entire face, limiting expressiveness." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.627, + 0.473, + 0.822 + ], + "angle": 0, + "content": "In more detail, our model starts from 3D query points, uses hash encoding [27] to gather the merged hash table embeddings from \\( k \\)-nearest-neighbor vertices around the query point, and predicts the density and color via a small MLP. The hash encoding [27] allows us to use a very lightweight MLP to significantly reduce computation, leading to efficient inference. Additionally, the vertex-attached hash table blendshapes represent a 3DMM-anchored neural radiance field (NeRF), which can be easily controlled by the underlying 3DMM and produce high fidelity renderings as demonstrated by MonoAvatar [5]. To further accelerate our rendering speed, we propose a hierarchical \\( k \\)-nearest-neighbor search method." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.473, + 0.903 + ], + "angle": 0, + "content": "Our contributions are summarized as follows. We propose a novel approach for high quality and efficient 3D neural implicit head avatars. At the core of our model, vertex-attached local hash table blendshapes are proposed to support efficient rendering, controllability, and capturing fine" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.245 + ], + "angle": 0, + "content": "grained rendering details in dynamic facial performances. We also design a hierarchical querying solution to speed up the \\(k\\)-nearest-neighbor search when pulling hash table embeddings from neighbor vertices. Extensive experiments on multiple datasets verify that we are able to speed up avatar rendering to real-time (i.e., average \\(\\geq 30\\) FPS to render a \\(512 \\times 512\\) video) while maintaining comparable rendering quality with the state-of-the-art high quality 3D avatar [5] and being largely superior on challenging expressions than existing efficient 3D avatars [14, 41, 42]." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.257, + 0.642, + 0.273 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.282, + 0.895, + 0.359 + ], + "angle": 0, + "content": "Constructing photorealistic digital humans has been a extensively researched topic. Here, we focus on discussing prior work on implicit monocular head avatars and efficient rendering. We refer readers to state-of-the-art surveys [12, 32, 43] for a comprehensive literature review." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.366, + 0.895, + 0.713 + ], + "angle": 0, + "content": "High Quality Head Avatar. Traditionally, high-quality head avatars have been achieved under expensive equipment configurations, such as camera arrays [6, 23, 28], depth sensors [7], and light stages [15, 25], or require laborious manual intervention [1]. Recent research efforts have focused on constructing high quality avatars from monocular RGB videos. One typical class of approaches [4, 13, 35, 40] use implicit 3D representations (i.e., neural radiance fields (NeRFs), implicit occupancy fields) to build the head avatar, which are parameterized by Multilayer perceptrons (MLPs). Although reasonable results are obtained, their rendering quality is still unsatisfactory especially for more challenging expressions. More recently, Bai et al. [5] proposed a head avatar based on 3DMM-anchored NeRFs with expression-dependent features produced by a convolution neural network in UV space. Chen et al. [9] designed local deformation fields to capture expression-dependent deformations applied on canonical NeRFs. Despite the impressive results, their methods didn't demonstrate real-time rendering capability due to expensive inference using large MLPs. In contrast, with our proposed mesh-anchored hash table blendshapes (Sec. 3), we achieve much faster rendering speed while maintaining high fidelity results." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.897, + 0.903 + ], + "angle": 0, + "content": "Efficient Neural Radiance Fields. There has been a plethora of work in recent years attempting to accelerate rendering with neural implicit representations for static objects and scenes. SNeRG [22], DVGO [31] and Plenoxels [37] propose to directly optimize voxel grids of (neural or SH) features for faster performance. However, their approach still requires a large memory footprint to store pervoxel features in 3D space. KiloNeRF [30] dramatically accelerates the original NeRF by representing the scene with thousands of tiny MLPs, however, this approach requires a complex training strategy. TensoRF [8] factorizes the feature grid into compact components, resulting in sig" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.518, + 0.958 + ], + "angle": 0, + "content": "1976" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.078, + 0.087, + 0.891, + 0.273 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.278, + 0.893, + 0.361 + ], + "angle": 0, + "content": "Figure 2. Overview of our pipeline. Our core avatar representation is Mesh-anchored Hash Table Blendshapes (Sec. 3.1), where multiple small hash tables are attached to each 3DMM vertex. During inference, our method starts from a displacement map encoding the facial expression, which is then fed into a U-Net to predict hash table weights and per-vertex features. The predicted weights are used to linearly combine the multiple hash tables attached on each 3DMM vertex (Sec. 3.2). During volumetric rendering (Sec. 3.3 and Fig. 3), for each query point, we search its \\(k\\)-nearest-neighbor vertices, then pull embeddings from the merged hash tables and concatenate with the per-vertex feature to decode local density and color via a tiny MLP with two hidden layers." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.381, + 0.473, + 0.835 + ], + "angle": 0, + "content": "nificantly higher memory efficiency. Concurrently, Instant NGPs [27] utilizes multi-resolution hashing for efficient encoding, resulting in high compactness. MobileNeRF [10] propose to represent NeRF based on polygons, which allows leveraging a traditional polygon pipeline to enable their method to run in real-time on mobile devices. 3D Gaussian Splatting [17] represents a radiance field with 3D Gaussian point clouds, and leverages a point rasterization pipeline to enable fast rendering. These methods, however, cannot be easily extended to controllable dynamic contents. More recently, NeRFBlendshape [14] was proposed to handle controllable expressions by learning multiple hash tables for different global blendshapes and linearly combine them with expression codes. INSTA [42] transforms all expressions into a shared 3D canonical space, then adopts the vanilla Instant NGPs [27] conditioned on expression codes to model the head avatar. Despite the fast rendering speed achieved, their methods suffer from unsatisfactory rendering quality. PointAvatar [41] utilizes point clouds to represent the head avatar and uses large MLPs to predict the colors and motions of each point, leading to slow inference. Another type of works [11, 18, 19, 34, 36, 39] use 2D convolution neural networks to directly synthesize images (i.e. 2D neural rendering) from rasterized 3DMM meshes or low resolution feature maps generated by volumetric rendering. Despite their fast speed, the 2D CNNs may break the 3D consistency, leading to temporally unstable results especially for high frequency details. In contrast, our method simultaneously achieves controllability, high quality, and efficient rendering with a fully 3D representation." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.846, + 0.169, + 0.86 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.871, + 0.47, + 0.9 + ], + "angle": 0, + "content": "Given a monocular RGB video, our method learns a neural radiance field (NeRF) based head avatar, which can be ren-" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.381, + 0.893, + 0.471 + ], + "angle": 0, + "content": "dered under any specified cameras, articulated poses (i.e., neck, jaw, and eyes) and facial expressions defined by a face parametric model (i.e. 3DMM). We use FLAME [20] as the parametric model in this work, but our method can be generalized to any other mesh-based parametric models. Fig. 2 shows the overview of our method." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.509, + 0.893, + 0.901 + ], + "angle": 0, + "content": "Our goal is to design a 3D neural implicit head avatar architecture that can simultaneously achieve high image quality, controllability, and computationally efficient rendering. To achieve this, we propose mesh-anchored hash table blendshapes (Sec. 3.1) as a novel avatar representation that can leverage both advantages from recent high-quality (i.e., 3DMM-anchored NeRF [5]) and efficient (i.e., hash encoding [27]) frameworks. More specifically, we propose to attach multiple small hash tables on each 3DMM vertex. These vertex-attached hash tables form a set of local \"blendshapes\", which will be linearly merged with predicted blending weights (Sec. 3.2), and decoded into a 3DMM-anchored NeRF to support fine-grained control and high fidelity rendering. During NeRF decoding on a query point (Sec. 3.3), we pull the embeddings from the linearly merged hash tables attached on \\( k \\)-nearest-neighbor (\\( k \\)-NN) vertices from the 3DMM mesh. Using hash encoding allows us to use a very light weight MLP (only 2 hidden layers) to predict the final densities and colors, which is the key for efficient rendering. To further accelerate our approach to real-time, we leverage the fact that close query points likely share similar \\( k \\)-NN vertices, and thus propose to group the query points into voxels and hierarchically search for \\( k \\)-NN vertices (Sec. 3.4). Finally, our proposed avatar representation can be trained with only monocular RGB videos without any 3D scans or multi-view data (Sec. 3.5)." + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1977" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.077, + 0.091, + 0.426, + 0.108 + ], + "angle": 0, + "content": "3.1. Mesh-anchored Hash Table Blendshapes" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.114, + 0.47, + 0.265 + ], + "angle": 0, + "content": "The core of our model is an avatar representation that can represent a 3DMM-anchored neural radiance field (NeRF) while allowing us to adopt the hash encoding [27] technique for acceleration. Recent approaches manage to adopt hash encoding into head avatars with different avatar representations (i.e., global blendshapes [14] and canonical NeRF [42]). In contrast, our solution is built upon the most recent 3DMM-anchored NeRF [5], which is superior for high quality renderings as demonstrated in our experiments (Sec. 4.2)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.266, + 0.471, + 0.541 + ], + "angle": 0, + "content": "We propose mesh-anchored hash table blendshapes as the new avatar representation. Given a target expression \\( i \\) with 3DMM pose code \\( \\theta_{i} \\) and expression code \\( \\psi_{i} \\), we first get the deformed 3DMM mesh \\( \\mathbf{V}_{i} = \\mathcal{F}_{3\\mathrm{DMM}}(\\psi_{i},\\theta_{i}) \\) with \\( J \\) vertices. For the \\( j \\)-th 3DMM mesh vertex \\( \\mathbf{v}_{ij} \\), we attach \\( M \\) small hash tables \\( \\{\\mathbf{H}_j^{(m)}\\}_{M} \\) on it, where each hash table has multiple resolutions following instant NGPs [27]. Intuitively, these hash tables form a set of vertex-level \"blendshapes\" anchored on the mesh, where each \"blendshape\" is a hash table, whose embeddings encode the information of a local radiance field around vertex \\( \\mathbf{v}_{ij} \\). Given a target expression to render, these hash table blendshapes are linearly summed via expression-dependent weights (Sec. 3.2), such that the merged embeddings encode the fine details specific to the target expression. Simultaneously, the coarse motion of the target expression is captured by the 3DMM vertex movement, which moves the attached hash tables accordingly, and hence the corresponding local radiance field." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.548, + 0.388, + 0.565 + ], + "angle": 0, + "content": "3.2. Merge Mesh-anchored Blendshapes" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.571, + 0.47, + 0.771 + ], + "angle": 0, + "content": "We obtain per-vertex blending weights by running a convolution neural network (CNN) on the 3DMM deformation represented in UV atlas space. Specifically, we calculate the vertex displacements with respect to the neutral face \\(\\mathbf{D}_i = \\mathcal{F}_{\\mathrm{3DMM}}(\\psi_i,\\pmb {\\theta}_i) - \\mathcal{F}_{\\mathrm{3DMM}}(\\mathbf{0},\\mathbf{0})\\) . The displacements are then warped into the UV space and fed into a U-Net to predict a weights map in \\(\\mathbb{R}^{\\mathrm{H_t}\\times \\mathrm{W_t}\\times \\mathrm{M}}\\) , where \\(\\mathrm{H_t}\\times \\mathrm{W_t}\\) is the UV resolution, and \\(M\\) is the number of hash table blendshapes on each vertex (pre-defined as 5 in our experiments). The weights map is then sampled back to 3DMM vertices, serving as the expression-dependent weights \\(\\{w_{ij}^{(m)}\\}_{M}\\) to take a weighted sum of the embeddings in the hash tables on each vertex, which produces the merged hash tables" + }, + { + "type": "equation", + "bbox": [ + 0.192, + 0.778, + 0.47, + 0.819 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathbf {H}} _ {i j} = \\sum_ {m = 1} ^ {M} w _ {i j} ^ {(m)} \\mathbf {H} _ {j} ^ {(m)}. \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.47, + 0.9 + ], + "angle": 0, + "content": "The U-Net also produces a UV feature map. We sample a per-vertex feature \\(\\mathbf{f}_{ij}\\) from it, similar to MonoAvatar [5]. We empirically found this benefits the geometry quality. The mesh-anchored hash tables \\(\\hat{\\mathbf{H}}_{ij}\\) and features \\(\\mathbf{f}_{ij}\\) are decoded into a neural radiance field as described in Sec. 3.3." + }, + { + "type": "image", + "bbox": [ + 0.513, + 0.089, + 0.885, + 0.317 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.327, + 0.892, + 0.384 + ], + "angle": 0, + "content": "Figure 3. Pipeline of Neural Radiance Field Decoding (Sec. 3.3). Our method only needs a very lightweight MLP for NeRF inference, where a two-hidden-layer MLP is used to predict the density and the color of each query point." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.406, + 0.787, + 0.423 + ], + "angle": 0, + "content": "3.3. Neural Radiance Field Decoding" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.429, + 0.892, + 0.565 + ], + "angle": 0, + "content": "Given the mesh-anchored hash tables \\(\\hat{\\mathbf{H}}_{ij}\\) and features \\(\\mathbf{f}_{ij}\\) described in Sec. 3.2, the final step is to decode them into a Neural Radiance Field (NeRF) to render the output image as shown in Fig. 3. The key idea is to associate a query point to neighbor vertices, and pull the embeddings from the attached hash tables via hash encoding [27]. Finally, we decode these pulled embeddings and the nearest per-vertex feature into the color and the density of this query point, followed by volumetric rendering to obtain the output image." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.566, + 0.893, + 0.779 + ], + "angle": 0, + "content": "For a 3D query point \\(\\mathbf{q}\\) when rendering a particular facial expression \\(i\\), we first obtain its \\(k\\)-nearest-neighbors, denoted as \\(\\{\\mathbf{v}_{ik}\\}_{k\\in \\mathcal{N}_{\\mathbf{q}}^{K}}\\), from the 3DMM vertices, with \\(k^*\\) denoting the nearest vertex index. For each neighbor vertex \\(\\mathbf{v}_{ik}\\) with an attached hash table \\(\\hat{\\mathbf{H}}_{ik}\\), we denote \\(\\mathbf{q}_{ik}\\) as the coordinates of \\(\\mathbf{q}\\) in the tangent space of \\(\\mathbf{v}_{ik}\\). We then use \\(\\mathbf{q}_{ik}\\) to query the hash table \\(\\hat{\\mathbf{H}}_{ik}\\) using a hash encoding function \\(\\mathcal{H}(\\cdot)\\) and obtain the embedding \\(\\mathbf{h}_{ik} = \\mathcal{H}(\\mathbf{q}_{ik};\\hat{\\mathbf{H}}_{ik})\\). To interpolate the embeddings from all \\(k\\)-nearest-neighbors, we use the weighted sum of the inverse distances \\(z_{k} = 1 / \\| \\mathbf{q}_{ik}\\|_{2}\\). Next, the summed embedding, together with the nearest per-vertex feature \\(\\mathbf{f}_{ik^*}\\) and the query point tangent coordinate \\(\\mathbf{q}_{ik^*}\\) of the nearest vertex, are fed into a two-hidden-layer MLP to predict the density and color as" + }, + { + "type": "equation", + "bbox": [ + 0.537, + 0.783, + 0.891, + 0.82 + ], + "angle": 0, + "content": "\\[\n\\overline {{\\mathbf {h}}} _ {i} = \\sum_ {k \\in \\mathcal {N} _ {\\mathbf {q}} ^ {K}} \\bar {w} _ {k} \\mathbf {h} _ {i k}, \\text {w h e r e} \\bar {w} _ {k} = \\frac {z _ {k}}{\\sum_ {k ^ {\\prime} \\in \\mathcal {N} _ {\\mathbf {q}} ^ {K}} z _ {k ^ {\\prime}}}, \\tag {2}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.508, + 0.824, + 0.885, + 0.849 + ], + "angle": 0, + "content": "\\[\n\\left[ \\mathbf {c} _ {i} (\\mathbf {q}, \\mathbf {d}), \\sigma_ {i} (\\mathbf {q}) \\right] = \\mathcal {F} _ {\\mathrm {M L P}} \\left(\\overline {{\\mathbf {h}}} _ {i}, \\mathbf {f} _ {i k ^ {*}}, \\operatorname {P E} \\left(\\mathbf {q} _ {i k ^ {*}}\\right), \\operatorname {P E} (\\mathbf {d})\\right),\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.892, + 0.901 + ], + "angle": 0, + "content": "where \\(\\overline{w}_k\\) is the normalized inverse-distance based weight, \\(\\mathbf{d}\\) denotes the camera view direction, \\(\\mathrm{PE}(\\cdot)\\) denotes positional encoding, \\(\\mathbf{c}_i\\) denotes color, and \\(\\sigma_{i}\\) denotes density." + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1978" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.138 + ], + "angle": 0, + "content": "Finally, we render the output pixel with the given camera ray \\(\\mathbf{r}\\) by volumetric rendering, where we reparameterize the query point with samples on the ray \\(\\mathbf{q} = \\mathbf{r}(t) = \\mathbf{o} + t\\mathbf{d}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.155, + 0.144, + 0.429, + 0.181 + ], + "angle": 0, + "content": "\\[\n\\mathbf {C} _ {i} (\\mathbf {r}) = \\int_ {t _ {n}} ^ {t _ {f}} T (t) \\sigma_ {i} (\\mathbf {r} (t)) \\mathbf {c} _ {i} (\\mathbf {r} (t), \\mathbf {d}) d t,\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.12, + 0.183, + 0.47, + 0.218 + ], + "angle": 0, + "content": "\\[\n\\mathrm {w h e r e} T (t) = \\exp \\left(- \\int_ {t _ {n}} ^ {t} \\sigma_ {i} (\\mathbf {r} (s)) d s\\right). \\qquad (3)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.222, + 0.471, + 0.418 + ], + "angle": 0, + "content": "Following prior works [5, 16], we also introduce a per-frame error-correction warping field during training to reduce misalignments due to the noise in 3DMM tracking and unmodeled per-frame contents such as hair movements. We feed the query point \\(\\mathbf{q}\\), together with a per-frame latent code \\(\\mathbf{e}_i\\), into an MLP \\(\\mathcal{F}_{\\mathcal{E}}(\\cdot)\\) to obtain a rigid transformation applied on the original query point, denoted as \\(\\mathbf{q}' = \\mathcal{T}_i(\\mathbf{q}) = \\mathcal{F}_{\\mathcal{E}}(\\mathbf{q},\\mathbf{e}_i)\\). The warped query point \\(\\mathbf{q}'\\) is then used to compute the density and color for volumetric rendering. Since the warping fields are overfit to corresponding training frames, we disable the warping field during testing similar to previous works [5, 16], and hence \\(\\mathcal{F}_{\\mathcal{E}}\\) does not affect rendering efficiency." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.425, + 0.418, + 0.44 + ], + "angle": 0, + "content": "3.4. Hierarchical \\(k\\)-nearest-neighbor Search" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.448, + 0.47, + 0.766 + ], + "angle": 0, + "content": "As described in Sec. 3.3, our method involves a \\( k \\)-nearest-neighbor (\\( k \\)-NN) search, which is computationally expensive and cannot be naively accelerated with pre-calculated structures (e.g., KD-Tree) due to the dynamically changing search pool (i.e., the 3DMM vertices driven by poses and expressions). To speed up the process, we propose a hierarchical \\( k \\)-NN search algorithm following a coarse-to-fine strategy. The key idea is to group nearby query points into a cluster as they likely share similar nearest neighbors. Specifically, we use a 3D grid with resolution 64 and treat all query points that fall in each voxel as a cluster. For each cluster, we first search \\( K' \\) (where \\( K' > K \\)) nearest neighbors of the voxel center from all 3DMM vertices. Then, for each query point, we search \\( K \\) nearest neighbors from the \\( K' \\) nearest neighbors of the corresponding cluster. In practice, for a 3DMM with the vertices number of \\( J = 1772 \\), we set \\( K' = 12 \\) and \\( K = 3 \\). Our experiments empirically show that, with a proper grid resolution, this design significantly improves the nearest neighbor search speed, and does not introduce noticeable rendering artifacts, even though the \\( k \\)-NNs may not be accurate on some of the query points." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.772, + 0.182, + 0.787 + ], + "angle": 0, + "content": "3.5. Training" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.795, + 0.468, + 0.901 + ], + "angle": 0, + "content": "Only monocular RGB videos are required to train our model. Three losses are used during the training process: (1) a photometric loss that minimizes the \\(l_{2}\\)-norm distance between the rendered and ground truth pixel colors over all camera rays \\(\\mathbf{r}\\) from all training frames \\(i\\). Formally, we have \\(\\mathcal{L}_{\\mathrm{rgb}} = \\sum_{i} \\sum_{\\mathbf{r}} \\| \\mathbf{C}_i(\\mathbf{r}) - \\mathbf{I}_i(\\mathbf{r}) \\|_2\\); (2) a elastic regularization loss \\(\\mathcal{L}_{\\mathrm{elastic}}\\) applied on the learned error-correction" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.152 + ], + "angle": 0, + "content": "warping field \\(\\mathcal{T}(\\mathbf{q})\\), which is introduced in Nerfies [29]; (3) a magnitude regularization loss to encourage small warping fields, which is defined as \\(\\mathcal{L}_{\\mathrm{mag}} = \\sum_{\\mathbf{q}}\\| \\mathbf{q} - \\mathcal{T}(\\mathbf{q})\\| _2^2\\). Finally, we combine all three loss terms:" + }, + { + "type": "equation", + "bbox": [ + 0.558, + 0.165, + 0.892, + 0.181 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\mathcal {L} _ {\\mathrm {r g b}} + \\lambda_ {\\text {e l a s t i c}} \\mathcal {L} _ {\\text {e l a s t i c}} + \\lambda_ {\\text {m a g}} \\mathcal {L} _ {\\text {m a g}}, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.192, + 0.892, + 0.283 + ], + "angle": 0, + "content": "where we set \\(\\lambda_{\\mathrm{elastic}} = 10^{-4}\\) at the beginning of the training and decay it to \\(10^{-5}\\) after 150k iterations, and set \\(\\lambda_{\\mathrm{mag}} = 10^{-2}\\). To warm start training, we replace the \\(l_{2}\\)-norm distance in the photometric loss \\(\\mathcal{L}_{\\mathrm{rgb}}\\) with the \\(l_{2}\\) distance for the first 10k iterations. Please refer to the supplementary for more details." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.296, + 0.633, + 0.314 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.322, + 0.892, + 0.488 + ], + "angle": 0, + "content": "In this section, we first introduce the data and metrics used for training and evaluation (Sec. 4.1). Then, we show that our avatar model achieves real-time rendering speed, while producing superior rendering quality on challenging expressions than recent efficient avatars [14, 41, 42] and being comparable to previous high-quality approaches [5] (Sec. 4.2). Finally, we provide ablation studies to justify the design choices and hyper-parameters of our avatar representation, and demonstrate the rendering speed improvements contributed from each of our newly proposed algorithmic components." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.497, + 0.671, + 0.512 + ], + "angle": 0, + "content": "4.1. Data and Metrics" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.52, + 0.892, + 0.732 + ], + "angle": 0, + "content": "Data. We use monocular RGB videos of multiple subjects (i.e., one video for one subject) to train and evaluate our method, and compare to prior state-of-the-art (SOTA) approaches. Our dataset consists of 10 videos in total, which are a mix of videos captured by us, as well as videos from prior works including PointAvatar [41], INSTA [42], and MonoAvatar [5]. We filter out the background of the videos with off-the-shelf segmentation [24] and matting [21] methods, then crop and resize the videos into a VGA resolution that preserves the original aspect ratio. We compute the camera and 3DMM parameters from the videos following the 3DMM fitting optimization used in INSTA [42]. We reserve a short clip from the end of each video as the testing frames, and use the rest frames for training." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Metrics. Following prior arts [13], we use PSNR, SSIM (higher is preferable), and LPIPS (lower is preferable) to measure the image quality. As observed by Zhang et al. [38], LPIPS is a more effective metric in judging the perceptual quality compared to PSNR and SSIM. When computing PSNR and SSIM, we weigh the mean squared error map and the SSIM map with a foreground mask (eroded and smoothed), in order to focus on non-empty areas and avoid the inaccurate foreground segmentation from dominating the metrics." + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1979" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.258 + ], + "angle": 0, + "content": "To evaluate the computational cost, we measure the rendering speed in frames-per-second (FPS) on a RTX3090Ti and compare across different approaches with their available implementations. We also estimate the number of FLOPs (floating-point operations) of all methods as the theoretical measurement for the rendering speed. When estimating FLOPs, we fix the contribution of the ray-marching part to 16 points sampled along each camera ray. This simplifies the estimate, since the ray-marching varies the number of FLOPs needed across cameras and scenes, and it applies to all the considered methods." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.271, + 0.353, + 0.287 + ], + "angle": 0, + "content": "4.2. Comparison to State-of-the-art" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.295, + 0.47, + 0.431 + ], + "angle": 0, + "content": "We compare our method with several prior works, including: NeRFBlendshape [13], INSTA [42], PointAvatar [41], and MonoAvatar [5]. NeRFBlendshape [13] and INSTA [42] adopt hash encoding [27] into head avatars, leading to efficient renderings. PointAvatar [41] leverages point clouds to represent the head avatar. MonoAvatar [5] is based on a 3DMM-anchored NeRF, and produces high-quality renderings but is slow in speed. We use the same camera and 3DMM parameters to train and test all methods." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.432, + 0.473, + 0.765 + ], + "angle": 0, + "content": "From Tab. 1, ours and INSTA [42] are the only 2 methods that can achieve real-time rendering (i.e., \\(\\geq 30\\) mean FPS). However, INSTA is quantitatively inferior than our method by a large margin, and gives obvious artifacts in Fig. 4, especially for challenging expressions. PointAvatar [41] has the potential to run in real-time with an optimized implementation thanks to its point cloud representation, but their renderings are overall blurrier than ours, leading to worse quantitative results. Although NeRFBlendshape [14] gives relatively good numbers in Tab. 1, it produces severe artifacts in dynamically changing regions (e.g., mouth and eyebrows in Fig. 4) for several median and large expressions and also gives more floaters, resulting in implausible animations. We highly suggest readers to see the supplementary videos for more comparisons. MonoAvatar [5] gives good rendering qualities and animations, but is one order of magnitudes slower and slightly blurrier on high frequency details such as forehead wrinkles, presumably because that the hash encoding in our model can better capture high frequency contents. Among these compared approaches, our method is the only one that achieves real-time rendering while being one of the best on image qualities." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.471, + 0.903 + ], + "angle": 0, + "content": "We also compare the theoretical FLOPs of all methods in Tab. 1, where our method requires the least computation mostly because of the smaller MLPs we use (e.g., 2 hidden layers of ours vs. \\(\\geq 5\\) hidden layers of others). Note that INSTA [42] is implemented with a highly optimized pure \\(\\mathrm{C}++\\) and CUDA codebase, while other methods use python (tensorflow/pytorch) with customized CUDA kernels. This implementation advantage of INSTA makes it running in a high FPS even with a relatively larger FLOPs." + }, + { + "type": "table", + "bbox": [ + 0.502, + 0.09, + 0.894, + 0.188 + ], + "angle": 0, + "content": "
LPIPSSSIMPSNRMean FPSGFLOPs
PointAvatar [41]0.1170.72821.125.0933
INSTA [42]0.1490.75822.1246.2266
NeRFBlendshape [14]0.1100.79322.7711.2223
MonoAvatar [5]0.1140.79822.740.52385
Ours0.1000.79522.7735.9113
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.189, + 0.892, + 0.232 + ], + "angle": 0, + "content": "Table 1. Quantitative comparisons with state-of-the-art approaches. Our method achieves rendering quality among the best, while supports real-time rendering with a \\(512 \\times 512\\) resolution." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.246, + 0.655, + 0.263 + ], + "angle": 0, + "content": "4.3. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.269, + 0.892, + 0.33 + ], + "angle": 0, + "content": "In this section we show the impact of the proposed design choices, in particular proving the importance of mesh-anchored hash table blendshapes and the proposed hierarchical \\(k\\)-NN search." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.348, + 0.837, + 0.365 + ], + "angle": 0, + "content": "4.3.1 Mesh-anchored Hash Table Blendshapes" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.372, + 0.892, + 0.417 + ], + "angle": 0, + "content": "We hereby investigate alternative design choices and different hyper-parameters to justify the necessity of our mesh-anchored hash table blendshapes." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.418, + 0.893, + 0.614 + ], + "angle": 0, + "content": "Static Hash + 3DMM Param. We first build a naive alternative approach to incorporate hash encoding into 3DMM-anchored NeRF, by attaching a single hash table to each vertex. For a query point, we concatenate the 3DMM pose and expression codes \\((\\theta_{i},\\psi_{i})\\) with the embedding pulled from the hash tables, and send them into the MLP. Note that there is no convolution running in UV space. As shown in Tab. 2 and Fig. 5, obvious rendering artifacts show up, and the rendering quality metrics drop significantly compared to our full model. This is presumably because the lightweight MLP, which is crucial for good efficiency, does not have enough capacity to process detailed expression-dependent information from compact 3DMM codes." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.614, + 0.895, + 0.765 + ], + "angle": 0, + "content": "Static Hash + UV CNN. We then increase the model capacity by adding back the UV CNN branch, but use only one hash table per-vertex, i.e., single hash table without blend-shape formulation. As shown in Tab. 2 and Fig. 5, the rendered images show less artifacts over the previous case, but still contain blurry textures and floaters compared to our full model. This demonstrates that the blendshape formulation of mesh-anchored hash tables are necessary in order to obtain good expression dependent local embeddings, leading to a superior rendering quality." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Number of Hash Table Blendshapes. Here we investigate how the number of hash table blendshapes per vertex influences the final rendering quality. As shown in Tab. 2, we can see that more blendshapes per vertex leads to higher rendering qualities, which saturates as the number of tables increases. In Fig. 5, increasing the number of blendshapes also gives better details especially for eyelids and ears. Although further adding more blendshapes may produce a better quality, we choose to use 5 blendshapes per-vertex as our" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1980" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.085, + 0.088, + 0.891, + 0.808 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.816, + 0.897, + 0.873 + ], + "angle": 0, + "content": "Figure 4. Comparisons on the rendering quality to previous state-of-the-art methods. From left to right, each column contains the images of: 1) PointAvatar [41], 2) INSTA [42], 3) NeRFBlendshape [14], 4) MonoAvatar [5], 5) Ours, 6) Ground Truth. Our method faithfully reconstructs the personalized expressions and high-frequency details, achieving one of the best rendering quality with real-time rendering speed." + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1981" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.084, + 0.089, + 0.892, + 0.213 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.216, + 0.895, + 0.3 + ], + "angle": 0, + "content": "Figure 5. Qualitative results of our full model and various alternative design choices to demonstrate the necessity of our mesh-anchored hash table blendshapes. Please refer to Sec. 4.3.1 for details of each alternative design choice. \"Static Hash + 3DMM Param\" gives unfaithful expressions and obvious artifacts around eyes and ears. \"Static Hash + UV CNN\" produces more faithful expressions, but still suffers from floaters. By utilizing the blendshape formulation of hash tables, \"3 Blendshapes\" eliminates most of the artifacts, but produces blurriness on the details of eyes and ears. \"Ours\" gives the best rendering quality by increasing the number of blendshapes to 5, leading to cleaner details of the eyelids and ear boundary." + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.308, + 0.892, + 0.412 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.414, + 0.892, + 0.456 + ], + "angle": 0, + "content": "Figure 6. Qualitative results of using different 3D grid resolutions during hierarchical \\(k\\)-NN search and without hierarchical \\(k\\)-NN search for ablation studies. \"16 grid resolution\" results in blocky artifacts around mouth and nose. \"32 grid resolution\" removes blocky artifacts, but produces floaters inside mouth. \"64 grid resolution\" produces a similar level of visual quality as \"w/o hierarchical \\(k\\)-NN search\"." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.47, + 0.47, + 0.501 + ], + "angle": 0, + "content": "final setting to maintain a relatively small model size and computation cost." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.52, + 0.314, + 0.535 + ], + "angle": 0, + "content": "4.3.2 Hierarchical \\(k\\)-NN Search" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.544, + 0.47, + 0.679 + ], + "angle": 0, + "content": "We evaluate our model with and without hierarchical \\( k \\)-NN search in terms of speed and quality. From the comparison over rendering speeds (w/: 35.9 FPS; w/o: 26.4 FPS), we can see that hierarchical \\( k \\)-NN search gives around \\( 36\\% \\) improvements on the frame rate, which is crucial for achieving real-time rendering. From Fig. 6, we empirically find that enabling hierarchical \\( k \\)-NN search will not lead to observable drops on the rendering quality, as long as a proper grid resolution is used (i.e., 64 in our case)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.68, + 0.47, + 0.773 + ], + "angle": 0, + "content": "We also investigate the affect of using different 3D grid resolutions during hierarchical \\(k\\)-NN search. As shown in Fig. 6, we observe more artifacts around the mouth region when using smaller 3D grid resolutions (i.e., 32 and 16). Therefore, we choose to use 64 resolution in our final setting, which is a good trade-off between quality and speed." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.785, + 0.192, + 0.801 + ], + "angle": 0, + "content": "5. Discussion" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.81, + 0.471, + 0.903 + ], + "angle": 0, + "content": "We present a high quality 3D neural volumetric head avatar that can be rendered efficiently, while only requires monocular RGB videos for construction. We propose the mesh-anchored hash table blendshapes as our avatar representation, which enable a significantly faster rendering speed by utilizing hash encoding and lightweight MLPs, while still" + }, + { + "type": "table", + "bbox": [ + 0.503, + 0.468, + 0.892, + 0.553 + ], + "angle": 0, + "content": "
LPIPSSSIMPSNR
Static Hash + 3DMM Param0.1250.76321.99
1 Blendshape (Static Hash + UV CNN)0.1150.78522.52
3 Blendshapes0.1040.79122.70
5 Blendshapes (Ours)0.1000.79522.77
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.557, + 0.892, + 0.599 + ], + "angle": 0, + "content": "Table 2. Quantitative results for the ablation study. Our full model (i.e., mesh-anchored hash table blendshapes) consistently outperforms alternative design choices discussed in Sec. 4.3.1." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.614, + 0.892, + 0.72 + ], + "angle": 0, + "content": "maintaining superior controllability to support realistic facial animations, and producing vivid expression-dependent details thank to the local blendshape formulation of hash tables. The experiments indicate that our approach runs in real-time at a \\(512 \\times 512\\) resolution, while giving a rendering quality comparable to state-of-the-art, with better challenging expressions than prior efficient approaches." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.893, + 0.889 + ], + "angle": 0, + "content": "As a limitation, we observe floaters under camera viewpoints and expressions that are far from the training distribution, which is a common issue in instant NGPs [27] based approaches. We also notice that performance around the mouth interior regions tends to be less stable because of the relatively poor tracking in these areas on the training data. Fortunately, the fast rendering could enable the possibility to adopt more expensive training strategies, such as regularization terms, adversarial loss, or joint face fitting refinement during the training, which could potentially mitigate these issues and further improve the rendering expressiveness and quality." + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1982" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.105 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.116, + 0.468, + 0.155 + ], + "angle": 0, + "content": "[1] MetaHuman - Unreal Engine. https://www.unrealengine.com/en-US/metahuman. Accessed: 2022-10-17. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.158, + 0.468, + 0.184 + ], + "angle": 0, + "content": "[2] VRChat. https://hello.vrchat.com. Accessed: 2023-02-05.1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.187, + 0.468, + 0.226 + ], + "angle": 0, + "content": "[3] XHolo Virtual Assistant. https://www.digalix.com/en/virtual-assistant-augmented-reality-xholo. Accessed: 2023-03-01. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.229, + 0.468, + 0.269 + ], + "angle": 0, + "content": "[4] ShahRukh Athar, Zexiang Xu, Kalyan Sunkavalli, Eli Shechtman, and Zhixin Shu. RigNeRF: Fully Controllable Neural 3D Portraits. pages 20364-20373, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.271, + 0.468, + 0.352 + ], + "angle": 0, + "content": "[5] Ziqian Bai, Feitong Tan, Zeng Huang, Kripasindhu Sarkar, Danhang Tang, Di Qiu, Abhimitra Meka, Ruofei Du, Mingsong Dou, Sergio Orts-Escalano, et al. Learning personalized high quality volumetric head avatars from monocular rgb videos. In CVPR, pages 16890-16900, 2023. 1, 2, 3, 4, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.355, + 0.468, + 0.424 + ], + "angle": 0, + "content": "[6] Thabo Beeler, Fabian Hahn, Derek Bradley, Bernd Bickel, Paul Beardsley, Craig Gotsman, Robert W Sumner, and Markus Gross. High-Quality Passive Facial Performance Capture Using Anchor Frames. In ACM SIGGRAPH 2011 Papers, pages 1-10. 2011. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.426, + 0.468, + 0.493 + ], + "angle": 0, + "content": "[7] Chen Cao, Tomas Simon, Jin Kyu Kim, Gabe Schwartz, Michael Zollhoefer, Shun-Suke Saito, Stephen Lombardi, Shih-En Wei, Danielle Belko, Shouu-I Yu, Yaser Sheikh, and Jason Saragih. Authentic volumetric avatars from a phone scan. ACM TOG, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.495, + 0.468, + 0.564 + ], + "angle": 0, + "content": "[8] Anpei Chen, Zexiang Xu, Andreas Geiger, Jingyi Yu, and Hao Su. Tensorf: Tensorial radiance fields. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXXII, pages 333-350. Springer, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.565, + 0.468, + 0.606 + ], + "angle": 0, + "content": "[9] Chuhan Chen, Matthew O'Toole, Gaurav Bharaj, and Pablo Garrido. Implicit neural head synthesis via controllable local deformation fields. In CVPR, pages 416-426, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.608, + 0.468, + 0.674 + ], + "angle": 0, + "content": "[10] Zhiqin Chen, Thomas Funkhouser, Peter Hedman, and Andrea Tagliasacchi. Mobilenerf: Exploiting the polygon rasterization pipeline for efficient neural field rendering on mobile architectures. arXiv preprint arXiv:2208.00277, 2022.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.677, + 0.468, + 0.746 + ], + "angle": 0, + "content": "[11] Michail Christos Doukas, Mohammad Rami Koujan, Viktoriaa Sharmanska, Anastasios Roussos, and Stefanos Zafeiriou. Head2head++: Deep Facial Attributes Re-Targeting. IEEE Transactions on Biometrics, Behavior, and Identity Science, 3(1):31-43, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.748, + 0.468, + 0.815 + ], + "angle": 0, + "content": "[12] Bernhard Egger, William AP Smith, Ayush Tewari, Stefanie Wuhrer, Michael Zollhoefer, Thabo Beeler, Florian Bernard, Timo Bolkart, Adam Kortylewski, Sami Romdhani, et al. 3D Morphable Face Models—Past, Present, and Future. 39(5): 1-38, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.817, + 0.468, + 0.871 + ], + "angle": 0, + "content": "[13] Guy Gafni, Justus Thies, Michael Zollhofer, and Matthias Nießner. Dynamic neural radiance fields for monocular 4d facial avatar reconstruction. In CVPR, pages 8649-8658, 2021. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.468, + 0.901 + ], + "angle": 0, + "content": "[14] Xuan Gao, Chenglai Zhong, Jun Xiang, Yang Hong, Yudong Guo, and Juyong Zhang. Reconstructing person" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.116, + 0.468, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.094, + 0.892, + 0.134 + ], + "angle": 0, + "content": "alized semantic facial nerf models from monocular video. ACM Transactions on Graphics (Proceedings of SIGGRAPH Asia), 41(6), 2022. 1, 2, 3, 4, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.137, + 0.892, + 0.246 + ], + "angle": 0, + "content": "[15] Kaiwen Guo, Peter Lincoln, Philip Davidson, Jay Busch, Xueming Yu, Matt Whalen, Geoff Harvey, Sergio Orts-Escolano, Rohit Pandey, Jason Dourgarian, Danhang Tang, Anastasia Tkach, Adarsh Kowdle, Emily Cooper, Ming-song Dou, Sean Fanello, Graham Fyffe, Christoph Rhemann, Jonathan Taylor, Paul Debevec, and Shahram Izadi. The Relightables: Volumetric Performance Capture of Humans With Realistic Relighting. ACM TOG, 2019. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.249, + 0.892, + 0.289 + ], + "angle": 0, + "content": "[16] Wei Jiang, Kwang Moo Yi, Golnoosh Samei, Oncel Tuzel, and Anurag Ranjan. NeuMan: Neural Human Radiance Field From a Single Video. 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.292, + 0.892, + 0.332 + ], + "angle": 0, + "content": "[17] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM TOG, 42(4):1-14, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.335, + 0.892, + 0.389 + ], + "angle": 0, + "content": "[18] Hyeongwoo Kim, Pablo Garrido, Ayush Tewari, Weipeng Xu, Justus Thies, Matthias Niessner, Patrick Pérez, Christian Richardt, Michael Zollhöfer, and Christian Theobalt. Deep Video Portraits. ACM TOG, 37(4):1-14, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.392, + 0.892, + 0.46 + ], + "angle": 0, + "content": "[19] Mohammad Rami Koujan, Michail Christos Doukas, Anastasios Roussos, and Stefanos Zafeiriou. Head2head: Video-Based Neural Head Synthesis. In 2020 15th IEEE International Conference on Automatic Face and Gesture Recognition (FG 2020), pages 16-23. IEEE, IEEE, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.463, + 0.892, + 0.516 + ], + "angle": 0, + "content": "[20] Tianye Li, Timo Bolkart, Michael. J. Black, Hao Li, and Javier Romero. Learning a Model of Facial Shape and Expression From 4D Scans. ACM TOG, 36(6):194:1-194:17, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.519, + 0.892, + 0.586 + ], + "angle": 0, + "content": "[21] Shanchuan Lin, Linjie Yang, Imran Saleemi, and Soumyadip Sengupta. Robust high-resolution video matting with temporal guidance. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 238-247, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.591, + 0.892, + 0.643 + ], + "angle": 0, + "content": "[22] Lingjie Liu, Jiatao Gu, Kyaw Zaw Lin, Tat-Seng Chua, and Christian Theobalt. Neural sparse voxel fields. Advances in Neural Information Processing Systems, 33:15651-15663, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.647, + 0.892, + 0.701 + ], + "angle": 0, + "content": "[23] Stephen Lombardi, Tomas Simon, Gabriel Schwartz, Michael Zollhoefer, Yaser Sheikh, and Jason Saragih. Mixture of volumetric primitives for efficient neural rendering. ACM TOG, 40(4):1-13, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.704, + 0.892, + 0.773 + ], + "angle": 0, + "content": "[24] Camillo Lugaresi, Jiuqiang Tang, Hadon Nash, Chris McClanahan, Esha Uboweja, Michael Hays, Fan Zhang, Chuoling Chang, Ming Guang Yong, Juhyun Lee, et al. Mediapipe: A framework for building perception pipelines. arXiv preprint arXiv:1906.08172, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.776, + 0.892, + 0.855 + ], + "angle": 0, + "content": "[25] Abhimitra Meka, Rohit Pandey, Christian Haene, Sergio Orts-Escalano, Peter Barnum, Philip David-Son, Daniel Erickson, Yinda Zhang, Jonathan Taylor, Sofien Bouaziz, et al. Deep Relightable Textures: Volumetric Performance Capture With Neural Rendering. ACM TOG, 39(6):1-21, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.859, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[26] Abhinitra Meka, Rohit Pandey, Christian Haene, Sergio Orts-Escolano, Peter Barnum, Philip Davidson, Daniel Erickson, Yinda Zhang, Jonathan Taylor, Sofien Bouaziz," + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.094, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.946, + 0.515, + 0.957 + ], + "angle": 0, + "content": "1983" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.092, + 0.47, + 0.161 + ], + "angle": 0, + "content": "Chloe Legendre, Wan-Chun Ma, Ryan Overbeck, Thabo Beeler, Paul Debevec, Shahram Izadi, Christian Theobalt, Christoph Rhemann, and Sean Fanello. Deep relightable textures - volumetric performance capture with neural rendering. 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.163, + 0.47, + 0.217 + ], + "angle": 0, + "content": "[27] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM Trans. Graph., 41(4):102:1-102:15, 2022. 1, 2, 3, 4, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.218, + 0.471, + 0.355 + ], + "angle": 0, + "content": "[28] Sergio Orts-Escolano, Christoph Rhemann, Sean Fanello, Wayne Chang, Adarsh Kowdle, Yury Degtyarev, David Kim, Philip Davidson, Sameh Khamis, Mingsong Dou, Vladimir Tankovich, Charles Loop, Qin Cai, Philip Chou, Sarah Mennicken, Julien Valentin, Vivek Pradeep, Shenlong Wang, Sing Bing Kang, Pushmeet Kohli, Yuliya Lutchyn, Cem Keskin, and Shahram Izadi. Holoportation: Virtual 3D Teleportation in Real-Time. In Proceedings of the 29th Annual Symposium on User Interface Software and Technology. ACM, 2016. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.357, + 0.47, + 0.412 + ], + "angle": 0, + "content": "[29] Keunhong Park, Utkarsh Sinha, Jonathan T Barron, Sofien Bouaziz, Dan B Goldman, Steven M Seitz, and Ricardo Martin-Brualla. Nerfies: Deformable Neural Radiance Fields. pages 5865-5874, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.413, + 0.47, + 0.481 + ], + "angle": 0, + "content": "[30] Christian Reiser, Songyou Peng, Yiyi Liao, and Andreas Geiger. Kilonerf: Speeding up neural radiance fields with thousands of tiny mlps. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14335-14345, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.483, + 0.469, + 0.523 + ], + "angle": 0, + "content": "[31] Cheng Sun, Min Sun, and Hwann-Tzong Chen. Direct voxel grid optimization: Super-fast convergence for radiance fields reconstruction. In CVPR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.525, + 0.469, + 0.594 + ], + "angle": 0, + "content": "[32] Ayush Tewari, Justus Thies, Ben Mildenhall, Pratul Srinivasan, Edgar Tretschk, Wang Yifan, Christoph Lassner, Vincent Sitzmann, Ricardo Martin-Brualla, Stephen Lombardi, et al. Advances in neural rendering. In Computer Graphics Forum, pages 703-735. Wiley Online Library, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.595, + 0.469, + 0.621 + ], + "angle": 0, + "content": "[33] Zach Waggoner. My Avatar, My Self: Identity in Video Role-Playing Games. McFarland, 2009. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.622, + 0.469, + 0.691 + ], + "angle": 0, + "content": "[34] Lizhen Wang, Xiaochen Zhao, Jingxiang Sun, Yuxiang Zhang, Hongwen Zhang, Tao Yu, and Yebin Liu. Stylea-vatar: Real-time photo-realistic portrait avatar from a single video. In ACM SIGGRAPH 2023 Conference Proceedings, pages 1-10, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.692, + 0.469, + 0.759 + ], + "angle": 0, + "content": "[35] Yuelang Xu, Lizhen Wang, Xiaochen Zhao, Hongwen Zhang, and Yebin Liu. Avatarmav: Fast 3d head avatar reconstruction using motion-aware neural voxels. In ACM SIGGRAPH 2023 Conference Proceedings, pages 1-10, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.762, + 0.469, + 0.831 + ], + "angle": 0, + "content": "[36] Yuelang Xu, Hongwen Zhang, Lizhen Wang, Xiaochen Zhao, Han Huang, Guojun Qi, and Yebin Liu. Latentavatar: Learning latent expression code for expressive neural head avatar. In ACM SIGGRAPH 2023 Conference Proceedings, pages 1-10, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.832, + 0.469, + 0.899 + ], + "angle": 0, + "content": "[37] Alex Yu, Ruilong Li, Matthew Tancik, Hao Li, Ren Ng, and Angjoo Kanazawa. Plenoctrees for real-time rendering of neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5752-5761, 2021. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.471, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.135 + ], + "angle": 0, + "content": "[38] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The Unreasonable Effectiveness of Deep Features As a Perceptual Metric. pages 586-595, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.136, + 0.892, + 0.19 + ], + "angle": 0, + "content": "[39] Xiaochen Zhao, Lizhen Wang, Jingxiang Sun, Hongwen Zhang, Jinli Suo, and Yebin Liu. Havatar: High-fidelity head avatar via facial model conditioned neural radiance field. ACM TOG, 43(1):1-16, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.192, + 0.892, + 0.247 + ], + "angle": 0, + "content": "[40] Yufeng Zheng, Victoria Fernández Abrevaya, Marcel C Bühler, Xu Chen, Michael J Black, and Otmar Hilliges. Im avatar: Implicit morphable head avatars from videos. In CVPR, pages 13545-13555, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.249, + 0.892, + 0.303 + ], + "angle": 0, + "content": "[41] Yufeng Zheng, Wang Yifan, Gordon Wetzstein, Michael J Black, and Otmar Hilliges. Pointavatar: Deformable point-based head avatars from videos. In CVPR, pages 21057-21067, 2023. 1, 2, 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.305, + 0.892, + 0.346 + ], + "angle": 0, + "content": "[42] Wojciech Zielonka, Timo Bolkart, and Justus Thies. Instant volumetric head avatars. In CVPR, pages 4574-4584, 2023. 1, 2, 3, 4, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.348, + 0.892, + 0.417 + ], + "angle": 0, + "content": "[43] Michael Zollhöfer, Justus Thies, Pablo Garrido, Derek Bradley, Thabo Beeler, Patrick Pérez, Marc Stamminger, Matthias Nießner, and Christian Theobalt. State of the Art on Monocular 3D Face Reconstruction, Tracking, and Applications. pages 523-550, 2018. 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.417 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "1984" + } + ] +] \ No newline at end of file diff --git a/2024/Efficient 3D Implicit Head Avatar with Mesh-anchored Hash Table Blendshapes/afaaa8c5-698e-4fcb-9bde-21aa2469de9a_origin.pdf b/2024/Efficient 3D Implicit Head Avatar with Mesh-anchored Hash Table Blendshapes/afaaa8c5-698e-4fcb-9bde-21aa2469de9a_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..b56d52d260f53539cadb64a6a65b9fe57b749f59 --- /dev/null +++ b/2024/Efficient 3D Implicit Head Avatar with Mesh-anchored Hash Table Blendshapes/afaaa8c5-698e-4fcb-9bde-21aa2469de9a_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4e79488622abbcb6384644b0e155f097cfa569512e56d52cb21eabc5b358d75f +size 5438299 diff --git a/2024/Efficient 3D Implicit Head Avatar with Mesh-anchored Hash Table Blendshapes/full.md b/2024/Efficient 3D Implicit Head Avatar with Mesh-anchored Hash Table Blendshapes/full.md new file mode 100644 index 0000000000000000000000000000000000000000..18c61002e89620205303109e85db48fb3a1587d3 --- /dev/null +++ b/2024/Efficient 3D Implicit Head Avatar with Mesh-anchored Hash Table Blendshapes/full.md @@ -0,0 +1,244 @@ +# Efficient 3D Implicit Head Avatar with Mesh-anchored Hash Table Blendshapes + +Ziqian Bai $^{1,2*}$ Feitong Tan $^{1}$ Sean Fanello $^{1}$ Rohit Pandey $^{1}$ Mingsong Dou $^{1}$ Shichen Liu $^{1}$ Ping Tan $^{3}$ Yinda Zhang $^{1}$ + +Google + +$^{2}$ Simon Fraser University + +3 The Hong Kong University of Science and Technology + +![](images/5711f29f7cf67a1111456bdda84716b1dcc8c7f8a5f2a528900c44276fb04533.jpg) +Figure 1. Visual Quality vs. Rendering Speed comparisons between our monocular 3D head avatars and prior state-of-the-art methods including NeRFBlendshape [14], INSTA [42], PointAvatar [41] and MonoAvatar [5]. Our approach achieves real-time rendering (i.e. $>30$ mean FPS) with a $512 \times 512$ resolution, while produces comparable visual quality comparing to prior SOTAs [5], and gives significantly better results on challenging expressions than prior efficient avatars [14, 41, 42]. Webpage augmentedperception.github.io/monoavatar-plus + +![](images/e6c022d54e15cc4bdf00f6a0bc6e4efafe906c34aa5218377f999d8304c8f86d.jpg) + +![](images/a377184159282ef49a05f86238bbba261f699c7ecce41884e41d486ee32b5422.jpg) + +![](images/fc3e527d29141d1109789cdb8834ff2002fe41ca2761b9caa4d9a5e003d0f9ca.jpg) + +![](images/35ac6f899365e031ab2e1b353b6adc15bc15f472d6724242609bb85f11ee6651.jpg) + +![](images/2f4e777735712c9a94daba4203b3452828327dd0ef6d2e3ea2d471af69b60c91.jpg) + +# Abstract + +3D head avatars built with neural implicit volumetric representations have achieved unprecedented levels of photorealism. However, the computational cost of these methods remains a significant barrier to their widespread adoption, particularly in real-time applications such as virtual reality and teleconferencing. While attempts have been made to develop fast neural rendering approaches for static scenes, these methods cannot be simply employed to support realistic facial expressions, such as in the case of a dynamic facial performance. To address these challenges, we propose a novel fast 3D neural implicit head avatar model that achieves real-time rendering while maintaining fine-grained controllability and high rendering quality. Our key idea lies in the introduction of local hash table blendshapes, which are learned and attached to the vertices of an underlying face parametric model. These per-vertex hash-tables are linearly merged with weights predicted via a CNN, resulting in expression dependent embeddings. Our novel representation enables efficient density and color predictions using a lightweight MLP, which is further accelerated by a hierarchical nearest neighbor search method. Extensive experiments show that our approach runs in real-time while achieving comparable rendering quality to state-of-the-arts and decent results on challenging expressions. + +# 1. Introduction + +The demand of high performing photo-realistic human avatars has dramatically increased with emerging VR/AR applications, e.g. VR gaming [2, 33], virtual assistant [3], tele-presence [28], and 3D videos [15, 26]. How to build efficient high quality avatars from monocular RGB videos becomes a promising direction due to the convenience of monocular data acquisition. While early works mostly adopt surface-based models for convenient controllability, recent methods (e.g. MonoAvatar [5]) leverage a sophisticated pipeline to build human avatars on neural radiance fields, which delivers vivid animations as well as significantly better rendering quality, especially over challenging parts such as hairs and glasses. On the downside, these approaches tend to be prohibitively slow, and most of the computation is consumed by the neural radiance field inference with large Multilayer Perceptrons (MLPs). + +Recently, fast approaches for neural radiance fields (e.g. hash encoding in Instant NGPs [27]) have been proposed, which are designed mostly for static scenes or pre-recorded temporal sequences. Despite their great success, it is not straightforward to extend these approaches for human avatars, which requires real-time rendering of dynamic facial performances when controlling the avatar. NeRFblend + +shape [14] address these issues by learning multiple feature hash tables, one for each face blendshape. These hash tables are linearly combined with blending weights to render the target facial expression via hash encoding. However, the expressiveness of the avatar is compromised by the global nature of the blending shapes, which cannot accurately capture vertex-level local deformations. On the other hand, INSTA [42] proposes to build the appearance model in a canonical space using a vanilla Instant NGPs [27] with expression codes as the additional MLP input to capture dynamic details, which is then transformed into target expression via a face parametric model (i.e. 3D morphable model (3DMM)). However, the lightweight MLP used in the vanilla Instant NGPs limits their model capacity, resulting in inferior animation quality, especially on extreme expressions. + +In this work, we propose a novel 3D neural avatar system that achieves efficient inference while maintaining fine-grained controllability and high fidelity quality. To achieve this, we introduce mesh-anchored hash table blendshapes, where we attach multiple, small hash tables to each of the 3DMM mesh vertices. These hash tables act as per-vertex "local blendshapes" (i.e., each "blendshape" is controlled by one local hash table) and influence only a local region. The mesh-anchored blendshapes are linearly merged with per-vertex weights predicted by a convolutional neural network in UV space from avatar driving signals, such as expression and head rotation. This results in expression-dependent hash table embeddings which offer several advantages over a global linear combination of blendshapes. Indeed by associating hash tables with individual vertices, we enhance the expressiveness of the model, allowing for more localized and nuanced facial expressions. This contrasts with global blendshapes, which apply uniform transformations across the entire face, limiting expressiveness. + +In more detail, our model starts from 3D query points, uses hash encoding [27] to gather the merged hash table embeddings from $k$ -nearest-neighbor vertices around the query point, and predicts the density and color via a small MLP. The hash encoding [27] allows us to use a very lightweight MLP to significantly reduce computation, leading to efficient inference. Additionally, the vertex-attached hash table blendshapes represent a 3DMM-anchored neural radiance field (NeRF), which can be easily controlled by the underlying 3DMM and produce high fidelity renderings as demonstrated by MonoAvatar [5]. To further accelerate our rendering speed, we propose a hierarchical $k$ -nearest-neighbor search method. + +Our contributions are summarized as follows. We propose a novel approach for high quality and efficient 3D neural implicit head avatars. At the core of our model, vertex-attached local hash table blendshapes are proposed to support efficient rendering, controllability, and capturing fine + +grained rendering details in dynamic facial performances. We also design a hierarchical querying solution to speed up the $k$ -nearest-neighbor search when pulling hash table embeddings from neighbor vertices. Extensive experiments on multiple datasets verify that we are able to speed up avatar rendering to real-time (i.e., average $\geq 30$ FPS to render a $512 \times 512$ video) while maintaining comparable rendering quality with the state-of-the-art high quality 3D avatar [5] and being largely superior on challenging expressions than existing efficient 3D avatars [14, 41, 42]. + +# 2. Related Work + +Constructing photorealistic digital humans has been a extensively researched topic. Here, we focus on discussing prior work on implicit monocular head avatars and efficient rendering. We refer readers to state-of-the-art surveys [12, 32, 43] for a comprehensive literature review. + +High Quality Head Avatar. Traditionally, high-quality head avatars have been achieved under expensive equipment configurations, such as camera arrays [6, 23, 28], depth sensors [7], and light stages [15, 25], or require laborious manual intervention [1]. Recent research efforts have focused on constructing high quality avatars from monocular RGB videos. One typical class of approaches [4, 13, 35, 40] use implicit 3D representations (i.e., neural radiance fields (NeRFs), implicit occupancy fields) to build the head avatar, which are parameterized by Multilayer perceptrons (MLPs). Although reasonable results are obtained, their rendering quality is still unsatisfactory especially for more challenging expressions. More recently, Bai et al. [5] proposed a head avatar based on 3DMM-anchored NeRFs with expression-dependent features produced by a convolution neural network in UV space. Chen et al. [9] designed local deformation fields to capture expression-dependent deformations applied on canonical NeRFs. Despite the impressive results, their methods didn't demonstrate real-time rendering capability due to expensive inference using large MLPs. In contrast, with our proposed mesh-anchored hash table blendshapes (Sec. 3), we achieve much faster rendering speed while maintaining high fidelity results. + +Efficient Neural Radiance Fields. There has been a plethora of work in recent years attempting to accelerate rendering with neural implicit representations for static objects and scenes. SNeRG [22], DVGO [31] and Plenoxels [37] propose to directly optimize voxel grids of (neural or SH) features for faster performance. However, their approach still requires a large memory footprint to store pervoxel features in 3D space. KiloNeRF [30] dramatically accelerates the original NeRF by representing the scene with thousands of tiny MLPs, however, this approach requires a complex training strategy. TensoRF [8] factorizes the feature grid into compact components, resulting in sig + +![](images/44e43a35c15cca6871b74a7c81e903c099dfc1401b4ff664a9d59249d41a42a1.jpg) +Figure 2. Overview of our pipeline. Our core avatar representation is Mesh-anchored Hash Table Blendshapes (Sec. 3.1), where multiple small hash tables are attached to each 3DMM vertex. During inference, our method starts from a displacement map encoding the facial expression, which is then fed into a U-Net to predict hash table weights and per-vertex features. The predicted weights are used to linearly combine the multiple hash tables attached on each 3DMM vertex (Sec. 3.2). During volumetric rendering (Sec. 3.3 and Fig. 3), for each query point, we search its $k$ -nearest-neighbor vertices, then pull embeddings from the merged hash tables and concatenate with the per-vertex feature to decode local density and color via a tiny MLP with two hidden layers. + +nificantly higher memory efficiency. Concurrently, Instant NGPs [27] utilizes multi-resolution hashing for efficient encoding, resulting in high compactness. MobileNeRF [10] propose to represent NeRF based on polygons, which allows leveraging a traditional polygon pipeline to enable their method to run in real-time on mobile devices. 3D Gaussian Splatting [17] represents a radiance field with 3D Gaussian point clouds, and leverages a point rasterization pipeline to enable fast rendering. These methods, however, cannot be easily extended to controllable dynamic contents. More recently, NeRFBlendshape [14] was proposed to handle controllable expressions by learning multiple hash tables for different global blendshapes and linearly combine them with expression codes. INSTA [42] transforms all expressions into a shared 3D canonical space, then adopts the vanilla Instant NGPs [27] conditioned on expression codes to model the head avatar. Despite the fast rendering speed achieved, their methods suffer from unsatisfactory rendering quality. PointAvatar [41] utilizes point clouds to represent the head avatar and uses large MLPs to predict the colors and motions of each point, leading to slow inference. Another type of works [11, 18, 19, 34, 36, 39] use 2D convolution neural networks to directly synthesize images (i.e. 2D neural rendering) from rasterized 3DMM meshes or low resolution feature maps generated by volumetric rendering. Despite their fast speed, the 2D CNNs may break the 3D consistency, leading to temporally unstable results especially for high frequency details. In contrast, our method simultaneously achieves controllability, high quality, and efficient rendering with a fully 3D representation. + +# 3. Method + +Given a monocular RGB video, our method learns a neural radiance field (NeRF) based head avatar, which can be ren- + +dered under any specified cameras, articulated poses (i.e., neck, jaw, and eyes) and facial expressions defined by a face parametric model (i.e. 3DMM). We use FLAME [20] as the parametric model in this work, but our method can be generalized to any other mesh-based parametric models. Fig. 2 shows the overview of our method. + +Our goal is to design a 3D neural implicit head avatar architecture that can simultaneously achieve high image quality, controllability, and computationally efficient rendering. To achieve this, we propose mesh-anchored hash table blendshapes (Sec. 3.1) as a novel avatar representation that can leverage both advantages from recent high-quality (i.e., 3DMM-anchored NeRF [5]) and efficient (i.e., hash encoding [27]) frameworks. More specifically, we propose to attach multiple small hash tables on each 3DMM vertex. These vertex-attached hash tables form a set of local "blendshapes", which will be linearly merged with predicted blending weights (Sec. 3.2), and decoded into a 3DMM-anchored NeRF to support fine-grained control and high fidelity rendering. During NeRF decoding on a query point (Sec. 3.3), we pull the embeddings from the linearly merged hash tables attached on $k$ -nearest-neighbor ( $k$ -NN) vertices from the 3DMM mesh. Using hash encoding allows us to use a very light weight MLP (only 2 hidden layers) to predict the final densities and colors, which is the key for efficient rendering. To further accelerate our approach to real-time, we leverage the fact that close query points likely share similar $k$ -NN vertices, and thus propose to group the query points into voxels and hierarchically search for $k$ -NN vertices (Sec. 3.4). Finally, our proposed avatar representation can be trained with only monocular RGB videos without any 3D scans or multi-view data (Sec. 3.5). + +# 3.1. Mesh-anchored Hash Table Blendshapes + +The core of our model is an avatar representation that can represent a 3DMM-anchored neural radiance field (NeRF) while allowing us to adopt the hash encoding [27] technique for acceleration. Recent approaches manage to adopt hash encoding into head avatars with different avatar representations (i.e., global blendshapes [14] and canonical NeRF [42]). In contrast, our solution is built upon the most recent 3DMM-anchored NeRF [5], which is superior for high quality renderings as demonstrated in our experiments (Sec. 4.2). + +We propose mesh-anchored hash table blendshapes as the new avatar representation. Given a target expression $i$ with 3DMM pose code $\theta_{i}$ and expression code $\psi_{i}$ , we first get the deformed 3DMM mesh $\mathbf{V}_{i} = \mathcal{F}_{3\mathrm{DMM}}(\psi_{i},\theta_{i})$ with $J$ vertices. For the $j$ -th 3DMM mesh vertex $\mathbf{v}_{ij}$ , we attach $M$ small hash tables $\{\mathbf{H}_j^{(m)}\}_{M}$ on it, where each hash table has multiple resolutions following instant NGPs [27]. Intuitively, these hash tables form a set of vertex-level "blendshapes" anchored on the mesh, where each "blendshape" is a hash table, whose embeddings encode the information of a local radiance field around vertex $\mathbf{v}_{ij}$ . Given a target expression to render, these hash table blendshapes are linearly summed via expression-dependent weights (Sec. 3.2), such that the merged embeddings encode the fine details specific to the target expression. Simultaneously, the coarse motion of the target expression is captured by the 3DMM vertex movement, which moves the attached hash tables accordingly, and hence the corresponding local radiance field. + +# 3.2. Merge Mesh-anchored Blendshapes + +We obtain per-vertex blending weights by running a convolution neural network (CNN) on the 3DMM deformation represented in UV atlas space. Specifically, we calculate the vertex displacements with respect to the neutral face $\mathbf{D}_i = \mathcal{F}_{\mathrm{3DMM}}(\psi_i,\pmb {\theta}_i) - \mathcal{F}_{\mathrm{3DMM}}(\mathbf{0},\mathbf{0})$ . The displacements are then warped into the UV space and fed into a U-Net to predict a weights map in $\mathbb{R}^{\mathrm{H_t}\times \mathrm{W_t}\times \mathrm{M}}$ , where $\mathrm{H_t}\times \mathrm{W_t}$ is the UV resolution, and $M$ is the number of hash table blendshapes on each vertex (pre-defined as 5 in our experiments). The weights map is then sampled back to 3DMM vertices, serving as the expression-dependent weights $\{w_{ij}^{(m)}\}_{M}$ to take a weighted sum of the embeddings in the hash tables on each vertex, which produces the merged hash tables + +$$ +\hat {\mathbf {H}} _ {i j} = \sum_ {m = 1} ^ {M} w _ {i j} ^ {(m)} \mathbf {H} _ {j} ^ {(m)}. \tag {1} +$$ + +The U-Net also produces a UV feature map. We sample a per-vertex feature $\mathbf{f}_{ij}$ from it, similar to MonoAvatar [5]. We empirically found this benefits the geometry quality. The mesh-anchored hash tables $\hat{\mathbf{H}}_{ij}$ and features $\mathbf{f}_{ij}$ are decoded into a neural radiance field as described in Sec. 3.3. + +![](images/72db2b65c8180081479f4da2a586cfa6c43a25b8ea455a2da735e2d2c62df568.jpg) +Figure 3. Pipeline of Neural Radiance Field Decoding (Sec. 3.3). Our method only needs a very lightweight MLP for NeRF inference, where a two-hidden-layer MLP is used to predict the density and the color of each query point. + +# 3.3. Neural Radiance Field Decoding + +Given the mesh-anchored hash tables $\hat{\mathbf{H}}_{ij}$ and features $\mathbf{f}_{ij}$ described in Sec. 3.2, the final step is to decode them into a Neural Radiance Field (NeRF) to render the output image as shown in Fig. 3. The key idea is to associate a query point to neighbor vertices, and pull the embeddings from the attached hash tables via hash encoding [27]. Finally, we decode these pulled embeddings and the nearest per-vertex feature into the color and the density of this query point, followed by volumetric rendering to obtain the output image. + +For a 3D query point $\mathbf{q}$ when rendering a particular facial expression $i$ , we first obtain its $k$ -nearest-neighbors, denoted as $\{\mathbf{v}_{ik}\}_{k\in \mathcal{N}_{\mathbf{q}}^{K}}$ , from the 3DMM vertices, with $k^*$ denoting the nearest vertex index. For each neighbor vertex $\mathbf{v}_{ik}$ with an attached hash table $\hat{\mathbf{H}}_{ik}$ , we denote $\mathbf{q}_{ik}$ as the coordinates of $\mathbf{q}$ in the tangent space of $\mathbf{v}_{ik}$ . We then use $\mathbf{q}_{ik}$ to query the hash table $\hat{\mathbf{H}}_{ik}$ using a hash encoding function $\mathcal{H}(\cdot)$ and obtain the embedding $\mathbf{h}_{ik} = \mathcal{H}(\mathbf{q}_{ik};\hat{\mathbf{H}}_{ik})$ . To interpolate the embeddings from all $k$ -nearest-neighbors, we use the weighted sum of the inverse distances $z_{k} = 1 / \| \mathbf{q}_{ik}\|_{2}$ . Next, the summed embedding, together with the nearest per-vertex feature $\mathbf{f}_{ik^*}$ and the query point tangent coordinate $\mathbf{q}_{ik^*}$ of the nearest vertex, are fed into a two-hidden-layer MLP to predict the density and color as + +$$ +\overline {{\mathbf {h}}} _ {i} = \sum_ {k \in \mathcal {N} _ {\mathbf {q}} ^ {K}} \bar {w} _ {k} \mathbf {h} _ {i k}, \text {w h e r e} \bar {w} _ {k} = \frac {z _ {k}}{\sum_ {k ^ {\prime} \in \mathcal {N} _ {\mathbf {q}} ^ {K}} z _ {k ^ {\prime}}}, \tag {2} +$$ + +$$ +\left[ \mathbf {c} _ {i} (\mathbf {q}, \mathbf {d}), \sigma_ {i} (\mathbf {q}) \right] = \mathcal {F} _ {\mathrm {M L P}} \left(\overline {{\mathbf {h}}} _ {i}, \mathbf {f} _ {i k ^ {*}}, \operatorname {P E} \left(\mathbf {q} _ {i k ^ {*}}\right), \operatorname {P E} (\mathbf {d})\right), +$$ + +where $\overline{w}_k$ is the normalized inverse-distance based weight, $\mathbf{d}$ denotes the camera view direction, $\mathrm{PE}(\cdot)$ denotes positional encoding, $\mathbf{c}_i$ denotes color, and $\sigma_{i}$ denotes density. + +Finally, we render the output pixel with the given camera ray $\mathbf{r}$ by volumetric rendering, where we reparameterize the query point with samples on the ray $\mathbf{q} = \mathbf{r}(t) = \mathbf{o} + t\mathbf{d}$ : + +$$ +\mathbf {C} _ {i} (\mathbf {r}) = \int_ {t _ {n}} ^ {t _ {f}} T (t) \sigma_ {i} (\mathbf {r} (t)) \mathbf {c} _ {i} (\mathbf {r} (t), \mathbf {d}) d t, +$$ + +$$ +\mathrm {w h e r e} T (t) = \exp \left(- \int_ {t _ {n}} ^ {t} \sigma_ {i} (\mathbf {r} (s)) d s\right). \qquad (3) +$$ + +Following prior works [5, 16], we also introduce a per-frame error-correction warping field during training to reduce misalignments due to the noise in 3DMM tracking and unmodeled per-frame contents such as hair movements. We feed the query point $\mathbf{q}$ , together with a per-frame latent code $\mathbf{e}_i$ , into an MLP $\mathcal{F}_{\mathcal{E}}(\cdot)$ to obtain a rigid transformation applied on the original query point, denoted as $\mathbf{q}' = \mathcal{T}_i(\mathbf{q}) = \mathcal{F}_{\mathcal{E}}(\mathbf{q},\mathbf{e}_i)$ . The warped query point $\mathbf{q}'$ is then used to compute the density and color for volumetric rendering. Since the warping fields are overfit to corresponding training frames, we disable the warping field during testing similar to previous works [5, 16], and hence $\mathcal{F}_{\mathcal{E}}$ does not affect rendering efficiency. + +# 3.4. Hierarchical $k$ -nearest-neighbor Search + +As described in Sec. 3.3, our method involves a $k$ -nearest-neighbor ( $k$ -NN) search, which is computationally expensive and cannot be naively accelerated with pre-calculated structures (e.g., KD-Tree) due to the dynamically changing search pool (i.e., the 3DMM vertices driven by poses and expressions). To speed up the process, we propose a hierarchical $k$ -NN search algorithm following a coarse-to-fine strategy. The key idea is to group nearby query points into a cluster as they likely share similar nearest neighbors. Specifically, we use a 3D grid with resolution 64 and treat all query points that fall in each voxel as a cluster. For each cluster, we first search $K'$ (where $K' > K$ ) nearest neighbors of the voxel center from all 3DMM vertices. Then, for each query point, we search $K$ nearest neighbors from the $K'$ nearest neighbors of the corresponding cluster. In practice, for a 3DMM with the vertices number of $J = 1772$ , we set $K' = 12$ and $K = 3$ . Our experiments empirically show that, with a proper grid resolution, this design significantly improves the nearest neighbor search speed, and does not introduce noticeable rendering artifacts, even though the $k$ -NNs may not be accurate on some of the query points. + +# 3.5. Training + +Only monocular RGB videos are required to train our model. Three losses are used during the training process: (1) a photometric loss that minimizes the $l_{2}$ -norm distance between the rendered and ground truth pixel colors over all camera rays $\mathbf{r}$ from all training frames $i$ . Formally, we have $\mathcal{L}_{\mathrm{rgb}} = \sum_{i} \sum_{\mathbf{r}} \| \mathbf{C}_i(\mathbf{r}) - \mathbf{I}_i(\mathbf{r}) \|_2$ ; (2) a elastic regularization loss $\mathcal{L}_{\mathrm{elastic}}$ applied on the learned error-correction + +warping field $\mathcal{T}(\mathbf{q})$ , which is introduced in Nerfies [29]; (3) a magnitude regularization loss to encourage small warping fields, which is defined as $\mathcal{L}_{\mathrm{mag}} = \sum_{\mathbf{q}}\| \mathbf{q} - \mathcal{T}(\mathbf{q})\| _2^2$ . Finally, we combine all three loss terms: + +$$ +\mathcal {L} = \mathcal {L} _ {\mathrm {r g b}} + \lambda_ {\text {e l a s t i c}} \mathcal {L} _ {\text {e l a s t i c}} + \lambda_ {\text {m a g}} \mathcal {L} _ {\text {m a g}}, \tag {4} +$$ + +where we set $\lambda_{\mathrm{elastic}} = 10^{-4}$ at the beginning of the training and decay it to $10^{-5}$ after 150k iterations, and set $\lambda_{\mathrm{mag}} = 10^{-2}$ . To warm start training, we replace the $l_{2}$ -norm distance in the photometric loss $\mathcal{L}_{\mathrm{rgb}}$ with the $l_{2}$ distance for the first 10k iterations. Please refer to the supplementary for more details. + +# 4. Experiments + +In this section, we first introduce the data and metrics used for training and evaluation (Sec. 4.1). Then, we show that our avatar model achieves real-time rendering speed, while producing superior rendering quality on challenging expressions than recent efficient avatars [14, 41, 42] and being comparable to previous high-quality approaches [5] (Sec. 4.2). Finally, we provide ablation studies to justify the design choices and hyper-parameters of our avatar representation, and demonstrate the rendering speed improvements contributed from each of our newly proposed algorithmic components. + +# 4.1. Data and Metrics + +Data. We use monocular RGB videos of multiple subjects (i.e., one video for one subject) to train and evaluate our method, and compare to prior state-of-the-art (SOTA) approaches. Our dataset consists of 10 videos in total, which are a mix of videos captured by us, as well as videos from prior works including PointAvatar [41], INSTA [42], and MonoAvatar [5]. We filter out the background of the videos with off-the-shelf segmentation [24] and matting [21] methods, then crop and resize the videos into a VGA resolution that preserves the original aspect ratio. We compute the camera and 3DMM parameters from the videos following the 3DMM fitting optimization used in INSTA [42]. We reserve a short clip from the end of each video as the testing frames, and use the rest frames for training. + +Metrics. Following prior arts [13], we use PSNR, SSIM (higher is preferable), and LPIPS (lower is preferable) to measure the image quality. As observed by Zhang et al. [38], LPIPS is a more effective metric in judging the perceptual quality compared to PSNR and SSIM. When computing PSNR and SSIM, we weigh the mean squared error map and the SSIM map with a foreground mask (eroded and smoothed), in order to focus on non-empty areas and avoid the inaccurate foreground segmentation from dominating the metrics. + +To evaluate the computational cost, we measure the rendering speed in frames-per-second (FPS) on a RTX3090Ti and compare across different approaches with their available implementations. We also estimate the number of FLOPs (floating-point operations) of all methods as the theoretical measurement for the rendering speed. When estimating FLOPs, we fix the contribution of the ray-marching part to 16 points sampled along each camera ray. This simplifies the estimate, since the ray-marching varies the number of FLOPs needed across cameras and scenes, and it applies to all the considered methods. + +# 4.2. Comparison to State-of-the-art + +We compare our method with several prior works, including: NeRFBlendshape [13], INSTA [42], PointAvatar [41], and MonoAvatar [5]. NeRFBlendshape [13] and INSTA [42] adopt hash encoding [27] into head avatars, leading to efficient renderings. PointAvatar [41] leverages point clouds to represent the head avatar. MonoAvatar [5] is based on a 3DMM-anchored NeRF, and produces high-quality renderings but is slow in speed. We use the same camera and 3DMM parameters to train and test all methods. + +From Tab. 1, ours and INSTA [42] are the only 2 methods that can achieve real-time rendering (i.e., $\geq 30$ mean FPS). However, INSTA is quantitatively inferior than our method by a large margin, and gives obvious artifacts in Fig. 4, especially for challenging expressions. PointAvatar [41] has the potential to run in real-time with an optimized implementation thanks to its point cloud representation, but their renderings are overall blurrier than ours, leading to worse quantitative results. Although NeRFBlendshape [14] gives relatively good numbers in Tab. 1, it produces severe artifacts in dynamically changing regions (e.g., mouth and eyebrows in Fig. 4) for several median and large expressions and also gives more floaters, resulting in implausible animations. We highly suggest readers to see the supplementary videos for more comparisons. MonoAvatar [5] gives good rendering qualities and animations, but is one order of magnitudes slower and slightly blurrier on high frequency details such as forehead wrinkles, presumably because that the hash encoding in our model can better capture high frequency contents. Among these compared approaches, our method is the only one that achieves real-time rendering while being one of the best on image qualities. + +We also compare the theoretical FLOPs of all methods in Tab. 1, where our method requires the least computation mostly because of the smaller MLPs we use (e.g., 2 hidden layers of ours vs. $\geq 5$ hidden layers of others). Note that INSTA [42] is implemented with a highly optimized pure $\mathrm{C}++$ and CUDA codebase, while other methods use python (tensorflow/pytorch) with customized CUDA kernels. This implementation advantage of INSTA makes it running in a high FPS even with a relatively larger FLOPs. + +
LPIPSSSIMPSNRMean FPSGFLOPs
PointAvatar [41]0.1170.72821.125.0933
INSTA [42]0.1490.75822.1246.2266
NeRFBlendshape [14]0.1100.79322.7711.2223
MonoAvatar [5]0.1140.79822.740.52385
Ours0.1000.79522.7735.9113
+ +Table 1. Quantitative comparisons with state-of-the-art approaches. Our method achieves rendering quality among the best, while supports real-time rendering with a $512 \times 512$ resolution. + +# 4.3. Ablation Study + +In this section we show the impact of the proposed design choices, in particular proving the importance of mesh-anchored hash table blendshapes and the proposed hierarchical $k$ -NN search. + +# 4.3.1 Mesh-anchored Hash Table Blendshapes + +We hereby investigate alternative design choices and different hyper-parameters to justify the necessity of our mesh-anchored hash table blendshapes. + +Static Hash + 3DMM Param. We first build a naive alternative approach to incorporate hash encoding into 3DMM-anchored NeRF, by attaching a single hash table to each vertex. For a query point, we concatenate the 3DMM pose and expression codes $(\theta_{i},\psi_{i})$ with the embedding pulled from the hash tables, and send them into the MLP. Note that there is no convolution running in UV space. As shown in Tab. 2 and Fig. 5, obvious rendering artifacts show up, and the rendering quality metrics drop significantly compared to our full model. This is presumably because the lightweight MLP, which is crucial for good efficiency, does not have enough capacity to process detailed expression-dependent information from compact 3DMM codes. + +Static Hash + UV CNN. We then increase the model capacity by adding back the UV CNN branch, but use only one hash table per-vertex, i.e., single hash table without blend-shape formulation. As shown in Tab. 2 and Fig. 5, the rendered images show less artifacts over the previous case, but still contain blurry textures and floaters compared to our full model. This demonstrates that the blendshape formulation of mesh-anchored hash tables are necessary in order to obtain good expression dependent local embeddings, leading to a superior rendering quality. + +Number of Hash Table Blendshapes. Here we investigate how the number of hash table blendshapes per vertex influences the final rendering quality. As shown in Tab. 2, we can see that more blendshapes per vertex leads to higher rendering qualities, which saturates as the number of tables increases. In Fig. 5, increasing the number of blendshapes also gives better details especially for eyelids and ears. Although further adding more blendshapes may produce a better quality, we choose to use 5 blendshapes per-vertex as our + +![](images/7ccb91c29e5b751006d3df318f43de24172b32ed48cc011e3641c1362821c9fe.jpg) +Figure 4. Comparisons on the rendering quality to previous state-of-the-art methods. From left to right, each column contains the images of: 1) PointAvatar [41], 2) INSTA [42], 3) NeRFBlendshape [14], 4) MonoAvatar [5], 5) Ours, 6) Ground Truth. Our method faithfully reconstructs the personalized expressions and high-frequency details, achieving one of the best rendering quality with real-time rendering speed. + +![](images/75c55b298d7d131cb7a338c066782833fa72f3a48110563b9632b328914526b0.jpg) +Figure 5. Qualitative results of our full model and various alternative design choices to demonstrate the necessity of our mesh-anchored hash table blendshapes. Please refer to Sec. 4.3.1 for details of each alternative design choice. "Static Hash + 3DMM Param" gives unfaithful expressions and obvious artifacts around eyes and ears. "Static Hash + UV CNN" produces more faithful expressions, but still suffers from floaters. By utilizing the blendshape formulation of hash tables, "3 Blendshapes" eliminates most of the artifacts, but produces blurriness on the details of eyes and ears. "Ours" gives the best rendering quality by increasing the number of blendshapes to 5, leading to cleaner details of the eyelids and ear boundary. + +![](images/b74324f3e847ea6e233f71207889aff1afaf2aa8018d81e238365f14eab65d8e.jpg) +Figure 6. Qualitative results of using different 3D grid resolutions during hierarchical $k$ -NN search and without hierarchical $k$ -NN search for ablation studies. "16 grid resolution" results in blocky artifacts around mouth and nose. "32 grid resolution" removes blocky artifacts, but produces floaters inside mouth. "64 grid resolution" produces a similar level of visual quality as "w/o hierarchical $k$ -NN search". + +final setting to maintain a relatively small model size and computation cost. + +# 4.3.2 Hierarchical $k$ -NN Search + +We evaluate our model with and without hierarchical $k$ -NN search in terms of speed and quality. From the comparison over rendering speeds (w/: 35.9 FPS; w/o: 26.4 FPS), we can see that hierarchical $k$ -NN search gives around $36\%$ improvements on the frame rate, which is crucial for achieving real-time rendering. From Fig. 6, we empirically find that enabling hierarchical $k$ -NN search will not lead to observable drops on the rendering quality, as long as a proper grid resolution is used (i.e., 64 in our case). + +We also investigate the affect of using different 3D grid resolutions during hierarchical $k$ -NN search. As shown in Fig. 6, we observe more artifacts around the mouth region when using smaller 3D grid resolutions (i.e., 32 and 16). Therefore, we choose to use 64 resolution in our final setting, which is a good trade-off between quality and speed. + +# 5. Discussion + +We present a high quality 3D neural volumetric head avatar that can be rendered efficiently, while only requires monocular RGB videos for construction. We propose the mesh-anchored hash table blendshapes as our avatar representation, which enable a significantly faster rendering speed by utilizing hash encoding and lightweight MLPs, while still + +
LPIPSSSIMPSNR
Static Hash + 3DMM Param0.1250.76321.99
1 Blendshape (Static Hash + UV CNN)0.1150.78522.52
3 Blendshapes0.1040.79122.70
5 Blendshapes (Ours)0.1000.79522.77
+ +Table 2. Quantitative results for the ablation study. Our full model (i.e., mesh-anchored hash table blendshapes) consistently outperforms alternative design choices discussed in Sec. 4.3.1. + +maintaining superior controllability to support realistic facial animations, and producing vivid expression-dependent details thank to the local blendshape formulation of hash tables. The experiments indicate that our approach runs in real-time at a $512 \times 512$ resolution, while giving a rendering quality comparable to state-of-the-art, with better challenging expressions than prior efficient approaches. + +As a limitation, we observe floaters under camera viewpoints and expressions that are far from the training distribution, which is a common issue in instant NGPs [27] based approaches. We also notice that performance around the mouth interior regions tends to be less stable because of the relatively poor tracking in these areas on the training data. Fortunately, the fast rendering could enable the possibility to adopt more expensive training strategies, such as regularization terms, adversarial loss, or joint face fitting refinement during the training, which could potentially mitigate these issues and further improve the rendering expressiveness and quality. + +# References + +[1] MetaHuman - Unreal Engine. https://www.unrealengine.com/en-US/metahuman. Accessed: 2022-10-17. 2 +[2] VRChat. https://hello.vrchat.com. Accessed: 2023-02-05.1 +[3] XHolo Virtual Assistant. https://www.digalix.com/en/virtual-assistant-augmented-reality-xholo. Accessed: 2023-03-01. 1 +[4] ShahRukh Athar, Zexiang Xu, Kalyan Sunkavalli, Eli Shechtman, and Zhixin Shu. RigNeRF: Fully Controllable Neural 3D Portraits. pages 20364-20373, 2022. 2 +[5] Ziqian Bai, Feitong Tan, Zeng Huang, Kripasindhu Sarkar, Danhang Tang, Di Qiu, Abhimitra Meka, Ruofei Du, Mingsong Dou, Sergio Orts-Escalano, et al. Learning personalized high quality volumetric head avatars from monocular rgb videos. In CVPR, pages 16890-16900, 2023. 1, 2, 3, 4, 5, 6, 7 +[6] Thabo Beeler, Fabian Hahn, Derek Bradley, Bernd Bickel, Paul Beardsley, Craig Gotsman, Robert W Sumner, and Markus Gross. High-Quality Passive Facial Performance Capture Using Anchor Frames. In ACM SIGGRAPH 2011 Papers, pages 1-10. 2011. 2 +[7] Chen Cao, Tomas Simon, Jin Kyu Kim, Gabe Schwartz, Michael Zollhoefer, Shun-Suke Saito, Stephen Lombardi, Shih-En Wei, Danielle Belko, Shouu-I Yu, Yaser Sheikh, and Jason Saragih. Authentic volumetric avatars from a phone scan. ACM TOG, 2022. 2 +[8] Anpei Chen, Zexiang Xu, Andreas Geiger, Jingyi Yu, and Hao Su. Tensorf: Tensorial radiance fields. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXXII, pages 333-350. Springer, 2022. 2 +[9] Chuhan Chen, Matthew O'Toole, Gaurav Bharaj, and Pablo Garrido. Implicit neural head synthesis via controllable local deformation fields. In CVPR, pages 416-426, 2023. 2 +[10] Zhiqin Chen, Thomas Funkhouser, Peter Hedman, and Andrea Tagliasacchi. Mobilenerf: Exploiting the polygon rasterization pipeline for efficient neural field rendering on mobile architectures. arXiv preprint arXiv:2208.00277, 2022.3 +[11] Michail Christos Doukas, Mohammad Rami Koujan, Viktoriaa Sharmanska, Anastasios Roussos, and Stefanos Zafeiriou. Head2head++: Deep Facial Attributes Re-Targeting. IEEE Transactions on Biometrics, Behavior, and Identity Science, 3(1):31-43, 2021. 3 +[12] Bernhard Egger, William AP Smith, Ayush Tewari, Stefanie Wuhrer, Michael Zollhoefer, Thabo Beeler, Florian Bernard, Timo Bolkart, Adam Kortylewski, Sami Romdhani, et al. 3D Morphable Face Models—Past, Present, and Future. 39(5): 1-38, 2020. 2 +[13] Guy Gafni, Justus Thies, Michael Zollhofer, and Matthias Nießner. Dynamic neural radiance fields for monocular 4d facial avatar reconstruction. In CVPR, pages 8649-8658, 2021. 2, 5, 6 +[14] Xuan Gao, Chenglai Zhong, Jun Xiang, Yang Hong, Yudong Guo, and Juyong Zhang. Reconstructing person + +alized semantic facial nerf models from monocular video. ACM Transactions on Graphics (Proceedings of SIGGRAPH Asia), 41(6), 2022. 1, 2, 3, 4, 5, 6, 7 +[15] Kaiwen Guo, Peter Lincoln, Philip Davidson, Jay Busch, Xueming Yu, Matt Whalen, Geoff Harvey, Sergio Orts-Escolano, Rohit Pandey, Jason Dourgarian, Danhang Tang, Anastasia Tkach, Adarsh Kowdle, Emily Cooper, Ming-song Dou, Sean Fanello, Graham Fyffe, Christoph Rhemann, Jonathan Taylor, Paul Debevec, and Shahram Izadi. The Relightables: Volumetric Performance Capture of Humans With Realistic Relighting. ACM TOG, 2019. 1, 2 +[16] Wei Jiang, Kwang Moo Yi, Golnoosh Samei, Oncel Tuzel, and Anurag Ranjan. NeuMan: Neural Human Radiance Field From a Single Video. 2022. 5 +[17] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM TOG, 42(4):1-14, 2023. 3 +[18] Hyeongwoo Kim, Pablo Garrido, Ayush Tewari, Weipeng Xu, Justus Thies, Matthias Niessner, Patrick Pérez, Christian Richardt, Michael Zollhöfer, and Christian Theobalt. Deep Video Portraits. ACM TOG, 37(4):1-14, 2018. 3 +[19] Mohammad Rami Koujan, Michail Christos Doukas, Anastasios Roussos, and Stefanos Zafeiriou. Head2head: Video-Based Neural Head Synthesis. In 2020 15th IEEE International Conference on Automatic Face and Gesture Recognition (FG 2020), pages 16-23. IEEE, IEEE, 2020. 3 +[20] Tianye Li, Timo Bolkart, Michael. J. Black, Hao Li, and Javier Romero. Learning a Model of Facial Shape and Expression From 4D Scans. ACM TOG, 36(6):194:1-194:17, 2017. 3 +[21] Shanchuan Lin, Linjie Yang, Imran Saleemi, and Soumyadip Sengupta. Robust high-resolution video matting with temporal guidance. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 238-247, 2022. 5 +[22] Lingjie Liu, Jiatao Gu, Kyaw Zaw Lin, Tat-Seng Chua, and Christian Theobalt. Neural sparse voxel fields. Advances in Neural Information Processing Systems, 33:15651-15663, 2020. 2 +[23] Stephen Lombardi, Tomas Simon, Gabriel Schwartz, Michael Zollhoefer, Yaser Sheikh, and Jason Saragih. Mixture of volumetric primitives for efficient neural rendering. ACM TOG, 40(4):1-13, 2021. 2 +[24] Camillo Lugaresi, Jiuqiang Tang, Hadon Nash, Chris McClanahan, Esha Uboweja, Michael Hays, Fan Zhang, Chuoling Chang, Ming Guang Yong, Juhyun Lee, et al. Mediapipe: A framework for building perception pipelines. arXiv preprint arXiv:1906.08172, 2019. 5 +[25] Abhimitra Meka, Rohit Pandey, Christian Haene, Sergio Orts-Escalano, Peter Barnum, Philip David-Son, Daniel Erickson, Yinda Zhang, Jonathan Taylor, Sofien Bouaziz, et al. Deep Relightable Textures: Volumetric Performance Capture With Neural Rendering. ACM TOG, 39(6):1-21, 2020. 2 +[26] Abhinitra Meka, Rohit Pandey, Christian Haene, Sergio Orts-Escolano, Peter Barnum, Philip Davidson, Daniel Erickson, Yinda Zhang, Jonathan Taylor, Sofien Bouaziz, + +Chloe Legendre, Wan-Chun Ma, Ryan Overbeck, Thabo Beeler, Paul Debevec, Shahram Izadi, Christian Theobalt, Christoph Rhemann, and Sean Fanello. Deep relightable textures - volumetric performance capture with neural rendering. 2020. 1 +[27] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM Trans. Graph., 41(4):102:1-102:15, 2022. 1, 2, 3, 4, 6, 8 +[28] Sergio Orts-Escolano, Christoph Rhemann, Sean Fanello, Wayne Chang, Adarsh Kowdle, Yury Degtyarev, David Kim, Philip Davidson, Sameh Khamis, Mingsong Dou, Vladimir Tankovich, Charles Loop, Qin Cai, Philip Chou, Sarah Mennicken, Julien Valentin, Vivek Pradeep, Shenlong Wang, Sing Bing Kang, Pushmeet Kohli, Yuliya Lutchyn, Cem Keskin, and Shahram Izadi. Holoportation: Virtual 3D Teleportation in Real-Time. In Proceedings of the 29th Annual Symposium on User Interface Software and Technology. ACM, 2016. 1, 2 +[29] Keunhong Park, Utkarsh Sinha, Jonathan T Barron, Sofien Bouaziz, Dan B Goldman, Steven M Seitz, and Ricardo Martin-Brualla. Nerfies: Deformable Neural Radiance Fields. pages 5865-5874, 2021. 5 +[30] Christian Reiser, Songyou Peng, Yiyi Liao, and Andreas Geiger. Kilonerf: Speeding up neural radiance fields with thousands of tiny mlps. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14335-14345, 2021. 2 +[31] Cheng Sun, Min Sun, and Hwann-Tzong Chen. Direct voxel grid optimization: Super-fast convergence for radiance fields reconstruction. In CVPR, 2022. 2 +[32] Ayush Tewari, Justus Thies, Ben Mildenhall, Pratul Srinivasan, Edgar Tretschk, Wang Yifan, Christoph Lassner, Vincent Sitzmann, Ricardo Martin-Brualla, Stephen Lombardi, et al. Advances in neural rendering. In Computer Graphics Forum, pages 703-735. Wiley Online Library, 2022. 2 +[33] Zach Waggoner. My Avatar, My Self: Identity in Video Role-Playing Games. McFarland, 2009. 1 +[34] Lizhen Wang, Xiaochen Zhao, Jingxiang Sun, Yuxiang Zhang, Hongwen Zhang, Tao Yu, and Yebin Liu. Stylea-vatar: Real-time photo-realistic portrait avatar from a single video. In ACM SIGGRAPH 2023 Conference Proceedings, pages 1-10, 2023. 3 +[35] Yuelang Xu, Lizhen Wang, Xiaochen Zhao, Hongwen Zhang, and Yebin Liu. Avatarmav: Fast 3d head avatar reconstruction using motion-aware neural voxels. In ACM SIGGRAPH 2023 Conference Proceedings, pages 1-10, 2023. 2 +[36] Yuelang Xu, Hongwen Zhang, Lizhen Wang, Xiaochen Zhao, Han Huang, Guojun Qi, and Yebin Liu. Latentavatar: Learning latent expression code for expressive neural head avatar. In ACM SIGGRAPH 2023 Conference Proceedings, pages 1-10, 2023. 3 +[37] Alex Yu, Ruilong Li, Matthew Tancik, Hao Li, Ren Ng, and Angjoo Kanazawa. Plenoctrees for real-time rendering of neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5752-5761, 2021. 2 + +[38] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The Unreasonable Effectiveness of Deep Features As a Perceptual Metric. pages 586-595, 2018. 5 +[39] Xiaochen Zhao, Lizhen Wang, Jingxiang Sun, Hongwen Zhang, Jinli Suo, and Yebin Liu. Havatar: High-fidelity head avatar via facial model conditioned neural radiance field. ACM TOG, 43(1):1-16, 2023. 3 +[40] Yufeng Zheng, Victoria Fernández Abrevaya, Marcel C Bühler, Xu Chen, Michael J Black, and Otmar Hilliges. Im avatar: Implicit morphable head avatars from videos. In CVPR, pages 13545-13555, 2022. 2 +[41] Yufeng Zheng, Wang Yifan, Gordon Wetzstein, Michael J Black, and Otmar Hilliges. Pointavatar: Deformable point-based head avatars from videos. In CVPR, pages 21057-21067, 2023. 1, 2, 3, 5, 6, 7 +[42] Wojciech Zielonka, Timo Bolkart, and Justus Thies. Instant volumetric head avatars. In CVPR, pages 4574-4584, 2023. 1, 2, 3, 4, 5, 6, 7 +[43] Michael Zollhöfer, Justus Thies, Pablo Garrido, Derek Bradley, Thabo Beeler, Patrick Pérez, Marc Stamminger, Matthias Nießner, and Christian Theobalt. State of the Art on Monocular 3D Face Reconstruction, Tracking, and Applications. pages 523-550, 2018. 2 \ No newline at end of file diff --git a/2024/Efficient 3D Implicit Head Avatar with Mesh-anchored Hash Table Blendshapes/images.zip b/2024/Efficient 3D Implicit Head Avatar with Mesh-anchored Hash Table Blendshapes/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..a7b0e68593aeb747cefe167b20ffa74924987423 --- /dev/null +++ b/2024/Efficient 3D Implicit Head Avatar with Mesh-anchored Hash Table Blendshapes/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1bac3d06661b0ca2e3656f8b47266b26ec4f1a3121dd23811300a20c21219577 +size 619333 diff --git a/2024/Efficient 3D Implicit Head Avatar with Mesh-anchored Hash Table Blendshapes/layout.json b/2024/Efficient 3D Implicit Head Avatar with Mesh-anchored Hash Table Blendshapes/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..8d0da5f7acf7ca673fa0e8e01f0c00c987d88425 --- /dev/null +++ b/2024/Efficient 3D Implicit Head Avatar with Mesh-anchored Hash Table Blendshapes/layout.json @@ -0,0 +1,7625 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 51, + 103, + 543, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 51, + 103, + 543, + 121 + ], + "spans": [ + { + "bbox": [ + 51, + 103, + 543, + 121 + ], + "type": "text", + "content": "Efficient 3D Implicit Head Avatar with Mesh-anchored Hash Table Blendshapes" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 139, + 141, + 453, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 141, + 453, + 171 + ], + "spans": [ + { + "bbox": [ + 139, + 141, + 453, + 171 + ], + "type": "text", + "content": "Ziqian Bai" + }, + { + "bbox": [ + 139, + 141, + 453, + 171 + ], + "type": "inline_equation", + "content": "^{1,2*}" + }, + { + "bbox": [ + 139, + 141, + 453, + 171 + ], + "type": "text", + "content": " Feitong Tan" + }, + { + "bbox": [ + 139, + 141, + 453, + 171 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 139, + 141, + 453, + 171 + ], + "type": "text", + "content": " Sean Fanello" + }, + { + "bbox": [ + 139, + 141, + 453, + 171 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 139, + 141, + 453, + 171 + ], + "type": "text", + "content": " Rohit Pandey" + }, + { + "bbox": [ + 139, + 141, + 453, + 171 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 139, + 141, + 453, + 171 + ], + "type": "text", + "content": " Mingsong Dou" + }, + { + "bbox": [ + 139, + 141, + 453, + 171 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 139, + 141, + 453, + 171 + ], + "type": "text", + "content": " Shichen Liu" + }, + { + "bbox": [ + 139, + 141, + 453, + 171 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 139, + 141, + 453, + 171 + ], + "type": "text", + "content": " Ping Tan" + }, + { + "bbox": [ + 139, + 141, + 453, + 171 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 139, + 141, + 453, + 171 + ], + "type": "text", + "content": " Yinda Zhang" + }, + { + "bbox": [ + 139, + 141, + 453, + 171 + ], + "type": "inline_equation", + "content": "^{1}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 88, + 172, + 127, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 88, + 172, + 127, + 186 + ], + "spans": [ + { + "bbox": [ + 88, + 172, + 127, + 186 + ], + "type": "text", + "content": "Google" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 146, + 172, + 254, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 172, + 254, + 185 + ], + "spans": [ + { + "bbox": [ + 146, + 172, + 254, + 185 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 146, + 172, + 254, + 185 + ], + "type": "text", + "content": " Simon Fraser University" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 274, + 172, + 502, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 274, + 172, + 502, + 186 + ], + "spans": [ + { + "bbox": [ + 274, + 172, + 502, + 186 + ], + "type": "text", + "content": "3 The Hong Kong University of Science and Technology" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 48, + 215, + 168, + 332 + ], + "blocks": [ + { + "bbox": [ + 48, + 215, + 168, + 332 + ], + "lines": [ + { + "bbox": [ + 48, + 215, + 168, + 332 + ], + "spans": [ + { + "bbox": [ + 48, + 215, + 168, + 332 + ], + "type": "image", + "image_path": "5711f29f7cf67a1111456bdda84716b1dcc8c7f8a5f2a528900c44276fb04533.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 334, + 547, + 378 + ], + "lines": [ + { + "bbox": [ + 46, + 334, + 547, + 378 + ], + "spans": [ + { + "bbox": [ + 46, + 334, + 547, + 378 + ], + "type": "text", + "content": "Figure 1. Visual Quality vs. Rendering Speed comparisons between our monocular 3D head avatars and prior state-of-the-art methods including NeRFBlendshape [14], INSTA [42], PointAvatar [41] and MonoAvatar [5]. Our approach achieves real-time rendering (i.e. " + }, + { + "bbox": [ + 46, + 334, + 547, + 378 + ], + "type": "inline_equation", + "content": ">30" + }, + { + "bbox": [ + 46, + 334, + 547, + 378 + ], + "type": "text", + "content": " mean FPS) with a " + }, + { + "bbox": [ + 46, + 334, + 547, + 378 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 46, + 334, + 547, + 378 + ], + "type": "text", + "content": " resolution, while produces comparable visual quality comparing to prior SOTAs [5], and gives significantly better results on challenging expressions than prior efficient avatars [14, 41, 42]. Webpage augmentedperception.github.io/monoavatar-plus" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 173, + 220, + 246, + 327 + ], + "blocks": [ + { + "bbox": [ + 173, + 220, + 246, + 327 + ], + "lines": [ + { + "bbox": [ + 173, + 220, + 246, + 327 + ], + "spans": [ + { + "bbox": [ + 173, + 220, + 246, + 327 + ], + "type": "image", + "image_path": "e6c022d54e15cc4bdf00f6a0bc6e4efafe906c34aa5218377f999d8304c8f86d.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 248, + 220, + 320, + 327 + ], + "blocks": [ + { + "bbox": [ + 248, + 220, + 320, + 327 + ], + "lines": [ + { + "bbox": [ + 248, + 220, + 320, + 327 + ], + "spans": [ + { + "bbox": [ + 248, + 220, + 320, + 327 + ], + "type": "image", + "image_path": "a377184159282ef49a05f86238bbba261f699c7ecce41884e41d486ee32b5422.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 321, + 220, + 395, + 327 + ], + "blocks": [ + { + "bbox": [ + 321, + 220, + 395, + 327 + ], + "lines": [ + { + "bbox": [ + 321, + 220, + 395, + 327 + ], + "spans": [ + { + "bbox": [ + 321, + 220, + 395, + 327 + ], + "type": "image", + "image_path": "fc3e527d29141d1109789cdb8834ff2002fe41ca2761b9caa4d9a5e003d0f9ca.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 397, + 220, + 470, + 327 + ], + "blocks": [ + { + "bbox": [ + 397, + 220, + 470, + 327 + ], + "lines": [ + { + "bbox": [ + 397, + 220, + 470, + 327 + ], + "spans": [ + { + "bbox": [ + 397, + 220, + 470, + 327 + ], + "type": "image", + "image_path": "35ac6f899365e031ab2e1b353b6adc15bc15f472d6724242609bb85f11ee6651.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 471, + 221, + 545, + 327 + ], + "blocks": [ + { + "bbox": [ + 471, + 221, + 545, + 327 + ], + "lines": [ + { + "bbox": [ + 471, + 221, + 545, + 327 + ], + "spans": [ + { + "bbox": [ + 471, + 221, + 545, + 327 + ], + "type": "image", + "image_path": "2f4e777735712c9a94daba4203b3452828327dd0ef6d2e3ea2d471af69b60c91.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "bbox": [ + 143, + 387, + 192, + 399 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 387, + 192, + 399 + ], + "spans": [ + { + "bbox": [ + 143, + 387, + 192, + 399 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 46, + 412, + 289, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 412, + 289, + 700 + ], + "spans": [ + { + "bbox": [ + 46, + 412, + 289, + 700 + ], + "type": "text", + "content": "3D head avatars built with neural implicit volumetric representations have achieved unprecedented levels of photorealism. However, the computational cost of these methods remains a significant barrier to their widespread adoption, particularly in real-time applications such as virtual reality and teleconferencing. While attempts have been made to develop fast neural rendering approaches for static scenes, these methods cannot be simply employed to support realistic facial expressions, such as in the case of a dynamic facial performance. To address these challenges, we propose a novel fast 3D neural implicit head avatar model that achieves real-time rendering while maintaining fine-grained controllability and high rendering quality. Our key idea lies in the introduction of local hash table blendshapes, which are learned and attached to the vertices of an underlying face parametric model. These per-vertex hash-tables are linearly merged with weights predicted via a CNN, resulting in expression dependent embeddings. Our novel representation enables efficient density and color predictions using a lightweight MLP, which is further accelerated by a hierarchical nearest neighbor search method. Extensive experiments show that our approach runs in real-time while achieving comparable rendering quality to state-of-the-arts and decent results on challenging expressions." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 387, + 386, + 399 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 387, + 386, + 399 + ], + "spans": [ + { + "bbox": [ + 307, + 387, + 386, + 399 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 419, + 545, + 610 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 419, + 545, + 610 + ], + "spans": [ + { + "bbox": [ + 304, + 419, + 545, + 610 + ], + "type": "text", + "content": "The demand of high performing photo-realistic human avatars has dramatically increased with emerging VR/AR applications, e.g. VR gaming [2, 33], virtual assistant [3], tele-presence [28], and 3D videos [15, 26]. How to build efficient high quality avatars from monocular RGB videos becomes a promising direction due to the convenience of monocular data acquisition. While early works mostly adopt surface-based models for convenient controllability, recent methods (e.g. MonoAvatar [5]) leverage a sophisticated pipeline to build human avatars on neural radiance fields, which delivers vivid animations as well as significantly better rendering quality, especially over challenging parts such as hairs and glasses. On the downside, these approaches tend to be prohibitively slow, and most of the computation is consumed by the neural radiance field inference with large Multilayer Perceptrons (MLPs)." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 611, + 545, + 694 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 611, + 545, + 694 + ], + "spans": [ + { + "bbox": [ + 304, + 611, + 545, + 694 + ], + "type": "text", + "content": "Recently, fast approaches for neural radiance fields (e.g. hash encoding in Instant NGPs [27]) have been proposed, which are designed mostly for static scenes or pre-recorded temporal sequences. Despite their great success, it is not straightforward to extend these approaches for human avatars, which requires real-time rendering of dynamic facial performances when controlling the avatar. NeRFblend" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 320, + 702, + 529, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 702, + 529, + 713 + ], + "spans": [ + { + "bbox": [ + 320, + 702, + 529, + 713 + ], + "type": "text", + "content": "*Work was conducted while Ziqian Bai was an intern at Google." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1975" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 263 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 263 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 263 + ], + "type": "text", + "content": "shape [14] address these issues by learning multiple feature hash tables, one for each face blendshape. These hash tables are linearly combined with blending weights to render the target facial expression via hash encoding. However, the expressiveness of the avatar is compromised by the global nature of the blending shapes, which cannot accurately capture vertex-level local deformations. On the other hand, INSTA [42] proposes to build the appearance model in a canonical space using a vanilla Instant NGPs [27] with expression codes as the additional MLP input to capture dynamic details, which is then transformed into target expression via a face parametric model (i.e. 3D morphable model (3DMM)). However, the lightweight MLP used in the vanilla Instant NGPs limits their model capacity, resulting in inferior animation quality, especially on extreme expressions." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 266, + 289, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 266, + 289, + 495 + ], + "spans": [ + { + "bbox": [ + 46, + 266, + 289, + 495 + ], + "type": "text", + "content": "In this work, we propose a novel 3D neural avatar system that achieves efficient inference while maintaining fine-grained controllability and high fidelity quality. To achieve this, we introduce mesh-anchored hash table blendshapes, where we attach multiple, small hash tables to each of the 3DMM mesh vertices. These hash tables act as per-vertex \"local blendshapes\" (i.e., each \"blendshape\" is controlled by one local hash table) and influence only a local region. The mesh-anchored blendshapes are linearly merged with per-vertex weights predicted by a convolutional neural network in UV space from avatar driving signals, such as expression and head rotation. This results in expression-dependent hash table embeddings which offer several advantages over a global linear combination of blendshapes. Indeed by associating hash tables with individual vertices, we enhance the expressiveness of the model, allowing for more localized and nuanced facial expressions. This contrasts with global blendshapes, which apply uniform transformations across the entire face, limiting expressiveness." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 496, + 289, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 496, + 289, + 651 + ], + "spans": [ + { + "bbox": [ + 46, + 496, + 289, + 651 + ], + "type": "text", + "content": "In more detail, our model starts from 3D query points, uses hash encoding [27] to gather the merged hash table embeddings from " + }, + { + "bbox": [ + 46, + 496, + 289, + 651 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 496, + 289, + 651 + ], + "type": "text", + "content": "-nearest-neighbor vertices around the query point, and predicts the density and color via a small MLP. The hash encoding [27] allows us to use a very lightweight MLP to significantly reduce computation, leading to efficient inference. Additionally, the vertex-attached hash table blendshapes represent a 3DMM-anchored neural radiance field (NeRF), which can be easily controlled by the underlying 3DMM and produce high fidelity renderings as demonstrated by MonoAvatar [5]. To further accelerate our rendering speed, we propose a hierarchical " + }, + { + "bbox": [ + 46, + 496, + 289, + 651 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 496, + 289, + 651 + ], + "type": "text", + "content": "-nearest-neighbor search method." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 653, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 289, + 715 + ], + "type": "text", + "content": "Our contributions are summarized as follows. We propose a novel approach for high quality and efficient 3D neural implicit head avatars. At the core of our model, vertex-attached local hash table blendshapes are proposed to support efficient rendering, controllability, and capturing fine" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 304, + 72, + 547, + 194 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 194 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 194 + ], + "type": "text", + "content": "grained rendering details in dynamic facial performances. We also design a hierarchical querying solution to speed up the " + }, + { + "bbox": [ + 304, + 72, + 547, + 194 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 72, + 547, + 194 + ], + "type": "text", + "content": "-nearest-neighbor search when pulling hash table embeddings from neighbor vertices. Extensive experiments on multiple datasets verify that we are able to speed up avatar rendering to real-time (i.e., average " + }, + { + "bbox": [ + 304, + 72, + 547, + 194 + ], + "type": "inline_equation", + "content": "\\geq 30" + }, + { + "bbox": [ + 304, + 72, + 547, + 194 + ], + "type": "text", + "content": " FPS to render a " + }, + { + "bbox": [ + 304, + 72, + 547, + 194 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 304, + 72, + 547, + 194 + ], + "type": "text", + "content": " video) while maintaining comparable rendering quality with the state-of-the-art high quality 3D avatar [5] and being largely superior on challenging expressions than existing efficient 3D avatars [14, 41, 42]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 305, + 203, + 392, + 216 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 203, + 392, + 216 + ], + "spans": [ + { + "bbox": [ + 305, + 203, + 392, + 216 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 223, + 547, + 284 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 223, + 547, + 284 + ], + "spans": [ + { + "bbox": [ + 304, + 223, + 547, + 284 + ], + "type": "text", + "content": "Constructing photorealistic digital humans has been a extensively researched topic. Here, we focus on discussing prior work on implicit monocular head avatars and efficient rendering. We refer readers to state-of-the-art surveys [12, 32, 43] for a comprehensive literature review." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 289, + 547, + 564 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 289, + 547, + 564 + ], + "spans": [ + { + "bbox": [ + 304, + 289, + 547, + 564 + ], + "type": "text", + "content": "High Quality Head Avatar. Traditionally, high-quality head avatars have been achieved under expensive equipment configurations, such as camera arrays [6, 23, 28], depth sensors [7], and light stages [15, 25], or require laborious manual intervention [1]. Recent research efforts have focused on constructing high quality avatars from monocular RGB videos. One typical class of approaches [4, 13, 35, 40] use implicit 3D representations (i.e., neural radiance fields (NeRFs), implicit occupancy fields) to build the head avatar, which are parameterized by Multilayer perceptrons (MLPs). Although reasonable results are obtained, their rendering quality is still unsatisfactory especially for more challenging expressions. More recently, Bai et al. [5] proposed a head avatar based on 3DMM-anchored NeRFs with expression-dependent features produced by a convolution neural network in UV space. Chen et al. [9] designed local deformation fields to capture expression-dependent deformations applied on canonical NeRFs. Despite the impressive results, their methods didn't demonstrate real-time rendering capability due to expensive inference using large MLPs. In contrast, with our proposed mesh-anchored hash table blendshapes (Sec. 3), we achieve much faster rendering speed while maintaining high fidelity results." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 570, + 548, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 548, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 548, + 715 + ], + "type": "text", + "content": "Efficient Neural Radiance Fields. There has been a plethora of work in recent years attempting to accelerate rendering with neural implicit representations for static objects and scenes. SNeRG [22], DVGO [31] and Plenoxels [37] propose to directly optimize voxel grids of (neural or SH) features for faster performance. However, their approach still requires a large memory footprint to store pervoxel features in 3D space. KiloNeRF [30] dramatically accelerates the original NeRF by representing the scene with thousands of tiny MLPs, however, this approach requires a complex training strategy. TensoRF [8] factorizes the feature grid into compact components, resulting in sig" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 317, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 317, + 758 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 317, + 758 + ], + "type": "text", + "content": "1976" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 47, + 68, + 545, + 216 + ], + "blocks": [ + { + "bbox": [ + 47, + 68, + 545, + 216 + ], + "lines": [ + { + "bbox": [ + 47, + 68, + 545, + 216 + ], + "spans": [ + { + "bbox": [ + 47, + 68, + 545, + 216 + ], + "type": "image", + "image_path": "44e43a35c15cca6871b74a7c81e903c099dfc1401b4ff664a9d59249d41a42a1.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 220, + 546, + 285 + ], + "lines": [ + { + "bbox": [ + 46, + 220, + 546, + 285 + ], + "spans": [ + { + "bbox": [ + 46, + 220, + 546, + 285 + ], + "type": "text", + "content": "Figure 2. Overview of our pipeline. Our core avatar representation is Mesh-anchored Hash Table Blendshapes (Sec. 3.1), where multiple small hash tables are attached to each 3DMM vertex. During inference, our method starts from a displacement map encoding the facial expression, which is then fed into a U-Net to predict hash table weights and per-vertex features. The predicted weights are used to linearly combine the multiple hash tables attached on each 3DMM vertex (Sec. 3.2). During volumetric rendering (Sec. 3.3 and Fig. 3), for each query point, we search its " + }, + { + "bbox": [ + 46, + 220, + 546, + 285 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 220, + 546, + 285 + ], + "type": "text", + "content": "-nearest-neighbor vertices, then pull embeddings from the merged hash tables and concatenate with the per-vertex feature to decode local density and color via a tiny MLP with two hidden layers." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 301, + 289, + 661 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 301, + 289, + 661 + ], + "spans": [ + { + "bbox": [ + 46, + 301, + 289, + 661 + ], + "type": "text", + "content": "nificantly higher memory efficiency. Concurrently, Instant NGPs [27] utilizes multi-resolution hashing for efficient encoding, resulting in high compactness. MobileNeRF [10] propose to represent NeRF based on polygons, which allows leveraging a traditional polygon pipeline to enable their method to run in real-time on mobile devices. 3D Gaussian Splatting [17] represents a radiance field with 3D Gaussian point clouds, and leverages a point rasterization pipeline to enable fast rendering. These methods, however, cannot be easily extended to controllable dynamic contents. More recently, NeRFBlendshape [14] was proposed to handle controllable expressions by learning multiple hash tables for different global blendshapes and linearly combine them with expression codes. INSTA [42] transforms all expressions into a shared 3D canonical space, then adopts the vanilla Instant NGPs [27] conditioned on expression codes to model the head avatar. Despite the fast rendering speed achieved, their methods suffer from unsatisfactory rendering quality. PointAvatar [41] utilizes point clouds to represent the head avatar and uses large MLPs to predict the colors and motions of each point, leading to slow inference. Another type of works [11, 18, 19, 34, 36, 39] use 2D convolution neural networks to directly synthesize images (i.e. 2D neural rendering) from rasterized 3DMM meshes or low resolution feature maps generated by volumetric rendering. Despite their fast speed, the 2D CNNs may break the 3D consistency, leading to temporally unstable results especially for high frequency details. In contrast, our method simultaneously achieves controllability, high quality, and efficient rendering with a fully 3D representation." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 670, + 103, + 681 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 670, + 103, + 681 + ], + "spans": [ + { + "bbox": [ + 47, + 670, + 103, + 681 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 689, + 287, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 689, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 46, + 689, + 287, + 712 + ], + "type": "text", + "content": "Given a monocular RGB video, our method learns a neural radiance field (NeRF) based head avatar, which can be ren-" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 305, + 301, + 546, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 301, + 546, + 373 + ], + "spans": [ + { + "bbox": [ + 305, + 301, + 546, + 373 + ], + "type": "text", + "content": "dered under any specified cameras, articulated poses (i.e., neck, jaw, and eyes) and facial expressions defined by a face parametric model (i.e. 3DMM). We use FLAME [20] as the parametric model in this work, but our method can be generalized to any other mesh-based parametric models. Fig. 2 shows the overview of our method." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 403, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 403, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 403, + 546, + 713 + ], + "type": "text", + "content": "Our goal is to design a 3D neural implicit head avatar architecture that can simultaneously achieve high image quality, controllability, and computationally efficient rendering. To achieve this, we propose mesh-anchored hash table blendshapes (Sec. 3.1) as a novel avatar representation that can leverage both advantages from recent high-quality (i.e., 3DMM-anchored NeRF [5]) and efficient (i.e., hash encoding [27]) frameworks. More specifically, we propose to attach multiple small hash tables on each 3DMM vertex. These vertex-attached hash tables form a set of local \"blendshapes\", which will be linearly merged with predicted blending weights (Sec. 3.2), and decoded into a 3DMM-anchored NeRF to support fine-grained control and high fidelity rendering. During NeRF decoding on a query point (Sec. 3.3), we pull the embeddings from the linearly merged hash tables attached on " + }, + { + "bbox": [ + 304, + 403, + 546, + 713 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 403, + 546, + 713 + ], + "type": "text", + "content": "-nearest-neighbor (" + }, + { + "bbox": [ + 304, + 403, + 546, + 713 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 403, + 546, + 713 + ], + "type": "text", + "content": "-NN) vertices from the 3DMM mesh. Using hash encoding allows us to use a very light weight MLP (only 2 hidden layers) to predict the final densities and colors, which is the key for efficient rendering. To further accelerate our approach to real-time, we leverage the fact that close query points likely share similar " + }, + { + "bbox": [ + 304, + 403, + 546, + 713 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 403, + 546, + 713 + ], + "type": "text", + "content": "-NN vertices, and thus propose to group the query points into voxels and hierarchically search for " + }, + { + "bbox": [ + 304, + 403, + 546, + 713 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 403, + 546, + 713 + ], + "type": "text", + "content": "-NN vertices (Sec. 3.4). Finally, our proposed avatar representation can be trained with only monocular RGB videos without any 3D scans or multi-view data (Sec. 3.5)." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1977" + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 260, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 260, + 85 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 260, + 85 + ], + "type": "text", + "content": "3.1. Mesh-anchored Hash Table Blendshapes" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 90, + 287, + 209 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 90, + 287, + 209 + ], + "spans": [ + { + "bbox": [ + 46, + 90, + 287, + 209 + ], + "type": "text", + "content": "The core of our model is an avatar representation that can represent a 3DMM-anchored neural radiance field (NeRF) while allowing us to adopt the hash encoding [27] technique for acceleration. Recent approaches manage to adopt hash encoding into head avatars with different avatar representations (i.e., global blendshapes [14] and canonical NeRF [42]). In contrast, our solution is built upon the most recent 3DMM-anchored NeRF [5], which is superior for high quality renderings as demonstrated in our experiments (Sec. 4.2)." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 210, + 288, + 428 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 210, + 288, + 428 + ], + "spans": [ + { + "bbox": [ + 46, + 210, + 288, + 428 + ], + "type": "text", + "content": "We propose mesh-anchored hash table blendshapes as the new avatar representation. Given a target expression " + }, + { + "bbox": [ + 46, + 210, + 288, + 428 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 210, + 288, + 428 + ], + "type": "text", + "content": " with 3DMM pose code " + }, + { + "bbox": [ + 46, + 210, + 288, + 428 + ], + "type": "inline_equation", + "content": "\\theta_{i}" + }, + { + "bbox": [ + 46, + 210, + 288, + 428 + ], + "type": "text", + "content": " and expression code " + }, + { + "bbox": [ + 46, + 210, + 288, + 428 + ], + "type": "inline_equation", + "content": "\\psi_{i}" + }, + { + "bbox": [ + 46, + 210, + 288, + 428 + ], + "type": "text", + "content": ", we first get the deformed 3DMM mesh " + }, + { + "bbox": [ + 46, + 210, + 288, + 428 + ], + "type": "inline_equation", + "content": "\\mathbf{V}_{i} = \\mathcal{F}_{3\\mathrm{DMM}}(\\psi_{i},\\theta_{i})" + }, + { + "bbox": [ + 46, + 210, + 288, + 428 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 46, + 210, + 288, + 428 + ], + "type": "inline_equation", + "content": "J" + }, + { + "bbox": [ + 46, + 210, + 288, + 428 + ], + "type": "text", + "content": " vertices. For the " + }, + { + "bbox": [ + 46, + 210, + 288, + 428 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 46, + 210, + 288, + 428 + ], + "type": "text", + "content": "-th 3DMM mesh vertex " + }, + { + "bbox": [ + 46, + 210, + 288, + 428 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_{ij}" + }, + { + "bbox": [ + 46, + 210, + 288, + 428 + ], + "type": "text", + "content": ", we attach " + }, + { + "bbox": [ + 46, + 210, + 288, + 428 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 46, + 210, + 288, + 428 + ], + "type": "text", + "content": " small hash tables " + }, + { + "bbox": [ + 46, + 210, + 288, + 428 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{H}_j^{(m)}\\}_{M}" + }, + { + "bbox": [ + 46, + 210, + 288, + 428 + ], + "type": "text", + "content": " on it, where each hash table has multiple resolutions following instant NGPs [27]. Intuitively, these hash tables form a set of vertex-level \"blendshapes\" anchored on the mesh, where each \"blendshape\" is a hash table, whose embeddings encode the information of a local radiance field around vertex " + }, + { + "bbox": [ + 46, + 210, + 288, + 428 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_{ij}" + }, + { + "bbox": [ + 46, + 210, + 288, + 428 + ], + "type": "text", + "content": ". Given a target expression to render, these hash table blendshapes are linearly summed via expression-dependent weights (Sec. 3.2), such that the merged embeddings encode the fine details specific to the target expression. Simultaneously, the coarse motion of the target expression is captured by the 3DMM vertex movement, which moves the attached hash tables accordingly, and hence the corresponding local radiance field." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 434, + 237, + 447 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 434, + 237, + 447 + ], + "spans": [ + { + "bbox": [ + 47, + 434, + 237, + 447 + ], + "type": "text", + "content": "3.2. Merge Mesh-anchored Blendshapes" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 452, + 287, + 610 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 452, + 287, + 610 + ], + "spans": [ + { + "bbox": [ + 46, + 452, + 287, + 610 + ], + "type": "text", + "content": "We obtain per-vertex blending weights by running a convolution neural network (CNN) on the 3DMM deformation represented in UV atlas space. Specifically, we calculate the vertex displacements with respect to the neutral face " + }, + { + "bbox": [ + 46, + 452, + 287, + 610 + ], + "type": "inline_equation", + "content": "\\mathbf{D}_i = \\mathcal{F}_{\\mathrm{3DMM}}(\\psi_i,\\pmb {\\theta}_i) - \\mathcal{F}_{\\mathrm{3DMM}}(\\mathbf{0},\\mathbf{0})" + }, + { + "bbox": [ + 46, + 452, + 287, + 610 + ], + "type": "text", + "content": " . The displacements are then warped into the UV space and fed into a U-Net to predict a weights map in " + }, + { + "bbox": [ + 46, + 452, + 287, + 610 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^{\\mathrm{H_t}\\times \\mathrm{W_t}\\times \\mathrm{M}}" + }, + { + "bbox": [ + 46, + 452, + 287, + 610 + ], + "type": "text", + "content": " , where " + }, + { + "bbox": [ + 46, + 452, + 287, + 610 + ], + "type": "inline_equation", + "content": "\\mathrm{H_t}\\times \\mathrm{W_t}" + }, + { + "bbox": [ + 46, + 452, + 287, + 610 + ], + "type": "text", + "content": " is the UV resolution, and " + }, + { + "bbox": [ + 46, + 452, + 287, + 610 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 46, + 452, + 287, + 610 + ], + "type": "text", + "content": " is the number of hash table blendshapes on each vertex (pre-defined as 5 in our experiments). The weights map is then sampled back to 3DMM vertices, serving as the expression-dependent weights " + }, + { + "bbox": [ + 46, + 452, + 287, + 610 + ], + "type": "inline_equation", + "content": "\\{w_{ij}^{(m)}\\}_{M}" + }, + { + "bbox": [ + 46, + 452, + 287, + 610 + ], + "type": "text", + "content": " to take a weighted sum of the embeddings in the hash tables on each vertex, which produces the merged hash tables" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 117, + 616, + 287, + 648 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 117, + 616, + 287, + 648 + ], + "spans": [ + { + "bbox": [ + 117, + 616, + 287, + 648 + ], + "type": "interline_equation", + "content": "\\hat {\\mathbf {H}} _ {i j} = \\sum_ {m = 1} ^ {M} w _ {i j} ^ {(m)} \\mathbf {H} _ {j} ^ {(m)}. \\tag {1}", + "image_path": "9c0f06eb0e0f810ed9ee9cf36019dd3ee92eff2fd0736bb73fecf18dbb4010d8.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 653, + 287, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 287, + 712 + ], + "type": "text", + "content": "The U-Net also produces a UV feature map. We sample a per-vertex feature " + }, + { + "bbox": [ + 46, + 653, + 287, + 712 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_{ij}" + }, + { + "bbox": [ + 46, + 653, + 287, + 712 + ], + "type": "text", + "content": " from it, similar to MonoAvatar [5]. We empirically found this benefits the geometry quality. The mesh-anchored hash tables " + }, + { + "bbox": [ + 46, + 653, + 287, + 712 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{H}}_{ij}" + }, + { + "bbox": [ + 46, + 653, + 287, + 712 + ], + "type": "text", + "content": " and features " + }, + { + "bbox": [ + 46, + 653, + 287, + 712 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_{ij}" + }, + { + "bbox": [ + 46, + 653, + 287, + 712 + ], + "type": "text", + "content": " are decoded into a neural radiance field as described in Sec. 3.3." + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 313, + 70, + 541, + 251 + ], + "blocks": [ + { + "bbox": [ + 313, + 70, + 541, + 251 + ], + "lines": [ + { + "bbox": [ + 313, + 70, + 541, + 251 + ], + "spans": [ + { + "bbox": [ + 313, + 70, + 541, + 251 + ], + "type": "image", + "image_path": "72db2b65c8180081479f4da2a586cfa6c43a25b8ea455a2da735e2d2c62df568.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 258, + 545, + 304 + ], + "lines": [ + { + "bbox": [ + 304, + 258, + 545, + 304 + ], + "spans": [ + { + "bbox": [ + 304, + 258, + 545, + 304 + ], + "type": "text", + "content": "Figure 3. Pipeline of Neural Radiance Field Decoding (Sec. 3.3). Our method only needs a very lightweight MLP for NeRF inference, where a two-hidden-layer MLP is used to predict the density and the color of each query point." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 321, + 481, + 335 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 321, + 481, + 335 + ], + "spans": [ + { + "bbox": [ + 306, + 321, + 481, + 335 + ], + "type": "text", + "content": "3.3. Neural Radiance Field Decoding" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 339, + 545, + 447 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 339, + 545, + 447 + ], + "spans": [ + { + "bbox": [ + 304, + 339, + 545, + 447 + ], + "type": "text", + "content": "Given the mesh-anchored hash tables " + }, + { + "bbox": [ + 304, + 339, + 545, + 447 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{H}}_{ij}" + }, + { + "bbox": [ + 304, + 339, + 545, + 447 + ], + "type": "text", + "content": " and features " + }, + { + "bbox": [ + 304, + 339, + 545, + 447 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_{ij}" + }, + { + "bbox": [ + 304, + 339, + 545, + 447 + ], + "type": "text", + "content": " described in Sec. 3.2, the final step is to decode them into a Neural Radiance Field (NeRF) to render the output image as shown in Fig. 3. The key idea is to associate a query point to neighbor vertices, and pull the embeddings from the attached hash tables via hash encoding [27]. Finally, we decode these pulled embeddings and the nearest per-vertex feature into the color and the density of this query point, followed by volumetric rendering to obtain the output image." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 448, + 546, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 448, + 546, + 616 + ], + "spans": [ + { + "bbox": [ + 304, + 448, + 546, + 616 + ], + "type": "text", + "content": "For a 3D query point " + }, + { + "bbox": [ + 304, + 448, + 546, + 616 + ], + "type": "inline_equation", + "content": "\\mathbf{q}" + }, + { + "bbox": [ + 304, + 448, + 546, + 616 + ], + "type": "text", + "content": " when rendering a particular facial expression " + }, + { + "bbox": [ + 304, + 448, + 546, + 616 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 448, + 546, + 616 + ], + "type": "text", + "content": ", we first obtain its " + }, + { + "bbox": [ + 304, + 448, + 546, + 616 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 448, + 546, + 616 + ], + "type": "text", + "content": "-nearest-neighbors, denoted as " + }, + { + "bbox": [ + 304, + 448, + 546, + 616 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{v}_{ik}\\}_{k\\in \\mathcal{N}_{\\mathbf{q}}^{K}}" + }, + { + "bbox": [ + 304, + 448, + 546, + 616 + ], + "type": "text", + "content": ", from the 3DMM vertices, with " + }, + { + "bbox": [ + 304, + 448, + 546, + 616 + ], + "type": "inline_equation", + "content": "k^*" + }, + { + "bbox": [ + 304, + 448, + 546, + 616 + ], + "type": "text", + "content": " denoting the nearest vertex index. For each neighbor vertex " + }, + { + "bbox": [ + 304, + 448, + 546, + 616 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_{ik}" + }, + { + "bbox": [ + 304, + 448, + 546, + 616 + ], + "type": "text", + "content": " with an attached hash table " + }, + { + "bbox": [ + 304, + 448, + 546, + 616 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{H}}_{ik}" + }, + { + "bbox": [ + 304, + 448, + 546, + 616 + ], + "type": "text", + "content": ", we denote " + }, + { + "bbox": [ + 304, + 448, + 546, + 616 + ], + "type": "inline_equation", + "content": "\\mathbf{q}_{ik}" + }, + { + "bbox": [ + 304, + 448, + 546, + 616 + ], + "type": "text", + "content": " as the coordinates of " + }, + { + "bbox": [ + 304, + 448, + 546, + 616 + ], + "type": "inline_equation", + "content": "\\mathbf{q}" + }, + { + "bbox": [ + 304, + 448, + 546, + 616 + ], + "type": "text", + "content": " in the tangent space of " + }, + { + "bbox": [ + 304, + 448, + 546, + 616 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_{ik}" + }, + { + "bbox": [ + 304, + 448, + 546, + 616 + ], + "type": "text", + "content": ". We then use " + }, + { + "bbox": [ + 304, + 448, + 546, + 616 + ], + "type": "inline_equation", + "content": "\\mathbf{q}_{ik}" + }, + { + "bbox": [ + 304, + 448, + 546, + 616 + ], + "type": "text", + "content": " to query the hash table " + }, + { + "bbox": [ + 304, + 448, + 546, + 616 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{H}}_{ik}" + }, + { + "bbox": [ + 304, + 448, + 546, + 616 + ], + "type": "text", + "content": " using a hash encoding function " + }, + { + "bbox": [ + 304, + 448, + 546, + 616 + ], + "type": "inline_equation", + "content": "\\mathcal{H}(\\cdot)" + }, + { + "bbox": [ + 304, + 448, + 546, + 616 + ], + "type": "text", + "content": " and obtain the embedding " + }, + { + "bbox": [ + 304, + 448, + 546, + 616 + ], + "type": "inline_equation", + "content": "\\mathbf{h}_{ik} = \\mathcal{H}(\\mathbf{q}_{ik};\\hat{\\mathbf{H}}_{ik})" + }, + { + "bbox": [ + 304, + 448, + 546, + 616 + ], + "type": "text", + "content": ". To interpolate the embeddings from all " + }, + { + "bbox": [ + 304, + 448, + 546, + 616 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 448, + 546, + 616 + ], + "type": "text", + "content": "-nearest-neighbors, we use the weighted sum of the inverse distances " + }, + { + "bbox": [ + 304, + 448, + 546, + 616 + ], + "type": "inline_equation", + "content": "z_{k} = 1 / \\| \\mathbf{q}_{ik}\\|_{2}" + }, + { + "bbox": [ + 304, + 448, + 546, + 616 + ], + "type": "text", + "content": ". Next, the summed embedding, together with the nearest per-vertex feature " + }, + { + "bbox": [ + 304, + 448, + 546, + 616 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_{ik^*}" + }, + { + "bbox": [ + 304, + 448, + 546, + 616 + ], + "type": "text", + "content": " and the query point tangent coordinate " + }, + { + "bbox": [ + 304, + 448, + 546, + 616 + ], + "type": "inline_equation", + "content": "\\mathbf{q}_{ik^*}" + }, + { + "bbox": [ + 304, + 448, + 546, + 616 + ], + "type": "text", + "content": " of the nearest vertex, are fed into a two-hidden-layer MLP to predict the density and color as" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 328, + 620, + 545, + 649 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 328, + 620, + 545, + 649 + ], + "spans": [ + { + "bbox": [ + 328, + 620, + 545, + 649 + ], + "type": "interline_equation", + "content": "\\overline {{\\mathbf {h}}} _ {i} = \\sum_ {k \\in \\mathcal {N} _ {\\mathbf {q}} ^ {K}} \\bar {w} _ {k} \\mathbf {h} _ {i k}, \\text {w h e r e} \\bar {w} _ {k} = \\frac {z _ {k}}{\\sum_ {k ^ {\\prime} \\in \\mathcal {N} _ {\\mathbf {q}} ^ {K}} z _ {k ^ {\\prime}}}, \\tag {2}", + "image_path": "3248ce5f22ecc5cd08b1322e9801eb571828a9b7cc815d109de7c95d7df017e3.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 310, + 652, + 541, + 672 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 310, + 652, + 541, + 672 + ], + "spans": [ + { + "bbox": [ + 310, + 652, + 541, + 672 + ], + "type": "interline_equation", + "content": "\\left[ \\mathbf {c} _ {i} (\\mathbf {q}, \\mathbf {d}), \\sigma_ {i} (\\mathbf {q}) \\right] = \\mathcal {F} _ {\\mathrm {M L P}} \\left(\\overline {{\\mathbf {h}}} _ {i}, \\mathbf {f} _ {i k ^ {*}}, \\operatorname {P E} \\left(\\mathbf {q} _ {i k ^ {*}}\\right), \\operatorname {P E} (\\mathbf {d})\\right),", + "image_path": "7e02beeef51836cb28fc86974d8b8ec1e96ebf4a0d6026ea2eb64b4f547646e2.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\overline{w}_k" + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "content": " is the normalized inverse-distance based weight, " + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{d}" + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "content": " denotes the camera view direction, " + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{PE}(\\cdot)" + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "content": " denotes positional encoding, " + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{c}_i" + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "content": " denotes color, and " + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\sigma_{i}" + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "content": " denotes density." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1978" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 109 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 109 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 109 + ], + "type": "text", + "content": "Finally, we render the output pixel with the given camera ray " + }, + { + "bbox": [ + 46, + 72, + 287, + 109 + ], + "type": "inline_equation", + "content": "\\mathbf{r}" + }, + { + "bbox": [ + 46, + 72, + 287, + 109 + ], + "type": "text", + "content": " by volumetric rendering, where we reparameterize the query point with samples on the ray " + }, + { + "bbox": [ + 46, + 72, + 287, + 109 + ], + "type": "inline_equation", + "content": "\\mathbf{q} = \\mathbf{r}(t) = \\mathbf{o} + t\\mathbf{d}" + }, + { + "bbox": [ + 46, + 72, + 287, + 109 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 94, + 114, + 262, + 143 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 94, + 114, + 262, + 143 + ], + "spans": [ + { + "bbox": [ + 94, + 114, + 262, + 143 + ], + "type": "interline_equation", + "content": "\\mathbf {C} _ {i} (\\mathbf {r}) = \\int_ {t _ {n}} ^ {t _ {f}} T (t) \\sigma_ {i} (\\mathbf {r} (t)) \\mathbf {c} _ {i} (\\mathbf {r} (t), \\mathbf {d}) d t,", + "image_path": "ffdc354703fc9eaf19ec0b0b9ddd1d0129f8a752677dfe2dc73372b9709a9841.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 73, + 144, + 287, + 172 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 73, + 144, + 287, + 172 + ], + "spans": [ + { + "bbox": [ + 73, + 144, + 287, + 172 + ], + "type": "interline_equation", + "content": "\\mathrm {w h e r e} T (t) = \\exp \\left(- \\int_ {t _ {n}} ^ {t} \\sigma_ {i} (\\mathbf {r} (s)) d s\\right). \\qquad (3)", + "image_path": "04411dd18665b3b6fa49feba20a58acf9b557f6eb59c7a7042b4a3d53c91dca7.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 175, + 288, + 331 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 175, + 288, + 331 + ], + "spans": [ + { + "bbox": [ + 46, + 175, + 288, + 331 + ], + "type": "text", + "content": "Following prior works [5, 16], we also introduce a per-frame error-correction warping field during training to reduce misalignments due to the noise in 3DMM tracking and unmodeled per-frame contents such as hair movements. We feed the query point " + }, + { + "bbox": [ + 46, + 175, + 288, + 331 + ], + "type": "inline_equation", + "content": "\\mathbf{q}" + }, + { + "bbox": [ + 46, + 175, + 288, + 331 + ], + "type": "text", + "content": ", together with a per-frame latent code " + }, + { + "bbox": [ + 46, + 175, + 288, + 331 + ], + "type": "inline_equation", + "content": "\\mathbf{e}_i" + }, + { + "bbox": [ + 46, + 175, + 288, + 331 + ], + "type": "text", + "content": ", into an MLP " + }, + { + "bbox": [ + 46, + 175, + 288, + 331 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{\\mathcal{E}}(\\cdot)" + }, + { + "bbox": [ + 46, + 175, + 288, + 331 + ], + "type": "text", + "content": " to obtain a rigid transformation applied on the original query point, denoted as " + }, + { + "bbox": [ + 46, + 175, + 288, + 331 + ], + "type": "inline_equation", + "content": "\\mathbf{q}' = \\mathcal{T}_i(\\mathbf{q}) = \\mathcal{F}_{\\mathcal{E}}(\\mathbf{q},\\mathbf{e}_i)" + }, + { + "bbox": [ + 46, + 175, + 288, + 331 + ], + "type": "text", + "content": ". The warped query point " + }, + { + "bbox": [ + 46, + 175, + 288, + 331 + ], + "type": "inline_equation", + "content": "\\mathbf{q}'" + }, + { + "bbox": [ + 46, + 175, + 288, + 331 + ], + "type": "text", + "content": " is then used to compute the density and color for volumetric rendering. Since the warping fields are overfit to corresponding training frames, we disable the warping field during testing similar to previous works [5, 16], and hence " + }, + { + "bbox": [ + 46, + 175, + 288, + 331 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{\\mathcal{E}}" + }, + { + "bbox": [ + 46, + 175, + 288, + 331 + ], + "type": "text", + "content": " does not affect rendering efficiency." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 336, + 255, + 348 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 336, + 255, + 348 + ], + "spans": [ + { + "bbox": [ + 47, + 336, + 255, + 348 + ], + "type": "text", + "content": "3.4. Hierarchical " + }, + { + "bbox": [ + 47, + 336, + 255, + 348 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 47, + 336, + 255, + 348 + ], + "type": "text", + "content": "-nearest-neighbor Search" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 354, + 287, + 606 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 354, + 287, + 606 + ], + "spans": [ + { + "bbox": [ + 46, + 354, + 287, + 606 + ], + "type": "text", + "content": "As described in Sec. 3.3, our method involves a " + }, + { + "bbox": [ + 46, + 354, + 287, + 606 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 354, + 287, + 606 + ], + "type": "text", + "content": "-nearest-neighbor (" + }, + { + "bbox": [ + 46, + 354, + 287, + 606 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 354, + 287, + 606 + ], + "type": "text", + "content": "-NN) search, which is computationally expensive and cannot be naively accelerated with pre-calculated structures (e.g., KD-Tree) due to the dynamically changing search pool (i.e., the 3DMM vertices driven by poses and expressions). To speed up the process, we propose a hierarchical " + }, + { + "bbox": [ + 46, + 354, + 287, + 606 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 354, + 287, + 606 + ], + "type": "text", + "content": "-NN search algorithm following a coarse-to-fine strategy. The key idea is to group nearby query points into a cluster as they likely share similar nearest neighbors. Specifically, we use a 3D grid with resolution 64 and treat all query points that fall in each voxel as a cluster. For each cluster, we first search " + }, + { + "bbox": [ + 46, + 354, + 287, + 606 + ], + "type": "inline_equation", + "content": "K'" + }, + { + "bbox": [ + 46, + 354, + 287, + 606 + ], + "type": "text", + "content": " (where " + }, + { + "bbox": [ + 46, + 354, + 287, + 606 + ], + "type": "inline_equation", + "content": "K' > K" + }, + { + "bbox": [ + 46, + 354, + 287, + 606 + ], + "type": "text", + "content": ") nearest neighbors of the voxel center from all 3DMM vertices. Then, for each query point, we search " + }, + { + "bbox": [ + 46, + 354, + 287, + 606 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 46, + 354, + 287, + 606 + ], + "type": "text", + "content": " nearest neighbors from the " + }, + { + "bbox": [ + 46, + 354, + 287, + 606 + ], + "type": "inline_equation", + "content": "K'" + }, + { + "bbox": [ + 46, + 354, + 287, + 606 + ], + "type": "text", + "content": " nearest neighbors of the corresponding cluster. In practice, for a 3DMM with the vertices number of " + }, + { + "bbox": [ + 46, + 354, + 287, + 606 + ], + "type": "inline_equation", + "content": "J = 1772" + }, + { + "bbox": [ + 46, + 354, + 287, + 606 + ], + "type": "text", + "content": ", we set " + }, + { + "bbox": [ + 46, + 354, + 287, + 606 + ], + "type": "inline_equation", + "content": "K' = 12" + }, + { + "bbox": [ + 46, + 354, + 287, + 606 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 354, + 287, + 606 + ], + "type": "inline_equation", + "content": "K = 3" + }, + { + "bbox": [ + 46, + 354, + 287, + 606 + ], + "type": "text", + "content": ". Our experiments empirically show that, with a proper grid resolution, this design significantly improves the nearest neighbor search speed, and does not introduce noticeable rendering artifacts, even though the " + }, + { + "bbox": [ + 46, + 354, + 287, + 606 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 354, + 287, + 606 + ], + "type": "text", + "content": "-NNs may not be accurate on some of the query points." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 611, + 111, + 623 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 611, + 111, + 623 + ], + "spans": [ + { + "bbox": [ + 47, + 611, + 111, + 623 + ], + "type": "text", + "content": "3.5. Training" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 629, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 629, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 629, + 286, + 713 + ], + "type": "text", + "content": "Only monocular RGB videos are required to train our model. Three losses are used during the training process: (1) a photometric loss that minimizes the " + }, + { + "bbox": [ + 46, + 629, + 286, + 713 + ], + "type": "inline_equation", + "content": "l_{2}" + }, + { + "bbox": [ + 46, + 629, + 286, + 713 + ], + "type": "text", + "content": "-norm distance between the rendered and ground truth pixel colors over all camera rays " + }, + { + "bbox": [ + 46, + 629, + 286, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{r}" + }, + { + "bbox": [ + 46, + 629, + 286, + 713 + ], + "type": "text", + "content": " from all training frames " + }, + { + "bbox": [ + 46, + 629, + 286, + 713 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 629, + 286, + 713 + ], + "type": "text", + "content": ". Formally, we have " + }, + { + "bbox": [ + 46, + 629, + 286, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{rgb}} = \\sum_{i} \\sum_{\\mathbf{r}} \\| \\mathbf{C}_i(\\mathbf{r}) - \\mathbf{I}_i(\\mathbf{r}) \\|_2" + }, + { + "bbox": [ + 46, + 629, + 286, + 713 + ], + "type": "text", + "content": "; (2) a elastic regularization loss " + }, + { + "bbox": [ + 46, + 629, + 286, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{elastic}}" + }, + { + "bbox": [ + 46, + 629, + 286, + 713 + ], + "type": "text", + "content": " applied on the learned error-correction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "text", + "content": "warping field " + }, + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "inline_equation", + "content": "\\mathcal{T}(\\mathbf{q})" + }, + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "text", + "content": ", which is introduced in Nerfies [29]; (3) a magnitude regularization loss to encourage small warping fields, which is defined as " + }, + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{mag}} = \\sum_{\\mathbf{q}}\\| \\mathbf{q} - \\mathcal{T}(\\mathbf{q})\\| _2^2" + }, + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "text", + "content": ". Finally, we combine all three loss terms:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 341, + 130, + 545, + 143 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 341, + 130, + 545, + 143 + ], + "spans": [ + { + "bbox": [ + 341, + 130, + 545, + 143 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\mathcal {L} _ {\\mathrm {r g b}} + \\lambda_ {\\text {e l a s t i c}} \\mathcal {L} _ {\\text {e l a s t i c}} + \\lambda_ {\\text {m a g}} \\mathcal {L} _ {\\text {m a g}}, \\tag {4}", + "image_path": "23d0a465212b37c7019f0b7dcc7f565f0c0029f697a6ca129b51a097da72f3ba.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 152, + 545, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 152, + 545, + 224 + ], + "spans": [ + { + "bbox": [ + 304, + 152, + 545, + 224 + ], + "type": "text", + "content": "where we set " + }, + { + "bbox": [ + 304, + 152, + 545, + 224 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{elastic}} = 10^{-4}" + }, + { + "bbox": [ + 304, + 152, + 545, + 224 + ], + "type": "text", + "content": " at the beginning of the training and decay it to " + }, + { + "bbox": [ + 304, + 152, + 545, + 224 + ], + "type": "inline_equation", + "content": "10^{-5}" + }, + { + "bbox": [ + 304, + 152, + 545, + 224 + ], + "type": "text", + "content": " after 150k iterations, and set " + }, + { + "bbox": [ + 304, + 152, + 545, + 224 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{mag}} = 10^{-2}" + }, + { + "bbox": [ + 304, + 152, + 545, + 224 + ], + "type": "text", + "content": ". To warm start training, we replace the " + }, + { + "bbox": [ + 304, + 152, + 545, + 224 + ], + "type": "inline_equation", + "content": "l_{2}" + }, + { + "bbox": [ + 304, + 152, + 545, + 224 + ], + "type": "text", + "content": "-norm distance in the photometric loss " + }, + { + "bbox": [ + 304, + 152, + 545, + 224 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{rgb}}" + }, + { + "bbox": [ + 304, + 152, + 545, + 224 + ], + "type": "text", + "content": " with the " + }, + { + "bbox": [ + 304, + 152, + 545, + 224 + ], + "type": "inline_equation", + "content": "l_{2}" + }, + { + "bbox": [ + 304, + 152, + 545, + 224 + ], + "type": "text", + "content": " distance for the first 10k iterations. Please refer to the supplementary for more details." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 234, + 387, + 248 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 234, + 387, + 248 + ], + "spans": [ + { + "bbox": [ + 305, + 234, + 387, + 248 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 255, + 545, + 386 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 255, + 545, + 386 + ], + "spans": [ + { + "bbox": [ + 304, + 255, + 545, + 386 + ], + "type": "text", + "content": "In this section, we first introduce the data and metrics used for training and evaluation (Sec. 4.1). Then, we show that our avatar model achieves real-time rendering speed, while producing superior rendering quality on challenging expressions than recent efficient avatars [14, 41, 42] and being comparable to previous high-quality approaches [5] (Sec. 4.2). Finally, we provide ablation studies to justify the design choices and hyper-parameters of our avatar representation, and demonstrate the rendering speed improvements contributed from each of our newly proposed algorithmic components." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 393, + 410, + 405 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 393, + 410, + 405 + ], + "spans": [ + { + "bbox": [ + 305, + 393, + 410, + 405 + ], + "type": "text", + "content": "4.1. Data and Metrics" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 411, + 545, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 411, + 545, + 579 + ], + "spans": [ + { + "bbox": [ + 304, + 411, + 545, + 579 + ], + "type": "text", + "content": "Data. We use monocular RGB videos of multiple subjects (i.e., one video for one subject) to train and evaluate our method, and compare to prior state-of-the-art (SOTA) approaches. Our dataset consists of 10 videos in total, which are a mix of videos captured by us, as well as videos from prior works including PointAvatar [41], INSTA [42], and MonoAvatar [5]. We filter out the background of the videos with off-the-shelf segmentation [24] and matting [21] methods, then crop and resize the videos into a VGA resolution that preserves the original aspect ratio. We compute the camera and 3DMM parameters from the videos following the 3DMM fitting optimization used in INSTA [42]. We reserve a short clip from the end of each video as the testing frames, and use the rest frames for training." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "text", + "content": "Metrics. Following prior arts [13], we use PSNR, SSIM (higher is preferable), and LPIPS (lower is preferable) to measure the image quality. As observed by Zhang et al. [38], LPIPS is a more effective metric in judging the perceptual quality compared to PSNR and SSIM. When computing PSNR and SSIM, we weigh the mean squared error map and the SSIM map with a foreground mask (eroded and smoothed), in order to focus on non-empty areas and avoid the inaccurate foreground segmentation from dominating the metrics." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1979" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 204 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 204 + ], + "type": "text", + "content": "To evaluate the computational cost, we measure the rendering speed in frames-per-second (FPS) on a RTX3090Ti and compare across different approaches with their available implementations. We also estimate the number of FLOPs (floating-point operations) of all methods as the theoretical measurement for the rendering speed. When estimating FLOPs, we fix the contribution of the ray-marching part to 16 points sampled along each camera ray. This simplifies the estimate, since the ray-marching varies the number of FLOPs needed across cameras and scenes, and it applies to all the considered methods." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 214, + 216, + 227 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 214, + 216, + 227 + ], + "spans": [ + { + "bbox": [ + 47, + 214, + 216, + 227 + ], + "type": "text", + "content": "4.2. Comparison to State-of-the-art" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 233, + 287, + 341 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 233, + 287, + 341 + ], + "spans": [ + { + "bbox": [ + 46, + 233, + 287, + 341 + ], + "type": "text", + "content": "We compare our method with several prior works, including: NeRFBlendshape [13], INSTA [42], PointAvatar [41], and MonoAvatar [5]. NeRFBlendshape [13] and INSTA [42] adopt hash encoding [27] into head avatars, leading to efficient renderings. PointAvatar [41] leverages point clouds to represent the head avatar. MonoAvatar [5] is based on a 3DMM-anchored NeRF, and produces high-quality renderings but is slow in speed. We use the same camera and 3DMM parameters to train and test all methods." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 342, + 289, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 342, + 289, + 605 + ], + "spans": [ + { + "bbox": [ + 46, + 342, + 289, + 605 + ], + "type": "text", + "content": "From Tab. 1, ours and INSTA [42] are the only 2 methods that can achieve real-time rendering (i.e., " + }, + { + "bbox": [ + 46, + 342, + 289, + 605 + ], + "type": "inline_equation", + "content": "\\geq 30" + }, + { + "bbox": [ + 46, + 342, + 289, + 605 + ], + "type": "text", + "content": " mean FPS). However, INSTA is quantitatively inferior than our method by a large margin, and gives obvious artifacts in Fig. 4, especially for challenging expressions. PointAvatar [41] has the potential to run in real-time with an optimized implementation thanks to its point cloud representation, but their renderings are overall blurrier than ours, leading to worse quantitative results. Although NeRFBlendshape [14] gives relatively good numbers in Tab. 1, it produces severe artifacts in dynamically changing regions (e.g., mouth and eyebrows in Fig. 4) for several median and large expressions and also gives more floaters, resulting in implausible animations. We highly suggest readers to see the supplementary videos for more comparisons. MonoAvatar [5] gives good rendering qualities and animations, but is one order of magnitudes slower and slightly blurrier on high frequency details such as forehead wrinkles, presumably because that the hash encoding in our model can better capture high frequency contents. Among these compared approaches, our method is the only one that achieves real-time rendering while being one of the best on image qualities." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "text", + "content": "We also compare the theoretical FLOPs of all methods in Tab. 1, where our method requires the least computation mostly because of the smaller MLPs we use (e.g., 2 hidden layers of ours vs. " + }, + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\geq 5" + }, + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "text", + "content": " hidden layers of others). Note that INSTA [42] is implemented with a highly optimized pure " + }, + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\mathrm{C}++" + }, + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "text", + "content": " and CUDA codebase, while other methods use python (tensorflow/pytorch) with customized CUDA kernels. This implementation advantage of INSTA makes it running in a high FPS even with a relatively larger FLOPs." + } + ] + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 307, + 71, + 547, + 148 + ], + "blocks": [ + { + "bbox": [ + 307, + 71, + 547, + 148 + ], + "lines": [ + { + "bbox": [ + 307, + 71, + 547, + 148 + ], + "spans": [ + { + "bbox": [ + 307, + 71, + 547, + 148 + ], + "type": "table", + "html": "
LPIPSSSIMPSNRMean FPSGFLOPs
PointAvatar [41]0.1170.72821.125.0933
INSTA [42]0.1490.75822.1246.2266
NeRFBlendshape [14]0.1100.79322.7711.2223
MonoAvatar [5]0.1140.79822.740.52385
Ours0.1000.79522.7735.9113
", + "image_path": "21032d3b9ac2a6e87ce9fd6ec3c97548436808291297b9d9f706be99286ae919.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 305, + 149, + 545, + 183 + ], + "lines": [ + { + "bbox": [ + 305, + 149, + 545, + 183 + ], + "spans": [ + { + "bbox": [ + 305, + 149, + 545, + 183 + ], + "type": "text", + "content": "Table 1. Quantitative comparisons with state-of-the-art approaches. Our method achieves rendering quality among the best, while supports real-time rendering with a " + }, + { + "bbox": [ + 305, + 149, + 545, + 183 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 305, + 149, + 545, + 183 + ], + "type": "text", + "content": " resolution." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 306, + 194, + 400, + 208 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 194, + 400, + 208 + ], + "spans": [ + { + "bbox": [ + 306, + 194, + 400, + 208 + ], + "type": "text", + "content": "4.3. Ablation Study" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 213, + 545, + 261 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 213, + 545, + 261 + ], + "spans": [ + { + "bbox": [ + 305, + 213, + 545, + 261 + ], + "type": "text", + "content": "In this section we show the impact of the proposed design choices, in particular proving the importance of mesh-anchored hash table blendshapes and the proposed hierarchical " + }, + { + "bbox": [ + 305, + 213, + 545, + 261 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 305, + 213, + 545, + 261 + ], + "type": "text", + "content": "-NN search." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 275, + 512, + 289 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 275, + 512, + 289 + ], + "spans": [ + { + "bbox": [ + 305, + 275, + 512, + 289 + ], + "type": "text", + "content": "4.3.1 Mesh-anchored Hash Table Blendshapes" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 294, + 545, + 330 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 294, + 545, + 330 + ], + "spans": [ + { + "bbox": [ + 304, + 294, + 545, + 330 + ], + "type": "text", + "content": "We hereby investigate alternative design choices and different hyper-parameters to justify the necessity of our mesh-anchored hash table blendshapes." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 331, + 546, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 331, + 546, + 486 + ], + "spans": [ + { + "bbox": [ + 304, + 331, + 546, + 486 + ], + "type": "text", + "content": "Static Hash + 3DMM Param. We first build a naive alternative approach to incorporate hash encoding into 3DMM-anchored NeRF, by attaching a single hash table to each vertex. For a query point, we concatenate the 3DMM pose and expression codes " + }, + { + "bbox": [ + 304, + 331, + 546, + 486 + ], + "type": "inline_equation", + "content": "(\\theta_{i},\\psi_{i})" + }, + { + "bbox": [ + 304, + 331, + 546, + 486 + ], + "type": "text", + "content": " with the embedding pulled from the hash tables, and send them into the MLP. Note that there is no convolution running in UV space. As shown in Tab. 2 and Fig. 5, obvious rendering artifacts show up, and the rendering quality metrics drop significantly compared to our full model. This is presumably because the lightweight MLP, which is crucial for good efficiency, does not have enough capacity to process detailed expression-dependent information from compact 3DMM codes." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 486, + 547, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 486, + 547, + 605 + ], + "spans": [ + { + "bbox": [ + 304, + 486, + 547, + 605 + ], + "type": "text", + "content": "Static Hash + UV CNN. We then increase the model capacity by adding back the UV CNN branch, but use only one hash table per-vertex, i.e., single hash table without blend-shape formulation. As shown in Tab. 2 and Fig. 5, the rendered images show less artifacts over the previous case, but still contain blurry textures and floaters compared to our full model. This demonstrates that the blendshape formulation of mesh-anchored hash tables are necessary in order to obtain good expression dependent local embeddings, leading to a superior rendering quality." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "type": "text", + "content": "Number of Hash Table Blendshapes. Here we investigate how the number of hash table blendshapes per vertex influences the final rendering quality. As shown in Tab. 2, we can see that more blendshapes per vertex leads to higher rendering qualities, which saturates as the number of tables increases. In Fig. 5, increasing the number of blendshapes also gives better details especially for eyelids and ears. Although further adding more blendshapes may produce a better quality, we choose to use 5 blendshapes per-vertex as our" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1980" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 52, + 69, + 545, + 639 + ], + "blocks": [ + { + "bbox": [ + 52, + 69, + 545, + 639 + ], + "lines": [ + { + "bbox": [ + 52, + 69, + 545, + 639 + ], + "spans": [ + { + "bbox": [ + 52, + 69, + 545, + 639 + ], + "type": "image", + "image_path": "7ccb91c29e5b751006d3df318f43de24172b32ed48cc011e3641c1362821c9fe.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 646, + 548, + 691 + ], + "lines": [ + { + "bbox": [ + 46, + 646, + 548, + 691 + ], + "spans": [ + { + "bbox": [ + 46, + 646, + 548, + 691 + ], + "type": "text", + "content": "Figure 4. Comparisons on the rendering quality to previous state-of-the-art methods. From left to right, each column contains the images of: 1) PointAvatar [41], 2) INSTA [42], 3) NeRFBlendshape [14], 4) MonoAvatar [5], 5) Ours, 6) Ground Truth. Our method faithfully reconstructs the personalized expressions and high-frequency details, achieving one of the best rendering quality with real-time rendering speed." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "type": "text", + "content": "1981" + } + ] + } + ], + "index": 2 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 51, + 70, + 545, + 168 + ], + "blocks": [ + { + "bbox": [ + 51, + 70, + 545, + 168 + ], + "lines": [ + { + "bbox": [ + 51, + 70, + 545, + 168 + ], + "spans": [ + { + "bbox": [ + 51, + 70, + 545, + 168 + ], + "type": "image", + "image_path": "75c55b298d7d131cb7a338c066782833fa72f3a48110563b9632b328914526b0.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 171, + 547, + 237 + ], + "lines": [ + { + "bbox": [ + 46, + 171, + 547, + 237 + ], + "spans": [ + { + "bbox": [ + 46, + 171, + 547, + 237 + ], + "type": "text", + "content": "Figure 5. Qualitative results of our full model and various alternative design choices to demonstrate the necessity of our mesh-anchored hash table blendshapes. Please refer to Sec. 4.3.1 for details of each alternative design choice. \"Static Hash + 3DMM Param\" gives unfaithful expressions and obvious artifacts around eyes and ears. \"Static Hash + UV CNN\" produces more faithful expressions, but still suffers from floaters. By utilizing the blendshape formulation of hash tables, \"3 Blendshapes\" eliminates most of the artifacts, but produces blurriness on the details of eyes and ears. \"Ours\" gives the best rendering quality by increasing the number of blendshapes to 5, leading to cleaner details of the eyelids and ear boundary." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 50, + 243, + 545, + 326 + ], + "blocks": [ + { + "bbox": [ + 50, + 243, + 545, + 326 + ], + "lines": [ + { + "bbox": [ + 50, + 243, + 545, + 326 + ], + "spans": [ + { + "bbox": [ + 50, + 243, + 545, + 326 + ], + "type": "image", + "image_path": "b74324f3e847ea6e233f71207889aff1afaf2aa8018d81e238365f14eab65d8e.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 327, + 545, + 361 + ], + "lines": [ + { + "bbox": [ + 46, + 327, + 545, + 361 + ], + "spans": [ + { + "bbox": [ + 46, + 327, + 545, + 361 + ], + "type": "text", + "content": "Figure 6. Qualitative results of using different 3D grid resolutions during hierarchical " + }, + { + "bbox": [ + 46, + 327, + 545, + 361 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 327, + 545, + 361 + ], + "type": "text", + "content": "-NN search and without hierarchical " + }, + { + "bbox": [ + 46, + 327, + 545, + 361 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 327, + 545, + 361 + ], + "type": "text", + "content": "-NN search for ablation studies. \"16 grid resolution\" results in blocky artifacts around mouth and nose. \"32 grid resolution\" removes blocky artifacts, but produces floaters inside mouth. \"64 grid resolution\" produces a similar level of visual quality as \"w/o hierarchical " + }, + { + "bbox": [ + 46, + 327, + 545, + 361 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 327, + 545, + 361 + ], + "type": "text", + "content": "-NN search\"." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 372, + 287, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 372, + 287, + 396 + ], + "spans": [ + { + "bbox": [ + 46, + 372, + 287, + 396 + ], + "type": "text", + "content": "final setting to maintain a relatively small model size and computation cost." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 411, + 192, + 423 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 411, + 192, + 423 + ], + "spans": [ + { + "bbox": [ + 47, + 411, + 192, + 423 + ], + "type": "text", + "content": "4.3.2 Hierarchical " + }, + { + "bbox": [ + 47, + 411, + 192, + 423 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 47, + 411, + 192, + 423 + ], + "type": "text", + "content": "-NN Search" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 430, + 287, + 537 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 430, + 287, + 537 + ], + "spans": [ + { + "bbox": [ + 46, + 430, + 287, + 537 + ], + "type": "text", + "content": "We evaluate our model with and without hierarchical " + }, + { + "bbox": [ + 46, + 430, + 287, + 537 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 430, + 287, + 537 + ], + "type": "text", + "content": "-NN search in terms of speed and quality. From the comparison over rendering speeds (w/: 35.9 FPS; w/o: 26.4 FPS), we can see that hierarchical " + }, + { + "bbox": [ + 46, + 430, + 287, + 537 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 430, + 287, + 537 + ], + "type": "text", + "content": "-NN search gives around " + }, + { + "bbox": [ + 46, + 430, + 287, + 537 + ], + "type": "inline_equation", + "content": "36\\%" + }, + { + "bbox": [ + 46, + 430, + 287, + 537 + ], + "type": "text", + "content": " improvements on the frame rate, which is crucial for achieving real-time rendering. From Fig. 6, we empirically find that enabling hierarchical " + }, + { + "bbox": [ + 46, + 430, + 287, + 537 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 430, + 287, + 537 + ], + "type": "text", + "content": "-NN search will not lead to observable drops on the rendering quality, as long as a proper grid resolution is used (i.e., 64 in our case)." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 538, + 287, + 612 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 538, + 287, + 612 + ], + "spans": [ + { + "bbox": [ + 46, + 538, + 287, + 612 + ], + "type": "text", + "content": "We also investigate the affect of using different 3D grid resolutions during hierarchical " + }, + { + "bbox": [ + 46, + 538, + 287, + 612 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 538, + 287, + 612 + ], + "type": "text", + "content": "-NN search. As shown in Fig. 6, we observe more artifacts around the mouth region when using smaller 3D grid resolutions (i.e., 32 and 16). Therefore, we choose to use 64 resolution in our final setting, which is a good trade-off between quality and speed." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 621, + 117, + 634 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 621, + 117, + 634 + ], + "spans": [ + { + "bbox": [ + 47, + 621, + 117, + 634 + ], + "type": "text", + "content": "5. Discussion" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 641, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 641, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 641, + 288, + 715 + ], + "type": "text", + "content": "We present a high quality 3D neural volumetric head avatar that can be rendered efficiently, while only requires monocular RGB videos for construction. We propose the mesh-anchored hash table blendshapes as our avatar representation, which enable a significantly faster rendering speed by utilizing hash encoding and lightweight MLPs, while still" + } + ] + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 307, + 370, + 545, + 437 + ], + "blocks": [ + { + "bbox": [ + 307, + 370, + 545, + 437 + ], + "lines": [ + { + "bbox": [ + 307, + 370, + 545, + 437 + ], + "spans": [ + { + "bbox": [ + 307, + 370, + 545, + 437 + ], + "type": "table", + "html": "
LPIPSSSIMPSNR
Static Hash + 3DMM Param0.1250.76321.99
1 Blendshape (Static Hash + UV CNN)0.1150.78522.52
3 Blendshapes0.1040.79122.70
5 Blendshapes (Ours)0.1000.79522.77
", + "image_path": "414a31889138248f198c203ddfad31931b450d380e336321e8a9ef91c417dfce.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 441, + 545, + 474 + ], + "lines": [ + { + "bbox": [ + 305, + 441, + 545, + 474 + ], + "spans": [ + { + "bbox": [ + 305, + 441, + 545, + 474 + ], + "type": "text", + "content": "Table 2. Quantitative results for the ablation study. Our full model (i.e., mesh-anchored hash table blendshapes) consistently outperforms alternative design choices discussed in Sec. 4.3.1." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 486, + 545, + 570 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 486, + 545, + 570 + ], + "spans": [ + { + "bbox": [ + 304, + 486, + 545, + 570 + ], + "type": "text", + "content": "maintaining superior controllability to support realistic facial animations, and producing vivid expression-dependent details thank to the local blendshape formulation of hash tables. The experiments indicate that our approach runs in real-time at a " + }, + { + "bbox": [ + 304, + 486, + 545, + 570 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 304, + 486, + 545, + 570 + ], + "type": "text", + "content": " resolution, while giving a rendering quality comparable to state-of-the-art, with better challenging expressions than prior efficient approaches." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 570, + 546, + 704 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 546, + 704 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 546, + 704 + ], + "type": "text", + "content": "As a limitation, we observe floaters under camera viewpoints and expressions that are far from the training distribution, which is a common issue in instant NGPs [27] based approaches. We also notice that performance around the mouth interior regions tends to be less stable because of the relatively poor tracking in these areas on the training data. Fortunately, the fast rendering could enable the possibility to adopt more expensive training strategies, such as regularization terms, adversarial loss, or joint face fitting refinement during the training, which could potentially mitigate these issues and further improve the rendering expressiveness and quality." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1982" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 286, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 53, + 91, + 286, + 122 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 286, + 122 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 286, + 122 + ], + "type": "text", + "content": "[1] MetaHuman - Unreal Engine. https://www.unrealengine.com/en-US/metahuman. Accessed: 2022-10-17. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 125, + 286, + 145 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 125, + 286, + 145 + ], + "spans": [ + { + "bbox": [ + 53, + 125, + 286, + 145 + ], + "type": "text", + "content": "[2] VRChat. https://hello.vrchat.com. Accessed: 2023-02-05.1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 148, + 286, + 178 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 148, + 286, + 178 + ], + "spans": [ + { + "bbox": [ + 53, + 148, + 286, + 178 + ], + "type": "text", + "content": "[3] XHolo Virtual Assistant. https://www.digalix.com/en/virtual-assistant-augmented-reality-xholo. Accessed: 2023-03-01. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 181, + 286, + 213 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 181, + 286, + 213 + ], + "spans": [ + { + "bbox": [ + 53, + 181, + 286, + 213 + ], + "type": "text", + "content": "[4] ShahRukh Athar, Zexiang Xu, Kalyan Sunkavalli, Eli Shechtman, and Zhixin Shu. RigNeRF: Fully Controllable Neural 3D Portraits. pages 20364-20373, 2022. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 214, + 286, + 278 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 214, + 286, + 278 + ], + "spans": [ + { + "bbox": [ + 53, + 214, + 286, + 278 + ], + "type": "text", + "content": "[5] Ziqian Bai, Feitong Tan, Zeng Huang, Kripasindhu Sarkar, Danhang Tang, Di Qiu, Abhimitra Meka, Ruofei Du, Mingsong Dou, Sergio Orts-Escalano, et al. Learning personalized high quality volumetric head avatars from monocular rgb videos. In CVPR, pages 16890-16900, 2023. 1, 2, 3, 4, 5, 6, 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 281, + 286, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 281, + 286, + 335 + ], + "spans": [ + { + "bbox": [ + 53, + 281, + 286, + 335 + ], + "type": "text", + "content": "[6] Thabo Beeler, Fabian Hahn, Derek Bradley, Bernd Bickel, Paul Beardsley, Craig Gotsman, Robert W Sumner, and Markus Gross. High-Quality Passive Facial Performance Capture Using Anchor Frames. In ACM SIGGRAPH 2011 Papers, pages 1-10. 2011. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 337, + 286, + 390 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 337, + 286, + 390 + ], + "spans": [ + { + "bbox": [ + 53, + 337, + 286, + 390 + ], + "type": "text", + "content": "[7] Chen Cao, Tomas Simon, Jin Kyu Kim, Gabe Schwartz, Michael Zollhoefer, Shun-Suke Saito, Stephen Lombardi, Shih-En Wei, Danielle Belko, Shouu-I Yu, Yaser Sheikh, and Jason Saragih. Authentic volumetric avatars from a phone scan. ACM TOG, 2022. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 392, + 286, + 446 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 392, + 286, + 446 + ], + "spans": [ + { + "bbox": [ + 53, + 392, + 286, + 446 + ], + "type": "text", + "content": "[8] Anpei Chen, Zexiang Xu, Andreas Geiger, Jingyi Yu, and Hao Su. Tensorf: Tensorial radiance fields. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXXII, pages 333-350. Springer, 2022. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 447, + 286, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 447, + 286, + 479 + ], + "spans": [ + { + "bbox": [ + 53, + 447, + 286, + 479 + ], + "type": "text", + "content": "[9] Chuhan Chen, Matthew O'Toole, Gaurav Bharaj, and Pablo Garrido. Implicit neural head synthesis via controllable local deformation fields. In CVPR, pages 416-426, 2023. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 481, + 286, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 481, + 286, + 533 + ], + "spans": [ + { + "bbox": [ + 48, + 481, + 286, + 533 + ], + "type": "text", + "content": "[10] Zhiqin Chen, Thomas Funkhouser, Peter Hedman, and Andrea Tagliasacchi. Mobilenerf: Exploiting the polygon rasterization pipeline for efficient neural field rendering on mobile architectures. arXiv preprint arXiv:2208.00277, 2022.3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 536, + 286, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 536, + 286, + 590 + ], + "spans": [ + { + "bbox": [ + 48, + 536, + 286, + 590 + ], + "type": "text", + "content": "[11] Michail Christos Doukas, Mohammad Rami Koujan, Viktoriaa Sharmanska, Anastasios Roussos, and Stefanos Zafeiriou. Head2head++: Deep Facial Attributes Re-Targeting. IEEE Transactions on Biometrics, Behavior, and Identity Science, 3(1):31-43, 2021. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 592, + 286, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 592, + 286, + 645 + ], + "spans": [ + { + "bbox": [ + 48, + 592, + 286, + 645 + ], + "type": "text", + "content": "[12] Bernhard Egger, William AP Smith, Ayush Tewari, Stefanie Wuhrer, Michael Zollhoefer, Thabo Beeler, Florian Bernard, Timo Bolkart, Adam Kortylewski, Sami Romdhani, et al. 3D Morphable Face Models—Past, Present, and Future. 39(5): 1-38, 2020. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 647, + 286, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 647, + 286, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 647, + 286, + 689 + ], + "type": "text", + "content": "[13] Guy Gafni, Justus Thies, Michael Zollhofer, and Matthias Nießner. Dynamic neural radiance fields for monocular 4d facial avatar reconstruction. In CVPR, pages 8649-8658, 2021. 2, 5, 6" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 691, + 286, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 286, + 713 + ], + "type": "text", + "content": "[14] Xuan Gao, Chenglai Zhong, Jun Xiang, Yang Hong, Yudong Guo, and Juyong Zhang. Reconstructing person" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 74, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 326, + 74, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 74, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 326, + 74, + 545, + 106 + ], + "type": "text", + "content": "alized semantic facial nerf models from monocular video. ACM Transactions on Graphics (Proceedings of SIGGRAPH Asia), 41(6), 2022. 1, 2, 3, 4, 5, 6, 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 108, + 545, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 108, + 545, + 194 + ], + "spans": [ + { + "bbox": [ + 308, + 108, + 545, + 194 + ], + "type": "text", + "content": "[15] Kaiwen Guo, Peter Lincoln, Philip Davidson, Jay Busch, Xueming Yu, Matt Whalen, Geoff Harvey, Sergio Orts-Escolano, Rohit Pandey, Jason Dourgarian, Danhang Tang, Anastasia Tkach, Adarsh Kowdle, Emily Cooper, Ming-song Dou, Sean Fanello, Graham Fyffe, Christoph Rhemann, Jonathan Taylor, Paul Debevec, and Shahram Izadi. The Relightables: Volumetric Performance Capture of Humans With Realistic Relighting. ACM TOG, 2019. 1, 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 197, + 545, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 197, + 545, + 228 + ], + "spans": [ + { + "bbox": [ + 308, + 197, + 545, + 228 + ], + "type": "text", + "content": "[16] Wei Jiang, Kwang Moo Yi, Golnoosh Samei, Oncel Tuzel, and Anurag Ranjan. NeuMan: Neural Human Radiance Field From a Single Video. 2022. 5" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 231, + 545, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 231, + 545, + 262 + ], + "spans": [ + { + "bbox": [ + 308, + 231, + 545, + 262 + ], + "type": "text", + "content": "[17] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM TOG, 42(4):1-14, 2023. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 265, + 545, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 265, + 545, + 308 + ], + "spans": [ + { + "bbox": [ + 308, + 265, + 545, + 308 + ], + "type": "text", + "content": "[18] Hyeongwoo Kim, Pablo Garrido, Ayush Tewari, Weipeng Xu, Justus Thies, Matthias Niessner, Patrick Pérez, Christian Richardt, Michael Zollhöfer, and Christian Theobalt. Deep Video Portraits. ACM TOG, 37(4):1-14, 2018. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 310, + 545, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 310, + 545, + 364 + ], + "spans": [ + { + "bbox": [ + 308, + 310, + 545, + 364 + ], + "type": "text", + "content": "[19] Mohammad Rami Koujan, Michail Christos Doukas, Anastasios Roussos, and Stefanos Zafeiriou. Head2head: Video-Based Neural Head Synthesis. In 2020 15th IEEE International Conference on Automatic Face and Gesture Recognition (FG 2020), pages 16-23. IEEE, IEEE, 2020. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 366, + 545, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 366, + 545, + 408 + ], + "spans": [ + { + "bbox": [ + 308, + 366, + 545, + 408 + ], + "type": "text", + "content": "[20] Tianye Li, Timo Bolkart, Michael. J. Black, Hao Li, and Javier Romero. Learning a Model of Facial Shape and Expression From 4D Scans. ACM TOG, 36(6):194:1-194:17, 2017. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 411, + 545, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 411, + 545, + 464 + ], + "spans": [ + { + "bbox": [ + 308, + 411, + 545, + 464 + ], + "type": "text", + "content": "[21] Shanchuan Lin, Linjie Yang, Imran Saleemi, and Soumyadip Sengupta. Robust high-resolution video matting with temporal guidance. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 238-247, 2022. 5" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 468, + 545, + 509 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 468, + 545, + 509 + ], + "spans": [ + { + "bbox": [ + 308, + 468, + 545, + 509 + ], + "type": "text", + "content": "[22] Lingjie Liu, Jiatao Gu, Kyaw Zaw Lin, Tat-Seng Chua, and Christian Theobalt. Neural sparse voxel fields. Advances in Neural Information Processing Systems, 33:15651-15663, 2020. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 512, + 545, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 512, + 545, + 555 + ], + "spans": [ + { + "bbox": [ + 308, + 512, + 545, + 555 + ], + "type": "text", + "content": "[23] Stephen Lombardi, Tomas Simon, Gabriel Schwartz, Michael Zollhoefer, Yaser Sheikh, and Jason Saragih. Mixture of volumetric primitives for efficient neural rendering. ACM TOG, 40(4):1-13, 2021. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 557, + 545, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 557, + 545, + 612 + ], + "spans": [ + { + "bbox": [ + 308, + 557, + 545, + 612 + ], + "type": "text", + "content": "[24] Camillo Lugaresi, Jiuqiang Tang, Hadon Nash, Chris McClanahan, Esha Uboweja, Michael Hays, Fan Zhang, Chuoling Chang, Ming Guang Yong, Juhyun Lee, et al. Mediapipe: A framework for building perception pipelines. arXiv preprint arXiv:1906.08172, 2019. 5" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 614, + 545, + 677 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 614, + 545, + 677 + ], + "spans": [ + { + "bbox": [ + 308, + 614, + 545, + 677 + ], + "type": "text", + "content": "[25] Abhimitra Meka, Rohit Pandey, Christian Haene, Sergio Orts-Escalano, Peter Barnum, Philip David-Son, Daniel Erickson, Yinda Zhang, Jonathan Taylor, Sofien Bouaziz, et al. Deep Relightable Textures: Volumetric Performance Capture With Neural Rendering. ACM TOG, 39(6):1-21, 2020. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 680, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 680, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 680, + 545, + 713 + ], + "type": "text", + "content": "[26] Abhinitra Meka, Rohit Pandey, Christian Haene, Sergio Orts-Escolano, Peter Barnum, Philip Davidson, Daniel Erickson, Yinda Zhang, Jonathan Taylor, Sofien Bouaziz," + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "type": "text", + "content": "1983" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 66, + 72, + 287, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 72, + 287, + 127 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 287, + 127 + ], + "type": "text", + "content": "Chloe Legendre, Wan-Chun Ma, Ryan Overbeck, Thabo Beeler, Paul Debevec, Shahram Izadi, Christian Theobalt, Christoph Rhemann, and Sean Fanello. Deep relightable textures - volumetric performance capture with neural rendering. 2020. 1" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 129, + 287, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 129, + 287, + 171 + ], + "spans": [ + { + "bbox": [ + 48, + 129, + 287, + 171 + ], + "type": "text", + "content": "[27] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM Trans. Graph., 41(4):102:1-102:15, 2022. 1, 2, 3, 4, 6, 8" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 172, + 288, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 172, + 288, + 281 + ], + "spans": [ + { + "bbox": [ + 48, + 172, + 288, + 281 + ], + "type": "text", + "content": "[28] Sergio Orts-Escolano, Christoph Rhemann, Sean Fanello, Wayne Chang, Adarsh Kowdle, Yury Degtyarev, David Kim, Philip Davidson, Sameh Khamis, Mingsong Dou, Vladimir Tankovich, Charles Loop, Qin Cai, Philip Chou, Sarah Mennicken, Julien Valentin, Vivek Pradeep, Shenlong Wang, Sing Bing Kang, Pushmeet Kohli, Yuliya Lutchyn, Cem Keskin, and Shahram Izadi. Holoportation: Virtual 3D Teleportation in Real-Time. In Proceedings of the 29th Annual Symposium on User Interface Software and Technology. ACM, 2016. 1, 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 282, + 287, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 282, + 287, + 326 + ], + "spans": [ + { + "bbox": [ + 48, + 282, + 287, + 326 + ], + "type": "text", + "content": "[29] Keunhong Park, Utkarsh Sinha, Jonathan T Barron, Sofien Bouaziz, Dan B Goldman, Steven M Seitz, and Ricardo Martin-Brualla. Nerfies: Deformable Neural Radiance Fields. pages 5865-5874, 2021. 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 327, + 287, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 327, + 287, + 380 + ], + "spans": [ + { + "bbox": [ + 48, + 327, + 287, + 380 + ], + "type": "text", + "content": "[30] Christian Reiser, Songyou Peng, Yiyi Liao, and Andreas Geiger. Kilonerf: Speeding up neural radiance fields with thousands of tiny mlps. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14335-14345, 2021. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 382, + 287, + 414 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 382, + 287, + 414 + ], + "spans": [ + { + "bbox": [ + 48, + 382, + 287, + 414 + ], + "type": "text", + "content": "[31] Cheng Sun, Min Sun, and Hwann-Tzong Chen. Direct voxel grid optimization: Super-fast convergence for radiance fields reconstruction. In CVPR, 2022. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 415, + 287, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 415, + 287, + 470 + ], + "spans": [ + { + "bbox": [ + 48, + 415, + 287, + 470 + ], + "type": "text", + "content": "[32] Ayush Tewari, Justus Thies, Ben Mildenhall, Pratul Srinivasan, Edgar Tretschk, Wang Yifan, Christoph Lassner, Vincent Sitzmann, Ricardo Martin-Brualla, Stephen Lombardi, et al. Advances in neural rendering. In Computer Graphics Forum, pages 703-735. Wiley Online Library, 2022. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 471, + 287, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 471, + 287, + 491 + ], + "spans": [ + { + "bbox": [ + 48, + 471, + 287, + 491 + ], + "type": "text", + "content": "[33] Zach Waggoner. My Avatar, My Self: Identity in Video Role-Playing Games. McFarland, 2009. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 492, + 287, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 492, + 287, + 547 + ], + "spans": [ + { + "bbox": [ + 48, + 492, + 287, + 547 + ], + "type": "text", + "content": "[34] Lizhen Wang, Xiaochen Zhao, Jingxiang Sun, Yuxiang Zhang, Hongwen Zhang, Tao Yu, and Yebin Liu. Stylea-vatar: Real-time photo-realistic portrait avatar from a single video. In ACM SIGGRAPH 2023 Conference Proceedings, pages 1-10, 2023. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 548, + 287, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 548, + 287, + 601 + ], + "spans": [ + { + "bbox": [ + 48, + 548, + 287, + 601 + ], + "type": "text", + "content": "[35] Yuelang Xu, Lizhen Wang, Xiaochen Zhao, Hongwen Zhang, and Yebin Liu. Avatarmav: Fast 3d head avatar reconstruction using motion-aware neural voxels. In ACM SIGGRAPH 2023 Conference Proceedings, pages 1-10, 2023. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 603, + 287, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 603, + 287, + 658 + ], + "spans": [ + { + "bbox": [ + 48, + 603, + 287, + 658 + ], + "type": "text", + "content": "[36] Yuelang Xu, Hongwen Zhang, Lizhen Wang, Xiaochen Zhao, Han Huang, Guojun Qi, and Yebin Liu. Latentavatar: Learning latent expression code for expressive neural head avatar. In ACM SIGGRAPH 2023 Conference Proceedings, pages 1-10, 2023. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 658, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 287, + 712 + ], + "type": "text", + "content": "[37] Alex Yu, Ruilong Li, Matthew Tancik, Hao Li, Ren Ng, and Angjoo Kanazawa. Plenoctrees for real-time rendering of neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5752-5761, 2021. 2" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 330 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "text", + "content": "[38] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The Unreasonable Effectiveness of Deep Features As a Perceptual Metric. pages 586-595, 2018. 5" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 307, + 107, + 545, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 107, + 545, + 150 + ], + "spans": [ + { + "bbox": [ + 307, + 107, + 545, + 150 + ], + "type": "text", + "content": "[39] Xiaochen Zhao, Lizhen Wang, Jingxiang Sun, Hongwen Zhang, Jinli Suo, and Yebin Liu. Havatar: High-fidelity head avatar via facial model conditioned neural radiance field. ACM TOG, 43(1):1-16, 2023. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 152, + 545, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 152, + 545, + 195 + ], + "spans": [ + { + "bbox": [ + 307, + 152, + 545, + 195 + ], + "type": "text", + "content": "[40] Yufeng Zheng, Victoria Fernández Abrevaya, Marcel C Bühler, Xu Chen, Michael J Black, and Otmar Hilliges. Im avatar: Implicit morphable head avatars from videos. In CVPR, pages 13545-13555, 2022. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 197, + 545, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 197, + 545, + 239 + ], + "spans": [ + { + "bbox": [ + 307, + 197, + 545, + 239 + ], + "type": "text", + "content": "[41] Yufeng Zheng, Wang Yifan, Gordon Wetzstein, Michael J Black, and Otmar Hilliges. Pointavatar: Deformable point-based head avatars from videos. In CVPR, pages 21057-21067, 2023. 1, 2, 3, 5, 6, 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 241, + 545, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 241, + 545, + 274 + ], + "spans": [ + { + "bbox": [ + 307, + 241, + 545, + 274 + ], + "type": "text", + "content": "[42] Wojciech Zielonka, Timo Bolkart, and Justus Thies. Instant volumetric head avatars. In CVPR, pages 4574-4584, 2023. 1, 2, 3, 4, 5, 6, 7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 275, + 545, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 275, + 545, + 330 + ], + "spans": [ + { + "bbox": [ + 308, + 275, + 545, + 330 + ], + "type": "text", + "content": "[43] Michael Zollhöfer, Justus Thies, Pablo Garrido, Derek Bradley, Thabo Beeler, Patrick Pérez, Marc Stamminger, Matthias Nießner, and Christian Theobalt. State of the Art on Monocular 3D Face Reconstruction, Tracking, and Applications. pages 523-550, 2018. 2" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1984" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Efficient Dataset Distillation via Minimax Diffusion/75146b44-b76a-42b0-8456-4c37446317b7_content_list.json b/2024/Efficient Dataset Distillation via Minimax Diffusion/75146b44-b76a-42b0-8456-4c37446317b7_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..209052846b4bfe6660cfe316a116214c15e1e654 --- /dev/null +++ b/2024/Efficient Dataset Distillation via Minimax Diffusion/75146b44-b76a-42b0-8456-4c37446317b7_content_list.json @@ -0,0 +1,1625 @@ +[ + { + "type": "text", + "text": "Efficient Dataset Distillation via Minimax Diffusion", + "text_level": 1, + "bbox": [ + 225, + 130, + 743, + 150 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Jianyang Gu1 Saeed Vahidian2 Vyacheslav Kungurtsev3", + "bbox": [ + 246, + 179, + 720, + 198 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Haonan Wang $^{4}$ Wei Jiang $^{1*}$ Yang You $^{4}$ Yiran Chen $^{2}$", + "bbox": [ + 251, + 198, + 697, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Zhejiang University $^{2}$ Duke University $^{3}$ Czech Technical University", + "bbox": [ + 200, + 215, + 777, + 234 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{4}$ National University of Singapore", + "bbox": [ + 349, + 234, + 620, + 252 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{gu_jianyang, jiangwei_zju}@zju.edu.cn", + "bbox": [ + 318, + 253, + 645, + 268 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 304, + 312, + 319 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Dataset distillation reduces the storage and computational consumption of training a network by generating a small surrogate dataset that encapsulates rich information of the original large-scale one. However, previous distillation methods heavily rely on the sample-wise iterative optimization scheme. As the images-per-class (IPC) setting or image resolution grows larger, the necessary computation will demand overwhelming time and resources. In this work, we intend to incorporate generative diffusion techniques for computing the surrogate dataset. Observing that key factors for constructing an effective surrogate dataset are representativeness and diversity, we design additional minimax criteria in the generative training to enhance these facets for the generated images of diffusion models. We present a theoretical model of the process as hierarchical diffusion control demonstrating the flexibility of the diffusion process to target these criteria without jeopardizing the faithfulness of the sample to the desired distribution. The proposed method achieves state-of-the-art validation performance while demanding much less computational resources. Under the 100-IPC setting on ImageWoof, our method requires less than one-twentieth the distillation time of previous methods, yet yields even better performance. Source code and generated data are available in https://github.com/vimar-gu/MinimaxDiffusion.", + "bbox": [ + 76, + 335, + 472, + 713 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 741, + 209, + 756 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Data, as a necessary resource for deep learning, has concurrently promoted algorithmic advancements while imposing challenges on researchers due to heavy demands on storage and computational resources [6, 10, 18, 47]. Confronted with the conflict between the requirement for high-precision", + "bbox": [ + 75, + 766, + 468, + 843 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/552d665e9aa78a478bcd3eadac76eca61fc7600fb0dcb58363cc269e51885137.jpg", + "image_caption": [ + "Figure 1. The validation accuracy and distillation time of different methods on ImageWoof [15], with a number following each method denoting the Image-Per-Class (IPC) setting. Previous methods are restricted by the heavier running time and memory consumption as IPC grows larger. In comparison, our proposed method notably reduces the demanding computational resources and also achieves state-of-the-art validation performance." + ], + "image_footnote": [], + "bbox": [ + 514, + 305, + 875, + 503 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "models and overwhelming resource demands, dataset distillation is proposed to condense the rich information of a large-scale dataset into a small surrogate one [5, 21, 47, 57]. Such a surrogate dataset is expected to achieve training performance comparable to that attained with the original one.", + "bbox": [ + 496, + 642, + 893, + 718 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Previous dataset distillation methods mostly engage in iterative optimization on fixed-number samples at the pixel level [21, 26, 30, 31, 44, 54, 57] or embedding level [4, 55]. However, the sample-wise iterative optimization scheme suffers from problems of two perspectives. (1) The parameter space of optimization is positively correlated with the size of the target surrogate dataset and the image resolution [3, 57]. Consequently, substantial time and computational resources are required for distilling larger datasets. As shown in Fig. 1, IDC-1 [21] takes over 90 hours to distill a 100-image-per-class (IPC) set from ImageWoof [15], while training on ImageWoof itself only requires a matter of", + "bbox": [ + 496, + 719, + 893, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 44 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author", + "bbox": [ + 94, + 851, + 220, + 863 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "This work has received funding from the European Union's Horizon Europe research and innovation program under grant agreement No. 101084642.", + "bbox": [ + 76, + 863, + 468, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "15793", + "bbox": [ + 478, + 924, + 517, + 936 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/d1a04696a1ff47117b25ec1488c8a7523afe466bfb34236d84cdbc1ed99caf80.jpg", + "image_caption": [ + "Figure 2. Sample images distilled by the pixel-level sample-wise optimization method DM [56] on ImageWoof. As the parameter space increases along with the Image-Per-Class (IPC) setting, with the same initialization, the appearance disparity between original and distilled images is smaller." + ], + "image_footnote": [], + "bbox": [ + 83, + 88, + 460, + 260 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "hours. (2) The expanded parameter space also increases the optimization complexity. As shown in Fig. 2, while distillation yields significant information condensation under small IPC settings, the pixel modification diminishes when distilling larger-IPC datasets. The reduced disparity also leads to smaller performance gain compared with original images, with instances where the distilled set even performs worse. Especially when distilling data of fine-grained classes, the sample-wise optimization scheme fails to provide adequate discriminative information. These constraints severely hinder individual researchers from distilling personalized data. A more practical training scheme is urgently needed to facilitate the broader application of dataset distillation.", + "bbox": [ + 75, + 369, + 467, + 566 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this work, we explore the possibility of incorporating generative diffusion techniques [20, 22, 32] to efficiently compute effective surrogate datasets. We first conduct empirical analysis on the suitability of data generated by raw diffusion models for training networks. Based on the observations, we conclude that constructing an effective surrogate dataset hinges on two key factors: representativeness and diversity. Accordingly, we design extra minimax criteria for the generative training to enhance the capability of generating more effective surrogate datasets without explicit prompt designs. The minimax criteria involve two aspects: enforcing the generated sample to be close to the farthest real sample, while being far away from the most similar generated one. We provide theoretical analysis to support that the proposed minimax scheme aims to solve a well defined problem with all the criteria, including the generative accuracy and the minimax criteria, can be targeted simultaneously without detriment to the others.", + "bbox": [ + 75, + 566, + 467, + 838 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Compared with the astronomical training time consumption of the sample-wise iterative optimization schemes, the proposed method takes less than 1 hour to distill a 100-IPC surrogate dataset for a 10-class ImageNet subset, including", + "bbox": [ + 76, + 839, + 467, + 900 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/d0d68cd486e76069a23a5fac587f4be503f4f213d019b2eacbb53e1e11581cc9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 506, + 90, + 689, + 213 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/92b58d4e47e34ee50d933a4ead78330b9d2721a6dd86e8615878255391c2d23a.jpg", + "image_caption": [ + "Figure 3. The feature distribution comparison of different image generation methods with the original set. The validation performance of each surrogate set is listed in the upper-right corner." + ], + "image_footnote": [], + "bbox": [ + 504, + 214, + 687, + 335 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/16b215162ea890c298ab7256ef36a046addb177a11233a288ec15eef4ba2983d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 696, + 90, + 880, + 213 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/1265cd04ab1e8d2ea1b2eecd3819ac9f56eacc6a98787ac2b8810a85d79ff094.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 696, + 214, + 880, + 335 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "the fine-tuning and image generation processes. Remarkably, the GPU consumption remains consistent across all IPC settings. Furthermore, the distilled surrogate dataset attains superior validation performance compared with other state-of-the-art methods. Especially on the challenging fine-grained ImageWoof subset, the proposed method outperforms the second-best DD method by $5.5\\%$ and $8.1\\%$ under the IPC settings of 70 and 100, respectively. The source code is provided in the supplementary material.", + "bbox": [ + 496, + 421, + 890, + 556 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The contributions of this work are summarized into:", + "bbox": [ + 517, + 559, + 864, + 571 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We analyze the data generated by diffusion models, and emphasize the importance of representativeness and diversity for constructing effective surrogate datasets.", + "- We propose a novel dataset distillation scheme based on extra minimax criteria for diffusion models targeting the representativeness and diversity of generated data.", + "- We theoretically justify the proposed minimax criteria as enforceable without trade-offs in the generation quality of the individual data points.", + "- We conduct extensive experiments to validate that our proposed method achieves state-of-the-art performance while demanding significantly reduced training time in comparison to previous dataset distillation methods." + ], + "bbox": [ + 500, + 574, + 890, + 768 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Method", + "text_level": 1, + "bbox": [ + 500, + 787, + 589, + 803 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Problem Definition", + "text_level": 1, + "bbox": [ + 500, + 814, + 681, + 828 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The general purpose of dataset distillation is to generate a small surrogate dataset $S = \\{(\\mathbf{x}_i,y_i)\\}_{i = 1}^{N_S}$ from a largescale one $\\mathcal{T} = \\{(\\mathbf{x}_i,y_i)\\}_{i = 1}^{N_T}$ [47, 57]. Here each $\\mathbf{x}_i$ denotes an image with a corresponding class label $y_{i}$ , and", + "bbox": [ + 498, + 839, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "15794", + "bbox": [ + 480, + 924, + 517, + 935 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "$N_{S} \\ll N_{T}$ . The surrogate dataset $S$ is expected to encapsulate substantial information from the original $\\mathcal{T}$ , such that training a model on $S$ achieves performance comparable with that on $\\mathcal{T}$ . After distilling, we train network models on $S$ and validate the performance on the original test set.", + "bbox": [ + 76, + 90, + 468, + 167 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. Diffusion for Distillation", + "text_level": 1, + "bbox": [ + 76, + 174, + 302, + 189 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Diffusion models learn a dataset distribution by gradually adding Gaussian noise to images and reversing back. Taking the latent diffusion model (LDM) as an example, given a training image $\\mathbf{x}$ , the training process is separated into two parts. An encoder $E$ transforms the image into the latent space $\\mathbf{z} = E(\\mathbf{x})$ and a decoder $D$ reconstructs a latent code back to the image space $\\hat{\\mathbf{x}} = D(\\mathbf{z})$ . The forward noisng process gradually adds noise $\\epsilon \\sim \\mathcal{N}(\\mathbf{0},\\mathbf{I})$ to the original latent code $\\mathbf{z}_0$ : $\\mathbf{z}_t = \\sqrt{\\bar{\\alpha}_t}\\mathbf{z}_0 + \\sqrt{1 - \\bar{\\alpha}_t\\epsilon}$ , where $\\bar{\\alpha}_t$ is a hyper-parameter. Provided with a conditioning vector $\\mathbf{c}$ encoded with class labels, the diffusion models are trained by the squared error between the predicted noise $\\epsilon_{\\theta}(\\mathbf{z}_t,\\mathbf{c})$ and the ground truth $\\epsilon$ :", + "bbox": [ + 75, + 198, + 470, + 393 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {s i m p l e}} = \\left| \\left| \\epsilon_ {\\theta} \\left(\\mathbf {z} _ {t}, \\mathbf {c}\\right) - \\epsilon \\right| \\right| _ {2} ^ {2}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 176, + 402, + 468, + 422 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\epsilon_{\\theta}$ is a noise prediction network parameterized by $\\theta$ . Diffusion models are proven to generate images of higher quality compared with GANs [8]. There are also some Parameter Efficient Fine-Tuning (PEFT) methods updating a small number of model parameters in order for the model to be better adapted to specific data domains [34, 50]. We adopt DiT [33] as the baseline and Difffit [50] as the naive fine-tuning method for image generation. The generated images are compared with the original data from the perspective of embedding distribution in Fig. 3.", + "bbox": [ + 75, + 430, + 468, + 580 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The samples of random selection and pre-trained diffusion models present two extreme ends of the distribution. Random selection faithfully reflects the original distribution, yet fails to emphasize some high-density regions. In contrast, diffusion models are over-fitted to those dense areas, leaving a large part of the original distribution uncovered. We attribute these two distributions to two properties, respectively. The randomly selected data holds extraordinary diversity, and the diffusion-generated data shows representativeness to the original distribution. We claim that both properties are essential for constructing an effective surrogate dataset. By naive fine-tuning, Difffit better captures the representative regions but leaves more regions uncovered. To this end, we propose extra minimax criteria for the diffusion model to enhance both of the properties.", + "bbox": [ + 75, + 582, + 468, + 809 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.3. Minimax Diffusion Criteria", + "text_level": 1, + "bbox": [ + 76, + 816, + 326, + 832 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Based on the observation that representativeness and diversity are two key factors to construct an effective surrogate dataset, we accordingly design extra minimax criteria to enhance these two essential properties for the diffusion model.", + "bbox": [ + 75, + 839, + 468, + 900 + ], + "page_idx": 2 + }, + { + "type": "code", + "sub_type": "algorithm", + "code_caption": [ + "Algorithm 1: Minimax Diffusion Fine-tuning" + ], + "code_body": "Input: initialized model parameter $\\theta$ original dataset $\\mathcal{T} = \\{(\\mathbf{x},y)\\}$ , encoder $E$ class encoder $E_{c}$ time step $t$ variance schedule $\\bar{\\alpha}_t$ real embedding memory $\\mathcal{M}$ predicted embedding memory $\\mathcal{D}$ \nOutput: optimized model parameter $\\theta^{*}$ \nfor each step do Obtain the original embedding: $\\mathbf{z}_0 = E(\\mathbf{x})$ Obtain the class embedding: $\\mathbf{c} = E_c(y)$ Sample random noise: $\\epsilon \\sim \\mathcal{N}(\\mathbf{0},\\mathbf{I})$ Add noise to the embedding: $\\mathbf{z}_t = \\sqrt{\\bar{\\alpha}_t}\\mathbf{z}_0 + \\sqrt{1 - \\bar{\\alpha}_t}\\epsilon$ Predict the noise $\\epsilon_{\\theta}(\\mathbf{z}_{t},\\mathbf{c})$ and recovered embedding $\\hat{\\mathbf{z}}_{\\theta}(\\mathbf{z}_{t},\\mathbf{c}) = \\mathbf{z}_{t} - \\epsilon_{\\theta}(\\mathbf{z}_{t},\\mathbf{c})$ Update the model parameter with Eq. (5) Enqueue the real embedding: $\\mathcal{M}_r\\gets \\mathbf{z}_0$ Enqueue the predicted embedding: $\\mathcal{M}_d\\gets \\hat{\\mathbf{z}}_\\theta (\\mathbf{z}_t,\\mathbf{c})$ \nend", + "bbox": [ + 506, + 111, + 890, + 349 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Representativeness It is essential for the small surrogate dataset to sufficiently represent the original data. A naive approach to improve the representativeness is aligning the embedding distribution between synthetic and real samples:", + "bbox": [ + 498, + 383, + 890, + 444 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {r} = \\underset {\\theta} {\\arg \\max } \\sigma \\left(\\hat {\\mathbf {z}} _ {\\theta} \\left(\\mathbf {z} _ {t}, \\mathbf {c}\\right), \\frac {1}{N _ {B}} \\sum_ {i = 0} ^ {N _ {B}} \\mathbf {z} _ {i}\\right), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 552, + 454, + 890, + 496 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\sigma (\\cdot ,\\cdot)$ is the cosine similarity, $\\hat{\\mathbf{z}}_{\\theta}(\\mathbf{z}_t,\\mathbf{c})$ is the predicted original embedding by subtracting the noise from the noisy embedding $\\hat{\\mathbf{z}}_{\\theta}(\\mathbf{z}_t,\\mathbf{c}) = \\mathbf{z}_t - \\epsilon_{\\theta}(\\mathbf{z}_t,\\mathbf{c})$ , and $N_B$ is the size of the sampled real sample mini-batch. However, the naive alignment tends to draw the predicted embedding towards the center of the real distribution, which severely limits the diversity. Therefore, we propose to maintain an auxiliary memory $\\mathcal{M} = \\{\\mathbf{z}_m\\}_{m = 1}^{N_M}$ to store the real samples utilized in adjacent iterations, and design a minimax optimization objective as:", + "bbox": [ + 496, + 502, + 890, + 654 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {r} = \\arg \\max _ {\\theta} \\min _ {m \\in \\left[ N _ {M} \\right]} \\sigma \\left(\\hat {\\mathbf {z}} _ {\\theta} \\left(\\mathbf {z} _ {t}, \\mathbf {c}\\right), \\mathbf {z} _ {m}\\right). \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 557, + 665, + 890, + 690 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "By pulling close the least similar sample pairs, the diffusion model is encouraged to generate images that better cover the original distribution. It is notable that the diffusion training objective $\\mathcal{L}_{simple}$ itself encourages the generated images to resemble the original ones. Thus, the minimax criterion allows the preservation of diversity to the maximum extent.", + "bbox": [ + 496, + 700, + 890, + 791 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Diversity Although the pre-trained diffusion models already achieve satisfactory generation quality, the remaining defect is limited diversity compared with the original data, as shown in Fig. 3. We expect the data generated by the diffusion model can accurately reflect the original distribution, while simultaneously being different from", + "bbox": [ + 496, + 809, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "15795", + "bbox": [ + 480, + 925, + 517, + 936 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "each other. Hence, we maintain another auxiliary memory $\\mathcal{D} = \\{\\mathbf{z}_d\\}_{d=1}^{N_D}$ for the predicted embeddings of adjacent iterations and design another minimax objective to explicitly enhance the sample diversity as:", + "bbox": [ + 76, + 90, + 468, + 151 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {d} = \\arg \\min _ {\\theta} \\max _ {d \\in [ N _ {D} ]} \\sigma (\\hat {\\mathbf {z}} _ {\\theta} (\\mathbf {z} _ {t}, \\mathbf {c}), \\mathbf {z} _ {d}). \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 142, + 160, + 468, + 185 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The diversity term has an opposite optimization target compared with the representativeness term, where the predicted embedding is pushed away from the most similar one stored in the memory bank. Although diversity is essential for an effective surrogate set, too much of it will cause the generated data to lose representativeness. The proposed minimax optimization enhances the diversity in a gentle way, with less influence on the class-related features.", + "bbox": [ + 75, + 191, + 468, + 311 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Combining all the components, we summarize the training process in Algorithm 1. The complete training objective can be formulated as:", + "bbox": [ + 75, + 313, + 468, + 357 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\mathcal {L} _ {\\text {s i m p l e}} + \\lambda_ {r} \\mathcal {L} _ {r} + \\lambda_ {d} \\mathcal {L} _ {d}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 169, + 368, + 468, + 383 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\lambda_r$ and $\\lambda_d$ are weighting hyper-parameters.", + "bbox": [ + 76, + 393, + 406, + 407 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3. Theoretical Analysis", + "text_level": 1, + "bbox": [ + 76, + 420, + 272, + 436 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Assume that $\\mu$ is the real distribution of the latent variables $\\mathbf{z}$ associated with the target dataset $\\mathcal{T}$ . We rewrite the optimization problem presented in Eq. (5) in a modified form:", + "bbox": [ + 76, + 445, + 468, + 491 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\min_{\\{\\theta^{(i)}\\}_{i\\in [N_{D}]}}\\lambda_{d}\\max_{i,j = 1,\\ldots ,N_{D}}\\sigma \\left(\\hat{\\mathbf{z}} (\\theta^{(i)}),\\hat{\\mathbf{z}} (\\theta^{(j)})\\right) \\\\ + \\sum_ {i = 1} ^ {N _ {D}} \\left\\{- \\lambda_ {r} Q _ {\\tilde {q}, w \\sim \\mu} \\left[ \\sigma \\left(\\hat {\\mathbf {z}} \\left(\\theta^ {(i)}\\right), w\\right) \\right] + \\| \\hat {\\mathbf {z}} \\left(\\theta^ {(i)}\\right) - \\mathbf {z} _ {0} ^ {(i)} \\| ^ {2} \\right\\}, \\tag {6} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 84, + 510, + 470, + 579 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $Q_{\\tilde{q}}[\\cdot]$ denotes the quantile function with $\\tilde{q}$ as the quantile percentage. Note that here we consider a theoretical idealized variant of our algorithm wherein we perform simultaneous generation of all the embeddings $\\{\\hat{\\mathbf{z}}(\\theta^{(i)})\\}$ , rather than sample by sample. Hence the objectives turn to the sum of pairwise similarities rather than the form in Eq. (4). And we minimize the negative to aim for maximal representativeness, as in Eq. (5).", + "bbox": [ + 75, + 580, + 468, + 700 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "It can be considered as a scalarized solution to a multi-objective optimization problem, wherein multiple criteria are weighed (see, e.g. [13]). This perspective aligns with a Pareto front with trade-offs. It means that one objective decreasing will by necessity result in another increasing.", + "bbox": [ + 75, + 702, + 468, + 777 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "However, consider that any solution to the following tri-level optimization problem is also a solution for Eq. (6):", + "bbox": [ + 76, + 777, + 468, + 808 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\min _ {\\{\\theta^ {(i)} \\} _ {i \\in [ N _ {D} ]}} \\max _ {i, j = 1,.., N _ {D}} \\sigma \\left(\\hat {\\mathbf {z}} (\\theta^ {(i)}), \\hat {\\mathbf {z}} (\\theta^ {(j)})\\right) \\\\ \\text {s u b j . t o} \\quad \\left\\{\\theta^ {(i)} \\right\\} \\in \\arg \\max \\sum_ {i = 1} ^ {N _ {D}} Q _ {\\tilde {q}, w \\sim \\mu} \\left[ \\sigma (\\hat {\\mathbf {z}} (\\theta^ {(i)}), w) \\right] \\\\ \\text {s u b j . t o} \\quad \\theta^ {(i)} \\in \\arg \\min \\| \\hat {\\mathbf {z}} (\\theta) - \\mathbf {z} _ {0} ^ {(i)} \\| ^ {2}, \\forall i \\in [ N _ {D} ]. \\tag {7} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 84, + 814, + 468, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "If a solution to Eq. (7) is discovered, either incidentally through solving Eq. (6) or by careful tuning of step sizes, the set of minimizers will be sufficiently large at both levels, with no trade-offs involved. However, can we justify the presumption that there exists a meaningful set of potential minimizers?", + "bbox": [ + 496, + 90, + 890, + 180 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Diffusion Process Model One popular framework for the mathematical analysis of diffusion involves analyzing the convergence and asymptotic properties of, appropriately homonymous, diffusion processes. These processes are characterized by the standard stochastic differential equation with a drift and diffusion term. For a time-dependent random variable $Z_{t}$ ,", + "bbox": [ + 496, + 199, + 890, + 304 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nd Z _ {t} = V \\left(Z _ {t}\\right) d t + d W _ {t} \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 616, + 310, + 890, + 327 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $V$ is a drift function dependent on the current $Z_{t}$ and $dW_{t}$ is a Wiener (Brownian noise) process. This equation serves as an appropriate continuous approximation of generative diffusion, given that Brownian noise is a continuous limit of adding normal random variables. Consequently, we aim for any realization $\\mathbf{z} \\sim Z_{t}$ to have certain desired properties that reflect generative modeling with high probability.", + "bbox": [ + 496, + 333, + 890, + 436 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The work [43] established a theoretical model utilizing concepts in the control of these diffusions, demonstrating how it can result in sampling from the distribution of a desired data set. In the supplementary material we present a description of their framework and present an argument supporting the well-defined nature of the following problem, indicating that it has non-trivial solutions.", + "bbox": [ + 496, + 438, + 890, + 542 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "When sampling optimally from the population dataset, we consider a stochastic control problem wherein $V$ depends also on some chosen control $u(\\mathbf{z},t)$ . This control aims to find the most representative samples and, among the possible collection of such samples, to obtain the most diverse one while sampling from the desired dataset $\\mu$ . This involves solving:", + "bbox": [ + 496, + 544, + 890, + 648 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {u (x, t)} \\max _ {i, j = 1,.., N _ {D}} \\sigma \\left(Z _ {1} ^ {u, (i)}, Z _ {1} ^ {u, (j)}\\right)\n$$\n", + "text_format": "latex", + "bbox": [ + 524, + 654, + 751, + 679 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathrm {s u b j . t o} \\quad u \\in \\arg \\max \\sum_ {i = 1} ^ {N _ {D}} \\int_ {0} ^ {1} \\mathbb {E} _ {Z _ {t} ^ {(i)}} Q _ {\\tilde {q}, w \\sim \\mu} \\left[ \\sigma \\left(Z _ {t} ^ {(i)}, w\\right) \\right] d s \\\\ Z _ {1} \\sim \\mu , \\\\ d Z _ {t} ^ {u, (i)} = u \\left(Z _ {t} ^ {u, (i)}, t\\right) d t + d W _ {t}, t \\in [ 0, 1 ]; \\\\ Z _ {0} = \\mathbf {z} _ {0}. \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 509, + 680, + 888, + 750 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "(9)", + "bbox": [ + 870, + 751, + 890, + 763 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "This problem poses a bi-level stochastic control challenge where employing a layered dynamic programming is far from tractable. Additionally, a multi-stage stochastic programming approximation would also be infeasible given the scale of the datasets involved. Instead, we opt for parameterization with a neural network, forego exact sampling, discretize the problem, redefine the criteria to be time independent, and seek to solve an approximate solution for the tri-level optimization problem Eq. (7).", + "bbox": [ + 496, + 763, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "15796", + "bbox": [ + 478, + 924, + 517, + 936 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/939fb4f0c2c36773432cb3723f5d6bb26c59ba228c09d8ee5169e0049bce1403.jpg", + "table_caption": [ + "Table 1. Performance comparison with pre-trained diffusion models and other state-of-the-art methods on ImageWoof. All the results are reproduced by us on the $256\\times 256$ resolution. The missing results are due to out-of-memory. The best results are marked as bold." + ], + "table_footnote": [], + "table_body": "
IPC (Ratio)Test ModelRandomK-Center [37]Herding [48]DiT [33]DM [56]IDC-1 [21]GLaD [4]OursFull
10 (0.8%)ConvNet-624.3±1.119.4±0.926.7±0.534.2±1.126.9±1.233.3±1.133.8±0.937.0±1.086.4±0.2
ResNetAP-1029.4±0.822.1±0.132.0±0.334.7±0.530.3±1.239.1±0.532.9±0.939.2±1.387.5±0.5
ResNet-1827.7±0.921.1±0.430.2±1.234.7±0.433.4±0.737.3±0.231.7±0.837.6±0.989.3±1.2
20 (1.6%)ConvNet-629.1±0.721.5±0.829.5±0.336.1±0.829.9±1.035.5±0.8-37.6±0.286.4±0.2
ResNetAP-1032.7±0.425.1±0.734.9±0.141.1±0.835.2±0.643.4±0.3-45.8±0.587.5±0.5
ResNet-1829.7±0.523.6±0.332.2±0.640.5±0.529.8±1.738.6±0.2-42.5±0.689.3±1.2
50 (3.8%)ConvNet-641.3±0.636.5±1.040.3±0.746.5±0.844.4±1.043.9±1.2-53.9±0.686.4±0.2
ResNetAP-1047.2±1.340.6±0.449.1±0.749.3±0.247.1±1.148.3±1.0-56.3±1.087.5±0.5
ResNet-1847.9±1.839.6±1.048.3±1.250.1±0.546.2±0.648.3±0.8-57.1±0.689.3±1.2
70 (5.4%)ConvNet-646.3±0.638.6±0.746.2±0.650.1±1.247.5±0.848.9±0.7-55.7±0.986.4±0.2
ResNetAP-1050.8±0.645.9±1.553.4±1.454.3±0.951.7±0.852.8±1.8-58.3±0.287.5±0.5
ResNet-1852.1±1.044.6±1.149.7±0.851.5±1.051.9±0.851.1±1.7-58.8±0.789.3±1.2
100 (7.7%)ConvNet-652.2±0.445.1±0.554.4±1.153.4±0.355.0±1.353.2±0.9-61.1±0.786.4±0.2
ResNetAP-1059.4±1.054.8±0.261.7±0.958.3±0.856.4±0.856.1±0.9-64.5±0.287.5±0.5
ResNet-1861.5±1.350.4±0.459.3±0.758.9±1.360.2±1.058.3±1.2-65.7±0.489.3±1.2
", + "bbox": [ + 80, + 128, + 890, + 383 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In the supplementary material we provide a rationale for the meaningfulness of the problem in Eq. (9) based on the model of generative diffusion [43]. Specifically, we argue that the set of controls that leads to the desired final distribution and the set of minimizers, is sufficiently large for a low value of the objective at the top layer.", + "bbox": [ + 75, + 407, + 470, + 500 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 76, + 513, + 209, + 531 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Implementation Details", + "text_level": 1, + "bbox": [ + 76, + 539, + 294, + 556 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For the diffusion model, we adopt pre-trained DiT [33] as the baseline and conduct PEFT with Difffit [50]. $\\lambda_r$ and $\\lambda_d$ are set as 0.002 and 0.008 for Eq. (5), respectively. The image size for the diffusion fine-tuning and sample generation is set as $256\\times 256$ . The fine-tuning mini-batch size is set as 8, and the fine-tuning lasts 8 epochs. The learning rate is set as 1e-3 for an AdamW optimizer. After fine-tuning, the images are generated by 50 denoising steps on a pre-defined number of random noise, according to the IPC setting. All the experiments are conducted on a single RTX 4090 GPU.", + "bbox": [ + 75, + 561, + 468, + 715 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Datasets and Evaluation Metric", + "text_level": 1, + "bbox": [ + 76, + 724, + 357, + 739 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For practical applicability, the experiments are exclusively conducted on full-sized ImageNet [6] subsets in this work. The selected subsets include ImageWoof, ImageNette [15] and the 10-class split adopted in [21, 42], denoted as ImageIDC afterward. ImageWoof is a challenging subset, containing only classes of dog breeds, while ImageNette and ImageIDC contain classes with less similarity, and hence are easier to discriminate. For evaluation, we adopt the same setting as in [21]. The surrogate dataset is trained on different model architectures, with a learning rate of 0.01,", + "bbox": [ + 75, + 750, + 470, + 902 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/9d6fe3cc879402975ab700526d893bfc900a554e120e884afd059880c972168c.jpg", + "table_caption": [ + "Table 2. The Maximum Mean Discrepancy (MMD) between the extracted features of surrogate dataset and the original one." + ], + "table_footnote": [], + "table_body": "
IPCDiT [33]Diffit [50]DM [56]IDC-1 [21]Ours
505.45.44.86.74.0
1005.55.34.06.44.3
", + "bbox": [ + 504, + 445, + 888, + 503 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "and a scheduler decaying the learning rate at 2/3 and 5/6 of the whole training iterations. The top-1 accuracy on the original testing set is reported to illustrate the performance.", + "bbox": [ + 498, + 530, + 890, + 578 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.3. Comparison with State-of-the-art Methods", + "text_level": 1, + "bbox": [ + 498, + 588, + 864, + 604 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We compare our method with other state-of-the-art methods across different IPC settings and model architectures. For a fair comparison, the results are all reproduced by us under the same evaluation protocol. ResNet-10 [18] with average pooling is adopted for matching the feature distribution (DM [56], GLaD [4]) and training gradients (IDC-1 [21]). DM is implemented on IDC-1 by only modifying the matching objective from training gradients to feature distribution, such that better performance is achieved. Each experiment is conducted 3 times, with the mean value and standard variance reported. Firstly, we present the validation results on the challenging ImageWoof subset [15] in Tab. 1.", + "bbox": [ + 496, + 612, + 890, + 792 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "With the target of distilling surrogate datasets of small IPCs (e.g., 10 and 20), the pixel-level optimization method IDC-1 [21] demonstrates outstanding performance gain over random original images. However, as the IPC increases, the performance gain drastically drops. Especially under the 100-IPC setting, the distilled dataset even performs worse than random original images. This observation", + "bbox": [ + 496, + 795, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "15797", + "bbox": [ + 478, + 925, + 517, + 936 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/ddce42543c1a7adc18c94afcf8e9c7eb4d4b54fbd55e499143cd635bbcab4de0.jpg", + "table_caption": [ + "Table 3. Performance comparison with pre-trained diffusion models and state-of-the-art methods on more ImageNet subsets. The results are obtained on ResNet-10 with average pooling. The best results are marked as **bold**." + ], + "table_footnote": [], + "table_body": "
IPCRandomDiT [33]DM [56]Ours
Nette1054.2±1.659.1±0.760.8±0.662.0±0.2
2063.5±0.564.8±1.266.5±1.166.8±0.4
5076.1±1.173.3±0.976.2±0.476.6±0.2
IDC1048.1±0.854.1±0.452.8±0.553.1±0.2
2052.5±0.958.9±0.258.5±0.459.0±0.4
5068.1±0.764.3±0.669.1±0.869.6±0.2
", + "bbox": [ + 81, + 156, + 467, + 276 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "aligns with the empirical findings in Fig. 2, where pixel-level methods struggle to optimize the expanded parameter space of large IPCs. The embedding-level optimization method GLaD [4] yields good performance under the 10-IPC setting. However, it requires overwhelming GPU resources for larger IPC settings, which is inapplicable for resource-restricted scenarios. It is also notable that on large IPCs, the coreset method Herding [48] surpasses previous DD methods with far less computational cost.", + "bbox": [ + 75, + 305, + 468, + 441 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The pre-trained DiT [33] here serves as the baseline for generative diffusion techniques. Under the 50-IPC setting, DiT outperforms both random original images and IDC-1. However, the insufficiency of representativeness and diversity restricts its performance on smaller and larger IPCs, respectively. In contrast, our proposed minimax diffusion consistently provides superior performance across all the IPCs over both original images and Herding. Besides, the proposed method eliminates the need of specific network architectures for matching training metrics. Consequently, the cross-architecture generalization is significantly improved. Under most IPC settings, the performance gap between ConvNet-6 and ResNetAP-10 is even smaller than that of the original images. It validates the universality of the rich information learned by the minimax fine-tuning process.", + "bbox": [ + 75, + 443, + 468, + 671 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Furthermore, we extensively assess the Maximum Mean Discrepancy (MMD) between the embedded features of the selected/generated surrogate dataset and the original one in Tab. 2. The features are extracted by a ResNet-10 network pre-trained on the full original dataset. Our method achieves the lowest discrepancy by average, where DM [56] directly sets MMD as the optimization target, proving the validity of extra minimax criteria in fitting distributions.", + "bbox": [ + 75, + 671, + 468, + 792 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Moreover, we show the performance comparison on ImageNette [15] and ImageIDC [21] in Tab. 3. The performance trend generally aligns with that on ImageWoof. More specifically, on these two easier subsets, DiT quickly loses the advantage over original images as IPC increases. Conversely, our proposed minimax diffusion method consistently demonstrates state-of-the-art performance.", + "bbox": [ + 75, + 795, + 467, + 900 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/e4d7a2c3908f462b109b909137737ebc5af7c07e23ad70082a217fda477dc655.jpg", + "table_caption": [ + "Table 4. Performance comparison on ImageNet-1K." + ], + "table_footnote": [], + "table_body": "
IPCSRe2L [53]RDED [41]DiTOurs
1021.3±0.642.0±0.139.6±0.444.3±0.5
5046.8±0.256.5±0.152.9±0.658.6±0.3
", + "bbox": [ + 514, + 114, + 879, + 172 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/eb4fcacef452f9e119c89cf526ff2f9fe92bfa55af573b4c3d94b37996e7e559.jpg", + "image_caption": [ + "Figure 4. With the help of the minimax diffusion, the proposed method significantly enhances the representativeness and diversity of the generated images. Thereby it consistently provides superior performance compared with random selection and baseline diffusion models by a large margin across different IPC settings." + ], + "image_footnote": [], + "bbox": [ + 506, + 189, + 704, + 406 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/a12b7a4869b423baeb051db78d08ae72416aface9a0e686a33358074d414aa48.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 705, + 189, + 883, + 406 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Experiments on ImageNet-1K. We further conduct experiments on the full ImageNet-1K with the validation protocol of RDED [41] and present the results in Tab. 4. The synthetic images are resized to $224 \\times 224$ for evaluation. The significant performance advantage over the compared works validates the scalability of the proposed method.", + "bbox": [ + 498, + 520, + 890, + 611 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.4. Ablation Study", + "text_level": 1, + "bbox": [ + 500, + 619, + 653, + 636 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Component Analysis. We compare the performance with baseline diffusion models to validate the effectiveness of proposed minimax criteria in Fig. 4. The experiments are conducted on ImageWoof and ImageIDC to evaluate the effect on challenging and easy tasks, respectively. Under the IPC of 10 and 20, the raw diffusion models (DiT) generate informative images, with validation performance much higher than randomly selected original samples. However, as the IPC is continuously increased, the performance gap diminishes for ImageWoof, and random original images surpass the DiT-generated ones at the IPC of 100. On ImageIDC the intersection occurs even earlier at the IPC of 50. The main reason is reflected in Fig. 3, where the sample diversity remains limited without external guidance. The naive Difffit fine-tuning adapts the model to specific domains, yet on large IPCs, the over-fitted generative model still yields inferior performance than the original images.", + "bbox": [ + 496, + 643, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "15798", + "bbox": [ + 480, + 925, + 517, + 936 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/4333973511500f8d9b9119d19e8564fce33e79c0b54d97b7940598d07f3635a1.jpg", + "image_caption": [ + "Figure 5. Visualization of random original images, images generated by baseline diffusion models (DiT [33] and Difffit [50]) and our proposed method. For each column, the generated images are based on the same random seed. Comparatively, our method significantly enhances the coverage of original data distribution and the diversity of the surrogate dataset." + ], + "image_footnote": [], + "bbox": [ + 102, + 89, + 869, + 330 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/f4ff12dcfd24b2eea5fbc6c410c218b769b893939bb578d4f299ad79154c9144.jpg", + "table_caption": [ + "Table 5. The ablation study of the proposed minimax scheme. The result are obtained with ResNet-10 on ImageWoof and ImageIDC. m denotes the proposed minimax optimization scheme." + ], + "table_footnote": [], + "table_body": "
\\( {\\mathcal{L}}_{r} \\)\\( {\\mathcal{L}}_{r} \\) w\\m\\( {\\mathcal{L}}_{d} \\)\\( {\\mathcal{L}}_{d} \\) w\\mImageWoofImageIDC
10-IPC50-IPC10-IPC50-IPC
----35.6±0.951.0±0.953.5±0.266.3±0.2
✓---34.4±1.147.1±0.549.6±0.760.2±1.2
-✓--37.4±0.449.5±1.054.5±1.265.0±0.8
--✓-35.7±0.848.3±0.651.5±0.664.8±0.8
---✓38.7±0.954.9±0.752.2±0.668.4±0.7
✓-✓-38.3±0.554.9±0.453.3±0.566.8±0.5
-✓-✓39.2±1.356.3±1.053.1±0.269.6±0.2
", + "bbox": [ + 78, + 462, + 468, + 604 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The addition of representativeness constraint to the training process further enhances the effect of distribution fitting. At small IPCs, the generated images contain richer information, yet for larger IPCs, the lack of diversity brings a negative influence. The diversity constraint, in contrast, significantly boosts the information contained in the generated surrogate dataset. Despite the performance advantage of $\\mathcal{L}_d$ over $\\mathcal{L}_r$ , combining them still brings stable improvement as our full method. Especially on the easier ImageIDC task, grouping these two constraints together contributes to a consistent performance margin over random original images. The experimental results validate that both representativeness and diversity play essential parts in constructing effective surrogate datasets.", + "bbox": [ + 75, + 638, + 468, + 851 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Minimax Scheme. In this work, we propose to enhance the representativeness and diversity each with a minimax objective. We compare the distillation result with or with-", + "bbox": [ + 75, + 854, + 470, + 902 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "out the minimax operation in Tab. 5. The first row presents the performance of naive Difffit fine-tuning. Matching the embeddings to the distribution center as in Eq. (2) severely degrades the validation performance across all IPCs. In contrast, the minimax version constraint as in Eq. (3) encourages better coverage, where the performance on small IPCs is improved. The effects of diversity constraint and the full method show similar trends. The superior performance suggests the effectiveness in enhancing the essential properties of the generative diffusion techniques.", + "bbox": [ + 496, + 411, + 893, + 561 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.5. Visualization", + "text_level": 1, + "bbox": [ + 500, + 573, + 635, + 588 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Sample Distribution Visualization. The target of our proposed method is to construct a surrogate dataset with both representativeness and diversity. We visualize the t-SNE distribution of samples generated by our proposed method in Fig. 3. In comparison with random original images and baseline diffusion models, our method demonstrates a more thorough coverage over the entire data distribution while maintaining consistency in sample density. At the original high-density region, the generated images also form a dense sub-cluster, which is not reflected by random sampling. On the other hand, at the original sparse regions, our method exhibits better coverage than baseline diffusion models. By simultaneously enhancing the representativeness and diversity in the generative model, the proposed method manages to significantly improve the validation performance of the generated surrogate dataset.", + "bbox": [ + 496, + 597, + 890, + 839 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Generated Sample Comparison. The proposed method notably enhances the representativeness and diversity of the generated surrogate dataset. We compare the samples generated with the same random noise (for each column) of", + "bbox": [ + 496, + 839, + 893, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "15799", + "bbox": [ + 480, + 925, + 517, + 936 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e9068d1bbb0364dd4adf568a30d710136b29428e2397cf17d941585ee0bb2151.jpg", + "image_caption": [ + "Figure 6. Hyper-parameter analysis on (a) the training epochs; (b) the representativeness weight $\\lambda_r$ ; (c) the diversity weight $\\lambda_d$ ; (d) the memory size $N_M$ . The results are obtained with ResNetAP-10 on ImageWoof. The dashed line indicates the value adopted in this work." + ], + "image_footnote": [], + "bbox": [ + 81, + 90, + 272, + 255 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/7dc98ca70bc837ad08622068ccb37cb43da590746fc6ae5e1decc544f50697b2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 279, + 90, + 468, + 255 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/14db85d19c673fcc64aa8565f40e91bd574b85beb356d2d625ce97bc30380486.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 477, + 90, + 666, + 255 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/4c03a4a19a2973f631050f9ce5731e8da4d0ee1d105d7c548928f6b181327eab.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 676, + 90, + 867, + 255 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "different generative methods in Fig. 5 to explicitly demonstrate the improved properties.", + "bbox": [ + 75, + 321, + 467, + 351 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The images generated by baseline DiT exhibit a realistic high-quality appearance. However, the images tend to share similar poses and only present the most prominent features of the objects. In the golden retriever case, the generated images mostly present the head part, while for the churches the exterior appearance. Difffit fine-tuning further fits the model to the distribution, but in most cases, the differences only lie in small details. Comparatively, the proposed minimax criteria significantly enhance both the representativeness and diversity of the generated images. On the one hand, there occurs more class-related content in the generated images. The golden retriever images include more body parts and the church images encompass the interior layout. The minimax optimization leads to better coverage over the entire original distribution, with more related features encapsulated. On the other hand, the diversity is significantly enhanced, including variations in pose, background, and appearance styles. In such a way the surrogate dataset better represents the original large-scale one, leading to superior validation performance.", + "bbox": [ + 75, + 352, + 467, + 654 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Training Curve Visualization. We visualize the accuracy curve during the training process in Fig. 6a. The validation performance is rapidly improved as the fine-tuning process starts. After four epochs, the model tends to converge and reaches the highest performance at the 8th epoch. Further extending the training epochs injects excessive diversity into the model, leading to performance degradation. We demonstrate the influence of training epochs on the generated images in supplementary material.", + "bbox": [ + 75, + 655, + 467, + 790 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.6. Parameter Analysis", + "text_level": 1, + "bbox": [ + 76, + 801, + 264, + 816 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Objective Weight $\\lambda_r\\lambda_d$ .We show the influence of representativeness weight $\\lambda_r$ and diversity weight $\\lambda_{d}$ in Fig. 6b and Fig. 6c, respectively. The $\\lambda_r$ variation only produces negligible performance fluctuation on small IPCs, while on large IPCs the performance is also relatively stable. For", + "bbox": [ + 75, + 824, + 467, + 898 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "$\\lambda_{d}$ , at a proper variation range, the performance is stable. However, continuously increasing the diversity of the generated dataset leads to a lack of representativeness, which results in a negative impact. The negative impact of overdiversity can also validated by the poor performance of K-Center in Tab. 1. A uniform performance decrease is observed as $\\lambda_{d}$ reaches 0.03. Based on the performance of 100 IPC, we set $\\lambda_{r}$ as 0.002 and $\\lambda_{d}$ as 0.008.", + "bbox": [ + 496, + 321, + 890, + 441 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Memory Size $N_{M}$ . The memory size $N_{M}$ influences the number of samples involved in the objective calculation. We investigate its influence in Fig. 6d. When the memory is extremely small ( $N_{M} = 16$ ), the provided supervision is also limited, yet the performance is already higher than naive fine-tuning. As the memory size is increased in a proper range, the model yields stable performance improvement. It is notable that with a larger memory, the performance under the IPC of 10 is better. It can be explained by that a larger memory contains more representative information. Out of the consideration of performance as well as storage burden, we select the memory of 64 in the other experiments.", + "bbox": [ + 496, + 444, + 890, + 625 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 643, + 617, + 659 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this work, we propose a novel dataset distillation method based on generative diffusion techniques. Through extra minimax criteria, the proposed method significantly enhances the representativeness and diversity of the generated surrogate dataset. With much less computational time consumption, the proposed method achieves state-of-the-art validation performance on challenging ImageNet subsets. It reduces the resource dependency of previous dataset distillation methods and opens up new possibilities for more practical applications for distilling personalized data.", + "bbox": [ + 496, + 670, + 890, + 821 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Limitations and Future Works. In addition to object classification, we will explore the possibility of incorporating generative techniques for more specific data domains.", + "bbox": [ + 496, + 823, + 890, + 868 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgement. We would like to specially thank Olga Russakovsky for invaluable advice on the manuscript.", + "bbox": [ + 500, + 869, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "15800", + "bbox": [ + 480, + 925, + 517, + 936 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 174, + 106 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Shekoofeh Azizi, Simon Kornblith, Chitwan Sahara, Mohammad Norouzi, and David J Fleet. Synthetic data from diffusion models improves imagenet classification. arXiv preprint arXiv:2304.08466, 2023. 2", + "[2] Victor Besnier, Himalaya Jain, Andrei Bursuc, Matthieu Cord, and Patrick Pérez. This Dataset Does Not Exist: Training Models from Generated Images. In ICASSP, pages 1-5, 2020. 2", + "[3] George Cazenavette, Tongzhou Wang, Antonio Torralba, Alexei A Efros, and Jun-Yan Zhu. Dataset distillation by matching training trajectories. In CVPR, pages 4750-4759, 2022. 1", + "[4] George Cazenavette, Tongzhou Wang, Antonio Torralba, Alexei A Efros, and Jun-Yan Zhu. Generalizing dataset distillation via deep generative prior. In CVPR, pages 3739-3748, 2023. 1, 5, 6", + "[5] Justin Cui, Ruochen Wang, Si Si, and Cho-Jui Hsieh. Dc-bench: Dataset condensation benchmark. NeurIPS, 35:810-822, 2022. 1", + "[6] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. ImageNet: A large-scale hierarchical image database. In CVPR, pages 248-255, 2009. 1, 5", + "[7] Zhiwei Deng and Olga Russakovsky. Remember the Past: Distilling Datasets into Addressable Memories for Neural Networks. In NeurIPS, 2022. 1", + "[8] Prafulla Dhariwal and Alexander Nichol. Diffusion Models Beat GANs on Image Synthesis. In NeurIPS, pages 8780-8794, 2021. 3, 2", + "[9] Thinh T Doan. Nonlinear two-time-scale stochastic approximation convergence and finite-time performance. IEEE Transactions on Automatic Control, 2022. 1", + "[10] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale. In ICLR, 2022. 1", + "[11] Jiawei Du, Yidi Jiang, Vincent Y.F. Tan, Joey Tianyi Zhou, and Haizhou Li. Minimizing the Accumulated Trajectory Error to Improve Dataset Distillation. In CVPR, pages 3749-3758, 2023. 1", + "[12] Lisa Dunlap, Alyssa Umino, Han Zhang, Jiezhi Yang, Joseph E Gonzalez, and Trevor Darrell. Diversify your vision datasets with automatic diffusion-based augmentation. arXiv preprint arXiv:2305.16289, 2023. 2", + "[13] Gabriele Eichfelder. Scalarizations for adaptively solving multi-objective optimization problems. Com-" + ], + "bbox": [ + 78, + 114, + 470, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "putational Optimization and Applications, 44:249-273, 2009. 4", + "[14] Ronen Eldan and James R Lee. Regularization under diffusion and anticoncentration of the information content. Duke Mathematical Journal, 167(5):969-993, 2018. 1", + "[15] Fastai. Fastai/imagenette: A smaller subset of 10 easily classified classes from imagenet, and a little more french. 1, 5, 6", + "[16] Jianyang Gu, Kai Wang, Wei Jiang, and Yang You. Summarizing stream data for memory-restricted online continual learning. arXiv preprint arXiv:2305.16645, 2023. 1", + "[17] Swaminathan Gurumurthy, Ravi Kiran Sarvadevabhatla, and R. Venkatesh Babu. DeLiGAN: Generative Adversarial Networks for Diverse and Limited Data. In CVPR, pages 4941-4949, 2017. 2", + "[18] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, pages 770-778, 2016. 1, 5, 2", + "[19] Ruifei He, Shuyang Sun, Xin Yu, Chuhui Xue, Wenqing Zhang, Philip Torr, Song Bai, and Xiaojuan Qi. Is Synthetic Data from Generative Models Ready for Image Recognition? In ICLR, 2022. 2", + "[20] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising Diffusion Probabilistic Models. In NeurIPS, pages 6840-6851, 2020. 2", + "[21] Jang-Hyun Kim, Jinuk Kim, Seong Joon Oh, Sang-doo Yun, Hwanjun Song, Joonhyun Jeong, Jung-Woo Ha, and Hyun Oh Song. Dataset condensation via efficient synthetic-data parameterization. In ICML, pages 11102-11118, 2022. 1, 5, 6, 2, 3", + "[22] Diederik Kingma, Tim Salimans, Ben Poole, and Jonathan Ho. Variational Diffusion Models. In NeurIPS, pages 21696-21707, 2021. 2", + "[23] Saehyung Lee, Sanghyuk Chun, Sangwon Jung, Sang-doo Yun, and Sungroh Yoon. Dataset condensation with contrastive signals. In ICML, pages 12352-12364, 2022. 1", + "[24] Joseph Lehec. Representation formula for the entropy and functional inequalities. In Annales de l'IHP Probabilités et statistiques, pages 885-899, 2013. 1", + "[25] Ping Liu, Xin Yu, and Joey Tianyi Zhou. Meta Knowledge Condensation for Federated Learning. In ICLR, 2022. 1", + "[26] Yanqing Liu, Jianyang Gu, Kai Wang, Zheng Zhu, Wei Jiang, and Yang You. Dream: Efficient dataset distillation by representative matching. In ICCV, 2023. 1, 2", + "[27] Noel Loo, Ramin Hasani, Alexander Amini, and Daniela Rus. Efficient dataset distillation using random feature approximation. NeurIPS, 35:13877-13891, 2022. 1" + ], + "bbox": [ + 503, + 92, + 893, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "15801", + "bbox": [ + 480, + 926, + 517, + 936 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[28] Noel Loo, Ramin Hasani, Mathias Lechner, and Daniela Rus. Dataset distillation with convexified implicit gradients. arXiv preprint arXiv:2302.06755, 2023.1", + "[29] Qi Mao, Hsin-Ying Lee, Hung-Yu Tseng, Siwei Ma, and Ming-Hsuan Yang. Mode Seeking Generative Adversarial Networks for Diverse Image Synthesis. In CVPR, pages 1429-1437, 2019. 2", + "[30] Timothy Nguyen, Zhourong Chen, and Jaehoon Lee. Dataset meta-learning from kernel ridge-regression. In ICLR, 2021. 1", + "[31] Timothy Nguyen, Roman Novak, Lechao Xiao, and Jaehoon Lee. Dataset distillation with infinitely wide convolutional networks. NeurIPS, 34:5186-5198, 2021. 1", + "[32] Alexander Quinn Nichol and Prafulla Dhariwal. Improved Denoising Diffusion Probabilistic Models. In ICML, pages 8162-8171, 2021. 2", + "[33] William Peebles and Saining Xie. Scalable diffusion models with transformers. In ICCV, pages 4195-4205, 2023. 3, 5, 6, 7", + "[34] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In CVPR, pages 22500-22510, 2023. 3", + "[35] Ahmad Sajedi, Samir Khaki, Ehsan Amjadian, Lucy Z Liu, Yuri A Lawryshyn, and Konstantinos N Plataniotis. DataDAM: Efficient Dataset Distillation with Attention Matching. In ICCV, 2023. 1", + "[36] Mert Bulent Sariyildiz, Karteek Alahari, Diane Larlus, and Yannis Kalantidis. Fake it Till You Make it: Learning Transferable Representations from Synthetic ImageNet Clones. In CVPR, pages 8011-8021, 2023. 2", + "[37] Ozan Sener and Silvio Savarese. Active learning for convolutional neural networks: A core-set approach. In International Conference on Learning Representations, 2018. 5", + "[38] Allahkaram Shafiei, Vyacheslav Kungurtsev, and Jakub Marecek. Trilevel and multilevel optimization using monotone operator theory. arXiv preprint arXiv:2105.09407, 2021. 1", + "[39] Jordan Shipard, Arnold Wiliem, Kien Nguyen Thanh, Wei Xiang, and Clinton Fookes. Diversity Is Definitely Needed: Improving Model-Agnostic Zero-Shot Classification via Stable Diffusion. In CVPR, pages 769-778, 2023. 2", + "[40] Felipe Petroski Such, Aditya Rawal, Joel Lehman, Kenneth Stanley, and Jeffrey Clune. Generative Teaching Networks: Accelerating Neural Architecture Search by Learning to Generate Synthetic Training Data. In ICML, pages 9206-9216, 2020. 1" + ], + "bbox": [ + 78, + 90, + 470, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[41] Peng Sun, Bei Shi, Daiwei Yu, and Tao Lin. On the diversity and realism of distilled dataset: An efficient dataset distillation paradigm. In CVPR, 2024. 6", + "[42] Yonglong Tian, Dilip Krishnan, and Phillip Isola. Contrastive multiview coding. In ECCV, pages 776-794, 2020. 5", + "[43] Belinda Tzen and Maxim Raginsky. Theoretical guarantees for sampling and inference in generative models with latent diffusions. In $COLT$ , pages 3084-3114, 2019. 4, 5, 1", + "[44] Saeed Vahidian, Mingyu Wang, Jianyang Gu, Vyacheslav Kungurtsev, Wei Jiang, and Yiran Chen. Group distributionally robust dataset distillation with risk minimization. arXiv preprint arXiv:2402.04676, 2024.1", + "[45] Kai Wang, Bo Zhao, Xiangyu Peng, Zheng Zhu, Shuo Yang, Shuo Wang, Guan Huang, Hakan Bilen, Xin chao Wang, and Yang You. Cafe: Learning to condense dataset by aligning features. In CVPR, pages 12196-12205, 2022. 1", + "[46] Kai Wang, Jianyang Gu, Daquan Zhou, Zheng Zhu, Wei Jiang, and Yang You. Dim: Distilling dataset into generative model. arXiv preprint arXiv:2303.04707, 2023.1", + "[47] Tongzhou Wang, Jun-Yan Zhu, Antonio Torralba, and Alexei A Efros. Dataset distillation. arXiv preprint arXiv:1811.10959, 2018. 1, 2", + "[48] Max Welling. Herding dynamical weights to learn. In ICML, pages 1121-1128, 2009. 5, 6, 3", + "[49] Xindi Wu, Zhiwei Deng, and Olga Russakovsky. Multimodal dataset distillation for image-text retrieval. arXiv preprint arXiv:2308.07545, 2023. 1", + "[50] Enze Xie, Lewei Yao, Han Shi, Zhili Liu, Daquan Zhou, Zhaoqiang Liu, Jiawei Li, and Zhenguo Li. Difffit: Unlocking transferability of large diffusion models via simple parameter-efficient fine-tuning. arXiv preprint arXiv:2304.06648, 2023. 3, 5, 7, 2", + "[51] Yuanhao Xiong, Ruochen Wang, Minhao Cheng, Felix Yu, and Cho-Jui Hsieh. FedDM: Iterative Distribution Matching for Communication-Efficient Federated Learning. In CVPR, pages 16323-16332, 2023. 1", + "[52] Dingdong Yang, Seunghoon Hong, Yunseok Jang, Tianchen Zhao, and Honglak Lee. Diversity-Sensitive Conditional Generative Adversarial Networks. In ICLR, 2018. 2", + "[53] Zeyuan Yin, Eric Xing, and Zhiqiang Shen. Squeeze, recover and relabel: Dataset condensation atImagenet scale from a new perspective. In NeurIPS, 2023. 6", + "[54] Bo Zhao and Hakan Bilen. Dataset condensation with differentiable siamese augmentation. In ICML, pages 12674-12685, 2021. 1, 2" + ], + "bbox": [ + 501, + 90, + 890, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "15802", + "bbox": [ + 480, + 926, + 519, + 936 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[55] Bo Zhao and Hakan Bilen. Synthesizing informative training samples with gan. arXiv preprint arXiv:2204.07513, 2022. 1", + "[56] Bo Zhao and Hakan Bilen. Dataset condensation with distribution matching. In WACV, pages 6514-6523, 2023. 2, 5, 6, 1", + "[57] Bo Zhao, Konda Reddy Mopuri, and Hakan Bilen. Dataset condensation with gradient matching. In ICLR, 2021. 1, 2", + "[58] Ganlong Zhao, Guanbin Li, Yipeng Qin, and Yizhou Yu. Improved Distribution Matching for Dataset Condensation. In CVPR, pages 7856-7865, 2023. 1", + "[59] Yongchao Zhou, Ehsan Nezhadarya, and Jimmy Ba. Dataset distillation using neural feature regression. NeurIPS, 35:9813-9827, 2022. 1", + "[60] Yongchao Zhou, Hshmat Sahak, and Jimmy Ba. Training on thin air: Improve image classification with generated data. arXiv preprint arXiv:2305.15316, 2023.2" + ], + "bbox": [ + 78, + 90, + 468, + 382 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "15803", + "bbox": [ + 480, + 926, + 517, + 936 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2024/Efficient Dataset Distillation via Minimax Diffusion/75146b44-b76a-42b0-8456-4c37446317b7_model.json b/2024/Efficient Dataset Distillation via Minimax Diffusion/75146b44-b76a-42b0-8456-4c37446317b7_model.json new file mode 100644 index 0000000000000000000000000000000000000000..be81ad316dfbc3a53b95587544cdaadb7ee73a4f --- /dev/null +++ b/2024/Efficient Dataset Distillation via Minimax Diffusion/75146b44-b76a-42b0-8456-4c37446317b7_model.json @@ -0,0 +1,2334 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.0, + 0.812, + 0.045 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.226, + 0.131, + 0.745, + 0.151 + ], + "angle": 0, + "content": "Efficient Dataset Distillation via Minimax Diffusion" + }, + { + "type": "text", + "bbox": [ + 0.248, + 0.18, + 0.721, + 0.199 + ], + "angle": 0, + "content": "Jianyang Gu1 Saeed Vahidian2 Vyacheslav Kungurtsev3" + }, + { + "type": "text", + "bbox": [ + 0.252, + 0.199, + 0.699, + 0.216 + ], + "angle": 0, + "content": "Haonan Wang\\(^{4}\\) Wei Jiang\\(^{1*}\\) Yang You\\(^{4}\\) Yiran Chen\\(^{2}\\)" + }, + { + "type": "text", + "bbox": [ + 0.202, + 0.216, + 0.778, + 0.235 + ], + "angle": 0, + "content": "\\(^{1}\\)Zhejiang University \\(^{2}\\)Duke University \\(^{3}\\)Czech Technical University" + }, + { + "type": "text", + "bbox": [ + 0.351, + 0.235, + 0.621, + 0.253 + ], + "angle": 0, + "content": "\\(^{4}\\)National University of Singapore" + }, + { + "type": "text", + "bbox": [ + 0.32, + 0.255, + 0.647, + 0.27 + ], + "angle": 0, + "content": "{gu_jianyang, jiangwei_zju}@zju.edu.cn" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.305, + 0.313, + 0.32 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.336, + 0.473, + 0.714 + ], + "angle": 0, + "content": "Dataset distillation reduces the storage and computational consumption of training a network by generating a small surrogate dataset that encapsulates rich information of the original large-scale one. However, previous distillation methods heavily rely on the sample-wise iterative optimization scheme. As the images-per-class (IPC) setting or image resolution grows larger, the necessary computation will demand overwhelming time and resources. In this work, we intend to incorporate generative diffusion techniques for computing the surrogate dataset. Observing that key factors for constructing an effective surrogate dataset are representativeness and diversity, we design additional minimax criteria in the generative training to enhance these facets for the generated images of diffusion models. We present a theoretical model of the process as hierarchical diffusion control demonstrating the flexibility of the diffusion process to target these criteria without jeopardizing the faithfulness of the sample to the desired distribution. The proposed method achieves state-of-the-art validation performance while demanding much less computational resources. Under the 100-IPC setting on ImageWoof, our method requires less than one-twentieth the distillation time of previous methods, yet yields even better performance. Source code and generated data are available in https://github.com/vimar-gu/MinimaxDiffusion." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.742, + 0.21, + 0.757 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.767, + 0.47, + 0.844 + ], + "angle": 0, + "content": "Data, as a necessary resource for deep learning, has concurrently promoted algorithmic advancements while imposing challenges on researchers due to heavy demands on storage and computational resources [6, 10, 18, 47]. Confronted with the conflict between the requirement for high-precision" + }, + { + "type": "image", + "bbox": [ + 0.516, + 0.306, + 0.877, + 0.504 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.517, + 0.895, + 0.616 + ], + "angle": 0, + "content": "Figure 1. The validation accuracy and distillation time of different methods on ImageWoof [15], with a number following each method denoting the Image-Per-Class (IPC) setting. Previous methods are restricted by the heavier running time and memory consumption as IPC grows larger. In comparison, our proposed method notably reduces the demanding computational resources and also achieves state-of-the-art validation performance." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.643, + 0.894, + 0.719 + ], + "angle": 0, + "content": "models and overwhelming resource demands, dataset distillation is proposed to condense the rich information of a large-scale dataset into a small surrogate one [5, 21, 47, 57]. Such a surrogate dataset is expected to achieve training performance comparable to that attained with the original one." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.895, + 0.901 + ], + "angle": 0, + "content": "Previous dataset distillation methods mostly engage in iterative optimization on fixed-number samples at the pixel level [21, 26, 30, 31, 44, 54, 57] or embedding level [4, 55]. However, the sample-wise iterative optimization scheme suffers from problems of two perspectives. (1) The parameter space of optimization is positively correlated with the size of the target surrogate dataset and the image resolution [3, 57]. Consequently, substantial time and computational resources are required for distilling larger datasets. As shown in Fig. 1, IDC-1 [21] takes over 90 hours to distill a 100-image-per-class (IPC) set from ImageWoof [15], while training on ImageWoof itself only requires a matter of" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.852, + 0.222, + 0.864 + ], + "angle": 0, + "content": "*Corresponding author" + }, + { + "type": "page_footnote", + "bbox": [ + 0.078, + 0.864, + 0.469, + 0.899 + ], + "angle": 0, + "content": "This work has received funding from the European Union's Horizon Europe research and innovation program under grant agreement No. 101084642." + }, + { + "type": "list", + "bbox": [ + 0.078, + 0.852, + 0.469, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.925, + 0.519, + 0.937 + ], + "angle": 0, + "content": "15793" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.084, + 0.089, + 0.462, + 0.261 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.273, + 0.47, + 0.343 + ], + "angle": 0, + "content": "Figure 2. Sample images distilled by the pixel-level sample-wise optimization method DM [56] on ImageWoof. As the parameter space increases along with the Image-Per-Class (IPC) setting, with the same initialization, the appearance disparity between original and distilled images is smaller." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.371, + 0.468, + 0.567 + ], + "angle": 0, + "content": "hours. (2) The expanded parameter space also increases the optimization complexity. As shown in Fig. 2, while distillation yields significant information condensation under small IPC settings, the pixel modification diminishes when distilling larger-IPC datasets. The reduced disparity also leads to smaller performance gain compared with original images, with instances where the distilled set even performs worse. Especially when distilling data of fine-grained classes, the sample-wise optimization scheme fails to provide adequate discriminative information. These constraints severely hinder individual researchers from distilling personalized data. A more practical training scheme is urgently needed to facilitate the broader application of dataset distillation." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.568, + 0.468, + 0.839 + ], + "angle": 0, + "content": "In this work, we explore the possibility of incorporating generative diffusion techniques [20, 22, 32] to efficiently compute effective surrogate datasets. We first conduct empirical analysis on the suitability of data generated by raw diffusion models for training networks. Based on the observations, we conclude that constructing an effective surrogate dataset hinges on two key factors: representativeness and diversity. Accordingly, we design extra minimax criteria for the generative training to enhance the capability of generating more effective surrogate datasets without explicit prompt designs. The minimax criteria involve two aspects: enforcing the generated sample to be close to the farthest real sample, while being far away from the most similar generated one. We provide theoretical analysis to support that the proposed minimax scheme aims to solve a well defined problem with all the criteria, including the generative accuracy and the minimax criteria, can be targeted simultaneously without detriment to the others." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.84, + 0.468, + 0.901 + ], + "angle": 0, + "content": "Compared with the astronomical training time consumption of the sample-wise iterative optimization schemes, the proposed method takes less than 1 hour to distill a 100-IPC surrogate dataset for a 10-class ImageNet subset, including" + }, + { + "type": "image", + "bbox": [ + 0.507, + 0.092, + 0.691, + 0.214 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.506, + 0.215, + 0.689, + 0.337 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.697, + 0.092, + 0.882, + 0.214 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.697, + 0.215, + 0.882, + 0.337 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.351, + 0.892, + 0.393 + ], + "angle": 0, + "content": "Figure 3. The feature distribution comparison of different image generation methods with the original set. The validation performance of each surrogate set is listed in the upper-right corner." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.422, + 0.892, + 0.558 + ], + "angle": 0, + "content": "the fine-tuning and image generation processes. Remarkably, the GPU consumption remains consistent across all IPC settings. Furthermore, the distilled surrogate dataset attains superior validation performance compared with other state-of-the-art methods. Especially on the challenging fine-grained ImageWoof subset, the proposed method outperforms the second-best DD method by \\(5.5\\%\\) and \\(8.1\\%\\) under the IPC settings of 70 and 100, respectively. The source code is provided in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.56, + 0.865, + 0.573 + ], + "angle": 0, + "content": "The contributions of this work are summarized into:" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.575, + 0.891, + 0.619 + ], + "angle": 0, + "content": "- We analyze the data generated by diffusion models, and emphasize the importance of representativeness and diversity for constructing effective surrogate datasets." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.621, + 0.89, + 0.664 + ], + "angle": 0, + "content": "- We propose a novel dataset distillation scheme based on extra minimax criteria for diffusion models targeting the representativeness and diversity of generated data." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.666, + 0.891, + 0.71 + ], + "angle": 0, + "content": "- We theoretically justify the proposed minimax criteria as enforceable without trade-offs in the generation quality of the individual data points." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.712, + 0.89, + 0.77 + ], + "angle": 0, + "content": "- We conduct extensive experiments to validate that our proposed method achieves state-of-the-art performance while demanding significantly reduced training time in comparison to previous dataset distillation methods." + }, + { + "type": "list", + "bbox": [ + 0.5, + 0.575, + 0.891, + 0.77 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.789, + 0.59, + 0.804 + ], + "angle": 0, + "content": "2. Method" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.815, + 0.682, + 0.829 + ], + "angle": 0, + "content": "2.1. Problem Definition" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.84, + 0.892, + 0.901 + ], + "angle": 0, + "content": "The general purpose of dataset distillation is to generate a small surrogate dataset \\(S = \\{(\\mathbf{x}_i,y_i)\\}_{i = 1}^{N_S}\\) from a largescale one \\(\\mathcal{T} = \\{(\\mathbf{x}_i,y_i)\\}_{i = 1}^{N_T}\\) [47, 57]. Here each \\(\\mathbf{x}_i\\) denotes an image with a corresponding class label \\(y_{i}\\), and" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.925, + 0.518, + 0.936 + ], + "angle": 0, + "content": "15794" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.47, + 0.168 + ], + "angle": 0, + "content": "\\(N_{S} \\ll N_{T}\\). The surrogate dataset \\(S\\) is expected to encapsulate substantial information from the original \\(\\mathcal{T}\\), such that training a model on \\(S\\) achieves performance comparable with that on \\(\\mathcal{T}\\). After distilling, we train network models on \\(S\\) and validate the performance on the original test set." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.175, + 0.303, + 0.19 + ], + "angle": 0, + "content": "2.2. Diffusion for Distillation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.199, + 0.471, + 0.395 + ], + "angle": 0, + "content": "Diffusion models learn a dataset distribution by gradually adding Gaussian noise to images and reversing back. Taking the latent diffusion model (LDM) as an example, given a training image \\(\\mathbf{x}\\), the training process is separated into two parts. An encoder \\(E\\) transforms the image into the latent space \\(\\mathbf{z} = E(\\mathbf{x})\\) and a decoder \\(D\\) reconstructs a latent code back to the image space \\(\\hat{\\mathbf{x}} = D(\\mathbf{z})\\). The forward noisng process gradually adds noise \\(\\epsilon \\sim \\mathcal{N}(\\mathbf{0},\\mathbf{I})\\) to the original latent code \\(\\mathbf{z}_0\\): \\(\\mathbf{z}_t = \\sqrt{\\bar{\\alpha}_t}\\mathbf{z}_0 + \\sqrt{1 - \\bar{\\alpha}_t\\epsilon}\\), where \\(\\bar{\\alpha}_t\\) is a hyper-parameter. Provided with a conditioning vector \\(\\mathbf{c}\\) encoded with class labels, the diffusion models are trained by the squared error between the predicted noise \\(\\epsilon_{\\theta}(\\mathbf{z}_t,\\mathbf{c})\\) and the ground truth \\(\\epsilon\\):" + }, + { + "type": "equation", + "bbox": [ + 0.178, + 0.404, + 0.469, + 0.423 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {s i m p l e}} = \\left| \\left| \\epsilon_ {\\theta} \\left(\\mathbf {z} _ {t}, \\mathbf {c}\\right) - \\epsilon \\right| \\right| _ {2} ^ {2}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.431, + 0.47, + 0.582 + ], + "angle": 0, + "content": "where \\(\\epsilon_{\\theta}\\) is a noise prediction network parameterized by \\(\\theta\\). Diffusion models are proven to generate images of higher quality compared with GANs [8]. There are also some Parameter Efficient Fine-Tuning (PEFT) methods updating a small number of model parameters in order for the model to be better adapted to specific data domains [34, 50]. We adopt DiT [33] as the baseline and Difffit [50] as the naive fine-tuning method for image generation. The generated images are compared with the original data from the perspective of embedding distribution in Fig. 3." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.583, + 0.47, + 0.81 + ], + "angle": 0, + "content": "The samples of random selection and pre-trained diffusion models present two extreme ends of the distribution. Random selection faithfully reflects the original distribution, yet fails to emphasize some high-density regions. In contrast, diffusion models are over-fitted to those dense areas, leaving a large part of the original distribution uncovered. We attribute these two distributions to two properties, respectively. The randomly selected data holds extraordinary diversity, and the diffusion-generated data shows representativeness to the original distribution. We claim that both properties are essential for constructing an effective surrogate dataset. By naive fine-tuning, Difffit better captures the representative regions but leaves more regions uncovered. To this end, we propose extra minimax criteria for the diffusion model to enhance both of the properties." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.817, + 0.327, + 0.833 + ], + "angle": 0, + "content": "2.3. Minimax Diffusion Criteria" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Based on the observation that representativeness and diversity are two key factors to construct an effective surrogate dataset, we accordingly design extra minimax criteria to enhance these two essential properties for the diffusion model." + }, + { + "type": "code_caption", + "bbox": [ + 0.51, + 0.095, + 0.816, + 0.11 + ], + "angle": 0, + "content": "Algorithm 1: Minimax Diffusion Fine-tuning" + }, + { + "type": "algorithm", + "bbox": [ + 0.508, + 0.112, + 0.892, + 0.35 + ], + "angle": 0, + "content": "Input: initialized model parameter \\(\\theta\\) original dataset \\(\\mathcal{T} = \\{(\\mathbf{x},y)\\}\\) , encoder \\(E\\) class encoder \\(E_{c}\\) time step \\(t\\) variance schedule \\(\\bar{\\alpha}_t\\) real embedding memory \\(\\mathcal{M}\\) predicted embedding memory \\(\\mathcal{D}\\) \nOutput: optimized model parameter \\(\\theta^{*}\\) \nfor each step do Obtain the original embedding: \\(\\mathbf{z}_0 = E(\\mathbf{x})\\) Obtain the class embedding: \\(\\mathbf{c} = E_c(y)\\) Sample random noise: \\(\\epsilon \\sim \\mathcal{N}(\\mathbf{0},\\mathbf{I})\\) Add noise to the embedding: \\(\\mathbf{z}_t = \\sqrt{\\bar{\\alpha}_t}\\mathbf{z}_0 + \\sqrt{1 - \\bar{\\alpha}_t}\\epsilon\\) Predict the noise \\(\\epsilon_{\\theta}(\\mathbf{z}_{t},\\mathbf{c})\\) and recovered embedding \\(\\hat{\\mathbf{z}}_{\\theta}(\\mathbf{z}_{t},\\mathbf{c}) = \\mathbf{z}_{t} - \\epsilon_{\\theta}(\\mathbf{z}_{t},\\mathbf{c})\\) Update the model parameter with Eq. (5) Enqueue the real embedding: \\(\\mathcal{M}_r\\gets \\mathbf{z}_0\\) Enqueue the predicted embedding: \\(\\mathcal{M}_d\\gets \\hat{\\mathbf{z}}_\\theta (\\mathbf{z}_t,\\mathbf{c})\\) \nend" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.384, + 0.892, + 0.445 + ], + "angle": 0, + "content": "Representativeness It is essential for the small surrogate dataset to sufficiently represent the original data. A naive approach to improve the representativeness is aligning the embedding distribution between synthetic and real samples:" + }, + { + "type": "equation", + "bbox": [ + 0.553, + 0.455, + 0.892, + 0.497 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {r} = \\underset {\\theta} {\\arg \\max } \\sigma \\left(\\hat {\\mathbf {z}} _ {\\theta} \\left(\\mathbf {z} _ {t}, \\mathbf {c}\\right), \\frac {1}{N _ {B}} \\sum_ {i = 0} ^ {N _ {B}} \\mathbf {z} _ {i}\\right), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.503, + 0.892, + 0.655 + ], + "angle": 0, + "content": "where \\(\\sigma (\\cdot ,\\cdot)\\) is the cosine similarity, \\(\\hat{\\mathbf{z}}_{\\theta}(\\mathbf{z}_t,\\mathbf{c})\\) is the predicted original embedding by subtracting the noise from the noisy embedding \\(\\hat{\\mathbf{z}}_{\\theta}(\\mathbf{z}_t,\\mathbf{c}) = \\mathbf{z}_t - \\epsilon_{\\theta}(\\mathbf{z}_t,\\mathbf{c})\\), and \\(N_B\\) is the size of the sampled real sample mini-batch. However, the naive alignment tends to draw the predicted embedding towards the center of the real distribution, which severely limits the diversity. Therefore, we propose to maintain an auxiliary memory \\(\\mathcal{M} = \\{\\mathbf{z}_m\\}_{m = 1}^{N_M}\\) to store the real samples utilized in adjacent iterations, and design a minimax optimization objective as:" + }, + { + "type": "equation", + "bbox": [ + 0.558, + 0.666, + 0.891, + 0.691 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {r} = \\arg \\max _ {\\theta} \\min _ {m \\in \\left[ N _ {M} \\right]} \\sigma \\left(\\hat {\\mathbf {z}} _ {\\theta} \\left(\\mathbf {z} _ {t}, \\mathbf {c}\\right), \\mathbf {z} _ {m}\\right). \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.701, + 0.892, + 0.792 + ], + "angle": 0, + "content": "By pulling close the least similar sample pairs, the diffusion model is encouraged to generate images that better cover the original distribution. It is notable that the diffusion training objective \\(\\mathcal{L}_{simple}\\) itself encourages the generated images to resemble the original ones. Thus, the minimax criterion allows the preservation of diversity to the maximum extent." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.81, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Diversity Although the pre-trained diffusion models already achieve satisfactory generation quality, the remaining defect is limited diversity compared with the original data, as shown in Fig. 3. We expect the data generated by the diffusion model can accurately reflect the original distribution, while simultaneously being different from" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.926, + 0.519, + 0.938 + ], + "angle": 0, + "content": "15795" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.47, + 0.152 + ], + "angle": 0, + "content": "each other. Hence, we maintain another auxiliary memory \\(\\mathcal{D} = \\{\\mathbf{z}_d\\}_{d=1}^{N_D}\\) for the predicted embeddings of adjacent iterations and design another minimax objective to explicitly enhance the sample diversity as:" + }, + { + "type": "equation", + "bbox": [ + 0.143, + 0.161, + 0.469, + 0.186 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {d} = \\arg \\min _ {\\theta} \\max _ {d \\in [ N _ {D} ]} \\sigma (\\hat {\\mathbf {z}} _ {\\theta} (\\mathbf {z} _ {t}, \\mathbf {c}), \\mathbf {z} _ {d}). \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.193, + 0.469, + 0.312 + ], + "angle": 0, + "content": "The diversity term has an opposite optimization target compared with the representativeness term, where the predicted embedding is pushed away from the most similar one stored in the memory bank. Although diversity is essential for an effective surrogate set, too much of it will cause the generated data to lose representativeness. The proposed minimax optimization enhances the diversity in a gentle way, with less influence on the class-related features." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.314, + 0.469, + 0.358 + ], + "angle": 0, + "content": "Combining all the components, we summarize the training process in Algorithm 1. The complete training objective can be formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.17, + 0.369, + 0.469, + 0.385 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\mathcal {L} _ {\\text {s i m p l e}} + \\lambda_ {r} \\mathcal {L} _ {r} + \\lambda_ {d} \\mathcal {L} _ {d}, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.394, + 0.407, + 0.409 + ], + "angle": 0, + "content": "where \\(\\lambda_r\\) and \\(\\lambda_d\\) are weighting hyper-parameters." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.421, + 0.274, + 0.437 + ], + "angle": 0, + "content": "3. Theoretical Analysis" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.446, + 0.469, + 0.492 + ], + "angle": 0, + "content": "Assume that \\(\\mu\\) is the real distribution of the latent variables \\(\\mathbf{z}\\) associated with the target dataset \\(\\mathcal{T}\\). We rewrite the optimization problem presented in Eq. (5) in a modified form:" + }, + { + "type": "equation", + "bbox": [ + 0.086, + 0.511, + 0.472, + 0.58 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\min_{\\{\\theta^{(i)}\\}_{i\\in [N_{D}]}}\\lambda_{d}\\max_{i,j = 1,\\ldots ,N_{D}}\\sigma \\left(\\hat{\\mathbf{z}} (\\theta^{(i)}),\\hat{\\mathbf{z}} (\\theta^{(j)})\\right) \\\\ + \\sum_ {i = 1} ^ {N _ {D}} \\left\\{- \\lambda_ {r} Q _ {\\tilde {q}, w \\sim \\mu} \\left[ \\sigma \\left(\\hat {\\mathbf {z}} \\left(\\theta^ {(i)}\\right), w\\right) \\right] + \\| \\hat {\\mathbf {z}} \\left(\\theta^ {(i)}\\right) - \\mathbf {z} _ {0} ^ {(i)} \\| ^ {2} \\right\\}, \\tag {6} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.581, + 0.469, + 0.701 + ], + "angle": 0, + "content": "where \\( Q_{\\tilde{q}}[\\cdot] \\) denotes the quantile function with \\( \\tilde{q} \\) as the quantile percentage. Note that here we consider a theoretical idealized variant of our algorithm wherein we perform simultaneous generation of all the embeddings \\( \\{\\hat{\\mathbf{z}}(\\theta^{(i)})\\} \\), rather than sample by sample. Hence the objectives turn to the sum of pairwise similarities rather than the form in Eq. (4). And we minimize the negative to aim for maximal representativeness, as in Eq. (5)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.703, + 0.469, + 0.778 + ], + "angle": 0, + "content": "It can be considered as a scalarized solution to a multi-objective optimization problem, wherein multiple criteria are weighed (see, e.g. [13]). This perspective aligns with a Pareto front with trade-offs. It means that one objective decreasing will by necessity result in another increasing." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.779, + 0.469, + 0.809 + ], + "angle": 0, + "content": "However, consider that any solution to the following tri-level optimization problem is also a solution for Eq. (6):" + }, + { + "type": "equation", + "bbox": [ + 0.086, + 0.815, + 0.469, + 0.901 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\min _ {\\{\\theta^ {(i)} \\} _ {i \\in [ N _ {D} ]}} \\max _ {i, j = 1,.., N _ {D}} \\sigma \\left(\\hat {\\mathbf {z}} (\\theta^ {(i)}), \\hat {\\mathbf {z}} (\\theta^ {(j)})\\right) \\\\ \\text {s u b j . t o} \\quad \\left\\{\\theta^ {(i)} \\right\\} \\in \\arg \\max \\sum_ {i = 1} ^ {N _ {D}} Q _ {\\tilde {q}, w \\sim \\mu} \\left[ \\sigma (\\hat {\\mathbf {z}} (\\theta^ {(i)}), w) \\right] \\\\ \\text {s u b j . t o} \\quad \\theta^ {(i)} \\in \\arg \\min \\| \\hat {\\mathbf {z}} (\\theta) - \\mathbf {z} _ {0} ^ {(i)} \\| ^ {2}, \\forall i \\in [ N _ {D} ]. \\tag {7} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.181 + ], + "angle": 0, + "content": "If a solution to Eq. (7) is discovered, either incidentally through solving Eq. (6) or by careful tuning of step sizes, the set of minimizers will be sufficiently large at both levels, with no trade-offs involved. However, can we justify the presumption that there exists a meaningful set of potential minimizers?" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.2, + 0.892, + 0.305 + ], + "angle": 0, + "content": "Diffusion Process Model One popular framework for the mathematical analysis of diffusion involves analyzing the convergence and asymptotic properties of, appropriately homonymous, diffusion processes. These processes are characterized by the standard stochastic differential equation with a drift and diffusion term. For a time-dependent random variable \\( Z_{t} \\)," + }, + { + "type": "equation", + "bbox": [ + 0.617, + 0.311, + 0.891, + 0.328 + ], + "angle": 0, + "content": "\\[\nd Z _ {t} = V \\left(Z _ {t}\\right) d t + d W _ {t} \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.334, + 0.892, + 0.438 + ], + "angle": 0, + "content": "where \\(V\\) is a drift function dependent on the current \\(Z_{t}\\) and \\(dW_{t}\\) is a Wiener (Brownian noise) process. This equation serves as an appropriate continuous approximation of generative diffusion, given that Brownian noise is a continuous limit of adding normal random variables. Consequently, we aim for any realization \\(\\mathbf{z} \\sim Z_{t}\\) to have certain desired properties that reflect generative modeling with high probability." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.439, + 0.892, + 0.544 + ], + "angle": 0, + "content": "The work [43] established a theoretical model utilizing concepts in the control of these diffusions, demonstrating how it can result in sampling from the distribution of a desired data set. In the supplementary material we present a description of their framework and present an argument supporting the well-defined nature of the following problem, indicating that it has non-trivial solutions." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.545, + 0.892, + 0.65 + ], + "angle": 0, + "content": "When sampling optimally from the population dataset, we consider a stochastic control problem wherein \\(V\\) depends also on some chosen control \\(u(\\mathbf{z},t)\\). This control aims to find the most representative samples and, among the possible collection of such samples, to obtain the most diverse one while sampling from the desired dataset \\(\\mu\\). This involves solving:" + }, + { + "type": "equation", + "bbox": [ + 0.526, + 0.655, + 0.753, + 0.68 + ], + "angle": 0, + "content": "\\[\n\\min _ {u (x, t)} \\max _ {i, j = 1,.., N _ {D}} \\sigma \\left(Z _ {1} ^ {u, (i)}, Z _ {1} ^ {u, (j)}\\right)\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.51, + 0.681, + 0.889, + 0.751 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathrm {s u b j . t o} \\quad u \\in \\arg \\max \\sum_ {i = 1} ^ {N _ {D}} \\int_ {0} ^ {1} \\mathbb {E} _ {Z _ {t} ^ {(i)}} Q _ {\\tilde {q}, w \\sim \\mu} \\left[ \\sigma \\left(Z _ {t} ^ {(i)}, w\\right) \\right] d s \\\\ Z _ {1} \\sim \\mu , \\\\ d Z _ {t} ^ {u, (i)} = u \\left(Z _ {t} ^ {u, (i)}, t\\right) d t + d W _ {t}, t \\in [ 0, 1 ]; \\\\ Z _ {0} = \\mathbf {z} _ {0}. \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.872, + 0.752, + 0.891, + 0.764 + ], + "angle": 0, + "content": "(9)" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.892, + 0.901 + ], + "angle": 0, + "content": "This problem poses a bi-level stochastic control challenge where employing a layered dynamic programming is far from tractable. Additionally, a multi-stage stochastic programming approximation would also be infeasible given the scale of the datasets involved. Instead, we opt for parameterization with a neural network, forego exact sampling, discretize the problem, redefine the criteria to be time independent, and seek to solve an approximate solution for the tri-level optimization problem Eq. (7)." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.925, + 0.519, + 0.937 + ], + "angle": 0, + "content": "15796" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.089, + 0.892, + 0.119 + ], + "angle": 0, + "content": "Table 1. Performance comparison with pre-trained diffusion models and other state-of-the-art methods on ImageWoof. All the results are reproduced by us on the \\(256\\times 256\\) resolution. The missing results are due to out-of-memory. The best results are marked as bold." + }, + { + "type": "table", + "bbox": [ + 0.081, + 0.13, + 0.892, + 0.384 + ], + "angle": 0, + "content": "
IPC (Ratio)Test ModelRandomK-Center [37]Herding [48]DiT [33]DM [56]IDC-1 [21]GLaD [4]OursFull
10 (0.8%)ConvNet-624.3±1.119.4±0.926.7±0.534.2±1.126.9±1.233.3±1.133.8±0.937.0±1.086.4±0.2
ResNetAP-1029.4±0.822.1±0.132.0±0.334.7±0.530.3±1.239.1±0.532.9±0.939.2±1.387.5±0.5
ResNet-1827.7±0.921.1±0.430.2±1.234.7±0.433.4±0.737.3±0.231.7±0.837.6±0.989.3±1.2
20 (1.6%)ConvNet-629.1±0.721.5±0.829.5±0.336.1±0.829.9±1.035.5±0.8-37.6±0.286.4±0.2
ResNetAP-1032.7±0.425.1±0.734.9±0.141.1±0.835.2±0.643.4±0.3-45.8±0.587.5±0.5
ResNet-1829.7±0.523.6±0.332.2±0.640.5±0.529.8±1.738.6±0.2-42.5±0.689.3±1.2
50 (3.8%)ConvNet-641.3±0.636.5±1.040.3±0.746.5±0.844.4±1.043.9±1.2-53.9±0.686.4±0.2
ResNetAP-1047.2±1.340.6±0.449.1±0.749.3±0.247.1±1.148.3±1.0-56.3±1.087.5±0.5
ResNet-1847.9±1.839.6±1.048.3±1.250.1±0.546.2±0.648.3±0.8-57.1±0.689.3±1.2
70 (5.4%)ConvNet-646.3±0.638.6±0.746.2±0.650.1±1.247.5±0.848.9±0.7-55.7±0.986.4±0.2
ResNetAP-1050.8±0.645.9±1.553.4±1.454.3±0.951.7±0.852.8±1.8-58.3±0.287.5±0.5
ResNet-1852.1±1.044.6±1.149.7±0.851.5±1.051.9±0.851.1±1.7-58.8±0.789.3±1.2
100 (7.7%)ConvNet-652.2±0.445.1±0.554.4±1.153.4±0.355.0±1.353.2±0.9-61.1±0.786.4±0.2
ResNetAP-1059.4±1.054.8±0.261.7±0.958.3±0.856.4±0.856.1±0.9-64.5±0.287.5±0.5
ResNet-1861.5±1.350.4±0.459.3±0.758.9±1.360.2±1.058.3±1.2-65.7±0.489.3±1.2
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.408, + 0.471, + 0.501 + ], + "angle": 0, + "content": "In the supplementary material we provide a rationale for the meaningfulness of the problem in Eq. (9) based on the model of generative diffusion [43]. Specifically, we argue that the set of controls that leads to the desired final distribution and the set of minimizers, is sufficiently large for a low value of the objective at the top layer." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.515, + 0.21, + 0.532 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.54, + 0.295, + 0.557 + ], + "angle": 0, + "content": "4.1. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.563, + 0.47, + 0.716 + ], + "angle": 0, + "content": "For the diffusion model, we adopt pre-trained DiT [33] as the baseline and conduct PEFT with Difffit [50]. \\(\\lambda_r\\) and \\(\\lambda_d\\) are set as 0.002 and 0.008 for Eq. (5), respectively. The image size for the diffusion fine-tuning and sample generation is set as \\(256\\times 256\\). The fine-tuning mini-batch size is set as 8, and the fine-tuning lasts 8 epochs. The learning rate is set as 1e-3 for an AdamW optimizer. After fine-tuning, the images are generated by 50 denoising steps on a pre-defined number of random noise, according to the IPC setting. All the experiments are conducted on a single RTX 4090 GPU." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.726, + 0.358, + 0.741 + ], + "angle": 0, + "content": "4.2. Datasets and Evaluation Metric" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.471, + 0.903 + ], + "angle": 0, + "content": "For practical applicability, the experiments are exclusively conducted on full-sized ImageNet [6] subsets in this work. The selected subsets include ImageWoof, ImageNette [15] and the 10-class split adopted in [21, 42], denoted as ImageIDC afterward. ImageWoof is a challenging subset, containing only classes of dog breeds, while ImageNette and ImageIDC contain classes with less similarity, and hence are easier to discriminate. For evaluation, we adopt the same setting as in [21]. The surrogate dataset is trained on different model architectures, with a learning rate of 0.01," + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.407, + 0.892, + 0.435 + ], + "angle": 0, + "content": "Table 2. The Maximum Mean Discrepancy (MMD) between the extracted features of surrogate dataset and the original one." + }, + { + "type": "table", + "bbox": [ + 0.505, + 0.446, + 0.89, + 0.505 + ], + "angle": 0, + "content": "
IPCDiT [33]Diffit [50]DM [56]IDC-1 [21]Ours
505.45.44.86.74.0
1005.55.34.06.44.3
" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.531, + 0.892, + 0.579 + ], + "angle": 0, + "content": "and a scheduler decaying the learning rate at 2/3 and 5/6 of the whole training iterations. The top-1 accuracy on the original testing set is reported to illustrate the performance." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.589, + 0.866, + 0.605 + ], + "angle": 0, + "content": "4.3. Comparison with State-of-the-art Methods" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.613, + 0.892, + 0.794 + ], + "angle": 0, + "content": "We compare our method with other state-of-the-art methods across different IPC settings and model architectures. For a fair comparison, the results are all reproduced by us under the same evaluation protocol. ResNet-10 [18] with average pooling is adopted for matching the feature distribution (DM [56], GLaD [4]) and training gradients (IDC-1 [21]). DM is implemented on IDC-1 by only modifying the matching objective from training gradients to feature distribution, such that better performance is achieved. Each experiment is conducted 3 times, with the mean value and standard variance reported. Firstly, we present the validation results on the challenging ImageWoof subset [15] in Tab. 1." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.796, + 0.892, + 0.901 + ], + "angle": 0, + "content": "With the target of distilling surrogate datasets of small IPCs (e.g., 10 and 20), the pixel-level optimization method IDC-1 [21] demonstrates outstanding performance gain over random original images. However, as the IPC increases, the performance gain drastically drops. Especially under the 100-IPC setting, the distilled dataset even performs worse than random original images. This observation" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.926, + 0.519, + 0.938 + ], + "angle": 0, + "content": "15797" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.089, + 0.472, + 0.145 + ], + "angle": 0, + "content": "Table 3. Performance comparison with pre-trained diffusion models and state-of-the-art methods on more ImageNet subsets. The results are obtained on ResNet-10 with average pooling. The best results are marked as **bold**." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.157, + 0.468, + 0.277 + ], + "angle": 0, + "content": "
IPCRandomDiT [33]DM [56]Ours
Nette1054.2±1.659.1±0.760.8±0.662.0±0.2
2063.5±0.564.8±1.266.5±1.166.8±0.4
5076.1±1.173.3±0.976.2±0.476.6±0.2
IDC1048.1±0.854.1±0.452.8±0.553.1±0.2
2052.5±0.958.9±0.258.5±0.459.0±0.4
5068.1±0.764.3±0.669.1±0.869.6±0.2
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.306, + 0.47, + 0.442 + ], + "angle": 0, + "content": "aligns with the empirical findings in Fig. 2, where pixel-level methods struggle to optimize the expanded parameter space of large IPCs. The embedding-level optimization method GLaD [4] yields good performance under the 10-IPC setting. However, it requires overwhelming GPU resources for larger IPC settings, which is inapplicable for resource-restricted scenarios. It is also notable that on large IPCs, the coreset method Herding [48] surpasses previous DD methods with far less computational cost." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.444, + 0.47, + 0.672 + ], + "angle": 0, + "content": "The pre-trained DiT [33] here serves as the baseline for generative diffusion techniques. Under the 50-IPC setting, DiT outperforms both random original images and IDC-1. However, the insufficiency of representativeness and diversity restricts its performance on smaller and larger IPCs, respectively. In contrast, our proposed minimax diffusion consistently provides superior performance across all the IPCs over both original images and Herding. Besides, the proposed method eliminates the need of specific network architectures for matching training metrics. Consequently, the cross-architecture generalization is significantly improved. Under most IPC settings, the performance gap between ConvNet-6 and ResNetAP-10 is even smaller than that of the original images. It validates the universality of the rich information learned by the minimax fine-tuning process." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.672, + 0.47, + 0.794 + ], + "angle": 0, + "content": "Furthermore, we extensively assess the Maximum Mean Discrepancy (MMD) between the embedded features of the selected/generated surrogate dataset and the original one in Tab. 2. The features are extracted by a ResNet-10 network pre-trained on the full original dataset. Our method achieves the lowest discrepancy by average, where DM [56] directly sets MMD as the optimization target, proving the validity of extra minimax criteria in fitting distributions." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.796, + 0.468, + 0.901 + ], + "angle": 0, + "content": "Moreover, we show the performance comparison on ImageNette [15] and ImageIDC [21] in Tab. 3. The performance trend generally aligns with that on ImageWoof. More specifically, on these two easier subsets, DiT quickly loses the advantage over original images as IPC increases. Conversely, our proposed minimax diffusion method consistently demonstrates state-of-the-art performance." + }, + { + "type": "table_caption", + "bbox": [ + 0.542, + 0.09, + 0.851, + 0.104 + ], + "angle": 0, + "content": "Table 4. Performance comparison on ImageNet-1K." + }, + { + "type": "table", + "bbox": [ + 0.515, + 0.115, + 0.88, + 0.174 + ], + "angle": 0, + "content": "
IPCSRe2L [53]RDED [41]DiTOurs
1021.3±0.642.0±0.139.6±0.444.3±0.5
5046.8±0.256.5±0.152.9±0.658.6±0.3
" + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.19, + 0.705, + 0.407 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.706, + 0.19, + 0.885, + 0.407 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.424, + 0.892, + 0.493 + ], + "angle": 0, + "content": "Figure 4. With the help of the minimax diffusion, the proposed method significantly enhances the representativeness and diversity of the generated images. Thereby it consistently provides superior performance compared with random selection and baseline diffusion models by a large margin across different IPC settings." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.521, + 0.892, + 0.612 + ], + "angle": 0, + "content": "Experiments on ImageNet-1K. We further conduct experiments on the full ImageNet-1K with the validation protocol of RDED [41] and present the results in Tab. 4. The synthetic images are resized to \\(224 \\times 224\\) for evaluation. The significant performance advantage over the compared works validates the scalability of the proposed method." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.621, + 0.655, + 0.637 + ], + "angle": 0, + "content": "4.4. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.644, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Component Analysis. We compare the performance with baseline diffusion models to validate the effectiveness of proposed minimax criteria in Fig. 4. The experiments are conducted on ImageWoof and ImageIDC to evaluate the effect on challenging and easy tasks, respectively. Under the IPC of 10 and 20, the raw diffusion models (DiT) generate informative images, with validation performance much higher than randomly selected original samples. However, as the IPC is continuously increased, the performance gap diminishes for ImageWoof, and random original images surpass the DiT-generated ones at the IPC of 100. On ImageIDC the intersection occurs even earlier at the IPC of 50. The main reason is reflected in Fig. 3, where the sample diversity remains limited without external guidance. The naive Difffit fine-tuning adapts the model to specific domains, yet on large IPCs, the over-fitted generative model still yields inferior performance than the original images." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.926, + 0.519, + 0.938 + ], + "angle": 0, + "content": "15798" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.104, + 0.09, + 0.87, + 0.332 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.343, + 0.895, + 0.388 + ], + "angle": 0, + "content": "Figure 5. Visualization of random original images, images generated by baseline diffusion models (DiT [33] and Difffit [50]) and our proposed method. For each column, the generated images are based on the same random seed. Comparatively, our method significantly enhances the coverage of original data distribution and the diversity of the surrogate dataset." + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.41, + 0.471, + 0.452 + ], + "angle": 0, + "content": "Table 5. The ablation study of the proposed minimax scheme. The result are obtained with ResNet-10 on ImageWoof and ImageIDC. m denotes the proposed minimax optimization scheme." + }, + { + "type": "table", + "bbox": [ + 0.08, + 0.463, + 0.47, + 0.606 + ], + "angle": 0, + "content": "
\\( {\\mathcal{L}}_{r} \\)\\( {\\mathcal{L}}_{r} \\) w\\m\\( {\\mathcal{L}}_{d} \\)\\( {\\mathcal{L}}_{d} \\) w\\mImageWoofImageIDC
10-IPC50-IPC10-IPC50-IPC
----35.6±0.951.0±0.953.5±0.266.3±0.2
✓---34.4±1.147.1±0.549.6±0.760.2±1.2
-✓--37.4±0.449.5±1.054.5±1.265.0±0.8
--✓-35.7±0.848.3±0.651.5±0.664.8±0.8
---✓38.7±0.954.9±0.752.2±0.668.4±0.7
✓-✓-38.3±0.554.9±0.453.3±0.566.8±0.5
-✓-✓39.2±1.356.3±1.053.1±0.269.6±0.2
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.64, + 0.47, + 0.852 + ], + "angle": 0, + "content": "The addition of representativeness constraint to the training process further enhances the effect of distribution fitting. At small IPCs, the generated images contain richer information, yet for larger IPCs, the lack of diversity brings a negative influence. The diversity constraint, in contrast, significantly boosts the information contained in the generated surrogate dataset. Despite the performance advantage of \\(\\mathcal{L}_d\\) over \\(\\mathcal{L}_r\\), combining them still brings stable improvement as our full method. Especially on the easier ImageIDC task, grouping these two constraints together contributes to a consistent performance margin over random original images. The experimental results validate that both representativeness and diversity play essential parts in constructing effective surrogate datasets." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Minimax Scheme. In this work, we propose to enhance the representativeness and diversity each with a minimax objective. We compare the distillation result with or with-" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.412, + 0.895, + 0.563 + ], + "angle": 0, + "content": "out the minimax operation in Tab. 5. The first row presents the performance of naive Difffit fine-tuning. Matching the embeddings to the distribution center as in Eq. (2) severely degrades the validation performance across all IPCs. In contrast, the minimax version constraint as in Eq. (3) encourages better coverage, where the performance on small IPCs is improved. The effects of diversity constraint and the full method show similar trends. The superior performance suggests the effectiveness in enhancing the essential properties of the generative diffusion techniques." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.574, + 0.637, + 0.589 + ], + "angle": 0, + "content": "4.5. Visualization" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.598, + 0.892, + 0.84 + ], + "angle": 0, + "content": "Sample Distribution Visualization. The target of our proposed method is to construct a surrogate dataset with both representativeness and diversity. We visualize the t-SNE distribution of samples generated by our proposed method in Fig. 3. In comparison with random original images and baseline diffusion models, our method demonstrates a more thorough coverage over the entire data distribution while maintaining consistency in sample density. At the original high-density region, the generated images also form a dense sub-cluster, which is not reflected by random sampling. On the other hand, at the original sparse regions, our method exhibits better coverage than baseline diffusion models. By simultaneously enhancing the representativeness and diversity in the generative model, the proposed method manages to significantly improve the validation performance of the generated surrogate dataset." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.895, + 0.901 + ], + "angle": 0, + "content": "Generated Sample Comparison. The proposed method notably enhances the representativeness and diversity of the generated surrogate dataset. We compare the samples generated with the same random noise (for each column) of" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.926, + 0.519, + 0.938 + ], + "angle": 0, + "content": "15799" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.091, + 0.273, + 0.256 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.281, + 0.091, + 0.47, + 0.256 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.478, + 0.091, + 0.668, + 0.256 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.677, + 0.092, + 0.868, + 0.256 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.268, + 0.892, + 0.297 + ], + "angle": 0, + "content": "Figure 6. Hyper-parameter analysis on (a) the training epochs; (b) the representativeness weight \\(\\lambda_r\\); (c) the diversity weight \\(\\lambda_d\\); (d) the memory size \\(N_M\\). The results are obtained with ResNetAP-10 on ImageWoof. The dashed line indicates the value adopted in this work." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.323, + 0.468, + 0.352 + ], + "angle": 0, + "content": "different generative methods in Fig. 5 to explicitly demonstrate the improved properties." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.353, + 0.468, + 0.655 + ], + "angle": 0, + "content": "The images generated by baseline DiT exhibit a realistic high-quality appearance. However, the images tend to share similar poses and only present the most prominent features of the objects. In the golden retriever case, the generated images mostly present the head part, while for the churches the exterior appearance. Difffit fine-tuning further fits the model to the distribution, but in most cases, the differences only lie in small details. Comparatively, the proposed minimax criteria significantly enhance both the representativeness and diversity of the generated images. On the one hand, there occurs more class-related content in the generated images. The golden retriever images include more body parts and the church images encompass the interior layout. The minimax optimization leads to better coverage over the entire original distribution, with more related features encapsulated. On the other hand, the diversity is significantly enhanced, including variations in pose, background, and appearance styles. In such a way the surrogate dataset better represents the original large-scale one, leading to superior validation performance." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.656, + 0.468, + 0.791 + ], + "angle": 0, + "content": "Training Curve Visualization. We visualize the accuracy curve during the training process in Fig. 6a. The validation performance is rapidly improved as the fine-tuning process starts. After four epochs, the model tends to converge and reaches the highest performance at the 8th epoch. Further extending the training epochs injects excessive diversity into the model, leading to performance degradation. We demonstrate the influence of training epochs on the generated images in supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.802, + 0.265, + 0.818 + ], + "angle": 0, + "content": "4.6. Parameter Analysis" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.468, + 0.9 + ], + "angle": 0, + "content": "Objective Weight \\(\\lambda_r\\lambda_d\\) .We show the influence of representativeness weight \\(\\lambda_r\\) and diversity weight \\(\\lambda_{d}\\) in Fig. 6b and Fig. 6c, respectively. The \\(\\lambda_r\\) variation only produces negligible performance fluctuation on small IPCs, while on large IPCs the performance is also relatively stable. For" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.323, + 0.892, + 0.443 + ], + "angle": 0, + "content": "\\(\\lambda_{d}\\), at a proper variation range, the performance is stable. However, continuously increasing the diversity of the generated dataset leads to a lack of representativeness, which results in a negative impact. The negative impact of overdiversity can also validated by the poor performance of K-Center in Tab. 1. A uniform performance decrease is observed as \\(\\lambda_{d}\\) reaches 0.03. Based on the performance of 100 IPC, we set \\(\\lambda_{r}\\) as 0.002 and \\(\\lambda_{d}\\) as 0.008." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.445, + 0.892, + 0.626 + ], + "angle": 0, + "content": "Memory Size \\( N_{M} \\). The memory size \\( N_{M} \\) influences the number of samples involved in the objective calculation. We investigate its influence in Fig. 6d. When the memory is extremely small (\\( N_{M} = 16 \\)), the provided supervision is also limited, yet the performance is already higher than naive fine-tuning. As the memory size is increased in a proper range, the model yields stable performance improvement. It is notable that with a larger memory, the performance under the IPC of 10 is better. It can be explained by that a larger memory contains more representative information. Out of the consideration of performance as well as storage burden, we select the memory of 64 in the other experiments." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.644, + 0.618, + 0.66 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.671, + 0.892, + 0.822 + ], + "angle": 0, + "content": "In this work, we propose a novel dataset distillation method based on generative diffusion techniques. Through extra minimax criteria, the proposed method significantly enhances the representativeness and diversity of the generated surrogate dataset. With much less computational time consumption, the proposed method achieves state-of-the-art validation performance on challenging ImageNet subsets. It reduces the resource dependency of previous dataset distillation methods and opens up new possibilities for more practical applications for distilling personalized data." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.824, + 0.891, + 0.869 + ], + "angle": 0, + "content": "Limitations and Future Works. In addition to object classification, we will explore the possibility of incorporating generative techniques for more specific data domains." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.891, + 0.901 + ], + "angle": 0, + "content": "Acknowledgement. We would like to specially thank Olga Russakovsky for invaluable advice on the manuscript." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.926, + 0.519, + 0.938 + ], + "angle": 0, + "content": "15800" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.079, + 0.09, + 0.176, + 0.107 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.116, + 0.47, + 0.176 + ], + "angle": 0, + "content": "[1] Shekoofeh Azizi, Simon Kornblith, Chitwan Sahara, Mohammad Norouzi, and David J Fleet. Synthetic data from diffusion models improves imagenet classification. arXiv preprint arXiv:2304.08466, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.178, + 0.472, + 0.238 + ], + "angle": 0, + "content": "[2] Victor Besnier, Himalaya Jain, Andrei Bursuc, Matthieu Cord, and Patrick Pérez. This Dataset Does Not Exist: Training Models from Generated Images. In ICASSP, pages 1-5, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.089, + 0.24, + 0.471, + 0.3 + ], + "angle": 0, + "content": "[3] George Cazenavette, Tongzhou Wang, Antonio Torralba, Alexei A Efros, and Jun-Yan Zhu. Dataset distillation by matching training trajectories. In CVPR, pages 4750-4759, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.089, + 0.301, + 0.471, + 0.361 + ], + "angle": 0, + "content": "[4] George Cazenavette, Tongzhou Wang, Antonio Torralba, Alexei A Efros, and Jun-Yan Zhu. Generalizing dataset distillation via deep generative prior. In CVPR, pages 3739-3748, 2023. 1, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.089, + 0.362, + 0.47, + 0.407 + ], + "angle": 0, + "content": "[5] Justin Cui, Ruochen Wang, Si Si, and Cho-Jui Hsieh. Dc-bench: Dataset condensation benchmark. NeurIPS, 35:810-822, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.089, + 0.409, + 0.47, + 0.468 + ], + "angle": 0, + "content": "[6] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. ImageNet: A large-scale hierarchical image database. In CVPR, pages 248-255, 2009. 1, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.089, + 0.471, + 0.47, + 0.516 + ], + "angle": 0, + "content": "[7] Zhiwei Deng and Olga Russakovsky. Remember the Past: Distilling Datasets into Addressable Memories for Neural Networks. In NeurIPS, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.089, + 0.518, + 0.47, + 0.563 + ], + "angle": 0, + "content": "[8] Prafulla Dhariwal and Alexander Nichol. Diffusion Models Beat GANs on Image Synthesis. In NeurIPS, pages 8780-8794, 2021. 3, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.089, + 0.565, + 0.47, + 0.623 + ], + "angle": 0, + "content": "[9] Thinh T Doan. Nonlinear two-time-scale stochastic approximation convergence and finite-time performance. IEEE Transactions on Automatic Control, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.626, + 0.47, + 0.731 + ], + "angle": 0, + "content": "[10] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale. In ICLR, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.733, + 0.47, + 0.793 + ], + "angle": 0, + "content": "[11] Jiawei Du, Yidi Jiang, Vincent Y.F. Tan, Joey Tianyi Zhou, and Haizhou Li. Minimizing the Accumulated Trajectory Error to Improve Dataset Distillation. In CVPR, pages 3749-3758, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.794, + 0.47, + 0.868 + ], + "angle": 0, + "content": "[12] Lisa Dunlap, Alyssa Umino, Han Zhang, Jiezhi Yang, Joseph E Gonzalez, and Trevor Darrell. Diversify your vision datasets with automatic diffusion-based augmentation. arXiv preprint arXiv:2305.16289, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.871, + 0.47, + 0.902 + ], + "angle": 0, + "content": "[13] Gabriele Eichfelder. Scalarizations for adaptively solving multi-objective optimization problems. Com-" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.116, + 0.472, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.892, + 0.121 + ], + "angle": 0, + "content": "putational Optimization and Applications, 44:249-273, 2009. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.124, + 0.894, + 0.182 + ], + "angle": 0, + "content": "[14] Ronen Eldan and James R Lee. Regularization under diffusion and anticoncentration of the information content. Duke Mathematical Journal, 167(5):969-993, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.184, + 0.892, + 0.228 + ], + "angle": 0, + "content": "[15] Fastai. Fastai/imagenette: A smaller subset of 10 easily classified classes from imagenet, and a little more french. 1, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.231, + 0.894, + 0.289 + ], + "angle": 0, + "content": "[16] Jianyang Gu, Kai Wang, Wei Jiang, and Yang You. Summarizing stream data for memory-restricted online continual learning. arXiv preprint arXiv:2305.16645, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.291, + 0.892, + 0.351 + ], + "angle": 0, + "content": "[17] Swaminathan Gurumurthy, Ravi Kiran Sarvadevabhatla, and R. Venkatesh Babu. DeLiGAN: Generative Adversarial Networks for Diverse and Limited Data. In CVPR, pages 4941-4949, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.353, + 0.892, + 0.397 + ], + "angle": 0, + "content": "[18] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, pages 770-778, 2016. 1, 5, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.399, + 0.892, + 0.457 + ], + "angle": 0, + "content": "[19] Ruifei He, Shuyang Sun, Xin Yu, Chuhui Xue, Wenqing Zhang, Philip Torr, Song Bai, and Xiaojuan Qi. Is Synthetic Data from Generative Models Ready for Image Recognition? In ICLR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.459, + 0.892, + 0.503 + ], + "angle": 0, + "content": "[20] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising Diffusion Probabilistic Models. In NeurIPS, pages 6840-6851, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.505, + 0.892, + 0.579 + ], + "angle": 0, + "content": "[21] Jang-Hyun Kim, Jinuk Kim, Seong Joon Oh, Sang-doo Yun, Hwanjun Song, Joonhyun Jeong, Jung-Woo Ha, and Hyun Oh Song. Dataset condensation via efficient synthetic-data parameterization. In ICML, pages 11102-11118, 2022. 1, 5, 6, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.581, + 0.892, + 0.626 + ], + "angle": 0, + "content": "[22] Diederik Kingma, Tim Salimans, Ben Poole, and Jonathan Ho. Variational Diffusion Models. In NeurIPS, pages 21696-21707, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.628, + 0.892, + 0.686 + ], + "angle": 0, + "content": "[23] Saehyung Lee, Sanghyuk Chun, Sangwon Jung, Sang-doo Yun, and Sungroh Yoon. Dataset condensation with contrastive signals. In ICML, pages 12352-12364, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.688, + 0.892, + 0.734 + ], + "angle": 0, + "content": "[24] Joseph Lehec. Representation formula for the entropy and functional inequalities. In Annales de l'IHP Probabilités et statistiques, pages 885-899, 2013. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.735, + 0.892, + 0.778 + ], + "angle": 0, + "content": "[25] Ping Liu, Xin Yu, and Joey Tianyi Zhou. Meta Knowledge Condensation for Federated Learning. In ICLR, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.78, + 0.892, + 0.839 + ], + "angle": 0, + "content": "[26] Yanqing Liu, Jianyang Gu, Kai Wang, Zheng Zhu, Wei Jiang, and Yang You. Dream: Efficient dataset distillation by representative matching. In ICCV, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.841, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[27] Noel Loo, Ramin Hasani, Alexander Amini, and Daniela Rus. Efficient dataset distillation using random feature approximation. NeurIPS, 35:13877-13891, 2022. 1" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.894, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.927, + 0.518, + 0.938 + ], + "angle": 0, + "content": "15801" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.151 + ], + "angle": 0, + "content": "[28] Noel Loo, Ramin Hasani, Mathias Lechner, and Daniela Rus. Dataset distillation with convexified implicit gradients. arXiv preprint arXiv:2302.06755, 2023.1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.152, + 0.472, + 0.213 + ], + "angle": 0, + "content": "[29] Qi Mao, Hsin-Ying Lee, Hung-Yu Tseng, Siwei Ma, and Ming-Hsuan Yang. Mode Seeking Generative Adversarial Networks for Diverse Image Synthesis. In CVPR, pages 1429-1437, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.213, + 0.47, + 0.259 + ], + "angle": 0, + "content": "[30] Timothy Nguyen, Zhourong Chen, and Jaehoon Lee. Dataset meta-learning from kernel ridge-regression. In ICLR, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.26, + 0.47, + 0.32 + ], + "angle": 0, + "content": "[31] Timothy Nguyen, Roman Novak, Lechao Xiao, and Jaehoon Lee. Dataset distillation with infinitely wide convolutional networks. NeurIPS, 34:5186-5198, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.321, + 0.47, + 0.367 + ], + "angle": 0, + "content": "[32] Alexander Quinn Nichol and Prafulla Dhariwal. Improved Denoising Diffusion Probabilistic Models. In ICML, pages 8162-8171, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.367, + 0.47, + 0.412 + ], + "angle": 0, + "content": "[33] William Peebles and Saining Xie. Scalable diffusion models with transformers. In ICCV, pages 4195-4205, 2023. 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.413, + 0.47, + 0.489 + ], + "angle": 0, + "content": "[34] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In CVPR, pages 22500-22510, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.49, + 0.47, + 0.55 + ], + "angle": 0, + "content": "[35] Ahmad Sajedi, Samir Khaki, Ehsan Amjadian, Lucy Z Liu, Yuri A Lawryshyn, and Konstantinos N Plataniotis. DataDAM: Efficient Dataset Distillation with Attention Matching. In ICCV, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.55, + 0.469, + 0.625 + ], + "angle": 0, + "content": "[36] Mert Bulent Sariyildiz, Karteek Alahari, Diane Larlus, and Yannis Kalantidis. Fake it Till You Make it: Learning Transferable Representations from Synthetic ImageNet Clones. In CVPR, pages 8011-8021, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.627, + 0.47, + 0.687 + ], + "angle": 0, + "content": "[37] Ozan Sener and Silvio Savarese. Active learning for convolutional neural networks: A core-set approach. In International Conference on Learning Representations, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.688, + 0.47, + 0.748 + ], + "angle": 0, + "content": "[38] Allahkaram Shafiei, Vyacheslav Kungurtsev, and Jakub Marecek. Trilevel and multilevel optimization using monotone operator theory. arXiv preprint arXiv:2105.09407, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.749, + 0.47, + 0.824 + ], + "angle": 0, + "content": "[39] Jordan Shipard, Arnold Wiliem, Kien Nguyen Thanh, Wei Xiang, and Clinton Fookes. Diversity Is Definitely Needed: Improving Model-Agnostic Zero-Shot Classification via Stable Diffusion. In CVPR, pages 769-778, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.825, + 0.47, + 0.901 + ], + "angle": 0, + "content": "[40] Felipe Petroski Such, Aditya Rawal, Joel Lehman, Kenneth Stanley, and Jeffrey Clune. Generative Teaching Networks: Accelerating Neural Architecture Search by Learning to Generate Synthetic Training Data. In ICML, pages 9206-9216, 2020. 1" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.092, + 0.892, + 0.137 + ], + "angle": 0, + "content": "[41] Peng Sun, Bei Shi, Daiwei Yu, and Tao Lin. On the diversity and realism of distilled dataset: An efficient dataset distillation paradigm. In CVPR, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.14, + 0.892, + 0.184 + ], + "angle": 0, + "content": "[42] Yonglong Tian, Dilip Krishnan, and Phillip Isola. Contrastive multiview coding. In ECCV, pages 776-794, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.188, + 0.892, + 0.248 + ], + "angle": 0, + "content": "[43] Belinda Tzen and Maxim Raginsky. Theoretical guarantees for sampling and inference in generative models with latent diffusions. In \\(COLT\\), pages 3084-3114, 2019. 4, 5, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.251, + 0.892, + 0.327 + ], + "angle": 0, + "content": "[44] Saeed Vahidian, Mingyu Wang, Jianyang Gu, Vyacheslav Kungurtsev, Wei Jiang, and Yiran Chen. Group distributionally robust dataset distillation with risk minimization. arXiv preprint arXiv:2402.04676, 2024.1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.33, + 0.892, + 0.405 + ], + "angle": 0, + "content": "[45] Kai Wang, Bo Zhao, Xiangyu Peng, Zheng Zhu, Shuo Yang, Shuo Wang, Guan Huang, Hakan Bilen, Xin chao Wang, and Yang You. Cafe: Learning to condense dataset by aligning features. In CVPR, pages 12196-12205, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.409, + 0.892, + 0.468 + ], + "angle": 0, + "content": "[46] Kai Wang, Jianyang Gu, Daquan Zhou, Zheng Zhu, Wei Jiang, and Yang You. Dim: Distilling dataset into generative model. arXiv preprint arXiv:2303.04707, 2023.1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.472, + 0.892, + 0.517 + ], + "angle": 0, + "content": "[47] Tongzhou Wang, Jun-Yan Zhu, Antonio Torralba, and Alexei A Efros. Dataset distillation. arXiv preprint arXiv:1811.10959, 2018. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.52, + 0.892, + 0.551 + ], + "angle": 0, + "content": "[48] Max Welling. Herding dynamical weights to learn. In ICML, pages 1121-1128, 2009. 5, 6, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.553, + 0.892, + 0.599 + ], + "angle": 0, + "content": "[49] Xindi Wu, Zhiwei Deng, and Olga Russakovsky. Multimodal dataset distillation for image-text retrieval. arXiv preprint arXiv:2308.07545, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.602, + 0.892, + 0.678 + ], + "angle": 0, + "content": "[50] Enze Xie, Lewei Yao, Han Shi, Zhili Liu, Daquan Zhou, Zhaoqiang Liu, Jiawei Li, and Zhenguo Li. Difffit: Unlocking transferability of large diffusion models via simple parameter-efficient fine-tuning. arXiv preprint arXiv:2304.06648, 2023. 3, 5, 7, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.681, + 0.892, + 0.741 + ], + "angle": 0, + "content": "[51] Yuanhao Xiong, Ruochen Wang, Minhao Cheng, Felix Yu, and Cho-Jui Hsieh. FedDM: Iterative Distribution Matching for Communication-Efficient Federated Learning. In CVPR, pages 16323-16332, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.744, + 0.892, + 0.804 + ], + "angle": 0, + "content": "[52] Dingdong Yang, Seunghoon Hong, Yunseok Jang, Tianchen Zhao, and Honglak Lee. Diversity-Sensitive Conditional Generative Adversarial Networks. In ICLR, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.807, + 0.892, + 0.853 + ], + "angle": 0, + "content": "[53] Zeyuan Yin, Eric Xing, and Zhiqiang Shen. Squeeze, recover and relabel: Dataset condensation atImagenet scale from a new perspective. In NeurIPS, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.856, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[54] Bo Zhao and Hakan Bilen. Dataset condensation with differentiable siamese augmentation. In ICML, pages 12674-12685, 2021. 1, 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.092, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.927, + 0.52, + 0.938 + ], + "angle": 0, + "content": "15802" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.091, + 0.47, + 0.135 + ], + "angle": 0, + "content": "[55] Bo Zhao and Hakan Bilen. Synthesizing informative training samples with gan. arXiv preprint arXiv:2204.07513, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.138, + 0.47, + 0.182 + ], + "angle": 0, + "content": "[56] Bo Zhao and Hakan Bilen. Dataset condensation with distribution matching. In WACV, pages 6514-6523, 2023. 2, 5, 6, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.185, + 0.469, + 0.229 + ], + "angle": 0, + "content": "[57] Bo Zhao, Konda Reddy Mopuri, and Hakan Bilen. Dataset condensation with gradient matching. In ICLR, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.232, + 0.469, + 0.276 + ], + "angle": 0, + "content": "[58] Ganlong Zhao, Guanbin Li, Yipeng Qin, and Yizhou Yu. Improved Distribution Matching for Dataset Condensation. In CVPR, pages 7856-7865, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.278, + 0.468, + 0.322 + ], + "angle": 0, + "content": "[59] Yongchao Zhou, Ehsan Nezhadarya, and Jimmy Ba. Dataset distillation using neural feature regression. NeurIPS, 35:9813-9827, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.325, + 0.468, + 0.383 + ], + "angle": 0, + "content": "[60] Yongchao Zhou, Hshmat Sahak, and Jimmy Ba. Training on thin air: Improve image classification with generated data. arXiv preprint arXiv:2305.15316, 2023.2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.091, + 0.47, + 0.383 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.927, + 0.519, + 0.938 + ], + "angle": 0, + "content": "15803" + } + ] +] \ No newline at end of file diff --git a/2024/Efficient Dataset Distillation via Minimax Diffusion/75146b44-b76a-42b0-8456-4c37446317b7_origin.pdf b/2024/Efficient Dataset Distillation via Minimax Diffusion/75146b44-b76a-42b0-8456-4c37446317b7_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..e680b78d87c533e49d01f237818459f02e94a11b --- /dev/null +++ b/2024/Efficient Dataset Distillation via Minimax Diffusion/75146b44-b76a-42b0-8456-4c37446317b7_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b22c9f8fbbfaa9e96f98b56eee6c72128c387209f5a737acac9dc51e858f7f63 +size 2183983 diff --git a/2024/Efficient Dataset Distillation via Minimax Diffusion/full.md b/2024/Efficient Dataset Distillation via Minimax Diffusion/full.md new file mode 100644 index 0000000000000000000000000000000000000000..620745c358dd8af2f7cce86d8e4fb306176cb742 --- /dev/null +++ b/2024/Efficient Dataset Distillation via Minimax Diffusion/full.md @@ -0,0 +1,332 @@ +# Efficient Dataset Distillation via Minimax Diffusion + +Jianyang Gu1 Saeed Vahidian2 Vyacheslav Kungurtsev3 + +Haonan Wang $^{4}$ Wei Jiang $^{1*}$ Yang You $^{4}$ Yiran Chen $^{2}$ + +$^{1}$ Zhejiang University $^{2}$ Duke University $^{3}$ Czech Technical University + +$^{4}$ National University of Singapore + +{gu_jianyang, jiangwei_zju}@zju.edu.cn + +# Abstract + +Dataset distillation reduces the storage and computational consumption of training a network by generating a small surrogate dataset that encapsulates rich information of the original large-scale one. However, previous distillation methods heavily rely on the sample-wise iterative optimization scheme. As the images-per-class (IPC) setting or image resolution grows larger, the necessary computation will demand overwhelming time and resources. In this work, we intend to incorporate generative diffusion techniques for computing the surrogate dataset. Observing that key factors for constructing an effective surrogate dataset are representativeness and diversity, we design additional minimax criteria in the generative training to enhance these facets for the generated images of diffusion models. We present a theoretical model of the process as hierarchical diffusion control demonstrating the flexibility of the diffusion process to target these criteria without jeopardizing the faithfulness of the sample to the desired distribution. The proposed method achieves state-of-the-art validation performance while demanding much less computational resources. Under the 100-IPC setting on ImageWoof, our method requires less than one-twentieth the distillation time of previous methods, yet yields even better performance. Source code and generated data are available in https://github.com/vimar-gu/MinimaxDiffusion. + +# 1. Introduction + +Data, as a necessary resource for deep learning, has concurrently promoted algorithmic advancements while imposing challenges on researchers due to heavy demands on storage and computational resources [6, 10, 18, 47]. Confronted with the conflict between the requirement for high-precision + +![](images/552d665e9aa78a478bcd3eadac76eca61fc7600fb0dcb58363cc269e51885137.jpg) +Figure 1. The validation accuracy and distillation time of different methods on ImageWoof [15], with a number following each method denoting the Image-Per-Class (IPC) setting. Previous methods are restricted by the heavier running time and memory consumption as IPC grows larger. In comparison, our proposed method notably reduces the demanding computational resources and also achieves state-of-the-art validation performance. + +models and overwhelming resource demands, dataset distillation is proposed to condense the rich information of a large-scale dataset into a small surrogate one [5, 21, 47, 57]. Such a surrogate dataset is expected to achieve training performance comparable to that attained with the original one. + +Previous dataset distillation methods mostly engage in iterative optimization on fixed-number samples at the pixel level [21, 26, 30, 31, 44, 54, 57] or embedding level [4, 55]. However, the sample-wise iterative optimization scheme suffers from problems of two perspectives. (1) The parameter space of optimization is positively correlated with the size of the target surrogate dataset and the image resolution [3, 57]. Consequently, substantial time and computational resources are required for distilling larger datasets. As shown in Fig. 1, IDC-1 [21] takes over 90 hours to distill a 100-image-per-class (IPC) set from ImageWoof [15], while training on ImageWoof itself only requires a matter of + +![](images/d1a04696a1ff47117b25ec1488c8a7523afe466bfb34236d84cdbc1ed99caf80.jpg) +Figure 2. Sample images distilled by the pixel-level sample-wise optimization method DM [56] on ImageWoof. As the parameter space increases along with the Image-Per-Class (IPC) setting, with the same initialization, the appearance disparity between original and distilled images is smaller. + +hours. (2) The expanded parameter space also increases the optimization complexity. As shown in Fig. 2, while distillation yields significant information condensation under small IPC settings, the pixel modification diminishes when distilling larger-IPC datasets. The reduced disparity also leads to smaller performance gain compared with original images, with instances where the distilled set even performs worse. Especially when distilling data of fine-grained classes, the sample-wise optimization scheme fails to provide adequate discriminative information. These constraints severely hinder individual researchers from distilling personalized data. A more practical training scheme is urgently needed to facilitate the broader application of dataset distillation. + +In this work, we explore the possibility of incorporating generative diffusion techniques [20, 22, 32] to efficiently compute effective surrogate datasets. We first conduct empirical analysis on the suitability of data generated by raw diffusion models for training networks. Based on the observations, we conclude that constructing an effective surrogate dataset hinges on two key factors: representativeness and diversity. Accordingly, we design extra minimax criteria for the generative training to enhance the capability of generating more effective surrogate datasets without explicit prompt designs. The minimax criteria involve two aspects: enforcing the generated sample to be close to the farthest real sample, while being far away from the most similar generated one. We provide theoretical analysis to support that the proposed minimax scheme aims to solve a well defined problem with all the criteria, including the generative accuracy and the minimax criteria, can be targeted simultaneously without detriment to the others. + +Compared with the astronomical training time consumption of the sample-wise iterative optimization schemes, the proposed method takes less than 1 hour to distill a 100-IPC surrogate dataset for a 10-class ImageNet subset, including + +![](images/d0d68cd486e76069a23a5fac587f4be503f4f213d019b2eacbb53e1e11581cc9.jpg) + +![](images/92b58d4e47e34ee50d933a4ead78330b9d2721a6dd86e8615878255391c2d23a.jpg) +Figure 3. The feature distribution comparison of different image generation methods with the original set. The validation performance of each surrogate set is listed in the upper-right corner. + +![](images/16b215162ea890c298ab7256ef36a046addb177a11233a288ec15eef4ba2983d.jpg) + +![](images/1265cd04ab1e8d2ea1b2eecd3819ac9f56eacc6a98787ac2b8810a85d79ff094.jpg) + +the fine-tuning and image generation processes. Remarkably, the GPU consumption remains consistent across all IPC settings. Furthermore, the distilled surrogate dataset attains superior validation performance compared with other state-of-the-art methods. Especially on the challenging fine-grained ImageWoof subset, the proposed method outperforms the second-best DD method by $5.5\%$ and $8.1\%$ under the IPC settings of 70 and 100, respectively. The source code is provided in the supplementary material. + +The contributions of this work are summarized into: + +- We analyze the data generated by diffusion models, and emphasize the importance of representativeness and diversity for constructing effective surrogate datasets. +- We propose a novel dataset distillation scheme based on extra minimax criteria for diffusion models targeting the representativeness and diversity of generated data. +- We theoretically justify the proposed minimax criteria as enforceable without trade-offs in the generation quality of the individual data points. +- We conduct extensive experiments to validate that our proposed method achieves state-of-the-art performance while demanding significantly reduced training time in comparison to previous dataset distillation methods. + +# 2. Method + +# 2.1. Problem Definition + +The general purpose of dataset distillation is to generate a small surrogate dataset $S = \{(\mathbf{x}_i,y_i)\}_{i = 1}^{N_S}$ from a largescale one $\mathcal{T} = \{(\mathbf{x}_i,y_i)\}_{i = 1}^{N_T}$ [47, 57]. Here each $\mathbf{x}_i$ denotes an image with a corresponding class label $y_{i}$ , and + +$N_{S} \ll N_{T}$ . The surrogate dataset $S$ is expected to encapsulate substantial information from the original $\mathcal{T}$ , such that training a model on $S$ achieves performance comparable with that on $\mathcal{T}$ . After distilling, we train network models on $S$ and validate the performance on the original test set. + +# 2.2. Diffusion for Distillation + +Diffusion models learn a dataset distribution by gradually adding Gaussian noise to images and reversing back. Taking the latent diffusion model (LDM) as an example, given a training image $\mathbf{x}$ , the training process is separated into two parts. An encoder $E$ transforms the image into the latent space $\mathbf{z} = E(\mathbf{x})$ and a decoder $D$ reconstructs a latent code back to the image space $\hat{\mathbf{x}} = D(\mathbf{z})$ . The forward noisng process gradually adds noise $\epsilon \sim \mathcal{N}(\mathbf{0},\mathbf{I})$ to the original latent code $\mathbf{z}_0$ : $\mathbf{z}_t = \sqrt{\bar{\alpha}_t}\mathbf{z}_0 + \sqrt{1 - \bar{\alpha}_t\epsilon}$ , where $\bar{\alpha}_t$ is a hyper-parameter. Provided with a conditioning vector $\mathbf{c}$ encoded with class labels, the diffusion models are trained by the squared error between the predicted noise $\epsilon_{\theta}(\mathbf{z}_t,\mathbf{c})$ and the ground truth $\epsilon$ : + +$$ +\mathcal {L} _ {\text {s i m p l e}} = \left| \left| \epsilon_ {\theta} \left(\mathbf {z} _ {t}, \mathbf {c}\right) - \epsilon \right| \right| _ {2} ^ {2}, \tag {1} +$$ + +where $\epsilon_{\theta}$ is a noise prediction network parameterized by $\theta$ . Diffusion models are proven to generate images of higher quality compared with GANs [8]. There are also some Parameter Efficient Fine-Tuning (PEFT) methods updating a small number of model parameters in order for the model to be better adapted to specific data domains [34, 50]. We adopt DiT [33] as the baseline and Difffit [50] as the naive fine-tuning method for image generation. The generated images are compared with the original data from the perspective of embedding distribution in Fig. 3. + +The samples of random selection and pre-trained diffusion models present two extreme ends of the distribution. Random selection faithfully reflects the original distribution, yet fails to emphasize some high-density regions. In contrast, diffusion models are over-fitted to those dense areas, leaving a large part of the original distribution uncovered. We attribute these two distributions to two properties, respectively. The randomly selected data holds extraordinary diversity, and the diffusion-generated data shows representativeness to the original distribution. We claim that both properties are essential for constructing an effective surrogate dataset. By naive fine-tuning, Difffit better captures the representative regions but leaves more regions uncovered. To this end, we propose extra minimax criteria for the diffusion model to enhance both of the properties. + +# 2.3. Minimax Diffusion Criteria + +Based on the observation that representativeness and diversity are two key factors to construct an effective surrogate dataset, we accordingly design extra minimax criteria to enhance these two essential properties for the diffusion model. + +Algorithm 1: Minimax Diffusion Fine-tuning +Input: initialized model parameter $\theta$ original dataset $\mathcal{T} = \{(\mathbf{x},y)\}$ , encoder $E$ class encoder $E_{c}$ time step $t$ variance schedule $\bar{\alpha}_t$ real embedding memory $\mathcal{M}$ predicted embedding memory $\mathcal{D}$ +Output: optimized model parameter $\theta^{*}$ +for each step do Obtain the original embedding: $\mathbf{z}_0 = E(\mathbf{x})$ Obtain the class embedding: $\mathbf{c} = E_c(y)$ Sample random noise: $\epsilon \sim \mathcal{N}(\mathbf{0},\mathbf{I})$ Add noise to the embedding: $\mathbf{z}_t = \sqrt{\bar{\alpha}_t}\mathbf{z}_0 + \sqrt{1 - \bar{\alpha}_t}\epsilon$ Predict the noise $\epsilon_{\theta}(\mathbf{z}_{t},\mathbf{c})$ and recovered embedding $\hat{\mathbf{z}}_{\theta}(\mathbf{z}_{t},\mathbf{c}) = \mathbf{z}_{t} - \epsilon_{\theta}(\mathbf{z}_{t},\mathbf{c})$ Update the model parameter with Eq. (5) Enqueue the real embedding: $\mathcal{M}_r\gets \mathbf{z}_0$ Enqueue the predicted embedding: $\mathcal{M}_d\gets \hat{\mathbf{z}}_\theta (\mathbf{z}_t,\mathbf{c})$ +end + +Representativeness It is essential for the small surrogate dataset to sufficiently represent the original data. A naive approach to improve the representativeness is aligning the embedding distribution between synthetic and real samples: + +$$ +\mathcal {L} _ {r} = \underset {\theta} {\arg \max } \sigma \left(\hat {\mathbf {z}} _ {\theta} \left(\mathbf {z} _ {t}, \mathbf {c}\right), \frac {1}{N _ {B}} \sum_ {i = 0} ^ {N _ {B}} \mathbf {z} _ {i}\right), \tag {2} +$$ + +where $\sigma (\cdot ,\cdot)$ is the cosine similarity, $\hat{\mathbf{z}}_{\theta}(\mathbf{z}_t,\mathbf{c})$ is the predicted original embedding by subtracting the noise from the noisy embedding $\hat{\mathbf{z}}_{\theta}(\mathbf{z}_t,\mathbf{c}) = \mathbf{z}_t - \epsilon_{\theta}(\mathbf{z}_t,\mathbf{c})$ , and $N_B$ is the size of the sampled real sample mini-batch. However, the naive alignment tends to draw the predicted embedding towards the center of the real distribution, which severely limits the diversity. Therefore, we propose to maintain an auxiliary memory $\mathcal{M} = \{\mathbf{z}_m\}_{m = 1}^{N_M}$ to store the real samples utilized in adjacent iterations, and design a minimax optimization objective as: + +$$ +\mathcal {L} _ {r} = \arg \max _ {\theta} \min _ {m \in \left[ N _ {M} \right]} \sigma \left(\hat {\mathbf {z}} _ {\theta} \left(\mathbf {z} _ {t}, \mathbf {c}\right), \mathbf {z} _ {m}\right). \tag {3} +$$ + +By pulling close the least similar sample pairs, the diffusion model is encouraged to generate images that better cover the original distribution. It is notable that the diffusion training objective $\mathcal{L}_{simple}$ itself encourages the generated images to resemble the original ones. Thus, the minimax criterion allows the preservation of diversity to the maximum extent. + +Diversity Although the pre-trained diffusion models already achieve satisfactory generation quality, the remaining defect is limited diversity compared with the original data, as shown in Fig. 3. We expect the data generated by the diffusion model can accurately reflect the original distribution, while simultaneously being different from + +each other. Hence, we maintain another auxiliary memory $\mathcal{D} = \{\mathbf{z}_d\}_{d=1}^{N_D}$ for the predicted embeddings of adjacent iterations and design another minimax objective to explicitly enhance the sample diversity as: + +$$ +\mathcal {L} _ {d} = \arg \min _ {\theta} \max _ {d \in [ N _ {D} ]} \sigma (\hat {\mathbf {z}} _ {\theta} (\mathbf {z} _ {t}, \mathbf {c}), \mathbf {z} _ {d}). \tag {4} +$$ + +The diversity term has an opposite optimization target compared with the representativeness term, where the predicted embedding is pushed away from the most similar one stored in the memory bank. Although diversity is essential for an effective surrogate set, too much of it will cause the generated data to lose representativeness. The proposed minimax optimization enhances the diversity in a gentle way, with less influence on the class-related features. + +Combining all the components, we summarize the training process in Algorithm 1. The complete training objective can be formulated as: + +$$ +\mathcal {L} = \mathcal {L} _ {\text {s i m p l e}} + \lambda_ {r} \mathcal {L} _ {r} + \lambda_ {d} \mathcal {L} _ {d}, \tag {5} +$$ + +where $\lambda_r$ and $\lambda_d$ are weighting hyper-parameters. + +# 3. Theoretical Analysis + +Assume that $\mu$ is the real distribution of the latent variables $\mathbf{z}$ associated with the target dataset $\mathcal{T}$ . We rewrite the optimization problem presented in Eq. (5) in a modified form: + +$$ +\begin{array}{l} \min_{\{\theta^{(i)}\}_{i\in [N_{D}]}}\lambda_{d}\max_{i,j = 1,\ldots ,N_{D}}\sigma \left(\hat{\mathbf{z}} (\theta^{(i)}),\hat{\mathbf{z}} (\theta^{(j)})\right) \\ + \sum_ {i = 1} ^ {N _ {D}} \left\{- \lambda_ {r} Q _ {\tilde {q}, w \sim \mu} \left[ \sigma \left(\hat {\mathbf {z}} \left(\theta^ {(i)}\right), w\right) \right] + \| \hat {\mathbf {z}} \left(\theta^ {(i)}\right) - \mathbf {z} _ {0} ^ {(i)} \| ^ {2} \right\}, \tag {6} \\ \end{array} +$$ + +where $Q_{\tilde{q}}[\cdot]$ denotes the quantile function with $\tilde{q}$ as the quantile percentage. Note that here we consider a theoretical idealized variant of our algorithm wherein we perform simultaneous generation of all the embeddings $\{\hat{\mathbf{z}}(\theta^{(i)})\}$ , rather than sample by sample. Hence the objectives turn to the sum of pairwise similarities rather than the form in Eq. (4). And we minimize the negative to aim for maximal representativeness, as in Eq. (5). + +It can be considered as a scalarized solution to a multi-objective optimization problem, wherein multiple criteria are weighed (see, e.g. [13]). This perspective aligns with a Pareto front with trade-offs. It means that one objective decreasing will by necessity result in another increasing. + +However, consider that any solution to the following tri-level optimization problem is also a solution for Eq. (6): + +$$ +\begin{array}{l} \min _ {\{\theta^ {(i)} \} _ {i \in [ N _ {D} ]}} \max _ {i, j = 1,.., N _ {D}} \sigma \left(\hat {\mathbf {z}} (\theta^ {(i)}), \hat {\mathbf {z}} (\theta^ {(j)})\right) \\ \text {s u b j . t o} \quad \left\{\theta^ {(i)} \right\} \in \arg \max \sum_ {i = 1} ^ {N _ {D}} Q _ {\tilde {q}, w \sim \mu} \left[ \sigma (\hat {\mathbf {z}} (\theta^ {(i)}), w) \right] \\ \text {s u b j . t o} \quad \theta^ {(i)} \in \arg \min \| \hat {\mathbf {z}} (\theta) - \mathbf {z} _ {0} ^ {(i)} \| ^ {2}, \forall i \in [ N _ {D} ]. \tag {7} \\ \end{array} +$$ + +If a solution to Eq. (7) is discovered, either incidentally through solving Eq. (6) or by careful tuning of step sizes, the set of minimizers will be sufficiently large at both levels, with no trade-offs involved. However, can we justify the presumption that there exists a meaningful set of potential minimizers? + +Diffusion Process Model One popular framework for the mathematical analysis of diffusion involves analyzing the convergence and asymptotic properties of, appropriately homonymous, diffusion processes. These processes are characterized by the standard stochastic differential equation with a drift and diffusion term. For a time-dependent random variable $Z_{t}$ , + +$$ +d Z _ {t} = V \left(Z _ {t}\right) d t + d W _ {t} \tag {8} +$$ + +where $V$ is a drift function dependent on the current $Z_{t}$ and $dW_{t}$ is a Wiener (Brownian noise) process. This equation serves as an appropriate continuous approximation of generative diffusion, given that Brownian noise is a continuous limit of adding normal random variables. Consequently, we aim for any realization $\mathbf{z} \sim Z_{t}$ to have certain desired properties that reflect generative modeling with high probability. + +The work [43] established a theoretical model utilizing concepts in the control of these diffusions, demonstrating how it can result in sampling from the distribution of a desired data set. In the supplementary material we present a description of their framework and present an argument supporting the well-defined nature of the following problem, indicating that it has non-trivial solutions. + +When sampling optimally from the population dataset, we consider a stochastic control problem wherein $V$ depends also on some chosen control $u(\mathbf{z},t)$ . This control aims to find the most representative samples and, among the possible collection of such samples, to obtain the most diverse one while sampling from the desired dataset $\mu$ . This involves solving: + +$$ +\min _ {u (x, t)} \max _ {i, j = 1,.., N _ {D}} \sigma \left(Z _ {1} ^ {u, (i)}, Z _ {1} ^ {u, (j)}\right) +$$ + +$$ +\begin{array}{l} \mathrm {s u b j . t o} \quad u \in \arg \max \sum_ {i = 1} ^ {N _ {D}} \int_ {0} ^ {1} \mathbb {E} _ {Z _ {t} ^ {(i)}} Q _ {\tilde {q}, w \sim \mu} \left[ \sigma \left(Z _ {t} ^ {(i)}, w\right) \right] d s \\ Z _ {1} \sim \mu , \\ d Z _ {t} ^ {u, (i)} = u \left(Z _ {t} ^ {u, (i)}, t\right) d t + d W _ {t}, t \in [ 0, 1 ]; \\ Z _ {0} = \mathbf {z} _ {0}. \\ \end{array} +$$ + +(9) + +This problem poses a bi-level stochastic control challenge where employing a layered dynamic programming is far from tractable. Additionally, a multi-stage stochastic programming approximation would also be infeasible given the scale of the datasets involved. Instead, we opt for parameterization with a neural network, forego exact sampling, discretize the problem, redefine the criteria to be time independent, and seek to solve an approximate solution for the tri-level optimization problem Eq. (7). + +Table 1. Performance comparison with pre-trained diffusion models and other state-of-the-art methods on ImageWoof. All the results are reproduced by us on the $256\times 256$ resolution. The missing results are due to out-of-memory. The best results are marked as bold. + +
IPC (Ratio)Test ModelRandomK-Center [37]Herding [48]DiT [33]DM [56]IDC-1 [21]GLaD [4]OursFull
10 (0.8%)ConvNet-624.3±1.119.4±0.926.7±0.534.2±1.126.9±1.233.3±1.133.8±0.937.0±1.086.4±0.2
ResNetAP-1029.4±0.822.1±0.132.0±0.334.7±0.530.3±1.239.1±0.532.9±0.939.2±1.387.5±0.5
ResNet-1827.7±0.921.1±0.430.2±1.234.7±0.433.4±0.737.3±0.231.7±0.837.6±0.989.3±1.2
20 (1.6%)ConvNet-629.1±0.721.5±0.829.5±0.336.1±0.829.9±1.035.5±0.8-37.6±0.286.4±0.2
ResNetAP-1032.7±0.425.1±0.734.9±0.141.1±0.835.2±0.643.4±0.3-45.8±0.587.5±0.5
ResNet-1829.7±0.523.6±0.332.2±0.640.5±0.529.8±1.738.6±0.2-42.5±0.689.3±1.2
50 (3.8%)ConvNet-641.3±0.636.5±1.040.3±0.746.5±0.844.4±1.043.9±1.2-53.9±0.686.4±0.2
ResNetAP-1047.2±1.340.6±0.449.1±0.749.3±0.247.1±1.148.3±1.0-56.3±1.087.5±0.5
ResNet-1847.9±1.839.6±1.048.3±1.250.1±0.546.2±0.648.3±0.8-57.1±0.689.3±1.2
70 (5.4%)ConvNet-646.3±0.638.6±0.746.2±0.650.1±1.247.5±0.848.9±0.7-55.7±0.986.4±0.2
ResNetAP-1050.8±0.645.9±1.553.4±1.454.3±0.951.7±0.852.8±1.8-58.3±0.287.5±0.5
ResNet-1852.1±1.044.6±1.149.7±0.851.5±1.051.9±0.851.1±1.7-58.8±0.789.3±1.2
100 (7.7%)ConvNet-652.2±0.445.1±0.554.4±1.153.4±0.355.0±1.353.2±0.9-61.1±0.786.4±0.2
ResNetAP-1059.4±1.054.8±0.261.7±0.958.3±0.856.4±0.856.1±0.9-64.5±0.287.5±0.5
ResNet-1861.5±1.350.4±0.459.3±0.758.9±1.360.2±1.058.3±1.2-65.7±0.489.3±1.2
+ +In the supplementary material we provide a rationale for the meaningfulness of the problem in Eq. (9) based on the model of generative diffusion [43]. Specifically, we argue that the set of controls that leads to the desired final distribution and the set of minimizers, is sufficiently large for a low value of the objective at the top layer. + +# 4. Experiments + +# 4.1. Implementation Details + +For the diffusion model, we adopt pre-trained DiT [33] as the baseline and conduct PEFT with Difffit [50]. $\lambda_r$ and $\lambda_d$ are set as 0.002 and 0.008 for Eq. (5), respectively. The image size for the diffusion fine-tuning and sample generation is set as $256\times 256$ . The fine-tuning mini-batch size is set as 8, and the fine-tuning lasts 8 epochs. The learning rate is set as 1e-3 for an AdamW optimizer. After fine-tuning, the images are generated by 50 denoising steps on a pre-defined number of random noise, according to the IPC setting. All the experiments are conducted on a single RTX 4090 GPU. + +# 4.2. Datasets and Evaluation Metric + +For practical applicability, the experiments are exclusively conducted on full-sized ImageNet [6] subsets in this work. The selected subsets include ImageWoof, ImageNette [15] and the 10-class split adopted in [21, 42], denoted as ImageIDC afterward. ImageWoof is a challenging subset, containing only classes of dog breeds, while ImageNette and ImageIDC contain classes with less similarity, and hence are easier to discriminate. For evaluation, we adopt the same setting as in [21]. The surrogate dataset is trained on different model architectures, with a learning rate of 0.01, + +Table 2. The Maximum Mean Discrepancy (MMD) between the extracted features of surrogate dataset and the original one. + +
IPCDiT [33]Diffit [50]DM [56]IDC-1 [21]Ours
505.45.44.86.74.0
1005.55.34.06.44.3
+ +and a scheduler decaying the learning rate at 2/3 and 5/6 of the whole training iterations. The top-1 accuracy on the original testing set is reported to illustrate the performance. + +# 4.3. Comparison with State-of-the-art Methods + +We compare our method with other state-of-the-art methods across different IPC settings and model architectures. For a fair comparison, the results are all reproduced by us under the same evaluation protocol. ResNet-10 [18] with average pooling is adopted for matching the feature distribution (DM [56], GLaD [4]) and training gradients (IDC-1 [21]). DM is implemented on IDC-1 by only modifying the matching objective from training gradients to feature distribution, such that better performance is achieved. Each experiment is conducted 3 times, with the mean value and standard variance reported. Firstly, we present the validation results on the challenging ImageWoof subset [15] in Tab. 1. + +With the target of distilling surrogate datasets of small IPCs (e.g., 10 and 20), the pixel-level optimization method IDC-1 [21] demonstrates outstanding performance gain over random original images. However, as the IPC increases, the performance gain drastically drops. Especially under the 100-IPC setting, the distilled dataset even performs worse than random original images. This observation + +Table 3. Performance comparison with pre-trained diffusion models and state-of-the-art methods on more ImageNet subsets. The results are obtained on ResNet-10 with average pooling. The best results are marked as **bold**. + +
IPCRandomDiT [33]DM [56]Ours
Nette1054.2±1.659.1±0.760.8±0.662.0±0.2
2063.5±0.564.8±1.266.5±1.166.8±0.4
5076.1±1.173.3±0.976.2±0.476.6±0.2
IDC1048.1±0.854.1±0.452.8±0.553.1±0.2
2052.5±0.958.9±0.258.5±0.459.0±0.4
5068.1±0.764.3±0.669.1±0.869.6±0.2
+ +aligns with the empirical findings in Fig. 2, where pixel-level methods struggle to optimize the expanded parameter space of large IPCs. The embedding-level optimization method GLaD [4] yields good performance under the 10-IPC setting. However, it requires overwhelming GPU resources for larger IPC settings, which is inapplicable for resource-restricted scenarios. It is also notable that on large IPCs, the coreset method Herding [48] surpasses previous DD methods with far less computational cost. + +The pre-trained DiT [33] here serves as the baseline for generative diffusion techniques. Under the 50-IPC setting, DiT outperforms both random original images and IDC-1. However, the insufficiency of representativeness and diversity restricts its performance on smaller and larger IPCs, respectively. In contrast, our proposed minimax diffusion consistently provides superior performance across all the IPCs over both original images and Herding. Besides, the proposed method eliminates the need of specific network architectures for matching training metrics. Consequently, the cross-architecture generalization is significantly improved. Under most IPC settings, the performance gap between ConvNet-6 and ResNetAP-10 is even smaller than that of the original images. It validates the universality of the rich information learned by the minimax fine-tuning process. + +Furthermore, we extensively assess the Maximum Mean Discrepancy (MMD) between the embedded features of the selected/generated surrogate dataset and the original one in Tab. 2. The features are extracted by a ResNet-10 network pre-trained on the full original dataset. Our method achieves the lowest discrepancy by average, where DM [56] directly sets MMD as the optimization target, proving the validity of extra minimax criteria in fitting distributions. + +Moreover, we show the performance comparison on ImageNette [15] and ImageIDC [21] in Tab. 3. The performance trend generally aligns with that on ImageWoof. More specifically, on these two easier subsets, DiT quickly loses the advantage over original images as IPC increases. Conversely, our proposed minimax diffusion method consistently demonstrates state-of-the-art performance. + +Table 4. Performance comparison on ImageNet-1K. + +
IPCSRe2L [53]RDED [41]DiTOurs
1021.3±0.642.0±0.139.6±0.444.3±0.5
5046.8±0.256.5±0.152.9±0.658.6±0.3
+ +![](images/eb4fcacef452f9e119c89cf526ff2f9fe92bfa55af573b4c3d94b37996e7e559.jpg) +Figure 4. With the help of the minimax diffusion, the proposed method significantly enhances the representativeness and diversity of the generated images. Thereby it consistently provides superior performance compared with random selection and baseline diffusion models by a large margin across different IPC settings. + +![](images/a12b7a4869b423baeb051db78d08ae72416aface9a0e686a33358074d414aa48.jpg) + +Experiments on ImageNet-1K. We further conduct experiments on the full ImageNet-1K with the validation protocol of RDED [41] and present the results in Tab. 4. The synthetic images are resized to $224 \times 224$ for evaluation. The significant performance advantage over the compared works validates the scalability of the proposed method. + +# 4.4. Ablation Study + +Component Analysis. We compare the performance with baseline diffusion models to validate the effectiveness of proposed minimax criteria in Fig. 4. The experiments are conducted on ImageWoof and ImageIDC to evaluate the effect on challenging and easy tasks, respectively. Under the IPC of 10 and 20, the raw diffusion models (DiT) generate informative images, with validation performance much higher than randomly selected original samples. However, as the IPC is continuously increased, the performance gap diminishes for ImageWoof, and random original images surpass the DiT-generated ones at the IPC of 100. On ImageIDC the intersection occurs even earlier at the IPC of 50. The main reason is reflected in Fig. 3, where the sample diversity remains limited without external guidance. The naive Difffit fine-tuning adapts the model to specific domains, yet on large IPCs, the over-fitted generative model still yields inferior performance than the original images. + +![](images/4333973511500f8d9b9119d19e8564fce33e79c0b54d97b7940598d07f3635a1.jpg) +Figure 5. Visualization of random original images, images generated by baseline diffusion models (DiT [33] and Difffit [50]) and our proposed method. For each column, the generated images are based on the same random seed. Comparatively, our method significantly enhances the coverage of original data distribution and the diversity of the surrogate dataset. + +Table 5. The ablation study of the proposed minimax scheme. The result are obtained with ResNet-10 on ImageWoof and ImageIDC. m denotes the proposed minimax optimization scheme. + +
\( {\mathcal{L}}_{r} \)\( {\mathcal{L}}_{r} \) w\m\( {\mathcal{L}}_{d} \)\( {\mathcal{L}}_{d} \) w\mImageWoofImageIDC
10-IPC50-IPC10-IPC50-IPC
----35.6±0.951.0±0.953.5±0.266.3±0.2
✓---34.4±1.147.1±0.549.6±0.760.2±1.2
-✓--37.4±0.449.5±1.054.5±1.265.0±0.8
--✓-35.7±0.848.3±0.651.5±0.664.8±0.8
---✓38.7±0.954.9±0.752.2±0.668.4±0.7
✓-✓-38.3±0.554.9±0.453.3±0.566.8±0.5
-✓-✓39.2±1.356.3±1.053.1±0.269.6±0.2
+ +The addition of representativeness constraint to the training process further enhances the effect of distribution fitting. At small IPCs, the generated images contain richer information, yet for larger IPCs, the lack of diversity brings a negative influence. The diversity constraint, in contrast, significantly boosts the information contained in the generated surrogate dataset. Despite the performance advantage of $\mathcal{L}_d$ over $\mathcal{L}_r$ , combining them still brings stable improvement as our full method. Especially on the easier ImageIDC task, grouping these two constraints together contributes to a consistent performance margin over random original images. The experimental results validate that both representativeness and diversity play essential parts in constructing effective surrogate datasets. + +Minimax Scheme. In this work, we propose to enhance the representativeness and diversity each with a minimax objective. We compare the distillation result with or with- + +out the minimax operation in Tab. 5. The first row presents the performance of naive Difffit fine-tuning. Matching the embeddings to the distribution center as in Eq. (2) severely degrades the validation performance across all IPCs. In contrast, the minimax version constraint as in Eq. (3) encourages better coverage, where the performance on small IPCs is improved. The effects of diversity constraint and the full method show similar trends. The superior performance suggests the effectiveness in enhancing the essential properties of the generative diffusion techniques. + +# 4.5. Visualization + +Sample Distribution Visualization. The target of our proposed method is to construct a surrogate dataset with both representativeness and diversity. We visualize the t-SNE distribution of samples generated by our proposed method in Fig. 3. In comparison with random original images and baseline diffusion models, our method demonstrates a more thorough coverage over the entire data distribution while maintaining consistency in sample density. At the original high-density region, the generated images also form a dense sub-cluster, which is not reflected by random sampling. On the other hand, at the original sparse regions, our method exhibits better coverage than baseline diffusion models. By simultaneously enhancing the representativeness and diversity in the generative model, the proposed method manages to significantly improve the validation performance of the generated surrogate dataset. + +Generated Sample Comparison. The proposed method notably enhances the representativeness and diversity of the generated surrogate dataset. We compare the samples generated with the same random noise (for each column) of + +![](images/e9068d1bbb0364dd4adf568a30d710136b29428e2397cf17d941585ee0bb2151.jpg) +Figure 6. Hyper-parameter analysis on (a) the training epochs; (b) the representativeness weight $\lambda_r$ ; (c) the diversity weight $\lambda_d$ ; (d) the memory size $N_M$ . The results are obtained with ResNetAP-10 on ImageWoof. The dashed line indicates the value adopted in this work. + +![](images/7dc98ca70bc837ad08622068ccb37cb43da590746fc6ae5e1decc544f50697b2.jpg) + +![](images/14db85d19c673fcc64aa8565f40e91bd574b85beb356d2d625ce97bc30380486.jpg) + +![](images/4c03a4a19a2973f631050f9ce5731e8da4d0ee1d105d7c548928f6b181327eab.jpg) + +different generative methods in Fig. 5 to explicitly demonstrate the improved properties. + +The images generated by baseline DiT exhibit a realistic high-quality appearance. However, the images tend to share similar poses and only present the most prominent features of the objects. In the golden retriever case, the generated images mostly present the head part, while for the churches the exterior appearance. Difffit fine-tuning further fits the model to the distribution, but in most cases, the differences only lie in small details. Comparatively, the proposed minimax criteria significantly enhance both the representativeness and diversity of the generated images. On the one hand, there occurs more class-related content in the generated images. The golden retriever images include more body parts and the church images encompass the interior layout. The minimax optimization leads to better coverage over the entire original distribution, with more related features encapsulated. On the other hand, the diversity is significantly enhanced, including variations in pose, background, and appearance styles. In such a way the surrogate dataset better represents the original large-scale one, leading to superior validation performance. + +Training Curve Visualization. We visualize the accuracy curve during the training process in Fig. 6a. The validation performance is rapidly improved as the fine-tuning process starts. After four epochs, the model tends to converge and reaches the highest performance at the 8th epoch. Further extending the training epochs injects excessive diversity into the model, leading to performance degradation. We demonstrate the influence of training epochs on the generated images in supplementary material. + +# 4.6. Parameter Analysis + +Objective Weight $\lambda_r\lambda_d$ .We show the influence of representativeness weight $\lambda_r$ and diversity weight $\lambda_{d}$ in Fig. 6b and Fig. 6c, respectively. The $\lambda_r$ variation only produces negligible performance fluctuation on small IPCs, while on large IPCs the performance is also relatively stable. For + +$\lambda_{d}$ , at a proper variation range, the performance is stable. However, continuously increasing the diversity of the generated dataset leads to a lack of representativeness, which results in a negative impact. The negative impact of overdiversity can also validated by the poor performance of K-Center in Tab. 1. A uniform performance decrease is observed as $\lambda_{d}$ reaches 0.03. Based on the performance of 100 IPC, we set $\lambda_{r}$ as 0.002 and $\lambda_{d}$ as 0.008. + +Memory Size $N_{M}$ . The memory size $N_{M}$ influences the number of samples involved in the objective calculation. We investigate its influence in Fig. 6d. When the memory is extremely small ( $N_{M} = 16$ ), the provided supervision is also limited, yet the performance is already higher than naive fine-tuning. As the memory size is increased in a proper range, the model yields stable performance improvement. It is notable that with a larger memory, the performance under the IPC of 10 is better. It can be explained by that a larger memory contains more representative information. Out of the consideration of performance as well as storage burden, we select the memory of 64 in the other experiments. + +# 5. Conclusion + +In this work, we propose a novel dataset distillation method based on generative diffusion techniques. Through extra minimax criteria, the proposed method significantly enhances the representativeness and diversity of the generated surrogate dataset. With much less computational time consumption, the proposed method achieves state-of-the-art validation performance on challenging ImageNet subsets. It reduces the resource dependency of previous dataset distillation methods and opens up new possibilities for more practical applications for distilling personalized data. + +Limitations and Future Works. In addition to object classification, we will explore the possibility of incorporating generative techniques for more specific data domains. + +Acknowledgement. We would like to specially thank Olga Russakovsky for invaluable advice on the manuscript. + +# References + +[1] Shekoofeh Azizi, Simon Kornblith, Chitwan Sahara, Mohammad Norouzi, and David J Fleet. Synthetic data from diffusion models improves imagenet classification. arXiv preprint arXiv:2304.08466, 2023. 2 +[2] Victor Besnier, Himalaya Jain, Andrei Bursuc, Matthieu Cord, and Patrick Pérez. This Dataset Does Not Exist: Training Models from Generated Images. In ICASSP, pages 1-5, 2020. 2 +[3] George Cazenavette, Tongzhou Wang, Antonio Torralba, Alexei A Efros, and Jun-Yan Zhu. Dataset distillation by matching training trajectories. In CVPR, pages 4750-4759, 2022. 1 +[4] George Cazenavette, Tongzhou Wang, Antonio Torralba, Alexei A Efros, and Jun-Yan Zhu. Generalizing dataset distillation via deep generative prior. In CVPR, pages 3739-3748, 2023. 1, 5, 6 +[5] Justin Cui, Ruochen Wang, Si Si, and Cho-Jui Hsieh. Dc-bench: Dataset condensation benchmark. NeurIPS, 35:810-822, 2022. 1 +[6] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. ImageNet: A large-scale hierarchical image database. In CVPR, pages 248-255, 2009. 1, 5 +[7] Zhiwei Deng and Olga Russakovsky. Remember the Past: Distilling Datasets into Addressable Memories for Neural Networks. In NeurIPS, 2022. 1 +[8] Prafulla Dhariwal and Alexander Nichol. Diffusion Models Beat GANs on Image Synthesis. In NeurIPS, pages 8780-8794, 2021. 3, 2 +[9] Thinh T Doan. Nonlinear two-time-scale stochastic approximation convergence and finite-time performance. IEEE Transactions on Automatic Control, 2022. 1 +[10] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale. In ICLR, 2022. 1 +[11] Jiawei Du, Yidi Jiang, Vincent Y.F. Tan, Joey Tianyi Zhou, and Haizhou Li. Minimizing the Accumulated Trajectory Error to Improve Dataset Distillation. In CVPR, pages 3749-3758, 2023. 1 +[12] Lisa Dunlap, Alyssa Umino, Han Zhang, Jiezhi Yang, Joseph E Gonzalez, and Trevor Darrell. Diversify your vision datasets with automatic diffusion-based augmentation. arXiv preprint arXiv:2305.16289, 2023. 2 +[13] Gabriele Eichfelder. Scalarizations for adaptively solving multi-objective optimization problems. Com- + +putational Optimization and Applications, 44:249-273, 2009. 4 +[14] Ronen Eldan and James R Lee. Regularization under diffusion and anticoncentration of the information content. Duke Mathematical Journal, 167(5):969-993, 2018. 1 +[15] Fastai. Fastai/imagenette: A smaller subset of 10 easily classified classes from imagenet, and a little more french. 1, 5, 6 +[16] Jianyang Gu, Kai Wang, Wei Jiang, and Yang You. Summarizing stream data for memory-restricted online continual learning. arXiv preprint arXiv:2305.16645, 2023. 1 +[17] Swaminathan Gurumurthy, Ravi Kiran Sarvadevabhatla, and R. Venkatesh Babu. DeLiGAN: Generative Adversarial Networks for Diverse and Limited Data. In CVPR, pages 4941-4949, 2017. 2 +[18] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, pages 770-778, 2016. 1, 5, 2 +[19] Ruifei He, Shuyang Sun, Xin Yu, Chuhui Xue, Wenqing Zhang, Philip Torr, Song Bai, and Xiaojuan Qi. Is Synthetic Data from Generative Models Ready for Image Recognition? In ICLR, 2022. 2 +[20] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising Diffusion Probabilistic Models. In NeurIPS, pages 6840-6851, 2020. 2 +[21] Jang-Hyun Kim, Jinuk Kim, Seong Joon Oh, Sang-doo Yun, Hwanjun Song, Joonhyun Jeong, Jung-Woo Ha, and Hyun Oh Song. Dataset condensation via efficient synthetic-data parameterization. In ICML, pages 11102-11118, 2022. 1, 5, 6, 2, 3 +[22] Diederik Kingma, Tim Salimans, Ben Poole, and Jonathan Ho. Variational Diffusion Models. In NeurIPS, pages 21696-21707, 2021. 2 +[23] Saehyung Lee, Sanghyuk Chun, Sangwon Jung, Sang-doo Yun, and Sungroh Yoon. Dataset condensation with contrastive signals. In ICML, pages 12352-12364, 2022. 1 +[24] Joseph Lehec. Representation formula for the entropy and functional inequalities. In Annales de l'IHP Probabilités et statistiques, pages 885-899, 2013. 1 +[25] Ping Liu, Xin Yu, and Joey Tianyi Zhou. Meta Knowledge Condensation for Federated Learning. In ICLR, 2022. 1 +[26] Yanqing Liu, Jianyang Gu, Kai Wang, Zheng Zhu, Wei Jiang, and Yang You. Dream: Efficient dataset distillation by representative matching. In ICCV, 2023. 1, 2 +[27] Noel Loo, Ramin Hasani, Alexander Amini, and Daniela Rus. Efficient dataset distillation using random feature approximation. NeurIPS, 35:13877-13891, 2022. 1 + +[28] Noel Loo, Ramin Hasani, Mathias Lechner, and Daniela Rus. Dataset distillation with convexified implicit gradients. arXiv preprint arXiv:2302.06755, 2023.1 +[29] Qi Mao, Hsin-Ying Lee, Hung-Yu Tseng, Siwei Ma, and Ming-Hsuan Yang. Mode Seeking Generative Adversarial Networks for Diverse Image Synthesis. In CVPR, pages 1429-1437, 2019. 2 +[30] Timothy Nguyen, Zhourong Chen, and Jaehoon Lee. Dataset meta-learning from kernel ridge-regression. In ICLR, 2021. 1 +[31] Timothy Nguyen, Roman Novak, Lechao Xiao, and Jaehoon Lee. Dataset distillation with infinitely wide convolutional networks. NeurIPS, 34:5186-5198, 2021. 1 +[32] Alexander Quinn Nichol and Prafulla Dhariwal. Improved Denoising Diffusion Probabilistic Models. In ICML, pages 8162-8171, 2021. 2 +[33] William Peebles and Saining Xie. Scalable diffusion models with transformers. In ICCV, pages 4195-4205, 2023. 3, 5, 6, 7 +[34] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In CVPR, pages 22500-22510, 2023. 3 +[35] Ahmad Sajedi, Samir Khaki, Ehsan Amjadian, Lucy Z Liu, Yuri A Lawryshyn, and Konstantinos N Plataniotis. DataDAM: Efficient Dataset Distillation with Attention Matching. In ICCV, 2023. 1 +[36] Mert Bulent Sariyildiz, Karteek Alahari, Diane Larlus, and Yannis Kalantidis. Fake it Till You Make it: Learning Transferable Representations from Synthetic ImageNet Clones. In CVPR, pages 8011-8021, 2023. 2 +[37] Ozan Sener and Silvio Savarese. Active learning for convolutional neural networks: A core-set approach. In International Conference on Learning Representations, 2018. 5 +[38] Allahkaram Shafiei, Vyacheslav Kungurtsev, and Jakub Marecek. Trilevel and multilevel optimization using monotone operator theory. arXiv preprint arXiv:2105.09407, 2021. 1 +[39] Jordan Shipard, Arnold Wiliem, Kien Nguyen Thanh, Wei Xiang, and Clinton Fookes. Diversity Is Definitely Needed: Improving Model-Agnostic Zero-Shot Classification via Stable Diffusion. In CVPR, pages 769-778, 2023. 2 +[40] Felipe Petroski Such, Aditya Rawal, Joel Lehman, Kenneth Stanley, and Jeffrey Clune. Generative Teaching Networks: Accelerating Neural Architecture Search by Learning to Generate Synthetic Training Data. In ICML, pages 9206-9216, 2020. 1 + +[41] Peng Sun, Bei Shi, Daiwei Yu, and Tao Lin. On the diversity and realism of distilled dataset: An efficient dataset distillation paradigm. In CVPR, 2024. 6 +[42] Yonglong Tian, Dilip Krishnan, and Phillip Isola. Contrastive multiview coding. In ECCV, pages 776-794, 2020. 5 +[43] Belinda Tzen and Maxim Raginsky. Theoretical guarantees for sampling and inference in generative models with latent diffusions. In $COLT$ , pages 3084-3114, 2019. 4, 5, 1 +[44] Saeed Vahidian, Mingyu Wang, Jianyang Gu, Vyacheslav Kungurtsev, Wei Jiang, and Yiran Chen. Group distributionally robust dataset distillation with risk minimization. arXiv preprint arXiv:2402.04676, 2024.1 +[45] Kai Wang, Bo Zhao, Xiangyu Peng, Zheng Zhu, Shuo Yang, Shuo Wang, Guan Huang, Hakan Bilen, Xin chao Wang, and Yang You. Cafe: Learning to condense dataset by aligning features. In CVPR, pages 12196-12205, 2022. 1 +[46] Kai Wang, Jianyang Gu, Daquan Zhou, Zheng Zhu, Wei Jiang, and Yang You. Dim: Distilling dataset into generative model. arXiv preprint arXiv:2303.04707, 2023.1 +[47] Tongzhou Wang, Jun-Yan Zhu, Antonio Torralba, and Alexei A Efros. Dataset distillation. arXiv preprint arXiv:1811.10959, 2018. 1, 2 +[48] Max Welling. Herding dynamical weights to learn. In ICML, pages 1121-1128, 2009. 5, 6, 3 +[49] Xindi Wu, Zhiwei Deng, and Olga Russakovsky. Multimodal dataset distillation for image-text retrieval. arXiv preprint arXiv:2308.07545, 2023. 1 +[50] Enze Xie, Lewei Yao, Han Shi, Zhili Liu, Daquan Zhou, Zhaoqiang Liu, Jiawei Li, and Zhenguo Li. Difffit: Unlocking transferability of large diffusion models via simple parameter-efficient fine-tuning. arXiv preprint arXiv:2304.06648, 2023. 3, 5, 7, 2 +[51] Yuanhao Xiong, Ruochen Wang, Minhao Cheng, Felix Yu, and Cho-Jui Hsieh. FedDM: Iterative Distribution Matching for Communication-Efficient Federated Learning. In CVPR, pages 16323-16332, 2023. 1 +[52] Dingdong Yang, Seunghoon Hong, Yunseok Jang, Tianchen Zhao, and Honglak Lee. Diversity-Sensitive Conditional Generative Adversarial Networks. In ICLR, 2018. 2 +[53] Zeyuan Yin, Eric Xing, and Zhiqiang Shen. Squeeze, recover and relabel: Dataset condensation atImagenet scale from a new perspective. In NeurIPS, 2023. 6 +[54] Bo Zhao and Hakan Bilen. Dataset condensation with differentiable siamese augmentation. In ICML, pages 12674-12685, 2021. 1, 2 + +[55] Bo Zhao and Hakan Bilen. Synthesizing informative training samples with gan. arXiv preprint arXiv:2204.07513, 2022. 1 +[56] Bo Zhao and Hakan Bilen. Dataset condensation with distribution matching. In WACV, pages 6514-6523, 2023. 2, 5, 6, 1 +[57] Bo Zhao, Konda Reddy Mopuri, and Hakan Bilen. Dataset condensation with gradient matching. In ICLR, 2021. 1, 2 +[58] Ganlong Zhao, Guanbin Li, Yipeng Qin, and Yizhou Yu. Improved Distribution Matching for Dataset Condensation. In CVPR, pages 7856-7865, 2023. 1 +[59] Yongchao Zhou, Ehsan Nezhadarya, and Jimmy Ba. Dataset distillation using neural feature regression. NeurIPS, 35:9813-9827, 2022. 1 +[60] Yongchao Zhou, Hshmat Sahak, and Jimmy Ba. Training on thin air: Improve image classification with generated data. arXiv preprint arXiv:2305.15316, 2023.2 \ No newline at end of file diff --git a/2024/Efficient Dataset Distillation via Minimax Diffusion/images.zip b/2024/Efficient Dataset Distillation via Minimax Diffusion/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..ed8ca44f24e439cd8b2a70ea5383fe0a79670332 --- /dev/null +++ b/2024/Efficient Dataset Distillation via Minimax Diffusion/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:df23f0f7ed62a2fa8ea1ee7c42bba5a9cfc170633d8d541ef42b7370b9a4c117 +size 726938 diff --git a/2024/Efficient Dataset Distillation via Minimax Diffusion/layout.json b/2024/Efficient Dataset Distillation via Minimax Diffusion/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..1d03680bd395a462b1075cce295721f2047a4868 --- /dev/null +++ b/2024/Efficient Dataset Distillation via Minimax Diffusion/layout.json @@ -0,0 +1,9024 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 138, + 103, + 455, + 119 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 103, + 455, + 119 + ], + "spans": [ + { + "bbox": [ + 138, + 103, + 455, + 119 + ], + "type": "text", + "content": "Efficient Dataset Distillation via Minimax Diffusion" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 151, + 142, + 441, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 142, + 441, + 157 + ], + "spans": [ + { + "bbox": [ + 151, + 142, + 441, + 157 + ], + "type": "text", + "content": "Jianyang Gu1 Saeed Vahidian2 Vyacheslav Kungurtsev3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 154, + 157, + 427, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 154, + 157, + 427, + 171 + ], + "spans": [ + { + "bbox": [ + 154, + 157, + 427, + 171 + ], + "type": "text", + "content": "Haonan Wang" + }, + { + "bbox": [ + 154, + 157, + 427, + 171 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 154, + 157, + 427, + 171 + ], + "type": "text", + "content": " Wei Jiang" + }, + { + "bbox": [ + 154, + 157, + 427, + 171 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 154, + 157, + 427, + 171 + ], + "type": "text", + "content": " Yang You" + }, + { + "bbox": [ + 154, + 157, + 427, + 171 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 154, + 157, + 427, + 171 + ], + "type": "text", + "content": " Yiran Chen" + }, + { + "bbox": [ + 154, + 157, + 427, + 171 + ], + "type": "inline_equation", + "content": "^{2}" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 123, + 171, + 476, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 123, + 171, + 476, + 186 + ], + "spans": [ + { + "bbox": [ + 123, + 171, + 476, + 186 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 123, + 171, + 476, + 186 + ], + "type": "text", + "content": "Zhejiang University " + }, + { + "bbox": [ + 123, + 171, + 476, + 186 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 123, + 171, + 476, + 186 + ], + "type": "text", + "content": "Duke University " + }, + { + "bbox": [ + 123, + 171, + 476, + 186 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 123, + 171, + 476, + 186 + ], + "type": "text", + "content": "Czech Technical University" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 214, + 186, + 380, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 214, + 186, + 380, + 200 + ], + "spans": [ + { + "bbox": [ + 214, + 186, + 380, + 200 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 214, + 186, + 380, + 200 + ], + "type": "text", + "content": "National University of Singapore" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 195, + 201, + 395, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 195, + 201, + 395, + 213 + ], + "spans": [ + { + "bbox": [ + 195, + 201, + 395, + 213 + ], + "type": "text", + "content": "{gu_jianyang, jiangwei_zju}@zju.edu.cn" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 143, + 241, + 191, + 253 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 241, + 191, + 253 + ], + "spans": [ + { + "bbox": [ + 143, + 241, + 191, + 253 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 266, + 289, + 565 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 266, + 289, + 565 + ], + "spans": [ + { + "bbox": [ + 47, + 266, + 289, + 565 + ], + "type": "text", + "content": "Dataset distillation reduces the storage and computational consumption of training a network by generating a small surrogate dataset that encapsulates rich information of the original large-scale one. However, previous distillation methods heavily rely on the sample-wise iterative optimization scheme. As the images-per-class (IPC) setting or image resolution grows larger, the necessary computation will demand overwhelming time and resources. In this work, we intend to incorporate generative diffusion techniques for computing the surrogate dataset. Observing that key factors for constructing an effective surrogate dataset are representativeness and diversity, we design additional minimax criteria in the generative training to enhance these facets for the generated images of diffusion models. We present a theoretical model of the process as hierarchical diffusion control demonstrating the flexibility of the diffusion process to target these criteria without jeopardizing the faithfulness of the sample to the desired distribution. The proposed method achieves state-of-the-art validation performance while demanding much less computational resources. Under the 100-IPC setting on ImageWoof, our method requires less than one-twentieth the distillation time of previous methods, yet yields even better performance. Source code and generated data are available in https://github.com/vimar-gu/MinimaxDiffusion." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 587, + 128, + 599 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 587, + 128, + 599 + ], + "spans": [ + { + "bbox": [ + 47, + 587, + 128, + 599 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 607, + 287, + 668 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 607, + 287, + 668 + ], + "spans": [ + { + "bbox": [ + 46, + 607, + 287, + 668 + ], + "type": "text", + "content": "Data, as a necessary resource for deep learning, has concurrently promoted algorithmic advancements while imposing challenges on researchers due to heavy demands on storage and computational resources [6, 10, 18, 47]. Confronted with the conflict between the requirement for high-precision" + } + ] + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 315, + 242, + 536, + 399 + ], + "blocks": [ + { + "bbox": [ + 315, + 242, + 536, + 399 + ], + "lines": [ + { + "bbox": [ + 315, + 242, + 536, + 399 + ], + "spans": [ + { + "bbox": [ + 315, + 242, + 536, + 399 + ], + "type": "image", + "image_path": "552d665e9aa78a478bcd3eadac76eca61fc7600fb0dcb58363cc269e51885137.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 409, + 547, + 487 + ], + "lines": [ + { + "bbox": [ + 305, + 409, + 547, + 487 + ], + "spans": [ + { + "bbox": [ + 305, + 409, + 547, + 487 + ], + "type": "text", + "content": "Figure 1. The validation accuracy and distillation time of different methods on ImageWoof [15], with a number following each method denoting the Image-Per-Class (IPC) setting. Previous methods are restricted by the heavier running time and memory consumption as IPC grows larger. In comparison, our proposed method notably reduces the demanding computational resources and also achieves state-of-the-art validation performance." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 509, + 547, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 509, + 547, + 569 + ], + "spans": [ + { + "bbox": [ + 304, + 509, + 547, + 569 + ], + "type": "text", + "content": "models and overwhelming resource demands, dataset distillation is proposed to condense the rich information of a large-scale dataset into a small surrogate one [5, 21, 47, 57]. Such a surrogate dataset is expected to achieve training performance comparable to that attained with the original one." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 570, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 547, + 713 + ], + "type": "text", + "content": "Previous dataset distillation methods mostly engage in iterative optimization on fixed-number samples at the pixel level [21, 26, 30, 31, 44, 54, 57] or embedding level [4, 55]. However, the sample-wise iterative optimization scheme suffers from problems of two perspectives. (1) The parameter space of optimization is positively correlated with the size of the target surrogate dataset and the image resolution [3, 57]. Consequently, substantial time and computational resources are required for distilling larger datasets. As shown in Fig. 1, IDC-1 [21] takes over 90 hours to distill a 100-image-per-class (IPC) set from ImageWoof [15], while training on ImageWoof itself only requires a matter of" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 674, + 135, + 684 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 674, + 135, + 684 + ], + "spans": [ + { + "bbox": [ + 58, + 674, + 135, + 684 + ], + "type": "text", + "content": "*Corresponding author" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 47, + 684, + 287, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 684, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 47, + 684, + 287, + 712 + ], + "type": "text", + "content": "This work has received funding from the European Union's Horizon Europe research and innovation program under grant agreement No. 101084642." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 293, + 732, + 317, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 732, + 317, + 742 + ], + "spans": [ + { + "bbox": [ + 293, + 732, + 317, + 742 + ], + "type": "text", + "content": "15793" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 51, + 70, + 282, + 206 + ], + "blocks": [ + { + "bbox": [ + 51, + 70, + 282, + 206 + ], + "lines": [ + { + "bbox": [ + 51, + 70, + 282, + 206 + ], + "spans": [ + { + "bbox": [ + 51, + 70, + 282, + 206 + ], + "type": "image", + "image_path": "d1a04696a1ff47117b25ec1488c8a7523afe466bfb34236d84cdbc1ed99caf80.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 216, + 287, + 271 + ], + "lines": [ + { + "bbox": [ + 46, + 216, + 287, + 271 + ], + "spans": [ + { + "bbox": [ + 46, + 216, + 287, + 271 + ], + "type": "text", + "content": "Figure 2. Sample images distilled by the pixel-level sample-wise optimization method DM [56] on ImageWoof. As the parameter space increases along with the Image-Per-Class (IPC) setting, with the same initialization, the appearance disparity between original and distilled images is smaller." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 293, + 286, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 293, + 286, + 449 + ], + "spans": [ + { + "bbox": [ + 46, + 293, + 286, + 449 + ], + "type": "text", + "content": "hours. (2) The expanded parameter space also increases the optimization complexity. As shown in Fig. 2, while distillation yields significant information condensation under small IPC settings, the pixel modification diminishes when distilling larger-IPC datasets. The reduced disparity also leads to smaller performance gain compared with original images, with instances where the distilled set even performs worse. Especially when distilling data of fine-grained classes, the sample-wise optimization scheme fails to provide adequate discriminative information. These constraints severely hinder individual researchers from distilling personalized data. A more practical training scheme is urgently needed to facilitate the broader application of dataset distillation." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 449, + 286, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 449, + 286, + 664 + ], + "spans": [ + { + "bbox": [ + 46, + 449, + 286, + 664 + ], + "type": "text", + "content": "In this work, we explore the possibility of incorporating generative diffusion techniques [20, 22, 32] to efficiently compute effective surrogate datasets. We first conduct empirical analysis on the suitability of data generated by raw diffusion models for training networks. Based on the observations, we conclude that constructing an effective surrogate dataset hinges on two key factors: representativeness and diversity. Accordingly, we design extra minimax criteria for the generative training to enhance the capability of generating more effective surrogate datasets without explicit prompt designs. The minimax criteria involve two aspects: enforcing the generated sample to be close to the farthest real sample, while being far away from the most similar generated one. We provide theoretical analysis to support that the proposed minimax scheme aims to solve a well defined problem with all the criteria, including the generative accuracy and the minimax criteria, can be targeted simultaneously without detriment to the others." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 665, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 665, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 665, + 286, + 713 + ], + "type": "text", + "content": "Compared with the astronomical training time consumption of the sample-wise iterative optimization schemes, the proposed method takes less than 1 hour to distill a 100-IPC surrogate dataset for a 10-class ImageNet subset, including" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 310, + 72, + 422, + 169 + ], + "blocks": [ + { + "bbox": [ + 310, + 72, + 422, + 169 + ], + "lines": [ + { + "bbox": [ + 310, + 72, + 422, + 169 + ], + "spans": [ + { + "bbox": [ + 310, + 72, + 422, + 169 + ], + "type": "image", + "image_path": "d0d68cd486e76069a23a5fac587f4be503f4f213d019b2eacbb53e1e11581cc9.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 309, + 170, + 421, + 266 + ], + "blocks": [ + { + "bbox": [ + 309, + 170, + 421, + 266 + ], + "lines": [ + { + "bbox": [ + 309, + 170, + 421, + 266 + ], + "spans": [ + { + "bbox": [ + 309, + 170, + 421, + 266 + ], + "type": "image", + "image_path": "92b58d4e47e34ee50d933a4ead78330b9d2721a6dd86e8615878255391c2d23a.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 277, + 545, + 311 + ], + "lines": [ + { + "bbox": [ + 305, + 277, + 545, + 311 + ], + "spans": [ + { + "bbox": [ + 305, + 277, + 545, + 311 + ], + "type": "text", + "content": "Figure 3. The feature distribution comparison of different image generation methods with the original set. The validation performance of each surrogate set is listed in the upper-right corner." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 426, + 72, + 539, + 169 + ], + "blocks": [ + { + "bbox": [ + 426, + 72, + 539, + 169 + ], + "lines": [ + { + "bbox": [ + 426, + 72, + 539, + 169 + ], + "spans": [ + { + "bbox": [ + 426, + 72, + 539, + 169 + ], + "type": "image", + "image_path": "16b215162ea890c298ab7256ef36a046addb177a11233a288ec15eef4ba2983d.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 426, + 170, + 539, + 266 + ], + "blocks": [ + { + "bbox": [ + 426, + 170, + 539, + 266 + ], + "lines": [ + { + "bbox": [ + 426, + 170, + 539, + 266 + ], + "spans": [ + { + "bbox": [ + 426, + 170, + 539, + 266 + ], + "type": "image", + "image_path": "1265cd04ab1e8d2ea1b2eecd3819ac9f56eacc6a98787ac2b8810a85d79ff094.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 334, + 545, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 334, + 545, + 441 + ], + "spans": [ + { + "bbox": [ + 304, + 334, + 545, + 441 + ], + "type": "text", + "content": "the fine-tuning and image generation processes. Remarkably, the GPU consumption remains consistent across all IPC settings. Furthermore, the distilled surrogate dataset attains superior validation performance compared with other state-of-the-art methods. Especially on the challenging fine-grained ImageWoof subset, the proposed method outperforms the second-best DD method by " + }, + { + "bbox": [ + 304, + 334, + 545, + 441 + ], + "type": "inline_equation", + "content": "5.5\\%" + }, + { + "bbox": [ + 304, + 334, + 545, + 441 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 334, + 545, + 441 + ], + "type": "inline_equation", + "content": "8.1\\%" + }, + { + "bbox": [ + 304, + 334, + 545, + 441 + ], + "type": "text", + "content": " under the IPC settings of 70 and 100, respectively. The source code is provided in the supplementary material." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 317, + 443, + 529, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 443, + 529, + 453 + ], + "spans": [ + { + "bbox": [ + 317, + 443, + 529, + 453 + ], + "type": "text", + "content": "The contributions of this work are summarized into:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 455, + 545, + 609 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 306, + 455, + 545, + 490 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 455, + 545, + 490 + ], + "spans": [ + { + "bbox": [ + 306, + 455, + 545, + 490 + ], + "type": "text", + "content": "- We analyze the data generated by diffusion models, and emphasize the importance of representativeness and diversity for constructing effective surrogate datasets." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 491, + 544, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 491, + 544, + 525 + ], + "spans": [ + { + "bbox": [ + 306, + 491, + 544, + 525 + ], + "type": "text", + "content": "- We propose a novel dataset distillation scheme based on extra minimax criteria for diffusion models targeting the representativeness and diversity of generated data." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 527, + 545, + 562 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 527, + 545, + 562 + ], + "spans": [ + { + "bbox": [ + 306, + 527, + 545, + 562 + ], + "type": "text", + "content": "- We theoretically justify the proposed minimax criteria as enforceable without trade-offs in the generation quality of the individual data points." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 563, + 544, + 609 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 563, + 544, + 609 + ], + "spans": [ + { + "bbox": [ + 306, + 563, + 544, + 609 + ], + "type": "text", + "content": "- We conduct extensive experiments to validate that our proposed method achieves state-of-the-art performance while demanding significantly reduced training time in comparison to previous dataset distillation methods." + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 624, + 361, + 636 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 624, + 361, + 636 + ], + "spans": [ + { + "bbox": [ + 306, + 624, + 361, + 636 + ], + "type": "text", + "content": "2. Method" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 306, + 645, + 417, + 656 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 645, + 417, + 656 + ], + "spans": [ + { + "bbox": [ + 306, + 645, + 417, + 656 + ], + "type": "text", + "content": "2.1. Problem Definition" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 305, + 665, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 665, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 665, + 545, + 713 + ], + "type": "text", + "content": "The general purpose of dataset distillation is to generate a small surrogate dataset " + }, + { + "bbox": [ + 305, + 665, + 545, + 713 + ], + "type": "inline_equation", + "content": "S = \\{(\\mathbf{x}_i,y_i)\\}_{i = 1}^{N_S}" + }, + { + "bbox": [ + 305, + 665, + 545, + 713 + ], + "type": "text", + "content": " from a largescale one " + }, + { + "bbox": [ + 305, + 665, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{T} = \\{(\\mathbf{x}_i,y_i)\\}_{i = 1}^{N_T}" + }, + { + "bbox": [ + 305, + 665, + 545, + 713 + ], + "type": "text", + "content": " [47, 57]. Here each " + }, + { + "bbox": [ + 305, + 665, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_i" + }, + { + "bbox": [ + 305, + 665, + 545, + 713 + ], + "type": "text", + "content": " denotes an image with a corresponding class label " + }, + { + "bbox": [ + 305, + 665, + 545, + 713 + ], + "type": "inline_equation", + "content": "y_{i}" + }, + { + "bbox": [ + 305, + 665, + 545, + 713 + ], + "type": "text", + "content": ", and" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 732, + 317, + 741 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 732, + 317, + 741 + ], + "spans": [ + { + "bbox": [ + 294, + 732, + 317, + 741 + ], + "type": "text", + "content": "15794" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 287, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 287, + 133 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 287, + 133 + ], + "type": "inline_equation", + "content": "N_{S} \\ll N_{T}" + }, + { + "bbox": [ + 47, + 72, + 287, + 133 + ], + "type": "text", + "content": ". The surrogate dataset " + }, + { + "bbox": [ + 47, + 72, + 287, + 133 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 47, + 72, + 287, + 133 + ], + "type": "text", + "content": " is expected to encapsulate substantial information from the original " + }, + { + "bbox": [ + 47, + 72, + 287, + 133 + ], + "type": "inline_equation", + "content": "\\mathcal{T}" + }, + { + "bbox": [ + 47, + 72, + 287, + 133 + ], + "type": "text", + "content": ", such that training a model on " + }, + { + "bbox": [ + 47, + 72, + 287, + 133 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 47, + 72, + 287, + 133 + ], + "type": "text", + "content": " achieves performance comparable with that on " + }, + { + "bbox": [ + 47, + 72, + 287, + 133 + ], + "type": "inline_equation", + "content": "\\mathcal{T}" + }, + { + "bbox": [ + 47, + 72, + 287, + 133 + ], + "type": "text", + "content": ". After distilling, we train network models on " + }, + { + "bbox": [ + 47, + 72, + 287, + 133 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 47, + 72, + 287, + 133 + ], + "type": "text", + "content": " and validate the performance on the original test set." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 138, + 185, + 150 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 138, + 185, + 150 + ], + "spans": [ + { + "bbox": [ + 47, + 138, + 185, + 150 + ], + "type": "text", + "content": "2.2. Diffusion for Distillation" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 157, + 288, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 157, + 288, + 312 + ], + "spans": [ + { + "bbox": [ + 46, + 157, + 288, + 312 + ], + "type": "text", + "content": "Diffusion models learn a dataset distribution by gradually adding Gaussian noise to images and reversing back. Taking the latent diffusion model (LDM) as an example, given a training image " + }, + { + "bbox": [ + 46, + 157, + 288, + 312 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 46, + 157, + 288, + 312 + ], + "type": "text", + "content": ", the training process is separated into two parts. An encoder " + }, + { + "bbox": [ + 46, + 157, + 288, + 312 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 46, + 157, + 288, + 312 + ], + "type": "text", + "content": " transforms the image into the latent space " + }, + { + "bbox": [ + 46, + 157, + 288, + 312 + ], + "type": "inline_equation", + "content": "\\mathbf{z} = E(\\mathbf{x})" + }, + { + "bbox": [ + 46, + 157, + 288, + 312 + ], + "type": "text", + "content": " and a decoder " + }, + { + "bbox": [ + 46, + 157, + 288, + 312 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 46, + 157, + 288, + 312 + ], + "type": "text", + "content": " reconstructs a latent code back to the image space " + }, + { + "bbox": [ + 46, + 157, + 288, + 312 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{x}} = D(\\mathbf{z})" + }, + { + "bbox": [ + 46, + 157, + 288, + 312 + ], + "type": "text", + "content": ". The forward noisng process gradually adds noise " + }, + { + "bbox": [ + 46, + 157, + 288, + 312 + ], + "type": "inline_equation", + "content": "\\epsilon \\sim \\mathcal{N}(\\mathbf{0},\\mathbf{I})" + }, + { + "bbox": [ + 46, + 157, + 288, + 312 + ], + "type": "text", + "content": " to the original latent code " + }, + { + "bbox": [ + 46, + 157, + 288, + 312 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_0" + }, + { + "bbox": [ + 46, + 157, + 288, + 312 + ], + "type": "text", + "content": ": " + }, + { + "bbox": [ + 46, + 157, + 288, + 312 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_t = \\sqrt{\\bar{\\alpha}_t}\\mathbf{z}_0 + \\sqrt{1 - \\bar{\\alpha}_t\\epsilon}" + }, + { + "bbox": [ + 46, + 157, + 288, + 312 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 157, + 288, + 312 + ], + "type": "inline_equation", + "content": "\\bar{\\alpha}_t" + }, + { + "bbox": [ + 46, + 157, + 288, + 312 + ], + "type": "text", + "content": " is a hyper-parameter. Provided with a conditioning vector " + }, + { + "bbox": [ + 46, + 157, + 288, + 312 + ], + "type": "inline_equation", + "content": "\\mathbf{c}" + }, + { + "bbox": [ + 46, + 157, + 288, + 312 + ], + "type": "text", + "content": " encoded with class labels, the diffusion models are trained by the squared error between the predicted noise " + }, + { + "bbox": [ + 46, + 157, + 288, + 312 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}(\\mathbf{z}_t,\\mathbf{c})" + }, + { + "bbox": [ + 46, + 157, + 288, + 312 + ], + "type": "text", + "content": " and the ground truth " + }, + { + "bbox": [ + 46, + 157, + 288, + 312 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 46, + 157, + 288, + 312 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 108, + 319, + 287, + 335 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 108, + 319, + 287, + 335 + ], + "spans": [ + { + "bbox": [ + 108, + 319, + 287, + 335 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {s i m p l e}} = \\left| \\left| \\epsilon_ {\\theta} \\left(\\mathbf {z} _ {t}, \\mathbf {c}\\right) - \\epsilon \\right| \\right| _ {2} ^ {2}, \\tag {1}", + "image_path": "58812d4aa28eea26431dec8d432aefdfdd8538ee4c04ef94666cb5fd5e907d46.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 341, + 287, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 341, + 287, + 460 + ], + "spans": [ + { + "bbox": [ + 46, + 341, + 287, + 460 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 341, + 287, + 460 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}" + }, + { + "bbox": [ + 46, + 341, + 287, + 460 + ], + "type": "text", + "content": " is a noise prediction network parameterized by " + }, + { + "bbox": [ + 46, + 341, + 287, + 460 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 46, + 341, + 287, + 460 + ], + "type": "text", + "content": ". Diffusion models are proven to generate images of higher quality compared with GANs [8]. There are also some Parameter Efficient Fine-Tuning (PEFT) methods updating a small number of model parameters in order for the model to be better adapted to specific data domains [34, 50]. We adopt DiT [33] as the baseline and Difffit [50] as the naive fine-tuning method for image generation. The generated images are compared with the original data from the perspective of embedding distribution in Fig. 3." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 461, + 287, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 461, + 287, + 641 + ], + "spans": [ + { + "bbox": [ + 46, + 461, + 287, + 641 + ], + "type": "text", + "content": "The samples of random selection and pre-trained diffusion models present two extreme ends of the distribution. Random selection faithfully reflects the original distribution, yet fails to emphasize some high-density regions. In contrast, diffusion models are over-fitted to those dense areas, leaving a large part of the original distribution uncovered. We attribute these two distributions to two properties, respectively. The randomly selected data holds extraordinary diversity, and the diffusion-generated data shows representativeness to the original distribution. We claim that both properties are essential for constructing an effective surrogate dataset. By naive fine-tuning, Difffit better captures the representative regions but leaves more regions uncovered. To this end, we propose extra minimax criteria for the diffusion model to enhance both of the properties." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 647, + 200, + 659 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 647, + 200, + 659 + ], + "spans": [ + { + "bbox": [ + 47, + 647, + 200, + 659 + ], + "type": "text", + "content": "2.3. Minimax Diffusion Criteria" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 665, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 287, + 713 + ], + "type": "text", + "content": "Based on the observation that representativeness and diversity are two key factors to construct an effective surrogate dataset, we accordingly design extra minimax criteria to enhance these two essential properties for the diffusion model." + } + ] + } + ], + "index": 7 + }, + { + "type": "code", + "bbox": [ + 310, + 88, + 545, + 277 + ], + "blocks": [ + { + "bbox": [ + 312, + 75, + 499, + 87 + ], + "lines": [ + { + "bbox": [ + 312, + 75, + 499, + 87 + ], + "spans": [ + { + "bbox": [ + 312, + 75, + 499, + 87 + ], + "type": "text", + "content": "Algorithm 1: Minimax Diffusion Fine-tuning" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "code_caption" + }, + { + "bbox": [ + 310, + 88, + 545, + 277 + ], + "lines": [ + { + "bbox": [ + 310, + 88, + 545, + 277 + ], + "spans": [ + { + "bbox": [ + 310, + 88, + 545, + 277 + ], + "type": "text", + "content": "Input: initialized model parameter " + }, + { + "bbox": [ + 310, + 88, + 545, + 277 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 310, + 88, + 545, + 277 + ], + "type": "text", + "content": " original dataset " + }, + { + "bbox": [ + 310, + 88, + 545, + 277 + ], + "type": "inline_equation", + "content": "\\mathcal{T} = \\{(\\mathbf{x},y)\\}" + }, + { + "bbox": [ + 310, + 88, + 545, + 277 + ], + "type": "text", + "content": " , encoder " + }, + { + "bbox": [ + 310, + 88, + 545, + 277 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 310, + 88, + 545, + 277 + ], + "type": "text", + "content": " class encoder " + }, + { + "bbox": [ + 310, + 88, + 545, + 277 + ], + "type": "inline_equation", + "content": "E_{c}" + }, + { + "bbox": [ + 310, + 88, + 545, + 277 + ], + "type": "text", + "content": " time step " + }, + { + "bbox": [ + 310, + 88, + 545, + 277 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 310, + 88, + 545, + 277 + ], + "type": "text", + "content": " variance schedule " + }, + { + "bbox": [ + 310, + 88, + 545, + 277 + ], + "type": "inline_equation", + "content": "\\bar{\\alpha}_t" + }, + { + "bbox": [ + 310, + 88, + 545, + 277 + ], + "type": "text", + "content": " real embedding memory " + }, + { + "bbox": [ + 310, + 88, + 545, + 277 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 310, + 88, + 545, + 277 + ], + "type": "text", + "content": " predicted embedding memory " + }, + { + "bbox": [ + 310, + 88, + 545, + 277 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 310, + 88, + 545, + 277 + ], + "type": "text", + "content": " \nOutput: optimized model parameter " + }, + { + "bbox": [ + 310, + 88, + 545, + 277 + ], + "type": "inline_equation", + "content": "\\theta^{*}" + }, + { + "bbox": [ + 310, + 88, + 545, + 277 + ], + "type": "text", + "content": " \nfor each step do Obtain the original embedding: " + }, + { + "bbox": [ + 310, + 88, + 545, + 277 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_0 = E(\\mathbf{x})" + }, + { + "bbox": [ + 310, + 88, + 545, + 277 + ], + "type": "text", + "content": " Obtain the class embedding: " + }, + { + "bbox": [ + 310, + 88, + 545, + 277 + ], + "type": "inline_equation", + "content": "\\mathbf{c} = E_c(y)" + }, + { + "bbox": [ + 310, + 88, + 545, + 277 + ], + "type": "text", + "content": " Sample random noise: " + }, + { + "bbox": [ + 310, + 88, + 545, + 277 + ], + "type": "inline_equation", + "content": "\\epsilon \\sim \\mathcal{N}(\\mathbf{0},\\mathbf{I})" + }, + { + "bbox": [ + 310, + 88, + 545, + 277 + ], + "type": "text", + "content": " Add noise to the embedding: " + }, + { + "bbox": [ + 310, + 88, + 545, + 277 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_t = \\sqrt{\\bar{\\alpha}_t}\\mathbf{z}_0 + \\sqrt{1 - \\bar{\\alpha}_t}\\epsilon" + }, + { + "bbox": [ + 310, + 88, + 545, + 277 + ], + "type": "text", + "content": " Predict the noise " + }, + { + "bbox": [ + 310, + 88, + 545, + 277 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}(\\mathbf{z}_{t},\\mathbf{c})" + }, + { + "bbox": [ + 310, + 88, + 545, + 277 + ], + "type": "text", + "content": " and recovered embedding " + }, + { + "bbox": [ + 310, + 88, + 545, + 277 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{z}}_{\\theta}(\\mathbf{z}_{t},\\mathbf{c}) = \\mathbf{z}_{t} - \\epsilon_{\\theta}(\\mathbf{z}_{t},\\mathbf{c})" + }, + { + "bbox": [ + 310, + 88, + 545, + 277 + ], + "type": "text", + "content": " Update the model parameter with Eq. (5) Enqueue the real embedding: " + }, + { + "bbox": [ + 310, + 88, + 545, + 277 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_r\\gets \\mathbf{z}_0" + }, + { + "bbox": [ + 310, + 88, + 545, + 277 + ], + "type": "text", + "content": " Enqueue the predicted embedding: " + }, + { + "bbox": [ + 310, + 88, + 545, + 277 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_d\\gets \\hat{\\mathbf{z}}_\\theta (\\mathbf{z}_t,\\mathbf{c})" + }, + { + "bbox": [ + 310, + 88, + 545, + 277 + ], + "type": "text", + "content": " \nend" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "code_body" + } + ], + "index": 9, + "sub_type": "algorithm" + }, + { + "bbox": [ + 305, + 304, + 545, + 352 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 304, + 545, + 352 + ], + "spans": [ + { + "bbox": [ + 305, + 304, + 545, + 352 + ], + "type": "text", + "content": "Representativeness It is essential for the small surrogate dataset to sufficiently represent the original data. A naive approach to improve the representativeness is aligning the embedding distribution between synthetic and real samples:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 338, + 360, + 545, + 393 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 338, + 360, + 545, + 393 + ], + "spans": [ + { + "bbox": [ + 338, + 360, + 545, + 393 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {r} = \\underset {\\theta} {\\arg \\max } \\sigma \\left(\\hat {\\mathbf {z}} _ {\\theta} \\left(\\mathbf {z} _ {t}, \\mathbf {c}\\right), \\frac {1}{N _ {B}} \\sum_ {i = 0} ^ {N _ {B}} \\mathbf {z} _ {i}\\right), \\tag {2}", + "image_path": "b1853e41bf43f82ba7c16735e7efb719d4ee0853a9557685e34f4624be4c529b.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 398, + 545, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 398, + 545, + 518 + ], + "spans": [ + { + "bbox": [ + 304, + 398, + 545, + 518 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 398, + 545, + 518 + ], + "type": "inline_equation", + "content": "\\sigma (\\cdot ,\\cdot)" + }, + { + "bbox": [ + 304, + 398, + 545, + 518 + ], + "type": "text", + "content": " is the cosine similarity, " + }, + { + "bbox": [ + 304, + 398, + 545, + 518 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{z}}_{\\theta}(\\mathbf{z}_t,\\mathbf{c})" + }, + { + "bbox": [ + 304, + 398, + 545, + 518 + ], + "type": "text", + "content": " is the predicted original embedding by subtracting the noise from the noisy embedding " + }, + { + "bbox": [ + 304, + 398, + 545, + 518 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{z}}_{\\theta}(\\mathbf{z}_t,\\mathbf{c}) = \\mathbf{z}_t - \\epsilon_{\\theta}(\\mathbf{z}_t,\\mathbf{c})" + }, + { + "bbox": [ + 304, + 398, + 545, + 518 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 398, + 545, + 518 + ], + "type": "inline_equation", + "content": "N_B" + }, + { + "bbox": [ + 304, + 398, + 545, + 518 + ], + "type": "text", + "content": " is the size of the sampled real sample mini-batch. However, the naive alignment tends to draw the predicted embedding towards the center of the real distribution, which severely limits the diversity. Therefore, we propose to maintain an auxiliary memory " + }, + { + "bbox": [ + 304, + 398, + 545, + 518 + ], + "type": "inline_equation", + "content": "\\mathcal{M} = \\{\\mathbf{z}_m\\}_{m = 1}^{N_M}" + }, + { + "bbox": [ + 304, + 398, + 545, + 518 + ], + "type": "text", + "content": " to store the real samples utilized in adjacent iterations, and design a minimax optimization objective as:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 341, + 527, + 545, + 547 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 341, + 527, + 545, + 547 + ], + "spans": [ + { + "bbox": [ + 341, + 527, + 545, + 547 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {r} = \\arg \\max _ {\\theta} \\min _ {m \\in \\left[ N _ {M} \\right]} \\sigma \\left(\\hat {\\mathbf {z}} _ {\\theta} \\left(\\mathbf {z} _ {t}, \\mathbf {c}\\right), \\mathbf {z} _ {m}\\right). \\tag {3}", + "image_path": "d885afc23f4116dcc320f36d707116d324c68e11039e3616d1e86fbe8f806818.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 555, + 545, + 627 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 555, + 545, + 627 + ], + "spans": [ + { + "bbox": [ + 304, + 555, + 545, + 627 + ], + "type": "text", + "content": "By pulling close the least similar sample pairs, the diffusion model is encouraged to generate images that better cover the original distribution. It is notable that the diffusion training objective " + }, + { + "bbox": [ + 304, + 555, + 545, + 627 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{simple}" + }, + { + "bbox": [ + 304, + 555, + 545, + 627 + ], + "type": "text", + "content": " itself encourages the generated images to resemble the original ones. Thus, the minimax criterion allows the preservation of diversity to the maximum extent." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "type": "text", + "content": "Diversity Although the pre-trained diffusion models already achieve satisfactory generation quality, the remaining defect is limited diversity compared with the original data, as shown in Fig. 3. We expect the data generated by the diffusion model can accurately reflect the original distribution, while simultaneously being different from" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 733, + 317, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 733, + 317, + 742 + ], + "spans": [ + { + "bbox": [ + 294, + 733, + 317, + 742 + ], + "type": "text", + "content": "15795" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 287, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 287, + 120 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 287, + 120 + ], + "type": "text", + "content": "each other. Hence, we maintain another auxiliary memory " + }, + { + "bbox": [ + 47, + 72, + 287, + 120 + ], + "type": "inline_equation", + "content": "\\mathcal{D} = \\{\\mathbf{z}_d\\}_{d=1}^{N_D}" + }, + { + "bbox": [ + 47, + 72, + 287, + 120 + ], + "type": "text", + "content": " for the predicted embeddings of adjacent iterations and design another minimax objective to explicitly enhance the sample diversity as:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 87, + 127, + 287, + 147 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 87, + 127, + 287, + 147 + ], + "spans": [ + { + "bbox": [ + 87, + 127, + 287, + 147 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {d} = \\arg \\min _ {\\theta} \\max _ {d \\in [ N _ {D} ]} \\sigma (\\hat {\\mathbf {z}} _ {\\theta} (\\mathbf {z} _ {t}, \\mathbf {c}), \\mathbf {z} _ {d}). \\tag {4}", + "image_path": "a03caa49aca7de747090504880145cbda674750c438d82715f826813b06c7fb3.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 152, + 287, + 247 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 152, + 287, + 247 + ], + "spans": [ + { + "bbox": [ + 46, + 152, + 287, + 247 + ], + "type": "text", + "content": "The diversity term has an opposite optimization target compared with the representativeness term, where the predicted embedding is pushed away from the most similar one stored in the memory bank. Although diversity is essential for an effective surrogate set, too much of it will cause the generated data to lose representativeness. The proposed minimax optimization enhances the diversity in a gentle way, with less influence on the class-related features." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 248, + 287, + 283 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 248, + 287, + 283 + ], + "spans": [ + { + "bbox": [ + 46, + 248, + 287, + 283 + ], + "type": "text", + "content": "Combining all the components, we summarize the training process in Algorithm 1. The complete training objective can be formulated as:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 104, + 292, + 287, + 304 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 104, + 292, + 287, + 304 + ], + "spans": [ + { + "bbox": [ + 104, + 292, + 287, + 304 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\mathcal {L} _ {\\text {s i m p l e}} + \\lambda_ {r} \\mathcal {L} _ {r} + \\lambda_ {d} \\mathcal {L} _ {d}, \\tag {5}", + "image_path": "cb6994438d7be066768a4f541be8cbce947bac34c6671ee4499cd0bb8152415d.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 312, + 249, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 312, + 249, + 323 + ], + "spans": [ + { + "bbox": [ + 47, + 312, + 249, + 323 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 312, + 249, + 323 + ], + "type": "inline_equation", + "content": "\\lambda_r" + }, + { + "bbox": [ + 47, + 312, + 249, + 323 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 312, + 249, + 323 + ], + "type": "inline_equation", + "content": "\\lambda_d" + }, + { + "bbox": [ + 47, + 312, + 249, + 323 + ], + "type": "text", + "content": " are weighting hyper-parameters." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 333, + 167, + 346 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 333, + 167, + 346 + ], + "spans": [ + { + "bbox": [ + 47, + 333, + 167, + 346 + ], + "type": "text", + "content": "3. Theoretical Analysis" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 353, + 287, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 353, + 287, + 389 + ], + "spans": [ + { + "bbox": [ + 47, + 353, + 287, + 389 + ], + "type": "text", + "content": "Assume that " + }, + { + "bbox": [ + 47, + 353, + 287, + 389 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 47, + 353, + 287, + 389 + ], + "type": "text", + "content": " is the real distribution of the latent variables " + }, + { + "bbox": [ + 47, + 353, + 287, + 389 + ], + "type": "inline_equation", + "content": "\\mathbf{z}" + }, + { + "bbox": [ + 47, + 353, + 287, + 389 + ], + "type": "text", + "content": " associated with the target dataset " + }, + { + "bbox": [ + 47, + 353, + 287, + 389 + ], + "type": "inline_equation", + "content": "\\mathcal{T}" + }, + { + "bbox": [ + 47, + 353, + 287, + 389 + ], + "type": "text", + "content": ". We rewrite the optimization problem presented in Eq. (5) in a modified form:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 52, + 404, + 288, + 459 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 52, + 404, + 288, + 459 + ], + "spans": [ + { + "bbox": [ + 52, + 404, + 288, + 459 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\min_{\\{\\theta^{(i)}\\}_{i\\in [N_{D}]}}\\lambda_{d}\\max_{i,j = 1,\\ldots ,N_{D}}\\sigma \\left(\\hat{\\mathbf{z}} (\\theta^{(i)}),\\hat{\\mathbf{z}} (\\theta^{(j)})\\right) \\\\ + \\sum_ {i = 1} ^ {N _ {D}} \\left\\{- \\lambda_ {r} Q _ {\\tilde {q}, w \\sim \\mu} \\left[ \\sigma \\left(\\hat {\\mathbf {z}} \\left(\\theta^ {(i)}\\right), w\\right) \\right] + \\| \\hat {\\mathbf {z}} \\left(\\theta^ {(i)}\\right) - \\mathbf {z} _ {0} ^ {(i)} \\| ^ {2} \\right\\}, \\tag {6} \\\\ \\end{array}", + "image_path": "3c429f1aec0dccb29f67351bd2f172034fd922c4658e73053e00009f7525525a.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 460, + 287, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 460, + 287, + 555 + ], + "spans": [ + { + "bbox": [ + 46, + 460, + 287, + 555 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 460, + 287, + 555 + ], + "type": "inline_equation", + "content": "Q_{\\tilde{q}}[\\cdot]" + }, + { + "bbox": [ + 46, + 460, + 287, + 555 + ], + "type": "text", + "content": " denotes the quantile function with " + }, + { + "bbox": [ + 46, + 460, + 287, + 555 + ], + "type": "inline_equation", + "content": "\\tilde{q}" + }, + { + "bbox": [ + 46, + 460, + 287, + 555 + ], + "type": "text", + "content": " as the quantile percentage. Note that here we consider a theoretical idealized variant of our algorithm wherein we perform simultaneous generation of all the embeddings " + }, + { + "bbox": [ + 46, + 460, + 287, + 555 + ], + "type": "inline_equation", + "content": "\\{\\hat{\\mathbf{z}}(\\theta^{(i)})\\}" + }, + { + "bbox": [ + 46, + 460, + 287, + 555 + ], + "type": "text", + "content": ", rather than sample by sample. Hence the objectives turn to the sum of pairwise similarities rather than the form in Eq. (4). And we minimize the negative to aim for maximal representativeness, as in Eq. (5)." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 556, + 287, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 556, + 287, + 616 + ], + "spans": [ + { + "bbox": [ + 46, + 556, + 287, + 616 + ], + "type": "text", + "content": "It can be considered as a scalarized solution to a multi-objective optimization problem, wherein multiple criteria are weighed (see, e.g. [13]). This perspective aligns with a Pareto front with trade-offs. It means that one objective decreasing will by necessity result in another increasing." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 616, + 287, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 616, + 287, + 640 + ], + "spans": [ + { + "bbox": [ + 47, + 616, + 287, + 640 + ], + "type": "text", + "content": "However, consider that any solution to the following tri-level optimization problem is also a solution for Eq. (6):" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 52, + 645, + 287, + 713 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 52, + 645, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 52, + 645, + 287, + 713 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\min _ {\\{\\theta^ {(i)} \\} _ {i \\in [ N _ {D} ]}} \\max _ {i, j = 1,.., N _ {D}} \\sigma \\left(\\hat {\\mathbf {z}} (\\theta^ {(i)}), \\hat {\\mathbf {z}} (\\theta^ {(j)})\\right) \\\\ \\text {s u b j . t o} \\quad \\left\\{\\theta^ {(i)} \\right\\} \\in \\arg \\max \\sum_ {i = 1} ^ {N _ {D}} Q _ {\\tilde {q}, w \\sim \\mu} \\left[ \\sigma (\\hat {\\mathbf {z}} (\\theta^ {(i)}), w) \\right] \\\\ \\text {s u b j . t o} \\quad \\theta^ {(i)} \\in \\arg \\min \\| \\hat {\\mathbf {z}} (\\theta) - \\mathbf {z} _ {0} ^ {(i)} \\| ^ {2}, \\forall i \\in [ N _ {D} ]. \\tag {7} \\\\ \\end{array}", + "image_path": "7c25dbaa108934512f4f0d91dc8af39d4f2bca8fac3bacb1043a5837ba506a04.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 72, + 545, + 143 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 143 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 143 + ], + "type": "text", + "content": "If a solution to Eq. (7) is discovered, either incidentally through solving Eq. (6) or by careful tuning of step sizes, the set of minimizers will be sufficiently large at both levels, with no trade-offs involved. However, can we justify the presumption that there exists a meaningful set of potential minimizers?" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 158, + 545, + 241 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 158, + 545, + 241 + ], + "spans": [ + { + "bbox": [ + 304, + 158, + 545, + 241 + ], + "type": "text", + "content": "Diffusion Process Model One popular framework for the mathematical analysis of diffusion involves analyzing the convergence and asymptotic properties of, appropriately homonymous, diffusion processes. These processes are characterized by the standard stochastic differential equation with a drift and diffusion term. For a time-dependent random variable " + }, + { + "bbox": [ + 304, + 158, + 545, + 241 + ], + "type": "inline_equation", + "content": "Z_{t}" + }, + { + "bbox": [ + 304, + 158, + 545, + 241 + ], + "type": "text", + "content": "," + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 377, + 246, + 545, + 259 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 377, + 246, + 545, + 259 + ], + "spans": [ + { + "bbox": [ + 377, + 246, + 545, + 259 + ], + "type": "interline_equation", + "content": "d Z _ {t} = V \\left(Z _ {t}\\right) d t + d W _ {t} \\tag {8}", + "image_path": "9ae8d7631465492daab063f49242ee48c4cb6d55035453ca56cdbf885b8cbb1d.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 264, + 545, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 264, + 545, + 346 + ], + "spans": [ + { + "bbox": [ + 304, + 264, + 545, + 346 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 264, + 545, + 346 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 304, + 264, + 545, + 346 + ], + "type": "text", + "content": " is a drift function dependent on the current " + }, + { + "bbox": [ + 304, + 264, + 545, + 346 + ], + "type": "inline_equation", + "content": "Z_{t}" + }, + { + "bbox": [ + 304, + 264, + 545, + 346 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 264, + 545, + 346 + ], + "type": "inline_equation", + "content": "dW_{t}" + }, + { + "bbox": [ + 304, + 264, + 545, + 346 + ], + "type": "text", + "content": " is a Wiener (Brownian noise) process. This equation serves as an appropriate continuous approximation of generative diffusion, given that Brownian noise is a continuous limit of adding normal random variables. Consequently, we aim for any realization " + }, + { + "bbox": [ + 304, + 264, + 545, + 346 + ], + "type": "inline_equation", + "content": "\\mathbf{z} \\sim Z_{t}" + }, + { + "bbox": [ + 304, + 264, + 545, + 346 + ], + "type": "text", + "content": " to have certain desired properties that reflect generative modeling with high probability." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 347, + 545, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 347, + 545, + 430 + ], + "spans": [ + { + "bbox": [ + 304, + 347, + 545, + 430 + ], + "type": "text", + "content": "The work [43] established a theoretical model utilizing concepts in the control of these diffusions, demonstrating how it can result in sampling from the distribution of a desired data set. In the supplementary material we present a description of their framework and present an argument supporting the well-defined nature of the following problem, indicating that it has non-trivial solutions." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 431, + 545, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 431, + 545, + 514 + ], + "spans": [ + { + "bbox": [ + 304, + 431, + 545, + 514 + ], + "type": "text", + "content": "When sampling optimally from the population dataset, we consider a stochastic control problem wherein " + }, + { + "bbox": [ + 304, + 431, + 545, + 514 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 304, + 431, + 545, + 514 + ], + "type": "text", + "content": " depends also on some chosen control " + }, + { + "bbox": [ + 304, + 431, + 545, + 514 + ], + "type": "inline_equation", + "content": "u(\\mathbf{z},t)" + }, + { + "bbox": [ + 304, + 431, + 545, + 514 + ], + "type": "text", + "content": ". This control aims to find the most representative samples and, among the possible collection of such samples, to obtain the most diverse one while sampling from the desired dataset " + }, + { + "bbox": [ + 304, + 431, + 545, + 514 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 304, + 431, + 545, + 514 + ], + "type": "text", + "content": ". This involves solving:" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 321, + 518, + 460, + 538 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 518, + 460, + 538 + ], + "spans": [ + { + "bbox": [ + 321, + 518, + 460, + 538 + ], + "type": "interline_equation", + "content": "\\min _ {u (x, t)} \\max _ {i, j = 1,.., N _ {D}} \\sigma \\left(Z _ {1} ^ {u, (i)}, Z _ {1} ^ {u, (j)}\\right)", + "image_path": "8b2004a35c062c6bf6bf4de55b9120bde5b2ccd7b3772f76e68cecd687d407a2.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 312, + 539, + 544, + 594 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 539, + 544, + 594 + ], + "spans": [ + { + "bbox": [ + 312, + 539, + 544, + 594 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathrm {s u b j . t o} \\quad u \\in \\arg \\max \\sum_ {i = 1} ^ {N _ {D}} \\int_ {0} ^ {1} \\mathbb {E} _ {Z _ {t} ^ {(i)}} Q _ {\\tilde {q}, w \\sim \\mu} \\left[ \\sigma \\left(Z _ {t} ^ {(i)}, w\\right) \\right] d s \\\\ Z _ {1} \\sim \\mu , \\\\ d Z _ {t} ^ {u, (i)} = u \\left(Z _ {t} ^ {u, (i)}, t\\right) d t + d W _ {t}, t \\in [ 0, 1 ]; \\\\ Z _ {0} = \\mathbf {z} _ {0}. \\\\ \\end{array}", + "image_path": "533598cabf9d378cf64a0ea274d7f2a748225092c27bc78ca0ae7c58a1fcd711.jpg" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 533, + 595, + 545, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 533, + 595, + 545, + 605 + ], + "spans": [ + { + "bbox": [ + 533, + 595, + 545, + 605 + ], + "type": "text", + "content": "(9)" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": "This problem poses a bi-level stochastic control challenge where employing a layered dynamic programming is far from tractable. Additionally, a multi-stage stochastic programming approximation would also be infeasible given the scale of the datasets involved. Instead, we opt for parameterization with a neural network, forego exact sampling, discretize the problem, redefine the criteria to be time independent, and seek to solve an approximate solution for the tri-level optimization problem Eq. (7)." + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 732, + 317, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 732, + 317, + 742 + ], + "spans": [ + { + "bbox": [ + 293, + 732, + 317, + 742 + ], + "type": "text", + "content": "15796" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 49, + 102, + 545, + 304 + ], + "blocks": [ + { + "bbox": [ + 47, + 70, + 545, + 94 + ], + "lines": [ + { + "bbox": [ + 47, + 70, + 545, + 94 + ], + "spans": [ + { + "bbox": [ + 47, + 70, + 545, + 94 + ], + "type": "text", + "content": "Table 1. Performance comparison with pre-trained diffusion models and other state-of-the-art methods on ImageWoof. All the results are reproduced by us on the " + }, + { + "bbox": [ + 47, + 70, + 545, + 94 + ], + "type": "inline_equation", + "content": "256\\times 256" + }, + { + "bbox": [ + 47, + 70, + 545, + 94 + ], + "type": "text", + "content": " resolution. The missing results are due to out-of-memory. The best results are marked as bold." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 49, + 102, + 545, + 304 + ], + "lines": [ + { + "bbox": [ + 49, + 102, + 545, + 304 + ], + "spans": [ + { + "bbox": [ + 49, + 102, + 545, + 304 + ], + "type": "table", + "html": "
IPC (Ratio)Test ModelRandomK-Center [37]Herding [48]DiT [33]DM [56]IDC-1 [21]GLaD [4]OursFull
10 (0.8%)ConvNet-624.3±1.119.4±0.926.7±0.534.2±1.126.9±1.233.3±1.133.8±0.937.0±1.086.4±0.2
ResNetAP-1029.4±0.822.1±0.132.0±0.334.7±0.530.3±1.239.1±0.532.9±0.939.2±1.387.5±0.5
ResNet-1827.7±0.921.1±0.430.2±1.234.7±0.433.4±0.737.3±0.231.7±0.837.6±0.989.3±1.2
20 (1.6%)ConvNet-629.1±0.721.5±0.829.5±0.336.1±0.829.9±1.035.5±0.8-37.6±0.286.4±0.2
ResNetAP-1032.7±0.425.1±0.734.9±0.141.1±0.835.2±0.643.4±0.3-45.8±0.587.5±0.5
ResNet-1829.7±0.523.6±0.332.2±0.640.5±0.529.8±1.738.6±0.2-42.5±0.689.3±1.2
50 (3.8%)ConvNet-641.3±0.636.5±1.040.3±0.746.5±0.844.4±1.043.9±1.2-53.9±0.686.4±0.2
ResNetAP-1047.2±1.340.6±0.449.1±0.749.3±0.247.1±1.148.3±1.0-56.3±1.087.5±0.5
ResNet-1847.9±1.839.6±1.048.3±1.250.1±0.546.2±0.648.3±0.8-57.1±0.689.3±1.2
70 (5.4%)ConvNet-646.3±0.638.6±0.746.2±0.650.1±1.247.5±0.848.9±0.7-55.7±0.986.4±0.2
ResNetAP-1050.8±0.645.9±1.553.4±1.454.3±0.951.7±0.852.8±1.8-58.3±0.287.5±0.5
ResNet-1852.1±1.044.6±1.149.7±0.851.5±1.051.9±0.851.1±1.7-58.8±0.789.3±1.2
100 (7.7%)ConvNet-652.2±0.445.1±0.554.4±1.153.4±0.355.0±1.353.2±0.9-61.1±0.786.4±0.2
ResNetAP-1059.4±1.054.8±0.261.7±0.958.3±0.856.4±0.856.1±0.9-64.5±0.287.5±0.5
ResNet-1861.5±1.350.4±0.459.3±0.758.9±1.360.2±1.058.3±1.2-65.7±0.489.3±1.2
", + "image_path": "939fb4f0c2c36773432cb3723f5d6bb26c59ba228c09d8ee5169e0049bce1403.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 323, + 288, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 323, + 288, + 396 + ], + "spans": [ + { + "bbox": [ + 46, + 323, + 288, + 396 + ], + "type": "text", + "content": "In the supplementary material we provide a rationale for the meaningfulness of the problem in Eq. (9) based on the model of generative diffusion [43]. Specifically, we argue that the set of controls that leads to the desired final distribution and the set of minimizers, is sufficiently large for a low value of the objective at the top layer." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 407, + 128, + 421 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 407, + 128, + 421 + ], + "spans": [ + { + "bbox": [ + 47, + 407, + 128, + 421 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 427, + 180, + 441 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 427, + 180, + 441 + ], + "spans": [ + { + "bbox": [ + 47, + 427, + 180, + 441 + ], + "type": "text", + "content": "4.1. Implementation Details" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 445, + 287, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 445, + 287, + 567 + ], + "spans": [ + { + "bbox": [ + 46, + 445, + 287, + 567 + ], + "type": "text", + "content": "For the diffusion model, we adopt pre-trained DiT [33] as the baseline and conduct PEFT with Difffit [50]. " + }, + { + "bbox": [ + 46, + 445, + 287, + 567 + ], + "type": "inline_equation", + "content": "\\lambda_r" + }, + { + "bbox": [ + 46, + 445, + 287, + 567 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 445, + 287, + 567 + ], + "type": "inline_equation", + "content": "\\lambda_d" + }, + { + "bbox": [ + 46, + 445, + 287, + 567 + ], + "type": "text", + "content": " are set as 0.002 and 0.008 for Eq. (5), respectively. The image size for the diffusion fine-tuning and sample generation is set as " + }, + { + "bbox": [ + 46, + 445, + 287, + 567 + ], + "type": "inline_equation", + "content": "256\\times 256" + }, + { + "bbox": [ + 46, + 445, + 287, + 567 + ], + "type": "text", + "content": ". The fine-tuning mini-batch size is set as 8, and the fine-tuning lasts 8 epochs. The learning rate is set as 1e-3 for an AdamW optimizer. After fine-tuning, the images are generated by 50 denoising steps on a pre-defined number of random noise, according to the IPC setting. All the experiments are conducted on a single RTX 4090 GPU." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 574, + 219, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 574, + 219, + 586 + ], + "spans": [ + { + "bbox": [ + 47, + 574, + 219, + 586 + ], + "type": "text", + "content": "4.2. Datasets and Evaluation Metric" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 594, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 288, + 715 + ], + "type": "text", + "content": "For practical applicability, the experiments are exclusively conducted on full-sized ImageNet [6] subsets in this work. The selected subsets include ImageWoof, ImageNette [15] and the 10-class split adopted in [21, 42], denoted as ImageIDC afterward. ImageWoof is a challenging subset, containing only classes of dog breeds, while ImageNette and ImageIDC contain classes with less similarity, and hence are easier to discriminate. For evaluation, we adopt the same setting as in [21]. The surrogate dataset is trained on different model architectures, with a learning rate of 0.01," + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 309, + 353, + 544, + 399 + ], + "blocks": [ + { + "bbox": [ + 305, + 322, + 545, + 344 + ], + "lines": [ + { + "bbox": [ + 305, + 322, + 545, + 344 + ], + "spans": [ + { + "bbox": [ + 305, + 322, + 545, + 344 + ], + "type": "text", + "content": "Table 2. The Maximum Mean Discrepancy (MMD) between the extracted features of surrogate dataset and the original one." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 309, + 353, + 544, + 399 + ], + "lines": [ + { + "bbox": [ + 309, + 353, + 544, + 399 + ], + "spans": [ + { + "bbox": [ + 309, + 353, + 544, + 399 + ], + "type": "table", + "html": "
IPCDiT [33]Diffit [50]DM [56]IDC-1 [21]Ours
505.45.44.86.74.0
1005.55.34.06.44.3
", + "image_path": "9d6fe3cc879402975ab700526d893bfc900a554e120e884afd059880c972168c.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 420, + 545, + 458 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 420, + 545, + 458 + ], + "spans": [ + { + "bbox": [ + 305, + 420, + 545, + 458 + ], + "type": "text", + "content": "and a scheduler decaying the learning rate at 2/3 and 5/6 of the whole training iterations. The top-1 accuracy on the original testing set is reported to illustrate the performance." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 466, + 529, + 479 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 466, + 529, + 479 + ], + "spans": [ + { + "bbox": [ + 305, + 466, + 529, + 479 + ], + "type": "text", + "content": "4.3. Comparison with State-of-the-art Methods" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 485, + 545, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 485, + 545, + 628 + ], + "spans": [ + { + "bbox": [ + 304, + 485, + 545, + 628 + ], + "type": "text", + "content": "We compare our method with other state-of-the-art methods across different IPC settings and model architectures. For a fair comparison, the results are all reproduced by us under the same evaluation protocol. ResNet-10 [18] with average pooling is adopted for matching the feature distribution (DM [56], GLaD [4]) and training gradients (IDC-1 [21]). DM is implemented on IDC-1 by only modifying the matching objective from training gradients to feature distribution, such that better performance is achieved. Each experiment is conducted 3 times, with the mean value and standard variance reported. Firstly, we present the validation results on the challenging ImageWoof subset [15] in Tab. 1." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "content": "With the target of distilling surrogate datasets of small IPCs (e.g., 10 and 20), the pixel-level optimization method IDC-1 [21] demonstrates outstanding performance gain over random original images. However, as the IPC increases, the performance gain drastically drops. Especially under the 100-IPC setting, the distilled dataset even performs worse than random original images. This observation" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 733, + 317, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 733, + 317, + 742 + ], + "spans": [ + { + "bbox": [ + 293, + 733, + 317, + 742 + ], + "type": "text", + "content": "15797" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 50, + 124, + 286, + 219 + ], + "blocks": [ + { + "bbox": [ + 47, + 70, + 288, + 114 + ], + "lines": [ + { + "bbox": [ + 47, + 70, + 288, + 114 + ], + "spans": [ + { + "bbox": [ + 47, + 70, + 288, + 114 + ], + "type": "text", + "content": "Table 3. Performance comparison with pre-trained diffusion models and state-of-the-art methods on more ImageNet subsets. The results are obtained on ResNet-10 with average pooling. The best results are marked as **bold**." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 50, + 124, + 286, + 219 + ], + "lines": [ + { + "bbox": [ + 50, + 124, + 286, + 219 + ], + "spans": [ + { + "bbox": [ + 50, + 124, + 286, + 219 + ], + "type": "table", + "html": "
IPCRandomDiT [33]DM [56]Ours
Nette1054.2±1.659.1±0.760.8±0.662.0±0.2
2063.5±0.564.8±1.266.5±1.166.8±0.4
5076.1±1.173.3±0.976.2±0.476.6±0.2
IDC1048.1±0.854.1±0.452.8±0.553.1±0.2
2052.5±0.958.9±0.258.5±0.459.0±0.4
5068.1±0.764.3±0.669.1±0.869.6±0.2
", + "image_path": "ddce42543c1a7adc18c94afcf8e9c7eb4d4b54fbd55e499143cd635bbcab4de0.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 242, + 287, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 242, + 287, + 350 + ], + "spans": [ + { + "bbox": [ + 46, + 242, + 287, + 350 + ], + "type": "text", + "content": "aligns with the empirical findings in Fig. 2, where pixel-level methods struggle to optimize the expanded parameter space of large IPCs. The embedding-level optimization method GLaD [4] yields good performance under the 10-IPC setting. However, it requires overwhelming GPU resources for larger IPC settings, which is inapplicable for resource-restricted scenarios. It is also notable that on large IPCs, the coreset method Herding [48] surpasses previous DD methods with far less computational cost." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 351, + 287, + 532 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 351, + 287, + 532 + ], + "spans": [ + { + "bbox": [ + 46, + 351, + 287, + 532 + ], + "type": "text", + "content": "The pre-trained DiT [33] here serves as the baseline for generative diffusion techniques. Under the 50-IPC setting, DiT outperforms both random original images and IDC-1. However, the insufficiency of representativeness and diversity restricts its performance on smaller and larger IPCs, respectively. In contrast, our proposed minimax diffusion consistently provides superior performance across all the IPCs over both original images and Herding. Besides, the proposed method eliminates the need of specific network architectures for matching training metrics. Consequently, the cross-architecture generalization is significantly improved. Under most IPC settings, the performance gap between ConvNet-6 and ResNetAP-10 is even smaller than that of the original images. It validates the universality of the rich information learned by the minimax fine-tuning process." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 532, + 287, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 532, + 287, + 628 + ], + "spans": [ + { + "bbox": [ + 46, + 532, + 287, + 628 + ], + "type": "text", + "content": "Furthermore, we extensively assess the Maximum Mean Discrepancy (MMD) between the embedded features of the selected/generated surrogate dataset and the original one in Tab. 2. The features are extracted by a ResNet-10 network pre-trained on the full original dataset. Our method achieves the lowest discrepancy by average, where DM [56] directly sets MMD as the optimization target, proving the validity of extra minimax criteria in fitting distributions." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 630, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 630, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 630, + 286, + 713 + ], + "type": "text", + "content": "Moreover, we show the performance comparison on ImageNette [15] and ImageIDC [21] in Tab. 3. The performance trend generally aligns with that on ImageWoof. More specifically, on these two easier subsets, DiT quickly loses the advantage over original images as IPC increases. Conversely, our proposed minimax diffusion method consistently demonstrates state-of-the-art performance." + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 315, + 91, + 538, + 137 + ], + "blocks": [ + { + "bbox": [ + 331, + 71, + 520, + 82 + ], + "lines": [ + { + "bbox": [ + 331, + 71, + 520, + 82 + ], + "spans": [ + { + "bbox": [ + 331, + 71, + 520, + 82 + ], + "type": "text", + "content": "Table 4. Performance comparison on ImageNet-1K." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 315, + 91, + 538, + 137 + ], + "lines": [ + { + "bbox": [ + 315, + 91, + 538, + 137 + ], + "spans": [ + { + "bbox": [ + 315, + 91, + 538, + 137 + ], + "type": "table", + "html": "
IPCSRe2L [53]RDED [41]DiTOurs
1021.3±0.642.0±0.139.6±0.444.3±0.5
5046.8±0.256.5±0.152.9±0.658.6±0.3
", + "image_path": "e4d7a2c3908f462b109b909137737ebc5af7c07e23ad70082a217fda477dc655.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 310, + 150, + 431, + 322 + ], + "blocks": [ + { + "bbox": [ + 310, + 150, + 431, + 322 + ], + "lines": [ + { + "bbox": [ + 310, + 150, + 431, + 322 + ], + "spans": [ + { + "bbox": [ + 310, + 150, + 431, + 322 + ], + "type": "image", + "image_path": "eb4fcacef452f9e119c89cf526ff2f9fe92bfa55af573b4c3d94b37996e7e559.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 335, + 545, + 390 + ], + "lines": [ + { + "bbox": [ + 305, + 335, + 545, + 390 + ], + "spans": [ + { + "bbox": [ + 305, + 335, + 545, + 390 + ], + "type": "text", + "content": "Figure 4. With the help of the minimax diffusion, the proposed method significantly enhances the representativeness and diversity of the generated images. Thereby it consistently provides superior performance compared with random selection and baseline diffusion models by a large margin across different IPC settings." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 432, + 150, + 541, + 322 + ], + "blocks": [ + { + "bbox": [ + 432, + 150, + 541, + 322 + ], + "lines": [ + { + "bbox": [ + 432, + 150, + 541, + 322 + ], + "spans": [ + { + "bbox": [ + 432, + 150, + 541, + 322 + ], + "type": "image", + "image_path": "a12b7a4869b423baeb051db78d08ae72416aface9a0e686a33358074d414aa48.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 412, + 545, + 484 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 412, + 545, + 484 + ], + "spans": [ + { + "bbox": [ + 305, + 412, + 545, + 484 + ], + "type": "text", + "content": "Experiments on ImageNet-1K. We further conduct experiments on the full ImageNet-1K with the validation protocol of RDED [41] and present the results in Tab. 4. The synthetic images are resized to " + }, + { + "bbox": [ + 305, + 412, + 545, + 484 + ], + "type": "inline_equation", + "content": "224 \\times 224" + }, + { + "bbox": [ + 305, + 412, + 545, + 484 + ], + "type": "text", + "content": " for evaluation. The significant performance advantage over the compared works validates the scalability of the proposed method." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 491, + 400, + 504 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 491, + 400, + 504 + ], + "spans": [ + { + "bbox": [ + 306, + 491, + 400, + 504 + ], + "type": "text", + "content": "4.4. Ablation Study" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "text", + "content": "Component Analysis. We compare the performance with baseline diffusion models to validate the effectiveness of proposed minimax criteria in Fig. 4. The experiments are conducted on ImageWoof and ImageIDC to evaluate the effect on challenging and easy tasks, respectively. Under the IPC of 10 and 20, the raw diffusion models (DiT) generate informative images, with validation performance much higher than randomly selected original samples. However, as the IPC is continuously increased, the performance gap diminishes for ImageWoof, and random original images surpass the DiT-generated ones at the IPC of 100. On ImageIDC the intersection occurs even earlier at the IPC of 50. The main reason is reflected in Fig. 3, where the sample diversity remains limited without external guidance. The naive Difffit fine-tuning adapts the model to specific domains, yet on large IPCs, the over-fitted generative model still yields inferior performance than the original images." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 733, + 317, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 733, + 317, + 742 + ], + "spans": [ + { + "bbox": [ + 294, + 733, + 317, + 742 + ], + "type": "text", + "content": "15798" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 63, + 71, + 532, + 262 + ], + "blocks": [ + { + "bbox": [ + 63, + 71, + 532, + 262 + ], + "lines": [ + { + "bbox": [ + 63, + 71, + 532, + 262 + ], + "spans": [ + { + "bbox": [ + 63, + 71, + 532, + 262 + ], + "type": "image", + "image_path": "4333973511500f8d9b9119d19e8564fce33e79c0b54d97b7940598d07f3635a1.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 271, + 547, + 307 + ], + "lines": [ + { + "bbox": [ + 46, + 271, + 547, + 307 + ], + "spans": [ + { + "bbox": [ + 46, + 271, + 547, + 307 + ], + "type": "text", + "content": "Figure 5. Visualization of random original images, images generated by baseline diffusion models (DiT [33] and Difffit [50]) and our proposed method. For each column, the generated images are based on the same random seed. Comparatively, our method significantly enhances the coverage of original data distribution and the diversity of the surrogate dataset." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 48, + 366, + 287, + 479 + ], + "blocks": [ + { + "bbox": [ + 47, + 324, + 288, + 357 + ], + "lines": [ + { + "bbox": [ + 47, + 324, + 288, + 357 + ], + "spans": [ + { + "bbox": [ + 47, + 324, + 288, + 357 + ], + "type": "text", + "content": "Table 5. The ablation study of the proposed minimax scheme. The result are obtained with ResNet-10 on ImageWoof and ImageIDC. m denotes the proposed minimax optimization scheme." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 48, + 366, + 287, + 479 + ], + "lines": [ + { + "bbox": [ + 48, + 366, + 287, + 479 + ], + "spans": [ + { + "bbox": [ + 48, + 366, + 287, + 479 + ], + "type": "table", + "html": "
\\( {\\mathcal{L}}_{r} \\)\\( {\\mathcal{L}}_{r} \\) w\\m\\( {\\mathcal{L}}_{d} \\)\\( {\\mathcal{L}}_{d} \\) w\\mImageWoofImageIDC
10-IPC50-IPC10-IPC50-IPC
----35.6±0.951.0±0.953.5±0.266.3±0.2
✓---34.4±1.147.1±0.549.6±0.760.2±1.2
-✓--37.4±0.449.5±1.054.5±1.265.0±0.8
--✓-35.7±0.848.3±0.651.5±0.664.8±0.8
---✓38.7±0.954.9±0.752.2±0.668.4±0.7
✓-✓-38.3±0.554.9±0.453.3±0.566.8±0.5
-✓-✓39.2±1.356.3±1.053.1±0.269.6±0.2
", + "image_path": "f4ff12dcfd24b2eea5fbc6c410c218b769b893939bb578d4f299ad79154c9144.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 506, + 287, + 674 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 506, + 287, + 674 + ], + "spans": [ + { + "bbox": [ + 46, + 506, + 287, + 674 + ], + "type": "text", + "content": "The addition of representativeness constraint to the training process further enhances the effect of distribution fitting. At small IPCs, the generated images contain richer information, yet for larger IPCs, the lack of diversity brings a negative influence. The diversity constraint, in contrast, significantly boosts the information contained in the generated surrogate dataset. Despite the performance advantage of " + }, + { + "bbox": [ + 46, + 506, + 287, + 674 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_d" + }, + { + "bbox": [ + 46, + 506, + 287, + 674 + ], + "type": "text", + "content": " over " + }, + { + "bbox": [ + 46, + 506, + 287, + 674 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_r" + }, + { + "bbox": [ + 46, + 506, + 287, + 674 + ], + "type": "text", + "content": ", combining them still brings stable improvement as our full method. Especially on the easier ImageIDC task, grouping these two constraints together contributes to a consistent performance margin over random original images. The experimental results validate that both representativeness and diversity play essential parts in constructing effective surrogate datasets." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 677, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 288, + 715 + ], + "type": "text", + "content": "Minimax Scheme. In this work, we propose to enhance the representativeness and diversity each with a minimax objective. We compare the distillation result with or with-" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 326, + 547, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 326, + 547, + 445 + ], + "spans": [ + { + "bbox": [ + 304, + 326, + 547, + 445 + ], + "type": "text", + "content": "out the minimax operation in Tab. 5. The first row presents the performance of naive Difffit fine-tuning. Matching the embeddings to the distribution center as in Eq. (2) severely degrades the validation performance across all IPCs. In contrast, the minimax version constraint as in Eq. (3) encourages better coverage, where the performance on small IPCs is improved. The effects of diversity constraint and the full method show similar trends. The superior performance suggests the effectiveness in enhancing the essential properties of the generative diffusion techniques." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 454, + 389, + 466 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 454, + 389, + 466 + ], + "spans": [ + { + "bbox": [ + 306, + 454, + 389, + 466 + ], + "type": "text", + "content": "4.5. Visualization" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 473, + 545, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 473, + 545, + 665 + ], + "spans": [ + { + "bbox": [ + 304, + 473, + 545, + 665 + ], + "type": "text", + "content": "Sample Distribution Visualization. The target of our proposed method is to construct a surrogate dataset with both representativeness and diversity. We visualize the t-SNE distribution of samples generated by our proposed method in Fig. 3. In comparison with random original images and baseline diffusion models, our method demonstrates a more thorough coverage over the entire data distribution while maintaining consistency in sample density. At the original high-density region, the generated images also form a dense sub-cluster, which is not reflected by random sampling. On the other hand, at the original sparse regions, our method exhibits better coverage than baseline diffusion models. By simultaneously enhancing the representativeness and diversity in the generative model, the proposed method manages to significantly improve the validation performance of the generated surrogate dataset." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "type": "text", + "content": "Generated Sample Comparison. The proposed method notably enhances the representativeness and diversity of the generated surrogate dataset. We compare the samples generated with the same random noise (for each column) of" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 733, + 317, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 733, + 317, + 742 + ], + "spans": [ + { + "bbox": [ + 294, + 733, + 317, + 742 + ], + "type": "text", + "content": "15799" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 72, + 167, + 202 + ], + "blocks": [ + { + "bbox": [ + 50, + 72, + 167, + 202 + ], + "lines": [ + { + "bbox": [ + 50, + 72, + 167, + 202 + ], + "spans": [ + { + "bbox": [ + 50, + 72, + 167, + 202 + ], + "type": "image", + "image_path": "e9068d1bbb0364dd4adf568a30d710136b29428e2397cf17d941585ee0bb2151.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 212, + 545, + 235 + ], + "lines": [ + { + "bbox": [ + 46, + 212, + 545, + 235 + ], + "spans": [ + { + "bbox": [ + 46, + 212, + 545, + 235 + ], + "type": "text", + "content": "Figure 6. Hyper-parameter analysis on (a) the training epochs; (b) the representativeness weight " + }, + { + "bbox": [ + 46, + 212, + 545, + 235 + ], + "type": "inline_equation", + "content": "\\lambda_r" + }, + { + "bbox": [ + 46, + 212, + 545, + 235 + ], + "type": "text", + "content": "; (c) the diversity weight " + }, + { + "bbox": [ + 46, + 212, + 545, + 235 + ], + "type": "inline_equation", + "content": "\\lambda_d" + }, + { + "bbox": [ + 46, + 212, + 545, + 235 + ], + "type": "text", + "content": "; (d) the memory size " + }, + { + "bbox": [ + 46, + 212, + 545, + 235 + ], + "type": "inline_equation", + "content": "N_M" + }, + { + "bbox": [ + 46, + 212, + 545, + 235 + ], + "type": "text", + "content": ". The results are obtained with ResNetAP-10 on ImageWoof. The dashed line indicates the value adopted in this work." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 171, + 72, + 287, + 202 + ], + "blocks": [ + { + "bbox": [ + 171, + 72, + 287, + 202 + ], + "lines": [ + { + "bbox": [ + 171, + 72, + 287, + 202 + ], + "spans": [ + { + "bbox": [ + 171, + 72, + 287, + 202 + ], + "type": "image", + "image_path": "7dc98ca70bc837ad08622068ccb37cb43da590746fc6ae5e1decc544f50697b2.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 292, + 72, + 408, + 202 + ], + "blocks": [ + { + "bbox": [ + 292, + 72, + 408, + 202 + ], + "lines": [ + { + "bbox": [ + 292, + 72, + 408, + 202 + ], + "spans": [ + { + "bbox": [ + 292, + 72, + 408, + 202 + ], + "type": "image", + "image_path": "14db85d19c673fcc64aa8565f40e91bd574b85beb356d2d625ce97bc30380486.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 414, + 72, + 531, + 202 + ], + "blocks": [ + { + "bbox": [ + 414, + 72, + 531, + 202 + ], + "lines": [ + { + "bbox": [ + 414, + 72, + 531, + 202 + ], + "spans": [ + { + "bbox": [ + 414, + 72, + 531, + 202 + ], + "type": "image", + "image_path": "4c03a4a19a2973f631050f9ce5731e8da4d0ee1d105d7c548928f6b181327eab.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 255, + 286, + 278 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 255, + 286, + 278 + ], + "spans": [ + { + "bbox": [ + 46, + 255, + 286, + 278 + ], + "type": "text", + "content": "different generative methods in Fig. 5 to explicitly demonstrate the improved properties." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 279, + 286, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 279, + 286, + 518 + ], + "spans": [ + { + "bbox": [ + 46, + 279, + 286, + 518 + ], + "type": "text", + "content": "The images generated by baseline DiT exhibit a realistic high-quality appearance. However, the images tend to share similar poses and only present the most prominent features of the objects. In the golden retriever case, the generated images mostly present the head part, while for the churches the exterior appearance. Difffit fine-tuning further fits the model to the distribution, but in most cases, the differences only lie in small details. Comparatively, the proposed minimax criteria significantly enhance both the representativeness and diversity of the generated images. On the one hand, there occurs more class-related content in the generated images. The golden retriever images include more body parts and the church images encompass the interior layout. The minimax optimization leads to better coverage over the entire original distribution, with more related features encapsulated. On the other hand, the diversity is significantly enhanced, including variations in pose, background, and appearance styles. In such a way the surrogate dataset better represents the original large-scale one, leading to superior validation performance." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 519, + 286, + 626 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 519, + 286, + 626 + ], + "spans": [ + { + "bbox": [ + 46, + 519, + 286, + 626 + ], + "type": "text", + "content": "Training Curve Visualization. We visualize the accuracy curve during the training process in Fig. 6a. The validation performance is rapidly improved as the fine-tuning process starts. After four epochs, the model tends to converge and reaches the highest performance at the 8th epoch. Further extending the training epochs injects excessive diversity into the model, leading to performance degradation. We demonstrate the influence of training epochs on the generated images in supplementary material." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 635, + 162, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 635, + 162, + 647 + ], + "spans": [ + { + "bbox": [ + 47, + 635, + 162, + 647 + ], + "type": "text", + "content": "4.6. Parameter Analysis" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 653, + 286, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 286, + 712 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 286, + 712 + ], + "type": "text", + "content": "Objective Weight " + }, + { + "bbox": [ + 46, + 653, + 286, + 712 + ], + "type": "inline_equation", + "content": "\\lambda_r\\lambda_d" + }, + { + "bbox": [ + 46, + 653, + 286, + 712 + ], + "type": "text", + "content": " .We show the influence of representativeness weight " + }, + { + "bbox": [ + 46, + 653, + 286, + 712 + ], + "type": "inline_equation", + "content": "\\lambda_r" + }, + { + "bbox": [ + 46, + 653, + 286, + 712 + ], + "type": "text", + "content": " and diversity weight " + }, + { + "bbox": [ + 46, + 653, + 286, + 712 + ], + "type": "inline_equation", + "content": "\\lambda_{d}" + }, + { + "bbox": [ + 46, + 653, + 286, + 712 + ], + "type": "text", + "content": " in Fig. 6b and Fig. 6c, respectively. The " + }, + { + "bbox": [ + 46, + 653, + 286, + 712 + ], + "type": "inline_equation", + "content": "\\lambda_r" + }, + { + "bbox": [ + 46, + 653, + 286, + 712 + ], + "type": "text", + "content": " variation only produces negligible performance fluctuation on small IPCs, while on large IPCs the performance is also relatively stable. For" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 255, + 545, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 255, + 545, + 350 + ], + "spans": [ + { + "bbox": [ + 304, + 255, + 545, + 350 + ], + "type": "inline_equation", + "content": "\\lambda_{d}" + }, + { + "bbox": [ + 304, + 255, + 545, + 350 + ], + "type": "text", + "content": ", at a proper variation range, the performance is stable. However, continuously increasing the diversity of the generated dataset leads to a lack of representativeness, which results in a negative impact. The negative impact of overdiversity can also validated by the poor performance of K-Center in Tab. 1. A uniform performance decrease is observed as " + }, + { + "bbox": [ + 304, + 255, + 545, + 350 + ], + "type": "inline_equation", + "content": "\\lambda_{d}" + }, + { + "bbox": [ + 304, + 255, + 545, + 350 + ], + "type": "text", + "content": " reaches 0.03. Based on the performance of 100 IPC, we set " + }, + { + "bbox": [ + 304, + 255, + 545, + 350 + ], + "type": "inline_equation", + "content": "\\lambda_{r}" + }, + { + "bbox": [ + 304, + 255, + 545, + 350 + ], + "type": "text", + "content": " as 0.002 and " + }, + { + "bbox": [ + 304, + 255, + 545, + 350 + ], + "type": "inline_equation", + "content": "\\lambda_{d}" + }, + { + "bbox": [ + 304, + 255, + 545, + 350 + ], + "type": "text", + "content": " as 0.008." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 352, + 545, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 352, + 545, + 495 + ], + "spans": [ + { + "bbox": [ + 304, + 352, + 545, + 495 + ], + "type": "text", + "content": "Memory Size " + }, + { + "bbox": [ + 304, + 352, + 545, + 495 + ], + "type": "inline_equation", + "content": "N_{M}" + }, + { + "bbox": [ + 304, + 352, + 545, + 495 + ], + "type": "text", + "content": ". The memory size " + }, + { + "bbox": [ + 304, + 352, + 545, + 495 + ], + "type": "inline_equation", + "content": "N_{M}" + }, + { + "bbox": [ + 304, + 352, + 545, + 495 + ], + "type": "text", + "content": " influences the number of samples involved in the objective calculation. We investigate its influence in Fig. 6d. When the memory is extremely small (" + }, + { + "bbox": [ + 304, + 352, + 545, + 495 + ], + "type": "inline_equation", + "content": "N_{M} = 16" + }, + { + "bbox": [ + 304, + 352, + 545, + 495 + ], + "type": "text", + "content": "), the provided supervision is also limited, yet the performance is already higher than naive fine-tuning. As the memory size is increased in a proper range, the model yields stable performance improvement. It is notable that with a larger memory, the performance under the IPC of 10 is better. It can be explained by that a larger memory contains more representative information. Out of the consideration of performance as well as storage burden, we select the memory of 64 in the other experiments." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 510, + 378, + 522 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 510, + 378, + 522 + ], + "spans": [ + { + "bbox": [ + 306, + 510, + 378, + 522 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 531, + 545, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 531, + 545, + 651 + ], + "spans": [ + { + "bbox": [ + 304, + 531, + 545, + 651 + ], + "type": "text", + "content": "In this work, we propose a novel dataset distillation method based on generative diffusion techniques. Through extra minimax criteria, the proposed method significantly enhances the representativeness and diversity of the generated surrogate dataset. With much less computational time consumption, the proposed method achieves state-of-the-art validation performance on challenging ImageNet subsets. It reduces the resource dependency of previous dataset distillation methods and opens up new possibilities for more practical applications for distilling personalized data." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 652, + 545, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 652, + 545, + 688 + ], + "spans": [ + { + "bbox": [ + 304, + 652, + 545, + 688 + ], + "type": "text", + "content": "Limitations and Future Works. In addition to object classification, we will explore the possibility of incorporating generative techniques for more specific data domains." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "content": "Acknowledgement. We would like to specially thank Olga Russakovsky for invaluable advice on the manuscript." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 733, + 317, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 733, + 317, + 742 + ], + "spans": [ + { + "bbox": [ + 294, + 733, + 317, + 742 + ], + "type": "text", + "content": "15800" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 714 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 139 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 139 + ], + "type": "text", + "content": "[1] Shekoofeh Azizi, Simon Kornblith, Chitwan Sahara, Mohammad Norouzi, and David J Fleet. Synthetic data from diffusion models improves imagenet classification. arXiv preprint arXiv:2304.08466, 2023. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 140, + 288, + 188 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 140, + 288, + 188 + ], + "spans": [ + { + "bbox": [ + 53, + 140, + 288, + 188 + ], + "type": "text", + "content": "[2] Victor Besnier, Himalaya Jain, Andrei Bursuc, Matthieu Cord, and Patrick Pérez. This Dataset Does Not Exist: Training Models from Generated Images. In ICASSP, pages 1-5, 2020. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 190, + 288, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 190, + 288, + 237 + ], + "spans": [ + { + "bbox": [ + 54, + 190, + 288, + 237 + ], + "type": "text", + "content": "[3] George Cazenavette, Tongzhou Wang, Antonio Torralba, Alexei A Efros, and Jun-Yan Zhu. Dataset distillation by matching training trajectories. In CVPR, pages 4750-4759, 2022. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 238, + 288, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 238, + 288, + 285 + ], + "spans": [ + { + "bbox": [ + 54, + 238, + 288, + 285 + ], + "type": "text", + "content": "[4] George Cazenavette, Tongzhou Wang, Antonio Torralba, Alexei A Efros, and Jun-Yan Zhu. Generalizing dataset distillation via deep generative prior. In CVPR, pages 3739-3748, 2023. 1, 5, 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 286, + 287, + 322 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 286, + 287, + 322 + ], + "spans": [ + { + "bbox": [ + 54, + 286, + 287, + 322 + ], + "type": "text", + "content": "[5] Justin Cui, Ruochen Wang, Si Si, and Cho-Jui Hsieh. Dc-bench: Dataset condensation benchmark. NeurIPS, 35:810-822, 2022. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 323, + 287, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 323, + 287, + 370 + ], + "spans": [ + { + "bbox": [ + 54, + 323, + 287, + 370 + ], + "type": "text", + "content": "[6] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. ImageNet: A large-scale hierarchical image database. In CVPR, pages 248-255, 2009. 1, 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 54, + 373, + 287, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 373, + 287, + 408 + ], + "spans": [ + { + "bbox": [ + 54, + 373, + 287, + 408 + ], + "type": "text", + "content": "[7] Zhiwei Deng and Olga Russakovsky. Remember the Past: Distilling Datasets into Addressable Memories for Neural Networks. In NeurIPS, 2022. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 54, + 410, + 287, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 410, + 287, + 445 + ], + "spans": [ + { + "bbox": [ + 54, + 410, + 287, + 445 + ], + "type": "text", + "content": "[8] Prafulla Dhariwal and Alexander Nichol. Diffusion Models Beat GANs on Image Synthesis. In NeurIPS, pages 8780-8794, 2021. 3, 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 54, + 447, + 287, + 493 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 447, + 287, + 493 + ], + "spans": [ + { + "bbox": [ + 54, + 447, + 287, + 493 + ], + "type": "text", + "content": "[9] Thinh T Doan. Nonlinear two-time-scale stochastic approximation convergence and finite-time performance. IEEE Transactions on Automatic Control, 2022. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 495, + 287, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 495, + 287, + 578 + ], + "spans": [ + { + "bbox": [ + 48, + 495, + 287, + 578 + ], + "type": "text", + "content": "[10] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale. In ICLR, 2022. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 580, + 287, + 628 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 580, + 287, + 628 + ], + "spans": [ + { + "bbox": [ + 48, + 580, + 287, + 628 + ], + "type": "text", + "content": "[11] Jiawei Du, Yidi Jiang, Vincent Y.F. Tan, Joey Tianyi Zhou, and Haizhou Li. Minimizing the Accumulated Trajectory Error to Improve Dataset Distillation. In CVPR, pages 3749-3758, 2023. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 628, + 287, + 687 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 628, + 287, + 687 + ], + "spans": [ + { + "bbox": [ + 48, + 628, + 287, + 687 + ], + "type": "text", + "content": "[12] Lisa Dunlap, Alyssa Umino, Han Zhang, Jiezhi Yang, Joseph E Gonzalez, and Trevor Darrell. Diversify your vision datasets with automatic diffusion-based augmentation. arXiv preprint arXiv:2305.16289, 2023. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 689, + 287, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 689, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 48, + 689, + 287, + 714 + ], + "type": "text", + "content": "[13] Gabriele Eichfelder. Scalarizations for adaptively solving multi-objective optimization problems. Com-" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 547, + 713 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "type": "text", + "content": "putational Optimization and Applications, 44:249-273, 2009. 4" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 98, + 547, + 144 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 98, + 547, + 144 + ], + "spans": [ + { + "bbox": [ + 308, + 98, + 547, + 144 + ], + "type": "text", + "content": "[14] Ronen Eldan and James R Lee. Regularization under diffusion and anticoncentration of the information content. Duke Mathematical Journal, 167(5):969-993, 2018. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 145, + 545, + 180 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 145, + 545, + 180 + ], + "spans": [ + { + "bbox": [ + 308, + 145, + 545, + 180 + ], + "type": "text", + "content": "[15] Fastai. Fastai/imagenette: A smaller subset of 10 easily classified classes from imagenet, and a little more french. 1, 5, 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 182, + 547, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 182, + 547, + 228 + ], + "spans": [ + { + "bbox": [ + 308, + 182, + 547, + 228 + ], + "type": "text", + "content": "[16] Jianyang Gu, Kai Wang, Wei Jiang, and Yang You. Summarizing stream data for memory-restricted online continual learning. arXiv preprint arXiv:2305.16645, 2023. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 230, + 545, + 277 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 230, + 545, + 277 + ], + "spans": [ + { + "bbox": [ + 308, + 230, + 545, + 277 + ], + "type": "text", + "content": "[17] Swaminathan Gurumurthy, Ravi Kiran Sarvadevabhatla, and R. Venkatesh Babu. DeLiGAN: Generative Adversarial Networks for Diverse and Limited Data. In CVPR, pages 4941-4949, 2017. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 279, + 545, + 314 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 279, + 545, + 314 + ], + "spans": [ + { + "bbox": [ + 308, + 279, + 545, + 314 + ], + "type": "text", + "content": "[18] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, pages 770-778, 2016. 1, 5, 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 316, + 545, + 361 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 316, + 545, + 361 + ], + "spans": [ + { + "bbox": [ + 308, + 316, + 545, + 361 + ], + "type": "text", + "content": "[19] Ruifei He, Shuyang Sun, Xin Yu, Chuhui Xue, Wenqing Zhang, Philip Torr, Song Bai, and Xiaojuan Qi. Is Synthetic Data from Generative Models Ready for Image Recognition? In ICLR, 2022. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 363, + 545, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 363, + 545, + 398 + ], + "spans": [ + { + "bbox": [ + 308, + 363, + 545, + 398 + ], + "type": "text", + "content": "[20] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising Diffusion Probabilistic Models. In NeurIPS, pages 6840-6851, 2020. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 399, + 545, + 458 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 399, + 545, + 458 + ], + "spans": [ + { + "bbox": [ + 308, + 399, + 545, + 458 + ], + "type": "text", + "content": "[21] Jang-Hyun Kim, Jinuk Kim, Seong Joon Oh, Sang-doo Yun, Hwanjun Song, Joonhyun Jeong, Jung-Woo Ha, and Hyun Oh Song. Dataset condensation via efficient synthetic-data parameterization. In ICML, pages 11102-11118, 2022. 1, 5, 6, 2, 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 460, + 545, + 495 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 460, + 545, + 495 + ], + "spans": [ + { + "bbox": [ + 308, + 460, + 545, + 495 + ], + "type": "text", + "content": "[22] Diederik Kingma, Tim Salimans, Ben Poole, and Jonathan Ho. Variational Diffusion Models. In NeurIPS, pages 21696-21707, 2021. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 497, + 545, + 543 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 497, + 545, + 543 + ], + "spans": [ + { + "bbox": [ + 308, + 497, + 545, + 543 + ], + "type": "text", + "content": "[23] Saehyung Lee, Sanghyuk Chun, Sangwon Jung, Sang-doo Yun, and Sungroh Yoon. Dataset condensation with contrastive signals. In ICML, pages 12352-12364, 2022. 1" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 544, + 545, + 581 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 544, + 545, + 581 + ], + "spans": [ + { + "bbox": [ + 308, + 544, + 545, + 581 + ], + "type": "text", + "content": "[24] Joseph Lehec. Representation formula for the entropy and functional inequalities. In Annales de l'IHP Probabilités et statistiques, pages 885-899, 2013. 1" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 582, + 545, + 616 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 582, + 545, + 616 + ], + "spans": [ + { + "bbox": [ + 308, + 582, + 545, + 616 + ], + "type": "text", + "content": "[25] Ping Liu, Xin Yu, and Joey Tianyi Zhou. Meta Knowledge Condensation for Federated Learning. In ICLR, 2022. 1" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 617, + 545, + 664 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 617, + 545, + 664 + ], + "spans": [ + { + "bbox": [ + 308, + 617, + 545, + 664 + ], + "type": "text", + "content": "[26] Yanqing Liu, Jianyang Gu, Kai Wang, Zheng Zhu, Wei Jiang, and Yang You. Dream: Efficient dataset distillation by representative matching. In ICCV, 2023. 1, 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 666, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 666, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 666, + 545, + 713 + ], + "type": "text", + "content": "[27] Noel Loo, Ramin Hasani, Alexander Amini, and Daniela Rus. Efficient dataset distillation using random feature approximation. NeurIPS, 35:13877-13891, 2022. 1" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 734, + 317, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 734, + 317, + 742 + ], + "spans": [ + { + "bbox": [ + 294, + 734, + 317, + 742 + ], + "type": "text", + "content": "15801" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 119 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 119 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 119 + ], + "type": "text", + "content": "[28] Noel Loo, Ramin Hasani, Mathias Lechner, and Daniela Rus. Dataset distillation with convexified implicit gradients. arXiv preprint arXiv:2302.06755, 2023.1" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 120, + 288, + 168 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 120, + 288, + 168 + ], + "spans": [ + { + "bbox": [ + 48, + 120, + 288, + 168 + ], + "type": "text", + "content": "[29] Qi Mao, Hsin-Ying Lee, Hung-Yu Tseng, Siwei Ma, and Ming-Hsuan Yang. Mode Seeking Generative Adversarial Networks for Diverse Image Synthesis. In CVPR, pages 1429-1437, 2019. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 168, + 287, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 168, + 287, + 205 + ], + "spans": [ + { + "bbox": [ + 48, + 168, + 287, + 205 + ], + "type": "text", + "content": "[30] Timothy Nguyen, Zhourong Chen, and Jaehoon Lee. Dataset meta-learning from kernel ridge-regression. In ICLR, 2021. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 205, + 287, + 253 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 205, + 287, + 253 + ], + "spans": [ + { + "bbox": [ + 48, + 205, + 287, + 253 + ], + "type": "text", + "content": "[31] Timothy Nguyen, Roman Novak, Lechao Xiao, and Jaehoon Lee. Dataset distillation with infinitely wide convolutional networks. NeurIPS, 34:5186-5198, 2021. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 254, + 287, + 290 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 254, + 287, + 290 + ], + "spans": [ + { + "bbox": [ + 48, + 254, + 287, + 290 + ], + "type": "text", + "content": "[32] Alexander Quinn Nichol and Prafulla Dhariwal. Improved Denoising Diffusion Probabilistic Models. In ICML, pages 8162-8171, 2021. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 290, + 287, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 290, + 287, + 326 + ], + "spans": [ + { + "bbox": [ + 48, + 290, + 287, + 326 + ], + "type": "text", + "content": "[33] William Peebles and Saining Xie. Scalable diffusion models with transformers. In ICCV, pages 4195-4205, 2023. 3, 5, 6, 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 327, + 287, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 327, + 287, + 387 + ], + "spans": [ + { + "bbox": [ + 48, + 327, + 287, + 387 + ], + "type": "text", + "content": "[34] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In CVPR, pages 22500-22510, 2023. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 388, + 287, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 388, + 287, + 435 + ], + "spans": [ + { + "bbox": [ + 48, + 388, + 287, + 435 + ], + "type": "text", + "content": "[35] Ahmad Sajedi, Samir Khaki, Ehsan Amjadian, Lucy Z Liu, Yuri A Lawryshyn, and Konstantinos N Plataniotis. DataDAM: Efficient Dataset Distillation with Attention Matching. In ICCV, 2023. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 435, + 287, + 495 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 435, + 287, + 495 + ], + "spans": [ + { + "bbox": [ + 48, + 435, + 287, + 495 + ], + "type": "text", + "content": "[36] Mert Bulent Sariyildiz, Karteek Alahari, Diane Larlus, and Yannis Kalantidis. Fake it Till You Make it: Learning Transferable Representations from Synthetic ImageNet Clones. In CVPR, pages 8011-8021, 2023. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 496, + 287, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 496, + 287, + 544 + ], + "spans": [ + { + "bbox": [ + 48, + 496, + 287, + 544 + ], + "type": "text", + "content": "[37] Ozan Sener and Silvio Savarese. Active learning for convolutional neural networks: A core-set approach. In International Conference on Learning Representations, 2018. 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 544, + 287, + 592 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 544, + 287, + 592 + ], + "spans": [ + { + "bbox": [ + 48, + 544, + 287, + 592 + ], + "type": "text", + "content": "[38] Allahkaram Shafiei, Vyacheslav Kungurtsev, and Jakub Marecek. Trilevel and multilevel optimization using monotone operator theory. arXiv preprint arXiv:2105.09407, 2021. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 593, + 287, + 652 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 593, + 287, + 652 + ], + "spans": [ + { + "bbox": [ + 48, + 593, + 287, + 652 + ], + "type": "text", + "content": "[39] Jordan Shipard, Arnold Wiliem, Kien Nguyen Thanh, Wei Xiang, and Clinton Fookes. Diversity Is Definitely Needed: Improving Model-Agnostic Zero-Shot Classification via Stable Diffusion. In CVPR, pages 769-778, 2023. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 653, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 653, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 653, + 287, + 713 + ], + "type": "text", + "content": "[40] Felipe Petroski Such, Aditya Rawal, Joel Lehman, Kenneth Stanley, and Jeffrey Clune. Generative Teaching Networks: Accelerating Neural Architecture Search by Learning to Generate Synthetic Training Data. In ICML, pages 9206-9216, 2020. 1" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 72, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 307, + 72, + 545, + 108 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 72, + 545, + 108 + ], + "spans": [ + { + "bbox": [ + 307, + 72, + 545, + 108 + ], + "type": "text", + "content": "[41] Peng Sun, Bei Shi, Daiwei Yu, and Tao Lin. On the diversity and realism of distilled dataset: An efficient dataset distillation paradigm. In CVPR, 2024. 6" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 110, + 545, + 145 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 110, + 545, + 145 + ], + "spans": [ + { + "bbox": [ + 307, + 110, + 545, + 145 + ], + "type": "text", + "content": "[42] Yonglong Tian, Dilip Krishnan, and Phillip Isola. Contrastive multiview coding. In ECCV, pages 776-794, 2020. 5" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 148, + 545, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 148, + 545, + 196 + ], + "spans": [ + { + "bbox": [ + 308, + 148, + 545, + 196 + ], + "type": "text", + "content": "[43] Belinda Tzen and Maxim Raginsky. Theoretical guarantees for sampling and inference in generative models with latent diffusions. In " + }, + { + "bbox": [ + 308, + 148, + 545, + 196 + ], + "type": "inline_equation", + "content": "COLT" + }, + { + "bbox": [ + 308, + 148, + 545, + 196 + ], + "type": "text", + "content": ", pages 3084-3114, 2019. 4, 5, 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 198, + 545, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 198, + 545, + 258 + ], + "spans": [ + { + "bbox": [ + 308, + 198, + 545, + 258 + ], + "type": "text", + "content": "[44] Saeed Vahidian, Mingyu Wang, Jianyang Gu, Vyacheslav Kungurtsev, Wei Jiang, and Yiran Chen. Group distributionally robust dataset distillation with risk minimization. arXiv preprint arXiv:2402.04676, 2024.1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 261, + 545, + 320 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 261, + 545, + 320 + ], + "spans": [ + { + "bbox": [ + 308, + 261, + 545, + 320 + ], + "type": "text", + "content": "[45] Kai Wang, Bo Zhao, Xiangyu Peng, Zheng Zhu, Shuo Yang, Shuo Wang, Guan Huang, Hakan Bilen, Xin chao Wang, and Yang You. Cafe: Learning to condense dataset by aligning features. In CVPR, pages 12196-12205, 2022. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 323, + 545, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 323, + 545, + 370 + ], + "spans": [ + { + "bbox": [ + 308, + 323, + 545, + 370 + ], + "type": "text", + "content": "[46] Kai Wang, Jianyang Gu, Daquan Zhou, Zheng Zhu, Wei Jiang, and Yang You. Dim: Distilling dataset into generative model. arXiv preprint arXiv:2303.04707, 2023.1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 373, + 545, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 373, + 545, + 409 + ], + "spans": [ + { + "bbox": [ + 308, + 373, + 545, + 409 + ], + "type": "text", + "content": "[47] Tongzhou Wang, Jun-Yan Zhu, Antonio Torralba, and Alexei A Efros. Dataset distillation. arXiv preprint arXiv:1811.10959, 2018. 1, 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 411, + 545, + 436 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 411, + 545, + 436 + ], + "spans": [ + { + "bbox": [ + 308, + 411, + 545, + 436 + ], + "type": "text", + "content": "[48] Max Welling. Herding dynamical weights to learn. In ICML, pages 1121-1128, 2009. 5, 6, 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 437, + 545, + 474 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 437, + 545, + 474 + ], + "spans": [ + { + "bbox": [ + 308, + 437, + 545, + 474 + ], + "type": "text", + "content": "[49] Xindi Wu, Zhiwei Deng, and Olga Russakovsky. Multimodal dataset distillation for image-text retrieval. arXiv preprint arXiv:2308.07545, 2023. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 476, + 545, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 476, + 545, + 536 + ], + "spans": [ + { + "bbox": [ + 308, + 476, + 545, + 536 + ], + "type": "text", + "content": "[50] Enze Xie, Lewei Yao, Han Shi, Zhili Liu, Daquan Zhou, Zhaoqiang Liu, Jiawei Li, and Zhenguo Li. Difffit: Unlocking transferability of large diffusion models via simple parameter-efficient fine-tuning. arXiv preprint arXiv:2304.06648, 2023. 3, 5, 7, 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 539, + 545, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 539, + 545, + 586 + ], + "spans": [ + { + "bbox": [ + 308, + 539, + 545, + 586 + ], + "type": "text", + "content": "[51] Yuanhao Xiong, Ruochen Wang, Minhao Cheng, Felix Yu, and Cho-Jui Hsieh. FedDM: Iterative Distribution Matching for Communication-Efficient Federated Learning. In CVPR, pages 16323-16332, 2023. 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 589, + 545, + 636 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 589, + 545, + 636 + ], + "spans": [ + { + "bbox": [ + 308, + 589, + 545, + 636 + ], + "type": "text", + "content": "[52] Dingdong Yang, Seunghoon Hong, Yunseok Jang, Tianchen Zhao, and Honglak Lee. Diversity-Sensitive Conditional Generative Adversarial Networks. In ICLR, 2018. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 639, + 545, + 675 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 639, + 545, + 675 + ], + "spans": [ + { + "bbox": [ + 308, + 639, + 545, + 675 + ], + "type": "text", + "content": "[53] Zeyuan Yin, Eric Xing, and Zhiqiang Shen. Squeeze, recover and relabel: Dataset condensation atImagenet scale from a new perspective. In NeurIPS, 2023. 6" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 677, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 677, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 677, + 545, + 713 + ], + "type": "text", + "content": "[54] Bo Zhao and Hakan Bilen. Dataset condensation with differentiable siamese augmentation. In ICML, pages 12674-12685, 2021. 1, 2" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 734, + 318, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 734, + 318, + 742 + ], + "spans": [ + { + "bbox": [ + 294, + 734, + 318, + 742 + ], + "type": "text", + "content": "15802" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 303 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "type": "text", + "content": "[55] Bo Zhao and Hakan Bilen. Synthesizing informative training samples with gan. arXiv preprint arXiv:2204.07513, 2022. 1" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 109, + 287, + 144 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 109, + 287, + 144 + ], + "spans": [ + { + "bbox": [ + 49, + 109, + 287, + 144 + ], + "type": "text", + "content": "[56] Bo Zhao and Hakan Bilen. Dataset condensation with distribution matching. In WACV, pages 6514-6523, 2023. 2, 5, 6, 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 146, + 287, + 181 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 146, + 287, + 181 + ], + "spans": [ + { + "bbox": [ + 49, + 146, + 287, + 181 + ], + "type": "text", + "content": "[57] Bo Zhao, Konda Reddy Mopuri, and Hakan Bilen. Dataset condensation with gradient matching. In ICLR, 2021. 1, 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 183, + 287, + 218 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 183, + 287, + 218 + ], + "spans": [ + { + "bbox": [ + 49, + 183, + 287, + 218 + ], + "type": "text", + "content": "[58] Ganlong Zhao, Guanbin Li, Yipeng Qin, and Yizhou Yu. Improved Distribution Matching for Dataset Condensation. In CVPR, pages 7856-7865, 2023. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 49, + 220, + 286, + 255 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 220, + 286, + 255 + ], + "spans": [ + { + "bbox": [ + 49, + 220, + 286, + 255 + ], + "type": "text", + "content": "[59] Yongchao Zhou, Ehsan Nezhadarya, and Jimmy Ba. Dataset distillation using neural feature regression. NeurIPS, 35:9813-9827, 2022. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 49, + 257, + 286, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 257, + 286, + 303 + ], + "spans": [ + { + "bbox": [ + 49, + 257, + 286, + 303 + ], + "type": "text", + "content": "[60] Yongchao Zhou, Hshmat Sahak, and Jimmy Ba. Training on thin air: Improve image classification with generated data. arXiv preprint arXiv:2305.15316, 2023.2" + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 734, + 317, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 734, + 317, + 742 + ], + "spans": [ + { + "bbox": [ + 294, + 734, + 317, + 742 + ], + "type": "text", + "content": "15803" + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Efficient Deformable ConvNets_ Rethinking Dynamic and Sparse Operator for Vision Applications/d839f1b7-f6aa-4faa-bb5a-7c661a9aa4e5_content_list.json b/2024/Efficient Deformable ConvNets_ Rethinking Dynamic and Sparse Operator for Vision Applications/d839f1b7-f6aa-4faa-bb5a-7c661a9aa4e5_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..0d3ddc3cdea1d6d2f0dd9efdff12782f3325a251 --- /dev/null +++ b/2024/Efficient Deformable ConvNets_ Rethinking Dynamic and Sparse Operator for Vision Applications/d839f1b7-f6aa-4faa-bb5a-7c661a9aa4e5_content_list.json @@ -0,0 +1,1478 @@ +[ + { + "type": "text", + "text": "Efficient Deformable ConvNets: Rethinking Dynamic and Sparse Operator for Vision Applications", + "text_level": 1, + "bbox": [ + 156, + 128, + 816, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yuwen Xiong $^{*,1,2}$ Zhiqi Li $^{*,3,2}$ Yuntao Chen $^{*4}$ Feng Wang $^{*5}$", + "bbox": [ + 223, + 200, + 745, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Xizhou Zhu $^{5,6}$ Jiapeng Luo $^{6}$ Wenhai Wang $^{7,2}$ Tong Lu $^{3}$ Hongsheng Li $^{7}$", + "bbox": [ + 163, + 220, + 803, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yu Qiao² Lewei Lu⁶ Jie Zhou⁵ Jifeng Dai⁵,²", + "bbox": [ + 276, + 238, + 694, + 255 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ University of Toronto $^{2}$ OpenGVLab, Shanghai AI Laboratory", + "bbox": [ + 207, + 272, + 761, + 290 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{3}$ Nanjing University $^{4}$ CAIR, HKISI, CAS $^{5}$ Tsinghua University", + "bbox": [ + 174, + 290, + 792, + 306 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{6}$ SenseTime Research $^{7}$ The Chinese University of Hong Kong", + "bbox": [ + 212, + 306, + 754, + 325 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "https://github.com/OpenGVLab/DCNv4", + "bbox": [ + 279, + 325, + 681, + 342 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 264, + 377, + 341, + 393 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We introduce Deformable Convolution v4 (DCNv4), a highly efficient and effective operator designed for a broad spectrum of vision applications. DCNv4 addresses the limitations of its predecessor, DCNv3, with two key enhancements: 1. removing softmax normalization in spatial aggregation to enhance its dynamic property and expressive power and 2. optimizing memory access to minimize redundant operations for speedup. These improvements result in a significantly faster convergence compared to DCNv3 and a substantial increase in processing speed, with DCNv4 achieving more than three times the forward speed. DCNv4 demonstrates exceptional performance across various tasks, including image classification, instance and semantic segmentation, and notably, image generation. When integrated into generative models like U-Net in the latent diffusion model, DCNv4 outperforms its baseline, underscoring its possibility to enhance generative models. In practical applications, replacing DCNv3 with DCNv4 in the InternImage model to create FlashInternImage results in up to $80\\%$ speed increase and further performance improvement without further modifications. The advancements in speed and efficiency of DCNv4, combined with its robust performance across diverse vision tasks, show its potential as a foundational building block for future vision models.", + "bbox": [ + 73, + 409, + 473, + 770 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 790, + 207, + 806 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In the field of computer vision, there is an ongoing debate about whether convolutional networks (ConvNets) or Transformers offer superior performance. In recent years, Trans", + "bbox": [ + 75, + 816, + 470, + 861 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/0a203e8027418311ed674c2a129d80b85a2d8873b8ba87e04c7d45a298d9ee45.jpg", + "image_caption": [ + "(a)" + ], + "image_footnote": [], + "bbox": [ + 509, + 377, + 736, + 526 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/05016b33a235bfddc93728e971ce3b133882971c9880b5eec24bc9a4bd5b8091.jpg", + "image_caption": [ + "(b)", + "Figure 1. (a) We show relative runtime with DCNv3 as the baseline. DCNv4 shows significant speedup over DCNv3, and surpasses other common vision operators. (b) With the same network architecture, DCNv4 converges faster than other operators." + ], + "image_footnote": [], + "bbox": [ + 743, + 378, + 893, + 525 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "former models [12, 25, 44] have achieved remarkable results in large vision models with the attention mechanism, showing the potential to overtake ConvNets. However, recent works such as InternImage [38] and ConvNeXt [26] demonstrate that ConvNet-based vision models retain robust performance, efficiency, simplicity, and suitable inductive bias for various downstream tasks [15, 41]. Notably, in domains like image generation [29, 31], models containing both convolutions and transformers remain the preferred approach. This situation brings to light the enduring value of convolution-based approaches.", + "bbox": [ + 496, + 601, + 893, + 765 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Building on convolution's strengths, Deformable Convolution v3 (DCNv3) – the core operator of the advanced ConvNet model InternImage – innovatively combines a sparse attention mechanism with convolution: it processes each output location in a sliding window manner with a small window size (e.g. $3 \\times 3 = 9$ points) which acts as a local, sparse operator like convolution, while dynamically samples point with an adaptive range and aggregates the spatial features with input-dependent attention weights. With its", + "bbox": [ + 496, + 766, + 895, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Equal contribution", + "bbox": [ + 99, + 875, + 210, + 887 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "Corresponding author (daijifeng@tsinghua.edu.cn)", + "bbox": [ + 104, + 887, + 387, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "5652", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "small window size and ConvNet inductive bias, DCNv3 is expected to achieve a faster convergence rate and lower inference latency, especially when compared to dense global [12] or local window-based [25] attention methods.", + "bbox": [ + 75, + 90, + 470, + 150 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Despite these advantages, DCN has not become the go-to solution for vision backbone models. This observation led us to investigate the lingering limitations of the DCN operator. The first thing we notice is the running speed. The slow speed of DCN is known to be a long-standing problem [1], as it introduces extra overhead on sampling non-nearby locations, making it not fit modern convolution algorithms. Our comparative analysis, illustrated in Fig. 1a, reveals that DCNv3 can be slower than a properly optimized dense global attention [9], highlighting the need for further optimization. Moreover, we find DCNv3 even converges slower than global attention at the initial backbone training phase, as shown in Fig. 1b, which is counter-intuitive as DCNv3 is equipped with ConvNet inductive bias.", + "bbox": [ + 75, + 151, + 470, + 359 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To overcome these challenges, we propose Deformable Convolution v4 (DCNv4), an innovative advancement to optimize the sparse DCN operator for practical efficiency. DCNv4 comes with a much faster implementation and an improved operator design to enhance its performance, which we will elaborate on as follows:", + "bbox": [ + 75, + 361, + 470, + 449 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "First, we conduct instruction-level kernel profiling for existing implementation and find that DCNv3 is already lightweight. The compute cost is less than $1\\%$ , while memory access costs $99\\%$ . This motivates us to revisit the operator implementation and find that many memory accesses in the DCN forward process are redundant and thus can be optimized, leading to a much faster DCNv4 implementation.", + "bbox": [ + 75, + 450, + 470, + 556 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Second, drawing inspiration from convolution's unbounded weight range, we find that softmax in spatial aggregation, a standard operation in dense attention, is unnecessary in DCNv3, as it is not a requirement for operators with a dedicated aggregation window for each location. Intuitively, softmax puts a bounded $0 \\sim 1$ value range to the weight and will limit the expressive power of the aggregation weight. This insight led us to remove the softmax in DCNv4, enhancing its dynamic property and improving its performance.", + "bbox": [ + 75, + 556, + 470, + 705 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "As a result, DCNv4 converges significantly faster than DCNv3 and accelerates forward speed by more than $3 \\times$ . This improvement allows DCNv4 to fully leverage its sparse property and become one of the fastest core vision operators.", + "bbox": [ + 75, + 705, + 470, + 765 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We further replace DCNv3 in InternImage with DCNv4, creating FlashInternImage. Remarkably, FlashInternImage achieves a $50 \\sim 80\\%$ speed increase compared to InternImage without any additional modifications. This enhancement positions FlashInternImage as one of the fastest modern vision backbone networks while maintaining superior performance. With the help of DCNv4, FlashInternImage significantly improves the convergence speed in ImageNet classification [10] and transfer learning settings and further", + "bbox": [ + 75, + 766, + 470, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "demonstrates improved performance in downstream tasks.", + "bbox": [ + 498, + 90, + 883, + 106 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Furthermore, DCNv4 shows potential as a universal vision operator in various architectures and tasks. We integrate DCNv4 into other modern backbone architectures, including ConvNeXt [26] and ViT [12], replacing depthwise convolution [6] and dense self-attention layers [35]. Surprisingly, without any hyperparameter adjustments, these meticulously designed networks with DCNv4 perform on par while being much faster, showing the efficacy and efficiency of the dynamic, sparse DCNv4. Moreover, we explore the potential of DCNv4 in generative models as a new application domain. Specifically, we apply it in the U-Net [30] architecture used in diffusion models [29], replacing regular convolution with DCNv4. Our experimental results show that DCNv4 can work better than the baselines in image generation, showing great potential for using DCN to improve generative models.", + "bbox": [ + 496, + 107, + 893, + 330 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We have released the DCNv4 implementation to facilitate future research in the vision community.", + "bbox": [ + 498, + 330, + 893, + 361 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 368, + 640, + 383 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Core operators in vision models: The standard convolution [17] stands as the most prevalent and impactful operator, forming the backbone of the majority of computer vision architectures [14, 16, 32]. Nevertheless, a myriad of operators, each with unique characteristics, collectively play a crucial role in the development of computer vision. Depthwise separable convolution (DWConv) [6] separates the spatial and channel operations, and has been pivotal in developing lightweight and efficient models [26, 27]. RepLKNet [11] illustrates that a purely convolutional network, leveraging large-kernel depth-wise convolutions, can attain competitive performance in both efficiency and effectiveness. Deformable Convolution (DCN) series [7, 38, 47] significantly leaps the adaptability of convolution by adding learnable offsets to the convolutions kernels. Contrary to convolutions, attention mechanisms [35] possess the capacity to model long-range dependencies and have been successfully adopted in various computer vision tasks [3, 12, 24, 33]. Window attention [25, 36] reduces the computational complexity inherent in vanilla attention by restricting the attention operation to a fixed-size window. To mitigate the high computational complexity associated with vanilla attention, deformable attention [48] enables each query to concentrate on a select number of key sampling points, with dynamically determined locations and weights. This efficient method is widely used in the following arts perception methods [4, 19, 21, 22, 43, 45]. DynamicConv [40] and dynamic-DWNet [13] augment DWConv by incorporating dynamic weights, thereby enabling the use of instance-specific weights that adapt dynamically. For non-grid structured data, sparse operators [34, 37, 42] utilize dynamic weights obtained via bilinear interpolation or in a parametric way.", + "bbox": [ + 496, + 386, + 895, + 866 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Memory access cost in vision backbones: As underscored in previous studies [18, 27], FLOPs, although a fre", + "bbox": [ + 498, + 869, + 895, + 901 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "5653", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "table", + "img_path": "images/edb85164031ac5b3d3aa413971ac9aaf509c9ce20fbd33644dea91dbe04f7f34.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Model5th EP10th Ep20th Ep50th Ep300th Ep
ConvNeXt29.953.566.174.883.8
ConvNeXt + softmax8.5 (-21.4)25.3 (-28.2)51.1 (-15.0)69.1 (-5.7)81.6 (-2.2)
", + "bbox": [ + 81, + 88, + 467, + 143 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Table 1. ImageNet-1K accuracy at different training epochs. Softmax on the convolution weights significantly affects the convergence speed and the final performance for the ConvNeXt model.", + "bbox": [ + 76, + 145, + 472, + 186 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "sequently used metric to measure model complexity, do not accurately represent the model's speed or latency. In practical scenarios, the running speed of a model is influenced by multiple factors, not just FLOPs. Memory access cost plays a significant role in this context [27]. Flash-Attention [9], by reducing the number of accesses to High Bandwidth Memory (HBM), achieves a significantly faster speed in practice despite having higher FLOPs compared to vanilla attention. Although DCN operators do not exhibit a disadvantage in terms of FLOPs, their latency is considerably longer compared to DW-Conv, under the same FLOPs budget, predominantly due to substantial memory access costs. In this work, we conduct a thorough analysis and optimization of the memory access costs associated with the DCN operators, significantly accelerating the DCN's running speed.", + "bbox": [ + 75, + 191, + 473, + 417 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 76, + 426, + 168, + 441 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Rethinking the Dynamic Property in Deformable Convolution", + "text_level": 1, + "bbox": [ + 76, + 450, + 468, + 481 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Revisiting DCNv3: Given an input $\\mathbf{x} \\in \\mathbb{R}^{H \\times W \\times C}$ with height $H$ , width $W$ and channel $C$ , the DCNv3 operation with $K$ points is defined in Eq. (2) for each point $p_0$ :", + "bbox": [ + 76, + 488, + 470, + 536 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {y} _ {g} = \\sum_ {k = 1} ^ {K} \\mathbf {m} _ {g k} \\mathbf {x} _ {g} \\left(p _ {0} + p _ {k} + \\Delta p _ {g k}\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 142, + 542, + 468, + 584 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {y} = \\operatorname {c o n c a t} \\left(\\left[ \\mathbf {y} _ {1}, \\mathbf {y} _ {2}, \\dots , \\mathbf {y} _ {G} \\right], \\text {a x i s} = - 1\\right), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 151, + 587, + 468, + 606 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $G$ denotes the number of spatial aggregation groups. For the $g$ -th group, $\\mathbf{x}_g,\\mathbf{y}_g\\in \\mathbb{R}^{H\\times W\\times C'}$ represents the sliced input/output feature map with $C' = C / G$ represents the group dimension; $\\mathbf{m}_{gk}\\in \\mathbb{R}$ denotes the spatial aggregation weights (also known as modulation scalar) of the $k$ -th sampling point in the $g$ -th group, conditioned on the input $\\mathbf{x}$ and normalized by the softmax function along the dimension $K$ ; $p_k$ denotes the $k$ -th location of the pre-defined grid sampling $\\{(-1, - 1),(-1,0),\\dots,(0, + 1),\\dots,(+1, + 1)\\}$ as in regular convolutions and $\\Delta p_{gk}$ is the offset corresponding to the grid sampling location $p_k$ in the $g$ -th group. A $1\\times 1$ point-wise convolution on $\\mathbf{x}$ and $\\mathbf{y}$ can be applied before and after the DCNv3 operator to enhance the model's expressive power, following separable convolution [6]. DCNv3 is a combination of convolution and attention: on the one hand, it processes the input data in a sliding window manner, which follows convolution and inherent its inductive bias; on the other hand, the sampling offset $\\Delta p$ and spatial aggregation weight $\\mathbf{m}$ are dynamically predicted from the input feature,", + "bbox": [ + 75, + 616, + 472, + 902 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/0da0a2416ed9ee5e5ba643bd4e37433e46ab5f66b82645ab00246b074a464e2e.jpg", + "image_caption": [ + "(a) Attention" + ], + "image_footnote": [], + "bbox": [ + 504, + 90, + 669, + 165 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/8a4cd00fb02996c23fda3a714e41a20b56e5df9ee73adb56f7d181101e90237f.jpg", + "image_caption": [ + "Figure 2. Comparisons of core operators in spatial aggregation for query pixels on different locations within the same channel. (a) Attention and (b) DCNv3 use bounded (range from $0 \\sim 1$ ) dynamic weights to aggregate spatial features, while the window (sampling point set) for attention is the same, and DCNv3 uses a dedicated window for each location. (c) Convolution has a more flexible unbounded value range for aggregation weights and uses a dedicated sliding window for each location, but the window shape and aggregation weights are input-independent. (d) DCNv4 combines their advantages, using an adaptive aggregation window and dynamic aggregation weights with an unbounded value range." + ], + "image_footnote": [], + "bbox": [ + 504, + 172, + 684, + 247 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/33e00d201f52315ef0a993c478a5b8dacac3f7c24d1bcef7ad3f7498f1b2da04.jpg", + "image_caption": [ + "(b) DCNv3" + ], + "image_footnote": [], + "bbox": [ + 694, + 90, + 870, + 166 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/a00f4168748400aa3f4bd5805a6d86f8fe61347d8863fa8041ffaf80c1424ece.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 694, + 172, + 888, + 247 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "showing its dynamic property and making it more like an attention mechanism. We compare different operators where each has its own property, as illustrated in Fig. 2", + "bbox": [ + 496, + 411, + 892, + 455 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Softmax normalization: A key difference between convolution and DCNv3 is that DCNv3 normalizes $\\mathbf{m}$ , the spatial aggregation weights, with a softmax function, following the convention of scaled dot-product self-attention. Conversely, convolution does not employ softmax over its weights and still works well. The reason why attention needs a softmax is straightforward: scaled dot-product self-attention with $Q, K, V \\in \\mathbb{R}^{N \\times d}$ is defined with a formulation:", + "bbox": [ + 496, + 460, + 893, + 580 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {s o f t m a x} \\left(\\frac {1}{\\sqrt {d}} Q K ^ {\\top}\\right) V, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 617, + 588, + 892, + 621 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $N$ is the number of points in the same attention window (can be either global [12] or local [25]), $d$ is the hidden dimension, $Q, K, V$ are the query, key, and value matrices computed from the input. Softmax operation is required in Eq. (3) for attention; without softmax, $K^{\\top}V \\in \\mathbb{R}^{d \\times d}$ can be calculated first, and it degrades to a linear projection for all queries in the same attention window, resulting in degenerated performance. However, for convolutional operators like depthwise convolution and DCNv3 where each point has its own dedicated aggregation window and the values in each aggregation window are already different and there is no \"key\" concept, such degradation issues no longer exist, and the normalization becomes unnecessary. In fact, normalizing convolution weights within a fixed 0-1 range using softmax could impose a significant limitation on the operator's expressive power and make the learning slower.", + "bbox": [ + 496, + 630, + 893, + 869 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To confirm this hypothesis, we train a ConvNeXt model and apply softmax to the $7 \\times 7$ window of the depthwise", + "bbox": [ + 498, + 869, + 893, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "5654", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "convolution weights before convolution forward. We observe a remarkable decline in model performance as well as convergence speed from results in Tab. 1. This suggests that for operators with a dedicated aggregation window on each location like convolution or DCN, aggregation weights with an unbounded range offer better expressive power than softmax-normalized, bounded-range weights.", + "bbox": [ + 75, + 90, + 472, + 196 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Enhancing dynamic property: Motivated by this observation, we remove the softmax normalization in DCNv3, transforming the modulation scalars ranging from 0 to 1 to unbounded dynamic weights similar to convolution. As shown in Fig. 2, this alteration further amplifies the dynamic property of DCN, where other operators have certain limits, such as bounded value range (attention/DCNv3) or fixed aggregation window with input-independent aggregation weights (convolution). Fig. 1b shows that by making this change, DCNv4 converges significantly faster than DCNv3 and other common operators, including convolution and attention. Results in Sec. 4 further showcase that DCNv4 works well in both pre-training and transfer learning settings.", + "bbox": [ + 75, + 203, + 473, + 398 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Speeding up DCN", + "text_level": 1, + "bbox": [ + 76, + 402, + 251, + 419 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Theoretically, DCN, as a sparse operator with $3 \\times 3$ window, should act faster than other common operators that employ larger window sizes, like dense attention or $7 \\times 7$ depthwise convolution. However, we find that this is not the case, as shown in Fig. 1a. In this subsection, we first conduct a theoretical analysis of GPU efficiency, showing a large variance in memory access cost depending on how we read the memory. We further perform optimization based on our observations, significantly improving the speed of DCN by saving additional memory instruction and bringing the speed advantage of being a sparse operator into reality.", + "bbox": [ + 75, + 424, + 472, + 590 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Theoretical analysis of GPU efficiency Our study begins with a theoretical examination of the DCNv3 operator's computational behavior. We employ the roofline model to evaluate its performance, focusing on theoretical FLOPs and memory access cost. For an input and output tensor of shape $(H,W,C)$ , the DCNv3 operator requires $36HWC$ FLOPs, where $3\\times 3$ represents the convolution kernel's spatial dimensions and the factor of 4 accounts for the bilinear interpolation at each sampling point.", + "bbox": [ + 75, + 599, + 470, + 734 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Following the framework outlined in [27], DCNv3's memory access cost is calculated as $2HWC + 27HWG$ . The first term corresponds to the input/output feature map size and the second to the DCNv3's offset and aggregation weights with $G$ groups. We approximate $G$ as $C / 16$ assuming a group dimension of 16, resulting in approximately $3.7HWC$ memory access cost. However, this assumes an ideal scenario of infinite cache and a single memory read for each value, which is often unrealistic in parallel computing environments where concurrent thread execution necessitates simultaneous data access.", + "bbox": [ + 75, + 736, + 472, + 900 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/a17fa7dfce35b4a9b16f016b29f5033fa5c616e4468c2f79f26bc2189c681b73.jpg", + "image_caption": [ + "Figure 3. Illustration of our optimization. In DCNv4, we use one thread to process multiple channels in the same group that shares sampling offset and aggregation weights. Workloads like memory reading and bilinear interpolation coefficient computation can be reduced, and multiple memory access instructions can be merged." + ], + "image_footnote": [], + "bbox": [ + 501, + 90, + 893, + 218 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To estimate the maximum memory access requirement, we consider a scenario devoid of cache, where each output location requires fresh memory reads and involves 36 reads for bilinear interpolation, 27 for offset/aggregation weights, and one write operation, resulting in a memory access cost of $64HWC$ . This is 17 times larger than the ideal case.", + "bbox": [ + 496, + 301, + 893, + 391 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "This analysis reveals a substantial gap in the ratio of computation-to-memory access (ranging from 0.6 to 9.7), highlighting the significant potential for memory access optimization. Notably, despite DCNv3's use of input-dependent, dynamic offsets causing non-deterministic memory access, one deterministic thing is that channels within the same group share offset values. This leads us to propose a specific optimization strategy for enhancing DCNv3's speed.", + "bbox": [ + 496, + 391, + 893, + 512 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Eliminating redundant workload: In previous CUDA implementations of DCN kernel, for input with shape $(H,W,C)^1$ , offset $(H,W,G,K^2\\times 2)$ and aggregation weight $(H,W,G,K^2)$ , we will create $H\\times W\\times C$ threads in total to maximize parallelism, where each thread processes one channel for one output location. Notably, the $D = C / G$ channels within each group share the same sampling offset and aggregation weight values for each output location. Using multiple threads to process these $D$ channels on the same output location is wasteful as different threads will read the same sampling offset and aggregation weight values from GPU memory multiple times, which is critical for a memory-bound operator. Processing multiple channels within the same group on each output location with one thread can eliminate these redundant memory read requests, greatly reducing memory bandwidth usage. As the sampling locations are the same, we can also only calculate the bilinear interpolation coefficient used in DCN once. Specifically, if each thread processes $D^{\\prime}$ channels, the memory access cost for reading offset and aggregation weight, as well as the computation cost for calculating bilinear interpolation coefficient, can be reduced $D^{\\prime}$ times.", + "bbox": [ + 496, + 513, + 895, + 843 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Eliminating redundant memory instructions: In practice, solely reusing threads for multiple channels will not", + "bbox": [ + 498, + 847, + 893, + 877 + ], + "page_idx": 3 + }, + { + "type": "page_footnote", + "text": "1We assume the batch size is 1 and memory layout is channel-last.", + "bbox": [ + 517, + 886, + 867, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "5655", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/de434fa665229f69d642a31d1748b375d77c1ff611ac29f7a61e2f08bbfe600d.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
OperatorRuntime (ms)
56 × 56 × 12828 × 28 × 25614 × 14 × 5127 × 7 × 102414 × 14 × 768
Attention (torch)30.8 / 19.33.35 / 2.120.539 / 0.4480.446 / 0.1210.779 / 0.654
FlashAttention-2N/A / 2.46N/A / 0.451N/A / 0.123N/A / 0.0901N/A / 0.163
Window Attn (7 × 7)4.05 / 1.462.07 / 0.7701.08 / 0.4220.577 / 0.2391.58 / 0.604
DWConv (7 × 7, torch)2.02 / 1.981.03 / 1.000.515 / 0.5230.269 / 0.2610.779 / 0.773
DWConv (7 × 7, cuDNN)0.981 / 0.4380.522 / 0.2670.287 / 0.1530.199 / 0.1020.413 / 0.210
DCNv31.45 / 1.520.688 / 0.7110.294 / 0.2980.125 / 0.1260.528 / 0.548
DCNv40.606 / 0.4040.303 / 0.2300.145 / 0.1230.0730 / 0.06800.224 / 0.147
", + "bbox": [ + 119, + 88, + 846, + 222 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Table 2. Op-level benchmark on standard input shape with various downsample rates. FP32/FP16 results are reported when the implementation is available. Our new DCNv4 can surpass all other commonly used operators under different input resolutions.", + "bbox": [ + 75, + 223, + 890, + 250 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "see speed improvement. The reason is that when $D'$ increases, we create fewer threads and the workload of each thread now increases $D'$ times. This essentially reduces the degree of parallelism for the CUDA kernel. Luckily, the DCN kernel is now computationally lightweight as the bilinear interpolation only needs to be performed once for all $D'$ channels, and most of the workload is the memory instructions reading input values from different channels. When the memory layout is channel-last, and all $D'$ channel values are contiguous, we can leverage vectorized load: for example, to read four 32-bit float values from memory, instead of reading one 32-bit float value four times with four instructions, we can use a single instruction to load a 128-bit packed value at once, thus reducing the number of instructions and execution time of each thread. We can apply similar technique when writing the final results to GPU memory, minimizing the memory access time and increasing memory bandwidth utilization. Moreover, the modern half-precision data format (float16/bfloat16) halves the bytes that need to be loaded, which means the memory efficiency can be twice as much under the same memory bandwidth when using the half-precision format. However, we do not see speed improvement with half-precision data in the original DCNv3 implementation, possibly due to too much overhead on data access and computation, while in our new implementation, the speedup is significant. It is worth noting that the aforementioned optimization techniques can also be applied to DCNv1/v2 and deformable attention [48], as they share a similar performance bottleneck and issue.", + "bbox": [ + 73, + 258, + 472, + 691 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Micro design in DCN module: DCNv3 module introduces multiple micro designs; as the core kernel is optimized, their impact on the speed becomes non-negligible. We identify two points in DCNv3 designs that could be further optimized: first, after removing the softmax and transforming the modulation scalar into dynamic aggregation weights as mentioned in the previous paragraph. The linear layers for computing offset and dynamic weights can actually be combined into one linear layer. This reduces network fragmentation and eliminates extra overheads, such as kernel launching and synchronization, enhancing run-time efficiency on the GPU; second, in the original DCNv3 module design, a complex sub-network that consists of depthwise $3 \\times 3$ conv, layer", + "bbox": [ + 75, + 705, + 470, + 902 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "normalization (LN), GELU, and linear layer is used to compute offsets and dynamic weights. Following the design in Xception [6], we remove the additional LN-GELU layers and use the original separable convolution structure, further reducing running time. We empirically find that if latency is a higher priority, the depthwise convolution can also be removed with only a minor performance sacrifice.", + "bbox": [ + 496, + 258, + 893, + 364 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 498, + 368, + 633, + 386 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this section, we verify the effectiveness of our proposed DCNv4 module from both speed as well as performance perspective. We benchmark the operator-level speed and integrate DCNv4 into the backbone to test the system-level performance. All speed test results are obtained with an NVIDIA A100 80G SXM GPU. We include additional experimental results and implementation details in supp.", + "bbox": [ + 496, + 393, + 893, + 501 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Speed Benchmark for Operators", + "text_level": 1, + "bbox": [ + 498, + 506, + 785, + 523 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Settings: We conduct the op-level benchmark by solely measuring the running time of several representative operators building state-of-the-art vision backbone models, including full attention [35] implemented with PyTorch as well as the advanced FlashAttention-2 [8] implementation, window-based attention with window size $7 \\times 7$ [25], depthwise convolution with $7 \\times 7$ window, implemented by cuDNN [5] and ATen library from PyTorch [28], respectively. We only benchmark the core operation for spatial aggregation, and additional linear layers like qkv projection and output projection layers are excluded and not included in the runtime measurement. Please refer to supp. for a more comprehensive module-level comparison. We first consider a feature map shape generated from the standard $224 \\times 224$ input resolution for image classification with 4, 8, 16, $32 \\times$ downsample ratio as used by common hierarchical ConvNet/transformer backbones; we also add a feature map shape from isotropic backbone like ViT with a downsampling ratio 16 and larger hidden dimension. We further consider high-resolution inputs often used in downstream tasks like object detection. We set the input shape to be $800 \\times 1280$ and $1024 \\times 1024$ for the hierarchical feature map and isotropic feature map, respectively, as they are the common practice in object detection [15, 20]. Batch size is 64 and 1 for these two input sets, respectively. For operators with a head/group concept, we", + "bbox": [ + 496, + 526, + 893, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5656", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/29932da64e9391ff28d8aa823af265a03abd49cfeaaea5f0526aebc12874c911.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
OperatorRuntime (ms)
200 × 320 × 128100 × 160 × 25650 × 80 × 51225 × 40 × 102464 × 64 × 768
Attention (torch)OOM / OOM25.4 / 12.92.88 / 1.890.490 / 0.3094.17 / 2.57
FlashAttention-2N/A / 13.2N/A / 1.74N/A / 0.285N/A / 0.0797N/A / 0.437
Window Attn (7 × 7)1.33 / 0.5090.728 / 0.2910.426 / 0.1860.279 / 0.1650.673 / 0.272
DWConv (7 × 7, torch)0.634 / 0.6080.313 / 0.3150.167 / 0.1580.0943 / 0.08940.260 / 0.253
DWConv (7 × 7, cuDNN)0.331 / 0.2820.188 / 0.1680.114 / 0.1150.0817 / 0.08810.161 / 0.156
DCNv30.472 / 0.4930.244 / 0.2530.128 / 0.1320.0737 / 0.07670.194 / 0.199
DCNv40.210 / 0.1360.124 / 0.08950.0707 / 0.05890.0452 / 0.04260.103 / 0.0672
", + "bbox": [ + 122, + 88, + 848, + 218 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "set the dimension of each head/group to 32 and change the number of heads/groups when the hidden dimension varies.", + "bbox": [ + 75, + 263, + 470, + 294 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Results: We show the benchmark results on standard resolution and high-resolution input in Tab. 2 and Tab. 3. We report results with both FP32 and FP16 data formats unless the FP32 implementation is not available. Dense global attention implemented with PyTorch performs significantly slower when the input resolution is large and even out of memory. FlashAttention significantly improves the speed of attention and can be even faster than $7 \\times 7$ window attention in certain cases, indicating the importance of proper optimization. However, it does not change the quadratic complexity of attention; when the input resolution is high, it still falls behind local/sparse operators like window attention or convolution. While DCNv3 can be faster than DWConv with plain implementation, it is slower than the heavily optimized cuDNN version. Instead, our DCNv4 can provide more than $3 \\times$ speedup compared to DCNv3, greatly saving the running time. Moreover, DCNv4 can leverage the advantage of using a $3 \\times 3$ sparse window to perform significantly faster than other baselines under different settings.", + "bbox": [ + 75, + 296, + 473, + 582 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Image Classification", + "text_level": 1, + "bbox": [ + 76, + 588, + 269, + 604 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Settings: We evaluate the effectiveness of DCNv4 on ImageNet classification. We start from InternImage [38] as it shows state-of-the-art performance among ConvNet-based models. We replace the DCNv3 in InternImage with DCNv4 and create FlashInternImage. Other implementation details, including network architecture and hyperparameters, are kept the same as [38]. We compare Swin-Transformer and ConvNeXt which are two representative baselines in Transformer and ConvNet models. We follow the common practice [25, 26, 38] of training protocols, including data augmentation, preprocessing, optimizer and learning rate schedule, and train FlashInternImage-Tiny/Small/Base on ImageNet-1K for 300 epochs. FlashInternImage-Large is trained on ImageNet-22K for 90 epochs and then fine-tuned on ImageNet-1K for 20 epochs. Other baselines share the same setting for a fair comparison.", + "bbox": [ + 75, + 608, + 470, + 848 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Results: Tab. 4 shows the results of models at various scales. Besides the model size and training/inference resolution, we also report each model's overall throughput (number", + "bbox": [ + 75, + 854, + 472, + 902 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/24ed477e96994e9d152d8492f9d83aee63b62dcf71c0ad293622162e099c84b0.jpg", + "table_caption": [ + "Table 3. Op-level benchmark on high-resolution input shape with various downsample rates. DCNv4 performs well as a sparse operator, surpassing all other baselines, while dense global attention is slow under this scenario." + ], + "table_footnote": [], + "table_body": "
ModelSizeScaleAccThroughput
Swin-T29M224281.31989 / 3619
ConvNeXt-T29M224282.12485 / 4305
InternImage-T30M224283.51409 / 1746
FlashInternImage-T30M224283.62316 / 3154 (+64%/ +80%)
Swin-S50M224283.01167 / 2000
ConvNeXt-S50M224283.11645 / 2538
InternImage-S50M224284.21044 / 1321
FlashInternImage-S50M224284.41625 / 2396 (+56%/ +81%)
Swin-B88M224283.5934 / 1741
ConvNeXt-B89M224283.81241 / 1888
InternImage-B97M224284.9779 / 1030
FlashInternImage-B97M224284.91174 / 1816 (+51%/ +76%)
Swin-L197M384287.3206 / 301
ConvNeXt-L198M384287.5252 / 436
InternImage-L223M384287.7158 / 214
ConvNeXt-XL350M384287.8170 / 299
InternImage-XL335M384288.0125 / 174
FlashInternImage-L223M384288.1248 / 401 (+57%/ +87%)
", + "bbox": [ + 526, + 261, + 867, + 578 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 4. Image classification performance on ImageNet-1K. We show relative speedup between FlashInternImage w/ DCNv4 and its InternImage counterparts. DCNv4 significantly improves the speed while shows state-of-the-art performance.", + "bbox": [ + 496, + 582, + 893, + 638 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "of images per second) in FP32/FP16 data format. We use timm [39] implementation of ConvNeXt and Swin Transformer, which is faster than the original implementation. Equipped with DCNv4, FlashInternImage significantly improves the throughput of the InternImage counterpart over $50\\% \\sim 80\\%$ and slightly improves the model performance. FlashInternImage now matches the speed of ConvNeXt with higher accuracy. It is noteworthy that FlashInternImage-S can outperform ConvNeXt-B (84.4% vs. 83.8%) while being faster than it, showing a better speed-accuracy trade-off. Moreover, the FlashInternImage-L can surpass ConvNeXt-XL and InternImage-XL and being $30\\% \\sim 130\\%$ (401 vs. 174) faster, showing the effectiveness of our DCNv4 module.", + "bbox": [ + 496, + 650, + 895, + 845 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Downstream Tasks with High-Resolution Input We evaluate the performance of DCNv4 on representative downstream perception tasks with high-resolution input, in", + "bbox": [ + 496, + 852, + 893, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "5657", + "bbox": [ + 482, + 944, + 514, + 957 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/0f4a17346f2f07a4ee070d79acb1ee2d90f69189c3fbc7d019b675c9bf64c9b6.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Model#paramFPSMask R-CNN
1×3×+MS
APbAPmAPbAPm
Swin-T48M66 / 10642.739.346.041.6
ConvNeXt-T48M78 / 11344.240.146.241.7
InternImage-T49M54 / 6947.242.549.143.7
FlashInternImage-T49M72 / 10248.043.149.544.0
Swin-S69M45 / 7744.840.948.243.2
ConvNeXt-S70M54 / 8345.441.847.942.9
InternImage-S69M44 / 5647.843.349.744.5
FlashInternImage-S69M57 / 8349.244.050.544.9
Swin-B107M33 / 5946.942.348.643.3
ConvNeXt-B108M43 / 7047.042.748.543.5
InternImage-B115M33 / 4348.844.050.344.8
FlashInternImage-B115M44 / 6750.144.550.645.4
", + "bbox": [ + 84, + 89, + 464, + 287 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/ee8565477bd4bdd618637c0b5da4e85d51d3c712a39e22a87d495ed1930601be.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Model#paramFPSCascade Mask R-CNN
1×3×+MS
APbAPmAPbAPm
Swin-L253M20 / 2651.844.953.946.7
ConvNeXt-L255M26 / 4053.546.454.847.6
InternImage-L277M20 / 2654.947.756.148.5
ConvNeXt-XL407M21 / 3253.646.555.247.7
InternImage-XL387M16 / 2355.348.156.248.8
FlashInternImage-L277M26 / 3955.648.256.748.9
", + "bbox": [ + 84, + 299, + 464, + 419 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "cluding instance segmentation, semantic segmentation and 3D object detection. We keep all implementation details the same as InternImage and only change the backbone model. The backbone models are initialized from the ImageNet pretrained weights when training the downstream models.", + "bbox": [ + 76, + 534, + 470, + 609 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Instance Segmentation: We train FlashInternImage with two representative instance segmentation frameworks, Mask R-CNN [15] and Cascade Mask-RCNN [2], on COCO dataset [23] at $1 \\times (12$ epochs) and $3 \\times (36$ epochs) training schedules. The results are shown in Tab. 5. We also report FPS with batch size 16 in FP32/FP16 data format. FlashInternImage shows superior results on all model scales and training schedules, achieving a higher speed-accuracy tradeoff. For example, FlashInternImage-T/S surpasses all other models with the same scale and is on par with a larger InternImage-S/B while being $80\\% - 90\\%$ faster.", + "bbox": [ + 75, + 619, + 470, + 785 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Semantic Segmentation: We train FlashInternImage with UperNet [41] on ADE20K [46] dataset for 160K iterations. We can draw a similar conclusion as instance segmentation from the results in Tab. 6, with FPS reported with batch size 16 in FP32/FP16. FlashInternImage w/ DCNv4 can achieve significantly faster speed and further improve the performance of InternImage across different model scales, resulting in a new state-of-the-art.", + "bbox": [ + 75, + 791, + 470, + 912 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/e9447db2f16d972055e89f5aa24871f3076499b6addcfc248cae1a230890b0ab.jpg", + "table_caption": [ + "Table 5. Object detection and instance segmentation performance on COCO val2017. $\\mathrm{AP^b}$ and $\\mathrm{AP^m}$ denotes box AP and mask AP, respectively. \"MS\" means multi-scale training. FlashInternImage w/ DCNv4 models converge faster, clearly outperform other baselines with $1\\times$ training schedule, and still maintain a leading position when training $3\\times$ longer while being significantly faster than InternImage." + ], + "table_footnote": [], + "table_body": "
Modelcrop size#paramFPSmIoU (SS)mIoU (MS)
Swin-T512260M107 / 16844.545.8
ConvNeXt-T512260M120 / 18446.046.7
InternImage-T512259M100 / 13947.948.1
FlashInternImage-T512259M119 / 20649.350.3
Swin-S512281M89 / 14247.649.5
ConvNeXt-S512282M107 / 16448.749.6
InternImage-S512280M89 / 12350.150.9
FlashInternImage-S512280M107 / 18250.651.6
Swin-B5122121M77 / 12648.149.7
ConvNeXt-B5122122M95 / 14749.149.9
InternImage-B5122128M77 / 10450.851.3
FlashInternImage-B5122128M94 / 15752.052.6
Swin-L6402234M59 / 9952.153.5
ConvNeXt-L6402235M73 / 11753.253.7
InternImage-L6402256M56 / 7853.954.1
ConvNeXt-XL6402391M53 / 7553.654.0
InternImage-XL6402368M47 / 6755.055.3
FlashInternImage-L6402256M71 / 12255.656.0
", + "bbox": [ + 516, + 89, + 877, + 358 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/f38102fea077724c866ca67aa14c746b9cf552d65a0e517f5793e5938af9eb88.jpg", + "table_caption": [ + "Table 6. Semantic segmentation performance on the ADE20K validation set. All models are trained with UperNet. \"SS\" and \"MS\" denote single-scale and multi-scale testing, respectively. FPS is reported with single-scale testing. FlashInternImage w/ DCNv4 achieves superior performance with competitive speed." + ], + "table_footnote": [], + "table_body": "
ModelNDSmAPFPS†FPS
InternImage-B62.054.08.02.7
InternImage-XL63.455.64.02.0
FlashInternImage-S61.755.516.84.1
FlashInternImage-B63.157.412.13.8
", + "bbox": [ + 540, + 436, + 846, + 508 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 7. 3D detection performance of BEVFormer v2 on nuScenes test set. We report backbone FPS (denoted with $\\dagger$ ), overall FPS results with underoptimized head implementation are also added for reference. With on-par NDS and higher mAP results, FlashInternImage can be $50 - 90\\%$ faster than InternImage baselines or $200\\% - 300\\%$ when only considering the backbone.", + "bbox": [ + 496, + 512, + 893, + 595 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3D Detection: We further test DCNv4 on the camera-based 3D object detection task in the autonomous driving scenario, We train BEVFormer v2 [43], a state-of-the-art multi-camera 3D object detector, with FlashInternImage-Small and Base backbone models on the nuScenes dataset for 24 epochs. We report results on the nuScenes test set in Tab. 7 with FPS for each model. We note that the header parts, such as the BEV encoder and object decoder in BEV-Former v2, are underoptimized and take more than $50\\%$ of the running time (and even more with a faster backbone); thus, we also report the FPS for the backbone for a clearer illustration. Our results show that when only considering the backbone, FlashInternImage can be twice or even three times faster than the InternImage backbone with an on-par performance, greatly increasing the model efficiency.", + "bbox": [ + 496, + 612, + 893, + 837 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. DCNv4 as a Universal Operator", + "text_level": 1, + "bbox": [ + 500, + 847, + 782, + 864 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Drop-in replacement in other vision backbones : Weverify whether DCNv4 can still work well in architectures de-", + "bbox": [ + 500, + 869, + 893, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "5658", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/130974bdecf54d90cdddb3a1dfe3d5758068be568f78e5aeeabd497b8f5c0c7d.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelIN-1K AccThroughput
ConvNeXt-B83.81241 / 1888
ConvNeXt-B + DCNv484.01495 / 2513 (+20% / +33%)
ViT-B81.81683 / 2781†
ViT-B + DCNv481.92092 / 3261 (+24% / +17%)
", + "bbox": [ + 117, + 89, + 424, + 178 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/40fed1164be495c2d4b5d84bfea672f66785761adb63631f131c1e362e0704f9.jpg", + "table_caption": [ + "Table 8. DCNv4 in other architecture. We show supervised learning results on ImageNet-1K and throughput. DCNv4 achieves higher throughput with comparable accuracy. $\\dagger$ denotes testing with the advanced FlashAttention-2 implementation." + ], + "table_footnote": [], + "table_body": "
Model#paramFIDFPS
U-Net860M2.944.82
U-Net + DCNv4566M2.444.92
", + "bbox": [ + 119, + 246, + 424, + 301 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "signed with other operators, such as ConvNeXt and ViT. To achieve that, we replace the attention module and depthwise convolution layer with DCNv4 in ViT and ConvNeXt and perform supervised learning on ImageNet-1K without changing any other architecture and hyperparameters, similar to FlashInternImage and InternImage. The results are shown in Tab. 8. We can see that on these architectures, which are carefully tuned for the specific operators, our DCNv4 can perform equally well. Thanks to the fast speed of DCNv4, the new model can even achieve better throughput, showcasing the superior performance of DCNv4.", + "bbox": [ + 75, + 375, + 470, + 541 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Drop-in replacement in diffusion model: DCN has been recognized to be an effective operator for perception tasks. As generative models become a fundamental tool for AI-generated content (AIGC), we are also curious if DCNv4 can work well on generation tasks with diffusion-based generative models. Specifically, we choose the U-Net [30] used in Stable Diffusion [29] as our baselines and replace the attention module and regular $3 \\times 3$ convolution in U-Net. We use U-ViT's codebase, follow its training schedule, and train a latent diffusion model based on the image latent extracted from an image autoencoder provided by Stable Diffusion. We show the results in Tab. 9. We can see that DCNv4 also works well in generative modeling, achieving better results in terms of FID/Throughput with fewer parameters compared to regular convolution in U-Net. Notice that the architecture/hyperparameters may not be optimal for DCNv4, and it is possible that re-designing the models or searching for new hyperparameters for DCNv4 will give better results.", + "bbox": [ + 75, + 551, + 470, + 821 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.5. Ablation Study", + "text_level": 1, + "bbox": [ + 76, + 830, + 232, + 848 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We conduct ablation studies in our optimization choice described in Sec. 3.2. The results are shown in Tab. 10. The time in the table is obtained with $56 \\times 56 \\times 128$ input with", + "bbox": [ + 75, + 854, + 470, + 902 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/34393397696dbc313a2e1b9b9a7111411c5e47d979865320990896d17d34a300.jpg", + "table_caption": [ + "Table 9. Class conditional generation on ImageNet 256x256. Latent diffusion models are trained from scratch with U-Net. We replace the convolution layer in the models with DCNv4. DCNv4 can achieve better FID results without any hyperparameter tuning." + ], + "table_footnote": [], + "table_body": "
Implementation variantModuleKernel
Original DCNv33.281.45
- micro design2.121.45
- redundant memory access2.201.53
- redundant computation2.181.51
- redundant memory instr.1.280.606
- half-precision format0.8730.404
", + "bbox": [ + 532, + 89, + 857, + 199 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 10. Ablation studies of DCN's runtime (ms). We show how to achieve DCNv4 (gray row) from the original DCNv3 implementation and how different design choices affect the speed.", + "bbox": [ + 496, + 203, + 893, + 246 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "batch size 64 and 4 groups (32 channels per group). We first remove the softmax operation and improve the micro design, which means we merge the two linear layers into one and remove costly layer norm and GELU activation in offset/aggregation weight computing, simplifying the overall modules and increasing the speed. We then start modifying the kernel implementation. First, we change the parallel execution pattern and let each thread process 8 channels instead of 1 channel, thus, unnecessary memory access on loading sampling offset and aggregation weight values from the GPU memory can be saved. As we expected, solely applying this change will not increase the speed as the degree of parallelism decreases, and each thread's workload increases 8 times now. The latency is increased instead. Eliminating redundant computation by reusing the bilinear interpolation coefficient (4th row) can save some time but is insignificant. Removing the redundant memory instruction via vectorized load/store can greatly reduce the workload of each thread and largely accelerate the GPU kernel speed (5th row). Using a half-precision datatype, which halves the number of bytes the kernel needs to read/write, further increases the data throughput, as shown in the 6th row. In the end, we reach the final DCNv4 design, which is three times more efficient than the original implementation.", + "bbox": [ + 496, + 255, + 893, + 613 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 625, + 619, + 640 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We present Deformable Convolution v4 (DCNv4), an efficient dynamic and sparse operator. By rethinking the dynamic property in deformable convolution and streamlining memory access, DCNv4 emerges as a much faster and more effective operator than its predecessor DCNv3. DCNv4-equipped FlashInternImage backbone not only enhances speed but also improves performance across various vision tasks. We further show DCNv4's versatility and effectiveness as a universal operator by integrating it into state-of-the-art architecture like ConvNeXt and ViT with improved throughput and accuracy; and it also works well in latent diffusion model, showing its potential to enhance generative models.", + "bbox": [ + 496, + 646, + 893, + 827 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgement", + "text_level": 1, + "bbox": [ + 500, + 837, + 660, + 854 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The work is supported by the National Key R&D Program of China (NO. 2022ZD0161300), by the National Natural Science Foundation of China (62376134).", + "bbox": [ + 496, + 858, + 893, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "5659", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 119, + 173, + 135 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Saehyun Ahn, Jung-Woo Chang, and Suk-Ju Kang. An efficient accelerator design methodology for deformable convolutional networks. In 2020 IEEE International Conference on Image Processing (ICIP), pages 3075-3079. IEEE, 2020. 2", + "[2] Zhaowei Cai and Nuno Vasconcelos. Cascade r-cnn: Delving into high quality object detection. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6154-6162, 2018. 7", + "[3] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In European conference on computer vision, pages 213-229. Springer, 2020. 2", + "[4] Bowen Cheng, Ishan Misra, Alexander G Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1290–1299, 2022. 2", + "[5] Sharan Chetlur, Cliff Woolley, Philippe Vandermersch, Jonathan Cohen, John Tran, Bryan Catanzaro, and Evan Shelhamer. codnn: Efficient primitives for deep learning. arXiv preprint arXiv:1410.0759, 2014. 5", + "[6] François Chollet. Xception: Deep learning with depthwise separable convolutions. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1251-1258, 2017. 2, 3, 5", + "[7] Jifeng Dai, Haozhi Qi, Yuwen Xiong, Yi Li, Guodong Zhang, Han Hu, and Yichen Wei. Deformable convolutional networks. In Proceedings of the IEEE international conference on computer vision, pages 764-773, 2017. 2", + "[8] Tri Dao. Flashattention-2: Faster attention with better parallelism and work partitioning. arXiv preprint arXiv:2307.08691, 2023.5", + "[9] Tri Dao, Dan Fu, Stefano Ermon, Atri Rudra, and Christopher Ré. Flashattention: Fast and memory-efficient exact attention with io-awareness. Advances in Neural Information Processing Systems, 35:16344-16359, 2022. 2, 3", + "[10] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248–255. IEEE, 2009. 2", + "[11] Xiaohan Ding, Xiangyu Zhang, Jungong Han, and Guiguang Ding. Scaling up your kernels to 31x31: Revisiting large kernel design in cnns. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11963-11975, 2022. 2", + "[12] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations, 2020. 1, 2, 3", + "[13] Qi Han, Zejia Fan, Qi Dai, Lei Sun, Ming-Ming Cheng, Jiaying Liu, and Jingdong Wang. On the connection between" + ], + "bbox": [ + 78, + 143, + 470, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "local attention and dynamic depth-wise convolution. arXiv preprint arXiv:2106.04263, 2021. 2", + "[14] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 2", + "[15] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In Proceedings of the IEEE international conference on computer vision, 2017. 1, 5, 7", + "[16] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. ImageNet classification with deep convolutional neural networks. Advances in neural information processing systems, 2012. 2", + "[17] Yann LeCun, Léon Bottou, Yoshua Bengio, and Patrick Haffner. Gradient-based learning applied to document recognition. Proceedings of the IEEE, 86(11):2278-2324, 1998. 2", + "[18] Youngwan Lee, Joong-won Hwang, Sangrok Lee, Yuseok Bae, and Jongyoul Park. An energy andgpu-computation efficient backbone network for real-time object detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops, pages 0–0, 2019. 2", + "[19] Feng Li, Hao Zhang, Huaizhe Xu, Shilong Liu, Lei Zhang, Lionel M Ni, and Heung-Yeung Shum. Mask dino: Towards a unified transformer-based framework for object detection and segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3041-3050, 2023. 2", + "[20] Yanghao Li, Hanzi Mao, Ross Girshick, and Kaiming He. Exploring plain vision transformer backbones for object detection. In European Conference on Computer Vision, pages 280-296. Springer, 2022. 5", + "[21] Zhiqi Li, Wenhai Wang, Hongyang Li, Enze Xie, Chonghao Sima, Tong Lu, Yu Qiao, and Jifeng Dai. Bevformer: Learning bird's-eye-view representation from multi-camera images via spatiotemporal transformers. In European conference on computer vision, pages 1-18. Springer, 2022. 2", + "[22] Zhiqi Li, Wenhai Wang, Enze Xie, Zhiding Yu, Anima Anandkumar, Jose M Alvarez, Ping Luo, and Tong Lu. Panoptic segformer: Delving deeper into panoptic segmentation with transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1280-1289, 2022. 2", + "[23] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 7", + "[24] Yingfei Liu, Tiancai Wang, Xiangyu Zhang, and Jian Sun. Petr: Position embedding transformation for multi-view 3d object detection. In European Conference on Computer Vision, pages 531-548. Springer, 2022. 2", + "[25] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10012-10022, 2021. 1, 2, 3, 5, 6" + ], + "bbox": [ + 503, + 92, + 893, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "5660", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[26] Zhuang Liu, Hanzi Mao, Chao-Yuan Wu, Christoph Feichtenhofer, Trevor Darrell, and Saining Xie. A convnet for the 2020s. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11976-11986, 2022. 1, 2, 6", + "[27] Ningning Ma, Xiangyu Zhang, Hai-Tao Zheng, and Jian Sun. Shufflenet v2: Practical guidelines for efficient cnn architecture design. In ECCV, 2018. 2, 3, 4", + "[28] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. Advances in neural information processing systems, 32, 2019. 5", + "[29] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 1, 2, 8", + "[30] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical Image Computing and Computer-Assisted Intervention-MICCAI 2015: 18th International Conference, Munich, Germany, October 5-9, 2015, Proceedings, Part III 18, pages 234-241. Springer, 2015. 2, 8", + "[31] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in Neural Information Processing Systems, 35:36479-36494, 2022. 1", + "[32] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556, 2014. 2", + "[33] Robin Strudel, Ricardo Garcia, Ivan Laptev, and Cordelia Schmid. Segmenter: Transformer for semantic segmentation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 7262-7272, 2021. 2", + "[34] Hugues Thomas, Charles R Qi, Jean-Emmanuel Deschaud, Beatz Marcotegui, François Goulette, and Leonidas J Guibas. Kpconv: Flexible and deformable convolution for point clouds. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6411–6420, 2019. 2", + "[35] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 2, 5", + "[36] Ashish Vaswani, Prajit Ramachandran, Aravind Srinivas, Niki Parmar, Blake Hechtman, and Jonathon Shlens. Scaling local self-attention for parameter efficient visual backbones. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12894–12904, 2021. 2", + "[37] Shenlong Wang, Simon Suo, Wei-Chiu Ma, Andrei Pokrovsky, and Raquel Urtasun. Deep parametric continuous convolutional neural networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2589-2597, 2018. 2" + ], + "bbox": [ + 78, + 90, + 470, + 901 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[38] Wenhai Wang, Jifeng Dai, Zhe Chen, Zhenhang Huang, Zhiqi Li, Xizhou Zhu, Xiaowei Hu, Tong Lu, Lewei Lu, Hongsheng Li, et al. Internimage: Exploring large-scale vision foundation models with deformable convolutions. In CVPR, 2023. 1, 2, 6", + "[39] Ross Wightman. Pytorch image models. https://github.com/rwrightman/pytorch-image-models, 2019.6", + "[40] Felix Wu, Angela Fan, Alexei Baevski, Yann N Dauphin, and Michael Auli. Pay less attention with lightweight and dynamic convolutions. arXiv preprint arXiv:1901.10430, 2019. 2", + "[41] Tete Xiao, Yingcheng Liu, Bolei Zhou, Yuning Jiang, and Jian Sun. Unified perceptual parsing for scene understanding. In Proceedings of the European conference on computer vision (ECCV), pages 418-434, 2018. 1, 7", + "[42] Yuwen Xiong, Mengye Ren, Renjie Liao, Kelvin Wong, and Raquel Urtasun. Deformable filter convolution for point cloud reasoning. arXiv preprint arXiv:1907.13079, 2019. 2", + "[43] Chenyu Yang, Yuntao Chen, Hao Tian, Chenxin Tao, Xizhou Zhu, Zhaoxiang Zhang, Gao Huang, Hongyang Li, Yu Qiao, Lewei Lu, et al. Bevformer v2: Adapting modern image backbones to bird's-eye-view recognition via perspective supervision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17830-17839, 2023. 2, 7", + "[44] Xiaohua Zhai, Alexander Kolesnikov, Neil Houlsby, and Lucas Beyer. Scaling vision transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12104-12113, 2022. 1", + "[45] Hao Zhang, Feng Li, Shilong Liu, Lei Zhang, Hang Su, Jun Zhu, Lionel M Ni, and Heung-Yeung Shum. Dino: Detr with improved denoising anchor boxes for end-to-end object detection. arXiv preprint arXiv:2203.03605, 2022. 2", + "[46] Bolei Zhou, Hang Zhao, Xavier Puig, Sanja Fidler, Adela Barriuso, and Antonio Torralba. Scene parsing through ade20k dataset. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 633-641, 2017. 7", + "[47] Xizhou Zhu, Han Hu, Stephen Lin, and Jifeng Dai. Deformable convnets v2: More deformable, better results. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 2", + "[48] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable detr: Deformable transformers for end-to-end object detection. In International Conference on Learning Representations, 2020. 2, 5" + ], + "bbox": [ + 501, + 92, + 893, + 722 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "5661", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2024/Efficient Deformable ConvNets_ Rethinking Dynamic and Sparse Operator for Vision Applications/d839f1b7-f6aa-4faa-bb5a-7c661a9aa4e5_model.json b/2024/Efficient Deformable ConvNets_ Rethinking Dynamic and Sparse Operator for Vision Applications/d839f1b7-f6aa-4faa-bb5a-7c661a9aa4e5_model.json new file mode 100644 index 0000000000000000000000000000000000000000..dba7cd1b73c53a5a1f726b81d918ecc3b5dbcd91 --- /dev/null +++ b/2024/Efficient Deformable ConvNets_ Rethinking Dynamic and Sparse Operator for Vision Applications/d839f1b7-f6aa-4faa-bb5a-7c661a9aa4e5_model.json @@ -0,0 +1,2035 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.158, + 0.13, + 0.817, + 0.177 + ], + "angle": 0, + "content": "Efficient Deformable ConvNets: Rethinking Dynamic and Sparse Operator for Vision Applications" + }, + { + "type": "text", + "bbox": [ + 0.224, + 0.202, + 0.746, + 0.221 + ], + "angle": 0, + "content": "Yuwen Xiong\\(^{*,1,2}\\) Zhiqi Li\\(^{*,3,2}\\) Yuntao Chen\\(^{*4}\\) Feng Wang\\(^{*5}\\)" + }, + { + "type": "text", + "bbox": [ + 0.165, + 0.221, + 0.804, + 0.24 + ], + "angle": 0, + "content": "Xizhou Zhu\\(^{5,6}\\) Jiapeng Luo\\(^{6}\\) Wenhai Wang\\(^{7,2}\\) Tong Lu\\(^{3}\\) Hongsheng Li\\(^{7}\\)" + }, + { + "type": "text", + "bbox": [ + 0.277, + 0.239, + 0.695, + 0.256 + ], + "angle": 0, + "content": "Yu Qiao² Lewei Lu⁶ Jie Zhou⁵ Jifeng Dai⁵,²" + }, + { + "type": "text", + "bbox": [ + 0.209, + 0.273, + 0.762, + 0.291 + ], + "angle": 0, + "content": "\\(^{1}\\)University of Toronto \\(^{2}\\)OpenGVLab, Shanghai AI Laboratory" + }, + { + "type": "text", + "bbox": [ + 0.176, + 0.291, + 0.794, + 0.308 + ], + "angle": 0, + "content": "\\(^{3}\\)Nanjing University \\(^{4}\\)CAIR, HKISI, CAS \\(^{5}\\)Tsinghua University" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.308, + 0.756, + 0.327 + ], + "angle": 0, + "content": "\\(^{6}\\)SenseTime Research \\(^{7}\\)The Chinese University of Hong Kong" + }, + { + "type": "text", + "bbox": [ + 0.281, + 0.327, + 0.682, + 0.343 + ], + "angle": 0, + "content": "https://github.com/OpenGVLab/DCNv4" + }, + { + "type": "title", + "bbox": [ + 0.265, + 0.378, + 0.342, + 0.394 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.41, + 0.474, + 0.771 + ], + "angle": 0, + "content": "We introduce Deformable Convolution v4 (DCNv4), a highly efficient and effective operator designed for a broad spectrum of vision applications. DCNv4 addresses the limitations of its predecessor, DCNv3, with two key enhancements: 1. removing softmax normalization in spatial aggregation to enhance its dynamic property and expressive power and 2. optimizing memory access to minimize redundant operations for speedup. These improvements result in a significantly faster convergence compared to DCNv3 and a substantial increase in processing speed, with DCNv4 achieving more than three times the forward speed. DCNv4 demonstrates exceptional performance across various tasks, including image classification, instance and semantic segmentation, and notably, image generation. When integrated into generative models like U-Net in the latent diffusion model, DCNv4 outperforms its baseline, underscoring its possibility to enhance generative models. In practical applications, replacing DCNv3 with DCNv4 in the InternImage model to create FlashInternImage results in up to \\(80\\%\\) speed increase and further performance improvement without further modifications. The advancements in speed and efficiency of DCNv4, combined with its robust performance across diverse vision tasks, show its potential as a foundational building block for future vision models." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.791, + 0.208, + 0.807 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.817, + 0.472, + 0.862 + ], + "angle": 0, + "content": "In the field of computer vision, there is an ongoing debate about whether convolutional networks (ConvNets) or Transformers offer superior performance. In recent years, Trans" + }, + { + "type": "image", + "bbox": [ + 0.51, + 0.378, + 0.737, + 0.527 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.614, + 0.532, + 0.631, + 0.542 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.745, + 0.379, + 0.895, + 0.526 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.81, + 0.531, + 0.829, + 0.542 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.545, + 0.895, + 0.6 + ], + "angle": 0, + "content": "Figure 1. (a) We show relative runtime with DCNv3 as the baseline. DCNv4 shows significant speedup over DCNv3, and surpasses other common vision operators. (b) With the same network architecture, DCNv4 converges faster than other operators." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.602, + 0.895, + 0.766 + ], + "angle": 0, + "content": "former models [12, 25, 44] have achieved remarkable results in large vision models with the attention mechanism, showing the potential to overtake ConvNets. However, recent works such as InternImage [38] and ConvNeXt [26] demonstrate that ConvNet-based vision models retain robust performance, efficiency, simplicity, and suitable inductive bias for various downstream tasks [15, 41]. Notably, in domains like image generation [29, 31], models containing both convolutions and transformers remain the preferred approach. This situation brings to light the enduring value of convolution-based approaches." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.767, + 0.896, + 0.901 + ], + "angle": 0, + "content": "Building on convolution's strengths, Deformable Convolution v3 (DCNv3) – the core operator of the advanced ConvNet model InternImage – innovatively combines a sparse attention mechanism with convolution: it processes each output location in a sliding window manner with a small window size (e.g. \\( 3 \\times 3 = 9 \\) points) which acts as a local, sparse operator like convolution, while dynamically samples point with an adaptive range and aggregates the spatial features with input-dependent attention weights. With its" + }, + { + "type": "page_footnote", + "bbox": [ + 0.1, + 0.876, + 0.212, + 0.888 + ], + "angle": 0, + "content": "* Equal contribution" + }, + { + "type": "page_footnote", + "bbox": [ + 0.106, + 0.888, + 0.388, + 0.9 + ], + "angle": 0, + "content": "Corresponding author (daijifeng@tsinghua.edu.cn)" + }, + { + "type": "list", + "bbox": [ + 0.1, + 0.876, + 0.388, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5652" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.471, + 0.151 + ], + "angle": 0, + "content": "small window size and ConvNet inductive bias, DCNv3 is expected to achieve a faster convergence rate and lower inference latency, especially when compared to dense global [12] or local window-based [25] attention methods." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.152, + 0.472, + 0.36 + ], + "angle": 0, + "content": "Despite these advantages, DCN has not become the go-to solution for vision backbone models. This observation led us to investigate the lingering limitations of the DCN operator. The first thing we notice is the running speed. The slow speed of DCN is known to be a long-standing problem [1], as it introduces extra overhead on sampling non-nearby locations, making it not fit modern convolution algorithms. Our comparative analysis, illustrated in Fig. 1a, reveals that DCNv3 can be slower than a properly optimized dense global attention [9], highlighting the need for further optimization. Moreover, we find DCNv3 even converges slower than global attention at the initial backbone training phase, as shown in Fig. 1b, which is counter-intuitive as DCNv3 is equipped with ConvNet inductive bias." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.362, + 0.472, + 0.45 + ], + "angle": 0, + "content": "To overcome these challenges, we propose Deformable Convolution v4 (DCNv4), an innovative advancement to optimize the sparse DCN operator for practical efficiency. DCNv4 comes with a much faster implementation and an improved operator design to enhance its performance, which we will elaborate on as follows:" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.452, + 0.472, + 0.557 + ], + "angle": 0, + "content": "First, we conduct instruction-level kernel profiling for existing implementation and find that DCNv3 is already lightweight. The compute cost is less than \\(1\\%\\), while memory access costs \\(99\\%\\). This motivates us to revisit the operator implementation and find that many memory accesses in the DCN forward process are redundant and thus can be optimized, leading to a much faster DCNv4 implementation." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.557, + 0.472, + 0.706 + ], + "angle": 0, + "content": "Second, drawing inspiration from convolution's unbounded weight range, we find that softmax in spatial aggregation, a standard operation in dense attention, is unnecessary in DCNv3, as it is not a requirement for operators with a dedicated aggregation window for each location. Intuitively, softmax puts a bounded \\(0 \\sim 1\\) value range to the weight and will limit the expressive power of the aggregation weight. This insight led us to remove the softmax in DCNv4, enhancing its dynamic property and improving its performance." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.707, + 0.472, + 0.766 + ], + "angle": 0, + "content": "As a result, DCNv4 converges significantly faster than DCNv3 and accelerates forward speed by more than \\( 3 \\times \\). This improvement allows DCNv4 to fully leverage its sparse property and become one of the fastest core vision operators." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.767, + 0.472, + 0.902 + ], + "angle": 0, + "content": "We further replace DCNv3 in InternImage with DCNv4, creating FlashInternImage. Remarkably, FlashInternImage achieves a \\(50 \\sim 80\\%\\) speed increase compared to InternImage without any additional modifications. This enhancement positions FlashInternImage as one of the fastest modern vision backbone networks while maintaining superior performance. With the help of DCNv4, FlashInternImage significantly improves the convergence speed in ImageNet classification [10] and transfer learning settings and further" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.885, + 0.107 + ], + "angle": 0, + "content": "demonstrates improved performance in downstream tasks." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.108, + 0.895, + 0.331 + ], + "angle": 0, + "content": "Furthermore, DCNv4 shows potential as a universal vision operator in various architectures and tasks. We integrate DCNv4 into other modern backbone architectures, including ConvNeXt [26] and ViT [12], replacing depthwise convolution [6] and dense self-attention layers [35]. Surprisingly, without any hyperparameter adjustments, these meticulously designed networks with DCNv4 perform on par while being much faster, showing the efficacy and efficiency of the dynamic, sparse DCNv4. Moreover, we explore the potential of DCNv4 in generative models as a new application domain. Specifically, we apply it in the U-Net [30] architecture used in diffusion models [29], replacing regular convolution with DCNv4. Our experimental results show that DCNv4 can work better than the baselines in image generation, showing great potential for using DCN to improve generative models." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.332, + 0.895, + 0.362 + ], + "angle": 0, + "content": "We have released the DCNv4 implementation to facilitate future research in the vision community." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.369, + 0.642, + 0.384 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.387, + 0.897, + 0.867 + ], + "angle": 0, + "content": "Core operators in vision models: The standard convolution [17] stands as the most prevalent and impactful operator, forming the backbone of the majority of computer vision architectures [14, 16, 32]. Nevertheless, a myriad of operators, each with unique characteristics, collectively play a crucial role in the development of computer vision. Depthwise separable convolution (DWConv) [6] separates the spatial and channel operations, and has been pivotal in developing lightweight and efficient models [26, 27]. RepLKNet [11] illustrates that a purely convolutional network, leveraging large-kernel depth-wise convolutions, can attain competitive performance in both efficiency and effectiveness. Deformable Convolution (DCN) series [7, 38, 47] significantly leaps the adaptability of convolution by adding learnable offsets to the convolutions kernels. Contrary to convolutions, attention mechanisms [35] possess the capacity to model long-range dependencies and have been successfully adopted in various computer vision tasks [3, 12, 24, 33]. Window attention [25, 36] reduces the computational complexity inherent in vanilla attention by restricting the attention operation to a fixed-size window. To mitigate the high computational complexity associated with vanilla attention, deformable attention [48] enables each query to concentrate on a select number of key sampling points, with dynamically determined locations and weights. This efficient method is widely used in the following arts perception methods [4, 19, 21, 22, 43, 45]. DynamicConv [40] and dynamic-DWNet [13] augment DWConv by incorporating dynamic weights, thereby enabling the use of instance-specific weights that adapt dynamically. For non-grid structured data, sparse operators [34, 37, 42] utilize dynamic weights obtained via bilinear interpolation or in a parametric way." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.871, + 0.897, + 0.902 + ], + "angle": 0, + "content": "Memory access cost in vision backbones: As underscored in previous studies [18, 27], FLOPs, although a fre" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5653" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.082, + 0.089, + 0.468, + 0.144 + ], + "angle": 0, + "content": "
Model5th EP10th Ep20th Ep50th Ep300th Ep
ConvNeXt29.953.566.174.883.8
ConvNeXt + softmax8.5 (-21.4)25.3 (-28.2)51.1 (-15.0)69.1 (-5.7)81.6 (-2.2)
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.146, + 0.473, + 0.187 + ], + "angle": 0, + "content": "Table 1. ImageNet-1K accuracy at different training epochs. Softmax on the convolution weights significantly affects the convergence speed and the final performance for the ConvNeXt model." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.192, + 0.474, + 0.419 + ], + "angle": 0, + "content": "sequently used metric to measure model complexity, do not accurately represent the model's speed or latency. In practical scenarios, the running speed of a model is influenced by multiple factors, not just FLOPs. Memory access cost plays a significant role in this context [27]. Flash-Attention [9], by reducing the number of accesses to High Bandwidth Memory (HBM), achieves a significantly faster speed in practice despite having higher FLOPs compared to vanilla attention. Although DCN operators do not exhibit a disadvantage in terms of FLOPs, their latency is considerably longer compared to DW-Conv, under the same FLOPs budget, predominantly due to substantial memory access costs. In this work, we conduct a thorough analysis and optimization of the memory access costs associated with the DCN operators, significantly accelerating the DCN's running speed." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.427, + 0.169, + 0.442 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.452, + 0.47, + 0.482 + ], + "angle": 0, + "content": "3.1. Rethinking the Dynamic Property in Deformable Convolution" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.489, + 0.471, + 0.537 + ], + "angle": 0, + "content": "Revisiting DCNv3: Given an input \\(\\mathbf{x} \\in \\mathbb{R}^{H \\times W \\times C}\\) with height \\(H\\), width \\(W\\) and channel \\(C\\), the DCNv3 operation with \\(K\\) points is defined in Eq. (2) for each point \\(p_0\\):" + }, + { + "type": "equation", + "bbox": [ + 0.143, + 0.544, + 0.47, + 0.585 + ], + "angle": 0, + "content": "\\[\n\\mathbf {y} _ {g} = \\sum_ {k = 1} ^ {K} \\mathbf {m} _ {g k} \\mathbf {x} _ {g} \\left(p _ {0} + p _ {k} + \\Delta p _ {g k}\\right), \\tag {1}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.152, + 0.588, + 0.47, + 0.607 + ], + "angle": 0, + "content": "\\[\n\\mathbf {y} = \\operatorname {c o n c a t} \\left(\\left[ \\mathbf {y} _ {1}, \\mathbf {y} _ {2}, \\dots , \\mathbf {y} _ {G} \\right], \\text {a x i s} = - 1\\right), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.617, + 0.473, + 0.903 + ], + "angle": 0, + "content": "where \\(G\\) denotes the number of spatial aggregation groups. For the \\(g\\)-th group, \\(\\mathbf{x}_g,\\mathbf{y}_g\\in \\mathbb{R}^{H\\times W\\times C'}\\) represents the sliced input/output feature map with \\(C' = C / G\\) represents the group dimension; \\(\\mathbf{m}_{gk}\\in \\mathbb{R}\\) denotes the spatial aggregation weights (also known as modulation scalar) of the \\(k\\)-th sampling point in the \\(g\\)-th group, conditioned on the input \\(\\mathbf{x}\\) and normalized by the softmax function along the dimension \\(K\\); \\(p_k\\) denotes the \\(k\\)-th location of the pre-defined grid sampling \\(\\{(-1, - 1),(-1,0),\\dots,(0, + 1),\\dots,(+1, + 1)\\}\\) as in regular convolutions and \\(\\Delta p_{gk}\\) is the offset corresponding to the grid sampling location \\(p_k\\) in the \\(g\\)-th group. A \\(1\\times 1\\) point-wise convolution on \\(\\mathbf{x}\\) and \\(\\mathbf{y}\\) can be applied before and after the DCNv3 operator to enhance the model's expressive power, following separable convolution [6]. DCNv3 is a combination of convolution and attention: on the one hand, it processes the input data in a sliding window manner, which follows convolution and inherent its inductive bias; on the other hand, the sampling offset \\(\\Delta p\\) and spatial aggregation weight \\(\\mathbf{m}\\) are dynamically predicted from the input feature," + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.091, + 0.671, + 0.166 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.559, + 0.166, + 0.608, + 0.174 + ], + "angle": 0, + "content": "(a) Attention" + }, + { + "type": "image", + "bbox": [ + 0.506, + 0.174, + 0.685, + 0.248 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.695, + 0.091, + 0.871, + 0.167 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.752, + 0.167, + 0.796, + 0.175 + ], + "angle": 0, + "content": "(b) DCNv3" + }, + { + "type": "image", + "bbox": [ + 0.696, + 0.174, + 0.889, + 0.248 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.251, + 0.895, + 0.404 + ], + "angle": 0, + "content": "Figure 2. Comparisons of core operators in spatial aggregation for query pixels on different locations within the same channel. (a) Attention and (b) DCNv3 use bounded (range from \\(0 \\sim 1\\)) dynamic weights to aggregate spatial features, while the window (sampling point set) for attention is the same, and DCNv3 uses a dedicated window for each location. (c) Convolution has a more flexible unbounded value range for aggregation weights and uses a dedicated sliding window for each location, but the window shape and aggregation weights are input-independent. (d) DCNv4 combines their advantages, using an adaptive aggregation window and dynamic aggregation weights with an unbounded value range." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.412, + 0.893, + 0.457 + ], + "angle": 0, + "content": "showing its dynamic property and making it more like an attention mechanism. We compare different operators where each has its own property, as illustrated in Fig. 2" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.462, + 0.895, + 0.582 + ], + "angle": 0, + "content": "Softmax normalization: A key difference between convolution and DCNv3 is that DCNv3 normalizes \\(\\mathbf{m}\\), the spatial aggregation weights, with a softmax function, following the convention of scaled dot-product self-attention. Conversely, convolution does not employ softmax over its weights and still works well. The reason why attention needs a softmax is straightforward: scaled dot-product self-attention with \\(Q, K, V \\in \\mathbb{R}^{N \\times d}\\) is defined with a formulation:" + }, + { + "type": "equation", + "bbox": [ + 0.618, + 0.589, + 0.893, + 0.622 + ], + "angle": 0, + "content": "\\[\n\\operatorname {s o f t m a x} \\left(\\frac {1}{\\sqrt {d}} Q K ^ {\\top}\\right) V, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.631, + 0.895, + 0.871 + ], + "angle": 0, + "content": "where \\(N\\) is the number of points in the same attention window (can be either global [12] or local [25]), \\(d\\) is the hidden dimension, \\(Q, K, V\\) are the query, key, and value matrices computed from the input. Softmax operation is required in Eq. (3) for attention; without softmax, \\(K^{\\top}V \\in \\mathbb{R}^{d \\times d}\\) can be calculated first, and it degrades to a linear projection for all queries in the same attention window, resulting in degenerated performance. However, for convolutional operators like depthwise convolution and DCNv3 where each point has its own dedicated aggregation window and the values in each aggregation window are already different and there is no \"key\" concept, such degradation issues no longer exist, and the normalization becomes unnecessary. In fact, normalizing convolution weights within a fixed 0-1 range using softmax could impose a significant limitation on the operator's expressive power and make the learning slower." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.871, + 0.894, + 0.902 + ], + "angle": 0, + "content": "To confirm this hypothesis, we train a ConvNeXt model and apply softmax to the \\(7 \\times 7\\) window of the depthwise" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5654" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.198 + ], + "angle": 0, + "content": "convolution weights before convolution forward. We observe a remarkable decline in model performance as well as convergence speed from results in Tab. 1. This suggests that for operators with a dedicated aggregation window on each location like convolution or DCN, aggregation weights with an unbounded range offer better expressive power than softmax-normalized, bounded-range weights." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.204, + 0.474, + 0.399 + ], + "angle": 0, + "content": "Enhancing dynamic property: Motivated by this observation, we remove the softmax normalization in DCNv3, transforming the modulation scalars ranging from 0 to 1 to unbounded dynamic weights similar to convolution. As shown in Fig. 2, this alteration further amplifies the dynamic property of DCN, where other operators have certain limits, such as bounded value range (attention/DCNv3) or fixed aggregation window with input-independent aggregation weights (convolution). Fig. 1b shows that by making this change, DCNv4 converges significantly faster than DCNv3 and other common operators, including convolution and attention. Results in Sec. 4 further showcase that DCNv4 works well in both pre-training and transfer learning settings." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.404, + 0.253, + 0.42 + ], + "angle": 0, + "content": "3.2. Speeding up DCN" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.425, + 0.473, + 0.591 + ], + "angle": 0, + "content": "Theoretically, DCN, as a sparse operator with \\(3 \\times 3\\) window, should act faster than other common operators that employ larger window sizes, like dense attention or \\(7 \\times 7\\) depthwise convolution. However, we find that this is not the case, as shown in Fig. 1a. In this subsection, we first conduct a theoretical analysis of GPU efficiency, showing a large variance in memory access cost depending on how we read the memory. We further perform optimization based on our observations, significantly improving the speed of DCN by saving additional memory instruction and bringing the speed advantage of being a sparse operator into reality." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.601, + 0.471, + 0.736 + ], + "angle": 0, + "content": "Theoretical analysis of GPU efficiency Our study begins with a theoretical examination of the DCNv3 operator's computational behavior. We employ the roofline model to evaluate its performance, focusing on theoretical FLOPs and memory access cost. For an input and output tensor of shape \\((H,W,C)\\), the DCNv3 operator requires \\(36HWC\\) FLOPs, where \\(3\\times 3\\) represents the convolution kernel's spatial dimensions and the factor of 4 accounts for the bilinear interpolation at each sampling point." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.737, + 0.473, + 0.901 + ], + "angle": 0, + "content": "Following the framework outlined in [27], DCNv3's memory access cost is calculated as \\( 2HWC + 27HWG \\). The first term corresponds to the input/output feature map size and the second to the DCNv3's offset and aggregation weights with \\( G \\) groups. We approximate \\( G \\) as \\( C / 16 \\) assuming a group dimension of 16, resulting in approximately \\( 3.7HWC \\) memory access cost. However, this assumes an ideal scenario of infinite cache and a single memory read for each value, which is often unrealistic in parallel computing environments where concurrent thread execution necessitates simultaneous data access." + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.091, + 0.895, + 0.219 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.224, + 0.895, + 0.294 + ], + "angle": 0, + "content": "Figure 3. Illustration of our optimization. In DCNv4, we use one thread to process multiple channels in the same group that shares sampling offset and aggregation weights. Workloads like memory reading and bilinear interpolation coefficient computation can be reduced, and multiple memory access instructions can be merged." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.303, + 0.895, + 0.392 + ], + "angle": 0, + "content": "To estimate the maximum memory access requirement, we consider a scenario devoid of cache, where each output location requires fresh memory reads and involves 36 reads for bilinear interpolation, 27 for offset/aggregation weights, and one write operation, resulting in a memory access cost of \\(64HWC\\). This is 17 times larger than the ideal case." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.392, + 0.895, + 0.513 + ], + "angle": 0, + "content": "This analysis reveals a substantial gap in the ratio of computation-to-memory access (ranging from 0.6 to 9.7), highlighting the significant potential for memory access optimization. Notably, despite DCNv3's use of input-dependent, dynamic offsets causing non-deterministic memory access, one deterministic thing is that channels within the same group share offset values. This leads us to propose a specific optimization strategy for enhancing DCNv3's speed." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.514, + 0.896, + 0.844 + ], + "angle": 0, + "content": "Eliminating redundant workload: In previous CUDA implementations of DCN kernel, for input with shape \\((H,W,C)^1\\), offset \\((H,W,G,K^2\\times 2)\\) and aggregation weight \\((H,W,G,K^2)\\), we will create \\(H\\times W\\times C\\) threads in total to maximize parallelism, where each thread processes one channel for one output location. Notably, the \\(D = C / G\\) channels within each group share the same sampling offset and aggregation weight values for each output location. Using multiple threads to process these \\(D\\) channels on the same output location is wasteful as different threads will read the same sampling offset and aggregation weight values from GPU memory multiple times, which is critical for a memory-bound operator. Processing multiple channels within the same group on each output location with one thread can eliminate these redundant memory read requests, greatly reducing memory bandwidth usage. As the sampling locations are the same, we can also only calculate the bilinear interpolation coefficient used in DCN once. Specifically, if each thread processes \\(D^{\\prime}\\) channels, the memory access cost for reading offset and aggregation weight, as well as the computation cost for calculating bilinear interpolation coefficient, can be reduced \\(D^{\\prime}\\) times." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.848, + 0.895, + 0.878 + ], + "angle": 0, + "content": "Eliminating redundant memory instructions: In practice, solely reusing threads for multiple channels will not" + }, + { + "type": "page_footnote", + "bbox": [ + 0.518, + 0.887, + 0.869, + 0.901 + ], + "angle": 0, + "content": "1We assume the batch size is 1 and memory layout is channel-last." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5655" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.12, + 0.089, + 0.847, + 0.223 + ], + "angle": 0, + "content": "
OperatorRuntime (ms)
56 × 56 × 12828 × 28 × 25614 × 14 × 5127 × 7 × 102414 × 14 × 768
Attention (torch)30.8 / 19.33.35 / 2.120.539 / 0.4480.446 / 0.1210.779 / 0.654
FlashAttention-2N/A / 2.46N/A / 0.451N/A / 0.123N/A / 0.0901N/A / 0.163
Window Attn (7 × 7)4.05 / 1.462.07 / 0.7701.08 / 0.4220.577 / 0.2391.58 / 0.604
DWConv (7 × 7, torch)2.02 / 1.981.03 / 1.000.515 / 0.5230.269 / 0.2610.779 / 0.773
DWConv (7 × 7, cuDNN)0.981 / 0.4380.522 / 0.2670.287 / 0.1530.199 / 0.1020.413 / 0.210
DCNv31.45 / 1.520.688 / 0.7110.294 / 0.2980.125 / 0.1260.528 / 0.548
DCNv40.606 / 0.4040.303 / 0.2300.145 / 0.1230.0730 / 0.06800.224 / 0.147
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.224, + 0.892, + 0.251 + ], + "angle": 0, + "content": "Table 2. Op-level benchmark on standard input shape with various downsample rates. FP32/FP16 results are reported when the implementation is available. Our new DCNv4 can surpass all other commonly used operators under different input resolutions." + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.259, + 0.473, + 0.692 + ], + "angle": 0, + "content": "see speed improvement. The reason is that when \\( D' \\) increases, we create fewer threads and the workload of each thread now increases \\( D' \\) times. This essentially reduces the degree of parallelism for the CUDA kernel. Luckily, the DCN kernel is now computationally lightweight as the bilinear interpolation only needs to be performed once for all \\( D' \\) channels, and most of the workload is the memory instructions reading input values from different channels. When the memory layout is channel-last, and all \\( D' \\) channel values are contiguous, we can leverage vectorized load: for example, to read four 32-bit float values from memory, instead of reading one 32-bit float value four times with four instructions, we can use a single instruction to load a 128-bit packed value at once, thus reducing the number of instructions and execution time of each thread. We can apply similar technique when writing the final results to GPU memory, minimizing the memory access time and increasing memory bandwidth utilization. Moreover, the modern half-precision data format (float16/bfloat16) halves the bytes that need to be loaded, which means the memory efficiency can be twice as much under the same memory bandwidth when using the half-precision format. However, we do not see speed improvement with half-precision data in the original DCNv3 implementation, possibly due to too much overhead on data access and computation, while in our new implementation, the speedup is significant. It is worth noting that the aforementioned optimization techniques can also be applied to DCNv1/v2 and deformable attention [48], as they share a similar performance bottleneck and issue." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.706, + 0.472, + 0.903 + ], + "angle": 0, + "content": "Micro design in DCN module: DCNv3 module introduces multiple micro designs; as the core kernel is optimized, their impact on the speed becomes non-negligible. We identify two points in DCNv3 designs that could be further optimized: first, after removing the softmax and transforming the modulation scalar into dynamic aggregation weights as mentioned in the previous paragraph. The linear layers for computing offset and dynamic weights can actually be combined into one linear layer. This reduces network fragmentation and eliminates extra overheads, such as kernel launching and synchronization, enhancing run-time efficiency on the GPU; second, in the original DCNv3 module design, a complex sub-network that consists of depthwise \\( 3 \\times 3 \\) conv, layer" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.26, + 0.895, + 0.366 + ], + "angle": 0, + "content": "normalization (LN), GELU, and linear layer is used to compute offsets and dynamic weights. Following the design in Xception [6], we remove the additional LN-GELU layers and use the original separable convolution structure, further reducing running time. We empirically find that if latency is a higher priority, the depthwise convolution can also be removed with only a minor performance sacrifice." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.369, + 0.634, + 0.387 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.395, + 0.895, + 0.502 + ], + "angle": 0, + "content": "In this section, we verify the effectiveness of our proposed DCNv4 module from both speed as well as performance perspective. We benchmark the operator-level speed and integrate DCNv4 into the backbone to test the system-level performance. All speed test results are obtained with an NVIDIA A100 80G SXM GPU. We include additional experimental results and implementation details in supp." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.507, + 0.787, + 0.524 + ], + "angle": 0, + "content": "4.1. Speed Benchmark for Operators" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.527, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Settings: We conduct the op-level benchmark by solely measuring the running time of several representative operators building state-of-the-art vision backbone models, including full attention [35] implemented with PyTorch as well as the advanced FlashAttention-2 [8] implementation, window-based attention with window size \\(7 \\times 7\\) [25], depthwise convolution with \\(7 \\times 7\\) window, implemented by cuDNN [5] and ATen library from PyTorch [28], respectively. We only benchmark the core operation for spatial aggregation, and additional linear layers like qkv projection and output projection layers are excluded and not included in the runtime measurement. Please refer to supp. for a more comprehensive module-level comparison. We first consider a feature map shape generated from the standard \\(224 \\times 224\\) input resolution for image classification with 4, 8, 16, \\(32 \\times\\) downsample ratio as used by common hierarchical ConvNet/transformer backbones; we also add a feature map shape from isotropic backbone like ViT with a downsampling ratio 16 and larger hidden dimension. We further consider high-resolution inputs often used in downstream tasks like object detection. We set the input shape to be \\(800 \\times 1280\\) and \\(1024 \\times 1024\\) for the hierarchical feature map and isotropic feature map, respectively, as they are the common practice in object detection [15, 20]. Batch size is 64 and 1 for these two input sets, respectively. For operators with a head/group concept, we" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.957 + ], + "angle": 0, + "content": "5656" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.123, + 0.089, + 0.849, + 0.219 + ], + "angle": 0, + "content": "
OperatorRuntime (ms)
200 × 320 × 128100 × 160 × 25650 × 80 × 51225 × 40 × 102464 × 64 × 768
Attention (torch)OOM / OOM25.4 / 12.92.88 / 1.890.490 / 0.3094.17 / 2.57
FlashAttention-2N/A / 13.2N/A / 1.74N/A / 0.285N/A / 0.0797N/A / 0.437
Window Attn (7 × 7)1.33 / 0.5090.728 / 0.2910.426 / 0.1860.279 / 0.1650.673 / 0.272
DWConv (7 × 7, torch)0.634 / 0.6080.313 / 0.3150.167 / 0.1580.0943 / 0.08940.260 / 0.253
DWConv (7 × 7, cuDNN)0.331 / 0.2820.188 / 0.1680.114 / 0.1150.0817 / 0.08810.161 / 0.156
DCNv30.472 / 0.4930.244 / 0.2530.128 / 0.1320.0737 / 0.07670.194 / 0.199
DCNv40.210 / 0.1360.124 / 0.08950.0707 / 0.05890.0452 / 0.04260.103 / 0.0672
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.227, + 0.895, + 0.255 + ], + "angle": 0, + "content": "Table 3. Op-level benchmark on high-resolution input shape with various downsample rates. DCNv4 performs well as a sparse operator, surpassing all other baselines, while dense global attention is slow under this scenario." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.264, + 0.471, + 0.295 + ], + "angle": 0, + "content": "set the dimension of each head/group to 32 and change the number of heads/groups when the hidden dimension varies." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.297, + 0.474, + 0.583 + ], + "angle": 0, + "content": "Results: We show the benchmark results on standard resolution and high-resolution input in Tab. 2 and Tab. 3. We report results with both FP32 and FP16 data formats unless the FP32 implementation is not available. Dense global attention implemented with PyTorch performs significantly slower when the input resolution is large and even out of memory. FlashAttention significantly improves the speed of attention and can be even faster than \\(7 \\times 7\\) window attention in certain cases, indicating the importance of proper optimization. However, it does not change the quadratic complexity of attention; when the input resolution is high, it still falls behind local/sparse operators like window attention or convolution. While DCNv3 can be faster than DWConv with plain implementation, it is slower than the heavily optimized cuDNN version. Instead, our DCNv4 can provide more than \\(3 \\times\\) speedup compared to DCNv3, greatly saving the running time. Moreover, DCNv4 can leverage the advantage of using a \\(3 \\times 3\\) sparse window to perform significantly faster than other baselines under different settings." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.589, + 0.27, + 0.605 + ], + "angle": 0, + "content": "4.2. Image Classification" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.609, + 0.472, + 0.849 + ], + "angle": 0, + "content": "Settings: We evaluate the effectiveness of DCNv4 on ImageNet classification. We start from InternImage [38] as it shows state-of-the-art performance among ConvNet-based models. We replace the DCNv3 in InternImage with DCNv4 and create FlashInternImage. Other implementation details, including network architecture and hyperparameters, are kept the same as [38]. We compare Swin-Transformer and ConvNeXt which are two representative baselines in Transformer and ConvNet models. We follow the common practice [25, 26, 38] of training protocols, including data augmentation, preprocessing, optimizer and learning rate schedule, and train FlashInternImage-Tiny/Small/Base on ImageNet-1K for 300 epochs. FlashInternImage-Large is trained on ImageNet-22K for 90 epochs and then fine-tuned on ImageNet-1K for 20 epochs. Other baselines share the same setting for a fair comparison." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.473, + 0.903 + ], + "angle": 0, + "content": "Results: Tab. 4 shows the results of models at various scales. Besides the model size and training/inference resolution, we also report each model's overall throughput (number" + }, + { + "type": "table", + "bbox": [ + 0.527, + 0.262, + 0.869, + 0.579 + ], + "angle": 0, + "content": "
ModelSizeScaleAccThroughput
Swin-T29M224281.31989 / 3619
ConvNeXt-T29M224282.12485 / 4305
InternImage-T30M224283.51409 / 1746
FlashInternImage-T30M224283.62316 / 3154 (+64%/ +80%)
Swin-S50M224283.01167 / 2000
ConvNeXt-S50M224283.11645 / 2538
InternImage-S50M224284.21044 / 1321
FlashInternImage-S50M224284.41625 / 2396 (+56%/ +81%)
Swin-B88M224283.5934 / 1741
ConvNeXt-B89M224283.81241 / 1888
InternImage-B97M224284.9779 / 1030
FlashInternImage-B97M224284.91174 / 1816 (+51%/ +76%)
Swin-L197M384287.3206 / 301
ConvNeXt-L198M384287.5252 / 436
InternImage-L223M384287.7158 / 214
ConvNeXt-XL350M384287.8170 / 299
InternImage-XL335M384288.0125 / 174
FlashInternImage-L223M384288.1248 / 401 (+57%/ +87%)
" + }, + { + "type": "table_caption", + "bbox": [ + 0.498, + 0.583, + 0.895, + 0.639 + ], + "angle": 0, + "content": "Table 4. Image classification performance on ImageNet-1K. We show relative speedup between FlashInternImage w/ DCNv4 and its InternImage counterparts. DCNv4 significantly improves the speed while shows state-of-the-art performance." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.651, + 0.897, + 0.846 + ], + "angle": 0, + "content": "of images per second) in FP32/FP16 data format. We use timm [39] implementation of ConvNeXt and Swin Transformer, which is faster than the original implementation. Equipped with DCNv4, FlashInternImage significantly improves the throughput of the InternImage counterpart over \\(50\\% \\sim 80\\%\\) and slightly improves the model performance. FlashInternImage now matches the speed of ConvNeXt with higher accuracy. It is noteworthy that FlashInternImage-S can outperform ConvNeXt-B (84.4% vs. 83.8%) while being faster than it, showing a better speed-accuracy trade-off. Moreover, the FlashInternImage-L can surpass ConvNeXt-XL and InternImage-XL and being \\(30\\% \\sim 130\\%\\) (401 vs. 174) faster, showing the effectiveness of our DCNv4 module." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.853, + 0.895, + 0.903 + ], + "angle": 0, + "content": "4.3. Downstream Tasks with High-Resolution Input We evaluate the performance of DCNv4 on representative downstream perception tasks with high-resolution input, in" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.958 + ], + "angle": 0, + "content": "5657" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.086, + 0.09, + 0.465, + 0.289 + ], + "angle": 0, + "content": "
Model#paramFPSMask R-CNN
1×3×+MS
APbAPmAPbAPm
Swin-T48M66 / 10642.739.346.041.6
ConvNeXt-T48M78 / 11344.240.146.241.7
InternImage-T49M54 / 6947.242.549.143.7
FlashInternImage-T49M72 / 10248.043.149.544.0
Swin-S69M45 / 7744.840.948.243.2
ConvNeXt-S70M54 / 8345.441.847.942.9
InternImage-S69M44 / 5647.843.349.744.5
FlashInternImage-S69M57 / 8349.244.050.544.9
Swin-B107M33 / 5946.942.348.643.3
ConvNeXt-B108M43 / 7047.042.748.543.5
InternImage-B115M33 / 4348.844.050.344.8
FlashInternImage-B115M44 / 6750.144.550.645.4
" + }, + { + "type": "table", + "bbox": [ + 0.086, + 0.3, + 0.465, + 0.42 + ], + "angle": 0, + "content": "
Model#paramFPSCascade Mask R-CNN
1×3×+MS
APbAPmAPbAPm
Swin-L253M20 / 2651.844.953.946.7
ConvNeXt-L255M26 / 4053.546.454.847.6
InternImage-L277M20 / 2654.947.756.148.5
ConvNeXt-XL407M21 / 3253.646.555.247.7
InternImage-XL387M16 / 2355.348.156.248.8
FlashInternImage-L277M26 / 3955.648.256.748.9
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.427, + 0.472, + 0.524 + ], + "angle": 0, + "content": "Table 5. Object detection and instance segmentation performance on COCO val2017. \\(\\mathrm{AP^b}\\) and \\(\\mathrm{AP^m}\\) denotes box AP and mask AP, respectively. \"MS\" means multi-scale training. FlashInternImage w/ DCNv4 models converge faster, clearly outperform other baselines with \\(1\\times\\) training schedule, and still maintain a leading position when training \\(3\\times\\) longer while being significantly faster than InternImage." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.535, + 0.472, + 0.61 + ], + "angle": 0, + "content": "cluding instance segmentation, semantic segmentation and 3D object detection. We keep all implementation details the same as InternImage and only change the backbone model. The backbone models are initialized from the ImageNet pretrained weights when training the downstream models." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.621, + 0.472, + 0.786 + ], + "angle": 0, + "content": "Instance Segmentation: We train FlashInternImage with two representative instance segmentation frameworks, Mask R-CNN [15] and Cascade Mask-RCNN [2], on COCO dataset [23] at \\(1 \\times (12\\) epochs) and \\(3 \\times (36\\) epochs) training schedules. The results are shown in Tab. 5. We also report FPS with batch size 16 in FP32/FP16 data format. FlashInternImage shows superior results on all model scales and training schedules, achieving a higher speed-accuracy tradeoff. For example, FlashInternImage-T/S surpasses all other models with the same scale and is on par with a larger InternImage-S/B while being \\(80\\% - 90\\%\\) faster." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.792, + 0.472, + 0.913 + ], + "angle": 0, + "content": "Semantic Segmentation: We train FlashInternImage with UperNet [41] on ADE20K [46] dataset for 160K iterations. We can draw a similar conclusion as instance segmentation from the results in Tab. 6, with FPS reported with batch size 16 in FP32/FP16. FlashInternImage w/ DCNv4 can achieve significantly faster speed and further improve the performance of InternImage across different model scales, resulting in a new state-of-the-art." + }, + { + "type": "table", + "bbox": [ + 0.517, + 0.09, + 0.878, + 0.359 + ], + "angle": 0, + "content": "
Modelcrop size#paramFPSmIoU (SS)mIoU (MS)
Swin-T512260M107 / 16844.545.8
ConvNeXt-T512260M120 / 18446.046.7
InternImage-T512259M100 / 13947.948.1
FlashInternImage-T512259M119 / 20649.350.3
Swin-S512281M89 / 14247.649.5
ConvNeXt-S512282M107 / 16448.749.6
InternImage-S512280M89 / 12350.150.9
FlashInternImage-S512280M107 / 18250.651.6
Swin-B5122121M77 / 12648.149.7
ConvNeXt-B5122122M95 / 14749.149.9
InternImage-B5122128M77 / 10450.851.3
FlashInternImage-B5122128M94 / 15752.052.6
Swin-L6402234M59 / 9952.153.5
ConvNeXt-L6402235M73 / 11753.253.7
InternImage-L6402256M56 / 7853.954.1
ConvNeXt-XL6402391M53 / 7553.654.0
InternImage-XL6402368M47 / 6755.055.3
FlashInternImage-L6402256M71 / 12255.656.0
" + }, + { + "type": "table_caption", + "bbox": [ + 0.498, + 0.363, + 0.895, + 0.434 + ], + "angle": 0, + "content": "Table 6. Semantic segmentation performance on the ADE20K validation set. All models are trained with UperNet. \"SS\" and \"MS\" denote single-scale and multi-scale testing, respectively. FPS is reported with single-scale testing. FlashInternImage w/ DCNv4 achieves superior performance with competitive speed." + }, + { + "type": "table", + "bbox": [ + 0.541, + 0.437, + 0.848, + 0.509 + ], + "angle": 0, + "content": "
ModelNDSmAPFPS†FPS
InternImage-B62.054.08.02.7
InternImage-XL63.455.64.02.0
FlashInternImage-S61.755.516.84.1
FlashInternImage-B63.157.412.13.8
" + }, + { + "type": "table_caption", + "bbox": [ + 0.498, + 0.513, + 0.895, + 0.597 + ], + "angle": 0, + "content": "Table 7. 3D detection performance of BEVFormer v2 on nuScenes test set. We report backbone FPS (denoted with \\(\\dagger\\)), overall FPS results with underoptimized head implementation are also added for reference. With on-par NDS and higher mAP results, FlashInternImage can be \\(50 - 90\\%\\) faster than InternImage baselines or \\(200\\% - 300\\%\\) when only considering the backbone." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.613, + 0.895, + 0.838 + ], + "angle": 0, + "content": "3D Detection: We further test DCNv4 on the camera-based 3D object detection task in the autonomous driving scenario, We train BEVFormer v2 [43], a state-of-the-art multi-camera 3D object detector, with FlashInternImage-Small and Base backbone models on the nuScenes dataset for 24 epochs. We report results on the nuScenes test set in Tab. 7 with FPS for each model. We note that the header parts, such as the BEV encoder and object decoder in BEV-Former v2, are underoptimized and take more than \\(50\\%\\) of the running time (and even more with a faster backbone); thus, we also report the FPS for the backbone for a clearer illustration. Our results show that when only considering the backbone, FlashInternImage can be twice or even three times faster than the InternImage backbone with an on-par performance, greatly increasing the model efficiency." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.848, + 0.783, + 0.865 + ], + "angle": 0, + "content": "4.4. DCNv4 as a Universal Operator" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.895, + 0.901 + ], + "angle": 0, + "content": "Drop-in replacement in other vision backbones : Weverify whether DCNv4 can still work well in architectures de-" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5658" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.119, + 0.09, + 0.425, + 0.179 + ], + "angle": 0, + "content": "
ModelIN-1K AccThroughput
ConvNeXt-B83.81241 / 1888
ConvNeXt-B + DCNv484.01495 / 2513 (+20% / +33%)
ViT-B81.81683 / 2781†
ViT-B + DCNv481.92092 / 3261 (+24% / +17%)
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.183, + 0.471, + 0.239 + ], + "angle": 0, + "content": "Table 8. DCNv4 in other architecture. We show supervised learning results on ImageNet-1K and throughput. DCNv4 achieves higher throughput with comparable accuracy. \\(\\dagger\\) denotes testing with the advanced FlashAttention-2 implementation." + }, + { + "type": "table", + "bbox": [ + 0.12, + 0.247, + 0.425, + 0.303 + ], + "angle": 0, + "content": "
Model#paramFIDFPS
U-Net860M2.944.82
U-Net + DCNv4566M2.444.92
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.307, + 0.472, + 0.364 + ], + "angle": 0, + "content": "Table 9. Class conditional generation on ImageNet 256x256. Latent diffusion models are trained from scratch with U-Net. We replace the convolution layer in the models with DCNv4. DCNv4 can achieve better FID results without any hyperparameter tuning." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.376, + 0.472, + 0.542 + ], + "angle": 0, + "content": "signed with other operators, such as ConvNeXt and ViT. To achieve that, we replace the attention module and depthwise convolution layer with DCNv4 in ViT and ConvNeXt and perform supervised learning on ImageNet-1K without changing any other architecture and hyperparameters, similar to FlashInternImage and InternImage. The results are shown in Tab. 8. We can see that on these architectures, which are carefully tuned for the specific operators, our DCNv4 can perform equally well. Thanks to the fast speed of DCNv4, the new model can even achieve better throughput, showcasing the superior performance of DCNv4." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.552, + 0.472, + 0.822 + ], + "angle": 0, + "content": "Drop-in replacement in diffusion model: DCN has been recognized to be an effective operator for perception tasks. As generative models become a fundamental tool for AI-generated content (AIGC), we are also curious if DCNv4 can work well on generation tasks with diffusion-based generative models. Specifically, we choose the U-Net [30] used in Stable Diffusion [29] as our baselines and replace the attention module and regular \\(3 \\times 3\\) convolution in U-Net. We use U-ViT's codebase, follow its training schedule, and train a latent diffusion model based on the image latent extracted from an image autoencoder provided by Stable Diffusion. We show the results in Tab. 9. We can see that DCNv4 also works well in generative modeling, achieving better results in terms of FID/Throughput with fewer parameters compared to regular convolution in U-Net. Notice that the architecture/hyperparameters may not be optimal for DCNv4, and it is possible that re-designing the models or searching for new hyperparameters for DCNv4 will give better results." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.832, + 0.233, + 0.849 + ], + "angle": 0, + "content": "4.5. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.471, + 0.903 + ], + "angle": 0, + "content": "We conduct ablation studies in our optimization choice described in Sec. 3.2. The results are shown in Tab. 10. The time in the table is obtained with \\(56 \\times 56 \\times 128\\) input with" + }, + { + "type": "table", + "bbox": [ + 0.533, + 0.09, + 0.858, + 0.2 + ], + "angle": 0, + "content": "
Implementation variantModuleKernel
Original DCNv33.281.45
- micro design2.121.45
- redundant memory access2.201.53
- redundant computation2.181.51
- redundant memory instr.1.280.606
- half-precision format0.8730.404
" + }, + { + "type": "table_caption", + "bbox": [ + 0.498, + 0.204, + 0.895, + 0.247 + ], + "angle": 0, + "content": "Table 10. Ablation studies of DCN's runtime (ms). We show how to achieve DCNv4 (gray row) from the original DCNv3 implementation and how different design choices affect the speed." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.256, + 0.895, + 0.614 + ], + "angle": 0, + "content": "batch size 64 and 4 groups (32 channels per group). We first remove the softmax operation and improve the micro design, which means we merge the two linear layers into one and remove costly layer norm and GELU activation in offset/aggregation weight computing, simplifying the overall modules and increasing the speed. We then start modifying the kernel implementation. First, we change the parallel execution pattern and let each thread process 8 channels instead of 1 channel, thus, unnecessary memory access on loading sampling offset and aggregation weight values from the GPU memory can be saved. As we expected, solely applying this change will not increase the speed as the degree of parallelism decreases, and each thread's workload increases 8 times now. The latency is increased instead. Eliminating redundant computation by reusing the bilinear interpolation coefficient (4th row) can save some time but is insignificant. Removing the redundant memory instruction via vectorized load/store can greatly reduce the workload of each thread and largely accelerate the GPU kernel speed (5th row). Using a half-precision datatype, which halves the number of bytes the kernel needs to read/write, further increases the data throughput, as shown in the 6th row. In the end, we reach the final DCNv4 design, which is three times more efficient than the original implementation." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.626, + 0.62, + 0.641 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.647, + 0.895, + 0.828 + ], + "angle": 0, + "content": "We present Deformable Convolution v4 (DCNv4), an efficient dynamic and sparse operator. By rethinking the dynamic property in deformable convolution and streamlining memory access, DCNv4 emerges as a much faster and more effective operator than its predecessor DCNv3. DCNv4-equipped FlashInternImage backbone not only enhances speed but also improves performance across various vision tasks. We further show DCNv4's versatility and effectiveness as a universal operator by integrating it into state-of-the-art architecture like ConvNeXt and ViT with improved throughput and accuracy; and it also works well in latent diffusion model, showing its potential to enhance generative models." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.838, + 0.661, + 0.856 + ], + "angle": 0, + "content": "Acknowledgement" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.859, + 0.895, + 0.901 + ], + "angle": 0, + "content": "The work is supported by the National Key R&D Program of China (NO. 2022ZD0161300), by the National Natural Science Foundation of China (62376134)." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5659" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.12, + 0.174, + 0.136 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.145, + 0.47, + 0.2 + ], + "angle": 0, + "content": "[1] Saehyun Ahn, Jung-Woo Chang, and Suk-Ju Kang. An efficient accelerator design methodology for deformable convolutional networks. In 2020 IEEE International Conference on Image Processing (ICIP), pages 3075-3079. IEEE, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.202, + 0.472, + 0.256 + ], + "angle": 0, + "content": "[2] Zhaowei Cai and Nuno Vasconcelos. Cascade r-cnn: Delving into high quality object detection. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6154-6162, 2018. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.258, + 0.472, + 0.325 + ], + "angle": 0, + "content": "[3] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In European conference on computer vision, pages 213-229. Springer, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.328, + 0.472, + 0.396 + ], + "angle": 0, + "content": "[4] Bowen Cheng, Ishan Misra, Alexander G Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1290–1299, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.398, + 0.472, + 0.453 + ], + "angle": 0, + "content": "[5] Sharan Chetlur, Cliff Woolley, Philippe Vandermersch, Jonathan Cohen, John Tran, Bryan Catanzaro, and Evan Shelhamer. codnn: Efficient primitives for deep learning. arXiv preprint arXiv:1410.0759, 2014. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.454, + 0.472, + 0.508 + ], + "angle": 0, + "content": "[6] François Chollet. Xception: Deep learning with depthwise separable convolutions. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1251-1258, 2017. 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.51, + 0.472, + 0.564 + ], + "angle": 0, + "content": "[7] Jifeng Dai, Haozhi Qi, Yuwen Xiong, Yi Li, Guodong Zhang, Han Hu, and Yichen Wei. Deformable convolutional networks. In Proceedings of the IEEE international conference on computer vision, pages 764-773, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.566, + 0.472, + 0.606 + ], + "angle": 0, + "content": "[8] Tri Dao. Flashattention-2: Faster attention with better parallelism and work partitioning. arXiv preprint arXiv:2307.08691, 2023.5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.608, + 0.472, + 0.662 + ], + "angle": 0, + "content": "[9] Tri Dao, Dan Fu, Stefano Ermon, Atri Rudra, and Christopher Ré. Flashattention: Fast and memory-efficient exact attention with io-awareness. Advances in Neural Information Processing Systems, 35:16344-16359, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.664, + 0.47, + 0.718 + ], + "angle": 0, + "content": "[10] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248–255. IEEE, 2009. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.72, + 0.47, + 0.787 + ], + "angle": 0, + "content": "[11] Xiaohan Ding, Xiangyu Zhang, Jungong Han, and Guiguang Ding. Scaling up your kernels to 31x31: Revisiting large kernel design in cnns. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11963-11975, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.79, + 0.47, + 0.871 + ], + "angle": 0, + "content": "[12] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations, 2020. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.47, + 0.901 + ], + "angle": 0, + "content": "[13] Qi Han, Zejia Fan, Qi Dai, Lei Sun, Ming-Ming Cheng, Jiaying Liu, and Jingdong Wang. On the connection between" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.145, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "local attention and dynamic depth-wise convolution. arXiv preprint arXiv:2106.04263, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.123, + 0.894, + 0.178 + ], + "angle": 0, + "content": "[14] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.18, + 0.894, + 0.221 + ], + "angle": 0, + "content": "[15] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In Proceedings of the IEEE international conference on computer vision, 2017. 1, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.224, + 0.894, + 0.264 + ], + "angle": 0, + "content": "[16] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. ImageNet classification with deep convolutional neural networks. Advances in neural information processing systems, 2012. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.267, + 0.894, + 0.32 + ], + "angle": 0, + "content": "[17] Yann LeCun, Léon Bottou, Yoshua Bengio, and Patrick Haffner. Gradient-based learning applied to document recognition. Proceedings of the IEEE, 86(11):2278-2324, 1998. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.324, + 0.892, + 0.392 + ], + "angle": 0, + "content": "[18] Youngwan Lee, Joong-won Hwang, Sangrok Lee, Yuseok Bae, and Jongyoul Park. An energy andgpu-computation efficient backbone network for real-time object detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops, pages 0–0, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.394, + 0.894, + 0.475 + ], + "angle": 0, + "content": "[19] Feng Li, Hao Zhang, Huaizhe Xu, Shilong Liu, Lei Zhang, Lionel M Ni, and Heung-Yeung Shum. Mask dino: Towards a unified transformer-based framework for object detection and segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3041-3050, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.479, + 0.894, + 0.533 + ], + "angle": 0, + "content": "[20] Yanghao Li, Hanzi Mao, Ross Girshick, and Kaiming He. Exploring plain vision transformer backbones for object detection. In European Conference on Computer Vision, pages 280-296. Springer, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.536, + 0.894, + 0.604 + ], + "angle": 0, + "content": "[21] Zhiqi Li, Wenhai Wang, Hongyang Li, Enze Xie, Chonghao Sima, Tong Lu, Yu Qiao, and Jifeng Dai. Bevformer: Learning bird's-eye-view representation from multi-camera images via spatiotemporal transformers. In European conference on computer vision, pages 1-18. Springer, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.607, + 0.894, + 0.687 + ], + "angle": 0, + "content": "[22] Zhiqi Li, Wenhai Wang, Enze Xie, Zhiding Yu, Anima Anandkumar, Jose M Alvarez, Ping Luo, and Tong Lu. Panoptic segformer: Delving deeper into panoptic segmentation with transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1280-1289, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.691, + 0.894, + 0.773 + ], + "angle": 0, + "content": "[23] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.776, + 0.894, + 0.83 + ], + "angle": 0, + "content": "[24] Yingfei Liu, Tiancai Wang, Xiangyu Zhang, and Jian Sun. Petr: Position embedding transformation for multi-view 3d object detection. In European Conference on Computer Vision, pages 531-548. Springer, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.833, + 0.894, + 0.901 + ], + "angle": 0, + "content": "[25] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10012-10022, 2021. 1, 2, 3, 5, 6" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.894, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "5660" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.161 + ], + "angle": 0, + "content": "[26] Zhuang Liu, Hanzi Mao, Chao-Yuan Wu, Christoph Feichtenhofer, Trevor Darrell, and Saining Xie. A convnet for the 2020s. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11976-11986, 2022. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.164, + 0.472, + 0.207 + ], + "angle": 0, + "content": "[27] Ningning Ma, Xiangyu Zhang, Hai-Tao Zheng, and Jian Sun. Shufflenet v2: Practical guidelines for efficient cnn architecture design. In ECCV, 2018. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.208, + 0.472, + 0.288 + ], + "angle": 0, + "content": "[28] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. Advances in neural information processing systems, 32, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.292, + 0.472, + 0.362 + ], + "angle": 0, + "content": "[29] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 1, 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.363, + 0.472, + 0.446 + ], + "angle": 0, + "content": "[30] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical Image Computing and Computer-Assisted Intervention-MICCAI 2015: 18th International Conference, Munich, Germany, October 5-9, 2015, Proceedings, Part III 18, pages 234-241. Springer, 2015. 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.447, + 0.472, + 0.531 + ], + "angle": 0, + "content": "[31] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in Neural Information Processing Systems, 35:36479-36494, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.532, + 0.472, + 0.574 + ], + "angle": 0, + "content": "[32] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.576, + 0.472, + 0.632 + ], + "angle": 0, + "content": "[33] Robin Strudel, Ricardo Garcia, Ivan Laptev, and Cordelia Schmid. Segmenter: Transformer for semantic segmentation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 7262-7272, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.633, + 0.472, + 0.702 + ], + "angle": 0, + "content": "[34] Hugues Thomas, Charles R Qi, Jean-Emmanuel Deschaud, Beatz Marcotegui, François Goulette, and Leonidas J Guibas. Kpconv: Flexible and deformable convolution for point clouds. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6411–6420, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.704, + 0.472, + 0.76 + ], + "angle": 0, + "content": "[35] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.761, + 0.472, + 0.831 + ], + "angle": 0, + "content": "[36] Ashish Vaswani, Prajit Ramachandran, Aravind Srinivas, Niki Parmar, Blake Hechtman, and Jonathon Shlens. Scaling local self-attention for parameter efficient visual backbones. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12894–12904, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.832, + 0.472, + 0.902 + ], + "angle": 0, + "content": "[37] Shenlong Wang, Simon Suo, Wei-Chiu Ma, Andrei Pokrovsky, and Raquel Urtasun. Deep parametric continuous convolutional neural networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2589-2597, 2018. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.16 + ], + "angle": 0, + "content": "[38] Wenhai Wang, Jifeng Dai, Zhe Chen, Zhenhang Huang, Zhiqi Li, Xizhou Zhu, Xiaowei Hu, Tong Lu, Lewei Lu, Hongsheng Li, et al. Internimage: Exploring large-scale vision foundation models with deformable convolutions. In CVPR, 2023. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.163, + 0.894, + 0.203 + ], + "angle": 0, + "content": "[39] Ross Wightman. Pytorch image models. https://github.com/rwrightman/pytorch-image-models, 2019.6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.205, + 0.892, + 0.247 + ], + "angle": 0, + "content": "[40] Felix Wu, Angela Fan, Alexei Baevski, Yann N Dauphin, and Michael Auli. Pay less attention with lightweight and dynamic convolutions. arXiv preprint arXiv:1901.10430, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.248, + 0.892, + 0.302 + ], + "angle": 0, + "content": "[41] Tete Xiao, Yingcheng Liu, Bolei Zhou, Yuning Jiang, and Jian Sun. Unified perceptual parsing for scene understanding. In Proceedings of the European conference on computer vision (ECCV), pages 418-434, 2018. 1, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.304, + 0.892, + 0.344 + ], + "angle": 0, + "content": "[42] Yuwen Xiong, Mengye Ren, Renjie Liao, Kelvin Wong, and Raquel Urtasun. Deformable filter convolution for point cloud reasoning. arXiv preprint arXiv:1907.13079, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.346, + 0.894, + 0.442 + ], + "angle": 0, + "content": "[43] Chenyu Yang, Yuntao Chen, Hao Tian, Chenxin Tao, Xizhou Zhu, Zhaoxiang Zhang, Gao Huang, Hongyang Li, Yu Qiao, Lewei Lu, et al. Bevformer v2: Adapting modern image backbones to bird's-eye-view recognition via perspective supervision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17830-17839, 2023. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.444, + 0.894, + 0.498 + ], + "angle": 0, + "content": "[44] Xiaohua Zhai, Alexander Kolesnikov, Neil Houlsby, and Lucas Beyer. Scaling vision transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12104-12113, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.499, + 0.892, + 0.554 + ], + "angle": 0, + "content": "[45] Hao Zhang, Feng Li, Shilong Liu, Lei Zhang, Hang Su, Jun Zhu, Lionel M Ni, and Heung-Yeung Shum. Dino: Detr with improved denoising anchor boxes for end-to-end object detection. arXiv preprint arXiv:2203.03605, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.555, + 0.892, + 0.611 + ], + "angle": 0, + "content": "[46] Bolei Zhou, Hang Zhao, Xavier Puig, Sanja Fidler, Adela Barriuso, and Antonio Torralba. Scene parsing through ade20k dataset. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 633-641, 2017. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.611, + 0.892, + 0.667 + ], + "angle": 0, + "content": "[47] Xizhou Zhu, Han Hu, Stephen Lin, and Jifeng Dai. Deformable convnets v2: More deformable, better results. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.667, + 0.892, + 0.723 + ], + "angle": 0, + "content": "[48] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable detr: Deformable transformers for end-to-end object detection. In International Conference on Learning Representations, 2020. 2, 5" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.723 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "5661" + } + ] +] \ No newline at end of file diff --git a/2024/Efficient Deformable ConvNets_ Rethinking Dynamic and Sparse Operator for Vision Applications/d839f1b7-f6aa-4faa-bb5a-7c661a9aa4e5_origin.pdf b/2024/Efficient Deformable ConvNets_ Rethinking Dynamic and Sparse Operator for Vision Applications/d839f1b7-f6aa-4faa-bb5a-7c661a9aa4e5_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..df0c99ac217aefa2045c7e72b5be37e11a40806c --- /dev/null +++ b/2024/Efficient Deformable ConvNets_ Rethinking Dynamic and Sparse Operator for Vision Applications/d839f1b7-f6aa-4faa-bb5a-7c661a9aa4e5_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2526b6c63eb021da2aba0d0e8ff8377dc92db36c6ac9d8022b9552125a6dd0d6 +size 833245 diff --git a/2024/Efficient Deformable ConvNets_ Rethinking Dynamic and Sparse Operator for Vision Applications/full.md b/2024/Efficient Deformable ConvNets_ Rethinking Dynamic and Sparse Operator for Vision Applications/full.md new file mode 100644 index 0000000000000000000000000000000000000000..26cf771ed9be7de48f5d88137ead4f7febea73f1 --- /dev/null +++ b/2024/Efficient Deformable ConvNets_ Rethinking Dynamic and Sparse Operator for Vision Applications/full.md @@ -0,0 +1,279 @@ +# Efficient Deformable ConvNets: Rethinking Dynamic and Sparse Operator for Vision Applications + +Yuwen Xiong $^{*,1,2}$ Zhiqi Li $^{*,3,2}$ Yuntao Chen $^{*4}$ Feng Wang $^{*5}$ + +Xizhou Zhu $^{5,6}$ Jiapeng Luo $^{6}$ Wenhai Wang $^{7,2}$ Tong Lu $^{3}$ Hongsheng Li $^{7}$ + +Yu Qiao² Lewei Lu⁶ Jie Zhou⁵ Jifeng Dai⁵,² + +$^{1}$ University of Toronto $^{2}$ OpenGVLab, Shanghai AI Laboratory + +$^{3}$ Nanjing University $^{4}$ CAIR, HKISI, CAS $^{5}$ Tsinghua University + +$^{6}$ SenseTime Research $^{7}$ The Chinese University of Hong Kong + +https://github.com/OpenGVLab/DCNv4 + +# Abstract + +We introduce Deformable Convolution v4 (DCNv4), a highly efficient and effective operator designed for a broad spectrum of vision applications. DCNv4 addresses the limitations of its predecessor, DCNv3, with two key enhancements: 1. removing softmax normalization in spatial aggregation to enhance its dynamic property and expressive power and 2. optimizing memory access to minimize redundant operations for speedup. These improvements result in a significantly faster convergence compared to DCNv3 and a substantial increase in processing speed, with DCNv4 achieving more than three times the forward speed. DCNv4 demonstrates exceptional performance across various tasks, including image classification, instance and semantic segmentation, and notably, image generation. When integrated into generative models like U-Net in the latent diffusion model, DCNv4 outperforms its baseline, underscoring its possibility to enhance generative models. In practical applications, replacing DCNv3 with DCNv4 in the InternImage model to create FlashInternImage results in up to $80\%$ speed increase and further performance improvement without further modifications. The advancements in speed and efficiency of DCNv4, combined with its robust performance across diverse vision tasks, show its potential as a foundational building block for future vision models. + +# 1. Introduction + +In the field of computer vision, there is an ongoing debate about whether convolutional networks (ConvNets) or Transformers offer superior performance. In recent years, Trans + +![](images/0a203e8027418311ed674c2a129d80b85a2d8873b8ba87e04c7d45a298d9ee45.jpg) +(a) + +![](images/05016b33a235bfddc93728e971ce3b133882971c9880b5eec24bc9a4bd5b8091.jpg) +(b) +Figure 1. (a) We show relative runtime with DCNv3 as the baseline. DCNv4 shows significant speedup over DCNv3, and surpasses other common vision operators. (b) With the same network architecture, DCNv4 converges faster than other operators. + +former models [12, 25, 44] have achieved remarkable results in large vision models with the attention mechanism, showing the potential to overtake ConvNets. However, recent works such as InternImage [38] and ConvNeXt [26] demonstrate that ConvNet-based vision models retain robust performance, efficiency, simplicity, and suitable inductive bias for various downstream tasks [15, 41]. Notably, in domains like image generation [29, 31], models containing both convolutions and transformers remain the preferred approach. This situation brings to light the enduring value of convolution-based approaches. + +Building on convolution's strengths, Deformable Convolution v3 (DCNv3) – the core operator of the advanced ConvNet model InternImage – innovatively combines a sparse attention mechanism with convolution: it processes each output location in a sliding window manner with a small window size (e.g. $3 \times 3 = 9$ points) which acts as a local, sparse operator like convolution, while dynamically samples point with an adaptive range and aggregates the spatial features with input-dependent attention weights. With its + +small window size and ConvNet inductive bias, DCNv3 is expected to achieve a faster convergence rate and lower inference latency, especially when compared to dense global [12] or local window-based [25] attention methods. + +Despite these advantages, DCN has not become the go-to solution for vision backbone models. This observation led us to investigate the lingering limitations of the DCN operator. The first thing we notice is the running speed. The slow speed of DCN is known to be a long-standing problem [1], as it introduces extra overhead on sampling non-nearby locations, making it not fit modern convolution algorithms. Our comparative analysis, illustrated in Fig. 1a, reveals that DCNv3 can be slower than a properly optimized dense global attention [9], highlighting the need for further optimization. Moreover, we find DCNv3 even converges slower than global attention at the initial backbone training phase, as shown in Fig. 1b, which is counter-intuitive as DCNv3 is equipped with ConvNet inductive bias. + +To overcome these challenges, we propose Deformable Convolution v4 (DCNv4), an innovative advancement to optimize the sparse DCN operator for practical efficiency. DCNv4 comes with a much faster implementation and an improved operator design to enhance its performance, which we will elaborate on as follows: + +First, we conduct instruction-level kernel profiling for existing implementation and find that DCNv3 is already lightweight. The compute cost is less than $1\%$ , while memory access costs $99\%$ . This motivates us to revisit the operator implementation and find that many memory accesses in the DCN forward process are redundant and thus can be optimized, leading to a much faster DCNv4 implementation. + +Second, drawing inspiration from convolution's unbounded weight range, we find that softmax in spatial aggregation, a standard operation in dense attention, is unnecessary in DCNv3, as it is not a requirement for operators with a dedicated aggregation window for each location. Intuitively, softmax puts a bounded $0 \sim 1$ value range to the weight and will limit the expressive power of the aggregation weight. This insight led us to remove the softmax in DCNv4, enhancing its dynamic property and improving its performance. + +As a result, DCNv4 converges significantly faster than DCNv3 and accelerates forward speed by more than $3 \times$ . This improvement allows DCNv4 to fully leverage its sparse property and become one of the fastest core vision operators. + +We further replace DCNv3 in InternImage with DCNv4, creating FlashInternImage. Remarkably, FlashInternImage achieves a $50 \sim 80\%$ speed increase compared to InternImage without any additional modifications. This enhancement positions FlashInternImage as one of the fastest modern vision backbone networks while maintaining superior performance. With the help of DCNv4, FlashInternImage significantly improves the convergence speed in ImageNet classification [10] and transfer learning settings and further + +demonstrates improved performance in downstream tasks. + +Furthermore, DCNv4 shows potential as a universal vision operator in various architectures and tasks. We integrate DCNv4 into other modern backbone architectures, including ConvNeXt [26] and ViT [12], replacing depthwise convolution [6] and dense self-attention layers [35]. Surprisingly, without any hyperparameter adjustments, these meticulously designed networks with DCNv4 perform on par while being much faster, showing the efficacy and efficiency of the dynamic, sparse DCNv4. Moreover, we explore the potential of DCNv4 in generative models as a new application domain. Specifically, we apply it in the U-Net [30] architecture used in diffusion models [29], replacing regular convolution with DCNv4. Our experimental results show that DCNv4 can work better than the baselines in image generation, showing great potential for using DCN to improve generative models. + +We have released the DCNv4 implementation to facilitate future research in the vision community. + +# 2. Related Work + +Core operators in vision models: The standard convolution [17] stands as the most prevalent and impactful operator, forming the backbone of the majority of computer vision architectures [14, 16, 32]. Nevertheless, a myriad of operators, each with unique characteristics, collectively play a crucial role in the development of computer vision. Depthwise separable convolution (DWConv) [6] separates the spatial and channel operations, and has been pivotal in developing lightweight and efficient models [26, 27]. RepLKNet [11] illustrates that a purely convolutional network, leveraging large-kernel depth-wise convolutions, can attain competitive performance in both efficiency and effectiveness. Deformable Convolution (DCN) series [7, 38, 47] significantly leaps the adaptability of convolution by adding learnable offsets to the convolutions kernels. Contrary to convolutions, attention mechanisms [35] possess the capacity to model long-range dependencies and have been successfully adopted in various computer vision tasks [3, 12, 24, 33]. Window attention [25, 36] reduces the computational complexity inherent in vanilla attention by restricting the attention operation to a fixed-size window. To mitigate the high computational complexity associated with vanilla attention, deformable attention [48] enables each query to concentrate on a select number of key sampling points, with dynamically determined locations and weights. This efficient method is widely used in the following arts perception methods [4, 19, 21, 22, 43, 45]. DynamicConv [40] and dynamic-DWNet [13] augment DWConv by incorporating dynamic weights, thereby enabling the use of instance-specific weights that adapt dynamically. For non-grid structured data, sparse operators [34, 37, 42] utilize dynamic weights obtained via bilinear interpolation or in a parametric way. + +Memory access cost in vision backbones: As underscored in previous studies [18, 27], FLOPs, although a fre + +
Model5th EP10th Ep20th Ep50th Ep300th Ep
ConvNeXt29.953.566.174.883.8
ConvNeXt + softmax8.5 (-21.4)25.3 (-28.2)51.1 (-15.0)69.1 (-5.7)81.6 (-2.2)
+ +Table 1. ImageNet-1K accuracy at different training epochs. Softmax on the convolution weights significantly affects the convergence speed and the final performance for the ConvNeXt model. + +sequently used metric to measure model complexity, do not accurately represent the model's speed or latency. In practical scenarios, the running speed of a model is influenced by multiple factors, not just FLOPs. Memory access cost plays a significant role in this context [27]. Flash-Attention [9], by reducing the number of accesses to High Bandwidth Memory (HBM), achieves a significantly faster speed in practice despite having higher FLOPs compared to vanilla attention. Although DCN operators do not exhibit a disadvantage in terms of FLOPs, their latency is considerably longer compared to DW-Conv, under the same FLOPs budget, predominantly due to substantial memory access costs. In this work, we conduct a thorough analysis and optimization of the memory access costs associated with the DCN operators, significantly accelerating the DCN's running speed. + +# 3. Method + +# 3.1. Rethinking the Dynamic Property in Deformable Convolution + +Revisiting DCNv3: Given an input $\mathbf{x} \in \mathbb{R}^{H \times W \times C}$ with height $H$ , width $W$ and channel $C$ , the DCNv3 operation with $K$ points is defined in Eq. (2) for each point $p_0$ : + +$$ +\mathbf {y} _ {g} = \sum_ {k = 1} ^ {K} \mathbf {m} _ {g k} \mathbf {x} _ {g} \left(p _ {0} + p _ {k} + \Delta p _ {g k}\right), \tag {1} +$$ + +$$ +\mathbf {y} = \operatorname {c o n c a t} \left(\left[ \mathbf {y} _ {1}, \mathbf {y} _ {2}, \dots , \mathbf {y} _ {G} \right], \text {a x i s} = - 1\right), \tag {2} +$$ + +where $G$ denotes the number of spatial aggregation groups. For the $g$ -th group, $\mathbf{x}_g,\mathbf{y}_g\in \mathbb{R}^{H\times W\times C'}$ represents the sliced input/output feature map with $C' = C / G$ represents the group dimension; $\mathbf{m}_{gk}\in \mathbb{R}$ denotes the spatial aggregation weights (also known as modulation scalar) of the $k$ -th sampling point in the $g$ -th group, conditioned on the input $\mathbf{x}$ and normalized by the softmax function along the dimension $K$ ; $p_k$ denotes the $k$ -th location of the pre-defined grid sampling $\{(-1, - 1),(-1,0),\dots,(0, + 1),\dots,(+1, + 1)\}$ as in regular convolutions and $\Delta p_{gk}$ is the offset corresponding to the grid sampling location $p_k$ in the $g$ -th group. A $1\times 1$ point-wise convolution on $\mathbf{x}$ and $\mathbf{y}$ can be applied before and after the DCNv3 operator to enhance the model's expressive power, following separable convolution [6]. DCNv3 is a combination of convolution and attention: on the one hand, it processes the input data in a sliding window manner, which follows convolution and inherent its inductive bias; on the other hand, the sampling offset $\Delta p$ and spatial aggregation weight $\mathbf{m}$ are dynamically predicted from the input feature, + +![](images/0da0a2416ed9ee5e5ba643bd4e37433e46ab5f66b82645ab00246b074a464e2e.jpg) +(a) Attention + +![](images/8a4cd00fb02996c23fda3a714e41a20b56e5df9ee73adb56f7d181101e90237f.jpg) +Figure 2. Comparisons of core operators in spatial aggregation for query pixels on different locations within the same channel. (a) Attention and (b) DCNv3 use bounded (range from $0 \sim 1$ ) dynamic weights to aggregate spatial features, while the window (sampling point set) for attention is the same, and DCNv3 uses a dedicated window for each location. (c) Convolution has a more flexible unbounded value range for aggregation weights and uses a dedicated sliding window for each location, but the window shape and aggregation weights are input-independent. (d) DCNv4 combines their advantages, using an adaptive aggregation window and dynamic aggregation weights with an unbounded value range. + +![](images/33e00d201f52315ef0a993c478a5b8dacac3f7c24d1bcef7ad3f7498f1b2da04.jpg) +(b) DCNv3 + +![](images/a00f4168748400aa3f4bd5805a6d86f8fe61347d8863fa8041ffaf80c1424ece.jpg) + +showing its dynamic property and making it more like an attention mechanism. We compare different operators where each has its own property, as illustrated in Fig. 2 + +Softmax normalization: A key difference between convolution and DCNv3 is that DCNv3 normalizes $\mathbf{m}$ , the spatial aggregation weights, with a softmax function, following the convention of scaled dot-product self-attention. Conversely, convolution does not employ softmax over its weights and still works well. The reason why attention needs a softmax is straightforward: scaled dot-product self-attention with $Q, K, V \in \mathbb{R}^{N \times d}$ is defined with a formulation: + +$$ +\operatorname {s o f t m a x} \left(\frac {1}{\sqrt {d}} Q K ^ {\top}\right) V, \tag {3} +$$ + +where $N$ is the number of points in the same attention window (can be either global [12] or local [25]), $d$ is the hidden dimension, $Q, K, V$ are the query, key, and value matrices computed from the input. Softmax operation is required in Eq. (3) for attention; without softmax, $K^{\top}V \in \mathbb{R}^{d \times d}$ can be calculated first, and it degrades to a linear projection for all queries in the same attention window, resulting in degenerated performance. However, for convolutional operators like depthwise convolution and DCNv3 where each point has its own dedicated aggregation window and the values in each aggregation window are already different and there is no "key" concept, such degradation issues no longer exist, and the normalization becomes unnecessary. In fact, normalizing convolution weights within a fixed 0-1 range using softmax could impose a significant limitation on the operator's expressive power and make the learning slower. + +To confirm this hypothesis, we train a ConvNeXt model and apply softmax to the $7 \times 7$ window of the depthwise + +convolution weights before convolution forward. We observe a remarkable decline in model performance as well as convergence speed from results in Tab. 1. This suggests that for operators with a dedicated aggregation window on each location like convolution or DCN, aggregation weights with an unbounded range offer better expressive power than softmax-normalized, bounded-range weights. + +Enhancing dynamic property: Motivated by this observation, we remove the softmax normalization in DCNv3, transforming the modulation scalars ranging from 0 to 1 to unbounded dynamic weights similar to convolution. As shown in Fig. 2, this alteration further amplifies the dynamic property of DCN, where other operators have certain limits, such as bounded value range (attention/DCNv3) or fixed aggregation window with input-independent aggregation weights (convolution). Fig. 1b shows that by making this change, DCNv4 converges significantly faster than DCNv3 and other common operators, including convolution and attention. Results in Sec. 4 further showcase that DCNv4 works well in both pre-training and transfer learning settings. + +# 3.2. Speeding up DCN + +Theoretically, DCN, as a sparse operator with $3 \times 3$ window, should act faster than other common operators that employ larger window sizes, like dense attention or $7 \times 7$ depthwise convolution. However, we find that this is not the case, as shown in Fig. 1a. In this subsection, we first conduct a theoretical analysis of GPU efficiency, showing a large variance in memory access cost depending on how we read the memory. We further perform optimization based on our observations, significantly improving the speed of DCN by saving additional memory instruction and bringing the speed advantage of being a sparse operator into reality. + +Theoretical analysis of GPU efficiency Our study begins with a theoretical examination of the DCNv3 operator's computational behavior. We employ the roofline model to evaluate its performance, focusing on theoretical FLOPs and memory access cost. For an input and output tensor of shape $(H,W,C)$ , the DCNv3 operator requires $36HWC$ FLOPs, where $3\times 3$ represents the convolution kernel's spatial dimensions and the factor of 4 accounts for the bilinear interpolation at each sampling point. + +Following the framework outlined in [27], DCNv3's memory access cost is calculated as $2HWC + 27HWG$ . The first term corresponds to the input/output feature map size and the second to the DCNv3's offset and aggregation weights with $G$ groups. We approximate $G$ as $C / 16$ assuming a group dimension of 16, resulting in approximately $3.7HWC$ memory access cost. However, this assumes an ideal scenario of infinite cache and a single memory read for each value, which is often unrealistic in parallel computing environments where concurrent thread execution necessitates simultaneous data access. + +![](images/a17fa7dfce35b4a9b16f016b29f5033fa5c616e4468c2f79f26bc2189c681b73.jpg) +Figure 3. Illustration of our optimization. In DCNv4, we use one thread to process multiple channels in the same group that shares sampling offset and aggregation weights. Workloads like memory reading and bilinear interpolation coefficient computation can be reduced, and multiple memory access instructions can be merged. + +To estimate the maximum memory access requirement, we consider a scenario devoid of cache, where each output location requires fresh memory reads and involves 36 reads for bilinear interpolation, 27 for offset/aggregation weights, and one write operation, resulting in a memory access cost of $64HWC$ . This is 17 times larger than the ideal case. + +This analysis reveals a substantial gap in the ratio of computation-to-memory access (ranging from 0.6 to 9.7), highlighting the significant potential for memory access optimization. Notably, despite DCNv3's use of input-dependent, dynamic offsets causing non-deterministic memory access, one deterministic thing is that channels within the same group share offset values. This leads us to propose a specific optimization strategy for enhancing DCNv3's speed. + +Eliminating redundant workload: In previous CUDA implementations of DCN kernel, for input with shape $(H,W,C)^1$ , offset $(H,W,G,K^2\times 2)$ and aggregation weight $(H,W,G,K^2)$ , we will create $H\times W\times C$ threads in total to maximize parallelism, where each thread processes one channel for one output location. Notably, the $D = C / G$ channels within each group share the same sampling offset and aggregation weight values for each output location. Using multiple threads to process these $D$ channels on the same output location is wasteful as different threads will read the same sampling offset and aggregation weight values from GPU memory multiple times, which is critical for a memory-bound operator. Processing multiple channels within the same group on each output location with one thread can eliminate these redundant memory read requests, greatly reducing memory bandwidth usage. As the sampling locations are the same, we can also only calculate the bilinear interpolation coefficient used in DCN once. Specifically, if each thread processes $D^{\prime}$ channels, the memory access cost for reading offset and aggregation weight, as well as the computation cost for calculating bilinear interpolation coefficient, can be reduced $D^{\prime}$ times. + +Eliminating redundant memory instructions: In practice, solely reusing threads for multiple channels will not + +
OperatorRuntime (ms)
56 × 56 × 12828 × 28 × 25614 × 14 × 5127 × 7 × 102414 × 14 × 768
Attention (torch)30.8 / 19.33.35 / 2.120.539 / 0.4480.446 / 0.1210.779 / 0.654
FlashAttention-2N/A / 2.46N/A / 0.451N/A / 0.123N/A / 0.0901N/A / 0.163
Window Attn (7 × 7)4.05 / 1.462.07 / 0.7701.08 / 0.4220.577 / 0.2391.58 / 0.604
DWConv (7 × 7, torch)2.02 / 1.981.03 / 1.000.515 / 0.5230.269 / 0.2610.779 / 0.773
DWConv (7 × 7, cuDNN)0.981 / 0.4380.522 / 0.2670.287 / 0.1530.199 / 0.1020.413 / 0.210
DCNv31.45 / 1.520.688 / 0.7110.294 / 0.2980.125 / 0.1260.528 / 0.548
DCNv40.606 / 0.4040.303 / 0.2300.145 / 0.1230.0730 / 0.06800.224 / 0.147
+ +Table 2. Op-level benchmark on standard input shape with various downsample rates. FP32/FP16 results are reported when the implementation is available. Our new DCNv4 can surpass all other commonly used operators under different input resolutions. + +see speed improvement. The reason is that when $D'$ increases, we create fewer threads and the workload of each thread now increases $D'$ times. This essentially reduces the degree of parallelism for the CUDA kernel. Luckily, the DCN kernel is now computationally lightweight as the bilinear interpolation only needs to be performed once for all $D'$ channels, and most of the workload is the memory instructions reading input values from different channels. When the memory layout is channel-last, and all $D'$ channel values are contiguous, we can leverage vectorized load: for example, to read four 32-bit float values from memory, instead of reading one 32-bit float value four times with four instructions, we can use a single instruction to load a 128-bit packed value at once, thus reducing the number of instructions and execution time of each thread. We can apply similar technique when writing the final results to GPU memory, minimizing the memory access time and increasing memory bandwidth utilization. Moreover, the modern half-precision data format (float16/bfloat16) halves the bytes that need to be loaded, which means the memory efficiency can be twice as much under the same memory bandwidth when using the half-precision format. However, we do not see speed improvement with half-precision data in the original DCNv3 implementation, possibly due to too much overhead on data access and computation, while in our new implementation, the speedup is significant. It is worth noting that the aforementioned optimization techniques can also be applied to DCNv1/v2 and deformable attention [48], as they share a similar performance bottleneck and issue. + +Micro design in DCN module: DCNv3 module introduces multiple micro designs; as the core kernel is optimized, their impact on the speed becomes non-negligible. We identify two points in DCNv3 designs that could be further optimized: first, after removing the softmax and transforming the modulation scalar into dynamic aggregation weights as mentioned in the previous paragraph. The linear layers for computing offset and dynamic weights can actually be combined into one linear layer. This reduces network fragmentation and eliminates extra overheads, such as kernel launching and synchronization, enhancing run-time efficiency on the GPU; second, in the original DCNv3 module design, a complex sub-network that consists of depthwise $3 \times 3$ conv, layer + +normalization (LN), GELU, and linear layer is used to compute offsets and dynamic weights. Following the design in Xception [6], we remove the additional LN-GELU layers and use the original separable convolution structure, further reducing running time. We empirically find that if latency is a higher priority, the depthwise convolution can also be removed with only a minor performance sacrifice. + +# 4. Experiments + +In this section, we verify the effectiveness of our proposed DCNv4 module from both speed as well as performance perspective. We benchmark the operator-level speed and integrate DCNv4 into the backbone to test the system-level performance. All speed test results are obtained with an NVIDIA A100 80G SXM GPU. We include additional experimental results and implementation details in supp. + +# 4.1. Speed Benchmark for Operators + +Settings: We conduct the op-level benchmark by solely measuring the running time of several representative operators building state-of-the-art vision backbone models, including full attention [35] implemented with PyTorch as well as the advanced FlashAttention-2 [8] implementation, window-based attention with window size $7 \times 7$ [25], depthwise convolution with $7 \times 7$ window, implemented by cuDNN [5] and ATen library from PyTorch [28], respectively. We only benchmark the core operation for spatial aggregation, and additional linear layers like qkv projection and output projection layers are excluded and not included in the runtime measurement. Please refer to supp. for a more comprehensive module-level comparison. We first consider a feature map shape generated from the standard $224 \times 224$ input resolution for image classification with 4, 8, 16, $32 \times$ downsample ratio as used by common hierarchical ConvNet/transformer backbones; we also add a feature map shape from isotropic backbone like ViT with a downsampling ratio 16 and larger hidden dimension. We further consider high-resolution inputs often used in downstream tasks like object detection. We set the input shape to be $800 \times 1280$ and $1024 \times 1024$ for the hierarchical feature map and isotropic feature map, respectively, as they are the common practice in object detection [15, 20]. Batch size is 64 and 1 for these two input sets, respectively. For operators with a head/group concept, we + +
OperatorRuntime (ms)
200 × 320 × 128100 × 160 × 25650 × 80 × 51225 × 40 × 102464 × 64 × 768
Attention (torch)OOM / OOM25.4 / 12.92.88 / 1.890.490 / 0.3094.17 / 2.57
FlashAttention-2N/A / 13.2N/A / 1.74N/A / 0.285N/A / 0.0797N/A / 0.437
Window Attn (7 × 7)1.33 / 0.5090.728 / 0.2910.426 / 0.1860.279 / 0.1650.673 / 0.272
DWConv (7 × 7, torch)0.634 / 0.6080.313 / 0.3150.167 / 0.1580.0943 / 0.08940.260 / 0.253
DWConv (7 × 7, cuDNN)0.331 / 0.2820.188 / 0.1680.114 / 0.1150.0817 / 0.08810.161 / 0.156
DCNv30.472 / 0.4930.244 / 0.2530.128 / 0.1320.0737 / 0.07670.194 / 0.199
DCNv40.210 / 0.1360.124 / 0.08950.0707 / 0.05890.0452 / 0.04260.103 / 0.0672
+ +set the dimension of each head/group to 32 and change the number of heads/groups when the hidden dimension varies. + +Results: We show the benchmark results on standard resolution and high-resolution input in Tab. 2 and Tab. 3. We report results with both FP32 and FP16 data formats unless the FP32 implementation is not available. Dense global attention implemented with PyTorch performs significantly slower when the input resolution is large and even out of memory. FlashAttention significantly improves the speed of attention and can be even faster than $7 \times 7$ window attention in certain cases, indicating the importance of proper optimization. However, it does not change the quadratic complexity of attention; when the input resolution is high, it still falls behind local/sparse operators like window attention or convolution. While DCNv3 can be faster than DWConv with plain implementation, it is slower than the heavily optimized cuDNN version. Instead, our DCNv4 can provide more than $3 \times$ speedup compared to DCNv3, greatly saving the running time. Moreover, DCNv4 can leverage the advantage of using a $3 \times 3$ sparse window to perform significantly faster than other baselines under different settings. + +# 4.2. Image Classification + +Settings: We evaluate the effectiveness of DCNv4 on ImageNet classification. We start from InternImage [38] as it shows state-of-the-art performance among ConvNet-based models. We replace the DCNv3 in InternImage with DCNv4 and create FlashInternImage. Other implementation details, including network architecture and hyperparameters, are kept the same as [38]. We compare Swin-Transformer and ConvNeXt which are two representative baselines in Transformer and ConvNet models. We follow the common practice [25, 26, 38] of training protocols, including data augmentation, preprocessing, optimizer and learning rate schedule, and train FlashInternImage-Tiny/Small/Base on ImageNet-1K for 300 epochs. FlashInternImage-Large is trained on ImageNet-22K for 90 epochs and then fine-tuned on ImageNet-1K for 20 epochs. Other baselines share the same setting for a fair comparison. + +Results: Tab. 4 shows the results of models at various scales. Besides the model size and training/inference resolution, we also report each model's overall throughput (number + +Table 3. Op-level benchmark on high-resolution input shape with various downsample rates. DCNv4 performs well as a sparse operator, surpassing all other baselines, while dense global attention is slow under this scenario. + +
ModelSizeScaleAccThroughput
Swin-T29M224281.31989 / 3619
ConvNeXt-T29M224282.12485 / 4305
InternImage-T30M224283.51409 / 1746
FlashInternImage-T30M224283.62316 / 3154 (+64%/ +80%)
Swin-S50M224283.01167 / 2000
ConvNeXt-S50M224283.11645 / 2538
InternImage-S50M224284.21044 / 1321
FlashInternImage-S50M224284.41625 / 2396 (+56%/ +81%)
Swin-B88M224283.5934 / 1741
ConvNeXt-B89M224283.81241 / 1888
InternImage-B97M224284.9779 / 1030
FlashInternImage-B97M224284.91174 / 1816 (+51%/ +76%)
Swin-L197M384287.3206 / 301
ConvNeXt-L198M384287.5252 / 436
InternImage-L223M384287.7158 / 214
ConvNeXt-XL350M384287.8170 / 299
InternImage-XL335M384288.0125 / 174
FlashInternImage-L223M384288.1248 / 401 (+57%/ +87%)
+ +Table 4. Image classification performance on ImageNet-1K. We show relative speedup between FlashInternImage w/ DCNv4 and its InternImage counterparts. DCNv4 significantly improves the speed while shows state-of-the-art performance. + +of images per second) in FP32/FP16 data format. We use timm [39] implementation of ConvNeXt and Swin Transformer, which is faster than the original implementation. Equipped with DCNv4, FlashInternImage significantly improves the throughput of the InternImage counterpart over $50\% \sim 80\%$ and slightly improves the model performance. FlashInternImage now matches the speed of ConvNeXt with higher accuracy. It is noteworthy that FlashInternImage-S can outperform ConvNeXt-B (84.4% vs. 83.8%) while being faster than it, showing a better speed-accuracy trade-off. Moreover, the FlashInternImage-L can surpass ConvNeXt-XL and InternImage-XL and being $30\% \sim 130\%$ (401 vs. 174) faster, showing the effectiveness of our DCNv4 module. + +4.3. Downstream Tasks with High-Resolution Input We evaluate the performance of DCNv4 on representative downstream perception tasks with high-resolution input, in + +
Model#paramFPSMask R-CNN
1×3×+MS
APbAPmAPbAPm
Swin-T48M66 / 10642.739.346.041.6
ConvNeXt-T48M78 / 11344.240.146.241.7
InternImage-T49M54 / 6947.242.549.143.7
FlashInternImage-T49M72 / 10248.043.149.544.0
Swin-S69M45 / 7744.840.948.243.2
ConvNeXt-S70M54 / 8345.441.847.942.9
InternImage-S69M44 / 5647.843.349.744.5
FlashInternImage-S69M57 / 8349.244.050.544.9
Swin-B107M33 / 5946.942.348.643.3
ConvNeXt-B108M43 / 7047.042.748.543.5
InternImage-B115M33 / 4348.844.050.344.8
FlashInternImage-B115M44 / 6750.144.550.645.4
+ +
Model#paramFPSCascade Mask R-CNN
1×3×+MS
APbAPmAPbAPm
Swin-L253M20 / 2651.844.953.946.7
ConvNeXt-L255M26 / 4053.546.454.847.6
InternImage-L277M20 / 2654.947.756.148.5
ConvNeXt-XL407M21 / 3253.646.555.247.7
InternImage-XL387M16 / 2355.348.156.248.8
FlashInternImage-L277M26 / 3955.648.256.748.9
+ +cluding instance segmentation, semantic segmentation and 3D object detection. We keep all implementation details the same as InternImage and only change the backbone model. The backbone models are initialized from the ImageNet pretrained weights when training the downstream models. + +Instance Segmentation: We train FlashInternImage with two representative instance segmentation frameworks, Mask R-CNN [15] and Cascade Mask-RCNN [2], on COCO dataset [23] at $1 \times (12$ epochs) and $3 \times (36$ epochs) training schedules. The results are shown in Tab. 5. We also report FPS with batch size 16 in FP32/FP16 data format. FlashInternImage shows superior results on all model scales and training schedules, achieving a higher speed-accuracy tradeoff. For example, FlashInternImage-T/S surpasses all other models with the same scale and is on par with a larger InternImage-S/B while being $80\% - 90\%$ faster. + +Semantic Segmentation: We train FlashInternImage with UperNet [41] on ADE20K [46] dataset for 160K iterations. We can draw a similar conclusion as instance segmentation from the results in Tab. 6, with FPS reported with batch size 16 in FP32/FP16. FlashInternImage w/ DCNv4 can achieve significantly faster speed and further improve the performance of InternImage across different model scales, resulting in a new state-of-the-art. + +Table 5. Object detection and instance segmentation performance on COCO val2017. $\mathrm{AP^b}$ and $\mathrm{AP^m}$ denotes box AP and mask AP, respectively. "MS" means multi-scale training. FlashInternImage w/ DCNv4 models converge faster, clearly outperform other baselines with $1\times$ training schedule, and still maintain a leading position when training $3\times$ longer while being significantly faster than InternImage. + +
Modelcrop size#paramFPSmIoU (SS)mIoU (MS)
Swin-T512260M107 / 16844.545.8
ConvNeXt-T512260M120 / 18446.046.7
InternImage-T512259M100 / 13947.948.1
FlashInternImage-T512259M119 / 20649.350.3
Swin-S512281M89 / 14247.649.5
ConvNeXt-S512282M107 / 16448.749.6
InternImage-S512280M89 / 12350.150.9
FlashInternImage-S512280M107 / 18250.651.6
Swin-B5122121M77 / 12648.149.7
ConvNeXt-B5122122M95 / 14749.149.9
InternImage-B5122128M77 / 10450.851.3
FlashInternImage-B5122128M94 / 15752.052.6
Swin-L6402234M59 / 9952.153.5
ConvNeXt-L6402235M73 / 11753.253.7
InternImage-L6402256M56 / 7853.954.1
ConvNeXt-XL6402391M53 / 7553.654.0
InternImage-XL6402368M47 / 6755.055.3
FlashInternImage-L6402256M71 / 12255.656.0
+ +Table 6. Semantic segmentation performance on the ADE20K validation set. All models are trained with UperNet. "SS" and "MS" denote single-scale and multi-scale testing, respectively. FPS is reported with single-scale testing. FlashInternImage w/ DCNv4 achieves superior performance with competitive speed. + +
ModelNDSmAPFPS†FPS
InternImage-B62.054.08.02.7
InternImage-XL63.455.64.02.0
FlashInternImage-S61.755.516.84.1
FlashInternImage-B63.157.412.13.8
+ +Table 7. 3D detection performance of BEVFormer v2 on nuScenes test set. We report backbone FPS (denoted with $\dagger$ ), overall FPS results with underoptimized head implementation are also added for reference. With on-par NDS and higher mAP results, FlashInternImage can be $50 - 90\%$ faster than InternImage baselines or $200\% - 300\%$ when only considering the backbone. + +3D Detection: We further test DCNv4 on the camera-based 3D object detection task in the autonomous driving scenario, We train BEVFormer v2 [43], a state-of-the-art multi-camera 3D object detector, with FlashInternImage-Small and Base backbone models on the nuScenes dataset for 24 epochs. We report results on the nuScenes test set in Tab. 7 with FPS for each model. We note that the header parts, such as the BEV encoder and object decoder in BEV-Former v2, are underoptimized and take more than $50\%$ of the running time (and even more with a faster backbone); thus, we also report the FPS for the backbone for a clearer illustration. Our results show that when only considering the backbone, FlashInternImage can be twice or even three times faster than the InternImage backbone with an on-par performance, greatly increasing the model efficiency. + +# 4.4. DCNv4 as a Universal Operator + +Drop-in replacement in other vision backbones : Weverify whether DCNv4 can still work well in architectures de- + +
ModelIN-1K AccThroughput
ConvNeXt-B83.81241 / 1888
ConvNeXt-B + DCNv484.01495 / 2513 (+20% / +33%)
ViT-B81.81683 / 2781†
ViT-B + DCNv481.92092 / 3261 (+24% / +17%)
+ +Table 8. DCNv4 in other architecture. We show supervised learning results on ImageNet-1K and throughput. DCNv4 achieves higher throughput with comparable accuracy. $\dagger$ denotes testing with the advanced FlashAttention-2 implementation. + +
Model#paramFIDFPS
U-Net860M2.944.82
U-Net + DCNv4566M2.444.92
+ +signed with other operators, such as ConvNeXt and ViT. To achieve that, we replace the attention module and depthwise convolution layer with DCNv4 in ViT and ConvNeXt and perform supervised learning on ImageNet-1K without changing any other architecture and hyperparameters, similar to FlashInternImage and InternImage. The results are shown in Tab. 8. We can see that on these architectures, which are carefully tuned for the specific operators, our DCNv4 can perform equally well. Thanks to the fast speed of DCNv4, the new model can even achieve better throughput, showcasing the superior performance of DCNv4. + +Drop-in replacement in diffusion model: DCN has been recognized to be an effective operator for perception tasks. As generative models become a fundamental tool for AI-generated content (AIGC), we are also curious if DCNv4 can work well on generation tasks with diffusion-based generative models. Specifically, we choose the U-Net [30] used in Stable Diffusion [29] as our baselines and replace the attention module and regular $3 \times 3$ convolution in U-Net. We use U-ViT's codebase, follow its training schedule, and train a latent diffusion model based on the image latent extracted from an image autoencoder provided by Stable Diffusion. We show the results in Tab. 9. We can see that DCNv4 also works well in generative modeling, achieving better results in terms of FID/Throughput with fewer parameters compared to regular convolution in U-Net. Notice that the architecture/hyperparameters may not be optimal for DCNv4, and it is possible that re-designing the models or searching for new hyperparameters for DCNv4 will give better results. + +# 4.5. Ablation Study + +We conduct ablation studies in our optimization choice described in Sec. 3.2. The results are shown in Tab. 10. The time in the table is obtained with $56 \times 56 \times 128$ input with + +Table 9. Class conditional generation on ImageNet 256x256. Latent diffusion models are trained from scratch with U-Net. We replace the convolution layer in the models with DCNv4. DCNv4 can achieve better FID results without any hyperparameter tuning. + +
Implementation variantModuleKernel
Original DCNv33.281.45
- micro design2.121.45
- redundant memory access2.201.53
- redundant computation2.181.51
- redundant memory instr.1.280.606
- half-precision format0.8730.404
+ +Table 10. Ablation studies of DCN's runtime (ms). We show how to achieve DCNv4 (gray row) from the original DCNv3 implementation and how different design choices affect the speed. + +batch size 64 and 4 groups (32 channels per group). We first remove the softmax operation and improve the micro design, which means we merge the two linear layers into one and remove costly layer norm and GELU activation in offset/aggregation weight computing, simplifying the overall modules and increasing the speed. We then start modifying the kernel implementation. First, we change the parallel execution pattern and let each thread process 8 channels instead of 1 channel, thus, unnecessary memory access on loading sampling offset and aggregation weight values from the GPU memory can be saved. As we expected, solely applying this change will not increase the speed as the degree of parallelism decreases, and each thread's workload increases 8 times now. The latency is increased instead. Eliminating redundant computation by reusing the bilinear interpolation coefficient (4th row) can save some time but is insignificant. Removing the redundant memory instruction via vectorized load/store can greatly reduce the workload of each thread and largely accelerate the GPU kernel speed (5th row). Using a half-precision datatype, which halves the number of bytes the kernel needs to read/write, further increases the data throughput, as shown in the 6th row. In the end, we reach the final DCNv4 design, which is three times more efficient than the original implementation. + +# 5. Conclusion + +We present Deformable Convolution v4 (DCNv4), an efficient dynamic and sparse operator. By rethinking the dynamic property in deformable convolution and streamlining memory access, DCNv4 emerges as a much faster and more effective operator than its predecessor DCNv3. DCNv4-equipped FlashInternImage backbone not only enhances speed but also improves performance across various vision tasks. We further show DCNv4's versatility and effectiveness as a universal operator by integrating it into state-of-the-art architecture like ConvNeXt and ViT with improved throughput and accuracy; and it also works well in latent diffusion model, showing its potential to enhance generative models. + +# Acknowledgement + +The work is supported by the National Key R&D Program of China (NO. 2022ZD0161300), by the National Natural Science Foundation of China (62376134). + +# References + +[1] Saehyun Ahn, Jung-Woo Chang, and Suk-Ju Kang. An efficient accelerator design methodology for deformable convolutional networks. In 2020 IEEE International Conference on Image Processing (ICIP), pages 3075-3079. IEEE, 2020. 2 +[2] Zhaowei Cai and Nuno Vasconcelos. Cascade r-cnn: Delving into high quality object detection. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6154-6162, 2018. 7 +[3] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In European conference on computer vision, pages 213-229. Springer, 2020. 2 +[4] Bowen Cheng, Ishan Misra, Alexander G Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1290–1299, 2022. 2 +[5] Sharan Chetlur, Cliff Woolley, Philippe Vandermersch, Jonathan Cohen, John Tran, Bryan Catanzaro, and Evan Shelhamer. codnn: Efficient primitives for deep learning. arXiv preprint arXiv:1410.0759, 2014. 5 +[6] François Chollet. Xception: Deep learning with depthwise separable convolutions. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1251-1258, 2017. 2, 3, 5 +[7] Jifeng Dai, Haozhi Qi, Yuwen Xiong, Yi Li, Guodong Zhang, Han Hu, and Yichen Wei. Deformable convolutional networks. In Proceedings of the IEEE international conference on computer vision, pages 764-773, 2017. 2 +[8] Tri Dao. Flashattention-2: Faster attention with better parallelism and work partitioning. arXiv preprint arXiv:2307.08691, 2023.5 +[9] Tri Dao, Dan Fu, Stefano Ermon, Atri Rudra, and Christopher Ré. Flashattention: Fast and memory-efficient exact attention with io-awareness. Advances in Neural Information Processing Systems, 35:16344-16359, 2022. 2, 3 +[10] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248–255. IEEE, 2009. 2 +[11] Xiaohan Ding, Xiangyu Zhang, Jungong Han, and Guiguang Ding. Scaling up your kernels to 31x31: Revisiting large kernel design in cnns. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11963-11975, 2022. 2 +[12] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations, 2020. 1, 2, 3 +[13] Qi Han, Zejia Fan, Qi Dai, Lei Sun, Ming-Ming Cheng, Jiaying Liu, and Jingdong Wang. On the connection between + +local attention and dynamic depth-wise convolution. arXiv preprint arXiv:2106.04263, 2021. 2 +[14] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 2 +[15] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In Proceedings of the IEEE international conference on computer vision, 2017. 1, 5, 7 +[16] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. ImageNet classification with deep convolutional neural networks. Advances in neural information processing systems, 2012. 2 +[17] Yann LeCun, Léon Bottou, Yoshua Bengio, and Patrick Haffner. Gradient-based learning applied to document recognition. Proceedings of the IEEE, 86(11):2278-2324, 1998. 2 +[18] Youngwan Lee, Joong-won Hwang, Sangrok Lee, Yuseok Bae, and Jongyoul Park. An energy andgpu-computation efficient backbone network for real-time object detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops, pages 0–0, 2019. 2 +[19] Feng Li, Hao Zhang, Huaizhe Xu, Shilong Liu, Lei Zhang, Lionel M Ni, and Heung-Yeung Shum. Mask dino: Towards a unified transformer-based framework for object detection and segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3041-3050, 2023. 2 +[20] Yanghao Li, Hanzi Mao, Ross Girshick, and Kaiming He. Exploring plain vision transformer backbones for object detection. In European Conference on Computer Vision, pages 280-296. Springer, 2022. 5 +[21] Zhiqi Li, Wenhai Wang, Hongyang Li, Enze Xie, Chonghao Sima, Tong Lu, Yu Qiao, and Jifeng Dai. Bevformer: Learning bird's-eye-view representation from multi-camera images via spatiotemporal transformers. In European conference on computer vision, pages 1-18. Springer, 2022. 2 +[22] Zhiqi Li, Wenhai Wang, Enze Xie, Zhiding Yu, Anima Anandkumar, Jose M Alvarez, Ping Luo, and Tong Lu. Panoptic segformer: Delving deeper into panoptic segmentation with transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1280-1289, 2022. 2 +[23] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 7 +[24] Yingfei Liu, Tiancai Wang, Xiangyu Zhang, and Jian Sun. Petr: Position embedding transformation for multi-view 3d object detection. In European Conference on Computer Vision, pages 531-548. Springer, 2022. 2 +[25] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10012-10022, 2021. 1, 2, 3, 5, 6 + +[26] Zhuang Liu, Hanzi Mao, Chao-Yuan Wu, Christoph Feichtenhofer, Trevor Darrell, and Saining Xie. A convnet for the 2020s. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11976-11986, 2022. 1, 2, 6 +[27] Ningning Ma, Xiangyu Zhang, Hai-Tao Zheng, and Jian Sun. Shufflenet v2: Practical guidelines for efficient cnn architecture design. In ECCV, 2018. 2, 3, 4 +[28] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. Advances in neural information processing systems, 32, 2019. 5 +[29] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 1, 2, 8 +[30] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical Image Computing and Computer-Assisted Intervention-MICCAI 2015: 18th International Conference, Munich, Germany, October 5-9, 2015, Proceedings, Part III 18, pages 234-241. Springer, 2015. 2, 8 +[31] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in Neural Information Processing Systems, 35:36479-36494, 2022. 1 +[32] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556, 2014. 2 +[33] Robin Strudel, Ricardo Garcia, Ivan Laptev, and Cordelia Schmid. Segmenter: Transformer for semantic segmentation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 7262-7272, 2021. 2 +[34] Hugues Thomas, Charles R Qi, Jean-Emmanuel Deschaud, Beatz Marcotegui, François Goulette, and Leonidas J Guibas. Kpconv: Flexible and deformable convolution for point clouds. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6411–6420, 2019. 2 +[35] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 2, 5 +[36] Ashish Vaswani, Prajit Ramachandran, Aravind Srinivas, Niki Parmar, Blake Hechtman, and Jonathon Shlens. Scaling local self-attention for parameter efficient visual backbones. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12894–12904, 2021. 2 +[37] Shenlong Wang, Simon Suo, Wei-Chiu Ma, Andrei Pokrovsky, and Raquel Urtasun. Deep parametric continuous convolutional neural networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2589-2597, 2018. 2 + +[38] Wenhai Wang, Jifeng Dai, Zhe Chen, Zhenhang Huang, Zhiqi Li, Xizhou Zhu, Xiaowei Hu, Tong Lu, Lewei Lu, Hongsheng Li, et al. Internimage: Exploring large-scale vision foundation models with deformable convolutions. In CVPR, 2023. 1, 2, 6 +[39] Ross Wightman. Pytorch image models. https://github.com/rwrightman/pytorch-image-models, 2019.6 +[40] Felix Wu, Angela Fan, Alexei Baevski, Yann N Dauphin, and Michael Auli. Pay less attention with lightweight and dynamic convolutions. arXiv preprint arXiv:1901.10430, 2019. 2 +[41] Tete Xiao, Yingcheng Liu, Bolei Zhou, Yuning Jiang, and Jian Sun. Unified perceptual parsing for scene understanding. In Proceedings of the European conference on computer vision (ECCV), pages 418-434, 2018. 1, 7 +[42] Yuwen Xiong, Mengye Ren, Renjie Liao, Kelvin Wong, and Raquel Urtasun. Deformable filter convolution for point cloud reasoning. arXiv preprint arXiv:1907.13079, 2019. 2 +[43] Chenyu Yang, Yuntao Chen, Hao Tian, Chenxin Tao, Xizhou Zhu, Zhaoxiang Zhang, Gao Huang, Hongyang Li, Yu Qiao, Lewei Lu, et al. Bevformer v2: Adapting modern image backbones to bird's-eye-view recognition via perspective supervision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17830-17839, 2023. 2, 7 +[44] Xiaohua Zhai, Alexander Kolesnikov, Neil Houlsby, and Lucas Beyer. Scaling vision transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12104-12113, 2022. 1 +[45] Hao Zhang, Feng Li, Shilong Liu, Lei Zhang, Hang Su, Jun Zhu, Lionel M Ni, and Heung-Yeung Shum. Dino: Detr with improved denoising anchor boxes for end-to-end object detection. arXiv preprint arXiv:2203.03605, 2022. 2 +[46] Bolei Zhou, Hang Zhao, Xavier Puig, Sanja Fidler, Adela Barriuso, and Antonio Torralba. Scene parsing through ade20k dataset. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 633-641, 2017. 7 +[47] Xizhou Zhu, Han Hu, Stephen Lin, and Jifeng Dai. Deformable convnets v2: More deformable, better results. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 2 +[48] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable detr: Deformable transformers for end-to-end object detection. In International Conference on Learning Representations, 2020. 2, 5 \ No newline at end of file diff --git a/2024/Efficient Deformable ConvNets_ Rethinking Dynamic and Sparse Operator for Vision Applications/images.zip b/2024/Efficient Deformable ConvNets_ Rethinking Dynamic and Sparse Operator for Vision Applications/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..b410ce9ac0f23bd9d521790a36f4631872162c95 --- /dev/null +++ b/2024/Efficient Deformable ConvNets_ Rethinking Dynamic and Sparse Operator for Vision Applications/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:cdaa6da517fbbb324ddb432de6577933f4283777f8bd951bda01310565160a53 +size 616200 diff --git a/2024/Efficient Deformable ConvNets_ Rethinking Dynamic and Sparse Operator for Vision Applications/layout.json b/2024/Efficient Deformable ConvNets_ Rethinking Dynamic and Sparse Operator for Vision Applications/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..7a782c54b9883ad9b703c4a64f7480c58079c8e1 --- /dev/null +++ b/2024/Efficient Deformable ConvNets_ Rethinking Dynamic and Sparse Operator for Vision Applications/layout.json @@ -0,0 +1,8346 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 96, + 102, + 500, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 96, + 102, + 500, + 140 + ], + "spans": [ + { + "bbox": [ + 96, + 102, + 500, + 140 + ], + "type": "text", + "content": "Efficient Deformable ConvNets: Rethinking Dynamic and Sparse Operator for Vision Applications" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 137, + 159, + 456, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 159, + 456, + 175 + ], + "spans": [ + { + "bbox": [ + 137, + 159, + 456, + 175 + ], + "type": "text", + "content": "Yuwen Xiong" + }, + { + "bbox": [ + 137, + 159, + 456, + 175 + ], + "type": "inline_equation", + "content": "^{*,1,2}" + }, + { + "bbox": [ + 137, + 159, + 456, + 175 + ], + "type": "text", + "content": " Zhiqi Li" + }, + { + "bbox": [ + 137, + 159, + 456, + 175 + ], + "type": "inline_equation", + "content": "^{*,3,2}" + }, + { + "bbox": [ + 137, + 159, + 456, + 175 + ], + "type": "text", + "content": " Yuntao Chen" + }, + { + "bbox": [ + 137, + 159, + 456, + 175 + ], + "type": "inline_equation", + "content": "^{*4}" + }, + { + "bbox": [ + 137, + 159, + 456, + 175 + ], + "type": "text", + "content": " Feng Wang" + }, + { + "bbox": [ + 137, + 159, + 456, + 175 + ], + "type": "inline_equation", + "content": "^{*5}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 100, + 175, + 492, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 100, + 175, + 492, + 190 + ], + "spans": [ + { + "bbox": [ + 100, + 175, + 492, + 190 + ], + "type": "text", + "content": "Xizhou Zhu" + }, + { + "bbox": [ + 100, + 175, + 492, + 190 + ], + "type": "inline_equation", + "content": "^{5,6}" + }, + { + "bbox": [ + 100, + 175, + 492, + 190 + ], + "type": "text", + "content": " Jiapeng Luo" + }, + { + "bbox": [ + 100, + 175, + 492, + 190 + ], + "type": "inline_equation", + "content": "^{6}" + }, + { + "bbox": [ + 100, + 175, + 492, + 190 + ], + "type": "text", + "content": " Wenhai Wang" + }, + { + "bbox": [ + 100, + 175, + 492, + 190 + ], + "type": "inline_equation", + "content": "^{7,2}" + }, + { + "bbox": [ + 100, + 175, + 492, + 190 + ], + "type": "text", + "content": " Tong Lu" + }, + { + "bbox": [ + 100, + 175, + 492, + 190 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 100, + 175, + 492, + 190 + ], + "type": "text", + "content": " Hongsheng Li" + }, + { + "bbox": [ + 100, + 175, + 492, + 190 + ], + "type": "inline_equation", + "content": "^{7}" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 169, + 189, + 425, + 202 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 169, + 189, + 425, + 202 + ], + "spans": [ + { + "bbox": [ + 169, + 189, + 425, + 202 + ], + "type": "text", + "content": "Yu Qiao² Lewei Lu⁶ Jie Zhou⁵ Jifeng Dai⁵,²" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 127, + 216, + 466, + 230 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 127, + 216, + 466, + 230 + ], + "spans": [ + { + "bbox": [ + 127, + 216, + 466, + 230 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 127, + 216, + 466, + 230 + ], + "type": "text", + "content": "University of Toronto " + }, + { + "bbox": [ + 127, + 216, + 466, + 230 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 127, + 216, + 466, + 230 + ], + "type": "text", + "content": "OpenGVLab, Shanghai AI Laboratory" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 107, + 230, + 485, + 243 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 107, + 230, + 485, + 243 + ], + "spans": [ + { + "bbox": [ + 107, + 230, + 485, + 243 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 107, + 230, + 485, + 243 + ], + "type": "text", + "content": "Nanjing University " + }, + { + "bbox": [ + 107, + 230, + 485, + 243 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 107, + 230, + 485, + 243 + ], + "type": "text", + "content": "CAIR, HKISI, CAS " + }, + { + "bbox": [ + 107, + 230, + 485, + 243 + ], + "type": "inline_equation", + "content": "^{5}" + }, + { + "bbox": [ + 107, + 230, + 485, + 243 + ], + "type": "text", + "content": "Tsinghua University" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 243, + 462, + 258 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 243, + 462, + 258 + ], + "spans": [ + { + "bbox": [ + 130, + 243, + 462, + 258 + ], + "type": "inline_equation", + "content": "^{6}" + }, + { + "bbox": [ + 130, + 243, + 462, + 258 + ], + "type": "text", + "content": "SenseTime Research " + }, + { + "bbox": [ + 130, + 243, + 462, + 258 + ], + "type": "inline_equation", + "content": "^{7}" + }, + { + "bbox": [ + 130, + 243, + 462, + 258 + ], + "type": "text", + "content": "The Chinese University of Hong Kong" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 171, + 258, + 417, + 271 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 171, + 258, + 417, + 271 + ], + "spans": [ + { + "bbox": [ + 171, + 258, + 417, + 271 + ], + "type": "text", + "content": "https://github.com/OpenGVLab/DCNv4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 162, + 299, + 209, + 312 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 162, + 299, + 209, + 312 + ], + "spans": [ + { + "bbox": [ + 162, + 299, + 209, + 312 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 45, + 324, + 290, + 610 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 324, + 290, + 610 + ], + "spans": [ + { + "bbox": [ + 45, + 324, + 290, + 610 + ], + "type": "text", + "content": "We introduce Deformable Convolution v4 (DCNv4), a highly efficient and effective operator designed for a broad spectrum of vision applications. DCNv4 addresses the limitations of its predecessor, DCNv3, with two key enhancements: 1. removing softmax normalization in spatial aggregation to enhance its dynamic property and expressive power and 2. optimizing memory access to minimize redundant operations for speedup. These improvements result in a significantly faster convergence compared to DCNv3 and a substantial increase in processing speed, with DCNv4 achieving more than three times the forward speed. DCNv4 demonstrates exceptional performance across various tasks, including image classification, instance and semantic segmentation, and notably, image generation. When integrated into generative models like U-Net in the latent diffusion model, DCNv4 outperforms its baseline, underscoring its possibility to enhance generative models. In practical applications, replacing DCNv3 with DCNv4 in the InternImage model to create FlashInternImage results in up to " + }, + { + "bbox": [ + 45, + 324, + 290, + 610 + ], + "type": "inline_equation", + "content": "80\\%" + }, + { + "bbox": [ + 45, + 324, + 290, + 610 + ], + "type": "text", + "content": " speed increase and further performance improvement without further modifications. The advancements in speed and efficiency of DCNv4, combined with its robust performance across diverse vision tasks, show its potential as a foundational building block for future vision models." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 47, + 626, + 127, + 639 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 626, + 127, + 639 + ], + "spans": [ + { + "bbox": [ + 47, + 626, + 127, + 639 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 46, + 647, + 288, + 682 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 647, + 288, + 682 + ], + "spans": [ + { + "bbox": [ + 46, + 647, + 288, + 682 + ], + "type": "text", + "content": "In the field of computer vision, there is an ongoing debate about whether convolutional networks (ConvNets) or Transformers offer superior performance. In recent years, Trans" + } + ] + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 312, + 299, + 451, + 417 + ], + "blocks": [ + { + "bbox": [ + 312, + 299, + 451, + 417 + ], + "lines": [ + { + "bbox": [ + 312, + 299, + 451, + 417 + ], + "spans": [ + { + "bbox": [ + 312, + 299, + 451, + 417 + ], + "type": "image", + "image_path": "0a203e8027418311ed674c2a129d80b85a2d8873b8ba87e04c7d45a298d9ee45.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 375, + 421, + 386, + 429 + ], + "lines": [ + { + "bbox": [ + 375, + 421, + 386, + 429 + ], + "spans": [ + { + "bbox": [ + 375, + 421, + 386, + 429 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 455, + 300, + 547, + 416 + ], + "blocks": [ + { + "bbox": [ + 455, + 300, + 547, + 416 + ], + "lines": [ + { + "bbox": [ + 455, + 300, + 547, + 416 + ], + "spans": [ + { + "bbox": [ + 455, + 300, + 547, + 416 + ], + "type": "image", + "image_path": "05016b33a235bfddc93728e971ce3b133882971c9880b5eec24bc9a4bd5b8091.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 495, + 420, + 507, + 429 + ], + "lines": [ + { + "bbox": [ + 495, + 420, + 507, + 429 + ], + "spans": [ + { + "bbox": [ + 495, + 420, + 507, + 429 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 304, + 431, + 547, + 475 + ], + "lines": [ + { + "bbox": [ + 304, + 431, + 547, + 475 + ], + "spans": [ + { + "bbox": [ + 304, + 431, + 547, + 475 + ], + "type": "text", + "content": "Figure 1. (a) We show relative runtime with DCNv3 as the baseline. DCNv4 shows significant speedup over DCNv3, and surpasses other common vision operators. (b) With the same network architecture, DCNv4 converges faster than other operators." + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 476, + 547, + 606 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 476, + 547, + 606 + ], + "spans": [ + { + "bbox": [ + 304, + 476, + 547, + 606 + ], + "type": "text", + "content": "former models [12, 25, 44] have achieved remarkable results in large vision models with the attention mechanism, showing the potential to overtake ConvNets. However, recent works such as InternImage [38] and ConvNeXt [26] demonstrate that ConvNet-based vision models retain robust performance, efficiency, simplicity, and suitable inductive bias for various downstream tasks [15, 41]. Notably, in domains like image generation [29, 31], models containing both convolutions and transformers remain the preferred approach. This situation brings to light the enduring value of convolution-based approaches." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 607, + 548, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 607, + 548, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 607, + 548, + 713 + ], + "type": "text", + "content": "Building on convolution's strengths, Deformable Convolution v3 (DCNv3) – the core operator of the advanced ConvNet model InternImage – innovatively combines a sparse attention mechanism with convolution: it processes each output location in a sliding window manner with a small window size (e.g. " + }, + { + "bbox": [ + 304, + 607, + 548, + 713 + ], + "type": "inline_equation", + "content": "3 \\times 3 = 9" + }, + { + "bbox": [ + 304, + 607, + 548, + 713 + ], + "type": "text", + "content": " points) which acts as a local, sparse operator like convolution, while dynamically samples point with an adaptive range and aggregates the spatial features with input-dependent attention weights. With its" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 693, + 129, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 693, + 129, + 703 + ], + "spans": [ + { + "bbox": [ + 61, + 693, + 129, + 703 + ], + "type": "text", + "content": "* Equal contribution" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 64, + 703, + 237, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 703, + 237, + 712 + ], + "spans": [ + { + "bbox": [ + 64, + 703, + 237, + 712 + ], + "type": "text", + "content": "Corresponding author (daijifeng@tsinghua.edu.cn)" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5652" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 288, + 119 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 288, + 119 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 288, + 119 + ], + "type": "text", + "content": "small window size and ConvNet inductive bias, DCNv3 is expected to achieve a faster convergence rate and lower inference latency, especially when compared to dense global [12] or local window-based [25] attention methods." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 120, + 288, + 285 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 120, + 288, + 285 + ], + "spans": [ + { + "bbox": [ + 46, + 120, + 288, + 285 + ], + "type": "text", + "content": "Despite these advantages, DCN has not become the go-to solution for vision backbone models. This observation led us to investigate the lingering limitations of the DCN operator. The first thing we notice is the running speed. The slow speed of DCN is known to be a long-standing problem [1], as it introduces extra overhead on sampling non-nearby locations, making it not fit modern convolution algorithms. Our comparative analysis, illustrated in Fig. 1a, reveals that DCNv3 can be slower than a properly optimized dense global attention [9], highlighting the need for further optimization. Moreover, we find DCNv3 even converges slower than global attention at the initial backbone training phase, as shown in Fig. 1b, which is counter-intuitive as DCNv3 is equipped with ConvNet inductive bias." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 286, + 288, + 356 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 286, + 288, + 356 + ], + "spans": [ + { + "bbox": [ + 46, + 286, + 288, + 356 + ], + "type": "text", + "content": "To overcome these challenges, we propose Deformable Convolution v4 (DCNv4), an innovative advancement to optimize the sparse DCN operator for practical efficiency. DCNv4 comes with a much faster implementation and an improved operator design to enhance its performance, which we will elaborate on as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 357, + 288, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 357, + 288, + 441 + ], + "spans": [ + { + "bbox": [ + 46, + 357, + 288, + 441 + ], + "type": "text", + "content": "First, we conduct instruction-level kernel profiling for existing implementation and find that DCNv3 is already lightweight. The compute cost is less than " + }, + { + "bbox": [ + 46, + 357, + 288, + 441 + ], + "type": "inline_equation", + "content": "1\\%" + }, + { + "bbox": [ + 46, + 357, + 288, + 441 + ], + "type": "text", + "content": ", while memory access costs " + }, + { + "bbox": [ + 46, + 357, + 288, + 441 + ], + "type": "inline_equation", + "content": "99\\%" + }, + { + "bbox": [ + 46, + 357, + 288, + 441 + ], + "type": "text", + "content": ". This motivates us to revisit the operator implementation and find that many memory accesses in the DCN forward process are redundant and thus can be optimized, leading to a much faster DCNv4 implementation." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 441, + 288, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 441, + 288, + 559 + ], + "spans": [ + { + "bbox": [ + 46, + 441, + 288, + 559 + ], + "type": "text", + "content": "Second, drawing inspiration from convolution's unbounded weight range, we find that softmax in spatial aggregation, a standard operation in dense attention, is unnecessary in DCNv3, as it is not a requirement for operators with a dedicated aggregation window for each location. Intuitively, softmax puts a bounded " + }, + { + "bbox": [ + 46, + 441, + 288, + 559 + ], + "type": "inline_equation", + "content": "0 \\sim 1" + }, + { + "bbox": [ + 46, + 441, + 288, + 559 + ], + "type": "text", + "content": " value range to the weight and will limit the expressive power of the aggregation weight. This insight led us to remove the softmax in DCNv4, enhancing its dynamic property and improving its performance." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 559, + 288, + 606 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 559, + 288, + 606 + ], + "spans": [ + { + "bbox": [ + 46, + 559, + 288, + 606 + ], + "type": "text", + "content": "As a result, DCNv4 converges significantly faster than DCNv3 and accelerates forward speed by more than " + }, + { + "bbox": [ + 46, + 559, + 288, + 606 + ], + "type": "inline_equation", + "content": "3 \\times" + }, + { + "bbox": [ + 46, + 559, + 288, + 606 + ], + "type": "text", + "content": ". This improvement allows DCNv4 to fully leverage its sparse property and become one of the fastest core vision operators." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 607, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 607, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 607, + 288, + 714 + ], + "type": "text", + "content": "We further replace DCNv3 in InternImage with DCNv4, creating FlashInternImage. Remarkably, FlashInternImage achieves a " + }, + { + "bbox": [ + 46, + 607, + 288, + 714 + ], + "type": "inline_equation", + "content": "50 \\sim 80\\%" + }, + { + "bbox": [ + 46, + 607, + 288, + 714 + ], + "type": "text", + "content": " speed increase compared to InternImage without any additional modifications. This enhancement positions FlashInternImage as one of the fastest modern vision backbone networks while maintaining superior performance. With the help of DCNv4, FlashInternImage significantly improves the convergence speed in ImageNet classification [10] and transfer learning settings and further" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 72, + 541, + 84 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 541, + 84 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 541, + 84 + ], + "type": "text", + "content": "demonstrates improved performance in downstream tasks." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 85, + 547, + 262 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 85, + 547, + 262 + ], + "spans": [ + { + "bbox": [ + 304, + 85, + 547, + 262 + ], + "type": "text", + "content": "Furthermore, DCNv4 shows potential as a universal vision operator in various architectures and tasks. We integrate DCNv4 into other modern backbone architectures, including ConvNeXt [26] and ViT [12], replacing depthwise convolution [6] and dense self-attention layers [35]. Surprisingly, without any hyperparameter adjustments, these meticulously designed networks with DCNv4 perform on par while being much faster, showing the efficacy and efficiency of the dynamic, sparse DCNv4. Moreover, we explore the potential of DCNv4 in generative models as a new application domain. Specifically, we apply it in the U-Net [30] architecture used in diffusion models [29], replacing regular convolution with DCNv4. Our experimental results show that DCNv4 can work better than the baselines in image generation, showing great potential for using DCN to improve generative models." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 262, + 547, + 286 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 262, + 547, + 286 + ], + "spans": [ + { + "bbox": [ + 305, + 262, + 547, + 286 + ], + "type": "text", + "content": "We have released the DCNv4 implementation to facilitate future research in the vision community." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 292, + 392, + 304 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 292, + 392, + 304 + ], + "spans": [ + { + "bbox": [ + 306, + 292, + 392, + 304 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 306, + 548, + 686 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 306, + 548, + 686 + ], + "spans": [ + { + "bbox": [ + 304, + 306, + 548, + 686 + ], + "type": "text", + "content": "Core operators in vision models: The standard convolution [17] stands as the most prevalent and impactful operator, forming the backbone of the majority of computer vision architectures [14, 16, 32]. Nevertheless, a myriad of operators, each with unique characteristics, collectively play a crucial role in the development of computer vision. Depthwise separable convolution (DWConv) [6] separates the spatial and channel operations, and has been pivotal in developing lightweight and efficient models [26, 27]. RepLKNet [11] illustrates that a purely convolutional network, leveraging large-kernel depth-wise convolutions, can attain competitive performance in both efficiency and effectiveness. Deformable Convolution (DCN) series [7, 38, 47] significantly leaps the adaptability of convolution by adding learnable offsets to the convolutions kernels. Contrary to convolutions, attention mechanisms [35] possess the capacity to model long-range dependencies and have been successfully adopted in various computer vision tasks [3, 12, 24, 33]. Window attention [25, 36] reduces the computational complexity inherent in vanilla attention by restricting the attention operation to a fixed-size window. To mitigate the high computational complexity associated with vanilla attention, deformable attention [48] enables each query to concentrate on a select number of key sampling points, with dynamically determined locations and weights. This efficient method is widely used in the following arts perception methods [4, 19, 21, 22, 43, 45]. DynamicConv [40] and dynamic-DWNet [13] augment DWConv by incorporating dynamic weights, thereby enabling the use of instance-specific weights that adapt dynamically. For non-grid structured data, sparse operators [34, 37, 42] utilize dynamic weights obtained via bilinear interpolation or in a parametric way." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 689, + 548, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 689, + 548, + 714 + ], + "spans": [ + { + "bbox": [ + 305, + 689, + 548, + 714 + ], + "type": "text", + "content": "Memory access cost in vision backbones: As underscored in previous studies [18, 27], FLOPs, although a fre" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5653" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 50, + 70, + 286, + 114 + ], + "blocks": [ + { + "bbox": [ + 50, + 70, + 286, + 114 + ], + "lines": [ + { + "bbox": [ + 50, + 70, + 286, + 114 + ], + "spans": [ + { + "bbox": [ + 50, + 70, + 286, + 114 + ], + "type": "table", + "html": "
Model5th EP10th Ep20th Ep50th Ep300th Ep
ConvNeXt29.953.566.174.883.8
ConvNeXt + softmax8.5 (-21.4)25.3 (-28.2)51.1 (-15.0)69.1 (-5.7)81.6 (-2.2)
", + "image_path": "edb85164031ac5b3d3aa413971ac9aaf509c9ce20fbd33644dea91dbe04f7f34.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 115, + 289, + 148 + ], + "lines": [ + { + "bbox": [ + 47, + 115, + 289, + 148 + ], + "spans": [ + { + "bbox": [ + 47, + 115, + 289, + 148 + ], + "type": "text", + "content": "Table 1. ImageNet-1K accuracy at different training epochs. Softmax on the convolution weights significantly affects the convergence speed and the final performance for the ConvNeXt model." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 152, + 290, + 331 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 152, + 290, + 331 + ], + "spans": [ + { + "bbox": [ + 46, + 152, + 290, + 331 + ], + "type": "text", + "content": "sequently used metric to measure model complexity, do not accurately represent the model's speed or latency. In practical scenarios, the running speed of a model is influenced by multiple factors, not just FLOPs. Memory access cost plays a significant role in this context [27]. Flash-Attention [9], by reducing the number of accesses to High Bandwidth Memory (HBM), achieves a significantly faster speed in practice despite having higher FLOPs compared to vanilla attention. Although DCN operators do not exhibit a disadvantage in terms of FLOPs, their latency is considerably longer compared to DW-Conv, under the same FLOPs budget, predominantly due to substantial memory access costs. In this work, we conduct a thorough analysis and optimization of the memory access costs associated with the DCN operators, significantly accelerating the DCN's running speed." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 338, + 103, + 350 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 338, + 103, + 350 + ], + "spans": [ + { + "bbox": [ + 47, + 338, + 103, + 350 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 357, + 287, + 381 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 357, + 287, + 381 + ], + "spans": [ + { + "bbox": [ + 47, + 357, + 287, + 381 + ], + "type": "text", + "content": "3.1. Rethinking the Dynamic Property in Deformable Convolution" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 387, + 288, + 425 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 387, + 288, + 425 + ], + "spans": [ + { + "bbox": [ + 47, + 387, + 288, + 425 + ], + "type": "text", + "content": "Revisiting DCNv3: Given an input " + }, + { + "bbox": [ + 47, + 387, + 288, + 425 + ], + "type": "inline_equation", + "content": "\\mathbf{x} \\in \\mathbb{R}^{H \\times W \\times C}" + }, + { + "bbox": [ + 47, + 387, + 288, + 425 + ], + "type": "text", + "content": " with height " + }, + { + "bbox": [ + 47, + 387, + 288, + 425 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 47, + 387, + 288, + 425 + ], + "type": "text", + "content": ", width " + }, + { + "bbox": [ + 47, + 387, + 288, + 425 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 47, + 387, + 288, + 425 + ], + "type": "text", + "content": " and channel " + }, + { + "bbox": [ + 47, + 387, + 288, + 425 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 47, + 387, + 288, + 425 + ], + "type": "text", + "content": ", the DCNv3 operation with " + }, + { + "bbox": [ + 47, + 387, + 288, + 425 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 47, + 387, + 288, + 425 + ], + "type": "text", + "content": " points is defined in Eq. (2) for each point " + }, + { + "bbox": [ + 47, + 387, + 288, + 425 + ], + "type": "inline_equation", + "content": "p_0" + }, + { + "bbox": [ + 47, + 387, + 288, + 425 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 87, + 430, + 287, + 463 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 87, + 430, + 287, + 463 + ], + "spans": [ + { + "bbox": [ + 87, + 430, + 287, + 463 + ], + "type": "interline_equation", + "content": "\\mathbf {y} _ {g} = \\sum_ {k = 1} ^ {K} \\mathbf {m} _ {g k} \\mathbf {x} _ {g} \\left(p _ {0} + p _ {k} + \\Delta p _ {g k}\\right), \\tag {1}", + "image_path": "af67301a6d512a5f0c190572a2f053303cda7b8209de4b4878a8bbe63952145c.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 93, + 465, + 287, + 480 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 93, + 465, + 287, + 480 + ], + "spans": [ + { + "bbox": [ + 93, + 465, + 287, + 480 + ], + "type": "interline_equation", + "content": "\\mathbf {y} = \\operatorname {c o n c a t} \\left(\\left[ \\mathbf {y} _ {1}, \\mathbf {y} _ {2}, \\dots , \\mathbf {y} _ {G} \\right], \\text {a x i s} = - 1\\right), \\tag {2}", + "image_path": "a8d8dcb06cdcf510683680c514433800d8f7d5cc53e97ba33e318f0244cf9289.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 488, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 488, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 488, + 289, + 715 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 488, + 289, + 715 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 46, + 488, + 289, + 715 + ], + "type": "text", + "content": " denotes the number of spatial aggregation groups. For the " + }, + { + "bbox": [ + 46, + 488, + 289, + 715 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 46, + 488, + 289, + 715 + ], + "type": "text", + "content": "-th group, " + }, + { + "bbox": [ + 46, + 488, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_g,\\mathbf{y}_g\\in \\mathbb{R}^{H\\times W\\times C'}" + }, + { + "bbox": [ + 46, + 488, + 289, + 715 + ], + "type": "text", + "content": " represents the sliced input/output feature map with " + }, + { + "bbox": [ + 46, + 488, + 289, + 715 + ], + "type": "inline_equation", + "content": "C' = C / G" + }, + { + "bbox": [ + 46, + 488, + 289, + 715 + ], + "type": "text", + "content": " represents the group dimension; " + }, + { + "bbox": [ + 46, + 488, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{m}_{gk}\\in \\mathbb{R}" + }, + { + "bbox": [ + 46, + 488, + 289, + 715 + ], + "type": "text", + "content": " denotes the spatial aggregation weights (also known as modulation scalar) of the " + }, + { + "bbox": [ + 46, + 488, + 289, + 715 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 488, + 289, + 715 + ], + "type": "text", + "content": "-th sampling point in the " + }, + { + "bbox": [ + 46, + 488, + 289, + 715 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 46, + 488, + 289, + 715 + ], + "type": "text", + "content": "-th group, conditioned on the input " + }, + { + "bbox": [ + 46, + 488, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 46, + 488, + 289, + 715 + ], + "type": "text", + "content": " and normalized by the softmax function along the dimension " + }, + { + "bbox": [ + 46, + 488, + 289, + 715 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 46, + 488, + 289, + 715 + ], + "type": "text", + "content": "; " + }, + { + "bbox": [ + 46, + 488, + 289, + 715 + ], + "type": "inline_equation", + "content": "p_k" + }, + { + "bbox": [ + 46, + 488, + 289, + 715 + ], + "type": "text", + "content": " denotes the " + }, + { + "bbox": [ + 46, + 488, + 289, + 715 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 488, + 289, + 715 + ], + "type": "text", + "content": "-th location of the pre-defined grid sampling " + }, + { + "bbox": [ + 46, + 488, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\{(-1, - 1),(-1,0),\\dots,(0, + 1),\\dots,(+1, + 1)\\}" + }, + { + "bbox": [ + 46, + 488, + 289, + 715 + ], + "type": "text", + "content": " as in regular convolutions and " + }, + { + "bbox": [ + 46, + 488, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\Delta p_{gk}" + }, + { + "bbox": [ + 46, + 488, + 289, + 715 + ], + "type": "text", + "content": " is the offset corresponding to the grid sampling location " + }, + { + "bbox": [ + 46, + 488, + 289, + 715 + ], + "type": "inline_equation", + "content": "p_k" + }, + { + "bbox": [ + 46, + 488, + 289, + 715 + ], + "type": "text", + "content": " in the " + }, + { + "bbox": [ + 46, + 488, + 289, + 715 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 46, + 488, + 289, + 715 + ], + "type": "text", + "content": "-th group. A " + }, + { + "bbox": [ + 46, + 488, + 289, + 715 + ], + "type": "inline_equation", + "content": "1\\times 1" + }, + { + "bbox": [ + 46, + 488, + 289, + 715 + ], + "type": "text", + "content": " point-wise convolution on " + }, + { + "bbox": [ + 46, + 488, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 46, + 488, + 289, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 488, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{y}" + }, + { + "bbox": [ + 46, + 488, + 289, + 715 + ], + "type": "text", + "content": " can be applied before and after the DCNv3 operator to enhance the model's expressive power, following separable convolution [6]. DCNv3 is a combination of convolution and attention: on the one hand, it processes the input data in a sliding window manner, which follows convolution and inherent its inductive bias; on the other hand, the sampling offset " + }, + { + "bbox": [ + 46, + 488, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\Delta p" + }, + { + "bbox": [ + 46, + 488, + 289, + 715 + ], + "type": "text", + "content": " and spatial aggregation weight " + }, + { + "bbox": [ + 46, + 488, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{m}" + }, + { + "bbox": [ + 46, + 488, + 289, + 715 + ], + "type": "text", + "content": " are dynamically predicted from the input feature," + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 309, + 72, + 410, + 131 + ], + "blocks": [ + { + "bbox": [ + 309, + 72, + 410, + 131 + ], + "lines": [ + { + "bbox": [ + 309, + 72, + 410, + 131 + ], + "spans": [ + { + "bbox": [ + 309, + 72, + 410, + 131 + ], + "type": "image", + "image_path": "0da0a2416ed9ee5e5ba643bd4e37433e46ab5f66b82645ab00246b074a464e2e.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 342, + 131, + 372, + 137 + ], + "lines": [ + { + "bbox": [ + 342, + 131, + 372, + 137 + ], + "spans": [ + { + "bbox": [ + 342, + 131, + 372, + 137 + ], + "type": "text", + "content": "(a) Attention" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 309, + 137, + 419, + 196 + ], + "blocks": [ + { + "bbox": [ + 309, + 137, + 419, + 196 + ], + "lines": [ + { + "bbox": [ + 309, + 137, + 419, + 196 + ], + "spans": [ + { + "bbox": [ + 309, + 137, + 419, + 196 + ], + "type": "image", + "image_path": "8a4cd00fb02996c23fda3a714e41a20b56e5df9ee73adb56f7d181101e90237f.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 198, + 547, + 319 + ], + "lines": [ + { + "bbox": [ + 304, + 198, + 547, + 319 + ], + "spans": [ + { + "bbox": [ + 304, + 198, + 547, + 319 + ], + "type": "text", + "content": "Figure 2. Comparisons of core operators in spatial aggregation for query pixels on different locations within the same channel. (a) Attention and (b) DCNv3 use bounded (range from " + }, + { + "bbox": [ + 304, + 198, + 547, + 319 + ], + "type": "inline_equation", + "content": "0 \\sim 1" + }, + { + "bbox": [ + 304, + 198, + 547, + 319 + ], + "type": "text", + "content": ") dynamic weights to aggregate spatial features, while the window (sampling point set) for attention is the same, and DCNv3 uses a dedicated window for each location. (c) Convolution has a more flexible unbounded value range for aggregation weights and uses a dedicated sliding window for each location, but the window shape and aggregation weights are input-independent. (d) DCNv4 combines their advantages, using an adaptive aggregation window and dynamic aggregation weights with an unbounded value range." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 425, + 72, + 533, + 132 + ], + "blocks": [ + { + "bbox": [ + 425, + 72, + 533, + 132 + ], + "lines": [ + { + "bbox": [ + 425, + 72, + 533, + 132 + ], + "spans": [ + { + "bbox": [ + 425, + 72, + 533, + 132 + ], + "type": "image", + "image_path": "33e00d201f52315ef0a993c478a5b8dacac3f7c24d1bcef7ad3f7498f1b2da04.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 460, + 132, + 487, + 138 + ], + "lines": [ + { + "bbox": [ + 460, + 132, + 487, + 138 + ], + "spans": [ + { + "bbox": [ + 460, + 132, + 487, + 138 + ], + "type": "text", + "content": "(b) DCNv3" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 425, + 137, + 544, + 196 + ], + "blocks": [ + { + "bbox": [ + 425, + 137, + 544, + 196 + ], + "lines": [ + { + "bbox": [ + 425, + 137, + 544, + 196 + ], + "spans": [ + { + "bbox": [ + 425, + 137, + 544, + 196 + ], + "type": "image", + "image_path": "a00f4168748400aa3f4bd5805a6d86f8fe61347d8863fa8041ffaf80c1424ece.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 326, + 546, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 326, + 546, + 361 + ], + "spans": [ + { + "bbox": [ + 304, + 326, + 546, + 361 + ], + "type": "text", + "content": "showing its dynamic property and making it more like an attention mechanism. We compare different operators where each has its own property, as illustrated in Fig. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 365, + 547, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 365, + 547, + 460 + ], + "spans": [ + { + "bbox": [ + 304, + 365, + 547, + 460 + ], + "type": "text", + "content": "Softmax normalization: A key difference between convolution and DCNv3 is that DCNv3 normalizes " + }, + { + "bbox": [ + 304, + 365, + 547, + 460 + ], + "type": "inline_equation", + "content": "\\mathbf{m}" + }, + { + "bbox": [ + 304, + 365, + 547, + 460 + ], + "type": "text", + "content": ", the spatial aggregation weights, with a softmax function, following the convention of scaled dot-product self-attention. Conversely, convolution does not employ softmax over its weights and still works well. The reason why attention needs a softmax is straightforward: scaled dot-product self-attention with " + }, + { + "bbox": [ + 304, + 365, + 547, + 460 + ], + "type": "inline_equation", + "content": "Q, K, V \\in \\mathbb{R}^{N \\times d}" + }, + { + "bbox": [ + 304, + 365, + 547, + 460 + ], + "type": "text", + "content": " is defined with a formulation:" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 378, + 466, + 546, + 492 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 378, + 466, + 546, + 492 + ], + "spans": [ + { + "bbox": [ + 378, + 466, + 546, + 492 + ], + "type": "interline_equation", + "content": "\\operatorname {s o f t m a x} \\left(\\frac {1}{\\sqrt {d}} Q K ^ {\\top}\\right) V, \\tag {3}", + "image_path": "4c1ff3663669c3a24a7c37276110aa28ca3aab612f805acdaee60cc19dc54e52.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 499, + 547, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 499, + 547, + 689 + ], + "spans": [ + { + "bbox": [ + 304, + 499, + 547, + 689 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 499, + 547, + 689 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 499, + 547, + 689 + ], + "type": "text", + "content": " is the number of points in the same attention window (can be either global [12] or local [25]), " + }, + { + "bbox": [ + 304, + 499, + 547, + 689 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 304, + 499, + 547, + 689 + ], + "type": "text", + "content": " is the hidden dimension, " + }, + { + "bbox": [ + 304, + 499, + 547, + 689 + ], + "type": "inline_equation", + "content": "Q, K, V" + }, + { + "bbox": [ + 304, + 499, + 547, + 689 + ], + "type": "text", + "content": " are the query, key, and value matrices computed from the input. Softmax operation is required in Eq. (3) for attention; without softmax, " + }, + { + "bbox": [ + 304, + 499, + 547, + 689 + ], + "type": "inline_equation", + "content": "K^{\\top}V \\in \\mathbb{R}^{d \\times d}" + }, + { + "bbox": [ + 304, + 499, + 547, + 689 + ], + "type": "text", + "content": " can be calculated first, and it degrades to a linear projection for all queries in the same attention window, resulting in degenerated performance. However, for convolutional operators like depthwise convolution and DCNv3 where each point has its own dedicated aggregation window and the values in each aggregation window are already different and there is no \"key\" concept, such degradation issues no longer exist, and the normalization becomes unnecessary. In fact, normalizing convolution weights within a fixed 0-1 range using softmax could impose a significant limitation on the operator's expressive power and make the learning slower." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 305, + 689, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 689, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 305, + 689, + 547, + 714 + ], + "type": "text", + "content": "To confirm this hypothesis, we train a ConvNeXt model and apply softmax to the " + }, + { + "bbox": [ + 305, + 689, + 547, + 714 + ], + "type": "inline_equation", + "content": "7 \\times 7" + }, + { + "bbox": [ + 305, + 689, + 547, + 714 + ], + "type": "text", + "content": " window of the depthwise" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5654" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "type": "text", + "content": "convolution weights before convolution forward. We observe a remarkable decline in model performance as well as convergence speed from results in Tab. 1. This suggests that for operators with a dedicated aggregation window on each location like convolution or DCN, aggregation weights with an unbounded range offer better expressive power than softmax-normalized, bounded-range weights." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 161, + 290, + 316 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 161, + 290, + 316 + ], + "spans": [ + { + "bbox": [ + 46, + 161, + 290, + 316 + ], + "type": "text", + "content": "Enhancing dynamic property: Motivated by this observation, we remove the softmax normalization in DCNv3, transforming the modulation scalars ranging from 0 to 1 to unbounded dynamic weights similar to convolution. As shown in Fig. 2, this alteration further amplifies the dynamic property of DCN, where other operators have certain limits, such as bounded value range (attention/DCNv3) or fixed aggregation window with input-independent aggregation weights (convolution). Fig. 1b shows that by making this change, DCNv4 converges significantly faster than DCNv3 and other common operators, including convolution and attention. Results in Sec. 4 further showcase that DCNv4 works well in both pre-training and transfer learning settings." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 319, + 154, + 332 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 319, + 154, + 332 + ], + "spans": [ + { + "bbox": [ + 47, + 319, + 154, + 332 + ], + "type": "text", + "content": "3.2. Speeding up DCN" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 336, + 289, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 336, + 289, + 468 + ], + "spans": [ + { + "bbox": [ + 46, + 336, + 289, + 468 + ], + "type": "text", + "content": "Theoretically, DCN, as a sparse operator with " + }, + { + "bbox": [ + 46, + 336, + 289, + 468 + ], + "type": "inline_equation", + "content": "3 \\times 3" + }, + { + "bbox": [ + 46, + 336, + 289, + 468 + ], + "type": "text", + "content": " window, should act faster than other common operators that employ larger window sizes, like dense attention or " + }, + { + "bbox": [ + 46, + 336, + 289, + 468 + ], + "type": "inline_equation", + "content": "7 \\times 7" + }, + { + "bbox": [ + 46, + 336, + 289, + 468 + ], + "type": "text", + "content": " depthwise convolution. However, we find that this is not the case, as shown in Fig. 1a. In this subsection, we first conduct a theoretical analysis of GPU efficiency, showing a large variance in memory access cost depending on how we read the memory. We further perform optimization based on our observations, significantly improving the speed of DCN by saving additional memory instruction and bringing the speed advantage of being a sparse operator into reality." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 475, + 288, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 475, + 288, + 582 + ], + "spans": [ + { + "bbox": [ + 46, + 475, + 288, + 582 + ], + "type": "text", + "content": "Theoretical analysis of GPU efficiency Our study begins with a theoretical examination of the DCNv3 operator's computational behavior. We employ the roofline model to evaluate its performance, focusing on theoretical FLOPs and memory access cost. For an input and output tensor of shape " + }, + { + "bbox": [ + 46, + 475, + 288, + 582 + ], + "type": "inline_equation", + "content": "(H,W,C)" + }, + { + "bbox": [ + 46, + 475, + 288, + 582 + ], + "type": "text", + "content": ", the DCNv3 operator requires " + }, + { + "bbox": [ + 46, + 475, + 288, + 582 + ], + "type": "inline_equation", + "content": "36HWC" + }, + { + "bbox": [ + 46, + 475, + 288, + 582 + ], + "type": "text", + "content": " FLOPs, where " + }, + { + "bbox": [ + 46, + 475, + 288, + 582 + ], + "type": "inline_equation", + "content": "3\\times 3" + }, + { + "bbox": [ + 46, + 475, + 288, + 582 + ], + "type": "text", + "content": " represents the convolution kernel's spatial dimensions and the factor of 4 accounts for the bilinear interpolation at each sampling point." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 583, + 289, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 583, + 289, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 583, + 289, + 713 + ], + "type": "text", + "content": "Following the framework outlined in [27], DCNv3's memory access cost is calculated as " + }, + { + "bbox": [ + 46, + 583, + 289, + 713 + ], + "type": "inline_equation", + "content": "2HWC + 27HWG" + }, + { + "bbox": [ + 46, + 583, + 289, + 713 + ], + "type": "text", + "content": ". The first term corresponds to the input/output feature map size and the second to the DCNv3's offset and aggregation weights with " + }, + { + "bbox": [ + 46, + 583, + 289, + 713 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 46, + 583, + 289, + 713 + ], + "type": "text", + "content": " groups. We approximate " + }, + { + "bbox": [ + 46, + 583, + 289, + 713 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 46, + 583, + 289, + 713 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 46, + 583, + 289, + 713 + ], + "type": "inline_equation", + "content": "C / 16" + }, + { + "bbox": [ + 46, + 583, + 289, + 713 + ], + "type": "text", + "content": " assuming a group dimension of 16, resulting in approximately " + }, + { + "bbox": [ + 46, + 583, + 289, + 713 + ], + "type": "inline_equation", + "content": "3.7HWC" + }, + { + "bbox": [ + 46, + 583, + 289, + 713 + ], + "type": "text", + "content": " memory access cost. However, this assumes an ideal scenario of infinite cache and a single memory read for each value, which is often unrealistic in parallel computing environments where concurrent thread execution necessitates simultaneous data access." + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 307, + 72, + 547, + 173 + ], + "blocks": [ + { + "bbox": [ + 307, + 72, + 547, + 173 + ], + "lines": [ + { + "bbox": [ + 307, + 72, + 547, + 173 + ], + "spans": [ + { + "bbox": [ + 307, + 72, + 547, + 173 + ], + "type": "image", + "image_path": "a17fa7dfce35b4a9b16f016b29f5033fa5c616e4468c2f79f26bc2189c681b73.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 177, + 547, + 232 + ], + "lines": [ + { + "bbox": [ + 304, + 177, + 547, + 232 + ], + "spans": [ + { + "bbox": [ + 304, + 177, + 547, + 232 + ], + "type": "text", + "content": "Figure 3. Illustration of our optimization. In DCNv4, we use one thread to process multiple channels in the same group that shares sampling offset and aggregation weights. Workloads like memory reading and bilinear interpolation coefficient computation can be reduced, and multiple memory access instructions can be merged." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 239, + 547, + 310 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 239, + 547, + 310 + ], + "spans": [ + { + "bbox": [ + 304, + 239, + 547, + 310 + ], + "type": "text", + "content": "To estimate the maximum memory access requirement, we consider a scenario devoid of cache, where each output location requires fresh memory reads and involves 36 reads for bilinear interpolation, 27 for offset/aggregation weights, and one write operation, resulting in a memory access cost of " + }, + { + "bbox": [ + 304, + 239, + 547, + 310 + ], + "type": "inline_equation", + "content": "64HWC" + }, + { + "bbox": [ + 304, + 239, + 547, + 310 + ], + "type": "text", + "content": ". This is 17 times larger than the ideal case." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 310, + 547, + 406 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 310, + 547, + 406 + ], + "spans": [ + { + "bbox": [ + 304, + 310, + 547, + 406 + ], + "type": "text", + "content": "This analysis reveals a substantial gap in the ratio of computation-to-memory access (ranging from 0.6 to 9.7), highlighting the significant potential for memory access optimization. Notably, despite DCNv3's use of input-dependent, dynamic offsets causing non-deterministic memory access, one deterministic thing is that channels within the same group share offset values. This leads us to propose a specific optimization strategy for enhancing DCNv3's speed." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 407, + 548, + 668 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 407, + 548, + 668 + ], + "spans": [ + { + "bbox": [ + 304, + 407, + 548, + 668 + ], + "type": "text", + "content": "Eliminating redundant workload: In previous CUDA implementations of DCN kernel, for input with shape " + }, + { + "bbox": [ + 304, + 407, + 548, + 668 + ], + "type": "inline_equation", + "content": "(H,W,C)^1" + }, + { + "bbox": [ + 304, + 407, + 548, + 668 + ], + "type": "text", + "content": ", offset " + }, + { + "bbox": [ + 304, + 407, + 548, + 668 + ], + "type": "inline_equation", + "content": "(H,W,G,K^2\\times 2)" + }, + { + "bbox": [ + 304, + 407, + 548, + 668 + ], + "type": "text", + "content": " and aggregation weight " + }, + { + "bbox": [ + 304, + 407, + 548, + 668 + ], + "type": "inline_equation", + "content": "(H,W,G,K^2)" + }, + { + "bbox": [ + 304, + 407, + 548, + 668 + ], + "type": "text", + "content": ", we will create " + }, + { + "bbox": [ + 304, + 407, + 548, + 668 + ], + "type": "inline_equation", + "content": "H\\times W\\times C" + }, + { + "bbox": [ + 304, + 407, + 548, + 668 + ], + "type": "text", + "content": " threads in total to maximize parallelism, where each thread processes one channel for one output location. Notably, the " + }, + { + "bbox": [ + 304, + 407, + 548, + 668 + ], + "type": "inline_equation", + "content": "D = C / G" + }, + { + "bbox": [ + 304, + 407, + 548, + 668 + ], + "type": "text", + "content": " channels within each group share the same sampling offset and aggregation weight values for each output location. Using multiple threads to process these " + }, + { + "bbox": [ + 304, + 407, + 548, + 668 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 304, + 407, + 548, + 668 + ], + "type": "text", + "content": " channels on the same output location is wasteful as different threads will read the same sampling offset and aggregation weight values from GPU memory multiple times, which is critical for a memory-bound operator. Processing multiple channels within the same group on each output location with one thread can eliminate these redundant memory read requests, greatly reducing memory bandwidth usage. As the sampling locations are the same, we can also only calculate the bilinear interpolation coefficient used in DCN once. Specifically, if each thread processes " + }, + { + "bbox": [ + 304, + 407, + 548, + 668 + ], + "type": "inline_equation", + "content": "D^{\\prime}" + }, + { + "bbox": [ + 304, + 407, + 548, + 668 + ], + "type": "text", + "content": " channels, the memory access cost for reading offset and aggregation weight, as well as the computation cost for calculating bilinear interpolation coefficient, can be reduced " + }, + { + "bbox": [ + 304, + 407, + 548, + 668 + ], + "type": "inline_equation", + "content": "D^{\\prime}" + }, + { + "bbox": [ + 304, + 407, + 548, + 668 + ], + "type": "text", + "content": " times." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 671, + 547, + 695 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 671, + 547, + 695 + ], + "spans": [ + { + "bbox": [ + 305, + 671, + 547, + 695 + ], + "type": "text", + "content": "Eliminating redundant memory instructions: In practice, solely reusing threads for multiple channels will not" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 317, + 702, + 531, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 702, + 531, + 713 + ], + "spans": [ + { + "bbox": [ + 317, + 702, + 531, + 713 + ], + "type": "text", + "content": "1We assume the batch size is 1 and memory layout is channel-last." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5655" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 73, + 70, + 518, + 176 + ], + "blocks": [ + { + "bbox": [ + 73, + 70, + 518, + 176 + ], + "lines": [ + { + "bbox": [ + 73, + 70, + 518, + 176 + ], + "spans": [ + { + "bbox": [ + 73, + 70, + 518, + 176 + ], + "type": "table", + "html": "
OperatorRuntime (ms)
56 × 56 × 12828 × 28 × 25614 × 14 × 5127 × 7 × 102414 × 14 × 768
Attention (torch)30.8 / 19.33.35 / 2.120.539 / 0.4480.446 / 0.1210.779 / 0.654
FlashAttention-2N/A / 2.46N/A / 0.451N/A / 0.123N/A / 0.0901N/A / 0.163
Window Attn (7 × 7)4.05 / 1.462.07 / 0.7701.08 / 0.4220.577 / 0.2391.58 / 0.604
DWConv (7 × 7, torch)2.02 / 1.981.03 / 1.000.515 / 0.5230.269 / 0.2610.779 / 0.773
DWConv (7 × 7, cuDNN)0.981 / 0.4380.522 / 0.2670.287 / 0.1530.199 / 0.1020.413 / 0.210
DCNv31.45 / 1.520.688 / 0.7110.294 / 0.2980.125 / 0.1260.528 / 0.548
DCNv40.606 / 0.4040.303 / 0.2300.145 / 0.1230.0730 / 0.06800.224 / 0.147
", + "image_path": "de434fa665229f69d642a31d1748b375d77c1ff611ac29f7a61e2f08bbfe600d.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 177, + 545, + 198 + ], + "lines": [ + { + "bbox": [ + 46, + 177, + 545, + 198 + ], + "spans": [ + { + "bbox": [ + 46, + 177, + 545, + 198 + ], + "type": "text", + "content": "Table 2. Op-level benchmark on standard input shape with various downsample rates. FP32/FP16 results are reported when the implementation is available. Our new DCNv4 can surpass all other commonly used operators under different input resolutions." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 45, + 205, + 289, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 205, + 289, + 548 + ], + "spans": [ + { + "bbox": [ + 45, + 205, + 289, + 548 + ], + "type": "text", + "content": "see speed improvement. The reason is that when " + }, + { + "bbox": [ + 45, + 205, + 289, + 548 + ], + "type": "inline_equation", + "content": "D'" + }, + { + "bbox": [ + 45, + 205, + 289, + 548 + ], + "type": "text", + "content": " increases, we create fewer threads and the workload of each thread now increases " + }, + { + "bbox": [ + 45, + 205, + 289, + 548 + ], + "type": "inline_equation", + "content": "D'" + }, + { + "bbox": [ + 45, + 205, + 289, + 548 + ], + "type": "text", + "content": " times. This essentially reduces the degree of parallelism for the CUDA kernel. Luckily, the DCN kernel is now computationally lightweight as the bilinear interpolation only needs to be performed once for all " + }, + { + "bbox": [ + 45, + 205, + 289, + 548 + ], + "type": "inline_equation", + "content": "D'" + }, + { + "bbox": [ + 45, + 205, + 289, + 548 + ], + "type": "text", + "content": " channels, and most of the workload is the memory instructions reading input values from different channels. When the memory layout is channel-last, and all " + }, + { + "bbox": [ + 45, + 205, + 289, + 548 + ], + "type": "inline_equation", + "content": "D'" + }, + { + "bbox": [ + 45, + 205, + 289, + 548 + ], + "type": "text", + "content": " channel values are contiguous, we can leverage vectorized load: for example, to read four 32-bit float values from memory, instead of reading one 32-bit float value four times with four instructions, we can use a single instruction to load a 128-bit packed value at once, thus reducing the number of instructions and execution time of each thread. We can apply similar technique when writing the final results to GPU memory, minimizing the memory access time and increasing memory bandwidth utilization. Moreover, the modern half-precision data format (float16/bfloat16) halves the bytes that need to be loaded, which means the memory efficiency can be twice as much under the same memory bandwidth when using the half-precision format. However, we do not see speed improvement with half-precision data in the original DCNv3 implementation, possibly due to too much overhead on data access and computation, while in our new implementation, the speedup is significant. It is worth noting that the aforementioned optimization techniques can also be applied to DCNv1/v2 and deformable attention [48], as they share a similar performance bottleneck and issue." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 559, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 559, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 559, + 288, + 715 + ], + "type": "text", + "content": "Micro design in DCN module: DCNv3 module introduces multiple micro designs; as the core kernel is optimized, their impact on the speed becomes non-negligible. We identify two points in DCNv3 designs that could be further optimized: first, after removing the softmax and transforming the modulation scalar into dynamic aggregation weights as mentioned in the previous paragraph. The linear layers for computing offset and dynamic weights can actually be combined into one linear layer. This reduces network fragmentation and eliminates extra overheads, such as kernel launching and synchronization, enhancing run-time efficiency on the GPU; second, in the original DCNv3 module design, a complex sub-network that consists of depthwise " + }, + { + "bbox": [ + 46, + 559, + 288, + 715 + ], + "type": "inline_equation", + "content": "3 \\times 3" + }, + { + "bbox": [ + 46, + 559, + 288, + 715 + ], + "type": "text", + "content": " conv, layer" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 304, + 205, + 547, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 205, + 547, + 289 + ], + "spans": [ + { + "bbox": [ + 304, + 205, + 547, + 289 + ], + "type": "text", + "content": "normalization (LN), GELU, and linear layer is used to compute offsets and dynamic weights. Following the design in Xception [6], we remove the additional LN-GELU layers and use the original separable convolution structure, further reducing running time. We empirically find that if latency is a higher priority, the depthwise convolution can also be removed with only a minor performance sacrifice." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 305, + 292, + 388, + 306 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 292, + 388, + 306 + ], + "spans": [ + { + "bbox": [ + 305, + 292, + 388, + 306 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 312, + 547, + 397 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 312, + 547, + 397 + ], + "spans": [ + { + "bbox": [ + 304, + 312, + 547, + 397 + ], + "type": "text", + "content": "In this section, we verify the effectiveness of our proposed DCNv4 module from both speed as well as performance perspective. We benchmark the operator-level speed and integrate DCNv4 into the backbone to test the system-level performance. All speed test results are obtained with an NVIDIA A100 80G SXM GPU. We include additional experimental results and implementation details in supp." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 401, + 481, + 415 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 401, + 481, + 415 + ], + "spans": [ + { + "bbox": [ + 305, + 401, + 481, + 415 + ], + "type": "text", + "content": "4.1. Speed Benchmark for Operators" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 417, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 417, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 417, + 547, + 715 + ], + "type": "text", + "content": "Settings: We conduct the op-level benchmark by solely measuring the running time of several representative operators building state-of-the-art vision backbone models, including full attention [35] implemented with PyTorch as well as the advanced FlashAttention-2 [8] implementation, window-based attention with window size " + }, + { + "bbox": [ + 304, + 417, + 547, + 715 + ], + "type": "inline_equation", + "content": "7 \\times 7" + }, + { + "bbox": [ + 304, + 417, + 547, + 715 + ], + "type": "text", + "content": " [25], depthwise convolution with " + }, + { + "bbox": [ + 304, + 417, + 547, + 715 + ], + "type": "inline_equation", + "content": "7 \\times 7" + }, + { + "bbox": [ + 304, + 417, + 547, + 715 + ], + "type": "text", + "content": " window, implemented by cuDNN [5] and ATen library from PyTorch [28], respectively. We only benchmark the core operation for spatial aggregation, and additional linear layers like qkv projection and output projection layers are excluded and not included in the runtime measurement. Please refer to supp. for a more comprehensive module-level comparison. We first consider a feature map shape generated from the standard " + }, + { + "bbox": [ + 304, + 417, + 547, + 715 + ], + "type": "inline_equation", + "content": "224 \\times 224" + }, + { + "bbox": [ + 304, + 417, + 547, + 715 + ], + "type": "text", + "content": " input resolution for image classification with 4, 8, 16, " + }, + { + "bbox": [ + 304, + 417, + 547, + 715 + ], + "type": "inline_equation", + "content": "32 \\times" + }, + { + "bbox": [ + 304, + 417, + 547, + 715 + ], + "type": "text", + "content": " downsample ratio as used by common hierarchical ConvNet/transformer backbones; we also add a feature map shape from isotropic backbone like ViT with a downsampling ratio 16 and larger hidden dimension. We further consider high-resolution inputs often used in downstream tasks like object detection. We set the input shape to be " + }, + { + "bbox": [ + 304, + 417, + 547, + 715 + ], + "type": "inline_equation", + "content": "800 \\times 1280" + }, + { + "bbox": [ + 304, + 417, + 547, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 417, + 547, + 715 + ], + "type": "inline_equation", + "content": "1024 \\times 1024" + }, + { + "bbox": [ + 304, + 417, + 547, + 715 + ], + "type": "text", + "content": " for the hierarchical feature map and isotropic feature map, respectively, as they are the common practice in object detection [15, 20]. Batch size is 64 and 1 for these two input sets, respectively. For operators with a head/group concept, we" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "5656" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 75, + 70, + 519, + 173 + ], + "blocks": [ + { + "bbox": [ + 75, + 70, + 519, + 173 + ], + "lines": [ + { + "bbox": [ + 75, + 70, + 519, + 173 + ], + "spans": [ + { + "bbox": [ + 75, + 70, + 519, + 173 + ], + "type": "table", + "html": "
OperatorRuntime (ms)
200 × 320 × 128100 × 160 × 25650 × 80 × 51225 × 40 × 102464 × 64 × 768
Attention (torch)OOM / OOM25.4 / 12.92.88 / 1.890.490 / 0.3094.17 / 2.57
FlashAttention-2N/A / 13.2N/A / 1.74N/A / 0.285N/A / 0.0797N/A / 0.437
Window Attn (7 × 7)1.33 / 0.5090.728 / 0.2910.426 / 0.1860.279 / 0.1650.673 / 0.272
DWConv (7 × 7, torch)0.634 / 0.6080.313 / 0.3150.167 / 0.1580.0943 / 0.08940.260 / 0.253
DWConv (7 × 7, cuDNN)0.331 / 0.2820.188 / 0.1680.114 / 0.1150.0817 / 0.08810.161 / 0.156
DCNv30.472 / 0.4930.244 / 0.2530.128 / 0.1320.0737 / 0.07670.194 / 0.199
DCNv40.210 / 0.1360.124 / 0.08950.0707 / 0.05890.0452 / 0.04260.103 / 0.0672
", + "image_path": "29932da64e9391ff28d8aa823af265a03abd49cfeaaea5f0526aebc12874c911.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 209, + 288, + 233 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 209, + 288, + 233 + ], + "spans": [ + { + "bbox": [ + 46, + 209, + 288, + 233 + ], + "type": "text", + "content": "set the dimension of each head/group to 32 and change the number of heads/groups when the hidden dimension varies." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 235, + 290, + 461 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 235, + 290, + 461 + ], + "spans": [ + { + "bbox": [ + 46, + 235, + 290, + 461 + ], + "type": "text", + "content": "Results: We show the benchmark results on standard resolution and high-resolution input in Tab. 2 and Tab. 3. We report results with both FP32 and FP16 data formats unless the FP32 implementation is not available. Dense global attention implemented with PyTorch performs significantly slower when the input resolution is large and even out of memory. FlashAttention significantly improves the speed of attention and can be even faster than " + }, + { + "bbox": [ + 46, + 235, + 290, + 461 + ], + "type": "inline_equation", + "content": "7 \\times 7" + }, + { + "bbox": [ + 46, + 235, + 290, + 461 + ], + "type": "text", + "content": " window attention in certain cases, indicating the importance of proper optimization. However, it does not change the quadratic complexity of attention; when the input resolution is high, it still falls behind local/sparse operators like window attention or convolution. While DCNv3 can be faster than DWConv with plain implementation, it is slower than the heavily optimized cuDNN version. Instead, our DCNv4 can provide more than " + }, + { + "bbox": [ + 46, + 235, + 290, + 461 + ], + "type": "inline_equation", + "content": "3 \\times" + }, + { + "bbox": [ + 46, + 235, + 290, + 461 + ], + "type": "text", + "content": " speedup compared to DCNv3, greatly saving the running time. Moreover, DCNv4 can leverage the advantage of using a " + }, + { + "bbox": [ + 46, + 235, + 290, + 461 + ], + "type": "inline_equation", + "content": "3 \\times 3" + }, + { + "bbox": [ + 46, + 235, + 290, + 461 + ], + "type": "text", + "content": " sparse window to perform significantly faster than other baselines under different settings." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 466, + 165, + 479 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 466, + 165, + 479 + ], + "spans": [ + { + "bbox": [ + 47, + 466, + 165, + 479 + ], + "type": "text", + "content": "4.2. Image Classification" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 482, + 288, + 672 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 482, + 288, + 672 + ], + "spans": [ + { + "bbox": [ + 46, + 482, + 288, + 672 + ], + "type": "text", + "content": "Settings: We evaluate the effectiveness of DCNv4 on ImageNet classification. We start from InternImage [38] as it shows state-of-the-art performance among ConvNet-based models. We replace the DCNv3 in InternImage with DCNv4 and create FlashInternImage. Other implementation details, including network architecture and hyperparameters, are kept the same as [38]. We compare Swin-Transformer and ConvNeXt which are two representative baselines in Transformer and ConvNet models. We follow the common practice [25, 26, 38] of training protocols, including data augmentation, preprocessing, optimizer and learning rate schedule, and train FlashInternImage-Tiny/Small/Base on ImageNet-1K for 300 epochs. FlashInternImage-Large is trained on ImageNet-22K for 90 epochs and then fine-tuned on ImageNet-1K for 20 epochs. Other baselines share the same setting for a fair comparison." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 677, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 289, + 715 + ], + "type": "text", + "content": "Results: Tab. 4 shows the results of models at various scales. Besides the model size and training/inference resolution, we also report each model's overall throughput (number" + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 322, + 207, + 531, + 458 + ], + "blocks": [ + { + "bbox": [ + 46, + 179, + 547, + 201 + ], + "lines": [ + { + "bbox": [ + 46, + 179, + 547, + 201 + ], + "spans": [ + { + "bbox": [ + 46, + 179, + 547, + 201 + ], + "type": "text", + "content": "Table 3. Op-level benchmark on high-resolution input shape with various downsample rates. DCNv4 performs well as a sparse operator, surpassing all other baselines, while dense global attention is slow under this scenario." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 322, + 207, + 531, + 458 + ], + "lines": [ + { + "bbox": [ + 322, + 207, + 531, + 458 + ], + "spans": [ + { + "bbox": [ + 322, + 207, + 531, + 458 + ], + "type": "table", + "html": "
ModelSizeScaleAccThroughput
Swin-T29M224281.31989 / 3619
ConvNeXt-T29M224282.12485 / 4305
InternImage-T30M224283.51409 / 1746
FlashInternImage-T30M224283.62316 / 3154 (+64%/ +80%)
Swin-S50M224283.01167 / 2000
ConvNeXt-S50M224283.11645 / 2538
InternImage-S50M224284.21044 / 1321
FlashInternImage-S50M224284.41625 / 2396 (+56%/ +81%)
Swin-B88M224283.5934 / 1741
ConvNeXt-B89M224283.81241 / 1888
InternImage-B97M224284.9779 / 1030
FlashInternImage-B97M224284.91174 / 1816 (+51%/ +76%)
Swin-L197M384287.3206 / 301
ConvNeXt-L198M384287.5252 / 436
InternImage-L223M384287.7158 / 214
ConvNeXt-XL350M384287.8170 / 299
InternImage-XL335M384288.0125 / 174
FlashInternImage-L223M384288.1248 / 401 (+57%/ +87%)
", + "image_path": "24ed477e96994e9d152d8492f9d83aee63b62dcf71c0ad293622162e099c84b0.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 461, + 547, + 506 + ], + "lines": [ + { + "bbox": [ + 304, + 461, + 547, + 506 + ], + "spans": [ + { + "bbox": [ + 304, + 461, + 547, + 506 + ], + "type": "text", + "content": "Table 4. Image classification performance on ImageNet-1K. We show relative speedup between FlashInternImage w/ DCNv4 and its InternImage counterparts. DCNv4 significantly improves the speed while shows state-of-the-art performance." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 515, + 548, + 670 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 515, + 548, + 670 + ], + "spans": [ + { + "bbox": [ + 304, + 515, + 548, + 670 + ], + "type": "text", + "content": "of images per second) in FP32/FP16 data format. We use timm [39] implementation of ConvNeXt and Swin Transformer, which is faster than the original implementation. Equipped with DCNv4, FlashInternImage significantly improves the throughput of the InternImage counterpart over " + }, + { + "bbox": [ + 304, + 515, + 548, + 670 + ], + "type": "inline_equation", + "content": "50\\% \\sim 80\\%" + }, + { + "bbox": [ + 304, + 515, + 548, + 670 + ], + "type": "text", + "content": " and slightly improves the model performance. FlashInternImage now matches the speed of ConvNeXt with higher accuracy. It is noteworthy that FlashInternImage-S can outperform ConvNeXt-B (84.4% vs. 83.8%) while being faster than it, showing a better speed-accuracy trade-off. Moreover, the FlashInternImage-L can surpass ConvNeXt-XL and InternImage-XL and being " + }, + { + "bbox": [ + 304, + 515, + 548, + 670 + ], + "type": "inline_equation", + "content": "30\\% \\sim 130\\%" + }, + { + "bbox": [ + 304, + 515, + 548, + 670 + ], + "type": "text", + "content": " (401 vs. 174) faster, showing the effectiveness of our DCNv4 module." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 675, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 675, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 675, + 547, + 715 + ], + "type": "text", + "content": "4.3. Downstream Tasks with High-Resolution Input We evaluate the performance of DCNv4 on representative downstream perception tasks with high-resolution input, in" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 758 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 758 + ], + "type": "text", + "content": "5657" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 52, + 71, + 284, + 228 + ], + "blocks": [ + { + "bbox": [ + 52, + 71, + 284, + 228 + ], + "lines": [ + { + "bbox": [ + 52, + 71, + 284, + 228 + ], + "spans": [ + { + "bbox": [ + 52, + 71, + 284, + 228 + ], + "type": "table", + "html": "
Model#paramFPSMask R-CNN
1×3×+MS
APbAPmAPbAPm
Swin-T48M66 / 10642.739.346.041.6
ConvNeXt-T48M78 / 11344.240.146.241.7
InternImage-T49M54 / 6947.242.549.143.7
FlashInternImage-T49M72 / 10248.043.149.544.0
Swin-S69M45 / 7744.840.948.243.2
ConvNeXt-S70M54 / 8345.441.847.942.9
InternImage-S69M44 / 5647.843.349.744.5
FlashInternImage-S69M57 / 8349.244.050.544.9
Swin-B107M33 / 5946.942.348.643.3
ConvNeXt-B108M43 / 7047.042.748.543.5
InternImage-B115M33 / 4348.844.050.344.8
FlashInternImage-B115M44 / 6750.144.550.645.4
", + "image_path": "0f4a17346f2f07a4ee070d79acb1ee2d90f69189c3fbc7d019b675c9bf64c9b6.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 52, + 237, + 284, + 332 + ], + "blocks": [ + { + "bbox": [ + 52, + 237, + 284, + 332 + ], + "lines": [ + { + "bbox": [ + 52, + 237, + 284, + 332 + ], + "spans": [ + { + "bbox": [ + 52, + 237, + 284, + 332 + ], + "type": "table", + "html": "
Model#paramFPSCascade Mask R-CNN
1×3×+MS
APbAPmAPbAPm
Swin-L253M20 / 2651.844.953.946.7
ConvNeXt-L255M26 / 4053.546.454.847.6
InternImage-L277M20 / 2654.947.756.148.5
ConvNeXt-XL407M21 / 3253.646.555.247.7
InternImage-XL387M16 / 2355.348.156.248.8
FlashInternImage-L277M26 / 3955.648.256.748.9
", + "image_path": "ee8565477bd4bdd618637c0b5da4e85d51d3c712a39e22a87d495ed1930601be.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 423, + 288, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 423, + 288, + 483 + ], + "spans": [ + { + "bbox": [ + 47, + 423, + 288, + 483 + ], + "type": "text", + "content": "cluding instance segmentation, semantic segmentation and 3D object detection. We keep all implementation details the same as InternImage and only change the backbone model. The backbone models are initialized from the ImageNet pretrained weights when training the downstream models." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 491, + 288, + 622 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 491, + 288, + 622 + ], + "spans": [ + { + "bbox": [ + 46, + 491, + 288, + 622 + ], + "type": "text", + "content": "Instance Segmentation: We train FlashInternImage with two representative instance segmentation frameworks, Mask R-CNN [15] and Cascade Mask-RCNN [2], on COCO dataset [23] at " + }, + { + "bbox": [ + 46, + 491, + 288, + 622 + ], + "type": "inline_equation", + "content": "1 \\times (12" + }, + { + "bbox": [ + 46, + 491, + 288, + 622 + ], + "type": "text", + "content": " epochs) and " + }, + { + "bbox": [ + 46, + 491, + 288, + 622 + ], + "type": "inline_equation", + "content": "3 \\times (36" + }, + { + "bbox": [ + 46, + 491, + 288, + 622 + ], + "type": "text", + "content": " epochs) training schedules. The results are shown in Tab. 5. We also report FPS with batch size 16 in FP32/FP16 data format. FlashInternImage shows superior results on all model scales and training schedules, achieving a higher speed-accuracy tradeoff. For example, FlashInternImage-T/S surpasses all other models with the same scale and is on par with a larger InternImage-S/B while being " + }, + { + "bbox": [ + 46, + 491, + 288, + 622 + ], + "type": "inline_equation", + "content": "80\\% - 90\\%" + }, + { + "bbox": [ + 46, + 491, + 288, + 622 + ], + "type": "text", + "content": " faster." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 627, + 288, + 723 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 627, + 288, + 723 + ], + "spans": [ + { + "bbox": [ + 46, + 627, + 288, + 723 + ], + "type": "text", + "content": "Semantic Segmentation: We train FlashInternImage with UperNet [41] on ADE20K [46] dataset for 160K iterations. We can draw a similar conclusion as instance segmentation from the results in Tab. 6, with FPS reported with batch size 16 in FP32/FP16. FlashInternImage w/ DCNv4 can achieve significantly faster speed and further improve the performance of InternImage across different model scales, resulting in a new state-of-the-art." + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 316, + 71, + 537, + 284 + ], + "blocks": [ + { + "bbox": [ + 47, + 338, + 288, + 415 + ], + "lines": [ + { + "bbox": [ + 47, + 338, + 288, + 415 + ], + "spans": [ + { + "bbox": [ + 47, + 338, + 288, + 415 + ], + "type": "text", + "content": "Table 5. Object detection and instance segmentation performance on COCO val2017. " + }, + { + "bbox": [ + 47, + 338, + 288, + 415 + ], + "type": "inline_equation", + "content": "\\mathrm{AP^b}" + }, + { + "bbox": [ + 47, + 338, + 288, + 415 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 338, + 288, + 415 + ], + "type": "inline_equation", + "content": "\\mathrm{AP^m}" + }, + { + "bbox": [ + 47, + 338, + 288, + 415 + ], + "type": "text", + "content": " denotes box AP and mask AP, respectively. \"MS\" means multi-scale training. FlashInternImage w/ DCNv4 models converge faster, clearly outperform other baselines with " + }, + { + "bbox": [ + 47, + 338, + 288, + 415 + ], + "type": "inline_equation", + "content": "1\\times" + }, + { + "bbox": [ + 47, + 338, + 288, + 415 + ], + "type": "text", + "content": " training schedule, and still maintain a leading position when training " + }, + { + "bbox": [ + 47, + 338, + 288, + 415 + ], + "type": "inline_equation", + "content": "3\\times" + }, + { + "bbox": [ + 47, + 338, + 288, + 415 + ], + "type": "text", + "content": " longer while being significantly faster than InternImage." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 316, + 71, + 537, + 284 + ], + "lines": [ + { + "bbox": [ + 316, + 71, + 537, + 284 + ], + "spans": [ + { + "bbox": [ + 316, + 71, + 537, + 284 + ], + "type": "table", + "html": "
Modelcrop size#paramFPSmIoU (SS)mIoU (MS)
Swin-T512260M107 / 16844.545.8
ConvNeXt-T512260M120 / 18446.046.7
InternImage-T512259M100 / 13947.948.1
FlashInternImage-T512259M119 / 20649.350.3
Swin-S512281M89 / 14247.649.5
ConvNeXt-S512282M107 / 16448.749.6
InternImage-S512280M89 / 12350.150.9
FlashInternImage-S512280M107 / 18250.651.6
Swin-B5122121M77 / 12648.149.7
ConvNeXt-B5122122M95 / 14749.149.9
InternImage-B5122128M77 / 10450.851.3
FlashInternImage-B5122128M94 / 15752.052.6
Swin-L6402234M59 / 9952.153.5
ConvNeXt-L6402235M73 / 11753.253.7
InternImage-L6402256M56 / 7853.954.1
ConvNeXt-XL6402391M53 / 7553.654.0
InternImage-XL6402368M47 / 6755.055.3
FlashInternImage-L6402256M71 / 12255.656.0
", + "image_path": "e9447db2f16d972055e89f5aa24871f3076499b6addcfc248cae1a230890b0ab.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 331, + 346, + 518, + 403 + ], + "blocks": [ + { + "bbox": [ + 304, + 287, + 547, + 343 + ], + "lines": [ + { + "bbox": [ + 304, + 287, + 547, + 343 + ], + "spans": [ + { + "bbox": [ + 304, + 287, + 547, + 343 + ], + "type": "text", + "content": "Table 6. Semantic segmentation performance on the ADE20K validation set. All models are trained with UperNet. \"SS\" and \"MS\" denote single-scale and multi-scale testing, respectively. FPS is reported with single-scale testing. FlashInternImage w/ DCNv4 achieves superior performance with competitive speed." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 331, + 346, + 518, + 403 + ], + "lines": [ + { + "bbox": [ + 331, + 346, + 518, + 403 + ], + "spans": [ + { + "bbox": [ + 331, + 346, + 518, + 403 + ], + "type": "table", + "html": "
ModelNDSmAPFPS†FPS
InternImage-B62.054.08.02.7
InternImage-XL63.455.64.02.0
FlashInternImage-S61.755.516.84.1
FlashInternImage-B63.157.412.13.8
", + "image_path": "f38102fea077724c866ca67aa14c746b9cf552d65a0e517f5793e5938af9eb88.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 406, + 547, + 472 + ], + "lines": [ + { + "bbox": [ + 304, + 406, + 547, + 472 + ], + "spans": [ + { + "bbox": [ + 304, + 406, + 547, + 472 + ], + "type": "text", + "content": "Table 7. 3D detection performance of BEVFormer v2 on nuScenes test set. We report backbone FPS (denoted with " + }, + { + "bbox": [ + 304, + 406, + 547, + 472 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 304, + 406, + 547, + 472 + ], + "type": "text", + "content": "), overall FPS results with underoptimized head implementation are also added for reference. With on-par NDS and higher mAP results, FlashInternImage can be " + }, + { + "bbox": [ + 304, + 406, + 547, + 472 + ], + "type": "inline_equation", + "content": "50 - 90\\%" + }, + { + "bbox": [ + 304, + 406, + 547, + 472 + ], + "type": "text", + "content": " faster than InternImage baselines or " + }, + { + "bbox": [ + 304, + 406, + 547, + 472 + ], + "type": "inline_equation", + "content": "200\\% - 300\\%" + }, + { + "bbox": [ + 304, + 406, + 547, + 472 + ], + "type": "text", + "content": " when only considering the backbone." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 485, + 547, + 663 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 485, + 547, + 663 + ], + "spans": [ + { + "bbox": [ + 304, + 485, + 547, + 663 + ], + "type": "text", + "content": "3D Detection: We further test DCNv4 on the camera-based 3D object detection task in the autonomous driving scenario, We train BEVFormer v2 [43], a state-of-the-art multi-camera 3D object detector, with FlashInternImage-Small and Base backbone models on the nuScenes dataset for 24 epochs. We report results on the nuScenes test set in Tab. 7 with FPS for each model. We note that the header parts, such as the BEV encoder and object decoder in BEV-Former v2, are underoptimized and take more than " + }, + { + "bbox": [ + 304, + 485, + 547, + 663 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 304, + 485, + 547, + 663 + ], + "type": "text", + "content": " of the running time (and even more with a faster backbone); thus, we also report the FPS for the backbone for a clearer illustration. Our results show that when only considering the backbone, FlashInternImage can be twice or even three times faster than the InternImage backbone with an on-par performance, greatly increasing the model efficiency." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 671, + 479, + 685 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 671, + 479, + 685 + ], + "spans": [ + { + "bbox": [ + 306, + 671, + 479, + 685 + ], + "type": "text", + "content": "4.4. DCNv4 as a Universal Operator" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 689, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 547, + 713 + ], + "type": "text", + "content": "Drop-in replacement in other vision backbones : Weverify whether DCNv4 can still work well in architectures de-" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5658" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 72, + 71, + 260, + 141 + ], + "blocks": [ + { + "bbox": [ + 72, + 71, + 260, + 141 + ], + "lines": [ + { + "bbox": [ + 72, + 71, + 260, + 141 + ], + "spans": [ + { + "bbox": [ + 72, + 71, + 260, + 141 + ], + "type": "table", + "html": "
ModelIN-1K AccThroughput
ConvNeXt-B83.81241 / 1888
ConvNeXt-B + DCNv484.01495 / 2513 (+20% / +33%)
ViT-B81.81683 / 2781†
ViT-B + DCNv481.92092 / 3261 (+24% / +17%)
", + "image_path": "130974bdecf54d90cdddb3a1dfe3d5758068be568f78e5aeeabd497b8f5c0c7d.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 73, + 195, + 260, + 239 + ], + "blocks": [ + { + "bbox": [ + 46, + 144, + 288, + 189 + ], + "lines": [ + { + "bbox": [ + 46, + 144, + 288, + 189 + ], + "spans": [ + { + "bbox": [ + 46, + 144, + 288, + 189 + ], + "type": "text", + "content": "Table 8. DCNv4 in other architecture. We show supervised learning results on ImageNet-1K and throughput. DCNv4 achieves higher throughput with comparable accuracy. " + }, + { + "bbox": [ + 46, + 144, + 288, + 189 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 46, + 144, + 288, + 189 + ], + "type": "text", + "content": " denotes testing with the advanced FlashAttention-2 implementation." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 73, + 195, + 260, + 239 + ], + "lines": [ + { + "bbox": [ + 73, + 195, + 260, + 239 + ], + "spans": [ + { + "bbox": [ + 73, + 195, + 260, + 239 + ], + "type": "table", + "html": "
Model#paramFIDFPS
U-Net860M2.944.82
U-Net + DCNv4566M2.444.92
", + "image_path": "40fed1164be495c2d4b5d84bfea672f66785761adb63631f131c1e362e0704f9.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 297, + 288, + 429 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 297, + 288, + 429 + ], + "spans": [ + { + "bbox": [ + 46, + 297, + 288, + 429 + ], + "type": "text", + "content": "signed with other operators, such as ConvNeXt and ViT. To achieve that, we replace the attention module and depthwise convolution layer with DCNv4 in ViT and ConvNeXt and perform supervised learning on ImageNet-1K without changing any other architecture and hyperparameters, similar to FlashInternImage and InternImage. The results are shown in Tab. 8. We can see that on these architectures, which are carefully tuned for the specific operators, our DCNv4 can perform equally well. Thanks to the fast speed of DCNv4, the new model can even achieve better throughput, showcasing the superior performance of DCNv4." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 437, + 288, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 437, + 288, + 651 + ], + "spans": [ + { + "bbox": [ + 46, + 437, + 288, + 651 + ], + "type": "text", + "content": "Drop-in replacement in diffusion model: DCN has been recognized to be an effective operator for perception tasks. As generative models become a fundamental tool for AI-generated content (AIGC), we are also curious if DCNv4 can work well on generation tasks with diffusion-based generative models. Specifically, we choose the U-Net [30] used in Stable Diffusion [29] as our baselines and replace the attention module and regular " + }, + { + "bbox": [ + 46, + 437, + 288, + 651 + ], + "type": "inline_equation", + "content": "3 \\times 3" + }, + { + "bbox": [ + 46, + 437, + 288, + 651 + ], + "type": "text", + "content": " convolution in U-Net. We use U-ViT's codebase, follow its training schedule, and train a latent diffusion model based on the image latent extracted from an image autoencoder provided by Stable Diffusion. We show the results in Tab. 9. We can see that DCNv4 also works well in generative modeling, achieving better results in terms of FID/Throughput with fewer parameters compared to regular convolution in U-Net. Notice that the architecture/hyperparameters may not be optimal for DCNv4, and it is possible that re-designing the models or searching for new hyperparameters for DCNv4 will give better results." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 658, + 142, + 672 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 658, + 142, + 672 + ], + "spans": [ + { + "bbox": [ + 47, + 658, + 142, + 672 + ], + "type": "text", + "content": "4.5. Ablation Study" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 677, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 288, + 715 + ], + "type": "text", + "content": "We conduct ablation studies in our optimization choice described in Sec. 3.2. The results are shown in Tab. 10. The time in the table is obtained with " + }, + { + "bbox": [ + 46, + 677, + 288, + 715 + ], + "type": "inline_equation", + "content": "56 \\times 56 \\times 128" + }, + { + "bbox": [ + 46, + 677, + 288, + 715 + ], + "type": "text", + "content": " input with" + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 326, + 71, + 525, + 158 + ], + "blocks": [ + { + "bbox": [ + 46, + 243, + 288, + 288 + ], + "lines": [ + { + "bbox": [ + 46, + 243, + 288, + 288 + ], + "spans": [ + { + "bbox": [ + 46, + 243, + 288, + 288 + ], + "type": "text", + "content": "Table 9. Class conditional generation on ImageNet 256x256. Latent diffusion models are trained from scratch with U-Net. We replace the convolution layer in the models with DCNv4. DCNv4 can achieve better FID results without any hyperparameter tuning." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 326, + 71, + 525, + 158 + ], + "lines": [ + { + "bbox": [ + 326, + 71, + 525, + 158 + ], + "spans": [ + { + "bbox": [ + 326, + 71, + 525, + 158 + ], + "type": "table", + "html": "
Implementation variantModuleKernel
Original DCNv33.281.45
- micro design2.121.45
- redundant memory access2.201.53
- redundant computation2.181.51
- redundant memory instr.1.280.606
- half-precision format0.8730.404
", + "image_path": "34393397696dbc313a2e1b9b9a7111411c5e47d979865320990896d17d34a300.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 161, + 547, + 195 + ], + "lines": [ + { + "bbox": [ + 304, + 161, + 547, + 195 + ], + "spans": [ + { + "bbox": [ + 304, + 161, + 547, + 195 + ], + "type": "text", + "content": "Table 10. Ablation studies of DCN's runtime (ms). We show how to achieve DCNv4 (gray row) from the original DCNv3 implementation and how different design choices affect the speed." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 202, + 547, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 202, + 547, + 486 + ], + "spans": [ + { + "bbox": [ + 304, + 202, + 547, + 486 + ], + "type": "text", + "content": "batch size 64 and 4 groups (32 channels per group). We first remove the softmax operation and improve the micro design, which means we merge the two linear layers into one and remove costly layer norm and GELU activation in offset/aggregation weight computing, simplifying the overall modules and increasing the speed. We then start modifying the kernel implementation. First, we change the parallel execution pattern and let each thread process 8 channels instead of 1 channel, thus, unnecessary memory access on loading sampling offset and aggregation weight values from the GPU memory can be saved. As we expected, solely applying this change will not increase the speed as the degree of parallelism decreases, and each thread's workload increases 8 times now. The latency is increased instead. Eliminating redundant computation by reusing the bilinear interpolation coefficient (4th row) can save some time but is insignificant. Removing the redundant memory instruction via vectorized load/store can greatly reduce the workload of each thread and largely accelerate the GPU kernel speed (5th row). Using a half-precision datatype, which halves the number of bytes the kernel needs to read/write, further increases the data throughput, as shown in the 6th row. In the end, we reach the final DCNv4 design, which is three times more efficient than the original implementation." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 495, + 379, + 507 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 495, + 379, + 507 + ], + "spans": [ + { + "bbox": [ + 306, + 495, + 379, + 507 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 512, + 547, + 655 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 512, + 547, + 655 + ], + "spans": [ + { + "bbox": [ + 304, + 512, + 547, + 655 + ], + "type": "text", + "content": "We present Deformable Convolution v4 (DCNv4), an efficient dynamic and sparse operator. By rethinking the dynamic property in deformable convolution and streamlining memory access, DCNv4 emerges as a much faster and more effective operator than its predecessor DCNv3. DCNv4-equipped FlashInternImage backbone not only enhances speed but also improves performance across various vision tasks. We further show DCNv4's versatility and effectiveness as a universal operator by integrating it into state-of-the-art architecture like ConvNeXt and ViT with improved throughput and accuracy; and it also works well in latent diffusion model, showing its potential to enhance generative models." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 663, + 404, + 677 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 663, + 404, + 677 + ], + "spans": [ + { + "bbox": [ + 306, + 663, + 404, + 677 + ], + "type": "text", + "content": "Acknowledgement" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 680, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 680, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 680, + 547, + 713 + ], + "type": "text", + "content": "The work is supported by the National Key R&D Program of China (NO. 2022ZD0161300), by the National Natural Science Foundation of China (62376134)." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5659" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 95, + 106, + 107 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 95, + 106, + 107 + ], + "spans": [ + { + "bbox": [ + 48, + 95, + 106, + 107 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 114, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 53, + 114, + 287, + 158 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 114, + 287, + 158 + ], + "spans": [ + { + "bbox": [ + 53, + 114, + 287, + 158 + ], + "type": "text", + "content": "[1] Saehyun Ahn, Jung-Woo Chang, and Suk-Ju Kang. An efficient accelerator design methodology for deformable convolutional networks. In 2020 IEEE International Conference on Image Processing (ICIP), pages 3075-3079. IEEE, 2020. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 159, + 288, + 202 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 159, + 288, + 202 + ], + "spans": [ + { + "bbox": [ + 53, + 159, + 288, + 202 + ], + "type": "text", + "content": "[2] Zhaowei Cai and Nuno Vasconcelos. Cascade r-cnn: Delving into high quality object detection. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6154-6162, 2018. 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 204, + 288, + 257 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 204, + 288, + 257 + ], + "spans": [ + { + "bbox": [ + 53, + 204, + 288, + 257 + ], + "type": "text", + "content": "[3] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In European conference on computer vision, pages 213-229. Springer, 2020. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 259, + 288, + 313 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 259, + 288, + 313 + ], + "spans": [ + { + "bbox": [ + 53, + 259, + 288, + 313 + ], + "type": "text", + "content": "[4] Bowen Cheng, Ishan Misra, Alexander G Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1290–1299, 2022. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 315, + 288, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 315, + 288, + 358 + ], + "spans": [ + { + "bbox": [ + 53, + 315, + 288, + 358 + ], + "type": "text", + "content": "[5] Sharan Chetlur, Cliff Woolley, Philippe Vandermersch, Jonathan Cohen, John Tran, Bryan Catanzaro, and Evan Shelhamer. codnn: Efficient primitives for deep learning. arXiv preprint arXiv:1410.0759, 2014. 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 359, + 288, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 359, + 288, + 402 + ], + "spans": [ + { + "bbox": [ + 53, + 359, + 288, + 402 + ], + "type": "text", + "content": "[6] François Chollet. Xception: Deep learning with depthwise separable convolutions. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1251-1258, 2017. 2, 3, 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 403, + 288, + 446 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 403, + 288, + 446 + ], + "spans": [ + { + "bbox": [ + 53, + 403, + 288, + 446 + ], + "type": "text", + "content": "[7] Jifeng Dai, Haozhi Qi, Yuwen Xiong, Yi Li, Guodong Zhang, Han Hu, and Yichen Wei. Deformable convolutional networks. In Proceedings of the IEEE international conference on computer vision, pages 764-773, 2017. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 448, + 288, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 448, + 288, + 479 + ], + "spans": [ + { + "bbox": [ + 53, + 448, + 288, + 479 + ], + "type": "text", + "content": "[8] Tri Dao. Flashattention-2: Faster attention with better parallelism and work partitioning. arXiv preprint arXiv:2307.08691, 2023.5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 481, + 288, + 524 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 481, + 288, + 524 + ], + "spans": [ + { + "bbox": [ + 53, + 481, + 288, + 524 + ], + "type": "text", + "content": "[9] Tri Dao, Dan Fu, Stefano Ermon, Atri Rudra, and Christopher Ré. Flashattention: Fast and memory-efficient exact attention with io-awareness. Advances in Neural Information Processing Systems, 35:16344-16359, 2022. 2, 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 525, + 287, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 525, + 287, + 568 + ], + "spans": [ + { + "bbox": [ + 48, + 525, + 287, + 568 + ], + "type": "text", + "content": "[10] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248–255. IEEE, 2009. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 570, + 287, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 570, + 287, + 623 + ], + "spans": [ + { + "bbox": [ + 48, + 570, + 287, + 623 + ], + "type": "text", + "content": "[11] Xiaohan Ding, Xiangyu Zhang, Jungong Han, and Guiguang Ding. Scaling up your kernels to 31x31: Revisiting large kernel design in cnns. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11963-11975, 2022. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 625, + 287, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 625, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 625, + 287, + 689 + ], + "type": "text", + "content": "[12] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations, 2020. 1, 2, 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "type": "text", + "content": "[13] Qi Han, Zejia Fan, Qi Dai, Lei Sun, Ming-Ming Cheng, Jiaying Liu, and Jingdong Wang. On the connection between" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 547, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "text", + "content": "local attention and dynamic depth-wise convolution. arXiv preprint arXiv:2106.04263, 2021. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 97, + 547, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 97, + 547, + 140 + ], + "spans": [ + { + "bbox": [ + 308, + 97, + 547, + 140 + ], + "type": "text", + "content": "[14] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 142, + 547, + 175 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 142, + 547, + 175 + ], + "spans": [ + { + "bbox": [ + 308, + 142, + 547, + 175 + ], + "type": "text", + "content": "[15] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In Proceedings of the IEEE international conference on computer vision, 2017. 1, 5, 7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 177, + 547, + 209 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 177, + 547, + 209 + ], + "spans": [ + { + "bbox": [ + 308, + 177, + 547, + 209 + ], + "type": "text", + "content": "[16] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. ImageNet classification with deep convolutional neural networks. Advances in neural information processing systems, 2012. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 211, + 547, + 253 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 211, + 547, + 253 + ], + "spans": [ + { + "bbox": [ + 308, + 211, + 547, + 253 + ], + "type": "text", + "content": "[17] Yann LeCun, Léon Bottou, Yoshua Bengio, and Patrick Haffner. Gradient-based learning applied to document recognition. Proceedings of the IEEE, 86(11):2278-2324, 1998. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 256, + 545, + 310 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 256, + 545, + 310 + ], + "spans": [ + { + "bbox": [ + 308, + 256, + 545, + 310 + ], + "type": "text", + "content": "[18] Youngwan Lee, Joong-won Hwang, Sangrok Lee, Yuseok Bae, and Jongyoul Park. An energy andgpu-computation efficient backbone network for real-time object detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops, pages 0–0, 2019. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 312, + 547, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 312, + 547, + 376 + ], + "spans": [ + { + "bbox": [ + 308, + 312, + 547, + 376 + ], + "type": "text", + "content": "[19] Feng Li, Hao Zhang, Huaizhe Xu, Shilong Liu, Lei Zhang, Lionel M Ni, and Heung-Yeung Shum. Mask dino: Towards a unified transformer-based framework for object detection and segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3041-3050, 2023. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 379, + 547, + 422 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 379, + 547, + 422 + ], + "spans": [ + { + "bbox": [ + 308, + 379, + 547, + 422 + ], + "type": "text", + "content": "[20] Yanghao Li, Hanzi Mao, Ross Girshick, and Kaiming He. Exploring plain vision transformer backbones for object detection. In European Conference on Computer Vision, pages 280-296. Springer, 2022. 5" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 424, + 547, + 478 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 424, + 547, + 478 + ], + "spans": [ + { + "bbox": [ + 308, + 424, + 547, + 478 + ], + "type": "text", + "content": "[21] Zhiqi Li, Wenhai Wang, Hongyang Li, Enze Xie, Chonghao Sima, Tong Lu, Yu Qiao, and Jifeng Dai. Bevformer: Learning bird's-eye-view representation from multi-camera images via spatiotemporal transformers. In European conference on computer vision, pages 1-18. Springer, 2022. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 480, + 547, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 480, + 547, + 544 + ], + "spans": [ + { + "bbox": [ + 308, + 480, + 547, + 544 + ], + "type": "text", + "content": "[22] Zhiqi Li, Wenhai Wang, Enze Xie, Zhiding Yu, Anima Anandkumar, Jose M Alvarez, Ping Luo, and Tong Lu. Panoptic segformer: Delving deeper into panoptic segmentation with transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1280-1289, 2022. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 547, + 547, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 547, + 547, + 612 + ], + "spans": [ + { + "bbox": [ + 308, + 547, + 547, + 612 + ], + "type": "text", + "content": "[23] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 7" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 614, + 547, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 614, + 547, + 657 + ], + "spans": [ + { + "bbox": [ + 308, + 614, + 547, + 657 + ], + "type": "text", + "content": "[24] Yingfei Liu, Tiancai Wang, Xiangyu Zhang, and Jian Sun. Petr: Position embedding transformation for multi-view 3d object detection. In European Conference on Computer Vision, pages 531-548. Springer, 2022. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 659, + 547, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 659, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 659, + 547, + 713 + ], + "type": "text", + "content": "[25] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10012-10022, 2021. 1, 2, 3, 5, 6" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5660" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 714 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 288, + 127 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 288, + 127 + ], + "type": "text", + "content": "[26] Zhuang Liu, Hanzi Mao, Chao-Yuan Wu, Christoph Feichtenhofer, Trevor Darrell, and Saining Xie. A convnet for the 2020s. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11976-11986, 2022. 1, 2, 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 129, + 288, + 163 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 129, + 288, + 163 + ], + "spans": [ + { + "bbox": [ + 48, + 129, + 288, + 163 + ], + "type": "text", + "content": "[27] Ningning Ma, Xiangyu Zhang, Hai-Tao Zheng, and Jian Sun. Shufflenet v2: Practical guidelines for efficient cnn architecture design. In ECCV, 2018. 2, 3, 4" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 164, + 288, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 164, + 288, + 228 + ], + "spans": [ + { + "bbox": [ + 48, + 164, + 288, + 228 + ], + "type": "text", + "content": "[28] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. Advances in neural information processing systems, 32, 2019. 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 231, + 288, + 286 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 231, + 288, + 286 + ], + "spans": [ + { + "bbox": [ + 48, + 231, + 288, + 286 + ], + "type": "text", + "content": "[29] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 1, 2, 8" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 287, + 288, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 287, + 288, + 353 + ], + "spans": [ + { + "bbox": [ + 48, + 287, + 288, + 353 + ], + "type": "text", + "content": "[30] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical Image Computing and Computer-Assisted Intervention-MICCAI 2015: 18th International Conference, Munich, Germany, October 5-9, 2015, Proceedings, Part III 18, pages 234-241. Springer, 2015. 2, 8" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 354, + 288, + 420 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 354, + 288, + 420 + ], + "spans": [ + { + "bbox": [ + 48, + 354, + 288, + 420 + ], + "type": "text", + "content": "[31] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in Neural Information Processing Systems, 35:36479-36494, 2022. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 421, + 288, + 454 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 421, + 288, + 454 + ], + "spans": [ + { + "bbox": [ + 48, + 421, + 288, + 454 + ], + "type": "text", + "content": "[32] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556, 2014. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 456, + 288, + 500 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 456, + 288, + 500 + ], + "spans": [ + { + "bbox": [ + 48, + 456, + 288, + 500 + ], + "type": "text", + "content": "[33] Robin Strudel, Ricardo Garcia, Ivan Laptev, and Cordelia Schmid. Segmenter: Transformer for semantic segmentation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 7262-7272, 2021. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 501, + 288, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 501, + 288, + 555 + ], + "spans": [ + { + "bbox": [ + 48, + 501, + 288, + 555 + ], + "type": "text", + "content": "[34] Hugues Thomas, Charles R Qi, Jean-Emmanuel Deschaud, Beatz Marcotegui, François Goulette, and Leonidas J Guibas. Kpconv: Flexible and deformable convolution for point clouds. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6411–6420, 2019. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 557, + 288, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 557, + 288, + 601 + ], + "spans": [ + { + "bbox": [ + 48, + 557, + 288, + 601 + ], + "type": "text", + "content": "[35] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 2, 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 602, + 288, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 602, + 288, + 658 + ], + "spans": [ + { + "bbox": [ + 48, + 602, + 288, + 658 + ], + "type": "text", + "content": "[36] Ashish Vaswani, Prajit Ramachandran, Aravind Srinivas, Niki Parmar, Blake Hechtman, and Jonathon Shlens. Scaling local self-attention for parameter efficient visual backbones. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12894–12904, 2021. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 658, + 288, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 288, + 714 + ], + "type": "text", + "content": "[37] Shenlong Wang, Simon Suo, Wei-Chiu Ma, Andrei Pokrovsky, and Raquel Urtasun. Deep parametric continuous convolutional neural networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2589-2597, 2018. 2" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 572 + ], + "type": "list", + "angle": 0, + "index": 24, + "blocks": [ + { + "bbox": [ + 307, + 73, + 547, + 126 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 547, + 126 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 547, + 126 + ], + "type": "text", + "content": "[38] Wenhai Wang, Jifeng Dai, Zhe Chen, Zhenhang Huang, Zhiqi Li, Xizhou Zhu, Xiaowei Hu, Tong Lu, Lewei Lu, Hongsheng Li, et al. Internimage: Exploring large-scale vision foundation models with deformable convolutions. In CVPR, 2023. 1, 2, 6" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 307, + 129, + 547, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 129, + 547, + 160 + ], + "spans": [ + { + "bbox": [ + 307, + 129, + 547, + 160 + ], + "type": "text", + "content": "[39] Ross Wightman. Pytorch image models. https://github.com/rwrightman/pytorch-image-models, 2019.6" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 162, + 545, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 162, + 545, + 195 + ], + "spans": [ + { + "bbox": [ + 307, + 162, + 545, + 195 + ], + "type": "text", + "content": "[40] Felix Wu, Angela Fan, Alexei Baevski, Yann N Dauphin, and Michael Auli. Pay less attention with lightweight and dynamic convolutions. arXiv preprint arXiv:1901.10430, 2019. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 196, + 545, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 196, + 545, + 239 + ], + "spans": [ + { + "bbox": [ + 307, + 196, + 545, + 239 + ], + "type": "text", + "content": "[41] Tete Xiao, Yingcheng Liu, Bolei Zhou, Yuning Jiang, and Jian Sun. Unified perceptual parsing for scene understanding. In Proceedings of the European conference on computer vision (ECCV), pages 418-434, 2018. 1, 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 240, + 545, + 272 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 240, + 545, + 272 + ], + "spans": [ + { + "bbox": [ + 307, + 240, + 545, + 272 + ], + "type": "text", + "content": "[42] Yuwen Xiong, Mengye Ren, Renjie Liao, Kelvin Wong, and Raquel Urtasun. Deformable filter convolution for point cloud reasoning. arXiv preprint arXiv:1907.13079, 2019. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 274, + 547, + 350 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 274, + 547, + 350 + ], + "spans": [ + { + "bbox": [ + 308, + 274, + 547, + 350 + ], + "type": "text", + "content": "[43] Chenyu Yang, Yuntao Chen, Hao Tian, Chenxin Tao, Xizhou Zhu, Zhaoxiang Zhang, Gao Huang, Hongyang Li, Yu Qiao, Lewei Lu, et al. Bevformer v2: Adapting modern image backbones to bird's-eye-view recognition via perspective supervision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17830-17839, 2023. 2, 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 351, + 547, + 394 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 351, + 547, + 394 + ], + "spans": [ + { + "bbox": [ + 308, + 351, + 547, + 394 + ], + "type": "text", + "content": "[44] Xiaohua Zhai, Alexander Kolesnikov, Neil Houlsby, and Lucas Beyer. Scaling vision transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12104-12113, 2022. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 395, + 545, + 438 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 395, + 545, + 438 + ], + "spans": [ + { + "bbox": [ + 308, + 395, + 545, + 438 + ], + "type": "text", + "content": "[45] Hao Zhang, Feng Li, Shilong Liu, Lei Zhang, Hang Su, Jun Zhu, Lionel M Ni, and Heung-Yeung Shum. Dino: Detr with improved denoising anchor boxes for end-to-end object detection. arXiv preprint arXiv:2203.03605, 2022. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 439, + 545, + 483 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 439, + 545, + 483 + ], + "spans": [ + { + "bbox": [ + 308, + 439, + 545, + 483 + ], + "type": "text", + "content": "[46] Bolei Zhou, Hang Zhao, Xavier Puig, Sanja Fidler, Adela Barriuso, and Antonio Torralba. Scene parsing through ade20k dataset. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 633-641, 2017. 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 483, + 545, + 528 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 483, + 545, + 528 + ], + "spans": [ + { + "bbox": [ + 308, + 483, + 545, + 528 + ], + "type": "text", + "content": "[47] Xizhou Zhu, Han Hu, Stephen Lin, and Jifeng Dai. Deformable convnets v2: More deformable, better results. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 528, + 545, + 572 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 528, + 545, + 572 + ], + "spans": [ + { + "bbox": [ + 308, + 528, + 545, + 572 + ], + "type": "text", + "content": "[48] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable detr: Deformable transformers for end-to-end object detection. In International Conference on Learning Representations, 2020. 2, 5" + } + ] + } + ], + "index": 23 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "5661" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Efficient Detection of Long Consistent Cycles and its Application to Distributed Synchronization/c0dd6b15-09a8-4eba-bcfb-24b8083dd375_content_list.json b/2024/Efficient Detection of Long Consistent Cycles and its Application to Distributed Synchronization/c0dd6b15-09a8-4eba-bcfb-24b8083dd375_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..bbc900ec00017ef2b51ea5d5e864df04bdc92667 --- /dev/null +++ b/2024/Efficient Detection of Long Consistent Cycles and its Application to Distributed Synchronization/c0dd6b15-09a8-4eba-bcfb-24b8083dd375_content_list.json @@ -0,0 +1,1834 @@ +[ + { + "type": "text", + "text": "Efficient Detection of Long Consistent Cycles and its Application to Distributed Synchronization *", + "text_level": 1, + "bbox": [ + 240, + 130, + 736, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Shaohan Li \nUniversity of Minnesota \nli000743@umn.edu", + "bbox": [ + 140, + 203, + 326, + 256 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yunpeng Shi \nUniversity of California, Davis \nypshi@ucdavis.edu", + "bbox": [ + 367, + 203, + 602, + 257 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Gilad Lerman \nUniversity of Minnesota \nlerman@umn.edu", + "bbox": [ + 643, + 203, + 830, + 256 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 236, + 291, + 310, + 306 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Group synchronization plays a crucial role in global pipelines for Structure from Motion (SfM). Its formulation is nonconvex and it is faced with highly corrupted measurements. Cycle consistency has been effective in addressing these challenges. However, computationally efficient solutions are needed for cycles longer than three, especially in practical scenarios where 3-cycles are unavailable. To overcome this computational bottleneck, we propose an algorithm for group synchronization that leverages information from cycles of lengths ranging from three to six with a time complexity of order $O(n^{3})$ (or $O(n^{2.373})$ when using a faster matrix multiplication algorithm). We establish non-trivial theory for this and related methods that achieves competitive sample complexity, assuming the uniform corruption model. To advocate the practical need for our method, we consider distributed group synchronization, which requires at least 4-cycles, and we illustrate state-of-the-art performance by our method in this context.", + "bbox": [ + 75, + 323, + 473, + 580 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 611, + 202, + 627 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Structure from Motion (SfM) asks to recover the 3D structure of a stationary scene from multiple images taken by cameras from different orientations and locations. In the past decade, the global SfM pipeline has become increasingly popular due to its several advantages over the incremental pipelines [17, 31]. First of all, global SfM requires only one implementation of bundle adjustment, making it more efficient in computation. Second, it estimates camera poses in a global optimization framework which mitigates the drifting issue of the incremental pipelines. Despite the popularity of global SfM pipelines, the estimation of global camera poses (e.g., orientations) remains a highly challenging problem. For instance, estimating camera orientations from their relative measurements, often called rotation synchronization, is a highly nonconvex graph optimization problem. In typical scenarios of highly noisy or corrupted measurements", + "bbox": [ + 75, + 636, + 470, + 864 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "of relative orientations, many common solutions of rotation synchronization have poor accuracy and slow convergence.", + "bbox": [ + 500, + 292, + 890, + 323 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Given these challenges, theoretical developments have demonstrated the critical role of cycle-consistency information in inferring corrupted measurements [25]. In practice, the consistency constraint on 3-cycles was utilized to estimate the error of each measured relative orientation. It also helped nonconvex iterative rotation synchronization solvers avoid spurious local minima and achieve significantly higher accuracy [39]. However, the usage of 3-cycles largely limits the application of these improved algorithms to other important scenarios. One scenario involves a sparse viewing graph lacking sufficient 3-cycles. This often occurs when the size of the graph is too large to densely measure the relative orientations on its edges, which could happen in certain case for the molecular orientation estimation in cryo-electron microscopy imaging. Another scenario is orientation estimation for each piece of jigsaw puzzles, where the graph is a 2D lattice and 3-cycle does not exist. Lastly, in distributed SfM, edges between any two clusters of nodes form a bipartite graph, and cycles of odd length do not exist. Our numerical results primarily emphasize the practical scenario of distributed SfM, which holds particular relevance for the broader computer vision community.", + "bbox": [ + 496, + 324, + 893, + 641 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Despite the multiple critical applications of long-cycle consistency, inferring measurement noise from long cycles is challenging in both computation and theory. First of all, the number of cycles grows exponentially with the cycle length, and measuring cycle inconsistencies for each long cycle is computationally intractable. Moreover, developing theoretical guarantees for long-cycle inference methods is fundamentally more difficult than the 3-cycle case. Indeed, in a random graph setting, a set of longer cycles are more likely to share common edges, making their consistency score highly correlated. Therefore, new tools are required to handle the correlated empirical process.", + "bbox": [ + 496, + 642, + 895, + 808 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this work, we propose the first practical method, LongSync, for inferring edge corruption levels from long cycle consistency information. For this purpose, we carefully modify and vectorize the Cycle Edge Message Passing (CEMP) method [25]. This nontrivial modification drastically reduces its computational complexity when using longer cycles. Moreover,", + "bbox": [ + 496, + 810, + 895, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*This work was supported by NSF award DMS-2152766. Supplementary code: https://github.com/sli743/LongSync", + "bbox": [ + 75, + 875, + 387, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "5260", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "by employing a more delicate analysis and incorporating new tools from probability theory and combinatorics, we develop a significantly stronger exact recovery result with a general cycle length under a popular probabilistic model. The sample complexity in our theory is the lowest among all practical rotation synchronization methods. Although we limit our scope to the application of distributed SfM, our algorithm and theory applies to any finite-dimensional linear group in group synchronization, and not just $SO(3)$ in rotation synchronization.", + "bbox": [ + 75, + 90, + 472, + 227 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "1.1. Related Work", + "text_level": 1, + "bbox": [ + 76, + 239, + 220, + 253 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Earlier rotation synchronization methods [2, 10-13, 18, 33, 42, 43] seek to minimize a least squares energy function. They can be described as relaxed versions of the maximum likelihood estimator under the additive Gaussian noise model, but they are not robust in the presence of outliers or heavy-tailed noise. Nevertheless, in the case of global SfM, the initially estimated relative camera rotations can be severely corrupted due to the erroneous keypoint matches and the subsequent poor estimation of fundamental matrices.", + "bbox": [ + 75, + 263, + 470, + 398 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To handle outliers, robust rotation synchronization methods either minimize a robust energy function or reweigh/trim the viewing graph based on the corruption levels of the edges. Wang and Singer [49] minimizes a corresponding $\\ell_1$ objective function using semidefinite programming (SDP) relaxation, which is slow in practice. Other energy minimization methods are typically nonconvex, which include the Weiszfeld algorithm [20] and the Riemannian subgradient method [29] for $\\ell_1$ minimization, and the iteratively reweighted least squares (IRLS) for minimizing general $\\ell_p$ [5] and Geman-McClure [4] loss functions. However, all these methods heavily rely on good initialization. Sidhartha and Govindu [41] partially remedy the issue using adaptive Geman-McClure loss functions, but their approach remains sensitive to the initialized weights. Maunu and Lerman [30] propose to solve rotation synchronization by an iterative robust averaging method that utilizes Tukey depth, but they have not demonstrated effective performance for real SfM applications. Arrigoni et al. [1] applies a low-rank and sparse matrix decomposition method to SO(3) and SE(3) synchronization, but it is even less robust to outliers than IRLS.", + "bbox": [ + 75, + 401, + 470, + 703 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Instead of employing a robust objective function, Shen et al. [36] and Zach et al. [51] uses the 3-cycle consistency constraint to detect and remove corrupted relative orientations. Lerman and Shi [25] take one step further to estimate the corruption level of each relative measurement by a novel cycle-edge message passing (CEMP) algorithm. They then use the estimated corruption levels to reweigh the graph and solve rotation synchronization using a weighted least squares method. This message passing procedure was further combined with IRLS to boost its accuracy in [39]. Particular versions and extensions of this procedure for permutation and partial permutation synchronization, which are relevant to the matching component of SfM, were discussed in [27, 40].", + "bbox": [ + 75, + 704, + 470, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "However, all the previously mentioned cycle-based methods [25, 27, 36, 38-40, 51] only use 3-cycles in practice, limiting their application for distributed synchronization. Indeed, the standard distributed synchronization often requires \"stitching\" local solutions by synchronizing the relative rotations between clusters. Each of these inter-cluster rotations is estimated by \"averaging\" the edges between the two clusters. These edges form a bipartite graph, and the minimal cycle length is 4. As pointed in [51], the number of operations for computing long cycle consistency information scales exponentially with the cycle length. Therefore, none of the existing distributed rotation synchronization methods directly exploits long cycle information due to this computational challenge.", + "bbox": [ + 496, + 90, + 893, + 287 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The earlier distributed methods for $SO(d)$ synchronization, such as [45] and [44], minimize a least squares energy and are not robust to outliers. A series of distributed SfM methods [14-16] implement incremental SfM algorithms for each cluster. However, these methods do not employ a standard rotation synchronization algorithm, as they require additional information such as the number of keypoint matches between images. Moreover, the incremental methods are slower since they require multiple rounds of global rotation synchronization. MultiSync [9] synchronizes the inter-cluster rotations directly using all inter-cluster edges among all clusters, by formulating a novel synchronization problem on a multi-graph. Although it utilizes a more unified formulation, its objective function is least squares which largely limits its robustness to outliers.", + "bbox": [ + 496, + 287, + 893, + 500 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "A recent and different type of methods for rotation synchronization use deep learning [21, 26, 34]. However, these methods are supervised and thus may not generalize well when switching datasets. Moreover, like many other previous methods, they lack theoretical guarantees.", + "bbox": [ + 496, + 501, + 890, + 575 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "A common theoretical setting to assess the performance of rotation synchronization algorithms is the uniform corruption model (UCM) described in §4. We provide the best sample complexity for LongSync, even with only 3 cycles, among all previously established estimates for the UCM model.", + "bbox": [ + 496, + 578, + 890, + 652 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "1.2. Contributions of This Work", + "text_level": 1, + "bbox": [ + 500, + 665, + 750, + 680 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose the first practical algorithm that infers edge corruption levels from long cycle consistency information. The computation complexity of our method is reduced from $O(n^{c})$ to $O(n^{3})$ (or possibly $O(n^{2.373})$ ) for cycle length $c \\leq 6$ and $O(n^{[(c + 3) / 2]})$ for $c > 6$ .", + "- We establish sample complexity estimates for our method under the uniform corruption model, where we get closer to the information theoretic bound than any other existing work. Our proof requires delicate analysis and it also improves previous estimates for the CEMP algorithm.", + "- We introduce a new graph partition and graph preprocessing method that utilizes our inference method, and demonstrate the effectiveness of our pipeline in boosting the performance of distributed synchronization." + ], + "bbox": [ + 500, + 688, + 893, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "5261", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "- Extensive numerical experiments demonstrate the outstanding performance of our method.", + "bbox": [ + 76, + 90, + 468, + 121 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2. Problem Formulation and Preliminaries", + "text_level": 1, + "bbox": [ + 76, + 133, + 419, + 148 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Assume a graph $G = ([n], E)$ where $[n]$ is the set of nodes indexed by $\\{1, 2, \\dots, n\\}$ and $E$ is the set of edges. Given a mathematical group $\\mathcal{G}$ , each graph node is assigned an underlying ground truth group element $R_{i}^{*}$ , where $R_{i}^{*} \\in \\mathcal{G}$ and we use star superscript to emphasize the ground truth. For each edge $ij \\in E$ , we observe a relative group ratio $R_{ij} \\in \\mathcal{G}$ , whose clean counterpart is $R_{ij}^{*} = R_{i}^{*}R_{j}^{*-1}$ . Group synchronization aims to recover the ground truth group elements $\\{R_{i}^{*}\\}_{i \\in [n]}$ from the possibly noisy and corrupted measurements $\\{R_{ij}\\}_{ij \\in E}$ . In this paper, we focus on the case of rotation synchronization, which is a special case of group synchronization with $\\mathcal{G} = \\mathrm{SO}(d)$ . For applications in camera orientation synchronization ( $d = 3$ ), we estimate absolute rotations for each node $i \\in [n]$ from measured relative rotations of edges in $E$ . Note that since $\\{R_{i}^{*}\\}_{i \\in [n]}$ and $\\{R_{i}^{*}R_{0}\\}_{i \\in [n]}$ generate the same set of relative rotations, one can only estimate $\\{R_{i}^{*}\\}_{i \\in [n]}$ up to a global rotation. The generalization to any linear groups is discussed in the supplementary material.", + "bbox": [ + 76, + 159, + 472, + 431 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.1. Notations and Definitions", + "text_level": 1, + "bbox": [ + 76, + 439, + 308, + 454 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We denote the adjacency matrix of graph $G$ as $\\mathbf{A}$ , and form a pairwise observation matrix $\\mathbf{R} \\in \\mathbb{R}^{dn \\times dn}$ by stacking the $\\mathbf{R}_{ij}$ 's (for $ij \\notin E$ , set $\\mathbf{R}_{ij} = \\mathbf{0}_{3 \\times 3}$ ):", + "bbox": [ + 75, + 462, + 468, + 508 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {R} := \\left( \\begin{array}{c c c c} \\boldsymbol {R} _ {1 1} & \\boldsymbol {R} _ {1 2} & \\dots & \\boldsymbol {R} _ {1 n} \\\\ \\boldsymbol {R} _ {2 1} & \\boldsymbol {R} _ {2 2} & \\dots & \\boldsymbol {R} _ {2 n} \\\\ \\vdots & \\vdots & \\ddots & \\vdots \\\\ \\boldsymbol {R} _ {n 1} & \\boldsymbol {R} _ {n 2} & \\dots & \\boldsymbol {R} _ {n n} \\end{array} \\right).\n$$\n", + "text_format": "latex", + "bbox": [ + 161, + 517, + 385, + 589 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We list the matrix operations used in the paper. For matrices $X$ and $Y$ , the operations $X \\otimes Y$ , $X \\odot Y$ , $X \\oslash Y$ respectively denote the Kronecker product, Hardmard (element-wise) multiplication and Harnard division between $X$ and $Y$ . $X^{\\odot k}$ denotes the element-wise matrix $k$ -power. For block matrices, $\\langle X, Y \\rangle_{\\mathrm{block}}$ denotes the blockwise inner product of $X$ and $Y$ , i.e. $\\langle X, Y \\rangle_{\\mathrm{block}}(i,j) = \\langle X[i,j], Y[i,j] \\rangle$ , where $[i,j]$ refers to the corresponding block of the matrix.", + "bbox": [ + 76, + 597, + 470, + 717 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. Review of CEMP for $c$ -Cycles", + "text_level": 1, + "bbox": [ + 76, + 726, + 341, + 742 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We assume the above setting of $SO(d)$ synchronization. Let $\\mathcal{D}$ be any bi-invariant metric on $SO(d)$ . We assume a fixed number of cycles, $c$ , and denote by $N_{ij}^{c}$ the set of simple cycles of length $c$ (or simple $c$ -cycles) containing edge $ij$ . CEMP [25] aims to estimate for each edge $ij$ the corruption level", + "bbox": [ + 76, + 750, + 468, + 825 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\ns _ {i j} ^ {*} = \\mathcal {D} \\left(\\boldsymbol {R} _ {i j}, \\boldsymbol {R} _ {i j} ^ {*}\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 210, + 834, + 468, + 853 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "from the set of cycle inconsistency measures", + "bbox": [ + 76, + 859, + 357, + 875 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nd _ {L} = \\mathcal {D} \\left(\\boldsymbol {R} _ {L}, \\boldsymbol {R} _ {i j}\\right) \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 214, + 885, + 468, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where cycle $L = (ik_{1}, k_{1}k_{2}, \\dots, k_{c-2}j, ji) \\in N_{ij}^{c}$ and $R_{L} := R_{ik_{1}}R_{k_{1}k_{2}}\\dots R_{k_{c-2}j}$ . The estimated $s_{ij}^{*}$ can then be used for extracting a clean subgraph, or to implement a weighted least squares solver where higher weights are assigned to cleaner edges.", + "bbox": [ + 498, + 90, + 890, + 166 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "It is obvious that if all the edges in $L$ are clean then $d_{L} = 0$ . Moreover, due to bi-invariance of $\\mathcal{D}$ , the following holds true", + "bbox": [ + 498, + 167, + 893, + 196 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nd _ {L} = s _ {i j} ^ {*} \\text {w h e n e v e r} L \\in G _ {i j} ^ {c}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 606, + 205, + 890, + 224 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $G_{ij}^{c}$ is the set of good $c$ -cycles with respect to $ij$ , i.e. the set of cycles $L \\in N_{ij}^{c}$ such that $ik_1, \\dots, k_{c-2j}$ are clean. This gives a sufficient condition for $d_L$ to be an exact estimator of $s_{ij}^{*}$ .", + "bbox": [ + 496, + 231, + 893, + 276 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To estimate the corruption levels of each edge $ij$ , CEMP initializes the edge weight of each $ij \\in E$ as $w_{ij}^{(0)} = 1$ . It then iteratively updates the corruption level estimate as the following convex combination of $d_L$ 's:", + "bbox": [ + 496, + 277, + 890, + 339 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\ns _ {i j} ^ {(t)} = \\sum_ {L \\in N _ {i j} ^ {c}} w _ {L} ^ {(t)} d _ {L} / z _ {i j} ^ {(t)} \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 616, + 345, + 890, + 381 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $z_{ij}^{(t)} = \\sum_{L\\in N_{ij}^c}w_L^{(t)}$ . The cycle weights $w_{L}^{(t + 1)}$ are computed from the edge weights $w_{e}^{(t + 1)} = e^{-\\beta_{t}s_{e}^{(t)}}$ :", + "bbox": [ + 496, + 390, + 890, + 431 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nw _ {L} ^ {(t + 1)} = \\prod_ {e \\in L \\backslash \\{i j \\}} w _ {e} ^ {(t + 1)} = \\prod_ {e \\in L \\backslash \\{i j \\}} e ^ {- \\beta_ {t} s _ {e} ^ {(t)}}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 555, + 439, + 890, + 474 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "so that $w_{L}^{(t + 1)}$ focuses on good cycles. The cycle weights and edge corruption levels are alternatingly updated and improved from each other. Interestingly, it is proved in [25] under two different corruption models that CEMP converges linearly to the ground truth corruption estimates under mild conditions for $c = 3$ . In practice, CEMP only uses 3-cycles for consideration of efficiency. For longer cycles, the complexity of CEMP scales exponentially with the cycle length $c$ (which is discussed in §3.2), and the convergence guarantee of CEMP remains unknown.", + "bbox": [ + 496, + 484, + 893, + 623 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Our method: LongSync", + "text_level": 1, + "bbox": [ + 498, + 635, + 714, + 652 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. LongSync: Modification of CEMP", + "text_level": 1, + "bbox": [ + 498, + 660, + 800, + 676 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our goal is to develop a scalable variant of CEMP for any fixed number of cycles, $c \\geq 3$ . The main computational bottleneck of step (4) in CEMP is that computing and summing the cycle inconsistency measures takes $\\sum_{ij \\in E} |N_{ij}^c| = O(n^c)$ operations and memory. Therefore, to develop a scalable algorithm, we aim to take weighted average over $d_L$ without explicitly computing and storing each $d_L$ . To achieve this, we propose the following specification and modification on CEMP:", + "bbox": [ + 496, + 683, + 890, + 803 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "- Use Chordal distance on SO(d). We suggest the distance function", + "bbox": [ + 500, + 804, + 890, + 832 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {D} \\left(\\boldsymbol {R} _ {1}, \\boldsymbol {R} _ {2}\\right) = \\sqrt {1 - \\left\\langle \\boldsymbol {R} _ {1} , \\boldsymbol {R} _ {2} \\right\\rangle / d}.\n$$\n", + "text_format": "latex", + "bbox": [ + 598, + 832, + 805, + 849 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "This distance is proportional to the Chordal distance on $SO(3)$ , which is the Euclidean distance between two rotations embedded in $\\mathbb{R}^{d\\times d}$ .", + "bbox": [ + 511, + 854, + 890, + 898 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "5262", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "- Use weighted quadratic average for corruption level update. Instead of updating the corruption level estimates by a weighted average of $d_{L}$ , we use the weighted quadratic average of $d_{L}$ , namely", + "bbox": [ + 76, + 90, + 468, + 151 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\ns _ {i j} ^ {(t)} = \\sqrt {\\sum_ {L \\in N _ {i j} ^ {c}} w _ {L} ^ {(t)} d _ {L} ^ {2} / z _ {i j} ^ {(t)}} \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 192, + 162, + 468, + 202 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where the update rule of cycle weights remains the same:", + "bbox": [ + 89, + 205, + 447, + 220 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nw _ {L} ^ {(t + 1)} = \\prod_ {e \\in L \\backslash \\{i j \\}} w _ {e} ^ {(t + 1)} = \\prod_ {e \\in L \\backslash \\{i j \\}} e ^ {- \\beta_ {t} s _ {e} ^ {(t)}}. \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 138, + 227, + 468, + 263 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As a result, $d_{L}^{2} = d^{2}(\\pmb{R}_{L},\\pmb{R}_{ij}) = 1 - \\langle \\pmb{R}_{L},\\pmb{R}_{ij}\\rangle /d$ is linear in both $R_{L}$ and $R_{ij}$ . Therefore one can switch the order of $d^{2}$ and the weighted summation, so that the computation of $s_{ij}^{(t)}$ can be vectorized. Indeed, by this linearity and equations (4) and (2), and note that $z_{ij}^{(t)} = \\sum_{L\\in N_{ij}^{c}}w_{L}^{(t)}$ , we obtain the following equation:", + "bbox": [ + 76, + 271, + 468, + 372 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} s _ {i j} ^ {(t)} = \\left(\\sum_ {L \\in N _ {i j} ^ {c}} w _ {L} ^ {(t)} d _ {L} ^ {2} / z _ {i j} ^ {(t)}\\right) ^ {1 / 2} \\\\ = \\Big (\\sum_ {L \\in N _ {i j} ^ {c}} w _ {L} ^ {(t)} \\mathcal {D} ^ {2} (\\boldsymbol {R} _ {L}, \\boldsymbol {R} _ {i j}) / z _ {i j} ^ {(t)} \\Big) ^ {1 / 2} \\\\ = \\left(\\mathcal {D} ^ {2} \\Big (\\sum_ {L \\in N _ {i j} ^ {c}} w _ {L} ^ {(t)} \\boldsymbol {R} _ {L}, \\boldsymbol {R} _ {i j} \\Big) / z _ {i j} ^ {(t)}\\right) ^ {1 / 2} \\\\ = \\left(1 - \\left\\langle \\sum_ {L \\in N _ {i j} ^ {c}} w _ {L} ^ {(t)} \\boldsymbol {R} _ {L}, \\boldsymbol {R} _ {i j} \\right\\rangle / d \\sum_ {L \\in N _ {i j} ^ {c}} w _ {L} ^ {(t)}\\right) ^ {1 / 2} \\tag {8} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 112, + 378, + 468, + 550 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Equation (8) can be vectorized using the trick of matrix power if we allow repeated nodes for each cycle. That is, one can stack the $s_{ij}^{(t)}$ 's and $w_{ij}^{(t)}$ 's into matrices $\\mathbf{S}^{(t)}$ and $\\mathbf{W}^{(t)}$ , and vectorize (8) as", + "bbox": [ + 76, + 556, + 468, + 619 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\boldsymbol {S} ^ {(t)} = \\left(\\boldsymbol {A} - \\left\\langle \\left(\\boldsymbol {W} ^ {(t)} \\otimes 1 _ {d} \\odot \\boldsymbol {R}\\right) ^ {c - 1}, \\boldsymbol {R} \\right\\rangle_ {\\text {b l o c k}} \\right. \\\\ \\left. \\odot d \\left(\\boldsymbol {W} ^ {(t)}\\right) ^ {c - 1}\\right) ^ {\\odot 1 / 2}. \\tag {9} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 130, + 638, + 468, + 699 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Indeed, by (7) and the definition of $\\pmb{R}_L$ , $\\sum_{L \\in C_{ij}^c} w_L^{(t)} \\pmb{R}_L$ is the $ij$ -th block of $(\\pmb{W}^{(t)} \\otimes 1_d \\odot \\pmb{R})^{c-1}$ , and $\\sum_{L \\in C_{ij}^c} w_L^{(t)}$ is the $ij$ -th element of $\\pmb{W}^{(t)c-1}$ , where $C_{ij}^c$ is the set of $c$ -cycles containing $ij$ with possibly repeated nodes.", + "bbox": [ + 76, + 708, + 468, + 782 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In the case of utilizing only simple cycles, (8) and (9) are not equivalent and we need to correct (9) to remove the cycles with repeated nodes, so that only simple cycles in $N_{ij}^{c}$ remain. Let $g_{c}(\\boldsymbol{W},\\boldsymbol{R})$ be the matrix valued function where $g_{c}(\\boldsymbol{W},\\boldsymbol{R})(i,j) = \\sum_{L\\in N_{ij}^{c}}(\\prod_{e\\in L\\setminus \\{ij\\}}w_{e}^{(t)})\\boldsymbol{R}_{L}$ . Let $f_{c}(\\boldsymbol{W})$ be the matrix valued function where $f_{c}(\\boldsymbol{W})(i,j) = \\sum_{L\\in N_{ij}^{c}}(\\prod_{e\\in L\\setminus \\{ij\\}}w_{e}^{(t)})$ . The following result holds:", + "bbox": [ + 76, + 782, + 468, + 904 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Proposition 3.1. The update rule of LongSync (8) is equivalent to the following matrix equations:", + "bbox": [ + 498, + 90, + 890, + 121 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {S} ^ {(t)} = \\left(\\boldsymbol {A} - \\left\\langle h _ {c} (\\boldsymbol {W} ^ {(t)}, \\boldsymbol {R}), \\boldsymbol {R} \\right\\rangle_ {\\text {b l o c k}} \\odot \\boldsymbol {A}\\right) ^ {\\odot 1 / 2}, \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 535, + 131, + 890, + 162 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\pmb{W}^{(t + 1)} = \\pmb{A}\\odot \\exp (-\\beta_t\\pmb{S}^{(t)})$ and", + "bbox": [ + 500, + 172, + 758, + 189 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nh _ {c} \\left(\\boldsymbol {W} ^ {(t)}, \\boldsymbol {R}\\right) := g _ {c} \\left(\\boldsymbol {W} ^ {(t)}, \\boldsymbol {R}\\right) \\oslash \\left(d \\cdot f _ {c} \\left(\\boldsymbol {W} ^ {(t)} \\otimes 1 _ {d}\\right)\\right).\n$$\n", + "text_format": "latex", + "bbox": [ + 535, + 198, + 852, + 217 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Here exp denotes the elementwise exponential function.", + "bbox": [ + 517, + 228, + 866, + 244 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We use equation (10) as the update rule of LongSync and propose the vectorized LongSync algorithm in algorithm 1.", + "bbox": [ + 500, + 253, + 890, + 284 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Algorithm 1 (LongSync)", + "text_level": 1, + "bbox": [ + 501, + 297, + 663, + 314 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Input: pairwise rotation matrix $\\mathbf{R} \\in \\mathbb{R}^{dn \\times dn}$ , adjacency matrix $\\mathbf{A} \\in [0,1]^{n \\times n}$ , cycle length $c$ , positive parameters $\\{\\beta_t\\}_{t \\geq 1}$ , time step $T$", + "bbox": [ + 500, + 319, + 890, + 364 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {W} ^ {(0)} (i, j) \\leftarrow \\boldsymbol {A}\n$$\n", + "text_format": "latex", + "bbox": [ + 517, + 364, + 617, + 378 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "for $t = 0:T$ do", + "bbox": [ + 517, + 381, + 612, + 393 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {S} ^ {(t)} \\leftarrow \\left(\\boldsymbol {A} - \\left\\langle h _ {c} (\\boldsymbol {W} ^ {(t)}, \\boldsymbol {R}), \\boldsymbol {R} \\right\\rangle_ {\\text {b l o c k}} \\odot \\boldsymbol {A}\\right) ^ {\\odot 1 / 2} \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 542, + 402, + 890, + 433 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {W} ^ {(t + 1)} \\leftarrow \\boldsymbol {A} \\odot \\exp (- \\beta_ {t} \\boldsymbol {S} ^ {(t)}) \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 544, + 435, + 890, + 453 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "end for", + "text_level": 1, + "bbox": [ + 517, + 465, + 568, + 478 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Output: edge weights $\\mathbf{W}^{(T + 1)}$ , corruption levels $\\mathbf{S}^{(T)}$", + "bbox": [ + 500, + 479, + 849, + 494 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We claim that $g_{c}$ and $f_{c}$ can be computed with a sequence of matrix operations, thus greatly reducing the time and space consumption of LongSync compared to its original form. For $c \\leq 6$ , the time complexity of computing $g_{c}$ and $f_{c}$ is $O(r(dn))$ , where $r(n)$ is the complexity for multiplying two $n \\times n$ matrices; for $c \\geq 7$ the time complexity is at most $O(n^{[(c + 3) / 2]})$ . This claim is proved in the supplementary material. We list the formula for $g_{c}$ and $f_{c}$ for $c = 3,4,5,6$ inspired by [35, 46]. The formula for $c = 6$ is moved to the supplementary material due to the space limit. For $c \\geq 7$ the formula becomes extremely complicated. We remark that in practice, the cycles of length greater than 6 are often not used.", + "bbox": [ + 496, + 516, + 890, + 696 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We finally remark that although Algorithm 1 only utilizes cycles of a fixed length, one can easily generalize it to incorporate cycles of different lengths. Indeed, the equation (10) could use a convex combination of $h_c$ 's that corresponds to different values of $c$ . That is, for a preselected set of cycle lengths $C$ , the equation (11) in Algorithm 1 is replaced by", + "bbox": [ + 496, + 696, + 890, + 787 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {S} ^ {(t)} = \\left(\\boldsymbol {A} - \\left\\langle \\sum_ {c \\in C} \\lambda_ {c} h _ {c} \\left(\\boldsymbol {W} ^ {(t)}, \\boldsymbol {R}\\right), \\boldsymbol {R} \\right\\rangle_ {\\text {b l o c k}} \\odot \\boldsymbol {A}\\right) ^ {\\odot 1 / 2} \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 509, + 797, + 890, + 844 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where the coefficients $\\lambda_{c}$ satisfies $\\sum_{c\\in C}\\lambda_c = 1$ to ensure a convex combination. Here each $\\lambda_{c}$ is user-specified to reflect the importance of the cycles of length $c$ . However, the optimal", + "bbox": [ + 496, + 854, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "5263", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "choice of these parameters under certain statistical model remains unclear.", + "bbox": [ + 75, + 90, + 468, + 119 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For simplicity, in the experiments we only use a fixed cycle length to avoid choosing $\\lambda_{c}$ . We have observed that such simple choice still yields satisfying accuracy in camera orientation estimation on both synthetic and real data. We refer the readers to §5 and 6 for more details.", + "bbox": [ + 75, + 121, + 467, + 195 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/c998bbdd513673af74c82589d6d8e30a5ee5a35096a9e1f2bf87ee12a6c0c53e.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
cFormula of fc(W)Formula of gc(W,R)
3W2P2
4W3-d(W2)WP3-dblock(P2)P
-Wd(W2)+W⊙3-Pdblock(P2)+P⊙3
5W4-d(W3)WP4-dblock(P3)P
-d(W2)W2-dblock(P2)P2
-W2d(W2)-P2dblock(P2)
-Wd(W2)W-Pdblock(P2)P
+3W⊙2W2+3P⊙2P2
+WW⊙3+W⊙3W+PP⊙3+P⊙3P
", + "bbox": [ + 96, + 205, + 450, + 363 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Table 1. Formulas for $f_{c}$ and $g_{c}$ . Here we let $P = (W \\otimes 1_d) \\odot R$ for shorter notation. $\\mathrm{d}(X)$ returns the diagonal of matrix $X$ , $\\mathrm{d}_{\\mathrm{block}}(X)$ returns the diagonal block matrix from the $d \\times d$ diagonal blocks of matrix $X$ .", + "bbox": [ + 75, + 372, + 468, + 428 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2. Computational Complexity", + "text_level": 1, + "bbox": [ + 76, + 450, + 321, + 468 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We derive the space and time complexity for LongSync, and demonstrate its advantages over CEMP. The initialization step involves setting the weights of all edges to 1, which takes time $O(|E|)$ and space $O(n^{2})$ . For each iteration, LongSync updates the matrices $\\pmb{S}^{(t)}$ and $\\pmb{W}^{(t)}$ with equations (11) and (12), respectively. Computing $\\pmb{W}^{(t + 1)}$ involves two matrix subtractions, one scalar-matrix multiplication and one element-wise matrix exponential operation on $\\pmb{S}^{(t)} \\in \\mathbb{R}^{n \\times n}$ . Therefore the update step (12) takes at most $O(n^{2})$ time and space. Equation (11), on the other hand, involves a sequence of matrix operations on $\\pmb{P}^{(t)} = (\\pmb{W}^{(t)} \\otimes 1_{d}) \\odot \\pmb{R} \\in \\mathbb{R}^{dn \\times dn}$ and $\\pmb{W}^{(t)} \\in \\mathbb{R}^{n \\times n}$ , including matrix multiplications, element-wise multiplications and diagonal block selections. Computing $\\pmb{P}^{(t)}$ takes $O(d^{2}n^{2})$ memory and $O(d^{2}n^{2})$ time. The matrix operations on $\\pmb{P}^{(t)}$ take $O(K_{c}d^{3}n^{3})$ time and $O(d^{2}n^{2})$ space, and the matrix operations on $\\pmb{W}^{(t)}$ take $O(K_{c}n^{3})$ time and $O(n^{2})$ space, where $K_{c}$ is the number of terms in the equation for $f_{c}$ and $g_{c}$ . Therefore, each iteration over $t$ takes $O(K_{c}d^{3}n^{3})$ time and $O(d^{2}n^{2})$ space. To sum up, for LongSync, the time complexity is $O(K_{c}d^{3}n^{3})$ and the space complexity is $O(d^{2}n^{2})$ . For $c = 3,4,5,6$ , the number $K_{c}$ is equal to 1,3,9,32.", + "bbox": [ + 75, + 474, + 468, + 792 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In comparison, we consider the initialization step of CEMP. For each edge $ij \\in E$ and $L \\in N_{ij}^{c}$ , initializing CEMP involves computing and storing all the cycle inconsistency measures $d_{L}$ using equation (2). For each $L \\in N_{ij}^{c}$ , computing $d_{L}$ involves multiplying $c$ rotations, which takes $O(cd^3)$ time and $O(d^2)$ space. This step is repeated for each $ij \\in E$ and $L \\in N_{ij}^{c}$ , therefore the total time complexity is $O(cd^3\\sum_{ij \\in E}|N_{ij}^{c}|)$ and", + "bbox": [ + 75, + 792, + 470, + 902 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "the total space complexity is $O(d^{2}\\sum_{ij\\in E}|N_{ij}^{c}|)$ . Since for each edge there are $(n - 2)(n - 3)\\dots (n - c + 1) = O(n^{c - 2})$ cycle candidates, we know that $|N_{ij}^{c}|\\sim O(n^{c - 2})$ for each $ij\\in E$ in the worst case scenario of a dense graph. Therefore the initialization of CEMP takes $O(cd^3 n^{c - 2}|E|)$ time and $O(d^{2}n^{c - 2}|E|)$ space in the worst case. Given $c\\geq 4$ and $|E|\\sim n^2$ , CEMP requires much more time and space than LongSync.", + "bbox": [ + 496, + 90, + 890, + 199 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Theory for Uniform Corruption Model", + "text_level": 1, + "bbox": [ + 498, + 213, + 834, + 231 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this section, we present the exact recovery guarantee of LongSync under the uniform corruption model (UCM). UCM is a popular probabilistic model that is widely adopted for synthetic experiments of many previous works on group synchronization [8, 25, 28, 29, 32, 42]. The model $\\mathrm{UCM}(n,p,q_g)$ assumes that $G$ is an Erdős-Rényi graph with edge connection probability $p$ . For each edge $ij \\in E$ , $R_{ij}$ is generated independently as follows:", + "bbox": [ + 496, + 239, + 890, + 347 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {R} _ {i j} = \\left\\{ \\begin{array}{l l} \\boldsymbol {R} _ {i j} ^ {*} & \\text {w . p .} q _ {g}; \\\\ \\tilde {\\boldsymbol {R}} _ {i j} \\sim \\operatorname {H a a r} (\\mathcal {G}) & \\text {w . p .} 1 - q _ {g}. \\end{array} \\right.\n$$\n", + "text_format": "latex", + "bbox": [ + 575, + 359, + 812, + 398 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We also developed an exact recovery theory for a general model of adversarial corruption, which we include in section C.1 of the supplementary material. An informal version of our main result for UCM is stated in Theorem 4.1. Although the application of this paper is focused on rotation synchronization, the following theory for UCM is valid for any compact group $\\mathcal{G}$ , as explained in the supplementary material.", + "bbox": [ + 496, + 406, + 890, + 512 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Theorem 4.1. Let $0 < r < 1$ , $0 < q < 1$ , $0 < p \\leq 1$ , $\\mathcal{G} = SO(3)$ . Assume LongSync is applied with cycles of length $c$ , $n / \\log n \\sim p^{-(c-1) / (c-2-\\epsilon)} q_g^{-7(c-1) / 3(c-2)}$ for some $\\epsilon > 0$ and", + "bbox": [ + 496, + 523, + 890, + 571 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n1 / \\beta_ {t + 1} = r / \\beta_ {t} f o r a l l t \\geq 1.\n$$\n", + "text_format": "latex", + "bbox": [ + 607, + 585, + 781, + 602 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Then with appropriate choices of $\\beta_0, \\beta_1, r$ , and high probability, $\\max_{ij \\in E} |s_{ij}^* - s_{ij}^{(t)}| \\leq \\frac{1}{2c\\beta_t}$ for all $t \\geq 1$ .", + "bbox": [ + 496, + 614, + 890, + 650 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The major difficulty of proving Theorem 4.1 is the dependence in the cycle inconsistency measures for cycles in $N_{ij}^{c}$ when $c \\geq 4$ . Unlike the 3-cycle case, the cycle inconsistency measure of a 4-cycle $L_{1} = (ik_{1}, k_{1}k_{2}, k_{2}j)$ is correlated with that of $L_{2} = (ik_{1}, k_{1}k_{3}, k_{3}j)$ under UCM. Therefore the key concentration inequalities for the proof cannot be concluded from the standard Chernoff bounds. To overcome this theoretical obstacle, we have integrated various mathematical techniques from [3, 7, 22-24, 48] to prove the theorem, whose details are included in the supplementary material.", + "bbox": [ + 496, + 657, + 890, + 808 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Theorem 4.1 provides an upper bound of the sample complexity (the required graph size $n$ ) of LongSync for exact recovery of the ground truth solutions. This sample complexity is the closest to the information theoretic bound among all existing rotation synchronization methods. The comparison with previous works is summarized in Table 2.", + "bbox": [ + 496, + 810, + 890, + 898 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5264", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/1ed7eb623de572ff938b6db1ad89be5e8a112833468deb33a8f68e368b7759cc.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ReferenceSample Complexity
[25] for CEMPO(p-2qg-28/3)
[29] for ReSyncO(p-2qg-7)
Ours for CEMPO(p-2-εqg-14/3)
Ours for LongSync (c=3)O(p-2-εqg-14/3)
Ours for LongSync (c=4)O(p-1.5-εqg-3.5)
Ours for LongSync (any c)O(p-7(c-1/c-2-ε) qg-3(c-2))
Ours for LongSync (c→∞)O(p-1-εqg-7/3)
Information Theoretic Bound [6]O(p-1qg-2)
", + "bbox": [ + 80, + 88, + 467, + 253 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 2. Comparison of the sample complexity requirement. Lower absolute values of the powers on $p, q_{g}$ indicate better results. $\\epsilon$ is an arbitrarily small positive real number.", + "bbox": [ + 75, + 263, + 470, + 306 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Synthetic Data Experiment", + "text_level": 1, + "bbox": [ + 76, + 329, + 316, + 348 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We test LongSync on synthetic datasets generated with Uniform Corruption Model (UCM) and Uniform Bipartite Corruption Model (UBCM) respectively described in §5.1 and 5.2. For both models with their corresponding viewing graphs $G = ([n], E)$ , we sample the ground truth absolute rotation matrices $\\{R_i^*\\}_{i \\in [n]}$ independently from the Haar measure on SO(3), and we generate the observed relative rotations $\\{R_{ij}\\}_{ij \\in E}$ independently as follows:", + "bbox": [ + 75, + 354, + 468, + 477 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {R} _ {i j} = \\left\\{ \\begin{array}{l l} \\boldsymbol {R} _ {i} ^ {*} \\boldsymbol {R} _ {j} ^ {*} & \\text {w . p .} q _ {g}; \\\\ \\tilde {\\boldsymbol {R}} _ {i j} \\sim \\operatorname {H a a r} (S O (3)) & \\text {w . p .} 1 - q _ {g}. \\end{array} \\right.\n$$\n", + "text_format": "latex", + "bbox": [ + 135, + 487, + 406, + 527 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We use LongSync with cycle length $c$ , $\\beta_{t} = \\min (2^{t},20)$ and $T = 10$ and record the edge weights. For UCM we set $c = 4,5$ and for UBCM we only use $c = 4$ since no 5-cycles exist. For our method, we first build a weighted graph whose edge weights are estimated by LongSync. We then extract a maximum spanning tree (MST) of the resulting weighted graph. The resulting spanning tree is expected to be the cleanest possible spanning tree. To initialize our solution of absolute rotations, we first fix $\\pmb{R}_{1}$ as the identity rotation, and find the rest of $\\pmb{R}_{i}$ 's by consecutively multiplying the relative rotations along the spanning tree using the formula $\\pmb{R}_{i} = \\pmb{R}_{ij}\\pmb{R}_{j}$ . To refine our initialized solution, we apply IRLS with Geman-McClure [4] loss functions to minimize $\\sum_{ij\\in E}\\rho_{GM}(d_{\\angle}(\\pmb{R}_{ij},\\pmb{R}_i\\pmb{R}_j^T))$ , where $d_{\\angle}$ is the geodesic distance in $SO(3)$ . We refer to this method as LongSync+IRLS.", + "bbox": [ + 75, + 537, + 468, + 750 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To demonstrate the advantages of utilizing longer cycle information, we compare our method with IRLS initialized by other two different spanning trees. The first one is the random spanning tree, which uses no cycle information. The other one the MST extracted from the CEMP-estimated weights. Note that CEMP only uses 3-cycle information. We refer to these methods as IRLS and CEMP+IRLS respectively.", + "bbox": [ + 75, + 750, + 467, + 854 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Since the solution of absolute rotations is determined up to a global rotation, we align our estimated rotation $\\{\\hat{R}_i\\}$ with $\\{R_i^*\\}$ by $R_{\\mathrm{align}}$ that minimizes the $\\ell_1$ rotation alignment error", + "bbox": [ + 76, + 854, + 468, + 901 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/6968ab42c664c6e8f93e8d667cc771f80578269e0b14294c9b16ed4a19989a1c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 576, + 97, + 807, + 232 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/669a3c1c2a165ca8dab3004f7386cecf1dcb8051b66a15b01b38f281cb7c17b2.jpg", + "image_caption": [ + "Figure 1. Average errors for IRLS, CEMP+IRLS and LongSync +IRLS with $c = 4,5$ , using the uniform corruption (top) and uniform bipartite corruption (bottom) models. The mean errors are measured in degrees. LongSync4 and LongSync5 refer to LongSync with 4 and 5 cycles, respectively." + ], + "image_footnote": [], + "bbox": [ + 576, + 238, + 807, + 373 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "$\\sum_{i\\in [n]}\\| \\hat{\\pmb{R}}_i\\pmb{R}_{\\mathrm{align}} - \\pmb{R}_i^*\\| _F$ . We report the mean estimation error in degrees: $180\\cdot \\sum_{i\\in [n]}d_{\\angle}(\\hat{\\pmb{R}}_i\\pmb{R}_{\\mathrm{align}},\\pmb{R}_i^*) / n.$", + "bbox": [ + 500, + 482, + 890, + 521 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1. Uniform Corruption Model", + "text_level": 1, + "bbox": [ + 500, + 532, + 748, + 549 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For $\\mathrm{UCM}(n,p,q_g)$ , we take $n = 200$ and $p = 1$ and corruption probability $q = 1 - q_{g}$ ranges from 0.86 to 0.92. We report the average mean estimation error from 20 trials of the uniform corruption model in the top panel of Figure 1.", + "bbox": [ + 498, + 556, + 890, + 618 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We note that LongSync uniformly improves IRLS, and the mean error of LongSync decreases as the cycle length increases. When $q > 0.86$ , the expected number of clean 3-cycles for each edge is less than 4, and therefore longer cycles are helpful. The numerical result aligns with our theory that using longer cycles may tolerate higher corruption with fixed graph size $n$ .", + "bbox": [ + 498, + 619, + 890, + 710 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.2. Uniform Bipartite Corruption Model", + "text_level": 1, + "bbox": [ + 500, + 723, + 820, + 739 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For UBCM, we first generate the graph and relative rotations by $\\mathrm{UCM}(n,p,q_g)$ with $n = 200$ , $p = 1$ and $q = 1 - q_{g}$ ranging from 0.8 to 0.85. Then we split the nodes into two clusters of equal size and remove the intra-cluster edges for both clusters. The resulting graph is bipartite, where only cycles of even lengths exist. We report the mean estimation error from 20 trials in the bottom panel of Figure 1.", + "bbox": [ + 498, + 748, + 890, + 853 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We observe that LongSync with 4-cycles almost exactly recovers the rotations, while for other algorithms the rotation estimates are not even close to the ground truth.", + "bbox": [ + 498, + 854, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "5265", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "6. Real Data Experiment", + "text_level": 1, + "bbox": [ + 76, + 89, + 279, + 107 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We test distributed synchronization with LongSync on the PhotoTourism dataset [50] to demonstrate its advantages in accuracy and speed over other baselines. PhotoTourism is a large scale dataset consisting of 15 sets of images taken for 3D reconstruction. The smallest dataset consists of 247 cameras, and the largest dataset consists of 5433 cameras. The input graph and initial pairwise rotation estimates are provided in the dataset. In the following, we first explain the common steps for distributed synchronization, and our improvement using LongSync. We then describe our graph processing method for filtering bad nodes and edges, which also is applicable to other baseline methods.", + "bbox": [ + 75, + 118, + 472, + 284 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Steps in Distributed Synchronization:", + "text_level": 1, + "bbox": [ + 76, + 285, + 330, + 299 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "1. Graph partitioning. The first step involves partitioning the graph $G = ([n], E)$ into $K$ clusters $G_{i} = (V_{i}, E_{i}), i \\in [K]$ . In this paper we apply spectral clustering algorithm [37] on the adjacency matrix $G$ , where $K = 0.6\\sqrt{np}$ and $p = 2|E| / (n(n - 1))$ .", + "bbox": [ + 76, + 301, + 468, + 378 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "2. Synchronization within clusters. Run standard synchronization solvers for each cluster. In this work, we use the current state-of-the-art method MPLS [39]. Note that for each camera $p$ in cluster $k$ , one can only estimate the true rotation $R_{p}^{*}$ up to a global rotation $R_{k}$ . Namely, one only obtains $\\hat{R}^p \\approx R_p^* R_k^{-1}$ where $R_{k}$ is unknown and is the same for all cameras in cluster $k$ .", + "bbox": [ + 76, + 378, + 470, + 486 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3. Estimation of inter-cluster rotations. To find $\\pmb{R}_{p}^{*}$ of all cameras, one needs to solve $\\pmb{R}_{k}$ for all clusters. Namely, one needs to rotate and stitch the solutions of all clusters so that they are in the same reference frame. To do this, it is common to first estimate the inter-clsuter rotations $\\pmb{R}_{kl} := \\pmb{R}_{k}\\pmb{R}_{l}^{-1}$ between pairs of clusters $k,l$ , and then synchronize these relative rotations. To estimate each $\\pmb{R}_{kl}$ , we note that $\\pmb{R}_{kl} = \\pmb{R}_{p}^{*-1}\\pmb{R}_{pq}^{*}\\pmb{R}_{q}^{*}$ for each $p\\in V_k$ and $q\\in V_l$ . Therefore, one can use the rotations in the set $S_{kl} := \\{\\hat{\\pmb{R}}_p^{-1}\\pmb{R}_{pq}\\hat{\\pmb{R}}_q\\}_{p\\in V_k,q\\in V_l}$ to approximate $\\pmb{R}_{kl}$ . We remark that this step is crucial to the overall performance of distributed methods, and we compare the following methods for solving $\\pmb{R}_{kl}$ :", + "bbox": [ + 76, + 487, + 468, + 671 + ], + "page_idx": 6 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- MultSync [9]: Run synchronization on a multi-graph where each edge $kl$ is assigned a set of relative rotations $\\{\\hat{R}_p^{-1}R_{pq}\\hat{R}_q\\}_{p\\in V_k,q\\in V_l}$ . This combines the step 3 and 4 in a unified least squares formulation.", + "- Edge averaging using IRLS: We initialize $\\hat{R}_{kl}$ with the quaternion $\\ell_2$ mean of the set $S_{kl}$ and refine it using $\\ell_1$ -rotation averaging [19]. We refer to this method as IRLS in our comparison.", + "- Edge averaging using LongSync: We first perform LongSync with 4-cycles to estimate the weights of these inter-cluster edges (there are no 3-cycles for a bipartite graph). We next initialize $\\hat{R}_{kl}$ as the quaternion weighted $\\ell_2$ mean of $S_{kl}$ , using the edge weights from LongSync by their LongSync weights. Lastly, we refine the solution using [19]." + ], + "bbox": [ + 96, + 671, + 468, + 898 + ], + "page_idx": 6 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "4. Synchronization of inter-cluster rotations. This step is skipped for MultiSync. For other methods described in step 3, we find $\\pmb{R}_k$ (up to a rotation) for each cluster $k$ from the estimated $\\{R_{kl}\\}_{k,l\\in [K]}$ by MPLS.", + "5. Rotation merging. Finally, for each camera $p$ in cluster $k$ , the rotation estimate of $p$ is given by $R_p^{\\mathrm{final}} = \\hat{R}_p R_k^{-1}$ ." + ], + "bbox": [ + 498, + 90, + 893, + 181 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Next, we introduce our graph processing method to further boost the performance of all tested methods.", + "bbox": [ + 498, + 181, + 890, + 210 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Extra Improvement by Graph Processing:", + "text_level": 1, + "bbox": [ + 500, + 212, + 784, + 226 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "- Spectral clustering with Jaccard Index. For step 1, we use the Jaccard index matrix as the similarity matrix for spectral clustering, instead of the adjacency matrix. The $n \\times n$ Jaccard index matrix $A_J$ is defined as follows:", + "bbox": [ + 500, + 227, + 890, + 286 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {A} _ {J} (i, j) = \\left\\{ \\begin{array}{l l} 0 & i j \\notin E \\\\ \\frac {\\left| N _ {i} \\cap N _ {j} \\right|}{\\left| N _ {i} \\cup N _ {j} \\right|} & i j \\in E \\end{array} \\right. \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 604, + 294, + 890, + 335 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $N_{i}$ and $N_{j}$ denote the sets of neighboring nodes of node $i$ and $j$ , respectively. In this way, $\\mathbf{A}_{J}(i,j)$ is higher for the pair $ij$ contained in many 3-cycles, which is a more robust and nicely scaled statistics $(\\in [0,1])$ for measuring the local graph density around edge $ij$ .", + "bbox": [ + 509, + 340, + 890, + 417 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "- Refinement of intra-cluster edges and nodes. For step 2, after the MPLS step, we perform CEMP with 3-cycles to estimate the corruption level of the intra-cluster edges for each cluster. We remove a camera if the number of neighboring 'good' edges, i.e. the edges with corruption level less than 0.1, is less than 4. The numbers 4 and 0.1 are chosen to balance the number of remaining cameras and the quality of intra-cluster rotation estimates. In order to eliminate the sparsely connected components inside the cluster, we use the Matlab built-in hierarchical spectral clustering function on the remaining cameras with the 'cutoff' and 'depth' parameters as 2 and 4, and we keep the largest component. The absolute rotations for the remaining cameras are estimated by MPLS. We remark that one could replace CEMP by LongSync with 3-cycles. However, we have not observed significant difference in the performance.", + "bbox": [ + 500, + 417, + 892, + 657 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We respectively name MultiSync and IRLS with our new graph processing method as MultiSync(New) and IRLS(new). \"LongSync\" in our experiment refers to the full version of our algorithm: use LongSync weights for edge averaging in Step 3, with the graph processing step. We also compare with MPLS on the whole dataset, since it is a state-of-the-art non-distributed method, but we note that MPLS is significantly slower than all distributed methods. We report median error $180 \\cdot$ median( $\\{d_{\\angle}(\\hat{R}_i R_{\\text{align}}, R_i^*)\\}_{i \\in [n]}$ of the tested methods on 14 datasets in Figure 2. We exclude the result of Gendarmenmarkt since all methods return large estimation errors in the figure. The full results, including that of mean error are included in the supplementary material.", + "bbox": [ + 496, + 659, + 893, + 854 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In Figure 3, for each distributed method, we report the ratio (in percentage) between its total runtime on all datasets and that of the non-distributed MPLS. Namely, we compute", + "bbox": [ + 496, + 854, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "5266", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/bda3fb0b64824766e835611341624dac709532541ead448fa1660bf7d5e90c98.jpg", + "image_caption": [ + "Figure 2. Median error for rotations for each dataset measured in degrees." + ], + "image_footnote": [], + "bbox": [ + 86, + 88, + 893, + 272 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/baccd76e5c4d33968b5b02bea6b50a0629181129f71f2c34fc4ce00f4bdf8511.jpg", + "image_caption": [ + "Figure 3. Runtime ratio and average median and mean error gaps between the distributed methods and MPLS on the entire graph." + ], + "image_footnote": [], + "bbox": [ + 101, + 320, + 444, + 536 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "$\\sum_{d \\in D} t_{\\mathrm{dist},d} / \\sum_{d \\in D} t_{\\mathrm{MPLS},d}$ , where $D$ is the set of 15 datasets, and each $t_{\\mathrm{dist},d}$ and $t_{\\mathrm{MPLS},d}$ is respectively the runtime of the distributed method and MPLS on data $d$ . In the same figure, we present the mean/median error gap between each distributed method and MPLS. The mean and median error gap is respectively defined as $(\\bar{e}_{\\mathrm{dist}} - \\bar{e}_{\\mathrm{MPLS}}) / \\bar{e}_{\\mathrm{dist}}$ and $(\\hat{e}_{\\mathrm{dist}} - \\hat{e}_{\\mathrm{MPLS}}) / \\hat{e}_{\\mathrm{dist}}$ , where $\\bar{e}$ and $\\hat{e}$ respectively denote the mean and median error over all cameras.", + "bbox": [ + 75, + 608, + 468, + 731 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "From Figure 2 and 3, our method outperforms other distributed methods on 13 out of 15 datasets. The most significant improvement in mean error and median error are respectively $28.6\\%$ and $46.4\\%$ (in Notre Dame) compared to the best performing method between IRLS(new) and MultiSync(new). The improvement is even more significant when comparing to the original version of these baseline methods without our graph processing method. The only two datasets without improvement are Gendarmenmarkt and Union Square. Our method is comparable to others on Union Square, and all methods return large errors on Gendarmenmarkt due to many repetitive patterns in its 3D scene.", + "bbox": [ + 75, + 734, + 470, + 902 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The average mean and median error gap between our method and full MPLS are respectively $38.3\\%$ and $18.4\\%$ . Compared to the best performing method among others, our method reduces the average median error gap by $40.8\\%$ , and the average mean error gap by $9.6\\%$ . In terms of runtime, our method is uniformly faster than MultiSync and it is scalable on the largest dataset, taking less than $6\\%$ of the total runtime of full MPLS. In conclusion, our method significantly improves the result of distributed rotation synchronization without compromising runtime.", + "bbox": [ + 496, + 323, + 893, + 460 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In the supplementary material, we further demonstrate the improvement by our new graph processing method, which significantly improves the results of LongSync (without extra graph processing) in 14 of the 15 datasets. On these 14 datasets, the average reduction on mean error is $59.2\\%$ and the average reduction on median error is $28.5\\%$ .", + "bbox": [ + 496, + 467, + 893, + 556 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "7. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 595, + 612, + 611 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We propose LongSync, a robust and efficient algorithm for group synchronization. It modifies and vectorizes CEMP which enables efficient computation when using longer cycles. The theory we developed for LongSync is the strongest among all other existing results under UCM. Experiment shows that LongSync, together with our improved graph preprocessing method, achieves superior accuracy for distributed synchronization on large real datasets with competitive runtime. However, our method also has some limitations. First of all, in theory there is still a small gap of sample complexity from our method to the information theoretic one. Filling this gap is an open problem, which requires new tools and possibly more sophisticated analysis. Second, our graph preprocessing method is quite heuristic, and an automatic way of choosing parameters is needed. Our work also opens a door for some important future directions, including distributed partial permutation synchronization for multi-image matching, angular synchronization for Cryo-EM and Jigsaw Puzzles, and analysis of their algorithms.", + "bbox": [ + 496, + 628, + 893, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "5267", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 169, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Federica Arrigoni, Beatrice Rossi, Pasqualina Fragneto, and Andrea Fusiello. Robust synchronization in so(3) and se(3) via low-rank and sparse matrix decomposition. Computer Vision and Image Understanding, 174:95-113, 2018. 2", + "[2] A. S. Bandeira, N. Boumal, and A. Singer. Tightness of the maximum likelihood semidefinite relaxation for angular synchronization. arXiv preprint arXiv:1411.3272, 2014. 2", + "[3] Olivier Bousquet. A Bennett concentration inequality and its application to suprema of empirical processes. C. R. Math. Acad. Sci. Paris, 334(6):495-500, 2002. 5, 14, 16, 21", + "[4] Avishek Chatterjee and Venu Madhav Govindu. Efficient and robust large-scale rotation averaging. In IEEE International Conference on Computer Vision, ICCV 2013, Sydney, Australia, December 1-8, 2013, pages 521-528, 2013. 2, 6", + "[5] Avishek Chatterjee and Venu Madhav Govindu. Robust relative rotation averaging. IEEE Trans. Pattern Anal. Mach. Intell., 40(4):958-972, 2018. 2", + "[6] Yuxin Chen, Changho Suh, and Andrea J. Goldsmith. Information recovery from pairwise measurements. IEEE Trans. Inf. Theory, 62(10):5881-5905, 2016. 6", + "[7] Keresztély Corradi and András Hajnal. On the maximal number of independent circuits in a graph. Acta Mathematica Hungarica, 14(3-4):423-439, 1963. 5, 14", + "[8] Mihai Cucuringu. Synchronization over Z2 and community detection in signed multiplex networks with constraints. J. Complex Networks, 3(3):469-506, 2015. 5", + "[9] Andrea Porfiri Dal Cin, Luca Magri, Federica Arrigoni, Andrea Fusiello, and Giacomo Boracchi. Synchronization of group-labelled multi-graphs. In 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 6433-6443, 2021. 2, 7", + "[10] Frank Dellaert, David M. Rosen, Jing Wu, Robert E. Mahony, and Luca Carlone. Shonan rotation averaging: Global optimality by surfing $\\mathrm{so(p)^n}$ . In Andrea Vedaldi, Horst Bischof, Thomas Brox, and Jan-Michael Frahm, editors, Computer Vision - ECCV 2020 - 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part VI, volume 12351 of Lecture Notes in Computer Science, pages 292-308. Springer, 2020. 2", + "[11] Anders Eriksson, Carl Olsson, Fredrik Kahl, and Tat-Jun Chin. Rotation averaging and strong duality, 2017.", + "[12] Anders P. Eriksson, Carl Olsson, Fredrik Kahl, and Tat-Jun Chin. Rotation averaging and strong duality. In 2018 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2018, Salt Lake City, UT, USA, June 18-22, 2018, pages 127-135. Computer Vision Foundation / IEEE Computer Society, 2018.", + "[13] Anders P. Eriksson, Carl Olsson, Fredrik Kahl, and Tat-Jun Chin. Rotation averaging with the chordal distance: Global minimizers and strong duality. IEEE Trans. Pattern Anal. Mach. Intell., 43(1):256-268, 2021. 2", + "[14] Xiang Gao, Hainan Cui, Menghan Li, Zexiao Xie, and Shuhan Shen. Irav3: Hierarchical incremental rotation averaging on the fly. IEEE Trans. Circuits Syst. Video Technol., 33(4):2001-2006, 2023. 2", + "[15] Xiang Gao, Hainan Cui, and Shuhan Shen. Incremental rotation averaging revisited and more: A new rotation averaging benchmark. arXiv preprint, abs/2309.16924, 2023.", + "[16] Xiang Gao, Lingjie Zhu, Hainan Cui, Zexiao Xie, and Shuhan Shen. IRA++: distributed incremental rotation averaging. IEEE Trans. Circuits Syst. Video Technol., 32(7):4885-4892, 2022. 2" + ], + "bbox": [ + 78, + 114, + 470, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[17] Thomas Goldstein, Paul Hand, Choongbum Lee, Vladislav Voroninski, and Stefano Soatto. Shapefit and shapekick for robust, scalable structure from motion. In Computer Vision - ECCV 2016 - 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part VII, pages 289-304, 2016. 1", + "[18] V.M. Govindu. Lie-algebraic averaging for globally consistent motion estimation. In Proceedings of the 2004 IEEE Computer Society Conference on Computer Vision and Pattern Recognition, 2004. CVPR 2004., volume 1, pages I-I, 2004. 2", + "[19] R. Hartley, K. Aftab, and J. Trumpf. L1 rotation averaging using the weiszfeld algorithm. In Computer Vision and Pattern Recognition (CVPR), 2011 IEEE Conference on, pages 3041-3048. IEEE, 2011. 7", + "[20] Richard I. Hartley, Khurrum Aftab, and Jochen Trumpf. L1 rotation averaging using the weiszfeld algorithm. In The 24th IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2011, Colorado Springs, CO, USA, 20-25 June 2011, pages 3041-3048, 2011. 2", + "[21] Xiangru Huang, Zhenxiao Liang, Xiaowei Zhou, Yao Xie, Leonidas J. Guibas, and Qixing Huang. Learning transformation synchronization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2019. 2", + "[22] Svante Janson. Poisson approximation for large deviations. Random Structures & Algorithms, 1(2):221-229, 1990. 5, 14, 16", + "[23] Svante Janson. Large deviations for sums of partly dependent random variables. Random Structures & Algorithms, 24(3):234-248, 2004. 18, 21", + "[24] Jeong Han Kim and Van H Vu. Concentration of multivariate polynomials and its applications. Combinatorica, 20(3):417-434, 2000. 5, 14, 18", + "[25] Gilad Lerman and Yunpeng Shi. Robust group synchronization via cycle-edge message passing. Foundations of Computational Mathematics, 22(6):1665-1741, 2022. 1, 2, 3, 5, 6, 13, 14, 16, 19, 21", + "[26] Heng Li, Zhaopeng Cui, Shuaicheng Liu, and Ping Tan. RAGO: recurrent graph optimizer for multiple rotation averaging. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2022, New Orleans, LA, USA, June 18-24, 2022, pages 15766-15775. IEEE, 2022. 2", + "[27] Shaohan Li, Yunpeng Shi, and Gilad Lerman. Fast, accurate and memory-efficient partial permutation synchronization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 15735-15743, June 2022. 2", + "[28] Shuyang Ling. Near-optimal performance bounds for orthogonal and permutation group synchronization via spectral methods. arXiv preprint arXiv:2008.05341, 2020. 5", + "[29] Huikang Liu, Xiao Li, and Anthony Man-Cho So. Resync: Riemannian subgradient-based robust rotation synchronization. In Advances in Neural Information Processing Systems: 37th Annual Conference on Neural Information Processing Systems, 2023. 2, 5, 6", + "[30] Tyler Maunu and Gilad Lerman. Depth descent synchronization in SO(D). Int. J. Comput. Vis., 131(4):968-986, 2023. 2", + "[31] Onur Özyesil and Amit Singer. Robust camera location estimation by convex programming. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 2674-2683, 2015. 1", + "[32] Deepti Pachauri, Risi Kondor, and Vikas Singh. Solving the multi-way matching problem by permutation synchronization. In" + ], + "bbox": [ + 501, + 92, + 893, + 890 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "5268", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "C. J. C. Burges, L. Bottou, M. Welling, Z. Ghahramani, and K. Q. Weinberger, editors, Advances in Neural Information Processing Systems 26, pages 1860-1868. Curran Associates, Inc., 2013. 5", + "[33] Álvaro Parra, Shin-Fang Ch'ng, Tat-Jun Chin, Anders P. Eriksson, and Ian Reid. Rotation coordinate descent for fast globally optimal rotation averaging. In IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2021, virtual, June 19-25, 2021, pages 4298-4307. Computer Vision Foundation / IEEE, 2021. 2", + "[34] Pulak Purkait, Tat-Jun Chin, and Ian Reid. *Neurora: Neural robust rotation averaging*, 2020. 2", + "[35] Ian C Ross and Frank Harary. On the determination of redundancies in sociometric chains. Psychometrika, 17(2):195-208, 1952. 4, 12", + "[36] Tianwei Shen, Siyu Zhu, Tian Fang, Runze Zhang, and Long Quan. Graph-based consistent matching for structure-from-motion. In European Conference on Computer Vision, pages 139-155. Springer, 2016. 2", + "[37] Jianbo Shi and J. Malik. Normalized cuts and image segmentation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 22(8):888-905, 2000. 7", + "[38] Yunpeng Shi and Gilad Lerman. Estimation of camera locations in highly corrupted scenarios: All about that base, no shape trouble. In 2018 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2018, Salt Lake City, UT, USA, June 18-22, 2018, pages 2868-2876, 2018. 2", + "[39] Yunpeng Shi and Gilad Lerman. Message passing least squares framework and its application to rotation synchronization. In Proceedings of the 37th International Conference on Machine Learning (ICML), 2020. 1, 2, 7", + "[40] Yunpeng Shi, Shaohan Li, and Gilad Lerman. Robust multi-object matching via iterative reweighting of the graph connection laplacian. Advances in Neural Information Processing Systems, 2020-December, 2020. 2", + "[41] Chitturi Sidhartha and Venu Madhav Govindu. It is all in the weights: Robust rotation averaging revisited. In International Conference on 3D Vision, 3DV 2021, London, United Kingdom, December 1-3, 2021, pages 1134-1143. IEEE, 2021. 2", + "[42] Amit Singer. Angular synchronization by eigenvectors and semidefinite programming. Applied and computational harmonic analysis, 30(1):20-36, 2011. 2, 5", + "[43] A. Singer and Y. Shkolnisky. Angular synchronization by eigenvectors and semidefinite programming: Analysis and application to class averaging in cryo-electron microscopy, 2009. 2", + "[44] Johan Thunberg, Florian Bernard, and Jorge Goncalves. Distributed methods for synchronization of orthogonal matrices over graphs. Automatica, 80:243-252, jun 2017. 2", + "[45] Roberto Tron and René Vidal. Distributed image-based 3-d localization of camera sensor networks. In Proceedings of the 48th IEEE Conference on Decision and Control, CDC 2009, December 16-18, 2009, Shanghai, China, pages 901-908, 2009. 2", + "[46] Anton Voropaev and Sergey Perepechko. The number of fixed length cycles in undirected graph. explicit formulae in case of small lengths. Bulletin of PFUR. Series Mathematics, Information Sciences, Physics, pages 5-11, 01 2012. 4", + "[47] Anton Voropaev and Sergey Perepechko. The number of fixed length cycles in undirected graph. explicit formulae in case of small lengths. Bulletin of PFUR. Series Mathematics, Information Sciences, Physics, pages 5-11, 01 2012. 13", + "[48] Van H Vu. A large deviation result on the number of small" + ], + "bbox": [ + 78, + 90, + 470, + 888 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "subgraphs of a random graph. Combinatorics, Probability and Computing, 10(1):79-94, 2001. 5, 14, 17", + "[49] Lanhui Wang and Amit Singer. Exact and stable recovery of rotations for robust synchronization. Information and Inference, 2013. 2", + "[50] Kyle Wilson and Noah Snavely. Robust global translations with 1dsfm. In Computer Vision - ECCV 2014 - 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part III, pages 61-75, 2014. 7", + "[51] Christopher Zach, Manfred Klopschitz, and Marc Pollefeys. Disambiguating visual relations using loop constraints. In The Twenty-Third IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2010, San Francisco, CA, USA, 13-18 June 2010, pages 1426-1433, 2010. 2" + ], + "bbox": [ + 501, + 92, + 893, + 282 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "5269", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2024/Efficient Detection of Long Consistent Cycles and its Application to Distributed Synchronization/c0dd6b15-09a8-4eba-bcfb-24b8083dd375_model.json b/2024/Efficient Detection of Long Consistent Cycles and its Application to Distributed Synchronization/c0dd6b15-09a8-4eba-bcfb-24b8083dd375_model.json new file mode 100644 index 0000000000000000000000000000000000000000..41d0e24a59110e2adb81a0c9f034d1c84bd7569b --- /dev/null +++ b/2024/Efficient Detection of Long Consistent Cycles and its Application to Distributed Synchronization/c0dd6b15-09a8-4eba-bcfb-24b8083dd375_model.json @@ -0,0 +1,2420 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.0, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.241, + 0.131, + 0.737, + 0.177 + ], + "angle": 0, + "content": "Efficient Detection of Long Consistent Cycles and its Application to Distributed Synchronization *" + }, + { + "type": "text", + "bbox": [ + 0.141, + 0.204, + 0.327, + 0.257 + ], + "angle": 0, + "content": "Shaohan Li \nUniversity of Minnesota \nli000743@umn.edu" + }, + { + "type": "text", + "bbox": [ + 0.369, + 0.204, + 0.603, + 0.258 + ], + "angle": 0, + "content": "Yunpeng Shi \nUniversity of California, Davis \nypshi@ucdavis.edu" + }, + { + "type": "text", + "bbox": [ + 0.645, + 0.204, + 0.831, + 0.257 + ], + "angle": 0, + "content": "Gilad Lerman \nUniversity of Minnesota \nlerman@umn.edu" + }, + { + "type": "title", + "bbox": [ + 0.237, + 0.292, + 0.312, + 0.308 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.324, + 0.474, + 0.581 + ], + "angle": 0, + "content": "Group synchronization plays a crucial role in global pipelines for Structure from Motion (SfM). Its formulation is nonconvex and it is faced with highly corrupted measurements. Cycle consistency has been effective in addressing these challenges. However, computationally efficient solutions are needed for cycles longer than three, especially in practical scenarios where 3-cycles are unavailable. To overcome this computational bottleneck, we propose an algorithm for group synchronization that leverages information from cycles of lengths ranging from three to six with a time complexity of order \\( O(n^{3}) \\) (or \\( O(n^{2.373}) \\) when using a faster matrix multiplication algorithm). We establish non-trivial theory for this and related methods that achieves competitive sample complexity, assuming the uniform corruption model. To advocate the practical need for our method, we consider distributed group synchronization, which requires at least 4-cycles, and we illustrate state-of-the-art performance by our method in this context." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.612, + 0.203, + 0.628 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.637, + 0.472, + 0.865 + ], + "angle": 0, + "content": "Structure from Motion (SfM) asks to recover the 3D structure of a stationary scene from multiple images taken by cameras from different orientations and locations. In the past decade, the global SfM pipeline has become increasingly popular due to its several advantages over the incremental pipelines [17, 31]. First of all, global SfM requires only one implementation of bundle adjustment, making it more efficient in computation. Second, it estimates camera poses in a global optimization framework which mitigates the drifting issue of the incremental pipelines. Despite the popularity of global SfM pipelines, the estimation of global camera poses (e.g., orientations) remains a highly challenging problem. For instance, estimating camera orientations from their relative measurements, often called rotation synchronization, is a highly nonconvex graph optimization problem. In typical scenarios of highly noisy or corrupted measurements" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.293, + 0.892, + 0.324 + ], + "angle": 0, + "content": "of relative orientations, many common solutions of rotation synchronization have poor accuracy and slow convergence." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.325, + 0.895, + 0.642 + ], + "angle": 0, + "content": "Given these challenges, theoretical developments have demonstrated the critical role of cycle-consistency information in inferring corrupted measurements [25]. In practice, the consistency constraint on 3-cycles was utilized to estimate the error of each measured relative orientation. It also helped nonconvex iterative rotation synchronization solvers avoid spurious local minima and achieve significantly higher accuracy [39]. However, the usage of 3-cycles largely limits the application of these improved algorithms to other important scenarios. One scenario involves a sparse viewing graph lacking sufficient 3-cycles. This often occurs when the size of the graph is too large to densely measure the relative orientations on its edges, which could happen in certain case for the molecular orientation estimation in cryo-electron microscopy imaging. Another scenario is orientation estimation for each piece of jigsaw puzzles, where the graph is a 2D lattice and 3-cycle does not exist. Lastly, in distributed SfM, edges between any two clusters of nodes form a bipartite graph, and cycles of odd length do not exist. Our numerical results primarily emphasize the practical scenario of distributed SfM, which holds particular relevance for the broader computer vision community." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.643, + 0.896, + 0.809 + ], + "angle": 0, + "content": "Despite the multiple critical applications of long-cycle consistency, inferring measurement noise from long cycles is challenging in both computation and theory. First of all, the number of cycles grows exponentially with the cycle length, and measuring cycle inconsistencies for each long cycle is computationally intractable. Moreover, developing theoretical guarantees for long-cycle inference methods is fundamentally more difficult than the 3-cycle case. Indeed, in a random graph setting, a set of longer cycles are more likely to share common edges, making their consistency score highly correlated. Therefore, new tools are required to handle the correlated empirical process." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.811, + 0.897, + 0.902 + ], + "angle": 0, + "content": "In this work, we propose the first practical method, LongSync, for inferring edge corruption levels from long cycle consistency information. For this purpose, we carefully modify and vectorize the Cycle Edge Message Passing (CEMP) method [25]. This nontrivial modification drastically reduces its computational complexity when using longer cycles. Moreover," + }, + { + "type": "page_footnote", + "bbox": [ + 0.076, + 0.875, + 0.388, + 0.901 + ], + "angle": 0, + "content": "*This work was supported by NSF award DMS-2152766. Supplementary code: https://github.com/sli743/LongSync" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.957 + ], + "angle": 0, + "content": "5260" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.228 + ], + "angle": 0, + "content": "by employing a more delicate analysis and incorporating new tools from probability theory and combinatorics, we develop a significantly stronger exact recovery result with a general cycle length under a popular probabilistic model. The sample complexity in our theory is the lowest among all practical rotation synchronization methods. Although we limit our scope to the application of distributed SfM, our algorithm and theory applies to any finite-dimensional linear group in group synchronization, and not just \\(SO(3)\\) in rotation synchronization." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.24, + 0.221, + 0.255 + ], + "angle": 0, + "content": "1.1. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.265, + 0.472, + 0.4 + ], + "angle": 0, + "content": "Earlier rotation synchronization methods [2, 10-13, 18, 33, 42, 43] seek to minimize a least squares energy function. They can be described as relaxed versions of the maximum likelihood estimator under the additive Gaussian noise model, but they are not robust in the presence of outliers or heavy-tailed noise. Nevertheless, in the case of global SfM, the initially estimated relative camera rotations can be severely corrupted due to the erroneous keypoint matches and the subsequent poor estimation of fundamental matrices." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.402, + 0.472, + 0.704 + ], + "angle": 0, + "content": "To handle outliers, robust rotation synchronization methods either minimize a robust energy function or reweigh/trim the viewing graph based on the corruption levels of the edges. Wang and Singer [49] minimizes a corresponding \\(\\ell_1\\) objective function using semidefinite programming (SDP) relaxation, which is slow in practice. Other energy minimization methods are typically nonconvex, which include the Weiszfeld algorithm [20] and the Riemannian subgradient method [29] for \\(\\ell_1\\) minimization, and the iteratively reweighted least squares (IRLS) for minimizing general \\(\\ell_p\\) [5] and Geman-McClure [4] loss functions. However, all these methods heavily rely on good initialization. Sidhartha and Govindu [41] partially remedy the issue using adaptive Geman-McClure loss functions, but their approach remains sensitive to the initialized weights. Maunu and Lerman [30] propose to solve rotation synchronization by an iterative robust averaging method that utilizes Tukey depth, but they have not demonstrated effective performance for real SfM applications. Arrigoni et al. [1] applies a low-rank and sparse matrix decomposition method to SO(3) and SE(3) synchronization, but it is even less robust to outliers than IRLS." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.705, + 0.472, + 0.902 + ], + "angle": 0, + "content": "Instead of employing a robust objective function, Shen et al. [36] and Zach et al. [51] uses the 3-cycle consistency constraint to detect and remove corrupted relative orientations. Lerman and Shi [25] take one step further to estimate the corruption level of each relative measurement by a novel cycle-edge message passing (CEMP) algorithm. They then use the estimated corruption levels to reweigh the graph and solve rotation synchronization using a weighted least squares method. This message passing procedure was further combined with IRLS to boost its accuracy in [39]. Particular versions and extensions of this procedure for permutation and partial permutation synchronization, which are relevant to the matching component of SfM, were discussed in [27, 40]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.288 + ], + "angle": 0, + "content": "However, all the previously mentioned cycle-based methods [25, 27, 36, 38-40, 51] only use 3-cycles in practice, limiting their application for distributed synchronization. Indeed, the standard distributed synchronization often requires \"stitching\" local solutions by synchronizing the relative rotations between clusters. Each of these inter-cluster rotations is estimated by \"averaging\" the edges between the two clusters. These edges form a bipartite graph, and the minimal cycle length is 4. As pointed in [51], the number of operations for computing long cycle consistency information scales exponentially with the cycle length. Therefore, none of the existing distributed rotation synchronization methods directly exploits long cycle information due to this computational challenge." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.289, + 0.895, + 0.501 + ], + "angle": 0, + "content": "The earlier distributed methods for \\( SO(d) \\) synchronization, such as [45] and [44], minimize a least squares energy and are not robust to outliers. A series of distributed SfM methods [14-16] implement incremental SfM algorithms for each cluster. However, these methods do not employ a standard rotation synchronization algorithm, as they require additional information such as the number of keypoint matches between images. Moreover, the incremental methods are slower since they require multiple rounds of global rotation synchronization. MultiSync [9] synchronizes the inter-cluster rotations directly using all inter-cluster edges among all clusters, by formulating a novel synchronization problem on a multi-graph. Although it utilizes a more unified formulation, its objective function is least squares which largely limits its robustness to outliers." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.502, + 0.892, + 0.577 + ], + "angle": 0, + "content": "A recent and different type of methods for rotation synchronization use deep learning [21, 26, 34]. However, these methods are supervised and thus may not generalize well when switching datasets. Moreover, like many other previous methods, they lack theoretical guarantees." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.579, + 0.892, + 0.654 + ], + "angle": 0, + "content": "A common theoretical setting to assess the performance of rotation synchronization algorithms is the uniform corruption model (UCM) described in §4. We provide the best sample complexity for LongSync, even with only 3 cycles, among all previously established estimates for the UCM model." + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.666, + 0.75, + 0.681 + ], + "angle": 0, + "content": "1.2. Contributions of This Work" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.689, + 0.892, + 0.765 + ], + "angle": 0, + "content": "- We propose the first practical algorithm that infers edge corruption levels from long cycle consistency information. The computation complexity of our method is reduced from \\( O(n^{c}) \\) to \\( O(n^{3}) \\) (or possibly \\( O(n^{2.373}) \\)) for cycle length \\( c \\leq 6 \\) and \\( O(n^{[(c + 3) / 2]}) \\) for \\( c > 6 \\)." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.766, + 0.894, + 0.841 + ], + "angle": 0, + "content": "- We establish sample complexity estimates for our method under the uniform corruption model, where we get closer to the information theoretic bound than any other existing work. Our proof requires delicate analysis and it also improves previous estimates for the CEMP algorithm." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.841, + 0.892, + 0.901 + ], + "angle": 0, + "content": "- We introduce a new graph partition and graph preprocessing method that utilizes our inference method, and demonstrate the effectiveness of our pipeline in boosting the performance of distributed synchronization." + }, + { + "type": "list", + "bbox": [ + 0.5, + 0.689, + 0.894, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "5261" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.47, + 0.122 + ], + "angle": 0, + "content": "- Extensive numerical experiments demonstrate the outstanding performance of our method." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.134, + 0.421, + 0.15 + ], + "angle": 0, + "content": "2. Problem Formulation and Preliminaries" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.16, + 0.473, + 0.432 + ], + "angle": 0, + "content": "Assume a graph \\( G = ([n], E) \\) where \\( [n] \\) is the set of nodes indexed by \\( \\{1, 2, \\dots, n\\} \\) and \\( E \\) is the set of edges. Given a mathematical group \\( \\mathcal{G} \\), each graph node is assigned an underlying ground truth group element \\( R_{i}^{*} \\), where \\( R_{i}^{*} \\in \\mathcal{G} \\) and we use star superscript to emphasize the ground truth. For each edge \\( ij \\in E \\), we observe a relative group ratio \\( R_{ij} \\in \\mathcal{G} \\), whose clean counterpart is \\( R_{ij}^{*} = R_{i}^{*}R_{j}^{*-1} \\). Group synchronization aims to recover the ground truth group elements \\( \\{R_{i}^{*}\\}_{i \\in [n]} \\) from the possibly noisy and corrupted measurements \\( \\{R_{ij}\\}_{ij \\in E} \\). In this paper, we focus on the case of rotation synchronization, which is a special case of group synchronization with \\( \\mathcal{G} = \\mathrm{SO}(d) \\). For applications in camera orientation synchronization (\\( d = 3 \\)), we estimate absolute rotations for each node \\( i \\in [n] \\) from measured relative rotations of edges in \\( E \\). Note that since \\( \\{R_{i}^{*}\\}_{i \\in [n]} \\) and \\( \\{R_{i}^{*}R_{0}\\}_{i \\in [n]} \\) generate the same set of relative rotations, one can only estimate \\( \\{R_{i}^{*}\\}_{i \\in [n]} \\) up to a global rotation. The generalization to any linear groups is discussed in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.44, + 0.31, + 0.455 + ], + "angle": 0, + "content": "2.1. Notations and Definitions" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.463, + 0.47, + 0.51 + ], + "angle": 0, + "content": "We denote the adjacency matrix of graph \\(G\\) as \\(\\mathbf{A}\\), and form a pairwise observation matrix \\(\\mathbf{R} \\in \\mathbb{R}^{dn \\times dn}\\) by stacking the \\(\\mathbf{R}_{ij}\\)'s (for \\(ij \\notin E\\), set \\(\\mathbf{R}_{ij} = \\mathbf{0}_{3 \\times 3}\\)):" + }, + { + "type": "equation", + "bbox": [ + 0.162, + 0.518, + 0.386, + 0.59 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {R} := \\left( \\begin{array}{c c c c} \\boldsymbol {R} _ {1 1} & \\boldsymbol {R} _ {1 2} & \\dots & \\boldsymbol {R} _ {1 n} \\\\ \\boldsymbol {R} _ {2 1} & \\boldsymbol {R} _ {2 2} & \\dots & \\boldsymbol {R} _ {2 n} \\\\ \\vdots & \\vdots & \\ddots & \\vdots \\\\ \\boldsymbol {R} _ {n 1} & \\boldsymbol {R} _ {n 2} & \\dots & \\boldsymbol {R} _ {n n} \\end{array} \\right).\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.598, + 0.472, + 0.718 + ], + "angle": 0, + "content": "We list the matrix operations used in the paper. For matrices \\(X\\) and \\(Y\\), the operations \\(X \\otimes Y\\), \\(X \\odot Y\\), \\(X \\oslash Y\\) respectively denote the Kronecker product, Hardmard (element-wise) multiplication and Harnard division between \\(X\\) and \\(Y\\). \\(X^{\\odot k}\\) denotes the element-wise matrix \\(k\\)-power. For block matrices, \\(\\langle X, Y \\rangle_{\\mathrm{block}}\\) denotes the blockwise inner product of \\(X\\) and \\(Y\\), i.e. \\(\\langle X, Y \\rangle_{\\mathrm{block}}(i,j) = \\langle X[i,j], Y[i,j] \\rangle\\), where \\([i,j]\\) refers to the corresponding block of the matrix." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.727, + 0.342, + 0.743 + ], + "angle": 0, + "content": "2.2. Review of CEMP for \\(c\\)-Cycles" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.75, + 0.47, + 0.826 + ], + "angle": 0, + "content": "We assume the above setting of \\(SO(d)\\) synchronization. Let \\(\\mathcal{D}\\) be any bi-invariant metric on \\(SO(d)\\). We assume a fixed number of cycles, \\(c\\), and denote by \\(N_{ij}^{c}\\) the set of simple cycles of length \\(c\\) (or simple \\(c\\)-cycles) containing edge \\(ij\\). CEMP [25] aims to estimate for each edge \\(ij\\) the corruption level" + }, + { + "type": "equation", + "bbox": [ + 0.212, + 0.835, + 0.47, + 0.854 + ], + "angle": 0, + "content": "\\[\ns _ {i j} ^ {*} = \\mathcal {D} \\left(\\boldsymbol {R} _ {i j}, \\boldsymbol {R} _ {i j} ^ {*}\\right), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.861, + 0.359, + 0.876 + ], + "angle": 0, + "content": "from the set of cycle inconsistency measures" + }, + { + "type": "equation", + "bbox": [ + 0.215, + 0.886, + 0.47, + 0.903 + ], + "angle": 0, + "content": "\\[\nd _ {L} = \\mathcal {D} \\left(\\boldsymbol {R} _ {L}, \\boldsymbol {R} _ {i j}\\right) \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.892, + 0.167 + ], + "angle": 0, + "content": "where cycle \\( L = (ik_{1}, k_{1}k_{2}, \\dots, k_{c-2}j, ji) \\in N_{ij}^{c} \\) and \\( R_{L} := R_{ik_{1}}R_{k_{1}k_{2}}\\dots R_{k_{c-2}j} \\). The estimated \\( s_{ij}^{*} \\) can then be used for extracting a clean subgraph, or to implement a weighted least squares solver where higher weights are assigned to cleaner edges." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.168, + 0.894, + 0.198 + ], + "angle": 0, + "content": "It is obvious that if all the edges in \\(L\\) are clean then \\(d_{L} = 0\\). Moreover, due to bi-invariance of \\(\\mathcal{D}\\), the following holds true" + }, + { + "type": "equation", + "bbox": [ + 0.607, + 0.207, + 0.892, + 0.225 + ], + "angle": 0, + "content": "\\[\nd _ {L} = s _ {i j} ^ {*} \\text {w h e n e v e r} L \\in G _ {i j} ^ {c}, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.232, + 0.894, + 0.277 + ], + "angle": 0, + "content": "where \\( G_{ij}^{c} \\) is the set of good \\( c \\)-cycles with respect to \\( ij \\), i.e. the set of cycles \\( L \\in N_{ij}^{c} \\) such that \\( ik_1, \\dots, k_{c-2j} \\) are clean. This gives a sufficient condition for \\( d_L \\) to be an exact estimator of \\( s_{ij}^{*} \\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.278, + 0.892, + 0.34 + ], + "angle": 0, + "content": "To estimate the corruption levels of each edge \\( ij \\), CEMP initializes the edge weight of each \\( ij \\in E \\) as \\( w_{ij}^{(0)} = 1 \\). It then iteratively updates the corruption level estimate as the following convex combination of \\( d_L \\)'s:" + }, + { + "type": "equation", + "bbox": [ + 0.617, + 0.346, + 0.892, + 0.382 + ], + "angle": 0, + "content": "\\[\ns _ {i j} ^ {(t)} = \\sum_ {L \\in N _ {i j} ^ {c}} w _ {L} ^ {(t)} d _ {L} / z _ {i j} ^ {(t)} \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.391, + 0.892, + 0.433 + ], + "angle": 0, + "content": "where \\( z_{ij}^{(t)} = \\sum_{L\\in N_{ij}^c}w_L^{(t)} \\). The cycle weights \\( w_{L}^{(t + 1)} \\) are computed from the edge weights \\( w_{e}^{(t + 1)} = e^{-\\beta_{t}s_{e}^{(t)}} \\):" + }, + { + "type": "equation", + "bbox": [ + 0.556, + 0.44, + 0.892, + 0.476 + ], + "angle": 0, + "content": "\\[\nw _ {L} ^ {(t + 1)} = \\prod_ {e \\in L \\backslash \\{i j \\}} w _ {e} ^ {(t + 1)} = \\prod_ {e \\in L \\backslash \\{i j \\}} e ^ {- \\beta_ {t} s _ {e} ^ {(t)}}, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.486, + 0.895, + 0.624 + ], + "angle": 0, + "content": "so that \\( w_{L}^{(t + 1)} \\) focuses on good cycles. The cycle weights and edge corruption levels are alternatingly updated and improved from each other. Interestingly, it is proved in [25] under two different corruption models that CEMP converges linearly to the ground truth corruption estimates under mild conditions for \\( c = 3 \\). In practice, CEMP only uses 3-cycles for consideration of efficiency. For longer cycles, the complexity of CEMP scales exponentially with the cycle length \\( c \\) (which is discussed in §3.2), and the convergence guarantee of CEMP remains unknown." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.636, + 0.715, + 0.653 + ], + "angle": 0, + "content": "3. Our method: LongSync" + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.661, + 0.802, + 0.677 + ], + "angle": 0, + "content": "3.1. LongSync: Modification of CEMP" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.684, + 0.892, + 0.804 + ], + "angle": 0, + "content": "Our goal is to develop a scalable variant of CEMP for any fixed number of cycles, \\( c \\geq 3 \\). The main computational bottleneck of step (4) in CEMP is that computing and summing the cycle inconsistency measures takes \\( \\sum_{ij \\in E} |N_{ij}^c| = O(n^c) \\) operations and memory. Therefore, to develop a scalable algorithm, we aim to take weighted average over \\( d_L \\) without explicitly computing and storing each \\( d_L \\). To achieve this, we propose the following specification and modification on CEMP:" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.805, + 0.892, + 0.833 + ], + "angle": 0, + "content": "- Use Chordal distance on SO(d). We suggest the distance function" + }, + { + "type": "equation", + "bbox": [ + 0.599, + 0.833, + 0.807, + 0.851 + ], + "angle": 0, + "content": "\\[\n\\mathcal {D} \\left(\\boldsymbol {R} _ {1}, \\boldsymbol {R} _ {2}\\right) = \\sqrt {1 - \\left\\langle \\boldsymbol {R} _ {1} , \\boldsymbol {R} _ {2} \\right\\rangle / d}.\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.892, + 0.9 + ], + "angle": 0, + "content": "This distance is proportional to the Chordal distance on \\(SO(3)\\), which is the Euclidean distance between two rotations embedded in \\(\\mathbb{R}^{d\\times d}\\)." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5262" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.078, + 0.092, + 0.47, + 0.152 + ], + "angle": 0, + "content": "- Use weighted quadratic average for corruption level update. Instead of updating the corruption level estimates by a weighted average of \\( d_{L} \\), we use the weighted quadratic average of \\( d_{L} \\), namely" + }, + { + "type": "equation", + "bbox": [ + 0.194, + 0.163, + 0.47, + 0.203 + ], + "angle": 0, + "content": "\\[\ns _ {i j} ^ {(t)} = \\sqrt {\\sum_ {L \\in N _ {i j} ^ {c}} w _ {L} ^ {(t)} d _ {L} ^ {2} / z _ {i j} ^ {(t)}} \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.207, + 0.448, + 0.222 + ], + "angle": 0, + "content": "where the update rule of cycle weights remains the same:" + }, + { + "type": "equation", + "bbox": [ + 0.14, + 0.228, + 0.469, + 0.265 + ], + "angle": 0, + "content": "\\[\nw _ {L} ^ {(t + 1)} = \\prod_ {e \\in L \\backslash \\{i j \\}} w _ {e} ^ {(t + 1)} = \\prod_ {e \\in L \\backslash \\{i j \\}} e ^ {- \\beta_ {t} s _ {e} ^ {(t)}}. \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.272, + 0.47, + 0.373 + ], + "angle": 0, + "content": "As a result, \\( d_{L}^{2} = d^{2}(\\pmb{R}_{L},\\pmb{R}_{ij}) = 1 - \\langle \\pmb{R}_{L},\\pmb{R}_{ij}\\rangle /d \\) is linear in both \\( R_{L} \\) and \\( R_{ij} \\). Therefore one can switch the order of \\( d^{2} \\) and the weighted summation, so that the computation of \\( s_{ij}^{(t)} \\) can be vectorized. Indeed, by this linearity and equations (4) and (2), and note that \\( z_{ij}^{(t)} = \\sum_{L\\in N_{ij}^{c}}w_{L}^{(t)} \\), we obtain the following equation:" + }, + { + "type": "equation", + "bbox": [ + 0.113, + 0.38, + 0.469, + 0.551 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} s _ {i j} ^ {(t)} = \\left(\\sum_ {L \\in N _ {i j} ^ {c}} w _ {L} ^ {(t)} d _ {L} ^ {2} / z _ {i j} ^ {(t)}\\right) ^ {1 / 2} \\\\ = \\Big (\\sum_ {L \\in N _ {i j} ^ {c}} w _ {L} ^ {(t)} \\mathcal {D} ^ {2} (\\boldsymbol {R} _ {L}, \\boldsymbol {R} _ {i j}) / z _ {i j} ^ {(t)} \\Big) ^ {1 / 2} \\\\ = \\left(\\mathcal {D} ^ {2} \\Big (\\sum_ {L \\in N _ {i j} ^ {c}} w _ {L} ^ {(t)} \\boldsymbol {R} _ {L}, \\boldsymbol {R} _ {i j} \\Big) / z _ {i j} ^ {(t)}\\right) ^ {1 / 2} \\\\ = \\left(1 - \\left\\langle \\sum_ {L \\in N _ {i j} ^ {c}} w _ {L} ^ {(t)} \\boldsymbol {R} _ {L}, \\boldsymbol {R} _ {i j} \\right\\rangle / d \\sum_ {L \\in N _ {i j} ^ {c}} w _ {L} ^ {(t)}\\right) ^ {1 / 2} \\tag {8} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.558, + 0.469, + 0.62 + ], + "angle": 0, + "content": "Equation (8) can be vectorized using the trick of matrix power if we allow repeated nodes for each cycle. That is, one can stack the \\(s_{ij}^{(t)}\\)'s and \\(w_{ij}^{(t)}\\)'s into matrices \\(\\mathbf{S}^{(t)}\\) and \\(\\mathbf{W}^{(t)}\\), and vectorize (8) as" + }, + { + "type": "equation", + "bbox": [ + 0.131, + 0.639, + 0.469, + 0.7 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\boldsymbol {S} ^ {(t)} = \\left(\\boldsymbol {A} - \\left\\langle \\left(\\boldsymbol {W} ^ {(t)} \\otimes 1 _ {d} \\odot \\boldsymbol {R}\\right) ^ {c - 1}, \\boldsymbol {R} \\right\\rangle_ {\\text {b l o c k}} \\right. \\\\ \\left. \\odot d \\left(\\boldsymbol {W} ^ {(t)}\\right) ^ {c - 1}\\right) ^ {\\odot 1 / 2}. \\tag {9} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.709, + 0.469, + 0.784 + ], + "angle": 0, + "content": "Indeed, by (7) and the definition of \\( \\pmb{R}_L \\), \\( \\sum_{L \\in C_{ij}^c} w_L^{(t)} \\pmb{R}_L \\) is the \\( ij \\)-th block of \\( (\\pmb{W}^{(t)} \\otimes 1_d \\odot \\pmb{R})^{c-1} \\), and \\( \\sum_{L \\in C_{ij}^c} w_L^{(t)} \\) is the \\( ij \\)-th element of \\( \\pmb{W}^{(t)c-1} \\), where \\( C_{ij}^c \\) is the set of \\( c \\)-cycles containing \\( ij \\) with possibly repeated nodes." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.784, + 0.47, + 0.905 + ], + "angle": 0, + "content": "In the case of utilizing only simple cycles, (8) and (9) are not equivalent and we need to correct (9) to remove the cycles with repeated nodes, so that only simple cycles in \\( N_{ij}^{c} \\) remain. Let \\( g_{c}(\\boldsymbol{W},\\boldsymbol{R}) \\) be the matrix valued function where \\( g_{c}(\\boldsymbol{W},\\boldsymbol{R})(i,j) = \\sum_{L\\in N_{ij}^{c}}(\\prod_{e\\in L\\setminus \\{ij\\}}w_{e}^{(t)})\\boldsymbol{R}_{L} \\). Let \\( f_{c}(\\boldsymbol{W}) \\) be the matrix valued function where \\( f_{c}(\\boldsymbol{W})(i,j) = \\sum_{L\\in N_{ij}^{c}}(\\prod_{e\\in L\\setminus \\{ij\\}}w_{e}^{(t)}) \\). The following result holds:" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.892, + 0.122 + ], + "angle": 0, + "content": "Proposition 3.1. The update rule of LongSync (8) is equivalent to the following matrix equations:" + }, + { + "type": "equation", + "bbox": [ + 0.536, + 0.132, + 0.892, + 0.163 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {S} ^ {(t)} = \\left(\\boldsymbol {A} - \\left\\langle h _ {c} (\\boldsymbol {W} ^ {(t)}, \\boldsymbol {R}), \\boldsymbol {R} \\right\\rangle_ {\\text {b l o c k}} \\odot \\boldsymbol {A}\\right) ^ {\\odot 1 / 2}, \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.173, + 0.759, + 0.19 + ], + "angle": 0, + "content": "where \\(\\pmb{W}^{(t + 1)} = \\pmb{A}\\odot \\exp (-\\beta_t\\pmb{S}^{(t)})\\) and" + }, + { + "type": "equation", + "bbox": [ + 0.537, + 0.199, + 0.854, + 0.218 + ], + "angle": 0, + "content": "\\[\nh _ {c} \\left(\\boldsymbol {W} ^ {(t)}, \\boldsymbol {R}\\right) := g _ {c} \\left(\\boldsymbol {W} ^ {(t)}, \\boldsymbol {R}\\right) \\oslash \\left(d \\cdot f _ {c} \\left(\\boldsymbol {W} ^ {(t)} \\otimes 1 _ {d}\\right)\\right).\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.229, + 0.867, + 0.245 + ], + "angle": 0, + "content": "Here exp denotes the elementwise exponential function." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.254, + 0.892, + 0.285 + ], + "angle": 0, + "content": "We use equation (10) as the update rule of LongSync and propose the vectorized LongSync algorithm in algorithm 1." + }, + { + "type": "title", + "bbox": [ + 0.502, + 0.299, + 0.664, + 0.315 + ], + "angle": 0, + "content": "Algorithm 1 (LongSync)" + }, + { + "type": "text", + "bbox": [ + 0.501, + 0.32, + 0.892, + 0.365 + ], + "angle": 0, + "content": "Input: pairwise rotation matrix \\( \\mathbf{R} \\in \\mathbb{R}^{dn \\times dn} \\), adjacency matrix \\( \\mathbf{A} \\in [0,1]^{n \\times n} \\), cycle length \\( c \\), positive parameters \\( \\{\\beta_t\\}_{t \\geq 1} \\), time step \\( T \\)" + }, + { + "type": "equation", + "bbox": [ + 0.519, + 0.365, + 0.619, + 0.38 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {W} ^ {(0)} (i, j) \\leftarrow \\boldsymbol {A}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.382, + 0.614, + 0.394 + ], + "angle": 0, + "content": "for \\(t = 0:T\\) do" + }, + { + "type": "equation", + "bbox": [ + 0.543, + 0.404, + 0.891, + 0.434 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {S} ^ {(t)} \\leftarrow \\left(\\boldsymbol {A} - \\left\\langle h _ {c} (\\boldsymbol {W} ^ {(t)}, \\boldsymbol {R}), \\boldsymbol {R} \\right\\rangle_ {\\text {b l o c k}} \\odot \\boldsymbol {A}\\right) ^ {\\odot 1 / 2} \\tag {11}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.545, + 0.436, + 0.891, + 0.454 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {W} ^ {(t + 1)} \\leftarrow \\boldsymbol {A} \\odot \\exp (- \\beta_ {t} \\boldsymbol {S} ^ {(t)}) \\tag {12}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.518, + 0.467, + 0.57, + 0.479 + ], + "angle": 0, + "content": "end for" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.48, + 0.851, + 0.496 + ], + "angle": 0, + "content": "Output: edge weights \\( \\mathbf{W}^{(T + 1)} \\), corruption levels \\( \\mathbf{S}^{(T)} \\)" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.517, + 0.892, + 0.697 + ], + "angle": 0, + "content": "We claim that \\( g_{c} \\) and \\( f_{c} \\) can be computed with a sequence of matrix operations, thus greatly reducing the time and space consumption of LongSync compared to its original form. For \\( c \\leq 6 \\), the time complexity of computing \\( g_{c} \\) and \\( f_{c} \\) is \\( O(r(dn)) \\), where \\( r(n) \\) is the complexity for multiplying two \\( n \\times n \\) matrices; for \\( c \\geq 7 \\) the time complexity is at most \\( O(n^{[(c + 3) / 2]}) \\). This claim is proved in the supplementary material. We list the formula for \\( g_{c} \\) and \\( f_{c} \\) for \\( c = 3,4,5,6 \\) inspired by [35, 46]. The formula for \\( c = 6 \\) is moved to the supplementary material due to the space limit. For \\( c \\geq 7 \\) the formula becomes extremely complicated. We remark that in practice, the cycles of length greater than 6 are often not used." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.698, + 0.892, + 0.789 + ], + "angle": 0, + "content": "We finally remark that although Algorithm 1 only utilizes cycles of a fixed length, one can easily generalize it to incorporate cycles of different lengths. Indeed, the equation (10) could use a convex combination of \\( h_c \\)'s that corresponds to different values of \\( c \\). That is, for a preselected set of cycle lengths \\( C \\), the equation (11) in Algorithm 1 is replaced by" + }, + { + "type": "equation", + "bbox": [ + 0.511, + 0.799, + 0.891, + 0.845 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {S} ^ {(t)} = \\left(\\boldsymbol {A} - \\left\\langle \\sum_ {c \\in C} \\lambda_ {c} h _ {c} \\left(\\boldsymbol {W} ^ {(t)}, \\boldsymbol {R}\\right), \\boldsymbol {R} \\right\\rangle_ {\\text {b l o c k}} \\odot \\boldsymbol {A}\\right) ^ {\\odot 1 / 2} \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.892, + 0.901 + ], + "angle": 0, + "content": "where the coefficients \\(\\lambda_{c}\\) satisfies \\(\\sum_{c\\in C}\\lambda_c = 1\\) to ensure a convex combination. Here each \\(\\lambda_{c}\\) is user-specified to reflect the importance of the cycles of length \\(c\\). However, the optimal" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.515, + 0.957 + ], + "angle": 0, + "content": "5263" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.12 + ], + "angle": 0, + "content": "choice of these parameters under certain statistical model remains unclear." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.122, + 0.468, + 0.196 + ], + "angle": 0, + "content": "For simplicity, in the experiments we only use a fixed cycle length to avoid choosing \\(\\lambda_{c}\\). We have observed that such simple choice still yields satisfying accuracy in camera orientation estimation on both synthetic and real data. We refer the readers to §5 and 6 for more details." + }, + { + "type": "table", + "bbox": [ + 0.097, + 0.206, + 0.451, + 0.364 + ], + "angle": 0, + "content": "
cFormula of fc(W)Formula of gc(W,R)
3W2P2
4W3-d(W2)WP3-dblock(P2)P
-Wd(W2)+W⊙3-Pdblock(P2)+P⊙3
5W4-d(W3)WP4-dblock(P3)P
-d(W2)W2-dblock(P2)P2
-W2d(W2)-P2dblock(P2)
-Wd(W2)W-Pdblock(P2)P
+3W⊙2W2+3P⊙2P2
+WW⊙3+W⊙3W+PP⊙3+P⊙3P
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.373, + 0.47, + 0.429 + ], + "angle": 0, + "content": "Table 1. Formulas for \\( f_{c} \\) and \\( g_{c} \\). Here we let \\( P = (W \\otimes 1_d) \\odot R \\) for shorter notation. \\( \\mathrm{d}(X) \\) returns the diagonal of matrix \\( X \\), \\( \\mathrm{d}_{\\mathrm{block}}(X) \\) returns the diagonal block matrix from the \\( d \\times d \\) diagonal blocks of matrix \\( X \\)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.452, + 0.323, + 0.469 + ], + "angle": 0, + "content": "3.2. Computational Complexity" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.475, + 0.47, + 0.793 + ], + "angle": 0, + "content": "We derive the space and time complexity for LongSync, and demonstrate its advantages over CEMP. The initialization step involves setting the weights of all edges to 1, which takes time \\(O(|E|)\\) and space \\(O(n^{2})\\). For each iteration, LongSync updates the matrices \\(\\pmb{S}^{(t)}\\) and \\(\\pmb{W}^{(t)}\\) with equations (11) and (12), respectively. Computing \\(\\pmb{W}^{(t + 1)}\\) involves two matrix subtractions, one scalar-matrix multiplication and one element-wise matrix exponential operation on \\(\\pmb{S}^{(t)} \\in \\mathbb{R}^{n \\times n}\\). Therefore the update step (12) takes at most \\(O(n^{2})\\) time and space. Equation (11), on the other hand, involves a sequence of matrix operations on \\(\\pmb{P}^{(t)} = (\\pmb{W}^{(t)} \\otimes 1_{d}) \\odot \\pmb{R} \\in \\mathbb{R}^{dn \\times dn}\\) and \\(\\pmb{W}^{(t)} \\in \\mathbb{R}^{n \\times n}\\), including matrix multiplications, element-wise multiplications and diagonal block selections. Computing \\(\\pmb{P}^{(t)}\\) takes \\(O(d^{2}n^{2})\\) memory and \\(O(d^{2}n^{2})\\) time. The matrix operations on \\(\\pmb{P}^{(t)}\\) take \\(O(K_{c}d^{3}n^{3})\\) time and \\(O(d^{2}n^{2})\\) space, and the matrix operations on \\(\\pmb{W}^{(t)}\\) take \\(O(K_{c}n^{3})\\) time and \\(O(n^{2})\\) space, where \\(K_{c}\\) is the number of terms in the equation for \\(f_{c}\\) and \\(g_{c}\\). Therefore, each iteration over \\(t\\) takes \\(O(K_{c}d^{3}n^{3})\\) time and \\(O(d^{2}n^{2})\\) space. To sum up, for LongSync, the time complexity is \\(O(K_{c}d^{3}n^{3})\\) and the space complexity is \\(O(d^{2}n^{2})\\). For \\(c = 3,4,5,6\\), the number \\(K_{c}\\) is equal to 1,3,9,32." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.793, + 0.471, + 0.904 + ], + "angle": 0, + "content": "In comparison, we consider the initialization step of CEMP. For each edge \\( ij \\in E \\) and \\( L \\in N_{ij}^{c} \\), initializing CEMP involves computing and storing all the cycle inconsistency measures \\( d_{L} \\) using equation (2). For each \\( L \\in N_{ij}^{c} \\), computing \\( d_{L} \\) involves multiplying \\( c \\) rotations, which takes \\( O(cd^3) \\) time and \\( O(d^2) \\) space. This step is repeated for each \\( ij \\in E \\) and \\( L \\in N_{ij}^{c} \\), therefore the total time complexity is \\( O(cd^3\\sum_{ij \\in E}|N_{ij}^{c}|) \\) and" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.091, + 0.892, + 0.2 + ], + "angle": 0, + "content": "the total space complexity is \\( O(d^{2}\\sum_{ij\\in E}|N_{ij}^{c}|) \\). Since for each edge there are \\( (n - 2)(n - 3)\\dots (n - c + 1) = O(n^{c - 2}) \\) cycle candidates, we know that \\( |N_{ij}^{c}|\\sim O(n^{c - 2}) \\) for each \\( ij\\in E \\) in the worst case scenario of a dense graph. Therefore the initialization of CEMP takes \\( O(cd^3 n^{c - 2}|E|) \\) time and \\( O(d^{2}n^{c - 2}|E|) \\) space in the worst case. Given \\( c\\geq 4 \\) and \\( |E|\\sim n^2 \\), CEMP requires much more time and space than LongSync." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.214, + 0.835, + 0.232 + ], + "angle": 0, + "content": "4. Theory for Uniform Corruption Model" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.24, + 0.892, + 0.348 + ], + "angle": 0, + "content": "In this section, we present the exact recovery guarantee of LongSync under the uniform corruption model (UCM). UCM is a popular probabilistic model that is widely adopted for synthetic experiments of many previous works on group synchronization [8, 25, 28, 29, 32, 42]. The model \\(\\mathrm{UCM}(n,p,q_g)\\) assumes that \\(G\\) is an Erdős-Rényi graph with edge connection probability \\(p\\). For each edge \\(ij \\in E\\), \\(R_{ij}\\) is generated independently as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.576, + 0.36, + 0.813, + 0.4 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {R} _ {i j} = \\left\\{ \\begin{array}{l l} \\boldsymbol {R} _ {i j} ^ {*} & \\text {w . p .} q _ {g}; \\\\ \\tilde {\\boldsymbol {R}} _ {i j} \\sim \\operatorname {H a a r} (\\mathcal {G}) & \\text {w . p .} 1 - q _ {g}. \\end{array} \\right.\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.407, + 0.892, + 0.513 + ], + "angle": 0, + "content": "We also developed an exact recovery theory for a general model of adversarial corruption, which we include in section C.1 of the supplementary material. An informal version of our main result for UCM is stated in Theorem 4.1. Although the application of this paper is focused on rotation synchronization, the following theory for UCM is valid for any compact group \\(\\mathcal{G}\\), as explained in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.524, + 0.892, + 0.573 + ], + "angle": 0, + "content": "Theorem 4.1. Let \\(0 < r < 1\\), \\(0 < q < 1\\), \\(0 < p \\leq 1\\), \\(\\mathcal{G} = SO(3)\\). Assume LongSync is applied with cycles of length \\(c\\), \\(n / \\log n \\sim p^{-(c-1) / (c-2-\\epsilon)} q_g^{-7(c-1) / 3(c-2)}\\) for some \\(\\epsilon > 0\\) and" + }, + { + "type": "equation", + "bbox": [ + 0.608, + 0.586, + 0.782, + 0.603 + ], + "angle": 0, + "content": "\\[\n1 / \\beta_ {t + 1} = r / \\beta_ {t} f o r a l l t \\geq 1.\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.615, + 0.892, + 0.651 + ], + "angle": 0, + "content": "Then with appropriate choices of \\(\\beta_0, \\beta_1, r\\), and high probability, \\(\\max_{ij \\in E} |s_{ij}^* - s_{ij}^{(t)}| \\leq \\frac{1}{2c\\beta_t}\\) for all \\(t \\geq 1\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.659, + 0.892, + 0.809 + ], + "angle": 0, + "content": "The major difficulty of proving Theorem 4.1 is the dependence in the cycle inconsistency measures for cycles in \\( N_{ij}^{c} \\) when \\( c \\geq 4 \\). Unlike the 3-cycle case, the cycle inconsistency measure of a 4-cycle \\( L_{1} = (ik_{1}, k_{1}k_{2}, k_{2}j) \\) is correlated with that of \\( L_{2} = (ik_{1}, k_{1}k_{3}, k_{3}j) \\) under UCM. Therefore the key concentration inequalities for the proof cannot be concluded from the standard Chernoff bounds. To overcome this theoretical obstacle, we have integrated various mathematical techniques from [3, 7, 22-24, 48] to prove the theorem, whose details are included in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.811, + 0.892, + 0.9 + ], + "angle": 0, + "content": "Theorem 4.1 provides an upper bound of the sample complexity (the required graph size \\( n \\)) of LongSync for exact recovery of the ground truth solutions. This sample complexity is the closest to the information theoretic bound among all existing rotation synchronization methods. The comparison with previous works is summarized in Table 2." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5264" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.081, + 0.089, + 0.468, + 0.254 + ], + "angle": 0, + "content": "
ReferenceSample Complexity
[25] for CEMPO(p-2qg-28/3)
[29] for ReSyncO(p-2qg-7)
Ours for CEMPO(p-2-εqg-14/3)
Ours for LongSync (c=3)O(p-2-εqg-14/3)
Ours for LongSync (c=4)O(p-1.5-εqg-3.5)
Ours for LongSync (any c)O(p-7(c-1/c-2-ε) qg-3(c-2))
Ours for LongSync (c→∞)O(p-1-εqg-7/3)
Information Theoretic Bound [6]O(p-1qg-2)
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.264, + 0.471, + 0.307 + ], + "angle": 0, + "content": "Table 2. Comparison of the sample complexity requirement. Lower absolute values of the powers on \\( p, q_{g} \\) indicate better results. \\( \\epsilon \\) is an arbitrarily small positive real number." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.33, + 0.317, + 0.349 + ], + "angle": 0, + "content": "5. Synthetic Data Experiment" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.356, + 0.47, + 0.478 + ], + "angle": 0, + "content": "We test LongSync on synthetic datasets generated with Uniform Corruption Model (UCM) and Uniform Bipartite Corruption Model (UBCM) respectively described in §5.1 and 5.2. For both models with their corresponding viewing graphs \\( G = ([n], E) \\), we sample the ground truth absolute rotation matrices \\( \\{R_i^*\\}_{i \\in [n]} \\) independently from the Haar measure on SO(3), and we generate the observed relative rotations \\( \\{R_{ij}\\}_{ij \\in E} \\) independently as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.137, + 0.488, + 0.407, + 0.528 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {R} _ {i j} = \\left\\{ \\begin{array}{l l} \\boldsymbol {R} _ {i} ^ {*} \\boldsymbol {R} _ {j} ^ {*} & \\text {w . p .} q _ {g}; \\\\ \\tilde {\\boldsymbol {R}} _ {i j} \\sim \\operatorname {H a a r} (S O (3)) & \\text {w . p .} 1 - q _ {g}. \\end{array} \\right.\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.538, + 0.47, + 0.75 + ], + "angle": 0, + "content": "We use LongSync with cycle length \\(c\\), \\(\\beta_{t} = \\min (2^{t},20)\\) and \\(T = 10\\) and record the edge weights. For UCM we set \\(c = 4,5\\) and for UBCM we only use \\(c = 4\\) since no 5-cycles exist. For our method, we first build a weighted graph whose edge weights are estimated by LongSync. We then extract a maximum spanning tree (MST) of the resulting weighted graph. The resulting spanning tree is expected to be the cleanest possible spanning tree. To initialize our solution of absolute rotations, we first fix \\(\\pmb{R}_{1}\\) as the identity rotation, and find the rest of \\(\\pmb{R}_{i}\\)'s by consecutively multiplying the relative rotations along the spanning tree using the formula \\(\\pmb{R}_{i} = \\pmb{R}_{ij}\\pmb{R}_{j}\\). To refine our initialized solution, we apply IRLS with Geman-McClure [4] loss functions to minimize \\(\\sum_{ij\\in E}\\rho_{GM}(d_{\\angle}(\\pmb{R}_{ij},\\pmb{R}_i\\pmb{R}_j^T))\\), where \\(d_{\\angle}\\) is the geodesic distance in \\(SO(3)\\). We refer to this method as LongSync+IRLS." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.468, + 0.856 + ], + "angle": 0, + "content": "To demonstrate the advantages of utilizing longer cycle information, we compare our method with IRLS initialized by other two different spanning trees. The first one is the random spanning tree, which uses no cycle information. The other one the MST extracted from the CEMP-estimated weights. Note that CEMP only uses 3-cycle information. We refer to these methods as IRLS and CEMP+IRLS respectively." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.856, + 0.47, + 0.902 + ], + "angle": 0, + "content": "Since the solution of absolute rotations is determined up to a global rotation, we align our estimated rotation \\(\\{\\hat{R}_i\\}\\) with \\(\\{R_i^*\\}\\) by \\(R_{\\mathrm{align}}\\) that minimizes the \\(\\ell_1\\) rotation alignment error" + }, + { + "type": "image", + "bbox": [ + 0.578, + 0.098, + 0.808, + 0.233 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.578, + 0.239, + 0.808, + 0.374 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.385, + 0.892, + 0.456 + ], + "angle": 0, + "content": "Figure 1. Average errors for IRLS, CEMP+IRLS and LongSync +IRLS with \\( c = 4,5 \\), using the uniform corruption (top) and uniform bipartite corruption (bottom) models. The mean errors are measured in degrees. LongSync4 and LongSync5 refer to LongSync with 4 and 5 cycles, respectively." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.483, + 0.892, + 0.522 + ], + "angle": 0, + "content": "\\(\\sum_{i\\in [n]}\\| \\hat{\\pmb{R}}_i\\pmb{R}_{\\mathrm{align}} - \\pmb{R}_i^*\\| _F\\) . We report the mean estimation error in degrees: \\(180\\cdot \\sum_{i\\in [n]}d_{\\angle}(\\hat{\\pmb{R}}_i\\pmb{R}_{\\mathrm{align}},\\pmb{R}_i^*) / n.\\)" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.533, + 0.749, + 0.55 + ], + "angle": 0, + "content": "5.1. Uniform Corruption Model" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.558, + 0.892, + 0.619 + ], + "angle": 0, + "content": "For \\(\\mathrm{UCM}(n,p,q_g)\\), we take \\(n = 200\\) and \\(p = 1\\) and corruption probability \\(q = 1 - q_{g}\\) ranges from 0.86 to 0.92. We report the average mean estimation error from 20 trials of the uniform corruption model in the top panel of Figure 1." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.62, + 0.892, + 0.711 + ], + "angle": 0, + "content": "We note that LongSync uniformly improves IRLS, and the mean error of LongSync decreases as the cycle length increases. When \\( q > 0.86 \\), the expected number of clean 3-cycles for each edge is less than 4, and therefore longer cycles are helpful. The numerical result aligns with our theory that using longer cycles may tolerate higher corruption with fixed graph size \\( n \\)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.724, + 0.821, + 0.74 + ], + "angle": 0, + "content": "5.2. Uniform Bipartite Corruption Model" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.749, + 0.892, + 0.854 + ], + "angle": 0, + "content": "For UBCM, we first generate the graph and relative rotations by \\(\\mathrm{UCM}(n,p,q_g)\\) with \\(n = 200\\), \\(p = 1\\) and \\(q = 1 - q_{g}\\) ranging from 0.8 to 0.85. Then we split the nodes into two clusters of equal size and remove the intra-cluster edges for both clusters. The resulting graph is bipartite, where only cycles of even lengths exist. We report the mean estimation error from 20 trials in the bottom panel of Figure 1." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.856, + 0.892, + 0.901 + ], + "angle": 0, + "content": "We observe that LongSync with 4-cycles almost exactly recovers the rotations, while for other algorithms the rotation estimates are not even close to the ground truth." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.515, + 0.957 + ], + "angle": 0, + "content": "5265" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.077, + 0.09, + 0.281, + 0.108 + ], + "angle": 0, + "content": "6. Real Data Experiment" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.119, + 0.473, + 0.285 + ], + "angle": 0, + "content": "We test distributed synchronization with LongSync on the PhotoTourism dataset [50] to demonstrate its advantages in accuracy and speed over other baselines. PhotoTourism is a large scale dataset consisting of 15 sets of images taken for 3D reconstruction. The smallest dataset consists of 247 cameras, and the largest dataset consists of 5433 cameras. The input graph and initial pairwise rotation estimates are provided in the dataset. In the following, we first explain the common steps for distributed synchronization, and our improvement using LongSync. We then describe our graph processing method for filtering bad nodes and edges, which also is applicable to other baseline methods." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.286, + 0.331, + 0.3 + ], + "angle": 0, + "content": "Steps in Distributed Synchronization:" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.303, + 0.47, + 0.379 + ], + "angle": 0, + "content": "1. Graph partitioning. The first step involves partitioning the graph \\( G = ([n], E) \\) into \\( K \\) clusters \\( G_{i} = (V_{i}, E_{i}), i \\in [K] \\). In this paper we apply spectral clustering algorithm [37] on the adjacency matrix \\( G \\), where \\( K = 0.6\\sqrt{np} \\) and \\( p = 2|E| / (n(n - 1)) \\)." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.38, + 0.471, + 0.487 + ], + "angle": 0, + "content": "2. Synchronization within clusters. Run standard synchronization solvers for each cluster. In this work, we use the current state-of-the-art method MPLS [39]. Note that for each camera \\( p \\) in cluster \\( k \\), one can only estimate the true rotation \\( R_{p}^{*} \\) up to a global rotation \\( R_{k} \\). Namely, one only obtains \\( \\hat{R}^p \\approx R_p^* R_k^{-1} \\) where \\( R_{k} \\) is unknown and is the same for all cameras in cluster \\( k \\)." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.488, + 0.47, + 0.672 + ], + "angle": 0, + "content": "3. Estimation of inter-cluster rotations. To find \\( \\pmb{R}_{p}^{*} \\) of all cameras, one needs to solve \\( \\pmb{R}_{k} \\) for all clusters. Namely, one needs to rotate and stitch the solutions of all clusters so that they are in the same reference frame. To do this, it is common to first estimate the inter-clsuter rotations \\( \\pmb{R}_{kl} := \\pmb{R}_{k}\\pmb{R}_{l}^{-1} \\) between pairs of clusters \\( k,l \\), and then synchronize these relative rotations. To estimate each \\( \\pmb{R}_{kl} \\), we note that \\( \\pmb{R}_{kl} = \\pmb{R}_{p}^{*-1}\\pmb{R}_{pq}^{*}\\pmb{R}_{q}^{*} \\) for each \\( p\\in V_k \\) and \\( q\\in V_l \\). Therefore, one can use the rotations in the set \\( S_{kl} := \\{\\hat{\\pmb{R}}_p^{-1}\\pmb{R}_{pq}\\hat{\\pmb{R}}_q\\}_{p\\in V_k,q\\in V_l} \\) to approximate \\( \\pmb{R}_{kl} \\). We remark that this step is crucial to the overall performance of distributed methods, and we compare the following methods for solving \\( \\pmb{R}_{kl} \\):" + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.672, + 0.469, + 0.732 + ], + "angle": 0, + "content": "- MultSync [9]: Run synchronization on a multi-graph where each edge \\(kl\\) is assigned a set of relative rotations \\(\\{\\hat{R}_p^{-1}R_{pq}\\hat{R}_q\\}_{p\\in V_k,q\\in V_l}\\). This combines the step 3 and 4 in a unified least squares formulation." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.732, + 0.469, + 0.793 + ], + "angle": 0, + "content": "- Edge averaging using IRLS: We initialize \\(\\hat{R}_{kl}\\) with the quaternion \\(\\ell_2\\) mean of the set \\(S_{kl}\\) and refine it using \\(\\ell_1\\)-rotation averaging [19]. We refer to this method as IRLS in our comparison." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.793, + 0.469, + 0.899 + ], + "angle": 0, + "content": "- Edge averaging using LongSync: We first perform LongSync with 4-cycles to estimate the weights of these inter-cluster edges (there are no 3-cycles for a bipartite graph). We next initialize \\(\\hat{R}_{kl}\\) as the quaternion weighted \\(\\ell_2\\) mean of \\(S_{kl}\\), using the edge weights from LongSync by their LongSync weights. Lastly, we refine the solution using [19]." + }, + { + "type": "list", + "bbox": [ + 0.097, + 0.672, + 0.469, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.892, + 0.152 + ], + "angle": 0, + "content": "4. Synchronization of inter-cluster rotations. This step is skipped for MultiSync. For other methods described in step 3, we find \\( \\pmb{R}_k \\) (up to a rotation) for each cluster \\( k \\) from the estimated \\( \\{R_{kl}\\}_{k,l\\in [K]} \\) by MPLS." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.153, + 0.894, + 0.183 + ], + "angle": 0, + "content": "5. Rotation merging. Finally, for each camera \\( p \\) in cluster \\( k \\), the rotation estimate of \\( p \\) is given by \\( R_p^{\\mathrm{final}} = \\hat{R}_p R_k^{-1} \\)." + }, + { + "type": "list", + "bbox": [ + 0.499, + 0.092, + 0.894, + 0.183 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.183, + 0.892, + 0.212 + ], + "angle": 0, + "content": "Next, we introduce our graph processing method to further boost the performance of all tested methods." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.213, + 0.785, + 0.227 + ], + "angle": 0, + "content": "Extra Improvement by Graph Processing:" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.228, + 0.892, + 0.287 + ], + "angle": 0, + "content": "- Spectral clustering with Jaccard Index. For step 1, we use the Jaccard index matrix as the similarity matrix for spectral clustering, instead of the adjacency matrix. The \\( n \\times n \\) Jaccard index matrix \\( A_J \\) is defined as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.606, + 0.295, + 0.892, + 0.337 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {A} _ {J} (i, j) = \\left\\{ \\begin{array}{l l} 0 & i j \\notin E \\\\ \\frac {\\left| N _ {i} \\cap N _ {j} \\right|}{\\left| N _ {i} \\cup N _ {j} \\right|} & i j \\in E \\end{array} \\right. \\tag {14}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.342, + 0.892, + 0.418 + ], + "angle": 0, + "content": "where \\(N_{i}\\) and \\(N_{j}\\) denote the sets of neighboring nodes of node \\(i\\) and \\(j\\), respectively. In this way, \\(\\mathbf{A}_{J}(i,j)\\) is higher for the pair \\(ij\\) contained in many 3-cycles, which is a more robust and nicely scaled statistics \\((\\in [0,1])\\) for measuring the local graph density around edge \\(ij\\)." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.419, + 0.893, + 0.659 + ], + "angle": 0, + "content": "- Refinement of intra-cluster edges and nodes. For step 2, after the MPLS step, we perform CEMP with 3-cycles to estimate the corruption level of the intra-cluster edges for each cluster. We remove a camera if the number of neighboring 'good' edges, i.e. the edges with corruption level less than 0.1, is less than 4. The numbers 4 and 0.1 are chosen to balance the number of remaining cameras and the quality of intra-cluster rotation estimates. In order to eliminate the sparsely connected components inside the cluster, we use the Matlab built-in hierarchical spectral clustering function on the remaining cameras with the 'cutoff' and 'depth' parameters as 2 and 4, and we keep the largest component. The absolute rotations for the remaining cameras are estimated by MPLS. We remark that one could replace CEMP by LongSync with 3-cycles. However, we have not observed significant difference in the performance." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.66, + 0.894, + 0.856 + ], + "angle": 0, + "content": "We respectively name MultiSync and IRLS with our new graph processing method as MultiSync(New) and IRLS(new). \"LongSync\" in our experiment refers to the full version of our algorithm: use LongSync weights for edge averaging in Step 3, with the graph processing step. We also compare with MPLS on the whole dataset, since it is a state-of-the-art non-distributed method, but we note that MPLS is significantly slower than all distributed methods. We report median error \\(180 \\cdot\\) median(\\(\\{d_{\\angle}(\\hat{R}_i R_{\\text{align}}, R_i^*)\\}_{i \\in [n]}\\) of the tested methods on 14 datasets in Figure 2. We exclude the result of Gendarmenmarkt since all methods return large estimation errors in the figure. The full results, including that of mean error are included in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.892, + 0.901 + ], + "angle": 0, + "content": "In Figure 3, for each distributed method, we report the ratio (in percentage) between its total runtime on all datasets and that of the non-distributed MPLS. Namely, we compute" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5266" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.088, + 0.089, + 0.894, + 0.273 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.277, + 0.283, + 0.694, + 0.298 + ], + "angle": 0, + "content": "Figure 2. Median error for rotations for each dataset measured in degrees." + }, + { + "type": "image", + "bbox": [ + 0.102, + 0.321, + 0.446, + 0.537 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.547, + 0.47, + 0.577 + ], + "angle": 0, + "content": "Figure 3. Runtime ratio and average median and mean error gaps between the distributed methods and MPLS on the entire graph." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.609, + 0.47, + 0.732 + ], + "angle": 0, + "content": "\\(\\sum_{d \\in D} t_{\\mathrm{dist},d} / \\sum_{d \\in D} t_{\\mathrm{MPLS},d}\\), where \\(D\\) is the set of 15 datasets, and each \\(t_{\\mathrm{dist},d}\\) and \\(t_{\\mathrm{MPLS},d}\\) is respectively the runtime of the distributed method and MPLS on data \\(d\\). In the same figure, we present the mean/median error gap between each distributed method and MPLS. The mean and median error gap is respectively defined as \\((\\bar{e}_{\\mathrm{dist}} - \\bar{e}_{\\mathrm{MPLS}}) / \\bar{e}_{\\mathrm{dist}}\\) and \\((\\hat{e}_{\\mathrm{dist}} - \\hat{e}_{\\mathrm{MPLS}}) / \\hat{e}_{\\mathrm{dist}}\\), where \\(\\bar{e}\\) and \\(\\hat{e}\\) respectively denote the mean and median error over all cameras." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.735, + 0.471, + 0.903 + ], + "angle": 0, + "content": "From Figure 2 and 3, our method outperforms other distributed methods on 13 out of 15 datasets. The most significant improvement in mean error and median error are respectively \\(28.6\\%\\) and \\(46.4\\%\\) (in Notre Dame) compared to the best performing method between IRLS(new) and MultiSync(new). The improvement is even more significant when comparing to the original version of these baseline methods without our graph processing method. The only two datasets without improvement are Gendarmenmarkt and Union Square. Our method is comparable to others on Union Square, and all methods return large errors on Gendarmenmarkt due to many repetitive patterns in its 3D scene." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.324, + 0.895, + 0.462 + ], + "angle": 0, + "content": "The average mean and median error gap between our method and full MPLS are respectively \\(38.3\\%\\) and \\(18.4\\%\\). Compared to the best performing method among others, our method reduces the average median error gap by \\(40.8\\%\\), and the average mean error gap by \\(9.6\\%\\). In terms of runtime, our method is uniformly faster than MultiSync and it is scalable on the largest dataset, taking less than \\(6\\%\\) of the total runtime of full MPLS. In conclusion, our method significantly improves the result of distributed rotation synchronization without compromising runtime." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.468, + 0.895, + 0.558 + ], + "angle": 0, + "content": "In the supplementary material, we further demonstrate the improvement by our new graph processing method, which significantly improves the results of LongSync (without extra graph processing) in 14 of the 15 datasets. On these 14 datasets, the average reduction on mean error is \\(59.2\\%\\) and the average reduction on median error is \\(28.5\\%\\)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.596, + 0.614, + 0.612 + ], + "angle": 0, + "content": "7. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.629, + 0.895, + 0.903 + ], + "angle": 0, + "content": "We propose LongSync, a robust and efficient algorithm for group synchronization. It modifies and vectorizes CEMP which enables efficient computation when using longer cycles. The theory we developed for LongSync is the strongest among all other existing results under UCM. Experiment shows that LongSync, together with our improved graph preprocessing method, achieves superior accuracy for distributed synchronization on large real datasets with competitive runtime. However, our method also has some limitations. First of all, in theory there is still a small gap of sample complexity from our method to the information theoretic one. Filling this gap is an open problem, which requires new tools and possibly more sophisticated analysis. Second, our graph preprocessing method is quite heuristic, and an automatic way of choosing parameters is needed. Our work also opens a door for some important future directions, including distributed partial permutation synchronization for multi-image matching, angular synchronization for Cryo-EM and Jigsaw Puzzles, and analysis of their algorithms." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5267" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.079, + 0.09, + 0.17, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.17 + ], + "angle": 0, + "content": "[1] Federica Arrigoni, Beatrice Rossi, Pasqualina Fragneto, and Andrea Fusiello. Robust synchronization in so(3) and se(3) via low-rank and sparse matrix decomposition. Computer Vision and Image Understanding, 174:95-113, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.17, + 0.471, + 0.21 + ], + "angle": 0, + "content": "[2] A. S. Bandeira, N. Boumal, and A. Singer. Tightness of the maximum likelihood semidefinite relaxation for angular synchronization. arXiv preprint arXiv:1411.3272, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.21, + 0.471, + 0.251 + ], + "angle": 0, + "content": "[3] Olivier Bousquet. A Bennett concentration inequality and its application to suprema of empirical processes. C. R. Math. Acad. Sci. Paris, 334(6):495-500, 2002. 5, 14, 16, 21" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.251, + 0.471, + 0.305 + ], + "angle": 0, + "content": "[4] Avishek Chatterjee and Venu Madhav Govindu. Efficient and robust large-scale rotation averaging. In IEEE International Conference on Computer Vision, ICCV 2013, Sydney, Australia, December 1-8, 2013, pages 521-528, 2013. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.305, + 0.471, + 0.344 + ], + "angle": 0, + "content": "[5] Avishek Chatterjee and Venu Madhav Govindu. Robust relative rotation averaging. IEEE Trans. Pattern Anal. Mach. Intell., 40(4):958-972, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.344, + 0.471, + 0.385 + ], + "angle": 0, + "content": "[6] Yuxin Chen, Changho Suh, and Andrea J. Goldsmith. Information recovery from pairwise measurements. IEEE Trans. Inf. Theory, 62(10):5881-5905, 2016. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.385, + 0.471, + 0.425 + ], + "angle": 0, + "content": "[7] Keresztély Corradi and András Hajnal. On the maximal number of independent circuits in a graph. Acta Mathematica Hungarica, 14(3-4):423-439, 1963. 5, 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.425, + 0.471, + 0.465 + ], + "angle": 0, + "content": "[8] Mihai Cucuringu. Synchronization over Z2 and community detection in signed multiplex networks with constraints. J. Complex Networks, 3(3):469-506, 2015. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.466, + 0.471, + 0.52 + ], + "angle": 0, + "content": "[9] Andrea Porfiri Dal Cin, Luca Magri, Federica Arrigoni, Andrea Fusiello, and Giacomo Boracchi. Synchronization of group-labelled multi-graphs. In 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 6433-6443, 2021. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.52, + 0.471, + 0.615 + ], + "angle": 0, + "content": "[10] Frank Dellaert, David M. Rosen, Jing Wu, Robert E. Mahony, and Luca Carlone. Shonan rotation averaging: Global optimality by surfing \\(\\mathrm{so(p)^n}\\). In Andrea Vedaldi, Horst Bischof, Thomas Brox, and Jan-Michael Frahm, editors, Computer Vision - ECCV 2020 - 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part VI, volume 12351 of Lecture Notes in Computer Science, pages 292-308. Springer, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.615, + 0.471, + 0.642 + ], + "angle": 0, + "content": "[11] Anders Eriksson, Carl Olsson, Fredrik Kahl, and Tat-Jun Chin. Rotation averaging and strong duality, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.642, + 0.471, + 0.71 + ], + "angle": 0, + "content": "[12] Anders P. Eriksson, Carl Olsson, Fredrik Kahl, and Tat-Jun Chin. Rotation averaging and strong duality. In 2018 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2018, Salt Lake City, UT, USA, June 18-22, 2018, pages 127-135. Computer Vision Foundation / IEEE Computer Society, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.71, + 0.471, + 0.763 + ], + "angle": 0, + "content": "[13] Anders P. Eriksson, Carl Olsson, Fredrik Kahl, and Tat-Jun Chin. Rotation averaging with the chordal distance: Global minimizers and strong duality. IEEE Trans. Pattern Anal. Mach. Intell., 43(1):256-268, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.763, + 0.471, + 0.817 + ], + "angle": 0, + "content": "[14] Xiang Gao, Hainan Cui, Menghan Li, Zexiao Xie, and Shuhan Shen. Irav3: Hierarchical incremental rotation averaging on the fly. IEEE Trans. Circuits Syst. Video Technol., 33(4):2001-2006, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.818, + 0.471, + 0.859 + ], + "angle": 0, + "content": "[15] Xiang Gao, Hainan Cui, and Shuhan Shen. Incremental rotation averaging revisited and more: A new rotation averaging benchmark. arXiv preprint, abs/2309.16924, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.859, + 0.471, + 0.899 + ], + "angle": 0, + "content": "[16] Xiang Gao, Lingjie Zhu, Hainan Cui, Zexiao Xie, and Shuhan Shen. IRA++: distributed incremental rotation averaging. IEEE Trans. Circuits Syst. Video Technol., 32(7):4885-4892, 2022. 2" + }, + { + "type": "list", + "bbox": [ + 0.079, + 0.115, + 0.471, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.161 + ], + "angle": 0, + "content": "[17] Thomas Goldstein, Paul Hand, Choongbum Lee, Vladislav Voroninski, and Stefano Soatto. Shapefit and shapekick for robust, scalable structure from motion. In Computer Vision - ECCV 2016 - 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part VII, pages 289-304, 2016. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.161, + 0.894, + 0.216 + ], + "angle": 0, + "content": "[18] V.M. Govindu. Lie-algebraic averaging for globally consistent motion estimation. In Proceedings of the 2004 IEEE Computer Society Conference on Computer Vision and Pattern Recognition, 2004. CVPR 2004., volume 1, pages I-I, 2004. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.216, + 0.894, + 0.269 + ], + "angle": 0, + "content": "[19] R. Hartley, K. Aftab, and J. Trumpf. L1 rotation averaging using the weiszfeld algorithm. In Computer Vision and Pattern Recognition (CVPR), 2011 IEEE Conference on, pages 3041-3048. IEEE, 2011. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.269, + 0.894, + 0.338 + ], + "angle": 0, + "content": "[20] Richard I. Hartley, Khurrum Aftab, and Jochen Trumpf. L1 rotation averaging using the weiszfeld algorithm. In The 24th IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2011, Colorado Springs, CO, USA, 20-25 June 2011, pages 3041-3048, 2011. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.338, + 0.894, + 0.392 + ], + "angle": 0, + "content": "[21] Xiangru Huang, Zhenxiao Liang, Xiaowei Zhou, Yao Xie, Leonidas J. Guibas, and Qixing Huang. Learning transformation synchronization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.392, + 0.894, + 0.418 + ], + "angle": 0, + "content": "[22] Svante Janson. Poisson approximation for large deviations. Random Structures & Algorithms, 1(2):221-229, 1990. 5, 14, 16" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.418, + 0.894, + 0.459 + ], + "angle": 0, + "content": "[23] Svante Janson. Large deviations for sums of partly dependent random variables. Random Structures & Algorithms, 24(3):234-248, 2004. 18, 21" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.459, + 0.894, + 0.498 + ], + "angle": 0, + "content": "[24] Jeong Han Kim and Van H Vu. Concentration of multivariate polynomials and its applications. Combinatorica, 20(3):417-434, 2000. 5, 14, 18" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.499, + 0.894, + 0.551 + ], + "angle": 0, + "content": "[25] Gilad Lerman and Yunpeng Shi. Robust group synchronization via cycle-edge message passing. Foundations of Computational Mathematics, 22(6):1665-1741, 2022. 1, 2, 3, 5, 6, 13, 14, 16, 19, 21" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.552, + 0.894, + 0.621 + ], + "angle": 0, + "content": "[26] Heng Li, Zhaopeng Cui, Shuaicheng Liu, and Ping Tan. RAGO: recurrent graph optimizer for multiple rotation averaging. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2022, New Orleans, LA, USA, June 18-24, 2022, pages 15766-15775. IEEE, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.621, + 0.894, + 0.675 + ], + "angle": 0, + "content": "[27] Shaohan Li, Yunpeng Shi, and Gilad Lerman. Fast, accurate and memory-efficient partial permutation synchronization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 15735-15743, June 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.675, + 0.894, + 0.715 + ], + "angle": 0, + "content": "[28] Shuyang Ling. Near-optimal performance bounds for orthogonal and permutation group synchronization via spectral methods. arXiv preprint arXiv:2008.05341, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.715, + 0.894, + 0.782 + ], + "angle": 0, + "content": "[29] Huikang Liu, Xiao Li, and Anthony Man-Cho So. Resync: Riemannian subgradient-based robust rotation synchronization. In Advances in Neural Information Processing Systems: 37th Annual Conference on Neural Information Processing Systems, 2023. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.782, + 0.894, + 0.809 + ], + "angle": 0, + "content": "[30] Tyler Maunu and Gilad Lerman. Depth descent synchronization in SO(D). Int. J. Comput. Vis., 131(4):968-986, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.809, + 0.894, + 0.863 + ], + "angle": 0, + "content": "[31] Onur Özyesil and Amit Singer. Robust camera location estimation by convex programming. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 2674-2683, 2015. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.863, + 0.894, + 0.891 + ], + "angle": 0, + "content": "[32] Deepti Pachauri, Risi Kondor, and Vikas Singh. Solving the multi-way matching problem by permutation synchronization. In" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.891 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5268" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.092, + 0.47, + 0.134 + ], + "angle": 0, + "content": "C. J. C. Burges, L. Bottou, M. Welling, Z. Ghahramani, and K. Q. Weinberger, editors, Advances in Neural Information Processing Systems 26, pages 1860-1868. Curran Associates, Inc., 2013. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.134, + 0.472, + 0.202 + ], + "angle": 0, + "content": "[33] Álvaro Parra, Shin-Fang Ch'ng, Tat-Jun Chin, Anders P. Eriksson, and Ian Reid. Rotation coordinate descent for fast globally optimal rotation averaging. In IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2021, virtual, June 19-25, 2021, pages 4298-4307. Computer Vision Foundation / IEEE, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.202, + 0.471, + 0.227 + ], + "angle": 0, + "content": "[34] Pulak Purkait, Tat-Jun Chin, and Ian Reid. *Neurora: Neural robust rotation averaging*, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.228, + 0.471, + 0.268 + ], + "angle": 0, + "content": "[35] Ian C Ross and Frank Harary. On the determination of redundancies in sociometric chains. Psychometrika, 17(2):195-208, 1952. 4, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.268, + 0.471, + 0.322 + ], + "angle": 0, + "content": "[36] Tianwei Shen, Siyu Zhu, Tian Fang, Runze Zhang, and Long Quan. Graph-based consistent matching for structure-from-motion. In European Conference on Computer Vision, pages 139-155. Springer, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.322, + 0.471, + 0.362 + ], + "angle": 0, + "content": "[37] Jianbo Shi and J. Malik. Normalized cuts and image segmentation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 22(8):888-905, 2000. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.363, + 0.471, + 0.431 + ], + "angle": 0, + "content": "[38] Yunpeng Shi and Gilad Lerman. Estimation of camera locations in highly corrupted scenarios: All about that base, no shape trouble. In 2018 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2018, Salt Lake City, UT, USA, June 18-22, 2018, pages 2868-2876, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.431, + 0.471, + 0.485 + ], + "angle": 0, + "content": "[39] Yunpeng Shi and Gilad Lerman. Message passing least squares framework and its application to rotation synchronization. In Proceedings of the 37th International Conference on Machine Learning (ICML), 2020. 1, 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.485, + 0.471, + 0.538 + ], + "angle": 0, + "content": "[40] Yunpeng Shi, Shaohan Li, and Gilad Lerman. Robust multi-object matching via iterative reweighting of the graph connection laplacian. Advances in Neural Information Processing Systems, 2020-December, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.539, + 0.471, + 0.593 + ], + "angle": 0, + "content": "[41] Chitturi Sidhartha and Venu Madhav Govindu. It is all in the weights: Robust rotation averaging revisited. In International Conference on 3D Vision, 3DV 2021, London, United Kingdom, December 1-3, 2021, pages 1134-1143. IEEE, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.593, + 0.471, + 0.633 + ], + "angle": 0, + "content": "[42] Amit Singer. Angular synchronization by eigenvectors and semidefinite programming. Applied and computational harmonic analysis, 30(1):20-36, 2011. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.633, + 0.471, + 0.673 + ], + "angle": 0, + "content": "[43] A. Singer and Y. Shkolnisky. Angular synchronization by eigenvectors and semidefinite programming: Analysis and application to class averaging in cryo-electron microscopy, 2009. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.673, + 0.471, + 0.713 + ], + "angle": 0, + "content": "[44] Johan Thunberg, Florian Bernard, and Jorge Goncalves. Distributed methods for synchronization of orthogonal matrices over graphs. Automatica, 80:243-252, jun 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.714, + 0.471, + 0.768 + ], + "angle": 0, + "content": "[45] Roberto Tron and René Vidal. Distributed image-based 3-d localization of camera sensor networks. In Proceedings of the 48th IEEE Conference on Decision and Control, CDC 2009, December 16-18, 2009, Shanghai, China, pages 901-908, 2009. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.768, + 0.471, + 0.822 + ], + "angle": 0, + "content": "[46] Anton Voropaev and Sergey Perepechko. The number of fixed length cycles in undirected graph. explicit formulae in case of small lengths. Bulletin of PFUR. Series Mathematics, Information Sciences, Physics, pages 5-11, 01 2012. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.822, + 0.471, + 0.876 + ], + "angle": 0, + "content": "[47] Anton Voropaev and Sergey Perepechko. The number of fixed length cycles in undirected graph. explicit formulae in case of small lengths. Bulletin of PFUR. Series Mathematics, Information Sciences, Physics, pages 5-11, 01 2012. 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.876, + 0.471, + 0.89 + ], + "angle": 0, + "content": "[48] Van H Vu. A large deviation result on the number of small" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.89 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "subgraphs of a random graph. Combinatorics, Probability and Computing, 10(1):79-94, 2001. 5, 14, 17" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.12, + 0.894, + 0.159 + ], + "angle": 0, + "content": "[49] Lanhui Wang and Amit Singer. Exact and stable recovery of rotations for robust synchronization. Information and Inference, 2013. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.16, + 0.894, + 0.214 + ], + "angle": 0, + "content": "[50] Kyle Wilson and Noah Snavely. Robust global translations with 1dsfm. In Computer Vision - ECCV 2014 - 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part III, pages 61-75, 2014. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.214, + 0.894, + 0.283 + ], + "angle": 0, + "content": "[51] Christopher Zach, Manfred Klopschitz, and Marc Pollefeys. Disambiguating visual relations using loop constraints. In The Twenty-Third IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2010, San Francisco, CA, USA, 13-18 June 2010, pages 1426-1433, 2010. 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.283 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.957 + ], + "angle": 0, + "content": "5269" + } + ] +] \ No newline at end of file diff --git a/2024/Efficient Detection of Long Consistent Cycles and its Application to Distributed Synchronization/c0dd6b15-09a8-4eba-bcfb-24b8083dd375_origin.pdf b/2024/Efficient Detection of Long Consistent Cycles and its Application to Distributed Synchronization/c0dd6b15-09a8-4eba-bcfb-24b8083dd375_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..46f7fcbe6d0a2d817ab0e26778c26143cbfb8020 --- /dev/null +++ b/2024/Efficient Detection of Long Consistent Cycles and its Application to Distributed Synchronization/c0dd6b15-09a8-4eba-bcfb-24b8083dd375_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:46ab07a4103e56c286724b5221c645a873f6435df22f3cbec20bbad07cc4fcb5 +size 426815 diff --git a/2024/Efficient Detection of Long Consistent Cycles and its Application to Distributed Synchronization/full.md b/2024/Efficient Detection of Long Consistent Cycles and its Application to Distributed Synchronization/full.md new file mode 100644 index 0000000000000000000000000000000000000000..b81b8ccc2ac7b74bd4f2751a5ccaf26b715ceef0 --- /dev/null +++ b/2024/Efficient Detection of Long Consistent Cycles and its Application to Distributed Synchronization/full.md @@ -0,0 +1,386 @@ +# Efficient Detection of Long Consistent Cycles and its Application to Distributed Synchronization * + +Shaohan Li +University of Minnesota +li000743@umn.edu + +Yunpeng Shi +University of California, Davis +ypshi@ucdavis.edu + +Gilad Lerman +University of Minnesota +lerman@umn.edu + +# Abstract + +Group synchronization plays a crucial role in global pipelines for Structure from Motion (SfM). Its formulation is nonconvex and it is faced with highly corrupted measurements. Cycle consistency has been effective in addressing these challenges. However, computationally efficient solutions are needed for cycles longer than three, especially in practical scenarios where 3-cycles are unavailable. To overcome this computational bottleneck, we propose an algorithm for group synchronization that leverages information from cycles of lengths ranging from three to six with a time complexity of order $O(n^{3})$ (or $O(n^{2.373})$ when using a faster matrix multiplication algorithm). We establish non-trivial theory for this and related methods that achieves competitive sample complexity, assuming the uniform corruption model. To advocate the practical need for our method, we consider distributed group synchronization, which requires at least 4-cycles, and we illustrate state-of-the-art performance by our method in this context. + +# 1. Introduction + +Structure from Motion (SfM) asks to recover the 3D structure of a stationary scene from multiple images taken by cameras from different orientations and locations. In the past decade, the global SfM pipeline has become increasingly popular due to its several advantages over the incremental pipelines [17, 31]. First of all, global SfM requires only one implementation of bundle adjustment, making it more efficient in computation. Second, it estimates camera poses in a global optimization framework which mitigates the drifting issue of the incremental pipelines. Despite the popularity of global SfM pipelines, the estimation of global camera poses (e.g., orientations) remains a highly challenging problem. For instance, estimating camera orientations from their relative measurements, often called rotation synchronization, is a highly nonconvex graph optimization problem. In typical scenarios of highly noisy or corrupted measurements + +of relative orientations, many common solutions of rotation synchronization have poor accuracy and slow convergence. + +Given these challenges, theoretical developments have demonstrated the critical role of cycle-consistency information in inferring corrupted measurements [25]. In practice, the consistency constraint on 3-cycles was utilized to estimate the error of each measured relative orientation. It also helped nonconvex iterative rotation synchronization solvers avoid spurious local minima and achieve significantly higher accuracy [39]. However, the usage of 3-cycles largely limits the application of these improved algorithms to other important scenarios. One scenario involves a sparse viewing graph lacking sufficient 3-cycles. This often occurs when the size of the graph is too large to densely measure the relative orientations on its edges, which could happen in certain case for the molecular orientation estimation in cryo-electron microscopy imaging. Another scenario is orientation estimation for each piece of jigsaw puzzles, where the graph is a 2D lattice and 3-cycle does not exist. Lastly, in distributed SfM, edges between any two clusters of nodes form a bipartite graph, and cycles of odd length do not exist. Our numerical results primarily emphasize the practical scenario of distributed SfM, which holds particular relevance for the broader computer vision community. + +Despite the multiple critical applications of long-cycle consistency, inferring measurement noise from long cycles is challenging in both computation and theory. First of all, the number of cycles grows exponentially with the cycle length, and measuring cycle inconsistencies for each long cycle is computationally intractable. Moreover, developing theoretical guarantees for long-cycle inference methods is fundamentally more difficult than the 3-cycle case. Indeed, in a random graph setting, a set of longer cycles are more likely to share common edges, making their consistency score highly correlated. Therefore, new tools are required to handle the correlated empirical process. + +In this work, we propose the first practical method, LongSync, for inferring edge corruption levels from long cycle consistency information. For this purpose, we carefully modify and vectorize the Cycle Edge Message Passing (CEMP) method [25]. This nontrivial modification drastically reduces its computational complexity when using longer cycles. Moreover, + +by employing a more delicate analysis and incorporating new tools from probability theory and combinatorics, we develop a significantly stronger exact recovery result with a general cycle length under a popular probabilistic model. The sample complexity in our theory is the lowest among all practical rotation synchronization methods. Although we limit our scope to the application of distributed SfM, our algorithm and theory applies to any finite-dimensional linear group in group synchronization, and not just $SO(3)$ in rotation synchronization. + +# 1.1. Related Work + +Earlier rotation synchronization methods [2, 10-13, 18, 33, 42, 43] seek to minimize a least squares energy function. They can be described as relaxed versions of the maximum likelihood estimator under the additive Gaussian noise model, but they are not robust in the presence of outliers or heavy-tailed noise. Nevertheless, in the case of global SfM, the initially estimated relative camera rotations can be severely corrupted due to the erroneous keypoint matches and the subsequent poor estimation of fundamental matrices. + +To handle outliers, robust rotation synchronization methods either minimize a robust energy function or reweigh/trim the viewing graph based on the corruption levels of the edges. Wang and Singer [49] minimizes a corresponding $\ell_1$ objective function using semidefinite programming (SDP) relaxation, which is slow in practice. Other energy minimization methods are typically nonconvex, which include the Weiszfeld algorithm [20] and the Riemannian subgradient method [29] for $\ell_1$ minimization, and the iteratively reweighted least squares (IRLS) for minimizing general $\ell_p$ [5] and Geman-McClure [4] loss functions. However, all these methods heavily rely on good initialization. Sidhartha and Govindu [41] partially remedy the issue using adaptive Geman-McClure loss functions, but their approach remains sensitive to the initialized weights. Maunu and Lerman [30] propose to solve rotation synchronization by an iterative robust averaging method that utilizes Tukey depth, but they have not demonstrated effective performance for real SfM applications. Arrigoni et al. [1] applies a low-rank and sparse matrix decomposition method to SO(3) and SE(3) synchronization, but it is even less robust to outliers than IRLS. + +Instead of employing a robust objective function, Shen et al. [36] and Zach et al. [51] uses the 3-cycle consistency constraint to detect and remove corrupted relative orientations. Lerman and Shi [25] take one step further to estimate the corruption level of each relative measurement by a novel cycle-edge message passing (CEMP) algorithm. They then use the estimated corruption levels to reweigh the graph and solve rotation synchronization using a weighted least squares method. This message passing procedure was further combined with IRLS to boost its accuracy in [39]. Particular versions and extensions of this procedure for permutation and partial permutation synchronization, which are relevant to the matching component of SfM, were discussed in [27, 40]. + +However, all the previously mentioned cycle-based methods [25, 27, 36, 38-40, 51] only use 3-cycles in practice, limiting their application for distributed synchronization. Indeed, the standard distributed synchronization often requires "stitching" local solutions by synchronizing the relative rotations between clusters. Each of these inter-cluster rotations is estimated by "averaging" the edges between the two clusters. These edges form a bipartite graph, and the minimal cycle length is 4. As pointed in [51], the number of operations for computing long cycle consistency information scales exponentially with the cycle length. Therefore, none of the existing distributed rotation synchronization methods directly exploits long cycle information due to this computational challenge. + +The earlier distributed methods for $SO(d)$ synchronization, such as [45] and [44], minimize a least squares energy and are not robust to outliers. A series of distributed SfM methods [14-16] implement incremental SfM algorithms for each cluster. However, these methods do not employ a standard rotation synchronization algorithm, as they require additional information such as the number of keypoint matches between images. Moreover, the incremental methods are slower since they require multiple rounds of global rotation synchronization. MultiSync [9] synchronizes the inter-cluster rotations directly using all inter-cluster edges among all clusters, by formulating a novel synchronization problem on a multi-graph. Although it utilizes a more unified formulation, its objective function is least squares which largely limits its robustness to outliers. + +A recent and different type of methods for rotation synchronization use deep learning [21, 26, 34]. However, these methods are supervised and thus may not generalize well when switching datasets. Moreover, like many other previous methods, they lack theoretical guarantees. + +A common theoretical setting to assess the performance of rotation synchronization algorithms is the uniform corruption model (UCM) described in §4. We provide the best sample complexity for LongSync, even with only 3 cycles, among all previously established estimates for the UCM model. + +# 1.2. Contributions of This Work + +- We propose the first practical algorithm that infers edge corruption levels from long cycle consistency information. The computation complexity of our method is reduced from $O(n^{c})$ to $O(n^{3})$ (or possibly $O(n^{2.373})$ ) for cycle length $c \leq 6$ and $O(n^{[(c + 3) / 2]})$ for $c > 6$ . +- We establish sample complexity estimates for our method under the uniform corruption model, where we get closer to the information theoretic bound than any other existing work. Our proof requires delicate analysis and it also improves previous estimates for the CEMP algorithm. +- We introduce a new graph partition and graph preprocessing method that utilizes our inference method, and demonstrate the effectiveness of our pipeline in boosting the performance of distributed synchronization. + +- Extensive numerical experiments demonstrate the outstanding performance of our method. + +# 2. Problem Formulation and Preliminaries + +Assume a graph $G = ([n], E)$ where $[n]$ is the set of nodes indexed by $\{1, 2, \dots, n\}$ and $E$ is the set of edges. Given a mathematical group $\mathcal{G}$ , each graph node is assigned an underlying ground truth group element $R_{i}^{*}$ , where $R_{i}^{*} \in \mathcal{G}$ and we use star superscript to emphasize the ground truth. For each edge $ij \in E$ , we observe a relative group ratio $R_{ij} \in \mathcal{G}$ , whose clean counterpart is $R_{ij}^{*} = R_{i}^{*}R_{j}^{*-1}$ . Group synchronization aims to recover the ground truth group elements $\{R_{i}^{*}\}_{i \in [n]}$ from the possibly noisy and corrupted measurements $\{R_{ij}\}_{ij \in E}$ . In this paper, we focus on the case of rotation synchronization, which is a special case of group synchronization with $\mathcal{G} = \mathrm{SO}(d)$ . For applications in camera orientation synchronization ( $d = 3$ ), we estimate absolute rotations for each node $i \in [n]$ from measured relative rotations of edges in $E$ . Note that since $\{R_{i}^{*}\}_{i \in [n]}$ and $\{R_{i}^{*}R_{0}\}_{i \in [n]}$ generate the same set of relative rotations, one can only estimate $\{R_{i}^{*}\}_{i \in [n]}$ up to a global rotation. The generalization to any linear groups is discussed in the supplementary material. + +# 2.1. Notations and Definitions + +We denote the adjacency matrix of graph $G$ as $\mathbf{A}$ , and form a pairwise observation matrix $\mathbf{R} \in \mathbb{R}^{dn \times dn}$ by stacking the $\mathbf{R}_{ij}$ 's (for $ij \notin E$ , set $\mathbf{R}_{ij} = \mathbf{0}_{3 \times 3}$ ): + +$$ +\boldsymbol {R} := \left( \begin{array}{c c c c} \boldsymbol {R} _ {1 1} & \boldsymbol {R} _ {1 2} & \dots & \boldsymbol {R} _ {1 n} \\ \boldsymbol {R} _ {2 1} & \boldsymbol {R} _ {2 2} & \dots & \boldsymbol {R} _ {2 n} \\ \vdots & \vdots & \ddots & \vdots \\ \boldsymbol {R} _ {n 1} & \boldsymbol {R} _ {n 2} & \dots & \boldsymbol {R} _ {n n} \end{array} \right). +$$ + +We list the matrix operations used in the paper. For matrices $X$ and $Y$ , the operations $X \otimes Y$ , $X \odot Y$ , $X \oslash Y$ respectively denote the Kronecker product, Hardmard (element-wise) multiplication and Harnard division between $X$ and $Y$ . $X^{\odot k}$ denotes the element-wise matrix $k$ -power. For block matrices, $\langle X, Y \rangle_{\mathrm{block}}$ denotes the blockwise inner product of $X$ and $Y$ , i.e. $\langle X, Y \rangle_{\mathrm{block}}(i,j) = \langle X[i,j], Y[i,j] \rangle$ , where $[i,j]$ refers to the corresponding block of the matrix. + +# 2.2. Review of CEMP for $c$ -Cycles + +We assume the above setting of $SO(d)$ synchronization. Let $\mathcal{D}$ be any bi-invariant metric on $SO(d)$ . We assume a fixed number of cycles, $c$ , and denote by $N_{ij}^{c}$ the set of simple cycles of length $c$ (or simple $c$ -cycles) containing edge $ij$ . CEMP [25] aims to estimate for each edge $ij$ the corruption level + +$$ +s _ {i j} ^ {*} = \mathcal {D} \left(\boldsymbol {R} _ {i j}, \boldsymbol {R} _ {i j} ^ {*}\right), \tag {1} +$$ + +from the set of cycle inconsistency measures + +$$ +d _ {L} = \mathcal {D} \left(\boldsymbol {R} _ {L}, \boldsymbol {R} _ {i j}\right) \tag {2} +$$ + +where cycle $L = (ik_{1}, k_{1}k_{2}, \dots, k_{c-2}j, ji) \in N_{ij}^{c}$ and $R_{L} := R_{ik_{1}}R_{k_{1}k_{2}}\dots R_{k_{c-2}j}$ . The estimated $s_{ij}^{*}$ can then be used for extracting a clean subgraph, or to implement a weighted least squares solver where higher weights are assigned to cleaner edges. + +It is obvious that if all the edges in $L$ are clean then $d_{L} = 0$ . Moreover, due to bi-invariance of $\mathcal{D}$ , the following holds true + +$$ +d _ {L} = s _ {i j} ^ {*} \text {w h e n e v e r} L \in G _ {i j} ^ {c}, \tag {3} +$$ + +where $G_{ij}^{c}$ is the set of good $c$ -cycles with respect to $ij$ , i.e. the set of cycles $L \in N_{ij}^{c}$ such that $ik_1, \dots, k_{c-2j}$ are clean. This gives a sufficient condition for $d_L$ to be an exact estimator of $s_{ij}^{*}$ . + +To estimate the corruption levels of each edge $ij$ , CEMP initializes the edge weight of each $ij \in E$ as $w_{ij}^{(0)} = 1$ . It then iteratively updates the corruption level estimate as the following convex combination of $d_L$ 's: + +$$ +s _ {i j} ^ {(t)} = \sum_ {L \in N _ {i j} ^ {c}} w _ {L} ^ {(t)} d _ {L} / z _ {i j} ^ {(t)} \tag {4} +$$ + +where $z_{ij}^{(t)} = \sum_{L\in N_{ij}^c}w_L^{(t)}$ . The cycle weights $w_{L}^{(t + 1)}$ are computed from the edge weights $w_{e}^{(t + 1)} = e^{-\beta_{t}s_{e}^{(t)}}$ : + +$$ +w _ {L} ^ {(t + 1)} = \prod_ {e \in L \backslash \{i j \}} w _ {e} ^ {(t + 1)} = \prod_ {e \in L \backslash \{i j \}} e ^ {- \beta_ {t} s _ {e} ^ {(t)}}, \tag {5} +$$ + +so that $w_{L}^{(t + 1)}$ focuses on good cycles. The cycle weights and edge corruption levels are alternatingly updated and improved from each other. Interestingly, it is proved in [25] under two different corruption models that CEMP converges linearly to the ground truth corruption estimates under mild conditions for $c = 3$ . In practice, CEMP only uses 3-cycles for consideration of efficiency. For longer cycles, the complexity of CEMP scales exponentially with the cycle length $c$ (which is discussed in §3.2), and the convergence guarantee of CEMP remains unknown. + +# 3. Our method: LongSync + +# 3.1. LongSync: Modification of CEMP + +Our goal is to develop a scalable variant of CEMP for any fixed number of cycles, $c \geq 3$ . The main computational bottleneck of step (4) in CEMP is that computing and summing the cycle inconsistency measures takes $\sum_{ij \in E} |N_{ij}^c| = O(n^c)$ operations and memory. Therefore, to develop a scalable algorithm, we aim to take weighted average over $d_L$ without explicitly computing and storing each $d_L$ . To achieve this, we propose the following specification and modification on CEMP: + +- Use Chordal distance on SO(d). We suggest the distance function + +$$ +\mathcal {D} \left(\boldsymbol {R} _ {1}, \boldsymbol {R} _ {2}\right) = \sqrt {1 - \left\langle \boldsymbol {R} _ {1} , \boldsymbol {R} _ {2} \right\rangle / d}. +$$ + +This distance is proportional to the Chordal distance on $SO(3)$ , which is the Euclidean distance between two rotations embedded in $\mathbb{R}^{d\times d}$ . + +- Use weighted quadratic average for corruption level update. Instead of updating the corruption level estimates by a weighted average of $d_{L}$ , we use the weighted quadratic average of $d_{L}$ , namely + +$$ +s _ {i j} ^ {(t)} = \sqrt {\sum_ {L \in N _ {i j} ^ {c}} w _ {L} ^ {(t)} d _ {L} ^ {2} / z _ {i j} ^ {(t)}} \tag {6} +$$ + +where the update rule of cycle weights remains the same: + +$$ +w _ {L} ^ {(t + 1)} = \prod_ {e \in L \backslash \{i j \}} w _ {e} ^ {(t + 1)} = \prod_ {e \in L \backslash \{i j \}} e ^ {- \beta_ {t} s _ {e} ^ {(t)}}. \tag {7} +$$ + +As a result, $d_{L}^{2} = d^{2}(\pmb{R}_{L},\pmb{R}_{ij}) = 1 - \langle \pmb{R}_{L},\pmb{R}_{ij}\rangle /d$ is linear in both $R_{L}$ and $R_{ij}$ . Therefore one can switch the order of $d^{2}$ and the weighted summation, so that the computation of $s_{ij}^{(t)}$ can be vectorized. Indeed, by this linearity and equations (4) and (2), and note that $z_{ij}^{(t)} = \sum_{L\in N_{ij}^{c}}w_{L}^{(t)}$ , we obtain the following equation: + +$$ +\begin{array}{l} s _ {i j} ^ {(t)} = \left(\sum_ {L \in N _ {i j} ^ {c}} w _ {L} ^ {(t)} d _ {L} ^ {2} / z _ {i j} ^ {(t)}\right) ^ {1 / 2} \\ = \Big (\sum_ {L \in N _ {i j} ^ {c}} w _ {L} ^ {(t)} \mathcal {D} ^ {2} (\boldsymbol {R} _ {L}, \boldsymbol {R} _ {i j}) / z _ {i j} ^ {(t)} \Big) ^ {1 / 2} \\ = \left(\mathcal {D} ^ {2} \Big (\sum_ {L \in N _ {i j} ^ {c}} w _ {L} ^ {(t)} \boldsymbol {R} _ {L}, \boldsymbol {R} _ {i j} \Big) / z _ {i j} ^ {(t)}\right) ^ {1 / 2} \\ = \left(1 - \left\langle \sum_ {L \in N _ {i j} ^ {c}} w _ {L} ^ {(t)} \boldsymbol {R} _ {L}, \boldsymbol {R} _ {i j} \right\rangle / d \sum_ {L \in N _ {i j} ^ {c}} w _ {L} ^ {(t)}\right) ^ {1 / 2} \tag {8} \\ \end{array} +$$ + +Equation (8) can be vectorized using the trick of matrix power if we allow repeated nodes for each cycle. That is, one can stack the $s_{ij}^{(t)}$ 's and $w_{ij}^{(t)}$ 's into matrices $\mathbf{S}^{(t)}$ and $\mathbf{W}^{(t)}$ , and vectorize (8) as + +$$ +\begin{array}{l} \boldsymbol {S} ^ {(t)} = \left(\boldsymbol {A} - \left\langle \left(\boldsymbol {W} ^ {(t)} \otimes 1 _ {d} \odot \boldsymbol {R}\right) ^ {c - 1}, \boldsymbol {R} \right\rangle_ {\text {b l o c k}} \right. \\ \left. \odot d \left(\boldsymbol {W} ^ {(t)}\right) ^ {c - 1}\right) ^ {\odot 1 / 2}. \tag {9} \\ \end{array} +$$ + +Indeed, by (7) and the definition of $\pmb{R}_L$ , $\sum_{L \in C_{ij}^c} w_L^{(t)} \pmb{R}_L$ is the $ij$ -th block of $(\pmb{W}^{(t)} \otimes 1_d \odot \pmb{R})^{c-1}$ , and $\sum_{L \in C_{ij}^c} w_L^{(t)}$ is the $ij$ -th element of $\pmb{W}^{(t)c-1}$ , where $C_{ij}^c$ is the set of $c$ -cycles containing $ij$ with possibly repeated nodes. + +In the case of utilizing only simple cycles, (8) and (9) are not equivalent and we need to correct (9) to remove the cycles with repeated nodes, so that only simple cycles in $N_{ij}^{c}$ remain. Let $g_{c}(\boldsymbol{W},\boldsymbol{R})$ be the matrix valued function where $g_{c}(\boldsymbol{W},\boldsymbol{R})(i,j) = \sum_{L\in N_{ij}^{c}}(\prod_{e\in L\setminus \{ij\}}w_{e}^{(t)})\boldsymbol{R}_{L}$ . Let $f_{c}(\boldsymbol{W})$ be the matrix valued function where $f_{c}(\boldsymbol{W})(i,j) = \sum_{L\in N_{ij}^{c}}(\prod_{e\in L\setminus \{ij\}}w_{e}^{(t)})$ . The following result holds: + +Proposition 3.1. The update rule of LongSync (8) is equivalent to the following matrix equations: + +$$ +\boldsymbol {S} ^ {(t)} = \left(\boldsymbol {A} - \left\langle h _ {c} (\boldsymbol {W} ^ {(t)}, \boldsymbol {R}), \boldsymbol {R} \right\rangle_ {\text {b l o c k}} \odot \boldsymbol {A}\right) ^ {\odot 1 / 2}, \tag {10} +$$ + +where $\pmb{W}^{(t + 1)} = \pmb{A}\odot \exp (-\beta_t\pmb{S}^{(t)})$ and + +$$ +h _ {c} \left(\boldsymbol {W} ^ {(t)}, \boldsymbol {R}\right) := g _ {c} \left(\boldsymbol {W} ^ {(t)}, \boldsymbol {R}\right) \oslash \left(d \cdot f _ {c} \left(\boldsymbol {W} ^ {(t)} \otimes 1 _ {d}\right)\right). +$$ + +Here exp denotes the elementwise exponential function. + +We use equation (10) as the update rule of LongSync and propose the vectorized LongSync algorithm in algorithm 1. + +# Algorithm 1 (LongSync) + +Input: pairwise rotation matrix $\mathbf{R} \in \mathbb{R}^{dn \times dn}$ , adjacency matrix $\mathbf{A} \in [0,1]^{n \times n}$ , cycle length $c$ , positive parameters $\{\beta_t\}_{t \geq 1}$ , time step $T$ + +$$ +\boldsymbol {W} ^ {(0)} (i, j) \leftarrow \boldsymbol {A} +$$ + +for $t = 0:T$ do + +$$ +\boldsymbol {S} ^ {(t)} \leftarrow \left(\boldsymbol {A} - \left\langle h _ {c} (\boldsymbol {W} ^ {(t)}, \boldsymbol {R}), \boldsymbol {R} \right\rangle_ {\text {b l o c k}} \odot \boldsymbol {A}\right) ^ {\odot 1 / 2} \tag {11} +$$ + +$$ +\boldsymbol {W} ^ {(t + 1)} \leftarrow \boldsymbol {A} \odot \exp (- \beta_ {t} \boldsymbol {S} ^ {(t)}) \tag {12} +$$ + +# end for + +Output: edge weights $\mathbf{W}^{(T + 1)}$ , corruption levels $\mathbf{S}^{(T)}$ + +We claim that $g_{c}$ and $f_{c}$ can be computed with a sequence of matrix operations, thus greatly reducing the time and space consumption of LongSync compared to its original form. For $c \leq 6$ , the time complexity of computing $g_{c}$ and $f_{c}$ is $O(r(dn))$ , where $r(n)$ is the complexity for multiplying two $n \times n$ matrices; for $c \geq 7$ the time complexity is at most $O(n^{[(c + 3) / 2]})$ . This claim is proved in the supplementary material. We list the formula for $g_{c}$ and $f_{c}$ for $c = 3,4,5,6$ inspired by [35, 46]. The formula for $c = 6$ is moved to the supplementary material due to the space limit. For $c \geq 7$ the formula becomes extremely complicated. We remark that in practice, the cycles of length greater than 6 are often not used. + +We finally remark that although Algorithm 1 only utilizes cycles of a fixed length, one can easily generalize it to incorporate cycles of different lengths. Indeed, the equation (10) could use a convex combination of $h_c$ 's that corresponds to different values of $c$ . That is, for a preselected set of cycle lengths $C$ , the equation (11) in Algorithm 1 is replaced by + +$$ +\boldsymbol {S} ^ {(t)} = \left(\boldsymbol {A} - \left\langle \sum_ {c \in C} \lambda_ {c} h _ {c} \left(\boldsymbol {W} ^ {(t)}, \boldsymbol {R}\right), \boldsymbol {R} \right\rangle_ {\text {b l o c k}} \odot \boldsymbol {A}\right) ^ {\odot 1 / 2} \tag {13} +$$ + +where the coefficients $\lambda_{c}$ satisfies $\sum_{c\in C}\lambda_c = 1$ to ensure a convex combination. Here each $\lambda_{c}$ is user-specified to reflect the importance of the cycles of length $c$ . However, the optimal + +choice of these parameters under certain statistical model remains unclear. + +For simplicity, in the experiments we only use a fixed cycle length to avoid choosing $\lambda_{c}$ . We have observed that such simple choice still yields satisfying accuracy in camera orientation estimation on both synthetic and real data. We refer the readers to §5 and 6 for more details. + +
cFormula of fc(W)Formula of gc(W,R)
3W2P2
4W3-d(W2)WP3-dblock(P2)P
-Wd(W2)+W⊙3-Pdblock(P2)+P⊙3
5W4-d(W3)WP4-dblock(P3)P
-d(W2)W2-dblock(P2)P2
-W2d(W2)-P2dblock(P2)
-Wd(W2)W-Pdblock(P2)P
+3W⊙2W2+3P⊙2P2
+WW⊙3+W⊙3W+PP⊙3+P⊙3P
+ +Table 1. Formulas for $f_{c}$ and $g_{c}$ . Here we let $P = (W \otimes 1_d) \odot R$ for shorter notation. $\mathrm{d}(X)$ returns the diagonal of matrix $X$ , $\mathrm{d}_{\mathrm{block}}(X)$ returns the diagonal block matrix from the $d \times d$ diagonal blocks of matrix $X$ . + +# 3.2. Computational Complexity + +We derive the space and time complexity for LongSync, and demonstrate its advantages over CEMP. The initialization step involves setting the weights of all edges to 1, which takes time $O(|E|)$ and space $O(n^{2})$ . For each iteration, LongSync updates the matrices $\pmb{S}^{(t)}$ and $\pmb{W}^{(t)}$ with equations (11) and (12), respectively. Computing $\pmb{W}^{(t + 1)}$ involves two matrix subtractions, one scalar-matrix multiplication and one element-wise matrix exponential operation on $\pmb{S}^{(t)} \in \mathbb{R}^{n \times n}$ . Therefore the update step (12) takes at most $O(n^{2})$ time and space. Equation (11), on the other hand, involves a sequence of matrix operations on $\pmb{P}^{(t)} = (\pmb{W}^{(t)} \otimes 1_{d}) \odot \pmb{R} \in \mathbb{R}^{dn \times dn}$ and $\pmb{W}^{(t)} \in \mathbb{R}^{n \times n}$ , including matrix multiplications, element-wise multiplications and diagonal block selections. Computing $\pmb{P}^{(t)}$ takes $O(d^{2}n^{2})$ memory and $O(d^{2}n^{2})$ time. The matrix operations on $\pmb{P}^{(t)}$ take $O(K_{c}d^{3}n^{3})$ time and $O(d^{2}n^{2})$ space, and the matrix operations on $\pmb{W}^{(t)}$ take $O(K_{c}n^{3})$ time and $O(n^{2})$ space, where $K_{c}$ is the number of terms in the equation for $f_{c}$ and $g_{c}$ . Therefore, each iteration over $t$ takes $O(K_{c}d^{3}n^{3})$ time and $O(d^{2}n^{2})$ space. To sum up, for LongSync, the time complexity is $O(K_{c}d^{3}n^{3})$ and the space complexity is $O(d^{2}n^{2})$ . For $c = 3,4,5,6$ , the number $K_{c}$ is equal to 1,3,9,32. + +In comparison, we consider the initialization step of CEMP. For each edge $ij \in E$ and $L \in N_{ij}^{c}$ , initializing CEMP involves computing and storing all the cycle inconsistency measures $d_{L}$ using equation (2). For each $L \in N_{ij}^{c}$ , computing $d_{L}$ involves multiplying $c$ rotations, which takes $O(cd^3)$ time and $O(d^2)$ space. This step is repeated for each $ij \in E$ and $L \in N_{ij}^{c}$ , therefore the total time complexity is $O(cd^3\sum_{ij \in E}|N_{ij}^{c}|)$ and + +the total space complexity is $O(d^{2}\sum_{ij\in E}|N_{ij}^{c}|)$ . Since for each edge there are $(n - 2)(n - 3)\dots (n - c + 1) = O(n^{c - 2})$ cycle candidates, we know that $|N_{ij}^{c}|\sim O(n^{c - 2})$ for each $ij\in E$ in the worst case scenario of a dense graph. Therefore the initialization of CEMP takes $O(cd^3 n^{c - 2}|E|)$ time and $O(d^{2}n^{c - 2}|E|)$ space in the worst case. Given $c\geq 4$ and $|E|\sim n^2$ , CEMP requires much more time and space than LongSync. + +# 4. Theory for Uniform Corruption Model + +In this section, we present the exact recovery guarantee of LongSync under the uniform corruption model (UCM). UCM is a popular probabilistic model that is widely adopted for synthetic experiments of many previous works on group synchronization [8, 25, 28, 29, 32, 42]. The model $\mathrm{UCM}(n,p,q_g)$ assumes that $G$ is an Erdős-Rényi graph with edge connection probability $p$ . For each edge $ij \in E$ , $R_{ij}$ is generated independently as follows: + +$$ +\boldsymbol {R} _ {i j} = \left\{ \begin{array}{l l} \boldsymbol {R} _ {i j} ^ {*} & \text {w . p .} q _ {g}; \\ \tilde {\boldsymbol {R}} _ {i j} \sim \operatorname {H a a r} (\mathcal {G}) & \text {w . p .} 1 - q _ {g}. \end{array} \right. +$$ + +We also developed an exact recovery theory for a general model of adversarial corruption, which we include in section C.1 of the supplementary material. An informal version of our main result for UCM is stated in Theorem 4.1. Although the application of this paper is focused on rotation synchronization, the following theory for UCM is valid for any compact group $\mathcal{G}$ , as explained in the supplementary material. + +Theorem 4.1. Let $0 < r < 1$ , $0 < q < 1$ , $0 < p \leq 1$ , $\mathcal{G} = SO(3)$ . Assume LongSync is applied with cycles of length $c$ , $n / \log n \sim p^{-(c-1) / (c-2-\epsilon)} q_g^{-7(c-1) / 3(c-2)}$ for some $\epsilon > 0$ and + +$$ +1 / \beta_ {t + 1} = r / \beta_ {t} f o r a l l t \geq 1. +$$ + +Then with appropriate choices of $\beta_0, \beta_1, r$ , and high probability, $\max_{ij \in E} |s_{ij}^* - s_{ij}^{(t)}| \leq \frac{1}{2c\beta_t}$ for all $t \geq 1$ . + +The major difficulty of proving Theorem 4.1 is the dependence in the cycle inconsistency measures for cycles in $N_{ij}^{c}$ when $c \geq 4$ . Unlike the 3-cycle case, the cycle inconsistency measure of a 4-cycle $L_{1} = (ik_{1}, k_{1}k_{2}, k_{2}j)$ is correlated with that of $L_{2} = (ik_{1}, k_{1}k_{3}, k_{3}j)$ under UCM. Therefore the key concentration inequalities for the proof cannot be concluded from the standard Chernoff bounds. To overcome this theoretical obstacle, we have integrated various mathematical techniques from [3, 7, 22-24, 48] to prove the theorem, whose details are included in the supplementary material. + +Theorem 4.1 provides an upper bound of the sample complexity (the required graph size $n$ ) of LongSync for exact recovery of the ground truth solutions. This sample complexity is the closest to the information theoretic bound among all existing rotation synchronization methods. The comparison with previous works is summarized in Table 2. + +
ReferenceSample Complexity
[25] for CEMPO(p-2qg-28/3)
[29] for ReSyncO(p-2qg-7)
Ours for CEMPO(p-2-εqg-14/3)
Ours for LongSync (c=3)O(p-2-εqg-14/3)
Ours for LongSync (c=4)O(p-1.5-εqg-3.5)
Ours for LongSync (any c)O(p-7(c-1/c-2-ε) qg-3(c-2))
Ours for LongSync (c→∞)O(p-1-εqg-7/3)
Information Theoretic Bound [6]O(p-1qg-2)
+ +Table 2. Comparison of the sample complexity requirement. Lower absolute values of the powers on $p, q_{g}$ indicate better results. $\epsilon$ is an arbitrarily small positive real number. + +# 5. Synthetic Data Experiment + +We test LongSync on synthetic datasets generated with Uniform Corruption Model (UCM) and Uniform Bipartite Corruption Model (UBCM) respectively described in §5.1 and 5.2. For both models with their corresponding viewing graphs $G = ([n], E)$ , we sample the ground truth absolute rotation matrices $\{R_i^*\}_{i \in [n]}$ independently from the Haar measure on SO(3), and we generate the observed relative rotations $\{R_{ij}\}_{ij \in E}$ independently as follows: + +$$ +\boldsymbol {R} _ {i j} = \left\{ \begin{array}{l l} \boldsymbol {R} _ {i} ^ {*} \boldsymbol {R} _ {j} ^ {*} & \text {w . p .} q _ {g}; \\ \tilde {\boldsymbol {R}} _ {i j} \sim \operatorname {H a a r} (S O (3)) & \text {w . p .} 1 - q _ {g}. \end{array} \right. +$$ + +We use LongSync with cycle length $c$ , $\beta_{t} = \min (2^{t},20)$ and $T = 10$ and record the edge weights. For UCM we set $c = 4,5$ and for UBCM we only use $c = 4$ since no 5-cycles exist. For our method, we first build a weighted graph whose edge weights are estimated by LongSync. We then extract a maximum spanning tree (MST) of the resulting weighted graph. The resulting spanning tree is expected to be the cleanest possible spanning tree. To initialize our solution of absolute rotations, we first fix $\pmb{R}_{1}$ as the identity rotation, and find the rest of $\pmb{R}_{i}$ 's by consecutively multiplying the relative rotations along the spanning tree using the formula $\pmb{R}_{i} = \pmb{R}_{ij}\pmb{R}_{j}$ . To refine our initialized solution, we apply IRLS with Geman-McClure [4] loss functions to minimize $\sum_{ij\in E}\rho_{GM}(d_{\angle}(\pmb{R}_{ij},\pmb{R}_i\pmb{R}_j^T))$ , where $d_{\angle}$ is the geodesic distance in $SO(3)$ . We refer to this method as LongSync+IRLS. + +To demonstrate the advantages of utilizing longer cycle information, we compare our method with IRLS initialized by other two different spanning trees. The first one is the random spanning tree, which uses no cycle information. The other one the MST extracted from the CEMP-estimated weights. Note that CEMP only uses 3-cycle information. We refer to these methods as IRLS and CEMP+IRLS respectively. + +Since the solution of absolute rotations is determined up to a global rotation, we align our estimated rotation $\{\hat{R}_i\}$ with $\{R_i^*\}$ by $R_{\mathrm{align}}$ that minimizes the $\ell_1$ rotation alignment error + +![](images/6968ab42c664c6e8f93e8d667cc771f80578269e0b14294c9b16ed4a19989a1c.jpg) + +![](images/669a3c1c2a165ca8dab3004f7386cecf1dcb8051b66a15b01b38f281cb7c17b2.jpg) +Figure 1. Average errors for IRLS, CEMP+IRLS and LongSync +IRLS with $c = 4,5$ , using the uniform corruption (top) and uniform bipartite corruption (bottom) models. The mean errors are measured in degrees. LongSync4 and LongSync5 refer to LongSync with 4 and 5 cycles, respectively. + +$\sum_{i\in [n]}\| \hat{\pmb{R}}_i\pmb{R}_{\mathrm{align}} - \pmb{R}_i^*\| _F$ . We report the mean estimation error in degrees: $180\cdot \sum_{i\in [n]}d_{\angle}(\hat{\pmb{R}}_i\pmb{R}_{\mathrm{align}},\pmb{R}_i^*) / n.$ + +# 5.1. Uniform Corruption Model + +For $\mathrm{UCM}(n,p,q_g)$ , we take $n = 200$ and $p = 1$ and corruption probability $q = 1 - q_{g}$ ranges from 0.86 to 0.92. We report the average mean estimation error from 20 trials of the uniform corruption model in the top panel of Figure 1. + +We note that LongSync uniformly improves IRLS, and the mean error of LongSync decreases as the cycle length increases. When $q > 0.86$ , the expected number of clean 3-cycles for each edge is less than 4, and therefore longer cycles are helpful. The numerical result aligns with our theory that using longer cycles may tolerate higher corruption with fixed graph size $n$ . + +# 5.2. Uniform Bipartite Corruption Model + +For UBCM, we first generate the graph and relative rotations by $\mathrm{UCM}(n,p,q_g)$ with $n = 200$ , $p = 1$ and $q = 1 - q_{g}$ ranging from 0.8 to 0.85. Then we split the nodes into two clusters of equal size and remove the intra-cluster edges for both clusters. The resulting graph is bipartite, where only cycles of even lengths exist. We report the mean estimation error from 20 trials in the bottom panel of Figure 1. + +We observe that LongSync with 4-cycles almost exactly recovers the rotations, while for other algorithms the rotation estimates are not even close to the ground truth. + +# 6. Real Data Experiment + +We test distributed synchronization with LongSync on the PhotoTourism dataset [50] to demonstrate its advantages in accuracy and speed over other baselines. PhotoTourism is a large scale dataset consisting of 15 sets of images taken for 3D reconstruction. The smallest dataset consists of 247 cameras, and the largest dataset consists of 5433 cameras. The input graph and initial pairwise rotation estimates are provided in the dataset. In the following, we first explain the common steps for distributed synchronization, and our improvement using LongSync. We then describe our graph processing method for filtering bad nodes and edges, which also is applicable to other baseline methods. + +# Steps in Distributed Synchronization: + +1. Graph partitioning. The first step involves partitioning the graph $G = ([n], E)$ into $K$ clusters $G_{i} = (V_{i}, E_{i}), i \in [K]$ . In this paper we apply spectral clustering algorithm [37] on the adjacency matrix $G$ , where $K = 0.6\sqrt{np}$ and $p = 2|E| / (n(n - 1))$ . + +2. Synchronization within clusters. Run standard synchronization solvers for each cluster. In this work, we use the current state-of-the-art method MPLS [39]. Note that for each camera $p$ in cluster $k$ , one can only estimate the true rotation $R_{p}^{*}$ up to a global rotation $R_{k}$ . Namely, one only obtains $\hat{R}^p \approx R_p^* R_k^{-1}$ where $R_{k}$ is unknown and is the same for all cameras in cluster $k$ . + +3. Estimation of inter-cluster rotations. To find $\pmb{R}_{p}^{*}$ of all cameras, one needs to solve $\pmb{R}_{k}$ for all clusters. Namely, one needs to rotate and stitch the solutions of all clusters so that they are in the same reference frame. To do this, it is common to first estimate the inter-clsuter rotations $\pmb{R}_{kl} := \pmb{R}_{k}\pmb{R}_{l}^{-1}$ between pairs of clusters $k,l$ , and then synchronize these relative rotations. To estimate each $\pmb{R}_{kl}$ , we note that $\pmb{R}_{kl} = \pmb{R}_{p}^{*-1}\pmb{R}_{pq}^{*}\pmb{R}_{q}^{*}$ for each $p\in V_k$ and $q\in V_l$ . Therefore, one can use the rotations in the set $S_{kl} := \{\hat{\pmb{R}}_p^{-1}\pmb{R}_{pq}\hat{\pmb{R}}_q\}_{p\in V_k,q\in V_l}$ to approximate $\pmb{R}_{kl}$ . We remark that this step is crucial to the overall performance of distributed methods, and we compare the following methods for solving $\pmb{R}_{kl}$ : + +- MultSync [9]: Run synchronization on a multi-graph where each edge $kl$ is assigned a set of relative rotations $\{\hat{R}_p^{-1}R_{pq}\hat{R}_q\}_{p\in V_k,q\in V_l}$ . This combines the step 3 and 4 in a unified least squares formulation. +- Edge averaging using IRLS: We initialize $\hat{R}_{kl}$ with the quaternion $\ell_2$ mean of the set $S_{kl}$ and refine it using $\ell_1$ -rotation averaging [19]. We refer to this method as IRLS in our comparison. +- Edge averaging using LongSync: We first perform LongSync with 4-cycles to estimate the weights of these inter-cluster edges (there are no 3-cycles for a bipartite graph). We next initialize $\hat{R}_{kl}$ as the quaternion weighted $\ell_2$ mean of $S_{kl}$ , using the edge weights from LongSync by their LongSync weights. Lastly, we refine the solution using [19]. + +4. Synchronization of inter-cluster rotations. This step is skipped for MultiSync. For other methods described in step 3, we find $\pmb{R}_k$ (up to a rotation) for each cluster $k$ from the estimated $\{R_{kl}\}_{k,l\in [K]}$ by MPLS. +5. Rotation merging. Finally, for each camera $p$ in cluster $k$ , the rotation estimate of $p$ is given by $R_p^{\mathrm{final}} = \hat{R}_p R_k^{-1}$ . + +Next, we introduce our graph processing method to further boost the performance of all tested methods. + +# Extra Improvement by Graph Processing: + +- Spectral clustering with Jaccard Index. For step 1, we use the Jaccard index matrix as the similarity matrix for spectral clustering, instead of the adjacency matrix. The $n \times n$ Jaccard index matrix $A_J$ is defined as follows: + +$$ +\boldsymbol {A} _ {J} (i, j) = \left\{ \begin{array}{l l} 0 & i j \notin E \\ \frac {\left| N _ {i} \cap N _ {j} \right|}{\left| N _ {i} \cup N _ {j} \right|} & i j \in E \end{array} \right. \tag {14} +$$ + +where $N_{i}$ and $N_{j}$ denote the sets of neighboring nodes of node $i$ and $j$ , respectively. In this way, $\mathbf{A}_{J}(i,j)$ is higher for the pair $ij$ contained in many 3-cycles, which is a more robust and nicely scaled statistics $(\in [0,1])$ for measuring the local graph density around edge $ij$ . + +- Refinement of intra-cluster edges and nodes. For step 2, after the MPLS step, we perform CEMP with 3-cycles to estimate the corruption level of the intra-cluster edges for each cluster. We remove a camera if the number of neighboring 'good' edges, i.e. the edges with corruption level less than 0.1, is less than 4. The numbers 4 and 0.1 are chosen to balance the number of remaining cameras and the quality of intra-cluster rotation estimates. In order to eliminate the sparsely connected components inside the cluster, we use the Matlab built-in hierarchical spectral clustering function on the remaining cameras with the 'cutoff' and 'depth' parameters as 2 and 4, and we keep the largest component. The absolute rotations for the remaining cameras are estimated by MPLS. We remark that one could replace CEMP by LongSync with 3-cycles. However, we have not observed significant difference in the performance. + +We respectively name MultiSync and IRLS with our new graph processing method as MultiSync(New) and IRLS(new). "LongSync" in our experiment refers to the full version of our algorithm: use LongSync weights for edge averaging in Step 3, with the graph processing step. We also compare with MPLS on the whole dataset, since it is a state-of-the-art non-distributed method, but we note that MPLS is significantly slower than all distributed methods. We report median error $180 \cdot$ median( $\{d_{\angle}(\hat{R}_i R_{\text{align}}, R_i^*)\}_{i \in [n]}$ of the tested methods on 14 datasets in Figure 2. We exclude the result of Gendarmenmarkt since all methods return large estimation errors in the figure. The full results, including that of mean error are included in the supplementary material. + +In Figure 3, for each distributed method, we report the ratio (in percentage) between its total runtime on all datasets and that of the non-distributed MPLS. Namely, we compute + +![](images/bda3fb0b64824766e835611341624dac709532541ead448fa1660bf7d5e90c98.jpg) +Figure 2. Median error for rotations for each dataset measured in degrees. + +![](images/baccd76e5c4d33968b5b02bea6b50a0629181129f71f2c34fc4ce00f4bdf8511.jpg) +Figure 3. Runtime ratio and average median and mean error gaps between the distributed methods and MPLS on the entire graph. + +$\sum_{d \in D} t_{\mathrm{dist},d} / \sum_{d \in D} t_{\mathrm{MPLS},d}$ , where $D$ is the set of 15 datasets, and each $t_{\mathrm{dist},d}$ and $t_{\mathrm{MPLS},d}$ is respectively the runtime of the distributed method and MPLS on data $d$ . In the same figure, we present the mean/median error gap between each distributed method and MPLS. The mean and median error gap is respectively defined as $(\bar{e}_{\mathrm{dist}} - \bar{e}_{\mathrm{MPLS}}) / \bar{e}_{\mathrm{dist}}$ and $(\hat{e}_{\mathrm{dist}} - \hat{e}_{\mathrm{MPLS}}) / \hat{e}_{\mathrm{dist}}$ , where $\bar{e}$ and $\hat{e}$ respectively denote the mean and median error over all cameras. + +From Figure 2 and 3, our method outperforms other distributed methods on 13 out of 15 datasets. The most significant improvement in mean error and median error are respectively $28.6\%$ and $46.4\%$ (in Notre Dame) compared to the best performing method between IRLS(new) and MultiSync(new). The improvement is even more significant when comparing to the original version of these baseline methods without our graph processing method. The only two datasets without improvement are Gendarmenmarkt and Union Square. Our method is comparable to others on Union Square, and all methods return large errors on Gendarmenmarkt due to many repetitive patterns in its 3D scene. + +The average mean and median error gap between our method and full MPLS are respectively $38.3\%$ and $18.4\%$ . Compared to the best performing method among others, our method reduces the average median error gap by $40.8\%$ , and the average mean error gap by $9.6\%$ . In terms of runtime, our method is uniformly faster than MultiSync and it is scalable on the largest dataset, taking less than $6\%$ of the total runtime of full MPLS. In conclusion, our method significantly improves the result of distributed rotation synchronization without compromising runtime. + +In the supplementary material, we further demonstrate the improvement by our new graph processing method, which significantly improves the results of LongSync (without extra graph processing) in 14 of the 15 datasets. On these 14 datasets, the average reduction on mean error is $59.2\%$ and the average reduction on median error is $28.5\%$ . + +# 7. Conclusion + +We propose LongSync, a robust and efficient algorithm for group synchronization. It modifies and vectorizes CEMP which enables efficient computation when using longer cycles. The theory we developed for LongSync is the strongest among all other existing results under UCM. Experiment shows that LongSync, together with our improved graph preprocessing method, achieves superior accuracy for distributed synchronization on large real datasets with competitive runtime. However, our method also has some limitations. First of all, in theory there is still a small gap of sample complexity from our method to the information theoretic one. Filling this gap is an open problem, which requires new tools and possibly more sophisticated analysis. Second, our graph preprocessing method is quite heuristic, and an automatic way of choosing parameters is needed. Our work also opens a door for some important future directions, including distributed partial permutation synchronization for multi-image matching, angular synchronization for Cryo-EM and Jigsaw Puzzles, and analysis of their algorithms. + +# References + +[1] Federica Arrigoni, Beatrice Rossi, Pasqualina Fragneto, and Andrea Fusiello. Robust synchronization in so(3) and se(3) via low-rank and sparse matrix decomposition. Computer Vision and Image Understanding, 174:95-113, 2018. 2 +[2] A. S. Bandeira, N. Boumal, and A. Singer. Tightness of the maximum likelihood semidefinite relaxation for angular synchronization. arXiv preprint arXiv:1411.3272, 2014. 2 +[3] Olivier Bousquet. A Bennett concentration inequality and its application to suprema of empirical processes. C. R. Math. Acad. Sci. Paris, 334(6):495-500, 2002. 5, 14, 16, 21 +[4] Avishek Chatterjee and Venu Madhav Govindu. Efficient and robust large-scale rotation averaging. In IEEE International Conference on Computer Vision, ICCV 2013, Sydney, Australia, December 1-8, 2013, pages 521-528, 2013. 2, 6 +[5] Avishek Chatterjee and Venu Madhav Govindu. Robust relative rotation averaging. IEEE Trans. Pattern Anal. Mach. Intell., 40(4):958-972, 2018. 2 +[6] Yuxin Chen, Changho Suh, and Andrea J. Goldsmith. Information recovery from pairwise measurements. IEEE Trans. Inf. Theory, 62(10):5881-5905, 2016. 6 +[7] Keresztély Corradi and András Hajnal. On the maximal number of independent circuits in a graph. Acta Mathematica Hungarica, 14(3-4):423-439, 1963. 5, 14 +[8] Mihai Cucuringu. Synchronization over Z2 and community detection in signed multiplex networks with constraints. J. Complex Networks, 3(3):469-506, 2015. 5 +[9] Andrea Porfiri Dal Cin, Luca Magri, Federica Arrigoni, Andrea Fusiello, and Giacomo Boracchi. Synchronization of group-labelled multi-graphs. In 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 6433-6443, 2021. 2, 7 +[10] Frank Dellaert, David M. Rosen, Jing Wu, Robert E. Mahony, and Luca Carlone. Shonan rotation averaging: Global optimality by surfing $\mathrm{so(p)^n}$ . In Andrea Vedaldi, Horst Bischof, Thomas Brox, and Jan-Michael Frahm, editors, Computer Vision - ECCV 2020 - 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part VI, volume 12351 of Lecture Notes in Computer Science, pages 292-308. Springer, 2020. 2 +[11] Anders Eriksson, Carl Olsson, Fredrik Kahl, and Tat-Jun Chin. Rotation averaging and strong duality, 2017. +[12] Anders P. Eriksson, Carl Olsson, Fredrik Kahl, and Tat-Jun Chin. Rotation averaging and strong duality. In 2018 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2018, Salt Lake City, UT, USA, June 18-22, 2018, pages 127-135. Computer Vision Foundation / IEEE Computer Society, 2018. +[13] Anders P. Eriksson, Carl Olsson, Fredrik Kahl, and Tat-Jun Chin. Rotation averaging with the chordal distance: Global minimizers and strong duality. IEEE Trans. Pattern Anal. Mach. Intell., 43(1):256-268, 2021. 2 +[14] Xiang Gao, Hainan Cui, Menghan Li, Zexiao Xie, and Shuhan Shen. Irav3: Hierarchical incremental rotation averaging on the fly. IEEE Trans. Circuits Syst. Video Technol., 33(4):2001-2006, 2023. 2 +[15] Xiang Gao, Hainan Cui, and Shuhan Shen. Incremental rotation averaging revisited and more: A new rotation averaging benchmark. arXiv preprint, abs/2309.16924, 2023. +[16] Xiang Gao, Lingjie Zhu, Hainan Cui, Zexiao Xie, and Shuhan Shen. IRA++: distributed incremental rotation averaging. IEEE Trans. Circuits Syst. Video Technol., 32(7):4885-4892, 2022. 2 + +[17] Thomas Goldstein, Paul Hand, Choongbum Lee, Vladislav Voroninski, and Stefano Soatto. Shapefit and shapekick for robust, scalable structure from motion. In Computer Vision - ECCV 2016 - 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part VII, pages 289-304, 2016. 1 +[18] V.M. Govindu. Lie-algebraic averaging for globally consistent motion estimation. In Proceedings of the 2004 IEEE Computer Society Conference on Computer Vision and Pattern Recognition, 2004. CVPR 2004., volume 1, pages I-I, 2004. 2 +[19] R. Hartley, K. Aftab, and J. Trumpf. L1 rotation averaging using the weiszfeld algorithm. In Computer Vision and Pattern Recognition (CVPR), 2011 IEEE Conference on, pages 3041-3048. IEEE, 2011. 7 +[20] Richard I. Hartley, Khurrum Aftab, and Jochen Trumpf. L1 rotation averaging using the weiszfeld algorithm. In The 24th IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2011, Colorado Springs, CO, USA, 20-25 June 2011, pages 3041-3048, 2011. 2 +[21] Xiangru Huang, Zhenxiao Liang, Xiaowei Zhou, Yao Xie, Leonidas J. Guibas, and Qixing Huang. Learning transformation synchronization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2019. 2 +[22] Svante Janson. Poisson approximation for large deviations. Random Structures & Algorithms, 1(2):221-229, 1990. 5, 14, 16 +[23] Svante Janson. Large deviations for sums of partly dependent random variables. Random Structures & Algorithms, 24(3):234-248, 2004. 18, 21 +[24] Jeong Han Kim and Van H Vu. Concentration of multivariate polynomials and its applications. Combinatorica, 20(3):417-434, 2000. 5, 14, 18 +[25] Gilad Lerman and Yunpeng Shi. Robust group synchronization via cycle-edge message passing. Foundations of Computational Mathematics, 22(6):1665-1741, 2022. 1, 2, 3, 5, 6, 13, 14, 16, 19, 21 +[26] Heng Li, Zhaopeng Cui, Shuaicheng Liu, and Ping Tan. RAGO: recurrent graph optimizer for multiple rotation averaging. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2022, New Orleans, LA, USA, June 18-24, 2022, pages 15766-15775. IEEE, 2022. 2 +[27] Shaohan Li, Yunpeng Shi, and Gilad Lerman. Fast, accurate and memory-efficient partial permutation synchronization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 15735-15743, June 2022. 2 +[28] Shuyang Ling. Near-optimal performance bounds for orthogonal and permutation group synchronization via spectral methods. arXiv preprint arXiv:2008.05341, 2020. 5 +[29] Huikang Liu, Xiao Li, and Anthony Man-Cho So. Resync: Riemannian subgradient-based robust rotation synchronization. In Advances in Neural Information Processing Systems: 37th Annual Conference on Neural Information Processing Systems, 2023. 2, 5, 6 +[30] Tyler Maunu and Gilad Lerman. Depth descent synchronization in SO(D). Int. J. Comput. Vis., 131(4):968-986, 2023. 2 +[31] Onur Özyesil and Amit Singer. Robust camera location estimation by convex programming. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 2674-2683, 2015. 1 +[32] Deepti Pachauri, Risi Kondor, and Vikas Singh. Solving the multi-way matching problem by permutation synchronization. In + +C. J. C. Burges, L. Bottou, M. Welling, Z. Ghahramani, and K. Q. Weinberger, editors, Advances in Neural Information Processing Systems 26, pages 1860-1868. Curran Associates, Inc., 2013. 5 +[33] Álvaro Parra, Shin-Fang Ch'ng, Tat-Jun Chin, Anders P. Eriksson, and Ian Reid. Rotation coordinate descent for fast globally optimal rotation averaging. In IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2021, virtual, June 19-25, 2021, pages 4298-4307. Computer Vision Foundation / IEEE, 2021. 2 +[34] Pulak Purkait, Tat-Jun Chin, and Ian Reid. *Neurora: Neural robust rotation averaging*, 2020. 2 +[35] Ian C Ross and Frank Harary. On the determination of redundancies in sociometric chains. Psychometrika, 17(2):195-208, 1952. 4, 12 +[36] Tianwei Shen, Siyu Zhu, Tian Fang, Runze Zhang, and Long Quan. Graph-based consistent matching for structure-from-motion. In European Conference on Computer Vision, pages 139-155. Springer, 2016. 2 +[37] Jianbo Shi and J. Malik. Normalized cuts and image segmentation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 22(8):888-905, 2000. 7 +[38] Yunpeng Shi and Gilad Lerman. Estimation of camera locations in highly corrupted scenarios: All about that base, no shape trouble. In 2018 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2018, Salt Lake City, UT, USA, June 18-22, 2018, pages 2868-2876, 2018. 2 +[39] Yunpeng Shi and Gilad Lerman. Message passing least squares framework and its application to rotation synchronization. In Proceedings of the 37th International Conference on Machine Learning (ICML), 2020. 1, 2, 7 +[40] Yunpeng Shi, Shaohan Li, and Gilad Lerman. Robust multi-object matching via iterative reweighting of the graph connection laplacian. Advances in Neural Information Processing Systems, 2020-December, 2020. 2 +[41] Chitturi Sidhartha and Venu Madhav Govindu. It is all in the weights: Robust rotation averaging revisited. In International Conference on 3D Vision, 3DV 2021, London, United Kingdom, December 1-3, 2021, pages 1134-1143. IEEE, 2021. 2 +[42] Amit Singer. Angular synchronization by eigenvectors and semidefinite programming. Applied and computational harmonic analysis, 30(1):20-36, 2011. 2, 5 +[43] A. Singer and Y. Shkolnisky. Angular synchronization by eigenvectors and semidefinite programming: Analysis and application to class averaging in cryo-electron microscopy, 2009. 2 +[44] Johan Thunberg, Florian Bernard, and Jorge Goncalves. Distributed methods for synchronization of orthogonal matrices over graphs. Automatica, 80:243-252, jun 2017. 2 +[45] Roberto Tron and René Vidal. Distributed image-based 3-d localization of camera sensor networks. In Proceedings of the 48th IEEE Conference on Decision and Control, CDC 2009, December 16-18, 2009, Shanghai, China, pages 901-908, 2009. 2 +[46] Anton Voropaev and Sergey Perepechko. The number of fixed length cycles in undirected graph. explicit formulae in case of small lengths. Bulletin of PFUR. Series Mathematics, Information Sciences, Physics, pages 5-11, 01 2012. 4 +[47] Anton Voropaev and Sergey Perepechko. The number of fixed length cycles in undirected graph. explicit formulae in case of small lengths. Bulletin of PFUR. Series Mathematics, Information Sciences, Physics, pages 5-11, 01 2012. 13 +[48] Van H Vu. A large deviation result on the number of small + +subgraphs of a random graph. Combinatorics, Probability and Computing, 10(1):79-94, 2001. 5, 14, 17 +[49] Lanhui Wang and Amit Singer. Exact and stable recovery of rotations for robust synchronization. Information and Inference, 2013. 2 +[50] Kyle Wilson and Noah Snavely. Robust global translations with 1dsfm. In Computer Vision - ECCV 2014 - 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part III, pages 61-75, 2014. 7 +[51] Christopher Zach, Manfred Klopschitz, and Marc Pollefeys. Disambiguating visual relations using loop constraints. In The Twenty-Third IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2010, San Francisco, CA, USA, 13-18 June 2010, pages 1426-1433, 2010. 2 \ No newline at end of file diff --git a/2024/Efficient Detection of Long Consistent Cycles and its Application to Distributed Synchronization/images.zip b/2024/Efficient Detection of Long Consistent Cycles and its Application to Distributed Synchronization/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..07156ffbdd8e95760582bc0d488a0f81f1df6eda --- /dev/null +++ b/2024/Efficient Detection of Long Consistent Cycles and its Application to Distributed Synchronization/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e79a39ff5eccab1813790e1864a8ea56fe321435c14a0a335a3566d4a2f608ce +size 307762 diff --git a/2024/Efficient Detection of Long Consistent Cycles and its Application to Distributed Synchronization/layout.json b/2024/Efficient Detection of Long Consistent Cycles and its Application to Distributed Synchronization/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..8a548fd335a245033a66259abaaeee3983d8ec29 --- /dev/null +++ b/2024/Efficient Detection of Long Consistent Cycles and its Application to Distributed Synchronization/layout.json @@ -0,0 +1,13822 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 147, + 103, + 451, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 147, + 103, + 451, + 140 + ], + "spans": [ + { + "bbox": [ + 147, + 103, + 451, + 140 + ], + "type": "text", + "content": "Efficient Detection of Long Consistent Cycles and its Application to Distributed Synchronization *" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 86, + 161, + 200, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 86, + 161, + 200, + 203 + ], + "spans": [ + { + "bbox": [ + 86, + 161, + 200, + 203 + ], + "type": "text", + "content": "Shaohan Li \nUniversity of Minnesota \nli000743@umn.edu" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 225, + 161, + 369, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 225, + 161, + 369, + 204 + ], + "spans": [ + { + "bbox": [ + 225, + 161, + 369, + 204 + ], + "type": "text", + "content": "Yunpeng Shi \nUniversity of California, Davis \nypshi@ucdavis.edu" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 394, + 161, + 508, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 394, + 161, + 508, + 203 + ], + "spans": [ + { + "bbox": [ + 394, + 161, + 508, + 203 + ], + "type": "text", + "content": "Gilad Lerman \nUniversity of Minnesota \nlerman@umn.edu" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 145, + 231, + 190, + 243 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 231, + 190, + 243 + ], + "spans": [ + { + "bbox": [ + 145, + 231, + 190, + 243 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 256, + 290, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 256, + 290, + 460 + ], + "spans": [ + { + "bbox": [ + 46, + 256, + 290, + 460 + ], + "type": "text", + "content": "Group synchronization plays a crucial role in global pipelines for Structure from Motion (SfM). Its formulation is nonconvex and it is faced with highly corrupted measurements. Cycle consistency has been effective in addressing these challenges. However, computationally efficient solutions are needed for cycles longer than three, especially in practical scenarios where 3-cycles are unavailable. To overcome this computational bottleneck, we propose an algorithm for group synchronization that leverages information from cycles of lengths ranging from three to six with a time complexity of order " + }, + { + "bbox": [ + 46, + 256, + 290, + 460 + ], + "type": "inline_equation", + "content": "O(n^{3})" + }, + { + "bbox": [ + 46, + 256, + 290, + 460 + ], + "type": "text", + "content": " (or " + }, + { + "bbox": [ + 46, + 256, + 290, + 460 + ], + "type": "inline_equation", + "content": "O(n^{2.373})" + }, + { + "bbox": [ + 46, + 256, + 290, + 460 + ], + "type": "text", + "content": " when using a faster matrix multiplication algorithm). We establish non-trivial theory for this and related methods that achieves competitive sample complexity, assuming the uniform corruption model. To advocate the practical need for our method, we consider distributed group synchronization, which requires at least 4-cycles, and we illustrate state-of-the-art performance by our method in this context." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 484, + 124, + 497 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 484, + 124, + 497 + ], + "spans": [ + { + "bbox": [ + 47, + 484, + 124, + 497 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 504, + 288, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 504, + 288, + 685 + ], + "spans": [ + { + "bbox": [ + 46, + 504, + 288, + 685 + ], + "type": "text", + "content": "Structure from Motion (SfM) asks to recover the 3D structure of a stationary scene from multiple images taken by cameras from different orientations and locations. In the past decade, the global SfM pipeline has become increasingly popular due to its several advantages over the incremental pipelines [17, 31]. First of all, global SfM requires only one implementation of bundle adjustment, making it more efficient in computation. Second, it estimates camera poses in a global optimization framework which mitigates the drifting issue of the incremental pipelines. Despite the popularity of global SfM pipelines, the estimation of global camera poses (e.g., orientations) remains a highly challenging problem. For instance, estimating camera orientations from their relative measurements, often called rotation synchronization, is a highly nonconvex graph optimization problem. In typical scenarios of highly noisy or corrupted measurements" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 232, + 545, + 256 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 232, + 545, + 256 + ], + "spans": [ + { + "bbox": [ + 306, + 232, + 545, + 256 + ], + "type": "text", + "content": "of relative orientations, many common solutions of rotation synchronization have poor accuracy and slow convergence." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 257, + 547, + 508 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 257, + 547, + 508 + ], + "spans": [ + { + "bbox": [ + 304, + 257, + 547, + 508 + ], + "type": "text", + "content": "Given these challenges, theoretical developments have demonstrated the critical role of cycle-consistency information in inferring corrupted measurements [25]. In practice, the consistency constraint on 3-cycles was utilized to estimate the error of each measured relative orientation. It also helped nonconvex iterative rotation synchronization solvers avoid spurious local minima and achieve significantly higher accuracy [39]. However, the usage of 3-cycles largely limits the application of these improved algorithms to other important scenarios. One scenario involves a sparse viewing graph lacking sufficient 3-cycles. This often occurs when the size of the graph is too large to densely measure the relative orientations on its edges, which could happen in certain case for the molecular orientation estimation in cryo-electron microscopy imaging. Another scenario is orientation estimation for each piece of jigsaw puzzles, where the graph is a 2D lattice and 3-cycle does not exist. Lastly, in distributed SfM, edges between any two clusters of nodes form a bipartite graph, and cycles of odd length do not exist. Our numerical results primarily emphasize the practical scenario of distributed SfM, which holds particular relevance for the broader computer vision community." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 509, + 548, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 509, + 548, + 640 + ], + "spans": [ + { + "bbox": [ + 304, + 509, + 548, + 640 + ], + "type": "text", + "content": "Despite the multiple critical applications of long-cycle consistency, inferring measurement noise from long cycles is challenging in both computation and theory. First of all, the number of cycles grows exponentially with the cycle length, and measuring cycle inconsistencies for each long cycle is computationally intractable. Moreover, developing theoretical guarantees for long-cycle inference methods is fundamentally more difficult than the 3-cycle case. Indeed, in a random graph setting, a set of longer cycles are more likely to share common edges, making their consistency score highly correlated. Therefore, new tools are required to handle the correlated empirical process." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 642, + 548, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 642, + 548, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 642, + 548, + 714 + ], + "type": "text", + "content": "In this work, we propose the first practical method, LongSync, for inferring edge corruption levels from long cycle consistency information. For this purpose, we carefully modify and vectorize the Cycle Edge Message Passing (CEMP) method [25]. This nontrivial modification drastically reduces its computational complexity when using longer cycles. Moreover," + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 693, + 237, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 693, + 237, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 693, + 237, + 713 + ], + "type": "text", + "content": "*This work was supported by NSF award DMS-2152766. Supplementary code: https://github.com/sli743/LongSync" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "5260" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "type": "text", + "content": "by employing a more delicate analysis and incorporating new tools from probability theory and combinatorics, we develop a significantly stronger exact recovery result with a general cycle length under a popular probabilistic model. The sample complexity in our theory is the lowest among all practical rotation synchronization methods. Although we limit our scope to the application of distributed SfM, our algorithm and theory applies to any finite-dimensional linear group in group synchronization, and not just " + }, + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "type": "inline_equation", + "content": "SO(3)" + }, + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "type": "text", + "content": " in rotation synchronization." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 190, + 135, + 201 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 190, + 135, + 201 + ], + "spans": [ + { + "bbox": [ + 47, + 190, + 135, + 201 + ], + "type": "text", + "content": "1.1. Related Work" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 209, + 288, + 316 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 209, + 288, + 316 + ], + "spans": [ + { + "bbox": [ + 46, + 209, + 288, + 316 + ], + "type": "text", + "content": "Earlier rotation synchronization methods [2, 10-13, 18, 33, 42, 43] seek to minimize a least squares energy function. They can be described as relaxed versions of the maximum likelihood estimator under the additive Gaussian noise model, but they are not robust in the presence of outliers or heavy-tailed noise. Nevertheless, in the case of global SfM, the initially estimated relative camera rotations can be severely corrupted due to the erroneous keypoint matches and the subsequent poor estimation of fundamental matrices." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 318, + 288, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 318, + 288, + 557 + ], + "spans": [ + { + "bbox": [ + 46, + 318, + 288, + 557 + ], + "type": "text", + "content": "To handle outliers, robust rotation synchronization methods either minimize a robust energy function or reweigh/trim the viewing graph based on the corruption levels of the edges. Wang and Singer [49] minimizes a corresponding " + }, + { + "bbox": [ + 46, + 318, + 288, + 557 + ], + "type": "inline_equation", + "content": "\\ell_1" + }, + { + "bbox": [ + 46, + 318, + 288, + 557 + ], + "type": "text", + "content": " objective function using semidefinite programming (SDP) relaxation, which is slow in practice. Other energy minimization methods are typically nonconvex, which include the Weiszfeld algorithm [20] and the Riemannian subgradient method [29] for " + }, + { + "bbox": [ + 46, + 318, + 288, + 557 + ], + "type": "inline_equation", + "content": "\\ell_1" + }, + { + "bbox": [ + 46, + 318, + 288, + 557 + ], + "type": "text", + "content": " minimization, and the iteratively reweighted least squares (IRLS) for minimizing general " + }, + { + "bbox": [ + 46, + 318, + 288, + 557 + ], + "type": "inline_equation", + "content": "\\ell_p" + }, + { + "bbox": [ + 46, + 318, + 288, + 557 + ], + "type": "text", + "content": " [5] and Geman-McClure [4] loss functions. However, all these methods heavily rely on good initialization. Sidhartha and Govindu [41] partially remedy the issue using adaptive Geman-McClure loss functions, but their approach remains sensitive to the initialized weights. Maunu and Lerman [30] propose to solve rotation synchronization by an iterative robust averaging method that utilizes Tukey depth, but they have not demonstrated effective performance for real SfM applications. Arrigoni et al. [1] applies a low-rank and sparse matrix decomposition method to SO(3) and SE(3) synchronization, but it is even less robust to outliers than IRLS." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 558, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 558, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 558, + 288, + 714 + ], + "type": "text", + "content": "Instead of employing a robust objective function, Shen et al. [36] and Zach et al. [51] uses the 3-cycle consistency constraint to detect and remove corrupted relative orientations. Lerman and Shi [25] take one step further to estimate the corruption level of each relative measurement by a novel cycle-edge message passing (CEMP) algorithm. They then use the estimated corruption levels to reweigh the graph and solve rotation synchronization using a weighted least squares method. This message passing procedure was further combined with IRLS to boost its accuracy in [39]. Particular versions and extensions of this procedure for permutation and partial permutation synchronization, which are relevant to the matching component of SfM, were discussed in [27, 40]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 72, + 547, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 228 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 228 + ], + "type": "text", + "content": "However, all the previously mentioned cycle-based methods [25, 27, 36, 38-40, 51] only use 3-cycles in practice, limiting their application for distributed synchronization. Indeed, the standard distributed synchronization often requires \"stitching\" local solutions by synchronizing the relative rotations between clusters. Each of these inter-cluster rotations is estimated by \"averaging\" the edges between the two clusters. These edges form a bipartite graph, and the minimal cycle length is 4. As pointed in [51], the number of operations for computing long cycle consistency information scales exponentially with the cycle length. Therefore, none of the existing distributed rotation synchronization methods directly exploits long cycle information due to this computational challenge." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 228, + 547, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 228, + 547, + 396 + ], + "spans": [ + { + "bbox": [ + 304, + 228, + 547, + 396 + ], + "type": "text", + "content": "The earlier distributed methods for " + }, + { + "bbox": [ + 304, + 228, + 547, + 396 + ], + "type": "inline_equation", + "content": "SO(d)" + }, + { + "bbox": [ + 304, + 228, + 547, + 396 + ], + "type": "text", + "content": " synchronization, such as [45] and [44], minimize a least squares energy and are not robust to outliers. A series of distributed SfM methods [14-16] implement incremental SfM algorithms for each cluster. However, these methods do not employ a standard rotation synchronization algorithm, as they require additional information such as the number of keypoint matches between images. Moreover, the incremental methods are slower since they require multiple rounds of global rotation synchronization. MultiSync [9] synchronizes the inter-cluster rotations directly using all inter-cluster edges among all clusters, by formulating a novel synchronization problem on a multi-graph. Although it utilizes a more unified formulation, its objective function is least squares which largely limits its robustness to outliers." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 397, + 545, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 397, + 545, + 456 + ], + "spans": [ + { + "bbox": [ + 304, + 397, + 545, + 456 + ], + "type": "text", + "content": "A recent and different type of methods for rotation synchronization use deep learning [21, 26, 34]. However, these methods are supervised and thus may not generalize well when switching datasets. Moreover, like many other previous methods, they lack theoretical guarantees." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 458, + 545, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 458, + 545, + 517 + ], + "spans": [ + { + "bbox": [ + 304, + 458, + 545, + 517 + ], + "type": "text", + "content": "A common theoretical setting to assess the performance of rotation synchronization algorithms is the uniform corruption model (UCM) described in §4. We provide the best sample complexity for LongSync, even with only 3 cycles, among all previously established estimates for the UCM model." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 527, + 459, + 539 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 527, + 459, + 539 + ], + "spans": [ + { + "bbox": [ + 306, + 527, + 459, + 539 + ], + "type": "text", + "content": "1.2. Contributions of This Work" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 545, + 547, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 306, + 545, + 545, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 545, + 545, + 605 + ], + "spans": [ + { + "bbox": [ + 306, + 545, + 545, + 605 + ], + "type": "text", + "content": "- We propose the first practical algorithm that infers edge corruption levels from long cycle consistency information. The computation complexity of our method is reduced from " + }, + { + "bbox": [ + 306, + 545, + 545, + 605 + ], + "type": "inline_equation", + "content": "O(n^{c})" + }, + { + "bbox": [ + 306, + 545, + 545, + 605 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 306, + 545, + 545, + 605 + ], + "type": "inline_equation", + "content": "O(n^{3})" + }, + { + "bbox": [ + 306, + 545, + 545, + 605 + ], + "type": "text", + "content": " (or possibly " + }, + { + "bbox": [ + 306, + 545, + 545, + 605 + ], + "type": "inline_equation", + "content": "O(n^{2.373})" + }, + { + "bbox": [ + 306, + 545, + 545, + 605 + ], + "type": "text", + "content": ") for cycle length " + }, + { + "bbox": [ + 306, + 545, + 545, + 605 + ], + "type": "inline_equation", + "content": "c \\leq 6" + }, + { + "bbox": [ + 306, + 545, + 545, + 605 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 306, + 545, + 545, + 605 + ], + "type": "inline_equation", + "content": "O(n^{[(c + 3) / 2]})" + }, + { + "bbox": [ + 306, + 545, + 545, + 605 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 306, + 545, + 545, + 605 + ], + "type": "inline_equation", + "content": "c > 6" + }, + { + "bbox": [ + 306, + 545, + 545, + 605 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 606, + 547, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 606, + 547, + 666 + ], + "spans": [ + { + "bbox": [ + 306, + 606, + 547, + 666 + ], + "type": "text", + "content": "- We establish sample complexity estimates for our method under the uniform corruption model, where we get closer to the information theoretic bound than any other existing work. Our proof requires delicate analysis and it also improves previous estimates for the CEMP algorithm." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 666, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 666, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 666, + 545, + 713 + ], + "type": "text", + "content": "- We introduce a new graph partition and graph preprocessing method that utilizes our inference method, and demonstrate the effectiveness of our pipeline in boosting the performance of distributed synchronization." + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "5261" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 287, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 287, + 96 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 287, + 96 + ], + "type": "text", + "content": "- Extensive numerical experiments demonstrate the outstanding performance of our method." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 106, + 257, + 118 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 106, + 257, + 118 + ], + "spans": [ + { + "bbox": [ + 47, + 106, + 257, + 118 + ], + "type": "text", + "content": "2. Problem Formulation and Preliminaries" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 126, + 289, + 342 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 126, + 289, + 342 + ], + "spans": [ + { + "bbox": [ + 47, + 126, + 289, + 342 + ], + "type": "text", + "content": "Assume a graph " + }, + { + "bbox": [ + 47, + 126, + 289, + 342 + ], + "type": "inline_equation", + "content": "G = ([n], E)" + }, + { + "bbox": [ + 47, + 126, + 289, + 342 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 47, + 126, + 289, + 342 + ], + "type": "inline_equation", + "content": "[n]" + }, + { + "bbox": [ + 47, + 126, + 289, + 342 + ], + "type": "text", + "content": " is the set of nodes indexed by " + }, + { + "bbox": [ + 47, + 126, + 289, + 342 + ], + "type": "inline_equation", + "content": "\\{1, 2, \\dots, n\\}" + }, + { + "bbox": [ + 47, + 126, + 289, + 342 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 126, + 289, + 342 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 47, + 126, + 289, + 342 + ], + "type": "text", + "content": " is the set of edges. Given a mathematical group " + }, + { + "bbox": [ + 47, + 126, + 289, + 342 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 47, + 126, + 289, + 342 + ], + "type": "text", + "content": ", each graph node is assigned an underlying ground truth group element " + }, + { + "bbox": [ + 47, + 126, + 289, + 342 + ], + "type": "inline_equation", + "content": "R_{i}^{*}" + }, + { + "bbox": [ + 47, + 126, + 289, + 342 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 47, + 126, + 289, + 342 + ], + "type": "inline_equation", + "content": "R_{i}^{*} \\in \\mathcal{G}" + }, + { + "bbox": [ + 47, + 126, + 289, + 342 + ], + "type": "text", + "content": " and we use star superscript to emphasize the ground truth. For each edge " + }, + { + "bbox": [ + 47, + 126, + 289, + 342 + ], + "type": "inline_equation", + "content": "ij \\in E" + }, + { + "bbox": [ + 47, + 126, + 289, + 342 + ], + "type": "text", + "content": ", we observe a relative group ratio " + }, + { + "bbox": [ + 47, + 126, + 289, + 342 + ], + "type": "inline_equation", + "content": "R_{ij} \\in \\mathcal{G}" + }, + { + "bbox": [ + 47, + 126, + 289, + 342 + ], + "type": "text", + "content": ", whose clean counterpart is " + }, + { + "bbox": [ + 47, + 126, + 289, + 342 + ], + "type": "inline_equation", + "content": "R_{ij}^{*} = R_{i}^{*}R_{j}^{*-1}" + }, + { + "bbox": [ + 47, + 126, + 289, + 342 + ], + "type": "text", + "content": ". Group synchronization aims to recover the ground truth group elements " + }, + { + "bbox": [ + 47, + 126, + 289, + 342 + ], + "type": "inline_equation", + "content": "\\{R_{i}^{*}\\}_{i \\in [n]}" + }, + { + "bbox": [ + 47, + 126, + 289, + 342 + ], + "type": "text", + "content": " from the possibly noisy and corrupted measurements " + }, + { + "bbox": [ + 47, + 126, + 289, + 342 + ], + "type": "inline_equation", + "content": "\\{R_{ij}\\}_{ij \\in E}" + }, + { + "bbox": [ + 47, + 126, + 289, + 342 + ], + "type": "text", + "content": ". In this paper, we focus on the case of rotation synchronization, which is a special case of group synchronization with " + }, + { + "bbox": [ + 47, + 126, + 289, + 342 + ], + "type": "inline_equation", + "content": "\\mathcal{G} = \\mathrm{SO}(d)" + }, + { + "bbox": [ + 47, + 126, + 289, + 342 + ], + "type": "text", + "content": ". For applications in camera orientation synchronization (" + }, + { + "bbox": [ + 47, + 126, + 289, + 342 + ], + "type": "inline_equation", + "content": "d = 3" + }, + { + "bbox": [ + 47, + 126, + 289, + 342 + ], + "type": "text", + "content": "), we estimate absolute rotations for each node " + }, + { + "bbox": [ + 47, + 126, + 289, + 342 + ], + "type": "inline_equation", + "content": "i \\in [n]" + }, + { + "bbox": [ + 47, + 126, + 289, + 342 + ], + "type": "text", + "content": " from measured relative rotations of edges in " + }, + { + "bbox": [ + 47, + 126, + 289, + 342 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 47, + 126, + 289, + 342 + ], + "type": "text", + "content": ". Note that since " + }, + { + "bbox": [ + 47, + 126, + 289, + 342 + ], + "type": "inline_equation", + "content": "\\{R_{i}^{*}\\}_{i \\in [n]}" + }, + { + "bbox": [ + 47, + 126, + 289, + 342 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 126, + 289, + 342 + ], + "type": "inline_equation", + "content": "\\{R_{i}^{*}R_{0}\\}_{i \\in [n]}" + }, + { + "bbox": [ + 47, + 126, + 289, + 342 + ], + "type": "text", + "content": " generate the same set of relative rotations, one can only estimate " + }, + { + "bbox": [ + 47, + 126, + 289, + 342 + ], + "type": "inline_equation", + "content": "\\{R_{i}^{*}\\}_{i \\in [n]}" + }, + { + "bbox": [ + 47, + 126, + 289, + 342 + ], + "type": "text", + "content": " up to a global rotation. The generalization to any linear groups is discussed in the supplementary material." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 348, + 189, + 360 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 348, + 189, + 360 + ], + "spans": [ + { + "bbox": [ + 47, + 348, + 189, + 360 + ], + "type": "text", + "content": "2.1. Notations and Definitions" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 366, + 287, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 366, + 287, + 403 + ], + "spans": [ + { + "bbox": [ + 46, + 366, + 287, + 403 + ], + "type": "text", + "content": "We denote the adjacency matrix of graph " + }, + { + "bbox": [ + 46, + 366, + 287, + 403 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 46, + 366, + 287, + 403 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 46, + 366, + 287, + 403 + ], + "type": "inline_equation", + "content": "\\mathbf{A}" + }, + { + "bbox": [ + 46, + 366, + 287, + 403 + ], + "type": "text", + "content": ", and form a pairwise observation matrix " + }, + { + "bbox": [ + 46, + 366, + 287, + 403 + ], + "type": "inline_equation", + "content": "\\mathbf{R} \\in \\mathbb{R}^{dn \\times dn}" + }, + { + "bbox": [ + 46, + 366, + 287, + 403 + ], + "type": "text", + "content": " by stacking the " + }, + { + "bbox": [ + 46, + 366, + 287, + 403 + ], + "type": "inline_equation", + "content": "\\mathbf{R}_{ij}" + }, + { + "bbox": [ + 46, + 366, + 287, + 403 + ], + "type": "text", + "content": "'s (for " + }, + { + "bbox": [ + 46, + 366, + 287, + 403 + ], + "type": "inline_equation", + "content": "ij \\notin E" + }, + { + "bbox": [ + 46, + 366, + 287, + 403 + ], + "type": "text", + "content": ", set " + }, + { + "bbox": [ + 46, + 366, + 287, + 403 + ], + "type": "inline_equation", + "content": "\\mathbf{R}_{ij} = \\mathbf{0}_{3 \\times 3}" + }, + { + "bbox": [ + 46, + 366, + 287, + 403 + ], + "type": "text", + "content": "):" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 99, + 410, + 236, + 467 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 99, + 410, + 236, + 467 + ], + "spans": [ + { + "bbox": [ + 99, + 410, + 236, + 467 + ], + "type": "interline_equation", + "content": "\\boldsymbol {R} := \\left( \\begin{array}{c c c c} \\boldsymbol {R} _ {1 1} & \\boldsymbol {R} _ {1 2} & \\dots & \\boldsymbol {R} _ {1 n} \\\\ \\boldsymbol {R} _ {2 1} & \\boldsymbol {R} _ {2 2} & \\dots & \\boldsymbol {R} _ {2 n} \\\\ \\vdots & \\vdots & \\ddots & \\vdots \\\\ \\boldsymbol {R} _ {n 1} & \\boldsymbol {R} _ {n 2} & \\dots & \\boldsymbol {R} _ {n n} \\end{array} \\right).", + "image_path": "f1da903e9dd8f73f107f8e4bfe57e7d590e7864dd3d7bfaf10b94fc955943125.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 473, + 288, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 473, + 288, + 568 + ], + "spans": [ + { + "bbox": [ + 47, + 473, + 288, + 568 + ], + "type": "text", + "content": "We list the matrix operations used in the paper. For matrices " + }, + { + "bbox": [ + 47, + 473, + 288, + 568 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 47, + 473, + 288, + 568 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 473, + 288, + 568 + ], + "type": "inline_equation", + "content": "Y" + }, + { + "bbox": [ + 47, + 473, + 288, + 568 + ], + "type": "text", + "content": ", the operations " + }, + { + "bbox": [ + 47, + 473, + 288, + 568 + ], + "type": "inline_equation", + "content": "X \\otimes Y" + }, + { + "bbox": [ + 47, + 473, + 288, + 568 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 47, + 473, + 288, + 568 + ], + "type": "inline_equation", + "content": "X \\odot Y" + }, + { + "bbox": [ + 47, + 473, + 288, + 568 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 47, + 473, + 288, + 568 + ], + "type": "inline_equation", + "content": "X \\oslash Y" + }, + { + "bbox": [ + 47, + 473, + 288, + 568 + ], + "type": "text", + "content": " respectively denote the Kronecker product, Hardmard (element-wise) multiplication and Harnard division between " + }, + { + "bbox": [ + 47, + 473, + 288, + 568 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 47, + 473, + 288, + 568 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 473, + 288, + 568 + ], + "type": "inline_equation", + "content": "Y" + }, + { + "bbox": [ + 47, + 473, + 288, + 568 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 47, + 473, + 288, + 568 + ], + "type": "inline_equation", + "content": "X^{\\odot k}" + }, + { + "bbox": [ + 47, + 473, + 288, + 568 + ], + "type": "text", + "content": " denotes the element-wise matrix " + }, + { + "bbox": [ + 47, + 473, + 288, + 568 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 47, + 473, + 288, + 568 + ], + "type": "text", + "content": "-power. For block matrices, " + }, + { + "bbox": [ + 47, + 473, + 288, + 568 + ], + "type": "inline_equation", + "content": "\\langle X, Y \\rangle_{\\mathrm{block}}" + }, + { + "bbox": [ + 47, + 473, + 288, + 568 + ], + "type": "text", + "content": " denotes the blockwise inner product of " + }, + { + "bbox": [ + 47, + 473, + 288, + 568 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 47, + 473, + 288, + 568 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 473, + 288, + 568 + ], + "type": "inline_equation", + "content": "Y" + }, + { + "bbox": [ + 47, + 473, + 288, + 568 + ], + "type": "text", + "content": ", i.e. " + }, + { + "bbox": [ + 47, + 473, + 288, + 568 + ], + "type": "inline_equation", + "content": "\\langle X, Y \\rangle_{\\mathrm{block}}(i,j) = \\langle X[i,j], Y[i,j] \\rangle" + }, + { + "bbox": [ + 47, + 473, + 288, + 568 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 47, + 473, + 288, + 568 + ], + "type": "inline_equation", + "content": "[i,j]" + }, + { + "bbox": [ + 47, + 473, + 288, + 568 + ], + "type": "text", + "content": " refers to the corresponding block of the matrix." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 575, + 209, + 588 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 575, + 209, + 588 + ], + "spans": [ + { + "bbox": [ + 47, + 575, + 209, + 588 + ], + "type": "text", + "content": "2.2. Review of CEMP for " + }, + { + "bbox": [ + 47, + 575, + 209, + 588 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 47, + 575, + 209, + 588 + ], + "type": "text", + "content": "-Cycles" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 594, + 287, + 654 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 594, + 287, + 654 + ], + "spans": [ + { + "bbox": [ + 47, + 594, + 287, + 654 + ], + "type": "text", + "content": "We assume the above setting of " + }, + { + "bbox": [ + 47, + 594, + 287, + 654 + ], + "type": "inline_equation", + "content": "SO(d)" + }, + { + "bbox": [ + 47, + 594, + 287, + 654 + ], + "type": "text", + "content": " synchronization. Let " + }, + { + "bbox": [ + 47, + 594, + 287, + 654 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 47, + 594, + 287, + 654 + ], + "type": "text", + "content": " be any bi-invariant metric on " + }, + { + "bbox": [ + 47, + 594, + 287, + 654 + ], + "type": "inline_equation", + "content": "SO(d)" + }, + { + "bbox": [ + 47, + 594, + 287, + 654 + ], + "type": "text", + "content": ". We assume a fixed number of cycles, " + }, + { + "bbox": [ + 47, + 594, + 287, + 654 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 47, + 594, + 287, + 654 + ], + "type": "text", + "content": ", and denote by " + }, + { + "bbox": [ + 47, + 594, + 287, + 654 + ], + "type": "inline_equation", + "content": "N_{ij}^{c}" + }, + { + "bbox": [ + 47, + 594, + 287, + 654 + ], + "type": "text", + "content": " the set of simple cycles of length " + }, + { + "bbox": [ + 47, + 594, + 287, + 654 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 47, + 594, + 287, + 654 + ], + "type": "text", + "content": " (or simple " + }, + { + "bbox": [ + 47, + 594, + 287, + 654 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 47, + 594, + 287, + 654 + ], + "type": "text", + "content": "-cycles) containing edge " + }, + { + "bbox": [ + 47, + 594, + 287, + 654 + ], + "type": "inline_equation", + "content": "ij" + }, + { + "bbox": [ + 47, + 594, + 287, + 654 + ], + "type": "text", + "content": ". CEMP [25] aims to estimate for each edge " + }, + { + "bbox": [ + 47, + 594, + 287, + 654 + ], + "type": "inline_equation", + "content": "ij" + }, + { + "bbox": [ + 47, + 594, + 287, + 654 + ], + "type": "text", + "content": " the corruption level" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 129, + 661, + 287, + 676 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 129, + 661, + 287, + 676 + ], + "spans": [ + { + "bbox": [ + 129, + 661, + 287, + 676 + ], + "type": "interline_equation", + "content": "s _ {i j} ^ {*} = \\mathcal {D} \\left(\\boldsymbol {R} _ {i j}, \\boldsymbol {R} _ {i j} ^ {*}\\right), \\tag {1}", + "image_path": "8f6f1eaeff910fbc45f856edc91dedd6caf1e63908680d54346bd3567d619fdc.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 681, + 219, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 681, + 219, + 693 + ], + "spans": [ + { + "bbox": [ + 47, + 681, + 219, + 693 + ], + "type": "text", + "content": "from the set of cycle inconsistency measures" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 131, + 701, + 287, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 701, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 131, + 701, + 287, + 715 + ], + "type": "interline_equation", + "content": "d _ {L} = \\mathcal {D} \\left(\\boldsymbol {R} _ {L}, \\boldsymbol {R} _ {i j}\\right) \\tag {2}", + "image_path": "f4d3980a65043ca9becb9866f8ba1c673243059efeeb503f8d0728368924dc7e.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 72, + 545, + 132 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 132 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 132 + ], + "type": "text", + "content": "where cycle " + }, + { + "bbox": [ + 305, + 72, + 545, + 132 + ], + "type": "inline_equation", + "content": "L = (ik_{1}, k_{1}k_{2}, \\dots, k_{c-2}j, ji) \\in N_{ij}^{c}" + }, + { + "bbox": [ + 305, + 72, + 545, + 132 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 72, + 545, + 132 + ], + "type": "inline_equation", + "content": "R_{L} := R_{ik_{1}}R_{k_{1}k_{2}}\\dots R_{k_{c-2}j}" + }, + { + "bbox": [ + 305, + 72, + 545, + 132 + ], + "type": "text", + "content": ". The estimated " + }, + { + "bbox": [ + 305, + 72, + 545, + 132 + ], + "type": "inline_equation", + "content": "s_{ij}^{*}" + }, + { + "bbox": [ + 305, + 72, + 545, + 132 + ], + "type": "text", + "content": " can then be used for extracting a clean subgraph, or to implement a weighted least squares solver where higher weights are assigned to cleaner edges." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 133, + 547, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 133, + 547, + 156 + ], + "spans": [ + { + "bbox": [ + 305, + 133, + 547, + 156 + ], + "type": "text", + "content": "It is obvious that if all the edges in " + }, + { + "bbox": [ + 305, + 133, + 547, + 156 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 305, + 133, + 547, + 156 + ], + "type": "text", + "content": " are clean then " + }, + { + "bbox": [ + 305, + 133, + 547, + 156 + ], + "type": "inline_equation", + "content": "d_{L} = 0" + }, + { + "bbox": [ + 305, + 133, + 547, + 156 + ], + "type": "text", + "content": ". Moreover, due to bi-invariance of " + }, + { + "bbox": [ + 305, + 133, + 547, + 156 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 305, + 133, + 547, + 156 + ], + "type": "text", + "content": ", the following holds true" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 371, + 163, + 545, + 178 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 371, + 163, + 545, + 178 + ], + "spans": [ + { + "bbox": [ + 371, + 163, + 545, + 178 + ], + "type": "interline_equation", + "content": "d _ {L} = s _ {i j} ^ {*} \\text {w h e n e v e r} L \\in G _ {i j} ^ {c}, \\tag {3}", + "image_path": "a345bcbfb619f117a3202dc3b823f5584c403ef34b33e489bdc4f63a656e560a.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 183, + 547, + 219 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 183, + 547, + 219 + ], + "spans": [ + { + "bbox": [ + 304, + 183, + 547, + 219 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 183, + 547, + 219 + ], + "type": "inline_equation", + "content": "G_{ij}^{c}" + }, + { + "bbox": [ + 304, + 183, + 547, + 219 + ], + "type": "text", + "content": " is the set of good " + }, + { + "bbox": [ + 304, + 183, + 547, + 219 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 304, + 183, + 547, + 219 + ], + "type": "text", + "content": "-cycles with respect to " + }, + { + "bbox": [ + 304, + 183, + 547, + 219 + ], + "type": "inline_equation", + "content": "ij" + }, + { + "bbox": [ + 304, + 183, + 547, + 219 + ], + "type": "text", + "content": ", i.e. the set of cycles " + }, + { + "bbox": [ + 304, + 183, + 547, + 219 + ], + "type": "inline_equation", + "content": "L \\in N_{ij}^{c}" + }, + { + "bbox": [ + 304, + 183, + 547, + 219 + ], + "type": "text", + "content": " such that " + }, + { + "bbox": [ + 304, + 183, + 547, + 219 + ], + "type": "inline_equation", + "content": "ik_1, \\dots, k_{c-2j}" + }, + { + "bbox": [ + 304, + 183, + 547, + 219 + ], + "type": "text", + "content": " are clean. This gives a sufficient condition for " + }, + { + "bbox": [ + 304, + 183, + 547, + 219 + ], + "type": "inline_equation", + "content": "d_L" + }, + { + "bbox": [ + 304, + 183, + 547, + 219 + ], + "type": "text", + "content": " to be an exact estimator of " + }, + { + "bbox": [ + 304, + 183, + 547, + 219 + ], + "type": "inline_equation", + "content": "s_{ij}^{*}" + }, + { + "bbox": [ + 304, + 183, + 547, + 219 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 220, + 545, + 269 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 220, + 545, + 269 + ], + "spans": [ + { + "bbox": [ + 304, + 220, + 545, + 269 + ], + "type": "text", + "content": "To estimate the corruption levels of each edge " + }, + { + "bbox": [ + 304, + 220, + 545, + 269 + ], + "type": "inline_equation", + "content": "ij" + }, + { + "bbox": [ + 304, + 220, + 545, + 269 + ], + "type": "text", + "content": ", CEMP initializes the edge weight of each " + }, + { + "bbox": [ + 304, + 220, + 545, + 269 + ], + "type": "inline_equation", + "content": "ij \\in E" + }, + { + "bbox": [ + 304, + 220, + 545, + 269 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 304, + 220, + 545, + 269 + ], + "type": "inline_equation", + "content": "w_{ij}^{(0)} = 1" + }, + { + "bbox": [ + 304, + 220, + 545, + 269 + ], + "type": "text", + "content": ". It then iteratively updates the corruption level estimate as the following convex combination of " + }, + { + "bbox": [ + 304, + 220, + 545, + 269 + ], + "type": "inline_equation", + "content": "d_L" + }, + { + "bbox": [ + 304, + 220, + 545, + 269 + ], + "type": "text", + "content": "'s:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 377, + 274, + 545, + 302 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 377, + 274, + 545, + 302 + ], + "spans": [ + { + "bbox": [ + 377, + 274, + 545, + 302 + ], + "type": "interline_equation", + "content": "s _ {i j} ^ {(t)} = \\sum_ {L \\in N _ {i j} ^ {c}} w _ {L} ^ {(t)} d _ {L} / z _ {i j} ^ {(t)} \\tag {4}", + "image_path": "ccfb676da6a0698937059f6958e653a064db77a72377bdcd0595e533cd496720.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 309, + 545, + 342 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 309, + 545, + 342 + ], + "spans": [ + { + "bbox": [ + 304, + 309, + 545, + 342 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 309, + 545, + 342 + ], + "type": "inline_equation", + "content": "z_{ij}^{(t)} = \\sum_{L\\in N_{ij}^c}w_L^{(t)}" + }, + { + "bbox": [ + 304, + 309, + 545, + 342 + ], + "type": "text", + "content": ". The cycle weights " + }, + { + "bbox": [ + 304, + 309, + 545, + 342 + ], + "type": "inline_equation", + "content": "w_{L}^{(t + 1)}" + }, + { + "bbox": [ + 304, + 309, + 545, + 342 + ], + "type": "text", + "content": " are computed from the edge weights " + }, + { + "bbox": [ + 304, + 309, + 545, + 342 + ], + "type": "inline_equation", + "content": "w_{e}^{(t + 1)} = e^{-\\beta_{t}s_{e}^{(t)}}" + }, + { + "bbox": [ + 304, + 309, + 545, + 342 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 340, + 348, + 545, + 376 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 340, + 348, + 545, + 376 + ], + "spans": [ + { + "bbox": [ + 340, + 348, + 545, + 376 + ], + "type": "interline_equation", + "content": "w _ {L} ^ {(t + 1)} = \\prod_ {e \\in L \\backslash \\{i j \\}} w _ {e} ^ {(t + 1)} = \\prod_ {e \\in L \\backslash \\{i j \\}} e ^ {- \\beta_ {t} s _ {e} ^ {(t)}}, \\tag {5}", + "image_path": "01e04240c42fb86ed7aba258f178e2c1c5c8d9106e070371f73a2b8a0a113037.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 384, + 547, + 494 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 384, + 547, + 494 + ], + "spans": [ + { + "bbox": [ + 304, + 384, + 547, + 494 + ], + "type": "text", + "content": "so that " + }, + { + "bbox": [ + 304, + 384, + 547, + 494 + ], + "type": "inline_equation", + "content": "w_{L}^{(t + 1)}" + }, + { + "bbox": [ + 304, + 384, + 547, + 494 + ], + "type": "text", + "content": " focuses on good cycles. The cycle weights and edge corruption levels are alternatingly updated and improved from each other. Interestingly, it is proved in [25] under two different corruption models that CEMP converges linearly to the ground truth corruption estimates under mild conditions for " + }, + { + "bbox": [ + 304, + 384, + 547, + 494 + ], + "type": "inline_equation", + "content": "c = 3" + }, + { + "bbox": [ + 304, + 384, + 547, + 494 + ], + "type": "text", + "content": ". In practice, CEMP only uses 3-cycles for consideration of efficiency. For longer cycles, the complexity of CEMP scales exponentially with the cycle length " + }, + { + "bbox": [ + 304, + 384, + 547, + 494 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 304, + 384, + 547, + 494 + ], + "type": "text", + "content": " (which is discussed in §3.2), and the convergence guarantee of CEMP remains unknown." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 305, + 503, + 437, + 517 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 503, + 437, + 517 + ], + "spans": [ + { + "bbox": [ + 305, + 503, + 437, + 517 + ], + "type": "text", + "content": "3. Our method: LongSync" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 305, + 523, + 490, + 536 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 523, + 490, + 536 + ], + "spans": [ + { + "bbox": [ + 305, + 523, + 490, + 536 + ], + "type": "text", + "content": "3.1. LongSync: Modification of CEMP" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 304, + 541, + 545, + 636 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 541, + 545, + 636 + ], + "spans": [ + { + "bbox": [ + 304, + 541, + 545, + 636 + ], + "type": "text", + "content": "Our goal is to develop a scalable variant of CEMP for any fixed number of cycles, " + }, + { + "bbox": [ + 304, + 541, + 545, + 636 + ], + "type": "inline_equation", + "content": "c \\geq 3" + }, + { + "bbox": [ + 304, + 541, + 545, + 636 + ], + "type": "text", + "content": ". The main computational bottleneck of step (4) in CEMP is that computing and summing the cycle inconsistency measures takes " + }, + { + "bbox": [ + 304, + 541, + 545, + 636 + ], + "type": "inline_equation", + "content": "\\sum_{ij \\in E} |N_{ij}^c| = O(n^c)" + }, + { + "bbox": [ + 304, + 541, + 545, + 636 + ], + "type": "text", + "content": " operations and memory. Therefore, to develop a scalable algorithm, we aim to take weighted average over " + }, + { + "bbox": [ + 304, + 541, + 545, + 636 + ], + "type": "inline_equation", + "content": "d_L" + }, + { + "bbox": [ + 304, + 541, + 545, + 636 + ], + "type": "text", + "content": " without explicitly computing and storing each " + }, + { + "bbox": [ + 304, + 541, + 545, + 636 + ], + "type": "inline_equation", + "content": "d_L" + }, + { + "bbox": [ + 304, + 541, + 545, + 636 + ], + "type": "text", + "content": ". To achieve this, we propose the following specification and modification on CEMP:" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 306, + 637, + 545, + 659 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 637, + 545, + 659 + ], + "spans": [ + { + "bbox": [ + 306, + 637, + 545, + 659 + ], + "type": "text", + "content": "- Use Chordal distance on SO(d). We suggest the distance function" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 366, + 659, + 493, + 673 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 366, + 659, + 493, + 673 + ], + "spans": [ + { + "bbox": [ + 366, + 659, + 493, + 673 + ], + "type": "interline_equation", + "content": "\\mathcal {D} \\left(\\boldsymbol {R} _ {1}, \\boldsymbol {R} _ {2}\\right) = \\sqrt {1 - \\left\\langle \\boldsymbol {R} _ {1} , \\boldsymbol {R} _ {2} \\right\\rangle / d}.", + "image_path": "1f4d57efed73ec4018221e5ea99bcb0de227707ce8a4d8cf4aa012b562adb9cd.jpg" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 313, + 677, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 545, + 712 + ], + "type": "text", + "content": "This distance is proportional to the Chordal distance on " + }, + { + "bbox": [ + 313, + 677, + 545, + 712 + ], + "type": "inline_equation", + "content": "SO(3)" + }, + { + "bbox": [ + 313, + 677, + 545, + 712 + ], + "type": "text", + "content": ", which is the Euclidean distance between two rotations embedded in " + }, + { + "bbox": [ + 313, + 677, + 545, + 712 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^{d\\times d}" + }, + { + "bbox": [ + 313, + 677, + 545, + 712 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 26 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5262" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 287, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 287, + 120 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 287, + 120 + ], + "type": "text", + "content": "- Use weighted quadratic average for corruption level update. Instead of updating the corruption level estimates by a weighted average of " + }, + { + "bbox": [ + 47, + 72, + 287, + 120 + ], + "type": "inline_equation", + "content": "d_{L}" + }, + { + "bbox": [ + 47, + 72, + 287, + 120 + ], + "type": "text", + "content": ", we use the weighted quadratic average of " + }, + { + "bbox": [ + 47, + 72, + 287, + 120 + ], + "type": "inline_equation", + "content": "d_{L}" + }, + { + "bbox": [ + 47, + 72, + 287, + 120 + ], + "type": "text", + "content": ", namely" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 118, + 129, + 287, + 160 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 118, + 129, + 287, + 160 + ], + "spans": [ + { + "bbox": [ + 118, + 129, + 287, + 160 + ], + "type": "interline_equation", + "content": "s _ {i j} ^ {(t)} = \\sqrt {\\sum_ {L \\in N _ {i j} ^ {c}} w _ {L} ^ {(t)} d _ {L} ^ {2} / z _ {i j} ^ {(t)}} \\tag {6}", + "image_path": "826c28ed76cb9ab608f5c6e2a8d3d55193631e6390e37ec88fb82798cc53b12b.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 163, + 274, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 163, + 274, + 175 + ], + "spans": [ + { + "bbox": [ + 55, + 163, + 274, + 175 + ], + "type": "text", + "content": "where the update rule of cycle weights remains the same:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 85, + 180, + 287, + 209 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 85, + 180, + 287, + 209 + ], + "spans": [ + { + "bbox": [ + 85, + 180, + 287, + 209 + ], + "type": "interline_equation", + "content": "w _ {L} ^ {(t + 1)} = \\prod_ {e \\in L \\backslash \\{i j \\}} w _ {e} ^ {(t + 1)} = \\prod_ {e \\in L \\backslash \\{i j \\}} e ^ {- \\beta_ {t} s _ {e} ^ {(t)}}. \\tag {7}", + "image_path": "9e72e58cb84556e4dcd3389b265e5d86b6ca298c1c36c6764b77846ee965f2dd.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 215, + 287, + 295 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 215, + 287, + 295 + ], + "spans": [ + { + "bbox": [ + 47, + 215, + 287, + 295 + ], + "type": "text", + "content": "As a result, " + }, + { + "bbox": [ + 47, + 215, + 287, + 295 + ], + "type": "inline_equation", + "content": "d_{L}^{2} = d^{2}(\\pmb{R}_{L},\\pmb{R}_{ij}) = 1 - \\langle \\pmb{R}_{L},\\pmb{R}_{ij}\\rangle /d" + }, + { + "bbox": [ + 47, + 215, + 287, + 295 + ], + "type": "text", + "content": " is linear in both " + }, + { + "bbox": [ + 47, + 215, + 287, + 295 + ], + "type": "inline_equation", + "content": "R_{L}" + }, + { + "bbox": [ + 47, + 215, + 287, + 295 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 215, + 287, + 295 + ], + "type": "inline_equation", + "content": "R_{ij}" + }, + { + "bbox": [ + 47, + 215, + 287, + 295 + ], + "type": "text", + "content": ". Therefore one can switch the order of " + }, + { + "bbox": [ + 47, + 215, + 287, + 295 + ], + "type": "inline_equation", + "content": "d^{2}" + }, + { + "bbox": [ + 47, + 215, + 287, + 295 + ], + "type": "text", + "content": " and the weighted summation, so that the computation of " + }, + { + "bbox": [ + 47, + 215, + 287, + 295 + ], + "type": "inline_equation", + "content": "s_{ij}^{(t)}" + }, + { + "bbox": [ + 47, + 215, + 287, + 295 + ], + "type": "text", + "content": " can be vectorized. Indeed, by this linearity and equations (4) and (2), and note that " + }, + { + "bbox": [ + 47, + 215, + 287, + 295 + ], + "type": "inline_equation", + "content": "z_{ij}^{(t)} = \\sum_{L\\in N_{ij}^{c}}w_{L}^{(t)}" + }, + { + "bbox": [ + 47, + 215, + 287, + 295 + ], + "type": "text", + "content": ", we obtain the following equation:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 69, + 300, + 287, + 436 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 300, + 287, + 436 + ], + "spans": [ + { + "bbox": [ + 69, + 300, + 287, + 436 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} s _ {i j} ^ {(t)} = \\left(\\sum_ {L \\in N _ {i j} ^ {c}} w _ {L} ^ {(t)} d _ {L} ^ {2} / z _ {i j} ^ {(t)}\\right) ^ {1 / 2} \\\\ = \\Big (\\sum_ {L \\in N _ {i j} ^ {c}} w _ {L} ^ {(t)} \\mathcal {D} ^ {2} (\\boldsymbol {R} _ {L}, \\boldsymbol {R} _ {i j}) / z _ {i j} ^ {(t)} \\Big) ^ {1 / 2} \\\\ = \\left(\\mathcal {D} ^ {2} \\Big (\\sum_ {L \\in N _ {i j} ^ {c}} w _ {L} ^ {(t)} \\boldsymbol {R} _ {L}, \\boldsymbol {R} _ {i j} \\Big) / z _ {i j} ^ {(t)}\\right) ^ {1 / 2} \\\\ = \\left(1 - \\left\\langle \\sum_ {L \\in N _ {i j} ^ {c}} w _ {L} ^ {(t)} \\boldsymbol {R} _ {L}, \\boldsymbol {R} _ {i j} \\right\\rangle / d \\sum_ {L \\in N _ {i j} ^ {c}} w _ {L} ^ {(t)}\\right) ^ {1 / 2} \\tag {8} \\\\ \\end{array}", + "image_path": "a9219dd0ad989cca4483f979a670e3884e07d04f1f88ceef555b14f0c0b32b42.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 441, + 287, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 441, + 287, + 491 + ], + "spans": [ + { + "bbox": [ + 47, + 441, + 287, + 491 + ], + "type": "text", + "content": "Equation (8) can be vectorized using the trick of matrix power if we allow repeated nodes for each cycle. That is, one can stack the " + }, + { + "bbox": [ + 47, + 441, + 287, + 491 + ], + "type": "inline_equation", + "content": "s_{ij}^{(t)}" + }, + { + "bbox": [ + 47, + 441, + 287, + 491 + ], + "type": "text", + "content": "'s and " + }, + { + "bbox": [ + 47, + 441, + 287, + 491 + ], + "type": "inline_equation", + "content": "w_{ij}^{(t)}" + }, + { + "bbox": [ + 47, + 441, + 287, + 491 + ], + "type": "text", + "content": "'s into matrices " + }, + { + "bbox": [ + 47, + 441, + 287, + 491 + ], + "type": "inline_equation", + "content": "\\mathbf{S}^{(t)}" + }, + { + "bbox": [ + 47, + 441, + 287, + 491 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 441, + 287, + 491 + ], + "type": "inline_equation", + "content": "\\mathbf{W}^{(t)}" + }, + { + "bbox": [ + 47, + 441, + 287, + 491 + ], + "type": "text", + "content": ", and vectorize (8) as" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 80, + 506, + 287, + 554 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 80, + 506, + 287, + 554 + ], + "spans": [ + { + "bbox": [ + 80, + 506, + 287, + 554 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\boldsymbol {S} ^ {(t)} = \\left(\\boldsymbol {A} - \\left\\langle \\left(\\boldsymbol {W} ^ {(t)} \\otimes 1 _ {d} \\odot \\boldsymbol {R}\\right) ^ {c - 1}, \\boldsymbol {R} \\right\\rangle_ {\\text {b l o c k}} \\right. \\\\ \\left. \\odot d \\left(\\boldsymbol {W} ^ {(t)}\\right) ^ {c - 1}\\right) ^ {\\odot 1 / 2}. \\tag {9} \\\\ \\end{array}", + "image_path": "dad44fd85945d57a1d240e2bbd3e802a9365b7eb9c4a1d58fe16315f5ada8a03.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 561, + 287, + 620 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 561, + 287, + 620 + ], + "spans": [ + { + "bbox": [ + 47, + 561, + 287, + 620 + ], + "type": "text", + "content": "Indeed, by (7) and the definition of " + }, + { + "bbox": [ + 47, + 561, + 287, + 620 + ], + "type": "inline_equation", + "content": "\\pmb{R}_L" + }, + { + "bbox": [ + 47, + 561, + 287, + 620 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 47, + 561, + 287, + 620 + ], + "type": "inline_equation", + "content": "\\sum_{L \\in C_{ij}^c} w_L^{(t)} \\pmb{R}_L" + }, + { + "bbox": [ + 47, + 561, + 287, + 620 + ], + "type": "text", + "content": " is the " + }, + { + "bbox": [ + 47, + 561, + 287, + 620 + ], + "type": "inline_equation", + "content": "ij" + }, + { + "bbox": [ + 47, + 561, + 287, + 620 + ], + "type": "text", + "content": "-th block of " + }, + { + "bbox": [ + 47, + 561, + 287, + 620 + ], + "type": "inline_equation", + "content": "(\\pmb{W}^{(t)} \\otimes 1_d \\odot \\pmb{R})^{c-1}" + }, + { + "bbox": [ + 47, + 561, + 287, + 620 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 47, + 561, + 287, + 620 + ], + "type": "inline_equation", + "content": "\\sum_{L \\in C_{ij}^c} w_L^{(t)}" + }, + { + "bbox": [ + 47, + 561, + 287, + 620 + ], + "type": "text", + "content": " is the " + }, + { + "bbox": [ + 47, + 561, + 287, + 620 + ], + "type": "inline_equation", + "content": "ij" + }, + { + "bbox": [ + 47, + 561, + 287, + 620 + ], + "type": "text", + "content": "-th element of " + }, + { + "bbox": [ + 47, + 561, + 287, + 620 + ], + "type": "inline_equation", + "content": "\\pmb{W}^{(t)c-1}" + }, + { + "bbox": [ + 47, + 561, + 287, + 620 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 47, + 561, + 287, + 620 + ], + "type": "inline_equation", + "content": "C_{ij}^c" + }, + { + "bbox": [ + 47, + 561, + 287, + 620 + ], + "type": "text", + "content": " is the set of " + }, + { + "bbox": [ + 47, + 561, + 287, + 620 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 47, + 561, + 287, + 620 + ], + "type": "text", + "content": "-cycles containing " + }, + { + "bbox": [ + 47, + 561, + 287, + 620 + ], + "type": "inline_equation", + "content": "ij" + }, + { + "bbox": [ + 47, + 561, + 287, + 620 + ], + "type": "text", + "content": " with possibly repeated nodes." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 620, + 287, + 716 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 620, + 287, + 716 + ], + "spans": [ + { + "bbox": [ + 47, + 620, + 287, + 716 + ], + "type": "text", + "content": "In the case of utilizing only simple cycles, (8) and (9) are not equivalent and we need to correct (9) to remove the cycles with repeated nodes, so that only simple cycles in " + }, + { + "bbox": [ + 47, + 620, + 287, + 716 + ], + "type": "inline_equation", + "content": "N_{ij}^{c}" + }, + { + "bbox": [ + 47, + 620, + 287, + 716 + ], + "type": "text", + "content": " remain. Let " + }, + { + "bbox": [ + 47, + 620, + 287, + 716 + ], + "type": "inline_equation", + "content": "g_{c}(\\boldsymbol{W},\\boldsymbol{R})" + }, + { + "bbox": [ + 47, + 620, + 287, + 716 + ], + "type": "text", + "content": " be the matrix valued function where " + }, + { + "bbox": [ + 47, + 620, + 287, + 716 + ], + "type": "inline_equation", + "content": "g_{c}(\\boldsymbol{W},\\boldsymbol{R})(i,j) = \\sum_{L\\in N_{ij}^{c}}(\\prod_{e\\in L\\setminus \\{ij\\}}w_{e}^{(t)})\\boldsymbol{R}_{L}" + }, + { + "bbox": [ + 47, + 620, + 287, + 716 + ], + "type": "text", + "content": ". Let " + }, + { + "bbox": [ + 47, + 620, + 287, + 716 + ], + "type": "inline_equation", + "content": "f_{c}(\\boldsymbol{W})" + }, + { + "bbox": [ + 47, + 620, + 287, + 716 + ], + "type": "text", + "content": " be the matrix valued function where " + }, + { + "bbox": [ + 47, + 620, + 287, + 716 + ], + "type": "inline_equation", + "content": "f_{c}(\\boldsymbol{W})(i,j) = \\sum_{L\\in N_{ij}^{c}}(\\prod_{e\\in L\\setminus \\{ij\\}}w_{e}^{(t)})" + }, + { + "bbox": [ + 47, + 620, + 287, + 716 + ], + "type": "text", + "content": ". The following result holds:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "type": "text", + "content": "Proposition 3.1. The update rule of LongSync (8) is equivalent to the following matrix equations:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 328, + 104, + 545, + 129 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 328, + 104, + 545, + 129 + ], + "spans": [ + { + "bbox": [ + 328, + 104, + 545, + 129 + ], + "type": "interline_equation", + "content": "\\boldsymbol {S} ^ {(t)} = \\left(\\boldsymbol {A} - \\left\\langle h _ {c} (\\boldsymbol {W} ^ {(t)}, \\boldsymbol {R}), \\boldsymbol {R} \\right\\rangle_ {\\text {b l o c k}} \\odot \\boldsymbol {A}\\right) ^ {\\odot 1 / 2}, \\tag {10}", + "image_path": "5cf24f1449a94a6feaa7dab4328a2338db52c51aba11c3a3879c7350784ac462.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 137, + 464, + 150 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 137, + 464, + 150 + ], + "spans": [ + { + "bbox": [ + 306, + 137, + 464, + 150 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 306, + 137, + 464, + 150 + ], + "type": "inline_equation", + "content": "\\pmb{W}^{(t + 1)} = \\pmb{A}\\odot \\exp (-\\beta_t\\pmb{S}^{(t)})" + }, + { + "bbox": [ + 306, + 137, + 464, + 150 + ], + "type": "text", + "content": " and" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 328, + 157, + 522, + 172 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 328, + 157, + 522, + 172 + ], + "spans": [ + { + "bbox": [ + 328, + 157, + 522, + 172 + ], + "type": "interline_equation", + "content": "h _ {c} \\left(\\boldsymbol {W} ^ {(t)}, \\boldsymbol {R}\\right) := g _ {c} \\left(\\boldsymbol {W} ^ {(t)}, \\boldsymbol {R}\\right) \\oslash \\left(d \\cdot f _ {c} \\left(\\boldsymbol {W} ^ {(t)} \\otimes 1 _ {d}\\right)\\right).", + "image_path": "c59698f9b673b3ed7b8b938a3bddd1b475a150189f0b04ceb183cfcaea2aacd5.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 317, + 181, + 530, + 194 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 181, + 530, + 194 + ], + "spans": [ + { + "bbox": [ + 317, + 181, + 530, + 194 + ], + "type": "text", + "content": "Here exp denotes the elementwise exponential function." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 201, + 545, + 225 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 201, + 545, + 225 + ], + "spans": [ + { + "bbox": [ + 306, + 201, + 545, + 225 + ], + "type": "text", + "content": "We use equation (10) as the update rule of LongSync and propose the vectorized LongSync algorithm in algorithm 1." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 236, + 406, + 249 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 236, + 406, + 249 + ], + "spans": [ + { + "bbox": [ + 307, + 236, + 406, + 249 + ], + "type": "text", + "content": "Algorithm 1 (LongSync)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 253, + 545, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 253, + 545, + 289 + ], + "spans": [ + { + "bbox": [ + 306, + 253, + 545, + 289 + ], + "type": "text", + "content": "Input: pairwise rotation matrix " + }, + { + "bbox": [ + 306, + 253, + 545, + 289 + ], + "type": "inline_equation", + "content": "\\mathbf{R} \\in \\mathbb{R}^{dn \\times dn}" + }, + { + "bbox": [ + 306, + 253, + 545, + 289 + ], + "type": "text", + "content": ", adjacency matrix " + }, + { + "bbox": [ + 306, + 253, + 545, + 289 + ], + "type": "inline_equation", + "content": "\\mathbf{A} \\in [0,1]^{n \\times n}" + }, + { + "bbox": [ + 306, + 253, + 545, + 289 + ], + "type": "text", + "content": ", cycle length " + }, + { + "bbox": [ + 306, + 253, + 545, + 289 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 306, + 253, + 545, + 289 + ], + "type": "text", + "content": ", positive parameters " + }, + { + "bbox": [ + 306, + 253, + 545, + 289 + ], + "type": "inline_equation", + "content": "\\{\\beta_t\\}_{t \\geq 1}" + }, + { + "bbox": [ + 306, + 253, + 545, + 289 + ], + "type": "text", + "content": ", time step " + }, + { + "bbox": [ + 306, + 253, + 545, + 289 + ], + "type": "inline_equation", + "content": "T" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 289, + 378, + 300 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 289, + 378, + 300 + ], + "spans": [ + { + "bbox": [ + 317, + 289, + 378, + 300 + ], + "type": "interline_equation", + "content": "\\boldsymbol {W} ^ {(0)} (i, j) \\leftarrow \\boldsymbol {A}", + "image_path": "431516466b38361185a2c7cb628bafe6ff31b9e87ed5a45121769fceb3ea7c97.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 302, + 375, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 302, + 375, + 312 + ], + "spans": [ + { + "bbox": [ + 317, + 302, + 375, + 312 + ], + "type": "text", + "content": "for " + }, + { + "bbox": [ + 317, + 302, + 375, + 312 + ], + "type": "inline_equation", + "content": "t = 0:T" + }, + { + "bbox": [ + 317, + 302, + 375, + 312 + ], + "type": "text", + "content": " do" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 332, + 319, + 545, + 343 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 332, + 319, + 545, + 343 + ], + "spans": [ + { + "bbox": [ + 332, + 319, + 545, + 343 + ], + "type": "interline_equation", + "content": "\\boldsymbol {S} ^ {(t)} \\leftarrow \\left(\\boldsymbol {A} - \\left\\langle h _ {c} (\\boldsymbol {W} ^ {(t)}, \\boldsymbol {R}), \\boldsymbol {R} \\right\\rangle_ {\\text {b l o c k}} \\odot \\boldsymbol {A}\\right) ^ {\\odot 1 / 2} \\tag {11}", + "image_path": "d98cb2d91e983f33e567d465b99d674d461026b4ade14a16c32a4e4c03195912.jpg" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 333, + 345, + 545, + 359 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 345, + 545, + 359 + ], + "spans": [ + { + "bbox": [ + 333, + 345, + 545, + 359 + ], + "type": "interline_equation", + "content": "\\boldsymbol {W} ^ {(t + 1)} \\leftarrow \\boldsymbol {A} \\odot \\exp (- \\beta_ {t} \\boldsymbol {S} ^ {(t)}) \\tag {12}", + "image_path": "f734eaa13bff2b2249655d55d86c17d23907b7c89c047203120ff86890457041.jpg" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 369, + 348, + 379 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 369, + 348, + 379 + ], + "spans": [ + { + "bbox": [ + 317, + 369, + 348, + 379 + ], + "type": "text", + "content": "end for" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 306, + 380, + 520, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 380, + 520, + 392 + ], + "spans": [ + { + "bbox": [ + 306, + 380, + 520, + 392 + ], + "type": "text", + "content": "Output: edge weights " + }, + { + "bbox": [ + 306, + 380, + 520, + 392 + ], + "type": "inline_equation", + "content": "\\mathbf{W}^{(T + 1)}" + }, + { + "bbox": [ + 306, + 380, + 520, + 392 + ], + "type": "text", + "content": ", corruption levels " + }, + { + "bbox": [ + 306, + 380, + 520, + 392 + ], + "type": "inline_equation", + "content": "\\mathbf{S}^{(T)}" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 304, + 409, + 545, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 409, + 545, + 552 + ], + "spans": [ + { + "bbox": [ + 304, + 409, + 545, + 552 + ], + "type": "text", + "content": "We claim that " + }, + { + "bbox": [ + 304, + 409, + 545, + 552 + ], + "type": "inline_equation", + "content": "g_{c}" + }, + { + "bbox": [ + 304, + 409, + 545, + 552 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 409, + 545, + 552 + ], + "type": "inline_equation", + "content": "f_{c}" + }, + { + "bbox": [ + 304, + 409, + 545, + 552 + ], + "type": "text", + "content": " can be computed with a sequence of matrix operations, thus greatly reducing the time and space consumption of LongSync compared to its original form. For " + }, + { + "bbox": [ + 304, + 409, + 545, + 552 + ], + "type": "inline_equation", + "content": "c \\leq 6" + }, + { + "bbox": [ + 304, + 409, + 545, + 552 + ], + "type": "text", + "content": ", the time complexity of computing " + }, + { + "bbox": [ + 304, + 409, + 545, + 552 + ], + "type": "inline_equation", + "content": "g_{c}" + }, + { + "bbox": [ + 304, + 409, + 545, + 552 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 409, + 545, + 552 + ], + "type": "inline_equation", + "content": "f_{c}" + }, + { + "bbox": [ + 304, + 409, + 545, + 552 + ], + "type": "text", + "content": " is " + }, + { + "bbox": [ + 304, + 409, + 545, + 552 + ], + "type": "inline_equation", + "content": "O(r(dn))" + }, + { + "bbox": [ + 304, + 409, + 545, + 552 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 409, + 545, + 552 + ], + "type": "inline_equation", + "content": "r(n)" + }, + { + "bbox": [ + 304, + 409, + 545, + 552 + ], + "type": "text", + "content": " is the complexity for multiplying two " + }, + { + "bbox": [ + 304, + 409, + 545, + 552 + ], + "type": "inline_equation", + "content": "n \\times n" + }, + { + "bbox": [ + 304, + 409, + 545, + 552 + ], + "type": "text", + "content": " matrices; for " + }, + { + "bbox": [ + 304, + 409, + 545, + 552 + ], + "type": "inline_equation", + "content": "c \\geq 7" + }, + { + "bbox": [ + 304, + 409, + 545, + 552 + ], + "type": "text", + "content": " the time complexity is at most " + }, + { + "bbox": [ + 304, + 409, + 545, + 552 + ], + "type": "inline_equation", + "content": "O(n^{[(c + 3) / 2]})" + }, + { + "bbox": [ + 304, + 409, + 545, + 552 + ], + "type": "text", + "content": ". This claim is proved in the supplementary material. We list the formula for " + }, + { + "bbox": [ + 304, + 409, + 545, + 552 + ], + "type": "inline_equation", + "content": "g_{c}" + }, + { + "bbox": [ + 304, + 409, + 545, + 552 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 409, + 545, + 552 + ], + "type": "inline_equation", + "content": "f_{c}" + }, + { + "bbox": [ + 304, + 409, + 545, + 552 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 304, + 409, + 545, + 552 + ], + "type": "inline_equation", + "content": "c = 3,4,5,6" + }, + { + "bbox": [ + 304, + 409, + 545, + 552 + ], + "type": "text", + "content": " inspired by [35, 46]. The formula for " + }, + { + "bbox": [ + 304, + 409, + 545, + 552 + ], + "type": "inline_equation", + "content": "c = 6" + }, + { + "bbox": [ + 304, + 409, + 545, + 552 + ], + "type": "text", + "content": " is moved to the supplementary material due to the space limit. For " + }, + { + "bbox": [ + 304, + 409, + 545, + 552 + ], + "type": "inline_equation", + "content": "c \\geq 7" + }, + { + "bbox": [ + 304, + 409, + 545, + 552 + ], + "type": "text", + "content": " the formula becomes extremely complicated. We remark that in practice, the cycles of length greater than 6 are often not used." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 304, + 552, + 545, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 552, + 545, + 624 + ], + "spans": [ + { + "bbox": [ + 304, + 552, + 545, + 624 + ], + "type": "text", + "content": "We finally remark that although Algorithm 1 only utilizes cycles of a fixed length, one can easily generalize it to incorporate cycles of different lengths. Indeed, the equation (10) could use a convex combination of " + }, + { + "bbox": [ + 304, + 552, + 545, + 624 + ], + "type": "inline_equation", + "content": "h_c" + }, + { + "bbox": [ + 304, + 552, + 545, + 624 + ], + "type": "text", + "content": "'s that corresponds to different values of " + }, + { + "bbox": [ + 304, + 552, + 545, + 624 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 304, + 552, + 545, + 624 + ], + "type": "text", + "content": ". That is, for a preselected set of cycle lengths " + }, + { + "bbox": [ + 304, + 552, + 545, + 624 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 304, + 552, + 545, + 624 + ], + "type": "text", + "content": ", the equation (11) in Algorithm 1 is replaced by" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 312, + 632, + 545, + 669 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 632, + 545, + 669 + ], + "spans": [ + { + "bbox": [ + 312, + 632, + 545, + 669 + ], + "type": "interline_equation", + "content": "\\boldsymbol {S} ^ {(t)} = \\left(\\boldsymbol {A} - \\left\\langle \\sum_ {c \\in C} \\lambda_ {c} h _ {c} \\left(\\boldsymbol {W} ^ {(t)}, \\boldsymbol {R}\\right), \\boldsymbol {R} \\right\\rangle_ {\\text {b l o c k}} \\odot \\boldsymbol {A}\\right) ^ {\\odot 1 / 2} \\tag {13}", + "image_path": "a4756c7c7dc9325652c2bec74190fa21f425cb48ee97a07633c7a17e1daeb6b7.jpg" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "content": "where the coefficients " + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\lambda_{c}" + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "content": " satisfies " + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\sum_{c\\in C}\\lambda_c = 1" + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "content": " to ensure a convex combination. Here each " + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\lambda_{c}" + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "content": " is user-specified to reflect the importance of the cycles of length " + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "content": ". However, the optimal" + } + ] + } + ], + "index": 27 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5263" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 95 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 95 + ], + "type": "text", + "content": "choice of these parameters under certain statistical model remains unclear." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 96, + 286, + 155 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 96, + 286, + 155 + ], + "spans": [ + { + "bbox": [ + 46, + 96, + 286, + 155 + ], + "type": "text", + "content": "For simplicity, in the experiments we only use a fixed cycle length to avoid choosing " + }, + { + "bbox": [ + 46, + 96, + 286, + 155 + ], + "type": "inline_equation", + "content": "\\lambda_{c}" + }, + { + "bbox": [ + 46, + 96, + 286, + 155 + ], + "type": "text", + "content": ". We have observed that such simple choice still yields satisfying accuracy in camera orientation estimation on both synthetic and real data. We refer the readers to §5 and 6 for more details." + } + ] + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 59, + 163, + 276, + 288 + ], + "blocks": [ + { + "bbox": [ + 59, + 163, + 276, + 288 + ], + "lines": [ + { + "bbox": [ + 59, + 163, + 276, + 288 + ], + "spans": [ + { + "bbox": [ + 59, + 163, + 276, + 288 + ], + "type": "table", + "html": "
cFormula of fc(W)Formula of gc(W,R)
3W2P2
4W3-d(W2)WP3-dblock(P2)P
-Wd(W2)+W⊙3-Pdblock(P2)+P⊙3
5W4-d(W3)WP4-dblock(P3)P
-d(W2)W2-dblock(P2)P2
-W2d(W2)-P2dblock(P2)
-Wd(W2)W-Pdblock(P2)P
+3W⊙2W2+3P⊙2P2
+WW⊙3+W⊙3W+PP⊙3+P⊙3P
", + "image_path": "c998bbdd513673af74c82589d6d8e30a5ee5a35096a9e1f2bf87ee12a6c0c53e.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 295, + 287, + 339 + ], + "lines": [ + { + "bbox": [ + 46, + 295, + 287, + 339 + ], + "spans": [ + { + "bbox": [ + 46, + 295, + 287, + 339 + ], + "type": "text", + "content": "Table 1. Formulas for " + }, + { + "bbox": [ + 46, + 295, + 287, + 339 + ], + "type": "inline_equation", + "content": "f_{c}" + }, + { + "bbox": [ + 46, + 295, + 287, + 339 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 295, + 287, + 339 + ], + "type": "inline_equation", + "content": "g_{c}" + }, + { + "bbox": [ + 46, + 295, + 287, + 339 + ], + "type": "text", + "content": ". Here we let " + }, + { + "bbox": [ + 46, + 295, + 287, + 339 + ], + "type": "inline_equation", + "content": "P = (W \\otimes 1_d) \\odot R" + }, + { + "bbox": [ + 46, + 295, + 287, + 339 + ], + "type": "text", + "content": " for shorter notation. " + }, + { + "bbox": [ + 46, + 295, + 287, + 339 + ], + "type": "inline_equation", + "content": "\\mathrm{d}(X)" + }, + { + "bbox": [ + 46, + 295, + 287, + 339 + ], + "type": "text", + "content": " returns the diagonal of matrix " + }, + { + "bbox": [ + 46, + 295, + 287, + 339 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 46, + 295, + 287, + 339 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 295, + 287, + 339 + ], + "type": "inline_equation", + "content": "\\mathrm{d}_{\\mathrm{block}}(X)" + }, + { + "bbox": [ + 46, + 295, + 287, + 339 + ], + "type": "text", + "content": " returns the diagonal block matrix from the " + }, + { + "bbox": [ + 46, + 295, + 287, + 339 + ], + "type": "inline_equation", + "content": "d \\times d" + }, + { + "bbox": [ + 46, + 295, + 287, + 339 + ], + "type": "text", + "content": " diagonal blocks of matrix " + }, + { + "bbox": [ + 46, + 295, + 287, + 339 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 46, + 295, + 287, + 339 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 47, + 357, + 197, + 371 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 357, + 197, + 371 + ], + "spans": [ + { + "bbox": [ + 47, + 357, + 197, + 371 + ], + "type": "text", + "content": "3.2. Computational Complexity" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 376, + 287, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 376, + 287, + 628 + ], + "spans": [ + { + "bbox": [ + 46, + 376, + 287, + 628 + ], + "type": "text", + "content": "We derive the space and time complexity for LongSync, and demonstrate its advantages over CEMP. The initialization step involves setting the weights of all edges to 1, which takes time " + }, + { + "bbox": [ + 46, + 376, + 287, + 628 + ], + "type": "inline_equation", + "content": "O(|E|)" + }, + { + "bbox": [ + 46, + 376, + 287, + 628 + ], + "type": "text", + "content": " and space " + }, + { + "bbox": [ + 46, + 376, + 287, + 628 + ], + "type": "inline_equation", + "content": "O(n^{2})" + }, + { + "bbox": [ + 46, + 376, + 287, + 628 + ], + "type": "text", + "content": ". For each iteration, LongSync updates the matrices " + }, + { + "bbox": [ + 46, + 376, + 287, + 628 + ], + "type": "inline_equation", + "content": "\\pmb{S}^{(t)}" + }, + { + "bbox": [ + 46, + 376, + 287, + 628 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 376, + 287, + 628 + ], + "type": "inline_equation", + "content": "\\pmb{W}^{(t)}" + }, + { + "bbox": [ + 46, + 376, + 287, + 628 + ], + "type": "text", + "content": " with equations (11) and (12), respectively. Computing " + }, + { + "bbox": [ + 46, + 376, + 287, + 628 + ], + "type": "inline_equation", + "content": "\\pmb{W}^{(t + 1)}" + }, + { + "bbox": [ + 46, + 376, + 287, + 628 + ], + "type": "text", + "content": " involves two matrix subtractions, one scalar-matrix multiplication and one element-wise matrix exponential operation on " + }, + { + "bbox": [ + 46, + 376, + 287, + 628 + ], + "type": "inline_equation", + "content": "\\pmb{S}^{(t)} \\in \\mathbb{R}^{n \\times n}" + }, + { + "bbox": [ + 46, + 376, + 287, + 628 + ], + "type": "text", + "content": ". Therefore the update step (12) takes at most " + }, + { + "bbox": [ + 46, + 376, + 287, + 628 + ], + "type": "inline_equation", + "content": "O(n^{2})" + }, + { + "bbox": [ + 46, + 376, + 287, + 628 + ], + "type": "text", + "content": " time and space. Equation (11), on the other hand, involves a sequence of matrix operations on " + }, + { + "bbox": [ + 46, + 376, + 287, + 628 + ], + "type": "inline_equation", + "content": "\\pmb{P}^{(t)} = (\\pmb{W}^{(t)} \\otimes 1_{d}) \\odot \\pmb{R} \\in \\mathbb{R}^{dn \\times dn}" + }, + { + "bbox": [ + 46, + 376, + 287, + 628 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 376, + 287, + 628 + ], + "type": "inline_equation", + "content": "\\pmb{W}^{(t)} \\in \\mathbb{R}^{n \\times n}" + }, + { + "bbox": [ + 46, + 376, + 287, + 628 + ], + "type": "text", + "content": ", including matrix multiplications, element-wise multiplications and diagonal block selections. Computing " + }, + { + "bbox": [ + 46, + 376, + 287, + 628 + ], + "type": "inline_equation", + "content": "\\pmb{P}^{(t)}" + }, + { + "bbox": [ + 46, + 376, + 287, + 628 + ], + "type": "text", + "content": " takes " + }, + { + "bbox": [ + 46, + 376, + 287, + 628 + ], + "type": "inline_equation", + "content": "O(d^{2}n^{2})" + }, + { + "bbox": [ + 46, + 376, + 287, + 628 + ], + "type": "text", + "content": " memory and " + }, + { + "bbox": [ + 46, + 376, + 287, + 628 + ], + "type": "inline_equation", + "content": "O(d^{2}n^{2})" + }, + { + "bbox": [ + 46, + 376, + 287, + 628 + ], + "type": "text", + "content": " time. The matrix operations on " + }, + { + "bbox": [ + 46, + 376, + 287, + 628 + ], + "type": "inline_equation", + "content": "\\pmb{P}^{(t)}" + }, + { + "bbox": [ + 46, + 376, + 287, + 628 + ], + "type": "text", + "content": " take " + }, + { + "bbox": [ + 46, + 376, + 287, + 628 + ], + "type": "inline_equation", + "content": "O(K_{c}d^{3}n^{3})" + }, + { + "bbox": [ + 46, + 376, + 287, + 628 + ], + "type": "text", + "content": " time and " + }, + { + "bbox": [ + 46, + 376, + 287, + 628 + ], + "type": "inline_equation", + "content": "O(d^{2}n^{2})" + }, + { + "bbox": [ + 46, + 376, + 287, + 628 + ], + "type": "text", + "content": " space, and the matrix operations on " + }, + { + "bbox": [ + 46, + 376, + 287, + 628 + ], + "type": "inline_equation", + "content": "\\pmb{W}^{(t)}" + }, + { + "bbox": [ + 46, + 376, + 287, + 628 + ], + "type": "text", + "content": " take " + }, + { + "bbox": [ + 46, + 376, + 287, + 628 + ], + "type": "inline_equation", + "content": "O(K_{c}n^{3})" + }, + { + "bbox": [ + 46, + 376, + 287, + 628 + ], + "type": "text", + "content": " time and " + }, + { + "bbox": [ + 46, + 376, + 287, + 628 + ], + "type": "inline_equation", + "content": "O(n^{2})" + }, + { + "bbox": [ + 46, + 376, + 287, + 628 + ], + "type": "text", + "content": " space, where " + }, + { + "bbox": [ + 46, + 376, + 287, + 628 + ], + "type": "inline_equation", + "content": "K_{c}" + }, + { + "bbox": [ + 46, + 376, + 287, + 628 + ], + "type": "text", + "content": " is the number of terms in the equation for " + }, + { + "bbox": [ + 46, + 376, + 287, + 628 + ], + "type": "inline_equation", + "content": "f_{c}" + }, + { + "bbox": [ + 46, + 376, + 287, + 628 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 376, + 287, + 628 + ], + "type": "inline_equation", + "content": "g_{c}" + }, + { + "bbox": [ + 46, + 376, + 287, + 628 + ], + "type": "text", + "content": ". Therefore, each iteration over " + }, + { + "bbox": [ + 46, + 376, + 287, + 628 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 376, + 287, + 628 + ], + "type": "text", + "content": " takes " + }, + { + "bbox": [ + 46, + 376, + 287, + 628 + ], + "type": "inline_equation", + "content": "O(K_{c}d^{3}n^{3})" + }, + { + "bbox": [ + 46, + 376, + 287, + 628 + ], + "type": "text", + "content": " time and " + }, + { + "bbox": [ + 46, + 376, + 287, + 628 + ], + "type": "inline_equation", + "content": "O(d^{2}n^{2})" + }, + { + "bbox": [ + 46, + 376, + 287, + 628 + ], + "type": "text", + "content": " space. To sum up, for LongSync, the time complexity is " + }, + { + "bbox": [ + 46, + 376, + 287, + 628 + ], + "type": "inline_equation", + "content": "O(K_{c}d^{3}n^{3})" + }, + { + "bbox": [ + 46, + 376, + 287, + 628 + ], + "type": "text", + "content": " and the space complexity is " + }, + { + "bbox": [ + 46, + 376, + 287, + 628 + ], + "type": "inline_equation", + "content": "O(d^{2}n^{2})" + }, + { + "bbox": [ + 46, + 376, + 287, + 628 + ], + "type": "text", + "content": ". For " + }, + { + "bbox": [ + 46, + 376, + 287, + 628 + ], + "type": "inline_equation", + "content": "c = 3,4,5,6" + }, + { + "bbox": [ + 46, + 376, + 287, + 628 + ], + "type": "text", + "content": ", the number " + }, + { + "bbox": [ + 46, + 376, + 287, + 628 + ], + "type": "inline_equation", + "content": "K_{c}" + }, + { + "bbox": [ + 46, + 376, + 287, + 628 + ], + "type": "text", + "content": " is equal to 1,3,9,32." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 628, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 628, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 628, + 288, + 715 + ], + "type": "text", + "content": "In comparison, we consider the initialization step of CEMP. For each edge " + }, + { + "bbox": [ + 46, + 628, + 288, + 715 + ], + "type": "inline_equation", + "content": "ij \\in E" + }, + { + "bbox": [ + 46, + 628, + 288, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 628, + 288, + 715 + ], + "type": "inline_equation", + "content": "L \\in N_{ij}^{c}" + }, + { + "bbox": [ + 46, + 628, + 288, + 715 + ], + "type": "text", + "content": ", initializing CEMP involves computing and storing all the cycle inconsistency measures " + }, + { + "bbox": [ + 46, + 628, + 288, + 715 + ], + "type": "inline_equation", + "content": "d_{L}" + }, + { + "bbox": [ + 46, + 628, + 288, + 715 + ], + "type": "text", + "content": " using equation (2). For each " + }, + { + "bbox": [ + 46, + 628, + 288, + 715 + ], + "type": "inline_equation", + "content": "L \\in N_{ij}^{c}" + }, + { + "bbox": [ + 46, + 628, + 288, + 715 + ], + "type": "text", + "content": ", computing " + }, + { + "bbox": [ + 46, + 628, + 288, + 715 + ], + "type": "inline_equation", + "content": "d_{L}" + }, + { + "bbox": [ + 46, + 628, + 288, + 715 + ], + "type": "text", + "content": " involves multiplying " + }, + { + "bbox": [ + 46, + 628, + 288, + 715 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 46, + 628, + 288, + 715 + ], + "type": "text", + "content": " rotations, which takes " + }, + { + "bbox": [ + 46, + 628, + 288, + 715 + ], + "type": "inline_equation", + "content": "O(cd^3)" + }, + { + "bbox": [ + 46, + 628, + 288, + 715 + ], + "type": "text", + "content": " time and " + }, + { + "bbox": [ + 46, + 628, + 288, + 715 + ], + "type": "inline_equation", + "content": "O(d^2)" + }, + { + "bbox": [ + 46, + 628, + 288, + 715 + ], + "type": "text", + "content": " space. This step is repeated for each " + }, + { + "bbox": [ + 46, + 628, + 288, + 715 + ], + "type": "inline_equation", + "content": "ij \\in E" + }, + { + "bbox": [ + 46, + 628, + 288, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 628, + 288, + 715 + ], + "type": "inline_equation", + "content": "L \\in N_{ij}^{c}" + }, + { + "bbox": [ + 46, + 628, + 288, + 715 + ], + "type": "text", + "content": ", therefore the total time complexity is " + }, + { + "bbox": [ + 46, + 628, + 288, + 715 + ], + "type": "inline_equation", + "content": "O(cd^3\\sum_{ij \\in E}|N_{ij}^{c}|)" + }, + { + "bbox": [ + 46, + 628, + 288, + 715 + ], + "type": "text", + "content": " and" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 72, + 545, + 158 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 158 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 158 + ], + "type": "text", + "content": "the total space complexity is " + }, + { + "bbox": [ + 304, + 72, + 545, + 158 + ], + "type": "inline_equation", + "content": "O(d^{2}\\sum_{ij\\in E}|N_{ij}^{c}|)" + }, + { + "bbox": [ + 304, + 72, + 545, + 158 + ], + "type": "text", + "content": ". Since for each edge there are " + }, + { + "bbox": [ + 304, + 72, + 545, + 158 + ], + "type": "inline_equation", + "content": "(n - 2)(n - 3)\\dots (n - c + 1) = O(n^{c - 2})" + }, + { + "bbox": [ + 304, + 72, + 545, + 158 + ], + "type": "text", + "content": " cycle candidates, we know that " + }, + { + "bbox": [ + 304, + 72, + 545, + 158 + ], + "type": "inline_equation", + "content": "|N_{ij}^{c}|\\sim O(n^{c - 2})" + }, + { + "bbox": [ + 304, + 72, + 545, + 158 + ], + "type": "text", + "content": " for each " + }, + { + "bbox": [ + 304, + 72, + 545, + 158 + ], + "type": "inline_equation", + "content": "ij\\in E" + }, + { + "bbox": [ + 304, + 72, + 545, + 158 + ], + "type": "text", + "content": " in the worst case scenario of a dense graph. Therefore the initialization of CEMP takes " + }, + { + "bbox": [ + 304, + 72, + 545, + 158 + ], + "type": "inline_equation", + "content": "O(cd^3 n^{c - 2}|E|)" + }, + { + "bbox": [ + 304, + 72, + 545, + 158 + ], + "type": "text", + "content": " time and " + }, + { + "bbox": [ + 304, + 72, + 545, + 158 + ], + "type": "inline_equation", + "content": "O(d^{2}n^{c - 2}|E|)" + }, + { + "bbox": [ + 304, + 72, + 545, + 158 + ], + "type": "text", + "content": " space in the worst case. Given " + }, + { + "bbox": [ + 304, + 72, + 545, + 158 + ], + "type": "inline_equation", + "content": "c\\geq 4" + }, + { + "bbox": [ + 304, + 72, + 545, + 158 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 72, + 545, + 158 + ], + "type": "inline_equation", + "content": "|E|\\sim n^2" + }, + { + "bbox": [ + 304, + 72, + 545, + 158 + ], + "type": "text", + "content": ", CEMP requires much more time and space than LongSync." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 169, + 511, + 183 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 169, + 511, + 183 + ], + "spans": [ + { + "bbox": [ + 305, + 169, + 511, + 183 + ], + "type": "text", + "content": "4. Theory for Uniform Corruption Model" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 190, + 545, + 275 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 190, + 545, + 275 + ], + "spans": [ + { + "bbox": [ + 304, + 190, + 545, + 275 + ], + "type": "text", + "content": "In this section, we present the exact recovery guarantee of LongSync under the uniform corruption model (UCM). UCM is a popular probabilistic model that is widely adopted for synthetic experiments of many previous works on group synchronization [8, 25, 28, 29, 32, 42]. The model " + }, + { + "bbox": [ + 304, + 190, + 545, + 275 + ], + "type": "inline_equation", + "content": "\\mathrm{UCM}(n,p,q_g)" + }, + { + "bbox": [ + 304, + 190, + 545, + 275 + ], + "type": "text", + "content": " assumes that " + }, + { + "bbox": [ + 304, + 190, + 545, + 275 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 304, + 190, + 545, + 275 + ], + "type": "text", + "content": " is an Erdős-Rényi graph with edge connection probability " + }, + { + "bbox": [ + 304, + 190, + 545, + 275 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 304, + 190, + 545, + 275 + ], + "type": "text", + "content": ". For each edge " + }, + { + "bbox": [ + 304, + 190, + 545, + 275 + ], + "type": "inline_equation", + "content": "ij \\in E" + }, + { + "bbox": [ + 304, + 190, + 545, + 275 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 190, + 545, + 275 + ], + "type": "inline_equation", + "content": "R_{ij}" + }, + { + "bbox": [ + 304, + 190, + 545, + 275 + ], + "type": "text", + "content": " is generated independently as follows:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 352, + 285, + 497, + 316 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 352, + 285, + 497, + 316 + ], + "spans": [ + { + "bbox": [ + 352, + 285, + 497, + 316 + ], + "type": "interline_equation", + "content": "\\boldsymbol {R} _ {i j} = \\left\\{ \\begin{array}{l l} \\boldsymbol {R} _ {i j} ^ {*} & \\text {w . p .} q _ {g}; \\\\ \\tilde {\\boldsymbol {R}} _ {i j} \\sim \\operatorname {H a a r} (\\mathcal {G}) & \\text {w . p .} 1 - q _ {g}. \\end{array} \\right.", + "image_path": "f922c91de9bd23eb988ab570165a6c881f46e2d9bf7d499d906b7ad4c9ce358a.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 322, + 545, + 406 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 322, + 545, + 406 + ], + "spans": [ + { + "bbox": [ + 304, + 322, + 545, + 406 + ], + "type": "text", + "content": "We also developed an exact recovery theory for a general model of adversarial corruption, which we include in section C.1 of the supplementary material. An informal version of our main result for UCM is stated in Theorem 4.1. Although the application of this paper is focused on rotation synchronization, the following theory for UCM is valid for any compact group " + }, + { + "bbox": [ + 304, + 322, + 545, + 406 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 304, + 322, + 545, + 406 + ], + "type": "text", + "content": ", as explained in the supplementary material." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 415, + 545, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 415, + 545, + 453 + ], + "spans": [ + { + "bbox": [ + 304, + 415, + 545, + 453 + ], + "type": "text", + "content": "Theorem 4.1. Let " + }, + { + "bbox": [ + 304, + 415, + 545, + 453 + ], + "type": "inline_equation", + "content": "0 < r < 1" + }, + { + "bbox": [ + 304, + 415, + 545, + 453 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 415, + 545, + 453 + ], + "type": "inline_equation", + "content": "0 < q < 1" + }, + { + "bbox": [ + 304, + 415, + 545, + 453 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 415, + 545, + 453 + ], + "type": "inline_equation", + "content": "0 < p \\leq 1" + }, + { + "bbox": [ + 304, + 415, + 545, + 453 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 415, + 545, + 453 + ], + "type": "inline_equation", + "content": "\\mathcal{G} = SO(3)" + }, + { + "bbox": [ + 304, + 415, + 545, + 453 + ], + "type": "text", + "content": ". Assume LongSync is applied with cycles of length " + }, + { + "bbox": [ + 304, + 415, + 545, + 453 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 304, + 415, + 545, + 453 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 415, + 545, + 453 + ], + "type": "inline_equation", + "content": "n / \\log n \\sim p^{-(c-1) / (c-2-\\epsilon)} q_g^{-7(c-1) / 3(c-2)}" + }, + { + "bbox": [ + 304, + 415, + 545, + 453 + ], + "type": "text", + "content": " for some " + }, + { + "bbox": [ + 304, + 415, + 545, + 453 + ], + "type": "inline_equation", + "content": "\\epsilon > 0" + }, + { + "bbox": [ + 304, + 415, + 545, + 453 + ], + "type": "text", + "content": " and" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 372, + 464, + 478, + 477 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 372, + 464, + 478, + 477 + ], + "spans": [ + { + "bbox": [ + 372, + 464, + 478, + 477 + ], + "type": "interline_equation", + "content": "1 / \\beta_ {t + 1} = r / \\beta_ {t} f o r a l l t \\geq 1.", + "image_path": "8344c401aa33713aaa368c14cab62e642dd50f12ad0edcaaf06fe6b04477d93d.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 487, + 545, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 487, + 545, + 515 + ], + "spans": [ + { + "bbox": [ + 304, + 487, + 545, + 515 + ], + "type": "text", + "content": "Then with appropriate choices of " + }, + { + "bbox": [ + 304, + 487, + 545, + 515 + ], + "type": "inline_equation", + "content": "\\beta_0, \\beta_1, r" + }, + { + "bbox": [ + 304, + 487, + 545, + 515 + ], + "type": "text", + "content": ", and high probability, " + }, + { + "bbox": [ + 304, + 487, + 545, + 515 + ], + "type": "inline_equation", + "content": "\\max_{ij \\in E} |s_{ij}^* - s_{ij}^{(t)}| \\leq \\frac{1}{2c\\beta_t}" + }, + { + "bbox": [ + 304, + 487, + 545, + 515 + ], + "type": "text", + "content": " for all " + }, + { + "bbox": [ + 304, + 487, + 545, + 515 + ], + "type": "inline_equation", + "content": "t \\geq 1" + }, + { + "bbox": [ + 304, + 487, + 545, + 515 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 521, + 545, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 521, + 545, + 640 + ], + "spans": [ + { + "bbox": [ + 304, + 521, + 545, + 640 + ], + "type": "text", + "content": "The major difficulty of proving Theorem 4.1 is the dependence in the cycle inconsistency measures for cycles in " + }, + { + "bbox": [ + 304, + 521, + 545, + 640 + ], + "type": "inline_equation", + "content": "N_{ij}^{c}" + }, + { + "bbox": [ + 304, + 521, + 545, + 640 + ], + "type": "text", + "content": " when " + }, + { + "bbox": [ + 304, + 521, + 545, + 640 + ], + "type": "inline_equation", + "content": "c \\geq 4" + }, + { + "bbox": [ + 304, + 521, + 545, + 640 + ], + "type": "text", + "content": ". Unlike the 3-cycle case, the cycle inconsistency measure of a 4-cycle " + }, + { + "bbox": [ + 304, + 521, + 545, + 640 + ], + "type": "inline_equation", + "content": "L_{1} = (ik_{1}, k_{1}k_{2}, k_{2}j)" + }, + { + "bbox": [ + 304, + 521, + 545, + 640 + ], + "type": "text", + "content": " is correlated with that of " + }, + { + "bbox": [ + 304, + 521, + 545, + 640 + ], + "type": "inline_equation", + "content": "L_{2} = (ik_{1}, k_{1}k_{3}, k_{3}j)" + }, + { + "bbox": [ + 304, + 521, + 545, + 640 + ], + "type": "text", + "content": " under UCM. Therefore the key concentration inequalities for the proof cannot be concluded from the standard Chernoff bounds. To overcome this theoretical obstacle, we have integrated various mathematical techniques from [3, 7, 22-24, 48] to prove the theorem, whose details are included in the supplementary material." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 642, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 642, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 642, + 545, + 712 + ], + "type": "text", + "content": "Theorem 4.1 provides an upper bound of the sample complexity (the required graph size " + }, + { + "bbox": [ + 304, + 642, + 545, + 712 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 304, + 642, + 545, + 712 + ], + "type": "text", + "content": ") of LongSync for exact recovery of the ground truth solutions. This sample complexity is the closest to the information theoretic bound among all existing rotation synchronization methods. The comparison with previous works is summarized in Table 2." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5264" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 49, + 70, + 286, + 201 + ], + "blocks": [ + { + "bbox": [ + 49, + 70, + 286, + 201 + ], + "lines": [ + { + "bbox": [ + 49, + 70, + 286, + 201 + ], + "spans": [ + { + "bbox": [ + 49, + 70, + 286, + 201 + ], + "type": "table", + "html": "
ReferenceSample Complexity
[25] for CEMPO(p-2qg-28/3)
[29] for ReSyncO(p-2qg-7)
Ours for CEMPO(p-2-εqg-14/3)
Ours for LongSync (c=3)O(p-2-εqg-14/3)
Ours for LongSync (c=4)O(p-1.5-εqg-3.5)
Ours for LongSync (any c)O(p-7(c-1/c-2-ε) qg-3(c-2))
Ours for LongSync (c→∞)O(p-1-εqg-7/3)
Information Theoretic Bound [6]O(p-1qg-2)
", + "image_path": "1ed7eb623de572ff938b6db1ad89be5e8a112833468deb33a8f68e368b7759cc.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 209, + 288, + 243 + ], + "lines": [ + { + "bbox": [ + 46, + 209, + 288, + 243 + ], + "spans": [ + { + "bbox": [ + 46, + 209, + 288, + 243 + ], + "type": "text", + "content": "Table 2. Comparison of the sample complexity requirement. Lower absolute values of the powers on " + }, + { + "bbox": [ + 46, + 209, + 288, + 243 + ], + "type": "inline_equation", + "content": "p, q_{g}" + }, + { + "bbox": [ + 46, + 209, + 288, + 243 + ], + "type": "text", + "content": " indicate better results. " + }, + { + "bbox": [ + 46, + 209, + 288, + 243 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 46, + 209, + 288, + 243 + ], + "type": "text", + "content": " is an arbitrarily small positive real number." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 47, + 261, + 194, + 276 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 261, + 194, + 276 + ], + "spans": [ + { + "bbox": [ + 47, + 261, + 194, + 276 + ], + "type": "text", + "content": "5. Synthetic Data Experiment" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 281, + 287, + 378 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 281, + 287, + 378 + ], + "spans": [ + { + "bbox": [ + 46, + 281, + 287, + 378 + ], + "type": "text", + "content": "We test LongSync on synthetic datasets generated with Uniform Corruption Model (UCM) and Uniform Bipartite Corruption Model (UBCM) respectively described in §5.1 and 5.2. For both models with their corresponding viewing graphs " + }, + { + "bbox": [ + 46, + 281, + 287, + 378 + ], + "type": "inline_equation", + "content": "G = ([n], E)" + }, + { + "bbox": [ + 46, + 281, + 287, + 378 + ], + "type": "text", + "content": ", we sample the ground truth absolute rotation matrices " + }, + { + "bbox": [ + 46, + 281, + 287, + 378 + ], + "type": "inline_equation", + "content": "\\{R_i^*\\}_{i \\in [n]}" + }, + { + "bbox": [ + 46, + 281, + 287, + 378 + ], + "type": "text", + "content": " independently from the Haar measure on SO(3), and we generate the observed relative rotations " + }, + { + "bbox": [ + 46, + 281, + 287, + 378 + ], + "type": "inline_equation", + "content": "\\{R_{ij}\\}_{ij \\in E}" + }, + { + "bbox": [ + 46, + 281, + 287, + 378 + ], + "type": "text", + "content": " independently as follows:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 83, + 386, + 249, + 418 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 83, + 386, + 249, + 418 + ], + "spans": [ + { + "bbox": [ + 83, + 386, + 249, + 418 + ], + "type": "interline_equation", + "content": "\\boldsymbol {R} _ {i j} = \\left\\{ \\begin{array}{l l} \\boldsymbol {R} _ {i} ^ {*} \\boldsymbol {R} _ {j} ^ {*} & \\text {w . p .} q _ {g}; \\\\ \\tilde {\\boldsymbol {R}} _ {i j} \\sim \\operatorname {H a a r} (S O (3)) & \\text {w . p .} 1 - q _ {g}. \\end{array} \\right.", + "image_path": "db8892600edabcce16da7068788a1faf9633491169bf23cc0c6de0f4175179d5.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 426, + 287, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 426, + 287, + 594 + ], + "spans": [ + { + "bbox": [ + 46, + 426, + 287, + 594 + ], + "type": "text", + "content": "We use LongSync with cycle length " + }, + { + "bbox": [ + 46, + 426, + 287, + 594 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 46, + 426, + 287, + 594 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 426, + 287, + 594 + ], + "type": "inline_equation", + "content": "\\beta_{t} = \\min (2^{t},20)" + }, + { + "bbox": [ + 46, + 426, + 287, + 594 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 426, + 287, + 594 + ], + "type": "inline_equation", + "content": "T = 10" + }, + { + "bbox": [ + 46, + 426, + 287, + 594 + ], + "type": "text", + "content": " and record the edge weights. For UCM we set " + }, + { + "bbox": [ + 46, + 426, + 287, + 594 + ], + "type": "inline_equation", + "content": "c = 4,5" + }, + { + "bbox": [ + 46, + 426, + 287, + 594 + ], + "type": "text", + "content": " and for UBCM we only use " + }, + { + "bbox": [ + 46, + 426, + 287, + 594 + ], + "type": "inline_equation", + "content": "c = 4" + }, + { + "bbox": [ + 46, + 426, + 287, + 594 + ], + "type": "text", + "content": " since no 5-cycles exist. For our method, we first build a weighted graph whose edge weights are estimated by LongSync. We then extract a maximum spanning tree (MST) of the resulting weighted graph. The resulting spanning tree is expected to be the cleanest possible spanning tree. To initialize our solution of absolute rotations, we first fix " + }, + { + "bbox": [ + 46, + 426, + 287, + 594 + ], + "type": "inline_equation", + "content": "\\pmb{R}_{1}" + }, + { + "bbox": [ + 46, + 426, + 287, + 594 + ], + "type": "text", + "content": " as the identity rotation, and find the rest of " + }, + { + "bbox": [ + 46, + 426, + 287, + 594 + ], + "type": "inline_equation", + "content": "\\pmb{R}_{i}" + }, + { + "bbox": [ + 46, + 426, + 287, + 594 + ], + "type": "text", + "content": "'s by consecutively multiplying the relative rotations along the spanning tree using the formula " + }, + { + "bbox": [ + 46, + 426, + 287, + 594 + ], + "type": "inline_equation", + "content": "\\pmb{R}_{i} = \\pmb{R}_{ij}\\pmb{R}_{j}" + }, + { + "bbox": [ + 46, + 426, + 287, + 594 + ], + "type": "text", + "content": ". To refine our initialized solution, we apply IRLS with Geman-McClure [4] loss functions to minimize " + }, + { + "bbox": [ + 46, + 426, + 287, + 594 + ], + "type": "inline_equation", + "content": "\\sum_{ij\\in E}\\rho_{GM}(d_{\\angle}(\\pmb{R}_{ij},\\pmb{R}_i\\pmb{R}_j^T))" + }, + { + "bbox": [ + 46, + 426, + 287, + 594 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 426, + 287, + 594 + ], + "type": "inline_equation", + "content": "d_{\\angle}" + }, + { + "bbox": [ + 46, + 426, + 287, + 594 + ], + "type": "text", + "content": " is the geodesic distance in " + }, + { + "bbox": [ + 46, + 426, + 287, + 594 + ], + "type": "inline_equation", + "content": "SO(3)" + }, + { + "bbox": [ + 46, + 426, + 287, + 594 + ], + "type": "text", + "content": ". We refer to this method as LongSync+IRLS." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 594, + 286, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 286, + 677 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 286, + 677 + ], + "type": "text", + "content": "To demonstrate the advantages of utilizing longer cycle information, we compare our method with IRLS initialized by other two different spanning trees. The first one is the random spanning tree, which uses no cycle information. The other one the MST extracted from the CEMP-estimated weights. Note that CEMP only uses 3-cycle information. We refer to these methods as IRLS and CEMP+IRLS respectively." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 677, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 677, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 677, + 287, + 714 + ], + "type": "text", + "content": "Since the solution of absolute rotations is determined up to a global rotation, we align our estimated rotation " + }, + { + "bbox": [ + 47, + 677, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\{\\hat{R}_i\\}" + }, + { + "bbox": [ + 47, + 677, + 287, + 714 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 47, + 677, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\{R_i^*\\}" + }, + { + "bbox": [ + 47, + 677, + 287, + 714 + ], + "type": "text", + "content": " by " + }, + { + "bbox": [ + 47, + 677, + 287, + 714 + ], + "type": "inline_equation", + "content": "R_{\\mathrm{align}}" + }, + { + "bbox": [ + 47, + 677, + 287, + 714 + ], + "type": "text", + "content": " that minimizes the " + }, + { + "bbox": [ + 47, + 677, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\ell_1" + }, + { + "bbox": [ + 47, + 677, + 287, + 714 + ], + "type": "text", + "content": " rotation alignment error" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 353, + 77, + 494, + 184 + ], + "blocks": [ + { + "bbox": [ + 353, + 77, + 494, + 184 + ], + "lines": [ + { + "bbox": [ + 353, + 77, + 494, + 184 + ], + "spans": [ + { + "bbox": [ + 353, + 77, + 494, + 184 + ], + "type": "image", + "image_path": "6968ab42c664c6e8f93e8d667cc771f80578269e0b14294c9b16ed4a19989a1c.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 353, + 189, + 494, + 296 + ], + "blocks": [ + { + "bbox": [ + 353, + 189, + 494, + 296 + ], + "lines": [ + { + "bbox": [ + 353, + 189, + 494, + 296 + ], + "spans": [ + { + "bbox": [ + 353, + 189, + 494, + 296 + ], + "type": "image", + "image_path": "669a3c1c2a165ca8dab3004f7386cecf1dcb8051b66a15b01b38f281cb7c17b2.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 304, + 545, + 361 + ], + "lines": [ + { + "bbox": [ + 305, + 304, + 545, + 361 + ], + "spans": [ + { + "bbox": [ + 305, + 304, + 545, + 361 + ], + "type": "text", + "content": "Figure 1. Average errors for IRLS, CEMP+IRLS and LongSync +IRLS with " + }, + { + "bbox": [ + 305, + 304, + 545, + 361 + ], + "type": "inline_equation", + "content": "c = 4,5" + }, + { + "bbox": [ + 305, + 304, + 545, + 361 + ], + "type": "text", + "content": ", using the uniform corruption (top) and uniform bipartite corruption (bottom) models. The mean errors are measured in degrees. LongSync4 and LongSync5 refer to LongSync with 4 and 5 cycles, respectively." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 382, + 545, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 382, + 545, + 413 + ], + "spans": [ + { + "bbox": [ + 306, + 382, + 545, + 413 + ], + "type": "inline_equation", + "content": "\\sum_{i\\in [n]}\\| \\hat{\\pmb{R}}_i\\pmb{R}_{\\mathrm{align}} - \\pmb{R}_i^*\\| _F" + }, + { + "bbox": [ + 306, + 382, + 545, + 413 + ], + "type": "text", + "content": " . We report the mean estimation error in degrees: " + }, + { + "bbox": [ + 306, + 382, + 545, + 413 + ], + "type": "inline_equation", + "content": "180\\cdot \\sum_{i\\in [n]}d_{\\angle}(\\hat{\\pmb{R}}_i\\pmb{R}_{\\mathrm{align}},\\pmb{R}_i^*) / n." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 422, + 458, + 435 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 422, + 458, + 435 + ], + "spans": [ + { + "bbox": [ + 306, + 422, + 458, + 435 + ], + "type": "text", + "content": "5.1. Uniform Corruption Model" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 441, + 545, + 490 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 441, + 545, + 490 + ], + "spans": [ + { + "bbox": [ + 305, + 441, + 545, + 490 + ], + "type": "text", + "content": "For " + }, + { + "bbox": [ + 305, + 441, + 545, + 490 + ], + "type": "inline_equation", + "content": "\\mathrm{UCM}(n,p,q_g)" + }, + { + "bbox": [ + 305, + 441, + 545, + 490 + ], + "type": "text", + "content": ", we take " + }, + { + "bbox": [ + 305, + 441, + 545, + 490 + ], + "type": "inline_equation", + "content": "n = 200" + }, + { + "bbox": [ + 305, + 441, + 545, + 490 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 441, + 545, + 490 + ], + "type": "inline_equation", + "content": "p = 1" + }, + { + "bbox": [ + 305, + 441, + 545, + 490 + ], + "type": "text", + "content": " and corruption probability " + }, + { + "bbox": [ + 305, + 441, + 545, + 490 + ], + "type": "inline_equation", + "content": "q = 1 - q_{g}" + }, + { + "bbox": [ + 305, + 441, + 545, + 490 + ], + "type": "text", + "content": " ranges from 0.86 to 0.92. We report the average mean estimation error from 20 trials of the uniform corruption model in the top panel of Figure 1." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 491, + 545, + 563 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 491, + 545, + 563 + ], + "spans": [ + { + "bbox": [ + 305, + 491, + 545, + 563 + ], + "type": "text", + "content": "We note that LongSync uniformly improves IRLS, and the mean error of LongSync decreases as the cycle length increases. When " + }, + { + "bbox": [ + 305, + 491, + 545, + 563 + ], + "type": "inline_equation", + "content": "q > 0.86" + }, + { + "bbox": [ + 305, + 491, + 545, + 563 + ], + "type": "text", + "content": ", the expected number of clean 3-cycles for each edge is less than 4, and therefore longer cycles are helpful. The numerical result aligns with our theory that using longer cycles may tolerate higher corruption with fixed graph size " + }, + { + "bbox": [ + 305, + 491, + 545, + 563 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 305, + 491, + 545, + 563 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 573, + 502, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 573, + 502, + 586 + ], + "spans": [ + { + "bbox": [ + 306, + 573, + 502, + 586 + ], + "type": "text", + "content": "5.2. Uniform Bipartite Corruption Model" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 305, + 593, + 545, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 593, + 545, + 676 + ], + "spans": [ + { + "bbox": [ + 305, + 593, + 545, + 676 + ], + "type": "text", + "content": "For UBCM, we first generate the graph and relative rotations by " + }, + { + "bbox": [ + 305, + 593, + 545, + 676 + ], + "type": "inline_equation", + "content": "\\mathrm{UCM}(n,p,q_g)" + }, + { + "bbox": [ + 305, + 593, + 545, + 676 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 305, + 593, + 545, + 676 + ], + "type": "inline_equation", + "content": "n = 200" + }, + { + "bbox": [ + 305, + 593, + 545, + 676 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 305, + 593, + 545, + 676 + ], + "type": "inline_equation", + "content": "p = 1" + }, + { + "bbox": [ + 305, + 593, + 545, + 676 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 593, + 545, + 676 + ], + "type": "inline_equation", + "content": "q = 1 - q_{g}" + }, + { + "bbox": [ + 305, + 593, + 545, + 676 + ], + "type": "text", + "content": " ranging from 0.8 to 0.85. Then we split the nodes into two clusters of equal size and remove the intra-cluster edges for both clusters. The resulting graph is bipartite, where only cycles of even lengths exist. We report the mean estimation error from 20 trials in the bottom panel of Figure 1." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 305, + 677, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 677, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 677, + 545, + 713 + ], + "type": "text", + "content": "We observe that LongSync with 4-cycles almost exactly recovers the rotations, while for other algorithms the rotation estimates are not even close to the ground truth." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5265" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 71, + 171, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 71, + 171, + 85 + ], + "spans": [ + { + "bbox": [ + 47, + 71, + 171, + 85 + ], + "type": "text", + "content": "6. Real Data Experiment" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 94, + 289, + 225 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 94, + 289, + 225 + ], + "spans": [ + { + "bbox": [ + 46, + 94, + 289, + 225 + ], + "type": "text", + "content": "We test distributed synchronization with LongSync on the PhotoTourism dataset [50] to demonstrate its advantages in accuracy and speed over other baselines. PhotoTourism is a large scale dataset consisting of 15 sets of images taken for 3D reconstruction. The smallest dataset consists of 247 cameras, and the largest dataset consists of 5433 cameras. The input graph and initial pairwise rotation estimates are provided in the dataset. In the following, we first explain the common steps for distributed synchronization, and our improvement using LongSync. We then describe our graph processing method for filtering bad nodes and edges, which also is applicable to other baseline methods." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 226, + 202, + 237 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 226, + 202, + 237 + ], + "spans": [ + { + "bbox": [ + 47, + 226, + 202, + 237 + ], + "type": "text", + "content": "Steps in Distributed Synchronization:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 239, + 287, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 239, + 287, + 300 + ], + "spans": [ + { + "bbox": [ + 47, + 239, + 287, + 300 + ], + "type": "text", + "content": "1. Graph partitioning. The first step involves partitioning the graph " + }, + { + "bbox": [ + 47, + 239, + 287, + 300 + ], + "type": "inline_equation", + "content": "G = ([n], E)" + }, + { + "bbox": [ + 47, + 239, + 287, + 300 + ], + "type": "text", + "content": " into " + }, + { + "bbox": [ + 47, + 239, + 287, + 300 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 47, + 239, + 287, + 300 + ], + "type": "text", + "content": " clusters " + }, + { + "bbox": [ + 47, + 239, + 287, + 300 + ], + "type": "inline_equation", + "content": "G_{i} = (V_{i}, E_{i}), i \\in [K]" + }, + { + "bbox": [ + 47, + 239, + 287, + 300 + ], + "type": "text", + "content": ". In this paper we apply spectral clustering algorithm [37] on the adjacency matrix " + }, + { + "bbox": [ + 47, + 239, + 287, + 300 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 47, + 239, + 287, + 300 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 47, + 239, + 287, + 300 + ], + "type": "inline_equation", + "content": "K = 0.6\\sqrt{np}" + }, + { + "bbox": [ + 47, + 239, + 287, + 300 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 239, + 287, + 300 + ], + "type": "inline_equation", + "content": "p = 2|E| / (n(n - 1))" + }, + { + "bbox": [ + 47, + 239, + 287, + 300 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 300, + 288, + 385 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 300, + 288, + 385 + ], + "spans": [ + { + "bbox": [ + 47, + 300, + 288, + 385 + ], + "type": "text", + "content": "2. Synchronization within clusters. Run standard synchronization solvers for each cluster. In this work, we use the current state-of-the-art method MPLS [39]. Note that for each camera " + }, + { + "bbox": [ + 47, + 300, + 288, + 385 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 47, + 300, + 288, + 385 + ], + "type": "text", + "content": " in cluster " + }, + { + "bbox": [ + 47, + 300, + 288, + 385 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 47, + 300, + 288, + 385 + ], + "type": "text", + "content": ", one can only estimate the true rotation " + }, + { + "bbox": [ + 47, + 300, + 288, + 385 + ], + "type": "inline_equation", + "content": "R_{p}^{*}" + }, + { + "bbox": [ + 47, + 300, + 288, + 385 + ], + "type": "text", + "content": " up to a global rotation " + }, + { + "bbox": [ + 47, + 300, + 288, + 385 + ], + "type": "inline_equation", + "content": "R_{k}" + }, + { + "bbox": [ + 47, + 300, + 288, + 385 + ], + "type": "text", + "content": ". Namely, one only obtains " + }, + { + "bbox": [ + 47, + 300, + 288, + 385 + ], + "type": "inline_equation", + "content": "\\hat{R}^p \\approx R_p^* R_k^{-1}" + }, + { + "bbox": [ + 47, + 300, + 288, + 385 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 47, + 300, + 288, + 385 + ], + "type": "inline_equation", + "content": "R_{k}" + }, + { + "bbox": [ + 47, + 300, + 288, + 385 + ], + "type": "text", + "content": " is unknown and is the same for all cameras in cluster " + }, + { + "bbox": [ + 47, + 300, + 288, + 385 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 47, + 300, + 288, + 385 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 386, + 287, + 532 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 386, + 287, + 532 + ], + "spans": [ + { + "bbox": [ + 47, + 386, + 287, + 532 + ], + "type": "text", + "content": "3. Estimation of inter-cluster rotations. To find " + }, + { + "bbox": [ + 47, + 386, + 287, + 532 + ], + "type": "inline_equation", + "content": "\\pmb{R}_{p}^{*}" + }, + { + "bbox": [ + 47, + 386, + 287, + 532 + ], + "type": "text", + "content": " of all cameras, one needs to solve " + }, + { + "bbox": [ + 47, + 386, + 287, + 532 + ], + "type": "inline_equation", + "content": "\\pmb{R}_{k}" + }, + { + "bbox": [ + 47, + 386, + 287, + 532 + ], + "type": "text", + "content": " for all clusters. Namely, one needs to rotate and stitch the solutions of all clusters so that they are in the same reference frame. To do this, it is common to first estimate the inter-clsuter rotations " + }, + { + "bbox": [ + 47, + 386, + 287, + 532 + ], + "type": "inline_equation", + "content": "\\pmb{R}_{kl} := \\pmb{R}_{k}\\pmb{R}_{l}^{-1}" + }, + { + "bbox": [ + 47, + 386, + 287, + 532 + ], + "type": "text", + "content": " between pairs of clusters " + }, + { + "bbox": [ + 47, + 386, + 287, + 532 + ], + "type": "inline_equation", + "content": "k,l" + }, + { + "bbox": [ + 47, + 386, + 287, + 532 + ], + "type": "text", + "content": ", and then synchronize these relative rotations. To estimate each " + }, + { + "bbox": [ + 47, + 386, + 287, + 532 + ], + "type": "inline_equation", + "content": "\\pmb{R}_{kl}" + }, + { + "bbox": [ + 47, + 386, + 287, + 532 + ], + "type": "text", + "content": ", we note that " + }, + { + "bbox": [ + 47, + 386, + 287, + 532 + ], + "type": "inline_equation", + "content": "\\pmb{R}_{kl} = \\pmb{R}_{p}^{*-1}\\pmb{R}_{pq}^{*}\\pmb{R}_{q}^{*}" + }, + { + "bbox": [ + 47, + 386, + 287, + 532 + ], + "type": "text", + "content": " for each " + }, + { + "bbox": [ + 47, + 386, + 287, + 532 + ], + "type": "inline_equation", + "content": "p\\in V_k" + }, + { + "bbox": [ + 47, + 386, + 287, + 532 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 386, + 287, + 532 + ], + "type": "inline_equation", + "content": "q\\in V_l" + }, + { + "bbox": [ + 47, + 386, + 287, + 532 + ], + "type": "text", + "content": ". Therefore, one can use the rotations in the set " + }, + { + "bbox": [ + 47, + 386, + 287, + 532 + ], + "type": "inline_equation", + "content": "S_{kl} := \\{\\hat{\\pmb{R}}_p^{-1}\\pmb{R}_{pq}\\hat{\\pmb{R}}_q\\}_{p\\in V_k,q\\in V_l}" + }, + { + "bbox": [ + 47, + 386, + 287, + 532 + ], + "type": "text", + "content": " to approximate " + }, + { + "bbox": [ + 47, + 386, + 287, + 532 + ], + "type": "inline_equation", + "content": "\\pmb{R}_{kl}" + }, + { + "bbox": [ + 47, + 386, + 287, + 532 + ], + "type": "text", + "content": ". We remark that this step is crucial to the overall performance of distributed methods, and we compare the following methods for solving " + }, + { + "bbox": [ + 47, + 386, + 287, + 532 + ], + "type": "inline_equation", + "content": "\\pmb{R}_{kl}" + }, + { + "bbox": [ + 47, + 386, + 287, + 532 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 59, + 532, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 59, + 532, + 287, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 532, + 287, + 579 + ], + "spans": [ + { + "bbox": [ + 59, + 532, + 287, + 579 + ], + "type": "text", + "content": "- MultSync [9]: Run synchronization on a multi-graph where each edge " + }, + { + "bbox": [ + 59, + 532, + 287, + 579 + ], + "type": "inline_equation", + "content": "kl" + }, + { + "bbox": [ + 59, + 532, + 287, + 579 + ], + "type": "text", + "content": " is assigned a set of relative rotations " + }, + { + "bbox": [ + 59, + 532, + 287, + 579 + ], + "type": "inline_equation", + "content": "\\{\\hat{R}_p^{-1}R_{pq}\\hat{R}_q\\}_{p\\in V_k,q\\in V_l}" + }, + { + "bbox": [ + 59, + 532, + 287, + 579 + ], + "type": "text", + "content": ". This combines the step 3 and 4 in a unified least squares formulation." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 59, + 579, + 287, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 579, + 287, + 628 + ], + "spans": [ + { + "bbox": [ + 59, + 579, + 287, + 628 + ], + "type": "text", + "content": "- Edge averaging using IRLS: We initialize " + }, + { + "bbox": [ + 59, + 579, + 287, + 628 + ], + "type": "inline_equation", + "content": "\\hat{R}_{kl}" + }, + { + "bbox": [ + 59, + 579, + 287, + 628 + ], + "type": "text", + "content": " with the quaternion " + }, + { + "bbox": [ + 59, + 579, + 287, + 628 + ], + "type": "inline_equation", + "content": "\\ell_2" + }, + { + "bbox": [ + 59, + 579, + 287, + 628 + ], + "type": "text", + "content": " mean of the set " + }, + { + "bbox": [ + 59, + 579, + 287, + 628 + ], + "type": "inline_equation", + "content": "S_{kl}" + }, + { + "bbox": [ + 59, + 579, + 287, + 628 + ], + "type": "text", + "content": " and refine it using " + }, + { + "bbox": [ + 59, + 579, + 287, + 628 + ], + "type": "inline_equation", + "content": "\\ell_1" + }, + { + "bbox": [ + 59, + 579, + 287, + 628 + ], + "type": "text", + "content": "-rotation averaging [19]. We refer to this method as IRLS in our comparison." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 59, + 628, + 287, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 628, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 59, + 628, + 287, + 712 + ], + "type": "text", + "content": "- Edge averaging using LongSync: We first perform LongSync with 4-cycles to estimate the weights of these inter-cluster edges (there are no 3-cycles for a bipartite graph). We next initialize " + }, + { + "bbox": [ + 59, + 628, + 287, + 712 + ], + "type": "inline_equation", + "content": "\\hat{R}_{kl}" + }, + { + "bbox": [ + 59, + 628, + 287, + 712 + ], + "type": "text", + "content": " as the quaternion weighted " + }, + { + "bbox": [ + 59, + 628, + 287, + 712 + ], + "type": "inline_equation", + "content": "\\ell_2" + }, + { + "bbox": [ + 59, + 628, + 287, + 712 + ], + "type": "text", + "content": " mean of " + }, + { + "bbox": [ + 59, + 628, + 287, + 712 + ], + "type": "inline_equation", + "content": "S_{kl}" + }, + { + "bbox": [ + 59, + 628, + 287, + 712 + ], + "type": "text", + "content": ", using the edge weights from LongSync by their LongSync weights. Lastly, we refine the solution using [19]." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 305, + 72, + 547, + 144 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 305, + 72, + 545, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 120 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 120 + ], + "type": "text", + "content": "4. Synchronization of inter-cluster rotations. This step is skipped for MultiSync. For other methods described in step 3, we find " + }, + { + "bbox": [ + 305, + 72, + 545, + 120 + ], + "type": "inline_equation", + "content": "\\pmb{R}_k" + }, + { + "bbox": [ + 305, + 72, + 545, + 120 + ], + "type": "text", + "content": " (up to a rotation) for each cluster " + }, + { + "bbox": [ + 305, + 72, + 545, + 120 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 305, + 72, + 545, + 120 + ], + "type": "text", + "content": " from the estimated " + }, + { + "bbox": [ + 305, + 72, + 545, + 120 + ], + "type": "inline_equation", + "content": "\\{R_{kl}\\}_{k,l\\in [K]}" + }, + { + "bbox": [ + 305, + 72, + 545, + 120 + ], + "type": "text", + "content": " by MPLS." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 121, + 547, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 121, + 547, + 144 + ], + "spans": [ + { + "bbox": [ + 305, + 121, + 547, + 144 + ], + "type": "text", + "content": "5. Rotation merging. Finally, for each camera " + }, + { + "bbox": [ + 305, + 121, + 547, + 144 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 305, + 121, + 547, + 144 + ], + "type": "text", + "content": " in cluster " + }, + { + "bbox": [ + 305, + 121, + 547, + 144 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 305, + 121, + 547, + 144 + ], + "type": "text", + "content": ", the rotation estimate of " + }, + { + "bbox": [ + 305, + 121, + 547, + 144 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 305, + 121, + 547, + 144 + ], + "type": "text", + "content": " is given by " + }, + { + "bbox": [ + 305, + 121, + 547, + 144 + ], + "type": "inline_equation", + "content": "R_p^{\\mathrm{final}} = \\hat{R}_p R_k^{-1}" + }, + { + "bbox": [ + 305, + 121, + 547, + 144 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 305, + 144, + 545, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 144, + 545, + 167 + ], + "spans": [ + { + "bbox": [ + 305, + 144, + 545, + 167 + ], + "type": "text", + "content": "Next, we introduce our graph processing method to further boost the performance of all tested methods." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 168, + 480, + 179 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 168, + 480, + 179 + ], + "spans": [ + { + "bbox": [ + 306, + 168, + 480, + 179 + ], + "type": "text", + "content": "Extra Improvement by Graph Processing:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 180, + 545, + 227 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 180, + 545, + 227 + ], + "spans": [ + { + "bbox": [ + 306, + 180, + 545, + 227 + ], + "type": "text", + "content": "- Spectral clustering with Jaccard Index. For step 1, we use the Jaccard index matrix as the similarity matrix for spectral clustering, instead of the adjacency matrix. The " + }, + { + "bbox": [ + 306, + 180, + 545, + 227 + ], + "type": "inline_equation", + "content": "n \\times n" + }, + { + "bbox": [ + 306, + 180, + 545, + 227 + ], + "type": "text", + "content": " Jaccard index matrix " + }, + { + "bbox": [ + 306, + 180, + 545, + 227 + ], + "type": "inline_equation", + "content": "A_J" + }, + { + "bbox": [ + 306, + 180, + 545, + 227 + ], + "type": "text", + "content": " is defined as follows:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 370, + 233, + 545, + 266 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 370, + 233, + 545, + 266 + ], + "spans": [ + { + "bbox": [ + 370, + 233, + 545, + 266 + ], + "type": "interline_equation", + "content": "\\boldsymbol {A} _ {J} (i, j) = \\left\\{ \\begin{array}{l l} 0 & i j \\notin E \\\\ \\frac {\\left| N _ {i} \\cap N _ {j} \\right|}{\\left| N _ {i} \\cup N _ {j} \\right|} & i j \\in E \\end{array} \\right. \\tag {14}", + "image_path": "11a647d708704b2836d120d7467b82c72f863671fcb7feca34cf508a18255ddd.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 312, + 270, + 545, + 331 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 270, + 545, + 331 + ], + "spans": [ + { + "bbox": [ + 312, + 270, + 545, + 331 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 312, + 270, + 545, + 331 + ], + "type": "inline_equation", + "content": "N_{i}" + }, + { + "bbox": [ + 312, + 270, + 545, + 331 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 312, + 270, + 545, + 331 + ], + "type": "inline_equation", + "content": "N_{j}" + }, + { + "bbox": [ + 312, + 270, + 545, + 331 + ], + "type": "text", + "content": " denote the sets of neighboring nodes of node " + }, + { + "bbox": [ + 312, + 270, + 545, + 331 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 312, + 270, + 545, + 331 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 312, + 270, + 545, + 331 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 312, + 270, + 545, + 331 + ], + "type": "text", + "content": ", respectively. In this way, " + }, + { + "bbox": [ + 312, + 270, + 545, + 331 + ], + "type": "inline_equation", + "content": "\\mathbf{A}_{J}(i,j)" + }, + { + "bbox": [ + 312, + 270, + 545, + 331 + ], + "type": "text", + "content": " is higher for the pair " + }, + { + "bbox": [ + 312, + 270, + 545, + 331 + ], + "type": "inline_equation", + "content": "ij" + }, + { + "bbox": [ + 312, + 270, + 545, + 331 + ], + "type": "text", + "content": " contained in many 3-cycles, which is a more robust and nicely scaled statistics " + }, + { + "bbox": [ + 312, + 270, + 545, + 331 + ], + "type": "inline_equation", + "content": "(\\in [0,1])" + }, + { + "bbox": [ + 312, + 270, + 545, + 331 + ], + "type": "text", + "content": " for measuring the local graph density around edge " + }, + { + "bbox": [ + 312, + 270, + 545, + 331 + ], + "type": "inline_equation", + "content": "ij" + }, + { + "bbox": [ + 312, + 270, + 545, + 331 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 306, + 331, + 546, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 331, + 546, + 521 + ], + "spans": [ + { + "bbox": [ + 306, + 331, + 546, + 521 + ], + "type": "text", + "content": "- Refinement of intra-cluster edges and nodes. For step 2, after the MPLS step, we perform CEMP with 3-cycles to estimate the corruption level of the intra-cluster edges for each cluster. We remove a camera if the number of neighboring 'good' edges, i.e. the edges with corruption level less than 0.1, is less than 4. The numbers 4 and 0.1 are chosen to balance the number of remaining cameras and the quality of intra-cluster rotation estimates. In order to eliminate the sparsely connected components inside the cluster, we use the Matlab built-in hierarchical spectral clustering function on the remaining cameras with the 'cutoff' and 'depth' parameters as 2 and 4, and we keep the largest component. The absolute rotations for the remaining cameras are estimated by MPLS. We remark that one could replace CEMP by LongSync with 3-cycles. However, we have not observed significant difference in the performance." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 522, + 547, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 522, + 547, + 677 + ], + "spans": [ + { + "bbox": [ + 304, + 522, + 547, + 677 + ], + "type": "text", + "content": "We respectively name MultiSync and IRLS with our new graph processing method as MultiSync(New) and IRLS(new). \"LongSync\" in our experiment refers to the full version of our algorithm: use LongSync weights for edge averaging in Step 3, with the graph processing step. We also compare with MPLS on the whole dataset, since it is a state-of-the-art non-distributed method, but we note that MPLS is significantly slower than all distributed methods. We report median error " + }, + { + "bbox": [ + 304, + 522, + 547, + 677 + ], + "type": "inline_equation", + "content": "180 \\cdot" + }, + { + "bbox": [ + 304, + 522, + 547, + 677 + ], + "type": "text", + "content": " median(" + }, + { + "bbox": [ + 304, + 522, + 547, + 677 + ], + "type": "inline_equation", + "content": "\\{d_{\\angle}(\\hat{R}_i R_{\\text{align}}, R_i^*)\\}_{i \\in [n]}" + }, + { + "bbox": [ + 304, + 522, + 547, + 677 + ], + "type": "text", + "content": " of the tested methods on 14 datasets in Figure 2. We exclude the result of Gendarmenmarkt since all methods return large estimation errors in the figure. The full results, including that of mean error are included in the supplementary material." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "content": "In Figure 3, for each distributed method, we report the ratio (in percentage) between its total runtime on all datasets and that of the non-distributed MPLS. Namely, we compute" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5266" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 53, + 70, + 547, + 216 + ], + "blocks": [ + { + "bbox": [ + 53, + 70, + 547, + 216 + ], + "lines": [ + { + "bbox": [ + 53, + 70, + 547, + 216 + ], + "spans": [ + { + "bbox": [ + 53, + 70, + 547, + 216 + ], + "type": "image", + "image_path": "bda3fb0b64824766e835611341624dac709532541ead448fa1660bf7d5e90c98.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 169, + 224, + 424, + 236 + ], + "lines": [ + { + "bbox": [ + 169, + 224, + 424, + 236 + ], + "spans": [ + { + "bbox": [ + 169, + 224, + 424, + 236 + ], + "type": "text", + "content": "Figure 2. Median error for rotations for each dataset measured in degrees." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 62, + 254, + 272, + 425 + ], + "blocks": [ + { + "bbox": [ + 62, + 254, + 272, + 425 + ], + "lines": [ + { + "bbox": [ + 62, + 254, + 272, + 425 + ], + "spans": [ + { + "bbox": [ + 62, + 254, + 272, + 425 + ], + "type": "image", + "image_path": "baccd76e5c4d33968b5b02bea6b50a0629181129f71f2c34fc4ce00f4bdf8511.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 433, + 287, + 456 + ], + "lines": [ + { + "bbox": [ + 46, + 433, + 287, + 456 + ], + "spans": [ + { + "bbox": [ + 46, + 433, + 287, + 456 + ], + "type": "text", + "content": "Figure 3. Runtime ratio and average median and mean error gaps between the distributed methods and MPLS on the entire graph." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 482, + 287, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 482, + 287, + 579 + ], + "spans": [ + { + "bbox": [ + 46, + 482, + 287, + 579 + ], + "type": "inline_equation", + "content": "\\sum_{d \\in D} t_{\\mathrm{dist},d} / \\sum_{d \\in D} t_{\\mathrm{MPLS},d}" + }, + { + "bbox": [ + 46, + 482, + 287, + 579 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 482, + 287, + 579 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 46, + 482, + 287, + 579 + ], + "type": "text", + "content": " is the set of 15 datasets, and each " + }, + { + "bbox": [ + 46, + 482, + 287, + 579 + ], + "type": "inline_equation", + "content": "t_{\\mathrm{dist},d}" + }, + { + "bbox": [ + 46, + 482, + 287, + 579 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 482, + 287, + 579 + ], + "type": "inline_equation", + "content": "t_{\\mathrm{MPLS},d}" + }, + { + "bbox": [ + 46, + 482, + 287, + 579 + ], + "type": "text", + "content": " is respectively the runtime of the distributed method and MPLS on data " + }, + { + "bbox": [ + 46, + 482, + 287, + 579 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 46, + 482, + 287, + 579 + ], + "type": "text", + "content": ". In the same figure, we present the mean/median error gap between each distributed method and MPLS. The mean and median error gap is respectively defined as " + }, + { + "bbox": [ + 46, + 482, + 287, + 579 + ], + "type": "inline_equation", + "content": "(\\bar{e}_{\\mathrm{dist}} - \\bar{e}_{\\mathrm{MPLS}}) / \\bar{e}_{\\mathrm{dist}}" + }, + { + "bbox": [ + 46, + 482, + 287, + 579 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 482, + 287, + 579 + ], + "type": "inline_equation", + "content": "(\\hat{e}_{\\mathrm{dist}} - \\hat{e}_{\\mathrm{MPLS}}) / \\hat{e}_{\\mathrm{dist}}" + }, + { + "bbox": [ + 46, + 482, + 287, + 579 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 482, + 287, + 579 + ], + "type": "inline_equation", + "content": "\\bar{e}" + }, + { + "bbox": [ + 46, + 482, + 287, + 579 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 482, + 287, + 579 + ], + "type": "inline_equation", + "content": "\\hat{e}" + }, + { + "bbox": [ + 46, + 482, + 287, + 579 + ], + "type": "text", + "content": " respectively denote the mean and median error over all cameras." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 582, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 582, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 582, + 288, + 715 + ], + "type": "text", + "content": "From Figure 2 and 3, our method outperforms other distributed methods on 13 out of 15 datasets. The most significant improvement in mean error and median error are respectively " + }, + { + "bbox": [ + 46, + 582, + 288, + 715 + ], + "type": "inline_equation", + "content": "28.6\\%" + }, + { + "bbox": [ + 46, + 582, + 288, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 582, + 288, + 715 + ], + "type": "inline_equation", + "content": "46.4\\%" + }, + { + "bbox": [ + 46, + 582, + 288, + 715 + ], + "type": "text", + "content": " (in Notre Dame) compared to the best performing method between IRLS(new) and MultiSync(new). The improvement is even more significant when comparing to the original version of these baseline methods without our graph processing method. The only two datasets without improvement are Gendarmenmarkt and Union Square. Our method is comparable to others on Union Square, and all methods return large errors on Gendarmenmarkt due to many repetitive patterns in its 3D scene." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 256, + 547, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 256, + 547, + 365 + ], + "spans": [ + { + "bbox": [ + 304, + 256, + 547, + 365 + ], + "type": "text", + "content": "The average mean and median error gap between our method and full MPLS are respectively " + }, + { + "bbox": [ + 304, + 256, + 547, + 365 + ], + "type": "inline_equation", + "content": "38.3\\%" + }, + { + "bbox": [ + 304, + 256, + 547, + 365 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 256, + 547, + 365 + ], + "type": "inline_equation", + "content": "18.4\\%" + }, + { + "bbox": [ + 304, + 256, + 547, + 365 + ], + "type": "text", + "content": ". Compared to the best performing method among others, our method reduces the average median error gap by " + }, + { + "bbox": [ + 304, + 256, + 547, + 365 + ], + "type": "inline_equation", + "content": "40.8\\%" + }, + { + "bbox": [ + 304, + 256, + 547, + 365 + ], + "type": "text", + "content": ", and the average mean error gap by " + }, + { + "bbox": [ + 304, + 256, + 547, + 365 + ], + "type": "inline_equation", + "content": "9.6\\%" + }, + { + "bbox": [ + 304, + 256, + 547, + 365 + ], + "type": "text", + "content": ". In terms of runtime, our method is uniformly faster than MultiSync and it is scalable on the largest dataset, taking less than " + }, + { + "bbox": [ + 304, + 256, + 547, + 365 + ], + "type": "inline_equation", + "content": "6\\%" + }, + { + "bbox": [ + 304, + 256, + 547, + 365 + ], + "type": "text", + "content": " of the total runtime of full MPLS. In conclusion, our method significantly improves the result of distributed rotation synchronization without compromising runtime." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 370, + 547, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 370, + 547, + 441 + ], + "spans": [ + { + "bbox": [ + 304, + 370, + 547, + 441 + ], + "type": "text", + "content": "In the supplementary material, we further demonstrate the improvement by our new graph processing method, which significantly improves the results of LongSync (without extra graph processing) in 14 of the 15 datasets. On these 14 datasets, the average reduction on mean error is " + }, + { + "bbox": [ + 304, + 370, + 547, + 441 + ], + "type": "inline_equation", + "content": "59.2\\%" + }, + { + "bbox": [ + 304, + 370, + 547, + 441 + ], + "type": "text", + "content": " and the average reduction on median error is " + }, + { + "bbox": [ + 304, + 370, + 547, + 441 + ], + "type": "inline_equation", + "content": "28.5\\%" + }, + { + "bbox": [ + 304, + 370, + 547, + 441 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 472, + 375, + 484 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 472, + 375, + 484 + ], + "spans": [ + { + "bbox": [ + 306, + 472, + 375, + 484 + ], + "type": "text", + "content": "7. Conclusion" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 498, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 498, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 498, + 547, + 715 + ], + "type": "text", + "content": "We propose LongSync, a robust and efficient algorithm for group synchronization. It modifies and vectorizes CEMP which enables efficient computation when using longer cycles. The theory we developed for LongSync is the strongest among all other existing results under UCM. Experiment shows that LongSync, together with our improved graph preprocessing method, achieves superior accuracy for distributed synchronization on large real datasets with competitive runtime. However, our method also has some limitations. First of all, in theory there is still a small gap of sample complexity from our method to the information theoretic one. Filling this gap is an open problem, which requires new tools and possibly more sophisticated analysis. Second, our graph preprocessing method is quite heuristic, and an automatic way of choosing parameters is needed. Our work also opens a door for some important future directions, including distributed partial permutation synchronization for multi-image matching, angular synchronization for Cryo-EM and Jigsaw Puzzles, and analysis of their algorithms." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5267" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 104, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 104, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 104, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 134 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 134 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 134 + ], + "type": "text", + "content": "[1] Federica Arrigoni, Beatrice Rossi, Pasqualina Fragneto, and Andrea Fusiello. Robust synchronization in so(3) and se(3) via low-rank and sparse matrix decomposition. Computer Vision and Image Understanding, 174:95-113, 2018. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 134, + 288, + 166 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 134, + 288, + 166 + ], + "spans": [ + { + "bbox": [ + 53, + 134, + 288, + 166 + ], + "type": "text", + "content": "[2] A. S. Bandeira, N. Boumal, and A. Singer. Tightness of the maximum likelihood semidefinite relaxation for angular synchronization. arXiv preprint arXiv:1411.3272, 2014. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 166, + 288, + 198 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 166, + 288, + 198 + ], + "spans": [ + { + "bbox": [ + 53, + 166, + 288, + 198 + ], + "type": "text", + "content": "[3] Olivier Bousquet. A Bennett concentration inequality and its application to suprema of empirical processes. C. R. Math. Acad. Sci. Paris, 334(6):495-500, 2002. 5, 14, 16, 21" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 198, + 288, + 241 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 198, + 288, + 241 + ], + "spans": [ + { + "bbox": [ + 53, + 198, + 288, + 241 + ], + "type": "text", + "content": "[4] Avishek Chatterjee and Venu Madhav Govindu. Efficient and robust large-scale rotation averaging. In IEEE International Conference on Computer Vision, ICCV 2013, Sydney, Australia, December 1-8, 2013, pages 521-528, 2013. 2, 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 241, + 288, + 272 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 241, + 288, + 272 + ], + "spans": [ + { + "bbox": [ + 53, + 241, + 288, + 272 + ], + "type": "text", + "content": "[5] Avishek Chatterjee and Venu Madhav Govindu. Robust relative rotation averaging. IEEE Trans. Pattern Anal. Mach. Intell., 40(4):958-972, 2018. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 272, + 288, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 272, + 288, + 304 + ], + "spans": [ + { + "bbox": [ + 53, + 272, + 288, + 304 + ], + "type": "text", + "content": "[6] Yuxin Chen, Changho Suh, and Andrea J. Goldsmith. Information recovery from pairwise measurements. IEEE Trans. Inf. Theory, 62(10):5881-5905, 2016. 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 304, + 288, + 336 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 304, + 288, + 336 + ], + "spans": [ + { + "bbox": [ + 53, + 304, + 288, + 336 + ], + "type": "text", + "content": "[7] Keresztély Corradi and András Hajnal. On the maximal number of independent circuits in a graph. Acta Mathematica Hungarica, 14(3-4):423-439, 1963. 5, 14" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 336, + 288, + 368 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 336, + 288, + 368 + ], + "spans": [ + { + "bbox": [ + 53, + 336, + 288, + 368 + ], + "type": "text", + "content": "[8] Mihai Cucuringu. Synchronization over Z2 and community detection in signed multiplex networks with constraints. J. Complex Networks, 3(3):469-506, 2015. 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 369, + 288, + 411 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 369, + 288, + 411 + ], + "spans": [ + { + "bbox": [ + 53, + 369, + 288, + 411 + ], + "type": "text", + "content": "[9] Andrea Porfiri Dal Cin, Luca Magri, Federica Arrigoni, Andrea Fusiello, and Giacomo Boracchi. Synchronization of group-labelled multi-graphs. In 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 6433-6443, 2021. 2, 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 411, + 288, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 411, + 288, + 487 + ], + "spans": [ + { + "bbox": [ + 48, + 411, + 288, + 487 + ], + "type": "text", + "content": "[10] Frank Dellaert, David M. Rosen, Jing Wu, Robert E. Mahony, and Luca Carlone. Shonan rotation averaging: Global optimality by surfing " + }, + { + "bbox": [ + 48, + 411, + 288, + 487 + ], + "type": "inline_equation", + "content": "\\mathrm{so(p)^n}" + }, + { + "bbox": [ + 48, + 411, + 288, + 487 + ], + "type": "text", + "content": ". In Andrea Vedaldi, Horst Bischof, Thomas Brox, and Jan-Michael Frahm, editors, Computer Vision - ECCV 2020 - 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part VI, volume 12351 of Lecture Notes in Computer Science, pages 292-308. Springer, 2020. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 487, + 288, + 508 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 487, + 288, + 508 + ], + "spans": [ + { + "bbox": [ + 48, + 487, + 288, + 508 + ], + "type": "text", + "content": "[11] Anders Eriksson, Carl Olsson, Fredrik Kahl, and Tat-Jun Chin. Rotation averaging and strong duality, 2017." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 508, + 288, + 562 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 508, + 288, + 562 + ], + "spans": [ + { + "bbox": [ + 48, + 508, + 288, + 562 + ], + "type": "text", + "content": "[12] Anders P. Eriksson, Carl Olsson, Fredrik Kahl, and Tat-Jun Chin. Rotation averaging and strong duality. In 2018 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2018, Salt Lake City, UT, USA, June 18-22, 2018, pages 127-135. Computer Vision Foundation / IEEE Computer Society, 2018." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 562, + 288, + 604 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 562, + 288, + 604 + ], + "spans": [ + { + "bbox": [ + 48, + 562, + 288, + 604 + ], + "type": "text", + "content": "[13] Anders P. Eriksson, Carl Olsson, Fredrik Kahl, and Tat-Jun Chin. Rotation averaging with the chordal distance: Global minimizers and strong duality. IEEE Trans. Pattern Anal. Mach. Intell., 43(1):256-268, 2021. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 604, + 288, + 647 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 604, + 288, + 647 + ], + "spans": [ + { + "bbox": [ + 48, + 604, + 288, + 647 + ], + "type": "text", + "content": "[14] Xiang Gao, Hainan Cui, Menghan Li, Zexiao Xie, and Shuhan Shen. Irav3: Hierarchical incremental rotation averaging on the fly. IEEE Trans. Circuits Syst. Video Technol., 33(4):2001-2006, 2023. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 647, + 288, + 680 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 647, + 288, + 680 + ], + "spans": [ + { + "bbox": [ + 48, + 647, + 288, + 680 + ], + "type": "text", + "content": "[15] Xiang Gao, Hainan Cui, and Shuhan Shen. Incremental rotation averaging revisited and more: A new rotation averaging benchmark. arXiv preprint, abs/2309.16924, 2023." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 48, + 680, + 288, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 680, + 288, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 680, + 288, + 712 + ], + "type": "text", + "content": "[16] Xiang Gao, Lingjie Zhu, Hainan Cui, Zexiao Xie, and Shuhan Shen. IRA++: distributed incremental rotation averaging. IEEE Trans. Circuits Syst. Video Technol., 32(7):4885-4892, 2022. 2" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 705 + ], + "type": "list", + "angle": 0, + "index": 34, + "blocks": [ + { + "bbox": [ + 307, + 73, + 547, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 547, + 127 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 547, + 127 + ], + "type": "text", + "content": "[17] Thomas Goldstein, Paul Hand, Choongbum Lee, Vladislav Voroninski, and Stefano Soatto. Shapefit and shapekick for robust, scalable structure from motion. In Computer Vision - ECCV 2016 - 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part VII, pages 289-304, 2016. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 127, + 547, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 127, + 547, + 171 + ], + "spans": [ + { + "bbox": [ + 307, + 127, + 547, + 171 + ], + "type": "text", + "content": "[18] V.M. Govindu. Lie-algebraic averaging for globally consistent motion estimation. In Proceedings of the 2004 IEEE Computer Society Conference on Computer Vision and Pattern Recognition, 2004. CVPR 2004., volume 1, pages I-I, 2004. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 171, + 547, + 213 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 171, + 547, + 213 + ], + "spans": [ + { + "bbox": [ + 307, + 171, + 547, + 213 + ], + "type": "text", + "content": "[19] R. Hartley, K. Aftab, and J. Trumpf. L1 rotation averaging using the weiszfeld algorithm. In Computer Vision and Pattern Recognition (CVPR), 2011 IEEE Conference on, pages 3041-3048. IEEE, 2011. 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 213, + 547, + 267 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 213, + 547, + 267 + ], + "spans": [ + { + "bbox": [ + 307, + 213, + 547, + 267 + ], + "type": "text", + "content": "[20] Richard I. Hartley, Khurrum Aftab, and Jochen Trumpf. L1 rotation averaging using the weiszfeld algorithm. In The 24th IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2011, Colorado Springs, CO, USA, 20-25 June 2011, pages 3041-3048, 2011. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 267, + 547, + 310 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 267, + 547, + 310 + ], + "spans": [ + { + "bbox": [ + 307, + 267, + 547, + 310 + ], + "type": "text", + "content": "[21] Xiangru Huang, Zhenxiao Liang, Xiaowei Zhou, Yao Xie, Leonidas J. Guibas, and Qixing Huang. Learning transformation synchronization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2019. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 310, + 547, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 310, + 547, + 331 + ], + "spans": [ + { + "bbox": [ + 307, + 310, + 547, + 331 + ], + "type": "text", + "content": "[22] Svante Janson. Poisson approximation for large deviations. Random Structures & Algorithms, 1(2):221-229, 1990. 5, 14, 16" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 331, + 547, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 331, + 547, + 363 + ], + "spans": [ + { + "bbox": [ + 307, + 331, + 547, + 363 + ], + "type": "text", + "content": "[23] Svante Janson. Large deviations for sums of partly dependent random variables. Random Structures & Algorithms, 24(3):234-248, 2004. 18, 21" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 363, + 547, + 394 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 363, + 547, + 394 + ], + "spans": [ + { + "bbox": [ + 307, + 363, + 547, + 394 + ], + "type": "text", + "content": "[24] Jeong Han Kim and Van H Vu. Concentration of multivariate polynomials and its applications. Combinatorica, 20(3):417-434, 2000. 5, 14, 18" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 395, + 547, + 436 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 395, + 547, + 436 + ], + "spans": [ + { + "bbox": [ + 307, + 395, + 547, + 436 + ], + "type": "text", + "content": "[25] Gilad Lerman and Yunpeng Shi. Robust group synchronization via cycle-edge message passing. Foundations of Computational Mathematics, 22(6):1665-1741, 2022. 1, 2, 3, 5, 6, 13, 14, 16, 19, 21" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 437, + 547, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 437, + 547, + 491 + ], + "spans": [ + { + "bbox": [ + 307, + 437, + 547, + 491 + ], + "type": "text", + "content": "[26] Heng Li, Zhaopeng Cui, Shuaicheng Liu, and Ping Tan. RAGO: recurrent graph optimizer for multiple rotation averaging. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2022, New Orleans, LA, USA, June 18-24, 2022, pages 15766-15775. IEEE, 2022. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 491, + 547, + 534 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 491, + 547, + 534 + ], + "spans": [ + { + "bbox": [ + 307, + 491, + 547, + 534 + ], + "type": "text", + "content": "[27] Shaohan Li, Yunpeng Shi, and Gilad Lerman. Fast, accurate and memory-efficient partial permutation synchronization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 15735-15743, June 2022. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 534, + 547, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 534, + 547, + 566 + ], + "spans": [ + { + "bbox": [ + 307, + 534, + 547, + 566 + ], + "type": "text", + "content": "[28] Shuyang Ling. Near-optimal performance bounds for orthogonal and permutation group synchronization via spectral methods. arXiv preprint arXiv:2008.05341, 2020. 5" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 566, + 547, + 619 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 566, + 547, + 619 + ], + "spans": [ + { + "bbox": [ + 307, + 566, + 547, + 619 + ], + "type": "text", + "content": "[29] Huikang Liu, Xiao Li, and Anthony Man-Cho So. Resync: Riemannian subgradient-based robust rotation synchronization. In Advances in Neural Information Processing Systems: 37th Annual Conference on Neural Information Processing Systems, 2023. 2, 5, 6" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 307, + 619, + 547, + 640 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 619, + 547, + 640 + ], + "spans": [ + { + "bbox": [ + 307, + 619, + 547, + 640 + ], + "type": "text", + "content": "[30] Tyler Maunu and Gilad Lerman. Depth descent synchronization in SO(D). Int. J. Comput. Vis., 131(4):968-986, 2023. 2" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 307, + 640, + 547, + 683 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 640, + 547, + 683 + ], + "spans": [ + { + "bbox": [ + 307, + 640, + 547, + 683 + ], + "type": "text", + "content": "[31] Onur Özyesil and Amit Singer. Robust camera location estimation by convex programming. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 2674-2683, 2015. 1" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 307, + 683, + 547, + 705 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 683, + 547, + 705 + ], + "spans": [ + { + "bbox": [ + 307, + 683, + 547, + 705 + ], + "type": "text", + "content": "[32] Deepti Pachauri, Risi Kondor, and Vikas Singh. Solving the multi-way matching problem by permutation synchronization. In" + } + ] + } + ], + "index": 33 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5268" + } + ] + } + ], + "index": 35 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 704 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 66, + 72, + 287, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 72, + 287, + 106 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 287, + 106 + ], + "type": "text", + "content": "C. J. C. Burges, L. Bottou, M. Welling, Z. Ghahramani, and K. Q. Weinberger, editors, Advances in Neural Information Processing Systems 26, pages 1860-1868. Curran Associates, Inc., 2013. 5" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 106, + 288, + 159 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 106, + 288, + 159 + ], + "spans": [ + { + "bbox": [ + 48, + 106, + 288, + 159 + ], + "type": "text", + "content": "[33] Álvaro Parra, Shin-Fang Ch'ng, Tat-Jun Chin, Anders P. Eriksson, and Ian Reid. Rotation coordinate descent for fast globally optimal rotation averaging. In IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2021, virtual, June 19-25, 2021, pages 4298-4307. Computer Vision Foundation / IEEE, 2021. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 159, + 288, + 179 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 159, + 288, + 179 + ], + "spans": [ + { + "bbox": [ + 48, + 159, + 288, + 179 + ], + "type": "text", + "content": "[34] Pulak Purkait, Tat-Jun Chin, and Ian Reid. *Neurora: Neural robust rotation averaging*, 2020. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 180, + 288, + 212 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 180, + 288, + 212 + ], + "spans": [ + { + "bbox": [ + 48, + 180, + 288, + 212 + ], + "type": "text", + "content": "[35] Ian C Ross and Frank Harary. On the determination of redundancies in sociometric chains. Psychometrika, 17(2):195-208, 1952. 4, 12" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 212, + 288, + 255 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 212, + 288, + 255 + ], + "spans": [ + { + "bbox": [ + 48, + 212, + 288, + 255 + ], + "type": "text", + "content": "[36] Tianwei Shen, Siyu Zhu, Tian Fang, Runze Zhang, and Long Quan. Graph-based consistent matching for structure-from-motion. In European Conference on Computer Vision, pages 139-155. Springer, 2016. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 255, + 288, + 286 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 255, + 288, + 286 + ], + "spans": [ + { + "bbox": [ + 48, + 255, + 288, + 286 + ], + "type": "text", + "content": "[37] Jianbo Shi and J. Malik. Normalized cuts and image segmentation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 22(8):888-905, 2000. 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 287, + 288, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 287, + 288, + 341 + ], + "spans": [ + { + "bbox": [ + 48, + 287, + 288, + 341 + ], + "type": "text", + "content": "[38] Yunpeng Shi and Gilad Lerman. Estimation of camera locations in highly corrupted scenarios: All about that base, no shape trouble. In 2018 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2018, Salt Lake City, UT, USA, June 18-22, 2018, pages 2868-2876, 2018. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 341, + 288, + 384 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 341, + 288, + 384 + ], + "spans": [ + { + "bbox": [ + 48, + 341, + 288, + 384 + ], + "type": "text", + "content": "[39] Yunpeng Shi and Gilad Lerman. Message passing least squares framework and its application to rotation synchronization. In Proceedings of the 37th International Conference on Machine Learning (ICML), 2020. 1, 2, 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 384, + 288, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 384, + 288, + 426 + ], + "spans": [ + { + "bbox": [ + 48, + 384, + 288, + 426 + ], + "type": "text", + "content": "[40] Yunpeng Shi, Shaohan Li, and Gilad Lerman. Robust multi-object matching via iterative reweighting of the graph connection laplacian. Advances in Neural Information Processing Systems, 2020-December, 2020. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 426, + 288, + 469 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 426, + 288, + 469 + ], + "spans": [ + { + "bbox": [ + 48, + 426, + 288, + 469 + ], + "type": "text", + "content": "[41] Chitturi Sidhartha and Venu Madhav Govindu. It is all in the weights: Robust rotation averaging revisited. In International Conference on 3D Vision, 3DV 2021, London, United Kingdom, December 1-3, 2021, pages 1134-1143. IEEE, 2021. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 469, + 288, + 501 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 469, + 288, + 501 + ], + "spans": [ + { + "bbox": [ + 48, + 469, + 288, + 501 + ], + "type": "text", + "content": "[42] Amit Singer. Angular synchronization by eigenvectors and semidefinite programming. Applied and computational harmonic analysis, 30(1):20-36, 2011. 2, 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 501, + 288, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 501, + 288, + 533 + ], + "spans": [ + { + "bbox": [ + 48, + 501, + 288, + 533 + ], + "type": "text", + "content": "[43] A. Singer and Y. Shkolnisky. Angular synchronization by eigenvectors and semidefinite programming: Analysis and application to class averaging in cryo-electron microscopy, 2009. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 533, + 288, + 564 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 533, + 288, + 564 + ], + "spans": [ + { + "bbox": [ + 48, + 533, + 288, + 564 + ], + "type": "text", + "content": "[44] Johan Thunberg, Florian Bernard, and Jorge Goncalves. Distributed methods for synchronization of orthogonal matrices over graphs. Automatica, 80:243-252, jun 2017. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 565, + 288, + 608 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 565, + 288, + 608 + ], + "spans": [ + { + "bbox": [ + 48, + 565, + 288, + 608 + ], + "type": "text", + "content": "[45] Roberto Tron and René Vidal. Distributed image-based 3-d localization of camera sensor networks. In Proceedings of the 48th IEEE Conference on Decision and Control, CDC 2009, December 16-18, 2009, Shanghai, China, pages 901-908, 2009. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 608, + 288, + 651 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 608, + 288, + 651 + ], + "spans": [ + { + "bbox": [ + 48, + 608, + 288, + 651 + ], + "type": "text", + "content": "[46] Anton Voropaev and Sergey Perepechko. The number of fixed length cycles in undirected graph. explicit formulae in case of small lengths. Bulletin of PFUR. Series Mathematics, Information Sciences, Physics, pages 5-11, 01 2012. 4" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 651, + 288, + 693 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 651, + 288, + 693 + ], + "spans": [ + { + "bbox": [ + 48, + 651, + 288, + 693 + ], + "type": "text", + "content": "[47] Anton Voropaev and Sergey Perepechko. The number of fixed length cycles in undirected graph. explicit formulae in case of small lengths. Bulletin of PFUR. Series Mathematics, Information Sciences, Physics, pages 5-11, 01 2012. 13" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 48, + 693, + 288, + 704 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 693, + 288, + 704 + ], + "spans": [ + { + "bbox": [ + 48, + 693, + 288, + 704 + ], + "type": "text", + "content": "[48] Van H Vu. A large deviation result on the number of small" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 224 + ], + "type": "list", + "angle": 0, + "index": 22, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "text", + "content": "subgraphs of a random graph. Combinatorics, Probability and Computing, 10(1):79-94, 2001. 5, 14, 17" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 95, + 547, + 125 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 95, + 547, + 125 + ], + "spans": [ + { + "bbox": [ + 307, + 95, + 547, + 125 + ], + "type": "text", + "content": "[49] Lanhui Wang and Amit Singer. Exact and stable recovery of rotations for robust synchronization. Information and Inference, 2013. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 126, + 547, + 169 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 126, + 547, + 169 + ], + "spans": [ + { + "bbox": [ + 307, + 126, + 547, + 169 + ], + "type": "text", + "content": "[50] Kyle Wilson and Noah Snavely. Robust global translations with 1dsfm. In Computer Vision - ECCV 2014 - 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part III, pages 61-75, 2014. 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 169, + 547, + 224 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 169, + 547, + 224 + ], + "spans": [ + { + "bbox": [ + 307, + 169, + 547, + 224 + ], + "type": "text", + "content": "[51] Christopher Zach, Manfred Klopschitz, and Marc Pollefeys. Disambiguating visual relations using loop constraints. In The Twenty-Third IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2010, San Francisco, CA, USA, 13-18 June 2010, pages 1426-1433, 2010. 2" + } + ] + } + ], + "index": 21 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "5269" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Efficient Hyperparameter Optimization with Adaptive Fidelity Identification/fa155004-89c9-4599-9ab7-9a5d6edc5fbc_content_list.json b/2024/Efficient Hyperparameter Optimization with Adaptive Fidelity Identification/fa155004-89c9-4599-9ab7-9a5d6edc5fbc_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..898a696a7d053b4add6cdc4c23e8f3be0d422a0f --- /dev/null +++ b/2024/Efficient Hyperparameter Optimization with Adaptive Fidelity Identification/fa155004-89c9-4599-9ab7-9a5d6edc5fbc_content_list.json @@ -0,0 +1,1325 @@ +[ + { + "type": "text", + "text": "Efficient Hyperparameter Optimization with Adaptive Fidelity Identification", + "text_level": 1, + "bbox": [ + 98, + 130, + 870, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Jiantong Jiang $^{1}$ , Zeyi Wen $^{2,3*}$ , Atif Mansoor $^{1}$ , Ajmal Mian $^{1}$ $^{1}$ The University of Western Australia, $^{2}$ HKUST (Guangzhou), $^{3}$ HKUST", + "bbox": [ + 199, + 179, + 769, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "jiantong.jiang@research.uwa.edu.au, wenzeyi@ust.hk, {atif.mansoor, ajmal.mian}@uwa.edu.au", + "bbox": [ + 89, + 219, + 874, + 234 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 268, + 313, + 284 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Hyperparameter Optimization and Neural Architecture Search are powerful in attaining state-of-the-art machine learning models, with Bayesian Optimization (BO) standing out as a mainstream method. Extending BO into the multi-fidelity setting has been an emerging research topic in this field, but faces the challenge of determining an appropriate fidelity for each hyperparameter configuration to fit the surrogate model. To tackle the challenge, we propose a multi-fidelity BO method named FastBO, which excels in adaptively deciding the fidelity for each configuration and providing strong performance while ensuring efficient resource usage. These advantages are achieved through our proposed techniques based on the concepts of efficient point and saturation point for each configuration, which can be obtained from the empirical learning curve of the configuration, estimated from early observations. Extensive experiments demonstrate FastBO's superior anytime performance and efficiency in identifying high-quality configurations and architectures. We also show that our method provides a way to extend any single-fidelity method to the multi-fidelity setting, highlighting the wide applicability of our approach.", + "bbox": [ + 75, + 300, + 473, + 619 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 654, + 209, + 671 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Hyperparameters are crucial in machine learning pipelines. Hyperparameter optimization (HPO) [11] and Neural Architecture Search (NAS) [9] aims to find the hyperparameters or architectures that can yield good performance without human experts. Among different HPO and NAS methods, Bayesian Optimization (BO) [2, 14, 40] is an effective model-based method that has shown remarkable success [8, 39]. BO maintains a surrogate model of the target performance metric based on past evaluations of hyperparameter configurations, which guides the choice of more promising configurations to evaluate.", + "bbox": [ + 75, + 680, + 468, + 845 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Despite its sample efficiency, standard BO requires a full evaluation of each configuration, involving full-scale", + "bbox": [ + 76, + 847, + 468, + 878 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "training and testing of models, which can be highly time-consuming, particularly with the recent trend to larger models. To avoid expensive full evaluations, multi-fidelity methods [4, 16, 25, 26] have been proposed, where the fidelity refers to the levels of performance metrics obtained under different resource levels. These methods follow the principle of successive halving (SHA) [16]: initially, they evaluate a set of randomly selected configurations using a small number of resources; then, based on the low-fidelity performances, the poorly-performing ones are successively eliminated, while the well-performing ones continue to be evaluated with increasing resources. Follow-up studies [10, 22, 27, 37, 47] propose model-based multi-fidelity methods, replacing the random configuration selection with a more informed model to improve sample efficiency.", + "bbox": [ + 496, + 268, + 893, + 496 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Nevertheless, current model-based multi-fidelity methods face a major limitation: they are built upon the SHA framework, which operates under the assumption that learning curves of different configurations rarely intersect. This assumption does not hold in practice [46], i.e., early performance observations cannot always indicate the final fidelity performance at the full resource level. This leads to a fundamental challenge when extending model-based methods to the multi-fidelity setting: What is the appropriate fidelity for each configuration to fit the surrogate model? In other words, which fidelity can provide performance observations that reliably indicate the final fidelity performance? Existing methods struggle to address this fundamental challenge. In particular, BOHB [10] and Hyper-Tune [27] fit separate surrogate models for different fidelities, failing to capture inter-fidelity correlations. FTBO [44] and A-BOHB [22] fit a joint model but require strong assumptions to remain tractable. Another work by Salinas et al. [37] suggests using the last observed fidelity performance to fit the surrogate model. However, it widens the gap between poorly- and well-performing configurations at the early stage, potentially leading to an inaccurate surrogate model.", + "bbox": [ + 496, + 502, + 893, + 835 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To this end, we propose a multi-fidelity extension of BO, namely FastBO, which tackles the challenge of deciding the appropriate fidelity for each configuration to fit the surrogate model. FastBO identifies a so-called efficient point for", + "bbox": [ + 496, + 839, + 893, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Zeyi Wen is the corresponding author.", + "bbox": [ + 94, + 886, + 303, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "26181", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "each configuration to be the fidelity. The point balances computational cost and performance quality while capturing valuable learning curve trends. In essence, FastBO selects the fidelity for each configuration instead of evaluating all the configurations at the same fidelity. Additionally, a saturation point for each configuration is identified to be an approximation of the final fidelity, leading to high-quality performance while reducing resource wastage. The two crucial points are adaptively derived from the estimated learning curve of each configuration. Furthermore, the warm-up and post-processing stages are carefully designed to enable judicious early-termination detection and efficient saturation level evaluation. Empirical evaluation against the state-of-the-art methods shows that FastBO has strong anytime performance and can considerably save up to $87\\%$ of the time required to identify a good configuration or architecture, lowering the barriers for engaging in HPO and NAS. In summary, we make the following major contributions.", + "bbox": [ + 75, + 90, + 472, + 363 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. We propose a multi-fidelity model-based HPO method that adaptively decides the fidelities for configurations and efficiently offers strong performance, thanks to the introduced concepts of efficient and saturation points.", + "2. We develop a learning curve modeling module to enable adaptive derivation of the key points, a warm-up stage to allow early-termination detection, and a post-processing stage to ensure efficient saturation-level evaluation.", + "3. We show that our strategy can be used to extend existing single-fidelity methods to the multi-fidelity setting, demonstrating the effectiveness and generality of our method and highlighting promising future opportunities." + ], + "bbox": [ + 76, + 367, + 468, + 547 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 76, + 558, + 218, + 574 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "HPO and NAS incur significant costs, especially considering the escalating model evaluation overhead. Despite efforts to directly accelerate computation [17-19, 52], smarter HPO strategies are still required for the widespread adoption of automation. Two crucial directions include model-based and multi-fidelity methods, which can also be combined. Here, we review the methods in these categories.", + "bbox": [ + 75, + 583, + 468, + 688 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Model-based methods. BO is the representative of model-based methods. Based on the surrogate model constructed by historical evaluation results, BO selects the configurations to evaluate via an acquisition function that balances exploration and exploitation. Commonly used surrogate models are Gaussian processes [40], random forests [14], tree-structured Parzen estimator [2], and deep networks [41, 42]. Popular acquisition functions include Expected Improvement [28], Knowledge Gradient [12], Upper Confidence Bound [43], and Predictive Entropy Search [13]. Recent studies on BO have explored the utilization of expert priors [15, 24, 33, 38] and derivative information [1, 34, 49]. There also has been a recent focus on enhancing the interpretability [5, 50, 51] of the HPO process [3, 31, 32].", + "bbox": [ + 75, + 688, + 470, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Multi-fidelity methods. Multi-fidelity methods exploit low and high fidelities for configurations to save the evaluation time. Successive halving (SHA) [16] runs a set of configurations using a small number of resources and promotes only the best-performing half of configurations to continue for twice as many resources. Hyperband [25] calls SHA as a sub-routine with varying maximum resources and introduces a reduction factor to control the fraction of promotion. ASHA [26] extends SHA to the asynchronous setting by aggressive early-stopping. Later, PASHA [4] further extends ASHA through more aggressive early-stopping based on the ranking of configurations during the tuning process.", + "bbox": [ + 496, + 90, + 890, + 272 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Combination of model-based and multi-fidelity methods. BOHB [10] and a parallel work [47] first propose to combine model-based and multi-fidelity methods by replacing the random sampling in Hyperband with BO. A BOHB [22] employs a joint GP surrogate over fidelities and supports asynchronous scheduling. Hyper-Tune [27] improves its Hyperband by a delayed strategy to decrease inaccurate promotions. Salinas et al. [37] proposed to extend methods to multi-fidelity settings by using the performance of the last fidelity in an ASHA running. DyHPO [48] and DPL [20] introduce new surrogates for multi-fidelity BO considering the learning curves; the former uses deep GP kernels while the latter integrates deep power law functions.", + "bbox": [ + 496, + 272, + 890, + 469 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3. Problem Formulation", + "text_level": 1, + "bbox": [ + 498, + 482, + 705, + 498 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Given an algorithm having hyperparameters $\\lambda_1, \\dots, \\lambda_m$ with domains $\\Lambda_1, \\dots, \\Lambda_m$ , we define its hyperparameter space as $\\Lambda = \\Lambda_1 \\times \\dots \\times \\Lambda_m$ . Here, we define the problem and outline the key challenge related to hyperparameter optimization (HPO). Notations are in Supp. 7 for reference.", + "bbox": [ + 496, + 507, + 890, + 583 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Single-fidelity setting. For each hyperparameter configuration $\\lambda$ , we denote $f(\\lambda)$ as the performance achieved using $\\lambda$ . For consistency, the metric in this paper refers to descending metrics like validation loss, with ascending metrics being treated similarly. In the single-fidelity HPO setting, we aim to find $\\lambda^*$ minimizing function $f(\\lambda)$ , i.e., $\\lambda^* = \\arg \\min_{\\lambda \\in \\Lambda} f(\\lambda)$ . BO is one of the most popular single-fidelity HPO methods. The vanilla BO has two key components: a surrogate model $\\mathcal{M}$ to approximate the objective function $f(\\lambda)$ , and an acquisition function $a$ to identify a promising configuration from search space. With these ingredients, BO iterates three steps: (i) select a configuration $\\lambda_i$ by maximizing the acquisition function; (ii) evaluate $\\lambda_i$ to get $y_i$ and add the data $(\\lambda_i, y_i)$ into the current observation set $\\mathcal{D}_{i-1} = \\{(\\lambda_1, y_1), \\dots, (\\lambda_{i-1}, y_{i-1})\\}$ ; (iii) update the surrogate model and the acquisition function based on the augmented $\\mathcal{D}_i$ . In this work, $\\mathcal{M}$ is a Gaussian Process and $a$ is Expected Improvement.", + "bbox": [ + 496, + 583, + 890, + 853 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Multi-fidelity setting. Multi-fidelity methods consider resource information, such as training epochs or training subset ratios. Evaluations at various resource levels results in", + "bbox": [ + 496, + 854, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "26182", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "different performance levels, known as the fidelities. Different fidelities provide a way to balance computational cost and performance quality. In multi-fidelity HPO problems, the target is extended to $\\lambda^{*} = \\arg \\min_{\\lambda \\in \\Lambda}f(\\lambda ,r)$ , where $f(\\lambda ,r)$ is the objective function obtained for $\\lambda$ at $r$ . We use $r$ to denote the resource level, which can also be interpreted as the fidelity, and $r\\in \\{r_{min},\\dots,r_{max}\\}$ .", + "bbox": [ + 75, + 90, + 468, + 196 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Extending single-fidelity methods to the multi-fidelity setting. The inefficiency of single-fidelity methods stems from their reliance on the final fidelity evaluation of $f(\\lambda, r_{max})$ to be the evaluation of its objective $f(\\lambda, r)$ . Fitting surrogate models by such final fidelity evaluations incurs high cost due to the full evaluation of the configurations. Notably, low-fidelity evaluations at $r < r_{max}$ provide informative insights into the objective but are computationally cheaper, which is valuable to the optimization process. Therefore, we seek an effective way to extend single-fidelity methods like BO to the multi-fidelity setting. More specifically, recalling the earlier steps of BO, when evaluating the configuration $\\lambda_i$ in the second step, we instead acquire its low-fidelity performance $y_i^{r_i}$ at $r_i$ , where $r_i$ denotes the fidelity used for $\\lambda_i$ to fit the surrogate model. The observations $\\mathcal{D}_i$ then become $\\{(\\lambda_1, y_1^{r_1}), \\dots, (\\lambda_i, y_i^{r_i})\\}$ . To conclude, in order to extend single-fidelity methods to the multi-fidelity setting, the key challenge to be addressed is to determine $r_i$ for each $\\lambda_i$ .", + "bbox": [ + 75, + 198, + 472, + 484 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4. Methodology", + "text_level": 1, + "bbox": [ + 76, + 500, + 212, + 517 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we propose a novel multi-fidelity model-based algorithm FastBO. We first propose the key concepts of efficient point and saturation point, which are crucial in deciding the fidelity level to fit the surrogate model and to approximate the final fidelity respectively. Secondly, we elaborate on the details of learning curve modeling, where the two crucial points can be extracted. Then, we present the techniques associated with the auxiliary warm-up and post-processing stages. Finally, we summarize FastBO and discuss its wide applicability to any single-fidelity methods.", + "bbox": [ + 75, + 525, + 468, + 676 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4.1. Estimation of Efficient and Saturation Points", + "text_level": 1, + "bbox": [ + 76, + 686, + 457, + 702 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In our method, we adaptively identify efficient and saturation points for each configuration. The two points are crucial in the optimization process. We first formally define the efficient point as follows.", + "bbox": [ + 75, + 710, + 467, + 771 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Definition 1 (Efficient point). For a given learning curve $\\mathcal{C}_i(r)$ of hyperparameter configuration $\\lambda_{i}$ , where $r$ represents the resource level (also referred to as fidelity), the efficient point $e_i$ of $\\lambda_{i}$ is defined as: $e_i = \\min \\{r\\mid \\mathcal{C}_i(r) - \\mathcal{C}_i(2r) < \\delta_1\\}$ , where $\\delta_{1}$ is a predefined small threshold.", + "bbox": [ + 75, + 782, + 467, + 859 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The semantic of Definition 1 is that starting from the efficient point onwards, when the resources are doubled (i.e.,", + "bbox": [ + 76, + 869, + 467, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "from $r$ to $2r$ ), the performance improvement falls below a small threshold $\\delta_1$ . Consequently, this point characterizes the fidelity at which a configuration demonstrates strong performance while still efficiently utilizing resources. In simpler terms, it signifies an appropriate fidelity of performance that can be achieved with comparably efficient resource usage. Therefore, we make the following remark.", + "bbox": [ + 496, + 90, + 890, + 196 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Remark 1. The efficient points of the hyperparameter configurations can serve as their appropriate fidelities used for fitting the surrogate model. This is due to their (i) optimal resource-to-performance balance, (ii) ability to capture valuable learning curve trends, and (iii) customization for different hyperparameter configurations.", + "bbox": [ + 496, + 207, + 893, + 297 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We elaborate on the reasons in Remark 1 as follows. Firstly, efficient points balance the trade-off between computational cost and result quality. Beyond the efficient point of a given configuration, allocating additional resources to that configuration becomes less efficient. Secondly, efficient points capture valuable trends within the learning curves. For example, the learning rate influences the shape of learning curves; the identification of efficient points for configurations with smaller learning rates often occurs at later stages. The insights into learning curve behaviors enable more informed decision-making. Thirdly, the ability to customize the fidelity for each specific configuration is a significant advantage. This adaptive approach is more reasonable than previous studies that use a fixed fidelity for all configurations, as it better accounts for the unique characteristics of individual learning curves.", + "bbox": [ + 496, + 306, + 892, + 547 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "This insight leads us to use the efficient point $e_i$ identified for each configuration $\\lambda_i$ as its fidelity used to fit the surrogate model. Specifically, we evaluate $\\lambda_i$ until reaching $e_i$ and obtain the observed performance $y_i^{e_i}$ . The resulting data point $(\\lambda_i, y_i^{e_i})$ is then added into the current observation set $\\mathcal{D}_{i-1}$ to refit the surrogate model. We proof the superiority of FastBO over SHA-based methods in Supp. 8.", + "bbox": [ + 496, + 549, + 890, + 654 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Besides efficient points, we identify saturation points for all configurations from their learning curves as well. We provide the definition of the saturation point as follows.", + "bbox": [ + 496, + 655, + 890, + 699 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Definition 2 (Saturation point). For a given learning curve $\\mathcal{C}_i(r)$ of configuration $\\lambda_i$ , where $r$ represents the resource level (also referred to as fidelity), the saturation point $s_i$ of $\\lambda_i$ is defined as: $s_i = \\min \\{r \\mid \\forall r' > r, |\\mathcal{C}_i(r') - \\mathcal{C}_i(r)| < \\delta_2\\}$ , where $\\delta_2$ is a predefined small threshold.", + "bbox": [ + 496, + 709, + 893, + 785 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The semantic of Definition 2 is that beyond the saturation point, the observed performance no longer exhibits notable variations with more resources. Thus, this point characterizes the fidelity at which the performance of a configuration stabilizes. The concept of saturation point is well-recognized within the machine learning community. Building on the above definition, we make the following remark.", + "bbox": [ + 496, + 795, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "26183", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Remark 2. The saturation points of the hyperparameter configurations can serve as their approximate final fidelities, as they provide performance results that meet predefined quality thresholds while reducing resource wastage.", + "bbox": [ + 73, + 90, + 468, + 154 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "This insight leads us to use the saturation point $s_i$ identified for each configuration $\\lambda_i$ as its final fidelity approximation. The point is used in the post-processing stage for promoting some well-performing configurations to get higher-fidelity performances. In essence, when aiming for a full evaluation of the configurations, we suggest that terminating the evaluation at the saturation point is sufficient. A more intuitive illustration of the concepts of efficient and saturation points is provided in Supp. 9.", + "bbox": [ + 75, + 159, + 470, + 297 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.2. Learning Curve Modeling", + "text_level": 1, + "bbox": [ + 75, + 303, + 313, + 320 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "From Definitions 1 and 2, we can extract the efficient and saturation points of configurations from their learning curves. The curve $\\mathcal{C}_i(r)$ corresponds to configuration $\\lambda_{i}$ and describes the predictive performance with $\\lambda_{i}$ as a function of the fidelity $r$ . Here, $r$ can be either the number of training instances or the number of training epochs. In the context of learning curves, the former is referred to as observation learning curves, while the latter is iteration learning curves [29]. Both types are applicable to FastBO, so we use the term learning curve to encompass both. Given the observation set $\\mathcal{O}_i^w = \\{(r,y_i^r)\\}_{r = r_{min},\\dots,w}$ for $\\lambda_{i}$ , which comprises pairs of data points representing fidelities $r\\in \\{r_{min},\\dots,w\\}$ and the corresponding evaluations $y_{i}^{r}$ , where $w$ is a pre-defined warm-up point to stop collecting data, FastBO can estimate a learning curve for $\\lambda_{i}$ based on $\\mathcal{O}_i^w$ by first constructing a parametric learning curve model, then estimating the parameters.", + "bbox": [ + 73, + 327, + 468, + 583 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Constructing a parametric learning curve model. Empirical learning curves can be modeled with function classes relying on some parameters. Viering and Loog [46] comprehensively summarized the parametric models studied in machine learning. In practice, different problems have different learning curves; even under the same problem, different hyperparameter configurations (e.g., learning rate, regularization, etc.) may lead to significantly different learning curves. Since one single parametric model is not enough to characterize all the learning curves by itself, we consider combining different parametric models into a single model. Specifically, we consider three parametric models POW3, EXP3 and LOG2, as listed in Tab. 1, which have shown", + "bbox": [ + 73, + 584, + 470, + 780 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/75d20c7159029cc3ddc01d6d90eb446c36175672fa38b1750ef219dae8207d7c.jpg", + "table_caption": [ + "Table 1. Parametric learning curve models used." + ], + "table_footnote": [], + "table_body": "
ModelFormulaFamily
POW3y = d + ax-αPower law
EXP3y = d + e-ax+bExponential
LOG2y = d + a log(x)Logarithmic
", + "bbox": [ + 125, + 816, + 421, + 896 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "good fitting and predicting performance in previous empirical studies [29, 46]. We provide detailed discussions on the choice of parametric models in Supp. 10.", + "bbox": [ + 496, + 90, + 890, + 136 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Here, we denote each parametric model as $c_{j}(r|\\theta_{j})$ with parameters $\\theta_{j}$ , where the independent variable $r$ represents the fidelity. We combine three models into one model via a weighted linear combination:", + "bbox": [ + 496, + 136, + 890, + 196 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {C} (r | \\phi) = \\sum_ {j \\in \\{1, 2, 3 \\}} \\omega_ {j} c _ {j} (r | \\boldsymbol {\\theta} _ {j}), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 576, + 209, + 890, + 234 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\phi = \\{\\omega_1,\\omega_2,\\omega_3,\\theta_1,\\theta_2,\\theta_3\\}$ is the parameter of the combined model, which consists of parameters $\\{\\pmb {\\theta}_1,\\pmb {\\theta}_2,\\pmb {\\theta}_3\\}$ and weight $\\{\\omega_{1},\\omega_{2},\\omega_{3}\\}$ of every single model. Therefore, each pair of observations $(r,y_i^r)$ in $O_{i}^{w}$ can be modeled by the combined model as $y_{i}^{r} = \\mathcal{C}(r|\\phi) + \\epsilon$ , where $y_{i}^{r}$ is the observed dependent variable and $\\epsilon$ represents the error term.", + "bbox": [ + 496, + 246, + 890, + 335 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Estimating parameters in the parametric learning curve model. We employ maximum likelihood estimation to estimate the parameters $\\phi$ in the parametric model $\\mathcal{C}(r|\\phi)$ . Assuming that $\\epsilon \\sim \\mathcal{N}(0,\\sigma^2)$ , the probability of an observed performance $y_{i}^{r}$ under parameters is given by $p(y_i^r |\\phi ,\\sigma^2) = \\mathcal{N}(y_i^r;\\mathcal{C}(r|\\phi),\\sigma^2)$ . Given the observations $\\mathcal{O}_i^w$ of $\\lambda_{i}$ that contains a set of observed data points $(r,y_i^r)$ , the likelihood function can be expressed as:", + "bbox": [ + 496, + 338, + 890, + 458 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} (\\boldsymbol {\\phi}, \\sigma^ {2}; \\mathbf {r}, \\mathbf {y} _ {i} ^ {r}) = \\prod p (y _ {i} ^ {k} | \\boldsymbol {\\phi}, \\sigma^ {2}) \\\\ = \\prod_ {k = r _ {\\text {m i n}}} ^ {w} \\frac {1}{\\sigma \\sqrt {2 \\pi}} \\exp \\left(- \\frac {\\left(y _ {i} ^ {r} - \\mathcal {C} (r = k | \\phi)\\right) ^ {2}}{2 \\sigma^ {2}}\\right). \\tag {2} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 524, + 470, + 890, + 536 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We estimate $\\phi$ by maximizing log-likelihood function, which is easily calculated given Eq. 2.", + "bbox": [ + 496, + 547, + 888, + 577 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "An existing model-free method [7] also considers using learning curves for the HPO problem. However, it targets predicting the high-fidelity performance from the low-fidelity observations and thus stopping configurations that are unlikely to beat the current best values, which is different from our main target of identifying appropriate fidelity levels for the configurations to fit the surrogate model from their estimated learning curves.", + "bbox": [ + 496, + 579, + 890, + 700 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.3. Warm-up And Post-processing Stages", + "text_level": 1, + "bbox": [ + 498, + 710, + 823, + 727 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In addition to its core components, FastBO incorporates two auxiliary stages: the warm-up and post-processing stages. For the completeness of our method, we provide an overview of these stages, outlining their targets and presenting the key techniques of early-termination detection and saturation-level evaluation that are applied within.", + "bbox": [ + 496, + 734, + 890, + 824 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Warm-up stage. The warm-up stage prepares the early observation set $\\mathcal{O}_i^w$ for each configuration $\\lambda_{i}$ that is used to estimate its learning curve, as discussed in $\\S 4.2$ . Here $w\\in (r_{min},r_{max})$ is a pre-determined fidelity, denoted as warm-up point. Specifically, we initiate the evaluation of", + "bbox": [ + 496, + 825, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "26184", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "each newly selected $\\lambda_{i}$ , proceeding until reaching $w$ . During this process, we record each fidelity $r$ and its evaluation result $y_{i}^{r}$ , forming pairs $(r,y_i^r)$ . Upon reaching $w$ , we pause the evaluation for $\\lambda_{i}$ and obtain its early observation set $\\mathcal{O}_i^w = \\{(r,y_i^r)\\}_{r = r_{min},\\ldots ,w}$ , and start modeling the learning curve. During the warm-up stage, we monitor the performance changes across every two continuous fidelities. If we detect that the performance of $\\lambda_{i}$ has consecutively dropped twice by more than a ratio $\\alpha$ , i.e., $(y_{i}^{r - 1} - y_{i}^{r - 2}) > \\alpha y_{i}^{r - 2}$ and $(y_i^r -y_i^{r - 1}) > \\alpha y_i^{r - 1}$ , we promptly terminate the evaluation for $\\lambda_{i}$ at its current fidelity $r$ , because such consecutive performance deterioration indicates $\\lambda_{i}$ is unlikely to achieve satisfactory performance. Once terminated, we directly incorporate the current performance $y_{i}^{r}$ of $\\lambda_{i}$ into $\\mathcal{D}_{i - 1}$ that is used for updating the surrogate model. Thus, further operations like learning curve modeling are discontinued for $\\lambda_{i}$ . Moreover, if we observe a single case of performance drop without subsequent occurrences, i.e., $y_{i}^{r - 1} - y_{i}^{r - 2} > \\alpha y_{i}^{r - 2}$ and $y_{i}^{r} - y_{i}^{r - 1}\\leq \\alpha y_{i}^{r - 1}$ , we opt not to include data from fidelity $r - 1$ in $\\mathcal{O}_i^w$ . This is to manually filter out potential noise in the data that may adversely affect the fitting of the learning curve.", + "bbox": [ + 76, + 90, + 472, + 422 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Post-processing stage. The post-processing stage aims at two tasks: promoting the well-performing configurations for saturation-level evaluations and identifying the best configuration and its performance. Firstly, FastBO promotes the top- $k$ well-performing configurations and evaluates them to their saturation points to ensure high-quality performance while maintaining efficient resource utilization. We set $k$ to be always less than or equal to the number of parallel workers available, ensuring a manageable overhead of saturation-level evaluations. It is worth noting that the additional time required is factored into the overall time. Secondly, FastBO finds the best configuration along with its performance achieved so far, which is a standard final step in most HPO methods. However, an increase in fidelities does not always result in performance improvement, possibly due to overfitting, resource saturation, or problem complexity. Therefore, we treat the evaluation at each fidelity as an individual task and record all these intermediate evaluation results, which is also a common practice in recent implementations. In this way, FastBO finds the best performance by considering all the results, rather than relying solely on the highest-fidelity performances of the configurations. In the parallel setting, treating each fidelity evaluation as an individual task offers an added benefit due to its finer granularity. More specifically, when a worker is idle, it takes on a new task of evaluating a configuration at a specific fidelity, rather than evaluating an entire configuration.", + "bbox": [ + 76, + 425, + 472, + 834 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.4. FastBO and Generalization", + "text_level": 1, + "bbox": [ + 76, + 845, + 323, + 861 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Algorithm 1 summarizes our proposed FastBO. It takes surrogate model $\\mathcal{M}$ , acquisition function $a$ , warm-up point $w$ ,", + "bbox": [ + 76, + 869, + 468, + 902 + ], + "page_idx": 4 + }, + { + "type": "code", + "sub_type": "algorithm", + "code_caption": [ + "Algorithm 1: FastBO algorithm" + ], + "code_body": "input : $\\mathcal{M},a,w,\\alpha ,k,\\delta_1,\\delta_2$ output: $\\lambda^{*},y^{*}$ \n1 $i\\gets 0,D\\gets \\emptyset$ \n2 while not meet the stop criterion do \n3 find $\\lambda_{i}\\gets \\arg \\max_{\\lambda \\in \\Lambda}a(\\lambda ,\\mathcal{M}_{i - 1})$ \n4 $\\mathcal{O}_i^w,t\\gets$ warm-up given $w,\\alpha$ // cf. $\\S 4.3$ \n5 if $\\mathcal{O}_i^w$ is not empty then \n6 fit $\\mathcal{C}_i(r)$ to $\\mathcal{O}_i^w$ // cf. $\\S 4.2$ \n7 find $e_i,s_i$ given $\\mathcal{C}_i(r),\\delta_1,\\delta_2$ // cf. $\\S 4.1$ \n8 $y_{i}^{e_{i}}\\gets$ continue evaluating $\\lambda_{i}$ to $e_i$ \n9 else \n10 $e_i\\gets t,s_i\\gets r_{max}$ \n11 $\\mathcal{D}_i\\gets \\mathcal{D}_{i - 1}\\cup (\\lambda_i,y_i^{e_i})$ \n12 refit $\\mathcal{M}_i$ to $\\mathcal{D}_i$ \n13 $i\\gets i + 1$ \n14 $\\lambda^{*},y^{*}\\gets$ post-process given s = $\\{s_i\\} ,k / / c f.\\S 4.3$", + "bbox": [ + 500, + 111, + 890, + 373 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "performance decrease ratio $\\alpha$ , promotion number $k$ , and thresholds $\\delta_{1}$ , $\\delta_{2}$ as inputs, and output the best-founded configuration $\\lambda^{*}$ and its performance $y^{*}$ . FastBO follows a similar iterative process of model-based methods but replaces the expensive full evaluations with a more intelligent alternative (cf. Lines 4-10). Specifically, each configuration $\\lambda_{i}$ first enters a warm-up stage to collect its early observation set $\\mathcal{O}_i^w$ and to be detected and terminated if it exhibits consecutive performance deterioration (cf. Line 4). If $\\lambda_{i}$ is not terminated, FastBO then estimates a learning curve $\\mathcal{C}_i(r)$ for $\\lambda_{i}$ based on $\\mathcal{O}_i^w$ (cf. Line 6), and thus the efficient point and saturation point of $\\lambda_{i}$ can be obtained (cf. Line 7). After that, $\\lambda_{i}$ continues to be evaluated until reaching $e_i$ (cf. Line 8); the result is added to the observation set $\\mathcal{D}$ (cf. Line 11) that is used for updating $\\mathcal{M}$ (cf. Line 12). On the other hand, the poorly-performing configuration will be terminated early at fidelity $t$ with its result being added directly to $\\mathcal{D}$ (cf. Lines 10, 11). Finally, the post-processing stage promotes the most promising configurations to their saturation points and finds the best-founded configuration $\\lambda^{*}$ and its performance $y^{*}$ (cf. Line 14).", + "bbox": [ + 496, + 412, + 893, + 731 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Generalizing FastBO to single-fidelity methods. The core of FastBO is to tackle the key challenge of deciding an appropriate fidelity for each configuration to fit the surrogate model by adaptively identifying its efficient point. This strategy of using the efficient point performances for surrogate model fitting also provides a simple but effective way to bridge the gap between single- and multi-fidelity methods. While it is primarily described in the context of model-based methods, the strategy can be generalized to various single-fidelity methods. For example, when evaluating configurations within the population for an evolution-", + "bbox": [ + 496, + 734, + 893, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "26185", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/f02b00b4e9beacdb5607cfb0cc7957f340ddd17d4c6cf9731807ca0c3e39398f.jpg", + "image_caption": [ + "Figure 1. Performance of average validation accuracy on the LCBench benchmark." + ], + "image_footnote": [], + "bbox": [ + 76, + 88, + 893, + 238 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/c6bd55335731ea34e417923de43d229dc73256aba6d6635452af1a8c6aa9e4be.jpg", + "image_caption": [ + "Figure 2. Performance of (a) average validation error on NAS-Bench-201 and (b) average validation loss on FCNet." + ], + "image_footnote": [], + "bbox": [ + 76, + 273, + 890, + 435 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "ary algorithm-based HPO method, we can similarly evaluate the efficient point performances instead of the final performances of these configurations and integrate the performances in the subsequent processes, such as selection and variation. Relying on the efficient point rather than the final fidelity or all fidelities available simplifies the extension of the single-fidelity methods to the multi-fidelity setting. The rationale behind this adaptive fidelity identification strategy is discussed in Remark 1. We also demonstrate in our experiments the efficacy of this strategy in extending a range of single-fidelity methods to the multi-fidelity setting.", + "bbox": [ + 75, + 478, + 472, + 645 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 76, + 661, + 209, + 679 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We empirically evaluate the performance of FastBO and compare it with the random search baseline (RS) and 9 competitive baselines from 3 related categories, including (i) model-based methods: standard Gaussian Process-based BO [40]; (ii) multi-fidelity methods: ASHA [26], Hyperband [25], PASHA [4]; and (iii) model-based multifidelity methods: A-BOHB [22], A-CQR [37], BOHB [10], DyHPO [48], Hyper-Tune [27]. RS and BO are single-fidelity baselines, while the others are multi-fidelity ones.", + "bbox": [ + 75, + 686, + 468, + 823 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Our experiments are conducted on 10 datasets from 3 popular benchmarks LCBench [53], NAS-Bench-201 [8] and FCNet [21]. Detailed information on the benchmarks is provided in Supp. 13.1. All the experiments are evaluated with four parallel workers and 10 random seeds. We", + "bbox": [ + 75, + 824, + 470, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "allocate $20\\%$ total budget for warm-up, i.e., $w = r_{min} + 0.2 \\cdot (r_{max} - r_{min})$ . Ratio $\\alpha$ is set to 0.1; thresholds $\\delta_1$ and $\\delta_2$ are set to 0.001 and $0.0005^1$ . We set $k$ based on the number of workers $\\# w$ and the number of started configurations $\\# c$ : $k = \\max \\{\\lceil \\# c / 10 \\rceil, \\# w\\}$ . We provide more experiments and discussions on the hyperparameters in Supp. 12. We use implementations of the baselines in Syne Tune [36]. Details of the baseline settings are in Supp. 13.2.", + "bbox": [ + 496, + 478, + 893, + 602 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1. Anytime Performance", + "text_level": 1, + "bbox": [ + 498, + 611, + 705, + 628 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To evaluate the anytime performance, we compare FastBO against the baselines on wall-clock time. For fair comparisons, all the baselines, even single-fidelity BO and RS, are extended to consider intermediate results at all the fidelities when identifying the configuration, akin to FastBO as discussed in § 4.3. Consequently, all the baselines are able to achieve their best possible anytime performance.", + "bbox": [ + 496, + 633, + 893, + 739 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The results on LCBench, NAS-Bench-201, and FCNet are shown in Figs. 1 and 2. We report the validation accuracy, error, and loss over wall-clock time for the three benchmarks, as provided by the benchmarks. We provide the results on NAS-Bench-301 [39] in Supp. 11.1. Overall, FastBO can handle various performance metrics and shows strong anytime performance. We can observe that FastBO gains an advantage earlier than other methods, rapidly con", + "bbox": [ + 496, + 741, + 893, + 863 + ], + "page_idx": 5 + }, + { + "type": "page_footnote", + "text": "1Parameters $\\delta_{1}$ and $\\delta_{2}$ given here are derived after standardizing metrics to a uniform scale from 0 to 1.", + "bbox": [ + 500, + 875, + 890, + 898 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "26186", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/58fad66fa14733261cdc6443ac1f82eafe16d6c256249f8a86fa84c46478138d.jpg", + "table_caption": [ + "Table 2. Comparison of relative efficiency on configuration identification. FastBO is set as the baseline with a relative efficiency of 1.00. Wall-clock time (abbr. WC time) reports the elapsed time spent for each method on finding configurations with similar performance metrics, i.e., validation error $(\\times 10^{-2})$ for Covertype and ImageNet16-120 and validation loss $(\\times 10^{-5})$ for Slice." + ], + "table_footnote": [], + "table_body": "
Metric DatasetMethodFastBOBOPASHAA-BOHBA-CQRBOHBDyHPOHyper-Tune
CovertypeVal. error22.9±0.223.0±0.325.1±2.523.5±1.131.6±1.932.5±0.823.0±0.323.0±0.2
WC time (h)0.7±0.32.9±0.73.9±1.02.0±1.03.9±0.22.5±1.01.7±0.61.8±0.7
Rel. efficiency1.000.250.180.370.190.290.410.40
ImageNet 16-120Val. error55.3±0.257.4±1.255.7±0.355.8±1.655.5±0.955.5±1.155.5±1.055.3±2.0
WC time (h)2.2±0.76.6±0.92.5±1.25.9±1.16.0±1.33.2±0.74.3±1.03.4±1.1
Rel. efficiency1.000.340.900.380.370.680.510.67
SliceVal. loss26.3±2.626.4±4.426.8±9.526.3±6.327.1±4.226.8±5.627.4±2.328.7±1.3
WC time (h)0.4±0.13.1±0.71.2±0.92.1±0.72.5±0.72.2±0.92.5±0.51.8±0.6
Rel. efficiency1.000.130.350.200.170.190.170.24
", + "bbox": [ + 84, + 143, + 885, + 353 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "verging to the global optimum after the initial phase.", + "bbox": [ + 76, + 381, + 426, + 397 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The superiority can be attributed to two main factors. Firstly, FastBO maintains, and in some cases even surpasses, the sample efficiency of vanilla BO, thanks to our techniques that enable quick and precise identification of the fidelities for configurations to update the surrogate model. We provide more explanations and experiments on sample efficiency in Supp. 11.2. Secondly, the multi-fidelity extension speeds up the evaluations, contributing to its overall efficiency. In contrast, the single-fidelity baselines tend to waste more time on the full evaluations. While the multifidelity baselines efficiently explore numerous configurations, they limit their evaluations to only constrained fidelities for some time, thus struggling to provide relatively high performance in a short time. This issue in multi-fidelity methods is particularly pronounced in PASHA when applied to NAS-Bench-201 and FCNet, as shown in Fig. 2. In Supp. 11.3, we further provide the ranks of all methods and statistically show FastBO's superiority on an early stage. It is worth noting that all the additional overhead introduced by FastBO is taken into account in the wall-clock time.", + "bbox": [ + 75, + 398, + 470, + 700 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Regarding the final performance, most methods are able to converge to satisfactory solutions, with negligible differences among them in most cases. Although our goal is not to offer the best final performance as we limit the evaluations to at most the saturation point even for those we consider most promising, FastBO still achieves top-2 final performance on 8 out of 10 datasets. In contrast, model-free methods sometimes cannot obtain a satisfactory final performance because they randomly select the configurations. For example, on the \"Covertype\" dataset, only 3 out of 2000 configurations yield a validation accuracy exceeding $75\\%$ . As a result, all the model-free methods face challenges in converging to a satisfactory final performance.", + "bbox": [ + 75, + 704, + 470, + 902 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.2. Efficiency on Configuration Identification", + "text_level": 1, + "bbox": [ + 498, + 380, + 856, + 398 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "One explanation for PASHA's suboptimal anytime performance (cf. Fig. 2) lies in its primary goal [4]: the goal of PASHA is not high accuracy but to identify the best configuration more quickly. To ensure equitable comparisons, we report the time spent for each method on identifying a satisfactory configuration, consistent with the experiments described in PASHA [4]. Results on three expensive datasets \"Covertype\", \"ImageNet16-120\", and \"Slice\" of the three benchmarks are shown in Tab. 2. Similar results on additional datasets can be found in Supp. 11.4. Besides PASHA, results of other model-free multi-fidelity methods are not included, as PASHA demonstrates its superiority over them.", + "bbox": [ + 496, + 404, + 890, + 585 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Tab. 2 shows that FastBO saves $10\\%$ to $87\\%$ wall-clock time over other methods when achieving up to $9.6\\%$ better performance values. It can be observed from the \"rel. efficiency\" rows, where we set FastBO as the baseline with a relative efficiency of 1.00 and report the efficiency of other methods relative to ours. When compared with vanilla BO, FastBO significantly shortens the time in identifying a good configuration by a factor of 3 to 8, because FastBO pauses a configuration earlier at an appropriate fidelity and fits the surrogate model to guide the next configuration search. This advantage creates opportunities to efficiently explore more configurations. Another observation is that PASHA always gets a relatively high variance in wall-clock time. This is due to the fact that different random seeds can have a larger impact on such model-free methods.", + "bbox": [ + 496, + 585, + 892, + 813 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.3. Effectiveness of Adaptive Fidelity Identification", + "text_level": 1, + "bbox": [ + 498, + 821, + 890, + 839 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "As discussed in § 4.1, FastBO is able to adaptively identify the efficient point $e_i$ for each configuration $\\lambda_i$ and serves", + "bbox": [ + 496, + 845, + 890, + 878 + ], + "page_idx": 6 + }, + { + "type": "page_footnote", + "text": "2We convert the accuracy of \"Covertype\" into error for readability.", + "bbox": [ + 514, + 886, + 870, + 901 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "26187", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/5871895093db6b0070b54a1377b6b9d466bff987012e8951815add0452787815.jpg", + "image_caption": [ + "Figure 3. Performance comparison: (a) Performance of FastBO that adaptively sets $r_i = e_i$ with the schemes that use fixed $r_i$ for all configurations. (b) Performance of single-fidelity methods CQR, BORE, REA and their multi-fidelity variants using our extension method." + ], + "image_footnote": [], + "bbox": [ + 63, + 85, + 903, + 210 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "$e_i$ as its fidelity $r_i$ for surrogate model fitting. To investigate the effectiveness of the adaptive fidelity identification strategy, we conduct an ablation study to compare the performance achieved with and without applying this strategy. Specifically, we compare FastBO, where $r_i$ is adaptively set to $e_i$ , with the partial evaluation schemes that employ fixed predefined values as the fidelity for all the configurations to fit the surrogate model. We consider three representative fixed fidelities, including $25\\%$ , $50\\%$ , and $75\\%$ of the total resource budget. In addition, we include a comparison with vanilla BO that can be viewed as using $100\\%$ resource budget as the fixed fidelity for all configurations.", + "bbox": [ + 75, + 267, + 472, + 448 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We provide the results on three representative datasets in Fig. 3(a), with more results available in Supp. 11.5. We have three main observations. Firstly, FastBO always outperforms the partial evaluation baselines that use a fixed fidelity, indicating the effectiveness of the adaptive strategy. Secondly, FastBO shows stronger performance than vanilla BO. The limitation of vanilla BO lies in the additional time required for full evaluations. Secondly, compared to the vanilla BO, partial evaluation schemes with fixed $r_i$ converge faster in the initial stage due to their ability to evaluate more configurations promptly, but this advantage is gradually offset over time because they fail to find appropriate fidelities to create an accurate surrogate model. This causes a suboptimal final performance compared to vanilla BO, as shown in the first two figures in Figs. 3(a). In the case of the last one, we can observe a noticeable upward trend exhibited by the vanilla BO towards the end of the evaluation, indicating its potential to improve the final performance given abundant time. The comparison between the partial evaluation baselines and vanilla BO also demonstrates the importance of our adaptive strategy, which ensures that the fidelities align optimally with each configuration.", + "bbox": [ + 75, + 450, + 472, + 785 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.4. Generality of The Proposed Extension Method", + "text_level": 1, + "bbox": [ + 76, + 799, + 468, + 815 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The adaptive fidelity identification strategy provides a simple way to extend single-fidelity methods to the multifidelity setting, as discussed in § 4.4. To examine the ability of our extension method, we conduct experiments using three popular single-fidelity methods CQR [37], BORE [45]", + "bbox": [ + 75, + 825, + 468, + 902 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "and REA [35], extending them to the multi-fidelity variants with our extension method, referred to as FastCQR, FastBORE, and FastREA respectively. Similar to FastBO, all the multi-fidelity extensions evaluate the configurations to the adaptively identified efficient point and use the corresponding performances for the subsequent operations. The results on three datasets are illustrated in Fig. 3(b) and similar results on other datasets are in Supp. 11.6. We can clearly observe that the multi-fidelity variants with our extension method always outperform their single-fidelity counterparts. It is worth noting that REA is an evolutionary algorithm-based HPO method and is also significantly improved by our extension. The observation highlights the ability of the proposed adaptive strategy to extend any single-fidelity method to the multi-fidelity setting. It also suggests future opportunities to extend other advanced single-fidelity techniques into the multi-fidelity setting.", + "bbox": [ + 496, + 267, + 893, + 525 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 542, + 617, + 556 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we propose a model-based multi-fidelity HPO method FastBO, which adaptively identifies the appropriate fidelity for each configuration to fit the surrogate model and offers high-quality performance while ensuring efficient resource utilization. The advantages are achieved through our concepts of efficient and saturation point, the proposed techniques of learning curve modeling, and well-designed warm-up and post-processing stages with judicious early-termination detection and efficient saturation-level evaluation. Moreover, the proposed adaptive fidelity identification strategy provides a simple way to extend any single-fidelity method to the multi-fidelity setting. Experiments demonstrate the effectiveness and wide generality of our proposed techniques. FastBO source code is freely available at https://github.com/jjiantong/FastBO.", + "bbox": [ + 496, + 568, + 893, + 796 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgment", + "text_level": 1, + "bbox": [ + 500, + 813, + 651, + 830 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This research was funded by ARC Grant number DP190102443. Ajmal Mian is the recipient of an Australian Research Council Future Fellowship Award (project number FT210100268) funded by the Australian Government.", + "bbox": [ + 496, + 839, + 893, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "26188", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Sebastian E Ament and Carla P Gomes. Scalable first-order Bayesian Optimization via structured automatic differentiation. In International Conference on Machine Learning, pages 500-516. PMLR, 2022. 2", + "[2] James Bergstra, Rémi Bardenet, Yoshua Bengio, and Balázs Kégl. Algorithms for hyper-parameter optimization. Advances in Neural Information Processing Systems, 24, 2011. 1, 2", + "[3] Bernd Bischl, Martin Binder, Michel Lang, Tobias Pielok, Jakob Richter, Stefan Coors, Janek Thomas, Theresa Ullmann, Marc Becker, Anne-Laure Boulesteix, et al. Hyperparameter optimization: Foundations, algorithms, best practices, and open challenges. Wiley Interdisciplinary Reviews: Data Mining and Knowledge Discovery, 13(2):e1484, 2023. 2", + "[4] Ondrej Bohdal, Lukas Balles, Martin Wistuba, Beyza Ermis, Cedric Archambeau, and Giovanni Zappella. PASHA: efficient HPO and NAS with progressive resource allocation. In International Conference on Learning Representations. OpenReview.net, 2023. 1, 2, 6, 7, 4, 9", + "[5] Chaofan Chen, Oscar Li, Daniel Tao, Alina Barnett, Cynthia Rudin, and Jonathan K Su. This looks like that: deep learning for interpretable image recognition. Advances in Neural Information Processing Systems, 32, 2019. 2", + "[6] Tianqi Chen and Carlos Guestrin. Xgboost: A scalable tree boosting system. In Proceedings of the 22nd acm sigkdd international conference on knowledge discovery and data mining, pages 785-794, 2016. 9", + "[7] Tobias Domhan, Jost Tobias Springenberg, and Frank Hutter. Speeding up automatic hyperparameter optimization of deep neural networks by extrapolation of learning curves. In International Joint Conference on Artificial Intelligence, 2015. 4", + "[8] Xuanyi Dong and Yi Yang. NAS-Bench-201: Extending the scope of reproducible neural architecture search. In International Conference on Learning Representations, 2020. 1, 6, 8", + "[9] Thomas Elsken, Jan Hendrik Metzen, and Frank Hutter. Neural architecture search: A survey. The Journal of Machine Learning Research, 20(1):1997-2017, 2019. 1", + "[10] Stefan Falkner, Aaron Klein, and Frank Hutter. BOHB: Robust and efficient hyperparameter optimization at scale. In International Conference on Machine Learning, pages 1437-1446. PMLR, 2018. 1, 2, 6, 9", + "[11] Matthias Feurer and Frank Hutter. Hyperparameter optimization. Automated Machine Learning: Methods, Systems, Challenges, pages 3-33, 2019. 1", + "[12] Peter I Frazier, Warren B Powell, and Savas Dayanik. A knowledge-gradient policy for sequential information collection. SIAM Journal on Control and Optimization, 47(5): 2410-2439, 2008. 2", + "[13] José Miguel Hernández-Lobato, Matthew W Hoffman, and Zoubin Ghahramani. Predictive entropy search for efficient global optimization of black-box functions. Advances in Neural Information Processing systems, 27, 2014. 2" + ], + "bbox": [ + 78, + 114, + 470, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[14] Frank Hutter, Holger H Hoos, and Kevin Leyton-Brown. Sequential model-based optimization for general algorithm configuration. In Learning and Intelligent Optimization, pages 507-523. Springer, 2011. 1, 2", + "[15] Carl Hvarfner, Danny Stoll, Artur L. F. Souza, Marius Lindauer, Frank Hutter, and Luigi Nardi. $\\S$ pi $\\S$ BO: Augmenting acquisition functions with user beliefs for bayesian optimization. In International Conference on Learning Representations. OpenReview.net, 2022. 2", + "[16] Kevin Jamieson and Ameet Talwalkar. Non-stochastic best arm identification and hyperparameter optimization. In Artificial Intelligence and Statistics, pages 240-248. PMLR, 2016. 1, 2, 9", + "[17] Kasra Jamshidi, Harry Xu, and Keval Vora. Accelerating graph mining systems with subgraph morphing. In European Conference on Computer Systems, pages 162-181, 2023. 2", + "[18] Jiantong Jiang, Zeyi Wen, Zeke Wang, Bingsheng He, and Jian Chen. Parallel and distributed structuredsvm training. IEEE Transactions on Parallel and Distributed Systems, 33 (5):1084-1096, 2021.", + "[19] Jiantong Jiang, Zeyi Wen, and Ajmal Mian. Fast parallel bayesian network structure learning. In IEEE International Parallel and Distributed Processing Symposium, pages 617-627. IEEE, 2022. 2", + "[20] Arlind Kadra, Maciej Janowski, Martin Wistuba, and Josif Grabocka. Scaling laws for hyperparameter optimization. In Advances in Neural Information Processing Systems, 2023. 2", + "[21] Aaron Klein and Frank Hutter. Tabular benchmarks for joint architecture and hyperparameter optimization. arXiv preprint arXiv:1905.04970, 2019. 6, 8", + "[22] Aaron Klein, Louis C Tiao, Thibaut Lienart, Cedric Archambeau, and Matthias Seeger. Model-based asynchronous hyperparameter and neural architecture search. arXiv preprint arXiv:2003.10865, 2020. 1, 2, 6, 9", + "[23] Prasanth Kolachina, Nicola Cancedda, Marc Dymetman, and Sriram Venkatapathy. Prediction of learning curves in machine translation. In Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 22-30, 2012. 2", + "[24] Cheng Li, Santu Rana, Sunil Gupta, Vu Nguyen, Svetha Venkatesh, Alessandra Sutti, David Rubin de Celis Leal, Teo Slezak, Murray Height, Mazher Mohammed, and Ian Gibson. Accelerating experimental design by incorporating experimenter hunches. In International Conference on Data Mining, pages 257-266. IEEE Computer Society, 2018. 2", + "[25] Lisha Li, Kevin Jamieson, Giulia DeSalvo, Afshin Rostamizadeh, and Ameet Talwalkar. Hyperband: A novel bandit-based approach to hyperparameter optimization. Journal of Machine Learning Research, 18(1):6765-6816, 2017. 1, 2, 6, 9", + "[26] Liam Li, Kevin Jamieson, Afshin Rostamizadeh, Ekaterina Gonina, Jonathan Ben-Tzur, Moritz Hardt, Benjamin Recht, and Ameet Talwalkar. A system for massively parallel hyperparameter tuning. Proceedings of Machine Learning and Systems, 2:230-246, 2020. 1, 2, 6, 9", + "[27] Yang Li, Yu Shen, Huaijun Jiang, Wentao Zhang, Jixiang Li, Ji Liu, Ce Zhang, and Bin Cui. Hyper-tune: Towards" + ], + "bbox": [ + 501, + 92, + 893, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "26189", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "efficient hyper-parameter tuning at scale. Proceedings of the VLDB Endowment, 15(6):1256-1265, 2022. 1, 2, 6, 9", + "[28] Jonas Mockus. The application of Bayesian methods for seeking the extremum. Towards global optimization, 2:117, 1998. 2", + "[29] Felix Mohr and Jan N van Rijn. Learning curves for decision making in supervised machine learning-a survey. arXiv preprint arXiv:2201.12150, 2022. 4, 2", + "[30] Felix Mohr, Tom J Viering, Marco Loog, and Jan N van Rijn. Lcdb 1.0: An extensive learning curves database for classification tasks. In Joint European Conference on Machine Learning and Knowledge Discovery in Databases, pages 3-19. Springer, 2022. 2", + "[31] Julia Moosbauer, Julia Herbinger, Giuseppe Casalicchio, Marius Lindauer, and Bernd Bischl. Explaining hyperparameter optimization via partial dependence plots. Advances in Neural Information Processing Systems, 34:2280-2291, 2021. 2", + "[32] Julia Moosbauer, Giuseppe Casalicchio, Marius Lindauer, and Bernd Bischl. Improving accuracy of interpretability measures in hyperparameter optimization via Bayesian algorithm execution. arXiv preprint arXiv:2206.05447, 2022. 2", + "[33] ChangYong Oh, Efstratios Gavves, and Max Welling. BOCK: Bayesian optimization with cylindrical kernels. In International Conference on Machine Learning, pages 3868-3877. PMLR, 2018. 2", + "[34] Misha Padidar, Xinran Zhu, Leo Huang, Jacob Gardner, and David Bindel. Scaling gaussian processes with derivative information using variational inference. Advances in Neural Information Processing Systems, 34:6442-6453, 2021. 2", + "[35] Esteban Real, Alok Aggarwal, Yanping Huang, and Quoc V Le. Regularized evolution for image classifier architecture search. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 4780-4789, 2019. 8, 5, 9", + "[36] David Salinas, Matthias Seeger, Aaron Klein, Valerio Perrone, Martin Wistuba, and Cedric Archambeau. Syne tune: A library for large scale hyperparameter tuning and reproducible research. In International Conference on Automated Machine Learning, pages 16-1. PMLR, 2022. 6, 9", + "[37] David Salinas, Jacek Golebiowski, Aaron Klein, Matthias W. Seeger, and Cedric Archambeau. Optimizing hyperparameters with conformal quantile regression. In International Conference on Machine Learning, pages 29876-29893. PMLR, 2023. 1, 2, 6, 8, 5, 9", + "[38] Bobak Shahriari, Alexandre Bouchard-Côté, and Nando Freitas. Unbounded Bayesian Optimization via regularization. In Artificial intelligence and statistics, pages 1168-1176. PMLR, 2016. 2", + "[39] Julien Siems, Lucas Zimmer, Arber Zela, Jovita Lukasik, Margret Keuper, and Frank Hutter. Nas-bench-301 and the case for surrogate benchmarks for neural architecture search. CoRR, abs/2008.09777, 2020. 1, 6, 3", + "[40] Jasper Snoek, Hugo Larochelle, and Ryan P Adams. Practical Bayesian optimization of machine learning algorithms. Advances in Neural Information Processing Systems, 25, 2012. 1, 2, 6, 9" + ], + "bbox": [ + 78, + 90, + 468, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[41] Jasper Snoek, Oren Rippel, Kevin Swersky, Ryan Kiros, Nadathur Satish, Narayanan Sundaram, Mostofa Patwary, Mr Prabhat, and Ryan Adams. Scalable Bayesian optimization using deep neural networks. In International Conference on Machine Learning, pages 2171-2180. PMLR, 2015. 2", + "[42] Jost Tobias Springenberg, Aaron Klein, Stefan Falkner, and Frank Hutter. Bayesian optimization with robust bayesian neural networks. Advances in Neural Information Processing Systems, 29, 2016. 2", + "[43] Niranjan Srinivas, Andreas Krause, Sham M Kakade, and Matthias Seeger. Gaussian process optimization in the bandit setting: No regret and experimental design. arXiv preprint arXiv:0912.3995, 2009. 2", + "[44] Kevin Swersky, Jasper Snoek, and Ryan Prescott Adams. Freeze-thaw Bayesian optimization. arXiv preprint arXiv:1406.3896, 2014. 1", + "[45] Louis C Tiao, Aaron Klein, Matthias W Seeger, Edwin V Bonilla, Cedric Archambeau, and Fabio Ramos. BORE: Bayesian optimization by density-ratio estimation. In International Conference on Machine Learning, pages 10289-10300. PMLR, 2021. 8, 5, 9", + "[46] Tom Viering and Marco Loog. The shape of learning curves: a review. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2022. 1, 4, 2", + "[47] Jiazhuo Wang, Jason Xu, and Xuejun Wang. Combination of hyperband and bayesian optimization for hyperparameter optimization in deep learning. arXiv preprint arXiv:1801.01596, 2018. 1, 2", + "[48] Martin Wistuba, Arlind Kadra, and Josif Grabocka. Supervising the multi-fidelity race of hyperparameter configurations. Advances in Neural Information Processing Systems, 35:13470-13484, 2022. 2, 6, 9", + "[49] Jian Wu, Matthias Poloczek, Andrew G Wilson, and Peter Frazier. Bayesian optimization with gradients. Advances in neural information processing systems, 30, 2017. 2", + "[50] Peiyu Yang, Naveed Akhtar, Zeyi Wen, Mubarak Shah, and Ajmal Mian. Re-calibrating feature attributions for model interpretation. In International Conference on Learning Representations, 2022. 2", + "[51] Peiyu Yang, Naveed Akhtar, Zeyi Wen, and Ajmal Mian. Local path integration for attribution. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 3173-3180, 2023. 2", + "[52] Xueying Zhu, Jie Sun, Zhenhao He, Jiantong Jiang, and Zeke Wang. Staleness-reduction mini-batch $k$ -means. IEEE Transactions on Neural Networks and Learning Systems, 2023. 2", + "[53] Lucas Zimmer, Marius Thomas Lindauer, and Frank Hutter. Auto-Pytorch: Multi-fidelity metalearning for efficient and robust AutoDL. IEEE Transactions on Pattern Analysis and Machine Intelligence, 43:3079-3090, 2021. 6, 8" + ], + "bbox": [ + 501, + 92, + 890, + 811 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "26190", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2024/Efficient Hyperparameter Optimization with Adaptive Fidelity Identification/fa155004-89c9-4599-9ab7-9a5d6edc5fbc_model.json b/2024/Efficient Hyperparameter Optimization with Adaptive Fidelity Identification/fa155004-89c9-4599-9ab7-9a5d6edc5fbc_model.json new file mode 100644 index 0000000000000000000000000000000000000000..01e750dfb5504702eab433911ded4bd358810224 --- /dev/null +++ b/2024/Efficient Hyperparameter Optimization with Adaptive Fidelity Identification/fa155004-89c9-4599-9ab7-9a5d6edc5fbc_model.json @@ -0,0 +1,1925 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.099, + 0.131, + 0.872, + 0.153 + ], + "angle": 0, + "content": "Efficient Hyperparameter Optimization with Adaptive Fidelity Identification" + }, + { + "type": "text", + "bbox": [ + 0.2, + 0.18, + 0.77, + 0.217 + ], + "angle": 0, + "content": "Jiantong Jiang\\(^{1}\\), Zeyi Wen\\(^{2,3*}\\), Atif Mansoor\\(^{1}\\), Ajmal Mian\\(^{1}\\) \n\\(^{1}\\)The University of Western Australia, \\(^{2}\\)HKUST (Guangzhou), \\(^{3}\\)HKUST" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.22, + 0.875, + 0.235 + ], + "angle": 0, + "content": "jiantong.jiang@research.uwa.edu.au, wenzeyi@ust.hk, {atif.mansoor, ajmal.mian}@uwa.edu.au" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.269, + 0.314, + 0.285 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.301, + 0.474, + 0.62 + ], + "angle": 0, + "content": "Hyperparameter Optimization and Neural Architecture Search are powerful in attaining state-of-the-art machine learning models, with Bayesian Optimization (BO) standing out as a mainstream method. Extending BO into the multi-fidelity setting has been an emerging research topic in this field, but faces the challenge of determining an appropriate fidelity for each hyperparameter configuration to fit the surrogate model. To tackle the challenge, we propose a multi-fidelity BO method named FastBO, which excels in adaptively deciding the fidelity for each configuration and providing strong performance while ensuring efficient resource usage. These advantages are achieved through our proposed techniques based on the concepts of efficient point and saturation point for each configuration, which can be obtained from the empirical learning curve of the configuration, estimated from early observations. Extensive experiments demonstrate FastBO's superior anytime performance and efficiency in identifying high-quality configurations and architectures. We also show that our method provides a way to extend any single-fidelity method to the multi-fidelity setting, highlighting the wide applicability of our approach." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.655, + 0.21, + 0.672 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.681, + 0.47, + 0.846 + ], + "angle": 0, + "content": "Hyperparameters are crucial in machine learning pipelines. Hyperparameter optimization (HPO) [11] and Neural Architecture Search (NAS) [9] aims to find the hyperparameters or architectures that can yield good performance without human experts. Among different HPO and NAS methods, Bayesian Optimization (BO) [2, 14, 40] is an effective model-based method that has shown remarkable success [8, 39]. BO maintains a surrogate model of the target performance metric based on past evaluations of hyperparameter configurations, which guides the choice of more promising configurations to evaluate." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.848, + 0.47, + 0.879 + ], + "angle": 0, + "content": "Despite its sample efficiency, standard BO requires a full evaluation of each configuration, involving full-scale" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.27, + 0.895, + 0.497 + ], + "angle": 0, + "content": "training and testing of models, which can be highly time-consuming, particularly with the recent trend to larger models. To avoid expensive full evaluations, multi-fidelity methods [4, 16, 25, 26] have been proposed, where the fidelity refers to the levels of performance metrics obtained under different resource levels. These methods follow the principle of successive halving (SHA) [16]: initially, they evaluate a set of randomly selected configurations using a small number of resources; then, based on the low-fidelity performances, the poorly-performing ones are successively eliminated, while the well-performing ones continue to be evaluated with increasing resources. Follow-up studies [10, 22, 27, 37, 47] propose model-based multi-fidelity methods, replacing the random configuration selection with a more informed model to improve sample efficiency." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.503, + 0.895, + 0.837 + ], + "angle": 0, + "content": "Nevertheless, current model-based multi-fidelity methods face a major limitation: they are built upon the SHA framework, which operates under the assumption that learning curves of different configurations rarely intersect. This assumption does not hold in practice [46], i.e., early performance observations cannot always indicate the final fidelity performance at the full resource level. This leads to a fundamental challenge when extending model-based methods to the multi-fidelity setting: What is the appropriate fidelity for each configuration to fit the surrogate model? In other words, which fidelity can provide performance observations that reliably indicate the final fidelity performance? Existing methods struggle to address this fundamental challenge. In particular, BOHB [10] and Hyper-Tune [27] fit separate surrogate models for different fidelities, failing to capture inter-fidelity correlations. FTBO [44] and A-BOHB [22] fit a joint model but require strong assumptions to remain tractable. Another work by Salinas et al. [37] suggests using the last observed fidelity performance to fit the surrogate model. However, it widens the gap between poorly- and well-performing configurations at the early stage, potentially leading to an inaccurate surrogate model." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.895, + 0.903 + ], + "angle": 0, + "content": "To this end, we propose a multi-fidelity extension of BO, namely FastBO, which tackles the challenge of deciding the appropriate fidelity for each configuration to fit the surrogate model. FastBO identifies a so-called efficient point for" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.887, + 0.305, + 0.901 + ], + "angle": 0, + "content": "*Zeyi Wen is the corresponding author." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "26181" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.364 + ], + "angle": 0, + "content": "each configuration to be the fidelity. The point balances computational cost and performance quality while capturing valuable learning curve trends. In essence, FastBO selects the fidelity for each configuration instead of evaluating all the configurations at the same fidelity. Additionally, a saturation point for each configuration is identified to be an approximation of the final fidelity, leading to high-quality performance while reducing resource wastage. The two crucial points are adaptively derived from the estimated learning curve of each configuration. Furthermore, the warm-up and post-processing stages are carefully designed to enable judicious early-termination detection and efficient saturation level evaluation. Empirical evaluation against the state-of-the-art methods shows that FastBO has strong anytime performance and can considerably save up to \\(87\\%\\) of the time required to identify a good configuration or architecture, lowering the barriers for engaging in HPO and NAS. In summary, we make the following major contributions." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.368, + 0.47, + 0.429 + ], + "angle": 0, + "content": "1. We propose a multi-fidelity model-based HPO method that adaptively decides the fidelities for configurations and efficiently offers strong performance, thanks to the introduced concepts of efficient and saturation points." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.429, + 0.47, + 0.488 + ], + "angle": 0, + "content": "2. We develop a learning curve modeling module to enable adaptive derivation of the key points, a warm-up stage to allow early-termination detection, and a post-processing stage to ensure efficient saturation-level evaluation." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.488, + 0.47, + 0.548 + ], + "angle": 0, + "content": "3. We show that our strategy can be used to extend existing single-fidelity methods to the multi-fidelity setting, demonstrating the effectiveness and generality of our method and highlighting promising future opportunities." + }, + { + "type": "list", + "bbox": [ + 0.078, + 0.368, + 0.47, + 0.548 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.559, + 0.22, + 0.575 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.584, + 0.47, + 0.689 + ], + "angle": 0, + "content": "HPO and NAS incur significant costs, especially considering the escalating model evaluation overhead. Despite efforts to directly accelerate computation [17-19, 52], smarter HPO strategies are still required for the widespread adoption of automation. Two crucial directions include model-based and multi-fidelity methods, which can also be combined. Here, we review the methods in these categories." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.689, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Model-based methods. BO is the representative of model-based methods. Based on the surrogate model constructed by historical evaluation results, BO selects the configurations to evaluate via an acquisition function that balances exploration and exploitation. Commonly used surrogate models are Gaussian processes [40], random forests [14], tree-structured Parzen estimator [2], and deep networks [41, 42]. Popular acquisition functions include Expected Improvement [28], Knowledge Gradient [12], Upper Confidence Bound [43], and Predictive Entropy Search [13]. Recent studies on BO have explored the utilization of expert priors [15, 24, 33, 38] and derivative information [1, 34, 49]. There also has been a recent focus on enhancing the interpretability [5, 50, 51] of the HPO process [3, 31, 32]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.273 + ], + "angle": 0, + "content": "Multi-fidelity methods. Multi-fidelity methods exploit low and high fidelities for configurations to save the evaluation time. Successive halving (SHA) [16] runs a set of configurations using a small number of resources and promotes only the best-performing half of configurations to continue for twice as many resources. Hyperband [25] calls SHA as a sub-routine with varying maximum resources and introduces a reduction factor to control the fraction of promotion. ASHA [26] extends SHA to the asynchronous setting by aggressive early-stopping. Later, PASHA [4] further extends ASHA through more aggressive early-stopping based on the ranking of configurations during the tuning process." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.273, + 0.892, + 0.47 + ], + "angle": 0, + "content": "Combination of model-based and multi-fidelity methods. BOHB [10] and a parallel work [47] first propose to combine model-based and multi-fidelity methods by replacing the random sampling in Hyperband with BO. A BOHB [22] employs a joint GP surrogate over fidelities and supports asynchronous scheduling. Hyper-Tune [27] improves its Hyperband by a delayed strategy to decrease inaccurate promotions. Salinas et al. [37] proposed to extend methods to multi-fidelity settings by using the performance of the last fidelity in an ASHA running. DyHPO [48] and DPL [20] introduce new surrogates for multi-fidelity BO considering the learning curves; the former uses deep GP kernels while the latter integrates deep power law functions." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.483, + 0.707, + 0.499 + ], + "angle": 0, + "content": "3. Problem Formulation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.508, + 0.892, + 0.584 + ], + "angle": 0, + "content": "Given an algorithm having hyperparameters \\(\\lambda_1, \\dots, \\lambda_m\\) with domains \\(\\Lambda_1, \\dots, \\Lambda_m\\), we define its hyperparameter space as \\(\\Lambda = \\Lambda_1 \\times \\dots \\times \\Lambda_m\\). Here, we define the problem and outline the key challenge related to hyperparameter optimization (HPO). Notations are in Supp. 7 for reference." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.584, + 0.892, + 0.854 + ], + "angle": 0, + "content": "Single-fidelity setting. For each hyperparameter configuration \\(\\lambda\\), we denote \\(f(\\lambda)\\) as the performance achieved using \\(\\lambda\\). For consistency, the metric in this paper refers to descending metrics like validation loss, with ascending metrics being treated similarly. In the single-fidelity HPO setting, we aim to find \\(\\lambda^*\\) minimizing function \\(f(\\lambda)\\), i.e., \\(\\lambda^* = \\arg \\min_{\\lambda \\in \\Lambda} f(\\lambda)\\). BO is one of the most popular single-fidelity HPO methods. The vanilla BO has two key components: a surrogate model \\(\\mathcal{M}\\) to approximate the objective function \\(f(\\lambda)\\), and an acquisition function \\(a\\) to identify a promising configuration from search space. With these ingredients, BO iterates three steps: (i) select a configuration \\(\\lambda_i\\) by maximizing the acquisition function; (ii) evaluate \\(\\lambda_i\\) to get \\(y_i\\) and add the data \\((\\lambda_i, y_i)\\) into the current observation set \\(\\mathcal{D}_{i-1} = \\{(\\lambda_1, y_1), \\dots, (\\lambda_{i-1}, y_{i-1})\\}\\); (iii) update the surrogate model and the acquisition function based on the augmented \\(\\mathcal{D}_i\\). In this work, \\(\\mathcal{M}\\) is a Gaussian Process and \\(a\\) is Expected Improvement." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Multi-fidelity setting. Multi-fidelity methods consider resource information, such as training epochs or training subset ratios. Evaluations at various resource levels results in" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "26182" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.198 + ], + "angle": 0, + "content": "different performance levels, known as the fidelities. Different fidelities provide a way to balance computational cost and performance quality. In multi-fidelity HPO problems, the target is extended to \\(\\lambda^{*} = \\arg \\min_{\\lambda \\in \\Lambda}f(\\lambda ,r)\\), where \\(f(\\lambda ,r)\\) is the objective function obtained for \\(\\lambda\\) at \\(r\\). We use \\(r\\) to denote the resource level, which can also be interpreted as the fidelity, and \\(r\\in \\{r_{min},\\dots,r_{max}\\}\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.199, + 0.473, + 0.485 + ], + "angle": 0, + "content": "Extending single-fidelity methods to the multi-fidelity setting. The inefficiency of single-fidelity methods stems from their reliance on the final fidelity evaluation of \\( f(\\lambda, r_{max}) \\) to be the evaluation of its objective \\( f(\\lambda, r) \\). Fitting surrogate models by such final fidelity evaluations incurs high cost due to the full evaluation of the configurations. Notably, low-fidelity evaluations at \\( r < r_{max} \\) provide informative insights into the objective but are computationally cheaper, which is valuable to the optimization process. Therefore, we seek an effective way to extend single-fidelity methods like BO to the multi-fidelity setting. More specifically, recalling the earlier steps of BO, when evaluating the configuration \\( \\lambda_i \\) in the second step, we instead acquire its low-fidelity performance \\( y_i^{r_i} \\) at \\( r_i \\), where \\( r_i \\) denotes the fidelity used for \\( \\lambda_i \\) to fit the surrogate model. The observations \\( \\mathcal{D}_i \\) then become \\( \\{(\\lambda_1, y_1^{r_1}), \\dots, (\\lambda_i, y_i^{r_i})\\} \\). To conclude, in order to extend single-fidelity methods to the multi-fidelity setting, the key challenge to be addressed is to determine \\( r_i \\) for each \\( \\lambda_i \\)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.5, + 0.214, + 0.518 + ], + "angle": 0, + "content": "4. Methodology" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.526, + 0.47, + 0.678 + ], + "angle": 0, + "content": "In this section, we propose a novel multi-fidelity model-based algorithm FastBO. We first propose the key concepts of efficient point and saturation point, which are crucial in deciding the fidelity level to fit the surrogate model and to approximate the final fidelity respectively. Secondly, we elaborate on the details of learning curve modeling, where the two crucial points can be extracted. Then, we present the techniques associated with the auxiliary warm-up and post-processing stages. Finally, we summarize FastBO and discuss its wide applicability to any single-fidelity methods." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.688, + 0.459, + 0.703 + ], + "angle": 0, + "content": "4.1. Estimation of Efficient and Saturation Points" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.712, + 0.468, + 0.772 + ], + "angle": 0, + "content": "In our method, we adaptively identify efficient and saturation points for each configuration. The two points are crucial in the optimization process. We first formally define the efficient point as follows." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.784, + 0.468, + 0.86 + ], + "angle": 0, + "content": "Definition 1 (Efficient point). For a given learning curve \\(\\mathcal{C}_i(r)\\) of hyperparameter configuration \\(\\lambda_{i}\\), where \\(r\\) represents the resource level (also referred to as fidelity), the efficient point \\(e_i\\) of \\(\\lambda_{i}\\) is defined as: \\(e_i = \\min \\{r\\mid \\mathcal{C}_i(r) - \\mathcal{C}_i(2r) < \\delta_1\\}\\), where \\(\\delta_{1}\\) is a predefined small threshold." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.468, + 0.901 + ], + "angle": 0, + "content": "The semantic of Definition 1 is that starting from the efficient point onwards, when the resources are doubled (i.e.," + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.198 + ], + "angle": 0, + "content": "from \\( r \\) to \\( 2r \\)), the performance improvement falls below a small threshold \\( \\delta_1 \\). Consequently, this point characterizes the fidelity at which a configuration demonstrates strong performance while still efficiently utilizing resources. In simpler terms, it signifies an appropriate fidelity of performance that can be achieved with comparably efficient resource usage. Therefore, we make the following remark." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.208, + 0.894, + 0.299 + ], + "angle": 0, + "content": "Remark 1. The efficient points of the hyperparameter configurations can serve as their appropriate fidelities used for fitting the surrogate model. This is due to their (i) optimal resource-to-performance balance, (ii) ability to capture valuable learning curve trends, and (iii) customization for different hyperparameter configurations." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.308, + 0.893, + 0.549 + ], + "angle": 0, + "content": "We elaborate on the reasons in Remark 1 as follows. Firstly, efficient points balance the trade-off between computational cost and result quality. Beyond the efficient point of a given configuration, allocating additional resources to that configuration becomes less efficient. Secondly, efficient points capture valuable trends within the learning curves. For example, the learning rate influences the shape of learning curves; the identification of efficient points for configurations with smaller learning rates often occurs at later stages. The insights into learning curve behaviors enable more informed decision-making. Thirdly, the ability to customize the fidelity for each specific configuration is a significant advantage. This adaptive approach is more reasonable than previous studies that use a fixed fidelity for all configurations, as it better accounts for the unique characteristics of individual learning curves." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.55, + 0.892, + 0.655 + ], + "angle": 0, + "content": "This insight leads us to use the efficient point \\( e_i \\) identified for each configuration \\( \\lambda_i \\) as its fidelity used to fit the surrogate model. Specifically, we evaluate \\( \\lambda_i \\) until reaching \\( e_i \\) and obtain the observed performance \\( y_i^{e_i} \\). The resulting data point \\( (\\lambda_i, y_i^{e_i}) \\) is then added into the current observation set \\( \\mathcal{D}_{i-1} \\) to refit the surrogate model. We proof the superiority of FastBO over SHA-based methods in Supp. 8." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.656, + 0.892, + 0.7 + ], + "angle": 0, + "content": "Besides efficient points, we identify saturation points for all configurations from their learning curves as well. We provide the definition of the saturation point as follows." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.71, + 0.894, + 0.786 + ], + "angle": 0, + "content": "Definition 2 (Saturation point). For a given learning curve \\(\\mathcal{C}_i(r)\\) of configuration \\(\\lambda_i\\), where \\(r\\) represents the resource level (also referred to as fidelity), the saturation point \\(s_i\\) of \\(\\lambda_i\\) is defined as: \\(s_i = \\min \\{r \\mid \\forall r' > r, |\\mathcal{C}_i(r') - \\mathcal{C}_i(r)| < \\delta_2\\}\\), where \\(\\delta_2\\) is a predefined small threshold." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.796, + 0.892, + 0.901 + ], + "angle": 0, + "content": "The semantic of Definition 2 is that beyond the saturation point, the observed performance no longer exhibits notable variations with more resources. Thus, this point characterizes the fidelity at which the performance of a configuration stabilizes. The concept of saturation point is well-recognized within the machine learning community. Building on the above definition, we make the following remark." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "26183" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.075, + 0.092, + 0.47, + 0.155 + ], + "angle": 0, + "content": "Remark 2. The saturation points of the hyperparameter configurations can serve as their approximate final fidelities, as they provide performance results that meet predefined quality thresholds while reducing resource wastage." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.16, + 0.471, + 0.298 + ], + "angle": 0, + "content": "This insight leads us to use the saturation point \\( s_i \\) identified for each configuration \\( \\lambda_i \\) as its final fidelity approximation. The point is used in the post-processing stage for promoting some well-performing configurations to get higher-fidelity performances. In essence, when aiming for a full evaluation of the configurations, we suggest that terminating the evaluation at the saturation point is sufficient. A more intuitive illustration of the concepts of efficient and saturation points is provided in Supp. 9." + }, + { + "type": "title", + "bbox": [ + 0.076, + 0.304, + 0.315, + 0.321 + ], + "angle": 0, + "content": "4.2. Learning Curve Modeling" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.328, + 0.47, + 0.584 + ], + "angle": 0, + "content": "From Definitions 1 and 2, we can extract the efficient and saturation points of configurations from their learning curves. The curve \\(\\mathcal{C}_i(r)\\) corresponds to configuration \\(\\lambda_{i}\\) and describes the predictive performance with \\(\\lambda_{i}\\) as a function of the fidelity \\(r\\). Here, \\(r\\) can be either the number of training instances or the number of training epochs. In the context of learning curves, the former is referred to as observation learning curves, while the latter is iteration learning curves [29]. Both types are applicable to FastBO, so we use the term learning curve to encompass both. Given the observation set \\(\\mathcal{O}_i^w = \\{(r,y_i^r)\\}_{r = r_{min},\\dots,w}\\) for \\(\\lambda_{i}\\), which comprises pairs of data points representing fidelities \\(r\\in \\{r_{min},\\dots,w\\}\\) and the corresponding evaluations \\(y_{i}^{r}\\), where \\(w\\) is a pre-defined warm-up point to stop collecting data, FastBO can estimate a learning curve for \\(\\lambda_{i}\\) based on \\(\\mathcal{O}_i^w\\) by first constructing a parametric learning curve model, then estimating the parameters." + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.585, + 0.471, + 0.781 + ], + "angle": 0, + "content": "Constructing a parametric learning curve model. Empirical learning curves can be modeled with function classes relying on some parameters. Viering and Loog [46] comprehensively summarized the parametric models studied in machine learning. In practice, different problems have different learning curves; even under the same problem, different hyperparameter configurations (e.g., learning rate, regularization, etc.) may lead to significantly different learning curves. Since one single parametric model is not enough to characterize all the learning curves by itself, we consider combining different parametric models into a single model. Specifically, we consider three parametric models POW3, EXP3 and LOG2, as listed in Tab. 1, which have shown" + }, + { + "type": "table_caption", + "bbox": [ + 0.129, + 0.793, + 0.416, + 0.807 + ], + "angle": 0, + "content": "Table 1. Parametric learning curve models used." + }, + { + "type": "table", + "bbox": [ + 0.126, + 0.818, + 0.422, + 0.897 + ], + "angle": 0, + "content": "
ModelFormulaFamily
POW3y = d + ax-αPower law
EXP3y = d + e-ax+bExponential
LOG2y = d + a log(x)Logarithmic
" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.891, + 0.137 + ], + "angle": 0, + "content": "good fitting and predicting performance in previous empirical studies [29, 46]. We provide detailed discussions on the choice of parametric models in Supp. 10." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.137, + 0.892, + 0.198 + ], + "angle": 0, + "content": "Here, we denote each parametric model as \\(c_{j}(r|\\theta_{j})\\) with parameters \\(\\theta_{j}\\), where the independent variable \\(r\\) represents the fidelity. We combine three models into one model via a weighted linear combination:" + }, + { + "type": "equation", + "bbox": [ + 0.578, + 0.21, + 0.892, + 0.236 + ], + "angle": 0, + "content": "\\[\n\\mathcal {C} (r | \\phi) = \\sum_ {j \\in \\{1, 2, 3 \\}} \\omega_ {j} c _ {j} (r | \\boldsymbol {\\theta} _ {j}), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.247, + 0.892, + 0.337 + ], + "angle": 0, + "content": "where \\(\\phi = \\{\\omega_1,\\omega_2,\\omega_3,\\theta_1,\\theta_2,\\theta_3\\}\\) is the parameter of the combined model, which consists of parameters \\(\\{\\pmb {\\theta}_1,\\pmb {\\theta}_2,\\pmb {\\theta}_3\\}\\) and weight \\(\\{\\omega_{1},\\omega_{2},\\omega_{3}\\}\\) of every single model. Therefore, each pair of observations \\((r,y_i^r)\\) in \\(O_{i}^{w}\\) can be modeled by the combined model as \\(y_{i}^{r} = \\mathcal{C}(r|\\phi) + \\epsilon\\) , where \\(y_{i}^{r}\\) is the observed dependent variable and \\(\\epsilon\\) represents the error term." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.339, + 0.892, + 0.459 + ], + "angle": 0, + "content": "Estimating parameters in the parametric learning curve model. We employ maximum likelihood estimation to estimate the parameters \\(\\phi\\) in the parametric model \\(\\mathcal{C}(r|\\phi)\\). Assuming that \\(\\epsilon \\sim \\mathcal{N}(0,\\sigma^2)\\), the probability of an observed performance \\(y_{i}^{r}\\) under parameters is given by \\(p(y_i^r |\\phi ,\\sigma^2) = \\mathcal{N}(y_i^r;\\mathcal{C}(r|\\phi),\\sigma^2)\\). Given the observations \\(\\mathcal{O}_i^w\\) of \\(\\lambda_{i}\\) that contains a set of observed data points \\((r,y_i^r)\\), the likelihood function can be expressed as:" + }, + { + "type": "equation", + "bbox": [ + 0.525, + 0.472, + 0.891, + 0.537 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} (\\boldsymbol {\\phi}, \\sigma^ {2}; \\mathbf {r}, \\mathbf {y} _ {i} ^ {r}) = \\prod p (y _ {i} ^ {k} | \\boldsymbol {\\phi}, \\sigma^ {2}) \\\\ = \\prod_ {k = r _ {\\text {m i n}}} ^ {w} \\frac {1}{\\sigma \\sqrt {2 \\pi}} \\exp \\left(- \\frac {\\left(y _ {i} ^ {r} - \\mathcal {C} (r = k | \\phi)\\right) ^ {2}}{2 \\sigma^ {2}}\\right). \\tag {2} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.548, + 0.89, + 0.578 + ], + "angle": 0, + "content": "We estimate \\(\\phi\\) by maximizing log-likelihood function, which is easily calculated given Eq. 2." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.58, + 0.892, + 0.701 + ], + "angle": 0, + "content": "An existing model-free method [7] also considers using learning curves for the HPO problem. However, it targets predicting the high-fidelity performance from the low-fidelity observations and thus stopping configurations that are unlikely to beat the current best values, which is different from our main target of identifying appropriate fidelity levels for the configurations to fit the surrogate model from their estimated learning curves." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.711, + 0.825, + 0.728 + ], + "angle": 0, + "content": "4.3. Warm-up And Post-processing Stages" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.735, + 0.892, + 0.825 + ], + "angle": 0, + "content": "In addition to its core components, FastBO incorporates two auxiliary stages: the warm-up and post-processing stages. For the completeness of our method, we provide an overview of these stages, outlining their targets and presenting the key techniques of early-termination detection and saturation-level evaluation that are applied within." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.826, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Warm-up stage. The warm-up stage prepares the early observation set \\(\\mathcal{O}_i^w\\) for each configuration \\(\\lambda_{i}\\) that is used to estimate its learning curve, as discussed in \\(\\S 4.2\\). Here \\(w\\in (r_{min},r_{max})\\) is a pre-determined fidelity, denoted as warm-up point. Specifically, we initiate the evaluation of" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "26184" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.473, + 0.424 + ], + "angle": 0, + "content": "each newly selected \\(\\lambda_{i}\\), proceeding until reaching \\(w\\). During this process, we record each fidelity \\(r\\) and its evaluation result \\(y_{i}^{r}\\), forming pairs \\((r,y_i^r)\\). Upon reaching \\(w\\), we pause the evaluation for \\(\\lambda_{i}\\) and obtain its early observation set \\(\\mathcal{O}_i^w = \\{(r,y_i^r)\\}_{r = r_{min},\\ldots ,w}\\), and start modeling the learning curve. During the warm-up stage, we monitor the performance changes across every two continuous fidelities. If we detect that the performance of \\(\\lambda_{i}\\) has consecutively dropped twice by more than a ratio \\(\\alpha\\), i.e., \\((y_{i}^{r - 1} - y_{i}^{r - 2}) > \\alpha y_{i}^{r - 2}\\) and \\((y_i^r -y_i^{r - 1}) > \\alpha y_i^{r - 1}\\), we promptly terminate the evaluation for \\(\\lambda_{i}\\) at its current fidelity \\(r\\), because such consecutive performance deterioration indicates \\(\\lambda_{i}\\) is unlikely to achieve satisfactory performance. Once terminated, we directly incorporate the current performance \\(y_{i}^{r}\\) of \\(\\lambda_{i}\\) into \\(\\mathcal{D}_{i - 1}\\) that is used for updating the surrogate model. Thus, further operations like learning curve modeling are discontinued for \\(\\lambda_{i}\\). Moreover, if we observe a single case of performance drop without subsequent occurrences, i.e., \\(y_{i}^{r - 1} - y_{i}^{r - 2} > \\alpha y_{i}^{r - 2}\\) and \\(y_{i}^{r} - y_{i}^{r - 1}\\leq \\alpha y_{i}^{r - 1}\\), we opt not to include data from fidelity \\(r - 1\\) in \\(\\mathcal{O}_i^w\\). This is to manually filter out potential noise in the data that may adversely affect the fitting of the learning curve." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.426, + 0.473, + 0.835 + ], + "angle": 0, + "content": "Post-processing stage. The post-processing stage aims at two tasks: promoting the well-performing configurations for saturation-level evaluations and identifying the best configuration and its performance. Firstly, FastBO promotes the top-\\(k\\) well-performing configurations and evaluates them to their saturation points to ensure high-quality performance while maintaining efficient resource utilization. We set \\(k\\) to be always less than or equal to the number of parallel workers available, ensuring a manageable overhead of saturation-level evaluations. It is worth noting that the additional time required is factored into the overall time. Secondly, FastBO finds the best configuration along with its performance achieved so far, which is a standard final step in most HPO methods. However, an increase in fidelities does not always result in performance improvement, possibly due to overfitting, resource saturation, or problem complexity. Therefore, we treat the evaluation at each fidelity as an individual task and record all these intermediate evaluation results, which is also a common practice in recent implementations. In this way, FastBO finds the best performance by considering all the results, rather than relying solely on the highest-fidelity performances of the configurations. In the parallel setting, treating each fidelity evaluation as an individual task offers an added benefit due to its finer granularity. More specifically, when a worker is idle, it takes on a new task of evaluating a configuration at a specific fidelity, rather than evaluating an entire configuration." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.847, + 0.325, + 0.862 + ], + "angle": 0, + "content": "4.4. FastBO and Generalization" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.47, + 0.903 + ], + "angle": 0, + "content": "Algorithm 1 summarizes our proposed FastBO. It takes surrogate model \\(\\mathcal{M}\\), acquisition function \\(a\\), warm-up point \\(w\\)," + }, + { + "type": "code_caption", + "bbox": [ + 0.508, + 0.094, + 0.728, + 0.111 + ], + "angle": 0, + "content": "Algorithm 1: FastBO algorithm" + }, + { + "type": "algorithm", + "bbox": [ + 0.5, + 0.112, + 0.892, + 0.374 + ], + "angle": 0, + "content": "input : \\(\\mathcal{M},a,w,\\alpha ,k,\\delta_1,\\delta_2\\) output: \\(\\lambda^{*},y^{*}\\) \n1 \\(i\\gets 0,D\\gets \\emptyset\\) \n2 while not meet the stop criterion do \n3 find \\(\\lambda_{i}\\gets \\arg \\max_{\\lambda \\in \\Lambda}a(\\lambda ,\\mathcal{M}_{i - 1})\\) \n4 \\(\\mathcal{O}_i^w,t\\gets\\) warm-up given \\(w,\\alpha\\) // cf. \\(\\S 4.3\\) \n5 if \\(\\mathcal{O}_i^w\\) is not empty then \n6 fit \\(\\mathcal{C}_i(r)\\) to \\(\\mathcal{O}_i^w\\) // cf. \\(\\S 4.2\\) \n7 find \\(e_i,s_i\\) given \\(\\mathcal{C}_i(r),\\delta_1,\\delta_2\\) // cf. \\(\\S 4.1\\) \n8 \\(y_{i}^{e_{i}}\\gets\\) continue evaluating \\(\\lambda_{i}\\) to \\(e_i\\) \n9 else \n10 \\(e_i\\gets t,s_i\\gets r_{max}\\) \n11 \\(\\mathcal{D}_i\\gets \\mathcal{D}_{i - 1}\\cup (\\lambda_i,y_i^{e_i})\\) \n12 refit \\(\\mathcal{M}_i\\) to \\(\\mathcal{D}_i\\) \n13 \\(i\\gets i + 1\\) \n14 \\(\\lambda^{*},y^{*}\\gets\\) post-process given s = \\(\\{s_i\\} ,k / / c f.\\S 4.3\\)" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.413, + 0.895, + 0.732 + ], + "angle": 0, + "content": "performance decrease ratio \\(\\alpha\\), promotion number \\(k\\), and thresholds \\(\\delta_{1}\\), \\(\\delta_{2}\\) as inputs, and output the best-founded configuration \\(\\lambda^{*}\\) and its performance \\(y^{*}\\). FastBO follows a similar iterative process of model-based methods but replaces the expensive full evaluations with a more intelligent alternative (cf. Lines 4-10). Specifically, each configuration \\(\\lambda_{i}\\) first enters a warm-up stage to collect its early observation set \\(\\mathcal{O}_i^w\\) and to be detected and terminated if it exhibits consecutive performance deterioration (cf. Line 4). If \\(\\lambda_{i}\\) is not terminated, FastBO then estimates a learning curve \\(\\mathcal{C}_i(r)\\) for \\(\\lambda_{i}\\) based on \\(\\mathcal{O}_i^w\\) (cf. Line 6), and thus the efficient point and saturation point of \\(\\lambda_{i}\\) can be obtained (cf. Line 7). After that, \\(\\lambda_{i}\\) continues to be evaluated until reaching \\(e_i\\) (cf. Line 8); the result is added to the observation set \\(\\mathcal{D}\\) (cf. Line 11) that is used for updating \\(\\mathcal{M}\\) (cf. Line 12). On the other hand, the poorly-performing configuration will be terminated early at fidelity \\(t\\) with its result being added directly to \\(\\mathcal{D}\\) (cf. Lines 10, 11). Finally, the post-processing stage promotes the most promising configurations to their saturation points and finds the best-founded configuration \\(\\lambda^{*}\\) and its performance \\(y^{*}\\) (cf. Line 14)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.735, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Generalizing FastBO to single-fidelity methods. The core of FastBO is to tackle the key challenge of deciding an appropriate fidelity for each configuration to fit the surrogate model by adaptively identifying its efficient point. This strategy of using the efficient point performances for surrogate model fitting also provides a simple but effective way to bridge the gap between single- and multi-fidelity methods. While it is primarily described in the context of model-based methods, the strategy can be generalized to various single-fidelity methods. For example, when evaluating configurations within the population for an evolution-" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "26185" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.078, + 0.089, + 0.895, + 0.239 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.238, + 0.246, + 0.731, + 0.261 + ], + "angle": 0, + "content": "Figure 1. Performance of average validation accuracy on the LCBench benchmark." + }, + { + "type": "image", + "bbox": [ + 0.077, + 0.275, + 0.891, + 0.436 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.141, + 0.445, + 0.828, + 0.46 + ], + "angle": 0, + "content": "Figure 2. Performance of (a) average validation error on NAS-Bench-201 and (b) average validation loss on FCNet." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.479, + 0.473, + 0.646 + ], + "angle": 0, + "content": "ary algorithm-based HPO method, we can similarly evaluate the efficient point performances instead of the final performances of these configurations and integrate the performances in the subsequent processes, such as selection and variation. Relying on the efficient point rather than the final fidelity or all fidelities available simplifies the extension of the single-fidelity methods to the multi-fidelity setting. The rationale behind this adaptive fidelity identification strategy is discussed in Remark 1. We also demonstrate in our experiments the efficacy of this strategy in extending a range of single-fidelity methods to the multi-fidelity setting." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.662, + 0.21, + 0.68 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.688, + 0.47, + 0.824 + ], + "angle": 0, + "content": "We empirically evaluate the performance of FastBO and compare it with the random search baseline (RS) and 9 competitive baselines from 3 related categories, including (i) model-based methods: standard Gaussian Process-based BO [40]; (ii) multi-fidelity methods: ASHA [26], Hyperband [25], PASHA [4]; and (iii) model-based multifidelity methods: A-BOHB [22], A-CQR [37], BOHB [10], DyHPO [48], Hyper-Tune [27]. RS and BO are single-fidelity baselines, while the others are multi-fidelity ones." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Our experiments are conducted on 10 datasets from 3 popular benchmarks LCBench [53], NAS-Bench-201 [8] and FCNet [21]. Detailed information on the benchmarks is provided in Supp. 13.1. All the experiments are evaluated with four parallel workers and 10 random seeds. We" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.479, + 0.895, + 0.603 + ], + "angle": 0, + "content": "allocate \\(20\\%\\) total budget for warm-up, i.e., \\(w = r_{min} + 0.2 \\cdot (r_{max} - r_{min})\\). Ratio \\(\\alpha\\) is set to 0.1; thresholds \\(\\delta_1\\) and \\(\\delta_2\\) are set to 0.001 and \\(0.0005^1\\). We set \\(k\\) based on the number of workers \\(\\# w\\) and the number of started configurations \\(\\# c\\): \\(k = \\max \\{\\lceil \\# c / 10 \\rceil, \\# w\\}\\). We provide more experiments and discussions on the hyperparameters in Supp. 12. We use implementations of the baselines in Syne Tune [36]. Details of the baseline settings are in Supp. 13.2." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.612, + 0.707, + 0.629 + ], + "angle": 0, + "content": "5.1. Anytime Performance" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.635, + 0.894, + 0.741 + ], + "angle": 0, + "content": "To evaluate the anytime performance, we compare FastBO against the baselines on wall-clock time. For fair comparisons, all the baselines, even single-fidelity BO and RS, are extended to consider intermediate results at all the fidelities when identifying the configuration, akin to FastBO as discussed in § 4.3. Consequently, all the baselines are able to achieve their best possible anytime performance." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.742, + 0.895, + 0.864 + ], + "angle": 0, + "content": "The results on LCBench, NAS-Bench-201, and FCNet are shown in Figs. 1 and 2. We report the validation accuracy, error, and loss over wall-clock time for the three benchmarks, as provided by the benchmarks. We provide the results on NAS-Bench-301 [39] in Supp. 11.1. Overall, FastBO can handle various performance metrics and shows strong anytime performance. We can observe that FastBO gains an advantage earlier than other methods, rapidly con" + }, + { + "type": "page_footnote", + "bbox": [ + 0.5, + 0.875, + 0.892, + 0.9 + ], + "angle": 0, + "content": "1Parameters \\(\\delta_{1}\\) and \\(\\delta_{2}\\) given here are derived after standardizing metrics to a uniform scale from 0 to 1." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "26186" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.089, + 0.895, + 0.133 + ], + "angle": 0, + "content": "Table 2. Comparison of relative efficiency on configuration identification. FastBO is set as the baseline with a relative efficiency of 1.00. Wall-clock time (abbr. WC time) reports the elapsed time spent for each method on finding configurations with similar performance metrics, i.e., validation error \\((\\times 10^{-2})\\) for Covertype and ImageNet16-120 and validation loss \\((\\times 10^{-5})\\) for Slice." + }, + { + "type": "table", + "bbox": [ + 0.086, + 0.144, + 0.887, + 0.354 + ], + "angle": 0, + "content": "
Metric DatasetMethodFastBOBOPASHAA-BOHBA-CQRBOHBDyHPOHyper-Tune
CovertypeVal. error22.9±0.223.0±0.325.1±2.523.5±1.131.6±1.932.5±0.823.0±0.323.0±0.2
WC time (h)0.7±0.32.9±0.73.9±1.02.0±1.03.9±0.22.5±1.01.7±0.61.8±0.7
Rel. efficiency1.000.250.180.370.190.290.410.40
ImageNet 16-120Val. error55.3±0.257.4±1.255.7±0.355.8±1.655.5±0.955.5±1.155.5±1.055.3±2.0
WC time (h)2.2±0.76.6±0.92.5±1.25.9±1.16.0±1.33.2±0.74.3±1.03.4±1.1
Rel. efficiency1.000.340.900.380.370.680.510.67
SliceVal. loss26.3±2.626.4±4.426.8±9.526.3±6.327.1±4.226.8±5.627.4±2.328.7±1.3
WC time (h)0.4±0.13.1±0.71.2±0.92.1±0.72.5±0.72.2±0.92.5±0.51.8±0.6
Rel. efficiency1.000.130.350.200.170.190.170.24
" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.382, + 0.427, + 0.398 + ], + "angle": 0, + "content": "verging to the global optimum after the initial phase." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.4, + 0.471, + 0.702 + ], + "angle": 0, + "content": "The superiority can be attributed to two main factors. Firstly, FastBO maintains, and in some cases even surpasses, the sample efficiency of vanilla BO, thanks to our techniques that enable quick and precise identification of the fidelities for configurations to update the surrogate model. We provide more explanations and experiments on sample efficiency in Supp. 11.2. Secondly, the multi-fidelity extension speeds up the evaluations, contributing to its overall efficiency. In contrast, the single-fidelity baselines tend to waste more time on the full evaluations. While the multifidelity baselines efficiently explore numerous configurations, they limit their evaluations to only constrained fidelities for some time, thus struggling to provide relatively high performance in a short time. This issue in multi-fidelity methods is particularly pronounced in PASHA when applied to NAS-Bench-201 and FCNet, as shown in Fig. 2. In Supp. 11.3, we further provide the ranks of all methods and statistically show FastBO's superiority on an early stage. It is worth noting that all the additional overhead introduced by FastBO is taken into account in the wall-clock time." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.705, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Regarding the final performance, most methods are able to converge to satisfactory solutions, with negligible differences among them in most cases. Although our goal is not to offer the best final performance as we limit the evaluations to at most the saturation point even for those we consider most promising, FastBO still achieves top-2 final performance on 8 out of 10 datasets. In contrast, model-free methods sometimes cannot obtain a satisfactory final performance because they randomly select the configurations. For example, on the \"Covertype\" dataset, only 3 out of 2000 configurations yield a validation accuracy exceeding \\(75\\%\\). As a result, all the model-free methods face challenges in converging to a satisfactory final performance." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.381, + 0.857, + 0.399 + ], + "angle": 0, + "content": "5.2. Efficiency on Configuration Identification" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.405, + 0.892, + 0.586 + ], + "angle": 0, + "content": "One explanation for PASHA's suboptimal anytime performance (cf. Fig. 2) lies in its primary goal [4]: the goal of PASHA is not high accuracy but to identify the best configuration more quickly. To ensure equitable comparisons, we report the time spent for each method on identifying a satisfactory configuration, consistent with the experiments described in PASHA [4]. Results on three expensive datasets \"Covertype\", \"ImageNet16-120\", and \"Slice\" of the three benchmarks are shown in Tab. 2. Similar results on additional datasets can be found in Supp. 11.4. Besides PASHA, results of other model-free multi-fidelity methods are not included, as PASHA demonstrates its superiority over them." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.587, + 0.893, + 0.814 + ], + "angle": 0, + "content": "Tab. 2 shows that FastBO saves \\(10\\%\\) to \\(87\\%\\) wall-clock time over other methods when achieving up to \\(9.6\\%\\) better performance values. It can be observed from the \"rel. efficiency\" rows, where we set FastBO as the baseline with a relative efficiency of 1.00 and report the efficiency of other methods relative to ours. When compared with vanilla BO, FastBO significantly shortens the time in identifying a good configuration by a factor of 3 to 8, because FastBO pauses a configuration earlier at an appropriate fidelity and fits the surrogate model to guide the next configuration search. This advantage creates opportunities to efficiently explore more configurations. Another observation is that PASHA always gets a relatively high variance in wall-clock time. This is due to the fact that different random seeds can have a larger impact on such model-free methods." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.823, + 0.892, + 0.84 + ], + "angle": 0, + "content": "5.3. Effectiveness of Adaptive Fidelity Identification" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.847, + 0.892, + 0.879 + ], + "angle": 0, + "content": "As discussed in § 4.1, FastBO is able to adaptively identify the efficient point \\( e_i \\) for each configuration \\( \\lambda_i \\) and serves" + }, + { + "type": "page_footnote", + "bbox": [ + 0.516, + 0.887, + 0.871, + 0.902 + ], + "angle": 0, + "content": "2We convert the accuracy of \"Covertype\" into error for readability." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "26187" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.065, + 0.087, + 0.905, + 0.212 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.22, + 0.893, + 0.25 + ], + "angle": 0, + "content": "Figure 3. Performance comparison: (a) Performance of FastBO that adaptively sets \\( r_i = e_i \\) with the schemes that use fixed \\( r_i \\) for all configurations. (b) Performance of single-fidelity methods CQR, BORE, REA and their multi-fidelity variants using our extension method." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.268, + 0.473, + 0.449 + ], + "angle": 0, + "content": "\\(e_i\\) as its fidelity \\(r_i\\) for surrogate model fitting. To investigate the effectiveness of the adaptive fidelity identification strategy, we conduct an ablation study to compare the performance achieved with and without applying this strategy. Specifically, we compare FastBO, where \\(r_i\\) is adaptively set to \\(e_i\\), with the partial evaluation schemes that employ fixed predefined values as the fidelity for all the configurations to fit the surrogate model. We consider three representative fixed fidelities, including \\(25\\%\\), \\(50\\%\\), and \\(75\\%\\) of the total resource budget. In addition, we include a comparison with vanilla BO that can be viewed as using \\(100\\%\\) resource budget as the fixed fidelity for all configurations." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.452, + 0.473, + 0.786 + ], + "angle": 0, + "content": "We provide the results on three representative datasets in Fig. 3(a), with more results available in Supp. 11.5. We have three main observations. Firstly, FastBO always outperforms the partial evaluation baselines that use a fixed fidelity, indicating the effectiveness of the adaptive strategy. Secondly, FastBO shows stronger performance than vanilla BO. The limitation of vanilla BO lies in the additional time required for full evaluations. Secondly, compared to the vanilla BO, partial evaluation schemes with fixed \\( r_i \\) converge faster in the initial stage due to their ability to evaluate more configurations promptly, but this advantage is gradually offset over time because they fail to find appropriate fidelities to create an accurate surrogate model. This causes a suboptimal final performance compared to vanilla BO, as shown in the first two figures in Figs. 3(a). In the case of the last one, we can observe a noticeable upward trend exhibited by the vanilla BO towards the end of the evaluation, indicating its potential to improve the final performance given abundant time. The comparison between the partial evaluation baselines and vanilla BO also demonstrates the importance of our adaptive strategy, which ensures that the fidelities align optimally with each configuration." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.8, + 0.47, + 0.816 + ], + "angle": 0, + "content": "5.4. Generality of The Proposed Extension Method" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.826, + 0.47, + 0.903 + ], + "angle": 0, + "content": "The adaptive fidelity identification strategy provides a simple way to extend single-fidelity methods to the multifidelity setting, as discussed in § 4.4. To examine the ability of our extension method, we conduct experiments using three popular single-fidelity methods CQR [37], BORE [45]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.268, + 0.895, + 0.526 + ], + "angle": 0, + "content": "and REA [35], extending them to the multi-fidelity variants with our extension method, referred to as FastCQR, FastBORE, and FastREA respectively. Similar to FastBO, all the multi-fidelity extensions evaluate the configurations to the adaptively identified efficient point and use the corresponding performances for the subsequent operations. The results on three datasets are illustrated in Fig. 3(b) and similar results on other datasets are in Supp. 11.6. We can clearly observe that the multi-fidelity variants with our extension method always outperform their single-fidelity counterparts. It is worth noting that REA is an evolutionary algorithm-based HPO method and is also significantly improved by our extension. The observation highlights the ability of the proposed adaptive strategy to extend any single-fidelity method to the multi-fidelity setting. It also suggests future opportunities to extend other advanced single-fidelity techniques into the multi-fidelity setting." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.543, + 0.619, + 0.558 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.569, + 0.895, + 0.797 + ], + "angle": 0, + "content": "In this paper, we propose a model-based multi-fidelity HPO method FastBO, which adaptively identifies the appropriate fidelity for each configuration to fit the surrogate model and offers high-quality performance while ensuring efficient resource utilization. The advantages are achieved through our concepts of efficient and saturation point, the proposed techniques of learning curve modeling, and well-designed warm-up and post-processing stages with judicious early-termination detection and efficient saturation-level evaluation. Moreover, the proposed adaptive fidelity identification strategy provides a simple way to extend any single-fidelity method to the multi-fidelity setting. Experiments demonstrate the effectiveness and wide generality of our proposed techniques. FastBO source code is freely available at https://github.com/jjiantong/FastBO." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.814, + 0.652, + 0.831 + ], + "angle": 0, + "content": "Acknowledgment" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.894, + 0.901 + ], + "angle": 0, + "content": "This research was funded by ARC Grant number DP190102443. Ajmal Mian is the recipient of an Australian Research Council Future Fellowship Award (project number FT210100268) funded by the Australian Government." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "26188" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.171 + ], + "angle": 0, + "content": "[1] Sebastian E Ament and Carla P Gomes. Scalable first-order Bayesian Optimization via structured automatic differentiation. In International Conference on Machine Learning, pages 500-516. PMLR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.173, + 0.472, + 0.227 + ], + "angle": 0, + "content": "[2] James Bergstra, Rémi Bardenet, Yoshua Bengio, and Balázs Kégl. Algorithms for hyper-parameter optimization. Advances in Neural Information Processing Systems, 24, 2011. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.23, + 0.471, + 0.327 + ], + "angle": 0, + "content": "[3] Bernd Bischl, Martin Binder, Michel Lang, Tobias Pielok, Jakob Richter, Stefan Coors, Janek Thomas, Theresa Ullmann, Marc Becker, Anne-Laure Boulesteix, et al. Hyperparameter optimization: Foundations, algorithms, best practices, and open challenges. Wiley Interdisciplinary Reviews: Data Mining and Knowledge Discovery, 13(2):e1484, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.329, + 0.47, + 0.398 + ], + "angle": 0, + "content": "[4] Ondrej Bohdal, Lukas Balles, Martin Wistuba, Beyza Ermis, Cedric Archambeau, and Giovanni Zappella. PASHA: efficient HPO and NAS with progressive resource allocation. In International Conference on Learning Representations. OpenReview.net, 2023. 1, 2, 6, 7, 4, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.401, + 0.471, + 0.457 + ], + "angle": 0, + "content": "[5] Chaofan Chen, Oscar Li, Daniel Tao, Alina Barnett, Cynthia Rudin, and Jonathan K Su. This looks like that: deep learning for interpretable image recognition. Advances in Neural Information Processing Systems, 32, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.458, + 0.471, + 0.514 + ], + "angle": 0, + "content": "[6] Tianqi Chen and Carlos Guestrin. Xgboost: A scalable tree boosting system. In Proceedings of the 22nd acm sigkdd international conference on knowledge discovery and data mining, pages 785-794, 2016. 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.515, + 0.47, + 0.583 + ], + "angle": 0, + "content": "[7] Tobias Domhan, Jost Tobias Springenberg, and Frank Hutter. Speeding up automatic hyperparameter optimization of deep neural networks by extrapolation of learning curves. In International Joint Conference on Artificial Intelligence, 2015. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.586, + 0.471, + 0.641 + ], + "angle": 0, + "content": "[8] Xuanyi Dong and Yi Yang. NAS-Bench-201: Extending the scope of reproducible neural architecture search. In International Conference on Learning Representations, 2020. 1, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.644, + 0.47, + 0.685 + ], + "angle": 0, + "content": "[9] Thomas Elsken, Jan Hendrik Metzen, and Frank Hutter. Neural architecture search: A survey. The Journal of Machine Learning Research, 20(1):1997-2017, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.687, + 0.47, + 0.743 + ], + "angle": 0, + "content": "[10] Stefan Falkner, Aaron Klein, and Frank Hutter. BOHB: Robust and efficient hyperparameter optimization at scale. In International Conference on Machine Learning, pages 1437-1446. PMLR, 2018. 1, 2, 6, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.745, + 0.469, + 0.787 + ], + "angle": 0, + "content": "[11] Matthias Feurer and Frank Hutter. Hyperparameter optimization. Automated Machine Learning: Methods, Systems, Challenges, pages 3-33, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.788, + 0.47, + 0.843 + ], + "angle": 0, + "content": "[12] Peter I Frazier, Warren B Powell, and Savas Dayanik. A knowledge-gradient policy for sequential information collection. SIAM Journal on Control and Optimization, 47(5): 2410-2439, 2008. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.845, + 0.47, + 0.901 + ], + "angle": 0, + "content": "[13] José Miguel Hernández-Lobato, Matthew W Hoffman, and Zoubin Ghahramani. Predictive entropy search for efficient global optimization of black-box functions. Advances in Neural Information Processing systems, 27, 2014. 2" + }, + { + "type": "list", + "bbox": [ + 0.079, + 0.115, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.148 + ], + "angle": 0, + "content": "[14] Frank Hutter, Holger H Hoos, and Kevin Leyton-Brown. Sequential model-based optimization for general algorithm configuration. In Learning and Intelligent Optimization, pages 507-523. Springer, 2011. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.149, + 0.894, + 0.218 + ], + "angle": 0, + "content": "[15] Carl Hvarfner, Danny Stoll, Artur L. F. Souza, Marius Lindauer, Frank Hutter, and Luigi Nardi. \\(\\S\\) pi\\(\\S\\)BO: Augmenting acquisition functions with user beliefs for bayesian optimization. In International Conference on Learning Representations. OpenReview.net, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.219, + 0.892, + 0.272 + ], + "angle": 0, + "content": "[16] Kevin Jamieson and Ameet Talwalkar. Non-stochastic best arm identification and hyperparameter optimization. In Artificial Intelligence and Statistics, pages 240-248. PMLR, 2016. 1, 2, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.274, + 0.892, + 0.316 + ], + "angle": 0, + "content": "[17] Kasra Jamshidi, Harry Xu, and Keval Vora. Accelerating graph mining systems with subgraph morphing. In European Conference on Computer Systems, pages 162-181, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.317, + 0.892, + 0.371 + ], + "angle": 0, + "content": "[18] Jiantong Jiang, Zeyi Wen, Zeke Wang, Bingsheng He, and Jian Chen. Parallel and distributed structuredsvm training. IEEE Transactions on Parallel and Distributed Systems, 33 (5):1084-1096, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.372, + 0.892, + 0.426 + ], + "angle": 0, + "content": "[19] Jiantong Jiang, Zeyi Wen, and Ajmal Mian. Fast parallel bayesian network structure learning. In IEEE International Parallel and Distributed Processing Symposium, pages 617-627. IEEE, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.428, + 0.892, + 0.481 + ], + "angle": 0, + "content": "[20] Arlind Kadra, Maciej Janowski, Martin Wistuba, and Josif Grabocka. Scaling laws for hyperparameter optimization. In Advances in Neural Information Processing Systems, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.484, + 0.892, + 0.525 + ], + "angle": 0, + "content": "[21] Aaron Klein and Frank Hutter. Tabular benchmarks for joint architecture and hyperparameter optimization. arXiv preprint arXiv:1905.04970, 2019. 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.526, + 0.892, + 0.58 + ], + "angle": 0, + "content": "[22] Aaron Klein, Louis C Tiao, Thibaut Lienart, Cedric Archambeau, and Matthias Seeger. Model-based asynchronous hyperparameter and neural architecture search. arXiv preprint arXiv:2003.10865, 2020. 1, 2, 6, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.581, + 0.892, + 0.649 + ], + "angle": 0, + "content": "[23] Prasanth Kolachina, Nicola Cancedda, Marc Dymetman, and Sriram Venkatapathy. Prediction of learning curves in machine translation. In Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 22-30, 2012. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.651, + 0.892, + 0.734 + ], + "angle": 0, + "content": "[24] Cheng Li, Santu Rana, Sunil Gupta, Vu Nguyen, Svetha Venkatesh, Alessandra Sutti, David Rubin de Celis Leal, Teo Slezak, Murray Height, Mazher Mohammed, and Ian Gibson. Accelerating experimental design by incorporating experimenter hunches. In International Conference on Data Mining, pages 257-266. IEEE Computer Society, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.735, + 0.892, + 0.802 + ], + "angle": 0, + "content": "[25] Lisha Li, Kevin Jamieson, Giulia DeSalvo, Afshin Rostamizadeh, and Ameet Talwalkar. Hyperband: A novel bandit-based approach to hyperparameter optimization. Journal of Machine Learning Research, 18(1):6765-6816, 2017. 1, 2, 6, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.804, + 0.892, + 0.872 + ], + "angle": 0, + "content": "[26] Liam Li, Kevin Jamieson, Afshin Rostamizadeh, Ekaterina Gonina, Jonathan Ben-Tzur, Moritz Hardt, Benjamin Recht, and Ameet Talwalkar. A system for massively parallel hyperparameter tuning. Proceedings of Machine Learning and Systems, 2:230-246, 2020. 1, 2, 6, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.873, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[27] Yang Li, Yu Shen, Huaijun Jiang, Wentao Zhang, Jixiang Li, Ji Liu, Ce Zhang, and Bin Cui. Hyper-tune: Towards" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "26189" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.092, + 0.469, + 0.12 + ], + "angle": 0, + "content": "efficient hyper-parameter tuning at scale. Proceedings of the VLDB Endowment, 15(6):1256-1265, 2022. 1, 2, 6, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.122, + 0.47, + 0.162 + ], + "angle": 0, + "content": "[28] Jonas Mockus. The application of Bayesian methods for seeking the extremum. Towards global optimization, 2:117, 1998. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.165, + 0.469, + 0.207 + ], + "angle": 0, + "content": "[29] Felix Mohr and Jan N van Rijn. Learning curves for decision making in supervised machine learning-a survey. arXiv preprint arXiv:2201.12150, 2022. 4, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.208, + 0.469, + 0.277 + ], + "angle": 0, + "content": "[30] Felix Mohr, Tom J Viering, Marco Loog, and Jan N van Rijn. Lcdb 1.0: An extensive learning curves database for classification tasks. In Joint European Conference on Machine Learning and Knowledge Discovery in Databases, pages 3-19. Springer, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.279, + 0.469, + 0.346 + ], + "angle": 0, + "content": "[31] Julia Moosbauer, Julia Herbinger, Giuseppe Casalicchio, Marius Lindauer, and Bernd Bischl. Explaining hyperparameter optimization via partial dependence plots. Advances in Neural Information Processing Systems, 34:2280-2291, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.349, + 0.469, + 0.417 + ], + "angle": 0, + "content": "[32] Julia Moosbauer, Giuseppe Casalicchio, Marius Lindauer, and Bernd Bischl. Improving accuracy of interpretability measures in hyperparameter optimization via Bayesian algorithm execution. arXiv preprint arXiv:2206.05447, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.42, + 0.469, + 0.475 + ], + "angle": 0, + "content": "[33] ChangYong Oh, Efstratios Gavves, and Max Welling. BOCK: Bayesian optimization with cylindrical kernels. In International Conference on Machine Learning, pages 3868-3877. PMLR, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.477, + 0.469, + 0.532 + ], + "angle": 0, + "content": "[34] Misha Padidar, Xinran Zhu, Leo Huang, Jacob Gardner, and David Bindel. Scaling gaussian processes with derivative information using variational inference. Advances in Neural Information Processing Systems, 34:6442-6453, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.534, + 0.469, + 0.589 + ], + "angle": 0, + "content": "[35] Esteban Real, Alok Aggarwal, Yanping Huang, and Quoc V Le. Regularized evolution for image classifier architecture search. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 4780-4789, 2019. 8, 5, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.591, + 0.469, + 0.659 + ], + "angle": 0, + "content": "[36] David Salinas, Matthias Seeger, Aaron Klein, Valerio Perrone, Martin Wistuba, and Cedric Archambeau. Syne tune: A library for large scale hyperparameter tuning and reproducible research. In International Conference on Automated Machine Learning, pages 16-1. PMLR, 2022. 6, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.661, + 0.469, + 0.73 + ], + "angle": 0, + "content": "[37] David Salinas, Jacek Golebiowski, Aaron Klein, Matthias W. Seeger, and Cedric Archambeau. Optimizing hyperparameters with conformal quantile regression. In International Conference on Machine Learning, pages 29876-29893. PMLR, 2023. 1, 2, 6, 8, 5, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.732, + 0.469, + 0.786 + ], + "angle": 0, + "content": "[38] Bobak Shahriari, Alexandre Bouchard-Côté, and Nando Freitas. Unbounded Bayesian Optimization via regularization. In Artificial intelligence and statistics, pages 1168-1176. PMLR, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.789, + 0.469, + 0.843 + ], + "angle": 0, + "content": "[39] Julien Siems, Lucas Zimmer, Arber Zela, Jovita Lukasik, Margret Keuper, and Frank Hutter. Nas-bench-301 and the case for surrogate benchmarks for neural architecture search. CoRR, abs/2008.09777, 2020. 1, 6, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.846, + 0.469, + 0.9 + ], + "angle": 0, + "content": "[40] Jasper Snoek, Hugo Larochelle, and Ryan P Adams. Practical Bayesian optimization of machine learning algorithms. Advances in Neural Information Processing Systems, 25, 2012. 1, 2, 6, 9" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.162 + ], + "angle": 0, + "content": "[41] Jasper Snoek, Oren Rippel, Kevin Swersky, Ryan Kiros, Nadathur Satish, Narayanan Sundaram, Mostofa Patwary, Mr Prabhat, and Ryan Adams. Scalable Bayesian optimization using deep neural networks. In International Conference on Machine Learning, pages 2171-2180. PMLR, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.164, + 0.892, + 0.219 + ], + "angle": 0, + "content": "[42] Jost Tobias Springenberg, Aaron Klein, Stefan Falkner, and Frank Hutter. Bayesian optimization with robust bayesian neural networks. Advances in Neural Information Processing Systems, 29, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.221, + 0.892, + 0.275 + ], + "angle": 0, + "content": "[43] Niranjan Srinivas, Andreas Krause, Sham M Kakade, and Matthias Seeger. Gaussian process optimization in the bandit setting: No regret and experimental design. arXiv preprint arXiv:0912.3995, 2009. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.277, + 0.892, + 0.317 + ], + "angle": 0, + "content": "[44] Kevin Swersky, Jasper Snoek, and Ryan Prescott Adams. Freeze-thaw Bayesian optimization. arXiv preprint arXiv:1406.3896, 2014. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.319, + 0.892, + 0.387 + ], + "angle": 0, + "content": "[45] Louis C Tiao, Aaron Klein, Matthias W Seeger, Edwin V Bonilla, Cedric Archambeau, and Fabio Ramos. BORE: Bayesian optimization by density-ratio estimation. In International Conference on Machine Learning, pages 10289-10300. PMLR, 2021. 8, 5, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.39, + 0.892, + 0.432 + ], + "angle": 0, + "content": "[46] Tom Viering and Marco Loog. The shape of learning curves: a review. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2022. 1, 4, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.434, + 0.892, + 0.487 + ], + "angle": 0, + "content": "[47] Jiazhuo Wang, Jason Xu, and Xuejun Wang. Combination of hyperband and bayesian optimization for hyperparameter optimization in deep learning. arXiv preprint arXiv:1801.01596, 2018. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.489, + 0.892, + 0.544 + ], + "angle": 0, + "content": "[48] Martin Wistuba, Arlind Kadra, and Josif Grabocka. Supervising the multi-fidelity race of hyperparameter configurations. Advances in Neural Information Processing Systems, 35:13470-13484, 2022. 2, 6, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.546, + 0.892, + 0.587 + ], + "angle": 0, + "content": "[49] Jian Wu, Matthias Poloczek, Andrew G Wilson, and Peter Frazier. Bayesian optimization with gradients. Advances in neural information processing systems, 30, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.589, + 0.892, + 0.642 + ], + "angle": 0, + "content": "[50] Peiyu Yang, Naveed Akhtar, Zeyi Wen, Mubarak Shah, and Ajmal Mian. Re-calibrating feature attributions for model interpretation. In International Conference on Learning Representations, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.645, + 0.892, + 0.699 + ], + "angle": 0, + "content": "[51] Peiyu Yang, Naveed Akhtar, Zeyi Wen, and Ajmal Mian. Local path integration for attribution. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 3173-3180, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.702, + 0.892, + 0.756 + ], + "angle": 0, + "content": "[52] Xueying Zhu, Jie Sun, Zhenhao He, Jiantong Jiang, and Zeke Wang. Staleness-reduction mini-batch \\(k\\)-means. IEEE Transactions on Neural Networks and Learning Systems, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.758, + 0.892, + 0.813 + ], + "angle": 0, + "content": "[53] Lucas Zimmer, Marius Thomas Lindauer, and Frank Hutter. Auto-Pytorch: Multi-fidelity metalearning for efficient and robust AutoDL. IEEE Transactions on Pattern Analysis and Machine Intelligence, 43:3079-3090, 2021. 6, 8" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.813 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "26190" + } + ] +] \ No newline at end of file diff --git a/2024/Efficient Hyperparameter Optimization with Adaptive Fidelity Identification/fa155004-89c9-4599-9ab7-9a5d6edc5fbc_origin.pdf b/2024/Efficient Hyperparameter Optimization with Adaptive Fidelity Identification/fa155004-89c9-4599-9ab7-9a5d6edc5fbc_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..462f55e01f1f744a95d969117ec4407c1c0aee3e --- /dev/null +++ b/2024/Efficient Hyperparameter Optimization with Adaptive Fidelity Identification/fa155004-89c9-4599-9ab7-9a5d6edc5fbc_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2dd38b6728f75d5eeed64032ec90b8c16e6514d2ea58fd7239a7fc1dd7298924 +size 959606 diff --git a/2024/Efficient Hyperparameter Optimization with Adaptive Fidelity Identification/full.md b/2024/Efficient Hyperparameter Optimization with Adaptive Fidelity Identification/full.md new file mode 100644 index 0000000000000000000000000000000000000000..fb9676e6440fc4f93ad7886997c6fbdef44642d5 --- /dev/null +++ b/2024/Efficient Hyperparameter Optimization with Adaptive Fidelity Identification/full.md @@ -0,0 +1,267 @@ +# Efficient Hyperparameter Optimization with Adaptive Fidelity Identification + +Jiantong Jiang $^{1}$ , Zeyi Wen $^{2,3*}$ , Atif Mansoor $^{1}$ , Ajmal Mian $^{1}$ $^{1}$ The University of Western Australia, $^{2}$ HKUST (Guangzhou), $^{3}$ HKUST + +jiantong.jiang@research.uwa.edu.au, wenzeyi@ust.hk, {atif.mansoor, ajmal.mian}@uwa.edu.au + +# Abstract + +Hyperparameter Optimization and Neural Architecture Search are powerful in attaining state-of-the-art machine learning models, with Bayesian Optimization (BO) standing out as a mainstream method. Extending BO into the multi-fidelity setting has been an emerging research topic in this field, but faces the challenge of determining an appropriate fidelity for each hyperparameter configuration to fit the surrogate model. To tackle the challenge, we propose a multi-fidelity BO method named FastBO, which excels in adaptively deciding the fidelity for each configuration and providing strong performance while ensuring efficient resource usage. These advantages are achieved through our proposed techniques based on the concepts of efficient point and saturation point for each configuration, which can be obtained from the empirical learning curve of the configuration, estimated from early observations. Extensive experiments demonstrate FastBO's superior anytime performance and efficiency in identifying high-quality configurations and architectures. We also show that our method provides a way to extend any single-fidelity method to the multi-fidelity setting, highlighting the wide applicability of our approach. + +# 1. Introduction + +Hyperparameters are crucial in machine learning pipelines. Hyperparameter optimization (HPO) [11] and Neural Architecture Search (NAS) [9] aims to find the hyperparameters or architectures that can yield good performance without human experts. Among different HPO and NAS methods, Bayesian Optimization (BO) [2, 14, 40] is an effective model-based method that has shown remarkable success [8, 39]. BO maintains a surrogate model of the target performance metric based on past evaluations of hyperparameter configurations, which guides the choice of more promising configurations to evaluate. + +Despite its sample efficiency, standard BO requires a full evaluation of each configuration, involving full-scale + +training and testing of models, which can be highly time-consuming, particularly with the recent trend to larger models. To avoid expensive full evaluations, multi-fidelity methods [4, 16, 25, 26] have been proposed, where the fidelity refers to the levels of performance metrics obtained under different resource levels. These methods follow the principle of successive halving (SHA) [16]: initially, they evaluate a set of randomly selected configurations using a small number of resources; then, based on the low-fidelity performances, the poorly-performing ones are successively eliminated, while the well-performing ones continue to be evaluated with increasing resources. Follow-up studies [10, 22, 27, 37, 47] propose model-based multi-fidelity methods, replacing the random configuration selection with a more informed model to improve sample efficiency. + +Nevertheless, current model-based multi-fidelity methods face a major limitation: they are built upon the SHA framework, which operates under the assumption that learning curves of different configurations rarely intersect. This assumption does not hold in practice [46], i.e., early performance observations cannot always indicate the final fidelity performance at the full resource level. This leads to a fundamental challenge when extending model-based methods to the multi-fidelity setting: What is the appropriate fidelity for each configuration to fit the surrogate model? In other words, which fidelity can provide performance observations that reliably indicate the final fidelity performance? Existing methods struggle to address this fundamental challenge. In particular, BOHB [10] and Hyper-Tune [27] fit separate surrogate models for different fidelities, failing to capture inter-fidelity correlations. FTBO [44] and A-BOHB [22] fit a joint model but require strong assumptions to remain tractable. Another work by Salinas et al. [37] suggests using the last observed fidelity performance to fit the surrogate model. However, it widens the gap between poorly- and well-performing configurations at the early stage, potentially leading to an inaccurate surrogate model. + +To this end, we propose a multi-fidelity extension of BO, namely FastBO, which tackles the challenge of deciding the appropriate fidelity for each configuration to fit the surrogate model. FastBO identifies a so-called efficient point for + +each configuration to be the fidelity. The point balances computational cost and performance quality while capturing valuable learning curve trends. In essence, FastBO selects the fidelity for each configuration instead of evaluating all the configurations at the same fidelity. Additionally, a saturation point for each configuration is identified to be an approximation of the final fidelity, leading to high-quality performance while reducing resource wastage. The two crucial points are adaptively derived from the estimated learning curve of each configuration. Furthermore, the warm-up and post-processing stages are carefully designed to enable judicious early-termination detection and efficient saturation level evaluation. Empirical evaluation against the state-of-the-art methods shows that FastBO has strong anytime performance and can considerably save up to $87\%$ of the time required to identify a good configuration or architecture, lowering the barriers for engaging in HPO and NAS. In summary, we make the following major contributions. + +1. We propose a multi-fidelity model-based HPO method that adaptively decides the fidelities for configurations and efficiently offers strong performance, thanks to the introduced concepts of efficient and saturation points. +2. We develop a learning curve modeling module to enable adaptive derivation of the key points, a warm-up stage to allow early-termination detection, and a post-processing stage to ensure efficient saturation-level evaluation. +3. We show that our strategy can be used to extend existing single-fidelity methods to the multi-fidelity setting, demonstrating the effectiveness and generality of our method and highlighting promising future opportunities. + +# 2. Related Work + +HPO and NAS incur significant costs, especially considering the escalating model evaluation overhead. Despite efforts to directly accelerate computation [17-19, 52], smarter HPO strategies are still required for the widespread adoption of automation. Two crucial directions include model-based and multi-fidelity methods, which can also be combined. Here, we review the methods in these categories. + +Model-based methods. BO is the representative of model-based methods. Based on the surrogate model constructed by historical evaluation results, BO selects the configurations to evaluate via an acquisition function that balances exploration and exploitation. Commonly used surrogate models are Gaussian processes [40], random forests [14], tree-structured Parzen estimator [2], and deep networks [41, 42]. Popular acquisition functions include Expected Improvement [28], Knowledge Gradient [12], Upper Confidence Bound [43], and Predictive Entropy Search [13]. Recent studies on BO have explored the utilization of expert priors [15, 24, 33, 38] and derivative information [1, 34, 49]. There also has been a recent focus on enhancing the interpretability [5, 50, 51] of the HPO process [3, 31, 32]. + +Multi-fidelity methods. Multi-fidelity methods exploit low and high fidelities for configurations to save the evaluation time. Successive halving (SHA) [16] runs a set of configurations using a small number of resources and promotes only the best-performing half of configurations to continue for twice as many resources. Hyperband [25] calls SHA as a sub-routine with varying maximum resources and introduces a reduction factor to control the fraction of promotion. ASHA [26] extends SHA to the asynchronous setting by aggressive early-stopping. Later, PASHA [4] further extends ASHA through more aggressive early-stopping based on the ranking of configurations during the tuning process. + +Combination of model-based and multi-fidelity methods. BOHB [10] and a parallel work [47] first propose to combine model-based and multi-fidelity methods by replacing the random sampling in Hyperband with BO. A BOHB [22] employs a joint GP surrogate over fidelities and supports asynchronous scheduling. Hyper-Tune [27] improves its Hyperband by a delayed strategy to decrease inaccurate promotions. Salinas et al. [37] proposed to extend methods to multi-fidelity settings by using the performance of the last fidelity in an ASHA running. DyHPO [48] and DPL [20] introduce new surrogates for multi-fidelity BO considering the learning curves; the former uses deep GP kernels while the latter integrates deep power law functions. + +# 3. Problem Formulation + +Given an algorithm having hyperparameters $\lambda_1, \dots, \lambda_m$ with domains $\Lambda_1, \dots, \Lambda_m$ , we define its hyperparameter space as $\Lambda = \Lambda_1 \times \dots \times \Lambda_m$ . Here, we define the problem and outline the key challenge related to hyperparameter optimization (HPO). Notations are in Supp. 7 for reference. + +Single-fidelity setting. For each hyperparameter configuration $\lambda$ , we denote $f(\lambda)$ as the performance achieved using $\lambda$ . For consistency, the metric in this paper refers to descending metrics like validation loss, with ascending metrics being treated similarly. In the single-fidelity HPO setting, we aim to find $\lambda^*$ minimizing function $f(\lambda)$ , i.e., $\lambda^* = \arg \min_{\lambda \in \Lambda} f(\lambda)$ . BO is one of the most popular single-fidelity HPO methods. The vanilla BO has two key components: a surrogate model $\mathcal{M}$ to approximate the objective function $f(\lambda)$ , and an acquisition function $a$ to identify a promising configuration from search space. With these ingredients, BO iterates three steps: (i) select a configuration $\lambda_i$ by maximizing the acquisition function; (ii) evaluate $\lambda_i$ to get $y_i$ and add the data $(\lambda_i, y_i)$ into the current observation set $\mathcal{D}_{i-1} = \{(\lambda_1, y_1), \dots, (\lambda_{i-1}, y_{i-1})\}$ ; (iii) update the surrogate model and the acquisition function based on the augmented $\mathcal{D}_i$ . In this work, $\mathcal{M}$ is a Gaussian Process and $a$ is Expected Improvement. + +Multi-fidelity setting. Multi-fidelity methods consider resource information, such as training epochs or training subset ratios. Evaluations at various resource levels results in + +different performance levels, known as the fidelities. Different fidelities provide a way to balance computational cost and performance quality. In multi-fidelity HPO problems, the target is extended to $\lambda^{*} = \arg \min_{\lambda \in \Lambda}f(\lambda ,r)$ , where $f(\lambda ,r)$ is the objective function obtained for $\lambda$ at $r$ . We use $r$ to denote the resource level, which can also be interpreted as the fidelity, and $r\in \{r_{min},\dots,r_{max}\}$ . + +Extending single-fidelity methods to the multi-fidelity setting. The inefficiency of single-fidelity methods stems from their reliance on the final fidelity evaluation of $f(\lambda, r_{max})$ to be the evaluation of its objective $f(\lambda, r)$ . Fitting surrogate models by such final fidelity evaluations incurs high cost due to the full evaluation of the configurations. Notably, low-fidelity evaluations at $r < r_{max}$ provide informative insights into the objective but are computationally cheaper, which is valuable to the optimization process. Therefore, we seek an effective way to extend single-fidelity methods like BO to the multi-fidelity setting. More specifically, recalling the earlier steps of BO, when evaluating the configuration $\lambda_i$ in the second step, we instead acquire its low-fidelity performance $y_i^{r_i}$ at $r_i$ , where $r_i$ denotes the fidelity used for $\lambda_i$ to fit the surrogate model. The observations $\mathcal{D}_i$ then become $\{(\lambda_1, y_1^{r_1}), \dots, (\lambda_i, y_i^{r_i})\}$ . To conclude, in order to extend single-fidelity methods to the multi-fidelity setting, the key challenge to be addressed is to determine $r_i$ for each $\lambda_i$ . + +# 4. Methodology + +In this section, we propose a novel multi-fidelity model-based algorithm FastBO. We first propose the key concepts of efficient point and saturation point, which are crucial in deciding the fidelity level to fit the surrogate model and to approximate the final fidelity respectively. Secondly, we elaborate on the details of learning curve modeling, where the two crucial points can be extracted. Then, we present the techniques associated with the auxiliary warm-up and post-processing stages. Finally, we summarize FastBO and discuss its wide applicability to any single-fidelity methods. + +# 4.1. Estimation of Efficient and Saturation Points + +In our method, we adaptively identify efficient and saturation points for each configuration. The two points are crucial in the optimization process. We first formally define the efficient point as follows. + +Definition 1 (Efficient point). For a given learning curve $\mathcal{C}_i(r)$ of hyperparameter configuration $\lambda_{i}$ , where $r$ represents the resource level (also referred to as fidelity), the efficient point $e_i$ of $\lambda_{i}$ is defined as: $e_i = \min \{r\mid \mathcal{C}_i(r) - \mathcal{C}_i(2r) < \delta_1\}$ , where $\delta_{1}$ is a predefined small threshold. + +The semantic of Definition 1 is that starting from the efficient point onwards, when the resources are doubled (i.e., + +from $r$ to $2r$ ), the performance improvement falls below a small threshold $\delta_1$ . Consequently, this point characterizes the fidelity at which a configuration demonstrates strong performance while still efficiently utilizing resources. In simpler terms, it signifies an appropriate fidelity of performance that can be achieved with comparably efficient resource usage. Therefore, we make the following remark. + +Remark 1. The efficient points of the hyperparameter configurations can serve as their appropriate fidelities used for fitting the surrogate model. This is due to their (i) optimal resource-to-performance balance, (ii) ability to capture valuable learning curve trends, and (iii) customization for different hyperparameter configurations. + +We elaborate on the reasons in Remark 1 as follows. Firstly, efficient points balance the trade-off between computational cost and result quality. Beyond the efficient point of a given configuration, allocating additional resources to that configuration becomes less efficient. Secondly, efficient points capture valuable trends within the learning curves. For example, the learning rate influences the shape of learning curves; the identification of efficient points for configurations with smaller learning rates often occurs at later stages. The insights into learning curve behaviors enable more informed decision-making. Thirdly, the ability to customize the fidelity for each specific configuration is a significant advantage. This adaptive approach is more reasonable than previous studies that use a fixed fidelity for all configurations, as it better accounts for the unique characteristics of individual learning curves. + +This insight leads us to use the efficient point $e_i$ identified for each configuration $\lambda_i$ as its fidelity used to fit the surrogate model. Specifically, we evaluate $\lambda_i$ until reaching $e_i$ and obtain the observed performance $y_i^{e_i}$ . The resulting data point $(\lambda_i, y_i^{e_i})$ is then added into the current observation set $\mathcal{D}_{i-1}$ to refit the surrogate model. We proof the superiority of FastBO over SHA-based methods in Supp. 8. + +Besides efficient points, we identify saturation points for all configurations from their learning curves as well. We provide the definition of the saturation point as follows. + +Definition 2 (Saturation point). For a given learning curve $\mathcal{C}_i(r)$ of configuration $\lambda_i$ , where $r$ represents the resource level (also referred to as fidelity), the saturation point $s_i$ of $\lambda_i$ is defined as: $s_i = \min \{r \mid \forall r' > r, |\mathcal{C}_i(r') - \mathcal{C}_i(r)| < \delta_2\}$ , where $\delta_2$ is a predefined small threshold. + +The semantic of Definition 2 is that beyond the saturation point, the observed performance no longer exhibits notable variations with more resources. Thus, this point characterizes the fidelity at which the performance of a configuration stabilizes. The concept of saturation point is well-recognized within the machine learning community. Building on the above definition, we make the following remark. + +Remark 2. The saturation points of the hyperparameter configurations can serve as their approximate final fidelities, as they provide performance results that meet predefined quality thresholds while reducing resource wastage. + +This insight leads us to use the saturation point $s_i$ identified for each configuration $\lambda_i$ as its final fidelity approximation. The point is used in the post-processing stage for promoting some well-performing configurations to get higher-fidelity performances. In essence, when aiming for a full evaluation of the configurations, we suggest that terminating the evaluation at the saturation point is sufficient. A more intuitive illustration of the concepts of efficient and saturation points is provided in Supp. 9. + +# 4.2. Learning Curve Modeling + +From Definitions 1 and 2, we can extract the efficient and saturation points of configurations from their learning curves. The curve $\mathcal{C}_i(r)$ corresponds to configuration $\lambda_{i}$ and describes the predictive performance with $\lambda_{i}$ as a function of the fidelity $r$ . Here, $r$ can be either the number of training instances or the number of training epochs. In the context of learning curves, the former is referred to as observation learning curves, while the latter is iteration learning curves [29]. Both types are applicable to FastBO, so we use the term learning curve to encompass both. Given the observation set $\mathcal{O}_i^w = \{(r,y_i^r)\}_{r = r_{min},\dots,w}$ for $\lambda_{i}$ , which comprises pairs of data points representing fidelities $r\in \{r_{min},\dots,w\}$ and the corresponding evaluations $y_{i}^{r}$ , where $w$ is a pre-defined warm-up point to stop collecting data, FastBO can estimate a learning curve for $\lambda_{i}$ based on $\mathcal{O}_i^w$ by first constructing a parametric learning curve model, then estimating the parameters. + +Constructing a parametric learning curve model. Empirical learning curves can be modeled with function classes relying on some parameters. Viering and Loog [46] comprehensively summarized the parametric models studied in machine learning. In practice, different problems have different learning curves; even under the same problem, different hyperparameter configurations (e.g., learning rate, regularization, etc.) may lead to significantly different learning curves. Since one single parametric model is not enough to characterize all the learning curves by itself, we consider combining different parametric models into a single model. Specifically, we consider three parametric models POW3, EXP3 and LOG2, as listed in Tab. 1, which have shown + +Table 1. Parametric learning curve models used. + +
ModelFormulaFamily
POW3y = d + ax-αPower law
EXP3y = d + e-ax+bExponential
LOG2y = d + a log(x)Logarithmic
+ +good fitting and predicting performance in previous empirical studies [29, 46]. We provide detailed discussions on the choice of parametric models in Supp. 10. + +Here, we denote each parametric model as $c_{j}(r|\theta_{j})$ with parameters $\theta_{j}$ , where the independent variable $r$ represents the fidelity. We combine three models into one model via a weighted linear combination: + +$$ +\mathcal {C} (r | \phi) = \sum_ {j \in \{1, 2, 3 \}} \omega_ {j} c _ {j} (r | \boldsymbol {\theta} _ {j}), \tag {1} +$$ + +where $\phi = \{\omega_1,\omega_2,\omega_3,\theta_1,\theta_2,\theta_3\}$ is the parameter of the combined model, which consists of parameters $\{\pmb {\theta}_1,\pmb {\theta}_2,\pmb {\theta}_3\}$ and weight $\{\omega_{1},\omega_{2},\omega_{3}\}$ of every single model. Therefore, each pair of observations $(r,y_i^r)$ in $O_{i}^{w}$ can be modeled by the combined model as $y_{i}^{r} = \mathcal{C}(r|\phi) + \epsilon$ , where $y_{i}^{r}$ is the observed dependent variable and $\epsilon$ represents the error term. + +Estimating parameters in the parametric learning curve model. We employ maximum likelihood estimation to estimate the parameters $\phi$ in the parametric model $\mathcal{C}(r|\phi)$ . Assuming that $\epsilon \sim \mathcal{N}(0,\sigma^2)$ , the probability of an observed performance $y_{i}^{r}$ under parameters is given by $p(y_i^r |\phi ,\sigma^2) = \mathcal{N}(y_i^r;\mathcal{C}(r|\phi),\sigma^2)$ . Given the observations $\mathcal{O}_i^w$ of $\lambda_{i}$ that contains a set of observed data points $(r,y_i^r)$ , the likelihood function can be expressed as: + +$$ +\begin{array}{l} \mathcal {L} (\boldsymbol {\phi}, \sigma^ {2}; \mathbf {r}, \mathbf {y} _ {i} ^ {r}) = \prod p (y _ {i} ^ {k} | \boldsymbol {\phi}, \sigma^ {2}) \\ = \prod_ {k = r _ {\text {m i n}}} ^ {w} \frac {1}{\sigma \sqrt {2 \pi}} \exp \left(- \frac {\left(y _ {i} ^ {r} - \mathcal {C} (r = k | \phi)\right) ^ {2}}{2 \sigma^ {2}}\right). \tag {2} \\ \end{array} +$$ + +We estimate $\phi$ by maximizing log-likelihood function, which is easily calculated given Eq. 2. + +An existing model-free method [7] also considers using learning curves for the HPO problem. However, it targets predicting the high-fidelity performance from the low-fidelity observations and thus stopping configurations that are unlikely to beat the current best values, which is different from our main target of identifying appropriate fidelity levels for the configurations to fit the surrogate model from their estimated learning curves. + +# 4.3. Warm-up And Post-processing Stages + +In addition to its core components, FastBO incorporates two auxiliary stages: the warm-up and post-processing stages. For the completeness of our method, we provide an overview of these stages, outlining their targets and presenting the key techniques of early-termination detection and saturation-level evaluation that are applied within. + +Warm-up stage. The warm-up stage prepares the early observation set $\mathcal{O}_i^w$ for each configuration $\lambda_{i}$ that is used to estimate its learning curve, as discussed in $\S 4.2$ . Here $w\in (r_{min},r_{max})$ is a pre-determined fidelity, denoted as warm-up point. Specifically, we initiate the evaluation of + +each newly selected $\lambda_{i}$ , proceeding until reaching $w$ . During this process, we record each fidelity $r$ and its evaluation result $y_{i}^{r}$ , forming pairs $(r,y_i^r)$ . Upon reaching $w$ , we pause the evaluation for $\lambda_{i}$ and obtain its early observation set $\mathcal{O}_i^w = \{(r,y_i^r)\}_{r = r_{min},\ldots ,w}$ , and start modeling the learning curve. During the warm-up stage, we monitor the performance changes across every two continuous fidelities. If we detect that the performance of $\lambda_{i}$ has consecutively dropped twice by more than a ratio $\alpha$ , i.e., $(y_{i}^{r - 1} - y_{i}^{r - 2}) > \alpha y_{i}^{r - 2}$ and $(y_i^r -y_i^{r - 1}) > \alpha y_i^{r - 1}$ , we promptly terminate the evaluation for $\lambda_{i}$ at its current fidelity $r$ , because such consecutive performance deterioration indicates $\lambda_{i}$ is unlikely to achieve satisfactory performance. Once terminated, we directly incorporate the current performance $y_{i}^{r}$ of $\lambda_{i}$ into $\mathcal{D}_{i - 1}$ that is used for updating the surrogate model. Thus, further operations like learning curve modeling are discontinued for $\lambda_{i}$ . Moreover, if we observe a single case of performance drop without subsequent occurrences, i.e., $y_{i}^{r - 1} - y_{i}^{r - 2} > \alpha y_{i}^{r - 2}$ and $y_{i}^{r} - y_{i}^{r - 1}\leq \alpha y_{i}^{r - 1}$ , we opt not to include data from fidelity $r - 1$ in $\mathcal{O}_i^w$ . This is to manually filter out potential noise in the data that may adversely affect the fitting of the learning curve. + +Post-processing stage. The post-processing stage aims at two tasks: promoting the well-performing configurations for saturation-level evaluations and identifying the best configuration and its performance. Firstly, FastBO promotes the top- $k$ well-performing configurations and evaluates them to their saturation points to ensure high-quality performance while maintaining efficient resource utilization. We set $k$ to be always less than or equal to the number of parallel workers available, ensuring a manageable overhead of saturation-level evaluations. It is worth noting that the additional time required is factored into the overall time. Secondly, FastBO finds the best configuration along with its performance achieved so far, which is a standard final step in most HPO methods. However, an increase in fidelities does not always result in performance improvement, possibly due to overfitting, resource saturation, or problem complexity. Therefore, we treat the evaluation at each fidelity as an individual task and record all these intermediate evaluation results, which is also a common practice in recent implementations. In this way, FastBO finds the best performance by considering all the results, rather than relying solely on the highest-fidelity performances of the configurations. In the parallel setting, treating each fidelity evaluation as an individual task offers an added benefit due to its finer granularity. More specifically, when a worker is idle, it takes on a new task of evaluating a configuration at a specific fidelity, rather than evaluating an entire configuration. + +# 4.4. FastBO and Generalization + +Algorithm 1 summarizes our proposed FastBO. It takes surrogate model $\mathcal{M}$ , acquisition function $a$ , warm-up point $w$ , + +Algorithm 1: FastBO algorithm +input : $\mathcal{M},a,w,\alpha ,k,\delta_1,\delta_2$ output: $\lambda^{*},y^{*}$ +1 $i\gets 0,D\gets \emptyset$ +2 while not meet the stop criterion do +3 find $\lambda_{i}\gets \arg \max_{\lambda \in \Lambda}a(\lambda ,\mathcal{M}_{i - 1})$ +4 $\mathcal{O}_i^w,t\gets$ warm-up given $w,\alpha$ // cf. $\S 4.3$ +5 if $\mathcal{O}_i^w$ is not empty then +6 fit $\mathcal{C}_i(r)$ to $\mathcal{O}_i^w$ // cf. $\S 4.2$ +7 find $e_i,s_i$ given $\mathcal{C}_i(r),\delta_1,\delta_2$ // cf. $\S 4.1$ +8 $y_{i}^{e_{i}}\gets$ continue evaluating $\lambda_{i}$ to $e_i$ +9 else +10 $e_i\gets t,s_i\gets r_{max}$ +11 $\mathcal{D}_i\gets \mathcal{D}_{i - 1}\cup (\lambda_i,y_i^{e_i})$ +12 refit $\mathcal{M}_i$ to $\mathcal{D}_i$ +13 $i\gets i + 1$ +14 $\lambda^{*},y^{*}\gets$ post-process given s = $\{s_i\} ,k / / c f.\S 4.3$ + +performance decrease ratio $\alpha$ , promotion number $k$ , and thresholds $\delta_{1}$ , $\delta_{2}$ as inputs, and output the best-founded configuration $\lambda^{*}$ and its performance $y^{*}$ . FastBO follows a similar iterative process of model-based methods but replaces the expensive full evaluations with a more intelligent alternative (cf. Lines 4-10). Specifically, each configuration $\lambda_{i}$ first enters a warm-up stage to collect its early observation set $\mathcal{O}_i^w$ and to be detected and terminated if it exhibits consecutive performance deterioration (cf. Line 4). If $\lambda_{i}$ is not terminated, FastBO then estimates a learning curve $\mathcal{C}_i(r)$ for $\lambda_{i}$ based on $\mathcal{O}_i^w$ (cf. Line 6), and thus the efficient point and saturation point of $\lambda_{i}$ can be obtained (cf. Line 7). After that, $\lambda_{i}$ continues to be evaluated until reaching $e_i$ (cf. Line 8); the result is added to the observation set $\mathcal{D}$ (cf. Line 11) that is used for updating $\mathcal{M}$ (cf. Line 12). On the other hand, the poorly-performing configuration will be terminated early at fidelity $t$ with its result being added directly to $\mathcal{D}$ (cf. Lines 10, 11). Finally, the post-processing stage promotes the most promising configurations to their saturation points and finds the best-founded configuration $\lambda^{*}$ and its performance $y^{*}$ (cf. Line 14). + +Generalizing FastBO to single-fidelity methods. The core of FastBO is to tackle the key challenge of deciding an appropriate fidelity for each configuration to fit the surrogate model by adaptively identifying its efficient point. This strategy of using the efficient point performances for surrogate model fitting also provides a simple but effective way to bridge the gap between single- and multi-fidelity methods. While it is primarily described in the context of model-based methods, the strategy can be generalized to various single-fidelity methods. For example, when evaluating configurations within the population for an evolution- + +![](images/f02b00b4e9beacdb5607cfb0cc7957f340ddd17d4c6cf9731807ca0c3e39398f.jpg) +Figure 1. Performance of average validation accuracy on the LCBench benchmark. + +![](images/c6bd55335731ea34e417923de43d229dc73256aba6d6635452af1a8c6aa9e4be.jpg) +Figure 2. Performance of (a) average validation error on NAS-Bench-201 and (b) average validation loss on FCNet. + +ary algorithm-based HPO method, we can similarly evaluate the efficient point performances instead of the final performances of these configurations and integrate the performances in the subsequent processes, such as selection and variation. Relying on the efficient point rather than the final fidelity or all fidelities available simplifies the extension of the single-fidelity methods to the multi-fidelity setting. The rationale behind this adaptive fidelity identification strategy is discussed in Remark 1. We also demonstrate in our experiments the efficacy of this strategy in extending a range of single-fidelity methods to the multi-fidelity setting. + +# 5. Experiments + +We empirically evaluate the performance of FastBO and compare it with the random search baseline (RS) and 9 competitive baselines from 3 related categories, including (i) model-based methods: standard Gaussian Process-based BO [40]; (ii) multi-fidelity methods: ASHA [26], Hyperband [25], PASHA [4]; and (iii) model-based multifidelity methods: A-BOHB [22], A-CQR [37], BOHB [10], DyHPO [48], Hyper-Tune [27]. RS and BO are single-fidelity baselines, while the others are multi-fidelity ones. + +Our experiments are conducted on 10 datasets from 3 popular benchmarks LCBench [53], NAS-Bench-201 [8] and FCNet [21]. Detailed information on the benchmarks is provided in Supp. 13.1. All the experiments are evaluated with four parallel workers and 10 random seeds. We + +allocate $20\%$ total budget for warm-up, i.e., $w = r_{min} + 0.2 \cdot (r_{max} - r_{min})$ . Ratio $\alpha$ is set to 0.1; thresholds $\delta_1$ and $\delta_2$ are set to 0.001 and $0.0005^1$ . We set $k$ based on the number of workers $\# w$ and the number of started configurations $\# c$ : $k = \max \{\lceil \# c / 10 \rceil, \# w\}$ . We provide more experiments and discussions on the hyperparameters in Supp. 12. We use implementations of the baselines in Syne Tune [36]. Details of the baseline settings are in Supp. 13.2. + +# 5.1. Anytime Performance + +To evaluate the anytime performance, we compare FastBO against the baselines on wall-clock time. For fair comparisons, all the baselines, even single-fidelity BO and RS, are extended to consider intermediate results at all the fidelities when identifying the configuration, akin to FastBO as discussed in § 4.3. Consequently, all the baselines are able to achieve their best possible anytime performance. + +The results on LCBench, NAS-Bench-201, and FCNet are shown in Figs. 1 and 2. We report the validation accuracy, error, and loss over wall-clock time for the three benchmarks, as provided by the benchmarks. We provide the results on NAS-Bench-301 [39] in Supp. 11.1. Overall, FastBO can handle various performance metrics and shows strong anytime performance. We can observe that FastBO gains an advantage earlier than other methods, rapidly con + +Table 2. Comparison of relative efficiency on configuration identification. FastBO is set as the baseline with a relative efficiency of 1.00. Wall-clock time (abbr. WC time) reports the elapsed time spent for each method on finding configurations with similar performance metrics, i.e., validation error $(\times 10^{-2})$ for Covertype and ImageNet16-120 and validation loss $(\times 10^{-5})$ for Slice. + +
Metric DatasetMethodFastBOBOPASHAA-BOHBA-CQRBOHBDyHPOHyper-Tune
CovertypeVal. error22.9±0.223.0±0.325.1±2.523.5±1.131.6±1.932.5±0.823.0±0.323.0±0.2
WC time (h)0.7±0.32.9±0.73.9±1.02.0±1.03.9±0.22.5±1.01.7±0.61.8±0.7
Rel. efficiency1.000.250.180.370.190.290.410.40
ImageNet 16-120Val. error55.3±0.257.4±1.255.7±0.355.8±1.655.5±0.955.5±1.155.5±1.055.3±2.0
WC time (h)2.2±0.76.6±0.92.5±1.25.9±1.16.0±1.33.2±0.74.3±1.03.4±1.1
Rel. efficiency1.000.340.900.380.370.680.510.67
SliceVal. loss26.3±2.626.4±4.426.8±9.526.3±6.327.1±4.226.8±5.627.4±2.328.7±1.3
WC time (h)0.4±0.13.1±0.71.2±0.92.1±0.72.5±0.72.2±0.92.5±0.51.8±0.6
Rel. efficiency1.000.130.350.200.170.190.170.24
+ +verging to the global optimum after the initial phase. + +The superiority can be attributed to two main factors. Firstly, FastBO maintains, and in some cases even surpasses, the sample efficiency of vanilla BO, thanks to our techniques that enable quick and precise identification of the fidelities for configurations to update the surrogate model. We provide more explanations and experiments on sample efficiency in Supp. 11.2. Secondly, the multi-fidelity extension speeds up the evaluations, contributing to its overall efficiency. In contrast, the single-fidelity baselines tend to waste more time on the full evaluations. While the multifidelity baselines efficiently explore numerous configurations, they limit their evaluations to only constrained fidelities for some time, thus struggling to provide relatively high performance in a short time. This issue in multi-fidelity methods is particularly pronounced in PASHA when applied to NAS-Bench-201 and FCNet, as shown in Fig. 2. In Supp. 11.3, we further provide the ranks of all methods and statistically show FastBO's superiority on an early stage. It is worth noting that all the additional overhead introduced by FastBO is taken into account in the wall-clock time. + +Regarding the final performance, most methods are able to converge to satisfactory solutions, with negligible differences among them in most cases. Although our goal is not to offer the best final performance as we limit the evaluations to at most the saturation point even for those we consider most promising, FastBO still achieves top-2 final performance on 8 out of 10 datasets. In contrast, model-free methods sometimes cannot obtain a satisfactory final performance because they randomly select the configurations. For example, on the "Covertype" dataset, only 3 out of 2000 configurations yield a validation accuracy exceeding $75\%$ . As a result, all the model-free methods face challenges in converging to a satisfactory final performance. + +# 5.2. Efficiency on Configuration Identification + +One explanation for PASHA's suboptimal anytime performance (cf. Fig. 2) lies in its primary goal [4]: the goal of PASHA is not high accuracy but to identify the best configuration more quickly. To ensure equitable comparisons, we report the time spent for each method on identifying a satisfactory configuration, consistent with the experiments described in PASHA [4]. Results on three expensive datasets "Covertype", "ImageNet16-120", and "Slice" of the three benchmarks are shown in Tab. 2. Similar results on additional datasets can be found in Supp. 11.4. Besides PASHA, results of other model-free multi-fidelity methods are not included, as PASHA demonstrates its superiority over them. + +Tab. 2 shows that FastBO saves $10\%$ to $87\%$ wall-clock time over other methods when achieving up to $9.6\%$ better performance values. It can be observed from the "rel. efficiency" rows, where we set FastBO as the baseline with a relative efficiency of 1.00 and report the efficiency of other methods relative to ours. When compared with vanilla BO, FastBO significantly shortens the time in identifying a good configuration by a factor of 3 to 8, because FastBO pauses a configuration earlier at an appropriate fidelity and fits the surrogate model to guide the next configuration search. This advantage creates opportunities to efficiently explore more configurations. Another observation is that PASHA always gets a relatively high variance in wall-clock time. This is due to the fact that different random seeds can have a larger impact on such model-free methods. + +# 5.3. Effectiveness of Adaptive Fidelity Identification + +As discussed in § 4.1, FastBO is able to adaptively identify the efficient point $e_i$ for each configuration $\lambda_i$ and serves + +![](images/5871895093db6b0070b54a1377b6b9d466bff987012e8951815add0452787815.jpg) +Figure 3. Performance comparison: (a) Performance of FastBO that adaptively sets $r_i = e_i$ with the schemes that use fixed $r_i$ for all configurations. (b) Performance of single-fidelity methods CQR, BORE, REA and their multi-fidelity variants using our extension method. + +$e_i$ as its fidelity $r_i$ for surrogate model fitting. To investigate the effectiveness of the adaptive fidelity identification strategy, we conduct an ablation study to compare the performance achieved with and without applying this strategy. Specifically, we compare FastBO, where $r_i$ is adaptively set to $e_i$ , with the partial evaluation schemes that employ fixed predefined values as the fidelity for all the configurations to fit the surrogate model. We consider three representative fixed fidelities, including $25\%$ , $50\%$ , and $75\%$ of the total resource budget. In addition, we include a comparison with vanilla BO that can be viewed as using $100\%$ resource budget as the fixed fidelity for all configurations. + +We provide the results on three representative datasets in Fig. 3(a), with more results available in Supp. 11.5. We have three main observations. Firstly, FastBO always outperforms the partial evaluation baselines that use a fixed fidelity, indicating the effectiveness of the adaptive strategy. Secondly, FastBO shows stronger performance than vanilla BO. The limitation of vanilla BO lies in the additional time required for full evaluations. Secondly, compared to the vanilla BO, partial evaluation schemes with fixed $r_i$ converge faster in the initial stage due to their ability to evaluate more configurations promptly, but this advantage is gradually offset over time because they fail to find appropriate fidelities to create an accurate surrogate model. This causes a suboptimal final performance compared to vanilla BO, as shown in the first two figures in Figs. 3(a). In the case of the last one, we can observe a noticeable upward trend exhibited by the vanilla BO towards the end of the evaluation, indicating its potential to improve the final performance given abundant time. The comparison between the partial evaluation baselines and vanilla BO also demonstrates the importance of our adaptive strategy, which ensures that the fidelities align optimally with each configuration. + +# 5.4. Generality of The Proposed Extension Method + +The adaptive fidelity identification strategy provides a simple way to extend single-fidelity methods to the multifidelity setting, as discussed in § 4.4. To examine the ability of our extension method, we conduct experiments using three popular single-fidelity methods CQR [37], BORE [45] + +and REA [35], extending them to the multi-fidelity variants with our extension method, referred to as FastCQR, FastBORE, and FastREA respectively. Similar to FastBO, all the multi-fidelity extensions evaluate the configurations to the adaptively identified efficient point and use the corresponding performances for the subsequent operations. The results on three datasets are illustrated in Fig. 3(b) and similar results on other datasets are in Supp. 11.6. We can clearly observe that the multi-fidelity variants with our extension method always outperform their single-fidelity counterparts. It is worth noting that REA is an evolutionary algorithm-based HPO method and is also significantly improved by our extension. The observation highlights the ability of the proposed adaptive strategy to extend any single-fidelity method to the multi-fidelity setting. It also suggests future opportunities to extend other advanced single-fidelity techniques into the multi-fidelity setting. + +# 6. Conclusion + +In this paper, we propose a model-based multi-fidelity HPO method FastBO, which adaptively identifies the appropriate fidelity for each configuration to fit the surrogate model and offers high-quality performance while ensuring efficient resource utilization. The advantages are achieved through our concepts of efficient and saturation point, the proposed techniques of learning curve modeling, and well-designed warm-up and post-processing stages with judicious early-termination detection and efficient saturation-level evaluation. Moreover, the proposed adaptive fidelity identification strategy provides a simple way to extend any single-fidelity method to the multi-fidelity setting. Experiments demonstrate the effectiveness and wide generality of our proposed techniques. FastBO source code is freely available at https://github.com/jjiantong/FastBO. + +# Acknowledgment + +This research was funded by ARC Grant number DP190102443. Ajmal Mian is the recipient of an Australian Research Council Future Fellowship Award (project number FT210100268) funded by the Australian Government. + +# References + +[1] Sebastian E Ament and Carla P Gomes. Scalable first-order Bayesian Optimization via structured automatic differentiation. In International Conference on Machine Learning, pages 500-516. PMLR, 2022. 2 +[2] James Bergstra, Rémi Bardenet, Yoshua Bengio, and Balázs Kégl. Algorithms for hyper-parameter optimization. Advances in Neural Information Processing Systems, 24, 2011. 1, 2 +[3] Bernd Bischl, Martin Binder, Michel Lang, Tobias Pielok, Jakob Richter, Stefan Coors, Janek Thomas, Theresa Ullmann, Marc Becker, Anne-Laure Boulesteix, et al. Hyperparameter optimization: Foundations, algorithms, best practices, and open challenges. Wiley Interdisciplinary Reviews: Data Mining and Knowledge Discovery, 13(2):e1484, 2023. 2 +[4] Ondrej Bohdal, Lukas Balles, Martin Wistuba, Beyza Ermis, Cedric Archambeau, and Giovanni Zappella. PASHA: efficient HPO and NAS with progressive resource allocation. In International Conference on Learning Representations. OpenReview.net, 2023. 1, 2, 6, 7, 4, 9 +[5] Chaofan Chen, Oscar Li, Daniel Tao, Alina Barnett, Cynthia Rudin, and Jonathan K Su. This looks like that: deep learning for interpretable image recognition. Advances in Neural Information Processing Systems, 32, 2019. 2 +[6] Tianqi Chen and Carlos Guestrin. Xgboost: A scalable tree boosting system. In Proceedings of the 22nd acm sigkdd international conference on knowledge discovery and data mining, pages 785-794, 2016. 9 +[7] Tobias Domhan, Jost Tobias Springenberg, and Frank Hutter. Speeding up automatic hyperparameter optimization of deep neural networks by extrapolation of learning curves. In International Joint Conference on Artificial Intelligence, 2015. 4 +[8] Xuanyi Dong and Yi Yang. NAS-Bench-201: Extending the scope of reproducible neural architecture search. In International Conference on Learning Representations, 2020. 1, 6, 8 +[9] Thomas Elsken, Jan Hendrik Metzen, and Frank Hutter. Neural architecture search: A survey. The Journal of Machine Learning Research, 20(1):1997-2017, 2019. 1 +[10] Stefan Falkner, Aaron Klein, and Frank Hutter. BOHB: Robust and efficient hyperparameter optimization at scale. In International Conference on Machine Learning, pages 1437-1446. PMLR, 2018. 1, 2, 6, 9 +[11] Matthias Feurer and Frank Hutter. Hyperparameter optimization. Automated Machine Learning: Methods, Systems, Challenges, pages 3-33, 2019. 1 +[12] Peter I Frazier, Warren B Powell, and Savas Dayanik. A knowledge-gradient policy for sequential information collection. SIAM Journal on Control and Optimization, 47(5): 2410-2439, 2008. 2 +[13] José Miguel Hernández-Lobato, Matthew W Hoffman, and Zoubin Ghahramani. Predictive entropy search for efficient global optimization of black-box functions. Advances in Neural Information Processing systems, 27, 2014. 2 + +[14] Frank Hutter, Holger H Hoos, and Kevin Leyton-Brown. Sequential model-based optimization for general algorithm configuration. In Learning and Intelligent Optimization, pages 507-523. Springer, 2011. 1, 2 +[15] Carl Hvarfner, Danny Stoll, Artur L. F. Souza, Marius Lindauer, Frank Hutter, and Luigi Nardi. $\S$ pi $\S$ BO: Augmenting acquisition functions with user beliefs for bayesian optimization. In International Conference on Learning Representations. OpenReview.net, 2022. 2 +[16] Kevin Jamieson and Ameet Talwalkar. Non-stochastic best arm identification and hyperparameter optimization. In Artificial Intelligence and Statistics, pages 240-248. PMLR, 2016. 1, 2, 9 +[17] Kasra Jamshidi, Harry Xu, and Keval Vora. Accelerating graph mining systems with subgraph morphing. In European Conference on Computer Systems, pages 162-181, 2023. 2 +[18] Jiantong Jiang, Zeyi Wen, Zeke Wang, Bingsheng He, and Jian Chen. Parallel and distributed structuredsvm training. IEEE Transactions on Parallel and Distributed Systems, 33 (5):1084-1096, 2021. +[19] Jiantong Jiang, Zeyi Wen, and Ajmal Mian. Fast parallel bayesian network structure learning. In IEEE International Parallel and Distributed Processing Symposium, pages 617-627. IEEE, 2022. 2 +[20] Arlind Kadra, Maciej Janowski, Martin Wistuba, and Josif Grabocka. Scaling laws for hyperparameter optimization. In Advances in Neural Information Processing Systems, 2023. 2 +[21] Aaron Klein and Frank Hutter. Tabular benchmarks for joint architecture and hyperparameter optimization. arXiv preprint arXiv:1905.04970, 2019. 6, 8 +[22] Aaron Klein, Louis C Tiao, Thibaut Lienart, Cedric Archambeau, and Matthias Seeger. Model-based asynchronous hyperparameter and neural architecture search. arXiv preprint arXiv:2003.10865, 2020. 1, 2, 6, 9 +[23] Prasanth Kolachina, Nicola Cancedda, Marc Dymetman, and Sriram Venkatapathy. Prediction of learning curves in machine translation. In Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 22-30, 2012. 2 +[24] Cheng Li, Santu Rana, Sunil Gupta, Vu Nguyen, Svetha Venkatesh, Alessandra Sutti, David Rubin de Celis Leal, Teo Slezak, Murray Height, Mazher Mohammed, and Ian Gibson. Accelerating experimental design by incorporating experimenter hunches. In International Conference on Data Mining, pages 257-266. IEEE Computer Society, 2018. 2 +[25] Lisha Li, Kevin Jamieson, Giulia DeSalvo, Afshin Rostamizadeh, and Ameet Talwalkar. Hyperband: A novel bandit-based approach to hyperparameter optimization. Journal of Machine Learning Research, 18(1):6765-6816, 2017. 1, 2, 6, 9 +[26] Liam Li, Kevin Jamieson, Afshin Rostamizadeh, Ekaterina Gonina, Jonathan Ben-Tzur, Moritz Hardt, Benjamin Recht, and Ameet Talwalkar. A system for massively parallel hyperparameter tuning. Proceedings of Machine Learning and Systems, 2:230-246, 2020. 1, 2, 6, 9 +[27] Yang Li, Yu Shen, Huaijun Jiang, Wentao Zhang, Jixiang Li, Ji Liu, Ce Zhang, and Bin Cui. Hyper-tune: Towards + +efficient hyper-parameter tuning at scale. Proceedings of the VLDB Endowment, 15(6):1256-1265, 2022. 1, 2, 6, 9 +[28] Jonas Mockus. The application of Bayesian methods for seeking the extremum. Towards global optimization, 2:117, 1998. 2 +[29] Felix Mohr and Jan N van Rijn. Learning curves for decision making in supervised machine learning-a survey. arXiv preprint arXiv:2201.12150, 2022. 4, 2 +[30] Felix Mohr, Tom J Viering, Marco Loog, and Jan N van Rijn. Lcdb 1.0: An extensive learning curves database for classification tasks. In Joint European Conference on Machine Learning and Knowledge Discovery in Databases, pages 3-19. Springer, 2022. 2 +[31] Julia Moosbauer, Julia Herbinger, Giuseppe Casalicchio, Marius Lindauer, and Bernd Bischl. Explaining hyperparameter optimization via partial dependence plots. Advances in Neural Information Processing Systems, 34:2280-2291, 2021. 2 +[32] Julia Moosbauer, Giuseppe Casalicchio, Marius Lindauer, and Bernd Bischl. Improving accuracy of interpretability measures in hyperparameter optimization via Bayesian algorithm execution. arXiv preprint arXiv:2206.05447, 2022. 2 +[33] ChangYong Oh, Efstratios Gavves, and Max Welling. BOCK: Bayesian optimization with cylindrical kernels. In International Conference on Machine Learning, pages 3868-3877. PMLR, 2018. 2 +[34] Misha Padidar, Xinran Zhu, Leo Huang, Jacob Gardner, and David Bindel. Scaling gaussian processes with derivative information using variational inference. Advances in Neural Information Processing Systems, 34:6442-6453, 2021. 2 +[35] Esteban Real, Alok Aggarwal, Yanping Huang, and Quoc V Le. Regularized evolution for image classifier architecture search. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 4780-4789, 2019. 8, 5, 9 +[36] David Salinas, Matthias Seeger, Aaron Klein, Valerio Perrone, Martin Wistuba, and Cedric Archambeau. Syne tune: A library for large scale hyperparameter tuning and reproducible research. In International Conference on Automated Machine Learning, pages 16-1. PMLR, 2022. 6, 9 +[37] David Salinas, Jacek Golebiowski, Aaron Klein, Matthias W. Seeger, and Cedric Archambeau. Optimizing hyperparameters with conformal quantile regression. In International Conference on Machine Learning, pages 29876-29893. PMLR, 2023. 1, 2, 6, 8, 5, 9 +[38] Bobak Shahriari, Alexandre Bouchard-Côté, and Nando Freitas. Unbounded Bayesian Optimization via regularization. In Artificial intelligence and statistics, pages 1168-1176. PMLR, 2016. 2 +[39] Julien Siems, Lucas Zimmer, Arber Zela, Jovita Lukasik, Margret Keuper, and Frank Hutter. Nas-bench-301 and the case for surrogate benchmarks for neural architecture search. CoRR, abs/2008.09777, 2020. 1, 6, 3 +[40] Jasper Snoek, Hugo Larochelle, and Ryan P Adams. Practical Bayesian optimization of machine learning algorithms. Advances in Neural Information Processing Systems, 25, 2012. 1, 2, 6, 9 + +[41] Jasper Snoek, Oren Rippel, Kevin Swersky, Ryan Kiros, Nadathur Satish, Narayanan Sundaram, Mostofa Patwary, Mr Prabhat, and Ryan Adams. Scalable Bayesian optimization using deep neural networks. In International Conference on Machine Learning, pages 2171-2180. PMLR, 2015. 2 +[42] Jost Tobias Springenberg, Aaron Klein, Stefan Falkner, and Frank Hutter. Bayesian optimization with robust bayesian neural networks. Advances in Neural Information Processing Systems, 29, 2016. 2 +[43] Niranjan Srinivas, Andreas Krause, Sham M Kakade, and Matthias Seeger. Gaussian process optimization in the bandit setting: No regret and experimental design. arXiv preprint arXiv:0912.3995, 2009. 2 +[44] Kevin Swersky, Jasper Snoek, and Ryan Prescott Adams. Freeze-thaw Bayesian optimization. arXiv preprint arXiv:1406.3896, 2014. 1 +[45] Louis C Tiao, Aaron Klein, Matthias W Seeger, Edwin V Bonilla, Cedric Archambeau, and Fabio Ramos. BORE: Bayesian optimization by density-ratio estimation. In International Conference on Machine Learning, pages 10289-10300. PMLR, 2021. 8, 5, 9 +[46] Tom Viering and Marco Loog. The shape of learning curves: a review. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2022. 1, 4, 2 +[47] Jiazhuo Wang, Jason Xu, and Xuejun Wang. Combination of hyperband and bayesian optimization for hyperparameter optimization in deep learning. arXiv preprint arXiv:1801.01596, 2018. 1, 2 +[48] Martin Wistuba, Arlind Kadra, and Josif Grabocka. Supervising the multi-fidelity race of hyperparameter configurations. Advances in Neural Information Processing Systems, 35:13470-13484, 2022. 2, 6, 9 +[49] Jian Wu, Matthias Poloczek, Andrew G Wilson, and Peter Frazier. Bayesian optimization with gradients. Advances in neural information processing systems, 30, 2017. 2 +[50] Peiyu Yang, Naveed Akhtar, Zeyi Wen, Mubarak Shah, and Ajmal Mian. Re-calibrating feature attributions for model interpretation. In International Conference on Learning Representations, 2022. 2 +[51] Peiyu Yang, Naveed Akhtar, Zeyi Wen, and Ajmal Mian. Local path integration for attribution. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 3173-3180, 2023. 2 +[52] Xueying Zhu, Jie Sun, Zhenhao He, Jiantong Jiang, and Zeke Wang. Staleness-reduction mini-batch $k$ -means. IEEE Transactions on Neural Networks and Learning Systems, 2023. 2 +[53] Lucas Zimmer, Marius Thomas Lindauer, and Frank Hutter. Auto-Pytorch: Multi-fidelity metalearning for efficient and robust AutoDL. IEEE Transactions on Pattern Analysis and Machine Intelligence, 43:3079-3090, 2021. 6, 8 \ No newline at end of file diff --git a/2024/Efficient Hyperparameter Optimization with Adaptive Fidelity Identification/images.zip b/2024/Efficient Hyperparameter Optimization with Adaptive Fidelity Identification/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..94e2c717d32c9e3ea42e6978a8bc5cc267debf83 --- /dev/null +++ b/2024/Efficient Hyperparameter Optimization with Adaptive Fidelity Identification/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6bc9bef7f859357bff949a7c685f4ed10cac37e2e37033f536b545b3382bb175 +size 372194 diff --git a/2024/Efficient Hyperparameter Optimization with Adaptive Fidelity Identification/layout.json b/2024/Efficient Hyperparameter Optimization with Adaptive Fidelity Identification/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..07ba337a3e36f58af0638794cd2f762b57f75de9 --- /dev/null +++ b/2024/Efficient Hyperparameter Optimization with Adaptive Fidelity Identification/layout.json @@ -0,0 +1,10077 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 60, + 103, + 533, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 60, + 103, + 533, + 121 + ], + "spans": [ + { + "bbox": [ + 60, + 103, + 533, + 121 + ], + "type": "text", + "content": "Efficient Hyperparameter Optimization with Adaptive Fidelity Identification" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 122, + 142, + 471, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 122, + 142, + 471, + 171 + ], + "spans": [ + { + "bbox": [ + 122, + 142, + 471, + 171 + ], + "type": "text", + "content": "Jiantong Jiang" + }, + { + "bbox": [ + 122, + 142, + 471, + 171 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 122, + 142, + 471, + 171 + ], + "type": "text", + "content": ", Zeyi Wen" + }, + { + "bbox": [ + 122, + 142, + 471, + 171 + ], + "type": "inline_equation", + "content": "^{2,3*}" + }, + { + "bbox": [ + 122, + 142, + 471, + 171 + ], + "type": "text", + "content": ", Atif Mansoor" + }, + { + "bbox": [ + 122, + 142, + 471, + 171 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 122, + 142, + 471, + 171 + ], + "type": "text", + "content": ", Ajmal Mian" + }, + { + "bbox": [ + 122, + 142, + 471, + 171 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 122, + 142, + 471, + 171 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 122, + 142, + 471, + 171 + ], + "type": "text", + "content": "The University of Western Australia, " + }, + { + "bbox": [ + 122, + 142, + 471, + 171 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 122, + 142, + 471, + 171 + ], + "type": "text", + "content": "HKUST (Guangzhou), " + }, + { + "bbox": [ + 122, + 142, + 471, + 171 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 122, + 142, + 471, + 171 + ], + "type": "text", + "content": "HKUST" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 174, + 535, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 174, + 535, + 186 + ], + "spans": [ + { + "bbox": [ + 55, + 174, + 535, + 186 + ], + "type": "text", + "content": "jiantong.jiang@research.uwa.edu.au, wenzeyi@ust.hk, {atif.mansoor, ajmal.mian}@uwa.edu.au" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "spans": [ + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 238, + 290, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 238, + 290, + 491 + ], + "spans": [ + { + "bbox": [ + 46, + 238, + 290, + 491 + ], + "type": "text", + "content": "Hyperparameter Optimization and Neural Architecture Search are powerful in attaining state-of-the-art machine learning models, with Bayesian Optimization (BO) standing out as a mainstream method. Extending BO into the multi-fidelity setting has been an emerging research topic in this field, but faces the challenge of determining an appropriate fidelity for each hyperparameter configuration to fit the surrogate model. To tackle the challenge, we propose a multi-fidelity BO method named FastBO, which excels in adaptively deciding the fidelity for each configuration and providing strong performance while ensuring efficient resource usage. These advantages are achieved through our proposed techniques based on the concepts of efficient point and saturation point for each configuration, which can be obtained from the empirical learning curve of the configuration, estimated from early observations. Extensive experiments demonstrate FastBO's superior anytime performance and efficiency in identifying high-quality configurations and architectures. We also show that our method provides a way to extend any single-fidelity method to the multi-fidelity setting, highlighting the wide applicability of our approach." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 518, + 128, + 532 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 518, + 128, + 532 + ], + "spans": [ + { + "bbox": [ + 47, + 518, + 128, + 532 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 539, + 287, + 670 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 539, + 287, + 670 + ], + "spans": [ + { + "bbox": [ + 46, + 539, + 287, + 670 + ], + "type": "text", + "content": "Hyperparameters are crucial in machine learning pipelines. Hyperparameter optimization (HPO) [11] and Neural Architecture Search (NAS) [9] aims to find the hyperparameters or architectures that can yield good performance without human experts. Among different HPO and NAS methods, Bayesian Optimization (BO) [2, 14, 40] is an effective model-based method that has shown remarkable success [8, 39]. BO maintains a surrogate model of the target performance metric based on past evaluations of hyperparameter configurations, which guides the choice of more promising configurations to evaluate." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 671, + 287, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 671, + 287, + 696 + ], + "spans": [ + { + "bbox": [ + 47, + 671, + 287, + 696 + ], + "type": "text", + "content": "Despite its sample efficiency, standard BO requires a full evaluation of each configuration, involving full-scale" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 213, + 547, + 393 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 213, + 547, + 393 + ], + "spans": [ + { + "bbox": [ + 304, + 213, + 547, + 393 + ], + "type": "text", + "content": "training and testing of models, which can be highly time-consuming, particularly with the recent trend to larger models. To avoid expensive full evaluations, multi-fidelity methods [4, 16, 25, 26] have been proposed, where the fidelity refers to the levels of performance metrics obtained under different resource levels. These methods follow the principle of successive halving (SHA) [16]: initially, they evaluate a set of randomly selected configurations using a small number of resources; then, based on the low-fidelity performances, the poorly-performing ones are successively eliminated, while the well-performing ones continue to be evaluated with increasing resources. Follow-up studies [10, 22, 27, 37, 47] propose model-based multi-fidelity methods, replacing the random configuration selection with a more informed model to improve sample efficiency." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 398, + 547, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 398, + 547, + 662 + ], + "spans": [ + { + "bbox": [ + 304, + 398, + 547, + 662 + ], + "type": "text", + "content": "Nevertheless, current model-based multi-fidelity methods face a major limitation: they are built upon the SHA framework, which operates under the assumption that learning curves of different configurations rarely intersect. This assumption does not hold in practice [46], i.e., early performance observations cannot always indicate the final fidelity performance at the full resource level. This leads to a fundamental challenge when extending model-based methods to the multi-fidelity setting: What is the appropriate fidelity for each configuration to fit the surrogate model? In other words, which fidelity can provide performance observations that reliably indicate the final fidelity performance? Existing methods struggle to address this fundamental challenge. In particular, BOHB [10] and Hyper-Tune [27] fit separate surrogate models for different fidelities, failing to capture inter-fidelity correlations. FTBO [44] and A-BOHB [22] fit a joint model but require strong assumptions to remain tractable. Another work by Salinas et al. [37] suggests using the last observed fidelity performance to fit the surrogate model. However, it widens the gap between poorly- and well-performing configurations at the early stage, potentially leading to an inaccurate surrogate model." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 665, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 547, + 715 + ], + "type": "text", + "content": "To this end, we propose a multi-fidelity extension of BO, namely FastBO, which tackles the challenge of deciding the appropriate fidelity for each configuration to fit the surrogate model. FastBO identifies a so-called efficient point for" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 702, + 186, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 702, + 186, + 713 + ], + "spans": [ + { + "bbox": [ + 58, + 702, + 186, + 713 + ], + "type": "text", + "content": "*Zeyi Wen is the corresponding author." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "26181" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 288 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 288 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 288 + ], + "type": "text", + "content": "each configuration to be the fidelity. The point balances computational cost and performance quality while capturing valuable learning curve trends. In essence, FastBO selects the fidelity for each configuration instead of evaluating all the configurations at the same fidelity. Additionally, a saturation point for each configuration is identified to be an approximation of the final fidelity, leading to high-quality performance while reducing resource wastage. The two crucial points are adaptively derived from the estimated learning curve of each configuration. Furthermore, the warm-up and post-processing stages are carefully designed to enable judicious early-termination detection and efficient saturation level evaluation. Empirical evaluation against the state-of-the-art methods shows that FastBO has strong anytime performance and can considerably save up to " + }, + { + "bbox": [ + 46, + 72, + 289, + 288 + ], + "type": "inline_equation", + "content": "87\\%" + }, + { + "bbox": [ + 46, + 72, + 289, + 288 + ], + "type": "text", + "content": " of the time required to identify a good configuration or architecture, lowering the barriers for engaging in HPO and NAS. In summary, we make the following major contributions." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 291, + 287, + 434 + ], + "type": "list", + "angle": 0, + "index": 4, + "blocks": [ + { + "bbox": [ + 47, + 291, + 287, + 339 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 291, + 287, + 339 + ], + "spans": [ + { + "bbox": [ + 47, + 291, + 287, + 339 + ], + "type": "text", + "content": "1. We propose a multi-fidelity model-based HPO method that adaptively decides the fidelities for configurations and efficiently offers strong performance, thanks to the introduced concepts of efficient and saturation points." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 339, + 287, + 386 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 339, + 287, + 386 + ], + "spans": [ + { + "bbox": [ + 47, + 339, + 287, + 386 + ], + "type": "text", + "content": "2. We develop a learning curve modeling module to enable adaptive derivation of the key points, a warm-up stage to allow early-termination detection, and a post-processing stage to ensure efficient saturation-level evaluation." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 386, + 287, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 386, + 287, + 434 + ], + "spans": [ + { + "bbox": [ + 47, + 386, + 287, + 434 + ], + "type": "text", + "content": "3. We show that our strategy can be used to extend existing single-fidelity methods to the multi-fidelity setting, demonstrating the effectiveness and generality of our method and highlighting promising future opportunities." + } + ] + } + ], + "index": 3 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 47, + 442, + 134, + 455 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 442, + 134, + 455 + ], + "spans": [ + { + "bbox": [ + 47, + 442, + 134, + 455 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 462, + 287, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 462, + 287, + 545 + ], + "spans": [ + { + "bbox": [ + 46, + 462, + 287, + 545 + ], + "type": "text", + "content": "HPO and NAS incur significant costs, especially considering the escalating model evaluation overhead. Despite efforts to directly accelerate computation [17-19, 52], smarter HPO strategies are still required for the widespread adoption of automation. Two crucial directions include model-based and multi-fidelity methods, which can also be combined. Here, we review the methods in these categories." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 545, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 545, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 545, + 288, + 714 + ], + "type": "text", + "content": "Model-based methods. BO is the representative of model-based methods. Based on the surrogate model constructed by historical evaluation results, BO selects the configurations to evaluate via an acquisition function that balances exploration and exploitation. Commonly used surrogate models are Gaussian processes [40], random forests [14], tree-structured Parzen estimator [2], and deep networks [41, 42]. Popular acquisition functions include Expected Improvement [28], Knowledge Gradient [12], Upper Confidence Bound [43], and Predictive Entropy Search [13]. Recent studies on BO have explored the utilization of expert priors [15, 24, 33, 38] and derivative information [1, 34, 49]. There also has been a recent focus on enhancing the interpretability [5, 50, 51] of the HPO process [3, 31, 32]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 72, + 545, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 216 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 216 + ], + "type": "text", + "content": "Multi-fidelity methods. Multi-fidelity methods exploit low and high fidelities for configurations to save the evaluation time. Successive halving (SHA) [16] runs a set of configurations using a small number of resources and promotes only the best-performing half of configurations to continue for twice as many resources. Hyperband [25] calls SHA as a sub-routine with varying maximum resources and introduces a reduction factor to control the fraction of promotion. ASHA [26] extends SHA to the asynchronous setting by aggressive early-stopping. Later, PASHA [4] further extends ASHA through more aggressive early-stopping based on the ranking of configurations during the tuning process." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 216, + 545, + 372 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 216, + 545, + 372 + ], + "spans": [ + { + "bbox": [ + 304, + 216, + 545, + 372 + ], + "type": "text", + "content": "Combination of model-based and multi-fidelity methods. BOHB [10] and a parallel work [47] first propose to combine model-based and multi-fidelity methods by replacing the random sampling in Hyperband with BO. A BOHB [22] employs a joint GP surrogate over fidelities and supports asynchronous scheduling. Hyper-Tune [27] improves its Hyperband by a delayed strategy to decrease inaccurate promotions. Salinas et al. [37] proposed to extend methods to multi-fidelity settings by using the performance of the last fidelity in an ASHA running. DyHPO [48] and DPL [20] introduce new surrogates for multi-fidelity BO considering the learning curves; the former uses deep GP kernels while the latter integrates deep power law functions." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 382, + 432, + 395 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 382, + 432, + 395 + ], + "spans": [ + { + "bbox": [ + 305, + 382, + 432, + 395 + ], + "type": "text", + "content": "3. Problem Formulation" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 402, + 545, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 402, + 545, + 462 + ], + "spans": [ + { + "bbox": [ + 304, + 402, + 545, + 462 + ], + "type": "text", + "content": "Given an algorithm having hyperparameters " + }, + { + "bbox": [ + 304, + 402, + 545, + 462 + ], + "type": "inline_equation", + "content": "\\lambda_1, \\dots, \\lambda_m" + }, + { + "bbox": [ + 304, + 402, + 545, + 462 + ], + "type": "text", + "content": " with domains " + }, + { + "bbox": [ + 304, + 402, + 545, + 462 + ], + "type": "inline_equation", + "content": "\\Lambda_1, \\dots, \\Lambda_m" + }, + { + "bbox": [ + 304, + 402, + 545, + 462 + ], + "type": "text", + "content": ", we define its hyperparameter space as " + }, + { + "bbox": [ + 304, + 402, + 545, + 462 + ], + "type": "inline_equation", + "content": "\\Lambda = \\Lambda_1 \\times \\dots \\times \\Lambda_m" + }, + { + "bbox": [ + 304, + 402, + 545, + 462 + ], + "type": "text", + "content": ". Here, we define the problem and outline the key challenge related to hyperparameter optimization (HPO). Notations are in Supp. 7 for reference." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 462, + 545, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 462, + 545, + 676 + ], + "spans": [ + { + "bbox": [ + 304, + 462, + 545, + 676 + ], + "type": "text", + "content": "Single-fidelity setting. For each hyperparameter configuration " + }, + { + "bbox": [ + 304, + 462, + 545, + 676 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 304, + 462, + 545, + 676 + ], + "type": "text", + "content": ", we denote " + }, + { + "bbox": [ + 304, + 462, + 545, + 676 + ], + "type": "inline_equation", + "content": "f(\\lambda)" + }, + { + "bbox": [ + 304, + 462, + 545, + 676 + ], + "type": "text", + "content": " as the performance achieved using " + }, + { + "bbox": [ + 304, + 462, + 545, + 676 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 304, + 462, + 545, + 676 + ], + "type": "text", + "content": ". For consistency, the metric in this paper refers to descending metrics like validation loss, with ascending metrics being treated similarly. In the single-fidelity HPO setting, we aim to find " + }, + { + "bbox": [ + 304, + 462, + 545, + 676 + ], + "type": "inline_equation", + "content": "\\lambda^*" + }, + { + "bbox": [ + 304, + 462, + 545, + 676 + ], + "type": "text", + "content": " minimizing function " + }, + { + "bbox": [ + 304, + 462, + 545, + 676 + ], + "type": "inline_equation", + "content": "f(\\lambda)" + }, + { + "bbox": [ + 304, + 462, + 545, + 676 + ], + "type": "text", + "content": ", i.e., " + }, + { + "bbox": [ + 304, + 462, + 545, + 676 + ], + "type": "inline_equation", + "content": "\\lambda^* = \\arg \\min_{\\lambda \\in \\Lambda} f(\\lambda)" + }, + { + "bbox": [ + 304, + 462, + 545, + 676 + ], + "type": "text", + "content": ". BO is one of the most popular single-fidelity HPO methods. The vanilla BO has two key components: a surrogate model " + }, + { + "bbox": [ + 304, + 462, + 545, + 676 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 304, + 462, + 545, + 676 + ], + "type": "text", + "content": " to approximate the objective function " + }, + { + "bbox": [ + 304, + 462, + 545, + 676 + ], + "type": "inline_equation", + "content": "f(\\lambda)" + }, + { + "bbox": [ + 304, + 462, + 545, + 676 + ], + "type": "text", + "content": ", and an acquisition function " + }, + { + "bbox": [ + 304, + 462, + 545, + 676 + ], + "type": "inline_equation", + "content": "a" + }, + { + "bbox": [ + 304, + 462, + 545, + 676 + ], + "type": "text", + "content": " to identify a promising configuration from search space. With these ingredients, BO iterates three steps: (i) select a configuration " + }, + { + "bbox": [ + 304, + 462, + 545, + 676 + ], + "type": "inline_equation", + "content": "\\lambda_i" + }, + { + "bbox": [ + 304, + 462, + 545, + 676 + ], + "type": "text", + "content": " by maximizing the acquisition function; (ii) evaluate " + }, + { + "bbox": [ + 304, + 462, + 545, + 676 + ], + "type": "inline_equation", + "content": "\\lambda_i" + }, + { + "bbox": [ + 304, + 462, + 545, + 676 + ], + "type": "text", + "content": " to get " + }, + { + "bbox": [ + 304, + 462, + 545, + 676 + ], + "type": "inline_equation", + "content": "y_i" + }, + { + "bbox": [ + 304, + 462, + 545, + 676 + ], + "type": "text", + "content": " and add the data " + }, + { + "bbox": [ + 304, + 462, + 545, + 676 + ], + "type": "inline_equation", + "content": "(\\lambda_i, y_i)" + }, + { + "bbox": [ + 304, + 462, + 545, + 676 + ], + "type": "text", + "content": " into the current observation set " + }, + { + "bbox": [ + 304, + 462, + 545, + 676 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{i-1} = \\{(\\lambda_1, y_1), \\dots, (\\lambda_{i-1}, y_{i-1})\\}" + }, + { + "bbox": [ + 304, + 462, + 545, + 676 + ], + "type": "text", + "content": "; (iii) update the surrogate model and the acquisition function based on the augmented " + }, + { + "bbox": [ + 304, + 462, + 545, + 676 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_i" + }, + { + "bbox": [ + 304, + 462, + 545, + 676 + ], + "type": "text", + "content": ". In this work, " + }, + { + "bbox": [ + 304, + 462, + 545, + 676 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 304, + 462, + 545, + 676 + ], + "type": "text", + "content": " is a Gaussian Process and " + }, + { + "bbox": [ + 304, + 462, + 545, + 676 + ], + "type": "inline_equation", + "content": "a" + }, + { + "bbox": [ + 304, + 462, + 545, + 676 + ], + "type": "text", + "content": " is Expected Improvement." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "content": "Multi-fidelity setting. Multi-fidelity methods consider resource information, such as training epochs or training subset ratios. Evaluations at various resource levels results in" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "26182" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 156 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 156 + ], + "type": "text", + "content": "different performance levels, known as the fidelities. Different fidelities provide a way to balance computational cost and performance quality. In multi-fidelity HPO problems, the target is extended to " + }, + { + "bbox": [ + 46, + 72, + 287, + 156 + ], + "type": "inline_equation", + "content": "\\lambda^{*} = \\arg \\min_{\\lambda \\in \\Lambda}f(\\lambda ,r)" + }, + { + "bbox": [ + 46, + 72, + 287, + 156 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 72, + 287, + 156 + ], + "type": "inline_equation", + "content": "f(\\lambda ,r)" + }, + { + "bbox": [ + 46, + 72, + 287, + 156 + ], + "type": "text", + "content": " is the objective function obtained for " + }, + { + "bbox": [ + 46, + 72, + 287, + 156 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 46, + 72, + 287, + 156 + ], + "type": "text", + "content": " at " + }, + { + "bbox": [ + 46, + 72, + 287, + 156 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 46, + 72, + 287, + 156 + ], + "type": "text", + "content": ". We use " + }, + { + "bbox": [ + 46, + 72, + 287, + 156 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 46, + 72, + 287, + 156 + ], + "type": "text", + "content": " to denote the resource level, which can also be interpreted as the fidelity, and " + }, + { + "bbox": [ + 46, + 72, + 287, + 156 + ], + "type": "inline_equation", + "content": "r\\in \\{r_{min},\\dots,r_{max}\\}" + }, + { + "bbox": [ + 46, + 72, + 287, + 156 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 157, + 289, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 157, + 289, + 384 + ], + "spans": [ + { + "bbox": [ + 46, + 157, + 289, + 384 + ], + "type": "text", + "content": "Extending single-fidelity methods to the multi-fidelity setting. The inefficiency of single-fidelity methods stems from their reliance on the final fidelity evaluation of " + }, + { + "bbox": [ + 46, + 157, + 289, + 384 + ], + "type": "inline_equation", + "content": "f(\\lambda, r_{max})" + }, + { + "bbox": [ + 46, + 157, + 289, + 384 + ], + "type": "text", + "content": " to be the evaluation of its objective " + }, + { + "bbox": [ + 46, + 157, + 289, + 384 + ], + "type": "inline_equation", + "content": "f(\\lambda, r)" + }, + { + "bbox": [ + 46, + 157, + 289, + 384 + ], + "type": "text", + "content": ". Fitting surrogate models by such final fidelity evaluations incurs high cost due to the full evaluation of the configurations. Notably, low-fidelity evaluations at " + }, + { + "bbox": [ + 46, + 157, + 289, + 384 + ], + "type": "inline_equation", + "content": "r < r_{max}" + }, + { + "bbox": [ + 46, + 157, + 289, + 384 + ], + "type": "text", + "content": " provide informative insights into the objective but are computationally cheaper, which is valuable to the optimization process. Therefore, we seek an effective way to extend single-fidelity methods like BO to the multi-fidelity setting. More specifically, recalling the earlier steps of BO, when evaluating the configuration " + }, + { + "bbox": [ + 46, + 157, + 289, + 384 + ], + "type": "inline_equation", + "content": "\\lambda_i" + }, + { + "bbox": [ + 46, + 157, + 289, + 384 + ], + "type": "text", + "content": " in the second step, we instead acquire its low-fidelity performance " + }, + { + "bbox": [ + 46, + 157, + 289, + 384 + ], + "type": "inline_equation", + "content": "y_i^{r_i}" + }, + { + "bbox": [ + 46, + 157, + 289, + 384 + ], + "type": "text", + "content": " at " + }, + { + "bbox": [ + 46, + 157, + 289, + 384 + ], + "type": "inline_equation", + "content": "r_i" + }, + { + "bbox": [ + 46, + 157, + 289, + 384 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 157, + 289, + 384 + ], + "type": "inline_equation", + "content": "r_i" + }, + { + "bbox": [ + 46, + 157, + 289, + 384 + ], + "type": "text", + "content": " denotes the fidelity used for " + }, + { + "bbox": [ + 46, + 157, + 289, + 384 + ], + "type": "inline_equation", + "content": "\\lambda_i" + }, + { + "bbox": [ + 46, + 157, + 289, + 384 + ], + "type": "text", + "content": " to fit the surrogate model. The observations " + }, + { + "bbox": [ + 46, + 157, + 289, + 384 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_i" + }, + { + "bbox": [ + 46, + 157, + 289, + 384 + ], + "type": "text", + "content": " then become " + }, + { + "bbox": [ + 46, + 157, + 289, + 384 + ], + "type": "inline_equation", + "content": "\\{(\\lambda_1, y_1^{r_1}), \\dots, (\\lambda_i, y_i^{r_i})\\}" + }, + { + "bbox": [ + 46, + 157, + 289, + 384 + ], + "type": "text", + "content": ". To conclude, in order to extend single-fidelity methods to the multi-fidelity setting, the key challenge to be addressed is to determine " + }, + { + "bbox": [ + 46, + 157, + 289, + 384 + ], + "type": "inline_equation", + "content": "r_i" + }, + { + "bbox": [ + 46, + 157, + 289, + 384 + ], + "type": "text", + "content": " for each " + }, + { + "bbox": [ + 46, + 157, + 289, + 384 + ], + "type": "inline_equation", + "content": "\\lambda_i" + }, + { + "bbox": [ + 46, + 157, + 289, + 384 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 396, + 130, + 410 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 396, + 130, + 410 + ], + "spans": [ + { + "bbox": [ + 47, + 396, + 130, + 410 + ], + "type": "text", + "content": "4. Methodology" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 416, + 287, + 536 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 416, + 287, + 536 + ], + "spans": [ + { + "bbox": [ + 46, + 416, + 287, + 536 + ], + "type": "text", + "content": "In this section, we propose a novel multi-fidelity model-based algorithm FastBO. We first propose the key concepts of efficient point and saturation point, which are crucial in deciding the fidelity level to fit the surrogate model and to approximate the final fidelity respectively. Secondly, we elaborate on the details of learning curve modeling, where the two crucial points can be extracted. Then, we present the techniques associated with the auxiliary warm-up and post-processing stages. Finally, we summarize FastBO and discuss its wide applicability to any single-fidelity methods." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 544, + 280, + 556 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 544, + 280, + 556 + ], + "spans": [ + { + "bbox": [ + 47, + 544, + 280, + 556 + ], + "type": "text", + "content": "4.1. Estimation of Efficient and Saturation Points" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 563, + 286, + 611 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 563, + 286, + 611 + ], + "spans": [ + { + "bbox": [ + 46, + 563, + 286, + 611 + ], + "type": "text", + "content": "In our method, we adaptively identify efficient and saturation points for each configuration. The two points are crucial in the optimization process. We first formally define the efficient point as follows." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 620, + 286, + 681 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 620, + 286, + 681 + ], + "spans": [ + { + "bbox": [ + 46, + 620, + 286, + 681 + ], + "type": "text", + "content": "Definition 1 (Efficient point). For a given learning curve " + }, + { + "bbox": [ + 46, + 620, + 286, + 681 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_i(r)" + }, + { + "bbox": [ + 46, + 620, + 286, + 681 + ], + "type": "text", + "content": " of hyperparameter configuration " + }, + { + "bbox": [ + 46, + 620, + 286, + 681 + ], + "type": "inline_equation", + "content": "\\lambda_{i}" + }, + { + "bbox": [ + 46, + 620, + 286, + 681 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 620, + 286, + 681 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 46, + 620, + 286, + 681 + ], + "type": "text", + "content": " represents the resource level (also referred to as fidelity), the efficient point " + }, + { + "bbox": [ + 46, + 620, + 286, + 681 + ], + "type": "inline_equation", + "content": "e_i" + }, + { + "bbox": [ + 46, + 620, + 286, + 681 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 46, + 620, + 286, + 681 + ], + "type": "inline_equation", + "content": "\\lambda_{i}" + }, + { + "bbox": [ + 46, + 620, + 286, + 681 + ], + "type": "text", + "content": " is defined as: " + }, + { + "bbox": [ + 46, + 620, + 286, + 681 + ], + "type": "inline_equation", + "content": "e_i = \\min \\{r\\mid \\mathcal{C}_i(r) - \\mathcal{C}_i(2r) < \\delta_1\\}" + }, + { + "bbox": [ + 46, + 620, + 286, + 681 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 620, + 286, + 681 + ], + "type": "inline_equation", + "content": "\\delta_{1}" + }, + { + "bbox": [ + 46, + 620, + 286, + 681 + ], + "type": "text", + "content": " is a predefined small threshold." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 689, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 286, + 713 + ], + "type": "text", + "content": "The semantic of Definition 1 is that starting from the efficient point onwards, when the resources are doubled (i.e.," + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "text", + "content": "from " + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "inline_equation", + "content": "2r" + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "text", + "content": "), the performance improvement falls below a small threshold " + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "inline_equation", + "content": "\\delta_1" + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "text", + "content": ". Consequently, this point characterizes the fidelity at which a configuration demonstrates strong performance while still efficiently utilizing resources. In simpler terms, it signifies an appropriate fidelity of performance that can be achieved with comparably efficient resource usage. Therefore, we make the following remark." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 164, + 547, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 164, + 547, + 236 + ], + "spans": [ + { + "bbox": [ + 304, + 164, + 547, + 236 + ], + "type": "text", + "content": "Remark 1. The efficient points of the hyperparameter configurations can serve as their appropriate fidelities used for fitting the surrogate model. This is due to their (i) optimal resource-to-performance balance, (ii) ability to capture valuable learning curve trends, and (iii) customization for different hyperparameter configurations." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 243, + 546, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 243, + 546, + 434 + ], + "spans": [ + { + "bbox": [ + 304, + 243, + 546, + 434 + ], + "type": "text", + "content": "We elaborate on the reasons in Remark 1 as follows. Firstly, efficient points balance the trade-off between computational cost and result quality. Beyond the efficient point of a given configuration, allocating additional resources to that configuration becomes less efficient. Secondly, efficient points capture valuable trends within the learning curves. For example, the learning rate influences the shape of learning curves; the identification of efficient points for configurations with smaller learning rates often occurs at later stages. The insights into learning curve behaviors enable more informed decision-making. Thirdly, the ability to customize the fidelity for each specific configuration is a significant advantage. This adaptive approach is more reasonable than previous studies that use a fixed fidelity for all configurations, as it better accounts for the unique characteristics of individual learning curves." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 435, + 545, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 435, + 545, + 518 + ], + "spans": [ + { + "bbox": [ + 304, + 435, + 545, + 518 + ], + "type": "text", + "content": "This insight leads us to use the efficient point " + }, + { + "bbox": [ + 304, + 435, + 545, + 518 + ], + "type": "inline_equation", + "content": "e_i" + }, + { + "bbox": [ + 304, + 435, + 545, + 518 + ], + "type": "text", + "content": " identified for each configuration " + }, + { + "bbox": [ + 304, + 435, + 545, + 518 + ], + "type": "inline_equation", + "content": "\\lambda_i" + }, + { + "bbox": [ + 304, + 435, + 545, + 518 + ], + "type": "text", + "content": " as its fidelity used to fit the surrogate model. Specifically, we evaluate " + }, + { + "bbox": [ + 304, + 435, + 545, + 518 + ], + "type": "inline_equation", + "content": "\\lambda_i" + }, + { + "bbox": [ + 304, + 435, + 545, + 518 + ], + "type": "text", + "content": " until reaching " + }, + { + "bbox": [ + 304, + 435, + 545, + 518 + ], + "type": "inline_equation", + "content": "e_i" + }, + { + "bbox": [ + 304, + 435, + 545, + 518 + ], + "type": "text", + "content": " and obtain the observed performance " + }, + { + "bbox": [ + 304, + 435, + 545, + 518 + ], + "type": "inline_equation", + "content": "y_i^{e_i}" + }, + { + "bbox": [ + 304, + 435, + 545, + 518 + ], + "type": "text", + "content": ". The resulting data point " + }, + { + "bbox": [ + 304, + 435, + 545, + 518 + ], + "type": "inline_equation", + "content": "(\\lambda_i, y_i^{e_i})" + }, + { + "bbox": [ + 304, + 435, + 545, + 518 + ], + "type": "text", + "content": " is then added into the current observation set " + }, + { + "bbox": [ + 304, + 435, + 545, + 518 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{i-1}" + }, + { + "bbox": [ + 304, + 435, + 545, + 518 + ], + "type": "text", + "content": " to refit the surrogate model. We proof the superiority of FastBO over SHA-based methods in Supp. 8." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 519, + 545, + 554 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 519, + 545, + 554 + ], + "spans": [ + { + "bbox": [ + 304, + 519, + 545, + 554 + ], + "type": "text", + "content": "Besides efficient points, we identify saturation points for all configurations from their learning curves as well. We provide the definition of the saturation point as follows." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 562, + 547, + 622 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 562, + 547, + 622 + ], + "spans": [ + { + "bbox": [ + 304, + 562, + 547, + 622 + ], + "type": "text", + "content": "Definition 2 (Saturation point). For a given learning curve " + }, + { + "bbox": [ + 304, + 562, + 547, + 622 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_i(r)" + }, + { + "bbox": [ + 304, + 562, + 547, + 622 + ], + "type": "text", + "content": " of configuration " + }, + { + "bbox": [ + 304, + 562, + 547, + 622 + ], + "type": "inline_equation", + "content": "\\lambda_i" + }, + { + "bbox": [ + 304, + 562, + 547, + 622 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 562, + 547, + 622 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 304, + 562, + 547, + 622 + ], + "type": "text", + "content": " represents the resource level (also referred to as fidelity), the saturation point " + }, + { + "bbox": [ + 304, + 562, + 547, + 622 + ], + "type": "inline_equation", + "content": "s_i" + }, + { + "bbox": [ + 304, + 562, + 547, + 622 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 304, + 562, + 547, + 622 + ], + "type": "inline_equation", + "content": "\\lambda_i" + }, + { + "bbox": [ + 304, + 562, + 547, + 622 + ], + "type": "text", + "content": " is defined as: " + }, + { + "bbox": [ + 304, + 562, + 547, + 622 + ], + "type": "inline_equation", + "content": "s_i = \\min \\{r \\mid \\forall r' > r, |\\mathcal{C}_i(r') - \\mathcal{C}_i(r)| < \\delta_2\\}" + }, + { + "bbox": [ + 304, + 562, + 547, + 622 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 562, + 547, + 622 + ], + "type": "inline_equation", + "content": "\\delta_2" + }, + { + "bbox": [ + 304, + 562, + 547, + 622 + ], + "type": "text", + "content": " is a predefined small threshold." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "content": "The semantic of Definition 2 is that beyond the saturation point, the observed performance no longer exhibits notable variations with more resources. Thus, this point characterizes the fidelity at which the performance of a configuration stabilizes. The concept of saturation point is well-recognized within the machine learning community. Building on the above definition, we make the following remark." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "26183" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 45, + 72, + 287, + 122 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 72, + 287, + 122 + ], + "spans": [ + { + "bbox": [ + 45, + 72, + 287, + 122 + ], + "type": "text", + "content": "Remark 2. The saturation points of the hyperparameter configurations can serve as their approximate final fidelities, as they provide performance results that meet predefined quality thresholds while reducing resource wastage." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 126, + 288, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 126, + 288, + 236 + ], + "spans": [ + { + "bbox": [ + 46, + 126, + 288, + 236 + ], + "type": "text", + "content": "This insight leads us to use the saturation point " + }, + { + "bbox": [ + 46, + 126, + 288, + 236 + ], + "type": "inline_equation", + "content": "s_i" + }, + { + "bbox": [ + 46, + 126, + 288, + 236 + ], + "type": "text", + "content": " identified for each configuration " + }, + { + "bbox": [ + 46, + 126, + 288, + 236 + ], + "type": "inline_equation", + "content": "\\lambda_i" + }, + { + "bbox": [ + 46, + 126, + 288, + 236 + ], + "type": "text", + "content": " as its final fidelity approximation. The point is used in the post-processing stage for promoting some well-performing configurations to get higher-fidelity performances. In essence, when aiming for a full evaluation of the configurations, we suggest that terminating the evaluation at the saturation point is sufficient. A more intuitive illustration of the concepts of efficient and saturation points is provided in Supp. 9." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 240, + 192, + 254 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 240, + 192, + 254 + ], + "spans": [ + { + "bbox": [ + 46, + 240, + 192, + 254 + ], + "type": "text", + "content": "4.2. Learning Curve Modeling" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 45, + 259, + 287, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 259, + 287, + 462 + ], + "spans": [ + { + "bbox": [ + 45, + 259, + 287, + 462 + ], + "type": "text", + "content": "From Definitions 1 and 2, we can extract the efficient and saturation points of configurations from their learning curves. The curve " + }, + { + "bbox": [ + 45, + 259, + 287, + 462 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_i(r)" + }, + { + "bbox": [ + 45, + 259, + 287, + 462 + ], + "type": "text", + "content": " corresponds to configuration " + }, + { + "bbox": [ + 45, + 259, + 287, + 462 + ], + "type": "inline_equation", + "content": "\\lambda_{i}" + }, + { + "bbox": [ + 45, + 259, + 287, + 462 + ], + "type": "text", + "content": " and describes the predictive performance with " + }, + { + "bbox": [ + 45, + 259, + 287, + 462 + ], + "type": "inline_equation", + "content": "\\lambda_{i}" + }, + { + "bbox": [ + 45, + 259, + 287, + 462 + ], + "type": "text", + "content": " as a function of the fidelity " + }, + { + "bbox": [ + 45, + 259, + 287, + 462 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 45, + 259, + 287, + 462 + ], + "type": "text", + "content": ". Here, " + }, + { + "bbox": [ + 45, + 259, + 287, + 462 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 45, + 259, + 287, + 462 + ], + "type": "text", + "content": " can be either the number of training instances or the number of training epochs. In the context of learning curves, the former is referred to as observation learning curves, while the latter is iteration learning curves [29]. Both types are applicable to FastBO, so we use the term learning curve to encompass both. Given the observation set " + }, + { + "bbox": [ + 45, + 259, + 287, + 462 + ], + "type": "inline_equation", + "content": "\\mathcal{O}_i^w = \\{(r,y_i^r)\\}_{r = r_{min},\\dots,w}" + }, + { + "bbox": [ + 45, + 259, + 287, + 462 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 45, + 259, + 287, + 462 + ], + "type": "inline_equation", + "content": "\\lambda_{i}" + }, + { + "bbox": [ + 45, + 259, + 287, + 462 + ], + "type": "text", + "content": ", which comprises pairs of data points representing fidelities " + }, + { + "bbox": [ + 45, + 259, + 287, + 462 + ], + "type": "inline_equation", + "content": "r\\in \\{r_{min},\\dots,w\\}" + }, + { + "bbox": [ + 45, + 259, + 287, + 462 + ], + "type": "text", + "content": " and the corresponding evaluations " + }, + { + "bbox": [ + 45, + 259, + 287, + 462 + ], + "type": "inline_equation", + "content": "y_{i}^{r}" + }, + { + "bbox": [ + 45, + 259, + 287, + 462 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 45, + 259, + 287, + 462 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 45, + 259, + 287, + 462 + ], + "type": "text", + "content": " is a pre-defined warm-up point to stop collecting data, FastBO can estimate a learning curve for " + }, + { + "bbox": [ + 45, + 259, + 287, + 462 + ], + "type": "inline_equation", + "content": "\\lambda_{i}" + }, + { + "bbox": [ + 45, + 259, + 287, + 462 + ], + "type": "text", + "content": " based on " + }, + { + "bbox": [ + 45, + 259, + 287, + 462 + ], + "type": "inline_equation", + "content": "\\mathcal{O}_i^w" + }, + { + "bbox": [ + 45, + 259, + 287, + 462 + ], + "type": "text", + "content": " by first constructing a parametric learning curve model, then estimating the parameters." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 45, + 463, + 288, + 618 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 463, + 288, + 618 + ], + "spans": [ + { + "bbox": [ + 45, + 463, + 288, + 618 + ], + "type": "text", + "content": "Constructing a parametric learning curve model. Empirical learning curves can be modeled with function classes relying on some parameters. Viering and Loog [46] comprehensively summarized the parametric models studied in machine learning. In practice, different problems have different learning curves; even under the same problem, different hyperparameter configurations (e.g., learning rate, regularization, etc.) may lead to significantly different learning curves. Since one single parametric model is not enough to characterize all the learning curves by itself, we consider combining different parametric models into a single model. Specifically, we consider three parametric models POW3, EXP3 and LOG2, as listed in Tab. 1, which have shown" + } + ] + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 77, + 647, + 258, + 710 + ], + "blocks": [ + { + "bbox": [ + 78, + 628, + 254, + 639 + ], + "lines": [ + { + "bbox": [ + 78, + 628, + 254, + 639 + ], + "spans": [ + { + "bbox": [ + 78, + 628, + 254, + 639 + ], + "type": "text", + "content": "Table 1. Parametric learning curve models used." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 77, + 647, + 258, + 710 + ], + "lines": [ + { + "bbox": [ + 77, + 647, + 258, + 710 + ], + "spans": [ + { + "bbox": [ + 77, + 647, + 258, + 710 + ], + "type": "table", + "html": "
ModelFormulaFamily
POW3y = d + ax-αPower law
EXP3y = d + e-ax+bExponential
LOG2y = d + a log(x)Logarithmic
", + "image_path": "75d20c7159029cc3ddc01d6d90eb446c36175672fa38b1750ef219dae8207d7c.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 72, + 545, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 108 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 108 + ], + "type": "text", + "content": "good fitting and predicting performance in previous empirical studies [29, 46]. We provide detailed discussions on the choice of parametric models in Supp. 10." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 108, + 545, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 108, + 545, + 156 + ], + "spans": [ + { + "bbox": [ + 304, + 108, + 545, + 156 + ], + "type": "text", + "content": "Here, we denote each parametric model as " + }, + { + "bbox": [ + 304, + 108, + 545, + 156 + ], + "type": "inline_equation", + "content": "c_{j}(r|\\theta_{j})" + }, + { + "bbox": [ + 304, + 108, + 545, + 156 + ], + "type": "text", + "content": " with parameters " + }, + { + "bbox": [ + 304, + 108, + 545, + 156 + ], + "type": "inline_equation", + "content": "\\theta_{j}" + }, + { + "bbox": [ + 304, + 108, + 545, + 156 + ], + "type": "text", + "content": ", where the independent variable " + }, + { + "bbox": [ + 304, + 108, + 545, + 156 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 304, + 108, + 545, + 156 + ], + "type": "text", + "content": " represents the fidelity. We combine three models into one model via a weighted linear combination:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 353, + 166, + 545, + 186 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 353, + 166, + 545, + 186 + ], + "spans": [ + { + "bbox": [ + 353, + 166, + 545, + 186 + ], + "type": "interline_equation", + "content": "\\mathcal {C} (r | \\phi) = \\sum_ {j \\in \\{1, 2, 3 \\}} \\omega_ {j} c _ {j} (r | \\boldsymbol {\\theta} _ {j}), \\tag {1}", + "image_path": "bc7333f9621018a84a342f3d994cb34dc9627e03b8094100382be8fe9043e65b.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 195, + 545, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 195, + 545, + 266 + ], + "spans": [ + { + "bbox": [ + 304, + 195, + 545, + 266 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 195, + 545, + 266 + ], + "type": "inline_equation", + "content": "\\phi = \\{\\omega_1,\\omega_2,\\omega_3,\\theta_1,\\theta_2,\\theta_3\\}" + }, + { + "bbox": [ + 304, + 195, + 545, + 266 + ], + "type": "text", + "content": " is the parameter of the combined model, which consists of parameters " + }, + { + "bbox": [ + 304, + 195, + 545, + 266 + ], + "type": "inline_equation", + "content": "\\{\\pmb {\\theta}_1,\\pmb {\\theta}_2,\\pmb {\\theta}_3\\}" + }, + { + "bbox": [ + 304, + 195, + 545, + 266 + ], + "type": "text", + "content": " and weight " + }, + { + "bbox": [ + 304, + 195, + 545, + 266 + ], + "type": "inline_equation", + "content": "\\{\\omega_{1},\\omega_{2},\\omega_{3}\\}" + }, + { + "bbox": [ + 304, + 195, + 545, + 266 + ], + "type": "text", + "content": " of every single model. Therefore, each pair of observations " + }, + { + "bbox": [ + 304, + 195, + 545, + 266 + ], + "type": "inline_equation", + "content": "(r,y_i^r)" + }, + { + "bbox": [ + 304, + 195, + 545, + 266 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 304, + 195, + 545, + 266 + ], + "type": "inline_equation", + "content": "O_{i}^{w}" + }, + { + "bbox": [ + 304, + 195, + 545, + 266 + ], + "type": "text", + "content": " can be modeled by the combined model as " + }, + { + "bbox": [ + 304, + 195, + 545, + 266 + ], + "type": "inline_equation", + "content": "y_{i}^{r} = \\mathcal{C}(r|\\phi) + \\epsilon" + }, + { + "bbox": [ + 304, + 195, + 545, + 266 + ], + "type": "text", + "content": " , where " + }, + { + "bbox": [ + 304, + 195, + 545, + 266 + ], + "type": "inline_equation", + "content": "y_{i}^{r}" + }, + { + "bbox": [ + 304, + 195, + 545, + 266 + ], + "type": "text", + "content": " is the observed dependent variable and " + }, + { + "bbox": [ + 304, + 195, + 545, + 266 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 304, + 195, + 545, + 266 + ], + "type": "text", + "content": " represents the error term." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 268, + 545, + 363 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 268, + 545, + 363 + ], + "spans": [ + { + "bbox": [ + 304, + 268, + 545, + 363 + ], + "type": "text", + "content": "Estimating parameters in the parametric learning curve model. We employ maximum likelihood estimation to estimate the parameters " + }, + { + "bbox": [ + 304, + 268, + 545, + 363 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 304, + 268, + 545, + 363 + ], + "type": "text", + "content": " in the parametric model " + }, + { + "bbox": [ + 304, + 268, + 545, + 363 + ], + "type": "inline_equation", + "content": "\\mathcal{C}(r|\\phi)" + }, + { + "bbox": [ + 304, + 268, + 545, + 363 + ], + "type": "text", + "content": ". Assuming that " + }, + { + "bbox": [ + 304, + 268, + 545, + 363 + ], + "type": "inline_equation", + "content": "\\epsilon \\sim \\mathcal{N}(0,\\sigma^2)" + }, + { + "bbox": [ + 304, + 268, + 545, + 363 + ], + "type": "text", + "content": ", the probability of an observed performance " + }, + { + "bbox": [ + 304, + 268, + 545, + 363 + ], + "type": "inline_equation", + "content": "y_{i}^{r}" + }, + { + "bbox": [ + 304, + 268, + 545, + 363 + ], + "type": "text", + "content": " under parameters is given by " + }, + { + "bbox": [ + 304, + 268, + 545, + 363 + ], + "type": "inline_equation", + "content": "p(y_i^r |\\phi ,\\sigma^2) = \\mathcal{N}(y_i^r;\\mathcal{C}(r|\\phi),\\sigma^2)" + }, + { + "bbox": [ + 304, + 268, + 545, + 363 + ], + "type": "text", + "content": ". Given the observations " + }, + { + "bbox": [ + 304, + 268, + 545, + 363 + ], + "type": "inline_equation", + "content": "\\mathcal{O}_i^w" + }, + { + "bbox": [ + 304, + 268, + 545, + 363 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 304, + 268, + 545, + 363 + ], + "type": "inline_equation", + "content": "\\lambda_{i}" + }, + { + "bbox": [ + 304, + 268, + 545, + 363 + ], + "type": "text", + "content": " that contains a set of observed data points " + }, + { + "bbox": [ + 304, + 268, + 545, + 363 + ], + "type": "inline_equation", + "content": "(r,y_i^r)" + }, + { + "bbox": [ + 304, + 268, + 545, + 363 + ], + "type": "text", + "content": ", the likelihood function can be expressed as:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 321, + 373, + 545, + 425 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 373, + 545, + 425 + ], + "spans": [ + { + "bbox": [ + 321, + 373, + 545, + 425 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} (\\boldsymbol {\\phi}, \\sigma^ {2}; \\mathbf {r}, \\mathbf {y} _ {i} ^ {r}) = \\prod p (y _ {i} ^ {k} | \\boldsymbol {\\phi}, \\sigma^ {2}) \\\\ = \\prod_ {k = r _ {\\text {m i n}}} ^ {w} \\frac {1}{\\sigma \\sqrt {2 \\pi}} \\exp \\left(- \\frac {\\left(y _ {i} ^ {r} - \\mathcal {C} (r = k | \\phi)\\right) ^ {2}}{2 \\sigma^ {2}}\\right). \\tag {2} \\\\ \\end{array}", + "image_path": "655ba83d1a67812986bde9ecdaa2075bcacc49cd4d77f9310e2dd7e9941ba5a7.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 434, + 544, + 457 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 434, + 544, + 457 + ], + "spans": [ + { + "bbox": [ + 304, + 434, + 544, + 457 + ], + "type": "text", + "content": "We estimate " + }, + { + "bbox": [ + 304, + 434, + 544, + 457 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 304, + 434, + 544, + 457 + ], + "type": "text", + "content": " by maximizing log-likelihood function, which is easily calculated given Eq. 2." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 459, + 545, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 459, + 545, + 555 + ], + "spans": [ + { + "bbox": [ + 304, + 459, + 545, + 555 + ], + "type": "text", + "content": "An existing model-free method [7] also considers using learning curves for the HPO problem. However, it targets predicting the high-fidelity performance from the low-fidelity observations and thus stopping configurations that are unlikely to beat the current best values, which is different from our main target of identifying appropriate fidelity levels for the configurations to fit the surrogate model from their estimated learning curves." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 563, + 504, + 576 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 563, + 504, + 576 + ], + "spans": [ + { + "bbox": [ + 305, + 563, + 504, + 576 + ], + "type": "text", + "content": "4.3. Warm-up And Post-processing Stages" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 582, + 545, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 582, + 545, + 653 + ], + "spans": [ + { + "bbox": [ + 304, + 582, + 545, + 653 + ], + "type": "text", + "content": "In addition to its core components, FastBO incorporates two auxiliary stages: the warm-up and post-processing stages. For the completeness of our method, we provide an overview of these stages, outlining their targets and presenting the key techniques of early-termination detection and saturation-level evaluation that are applied within." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "content": "Warm-up stage. The warm-up stage prepares the early observation set " + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{O}_i^w" + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "content": " for each configuration " + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\lambda_{i}" + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "content": " that is used to estimate its learning curve, as discussed in " + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\S 4.2" + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "content": ". Here " + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "inline_equation", + "content": "w\\in (r_{min},r_{max})" + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "content": " is a pre-determined fidelity, denoted as warm-up point. Specifically, we initiate the evaluation of" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "26184" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 289, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 289, + 335 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 289, + 335 + ], + "type": "text", + "content": "each newly selected " + }, + { + "bbox": [ + 47, + 72, + 289, + 335 + ], + "type": "inline_equation", + "content": "\\lambda_{i}" + }, + { + "bbox": [ + 47, + 72, + 289, + 335 + ], + "type": "text", + "content": ", proceeding until reaching " + }, + { + "bbox": [ + 47, + 72, + 289, + 335 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 47, + 72, + 289, + 335 + ], + "type": "text", + "content": ". During this process, we record each fidelity " + }, + { + "bbox": [ + 47, + 72, + 289, + 335 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 47, + 72, + 289, + 335 + ], + "type": "text", + "content": " and its evaluation result " + }, + { + "bbox": [ + 47, + 72, + 289, + 335 + ], + "type": "inline_equation", + "content": "y_{i}^{r}" + }, + { + "bbox": [ + 47, + 72, + 289, + 335 + ], + "type": "text", + "content": ", forming pairs " + }, + { + "bbox": [ + 47, + 72, + 289, + 335 + ], + "type": "inline_equation", + "content": "(r,y_i^r)" + }, + { + "bbox": [ + 47, + 72, + 289, + 335 + ], + "type": "text", + "content": ". Upon reaching " + }, + { + "bbox": [ + 47, + 72, + 289, + 335 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 47, + 72, + 289, + 335 + ], + "type": "text", + "content": ", we pause the evaluation for " + }, + { + "bbox": [ + 47, + 72, + 289, + 335 + ], + "type": "inline_equation", + "content": "\\lambda_{i}" + }, + { + "bbox": [ + 47, + 72, + 289, + 335 + ], + "type": "text", + "content": " and obtain its early observation set " + }, + { + "bbox": [ + 47, + 72, + 289, + 335 + ], + "type": "inline_equation", + "content": "\\mathcal{O}_i^w = \\{(r,y_i^r)\\}_{r = r_{min},\\ldots ,w}" + }, + { + "bbox": [ + 47, + 72, + 289, + 335 + ], + "type": "text", + "content": ", and start modeling the learning curve. During the warm-up stage, we monitor the performance changes across every two continuous fidelities. If we detect that the performance of " + }, + { + "bbox": [ + 47, + 72, + 289, + 335 + ], + "type": "inline_equation", + "content": "\\lambda_{i}" + }, + { + "bbox": [ + 47, + 72, + 289, + 335 + ], + "type": "text", + "content": " has consecutively dropped twice by more than a ratio " + }, + { + "bbox": [ + 47, + 72, + 289, + 335 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 47, + 72, + 289, + 335 + ], + "type": "text", + "content": ", i.e., " + }, + { + "bbox": [ + 47, + 72, + 289, + 335 + ], + "type": "inline_equation", + "content": "(y_{i}^{r - 1} - y_{i}^{r - 2}) > \\alpha y_{i}^{r - 2}" + }, + { + "bbox": [ + 47, + 72, + 289, + 335 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 72, + 289, + 335 + ], + "type": "inline_equation", + "content": "(y_i^r -y_i^{r - 1}) > \\alpha y_i^{r - 1}" + }, + { + "bbox": [ + 47, + 72, + 289, + 335 + ], + "type": "text", + "content": ", we promptly terminate the evaluation for " + }, + { + "bbox": [ + 47, + 72, + 289, + 335 + ], + "type": "inline_equation", + "content": "\\lambda_{i}" + }, + { + "bbox": [ + 47, + 72, + 289, + 335 + ], + "type": "text", + "content": " at its current fidelity " + }, + { + "bbox": [ + 47, + 72, + 289, + 335 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 47, + 72, + 289, + 335 + ], + "type": "text", + "content": ", because such consecutive performance deterioration indicates " + }, + { + "bbox": [ + 47, + 72, + 289, + 335 + ], + "type": "inline_equation", + "content": "\\lambda_{i}" + }, + { + "bbox": [ + 47, + 72, + 289, + 335 + ], + "type": "text", + "content": " is unlikely to achieve satisfactory performance. Once terminated, we directly incorporate the current performance " + }, + { + "bbox": [ + 47, + 72, + 289, + 335 + ], + "type": "inline_equation", + "content": "y_{i}^{r}" + }, + { + "bbox": [ + 47, + 72, + 289, + 335 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 47, + 72, + 289, + 335 + ], + "type": "inline_equation", + "content": "\\lambda_{i}" + }, + { + "bbox": [ + 47, + 72, + 289, + 335 + ], + "type": "text", + "content": " into " + }, + { + "bbox": [ + 47, + 72, + 289, + 335 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{i - 1}" + }, + { + "bbox": [ + 47, + 72, + 289, + 335 + ], + "type": "text", + "content": " that is used for updating the surrogate model. Thus, further operations like learning curve modeling are discontinued for " + }, + { + "bbox": [ + 47, + 72, + 289, + 335 + ], + "type": "inline_equation", + "content": "\\lambda_{i}" + }, + { + "bbox": [ + 47, + 72, + 289, + 335 + ], + "type": "text", + "content": ". Moreover, if we observe a single case of performance drop without subsequent occurrences, i.e., " + }, + { + "bbox": [ + 47, + 72, + 289, + 335 + ], + "type": "inline_equation", + "content": "y_{i}^{r - 1} - y_{i}^{r - 2} > \\alpha y_{i}^{r - 2}" + }, + { + "bbox": [ + 47, + 72, + 289, + 335 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 72, + 289, + 335 + ], + "type": "inline_equation", + "content": "y_{i}^{r} - y_{i}^{r - 1}\\leq \\alpha y_{i}^{r - 1}" + }, + { + "bbox": [ + 47, + 72, + 289, + 335 + ], + "type": "text", + "content": ", we opt not to include data from fidelity " + }, + { + "bbox": [ + 47, + 72, + 289, + 335 + ], + "type": "inline_equation", + "content": "r - 1" + }, + { + "bbox": [ + 47, + 72, + 289, + 335 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 47, + 72, + 289, + 335 + ], + "type": "inline_equation", + "content": "\\mathcal{O}_i^w" + }, + { + "bbox": [ + 47, + 72, + 289, + 335 + ], + "type": "text", + "content": ". This is to manually filter out potential noise in the data that may adversely affect the fitting of the learning curve." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 337, + 289, + 661 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 337, + 289, + 661 + ], + "spans": [ + { + "bbox": [ + 47, + 337, + 289, + 661 + ], + "type": "text", + "content": "Post-processing stage. The post-processing stage aims at two tasks: promoting the well-performing configurations for saturation-level evaluations and identifying the best configuration and its performance. Firstly, FastBO promotes the top-" + }, + { + "bbox": [ + 47, + 337, + 289, + 661 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 47, + 337, + 289, + 661 + ], + "type": "text", + "content": " well-performing configurations and evaluates them to their saturation points to ensure high-quality performance while maintaining efficient resource utilization. We set " + }, + { + "bbox": [ + 47, + 337, + 289, + 661 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 47, + 337, + 289, + 661 + ], + "type": "text", + "content": " to be always less than or equal to the number of parallel workers available, ensuring a manageable overhead of saturation-level evaluations. It is worth noting that the additional time required is factored into the overall time. Secondly, FastBO finds the best configuration along with its performance achieved so far, which is a standard final step in most HPO methods. However, an increase in fidelities does not always result in performance improvement, possibly due to overfitting, resource saturation, or problem complexity. Therefore, we treat the evaluation at each fidelity as an individual task and record all these intermediate evaluation results, which is also a common practice in recent implementations. In this way, FastBO finds the best performance by considering all the results, rather than relying solely on the highest-fidelity performances of the configurations. In the parallel setting, treating each fidelity evaluation as an individual task offers an added benefit due to its finer granularity. More specifically, when a worker is idle, it takes on a new task of evaluating a configuration at a specific fidelity, rather than evaluating an entire configuration." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 670, + 198, + 682 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 670, + 198, + 682 + ], + "spans": [ + { + "bbox": [ + 47, + 670, + 198, + 682 + ], + "type": "text", + "content": "4.4. FastBO and Generalization" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 689, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 287, + 715 + ], + "type": "text", + "content": "Algorithm 1 summarizes our proposed FastBO. It takes surrogate model " + }, + { + "bbox": [ + 47, + 689, + 287, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 47, + 689, + 287, + 715 + ], + "type": "text", + "content": ", acquisition function " + }, + { + "bbox": [ + 47, + 689, + 287, + 715 + ], + "type": "inline_equation", + "content": "a" + }, + { + "bbox": [ + 47, + 689, + 287, + 715 + ], + "type": "text", + "content": ", warm-up point " + }, + { + "bbox": [ + 47, + 689, + 287, + 715 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 47, + 689, + 287, + 715 + ], + "type": "text", + "content": "," + } + ] + } + ], + "index": 3 + }, + { + "type": "code", + "bbox": [ + 306, + 88, + 545, + 296 + ], + "blocks": [ + { + "bbox": [ + 310, + 74, + 445, + 87 + ], + "lines": [ + { + "bbox": [ + 310, + 74, + 445, + 87 + ], + "spans": [ + { + "bbox": [ + 310, + 74, + 445, + 87 + ], + "type": "text", + "content": "Algorithm 1: FastBO algorithm" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "code_caption" + }, + { + "bbox": [ + 306, + 88, + 545, + 296 + ], + "lines": [ + { + "bbox": [ + 306, + 88, + 545, + 296 + ], + "spans": [ + { + "bbox": [ + 306, + 88, + 545, + 296 + ], + "type": "text", + "content": "input : " + }, + { + "bbox": [ + 306, + 88, + 545, + 296 + ], + "type": "inline_equation", + "content": "\\mathcal{M},a,w,\\alpha ,k,\\delta_1,\\delta_2" + }, + { + "bbox": [ + 306, + 88, + 545, + 296 + ], + "type": "text", + "content": " output: " + }, + { + "bbox": [ + 306, + 88, + 545, + 296 + ], + "type": "inline_equation", + "content": "\\lambda^{*},y^{*}" + }, + { + "bbox": [ + 306, + 88, + 545, + 296 + ], + "type": "text", + "content": " \n1 " + }, + { + "bbox": [ + 306, + 88, + 545, + 296 + ], + "type": "inline_equation", + "content": "i\\gets 0,D\\gets \\emptyset" + }, + { + "bbox": [ + 306, + 88, + 545, + 296 + ], + "type": "text", + "content": " \n2 while not meet the stop criterion do \n3 find " + }, + { + "bbox": [ + 306, + 88, + 545, + 296 + ], + "type": "inline_equation", + "content": "\\lambda_{i}\\gets \\arg \\max_{\\lambda \\in \\Lambda}a(\\lambda ,\\mathcal{M}_{i - 1})" + }, + { + "bbox": [ + 306, + 88, + 545, + 296 + ], + "type": "text", + "content": " \n4 " + }, + { + "bbox": [ + 306, + 88, + 545, + 296 + ], + "type": "inline_equation", + "content": "\\mathcal{O}_i^w,t\\gets" + }, + { + "bbox": [ + 306, + 88, + 545, + 296 + ], + "type": "text", + "content": " warm-up given " + }, + { + "bbox": [ + 306, + 88, + 545, + 296 + ], + "type": "inline_equation", + "content": "w,\\alpha" + }, + { + "bbox": [ + 306, + 88, + 545, + 296 + ], + "type": "text", + "content": " // cf. " + }, + { + "bbox": [ + 306, + 88, + 545, + 296 + ], + "type": "inline_equation", + "content": "\\S 4.3" + }, + { + "bbox": [ + 306, + 88, + 545, + 296 + ], + "type": "text", + "content": " \n5 if " + }, + { + "bbox": [ + 306, + 88, + 545, + 296 + ], + "type": "inline_equation", + "content": "\\mathcal{O}_i^w" + }, + { + "bbox": [ + 306, + 88, + 545, + 296 + ], + "type": "text", + "content": " is not empty then \n6 fit " + }, + { + "bbox": [ + 306, + 88, + 545, + 296 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_i(r)" + }, + { + "bbox": [ + 306, + 88, + 545, + 296 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 306, + 88, + 545, + 296 + ], + "type": "inline_equation", + "content": "\\mathcal{O}_i^w" + }, + { + "bbox": [ + 306, + 88, + 545, + 296 + ], + "type": "text", + "content": " // cf. " + }, + { + "bbox": [ + 306, + 88, + 545, + 296 + ], + "type": "inline_equation", + "content": "\\S 4.2" + }, + { + "bbox": [ + 306, + 88, + 545, + 296 + ], + "type": "text", + "content": " \n7 find " + }, + { + "bbox": [ + 306, + 88, + 545, + 296 + ], + "type": "inline_equation", + "content": "e_i,s_i" + }, + { + "bbox": [ + 306, + 88, + 545, + 296 + ], + "type": "text", + "content": " given " + }, + { + "bbox": [ + 306, + 88, + 545, + 296 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_i(r),\\delta_1,\\delta_2" + }, + { + "bbox": [ + 306, + 88, + 545, + 296 + ], + "type": "text", + "content": " // cf. " + }, + { + "bbox": [ + 306, + 88, + 545, + 296 + ], + "type": "inline_equation", + "content": "\\S 4.1" + }, + { + "bbox": [ + 306, + 88, + 545, + 296 + ], + "type": "text", + "content": " \n8 " + }, + { + "bbox": [ + 306, + 88, + 545, + 296 + ], + "type": "inline_equation", + "content": "y_{i}^{e_{i}}\\gets" + }, + { + "bbox": [ + 306, + 88, + 545, + 296 + ], + "type": "text", + "content": " continue evaluating " + }, + { + "bbox": [ + 306, + 88, + 545, + 296 + ], + "type": "inline_equation", + "content": "\\lambda_{i}" + }, + { + "bbox": [ + 306, + 88, + 545, + 296 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 306, + 88, + 545, + 296 + ], + "type": "inline_equation", + "content": "e_i" + }, + { + "bbox": [ + 306, + 88, + 545, + 296 + ], + "type": "text", + "content": " \n9 else \n10 " + }, + { + "bbox": [ + 306, + 88, + 545, + 296 + ], + "type": "inline_equation", + "content": "e_i\\gets t,s_i\\gets r_{max}" + }, + { + "bbox": [ + 306, + 88, + 545, + 296 + ], + "type": "text", + "content": " \n11 " + }, + { + "bbox": [ + 306, + 88, + 545, + 296 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_i\\gets \\mathcal{D}_{i - 1}\\cup (\\lambda_i,y_i^{e_i})" + }, + { + "bbox": [ + 306, + 88, + 545, + 296 + ], + "type": "text", + "content": " \n12 refit " + }, + { + "bbox": [ + 306, + 88, + 545, + 296 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_i" + }, + { + "bbox": [ + 306, + 88, + 545, + 296 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 306, + 88, + 545, + 296 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_i" + }, + { + "bbox": [ + 306, + 88, + 545, + 296 + ], + "type": "text", + "content": " \n13 " + }, + { + "bbox": [ + 306, + 88, + 545, + 296 + ], + "type": "inline_equation", + "content": "i\\gets i + 1" + }, + { + "bbox": [ + 306, + 88, + 545, + 296 + ], + "type": "text", + "content": " \n14 " + }, + { + "bbox": [ + 306, + 88, + 545, + 296 + ], + "type": "inline_equation", + "content": "\\lambda^{*},y^{*}\\gets" + }, + { + "bbox": [ + 306, + 88, + 545, + 296 + ], + "type": "text", + "content": " post-process given s = " + }, + { + "bbox": [ + 306, + 88, + 545, + 296 + ], + "type": "inline_equation", + "content": "\\{s_i\\} ,k / / c f.\\S 4.3" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "code_body" + } + ], + "index": 5, + "sub_type": "algorithm" + }, + { + "bbox": [ + 304, + 327, + 547, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 327, + 547, + 579 + ], + "spans": [ + { + "bbox": [ + 304, + 327, + 547, + 579 + ], + "type": "text", + "content": "performance decrease ratio " + }, + { + "bbox": [ + 304, + 327, + 547, + 579 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 304, + 327, + 547, + 579 + ], + "type": "text", + "content": ", promotion number " + }, + { + "bbox": [ + 304, + 327, + 547, + 579 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 327, + 547, + 579 + ], + "type": "text", + "content": ", and thresholds " + }, + { + "bbox": [ + 304, + 327, + 547, + 579 + ], + "type": "inline_equation", + "content": "\\delta_{1}" + }, + { + "bbox": [ + 304, + 327, + 547, + 579 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 327, + 547, + 579 + ], + "type": "inline_equation", + "content": "\\delta_{2}" + }, + { + "bbox": [ + 304, + 327, + 547, + 579 + ], + "type": "text", + "content": " as inputs, and output the best-founded configuration " + }, + { + "bbox": [ + 304, + 327, + 547, + 579 + ], + "type": "inline_equation", + "content": "\\lambda^{*}" + }, + { + "bbox": [ + 304, + 327, + 547, + 579 + ], + "type": "text", + "content": " and its performance " + }, + { + "bbox": [ + 304, + 327, + 547, + 579 + ], + "type": "inline_equation", + "content": "y^{*}" + }, + { + "bbox": [ + 304, + 327, + 547, + 579 + ], + "type": "text", + "content": ". FastBO follows a similar iterative process of model-based methods but replaces the expensive full evaluations with a more intelligent alternative (cf. Lines 4-10). Specifically, each configuration " + }, + { + "bbox": [ + 304, + 327, + 547, + 579 + ], + "type": "inline_equation", + "content": "\\lambda_{i}" + }, + { + "bbox": [ + 304, + 327, + 547, + 579 + ], + "type": "text", + "content": " first enters a warm-up stage to collect its early observation set " + }, + { + "bbox": [ + 304, + 327, + 547, + 579 + ], + "type": "inline_equation", + "content": "\\mathcal{O}_i^w" + }, + { + "bbox": [ + 304, + 327, + 547, + 579 + ], + "type": "text", + "content": " and to be detected and terminated if it exhibits consecutive performance deterioration (cf. Line 4). If " + }, + { + "bbox": [ + 304, + 327, + 547, + 579 + ], + "type": "inline_equation", + "content": "\\lambda_{i}" + }, + { + "bbox": [ + 304, + 327, + 547, + 579 + ], + "type": "text", + "content": " is not terminated, FastBO then estimates a learning curve " + }, + { + "bbox": [ + 304, + 327, + 547, + 579 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_i(r)" + }, + { + "bbox": [ + 304, + 327, + 547, + 579 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 304, + 327, + 547, + 579 + ], + "type": "inline_equation", + "content": "\\lambda_{i}" + }, + { + "bbox": [ + 304, + 327, + 547, + 579 + ], + "type": "text", + "content": " based on " + }, + { + "bbox": [ + 304, + 327, + 547, + 579 + ], + "type": "inline_equation", + "content": "\\mathcal{O}_i^w" + }, + { + "bbox": [ + 304, + 327, + 547, + 579 + ], + "type": "text", + "content": " (cf. Line 6), and thus the efficient point and saturation point of " + }, + { + "bbox": [ + 304, + 327, + 547, + 579 + ], + "type": "inline_equation", + "content": "\\lambda_{i}" + }, + { + "bbox": [ + 304, + 327, + 547, + 579 + ], + "type": "text", + "content": " can be obtained (cf. Line 7). After that, " + }, + { + "bbox": [ + 304, + 327, + 547, + 579 + ], + "type": "inline_equation", + "content": "\\lambda_{i}" + }, + { + "bbox": [ + 304, + 327, + 547, + 579 + ], + "type": "text", + "content": " continues to be evaluated until reaching " + }, + { + "bbox": [ + 304, + 327, + 547, + 579 + ], + "type": "inline_equation", + "content": "e_i" + }, + { + "bbox": [ + 304, + 327, + 547, + 579 + ], + "type": "text", + "content": " (cf. Line 8); the result is added to the observation set " + }, + { + "bbox": [ + 304, + 327, + 547, + 579 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 304, + 327, + 547, + 579 + ], + "type": "text", + "content": " (cf. Line 11) that is used for updating " + }, + { + "bbox": [ + 304, + 327, + 547, + 579 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 304, + 327, + 547, + 579 + ], + "type": "text", + "content": " (cf. Line 12). On the other hand, the poorly-performing configuration will be terminated early at fidelity " + }, + { + "bbox": [ + 304, + 327, + 547, + 579 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 327, + 547, + 579 + ], + "type": "text", + "content": " with its result being added directly to " + }, + { + "bbox": [ + 304, + 327, + 547, + 579 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 304, + 327, + 547, + 579 + ], + "type": "text", + "content": " (cf. Lines 10, 11). Finally, the post-processing stage promotes the most promising configurations to their saturation points and finds the best-founded configuration " + }, + { + "bbox": [ + 304, + 327, + 547, + 579 + ], + "type": "inline_equation", + "content": "\\lambda^{*}" + }, + { + "bbox": [ + 304, + 327, + 547, + 579 + ], + "type": "text", + "content": " and its performance " + }, + { + "bbox": [ + 304, + 327, + 547, + 579 + ], + "type": "inline_equation", + "content": "y^{*}" + }, + { + "bbox": [ + 304, + 327, + 547, + 579 + ], + "type": "text", + "content": " (cf. Line 14)." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 582, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 582, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 582, + 547, + 715 + ], + "type": "text", + "content": "Generalizing FastBO to single-fidelity methods. The core of FastBO is to tackle the key challenge of deciding an appropriate fidelity for each configuration to fit the surrogate model by adaptively identifying its efficient point. This strategy of using the efficient point performances for surrogate model fitting also provides a simple but effective way to bridge the gap between single- and multi-fidelity methods. While it is primarily described in the context of model-based methods, the strategy can be generalized to various single-fidelity methods. For example, when evaluating configurations within the population for an evolution-" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "26185" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 47, + 70, + 547, + 189 + ], + "blocks": [ + { + "bbox": [ + 47, + 70, + 547, + 189 + ], + "lines": [ + { + "bbox": [ + 47, + 70, + 547, + 189 + ], + "spans": [ + { + "bbox": [ + 47, + 70, + 547, + 189 + ], + "type": "image", + "image_path": "f02b00b4e9beacdb5607cfb0cc7957f340ddd17d4c6cf9731807ca0c3e39398f.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 145, + 194, + 447, + 206 + ], + "lines": [ + { + "bbox": [ + 145, + 194, + 447, + 206 + ], + "spans": [ + { + "bbox": [ + 145, + 194, + 447, + 206 + ], + "type": "text", + "content": "Figure 1. Performance of average validation accuracy on the LCBench benchmark." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 47, + 217, + 545, + 345 + ], + "blocks": [ + { + "bbox": [ + 47, + 217, + 545, + 345 + ], + "lines": [ + { + "bbox": [ + 47, + 217, + 545, + 345 + ], + "spans": [ + { + "bbox": [ + 47, + 217, + 545, + 345 + ], + "type": "image", + "image_path": "c6bd55335731ea34e417923de43d229dc73256aba6d6635452af1a8c6aa9e4be.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 86, + 352, + 506, + 364 + ], + "lines": [ + { + "bbox": [ + 86, + 352, + 506, + 364 + ], + "spans": [ + { + "bbox": [ + 86, + 352, + 506, + 364 + ], + "type": "text", + "content": "Figure 2. Performance of (a) average validation error on NAS-Bench-201 and (b) average validation loss on FCNet." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 379, + 289, + 511 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 379, + 289, + 511 + ], + "spans": [ + { + "bbox": [ + 46, + 379, + 289, + 511 + ], + "type": "text", + "content": "ary algorithm-based HPO method, we can similarly evaluate the efficient point performances instead of the final performances of these configurations and integrate the performances in the subsequent processes, such as selection and variation. Relying on the efficient point rather than the final fidelity or all fidelities available simplifies the extension of the single-fidelity methods to the multi-fidelity setting. The rationale behind this adaptive fidelity identification strategy is discussed in Remark 1. We also demonstrate in our experiments the efficacy of this strategy in extending a range of single-fidelity methods to the multi-fidelity setting." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 524, + 128, + 538 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 524, + 128, + 538 + ], + "spans": [ + { + "bbox": [ + 47, + 524, + 128, + 538 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 544, + 287, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 544, + 287, + 652 + ], + "spans": [ + { + "bbox": [ + 46, + 544, + 287, + 652 + ], + "type": "text", + "content": "We empirically evaluate the performance of FastBO and compare it with the random search baseline (RS) and 9 competitive baselines from 3 related categories, including (i) model-based methods: standard Gaussian Process-based BO [40]; (ii) multi-fidelity methods: ASHA [26], Hyperband [25], PASHA [4]; and (iii) model-based multifidelity methods: A-BOHB [22], A-CQR [37], BOHB [10], DyHPO [48], Hyper-Tune [27]. RS and BO are single-fidelity baselines, while the others are multi-fidelity ones." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "type": "text", + "content": "Our experiments are conducted on 10 datasets from 3 popular benchmarks LCBench [53], NAS-Bench-201 [8] and FCNet [21]. Detailed information on the benchmarks is provided in Supp. 13.1. All the experiments are evaluated with four parallel workers and 10 random seeds. We" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 379, + 547, + 477 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 379, + 547, + 477 + ], + "spans": [ + { + "bbox": [ + 304, + 379, + 547, + 477 + ], + "type": "text", + "content": "allocate " + }, + { + "bbox": [ + 304, + 379, + 547, + 477 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 304, + 379, + 547, + 477 + ], + "type": "text", + "content": " total budget for warm-up, i.e., " + }, + { + "bbox": [ + 304, + 379, + 547, + 477 + ], + "type": "inline_equation", + "content": "w = r_{min} + 0.2 \\cdot (r_{max} - r_{min})" + }, + { + "bbox": [ + 304, + 379, + 547, + 477 + ], + "type": "text", + "content": ". Ratio " + }, + { + "bbox": [ + 304, + 379, + 547, + 477 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 304, + 379, + 547, + 477 + ], + "type": "text", + "content": " is set to 0.1; thresholds " + }, + { + "bbox": [ + 304, + 379, + 547, + 477 + ], + "type": "inline_equation", + "content": "\\delta_1" + }, + { + "bbox": [ + 304, + 379, + 547, + 477 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 379, + 547, + 477 + ], + "type": "inline_equation", + "content": "\\delta_2" + }, + { + "bbox": [ + 304, + 379, + 547, + 477 + ], + "type": "text", + "content": " are set to 0.001 and " + }, + { + "bbox": [ + 304, + 379, + 547, + 477 + ], + "type": "inline_equation", + "content": "0.0005^1" + }, + { + "bbox": [ + 304, + 379, + 547, + 477 + ], + "type": "text", + "content": ". We set " + }, + { + "bbox": [ + 304, + 379, + 547, + 477 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 379, + 547, + 477 + ], + "type": "text", + "content": " based on the number of workers " + }, + { + "bbox": [ + 304, + 379, + 547, + 477 + ], + "type": "inline_equation", + "content": "\\# w" + }, + { + "bbox": [ + 304, + 379, + 547, + 477 + ], + "type": "text", + "content": " and the number of started configurations " + }, + { + "bbox": [ + 304, + 379, + 547, + 477 + ], + "type": "inline_equation", + "content": "\\# c" + }, + { + "bbox": [ + 304, + 379, + 547, + 477 + ], + "type": "text", + "content": ": " + }, + { + "bbox": [ + 304, + 379, + 547, + 477 + ], + "type": "inline_equation", + "content": "k = \\max \\{\\lceil \\# c / 10 \\rceil, \\# w\\}" + }, + { + "bbox": [ + 304, + 379, + 547, + 477 + ], + "type": "text", + "content": ". We provide more experiments and discussions on the hyperparameters in Supp. 12. We use implementations of the baselines in Syne Tune [36]. Details of the baseline settings are in Supp. 13.2." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 484, + 432, + 498 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 484, + 432, + 498 + ], + "spans": [ + { + "bbox": [ + 305, + 484, + 432, + 498 + ], + "type": "text", + "content": "5.1. Anytime Performance" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 502, + 547, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 502, + 547, + 586 + ], + "spans": [ + { + "bbox": [ + 304, + 502, + 547, + 586 + ], + "type": "text", + "content": "To evaluate the anytime performance, we compare FastBO against the baselines on wall-clock time. For fair comparisons, all the baselines, even single-fidelity BO and RS, are extended to consider intermediate results at all the fidelities when identifying the configuration, akin to FastBO as discussed in § 4.3. Consequently, all the baselines are able to achieve their best possible anytime performance." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 587, + 547, + 684 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 587, + 547, + 684 + ], + "spans": [ + { + "bbox": [ + 304, + 587, + 547, + 684 + ], + "type": "text", + "content": "The results on LCBench, NAS-Bench-201, and FCNet are shown in Figs. 1 and 2. We report the validation accuracy, error, and loss over wall-clock time for the three benchmarks, as provided by the benchmarks. We provide the results on NAS-Bench-301 [39] in Supp. 11.1. Overall, FastBO can handle various performance metrics and shows strong anytime performance. We can observe that FastBO gains an advantage earlier than other methods, rapidly con" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 306, + 693, + 545, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 693, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 306, + 693, + 545, + 712 + ], + "type": "text", + "content": "1Parameters " + }, + { + "bbox": [ + 306, + 693, + 545, + 712 + ], + "type": "inline_equation", + "content": "\\delta_{1}" + }, + { + "bbox": [ + 306, + 693, + 545, + 712 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 306, + 693, + 545, + 712 + ], + "type": "inline_equation", + "content": "\\delta_{2}" + }, + { + "bbox": [ + 306, + 693, + 545, + 712 + ], + "type": "text", + "content": " given here are derived after standardizing metrics to a uniform scale from 0 to 1." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "26186" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 52, + 114, + 542, + 280 + ], + "blocks": [ + { + "bbox": [ + 47, + 70, + 547, + 105 + ], + "lines": [ + { + "bbox": [ + 47, + 70, + 547, + 105 + ], + "spans": [ + { + "bbox": [ + 47, + 70, + 547, + 105 + ], + "type": "text", + "content": "Table 2. Comparison of relative efficiency on configuration identification. FastBO is set as the baseline with a relative efficiency of 1.00. Wall-clock time (abbr. WC time) reports the elapsed time spent for each method on finding configurations with similar performance metrics, i.e., validation error " + }, + { + "bbox": [ + 47, + 70, + 547, + 105 + ], + "type": "inline_equation", + "content": "(\\times 10^{-2})" + }, + { + "bbox": [ + 47, + 70, + 547, + 105 + ], + "type": "text", + "content": " for Covertype and ImageNet16-120 and validation loss " + }, + { + "bbox": [ + 47, + 70, + 547, + 105 + ], + "type": "inline_equation", + "content": "(\\times 10^{-5})" + }, + { + "bbox": [ + 47, + 70, + 547, + 105 + ], + "type": "text", + "content": " for Slice." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 52, + 114, + 542, + 280 + ], + "lines": [ + { + "bbox": [ + 52, + 114, + 542, + 280 + ], + "spans": [ + { + "bbox": [ + 52, + 114, + 542, + 280 + ], + "type": "table", + "html": "
Metric DatasetMethodFastBOBOPASHAA-BOHBA-CQRBOHBDyHPOHyper-Tune
CovertypeVal. error22.9±0.223.0±0.325.1±2.523.5±1.131.6±1.932.5±0.823.0±0.323.0±0.2
WC time (h)0.7±0.32.9±0.73.9±1.02.0±1.03.9±0.22.5±1.01.7±0.61.8±0.7
Rel. efficiency1.000.250.180.370.190.290.410.40
ImageNet 16-120Val. error55.3±0.257.4±1.255.7±0.355.8±1.655.5±0.955.5±1.155.5±1.055.3±2.0
WC time (h)2.2±0.76.6±0.92.5±1.25.9±1.16.0±1.33.2±0.74.3±1.03.4±1.1
Rel. efficiency1.000.340.900.380.370.680.510.67
SliceVal. loss26.3±2.626.4±4.426.8±9.526.3±6.327.1±4.226.8±5.627.4±2.328.7±1.3
WC time (h)0.4±0.13.1±0.71.2±0.92.1±0.72.5±0.72.2±0.92.5±0.51.8±0.6
Rel. efficiency1.000.130.350.200.170.190.170.24
", + "image_path": "58fad66fa14733261cdc6443ac1f82eafe16d6c256249f8a86fa84c46478138d.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 302, + 261, + 315 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 302, + 261, + 315 + ], + "spans": [ + { + "bbox": [ + 47, + 302, + 261, + 315 + ], + "type": "text", + "content": "verging to the global optimum after the initial phase." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 316, + 288, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 316, + 288, + 555 + ], + "spans": [ + { + "bbox": [ + 46, + 316, + 288, + 555 + ], + "type": "text", + "content": "The superiority can be attributed to two main factors. Firstly, FastBO maintains, and in some cases even surpasses, the sample efficiency of vanilla BO, thanks to our techniques that enable quick and precise identification of the fidelities for configurations to update the surrogate model. We provide more explanations and experiments on sample efficiency in Supp. 11.2. Secondly, the multi-fidelity extension speeds up the evaluations, contributing to its overall efficiency. In contrast, the single-fidelity baselines tend to waste more time on the full evaluations. While the multifidelity baselines efficiently explore numerous configurations, they limit their evaluations to only constrained fidelities for some time, thus struggling to provide relatively high performance in a short time. This issue in multi-fidelity methods is particularly pronounced in PASHA when applied to NAS-Bench-201 and FCNet, as shown in Fig. 2. In Supp. 11.3, we further provide the ranks of all methods and statistically show FastBO's superiority on an early stage. It is worth noting that all the additional overhead introduced by FastBO is taken into account in the wall-clock time." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 558, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 558, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 558, + 288, + 715 + ], + "type": "text", + "content": "Regarding the final performance, most methods are able to converge to satisfactory solutions, with negligible differences among them in most cases. Although our goal is not to offer the best final performance as we limit the evaluations to at most the saturation point even for those we consider most promising, FastBO still achieves top-2 final performance on 8 out of 10 datasets. In contrast, model-free methods sometimes cannot obtain a satisfactory final performance because they randomly select the configurations. For example, on the \"Covertype\" dataset, only 3 out of 2000 configurations yield a validation accuracy exceeding " + }, + { + "bbox": [ + 46, + 558, + 288, + 715 + ], + "type": "inline_equation", + "content": "75\\%" + }, + { + "bbox": [ + 46, + 558, + 288, + 715 + ], + "type": "text", + "content": ". As a result, all the model-free methods face challenges in converging to a satisfactory final performance." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 305, + 301, + 524, + 316 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 301, + 524, + 316 + ], + "spans": [ + { + "bbox": [ + 305, + 301, + 524, + 316 + ], + "type": "text", + "content": "5.2. Efficiency on Configuration Identification" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 320, + 545, + 464 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 320, + 545, + 464 + ], + "spans": [ + { + "bbox": [ + 304, + 320, + 545, + 464 + ], + "type": "text", + "content": "One explanation for PASHA's suboptimal anytime performance (cf. Fig. 2) lies in its primary goal [4]: the goal of PASHA is not high accuracy but to identify the best configuration more quickly. To ensure equitable comparisons, we report the time spent for each method on identifying a satisfactory configuration, consistent with the experiments described in PASHA [4]. Results on three expensive datasets \"Covertype\", \"ImageNet16-120\", and \"Slice\" of the three benchmarks are shown in Tab. 2. Similar results on additional datasets can be found in Supp. 11.4. Besides PASHA, results of other model-free multi-fidelity methods are not included, as PASHA demonstrates its superiority over them." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 464, + 546, + 644 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 464, + 546, + 644 + ], + "spans": [ + { + "bbox": [ + 304, + 464, + 546, + 644 + ], + "type": "text", + "content": "Tab. 2 shows that FastBO saves " + }, + { + "bbox": [ + 304, + 464, + 546, + 644 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 304, + 464, + 546, + 644 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 304, + 464, + 546, + 644 + ], + "type": "inline_equation", + "content": "87\\%" + }, + { + "bbox": [ + 304, + 464, + 546, + 644 + ], + "type": "text", + "content": " wall-clock time over other methods when achieving up to " + }, + { + "bbox": [ + 304, + 464, + 546, + 644 + ], + "type": "inline_equation", + "content": "9.6\\%" + }, + { + "bbox": [ + 304, + 464, + 546, + 644 + ], + "type": "text", + "content": " better performance values. It can be observed from the \"rel. efficiency\" rows, where we set FastBO as the baseline with a relative efficiency of 1.00 and report the efficiency of other methods relative to ours. When compared with vanilla BO, FastBO significantly shortens the time in identifying a good configuration by a factor of 3 to 8, because FastBO pauses a configuration earlier at an appropriate fidelity and fits the surrogate model to guide the next configuration search. This advantage creates opportunities to efficiently explore more configurations. Another observation is that PASHA always gets a relatively high variance in wall-clock time. This is due to the fact that different random seeds can have a larger impact on such model-free methods." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 651, + 545, + 665 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 651, + 545, + 665 + ], + "spans": [ + { + "bbox": [ + 305, + 651, + 545, + 665 + ], + "type": "text", + "content": "5.3. Effectiveness of Adaptive Fidelity Identification" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 670, + 545, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 670, + 545, + 696 + ], + "spans": [ + { + "bbox": [ + 304, + 670, + 545, + 696 + ], + "type": "text", + "content": "As discussed in § 4.1, FastBO is able to adaptively identify the efficient point " + }, + { + "bbox": [ + 304, + 670, + 545, + 696 + ], + "type": "inline_equation", + "content": "e_i" + }, + { + "bbox": [ + 304, + 670, + 545, + 696 + ], + "type": "text", + "content": " for each configuration " + }, + { + "bbox": [ + 304, + 670, + 545, + 696 + ], + "type": "inline_equation", + "content": "\\lambda_i" + }, + { + "bbox": [ + 304, + 670, + 545, + 696 + ], + "type": "text", + "content": " and serves" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 315, + 702, + 533, + 714 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 702, + 533, + 714 + ], + "spans": [ + { + "bbox": [ + 315, + 702, + 533, + 714 + ], + "type": "text", + "content": "2We convert the accuracy of \"Covertype\" into error for readability." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "26187" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 39, + 68, + 553, + 167 + ], + "blocks": [ + { + "bbox": [ + 39, + 68, + 553, + 167 + ], + "lines": [ + { + "bbox": [ + 39, + 68, + 553, + 167 + ], + "spans": [ + { + "bbox": [ + 39, + 68, + 553, + 167 + ], + "type": "image", + "image_path": "5871895093db6b0070b54a1377b6b9d466bff987012e8951815add0452787815.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 174, + 546, + 198 + ], + "lines": [ + { + "bbox": [ + 46, + 174, + 546, + 198 + ], + "spans": [ + { + "bbox": [ + 46, + 174, + 546, + 198 + ], + "type": "text", + "content": "Figure 3. Performance comparison: (a) Performance of FastBO that adaptively sets " + }, + { + "bbox": [ + 46, + 174, + 546, + 198 + ], + "type": "inline_equation", + "content": "r_i = e_i" + }, + { + "bbox": [ + 46, + 174, + 546, + 198 + ], + "type": "text", + "content": " with the schemes that use fixed " + }, + { + "bbox": [ + 46, + 174, + 546, + 198 + ], + "type": "inline_equation", + "content": "r_i" + }, + { + "bbox": [ + 46, + 174, + 546, + 198 + ], + "type": "text", + "content": " for all configurations. (b) Performance of single-fidelity methods CQR, BORE, REA and their multi-fidelity variants using our extension method." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 212, + 289, + 355 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 212, + 289, + 355 + ], + "spans": [ + { + "bbox": [ + 46, + 212, + 289, + 355 + ], + "type": "inline_equation", + "content": "e_i" + }, + { + "bbox": [ + 46, + 212, + 289, + 355 + ], + "type": "text", + "content": " as its fidelity " + }, + { + "bbox": [ + 46, + 212, + 289, + 355 + ], + "type": "inline_equation", + "content": "r_i" + }, + { + "bbox": [ + 46, + 212, + 289, + 355 + ], + "type": "text", + "content": " for surrogate model fitting. To investigate the effectiveness of the adaptive fidelity identification strategy, we conduct an ablation study to compare the performance achieved with and without applying this strategy. Specifically, we compare FastBO, where " + }, + { + "bbox": [ + 46, + 212, + 289, + 355 + ], + "type": "inline_equation", + "content": "r_i" + }, + { + "bbox": [ + 46, + 212, + 289, + 355 + ], + "type": "text", + "content": " is adaptively set to " + }, + { + "bbox": [ + 46, + 212, + 289, + 355 + ], + "type": "inline_equation", + "content": "e_i" + }, + { + "bbox": [ + 46, + 212, + 289, + 355 + ], + "type": "text", + "content": ", with the partial evaluation schemes that employ fixed predefined values as the fidelity for all the configurations to fit the surrogate model. We consider three representative fixed fidelities, including " + }, + { + "bbox": [ + 46, + 212, + 289, + 355 + ], + "type": "inline_equation", + "content": "25\\%" + }, + { + "bbox": [ + 46, + 212, + 289, + 355 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 212, + 289, + 355 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 46, + 212, + 289, + 355 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 212, + 289, + 355 + ], + "type": "inline_equation", + "content": "75\\%" + }, + { + "bbox": [ + 46, + 212, + 289, + 355 + ], + "type": "text", + "content": " of the total resource budget. In addition, we include a comparison with vanilla BO that can be viewed as using " + }, + { + "bbox": [ + 46, + 212, + 289, + 355 + ], + "type": "inline_equation", + "content": "100\\%" + }, + { + "bbox": [ + 46, + 212, + 289, + 355 + ], + "type": "text", + "content": " resource budget as the fixed fidelity for all configurations." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 357, + 289, + 622 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 357, + 289, + 622 + ], + "spans": [ + { + "bbox": [ + 46, + 357, + 289, + 622 + ], + "type": "text", + "content": "We provide the results on three representative datasets in Fig. 3(a), with more results available in Supp. 11.5. We have three main observations. Firstly, FastBO always outperforms the partial evaluation baselines that use a fixed fidelity, indicating the effectiveness of the adaptive strategy. Secondly, FastBO shows stronger performance than vanilla BO. The limitation of vanilla BO lies in the additional time required for full evaluations. Secondly, compared to the vanilla BO, partial evaluation schemes with fixed " + }, + { + "bbox": [ + 46, + 357, + 289, + 622 + ], + "type": "inline_equation", + "content": "r_i" + }, + { + "bbox": [ + 46, + 357, + 289, + 622 + ], + "type": "text", + "content": " converge faster in the initial stage due to their ability to evaluate more configurations promptly, but this advantage is gradually offset over time because they fail to find appropriate fidelities to create an accurate surrogate model. This causes a suboptimal final performance compared to vanilla BO, as shown in the first two figures in Figs. 3(a). In the case of the last one, we can observe a noticeable upward trend exhibited by the vanilla BO towards the end of the evaluation, indicating its potential to improve the final performance given abundant time. The comparison between the partial evaluation baselines and vanilla BO also demonstrates the importance of our adaptive strategy, which ensures that the fidelities align optimally with each configuration." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 633, + 287, + 646 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 633, + 287, + 646 + ], + "spans": [ + { + "bbox": [ + 47, + 633, + 287, + 646 + ], + "type": "text", + "content": "5.4. Generality of The Proposed Extension Method" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 654, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 654, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 654, + 287, + 715 + ], + "type": "text", + "content": "The adaptive fidelity identification strategy provides a simple way to extend single-fidelity methods to the multifidelity setting, as discussed in § 4.4. To examine the ability of our extension method, we conduct experiments using three popular single-fidelity methods CQR [37], BORE [45]" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 212, + 547, + 416 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 212, + 547, + 416 + ], + "spans": [ + { + "bbox": [ + 304, + 212, + 547, + 416 + ], + "type": "text", + "content": "and REA [35], extending them to the multi-fidelity variants with our extension method, referred to as FastCQR, FastBORE, and FastREA respectively. Similar to FastBO, all the multi-fidelity extensions evaluate the configurations to the adaptively identified efficient point and use the corresponding performances for the subsequent operations. The results on three datasets are illustrated in Fig. 3(b) and similar results on other datasets are in Supp. 11.6. We can clearly observe that the multi-fidelity variants with our extension method always outperform their single-fidelity counterparts. It is worth noting that REA is an evolutionary algorithm-based HPO method and is also significantly improved by our extension. The observation highlights the ability of the proposed adaptive strategy to extend any single-fidelity method to the multi-fidelity setting. It also suggests future opportunities to extend other advanced single-fidelity techniques into the multi-fidelity setting." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 430, + 378, + 441 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 430, + 378, + 441 + ], + "spans": [ + { + "bbox": [ + 306, + 430, + 378, + 441 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 450, + 547, + 631 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 450, + 547, + 631 + ], + "spans": [ + { + "bbox": [ + 304, + 450, + 547, + 631 + ], + "type": "text", + "content": "In this paper, we propose a model-based multi-fidelity HPO method FastBO, which adaptively identifies the appropriate fidelity for each configuration to fit the surrogate model and offers high-quality performance while ensuring efficient resource utilization. The advantages are achieved through our concepts of efficient and saturation point, the proposed techniques of learning curve modeling, and well-designed warm-up and post-processing stages with judicious early-termination detection and efficient saturation-level evaluation. Moreover, the proposed adaptive fidelity identification strategy provides a simple way to extend any single-fidelity method to the multi-fidelity setting. Experiments demonstrate the effectiveness and wide generality of our proposed techniques. FastBO source code is freely available at https://github.com/jjiantong/FastBO." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 644, + 399, + 658 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 644, + 399, + 658 + ], + "spans": [ + { + "bbox": [ + 306, + 644, + 399, + 658 + ], + "type": "text", + "content": "Acknowledgment" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "type": "text", + "content": "This research was funded by ARC Grant number DP190102443. Ajmal Mian is the recipient of an Australian Research Council Future Fellowship Award (project number FT210100268) funded by the Australian Government." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "26188" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 135 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 135 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 135 + ], + "type": "text", + "content": "[1] Sebastian E Ament and Carla P Gomes. Scalable first-order Bayesian Optimization via structured automatic differentiation. In International Conference on Machine Learning, pages 500-516. PMLR, 2022. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 137, + 288, + 179 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 137, + 288, + 179 + ], + "spans": [ + { + "bbox": [ + 53, + 137, + 288, + 179 + ], + "type": "text", + "content": "[2] James Bergstra, Rémi Bardenet, Yoshua Bengio, and Balázs Kégl. Algorithms for hyper-parameter optimization. Advances in Neural Information Processing Systems, 24, 2011. 1, 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 182, + 288, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 182, + 288, + 258 + ], + "spans": [ + { + "bbox": [ + 53, + 182, + 288, + 258 + ], + "type": "text", + "content": "[3] Bernd Bischl, Martin Binder, Michel Lang, Tobias Pielok, Jakob Richter, Stefan Coors, Janek Thomas, Theresa Ullmann, Marc Becker, Anne-Laure Boulesteix, et al. Hyperparameter optimization: Foundations, algorithms, best practices, and open challenges. Wiley Interdisciplinary Reviews: Data Mining and Knowledge Discovery, 13(2):e1484, 2023. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 260, + 287, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 260, + 287, + 315 + ], + "spans": [ + { + "bbox": [ + 53, + 260, + 287, + 315 + ], + "type": "text", + "content": "[4] Ondrej Bohdal, Lukas Balles, Martin Wistuba, Beyza Ermis, Cedric Archambeau, and Giovanni Zappella. PASHA: efficient HPO and NAS with progressive resource allocation. In International Conference on Learning Representations. OpenReview.net, 2023. 1, 2, 6, 7, 4, 9" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 317, + 288, + 361 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 317, + 288, + 361 + ], + "spans": [ + { + "bbox": [ + 53, + 317, + 288, + 361 + ], + "type": "text", + "content": "[5] Chaofan Chen, Oscar Li, Daniel Tao, Alina Barnett, Cynthia Rudin, and Jonathan K Su. This looks like that: deep learning for interpretable image recognition. Advances in Neural Information Processing Systems, 32, 2019. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 362, + 288, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 362, + 288, + 407 + ], + "spans": [ + { + "bbox": [ + 53, + 362, + 288, + 407 + ], + "type": "text", + "content": "[6] Tianqi Chen and Carlos Guestrin. Xgboost: A scalable tree boosting system. In Proceedings of the 22nd acm sigkdd international conference on knowledge discovery and data mining, pages 785-794, 2016. 9" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 407, + 287, + 461 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 407, + 287, + 461 + ], + "spans": [ + { + "bbox": [ + 53, + 407, + 287, + 461 + ], + "type": "text", + "content": "[7] Tobias Domhan, Jost Tobias Springenberg, and Frank Hutter. Speeding up automatic hyperparameter optimization of deep neural networks by extrapolation of learning curves. In International Joint Conference on Artificial Intelligence, 2015. 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 464, + 288, + 507 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 464, + 288, + 507 + ], + "spans": [ + { + "bbox": [ + 53, + 464, + 288, + 507 + ], + "type": "text", + "content": "[8] Xuanyi Dong and Yi Yang. NAS-Bench-201: Extending the scope of reproducible neural architecture search. In International Conference on Learning Representations, 2020. 1, 6, 8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 510, + 287, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 510, + 287, + 542 + ], + "spans": [ + { + "bbox": [ + 53, + 510, + 287, + 542 + ], + "type": "text", + "content": "[9] Thomas Elsken, Jan Hendrik Metzen, and Frank Hutter. Neural architecture search: A survey. The Journal of Machine Learning Research, 20(1):1997-2017, 2019. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 544, + 287, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 544, + 287, + 588 + ], + "spans": [ + { + "bbox": [ + 48, + 544, + 287, + 588 + ], + "type": "text", + "content": "[10] Stefan Falkner, Aaron Klein, and Frank Hutter. BOHB: Robust and efficient hyperparameter optimization at scale. In International Conference on Machine Learning, pages 1437-1446. PMLR, 2018. 1, 2, 6, 9" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 590, + 287, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 590, + 287, + 623 + ], + "spans": [ + { + "bbox": [ + 48, + 590, + 287, + 623 + ], + "type": "text", + "content": "[11] Matthias Feurer and Frank Hutter. Hyperparameter optimization. Automated Machine Learning: Methods, Systems, Challenges, pages 3-33, 2019. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 624, + 287, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 624, + 287, + 667 + ], + "spans": [ + { + "bbox": [ + 48, + 624, + 287, + 667 + ], + "type": "text", + "content": "[12] Peter I Frazier, Warren B Powell, and Savas Dayanik. A knowledge-gradient policy for sequential information collection. SIAM Journal on Control and Optimization, 47(5): 2410-2439, 2008. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 669, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 669, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 669, + 287, + 713 + ], + "type": "text", + "content": "[13] José Miguel Hernández-Lobato, Matthew W Hoffman, and Zoubin Ghahramani. Predictive entropy search for efficient global optimization of black-box functions. Advances in Neural Information Processing systems, 27, 2014. 2" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 713 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 117 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 117 + ], + "type": "text", + "content": "[14] Frank Hutter, Holger H Hoos, and Kevin Leyton-Brown. Sequential model-based optimization for general algorithm configuration. In Learning and Intelligent Optimization, pages 507-523. Springer, 2011. 1, 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 118, + 547, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 547, + 172 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 547, + 172 + ], + "type": "text", + "content": "[15] Carl Hvarfner, Danny Stoll, Artur L. F. Souza, Marius Lindauer, Frank Hutter, and Luigi Nardi. " + }, + { + "bbox": [ + 307, + 118, + 547, + 172 + ], + "type": "inline_equation", + "content": "\\S" + }, + { + "bbox": [ + 307, + 118, + 547, + 172 + ], + "type": "text", + "content": " pi" + }, + { + "bbox": [ + 307, + 118, + 547, + 172 + ], + "type": "inline_equation", + "content": "\\S" + }, + { + "bbox": [ + 307, + 118, + 547, + 172 + ], + "type": "text", + "content": "BO: Augmenting acquisition functions with user beliefs for bayesian optimization. In International Conference on Learning Representations. OpenReview.net, 2022. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 173, + 545, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 173, + 545, + 215 + ], + "spans": [ + { + "bbox": [ + 307, + 173, + 545, + 215 + ], + "type": "text", + "content": "[16] Kevin Jamieson and Ameet Talwalkar. Non-stochastic best arm identification and hyperparameter optimization. In Artificial Intelligence and Statistics, pages 240-248. PMLR, 2016. 1, 2, 9" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 217, + 545, + 250 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 217, + 545, + 250 + ], + "spans": [ + { + "bbox": [ + 307, + 217, + 545, + 250 + ], + "type": "text", + "content": "[17] Kasra Jamshidi, Harry Xu, and Keval Vora. Accelerating graph mining systems with subgraph morphing. In European Conference on Computer Systems, pages 162-181, 2023. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 251, + 545, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 251, + 545, + 293 + ], + "spans": [ + { + "bbox": [ + 307, + 251, + 545, + 293 + ], + "type": "text", + "content": "[18] Jiantong Jiang, Zeyi Wen, Zeke Wang, Bingsheng He, and Jian Chen. Parallel and distributed structuredsvm training. IEEE Transactions on Parallel and Distributed Systems, 33 (5):1084-1096, 2021." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 294, + 545, + 337 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 294, + 545, + 337 + ], + "spans": [ + { + "bbox": [ + 307, + 294, + 545, + 337 + ], + "type": "text", + "content": "[19] Jiantong Jiang, Zeyi Wen, and Ajmal Mian. Fast parallel bayesian network structure learning. In IEEE International Parallel and Distributed Processing Symposium, pages 617-627. IEEE, 2022. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 338, + 545, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 338, + 545, + 380 + ], + "spans": [ + { + "bbox": [ + 307, + 338, + 545, + 380 + ], + "type": "text", + "content": "[20] Arlind Kadra, Maciej Janowski, Martin Wistuba, and Josif Grabocka. Scaling laws for hyperparameter optimization. In Advances in Neural Information Processing Systems, 2023. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 383, + 545, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 383, + 545, + 415 + ], + "spans": [ + { + "bbox": [ + 307, + 383, + 545, + 415 + ], + "type": "text", + "content": "[21] Aaron Klein and Frank Hutter. Tabular benchmarks for joint architecture and hyperparameter optimization. arXiv preprint arXiv:1905.04970, 2019. 6, 8" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 416, + 545, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 416, + 545, + 459 + ], + "spans": [ + { + "bbox": [ + 307, + 416, + 545, + 459 + ], + "type": "text", + "content": "[22] Aaron Klein, Louis C Tiao, Thibaut Lienart, Cedric Archambeau, and Matthias Seeger. Model-based asynchronous hyperparameter and neural architecture search. arXiv preprint arXiv:2003.10865, 2020. 1, 2, 6, 9" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 460, + 545, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 460, + 545, + 514 + ], + "spans": [ + { + "bbox": [ + 307, + 460, + 545, + 514 + ], + "type": "text", + "content": "[23] Prasanth Kolachina, Nicola Cancedda, Marc Dymetman, and Sriram Venkatapathy. Prediction of learning curves in machine translation. In Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 22-30, 2012. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 515, + 545, + 581 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 515, + 545, + 581 + ], + "spans": [ + { + "bbox": [ + 307, + 515, + 545, + 581 + ], + "type": "text", + "content": "[24] Cheng Li, Santu Rana, Sunil Gupta, Vu Nguyen, Svetha Venkatesh, Alessandra Sutti, David Rubin de Celis Leal, Teo Slezak, Murray Height, Mazher Mohammed, and Ian Gibson. Accelerating experimental design by incorporating experimenter hunches. In International Conference on Data Mining, pages 257-266. IEEE Computer Society, 2018. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 582, + 545, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 582, + 545, + 635 + ], + "spans": [ + { + "bbox": [ + 307, + 582, + 545, + 635 + ], + "type": "text", + "content": "[25] Lisha Li, Kevin Jamieson, Giulia DeSalvo, Afshin Rostamizadeh, and Ameet Talwalkar. Hyperband: A novel bandit-based approach to hyperparameter optimization. Journal of Machine Learning Research, 18(1):6765-6816, 2017. 1, 2, 6, 9" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 636, + 545, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 636, + 545, + 690 + ], + "spans": [ + { + "bbox": [ + 307, + 636, + 545, + 690 + ], + "type": "text", + "content": "[26] Liam Li, Kevin Jamieson, Afshin Rostamizadeh, Ekaterina Gonina, Jonathan Ben-Tzur, Moritz Hardt, Benjamin Recht, and Ameet Talwalkar. A system for massively parallel hyperparameter tuning. Proceedings of Machine Learning and Systems, 2:230-246, 2020. 1, 2, 6, 9" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 691, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 691, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 691, + 545, + 713 + ], + "type": "text", + "content": "[27] Yang Li, Yu Shen, Huaijun Jiang, Wentao Zhang, Jixiang Li, Ji Liu, Ce Zhang, and Bin Cui. Hyper-tune: Towards" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "26189" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 66, + 72, + 287, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 72, + 287, + 95 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 287, + 95 + ], + "type": "text", + "content": "efficient hyper-parameter tuning at scale. Proceedings of the VLDB Endowment, 15(6):1256-1265, 2022. 1, 2, 6, 9" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 96, + 287, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 96, + 287, + 128 + ], + "spans": [ + { + "bbox": [ + 48, + 96, + 287, + 128 + ], + "type": "text", + "content": "[28] Jonas Mockus. The application of Bayesian methods for seeking the extremum. Towards global optimization, 2:117, 1998. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 130, + 287, + 163 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 130, + 287, + 163 + ], + "spans": [ + { + "bbox": [ + 48, + 130, + 287, + 163 + ], + "type": "text", + "content": "[29] Felix Mohr and Jan N van Rijn. Learning curves for decision making in supervised machine learning-a survey. arXiv preprint arXiv:2201.12150, 2022. 4, 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 164, + 287, + 219 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 164, + 287, + 219 + ], + "spans": [ + { + "bbox": [ + 48, + 164, + 287, + 219 + ], + "type": "text", + "content": "[30] Felix Mohr, Tom J Viering, Marco Loog, and Jan N van Rijn. Lcdb 1.0: An extensive learning curves database for classification tasks. In Joint European Conference on Machine Learning and Knowledge Discovery in Databases, pages 3-19. Springer, 2022. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 220, + 287, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 220, + 287, + 274 + ], + "spans": [ + { + "bbox": [ + 48, + 220, + 287, + 274 + ], + "type": "text", + "content": "[31] Julia Moosbauer, Julia Herbinger, Giuseppe Casalicchio, Marius Lindauer, and Bernd Bischl. Explaining hyperparameter optimization via partial dependence plots. Advances in Neural Information Processing Systems, 34:2280-2291, 2021. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 276, + 287, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 276, + 287, + 330 + ], + "spans": [ + { + "bbox": [ + 48, + 276, + 287, + 330 + ], + "type": "text", + "content": "[32] Julia Moosbauer, Giuseppe Casalicchio, Marius Lindauer, and Bernd Bischl. Improving accuracy of interpretability measures in hyperparameter optimization via Bayesian algorithm execution. arXiv preprint arXiv:2206.05447, 2022. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 332, + 287, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 332, + 287, + 376 + ], + "spans": [ + { + "bbox": [ + 48, + 332, + 287, + 376 + ], + "type": "text", + "content": "[33] ChangYong Oh, Efstratios Gavves, and Max Welling. BOCK: Bayesian optimization with cylindrical kernels. In International Conference on Machine Learning, pages 3868-3877. PMLR, 2018. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 377, + 287, + 421 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 377, + 287, + 421 + ], + "spans": [ + { + "bbox": [ + 48, + 377, + 287, + 421 + ], + "type": "text", + "content": "[34] Misha Padidar, Xinran Zhu, Leo Huang, Jacob Gardner, and David Bindel. Scaling gaussian processes with derivative information using variational inference. Advances in Neural Information Processing Systems, 34:6442-6453, 2021. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 422, + 287, + 466 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 422, + 287, + 466 + ], + "spans": [ + { + "bbox": [ + 48, + 422, + 287, + 466 + ], + "type": "text", + "content": "[35] Esteban Real, Alok Aggarwal, Yanping Huang, and Quoc V Le. Regularized evolution for image classifier architecture search. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 4780-4789, 2019. 8, 5, 9" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 468, + 287, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 468, + 287, + 521 + ], + "spans": [ + { + "bbox": [ + 48, + 468, + 287, + 521 + ], + "type": "text", + "content": "[36] David Salinas, Matthias Seeger, Aaron Klein, Valerio Perrone, Martin Wistuba, and Cedric Archambeau. Syne tune: A library for large scale hyperparameter tuning and reproducible research. In International Conference on Automated Machine Learning, pages 16-1. PMLR, 2022. 6, 9" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 523, + 287, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 523, + 287, + 578 + ], + "spans": [ + { + "bbox": [ + 48, + 523, + 287, + 578 + ], + "type": "text", + "content": "[37] David Salinas, Jacek Golebiowski, Aaron Klein, Matthias W. Seeger, and Cedric Archambeau. Optimizing hyperparameters with conformal quantile regression. In International Conference on Machine Learning, pages 29876-29893. PMLR, 2023. 1, 2, 6, 8, 5, 9" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 579, + 287, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 579, + 287, + 622 + ], + "spans": [ + { + "bbox": [ + 48, + 579, + 287, + 622 + ], + "type": "text", + "content": "[38] Bobak Shahriari, Alexandre Bouchard-Côté, and Nando Freitas. Unbounded Bayesian Optimization via regularization. In Artificial intelligence and statistics, pages 1168-1176. PMLR, 2016. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 624, + 287, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 624, + 287, + 667 + ], + "spans": [ + { + "bbox": [ + 48, + 624, + 287, + 667 + ], + "type": "text", + "content": "[39] Julien Siems, Lucas Zimmer, Arber Zela, Jovita Lukasik, Margret Keuper, and Frank Hutter. Nas-bench-301 and the case for surrogate benchmarks for neural architecture search. CoRR, abs/2008.09777, 2020. 1, 6, 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "type": "text", + "content": "[40] Jasper Snoek, Hugo Larochelle, and Ryan P Adams. Practical Bayesian optimization of machine learning algorithms. Advances in Neural Information Processing Systems, 25, 2012. 1, 2, 6, 9" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 643 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 128 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 128 + ], + "type": "text", + "content": "[41] Jasper Snoek, Oren Rippel, Kevin Swersky, Ryan Kiros, Nadathur Satish, Narayanan Sundaram, Mostofa Patwary, Mr Prabhat, and Ryan Adams. Scalable Bayesian optimization using deep neural networks. In International Conference on Machine Learning, pages 2171-2180. PMLR, 2015. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 129, + 545, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 129, + 545, + 173 + ], + "spans": [ + { + "bbox": [ + 307, + 129, + 545, + 173 + ], + "type": "text", + "content": "[42] Jost Tobias Springenberg, Aaron Klein, Stefan Falkner, and Frank Hutter. Bayesian optimization with robust bayesian neural networks. Advances in Neural Information Processing Systems, 29, 2016. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 175, + 545, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 175, + 545, + 217 + ], + "spans": [ + { + "bbox": [ + 307, + 175, + 545, + 217 + ], + "type": "text", + "content": "[43] Niranjan Srinivas, Andreas Krause, Sham M Kakade, and Matthias Seeger. Gaussian process optimization in the bandit setting: No regret and experimental design. arXiv preprint arXiv:0912.3995, 2009. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 219, + 545, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 219, + 545, + 251 + ], + "spans": [ + { + "bbox": [ + 307, + 219, + 545, + 251 + ], + "type": "text", + "content": "[44] Kevin Swersky, Jasper Snoek, and Ryan Prescott Adams. Freeze-thaw Bayesian optimization. arXiv preprint arXiv:1406.3896, 2014. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 252, + 545, + 306 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 252, + 545, + 306 + ], + "spans": [ + { + "bbox": [ + 307, + 252, + 545, + 306 + ], + "type": "text", + "content": "[45] Louis C Tiao, Aaron Klein, Matthias W Seeger, Edwin V Bonilla, Cedric Archambeau, and Fabio Ramos. BORE: Bayesian optimization by density-ratio estimation. In International Conference on Machine Learning, pages 10289-10300. PMLR, 2021. 8, 5, 9" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 308, + 545, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 308, + 545, + 342 + ], + "spans": [ + { + "bbox": [ + 307, + 308, + 545, + 342 + ], + "type": "text", + "content": "[46] Tom Viering and Marco Loog. The shape of learning curves: a review. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2022. 1, 4, 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 343, + 545, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 343, + 545, + 385 + ], + "spans": [ + { + "bbox": [ + 307, + 343, + 545, + 385 + ], + "type": "text", + "content": "[47] Jiazhuo Wang, Jason Xu, and Xuejun Wang. Combination of hyperband and bayesian optimization for hyperparameter optimization in deep learning. arXiv preprint arXiv:1801.01596, 2018. 1, 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 387, + 545, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 387, + 545, + 430 + ], + "spans": [ + { + "bbox": [ + 307, + 387, + 545, + 430 + ], + "type": "text", + "content": "[48] Martin Wistuba, Arlind Kadra, and Josif Grabocka. Supervising the multi-fidelity race of hyperparameter configurations. Advances in Neural Information Processing Systems, 35:13470-13484, 2022. 2, 6, 9" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 432, + 545, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 432, + 545, + 464 + ], + "spans": [ + { + "bbox": [ + 307, + 432, + 545, + 464 + ], + "type": "text", + "content": "[49] Jian Wu, Matthias Poloczek, Andrew G Wilson, and Peter Frazier. Bayesian optimization with gradients. Advances in neural information processing systems, 30, 2017. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 466, + 545, + 508 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 466, + 545, + 508 + ], + "spans": [ + { + "bbox": [ + 307, + 466, + 545, + 508 + ], + "type": "text", + "content": "[50] Peiyu Yang, Naveed Akhtar, Zeyi Wen, Mubarak Shah, and Ajmal Mian. Re-calibrating feature attributions for model interpretation. In International Conference on Learning Representations, 2022. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 510, + 545, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 510, + 545, + 553 + ], + "spans": [ + { + "bbox": [ + 307, + 510, + 545, + 553 + ], + "type": "text", + "content": "[51] Peiyu Yang, Naveed Akhtar, Zeyi Wen, and Ajmal Mian. Local path integration for attribution. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 3173-3180, 2023. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 555, + 545, + 598 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 555, + 545, + 598 + ], + "spans": [ + { + "bbox": [ + 307, + 555, + 545, + 598 + ], + "type": "text", + "content": "[52] Xueying Zhu, Jie Sun, Zhenhao He, Jiantong Jiang, and Zeke Wang. Staleness-reduction mini-batch " + }, + { + "bbox": [ + 307, + 555, + 545, + 598 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 307, + 555, + 545, + 598 + ], + "type": "text", + "content": "-means. IEEE Transactions on Neural Networks and Learning Systems, 2023. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 600, + 545, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 600, + 545, + 643 + ], + "spans": [ + { + "bbox": [ + 307, + 600, + 545, + 643 + ], + "type": "text", + "content": "[53] Lucas Zimmer, Marius Thomas Lindauer, and Frank Hutter. Auto-Pytorch: Multi-fidelity metalearning for efficient and robust AutoDL. IEEE Transactions on Pattern Analysis and Machine Intelligence, 43:3079-3090, 2021. 6, 8" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "26190" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Efficient LoFTR_ Semi-Dense Local Feature Matching with Sparse-Like Speed/14a55345-a1a3-4ac4-a5a2-7c153c54b832_content_list.json b/2024/Efficient LoFTR_ Semi-Dense Local Feature Matching with Sparse-Like Speed/14a55345-a1a3-4ac4-a5a2-7c153c54b832_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..cf01ecc96e647def945a75148bb993a6a0d3fb9a --- /dev/null +++ b/2024/Efficient LoFTR_ Semi-Dense Local Feature Matching with Sparse-Like Speed/14a55345-a1a3-4ac4-a5a2-7c153c54b832_content_list.json @@ -0,0 +1,1532 @@ +[ + { + "type": "text", + "text": "Efficient LoFTR: Semi-Dense Local Feature Matching with Sparse-Like Speed", + "text_level": 1, + "bbox": [ + 88, + 130, + 880, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yifan Wang* Xingyi He* Sida Peng† Dongli Tan Xiaowei Zhou", + "bbox": [ + 200, + 172, + 781, + 191 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Zhejiang University", + "bbox": [ + 390, + 205, + 553, + 224 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 243, + 252, + 321, + 268 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We present a novel method for efficiently producing semidense matches across images. Previous detector-free matcher LoFTR has shown remarkable matching capability in handling large-viewpoint change and texture-poor scenarios but suffers from low efficiency. We revisit its design choices and derive multiple improvements for both efficiency and accuracy. One key observation is that performing the transformer over the entire feature map is redundant due to shared local information, therefore we propose an aggregated attention mechanism with adaptive token selection for efficiency. Furthermore, we find spatial variance exists in LoFTR's fine correlation module, which is adverse to matching accuracy. A novel two-stage correlation layer is proposed to achieve accurate subpixel correspondences for accuracy improvement. Our efficiency optimized model is $\\sim 2.5\\times$ faster than LoFTR which can even surpass state-of-the-art efficient sparse matching pipeline SuperPoint + LightGlue. Moreover, extensive experiments show that our method can achieve higher accuracy compared with competitive semi-dense matchers, with considerable efficiency benefits. This opens up exciting prospects for large-scale or latency-sensitive applications such as image retrieval and 3D reconstruction. Project page: https://zju3dv.github.io/efficientloftr/.", + "bbox": [ + 76, + 279, + 473, + 642 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 667, + 209, + 684 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Image matching is the cornerstone of many 3D computer vision tasks, which aim to find a set of highly accurate correspondences given an image pair. The established matches between images have broad usages such as reconstructing the 3D world by structure from motion (SfM) [1, 18, 25, 43] or SLAM system [30, 31], and visual localization [38, 40], etc. Previous methods typically follow a two-stage pipeline: they first detect [37] and describe [49] a set of keypoints on each image, and then establish keypoint correspondences by handcrafted [27] or learning-based matchers [26, 39]. These detector-based methods are efficient but suffer from robustly", + "bbox": [ + 75, + 694, + 470, + 861 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/dcc10e8a010fd52aed100c9bace3e941ad983eee32cb58fdce6c1aee64e93e25.jpg", + "image_caption": [ + "Figure 1. Matching Accuracy and Efficiency Comparisons. Our method achieves competitive accuracy compared with semi-dense matchers $(\\odot)$ at a significantly higher speed. Compared with dense matcher ROMA $(\\odot)$ , our method is $\\sim 7.5\\times$ faster. Moreover, our efficiency optimized model $(\\star)$ can surpass the robust sparse matching pipeline $(\\odot)$ SuperPoint (SP) + LightGlue (LG) on efficiency with considerably better accuracy." + ], + "image_footnote": [], + "bbox": [ + 506, + 252, + 890, + 515 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "detecting repeatable keypoints across challenging pairs, such as extreme viewpoint changes and texture-poor regions.", + "bbox": [ + 496, + 626, + 892, + 656 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recently, LoFTR [46] introduces a detector-free matching paradigm with transformer to directly establish semi-dense correspondences between two images without detecting keypoints. With the help of the transformer mechanism to capture the global image context and the detector-free design, LoFTR shows a strong capability of matching challenging pairs, especially in texture-poor scenarios. To reduce the computation burden, LoFTR adopts a coarse-to-fine pipeline by first performing dense matching on downsampled coarse features maps, where transformer is applied. Then, the feature locations of coarse matches on one image are fixed, while their subpixel correspondences are searched on the other image by cropping feature patches based on coarse match, performing the feature correlation, and calculating expectation over the correlation patch.", + "bbox": [ + 496, + 657, + 893, + 883 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Despite its impressive matching performance, LoFTR", + "bbox": [ + 517, + 885, + 890, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 44 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Equal contribution. † Corresponding author: Sida Peng. The authors from Zhejiang University are affiliated with the State Key Lab of CAD&CG.", + "bbox": [ + 76, + 875, + 468, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "21666", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "suffers from limited efficiency due to the large token size of performing transformer on the entire coarse feature map, which significantly barricades practical large-scale usages such as image retrieval [16] and SfM [43]. A large bunch of LoFTR's follow-up works [6, 15, 32, 48, 55] have attempted to improve its matching accuracy. However, there are rare methods that focus on matching efficiency of detector-free matching. QuadTree Attention [48] incorporates multi-scale transformation with a gradually narrowed attention span to avoid performing attention on large feature maps. This strategy can reduce the computation cost, but it also divides a single coarse attention process into multiple steps, leading to increased latency.", + "bbox": [ + 75, + 90, + 472, + 287 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we revisit the design decisions of the detector-free matcher LoFTR, and propose a new matching algorithm that squeezes out redundant computations for significantly better efficiency while further improving the accuracy. As shown in Fig. 1, our approach achieves the best inference speed compared with recent image matching methods while being competitive in terms of accuracy. Our key innovations lie in introducing a token aggregation mechanism for efficient feature transformation and a two-stage correlation layer for correspondence refinement. Specifically, we find that densely performing global attention over the entire coarse feature map as in LoFTR is unnecessary, as the attention information is similar and shared in the local region. Therefore, we devise an aggregated attention mechanism to perform feature transformation on adaptively selected tokens, which is significantly compact and effectively reduces the cost of local feature transformation.", + "bbox": [ + 75, + 287, + 472, + 545 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In addition, we observe that there can be spatial variance in the matching refinement phase of LoFTR, which is caused by the expectation over the entire correlation patch when noisy feature correlation exists. To solve this issue, our approach designs a two-stage correlation layer that first locates pixel-level matches with the accurate mutual-nearest-neighbor matching on fine feature patches, and then further refines matches for subpixel-level by performing the correlation and expectation locally within tiny patches.", + "bbox": [ + 75, + 547, + 472, + 683 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Extensive experiments are conducted on multiple tasks, including homography estimation, relative pose recovery, as well as visual localization, to show the efficacy of our method. Our pipeline pushes detector-free matching to unprecedented efficiency, which is $\\sim 2.5$ times faster than LoFTR and can even surpass the current state-of-the-art efficient sparse matcher LightGlue [26]. Moreover, our framework can achieve comparable or even better matching accuracy compared with competitive detector-free baselines [6, 13, 14] with considerably higher efficiency.", + "bbox": [ + 75, + 685, + 472, + 835 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In summary, this paper has the following contributions:", + "bbox": [ + 94, + 838, + 460, + 853 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "- A new detector-free matching pipeline with multiple improvements based on the comprehensive revisiting of LoFTR, which is significantly more efficient and with", + "bbox": [ + 76, + 854, + 472, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "better accuracy.", + "bbox": [ + 511, + 92, + 620, + 106 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- A novel aggregated attention network for efficient local feature transformation.", + "- A novel two-stage correlation refinement layer for accurate and subpixel-level refined correspondences." + ], + "bbox": [ + 500, + 106, + 890, + 166 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 185, + 640, + 200 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Detector-Based Image Matching. Classical image matching methods [3, 27, 37] adopt handcrafted critics for detecting keypoints, describing and then matching them. Recent methods draw benefits from deep neural networks for both detection [21, 37, 42] and description [12, 29, 49, 50], where the robustness and discriminativeness of local descriptors are significantly improved. Besides, some methods [9, 11, 28, 34, 51] managed to learn the detector and descriptor together. SuperGlue [39] is a pioneering method that first introduces the transformer mechanism into matching, which has shown notable improvements over classical handcrafted matchers. As a side effect, it also costs more time, especially with many keypoints to match. To improve the efficiency, some subsequent works, such as [5, 44], endeavor to reduce the size of the attention mechanism, albeit at the cost of sacrificing performance. LightGlue [26] introduces a new scheme for efficient sparse matching that is adaptive to the matching difficulty, where the attention process can be stopped earlier for easy pairs. It is faster than SuperGlue and can achieve competitive performance. However, robustly detecting keypoints across images is still challenging, especially for texture-poor regions. Unlike them, our method focuses on the efficiency of the detector-free method, which eliminates the restriction of keypoint detection and shows superior performance for challenging pairs.", + "bbox": [ + 496, + 212, + 893, + 590 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Detector-Free Image Matching. Detector-free methods directly match images instead of relying on a set of detected keypoints, producing semi-dense or dense matches. NC-Net [35] represents all features and possible matches as a 4D correlation volume. Sparse NC-Net [36] utilizes sparse correlation layers to ease resolution limitations. Subsequently, DRC-Net [23] improves efficiency and further improves performance in a coarse-to-fine manner.", + "bbox": [ + 496, + 595, + 893, + 715 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "LoFTR [46] first employs the Transformer in detector-free matching to model the long-range dependencies. It shows remarkable matching capabilities, however, suffers from low efficiency due to the huge computation of densely transforming entire coarse feature maps. Many follow-up works further improve the matching accuracy. Matchformer [55] and AspanFormer [6] perform attention on multiscale features, where local attention regions of [6] are found with the help of estimated flow. QuadTree [48] gradually restricts the attention span during hierarchical attention to relevant areas, which can reduce overall computation. However, these designs contribute marginally or even decrease", + "bbox": [ + 496, + 719, + 895, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "21667", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/b0dbd617096a4e6fca43c86b88ec1450b28e3b789be6a54b6d8f574d4097a7e8.jpg", + "image_caption": [ + "Figure 2. Pipeline Overview. (1) Given an image pair, a CNN network extracts coarse feature maps $\\tilde{\\mathbf{F}}_A$ and $\\tilde{\\mathbf{F}}_B$ , as well as fine features. (2) Then, we transform coarse features for more discriminative feature maps by interleaving our aggregated self- and cross-attention $N$ times, where adaptively feature aggregation is performed to reduce token size before each attention for efficiency. (3) Transformed coarse features are correlated for the score matrix $S$ . Mutual-nearest-neighbor (MNN) searching is followed to establish coarse matches $\\{\\mathcal{M}_c\\}$ . (4) To refine coarse matches, discriminative fine features $\\hat{\\mathbf{F}}_A^t$ , $\\hat{\\mathbf{F}}_B^t$ in full resolution are obtained by fusing transformed coarse features $\\tilde{\\mathbf{F}}_A^t$ , $\\tilde{\\mathbf{F}}_B^t$ with backbone features. Feature patches are then cropped centered at each coarse match $\\mathcal{M}_c$ . A two-stage refinement is followed to obtain sub-pixel correspondence $\\mathcal{M}_f$ ." + ], + "image_footnote": [], + "bbox": [ + 76, + 82, + 893, + 357 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "efficiency, since the hierarchical nature of multi-scale attention will further introduce latencies. TopicFM [15] first assigns features with similar semantic meanings to the same topic, where attention is conducted within each topic for efficiency. Since it needs to sequentially process each token's features for transformation, the efficiency improvement is limited. Moreover, performing local attention within topics can potentially restrict the capability of modeling long-range dependencies. Compared with them, the proposed aggregated attention module in our method significantly improves efficiency while achieving better accuracy.", + "bbox": [ + 75, + 486, + 470, + 652 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Dense matching methods [13, 14, 52] are designed to estimate all possible correspondences between two images, which show strong robustness. However, they are generally much slower compared with sparse and semi-dense methods. Unlike them, our method produces semi-dense matches with competitive performance and considerably better efficiency.", + "bbox": [ + 75, + 652, + 470, + 744 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Transformer has been broadly used in multiple vision tasks, including feature matching. The efficiency and memory footprint of handling large token sizes are the main limitations of transformer [54], where some methods [19, 20, 56] attempt to reduce the complexity to a linear scale to alleviate these problems. Some methods [8, 22] propose optimizing transformer models for specific hardware architectures for memory and running-time efficiency. They are orthogonal to our method and can be naturally adapted into the pipeline for further efficiency improvement.", + "bbox": [ + 75, + 750, + 472, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 500, + 484, + 589, + 500 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given a pair of images $\\mathbf{I}_A, \\mathbf{I}_B$ , our objective is to establish a set of reliable correspondences between them. We achieve this by a coarse-to-fine matching pipeline, which first establishes coarse matches on downsampled feature maps and then refines them for high accuracy. An overview of our pipeline is shown in Fig. 2.", + "bbox": [ + 496, + 513, + 893, + 606 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Local Feature Extraction", + "text_level": 1, + "bbox": [ + 500, + 627, + 728, + 642 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Image feature maps are first extracted by a lightweight backbone for later transformation and matching. Unlike LoFTR and many other detector-free matchers that use a heavy multibranch ResNet [17] network for feature extraction, we alternate to a lightweight single-branch network with reparameterization [10] to achieve better inference efficiency while preserving the model performance.", + "bbox": [ + 496, + 654, + 893, + 760 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In particular, a multi-branch CNN network with residual connections is applied during training for maximum representational power. At inference time, we losslessly convert the feature backbone into an efficient single-branch network by adopting the reparameterization technique [10], which is achieved by fusing parallel convolution kernels into a single one. Then, the intermediate $\\frac{1}{8}$ down-sampled coarse features $\\tilde{\\mathbf{F}}_A, \\tilde{\\mathbf{F}}_B$ and fine features in $\\frac{1}{4}$ and $\\frac{1}{2}$ resolutions are extracted efficiently for later coarse-to-fine matching.", + "bbox": [ + 496, + 763, + 895, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "21668", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Efficient Local Feature Transformation", + "text_level": 1, + "bbox": [ + 76, + 90, + 415, + 107 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "After the feature extraction, the coarse-level feature maps $\\tilde{\\mathbf{F}}_A$ and $\\tilde{\\mathbf{F}}_B$ are transformed by interleaving self- and crossattention $^1 n$ times to improve discriminativeness. The transformed features are denoted as $\\tilde{\\mathbf{F}}_A^t,\\tilde{\\mathbf{F}}_B^t$", + "bbox": [ + 75, + 113, + 470, + 176 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Previous methods often perform attention on the entire coarse-level feature maps, where linear attention instead of vanilla attention is applied to ensure a manageable computation cost. However, the efficiency is still limited due to the large token size of coarse features. Moreover, the usage of linear attention leads to sub-optimal model capability. Unlike them, we propose efficient aggregated attention for both efficiency and performance.", + "bbox": [ + 75, + 178, + 472, + 297 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Preliminaries. First, we provide a brief overview of the commonly used vanilla attention and linear attention. Vanilla attention is a core mechanism in transformer encoder layer, relying on three inputs: query Q, key K, and value V. The resultant output is a weighted sum of the value, where the weighted matrix is determined by the query and its corresponding key. Formally, the attention function is defined as follows:", + "bbox": [ + 75, + 315, + 470, + 435 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {V a n i l l a A t t e n t i o n} (Q, K, V) = \\operatorname {s o f t m a x} (Q K ^ {T}) V. \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 89, + 443, + 470, + 460 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "However, applying the vanilla attention directly to dense local features is impractical due to the significant token size. To address this issue, previous methods use linear attention to reduce the computational complexity from quadratic to linear:", + "bbox": [ + 75, + 468, + 470, + 542 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {L i n e a r A t t e n t i o n} (Q, K, V) = \\phi (Q) (\\phi (K) ^ {T} \\phi (V)). \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 86, + 550, + 468, + 569 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\phi (\\cdot) = elu(\\cdot) + 1$ . However, it comes at the cost of reduced representational power, which is also observed by [4].", + "bbox": [ + 75, + 575, + 470, + 622 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Aggregated Attention Module. After comprehensively investigating the mechanism of the transformer on coarse feature maps, we have two observations that motivate us to devise a new efficient aggregated attention. First, the attention regions of neighboring query tokens are similar, thus we can aggregate the neighboring tokens of $f_{i}$ to prevent the redundant computation. Second, most of the attention weights of each query token are concentrated on a small number of key tokens, hence we can select the salient tokens of $f_{j}$ before attention to reduce the computation.", + "bbox": [ + 75, + 638, + 468, + 787 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Therefore, we propose to first aggregate the $f_{i}$ token utilizing a depth-wise convolution network, and $f_{j}$ is aggregated by a max pooling layer to get reduced salient tokens:", + "bbox": [ + 75, + 789, + 470, + 835 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nf _ {i} ^ {\\prime} = \\operatorname {C o n v 2 D} \\left(f _ {i}\\right), f _ {j} ^ {\\prime} = \\operatorname {M a x P o o l} \\left(f _ {j}\\right) ， \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 132, + 849, + 468, + 867 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/e1b0da0f135ead94bdcbe978d070b3778d2f98147bf05222106c9a358c772783.jpg", + "image_caption": [ + "Figure 3. Detailed Transformer Module Comparison. Unlike LoFTR which uses all tokens of feature maps to compute attention and resort to linear attention to reduce the computational cost, the proposed attention module first aggregates features for salient tokens, which is significantly more efficient for attention. Then the vanilla attention is utilized to transform aggregated features, where relative positional encoding is inserted to capture the spatial information. Transformed features are upsampled and fused with the original features to form the final features." + ], + "image_footnote": [], + "bbox": [ + 501, + 61, + 671, + 358 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/473053f9f31671e5aa9fe3086f2e55c7e197fb322786d9cf7097f4ebb0663bdd.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 683, + 60, + 887, + 359 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where Conv2D is implemented by a strided depthwise convolution with a kernel size of $s \\times s$ , identical to that of the max-pooling layer. Then positional encoding and vanilla attention are followed to process reduced tokens. Positional encoding (PE) can help to model the spatial location contexts, where RoPE [45] is adopted in practice to account for more robust relative positions, inspired by [26]. Note that the PE layer is enabled exclusively for self-attention and skipped during cross-attention. The transformed feature map is then upsampled and fused with $f_{i}$ for the final feature map. Due to the aggregation and selection, the number of tokens in $f_{i}'$ and $f_{j}'$ is reduced by $s^{2}$ , which contributes to the efficiency of the attention phase.", + "bbox": [ + 496, + 497, + 892, + 693 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Coarse-level Matching Module", + "text_level": 1, + "bbox": [ + 500, + 702, + 772, + 718 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We establish coarse-level matches based on the previously transformed coarse feature maps $\\tilde{\\mathbf{F}}_A^t$ , $\\tilde{\\mathbf{F}}_B^t$ . Coarse correspondences indicate rough match regions for later subpixel-level matching in the refinement phase. To achieve this, $\\tilde{\\mathbf{F}}_A^t$ and $\\tilde{\\mathbf{F}}_B^t$ are densely correlated to build a score matrix $S$ . The softmax operator on both $S$ dimensions (referred to as dual-softmax) is then applied to obtain the probability of mutual nearest matching, which is commonly used in [35, 46, 53]. The coarse correspondences $\\{\\mathcal{M}_c\\}$ are established by selecting matches above the score threshold $\\tau$ while satisfying the mutual-nearest-neighbor (MNN) constraint.", + "bbox": [ + 496, + 724, + 893, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_footnote", + "text": "1We feed feature of one image as query and feature of the other image as key and value into cross-attention, similar to SG [39] and LoFTR [46].", + "bbox": [ + 75, + 875, + 468, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "21669", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Efficient Inference Strategy. We observe that the dual-softmax operator in the coarse matching can significantly restrict the efficiency in inference due to the large token size, especially for high-resolution images. Moreover, we find that the dual-softmax operator is crucial for training, dropping it at inference time while directly using the score matrix $S$ for MNN matching can also work well with better efficiency.", + "bbox": [ + 75, + 90, + 472, + 210 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The reason for using the dual-softmax operator in training is that it can help to train discriminative features. Intuitively, with the softmax operation, the matching score between two pixels can also conditioned on other pixels. This mechanism forces the network to improve feature similarity of true correspondences while suppressing similarity with irrelevant points. With trained discriminative features, the softmax operation can be potentially eliminated during inference.", + "bbox": [ + 75, + 212, + 472, + 333 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We denote the model skipping dual-softmax layer in inference as efficiency optimized model. Results in Tab. 1 demonstrate the effectiveness of this design.", + "bbox": [ + 75, + 333, + 472, + 378 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Subpixel-Level Refinement Module", + "text_level": 1, + "bbox": [ + 76, + 388, + 383, + 404 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As overviewed in Fig. 2 (4), with established coarse matches $\\{\\mathcal{M}_c\\}$ , we refine them for sub-pixel accuracy with our refinement module. It is composed of an efficient feature patch extractor for discriminative fine features, followed by a two-stage feature correlation layer for final matches $\\{\\mathcal{M}_f\\}$ .", + "bbox": [ + 76, + 412, + 472, + 491 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Efficient Fine Feature Extraction. We first extract discriminative fine feature patches centered at each coarse match $\\mathcal{M}_c$ by an efficient fusion network for later match refinement. For efficiency, our key idea here is to re-ileverage the previously transformed coarse features $\\tilde{\\mathbf{F}}_A^t$ , $\\tilde{\\mathbf{F}}_B^t$ to obtain cross-view attended discriminative fine features, instead of introducing additional feature transform networks as in LoFTR [46].", + "bbox": [ + 75, + 508, + 472, + 632 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To be specific, $\\tilde{\\mathbf{F}}_A^t$ and $\\tilde{\\mathbf{F}}_B^t$ are adaptively fused with $1/4$ and $1/2$ resolution backbone features by convolution and upsampling to obtain fine feature maps $\\hat{\\mathbf{F}}_A^t$ , $\\hat{\\mathbf{F}}_B^t$ in the original image resolution. Then $8 \\times 8$ local feature patches are cropped on fine feature maps centered at each coarse match. Since only shallow feed-forward networks are included, our fine feature fusion network is remarkably efficient.", + "bbox": [ + 75, + 632, + 472, + 744 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Two-Stage Correlation for Refinement. Based on the extracted fine local feature patches of coarse matches, we search for high-accurate sub-pixel matches. To refine a coarse match, a commonly used strategy [6, 15, 46] is to select the center-patch feature of $\\mathbf{I}_A$ as a fixed reference point, and perform feature correlation and expectation on the entire corresponding feature patch for its fine match. However, this refinement-by-expectations will introduce location variance to the final match, because irrelevant regions also", + "bbox": [ + 75, + 763, + 472, + 902 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "have weights and can affect results. Therefore, we propose a novel two-stage correlation module to solve this problem.", + "bbox": [ + 496, + 90, + 890, + 121 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our idea is to utilize a mutual-nearest-neighbor (MNN) matching to get intermediate pixel-level refined matches in the first stage, and then refine them for subpixel accuracy by correlation and expectation. Motivations are that MNN matching don't have spatial variance since matches are selected by directly indexing pixels with maximum scores, but cannot achieve sub-pixel accuracy. Conversely, refinement-by-expectation can achieve sub-pixel accuracy but variance exists. The proposed two-stage refinement can draw benefits by combining the best of both worlds.", + "bbox": [ + 496, + 122, + 893, + 272 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Specifically, to refine a coarse-level correspondence $\\mathcal{M}_c$ , the first-stage refinement phase densely correlates their fine feature patches to obtain the local patch score matrix $S_l$ . MNN searching is then applied on $S_l$ to get intermediate pixel-level fine matches. To limit the overall match number, we select the top-1 fine match for one coarse match by sorting the correlation scores.", + "bbox": [ + 496, + 273, + 893, + 378 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Then, we further refine these pixel-level matches for sub-pixel accuracy by our second-stage refinement. Since the matching accuracy has already significantly improved in first-stage refinement, now we can use a tiny local window for correlation and expectation with a maximum suppression of location variance. In practice, we correlate the feature of each point in $\\mathbf{I}_A$ with a $3\\times 3$ feature patch centered at its fine match in $\\mathbf{I}_B$ . The softmax operator is then applied to get a match distribution matrix and the final refined match is obtained by calculating expectations.", + "bbox": [ + 496, + 378, + 893, + 532 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.5. Supervision", + "text_level": 1, + "bbox": [ + 500, + 542, + 627, + 558 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The entire pipeline is trained end-to-end by supervising the coarse and refinement matching modules separately.", + "bbox": [ + 496, + 566, + 890, + 598 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Coarse-Level Matching Supervision. The coarse ground truth matches $\\{\\mathcal{M}_c\\}_{gt}$ with a total number of $N$ are built by warping grid-level points from $\\mathbf{I}_A$ to $\\mathbf{I}_B$ via depth maps and image poses following previous methods [39, 46]. The produced correlation score matrix $\\mathcal{S}$ in coarse matching is supervised by minimizing the log-likelihood loss over locations of $\\{\\mathcal{M}_c\\}_{gt}$ :", + "bbox": [ + 496, + 618, + 893, + 727 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {c} = - \\frac {1}{N} \\sum_ {(\\tilde {i}, \\tilde {j}) \\in \\left\\{\\mathcal {M} _ {c} \\right\\} _ {g t}} \\log \\mathcal {S} (\\tilde {i}, \\tilde {j}). \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 550, + 737, + 893, + 760 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Fine-Level Matching Supervision. We train the proposed two-stage fine-level matching module by separately supervising the two phases. The first stage fine loss $\\mathcal{L}_{f1}$ is to minimize the log-likelihood loss of each fine local score matrix $S_{l}$ based on the pixel-level ground truth fine matches, similar to coarse loss. The second stage is trained by $\\mathcal{L}_{f2}$ that calculates the $\\ell_2$ loss between the final subpixel matches $\\{\\mathcal{M}_f\\}$ and ground truth fine matches $\\{\\mathcal{M}_f\\}_{gt}$ .", + "bbox": [ + 496, + 779, + 893, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "21670", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/4427a473cf42cd32049546867b637e7fb785ca19d44f46d427031e05f578cef8.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
CategoryMethodMegaDepth DatasetScanNet DatasetTime (ms)
AUC@5°AUC@10°AUC@20°AUC@5°AUC@10°AUC@20°
SparseSP + NN31.746.860.17.518.632.110.8
SP + SG49.767.180.616.232.849.748.3
SP + LG49.967.080.114.830.847.531.9/30.7
Semi-DenseDRC-Net27.042.958.37.717.930.5328.0
LoFTR52.869.281.216.933.650.666.2
QuadTree54.670.582.219.037.353.5100.7
MatchFormer53.369.781.815.832.048.0128.9
TopicFM54.170.181.617.335.550.966.4
AspanFormer55.371.583.119.637.754.481.6
Ours56.472.283.519.237.053.640.1/34.4
Ours (Optimized)55.471.482.917.434.451.235.6/27.0
DenseDKM60.474.985.126.6447.0764.17210.8
ROMA62.676.786.328.950.468.3302.7
", + "bbox": [ + 122, + 70, + 848, + 256 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1. Results of Relative Pose Estimation on MegaDepth Dataset and ScanNet Dataset. We use the models trained on the MegaDepth dataset to evaluate all methods on both datasets, which can show the intra- and inter-dataset generalization abilities. The AUC of pose error at different thresholds, along with the processing time for matching image pair at a resolution of $640 \\times 480$ , is presented. For SP + LG, Ours, and Ours (Optimized), the running times of the model using FP32/Mixed-Precision numerical precisions (without Efficient Attention [8, 22]) are shown.", + "bbox": [ + 75, + 271, + 890, + 340 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The total loss is the weighted sum of all supervisions: $\\mathcal{L} = \\mathcal{L}_c + \\alpha \\mathcal{L}_{f1} + \\beta \\mathcal{L}_{f2}$", + "bbox": [ + 76, + 356, + 470, + 387 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 76, + 398, + 209, + 416 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this section, we evaluate the performance of our method on several downstream tasks, including homography estimation, pairwise pose estimation and visual localization. Furthermore, we evaluate the effectiveness of our design by conducting detailed ablation studies.", + "bbox": [ + 75, + 424, + 470, + 500 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Implementation Details", + "text_level": 1, + "bbox": [ + 76, + 508, + 294, + 523 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We adopt RepVGG [10] as our feature backbone, and self- and cross-attention are interleaved for $N = 4$ times to transform coarse features. For each attention, we aggregate features by a depth-wise convolution layer and a max-pooling layer, both with a kernel size of $4 \\times 4$ . Our model is trained on the MegaDepth dataset [24], which is a large-scale outdoor dataset. The test scenes are separated from training data following [46]. The loss function's weights $\\alpha$ and $\\beta$ are set to 1.0 and 0.25, respectively. We use the AdamW optimizer with an initial learning rate of $4 \\times 10^{-3}$ . The network training takes about 15 hours with a batch size of 20 on 4 NVIDIA V100 GPUs. And the coarse and fine stages are trained together from scratch. The trained model on MegaDepth is used to evaluate all datasets and tasks in our experiments to demonstrate the generalization ability.", + "bbox": [ + 73, + 531, + 470, + 758 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Relative Pose Estimation", + "text_level": 1, + "bbox": [ + 76, + 766, + 303, + 781 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Datasets. We use the outdoor MegaDepth [24] dataset and indoor ScanNet [7] dataset for the evaluation of relative pose estimation to demonstrate the efficacy of our method.", + "bbox": [ + 75, + 794, + 468, + 839 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "MegaDepth dataset is a large-scale dataset containing sparse 3D reconstructions from 196 scenes. The key challenges on this dataset are large viewpoints and illumination changes, as well as repetitive patterns. We follow the test", + "bbox": [ + 75, + 839, + 470, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "split of the previous method [46] that uses 1500 sampled pairs from scenes \"Sacre Coeur\" and \"St. Peter's Square\" for evaluation. Images are resized so that the longest edge equals 1200 for all semi-dense and dense methods. Following [26], sparse methods are provided resized images with longest edge equals 1600.", + "bbox": [ + 496, + 354, + 893, + 445 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "ScanNet dataset contains 1613 sequences with ground-truth depth maps and camera poses. They depict indoor scenes with viewpoint changes and texture-less regions. We use the sampled test pairs from [39] for the evaluation, where images are resized to $640 \\times 480$ for all methods.", + "bbox": [ + 496, + 448, + 893, + 523 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Baselines. We compare the proposed method with three categories of methods: 1) sparse keypoint detection and matching methods, including SuperPoint [9] with Nearest-Neighbor (NN), SuperGlue (SG) [39], LightGlue (LG) [26] matchers. 2) semi-dense matchers, including DRC-Net [23], LoFTR [46], QuadTree Attention [48], MatchFormer [55], AspanFormer [6], TopicFM [15], and 3) state-of-the-art dense matcher ROMA [14] that predict matches for each pixel.", + "bbox": [ + 496, + 530, + 893, + 667 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation protocol. Following previous methods, the recovered relative poses by matches are evaluated for reflecting matching accuracy. The pose error is defined as the maximum of angular errors in rotation and translation. We report the AUC of the pose error at thresholds $(5^{\\circ}, 10^{\\circ}$ , and $20^{\\circ})$ . Moreover, the running time of matching each image pair in the ScanNet dataset on a single NVIDIA RTX 3090 is reported for comprehensively understanding the matching accuracy and efficiency balance. We also include the running time of feature extraction and keypoint detection (SuperPoint) for the sparse method to compare with dense methods, which are end-to-end (from image input to matches output). LG [26]'s SuperPoint [9] parameters (max keypoints=2048, keypoint threshold=0.0005, nms=4) are uniformly used for all sparse methods.", + "bbox": [ + 496, + 672, + 895, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "21671", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/fb808f9295b033503722733f24114686b7896512c0118aa6192b9ce804a51320.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
CategoryMethodHomography est. AUC
@3px@5px@10px
SparseD2Net + NN23.235.953.6
R2D2 + NN50.663.976.8
DISK + NN52.364.978.9
SP + SG53.968.381.7
Semi-DenseSparse-NCNet48.954.267.1
DRC-Net50.656.268.3
LoFTR65.975.684.6
Ours66.576.485.5
", + "bbox": [ + 83, + 70, + 467, + 191 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Results. As shown in Tab. 1, the proposed method achieves competitive performances compared with sparse and semidense methods on both datasets. Qualitative comparisons are shown in Fig. 4. Specifically, our method outperforms the best semi-dense baseline AspanFormer on all metrics of the MegaDepth dataset and has lower but comparable performance on the ScanNet dataset, with $\\sim 2$ times faster. Our optimized model that eliminates the dual-softmax operator in coarse-level matching further brings efficiency improvements, with slight performance decreases. Using this strategy, our method can outperform the efficient and robust sparse method $\\mathrm{SP} + \\mathrm{LG}$ in efficiency with significantly higher accuracy. Dense matcher ROMA shows remarkable matching capability but is slow for applications in practice. Moreover, since ROMA utilizes the pre-trained DINOv2 [33] backbone, its strong generalizability on ScanNet may be attributed to the similar indoor training data in DINOv2, where other methods are trained on outdoor MegaDepth only. Compared with it, our method is $\\sim 7.5$ times faster, which has a good balance between accuracy and efficiency.", + "bbox": [ + 75, + 271, + 472, + 574 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Homography Estimation", + "text_level": 1, + "bbox": [ + 76, + 585, + 303, + 603 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Dataset. We evaluate our method on HPatches dataset [2]. HPatches dataset depicts planar scenes divided into sequences. Images are taken under different viewpoints or illumination changes.", + "bbox": [ + 75, + 616, + 470, + 676 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Baselines. We compare our method with sparse methods including D2Net [11], R2D2 [34], DISK [53] detectors with NN matcher, and SuperPoint [9] + SuperGlue [39]. As for semi-dense methods, we compare with Sparse-NCNet [36], DRC-Net [23], and LoFTR [46]. For SuperGlue and all semidense methods, we use their models trained on MegaDepth dataset for evaluation.", + "bbox": [ + 75, + 681, + 468, + 787 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Evaluation Protocol. Following SuperGlue [39] and LoFTR [46], we resize all images for matching so that their smallest edge equals 480 pixels. We collect the mean reprojection error of corner points, and report the area under the cumulative curve (AUC) under 3 different thresholds, including 3 px, 5 px, and 10 px. For all baselines, we employ the same RANSAC method as a robust homography estimator", + "bbox": [ + 75, + 794, + 470, + 901 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/43bb865f1e053b2ef07e8e271d899228814e6937e479000998b87bb724f06dad.jpg", + "table_caption": [ + "Table 2. Results of Homography Estimation on HPatches Dataset. Our method is compared with sparse and semi-dense methods. The AUC of reprojection error of corner points at different thresholds is reported." + ], + "table_footnote": [], + "table_body": "
MethodDUC1DUC2
(0.25m,2°)/(0.5m,5°)/(1.0m,10°)
SP+SG49.0 / 68.7 / 80.853.4 / 77.1 / 82.4
LoFTR47.5 / 72.2 / 84.854.2 / 74.8 / 85.5
TopicFM52.0 / 74.7 / 87.453.4 / 74.8 / 83.2
PATS55.6 / 71.2 / 81.058.8 / 80.9 / 85.5
AspanFormer51.5 / 73.7 / 86.055.0 / 74.0 / 81.7
Ours52.0 / 74.7 / 86.958.0 / 80.9 / 89.3
", + "bbox": [ + 506, + 70, + 888, + 169 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/c41257502ba55398c518460753f6776e1807e41356fdfbb17f679c04ef75d50b.jpg", + "table_caption": [ + "Table 3. Results of Visual Localization on InLoc Dataset." + ], + "table_footnote": [], + "table_body": "
MethodDayNight
(0.25m,2°)/(0.5m,5°)/(1.0m,10°)
SP+SG89.8 / 96.1 / 99.477.0 / 90.6 / 100.0
LoFTR88.7 / 95.6 / 99.078.5 / 90.6 / 99.0
TopicFM90.2 / 95.9 / 98.977.5 / 91.1 / 99.5
PATS89.6 / 95.8 / 99.373.8 / 92.1 / 99.5
AspanFormer89.4 / 95.6 / 99.077.5 / 91.6 / 99.5
Ours89.6 / 96.2 / 99.077.0 / 91.1 / 99.5
", + "bbox": [ + 506, + 199, + 888, + 297 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 4. Results of Visual Localization on Aachen v1.1 Dataset.", + "bbox": [ + 498, + 303, + 890, + 316 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "for a fair comparison. Following LoFTR, we select only the top 1000 predicted matches of semi-dense methods for the sake of fairness.", + "bbox": [ + 498, + 333, + 890, + 378 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Results. As shown in Tab. 2, even though the number of matches is restricted, our method can also work remarkably well and outperform sparse methods significantly. Compared with semi-dense, our method can surpass them with significantly higher efficiency. We attribute this to the effectiveness of two-stage refinement for accuracy improvement and proposed aggregation module for efficiency.", + "bbox": [ + 498, + 383, + 893, + 491 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Visual Localization", + "text_level": 1, + "bbox": [ + 500, + 502, + 683, + 517 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Datasets and Evaluation Protocols. Visual localization is an important downstream task of image matching, which aims to estimate the 6-DoF poses of query images based on the 3D scene model. We conduct experiments on two commonly used benchmarks, including InLoc [47] dataset and Aachen v1.1 [41] dataset, for evaluation to demonstrate the superiority of our method. InLoc dataset is captured on indoor scenes with plenty of repetitive structures and textureless regions, where each database image has a corresponding depth map. Aachen v1.1 is a challenging large-scale outdoor dataset for localization with large-viewpoint and day-and-night illumination changes, which particularly relies on the robustness of matching methods. We adopt its full localization track for benchmarking.", + "bbox": [ + 496, + 531, + 893, + 743 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Following [6, 46], the open-sourced localization framework HLoc [38] is utilized. For both datasets, the percentage of pose errors satisfying both angular and distance thresholds is reported following the benchmarks, where different thresholds are used. For the InLoc dataset, the metrics of two test scenes including DUC1 and DUC2 are separately reported. As for the Aachen v1.1 dataset, the metrics corresponding to the daytime and nighttime divisions are reported.", + "bbox": [ + 496, + 743, + 893, + 864 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Baselines. We compare the proposed method with both detector-based method SuperPoint [9] + SuperGlue [39] and", + "bbox": [ + 500, + 869, + 890, + 901 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "21672", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c381cf5d3a60fdbc0fdc9bca1154fc524231d28359a4512b136d2d04f96390e1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 143, + 47, + 377, + 247 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/e592b78f9f7fe6f413684e84894725f215261dc9270936e8fed1412663c13b56.jpg", + "image_caption": [ + "Figure 4. Qualitative Results. Our method is compared with the sparse matching pipeline SuperPoint [9] $^+$ LightGlue [26], semi-dense \nmacher AspanFormer [6]. Image pairs with texture-poor regions and large-viewpoint changes can be robustly matched by our method. The \nred color indicates epipolar error beyond $5\\times 10^{-4}$ (in the normalized image coordinates)." + ], + "image_footnote": [], + "bbox": [ + 388, + 47, + 604, + 250 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/9b21e807ba425ad2643d05931505c51b460341d8f5a1f3829da0f4372e2bc8ec.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 617, + 50, + 830, + 250 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/838e5501e4a0826fb5973c1508383a04bd832c1c13e551d9d45685f8c1f689b9.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodPose Estimation AUCTime (ms)
@5°@10°@20°
Ours Full56.472.283.5139.2
1) Ours Optimal (w/o dual-softmax)55.471.482.9102.0
2) Replace Agg. Attention to LoFTR's Trans.54.770.582.2171.4
3) Replace two-stage refine. to LoFTR's refine.54.770.982.7135.3
4) No second-stage refinement55.871.883.3138.1
5) Replace RepVGG with ResNet55.471.482.9156.2
", + "bbox": [ + 81, + 300, + 467, + 402 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 5. Ablation Studies. The components of our method are ablated on the MegaDepth dataset for a comprehensive understanding of our method, where averaged running times for an image pair with high-resolution $1200 \\times 1200$ are reported.", + "bbox": [ + 75, + 407, + 470, + 465 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "detector-free methods including LoFTR [46], TopicFM [15], PATS [32] and Aspanformer [6].", + "bbox": [ + 75, + 477, + 470, + 508 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Results. We adhere to the pipeline and evaluation settings of the online visual localization benchmark² to ensure fairness. As presented in Tab. 3, the proposed method achieves competitive results, taking both detector-based and detector-free methods into account. Being a method primarily geared towards efficiency, our approach can deliver results comparable to those of many accuracy-oriented methods. As depicted in Tab. 4, our method also demonstrates performance on par with the best-performing approaches.", + "bbox": [ + 75, + 512, + 472, + 650 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.5. Ablation Studies", + "text_level": 1, + "bbox": [ + 76, + 657, + 243, + 674 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this part, we conduct detailed ablation studies to analyze the effectiveness of our proposed modules with results shown in Tab. 5. 1) Without dual-softmax, our optimal model can bring huge efficiency improvement in high-resolution images. 2) In coarse feature transformation, replacing the proposed aggregated attention module with LoFTR's transformer can bring significant efficiency dropping, as well as accuracy decrease. Note that the replaced transformer is also equipped with RoPE same as ours for fair comparison. This demonstrates the efficacy of the proposed module that performing vanilla attention on aggregated features can achieve higher efficiency with even better matching accuracy. 3) Compared with using LoFTR's refinement that performs", + "bbox": [ + 75, + 681, + 472, + 878 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "expectation on the entire correlation patch, the proposed two-stage refinement layer can bring accuracy improvement with negligible latency. We attribute this to the two-stage refinement's property that can maximize the suppression of location variance in correlation refinement. 4) Dropping the second refinement stage will lead to degraded pose accuracy with minor efficiency changes, especially on the strict AUC@5° metric. 5) Changing the backbone from reparameterized VGG [10] back to multi-branch ResNet [17] leads to decreased efficiency with similar accuracy, which demonstrates the effectiveness of our design choice for efficiency.", + "bbox": [ + 496, + 303, + 895, + 470 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusions", + "text_level": 1, + "bbox": [ + 500, + 482, + 627, + 498 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This paper introduces a new semi-dense local feature matcher based on the success of LoFTR. We revisit its designs and propose several improvements for both efficiency and matching accuracy. A key observation is that performing the Transformer on the entire coarse feature map is redundant due to the similar local information, where an aggregated attention module is proposed to perform transformer on reduced tokens with significantly better efficiency and competitive performance. Moreover, a two-stage correlation layer is devised to solve the location variance problem in LoFTR's refinement design, which further brings accuracy improvements. As a result, our method can achieve $\\sim 2.5$ times faster compared with LoFTR with better matching accuracy. Moreover, as a semi-dense matching method, the proposed method can achieve comparable efficiency with the recent robust sparse feature matcher LightGlue [26]. We believe this opens up the applications of our method in large-scale or latency-sensitive downstream tasks, such as image retrieval and 3D reconstruction. Please refer to the supplementary material for discussions about limitations and future works.", + "bbox": [ + 496, + 507, + 895, + 811 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgement The authors would like to acknowledge support from NSFC (No. 62172364) and Information Technology Center and State Key Lab of CAD&CG, Zhejiang University.", + "bbox": [ + 496, + 828, + 895, + 888 + ], + "page_idx": 7 + }, + { + "type": "page_footnote", + "text": "$^{2}$ https://wwwVisualization.net/benchmark", + "bbox": [ + 94, + 886, + 346, + 898 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "21673", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Sameer Agarwal, Yasutaka Furukawa, Noah Snavely, Ian Simon, Brian Curless, Steven M. Seitz, and Richard Szeliski. Building rome in a day. ICCV, 2009. 1", + "[2] Vassileios Balntas, Karel Lenc, Andrea Vedaldi, and Krystian Mikolajczyk. Hpatches: A benchmark and evaluation of handcrafted and learned local descriptors. In CVPR, pages 5173-5182, 2017. 7", + "[3] Herbert Bay, Andreas Ess, Tinne Tuytelaars, and Luc Van Gool. Speeded-up robust features (surf). CVIU, 110 (3):346-359, 2008. 2", + "[4] Han Cai, Chuang Gan, and Song Han. Efficientvit: Enhanced linear attention for high-resolution low-computation visual recognition. arXiv preprint arXiv:2205.14756, 2022. 4", + "[5] Hongkai Chen, Zixin Luo, Jiahui Zhang, Lei Zhou, Xuyang Bai, Zeyu Hu, Chiew-Lan Tai, and Long Quan. Learning to match features with seeded graph matching network. In ICCV, pages 6301-6310, 2021. 2", + "[6] Hongkai Chen, Zixin Luo, Lei Zhou, Yurun Tian, Mingmin Zhen, Tian Fang, David McKinnon, Yanghai Tsin, and Long Quan. Aspanformer: Detector-free image matching with adaptive span transformer. In ECCV, pages 20-36. Springer, 2022. 2, 5, 6, 7, 8", + "[7] Angela Dai, Angel X Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In CVPR, pages 5828-5839, 2017. 6", + "[8] Tri Dao, Daniel Y. Fu, Stefano Ermon, Atri Rudra, and Christopher Ré. FlashAttention: Fast and memory-efficient exact attention with IO-awareness. In NeurIPS, 2022. 3, 6", + "[9] Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. Superpoint: Self-supervised interest point detection and description. CVPRW, 2018. 2, 6, 7, 8", + "[10] Xiaohan Ding, Xiangyu Zhang, Ningning Ma, Jungong Han, Guiguang Ding, and Jian Sun. Repvgg: Making vgg-style convnets great again. In CVPR, pages 13733-13742, 2021. 3, 6, 8", + "[11] Mihai Dusmanu, Ignacio Rocco, Tomas Pajdla, Marc Pollefeys, Josef Sivic, Akihiko Torii, and Torsten Sattler. D2-net: A trainable cnn for joint description and detection of local features. In CVPR, pages 8092-8101, 2019. 2, 7", + "[12] Patrick Ebel, Anastasiia Mishchuk, Kwang Moo Yi, Pascal Fua, and Eduard Trulls. Beyond cartesian representations for local descriptors. In ICCV, pages 253-262, 2019. 2", + "[13] Johan Edstedt, Ioannis Athanasiadis, Mårten Wadenbäck, and Michael Felsberg. Dkm: Dense kernelized feature matching for geometry estimation. In CVPR, pages 17765-17775, 2023. 2, 3", + "[14] Johan Edstedt, Qiyu Sun, Georg Bokman, Mårten Wadenback, and Michael Felsberg. Roma: Revisiting robust losses for dense feature matching. arXiv preprint arXiv:2305.15404, 2023. 2, 3, 6", + "[15] Khang Truong Giang, Soohwan Song, and Sung-Guk Jo. Top-icfm: Robust and interpretable topic-assisted feature matching. In AAAI, 2023. 2, 3, 5, 6, 8", + "[16] Stephen Hausler, Sourav Garg, Ming Xu, Michael Milford, and Tobias Fischer. Patch-netvlad: Multi-scale fusion of" + ], + "bbox": [ + 78, + 114, + 470, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "locally-global descriptors for place recognition. CVPR, pages 14136-14147, 2021. 2", + "[17] Kaiming He, X. Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. CVPR, pages 770-778, 2015. 3, 8", + "[18] Xingyi He, Jiaming Sun, Yifan Wang, Sida Peng, Qixing Huang, Hujun Bao, and Xiaowei Zhou. Detector-free structure from motion. In CVPR, 2024. 1", + "[19] Angelos Katharopoulos, Apoorv Vyas, Nikolaos Pappas, and Francois Fleuret. Transformers are rnns: Fast autoregressive transformers with linear attention. In ICML, 2020. 3", + "[20] Nikita Kitaev, Lukasz Kaiser, and Anselm Levskaya. Reformer: The efficient transformer. ArXiv, abs/2001.04451, 2020.3", + "[21] Axel Barroso Laguna, Edgar Riba, Daniel Ponsa, and Krystian Mikolajczyk. Key.net: Keypoint detection by handcrafted and learned cnn filters. ICCV, pages 5835-5843, 2019. 2", + "[22] Benjamin Lefaudeau, Francisco Massa, Diana Liskovich, Wenhan Xiong, Vittorio Caggiano, Sean Naren, Min Xu, Jieru Hu, Marta Tintore, Susan Zhang, Patrick Labatut, and Daniel Haziza. xformers: A modular and hackable transformer modelling library. https://github.com/facebookresearch/xformers, 2022.3.6", + "[23] Xinghui Li, Kai Han, Shuda Li, and Victor Prisacariu. Dualresolution correspondence networks. In NeurlPS, 2020. 2, 6, 7", + "[24] Zhengqi Li and Noah Snavely. Megadepth: Learning single-view depth prediction from internet photos. In CVPR, pages 2041-2050, 2018. 6", + "[25] Philipp Lindenberger, Paul-Edouard Sarlin, Viktor Larsson, and Marc Pollefeys. Pixel-perfect structure-from-motion with featuremetric refinement. ICCV, pages 5967-5977, 2021. 1", + "[26] Philipp Lindenberger, Paul-Edouard Sarlin, and Marc Pollefeys. LightGlue: Local Feature Matching at Light Speed. In ICCV, 2023. 1, 2, 4, 6, 8", + "[27] G LoweDavid. Distinctive image features from scale-invariant keypoints. IJCV, 2004. 1, 2", + "[28] Zixin Luo, Lei Zhou, Xuyang Bai, Hongkai Chen, Jiahui Zhang, Yao Yao, Shiwei Li, Tian Fang, and Long Quan. Aslfeat: Learning local features of accurate shape and localization. In CVPR, pages 6589-6598, 2020. 2", + "[29] Anastasiia Mishchuk, Dmytro Mishkin, Filip Radenovic, and Jiri Matas. Working hard to know your neighbor's margins: Local descriptor learning loss. NeurIPS, 30, 2017. 2", + "[30] Raul Mur-Artal and Juan D. Tardós. Orb-slam2: An open-source slam system for monocular, stereo, and rgb-d cameras. TR, 33:1255-1262, 2016. 1", + "[31] Raul Mur-Artal, José M. M. Montiel, and Juan D. Tardós. Orb-slam: A versatile and accurate monocular slam system. TR, 31:1147-1163, 2015. 1", + "[32] Junjie Ni, Yijin Li, Zhaoyang Huang, Hongsheng Li, Hujun Bao, Zhaopeng Cui, and Guofeng Zhang. Pats: Patch area transportation with subdivision for local feature matching. CVPR, pages 17776-17786, 2023. 2, 8", + "[33] Maxime Oquab, Timothée Darcet, Theo Moutakanni, Huy V. Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, Russell" + ], + "bbox": [ + 503, + 92, + 893, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "21674", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Howes, Po-Yao Huang, Hu Xu, Vasu Sharma, Shang-Wen Li, Wojciech Galuba, Mike Rabbat, Mido Assran, Nicolas Ballas, Gabriel Synnaeve, Ishan Misra, Herve Jegou, Julien Mairal, Patrick Labatut, Armand Joulin, and Piotr Bojanowski. Dinov2: Learning robust visual features without supervision, 2023. 7", + "[34] Jérôme Revaud, César Roberto de Souza, M. Humenberger, and Philippe Weinzaepfel. R2d2: Reliable and repeatable detector and descriptor. In NeurIPS, 2019. 2, 7", + "[35] Ignacio Rocco, Mircea Cimpoi, Relja Arandjelović, Akihiko Torii, Tomas Pajdla, and Josef Sivic. Neighbourhood consensus networks. *NeurIPS*, 31, 2018. 2, 4", + "[36] Ignacio Rocco, Relja Arandjelović, and Josef Sivic. Efficient neighbourhood consensus networks via submanifold sparse convolutions. In ECCV, pages 605-621. Springer, 2020. 2, 7", + "[37] Edward Rosten and Tom Drummond. Machine learning for high-speed corner detection. In ECCV, pages 430-443. Springer, 2006. 1, 2", + "[38] Paul-Edouard Sarlin, César Cadena, Roland Y. Siegwart, and Marcin Dymczyk. From coarse to fine: Robust hierarchical localization at large scale. CVPR, pages 12708–12717, 2019. 1, 7", + "[39] Paul-Edouard Sarlin, Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. SuperGlue: Learning feature matching with graph neural networks. In CVPR, 2020. 1, 2, 4, 5, 6, 7", + "[40] Paul-Edouard Sarlin, Ajaykumar Unagar, Måns Larsson, Hugo Germain, Carl Toft, Victor Larsson, Marc Pollefeys, Vincent Lepetit, Lars Hammarstrand, Fredrik Kahl, and Torsten Sattler. Back to the Feature: Learning Robust Camera Localization from Pixels to Pose. In CVPR, 2021. 1", + "[41] Torsten Sattler, Will Maddern, Carl Toft, Akihiko Torii, Lars Hammarstrand, Erik Stenberg, Daniel Safari, Masatoshi Okutomi, Marc Pollefeys, Josef Sivic, et al. Benchmarking 6dof outdoor visual localization in changing conditions. In CVPR, pages 8601-8610, 2018. 7", + "[42] Nikolay Savinov, Akihito Seki, Lubor Ladicky, Torsten Sattler, and Marc Pollefeys. Quad-networks: unsupervised learning to rank for interest point detection. In CVPR, pages 1822-1830, 2017. 2", + "[43] Johannes L. Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In CVPR, 2016. 1, 2", + "[44] Yan Shi, Jun-Xiong Cai, Yoli Shavit, Tai-Jiang Mu, Wensen Feng, and Kai Zhang. Clustergnn: Cluster-based coarse-to-fine graph neural network for efficient feature matching. In CVPR, pages 12517-12526, 2022. 2", + "[45] Jianlin Su, Yu Lu, Shengfeng Pan, Bo Wen, and Yunfeng Liu. Roformer: Enhanced transformer with rotary position embedding. ArXiv, abs/2104.09864, 2021. 4", + "[46] Jiaming Sun, Zehong Shen, Yuang Wang, Hujun Bao, and Xiaowei Zhou. Loftr: Detector-free local feature matching with transformers. In CVPR, pages 8922-8931, 2021. 1, 2, 4, 5, 6, 7, 8", + "[47] Hajime Taira, Masatoshi Okutomi, Torsten Sattler, Mircea Cimpoi, Marc Pollefeys, Josef Sivic, Tomas Pajdla, and Akihiko Torii. Inloc: Indoor visual localization with dense matching and view synthesis. In CVPR, pages 7199-7209, 2018. 7" + ], + "bbox": [ + 78, + 90, + 470, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[48] Shitao Tang, Jiahui Zhang, Siyu Zhu, and Ping Tan. Quadtree attention for vision transformers. *ICLR*, 2022. 2, 6", + "[49] Yurun Tian, Bin Fan, and Fuchao Wu. L2-net: Deep learning of discriminative patch descriptor in euclidean space. In CVPR, pages 661-669, 2017. 1, 2", + "[50] Yurun Tian, Xin Yu, Bin Fan, Fuchao Wu, Huub Heijnen, and Vassileios Balntas. Sosnet: Second order similarity regularization for local descriptor learning. In CVPR, pages 11016-11025, 2019. 2", + "[51] Yurun Tian, Vassileios Balntas, Tony Ng, Axel Barroso-Laguna, Yiannis Demiris, and Krystian Mikolajczyk. D2d: Keypoint extraction with describe to detect approach. In ACCV, 2020. 2", + "[52] Prune Truong, Martin Danelljan, Luc Van Gool, and Radu Timofte. Learning accurate dense correspondences and when to trust them. In CVPR, pages 5714-5724, 2021. 3", + "[53] Michal Tyszkiiewicz, Pascal Fua, and Eduard Trulls. Disk: Learning local features with policy gradient. NeurIPS, 2020. 4, 7", + "[54] Ashish Vaswani, Noam M. Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In NeurIPS, 2017. 3", + "[55] Qing Wang, Jiaming Zhang, Kailun Yang, Kunyu Peng, and Rainer Stiefelhagen. Matchformer: Interleaving attention in transformers for feature matching. In ACCV, 2022. 2, 6", + "[56] Sinong Wang, Belinda Z. Li, Madian Khabsa, Han Fang, and Hao Ma. Linformer: Self-attention with linear complexity. ArXiv, abs/2006.04768, 2020. 3" + ], + "bbox": [ + 501, + 92, + 890, + 502 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "21675", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2024/Efficient LoFTR_ Semi-Dense Local Feature Matching with Sparse-Like Speed/14a55345-a1a3-4ac4-a5a2-7c153c54b832_model.json b/2024/Efficient LoFTR_ Semi-Dense Local Feature Matching with Sparse-Like Speed/14a55345-a1a3-4ac4-a5a2-7c153c54b832_model.json new file mode 100644 index 0000000000000000000000000000000000000000..a3f7b73dcb82ab487adfd9c0739c20538cb50587 --- /dev/null +++ b/2024/Efficient LoFTR_ Semi-Dense Local Feature Matching with Sparse-Like Speed/14a55345-a1a3-4ac4-a5a2-7c153c54b832_model.json @@ -0,0 +1,2145 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.0, + 0.812, + 0.045 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.089, + 0.131, + 0.882, + 0.153 + ], + "angle": 0, + "content": "Efficient LoFTR: Semi-Dense Local Feature Matching with Sparse-Like Speed" + }, + { + "type": "text", + "bbox": [ + 0.202, + 0.173, + 0.782, + 0.192 + ], + "angle": 0, + "content": "Yifan Wang* Xingyi He* Sida Peng† Dongli Tan Xiaowei Zhou" + }, + { + "type": "text", + "bbox": [ + 0.391, + 0.207, + 0.554, + 0.225 + ], + "angle": 0, + "content": "Zhejiang University" + }, + { + "type": "title", + "bbox": [ + 0.244, + 0.253, + 0.323, + 0.269 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.28, + 0.474, + 0.643 + ], + "angle": 0, + "content": "We present a novel method for efficiently producing semidense matches across images. Previous detector-free matcher LoFTR has shown remarkable matching capability in handling large-viewpoint change and texture-poor scenarios but suffers from low efficiency. We revisit its design choices and derive multiple improvements for both efficiency and accuracy. One key observation is that performing the transformer over the entire feature map is redundant due to shared local information, therefore we propose an aggregated attention mechanism with adaptive token selection for efficiency. Furthermore, we find spatial variance exists in LoFTR's fine correlation module, which is adverse to matching accuracy. A novel two-stage correlation layer is proposed to achieve accurate subpixel correspondences for accuracy improvement. Our efficiency optimized model is \\(\\sim 2.5\\times\\) faster than LoFTR which can even surpass state-of-the-art efficient sparse matching pipeline SuperPoint + LightGlue. Moreover, extensive experiments show that our method can achieve higher accuracy compared with competitive semi-dense matchers, with considerable efficiency benefits. This opens up exciting prospects for large-scale or latency-sensitive applications such as image retrieval and 3D reconstruction. Project page: https://zju3dv.github.io/efficientloftr/." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.669, + 0.21, + 0.685 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.695, + 0.471, + 0.862 + ], + "angle": 0, + "content": "Image matching is the cornerstone of many 3D computer vision tasks, which aim to find a set of highly accurate correspondences given an image pair. The established matches between images have broad usages such as reconstructing the 3D world by structure from motion (SfM) [1, 18, 25, 43] or SLAM system [30, 31], and visual localization [38, 40], etc. Previous methods typically follow a two-stage pipeline: they first detect [37] and describe [49] a set of keypoints on each image, and then establish keypoint correspondences by handcrafted [27] or learning-based matchers [26, 39]. These detector-based methods are efficient but suffer from robustly" + }, + { + "type": "image", + "bbox": [ + 0.507, + 0.253, + 0.891, + 0.516 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.519, + 0.895, + 0.618 + ], + "angle": 0, + "content": "Figure 1. Matching Accuracy and Efficiency Comparisons. Our method achieves competitive accuracy compared with semi-dense matchers \\((\\odot)\\) at a significantly higher speed. Compared with dense matcher ROMA \\((\\odot)\\), our method is \\(\\sim 7.5\\times\\) faster. Moreover, our efficiency optimized model \\((\\star)\\) can surpass the robust sparse matching pipeline \\((\\odot)\\) SuperPoint (SP) + LightGlue (LG) on efficiency with considerably better accuracy." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.627, + 0.893, + 0.657 + ], + "angle": 0, + "content": "detecting repeatable keypoints across challenging pairs, such as extreme viewpoint changes and texture-poor regions." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.658, + 0.895, + 0.884 + ], + "angle": 0, + "content": "Recently, LoFTR [46] introduces a detector-free matching paradigm with transformer to directly establish semi-dense correspondences between two images without detecting keypoints. With the help of the transformer mechanism to capture the global image context and the detector-free design, LoFTR shows a strong capability of matching challenging pairs, especially in texture-poor scenarios. To reduce the computation burden, LoFTR adopts a coarse-to-fine pipeline by first performing dense matching on downsampled coarse features maps, where transformer is applied. Then, the feature locations of coarse matches on one image are fixed, while their subpixel correspondences are searched on the other image by cropping feature patches based on coarse match, performing the feature correlation, and calculating expectation over the correlation patch." + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.886, + 0.892, + 0.902 + ], + "angle": 0, + "content": "Despite its impressive matching performance, LoFTR" + }, + { + "type": "page_footnote", + "bbox": [ + 0.077, + 0.875, + 0.47, + 0.901 + ], + "angle": 0, + "content": "*Equal contribution. † Corresponding author: Sida Peng. The authors from Zhejiang University are affiliated with the State Key Lab of CAD&CG." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.946, + 0.521, + 0.957 + ], + "angle": 0, + "content": "21666" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.288 + ], + "angle": 0, + "content": "suffers from limited efficiency due to the large token size of performing transformer on the entire coarse feature map, which significantly barricades practical large-scale usages such as image retrieval [16] and SfM [43]. A large bunch of LoFTR's follow-up works [6, 15, 32, 48, 55] have attempted to improve its matching accuracy. However, there are rare methods that focus on matching efficiency of detector-free matching. QuadTree Attention [48] incorporates multi-scale transformation with a gradually narrowed attention span to avoid performing attention on large feature maps. This strategy can reduce the computation cost, but it also divides a single coarse attention process into multiple steps, leading to increased latency." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.289, + 0.473, + 0.546 + ], + "angle": 0, + "content": "In this paper, we revisit the design decisions of the detector-free matcher LoFTR, and propose a new matching algorithm that squeezes out redundant computations for significantly better efficiency while further improving the accuracy. As shown in Fig. 1, our approach achieves the best inference speed compared with recent image matching methods while being competitive in terms of accuracy. Our key innovations lie in introducing a token aggregation mechanism for efficient feature transformation and a two-stage correlation layer for correspondence refinement. Specifically, we find that densely performing global attention over the entire coarse feature map as in LoFTR is unnecessary, as the attention information is similar and shared in the local region. Therefore, we devise an aggregated attention mechanism to perform feature transformation on adaptively selected tokens, which is significantly compact and effectively reduces the cost of local feature transformation." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.548, + 0.473, + 0.684 + ], + "angle": 0, + "content": "In addition, we observe that there can be spatial variance in the matching refinement phase of LoFTR, which is caused by the expectation over the entire correlation patch when noisy feature correlation exists. To solve this issue, our approach designs a two-stage correlation layer that first locates pixel-level matches with the accurate mutual-nearest-neighbor matching on fine feature patches, and then further refines matches for subpixel-level by performing the correlation and expectation locally within tiny patches." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.686, + 0.473, + 0.837 + ], + "angle": 0, + "content": "Extensive experiments are conducted on multiple tasks, including homography estimation, relative pose recovery, as well as visual localization, to show the efficacy of our method. Our pipeline pushes detector-free matching to unprecedented efficiency, which is \\(\\sim 2.5\\) times faster than LoFTR and can even surpass the current state-of-the-art efficient sparse matcher LightGlue [26]. Moreover, our framework can achieve comparable or even better matching accuracy compared with competitive detector-free baselines [6, 13, 14] with considerably higher efficiency." + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.839, + 0.462, + 0.854 + ], + "angle": 0, + "content": "In summary, this paper has the following contributions:" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.856, + 0.473, + 0.901 + ], + "angle": 0, + "content": "- A new detector-free matching pipeline with multiple improvements based on the comprehensive revisiting of LoFTR, which is significantly more efficient and with" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.093, + 0.622, + 0.107 + ], + "angle": 0, + "content": "better accuracy." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.107, + 0.892, + 0.135 + ], + "angle": 0, + "content": "- A novel aggregated attention network for efficient local feature transformation." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.137, + 0.892, + 0.167 + ], + "angle": 0, + "content": "- A novel two-stage correlation refinement layer for accurate and subpixel-level refined correspondences." + }, + { + "type": "list", + "bbox": [ + 0.5, + 0.107, + 0.892, + 0.167 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.186, + 0.642, + 0.202 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.213, + 0.895, + 0.591 + ], + "angle": 0, + "content": "Detector-Based Image Matching. Classical image matching methods [3, 27, 37] adopt handcrafted critics for detecting keypoints, describing and then matching them. Recent methods draw benefits from deep neural networks for both detection [21, 37, 42] and description [12, 29, 49, 50], where the robustness and discriminativeness of local descriptors are significantly improved. Besides, some methods [9, 11, 28, 34, 51] managed to learn the detector and descriptor together. SuperGlue [39] is a pioneering method that first introduces the transformer mechanism into matching, which has shown notable improvements over classical handcrafted matchers. As a side effect, it also costs more time, especially with many keypoints to match. To improve the efficiency, some subsequent works, such as [5, 44], endeavor to reduce the size of the attention mechanism, albeit at the cost of sacrificing performance. LightGlue [26] introduces a new scheme for efficient sparse matching that is adaptive to the matching difficulty, where the attention process can be stopped earlier for easy pairs. It is faster than SuperGlue and can achieve competitive performance. However, robustly detecting keypoints across images is still challenging, especially for texture-poor regions. Unlike them, our method focuses on the efficiency of the detector-free method, which eliminates the restriction of keypoint detection and shows superior performance for challenging pairs." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.597, + 0.895, + 0.717 + ], + "angle": 0, + "content": "Detector-Free Image Matching. Detector-free methods directly match images instead of relying on a set of detected keypoints, producing semi-dense or dense matches. NC-Net [35] represents all features and possible matches as a 4D correlation volume. Sparse NC-Net [36] utilizes sparse correlation layers to ease resolution limitations. Subsequently, DRC-Net [23] improves efficiency and further improves performance in a coarse-to-fine manner." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.897, + 0.901 + ], + "angle": 0, + "content": "LoFTR [46] first employs the Transformer in detector-free matching to model the long-range dependencies. It shows remarkable matching capabilities, however, suffers from low efficiency due to the huge computation of densely transforming entire coarse feature maps. Many follow-up works further improve the matching accuracy. Matchformer [55] and AspanFormer [6] perform attention on multiscale features, where local attention regions of [6] are found with the help of estimated flow. QuadTree [48] gradually restricts the attention span during hierarchical attention to relevant areas, which can reduce overall computation. However, these designs contribute marginally or even decrease" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "21667" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.078, + 0.083, + 0.895, + 0.358 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.373, + 0.897, + 0.473 + ], + "angle": 0, + "content": "Figure 2. Pipeline Overview. (1) Given an image pair, a CNN network extracts coarse feature maps \\(\\tilde{\\mathbf{F}}_A\\) and \\(\\tilde{\\mathbf{F}}_B\\), as well as fine features. (2) Then, we transform coarse features for more discriminative feature maps by interleaving our aggregated self- and cross-attention \\(N\\) times, where adaptively feature aggregation is performed to reduce token size before each attention for efficiency. (3) Transformed coarse features are correlated for the score matrix \\(S\\). Mutual-nearest-neighbor (MNN) searching is followed to establish coarse matches \\(\\{\\mathcal{M}_c\\}\\). (4) To refine coarse matches, discriminative fine features \\(\\hat{\\mathbf{F}}_A^t\\), \\(\\hat{\\mathbf{F}}_B^t\\) in full resolution are obtained by fusing transformed coarse features \\(\\tilde{\\mathbf{F}}_A^t\\), \\(\\tilde{\\mathbf{F}}_B^t\\) with backbone features. Feature patches are then cropped centered at each coarse match \\(\\mathcal{M}_c\\). A two-stage refinement is followed to obtain sub-pixel correspondence \\(\\mathcal{M}_f\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.487, + 0.472, + 0.653 + ], + "angle": 0, + "content": "efficiency, since the hierarchical nature of multi-scale attention will further introduce latencies. TopicFM [15] first assigns features with similar semantic meanings to the same topic, where attention is conducted within each topic for efficiency. Since it needs to sequentially process each token's features for transformation, the efficiency improvement is limited. Moreover, performing local attention within topics can potentially restrict the capability of modeling long-range dependencies. Compared with them, the proposed aggregated attention module in our method significantly improves efficiency while achieving better accuracy." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.654, + 0.472, + 0.746 + ], + "angle": 0, + "content": "Dense matching methods [13, 14, 52] are designed to estimate all possible correspondences between two images, which show strong robustness. However, they are generally much slower compared with sparse and semi-dense methods. Unlike them, our method produces semi-dense matches with competitive performance and considerably better efficiency." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.473, + 0.902 + ], + "angle": 0, + "content": "Transformer has been broadly used in multiple vision tasks, including feature matching. The efficiency and memory footprint of handling large token sizes are the main limitations of transformer [54], where some methods [19, 20, 56] attempt to reduce the complexity to a linear scale to alleviate these problems. Some methods [8, 22] propose optimizing transformer models for specific hardware architectures for memory and running-time efficiency. They are orthogonal to our method and can be naturally adapted into the pipeline for further efficiency improvement." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.485, + 0.591, + 0.501 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.515, + 0.895, + 0.607 + ], + "angle": 0, + "content": "Given a pair of images \\(\\mathbf{I}_A, \\mathbf{I}_B\\), our objective is to establish a set of reliable correspondences between them. We achieve this by a coarse-to-fine matching pipeline, which first establishes coarse matches on downsampled feature maps and then refines them for high accuracy. An overview of our pipeline is shown in Fig. 2." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.628, + 0.73, + 0.643 + ], + "angle": 0, + "content": "3.1. Local Feature Extraction" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.655, + 0.895, + 0.761 + ], + "angle": 0, + "content": "Image feature maps are first extracted by a lightweight backbone for later transformation and matching. Unlike LoFTR and many other detector-free matchers that use a heavy multibranch ResNet [17] network for feature extraction, we alternate to a lightweight single-branch network with reparameterization [10] to achieve better inference efficiency while preserving the model performance." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.897, + 0.902 + ], + "angle": 0, + "content": "In particular, a multi-branch CNN network with residual connections is applied during training for maximum representational power. At inference time, we losslessly convert the feature backbone into an efficient single-branch network by adopting the reparameterization technique [10], which is achieved by fusing parallel convolution kernels into a single one. Then, the intermediate \\( \\frac{1}{8} \\) down-sampled coarse features \\( \\tilde{\\mathbf{F}}_A, \\tilde{\\mathbf{F}}_B \\) and fine features in \\( \\frac{1}{4} \\) and \\( \\frac{1}{2} \\) resolutions are extracted efficiently for later coarse-to-fine matching." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "21668" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.077, + 0.091, + 0.416, + 0.108 + ], + "angle": 0, + "content": "3.2. Efficient Local Feature Transformation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.114, + 0.471, + 0.178 + ], + "angle": 0, + "content": "After the feature extraction, the coarse-level feature maps \\(\\tilde{\\mathbf{F}}_A\\) and \\(\\tilde{\\mathbf{F}}_B\\) are transformed by interleaving self- and crossattention \\(^1 n\\) times to improve discriminativeness. The transformed features are denoted as \\(\\tilde{\\mathbf{F}}_A^t,\\tilde{\\mathbf{F}}_B^t\\)" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.179, + 0.473, + 0.299 + ], + "angle": 0, + "content": "Previous methods often perform attention on the entire coarse-level feature maps, where linear attention instead of vanilla attention is applied to ensure a manageable computation cost. However, the efficiency is still limited due to the large token size of coarse features. Moreover, the usage of linear attention leads to sub-optimal model capability. Unlike them, we propose efficient aggregated attention for both efficiency and performance." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.316, + 0.471, + 0.436 + ], + "angle": 0, + "content": "Preliminaries. First, we provide a brief overview of the commonly used vanilla attention and linear attention. Vanilla attention is a core mechanism in transformer encoder layer, relying on three inputs: query Q, key K, and value V. The resultant output is a weighted sum of the value, where the weighted matrix is determined by the query and its corresponding key. Formally, the attention function is defined as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.091, + 0.444, + 0.471, + 0.462 + ], + "angle": 0, + "content": "\\[\n\\operatorname {V a n i l l a A t t e n t i o n} (Q, K, V) = \\operatorname {s o f t m a x} (Q K ^ {T}) V. \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.469, + 0.471, + 0.543 + ], + "angle": 0, + "content": "However, applying the vanilla attention directly to dense local features is impractical due to the significant token size. To address this issue, previous methods use linear attention to reduce the computational complexity from quadratic to linear:" + }, + { + "type": "equation", + "bbox": [ + 0.088, + 0.551, + 0.47, + 0.57 + ], + "angle": 0, + "content": "\\[\n\\operatorname {L i n e a r A t t e n t i o n} (Q, K, V) = \\phi (Q) (\\phi (K) ^ {T} \\phi (V)). \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.576, + 0.471, + 0.623 + ], + "angle": 0, + "content": "where \\(\\phi (\\cdot) = elu(\\cdot) + 1\\). However, it comes at the cost of reduced representational power, which is also observed by [4]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.64, + 0.47, + 0.789 + ], + "angle": 0, + "content": "Aggregated Attention Module. After comprehensively investigating the mechanism of the transformer on coarse feature maps, we have two observations that motivate us to devise a new efficient aggregated attention. First, the attention regions of neighboring query tokens are similar, thus we can aggregate the neighboring tokens of \\( f_{i} \\) to prevent the redundant computation. Second, most of the attention weights of each query token are concentrated on a small number of key tokens, hence we can select the salient tokens of \\( f_{j} \\) before attention to reduce the computation." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.79, + 0.471, + 0.837 + ], + "angle": 0, + "content": "Therefore, we propose to first aggregate the \\( f_{i} \\) token utilizing a depth-wise convolution network, and \\( f_{j} \\) is aggregated by a max pooling layer to get reduced salient tokens:" + }, + { + "type": "equation", + "bbox": [ + 0.133, + 0.85, + 0.47, + 0.868 + ], + "angle": 0, + "content": "\\[\nf _ {i} ^ {\\prime} = \\operatorname {C o n v 2 D} \\left(f _ {i}\\right), f _ {j} ^ {\\prime} = \\operatorname {M a x P o o l} \\left(f _ {j}\\right) ， \\tag {3}\n\\]" + }, + { + "type": "page_footnote", + "bbox": [ + 0.076, + 0.875, + 0.469, + 0.901 + ], + "angle": 0, + "content": "1We feed feature of one image as query and feature of the other image as key and value into cross-attention, similar to SG [39] and LoFTR [46]." + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.062, + 0.672, + 0.359 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.684, + 0.061, + 0.888, + 0.36 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.372, + 0.892, + 0.496 + ], + "angle": 0, + "content": "Figure 3. Detailed Transformer Module Comparison. Unlike LoFTR which uses all tokens of feature maps to compute attention and resort to linear attention to reduce the computational cost, the proposed attention module first aggregates features for salient tokens, which is significantly more efficient for attention. Then the vanilla attention is utilized to transform aggregated features, where relative positional encoding is inserted to capture the spatial information. Transformed features are upsampled and fused with the original features to form the final features." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.498, + 0.893, + 0.694 + ], + "angle": 0, + "content": "where Conv2D is implemented by a strided depthwise convolution with a kernel size of \\( s \\times s \\), identical to that of the max-pooling layer. Then positional encoding and vanilla attention are followed to process reduced tokens. Positional encoding (PE) can help to model the spatial location contexts, where RoPE [45] is adopted in practice to account for more robust relative positions, inspired by [26]. Note that the PE layer is enabled exclusively for self-attention and skipped during cross-attention. The transformed feature map is then upsampled and fused with \\( f_{i} \\) for the final feature map. Due to the aggregation and selection, the number of tokens in \\( f_{i}' \\) and \\( f_{j}' \\) is reduced by \\( s^{2} \\), which contributes to the efficiency of the attention phase." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.703, + 0.773, + 0.719 + ], + "angle": 0, + "content": "3.3. Coarse-level Matching Module" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.726, + 0.895, + 0.901 + ], + "angle": 0, + "content": "We establish coarse-level matches based on the previously transformed coarse feature maps \\(\\tilde{\\mathbf{F}}_A^t\\), \\(\\tilde{\\mathbf{F}}_B^t\\). Coarse correspondences indicate rough match regions for later subpixel-level matching in the refinement phase. To achieve this, \\(\\tilde{\\mathbf{F}}_A^t\\) and \\(\\tilde{\\mathbf{F}}_B^t\\) are densely correlated to build a score matrix \\(S\\). The softmax operator on both \\(S\\) dimensions (referred to as dual-softmax) is then applied to obtain the probability of mutual nearest matching, which is commonly used in [35, 46, 53]. The coarse correspondences \\(\\{\\mathcal{M}_c\\}\\) are established by selecting matches above the score threshold \\(\\tau\\) while satisfying the mutual-nearest-neighbor (MNN) constraint." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "21669" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.212 + ], + "angle": 0, + "content": "Efficient Inference Strategy. We observe that the dual-softmax operator in the coarse matching can significantly restrict the efficiency in inference due to the large token size, especially for high-resolution images. Moreover, we find that the dual-softmax operator is crucial for training, dropping it at inference time while directly using the score matrix \\( S \\) for MNN matching can also work well with better efficiency." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.213, + 0.473, + 0.334 + ], + "angle": 0, + "content": "The reason for using the dual-softmax operator in training is that it can help to train discriminative features. Intuitively, with the softmax operation, the matching score between two pixels can also conditioned on other pixels. This mechanism forces the network to improve feature similarity of true correspondences while suppressing similarity with irrelevant points. With trained discriminative features, the softmax operation can be potentially eliminated during inference." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.334, + 0.473, + 0.38 + ], + "angle": 0, + "content": "We denote the model skipping dual-softmax layer in inference as efficiency optimized model. Results in Tab. 1 demonstrate the effectiveness of this design." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.39, + 0.385, + 0.405 + ], + "angle": 0, + "content": "3.4. Subpixel-Level Refinement Module" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.413, + 0.473, + 0.492 + ], + "angle": 0, + "content": "As overviewed in Fig. 2 (4), with established coarse matches \\(\\{\\mathcal{M}_c\\}\\), we refine them for sub-pixel accuracy with our refinement module. It is composed of an efficient feature patch extractor for discriminative fine features, followed by a two-stage feature correlation layer for final matches \\(\\{\\mathcal{M}_f\\}\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.51, + 0.473, + 0.633 + ], + "angle": 0, + "content": "Efficient Fine Feature Extraction. We first extract discriminative fine feature patches centered at each coarse match \\(\\mathcal{M}_c\\) by an efficient fusion network for later match refinement. For efficiency, our key idea here is to re-ileverage the previously transformed coarse features \\(\\tilde{\\mathbf{F}}_A^t\\), \\(\\tilde{\\mathbf{F}}_B^t\\) to obtain cross-view attended discriminative fine features, instead of introducing additional feature transform networks as in LoFTR [46]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.633, + 0.473, + 0.746 + ], + "angle": 0, + "content": "To be specific, \\(\\tilde{\\mathbf{F}}_A^t\\) and \\(\\tilde{\\mathbf{F}}_B^t\\) are adaptively fused with \\(1/4\\) and \\(1/2\\) resolution backbone features by convolution and upsampling to obtain fine feature maps \\(\\hat{\\mathbf{F}}_A^t\\), \\(\\hat{\\mathbf{F}}_B^t\\) in the original image resolution. Then \\(8 \\times 8\\) local feature patches are cropped on fine feature maps centered at each coarse match. Since only shallow feed-forward networks are included, our fine feature fusion network is remarkably efficient." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.473, + 0.903 + ], + "angle": 0, + "content": "Two-Stage Correlation for Refinement. Based on the extracted fine local feature patches of coarse matches, we search for high-accurate sub-pixel matches. To refine a coarse match, a commonly used strategy [6, 15, 46] is to select the center-patch feature of \\(\\mathbf{I}_A\\) as a fixed reference point, and perform feature correlation and expectation on the entire corresponding feature patch for its fine match. However, this refinement-by-expectations will introduce location variance to the final match, because irrelevant regions also" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.122 + ], + "angle": 0, + "content": "have weights and can affect results. Therefore, we propose a novel two-stage correlation module to solve this problem." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.123, + 0.895, + 0.273 + ], + "angle": 0, + "content": "Our idea is to utilize a mutual-nearest-neighbor (MNN) matching to get intermediate pixel-level refined matches in the first stage, and then refine them for subpixel accuracy by correlation and expectation. Motivations are that MNN matching don't have spatial variance since matches are selected by directly indexing pixels with maximum scores, but cannot achieve sub-pixel accuracy. Conversely, refinement-by-expectation can achieve sub-pixel accuracy but variance exists. The proposed two-stage refinement can draw benefits by combining the best of both worlds." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.274, + 0.895, + 0.38 + ], + "angle": 0, + "content": "Specifically, to refine a coarse-level correspondence \\(\\mathcal{M}_c\\), the first-stage refinement phase densely correlates their fine feature patches to obtain the local patch score matrix \\(S_l\\). MNN searching is then applied on \\(S_l\\) to get intermediate pixel-level fine matches. To limit the overall match number, we select the top-1 fine match for one coarse match by sorting the correlation scores." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.38, + 0.895, + 0.533 + ], + "angle": 0, + "content": "Then, we further refine these pixel-level matches for sub-pixel accuracy by our second-stage refinement. Since the matching accuracy has already significantly improved in first-stage refinement, now we can use a tiny local window for correlation and expectation with a maximum suppression of location variance. In practice, we correlate the feature of each point in \\(\\mathbf{I}_A\\) with a \\(3\\times 3\\) feature patch centered at its fine match in \\(\\mathbf{I}_B\\). The softmax operator is then applied to get a match distribution matrix and the final refined match is obtained by calculating expectations." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.543, + 0.629, + 0.559 + ], + "angle": 0, + "content": "3.5. Supervision" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.567, + 0.892, + 0.599 + ], + "angle": 0, + "content": "The entire pipeline is trained end-to-end by supervising the coarse and refinement matching modules separately." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.619, + 0.895, + 0.728 + ], + "angle": 0, + "content": "Coarse-Level Matching Supervision. The coarse ground truth matches \\(\\{\\mathcal{M}_c\\}_{gt}\\) with a total number of \\(N\\) are built by warping grid-level points from \\(\\mathbf{I}_A\\) to \\(\\mathbf{I}_B\\) via depth maps and image poses following previous methods [39, 46]. The produced correlation score matrix \\(\\mathcal{S}\\) in coarse matching is supervised by minimizing the log-likelihood loss over locations of \\(\\{\\mathcal{M}_c\\}_{gt}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.552, + 0.738, + 0.895, + 0.761 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {c} = - \\frac {1}{N} \\sum_ {(\\tilde {i}, \\tilde {j}) \\in \\left\\{\\mathcal {M} _ {c} \\right\\} _ {g t}} \\log \\mathcal {S} (\\tilde {i}, \\tilde {j}). \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.78, + 0.895, + 0.904 + ], + "angle": 0, + "content": "Fine-Level Matching Supervision. We train the proposed two-stage fine-level matching module by separately supervising the two phases. The first stage fine loss \\(\\mathcal{L}_{f1}\\) is to minimize the log-likelihood loss of each fine local score matrix \\(S_{l}\\) based on the pixel-level ground truth fine matches, similar to coarse loss. The second stage is trained by \\(\\mathcal{L}_{f2}\\) that calculates the \\(\\ell_2\\) loss between the final subpixel matches \\(\\{\\mathcal{M}_f\\}\\) and ground truth fine matches \\(\\{\\mathcal{M}_f\\}_{gt}\\)." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "21670" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.124, + 0.071, + 0.849, + 0.257 + ], + "angle": 0, + "content": "
CategoryMethodMegaDepth DatasetScanNet DatasetTime (ms)
AUC@5°AUC@10°AUC@20°AUC@5°AUC@10°AUC@20°
SparseSP + NN31.746.860.17.518.632.110.8
SP + SG49.767.180.616.232.849.748.3
SP + LG49.967.080.114.830.847.531.9/30.7
Semi-DenseDRC-Net27.042.958.37.717.930.5328.0
LoFTR52.869.281.216.933.650.666.2
QuadTree54.670.582.219.037.353.5100.7
MatchFormer53.369.781.815.832.048.0128.9
TopicFM54.170.181.617.335.550.966.4
AspanFormer55.371.583.119.637.754.481.6
Ours56.472.283.519.237.053.640.1/34.4
Ours (Optimized)55.471.482.917.434.451.235.6/27.0
DenseDKM60.474.985.126.6447.0764.17210.8
ROMA62.676.786.328.950.468.3302.7
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.272, + 0.892, + 0.342 + ], + "angle": 0, + "content": "Table 1. Results of Relative Pose Estimation on MegaDepth Dataset and ScanNet Dataset. We use the models trained on the MegaDepth dataset to evaluate all methods on both datasets, which can show the intra- and inter-dataset generalization abilities. The AUC of pose error at different thresholds, along with the processing time for matching image pair at a resolution of \\(640 \\times 480\\), is presented. For SP + LG, Ours, and Ours (Optimized), the running times of the model using FP32/Mixed-Precision numerical precisions (without Efficient Attention [8, 22]) are shown." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.357, + 0.471, + 0.388 + ], + "angle": 0, + "content": "The total loss is the weighted sum of all supervisions: \\(\\mathcal{L} = \\mathcal{L}_c + \\alpha \\mathcal{L}_{f1} + \\beta \\mathcal{L}_{f2}\\)" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.399, + 0.21, + 0.417 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.425, + 0.471, + 0.501 + ], + "angle": 0, + "content": "In this section, we evaluate the performance of our method on several downstream tasks, including homography estimation, pairwise pose estimation and visual localization. Furthermore, we evaluate the effectiveness of our design by conducting detailed ablation studies." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.509, + 0.295, + 0.525 + ], + "angle": 0, + "content": "4.1. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.532, + 0.472, + 0.759 + ], + "angle": 0, + "content": "We adopt RepVGG [10] as our feature backbone, and self- and cross-attention are interleaved for \\( N = 4 \\) times to transform coarse features. For each attention, we aggregate features by a depth-wise convolution layer and a max-pooling layer, both with a kernel size of \\( 4 \\times 4 \\). Our model is trained on the MegaDepth dataset [24], which is a large-scale outdoor dataset. The test scenes are separated from training data following [46]. The loss function's weights \\( \\alpha \\) and \\( \\beta \\) are set to 1.0 and 0.25, respectively. We use the AdamW optimizer with an initial learning rate of \\( 4 \\times 10^{-3} \\). The network training takes about 15 hours with a batch size of 20 on 4 NVIDIA V100 GPUs. And the coarse and fine stages are trained together from scratch. The trained model on MegaDepth is used to evaluate all datasets and tasks in our experiments to demonstrate the generalization ability." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.767, + 0.304, + 0.782 + ], + "angle": 0, + "content": "4.2. Relative Pose Estimation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.795, + 0.469, + 0.84 + ], + "angle": 0, + "content": "Datasets. We use the outdoor MegaDepth [24] dataset and indoor ScanNet [7] dataset for the evaluation of relative pose estimation to demonstrate the efficacy of our method." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.471, + 0.901 + ], + "angle": 0, + "content": "MegaDepth dataset is a large-scale dataset containing sparse 3D reconstructions from 196 scenes. The key challenges on this dataset are large viewpoints and illumination changes, as well as repetitive patterns. We follow the test" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.356, + 0.895, + 0.446 + ], + "angle": 0, + "content": "split of the previous method [46] that uses 1500 sampled pairs from scenes \"Sacre Coeur\" and \"St. Peter's Square\" for evaluation. Images are resized so that the longest edge equals 1200 for all semi-dense and dense methods. Following [26], sparse methods are provided resized images with longest edge equals 1600." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.449, + 0.895, + 0.525 + ], + "angle": 0, + "content": "ScanNet dataset contains 1613 sequences with ground-truth depth maps and camera poses. They depict indoor scenes with viewpoint changes and texture-less regions. We use the sampled test pairs from [39] for the evaluation, where images are resized to \\(640 \\times 480\\) for all methods." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.531, + 0.895, + 0.668 + ], + "angle": 0, + "content": "Baselines. We compare the proposed method with three categories of methods: 1) sparse keypoint detection and matching methods, including SuperPoint [9] with Nearest-Neighbor (NN), SuperGlue (SG) [39], LightGlue (LG) [26] matchers. 2) semi-dense matchers, including DRC-Net [23], LoFTR [46], QuadTree Attention [48], MatchFormer [55], AspanFormer [6], TopicFM [15], and 3) state-of-the-art dense matcher ROMA [14] that predict matches for each pixel." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.674, + 0.897, + 0.901 + ], + "angle": 0, + "content": "Evaluation protocol. Following previous methods, the recovered relative poses by matches are evaluated for reflecting matching accuracy. The pose error is defined as the maximum of angular errors in rotation and translation. We report the AUC of the pose error at thresholds \\((5^{\\circ}, 10^{\\circ}\\), and \\(20^{\\circ})\\). Moreover, the running time of matching each image pair in the ScanNet dataset on a single NVIDIA RTX 3090 is reported for comprehensively understanding the matching accuracy and efficiency balance. We also include the running time of feature extraction and keypoint detection (SuperPoint) for the sparse method to compare with dense methods, which are end-to-end (from image input to matches output). LG [26]'s SuperPoint [9] parameters (max keypoints=2048, keypoint threshold=0.0005, nms=4) are uniformly used for all sparse methods." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "21671" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.084, + 0.071, + 0.468, + 0.193 + ], + "angle": 0, + "content": "
CategoryMethodHomography est. AUC
@3px@5px@10px
SparseD2Net + NN23.235.953.6
R2D2 + NN50.663.976.8
DISK + NN52.364.978.9
SP + SG53.968.381.7
Semi-DenseSparse-NCNet48.954.267.1
DRC-Net50.656.268.3
LoFTR65.975.684.6
Ours66.576.485.5
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.199, + 0.471, + 0.255 + ], + "angle": 0, + "content": "Table 2. Results of Homography Estimation on HPatches Dataset. Our method is compared with sparse and semi-dense methods. The AUC of reprojection error of corner points at different thresholds is reported." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.272, + 0.473, + 0.575 + ], + "angle": 0, + "content": "Results. As shown in Tab. 1, the proposed method achieves competitive performances compared with sparse and semidense methods on both datasets. Qualitative comparisons are shown in Fig. 4. Specifically, our method outperforms the best semi-dense baseline AspanFormer on all metrics of the MegaDepth dataset and has lower but comparable performance on the ScanNet dataset, with \\(\\sim 2\\) times faster. Our optimized model that eliminates the dual-softmax operator in coarse-level matching further brings efficiency improvements, with slight performance decreases. Using this strategy, our method can outperform the efficient and robust sparse method \\(\\mathrm{SP} + \\mathrm{LG}\\) in efficiency with significantly higher accuracy. Dense matcher ROMA shows remarkable matching capability but is slow for applications in practice. Moreover, since ROMA utilizes the pre-trained DINOv2 [33] backbone, its strong generalizability on ScanNet may be attributed to the similar indoor training data in DINOv2, where other methods are trained on outdoor MegaDepth only. Compared with it, our method is \\(\\sim 7.5\\) times faster, which has a good balance between accuracy and efficiency." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.587, + 0.304, + 0.604 + ], + "angle": 0, + "content": "4.3. Homography Estimation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.617, + 0.471, + 0.678 + ], + "angle": 0, + "content": "Dataset. We evaluate our method on HPatches dataset [2]. HPatches dataset depicts planar scenes divided into sequences. Images are taken under different viewpoints or illumination changes." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.683, + 0.47, + 0.789 + ], + "angle": 0, + "content": "Baselines. We compare our method with sparse methods including D2Net [11], R2D2 [34], DISK [53] detectors with NN matcher, and SuperPoint [9] + SuperGlue [39]. As for semi-dense methods, we compare with Sparse-NCNet [36], DRC-Net [23], and LoFTR [46]. For SuperGlue and all semidense methods, we use their models trained on MegaDepth dataset for evaluation." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.795, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Evaluation Protocol. Following SuperGlue [39] and LoFTR [46], we resize all images for matching so that their smallest edge equals 480 pixels. We collect the mean reprojection error of corner points, and report the area under the cumulative curve (AUC) under 3 different thresholds, including 3 px, 5 px, and 10 px. For all baselines, we employ the same RANSAC method as a robust homography estimator" + }, + { + "type": "table", + "bbox": [ + 0.508, + 0.071, + 0.89, + 0.17 + ], + "angle": 0, + "content": "
MethodDUC1DUC2
(0.25m,2°)/(0.5m,5°)/(1.0m,10°)
SP+SG49.0 / 68.7 / 80.853.4 / 77.1 / 82.4
LoFTR47.5 / 72.2 / 84.854.2 / 74.8 / 85.5
TopicFM52.0 / 74.7 / 87.453.4 / 74.8 / 83.2
PATS55.6 / 71.2 / 81.058.8 / 80.9 / 85.5
AspanFormer51.5 / 73.7 / 86.055.0 / 74.0 / 81.7
Ours52.0 / 74.7 / 86.958.0 / 80.9 / 89.3
" + }, + { + "type": "table_caption", + "bbox": [ + 0.517, + 0.176, + 0.874, + 0.189 + ], + "angle": 0, + "content": "Table 3. Results of Visual Localization on InLoc Dataset." + }, + { + "type": "table", + "bbox": [ + 0.508, + 0.2, + 0.89, + 0.298 + ], + "angle": 0, + "content": "
MethodDayNight
(0.25m,2°)/(0.5m,5°)/(1.0m,10°)
SP+SG89.8 / 96.1 / 99.477.0 / 90.6 / 100.0
LoFTR88.7 / 95.6 / 99.078.5 / 90.6 / 99.0
TopicFM90.2 / 95.9 / 98.977.5 / 91.1 / 99.5
PATS89.6 / 95.8 / 99.373.8 / 92.1 / 99.5
AspanFormer89.4 / 95.6 / 99.077.5 / 91.6 / 99.5
Ours89.6 / 96.2 / 99.077.0 / 91.1 / 99.5
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.304, + 0.892, + 0.318 + ], + "angle": 0, + "content": "Table 4. Results of Visual Localization on Aachen v1.1 Dataset." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.334, + 0.892, + 0.379 + ], + "angle": 0, + "content": "for a fair comparison. Following LoFTR, we select only the top 1000 predicted matches of semi-dense methods for the sake of fairness." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.385, + 0.895, + 0.492 + ], + "angle": 0, + "content": "Results. As shown in Tab. 2, even though the number of matches is restricted, our method can also work remarkably well and outperform sparse methods significantly. Compared with semi-dense, our method can surpass them with significantly higher efficiency. We attribute this to the effectiveness of two-stage refinement for accuracy improvement and proposed aggregation module for efficiency." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.503, + 0.684, + 0.518 + ], + "angle": 0, + "content": "4.4. Visual Localization" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.532, + 0.895, + 0.744 + ], + "angle": 0, + "content": "Datasets and Evaluation Protocols. Visual localization is an important downstream task of image matching, which aims to estimate the 6-DoF poses of query images based on the 3D scene model. We conduct experiments on two commonly used benchmarks, including InLoc [47] dataset and Aachen v1.1 [41] dataset, for evaluation to demonstrate the superiority of our method. InLoc dataset is captured on indoor scenes with plenty of repetitive structures and textureless regions, where each database image has a corresponding depth map. Aachen v1.1 is a challenging large-scale outdoor dataset for localization with large-viewpoint and day-and-night illumination changes, which particularly relies on the robustness of matching methods. We adopt its full localization track for benchmarking." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.744, + 0.895, + 0.866 + ], + "angle": 0, + "content": "Following [6, 46], the open-sourced localization framework HLoc [38] is utilized. For both datasets, the percentage of pose errors satisfying both angular and distance thresholds is reported following the benchmarks, where different thresholds are used. For the InLoc dataset, the metrics of two test scenes including DUC1 and DUC2 are separately reported. As for the Aachen v1.1 dataset, the metrics corresponding to the daytime and nighttime divisions are reported." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.892, + 0.902 + ], + "angle": 0, + "content": "Baselines. We compare the proposed method with both detector-based method SuperPoint [9] + SuperGlue [39] and" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "21672" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.145, + 0.049, + 0.378, + 0.248 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.39, + 0.049, + 0.605, + 0.25 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.619, + 0.051, + 0.831, + 0.25 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.25, + 0.895, + 0.293 + ], + "angle": 0, + "content": "Figure 4. Qualitative Results. Our method is compared with the sparse matching pipeline SuperPoint [9] \\(^+\\) LightGlue [26], semi-dense \nmacher AspanFormer [6]. Image pairs with texture-poor regions and large-viewpoint changes can be robustly matched by our method. The \nred color indicates epipolar error beyond \\(5\\times 10^{-4}\\) (in the normalized image coordinates)." + }, + { + "type": "table", + "bbox": [ + 0.083, + 0.301, + 0.468, + 0.404 + ], + "angle": 0, + "content": "
MethodPose Estimation AUCTime (ms)
@5°@10°@20°
Ours Full56.472.283.5139.2
1) Ours Optimal (w/o dual-softmax)55.471.482.9102.0
2) Replace Agg. Attention to LoFTR's Trans.54.770.582.2171.4
3) Replace two-stage refine. to LoFTR's refine.54.770.982.7135.3
4) No second-stage refinement55.871.883.3138.1
5) Replace RepVGG with ResNet55.471.482.9156.2
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.408, + 0.472, + 0.466 + ], + "angle": 0, + "content": "Table 5. Ablation Studies. The components of our method are ablated on the MegaDepth dataset for a comprehensive understanding of our method, where averaged running times for an image pair with high-resolution \\(1200 \\times 1200\\) are reported." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.478, + 0.471, + 0.509 + ], + "angle": 0, + "content": "detector-free methods including LoFTR [46], TopicFM [15], PATS [32] and Aspanformer [6]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.513, + 0.473, + 0.651 + ], + "angle": 0, + "content": "Results. We adhere to the pipeline and evaluation settings of the online visual localization benchmark² to ensure fairness. As presented in Tab. 3, the proposed method achieves competitive results, taking both detector-based and detector-free methods into account. Being a method primarily geared towards efficiency, our approach can deliver results comparable to those of many accuracy-oriented methods. As depicted in Tab. 4, our method also demonstrates performance on par with the best-performing approaches." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.659, + 0.245, + 0.675 + ], + "angle": 0, + "content": "4.5. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.682, + 0.473, + 0.879 + ], + "angle": 0, + "content": "In this part, we conduct detailed ablation studies to analyze the effectiveness of our proposed modules with results shown in Tab. 5. 1) Without dual-softmax, our optimal model can bring huge efficiency improvement in high-resolution images. 2) In coarse feature transformation, replacing the proposed aggregated attention module with LoFTR's transformer can bring significant efficiency dropping, as well as accuracy decrease. Note that the replaced transformer is also equipped with RoPE same as ours for fair comparison. This demonstrates the efficacy of the proposed module that performing vanilla attention on aggregated features can achieve higher efficiency with even better matching accuracy. 3) Compared with using LoFTR's refinement that performs" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.304, + 0.896, + 0.471 + ], + "angle": 0, + "content": "expectation on the entire correlation patch, the proposed two-stage refinement layer can bring accuracy improvement with negligible latency. We attribute this to the two-stage refinement's property that can maximize the suppression of location variance in correlation refinement. 4) Dropping the second refinement stage will lead to degraded pose accuracy with minor efficiency changes, especially on the strict AUC@5° metric. 5) Changing the backbone from reparameterized VGG [10] back to multi-branch ResNet [17] leads to decreased efficiency with similar accuracy, which demonstrates the effectiveness of our design choice for efficiency." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.483, + 0.628, + 0.499 + ], + "angle": 0, + "content": "5. Conclusions" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.508, + 0.896, + 0.812 + ], + "angle": 0, + "content": "This paper introduces a new semi-dense local feature matcher based on the success of LoFTR. We revisit its designs and propose several improvements for both efficiency and matching accuracy. A key observation is that performing the Transformer on the entire coarse feature map is redundant due to the similar local information, where an aggregated attention module is proposed to perform transformer on reduced tokens with significantly better efficiency and competitive performance. Moreover, a two-stage correlation layer is devised to solve the location variance problem in LoFTR's refinement design, which further brings accuracy improvements. As a result, our method can achieve \\(\\sim 2.5\\) times faster compared with LoFTR with better matching accuracy. Moreover, as a semi-dense matching method, the proposed method can achieve comparable efficiency with the recent robust sparse feature matcher LightGlue [26]. We believe this opens up the applications of our method in large-scale or latency-sensitive downstream tasks, such as image retrieval and 3D reconstruction. Please refer to the supplementary material for discussions about limitations and future works." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.829, + 0.897, + 0.89 + ], + "angle": 0, + "content": "Acknowledgement The authors would like to acknowledge support from NSFC (No. 62172364) and Information Technology Center and State Key Lab of CAD&CG, Zhejiang University." + }, + { + "type": "page_footnote", + "bbox": [ + 0.095, + 0.887, + 0.348, + 0.9 + ], + "angle": 0, + "content": "\\(^{2}\\)https://wwwVisualization.net/benchmark" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "21673" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.471, + 0.156 + ], + "angle": 0, + "content": "[1] Sameer Agarwal, Yasutaka Furukawa, Noah Snavely, Ian Simon, Brian Curless, Steven M. Seitz, and Richard Szeliski. Building rome in a day. ICCV, 2009. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.157, + 0.472, + 0.212 + ], + "angle": 0, + "content": "[2] Vassileios Balntas, Karel Lenc, Andrea Vedaldi, and Krystian Mikolajczyk. Hpatches: A benchmark and evaluation of handcrafted and learned local descriptors. In CVPR, pages 5173-5182, 2017. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.213, + 0.471, + 0.255 + ], + "angle": 0, + "content": "[3] Herbert Bay, Andreas Ess, Tinne Tuytelaars, and Luc Van Gool. Speeded-up robust features (surf). CVIU, 110 (3):346-359, 2008. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.256, + 0.471, + 0.297 + ], + "angle": 0, + "content": "[4] Han Cai, Chuang Gan, and Song Han. Efficientvit: Enhanced linear attention for high-resolution low-computation visual recognition. arXiv preprint arXiv:2205.14756, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.298, + 0.471, + 0.354 + ], + "angle": 0, + "content": "[5] Hongkai Chen, Zixin Luo, Jiahui Zhang, Lei Zhou, Xuyang Bai, Zeyu Hu, Chiew-Lan Tai, and Long Quan. Learning to match features with seeded graph matching network. In ICCV, pages 6301-6310, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.354, + 0.471, + 0.422 + ], + "angle": 0, + "content": "[6] Hongkai Chen, Zixin Luo, Lei Zhou, Yurun Tian, Mingmin Zhen, Tian Fang, David McKinnon, Yanghai Tsin, and Long Quan. Aspanformer: Detector-free image matching with adaptive span transformer. In ECCV, pages 20-36. Springer, 2022. 2, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.424, + 0.471, + 0.479 + ], + "angle": 0, + "content": "[7] Angela Dai, Angel X Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In CVPR, pages 5828-5839, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.48, + 0.471, + 0.521 + ], + "angle": 0, + "content": "[8] Tri Dao, Daniel Y. Fu, Stefano Ermon, Atri Rudra, and Christopher Ré. FlashAttention: Fast and memory-efficient exact attention with IO-awareness. In NeurIPS, 2022. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.522, + 0.471, + 0.564 + ], + "angle": 0, + "content": "[9] Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. Superpoint: Self-supervised interest point detection and description. CVPRW, 2018. 2, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.565, + 0.471, + 0.619 + ], + "angle": 0, + "content": "[10] Xiaohan Ding, Xiangyu Zhang, Ningning Ma, Jungong Han, Guiguang Ding, and Jian Sun. Repvgg: Making vgg-style convnets great again. In CVPR, pages 13733-13742, 2021. 3, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.621, + 0.471, + 0.676 + ], + "angle": 0, + "content": "[11] Mihai Dusmanu, Ignacio Rocco, Tomas Pajdla, Marc Pollefeys, Josef Sivic, Akihiko Torii, and Torsten Sattler. D2-net: A trainable cnn for joint description and detection of local features. In CVPR, pages 8092-8101, 2019. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.677, + 0.471, + 0.718 + ], + "angle": 0, + "content": "[12] Patrick Ebel, Anastasiia Mishchuk, Kwang Moo Yi, Pascal Fua, and Eduard Trulls. Beyond cartesian representations for local descriptors. In ICCV, pages 253-262, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.719, + 0.471, + 0.773 + ], + "angle": 0, + "content": "[13] Johan Edstedt, Ioannis Athanasiadis, Mårten Wadenbäck, and Michael Felsberg. Dkm: Dense kernelized feature matching for geometry estimation. In CVPR, pages 17765-17775, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.775, + 0.471, + 0.829 + ], + "angle": 0, + "content": "[14] Johan Edstedt, Qiyu Sun, Georg Bokman, Mårten Wadenback, and Michael Felsberg. Roma: Revisiting robust losses for dense feature matching. arXiv preprint arXiv:2305.15404, 2023. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.831, + 0.471, + 0.872 + ], + "angle": 0, + "content": "[15] Khang Truong Giang, Soohwan Song, and Sung-Guk Jo. Top-icfm: Robust and interpretable topic-assisted feature matching. In AAAI, 2023. 2, 3, 5, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.471, + 0.9 + ], + "angle": 0, + "content": "[16] Stephen Hausler, Sourav Garg, Ming Xu, Michael Milford, and Tobias Fischer. Patch-netvlad: Multi-scale fusion of" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.472, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "locally-global descriptors for place recognition. CVPR, pages 14136-14147, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.122, + 0.894, + 0.162 + ], + "angle": 0, + "content": "[17] Kaiming He, X. Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. CVPR, pages 770-778, 2015. 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.164, + 0.894, + 0.205 + ], + "angle": 0, + "content": "[18] Xingyi He, Jiaming Sun, Yifan Wang, Sida Peng, Qixing Huang, Hujun Bao, and Xiaowei Zhou. Detector-free structure from motion. In CVPR, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.207, + 0.892, + 0.248 + ], + "angle": 0, + "content": "[19] Angelos Katharopoulos, Apoorv Vyas, Nikolaos Pappas, and Francois Fleuret. Transformers are rnns: Fast autoregressive transformers with linear attention. In ICML, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.25, + 0.894, + 0.289 + ], + "angle": 0, + "content": "[20] Nikita Kitaev, Lukasz Kaiser, and Anselm Levskaya. Reformer: The efficient transformer. ArXiv, abs/2001.04451, 2020.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.291, + 0.892, + 0.333 + ], + "angle": 0, + "content": "[21] Axel Barroso Laguna, Edgar Riba, Daniel Ponsa, and Krystian Mikolajczyk. Key.net: Keypoint detection by handcrafted and learned cnn filters. ICCV, pages 5835-5843, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.335, + 0.894, + 0.417 + ], + "angle": 0, + "content": "[22] Benjamin Lefaudeau, Francisco Massa, Diana Liskovich, Wenhan Xiong, Vittorio Caggiano, Sean Naren, Min Xu, Jieru Hu, Marta Tintore, Susan Zhang, Patrick Labatut, and Daniel Haziza. xformers: A modular and hackable transformer modelling library. https://github.com/facebookresearch/xformers, 2022.3.6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.419, + 0.894, + 0.459 + ], + "angle": 0, + "content": "[23] Xinghui Li, Kai Han, Shuda Li, and Victor Prisacariu. Dualresolution correspondence networks. In NeurlPS, 2020. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.461, + 0.892, + 0.502 + ], + "angle": 0, + "content": "[24] Zhengqi Li and Noah Snavely. Megadepth: Learning single-view depth prediction from internet photos. In CVPR, pages 2041-2050, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.504, + 0.892, + 0.546 + ], + "angle": 0, + "content": "[25] Philipp Lindenberger, Paul-Edouard Sarlin, Viktor Larsson, and Marc Pollefeys. Pixel-perfect structure-from-motion with featuremetric refinement. ICCV, pages 5967-5977, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.547, + 0.892, + 0.588 + ], + "angle": 0, + "content": "[26] Philipp Lindenberger, Paul-Edouard Sarlin, and Marc Pollefeys. LightGlue: Local Feature Matching at Light Speed. In ICCV, 2023. 1, 2, 4, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.589, + 0.892, + 0.617 + ], + "angle": 0, + "content": "[27] G LoweDavid. Distinctive image features from scale-invariant keypoints. IJCV, 2004. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.618, + 0.892, + 0.673 + ], + "angle": 0, + "content": "[28] Zixin Luo, Lei Zhou, Xuyang Bai, Hongkai Chen, Jiahui Zhang, Yao Yao, Shiwei Li, Tian Fang, and Long Quan. Aslfeat: Learning local features of accurate shape and localization. In CVPR, pages 6589-6598, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.675, + 0.892, + 0.716 + ], + "angle": 0, + "content": "[29] Anastasiia Mishchuk, Dmytro Mishkin, Filip Radenovic, and Jiri Matas. Working hard to know your neighbor's margins: Local descriptor learning loss. NeurIPS, 30, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.718, + 0.892, + 0.758 + ], + "angle": 0, + "content": "[30] Raul Mur-Artal and Juan D. Tardós. Orb-slam2: An open-source slam system for monocular, stereo, and rgb-d cameras. TR, 33:1255-1262, 2016. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.76, + 0.892, + 0.801 + ], + "angle": 0, + "content": "[31] Raul Mur-Artal, José M. M. Montiel, and Juan D. Tardós. Orb-slam: A versatile and accurate monocular slam system. TR, 31:1147-1163, 2015. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.802, + 0.892, + 0.858 + ], + "angle": 0, + "content": "[32] Junjie Ni, Yijin Li, Zhaoyang Huang, Hongsheng Li, Hujun Bao, Zhaopeng Cui, and Guofeng Zhang. Pats: Patch area transportation with subdivision for local feature matching. CVPR, pages 17776-17786, 2023. 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.859, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[33] Maxime Oquab, Timothée Darcet, Theo Moutakanni, Huy V. Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, Russell" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.894, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "21674" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.092, + 0.472, + 0.174 + ], + "angle": 0, + "content": "Howes, Po-Yao Huang, Hu Xu, Vasu Sharma, Shang-Wen Li, Wojciech Galuba, Mike Rabbat, Mido Assran, Nicolas Ballas, Gabriel Synnaeve, Ishan Misra, Herve Jegou, Julien Mairal, Patrick Labatut, Armand Joulin, and Piotr Bojanowski. Dinov2: Learning robust visual features without supervision, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.176, + 0.472, + 0.218 + ], + "angle": 0, + "content": "[34] Jérôme Revaud, César Roberto de Souza, M. Humenberger, and Philippe Weinzaepfel. R2d2: Reliable and repeatable detector and descriptor. In NeurIPS, 2019. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.218, + 0.472, + 0.259 + ], + "angle": 0, + "content": "[35] Ignacio Rocco, Mircea Cimpoi, Relja Arandjelović, Akihiko Torii, Tomas Pajdla, and Josef Sivic. Neighbourhood consensus networks. *NeurIPS*, 31, 2018. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.26, + 0.472, + 0.302 + ], + "angle": 0, + "content": "[36] Ignacio Rocco, Relja Arandjelović, and Josef Sivic. Efficient neighbourhood consensus networks via submanifold sparse convolutions. In ECCV, pages 605-621. Springer, 2020. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.302, + 0.472, + 0.344 + ], + "angle": 0, + "content": "[37] Edward Rosten and Tom Drummond. Machine learning for high-speed corner detection. In ECCV, pages 430-443. Springer, 2006. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.344, + 0.472, + 0.398 + ], + "angle": 0, + "content": "[38] Paul-Edouard Sarlin, César Cadena, Roland Y. Siegwart, and Marcin Dymczyk. From coarse to fine: Robust hierarchical localization at large scale. CVPR, pages 12708–12717, 2019. 1, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.4, + 0.472, + 0.453 + ], + "angle": 0, + "content": "[39] Paul-Edouard Sarlin, Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. SuperGlue: Learning feature matching with graph neural networks. In CVPR, 2020. 1, 2, 4, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.455, + 0.472, + 0.524 + ], + "angle": 0, + "content": "[40] Paul-Edouard Sarlin, Ajaykumar Unagar, Måns Larsson, Hugo Germain, Carl Toft, Victor Larsson, Marc Pollefeys, Vincent Lepetit, Lars Hammarstrand, Fredrik Kahl, and Torsten Sattler. Back to the Feature: Learning Robust Camera Localization from Pixels to Pose. In CVPR, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.525, + 0.472, + 0.594 + ], + "angle": 0, + "content": "[41] Torsten Sattler, Will Maddern, Carl Toft, Akihiko Torii, Lars Hammarstrand, Erik Stenberg, Daniel Safari, Masatoshi Okutomi, Marc Pollefeys, Josef Sivic, et al. Benchmarking 6dof outdoor visual localization in changing conditions. In CVPR, pages 8601-8610, 2018. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.595, + 0.472, + 0.649 + ], + "angle": 0, + "content": "[42] Nikolay Savinov, Akihito Seki, Lubor Ladicky, Torsten Sattler, and Marc Pollefeys. Quad-networks: unsupervised learning to rank for interest point detection. In CVPR, pages 1822-1830, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.65, + 0.472, + 0.677 + ], + "angle": 0, + "content": "[43] Johannes L. Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In CVPR, 2016. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.679, + 0.472, + 0.734 + ], + "angle": 0, + "content": "[44] Yan Shi, Jun-Xiong Cai, Yoli Shavit, Tai-Jiang Mu, Wensen Feng, and Kai Zhang. Clustergnn: Cluster-based coarse-to-fine graph neural network for efficient feature matching. In CVPR, pages 12517-12526, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.735, + 0.472, + 0.775 + ], + "angle": 0, + "content": "[45] Jianlin Su, Yu Lu, Shengfeng Pan, Bo Wen, and Yunfeng Liu. Roformer: Enhanced transformer with rotary position embedding. ArXiv, abs/2104.09864, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.776, + 0.472, + 0.831 + ], + "angle": 0, + "content": "[46] Jiaming Sun, Zehong Shen, Yuang Wang, Hujun Bao, and Xiaowei Zhou. Loftr: Detector-free local feature matching with transformers. In CVPR, pages 8922-8931, 2021. 1, 2, 4, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.832, + 0.472, + 0.9 + ], + "angle": 0, + "content": "[47] Hajime Taira, Masatoshi Okutomi, Torsten Sattler, Mircea Cimpoi, Marc Pollefeys, Josef Sivic, Tomas Pajdla, and Akihiko Torii. Inloc: Indoor visual localization with dense matching and view synthesis. In CVPR, pages 7199-7209, 2018. 7" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "[48] Shitao Tang, Jiahui Zhang, Siyu Zhu, and Ping Tan. Quadtree attention for vision transformers. *ICLR*, 2022. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.122, + 0.892, + 0.163 + ], + "angle": 0, + "content": "[49] Yurun Tian, Bin Fan, and Fuchao Wu. L2-net: Deep learning of discriminative patch descriptor in euclidean space. In CVPR, pages 661-669, 2017. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.164, + 0.892, + 0.219 + ], + "angle": 0, + "content": "[50] Yurun Tian, Xin Yu, Bin Fan, Fuchao Wu, Huub Heijnen, and Vassileios Balntas. Sosnet: Second order similarity regularization for local descriptor learning. In CVPR, pages 11016-11025, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.221, + 0.892, + 0.275 + ], + "angle": 0, + "content": "[51] Yurun Tian, Vassileios Balntas, Tony Ng, Axel Barroso-Laguna, Yiannis Demiris, and Krystian Mikolajczyk. D2d: Keypoint extraction with describe to detect approach. In ACCV, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.278, + 0.892, + 0.319 + ], + "angle": 0, + "content": "[52] Prune Truong, Martin Danelljan, Luc Van Gool, and Radu Timofte. Learning accurate dense correspondences and when to trust them. In CVPR, pages 5714-5724, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.321, + 0.892, + 0.36 + ], + "angle": 0, + "content": "[53] Michal Tyszkiiewicz, Pascal Fua, and Eduard Trulls. Disk: Learning local features with policy gradient. NeurIPS, 2020. 4, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.363, + 0.892, + 0.417 + ], + "angle": 0, + "content": "[54] Ashish Vaswani, Noam M. Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In NeurIPS, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.42, + 0.892, + 0.461 + ], + "angle": 0, + "content": "[55] Qing Wang, Jiaming Zhang, Kailun Yang, Kunyu Peng, and Rainer Stiefelhagen. Matchformer: Interleaving attention in transformers for feature matching. In ACCV, 2022. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.462, + 0.892, + 0.503 + ], + "angle": 0, + "content": "[56] Sinong Wang, Belinda Z. Li, Madian Khabsa, Han Fang, and Hao Ma. Linformer: Self-attention with linear complexity. ArXiv, abs/2006.04768, 2020. 3" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.503 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "21675" + } + ] +] \ No newline at end of file diff --git a/2024/Efficient LoFTR_ Semi-Dense Local Feature Matching with Sparse-Like Speed/14a55345-a1a3-4ac4-a5a2-7c153c54b832_origin.pdf b/2024/Efficient LoFTR_ Semi-Dense Local Feature Matching with Sparse-Like Speed/14a55345-a1a3-4ac4-a5a2-7c153c54b832_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..0c08c198031d8a28464e37f1633bb28cc31efb6b --- /dev/null +++ b/2024/Efficient LoFTR_ Semi-Dense Local Feature Matching with Sparse-Like Speed/14a55345-a1a3-4ac4-a5a2-7c153c54b832_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:fbac646e9320ae1b6bd31182287418a0876d36ec95767f9f3f0a6138a3700916 +size 5715512 diff --git a/2024/Efficient LoFTR_ Semi-Dense Local Feature Matching with Sparse-Like Speed/full.md b/2024/Efficient LoFTR_ Semi-Dense Local Feature Matching with Sparse-Like Speed/full.md new file mode 100644 index 0000000000000000000000000000000000000000..1f7ad3d42d11a8040d716bd0cff67de4a833c308 --- /dev/null +++ b/2024/Efficient LoFTR_ Semi-Dense Local Feature Matching with Sparse-Like Speed/full.md @@ -0,0 +1,296 @@ +# Efficient LoFTR: Semi-Dense Local Feature Matching with Sparse-Like Speed + +Yifan Wang* Xingyi He* Sida Peng† Dongli Tan Xiaowei Zhou + +Zhejiang University + +# Abstract + +We present a novel method for efficiently producing semidense matches across images. Previous detector-free matcher LoFTR has shown remarkable matching capability in handling large-viewpoint change and texture-poor scenarios but suffers from low efficiency. We revisit its design choices and derive multiple improvements for both efficiency and accuracy. One key observation is that performing the transformer over the entire feature map is redundant due to shared local information, therefore we propose an aggregated attention mechanism with adaptive token selection for efficiency. Furthermore, we find spatial variance exists in LoFTR's fine correlation module, which is adverse to matching accuracy. A novel two-stage correlation layer is proposed to achieve accurate subpixel correspondences for accuracy improvement. Our efficiency optimized model is $\sim 2.5\times$ faster than LoFTR which can even surpass state-of-the-art efficient sparse matching pipeline SuperPoint + LightGlue. Moreover, extensive experiments show that our method can achieve higher accuracy compared with competitive semi-dense matchers, with considerable efficiency benefits. This opens up exciting prospects for large-scale or latency-sensitive applications such as image retrieval and 3D reconstruction. Project page: https://zju3dv.github.io/efficientloftr/. + +# 1. Introduction + +Image matching is the cornerstone of many 3D computer vision tasks, which aim to find a set of highly accurate correspondences given an image pair. The established matches between images have broad usages such as reconstructing the 3D world by structure from motion (SfM) [1, 18, 25, 43] or SLAM system [30, 31], and visual localization [38, 40], etc. Previous methods typically follow a two-stage pipeline: they first detect [37] and describe [49] a set of keypoints on each image, and then establish keypoint correspondences by handcrafted [27] or learning-based matchers [26, 39]. These detector-based methods are efficient but suffer from robustly + +![](images/dcc10e8a010fd52aed100c9bace3e941ad983eee32cb58fdce6c1aee64e93e25.jpg) +Figure 1. Matching Accuracy and Efficiency Comparisons. Our method achieves competitive accuracy compared with semi-dense matchers $(\odot)$ at a significantly higher speed. Compared with dense matcher ROMA $(\odot)$ , our method is $\sim 7.5\times$ faster. Moreover, our efficiency optimized model $(\star)$ can surpass the robust sparse matching pipeline $(\odot)$ SuperPoint (SP) + LightGlue (LG) on efficiency with considerably better accuracy. + +detecting repeatable keypoints across challenging pairs, such as extreme viewpoint changes and texture-poor regions. + +Recently, LoFTR [46] introduces a detector-free matching paradigm with transformer to directly establish semi-dense correspondences between two images without detecting keypoints. With the help of the transformer mechanism to capture the global image context and the detector-free design, LoFTR shows a strong capability of matching challenging pairs, especially in texture-poor scenarios. To reduce the computation burden, LoFTR adopts a coarse-to-fine pipeline by first performing dense matching on downsampled coarse features maps, where transformer is applied. Then, the feature locations of coarse matches on one image are fixed, while their subpixel correspondences are searched on the other image by cropping feature patches based on coarse match, performing the feature correlation, and calculating expectation over the correlation patch. + +Despite its impressive matching performance, LoFTR + +suffers from limited efficiency due to the large token size of performing transformer on the entire coarse feature map, which significantly barricades practical large-scale usages such as image retrieval [16] and SfM [43]. A large bunch of LoFTR's follow-up works [6, 15, 32, 48, 55] have attempted to improve its matching accuracy. However, there are rare methods that focus on matching efficiency of detector-free matching. QuadTree Attention [48] incorporates multi-scale transformation with a gradually narrowed attention span to avoid performing attention on large feature maps. This strategy can reduce the computation cost, but it also divides a single coarse attention process into multiple steps, leading to increased latency. + +In this paper, we revisit the design decisions of the detector-free matcher LoFTR, and propose a new matching algorithm that squeezes out redundant computations for significantly better efficiency while further improving the accuracy. As shown in Fig. 1, our approach achieves the best inference speed compared with recent image matching methods while being competitive in terms of accuracy. Our key innovations lie in introducing a token aggregation mechanism for efficient feature transformation and a two-stage correlation layer for correspondence refinement. Specifically, we find that densely performing global attention over the entire coarse feature map as in LoFTR is unnecessary, as the attention information is similar and shared in the local region. Therefore, we devise an aggregated attention mechanism to perform feature transformation on adaptively selected tokens, which is significantly compact and effectively reduces the cost of local feature transformation. + +In addition, we observe that there can be spatial variance in the matching refinement phase of LoFTR, which is caused by the expectation over the entire correlation patch when noisy feature correlation exists. To solve this issue, our approach designs a two-stage correlation layer that first locates pixel-level matches with the accurate mutual-nearest-neighbor matching on fine feature patches, and then further refines matches for subpixel-level by performing the correlation and expectation locally within tiny patches. + +Extensive experiments are conducted on multiple tasks, including homography estimation, relative pose recovery, as well as visual localization, to show the efficacy of our method. Our pipeline pushes detector-free matching to unprecedented efficiency, which is $\sim 2.5$ times faster than LoFTR and can even surpass the current state-of-the-art efficient sparse matcher LightGlue [26]. Moreover, our framework can achieve comparable or even better matching accuracy compared with competitive detector-free baselines [6, 13, 14] with considerably higher efficiency. + +In summary, this paper has the following contributions: + +- A new detector-free matching pipeline with multiple improvements based on the comprehensive revisiting of LoFTR, which is significantly more efficient and with + +better accuracy. + +- A novel aggregated attention network for efficient local feature transformation. +- A novel two-stage correlation refinement layer for accurate and subpixel-level refined correspondences. + +# 2. Related Work + +Detector-Based Image Matching. Classical image matching methods [3, 27, 37] adopt handcrafted critics for detecting keypoints, describing and then matching them. Recent methods draw benefits from deep neural networks for both detection [21, 37, 42] and description [12, 29, 49, 50], where the robustness and discriminativeness of local descriptors are significantly improved. Besides, some methods [9, 11, 28, 34, 51] managed to learn the detector and descriptor together. SuperGlue [39] is a pioneering method that first introduces the transformer mechanism into matching, which has shown notable improvements over classical handcrafted matchers. As a side effect, it also costs more time, especially with many keypoints to match. To improve the efficiency, some subsequent works, such as [5, 44], endeavor to reduce the size of the attention mechanism, albeit at the cost of sacrificing performance. LightGlue [26] introduces a new scheme for efficient sparse matching that is adaptive to the matching difficulty, where the attention process can be stopped earlier for easy pairs. It is faster than SuperGlue and can achieve competitive performance. However, robustly detecting keypoints across images is still challenging, especially for texture-poor regions. Unlike them, our method focuses on the efficiency of the detector-free method, which eliminates the restriction of keypoint detection and shows superior performance for challenging pairs. + +Detector-Free Image Matching. Detector-free methods directly match images instead of relying on a set of detected keypoints, producing semi-dense or dense matches. NC-Net [35] represents all features and possible matches as a 4D correlation volume. Sparse NC-Net [36] utilizes sparse correlation layers to ease resolution limitations. Subsequently, DRC-Net [23] improves efficiency and further improves performance in a coarse-to-fine manner. + +LoFTR [46] first employs the Transformer in detector-free matching to model the long-range dependencies. It shows remarkable matching capabilities, however, suffers from low efficiency due to the huge computation of densely transforming entire coarse feature maps. Many follow-up works further improve the matching accuracy. Matchformer [55] and AspanFormer [6] perform attention on multiscale features, where local attention regions of [6] are found with the help of estimated flow. QuadTree [48] gradually restricts the attention span during hierarchical attention to relevant areas, which can reduce overall computation. However, these designs contribute marginally or even decrease + +![](images/b0dbd617096a4e6fca43c86b88ec1450b28e3b789be6a54b6d8f574d4097a7e8.jpg) +Figure 2. Pipeline Overview. (1) Given an image pair, a CNN network extracts coarse feature maps $\tilde{\mathbf{F}}_A$ and $\tilde{\mathbf{F}}_B$ , as well as fine features. (2) Then, we transform coarse features for more discriminative feature maps by interleaving our aggregated self- and cross-attention $N$ times, where adaptively feature aggregation is performed to reduce token size before each attention for efficiency. (3) Transformed coarse features are correlated for the score matrix $S$ . Mutual-nearest-neighbor (MNN) searching is followed to establish coarse matches $\{\mathcal{M}_c\}$ . (4) To refine coarse matches, discriminative fine features $\hat{\mathbf{F}}_A^t$ , $\hat{\mathbf{F}}_B^t$ in full resolution are obtained by fusing transformed coarse features $\tilde{\mathbf{F}}_A^t$ , $\tilde{\mathbf{F}}_B^t$ with backbone features. Feature patches are then cropped centered at each coarse match $\mathcal{M}_c$ . A two-stage refinement is followed to obtain sub-pixel correspondence $\mathcal{M}_f$ . + +efficiency, since the hierarchical nature of multi-scale attention will further introduce latencies. TopicFM [15] first assigns features with similar semantic meanings to the same topic, where attention is conducted within each topic for efficiency. Since it needs to sequentially process each token's features for transformation, the efficiency improvement is limited. Moreover, performing local attention within topics can potentially restrict the capability of modeling long-range dependencies. Compared with them, the proposed aggregated attention module in our method significantly improves efficiency while achieving better accuracy. + +Dense matching methods [13, 14, 52] are designed to estimate all possible correspondences between two images, which show strong robustness. However, they are generally much slower compared with sparse and semi-dense methods. Unlike them, our method produces semi-dense matches with competitive performance and considerably better efficiency. + +Transformer has been broadly used in multiple vision tasks, including feature matching. The efficiency and memory footprint of handling large token sizes are the main limitations of transformer [54], where some methods [19, 20, 56] attempt to reduce the complexity to a linear scale to alleviate these problems. Some methods [8, 22] propose optimizing transformer models for specific hardware architectures for memory and running-time efficiency. They are orthogonal to our method and can be naturally adapted into the pipeline for further efficiency improvement. + +# 3. Method + +Given a pair of images $\mathbf{I}_A, \mathbf{I}_B$ , our objective is to establish a set of reliable correspondences between them. We achieve this by a coarse-to-fine matching pipeline, which first establishes coarse matches on downsampled feature maps and then refines them for high accuracy. An overview of our pipeline is shown in Fig. 2. + +# 3.1. Local Feature Extraction + +Image feature maps are first extracted by a lightweight backbone for later transformation and matching. Unlike LoFTR and many other detector-free matchers that use a heavy multibranch ResNet [17] network for feature extraction, we alternate to a lightweight single-branch network with reparameterization [10] to achieve better inference efficiency while preserving the model performance. + +In particular, a multi-branch CNN network with residual connections is applied during training for maximum representational power. At inference time, we losslessly convert the feature backbone into an efficient single-branch network by adopting the reparameterization technique [10], which is achieved by fusing parallel convolution kernels into a single one. Then, the intermediate $\frac{1}{8}$ down-sampled coarse features $\tilde{\mathbf{F}}_A, \tilde{\mathbf{F}}_B$ and fine features in $\frac{1}{4}$ and $\frac{1}{2}$ resolutions are extracted efficiently for later coarse-to-fine matching. + +# 3.2. Efficient Local Feature Transformation + +After the feature extraction, the coarse-level feature maps $\tilde{\mathbf{F}}_A$ and $\tilde{\mathbf{F}}_B$ are transformed by interleaving self- and crossattention $^1 n$ times to improve discriminativeness. The transformed features are denoted as $\tilde{\mathbf{F}}_A^t,\tilde{\mathbf{F}}_B^t$ + +Previous methods often perform attention on the entire coarse-level feature maps, where linear attention instead of vanilla attention is applied to ensure a manageable computation cost. However, the efficiency is still limited due to the large token size of coarse features. Moreover, the usage of linear attention leads to sub-optimal model capability. Unlike them, we propose efficient aggregated attention for both efficiency and performance. + +Preliminaries. First, we provide a brief overview of the commonly used vanilla attention and linear attention. Vanilla attention is a core mechanism in transformer encoder layer, relying on three inputs: query Q, key K, and value V. The resultant output is a weighted sum of the value, where the weighted matrix is determined by the query and its corresponding key. Formally, the attention function is defined as follows: + +$$ +\operatorname {V a n i l l a A t t e n t i o n} (Q, K, V) = \operatorname {s o f t m a x} (Q K ^ {T}) V. \tag {1} +$$ + +However, applying the vanilla attention directly to dense local features is impractical due to the significant token size. To address this issue, previous methods use linear attention to reduce the computational complexity from quadratic to linear: + +$$ +\operatorname {L i n e a r A t t e n t i o n} (Q, K, V) = \phi (Q) (\phi (K) ^ {T} \phi (V)). \tag {2} +$$ + +where $\phi (\cdot) = elu(\cdot) + 1$ . However, it comes at the cost of reduced representational power, which is also observed by [4]. + +Aggregated Attention Module. After comprehensively investigating the mechanism of the transformer on coarse feature maps, we have two observations that motivate us to devise a new efficient aggregated attention. First, the attention regions of neighboring query tokens are similar, thus we can aggregate the neighboring tokens of $f_{i}$ to prevent the redundant computation. Second, most of the attention weights of each query token are concentrated on a small number of key tokens, hence we can select the salient tokens of $f_{j}$ before attention to reduce the computation. + +Therefore, we propose to first aggregate the $f_{i}$ token utilizing a depth-wise convolution network, and $f_{j}$ is aggregated by a max pooling layer to get reduced salient tokens: + +$$ +f _ {i} ^ {\prime} = \operatorname {C o n v 2 D} \left(f _ {i}\right), f _ {j} ^ {\prime} = \operatorname {M a x P o o l} \left(f _ {j}\right) ， \tag {3} +$$ + +![](images/e1b0da0f135ead94bdcbe978d070b3778d2f98147bf05222106c9a358c772783.jpg) +Figure 3. Detailed Transformer Module Comparison. Unlike LoFTR which uses all tokens of feature maps to compute attention and resort to linear attention to reduce the computational cost, the proposed attention module first aggregates features for salient tokens, which is significantly more efficient for attention. Then the vanilla attention is utilized to transform aggregated features, where relative positional encoding is inserted to capture the spatial information. Transformed features are upsampled and fused with the original features to form the final features. + +![](images/473053f9f31671e5aa9fe3086f2e55c7e197fb322786d9cf7097f4ebb0663bdd.jpg) + +where Conv2D is implemented by a strided depthwise convolution with a kernel size of $s \times s$ , identical to that of the max-pooling layer. Then positional encoding and vanilla attention are followed to process reduced tokens. Positional encoding (PE) can help to model the spatial location contexts, where RoPE [45] is adopted in practice to account for more robust relative positions, inspired by [26]. Note that the PE layer is enabled exclusively for self-attention and skipped during cross-attention. The transformed feature map is then upsampled and fused with $f_{i}$ for the final feature map. Due to the aggregation and selection, the number of tokens in $f_{i}'$ and $f_{j}'$ is reduced by $s^{2}$ , which contributes to the efficiency of the attention phase. + +# 3.3. Coarse-level Matching Module + +We establish coarse-level matches based on the previously transformed coarse feature maps $\tilde{\mathbf{F}}_A^t$ , $\tilde{\mathbf{F}}_B^t$ . Coarse correspondences indicate rough match regions for later subpixel-level matching in the refinement phase. To achieve this, $\tilde{\mathbf{F}}_A^t$ and $\tilde{\mathbf{F}}_B^t$ are densely correlated to build a score matrix $S$ . The softmax operator on both $S$ dimensions (referred to as dual-softmax) is then applied to obtain the probability of mutual nearest matching, which is commonly used in [35, 46, 53]. The coarse correspondences $\{\mathcal{M}_c\}$ are established by selecting matches above the score threshold $\tau$ while satisfying the mutual-nearest-neighbor (MNN) constraint. + +Efficient Inference Strategy. We observe that the dual-softmax operator in the coarse matching can significantly restrict the efficiency in inference due to the large token size, especially for high-resolution images. Moreover, we find that the dual-softmax operator is crucial for training, dropping it at inference time while directly using the score matrix $S$ for MNN matching can also work well with better efficiency. + +The reason for using the dual-softmax operator in training is that it can help to train discriminative features. Intuitively, with the softmax operation, the matching score between two pixels can also conditioned on other pixels. This mechanism forces the network to improve feature similarity of true correspondences while suppressing similarity with irrelevant points. With trained discriminative features, the softmax operation can be potentially eliminated during inference. + +We denote the model skipping dual-softmax layer in inference as efficiency optimized model. Results in Tab. 1 demonstrate the effectiveness of this design. + +# 3.4. Subpixel-Level Refinement Module + +As overviewed in Fig. 2 (4), with established coarse matches $\{\mathcal{M}_c\}$ , we refine them for sub-pixel accuracy with our refinement module. It is composed of an efficient feature patch extractor for discriminative fine features, followed by a two-stage feature correlation layer for final matches $\{\mathcal{M}_f\}$ . + +Efficient Fine Feature Extraction. We first extract discriminative fine feature patches centered at each coarse match $\mathcal{M}_c$ by an efficient fusion network for later match refinement. For efficiency, our key idea here is to re-ileverage the previously transformed coarse features $\tilde{\mathbf{F}}_A^t$ , $\tilde{\mathbf{F}}_B^t$ to obtain cross-view attended discriminative fine features, instead of introducing additional feature transform networks as in LoFTR [46]. + +To be specific, $\tilde{\mathbf{F}}_A^t$ and $\tilde{\mathbf{F}}_B^t$ are adaptively fused with $1/4$ and $1/2$ resolution backbone features by convolution and upsampling to obtain fine feature maps $\hat{\mathbf{F}}_A^t$ , $\hat{\mathbf{F}}_B^t$ in the original image resolution. Then $8 \times 8$ local feature patches are cropped on fine feature maps centered at each coarse match. Since only shallow feed-forward networks are included, our fine feature fusion network is remarkably efficient. + +Two-Stage Correlation for Refinement. Based on the extracted fine local feature patches of coarse matches, we search for high-accurate sub-pixel matches. To refine a coarse match, a commonly used strategy [6, 15, 46] is to select the center-patch feature of $\mathbf{I}_A$ as a fixed reference point, and perform feature correlation and expectation on the entire corresponding feature patch for its fine match. However, this refinement-by-expectations will introduce location variance to the final match, because irrelevant regions also + +have weights and can affect results. Therefore, we propose a novel two-stage correlation module to solve this problem. + +Our idea is to utilize a mutual-nearest-neighbor (MNN) matching to get intermediate pixel-level refined matches in the first stage, and then refine them for subpixel accuracy by correlation and expectation. Motivations are that MNN matching don't have spatial variance since matches are selected by directly indexing pixels with maximum scores, but cannot achieve sub-pixel accuracy. Conversely, refinement-by-expectation can achieve sub-pixel accuracy but variance exists. The proposed two-stage refinement can draw benefits by combining the best of both worlds. + +Specifically, to refine a coarse-level correspondence $\mathcal{M}_c$ , the first-stage refinement phase densely correlates their fine feature patches to obtain the local patch score matrix $S_l$ . MNN searching is then applied on $S_l$ to get intermediate pixel-level fine matches. To limit the overall match number, we select the top-1 fine match for one coarse match by sorting the correlation scores. + +Then, we further refine these pixel-level matches for sub-pixel accuracy by our second-stage refinement. Since the matching accuracy has already significantly improved in first-stage refinement, now we can use a tiny local window for correlation and expectation with a maximum suppression of location variance. In practice, we correlate the feature of each point in $\mathbf{I}_A$ with a $3\times 3$ feature patch centered at its fine match in $\mathbf{I}_B$ . The softmax operator is then applied to get a match distribution matrix and the final refined match is obtained by calculating expectations. + +# 3.5. Supervision + +The entire pipeline is trained end-to-end by supervising the coarse and refinement matching modules separately. + +Coarse-Level Matching Supervision. The coarse ground truth matches $\{\mathcal{M}_c\}_{gt}$ with a total number of $N$ are built by warping grid-level points from $\mathbf{I}_A$ to $\mathbf{I}_B$ via depth maps and image poses following previous methods [39, 46]. The produced correlation score matrix $\mathcal{S}$ in coarse matching is supervised by minimizing the log-likelihood loss over locations of $\{\mathcal{M}_c\}_{gt}$ : + +$$ +\mathcal {L} _ {c} = - \frac {1}{N} \sum_ {(\tilde {i}, \tilde {j}) \in \left\{\mathcal {M} _ {c} \right\} _ {g t}} \log \mathcal {S} (\tilde {i}, \tilde {j}). \tag {4} +$$ + +Fine-Level Matching Supervision. We train the proposed two-stage fine-level matching module by separately supervising the two phases. The first stage fine loss $\mathcal{L}_{f1}$ is to minimize the log-likelihood loss of each fine local score matrix $S_{l}$ based on the pixel-level ground truth fine matches, similar to coarse loss. The second stage is trained by $\mathcal{L}_{f2}$ that calculates the $\ell_2$ loss between the final subpixel matches $\{\mathcal{M}_f\}$ and ground truth fine matches $\{\mathcal{M}_f\}_{gt}$ . + +
CategoryMethodMegaDepth DatasetScanNet DatasetTime (ms)
AUC@5°AUC@10°AUC@20°AUC@5°AUC@10°AUC@20°
SparseSP + NN31.746.860.17.518.632.110.8
SP + SG49.767.180.616.232.849.748.3
SP + LG49.967.080.114.830.847.531.9/30.7
Semi-DenseDRC-Net27.042.958.37.717.930.5328.0
LoFTR52.869.281.216.933.650.666.2
QuadTree54.670.582.219.037.353.5100.7
MatchFormer53.369.781.815.832.048.0128.9
TopicFM54.170.181.617.335.550.966.4
AspanFormer55.371.583.119.637.754.481.6
Ours56.472.283.519.237.053.640.1/34.4
Ours (Optimized)55.471.482.917.434.451.235.6/27.0
DenseDKM60.474.985.126.6447.0764.17210.8
ROMA62.676.786.328.950.468.3302.7
+ +Table 1. Results of Relative Pose Estimation on MegaDepth Dataset and ScanNet Dataset. We use the models trained on the MegaDepth dataset to evaluate all methods on both datasets, which can show the intra- and inter-dataset generalization abilities. The AUC of pose error at different thresholds, along with the processing time for matching image pair at a resolution of $640 \times 480$ , is presented. For SP + LG, Ours, and Ours (Optimized), the running times of the model using FP32/Mixed-Precision numerical precisions (without Efficient Attention [8, 22]) are shown. + +The total loss is the weighted sum of all supervisions: $\mathcal{L} = \mathcal{L}_c + \alpha \mathcal{L}_{f1} + \beta \mathcal{L}_{f2}$ + +# 4. Experiments + +In this section, we evaluate the performance of our method on several downstream tasks, including homography estimation, pairwise pose estimation and visual localization. Furthermore, we evaluate the effectiveness of our design by conducting detailed ablation studies. + +# 4.1. Implementation Details + +We adopt RepVGG [10] as our feature backbone, and self- and cross-attention are interleaved for $N = 4$ times to transform coarse features. For each attention, we aggregate features by a depth-wise convolution layer and a max-pooling layer, both with a kernel size of $4 \times 4$ . Our model is trained on the MegaDepth dataset [24], which is a large-scale outdoor dataset. The test scenes are separated from training data following [46]. The loss function's weights $\alpha$ and $\beta$ are set to 1.0 and 0.25, respectively. We use the AdamW optimizer with an initial learning rate of $4 \times 10^{-3}$ . The network training takes about 15 hours with a batch size of 20 on 4 NVIDIA V100 GPUs. And the coarse and fine stages are trained together from scratch. The trained model on MegaDepth is used to evaluate all datasets and tasks in our experiments to demonstrate the generalization ability. + +# 4.2. Relative Pose Estimation + +Datasets. We use the outdoor MegaDepth [24] dataset and indoor ScanNet [7] dataset for the evaluation of relative pose estimation to demonstrate the efficacy of our method. + +MegaDepth dataset is a large-scale dataset containing sparse 3D reconstructions from 196 scenes. The key challenges on this dataset are large viewpoints and illumination changes, as well as repetitive patterns. We follow the test + +split of the previous method [46] that uses 1500 sampled pairs from scenes "Sacre Coeur" and "St. Peter's Square" for evaluation. Images are resized so that the longest edge equals 1200 for all semi-dense and dense methods. Following [26], sparse methods are provided resized images with longest edge equals 1600. + +ScanNet dataset contains 1613 sequences with ground-truth depth maps and camera poses. They depict indoor scenes with viewpoint changes and texture-less regions. We use the sampled test pairs from [39] for the evaluation, where images are resized to $640 \times 480$ for all methods. + +Baselines. We compare the proposed method with three categories of methods: 1) sparse keypoint detection and matching methods, including SuperPoint [9] with Nearest-Neighbor (NN), SuperGlue (SG) [39], LightGlue (LG) [26] matchers. 2) semi-dense matchers, including DRC-Net [23], LoFTR [46], QuadTree Attention [48], MatchFormer [55], AspanFormer [6], TopicFM [15], and 3) state-of-the-art dense matcher ROMA [14] that predict matches for each pixel. + +Evaluation protocol. Following previous methods, the recovered relative poses by matches are evaluated for reflecting matching accuracy. The pose error is defined as the maximum of angular errors in rotation and translation. We report the AUC of the pose error at thresholds $(5^{\circ}, 10^{\circ}$ , and $20^{\circ})$ . Moreover, the running time of matching each image pair in the ScanNet dataset on a single NVIDIA RTX 3090 is reported for comprehensively understanding the matching accuracy and efficiency balance. We also include the running time of feature extraction and keypoint detection (SuperPoint) for the sparse method to compare with dense methods, which are end-to-end (from image input to matches output). LG [26]'s SuperPoint [9] parameters (max keypoints=2048, keypoint threshold=0.0005, nms=4) are uniformly used for all sparse methods. + +
CategoryMethodHomography est. AUC
@3px@5px@10px
SparseD2Net + NN23.235.953.6
R2D2 + NN50.663.976.8
DISK + NN52.364.978.9
SP + SG53.968.381.7
Semi-DenseSparse-NCNet48.954.267.1
DRC-Net50.656.268.3
LoFTR65.975.684.6
Ours66.576.485.5
+ +Results. As shown in Tab. 1, the proposed method achieves competitive performances compared with sparse and semidense methods on both datasets. Qualitative comparisons are shown in Fig. 4. Specifically, our method outperforms the best semi-dense baseline AspanFormer on all metrics of the MegaDepth dataset and has lower but comparable performance on the ScanNet dataset, with $\sim 2$ times faster. Our optimized model that eliminates the dual-softmax operator in coarse-level matching further brings efficiency improvements, with slight performance decreases. Using this strategy, our method can outperform the efficient and robust sparse method $\mathrm{SP} + \mathrm{LG}$ in efficiency with significantly higher accuracy. Dense matcher ROMA shows remarkable matching capability but is slow for applications in practice. Moreover, since ROMA utilizes the pre-trained DINOv2 [33] backbone, its strong generalizability on ScanNet may be attributed to the similar indoor training data in DINOv2, where other methods are trained on outdoor MegaDepth only. Compared with it, our method is $\sim 7.5$ times faster, which has a good balance between accuracy and efficiency. + +# 4.3. Homography Estimation + +Dataset. We evaluate our method on HPatches dataset [2]. HPatches dataset depicts planar scenes divided into sequences. Images are taken under different viewpoints or illumination changes. + +Baselines. We compare our method with sparse methods including D2Net [11], R2D2 [34], DISK [53] detectors with NN matcher, and SuperPoint [9] + SuperGlue [39]. As for semi-dense methods, we compare with Sparse-NCNet [36], DRC-Net [23], and LoFTR [46]. For SuperGlue and all semidense methods, we use their models trained on MegaDepth dataset for evaluation. + +Evaluation Protocol. Following SuperGlue [39] and LoFTR [46], we resize all images for matching so that their smallest edge equals 480 pixels. We collect the mean reprojection error of corner points, and report the area under the cumulative curve (AUC) under 3 different thresholds, including 3 px, 5 px, and 10 px. For all baselines, we employ the same RANSAC method as a robust homography estimator + +Table 2. Results of Homography Estimation on HPatches Dataset. Our method is compared with sparse and semi-dense methods. The AUC of reprojection error of corner points at different thresholds is reported. + +
MethodDUC1DUC2
(0.25m,2°)/(0.5m,5°)/(1.0m,10°)
SP+SG49.0 / 68.7 / 80.853.4 / 77.1 / 82.4
LoFTR47.5 / 72.2 / 84.854.2 / 74.8 / 85.5
TopicFM52.0 / 74.7 / 87.453.4 / 74.8 / 83.2
PATS55.6 / 71.2 / 81.058.8 / 80.9 / 85.5
AspanFormer51.5 / 73.7 / 86.055.0 / 74.0 / 81.7
Ours52.0 / 74.7 / 86.958.0 / 80.9 / 89.3
+ +Table 3. Results of Visual Localization on InLoc Dataset. + +
MethodDayNight
(0.25m,2°)/(0.5m,5°)/(1.0m,10°)
SP+SG89.8 / 96.1 / 99.477.0 / 90.6 / 100.0
LoFTR88.7 / 95.6 / 99.078.5 / 90.6 / 99.0
TopicFM90.2 / 95.9 / 98.977.5 / 91.1 / 99.5
PATS89.6 / 95.8 / 99.373.8 / 92.1 / 99.5
AspanFormer89.4 / 95.6 / 99.077.5 / 91.6 / 99.5
Ours89.6 / 96.2 / 99.077.0 / 91.1 / 99.5
+ +Table 4. Results of Visual Localization on Aachen v1.1 Dataset. + +for a fair comparison. Following LoFTR, we select only the top 1000 predicted matches of semi-dense methods for the sake of fairness. + +Results. As shown in Tab. 2, even though the number of matches is restricted, our method can also work remarkably well and outperform sparse methods significantly. Compared with semi-dense, our method can surpass them with significantly higher efficiency. We attribute this to the effectiveness of two-stage refinement for accuracy improvement and proposed aggregation module for efficiency. + +# 4.4. Visual Localization + +Datasets and Evaluation Protocols. Visual localization is an important downstream task of image matching, which aims to estimate the 6-DoF poses of query images based on the 3D scene model. We conduct experiments on two commonly used benchmarks, including InLoc [47] dataset and Aachen v1.1 [41] dataset, for evaluation to demonstrate the superiority of our method. InLoc dataset is captured on indoor scenes with plenty of repetitive structures and textureless regions, where each database image has a corresponding depth map. Aachen v1.1 is a challenging large-scale outdoor dataset for localization with large-viewpoint and day-and-night illumination changes, which particularly relies on the robustness of matching methods. We adopt its full localization track for benchmarking. + +Following [6, 46], the open-sourced localization framework HLoc [38] is utilized. For both datasets, the percentage of pose errors satisfying both angular and distance thresholds is reported following the benchmarks, where different thresholds are used. For the InLoc dataset, the metrics of two test scenes including DUC1 and DUC2 are separately reported. As for the Aachen v1.1 dataset, the metrics corresponding to the daytime and nighttime divisions are reported. + +Baselines. We compare the proposed method with both detector-based method SuperPoint [9] + SuperGlue [39] and + +![](images/c381cf5d3a60fdbc0fdc9bca1154fc524231d28359a4512b136d2d04f96390e1.jpg) + +![](images/e592b78f9f7fe6f413684e84894725f215261dc9270936e8fed1412663c13b56.jpg) +Figure 4. Qualitative Results. Our method is compared with the sparse matching pipeline SuperPoint [9] $^+$ LightGlue [26], semi-dense +macher AspanFormer [6]. Image pairs with texture-poor regions and large-viewpoint changes can be robustly matched by our method. The +red color indicates epipolar error beyond $5\times 10^{-4}$ (in the normalized image coordinates). + +![](images/9b21e807ba425ad2643d05931505c51b460341d8f5a1f3829da0f4372e2bc8ec.jpg) + +
MethodPose Estimation AUCTime (ms)
@5°@10°@20°
Ours Full56.472.283.5139.2
1) Ours Optimal (w/o dual-softmax)55.471.482.9102.0
2) Replace Agg. Attention to LoFTR's Trans.54.770.582.2171.4
3) Replace two-stage refine. to LoFTR's refine.54.770.982.7135.3
4) No second-stage refinement55.871.883.3138.1
5) Replace RepVGG with ResNet55.471.482.9156.2
+ +Table 5. Ablation Studies. The components of our method are ablated on the MegaDepth dataset for a comprehensive understanding of our method, where averaged running times for an image pair with high-resolution $1200 \times 1200$ are reported. + +detector-free methods including LoFTR [46], TopicFM [15], PATS [32] and Aspanformer [6]. + +Results. We adhere to the pipeline and evaluation settings of the online visual localization benchmark² to ensure fairness. As presented in Tab. 3, the proposed method achieves competitive results, taking both detector-based and detector-free methods into account. Being a method primarily geared towards efficiency, our approach can deliver results comparable to those of many accuracy-oriented methods. As depicted in Tab. 4, our method also demonstrates performance on par with the best-performing approaches. + +# 4.5. Ablation Studies + +In this part, we conduct detailed ablation studies to analyze the effectiveness of our proposed modules with results shown in Tab. 5. 1) Without dual-softmax, our optimal model can bring huge efficiency improvement in high-resolution images. 2) In coarse feature transformation, replacing the proposed aggregated attention module with LoFTR's transformer can bring significant efficiency dropping, as well as accuracy decrease. Note that the replaced transformer is also equipped with RoPE same as ours for fair comparison. This demonstrates the efficacy of the proposed module that performing vanilla attention on aggregated features can achieve higher efficiency with even better matching accuracy. 3) Compared with using LoFTR's refinement that performs + +expectation on the entire correlation patch, the proposed two-stage refinement layer can bring accuracy improvement with negligible latency. We attribute this to the two-stage refinement's property that can maximize the suppression of location variance in correlation refinement. 4) Dropping the second refinement stage will lead to degraded pose accuracy with minor efficiency changes, especially on the strict AUC@5° metric. 5) Changing the backbone from reparameterized VGG [10] back to multi-branch ResNet [17] leads to decreased efficiency with similar accuracy, which demonstrates the effectiveness of our design choice for efficiency. + +# 5. Conclusions + +This paper introduces a new semi-dense local feature matcher based on the success of LoFTR. We revisit its designs and propose several improvements for both efficiency and matching accuracy. A key observation is that performing the Transformer on the entire coarse feature map is redundant due to the similar local information, where an aggregated attention module is proposed to perform transformer on reduced tokens with significantly better efficiency and competitive performance. Moreover, a two-stage correlation layer is devised to solve the location variance problem in LoFTR's refinement design, which further brings accuracy improvements. As a result, our method can achieve $\sim 2.5$ times faster compared with LoFTR with better matching accuracy. Moreover, as a semi-dense matching method, the proposed method can achieve comparable efficiency with the recent robust sparse feature matcher LightGlue [26]. We believe this opens up the applications of our method in large-scale or latency-sensitive downstream tasks, such as image retrieval and 3D reconstruction. Please refer to the supplementary material for discussions about limitations and future works. + +Acknowledgement The authors would like to acknowledge support from NSFC (No. 62172364) and Information Technology Center and State Key Lab of CAD&CG, Zhejiang University. + +# References + +[1] Sameer Agarwal, Yasutaka Furukawa, Noah Snavely, Ian Simon, Brian Curless, Steven M. Seitz, and Richard Szeliski. Building rome in a day. ICCV, 2009. 1 +[2] Vassileios Balntas, Karel Lenc, Andrea Vedaldi, and Krystian Mikolajczyk. Hpatches: A benchmark and evaluation of handcrafted and learned local descriptors. In CVPR, pages 5173-5182, 2017. 7 +[3] Herbert Bay, Andreas Ess, Tinne Tuytelaars, and Luc Van Gool. Speeded-up robust features (surf). CVIU, 110 (3):346-359, 2008. 2 +[4] Han Cai, Chuang Gan, and Song Han. Efficientvit: Enhanced linear attention for high-resolution low-computation visual recognition. arXiv preprint arXiv:2205.14756, 2022. 4 +[5] Hongkai Chen, Zixin Luo, Jiahui Zhang, Lei Zhou, Xuyang Bai, Zeyu Hu, Chiew-Lan Tai, and Long Quan. Learning to match features with seeded graph matching network. In ICCV, pages 6301-6310, 2021. 2 +[6] Hongkai Chen, Zixin Luo, Lei Zhou, Yurun Tian, Mingmin Zhen, Tian Fang, David McKinnon, Yanghai Tsin, and Long Quan. Aspanformer: Detector-free image matching with adaptive span transformer. In ECCV, pages 20-36. Springer, 2022. 2, 5, 6, 7, 8 +[7] Angela Dai, Angel X Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In CVPR, pages 5828-5839, 2017. 6 +[8] Tri Dao, Daniel Y. Fu, Stefano Ermon, Atri Rudra, and Christopher Ré. FlashAttention: Fast and memory-efficient exact attention with IO-awareness. In NeurIPS, 2022. 3, 6 +[9] Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. Superpoint: Self-supervised interest point detection and description. CVPRW, 2018. 2, 6, 7, 8 +[10] Xiaohan Ding, Xiangyu Zhang, Ningning Ma, Jungong Han, Guiguang Ding, and Jian Sun. Repvgg: Making vgg-style convnets great again. In CVPR, pages 13733-13742, 2021. 3, 6, 8 +[11] Mihai Dusmanu, Ignacio Rocco, Tomas Pajdla, Marc Pollefeys, Josef Sivic, Akihiko Torii, and Torsten Sattler. D2-net: A trainable cnn for joint description and detection of local features. In CVPR, pages 8092-8101, 2019. 2, 7 +[12] Patrick Ebel, Anastasiia Mishchuk, Kwang Moo Yi, Pascal Fua, and Eduard Trulls. Beyond cartesian representations for local descriptors. In ICCV, pages 253-262, 2019. 2 +[13] Johan Edstedt, Ioannis Athanasiadis, Mårten Wadenbäck, and Michael Felsberg. Dkm: Dense kernelized feature matching for geometry estimation. In CVPR, pages 17765-17775, 2023. 2, 3 +[14] Johan Edstedt, Qiyu Sun, Georg Bokman, Mårten Wadenback, and Michael Felsberg. Roma: Revisiting robust losses for dense feature matching. arXiv preprint arXiv:2305.15404, 2023. 2, 3, 6 +[15] Khang Truong Giang, Soohwan Song, and Sung-Guk Jo. Top-icfm: Robust and interpretable topic-assisted feature matching. In AAAI, 2023. 2, 3, 5, 6, 8 +[16] Stephen Hausler, Sourav Garg, Ming Xu, Michael Milford, and Tobias Fischer. Patch-netvlad: Multi-scale fusion of + +locally-global descriptors for place recognition. CVPR, pages 14136-14147, 2021. 2 +[17] Kaiming He, X. Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. CVPR, pages 770-778, 2015. 3, 8 +[18] Xingyi He, Jiaming Sun, Yifan Wang, Sida Peng, Qixing Huang, Hujun Bao, and Xiaowei Zhou. Detector-free structure from motion. In CVPR, 2024. 1 +[19] Angelos Katharopoulos, Apoorv Vyas, Nikolaos Pappas, and Francois Fleuret. Transformers are rnns: Fast autoregressive transformers with linear attention. In ICML, 2020. 3 +[20] Nikita Kitaev, Lukasz Kaiser, and Anselm Levskaya. Reformer: The efficient transformer. ArXiv, abs/2001.04451, 2020.3 +[21] Axel Barroso Laguna, Edgar Riba, Daniel Ponsa, and Krystian Mikolajczyk. Key.net: Keypoint detection by handcrafted and learned cnn filters. ICCV, pages 5835-5843, 2019. 2 +[22] Benjamin Lefaudeau, Francisco Massa, Diana Liskovich, Wenhan Xiong, Vittorio Caggiano, Sean Naren, Min Xu, Jieru Hu, Marta Tintore, Susan Zhang, Patrick Labatut, and Daniel Haziza. xformers: A modular and hackable transformer modelling library. https://github.com/facebookresearch/xformers, 2022.3.6 +[23] Xinghui Li, Kai Han, Shuda Li, and Victor Prisacariu. Dualresolution correspondence networks. In NeurlPS, 2020. 2, 6, 7 +[24] Zhengqi Li and Noah Snavely. Megadepth: Learning single-view depth prediction from internet photos. In CVPR, pages 2041-2050, 2018. 6 +[25] Philipp Lindenberger, Paul-Edouard Sarlin, Viktor Larsson, and Marc Pollefeys. Pixel-perfect structure-from-motion with featuremetric refinement. ICCV, pages 5967-5977, 2021. 1 +[26] Philipp Lindenberger, Paul-Edouard Sarlin, and Marc Pollefeys. LightGlue: Local Feature Matching at Light Speed. In ICCV, 2023. 1, 2, 4, 6, 8 +[27] G LoweDavid. Distinctive image features from scale-invariant keypoints. IJCV, 2004. 1, 2 +[28] Zixin Luo, Lei Zhou, Xuyang Bai, Hongkai Chen, Jiahui Zhang, Yao Yao, Shiwei Li, Tian Fang, and Long Quan. Aslfeat: Learning local features of accurate shape and localization. In CVPR, pages 6589-6598, 2020. 2 +[29] Anastasiia Mishchuk, Dmytro Mishkin, Filip Radenovic, and Jiri Matas. Working hard to know your neighbor's margins: Local descriptor learning loss. NeurIPS, 30, 2017. 2 +[30] Raul Mur-Artal and Juan D. Tardós. Orb-slam2: An open-source slam system for monocular, stereo, and rgb-d cameras. TR, 33:1255-1262, 2016. 1 +[31] Raul Mur-Artal, José M. M. Montiel, and Juan D. Tardós. Orb-slam: A versatile and accurate monocular slam system. TR, 31:1147-1163, 2015. 1 +[32] Junjie Ni, Yijin Li, Zhaoyang Huang, Hongsheng Li, Hujun Bao, Zhaopeng Cui, and Guofeng Zhang. Pats: Patch area transportation with subdivision for local feature matching. CVPR, pages 17776-17786, 2023. 2, 8 +[33] Maxime Oquab, Timothée Darcet, Theo Moutakanni, Huy V. Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, Russell + +Howes, Po-Yao Huang, Hu Xu, Vasu Sharma, Shang-Wen Li, Wojciech Galuba, Mike Rabbat, Mido Assran, Nicolas Ballas, Gabriel Synnaeve, Ishan Misra, Herve Jegou, Julien Mairal, Patrick Labatut, Armand Joulin, and Piotr Bojanowski. Dinov2: Learning robust visual features without supervision, 2023. 7 +[34] Jérôme Revaud, César Roberto de Souza, M. Humenberger, and Philippe Weinzaepfel. R2d2: Reliable and repeatable detector and descriptor. In NeurIPS, 2019. 2, 7 +[35] Ignacio Rocco, Mircea Cimpoi, Relja Arandjelović, Akihiko Torii, Tomas Pajdla, and Josef Sivic. Neighbourhood consensus networks. *NeurIPS*, 31, 2018. 2, 4 +[36] Ignacio Rocco, Relja Arandjelović, and Josef Sivic. Efficient neighbourhood consensus networks via submanifold sparse convolutions. In ECCV, pages 605-621. Springer, 2020. 2, 7 +[37] Edward Rosten and Tom Drummond. Machine learning for high-speed corner detection. In ECCV, pages 430-443. Springer, 2006. 1, 2 +[38] Paul-Edouard Sarlin, César Cadena, Roland Y. Siegwart, and Marcin Dymczyk. From coarse to fine: Robust hierarchical localization at large scale. CVPR, pages 12708–12717, 2019. 1, 7 +[39] Paul-Edouard Sarlin, Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. SuperGlue: Learning feature matching with graph neural networks. In CVPR, 2020. 1, 2, 4, 5, 6, 7 +[40] Paul-Edouard Sarlin, Ajaykumar Unagar, Måns Larsson, Hugo Germain, Carl Toft, Victor Larsson, Marc Pollefeys, Vincent Lepetit, Lars Hammarstrand, Fredrik Kahl, and Torsten Sattler. Back to the Feature: Learning Robust Camera Localization from Pixels to Pose. In CVPR, 2021. 1 +[41] Torsten Sattler, Will Maddern, Carl Toft, Akihiko Torii, Lars Hammarstrand, Erik Stenberg, Daniel Safari, Masatoshi Okutomi, Marc Pollefeys, Josef Sivic, et al. Benchmarking 6dof outdoor visual localization in changing conditions. In CVPR, pages 8601-8610, 2018. 7 +[42] Nikolay Savinov, Akihito Seki, Lubor Ladicky, Torsten Sattler, and Marc Pollefeys. Quad-networks: unsupervised learning to rank for interest point detection. In CVPR, pages 1822-1830, 2017. 2 +[43] Johannes L. Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In CVPR, 2016. 1, 2 +[44] Yan Shi, Jun-Xiong Cai, Yoli Shavit, Tai-Jiang Mu, Wensen Feng, and Kai Zhang. Clustergnn: Cluster-based coarse-to-fine graph neural network for efficient feature matching. In CVPR, pages 12517-12526, 2022. 2 +[45] Jianlin Su, Yu Lu, Shengfeng Pan, Bo Wen, and Yunfeng Liu. Roformer: Enhanced transformer with rotary position embedding. ArXiv, abs/2104.09864, 2021. 4 +[46] Jiaming Sun, Zehong Shen, Yuang Wang, Hujun Bao, and Xiaowei Zhou. Loftr: Detector-free local feature matching with transformers. In CVPR, pages 8922-8931, 2021. 1, 2, 4, 5, 6, 7, 8 +[47] Hajime Taira, Masatoshi Okutomi, Torsten Sattler, Mircea Cimpoi, Marc Pollefeys, Josef Sivic, Tomas Pajdla, and Akihiko Torii. Inloc: Indoor visual localization with dense matching and view synthesis. In CVPR, pages 7199-7209, 2018. 7 + +[48] Shitao Tang, Jiahui Zhang, Siyu Zhu, and Ping Tan. Quadtree attention for vision transformers. *ICLR*, 2022. 2, 6 +[49] Yurun Tian, Bin Fan, and Fuchao Wu. L2-net: Deep learning of discriminative patch descriptor in euclidean space. In CVPR, pages 661-669, 2017. 1, 2 +[50] Yurun Tian, Xin Yu, Bin Fan, Fuchao Wu, Huub Heijnen, and Vassileios Balntas. Sosnet: Second order similarity regularization for local descriptor learning. In CVPR, pages 11016-11025, 2019. 2 +[51] Yurun Tian, Vassileios Balntas, Tony Ng, Axel Barroso-Laguna, Yiannis Demiris, and Krystian Mikolajczyk. D2d: Keypoint extraction with describe to detect approach. In ACCV, 2020. 2 +[52] Prune Truong, Martin Danelljan, Luc Van Gool, and Radu Timofte. Learning accurate dense correspondences and when to trust them. In CVPR, pages 5714-5724, 2021. 3 +[53] Michal Tyszkiiewicz, Pascal Fua, and Eduard Trulls. Disk: Learning local features with policy gradient. NeurIPS, 2020. 4, 7 +[54] Ashish Vaswani, Noam M. Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In NeurIPS, 2017. 3 +[55] Qing Wang, Jiaming Zhang, Kailun Yang, Kunyu Peng, and Rainer Stiefelhagen. Matchformer: Interleaving attention in transformers for feature matching. In ACCV, 2022. 2, 6 +[56] Sinong Wang, Belinda Z. Li, Madian Khabsa, Han Fang, and Hao Ma. Linformer: Self-attention with linear complexity. ArXiv, abs/2006.04768, 2020. 3 \ No newline at end of file diff --git a/2024/Efficient LoFTR_ Semi-Dense Local Feature Matching with Sparse-Like Speed/images.zip b/2024/Efficient LoFTR_ Semi-Dense Local Feature Matching with Sparse-Like Speed/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..bcacd5f9e5568d1cb13a7b913a15bd1003e2c574 --- /dev/null +++ b/2024/Efficient LoFTR_ Semi-Dense Local Feature Matching with Sparse-Like Speed/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b6f038c0fc49c3dd3074b8d86b5378cae6a72e386111d69faad599b407108003 +size 472947 diff --git a/2024/Efficient LoFTR_ Semi-Dense Local Feature Matching with Sparse-Like Speed/layout.json b/2024/Efficient LoFTR_ Semi-Dense Local Feature Matching with Sparse-Like Speed/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..5521d43442e5efa200286246013a4112677799b2 --- /dev/null +++ b/2024/Efficient LoFTR_ Semi-Dense Local Feature Matching with Sparse-Like Speed/layout.json @@ -0,0 +1,8413 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 54, + 103, + 539, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 103, + 539, + 121 + ], + "spans": [ + { + "bbox": [ + 54, + 103, + 539, + 121 + ], + "type": "text", + "content": "Efficient LoFTR: Semi-Dense Local Feature Matching with Sparse-Like Speed" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 123, + 137, + 478, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 123, + 137, + 478, + 152 + ], + "spans": [ + { + "bbox": [ + 123, + 137, + 478, + 152 + ], + "type": "text", + "content": "Yifan Wang* Xingyi He* Sida Peng† Dongli Tan Xiaowei Zhou" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 239, + 163, + 339, + 178 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 239, + 163, + 339, + 178 + ], + "spans": [ + { + "bbox": [ + 239, + 163, + 339, + 178 + ], + "type": "text", + "content": "Zhejiang University" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 149, + 200, + 197, + 213 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 149, + 200, + 197, + 213 + ], + "spans": [ + { + "bbox": [ + 149, + 200, + 197, + 213 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 221, + 290, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 221, + 290, + 509 + ], + "spans": [ + { + "bbox": [ + 47, + 221, + 290, + 509 + ], + "type": "text", + "content": "We present a novel method for efficiently producing semidense matches across images. Previous detector-free matcher LoFTR has shown remarkable matching capability in handling large-viewpoint change and texture-poor scenarios but suffers from low efficiency. We revisit its design choices and derive multiple improvements for both efficiency and accuracy. One key observation is that performing the transformer over the entire feature map is redundant due to shared local information, therefore we propose an aggregated attention mechanism with adaptive token selection for efficiency. Furthermore, we find spatial variance exists in LoFTR's fine correlation module, which is adverse to matching accuracy. A novel two-stage correlation layer is proposed to achieve accurate subpixel correspondences for accuracy improvement. Our efficiency optimized model is " + }, + { + "bbox": [ + 47, + 221, + 290, + 509 + ], + "type": "inline_equation", + "content": "\\sim 2.5\\times" + }, + { + "bbox": [ + 47, + 221, + 290, + 509 + ], + "type": "text", + "content": " faster than LoFTR which can even surpass state-of-the-art efficient sparse matching pipeline SuperPoint + LightGlue. Moreover, extensive experiments show that our method can achieve higher accuracy compared with competitive semi-dense matchers, with considerable efficiency benefits. This opens up exciting prospects for large-scale or latency-sensitive applications such as image retrieval and 3D reconstruction. Project page: https://zju3dv.github.io/efficientloftr/." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 529, + 128, + 542 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 529, + 128, + 542 + ], + "spans": [ + { + "bbox": [ + 47, + 529, + 128, + 542 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 550, + 288, + 682 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 550, + 288, + 682 + ], + "spans": [ + { + "bbox": [ + 46, + 550, + 288, + 682 + ], + "type": "text", + "content": "Image matching is the cornerstone of many 3D computer vision tasks, which aim to find a set of highly accurate correspondences given an image pair. The established matches between images have broad usages such as reconstructing the 3D world by structure from motion (SfM) [1, 18, 25, 43] or SLAM system [30, 31], and visual localization [38, 40], etc. Previous methods typically follow a two-stage pipeline: they first detect [37] and describe [49] a set of keypoints on each image, and then establish keypoint correspondences by handcrafted [27] or learning-based matchers [26, 39]. These detector-based methods are efficient but suffer from robustly" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 310, + 200, + 545, + 408 + ], + "blocks": [ + { + "bbox": [ + 310, + 200, + 545, + 408 + ], + "lines": [ + { + "bbox": [ + 310, + 200, + 545, + 408 + ], + "spans": [ + { + "bbox": [ + 310, + 200, + 545, + 408 + ], + "type": "image", + "image_path": "dcc10e8a010fd52aed100c9bace3e941ad983eee32cb58fdce6c1aee64e93e25.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 411, + 547, + 489 + ], + "lines": [ + { + "bbox": [ + 304, + 411, + 547, + 489 + ], + "spans": [ + { + "bbox": [ + 304, + 411, + 547, + 489 + ], + "type": "text", + "content": "Figure 1. Matching Accuracy and Efficiency Comparisons. Our method achieves competitive accuracy compared with semi-dense matchers " + }, + { + "bbox": [ + 304, + 411, + 547, + 489 + ], + "type": "inline_equation", + "content": "(\\odot)" + }, + { + "bbox": [ + 304, + 411, + 547, + 489 + ], + "type": "text", + "content": " at a significantly higher speed. Compared with dense matcher ROMA " + }, + { + "bbox": [ + 304, + 411, + 547, + 489 + ], + "type": "inline_equation", + "content": "(\\odot)" + }, + { + "bbox": [ + 304, + 411, + 547, + 489 + ], + "type": "text", + "content": ", our method is " + }, + { + "bbox": [ + 304, + 411, + 547, + 489 + ], + "type": "inline_equation", + "content": "\\sim 7.5\\times" + }, + { + "bbox": [ + 304, + 411, + 547, + 489 + ], + "type": "text", + "content": " faster. Moreover, our efficiency optimized model " + }, + { + "bbox": [ + 304, + 411, + 547, + 489 + ], + "type": "inline_equation", + "content": "(\\star)" + }, + { + "bbox": [ + 304, + 411, + 547, + 489 + ], + "type": "text", + "content": " can surpass the robust sparse matching pipeline " + }, + { + "bbox": [ + 304, + 411, + 547, + 489 + ], + "type": "inline_equation", + "content": "(\\odot)" + }, + { + "bbox": [ + 304, + 411, + 547, + 489 + ], + "type": "text", + "content": " SuperPoint (SP) + LightGlue (LG) on efficiency with considerably better accuracy." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 496, + 546, + 520 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 496, + 546, + 520 + ], + "spans": [ + { + "bbox": [ + 304, + 496, + 546, + 520 + ], + "type": "text", + "content": "detecting repeatable keypoints across challenging pairs, such as extreme viewpoint changes and texture-poor regions." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 521, + 547, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 521, + 547, + 700 + ], + "spans": [ + { + "bbox": [ + 304, + 521, + 547, + 700 + ], + "type": "text", + "content": "Recently, LoFTR [46] introduces a detector-free matching paradigm with transformer to directly establish semi-dense correspondences between two images without detecting keypoints. With the help of the transformer mechanism to capture the global image context and the detector-free design, LoFTR shows a strong capability of matching challenging pairs, especially in texture-poor scenarios. To reduce the computation burden, LoFTR adopts a coarse-to-fine pipeline by first performing dense matching on downsampled coarse features maps, where transformer is applied. Then, the feature locations of coarse matches on one image are fixed, while their subpixel correspondences are searched on the other image by cropping feature patches based on coarse match, performing the feature correlation, and calculating expectation over the correlation patch." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 317, + 701, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 701, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 317, + 701, + 545, + 714 + ], + "type": "text", + "content": "Despite its impressive matching performance, LoFTR" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 693, + 287, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 693, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 693, + 287, + 713 + ], + "type": "text", + "content": "*Equal contribution. † Corresponding author: Sida Peng. The authors from Zhejiang University are affiliated with the State Key Lab of CAD&CG." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "21666" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 228 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 228 + ], + "type": "text", + "content": "suffers from limited efficiency due to the large token size of performing transformer on the entire coarse feature map, which significantly barricades practical large-scale usages such as image retrieval [16] and SfM [43]. A large bunch of LoFTR's follow-up works [6, 15, 32, 48, 55] have attempted to improve its matching accuracy. However, there are rare methods that focus on matching efficiency of detector-free matching. QuadTree Attention [48] incorporates multi-scale transformation with a gradually narrowed attention span to avoid performing attention on large feature maps. This strategy can reduce the computation cost, but it also divides a single coarse attention process into multiple steps, leading to increased latency." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 228, + 289, + 432 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 228, + 289, + 432 + ], + "spans": [ + { + "bbox": [ + 46, + 228, + 289, + 432 + ], + "type": "text", + "content": "In this paper, we revisit the design decisions of the detector-free matcher LoFTR, and propose a new matching algorithm that squeezes out redundant computations for significantly better efficiency while further improving the accuracy. As shown in Fig. 1, our approach achieves the best inference speed compared with recent image matching methods while being competitive in terms of accuracy. Our key innovations lie in introducing a token aggregation mechanism for efficient feature transformation and a two-stage correlation layer for correspondence refinement. Specifically, we find that densely performing global attention over the entire coarse feature map as in LoFTR is unnecessary, as the attention information is similar and shared in the local region. Therefore, we devise an aggregated attention mechanism to perform feature transformation on adaptively selected tokens, which is significantly compact and effectively reduces the cost of local feature transformation." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 434, + 289, + 541 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 434, + 289, + 541 + ], + "spans": [ + { + "bbox": [ + 46, + 434, + 289, + 541 + ], + "type": "text", + "content": "In addition, we observe that there can be spatial variance in the matching refinement phase of LoFTR, which is caused by the expectation over the entire correlation patch when noisy feature correlation exists. To solve this issue, our approach designs a two-stage correlation layer that first locates pixel-level matches with the accurate mutual-nearest-neighbor matching on fine feature patches, and then further refines matches for subpixel-level by performing the correlation and expectation locally within tiny patches." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 543, + 289, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 543, + 289, + 662 + ], + "spans": [ + { + "bbox": [ + 46, + 543, + 289, + 662 + ], + "type": "text", + "content": "Extensive experiments are conducted on multiple tasks, including homography estimation, relative pose recovery, as well as visual localization, to show the efficacy of our method. Our pipeline pushes detector-free matching to unprecedented efficiency, which is " + }, + { + "bbox": [ + 46, + 543, + 289, + 662 + ], + "type": "inline_equation", + "content": "\\sim 2.5" + }, + { + "bbox": [ + 46, + 543, + 289, + 662 + ], + "type": "text", + "content": " times faster than LoFTR and can even surpass the current state-of-the-art efficient sparse matcher LightGlue [26]. Moreover, our framework can achieve comparable or even better matching accuracy compared with competitive detector-free baselines [6, 13, 14] with considerably higher efficiency." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 58, + 664, + 282, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 664, + 282, + 676 + ], + "spans": [ + { + "bbox": [ + 58, + 664, + 282, + 676 + ], + "type": "text", + "content": "In summary, this paper has the following contributions:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 677, + 289, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 677, + 289, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 677, + 289, + 713 + ], + "type": "text", + "content": "- A new detector-free matching pipeline with multiple improvements based on the comprehensive revisiting of LoFTR, which is significantly more efficient and with" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 73, + 380, + 84 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 73, + 380, + 84 + ], + "spans": [ + { + "bbox": [ + 313, + 73, + 380, + 84 + ], + "type": "text", + "content": "better accuracy." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 84, + 545, + 132 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 306, + 84, + 545, + 106 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 84, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 306, + 84, + 545, + 106 + ], + "type": "text", + "content": "- A novel aggregated attention network for efficient local feature transformation." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 108, + 545, + 132 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 108, + 545, + 132 + ], + "spans": [ + { + "bbox": [ + 306, + 108, + 545, + 132 + ], + "type": "text", + "content": "- A novel two-stage correlation refinement layer for accurate and subpixel-level refined correspondences." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 147, + 392, + 159 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 147, + 392, + 159 + ], + "spans": [ + { + "bbox": [ + 306, + 147, + 392, + 159 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 168, + 547, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 168, + 547, + 468 + ], + "spans": [ + { + "bbox": [ + 304, + 168, + 547, + 468 + ], + "type": "text", + "content": "Detector-Based Image Matching. Classical image matching methods [3, 27, 37] adopt handcrafted critics for detecting keypoints, describing and then matching them. Recent methods draw benefits from deep neural networks for both detection [21, 37, 42] and description [12, 29, 49, 50], where the robustness and discriminativeness of local descriptors are significantly improved. Besides, some methods [9, 11, 28, 34, 51] managed to learn the detector and descriptor together. SuperGlue [39] is a pioneering method that first introduces the transformer mechanism into matching, which has shown notable improvements over classical handcrafted matchers. As a side effect, it also costs more time, especially with many keypoints to match. To improve the efficiency, some subsequent works, such as [5, 44], endeavor to reduce the size of the attention mechanism, albeit at the cost of sacrificing performance. LightGlue [26] introduces a new scheme for efficient sparse matching that is adaptive to the matching difficulty, where the attention process can be stopped earlier for easy pairs. It is faster than SuperGlue and can achieve competitive performance. However, robustly detecting keypoints across images is still challenging, especially for texture-poor regions. Unlike them, our method focuses on the efficiency of the detector-free method, which eliminates the restriction of keypoint detection and shows superior performance for challenging pairs." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 472, + 547, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 472, + 547, + 567 + ], + "spans": [ + { + "bbox": [ + 304, + 472, + 547, + 567 + ], + "type": "text", + "content": "Detector-Free Image Matching. Detector-free methods directly match images instead of relying on a set of detected keypoints, producing semi-dense or dense matches. NC-Net [35] represents all features and possible matches as a 4D correlation volume. Sparse NC-Net [36] utilizes sparse correlation layers to ease resolution limitations. Subsequently, DRC-Net [23] improves efficiency and further improves performance in a coarse-to-fine manner." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 570, + 548, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 548, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 548, + 713 + ], + "type": "text", + "content": "LoFTR [46] first employs the Transformer in detector-free matching to model the long-range dependencies. It shows remarkable matching capabilities, however, suffers from low efficiency due to the huge computation of densely transforming entire coarse feature maps. Many follow-up works further improve the matching accuracy. Matchformer [55] and AspanFormer [6] perform attention on multiscale features, where local attention regions of [6] are found with the help of estimated flow. QuadTree [48] gradually restricts the attention span during hierarchical attention to relevant areas, which can reduce overall computation. However, these designs contribute marginally or even decrease" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "21667" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 47, + 65, + 547, + 283 + ], + "blocks": [ + { + "bbox": [ + 47, + 65, + 547, + 283 + ], + "lines": [ + { + "bbox": [ + 47, + 65, + 547, + 283 + ], + "spans": [ + { + "bbox": [ + 47, + 65, + 547, + 283 + ], + "type": "image", + "image_path": "b0dbd617096a4e6fca43c86b88ec1450b28e3b789be6a54b6d8f574d4097a7e8.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 295, + 548, + 374 + ], + "lines": [ + { + "bbox": [ + 46, + 295, + 548, + 374 + ], + "spans": [ + { + "bbox": [ + 46, + 295, + 548, + 374 + ], + "type": "text", + "content": "Figure 2. Pipeline Overview. (1) Given an image pair, a CNN network extracts coarse feature maps " + }, + { + "bbox": [ + 46, + 295, + 548, + 374 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{F}}_A" + }, + { + "bbox": [ + 46, + 295, + 548, + 374 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 295, + 548, + 374 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{F}}_B" + }, + { + "bbox": [ + 46, + 295, + 548, + 374 + ], + "type": "text", + "content": ", as well as fine features. (2) Then, we transform coarse features for more discriminative feature maps by interleaving our aggregated self- and cross-attention " + }, + { + "bbox": [ + 46, + 295, + 548, + 374 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 295, + 548, + 374 + ], + "type": "text", + "content": " times, where adaptively feature aggregation is performed to reduce token size before each attention for efficiency. (3) Transformed coarse features are correlated for the score matrix " + }, + { + "bbox": [ + 46, + 295, + 548, + 374 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 46, + 295, + 548, + 374 + ], + "type": "text", + "content": ". Mutual-nearest-neighbor (MNN) searching is followed to establish coarse matches " + }, + { + "bbox": [ + 46, + 295, + 548, + 374 + ], + "type": "inline_equation", + "content": "\\{\\mathcal{M}_c\\}" + }, + { + "bbox": [ + 46, + 295, + 548, + 374 + ], + "type": "text", + "content": ". (4) To refine coarse matches, discriminative fine features " + }, + { + "bbox": [ + 46, + 295, + 548, + 374 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{F}}_A^t" + }, + { + "bbox": [ + 46, + 295, + 548, + 374 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 295, + 548, + 374 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{F}}_B^t" + }, + { + "bbox": [ + 46, + 295, + 548, + 374 + ], + "type": "text", + "content": " in full resolution are obtained by fusing transformed coarse features " + }, + { + "bbox": [ + 46, + 295, + 548, + 374 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{F}}_A^t" + }, + { + "bbox": [ + 46, + 295, + 548, + 374 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 295, + 548, + 374 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{F}}_B^t" + }, + { + "bbox": [ + 46, + 295, + 548, + 374 + ], + "type": "text", + "content": " with backbone features. Feature patches are then cropped centered at each coarse match " + }, + { + "bbox": [ + 46, + 295, + 548, + 374 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_c" + }, + { + "bbox": [ + 46, + 295, + 548, + 374 + ], + "type": "text", + "content": ". A two-stage refinement is followed to obtain sub-pixel correspondence " + }, + { + "bbox": [ + 46, + 295, + 548, + 374 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_f" + }, + { + "bbox": [ + 46, + 295, + 548, + 374 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 385, + 288, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 385, + 288, + 517 + ], + "spans": [ + { + "bbox": [ + 46, + 385, + 288, + 517 + ], + "type": "text", + "content": "efficiency, since the hierarchical nature of multi-scale attention will further introduce latencies. TopicFM [15] first assigns features with similar semantic meanings to the same topic, where attention is conducted within each topic for efficiency. Since it needs to sequentially process each token's features for transformation, the efficiency improvement is limited. Moreover, performing local attention within topics can potentially restrict the capability of modeling long-range dependencies. Compared with them, the proposed aggregated attention module in our method significantly improves efficiency while achieving better accuracy." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 517, + 288, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 517, + 288, + 590 + ], + "spans": [ + { + "bbox": [ + 46, + 517, + 288, + 590 + ], + "type": "text", + "content": "Dense matching methods [13, 14, 52] are designed to estimate all possible correspondences between two images, which show strong robustness. However, they are generally much slower compared with sparse and semi-dense methods. Unlike them, our method produces semi-dense matches with competitive performance and considerably better efficiency." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 594, + 289, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 289, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 289, + 714 + ], + "type": "text", + "content": "Transformer has been broadly used in multiple vision tasks, including feature matching. The efficiency and memory footprint of handling large token sizes are the main limitations of transformer [54], where some methods [19, 20, 56] attempt to reduce the complexity to a linear scale to alleviate these problems. Some methods [8, 22] propose optimizing transformer models for specific hardware architectures for memory and running-time efficiency. They are orthogonal to our method and can be naturally adapted into the pipeline for further efficiency improvement." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 306, + 384, + 361, + 396 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 384, + 361, + 396 + ], + "spans": [ + { + "bbox": [ + 306, + 384, + 361, + 396 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 407, + 547, + 480 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 407, + 547, + 480 + ], + "spans": [ + { + "bbox": [ + 304, + 407, + 547, + 480 + ], + "type": "text", + "content": "Given a pair of images " + }, + { + "bbox": [ + 304, + 407, + 547, + 480 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_A, \\mathbf{I}_B" + }, + { + "bbox": [ + 304, + 407, + 547, + 480 + ], + "type": "text", + "content": ", our objective is to establish a set of reliable correspondences between them. We achieve this by a coarse-to-fine matching pipeline, which first establishes coarse matches on downsampled feature maps and then refines them for high accuracy. An overview of our pipeline is shown in Fig. 2." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 497, + 446, + 509 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 497, + 446, + 509 + ], + "spans": [ + { + "bbox": [ + 306, + 497, + 446, + 509 + ], + "type": "text", + "content": "3.1. Local Feature Extraction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 518, + 547, + 602 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 518, + 547, + 602 + ], + "spans": [ + { + "bbox": [ + 304, + 518, + 547, + 602 + ], + "type": "text", + "content": "Image feature maps are first extracted by a lightweight backbone for later transformation and matching. Unlike LoFTR and many other detector-free matchers that use a heavy multibranch ResNet [17] network for feature extraction, we alternate to a lightweight single-branch network with reparameterization [10] to achieve better inference efficiency while preserving the model performance." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 605, + 548, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 548, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 548, + 714 + ], + "type": "text", + "content": "In particular, a multi-branch CNN network with residual connections is applied during training for maximum representational power. At inference time, we losslessly convert the feature backbone into an efficient single-branch network by adopting the reparameterization technique [10], which is achieved by fusing parallel convolution kernels into a single one. Then, the intermediate " + }, + { + "bbox": [ + 304, + 605, + 548, + 714 + ], + "type": "inline_equation", + "content": "\\frac{1}{8}" + }, + { + "bbox": [ + 304, + 605, + 548, + 714 + ], + "type": "text", + "content": " down-sampled coarse features " + }, + { + "bbox": [ + 304, + 605, + 548, + 714 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{F}}_A, \\tilde{\\mathbf{F}}_B" + }, + { + "bbox": [ + 304, + 605, + 548, + 714 + ], + "type": "text", + "content": " and fine features in " + }, + { + "bbox": [ + 304, + 605, + 548, + 714 + ], + "type": "inline_equation", + "content": "\\frac{1}{4}" + }, + { + "bbox": [ + 304, + 605, + 548, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 605, + 548, + 714 + ], + "type": "inline_equation", + "content": "\\frac{1}{2}" + }, + { + "bbox": [ + 304, + 605, + 548, + 714 + ], + "type": "text", + "content": " resolutions are extracted efficiently for later coarse-to-fine matching." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "21668" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 254, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 254, + 85 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 254, + 85 + ], + "type": "text", + "content": "3.2. Efficient Local Feature Transformation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 90, + 288, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 90, + 288, + 140 + ], + "spans": [ + { + "bbox": [ + 46, + 90, + 288, + 140 + ], + "type": "text", + "content": "After the feature extraction, the coarse-level feature maps " + }, + { + "bbox": [ + 46, + 90, + 288, + 140 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{F}}_A" + }, + { + "bbox": [ + 46, + 90, + 288, + 140 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 90, + 288, + 140 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{F}}_B" + }, + { + "bbox": [ + 46, + 90, + 288, + 140 + ], + "type": "text", + "content": " are transformed by interleaving self- and crossattention " + }, + { + "bbox": [ + 46, + 90, + 288, + 140 + ], + "type": "inline_equation", + "content": "^1 n" + }, + { + "bbox": [ + 46, + 90, + 288, + 140 + ], + "type": "text", + "content": " times to improve discriminativeness. The transformed features are denoted as " + }, + { + "bbox": [ + 46, + 90, + 288, + 140 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{F}}_A^t,\\tilde{\\mathbf{F}}_B^t" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 141, + 289, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 141, + 289, + 236 + ], + "spans": [ + { + "bbox": [ + 46, + 141, + 289, + 236 + ], + "type": "text", + "content": "Previous methods often perform attention on the entire coarse-level feature maps, where linear attention instead of vanilla attention is applied to ensure a manageable computation cost. However, the efficiency is still limited due to the large token size of coarse features. Moreover, the usage of linear attention leads to sub-optimal model capability. Unlike them, we propose efficient aggregated attention for both efficiency and performance." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 250, + 288, + 345 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 250, + 288, + 345 + ], + "spans": [ + { + "bbox": [ + 46, + 250, + 288, + 345 + ], + "type": "text", + "content": "Preliminaries. First, we provide a brief overview of the commonly used vanilla attention and linear attention. Vanilla attention is a core mechanism in transformer encoder layer, relying on three inputs: query Q, key K, and value V. The resultant output is a weighted sum of the value, where the weighted matrix is determined by the query and its corresponding key. Formally, the attention function is defined as follows:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 351, + 288, + 365 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 351, + 288, + 365 + ], + "spans": [ + { + "bbox": [ + 55, + 351, + 288, + 365 + ], + "type": "interline_equation", + "content": "\\operatorname {V a n i l l a A t t e n t i o n} (Q, K, V) = \\operatorname {s o f t m a x} (Q K ^ {T}) V. \\tag {1}", + "image_path": "c679521692c549b876402ec72c56d1d504beb22f1987621ef1d763b81f541c0a.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 371, + 288, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 371, + 288, + 430 + ], + "spans": [ + { + "bbox": [ + 46, + 371, + 288, + 430 + ], + "type": "text", + "content": "However, applying the vanilla attention directly to dense local features is impractical due to the significant token size. To address this issue, previous methods use linear attention to reduce the computational complexity from quadratic to linear:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 436, + 287, + 451 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 436, + 287, + 451 + ], + "spans": [ + { + "bbox": [ + 53, + 436, + 287, + 451 + ], + "type": "interline_equation", + "content": "\\operatorname {L i n e a r A t t e n t i o n} (Q, K, V) = \\phi (Q) (\\phi (K) ^ {T} \\phi (V)). \\tag {2}", + "image_path": "56072b50a2a7449bd26de427c3210aa3757ee4dbd4c9db0651f016c3ba71bfc9.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 456, + 288, + 493 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 456, + 288, + 493 + ], + "spans": [ + { + "bbox": [ + 46, + 456, + 288, + 493 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 456, + 288, + 493 + ], + "type": "inline_equation", + "content": "\\phi (\\cdot) = elu(\\cdot) + 1" + }, + { + "bbox": [ + 46, + 456, + 288, + 493 + ], + "type": "text", + "content": ". However, it comes at the cost of reduced representational power, which is also observed by [4]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 506, + 287, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 506, + 287, + 624 + ], + "spans": [ + { + "bbox": [ + 46, + 506, + 287, + 624 + ], + "type": "text", + "content": "Aggregated Attention Module. After comprehensively investigating the mechanism of the transformer on coarse feature maps, we have two observations that motivate us to devise a new efficient aggregated attention. First, the attention regions of neighboring query tokens are similar, thus we can aggregate the neighboring tokens of " + }, + { + "bbox": [ + 46, + 506, + 287, + 624 + ], + "type": "inline_equation", + "content": "f_{i}" + }, + { + "bbox": [ + 46, + 506, + 287, + 624 + ], + "type": "text", + "content": " to prevent the redundant computation. Second, most of the attention weights of each query token are concentrated on a small number of key tokens, hence we can select the salient tokens of " + }, + { + "bbox": [ + 46, + 506, + 287, + 624 + ], + "type": "inline_equation", + "content": "f_{j}" + }, + { + "bbox": [ + 46, + 506, + 287, + 624 + ], + "type": "text", + "content": " before attention to reduce the computation." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 625, + 288, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 625, + 288, + 662 + ], + "spans": [ + { + "bbox": [ + 46, + 625, + 288, + 662 + ], + "type": "text", + "content": "Therefore, we propose to first aggregate the " + }, + { + "bbox": [ + 46, + 625, + 288, + 662 + ], + "type": "inline_equation", + "content": "f_{i}" + }, + { + "bbox": [ + 46, + 625, + 288, + 662 + ], + "type": "text", + "content": " token utilizing a depth-wise convolution network, and " + }, + { + "bbox": [ + 46, + 625, + 288, + 662 + ], + "type": "inline_equation", + "content": "f_{j}" + }, + { + "bbox": [ + 46, + 625, + 288, + 662 + ], + "type": "text", + "content": " is aggregated by a max pooling layer to get reduced salient tokens:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 81, + 673, + 287, + 687 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 81, + 673, + 287, + 687 + ], + "spans": [ + { + "bbox": [ + 81, + 673, + 287, + 687 + ], + "type": "interline_equation", + "content": "f _ {i} ^ {\\prime} = \\operatorname {C o n v 2 D} \\left(f _ {i}\\right), f _ {j} ^ {\\prime} = \\operatorname {M a x P o o l} \\left(f _ {j}\\right) ， \\tag {3}", + "image_path": "b72c1eceb8a4630a4911104eab11e20878d81e674a606190c54eb28e88c13db1.jpg" + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 307, + 49, + 411, + 284 + ], + "blocks": [ + { + "bbox": [ + 307, + 49, + 411, + 284 + ], + "lines": [ + { + "bbox": [ + 307, + 49, + 411, + 284 + ], + "spans": [ + { + "bbox": [ + 307, + 49, + 411, + 284 + ], + "type": "image", + "image_path": "e1b0da0f135ead94bdcbe978d070b3778d2f98147bf05222106c9a358c772783.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 294, + 545, + 392 + ], + "lines": [ + { + "bbox": [ + 304, + 294, + 545, + 392 + ], + "spans": [ + { + "bbox": [ + 304, + 294, + 545, + 392 + ], + "type": "text", + "content": "Figure 3. Detailed Transformer Module Comparison. Unlike LoFTR which uses all tokens of feature maps to compute attention and resort to linear attention to reduce the computational cost, the proposed attention module first aggregates features for salient tokens, which is significantly more efficient for attention. Then the vanilla attention is utilized to transform aggregated features, where relative positional encoding is inserted to capture the spatial information. Transformed features are upsampled and fused with the original features to form the final features." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 418, + 48, + 543, + 285 + ], + "blocks": [ + { + "bbox": [ + 418, + 48, + 543, + 285 + ], + "lines": [ + { + "bbox": [ + 418, + 48, + 543, + 285 + ], + "spans": [ + { + "bbox": [ + 418, + 48, + 543, + 285 + ], + "type": "image", + "image_path": "473053f9f31671e5aa9fe3086f2e55c7e197fb322786d9cf7097f4ebb0663bdd.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 394, + 546, + 549 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 394, + 546, + 549 + ], + "spans": [ + { + "bbox": [ + 304, + 394, + 546, + 549 + ], + "type": "text", + "content": "where Conv2D is implemented by a strided depthwise convolution with a kernel size of " + }, + { + "bbox": [ + 304, + 394, + 546, + 549 + ], + "type": "inline_equation", + "content": "s \\times s" + }, + { + "bbox": [ + 304, + 394, + 546, + 549 + ], + "type": "text", + "content": ", identical to that of the max-pooling layer. Then positional encoding and vanilla attention are followed to process reduced tokens. Positional encoding (PE) can help to model the spatial location contexts, where RoPE [45] is adopted in practice to account for more robust relative positions, inspired by [26]. Note that the PE layer is enabled exclusively for self-attention and skipped during cross-attention. The transformed feature map is then upsampled and fused with " + }, + { + "bbox": [ + 304, + 394, + 546, + 549 + ], + "type": "inline_equation", + "content": "f_{i}" + }, + { + "bbox": [ + 304, + 394, + 546, + 549 + ], + "type": "text", + "content": " for the final feature map. Due to the aggregation and selection, the number of tokens in " + }, + { + "bbox": [ + 304, + 394, + 546, + 549 + ], + "type": "inline_equation", + "content": "f_{i}'" + }, + { + "bbox": [ + 304, + 394, + 546, + 549 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 394, + 546, + 549 + ], + "type": "inline_equation", + "content": "f_{j}'" + }, + { + "bbox": [ + 304, + 394, + 546, + 549 + ], + "type": "text", + "content": " is reduced by " + }, + { + "bbox": [ + 304, + 394, + 546, + 549 + ], + "type": "inline_equation", + "content": "s^{2}" + }, + { + "bbox": [ + 304, + 394, + 546, + 549 + ], + "type": "text", + "content": ", which contributes to the efficiency of the attention phase." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 556, + 473, + 569 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 556, + 473, + 569 + ], + "spans": [ + { + "bbox": [ + 306, + 556, + 473, + 569 + ], + "type": "text", + "content": "3.3. Coarse-level Matching Module" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 574, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 574, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 574, + 547, + 713 + ], + "type": "text", + "content": "We establish coarse-level matches based on the previously transformed coarse feature maps " + }, + { + "bbox": [ + 304, + 574, + 547, + 713 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{F}}_A^t" + }, + { + "bbox": [ + 304, + 574, + 547, + 713 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 574, + 547, + 713 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{F}}_B^t" + }, + { + "bbox": [ + 304, + 574, + 547, + 713 + ], + "type": "text", + "content": ". Coarse correspondences indicate rough match regions for later subpixel-level matching in the refinement phase. To achieve this, " + }, + { + "bbox": [ + 304, + 574, + 547, + 713 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{F}}_A^t" + }, + { + "bbox": [ + 304, + 574, + 547, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 574, + 547, + 713 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{F}}_B^t" + }, + { + "bbox": [ + 304, + 574, + 547, + 713 + ], + "type": "text", + "content": " are densely correlated to build a score matrix " + }, + { + "bbox": [ + 304, + 574, + 547, + 713 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 304, + 574, + 547, + 713 + ], + "type": "text", + "content": ". The softmax operator on both " + }, + { + "bbox": [ + 304, + 574, + 547, + 713 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 304, + 574, + 547, + 713 + ], + "type": "text", + "content": " dimensions (referred to as dual-softmax) is then applied to obtain the probability of mutual nearest matching, which is commonly used in [35, 46, 53]. The coarse correspondences " + }, + { + "bbox": [ + 304, + 574, + 547, + 713 + ], + "type": "inline_equation", + "content": "\\{\\mathcal{M}_c\\}" + }, + { + "bbox": [ + 304, + 574, + 547, + 713 + ], + "type": "text", + "content": " are established by selecting matches above the score threshold " + }, + { + "bbox": [ + 304, + 574, + 547, + 713 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 304, + 574, + 547, + 713 + ], + "type": "text", + "content": " while satisfying the mutual-nearest-neighbor (MNN) constraint." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 46, + 693, + 287, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 693, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 693, + 287, + 713 + ], + "type": "text", + "content": "1We feed feature of one image as query and feature of the other image as key and value into cross-attention, similar to SG [39] and LoFTR [46]." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "21669" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 167 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 167 + ], + "type": "text", + "content": "Efficient Inference Strategy. We observe that the dual-softmax operator in the coarse matching can significantly restrict the efficiency in inference due to the large token size, especially for high-resolution images. Moreover, we find that the dual-softmax operator is crucial for training, dropping it at inference time while directly using the score matrix " + }, + { + "bbox": [ + 46, + 72, + 289, + 167 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 46, + 72, + 289, + 167 + ], + "type": "text", + "content": " for MNN matching can also work well with better efficiency." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 168, + 289, + 264 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 168, + 289, + 264 + ], + "spans": [ + { + "bbox": [ + 46, + 168, + 289, + 264 + ], + "type": "text", + "content": "The reason for using the dual-softmax operator in training is that it can help to train discriminative features. Intuitively, with the softmax operation, the matching score between two pixels can also conditioned on other pixels. This mechanism forces the network to improve feature similarity of true correspondences while suppressing similarity with irrelevant points. With trained discriminative features, the softmax operation can be potentially eliminated during inference." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 264, + 289, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 264, + 289, + 300 + ], + "spans": [ + { + "bbox": [ + 46, + 264, + 289, + 300 + ], + "type": "text", + "content": "We denote the model skipping dual-softmax layer in inference as efficiency optimized model. Results in Tab. 1 demonstrate the effectiveness of this design." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 308, + 235, + 320 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 308, + 235, + 320 + ], + "spans": [ + { + "bbox": [ + 47, + 308, + 235, + 320 + ], + "type": "text", + "content": "3.4. Subpixel-Level Refinement Module" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 327, + 289, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 327, + 289, + 389 + ], + "spans": [ + { + "bbox": [ + 47, + 327, + 289, + 389 + ], + "type": "text", + "content": "As overviewed in Fig. 2 (4), with established coarse matches " + }, + { + "bbox": [ + 47, + 327, + 289, + 389 + ], + "type": "inline_equation", + "content": "\\{\\mathcal{M}_c\\}" + }, + { + "bbox": [ + 47, + 327, + 289, + 389 + ], + "type": "text", + "content": ", we refine them for sub-pixel accuracy with our refinement module. It is composed of an efficient feature patch extractor for discriminative fine features, followed by a two-stage feature correlation layer for final matches " + }, + { + "bbox": [ + 47, + 327, + 289, + 389 + ], + "type": "inline_equation", + "content": "\\{\\mathcal{M}_f\\}" + }, + { + "bbox": [ + 47, + 327, + 289, + 389 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 403, + 289, + 501 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 403, + 289, + 501 + ], + "spans": [ + { + "bbox": [ + 46, + 403, + 289, + 501 + ], + "type": "text", + "content": "Efficient Fine Feature Extraction. We first extract discriminative fine feature patches centered at each coarse match " + }, + { + "bbox": [ + 46, + 403, + 289, + 501 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_c" + }, + { + "bbox": [ + 46, + 403, + 289, + 501 + ], + "type": "text", + "content": " by an efficient fusion network for later match refinement. For efficiency, our key idea here is to re-ileverage the previously transformed coarse features " + }, + { + "bbox": [ + 46, + 403, + 289, + 501 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{F}}_A^t" + }, + { + "bbox": [ + 46, + 403, + 289, + 501 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 403, + 289, + 501 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{F}}_B^t" + }, + { + "bbox": [ + 46, + 403, + 289, + 501 + ], + "type": "text", + "content": " to obtain cross-view attended discriminative fine features, instead of introducing additional feature transform networks as in LoFTR [46]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 501, + 289, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 501, + 289, + 590 + ], + "spans": [ + { + "bbox": [ + 46, + 501, + 289, + 590 + ], + "type": "text", + "content": "To be specific, " + }, + { + "bbox": [ + 46, + 501, + 289, + 590 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{F}}_A^t" + }, + { + "bbox": [ + 46, + 501, + 289, + 590 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 501, + 289, + 590 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{F}}_B^t" + }, + { + "bbox": [ + 46, + 501, + 289, + 590 + ], + "type": "text", + "content": " are adaptively fused with " + }, + { + "bbox": [ + 46, + 501, + 289, + 590 + ], + "type": "inline_equation", + "content": "1/4" + }, + { + "bbox": [ + 46, + 501, + 289, + 590 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 501, + 289, + 590 + ], + "type": "inline_equation", + "content": "1/2" + }, + { + "bbox": [ + 46, + 501, + 289, + 590 + ], + "type": "text", + "content": " resolution backbone features by convolution and upsampling to obtain fine feature maps " + }, + { + "bbox": [ + 46, + 501, + 289, + 590 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{F}}_A^t" + }, + { + "bbox": [ + 46, + 501, + 289, + 590 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 501, + 289, + 590 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{F}}_B^t" + }, + { + "bbox": [ + 46, + 501, + 289, + 590 + ], + "type": "text", + "content": " in the original image resolution. Then " + }, + { + "bbox": [ + 46, + 501, + 289, + 590 + ], + "type": "inline_equation", + "content": "8 \\times 8" + }, + { + "bbox": [ + 46, + 501, + 289, + 590 + ], + "type": "text", + "content": " local feature patches are cropped on fine feature maps centered at each coarse match. Since only shallow feed-forward networks are included, our fine feature fusion network is remarkably efficient." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 605, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 289, + 715 + ], + "type": "text", + "content": "Two-Stage Correlation for Refinement. Based on the extracted fine local feature patches of coarse matches, we search for high-accurate sub-pixel matches. To refine a coarse match, a commonly used strategy [6, 15, 46] is to select the center-patch feature of " + }, + { + "bbox": [ + 46, + 605, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_A" + }, + { + "bbox": [ + 46, + 605, + 289, + 715 + ], + "type": "text", + "content": " as a fixed reference point, and perform feature correlation and expectation on the entire corresponding feature patch for its fine match. However, this refinement-by-expectations will introduce location variance to the final match, because irrelevant regions also" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 72, + 545, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 96 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 96 + ], + "type": "text", + "content": "have weights and can affect results. Therefore, we propose a novel two-stage correlation module to solve this problem." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 97, + 547, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 97, + 547, + 216 + ], + "spans": [ + { + "bbox": [ + 304, + 97, + 547, + 216 + ], + "type": "text", + "content": "Our idea is to utilize a mutual-nearest-neighbor (MNN) matching to get intermediate pixel-level refined matches in the first stage, and then refine them for subpixel accuracy by correlation and expectation. Motivations are that MNN matching don't have spatial variance since matches are selected by directly indexing pixels with maximum scores, but cannot achieve sub-pixel accuracy. Conversely, refinement-by-expectation can achieve sub-pixel accuracy but variance exists. The proposed two-stage refinement can draw benefits by combining the best of both worlds." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 217, + 547, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 217, + 547, + 300 + ], + "spans": [ + { + "bbox": [ + 304, + 217, + 547, + 300 + ], + "type": "text", + "content": "Specifically, to refine a coarse-level correspondence " + }, + { + "bbox": [ + 304, + 217, + 547, + 300 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_c" + }, + { + "bbox": [ + 304, + 217, + 547, + 300 + ], + "type": "text", + "content": ", the first-stage refinement phase densely correlates their fine feature patches to obtain the local patch score matrix " + }, + { + "bbox": [ + 304, + 217, + 547, + 300 + ], + "type": "inline_equation", + "content": "S_l" + }, + { + "bbox": [ + 304, + 217, + 547, + 300 + ], + "type": "text", + "content": ". MNN searching is then applied on " + }, + { + "bbox": [ + 304, + 217, + 547, + 300 + ], + "type": "inline_equation", + "content": "S_l" + }, + { + "bbox": [ + 304, + 217, + 547, + 300 + ], + "type": "text", + "content": " to get intermediate pixel-level fine matches. To limit the overall match number, we select the top-1 fine match for one coarse match by sorting the correlation scores." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 300, + 547, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 300, + 547, + 422 + ], + "spans": [ + { + "bbox": [ + 304, + 300, + 547, + 422 + ], + "type": "text", + "content": "Then, we further refine these pixel-level matches for sub-pixel accuracy by our second-stage refinement. Since the matching accuracy has already significantly improved in first-stage refinement, now we can use a tiny local window for correlation and expectation with a maximum suppression of location variance. In practice, we correlate the feature of each point in " + }, + { + "bbox": [ + 304, + 300, + 547, + 422 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_A" + }, + { + "bbox": [ + 304, + 300, + 547, + 422 + ], + "type": "text", + "content": " with a " + }, + { + "bbox": [ + 304, + 300, + 547, + 422 + ], + "type": "inline_equation", + "content": "3\\times 3" + }, + { + "bbox": [ + 304, + 300, + 547, + 422 + ], + "type": "text", + "content": " feature patch centered at its fine match in " + }, + { + "bbox": [ + 304, + 300, + 547, + 422 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_B" + }, + { + "bbox": [ + 304, + 300, + 547, + 422 + ], + "type": "text", + "content": ". The softmax operator is then applied to get a match distribution matrix and the final refined match is obtained by calculating expectations." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 430, + 384, + 442 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 430, + 384, + 442 + ], + "spans": [ + { + "bbox": [ + 306, + 430, + 384, + 442 + ], + "type": "text", + "content": "3.5. Supervision" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 449, + 545, + 474 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 449, + 545, + 474 + ], + "spans": [ + { + "bbox": [ + 304, + 449, + 545, + 474 + ], + "type": "text", + "content": "The entire pipeline is trained end-to-end by supervising the coarse and refinement matching modules separately." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 490, + 547, + 576 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 490, + 547, + 576 + ], + "spans": [ + { + "bbox": [ + 304, + 490, + 547, + 576 + ], + "type": "text", + "content": "Coarse-Level Matching Supervision. The coarse ground truth matches " + }, + { + "bbox": [ + 304, + 490, + 547, + 576 + ], + "type": "inline_equation", + "content": "\\{\\mathcal{M}_c\\}_{gt}" + }, + { + "bbox": [ + 304, + 490, + 547, + 576 + ], + "type": "text", + "content": " with a total number of " + }, + { + "bbox": [ + 304, + 490, + 547, + 576 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 490, + 547, + 576 + ], + "type": "text", + "content": " are built by warping grid-level points from " + }, + { + "bbox": [ + 304, + 490, + 547, + 576 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_A" + }, + { + "bbox": [ + 304, + 490, + 547, + 576 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 304, + 490, + 547, + 576 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_B" + }, + { + "bbox": [ + 304, + 490, + 547, + 576 + ], + "type": "text", + "content": " via depth maps and image poses following previous methods [39, 46]. The produced correlation score matrix " + }, + { + "bbox": [ + 304, + 490, + 547, + 576 + ], + "type": "inline_equation", + "content": "\\mathcal{S}" + }, + { + "bbox": [ + 304, + 490, + 547, + 576 + ], + "type": "text", + "content": " in coarse matching is supervised by minimizing the log-likelihood loss over locations of " + }, + { + "bbox": [ + 304, + 490, + 547, + 576 + ], + "type": "inline_equation", + "content": "\\{\\mathcal{M}_c\\}_{gt}" + }, + { + "bbox": [ + 304, + 490, + 547, + 576 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 337, + 584, + 547, + 602 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 337, + 584, + 547, + 602 + ], + "spans": [ + { + "bbox": [ + 337, + 584, + 547, + 602 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {c} = - \\frac {1}{N} \\sum_ {(\\tilde {i}, \\tilde {j}) \\in \\left\\{\\mathcal {M} _ {c} \\right\\} _ {g t}} \\log \\mathcal {S} (\\tilde {i}, \\tilde {j}). \\tag {4}", + "image_path": "34522aecd8cd0b587bb0670438fe045e9e1e2b8f5cf43e622d467a979de54678.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 617, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 617, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 617, + 547, + 715 + ], + "type": "text", + "content": "Fine-Level Matching Supervision. We train the proposed two-stage fine-level matching module by separately supervising the two phases. The first stage fine loss " + }, + { + "bbox": [ + 304, + 617, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{f1}" + }, + { + "bbox": [ + 304, + 617, + 547, + 715 + ], + "type": "text", + "content": " is to minimize the log-likelihood loss of each fine local score matrix " + }, + { + "bbox": [ + 304, + 617, + 547, + 715 + ], + "type": "inline_equation", + "content": "S_{l}" + }, + { + "bbox": [ + 304, + 617, + 547, + 715 + ], + "type": "text", + "content": " based on the pixel-level ground truth fine matches, similar to coarse loss. The second stage is trained by " + }, + { + "bbox": [ + 304, + 617, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{f2}" + }, + { + "bbox": [ + 304, + 617, + 547, + 715 + ], + "type": "text", + "content": " that calculates the " + }, + { + "bbox": [ + 304, + 617, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\ell_2" + }, + { + "bbox": [ + 304, + 617, + 547, + 715 + ], + "type": "text", + "content": " loss between the final subpixel matches " + }, + { + "bbox": [ + 304, + 617, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\{\\mathcal{M}_f\\}" + }, + { + "bbox": [ + 304, + 617, + 547, + 715 + ], + "type": "text", + "content": " and ground truth fine matches " + }, + { + "bbox": [ + 304, + 617, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\{\\mathcal{M}_f\\}_{gt}" + }, + { + "bbox": [ + 304, + 617, + 547, + 715 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "21670" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 75, + 56, + 519, + 203 + ], + "blocks": [ + { + "bbox": [ + 75, + 56, + 519, + 203 + ], + "lines": [ + { + "bbox": [ + 75, + 56, + 519, + 203 + ], + "spans": [ + { + "bbox": [ + 75, + 56, + 519, + 203 + ], + "type": "table", + "html": "
CategoryMethodMegaDepth DatasetScanNet DatasetTime (ms)
AUC@5°AUC@10°AUC@20°AUC@5°AUC@10°AUC@20°
SparseSP + NN31.746.860.17.518.632.110.8
SP + SG49.767.180.616.232.849.748.3
SP + LG49.967.080.114.830.847.531.9/30.7
Semi-DenseDRC-Net27.042.958.37.717.930.5328.0
LoFTR52.869.281.216.933.650.666.2
QuadTree54.670.582.219.037.353.5100.7
MatchFormer53.369.781.815.832.048.0128.9
TopicFM54.170.181.617.335.550.966.4
AspanFormer55.371.583.119.637.754.481.6
Ours56.472.283.519.237.053.640.1/34.4
Ours (Optimized)55.471.482.917.434.451.235.6/27.0
DenseDKM60.474.985.126.6447.0764.17210.8
ROMA62.676.786.328.950.468.3302.7
", + "image_path": "4427a473cf42cd32049546867b637e7fb785ca19d44f46d427031e05f578cef8.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 215, + 545, + 270 + ], + "lines": [ + { + "bbox": [ + 46, + 215, + 545, + 270 + ], + "spans": [ + { + "bbox": [ + 46, + 215, + 545, + 270 + ], + "type": "text", + "content": "Table 1. Results of Relative Pose Estimation on MegaDepth Dataset and ScanNet Dataset. We use the models trained on the MegaDepth dataset to evaluate all methods on both datasets, which can show the intra- and inter-dataset generalization abilities. The AUC of pose error at different thresholds, along with the processing time for matching image pair at a resolution of " + }, + { + "bbox": [ + 46, + 215, + 545, + 270 + ], + "type": "inline_equation", + "content": "640 \\times 480" + }, + { + "bbox": [ + 46, + 215, + 545, + 270 + ], + "type": "text", + "content": ", is presented. For SP + LG, Ours, and Ours (Optimized), the running times of the model using FP32/Mixed-Precision numerical precisions (without Efficient Attention [8, 22]) are shown." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 47, + 282, + 288, + 307 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 282, + 288, + 307 + ], + "spans": [ + { + "bbox": [ + 47, + 282, + 288, + 307 + ], + "type": "text", + "content": "The total loss is the weighted sum of all supervisions: " + }, + { + "bbox": [ + 47, + 282, + 288, + 307 + ], + "type": "inline_equation", + "content": "\\mathcal{L} = \\mathcal{L}_c + \\alpha \\mathcal{L}_{f1} + \\beta \\mathcal{L}_{f2}" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 316, + 128, + 330 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 316, + 128, + 330 + ], + "spans": [ + { + "bbox": [ + 47, + 316, + 128, + 330 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 336, + 288, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 336, + 288, + 396 + ], + "spans": [ + { + "bbox": [ + 46, + 336, + 288, + 396 + ], + "type": "text", + "content": "In this section, we evaluate the performance of our method on several downstream tasks, including homography estimation, pairwise pose estimation and visual localization. Furthermore, we evaluate the effectiveness of our design by conducting detailed ablation studies." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 403, + 180, + 415 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 403, + 180, + 415 + ], + "spans": [ + { + "bbox": [ + 47, + 403, + 180, + 415 + ], + "type": "text", + "content": "4.1. Implementation Details" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 45, + 421, + 288, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 421, + 288, + 601 + ], + "spans": [ + { + "bbox": [ + 45, + 421, + 288, + 601 + ], + "type": "text", + "content": "We adopt RepVGG [10] as our feature backbone, and self- and cross-attention are interleaved for " + }, + { + "bbox": [ + 45, + 421, + 288, + 601 + ], + "type": "inline_equation", + "content": "N = 4" + }, + { + "bbox": [ + 45, + 421, + 288, + 601 + ], + "type": "text", + "content": " times to transform coarse features. For each attention, we aggregate features by a depth-wise convolution layer and a max-pooling layer, both with a kernel size of " + }, + { + "bbox": [ + 45, + 421, + 288, + 601 + ], + "type": "inline_equation", + "content": "4 \\times 4" + }, + { + "bbox": [ + 45, + 421, + 288, + 601 + ], + "type": "text", + "content": ". Our model is trained on the MegaDepth dataset [24], which is a large-scale outdoor dataset. The test scenes are separated from training data following [46]. The loss function's weights " + }, + { + "bbox": [ + 45, + 421, + 288, + 601 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 45, + 421, + 288, + 601 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 45, + 421, + 288, + 601 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 45, + 421, + 288, + 601 + ], + "type": "text", + "content": " are set to 1.0 and 0.25, respectively. We use the AdamW optimizer with an initial learning rate of " + }, + { + "bbox": [ + 45, + 421, + 288, + 601 + ], + "type": "inline_equation", + "content": "4 \\times 10^{-3}" + }, + { + "bbox": [ + 45, + 421, + 288, + 601 + ], + "type": "text", + "content": ". The network training takes about 15 hours with a batch size of 20 on 4 NVIDIA V100 GPUs. And the coarse and fine stages are trained together from scratch. The trained model on MegaDepth is used to evaluate all datasets and tasks in our experiments to demonstrate the generalization ability." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 607, + 186, + 619 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 607, + 186, + 619 + ], + "spans": [ + { + "bbox": [ + 47, + 607, + 186, + 619 + ], + "type": "text", + "content": "4.2. Relative Pose Estimation" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 629, + 287, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 629, + 287, + 665 + ], + "spans": [ + { + "bbox": [ + 46, + 629, + 287, + 665 + ], + "type": "text", + "content": "Datasets. We use the outdoor MegaDepth [24] dataset and indoor ScanNet [7] dataset for the evaluation of relative pose estimation to demonstrate the efficacy of our method." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 665, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 288, + 713 + ], + "type": "text", + "content": "MegaDepth dataset is a large-scale dataset containing sparse 3D reconstructions from 196 scenes. The key challenges on this dataset are large viewpoints and illumination changes, as well as repetitive patterns. We follow the test" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 281, + 547, + 353 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 281, + 547, + 353 + ], + "spans": [ + { + "bbox": [ + 304, + 281, + 547, + 353 + ], + "type": "text", + "content": "split of the previous method [46] that uses 1500 sampled pairs from scenes \"Sacre Coeur\" and \"St. Peter's Square\" for evaluation. Images are resized so that the longest edge equals 1200 for all semi-dense and dense methods. Following [26], sparse methods are provided resized images with longest edge equals 1600." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 355, + 547, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 355, + 547, + 415 + ], + "spans": [ + { + "bbox": [ + 304, + 355, + 547, + 415 + ], + "type": "text", + "content": "ScanNet dataset contains 1613 sequences with ground-truth depth maps and camera poses. They depict indoor scenes with viewpoint changes and texture-less regions. We use the sampled test pairs from [39] for the evaluation, where images are resized to " + }, + { + "bbox": [ + 304, + 355, + 547, + 415 + ], + "type": "inline_equation", + "content": "640 \\times 480" + }, + { + "bbox": [ + 304, + 355, + 547, + 415 + ], + "type": "text", + "content": " for all methods." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 420, + 547, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 420, + 547, + 529 + ], + "spans": [ + { + "bbox": [ + 304, + 420, + 547, + 529 + ], + "type": "text", + "content": "Baselines. We compare the proposed method with three categories of methods: 1) sparse keypoint detection and matching methods, including SuperPoint [9] with Nearest-Neighbor (NN), SuperGlue (SG) [39], LightGlue (LG) [26] matchers. 2) semi-dense matchers, including DRC-Net [23], LoFTR [46], QuadTree Attention [48], MatchFormer [55], AspanFormer [6], TopicFM [15], and 3) state-of-the-art dense matcher ROMA [14] that predict matches for each pixel." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 533, + 548, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 533, + 548, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 533, + 548, + 713 + ], + "type": "text", + "content": "Evaluation protocol. Following previous methods, the recovered relative poses by matches are evaluated for reflecting matching accuracy. The pose error is defined as the maximum of angular errors in rotation and translation. We report the AUC of the pose error at thresholds " + }, + { + "bbox": [ + 304, + 533, + 548, + 713 + ], + "type": "inline_equation", + "content": "(5^{\\circ}, 10^{\\circ}" + }, + { + "bbox": [ + 304, + 533, + 548, + 713 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 533, + 548, + 713 + ], + "type": "inline_equation", + "content": "20^{\\circ})" + }, + { + "bbox": [ + 304, + 533, + 548, + 713 + ], + "type": "text", + "content": ". Moreover, the running time of matching each image pair in the ScanNet dataset on a single NVIDIA RTX 3090 is reported for comprehensively understanding the matching accuracy and efficiency balance. We also include the running time of feature extraction and keypoint detection (SuperPoint) for the sparse method to compare with dense methods, which are end-to-end (from image input to matches output). LG [26]'s SuperPoint [9] parameters (max keypoints=2048, keypoint threshold=0.0005, nms=4) are uniformly used for all sparse methods." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "21671" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 51, + 56, + 286, + 152 + ], + "blocks": [ + { + "bbox": [ + 51, + 56, + 286, + 152 + ], + "lines": [ + { + "bbox": [ + 51, + 56, + 286, + 152 + ], + "spans": [ + { + "bbox": [ + 51, + 56, + 286, + 152 + ], + "type": "table", + "html": "
CategoryMethodHomography est. AUC
@3px@5px@10px
SparseD2Net + NN23.235.953.6
R2D2 + NN50.663.976.8
DISK + NN52.364.978.9
SP + SG53.968.381.7
Semi-DenseSparse-NCNet48.954.267.1
DRC-Net50.656.268.3
LoFTR65.975.684.6
Ours66.576.485.5
", + "image_path": "fb808f9295b033503722733f24114686b7896512c0118aa6192b9ce804a51320.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 215, + 289, + 455 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 215, + 289, + 455 + ], + "spans": [ + { + "bbox": [ + 46, + 215, + 289, + 455 + ], + "type": "text", + "content": "Results. As shown in Tab. 1, the proposed method achieves competitive performances compared with sparse and semidense methods on both datasets. Qualitative comparisons are shown in Fig. 4. Specifically, our method outperforms the best semi-dense baseline AspanFormer on all metrics of the MegaDepth dataset and has lower but comparable performance on the ScanNet dataset, with " + }, + { + "bbox": [ + 46, + 215, + 289, + 455 + ], + "type": "inline_equation", + "content": "\\sim 2" + }, + { + "bbox": [ + 46, + 215, + 289, + 455 + ], + "type": "text", + "content": " times faster. Our optimized model that eliminates the dual-softmax operator in coarse-level matching further brings efficiency improvements, with slight performance decreases. Using this strategy, our method can outperform the efficient and robust sparse method " + }, + { + "bbox": [ + 46, + 215, + 289, + 455 + ], + "type": "inline_equation", + "content": "\\mathrm{SP} + \\mathrm{LG}" + }, + { + "bbox": [ + 46, + 215, + 289, + 455 + ], + "type": "text", + "content": " in efficiency with significantly higher accuracy. Dense matcher ROMA shows remarkable matching capability but is slow for applications in practice. Moreover, since ROMA utilizes the pre-trained DINOv2 [33] backbone, its strong generalizability on ScanNet may be attributed to the similar indoor training data in DINOv2, where other methods are trained on outdoor MegaDepth only. Compared with it, our method is " + }, + { + "bbox": [ + 46, + 215, + 289, + 455 + ], + "type": "inline_equation", + "content": "\\sim 7.5" + }, + { + "bbox": [ + 46, + 215, + 289, + 455 + ], + "type": "text", + "content": " times faster, which has a good balance between accuracy and efficiency." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 464, + 186, + 478 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 464, + 186, + 478 + ], + "spans": [ + { + "bbox": [ + 47, + 464, + 186, + 478 + ], + "type": "text", + "content": "4.3. Homography Estimation" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 488, + 288, + 536 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 488, + 288, + 536 + ], + "spans": [ + { + "bbox": [ + 46, + 488, + 288, + 536 + ], + "type": "text", + "content": "Dataset. We evaluate our method on HPatches dataset [2]. HPatches dataset depicts planar scenes divided into sequences. Images are taken under different viewpoints or illumination changes." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 540, + 287, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 540, + 287, + 624 + ], + "spans": [ + { + "bbox": [ + 46, + 540, + 287, + 624 + ], + "type": "text", + "content": "Baselines. We compare our method with sparse methods including D2Net [11], R2D2 [34], DISK [53] detectors with NN matcher, and SuperPoint [9] + SuperGlue [39]. As for semi-dense methods, we compare with Sparse-NCNet [36], DRC-Net [23], and LoFTR [46]. For SuperGlue and all semidense methods, we use their models trained on MegaDepth dataset for evaluation." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 629, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 629, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 629, + 288, + 714 + ], + "type": "text", + "content": "Evaluation Protocol. Following SuperGlue [39] and LoFTR [46], we resize all images for matching so that their smallest edge equals 480 pixels. We collect the mean reprojection error of corner points, and report the area under the cumulative curve (AUC) under 3 different thresholds, including 3 px, 5 px, and 10 px. For all baselines, we employ the same RANSAC method as a robust homography estimator" + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 310, + 56, + 544, + 134 + ], + "blocks": [ + { + "bbox": [ + 46, + 157, + 288, + 201 + ], + "lines": [ + { + "bbox": [ + 46, + 157, + 288, + 201 + ], + "spans": [ + { + "bbox": [ + 46, + 157, + 288, + 201 + ], + "type": "text", + "content": "Table 2. Results of Homography Estimation on HPatches Dataset. Our method is compared with sparse and semi-dense methods. The AUC of reprojection error of corner points at different thresholds is reported." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 310, + 56, + 544, + 134 + ], + "lines": [ + { + "bbox": [ + 310, + 56, + 544, + 134 + ], + "spans": [ + { + "bbox": [ + 310, + 56, + 544, + 134 + ], + "type": "table", + "html": "
MethodDUC1DUC2
(0.25m,2°)/(0.5m,5°)/(1.0m,10°)
SP+SG49.0 / 68.7 / 80.853.4 / 77.1 / 82.4
LoFTR47.5 / 72.2 / 84.854.2 / 74.8 / 85.5
TopicFM52.0 / 74.7 / 87.453.4 / 74.8 / 83.2
PATS55.6 / 71.2 / 81.058.8 / 80.9 / 85.5
AspanFormer51.5 / 73.7 / 86.055.0 / 74.0 / 81.7
Ours52.0 / 74.7 / 86.958.0 / 80.9 / 89.3
", + "image_path": "43bb865f1e053b2ef07e8e271d899228814e6937e479000998b87bb724f06dad.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 310, + 158, + 544, + 236 + ], + "blocks": [ + { + "bbox": [ + 316, + 139, + 534, + 149 + ], + "lines": [ + { + "bbox": [ + 316, + 139, + 534, + 149 + ], + "spans": [ + { + "bbox": [ + 316, + 139, + 534, + 149 + ], + "type": "text", + "content": "Table 3. Results of Visual Localization on InLoc Dataset." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 310, + 158, + 544, + 236 + ], + "lines": [ + { + "bbox": [ + 310, + 158, + 544, + 236 + ], + "spans": [ + { + "bbox": [ + 310, + 158, + 544, + 236 + ], + "type": "table", + "html": "
MethodDayNight
(0.25m,2°)/(0.5m,5°)/(1.0m,10°)
SP+SG89.8 / 96.1 / 99.477.0 / 90.6 / 100.0
LoFTR88.7 / 95.6 / 99.078.5 / 90.6 / 99.0
TopicFM90.2 / 95.9 / 98.977.5 / 91.1 / 99.5
PATS89.6 / 95.8 / 99.373.8 / 92.1 / 99.5
AspanFormer89.4 / 95.6 / 99.077.5 / 91.6 / 99.5
Ours89.6 / 96.2 / 99.077.0 / 91.1 / 99.5
", + "image_path": "c41257502ba55398c518460753f6776e1807e41356fdfbb17f679c04ef75d50b.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 240, + 545, + 251 + ], + "lines": [ + { + "bbox": [ + 305, + 240, + 545, + 251 + ], + "spans": [ + { + "bbox": [ + 305, + 240, + 545, + 251 + ], + "type": "text", + "content": "Table 4. Results of Visual Localization on Aachen v1.1 Dataset." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 305, + 264, + 545, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 264, + 545, + 300 + ], + "spans": [ + { + "bbox": [ + 305, + 264, + 545, + 300 + ], + "type": "text", + "content": "for a fair comparison. Following LoFTR, we select only the top 1000 predicted matches of semi-dense methods for the sake of fairness." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 304, + 547, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 304, + 547, + 389 + ], + "spans": [ + { + "bbox": [ + 305, + 304, + 547, + 389 + ], + "type": "text", + "content": "Results. As shown in Tab. 2, even though the number of matches is restricted, our method can also work remarkably well and outperform sparse methods significantly. Compared with semi-dense, our method can surpass them with significantly higher efficiency. We attribute this to the effectiveness of two-stage refinement for accuracy improvement and proposed aggregation module for efficiency." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 398, + 418, + 410 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 398, + 418, + 410 + ], + "spans": [ + { + "bbox": [ + 306, + 398, + 418, + 410 + ], + "type": "text", + "content": "4.4. Visual Localization" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 421, + 547, + 589 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 421, + 547, + 589 + ], + "spans": [ + { + "bbox": [ + 304, + 421, + 547, + 589 + ], + "type": "text", + "content": "Datasets and Evaluation Protocols. Visual localization is an important downstream task of image matching, which aims to estimate the 6-DoF poses of query images based on the 3D scene model. We conduct experiments on two commonly used benchmarks, including InLoc [47] dataset and Aachen v1.1 [41] dataset, for evaluation to demonstrate the superiority of our method. InLoc dataset is captured on indoor scenes with plenty of repetitive structures and textureless regions, where each database image has a corresponding depth map. Aachen v1.1 is a challenging large-scale outdoor dataset for localization with large-viewpoint and day-and-night illumination changes, which particularly relies on the robustness of matching methods. We adopt its full localization track for benchmarking." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 589, + 547, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 589, + 547, + 685 + ], + "spans": [ + { + "bbox": [ + 304, + 589, + 547, + 685 + ], + "type": "text", + "content": "Following [6, 46], the open-sourced localization framework HLoc [38] is utilized. For both datasets, the percentage of pose errors satisfying both angular and distance thresholds is reported following the benchmarks, where different thresholds are used. For the InLoc dataset, the metrics of two test scenes including DUC1 and DUC2 are separately reported. As for the Aachen v1.1 dataset, the metrics corresponding to the daytime and nighttime divisions are reported." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 689, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 545, + 714 + ], + "type": "text", + "content": "Baselines. We compare the proposed method with both detector-based method SuperPoint [9] + SuperGlue [39] and" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "21672" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 88, + 38, + 231, + 196 + ], + "blocks": [ + { + "bbox": [ + 88, + 38, + 231, + 196 + ], + "lines": [ + { + "bbox": [ + 88, + 38, + 231, + 196 + ], + "spans": [ + { + "bbox": [ + 88, + 38, + 231, + 196 + ], + "type": "image", + "image_path": "c381cf5d3a60fdbc0fdc9bca1154fc524231d28359a4512b136d2d04f96390e1.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 238, + 38, + 370, + 198 + ], + "blocks": [ + { + "bbox": [ + 238, + 38, + 370, + 198 + ], + "lines": [ + { + "bbox": [ + 238, + 38, + 370, + 198 + ], + "spans": [ + { + "bbox": [ + 238, + 38, + 370, + 198 + ], + "type": "image", + "image_path": "e592b78f9f7fe6f413684e84894725f215261dc9270936e8fed1412663c13b56.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 198, + 547, + 232 + ], + "lines": [ + { + "bbox": [ + 46, + 198, + 547, + 232 + ], + "spans": [ + { + "bbox": [ + 46, + 198, + 547, + 232 + ], + "type": "text", + "content": "Figure 4. Qualitative Results. Our method is compared with the sparse matching pipeline SuperPoint [9] " + }, + { + "bbox": [ + 46, + 198, + 547, + 232 + ], + "type": "inline_equation", + "content": "^+" + }, + { + "bbox": [ + 46, + 198, + 547, + 232 + ], + "type": "text", + "content": " LightGlue [26], semi-dense \nmacher AspanFormer [6]. Image pairs with texture-poor regions and large-viewpoint changes can be robustly matched by our method. The \nred color indicates epipolar error beyond " + }, + { + "bbox": [ + 46, + 198, + 547, + 232 + ], + "type": "inline_equation", + "content": "5\\times 10^{-4}" + }, + { + "bbox": [ + 46, + 198, + 547, + 232 + ], + "type": "text", + "content": " (in the normalized image coordinates)." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 378, + 40, + 508, + 198 + ], + "blocks": [ + { + "bbox": [ + 378, + 40, + 508, + 198 + ], + "lines": [ + { + "bbox": [ + 378, + 40, + 508, + 198 + ], + "spans": [ + { + "bbox": [ + 378, + 40, + 508, + 198 + ], + "type": "image", + "image_path": "9b21e807ba425ad2643d05931505c51b460341d8f5a1f3829da0f4372e2bc8ec.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 50, + 238, + 286, + 319 + ], + "blocks": [ + { + "bbox": [ + 50, + 238, + 286, + 319 + ], + "lines": [ + { + "bbox": [ + 50, + 238, + 286, + 319 + ], + "spans": [ + { + "bbox": [ + 50, + 238, + 286, + 319 + ], + "type": "table", + "html": "
MethodPose Estimation AUCTime (ms)
@5°@10°@20°
Ours Full56.472.283.5139.2
1) Ours Optimal (w/o dual-softmax)55.471.482.9102.0
2) Replace Agg. Attention to LoFTR's Trans.54.770.582.2171.4
3) Replace two-stage refine. to LoFTR's refine.54.770.982.7135.3
4) No second-stage refinement55.871.883.3138.1
5) Replace RepVGG with ResNet55.471.482.9156.2
", + "image_path": "838e5501e4a0826fb5973c1508383a04bd832c1c13e551d9d45685f8c1f689b9.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 323, + 288, + 369 + ], + "lines": [ + { + "bbox": [ + 46, + 323, + 288, + 369 + ], + "spans": [ + { + "bbox": [ + 46, + 323, + 288, + 369 + ], + "type": "text", + "content": "Table 5. Ablation Studies. The components of our method are ablated on the MegaDepth dataset for a comprehensive understanding of our method, where averaged running times for an image pair with high-resolution " + }, + { + "bbox": [ + 46, + 323, + 288, + 369 + ], + "type": "inline_equation", + "content": "1200 \\times 1200" + }, + { + "bbox": [ + 46, + 323, + 288, + 369 + ], + "type": "text", + "content": " are reported." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 378, + 288, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 378, + 288, + 403 + ], + "spans": [ + { + "bbox": [ + 46, + 378, + 288, + 403 + ], + "type": "text", + "content": "detector-free methods including LoFTR [46], TopicFM [15], PATS [32] and Aspanformer [6]." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 406, + 289, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 406, + 289, + 515 + ], + "spans": [ + { + "bbox": [ + 46, + 406, + 289, + 515 + ], + "type": "text", + "content": "Results. We adhere to the pipeline and evaluation settings of the online visual localization benchmark² to ensure fairness. As presented in Tab. 3, the proposed method achieves competitive results, taking both detector-based and detector-free methods into account. Being a method primarily geared towards efficiency, our approach can deliver results comparable to those of many accuracy-oriented methods. As depicted in Tab. 4, our method also demonstrates performance on par with the best-performing approaches." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 521, + 149, + 534 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 521, + 149, + 534 + ], + "spans": [ + { + "bbox": [ + 47, + 521, + 149, + 534 + ], + "type": "text", + "content": "4.5. Ablation Studies" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 540, + 289, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 540, + 289, + 696 + ], + "spans": [ + { + "bbox": [ + 46, + 540, + 289, + 696 + ], + "type": "text", + "content": "In this part, we conduct detailed ablation studies to analyze the effectiveness of our proposed modules with results shown in Tab. 5. 1) Without dual-softmax, our optimal model can bring huge efficiency improvement in high-resolution images. 2) In coarse feature transformation, replacing the proposed aggregated attention module with LoFTR's transformer can bring significant efficiency dropping, as well as accuracy decrease. Note that the replaced transformer is also equipped with RoPE same as ours for fair comparison. This demonstrates the efficacy of the proposed module that performing vanilla attention on aggregated features can achieve higher efficiency with even better matching accuracy. 3) Compared with using LoFTR's refinement that performs" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 240, + 548, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 240, + 548, + 373 + ], + "spans": [ + { + "bbox": [ + 304, + 240, + 548, + 373 + ], + "type": "text", + "content": "expectation on the entire correlation patch, the proposed two-stage refinement layer can bring accuracy improvement with negligible latency. We attribute this to the two-stage refinement's property that can maximize the suppression of location variance in correlation refinement. 4) Dropping the second refinement stage will lead to degraded pose accuracy with minor efficiency changes, especially on the strict AUC@5° metric. 5) Changing the backbone from reparameterized VGG [10] back to multi-branch ResNet [17] leads to decreased efficiency with similar accuracy, which demonstrates the effectiveness of our design choice for efficiency." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 382, + 384, + 395 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 382, + 384, + 395 + ], + "spans": [ + { + "bbox": [ + 306, + 382, + 384, + 395 + ], + "type": "text", + "content": "5. Conclusions" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 402, + 548, + 643 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 402, + 548, + 643 + ], + "spans": [ + { + "bbox": [ + 304, + 402, + 548, + 643 + ], + "type": "text", + "content": "This paper introduces a new semi-dense local feature matcher based on the success of LoFTR. We revisit its designs and propose several improvements for both efficiency and matching accuracy. A key observation is that performing the Transformer on the entire coarse feature map is redundant due to the similar local information, where an aggregated attention module is proposed to perform transformer on reduced tokens with significantly better efficiency and competitive performance. Moreover, a two-stage correlation layer is devised to solve the location variance problem in LoFTR's refinement design, which further brings accuracy improvements. As a result, our method can achieve " + }, + { + "bbox": [ + 304, + 402, + 548, + 643 + ], + "type": "inline_equation", + "content": "\\sim 2.5" + }, + { + "bbox": [ + 304, + 402, + 548, + 643 + ], + "type": "text", + "content": " times faster compared with LoFTR with better matching accuracy. Moreover, as a semi-dense matching method, the proposed method can achieve comparable efficiency with the recent robust sparse feature matcher LightGlue [26]. We believe this opens up the applications of our method in large-scale or latency-sensitive downstream tasks, such as image retrieval and 3D reconstruction. Please refer to the supplementary material for discussions about limitations and future works." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 656, + 548, + 704 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 656, + 548, + 704 + ], + "spans": [ + { + "bbox": [ + 304, + 656, + 548, + 704 + ], + "type": "text", + "content": "Acknowledgement The authors would like to acknowledge support from NSFC (No. 62172364) and Information Technology Center and State Key Lab of CAD&CG, Zhejiang University." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 58, + 702, + 212, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 702, + 212, + 712 + ], + "spans": [ + { + "bbox": [ + 58, + 702, + 212, + 712 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 58, + 702, + 212, + 712 + ], + "type": "text", + "content": "https://wwwVisualization.net/benchmark" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "21673" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 53, + 91, + 288, + 123 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 288, + 123 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 288, + 123 + ], + "type": "text", + "content": "[1] Sameer Agarwal, Yasutaka Furukawa, Noah Snavely, Ian Simon, Brian Curless, Steven M. Seitz, and Richard Szeliski. Building rome in a day. ICCV, 2009. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 124, + 288, + 167 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 124, + 288, + 167 + ], + "spans": [ + { + "bbox": [ + 53, + 124, + 288, + 167 + ], + "type": "text", + "content": "[2] Vassileios Balntas, Karel Lenc, Andrea Vedaldi, and Krystian Mikolajczyk. Hpatches: A benchmark and evaluation of handcrafted and learned local descriptors. In CVPR, pages 5173-5182, 2017. 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 168, + 288, + 201 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 168, + 288, + 201 + ], + "spans": [ + { + "bbox": [ + 53, + 168, + 288, + 201 + ], + "type": "text", + "content": "[3] Herbert Bay, Andreas Ess, Tinne Tuytelaars, and Luc Van Gool. Speeded-up robust features (surf). CVIU, 110 (3):346-359, 2008. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 202, + 288, + 235 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 202, + 288, + 235 + ], + "spans": [ + { + "bbox": [ + 53, + 202, + 288, + 235 + ], + "type": "text", + "content": "[4] Han Cai, Chuang Gan, and Song Han. Efficientvit: Enhanced linear attention for high-resolution low-computation visual recognition. arXiv preprint arXiv:2205.14756, 2022. 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 236, + 288, + 280 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 236, + 288, + 280 + ], + "spans": [ + { + "bbox": [ + 53, + 236, + 288, + 280 + ], + "type": "text", + "content": "[5] Hongkai Chen, Zixin Luo, Jiahui Zhang, Lei Zhou, Xuyang Bai, Zeyu Hu, Chiew-Lan Tai, and Long Quan. Learning to match features with seeded graph matching network. In ICCV, pages 6301-6310, 2021. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 280, + 288, + 334 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 280, + 288, + 334 + ], + "spans": [ + { + "bbox": [ + 53, + 280, + 288, + 334 + ], + "type": "text", + "content": "[6] Hongkai Chen, Zixin Luo, Lei Zhou, Yurun Tian, Mingmin Zhen, Tian Fang, David McKinnon, Yanghai Tsin, and Long Quan. Aspanformer: Detector-free image matching with adaptive span transformer. In ECCV, pages 20-36. Springer, 2022. 2, 5, 6, 7, 8" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 335, + 288, + 379 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 335, + 288, + 379 + ], + "spans": [ + { + "bbox": [ + 53, + 335, + 288, + 379 + ], + "type": "text", + "content": "[7] Angela Dai, Angel X Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In CVPR, pages 5828-5839, 2017. 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 380, + 288, + 412 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 380, + 288, + 412 + ], + "spans": [ + { + "bbox": [ + 53, + 380, + 288, + 412 + ], + "type": "text", + "content": "[8] Tri Dao, Daniel Y. Fu, Stefano Ermon, Atri Rudra, and Christopher Ré. FlashAttention: Fast and memory-efficient exact attention with IO-awareness. In NeurIPS, 2022. 3, 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 413, + 288, + 446 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 413, + 288, + 446 + ], + "spans": [ + { + "bbox": [ + 53, + 413, + 288, + 446 + ], + "type": "text", + "content": "[9] Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. Superpoint: Self-supervised interest point detection and description. CVPRW, 2018. 2, 6, 7, 8" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 447, + 288, + 490 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 447, + 288, + 490 + ], + "spans": [ + { + "bbox": [ + 48, + 447, + 288, + 490 + ], + "type": "text", + "content": "[10] Xiaohan Ding, Xiangyu Zhang, Ningning Ma, Jungong Han, Guiguang Ding, and Jian Sun. Repvgg: Making vgg-style convnets great again. In CVPR, pages 13733-13742, 2021. 3, 6, 8" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 491, + 288, + 535 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 491, + 288, + 535 + ], + "spans": [ + { + "bbox": [ + 48, + 491, + 288, + 535 + ], + "type": "text", + "content": "[11] Mihai Dusmanu, Ignacio Rocco, Tomas Pajdla, Marc Pollefeys, Josef Sivic, Akihiko Torii, and Torsten Sattler. D2-net: A trainable cnn for joint description and detection of local features. In CVPR, pages 8092-8101, 2019. 2, 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 536, + 288, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 536, + 288, + 568 + ], + "spans": [ + { + "bbox": [ + 48, + 536, + 288, + 568 + ], + "type": "text", + "content": "[12] Patrick Ebel, Anastasiia Mishchuk, Kwang Moo Yi, Pascal Fua, and Eduard Trulls. Beyond cartesian representations for local descriptors. In ICCV, pages 253-262, 2019. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 569, + 288, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 569, + 288, + 612 + ], + "spans": [ + { + "bbox": [ + 48, + 569, + 288, + 612 + ], + "type": "text", + "content": "[13] Johan Edstedt, Ioannis Athanasiadis, Mårten Wadenbäck, and Michael Felsberg. Dkm: Dense kernelized feature matching for geometry estimation. In CVPR, pages 17765-17775, 2023. 2, 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 613, + 288, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 613, + 288, + 656 + ], + "spans": [ + { + "bbox": [ + 48, + 613, + 288, + 656 + ], + "type": "text", + "content": "[14] Johan Edstedt, Qiyu Sun, Georg Bokman, Mårten Wadenback, and Michael Felsberg. Roma: Revisiting robust losses for dense feature matching. arXiv preprint arXiv:2305.15404, 2023. 2, 3, 6" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 658, + 288, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 288, + 690 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 288, + 690 + ], + "type": "text", + "content": "[15] Khang Truong Giang, Soohwan Song, and Sung-Guk Jo. Top-icfm: Robust and interpretable topic-assisted feature matching. In AAAI, 2023. 2, 3, 5, 6, 8" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 48, + 691, + 288, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 288, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 288, + 712 + ], + "type": "text", + "content": "[16] Stephen Hausler, Sourav Garg, Ming Xu, Michael Milford, and Tobias Fischer. Patch-netvlad: Multi-scale fusion of" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 547, + 713 + ], + "type": "list", + "angle": 0, + "index": 36, + "blocks": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "type": "text", + "content": "locally-global descriptors for place recognition. CVPR, pages 14136-14147, 2021. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 96, + 547, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 96, + 547, + 128 + ], + "spans": [ + { + "bbox": [ + 308, + 96, + 547, + 128 + ], + "type": "text", + "content": "[17] Kaiming He, X. Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. CVPR, pages 770-778, 2015. 3, 8" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 129, + 547, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 129, + 547, + 162 + ], + "spans": [ + { + "bbox": [ + 308, + 129, + 547, + 162 + ], + "type": "text", + "content": "[18] Xingyi He, Jiaming Sun, Yifan Wang, Sida Peng, Qixing Huang, Hujun Bao, and Xiaowei Zhou. Detector-free structure from motion. In CVPR, 2024. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 163, + 545, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 163, + 545, + 196 + ], + "spans": [ + { + "bbox": [ + 308, + 163, + 545, + 196 + ], + "type": "text", + "content": "[19] Angelos Katharopoulos, Apoorv Vyas, Nikolaos Pappas, and Francois Fleuret. Transformers are rnns: Fast autoregressive transformers with linear attention. In ICML, 2020. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 198, + 547, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 198, + 547, + 228 + ], + "spans": [ + { + "bbox": [ + 308, + 198, + 547, + 228 + ], + "type": "text", + "content": "[20] Nikita Kitaev, Lukasz Kaiser, and Anselm Levskaya. Reformer: The efficient transformer. ArXiv, abs/2001.04451, 2020.3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 230, + 545, + 263 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 230, + 545, + 263 + ], + "spans": [ + { + "bbox": [ + 308, + 230, + 545, + 263 + ], + "type": "text", + "content": "[21] Axel Barroso Laguna, Edgar Riba, Daniel Ponsa, and Krystian Mikolajczyk. Key.net: Keypoint detection by handcrafted and learned cnn filters. ICCV, pages 5835-5843, 2019. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 265, + 547, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 265, + 547, + 330 + ], + "spans": [ + { + "bbox": [ + 308, + 265, + 547, + 330 + ], + "type": "text", + "content": "[22] Benjamin Lefaudeau, Francisco Massa, Diana Liskovich, Wenhan Xiong, Vittorio Caggiano, Sean Naren, Min Xu, Jieru Hu, Marta Tintore, Susan Zhang, Patrick Labatut, and Daniel Haziza. xformers: A modular and hackable transformer modelling library. https://github.com/facebookresearch/xformers, 2022.3.6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 331, + 547, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 331, + 547, + 363 + ], + "spans": [ + { + "bbox": [ + 308, + 331, + 547, + 363 + ], + "type": "text", + "content": "[23] Xinghui Li, Kai Han, Shuda Li, and Victor Prisacariu. Dualresolution correspondence networks. In NeurlPS, 2020. 2, 6, 7" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 365, + 545, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 365, + 545, + 397 + ], + "spans": [ + { + "bbox": [ + 308, + 365, + 545, + 397 + ], + "type": "text", + "content": "[24] Zhengqi Li and Noah Snavely. Megadepth: Learning single-view depth prediction from internet photos. In CVPR, pages 2041-2050, 2018. 6" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 399, + 545, + 432 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 399, + 545, + 432 + ], + "spans": [ + { + "bbox": [ + 308, + 399, + 545, + 432 + ], + "type": "text", + "content": "[25] Philipp Lindenberger, Paul-Edouard Sarlin, Viktor Larsson, and Marc Pollefeys. Pixel-perfect structure-from-motion with featuremetric refinement. ICCV, pages 5967-5977, 2021. 1" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 433, + 545, + 465 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 433, + 545, + 465 + ], + "spans": [ + { + "bbox": [ + 308, + 433, + 545, + 465 + ], + "type": "text", + "content": "[26] Philipp Lindenberger, Paul-Edouard Sarlin, and Marc Pollefeys. LightGlue: Local Feature Matching at Light Speed. In ICCV, 2023. 1, 2, 4, 6, 8" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 466, + 545, + 488 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 466, + 545, + 488 + ], + "spans": [ + { + "bbox": [ + 308, + 466, + 545, + 488 + ], + "type": "text", + "content": "[27] G LoweDavid. Distinctive image features from scale-invariant keypoints. IJCV, 2004. 1, 2" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 308, + 489, + 545, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 489, + 545, + 533 + ], + "spans": [ + { + "bbox": [ + 308, + 489, + 545, + 533 + ], + "type": "text", + "content": "[28] Zixin Luo, Lei Zhou, Xuyang Bai, Hongkai Chen, Jiahui Zhang, Yao Yao, Shiwei Li, Tian Fang, and Long Quan. Aslfeat: Learning local features of accurate shape and localization. In CVPR, pages 6589-6598, 2020. 2" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 308, + 534, + 545, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 534, + 545, + 567 + ], + "spans": [ + { + "bbox": [ + 308, + 534, + 545, + 567 + ], + "type": "text", + "content": "[29] Anastasiia Mishchuk, Dmytro Mishkin, Filip Radenovic, and Jiri Matas. Working hard to know your neighbor's margins: Local descriptor learning loss. NeurIPS, 30, 2017. 2" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 308, + 568, + 545, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 568, + 545, + 600 + ], + "spans": [ + { + "bbox": [ + 308, + 568, + 545, + 600 + ], + "type": "text", + "content": "[30] Raul Mur-Artal and Juan D. Tardós. Orb-slam2: An open-source slam system for monocular, stereo, and rgb-d cameras. TR, 33:1255-1262, 2016. 1" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 308, + 601, + 545, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 601, + 545, + 634 + ], + "spans": [ + { + "bbox": [ + 308, + 601, + 545, + 634 + ], + "type": "text", + "content": "[31] Raul Mur-Artal, José M. M. Montiel, and Juan D. Tardós. Orb-slam: A versatile and accurate monocular slam system. TR, 31:1147-1163, 2015. 1" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 308, + 635, + 545, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 635, + 545, + 679 + ], + "spans": [ + { + "bbox": [ + 308, + 635, + 545, + 679 + ], + "type": "text", + "content": "[32] Junjie Ni, Yijin Li, Zhaoyang Huang, Hongsheng Li, Hujun Bao, Zhaopeng Cui, and Guofeng Zhang. Pats: Patch area transportation with subdivision for local feature matching. CVPR, pages 17776-17786, 2023. 2, 8" + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 308, + 680, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 680, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 680, + 545, + 713 + ], + "type": "text", + "content": "[33] Maxime Oquab, Timothée Darcet, Theo Moutakanni, Huy V. Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, Russell" + } + ] + } + ], + "index": 35 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "21674" + } + ] + } + ], + "index": 37 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 66, + 72, + 288, + 137 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 72, + 288, + 137 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 288, + 137 + ], + "type": "text", + "content": "Howes, Po-Yao Huang, Hu Xu, Vasu Sharma, Shang-Wen Li, Wojciech Galuba, Mike Rabbat, Mido Assran, Nicolas Ballas, Gabriel Synnaeve, Ishan Misra, Herve Jegou, Julien Mairal, Patrick Labatut, Armand Joulin, and Piotr Bojanowski. Dinov2: Learning robust visual features without supervision, 2023. 7" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 139, + 288, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 139, + 288, + 172 + ], + "spans": [ + { + "bbox": [ + 48, + 139, + 288, + 172 + ], + "type": "text", + "content": "[34] Jérôme Revaud, César Roberto de Souza, M. Humenberger, and Philippe Weinzaepfel. R2d2: Reliable and repeatable detector and descriptor. In NeurIPS, 2019. 2, 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 172, + 288, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 172, + 288, + 205 + ], + "spans": [ + { + "bbox": [ + 48, + 172, + 288, + 205 + ], + "type": "text", + "content": "[35] Ignacio Rocco, Mircea Cimpoi, Relja Arandjelović, Akihiko Torii, Tomas Pajdla, and Josef Sivic. Neighbourhood consensus networks. *NeurIPS*, 31, 2018. 2, 4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 205, + 288, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 205, + 288, + 239 + ], + "spans": [ + { + "bbox": [ + 48, + 205, + 288, + 239 + ], + "type": "text", + "content": "[36] Ignacio Rocco, Relja Arandjelović, and Josef Sivic. Efficient neighbourhood consensus networks via submanifold sparse convolutions. In ECCV, pages 605-621. Springer, 2020. 2, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 239, + 288, + 272 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 239, + 288, + 272 + ], + "spans": [ + { + "bbox": [ + 48, + 239, + 288, + 272 + ], + "type": "text", + "content": "[37] Edward Rosten and Tom Drummond. Machine learning for high-speed corner detection. In ECCV, pages 430-443. Springer, 2006. 1, 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 272, + 288, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 272, + 288, + 315 + ], + "spans": [ + { + "bbox": [ + 48, + 272, + 288, + 315 + ], + "type": "text", + "content": "[38] Paul-Edouard Sarlin, César Cadena, Roland Y. Siegwart, and Marcin Dymczyk. From coarse to fine: Robust hierarchical localization at large scale. CVPR, pages 12708–12717, 2019. 1, 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 316, + 288, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 316, + 288, + 358 + ], + "spans": [ + { + "bbox": [ + 48, + 316, + 288, + 358 + ], + "type": "text", + "content": "[39] Paul-Edouard Sarlin, Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. SuperGlue: Learning feature matching with graph neural networks. In CVPR, 2020. 1, 2, 4, 5, 6, 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 360, + 288, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 360, + 288, + 415 + ], + "spans": [ + { + "bbox": [ + 48, + 360, + 288, + 415 + ], + "type": "text", + "content": "[40] Paul-Edouard Sarlin, Ajaykumar Unagar, Måns Larsson, Hugo Germain, Carl Toft, Victor Larsson, Marc Pollefeys, Vincent Lepetit, Lars Hammarstrand, Fredrik Kahl, and Torsten Sattler. Back to the Feature: Learning Robust Camera Localization from Pixels to Pose. In CVPR, 2021. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 415, + 288, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 415, + 288, + 470 + ], + "spans": [ + { + "bbox": [ + 48, + 415, + 288, + 470 + ], + "type": "text", + "content": "[41] Torsten Sattler, Will Maddern, Carl Toft, Akihiko Torii, Lars Hammarstrand, Erik Stenberg, Daniel Safari, Masatoshi Okutomi, Marc Pollefeys, Josef Sivic, et al. Benchmarking 6dof outdoor visual localization in changing conditions. In CVPR, pages 8601-8610, 2018. 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 471, + 288, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 471, + 288, + 514 + ], + "spans": [ + { + "bbox": [ + 48, + 471, + 288, + 514 + ], + "type": "text", + "content": "[42] Nikolay Savinov, Akihito Seki, Lubor Ladicky, Torsten Sattler, and Marc Pollefeys. Quad-networks: unsupervised learning to rank for interest point detection. In CVPR, pages 1822-1830, 2017. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 514, + 288, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 514, + 288, + 536 + ], + "spans": [ + { + "bbox": [ + 48, + 514, + 288, + 536 + ], + "type": "text", + "content": "[43] Johannes L. Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In CVPR, 2016. 1, 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 537, + 288, + 581 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 537, + 288, + 581 + ], + "spans": [ + { + "bbox": [ + 48, + 537, + 288, + 581 + ], + "type": "text", + "content": "[44] Yan Shi, Jun-Xiong Cai, Yoli Shavit, Tai-Jiang Mu, Wensen Feng, and Kai Zhang. Clustergnn: Cluster-based coarse-to-fine graph neural network for efficient feature matching. In CVPR, pages 12517-12526, 2022. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 582, + 288, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 582, + 288, + 613 + ], + "spans": [ + { + "bbox": [ + 48, + 582, + 288, + 613 + ], + "type": "text", + "content": "[45] Jianlin Su, Yu Lu, Shengfeng Pan, Bo Wen, and Yunfeng Liu. Roformer: Enhanced transformer with rotary position embedding. ArXiv, abs/2104.09864, 2021. 4" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 614, + 288, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 614, + 288, + 658 + ], + "spans": [ + { + "bbox": [ + 48, + 614, + 288, + 658 + ], + "type": "text", + "content": "[46] Jiaming Sun, Zehong Shen, Yuang Wang, Hujun Bao, and Xiaowei Zhou. Loftr: Detector-free local feature matching with transformers. In CVPR, pages 8922-8931, 2021. 1, 2, 4, 5, 6, 7, 8" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 658, + 288, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 288, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 288, + 712 + ], + "type": "text", + "content": "[47] Hajime Taira, Masatoshi Okutomi, Torsten Sattler, Mircea Cimpoi, Marc Pollefeys, Josef Sivic, Tomas Pajdla, and Akihiko Torii. Inloc: Indoor visual localization with dense matching and view synthesis. In CVPR, pages 7199-7209, 2018. 7" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 398 + ], + "type": "list", + "angle": 0, + "index": 25, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 95 + ], + "type": "text", + "content": "[48] Shitao Tang, Jiahui Zhang, Siyu Zhu, and Ping Tan. Quadtree attention for vision transformers. *ICLR*, 2022. 2, 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 96, + 545, + 129 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 96, + 545, + 129 + ], + "spans": [ + { + "bbox": [ + 307, + 96, + 545, + 129 + ], + "type": "text", + "content": "[49] Yurun Tian, Bin Fan, and Fuchao Wu. L2-net: Deep learning of discriminative patch descriptor in euclidean space. In CVPR, pages 661-669, 2017. 1, 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 129, + 545, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 129, + 545, + 173 + ], + "spans": [ + { + "bbox": [ + 307, + 129, + 545, + 173 + ], + "type": "text", + "content": "[50] Yurun Tian, Xin Yu, Bin Fan, Fuchao Wu, Huub Heijnen, and Vassileios Balntas. Sosnet: Second order similarity regularization for local descriptor learning. In CVPR, pages 11016-11025, 2019. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 175, + 545, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 175, + 545, + 217 + ], + "spans": [ + { + "bbox": [ + 307, + 175, + 545, + 217 + ], + "type": "text", + "content": "[51] Yurun Tian, Vassileios Balntas, Tony Ng, Axel Barroso-Laguna, Yiannis Demiris, and Krystian Mikolajczyk. D2d: Keypoint extraction with describe to detect approach. In ACCV, 2020. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 220, + 545, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 220, + 545, + 252 + ], + "spans": [ + { + "bbox": [ + 307, + 220, + 545, + 252 + ], + "type": "text", + "content": "[52] Prune Truong, Martin Danelljan, Luc Van Gool, and Radu Timofte. Learning accurate dense correspondences and when to trust them. In CVPR, pages 5714-5724, 2021. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 254, + 545, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 254, + 545, + 285 + ], + "spans": [ + { + "bbox": [ + 307, + 254, + 545, + 285 + ], + "type": "text", + "content": "[53] Michal Tyszkiiewicz, Pascal Fua, and Eduard Trulls. Disk: Learning local features with policy gradient. NeurIPS, 2020. 4, 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 287, + 545, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 287, + 545, + 330 + ], + "spans": [ + { + "bbox": [ + 307, + 287, + 545, + 330 + ], + "type": "text", + "content": "[54] Ashish Vaswani, Noam M. Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In NeurIPS, 2017. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 332, + 545, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 332, + 545, + 365 + ], + "spans": [ + { + "bbox": [ + 307, + 332, + 545, + 365 + ], + "type": "text", + "content": "[55] Qing Wang, Jiaming Zhang, Kailun Yang, Kunyu Peng, and Rainer Stiefelhagen. Matchformer: Interleaving attention in transformers for feature matching. In ACCV, 2022. 2, 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 365, + 545, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 365, + 545, + 398 + ], + "spans": [ + { + "bbox": [ + 307, + 365, + 545, + 398 + ], + "type": "text", + "content": "[56] Sinong Wang, Belinda Z. Li, Madian Khabsa, Han Fang, and Hao Ma. Linformer: Self-attention with linear complexity. ArXiv, abs/2006.04768, 2020. 3" + } + ] + } + ], + "index": 24 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "21675" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Efficient Meshflow and Optical Flow Estimation from Event Cameras/141a2ee1-d1ef-416d-8c61-bc462b1074b9_content_list.json b/2024/Efficient Meshflow and Optical Flow Estimation from Event Cameras/141a2ee1-d1ef-416d-8c61-bc462b1074b9_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..2bb0c0d7ca34ae972fed249d3bcffc0cac7cfbc1 --- /dev/null +++ b/2024/Efficient Meshflow and Optical Flow Estimation from Event Cameras/141a2ee1-d1ef-416d-8c61-bc462b1074b9_content_list.json @@ -0,0 +1,2180 @@ +[ + { + "type": "text", + "text": "Efficient Meshflow and Optical Flow Estimation from Event Cameras", + "text_level": 1, + "bbox": [ + 133, + 130, + 836, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Xinglong Luo $^{1,4*}$ , Ao Luo $^{2,4*}$ , Zhengning Wang $^{1\\dagger}$ , Chunyu Lin $^{3}$ , Bing Zeng $^{1}$ , and Shuaicheng Liu $^{1,4\\dagger}$ $^{1}$ University of Electronic Science and Technology of China \n $^{2}$ Southwest Jiaotong University \n $^{3}$ Beijing Jiaotong University \n $^{4}$ Megvii Technology \n{luoboom@std.,zhengning.wang@,eezeng@,liushuaicheng@}uestc.edu.cn \naoluo@swjtu.edu.cn cylin@bjtu.edu.cn", + "bbox": [ + 101, + 178, + 875, + 268 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 304, + 313, + 320 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this paper, we explore the problem of event-based meshflow estimation, a novel task that involves predicting a spatially smooth sparse motion field from event cameras. To start, we generate a large-scale High-Resolution Event Meshflow (HREM) dataset, which showcases its superiority by encompassing the merits of high resolution at $1280 \\times 720$ , handling dynamic objects and complex motion patterns, and offering both optical flow and meshflow labels. These aspects have not been fully explored in previous works. Besides, we propose Efficient Event-based MeshFlow (EEMFlow) network, a lightweight model featuring a specially crafted encoder-decoder architecture to facilitate swift and accurate meshflow estimation. Furthermore, we upgrade EEMFlow network to support dense event optical flow, in which a Confidence-induced Detail Completion (CDC) module is proposed to preserve sharp motion boundaries. We conduct comprehensive experiments to show the exceptional performance and runtime efficiency $(39 \\times$ faster) of our EEMFlow model compared to recent state-of-the-art flow methods. Our code is available at https://github.com/boomluo02/EEMFlow.", + "bbox": [ + 75, + 335, + 473, + 652 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 670, + 209, + 686 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Meshflow, a spatially smooth sparse motion field, represents motion vectors exclusively at mesh vertices [28, 49], which has been widely applied in various vision applications, such as image alignment [29, 30, 34], video stabilization [26, 27, 45] and high dynamic range (HDR) imaging [31, 46]. This motion representation combines the benefits of optical flow and global homography, effectively reducing redundancy in motion information and computational costs, while also accommodating non-rigid motions beyond single-plane movements. However, meshflow estimation on RGB images often encounter challenges under scenarios such as low-light and rapid motions. This is due", + "bbox": [ + 75, + 695, + 470, + 878 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/6a7187498c849c3a671e8f295bb00308fee81606ff12877504673bf926e9b91d.jpg", + "image_caption": [ + "Figure 1. Comparison of computational overhead and accuracy metrics. The x-axis represents inference time, while the y-axis corresponds to the end-point error. The size of each circle indicates the number of model parameters. Lower values for all metrics are considered better." + ], + "image_footnote": [], + "bbox": [ + 501, + 303, + 893, + 531 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "to the loss of fine image texture details and motion blurs.", + "bbox": [ + 498, + 622, + 875, + 638 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In contrast, event cameras are well-suited for motion estimation under such situations [7, 11]. Equipped with bio-inspired vision sensors, event cameras can generate sequences of events with microsecond accuracy triggered by changes in log intensity. In particular, when a change is detected in a pixel, the camera returns an event in the form $e = (x,y,t,p)$ immediately, where $x,y$ stands for the spatial location, $t$ refers to the timestamp in microseconds, and $p$ is the polarity of the change, either positive or negative. The advantages of high temporal resolution and high dynamic range make event cameras highly effective for analyzing dynamic scenes.", + "bbox": [ + 496, + 641, + 890, + 821 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this work, we study a new problem that estimates meshflow from event camera data. To start, we create a large-scale High-Resolution Event Meshflow (HREM) dataset, which contains 20k train and 8k test samples. We build 100 virtual scenes in Blender to render the dataset,", + "bbox": [ + 496, + 825, + 893, + 898 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Equal contribution. † Corresponding Author.", + "bbox": [ + 99, + 886, + 349, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "19198", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "which can provide accurate physically-based events along with dense meshflow label pairs. Based on the dataset, we further propose an Efficient Event-based MeshFlow (EEMFlow) network to estimate high-resolution meshflow from event data. Unlike recent flow networks relying on recurrent refinement structure [11, 17, 42], our network is developed on an encoder-decoder architecture with multi-scale global optimizing scheme, which can produce full-resolution meshflow with minimal computational overhead. Specifically, our EEMFlow achieves efficiency by employing the lightweight encoder, building cost volume with dilated feature correlation, and using group shuffle convolutions during decoding. We select recent top-ranked flow networks, including ERAFT [11], SK-Flow [41], KPAFlow [32], and FlowFormer [15], which is trained on our HREM dataset. The results show that our approach achieves state-of-the-art performance while maintaining a fast inference speed of 142.9 FPS, outperforming previous works by a relatively large margin (see Fig. 1).", + "bbox": [ + 76, + 90, + 472, + 380 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Additionally, we empirically demonstrate that the proposed new pipeline has the capability to effectively handle various motion patterns. Its lightweight design and runtime efficiency further contribute significantly to the field of optical flow estimation. Specifically, we refine the optical flow progressively during decoding using the coarse to fine residual approach. A Confidence-induced Detail Completion (CDC) module is proposed to preserve motion boundary details during flow upsampling. We also perform comparative experiments with recent event flow networks [21, 25, 33] to illustrate its superiority. The enhanced flow network is referred to as EEMFlow+, demonstrating impressive performance on the reputable DSEC [11] online test benchmark, with the fastest inference speed reaching 39.2 FPS. Our contributions are summarized as:", + "bbox": [ + 76, + 393, + 472, + 621 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We build HREM, the first event-based meshflow dataset, superior in the highest resolution at $1280 \\times 720$ , dynamic scenes, complex motion patterns, as well as physically correct accurate events paired with meshflow and optical flow labels.", + "- We propose EEMFlow, which achieves SOTA performances when compared to top-ranked optical flow networks when trained on our meshflow dataset. Moreover, it achieves inference speed of 142.9 FPS, which is 25.5 to 38.7 times faster than compared methods.", + "- We propose CDC, a confidence-induced detail completion module that empowers EEMFlow to make a meaningful contribution to the optical flow community. The upgraded model achieving SOTA performance when compared to representative methods, while also boasting the fastest inference speed to date." + ], + "bbox": [ + 76, + 643, + 470, + 898 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 89, + 642, + 106 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Image-based Meshflow Warping", + "text_level": 1, + "bbox": [ + 500, + 114, + 784, + 131 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Meshflow is a lightweight and spatially smooth sparse motion field with motions only located at mesh vertices [28]. Meshflow is different from dense optical flow, where optical flow estimates motions of every pixel of an image while meshflow only concentrates on the global motion, rejecting motions of any dynamic contents. Meshflow is also different from a global homography, where local motions from nonplanar depth variations can be well reflected. Mesh-based methods proofs to be effective in various applications, such as high dynamic range (HDR) imaging [46], burst image denoising [47], video denoising [38], image/video stitching [23, 35] and video stabilization [45]. It is worth noting that directly downsampling the optical flow may yield flow fields of the same resolution as meshflow, but they differ significantly from meshflow and perform poorly in terms of warping effects [49]. Direct downsampling ignores motion outliers from different dynamic objects and the global consistency, which can be effectively addressed by meshflow through motion propagation.", + "bbox": [ + 496, + 137, + 890, + 426 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Event-based Optical Flow Estimation", + "text_level": 1, + "bbox": [ + 500, + 435, + 823, + 450 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Optical flow estimation from event cameras has received significant attention in recent years. Early approaches, such as [1], could only estimate optical flows at the regions where events are triggered. Recently, deep methods can estimate optical flows from event data, even for the regions without triggered events. For example, EV-FlowNet [52] learns event and flow labels in a self-supervised manner by minimizing photometric distances of grey images acquired by DAVIS [2]. Various event representations, including EST [8] and Matrix-LSTM [3], have been explored, and different network structures, such as SpikeFlowNet [19], LIF-EV-FlowNet [13], STE-FlowNet [5], Li et al. [22], ERAFT [11], Yang et al. [48], EVA-Flow [50], ADMFlow [33], E-FlowFormer [21], and TMA [25] have been proposed to improve performances. Some methods even use both events and images as input for flow estimation, such as Fusion-FlowNet [20], Pan et al. [36], DCEI-Flow [43] and RPEFlow [44]. In this work, we study a new problem of event-based meshflow estimation, proposing an efficient event-based meshflow network.", + "bbox": [ + 496, + 458, + 890, + 760 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.3. Event-based Optical Flow Dataset", + "text_level": 1, + "bbox": [ + 500, + 771, + 797, + 787 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Early works synthesize events by thresholding rgb images [18] and applying interpolation for high framerate [9]. However, the timestamp of synthesized events is inaccurate, let alone interpolation artifacts. The DAVIS event camera [2] can capture both images and real events, resulting in some event datasets: DVSFLOW [24], MVSEC [53] and DSEC [10], based on which EV-Flownet [52] and ER", + "bbox": [ + 496, + 794, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "19199", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 1 + }, + { + "type": "table", + "img_path": "images/a7e43710e4ec54d14b940282a4ae39839e0fbd9d4c0df1fc4b5e9679d5bf3acb.jpg", + "table_caption": [ + "Table 1. Comparison of available datasets. 'Dense OF' and Meshflow' indicate whether the dataset has dense optical flow labels and meshflow labels." + ], + "table_footnote": [], + "table_body": "
DatasetResolutionDynamic ObjectsExtreme ConditionsDense OFMeshflowMotion Pattern
DVSFLOW [24]180×240XXXXRotation
MVSEC [53]260×346X✓XXDrone
DSEC [10]480×640X✓XXCar
MDR [33]260×346XX✓XCar
BlinkFlow [21]480×640✓X✓XRandom
Ekubric [44]720×1280✓X✓XFalling
HREM(Ours)720×1280✓✓✓✓Random
", + "bbox": [ + 83, + 107, + 885, + 208 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "AFT [11] compute sparse optical flow. In this way, however, flows can only locate on sparse event regions [43]. Recently, Luo et al. [33] proposed to render the MDR dataset from graphics, but it only contains static scenes. Wan et al. [44] synthesized the Ekubric dataset based on the Kubirc toolbox [12], which only includes a single falling motion pattern. Li et al. [21] considered non-rigid motions simulating dancing but with lower resolution and without extreme condition scenarios. None of the aforementioned datasets involve the estimation of meshflow. In this work, we render a comprehensive dataset that can support both meshflow and optical flow estimation, with a higher resolution, rich dynamic scenes, as well as extreme conditions like relatively low light and motion blur, as shown in Table 1.", + "bbox": [ + 75, + 227, + 472, + 440 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Algorithm", + "text_level": 1, + "bbox": [ + 76, + 455, + 191, + 473 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. High-Resolution Event Meshflow Dataset", + "text_level": 1, + "bbox": [ + 76, + 482, + 429, + 498 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We create the high-resolution event meshflow dataset (HREM), consisting of 100 virtual scenes that accurately mimic real-world environments, both indoors and outdoors. In these scenes, we put dynamic objects to simulate intricate object motions. Camera is programmed to track these movements, ensuring a realistic portrayal of motion. The Blender rendering engine was utilized to create high frame rate videos and dense optical flow labels. For event data generation, we employed three advanced simulators: ESIM [37], V2E [14], and DVS-Voltmeter [24]. The simulator that provided the highest contrast in warped events images was chosen for our dataset. We also process the dense optical flow using motion propagation and median filters, enabling the generation of meshflow labels. Our comparisons with existing datasets is shown in Fig. 2, emphasize the superiority of HREM in high-resolution, dynamic scenes, complex motion patterns, and comprehensive labeling. Following [28], we generate meshflow from dense optical flow, as depicted in Fig. 3.", + "bbox": [ + 75, + 506, + 468, + 792 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Motion Propagation. Given a dense optical flow $\\mathbf{F}$ , we place a uniform mesh of $16 \\times 16$ regular cells on its image plane and then select the motion of the middle point $p$ in each cell as the local motion $v_{p}$ . Since the vertices of the mesh near point $p$ should have a similar motion to $v_{p}$ , we define a rectangle that covers $3 \\times 3$ cells centered at $p$ , and assign to all the vertices within the rectangle, ensuring the", + "bbox": [ + 75, + 795, + 470, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "propagation of the local motion $v_{p}$ across the image plane.", + "bbox": [ + 496, + 227, + 883, + 243 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Median Filters. The local motions of the middle points in all cells are propagated to their nearby mesh vertices, resulting in each vertex potentially receiving multiple motion vectors. To select the most appropriate motion vector for a given vertex, we apply a median filter $f_{1}$ to filter the candidate motions. The response of the filter is then assigned to the corresponding vertex. The median filter is a widely used technique in optical flow estimation and has been shown to produce high-quality flow estimates [39]. Therefore, we use the median filter for sparse motion regularization. However, due to dynamic objects, the motion field may contain noise and needs to be spatially smoothed. To address this issue, we apply another median filter $f_{2}$ that covers $3 \\times 3$ cells neighborhood to suppress the noise in the motion field. This second median filter produces a spatially-smooth sparse motion field, which is what we called as meshflow. Ultimately, we use the meshflow (generally $16 \\times 16$ ) as the label, which contains global motion information. However, we upsample the meshflow to the full image resolution for intuitive display and alignment applications.", + "bbox": [ + 496, + 243, + 890, + 546 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Estimation for Meshflow and Optical Flow", + "text_level": 1, + "bbox": [ + 500, + 556, + 862, + 571 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Following [11] which estimates optical flow from two consecutive event sequences, we estimate the meshflow $MF_{k\\to k + 1}$ and optical flow $F_{k\\rightarrow k + 1}$ from event sequences $E(t_{k - 1},t_k)$ and $E(t_{k},t_{k + 1})$ , and overall architecture of our network is shown in Fig. 4. Fellow [54], we convert the inputs $E(t_{k - 1},t_k)$ and $E(t_{k},t_{k + 1})$ to the 3D volumes $V_{k - 1\\to k}$ and $V_{k\\to k + 1}$", + "bbox": [ + 498, + 580, + 890, + 686 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2.1 Efficient Event-based Meshflow Network", + "text_level": 1, + "bbox": [ + 500, + 694, + 839, + 709 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We propose EEMFlow to directly output results of the same resolution as the ground truth meshflow $MF_{GT}$ for supervised regression, fully leveraging the advantage about low-parameter and high-motion-information of meshflow.", + "bbox": [ + 496, + 719, + 893, + 779 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Overall Structure of Meshflow Estimation. Since meshflow focuses more on global large motion rather than local detailed motion, the EEMFlow we designed does not require excessively deep network layers or refinement operations from coarse to fine. Firstly, EEMFlow employs a three-level pyramid feature encoder to extracting features $(V_{t_{k - 1}}^i)^N$ and $(V_{t_k}^i)^N$ from $V_{k - 1\\rightarrow k}$ and $V_{k\\rightarrow k + 1}$ , the convolutional layers within the $i$ -th level share weights for", + "bbox": [ + 496, + 780, + 893, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "19200", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/33b37661ca91328013c7b981dcb7665f1f604e5388fa28542ea364690667490e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 181, + 89, + 259, + 136 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/524953267b60898dd498f1c6910ed2b2b00942ba01d6f74a061deaf9306589f8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 263, + 90, + 366, + 136 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/98485f598180dc52b2da7a9ca4343b5785631b13062c9f91a090a92b91ced924.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 367, + 90, + 470, + 136 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/296bbbf2038ecd81415d972fd646e69c5526cfb5e6b3cadf82def1cd576156b3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 472, + 90, + 576, + 136 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/ae38838da1ba53a93b669380d73b4aa8ea445d23da337d69163c46ab90d30b32.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 576, + 90, + 679, + 136 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/3c64997abcf1dc0f4ae722197050cf63da35f51e3c074437e871150fbdb1ed83.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 679, + 90, + 782, + 136 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/42824c3e67f0fe77fdb1b91057b4ecd6abf4dad327616be3e97262a3143cbe18.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 785, + 90, + 888, + 136 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/e3085ee4e352aee8dfaa35cb0b6c6c4fef12141db8871d3876f8bd257503e64b.jpg", + "image_caption": [ + "(a)", + "(b)", + "(c)", + "=" + ], + "image_footnote": [], + "bbox": [ + 101, + 138, + 179, + 185 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/ec1bef5238fcd6e20a8974d3a4399269eb7d129f9ce878c8f7e5d0e6180c9115.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 181, + 138, + 261, + 185 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/4f5da1305ea57a74698431b28d27e8cf3023b3cf28d649bc03958a6c9bc9b908.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 264, + 138, + 366, + 185 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/b6cb352a7624bacec943dac5d04832baf37279a5e000eef99ca709dd54233b98.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 367, + 138, + 470, + 185 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/34114591dd5335d2594b39256e06abe090f974ce36a0102e11e1d941d4c736a1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 472, + 138, + 575, + 185 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/4c041d142d57b2125cab9c9617fefffe5632cd0ed2e821e70e25c0171efa2067.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 576, + 138, + 679, + 185 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/a9eb40e1bcf52234e02028eb35c8e30c6d3610f3ee23d5ca0fbf3baa4a7ba829.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 679, + 138, + 784, + 185 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/c17957b13cb923135c3fc2b38d2b70f18641160ed975f17e25faac8c5f3459f8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 785, + 138, + 888, + 185 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/5403db35b83329e7d27bedc5cd657e5aa7c9f3c1153c99a8c9fbb4d4c4397a44.jpg", + "image_caption": [ + "MVSEC", + "DSEC", + "Figure 2. Examples from MVSEC [53], DSEC [10] and our dataset. (a): Event data, (b): Optical flow, (c): Upsampled meshflow. The meshflows of MVSEC and DSEC contain a significant number of ineffective areas due to the sparsity of optical flow. In contrast, our meshflow provides a complete global motion field. Best viewed on a color screen in high resolution." + ], + "image_footnote": [], + "bbox": [ + 181, + 186, + 261, + 234 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/05f11b9af16dd20ed103d3731b84580f7f0b85504c11864375085c44485527ff.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 264, + 186, + 367, + 234 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/4d13999406ad78d44cbf5994bbe6f7ca0dd8b791620e4eded31b81cafe09cf15.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 369, + 186, + 470, + 234 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/6f29cf738b64611c320f3be1c33abf69738dd92940d35b5b5c8f86f71f6e2de4.jpg", + "image_caption": [ + "HREM(Ours)" + ], + "image_footnote": [], + "bbox": [ + 473, + 186, + 575, + 234 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/bad93513d554583ceb6ec9093f764d952a6e20c7d1b224524ddfa13c6d4cb777.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 576, + 186, + 679, + 234 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/e3e2e8af64738c9372cfb15a4190afefe67c589c7d7bd3a9f5fa27e68f4284c2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 681, + 186, + 784, + 234 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/ca48efec2a148cbd2aa73d2cda1d2ff2b19e34631b7dfc3310cb19ada99f8c16.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 785, + 186, + 888, + 234 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/41c32b84fab9b46c89f32de5b57f75dd06a3d3b6b99de90272e0161bdcd7f680.jpg", + "image_caption": [ + "(a)", + "Figure 3. The process of generating meshflow from dense optical flow. (a) Propagate the motion vector of each grid center to the grid vertices. (b) Apply median filter $f_{1}$ to multiple motion vectors of each vertex to select the most appropriate motion for that vertex. (c) Use median filtering $f_{2}$ to smooth the motion field in the mesh grid. For ease of visualization, we present the $8 \\times 8$ grid mesh in this paper." + ], + "image_footnote": [], + "bbox": [ + 78, + 314, + 205, + 369 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/5ed42728f070a615355e62212338e2db4a1f3b4a4db9b7d464dcedc4afe647c7.jpg", + "image_caption": [ + "(b)" + ], + "image_footnote": [], + "bbox": [ + 207, + 315, + 334, + 369 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/03e081b20510d66517837d7e44b658e9805d21ab18742ce6b1f3211502508cc7.jpg", + "image_caption": [ + "(c)" + ], + "image_footnote": [], + "bbox": [ + 336, + 315, + 467, + 369 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "$V_{t_{k-1}}^i$ and $V_{t_k}^i$ . Secondly, EEMFlow utilizes multi-scale feature correlation to builds the cost volumes for meshflow estimation. Features $(V_{t_{k-1}}^i)^N$ and $(V_{t_k}^i)^N$ undergo average pooling operation $\\mathcal{P}$ to the same resolution (1/64 resolution of $V_{k\\rightarrow k+1}$ ), and then use correlation to capture relative motion information and output cost volumes $(C^i)^N$ . Specifically, we employ the dilated feature correlation (DFC) to increase the search area while reducing computational parameters. Finally, we stack the cost volumes $(C^i)^N$ and features $(V_{t_{k-1}}^i)^N$ after pooling, fuse them with a weighted-sum operation, and then feed them into the decoders to regress the meshflow $MF_{k\\rightarrow k+1}$ at the same resolution as $MF_{GT}$ . Inspiring by [51], we replace the conventional convolutions with group shuffle convolutions, which leads to efficient computation while maintaining high accuracy.", + "bbox": [ + 75, + 491, + 467, + 718 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Dilated Feature Correlation. We use the inner product to calculate correlation between $V_{t_{k-1}}^i$ and $V_{t_k}^i$ for meshflow estimation:", + "bbox": [ + 75, + 719, + 467, + 763 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nC ^ {i} (\\boldsymbol {u}, \\boldsymbol {d}) = V _ {t _ {k}} ^ {i} (\\boldsymbol {u}) \\cdot V _ {t _ {k - 1}} ^ {i} (\\boldsymbol {u} + \\boldsymbol {d}) / M, \\boldsymbol {d} \\in \\mathcal {N}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 104, + 777, + 468, + 796 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\pmb{u}$ represents the spatial coordinates on $V_{t_k}^i$ , $\\mathcal{N}$ represents the search grid of coordinate $\\pmb{u}$ in feature $V_{t_{k - 1}}^i$ , $M$ represents the number of elements in $\\mathcal{N}$ , $\\pmb{d}$ represents the offset coordinates of the elements in $\\mathcal{N}$ , and $\\cdot$ represents the inner product calculation. Many methods like [16, 40] simply define $\\mathcal{N}$ as a square range of size $(2r + 1)\\times (2r + 1)$", + "bbox": [ + 75, + 809, + 468, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "and observe that increasing the radius $r$ can reduce errors but increase computational overhead. We propose dilated feature correlation that samples densely around the center and sparsely at farther distances, thereby reducing computation while enabling larger radius cost volumes, as Eq. 2:", + "bbox": [ + 496, + 316, + 890, + 393 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {N} \\left(d _ {x}, d _ {y}\\right) = \\left\\{ \\begin{array}{l} 0, \\text {i f} | d _ {x} | + | d _ {y} | = 2 k, k \\in [ 2, r ], \\\\ 1, \\text {o t h e r s ,} \\end{array} \\right. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 514, + 414, + 890, + 454 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\pmb{d} = (d_x, d_y)$ represents the relative position coordinates within the neighborhood $\\mathcal{N}$ , where $\\mathcal{N}(d_x, d_y) = 1$ indicates the computation of correlation, and $\\mathcal{N}(d_x, d_y) = 0$ signifies no computation.", + "bbox": [ + 496, + 460, + 890, + 522 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2.2 Event-based Optical Flow Network", + "text_level": 1, + "bbox": [ + 500, + 540, + 797, + 556 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Based on EEMFlow, we make some improvements for accurate estimation of optical flow $F_{k \\to k + 1}$ , upgrading it to EEMFlow+. Since optical flow focuses more on local motion and pays attention to object edge details, we employ the coarse to fine residual approach to progressively refine the flow, which can be expressed as Eq. 3.", + "bbox": [ + 496, + 564, + 890, + 655 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nF ^ {i + 1} = \\operatorname {C o n v} ^ {i} \\left(\\mathcal {C} \\left(V _ {t _ {k - 1}} ^ {i}, \\mathcal {W} \\left(V _ {t _ {k}} ^ {i}, F _ {\\uparrow} ^ {i}\\right)\\right)\\right) + F _ {\\uparrow} ^ {i}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 539, + 662, + 890, + 681 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathcal{C}(\\cdot, \\cdot), \\mathcal{W}(\\cdot, \\cdot)$ and $\\uparrow$ donates our dilated feature correlation, the warping operation and upsampling, respectively. We employ the pyramid decoders for optical flow, thus $F^i$ is the output flow of the $i$ -th decoder, $F^{i+1}$ and $\\mathrm{Conv}^i$ are respectively the output flow and convolutions in the $i+1$ -th decoder. The most noteworthy aspect is upsampling $F^i$ to $F_{\\uparrow}^i$ . Many methods [6, 16, 40, 52] use bilinear interpolation for upsampling, but this can lead to the mixing of incorrect motions at object edges, resulting in blurring. Therefore, we propose the confidence-induced detail completion module for upsampling to enhance edge details.", + "bbox": [ + 496, + 688, + 890, + 854 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Confidence-induced Detail Completion Module. We propose the confidence-induced detail completion module", + "bbox": [ + 498, + 869, + 890, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "19201", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/77f3abc72bc1cb2188873eb9d46e2ab42e3ec3aa18c7b4727ce9821bc1397469.jpg", + "image_caption": [ + "Figure 4. Our proposed network architecture. We employ pyramid encoders to extract multi-scale features from $V_{k-1 \\to k}$ and $V_{k \\to k+1}$ , then use dilated feature correlation to compute the cost volume between each layer of features, followed by decoding to output the results. For meshflow estimation, we utilize the decoders with group shuffle convolutions to output predictions $MF_{k \\to k+1}$ , upsampled to $MF \\uparrow_{k \\to k+1}$ . For optical flow estimation, we refine flows using a coarse-to-fine residual approach and confidence-induced detail completion module, finally outputting the optical flow $F_{k \\to k+1}$ ." + ], + "image_footnote": [], + "bbox": [ + 80, + 89, + 893, + 231 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/dde6960db194bab40144fcb44dd2a88573264479b9b3ba324a806464bea22868.jpg", + "image_caption": [ + "Figure 5. The structure of CDC. CDC employs self-corrector based on a dense convolutional network and self-correlation based on a self-attention mechanism to correct the flow obtained from bilinear upsampling. $\\mathcal{B}$ represents bilinear upsampling." + ], + "image_footnote": [], + "bbox": [ + 78, + 313, + 468, + 459 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "(CDC) to eliminate the blurring of object edges caused by the mixing of multiple motions at the junctions of different movements during upsampling. The detailed structure of our CDC is shown in Fig. 5. Given a small-scale flow $F^i$ from the $i$ -th level, our CDC first generates an initial flow $\\bar{F}^i$ for the $i + 1$ -th level through bilinear interpolation, and then arranges the self-corrector and self-correlation branches to correct it. Self-corrector is based on a dense convolutional network with a five-layer structure. It captures motion information within the edge neighborhood through dense convolution from the concatenated feature $V_{t_{k-1}}^i$ and $V_{t_k}^i$ , outputting the corrected flow $\\Delta F^i$ and the corrected confidence map $W^i$ . Self-correlation is based on a self-attention mechanism, using a large receptive field to find the fine regions in features $V_{t_k}^i$ that are identical to the motion of the error region in $\\bar{F}^i$ . It outputs self-attention weights $A^i$ , multiplied with the initial flow $\\bar{F}^i$ . With the corrected flow $\\Delta F^i$ and self-attention weights $A^i$ , we can generate the fine flow $\\tilde{F}^i$ :", + "bbox": [ + 75, + 542, + 472, + 833 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {F} ^ {i} = \\alpha \\mathcal {W} (\\bar {F} ^ {i}, \\Delta F ^ {i}) + (1 - \\alpha) \\left(A ^ {i} \\otimes \\bar {F} ^ {i}\\right), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 129, + 840, + 468, + 861 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathcal{W}(\\cdot, \\cdot)$ means the warping operation, $\\otimes$ donates multiplication and $\\alpha \\in [0,1]$ is the weight coefficient. We iden", + "bbox": [ + 76, + 869, + 470, + 902 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "tify error-prone object edge areas based on the corrected confidence map $W^{i}$ , as weights to fuse the initial flow $\\bar{F}^{i}$ and the fine flow $\\tilde{F}^{i}$ , obtaining the final corrected flow $F_{\\uparrow}^{i}$ :", + "bbox": [ + 498, + 315, + 892, + 362 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nF _ {\\uparrow} ^ {i} = W ^ {i} \\odot \\bar {F} ^ {i} + (1 - W ^ {i}) \\odot \\tilde {F} ^ {i}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 578, + 369, + 892, + 390 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\odot$ donates the element-wise multiplier.", + "bbox": [ + 500, + 397, + 803, + 412 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2.3 Loss Function", + "text_level": 1, + "bbox": [ + 500, + 431, + 653, + 445 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For both meshflow and optical flow estimation, we use L1 loss for supervised regression during training. Our EEMFlow, used for estimating meshflow, directly outputs results at the same resolution as the meshflow GT, allowing for direct loss calculation. Similarly, the results outputted by our EEMFlow+ for optical flow estimation are at the same resolution as the network input, enabling direct calculation with the optical flow GT.", + "bbox": [ + 496, + 454, + 892, + 575 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 500, + 587, + 633, + 604 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Implementation Details", + "text_level": 1, + "bbox": [ + 500, + 612, + 718, + 628 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1.1 Datasets", + "text_level": 1, + "bbox": [ + 500, + 635, + 614, + 648 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our HREM dataset includes 100 indoor and outdoor scenes, with a resolution of $1280 \\times 720$ . We randomly select 70 scenes for training and reserve the remaining 30 for testing. The training set comprises 20,000 samples, while the test set contains 8,000 samples. Additionally, we further divide the test set by scene type (outdoor vs. indoor) and camera motion speed during rendering, resulting in four sub-sequences (outdoor Slow, outdoor fast, indoor slow, indoor fast), with mean motion magnitudes ranging from, $0 - 30$ , $30 - 100$ , $0 - 20$ and $20 - 100$ pixels, respectively. Moreover, similar to [52], we employ two data input modes: $dt = 1$ and $dt = 4$ . $dt = 1$ uses the event sequence between two consecutive frames of RGB images as input, with a meshflow generation frequency of $60~\\mathrm{Hz}$ while $dt = 4$ uses the event sequence between four consecutive frames of RGB images as input, with a meshflow", + "bbox": [ + 496, + 657, + 893, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "19202", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/06ac50e76b3d9ff6415a5f67a9dd86a0cb87cd499b50c614be298af420265e46.jpg", + "table_caption": [ + "Table 2. Quantitative comparison of our EEMFlow with other advanced flow networks on our HREM dataset. The evaluation metric used is End-Point Error (EPE). \"Parameters\" and \"Time\" respectively indicate the network parameter count and inference time. $\\Delta P$ and $\\Delta T$ represent the change in parameter count and inference time relative to ERAFT [42] for other networks. Smaller values are desirable for all metrics. We highlight the best results in red and the second-best results in blue." + ], + "table_footnote": [], + "table_body": "
Method dt = 1Parameters (M)Time (ms)OutdoorIndoorAvg
SlowFastSlowFast
EVFlownet [52]38.2463.5516.162.9311.658.57
PWCNet [40]3.36423.9114.492.8611.898.29
ERAFT [11]5.27934.1513.322.9110.347.68
SKFlow [41]6.281453.7611.787.248.817.24
GMA [17]5.891082.1812.072.029.346.40
KPAFlow [32]6.001842.0312.251.959.026.31
FlowFormer [15]9.872812.0611.711.888.666.08
EEMFlow(Ours)1.2472.429.092.008.465.50
Method dt = 4ΔP (M)ΔT (ms)OutdoorIndoorAvg
SlowFastSlowFast
EVFlownet [52]+624%+51%18.2549.3216.1647.1932.73
PWCNet [40]-36%-55%16.4046.1714.4940.9029.49
ERAFT [11]0%0%15.2140.8313.3239.6127.24
SKFlow [41]+19%+56%14.9339.2411.7139.2226.28
GMA [17]+11%+16%14.1338.8912.0737.6825.69
KPAFlow [32]+14%+99%14.0438.0312.2537.2025.38
FlowFormer [15]+88%+202%13.8938.5510.7738.5325.44
EEMFlow(Ours)-76%-92%13.9737.3312.0934.3924.45
", + "bbox": [ + 161, + 148, + 810, + 393 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/9fc0c88cacfe6e458c8cfb59ca03cda2904196b652f4d696e78cab8325f790d2.jpg", + "image_caption": [ + "Image Overlaid", + "Figure 6. Subjective results of image registration using meshflow estimated by other methods and ours. We estimate meshflow from the event sequences and then warp image $I_{t_1}$ onto image $I_{t_2}$ for fusion, showcasing the fused result. The fewer blue or red ghosting artifacts indicate better alignment performance." + ], + "image_footnote": [], + "bbox": [ + 83, + 405, + 205, + 515 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/f280d6c275454193d148eb0324e69dc68af1773c37875506ff7ef3732c8ceec1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 220, + 406, + 341, + 460 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/7cf3e046c10ea117aa8290abb1e1435be49017bb72252e4bbd7743e14a1ec991.jpg", + "image_caption": [ + "EVFlownet" + ], + "image_footnote": [], + "bbox": [ + 220, + 460, + 341, + 515 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/2e850dc37c6dc0639f6ab80da4b2653feccfda8fc021fcbba7f6e2f0aebd8f4d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 357, + 406, + 478, + 460 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/ccd3554bfda5da10a8e9166277aa1d8e8ada1cb211a4379a13572f3e8a16f6de.jpg", + "image_caption": [ + "ERRAFT" + ], + "image_footnote": [], + "bbox": [ + 357, + 460, + 478, + 515 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/ef632fbee6fec007d58cdc19bfefd4e63501bf5f9654b0328cad5b6bd2c07f8c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 493, + 406, + 614, + 460 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/d83a0811619fbfaf618e1e8b7653469de9a27f755c92ae3ed7fc963efaf4ef88.jpg", + "image_caption": [ + "KPAFlow" + ], + "image_footnote": [], + "bbox": [ + 493, + 460, + 614, + 513 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/fc59d6b08c2ff9b1b3d3365ca810a5655c76ec127f9c7f4a0380943935d6c3d5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 629, + 406, + 751, + 460 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/694eb12bb2cc5e279cfeda3c5636c85c56b9e28ed869030cf31f04b4d1ef9846.jpg", + "image_caption": [ + "Ours" + ], + "image_footnote": [], + "bbox": [ + 629, + 460, + 751, + 515 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/bf17103592adf34b08c77e58c71bc7df60d2c48ff7ec6cc8e06547915d1bf1c7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 766, + 406, + 887, + 460 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/1d33704e70b973568640502fbbc9ed0de00a48367e30881d1869afb8179708ac.jpg", + "image_caption": [ + "Meshflow(GT)" + ], + "image_footnote": [], + "bbox": [ + 766, + 460, + 887, + 513 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "generation frequency of $15\\mathrm{Hz}$ . DSEC [10] is a commonly used dataset for event-based optical flow estimation, which consists of real-world data captured by real event cameras mounted on cars. We also conduct experiments on DSEC to compare the performance for optical flow estimation.", + "bbox": [ + 75, + 589, + 468, + 665 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1.2 Training details", + "text_level": 1, + "bbox": [ + 76, + 683, + 241, + 698 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We conduct experiments using the PyTorch framework on two NVIDIA 2080Ti GPUs. We train all networks with the same parameters on our HRDM dataset. We employ the AdamW optimizer and OneCycle policy with a learning rate of $5 \\times 10^{-4}$ , weight decay of $5 \\times 10^{-5}$ , and other default parameters set to $\\beta_{1} = 0.9$ , $\\beta_{2} = 0.99$ , $\\epsilon = 1 \\times 10^{-4}$ . We train all networks up to $100k$ iterations to reach convergence.", + "bbox": [ + 75, + 707, + 468, + 811 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1.3 Evaluation metrics", + "text_level": 1, + "bbox": [ + 76, + 830, + 263, + 845 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Following EV-FlowNet [52], we use the average Endpoint Error (EPE) as the metric. When evaluating meshflow, we upsample both the prediction and the ground truth to the", + "bbox": [ + 75, + 854, + 468, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "input resolution to calculate metrics. In addition, the DSEC dataset uses $N$ PE to measure the percentage of flow errors higher than $N$ pixels in magnitude (e.g., 3PE, 2PE, 1PE) for flow outliers analysis, and employs the Angular Error (AE) to assess the directional accuracy.", + "bbox": [ + 496, + 589, + 892, + 665 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Comparison with State-of-the-Arts", + "text_level": 1, + "bbox": [ + 500, + 672, + 803, + 688 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2.1 Results for Event-based Meshflow Estimation", + "text_level": 1, + "bbox": [ + 500, + 696, + 870, + 710 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In Table 2, We train and test our EEMFlow on our HREM dataset, compared with some advanced networks, e.g., EVFlownet [54], ERAFT [11], PwCNet [40], SKFlow [41], GMA [17], KPAFlow [32], and FlowFormer [15]. Since these networks are all structured to output prediction with the same resolution of input, we upsample the ground truth $MF_{GT}$ to supervise these networks during training. However, EEMFlow can directly output prediction aligned $MF_{GT}$ in resolution. But for fair comparison, we upsample the predictions of all networks to a uniform resolution during evaluation. We train and test all networks using two input modes ( $dt = 1$ and $dt = 4$ ), and", + "bbox": [ + 496, + 719, + 893, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "19203", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/5a8c0f47f9e71a046ccb10fae5a783c7464652c6e285efe1cbd1b6288bf717ad.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 81, + 88, + 207, + 143 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/303579b082eac16c49178a78c7ae1ef0e09d1effdd845ba3c55e62c66aad3106.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 83, + 145, + 205, + 199 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/0dee9e03d491fbbe1585d914ddf12cd982e46a0f808e251c6153d7fad7ef55c3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 218, + 90, + 341, + 143 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/3b898acb02f98b32b54be43c95aa0f3af255f18269dbd6c15aaf89fa5f63765c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 218, + 143, + 343, + 199 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/5f4a5608ec7743aea9011075eeccbca48b75df2ebb526d7558ae79bbd5961930.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 354, + 90, + 478, + 143 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a09caf932e6a7f206ff8751da8f0b62c09fed60ff347aff61742d6e05893551e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 354, + 143, + 478, + 199 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/4f1449191a3f8e0360017f82e0d8d4c1065dd08d1c3e3f460d01a264dbf00e34.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 491, + 90, + 614, + 143 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/2a4b3087f83298301bdbf5e9ab908279fecac0daac1a3ef0fc67f89b3d761898.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 491, + 143, + 614, + 199 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/dde18daf6e3dc43071308ecae8c8601e4f3f29157e46d584a386ad2d454645c8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 627, + 90, + 751, + 143 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/8668f53d774ed290cc97deefb29b2f23456c761e37c7ac18ca38980638a5d5ad.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 627, + 143, + 751, + 199 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/fa5d9085616fdc7563c28349ab0c7c4a681ac8c04f7c1a5c50b5446e5d0f03f3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 764, + 90, + 887, + 143 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/8daba0d062a0b561ed461b2a3d2e0666e25095c66432f744e2a7a28cc5652a4b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 764, + 143, + 887, + 199 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/81d3b1f686b36f9233539ce65f98a0d56a7c96140969bdac66195b05b1aa5664.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 83, + 205, + 205, + 261 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/4332fd43de910a4058b61d313e838e981065761f2cb54436398902268d8223f8.jpg", + "image_caption": [ + "Reference" + ], + "image_footnote": [], + "bbox": [ + 83, + 261, + 205, + 316 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c385d0c5272a8683530646dd4a3e64b79abf569033a161965c7318bddf189f70.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 218, + 205, + 341, + 261 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/393596e3a19cd222d3e50dcdd2549c1bc1bcbfa89588c8b3b3ffb17759850f28.jpg", + "image_caption": [ + "EVFlownet" + ], + "image_footnote": [], + "bbox": [ + 218, + 261, + 341, + 316 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d6a7cd9d93821cdb4e5c5798ea0def3ec4b6924268c107ecc513356bcab3f497.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 354, + 205, + 478, + 261 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/530f474e613766ac8ff207f80d502ae2e95b85022db741752ca88dc704989483.jpg", + "image_caption": [ + "ERRAFT" + ], + "image_footnote": [], + "bbox": [ + 354, + 261, + 478, + 316 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/2574e57faea5804beb100cbca3076322da371bb0a2cb892e9c813da73e8c55d0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 491, + 205, + 614, + 261 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/9d356c7b34313c8838f65cf45d3df7bb64d06ea0e0144c5496adb2856deeea6f.jpg", + "image_caption": [ + "KPAFlow" + ], + "image_footnote": [], + "bbox": [ + 491, + 261, + 614, + 316 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/cf1de7783ed2ab4e7f427e7782088aaf6c7db3854ea12c4dcbd292719b5e07f0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 627, + 205, + 751, + 261 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d54f1c76ddaee17679afaefa9ace9198fcc4f1006d1054d6a2cf47e516da16e8.jpg", + "image_caption": [ + "Ours" + ], + "image_footnote": [], + "bbox": [ + 627, + 261, + 751, + 316 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a6768023085dc78983a637802e01f1a4abeeced6c0e06c5ebd2869d278113461.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 764, + 205, + 885, + 262 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ee6b5dde4d2b965b0cf76a1e2a5a03964507a8b64c20a7e627b497679540d6bc.jpg", + "image_caption": [ + "Meshflow(GT)" + ], + "image_footnote": [], + "bbox": [ + 764, + 262, + 885, + 316 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/5cab1699d76658dc9fd965632df748afd3a292fa2379222b96c4520f577084e0.jpg", + "table_caption": [ + "Table 3. Results on DSEC dataset. All networks are trained and tested on the DSEC dataset for optical flow estimation." + ], + "table_footnote": [], + "table_body": "
MethodsFPS↑1PE↓2PE↓3PE↓EPE↓AE↓
MutilCM [39]-76.648.530.93.4714.0
EV-Flownet [52]22.355.429.818.62.328.12
OF-EV-SNN [4]-53.720.210.31.716.34
EVA-Flow [50]-15.9-3.200.883.31
ERRAFT [11]11.412.74.742.680.792.85
ADMFlow [33]9.8812.54.672.650.782.84
EFlowformer [21]-11.24.102.450.762.68
TMA [25]7.5510.93.972.300.742.68
EEMFlow+(Ours)39.211.43.932.150.752.67
", + "bbox": [ + 84, + 422, + 460, + 547 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "present their performance metrics in four test sub-sequences (outdoor Slow, outdoor fast, indoor slow, indoor fast). In addition, we average the metric scores across all four test sub-sequences as shown in the \"Avg\" column.", + "bbox": [ + 75, + 580, + 468, + 640 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 2 shows that our EEMFlow achieves the lowest EPE score for both input modes $dt = 1$ and $dt = 4$ in the \"outdoor_fast\" and \"indoor_fast\" test sub-sequences, demonstrating its superior performance on high-speed and large-movement sequences. EEMFlow also exhibits great potential to outperform other flow networks on the \"outdoor Slow\" and \"indoor Slow\" test sub-sequences. Notably, EEMFlow achieves the lowest EPE scores in the \"Avg\" column for both input modes $dt = 1$ and $dt = 4$ . Besides, we also achieve the least number of parameters and the fastest inference speed. Compared to ERAFT [11], our EEMFlow reduces the parameter count by $76\\%$ (from 5.27M to 1.24M), reduces the inference time by $92\\%$ (from 93ms to 7ms), and improves average EPE in $dt = 4$ by $8\\%$ (from 27.24 to 25.18). EEMFlow achieves comparable EPE performance to FlowFormer [15] but exhibits a $38.7\\times$ increase in inference speed.", + "bbox": [ + 75, + 643, + 468, + 900 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a8d61c4db4aca6afe377c9452a17380f685f8f490498857f122a64dde872bc88.jpg", + "image_caption": [ + "Figure 7. Qualitative comparison of our proposed EEMFlow with other advanced flow networks on our HREM dataset. The subjective images of events and dense optical flow on the left side serve as references. The areas enclosed by red rectangles are zoomed in." + ], + "image_footnote": [], + "bbox": [ + 503, + 387, + 888, + 433 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/53cf844b4b98f4a9add46d018969669764dabdf077c036ee3109367c814fe87e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 503, + 433, + 888, + 478 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/4757f04649ab5f374f031e50ae7c3535afe04acf61d5c9f29cd36a7a1cce1619.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 503, + 478, + 888, + 523 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e3b3b30b2c3dad0b9d3e9e396d776cf60ee347b3f3d8c62a8bf81431d404f735.jpg", + "image_caption": [ + "Reference", + "EVFlownet", + "ERRAFT", + "TMA", + "Ours", + "Figure 8. Qualitative comparisons on the DSEC test set. We visualize the dense predictions and zoom in the areas where are apparent differences." + ], + "image_footnote": [], + "bbox": [ + 503, + 523, + 888, + 569 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In Fig. 7, we qualitatively compare our proposed EEMFlow with other flow networks, e.g., EVFlownet [54], ER-AFT [11], and KPAFlow [32]. To facilitate comparison, we upsample the meshflow estimation and ground truth to the same resolution of input and zoom in on the areas with the apparent differences. EVFlownet shows the worst performance, with many holes and color mixing. ERAFT and KPAFlow would exhibit block artifacts and appear coarse in nature. In contrast, our EEMFlow results are smoother with more natural color transitions and greater similarity to the upsampled ground truth. In Fig. 6, we present the subjective results of these networks for image registration. We estimate and upsample meshflow prediction from the event sequences $E(t_{1},t_{2})$ and then warp image $I_{t_1}$ onto image $I_{t_2}$ . We also zoom in on the challenging areas of alignment in the registered results, clearly demonstrating that our EEMFlow achieves excellent image registration performance.", + "bbox": [ + 496, + 643, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "19204", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/aebd1753d8341d94d1b4d7596543dfaea1fd6be8807c9858aa8abe6ba244bf69.jpg", + "image_caption": [ + "Figure 9. Visualization for meshflow and optical flow results." + ], + "image_footnote": [], + "bbox": [ + 80, + 89, + 465, + 186 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/67f9c7891317c47667481ddf5390d13766861b28d3aee9cfde3f75a0d6703fbe.jpg", + "table_caption": [ + "Table 4. Results of meshflow and optical flow on HREM for $dt = 1$" + ], + "table_footnote": [], + "table_body": "
TaskMethodOutdoorIndoorAvg
SlowFastSlowFast
Optical FlowFlowFormer6.2016.065.9915.2710.88
EEMFlow+3.8811.024.0310.927.46
Mesh-flowFlowFormer5.9915.125.7414.9510.45
EEMFlow2.429.092.008.465.50
", + "bbox": [ + 88, + 248, + 455, + 325 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.2.2 Results for event-based optical flow estimation", + "text_level": 1, + "bbox": [ + 76, + 357, + 452, + 372 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 3 presents the comparative results on DSEC dataset of our optical flow estimation network, EEMFlow+, against other event-based optical flow estimation networks. EEMFlow+ achieves state-of-the-art performance in the 2PE, 3PE, and AE metrics, and is on par with the best in the EPE metric. Notably, EEMFlow+ still maintains a significant advantage in inference speed. Compared to TMA [25], our EEMFlow+ increases the inference speed by $419\\%$ (from 7.55FPS to 39.2FPS). Fig. 8 displays the qualitative results of our optical flow estimation on the DSEC test set, compared with other advanced methods, including EVFlownet [52], ERAFT [42], and TMA [25]. By zooming into the object areas, it is clearly observable that our results exhibit more regular shapes and clearer boundaries, highlighting the detail enhancement capability of our proposed CDC module.", + "bbox": [ + 75, + 386, + 468, + 628 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.3. Ablation Studies", + "text_level": 1, + "bbox": [ + 76, + 648, + 240, + 662 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.3.1 The Advantages of Event-Meshflow Estimation", + "text_level": 1, + "bbox": [ + 76, + 674, + 457, + 689 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "According to the ERAFT [42], image-based approaches face challenges in handling difficult images due to the limited dynamic range of image sensors. As shown in Fig. 9, even advanced image-based techniques like Flow-Former struggle, while our event-based methods (EEMFlow, EEMFlow+) exhibit notably superior performance. Additional quantitative comparisons are provided in Table 4, our network of event-based meshflow estimation EEMFlow achieves the lowest EPE scores in HREM dataset for the $dt = 1$ input mode, which demonstrates the remarkable efficiency of our EEMFlow makes it suitable for real-time applications such as online video stabilization and autonomous driving.", + "bbox": [ + 75, + 704, + 468, + 900 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/58d4be2b18066e05ee7a08df68467e629fa513b9732d10e51c63e14893e39ea8.jpg", + "table_caption": [ + "Table 5. Ablation studies about CDC of EEMFlow+. CDC consists of two branches, the self-corrector and self-correlation." + ], + "table_footnote": [], + "table_body": "
ModelSelf-correctorSelf-correlationFPS↑3PE↓EPE↓AE↓
(a)XX60.43.520.893.11
(b)✓X55.62.770.812.92
(c)X✓46.32.650.792.78
(d)✓✓39.22.150.752.67
", + "bbox": [ + 506, + 125, + 880, + 200 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.3.2 Experiments for CDC of EEMFlow+", + "text_level": 1, + "bbox": [ + 498, + 231, + 808, + 244 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In Table 5, we also conduct ablation experiments on the CDC module of EEMFlow+ used for optical flow estimation, including its two branches, the self-corrector and self-correlation. We train and evaluate all models using the same settings on the DSEC dataset to show the individual impact of each branch in CDC module. Comparison of (a)&(b) demonstrates that the CDC with only the self-corrector branch can bring a significant increase in accuracy with a minimal loss in speed. Comparison of (b)&(c) shows that self-correlation, compared to self-corrector, can lead to a higher increase in accuracy, albeit at a further reduction in inference speed. Finally, the comparison of (a)&(d) shows that the CDC composed of both the self-corrector and self-correlation branches significantly improves the accuracy of optical flow estimation with an acceptable loss in speed.", + "bbox": [ + 496, + 258, + 890, + 486 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 498, + 507, + 617, + 522 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this work, we develop the first event-based meshflow dataset, termed HREM, where 100 indoor and outdoor virtual scenes with rich scene contents are rendered using Blender. Our HREM possess physically correct accurate events and meshflow label pairs with so far the highest resolution. Furthermore, we propose an efficient event-based meshflow network (EEMFlow), which achieves state-of-the-art performance on HREM dataset while maintaining high efficiency. Based on EEMFlow, we propose a confidence-induced detail completion module to upgrade it as EEMFlow+ for optical flow estimation, achieving SOTA results on DSEC dataset with the fastest inference speed.", + "bbox": [ + 496, + 535, + 890, + 717 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The integration of optical flow and depth data allows for the extension of the proposed framework to address 3D scene flow. Subsequently, our future work will involve expanding this methodology on the HREM dataset by integrating scene flow annotations and investigating scene flow estimation using event cameras.", + "bbox": [ + 496, + 719, + 890, + 810 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements. This work was supported by National Natural Science Foundation of China (NSFC) under grant No.62372091, Sichuan Science and Technology Program of China under grants Nos. 2023NSFSC0462, 2022YFQ0079 and 2023YFG0054.", + "bbox": [ + 496, + 825, + 890, + 898 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "19205", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Ryad Benosman, Sio-Hoi Ieng, Charles Clercq, Chiara Bartolozzi, and Mandyam Srinivasan. Asynchronous frameless event-based optical flow. Neural Networks, 27:32-37, 2012. 2", + "[2] Christian Brandli, Raphael Berner, Minhao Yang, Shih-Chii Liu, and Tobi Delbruck. A $240 \\times 180$ 130 db $3\\mu s$ latency global shutter spatiotemporal vision sensor. IEEE Journal of Solid-State Circuits, 49(10):2333-2341, 2014. 2", + "[3] Marco Cannici, Marco Ciccone, Andrea Romanoni, and Matteo Matteucci. A differentiable recurrent surface for asynchronous event-based data. In Proc. ECCV, pages 136-152. Springer, 2020. 2", + "[4] Javier Cuadrado, Ulysse Rançon, Benoit R Cotterreau, Francisco Barranco, and Timothee Masquelier. Optical flow estimation from event-based cameras and spiking neural networks. Frontiers in Neuroscience, 17:1160034, 2023. 7", + "[5] Ziluo Ding, Rui Zhao, Jiyuan Zhang, Tianxiao Gao, Ruiqin Xiong, Zhaofei Yu, and Tiejun Huang. Spatio-temporal recurrent networks for event-based optical flow estimation. In Proc. AAAI, pages 525-533, 2022. 2", + "[6] Alexey Dosovitskiy, Philipp Fischer, Eddy Ilg, Philip Hausser, Caner Hazirbas, Vladimir Golkov, Patrick Van Der Smagt, Daniel Cremers, and Thomas Brox. Flownet: Learning optical flow with convolutional networks. In Proc. ICCV, pages 2758-2766, 2015. 4", + "[7] Guillermo Gallego, Tobi Delbrück, Garrick Orchard, Chiara Bartolozzi, Brian Taba, Andrea Censi, Stefan Leutenegger, Andrew J Davison, Jörg Conradt, Kostas Daniilidis, et al. Event-based vision: A survey. IEEE transactions on pattern analysis and machine intelligence, 44(1):154-180, 2020. 1", + "[8] Daniel Gehrig, Antonio Loquercio, Konstantinos G Derpanis, and Davide Scaramuzza. End-to-end learning of representations for asynchronous event-based data. In Proc. ICCV, pages 5633-5643, 2019. 2", + "[9] Daniel Gehrig, Mathias Gehrig, Javier Hidalgo-Carrió, and Davide Scaramuzza. Video to events: Recycling video datasets for event cameras. In Proc. CVPR, pages 3586-3595, 2020. 2", + "[10] Mathias Gehrig, Willem Aarents, Daniel Gehrig, and Davide Scaramuzza. Dsec: A stereo event camera dataset for driving scenarios. IEEE Robotics and Automation Letters, 6(3): 4947-4954, 2021. 2, 3, 4, 6", + "[11] Mathias Gehrig, Mario Millhäuser, Daniel Gehrig, and Davide Scaramuzza. E-raft: Dense optical flow from event cameras. In International Conference on 3D Vision (3DV), pages 197–206, 2021. 1, 2, 3, 6, 7", + "[12] Klaus Greff, Francois Belletti, Lucas Beyer, Carl Doersch, Yilun Du, Daniel Duckworth, David J Fleet, Dan Gnanapragasam, Florian Golemo, Charles Herrmann, et al. Kubric: A scalable dataset generator. In Proc. CVPR, pages 3749-3761, 2022. 3", + "[13] Jesse Hagenaars, Federico Paredes-Vallés, and Guido De Croon. Self-supervised learning of event-based optical flow with spiking neural networks. Advances in Neural Information Processing Systems, 34:7167-7179, 2021. 2" + ], + "bbox": [ + 78, + 114, + 470, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[14] Yuhuang Hu, Shih-Chii Liu, and Tobi Delbruck. v2e: From video frames to realistic dvs events. In Proc. CVPR, pages 1312-1321, 2021. 3", + "[15] Zhaoyang Huang, Xiaoyu Shi, Chao Zhang, Qiang Wang, Ka Chun Cheung, Hongwei Qin, Jifeng Dai, and Hongsheng Li. FlowFormer: A transformer architecture for optical flow. ECCV, 2022. 2, 6, 7", + "[16] Junhwa Hur and Stefan Roth. Iterative residual refinement for joint optical flow and occlusion estimation. In Proc. CVPR, 2019. 4", + "[17] Shihao Jiang, Dylan Campbell, Yao Lu, Hongdong Li, and Richard Hartley. Learning to estimate hidden motions with global motion aggregation. In Proc. ICCV, pages 9772-9781, 2021. 2, 6", + "[18] Jacques Kaiser, J Camilo Vasquez Tieck, Christian Hubschneider, Peter Wolf, Michael Weber, Michael Hoff, Alexander Friedrich, Konrad Wojtasik, Arne Roennau, Ralf Kohlhaas, et al. Towards a framework for end-to-end control of a simulated vehicle with spiking neural networks. In IEEE International Conference on Simulation, Modeling, and Programming for Autonomous Robots (SIMPAR), pages 127-134, 2016. 2", + "[19] Chankyu Lee, Adarsh Kumar Kosta, Alex Zihao Zhu, Kenneth Chaney, Kostas Daniilidis, and Kaushik Roy. Spikeflownet: event-based optical flow estimation with energy-efficient hybrid neural networks. In Proc. ECCV, pages 366-382. Springer, 2020. 2", + "[20] Chankyu Lee, Adarsh Kumar Kosta, and Kaushik Roy. Fusion-flownet: Energy-efficient optical flow estimation using sensor fusion and deep fused spiking-analog network architectures. In 2022 International Conference on Robotics and Automation (ICRA), pages 6504-6510. IEEE, 2022. 2", + "[21] Yijin Li, Zhaoyang Huang, Shuo Chen, Xiaoyu Shi, Hongsheng Li, Hujun Bao, Zhaopeng Cui, and Guofeng Zhang. Blinkflow: A dataset to push the limits of event-based optical flow estimation. arXiv preprint arXiv:2303.07716, 2023. 2, 3, 7", + "[22] Zhuoyan Li, Jiawei Shen, and Ruitao Liu. A lightweight network to learn optical flow from event data. In 2020 25th International Conference on Pattern Recognition (ICPR), pages 1-7. IEEE, 2021. 2", + "[23] Kaimo Lin, Shuaicheng Liu, Loong-Fah Cheong, and Bing Zeng. Seamless video stitching from hand-held camera inputs. In Computer Graphics Forum, pages 479-487. Wiley Online Library, 2016. 2", + "[24] Songnan Lin, Ye Ma, Zhenhua Guo, and Bihan Wen. Dvsvoltmeter: Stochastic process-based event simulator for dynamic vision sensors. In Proc. ECCV, pages 578-593. Springer, 2022. 2, 3", + "[25] Haotian Liu, Guang Chen, Sanqing Qu, Yanping Zhang, Zhi-jun Li, Alois Knoll, and Changjun Jiang. Tma: Temporal motion aggregation for event-based optical flow. In Proc. ICCV, 2023. 2, 7, 8", + "[26] Shuaicheng Liu, Lu Yuan, Ping Tan, and Jian Sun. Bundled camera paths for video stabilization. ACM transactions on graphics (TOG), 32(4):1-10, 2013. 1" + ], + "bbox": [ + 501, + 92, + 890, + 901 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "19206", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[27] Shuaicheng Liu, Lu Yuan, Ping Tan, and Jian Sun. Steadyflow: Spatially smooth optical flow for video stabilization. In Proc. CVPR, pages 4209-4216, 2014. 1", + "[28] Shuaicheng Liu, Ping Tan, Lu Yuan, Jian Sun, and Bing Zeng. Meshflow: Minimum latency online video stabilization. In Proc. ECCV, pages 800-815. Springer, 2016. 1, 2, 3", + "[29] Shuaicheng Liu, Nianjin Ye, Chuan Wang, Jirong Zhang, Lanpeng Jia, Kunming Luo, Jue Wang, and Jian Sun. Content-aware unsupervised deep homography estimation and its extensions. 45(3):2849-2863, 2022. 1", + "[30] Shuaicheng Liu, Yuhang Lu, Hai Jiang, Nianjin Ye, Chuan Wang, and Bing Zeng. Unsupervised global and local homography estimation with motion basis learning. 45(6): 7885-7899, 2023. 1", + "[31] Zhen Liu, Yinglong Wang, Bing Zeng, and Shuaicheng Liu. Ghost-free high dynamic range imaging with context-aware transformer. In Proc. ECCV, pages 344-360. Springer, 2022. 1", + "[32] Ao Luo, Fan Yang, Xin Li, and Shuaicheng Liu. Learning optical flow with kernel patch attention. In Proc. CVPR, 2022. 2, 6, 7", + "[33] Xinglong Luo, Kunming Luo, Ao Luo, Zhengning Wang, Ping Tan, and Shuaicheng Liu. Learning optical flow from event camera with rendered dataset. In Proc. ICCV, 2023. 2, 3, 7", + "[34] Lang Nie, Chunyu Lin, Kang Liao, Shuaicheng Liu, and Yao Zhao. Depth-aware multi-grid deep homography estimation with contextual correlation. 32(7):4460-4472, 2022. 1", + "[35] Lang Nie, Chunyu Lin, Kang Liao, Shuaicheng Liu, and Yao Zhao. Parallax-tolerant unsupervised deep image stitching. In Proc. ICCV, pages 7399-7408, 2023. 2", + "[36] Liyuan Pan, Miaomiao Liu, and Richard Hartley. Single image optical flow estimation with an event camera. In Proc. CVPR, pages 1669-1678, 2020. 2", + "[37] Henri Rebecq, Daniel Gehrig, and Davide Scaramuzza. Esim: an open event camera simulator. In Conference on robot learning, pages 969-982. PMLR, 2018. 3", + "[38] Zhihang Ren, Jiajia Li, Shuaicheng Liu, and Bing Zeng. Meshflow video denoising. In Proc. ICIP, pages 2966-2970, 2017. 2", + "[39] Deqing Sun, Stefan Roth, and Michael J Black. Secrets of optical flow estimation and their principles. In Proc. CVPR, pages 2432-2439, 2010. 3, 7", + "[40] Deqing Sun, Xiaodong Yang, Ming-Yu Liu, and Jan Kautz. Pwc-net: Cnns for optical flow using pyramid, warping, and cost volume. In Proc. CVPR, pages 8934-8943, 2018. 4, 6", + "[41] Shangkun Sun, Yuanqi Chen, Yu Zhu, Guodong Guo, and Ge Li. Skflow: Learning optical flow with super kernels. Proc. NeurIPS, 35:11313-11326, 2022. 2, 6", + "[42] Zachary Teed and Jia Deng. Raft: Recurrent all-pairs field transforms for optical flow. In Proc. ECCV, 2020. 2, 6, 8", + "[43] Zhexiong Wan, Yuchao Dai, and Yuxin Mao. Learning dense and continuous optical flow from an event camera. IEEE Transactions on Image Processing, 31:7237-7251, 2022. 2, 3" + ], + "bbox": [ + 78, + 90, + 467, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[44] Zhexiong Wan, Yuxin Mao, Jing Zhang, and Yuchao Dai. Rpeflow: Multimodal fusion of rgb-pointcloud-event for joint optical flow and scene flow estimation. In Proc. ICCV, pages 10030-10040, 2023. 2, 3", + "[45] Yiming Wang, Qian Huang, Chuanxu Jiang, Jiwen Liu, Mingzhou Shang, and Zhuang Miao. Video stabilization: A comprehensive survey. Neurocomputing, 2022. 1, 2", + "[46] Qingsen Yan, Jinqiu Sun, Haisen Li, Yu Zhu, and Yanning Zhang. High dynamic range imaging by sparse representation. Neurocomputing, 269:160-169, 2017. 1, 2", + "[47] Min Yang, Jingkun Liang, Jianhai Zhang, Haidong Gao, Fanyong Meng, Li Xingdong, and Sung-Jin Song. Non-local means theory based perona-malik model for image denoising. Neurocomputing, 120:262-267, 2013. 2", + "[48] Yan Yang, Liyuan Pan, and Liu Liu. Event camera data pretraining. In Proc. ICCV, pages 10699-10709, 2023. 2", + "[49] Nianjin Ye, Chuan Wang, Shuaicheng Liu, Lanpeng Jia, Jue Wang, and Yongqing Cui. Deepmeshflow: Content adaptive mesh deformation for robust image registration. arXiv preprint arXiv:1912.05131, 2019. 1, 2", + "[50] Yaozu Ye, Hao Shi, Kailun Yang, Ze Wang, Xiaoting Yin, Yaonan Wang, and Kaiwei Wang. Towards anytime optical flow estimation with event cameras. arXiv preprint arXiv:2307.05033, 2023. 2, 7", + "[51] Xiangyu Zhang, Xinyu Zhou, Mengxiao Lin, and Jian Sun. Shufflenet: An extremely efficient convolutional neural network for mobile devices. In Proc. CVPR, pages 6848-6856, 2018. 4", + "[52] Alex Zhu, Liangzhe Yuan, Kenneth Chaney, and Kostas Daniilidis. Ev-flownet: Self-supervised optical flow estimation for event-based cameras. In Proceedings of Robotics: Science and Systems, 2018. 2, 4, 5, 6, 7, 8", + "[53] Alex Zihao Zhu, Dinesh Thakur, Tolga Ozaslan, Bernd Pfrommer, Vijay Kumar, and Kostas Daniilidis. The multi-vehicle stereo event camera dataset: An event camera dataset for 3d perception. IEEE Robotics and Automation Letters, 3 (3):2032-2039, 2018. 2, 3, 4", + "[54] Alex Zihao Zhu, Liangzhe Yuan, Kenneth Chaney, and Kostas Daniilidis. Unsupervised event-based learning of optical flow, depth, and egomotion. In Proc. CVPR, pages 989-997, 2019. 3, 6, 7" + ], + "bbox": [ + 501, + 92, + 890, + 671 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "19207", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2024/Efficient Meshflow and Optical Flow Estimation from Event Cameras/141a2ee1-d1ef-416d-8c61-bc462b1074b9_model.json b/2024/Efficient Meshflow and Optical Flow Estimation from Event Cameras/141a2ee1-d1ef-416d-8c61-bc462b1074b9_model.json new file mode 100644 index 0000000000000000000000000000000000000000..561f9f9db9209d6092374963f6beae5dffef1dea --- /dev/null +++ b/2024/Efficient Meshflow and Optical Flow Estimation from Event Cameras/141a2ee1-d1ef-416d-8c61-bc462b1074b9_model.json @@ -0,0 +1,2959 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.135, + 0.131, + 0.838, + 0.154 + ], + "angle": 0, + "content": "Efficient Meshflow and Optical Flow Estimation from Event Cameras" + }, + { + "type": "text", + "bbox": [ + 0.102, + 0.179, + 0.877, + 0.27 + ], + "angle": 0, + "content": "Xinglong Luo\\(^{1,4*}\\), Ao Luo\\(^{2,4*}\\), Zhengning Wang\\(^{1\\dagger}\\), Chunyu Lin\\(^{3}\\), Bing Zeng\\(^{1}\\), and Shuaicheng Liu\\(^{1,4\\dagger}\\) \n\\(^{1}\\)University of Electronic Science and Technology of China \n\\(^{2}\\)Southwest Jiaotong University \n\\(^{3}\\)Beijing Jiaotong University \n\\(^{4}\\)Megvii Technology \n{luoboom@std.,zhengning.wang@,eezeng@,liushuaicheng@}uestc.edu.cn \naoluo@swjtu.edu.cn cylin@bjtu.edu.cn" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.305, + 0.314, + 0.321 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.336, + 0.474, + 0.653 + ], + "angle": 0, + "content": "In this paper, we explore the problem of event-based meshflow estimation, a novel task that involves predicting a spatially smooth sparse motion field from event cameras. To start, we generate a large-scale High-Resolution Event Meshflow (HREM) dataset, which showcases its superiority by encompassing the merits of high resolution at \\(1280 \\times 720\\), handling dynamic objects and complex motion patterns, and offering both optical flow and meshflow labels. These aspects have not been fully explored in previous works. Besides, we propose Efficient Event-based MeshFlow (EEMFlow) network, a lightweight model featuring a specially crafted encoder-decoder architecture to facilitate swift and accurate meshflow estimation. Furthermore, we upgrade EEMFlow network to support dense event optical flow, in which a Confidence-induced Detail Completion (CDC) module is proposed to preserve sharp motion boundaries. We conduct comprehensive experiments to show the exceptional performance and runtime efficiency \\((39 \\times\\) faster) of our EEMFlow model compared to recent state-of-the-art flow methods. Our code is available at https://github.com/boomluo02/EEMFlow." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.671, + 0.21, + 0.687 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.696, + 0.471, + 0.879 + ], + "angle": 0, + "content": "Meshflow, a spatially smooth sparse motion field, represents motion vectors exclusively at mesh vertices [28, 49], which has been widely applied in various vision applications, such as image alignment [29, 30, 34], video stabilization [26, 27, 45] and high dynamic range (HDR) imaging [31, 46]. This motion representation combines the benefits of optical flow and global homography, effectively reducing redundancy in motion information and computational costs, while also accommodating non-rigid motions beyond single-plane movements. However, meshflow estimation on RGB images often encounter challenges under scenarios such as low-light and rapid motions. This is due" + }, + { + "type": "image", + "bbox": [ + 0.502, + 0.304, + 0.895, + 0.532 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.541, + 0.895, + 0.591 + ], + "angle": 0, + "content": "Figure 1. Comparison of computational overhead and accuracy metrics. The x-axis represents inference time, while the y-axis corresponds to the end-point error. The size of each circle indicates the number of model parameters. Lower values for all metrics are considered better." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.623, + 0.877, + 0.639 + ], + "angle": 0, + "content": "to the loss of fine image texture details and motion blurs." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.642, + 0.892, + 0.822 + ], + "angle": 0, + "content": "In contrast, event cameras are well-suited for motion estimation under such situations [7, 11]. Equipped with bio-inspired vision sensors, event cameras can generate sequences of events with microsecond accuracy triggered by changes in log intensity. In particular, when a change is detected in a pixel, the camera returns an event in the form \\( e = (x,y,t,p) \\) immediately, where \\( x,y \\) stands for the spatial location, \\( t \\) refers to the timestamp in microseconds, and \\( p \\) is the polarity of the change, either positive or negative. The advantages of high temporal resolution and high dynamic range make event cameras highly effective for analyzing dynamic scenes." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.826, + 0.895, + 0.9 + ], + "angle": 0, + "content": "In this work, we study a new problem that estimates meshflow from event camera data. To start, we create a large-scale High-Resolution Event Meshflow (HREM) dataset, which contains 20k train and 8k test samples. We build 100 virtual scenes in Blender to render the dataset," + }, + { + "type": "page_footnote", + "bbox": [ + 0.1, + 0.887, + 0.351, + 0.901 + ], + "angle": 0, + "content": "* Equal contribution. † Corresponding Author." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "19198" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.473, + 0.381 + ], + "angle": 0, + "content": "which can provide accurate physically-based events along with dense meshflow label pairs. Based on the dataset, we further propose an Efficient Event-based MeshFlow (EEMFlow) network to estimate high-resolution meshflow from event data. Unlike recent flow networks relying on recurrent refinement structure [11, 17, 42], our network is developed on an encoder-decoder architecture with multi-scale global optimizing scheme, which can produce full-resolution meshflow with minimal computational overhead. Specifically, our EEMFlow achieves efficiency by employing the lightweight encoder, building cost volume with dilated feature correlation, and using group shuffle convolutions during decoding. We select recent top-ranked flow networks, including ERAFT [11], SK-Flow [41], KPAFlow [32], and FlowFormer [15], which is trained on our HREM dataset. The results show that our approach achieves state-of-the-art performance while maintaining a fast inference speed of 142.9 FPS, outperforming previous works by a relatively large margin (see Fig. 1)." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.395, + 0.473, + 0.622 + ], + "angle": 0, + "content": "Additionally, we empirically demonstrate that the proposed new pipeline has the capability to effectively handle various motion patterns. Its lightweight design and runtime efficiency further contribute significantly to the field of optical flow estimation. Specifically, we refine the optical flow progressively during decoding using the coarse to fine residual approach. A Confidence-induced Detail Completion (CDC) module is proposed to preserve motion boundary details during flow upsampling. We also perform comparative experiments with recent event flow networks [21, 25, 33] to illustrate its superiority. The enhanced flow network is referred to as EEMFlow+, demonstrating impressive performance on the reputable DSEC [11] online test benchmark, with the fastest inference speed reaching 39.2 FPS. Our contributions are summarized as:" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.644, + 0.47, + 0.719 + ], + "angle": 0, + "content": "- We build HREM, the first event-based meshflow dataset, superior in the highest resolution at \\(1280 \\times 720\\), dynamic scenes, complex motion patterns, as well as physically correct accurate events paired with meshflow and optical flow labels." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.726, + 0.47, + 0.802 + ], + "angle": 0, + "content": "- We propose EEMFlow, which achieves SOTA performances when compared to top-ranked optical flow networks when trained on our meshflow dataset. Moreover, it achieves inference speed of 142.9 FPS, which is 25.5 to 38.7 times faster than compared methods." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.808, + 0.471, + 0.899 + ], + "angle": 0, + "content": "- We propose CDC, a confidence-induced detail completion module that empowers EEMFlow to make a meaningful contribution to the optical flow community. The upgraded model achieving SOTA performance when compared to representative methods, while also boasting the fastest inference speed to date." + }, + { + "type": "list", + "bbox": [ + 0.078, + 0.644, + 0.471, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.09, + 0.643, + 0.107 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.115, + 0.785, + 0.132 + ], + "angle": 0, + "content": "2.1. Image-based Meshflow Warping" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.138, + 0.892, + 0.427 + ], + "angle": 0, + "content": "Meshflow is a lightweight and spatially smooth sparse motion field with motions only located at mesh vertices [28]. Meshflow is different from dense optical flow, where optical flow estimates motions of every pixel of an image while meshflow only concentrates on the global motion, rejecting motions of any dynamic contents. Meshflow is also different from a global homography, where local motions from nonplanar depth variations can be well reflected. Mesh-based methods proofs to be effective in various applications, such as high dynamic range (HDR) imaging [46], burst image denoising [47], video denoising [38], image/video stitching [23, 35] and video stabilization [45]. It is worth noting that directly downsampling the optical flow may yield flow fields of the same resolution as meshflow, but they differ significantly from meshflow and perform poorly in terms of warping effects [49]. Direct downsampling ignores motion outliers from different dynamic objects and the global consistency, which can be effectively addressed by meshflow through motion propagation." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.436, + 0.824, + 0.452 + ], + "angle": 0, + "content": "2.2. Event-based Optical Flow Estimation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.459, + 0.892, + 0.761 + ], + "angle": 0, + "content": "Optical flow estimation from event cameras has received significant attention in recent years. Early approaches, such as [1], could only estimate optical flows at the regions where events are triggered. Recently, deep methods can estimate optical flows from event data, even for the regions without triggered events. For example, EV-FlowNet [52] learns event and flow labels in a self-supervised manner by minimizing photometric distances of grey images acquired by DAVIS [2]. Various event representations, including EST [8] and Matrix-LSTM [3], have been explored, and different network structures, such as SpikeFlowNet [19], LIF-EV-FlowNet [13], STE-FlowNet [5], Li et al. [22], ERAFT [11], Yang et al. [48], EVA-Flow [50], ADMFlow [33], E-FlowFormer [21], and TMA [25] have been proposed to improve performances. Some methods even use both events and images as input for flow estimation, such as Fusion-FlowNet [20], Pan et al. [36], DCEI-Flow [43] and RPEFlow [44]. In this work, we study a new problem of event-based meshflow estimation, proposing an efficient event-based meshflow network." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.772, + 0.798, + 0.788 + ], + "angle": 0, + "content": "2.3. Event-based Optical Flow Dataset" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.795, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Early works synthesize events by thresholding rgb images [18] and applying interpolation for high framerate [9]. However, the timestamp of synthesized events is inaccurate, let alone interpolation artifacts. The DAVIS event camera [2] can capture both images and real events, resulting in some event datasets: DVSFLOW [24], MVSEC [53] and DSEC [10], based on which EV-Flownet [52] and ER" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "19199" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.093, + 0.089, + 0.877, + 0.102 + ], + "angle": 0, + "content": "Table 1. Comparison of available datasets. 'Dense OF' and Meshflow' indicate whether the dataset has dense optical flow labels and meshflow labels." + }, + { + "type": "table", + "bbox": [ + 0.084, + 0.108, + 0.887, + 0.209 + ], + "angle": 0, + "content": "
DatasetResolutionDynamic ObjectsExtreme ConditionsDense OFMeshflowMotion Pattern
DVSFLOW [24]180×240XXXXRotation
MVSEC [53]260×346X✓XXDrone
DSEC [10]480×640X✓XXCar
MDR [33]260×346XX✓XCar
BlinkFlow [21]480×640✓X✓XRandom
Ekubric [44]720×1280✓X✓XFalling
HREM(Ours)720×1280✓✓✓✓Random
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.228, + 0.473, + 0.441 + ], + "angle": 0, + "content": "AFT [11] compute sparse optical flow. In this way, however, flows can only locate on sparse event regions [43]. Recently, Luo et al. [33] proposed to render the MDR dataset from graphics, but it only contains static scenes. Wan et al. [44] synthesized the Ekubric dataset based on the Kubirc toolbox [12], which only includes a single falling motion pattern. Li et al. [21] considered non-rigid motions simulating dancing but with lower resolution and without extreme condition scenarios. None of the aforementioned datasets involve the estimation of meshflow. In this work, we render a comprehensive dataset that can support both meshflow and optical flow estimation, with a higher resolution, rich dynamic scenes, as well as extreme conditions like relatively low light and motion blur, as shown in Table 1." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.457, + 0.192, + 0.474 + ], + "angle": 0, + "content": "3. Algorithm" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.483, + 0.43, + 0.499 + ], + "angle": 0, + "content": "3.1. High-Resolution Event Meshflow Dataset" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.507, + 0.47, + 0.794 + ], + "angle": 0, + "content": "We create the high-resolution event meshflow dataset (HREM), consisting of 100 virtual scenes that accurately mimic real-world environments, both indoors and outdoors. In these scenes, we put dynamic objects to simulate intricate object motions. Camera is programmed to track these movements, ensuring a realistic portrayal of motion. The Blender rendering engine was utilized to create high frame rate videos and dense optical flow labels. For event data generation, we employed three advanced simulators: ESIM [37], V2E [14], and DVS-Voltmeter [24]. The simulator that provided the highest contrast in warped events images was chosen for our dataset. We also process the dense optical flow using motion propagation and median filters, enabling the generation of meshflow labels. Our comparisons with existing datasets is shown in Fig. 2, emphasize the superiority of HREM in high-resolution, dynamic scenes, complex motion patterns, and comprehensive labeling. Following [28], we generate meshflow from dense optical flow, as depicted in Fig. 3." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.796, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Motion Propagation. Given a dense optical flow \\( \\mathbf{F} \\), we place a uniform mesh of \\( 16 \\times 16 \\) regular cells on its image plane and then select the motion of the middle point \\( p \\) in each cell as the local motion \\( v_{p} \\). Since the vertices of the mesh near point \\( p \\) should have a similar motion to \\( v_{p} \\), we define a rectangle that covers \\( 3 \\times 3 \\) cells centered at \\( p \\), and assign to all the vertices within the rectangle, ensuring the" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.228, + 0.885, + 0.244 + ], + "angle": 0, + "content": "propagation of the local motion \\( v_{p} \\) across the image plane." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.244, + 0.892, + 0.547 + ], + "angle": 0, + "content": "Median Filters. The local motions of the middle points in all cells are propagated to their nearby mesh vertices, resulting in each vertex potentially receiving multiple motion vectors. To select the most appropriate motion vector for a given vertex, we apply a median filter \\( f_{1} \\) to filter the candidate motions. The response of the filter is then assigned to the corresponding vertex. The median filter is a widely used technique in optical flow estimation and has been shown to produce high-quality flow estimates [39]. Therefore, we use the median filter for sparse motion regularization. However, due to dynamic objects, the motion field may contain noise and needs to be spatially smoothed. To address this issue, we apply another median filter \\( f_{2} \\) that covers \\( 3 \\times 3 \\) cells neighborhood to suppress the noise in the motion field. This second median filter produces a spatially-smooth sparse motion field, which is what we called as meshflow. Ultimately, we use the meshflow (generally \\( 16 \\times 16 \\)) as the label, which contains global motion information. However, we upsample the meshflow to the full image resolution for intuitive display and alignment applications." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.557, + 0.864, + 0.573 + ], + "angle": 0, + "content": "3.2. Estimation for Meshflow and Optical Flow" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.581, + 0.892, + 0.687 + ], + "angle": 0, + "content": "Following [11] which estimates optical flow from two consecutive event sequences, we estimate the meshflow \\(MF_{k\\to k + 1}\\) and optical flow \\(F_{k\\rightarrow k + 1}\\) from event sequences \\(E(t_{k - 1},t_k)\\) and \\(E(t_{k},t_{k + 1})\\) , and overall architecture of our network is shown in Fig. 4. Fellow [54], we convert the inputs \\(E(t_{k - 1},t_k)\\) and \\(E(t_{k},t_{k + 1})\\) to the 3D volumes \\(V_{k - 1\\to k}\\) and \\(V_{k\\to k + 1}\\)" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.695, + 0.84, + 0.71 + ], + "angle": 0, + "content": "3.2.1 Efficient Event-based Meshflow Network" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.894, + 0.78 + ], + "angle": 0, + "content": "We propose EEMFlow to directly output results of the same resolution as the ground truth meshflow \\(MF_{GT}\\) for supervised regression, fully leveraging the advantage about low-parameter and high-motion-information of meshflow." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.781, + 0.895, + 0.902 + ], + "angle": 0, + "content": "Overall Structure of Meshflow Estimation. Since meshflow focuses more on global large motion rather than local detailed motion, the EEMFlow we designed does not require excessively deep network layers or refinement operations from coarse to fine. Firstly, EEMFlow employs a three-level pyramid feature encoder to extracting features \\((V_{t_{k - 1}}^i)^N\\) and \\((V_{t_k}^i)^N\\) from \\(V_{k - 1\\rightarrow k}\\) and \\(V_{k\\rightarrow k + 1}\\), the convolutional layers within the \\(i\\)-th level share weights for" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "19200" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.082, + 0.109, + 0.099, + 0.121 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image_caption", + "bbox": [ + 0.082, + 0.157, + 0.099, + 0.169 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image_caption", + "bbox": [ + 0.082, + 0.206, + 0.098, + 0.217 + ], + "angle": 0, + "content": "(c)" + }, + { + "type": "image", + "bbox": [ + 0.182, + 0.09, + 0.261, + 0.137 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.264, + 0.091, + 0.367, + 0.137 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.369, + 0.091, + 0.471, + 0.137 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.473, + 0.091, + 0.577, + 0.137 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.578, + 0.091, + 0.68, + 0.137 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.681, + 0.091, + 0.784, + 0.137 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.786, + 0.091, + 0.89, + 0.137 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.102, + 0.139, + 0.181, + 0.186 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.183, + 0.139, + 0.262, + 0.186 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.265, + 0.139, + 0.367, + 0.186 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.369, + 0.139, + 0.471, + 0.186 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.473, + 0.139, + 0.576, + 0.186 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.578, + 0.139, + 0.68, + 0.186 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.681, + 0.139, + 0.785, + 0.186 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.787, + 0.139, + 0.89, + 0.186 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.102, + 0.206, + 0.125, + 0.217 + ], + "angle": 0, + "content": "=" + }, + { + "type": "image_caption", + "bbox": [ + 0.102, + 0.235, + 0.162, + 0.245 + ], + "angle": 0, + "content": "MVSEC" + }, + { + "type": "image", + "bbox": [ + 0.182, + 0.188, + 0.262, + 0.236 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.208, + 0.237, + 0.24, + 0.245 + ], + "angle": 0, + "content": "DSEC" + }, + { + "type": "image", + "bbox": [ + 0.265, + 0.188, + 0.368, + 0.235 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.37, + 0.188, + 0.472, + 0.235 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.474, + 0.188, + 0.576, + 0.235 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.578, + 0.188, + 0.681, + 0.235 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.682, + 0.188, + 0.785, + 0.235 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.787, + 0.188, + 0.89, + 0.235 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.546, + 0.236, + 0.612, + 0.246 + ], + "angle": 0, + "content": "HREM(Ours)" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.254, + 0.894, + 0.292 + ], + "angle": 0, + "content": "Figure 2. Examples from MVSEC [53], DSEC [10] and our dataset. (a): Event data, (b): Optical flow, (c): Upsampled meshflow. The meshflows of MVSEC and DSEC contain a significant number of ineffective areas due to the sparsity of optical flow. In contrast, our meshflow provides a complete global motion field. Best viewed on a color screen in high resolution." + }, + { + "type": "image", + "bbox": [ + 0.08, + 0.315, + 0.206, + 0.371 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.135, + 0.371, + 0.15, + 0.38 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.208, + 0.316, + 0.335, + 0.371 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.262, + 0.371, + 0.278, + 0.38 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image", + "bbox": [ + 0.337, + 0.316, + 0.468, + 0.371 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.393, + 0.371, + 0.408, + 0.38 + ], + "angle": 0, + "content": "(c)" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.392, + 0.47, + 0.466 + ], + "angle": 0, + "content": "Figure 3. The process of generating meshflow from dense optical flow. (a) Propagate the motion vector of each grid center to the grid vertices. (b) Apply median filter \\( f_{1} \\) to multiple motion vectors of each vertex to select the most appropriate motion for that vertex. (c) Use median filtering \\( f_{2} \\) to smooth the motion field in the mesh grid. For ease of visualization, we present the \\( 8 \\times 8 \\) grid mesh in this paper." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.492, + 0.468, + 0.719 + ], + "angle": 0, + "content": "\\(V_{t_{k-1}}^i\\) and \\(V_{t_k}^i\\). Secondly, EEMFlow utilizes multi-scale feature correlation to builds the cost volumes for meshflow estimation. Features \\((V_{t_{k-1}}^i)^N\\) and \\((V_{t_k}^i)^N\\) undergo average pooling operation \\(\\mathcal{P}\\) to the same resolution (1/64 resolution of \\(V_{k\\rightarrow k+1}\\)), and then use correlation to capture relative motion information and output cost volumes \\((C^i)^N\\). Specifically, we employ the dilated feature correlation (DFC) to increase the search area while reducing computational parameters. Finally, we stack the cost volumes \\((C^i)^N\\) and features \\((V_{t_{k-1}}^i)^N\\) after pooling, fuse them with a weighted-sum operation, and then feed them into the decoders to regress the meshflow \\(MF_{k\\rightarrow k+1}\\) at the same resolution as \\(MF_{GT}\\). Inspiring by [51], we replace the conventional convolutions with group shuffle convolutions, which leads to efficient computation while maintaining high accuracy." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.72, + 0.468, + 0.764 + ], + "angle": 0, + "content": "Dilated Feature Correlation. We use the inner product to calculate correlation between \\( V_{t_{k-1}}^i \\) and \\( V_{t_k}^i \\) for meshflow estimation:" + }, + { + "type": "equation", + "bbox": [ + 0.105, + 0.778, + 0.469, + 0.797 + ], + "angle": 0, + "content": "\\[\nC ^ {i} (\\boldsymbol {u}, \\boldsymbol {d}) = V _ {t _ {k}} ^ {i} (\\boldsymbol {u}) \\cdot V _ {t _ {k - 1}} ^ {i} (\\boldsymbol {u} + \\boldsymbol {d}) / M, \\boldsymbol {d} \\in \\mathcal {N}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.81, + 0.469, + 0.902 + ], + "angle": 0, + "content": "where \\(\\pmb{u}\\) represents the spatial coordinates on \\(V_{t_k}^i\\), \\(\\mathcal{N}\\) represents the search grid of coordinate \\(\\pmb{u}\\) in feature \\(V_{t_{k - 1}}^i\\), \\(M\\) represents the number of elements in \\(\\mathcal{N}\\), \\(\\pmb{d}\\) represents the offset coordinates of the elements in \\(\\mathcal{N}\\), and \\(\\cdot\\) represents the inner product calculation. Many methods like [16, 40] simply define \\(\\mathcal{N}\\) as a square range of size \\((2r + 1)\\times (2r + 1)\\)" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.318, + 0.892, + 0.394 + ], + "angle": 0, + "content": "and observe that increasing the radius \\( r \\) can reduce errors but increase computational overhead. We propose dilated feature correlation that samples densely around the center and sparsely at farther distances, thereby reducing computation while enabling larger radius cost volumes, as Eq. 2:" + }, + { + "type": "equation", + "bbox": [ + 0.515, + 0.415, + 0.892, + 0.455 + ], + "angle": 0, + "content": "\\[\n\\mathcal {N} \\left(d _ {x}, d _ {y}\\right) = \\left\\{ \\begin{array}{l} 0, \\text {i f} | d _ {x} | + | d _ {y} | = 2 k, k \\in [ 2, r ], \\\\ 1, \\text {o t h e r s ,} \\end{array} \\right. \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.462, + 0.892, + 0.523 + ], + "angle": 0, + "content": "where \\( \\pmb{d} = (d_x, d_y) \\) represents the relative position coordinates within the neighborhood \\( \\mathcal{N} \\), where \\( \\mathcal{N}(d_x, d_y) = 1 \\) indicates the computation of correlation, and \\( \\mathcal{N}(d_x, d_y) = 0 \\) signifies no computation." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.541, + 0.799, + 0.557 + ], + "angle": 0, + "content": "3.2.2 Event-based Optical Flow Network" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.565, + 0.892, + 0.656 + ], + "angle": 0, + "content": "Based on EEMFlow, we make some improvements for accurate estimation of optical flow \\( F_{k \\to k + 1} \\), upgrading it to EEMFlow+. Since optical flow focuses more on local motion and pays attention to object edge details, we employ the coarse to fine residual approach to progressively refine the flow, which can be expressed as Eq. 3." + }, + { + "type": "equation", + "bbox": [ + 0.54, + 0.663, + 0.891, + 0.683 + ], + "angle": 0, + "content": "\\[\nF ^ {i + 1} = \\operatorname {C o n v} ^ {i} \\left(\\mathcal {C} \\left(V _ {t _ {k - 1}} ^ {i}, \\mathcal {W} \\left(V _ {t _ {k}} ^ {i}, F _ {\\uparrow} ^ {i}\\right)\\right)\\right) + F _ {\\uparrow} ^ {i}, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.689, + 0.892, + 0.856 + ], + "angle": 0, + "content": "where \\(\\mathcal{C}(\\cdot, \\cdot), \\mathcal{W}(\\cdot, \\cdot)\\) and \\(\\uparrow\\) donates our dilated feature correlation, the warping operation and upsampling, respectively. We employ the pyramid decoders for optical flow, thus \\(F^i\\) is the output flow of the \\(i\\)-th decoder, \\(F^{i+1}\\) and \\(\\mathrm{Conv}^i\\) are respectively the output flow and convolutions in the \\(i+1\\)-th decoder. The most noteworthy aspect is upsampling \\(F^i\\) to \\(F_{\\uparrow}^i\\). Many methods [6, 16, 40, 52] use bilinear interpolation for upsampling, but this can lead to the mixing of incorrect motions at object edges, resulting in blurring. Therefore, we propose the confidence-induced detail completion module for upsampling to enhance edge details." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.871, + 0.892, + 0.902 + ], + "angle": 0, + "content": "Confidence-induced Detail Completion Module. We propose the confidence-induced detail completion module" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "19201" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.081, + 0.09, + 0.894, + 0.232 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.24, + 0.895, + 0.292 + ], + "angle": 0, + "content": "Figure 4. Our proposed network architecture. We employ pyramid encoders to extract multi-scale features from \\( V_{k-1 \\to k} \\) and \\( V_{k \\to k+1} \\), then use dilated feature correlation to compute the cost volume between each layer of features, followed by decoding to output the results. For meshflow estimation, we utilize the decoders with group shuffle convolutions to output predictions \\( MF_{k \\to k+1} \\), upsampled to \\( MF \\uparrow_{k \\to k+1} \\). For optical flow estimation, we refine flows using a coarse-to-fine residual approach and confidence-induced detail completion module, finally outputting the optical flow \\( F_{k \\to k+1} \\)." + }, + { + "type": "image", + "bbox": [ + 0.08, + 0.314, + 0.47, + 0.46 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.469, + 0.471, + 0.52 + ], + "angle": 0, + "content": "Figure 5. The structure of CDC. CDC employs self-corrector based on a dense convolutional network and self-correlation based on a self-attention mechanism to correct the flow obtained from bilinear upsampling. \\(\\mathcal{B}\\) represents bilinear upsampling." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.544, + 0.473, + 0.834 + ], + "angle": 0, + "content": "(CDC) to eliminate the blurring of object edges caused by the mixing of multiple motions at the junctions of different movements during upsampling. The detailed structure of our CDC is shown in Fig. 5. Given a small-scale flow \\( F^i \\) from the \\( i \\)-th level, our CDC first generates an initial flow \\( \\bar{F}^i \\) for the \\( i + 1 \\)-th level through bilinear interpolation, and then arranges the self-corrector and self-correlation branches to correct it. Self-corrector is based on a dense convolutional network with a five-layer structure. It captures motion information within the edge neighborhood through dense convolution from the concatenated feature \\( V_{t_{k-1}}^i \\) and \\( V_{t_k}^i \\), outputting the corrected flow \\( \\Delta F^i \\) and the corrected confidence map \\( W^i \\). Self-correlation is based on a self-attention mechanism, using a large receptive field to find the fine regions in features \\( V_{t_k}^i \\) that are identical to the motion of the error region in \\( \\bar{F}^i \\). It outputs self-attention weights \\( A^i \\), multiplied with the initial flow \\( \\bar{F}^i \\). With the corrected flow \\( \\Delta F^i \\) and self-attention weights \\( A^i \\), we can generate the fine flow \\( \\tilde{F}^i \\):" + }, + { + "type": "equation", + "bbox": [ + 0.13, + 0.841, + 0.47, + 0.862 + ], + "angle": 0, + "content": "\\[\n\\tilde {F} ^ {i} = \\alpha \\mathcal {W} (\\bar {F} ^ {i}, \\Delta F ^ {i}) + (1 - \\alpha) \\left(A ^ {i} \\otimes \\bar {F} ^ {i}\\right), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.471, + 0.903 + ], + "angle": 0, + "content": "where \\(\\mathcal{W}(\\cdot, \\cdot)\\) means the warping operation, \\(\\otimes\\) donates multiplication and \\(\\alpha \\in [0,1]\\) is the weight coefficient. We iden" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.316, + 0.893, + 0.363 + ], + "angle": 0, + "content": "tify error-prone object edge areas based on the corrected confidence map \\( W^{i} \\), as weights to fuse the initial flow \\( \\bar{F}^{i} \\) and the fine flow \\( \\tilde{F}^{i} \\), obtaining the final corrected flow \\( F_{\\uparrow}^{i} \\):" + }, + { + "type": "equation", + "bbox": [ + 0.579, + 0.371, + 0.893, + 0.391 + ], + "angle": 0, + "content": "\\[\nF _ {\\uparrow} ^ {i} = W ^ {i} \\odot \\bar {F} ^ {i} + (1 - W ^ {i}) \\odot \\tilde {F} ^ {i}, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.398, + 0.804, + 0.414 + ], + "angle": 0, + "content": "where \\(\\odot\\) donates the element-wise multiplier." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.432, + 0.654, + 0.446 + ], + "angle": 0, + "content": "3.2.3 Loss Function" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.455, + 0.893, + 0.577 + ], + "angle": 0, + "content": "For both meshflow and optical flow estimation, we use L1 loss for supervised regression during training. Our EEMFlow, used for estimating meshflow, directly outputs results at the same resolution as the meshflow GT, allowing for direct loss calculation. Similarly, the results outputted by our EEMFlow+ for optical flow estimation are at the same resolution as the network input, enabling direct calculation with the optical flow GT." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.588, + 0.634, + 0.605 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.613, + 0.719, + 0.629 + ], + "angle": 0, + "content": "4.1. Implementation Details" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.636, + 0.615, + 0.65 + ], + "angle": 0, + "content": "4.1.1 Datasets" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.659, + 0.895, + 0.902 + ], + "angle": 0, + "content": "Our HREM dataset includes 100 indoor and outdoor scenes, with a resolution of \\(1280 \\times 720\\). We randomly select 70 scenes for training and reserve the remaining 30 for testing. The training set comprises 20,000 samples, while the test set contains 8,000 samples. Additionally, we further divide the test set by scene type (outdoor vs. indoor) and camera motion speed during rendering, resulting in four sub-sequences (outdoor Slow, outdoor fast, indoor slow, indoor fast), with mean motion magnitudes ranging from, \\(0 - 30\\), \\(30 - 100\\), \\(0 - 20\\) and \\(20 - 100\\) pixels, respectively. Moreover, similar to [52], we employ two data input modes: \\(dt = 1\\) and \\(dt = 4\\). \\(dt = 1\\) uses the event sequence between two consecutive frames of RGB images as input, with a meshflow generation frequency of \\(60~\\mathrm{Hz}\\) while \\(dt = 4\\) uses the event sequence between four consecutive frames of RGB images as input, with a meshflow" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "19202" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.089, + 0.895, + 0.138 + ], + "angle": 0, + "content": "Table 2. Quantitative comparison of our EEMFlow with other advanced flow networks on our HREM dataset. The evaluation metric used is End-Point Error (EPE). \"Parameters\" and \"Time\" respectively indicate the network parameter count and inference time. \\( \\Delta P \\) and \\( \\Delta T \\) represent the change in parameter count and inference time relative to ERAFT [42] for other networks. Smaller values are desirable for all metrics. We highlight the best results in red and the second-best results in blue." + }, + { + "type": "table", + "bbox": [ + 0.162, + 0.149, + 0.812, + 0.395 + ], + "angle": 0, + "content": "
Method dt = 1Parameters (M)Time (ms)OutdoorIndoorAvg
SlowFastSlowFast
EVFlownet [52]38.2463.5516.162.9311.658.57
PWCNet [40]3.36423.9114.492.8611.898.29
ERAFT [11]5.27934.1513.322.9110.347.68
SKFlow [41]6.281453.7611.787.248.817.24
GMA [17]5.891082.1812.072.029.346.40
KPAFlow [32]6.001842.0312.251.959.026.31
FlowFormer [15]9.872812.0611.711.888.666.08
EEMFlow(Ours)1.2472.429.092.008.465.50
Method dt = 4ΔP (M)ΔT (ms)OutdoorIndoorAvg
SlowFastSlowFast
EVFlownet [52]+624%+51%18.2549.3216.1647.1932.73
PWCNet [40]-36%-55%16.4046.1714.4940.9029.49
ERAFT [11]0%0%15.2140.8313.3239.6127.24
SKFlow [41]+19%+56%14.9339.2411.7139.2226.28
GMA [17]+11%+16%14.1338.8912.0737.6825.69
KPAFlow [32]+14%+99%14.0438.0312.2537.2025.38
FlowFormer [15]+88%+202%13.8938.5510.7738.5325.44
EEMFlow(Ours)-76%-92%13.9737.3312.0934.3924.45
" + }, + { + "type": "image", + "bbox": [ + 0.084, + 0.406, + 0.207, + 0.516 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.101, + 0.516, + 0.187, + 0.527 + ], + "angle": 0, + "content": "Image Overlaid" + }, + { + "type": "image", + "bbox": [ + 0.221, + 0.407, + 0.343, + 0.461 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.221, + 0.462, + 0.343, + 0.516 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.25, + 0.516, + 0.314, + 0.526 + ], + "angle": 0, + "content": "EVFlownet" + }, + { + "type": "image", + "bbox": [ + 0.358, + 0.407, + 0.479, + 0.461 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.358, + 0.462, + 0.48, + 0.516 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.396, + 0.516, + 0.441, + 0.526 + ], + "angle": 0, + "content": "ERRAFT" + }, + { + "type": "image", + "bbox": [ + 0.494, + 0.407, + 0.615, + 0.461 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.494, + 0.462, + 0.616, + 0.515 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.527, + 0.516, + 0.582, + 0.526 + ], + "angle": 0, + "content": "KPAFlow" + }, + { + "type": "image", + "bbox": [ + 0.63, + 0.407, + 0.752, + 0.461 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.63, + 0.462, + 0.752, + 0.516 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.676, + 0.516, + 0.706, + 0.526 + ], + "angle": 0, + "content": "Ours" + }, + { + "type": "image", + "bbox": [ + 0.767, + 0.407, + 0.888, + 0.461 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.767, + 0.462, + 0.888, + 0.515 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.786, + 0.515, + 0.869, + 0.527 + ], + "angle": 0, + "content": "Meshflow(GT)" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.538, + 0.894, + 0.565 + ], + "angle": 0, + "content": "Figure 6. Subjective results of image registration using meshflow estimated by other methods and ours. We estimate meshflow from the event sequences and then warp image \\( I_{t_1} \\) onto image \\( I_{t_2} \\) for fusion, showcasing the fused result. The fewer blue or red ghosting artifacts indicate better alignment performance." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.59, + 0.47, + 0.666 + ], + "angle": 0, + "content": "generation frequency of \\(15\\mathrm{Hz}\\). DSEC [10] is a commonly used dataset for event-based optical flow estimation, which consists of real-world data captured by real event cameras mounted on cars. We also conduct experiments on DSEC to compare the performance for optical flow estimation." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.684, + 0.242, + 0.699 + ], + "angle": 0, + "content": "4.1.2 Training details" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.708, + 0.469, + 0.813 + ], + "angle": 0, + "content": "We conduct experiments using the PyTorch framework on two NVIDIA 2080Ti GPUs. We train all networks with the same parameters on our HRDM dataset. We employ the AdamW optimizer and OneCycle policy with a learning rate of \\(5 \\times 10^{-4}\\), weight decay of \\(5 \\times 10^{-5}\\), and other default parameters set to \\(\\beta_{1} = 0.9\\), \\(\\beta_{2} = 0.99\\), \\(\\epsilon = 1 \\times 10^{-4}\\). We train all networks up to \\(100k\\) iterations to reach convergence." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.832, + 0.264, + 0.846 + ], + "angle": 0, + "content": "4.1.3 Evaluation metrics" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.47, + 0.901 + ], + "angle": 0, + "content": "Following EV-FlowNet [52], we use the average Endpoint Error (EPE) as the metric. When evaluating meshflow, we upsample both the prediction and the ground truth to the" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.59, + 0.893, + 0.666 + ], + "angle": 0, + "content": "input resolution to calculate metrics. In addition, the DSEC dataset uses \\( N \\)PE to measure the percentage of flow errors higher than \\( N \\) pixels in magnitude (e.g., 3PE, 2PE, 1PE) for flow outliers analysis, and employs the Angular Error (AE) to assess the directional accuracy." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.673, + 0.805, + 0.689 + ], + "angle": 0, + "content": "4.2. Comparison with State-of-the-Arts" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.697, + 0.871, + 0.711 + ], + "angle": 0, + "content": "4.2.1 Results for Event-based Meshflow Estimation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.894, + 0.901 + ], + "angle": 0, + "content": "In Table 2, We train and test our EEMFlow on our HREM dataset, compared with some advanced networks, e.g., EVFlownet [54], ERAFT [11], PwCNet [40], SKFlow [41], GMA [17], KPAFlow [32], and FlowFormer [15]. Since these networks are all structured to output prediction with the same resolution of input, we upsample the ground truth \\(MF_{GT}\\) to supervise these networks during training. However, EEMFlow can directly output prediction aligned \\(MF_{GT}\\) in resolution. But for fair comparison, we upsample the predictions of all networks to a uniform resolution during evaluation. We train and test all networks using two input modes (\\(dt = 1\\) and \\(dt = 4\\)), and" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19203" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.083, + 0.089, + 0.208, + 0.145 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.084, + 0.146, + 0.207, + 0.2 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.219, + 0.091, + 0.343, + 0.145 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.219, + 0.145, + 0.344, + 0.2 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.356, + 0.091, + 0.479, + 0.145 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.356, + 0.145, + 0.48, + 0.2 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.493, + 0.091, + 0.616, + 0.145 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.493, + 0.145, + 0.616, + 0.2 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.629, + 0.091, + 0.752, + 0.145 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.629, + 0.145, + 0.752, + 0.2 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.765, + 0.091, + 0.888, + 0.145 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.765, + 0.145, + 0.888, + 0.2 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.084, + 0.206, + 0.207, + 0.262 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.084, + 0.262, + 0.207, + 0.317 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.117, + 0.318, + 0.174, + 0.328 + ], + "angle": 0, + "content": "Reference" + }, + { + "type": "image", + "bbox": [ + 0.219, + 0.206, + 0.343, + 0.262 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.219, + 0.262, + 0.343, + 0.317 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.25, + 0.318, + 0.314, + 0.328 + ], + "angle": 0, + "content": "EVFlownet" + }, + { + "type": "image", + "bbox": [ + 0.356, + 0.206, + 0.479, + 0.262 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.356, + 0.262, + 0.479, + 0.317 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.395, + 0.318, + 0.439, + 0.328 + ], + "angle": 0, + "content": "ERRAFT" + }, + { + "type": "image", + "bbox": [ + 0.493, + 0.206, + 0.616, + 0.262 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.493, + 0.262, + 0.616, + 0.317 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.526, + 0.318, + 0.581, + 0.328 + ], + "angle": 0, + "content": "KPAFlow" + }, + { + "type": "image", + "bbox": [ + 0.629, + 0.206, + 0.752, + 0.262 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.629, + 0.262, + 0.752, + 0.317 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.675, + 0.318, + 0.704, + 0.328 + ], + "angle": 0, + "content": "Ours" + }, + { + "type": "image", + "bbox": [ + 0.765, + 0.207, + 0.887, + 0.263 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.765, + 0.263, + 0.887, + 0.317 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.785, + 0.318, + 0.867, + 0.329 + ], + "angle": 0, + "content": "Meshflow(GT)" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.345, + 0.892, + 0.37 + ], + "angle": 0, + "content": "Figure 7. Qualitative comparison of our proposed EEMFlow with other advanced flow networks on our HREM dataset. The subjective images of events and dense optical flow on the left side serve as references. The areas enclosed by red rectangles are zoomed in." + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.388, + 0.469, + 0.412 + ], + "angle": 0, + "content": "Table 3. Results on DSEC dataset. All networks are trained and tested on the DSEC dataset for optical flow estimation." + }, + { + "type": "table", + "bbox": [ + 0.085, + 0.423, + 0.462, + 0.549 + ], + "angle": 0, + "content": "
MethodsFPS↑1PE↓2PE↓3PE↓EPE↓AE↓
MutilCM [39]-76.648.530.93.4714.0
EV-Flownet [52]22.355.429.818.62.328.12
OF-EV-SNN [4]-53.720.210.31.716.34
EVA-Flow [50]-15.9-3.200.883.31
ERRAFT [11]11.412.74.742.680.792.85
ADMFlow [33]9.8812.54.672.650.782.84
EFlowformer [21]-11.24.102.450.762.68
TMA [25]7.5510.93.972.300.742.68
EEMFlow+(Ours)39.211.43.932.150.752.67
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.581, + 0.469, + 0.641 + ], + "angle": 0, + "content": "present their performance metrics in four test sub-sequences (outdoor Slow, outdoor fast, indoor slow, indoor fast). In addition, we average the metric scores across all four test sub-sequences as shown in the \"Avg\" column." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.645, + 0.47, + 0.901 + ], + "angle": 0, + "content": "Table 2 shows that our EEMFlow achieves the lowest EPE score for both input modes \\( dt = 1 \\) and \\( dt = 4 \\) in the \"outdoor_fast\" and \"indoor_fast\" test sub-sequences, demonstrating its superior performance on high-speed and large-movement sequences. EEMFlow also exhibits great potential to outperform other flow networks on the \"outdoor Slow\" and \"indoor Slow\" test sub-sequences. Notably, EEMFlow achieves the lowest EPE scores in the \"Avg\" column for both input modes \\( dt = 1 \\) and \\( dt = 4 \\). Besides, we also achieve the least number of parameters and the fastest inference speed. Compared to ERAFT [11], our EEMFlow reduces the parameter count by \\( 76\\% \\) (from 5.27M to 1.24M), reduces the inference time by \\( 92\\% \\) (from 93ms to 7ms), and improves average EPE in \\( dt = 4 \\) by \\( 8\\% \\) (from 27.24 to 25.18). EEMFlow achieves comparable EPE performance to FlowFormer [15] but exhibits a \\( 38.7\\times \\) increase in inference speed." + }, + { + "type": "image", + "bbox": [ + 0.504, + 0.388, + 0.89, + 0.434 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.504, + 0.434, + 0.889, + 0.479 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.504, + 0.479, + 0.889, + 0.525 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.504, + 0.525, + 0.889, + 0.57 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.517, + 0.57, + 0.566, + 0.579 + ], + "angle": 0, + "content": "Reference" + }, + { + "type": "image_caption", + "bbox": [ + 0.592, + 0.569, + 0.647, + 0.579 + ], + "angle": 0, + "content": "EVFlownet" + }, + { + "type": "image_caption", + "bbox": [ + 0.68, + 0.569, + 0.719, + 0.579 + ], + "angle": 0, + "content": "ERRAFT" + }, + { + "type": "image_caption", + "bbox": [ + 0.762, + 0.569, + 0.79, + 0.579 + ], + "angle": 0, + "content": "TMA" + }, + { + "type": "image_caption", + "bbox": [ + 0.837, + 0.569, + 0.862, + 0.578 + ], + "angle": 0, + "content": "Ours" + }, + { + "type": "image_caption", + "bbox": [ + 0.5, + 0.59, + 0.891, + 0.616 + ], + "angle": 0, + "content": "Figure 8. Qualitative comparisons on the DSEC test set. We visualize the dense predictions and zoom in the areas where are apparent differences." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.644, + 0.892, + 0.901 + ], + "angle": 0, + "content": "In Fig. 7, we qualitatively compare our proposed EEMFlow with other flow networks, e.g., EVFlownet [54], ER-AFT [11], and KPAFlow [32]. To facilitate comparison, we upsample the meshflow estimation and ground truth to the same resolution of input and zoom in on the areas with the apparent differences. EVFlownet shows the worst performance, with many holes and color mixing. ERAFT and KPAFlow would exhibit block artifacts and appear coarse in nature. In contrast, our EEMFlow results are smoother with more natural color transitions and greater similarity to the upsampled ground truth. In Fig. 6, we present the subjective results of these networks for image registration. We estimate and upsample meshflow prediction from the event sequences \\(E(t_{1},t_{2})\\) and then warp image \\(I_{t_1}\\) onto image \\(I_{t_2}\\). We also zoom in on the challenging areas of alignment in the registered results, clearly demonstrating that our EEMFlow achieves excellent image registration performance." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19204" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.081, + 0.09, + 0.466, + 0.187 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.111, + 0.192, + 0.435, + 0.204 + ], + "angle": 0, + "content": "Figure 9. Visualization for meshflow and optical flow results." + }, + { + "type": "table_caption", + "bbox": [ + 0.091, + 0.225, + 0.454, + 0.237 + ], + "angle": 0, + "content": "Table 4. Results of meshflow and optical flow on HREM for \\( dt = 1 \\)" + }, + { + "type": "table", + "bbox": [ + 0.089, + 0.249, + 0.456, + 0.327 + ], + "angle": 0, + "content": "
TaskMethodOutdoorIndoorAvg
SlowFastSlowFast
Optical FlowFlowFormer6.2016.065.9915.2710.88
EEMFlow+3.8811.024.0310.927.46
Mesh-flowFlowFormer5.9915.125.7414.9510.45
EEMFlow2.429.092.008.465.50
" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.358, + 0.453, + 0.373 + ], + "angle": 0, + "content": "4.2.2 Results for event-based optical flow estimation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.387, + 0.47, + 0.629 + ], + "angle": 0, + "content": "Table 3 presents the comparative results on DSEC dataset of our optical flow estimation network, EEMFlow+, against other event-based optical flow estimation networks. EEMFlow+ achieves state-of-the-art performance in the 2PE, 3PE, and AE metrics, and is on par with the best in the EPE metric. Notably, EEMFlow+ still maintains a significant advantage in inference speed. Compared to TMA [25], our EEMFlow+ increases the inference speed by \\(419\\%\\) (from 7.55FPS to 39.2FPS). Fig. 8 displays the qualitative results of our optical flow estimation on the DSEC test set, compared with other advanced methods, including EVFlownet [52], ERAFT [42], and TMA [25]. By zooming into the object areas, it is clearly observable that our results exhibit more regular shapes and clearer boundaries, highlighting the detail enhancement capability of our proposed CDC module." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.649, + 0.241, + 0.663 + ], + "angle": 0, + "content": "4.3. Ablation Studies" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.675, + 0.458, + 0.69 + ], + "angle": 0, + "content": "4.3.1 The Advantages of Event-Meshflow Estimation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.705, + 0.47, + 0.901 + ], + "angle": 0, + "content": "According to the ERAFT [42], image-based approaches face challenges in handling difficult images due to the limited dynamic range of image sensors. As shown in Fig. 9, even advanced image-based techniques like Flow-Former struggle, while our event-based methods (EEMFlow, EEMFlow+) exhibit notably superior performance. Additional quantitative comparisons are provided in Table 4, our network of event-based meshflow estimation EEMFlow achieves the lowest EPE scores in HREM dataset for the \\( dt = 1 \\) input mode, which demonstrates the remarkable efficiency of our EEMFlow makes it suitable for real-time applications such as online video stabilization and autonomous driving." + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.089, + 0.892, + 0.113 + ], + "angle": 0, + "content": "Table 5. Ablation studies about CDC of EEMFlow+. CDC consists of two branches, the self-corrector and self-correlation." + }, + { + "type": "table", + "bbox": [ + 0.508, + 0.125, + 0.882, + 0.201 + ], + "angle": 0, + "content": "
ModelSelf-correctorSelf-correlationFPS↑3PE↓EPE↓AE↓
(a)XX60.43.520.893.11
(b)✓X55.62.770.812.92
(c)X✓46.32.650.792.78
(d)✓✓39.22.150.752.67
" + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.232, + 0.809, + 0.246 + ], + "angle": 0, + "content": "4.3.2 Experiments for CDC of EEMFlow+" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.26, + 0.892, + 0.487 + ], + "angle": 0, + "content": "In Table 5, we also conduct ablation experiments on the CDC module of EEMFlow+ used for optical flow estimation, including its two branches, the self-corrector and self-correlation. We train and evaluate all models using the same settings on the DSEC dataset to show the individual impact of each branch in CDC module. Comparison of (a)&(b) demonstrates that the CDC with only the self-corrector branch can bring a significant increase in accuracy with a minimal loss in speed. Comparison of (b)&(c) shows that self-correlation, compared to self-corrector, can lead to a higher increase in accuracy, albeit at a further reduction in inference speed. Finally, the comparison of (a)&(d) shows that the CDC composed of both the self-corrector and self-correlation branches significantly improves the accuracy of optical flow estimation with an acceptable loss in speed." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.508, + 0.619, + 0.523 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.536, + 0.892, + 0.718 + ], + "angle": 0, + "content": "In this work, we develop the first event-based meshflow dataset, termed HREM, where 100 indoor and outdoor virtual scenes with rich scene contents are rendered using Blender. Our HREM possess physically correct accurate events and meshflow label pairs with so far the highest resolution. Furthermore, we propose an efficient event-based meshflow network (EEMFlow), which achieves state-of-the-art performance on HREM dataset while maintaining high efficiency. Based on EEMFlow, we propose a confidence-induced detail completion module to upgrade it as EEMFlow+ for optical flow estimation, achieving SOTA results on DSEC dataset with the fastest inference speed." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.892, + 0.811 + ], + "angle": 0, + "content": "The integration of optical flow and depth data allows for the extension of the proposed framework to address 3D scene flow. Subsequently, our future work will involve expanding this methodology on the HREM dataset by integrating scene flow annotations and investigating scene flow estimation using event cameras." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.826, + 0.892, + 0.9 + ], + "angle": 0, + "content": "Acknowledgements. This work was supported by National Natural Science Foundation of China (NSFC) under grant No.62372091, Sichuan Science and Technology Program of China under grants Nos. 2023NSFSC0462, 2022YFQ0079 and 2023YFG0054." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "19205" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.169 + ], + "angle": 0, + "content": "[1] Ryad Benosman, Sio-Hoi Ieng, Charles Clercq, Chiara Bartolozzi, and Mandyam Srinivasan. Asynchronous frameless event-based optical flow. Neural Networks, 27:32-37, 2012. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.173, + 0.472, + 0.227 + ], + "angle": 0, + "content": "[2] Christian Brandli, Raphael Berner, Minhao Yang, Shih-Chii Liu, and Tobi Delbruck. A \\(240 \\times 180\\) 130 db \\(3\\mu s\\) latency global shutter spatiotemporal vision sensor. IEEE Journal of Solid-State Circuits, 49(10):2333-2341, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.23, + 0.471, + 0.286 + ], + "angle": 0, + "content": "[3] Marco Cannici, Marco Ciccone, Andrea Romanoni, and Matteo Matteucci. A differentiable recurrent surface for asynchronous event-based data. In Proc. ECCV, pages 136-152. Springer, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.288, + 0.471, + 0.343 + ], + "angle": 0, + "content": "[4] Javier Cuadrado, Ulysse Rançon, Benoit R Cotterreau, Francisco Barranco, and Timothee Masquelier. Optical flow estimation from event-based cameras and spiking neural networks. Frontiers in Neuroscience, 17:1160034, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.345, + 0.471, + 0.401 + ], + "angle": 0, + "content": "[5] Ziluo Ding, Rui Zhao, Jiyuan Zhang, Tianxiao Gao, Ruiqin Xiong, Zhaofei Yu, and Tiejun Huang. Spatio-temporal recurrent networks for event-based optical flow estimation. In Proc. AAAI, pages 525-533, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.403, + 0.471, + 0.471 + ], + "angle": 0, + "content": "[6] Alexey Dosovitskiy, Philipp Fischer, Eddy Ilg, Philip Hausser, Caner Hazirbas, Vladimir Golkov, Patrick Van Der Smagt, Daniel Cremers, and Thomas Brox. Flownet: Learning optical flow with convolutional networks. In Proc. ICCV, pages 2758-2766, 2015. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.473, + 0.471, + 0.543 + ], + "angle": 0, + "content": "[7] Guillermo Gallego, Tobi Delbrück, Garrick Orchard, Chiara Bartolozzi, Brian Taba, Andrea Censi, Stefan Leutenegger, Andrew J Davison, Jörg Conradt, Kostas Daniilidis, et al. Event-based vision: A survey. IEEE transactions on pattern analysis and machine intelligence, 44(1):154-180, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.545, + 0.471, + 0.6 + ], + "angle": 0, + "content": "[8] Daniel Gehrig, Antonio Loquercio, Konstantinos G Derpanis, and Davide Scaramuzza. End-to-end learning of representations for asynchronous event-based data. In Proc. ICCV, pages 5633-5643, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.602, + 0.471, + 0.657 + ], + "angle": 0, + "content": "[9] Daniel Gehrig, Mathias Gehrig, Javier Hidalgo-Carrió, and Davide Scaramuzza. Video to events: Recycling video datasets for event cameras. In Proc. CVPR, pages 3586-3595, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.66, + 0.47, + 0.715 + ], + "angle": 0, + "content": "[10] Mathias Gehrig, Willem Aarents, Daniel Gehrig, and Davide Scaramuzza. Dsec: A stereo event camera dataset for driving scenarios. IEEE Robotics and Automation Letters, 6(3): 4947-4954, 2021. 2, 3, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.717, + 0.47, + 0.771 + ], + "angle": 0, + "content": "[11] Mathias Gehrig, Mario Millhäuser, Daniel Gehrig, and Davide Scaramuzza. E-raft: Dense optical flow from event cameras. In International Conference on 3D Vision (3DV), pages 197–206, 2021. 1, 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.774, + 0.47, + 0.843 + ], + "angle": 0, + "content": "[12] Klaus Greff, Francois Belletti, Lucas Beyer, Carl Doersch, Yilun Du, Daniel Duckworth, David J Fleet, Dan Gnanapragasam, Florian Golemo, Charles Herrmann, et al. Kubric: A scalable dataset generator. In Proc. CVPR, pages 3749-3761, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.846, + 0.47, + 0.902 + ], + "angle": 0, + "content": "[13] Jesse Hagenaars, Federico Paredes-Vallés, and Guido De Croon. Self-supervised learning of event-based optical flow with spiking neural networks. Advances in Neural Information Processing Systems, 34:7167-7179, 2021. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.472, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.134 + ], + "angle": 0, + "content": "[14] Yuhuang Hu, Shih-Chii Liu, and Tobi Delbruck. v2e: From video frames to realistic dvs events. In Proc. CVPR, pages 1312-1321, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.137, + 0.892, + 0.192 + ], + "angle": 0, + "content": "[15] Zhaoyang Huang, Xiaoyu Shi, Chao Zhang, Qiang Wang, Ka Chun Cheung, Hongwei Qin, Jifeng Dai, and Hongsheng Li. FlowFormer: A transformer architecture for optical flow. ECCV, 2022. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.195, + 0.892, + 0.236 + ], + "angle": 0, + "content": "[16] Junhwa Hur and Stefan Roth. Iterative residual refinement for joint optical flow and occlusion estimation. In Proc. CVPR, 2019. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.239, + 0.892, + 0.294 + ], + "angle": 0, + "content": "[17] Shihao Jiang, Dylan Campbell, Yao Lu, Hongdong Li, and Richard Hartley. Learning to estimate hidden motions with global motion aggregation. In Proc. ICCV, pages 9772-9781, 2021. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.297, + 0.892, + 0.408 + ], + "angle": 0, + "content": "[18] Jacques Kaiser, J Camilo Vasquez Tieck, Christian Hubschneider, Peter Wolf, Michael Weber, Michael Hoff, Alexander Friedrich, Konrad Wojtasik, Arne Roennau, Ralf Kohlhaas, et al. Towards a framework for end-to-end control of a simulated vehicle with spiking neural networks. In IEEE International Conference on Simulation, Modeling, and Programming for Autonomous Robots (SIMPAR), pages 127-134, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.411, + 0.892, + 0.481 + ], + "angle": 0, + "content": "[19] Chankyu Lee, Adarsh Kumar Kosta, Alex Zihao Zhu, Kenneth Chaney, Kostas Daniilidis, and Kaushik Roy. Spikeflownet: event-based optical flow estimation with energy-efficient hybrid neural networks. In Proc. ECCV, pages 366-382. Springer, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.483, + 0.892, + 0.552 + ], + "angle": 0, + "content": "[20] Chankyu Lee, Adarsh Kumar Kosta, and Kaushik Roy. Fusion-flownet: Energy-efficient optical flow estimation using sensor fusion and deep fused spiking-analog network architectures. In 2022 International Conference on Robotics and Automation (ICRA), pages 6504-6510. IEEE, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.555, + 0.892, + 0.624 + ], + "angle": 0, + "content": "[21] Yijin Li, Zhaoyang Huang, Shuo Chen, Xiaoyu Shi, Hongsheng Li, Hujun Bao, Zhaopeng Cui, and Guofeng Zhang. Blinkflow: A dataset to push the limits of event-based optical flow estimation. arXiv preprint arXiv:2303.07716, 2023. 2, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.627, + 0.892, + 0.682 + ], + "angle": 0, + "content": "[22] Zhuoyan Li, Jiawei Shen, and Ruitao Liu. A lightweight network to learn optical flow from event data. In 2020 25th International Conference on Pattern Recognition (ICPR), pages 1-7. IEEE, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.686, + 0.892, + 0.741 + ], + "angle": 0, + "content": "[23] Kaimo Lin, Shuaicheng Liu, Loong-Fah Cheong, and Bing Zeng. Seamless video stitching from hand-held camera inputs. In Computer Graphics Forum, pages 479-487. Wiley Online Library, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.743, + 0.892, + 0.799 + ], + "angle": 0, + "content": "[24] Songnan Lin, Ye Ma, Zhenhua Guo, and Bihan Wen. Dvsvoltmeter: Stochastic process-based event simulator for dynamic vision sensors. In Proc. ECCV, pages 578-593. Springer, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.801, + 0.892, + 0.857 + ], + "angle": 0, + "content": "[25] Haotian Liu, Guang Chen, Sanqing Qu, Yanping Zhang, Zhi-jun Li, Alois Knoll, and Changjun Jiang. Tma: Temporal motion aggregation for event-based optical flow. In Proc. ICCV, 2023. 2, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.859, + 0.892, + 0.902 + ], + "angle": 0, + "content": "[26] Shuaicheng Liu, Lu Yuan, Ping Tan, and Jian Sun. Bundled camera paths for video stabilization. ACM transactions on graphics (TOG), 32(4):1-10, 2013. 1" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "19206" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.468, + 0.134 + ], + "angle": 0, + "content": "[27] Shuaicheng Liu, Lu Yuan, Ping Tan, and Jian Sun. Steadyflow: Spatially smooth optical flow for video stabilization. In Proc. CVPR, pages 4209-4216, 2014. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.136, + 0.468, + 0.19 + ], + "angle": 0, + "content": "[28] Shuaicheng Liu, Ping Tan, Lu Yuan, Jian Sun, and Bing Zeng. Meshflow: Minimum latency online video stabilization. In Proc. ECCV, pages 800-815. Springer, 2016. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.194, + 0.468, + 0.248 + ], + "angle": 0, + "content": "[29] Shuaicheng Liu, Nianjin Ye, Chuan Wang, Jirong Zhang, Lanpeng Jia, Kunming Luo, Jue Wang, and Jian Sun. Content-aware unsupervised deep homography estimation and its extensions. 45(3):2849-2863, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.251, + 0.468, + 0.306 + ], + "angle": 0, + "content": "[30] Shuaicheng Liu, Yuhang Lu, Hai Jiang, Nianjin Ye, Chuan Wang, and Bing Zeng. Unsupervised global and local homography estimation with motion basis learning. 45(6): 7885-7899, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.309, + 0.468, + 0.362 + ], + "angle": 0, + "content": "[31] Zhen Liu, Yinglong Wang, Bing Zeng, and Shuaicheng Liu. Ghost-free high dynamic range imaging with context-aware transformer. In Proc. ECCV, pages 344-360. Springer, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.366, + 0.468, + 0.407 + ], + "angle": 0, + "content": "[32] Ao Luo, Fan Yang, Xin Li, and Shuaicheng Liu. Learning optical flow with kernel patch attention. In Proc. CVPR, 2022. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.41, + 0.468, + 0.463 + ], + "angle": 0, + "content": "[33] Xinglong Luo, Kunming Luo, Ao Luo, Zhengning Wang, Ping Tan, and Shuaicheng Liu. Learning optical flow from event camera with rendered dataset. In Proc. ICCV, 2023. 2, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.467, + 0.468, + 0.509 + ], + "angle": 0, + "content": "[34] Lang Nie, Chunyu Lin, Kang Liao, Shuaicheng Liu, and Yao Zhao. Depth-aware multi-grid deep homography estimation with contextual correlation. 32(7):4460-4472, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.511, + 0.468, + 0.552 + ], + "angle": 0, + "content": "[35] Lang Nie, Chunyu Lin, Kang Liao, Shuaicheng Liu, and Yao Zhao. Parallax-tolerant unsupervised deep image stitching. In Proc. ICCV, pages 7399-7408, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.555, + 0.468, + 0.596 + ], + "angle": 0, + "content": "[36] Liyuan Pan, Miaomiao Liu, and Richard Hartley. Single image optical flow estimation with an event camera. In Proc. CVPR, pages 1669-1678, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.599, + 0.468, + 0.64 + ], + "angle": 0, + "content": "[37] Henri Rebecq, Daniel Gehrig, and Davide Scaramuzza. Esim: an open event camera simulator. In Conference on robot learning, pages 969-982. PMLR, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.642, + 0.468, + 0.682 + ], + "angle": 0, + "content": "[38] Zhihang Ren, Jiajia Li, Shuaicheng Liu, and Bing Zeng. Meshflow video denoising. In Proc. ICIP, pages 2966-2970, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.686, + 0.468, + 0.726 + ], + "angle": 0, + "content": "[39] Deqing Sun, Stefan Roth, and Michael J Black. Secrets of optical flow estimation and their principles. In Proc. CVPR, pages 2432-2439, 2010. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.729, + 0.468, + 0.77 + ], + "angle": 0, + "content": "[40] Deqing Sun, Xiaodong Yang, Ming-Yu Liu, and Jan Kautz. Pwc-net: Cnns for optical flow using pyramid, warping, and cost volume. In Proc. CVPR, pages 8934-8943, 2018. 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.773, + 0.468, + 0.813 + ], + "angle": 0, + "content": "[41] Shangkun Sun, Yuanqi Chen, Yu Zhu, Guodong Guo, and Ge Li. Skflow: Learning optical flow with super kernels. Proc. NeurIPS, 35:11313-11326, 2022. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.816, + 0.468, + 0.844 + ], + "angle": 0, + "content": "[42] Zachary Teed and Jia Deng. Raft: Recurrent all-pairs field transforms for optical flow. In Proc. ECCV, 2020. 2, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.846, + 0.468, + 0.899 + ], + "angle": 0, + "content": "[43] Zhexiong Wan, Yuchao Dai, and Yuxin Mao. Learning dense and continuous optical flow from an event camera. IEEE Transactions on Image Processing, 31:7237-7251, 2022. 2, 3" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.468, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.148 + ], + "angle": 0, + "content": "[44] Zhexiong Wan, Yuxin Mao, Jing Zhang, and Yuchao Dai. Rpeflow: Multimodal fusion of rgb-pointcloud-event for joint optical flow and scene flow estimation. In Proc. ICCV, pages 10030-10040, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.151, + 0.892, + 0.19 + ], + "angle": 0, + "content": "[45] Yiming Wang, Qian Huang, Chuanxu Jiang, Jiwen Liu, Mingzhou Shang, and Zhuang Miao. Video stabilization: A comprehensive survey. Neurocomputing, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.193, + 0.892, + 0.234 + ], + "angle": 0, + "content": "[46] Qingsen Yan, Jinqiu Sun, Haisen Li, Yu Zhu, and Yanning Zhang. High dynamic range imaging by sparse representation. Neurocomputing, 269:160-169, 2017. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.236, + 0.892, + 0.29 + ], + "angle": 0, + "content": "[47] Min Yang, Jingkun Liang, Jianhai Zhang, Haidong Gao, Fanyong Meng, Li Xingdong, and Sung-Jin Song. Non-local means theory based perona-malik model for image denoising. Neurocomputing, 120:262-267, 2013. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.293, + 0.892, + 0.319 + ], + "angle": 0, + "content": "[48] Yan Yang, Liyuan Pan, and Liu Liu. Event camera data pretraining. In Proc. ICCV, pages 10699-10709, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.321, + 0.892, + 0.375 + ], + "angle": 0, + "content": "[49] Nianjin Ye, Chuan Wang, Shuaicheng Liu, Lanpeng Jia, Jue Wang, and Yongqing Cui. Deepmeshflow: Content adaptive mesh deformation for robust image registration. arXiv preprint arXiv:1912.05131, 2019. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.378, + 0.892, + 0.431 + ], + "angle": 0, + "content": "[50] Yaozu Ye, Hao Shi, Kailun Yang, Ze Wang, Xiaoting Yin, Yaonan Wang, and Kaiwei Wang. Towards anytime optical flow estimation with event cameras. arXiv preprint arXiv:2307.05033, 2023. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.434, + 0.892, + 0.488 + ], + "angle": 0, + "content": "[51] Xiangyu Zhang, Xinyu Zhou, Mengxiao Lin, and Jian Sun. Shufflenet: An extremely efficient convolutional neural network for mobile devices. In Proc. CVPR, pages 6848-6856, 2018. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.491, + 0.892, + 0.545 + ], + "angle": 0, + "content": "[52] Alex Zhu, Liangzhe Yuan, Kenneth Chaney, and Kostas Daniilidis. Ev-flownet: Self-supervised optical flow estimation for event-based cameras. In Proceedings of Robotics: Science and Systems, 2018. 2, 4, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.548, + 0.892, + 0.615 + ], + "angle": 0, + "content": "[53] Alex Zihao Zhu, Dinesh Thakur, Tolga Ozaslan, Bernd Pfrommer, Vijay Kumar, and Kostas Daniilidis. The multi-vehicle stereo event camera dataset: An event camera dataset for 3d perception. IEEE Robotics and Automation Letters, 3 (3):2032-2039, 2018. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.618, + 0.892, + 0.672 + ], + "angle": 0, + "content": "[54] Alex Zihao Zhu, Liangzhe Yuan, Kenneth Chaney, and Kostas Daniilidis. Unsupervised event-based learning of optical flow, depth, and egomotion. In Proc. CVPR, pages 989-997, 2019. 3, 6, 7" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.672 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19207" + } + ] +] \ No newline at end of file diff --git a/2024/Efficient Meshflow and Optical Flow Estimation from Event Cameras/141a2ee1-d1ef-416d-8c61-bc462b1074b9_origin.pdf b/2024/Efficient Meshflow and Optical Flow Estimation from Event Cameras/141a2ee1-d1ef-416d-8c61-bc462b1074b9_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..8c47a5fa82ac0b8b933855cbd68b17f2f6fba17d --- /dev/null +++ b/2024/Efficient Meshflow and Optical Flow Estimation from Event Cameras/141a2ee1-d1ef-416d-8c61-bc462b1074b9_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9297321ad68a0036ebfc70a2f62ec978dd3f295b4a1f376bddb7597441b36bc2 +size 1303249 diff --git a/2024/Efficient Meshflow and Optical Flow Estimation from Event Cameras/full.md b/2024/Efficient Meshflow and Optical Flow Estimation from Event Cameras/full.md new file mode 100644 index 0000000000000000000000000000000000000000..67f9350904f89aa1c5c4210bab680d63435f3223 --- /dev/null +++ b/2024/Efficient Meshflow and Optical Flow Estimation from Event Cameras/full.md @@ -0,0 +1,424 @@ +# Efficient Meshflow and Optical Flow Estimation from Event Cameras + +Xinglong Luo $^{1,4*}$ , Ao Luo $^{2,4*}$ , Zhengning Wang $^{1\dagger}$ , Chunyu Lin $^{3}$ , Bing Zeng $^{1}$ , and Shuaicheng Liu $^{1,4\dagger}$ $^{1}$ University of Electronic Science and Technology of China + $^{2}$ Southwest Jiaotong University + $^{3}$ Beijing Jiaotong University + $^{4}$ Megvii Technology +{luoboom@std.,zhengning.wang@,eezeng@,liushuaicheng@}uestc.edu.cn +aoluo@swjtu.edu.cn cylin@bjtu.edu.cn + +# Abstract + +In this paper, we explore the problem of event-based meshflow estimation, a novel task that involves predicting a spatially smooth sparse motion field from event cameras. To start, we generate a large-scale High-Resolution Event Meshflow (HREM) dataset, which showcases its superiority by encompassing the merits of high resolution at $1280 \times 720$ , handling dynamic objects and complex motion patterns, and offering both optical flow and meshflow labels. These aspects have not been fully explored in previous works. Besides, we propose Efficient Event-based MeshFlow (EEMFlow) network, a lightweight model featuring a specially crafted encoder-decoder architecture to facilitate swift and accurate meshflow estimation. Furthermore, we upgrade EEMFlow network to support dense event optical flow, in which a Confidence-induced Detail Completion (CDC) module is proposed to preserve sharp motion boundaries. We conduct comprehensive experiments to show the exceptional performance and runtime efficiency $(39 \times$ faster) of our EEMFlow model compared to recent state-of-the-art flow methods. Our code is available at https://github.com/boomluo02/EEMFlow. + +# 1. Introduction + +Meshflow, a spatially smooth sparse motion field, represents motion vectors exclusively at mesh vertices [28, 49], which has been widely applied in various vision applications, such as image alignment [29, 30, 34], video stabilization [26, 27, 45] and high dynamic range (HDR) imaging [31, 46]. This motion representation combines the benefits of optical flow and global homography, effectively reducing redundancy in motion information and computational costs, while also accommodating non-rigid motions beyond single-plane movements. However, meshflow estimation on RGB images often encounter challenges under scenarios such as low-light and rapid motions. This is due + +![](images/6a7187498c849c3a671e8f295bb00308fee81606ff12877504673bf926e9b91d.jpg) +Figure 1. Comparison of computational overhead and accuracy metrics. The x-axis represents inference time, while the y-axis corresponds to the end-point error. The size of each circle indicates the number of model parameters. Lower values for all metrics are considered better. + +to the loss of fine image texture details and motion blurs. + +In contrast, event cameras are well-suited for motion estimation under such situations [7, 11]. Equipped with bio-inspired vision sensors, event cameras can generate sequences of events with microsecond accuracy triggered by changes in log intensity. In particular, when a change is detected in a pixel, the camera returns an event in the form $e = (x,y,t,p)$ immediately, where $x,y$ stands for the spatial location, $t$ refers to the timestamp in microseconds, and $p$ is the polarity of the change, either positive or negative. The advantages of high temporal resolution and high dynamic range make event cameras highly effective for analyzing dynamic scenes. + +In this work, we study a new problem that estimates meshflow from event camera data. To start, we create a large-scale High-Resolution Event Meshflow (HREM) dataset, which contains 20k train and 8k test samples. We build 100 virtual scenes in Blender to render the dataset, + +which can provide accurate physically-based events along with dense meshflow label pairs. Based on the dataset, we further propose an Efficient Event-based MeshFlow (EEMFlow) network to estimate high-resolution meshflow from event data. Unlike recent flow networks relying on recurrent refinement structure [11, 17, 42], our network is developed on an encoder-decoder architecture with multi-scale global optimizing scheme, which can produce full-resolution meshflow with minimal computational overhead. Specifically, our EEMFlow achieves efficiency by employing the lightweight encoder, building cost volume with dilated feature correlation, and using group shuffle convolutions during decoding. We select recent top-ranked flow networks, including ERAFT [11], SK-Flow [41], KPAFlow [32], and FlowFormer [15], which is trained on our HREM dataset. The results show that our approach achieves state-of-the-art performance while maintaining a fast inference speed of 142.9 FPS, outperforming previous works by a relatively large margin (see Fig. 1). + +Additionally, we empirically demonstrate that the proposed new pipeline has the capability to effectively handle various motion patterns. Its lightweight design and runtime efficiency further contribute significantly to the field of optical flow estimation. Specifically, we refine the optical flow progressively during decoding using the coarse to fine residual approach. A Confidence-induced Detail Completion (CDC) module is proposed to preserve motion boundary details during flow upsampling. We also perform comparative experiments with recent event flow networks [21, 25, 33] to illustrate its superiority. The enhanced flow network is referred to as EEMFlow+, demonstrating impressive performance on the reputable DSEC [11] online test benchmark, with the fastest inference speed reaching 39.2 FPS. Our contributions are summarized as: + +- We build HREM, the first event-based meshflow dataset, superior in the highest resolution at $1280 \times 720$ , dynamic scenes, complex motion patterns, as well as physically correct accurate events paired with meshflow and optical flow labels. +- We propose EEMFlow, which achieves SOTA performances when compared to top-ranked optical flow networks when trained on our meshflow dataset. Moreover, it achieves inference speed of 142.9 FPS, which is 25.5 to 38.7 times faster than compared methods. +- We propose CDC, a confidence-induced detail completion module that empowers EEMFlow to make a meaningful contribution to the optical flow community. The upgraded model achieving SOTA performance when compared to representative methods, while also boasting the fastest inference speed to date. + +# 2. Related Work + +# 2.1. Image-based Meshflow Warping + +Meshflow is a lightweight and spatially smooth sparse motion field with motions only located at mesh vertices [28]. Meshflow is different from dense optical flow, where optical flow estimates motions of every pixel of an image while meshflow only concentrates on the global motion, rejecting motions of any dynamic contents. Meshflow is also different from a global homography, where local motions from nonplanar depth variations can be well reflected. Mesh-based methods proofs to be effective in various applications, such as high dynamic range (HDR) imaging [46], burst image denoising [47], video denoising [38], image/video stitching [23, 35] and video stabilization [45]. It is worth noting that directly downsampling the optical flow may yield flow fields of the same resolution as meshflow, but they differ significantly from meshflow and perform poorly in terms of warping effects [49]. Direct downsampling ignores motion outliers from different dynamic objects and the global consistency, which can be effectively addressed by meshflow through motion propagation. + +# 2.2. Event-based Optical Flow Estimation + +Optical flow estimation from event cameras has received significant attention in recent years. Early approaches, such as [1], could only estimate optical flows at the regions where events are triggered. Recently, deep methods can estimate optical flows from event data, even for the regions without triggered events. For example, EV-FlowNet [52] learns event and flow labels in a self-supervised manner by minimizing photometric distances of grey images acquired by DAVIS [2]. Various event representations, including EST [8] and Matrix-LSTM [3], have been explored, and different network structures, such as SpikeFlowNet [19], LIF-EV-FlowNet [13], STE-FlowNet [5], Li et al. [22], ERAFT [11], Yang et al. [48], EVA-Flow [50], ADMFlow [33], E-FlowFormer [21], and TMA [25] have been proposed to improve performances. Some methods even use both events and images as input for flow estimation, such as Fusion-FlowNet [20], Pan et al. [36], DCEI-Flow [43] and RPEFlow [44]. In this work, we study a new problem of event-based meshflow estimation, proposing an efficient event-based meshflow network. + +# 2.3. Event-based Optical Flow Dataset + +Early works synthesize events by thresholding rgb images [18] and applying interpolation for high framerate [9]. However, the timestamp of synthesized events is inaccurate, let alone interpolation artifacts. The DAVIS event camera [2] can capture both images and real events, resulting in some event datasets: DVSFLOW [24], MVSEC [53] and DSEC [10], based on which EV-Flownet [52] and ER + +Table 1. Comparison of available datasets. 'Dense OF' and Meshflow' indicate whether the dataset has dense optical flow labels and meshflow labels. + +
DatasetResolutionDynamic ObjectsExtreme ConditionsDense OFMeshflowMotion Pattern
DVSFLOW [24]180×240XXXXRotation
MVSEC [53]260×346X✓XXDrone
DSEC [10]480×640X✓XXCar
MDR [33]260×346XX✓XCar
BlinkFlow [21]480×640✓X✓XRandom
Ekubric [44]720×1280✓X✓XFalling
HREM(Ours)720×1280✓✓✓✓Random
+ +AFT [11] compute sparse optical flow. In this way, however, flows can only locate on sparse event regions [43]. Recently, Luo et al. [33] proposed to render the MDR dataset from graphics, but it only contains static scenes. Wan et al. [44] synthesized the Ekubric dataset based on the Kubirc toolbox [12], which only includes a single falling motion pattern. Li et al. [21] considered non-rigid motions simulating dancing but with lower resolution and without extreme condition scenarios. None of the aforementioned datasets involve the estimation of meshflow. In this work, we render a comprehensive dataset that can support both meshflow and optical flow estimation, with a higher resolution, rich dynamic scenes, as well as extreme conditions like relatively low light and motion blur, as shown in Table 1. + +# 3. Algorithm + +# 3.1. High-Resolution Event Meshflow Dataset + +We create the high-resolution event meshflow dataset (HREM), consisting of 100 virtual scenes that accurately mimic real-world environments, both indoors and outdoors. In these scenes, we put dynamic objects to simulate intricate object motions. Camera is programmed to track these movements, ensuring a realistic portrayal of motion. The Blender rendering engine was utilized to create high frame rate videos and dense optical flow labels. For event data generation, we employed three advanced simulators: ESIM [37], V2E [14], and DVS-Voltmeter [24]. The simulator that provided the highest contrast in warped events images was chosen for our dataset. We also process the dense optical flow using motion propagation and median filters, enabling the generation of meshflow labels. Our comparisons with existing datasets is shown in Fig. 2, emphasize the superiority of HREM in high-resolution, dynamic scenes, complex motion patterns, and comprehensive labeling. Following [28], we generate meshflow from dense optical flow, as depicted in Fig. 3. + +Motion Propagation. Given a dense optical flow $\mathbf{F}$ , we place a uniform mesh of $16 \times 16$ regular cells on its image plane and then select the motion of the middle point $p$ in each cell as the local motion $v_{p}$ . Since the vertices of the mesh near point $p$ should have a similar motion to $v_{p}$ , we define a rectangle that covers $3 \times 3$ cells centered at $p$ , and assign to all the vertices within the rectangle, ensuring the + +propagation of the local motion $v_{p}$ across the image plane. + +Median Filters. The local motions of the middle points in all cells are propagated to their nearby mesh vertices, resulting in each vertex potentially receiving multiple motion vectors. To select the most appropriate motion vector for a given vertex, we apply a median filter $f_{1}$ to filter the candidate motions. The response of the filter is then assigned to the corresponding vertex. The median filter is a widely used technique in optical flow estimation and has been shown to produce high-quality flow estimates [39]. Therefore, we use the median filter for sparse motion regularization. However, due to dynamic objects, the motion field may contain noise and needs to be spatially smoothed. To address this issue, we apply another median filter $f_{2}$ that covers $3 \times 3$ cells neighborhood to suppress the noise in the motion field. This second median filter produces a spatially-smooth sparse motion field, which is what we called as meshflow. Ultimately, we use the meshflow (generally $16 \times 16$ ) as the label, which contains global motion information. However, we upsample the meshflow to the full image resolution for intuitive display and alignment applications. + +# 3.2. Estimation for Meshflow and Optical Flow + +Following [11] which estimates optical flow from two consecutive event sequences, we estimate the meshflow $MF_{k\to k + 1}$ and optical flow $F_{k\rightarrow k + 1}$ from event sequences $E(t_{k - 1},t_k)$ and $E(t_{k},t_{k + 1})$ , and overall architecture of our network is shown in Fig. 4. Fellow [54], we convert the inputs $E(t_{k - 1},t_k)$ and $E(t_{k},t_{k + 1})$ to the 3D volumes $V_{k - 1\to k}$ and $V_{k\to k + 1}$ + +# 3.2.1 Efficient Event-based Meshflow Network + +We propose EEMFlow to directly output results of the same resolution as the ground truth meshflow $MF_{GT}$ for supervised regression, fully leveraging the advantage about low-parameter and high-motion-information of meshflow. + +Overall Structure of Meshflow Estimation. Since meshflow focuses more on global large motion rather than local detailed motion, the EEMFlow we designed does not require excessively deep network layers or refinement operations from coarse to fine. Firstly, EEMFlow employs a three-level pyramid feature encoder to extracting features $(V_{t_{k - 1}}^i)^N$ and $(V_{t_k}^i)^N$ from $V_{k - 1\rightarrow k}$ and $V_{k\rightarrow k + 1}$ , the convolutional layers within the $i$ -th level share weights for + +![](images/33b37661ca91328013c7b981dcb7665f1f604e5388fa28542ea364690667490e.jpg) + +![](images/524953267b60898dd498f1c6910ed2b2b00942ba01d6f74a061deaf9306589f8.jpg) + +![](images/98485f598180dc52b2da7a9ca4343b5785631b13062c9f91a090a92b91ced924.jpg) + +![](images/296bbbf2038ecd81415d972fd646e69c5526cfb5e6b3cadf82def1cd576156b3.jpg) + +![](images/ae38838da1ba53a93b669380d73b4aa8ea445d23da337d69163c46ab90d30b32.jpg) + +![](images/3c64997abcf1dc0f4ae722197050cf63da35f51e3c074437e871150fbdb1ed83.jpg) + +![](images/42824c3e67f0fe77fdb1b91057b4ecd6abf4dad327616be3e97262a3143cbe18.jpg) + +![](images/e3085ee4e352aee8dfaa35cb0b6c6c4fef12141db8871d3876f8bd257503e64b.jpg) +(a) +(b) +(c) += + +![](images/ec1bef5238fcd6e20a8974d3a4399269eb7d129f9ce878c8f7e5d0e6180c9115.jpg) + +![](images/4f5da1305ea57a74698431b28d27e8cf3023b3cf28d649bc03958a6c9bc9b908.jpg) + +![](images/b6cb352a7624bacec943dac5d04832baf37279a5e000eef99ca709dd54233b98.jpg) + +![](images/34114591dd5335d2594b39256e06abe090f974ce36a0102e11e1d941d4c736a1.jpg) + +![](images/4c041d142d57b2125cab9c9617fefffe5632cd0ed2e821e70e25c0171efa2067.jpg) + +![](images/a9eb40e1bcf52234e02028eb35c8e30c6d3610f3ee23d5ca0fbf3baa4a7ba829.jpg) + +![](images/c17957b13cb923135c3fc2b38d2b70f18641160ed975f17e25faac8c5f3459f8.jpg) + +![](images/5403db35b83329e7d27bedc5cd657e5aa7c9f3c1153c99a8c9fbb4d4c4397a44.jpg) +MVSEC +DSEC +Figure 2. Examples from MVSEC [53], DSEC [10] and our dataset. (a): Event data, (b): Optical flow, (c): Upsampled meshflow. The meshflows of MVSEC and DSEC contain a significant number of ineffective areas due to the sparsity of optical flow. In contrast, our meshflow provides a complete global motion field. Best viewed on a color screen in high resolution. + +![](images/05f11b9af16dd20ed103d3731b84580f7f0b85504c11864375085c44485527ff.jpg) + +![](images/4d13999406ad78d44cbf5994bbe6f7ca0dd8b791620e4eded31b81cafe09cf15.jpg) + +![](images/6f29cf738b64611c320f3be1c33abf69738dd92940d35b5b5c8f86f71f6e2de4.jpg) +HREM(Ours) + +![](images/bad93513d554583ceb6ec9093f764d952a6e20c7d1b224524ddfa13c6d4cb777.jpg) + +![](images/e3e2e8af64738c9372cfb15a4190afefe67c589c7d7bd3a9f5fa27e68f4284c2.jpg) + +![](images/ca48efec2a148cbd2aa73d2cda1d2ff2b19e34631b7dfc3310cb19ada99f8c16.jpg) + +![](images/41c32b84fab9b46c89f32de5b57f75dd06a3d3b6b99de90272e0161bdcd7f680.jpg) +(a) +Figure 3. The process of generating meshflow from dense optical flow. (a) Propagate the motion vector of each grid center to the grid vertices. (b) Apply median filter $f_{1}$ to multiple motion vectors of each vertex to select the most appropriate motion for that vertex. (c) Use median filtering $f_{2}$ to smooth the motion field in the mesh grid. For ease of visualization, we present the $8 \times 8$ grid mesh in this paper. + +![](images/5ed42728f070a615355e62212338e2db4a1f3b4a4db9b7d464dcedc4afe647c7.jpg) +(b) + +![](images/03e081b20510d66517837d7e44b658e9805d21ab18742ce6b1f3211502508cc7.jpg) +(c) + +$V_{t_{k-1}}^i$ and $V_{t_k}^i$ . Secondly, EEMFlow utilizes multi-scale feature correlation to builds the cost volumes for meshflow estimation. Features $(V_{t_{k-1}}^i)^N$ and $(V_{t_k}^i)^N$ undergo average pooling operation $\mathcal{P}$ to the same resolution (1/64 resolution of $V_{k\rightarrow k+1}$ ), and then use correlation to capture relative motion information and output cost volumes $(C^i)^N$ . Specifically, we employ the dilated feature correlation (DFC) to increase the search area while reducing computational parameters. Finally, we stack the cost volumes $(C^i)^N$ and features $(V_{t_{k-1}}^i)^N$ after pooling, fuse them with a weighted-sum operation, and then feed them into the decoders to regress the meshflow $MF_{k\rightarrow k+1}$ at the same resolution as $MF_{GT}$ . Inspiring by [51], we replace the conventional convolutions with group shuffle convolutions, which leads to efficient computation while maintaining high accuracy. + +Dilated Feature Correlation. We use the inner product to calculate correlation between $V_{t_{k-1}}^i$ and $V_{t_k}^i$ for meshflow estimation: + +$$ +C ^ {i} (\boldsymbol {u}, \boldsymbol {d}) = V _ {t _ {k}} ^ {i} (\boldsymbol {u}) \cdot V _ {t _ {k - 1}} ^ {i} (\boldsymbol {u} + \boldsymbol {d}) / M, \boldsymbol {d} \in \mathcal {N}, \tag {1} +$$ + +where $\pmb{u}$ represents the spatial coordinates on $V_{t_k}^i$ , $\mathcal{N}$ represents the search grid of coordinate $\pmb{u}$ in feature $V_{t_{k - 1}}^i$ , $M$ represents the number of elements in $\mathcal{N}$ , $\pmb{d}$ represents the offset coordinates of the elements in $\mathcal{N}$ , and $\cdot$ represents the inner product calculation. Many methods like [16, 40] simply define $\mathcal{N}$ as a square range of size $(2r + 1)\times (2r + 1)$ + +and observe that increasing the radius $r$ can reduce errors but increase computational overhead. We propose dilated feature correlation that samples densely around the center and sparsely at farther distances, thereby reducing computation while enabling larger radius cost volumes, as Eq. 2: + +$$ +\mathcal {N} \left(d _ {x}, d _ {y}\right) = \left\{ \begin{array}{l} 0, \text {i f} | d _ {x} | + | d _ {y} | = 2 k, k \in [ 2, r ], \\ 1, \text {o t h e r s ,} \end{array} \right. \tag {2} +$$ + +where $\pmb{d} = (d_x, d_y)$ represents the relative position coordinates within the neighborhood $\mathcal{N}$ , where $\mathcal{N}(d_x, d_y) = 1$ indicates the computation of correlation, and $\mathcal{N}(d_x, d_y) = 0$ signifies no computation. + +# 3.2.2 Event-based Optical Flow Network + +Based on EEMFlow, we make some improvements for accurate estimation of optical flow $F_{k \to k + 1}$ , upgrading it to EEMFlow+. Since optical flow focuses more on local motion and pays attention to object edge details, we employ the coarse to fine residual approach to progressively refine the flow, which can be expressed as Eq. 3. + +$$ +F ^ {i + 1} = \operatorname {C o n v} ^ {i} \left(\mathcal {C} \left(V _ {t _ {k - 1}} ^ {i}, \mathcal {W} \left(V _ {t _ {k}} ^ {i}, F _ {\uparrow} ^ {i}\right)\right)\right) + F _ {\uparrow} ^ {i}, \tag {3} +$$ + +where $\mathcal{C}(\cdot, \cdot), \mathcal{W}(\cdot, \cdot)$ and $\uparrow$ donates our dilated feature correlation, the warping operation and upsampling, respectively. We employ the pyramid decoders for optical flow, thus $F^i$ is the output flow of the $i$ -th decoder, $F^{i+1}$ and $\mathrm{Conv}^i$ are respectively the output flow and convolutions in the $i+1$ -th decoder. The most noteworthy aspect is upsampling $F^i$ to $F_{\uparrow}^i$ . Many methods [6, 16, 40, 52] use bilinear interpolation for upsampling, but this can lead to the mixing of incorrect motions at object edges, resulting in blurring. Therefore, we propose the confidence-induced detail completion module for upsampling to enhance edge details. + +Confidence-induced Detail Completion Module. We propose the confidence-induced detail completion module + +![](images/77f3abc72bc1cb2188873eb9d46e2ab42e3ec3aa18c7b4727ce9821bc1397469.jpg) +Figure 4. Our proposed network architecture. We employ pyramid encoders to extract multi-scale features from $V_{k-1 \to k}$ and $V_{k \to k+1}$ , then use dilated feature correlation to compute the cost volume between each layer of features, followed by decoding to output the results. For meshflow estimation, we utilize the decoders with group shuffle convolutions to output predictions $MF_{k \to k+1}$ , upsampled to $MF \uparrow_{k \to k+1}$ . For optical flow estimation, we refine flows using a coarse-to-fine residual approach and confidence-induced detail completion module, finally outputting the optical flow $F_{k \to k+1}$ . + +![](images/dde6960db194bab40144fcb44dd2a88573264479b9b3ba324a806464bea22868.jpg) +Figure 5. The structure of CDC. CDC employs self-corrector based on a dense convolutional network and self-correlation based on a self-attention mechanism to correct the flow obtained from bilinear upsampling. $\mathcal{B}$ represents bilinear upsampling. + +(CDC) to eliminate the blurring of object edges caused by the mixing of multiple motions at the junctions of different movements during upsampling. The detailed structure of our CDC is shown in Fig. 5. Given a small-scale flow $F^i$ from the $i$ -th level, our CDC first generates an initial flow $\bar{F}^i$ for the $i + 1$ -th level through bilinear interpolation, and then arranges the self-corrector and self-correlation branches to correct it. Self-corrector is based on a dense convolutional network with a five-layer structure. It captures motion information within the edge neighborhood through dense convolution from the concatenated feature $V_{t_{k-1}}^i$ and $V_{t_k}^i$ , outputting the corrected flow $\Delta F^i$ and the corrected confidence map $W^i$ . Self-correlation is based on a self-attention mechanism, using a large receptive field to find the fine regions in features $V_{t_k}^i$ that are identical to the motion of the error region in $\bar{F}^i$ . It outputs self-attention weights $A^i$ , multiplied with the initial flow $\bar{F}^i$ . With the corrected flow $\Delta F^i$ and self-attention weights $A^i$ , we can generate the fine flow $\tilde{F}^i$ : + +$$ +\tilde {F} ^ {i} = \alpha \mathcal {W} (\bar {F} ^ {i}, \Delta F ^ {i}) + (1 - \alpha) \left(A ^ {i} \otimes \bar {F} ^ {i}\right), \tag {4} +$$ + +where $\mathcal{W}(\cdot, \cdot)$ means the warping operation, $\otimes$ donates multiplication and $\alpha \in [0,1]$ is the weight coefficient. We iden + +tify error-prone object edge areas based on the corrected confidence map $W^{i}$ , as weights to fuse the initial flow $\bar{F}^{i}$ and the fine flow $\tilde{F}^{i}$ , obtaining the final corrected flow $F_{\uparrow}^{i}$ : + +$$ +F _ {\uparrow} ^ {i} = W ^ {i} \odot \bar {F} ^ {i} + (1 - W ^ {i}) \odot \tilde {F} ^ {i}, \tag {5} +$$ + +where $\odot$ donates the element-wise multiplier. + +# 3.2.3 Loss Function + +For both meshflow and optical flow estimation, we use L1 loss for supervised regression during training. Our EEMFlow, used for estimating meshflow, directly outputs results at the same resolution as the meshflow GT, allowing for direct loss calculation. Similarly, the results outputted by our EEMFlow+ for optical flow estimation are at the same resolution as the network input, enabling direct calculation with the optical flow GT. + +# 4. Experiments + +# 4.1. Implementation Details + +# 4.1.1 Datasets + +Our HREM dataset includes 100 indoor and outdoor scenes, with a resolution of $1280 \times 720$ . We randomly select 70 scenes for training and reserve the remaining 30 for testing. The training set comprises 20,000 samples, while the test set contains 8,000 samples. Additionally, we further divide the test set by scene type (outdoor vs. indoor) and camera motion speed during rendering, resulting in four sub-sequences (outdoor Slow, outdoor fast, indoor slow, indoor fast), with mean motion magnitudes ranging from, $0 - 30$ , $30 - 100$ , $0 - 20$ and $20 - 100$ pixels, respectively. Moreover, similar to [52], we employ two data input modes: $dt = 1$ and $dt = 4$ . $dt = 1$ uses the event sequence between two consecutive frames of RGB images as input, with a meshflow generation frequency of $60~\mathrm{Hz}$ while $dt = 4$ uses the event sequence between four consecutive frames of RGB images as input, with a meshflow + +Table 2. Quantitative comparison of our EEMFlow with other advanced flow networks on our HREM dataset. The evaluation metric used is End-Point Error (EPE). "Parameters" and "Time" respectively indicate the network parameter count and inference time. $\Delta P$ and $\Delta T$ represent the change in parameter count and inference time relative to ERAFT [42] for other networks. Smaller values are desirable for all metrics. We highlight the best results in red and the second-best results in blue. + +
Method dt = 1Parameters (M)Time (ms)OutdoorIndoorAvg
SlowFastSlowFast
EVFlownet [52]38.2463.5516.162.9311.658.57
PWCNet [40]3.36423.9114.492.8611.898.29
ERAFT [11]5.27934.1513.322.9110.347.68
SKFlow [41]6.281453.7611.787.248.817.24
GMA [17]5.891082.1812.072.029.346.40
KPAFlow [32]6.001842.0312.251.959.026.31
FlowFormer [15]9.872812.0611.711.888.666.08
EEMFlow(Ours)1.2472.429.092.008.465.50
Method dt = 4ΔP (M)ΔT (ms)OutdoorIndoorAvg
SlowFastSlowFast
EVFlownet [52]+624%+51%18.2549.3216.1647.1932.73
PWCNet [40]-36%-55%16.4046.1714.4940.9029.49
ERAFT [11]0%0%15.2140.8313.3239.6127.24
SKFlow [41]+19%+56%14.9339.2411.7139.2226.28
GMA [17]+11%+16%14.1338.8912.0737.6825.69
KPAFlow [32]+14%+99%14.0438.0312.2537.2025.38
FlowFormer [15]+88%+202%13.8938.5510.7738.5325.44
EEMFlow(Ours)-76%-92%13.9737.3312.0934.3924.45
+ +![](images/9fc0c88cacfe6e458c8cfb59ca03cda2904196b652f4d696e78cab8325f790d2.jpg) +Image Overlaid +Figure 6. Subjective results of image registration using meshflow estimated by other methods and ours. We estimate meshflow from the event sequences and then warp image $I_{t_1}$ onto image $I_{t_2}$ for fusion, showcasing the fused result. The fewer blue or red ghosting artifacts indicate better alignment performance. + +![](images/f280d6c275454193d148eb0324e69dc68af1773c37875506ff7ef3732c8ceec1.jpg) + +![](images/7cf3e046c10ea117aa8290abb1e1435be49017bb72252e4bbd7743e14a1ec991.jpg) +EVFlownet + +![](images/2e850dc37c6dc0639f6ab80da4b2653feccfda8fc021fcbba7f6e2f0aebd8f4d.jpg) + +![](images/ccd3554bfda5da10a8e9166277aa1d8e8ada1cb211a4379a13572f3e8a16f6de.jpg) +ERRAFT + +![](images/ef632fbee6fec007d58cdc19bfefd4e63501bf5f9654b0328cad5b6bd2c07f8c.jpg) + +![](images/d83a0811619fbfaf618e1e8b7653469de9a27f755c92ae3ed7fc963efaf4ef88.jpg) +KPAFlow + +![](images/fc59d6b08c2ff9b1b3d3365ca810a5655c76ec127f9c7f4a0380943935d6c3d5.jpg) + +![](images/694eb12bb2cc5e279cfeda3c5636c85c56b9e28ed869030cf31f04b4d1ef9846.jpg) +Ours + +![](images/bf17103592adf34b08c77e58c71bc7df60d2c48ff7ec6cc8e06547915d1bf1c7.jpg) + +![](images/1d33704e70b973568640502fbbc9ed0de00a48367e30881d1869afb8179708ac.jpg) +Meshflow(GT) + +generation frequency of $15\mathrm{Hz}$ . DSEC [10] is a commonly used dataset for event-based optical flow estimation, which consists of real-world data captured by real event cameras mounted on cars. We also conduct experiments on DSEC to compare the performance for optical flow estimation. + +# 4.1.2 Training details + +We conduct experiments using the PyTorch framework on two NVIDIA 2080Ti GPUs. We train all networks with the same parameters on our HRDM dataset. We employ the AdamW optimizer and OneCycle policy with a learning rate of $5 \times 10^{-4}$ , weight decay of $5 \times 10^{-5}$ , and other default parameters set to $\beta_{1} = 0.9$ , $\beta_{2} = 0.99$ , $\epsilon = 1 \times 10^{-4}$ . We train all networks up to $100k$ iterations to reach convergence. + +# 4.1.3 Evaluation metrics + +Following EV-FlowNet [52], we use the average Endpoint Error (EPE) as the metric. When evaluating meshflow, we upsample both the prediction and the ground truth to the + +input resolution to calculate metrics. In addition, the DSEC dataset uses $N$ PE to measure the percentage of flow errors higher than $N$ pixels in magnitude (e.g., 3PE, 2PE, 1PE) for flow outliers analysis, and employs the Angular Error (AE) to assess the directional accuracy. + +# 4.2. Comparison with State-of-the-Arts + +# 4.2.1 Results for Event-based Meshflow Estimation + +In Table 2, We train and test our EEMFlow on our HREM dataset, compared with some advanced networks, e.g., EVFlownet [54], ERAFT [11], PwCNet [40], SKFlow [41], GMA [17], KPAFlow [32], and FlowFormer [15]. Since these networks are all structured to output prediction with the same resolution of input, we upsample the ground truth $MF_{GT}$ to supervise these networks during training. However, EEMFlow can directly output prediction aligned $MF_{GT}$ in resolution. But for fair comparison, we upsample the predictions of all networks to a uniform resolution during evaluation. We train and test all networks using two input modes ( $dt = 1$ and $dt = 4$ ), and + +![](images/5a8c0f47f9e71a046ccb10fae5a783c7464652c6e285efe1cbd1b6288bf717ad.jpg) + +![](images/303579b082eac16c49178a78c7ae1ef0e09d1effdd845ba3c55e62c66aad3106.jpg) + +![](images/0dee9e03d491fbbe1585d914ddf12cd982e46a0f808e251c6153d7fad7ef55c3.jpg) + +![](images/3b898acb02f98b32b54be43c95aa0f3af255f18269dbd6c15aaf89fa5f63765c.jpg) + +![](images/5f4a5608ec7743aea9011075eeccbca48b75df2ebb526d7558ae79bbd5961930.jpg) + +![](images/a09caf932e6a7f206ff8751da8f0b62c09fed60ff347aff61742d6e05893551e.jpg) + +![](images/4f1449191a3f8e0360017f82e0d8d4c1065dd08d1c3e3f460d01a264dbf00e34.jpg) + +![](images/2a4b3087f83298301bdbf5e9ab908279fecac0daac1a3ef0fc67f89b3d761898.jpg) + +![](images/dde18daf6e3dc43071308ecae8c8601e4f3f29157e46d584a386ad2d454645c8.jpg) + +![](images/8668f53d774ed290cc97deefb29b2f23456c761e37c7ac18ca38980638a5d5ad.jpg) + +![](images/fa5d9085616fdc7563c28349ab0c7c4a681ac8c04f7c1a5c50b5446e5d0f03f3.jpg) + +![](images/8daba0d062a0b561ed461b2a3d2e0666e25095c66432f744e2a7a28cc5652a4b.jpg) + +![](images/81d3b1f686b36f9233539ce65f98a0d56a7c96140969bdac66195b05b1aa5664.jpg) + +![](images/4332fd43de910a4058b61d313e838e981065761f2cb54436398902268d8223f8.jpg) +Reference + +![](images/c385d0c5272a8683530646dd4a3e64b79abf569033a161965c7318bddf189f70.jpg) + +![](images/393596e3a19cd222d3e50dcdd2549c1bc1bcbfa89588c8b3b3ffb17759850f28.jpg) +EVFlownet + +![](images/d6a7cd9d93821cdb4e5c5798ea0def3ec4b6924268c107ecc513356bcab3f497.jpg) + +![](images/530f474e613766ac8ff207f80d502ae2e95b85022db741752ca88dc704989483.jpg) +ERRAFT + +![](images/2574e57faea5804beb100cbca3076322da371bb0a2cb892e9c813da73e8c55d0.jpg) + +![](images/9d356c7b34313c8838f65cf45d3df7bb64d06ea0e0144c5496adb2856deeea6f.jpg) +KPAFlow + +![](images/cf1de7783ed2ab4e7f427e7782088aaf6c7db3854ea12c4dcbd292719b5e07f0.jpg) + +![](images/d54f1c76ddaee17679afaefa9ace9198fcc4f1006d1054d6a2cf47e516da16e8.jpg) +Ours + +![](images/a6768023085dc78983a637802e01f1a4abeeced6c0e06c5ebd2869d278113461.jpg) + +![](images/ee6b5dde4d2b965b0cf76a1e2a5a03964507a8b64c20a7e627b497679540d6bc.jpg) +Meshflow(GT) + +Table 3. Results on DSEC dataset. All networks are trained and tested on the DSEC dataset for optical flow estimation. + +
MethodsFPS↑1PE↓2PE↓3PE↓EPE↓AE↓
MutilCM [39]-76.648.530.93.4714.0
EV-Flownet [52]22.355.429.818.62.328.12
OF-EV-SNN [4]-53.720.210.31.716.34
EVA-Flow [50]-15.9-3.200.883.31
ERRAFT [11]11.412.74.742.680.792.85
ADMFlow [33]9.8812.54.672.650.782.84
EFlowformer [21]-11.24.102.450.762.68
TMA [25]7.5510.93.972.300.742.68
EEMFlow+(Ours)39.211.43.932.150.752.67
+ +present their performance metrics in four test sub-sequences (outdoor Slow, outdoor fast, indoor slow, indoor fast). In addition, we average the metric scores across all four test sub-sequences as shown in the "Avg" column. + +Table 2 shows that our EEMFlow achieves the lowest EPE score for both input modes $dt = 1$ and $dt = 4$ in the "outdoor_fast" and "indoor_fast" test sub-sequences, demonstrating its superior performance on high-speed and large-movement sequences. EEMFlow also exhibits great potential to outperform other flow networks on the "outdoor Slow" and "indoor Slow" test sub-sequences. Notably, EEMFlow achieves the lowest EPE scores in the "Avg" column for both input modes $dt = 1$ and $dt = 4$ . Besides, we also achieve the least number of parameters and the fastest inference speed. Compared to ERAFT [11], our EEMFlow reduces the parameter count by $76\%$ (from 5.27M to 1.24M), reduces the inference time by $92\%$ (from 93ms to 7ms), and improves average EPE in $dt = 4$ by $8\%$ (from 27.24 to 25.18). EEMFlow achieves comparable EPE performance to FlowFormer [15] but exhibits a $38.7\times$ increase in inference speed. + +![](images/a8d61c4db4aca6afe377c9452a17380f685f8f490498857f122a64dde872bc88.jpg) +Figure 7. Qualitative comparison of our proposed EEMFlow with other advanced flow networks on our HREM dataset. The subjective images of events and dense optical flow on the left side serve as references. The areas enclosed by red rectangles are zoomed in. + +![](images/53cf844b4b98f4a9add46d018969669764dabdf077c036ee3109367c814fe87e.jpg) + +![](images/4757f04649ab5f374f031e50ae7c3535afe04acf61d5c9f29cd36a7a1cce1619.jpg) + +![](images/e3b3b30b2c3dad0b9d3e9e396d776cf60ee347b3f3d8c62a8bf81431d404f735.jpg) +Reference +EVFlownet +ERRAFT +TMA +Ours +Figure 8. Qualitative comparisons on the DSEC test set. We visualize the dense predictions and zoom in the areas where are apparent differences. + +In Fig. 7, we qualitatively compare our proposed EEMFlow with other flow networks, e.g., EVFlownet [54], ER-AFT [11], and KPAFlow [32]. To facilitate comparison, we upsample the meshflow estimation and ground truth to the same resolution of input and zoom in on the areas with the apparent differences. EVFlownet shows the worst performance, with many holes and color mixing. ERAFT and KPAFlow would exhibit block artifacts and appear coarse in nature. In contrast, our EEMFlow results are smoother with more natural color transitions and greater similarity to the upsampled ground truth. In Fig. 6, we present the subjective results of these networks for image registration. We estimate and upsample meshflow prediction from the event sequences $E(t_{1},t_{2})$ and then warp image $I_{t_1}$ onto image $I_{t_2}$ . We also zoom in on the challenging areas of alignment in the registered results, clearly demonstrating that our EEMFlow achieves excellent image registration performance. + +![](images/aebd1753d8341d94d1b4d7596543dfaea1fd6be8807c9858aa8abe6ba244bf69.jpg) +Figure 9. Visualization for meshflow and optical flow results. + +Table 4. Results of meshflow and optical flow on HREM for $dt = 1$ + +
TaskMethodOutdoorIndoorAvg
SlowFastSlowFast
Optical FlowFlowFormer6.2016.065.9915.2710.88
EEMFlow+3.8811.024.0310.927.46
Mesh-flowFlowFormer5.9915.125.7414.9510.45
EEMFlow2.429.092.008.465.50
+ +# 4.2.2 Results for event-based optical flow estimation + +Table 3 presents the comparative results on DSEC dataset of our optical flow estimation network, EEMFlow+, against other event-based optical flow estimation networks. EEMFlow+ achieves state-of-the-art performance in the 2PE, 3PE, and AE metrics, and is on par with the best in the EPE metric. Notably, EEMFlow+ still maintains a significant advantage in inference speed. Compared to TMA [25], our EEMFlow+ increases the inference speed by $419\%$ (from 7.55FPS to 39.2FPS). Fig. 8 displays the qualitative results of our optical flow estimation on the DSEC test set, compared with other advanced methods, including EVFlownet [52], ERAFT [42], and TMA [25]. By zooming into the object areas, it is clearly observable that our results exhibit more regular shapes and clearer boundaries, highlighting the detail enhancement capability of our proposed CDC module. + +# 4.3. Ablation Studies + +# 4.3.1 The Advantages of Event-Meshflow Estimation + +According to the ERAFT [42], image-based approaches face challenges in handling difficult images due to the limited dynamic range of image sensors. As shown in Fig. 9, even advanced image-based techniques like Flow-Former struggle, while our event-based methods (EEMFlow, EEMFlow+) exhibit notably superior performance. Additional quantitative comparisons are provided in Table 4, our network of event-based meshflow estimation EEMFlow achieves the lowest EPE scores in HREM dataset for the $dt = 1$ input mode, which demonstrates the remarkable efficiency of our EEMFlow makes it suitable for real-time applications such as online video stabilization and autonomous driving. + +Table 5. Ablation studies about CDC of EEMFlow+. CDC consists of two branches, the self-corrector and self-correlation. + +
ModelSelf-correctorSelf-correlationFPS↑3PE↓EPE↓AE↓
(a)XX60.43.520.893.11
(b)✓X55.62.770.812.92
(c)X✓46.32.650.792.78
(d)✓✓39.22.150.752.67
+ +# 4.3.2 Experiments for CDC of EEMFlow+ + +In Table 5, we also conduct ablation experiments on the CDC module of EEMFlow+ used for optical flow estimation, including its two branches, the self-corrector and self-correlation. We train and evaluate all models using the same settings on the DSEC dataset to show the individual impact of each branch in CDC module. Comparison of (a)&(b) demonstrates that the CDC with only the self-corrector branch can bring a significant increase in accuracy with a minimal loss in speed. Comparison of (b)&(c) shows that self-correlation, compared to self-corrector, can lead to a higher increase in accuracy, albeit at a further reduction in inference speed. Finally, the comparison of (a)&(d) shows that the CDC composed of both the self-corrector and self-correlation branches significantly improves the accuracy of optical flow estimation with an acceptable loss in speed. + +# 5. Conclusion + +In this work, we develop the first event-based meshflow dataset, termed HREM, where 100 indoor and outdoor virtual scenes with rich scene contents are rendered using Blender. Our HREM possess physically correct accurate events and meshflow label pairs with so far the highest resolution. Furthermore, we propose an efficient event-based meshflow network (EEMFlow), which achieves state-of-the-art performance on HREM dataset while maintaining high efficiency. Based on EEMFlow, we propose a confidence-induced detail completion module to upgrade it as EEMFlow+ for optical flow estimation, achieving SOTA results on DSEC dataset with the fastest inference speed. + +The integration of optical flow and depth data allows for the extension of the proposed framework to address 3D scene flow. Subsequently, our future work will involve expanding this methodology on the HREM dataset by integrating scene flow annotations and investigating scene flow estimation using event cameras. + +Acknowledgements. This work was supported by National Natural Science Foundation of China (NSFC) under grant No.62372091, Sichuan Science and Technology Program of China under grants Nos. 2023NSFSC0462, 2022YFQ0079 and 2023YFG0054. + +# References + +[1] Ryad Benosman, Sio-Hoi Ieng, Charles Clercq, Chiara Bartolozzi, and Mandyam Srinivasan. Asynchronous frameless event-based optical flow. Neural Networks, 27:32-37, 2012. 2 +[2] Christian Brandli, Raphael Berner, Minhao Yang, Shih-Chii Liu, and Tobi Delbruck. A $240 \times 180$ 130 db $3\mu s$ latency global shutter spatiotemporal vision sensor. IEEE Journal of Solid-State Circuits, 49(10):2333-2341, 2014. 2 +[3] Marco Cannici, Marco Ciccone, Andrea Romanoni, and Matteo Matteucci. A differentiable recurrent surface for asynchronous event-based data. In Proc. ECCV, pages 136-152. Springer, 2020. 2 +[4] Javier Cuadrado, Ulysse Rançon, Benoit R Cotterreau, Francisco Barranco, and Timothee Masquelier. Optical flow estimation from event-based cameras and spiking neural networks. Frontiers in Neuroscience, 17:1160034, 2023. 7 +[5] Ziluo Ding, Rui Zhao, Jiyuan Zhang, Tianxiao Gao, Ruiqin Xiong, Zhaofei Yu, and Tiejun Huang. Spatio-temporal recurrent networks for event-based optical flow estimation. In Proc. AAAI, pages 525-533, 2022. 2 +[6] Alexey Dosovitskiy, Philipp Fischer, Eddy Ilg, Philip Hausser, Caner Hazirbas, Vladimir Golkov, Patrick Van Der Smagt, Daniel Cremers, and Thomas Brox. Flownet: Learning optical flow with convolutional networks. In Proc. ICCV, pages 2758-2766, 2015. 4 +[7] Guillermo Gallego, Tobi Delbrück, Garrick Orchard, Chiara Bartolozzi, Brian Taba, Andrea Censi, Stefan Leutenegger, Andrew J Davison, Jörg Conradt, Kostas Daniilidis, et al. Event-based vision: A survey. IEEE transactions on pattern analysis and machine intelligence, 44(1):154-180, 2020. 1 +[8] Daniel Gehrig, Antonio Loquercio, Konstantinos G Derpanis, and Davide Scaramuzza. End-to-end learning of representations for asynchronous event-based data. In Proc. ICCV, pages 5633-5643, 2019. 2 +[9] Daniel Gehrig, Mathias Gehrig, Javier Hidalgo-Carrió, and Davide Scaramuzza. Video to events: Recycling video datasets for event cameras. In Proc. CVPR, pages 3586-3595, 2020. 2 +[10] Mathias Gehrig, Willem Aarents, Daniel Gehrig, and Davide Scaramuzza. Dsec: A stereo event camera dataset for driving scenarios. IEEE Robotics and Automation Letters, 6(3): 4947-4954, 2021. 2, 3, 4, 6 +[11] Mathias Gehrig, Mario Millhäuser, Daniel Gehrig, and Davide Scaramuzza. E-raft: Dense optical flow from event cameras. In International Conference on 3D Vision (3DV), pages 197–206, 2021. 1, 2, 3, 6, 7 +[12] Klaus Greff, Francois Belletti, Lucas Beyer, Carl Doersch, Yilun Du, Daniel Duckworth, David J Fleet, Dan Gnanapragasam, Florian Golemo, Charles Herrmann, et al. Kubric: A scalable dataset generator. In Proc. CVPR, pages 3749-3761, 2022. 3 +[13] Jesse Hagenaars, Federico Paredes-Vallés, and Guido De Croon. Self-supervised learning of event-based optical flow with spiking neural networks. Advances in Neural Information Processing Systems, 34:7167-7179, 2021. 2 + +[14] Yuhuang Hu, Shih-Chii Liu, and Tobi Delbruck. v2e: From video frames to realistic dvs events. In Proc. CVPR, pages 1312-1321, 2021. 3 +[15] Zhaoyang Huang, Xiaoyu Shi, Chao Zhang, Qiang Wang, Ka Chun Cheung, Hongwei Qin, Jifeng Dai, and Hongsheng Li. FlowFormer: A transformer architecture for optical flow. ECCV, 2022. 2, 6, 7 +[16] Junhwa Hur and Stefan Roth. Iterative residual refinement for joint optical flow and occlusion estimation. In Proc. CVPR, 2019. 4 +[17] Shihao Jiang, Dylan Campbell, Yao Lu, Hongdong Li, and Richard Hartley. Learning to estimate hidden motions with global motion aggregation. In Proc. ICCV, pages 9772-9781, 2021. 2, 6 +[18] Jacques Kaiser, J Camilo Vasquez Tieck, Christian Hubschneider, Peter Wolf, Michael Weber, Michael Hoff, Alexander Friedrich, Konrad Wojtasik, Arne Roennau, Ralf Kohlhaas, et al. Towards a framework for end-to-end control of a simulated vehicle with spiking neural networks. In IEEE International Conference on Simulation, Modeling, and Programming for Autonomous Robots (SIMPAR), pages 127-134, 2016. 2 +[19] Chankyu Lee, Adarsh Kumar Kosta, Alex Zihao Zhu, Kenneth Chaney, Kostas Daniilidis, and Kaushik Roy. Spikeflownet: event-based optical flow estimation with energy-efficient hybrid neural networks. In Proc. ECCV, pages 366-382. Springer, 2020. 2 +[20] Chankyu Lee, Adarsh Kumar Kosta, and Kaushik Roy. Fusion-flownet: Energy-efficient optical flow estimation using sensor fusion and deep fused spiking-analog network architectures. In 2022 International Conference on Robotics and Automation (ICRA), pages 6504-6510. IEEE, 2022. 2 +[21] Yijin Li, Zhaoyang Huang, Shuo Chen, Xiaoyu Shi, Hongsheng Li, Hujun Bao, Zhaopeng Cui, and Guofeng Zhang. Blinkflow: A dataset to push the limits of event-based optical flow estimation. arXiv preprint arXiv:2303.07716, 2023. 2, 3, 7 +[22] Zhuoyan Li, Jiawei Shen, and Ruitao Liu. A lightweight network to learn optical flow from event data. In 2020 25th International Conference on Pattern Recognition (ICPR), pages 1-7. IEEE, 2021. 2 +[23] Kaimo Lin, Shuaicheng Liu, Loong-Fah Cheong, and Bing Zeng. Seamless video stitching from hand-held camera inputs. In Computer Graphics Forum, pages 479-487. Wiley Online Library, 2016. 2 +[24] Songnan Lin, Ye Ma, Zhenhua Guo, and Bihan Wen. Dvsvoltmeter: Stochastic process-based event simulator for dynamic vision sensors. In Proc. ECCV, pages 578-593. Springer, 2022. 2, 3 +[25] Haotian Liu, Guang Chen, Sanqing Qu, Yanping Zhang, Zhi-jun Li, Alois Knoll, and Changjun Jiang. Tma: Temporal motion aggregation for event-based optical flow. In Proc. ICCV, 2023. 2, 7, 8 +[26] Shuaicheng Liu, Lu Yuan, Ping Tan, and Jian Sun. Bundled camera paths for video stabilization. ACM transactions on graphics (TOG), 32(4):1-10, 2013. 1 + +[27] Shuaicheng Liu, Lu Yuan, Ping Tan, and Jian Sun. Steadyflow: Spatially smooth optical flow for video stabilization. In Proc. CVPR, pages 4209-4216, 2014. 1 +[28] Shuaicheng Liu, Ping Tan, Lu Yuan, Jian Sun, and Bing Zeng. Meshflow: Minimum latency online video stabilization. In Proc. ECCV, pages 800-815. Springer, 2016. 1, 2, 3 +[29] Shuaicheng Liu, Nianjin Ye, Chuan Wang, Jirong Zhang, Lanpeng Jia, Kunming Luo, Jue Wang, and Jian Sun. Content-aware unsupervised deep homography estimation and its extensions. 45(3):2849-2863, 2022. 1 +[30] Shuaicheng Liu, Yuhang Lu, Hai Jiang, Nianjin Ye, Chuan Wang, and Bing Zeng. Unsupervised global and local homography estimation with motion basis learning. 45(6): 7885-7899, 2023. 1 +[31] Zhen Liu, Yinglong Wang, Bing Zeng, and Shuaicheng Liu. Ghost-free high dynamic range imaging with context-aware transformer. In Proc. ECCV, pages 344-360. Springer, 2022. 1 +[32] Ao Luo, Fan Yang, Xin Li, and Shuaicheng Liu. Learning optical flow with kernel patch attention. In Proc. CVPR, 2022. 2, 6, 7 +[33] Xinglong Luo, Kunming Luo, Ao Luo, Zhengning Wang, Ping Tan, and Shuaicheng Liu. Learning optical flow from event camera with rendered dataset. In Proc. ICCV, 2023. 2, 3, 7 +[34] Lang Nie, Chunyu Lin, Kang Liao, Shuaicheng Liu, and Yao Zhao. Depth-aware multi-grid deep homography estimation with contextual correlation. 32(7):4460-4472, 2022. 1 +[35] Lang Nie, Chunyu Lin, Kang Liao, Shuaicheng Liu, and Yao Zhao. Parallax-tolerant unsupervised deep image stitching. In Proc. ICCV, pages 7399-7408, 2023. 2 +[36] Liyuan Pan, Miaomiao Liu, and Richard Hartley. Single image optical flow estimation with an event camera. In Proc. CVPR, pages 1669-1678, 2020. 2 +[37] Henri Rebecq, Daniel Gehrig, and Davide Scaramuzza. Esim: an open event camera simulator. In Conference on robot learning, pages 969-982. PMLR, 2018. 3 +[38] Zhihang Ren, Jiajia Li, Shuaicheng Liu, and Bing Zeng. Meshflow video denoising. In Proc. ICIP, pages 2966-2970, 2017. 2 +[39] Deqing Sun, Stefan Roth, and Michael J Black. Secrets of optical flow estimation and their principles. In Proc. CVPR, pages 2432-2439, 2010. 3, 7 +[40] Deqing Sun, Xiaodong Yang, Ming-Yu Liu, and Jan Kautz. Pwc-net: Cnns for optical flow using pyramid, warping, and cost volume. In Proc. CVPR, pages 8934-8943, 2018. 4, 6 +[41] Shangkun Sun, Yuanqi Chen, Yu Zhu, Guodong Guo, and Ge Li. Skflow: Learning optical flow with super kernels. Proc. NeurIPS, 35:11313-11326, 2022. 2, 6 +[42] Zachary Teed and Jia Deng. Raft: Recurrent all-pairs field transforms for optical flow. In Proc. ECCV, 2020. 2, 6, 8 +[43] Zhexiong Wan, Yuchao Dai, and Yuxin Mao. Learning dense and continuous optical flow from an event camera. IEEE Transactions on Image Processing, 31:7237-7251, 2022. 2, 3 + +[44] Zhexiong Wan, Yuxin Mao, Jing Zhang, and Yuchao Dai. Rpeflow: Multimodal fusion of rgb-pointcloud-event for joint optical flow and scene flow estimation. In Proc. ICCV, pages 10030-10040, 2023. 2, 3 +[45] Yiming Wang, Qian Huang, Chuanxu Jiang, Jiwen Liu, Mingzhou Shang, and Zhuang Miao. Video stabilization: A comprehensive survey. Neurocomputing, 2022. 1, 2 +[46] Qingsen Yan, Jinqiu Sun, Haisen Li, Yu Zhu, and Yanning Zhang. High dynamic range imaging by sparse representation. Neurocomputing, 269:160-169, 2017. 1, 2 +[47] Min Yang, Jingkun Liang, Jianhai Zhang, Haidong Gao, Fanyong Meng, Li Xingdong, and Sung-Jin Song. Non-local means theory based perona-malik model for image denoising. Neurocomputing, 120:262-267, 2013. 2 +[48] Yan Yang, Liyuan Pan, and Liu Liu. Event camera data pretraining. In Proc. ICCV, pages 10699-10709, 2023. 2 +[49] Nianjin Ye, Chuan Wang, Shuaicheng Liu, Lanpeng Jia, Jue Wang, and Yongqing Cui. Deepmeshflow: Content adaptive mesh deformation for robust image registration. arXiv preprint arXiv:1912.05131, 2019. 1, 2 +[50] Yaozu Ye, Hao Shi, Kailun Yang, Ze Wang, Xiaoting Yin, Yaonan Wang, and Kaiwei Wang. Towards anytime optical flow estimation with event cameras. arXiv preprint arXiv:2307.05033, 2023. 2, 7 +[51] Xiangyu Zhang, Xinyu Zhou, Mengxiao Lin, and Jian Sun. Shufflenet: An extremely efficient convolutional neural network for mobile devices. In Proc. CVPR, pages 6848-6856, 2018. 4 +[52] Alex Zhu, Liangzhe Yuan, Kenneth Chaney, and Kostas Daniilidis. Ev-flownet: Self-supervised optical flow estimation for event-based cameras. In Proceedings of Robotics: Science and Systems, 2018. 2, 4, 5, 6, 7, 8 +[53] Alex Zihao Zhu, Dinesh Thakur, Tolga Ozaslan, Bernd Pfrommer, Vijay Kumar, and Kostas Daniilidis. The multi-vehicle stereo event camera dataset: An event camera dataset for 3d perception. IEEE Robotics and Automation Letters, 3 (3):2032-2039, 2018. 2, 3, 4 +[54] Alex Zihao Zhu, Liangzhe Yuan, Kenneth Chaney, and Kostas Daniilidis. Unsupervised event-based learning of optical flow, depth, and egomotion. In Proc. CVPR, pages 989-997, 2019. 3, 6, 7 \ No newline at end of file diff --git a/2024/Efficient Meshflow and Optical Flow Estimation from Event Cameras/images.zip b/2024/Efficient Meshflow and Optical Flow Estimation from Event Cameras/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..f7b60f0918cf8f0a7ed4a0064330fa0071a2bf7b --- /dev/null +++ b/2024/Efficient Meshflow and Optical Flow Estimation from Event Cameras/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a0c51d11da43bd96b74224f9b9bc051e1f57ec1ecdde0783a258e8f622139aa4 +size 707130 diff --git a/2024/Efficient Meshflow and Optical Flow Estimation from Event Cameras/layout.json b/2024/Efficient Meshflow and Optical Flow Estimation from Event Cameras/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..e7a435509fc497fb35f66e2de006e20247f26fd1 --- /dev/null +++ b/2024/Efficient Meshflow and Optical Flow Estimation from Event Cameras/layout.json @@ -0,0 +1,12868 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 82, + 103, + 512, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 82, + 103, + 512, + 121 + ], + "spans": [ + { + "bbox": [ + 82, + 103, + 512, + 121 + ], + "type": "text", + "content": "Efficient Meshflow and Optical Flow Estimation from Event Cameras" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 141, + 536, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 141, + 536, + 213 + ], + "spans": [ + { + "bbox": [ + 62, + 141, + 536, + 213 + ], + "type": "text", + "content": "Xinglong Luo" + }, + { + "bbox": [ + 62, + 141, + 536, + 213 + ], + "type": "inline_equation", + "content": "^{1,4*}" + }, + { + "bbox": [ + 62, + 141, + 536, + 213 + ], + "type": "text", + "content": ", Ao Luo" + }, + { + "bbox": [ + 62, + 141, + 536, + 213 + ], + "type": "inline_equation", + "content": "^{2,4*}" + }, + { + "bbox": [ + 62, + 141, + 536, + 213 + ], + "type": "text", + "content": ", Zhengning Wang" + }, + { + "bbox": [ + 62, + 141, + 536, + 213 + ], + "type": "inline_equation", + "content": "^{1\\dagger}" + }, + { + "bbox": [ + 62, + 141, + 536, + 213 + ], + "type": "text", + "content": ", Chunyu Lin" + }, + { + "bbox": [ + 62, + 141, + 536, + 213 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 62, + 141, + 536, + 213 + ], + "type": "text", + "content": ", Bing Zeng" + }, + { + "bbox": [ + 62, + 141, + 536, + 213 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 62, + 141, + 536, + 213 + ], + "type": "text", + "content": ", and Shuaicheng Liu" + }, + { + "bbox": [ + 62, + 141, + 536, + 213 + ], + "type": "inline_equation", + "content": "^{1,4\\dagger}" + }, + { + "bbox": [ + 62, + 141, + 536, + 213 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 62, + 141, + 536, + 213 + ], + "type": "text", + "content": "University of Electronic Science and Technology of China \n" + }, + { + "bbox": [ + 62, + 141, + 536, + 213 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 62, + 141, + 536, + 213 + ], + "type": "text", + "content": "Southwest Jiaotong University \n" + }, + { + "bbox": [ + 62, + 141, + 536, + 213 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 62, + 141, + 536, + 213 + ], + "type": "text", + "content": "Beijing Jiaotong University \n" + }, + { + "bbox": [ + 62, + 141, + 536, + 213 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 62, + 141, + 536, + 213 + ], + "type": "text", + "content": "Megvii Technology \n{luoboom@std.,zhengning.wang@,eezeng@,liushuaicheng@}uestc.edu.cn \naoluo@swjtu.edu.cn cylin@bjtu.edu.cn" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 143, + 241, + 192, + 254 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 241, + 192, + 254 + ], + "spans": [ + { + "bbox": [ + 143, + 241, + 192, + 254 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 266, + 290, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 266, + 290, + 517 + ], + "spans": [ + { + "bbox": [ + 46, + 266, + 290, + 517 + ], + "type": "text", + "content": "In this paper, we explore the problem of event-based meshflow estimation, a novel task that involves predicting a spatially smooth sparse motion field from event cameras. To start, we generate a large-scale High-Resolution Event Meshflow (HREM) dataset, which showcases its superiority by encompassing the merits of high resolution at " + }, + { + "bbox": [ + 46, + 266, + 290, + 517 + ], + "type": "inline_equation", + "content": "1280 \\times 720" + }, + { + "bbox": [ + 46, + 266, + 290, + 517 + ], + "type": "text", + "content": ", handling dynamic objects and complex motion patterns, and offering both optical flow and meshflow labels. These aspects have not been fully explored in previous works. Besides, we propose Efficient Event-based MeshFlow (EEMFlow) network, a lightweight model featuring a specially crafted encoder-decoder architecture to facilitate swift and accurate meshflow estimation. Furthermore, we upgrade EEMFlow network to support dense event optical flow, in which a Confidence-induced Detail Completion (CDC) module is proposed to preserve sharp motion boundaries. We conduct comprehensive experiments to show the exceptional performance and runtime efficiency " + }, + { + "bbox": [ + 46, + 266, + 290, + 517 + ], + "type": "inline_equation", + "content": "(39 \\times" + }, + { + "bbox": [ + 46, + 266, + 290, + 517 + ], + "type": "text", + "content": " faster) of our EEMFlow model compared to recent state-of-the-art flow methods. Our code is available at https://github.com/boomluo02/EEMFlow." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 531, + 128, + 544 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 531, + 128, + 544 + ], + "spans": [ + { + "bbox": [ + 47, + 531, + 128, + 544 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 551, + 288, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 551, + 288, + 696 + ], + "spans": [ + { + "bbox": [ + 46, + 551, + 288, + 696 + ], + "type": "text", + "content": "Meshflow, a spatially smooth sparse motion field, represents motion vectors exclusively at mesh vertices [28, 49], which has been widely applied in various vision applications, such as image alignment [29, 30, 34], video stabilization [26, 27, 45] and high dynamic range (HDR) imaging [31, 46]. This motion representation combines the benefits of optical flow and global homography, effectively reducing redundancy in motion information and computational costs, while also accommodating non-rigid motions beyond single-plane movements. However, meshflow estimation on RGB images often encounter challenges under scenarios such as low-light and rapid motions. This is due" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 307, + 240, + 547, + 421 + ], + "blocks": [ + { + "bbox": [ + 307, + 240, + 547, + 421 + ], + "lines": [ + { + "bbox": [ + 307, + 240, + 547, + 421 + ], + "spans": [ + { + "bbox": [ + 307, + 240, + 547, + 421 + ], + "type": "image", + "image_path": "6a7187498c849c3a671e8f295bb00308fee81606ff12877504673bf926e9b91d.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 428, + 547, + 468 + ], + "lines": [ + { + "bbox": [ + 305, + 428, + 547, + 468 + ], + "spans": [ + { + "bbox": [ + 305, + 428, + 547, + 468 + ], + "type": "text", + "content": "Figure 1. Comparison of computational overhead and accuracy metrics. The x-axis represents inference time, while the y-axis corresponds to the end-point error. The size of each circle indicates the number of model parameters. Lower values for all metrics are considered better." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 493, + 536, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 493, + 536, + 506 + ], + "spans": [ + { + "bbox": [ + 305, + 493, + 536, + 506 + ], + "type": "text", + "content": "to the loss of fine image texture details and motion blurs." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 508, + 545, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 508, + 545, + 651 + ], + "spans": [ + { + "bbox": [ + 304, + 508, + 545, + 651 + ], + "type": "text", + "content": "In contrast, event cameras are well-suited for motion estimation under such situations [7, 11]. Equipped with bio-inspired vision sensors, event cameras can generate sequences of events with microsecond accuracy triggered by changes in log intensity. In particular, when a change is detected in a pixel, the camera returns an event in the form " + }, + { + "bbox": [ + 304, + 508, + 545, + 651 + ], + "type": "inline_equation", + "content": "e = (x,y,t,p)" + }, + { + "bbox": [ + 304, + 508, + 545, + 651 + ], + "type": "text", + "content": " immediately, where " + }, + { + "bbox": [ + 304, + 508, + 545, + 651 + ], + "type": "inline_equation", + "content": "x,y" + }, + { + "bbox": [ + 304, + 508, + 545, + 651 + ], + "type": "text", + "content": " stands for the spatial location, " + }, + { + "bbox": [ + 304, + 508, + 545, + 651 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 508, + 545, + 651 + ], + "type": "text", + "content": " refers to the timestamp in microseconds, and " + }, + { + "bbox": [ + 304, + 508, + 545, + 651 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 304, + 508, + 545, + 651 + ], + "type": "text", + "content": " is the polarity of the change, either positive or negative. The advantages of high temporal resolution and high dynamic range make event cameras highly effective for analyzing dynamic scenes." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 654, + 547, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 547, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 547, + 712 + ], + "type": "text", + "content": "In this work, we study a new problem that estimates meshflow from event camera data. To start, we create a large-scale High-Resolution Event Meshflow (HREM) dataset, which contains 20k train and 8k test samples. We build 100 virtual scenes in Blender to render the dataset," + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 702, + 214, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 702, + 214, + 713 + ], + "spans": [ + { + "bbox": [ + 61, + 702, + 214, + 713 + ], + "type": "text", + "content": "* Equal contribution. † Corresponding Author." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "19198" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 289, + 301 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 289, + 301 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 289, + 301 + ], + "type": "text", + "content": "which can provide accurate physically-based events along with dense meshflow label pairs. Based on the dataset, we further propose an Efficient Event-based MeshFlow (EEMFlow) network to estimate high-resolution meshflow from event data. Unlike recent flow networks relying on recurrent refinement structure [11, 17, 42], our network is developed on an encoder-decoder architecture with multi-scale global optimizing scheme, which can produce full-resolution meshflow with minimal computational overhead. Specifically, our EEMFlow achieves efficiency by employing the lightweight encoder, building cost volume with dilated feature correlation, and using group shuffle convolutions during decoding. We select recent top-ranked flow networks, including ERAFT [11], SK-Flow [41], KPAFlow [32], and FlowFormer [15], which is trained on our HREM dataset. The results show that our approach achieves state-of-the-art performance while maintaining a fast inference speed of 142.9 FPS, outperforming previous works by a relatively large margin (see Fig. 1)." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 312, + 289, + 492 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 312, + 289, + 492 + ], + "spans": [ + { + "bbox": [ + 47, + 312, + 289, + 492 + ], + "type": "text", + "content": "Additionally, we empirically demonstrate that the proposed new pipeline has the capability to effectively handle various motion patterns. Its lightweight design and runtime efficiency further contribute significantly to the field of optical flow estimation. Specifically, we refine the optical flow progressively during decoding using the coarse to fine residual approach. A Confidence-induced Detail Completion (CDC) module is proposed to preserve motion boundary details during flow upsampling. We also perform comparative experiments with recent event flow networks [21, 25, 33] to illustrate its superiority. The enhanced flow network is referred to as EEMFlow+, demonstrating impressive performance on the reputable DSEC [11] online test benchmark, with the fastest inference speed reaching 39.2 FPS. Our contributions are summarized as:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 510, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 47, + 510, + 287, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 510, + 287, + 569 + ], + "spans": [ + { + "bbox": [ + 47, + 510, + 287, + 569 + ], + "type": "text", + "content": "- We build HREM, the first event-based meshflow dataset, superior in the highest resolution at " + }, + { + "bbox": [ + 47, + 510, + 287, + 569 + ], + "type": "inline_equation", + "content": "1280 \\times 720" + }, + { + "bbox": [ + 47, + 510, + 287, + 569 + ], + "type": "text", + "content": ", dynamic scenes, complex motion patterns, as well as physically correct accurate events paired with meshflow and optical flow labels." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 574, + 287, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 574, + 287, + 635 + ], + "spans": [ + { + "bbox": [ + 47, + 574, + 287, + 635 + ], + "type": "text", + "content": "- We propose EEMFlow, which achieves SOTA performances when compared to top-ranked optical flow networks when trained on our meshflow dataset. Moreover, it achieves inference speed of 142.9 FPS, which is 25.5 to 38.7 times faster than compared methods." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 639, + 288, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 639, + 288, + 712 + ], + "spans": [ + { + "bbox": [ + 47, + 639, + 288, + 712 + ], + "type": "text", + "content": "- We propose CDC, a confidence-induced detail completion module that empowers EEMFlow to make a meaningful contribution to the optical flow community. The upgraded model achieving SOTA performance when compared to representative methods, while also boasting the fastest inference speed to date." + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 71, + 393, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 71, + 393, + 84 + ], + "spans": [ + { + "bbox": [ + 306, + 71, + 393, + 84 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 91, + 480, + 104 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 91, + 480, + 104 + ], + "spans": [ + { + "bbox": [ + 306, + 91, + 480, + 104 + ], + "type": "text", + "content": "2.1. Image-based Meshflow Warping" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 109, + 545, + 338 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 109, + 545, + 338 + ], + "spans": [ + { + "bbox": [ + 304, + 109, + 545, + 338 + ], + "type": "text", + "content": "Meshflow is a lightweight and spatially smooth sparse motion field with motions only located at mesh vertices [28]. Meshflow is different from dense optical flow, where optical flow estimates motions of every pixel of an image while meshflow only concentrates on the global motion, rejecting motions of any dynamic contents. Meshflow is also different from a global homography, where local motions from nonplanar depth variations can be well reflected. Mesh-based methods proofs to be effective in various applications, such as high dynamic range (HDR) imaging [46], burst image denoising [47], video denoising [38], image/video stitching [23, 35] and video stabilization [45]. It is worth noting that directly downsampling the optical flow may yield flow fields of the same resolution as meshflow, but they differ significantly from meshflow and perform poorly in terms of warping effects [49]. Direct downsampling ignores motion outliers from different dynamic objects and the global consistency, which can be effectively addressed by meshflow through motion propagation." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 345, + 504, + 357 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 345, + 504, + 357 + ], + "spans": [ + { + "bbox": [ + 306, + 345, + 504, + 357 + ], + "type": "text", + "content": "2.2. Event-based Optical Flow Estimation" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 363, + 545, + 602 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 363, + 545, + 602 + ], + "spans": [ + { + "bbox": [ + 304, + 363, + 545, + 602 + ], + "type": "text", + "content": "Optical flow estimation from event cameras has received significant attention in recent years. Early approaches, such as [1], could only estimate optical flows at the regions where events are triggered. Recently, deep methods can estimate optical flows from event data, even for the regions without triggered events. For example, EV-FlowNet [52] learns event and flow labels in a self-supervised manner by minimizing photometric distances of grey images acquired by DAVIS [2]. Various event representations, including EST [8] and Matrix-LSTM [3], have been explored, and different network structures, such as SpikeFlowNet [19], LIF-EV-FlowNet [13], STE-FlowNet [5], Li et al. [22], ERAFT [11], Yang et al. [48], EVA-Flow [50], ADMFlow [33], E-FlowFormer [21], and TMA [25] have been proposed to improve performances. Some methods even use both events and images as input for flow estimation, such as Fusion-FlowNet [20], Pan et al. [36], DCEI-Flow [43] and RPEFlow [44]. In this work, we study a new problem of event-based meshflow estimation, proposing an efficient event-based meshflow network." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 611, + 488, + 624 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 611, + 488, + 624 + ], + "spans": [ + { + "bbox": [ + 306, + 611, + 488, + 624 + ], + "type": "text", + "content": "2.3. Event-based Optical Flow Dataset" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "text", + "content": "Early works synthesize events by thresholding rgb images [18] and applying interpolation for high framerate [9]. However, the timestamp of synthesized events is inaccurate, let alone interpolation artifacts. The DAVIS event camera [2] can capture both images and real events, resulting in some event datasets: DVSFLOW [24], MVSEC [53] and DSEC [10], based on which EV-Flownet [52] and ER" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "19199" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 51, + 85, + 542, + 165 + ], + "blocks": [ + { + "bbox": [ + 56, + 70, + 536, + 80 + ], + "lines": [ + { + "bbox": [ + 56, + 70, + 536, + 80 + ], + "spans": [ + { + "bbox": [ + 56, + 70, + 536, + 80 + ], + "type": "text", + "content": "Table 1. Comparison of available datasets. 'Dense OF' and Meshflow' indicate whether the dataset has dense optical flow labels and meshflow labels." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 51, + 85, + 542, + 165 + ], + "lines": [ + { + "bbox": [ + 51, + 85, + 542, + 165 + ], + "spans": [ + { + "bbox": [ + 51, + 85, + 542, + 165 + ], + "type": "table", + "html": "
DatasetResolutionDynamic ObjectsExtreme ConditionsDense OFMeshflowMotion Pattern
DVSFLOW [24]180×240XXXXRotation
MVSEC [53]260×346X✓XXDrone
DSEC [10]480×640X✓XXCar
MDR [33]260×346XX✓XCar
BlinkFlow [21]480×640✓X✓XRandom
Ekubric [44]720×1280✓X✓XFalling
HREM(Ours)720×1280✓✓✓✓Random
", + "image_path": "a7e43710e4ec54d14b940282a4ae39839e0fbd9d4c0df1fc4b5e9679d5bf3acb.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 180, + 289, + 349 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 180, + 289, + 349 + ], + "spans": [ + { + "bbox": [ + 46, + 180, + 289, + 349 + ], + "type": "text", + "content": "AFT [11] compute sparse optical flow. In this way, however, flows can only locate on sparse event regions [43]. Recently, Luo et al. [33] proposed to render the MDR dataset from graphics, but it only contains static scenes. Wan et al. [44] synthesized the Ekubric dataset based on the Kubirc toolbox [12], which only includes a single falling motion pattern. Li et al. [21] considered non-rigid motions simulating dancing but with lower resolution and without extreme condition scenarios. None of the aforementioned datasets involve the estimation of meshflow. In this work, we render a comprehensive dataset that can support both meshflow and optical flow estimation, with a higher resolution, rich dynamic scenes, as well as extreme conditions like relatively low light and motion blur, as shown in Table 1." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 361, + 117, + 375 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 361, + 117, + 375 + ], + "spans": [ + { + "bbox": [ + 47, + 361, + 117, + 375 + ], + "type": "text", + "content": "3. Algorithm" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 382, + 263, + 395 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 382, + 263, + 395 + ], + "spans": [ + { + "bbox": [ + 47, + 382, + 263, + 395 + ], + "type": "text", + "content": "3.1. High-Resolution Event Meshflow Dataset" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 401, + 287, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 401, + 287, + 628 + ], + "spans": [ + { + "bbox": [ + 46, + 401, + 287, + 628 + ], + "type": "text", + "content": "We create the high-resolution event meshflow dataset (HREM), consisting of 100 virtual scenes that accurately mimic real-world environments, both indoors and outdoors. In these scenes, we put dynamic objects to simulate intricate object motions. Camera is programmed to track these movements, ensuring a realistic portrayal of motion. The Blender rendering engine was utilized to create high frame rate videos and dense optical flow labels. For event data generation, we employed three advanced simulators: ESIM [37], V2E [14], and DVS-Voltmeter [24]. The simulator that provided the highest contrast in warped events images was chosen for our dataset. We also process the dense optical flow using motion propagation and median filters, enabling the generation of meshflow labels. Our comparisons with existing datasets is shown in Fig. 2, emphasize the superiority of HREM in high-resolution, dynamic scenes, complex motion patterns, and comprehensive labeling. Following [28], we generate meshflow from dense optical flow, as depicted in Fig. 3." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 630, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 630, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 630, + 288, + 715 + ], + "type": "text", + "content": "Motion Propagation. Given a dense optical flow " + }, + { + "bbox": [ + 46, + 630, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{F}" + }, + { + "bbox": [ + 46, + 630, + 288, + 715 + ], + "type": "text", + "content": ", we place a uniform mesh of " + }, + { + "bbox": [ + 46, + 630, + 288, + 715 + ], + "type": "inline_equation", + "content": "16 \\times 16" + }, + { + "bbox": [ + 46, + 630, + 288, + 715 + ], + "type": "text", + "content": " regular cells on its image plane and then select the motion of the middle point " + }, + { + "bbox": [ + 46, + 630, + 288, + 715 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 46, + 630, + 288, + 715 + ], + "type": "text", + "content": " in each cell as the local motion " + }, + { + "bbox": [ + 46, + 630, + 288, + 715 + ], + "type": "inline_equation", + "content": "v_{p}" + }, + { + "bbox": [ + 46, + 630, + 288, + 715 + ], + "type": "text", + "content": ". Since the vertices of the mesh near point " + }, + { + "bbox": [ + 46, + 630, + 288, + 715 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 46, + 630, + 288, + 715 + ], + "type": "text", + "content": " should have a similar motion to " + }, + { + "bbox": [ + 46, + 630, + 288, + 715 + ], + "type": "inline_equation", + "content": "v_{p}" + }, + { + "bbox": [ + 46, + 630, + 288, + 715 + ], + "type": "text", + "content": ", we define a rectangle that covers " + }, + { + "bbox": [ + 46, + 630, + 288, + 715 + ], + "type": "inline_equation", + "content": "3 \\times 3" + }, + { + "bbox": [ + 46, + 630, + 288, + 715 + ], + "type": "text", + "content": " cells centered at " + }, + { + "bbox": [ + 46, + 630, + 288, + 715 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 46, + 630, + 288, + 715 + ], + "type": "text", + "content": ", and assign to all the vertices within the rectangle, ensuring the" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 180, + 541, + 193 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 180, + 541, + 193 + ], + "spans": [ + { + "bbox": [ + 304, + 180, + 541, + 193 + ], + "type": "text", + "content": "propagation of the local motion " + }, + { + "bbox": [ + 304, + 180, + 541, + 193 + ], + "type": "inline_equation", + "content": "v_{p}" + }, + { + "bbox": [ + 304, + 180, + 541, + 193 + ], + "type": "text", + "content": " across the image plane." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 193, + 545, + 433 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 193, + 545, + 433 + ], + "spans": [ + { + "bbox": [ + 304, + 193, + 545, + 433 + ], + "type": "text", + "content": "Median Filters. The local motions of the middle points in all cells are propagated to their nearby mesh vertices, resulting in each vertex potentially receiving multiple motion vectors. To select the most appropriate motion vector for a given vertex, we apply a median filter " + }, + { + "bbox": [ + 304, + 193, + 545, + 433 + ], + "type": "inline_equation", + "content": "f_{1}" + }, + { + "bbox": [ + 304, + 193, + 545, + 433 + ], + "type": "text", + "content": " to filter the candidate motions. The response of the filter is then assigned to the corresponding vertex. The median filter is a widely used technique in optical flow estimation and has been shown to produce high-quality flow estimates [39]. Therefore, we use the median filter for sparse motion regularization. However, due to dynamic objects, the motion field may contain noise and needs to be spatially smoothed. To address this issue, we apply another median filter " + }, + { + "bbox": [ + 304, + 193, + 545, + 433 + ], + "type": "inline_equation", + "content": "f_{2}" + }, + { + "bbox": [ + 304, + 193, + 545, + 433 + ], + "type": "text", + "content": " that covers " + }, + { + "bbox": [ + 304, + 193, + 545, + 433 + ], + "type": "inline_equation", + "content": "3 \\times 3" + }, + { + "bbox": [ + 304, + 193, + 545, + 433 + ], + "type": "text", + "content": " cells neighborhood to suppress the noise in the motion field. This second median filter produces a spatially-smooth sparse motion field, which is what we called as meshflow. Ultimately, we use the meshflow (generally " + }, + { + "bbox": [ + 304, + 193, + 545, + 433 + ], + "type": "inline_equation", + "content": "16 \\times 16" + }, + { + "bbox": [ + 304, + 193, + 545, + 433 + ], + "type": "text", + "content": ") as the label, which contains global motion information. However, we upsample the meshflow to the full image resolution for intuitive display and alignment applications." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 441, + 528, + 453 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 441, + 528, + 453 + ], + "spans": [ + { + "bbox": [ + 306, + 441, + 528, + 453 + ], + "type": "text", + "content": "3.2. Estimation for Meshflow and Optical Flow" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 460, + 545, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 460, + 545, + 544 + ], + "spans": [ + { + "bbox": [ + 305, + 460, + 545, + 544 + ], + "type": "text", + "content": "Following [11] which estimates optical flow from two consecutive event sequences, we estimate the meshflow " + }, + { + "bbox": [ + 305, + 460, + 545, + 544 + ], + "type": "inline_equation", + "content": "MF_{k\\to k + 1}" + }, + { + "bbox": [ + 305, + 460, + 545, + 544 + ], + "type": "text", + "content": " and optical flow " + }, + { + "bbox": [ + 305, + 460, + 545, + 544 + ], + "type": "inline_equation", + "content": "F_{k\\rightarrow k + 1}" + }, + { + "bbox": [ + 305, + 460, + 545, + 544 + ], + "type": "text", + "content": " from event sequences " + }, + { + "bbox": [ + 305, + 460, + 545, + 544 + ], + "type": "inline_equation", + "content": "E(t_{k - 1},t_k)" + }, + { + "bbox": [ + 305, + 460, + 545, + 544 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 460, + 545, + 544 + ], + "type": "inline_equation", + "content": "E(t_{k},t_{k + 1})" + }, + { + "bbox": [ + 305, + 460, + 545, + 544 + ], + "type": "text", + "content": " , and overall architecture of our network is shown in Fig. 4. Fellow [54], we convert the inputs " + }, + { + "bbox": [ + 305, + 460, + 545, + 544 + ], + "type": "inline_equation", + "content": "E(t_{k - 1},t_k)" + }, + { + "bbox": [ + 305, + 460, + 545, + 544 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 460, + 545, + 544 + ], + "type": "inline_equation", + "content": "E(t_{k},t_{k + 1})" + }, + { + "bbox": [ + 305, + 460, + 545, + 544 + ], + "type": "text", + "content": " to the 3D volumes " + }, + { + "bbox": [ + 305, + 460, + 545, + 544 + ], + "type": "inline_equation", + "content": "V_{k - 1\\to k}" + }, + { + "bbox": [ + 305, + 460, + 545, + 544 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 460, + 545, + 544 + ], + "type": "inline_equation", + "content": "V_{k\\to k + 1}" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 550, + 514, + 562 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 550, + 514, + 562 + ], + "spans": [ + { + "bbox": [ + 306, + 550, + 514, + 562 + ], + "type": "text", + "content": "3.2.1 Efficient Event-based Meshflow Network" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 570, + 547, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 547, + 617 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 547, + 617 + ], + "type": "text", + "content": "We propose EEMFlow to directly output results of the same resolution as the ground truth meshflow " + }, + { + "bbox": [ + 304, + 570, + 547, + 617 + ], + "type": "inline_equation", + "content": "MF_{GT}" + }, + { + "bbox": [ + 304, + 570, + 547, + 617 + ], + "type": "text", + "content": " for supervised regression, fully leveraging the advantage about low-parameter and high-motion-information of meshflow." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 618, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 618, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 618, + 547, + 714 + ], + "type": "text", + "content": "Overall Structure of Meshflow Estimation. Since meshflow focuses more on global large motion rather than local detailed motion, the EEMFlow we designed does not require excessively deep network layers or refinement operations from coarse to fine. Firstly, EEMFlow employs a three-level pyramid feature encoder to extracting features " + }, + { + "bbox": [ + 304, + 618, + 547, + 714 + ], + "type": "inline_equation", + "content": "(V_{t_{k - 1}}^i)^N" + }, + { + "bbox": [ + 304, + 618, + 547, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 618, + 547, + 714 + ], + "type": "inline_equation", + "content": "(V_{t_k}^i)^N" + }, + { + "bbox": [ + 304, + 618, + 547, + 714 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 304, + 618, + 547, + 714 + ], + "type": "inline_equation", + "content": "V_{k - 1\\rightarrow k}" + }, + { + "bbox": [ + 304, + 618, + 547, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 618, + 547, + 714 + ], + "type": "inline_equation", + "content": "V_{k\\rightarrow k + 1}" + }, + { + "bbox": [ + 304, + 618, + 547, + 714 + ], + "type": "text", + "content": ", the convolutional layers within the " + }, + { + "bbox": [ + 304, + 618, + 547, + 714 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 618, + 547, + 714 + ], + "type": "text", + "content": "-th level share weights for" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "19200" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 111, + 71, + 159, + 108 + ], + "blocks": [ + { + "bbox": [ + 111, + 71, + 159, + 108 + ], + "lines": [ + { + "bbox": [ + 111, + 71, + 159, + 108 + ], + "spans": [ + { + "bbox": [ + 111, + 71, + 159, + 108 + ], + "type": "image", + "image_path": "33b37661ca91328013c7b981dcb7665f1f604e5388fa28542ea364690667490e.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 161, + 72, + 224, + 108 + ], + "blocks": [ + { + "bbox": [ + 161, + 72, + 224, + 108 + ], + "lines": [ + { + "bbox": [ + 161, + 72, + 224, + 108 + ], + "spans": [ + { + "bbox": [ + 161, + 72, + 224, + 108 + ], + "type": "image", + "image_path": "524953267b60898dd498f1c6910ed2b2b00942ba01d6f74a061deaf9306589f8.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 225, + 72, + 288, + 108 + ], + "blocks": [ + { + "bbox": [ + 225, + 72, + 288, + 108 + ], + "lines": [ + { + "bbox": [ + 225, + 72, + 288, + 108 + ], + "spans": [ + { + "bbox": [ + 225, + 72, + 288, + 108 + ], + "type": "image", + "image_path": "98485f598180dc52b2da7a9ca4343b5785631b13062c9f91a090a92b91ced924.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 289, + 72, + 353, + 108 + ], + "blocks": [ + { + "bbox": [ + 289, + 72, + 353, + 108 + ], + "lines": [ + { + "bbox": [ + 289, + 72, + 353, + 108 + ], + "spans": [ + { + "bbox": [ + 289, + 72, + 353, + 108 + ], + "type": "image", + "image_path": "296bbbf2038ecd81415d972fd646e69c5526cfb5e6b3cadf82def1cd576156b3.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 353, + 72, + 416, + 108 + ], + "blocks": [ + { + "bbox": [ + 353, + 72, + 416, + 108 + ], + "lines": [ + { + "bbox": [ + 353, + 72, + 416, + 108 + ], + "spans": [ + { + "bbox": [ + 353, + 72, + 416, + 108 + ], + "type": "image", + "image_path": "ae38838da1ba53a93b669380d73b4aa8ea445d23da337d69163c46ab90d30b32.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 416, + 72, + 479, + 108 + ], + "blocks": [ + { + "bbox": [ + 416, + 72, + 479, + 108 + ], + "lines": [ + { + "bbox": [ + 416, + 72, + 479, + 108 + ], + "spans": [ + { + "bbox": [ + 416, + 72, + 479, + 108 + ], + "type": "image", + "image_path": "3c64997abcf1dc0f4ae722197050cf63da35f51e3c074437e871150fbdb1ed83.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 481, + 72, + 544, + 108 + ], + "blocks": [ + { + "bbox": [ + 481, + 72, + 544, + 108 + ], + "lines": [ + { + "bbox": [ + 481, + 72, + 544, + 108 + ], + "spans": [ + { + "bbox": [ + 481, + 72, + 544, + 108 + ], + "type": "image", + "image_path": "42824c3e67f0fe77fdb1b91057b4ecd6abf4dad327616be3e97262a3143cbe18.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 62, + 110, + 110, + 147 + ], + "blocks": [ + { + "bbox": [ + 50, + 86, + 60, + 95 + ], + "lines": [ + { + "bbox": [ + 50, + 86, + 60, + 95 + ], + "spans": [ + { + "bbox": [ + 50, + 86, + 60, + 95 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 50, + 124, + 60, + 133 + ], + "lines": [ + { + "bbox": [ + 50, + 124, + 60, + 133 + ], + "spans": [ + { + "bbox": [ + 50, + 124, + 60, + 133 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 50, + 163, + 59, + 171 + ], + "lines": [ + { + "bbox": [ + 50, + 163, + 59, + 171 + ], + "spans": [ + { + "bbox": [ + 50, + 163, + 59, + 171 + ], + "type": "text", + "content": "(c)" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 62, + 110, + 110, + 147 + ], + "lines": [ + { + "bbox": [ + 62, + 110, + 110, + 147 + ], + "spans": [ + { + "bbox": [ + 62, + 110, + 110, + 147 + ], + "type": "image", + "image_path": "e3085ee4e352aee8dfaa35cb0b6c6c4fef12141db8871d3876f8bd257503e64b.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 62, + 163, + 76, + 171 + ], + "lines": [ + { + "bbox": [ + 62, + 163, + 76, + 171 + ], + "spans": [ + { + "bbox": [ + 62, + 163, + 76, + 171 + ], + "type": "text", + "content": "=" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 111, + 110, + 160, + 147 + ], + "blocks": [ + { + "bbox": [ + 111, + 110, + 160, + 147 + ], + "lines": [ + { + "bbox": [ + 111, + 110, + 160, + 147 + ], + "spans": [ + { + "bbox": [ + 111, + 110, + 160, + 147 + ], + "type": "image", + "image_path": "ec1bef5238fcd6e20a8974d3a4399269eb7d129f9ce878c8f7e5d0e6180c9115.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 162, + 110, + 224, + 147 + ], + "blocks": [ + { + "bbox": [ + 162, + 110, + 224, + 147 + ], + "lines": [ + { + "bbox": [ + 162, + 110, + 224, + 147 + ], + "spans": [ + { + "bbox": [ + 162, + 110, + 224, + 147 + ], + "type": "image", + "image_path": "4f5da1305ea57a74698431b28d27e8cf3023b3cf28d649bc03958a6c9bc9b908.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 225, + 110, + 288, + 147 + ], + "blocks": [ + { + "bbox": [ + 225, + 110, + 288, + 147 + ], + "lines": [ + { + "bbox": [ + 225, + 110, + 288, + 147 + ], + "spans": [ + { + "bbox": [ + 225, + 110, + 288, + 147 + ], + "type": "image", + "image_path": "b6cb352a7624bacec943dac5d04832baf37279a5e000eef99ca709dd54233b98.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 289, + 110, + 352, + 147 + ], + "blocks": [ + { + "bbox": [ + 289, + 110, + 352, + 147 + ], + "lines": [ + { + "bbox": [ + 289, + 110, + 352, + 147 + ], + "spans": [ + { + "bbox": [ + 289, + 110, + 352, + 147 + ], + "type": "image", + "image_path": "34114591dd5335d2594b39256e06abe090f974ce36a0102e11e1d941d4c736a1.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 353, + 110, + 416, + 147 + ], + "blocks": [ + { + "bbox": [ + 353, + 110, + 416, + 147 + ], + "lines": [ + { + "bbox": [ + 353, + 110, + 416, + 147 + ], + "spans": [ + { + "bbox": [ + 353, + 110, + 416, + 147 + ], + "type": "image", + "image_path": "4c041d142d57b2125cab9c9617fefffe5632cd0ed2e821e70e25c0171efa2067.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 416, + 110, + 480, + 147 + ], + "blocks": [ + { + "bbox": [ + 416, + 110, + 480, + 147 + ], + "lines": [ + { + "bbox": [ + 416, + 110, + 480, + 147 + ], + "spans": [ + { + "bbox": [ + 416, + 110, + 480, + 147 + ], + "type": "image", + "image_path": "a9eb40e1bcf52234e02028eb35c8e30c6d3610f3ee23d5ca0fbf3baa4a7ba829.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 481, + 110, + 544, + 147 + ], + "blocks": [ + { + "bbox": [ + 481, + 110, + 544, + 147 + ], + "lines": [ + { + "bbox": [ + 481, + 110, + 544, + 147 + ], + "spans": [ + { + "bbox": [ + 481, + 110, + 544, + 147 + ], + "type": "image", + "image_path": "c17957b13cb923135c3fc2b38d2b70f18641160ed975f17e25faac8c5f3459f8.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 111, + 148, + 160, + 186 + ], + "blocks": [ + { + "bbox": [ + 62, + 186, + 99, + 194 + ], + "lines": [ + { + "bbox": [ + 62, + 186, + 99, + 194 + ], + "spans": [ + { + "bbox": [ + 62, + 186, + 99, + 194 + ], + "type": "text", + "content": "MVSEC" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 111, + 148, + 160, + 186 + ], + "lines": [ + { + "bbox": [ + 111, + 148, + 160, + 186 + ], + "spans": [ + { + "bbox": [ + 111, + 148, + 160, + 186 + ], + "type": "image", + "image_path": "5403db35b83329e7d27bedc5cd657e5aa7c9f3c1153c99a8c9fbb4d4c4397a44.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 127, + 187, + 146, + 194 + ], + "lines": [ + { + "bbox": [ + 127, + 187, + 146, + 194 + ], + "spans": [ + { + "bbox": [ + 127, + 187, + 146, + 194 + ], + "type": "text", + "content": "DSEC" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 201, + 547, + 231 + ], + "lines": [ + { + "bbox": [ + 46, + 201, + 547, + 231 + ], + "spans": [ + { + "bbox": [ + 46, + 201, + 547, + 231 + ], + "type": "text", + "content": "Figure 2. Examples from MVSEC [53], DSEC [10] and our dataset. (a): Event data, (b): Optical flow, (c): Upsampled meshflow. The meshflows of MVSEC and DSEC contain a significant number of ineffective areas due to the sparsity of optical flow. In contrast, our meshflow provides a complete global motion field. Best viewed on a color screen in high resolution." + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_caption" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 162, + 148, + 225, + 186 + ], + "blocks": [ + { + "bbox": [ + 162, + 148, + 225, + 186 + ], + "lines": [ + { + "bbox": [ + 162, + 148, + 225, + 186 + ], + "spans": [ + { + "bbox": [ + 162, + 148, + 225, + 186 + ], + "type": "image", + "image_path": "05f11b9af16dd20ed103d3731b84580f7f0b85504c11864375085c44485527ff.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 226, + 148, + 288, + 186 + ], + "blocks": [ + { + "bbox": [ + 226, + 148, + 288, + 186 + ], + "lines": [ + { + "bbox": [ + 226, + 148, + 288, + 186 + ], + "spans": [ + { + "bbox": [ + 226, + 148, + 288, + 186 + ], + "type": "image", + "image_path": "4d13999406ad78d44cbf5994bbe6f7ca0dd8b791620e4eded31b81cafe09cf15.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 290, + 148, + 352, + 186 + ], + "blocks": [ + { + "bbox": [ + 290, + 148, + 352, + 186 + ], + "lines": [ + { + "bbox": [ + 290, + 148, + 352, + 186 + ], + "spans": [ + { + "bbox": [ + 290, + 148, + 352, + 186 + ], + "type": "image", + "image_path": "6f29cf738b64611c320f3be1c33abf69738dd92940d35b5b5c8f86f71f6e2de4.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 334, + 186, + 374, + 194 + ], + "lines": [ + { + "bbox": [ + 334, + 186, + 374, + 194 + ], + "spans": [ + { + "bbox": [ + 334, + 186, + 374, + 194 + ], + "type": "text", + "content": "HREM(Ours)" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_caption" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 353, + 148, + 416, + 186 + ], + "blocks": [ + { + "bbox": [ + 353, + 148, + 416, + 186 + ], + "lines": [ + { + "bbox": [ + 353, + 148, + 416, + 186 + ], + "spans": [ + { + "bbox": [ + 353, + 148, + 416, + 186 + ], + "type": "image", + "image_path": "bad93513d554583ceb6ec9093f764d952a6e20c7d1b224524ddfa13c6d4cb777.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 417, + 148, + 480, + 186 + ], + "blocks": [ + { + "bbox": [ + 417, + 148, + 480, + 186 + ], + "lines": [ + { + "bbox": [ + 417, + 148, + 480, + 186 + ], + "spans": [ + { + "bbox": [ + 417, + 148, + 480, + 186 + ], + "type": "image", + "image_path": "e3e2e8af64738c9372cfb15a4190afefe67c589c7d7bd3a9f5fa27e68f4284c2.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 481, + 148, + 544, + 186 + ], + "blocks": [ + { + "bbox": [ + 481, + 148, + 544, + 186 + ], + "lines": [ + { + "bbox": [ + 481, + 148, + 544, + 186 + ], + "spans": [ + { + "bbox": [ + 481, + 148, + 544, + 186 + ], + "type": "image", + "image_path": "ca48efec2a148cbd2aa73d2cda1d2ff2b19e34631b7dfc3310cb19ada99f8c16.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_body" + } + ], + "index": 27 + }, + { + "type": "image", + "bbox": [ + 48, + 249, + 126, + 293 + ], + "blocks": [ + { + "bbox": [ + 48, + 249, + 126, + 293 + ], + "lines": [ + { + "bbox": [ + 48, + 249, + 126, + 293 + ], + "spans": [ + { + "bbox": [ + 48, + 249, + 126, + 293 + ], + "type": "image", + "image_path": "41c32b84fab9b46c89f32de5b57f75dd06a3d3b6b99de90272e0161bdcd7f680.jpg" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 82, + 293, + 91, + 300 + ], + "lines": [ + { + "bbox": [ + 82, + 293, + 91, + 300 + ], + "spans": [ + { + "bbox": [ + 82, + 293, + 91, + 300 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 310, + 287, + 369 + ], + "lines": [ + { + "bbox": [ + 46, + 310, + 287, + 369 + ], + "spans": [ + { + "bbox": [ + 46, + 310, + 287, + 369 + ], + "type": "text", + "content": "Figure 3. The process of generating meshflow from dense optical flow. (a) Propagate the motion vector of each grid center to the grid vertices. (b) Apply median filter " + }, + { + "bbox": [ + 46, + 310, + 287, + 369 + ], + "type": "inline_equation", + "content": "f_{1}" + }, + { + "bbox": [ + 46, + 310, + 287, + 369 + ], + "type": "text", + "content": " to multiple motion vectors of each vertex to select the most appropriate motion for that vertex. (c) Use median filtering " + }, + { + "bbox": [ + 46, + 310, + 287, + 369 + ], + "type": "inline_equation", + "content": "f_{2}" + }, + { + "bbox": [ + 46, + 310, + 287, + 369 + ], + "type": "text", + "content": " to smooth the motion field in the mesh grid. For ease of visualization, we present the " + }, + { + "bbox": [ + 46, + 310, + 287, + 369 + ], + "type": "inline_equation", + "content": "8 \\times 8" + }, + { + "bbox": [ + 46, + 310, + 287, + 369 + ], + "type": "text", + "content": " grid mesh in this paper." + } + ] + } + ], + "index": 36, + "angle": 0, + "type": "image_caption" + } + ], + "index": 30 + }, + { + "type": "image", + "bbox": [ + 127, + 250, + 205, + 293 + ], + "blocks": [ + { + "bbox": [ + 127, + 250, + 205, + 293 + ], + "lines": [ + { + "bbox": [ + 127, + 250, + 205, + 293 + ], + "spans": [ + { + "bbox": [ + 127, + 250, + 205, + 293 + ], + "type": "image", + "image_path": "5ed42728f070a615355e62212338e2db4a1f3b4a4db9b7d464dcedc4afe647c7.jpg" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 160, + 293, + 170, + 300 + ], + "lines": [ + { + "bbox": [ + 160, + 293, + 170, + 300 + ], + "spans": [ + { + "bbox": [ + 160, + 293, + 170, + 300 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_caption" + } + ], + "index": 32 + }, + { + "type": "image", + "bbox": [ + 206, + 250, + 286, + 293 + ], + "blocks": [ + { + "bbox": [ + 206, + 250, + 286, + 293 + ], + "lines": [ + { + "bbox": [ + 206, + 250, + 286, + 293 + ], + "spans": [ + { + "bbox": [ + 206, + 250, + 286, + 293 + ], + "type": "image", + "image_path": "03e081b20510d66517837d7e44b658e9805d21ab18742ce6b1f3211502508cc7.jpg" + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 240, + 293, + 249, + 300 + ], + "lines": [ + { + "bbox": [ + 240, + 293, + 249, + 300 + ], + "spans": [ + { + "bbox": [ + 240, + 293, + 249, + 300 + ], + "type": "text", + "content": "(c)" + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "image_caption" + } + ], + "index": 34 + }, + { + "bbox": [ + 46, + 389, + 286, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 389, + 286, + 569 + ], + "spans": [ + { + "bbox": [ + 46, + 389, + 286, + 569 + ], + "type": "inline_equation", + "content": "V_{t_{k-1}}^i" + }, + { + "bbox": [ + 46, + 389, + 286, + 569 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 389, + 286, + 569 + ], + "type": "inline_equation", + "content": "V_{t_k}^i" + }, + { + "bbox": [ + 46, + 389, + 286, + 569 + ], + "type": "text", + "content": ". Secondly, EEMFlow utilizes multi-scale feature correlation to builds the cost volumes for meshflow estimation. Features " + }, + { + "bbox": [ + 46, + 389, + 286, + 569 + ], + "type": "inline_equation", + "content": "(V_{t_{k-1}}^i)^N" + }, + { + "bbox": [ + 46, + 389, + 286, + 569 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 389, + 286, + 569 + ], + "type": "inline_equation", + "content": "(V_{t_k}^i)^N" + }, + { + "bbox": [ + 46, + 389, + 286, + 569 + ], + "type": "text", + "content": " undergo average pooling operation " + }, + { + "bbox": [ + 46, + 389, + 286, + 569 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 46, + 389, + 286, + 569 + ], + "type": "text", + "content": " to the same resolution (1/64 resolution of " + }, + { + "bbox": [ + 46, + 389, + 286, + 569 + ], + "type": "inline_equation", + "content": "V_{k\\rightarrow k+1}" + }, + { + "bbox": [ + 46, + 389, + 286, + 569 + ], + "type": "text", + "content": "), and then use correlation to capture relative motion information and output cost volumes " + }, + { + "bbox": [ + 46, + 389, + 286, + 569 + ], + "type": "inline_equation", + "content": "(C^i)^N" + }, + { + "bbox": [ + 46, + 389, + 286, + 569 + ], + "type": "text", + "content": ". Specifically, we employ the dilated feature correlation (DFC) to increase the search area while reducing computational parameters. Finally, we stack the cost volumes " + }, + { + "bbox": [ + 46, + 389, + 286, + 569 + ], + "type": "inline_equation", + "content": "(C^i)^N" + }, + { + "bbox": [ + 46, + 389, + 286, + 569 + ], + "type": "text", + "content": " and features " + }, + { + "bbox": [ + 46, + 389, + 286, + 569 + ], + "type": "inline_equation", + "content": "(V_{t_{k-1}}^i)^N" + }, + { + "bbox": [ + 46, + 389, + 286, + 569 + ], + "type": "text", + "content": " after pooling, fuse them with a weighted-sum operation, and then feed them into the decoders to regress the meshflow " + }, + { + "bbox": [ + 46, + 389, + 286, + 569 + ], + "type": "inline_equation", + "content": "MF_{k\\rightarrow k+1}" + }, + { + "bbox": [ + 46, + 389, + 286, + 569 + ], + "type": "text", + "content": " at the same resolution as " + }, + { + "bbox": [ + 46, + 389, + 286, + 569 + ], + "type": "inline_equation", + "content": "MF_{GT}" + }, + { + "bbox": [ + 46, + 389, + 286, + 569 + ], + "type": "text", + "content": ". Inspiring by [51], we replace the conventional convolutions with group shuffle convolutions, which leads to efficient computation while maintaining high accuracy." + } + ] + } + ], + "index": 37 + }, + { + "bbox": [ + 46, + 570, + 286, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 570, + 286, + 605 + ], + "spans": [ + { + "bbox": [ + 46, + 570, + 286, + 605 + ], + "type": "text", + "content": "Dilated Feature Correlation. We use the inner product to calculate correlation between " + }, + { + "bbox": [ + 46, + 570, + 286, + 605 + ], + "type": "inline_equation", + "content": "V_{t_{k-1}}^i" + }, + { + "bbox": [ + 46, + 570, + 286, + 605 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 570, + 286, + 605 + ], + "type": "inline_equation", + "content": "V_{t_k}^i" + }, + { + "bbox": [ + 46, + 570, + 286, + 605 + ], + "type": "text", + "content": " for meshflow estimation:" + } + ] + } + ], + "index": 38 + }, + { + "bbox": [ + 64, + 616, + 287, + 631 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 616, + 287, + 631 + ], + "spans": [ + { + "bbox": [ + 64, + 616, + 287, + 631 + ], + "type": "interline_equation", + "content": "C ^ {i} (\\boldsymbol {u}, \\boldsymbol {d}) = V _ {t _ {k}} ^ {i} (\\boldsymbol {u}) \\cdot V _ {t _ {k - 1}} ^ {i} (\\boldsymbol {u} + \\boldsymbol {d}) / M, \\boldsymbol {d} \\in \\mathcal {N}, \\tag {1}", + "image_path": "8b7c4feff24f0385c7c617cb84a2e034bb205af4dd2bae0d27a9705b1a220da1.jpg" + } + ] + } + ], + "index": 39 + }, + { + "bbox": [ + 46, + 641, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 641, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 641, + 287, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 641, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\pmb{u}" + }, + { + "bbox": [ + 46, + 641, + 287, + 714 + ], + "type": "text", + "content": " represents the spatial coordinates on " + }, + { + "bbox": [ + 46, + 641, + 287, + 714 + ], + "type": "inline_equation", + "content": "V_{t_k}^i" + }, + { + "bbox": [ + 46, + 641, + 287, + 714 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 641, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{N}" + }, + { + "bbox": [ + 46, + 641, + 287, + 714 + ], + "type": "text", + "content": " represents the search grid of coordinate " + }, + { + "bbox": [ + 46, + 641, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\pmb{u}" + }, + { + "bbox": [ + 46, + 641, + 287, + 714 + ], + "type": "text", + "content": " in feature " + }, + { + "bbox": [ + 46, + 641, + 287, + 714 + ], + "type": "inline_equation", + "content": "V_{t_{k - 1}}^i" + }, + { + "bbox": [ + 46, + 641, + 287, + 714 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 641, + 287, + 714 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 46, + 641, + 287, + 714 + ], + "type": "text", + "content": " represents the number of elements in " + }, + { + "bbox": [ + 46, + 641, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{N}" + }, + { + "bbox": [ + 46, + 641, + 287, + 714 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 641, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\pmb{d}" + }, + { + "bbox": [ + 46, + 641, + 287, + 714 + ], + "type": "text", + "content": " represents the offset coordinates of the elements in " + }, + { + "bbox": [ + 46, + 641, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{N}" + }, + { + "bbox": [ + 46, + 641, + 287, + 714 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 641, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 46, + 641, + 287, + 714 + ], + "type": "text", + "content": " represents the inner product calculation. Many methods like [16, 40] simply define " + }, + { + "bbox": [ + 46, + 641, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{N}" + }, + { + "bbox": [ + 46, + 641, + 287, + 714 + ], + "type": "text", + "content": " as a square range of size " + }, + { + "bbox": [ + 46, + 641, + 287, + 714 + ], + "type": "inline_equation", + "content": "(2r + 1)\\times (2r + 1)" + } + ] + } + ], + "index": 40 + }, + { + "bbox": [ + 304, + 251, + 545, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 251, + 545, + 312 + ], + "spans": [ + { + "bbox": [ + 304, + 251, + 545, + 312 + ], + "type": "text", + "content": "and observe that increasing the radius " + }, + { + "bbox": [ + 304, + 251, + 545, + 312 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 304, + 251, + 545, + 312 + ], + "type": "text", + "content": " can reduce errors but increase computational overhead. We propose dilated feature correlation that samples densely around the center and sparsely at farther distances, thereby reducing computation while enabling larger radius cost volumes, as Eq. 2:" + } + ] + } + ], + "index": 41 + }, + { + "bbox": [ + 315, + 328, + 545, + 360 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 328, + 545, + 360 + ], + "spans": [ + { + "bbox": [ + 315, + 328, + 545, + 360 + ], + "type": "interline_equation", + "content": "\\mathcal {N} \\left(d _ {x}, d _ {y}\\right) = \\left\\{ \\begin{array}{l} 0, \\text {i f} | d _ {x} | + | d _ {y} | = 2 k, k \\in [ 2, r ], \\\\ 1, \\text {o t h e r s ,} \\end{array} \\right. \\tag {2}", + "image_path": "a2df93b7410274115c5bc8cf79e958ad153daeb6854b9d22cc155b853755ed49.jpg" + } + ] + } + ], + "index": 42 + }, + { + "bbox": [ + 304, + 365, + 545, + 414 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 365, + 545, + 414 + ], + "spans": [ + { + "bbox": [ + 304, + 365, + 545, + 414 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 365, + 545, + 414 + ], + "type": "inline_equation", + "content": "\\pmb{d} = (d_x, d_y)" + }, + { + "bbox": [ + 304, + 365, + 545, + 414 + ], + "type": "text", + "content": " represents the relative position coordinates within the neighborhood " + }, + { + "bbox": [ + 304, + 365, + 545, + 414 + ], + "type": "inline_equation", + "content": "\\mathcal{N}" + }, + { + "bbox": [ + 304, + 365, + 545, + 414 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 365, + 545, + 414 + ], + "type": "inline_equation", + "content": "\\mathcal{N}(d_x, d_y) = 1" + }, + { + "bbox": [ + 304, + 365, + 545, + 414 + ], + "type": "text", + "content": " indicates the computation of correlation, and " + }, + { + "bbox": [ + 304, + 365, + 545, + 414 + ], + "type": "inline_equation", + "content": "\\mathcal{N}(d_x, d_y) = 0" + }, + { + "bbox": [ + 304, + 365, + 545, + 414 + ], + "type": "text", + "content": " signifies no computation." + } + ] + } + ], + "index": 43 + }, + { + "bbox": [ + 306, + 428, + 488, + 441 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 428, + 488, + 441 + ], + "spans": [ + { + "bbox": [ + 306, + 428, + 488, + 441 + ], + "type": "text", + "content": "3.2.2 Event-based Optical Flow Network" + } + ] + } + ], + "index": 44 + }, + { + "bbox": [ + 304, + 447, + 545, + 519 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 447, + 545, + 519 + ], + "spans": [ + { + "bbox": [ + 304, + 447, + 545, + 519 + ], + "type": "text", + "content": "Based on EEMFlow, we make some improvements for accurate estimation of optical flow " + }, + { + "bbox": [ + 304, + 447, + 545, + 519 + ], + "type": "inline_equation", + "content": "F_{k \\to k + 1}" + }, + { + "bbox": [ + 304, + 447, + 545, + 519 + ], + "type": "text", + "content": ", upgrading it to EEMFlow+. Since optical flow focuses more on local motion and pays attention to object edge details, we employ the coarse to fine residual approach to progressively refine the flow, which can be expressed as Eq. 3." + } + ] + } + ], + "index": 45 + }, + { + "bbox": [ + 330, + 525, + 545, + 540 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 330, + 525, + 545, + 540 + ], + "spans": [ + { + "bbox": [ + 330, + 525, + 545, + 540 + ], + "type": "interline_equation", + "content": "F ^ {i + 1} = \\operatorname {C o n v} ^ {i} \\left(\\mathcal {C} \\left(V _ {t _ {k - 1}} ^ {i}, \\mathcal {W} \\left(V _ {t _ {k}} ^ {i}, F _ {\\uparrow} ^ {i}\\right)\\right)\\right) + F _ {\\uparrow} ^ {i}, \\tag {3}", + "image_path": "58857970c7903e04132e2ed49455c69af3325bc6c6eda559320667f26c081cc1.jpg" + } + ] + } + ], + "index": 46 + }, + { + "bbox": [ + 304, + 545, + 545, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 545, + 545, + 677 + ], + "spans": [ + { + "bbox": [ + 304, + 545, + 545, + 677 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 545, + 545, + 677 + ], + "type": "inline_equation", + "content": "\\mathcal{C}(\\cdot, \\cdot), \\mathcal{W}(\\cdot, \\cdot)" + }, + { + "bbox": [ + 304, + 545, + 545, + 677 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 545, + 545, + 677 + ], + "type": "inline_equation", + "content": "\\uparrow" + }, + { + "bbox": [ + 304, + 545, + 545, + 677 + ], + "type": "text", + "content": " donates our dilated feature correlation, the warping operation and upsampling, respectively. We employ the pyramid decoders for optical flow, thus " + }, + { + "bbox": [ + 304, + 545, + 545, + 677 + ], + "type": "inline_equation", + "content": "F^i" + }, + { + "bbox": [ + 304, + 545, + 545, + 677 + ], + "type": "text", + "content": " is the output flow of the " + }, + { + "bbox": [ + 304, + 545, + 545, + 677 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 545, + 545, + 677 + ], + "type": "text", + "content": "-th decoder, " + }, + { + "bbox": [ + 304, + 545, + 545, + 677 + ], + "type": "inline_equation", + "content": "F^{i+1}" + }, + { + "bbox": [ + 304, + 545, + 545, + 677 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 545, + 545, + 677 + ], + "type": "inline_equation", + "content": "\\mathrm{Conv}^i" + }, + { + "bbox": [ + 304, + 545, + 545, + 677 + ], + "type": "text", + "content": " are respectively the output flow and convolutions in the " + }, + { + "bbox": [ + 304, + 545, + 545, + 677 + ], + "type": "inline_equation", + "content": "i+1" + }, + { + "bbox": [ + 304, + 545, + 545, + 677 + ], + "type": "text", + "content": "-th decoder. The most noteworthy aspect is upsampling " + }, + { + "bbox": [ + 304, + 545, + 545, + 677 + ], + "type": "inline_equation", + "content": "F^i" + }, + { + "bbox": [ + 304, + 545, + 545, + 677 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 304, + 545, + 545, + 677 + ], + "type": "inline_equation", + "content": "F_{\\uparrow}^i" + }, + { + "bbox": [ + 304, + 545, + 545, + 677 + ], + "type": "text", + "content": ". Many methods [6, 16, 40, 52] use bilinear interpolation for upsampling, but this can lead to the mixing of incorrect motions at object edges, resulting in blurring. Therefore, we propose the confidence-induced detail completion module for upsampling to enhance edge details." + } + ] + } + ], + "index": 47 + }, + { + "bbox": [ + 305, + 689, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 689, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 305, + 689, + 545, + 714 + ], + "type": "text", + "content": "Confidence-induced Detail Completion Module. We propose the confidence-induced detail completion module" + } + ] + } + ], + "index": 48 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19201" + } + ] + } + ], + "index": 49 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 49, + 71, + 547, + 183 + ], + "blocks": [ + { + "bbox": [ + 49, + 71, + 547, + 183 + ], + "lines": [ + { + "bbox": [ + 49, + 71, + 547, + 183 + ], + "spans": [ + { + "bbox": [ + 49, + 71, + 547, + 183 + ], + "type": "image", + "image_path": "77f3abc72bc1cb2188873eb9d46e2ab42e3ec3aa18c7b4727ce9821bc1397469.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 190, + 547, + 231 + ], + "lines": [ + { + "bbox": [ + 46, + 190, + 547, + 231 + ], + "spans": [ + { + "bbox": [ + 46, + 190, + 547, + 231 + ], + "type": "text", + "content": "Figure 4. Our proposed network architecture. We employ pyramid encoders to extract multi-scale features from " + }, + { + "bbox": [ + 46, + 190, + 547, + 231 + ], + "type": "inline_equation", + "content": "V_{k-1 \\to k}" + }, + { + "bbox": [ + 46, + 190, + 547, + 231 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 190, + 547, + 231 + ], + "type": "inline_equation", + "content": "V_{k \\to k+1}" + }, + { + "bbox": [ + 46, + 190, + 547, + 231 + ], + "type": "text", + "content": ", then use dilated feature correlation to compute the cost volume between each layer of features, followed by decoding to output the results. For meshflow estimation, we utilize the decoders with group shuffle convolutions to output predictions " + }, + { + "bbox": [ + 46, + 190, + 547, + 231 + ], + "type": "inline_equation", + "content": "MF_{k \\to k+1}" + }, + { + "bbox": [ + 46, + 190, + 547, + 231 + ], + "type": "text", + "content": ", upsampled to " + }, + { + "bbox": [ + 46, + 190, + 547, + 231 + ], + "type": "inline_equation", + "content": "MF \\uparrow_{k \\to k+1}" + }, + { + "bbox": [ + 46, + 190, + 547, + 231 + ], + "type": "text", + "content": ". For optical flow estimation, we refine flows using a coarse-to-fine residual approach and confidence-induced detail completion module, finally outputting the optical flow " + }, + { + "bbox": [ + 46, + 190, + 547, + 231 + ], + "type": "inline_equation", + "content": "F_{k \\to k+1}" + }, + { + "bbox": [ + 46, + 190, + 547, + 231 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 48, + 248, + 287, + 364 + ], + "blocks": [ + { + "bbox": [ + 48, + 248, + 287, + 364 + ], + "lines": [ + { + "bbox": [ + 48, + 248, + 287, + 364 + ], + "spans": [ + { + "bbox": [ + 48, + 248, + 287, + 364 + ], + "type": "image", + "image_path": "dde6960db194bab40144fcb44dd2a88573264479b9b3ba324a806464bea22868.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 371, + 288, + 411 + ], + "lines": [ + { + "bbox": [ + 46, + 371, + 288, + 411 + ], + "spans": [ + { + "bbox": [ + 46, + 371, + 288, + 411 + ], + "type": "text", + "content": "Figure 5. The structure of CDC. CDC employs self-corrector based on a dense convolutional network and self-correlation based on a self-attention mechanism to correct the flow obtained from bilinear upsampling. " + }, + { + "bbox": [ + 46, + 371, + 288, + 411 + ], + "type": "inline_equation", + "content": "\\mathcal{B}" + }, + { + "bbox": [ + 46, + 371, + 288, + 411 + ], + "type": "text", + "content": " represents bilinear upsampling." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 430, + 289, + 660 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 430, + 289, + 660 + ], + "spans": [ + { + "bbox": [ + 46, + 430, + 289, + 660 + ], + "type": "text", + "content": "(CDC) to eliminate the blurring of object edges caused by the mixing of multiple motions at the junctions of different movements during upsampling. The detailed structure of our CDC is shown in Fig. 5. Given a small-scale flow " + }, + { + "bbox": [ + 46, + 430, + 289, + 660 + ], + "type": "inline_equation", + "content": "F^i" + }, + { + "bbox": [ + 46, + 430, + 289, + 660 + ], + "type": "text", + "content": " from the " + }, + { + "bbox": [ + 46, + 430, + 289, + 660 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 430, + 289, + 660 + ], + "type": "text", + "content": "-th level, our CDC first generates an initial flow " + }, + { + "bbox": [ + 46, + 430, + 289, + 660 + ], + "type": "inline_equation", + "content": "\\bar{F}^i" + }, + { + "bbox": [ + 46, + 430, + 289, + 660 + ], + "type": "text", + "content": " for the " + }, + { + "bbox": [ + 46, + 430, + 289, + 660 + ], + "type": "inline_equation", + "content": "i + 1" + }, + { + "bbox": [ + 46, + 430, + 289, + 660 + ], + "type": "text", + "content": "-th level through bilinear interpolation, and then arranges the self-corrector and self-correlation branches to correct it. Self-corrector is based on a dense convolutional network with a five-layer structure. It captures motion information within the edge neighborhood through dense convolution from the concatenated feature " + }, + { + "bbox": [ + 46, + 430, + 289, + 660 + ], + "type": "inline_equation", + "content": "V_{t_{k-1}}^i" + }, + { + "bbox": [ + 46, + 430, + 289, + 660 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 430, + 289, + 660 + ], + "type": "inline_equation", + "content": "V_{t_k}^i" + }, + { + "bbox": [ + 46, + 430, + 289, + 660 + ], + "type": "text", + "content": ", outputting the corrected flow " + }, + { + "bbox": [ + 46, + 430, + 289, + 660 + ], + "type": "inline_equation", + "content": "\\Delta F^i" + }, + { + "bbox": [ + 46, + 430, + 289, + 660 + ], + "type": "text", + "content": " and the corrected confidence map " + }, + { + "bbox": [ + 46, + 430, + 289, + 660 + ], + "type": "inline_equation", + "content": "W^i" + }, + { + "bbox": [ + 46, + 430, + 289, + 660 + ], + "type": "text", + "content": ". Self-correlation is based on a self-attention mechanism, using a large receptive field to find the fine regions in features " + }, + { + "bbox": [ + 46, + 430, + 289, + 660 + ], + "type": "inline_equation", + "content": "V_{t_k}^i" + }, + { + "bbox": [ + 46, + 430, + 289, + 660 + ], + "type": "text", + "content": " that are identical to the motion of the error region in " + }, + { + "bbox": [ + 46, + 430, + 289, + 660 + ], + "type": "inline_equation", + "content": "\\bar{F}^i" + }, + { + "bbox": [ + 46, + 430, + 289, + 660 + ], + "type": "text", + "content": ". It outputs self-attention weights " + }, + { + "bbox": [ + 46, + 430, + 289, + 660 + ], + "type": "inline_equation", + "content": "A^i" + }, + { + "bbox": [ + 46, + 430, + 289, + 660 + ], + "type": "text", + "content": ", multiplied with the initial flow " + }, + { + "bbox": [ + 46, + 430, + 289, + 660 + ], + "type": "inline_equation", + "content": "\\bar{F}^i" + }, + { + "bbox": [ + 46, + 430, + 289, + 660 + ], + "type": "text", + "content": ". With the corrected flow " + }, + { + "bbox": [ + 46, + 430, + 289, + 660 + ], + "type": "inline_equation", + "content": "\\Delta F^i" + }, + { + "bbox": [ + 46, + 430, + 289, + 660 + ], + "type": "text", + "content": " and self-attention weights " + }, + { + "bbox": [ + 46, + 430, + 289, + 660 + ], + "type": "inline_equation", + "content": "A^i" + }, + { + "bbox": [ + 46, + 430, + 289, + 660 + ], + "type": "text", + "content": ", we can generate the fine flow " + }, + { + "bbox": [ + 46, + 430, + 289, + 660 + ], + "type": "inline_equation", + "content": "\\tilde{F}^i" + }, + { + "bbox": [ + 46, + 430, + 289, + 660 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 79, + 666, + 287, + 682 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 79, + 666, + 287, + 682 + ], + "spans": [ + { + "bbox": [ + 79, + 666, + 287, + 682 + ], + "type": "interline_equation", + "content": "\\tilde {F} ^ {i} = \\alpha \\mathcal {W} (\\bar {F} ^ {i}, \\Delta F ^ {i}) + (1 - \\alpha) \\left(A ^ {i} \\otimes \\bar {F} ^ {i}\\right), \\tag {4}", + "image_path": "5401cb82ee69df29dd4f301a66896e8a4be001ba68322f096d798a656f5558f4.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 689, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 288, + 715 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 689, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{W}(\\cdot, \\cdot)" + }, + { + "bbox": [ + 47, + 689, + 288, + 715 + ], + "type": "text", + "content": " means the warping operation, " + }, + { + "bbox": [ + 47, + 689, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\otimes" + }, + { + "bbox": [ + 47, + 689, + 288, + 715 + ], + "type": "text", + "content": " donates multiplication and " + }, + { + "bbox": [ + 47, + 689, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\alpha \\in [0,1]" + }, + { + "bbox": [ + 47, + 689, + 288, + 715 + ], + "type": "text", + "content": " is the weight coefficient. We iden" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 250, + 546, + 287 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 250, + 546, + 287 + ], + "spans": [ + { + "bbox": [ + 305, + 250, + 546, + 287 + ], + "type": "text", + "content": "tify error-prone object edge areas based on the corrected confidence map " + }, + { + "bbox": [ + 305, + 250, + 546, + 287 + ], + "type": "inline_equation", + "content": "W^{i}" + }, + { + "bbox": [ + 305, + 250, + 546, + 287 + ], + "type": "text", + "content": ", as weights to fuse the initial flow " + }, + { + "bbox": [ + 305, + 250, + 546, + 287 + ], + "type": "inline_equation", + "content": "\\bar{F}^{i}" + }, + { + "bbox": [ + 305, + 250, + 546, + 287 + ], + "type": "text", + "content": " and the fine flow " + }, + { + "bbox": [ + 305, + 250, + 546, + 287 + ], + "type": "inline_equation", + "content": "\\tilde{F}^{i}" + }, + { + "bbox": [ + 305, + 250, + 546, + 287 + ], + "type": "text", + "content": ", obtaining the final corrected flow " + }, + { + "bbox": [ + 305, + 250, + 546, + 287 + ], + "type": "inline_equation", + "content": "F_{\\uparrow}^{i}" + }, + { + "bbox": [ + 305, + 250, + 546, + 287 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 354, + 293, + 546, + 309 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 354, + 293, + 546, + 309 + ], + "spans": [ + { + "bbox": [ + 354, + 293, + 546, + 309 + ], + "type": "interline_equation", + "content": "F _ {\\uparrow} ^ {i} = W ^ {i} \\odot \\bar {F} ^ {i} + (1 - W ^ {i}) \\odot \\tilde {F} ^ {i}, \\tag {5}", + "image_path": "9b4b67c8d3712a8c15b389a2f70bb537e211d8fa15072c8f8a5e217ed66aa3d7.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 315, + 492, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 315, + 492, + 327 + ], + "spans": [ + { + "bbox": [ + 306, + 315, + 492, + 327 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 306, + 315, + 492, + 327 + ], + "type": "inline_equation", + "content": "\\odot" + }, + { + "bbox": [ + 306, + 315, + 492, + 327 + ], + "type": "text", + "content": " donates the element-wise multiplier." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 342, + 400, + 353 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 342, + 400, + 353 + ], + "spans": [ + { + "bbox": [ + 306, + 342, + 400, + 353 + ], + "type": "text", + "content": "3.2.3 Loss Function" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 360, + 546, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 360, + 546, + 456 + ], + "spans": [ + { + "bbox": [ + 304, + 360, + 546, + 456 + ], + "type": "text", + "content": "For both meshflow and optical flow estimation, we use L1 loss for supervised regression during training. Our EEMFlow, used for estimating meshflow, directly outputs results at the same resolution as the meshflow GT, allowing for direct loss calculation. Similarly, the results outputted by our EEMFlow+ for optical flow estimation are at the same resolution as the network input, enabling direct calculation with the optical flow GT." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 465, + 388, + 479 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 465, + 388, + 479 + ], + "spans": [ + { + "bbox": [ + 306, + 465, + 388, + 479 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 485, + 440, + 498 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 485, + 440, + 498 + ], + "spans": [ + { + "bbox": [ + 306, + 485, + 440, + 498 + ], + "type": "text", + "content": "4.1. Implementation Details" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 503, + 376, + 514 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 503, + 376, + 514 + ], + "spans": [ + { + "bbox": [ + 306, + 503, + 376, + 514 + ], + "type": "text", + "content": "4.1.1 Datasets" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 521, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 521, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 521, + 547, + 714 + ], + "type": "text", + "content": "Our HREM dataset includes 100 indoor and outdoor scenes, with a resolution of " + }, + { + "bbox": [ + 304, + 521, + 547, + 714 + ], + "type": "inline_equation", + "content": "1280 \\times 720" + }, + { + "bbox": [ + 304, + 521, + 547, + 714 + ], + "type": "text", + "content": ". We randomly select 70 scenes for training and reserve the remaining 30 for testing. The training set comprises 20,000 samples, while the test set contains 8,000 samples. Additionally, we further divide the test set by scene type (outdoor vs. indoor) and camera motion speed during rendering, resulting in four sub-sequences (outdoor Slow, outdoor fast, indoor slow, indoor fast), with mean motion magnitudes ranging from, " + }, + { + "bbox": [ + 304, + 521, + 547, + 714 + ], + "type": "inline_equation", + "content": "0 - 30" + }, + { + "bbox": [ + 304, + 521, + 547, + 714 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 521, + 547, + 714 + ], + "type": "inline_equation", + "content": "30 - 100" + }, + { + "bbox": [ + 304, + 521, + 547, + 714 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 521, + 547, + 714 + ], + "type": "inline_equation", + "content": "0 - 20" + }, + { + "bbox": [ + 304, + 521, + 547, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 521, + 547, + 714 + ], + "type": "inline_equation", + "content": "20 - 100" + }, + { + "bbox": [ + 304, + 521, + 547, + 714 + ], + "type": "text", + "content": " pixels, respectively. Moreover, similar to [52], we employ two data input modes: " + }, + { + "bbox": [ + 304, + 521, + 547, + 714 + ], + "type": "inline_equation", + "content": "dt = 1" + }, + { + "bbox": [ + 304, + 521, + 547, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 521, + 547, + 714 + ], + "type": "inline_equation", + "content": "dt = 4" + }, + { + "bbox": [ + 304, + 521, + 547, + 714 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 304, + 521, + 547, + 714 + ], + "type": "inline_equation", + "content": "dt = 1" + }, + { + "bbox": [ + 304, + 521, + 547, + 714 + ], + "type": "text", + "content": " uses the event sequence between two consecutive frames of RGB images as input, with a meshflow generation frequency of " + }, + { + "bbox": [ + 304, + 521, + 547, + 714 + ], + "type": "inline_equation", + "content": "60~\\mathrm{Hz}" + }, + { + "bbox": [ + 304, + 521, + 547, + 714 + ], + "type": "text", + "content": " while " + }, + { + "bbox": [ + 304, + 521, + 547, + 714 + ], + "type": "inline_equation", + "content": "dt = 4" + }, + { + "bbox": [ + 304, + 521, + 547, + 714 + ], + "type": "text", + "content": " uses the event sequence between four consecutive frames of RGB images as input, with a meshflow" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "19202" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 99, + 118, + 496, + 312 + ], + "blocks": [ + { + "bbox": [ + 46, + 70, + 547, + 109 + ], + "lines": [ + { + "bbox": [ + 46, + 70, + 547, + 109 + ], + "spans": [ + { + "bbox": [ + 46, + 70, + 547, + 109 + ], + "type": "text", + "content": "Table 2. Quantitative comparison of our EEMFlow with other advanced flow networks on our HREM dataset. The evaluation metric used is End-Point Error (EPE). \"Parameters\" and \"Time\" respectively indicate the network parameter count and inference time. " + }, + { + "bbox": [ + 46, + 70, + 547, + 109 + ], + "type": "inline_equation", + "content": "\\Delta P" + }, + { + "bbox": [ + 46, + 70, + 547, + 109 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 70, + 547, + 109 + ], + "type": "inline_equation", + "content": "\\Delta T" + }, + { + "bbox": [ + 46, + 70, + 547, + 109 + ], + "type": "text", + "content": " represent the change in parameter count and inference time relative to ERAFT [42] for other networks. Smaller values are desirable for all metrics. We highlight the best results in red and the second-best results in blue." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 99, + 118, + 496, + 312 + ], + "lines": [ + { + "bbox": [ + 99, + 118, + 496, + 312 + ], + "spans": [ + { + "bbox": [ + 99, + 118, + 496, + 312 + ], + "type": "table", + "html": "
Method dt = 1Parameters (M)Time (ms)OutdoorIndoorAvg
SlowFastSlowFast
EVFlownet [52]38.2463.5516.162.9311.658.57
PWCNet [40]3.36423.9114.492.8611.898.29
ERAFT [11]5.27934.1513.322.9110.347.68
SKFlow [41]6.281453.7611.787.248.817.24
GMA [17]5.891082.1812.072.029.346.40
KPAFlow [32]6.001842.0312.251.959.026.31
FlowFormer [15]9.872812.0611.711.888.666.08
EEMFlow(Ours)1.2472.429.092.008.465.50
Method dt = 4ΔP (M)ΔT (ms)OutdoorIndoorAvg
SlowFastSlowFast
EVFlownet [52]+624%+51%18.2549.3216.1647.1932.73
PWCNet [40]-36%-55%16.4046.1714.4940.9029.49
ERAFT [11]0%0%15.2140.8313.3239.6127.24
SKFlow [41]+19%+56%14.9339.2411.7139.2226.28
GMA [17]+11%+16%14.1338.8912.0737.6825.69
KPAFlow [32]+14%+99%14.0438.0312.2537.2025.38
FlowFormer [15]+88%+202%13.8938.5510.7738.5325.44
EEMFlow(Ours)-76%-92%13.9737.3312.0934.3924.45
", + "image_path": "06ac50e76b3d9ff6415a5f67a9dd86a0cb87cd499b50c614be298af420265e46.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 51, + 321, + 126, + 408 + ], + "blocks": [ + { + "bbox": [ + 51, + 321, + 126, + 408 + ], + "lines": [ + { + "bbox": [ + 51, + 321, + 126, + 408 + ], + "spans": [ + { + "bbox": [ + 51, + 321, + 126, + 408 + ], + "type": "image", + "image_path": "9fc0c88cacfe6e458c8cfb59ca03cda2904196b652f4d696e78cab8325f790d2.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 61, + 408, + 114, + 417 + ], + "lines": [ + { + "bbox": [ + 61, + 408, + 114, + 417 + ], + "spans": [ + { + "bbox": [ + 61, + 408, + 114, + 417 + ], + "type": "text", + "content": "Image Overlaid" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 426, + 547, + 447 + ], + "lines": [ + { + "bbox": [ + 46, + 426, + 547, + 447 + ], + "spans": [ + { + "bbox": [ + 46, + 426, + 547, + 447 + ], + "type": "text", + "content": "Figure 6. Subjective results of image registration using meshflow estimated by other methods and ours. We estimate meshflow from the event sequences and then warp image " + }, + { + "bbox": [ + 46, + 426, + 547, + 447 + ], + "type": "inline_equation", + "content": "I_{t_1}" + }, + { + "bbox": [ + 46, + 426, + 547, + 447 + ], + "type": "text", + "content": " onto image " + }, + { + "bbox": [ + 46, + 426, + 547, + 447 + ], + "type": "inline_equation", + "content": "I_{t_2}" + }, + { + "bbox": [ + 46, + 426, + 547, + 447 + ], + "type": "text", + "content": " for fusion, showcasing the fused result. The fewer blue or red ghosting artifacts indicate better alignment performance." + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 135, + 322, + 209, + 365 + ], + "blocks": [ + { + "bbox": [ + 135, + 322, + 209, + 365 + ], + "lines": [ + { + "bbox": [ + 135, + 322, + 209, + 365 + ], + "spans": [ + { + "bbox": [ + 135, + 322, + 209, + 365 + ], + "type": "image", + "image_path": "f280d6c275454193d148eb0324e69dc68af1773c37875506ff7ef3732c8ceec1.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 135, + 365, + 209, + 408 + ], + "blocks": [ + { + "bbox": [ + 135, + 365, + 209, + 408 + ], + "lines": [ + { + "bbox": [ + 135, + 365, + 209, + 408 + ], + "spans": [ + { + "bbox": [ + 135, + 365, + 209, + 408 + ], + "type": "image", + "image_path": "7cf3e046c10ea117aa8290abb1e1435be49017bb72252e4bbd7743e14a1ec991.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 153, + 408, + 192, + 416 + ], + "lines": [ + { + "bbox": [ + 153, + 408, + 192, + 416 + ], + "spans": [ + { + "bbox": [ + 153, + 408, + 192, + 416 + ], + "type": "text", + "content": "EVFlownet" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 219, + 322, + 293, + 365 + ], + "blocks": [ + { + "bbox": [ + 219, + 322, + 293, + 365 + ], + "lines": [ + { + "bbox": [ + 219, + 322, + 293, + 365 + ], + "spans": [ + { + "bbox": [ + 219, + 322, + 293, + 365 + ], + "type": "image", + "image_path": "2e850dc37c6dc0639f6ab80da4b2653feccfda8fc021fcbba7f6e2f0aebd8f4d.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 219, + 365, + 293, + 408 + ], + "blocks": [ + { + "bbox": [ + 219, + 365, + 293, + 408 + ], + "lines": [ + { + "bbox": [ + 219, + 365, + 293, + 408 + ], + "spans": [ + { + "bbox": [ + 219, + 365, + 293, + 408 + ], + "type": "image", + "image_path": "ccd3554bfda5da10a8e9166277aa1d8e8ada1cb211a4379a13572f3e8a16f6de.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 242, + 408, + 269, + 416 + ], + "lines": [ + { + "bbox": [ + 242, + 408, + 269, + 416 + ], + "spans": [ + { + "bbox": [ + 242, + 408, + 269, + 416 + ], + "type": "text", + "content": "ERRAFT" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 302, + 322, + 376, + 365 + ], + "blocks": [ + { + "bbox": [ + 302, + 322, + 376, + 365 + ], + "lines": [ + { + "bbox": [ + 302, + 322, + 376, + 365 + ], + "spans": [ + { + "bbox": [ + 302, + 322, + 376, + 365 + ], + "type": "image", + "image_path": "ef632fbee6fec007d58cdc19bfefd4e63501bf5f9654b0328cad5b6bd2c07f8c.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 302, + 365, + 376, + 407 + ], + "blocks": [ + { + "bbox": [ + 302, + 365, + 376, + 407 + ], + "lines": [ + { + "bbox": [ + 302, + 365, + 376, + 407 + ], + "spans": [ + { + "bbox": [ + 302, + 365, + 376, + 407 + ], + "type": "image", + "image_path": "d83a0811619fbfaf618e1e8b7653469de9a27f755c92ae3ed7fc963efaf4ef88.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 322, + 408, + 356, + 416 + ], + "lines": [ + { + "bbox": [ + 322, + 408, + 356, + 416 + ], + "spans": [ + { + "bbox": [ + 322, + 408, + 356, + 416 + ], + "type": "text", + "content": "KPAFlow" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 385, + 322, + 460, + 365 + ], + "blocks": [ + { + "bbox": [ + 385, + 322, + 460, + 365 + ], + "lines": [ + { + "bbox": [ + 385, + 322, + 460, + 365 + ], + "spans": [ + { + "bbox": [ + 385, + 322, + 460, + 365 + ], + "type": "image", + "image_path": "fc59d6b08c2ff9b1b3d3365ca810a5655c76ec127f9c7f4a0380943935d6c3d5.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 385, + 365, + 460, + 408 + ], + "blocks": [ + { + "bbox": [ + 385, + 365, + 460, + 408 + ], + "lines": [ + { + "bbox": [ + 385, + 365, + 460, + 408 + ], + "spans": [ + { + "bbox": [ + 385, + 365, + 460, + 408 + ], + "type": "image", + "image_path": "694eb12bb2cc5e279cfeda3c5636c85c56b9e28ed869030cf31f04b4d1ef9846.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 413, + 408, + 432, + 416 + ], + "lines": [ + { + "bbox": [ + 413, + 408, + 432, + 416 + ], + "spans": [ + { + "bbox": [ + 413, + 408, + 432, + 416 + ], + "type": "text", + "content": "Ours" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 469, + 322, + 543, + 365 + ], + "blocks": [ + { + "bbox": [ + 469, + 322, + 543, + 365 + ], + "lines": [ + { + "bbox": [ + 469, + 322, + 543, + 365 + ], + "spans": [ + { + "bbox": [ + 469, + 322, + 543, + 365 + ], + "type": "image", + "image_path": "bf17103592adf34b08c77e58c71bc7df60d2c48ff7ec6cc8e06547915d1bf1c7.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 469, + 365, + 543, + 407 + ], + "blocks": [ + { + "bbox": [ + 469, + 365, + 543, + 407 + ], + "lines": [ + { + "bbox": [ + 469, + 365, + 543, + 407 + ], + "spans": [ + { + "bbox": [ + 469, + 365, + 543, + 407 + ], + "type": "image", + "image_path": "1d33704e70b973568640502fbbc9ed0de00a48367e30881d1869afb8179708ac.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 481, + 407, + 531, + 417 + ], + "lines": [ + { + "bbox": [ + 481, + 407, + 531, + 417 + ], + "spans": [ + { + "bbox": [ + 481, + 407, + 531, + 417 + ], + "type": "text", + "content": "Meshflow(GT)" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "bbox": [ + 46, + 467, + 287, + 527 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 467, + 287, + 527 + ], + "spans": [ + { + "bbox": [ + 46, + 467, + 287, + 527 + ], + "type": "text", + "content": "generation frequency of " + }, + { + "bbox": [ + 46, + 467, + 287, + 527 + ], + "type": "inline_equation", + "content": "15\\mathrm{Hz}" + }, + { + "bbox": [ + 46, + 467, + 287, + 527 + ], + "type": "text", + "content": ". DSEC [10] is a commonly used dataset for event-based optical flow estimation, which consists of real-world data captured by real event cameras mounted on cars. We also conduct experiments on DSEC to compare the performance for optical flow estimation." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 47, + 541, + 148, + 553 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 541, + 148, + 553 + ], + "spans": [ + { + "bbox": [ + 47, + 541, + 148, + 553 + ], + "type": "text", + "content": "4.1.2 Training details" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 46, + 560, + 287, + 643 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 560, + 287, + 643 + ], + "spans": [ + { + "bbox": [ + 46, + 560, + 287, + 643 + ], + "type": "text", + "content": "We conduct experiments using the PyTorch framework on two NVIDIA 2080Ti GPUs. We train all networks with the same parameters on our HRDM dataset. We employ the AdamW optimizer and OneCycle policy with a learning rate of " + }, + { + "bbox": [ + 46, + 560, + 287, + 643 + ], + "type": "inline_equation", + "content": "5 \\times 10^{-4}" + }, + { + "bbox": [ + 46, + 560, + 287, + 643 + ], + "type": "text", + "content": ", weight decay of " + }, + { + "bbox": [ + 46, + 560, + 287, + 643 + ], + "type": "inline_equation", + "content": "5 \\times 10^{-5}" + }, + { + "bbox": [ + 46, + 560, + 287, + 643 + ], + "type": "text", + "content": ", and other default parameters set to " + }, + { + "bbox": [ + 46, + 560, + 287, + 643 + ], + "type": "inline_equation", + "content": "\\beta_{1} = 0.9" + }, + { + "bbox": [ + 46, + 560, + 287, + 643 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 560, + 287, + 643 + ], + "type": "inline_equation", + "content": "\\beta_{2} = 0.99" + }, + { + "bbox": [ + 46, + 560, + 287, + 643 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 560, + 287, + 643 + ], + "type": "inline_equation", + "content": "\\epsilon = 1 \\times 10^{-4}" + }, + { + "bbox": [ + 46, + 560, + 287, + 643 + ], + "type": "text", + "content": ". We train all networks up to " + }, + { + "bbox": [ + 46, + 560, + 287, + 643 + ], + "type": "inline_equation", + "content": "100k" + }, + { + "bbox": [ + 46, + 560, + 287, + 643 + ], + "type": "text", + "content": " iterations to reach convergence." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 47, + 658, + 161, + 670 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 658, + 161, + 670 + ], + "spans": [ + { + "bbox": [ + 47, + 658, + 161, + 670 + ], + "type": "text", + "content": "4.1.3 Evaluation metrics" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "text", + "content": "Following EV-FlowNet [52], we use the average Endpoint Error (EPE) as the metric. When evaluating meshflow, we upsample both the prediction and the ground truth to the" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 304, + 467, + 546, + 527 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 467, + 546, + 527 + ], + "spans": [ + { + "bbox": [ + 304, + 467, + 546, + 527 + ], + "type": "text", + "content": "input resolution to calculate metrics. In addition, the DSEC dataset uses " + }, + { + "bbox": [ + 304, + 467, + 546, + 527 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 467, + 546, + 527 + ], + "type": "text", + "content": "PE to measure the percentage of flow errors higher than " + }, + { + "bbox": [ + 304, + 467, + 546, + 527 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 467, + 546, + 527 + ], + "type": "text", + "content": " pixels in magnitude (e.g., 3PE, 2PE, 1PE) for flow outliers analysis, and employs the Angular Error (AE) to assess the directional accuracy." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 306, + 533, + 492, + 545 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 533, + 492, + 545 + ], + "spans": [ + { + "bbox": [ + 306, + 533, + 492, + 545 + ], + "type": "text", + "content": "4.2. Comparison with State-of-the-Arts" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 306, + 552, + 533, + 563 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 552, + 533, + 563 + ], + "spans": [ + { + "bbox": [ + 306, + 552, + 533, + 563 + ], + "type": "text", + "content": "4.2.1 Results for Event-based Meshflow Estimation" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 304, + 570, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 547, + 713 + ], + "type": "text", + "content": "In Table 2, We train and test our EEMFlow on our HREM dataset, compared with some advanced networks, e.g., EVFlownet [54], ERAFT [11], PwCNet [40], SKFlow [41], GMA [17], KPAFlow [32], and FlowFormer [15]. Since these networks are all structured to output prediction with the same resolution of input, we upsample the ground truth " + }, + { + "bbox": [ + 304, + 570, + 547, + 713 + ], + "type": "inline_equation", + "content": "MF_{GT}" + }, + { + "bbox": [ + 304, + 570, + 547, + 713 + ], + "type": "text", + "content": " to supervise these networks during training. However, EEMFlow can directly output prediction aligned " + }, + { + "bbox": [ + 304, + 570, + 547, + 713 + ], + "type": "inline_equation", + "content": "MF_{GT}" + }, + { + "bbox": [ + 304, + 570, + 547, + 713 + ], + "type": "text", + "content": " in resolution. But for fair comparison, we upsample the predictions of all networks to a uniform resolution during evaluation. We train and test all networks using two input modes (" + }, + { + "bbox": [ + 304, + 570, + 547, + 713 + ], + "type": "inline_equation", + "content": "dt = 1" + }, + { + "bbox": [ + 304, + 570, + 547, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 570, + 547, + 713 + ], + "type": "inline_equation", + "content": "dt = 4" + }, + { + "bbox": [ + 304, + 570, + 547, + 713 + ], + "type": "text", + "content": "), and" + } + ] + } + ], + "index": 28 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19203" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 70, + 127, + 114 + ], + "blocks": [ + { + "bbox": [ + 50, + 70, + 127, + 114 + ], + "lines": [ + { + "bbox": [ + 50, + 70, + 127, + 114 + ], + "spans": [ + { + "bbox": [ + 50, + 70, + 127, + 114 + ], + "type": "image", + "image_path": "5a8c0f47f9e71a046ccb10fae5a783c7464652c6e285efe1cbd1b6288bf717ad.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 51, + 115, + 126, + 158 + ], + "blocks": [ + { + "bbox": [ + 51, + 115, + 126, + 158 + ], + "lines": [ + { + "bbox": [ + 51, + 115, + 126, + 158 + ], + "spans": [ + { + "bbox": [ + 51, + 115, + 126, + 158 + ], + "type": "image", + "image_path": "303579b082eac16c49178a78c7ae1ef0e09d1effdd845ba3c55e62c66aad3106.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 134, + 72, + 209, + 114 + ], + "blocks": [ + { + "bbox": [ + 134, + 72, + 209, + 114 + ], + "lines": [ + { + "bbox": [ + 134, + 72, + 209, + 114 + ], + "spans": [ + { + "bbox": [ + 134, + 72, + 209, + 114 + ], + "type": "image", + "image_path": "0dee9e03d491fbbe1585d914ddf12cd982e46a0f808e251c6153d7fad7ef55c3.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 134, + 114, + 210, + 158 + ], + "blocks": [ + { + "bbox": [ + 134, + 114, + 210, + 158 + ], + "lines": [ + { + "bbox": [ + 134, + 114, + 210, + 158 + ], + "spans": [ + { + "bbox": [ + 134, + 114, + 210, + 158 + ], + "type": "image", + "image_path": "3b898acb02f98b32b54be43c95aa0f3af255f18269dbd6c15aaf89fa5f63765c.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 217, + 72, + 293, + 114 + ], + "blocks": [ + { + "bbox": [ + 217, + 72, + 293, + 114 + ], + "lines": [ + { + "bbox": [ + 217, + 72, + 293, + 114 + ], + "spans": [ + { + "bbox": [ + 217, + 72, + 293, + 114 + ], + "type": "image", + "image_path": "5f4a5608ec7743aea9011075eeccbca48b75df2ebb526d7558ae79bbd5961930.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 217, + 114, + 293, + 158 + ], + "blocks": [ + { + "bbox": [ + 217, + 114, + 293, + 158 + ], + "lines": [ + { + "bbox": [ + 217, + 114, + 293, + 158 + ], + "spans": [ + { + "bbox": [ + 217, + 114, + 293, + 158 + ], + "type": "image", + "image_path": "a09caf932e6a7f206ff8751da8f0b62c09fed60ff347aff61742d6e05893551e.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 301, + 72, + 376, + 114 + ], + "blocks": [ + { + "bbox": [ + 301, + 72, + 376, + 114 + ], + "lines": [ + { + "bbox": [ + 301, + 72, + 376, + 114 + ], + "spans": [ + { + "bbox": [ + 301, + 72, + 376, + 114 + ], + "type": "image", + "image_path": "4f1449191a3f8e0360017f82e0d8d4c1065dd08d1c3e3f460d01a264dbf00e34.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 301, + 114, + 376, + 158 + ], + "blocks": [ + { + "bbox": [ + 301, + 114, + 376, + 158 + ], + "lines": [ + { + "bbox": [ + 301, + 114, + 376, + 158 + ], + "spans": [ + { + "bbox": [ + 301, + 114, + 376, + 158 + ], + "type": "image", + "image_path": "2a4b3087f83298301bdbf5e9ab908279fecac0daac1a3ef0fc67f89b3d761898.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 384, + 72, + 460, + 114 + ], + "blocks": [ + { + "bbox": [ + 384, + 72, + 460, + 114 + ], + "lines": [ + { + "bbox": [ + 384, + 72, + 460, + 114 + ], + "spans": [ + { + "bbox": [ + 384, + 72, + 460, + 114 + ], + "type": "image", + "image_path": "dde18daf6e3dc43071308ecae8c8601e4f3f29157e46d584a386ad2d454645c8.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 384, + 114, + 460, + 158 + ], + "blocks": [ + { + "bbox": [ + 384, + 114, + 460, + 158 + ], + "lines": [ + { + "bbox": [ + 384, + 114, + 460, + 158 + ], + "spans": [ + { + "bbox": [ + 384, + 114, + 460, + 158 + ], + "type": "image", + "image_path": "8668f53d774ed290cc97deefb29b2f23456c761e37c7ac18ca38980638a5d5ad.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 468, + 72, + 543, + 114 + ], + "blocks": [ + { + "bbox": [ + 468, + 72, + 543, + 114 + ], + "lines": [ + { + "bbox": [ + 468, + 72, + 543, + 114 + ], + "spans": [ + { + "bbox": [ + 468, + 72, + 543, + 114 + ], + "type": "image", + "image_path": "fa5d9085616fdc7563c28349ab0c7c4a681ac8c04f7c1a5c50b5446e5d0f03f3.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 468, + 114, + 543, + 158 + ], + "blocks": [ + { + "bbox": [ + 468, + 114, + 543, + 158 + ], + "lines": [ + { + "bbox": [ + 468, + 114, + 543, + 158 + ], + "spans": [ + { + "bbox": [ + 468, + 114, + 543, + 158 + ], + "type": "image", + "image_path": "8daba0d062a0b561ed461b2a3d2e0666e25095c66432f744e2a7a28cc5652a4b.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 51, + 163, + 126, + 207 + ], + "blocks": [ + { + "bbox": [ + 51, + 163, + 126, + 207 + ], + "lines": [ + { + "bbox": [ + 51, + 163, + 126, + 207 + ], + "spans": [ + { + "bbox": [ + 51, + 163, + 126, + 207 + ], + "type": "image", + "image_path": "81d3b1f686b36f9233539ce65f98a0d56a7c96140969bdac66195b05b1aa5664.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 51, + 207, + 126, + 251 + ], + "blocks": [ + { + "bbox": [ + 51, + 207, + 126, + 251 + ], + "lines": [ + { + "bbox": [ + 51, + 207, + 126, + 251 + ], + "spans": [ + { + "bbox": [ + 51, + 207, + 126, + 251 + ], + "type": "image", + "image_path": "4332fd43de910a4058b61d313e838e981065761f2cb54436398902268d8223f8.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 71, + 251, + 106, + 259 + ], + "lines": [ + { + "bbox": [ + 71, + 251, + 106, + 259 + ], + "spans": [ + { + "bbox": [ + 71, + 251, + 106, + 259 + ], + "type": "text", + "content": "Reference" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 134, + 163, + 209, + 207 + ], + "blocks": [ + { + "bbox": [ + 134, + 163, + 209, + 207 + ], + "lines": [ + { + "bbox": [ + 134, + 163, + 209, + 207 + ], + "spans": [ + { + "bbox": [ + 134, + 163, + 209, + 207 + ], + "type": "image", + "image_path": "c385d0c5272a8683530646dd4a3e64b79abf569033a161965c7318bddf189f70.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 134, + 207, + 209, + 251 + ], + "blocks": [ + { + "bbox": [ + 134, + 207, + 209, + 251 + ], + "lines": [ + { + "bbox": [ + 134, + 207, + 209, + 251 + ], + "spans": [ + { + "bbox": [ + 134, + 207, + 209, + 251 + ], + "type": "image", + "image_path": "393596e3a19cd222d3e50dcdd2549c1bc1bcbfa89588c8b3b3ffb17759850f28.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 153, + 251, + 192, + 259 + ], + "lines": [ + { + "bbox": [ + 153, + 251, + 192, + 259 + ], + "spans": [ + { + "bbox": [ + 153, + 251, + 192, + 259 + ], + "type": "text", + "content": "EVFlownet" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 217, + 163, + 293, + 207 + ], + "blocks": [ + { + "bbox": [ + 217, + 163, + 293, + 207 + ], + "lines": [ + { + "bbox": [ + 217, + 163, + 293, + 207 + ], + "spans": [ + { + "bbox": [ + 217, + 163, + 293, + 207 + ], + "type": "image", + "image_path": "d6a7cd9d93821cdb4e5c5798ea0def3ec4b6924268c107ecc513356bcab3f497.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 217, + 207, + 293, + 251 + ], + "blocks": [ + { + "bbox": [ + 217, + 207, + 293, + 251 + ], + "lines": [ + { + "bbox": [ + 217, + 207, + 293, + 251 + ], + "spans": [ + { + "bbox": [ + 217, + 207, + 293, + 251 + ], + "type": "image", + "image_path": "530f474e613766ac8ff207f80d502ae2e95b85022db741752ca88dc704989483.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 241, + 251, + 268, + 259 + ], + "lines": [ + { + "bbox": [ + 241, + 251, + 268, + 259 + ], + "spans": [ + { + "bbox": [ + 241, + 251, + 268, + 259 + ], + "type": "text", + "content": "ERRAFT" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 301, + 163, + 376, + 207 + ], + "blocks": [ + { + "bbox": [ + 301, + 163, + 376, + 207 + ], + "lines": [ + { + "bbox": [ + 301, + 163, + 376, + 207 + ], + "spans": [ + { + "bbox": [ + 301, + 163, + 376, + 207 + ], + "type": "image", + "image_path": "2574e57faea5804beb100cbca3076322da371bb0a2cb892e9c813da73e8c55d0.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 301, + 207, + 376, + 251 + ], + "blocks": [ + { + "bbox": [ + 301, + 207, + 376, + 251 + ], + "lines": [ + { + "bbox": [ + 301, + 207, + 376, + 251 + ], + "spans": [ + { + "bbox": [ + 301, + 207, + 376, + 251 + ], + "type": "image", + "image_path": "9d356c7b34313c8838f65cf45d3df7bb64d06ea0e0144c5496adb2856deeea6f.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 321, + 251, + 355, + 259 + ], + "lines": [ + { + "bbox": [ + 321, + 251, + 355, + 259 + ], + "spans": [ + { + "bbox": [ + 321, + 251, + 355, + 259 + ], + "type": "text", + "content": "KPAFlow" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 384, + 163, + 460, + 207 + ], + "blocks": [ + { + "bbox": [ + 384, + 163, + 460, + 207 + ], + "lines": [ + { + "bbox": [ + 384, + 163, + 460, + 207 + ], + "spans": [ + { + "bbox": [ + 384, + 163, + 460, + 207 + ], + "type": "image", + "image_path": "cf1de7783ed2ab4e7f427e7782088aaf6c7db3854ea12c4dcbd292719b5e07f0.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 384, + 207, + 460, + 251 + ], + "blocks": [ + { + "bbox": [ + 384, + 207, + 460, + 251 + ], + "lines": [ + { + "bbox": [ + 384, + 207, + 460, + 251 + ], + "spans": [ + { + "bbox": [ + 384, + 207, + 460, + 251 + ], + "type": "image", + "image_path": "d54f1c76ddaee17679afaefa9ace9198fcc4f1006d1054d6a2cf47e516da16e8.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 413, + 251, + 430, + 259 + ], + "lines": [ + { + "bbox": [ + 413, + 251, + 430, + 259 + ], + "spans": [ + { + "bbox": [ + 413, + 251, + 430, + 259 + ], + "type": "text", + "content": "Ours" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_caption" + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 468, + 163, + 542, + 208 + ], + "blocks": [ + { + "bbox": [ + 468, + 163, + 542, + 208 + ], + "lines": [ + { + "bbox": [ + 468, + 163, + 542, + 208 + ], + "spans": [ + { + "bbox": [ + 468, + 163, + 542, + 208 + ], + "type": "image", + "image_path": "a6768023085dc78983a637802e01f1a4abeeced6c0e06c5ebd2869d278113461.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_body" + } + ], + "index": 27 + }, + { + "type": "image", + "bbox": [ + 468, + 208, + 542, + 251 + ], + "blocks": [ + { + "bbox": [ + 468, + 208, + 542, + 251 + ], + "lines": [ + { + "bbox": [ + 468, + 208, + 542, + 251 + ], + "spans": [ + { + "bbox": [ + 468, + 208, + 542, + 251 + ], + "type": "image", + "image_path": "ee6b5dde4d2b965b0cf76a1e2a5a03964507a8b64c20a7e627b497679540d6bc.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 480, + 251, + 530, + 260 + ], + "lines": [ + { + "bbox": [ + 480, + 251, + 530, + 260 + ], + "spans": [ + { + "bbox": [ + 480, + 251, + 530, + 260 + ], + "type": "text", + "content": "Meshflow(GT)" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_caption" + } + ], + "index": 28 + }, + { + "type": "table", + "bbox": [ + 52, + 335, + 282, + 434 + ], + "blocks": [ + { + "bbox": [ + 47, + 307, + 287, + 326 + ], + "lines": [ + { + "bbox": [ + 47, + 307, + 287, + 326 + ], + "spans": [ + { + "bbox": [ + 47, + 307, + 287, + 326 + ], + "type": "text", + "content": "Table 3. Results on DSEC dataset. All networks are trained and tested on the DSEC dataset for optical flow estimation." + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 52, + 335, + 282, + 434 + ], + "lines": [ + { + "bbox": [ + 52, + 335, + 282, + 434 + ], + "spans": [ + { + "bbox": [ + 52, + 335, + 282, + 434 + ], + "type": "table", + "html": "
MethodsFPS↑1PE↓2PE↓3PE↓EPE↓AE↓
MutilCM [39]-76.648.530.93.4714.0
EV-Flownet [52]22.355.429.818.62.328.12
OF-EV-SNN [4]-53.720.210.31.716.34
EVA-Flow [50]-15.9-3.200.883.31
ERRAFT [11]11.412.74.742.680.792.85
ADMFlow [33]9.8812.54.672.650.782.84
EFlowformer [21]-11.24.102.450.762.68
TMA [25]7.5510.93.972.300.742.68
EEMFlow+(Ours)39.211.43.932.150.752.67
", + "image_path": "5cab1699d76658dc9fd965632df748afd3a292fa2379222b96c4520f577084e0.jpg" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "table_body" + } + ], + "index": 32 + }, + { + "bbox": [ + 46, + 460, + 287, + 507 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 460, + 287, + 507 + ], + "spans": [ + { + "bbox": [ + 46, + 460, + 287, + 507 + ], + "type": "text", + "content": "present their performance metrics in four test sub-sequences (outdoor Slow, outdoor fast, indoor slow, indoor fast). In addition, we average the metric scores across all four test sub-sequences as shown in the \"Avg\" column." + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 46, + 510, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 510, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 510, + 287, + 713 + ], + "type": "text", + "content": "Table 2 shows that our EEMFlow achieves the lowest EPE score for both input modes " + }, + { + "bbox": [ + 46, + 510, + 287, + 713 + ], + "type": "inline_equation", + "content": "dt = 1" + }, + { + "bbox": [ + 46, + 510, + 287, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 510, + 287, + 713 + ], + "type": "inline_equation", + "content": "dt = 4" + }, + { + "bbox": [ + 46, + 510, + 287, + 713 + ], + "type": "text", + "content": " in the \"outdoor_fast\" and \"indoor_fast\" test sub-sequences, demonstrating its superior performance on high-speed and large-movement sequences. EEMFlow also exhibits great potential to outperform other flow networks on the \"outdoor Slow\" and \"indoor Slow\" test sub-sequences. Notably, EEMFlow achieves the lowest EPE scores in the \"Avg\" column for both input modes " + }, + { + "bbox": [ + 46, + 510, + 287, + 713 + ], + "type": "inline_equation", + "content": "dt = 1" + }, + { + "bbox": [ + 46, + 510, + 287, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 510, + 287, + 713 + ], + "type": "inline_equation", + "content": "dt = 4" + }, + { + "bbox": [ + 46, + 510, + 287, + 713 + ], + "type": "text", + "content": ". Besides, we also achieve the least number of parameters and the fastest inference speed. Compared to ERAFT [11], our EEMFlow reduces the parameter count by " + }, + { + "bbox": [ + 46, + 510, + 287, + 713 + ], + "type": "inline_equation", + "content": "76\\%" + }, + { + "bbox": [ + 46, + 510, + 287, + 713 + ], + "type": "text", + "content": " (from 5.27M to 1.24M), reduces the inference time by " + }, + { + "bbox": [ + 46, + 510, + 287, + 713 + ], + "type": "inline_equation", + "content": "92\\%" + }, + { + "bbox": [ + 46, + 510, + 287, + 713 + ], + "type": "text", + "content": " (from 93ms to 7ms), and improves average EPE in " + }, + { + "bbox": [ + 46, + 510, + 287, + 713 + ], + "type": "inline_equation", + "content": "dt = 4" + }, + { + "bbox": [ + 46, + 510, + 287, + 713 + ], + "type": "text", + "content": " by " + }, + { + "bbox": [ + 46, + 510, + 287, + 713 + ], + "type": "inline_equation", + "content": "8\\%" + }, + { + "bbox": [ + 46, + 510, + 287, + 713 + ], + "type": "text", + "content": " (from 27.24 to 25.18). EEMFlow achieves comparable EPE performance to FlowFormer [15] but exhibits a " + }, + { + "bbox": [ + 46, + 510, + 287, + 713 + ], + "type": "inline_equation", + "content": "38.7\\times" + }, + { + "bbox": [ + 46, + 510, + 287, + 713 + ], + "type": "text", + "content": " increase in inference speed." + } + ] + } + ], + "index": 34 + }, + { + "type": "image", + "bbox": [ + 308, + 307, + 544, + 343 + ], + "blocks": [ + { + "bbox": [ + 46, + 273, + 545, + 293 + ], + "lines": [ + { + "bbox": [ + 46, + 273, + 545, + 293 + ], + "spans": [ + { + "bbox": [ + 46, + 273, + 545, + 293 + ], + "type": "text", + "content": "Figure 7. Qualitative comparison of our proposed EEMFlow with other advanced flow networks on our HREM dataset. The subjective images of events and dense optical flow on the left side serve as references. The areas enclosed by red rectangles are zoomed in." + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 308, + 307, + 544, + 343 + ], + "lines": [ + { + "bbox": [ + 308, + 307, + 544, + 343 + ], + "spans": [ + { + "bbox": [ + 308, + 307, + 544, + 343 + ], + "type": "image", + "image_path": "a8d61c4db4aca6afe377c9452a17380f685f8f490498857f122a64dde872bc88.jpg" + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "image_body" + } + ], + "index": 35 + }, + { + "type": "image", + "bbox": [ + 308, + 343, + 544, + 379 + ], + "blocks": [ + { + "bbox": [ + 308, + 343, + 544, + 379 + ], + "lines": [ + { + "bbox": [ + 308, + 343, + 544, + 379 + ], + "spans": [ + { + "bbox": [ + 308, + 343, + 544, + 379 + ], + "type": "image", + "image_path": "53cf844b4b98f4a9add46d018969669764dabdf077c036ee3109367c814fe87e.jpg" + } + ] + } + ], + "index": 36, + "angle": 0, + "type": "image_body" + } + ], + "index": 36 + }, + { + "type": "image", + "bbox": [ + 308, + 379, + 544, + 415 + ], + "blocks": [ + { + "bbox": [ + 308, + 379, + 544, + 415 + ], + "lines": [ + { + "bbox": [ + 308, + 379, + 544, + 415 + ], + "spans": [ + { + "bbox": [ + 308, + 379, + 544, + 415 + ], + "type": "image", + "image_path": "4757f04649ab5f374f031e50ae7c3535afe04acf61d5c9f29cd36a7a1cce1619.jpg" + } + ] + } + ], + "index": 37, + "angle": 0, + "type": "image_body" + } + ], + "index": 37 + }, + { + "type": "image", + "bbox": [ + 308, + 415, + 544, + 451 + ], + "blocks": [ + { + "bbox": [ + 308, + 415, + 544, + 451 + ], + "lines": [ + { + "bbox": [ + 308, + 415, + 544, + 451 + ], + "spans": [ + { + "bbox": [ + 308, + 415, + 544, + 451 + ], + "type": "image", + "image_path": "e3b3b30b2c3dad0b9d3e9e396d776cf60ee347b3f3d8c62a8bf81431d404f735.jpg" + } + ] + } + ], + "index": 38, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 316, + 451, + 346, + 458 + ], + "lines": [ + { + "bbox": [ + 316, + 451, + 346, + 458 + ], + "spans": [ + { + "bbox": [ + 316, + 451, + 346, + 458 + ], + "type": "text", + "content": "Reference" + } + ] + } + ], + "index": 39, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 362, + 450, + 395, + 458 + ], + "lines": [ + { + "bbox": [ + 362, + 450, + 395, + 458 + ], + "spans": [ + { + "bbox": [ + 362, + 450, + 395, + 458 + ], + "type": "text", + "content": "EVFlownet" + } + ] + } + ], + "index": 40, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 416, + 450, + 440, + 458 + ], + "lines": [ + { + "bbox": [ + 416, + 450, + 440, + 458 + ], + "spans": [ + { + "bbox": [ + 416, + 450, + 440, + 458 + ], + "type": "text", + "content": "ERRAFT" + } + ] + } + ], + "index": 41, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 466, + 450, + 483, + 458 + ], + "lines": [ + { + "bbox": [ + 466, + 450, + 483, + 458 + ], + "spans": [ + { + "bbox": [ + 466, + 450, + 483, + 458 + ], + "type": "text", + "content": "TMA" + } + ] + } + ], + "index": 42, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 512, + 450, + 527, + 457 + ], + "lines": [ + { + "bbox": [ + 512, + 450, + 527, + 457 + ], + "spans": [ + { + "bbox": [ + 512, + 450, + 527, + 457 + ], + "type": "text", + "content": "Ours" + } + ] + } + ], + "index": 43, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 306, + 467, + 545, + 487 + ], + "lines": [ + { + "bbox": [ + 306, + 467, + 545, + 487 + ], + "spans": [ + { + "bbox": [ + 306, + 467, + 545, + 487 + ], + "type": "text", + "content": "Figure 8. Qualitative comparisons on the DSEC test set. We visualize the dense predictions and zoom in the areas where are apparent differences." + } + ] + } + ], + "index": 44, + "angle": 0, + "type": "image_caption" + } + ], + "index": 38 + }, + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "text", + "content": "In Fig. 7, we qualitatively compare our proposed EEMFlow with other flow networks, e.g., EVFlownet [54], ER-AFT [11], and KPAFlow [32]. To facilitate comparison, we upsample the meshflow estimation and ground truth to the same resolution of input and zoom in on the areas with the apparent differences. EVFlownet shows the worst performance, with many holes and color mixing. ERAFT and KPAFlow would exhibit block artifacts and appear coarse in nature. In contrast, our EEMFlow results are smoother with more natural color transitions and greater similarity to the upsampled ground truth. In Fig. 6, we present the subjective results of these networks for image registration. We estimate and upsample meshflow prediction from the event sequences " + }, + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "inline_equation", + "content": "E(t_{1},t_{2})" + }, + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "text", + "content": " and then warp image " + }, + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "inline_equation", + "content": "I_{t_1}" + }, + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "text", + "content": " onto image " + }, + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "inline_equation", + "content": "I_{t_2}" + }, + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "text", + "content": ". We also zoom in on the challenging areas of alignment in the registered results, clearly demonstrating that our EEMFlow achieves excellent image registration performance." + } + ] + } + ], + "index": 45 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19204" + } + ] + } + ], + "index": 46 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 49, + 71, + 285, + 148 + ], + "blocks": [ + { + "bbox": [ + 49, + 71, + 285, + 148 + ], + "lines": [ + { + "bbox": [ + 49, + 71, + 285, + 148 + ], + "spans": [ + { + "bbox": [ + 49, + 71, + 285, + 148 + ], + "type": "image", + "image_path": "aebd1753d8341d94d1b4d7596543dfaea1fd6be8807c9858aa8abe6ba244bf69.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 67, + 152, + 266, + 161 + ], + "lines": [ + { + "bbox": [ + 67, + 152, + 266, + 161 + ], + "spans": [ + { + "bbox": [ + 67, + 152, + 266, + 161 + ], + "type": "text", + "content": "Figure 9. Visualization for meshflow and optical flow results." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 54, + 197, + 279, + 258 + ], + "blocks": [ + { + "bbox": [ + 55, + 178, + 277, + 187 + ], + "lines": [ + { + "bbox": [ + 55, + 178, + 277, + 187 + ], + "spans": [ + { + "bbox": [ + 55, + 178, + 277, + 187 + ], + "type": "text", + "content": "Table 4. Results of meshflow and optical flow on HREM for " + }, + { + "bbox": [ + 55, + 178, + 277, + 187 + ], + "type": "inline_equation", + "content": "dt = 1" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 54, + 197, + 279, + 258 + ], + "lines": [ + { + "bbox": [ + 54, + 197, + 279, + 258 + ], + "spans": [ + { + "bbox": [ + 54, + 197, + 279, + 258 + ], + "type": "table", + "html": "
TaskMethodOutdoorIndoorAvg
SlowFastSlowFast
Optical FlowFlowFormer6.2016.065.9915.2710.88
EEMFlow+3.8811.024.0310.927.46
Mesh-flowFlowFormer5.9915.125.7414.9510.45
EEMFlow2.429.092.008.465.50
", + "image_path": "67f9c7891317c47667481ddf5390d13766861b28d3aee9cfde3f75a0d6703fbe.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 283, + 277, + 295 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 283, + 277, + 295 + ], + "spans": [ + { + "bbox": [ + 47, + 283, + 277, + 295 + ], + "type": "text", + "content": "4.2.2 Results for event-based optical flow estimation" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 306, + 287, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 306, + 287, + 498 + ], + "spans": [ + { + "bbox": [ + 46, + 306, + 287, + 498 + ], + "type": "text", + "content": "Table 3 presents the comparative results on DSEC dataset of our optical flow estimation network, EEMFlow+, against other event-based optical flow estimation networks. EEMFlow+ achieves state-of-the-art performance in the 2PE, 3PE, and AE metrics, and is on par with the best in the EPE metric. Notably, EEMFlow+ still maintains a significant advantage in inference speed. Compared to TMA [25], our EEMFlow+ increases the inference speed by " + }, + { + "bbox": [ + 46, + 306, + 287, + 498 + ], + "type": "inline_equation", + "content": "419\\%" + }, + { + "bbox": [ + 46, + 306, + 287, + 498 + ], + "type": "text", + "content": " (from 7.55FPS to 39.2FPS). Fig. 8 displays the qualitative results of our optical flow estimation on the DSEC test set, compared with other advanced methods, including EVFlownet [52], ERAFT [42], and TMA [25]. By zooming into the object areas, it is clearly observable that our results exhibit more regular shapes and clearer boundaries, highlighting the detail enhancement capability of our proposed CDC module." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 514, + 147, + 525 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 514, + 147, + 525 + ], + "spans": [ + { + "bbox": [ + 47, + 514, + 147, + 525 + ], + "type": "text", + "content": "4.3. Ablation Studies" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 534, + 280, + 546 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 534, + 280, + 546 + ], + "spans": [ + { + "bbox": [ + 47, + 534, + 280, + 546 + ], + "type": "text", + "content": "4.3.1 The Advantages of Event-Meshflow Estimation" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 558, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 558, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 558, + 287, + 713 + ], + "type": "text", + "content": "According to the ERAFT [42], image-based approaches face challenges in handling difficult images due to the limited dynamic range of image sensors. As shown in Fig. 9, even advanced image-based techniques like Flow-Former struggle, while our event-based methods (EEMFlow, EEMFlow+) exhibit notably superior performance. Additional quantitative comparisons are provided in Table 4, our network of event-based meshflow estimation EEMFlow achieves the lowest EPE scores in HREM dataset for the " + }, + { + "bbox": [ + 46, + 558, + 287, + 713 + ], + "type": "inline_equation", + "content": "dt = 1" + }, + { + "bbox": [ + 46, + 558, + 287, + 713 + ], + "type": "text", + "content": " input mode, which demonstrates the remarkable efficiency of our EEMFlow makes it suitable for real-time applications such as online video stabilization and autonomous driving." + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 310, + 99, + 539, + 159 + ], + "blocks": [ + { + "bbox": [ + 305, + 70, + 545, + 89 + ], + "lines": [ + { + "bbox": [ + 305, + 70, + 545, + 89 + ], + "spans": [ + { + "bbox": [ + 305, + 70, + 545, + 89 + ], + "type": "text", + "content": "Table 5. Ablation studies about CDC of EEMFlow+. CDC consists of two branches, the self-corrector and self-correlation." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 310, + 99, + 539, + 159 + ], + "lines": [ + { + "bbox": [ + 310, + 99, + 539, + 159 + ], + "spans": [ + { + "bbox": [ + 310, + 99, + 539, + 159 + ], + "type": "table", + "html": "
ModelSelf-correctorSelf-correlationFPS↑3PE↓EPE↓AE↓
(a)XX60.43.520.893.11
(b)✓X55.62.770.812.92
(c)X✓46.32.650.792.78
(d)✓✓39.22.150.752.67
", + "image_path": "58d4be2b18066e05ee7a08df68467e629fa513b9732d10e51c63e14893e39ea8.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 183, + 495, + 194 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 183, + 495, + 194 + ], + "spans": [ + { + "bbox": [ + 305, + 183, + 495, + 194 + ], + "type": "text", + "content": "4.3.2 Experiments for CDC of EEMFlow+" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 205, + 545, + 385 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 205, + 545, + 385 + ], + "spans": [ + { + "bbox": [ + 304, + 205, + 545, + 385 + ], + "type": "text", + "content": "In Table 5, we also conduct ablation experiments on the CDC module of EEMFlow+ used for optical flow estimation, including its two branches, the self-corrector and self-correlation. We train and evaluate all models using the same settings on the DSEC dataset to show the individual impact of each branch in CDC module. Comparison of (a)&(b) demonstrates that the CDC with only the self-corrector branch can bring a significant increase in accuracy with a minimal loss in speed. Comparison of (b)&(c) shows that self-correlation, compared to self-corrector, can lead to a higher increase in accuracy, albeit at a further reduction in inference speed. Finally, the comparison of (a)&(d) shows that the CDC composed of both the self-corrector and self-correlation branches significantly improves the accuracy of optical flow estimation with an acceptable loss in speed." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 402, + 378, + 414 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 402, + 378, + 414 + ], + "spans": [ + { + "bbox": [ + 305, + 402, + 378, + 414 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 424, + 545, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 424, + 545, + 568 + ], + "spans": [ + { + "bbox": [ + 304, + 424, + 545, + 568 + ], + "type": "text", + "content": "In this work, we develop the first event-based meshflow dataset, termed HREM, where 100 indoor and outdoor virtual scenes with rich scene contents are rendered using Blender. Our HREM possess physically correct accurate events and meshflow label pairs with so far the highest resolution. Furthermore, we propose an efficient event-based meshflow network (EEMFlow), which achieves state-of-the-art performance on HREM dataset while maintaining high efficiency. Based on EEMFlow, we propose a confidence-induced detail completion module to upgrade it as EEMFlow+ for optical flow estimation, achieving SOTA results on DSEC dataset with the fastest inference speed." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 570, + 545, + 642 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 545, + 642 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 545, + 642 + ], + "type": "text", + "content": "The integration of optical flow and depth data allows for the extension of the proposed framework to address 3D scene flow. Subsequently, our future work will involve expanding this methodology on the HREM dataset by integrating scene flow annotations and investigating scene flow estimation using event cameras." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 654, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 545, + 712 + ], + "type": "text", + "content": "Acknowledgements. This work was supported by National Natural Science Foundation of China (NSFC) under grant No.62372091, Sichuan Science and Technology Program of China under grants Nos. 2023NSFSC0462, 2022YFQ0079 and 2023YFG0054." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19205" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 714 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 133 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 133 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 133 + ], + "type": "text", + "content": "[1] Ryad Benosman, Sio-Hoi Ieng, Charles Clercq, Chiara Bartolozzi, and Mandyam Srinivasan. Asynchronous frameless event-based optical flow. Neural Networks, 27:32-37, 2012. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 137, + 288, + 179 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 137, + 288, + 179 + ], + "spans": [ + { + "bbox": [ + 53, + 137, + 288, + 179 + ], + "type": "text", + "content": "[2] Christian Brandli, Raphael Berner, Minhao Yang, Shih-Chii Liu, and Tobi Delbruck. A " + }, + { + "bbox": [ + 53, + 137, + 288, + 179 + ], + "type": "inline_equation", + "content": "240 \\times 180" + }, + { + "bbox": [ + 53, + 137, + 288, + 179 + ], + "type": "text", + "content": " 130 db " + }, + { + "bbox": [ + 53, + 137, + 288, + 179 + ], + "type": "inline_equation", + "content": "3\\mu s" + }, + { + "bbox": [ + 53, + 137, + 288, + 179 + ], + "type": "text", + "content": " latency global shutter spatiotemporal vision sensor. IEEE Journal of Solid-State Circuits, 49(10):2333-2341, 2014. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 182, + 288, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 182, + 288, + 226 + ], + "spans": [ + { + "bbox": [ + 53, + 182, + 288, + 226 + ], + "type": "text", + "content": "[3] Marco Cannici, Marco Ciccone, Andrea Romanoni, and Matteo Matteucci. A differentiable recurrent surface for asynchronous event-based data. In Proc. ECCV, pages 136-152. Springer, 2020. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 228, + 288, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 228, + 288, + 271 + ], + "spans": [ + { + "bbox": [ + 53, + 228, + 288, + 271 + ], + "type": "text", + "content": "[4] Javier Cuadrado, Ulysse Rançon, Benoit R Cotterreau, Francisco Barranco, and Timothee Masquelier. Optical flow estimation from event-based cameras and spiking neural networks. Frontiers in Neuroscience, 17:1160034, 2023. 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 273, + 288, + 317 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 273, + 288, + 317 + ], + "spans": [ + { + "bbox": [ + 53, + 273, + 288, + 317 + ], + "type": "text", + "content": "[5] Ziluo Ding, Rui Zhao, Jiyuan Zhang, Tianxiao Gao, Ruiqin Xiong, Zhaofei Yu, and Tiejun Huang. Spatio-temporal recurrent networks for event-based optical flow estimation. In Proc. AAAI, pages 525-533, 2022. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 319, + 288, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 319, + 288, + 373 + ], + "spans": [ + { + "bbox": [ + 53, + 319, + 288, + 373 + ], + "type": "text", + "content": "[6] Alexey Dosovitskiy, Philipp Fischer, Eddy Ilg, Philip Hausser, Caner Hazirbas, Vladimir Golkov, Patrick Van Der Smagt, Daniel Cremers, and Thomas Brox. Flownet: Learning optical flow with convolutional networks. In Proc. ICCV, pages 2758-2766, 2015. 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 374, + 288, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 374, + 288, + 430 + ], + "spans": [ + { + "bbox": [ + 53, + 374, + 288, + 430 + ], + "type": "text", + "content": "[7] Guillermo Gallego, Tobi Delbrück, Garrick Orchard, Chiara Bartolozzi, Brian Taba, Andrea Censi, Stefan Leutenegger, Andrew J Davison, Jörg Conradt, Kostas Daniilidis, et al. Event-based vision: A survey. IEEE transactions on pattern analysis and machine intelligence, 44(1):154-180, 2020. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 431, + 288, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 431, + 288, + 475 + ], + "spans": [ + { + "bbox": [ + 53, + 431, + 288, + 475 + ], + "type": "text", + "content": "[8] Daniel Gehrig, Antonio Loquercio, Konstantinos G Derpanis, and Davide Scaramuzza. End-to-end learning of representations for asynchronous event-based data. In Proc. ICCV, pages 5633-5643, 2019. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 476, + 288, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 476, + 288, + 520 + ], + "spans": [ + { + "bbox": [ + 53, + 476, + 288, + 520 + ], + "type": "text", + "content": "[9] Daniel Gehrig, Mathias Gehrig, Javier Hidalgo-Carrió, and Davide Scaramuzza. Video to events: Recycling video datasets for event cameras. In Proc. CVPR, pages 3586-3595, 2020. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 522, + 287, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 522, + 287, + 566 + ], + "spans": [ + { + "bbox": [ + 48, + 522, + 287, + 566 + ], + "type": "text", + "content": "[10] Mathias Gehrig, Willem Aarents, Daniel Gehrig, and Davide Scaramuzza. Dsec: A stereo event camera dataset for driving scenarios. IEEE Robotics and Automation Letters, 6(3): 4947-4954, 2021. 2, 3, 4, 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 567, + 287, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 567, + 287, + 610 + ], + "spans": [ + { + "bbox": [ + 48, + 567, + 287, + 610 + ], + "type": "text", + "content": "[11] Mathias Gehrig, Mario Millhäuser, Daniel Gehrig, and Davide Scaramuzza. E-raft: Dense optical flow from event cameras. In International Conference on 3D Vision (3DV), pages 197–206, 2021. 1, 2, 3, 6, 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 613, + 287, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 613, + 287, + 667 + ], + "spans": [ + { + "bbox": [ + 48, + 613, + 287, + 667 + ], + "type": "text", + "content": "[12] Klaus Greff, Francois Belletti, Lucas Beyer, Carl Doersch, Yilun Du, Daniel Duckworth, David J Fleet, Dan Gnanapragasam, Florian Golemo, Charles Herrmann, et al. Kubric: A scalable dataset generator. In Proc. CVPR, pages 3749-3761, 2022. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 670, + 287, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 670, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 48, + 670, + 287, + 714 + ], + "type": "text", + "content": "[13] Jesse Hagenaars, Federico Paredes-Vallés, and Guido De Croon. Self-supervised learning of event-based optical flow with spiking neural networks. Advances in Neural Information Processing Systems, 34:7167-7179, 2021. 2" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 714 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "text", + "content": "[14] Yuhuang Hu, Shih-Chii Liu, and Tobi Delbruck. v2e: From video frames to realistic dvs events. In Proc. CVPR, pages 1312-1321, 2021. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 108, + 545, + 152 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 108, + 545, + 152 + ], + "spans": [ + { + "bbox": [ + 307, + 108, + 545, + 152 + ], + "type": "text", + "content": "[15] Zhaoyang Huang, Xiaoyu Shi, Chao Zhang, Qiang Wang, Ka Chun Cheung, Hongwei Qin, Jifeng Dai, and Hongsheng Li. FlowFormer: A transformer architecture for optical flow. ECCV, 2022. 2, 6, 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 154, + 545, + 186 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 154, + 545, + 186 + ], + "spans": [ + { + "bbox": [ + 307, + 154, + 545, + 186 + ], + "type": "text", + "content": "[16] Junhwa Hur and Stefan Roth. Iterative residual refinement for joint optical flow and occlusion estimation. In Proc. CVPR, 2019. 4" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 189, + 545, + 232 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 189, + 545, + 232 + ], + "spans": [ + { + "bbox": [ + 307, + 189, + 545, + 232 + ], + "type": "text", + "content": "[17] Shihao Jiang, Dylan Campbell, Yao Lu, Hongdong Li, and Richard Hartley. Learning to estimate hidden motions with global motion aggregation. In Proc. ICCV, pages 9772-9781, 2021. 2, 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 235, + 545, + 323 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 235, + 545, + 323 + ], + "spans": [ + { + "bbox": [ + 307, + 235, + 545, + 323 + ], + "type": "text", + "content": "[18] Jacques Kaiser, J Camilo Vasquez Tieck, Christian Hubschneider, Peter Wolf, Michael Weber, Michael Hoff, Alexander Friedrich, Konrad Wojtasik, Arne Roennau, Ralf Kohlhaas, et al. Towards a framework for end-to-end control of a simulated vehicle with spiking neural networks. In IEEE International Conference on Simulation, Modeling, and Programming for Autonomous Robots (SIMPAR), pages 127-134, 2016. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 325, + 545, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 325, + 545, + 380 + ], + "spans": [ + { + "bbox": [ + 307, + 325, + 545, + 380 + ], + "type": "text", + "content": "[19] Chankyu Lee, Adarsh Kumar Kosta, Alex Zihao Zhu, Kenneth Chaney, Kostas Daniilidis, and Kaushik Roy. Spikeflownet: event-based optical flow estimation with energy-efficient hybrid neural networks. In Proc. ECCV, pages 366-382. Springer, 2020. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 382, + 545, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 382, + 545, + 437 + ], + "spans": [ + { + "bbox": [ + 307, + 382, + 545, + 437 + ], + "type": "text", + "content": "[20] Chankyu Lee, Adarsh Kumar Kosta, and Kaushik Roy. Fusion-flownet: Energy-efficient optical flow estimation using sensor fusion and deep fused spiking-analog network architectures. In 2022 International Conference on Robotics and Automation (ICRA), pages 6504-6510. IEEE, 2022. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 439, + 545, + 494 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 439, + 545, + 494 + ], + "spans": [ + { + "bbox": [ + 307, + 439, + 545, + 494 + ], + "type": "text", + "content": "[21] Yijin Li, Zhaoyang Huang, Shuo Chen, Xiaoyu Shi, Hongsheng Li, Hujun Bao, Zhaopeng Cui, and Guofeng Zhang. Blinkflow: A dataset to push the limits of event-based optical flow estimation. arXiv preprint arXiv:2303.07716, 2023. 2, 3, 7" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 496, + 545, + 540 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 496, + 545, + 540 + ], + "spans": [ + { + "bbox": [ + 307, + 496, + 545, + 540 + ], + "type": "text", + "content": "[22] Zhuoyan Li, Jiawei Shen, and Ruitao Liu. A lightweight network to learn optical flow from event data. In 2020 25th International Conference on Pattern Recognition (ICPR), pages 1-7. IEEE, 2021. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 543, + 545, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 543, + 545, + 586 + ], + "spans": [ + { + "bbox": [ + 307, + 543, + 545, + 586 + ], + "type": "text", + "content": "[23] Kaimo Lin, Shuaicheng Liu, Loong-Fah Cheong, and Bing Zeng. Seamless video stitching from hand-held camera inputs. In Computer Graphics Forum, pages 479-487. Wiley Online Library, 2016. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 588, + 545, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 588, + 545, + 632 + ], + "spans": [ + { + "bbox": [ + 307, + 588, + 545, + 632 + ], + "type": "text", + "content": "[24] Songnan Lin, Ye Ma, Zhenhua Guo, and Bihan Wen. Dvsvoltmeter: Stochastic process-based event simulator for dynamic vision sensors. In Proc. ECCV, pages 578-593. Springer, 2022. 2, 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 634, + 545, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 634, + 545, + 678 + ], + "spans": [ + { + "bbox": [ + 307, + 634, + 545, + 678 + ], + "type": "text", + "content": "[25] Haotian Liu, Guang Chen, Sanqing Qu, Yanping Zhang, Zhi-jun Li, Alois Knoll, and Changjun Jiang. Tma: Temporal motion aggregation for event-based optical flow. In Proc. ICCV, 2023. 2, 7, 8" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 680, + 545, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 680, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 307, + 680, + 545, + 714 + ], + "type": "text", + "content": "[26] Shuaicheng Liu, Lu Yuan, Ping Tan, and Jian Sun. Bundled camera paths for video stabilization. ACM transactions on graphics (TOG), 32(4):1-10, 2013. 1" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "19206" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 286, + 712 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 48, + 72, + 286, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 286, + 106 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 286, + 106 + ], + "type": "text", + "content": "[27] Shuaicheng Liu, Lu Yuan, Ping Tan, and Jian Sun. Steadyflow: Spatially smooth optical flow for video stabilization. In Proc. CVPR, pages 4209-4216, 2014. 1" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 107, + 286, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 107, + 286, + 150 + ], + "spans": [ + { + "bbox": [ + 49, + 107, + 286, + 150 + ], + "type": "text", + "content": "[28] Shuaicheng Liu, Ping Tan, Lu Yuan, Jian Sun, and Bing Zeng. Meshflow: Minimum latency online video stabilization. In Proc. ECCV, pages 800-815. Springer, 2016. 1, 2, 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 153, + 286, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 153, + 286, + 196 + ], + "spans": [ + { + "bbox": [ + 49, + 153, + 286, + 196 + ], + "type": "text", + "content": "[29] Shuaicheng Liu, Nianjin Ye, Chuan Wang, Jirong Zhang, Lanpeng Jia, Kunming Luo, Jue Wang, and Jian Sun. Content-aware unsupervised deep homography estimation and its extensions. 45(3):2849-2863, 2022. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 198, + 286, + 242 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 198, + 286, + 242 + ], + "spans": [ + { + "bbox": [ + 49, + 198, + 286, + 242 + ], + "type": "text", + "content": "[30] Shuaicheng Liu, Yuhang Lu, Hai Jiang, Nianjin Ye, Chuan Wang, and Bing Zeng. Unsupervised global and local homography estimation with motion basis learning. 45(6): 7885-7899, 2023. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 49, + 244, + 286, + 286 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 244, + 286, + 286 + ], + "spans": [ + { + "bbox": [ + 49, + 244, + 286, + 286 + ], + "type": "text", + "content": "[31] Zhen Liu, Yinglong Wang, Bing Zeng, and Shuaicheng Liu. Ghost-free high dynamic range imaging with context-aware transformer. In Proc. ECCV, pages 344-360. Springer, 2022. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 49, + 289, + 286, + 322 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 289, + 286, + 322 + ], + "spans": [ + { + "bbox": [ + 49, + 289, + 286, + 322 + ], + "type": "text", + "content": "[32] Ao Luo, Fan Yang, Xin Li, and Shuaicheng Liu. Learning optical flow with kernel patch attention. In Proc. CVPR, 2022. 2, 6, 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 49, + 324, + 286, + 366 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 324, + 286, + 366 + ], + "spans": [ + { + "bbox": [ + 49, + 324, + 286, + 366 + ], + "type": "text", + "content": "[33] Xinglong Luo, Kunming Luo, Ao Luo, Zhengning Wang, Ping Tan, and Shuaicheng Liu. Learning optical flow from event camera with rendered dataset. In Proc. ICCV, 2023. 2, 3, 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 49, + 369, + 286, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 369, + 286, + 403 + ], + "spans": [ + { + "bbox": [ + 49, + 369, + 286, + 403 + ], + "type": "text", + "content": "[34] Lang Nie, Chunyu Lin, Kang Liao, Shuaicheng Liu, and Yao Zhao. Depth-aware multi-grid deep homography estimation with contextual correlation. 32(7):4460-4472, 2022. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 49, + 404, + 286, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 404, + 286, + 437 + ], + "spans": [ + { + "bbox": [ + 49, + 404, + 286, + 437 + ], + "type": "text", + "content": "[35] Lang Nie, Chunyu Lin, Kang Liao, Shuaicheng Liu, and Yao Zhao. Parallax-tolerant unsupervised deep image stitching. In Proc. ICCV, pages 7399-7408, 2023. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 49, + 439, + 286, + 472 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 439, + 286, + 472 + ], + "spans": [ + { + "bbox": [ + 49, + 439, + 286, + 472 + ], + "type": "text", + "content": "[36] Liyuan Pan, Miaomiao Liu, and Richard Hartley. Single image optical flow estimation with an event camera. In Proc. CVPR, pages 1669-1678, 2020. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 49, + 474, + 286, + 506 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 474, + 286, + 506 + ], + "spans": [ + { + "bbox": [ + 49, + 474, + 286, + 506 + ], + "type": "text", + "content": "[37] Henri Rebecq, Daniel Gehrig, and Davide Scaramuzza. Esim: an open event camera simulator. In Conference on robot learning, pages 969-982. PMLR, 2018. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 49, + 508, + 286, + 540 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 508, + 286, + 540 + ], + "spans": [ + { + "bbox": [ + 49, + 508, + 286, + 540 + ], + "type": "text", + "content": "[38] Zhihang Ren, Jiajia Li, Shuaicheng Liu, and Bing Zeng. Meshflow video denoising. In Proc. ICIP, pages 2966-2970, 2017. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 49, + 543, + 286, + 574 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 543, + 286, + 574 + ], + "spans": [ + { + "bbox": [ + 49, + 543, + 286, + 574 + ], + "type": "text", + "content": "[39] Deqing Sun, Stefan Roth, and Michael J Black. Secrets of optical flow estimation and their principles. In Proc. CVPR, pages 2432-2439, 2010. 3, 7" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 49, + 577, + 286, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 577, + 286, + 609 + ], + "spans": [ + { + "bbox": [ + 49, + 577, + 286, + 609 + ], + "type": "text", + "content": "[40] Deqing Sun, Xiaodong Yang, Ming-Yu Liu, and Jan Kautz. Pwc-net: Cnns for optical flow using pyramid, warping, and cost volume. In Proc. CVPR, pages 8934-8943, 2018. 4, 6" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 49, + 612, + 286, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 612, + 286, + 643 + ], + "spans": [ + { + "bbox": [ + 49, + 612, + 286, + 643 + ], + "type": "text", + "content": "[41] Shangkun Sun, Yuanqi Chen, Yu Zhu, Guodong Guo, and Ge Li. Skflow: Learning optical flow with super kernels. Proc. NeurIPS, 35:11313-11326, 2022. 2, 6" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 49, + 646, + 286, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 646, + 286, + 668 + ], + "spans": [ + { + "bbox": [ + 49, + 646, + 286, + 668 + ], + "type": "text", + "content": "[42] Zachary Teed and Jia Deng. Raft: Recurrent all-pairs field transforms for optical flow. In Proc. ECCV, 2020. 2, 6, 8" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 49, + 670, + 286, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 670, + 286, + 712 + ], + "spans": [ + { + "bbox": [ + 49, + 670, + 286, + 712 + ], + "type": "text", + "content": "[43] Zhexiong Wan, Yuchao Dai, and Yuxin Mao. Learning dense and continuous optical flow from an event camera. IEEE Transactions on Image Processing, 31:7237-7251, 2022. 2, 3" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 532 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 117 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 117 + ], + "type": "text", + "content": "[44] Zhexiong Wan, Yuxin Mao, Jing Zhang, and Yuchao Dai. Rpeflow: Multimodal fusion of rgb-pointcloud-event for joint optical flow and scene flow estimation. In Proc. ICCV, pages 10030-10040, 2023. 2, 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 119, + 545, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 119, + 545, + 150 + ], + "spans": [ + { + "bbox": [ + 307, + 119, + 545, + 150 + ], + "type": "text", + "content": "[45] Yiming Wang, Qian Huang, Chuanxu Jiang, Jiwen Liu, Mingzhou Shang, and Zhuang Miao. Video stabilization: A comprehensive survey. Neurocomputing, 2022. 1, 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 152, + 545, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 152, + 545, + 185 + ], + "spans": [ + { + "bbox": [ + 307, + 152, + 545, + 185 + ], + "type": "text", + "content": "[46] Qingsen Yan, Jinqiu Sun, Haisen Li, Yu Zhu, and Yanning Zhang. High dynamic range imaging by sparse representation. Neurocomputing, 269:160-169, 2017. 1, 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 186, + 545, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 186, + 545, + 229 + ], + "spans": [ + { + "bbox": [ + 307, + 186, + 545, + 229 + ], + "type": "text", + "content": "[47] Min Yang, Jingkun Liang, Jianhai Zhang, Haidong Gao, Fanyong Meng, Li Xingdong, and Sung-Jin Song. Non-local means theory based perona-malik model for image denoising. Neurocomputing, 120:262-267, 2013. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 232, + 545, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 232, + 545, + 252 + ], + "spans": [ + { + "bbox": [ + 307, + 232, + 545, + 252 + ], + "type": "text", + "content": "[48] Yan Yang, Liyuan Pan, and Liu Liu. Event camera data pretraining. In Proc. ICCV, pages 10699-10709, 2023. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 254, + 545, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 254, + 545, + 297 + ], + "spans": [ + { + "bbox": [ + 307, + 254, + 545, + 297 + ], + "type": "text", + "content": "[49] Nianjin Ye, Chuan Wang, Shuaicheng Liu, Lanpeng Jia, Jue Wang, and Yongqing Cui. Deepmeshflow: Content adaptive mesh deformation for robust image registration. arXiv preprint arXiv:1912.05131, 2019. 1, 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 299, + 545, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 299, + 545, + 341 + ], + "spans": [ + { + "bbox": [ + 307, + 299, + 545, + 341 + ], + "type": "text", + "content": "[50] Yaozu Ye, Hao Shi, Kailun Yang, Ze Wang, Xiaoting Yin, Yaonan Wang, and Kaiwei Wang. Towards anytime optical flow estimation with event cameras. arXiv preprint arXiv:2307.05033, 2023. 2, 7" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 343, + 545, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 343, + 545, + 386 + ], + "spans": [ + { + "bbox": [ + 307, + 343, + 545, + 386 + ], + "type": "text", + "content": "[51] Xiangyu Zhang, Xinyu Zhou, Mengxiao Lin, and Jian Sun. Shufflenet: An extremely efficient convolutional neural network for mobile devices. In Proc. CVPR, pages 6848-6856, 2018. 4" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 388, + 545, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 388, + 545, + 431 + ], + "spans": [ + { + "bbox": [ + 307, + 388, + 545, + 431 + ], + "type": "text", + "content": "[52] Alex Zhu, Liangzhe Yuan, Kenneth Chaney, and Kostas Daniilidis. Ev-flownet: Self-supervised optical flow estimation for event-based cameras. In Proceedings of Robotics: Science and Systems, 2018. 2, 4, 5, 6, 7, 8" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 434, + 545, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 434, + 545, + 487 + ], + "spans": [ + { + "bbox": [ + 307, + 434, + 545, + 487 + ], + "type": "text", + "content": "[53] Alex Zihao Zhu, Dinesh Thakur, Tolga Ozaslan, Bernd Pfrommer, Vijay Kumar, and Kostas Daniilidis. The multi-vehicle stereo event camera dataset: An event camera dataset for 3d perception. IEEE Robotics and Automation Letters, 3 (3):2032-2039, 2018. 2, 3, 4" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 489, + 545, + 532 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 489, + 545, + 532 + ], + "spans": [ + { + "bbox": [ + 307, + 489, + 545, + 532 + ], + "type": "text", + "content": "[54] Alex Zihao Zhu, Liangzhe Yuan, Kenneth Chaney, and Kostas Daniilidis. Unsupervised event-based learning of optical flow, depth, and egomotion. In Proc. CVPR, pages 989-997, 2019. 3, 6, 7" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19207" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Efficient Model Stealing Defense with Noise Transition Matrix/5181a0c8-f9af-47c9-b829-4e394817b1e7_content_list.json b/2024/Efficient Model Stealing Defense with Noise Transition Matrix/5181a0c8-f9af-47c9-b829-4e394817b1e7_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..78881ff55d39598b4e8dd887351d2f2bb2fb2513 --- /dev/null +++ b/2024/Efficient Model Stealing Defense with Noise Transition Matrix/5181a0c8-f9af-47c9-b829-4e394817b1e7_content_list.json @@ -0,0 +1,1804 @@ +[ + { + "type": "text", + "text": "Efficient Model Stealing Defense with Noise Transition Matrix", + "text_level": 1, + "bbox": [ + 171, + 130, + 799, + 151 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Dong-Dong $\\mathsf{W u}^{1}$ , Chilin $\\mathsf{F u}^{2}$ , Weichang $\\mathsf{W u}^{2}$ , Wenwen $\\mathsf{X i a}^{2}$", + "bbox": [ + 235, + 179, + 730, + 198 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Xiaolu Zhang $^{2}$ , Jun Zhou $^{2*}$ , Min-Ling Zhang $^{1*}$", + "bbox": [ + 294, + 198, + 673, + 217 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ School of Computer Science and Engineering, Southeast University, Nanjing, China", + "bbox": [ + 150, + 215, + 818, + 234 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ Ant Group", + "bbox": [ + 437, + 232, + 532, + 252 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{dongdongwu,zhangml}@seu.edu.cn,{chilin.fcl,jiuyue.wwc,wenzhen.xww}@antgroup.com", + "bbox": [ + 124, + 253, + 844, + 268 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{yueyin.zxl, jun.zhoujun}@antfin.com", + "bbox": [ + 316, + 272, + 635, + 286 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 321, + 313, + 338 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "With the escalating complexity and investment cost of training deep neural networks, safeguarding them from unauthorized usage and intellectual property theft has become imperative. Especially the rampant misuse of prediction APIs to replicate models without access to the original data or architecture poses grave security threats. Diverse defense strategies have emerged to address these vulnerabilities, yet these defenses either incur heavy inference overheads or assume idealized attack scenarios. To address these challenges, we revisit the utilization of noise transition matrix as an efficient perturbation technique, which injects noise into predicted posteriors in a linear manner and integrates seamlessly into existing systems with minimal overhead, for model stealing defense. Provably, with such perturbed posteriors, the attacker's cloning process degrades into learning from noisy data. Toward optimizing the noise transition matrix, we proposed a novel bi-level optimization training framework, which performs fidelity on the victim model while the surrogate model adversarially. Comprehensive experimental results demonstrate that our method effectively thwarts model stealing attacks and achieves minimal utility trade-offs, outperforming existing state-of-the-art defenses.", + "bbox": [ + 75, + 353, + 473, + 686 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 715, + 209, + 732 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Machine learning (ML) models, especially deep neural networks (DNNs), have attained human-level capabilities across various domains, such as pattern recognition tasks [24, 27] and generation tasks [2, 52], becoming indispensable in facilitating convenience for society [53]. With the burgeoning of ML application scenarios and the proliferation of data, training large and sophisticated models and making them publicly accessible has become a new profit trend for compa", + "bbox": [ + 75, + 742, + 470, + 864 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "nies. However, considering the steep costs of tuning models, directly divulging model parameters potentially engenders no returns on investment and even adverse societal impacts from malicious use. To alleviate those concerns, prediction APIs have emerged as an alternative by enabling black-box inferences through querying. This allows models to provide services without exposing the underlying parameters and details, and has been harnessed in Machine Learning as a Service (MLaaS) platforms (e.g., Amazon AWS, Microsoft Azure, Google Cloud).", + "bbox": [ + 500, + 323, + 893, + 474 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Unfortunately, recent research has demonstrated that prediction APIs are still vulnerable to intellectual property theft through model stealing attacks [39, 51, 54, 62]. Despite not exposing model details, attackers can query the APIs using carefully-crafted samples or surrogate samples to obtain responded predictions. Through this, they can then train a clone model on the gathered query-response pairs to replicate the victim model's functionality with high accuracy and fidelity, without needing to access the original training data or model architecture [22, 39]. Such a clone model poses the following security threats: (a) acquiring a functionally similar copy of the victim model, illegally circumventing the expensive process of manual data curation and undermining the intellectual property of MLaaS; (b) extracting confidential data through membership inference attacks [47], model inversion attack [6] and property inference attack [7]; (c) conducting adversarial attacks [31, 33, 35, 36, 49, 55] by creating malicious examples based on the white-box substitute model [42].", + "bbox": [ + 496, + 476, + 895, + 762 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To address such vulnerabilities, considerable researches [3, 5, 9, 17, 29, 50, 64] have focused on thwarting model stealing attacks. These defenses can be categorized into three strategies: detection, perturbation, and post-hoc verification. Detection-based defenses [19, 20, 23, 23] are predicated on the idea that malicious queries exhibit aberrant statistical distributions compared to normal queries. If a query sequence appears anomalous, the defender may discard the requests or corrupt the outputs. However, if an attacker carefully con", + "bbox": [ + 496, + 763, + 895, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Corresponding author.", + "bbox": [ + 86, + 875, + 217, + 887 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "This work is done when Dong-Dong Wu works as an intern at Ant Group.", + "bbox": [ + 86, + 887, + 467, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "24305", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/27b8898ae862573b82ebeb905ba6a3ba9fe7846f0081b6bf335764e4c4751b5c.jpg", + "image_caption": [ + "Figure 1. Illustrations of the noise transition matrix serving for noisy label learning and model stealing defense. In noisy label learning, it transforms the predicted posteriors to align with the noisy label. In model stealing defense, it perturbs the predicted posterior to impede attacker training while preserving the original predicted label for benign users." + ], + "image_footnote": [], + "bbox": [ + 80, + 89, + 467, + 195 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "structures queries matching the benign distribution, detection techniques will fail to discern the attack, yielding excessive false positives. Perturbation-based defenses [30, 40, 51] can preserve or constraint model accuracy to foil model stealing attacks. The accuracy-preserving defense truncates predicted posteriors to a minuscule fraction of their original size, which may reduce model lucidity and utility for critical applications like autonomous driving. The accuracy-constraint defense [38, 40] introduces perturbation into predicted posteriors, which incurs additional computational burdens or substantially degrading utility. Furthermore, the latency resulting from such perturbation methods leads to extended user response times, making the approach impractical. Post-hoc verification techniques such as watermarking [15] remain useful for post-attack prevention and validation. However, their retroactive nature means they cannot take proactive measures before an attack occurs. Considering inference latency and utility trade-offs, in this paper, we concentrate on a novel and efficient model stealing defense strategy.", + "bbox": [ + 75, + 323, + 468, + 611 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To overcome the limitations of existing defenses, we explore an efficient method to inject noise into predicted posteriors via linear mapping, while preserving utility for benign users. Motivated by memorization effects [11, 63] that neural networks can easily overfit noisy data, thus hampering model generalization, we revisit the utilization of noise transition matrix (NTM) in noisy labels learning [43, 58], where the matrix models the noise process. Unlike related work, we aim to leverage the NTM to introduce noise into predicted posteriors for model defense, as shown in Figure 1. Our goal is to find an optimized NTM that injects noise while retaining original prediction labels with minimal utility trade-offs. Intuitively, regardless of whether the victim model produces extremely confident one-hot vectors or uncertain flat vectors, we can readily derive a uniform NTM where the diagonal elements remain maximal for each row to meet our defense needs. To estimate the NTM with maximum utility, we propose a bi-level optimization method that alternately optimizes it and updates the surrogate network", + "bbox": [ + 75, + 613, + 470, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "weights. Specifically, we first make a one-step-forward virtual optimization of weights, then optimize the matrix guided by the loss with the one-step-forward weights fixed. Finally, we optimize the unrolled weights with the updated matrix. The contributions of this work are as follows:", + "bbox": [ + 496, + 90, + 893, + 165 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose a novel perturbation-based defense that efficiently introduces noise through linear mapping, adding minimal overhead. Furthermore, it deteriorates attacker training to a level equivalent to training on noisy-labeled data with instance-wise weight.", + "- Our method pioneers using the transition matrix to thwart model stealing attacks. The explicit matrix helps understand noise relationships among classes and makes the defense strategy more interpretable", + "- A bi-level optimization approach is proposed to optimize the transition matrix, which is model-agnostic and can adapt to different backbone networks.", + "- Extensive experiments exhibit our method consistently mitigates stealing and outperforms state-of-the-art defenses." + ], + "bbox": [ + 500, + 167, + 893, + 378 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 397, + 640, + 412 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Model Stealing Attacks aim to replicate the functionality of a victim model by interrogating its prediction APIs. The attacker sends queries and collects the resultant predictions, typically probability vectors or labels, as query-response pairs. These collected pairs are used to train a clone model that approximates the victim model. Tramèr et al. [51] first proposed the idea of model stealing and demonstrated the feasibility of extracting simple models utilizing queries. Juuti et al. [19] presented extraction attacks leveraging jacobian-based data augmentation [42], synthesizing data by perturbing a small set of in-distribution samples. However, these attacks presume the ability to optimally probe the APIs, which could be detected by monitoring for anomalous query patterns [19, 41]. Therefore, we concentrate on scenarios without carefully engineered input queries or sequences. One classic scenario to achieve this is by utilizing knowledge distillation techniques [12], where some studies exhibit model stealing is viable using weakly correlated [39] or random [25] queries, even when the teacher model declares itself undistillable [13, 28].", + "bbox": [ + 496, + 424, + 893, + 727 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Model Stealing Defenses can be taxonomized into three primary strategies: detection, perturbation, and post-hoc verification. Detection strategies [20, 21] aim to identify malicious queries or nefarious users by scrutinizing the queries received by the victim model. For example, Kariyappa and Qureshi [20] regarded malicious queries as out-of-distribution samples and devised an adaptive misinformation (AM) defense to mislead attackers with modified outputs. However, such detection strategies rely on strong assumptions about adversarial query patterns that may not reflect reality. Perturbation strategies mitigate model stealing", + "bbox": [ + 496, + 734, + 895, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "24306", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "attacks by returning perturbed posterior, which can be further subcategorized into accuracy-preserving [30] and accuracy-constrained defenses [29, 38, 40]. Accuracy-preserving defenses guarantee that the maximal class of the perturbed predictions matches the original prediction. The most straightforward such defense is truncating the posterior or only outputting the predicted label [51]. However, this diminishes utility for benign users, precluding specific applications and harming external transparency. In contrast, accuracy-constrained defenses permit a higher level of perturbation, trading off benign accuracy for enhanced security against model stealing attacks. In general, existing defense strategies that poison the attacker's update gradient [38, 40] are computationally expensive, even if they yield a comparative defense. Post-hoc verification, including watermarking [15, 32, 50], dataset inference [37], and proof of learning [16], embeds concealed patterns in the model during training [1] or inference [50] to distinguish a stolen model. Unfortunately, these approaches only work after a model theft, requiring the model owner to obtain partial or full access to the stolen model. This means if an attacker uses the stolen model as a proxy to attack the victim model, it is impossible for the model owner to protect themselves with these defenses. Overall, while existing ante-hoc defensive approaches mitigate knowledge leakage, they rely on strong assumptions or incur high computational costs during inference.", + "bbox": [ + 76, + 90, + 472, + 484 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Noise Transition Matrix (NTM) is often utilized to solve classification with noisy labels or ambiguous labeled data [10, 46, 61], where the matrix reflects probabilities of true label flipping to the other label. A typical application is loss correction [43], which modifies the loss of each example by multiplying the estimated matrix by the model output. For NTM estimation, there are two main approaches. One is a two-step solution that pre-estimates the matrix with anchor point assumptions, then uses it to train the classifier [56, 65]. For instance, backward correction [43] initially approximates the NTM using the softmax output of the model trained without correction. It then retransmits the model while correcting the original loss based on the estimated matrix. The other jointly estimates the matrix and classifier parameters in a unified framework without employing anchor points [18, 48, 60]. Differently, our work aims to estimate and apply the transition amtrix to intentionally inject noise into predicted posteriors, thus defending against model stealing.", + "bbox": [ + 76, + 491, + 470, + 765 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Threat Model", + "text_level": 1, + "bbox": [ + 76, + 782, + 217, + 797 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We consider a model stealing scenario involving a defender and an attacker interacting on image classification tasks. The defender grants access to its black-box model through prediction API. Leveraging the API, the attacker collects query-response pairs, aiming to mimic the defender model (victim model) functionalities by training a clone model on", + "bbox": [ + 75, + 809, + 470, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "this dataset. To evade detection defenses [19], we assume the attacker adopts non-adaptive querying strategies without anomalous patterns [39, 42]. Therefore, we consider the knockoff nets attack without adaptive querying [39], which resembles knowledge distillation [12] where the victim model distills knowledge into the attacker model.", + "bbox": [ + 496, + 90, + 893, + 181 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Attacker's Objective", + "text_level": 1, + "bbox": [ + 498, + 188, + 694, + 205 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Let $\\theta_{f}$ and $\\theta_{h}$ denote the parameters of the defender's model $f$ and attacker's clone model $h$ , respectively. The attacker prepares a query dataset $\\mathcal{D}_{\\mathrm{query}}$ , and sends queries $\\pmb{x} \\in \\mathcal{D}_{\\mathrm{query}}$ in random order to obtain the defender's posterior $f(\\pmb{x}; \\pmb{\\theta}_{f})$ from the prediction API. The classification loss of the attacker's clone model on the query $\\pmb{x}$ is $\\mathcal{L}_{\\mathrm{kna}} = -\\sum_{i} f(\\pmb{x}; \\pmb{\\theta}_{f})_{i} \\log h(\\pmb{x}; \\pmb{\\theta}_{h})_{i}$ . The attacker aims to minimize the classification loss on the defender's test set. However, since the test set is unavailable, the attacker's learning objective becomes minimizing the loss based on the collected query-response dataset $\\mathcal{D}_{\\mathrm{pair}} = \\{(x, f(x; \\pmb{\\theta}_{f})) | x \\in \\mathcal{D}_{\\mathrm{query}}\\}$ . Therefore, the attacker's objective can be formulated as:", + "bbox": [ + 496, + 210, + 893, + 393 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {\\boldsymbol {\\theta} _ {h}} \\mathbb {E} _ {\\boldsymbol {x} \\sim \\mathcal {D} _ {\\text {p a i r}}} \\left[ \\mathcal {L} _ {\\mathrm {k n a}} \\left(h (\\boldsymbol {x}; \\boldsymbol {\\theta} _ {h}), f (\\boldsymbol {x}; \\boldsymbol {\\theta} _ {f})\\right) \\right] \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 563, + 398, + 890, + 422 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Query Distribution. In practice, due to the knowledge limitation of attackers about the defender's training set, the collected query dataset $\\mathcal{D}_{\\mathrm{query}}$ may be similar [30, 51] or unrelated [20, 40] to the defender's training distribution. Hence, we consider both possibilities in our threat model. We refer to attackers as distribution-aware if their queries share semantic content with the defender's training distribution. Attackers are referred to as knowledge-limited if their queries are semantically disjoint from the defender's training distribution.", + "bbox": [ + 496, + 429, + 893, + 580 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Defender's Objective", + "text_level": 1, + "bbox": [ + 498, + 587, + 697, + 603 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In defenses against model stealing, the defender aims at preventing its model functionality from being stolen. In real-world scenarios, attackers may disguise themselves as benign users, executing attacks anytime. Therefore, the defender must persistently mitigate model stealing attempts while preserving the API's utility for legitimate users. Specifically, the defender needs to keep its own small classification loss, while maximizing the classification loss of the attacker's cloned model $\\theta_h^*(\\theta_f)$ obtained from Eq. (1) on the test set $\\mathcal{D}_{\\mathrm{test}}$ . The defender's objective can be formulated as:", + "bbox": [ + 496, + 609, + 893, + 762 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {\\boldsymbol {\\theta} _ {f}} \\mathbb {E} _ {(\\boldsymbol {x}, y) \\sim \\mathcal {D} _ {\\text {t e s t}}} [ \\mathcal {L} _ {\\mathrm {C E}} (f (\\boldsymbol {x}; \\boldsymbol {\\theta} _ {f}), y) ] \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 575, + 767, + 890, + 796 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {s. t.} \\max _ {\\mathbf {x}, y) \\sim \\mathcal {D} _ {\\text {t e s t}}} [ \\mathcal {L} _ {\\mathrm {C E}} (h (\\boldsymbol {x}; \\boldsymbol {\\theta} _ {h} ^ {*} (\\boldsymbol {\\theta} _ {f})), y) ],\n$$\n", + "text_format": "latex", + "bbox": [ + 549, + 796, + 844, + 813 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mathcal{L}_{\\mathrm{CE}}$ means the cross-entropy loss function.", + "bbox": [ + 500, + 818, + 828, + 833 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4. Approach", + "text_level": 1, + "bbox": [ + 500, + 844, + 609, + 861 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To develop an efficacious and inference-efficient defense against model stealing, we propose an Efficient Model de", + "bbox": [ + 498, + 869, + 893, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "24307", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "fense with noise transition MMatrix (EMMA) to mitigate model stealing attacks. As depicted in the \"Defense phase\" of Figure 2, the attacker iteratively queries the victim model using unlabeled data to obtain posteriors via the prediction API. The victim model receives these queries and outputs corresponding predicted probability vectors. These posteriors then multiply the optimized matrix $T^n$ , becoming perturbed posteriors that remain useful to benign users, and returned to the attacker. However, if the attacker collects these query-responses as training data to reconstruct a clone model, in that case, this dataset is equivalent to noisy labeled data with instance-wise weight, thus impeding cloning model training. To optimize the transition matrix with maximum utility, as shown in the \"Optimizing phase\", we devise a bi-level optimization between the victim and surrogate model which is devised to represent the attacker model. In each round, given initial $T^t$ and surrogate weights $\\theta_g^t$ , the optimized matrix $T^{t+1}$ stems from two gradient descents (GD) on the attack loss and defense loss, respectively. The surrogate model then updates through a GD on the attack loss. This process loops until reaching the termination condition, yielding the final optimized matrix $T^n$ . In this section, we first elucidate the defense mechanism of incorporating the transition matrix during the defensive phase. Then, a bi-level optimization approach is expounded to alternately optimize the matrix and update the surrogate model.", + "bbox": [ + 75, + 90, + 472, + 484 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1. The underlying mechanism", + "text_level": 1, + "bbox": [ + 76, + 493, + 323, + 508 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To understand how the transition matrix defends against attacks, we analyze the classification risks of two scenarios: \"learning on noisy labels\" and \"learning on query-response pairs with perturbed posteriors\". It is found that the two classification risks differ only by an instance-wise weight coefficient multiplied to the latter. Let $\\tilde{y} \\in \\mathbf{Y}$ denote the noisy label, where $\\tilde{y}$ is selected based on the transition matrix and the true label. Theorem 1 formalizes the defensive mechanism.", + "bbox": [ + 75, + 517, + 470, + 652 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Theorem 1. Let $\\mathcal{L}(\\cdot, \\cdot)$ denote the classification loss function. With sufficiently many queries, the classification risk $\\Re(h)$ of learning with noisy labels is $\\mathbb{E}_{(\\boldsymbol{x}, \\tilde{\\boldsymbol{y}})}[\\mathcal{L}(h(\\boldsymbol{x}), \\tilde{\\boldsymbol{y}})]$ . The classification risk $\\Re_T(h)$ of learning with query-response pairs is $\\mathbb{E}_{(\\boldsymbol{x}, \\tilde{\\boldsymbol{y}})}[q(\\boldsymbol{x}, \\tilde{\\boldsymbol{y}}) \\cdot \\mathcal{L}(h(\\boldsymbol{x}), \\tilde{\\boldsymbol{y}})]$ , where $q(\\boldsymbol{x}, \\tilde{\\boldsymbol{y}})$ denotes the instance-wise weight, proportional to the closeness between the predicted posterior and the true label.", + "bbox": [ + 75, + 662, + 470, + 770 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The proof can be found in Appendix 8. Theorem 1 demonstrates that training a clone model on such query-response pairs with perturbed posteriors is equivalent to training it on noisy-labeled data with instance-wise weights. It can be further deduced that the more closely the victim model's predicted posteriors match the true label, the more this training process resembles learning with noisy labels. In addition, the clone model's generalization can be further degraded by", + "bbox": [ + 75, + 779, + 472, + 902 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/e62144276b3f8110e06bbcbef99d4c3cab0b71c7c0bfc522368791929c71f5d9.jpg", + "image_caption": [ + "Figure 2. An illustration of \"Optimizing phase\" and \"Defense phase\" in the framework. \"Optimizing phase\" demonstrates the process of updating the noise transition matrix $T$ and surrogate network weights $\\theta_g$ alternatively. \"Defensive phase\" shows the interaction process where the attacker tries to steal the functionality from the defender." + ], + "image_footnote": [], + "bbox": [ + 501, + 90, + 890, + 333 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "increasing the transition matrix noise rate or exploring more difficult noise type through inter-class correlations, such as using class-conditional noise instead of random noise.", + "bbox": [ + 496, + 454, + 890, + 500 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.2. The overall training framework", + "text_level": 1, + "bbox": [ + 498, + 508, + 779, + 523 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Since the attacker's network is unknown, we cannot directly measure its cloning performance. Instead, we represent the attacker with a surrogate model $g$ parameterized by $\\theta_g$ and design optimal perturbations for it. By using the surrogate model to simulate the cloning process, we can continuously optimize the transition matrix during this process to minimize the utility trade-off between attacker performance and defender utility. To achieve this, we first introduce the respective objective functions for the attacker and defender. Then, a bi-level optimization approach is proposed.", + "bbox": [ + 496, + 531, + 893, + 681 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Given the perturbed posterior $f(\\pmb{x}; \\pmb{\\theta}_f)T$ , the attacker's objective function is defined as the loss on the gathered query-response pairs, formulated as:", + "bbox": [ + 496, + 681, + 890, + 728 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {a t t a c k}} \\left(\\boldsymbol {\\theta} _ {g}, T\\right) = \\mathbb {E} _ {\\boldsymbol {x} \\sim \\mathcal {D} _ {\\text {p a i r}}} \\left[ \\mathcal {L} _ {\\text {k n a}} \\left(g (\\boldsymbol {x}; \\boldsymbol {\\theta} _ {g}), f (\\boldsymbol {x}; \\boldsymbol {\\theta} _ {f}) T\\right) \\right]. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 511, + 739, + 893, + 757 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Notably, the optimization of $\\theta_{g}$ is influenced by $T$ . To defend against model stealing, we would like $T$ to regularize $\\theta_{g}$ to have poor generalizability. To achieve this, we instantiate a misinformation loss that inhibits the attacker model's outputs on the victim model's original predicted label $k = \\arg \\max_{i} f(\\pmb{x}, \\pmb{\\theta}_{f})_{i}$ . Formally, we derive the following modified negative log-likelihood loss:", + "bbox": [ + 496, + 768, + 893, + 875 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\operatorname {m i s}} \\left(\\boldsymbol {\\theta} _ {g}\\right) = \\mathbb {E} _ {\\boldsymbol {x} \\sim \\mathcal {D} _ {\\text {p a i r}}} \\left[ - \\log \\left(1 - g (\\boldsymbol {x}; \\boldsymbol {\\theta} _ {g}) _ {k}\\right) \\right]. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 547, + 883, + 892, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "24308", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The underlying principle is quite simple: The log-likelihood of the original prediction label on the output should be minimized. Despite its simple form, optimizing this loss has been empirically shown to be an effective defense against model stealing attacks [20, 59].", + "bbox": [ + 75, + 90, + 468, + 166 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Moreover, to maintain fidelity for the victim model, a straightforward strategy is aligning the perturbed output with the predicted label $k$ . This is implemented by minimizing the divergence between the output and the predicted label, formulated as:", + "bbox": [ + 75, + 167, + 470, + 239 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {a l i g n}} (T) = \\mathbb {E} _ {\\boldsymbol {x} \\sim \\mathcal {D} _ {\\text {p a i r}}} [ - \\log (f (\\boldsymbol {x}; \\boldsymbol {\\theta} _ {f}) T) _ {k} ]. \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 132, + 253, + 468, + 271 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In summary, the overall objective function of the defender can be represented as the following defense loss:", + "bbox": [ + 76, + 281, + 468, + 311 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {d e f e n s e}} \\left(\\boldsymbol {\\theta} _ {g}, T\\right) = \\lambda \\cdot \\mathcal {L} _ {\\text {a l i g n}} (T) + (1 - \\lambda) \\cdot \\mathcal {L} _ {\\text {m i s}} \\left(\\boldsymbol {\\theta} _ {g}\\right), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 96, + 323, + 468, + 340 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where the multiplicative factor $\\lambda$ is used to balance the contribution of these two loss terms.", + "bbox": [ + 75, + 352, + 468, + 381 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Bi-level optimization. Given the defense loss $\\mathcal{L}_{\\mathrm{defense}}$ , the surrogate model weights $\\theta_{g}$ need to be updated first before optimizing the transition matrix $T$ . For a mini-batch sample, we propose that surrogate model weights could be treated as a latent variable and optimized alternatively with the matrix. This yields the following bi-level optimization problem:", + "bbox": [ + 75, + 387, + 470, + 477 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\underset {T} {\\arg \\min } \\mathcal {L} _ {\\text {d e f s e n s e}} \\left(\\boldsymbol {\\theta} _ {g} ^ {*}, T\\right)\n$$\n", + "text_format": "latex", + "bbox": [ + 143, + 488, + 302, + 513 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\text {s u b j e c t} \\theta_ {g} ^ {*} = \\underset {\\theta_ {g}} {\\arg \\min } \\mathcal {L} _ {\\text {a t t a c k}} \\left(\\theta_ {g}, T\\right) \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 143, + 510, + 468, + 542 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Bi-level optimization problem has been proved strongly NP-hard [14], so getting such an exact solution for problem (7) is impossible in polynomial time. Fortunately, the inner optimization problem in (7) can be transformed into an optimized result $\\theta_g(T)$ as a function of $T$ . Then $\\theta_g(T)$ can then substitute $\\theta_g$ in the outer optimization to acquire an objective solely in terms of $T$ . This allows approximating the solution using gradient descent. Therefore, we employ an alternating optimization strategy to approximately solve this problem. In each iteration $\\theta_g$ and $T$ are updated on-the-fly in an alternating manner.", + "bbox": [ + 75, + 553, + 470, + 719 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Optimizing the transition matrix. In step $t + 1$ of optimizing $T^t$ , guided by the attack loss, surrogate model weights $\\theta_g^t$ are updated in a one-step-forward manner to obtain $\\theta_g^{t + 1}$ . Specifically, with the fixed $T^t$ , then $\\theta_g^t$ is updated via gradient descent with learning rate $\\alpha$ as:", + "bbox": [ + 75, + 724, + 470, + 800 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {\\boldsymbol {\\theta}} _ {g} ^ {t + 1} \\left(T ^ {t}\\right) = \\boldsymbol {\\theta} _ {g} ^ {t} - \\alpha \\cdot \\bigtriangledown_ {\\boldsymbol {\\theta} _ {g} ^ {t}} \\mathcal {L} _ {\\text {a t t a c k}}. \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 163, + 810, + 468, + 830 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Note that the surrogate model $g$ does not actually update in this step, but makes preparations for estimating $T^{t + 1}$ . We then optimize $T^t$ with $\\tilde{\\theta}_g^{t + 1}$ fixed under the defense loss. The motivation is that we would like to find a $T^{t + 1}$ that achieves", + "bbox": [ + 75, + 839, + 468, + 898 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Algorithm 1 The overall training procedure", + "text_level": 1, + "bbox": [ + 500, + 90, + 790, + 106 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Input: Query dataset $\\mathcal{D}_{\\mathrm{query}}$", + "bbox": [ + 500, + 109, + 692, + 126 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The surrogate model $g$ and its parameters $\\theta_{g}$", + "bbox": [ + 549, + 126, + 848, + 141 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The noise transition matrix $T$ .", + "bbox": [ + 549, + 141, + 750, + 154 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The number of iterations $I$ ;", + "bbox": [ + 549, + 156, + 732, + 169 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The multiplicative factor $\\lambda$ ;", + "bbox": [ + 549, + 171, + 732, + 184 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Procedure:", + "text_level": 1, + "bbox": [ + 500, + 186, + 581, + 198 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1: Initialize $T$ and $\\theta_{g}$", + "2: for $i = 1$ to $I$ do", + "3: Fetch a random batch $\\mathcal{B}$ from $\\mathcal{D}_{\\mathrm{query}}$", + "4: Calculate the one-step-forward surrogate model weights via Eq. (8);", + "5: Update the transition matrix via Eq. (9-12);", + "6: Update surrogate model weights via Eq. (13);", + "7: end for" + ], + "bbox": [ + 509, + 203, + 890, + 321 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Output: Learned noise transition matrix.", + "bbox": [ + 500, + 325, + 772, + 340 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "minimal utility trade-offs between the defender and attacker. The transition matrix $T^{t}$ is updated via gradient descent with learning rate $\\beta$ :", + "bbox": [ + 498, + 369, + 893, + 414 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nw ^ {t + 1} = T ^ {t} - \\beta \\cdot \\bigtriangledown_ {T ^ {t}} \\mathcal {L} _ {\\text {D e f s e n s e}}. \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 596, + 422, + 890, + 440 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Apply chain rule to Eq. (9), we can get:", + "bbox": [ + 500, + 450, + 763, + 465 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} w ^ {t + 1} = T ^ {t} - \\beta \\cdot \\left\\{\\lambda \\cdot \\nabla_ {T ^ {t}} \\mathcal {L} _ {\\text {a l i g n}} \\right. \\\\ + (1 - \\lambda) \\cdot \\bigtriangledown_ {\\tilde {\\boldsymbol {\\theta}} _ {g} ^ {t + 1}} \\mathcal {L} _ {\\mathrm {m i s}} \\cdot \\left(- \\alpha \\bigtriangledown_ {\\boldsymbol {\\theta} _ {g} ^ {t}, T ^ {t}} ^ {2} \\mathcal {L} _ {\\text {a t t a c k}}\\right) \\}. \\tag {10} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 514, + 470, + 890, + 513 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Note that $w^{t + 1}$ is not the final transition matrix, as its entries may not satisfy the properties of a valid transition matrix. Specifically, all elements must be non-negative and each row must be normalized. To address this, we first replace any negative elements in $w^{t + 1}$ with zeros via:", + "bbox": [ + 498, + 521, + 893, + 597 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {w} ^ {t + 1} = \\max \\left(w ^ {t + 1}, 0\\right). \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 616, + 604, + 890, + 622 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Then, each row is normalized to get the final valid transition matrix. For the $j^{\\mathrm{th}}$ row of $\\tilde{w}^{t + 1}$ , which indicates the probabilities of class $j$ transitioning to other classes, a normalization is performed to obtain the final matrix $T_{j,i}^{t + 1}$ :", + "bbox": [ + 498, + 631, + 893, + 695 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nT _ {j, i} ^ {t + 1} = \\frac {\\tilde {w} _ {j , i} ^ {t + 1}}{\\sum_ {i} \\tilde {w} _ {j , i} ^ {t + 1} + \\delta \\left(\\sum_ {i} \\tilde {w} _ {j , i} ^ {t + 1}\\right)}, \\delta (a) = \\left\\{ \\begin{array}{l} 1, \\text {i f} a = 0 \\\\ 0, \\text {i f} a \\neq 0 \\end{array} \\right. \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 509, + 702, + 890, + 753 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\delta (\\cdot)$ is used to avoid division by 0.", + "bbox": [ + 500, + 755, + 769, + 771 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Updating surrogate model weights. After optimizing the matrix, we conduct the surrogate model update by keeping $T^{t + 1}$ fixed and minimizing the attack loss. New surrogate model weights $\\theta_{g}^{t + 1}$ are obtained via gradient descent with learning rate $\\gamma$ :", + "bbox": [ + 500, + 775, + 890, + 851 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {\\theta} _ {g} ^ {t + 1} = \\boldsymbol {\\theta} _ {g} ^ {t} - \\gamma \\cdot \\bigtriangledown \\boldsymbol {\\theta} _ {g} ^ {t} \\mathcal {L} _ {\\text {a t t a c k}}. \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 602, + 859, + 890, + 878 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The whole training procedure is summarized in Algorithm 1.", + "bbox": [ + 500, + 885, + 893, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "24309", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/f05755d3e179c3574b171dad9ce29b1ed641c449d5f783f723824d0e3b696da2.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodCIFAR100 → CIFAR10CIFAR10 → CIFAR100Caltech256 → CUB200
ΔAcc=1ΔAcc=2ΔAcc=5ΔAcc=1ΔAcc=2ΔAcc=5ΔAcc=1ΔAcc=2ΔAcc=5
Random84.1777.2174.2143.2339.8132.1233.5031.1428.01
AM77.1072.9357.5639.8232.9819.3626.1721.2714.31
MAD80.5777.5469.0736.5632.5926.6537.1635.2331.50
NASTY76.6574.3164.1938.5333.6126.9531.0628.2324.70
GARD²77.4273.8871.4041.2139.2935.9435.6933.6732.34
MIN-IP78.8873.4070.6642.3338.6734.2634.7332.2628.94
EMMA61.7952.9142.2128.0118.3413.3719.1015.7913.11
ImageNet-C10 → CIFAR10ImageNet-C100 → CIFAR100ImageNet-CUB200 → CUB200
ΔAcc=1ΔAcc=2ΔAcc=5ΔAcc=1ΔAcc=2ΔAcc=5ΔAcc=1ΔAcc=2ΔAcc=5
Random88.9988.1887.0763.1661.7658.6446.5844.0640.49
AM87.5584.6076.7260.5756.0947.1045.1239.6530.15
MAD88.6785.7078.4458.7456.2350.8550.9848.5744.47
NASTY87.3385.3381.6859.7956.0548.9546.5643.2337.76
GARD²87.2986.0685.9358.3956.7153.6350.2747.5543.11
MIN-IP87.4086.8184.4561.1756.3551.9850.0946.4542.06
EMMA84.1080.6077.5256.7349.1831.4836.3533.1026.33
", + "bbox": [ + 88, + 88, + 883, + 381 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/977833f155ce48897c67d0303f5dd5caed4798f1b8c08de9a202f444e62538b8.jpg", + "table_caption": [ + "Table 1. Accuracy (%) of attacker model on defender's test set under knowledge-limited (top row) and distribution-aware (bottom row) attack scenarios with varying budgets. The best result in each column is highlighted in bold." + ], + "table_footnote": [], + "table_body": "
Eval DataDefender AccuracyAttacker Accuracy
KLDA
CIFAR1095.33 ± 0.1385.17 ± 0.8391.78 ± 0.19
CIFAR10075.96 ± 0.0951.95 ± 1.1266.82 ± 0.16
CUB20081.44 ± 0.4648.93 ± 0.3762.12 ± 0.22
", + "bbox": [ + 78, + 428, + 468, + 511 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 2. Accuracy $(\\%)$ of the defender model and the attacker model with no defense applied.", + "bbox": [ + 75, + 513, + 470, + 544 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 76, + 556, + 209, + 573 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To demonstrate the effectiveness of our proposed defense method, we conduct experiments on multiple image datasets and present the results in this section.", + "bbox": [ + 75, + 582, + 468, + 628 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1. Experiment Setup", + "text_level": 1, + "bbox": [ + 76, + 641, + 253, + 657 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Datasets. We employ three datasets for evaluation: CIFAR10, CIFAR100 [26] and CUB200 [57]. For each dataset, we explore two attacker scenarios - knowledge-limited (KL) and distribution-aware (DA). The KL scenario uses query sets of CIFAR100, CIFAR10, and Caltech256 [8], paired with CIFAR10, CIFAR100, and CUB200 evaluation datasets respectively. The DA scenario employs tailored query sets from ImageNet-1K [4] manually selected by prior work [38], specifically ImageNet-C10, ImageNet-C100, and ImageNet-CUB200 containing 183,763, 161,653, and 30,000 examples matched to the evaluation datasets, with their further details provided in the Appendix 9.", + "bbox": [ + 75, + 666, + 470, + 848 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Compared methods. We compare our method against the following six model stealing defense baselines: (1) No Defense, a naive approach where no defense is attempted. (2)", + "bbox": [ + 73, + 854, + 470, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Random, an interpolation-based method that perturbs the defender's predicted posterior by interpolating it with a one-hot posterior vector, where the index of single non-zero entry is randomly selected from the non-argmax prediction labels. (3) AM [20], a detection-based method that identifies out-of-distribution queries and interpolates with posteriors from a misinformation network, where the interpolation weight is denoted by the confidence of the argmax prediction label. (4) MAD [40], an accuracy-constrained method that adds controlled noise to original posteriors so that it maximizes the angular deviation between the perturbed and original gradient signal on a surrogate model. The surrogate model is randomly initialized. (5) NASTY [34], a self-undermining knowledge distillation approach which trains a nasty teacher network by maximizing the difference between the output of the nasty teacher and a normal pretrained network. (5) $GARD^2$ [38], a coordinated gradient redirection defense method, which perturbs original posteriors by setting the target gradient direction to be the all-ones vector. This redirects the attacker to update towards a consistent target direction. (6) MIN-IP [38], an uncoordinated gradient redirection defense method, which perturbs original posteriors by setting the target gradient direction opposite the original gradient direction, redirecting the update gradient of attackers opposite to the original gradients. The surrogate model of MIN-IP and $GARD^2$ train on the attacker's queries with early stopping after 10 epochs.", + "bbox": [ + 496, + 426, + 893, + 833 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For fair comparisons, we basically use the same network architecture, learning rate, optimizer, and augmentation strategy across all the comparing methods.", + "bbox": [ + 496, + 834, + 893, + 880 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation. Our experiment pipeline has three stages:", + "bbox": [ + 500, + 885, + 893, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "24310", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/f22df67711c7e2a03a84d40efbea92433dd69078620db61350f83067e53fae56.jpg", + "image_caption": [ + "(a) Heatmap of transition matrix." + ], + "image_footnote": [], + "bbox": [ + 84, + 88, + 261, + 207 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/1179097a99be1a5469ca28c057017ddde9033864b7c0a9b836c7feb0b58dcae8.jpg", + "image_caption": [ + "(b) Heatmap of transition matrix." + ], + "image_footnote": [], + "bbox": [ + 277, + 88, + 459, + 207 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c9df9a5cb987d1faf76f39ed786414a0649152acfc0e99f4e72ff535f22715be.jpg", + "image_caption": [ + "(c) Average predicted posterior" + ], + "image_footnote": [], + "bbox": [ + 467, + 90, + 661, + 207 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/271bc120cbe6b8c77faeb5962cc6cddf7a388916675e414315814784e1d0fdcf.jpg", + "image_caption": [ + "(d) Average perturbed posterior" + ], + "image_footnote": [], + "bbox": [ + 681, + 90, + 875, + 207 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d1373fed305d5394c88c3e75fbe04b70003b7b716967e5b52f80a81f39625b03.jpg", + "image_caption": [ + "Figure 3. Qualitative analysis of the transition matrix. (a-b): Heatmaps of optimized transition matrixes for distribution-aware attacks (a) Imagenet-C10→CIFAR10 and (b) Imagenet-C100→CIFAR100, respectively. (c-d): Average (c) predicted and (d) perturbed posteriors across the query set of (top) Imagenet-C10 and (bottom) Imagenet-C100 datasets.", + "Figure 4. Visualizations of predicted and perturbed posteriors." + ], + "image_footnote": [], + "bbox": [ + 91, + 286, + 455, + 435 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "First, the defender model trains on each evaluation dataset. Second, defense methods generate perturbed posteriors for query dataset and evaluation dataset at various defense budgets. Finally, the attacker model trains on the collected query-response pairs. We denote attacks with query data $\\mathcal{Q}$ and evaluation data $\\mathcal{D}$ as $\\mathcal{Q} \\to \\mathcal{D}$ .", + "bbox": [ + 75, + 474, + 468, + 565 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We use ResNet34 for CIFAR10/100 and ResNet50 pretrained on ImageNet for CUB200. For all models training, we utilize SGD as the optimizer with a momentum of 0.9 and weight decay of 5e-4 for 100 epochs. Initial learning rates are set as 0.01 for CUB200 and 0.1 for other datasets, with a cosine annealing schedule. For optimizing the transition matrix, the learning rate $\\gamma$ is set as 0.3 for CIFAR10 and 1.0 for other datasets. The surrogate model's one-step-forward learning rate $\\alpha$ and the actual learning rate $\\beta$ are set the same value as specified in the model training settings. Additionally, an early stopping mechanism is set to terminates optimization when the average change in $T$ between consecutive epochs falls below a specified threshold. The threshold is 1e-2 for CIFAR10 and 1e-4 for other datasets. Furthermore, the balancing factor $\\lambda$ is adjusted to achieve different utility budgets. Recent work [38] proposed a novel approach to initialize the surrogate model. They suggested to train the surrogate model through distilling from the victim model with early stopping. This would make it more representative of the actual attacker's behavior. Following this approach, we initialize our surrogate model by training it on the query dataset with the same early stopping mechanism.", + "bbox": [ + 75, + 568, + 472, + 902 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.2. Experiment Results", + "text_level": 1, + "bbox": [ + 500, + 287, + 687, + 303 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 2 presents the accuracy of the defender model and the attacker model, without any defense applied under two different query distributions. This shows the upper bound for the usage performance of the defender and stealable performance of the attacker. Table 1 compares the accuracy of an attacker's clone model at different defender accuracy drop budgets $\\Delta$ Acc. The budget refers to the average decrease in accuracy the defender causes across the test set by adding perturbations. Corresponding visualizations of these results are provided in the Appendix 10. As shown, our defense method achieves the lowest utility trade-off across budgets. For example, in the CIFAR100 $\\rightarrow$ CIFAR10 attack scenario, our method reduces the attacker model's accuracy by $23\\%$ while only decreasing the defender's accuracy by $1\\%$ .", + "bbox": [ + 496, + 311, + 893, + 523 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.3. Qualitative Analysis", + "text_level": 1, + "bbox": [ + 500, + 536, + 691, + 551 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Perturbed posteriors make cloning difficult. We visualize the predicted and perturbed posteriors using t-SNE in Figure 4. The perturbed posteriors clearly exhibit degraded inter-class distances compared to the predicted posteriors. This degradation obfuscates the decision boundaries between classes, making it difficult for the attacker to clone the defender model by imitating the posteriors. Notably, such effective defenses are achieved solely through linear mapping of the posteriors, demonstrating the simplicity and efficiency of our proposed noise perturbation strategy.", + "bbox": [ + 496, + 560, + 893, + 712 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The transition matrix induces challenging noise. We analyze the defense details of two attack scenarios: Imagenet-C10→CIFAR10 and Imagenet-C100→CIFAR100 in Figure 3. Specifically, Figure 3 (a-b) visualizes corresponding optimized transition matrices. It can be observed that the diagonal elements are maximized for fidelity, while latent inter-label relationships are also exploited. As shown the optimized matrix in Figure 3 (a), despite it causes a $2.35\\%$ drop in the defender's accuracy on the test set, the optimization makes it have a noise rate as high as $77\\%$ . This negatively affects the cloning process of the attacker, reducing the cloned model's accuracy by $14\\%$ .", + "bbox": [ + 496, + 719, + 895, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "24311", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In Figures 3 (c-d), we visualize the average posterior of predicted and perturbed predictions. Given that ImageNet-C10/C100 follow long-tail distributions as illustrated in appendix 9, the predicted posteriors also exhibit long-tailed shapes as expected. However, the perturbations significantly increase the confidence of tail classes, distorting the true inter-class relationships. This distortion of class relationships also makes stealing more difficult for the attacker.", + "bbox": [ + 75, + 90, + 472, + 212 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.4. Ablation Study", + "text_level": 1, + "bbox": [ + 76, + 223, + 230, + 239 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effectiveness of optimized transition matrix. To validate the effectiveness of the optimized transition matrix, we compare the defense performance between it and the uniform transition matrix with an identical noise rate. The results in Table 3 show that the optimized transition matrix can further reduce the attacker's accuracy while maintaining the defender's accuracy. This demonstrates that the optimized matrix provides improved defense over the uniform matrix. Overall, this shows that capturing the inter-class relationships during optimization contributes to generating a matrix that is more adversarial against the attacker.", + "bbox": [ + 75, + 246, + 473, + 412 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/7b6234df0fda0c1a7a0bf5ff0856c9a6691c79776081dea9e7e05979aa21db03.jpg", + "table_caption": [], + "table_footnote": [ + "Table 3. Defender and attacker accuracy $(\\%)$ with optimized and uniform transition matrix under attack scenarios for CIFAR10, CIFAR100 and CUB200 evaluation datasets." + ], + "table_body": "
Attack ScenariosOptimized matrixUniform matrix
DefenderAttackerDefenderAttacker
CIFAR100→CIFAR1093.87±0.1432.41±7.5494.81±0.0671.43±2.26
CIFAR10→CIFAR10074.48±0.2515.65±0.7274.75±0.0735.67±3.44
Caltech256→CUB20080.45±0.1318.25±0.4380.89±0.2230.15±0.21
", + "bbox": [ + 78, + 425, + 468, + 498 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effects of the number of queries. We also verify the effectiveness of our defense against attacks with varying number of queries in Figure 5. The EMMA-top1 result is depicted as a reference, which denotes the victim model only outputting the predicted label rather than the full posterior. As formalized in Theorem 1, when the victim model assigns full confidence to the true label, the attacker's cloning task degrades to learning from noisy labels. Thus, EMMA-top1 signifies the optimal defense under the same transition matrix. EMMA defense consistently forces degradation of the attacker accuracy across all query set sizes, remaining close to the EMMA-top1. As the query increases, the attacker model obtained from the victim model suffers more accuracy drops, revealing that our defense induces more significant detrimental impacts when the attacker uses more queries.", + "bbox": [ + 75, + 561, + 470, + 789 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Transferability across attack scenarios. To explore the transition matrix's transferable defense capabilities, we test it against an attack type different from the one it was optimized for. As shown in Figure 6, we use the matrix optimized under DA attack to defend against KL attack, denoted as \"DA $\\rightarrow$ KL\". Compared to a matrix specifically optimized for KL attack, denoted as \"KL $\\rightarrow$ KL\", it still maintained", + "bbox": [ + 75, + 795, + 470, + 900 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/ed201633e1a9d3cebbb1470398daf17ce8a00e5ea90ce7abce80b21605326866.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 506, + 90, + 692, + 184 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/9028e23d26bd00004128021321af4a80f66f06940570ffcbe45544bf452a7585.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 696, + 90, + 883, + 184 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/b3db8ce82356725ea57631298e973ca3391674ef2bbc731709eec883175269c5.jpg", + "image_caption": [ + "Figure 5. Effects of the number of queries." + ], + "image_footnote": [ + "No defense EMMA EMMA-top1" + ], + "bbox": [ + 504, + 188, + 692, + 280 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/68222c69d426003e06e4838b38ac9bc999b9b7b32a9a33692a19320a5d9c0b34.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 696, + 188, + 883, + 280 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "considerable defensive efficacy, exhibiting the optimized transition matrix's robust transferable defense capabilities across different attack types.", + "bbox": [ + 496, + 332, + 890, + 378 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/293ba659bc74a9176cb97b1b15b3f5c9abe3d9f56361f9a94d345cd9f17369d0.jpg", + "image_caption": [ + "Figure 6. Transferability across attack scenarios." + ], + "image_footnote": [], + "bbox": [ + 519, + 388, + 696, + 523 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/7709f982bd19a6cdf67251b1665ad488ccef7430bbad1b683ea188d84c45cc6c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 702, + 388, + 867, + 523 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Due to the space limitation, additional experiments of our method are provided in Appendix.", + "bbox": [ + 500, + 554, + 890, + 584 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 598, + 617, + 614 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this work, we revisited the noise transition matrix in a novel way by applying it to model stealing defense for the first time. This deteriorates the attacker's cloning process to a level equivalent to training on noisy labels, hampering attacker model generalization. To optimize the transition matrix, we proposed a novel bi-level optimization training framework that incorporates fidelity training for the victim model and adversarial training for the surrogate model. Experimental results under various attack scenarios showed the effectiveness and robustness of our proposed defense method compared to current state-of-the-art methods, achieving strong defense with minimal utility trade-off.", + "bbox": [ + 496, + 623, + 893, + 805 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "7. Acknowledgments", + "text_level": 1, + "bbox": [ + 500, + 819, + 678, + 835 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The authors wish to thank the anonymous reviewers for their helpful comments and suggestions. This work was supported by the National Science Foundation of China (62225602), and the Big Data Computing Center of Southeast University.", + "bbox": [ + 496, + 844, + 893, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "24312", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 118, + 174, + 133 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Yossi Adi, Carsten Baum, Moustapha Cisse, Benny Pinkas, and Joseph Keshet. Turning your weakness into a strength: Watermarking deep neural networks by backdoorsing. In USENIX Security Symposium, pages 1615-1631, 2018. 3", + "[2] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. In International Conference on Advanced Neural Information Processing Systems, pages 1877-1901, 2020. 1", + "[3] Yanjiao Chen, Rui Guan, Xueluan Gong, Jianshuo Dong, and Meng Xue. D-dae: Defense-penetrating model extraction attacks. In IEEE Symposium on Security and Privacy, pages 382-399, 2023. 1", + "[4] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In IEEE/CVF Computer Vision and Pattern Recognition Conference, pages 248-255, 2009. 6", + "[5] Adam Dziedzic, Muhammad Ahmad Kaleem, Yu Shen Lu, and Nicolas Papernot. Increasing the cost of model extraction with calibrated proof of work. In International Conference on Learning Representations, 2022. 1", + "[6] Matt Fredrikson, Somesh Jha, and Thomas Ristenpart. Model inversion attacks that exploit confidence information and basic countermeasures. In ACM SIGSAC conference on computer and communications security, pages 1322-1333, 2015. 1", + "[7] Karan Ganju, Qi Wang, Wei Yang, Carl A Gunter, and Nikita Borisov. Property inference attacks on fully connected neural networks using permutation invariant representations. In ACM SIGSAC conference on computer and communications security, pages 619-633, 2018. 1", + "[8] Gregory Griffin, Alex Holub, and Pietro Perona. Caltech-256 object category dataset. 2007. 6", + "[9] Jun Guo, Aishan Liu, Xingyu Zheng, Siyuan Liang, Yisong Xiao, Yichao Wu, and Xianglong Liu. Isolation and induction: Training robust deep neural networks against model stealing attacks. In ACM International Conference on Multimedia, 2023. 1", + "[10] Bo Han, Quanming Yao, Tongliang Liu, Gang Niu, Ivor W Tsang, James T Kwok, and Masashi Sugiyama. A survey of label-noise representation learning: Past, present and future. arXiv preprint arXiv:2011.04406, 2020. 3", + "[11] Bo Han, Quanming Yao, Xingrui Yu, Gang Niu, Miao Xu, Weihua Hu, Ivor W. Tsang, and Masashi Sugiyama. Robust training of deep neural networks with extremely noisy labels. In International Conference on Advanced Neural Information Processing Systems, page 4, 2020. 2", + "[12] Geoffrey Hinton, Oriol Vinyals, and Jeff Dean. Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531, 2015. 2, 3", + "[13] Surgan Jandial, Yash Khasbage, Arghya Pal, Vineeth N Balasubramanian, and Balaji Krishnamurthy. Distilling the undistillable: Learning from a nasty teacher. In European Conference on Computer Vision, pages 587-603, 2022. 2" + ], + "bbox": [ + 78, + 143, + 470, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[14] Robert G Jeroslow. The polynomial hierarchy and a simple model for competitive analysis. Mathematical programming, pages 146-164, 1985. 5", + "[15] Hengrui Jia, Christopher A Choquette-Choo, Varun Chandrasekaran, and Nicolas Papernot. Entangled watermarks as a defense against model extraction. In USENIX Security Symposium, pages 1937-1954, 2021. 2, 3", + "[16] Hengrui Jia, Mohammad Yaghini, Christopher A Choquette-Choo, Natalie Dullerud, Anvith Thudi, Varun Chandrasekaran, and Nicolas Papernot. Proof-of-learning: Definitions and practice. In IEEE Symposium on Security and Privacy, pages 1039-1056, 2021. 3", + "[17] Wenbo Jiang, Hongwei Li, Guowen Xu, Tianwei Zhang, and Rongxing Lu. A comprehensive defense framework against model extraction attacks. IEEE Transactions on Dependable and Secure Computing, 2023. 1", + "[18] Ishan Jindal, Matthew Nokleby, and Xuewen Chen. Learning deep networks from noisy labels with dropout regularization. In IEEE International Conference on Data Mining, pages 967-972, 2016. 3", + "[19] Mika Juuti, Sebastian Szyller, Samuel Marchal, and N Asokan. Prada: protecting against dnn model stealing attacks. In IEEE European Symposium on Security and Privacy, pages 512-527, 2019. 1, 2, 3", + "[20] Sanjay Kariyappa and Moinuddin K Qureshi. Defending against model stealing attacks with adaptive misinformation. In IEEE/CVF Computer Vision and Pattern Recognition Conference, pages 770-778, 2020. 1, 2, 3, 5, 6", + "[21] Sanjay Kariyappa, Atul Prakash, and Moinuddin K Qureshi. Protecting dnns from theft using an ensemble of diverse models. In International Conference on Learning Representations, 2020. 2", + "[22] Sanjay Kariyappa, Atul Prakash, and Moinuddin K Qureshi. Maze: Data-free model stealing attack using zeroth-order gradient estimation. In CVPR, 2021. 1, 2", + "[23] Manish Kesarwani, Bhaskar Mukholy, Vijay Arya, and Sameep Mehta. Model extraction warning in mlaas paradigm. In Annual Computer Security Applications Conference, pages 371-380, 2018. 1", + "[24] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In International Conference on Computer Vision, 2023. 1", + "[25] Kalpesh Krishna, Gaurav Singh Tomar, Ankur P Parikh, Nicolas Papernot, and Mohit Iyyer. Thieves on sesame street! model extraction of bert-based apis. In International Conference on Learning Representations, 2020. 2", + "[26] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009. 6", + "[27] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. ImageNet classification with deep convolutional neural networks. In International Conference on Advanced Neural Information Processing Systems, pages 1106-1114, 2012. 1", + "[28] Souvik Kundu, Qirui Sun, Yao Fu, Massoud Pedram, and Peter Beerel. Analyzing the confidentiality of undistinguishable teachers in knowledge distillation. pages 9181-9192, 2021. 2" + ], + "bbox": [ + 501, + 92, + 893, + 901 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "24313", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[29] Jeonghyun Lee, Sungmin Han, and Sangkyun Lee. Model stealing defense against exploiting information leak through the interpretation of deep neural nets. In International Joint Conference on Artificial Intelligence, pages 710-716, 2022. 1, 3", + "[30] Taesung Lee, Benjamin Edwards, Ian Molloy, and Dong Su. Defending against neural network model stealing attacks using deceptive perturbations. In IEEE Security and Privacy Workshops, pages 43-49, 2019. 2, 3", + "[31] Simin Li, Shuning Zhang, Gujun Chen, Dong Wang, Pu Feng, Jiakai Wang, Aishan Liu, Xin Yi, and Xianglong Liu. Towards benchmarking and assessing visual naturalness of physical world adversarial attacks. In IEEE/CVF Computer Vision and Pattern Recognition Conference, pages 12324–12333, 2023. 1", + "[32] Yiming Li, Linghui Zhu, Xiaojun Jia, Yong Jiang, Shu-Tao Xia, and Xiaochun Cao. Defending against model stealing via verifying embedded external features. In Association for the Advance of Artificial Intelligence, pages 1464-1472, 2022. 3", + "[33] Shunchang Liu, Jiakai Wang, Aishan Liu, Yingwei Li, Yijie Gao, Xianglong Liu, and Dacheng Tao. Harnessing perceptual adversarial patches for crowd counting. In ACM SIGSAC Conference on Computer and Communications Security, pages 2055-2069, 2022. 1", + "[34] Haoyu Ma, Tianlong Chen, Ting-Kuei Hu, Chenyu You, Xiaohui Xie, and Zhangyang Wang. Undistillable: Making a nasty teacher that cannot teach students. In ICLR, 2021. 6", + "[35] Ke Ma, Qianqian Xu, Jinshan Zeng, Xiaochun Cao, and Qingming Huang. Poisoning attack against estimating from pairwise comparisons. IEEE Trans. Pattern Anal. Mach. Intell., pages 6393-6408, 2021. 1", + "[36] Ke Ma, Qianqian Xu, Jinshan Zeng, Guorong Li, Xiaochun Cao, and Qingming Huang. A tale of hodgerank and spectral method: Target attack against rank aggregation is the fixed point of adversarial game. IEEE Trans. Pattern Anal. Mach. Intell., pages 4090-4108, 2022. 1", + "[37] Pratyush Maini, Mohammad Yaghini, and Nicolas Papernot. Dataset inference: Ownership resolution in machine learning. In International Conference on Learning Representations, 2021. 3", + "[38] Mantas Mazeika, Bo Li, and David Forsyth. How to steer your adversary: Targeted and efficient model stealing defenses with gradient redirection. In International Conference on Machine Learning, pages 15241-15254, 2022. 2, 3, 6, 7", + "[39] Tribhuvanesh Orekondy, Bernt Schiele, and Mario Fritz. Knockoff nets: Stealing functionality of black-box models. In IEEE/CVF Computer Vision and Pattern Recognition Conference, pages 4954-4963, 2019. 1, 2, 3", + "[40] Tribhuvanesh Orekondy, Bernt Schiele, and Mario Fritz. Prediction poisoning: Towards defenses against dnn model stealing attacks. In International Conference on Learning Representations, 2020. 2, 3, 6", + "[41] Soham Pal, Yash Gupta, Aditya Shukla, Aditya Kanade, Shirish Shevade, and Vinod Ganapathy. Activethief: Model extraction using active learning and unannotated public data. In Association for the Advance of Artificial Intelligence, pages 865–872, 2020. 2" + ], + "bbox": [ + 78, + 90, + 470, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[42] Nicolas Papernot, Patrick McDaniel, Ian Goodfellow, Somesh Jha, Z Berkay Celik, and Ananthram Swami. Practical black-box attacks against machine learning. In ACM on Asia conference on computer and communications security, pages 506-519, 2017. 1, 2, 3", + "[43] Giorgio Patrini, Alessandro Rozza, Aditya Krishna Menon, Richard Nock, and Lizhen Qu. Making deep neural networks robust to label noise: A loss correction approach. In IEEE/CVF Computer Vision and Pattern Recognition Conference, pages 1944-1952, 2017. 2, 3", + "[44] Sunandini Sanyal, Sravanti Addepalli, and R Venkatesh Babu. Towards data-free model stealing in a hard label setting. In CVPR, 2022. 2", + "[45] Zeyang Sha et al. Can't steal? cont-steal! contrastive stealing attacks against image encoders. In CVPR, 2023. 2", + "[46] Yu Shi, Dong-Dong Wu, Xin Geng, and Min-Ling Zhang. Robust representation learning for unreliable partial label learning. arXiv preprint arXiv:2308.16718, 2023. 3", + "[47] Reza Shokri, Marco Stronati, Congzheng Song, and Vitaly Shmatikov. Membership inference attacks against machine learning models. In IEEE symposium on security and privacy, pages 3-18, 2017. 1", + "[48] Sainbayar Sukhbaatar, Joan Bruna, Manohar Paluri, Lubomir Bourdev, and Rob Fergus. Training convolutional networks with noisy labels. arXiv preprint arXiv:1406.2080, 2014. 3", + "[49] Christian Szegedy, Wojciech Zaremba, Ilya Sutskever, Joan Bruna, Dumitru Erhan, Ian Goodfellow, and Rob Fergus. Intriguing properties of neural networks. In International Conference on Learning Representations, 2013. 1", + "[50] Sebastian Szyller, Buse Gul Atli, Samuel Marchal, and N Asokan. Dawn: Dynamic adversarial watermarking of neural networks. In ACM International Conference on Multimedia, pages 4417-4425, 2021. 1, 3", + "[51] Florian Tramér, Fan Zhang, Ari Juels, Michael K Reiter, and Thomas Ristenpart. Stealing machine learning models via prediction {APIs}. In USENIX security symposium, pages 601-618, 2016. 1, 2, 3", + "[52] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In International Conference on Advanced Neural Information Processing Systems, 2017. 1", + "[53] M Mitchell Waldrop et al. No drivers required. Nature, page 20, 2015. 1", + "[54] Binghui Wang and Neil Zhenqiang Gong. Stealing hyperparameters in machine learning. In IEEE symposium on security and privacy, pages 36-52, 2018. 1", + "[55] Jiakai Wang, Aishan Liu, Zixin Yin, Shunchang Liu, Shiyu Tang, and Xianglong Liu. Dual attention suppression attack: Generate adversarial camouflage in physical world. In IEEE/CVF Computer Vision and Pattern Recognition Conference, pages 8565-8574, 2021. 1", + "[56] Zhen Wang, Guosheng Hu, and Qinghua Hu. Training noise-robust deep neural networks via meta-learning. In IEEE/CVF Computer Vision and Pattern Recognition Conference, pages 4524–4533, 2020. 3" + ], + "bbox": [ + 501, + 92, + 893, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "24314", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[57] Peter Welinder, Steve Branson, Takeshi Mita, Catherine Wah, Florian Schroff, Serge Belongie, and Pietro Perona. Caltech-ucsd birds 200. 2010. 6", + "[58] Hongwei Wen, Jingyi Cui, Hanyuan Hang, Jiabin Liu, Yisen Wang, and Zhouchen Lin. Leveraged weighted loss for partial label learning. In International Conference on Machine Learning, pages 11091-11100, 2021. 2", + "[59] Dong-Dong Wu, Deng-Bao Wang, and Min-Ling Zhang. Revisiting consistency regularization for deep partial label learning. In International Conference on Machine Learning, pages 24212-24225, 2022. 5", + "[60] Xiaobo Xia, Tongliang Liu, Nannan Wang, Bo Han, Chen Gong, Gang Niu, and Masashi Sugiyama. Are anchor points really indispensable in label-noise learning? In International Conference on Advanced Neural Information Processing Systems, pages 6835-6846, 2019. 3", + "[61] Hao Yang, You-Zhi Jin, Zi-Yin Li, Deng-Bao Wang, Xin Geng, and Min-Ling Zhang. Learning from noisy labels via dynamic loss thresholding. IEEE Transactions on Knowledge and Data Engineering, 2023. 3", + "[62] Honggang Yu, Kaichen Yang, Teng Zhang, Yun-Yun Tsai, Tsung-Yi Ho, and Yier Jin. Cloudleak: Large-scale deep learning models stealing through adversarial examples. In Network and Distributed System Security Symposium, 2020. 1", + "[63] Chiyuan Zhang, Samy Bengio, Moritz Hardt, Benjamin Recht, and Oriol Vinyals. Understanding deep learning (still) requires rethinking generalization. Communications of the ACM, pages 107-115, 2021. 2", + "[64] Jiliang Zhang, Shuang Peng, Yansong Gao, Zhi Zhang, and Qinghui Hong. Apmsa: adversarial perturbation against model stealing attacks. IEEE Transactions on Information Forensics and Security, pages 1667-1679, 2023. 1", + "[65] Guoqing Zheng, Ahmed Hassan Awadallah, and Susan Dumas. Meta label correction for noisy label learning. In Association for the Advance of Artificial Intelligence, pages 11053-11061, 2021. 3" + ], + "bbox": [ + 78, + 90, + 470, + 614 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "24315", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2024/Efficient Model Stealing Defense with Noise Transition Matrix/5181a0c8-f9af-47c9-b829-4e394817b1e7_model.json b/2024/Efficient Model Stealing Defense with Noise Transition Matrix/5181a0c8-f9af-47c9-b829-4e394817b1e7_model.json new file mode 100644 index 0000000000000000000000000000000000000000..3090c527328d4128a4c1aa7a8ada48f1dc3fdb67 --- /dev/null +++ b/2024/Efficient Model Stealing Defense with Noise Transition Matrix/5181a0c8-f9af-47c9-b829-4e394817b1e7_model.json @@ -0,0 +1,2631 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.172, + 0.131, + 0.8, + 0.152 + ], + "angle": 0, + "content": "Efficient Model Stealing Defense with Noise Transition Matrix" + }, + { + "type": "text", + "bbox": [ + 0.236, + 0.18, + 0.732, + 0.199 + ], + "angle": 0, + "content": "Dong-Dong \\(\\mathsf{W u}^{1}\\), Chilin \\(\\mathsf{F u}^{2}\\), Weichang \\(\\mathsf{W u}^{2}\\), Wenwen \\(\\mathsf{X i a}^{2}\\)" + }, + { + "type": "text", + "bbox": [ + 0.295, + 0.199, + 0.674, + 0.218 + ], + "angle": 0, + "content": "Xiaolu Zhang\\(^{2}\\), Jun Zhou\\(^{2*}\\), Min-Ling Zhang\\(^{1*}\\)" + }, + { + "type": "text", + "bbox": [ + 0.151, + 0.217, + 0.82, + 0.235 + ], + "angle": 0, + "content": "\\(^{1}\\)School of Computer Science and Engineering, Southeast University, Nanjing, China" + }, + { + "type": "text", + "bbox": [ + 0.438, + 0.233, + 0.533, + 0.253 + ], + "angle": 0, + "content": "\\(^{2}\\)Ant Group" + }, + { + "type": "text", + "bbox": [ + 0.125, + 0.255, + 0.845, + 0.27 + ], + "angle": 0, + "content": "{dongdongwu,zhangml}@seu.edu.cn,{chilin.fcl,jiuyue.wwc,wenzhen.xww}@antgroup.com" + }, + { + "type": "text", + "bbox": [ + 0.318, + 0.273, + 0.636, + 0.287 + ], + "angle": 0, + "content": "{yueyin.zxl, jun.zhoujun}@antfin.com" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.322, + 0.314, + 0.339 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.354, + 0.474, + 0.687 + ], + "angle": 0, + "content": "With the escalating complexity and investment cost of training deep neural networks, safeguarding them from unauthorized usage and intellectual property theft has become imperative. Especially the rampant misuse of prediction APIs to replicate models without access to the original data or architecture poses grave security threats. Diverse defense strategies have emerged to address these vulnerabilities, yet these defenses either incur heavy inference overheads or assume idealized attack scenarios. To address these challenges, we revisit the utilization of noise transition matrix as an efficient perturbation technique, which injects noise into predicted posteriors in a linear manner and integrates seamlessly into existing systems with minimal overhead, for model stealing defense. Provably, with such perturbed posteriors, the attacker's cloning process degrades into learning from noisy data. Toward optimizing the noise transition matrix, we proposed a novel bi-level optimization training framework, which performs fidelity on the victim model while the surrogate model adversarially. Comprehensive experimental results demonstrate that our method effectively thwarts model stealing attacks and achieves minimal utility trade-offs, outperforming existing state-of-the-art defenses." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.717, + 0.21, + 0.733 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.743, + 0.472, + 0.866 + ], + "angle": 0, + "content": "Machine learning (ML) models, especially deep neural networks (DNNs), have attained human-level capabilities across various domains, such as pattern recognition tasks [24, 27] and generation tasks [2, 52], becoming indispensable in facilitating convenience for society [53]. With the burgeoning of ML application scenarios and the proliferation of data, training large and sophisticated models and making them publicly accessible has become a new profit trend for compa" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.324, + 0.895, + 0.475 + ], + "angle": 0, + "content": "nies. However, considering the steep costs of tuning models, directly divulging model parameters potentially engenders no returns on investment and even adverse societal impacts from malicious use. To alleviate those concerns, prediction APIs have emerged as an alternative by enabling black-box inferences through querying. This allows models to provide services without exposing the underlying parameters and details, and has been harnessed in Machine Learning as a Service (MLaaS) platforms (e.g., Amazon AWS, Microsoft Azure, Google Cloud)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.477, + 0.896, + 0.763 + ], + "angle": 0, + "content": "Unfortunately, recent research has demonstrated that prediction APIs are still vulnerable to intellectual property theft through model stealing attacks [39, 51, 54, 62]. Despite not exposing model details, attackers can query the APIs using carefully-crafted samples or surrogate samples to obtain responded predictions. Through this, they can then train a clone model on the gathered query-response pairs to replicate the victim model's functionality with high accuracy and fidelity, without needing to access the original training data or model architecture [22, 39]. Such a clone model poses the following security threats: (a) acquiring a functionally similar copy of the victim model, illegally circumventing the expensive process of manual data curation and undermining the intellectual property of MLaaS; (b) extracting confidential data through membership inference attacks [47], model inversion attack [6] and property inference attack [7]; (c) conducting adversarial attacks [31, 33, 35, 36, 49, 55] by creating malicious examples based on the white-box substitute model [42]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.897, + 0.903 + ], + "angle": 0, + "content": "To address such vulnerabilities, considerable researches [3, 5, 9, 17, 29, 50, 64] have focused on thwarting model stealing attacks. These defenses can be categorized into three strategies: detection, perturbation, and post-hoc verification. Detection-based defenses [19, 20, 23, 23] are predicated on the idea that malicious queries exhibit aberrant statistical distributions compared to normal queries. If a query sequence appears anomalous, the defender may discard the requests or corrupt the outputs. However, if an attacker carefully con" + }, + { + "type": "page_footnote", + "bbox": [ + 0.088, + 0.876, + 0.218, + 0.888 + ], + "angle": 0, + "content": "* Corresponding author." + }, + { + "type": "page_footnote", + "bbox": [ + 0.088, + 0.888, + 0.468, + 0.901 + ], + "angle": 0, + "content": "This work is done when Dong-Dong Wu works as an intern at Ant Group." + }, + { + "type": "list", + "bbox": [ + 0.088, + 0.876, + 0.468, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "24305" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.081, + 0.09, + 0.468, + 0.196 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.208, + 0.471, + 0.293 + ], + "angle": 0, + "content": "Figure 1. Illustrations of the noise transition matrix serving for noisy label learning and model stealing defense. In noisy label learning, it transforms the predicted posteriors to align with the noisy label. In model stealing defense, it perturbs the predicted posterior to impede attacker training while preserving the original predicted label for benign users." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.324, + 0.47, + 0.612 + ], + "angle": 0, + "content": "structures queries matching the benign distribution, detection techniques will fail to discern the attack, yielding excessive false positives. Perturbation-based defenses [30, 40, 51] can preserve or constraint model accuracy to foil model stealing attacks. The accuracy-preserving defense truncates predicted posteriors to a minuscule fraction of their original size, which may reduce model lucidity and utility for critical applications like autonomous driving. The accuracy-constraint defense [38, 40] introduces perturbation into predicted posteriors, which incurs additional computational burdens or substantially degrading utility. Furthermore, the latency resulting from such perturbation methods leads to extended user response times, making the approach impractical. Post-hoc verification techniques such as watermarking [15] remain useful for post-attack prevention and validation. However, their retroactive nature means they cannot take proactive measures before an attack occurs. Considering inference latency and utility trade-offs, in this paper, we concentrate on a novel and efficient model stealing defense strategy." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.614, + 0.471, + 0.901 + ], + "angle": 0, + "content": "To overcome the limitations of existing defenses, we explore an efficient method to inject noise into predicted posteriors via linear mapping, while preserving utility for benign users. Motivated by memorization effects [11, 63] that neural networks can easily overfit noisy data, thus hampering model generalization, we revisit the utilization of noise transition matrix (NTM) in noisy labels learning [43, 58], where the matrix models the noise process. Unlike related work, we aim to leverage the NTM to introduce noise into predicted posteriors for model defense, as shown in Figure 1. Our goal is to find an optimized NTM that injects noise while retaining original prediction labels with minimal utility trade-offs. Intuitively, regardless of whether the victim model produces extremely confident one-hot vectors or uncertain flat vectors, we can readily derive a uniform NTM where the diagonal elements remain maximal for each row to meet our defense needs. To estimate the NTM with maximum utility, we propose a bi-level optimization method that alternately optimizes it and updates the surrogate network" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.166 + ], + "angle": 0, + "content": "weights. Specifically, we first make a one-step-forward virtual optimization of weights, then optimize the matrix guided by the loss with the one-step-forward weights fixed. Finally, we optimize the unrolled weights with the updated matrix. The contributions of this work are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.169, + 0.894, + 0.244 + ], + "angle": 0, + "content": "- We propose a novel perturbation-based defense that efficiently introduces noise through linear mapping, adding minimal overhead. Furthermore, it deteriorates attacker training to a level equivalent to training on noisy-labeled data with instance-wise weight." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.244, + 0.894, + 0.305 + ], + "angle": 0, + "content": "- Our method pioneers using the transition matrix to thwart model stealing attacks. The explicit matrix helps understand noise relationships among classes and makes the defense strategy more interpretable" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.305, + 0.892, + 0.349 + ], + "angle": 0, + "content": "- A bi-level optimization approach is proposed to optimize the transition matrix, which is model-agnostic and can adapt to different backbone networks." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.35, + 0.894, + 0.38 + ], + "angle": 0, + "content": "- Extensive experiments exhibit our method consistently mitigates stealing and outperforms state-of-the-art defenses." + }, + { + "type": "list", + "bbox": [ + 0.5, + 0.169, + 0.894, + 0.38 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.398, + 0.642, + 0.414 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.425, + 0.895, + 0.728 + ], + "angle": 0, + "content": "Model Stealing Attacks aim to replicate the functionality of a victim model by interrogating its prediction APIs. The attacker sends queries and collects the resultant predictions, typically probability vectors or labels, as query-response pairs. These collected pairs are used to train a clone model that approximates the victim model. Tramèr et al. [51] first proposed the idea of model stealing and demonstrated the feasibility of extracting simple models utilizing queries. Juuti et al. [19] presented extraction attacks leveraging jacobian-based data augmentation [42], synthesizing data by perturbing a small set of in-distribution samples. However, these attacks presume the ability to optimally probe the APIs, which could be detected by monitoring for anomalous query patterns [19, 41]. Therefore, we concentrate on scenarios without carefully engineered input queries or sequences. One classic scenario to achieve this is by utilizing knowledge distillation techniques [12], where some studies exhibit model stealing is viable using weakly correlated [39] or random [25] queries, even when the teacher model declares itself undistillable [13, 28]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.735, + 0.897, + 0.903 + ], + "angle": 0, + "content": "Model Stealing Defenses can be taxonomized into three primary strategies: detection, perturbation, and post-hoc verification. Detection strategies [20, 21] aim to identify malicious queries or nefarious users by scrutinizing the queries received by the victim model. For example, Kariyappa and Qureshi [20] regarded malicious queries as out-of-distribution samples and devised an adaptive misinformation (AM) defense to mislead attackers with modified outputs. However, such detection strategies rely on strong assumptions about adversarial query patterns that may not reflect reality. Perturbation strategies mitigate model stealing" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "24306" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.473, + 0.485 + ], + "angle": 0, + "content": "attacks by returning perturbed posterior, which can be further subcategorized into accuracy-preserving [30] and accuracy-constrained defenses [29, 38, 40]. Accuracy-preserving defenses guarantee that the maximal class of the perturbed predictions matches the original prediction. The most straightforward such defense is truncating the posterior or only outputting the predicted label [51]. However, this diminishes utility for benign users, precluding specific applications and harming external transparency. In contrast, accuracy-constrained defenses permit a higher level of perturbation, trading off benign accuracy for enhanced security against model stealing attacks. In general, existing defense strategies that poison the attacker's update gradient [38, 40] are computationally expensive, even if they yield a comparative defense. Post-hoc verification, including watermarking [15, 32, 50], dataset inference [37], and proof of learning [16], embeds concealed patterns in the model during training [1] or inference [50] to distinguish a stolen model. Unfortunately, these approaches only work after a model theft, requiring the model owner to obtain partial or full access to the stolen model. This means if an attacker uses the stolen model as a proxy to attack the victim model, it is impossible for the model owner to protect themselves with these defenses. Overall, while existing ante-hoc defensive approaches mitigate knowledge leakage, they rely on strong assumptions or incur high computational costs during inference." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.492, + 0.472, + 0.766 + ], + "angle": 0, + "content": "Noise Transition Matrix (NTM) is often utilized to solve classification with noisy labels or ambiguous labeled data [10, 46, 61], where the matrix reflects probabilities of true label flipping to the other label. A typical application is loss correction [43], which modifies the loss of each example by multiplying the estimated matrix by the model output. For NTM estimation, there are two main approaches. One is a two-step solution that pre-estimates the matrix with anchor point assumptions, then uses it to train the classifier [56, 65]. For instance, backward correction [43] initially approximates the NTM using the softmax output of the model trained without correction. It then retransmits the model while correcting the original loss based on the estimated matrix. The other jointly estimates the matrix and classifier parameters in a unified framework without employing anchor points [18, 48, 60]. Differently, our work aims to estimate and apply the transition amtrix to intentionally inject noise into predicted posteriors, thus defending against model stealing." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.783, + 0.218, + 0.799 + ], + "angle": 0, + "content": "3. Threat Model" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.81, + 0.471, + 0.902 + ], + "angle": 0, + "content": "We consider a model stealing scenario involving a defender and an attacker interacting on image classification tasks. The defender grants access to its black-box model through prediction API. Leveraging the API, the attacker collects query-response pairs, aiming to mimic the defender model (victim model) functionalities by training a clone model on" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.183 + ], + "angle": 0, + "content": "this dataset. To evade detection defenses [19], we assume the attacker adopts non-adaptive querying strategies without anomalous patterns [39, 42]. Therefore, we consider the knockoff nets attack without adaptive querying [39], which resembles knowledge distillation [12] where the victim model distills knowledge into the attacker model." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.189, + 0.696, + 0.206 + ], + "angle": 0, + "content": "3.1. Attacker's Objective" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.212, + 0.895, + 0.394 + ], + "angle": 0, + "content": "Let \\(\\theta_{f}\\) and \\(\\theta_{h}\\) denote the parameters of the defender's model \\(f\\) and attacker's clone model \\(h\\), respectively. The attacker prepares a query dataset \\(\\mathcal{D}_{\\mathrm{query}}\\), and sends queries \\(\\pmb{x} \\in \\mathcal{D}_{\\mathrm{query}}\\) in random order to obtain the defender's posterior \\(f(\\pmb{x}; \\pmb{\\theta}_{f})\\) from the prediction API. The classification loss of the attacker's clone model on the query \\(\\pmb{x}\\) is \\(\\mathcal{L}_{\\mathrm{kna}} = -\\sum_{i} f(\\pmb{x}; \\pmb{\\theta}_{f})_{i} \\log h(\\pmb{x}; \\pmb{\\theta}_{h})_{i}\\). The attacker aims to minimize the classification loss on the defender's test set. However, since the test set is unavailable, the attacker's learning objective becomes minimizing the loss based on the collected query-response dataset \\(\\mathcal{D}_{\\mathrm{pair}} = \\{(x, f(x; \\pmb{\\theta}_{f})) | x \\in \\mathcal{D}_{\\mathrm{query}}\\}\\). Therefore, the attacker's objective can be formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.564, + 0.4, + 0.892, + 0.424 + ], + "angle": 0, + "content": "\\[\n\\min _ {\\boldsymbol {\\theta} _ {h}} \\mathbb {E} _ {\\boldsymbol {x} \\sim \\mathcal {D} _ {\\text {p a i r}}} \\left[ \\mathcal {L} _ {\\mathrm {k n a}} \\left(h (\\boldsymbol {x}; \\boldsymbol {\\theta} _ {h}), f (\\boldsymbol {x}; \\boldsymbol {\\theta} _ {f})\\right) \\right] \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.43, + 0.895, + 0.581 + ], + "angle": 0, + "content": "Query Distribution. In practice, due to the knowledge limitation of attackers about the defender's training set, the collected query dataset \\(\\mathcal{D}_{\\mathrm{query}}\\) may be similar [30, 51] or unrelated [20, 40] to the defender's training distribution. Hence, we consider both possibilities in our threat model. We refer to attackers as distribution-aware if their queries share semantic content with the defender's training distribution. Attackers are referred to as knowledge-limited if their queries are semantically disjoint from the defender's training distribution." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.588, + 0.698, + 0.604 + ], + "angle": 0, + "content": "3.2. Defender's Objective" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.611, + 0.895, + 0.763 + ], + "angle": 0, + "content": "In defenses against model stealing, the defender aims at preventing its model functionality from being stolen. In real-world scenarios, attackers may disguise themselves as benign users, executing attacks anytime. Therefore, the defender must persistently mitigate model stealing attempts while preserving the API's utility for legitimate users. Specifically, the defender needs to keep its own small classification loss, while maximizing the classification loss of the attacker's cloned model \\(\\theta_h^*(\\theta_f)\\) obtained from Eq. (1) on the test set \\(\\mathcal{D}_{\\mathrm{test}}\\). The defender's objective can be formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.576, + 0.768, + 0.892, + 0.797 + ], + "angle": 0, + "content": "\\[\n\\min _ {\\boldsymbol {\\theta} _ {f}} \\mathbb {E} _ {(\\boldsymbol {x}, y) \\sim \\mathcal {D} _ {\\text {t e s t}}} [ \\mathcal {L} _ {\\mathrm {C E}} (f (\\boldsymbol {x}; \\boldsymbol {\\theta} _ {f}), y) ] \\tag {2}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.55, + 0.797, + 0.845, + 0.814 + ], + "angle": 0, + "content": "\\[\n\\operatorname {s. t.} \\max _ {\\mathbf {x}, y) \\sim \\mathcal {D} _ {\\text {t e s t}}} [ \\mathcal {L} _ {\\mathrm {C E}} (h (\\boldsymbol {x}; \\boldsymbol {\\theta} _ {h} ^ {*} (\\boldsymbol {\\theta} _ {f})), y) ],\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.819, + 0.829, + 0.834 + ], + "angle": 0, + "content": "where \\(\\mathcal{L}_{\\mathrm{CE}}\\) means the cross-entropy loss function." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.845, + 0.61, + 0.862 + ], + "angle": 0, + "content": "4. Approach" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.871, + 0.895, + 0.902 + ], + "angle": 0, + "content": "To develop an efficacious and inference-efficient defense against model stealing, we propose an Efficient Model de" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "24307" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.485 + ], + "angle": 0, + "content": "fense with noise transition MMatrix (EMMA) to mitigate model stealing attacks. As depicted in the \"Defense phase\" of Figure 2, the attacker iteratively queries the victim model using unlabeled data to obtain posteriors via the prediction API. The victim model receives these queries and outputs corresponding predicted probability vectors. These posteriors then multiply the optimized matrix \\( T^n \\), becoming perturbed posteriors that remain useful to benign users, and returned to the attacker. However, if the attacker collects these query-responses as training data to reconstruct a clone model, in that case, this dataset is equivalent to noisy labeled data with instance-wise weight, thus impeding cloning model training. To optimize the transition matrix with maximum utility, as shown in the \"Optimizing phase\", we devise a bi-level optimization between the victim and surrogate model which is devised to represent the attacker model. In each round, given initial \\( T^t \\) and surrogate weights \\( \\theta_g^t \\), the optimized matrix \\( T^{t+1} \\) stems from two gradient descents (GD) on the attack loss and defense loss, respectively. The surrogate model then updates through a GD on the attack loss. This process loops until reaching the termination condition, yielding the final optimized matrix \\( T^n \\). In this section, we first elucidate the defense mechanism of incorporating the transition matrix during the defensive phase. Then, a bi-level optimization approach is expounded to alternately optimize the matrix and update the surrogate model." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.494, + 0.325, + 0.51 + ], + "angle": 0, + "content": "4.1. The underlying mechanism" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.518, + 0.471, + 0.653 + ], + "angle": 0, + "content": "To understand how the transition matrix defends against attacks, we analyze the classification risks of two scenarios: \"learning on noisy labels\" and \"learning on query-response pairs with perturbed posteriors\". It is found that the two classification risks differ only by an instance-wise weight coefficient multiplied to the latter. Let \\(\\tilde{y} \\in \\mathbf{Y}\\) denote the noisy label, where \\(\\tilde{y}\\) is selected based on the transition matrix and the true label. Theorem 1 formalizes the defensive mechanism." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.664, + 0.472, + 0.771 + ], + "angle": 0, + "content": "Theorem 1. Let \\(\\mathcal{L}(\\cdot, \\cdot)\\) denote the classification loss function. With sufficiently many queries, the classification risk \\(\\Re(h)\\) of learning with noisy labels is \\(\\mathbb{E}_{(\\boldsymbol{x}, \\tilde{\\boldsymbol{y}})}[\\mathcal{L}(h(\\boldsymbol{x}), \\tilde{\\boldsymbol{y}})]\\). The classification risk \\(\\Re_T(h)\\) of learning with query-response pairs is \\(\\mathbb{E}_{(\\boldsymbol{x}, \\tilde{\\boldsymbol{y}})}[q(\\boldsymbol{x}, \\tilde{\\boldsymbol{y}}) \\cdot \\mathcal{L}(h(\\boldsymbol{x}), \\tilde{\\boldsymbol{y}})]\\), where \\(q(\\boldsymbol{x}, \\tilde{\\boldsymbol{y}})\\) denotes the instance-wise weight, proportional to the closeness between the predicted posterior and the true label." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.78, + 0.473, + 0.903 + ], + "angle": 0, + "content": "The proof can be found in Appendix 8. Theorem 1 demonstrates that training a clone model on such query-response pairs with perturbed posteriors is equivalent to training it on noisy-labeled data with instance-wise weights. It can be further deduced that the more closely the victim model's predicted posteriors match the true label, the more this training process resembles learning with noisy labels. In addition, the clone model's generalization can be further degraded by" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.091, + 0.891, + 0.334 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.345, + 0.893, + 0.429 + ], + "angle": 0, + "content": "Figure 2. An illustration of \"Optimizing phase\" and \"Defense phase\" in the framework. \"Optimizing phase\" demonstrates the process of updating the noise transition matrix \\( T \\) and surrogate network weights \\( \\theta_g \\) alternatively. \"Defensive phase\" shows the interaction process where the attacker tries to steal the functionality from the defender." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.455, + 0.892, + 0.501 + ], + "angle": 0, + "content": "increasing the transition matrix noise rate or exploring more difficult noise type through inter-class correlations, such as using class-conditional noise instead of random noise." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.509, + 0.78, + 0.525 + ], + "angle": 0, + "content": "4.2. The overall training framework" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.532, + 0.895, + 0.682 + ], + "angle": 0, + "content": "Since the attacker's network is unknown, we cannot directly measure its cloning performance. Instead, we represent the attacker with a surrogate model \\( g \\) parameterized by \\( \\theta_g \\) and design optimal perturbations for it. By using the surrogate model to simulate the cloning process, we can continuously optimize the transition matrix during this process to minimize the utility trade-off between attacker performance and defender utility. To achieve this, we first introduce the respective objective functions for the attacker and defender. Then, a bi-level optimization approach is proposed." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.683, + 0.892, + 0.729 + ], + "angle": 0, + "content": "Given the perturbed posterior \\( f(\\pmb{x}; \\pmb{\\theta}_f)T \\), the attacker's objective function is defined as the loss on the gathered query-response pairs, formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.512, + 0.74, + 0.894, + 0.758 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {a t t a c k}} \\left(\\boldsymbol {\\theta} _ {g}, T\\right) = \\mathbb {E} _ {\\boldsymbol {x} \\sim \\mathcal {D} _ {\\text {p a i r}}} \\left[ \\mathcal {L} _ {\\text {k n a}} \\left(g (\\boldsymbol {x}; \\boldsymbol {\\theta} _ {g}), f (\\boldsymbol {x}; \\boldsymbol {\\theta} _ {f}) T\\right) \\right]. \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.769, + 0.895, + 0.875 + ], + "angle": 0, + "content": "Notably, the optimization of \\(\\theta_{g}\\) is influenced by \\(T\\). To defend against model stealing, we would like \\(T\\) to regularize \\(\\theta_{g}\\) to have poor generalizability. To achieve this, we instantiate a misinformation loss that inhibits the attacker model's outputs on the victim model's original predicted label \\(k = \\arg \\max_{i} f(\\pmb{x}, \\pmb{\\theta}_{f})_{i}\\). Formally, we derive the following modified negative log-likelihood loss:" + }, + { + "type": "equation", + "bbox": [ + 0.549, + 0.885, + 0.893, + 0.903 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\operatorname {m i s}} \\left(\\boldsymbol {\\theta} _ {g}\\right) = \\mathbb {E} _ {\\boldsymbol {x} \\sim \\mathcal {D} _ {\\text {p a i r}}} \\left[ - \\log \\left(1 - g (\\boldsymbol {x}; \\boldsymbol {\\theta} _ {g}) _ {k}\\right) \\right]. \\tag {4}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "24308" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.167 + ], + "angle": 0, + "content": "The underlying principle is quite simple: The log-likelihood of the original prediction label on the output should be minimized. Despite its simple form, optimizing this loss has been empirically shown to be an effective defense against model stealing attacks [20, 59]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.168, + 0.471, + 0.241 + ], + "angle": 0, + "content": "Moreover, to maintain fidelity for the victim model, a straightforward strategy is aligning the perturbed output with the predicted label \\( k \\). This is implemented by minimizing the divergence between the output and the predicted label, formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.133, + 0.254, + 0.47, + 0.272 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {a l i g n}} (T) = \\mathbb {E} _ {\\boldsymbol {x} \\sim \\mathcal {D} _ {\\text {p a i r}}} [ - \\log (f (\\boldsymbol {x}; \\boldsymbol {\\theta} _ {f}) T) _ {k} ]. \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.282, + 0.47, + 0.313 + ], + "angle": 0, + "content": "In summary, the overall objective function of the defender can be represented as the following defense loss:" + }, + { + "type": "equation", + "bbox": [ + 0.097, + 0.324, + 0.47, + 0.342 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {d e f e n s e}} \\left(\\boldsymbol {\\theta} _ {g}, T\\right) = \\lambda \\cdot \\mathcal {L} _ {\\text {a l i g n}} (T) + (1 - \\lambda) \\cdot \\mathcal {L} _ {\\text {m i s}} \\left(\\boldsymbol {\\theta} _ {g}\\right), \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.353, + 0.47, + 0.382 + ], + "angle": 0, + "content": "where the multiplicative factor \\(\\lambda\\) is used to balance the contribution of these two loss terms." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.388, + 0.471, + 0.478 + ], + "angle": 0, + "content": "Bi-level optimization. Given the defense loss \\(\\mathcal{L}_{\\mathrm{defense}}\\), the surrogate model weights \\(\\theta_{g}\\) need to be updated first before optimizing the transition matrix \\(T\\). For a mini-batch sample, we propose that surrogate model weights could be treated as a latent variable and optimized alternatively with the matrix. This yields the following bi-level optimization problem:" + }, + { + "type": "equation", + "bbox": [ + 0.144, + 0.489, + 0.303, + 0.514 + ], + "angle": 0, + "content": "\\[\n\\underset {T} {\\arg \\min } \\mathcal {L} _ {\\text {d e f s e n s e}} \\left(\\boldsymbol {\\theta} _ {g} ^ {*}, T\\right)\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.144, + 0.511, + 0.47, + 0.543 + ], + "angle": 0, + "content": "\\[\n\\text {s u b j e c t} \\theta_ {g} ^ {*} = \\underset {\\theta_ {g}} {\\arg \\min } \\mathcal {L} _ {\\text {a t t a c k}} \\left(\\theta_ {g}, T\\right) \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.554, + 0.471, + 0.72 + ], + "angle": 0, + "content": "Bi-level optimization problem has been proved strongly NP-hard [14], so getting such an exact solution for problem (7) is impossible in polynomial time. Fortunately, the inner optimization problem in (7) can be transformed into an optimized result \\(\\theta_g(T)\\) as a function of \\(T\\). Then \\(\\theta_g(T)\\) can then substitute \\(\\theta_g\\) in the outer optimization to acquire an objective solely in terms of \\(T\\). This allows approximating the solution using gradient descent. Therefore, we employ an alternating optimization strategy to approximately solve this problem. In each iteration \\(\\theta_g\\) and \\(T\\) are updated on-the-fly in an alternating manner." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.726, + 0.471, + 0.801 + ], + "angle": 0, + "content": "Optimizing the transition matrix. In step \\(t + 1\\) of optimizing \\(T^t\\), guided by the attack loss, surrogate model weights \\(\\theta_g^t\\) are updated in a one-step-forward manner to obtain \\(\\theta_g^{t + 1}\\). Specifically, with the fixed \\(T^t\\), then \\(\\theta_g^t\\) is updated via gradient descent with learning rate \\(\\alpha\\) as:" + }, + { + "type": "equation", + "bbox": [ + 0.165, + 0.811, + 0.47, + 0.832 + ], + "angle": 0, + "content": "\\[\n\\tilde {\\boldsymbol {\\theta}} _ {g} ^ {t + 1} \\left(T ^ {t}\\right) = \\boldsymbol {\\theta} _ {g} ^ {t} - \\alpha \\cdot \\bigtriangledown_ {\\boldsymbol {\\theta} _ {g} ^ {t}} \\mathcal {L} _ {\\text {a t t a c k}}. \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.47, + 0.9 + ], + "angle": 0, + "content": "Note that the surrogate model \\( g \\) does not actually update in this step, but makes preparations for estimating \\( T^{t + 1} \\). We then optimize \\( T^t \\) with \\( \\tilde{\\theta}_g^{t + 1} \\) fixed under the defense loss. The motivation is that we would like to find a \\( T^{t + 1} \\) that achieves" + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.092, + 0.792, + 0.107 + ], + "angle": 0, + "content": "Algorithm 1 The overall training procedure" + }, + { + "type": "text", + "bbox": [ + 0.501, + 0.111, + 0.693, + 0.127 + ], + "angle": 0, + "content": "Input: Query dataset \\(\\mathcal{D}_{\\mathrm{query}}\\)" + }, + { + "type": "text", + "bbox": [ + 0.55, + 0.127, + 0.849, + 0.142 + ], + "angle": 0, + "content": "The surrogate model \\(g\\) and its parameters \\(\\theta_{g}\\)" + }, + { + "type": "text", + "bbox": [ + 0.55, + 0.142, + 0.75, + 0.155 + ], + "angle": 0, + "content": "The noise transition matrix \\(T\\)." + }, + { + "type": "text", + "bbox": [ + 0.55, + 0.157, + 0.733, + 0.17 + ], + "angle": 0, + "content": "The number of iterations \\(I\\);" + }, + { + "type": "text", + "bbox": [ + 0.55, + 0.172, + 0.733, + 0.185 + ], + "angle": 0, + "content": "The multiplicative factor \\(\\lambda\\);" + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.187, + 0.582, + 0.199 + ], + "angle": 0, + "content": "Procedure:" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.204, + 0.66, + 0.219 + ], + "angle": 0, + "content": "1: Initialize \\(T\\) and \\(\\theta_{g}\\)" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.22, + 0.645, + 0.233 + ], + "angle": 0, + "content": "2: for \\( i = 1 \\) to \\( I \\) do" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.235, + 0.789, + 0.249 + ], + "angle": 0, + "content": "3: Fetch a random batch \\(\\mathcal{B}\\) from \\(\\mathcal{D}_{\\mathrm{query}}\\)" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.25, + 0.891, + 0.279 + ], + "angle": 0, + "content": "4: Calculate the one-step-forward surrogate model weights via Eq. (8);" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.28, + 0.833, + 0.294 + ], + "angle": 0, + "content": "5: Update the transition matrix via Eq. (9-12);" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.295, + 0.847, + 0.309 + ], + "angle": 0, + "content": "6: Update surrogate model weights via Eq. (13);" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.311, + 0.584, + 0.323 + ], + "angle": 0, + "content": "7: end for" + }, + { + "type": "list", + "bbox": [ + 0.511, + 0.204, + 0.891, + 0.323 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.501, + 0.327, + 0.774, + 0.342 + ], + "angle": 0, + "content": "Output: Learned noise transition matrix." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.371, + 0.894, + 0.415 + ], + "angle": 0, + "content": "minimal utility trade-offs between the defender and attacker. The transition matrix \\( T^{t} \\) is updated via gradient descent with learning rate \\( \\beta \\):" + }, + { + "type": "equation", + "bbox": [ + 0.597, + 0.424, + 0.892, + 0.441 + ], + "angle": 0, + "content": "\\[\nw ^ {t + 1} = T ^ {t} - \\beta \\cdot \\bigtriangledown_ {T ^ {t}} \\mathcal {L} _ {\\text {D e f s e n s e}}. \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.451, + 0.764, + 0.466 + ], + "angle": 0, + "content": "Apply chain rule to Eq. (9), we can get:" + }, + { + "type": "equation", + "bbox": [ + 0.515, + 0.472, + 0.892, + 0.515 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} w ^ {t + 1} = T ^ {t} - \\beta \\cdot \\left\\{\\lambda \\cdot \\nabla_ {T ^ {t}} \\mathcal {L} _ {\\text {a l i g n}} \\right. \\\\ + (1 - \\lambda) \\cdot \\bigtriangledown_ {\\tilde {\\boldsymbol {\\theta}} _ {g} ^ {t + 1}} \\mathcal {L} _ {\\mathrm {m i s}} \\cdot \\left(- \\alpha \\bigtriangledown_ {\\boldsymbol {\\theta} _ {g} ^ {t}, T ^ {t}} ^ {2} \\mathcal {L} _ {\\text {a t t a c k}}\\right) \\}. \\tag {10} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.522, + 0.894, + 0.598 + ], + "angle": 0, + "content": "Note that \\( w^{t + 1} \\) is not the final transition matrix, as its entries may not satisfy the properties of a valid transition matrix. Specifically, all elements must be non-negative and each row must be normalized. To address this, we first replace any negative elements in \\( w^{t + 1} \\) with zeros via:" + }, + { + "type": "equation", + "bbox": [ + 0.617, + 0.605, + 0.892, + 0.623 + ], + "angle": 0, + "content": "\\[\n\\tilde {w} ^ {t + 1} = \\max \\left(w ^ {t + 1}, 0\\right). \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.632, + 0.894, + 0.696 + ], + "angle": 0, + "content": "Then, each row is normalized to get the final valid transition matrix. For the \\(j^{\\mathrm{th}}\\) row of \\(\\tilde{w}^{t + 1}\\), which indicates the probabilities of class \\(j\\) transitioning to other classes, a normalization is performed to obtain the final matrix \\(T_{j,i}^{t + 1}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.51, + 0.703, + 0.892, + 0.755 + ], + "angle": 0, + "content": "\\[\nT _ {j, i} ^ {t + 1} = \\frac {\\tilde {w} _ {j , i} ^ {t + 1}}{\\sum_ {i} \\tilde {w} _ {j , i} ^ {t + 1} + \\delta \\left(\\sum_ {i} \\tilde {w} _ {j , i} ^ {t + 1}\\right)}, \\delta (a) = \\left\\{ \\begin{array}{l} 1, \\text {i f} a = 0 \\\\ 0, \\text {i f} a \\neq 0 \\end{array} \\right. \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.756, + 0.771, + 0.772 + ], + "angle": 0, + "content": "where \\(\\delta (\\cdot)\\) is used to avoid division by 0." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.776, + 0.892, + 0.852 + ], + "angle": 0, + "content": "Updating surrogate model weights. After optimizing the matrix, we conduct the surrogate model update by keeping \\( T^{t + 1} \\) fixed and minimizing the attack loss. New surrogate model weights \\( \\theta_{g}^{t + 1} \\) are obtained via gradient descent with learning rate \\( \\gamma \\):" + }, + { + "type": "equation", + "bbox": [ + 0.604, + 0.86, + 0.892, + 0.88 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {\\theta} _ {g} ^ {t + 1} = \\boldsymbol {\\theta} _ {g} ^ {t} - \\gamma \\cdot \\bigtriangledown \\boldsymbol {\\theta} _ {g} ^ {t} \\mathcal {L} _ {\\text {a t t a c k}}. \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.886, + 0.894, + 0.901 + ], + "angle": 0, + "content": "The whole training procedure is summarized in Algorithm 1." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "24309" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.089, + 0.089, + 0.885, + 0.382 + ], + "angle": 0, + "content": "
MethodCIFAR100 → CIFAR10CIFAR10 → CIFAR100Caltech256 → CUB200
ΔAcc=1ΔAcc=2ΔAcc=5ΔAcc=1ΔAcc=2ΔAcc=5ΔAcc=1ΔAcc=2ΔAcc=5
Random84.1777.2174.2143.2339.8132.1233.5031.1428.01
AM77.1072.9357.5639.8232.9819.3626.1721.2714.31
MAD80.5777.5469.0736.5632.5926.6537.1635.2331.50
NASTY76.6574.3164.1938.5333.6126.9531.0628.2324.70
GARD²77.4273.8871.4041.2139.2935.9435.6933.6732.34
MIN-IP78.8873.4070.6642.3338.6734.2634.7332.2628.94
EMMA61.7952.9142.2128.0118.3413.3719.1015.7913.11
ImageNet-C10 → CIFAR10ImageNet-C100 → CIFAR100ImageNet-CUB200 → CUB200
ΔAcc=1ΔAcc=2ΔAcc=5ΔAcc=1ΔAcc=2ΔAcc=5ΔAcc=1ΔAcc=2ΔAcc=5
Random88.9988.1887.0763.1661.7658.6446.5844.0640.49
AM87.5584.6076.7260.5756.0947.1045.1239.6530.15
MAD88.6785.7078.4458.7456.2350.8550.9848.5744.47
NASTY87.3385.3381.6859.7956.0548.9546.5643.2337.76
GARD²87.2986.0685.9358.3956.7153.6350.2747.5543.11
MIN-IP87.4086.8184.4561.1756.3551.9850.0946.4542.06
EMMA84.1080.6077.5256.7349.1831.4836.3533.1026.33
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.386, + 0.895, + 0.416 + ], + "angle": 0, + "content": "Table 1. Accuracy (%) of attacker model on defender's test set under knowledge-limited (top row) and distribution-aware (bottom row) attack scenarios with varying budgets. The best result in each column is highlighted in bold." + }, + { + "type": "table", + "bbox": [ + 0.079, + 0.429, + 0.47, + 0.512 + ], + "angle": 0, + "content": "
Eval DataDefender AccuracyAttacker Accuracy
KLDA
CIFAR1095.33 ± 0.1385.17 ± 0.8391.78 ± 0.19
CIFAR10075.96 ± 0.0951.95 ± 1.1266.82 ± 0.16
CUB20081.44 ± 0.4648.93 ± 0.3762.12 ± 0.22
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.514, + 0.472, + 0.545 + ], + "angle": 0, + "content": "Table 2. Accuracy \\((\\%)\\) of the defender model and the attacker model with no defense applied." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.557, + 0.21, + 0.574 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.583, + 0.47, + 0.63 + ], + "angle": 0, + "content": "To demonstrate the effectiveness of our proposed defense method, we conduct experiments on multiple image datasets and present the results in this section." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.642, + 0.254, + 0.659 + ], + "angle": 0, + "content": "5.1. Experiment Setup" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.667, + 0.471, + 0.849 + ], + "angle": 0, + "content": "Datasets. We employ three datasets for evaluation: CIFAR10, CIFAR100 [26] and CUB200 [57]. For each dataset, we explore two attacker scenarios - knowledge-limited (KL) and distribution-aware (DA). The KL scenario uses query sets of CIFAR100, CIFAR10, and Caltech256 [8], paired with CIFAR10, CIFAR100, and CUB200 evaluation datasets respectively. The DA scenario employs tailored query sets from ImageNet-1K [4] manually selected by prior work [38], specifically ImageNet-C10, ImageNet-C100, and ImageNet-CUB200 containing 183,763, 161,653, and 30,000 examples matched to the evaluation datasets, with their further details provided in the Appendix 9." + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.856, + 0.472, + 0.903 + ], + "angle": 0, + "content": "Compared methods. We compare our method against the following six model stealing defense baselines: (1) No Defense, a naive approach where no defense is attempted. (2)" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.427, + 0.895, + 0.834 + ], + "angle": 0, + "content": "Random, an interpolation-based method that perturbs the defender's predicted posterior by interpolating it with a one-hot posterior vector, where the index of single non-zero entry is randomly selected from the non-argmax prediction labels. (3) AM [20], a detection-based method that identifies out-of-distribution queries and interpolates with posteriors from a misinformation network, where the interpolation weight is denoted by the confidence of the argmax prediction label. (4) MAD [40], an accuracy-constrained method that adds controlled noise to original posteriors so that it maximizes the angular deviation between the perturbed and original gradient signal on a surrogate model. The surrogate model is randomly initialized. (5) NASTY [34], a self-undermining knowledge distillation approach which trains a nasty teacher network by maximizing the difference between the output of the nasty teacher and a normal pretrained network. (5) \\(GARD^2\\) [38], a coordinated gradient redirection defense method, which perturbs original posteriors by setting the target gradient direction to be the all-ones vector. This redirects the attacker to update towards a consistent target direction. (6) MIN-IP [38], an uncoordinated gradient redirection defense method, which perturbs original posteriors by setting the target gradient direction opposite the original gradient direction, redirecting the update gradient of attackers opposite to the original gradients. The surrogate model of MIN-IP and \\(GARD^2\\) train on the attacker's queries with early stopping after 10 epochs." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.835, + 0.895, + 0.881 + ], + "angle": 0, + "content": "For fair comparisons, we basically use the same network architecture, learning rate, optimizer, and augmentation strategy across all the comparing methods." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.886, + 0.895, + 0.902 + ], + "angle": 0, + "content": "Implementation. Our experiment pipeline has three stages:" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "24310" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.086, + 0.089, + 0.262, + 0.208 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.097, + 0.212, + 0.251, + 0.224 + ], + "angle": 0, + "content": "(a) Heatmap of transition matrix." + }, + { + "type": "image", + "bbox": [ + 0.278, + 0.089, + 0.46, + 0.208 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.29, + 0.212, + 0.446, + 0.224 + ], + "angle": 0, + "content": "(b) Heatmap of transition matrix." + }, + { + "type": "image", + "bbox": [ + 0.468, + 0.091, + 0.663, + 0.208 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.497, + 0.213, + 0.644, + 0.224 + ], + "angle": 0, + "content": "(c) Average predicted posterior" + }, + { + "type": "image", + "bbox": [ + 0.682, + 0.092, + 0.876, + 0.208 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.711, + 0.212, + 0.859, + 0.224 + ], + "angle": 0, + "content": "(d) Average perturbed posterior" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.228, + 0.894, + 0.271 + ], + "angle": 0, + "content": "Figure 3. Qualitative analysis of the transition matrix. (a-b): Heatmaps of optimized transition matrixes for distribution-aware attacks (a) Imagenet-C10→CIFAR10 and (b) Imagenet-C100→CIFAR100, respectively. (c-d): Average (c) predicted and (d) perturbed posteriors across the query set of (top) Imagenet-C10 and (bottom) Imagenet-C100 datasets." + }, + { + "type": "image", + "bbox": [ + 0.093, + 0.287, + 0.457, + 0.436 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.088, + 0.447, + 0.457, + 0.461 + ], + "angle": 0, + "content": "Figure 4. Visualizations of predicted and perturbed posteriors." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.476, + 0.47, + 0.566 + ], + "angle": 0, + "content": "First, the defender model trains on each evaluation dataset. Second, defense methods generate perturbed posteriors for query dataset and evaluation dataset at various defense budgets. Finally, the attacker model trains on the collected query-response pairs. We denote attacks with query data \\(\\mathcal{Q}\\) and evaluation data \\(\\mathcal{D}\\) as \\(\\mathcal{Q} \\to \\mathcal{D}\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.569, + 0.473, + 0.903 + ], + "angle": 0, + "content": "We use ResNet34 for CIFAR10/100 and ResNet50 pretrained on ImageNet for CUB200. For all models training, we utilize SGD as the optimizer with a momentum of 0.9 and weight decay of 5e-4 for 100 epochs. Initial learning rates are set as 0.01 for CUB200 and 0.1 for other datasets, with a cosine annealing schedule. For optimizing the transition matrix, the learning rate \\(\\gamma\\) is set as 0.3 for CIFAR10 and 1.0 for other datasets. The surrogate model's one-step-forward learning rate \\(\\alpha\\) and the actual learning rate \\(\\beta\\) are set the same value as specified in the model training settings. Additionally, an early stopping mechanism is set to terminates optimization when the average change in \\(T\\) between consecutive epochs falls below a specified threshold. The threshold is 1e-2 for CIFAR10 and 1e-4 for other datasets. Furthermore, the balancing factor \\(\\lambda\\) is adjusted to achieve different utility budgets. Recent work [38] proposed a novel approach to initialize the surrogate model. They suggested to train the surrogate model through distilling from the victim model with early stopping. This would make it more representative of the actual attacker's behavior. Following this approach, we initialize our surrogate model by training it on the query dataset with the same early stopping mechanism." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.289, + 0.688, + 0.304 + ], + "angle": 0, + "content": "5.2. Experiment Results" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.312, + 0.895, + 0.525 + ], + "angle": 0, + "content": "Table 2 presents the accuracy of the defender model and the attacker model, without any defense applied under two different query distributions. This shows the upper bound for the usage performance of the defender and stealable performance of the attacker. Table 1 compares the accuracy of an attacker's clone model at different defender accuracy drop budgets \\(\\Delta\\)Acc. The budget refers to the average decrease in accuracy the defender causes across the test set by adding perturbations. Corresponding visualizations of these results are provided in the Appendix 10. As shown, our defense method achieves the lowest utility trade-off across budgets. For example, in the CIFAR100 \\(\\rightarrow\\) CIFAR10 attack scenario, our method reduces the attacker model's accuracy by \\(23\\%\\) while only decreasing the defender's accuracy by \\(1\\%\\)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.537, + 0.692, + 0.553 + ], + "angle": 0, + "content": "5.3. Qualitative Analysis" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.561, + 0.895, + 0.713 + ], + "angle": 0, + "content": "Perturbed posteriors make cloning difficult. We visualize the predicted and perturbed posteriors using t-SNE in Figure 4. The perturbed posteriors clearly exhibit degraded inter-class distances compared to the predicted posteriors. This degradation obfuscates the decision boundaries between classes, making it difficult for the attacker to clone the defender model by imitating the posteriors. Notably, such effective defenses are achieved solely through linear mapping of the posteriors, demonstrating the simplicity and efficiency of our proposed noise perturbation strategy." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.896, + 0.901 + ], + "angle": 0, + "content": "The transition matrix induces challenging noise. We analyze the defense details of two attack scenarios: Imagenet-C10→CIFAR10 and Imagenet-C100→CIFAR100 in Figure 3. Specifically, Figure 3 (a-b) visualizes corresponding optimized transition matrices. It can be observed that the diagonal elements are maximized for fidelity, while latent inter-label relationships are also exploited. As shown the optimized matrix in Figure 3 (a), despite it causes a \\(2.35\\%\\) drop in the defender's accuracy on the test set, the optimization makes it have a noise rate as high as \\(77\\%\\). This negatively affects the cloning process of the attacker, reducing the cloned model's accuracy by \\(14\\%\\)." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "24311" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.213 + ], + "angle": 0, + "content": "In Figures 3 (c-d), we visualize the average posterior of predicted and perturbed predictions. Given that ImageNet-C10/C100 follow long-tail distributions as illustrated in appendix 9, the predicted posteriors also exhibit long-tailed shapes as expected. However, the perturbations significantly increase the confidence of tail classes, distorting the true inter-class relationships. This distortion of class relationships also makes stealing more difficult for the attacker." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.224, + 0.231, + 0.24 + ], + "angle": 0, + "content": "5.4. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.247, + 0.474, + 0.414 + ], + "angle": 0, + "content": "Effectiveness of optimized transition matrix. To validate the effectiveness of the optimized transition matrix, we compare the defense performance between it and the uniform transition matrix with an identical noise rate. The results in Table 3 show that the optimized transition matrix can further reduce the attacker's accuracy while maintaining the defender's accuracy. This demonstrates that the optimized matrix provides improved defense over the uniform matrix. Overall, this shows that capturing the inter-class relationships during optimization contributes to generating a matrix that is more adversarial against the attacker." + }, + { + "type": "table", + "bbox": [ + 0.079, + 0.426, + 0.47, + 0.499 + ], + "angle": 0, + "content": "
Attack ScenariosOptimized matrixUniform matrix
DefenderAttackerDefenderAttacker
CIFAR100→CIFAR1093.87±0.1432.41±7.5494.81±0.0671.43±2.26
CIFAR10→CIFAR10074.48±0.2515.65±0.7274.75±0.0735.67±3.44
Caltech256→CUB20080.45±0.1318.25±0.4380.89±0.2230.15±0.21
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.076, + 0.503, + 0.471, + 0.545 + ], + "angle": 0, + "content": "Table 3. Defender and attacker accuracy \\((\\%)\\) with optimized and uniform transition matrix under attack scenarios for CIFAR10, CIFAR100 and CUB200 evaluation datasets." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.562, + 0.472, + 0.79 + ], + "angle": 0, + "content": "Effects of the number of queries. We also verify the effectiveness of our defense against attacks with varying number of queries in Figure 5. The EMMA-top1 result is depicted as a reference, which denotes the victim model only outputting the predicted label rather than the full posterior. As formalized in Theorem 1, when the victim model assigns full confidence to the true label, the attacker's cloning task degrades to learning from noisy labels. Thus, EMMA-top1 signifies the optimal defense under the same transition matrix. EMMA defense consistently forces degradation of the attacker accuracy across all query set sizes, remaining close to the EMMA-top1. As the query increases, the attacker model obtained from the victim model suffers more accuracy drops, revealing that our defense induces more significant detrimental impacts when the attacker uses more queries." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.796, + 0.471, + 0.901 + ], + "angle": 0, + "content": "Transferability across attack scenarios. To explore the transition matrix's transferable defense capabilities, we test it against an attack type different from the one it was optimized for. As shown in Figure 6, we use the matrix optimized under DA attack to defend against KL attack, denoted as \"DA \\(\\rightarrow\\) KL\". Compared to a matrix specifically optimized for KL attack, denoted as \"KL \\(\\rightarrow\\) KL\", it still maintained" + }, + { + "type": "image", + "bbox": [ + 0.507, + 0.092, + 0.693, + 0.185 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.697, + 0.092, + 0.884, + 0.185 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.506, + 0.189, + 0.693, + 0.281 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.697, + 0.189, + 0.884, + 0.281 + ], + "angle": 0, + "content": null + }, + { + "type": "image_footnote", + "bbox": [ + 0.549, + 0.283, + 0.844, + 0.299 + ], + "angle": 0, + "content": "No defense EMMA EMMA-top1" + }, + { + "type": "image_caption", + "bbox": [ + 0.568, + 0.307, + 0.824, + 0.322 + ], + "angle": 0, + "content": "Figure 5. Effects of the number of queries." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.333, + 0.892, + 0.38 + ], + "angle": 0, + "content": "considerable defensive efficacy, exhibiting the optimized transition matrix's robust transferable defense capabilities across different attack types." + }, + { + "type": "image", + "bbox": [ + 0.521, + 0.39, + 0.697, + 0.524 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.704, + 0.39, + 0.868, + 0.524 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.55, + 0.531, + 0.841, + 0.545 + ], + "angle": 0, + "content": "Figure 6. Transferability across attack scenarios." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.555, + 0.892, + 0.585 + ], + "angle": 0, + "content": "Due to the space limitation, additional experiments of our method are provided in Appendix." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.599, + 0.619, + 0.615 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.624, + 0.895, + 0.806 + ], + "angle": 0, + "content": "In this work, we revisited the noise transition matrix in a novel way by applying it to model stealing defense for the first time. This deteriorates the attacker's cloning process to a level equivalent to training on noisy labels, hampering attacker model generalization. To optimize the transition matrix, we proposed a novel bi-level optimization training framework that incorporates fidelity training for the victim model and adversarial training for the surrogate model. Experimental results under various attack scenarios showed the effectiveness and robustness of our proposed defense method compared to current state-of-the-art methods, achieving strong defense with minimal utility trade-off." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.82, + 0.679, + 0.837 + ], + "angle": 0, + "content": "7. Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.845, + 0.895, + 0.903 + ], + "angle": 0, + "content": "The authors wish to thank the anonymous reviewers for their helpful comments and suggestions. This work was supported by the National Science Foundation of China (62225602), and the Big Data Computing Center of Southeast University." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "24312" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.079, + 0.119, + 0.176, + 0.135 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.145, + 0.47, + 0.2 + ], + "angle": 0, + "content": "[1] Yossi Adi, Carsten Baum, Moustapha Cisse, Benny Pinkas, and Joseph Keshet. Turning your weakness into a strength: Watermarking deep neural networks by backdoorsing. In USENIX Security Symposium, pages 1615-1631, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.201, + 0.472, + 0.284 + ], + "angle": 0, + "content": "[2] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. In International Conference on Advanced Neural Information Processing Systems, pages 1877-1901, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.285, + 0.471, + 0.34 + ], + "angle": 0, + "content": "[3] Yanjiao Chen, Rui Guan, Xueluan Gong, Jianshuo Dong, and Meng Xue. D-dae: Defense-penetrating model extraction attacks. In IEEE Symposium on Security and Privacy, pages 382-399, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.341, + 0.471, + 0.396 + ], + "angle": 0, + "content": "[4] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In IEEE/CVF Computer Vision and Pattern Recognition Conference, pages 248-255, 2009. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.397, + 0.471, + 0.453 + ], + "angle": 0, + "content": "[5] Adam Dziedzic, Muhammad Ahmad Kaleem, Yu Shen Lu, and Nicolas Papernot. Increasing the cost of model extraction with calibrated proof of work. In International Conference on Learning Representations, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.453, + 0.471, + 0.509 + ], + "angle": 0, + "content": "[6] Matt Fredrikson, Somesh Jha, and Thomas Ristenpart. Model inversion attacks that exploit confidence information and basic countermeasures. In ACM SIGSAC conference on computer and communications security, pages 1322-1333, 2015. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.509, + 0.471, + 0.579 + ], + "angle": 0, + "content": "[7] Karan Ganju, Qi Wang, Wei Yang, Carl A Gunter, and Nikita Borisov. Property inference attacks on fully connected neural networks using permutation invariant representations. In ACM SIGSAC conference on computer and communications security, pages 619-633, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.579, + 0.471, + 0.607 + ], + "angle": 0, + "content": "[8] Gregory Griffin, Alex Holub, and Pietro Perona. Caltech-256 object category dataset. 2007. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.608, + 0.471, + 0.675 + ], + "angle": 0, + "content": "[9] Jun Guo, Aishan Liu, Xingyu Zheng, Siyuan Liang, Yisong Xiao, Yichao Wu, and Xianglong Liu. Isolation and induction: Training robust deep neural networks against model stealing attacks. In ACM International Conference on Multimedia, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.677, + 0.471, + 0.734 + ], + "angle": 0, + "content": "[10] Bo Han, Quanming Yao, Tongliang Liu, Gang Niu, Ivor W Tsang, James T Kwok, and Masashi Sugiyama. A survey of label-noise representation learning: Past, present and future. arXiv preprint arXiv:2011.04406, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.734, + 0.471, + 0.803 + ], + "angle": 0, + "content": "[11] Bo Han, Quanming Yao, Xingrui Yu, Gang Niu, Miao Xu, Weihua Hu, Ivor W. Tsang, and Masashi Sugiyama. Robust training of deep neural networks with extremely noisy labels. In International Conference on Advanced Neural Information Processing Systems, page 4, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.804, + 0.471, + 0.845 + ], + "angle": 0, + "content": "[12] Geoffrey Hinton, Oriol Vinyals, and Jeff Dean. Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531, 2015. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.846, + 0.471, + 0.902 + ], + "angle": 0, + "content": "[13] Surgan Jandial, Yash Khasbage, Arghya Pal, Vineeth N Balasubramanian, and Balaji Krishnamurthy. Distilling the undistillable: Learning from a nasty teacher. In European Conference on Computer Vision, pages 587-603, 2022. 2" + }, + { + "type": "list", + "bbox": [ + 0.079, + 0.145, + 0.472, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.893, + 0.135 + ], + "angle": 0, + "content": "[14] Robert G Jeroslow. The polynomial hierarchy and a simple model for competitive analysis. Mathematical programming, pages 146-164, 1985. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.136, + 0.894, + 0.192 + ], + "angle": 0, + "content": "[15] Hengrui Jia, Christopher A Choquette-Choo, Varun Chandrasekaran, and Nicolas Papernot. Entangled watermarks as a defense against model extraction. In USENIX Security Symposium, pages 1937-1954, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.192, + 0.894, + 0.262 + ], + "angle": 0, + "content": "[16] Hengrui Jia, Mohammad Yaghini, Christopher A Choquette-Choo, Natalie Dullerud, Anvith Thudi, Varun Chandrasekaran, and Nicolas Papernot. Proof-of-learning: Definitions and practice. In IEEE Symposium on Security and Privacy, pages 1039-1056, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.262, + 0.893, + 0.319 + ], + "angle": 0, + "content": "[17] Wenbo Jiang, Hongwei Li, Guowen Xu, Tianwei Zhang, and Rongxing Lu. A comprehensive defense framework against model extraction attacks. IEEE Transactions on Dependable and Secure Computing, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.32, + 0.893, + 0.375 + ], + "angle": 0, + "content": "[18] Ishan Jindal, Matthew Nokleby, and Xuewen Chen. Learning deep networks from noisy labels with dropout regularization. In IEEE International Conference on Data Mining, pages 967-972, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.376, + 0.893, + 0.432 + ], + "angle": 0, + "content": "[19] Mika Juuti, Sebastian Szyller, Samuel Marchal, and N Asokan. Prada: protecting against dnn model stealing attacks. In IEEE European Symposium on Security and Privacy, pages 512-527, 2019. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.433, + 0.894, + 0.49 + ], + "angle": 0, + "content": "[20] Sanjay Kariyappa and Moinuddin K Qureshi. Defending against model stealing attacks with adaptive misinformation. In IEEE/CVF Computer Vision and Pattern Recognition Conference, pages 770-778, 2020. 1, 2, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.49, + 0.893, + 0.545 + ], + "angle": 0, + "content": "[21] Sanjay Kariyappa, Atul Prakash, and Moinuddin K Qureshi. Protecting dnns from theft using an ensemble of diverse models. In International Conference on Learning Representations, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.546, + 0.893, + 0.589 + ], + "angle": 0, + "content": "[22] Sanjay Kariyappa, Atul Prakash, and Moinuddin K Qureshi. Maze: Data-free model stealing attack using zeroth-order gradient estimation. In CVPR, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.589, + 0.893, + 0.645 + ], + "angle": 0, + "content": "[23] Manish Kesarwani, Bhaskar Mukholy, Vijay Arya, and Sameep Mehta. Model extraction warning in mlaas paradigm. In Annual Computer Security Applications Conference, pages 371-380, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.646, + 0.893, + 0.715 + ], + "angle": 0, + "content": "[24] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In International Conference on Computer Vision, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.716, + 0.893, + 0.773 + ], + "angle": 0, + "content": "[25] Kalpesh Krishna, Gaurav Singh Tomar, Ankur P Parikh, Nicolas Papernot, and Mohit Iyyer. Thieves on sesame street! model extraction of bert-based apis. In International Conference on Learning Representations, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.773, + 0.893, + 0.803 + ], + "angle": 0, + "content": "[26] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.803, + 0.893, + 0.859 + ], + "angle": 0, + "content": "[27] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. ImageNet classification with deep convolutional neural networks. In International Conference on Advanced Neural Information Processing Systems, pages 1106-1114, 2012. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.859, + 0.893, + 0.902 + ], + "angle": 0, + "content": "[28] Souvik Kundu, Qirui Sun, Yao Fu, Massoud Pedram, and Peter Beerel. Analyzing the confidentiality of undistinguishable teachers in knowledge distillation. pages 9181-9192, 2021. 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "24313" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.161 + ], + "angle": 0, + "content": "[29] Jeonghyun Lee, Sungmin Han, and Sangkyun Lee. Model stealing defense against exploiting information leak through the interpretation of deep neural nets. In International Joint Conference on Artificial Intelligence, pages 710-716, 2022. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.163, + 0.472, + 0.22 + ], + "angle": 0, + "content": "[30] Taesung Lee, Benjamin Edwards, Ian Molloy, and Dong Su. Defending against neural network model stealing attacks using deceptive perturbations. In IEEE Security and Privacy Workshops, pages 43-49, 2019. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.221, + 0.472, + 0.302 + ], + "angle": 0, + "content": "[31] Simin Li, Shuning Zhang, Gujun Chen, Dong Wang, Pu Feng, Jiakai Wang, Aishan Liu, Xin Yi, and Xianglong Liu. Towards benchmarking and assessing visual naturalness of physical world adversarial attacks. In IEEE/CVF Computer Vision and Pattern Recognition Conference, pages 12324–12333, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.305, + 0.472, + 0.361 + ], + "angle": 0, + "content": "[32] Yiming Li, Linghui Zhu, Xiaojun Jia, Yong Jiang, Shu-Tao Xia, and Xiaochun Cao. Defending against model stealing via verifying embedded external features. In Association for the Advance of Artificial Intelligence, pages 1464-1472, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.362, + 0.472, + 0.431 + ], + "angle": 0, + "content": "[33] Shunchang Liu, Jiakai Wang, Aishan Liu, Yingwei Li, Yijie Gao, Xianglong Liu, and Dacheng Tao. Harnessing perceptual adversarial patches for crowd counting. In ACM SIGSAC Conference on Computer and Communications Security, pages 2055-2069, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.433, + 0.472, + 0.475 + ], + "angle": 0, + "content": "[34] Haoyu Ma, Tianlong Chen, Ting-Kuei Hu, Chenyu You, Xiaohui Xie, and Zhangyang Wang. Undistillable: Making a nasty teacher that cannot teach students. In ICLR, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.476, + 0.472, + 0.532 + ], + "angle": 0, + "content": "[35] Ke Ma, Qianqian Xu, Jinshan Zeng, Xiaochun Cao, and Qingming Huang. Poisoning attack against estimating from pairwise comparisons. IEEE Trans. Pattern Anal. Mach. Intell., pages 6393-6408, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.533, + 0.472, + 0.602 + ], + "angle": 0, + "content": "[36] Ke Ma, Qianqian Xu, Jinshan Zeng, Guorong Li, Xiaochun Cao, and Qingming Huang. A tale of hodgerank and spectral method: Target attack against rank aggregation is the fixed point of adversarial game. IEEE Trans. Pattern Anal. Mach. Intell., pages 4090-4108, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.604, + 0.472, + 0.658 + ], + "angle": 0, + "content": "[37] Pratyush Maini, Mohammad Yaghini, and Nicolas Papernot. Dataset inference: Ownership resolution in machine learning. In International Conference on Learning Representations, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.66, + 0.472, + 0.716 + ], + "angle": 0, + "content": "[38] Mantas Mazeika, Bo Li, and David Forsyth. How to steer your adversary: Targeted and efficient model stealing defenses with gradient redirection. In International Conference on Machine Learning, pages 15241-15254, 2022. 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.718, + 0.472, + 0.773 + ], + "angle": 0, + "content": "[39] Tribhuvanesh Orekondy, Bernt Schiele, and Mario Fritz. Knockoff nets: Stealing functionality of black-box models. In IEEE/CVF Computer Vision and Pattern Recognition Conference, pages 4954-4963, 2019. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.775, + 0.472, + 0.83 + ], + "angle": 0, + "content": "[40] Tribhuvanesh Orekondy, Bernt Schiele, and Mario Fritz. Prediction poisoning: Towards defenses against dnn model stealing attacks. In International Conference on Learning Representations, 2020. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.832, + 0.472, + 0.9 + ], + "angle": 0, + "content": "[41] Soham Pal, Yash Gupta, Aditya Shukla, Aditya Kanade, Shirish Shevade, and Vinod Ganapathy. Activethief: Model extraction using active learning and unannotated public data. In Association for the Advance of Artificial Intelligence, pages 865–872, 2020. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.161 + ], + "angle": 0, + "content": "[42] Nicolas Papernot, Patrick McDaniel, Ian Goodfellow, Somesh Jha, Z Berkay Celik, and Ananthram Swami. Practical black-box attacks against machine learning. In ACM on Asia conference on computer and communications security, pages 506-519, 2017. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.164, + 0.894, + 0.234 + ], + "angle": 0, + "content": "[43] Giorgio Patrini, Alessandro Rozza, Aditya Krishna Menon, Richard Nock, and Lizhen Qu. Making deep neural networks robust to label noise: A loss correction approach. In IEEE/CVF Computer Vision and Pattern Recognition Conference, pages 1944-1952, 2017. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.236, + 0.894, + 0.277 + ], + "angle": 0, + "content": "[44] Sunandini Sanyal, Sravanti Addepalli, and R Venkatesh Babu. Towards data-free model stealing in a hard label setting. In CVPR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.279, + 0.892, + 0.308 + ], + "angle": 0, + "content": "[45] Zeyang Sha et al. Can't steal? cont-steal! contrastive stealing attacks against image encoders. In CVPR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.31, + 0.892, + 0.352 + ], + "angle": 0, + "content": "[46] Yu Shi, Dong-Dong Wu, Xin Geng, and Min-Ling Zhang. Robust representation learning for unreliable partial label learning. arXiv preprint arXiv:2308.16718, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.353, + 0.892, + 0.41 + ], + "angle": 0, + "content": "[47] Reza Shokri, Marco Stronati, Congzheng Song, and Vitaly Shmatikov. Membership inference attacks against machine learning models. In IEEE symposium on security and privacy, pages 3-18, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.411, + 0.892, + 0.453 + ], + "angle": 0, + "content": "[48] Sainbayar Sukhbaatar, Joan Bruna, Manohar Paluri, Lubomir Bourdev, and Rob Fergus. Training convolutional networks with noisy labels. arXiv preprint arXiv:1406.2080, 2014. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.455, + 0.892, + 0.511 + ], + "angle": 0, + "content": "[49] Christian Szegedy, Wojciech Zaremba, Ilya Sutskever, Joan Bruna, Dumitru Erhan, Ian Goodfellow, and Rob Fergus. Intriguing properties of neural networks. In International Conference on Learning Representations, 2013. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.513, + 0.892, + 0.569 + ], + "angle": 0, + "content": "[50] Sebastian Szyller, Buse Gul Atli, Samuel Marchal, and N Asokan. Dawn: Dynamic adversarial watermarking of neural networks. In ACM International Conference on Multimedia, pages 4417-4425, 2021. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.571, + 0.892, + 0.626 + ], + "angle": 0, + "content": "[51] Florian Tramér, Fan Zhang, Ari Juels, Michael K Reiter, and Thomas Ristenpart. Stealing machine learning models via prediction {APIs}. In USENIX security symposium, pages 601-618, 2016. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.628, + 0.892, + 0.697 + ], + "angle": 0, + "content": "[52] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In International Conference on Advanced Neural Information Processing Systems, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.7, + 0.892, + 0.728 + ], + "angle": 0, + "content": "[53] M Mitchell Waldrop et al. No drivers required. Nature, page 20, 2015. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.73, + 0.892, + 0.772 + ], + "angle": 0, + "content": "[54] Binghui Wang and Neil Zhenqiang Gong. Stealing hyperparameters in machine learning. In IEEE symposium on security and privacy, pages 36-52, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.774, + 0.892, + 0.843 + ], + "angle": 0, + "content": "[55] Jiakai Wang, Aishan Liu, Zixin Yin, Shunchang Liu, Shiyu Tang, and Xianglong Liu. Dual attention suppression attack: Generate adversarial camouflage in physical world. In IEEE/CVF Computer Vision and Pattern Recognition Conference, pages 8565-8574, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.845, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[56] Zhen Wang, Guosheng Hu, and Qinghua Hu. Training noise-robust deep neural networks via meta-learning. In IEEE/CVF Computer Vision and Pattern Recognition Conference, pages 4524–4533, 2020. 3" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "24314" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.133 + ], + "angle": 0, + "content": "[57] Peter Welinder, Steve Branson, Takeshi Mita, Catherine Wah, Florian Schroff, Serge Belongie, and Pietro Perona. Caltech-ucsd birds 200. 2010. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.136, + 0.472, + 0.19 + ], + "angle": 0, + "content": "[58] Hongwei Wen, Jingyi Cui, Hanyuan Hang, Jiabin Liu, Yisen Wang, and Zhouchen Lin. Leveraged weighted loss for partial label learning. In International Conference on Machine Learning, pages 11091-11100, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.192, + 0.472, + 0.247 + ], + "angle": 0, + "content": "[59] Dong-Dong Wu, Deng-Bao Wang, and Min-Ling Zhang. Revisiting consistency regularization for deep partial label learning. In International Conference on Machine Learning, pages 24212-24225, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.249, + 0.472, + 0.318 + ], + "angle": 0, + "content": "[60] Xiaobo Xia, Tongliang Liu, Nannan Wang, Bo Han, Chen Gong, Gang Niu, and Masashi Sugiyama. Are anchor points really indispensable in label-noise learning? In International Conference on Advanced Neural Information Processing Systems, pages 6835-6846, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.32, + 0.471, + 0.375 + ], + "angle": 0, + "content": "[61] Hao Yang, You-Zhi Jin, Zi-Yin Li, Deng-Bao Wang, Xin Geng, and Min-Ling Zhang. Learning from noisy labels via dynamic loss thresholding. IEEE Transactions on Knowledge and Data Engineering, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.376, + 0.471, + 0.444 + ], + "angle": 0, + "content": "[62] Honggang Yu, Kaichen Yang, Teng Zhang, Yun-Yun Tsai, Tsung-Yi Ho, and Yier Jin. Cloudleak: Large-scale deep learning models stealing through adversarial examples. In Network and Distributed System Security Symposium, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.446, + 0.472, + 0.502 + ], + "angle": 0, + "content": "[63] Chiyuan Zhang, Samy Bengio, Moritz Hardt, Benjamin Recht, and Oriol Vinyals. Understanding deep learning (still) requires rethinking generalization. Communications of the ACM, pages 107-115, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.504, + 0.471, + 0.559 + ], + "angle": 0, + "content": "[64] Jiliang Zhang, Shuang Peng, Yansong Gao, Zhi Zhang, and Qinghui Hong. Apmsa: adversarial perturbation against model stealing attacks. IEEE Transactions on Information Forensics and Security, pages 1667-1679, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.56, + 0.471, + 0.615 + ], + "angle": 0, + "content": "[65] Guoqing Zheng, Ahmed Hassan Awadallah, and Susan Dumas. Meta label correction for noisy label learning. In Association for the Advance of Artificial Intelligence, pages 11053-11061, 2021. 3" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.615 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "24315" + } + ] +] \ No newline at end of file diff --git a/2024/Efficient Model Stealing Defense with Noise Transition Matrix/5181a0c8-f9af-47c9-b829-4e394817b1e7_origin.pdf b/2024/Efficient Model Stealing Defense with Noise Transition Matrix/5181a0c8-f9af-47c9-b829-4e394817b1e7_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..803e1b6d2b6a64a9c5950b730e58b92a5a2c7345 --- /dev/null +++ b/2024/Efficient Model Stealing Defense with Noise Transition Matrix/5181a0c8-f9af-47c9-b829-4e394817b1e7_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bda88058b6dfe6b10fbc667190d5fb828a23dbe9b20323fff28783e8d496d256 +size 1154272 diff --git a/2024/Efficient Model Stealing Defense with Noise Transition Matrix/full.md b/2024/Efficient Model Stealing Defense with Noise Transition Matrix/full.md new file mode 100644 index 0000000000000000000000000000000000000000..bacb238f40f8a27ff03854de1d7cdcc550b1f6e3 --- /dev/null +++ b/2024/Efficient Model Stealing Defense with Noise Transition Matrix/full.md @@ -0,0 +1,377 @@ +# Efficient Model Stealing Defense with Noise Transition Matrix + +Dong-Dong $\mathsf{W u}^{1}$ , Chilin $\mathsf{F u}^{2}$ , Weichang $\mathsf{W u}^{2}$ , Wenwen $\mathsf{X i a}^{2}$ + +Xiaolu Zhang $^{2}$ , Jun Zhou $^{2*}$ , Min-Ling Zhang $^{1*}$ + +$^{1}$ School of Computer Science and Engineering, Southeast University, Nanjing, China + +$^{2}$ Ant Group + +{dongdongwu,zhangml}@seu.edu.cn,{chilin.fcl,jiuyue.wwc,wenzhen.xww}@antgroup.com + +{yueyin.zxl, jun.zhoujun}@antfin.com + +# Abstract + +With the escalating complexity and investment cost of training deep neural networks, safeguarding them from unauthorized usage and intellectual property theft has become imperative. Especially the rampant misuse of prediction APIs to replicate models without access to the original data or architecture poses grave security threats. Diverse defense strategies have emerged to address these vulnerabilities, yet these defenses either incur heavy inference overheads or assume idealized attack scenarios. To address these challenges, we revisit the utilization of noise transition matrix as an efficient perturbation technique, which injects noise into predicted posteriors in a linear manner and integrates seamlessly into existing systems with minimal overhead, for model stealing defense. Provably, with such perturbed posteriors, the attacker's cloning process degrades into learning from noisy data. Toward optimizing the noise transition matrix, we proposed a novel bi-level optimization training framework, which performs fidelity on the victim model while the surrogate model adversarially. Comprehensive experimental results demonstrate that our method effectively thwarts model stealing attacks and achieves minimal utility trade-offs, outperforming existing state-of-the-art defenses. + +# 1. Introduction + +Machine learning (ML) models, especially deep neural networks (DNNs), have attained human-level capabilities across various domains, such as pattern recognition tasks [24, 27] and generation tasks [2, 52], becoming indispensable in facilitating convenience for society [53]. With the burgeoning of ML application scenarios and the proliferation of data, training large and sophisticated models and making them publicly accessible has become a new profit trend for compa + +nies. However, considering the steep costs of tuning models, directly divulging model parameters potentially engenders no returns on investment and even adverse societal impacts from malicious use. To alleviate those concerns, prediction APIs have emerged as an alternative by enabling black-box inferences through querying. This allows models to provide services without exposing the underlying parameters and details, and has been harnessed in Machine Learning as a Service (MLaaS) platforms (e.g., Amazon AWS, Microsoft Azure, Google Cloud). + +Unfortunately, recent research has demonstrated that prediction APIs are still vulnerable to intellectual property theft through model stealing attacks [39, 51, 54, 62]. Despite not exposing model details, attackers can query the APIs using carefully-crafted samples or surrogate samples to obtain responded predictions. Through this, they can then train a clone model on the gathered query-response pairs to replicate the victim model's functionality with high accuracy and fidelity, without needing to access the original training data or model architecture [22, 39]. Such a clone model poses the following security threats: (a) acquiring a functionally similar copy of the victim model, illegally circumventing the expensive process of manual data curation and undermining the intellectual property of MLaaS; (b) extracting confidential data through membership inference attacks [47], model inversion attack [6] and property inference attack [7]; (c) conducting adversarial attacks [31, 33, 35, 36, 49, 55] by creating malicious examples based on the white-box substitute model [42]. + +To address such vulnerabilities, considerable researches [3, 5, 9, 17, 29, 50, 64] have focused on thwarting model stealing attacks. These defenses can be categorized into three strategies: detection, perturbation, and post-hoc verification. Detection-based defenses [19, 20, 23, 23] are predicated on the idea that malicious queries exhibit aberrant statistical distributions compared to normal queries. If a query sequence appears anomalous, the defender may discard the requests or corrupt the outputs. However, if an attacker carefully con + +![](images/27b8898ae862573b82ebeb905ba6a3ba9fe7846f0081b6bf335764e4c4751b5c.jpg) +Figure 1. Illustrations of the noise transition matrix serving for noisy label learning and model stealing defense. In noisy label learning, it transforms the predicted posteriors to align with the noisy label. In model stealing defense, it perturbs the predicted posterior to impede attacker training while preserving the original predicted label for benign users. + +structures queries matching the benign distribution, detection techniques will fail to discern the attack, yielding excessive false positives. Perturbation-based defenses [30, 40, 51] can preserve or constraint model accuracy to foil model stealing attacks. The accuracy-preserving defense truncates predicted posteriors to a minuscule fraction of their original size, which may reduce model lucidity and utility for critical applications like autonomous driving. The accuracy-constraint defense [38, 40] introduces perturbation into predicted posteriors, which incurs additional computational burdens or substantially degrading utility. Furthermore, the latency resulting from such perturbation methods leads to extended user response times, making the approach impractical. Post-hoc verification techniques such as watermarking [15] remain useful for post-attack prevention and validation. However, their retroactive nature means they cannot take proactive measures before an attack occurs. Considering inference latency and utility trade-offs, in this paper, we concentrate on a novel and efficient model stealing defense strategy. + +To overcome the limitations of existing defenses, we explore an efficient method to inject noise into predicted posteriors via linear mapping, while preserving utility for benign users. Motivated by memorization effects [11, 63] that neural networks can easily overfit noisy data, thus hampering model generalization, we revisit the utilization of noise transition matrix (NTM) in noisy labels learning [43, 58], where the matrix models the noise process. Unlike related work, we aim to leverage the NTM to introduce noise into predicted posteriors for model defense, as shown in Figure 1. Our goal is to find an optimized NTM that injects noise while retaining original prediction labels with minimal utility trade-offs. Intuitively, regardless of whether the victim model produces extremely confident one-hot vectors or uncertain flat vectors, we can readily derive a uniform NTM where the diagonal elements remain maximal for each row to meet our defense needs. To estimate the NTM with maximum utility, we propose a bi-level optimization method that alternately optimizes it and updates the surrogate network + +weights. Specifically, we first make a one-step-forward virtual optimization of weights, then optimize the matrix guided by the loss with the one-step-forward weights fixed. Finally, we optimize the unrolled weights with the updated matrix. The contributions of this work are as follows: + +- We propose a novel perturbation-based defense that efficiently introduces noise through linear mapping, adding minimal overhead. Furthermore, it deteriorates attacker training to a level equivalent to training on noisy-labeled data with instance-wise weight. +- Our method pioneers using the transition matrix to thwart model stealing attacks. The explicit matrix helps understand noise relationships among classes and makes the defense strategy more interpretable +- A bi-level optimization approach is proposed to optimize the transition matrix, which is model-agnostic and can adapt to different backbone networks. +- Extensive experiments exhibit our method consistently mitigates stealing and outperforms state-of-the-art defenses. + +# 2. Related Work + +Model Stealing Attacks aim to replicate the functionality of a victim model by interrogating its prediction APIs. The attacker sends queries and collects the resultant predictions, typically probability vectors or labels, as query-response pairs. These collected pairs are used to train a clone model that approximates the victim model. Tramèr et al. [51] first proposed the idea of model stealing and demonstrated the feasibility of extracting simple models utilizing queries. Juuti et al. [19] presented extraction attacks leveraging jacobian-based data augmentation [42], synthesizing data by perturbing a small set of in-distribution samples. However, these attacks presume the ability to optimally probe the APIs, which could be detected by monitoring for anomalous query patterns [19, 41]. Therefore, we concentrate on scenarios without carefully engineered input queries or sequences. One classic scenario to achieve this is by utilizing knowledge distillation techniques [12], where some studies exhibit model stealing is viable using weakly correlated [39] or random [25] queries, even when the teacher model declares itself undistillable [13, 28]. + +Model Stealing Defenses can be taxonomized into three primary strategies: detection, perturbation, and post-hoc verification. Detection strategies [20, 21] aim to identify malicious queries or nefarious users by scrutinizing the queries received by the victim model. For example, Kariyappa and Qureshi [20] regarded malicious queries as out-of-distribution samples and devised an adaptive misinformation (AM) defense to mislead attackers with modified outputs. However, such detection strategies rely on strong assumptions about adversarial query patterns that may not reflect reality. Perturbation strategies mitigate model stealing + +attacks by returning perturbed posterior, which can be further subcategorized into accuracy-preserving [30] and accuracy-constrained defenses [29, 38, 40]. Accuracy-preserving defenses guarantee that the maximal class of the perturbed predictions matches the original prediction. The most straightforward such defense is truncating the posterior or only outputting the predicted label [51]. However, this diminishes utility for benign users, precluding specific applications and harming external transparency. In contrast, accuracy-constrained defenses permit a higher level of perturbation, trading off benign accuracy for enhanced security against model stealing attacks. In general, existing defense strategies that poison the attacker's update gradient [38, 40] are computationally expensive, even if they yield a comparative defense. Post-hoc verification, including watermarking [15, 32, 50], dataset inference [37], and proof of learning [16], embeds concealed patterns in the model during training [1] or inference [50] to distinguish a stolen model. Unfortunately, these approaches only work after a model theft, requiring the model owner to obtain partial or full access to the stolen model. This means if an attacker uses the stolen model as a proxy to attack the victim model, it is impossible for the model owner to protect themselves with these defenses. Overall, while existing ante-hoc defensive approaches mitigate knowledge leakage, they rely on strong assumptions or incur high computational costs during inference. + +Noise Transition Matrix (NTM) is often utilized to solve classification with noisy labels or ambiguous labeled data [10, 46, 61], where the matrix reflects probabilities of true label flipping to the other label. A typical application is loss correction [43], which modifies the loss of each example by multiplying the estimated matrix by the model output. For NTM estimation, there are two main approaches. One is a two-step solution that pre-estimates the matrix with anchor point assumptions, then uses it to train the classifier [56, 65]. For instance, backward correction [43] initially approximates the NTM using the softmax output of the model trained without correction. It then retransmits the model while correcting the original loss based on the estimated matrix. The other jointly estimates the matrix and classifier parameters in a unified framework without employing anchor points [18, 48, 60]. Differently, our work aims to estimate and apply the transition amtrix to intentionally inject noise into predicted posteriors, thus defending against model stealing. + +# 3. Threat Model + +We consider a model stealing scenario involving a defender and an attacker interacting on image classification tasks. The defender grants access to its black-box model through prediction API. Leveraging the API, the attacker collects query-response pairs, aiming to mimic the defender model (victim model) functionalities by training a clone model on + +this dataset. To evade detection defenses [19], we assume the attacker adopts non-adaptive querying strategies without anomalous patterns [39, 42]. Therefore, we consider the knockoff nets attack without adaptive querying [39], which resembles knowledge distillation [12] where the victim model distills knowledge into the attacker model. + +# 3.1. Attacker's Objective + +Let $\theta_{f}$ and $\theta_{h}$ denote the parameters of the defender's model $f$ and attacker's clone model $h$ , respectively. The attacker prepares a query dataset $\mathcal{D}_{\mathrm{query}}$ , and sends queries $\pmb{x} \in \mathcal{D}_{\mathrm{query}}$ in random order to obtain the defender's posterior $f(\pmb{x}; \pmb{\theta}_{f})$ from the prediction API. The classification loss of the attacker's clone model on the query $\pmb{x}$ is $\mathcal{L}_{\mathrm{kna}} = -\sum_{i} f(\pmb{x}; \pmb{\theta}_{f})_{i} \log h(\pmb{x}; \pmb{\theta}_{h})_{i}$ . The attacker aims to minimize the classification loss on the defender's test set. However, since the test set is unavailable, the attacker's learning objective becomes minimizing the loss based on the collected query-response dataset $\mathcal{D}_{\mathrm{pair}} = \{(x, f(x; \pmb{\theta}_{f})) | x \in \mathcal{D}_{\mathrm{query}}\}$ . Therefore, the attacker's objective can be formulated as: + +$$ +\min _ {\boldsymbol {\theta} _ {h}} \mathbb {E} _ {\boldsymbol {x} \sim \mathcal {D} _ {\text {p a i r}}} \left[ \mathcal {L} _ {\mathrm {k n a}} \left(h (\boldsymbol {x}; \boldsymbol {\theta} _ {h}), f (\boldsymbol {x}; \boldsymbol {\theta} _ {f})\right) \right] \tag {1} +$$ + +Query Distribution. In practice, due to the knowledge limitation of attackers about the defender's training set, the collected query dataset $\mathcal{D}_{\mathrm{query}}$ may be similar [30, 51] or unrelated [20, 40] to the defender's training distribution. Hence, we consider both possibilities in our threat model. We refer to attackers as distribution-aware if their queries share semantic content with the defender's training distribution. Attackers are referred to as knowledge-limited if their queries are semantically disjoint from the defender's training distribution. + +# 3.2. Defender's Objective + +In defenses against model stealing, the defender aims at preventing its model functionality from being stolen. In real-world scenarios, attackers may disguise themselves as benign users, executing attacks anytime. Therefore, the defender must persistently mitigate model stealing attempts while preserving the API's utility for legitimate users. Specifically, the defender needs to keep its own small classification loss, while maximizing the classification loss of the attacker's cloned model $\theta_h^*(\theta_f)$ obtained from Eq. (1) on the test set $\mathcal{D}_{\mathrm{test}}$ . The defender's objective can be formulated as: + +$$ +\min _ {\boldsymbol {\theta} _ {f}} \mathbb {E} _ {(\boldsymbol {x}, y) \sim \mathcal {D} _ {\text {t e s t}}} [ \mathcal {L} _ {\mathrm {C E}} (f (\boldsymbol {x}; \boldsymbol {\theta} _ {f}), y) ] \tag {2} +$$ + +$$ +\operatorname {s. t.} \max _ {\mathbf {x}, y) \sim \mathcal {D} _ {\text {t e s t}}} [ \mathcal {L} _ {\mathrm {C E}} (h (\boldsymbol {x}; \boldsymbol {\theta} _ {h} ^ {*} (\boldsymbol {\theta} _ {f})), y) ], +$$ + +where $\mathcal{L}_{\mathrm{CE}}$ means the cross-entropy loss function. + +# 4. Approach + +To develop an efficacious and inference-efficient defense against model stealing, we propose an Efficient Model de + +fense with noise transition MMatrix (EMMA) to mitigate model stealing attacks. As depicted in the "Defense phase" of Figure 2, the attacker iteratively queries the victim model using unlabeled data to obtain posteriors via the prediction API. The victim model receives these queries and outputs corresponding predicted probability vectors. These posteriors then multiply the optimized matrix $T^n$ , becoming perturbed posteriors that remain useful to benign users, and returned to the attacker. However, if the attacker collects these query-responses as training data to reconstruct a clone model, in that case, this dataset is equivalent to noisy labeled data with instance-wise weight, thus impeding cloning model training. To optimize the transition matrix with maximum utility, as shown in the "Optimizing phase", we devise a bi-level optimization between the victim and surrogate model which is devised to represent the attacker model. In each round, given initial $T^t$ and surrogate weights $\theta_g^t$ , the optimized matrix $T^{t+1}$ stems from two gradient descents (GD) on the attack loss and defense loss, respectively. The surrogate model then updates through a GD on the attack loss. This process loops until reaching the termination condition, yielding the final optimized matrix $T^n$ . In this section, we first elucidate the defense mechanism of incorporating the transition matrix during the defensive phase. Then, a bi-level optimization approach is expounded to alternately optimize the matrix and update the surrogate model. + +# 4.1. The underlying mechanism + +To understand how the transition matrix defends against attacks, we analyze the classification risks of two scenarios: "learning on noisy labels" and "learning on query-response pairs with perturbed posteriors". It is found that the two classification risks differ only by an instance-wise weight coefficient multiplied to the latter. Let $\tilde{y} \in \mathbf{Y}$ denote the noisy label, where $\tilde{y}$ is selected based on the transition matrix and the true label. Theorem 1 formalizes the defensive mechanism. + +Theorem 1. Let $\mathcal{L}(\cdot, \cdot)$ denote the classification loss function. With sufficiently many queries, the classification risk $\Re(h)$ of learning with noisy labels is $\mathbb{E}_{(\boldsymbol{x}, \tilde{\boldsymbol{y}})}[\mathcal{L}(h(\boldsymbol{x}), \tilde{\boldsymbol{y}})]$ . The classification risk $\Re_T(h)$ of learning with query-response pairs is $\mathbb{E}_{(\boldsymbol{x}, \tilde{\boldsymbol{y}})}[q(\boldsymbol{x}, \tilde{\boldsymbol{y}}) \cdot \mathcal{L}(h(\boldsymbol{x}), \tilde{\boldsymbol{y}})]$ , where $q(\boldsymbol{x}, \tilde{\boldsymbol{y}})$ denotes the instance-wise weight, proportional to the closeness between the predicted posterior and the true label. + +The proof can be found in Appendix 8. Theorem 1 demonstrates that training a clone model on such query-response pairs with perturbed posteriors is equivalent to training it on noisy-labeled data with instance-wise weights. It can be further deduced that the more closely the victim model's predicted posteriors match the true label, the more this training process resembles learning with noisy labels. In addition, the clone model's generalization can be further degraded by + +![](images/e62144276b3f8110e06bbcbef99d4c3cab0b71c7c0bfc522368791929c71f5d9.jpg) +Figure 2. An illustration of "Optimizing phase" and "Defense phase" in the framework. "Optimizing phase" demonstrates the process of updating the noise transition matrix $T$ and surrogate network weights $\theta_g$ alternatively. "Defensive phase" shows the interaction process where the attacker tries to steal the functionality from the defender. + +increasing the transition matrix noise rate or exploring more difficult noise type through inter-class correlations, such as using class-conditional noise instead of random noise. + +# 4.2. The overall training framework + +Since the attacker's network is unknown, we cannot directly measure its cloning performance. Instead, we represent the attacker with a surrogate model $g$ parameterized by $\theta_g$ and design optimal perturbations for it. By using the surrogate model to simulate the cloning process, we can continuously optimize the transition matrix during this process to minimize the utility trade-off between attacker performance and defender utility. To achieve this, we first introduce the respective objective functions for the attacker and defender. Then, a bi-level optimization approach is proposed. + +Given the perturbed posterior $f(\pmb{x}; \pmb{\theta}_f)T$ , the attacker's objective function is defined as the loss on the gathered query-response pairs, formulated as: + +$$ +\mathcal {L} _ {\text {a t t a c k}} \left(\boldsymbol {\theta} _ {g}, T\right) = \mathbb {E} _ {\boldsymbol {x} \sim \mathcal {D} _ {\text {p a i r}}} \left[ \mathcal {L} _ {\text {k n a}} \left(g (\boldsymbol {x}; \boldsymbol {\theta} _ {g}), f (\boldsymbol {x}; \boldsymbol {\theta} _ {f}) T\right) \right]. \tag {3} +$$ + +Notably, the optimization of $\theta_{g}$ is influenced by $T$ . To defend against model stealing, we would like $T$ to regularize $\theta_{g}$ to have poor generalizability. To achieve this, we instantiate a misinformation loss that inhibits the attacker model's outputs on the victim model's original predicted label $k = \arg \max_{i} f(\pmb{x}, \pmb{\theta}_{f})_{i}$ . Formally, we derive the following modified negative log-likelihood loss: + +$$ +\mathcal {L} _ {\operatorname {m i s}} \left(\boldsymbol {\theta} _ {g}\right) = \mathbb {E} _ {\boldsymbol {x} \sim \mathcal {D} _ {\text {p a i r}}} \left[ - \log \left(1 - g (\boldsymbol {x}; \boldsymbol {\theta} _ {g}) _ {k}\right) \right]. \tag {4} +$$ + +The underlying principle is quite simple: The log-likelihood of the original prediction label on the output should be minimized. Despite its simple form, optimizing this loss has been empirically shown to be an effective defense against model stealing attacks [20, 59]. + +Moreover, to maintain fidelity for the victim model, a straightforward strategy is aligning the perturbed output with the predicted label $k$ . This is implemented by minimizing the divergence between the output and the predicted label, formulated as: + +$$ +\mathcal {L} _ {\text {a l i g n}} (T) = \mathbb {E} _ {\boldsymbol {x} \sim \mathcal {D} _ {\text {p a i r}}} [ - \log (f (\boldsymbol {x}; \boldsymbol {\theta} _ {f}) T) _ {k} ]. \tag {5} +$$ + +In summary, the overall objective function of the defender can be represented as the following defense loss: + +$$ +\mathcal {L} _ {\text {d e f e n s e}} \left(\boldsymbol {\theta} _ {g}, T\right) = \lambda \cdot \mathcal {L} _ {\text {a l i g n}} (T) + (1 - \lambda) \cdot \mathcal {L} _ {\text {m i s}} \left(\boldsymbol {\theta} _ {g}\right), \tag {6} +$$ + +where the multiplicative factor $\lambda$ is used to balance the contribution of these two loss terms. + +Bi-level optimization. Given the defense loss $\mathcal{L}_{\mathrm{defense}}$ , the surrogate model weights $\theta_{g}$ need to be updated first before optimizing the transition matrix $T$ . For a mini-batch sample, we propose that surrogate model weights could be treated as a latent variable and optimized alternatively with the matrix. This yields the following bi-level optimization problem: + +$$ +\underset {T} {\arg \min } \mathcal {L} _ {\text {d e f s e n s e}} \left(\boldsymbol {\theta} _ {g} ^ {*}, T\right) +$$ + +$$ +\text {s u b j e c t} \theta_ {g} ^ {*} = \underset {\theta_ {g}} {\arg \min } \mathcal {L} _ {\text {a t t a c k}} \left(\theta_ {g}, T\right) \tag {7} +$$ + +Bi-level optimization problem has been proved strongly NP-hard [14], so getting such an exact solution for problem (7) is impossible in polynomial time. Fortunately, the inner optimization problem in (7) can be transformed into an optimized result $\theta_g(T)$ as a function of $T$ . Then $\theta_g(T)$ can then substitute $\theta_g$ in the outer optimization to acquire an objective solely in terms of $T$ . This allows approximating the solution using gradient descent. Therefore, we employ an alternating optimization strategy to approximately solve this problem. In each iteration $\theta_g$ and $T$ are updated on-the-fly in an alternating manner. + +Optimizing the transition matrix. In step $t + 1$ of optimizing $T^t$ , guided by the attack loss, surrogate model weights $\theta_g^t$ are updated in a one-step-forward manner to obtain $\theta_g^{t + 1}$ . Specifically, with the fixed $T^t$ , then $\theta_g^t$ is updated via gradient descent with learning rate $\alpha$ as: + +$$ +\tilde {\boldsymbol {\theta}} _ {g} ^ {t + 1} \left(T ^ {t}\right) = \boldsymbol {\theta} _ {g} ^ {t} - \alpha \cdot \bigtriangledown_ {\boldsymbol {\theta} _ {g} ^ {t}} \mathcal {L} _ {\text {a t t a c k}}. \tag {8} +$$ + +Note that the surrogate model $g$ does not actually update in this step, but makes preparations for estimating $T^{t + 1}$ . We then optimize $T^t$ with $\tilde{\theta}_g^{t + 1}$ fixed under the defense loss. The motivation is that we would like to find a $T^{t + 1}$ that achieves + +# Algorithm 1 The overall training procedure + +Input: Query dataset $\mathcal{D}_{\mathrm{query}}$ + +The surrogate model $g$ and its parameters $\theta_{g}$ + +The noise transition matrix $T$ . + +The number of iterations $I$ ; + +The multiplicative factor $\lambda$ ; + +# Procedure: + +1: Initialize $T$ and $\theta_{g}$ +2: for $i = 1$ to $I$ do +3: Fetch a random batch $\mathcal{B}$ from $\mathcal{D}_{\mathrm{query}}$ +4: Calculate the one-step-forward surrogate model weights via Eq. (8); +5: Update the transition matrix via Eq. (9-12); +6: Update surrogate model weights via Eq. (13); +7: end for + +Output: Learned noise transition matrix. + +minimal utility trade-offs between the defender and attacker. The transition matrix $T^{t}$ is updated via gradient descent with learning rate $\beta$ : + +$$ +w ^ {t + 1} = T ^ {t} - \beta \cdot \bigtriangledown_ {T ^ {t}} \mathcal {L} _ {\text {D e f s e n s e}}. \tag {9} +$$ + +Apply chain rule to Eq. (9), we can get: + +$$ +\begin{array}{l} w ^ {t + 1} = T ^ {t} - \beta \cdot \left\{\lambda \cdot \nabla_ {T ^ {t}} \mathcal {L} _ {\text {a l i g n}} \right. \\ + (1 - \lambda) \cdot \bigtriangledown_ {\tilde {\boldsymbol {\theta}} _ {g} ^ {t + 1}} \mathcal {L} _ {\mathrm {m i s}} \cdot \left(- \alpha \bigtriangledown_ {\boldsymbol {\theta} _ {g} ^ {t}, T ^ {t}} ^ {2} \mathcal {L} _ {\text {a t t a c k}}\right) \}. \tag {10} \\ \end{array} +$$ + +Note that $w^{t + 1}$ is not the final transition matrix, as its entries may not satisfy the properties of a valid transition matrix. Specifically, all elements must be non-negative and each row must be normalized. To address this, we first replace any negative elements in $w^{t + 1}$ with zeros via: + +$$ +\tilde {w} ^ {t + 1} = \max \left(w ^ {t + 1}, 0\right). \tag {11} +$$ + +Then, each row is normalized to get the final valid transition matrix. For the $j^{\mathrm{th}}$ row of $\tilde{w}^{t + 1}$ , which indicates the probabilities of class $j$ transitioning to other classes, a normalization is performed to obtain the final matrix $T_{j,i}^{t + 1}$ : + +$$ +T _ {j, i} ^ {t + 1} = \frac {\tilde {w} _ {j , i} ^ {t + 1}}{\sum_ {i} \tilde {w} _ {j , i} ^ {t + 1} + \delta \left(\sum_ {i} \tilde {w} _ {j , i} ^ {t + 1}\right)}, \delta (a) = \left\{ \begin{array}{l} 1, \text {i f} a = 0 \\ 0, \text {i f} a \neq 0 \end{array} \right. \tag {12} +$$ + +where $\delta (\cdot)$ is used to avoid division by 0. + +Updating surrogate model weights. After optimizing the matrix, we conduct the surrogate model update by keeping $T^{t + 1}$ fixed and minimizing the attack loss. New surrogate model weights $\theta_{g}^{t + 1}$ are obtained via gradient descent with learning rate $\gamma$ : + +$$ +\boldsymbol {\theta} _ {g} ^ {t + 1} = \boldsymbol {\theta} _ {g} ^ {t} - \gamma \cdot \bigtriangledown \boldsymbol {\theta} _ {g} ^ {t} \mathcal {L} _ {\text {a t t a c k}}. \tag {13} +$$ + +The whole training procedure is summarized in Algorithm 1. + +
MethodCIFAR100 → CIFAR10CIFAR10 → CIFAR100Caltech256 → CUB200
ΔAcc=1ΔAcc=2ΔAcc=5ΔAcc=1ΔAcc=2ΔAcc=5ΔAcc=1ΔAcc=2ΔAcc=5
Random84.1777.2174.2143.2339.8132.1233.5031.1428.01
AM77.1072.9357.5639.8232.9819.3626.1721.2714.31
MAD80.5777.5469.0736.5632.5926.6537.1635.2331.50
NASTY76.6574.3164.1938.5333.6126.9531.0628.2324.70
GARD²77.4273.8871.4041.2139.2935.9435.6933.6732.34
MIN-IP78.8873.4070.6642.3338.6734.2634.7332.2628.94
EMMA61.7952.9142.2128.0118.3413.3719.1015.7913.11
ImageNet-C10 → CIFAR10ImageNet-C100 → CIFAR100ImageNet-CUB200 → CUB200
ΔAcc=1ΔAcc=2ΔAcc=5ΔAcc=1ΔAcc=2ΔAcc=5ΔAcc=1ΔAcc=2ΔAcc=5
Random88.9988.1887.0763.1661.7658.6446.5844.0640.49
AM87.5584.6076.7260.5756.0947.1045.1239.6530.15
MAD88.6785.7078.4458.7456.2350.8550.9848.5744.47
NASTY87.3385.3381.6859.7956.0548.9546.5643.2337.76
GARD²87.2986.0685.9358.3956.7153.6350.2747.5543.11
MIN-IP87.4086.8184.4561.1756.3551.9850.0946.4542.06
EMMA84.1080.6077.5256.7349.1831.4836.3533.1026.33
+ +Table 1. Accuracy (%) of attacker model on defender's test set under knowledge-limited (top row) and distribution-aware (bottom row) attack scenarios with varying budgets. The best result in each column is highlighted in bold. + +
Eval DataDefender AccuracyAttacker Accuracy
KLDA
CIFAR1095.33 ± 0.1385.17 ± 0.8391.78 ± 0.19
CIFAR10075.96 ± 0.0951.95 ± 1.1266.82 ± 0.16
CUB20081.44 ± 0.4648.93 ± 0.3762.12 ± 0.22
+ +Table 2. Accuracy $(\%)$ of the defender model and the attacker model with no defense applied. + +# 5. Experiments + +To demonstrate the effectiveness of our proposed defense method, we conduct experiments on multiple image datasets and present the results in this section. + +# 5.1. Experiment Setup + +Datasets. We employ three datasets for evaluation: CIFAR10, CIFAR100 [26] and CUB200 [57]. For each dataset, we explore two attacker scenarios - knowledge-limited (KL) and distribution-aware (DA). The KL scenario uses query sets of CIFAR100, CIFAR10, and Caltech256 [8], paired with CIFAR10, CIFAR100, and CUB200 evaluation datasets respectively. The DA scenario employs tailored query sets from ImageNet-1K [4] manually selected by prior work [38], specifically ImageNet-C10, ImageNet-C100, and ImageNet-CUB200 containing 183,763, 161,653, and 30,000 examples matched to the evaluation datasets, with their further details provided in the Appendix 9. + +Compared methods. We compare our method against the following six model stealing defense baselines: (1) No Defense, a naive approach where no defense is attempted. (2) + +Random, an interpolation-based method that perturbs the defender's predicted posterior by interpolating it with a one-hot posterior vector, where the index of single non-zero entry is randomly selected from the non-argmax prediction labels. (3) AM [20], a detection-based method that identifies out-of-distribution queries and interpolates with posteriors from a misinformation network, where the interpolation weight is denoted by the confidence of the argmax prediction label. (4) MAD [40], an accuracy-constrained method that adds controlled noise to original posteriors so that it maximizes the angular deviation between the perturbed and original gradient signal on a surrogate model. The surrogate model is randomly initialized. (5) NASTY [34], a self-undermining knowledge distillation approach which trains a nasty teacher network by maximizing the difference between the output of the nasty teacher and a normal pretrained network. (5) $GARD^2$ [38], a coordinated gradient redirection defense method, which perturbs original posteriors by setting the target gradient direction to be the all-ones vector. This redirects the attacker to update towards a consistent target direction. (6) MIN-IP [38], an uncoordinated gradient redirection defense method, which perturbs original posteriors by setting the target gradient direction opposite the original gradient direction, redirecting the update gradient of attackers opposite to the original gradients. The surrogate model of MIN-IP and $GARD^2$ train on the attacker's queries with early stopping after 10 epochs. + +For fair comparisons, we basically use the same network architecture, learning rate, optimizer, and augmentation strategy across all the comparing methods. + +Implementation. Our experiment pipeline has three stages: + +![](images/f22df67711c7e2a03a84d40efbea92433dd69078620db61350f83067e53fae56.jpg) +(a) Heatmap of transition matrix. + +![](images/1179097a99be1a5469ca28c057017ddde9033864b7c0a9b836c7feb0b58dcae8.jpg) +(b) Heatmap of transition matrix. + +![](images/c9df9a5cb987d1faf76f39ed786414a0649152acfc0e99f4e72ff535f22715be.jpg) +(c) Average predicted posterior + +![](images/271bc120cbe6b8c77faeb5962cc6cddf7a388916675e414315814784e1d0fdcf.jpg) +(d) Average perturbed posterior + +![](images/d1373fed305d5394c88c3e75fbe04b70003b7b716967e5b52f80a81f39625b03.jpg) +Figure 3. Qualitative analysis of the transition matrix. (a-b): Heatmaps of optimized transition matrixes for distribution-aware attacks (a) Imagenet-C10→CIFAR10 and (b) Imagenet-C100→CIFAR100, respectively. (c-d): Average (c) predicted and (d) perturbed posteriors across the query set of (top) Imagenet-C10 and (bottom) Imagenet-C100 datasets. +Figure 4. Visualizations of predicted and perturbed posteriors. + +First, the defender model trains on each evaluation dataset. Second, defense methods generate perturbed posteriors for query dataset and evaluation dataset at various defense budgets. Finally, the attacker model trains on the collected query-response pairs. We denote attacks with query data $\mathcal{Q}$ and evaluation data $\mathcal{D}$ as $\mathcal{Q} \to \mathcal{D}$ . + +We use ResNet34 for CIFAR10/100 and ResNet50 pretrained on ImageNet for CUB200. For all models training, we utilize SGD as the optimizer with a momentum of 0.9 and weight decay of 5e-4 for 100 epochs. Initial learning rates are set as 0.01 for CUB200 and 0.1 for other datasets, with a cosine annealing schedule. For optimizing the transition matrix, the learning rate $\gamma$ is set as 0.3 for CIFAR10 and 1.0 for other datasets. The surrogate model's one-step-forward learning rate $\alpha$ and the actual learning rate $\beta$ are set the same value as specified in the model training settings. Additionally, an early stopping mechanism is set to terminates optimization when the average change in $T$ between consecutive epochs falls below a specified threshold. The threshold is 1e-2 for CIFAR10 and 1e-4 for other datasets. Furthermore, the balancing factor $\lambda$ is adjusted to achieve different utility budgets. Recent work [38] proposed a novel approach to initialize the surrogate model. They suggested to train the surrogate model through distilling from the victim model with early stopping. This would make it more representative of the actual attacker's behavior. Following this approach, we initialize our surrogate model by training it on the query dataset with the same early stopping mechanism. + +# 5.2. Experiment Results + +Table 2 presents the accuracy of the defender model and the attacker model, without any defense applied under two different query distributions. This shows the upper bound for the usage performance of the defender and stealable performance of the attacker. Table 1 compares the accuracy of an attacker's clone model at different defender accuracy drop budgets $\Delta$ Acc. The budget refers to the average decrease in accuracy the defender causes across the test set by adding perturbations. Corresponding visualizations of these results are provided in the Appendix 10. As shown, our defense method achieves the lowest utility trade-off across budgets. For example, in the CIFAR100 $\rightarrow$ CIFAR10 attack scenario, our method reduces the attacker model's accuracy by $23\%$ while only decreasing the defender's accuracy by $1\%$ . + +# 5.3. Qualitative Analysis + +Perturbed posteriors make cloning difficult. We visualize the predicted and perturbed posteriors using t-SNE in Figure 4. The perturbed posteriors clearly exhibit degraded inter-class distances compared to the predicted posteriors. This degradation obfuscates the decision boundaries between classes, making it difficult for the attacker to clone the defender model by imitating the posteriors. Notably, such effective defenses are achieved solely through linear mapping of the posteriors, demonstrating the simplicity and efficiency of our proposed noise perturbation strategy. + +The transition matrix induces challenging noise. We analyze the defense details of two attack scenarios: Imagenet-C10→CIFAR10 and Imagenet-C100→CIFAR100 in Figure 3. Specifically, Figure 3 (a-b) visualizes corresponding optimized transition matrices. It can be observed that the diagonal elements are maximized for fidelity, while latent inter-label relationships are also exploited. As shown the optimized matrix in Figure 3 (a), despite it causes a $2.35\%$ drop in the defender's accuracy on the test set, the optimization makes it have a noise rate as high as $77\%$ . This negatively affects the cloning process of the attacker, reducing the cloned model's accuracy by $14\%$ . + +In Figures 3 (c-d), we visualize the average posterior of predicted and perturbed predictions. Given that ImageNet-C10/C100 follow long-tail distributions as illustrated in appendix 9, the predicted posteriors also exhibit long-tailed shapes as expected. However, the perturbations significantly increase the confidence of tail classes, distorting the true inter-class relationships. This distortion of class relationships also makes stealing more difficult for the attacker. + +# 5.4. Ablation Study + +Effectiveness of optimized transition matrix. To validate the effectiveness of the optimized transition matrix, we compare the defense performance between it and the uniform transition matrix with an identical noise rate. The results in Table 3 show that the optimized transition matrix can further reduce the attacker's accuracy while maintaining the defender's accuracy. This demonstrates that the optimized matrix provides improved defense over the uniform matrix. Overall, this shows that capturing the inter-class relationships during optimization contributes to generating a matrix that is more adversarial against the attacker. + +
Attack ScenariosOptimized matrixUniform matrix
DefenderAttackerDefenderAttacker
CIFAR100→CIFAR1093.87±0.1432.41±7.5494.81±0.0671.43±2.26
CIFAR10→CIFAR10074.48±0.2515.65±0.7274.75±0.0735.67±3.44
Caltech256→CUB20080.45±0.1318.25±0.4380.89±0.2230.15±0.21
+ +Table 3. Defender and attacker accuracy $(\%)$ with optimized and uniform transition matrix under attack scenarios for CIFAR10, CIFAR100 and CUB200 evaluation datasets. + +Effects of the number of queries. We also verify the effectiveness of our defense against attacks with varying number of queries in Figure 5. The EMMA-top1 result is depicted as a reference, which denotes the victim model only outputting the predicted label rather than the full posterior. As formalized in Theorem 1, when the victim model assigns full confidence to the true label, the attacker's cloning task degrades to learning from noisy labels. Thus, EMMA-top1 signifies the optimal defense under the same transition matrix. EMMA defense consistently forces degradation of the attacker accuracy across all query set sizes, remaining close to the EMMA-top1. As the query increases, the attacker model obtained from the victim model suffers more accuracy drops, revealing that our defense induces more significant detrimental impacts when the attacker uses more queries. + +Transferability across attack scenarios. To explore the transition matrix's transferable defense capabilities, we test it against an attack type different from the one it was optimized for. As shown in Figure 6, we use the matrix optimized under DA attack to defend against KL attack, denoted as "DA $\rightarrow$ KL". Compared to a matrix specifically optimized for KL attack, denoted as "KL $\rightarrow$ KL", it still maintained + +![](images/ed201633e1a9d3cebbb1470398daf17ce8a00e5ea90ce7abce80b21605326866.jpg) + +![](images/9028e23d26bd00004128021321af4a80f66f06940570ffcbe45544bf452a7585.jpg) + +Figure 5. Effects of the number of queries. +![](images/b3db8ce82356725ea57631298e973ca3391674ef2bbc731709eec883175269c5.jpg) +No defense EMMA EMMA-top1 + +![](images/68222c69d426003e06e4838b38ac9bc999b9b7b32a9a33692a19320a5d9c0b34.jpg) + +considerable defensive efficacy, exhibiting the optimized transition matrix's robust transferable defense capabilities across different attack types. + +![](images/293ba659bc74a9176cb97b1b15b3f5c9abe3d9f56361f9a94d345cd9f17369d0.jpg) +Figure 6. Transferability across attack scenarios. + +![](images/7709f982bd19a6cdf67251b1665ad488ccef7430bbad1b683ea188d84c45cc6c.jpg) + +Due to the space limitation, additional experiments of our method are provided in Appendix. + +# 6. Conclusion + +In this work, we revisited the noise transition matrix in a novel way by applying it to model stealing defense for the first time. This deteriorates the attacker's cloning process to a level equivalent to training on noisy labels, hampering attacker model generalization. To optimize the transition matrix, we proposed a novel bi-level optimization training framework that incorporates fidelity training for the victim model and adversarial training for the surrogate model. Experimental results under various attack scenarios showed the effectiveness and robustness of our proposed defense method compared to current state-of-the-art methods, achieving strong defense with minimal utility trade-off. + +# 7. Acknowledgments + +The authors wish to thank the anonymous reviewers for their helpful comments and suggestions. This work was supported by the National Science Foundation of China (62225602), and the Big Data Computing Center of Southeast University. + +# References + +[1] Yossi Adi, Carsten Baum, Moustapha Cisse, Benny Pinkas, and Joseph Keshet. Turning your weakness into a strength: Watermarking deep neural networks by backdoorsing. In USENIX Security Symposium, pages 1615-1631, 2018. 3 +[2] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. In International Conference on Advanced Neural Information Processing Systems, pages 1877-1901, 2020. 1 +[3] Yanjiao Chen, Rui Guan, Xueluan Gong, Jianshuo Dong, and Meng Xue. D-dae: Defense-penetrating model extraction attacks. In IEEE Symposium on Security and Privacy, pages 382-399, 2023. 1 +[4] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In IEEE/CVF Computer Vision and Pattern Recognition Conference, pages 248-255, 2009. 6 +[5] Adam Dziedzic, Muhammad Ahmad Kaleem, Yu Shen Lu, and Nicolas Papernot. Increasing the cost of model extraction with calibrated proof of work. In International Conference on Learning Representations, 2022. 1 +[6] Matt Fredrikson, Somesh Jha, and Thomas Ristenpart. Model inversion attacks that exploit confidence information and basic countermeasures. In ACM SIGSAC conference on computer and communications security, pages 1322-1333, 2015. 1 +[7] Karan Ganju, Qi Wang, Wei Yang, Carl A Gunter, and Nikita Borisov. Property inference attacks on fully connected neural networks using permutation invariant representations. In ACM SIGSAC conference on computer and communications security, pages 619-633, 2018. 1 +[8] Gregory Griffin, Alex Holub, and Pietro Perona. Caltech-256 object category dataset. 2007. 6 +[9] Jun Guo, Aishan Liu, Xingyu Zheng, Siyuan Liang, Yisong Xiao, Yichao Wu, and Xianglong Liu. Isolation and induction: Training robust deep neural networks against model stealing attacks. In ACM International Conference on Multimedia, 2023. 1 +[10] Bo Han, Quanming Yao, Tongliang Liu, Gang Niu, Ivor W Tsang, James T Kwok, and Masashi Sugiyama. A survey of label-noise representation learning: Past, present and future. arXiv preprint arXiv:2011.04406, 2020. 3 +[11] Bo Han, Quanming Yao, Xingrui Yu, Gang Niu, Miao Xu, Weihua Hu, Ivor W. Tsang, and Masashi Sugiyama. Robust training of deep neural networks with extremely noisy labels. In International Conference on Advanced Neural Information Processing Systems, page 4, 2020. 2 +[12] Geoffrey Hinton, Oriol Vinyals, and Jeff Dean. Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531, 2015. 2, 3 +[13] Surgan Jandial, Yash Khasbage, Arghya Pal, Vineeth N Balasubramanian, and Balaji Krishnamurthy. Distilling the undistillable: Learning from a nasty teacher. In European Conference on Computer Vision, pages 587-603, 2022. 2 + +[14] Robert G Jeroslow. The polynomial hierarchy and a simple model for competitive analysis. Mathematical programming, pages 146-164, 1985. 5 +[15] Hengrui Jia, Christopher A Choquette-Choo, Varun Chandrasekaran, and Nicolas Papernot. Entangled watermarks as a defense against model extraction. In USENIX Security Symposium, pages 1937-1954, 2021. 2, 3 +[16] Hengrui Jia, Mohammad Yaghini, Christopher A Choquette-Choo, Natalie Dullerud, Anvith Thudi, Varun Chandrasekaran, and Nicolas Papernot. Proof-of-learning: Definitions and practice. In IEEE Symposium on Security and Privacy, pages 1039-1056, 2021. 3 +[17] Wenbo Jiang, Hongwei Li, Guowen Xu, Tianwei Zhang, and Rongxing Lu. A comprehensive defense framework against model extraction attacks. IEEE Transactions on Dependable and Secure Computing, 2023. 1 +[18] Ishan Jindal, Matthew Nokleby, and Xuewen Chen. Learning deep networks from noisy labels with dropout regularization. In IEEE International Conference on Data Mining, pages 967-972, 2016. 3 +[19] Mika Juuti, Sebastian Szyller, Samuel Marchal, and N Asokan. Prada: protecting against dnn model stealing attacks. In IEEE European Symposium on Security and Privacy, pages 512-527, 2019. 1, 2, 3 +[20] Sanjay Kariyappa and Moinuddin K Qureshi. Defending against model stealing attacks with adaptive misinformation. In IEEE/CVF Computer Vision and Pattern Recognition Conference, pages 770-778, 2020. 1, 2, 3, 5, 6 +[21] Sanjay Kariyappa, Atul Prakash, and Moinuddin K Qureshi. Protecting dnns from theft using an ensemble of diverse models. In International Conference on Learning Representations, 2020. 2 +[22] Sanjay Kariyappa, Atul Prakash, and Moinuddin K Qureshi. Maze: Data-free model stealing attack using zeroth-order gradient estimation. In CVPR, 2021. 1, 2 +[23] Manish Kesarwani, Bhaskar Mukholy, Vijay Arya, and Sameep Mehta. Model extraction warning in mlaas paradigm. In Annual Computer Security Applications Conference, pages 371-380, 2018. 1 +[24] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In International Conference on Computer Vision, 2023. 1 +[25] Kalpesh Krishna, Gaurav Singh Tomar, Ankur P Parikh, Nicolas Papernot, and Mohit Iyyer. Thieves on sesame street! model extraction of bert-based apis. In International Conference on Learning Representations, 2020. 2 +[26] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009. 6 +[27] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. ImageNet classification with deep convolutional neural networks. In International Conference on Advanced Neural Information Processing Systems, pages 1106-1114, 2012. 1 +[28] Souvik Kundu, Qirui Sun, Yao Fu, Massoud Pedram, and Peter Beerel. Analyzing the confidentiality of undistinguishable teachers in knowledge distillation. pages 9181-9192, 2021. 2 + +[29] Jeonghyun Lee, Sungmin Han, and Sangkyun Lee. Model stealing defense against exploiting information leak through the interpretation of deep neural nets. In International Joint Conference on Artificial Intelligence, pages 710-716, 2022. 1, 3 +[30] Taesung Lee, Benjamin Edwards, Ian Molloy, and Dong Su. Defending against neural network model stealing attacks using deceptive perturbations. In IEEE Security and Privacy Workshops, pages 43-49, 2019. 2, 3 +[31] Simin Li, Shuning Zhang, Gujun Chen, Dong Wang, Pu Feng, Jiakai Wang, Aishan Liu, Xin Yi, and Xianglong Liu. Towards benchmarking and assessing visual naturalness of physical world adversarial attacks. In IEEE/CVF Computer Vision and Pattern Recognition Conference, pages 12324–12333, 2023. 1 +[32] Yiming Li, Linghui Zhu, Xiaojun Jia, Yong Jiang, Shu-Tao Xia, and Xiaochun Cao. Defending against model stealing via verifying embedded external features. In Association for the Advance of Artificial Intelligence, pages 1464-1472, 2022. 3 +[33] Shunchang Liu, Jiakai Wang, Aishan Liu, Yingwei Li, Yijie Gao, Xianglong Liu, and Dacheng Tao. Harnessing perceptual adversarial patches for crowd counting. In ACM SIGSAC Conference on Computer and Communications Security, pages 2055-2069, 2022. 1 +[34] Haoyu Ma, Tianlong Chen, Ting-Kuei Hu, Chenyu You, Xiaohui Xie, and Zhangyang Wang. Undistillable: Making a nasty teacher that cannot teach students. In ICLR, 2021. 6 +[35] Ke Ma, Qianqian Xu, Jinshan Zeng, Xiaochun Cao, and Qingming Huang. Poisoning attack against estimating from pairwise comparisons. IEEE Trans. Pattern Anal. Mach. Intell., pages 6393-6408, 2021. 1 +[36] Ke Ma, Qianqian Xu, Jinshan Zeng, Guorong Li, Xiaochun Cao, and Qingming Huang. A tale of hodgerank and spectral method: Target attack against rank aggregation is the fixed point of adversarial game. IEEE Trans. Pattern Anal. Mach. Intell., pages 4090-4108, 2022. 1 +[37] Pratyush Maini, Mohammad Yaghini, and Nicolas Papernot. Dataset inference: Ownership resolution in machine learning. In International Conference on Learning Representations, 2021. 3 +[38] Mantas Mazeika, Bo Li, and David Forsyth. How to steer your adversary: Targeted and efficient model stealing defenses with gradient redirection. In International Conference on Machine Learning, pages 15241-15254, 2022. 2, 3, 6, 7 +[39] Tribhuvanesh Orekondy, Bernt Schiele, and Mario Fritz. Knockoff nets: Stealing functionality of black-box models. In IEEE/CVF Computer Vision and Pattern Recognition Conference, pages 4954-4963, 2019. 1, 2, 3 +[40] Tribhuvanesh Orekondy, Bernt Schiele, and Mario Fritz. Prediction poisoning: Towards defenses against dnn model stealing attacks. In International Conference on Learning Representations, 2020. 2, 3, 6 +[41] Soham Pal, Yash Gupta, Aditya Shukla, Aditya Kanade, Shirish Shevade, and Vinod Ganapathy. Activethief: Model extraction using active learning and unannotated public data. In Association for the Advance of Artificial Intelligence, pages 865–872, 2020. 2 + +[42] Nicolas Papernot, Patrick McDaniel, Ian Goodfellow, Somesh Jha, Z Berkay Celik, and Ananthram Swami. Practical black-box attacks against machine learning. In ACM on Asia conference on computer and communications security, pages 506-519, 2017. 1, 2, 3 +[43] Giorgio Patrini, Alessandro Rozza, Aditya Krishna Menon, Richard Nock, and Lizhen Qu. Making deep neural networks robust to label noise: A loss correction approach. In IEEE/CVF Computer Vision and Pattern Recognition Conference, pages 1944-1952, 2017. 2, 3 +[44] Sunandini Sanyal, Sravanti Addepalli, and R Venkatesh Babu. Towards data-free model stealing in a hard label setting. In CVPR, 2022. 2 +[45] Zeyang Sha et al. Can't steal? cont-steal! contrastive stealing attacks against image encoders. In CVPR, 2023. 2 +[46] Yu Shi, Dong-Dong Wu, Xin Geng, and Min-Ling Zhang. Robust representation learning for unreliable partial label learning. arXiv preprint arXiv:2308.16718, 2023. 3 +[47] Reza Shokri, Marco Stronati, Congzheng Song, and Vitaly Shmatikov. Membership inference attacks against machine learning models. In IEEE symposium on security and privacy, pages 3-18, 2017. 1 +[48] Sainbayar Sukhbaatar, Joan Bruna, Manohar Paluri, Lubomir Bourdev, and Rob Fergus. Training convolutional networks with noisy labels. arXiv preprint arXiv:1406.2080, 2014. 3 +[49] Christian Szegedy, Wojciech Zaremba, Ilya Sutskever, Joan Bruna, Dumitru Erhan, Ian Goodfellow, and Rob Fergus. Intriguing properties of neural networks. In International Conference on Learning Representations, 2013. 1 +[50] Sebastian Szyller, Buse Gul Atli, Samuel Marchal, and N Asokan. Dawn: Dynamic adversarial watermarking of neural networks. In ACM International Conference on Multimedia, pages 4417-4425, 2021. 1, 3 +[51] Florian Tramér, Fan Zhang, Ari Juels, Michael K Reiter, and Thomas Ristenpart. Stealing machine learning models via prediction {APIs}. In USENIX security symposium, pages 601-618, 2016. 1, 2, 3 +[52] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In International Conference on Advanced Neural Information Processing Systems, 2017. 1 +[53] M Mitchell Waldrop et al. No drivers required. Nature, page 20, 2015. 1 +[54] Binghui Wang and Neil Zhenqiang Gong. Stealing hyperparameters in machine learning. In IEEE symposium on security and privacy, pages 36-52, 2018. 1 +[55] Jiakai Wang, Aishan Liu, Zixin Yin, Shunchang Liu, Shiyu Tang, and Xianglong Liu. Dual attention suppression attack: Generate adversarial camouflage in physical world. In IEEE/CVF Computer Vision and Pattern Recognition Conference, pages 8565-8574, 2021. 1 +[56] Zhen Wang, Guosheng Hu, and Qinghua Hu. Training noise-robust deep neural networks via meta-learning. In IEEE/CVF Computer Vision and Pattern Recognition Conference, pages 4524–4533, 2020. 3 + +[57] Peter Welinder, Steve Branson, Takeshi Mita, Catherine Wah, Florian Schroff, Serge Belongie, and Pietro Perona. Caltech-ucsd birds 200. 2010. 6 +[58] Hongwei Wen, Jingyi Cui, Hanyuan Hang, Jiabin Liu, Yisen Wang, and Zhouchen Lin. Leveraged weighted loss for partial label learning. In International Conference on Machine Learning, pages 11091-11100, 2021. 2 +[59] Dong-Dong Wu, Deng-Bao Wang, and Min-Ling Zhang. Revisiting consistency regularization for deep partial label learning. In International Conference on Machine Learning, pages 24212-24225, 2022. 5 +[60] Xiaobo Xia, Tongliang Liu, Nannan Wang, Bo Han, Chen Gong, Gang Niu, and Masashi Sugiyama. Are anchor points really indispensable in label-noise learning? In International Conference on Advanced Neural Information Processing Systems, pages 6835-6846, 2019. 3 +[61] Hao Yang, You-Zhi Jin, Zi-Yin Li, Deng-Bao Wang, Xin Geng, and Min-Ling Zhang. Learning from noisy labels via dynamic loss thresholding. IEEE Transactions on Knowledge and Data Engineering, 2023. 3 +[62] Honggang Yu, Kaichen Yang, Teng Zhang, Yun-Yun Tsai, Tsung-Yi Ho, and Yier Jin. Cloudleak: Large-scale deep learning models stealing through adversarial examples. In Network and Distributed System Security Symposium, 2020. 1 +[63] Chiyuan Zhang, Samy Bengio, Moritz Hardt, Benjamin Recht, and Oriol Vinyals. Understanding deep learning (still) requires rethinking generalization. Communications of the ACM, pages 107-115, 2021. 2 +[64] Jiliang Zhang, Shuang Peng, Yansong Gao, Zhi Zhang, and Qinghui Hong. Apmsa: adversarial perturbation against model stealing attacks. IEEE Transactions on Information Forensics and Security, pages 1667-1679, 2023. 1 +[65] Guoqing Zheng, Ahmed Hassan Awadallah, and Susan Dumas. Meta label correction for noisy label learning. In Association for the Advance of Artificial Intelligence, pages 11053-11061, 2021. 3 \ No newline at end of file diff --git a/2024/Efficient Model Stealing Defense with Noise Transition Matrix/images.zip b/2024/Efficient Model Stealing Defense with Noise Transition Matrix/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..aae35dbd0e816e7f188b305e0792fa026dab282f --- /dev/null +++ b/2024/Efficient Model Stealing Defense with Noise Transition Matrix/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:10c742d5ffe81d618a187ce181580ce1c7feaa95ec74a34fce0bb594504fec35 +size 492093 diff --git a/2024/Efficient Model Stealing Defense with Noise Transition Matrix/layout.json b/2024/Efficient Model Stealing Defense with Noise Transition Matrix/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..63ba34c56036b48a7918f2f88aecfae26cac3c98 --- /dev/null +++ b/2024/Efficient Model Stealing Defense with Noise Transition Matrix/layout.json @@ -0,0 +1,10212 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 105, + 103, + 489, + 120 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 105, + 103, + 489, + 120 + ], + "spans": [ + { + "bbox": [ + 105, + 103, + 489, + 120 + ], + "type": "text", + "content": "Efficient Model Stealing Defense with Noise Transition Matrix" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 144, + 142, + 447, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 144, + 142, + 447, + 157 + ], + "spans": [ + { + "bbox": [ + 144, + 142, + 447, + 157 + ], + "type": "text", + "content": "Dong-Dong " + }, + { + "bbox": [ + 144, + 142, + 447, + 157 + ], + "type": "inline_equation", + "content": "\\mathsf{W u}^{1}" + }, + { + "bbox": [ + 144, + 142, + 447, + 157 + ], + "type": "text", + "content": ", Chilin " + }, + { + "bbox": [ + 144, + 142, + 447, + 157 + ], + "type": "inline_equation", + "content": "\\mathsf{F u}^{2}" + }, + { + "bbox": [ + 144, + 142, + 447, + 157 + ], + "type": "text", + "content": ", Weichang " + }, + { + "bbox": [ + 144, + 142, + 447, + 157 + ], + "type": "inline_equation", + "content": "\\mathsf{W u}^{2}" + }, + { + "bbox": [ + 144, + 142, + 447, + 157 + ], + "type": "text", + "content": ", Wenwen " + }, + { + "bbox": [ + 144, + 142, + 447, + 157 + ], + "type": "inline_equation", + "content": "\\mathsf{X i a}^{2}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 180, + 157, + 412, + 172 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 180, + 157, + 412, + 172 + ], + "spans": [ + { + "bbox": [ + 180, + 157, + 412, + 172 + ], + "type": "text", + "content": "Xiaolu Zhang" + }, + { + "bbox": [ + 180, + 157, + 412, + 172 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 180, + 157, + 412, + 172 + ], + "type": "text", + "content": ", Jun Zhou" + }, + { + "bbox": [ + 180, + 157, + 412, + 172 + ], + "type": "inline_equation", + "content": "^{2*}" + }, + { + "bbox": [ + 180, + 157, + 412, + 172 + ], + "type": "text", + "content": ", Min-Ling Zhang" + }, + { + "bbox": [ + 180, + 157, + 412, + 172 + ], + "type": "inline_equation", + "content": "^{1*}" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 92, + 171, + 501, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 92, + 171, + 501, + 186 + ], + "spans": [ + { + "bbox": [ + 92, + 171, + 501, + 186 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 92, + 171, + 501, + 186 + ], + "type": "text", + "content": "School of Computer Science and Engineering, Southeast University, Nanjing, China" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 268, + 184, + 326, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 268, + 184, + 326, + 200 + ], + "spans": [ + { + "bbox": [ + 268, + 184, + 326, + 200 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 268, + 184, + 326, + 200 + ], + "type": "text", + "content": "Ant Group" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 76, + 201, + 517, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 201, + 517, + 213 + ], + "spans": [ + { + "bbox": [ + 76, + 201, + 517, + 213 + ], + "type": "text", + "content": "{dongdongwu,zhangml}@seu.edu.cn,{chilin.fcl,jiuyue.wwc,wenzhen.xww}@antgroup.com" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 194, + 216, + 389, + 227 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 194, + 216, + 389, + 227 + ], + "spans": [ + { + "bbox": [ + 194, + 216, + 389, + 227 + ], + "type": "text", + "content": "{yueyin.zxl, jun.zhoujun}@antfin.com" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 143, + 255, + 192, + 268 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 255, + 192, + 268 + ], + "spans": [ + { + "bbox": [ + 143, + 255, + 192, + 268 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 280, + 290, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 280, + 290, + 544 + ], + "spans": [ + { + "bbox": [ + 46, + 280, + 290, + 544 + ], + "type": "text", + "content": "With the escalating complexity and investment cost of training deep neural networks, safeguarding them from unauthorized usage and intellectual property theft has become imperative. Especially the rampant misuse of prediction APIs to replicate models without access to the original data or architecture poses grave security threats. Diverse defense strategies have emerged to address these vulnerabilities, yet these defenses either incur heavy inference overheads or assume idealized attack scenarios. To address these challenges, we revisit the utilization of noise transition matrix as an efficient perturbation technique, which injects noise into predicted posteriors in a linear manner and integrates seamlessly into existing systems with minimal overhead, for model stealing defense. Provably, with such perturbed posteriors, the attacker's cloning process degrades into learning from noisy data. Toward optimizing the noise transition matrix, we proposed a novel bi-level optimization training framework, which performs fidelity on the victim model while the surrogate model adversarially. Comprehensive experimental results demonstrate that our method effectively thwarts model stealing attacks and achieves minimal utility trade-offs, outperforming existing state-of-the-art defenses." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 567, + 128, + 580 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 567, + 128, + 580 + ], + "spans": [ + { + "bbox": [ + 47, + 567, + 128, + 580 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 588, + 288, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 588, + 288, + 685 + ], + "spans": [ + { + "bbox": [ + 46, + 588, + 288, + 685 + ], + "type": "text", + "content": "Machine learning (ML) models, especially deep neural networks (DNNs), have attained human-level capabilities across various domains, such as pattern recognition tasks [24, 27] and generation tasks [2, 52], becoming indispensable in facilitating convenience for society [53]. With the burgeoning of ML application scenarios and the proliferation of data, training large and sophisticated models and making them publicly accessible has become a new profit trend for compa" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 256, + 547, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 256, + 547, + 376 + ], + "spans": [ + { + "bbox": [ + 306, + 256, + 547, + 376 + ], + "type": "text", + "content": "nies. However, considering the steep costs of tuning models, directly divulging model parameters potentially engenders no returns on investment and even adverse societal impacts from malicious use. To alleviate those concerns, prediction APIs have emerged as an alternative by enabling black-box inferences through querying. This allows models to provide services without exposing the underlying parameters and details, and has been harnessed in Machine Learning as a Service (MLaaS) platforms (e.g., Amazon AWS, Microsoft Azure, Google Cloud)." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 377, + 548, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 377, + 548, + 604 + ], + "spans": [ + { + "bbox": [ + 304, + 377, + 548, + 604 + ], + "type": "text", + "content": "Unfortunately, recent research has demonstrated that prediction APIs are still vulnerable to intellectual property theft through model stealing attacks [39, 51, 54, 62]. Despite not exposing model details, attackers can query the APIs using carefully-crafted samples or surrogate samples to obtain responded predictions. Through this, they can then train a clone model on the gathered query-response pairs to replicate the victim model's functionality with high accuracy and fidelity, without needing to access the original training data or model architecture [22, 39]. Such a clone model poses the following security threats: (a) acquiring a functionally similar copy of the victim model, illegally circumventing the expensive process of manual data curation and undermining the intellectual property of MLaaS; (b) extracting confidential data through membership inference attacks [47], model inversion attack [6] and property inference attack [7]; (c) conducting adversarial attacks [31, 33, 35, 36, 49, 55] by creating malicious examples based on the white-box substitute model [42]." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 605, + 548, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 548, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 548, + 715 + ], + "type": "text", + "content": "To address such vulnerabilities, considerable researches [3, 5, 9, 17, 29, 50, 64] have focused on thwarting model stealing attacks. These defenses can be categorized into three strategies: detection, perturbation, and post-hoc verification. Detection-based defenses [19, 20, 23, 23] are predicated on the idea that malicious queries exhibit aberrant statistical distributions compared to normal queries. If a query sequence appears anomalous, the defender may discard the requests or corrupt the outputs. However, if an attacker carefully con" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 693, + 133, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 693, + 133, + 703 + ], + "spans": [ + { + "bbox": [ + 53, + 693, + 133, + 703 + ], + "type": "text", + "content": "* Corresponding author." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 53, + 703, + 286, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 703, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 53, + 703, + 286, + 713 + ], + "type": "text", + "content": "This work is done when Dong-Dong Wu works as an intern at Ant Group." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "24305" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 49, + 71, + 286, + 155 + ], + "blocks": [ + { + "bbox": [ + 49, + 71, + 286, + 155 + ], + "lines": [ + { + "bbox": [ + 49, + 71, + 286, + 155 + ], + "spans": [ + { + "bbox": [ + 49, + 71, + 286, + 155 + ], + "type": "image", + "image_path": "27b8898ae862573b82ebeb905ba6a3ba9fe7846f0081b6bf335764e4c4751b5c.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 164, + 288, + 232 + ], + "lines": [ + { + "bbox": [ + 46, + 164, + 288, + 232 + ], + "spans": [ + { + "bbox": [ + 46, + 164, + 288, + 232 + ], + "type": "text", + "content": "Figure 1. Illustrations of the noise transition matrix serving for noisy label learning and model stealing defense. In noisy label learning, it transforms the predicted posteriors to align with the noisy label. In model stealing defense, it perturbs the predicted posterior to impede attacker training while preserving the original predicted label for benign users." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 256, + 287, + 484 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 256, + 287, + 484 + ], + "spans": [ + { + "bbox": [ + 46, + 256, + 287, + 484 + ], + "type": "text", + "content": "structures queries matching the benign distribution, detection techniques will fail to discern the attack, yielding excessive false positives. Perturbation-based defenses [30, 40, 51] can preserve or constraint model accuracy to foil model stealing attacks. The accuracy-preserving defense truncates predicted posteriors to a minuscule fraction of their original size, which may reduce model lucidity and utility for critical applications like autonomous driving. The accuracy-constraint defense [38, 40] introduces perturbation into predicted posteriors, which incurs additional computational burdens or substantially degrading utility. Furthermore, the latency resulting from such perturbation methods leads to extended user response times, making the approach impractical. Post-hoc verification techniques such as watermarking [15] remain useful for post-attack prevention and validation. However, their retroactive nature means they cannot take proactive measures before an attack occurs. Considering inference latency and utility trade-offs, in this paper, we concentrate on a novel and efficient model stealing defense strategy." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 486, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 486, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 486, + 288, + 713 + ], + "type": "text", + "content": "To overcome the limitations of existing defenses, we explore an efficient method to inject noise into predicted posteriors via linear mapping, while preserving utility for benign users. Motivated by memorization effects [11, 63] that neural networks can easily overfit noisy data, thus hampering model generalization, we revisit the utilization of noise transition matrix (NTM) in noisy labels learning [43, 58], where the matrix models the noise process. Unlike related work, we aim to leverage the NTM to introduce noise into predicted posteriors for model defense, as shown in Figure 1. Our goal is to find an optimized NTM that injects noise while retaining original prediction labels with minimal utility trade-offs. Intuitively, regardless of whether the victim model produces extremely confident one-hot vectors or uncertain flat vectors, we can readily derive a uniform NTM where the diagonal elements remain maximal for each row to meet our defense needs. To estimate the NTM with maximum utility, we propose a bi-level optimization method that alternately optimizes it and updates the surrogate network" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 304, + 72, + 547, + 131 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 131 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 131 + ], + "type": "text", + "content": "weights. Specifically, we first make a one-step-forward virtual optimization of weights, then optimize the matrix guided by the loss with the one-step-forward weights fixed. Finally, we optimize the unrolled weights with the updated matrix. The contributions of this work are as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 306, + 133, + 547, + 300 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 306, + 133, + 547, + 193 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 133, + 547, + 193 + ], + "spans": [ + { + "bbox": [ + 306, + 133, + 547, + 193 + ], + "type": "text", + "content": "- We propose a novel perturbation-based defense that efficiently introduces noise through linear mapping, adding minimal overhead. Furthermore, it deteriorates attacker training to a level equivalent to training on noisy-labeled data with instance-wise weight." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 193, + 547, + 241 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 193, + 547, + 241 + ], + "spans": [ + { + "bbox": [ + 306, + 193, + 547, + 241 + ], + "type": "text", + "content": "- Our method pioneers using the transition matrix to thwart model stealing attacks. The explicit matrix helps understand noise relationships among classes and makes the defense strategy more interpretable" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 241, + 545, + 276 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 241, + 545, + 276 + ], + "spans": [ + { + "bbox": [ + 306, + 241, + 545, + 276 + ], + "type": "text", + "content": "- A bi-level optimization approach is proposed to optimize the transition matrix, which is model-agnostic and can adapt to different backbone networks." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 277, + 547, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 277, + 547, + 300 + ], + "spans": [ + { + "bbox": [ + 306, + 277, + 547, + 300 + ], + "type": "text", + "content": "- Extensive experiments exhibit our method consistently mitigates stealing and outperforms state-of-the-art defenses." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 315, + 392, + 327 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 315, + 392, + 327 + ], + "spans": [ + { + "bbox": [ + 306, + 315, + 392, + 327 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 336, + 547, + 576 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 336, + 547, + 576 + ], + "spans": [ + { + "bbox": [ + 304, + 336, + 547, + 576 + ], + "type": "text", + "content": "Model Stealing Attacks aim to replicate the functionality of a victim model by interrogating its prediction APIs. The attacker sends queries and collects the resultant predictions, typically probability vectors or labels, as query-response pairs. These collected pairs are used to train a clone model that approximates the victim model. Tramèr et al. [51] first proposed the idea of model stealing and demonstrated the feasibility of extracting simple models utilizing queries. Juuti et al. [19] presented extraction attacks leveraging jacobian-based data augmentation [42], synthesizing data by perturbing a small set of in-distribution samples. However, these attacks presume the ability to optimally probe the APIs, which could be detected by monitoring for anomalous query patterns [19, 41]. Therefore, we concentrate on scenarios without carefully engineered input queries or sequences. One classic scenario to achieve this is by utilizing knowledge distillation techniques [12], where some studies exhibit model stealing is viable using weakly correlated [39] or random [25] queries, even when the teacher model declares itself undistillable [13, 28]." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 582, + 548, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 582, + 548, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 582, + 548, + 715 + ], + "type": "text", + "content": "Model Stealing Defenses can be taxonomized into three primary strategies: detection, perturbation, and post-hoc verification. Detection strategies [20, 21] aim to identify malicious queries or nefarious users by scrutinizing the queries received by the victim model. For example, Kariyappa and Qureshi [20] regarded malicious queries as out-of-distribution samples and devised an adaptive misinformation (AM) defense to mislead attackers with modified outputs. However, such detection strategies rely on strong assumptions about adversarial query patterns that may not reflect reality. Perturbation strategies mitigate model stealing" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "24306" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 289, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 289, + 384 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 289, + 384 + ], + "type": "text", + "content": "attacks by returning perturbed posterior, which can be further subcategorized into accuracy-preserving [30] and accuracy-constrained defenses [29, 38, 40]. Accuracy-preserving defenses guarantee that the maximal class of the perturbed predictions matches the original prediction. The most straightforward such defense is truncating the posterior or only outputting the predicted label [51]. However, this diminishes utility for benign users, precluding specific applications and harming external transparency. In contrast, accuracy-constrained defenses permit a higher level of perturbation, trading off benign accuracy for enhanced security against model stealing attacks. In general, existing defense strategies that poison the attacker's update gradient [38, 40] are computationally expensive, even if they yield a comparative defense. Post-hoc verification, including watermarking [15, 32, 50], dataset inference [37], and proof of learning [16], embeds concealed patterns in the model during training [1] or inference [50] to distinguish a stolen model. Unfortunately, these approaches only work after a model theft, requiring the model owner to obtain partial or full access to the stolen model. This means if an attacker uses the stolen model as a proxy to attack the victim model, it is impossible for the model owner to protect themselves with these defenses. Overall, while existing ante-hoc defensive approaches mitigate knowledge leakage, they rely on strong assumptions or incur high computational costs during inference." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 389, + 288, + 606 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 389, + 288, + 606 + ], + "spans": [ + { + "bbox": [ + 47, + 389, + 288, + 606 + ], + "type": "text", + "content": "Noise Transition Matrix (NTM) is often utilized to solve classification with noisy labels or ambiguous labeled data [10, 46, 61], where the matrix reflects probabilities of true label flipping to the other label. A typical application is loss correction [43], which modifies the loss of each example by multiplying the estimated matrix by the model output. For NTM estimation, there are two main approaches. One is a two-step solution that pre-estimates the matrix with anchor point assumptions, then uses it to train the classifier [56, 65]. For instance, backward correction [43] initially approximates the NTM using the softmax output of the model trained without correction. It then retransmits the model while correcting the original loss based on the estimated matrix. The other jointly estimates the matrix and classifier parameters in a unified framework without employing anchor points [18, 48, 60]. Differently, our work aims to estimate and apply the transition amtrix to intentionally inject noise into predicted posteriors, thus defending against model stealing." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 620, + 133, + 632 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 620, + 133, + 632 + ], + "spans": [ + { + "bbox": [ + 47, + 620, + 133, + 632 + ], + "type": "text", + "content": "3. Threat Model" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 641, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 641, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 641, + 288, + 714 + ], + "type": "text", + "content": "We consider a model stealing scenario involving a defender and an attacker interacting on image classification tasks. The defender grants access to its black-box model through prediction API. Leveraging the API, the attacker collects query-response pairs, aiming to mimic the defender model (victim model) functionalities by training a clone model on" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 304, + 72, + 547, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 144 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 144 + ], + "type": "text", + "content": "this dataset. To evade detection defenses [19], we assume the attacker adopts non-adaptive querying strategies without anomalous patterns [39, 42]. Therefore, we consider the knockoff nets attack without adaptive querying [39], which resembles knowledge distillation [12] where the victim model distills knowledge into the attacker model." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 305, + 149, + 425, + 163 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 149, + 425, + 163 + ], + "spans": [ + { + "bbox": [ + 305, + 149, + 425, + 163 + ], + "type": "text", + "content": "3.1. Attacker's Objective" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 167, + 547, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 167, + 547, + 312 + ], + "spans": [ + { + "bbox": [ + 304, + 167, + 547, + 312 + ], + "type": "text", + "content": "Let " + }, + { + "bbox": [ + 304, + 167, + 547, + 312 + ], + "type": "inline_equation", + "content": "\\theta_{f}" + }, + { + "bbox": [ + 304, + 167, + 547, + 312 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 167, + 547, + 312 + ], + "type": "inline_equation", + "content": "\\theta_{h}" + }, + { + "bbox": [ + 304, + 167, + 547, + 312 + ], + "type": "text", + "content": " denote the parameters of the defender's model " + }, + { + "bbox": [ + 304, + 167, + 547, + 312 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 304, + 167, + 547, + 312 + ], + "type": "text", + "content": " and attacker's clone model " + }, + { + "bbox": [ + 304, + 167, + 547, + 312 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 304, + 167, + 547, + 312 + ], + "type": "text", + "content": ", respectively. The attacker prepares a query dataset " + }, + { + "bbox": [ + 304, + 167, + 547, + 312 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{\\mathrm{query}}" + }, + { + "bbox": [ + 304, + 167, + 547, + 312 + ], + "type": "text", + "content": ", and sends queries " + }, + { + "bbox": [ + 304, + 167, + 547, + 312 + ], + "type": "inline_equation", + "content": "\\pmb{x} \\in \\mathcal{D}_{\\mathrm{query}}" + }, + { + "bbox": [ + 304, + 167, + 547, + 312 + ], + "type": "text", + "content": " in random order to obtain the defender's posterior " + }, + { + "bbox": [ + 304, + 167, + 547, + 312 + ], + "type": "inline_equation", + "content": "f(\\pmb{x}; \\pmb{\\theta}_{f})" + }, + { + "bbox": [ + 304, + 167, + 547, + 312 + ], + "type": "text", + "content": " from the prediction API. The classification loss of the attacker's clone model on the query " + }, + { + "bbox": [ + 304, + 167, + 547, + 312 + ], + "type": "inline_equation", + "content": "\\pmb{x}" + }, + { + "bbox": [ + 304, + 167, + 547, + 312 + ], + "type": "text", + "content": " is " + }, + { + "bbox": [ + 304, + 167, + 547, + 312 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{kna}} = -\\sum_{i} f(\\pmb{x}; \\pmb{\\theta}_{f})_{i} \\log h(\\pmb{x}; \\pmb{\\theta}_{h})_{i}" + }, + { + "bbox": [ + 304, + 167, + 547, + 312 + ], + "type": "text", + "content": ". The attacker aims to minimize the classification loss on the defender's test set. However, since the test set is unavailable, the attacker's learning objective becomes minimizing the loss based on the collected query-response dataset " + }, + { + "bbox": [ + 304, + 167, + 547, + 312 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{\\mathrm{pair}} = \\{(x, f(x; \\pmb{\\theta}_{f})) | x \\in \\mathcal{D}_{\\mathrm{query}}\\}" + }, + { + "bbox": [ + 304, + 167, + 547, + 312 + ], + "type": "text", + "content": ". Therefore, the attacker's objective can be formulated as:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 345, + 316, + 545, + 335 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 345, + 316, + 545, + 335 + ], + "spans": [ + { + "bbox": [ + 345, + 316, + 545, + 335 + ], + "type": "interline_equation", + "content": "\\min _ {\\boldsymbol {\\theta} _ {h}} \\mathbb {E} _ {\\boldsymbol {x} \\sim \\mathcal {D} _ {\\text {p a i r}}} \\left[ \\mathcal {L} _ {\\mathrm {k n a}} \\left(h (\\boldsymbol {x}; \\boldsymbol {\\theta} _ {h}), f (\\boldsymbol {x}; \\boldsymbol {\\theta} _ {f})\\right) \\right] \\tag {1}", + "image_path": "8ee08dd8b62504639f72df5126fcbc330b0fda04aded8f2dfd1cd4c63c892b4f.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 340, + 547, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 340, + 547, + 460 + ], + "spans": [ + { + "bbox": [ + 304, + 340, + 547, + 460 + ], + "type": "text", + "content": "Query Distribution. In practice, due to the knowledge limitation of attackers about the defender's training set, the collected query dataset " + }, + { + "bbox": [ + 304, + 340, + 547, + 460 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{\\mathrm{query}}" + }, + { + "bbox": [ + 304, + 340, + 547, + 460 + ], + "type": "text", + "content": " may be similar [30, 51] or unrelated [20, 40] to the defender's training distribution. Hence, we consider both possibilities in our threat model. We refer to attackers as distribution-aware if their queries share semantic content with the defender's training distribution. Attackers are referred to as knowledge-limited if their queries are semantically disjoint from the defender's training distribution." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 465, + 427, + 478 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 465, + 427, + 478 + ], + "spans": [ + { + "bbox": [ + 305, + 465, + 427, + 478 + ], + "type": "text", + "content": "3.2. Defender's Objective" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 483, + 547, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 483, + 547, + 604 + ], + "spans": [ + { + "bbox": [ + 304, + 483, + 547, + 604 + ], + "type": "text", + "content": "In defenses against model stealing, the defender aims at preventing its model functionality from being stolen. In real-world scenarios, attackers may disguise themselves as benign users, executing attacks anytime. Therefore, the defender must persistently mitigate model stealing attempts while preserving the API's utility for legitimate users. Specifically, the defender needs to keep its own small classification loss, while maximizing the classification loss of the attacker's cloned model " + }, + { + "bbox": [ + 304, + 483, + 547, + 604 + ], + "type": "inline_equation", + "content": "\\theta_h^*(\\theta_f)" + }, + { + "bbox": [ + 304, + 483, + 547, + 604 + ], + "type": "text", + "content": " obtained from Eq. (1) on the test set " + }, + { + "bbox": [ + 304, + 483, + 547, + 604 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{\\mathrm{test}}" + }, + { + "bbox": [ + 304, + 483, + 547, + 604 + ], + "type": "text", + "content": ". The defender's objective can be formulated as:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 352, + 608, + 545, + 631 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 352, + 608, + 545, + 631 + ], + "spans": [ + { + "bbox": [ + 352, + 608, + 545, + 631 + ], + "type": "interline_equation", + "content": "\\min _ {\\boldsymbol {\\theta} _ {f}} \\mathbb {E} _ {(\\boldsymbol {x}, y) \\sim \\mathcal {D} _ {\\text {t e s t}}} [ \\mathcal {L} _ {\\mathrm {C E}} (f (\\boldsymbol {x}; \\boldsymbol {\\theta} _ {f}), y) ] \\tag {2}", + "image_path": "baa48a0dd729516aabf36e115662a0abf3bb41caca63bc3c86711dcd314cd79f.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 336, + 631, + 517, + 644 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 336, + 631, + 517, + 644 + ], + "spans": [ + { + "bbox": [ + 336, + 631, + 517, + 644 + ], + "type": "interline_equation", + "content": "\\operatorname {s. t.} \\max _ {\\mathbf {x}, y) \\sim \\mathcal {D} _ {\\text {t e s t}}} [ \\mathcal {L} _ {\\mathrm {C E}} (h (\\boldsymbol {x}; \\boldsymbol {\\theta} _ {h} ^ {*} (\\boldsymbol {\\theta} _ {f})), y) ],", + "image_path": "9b0d1550e1b4d076510a8d61fa97a59b7476fd4b74f6cb4ea7c706d6fcd3c1ad.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 648, + 507, + 660 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 648, + 507, + 660 + ], + "spans": [ + { + "bbox": [ + 306, + 648, + 507, + 660 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 306, + 648, + 507, + 660 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{CE}}" + }, + { + "bbox": [ + 306, + 648, + 507, + 660 + ], + "type": "text", + "content": " means the cross-entropy loss function." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 669, + 373, + 682 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 669, + 373, + 682 + ], + "spans": [ + { + "bbox": [ + 306, + 669, + 373, + 682 + ], + "type": "text", + "content": "4. Approach" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 689, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 689, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 305, + 689, + 547, + 714 + ], + "type": "text", + "content": "To develop an efficacious and inference-efficient defense against model stealing, we propose an Efficient Model de" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "24307" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 384 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 384 + ], + "type": "text", + "content": "fense with noise transition MMatrix (EMMA) to mitigate model stealing attacks. As depicted in the \"Defense phase\" of Figure 2, the attacker iteratively queries the victim model using unlabeled data to obtain posteriors via the prediction API. The victim model receives these queries and outputs corresponding predicted probability vectors. These posteriors then multiply the optimized matrix " + }, + { + "bbox": [ + 46, + 72, + 289, + 384 + ], + "type": "inline_equation", + "content": "T^n" + }, + { + "bbox": [ + 46, + 72, + 289, + 384 + ], + "type": "text", + "content": ", becoming perturbed posteriors that remain useful to benign users, and returned to the attacker. However, if the attacker collects these query-responses as training data to reconstruct a clone model, in that case, this dataset is equivalent to noisy labeled data with instance-wise weight, thus impeding cloning model training. To optimize the transition matrix with maximum utility, as shown in the \"Optimizing phase\", we devise a bi-level optimization between the victim and surrogate model which is devised to represent the attacker model. In each round, given initial " + }, + { + "bbox": [ + 46, + 72, + 289, + 384 + ], + "type": "inline_equation", + "content": "T^t" + }, + { + "bbox": [ + 46, + 72, + 289, + 384 + ], + "type": "text", + "content": " and surrogate weights " + }, + { + "bbox": [ + 46, + 72, + 289, + 384 + ], + "type": "inline_equation", + "content": "\\theta_g^t" + }, + { + "bbox": [ + 46, + 72, + 289, + 384 + ], + "type": "text", + "content": ", the optimized matrix " + }, + { + "bbox": [ + 46, + 72, + 289, + 384 + ], + "type": "inline_equation", + "content": "T^{t+1}" + }, + { + "bbox": [ + 46, + 72, + 289, + 384 + ], + "type": "text", + "content": " stems from two gradient descents (GD) on the attack loss and defense loss, respectively. The surrogate model then updates through a GD on the attack loss. This process loops until reaching the termination condition, yielding the final optimized matrix " + }, + { + "bbox": [ + 46, + 72, + 289, + 384 + ], + "type": "inline_equation", + "content": "T^n" + }, + { + "bbox": [ + 46, + 72, + 289, + 384 + ], + "type": "text", + "content": ". In this section, we first elucidate the defense mechanism of incorporating the transition matrix during the defensive phase. Then, a bi-level optimization approach is expounded to alternately optimize the matrix and update the surrogate model." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 391, + 198, + 403 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 391, + 198, + 403 + ], + "spans": [ + { + "bbox": [ + 47, + 391, + 198, + 403 + ], + "type": "text", + "content": "4.1. The underlying mechanism" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 410, + 288, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 410, + 288, + 517 + ], + "spans": [ + { + "bbox": [ + 46, + 410, + 288, + 517 + ], + "type": "text", + "content": "To understand how the transition matrix defends against attacks, we analyze the classification risks of two scenarios: \"learning on noisy labels\" and \"learning on query-response pairs with perturbed posteriors\". It is found that the two classification risks differ only by an instance-wise weight coefficient multiplied to the latter. Let " + }, + { + "bbox": [ + 46, + 410, + 288, + 517 + ], + "type": "inline_equation", + "content": "\\tilde{y} \\in \\mathbf{Y}" + }, + { + "bbox": [ + 46, + 410, + 288, + 517 + ], + "type": "text", + "content": " denote the noisy label, where " + }, + { + "bbox": [ + 46, + 410, + 288, + 517 + ], + "type": "inline_equation", + "content": "\\tilde{y}" + }, + { + "bbox": [ + 46, + 410, + 288, + 517 + ], + "type": "text", + "content": " is selected based on the transition matrix and the true label. Theorem 1 formalizes the defensive mechanism." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 525, + 288, + 610 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 525, + 288, + 610 + ], + "spans": [ + { + "bbox": [ + 46, + 525, + 288, + 610 + ], + "type": "text", + "content": "Theorem 1. Let " + }, + { + "bbox": [ + 46, + 525, + 288, + 610 + ], + "type": "inline_equation", + "content": "\\mathcal{L}(\\cdot, \\cdot)" + }, + { + "bbox": [ + 46, + 525, + 288, + 610 + ], + "type": "text", + "content": " denote the classification loss function. With sufficiently many queries, the classification risk " + }, + { + "bbox": [ + 46, + 525, + 288, + 610 + ], + "type": "inline_equation", + "content": "\\Re(h)" + }, + { + "bbox": [ + 46, + 525, + 288, + 610 + ], + "type": "text", + "content": " of learning with noisy labels is " + }, + { + "bbox": [ + 46, + 525, + 288, + 610 + ], + "type": "inline_equation", + "content": "\\mathbb{E}_{(\\boldsymbol{x}, \\tilde{\\boldsymbol{y}})}[\\mathcal{L}(h(\\boldsymbol{x}), \\tilde{\\boldsymbol{y}})]" + }, + { + "bbox": [ + 46, + 525, + 288, + 610 + ], + "type": "text", + "content": ". The classification risk " + }, + { + "bbox": [ + 46, + 525, + 288, + 610 + ], + "type": "inline_equation", + "content": "\\Re_T(h)" + }, + { + "bbox": [ + 46, + 525, + 288, + 610 + ], + "type": "text", + "content": " of learning with query-response pairs is " + }, + { + "bbox": [ + 46, + 525, + 288, + 610 + ], + "type": "inline_equation", + "content": "\\mathbb{E}_{(\\boldsymbol{x}, \\tilde{\\boldsymbol{y}})}[q(\\boldsymbol{x}, \\tilde{\\boldsymbol{y}}) \\cdot \\mathcal{L}(h(\\boldsymbol{x}), \\tilde{\\boldsymbol{y}})]" + }, + { + "bbox": [ + 46, + 525, + 288, + 610 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 525, + 288, + 610 + ], + "type": "inline_equation", + "content": "q(\\boldsymbol{x}, \\tilde{\\boldsymbol{y}})" + }, + { + "bbox": [ + 46, + 525, + 288, + 610 + ], + "type": "text", + "content": " denotes the instance-wise weight, proportional to the closeness between the predicted posterior and the true label." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 617, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 617, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 617, + 289, + 715 + ], + "type": "text", + "content": "The proof can be found in Appendix 8. Theorem 1 demonstrates that training a clone model on such query-response pairs with perturbed posteriors is equivalent to training it on noisy-labeled data with instance-wise weights. It can be further deduced that the more closely the victim model's predicted posteriors match the true label, the more this training process resembles learning with noisy labels. In addition, the clone model's generalization can be further degraded by" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 307, + 72, + 545, + 264 + ], + "blocks": [ + { + "bbox": [ + 307, + 72, + 545, + 264 + ], + "lines": [ + { + "bbox": [ + 307, + 72, + 545, + 264 + ], + "spans": [ + { + "bbox": [ + 307, + 72, + 545, + 264 + ], + "type": "image", + "image_path": "e62144276b3f8110e06bbcbef99d4c3cab0b71c7c0bfc522368791929c71f5d9.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 273, + 546, + 339 + ], + "lines": [ + { + "bbox": [ + 304, + 273, + 546, + 339 + ], + "spans": [ + { + "bbox": [ + 304, + 273, + 546, + 339 + ], + "type": "text", + "content": "Figure 2. An illustration of \"Optimizing phase\" and \"Defense phase\" in the framework. \"Optimizing phase\" demonstrates the process of updating the noise transition matrix " + }, + { + "bbox": [ + 304, + 273, + 546, + 339 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 304, + 273, + 546, + 339 + ], + "type": "text", + "content": " and surrogate network weights " + }, + { + "bbox": [ + 304, + 273, + 546, + 339 + ], + "type": "inline_equation", + "content": "\\theta_g" + }, + { + "bbox": [ + 304, + 273, + 546, + 339 + ], + "type": "text", + "content": " alternatively. \"Defensive phase\" shows the interaction process where the attacker tries to steal the functionality from the defender." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 360, + 545, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 360, + 545, + 396 + ], + "spans": [ + { + "bbox": [ + 304, + 360, + 545, + 396 + ], + "type": "text", + "content": "increasing the transition matrix noise rate or exploring more difficult noise type through inter-class correlations, such as using class-conditional noise instead of random noise." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 403, + 477, + 415 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 403, + 477, + 415 + ], + "spans": [ + { + "bbox": [ + 305, + 403, + 477, + 415 + ], + "type": "text", + "content": "4.2. The overall training framework" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 421, + 547, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 421, + 547, + 540 + ], + "spans": [ + { + "bbox": [ + 304, + 421, + 547, + 540 + ], + "type": "text", + "content": "Since the attacker's network is unknown, we cannot directly measure its cloning performance. Instead, we represent the attacker with a surrogate model " + }, + { + "bbox": [ + 304, + 421, + 547, + 540 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 304, + 421, + 547, + 540 + ], + "type": "text", + "content": " parameterized by " + }, + { + "bbox": [ + 304, + 421, + 547, + 540 + ], + "type": "inline_equation", + "content": "\\theta_g" + }, + { + "bbox": [ + 304, + 421, + 547, + 540 + ], + "type": "text", + "content": " and design optimal perturbations for it. By using the surrogate model to simulate the cloning process, we can continuously optimize the transition matrix during this process to minimize the utility trade-off between attacker performance and defender utility. To achieve this, we first introduce the respective objective functions for the attacker and defender. Then, a bi-level optimization approach is proposed." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 540, + 545, + 577 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 540, + 545, + 577 + ], + "spans": [ + { + "bbox": [ + 304, + 540, + 545, + 577 + ], + "type": "text", + "content": "Given the perturbed posterior " + }, + { + "bbox": [ + 304, + 540, + 545, + 577 + ], + "type": "inline_equation", + "content": "f(\\pmb{x}; \\pmb{\\theta}_f)T" + }, + { + "bbox": [ + 304, + 540, + 545, + 577 + ], + "type": "text", + "content": ", the attacker's objective function is defined as the loss on the gathered query-response pairs, formulated as:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 586, + 547, + 600 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 586, + 547, + 600 + ], + "spans": [ + { + "bbox": [ + 313, + 586, + 547, + 600 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {a t t a c k}} \\left(\\boldsymbol {\\theta} _ {g}, T\\right) = \\mathbb {E} _ {\\boldsymbol {x} \\sim \\mathcal {D} _ {\\text {p a i r}}} \\left[ \\mathcal {L} _ {\\text {k n a}} \\left(g (\\boldsymbol {x}; \\boldsymbol {\\theta} _ {g}), f (\\boldsymbol {x}; \\boldsymbol {\\theta} _ {f}) T\\right) \\right]. \\tag {3}", + "image_path": "df0cb67b56470987ec71b73f8c5a8ae198fd31da5955c23f07b6870309cacb71.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 609, + 547, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 609, + 547, + 693 + ], + "spans": [ + { + "bbox": [ + 304, + 609, + 547, + 693 + ], + "type": "text", + "content": "Notably, the optimization of " + }, + { + "bbox": [ + 304, + 609, + 547, + 693 + ], + "type": "inline_equation", + "content": "\\theta_{g}" + }, + { + "bbox": [ + 304, + 609, + 547, + 693 + ], + "type": "text", + "content": " is influenced by " + }, + { + "bbox": [ + 304, + 609, + 547, + 693 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 304, + 609, + 547, + 693 + ], + "type": "text", + "content": ". To defend against model stealing, we would like " + }, + { + "bbox": [ + 304, + 609, + 547, + 693 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 304, + 609, + 547, + 693 + ], + "type": "text", + "content": " to regularize " + }, + { + "bbox": [ + 304, + 609, + 547, + 693 + ], + "type": "inline_equation", + "content": "\\theta_{g}" + }, + { + "bbox": [ + 304, + 609, + 547, + 693 + ], + "type": "text", + "content": " to have poor generalizability. To achieve this, we instantiate a misinformation loss that inhibits the attacker model's outputs on the victim model's original predicted label " + }, + { + "bbox": [ + 304, + 609, + 547, + 693 + ], + "type": "inline_equation", + "content": "k = \\arg \\max_{i} f(\\pmb{x}, \\pmb{\\theta}_{f})_{i}" + }, + { + "bbox": [ + 304, + 609, + 547, + 693 + ], + "type": "text", + "content": ". Formally, we derive the following modified negative log-likelihood loss:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 335, + 700, + 546, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 700, + 546, + 715 + ], + "spans": [ + { + "bbox": [ + 335, + 700, + 546, + 715 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\operatorname {m i s}} \\left(\\boldsymbol {\\theta} _ {g}\\right) = \\mathbb {E} _ {\\boldsymbol {x} \\sim \\mathcal {D} _ {\\text {p a i r}}} \\left[ - \\log \\left(1 - g (\\boldsymbol {x}; \\boldsymbol {\\theta} _ {g}) _ {k}\\right) \\right]. \\tag {4}", + "image_path": "5f8fb32c5e0a63aa8e600d6857d83a80b95c78962145ac4c522c1f1fce3524d4.jpg" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "24308" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 132 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 132 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 132 + ], + "type": "text", + "content": "The underlying principle is quite simple: The log-likelihood of the original prediction label on the output should be minimized. Despite its simple form, optimizing this loss has been empirically shown to be an effective defense against model stealing attacks [20, 59]." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 133, + 288, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 133, + 288, + 190 + ], + "spans": [ + { + "bbox": [ + 46, + 133, + 288, + 190 + ], + "type": "text", + "content": "Moreover, to maintain fidelity for the victim model, a straightforward strategy is aligning the perturbed output with the predicted label " + }, + { + "bbox": [ + 46, + 133, + 288, + 190 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 133, + 288, + 190 + ], + "type": "text", + "content": ". This is implemented by minimizing the divergence between the output and the predicted label, formulated as:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 81, + 201, + 287, + 215 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 81, + 201, + 287, + 215 + ], + "spans": [ + { + "bbox": [ + 81, + 201, + 287, + 215 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {a l i g n}} (T) = \\mathbb {E} _ {\\boldsymbol {x} \\sim \\mathcal {D} _ {\\text {p a i r}}} [ - \\log (f (\\boldsymbol {x}; \\boldsymbol {\\theta} _ {f}) T) _ {k} ]. \\tag {5}", + "image_path": "ca923d260efa89ef6e6f9b062c1b0cce0438d145490041cf557e6f6f97901543.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 223, + 287, + 247 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 223, + 287, + 247 + ], + "spans": [ + { + "bbox": [ + 47, + 223, + 287, + 247 + ], + "type": "text", + "content": "In summary, the overall objective function of the defender can be represented as the following defense loss:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 59, + 256, + 287, + 270 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 256, + 287, + 270 + ], + "spans": [ + { + "bbox": [ + 59, + 256, + 287, + 270 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {d e f e n s e}} \\left(\\boldsymbol {\\theta} _ {g}, T\\right) = \\lambda \\cdot \\mathcal {L} _ {\\text {a l i g n}} (T) + (1 - \\lambda) \\cdot \\mathcal {L} _ {\\text {m i s}} \\left(\\boldsymbol {\\theta} _ {g}\\right), \\tag {6}", + "image_path": "e5150d9bd3c5644c4ead2f60d883fda8784235bea686c5817cf412457a2b73d6.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 279, + 287, + 302 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 279, + 287, + 302 + ], + "spans": [ + { + "bbox": [ + 46, + 279, + 287, + 302 + ], + "type": "text", + "content": "where the multiplicative factor " + }, + { + "bbox": [ + 46, + 279, + 287, + 302 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 46, + 279, + 287, + 302 + ], + "type": "text", + "content": " is used to balance the contribution of these two loss terms." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 307, + 288, + 378 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 307, + 288, + 378 + ], + "spans": [ + { + "bbox": [ + 46, + 307, + 288, + 378 + ], + "type": "text", + "content": "Bi-level optimization. Given the defense loss " + }, + { + "bbox": [ + 46, + 307, + 288, + 378 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{defense}}" + }, + { + "bbox": [ + 46, + 307, + 288, + 378 + ], + "type": "text", + "content": ", the surrogate model weights " + }, + { + "bbox": [ + 46, + 307, + 288, + 378 + ], + "type": "inline_equation", + "content": "\\theta_{g}" + }, + { + "bbox": [ + 46, + 307, + 288, + 378 + ], + "type": "text", + "content": " need to be updated first before optimizing the transition matrix " + }, + { + "bbox": [ + 46, + 307, + 288, + 378 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 46, + 307, + 288, + 378 + ], + "type": "text", + "content": ". For a mini-batch sample, we propose that surrogate model weights could be treated as a latent variable and optimized alternatively with the matrix. This yields the following bi-level optimization problem:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 88, + 387, + 185, + 407 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 88, + 387, + 185, + 407 + ], + "spans": [ + { + "bbox": [ + 88, + 387, + 185, + 407 + ], + "type": "interline_equation", + "content": "\\underset {T} {\\arg \\min } \\mathcal {L} _ {\\text {d e f s e n s e}} \\left(\\boldsymbol {\\theta} _ {g} ^ {*}, T\\right)", + "image_path": "1be227061b77f33f7f4054b0d0769ac38de7ce32205ddb5876d3f7a66715c5ba.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 88, + 404, + 287, + 430 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 88, + 404, + 287, + 430 + ], + "spans": [ + { + "bbox": [ + 88, + 404, + 287, + 430 + ], + "type": "interline_equation", + "content": "\\text {s u b j e c t} \\theta_ {g} ^ {*} = \\underset {\\theta_ {g}} {\\arg \\min } \\mathcal {L} _ {\\text {a t t a c k}} \\left(\\theta_ {g}, T\\right) \\tag {7}", + "image_path": "2b7f57a8a57921ef7cb94457b9ba2fa9ef3762212b9c677233b556646f983ab7.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 438, + 288, + 570 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 438, + 288, + 570 + ], + "spans": [ + { + "bbox": [ + 46, + 438, + 288, + 570 + ], + "type": "text", + "content": "Bi-level optimization problem has been proved strongly NP-hard [14], so getting such an exact solution for problem (7) is impossible in polynomial time. Fortunately, the inner optimization problem in (7) can be transformed into an optimized result " + }, + { + "bbox": [ + 46, + 438, + 288, + 570 + ], + "type": "inline_equation", + "content": "\\theta_g(T)" + }, + { + "bbox": [ + 46, + 438, + 288, + 570 + ], + "type": "text", + "content": " as a function of " + }, + { + "bbox": [ + 46, + 438, + 288, + 570 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 46, + 438, + 288, + 570 + ], + "type": "text", + "content": ". Then " + }, + { + "bbox": [ + 46, + 438, + 288, + 570 + ], + "type": "inline_equation", + "content": "\\theta_g(T)" + }, + { + "bbox": [ + 46, + 438, + 288, + 570 + ], + "type": "text", + "content": " can then substitute " + }, + { + "bbox": [ + 46, + 438, + 288, + 570 + ], + "type": "inline_equation", + "content": "\\theta_g" + }, + { + "bbox": [ + 46, + 438, + 288, + 570 + ], + "type": "text", + "content": " in the outer optimization to acquire an objective solely in terms of " + }, + { + "bbox": [ + 46, + 438, + 288, + 570 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 46, + 438, + 288, + 570 + ], + "type": "text", + "content": ". This allows approximating the solution using gradient descent. Therefore, we employ an alternating optimization strategy to approximately solve this problem. In each iteration " + }, + { + "bbox": [ + 46, + 438, + 288, + 570 + ], + "type": "inline_equation", + "content": "\\theta_g" + }, + { + "bbox": [ + 46, + 438, + 288, + 570 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 438, + 288, + 570 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 46, + 438, + 288, + 570 + ], + "type": "text", + "content": " are updated on-the-fly in an alternating manner." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 574, + 288, + 634 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 574, + 288, + 634 + ], + "spans": [ + { + "bbox": [ + 46, + 574, + 288, + 634 + ], + "type": "text", + "content": "Optimizing the transition matrix. In step " + }, + { + "bbox": [ + 46, + 574, + 288, + 634 + ], + "type": "inline_equation", + "content": "t + 1" + }, + { + "bbox": [ + 46, + 574, + 288, + 634 + ], + "type": "text", + "content": " of optimizing " + }, + { + "bbox": [ + 46, + 574, + 288, + 634 + ], + "type": "inline_equation", + "content": "T^t" + }, + { + "bbox": [ + 46, + 574, + 288, + 634 + ], + "type": "text", + "content": ", guided by the attack loss, surrogate model weights " + }, + { + "bbox": [ + 46, + 574, + 288, + 634 + ], + "type": "inline_equation", + "content": "\\theta_g^t" + }, + { + "bbox": [ + 46, + 574, + 288, + 634 + ], + "type": "text", + "content": " are updated in a one-step-forward manner to obtain " + }, + { + "bbox": [ + 46, + 574, + 288, + 634 + ], + "type": "inline_equation", + "content": "\\theta_g^{t + 1}" + }, + { + "bbox": [ + 46, + 574, + 288, + 634 + ], + "type": "text", + "content": ". Specifically, with the fixed " + }, + { + "bbox": [ + 46, + 574, + 288, + 634 + ], + "type": "inline_equation", + "content": "T^t" + }, + { + "bbox": [ + 46, + 574, + 288, + 634 + ], + "type": "text", + "content": ", then " + }, + { + "bbox": [ + 46, + 574, + 288, + 634 + ], + "type": "inline_equation", + "content": "\\theta_g^t" + }, + { + "bbox": [ + 46, + 574, + 288, + 634 + ], + "type": "text", + "content": " is updated via gradient descent with learning rate " + }, + { + "bbox": [ + 46, + 574, + 288, + 634 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 46, + 574, + 288, + 634 + ], + "type": "text", + "content": " as:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 100, + 642, + 287, + 658 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 100, + 642, + 287, + 658 + ], + "spans": [ + { + "bbox": [ + 100, + 642, + 287, + 658 + ], + "type": "interline_equation", + "content": "\\tilde {\\boldsymbol {\\theta}} _ {g} ^ {t + 1} \\left(T ^ {t}\\right) = \\boldsymbol {\\theta} _ {g} ^ {t} - \\alpha \\cdot \\bigtriangledown_ {\\boldsymbol {\\theta} _ {g} ^ {t}} \\mathcal {L} _ {\\text {a t t a c k}}. \\tag {8}", + "image_path": "3f90e084c26030f9bed9c6e804b92ea7785d512f204820cf576cdd1eea2393d2.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 665, + 287, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 287, + 712 + ], + "type": "text", + "content": "Note that the surrogate model " + }, + { + "bbox": [ + 46, + 665, + 287, + 712 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 46, + 665, + 287, + 712 + ], + "type": "text", + "content": " does not actually update in this step, but makes preparations for estimating " + }, + { + "bbox": [ + 46, + 665, + 287, + 712 + ], + "type": "inline_equation", + "content": "T^{t + 1}" + }, + { + "bbox": [ + 46, + 665, + 287, + 712 + ], + "type": "text", + "content": ". We then optimize " + }, + { + "bbox": [ + 46, + 665, + 287, + 712 + ], + "type": "inline_equation", + "content": "T^t" + }, + { + "bbox": [ + 46, + 665, + 287, + 712 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 46, + 665, + 287, + 712 + ], + "type": "inline_equation", + "content": "\\tilde{\\theta}_g^{t + 1}" + }, + { + "bbox": [ + 46, + 665, + 287, + 712 + ], + "type": "text", + "content": " fixed under the defense loss. The motivation is that we would like to find a " + }, + { + "bbox": [ + 46, + 665, + 287, + 712 + ], + "type": "inline_equation", + "content": "T^{t + 1}" + }, + { + "bbox": [ + 46, + 665, + 287, + 712 + ], + "type": "text", + "content": " that achieves" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 72, + 484, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 484, + 84 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 484, + 84 + ], + "type": "text", + "content": "Algorithm 1 The overall training procedure" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 87, + 424, + 100 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 87, + 424, + 100 + ], + "spans": [ + { + "bbox": [ + 306, + 87, + 424, + 100 + ], + "type": "text", + "content": "Input: Query dataset " + }, + { + "bbox": [ + 306, + 87, + 424, + 100 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{\\mathrm{query}}" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 336, + 100, + 519, + 112 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 336, + 100, + 519, + 112 + ], + "spans": [ + { + "bbox": [ + 336, + 100, + 519, + 112 + ], + "type": "text", + "content": "The surrogate model " + }, + { + "bbox": [ + 336, + 100, + 519, + 112 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 336, + 100, + 519, + 112 + ], + "type": "text", + "content": " and its parameters " + }, + { + "bbox": [ + 336, + 100, + 519, + 112 + ], + "type": "inline_equation", + "content": "\\theta_{g}" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 336, + 112, + 459, + 122 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 336, + 112, + 459, + 122 + ], + "spans": [ + { + "bbox": [ + 336, + 112, + 459, + 122 + ], + "type": "text", + "content": "The noise transition matrix " + }, + { + "bbox": [ + 336, + 112, + 459, + 122 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 336, + 112, + 459, + 122 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 336, + 124, + 448, + 134 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 336, + 124, + 448, + 134 + ], + "spans": [ + { + "bbox": [ + 336, + 124, + 448, + 134 + ], + "type": "text", + "content": "The number of iterations " + }, + { + "bbox": [ + 336, + 124, + 448, + 134 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 336, + 124, + 448, + 134 + ], + "type": "text", + "content": ";" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 336, + 136, + 448, + 146 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 336, + 136, + 448, + 146 + ], + "spans": [ + { + "bbox": [ + 336, + 136, + 448, + 146 + ], + "type": "text", + "content": "The multiplicative factor " + }, + { + "bbox": [ + 336, + 136, + 448, + 146 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 336, + 136, + 448, + 146 + ], + "type": "text", + "content": ";" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 306, + 148, + 356, + 157 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 148, + 356, + 157 + ], + "spans": [ + { + "bbox": [ + 306, + 148, + 356, + 157 + ], + "type": "text", + "content": "Procedure:" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 312, + 161, + 545, + 255 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 312, + 161, + 403, + 173 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 161, + 403, + 173 + ], + "spans": [ + { + "bbox": [ + 312, + 161, + 403, + 173 + ], + "type": "text", + "content": "1: Initialize " + }, + { + "bbox": [ + 312, + 161, + 403, + 173 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 312, + 161, + 403, + 173 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 312, + 161, + 403, + 173 + ], + "type": "inline_equation", + "content": "\\theta_{g}" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 312, + 174, + 394, + 184 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 174, + 394, + 184 + ], + "spans": [ + { + "bbox": [ + 312, + 174, + 394, + 184 + ], + "type": "text", + "content": "2: for " + }, + { + "bbox": [ + 312, + 174, + 394, + 184 + ], + "type": "inline_equation", + "content": "i = 1" + }, + { + "bbox": [ + 312, + 174, + 394, + 184 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 312, + 174, + 394, + 184 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 312, + 174, + 394, + 184 + ], + "type": "text", + "content": " do" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 312, + 186, + 482, + 197 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 186, + 482, + 197 + ], + "spans": [ + { + "bbox": [ + 312, + 186, + 482, + 197 + ], + "type": "text", + "content": "3: Fetch a random batch " + }, + { + "bbox": [ + 312, + 186, + 482, + 197 + ], + "type": "inline_equation", + "content": "\\mathcal{B}" + }, + { + "bbox": [ + 312, + 186, + 482, + 197 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 312, + 186, + 482, + 197 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{\\mathrm{query}}" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 312, + 198, + 545, + 220 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 198, + 545, + 220 + ], + "spans": [ + { + "bbox": [ + 312, + 198, + 545, + 220 + ], + "type": "text", + "content": "4: Calculate the one-step-forward surrogate model weights via Eq. (8);" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 312, + 221, + 509, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 221, + 509, + 232 + ], + "spans": [ + { + "bbox": [ + 312, + 221, + 509, + 232 + ], + "type": "text", + "content": "5: Update the transition matrix via Eq. (9-12);" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 312, + 233, + 518, + 244 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 233, + 518, + 244 + ], + "spans": [ + { + "bbox": [ + 312, + 233, + 518, + 244 + ], + "type": "text", + "content": "6: Update surrogate model weights via Eq. (13);" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 312, + 246, + 357, + 255 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 246, + 357, + 255 + ], + "spans": [ + { + "bbox": [ + 312, + 246, + 357, + 255 + ], + "type": "text", + "content": "7: end for" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 258, + 473, + 270 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 258, + 473, + 270 + ], + "spans": [ + { + "bbox": [ + 306, + 258, + 473, + 270 + ], + "type": "text", + "content": "Output: Learned noise transition matrix." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 305, + 293, + 547, + 328 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 293, + 547, + 328 + ], + "spans": [ + { + "bbox": [ + 305, + 293, + 547, + 328 + ], + "type": "text", + "content": "minimal utility trade-offs between the defender and attacker. The transition matrix " + }, + { + "bbox": [ + 305, + 293, + 547, + 328 + ], + "type": "inline_equation", + "content": "T^{t}" + }, + { + "bbox": [ + 305, + 293, + 547, + 328 + ], + "type": "text", + "content": " is updated via gradient descent with learning rate " + }, + { + "bbox": [ + 305, + 293, + 547, + 328 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 305, + 293, + 547, + 328 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 365, + 335, + 545, + 349 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 365, + 335, + 545, + 349 + ], + "spans": [ + { + "bbox": [ + 365, + 335, + 545, + 349 + ], + "type": "interline_equation", + "content": "w ^ {t + 1} = T ^ {t} - \\beta \\cdot \\bigtriangledown_ {T ^ {t}} \\mathcal {L} _ {\\text {D e f s e n s e}}. \\tag {9}", + "image_path": "b21f4f37460c362d01d9e5d2b57fb22585c706715d46b0410381756bd2fd69c0.jpg" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 306, + 357, + 467, + 369 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 357, + 467, + 369 + ], + "spans": [ + { + "bbox": [ + 306, + 357, + 467, + 369 + ], + "type": "text", + "content": "Apply chain rule to Eq. (9), we can get:" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 315, + 373, + 545, + 407 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 373, + 545, + 407 + ], + "spans": [ + { + "bbox": [ + 315, + 373, + 545, + 407 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} w ^ {t + 1} = T ^ {t} - \\beta \\cdot \\left\\{\\lambda \\cdot \\nabla_ {T ^ {t}} \\mathcal {L} _ {\\text {a l i g n}} \\right. \\\\ + (1 - \\lambda) \\cdot \\bigtriangledown_ {\\tilde {\\boldsymbol {\\theta}} _ {g} ^ {t + 1}} \\mathcal {L} _ {\\mathrm {m i s}} \\cdot \\left(- \\alpha \\bigtriangledown_ {\\boldsymbol {\\theta} _ {g} ^ {t}, T ^ {t}} ^ {2} \\mathcal {L} _ {\\text {a t t a c k}}\\right) \\}. \\tag {10} \\\\ \\end{array}", + "image_path": "269cd91bae603a553ac776e7529fd9645056645dca53bf070d7a44add91412fb.jpg" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 305, + 413, + 547, + 473 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 413, + 547, + 473 + ], + "spans": [ + { + "bbox": [ + 305, + 413, + 547, + 473 + ], + "type": "text", + "content": "Note that " + }, + { + "bbox": [ + 305, + 413, + 547, + 473 + ], + "type": "inline_equation", + "content": "w^{t + 1}" + }, + { + "bbox": [ + 305, + 413, + 547, + 473 + ], + "type": "text", + "content": " is not the final transition matrix, as its entries may not satisfy the properties of a valid transition matrix. Specifically, all elements must be non-negative and each row must be normalized. To address this, we first replace any negative elements in " + }, + { + "bbox": [ + 305, + 413, + 547, + 473 + ], + "type": "inline_equation", + "content": "w^{t + 1}" + }, + { + "bbox": [ + 305, + 413, + 547, + 473 + ], + "type": "text", + "content": " with zeros via:" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 377, + 479, + 545, + 493 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 377, + 479, + 545, + 493 + ], + "spans": [ + { + "bbox": [ + 377, + 479, + 545, + 493 + ], + "type": "interline_equation", + "content": "\\tilde {w} ^ {t + 1} = \\max \\left(w ^ {t + 1}, 0\\right). \\tag {11}", + "image_path": "7e3062cfc08f00bb66b97acdd829892ad0c3b7be1031903607f02fd5f1bcd33d.jpg" + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 305, + 500, + 547, + 551 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 500, + 547, + 551 + ], + "spans": [ + { + "bbox": [ + 305, + 500, + 547, + 551 + ], + "type": "text", + "content": "Then, each row is normalized to get the final valid transition matrix. For the " + }, + { + "bbox": [ + 305, + 500, + 547, + 551 + ], + "type": "inline_equation", + "content": "j^{\\mathrm{th}}" + }, + { + "bbox": [ + 305, + 500, + 547, + 551 + ], + "type": "text", + "content": " row of " + }, + { + "bbox": [ + 305, + 500, + 547, + 551 + ], + "type": "inline_equation", + "content": "\\tilde{w}^{t + 1}" + }, + { + "bbox": [ + 305, + 500, + 547, + 551 + ], + "type": "text", + "content": ", which indicates the probabilities of class " + }, + { + "bbox": [ + 305, + 500, + 547, + 551 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 305, + 500, + 547, + 551 + ], + "type": "text", + "content": " transitioning to other classes, a normalization is performed to obtain the final matrix " + }, + { + "bbox": [ + 305, + 500, + 547, + 551 + ], + "type": "inline_equation", + "content": "T_{j,i}^{t + 1}" + }, + { + "bbox": [ + 305, + 500, + 547, + 551 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 35 + }, + { + "bbox": [ + 312, + 556, + 545, + 597 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 556, + 545, + 597 + ], + "spans": [ + { + "bbox": [ + 312, + 556, + 545, + 597 + ], + "type": "interline_equation", + "content": "T _ {j, i} ^ {t + 1} = \\frac {\\tilde {w} _ {j , i} ^ {t + 1}}{\\sum_ {i} \\tilde {w} _ {j , i} ^ {t + 1} + \\delta \\left(\\sum_ {i} \\tilde {w} _ {j , i} ^ {t + 1}\\right)}, \\delta (a) = \\left\\{ \\begin{array}{l} 1, \\text {i f} a = 0 \\\\ 0, \\text {i f} a \\neq 0 \\end{array} \\right. \\tag {12}", + "image_path": "ddfabe6882d19ea783f2ebebfec4296f530ed32732320375185f76b869f640dc.jpg" + } + ] + } + ], + "index": 36 + }, + { + "bbox": [ + 306, + 598, + 471, + 611 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 598, + 471, + 611 + ], + "spans": [ + { + "bbox": [ + 306, + 598, + 471, + 611 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 306, + 598, + 471, + 611 + ], + "type": "inline_equation", + "content": "\\delta (\\cdot)" + }, + { + "bbox": [ + 306, + 598, + 471, + 611 + ], + "type": "text", + "content": " is used to avoid division by 0." + } + ] + } + ], + "index": 37 + }, + { + "bbox": [ + 306, + 614, + 545, + 674 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 614, + 545, + 674 + ], + "spans": [ + { + "bbox": [ + 306, + 614, + 545, + 674 + ], + "type": "text", + "content": "Updating surrogate model weights. After optimizing the matrix, we conduct the surrogate model update by keeping " + }, + { + "bbox": [ + 306, + 614, + 545, + 674 + ], + "type": "inline_equation", + "content": "T^{t + 1}" + }, + { + "bbox": [ + 306, + 614, + 545, + 674 + ], + "type": "text", + "content": " fixed and minimizing the attack loss. New surrogate model weights " + }, + { + "bbox": [ + 306, + 614, + 545, + 674 + ], + "type": "inline_equation", + "content": "\\theta_{g}^{t + 1}" + }, + { + "bbox": [ + 306, + 614, + 545, + 674 + ], + "type": "text", + "content": " are obtained via gradient descent with learning rate " + }, + { + "bbox": [ + 306, + 614, + 545, + 674 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 306, + 614, + 545, + 674 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 38 + }, + { + "bbox": [ + 369, + 681, + 545, + 696 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 369, + 681, + 545, + 696 + ], + "spans": [ + { + "bbox": [ + 369, + 681, + 545, + 696 + ], + "type": "interline_equation", + "content": "\\boldsymbol {\\theta} _ {g} ^ {t + 1} = \\boldsymbol {\\theta} _ {g} ^ {t} - \\gamma \\cdot \\bigtriangledown \\boldsymbol {\\theta} _ {g} ^ {t} \\mathcal {L} _ {\\text {a t t a c k}}. \\tag {13}", + "image_path": "a98fad271b96b3cc3b2b97531b2ce5635407cd519a050e3cbb4f19b080353886.jpg" + } + ] + } + ], + "index": 39 + }, + { + "bbox": [ + 306, + 701, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 701, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 701, + 547, + 713 + ], + "type": "text", + "content": "The whole training procedure is summarized in Algorithm 1." + } + ] + } + ], + "index": 40 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "24309" + } + ] + } + ], + "index": 41 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 54, + 70, + 541, + 302 + ], + "blocks": [ + { + "bbox": [ + 54, + 70, + 541, + 302 + ], + "lines": [ + { + "bbox": [ + 54, + 70, + 541, + 302 + ], + "spans": [ + { + "bbox": [ + 54, + 70, + 541, + 302 + ], + "type": "table", + "html": "
MethodCIFAR100 → CIFAR10CIFAR10 → CIFAR100Caltech256 → CUB200
ΔAcc=1ΔAcc=2ΔAcc=5ΔAcc=1ΔAcc=2ΔAcc=5ΔAcc=1ΔAcc=2ΔAcc=5
Random84.1777.2174.2143.2339.8132.1233.5031.1428.01
AM77.1072.9357.5639.8232.9819.3626.1721.2714.31
MAD80.5777.5469.0736.5632.5926.6537.1635.2331.50
NASTY76.6574.3164.1938.5333.6126.9531.0628.2324.70
GARD²77.4273.8871.4041.2139.2935.9435.6933.6732.34
MIN-IP78.8873.4070.6642.3338.6734.2634.7332.2628.94
EMMA61.7952.9142.2128.0118.3413.3719.1015.7913.11
ImageNet-C10 → CIFAR10ImageNet-C100 → CIFAR100ImageNet-CUB200 → CUB200
ΔAcc=1ΔAcc=2ΔAcc=5ΔAcc=1ΔAcc=2ΔAcc=5ΔAcc=1ΔAcc=2ΔAcc=5
Random88.9988.1887.0763.1661.7658.6446.5844.0640.49
AM87.5584.6076.7260.5756.0947.1045.1239.6530.15
MAD88.6785.7078.4458.7456.2350.8550.9848.5744.47
NASTY87.3385.3381.6859.7956.0548.9546.5643.2337.76
GARD²87.2986.0685.9358.3956.7153.6350.2747.5543.11
MIN-IP87.4086.8184.4561.1756.3551.9850.0946.4542.06
EMMA84.1080.6077.5256.7349.1831.4836.3533.1026.33
", + "image_path": "f05755d3e179c3574b171dad9ce29b1ed641c449d5f783f723824d0e3b696da2.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 48, + 339, + 287, + 405 + ], + "blocks": [ + { + "bbox": [ + 46, + 305, + 547, + 329 + ], + "lines": [ + { + "bbox": [ + 46, + 305, + 547, + 329 + ], + "spans": [ + { + "bbox": [ + 46, + 305, + 547, + 329 + ], + "type": "text", + "content": "Table 1. Accuracy (%) of attacker model on defender's test set under knowledge-limited (top row) and distribution-aware (bottom row) attack scenarios with varying budgets. The best result in each column is highlighted in bold." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 48, + 339, + 287, + 405 + ], + "lines": [ + { + "bbox": [ + 48, + 339, + 287, + 405 + ], + "spans": [ + { + "bbox": [ + 48, + 339, + 287, + 405 + ], + "type": "table", + "html": "
Eval DataDefender AccuracyAttacker Accuracy
KLDA
CIFAR1095.33 ± 0.1385.17 ± 0.8391.78 ± 0.19
CIFAR10075.96 ± 0.0951.95 ± 1.1266.82 ± 0.16
CUB20081.44 ± 0.4648.93 ± 0.3762.12 ± 0.22
", + "image_path": "977833f155ce48897c67d0303f5dd5caed4798f1b8c08de9a202f444e62538b8.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 407, + 288, + 431 + ], + "lines": [ + { + "bbox": [ + 46, + 407, + 288, + 431 + ], + "spans": [ + { + "bbox": [ + 46, + 407, + 288, + 431 + ], + "type": "text", + "content": "Table 2. Accuracy " + }, + { + "bbox": [ + 46, + 407, + 288, + 431 + ], + "type": "inline_equation", + "content": "(\\%)" + }, + { + "bbox": [ + 46, + 407, + 288, + 431 + ], + "type": "text", + "content": " of the defender model and the attacker model with no defense applied." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 47, + 441, + 128, + 454 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 441, + 128, + 454 + ], + "spans": [ + { + "bbox": [ + 47, + 441, + 128, + 454 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 461, + 287, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 461, + 287, + 498 + ], + "spans": [ + { + "bbox": [ + 46, + 461, + 287, + 498 + ], + "type": "text", + "content": "To demonstrate the effectiveness of our proposed defense method, we conduct experiments on multiple image datasets and present the results in this section." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 508, + 155, + 521 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 508, + 155, + 521 + ], + "spans": [ + { + "bbox": [ + 47, + 508, + 155, + 521 + ], + "type": "text", + "content": "5.1. Experiment Setup" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 528, + 288, + 672 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 528, + 288, + 672 + ], + "spans": [ + { + "bbox": [ + 46, + 528, + 288, + 672 + ], + "type": "text", + "content": "Datasets. We employ three datasets for evaluation: CIFAR10, CIFAR100 [26] and CUB200 [57]. For each dataset, we explore two attacker scenarios - knowledge-limited (KL) and distribution-aware (DA). The KL scenario uses query sets of CIFAR100, CIFAR10, and Caltech256 [8], paired with CIFAR10, CIFAR100, and CUB200 evaluation datasets respectively. The DA scenario employs tailored query sets from ImageNet-1K [4] manually selected by prior work [38], specifically ImageNet-C10, ImageNet-C100, and ImageNet-CUB200 containing 183,763, 161,653, and 30,000 examples matched to the evaluation datasets, with their further details provided in the Appendix 9." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 45, + 677, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 677, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 45, + 677, + 288, + 715 + ], + "type": "text", + "content": "Compared methods. We compare our method against the following six model stealing defense baselines: (1) No Defense, a naive approach where no defense is attempted. (2)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 338, + 547, + 660 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 338, + 547, + 660 + ], + "spans": [ + { + "bbox": [ + 304, + 338, + 547, + 660 + ], + "type": "text", + "content": "Random, an interpolation-based method that perturbs the defender's predicted posterior by interpolating it with a one-hot posterior vector, where the index of single non-zero entry is randomly selected from the non-argmax prediction labels. (3) AM [20], a detection-based method that identifies out-of-distribution queries and interpolates with posteriors from a misinformation network, where the interpolation weight is denoted by the confidence of the argmax prediction label. (4) MAD [40], an accuracy-constrained method that adds controlled noise to original posteriors so that it maximizes the angular deviation between the perturbed and original gradient signal on a surrogate model. The surrogate model is randomly initialized. (5) NASTY [34], a self-undermining knowledge distillation approach which trains a nasty teacher network by maximizing the difference between the output of the nasty teacher and a normal pretrained network. (5) " + }, + { + "bbox": [ + 304, + 338, + 547, + 660 + ], + "type": "inline_equation", + "content": "GARD^2" + }, + { + "bbox": [ + 304, + 338, + 547, + 660 + ], + "type": "text", + "content": " [38], a coordinated gradient redirection defense method, which perturbs original posteriors by setting the target gradient direction to be the all-ones vector. This redirects the attacker to update towards a consistent target direction. (6) MIN-IP [38], an uncoordinated gradient redirection defense method, which perturbs original posteriors by setting the target gradient direction opposite the original gradient direction, redirecting the update gradient of attackers opposite to the original gradients. The surrogate model of MIN-IP and " + }, + { + "bbox": [ + 304, + 338, + 547, + 660 + ], + "type": "inline_equation", + "content": "GARD^2" + }, + { + "bbox": [ + 304, + 338, + 547, + 660 + ], + "type": "text", + "content": " train on the attacker's queries with early stopping after 10 epochs." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 661, + 547, + 697 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 661, + 547, + 697 + ], + "spans": [ + { + "bbox": [ + 304, + 661, + 547, + 697 + ], + "type": "text", + "content": "For fair comparisons, we basically use the same network architecture, learning rate, optimizer, and augmentation strategy across all the comparing methods." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 701, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 701, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 306, + 701, + 547, + 714 + ], + "type": "text", + "content": "Implementation. Our experiment pipeline has three stages:" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "24310" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 52, + 70, + 160, + 164 + ], + "blocks": [ + { + "bbox": [ + 52, + 70, + 160, + 164 + ], + "lines": [ + { + "bbox": [ + 52, + 70, + 160, + 164 + ], + "spans": [ + { + "bbox": [ + 52, + 70, + 160, + 164 + ], + "type": "image", + "image_path": "f22df67711c7e2a03a84d40efbea92433dd69078620db61350f83067e53fae56.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 59, + 167, + 153, + 177 + ], + "lines": [ + { + "bbox": [ + 59, + 167, + 153, + 177 + ], + "spans": [ + { + "bbox": [ + 59, + 167, + 153, + 177 + ], + "type": "text", + "content": "(a) Heatmap of transition matrix." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 170, + 70, + 281, + 164 + ], + "blocks": [ + { + "bbox": [ + 170, + 70, + 281, + 164 + ], + "lines": [ + { + "bbox": [ + 170, + 70, + 281, + 164 + ], + "spans": [ + { + "bbox": [ + 170, + 70, + 281, + 164 + ], + "type": "image", + "image_path": "1179097a99be1a5469ca28c057017ddde9033864b7c0a9b836c7feb0b58dcae8.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 177, + 167, + 272, + 177 + ], + "lines": [ + { + "bbox": [ + 177, + 167, + 272, + 177 + ], + "spans": [ + { + "bbox": [ + 177, + 167, + 272, + 177 + ], + "type": "text", + "content": "(b) Heatmap of transition matrix." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 286, + 72, + 405, + 164 + ], + "blocks": [ + { + "bbox": [ + 286, + 72, + 405, + 164 + ], + "lines": [ + { + "bbox": [ + 286, + 72, + 405, + 164 + ], + "spans": [ + { + "bbox": [ + 286, + 72, + 405, + 164 + ], + "type": "image", + "image_path": "c9df9a5cb987d1faf76f39ed786414a0649152acfc0e99f4e72ff535f22715be.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 168, + 394, + 177 + ], + "lines": [ + { + "bbox": [ + 304, + 168, + 394, + 177 + ], + "spans": [ + { + "bbox": [ + 304, + 168, + 394, + 177 + ], + "type": "text", + "content": "(c) Average predicted posterior" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 417, + 72, + 536, + 164 + ], + "blocks": [ + { + "bbox": [ + 417, + 72, + 536, + 164 + ], + "lines": [ + { + "bbox": [ + 417, + 72, + 536, + 164 + ], + "spans": [ + { + "bbox": [ + 417, + 72, + 536, + 164 + ], + "type": "image", + "image_path": "271bc120cbe6b8c77faeb5962cc6cddf7a388916675e414315814784e1d0fdcf.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 435, + 167, + 525, + 177 + ], + "lines": [ + { + "bbox": [ + 435, + 167, + 525, + 177 + ], + "spans": [ + { + "bbox": [ + 435, + 167, + 525, + 177 + ], + "type": "text", + "content": "(d) Average perturbed posterior" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 56, + 227, + 279, + 345 + ], + "blocks": [ + { + "bbox": [ + 46, + 180, + 547, + 214 + ], + "lines": [ + { + "bbox": [ + 46, + 180, + 547, + 214 + ], + "spans": [ + { + "bbox": [ + 46, + 180, + 547, + 214 + ], + "type": "text", + "content": "Figure 3. Qualitative analysis of the transition matrix. (a-b): Heatmaps of optimized transition matrixes for distribution-aware attacks (a) Imagenet-C10→CIFAR10 and (b) Imagenet-C100→CIFAR100, respectively. (c-d): Average (c) predicted and (d) perturbed posteriors across the query set of (top) Imagenet-C10 and (bottom) Imagenet-C100 datasets." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 56, + 227, + 279, + 345 + ], + "lines": [ + { + "bbox": [ + 56, + 227, + 279, + 345 + ], + "spans": [ + { + "bbox": [ + 56, + 227, + 279, + 345 + ], + "type": "image", + "image_path": "d1373fed305d5394c88c3e75fbe04b70003b7b716967e5b52f80a81f39625b03.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 53, + 354, + 279, + 365 + ], + "lines": [ + { + "bbox": [ + 53, + 354, + 279, + 365 + ], + "spans": [ + { + "bbox": [ + 53, + 354, + 279, + 365 + ], + "type": "text", + "content": "Figure 4. Visualizations of predicted and perturbed posteriors." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 376, + 287, + 448 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 376, + 287, + 448 + ], + "spans": [ + { + "bbox": [ + 46, + 376, + 287, + 448 + ], + "type": "text", + "content": "First, the defender model trains on each evaluation dataset. Second, defense methods generate perturbed posteriors for query dataset and evaluation dataset at various defense budgets. Finally, the attacker model trains on the collected query-response pairs. We denote attacks with query data " + }, + { + "bbox": [ + 46, + 376, + 287, + 448 + ], + "type": "inline_equation", + "content": "\\mathcal{Q}" + }, + { + "bbox": [ + 46, + 376, + 287, + 448 + ], + "type": "text", + "content": " and evaluation data " + }, + { + "bbox": [ + 46, + 376, + 287, + 448 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 46, + 376, + 287, + 448 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 46, + 376, + 287, + 448 + ], + "type": "inline_equation", + "content": "\\mathcal{Q} \\to \\mathcal{D}" + }, + { + "bbox": [ + 46, + 376, + 287, + 448 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 450, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 450, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 450, + 289, + 715 + ], + "type": "text", + "content": "We use ResNet34 for CIFAR10/100 and ResNet50 pretrained on ImageNet for CUB200. For all models training, we utilize SGD as the optimizer with a momentum of 0.9 and weight decay of 5e-4 for 100 epochs. Initial learning rates are set as 0.01 for CUB200 and 0.1 for other datasets, with a cosine annealing schedule. For optimizing the transition matrix, the learning rate " + }, + { + "bbox": [ + 46, + 450, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 46, + 450, + 289, + 715 + ], + "type": "text", + "content": " is set as 0.3 for CIFAR10 and 1.0 for other datasets. The surrogate model's one-step-forward learning rate " + }, + { + "bbox": [ + 46, + 450, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 46, + 450, + 289, + 715 + ], + "type": "text", + "content": " and the actual learning rate " + }, + { + "bbox": [ + 46, + 450, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 46, + 450, + 289, + 715 + ], + "type": "text", + "content": " are set the same value as specified in the model training settings. Additionally, an early stopping mechanism is set to terminates optimization when the average change in " + }, + { + "bbox": [ + 46, + 450, + 289, + 715 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 46, + 450, + 289, + 715 + ], + "type": "text", + "content": " between consecutive epochs falls below a specified threshold. The threshold is 1e-2 for CIFAR10 and 1e-4 for other datasets. Furthermore, the balancing factor " + }, + { + "bbox": [ + 46, + 450, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 46, + 450, + 289, + 715 + ], + "type": "text", + "content": " is adjusted to achieve different utility budgets. Recent work [38] proposed a novel approach to initialize the surrogate model. They suggested to train the surrogate model through distilling from the victim model with early stopping. This would make it more representative of the actual attacker's behavior. Following this approach, we initialize our surrogate model by training it on the query dataset with the same early stopping mechanism." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 228, + 421, + 240 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 228, + 421, + 240 + ], + "spans": [ + { + "bbox": [ + 306, + 228, + 421, + 240 + ], + "type": "text", + "content": "5.2. Experiment Results" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 247, + 547, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 247, + 547, + 415 + ], + "spans": [ + { + "bbox": [ + 304, + 247, + 547, + 415 + ], + "type": "text", + "content": "Table 2 presents the accuracy of the defender model and the attacker model, without any defense applied under two different query distributions. This shows the upper bound for the usage performance of the defender and stealable performance of the attacker. Table 1 compares the accuracy of an attacker's clone model at different defender accuracy drop budgets " + }, + { + "bbox": [ + 304, + 247, + 547, + 415 + ], + "type": "inline_equation", + "content": "\\Delta" + }, + { + "bbox": [ + 304, + 247, + 547, + 415 + ], + "type": "text", + "content": "Acc. The budget refers to the average decrease in accuracy the defender causes across the test set by adding perturbations. Corresponding visualizations of these results are provided in the Appendix 10. As shown, our defense method achieves the lowest utility trade-off across budgets. For example, in the CIFAR100 " + }, + { + "bbox": [ + 304, + 247, + 547, + 415 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 304, + 247, + 547, + 415 + ], + "type": "text", + "content": " CIFAR10 attack scenario, our method reduces the attacker model's accuracy by " + }, + { + "bbox": [ + 304, + 247, + 547, + 415 + ], + "type": "inline_equation", + "content": "23\\%" + }, + { + "bbox": [ + 304, + 247, + 547, + 415 + ], + "type": "text", + "content": " while only decreasing the defender's accuracy by " + }, + { + "bbox": [ + 304, + 247, + 547, + 415 + ], + "type": "inline_equation", + "content": "1\\%" + }, + { + "bbox": [ + 304, + 247, + 547, + 415 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 425, + 423, + 437 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 425, + 423, + 437 + ], + "spans": [ + { + "bbox": [ + 306, + 425, + 423, + 437 + ], + "type": "text", + "content": "5.3. Qualitative Analysis" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 444, + 547, + 564 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 444, + 547, + 564 + ], + "spans": [ + { + "bbox": [ + 304, + 444, + 547, + 564 + ], + "type": "text", + "content": "Perturbed posteriors make cloning difficult. We visualize the predicted and perturbed posteriors using t-SNE in Figure 4. The perturbed posteriors clearly exhibit degraded inter-class distances compared to the predicted posteriors. This degradation obfuscates the decision boundaries between classes, making it difficult for the attacker to clone the defender model by imitating the posteriors. Notably, such effective defenses are achieved solely through linear mapping of the posteriors, demonstrating the simplicity and efficiency of our proposed noise perturbation strategy." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 570, + 548, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 548, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 548, + 713 + ], + "type": "text", + "content": "The transition matrix induces challenging noise. We analyze the defense details of two attack scenarios: Imagenet-C10→CIFAR10 and Imagenet-C100→CIFAR100 in Figure 3. Specifically, Figure 3 (a-b) visualizes corresponding optimized transition matrices. It can be observed that the diagonal elements are maximized for fidelity, while latent inter-label relationships are also exploited. As shown the optimized matrix in Figure 3 (a), despite it causes a " + }, + { + "bbox": [ + 304, + 570, + 548, + 713 + ], + "type": "inline_equation", + "content": "2.35\\%" + }, + { + "bbox": [ + 304, + 570, + 548, + 713 + ], + "type": "text", + "content": " drop in the defender's accuracy on the test set, the optimization makes it have a noise rate as high as " + }, + { + "bbox": [ + 304, + 570, + 548, + 713 + ], + "type": "inline_equation", + "content": "77\\%" + }, + { + "bbox": [ + 304, + 570, + 548, + 713 + ], + "type": "text", + "content": ". This negatively affects the cloning process of the attacker, reducing the cloned model's accuracy by " + }, + { + "bbox": [ + 304, + 570, + 548, + 713 + ], + "type": "inline_equation", + "content": "14\\%" + }, + { + "bbox": [ + 304, + 570, + 548, + 713 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "24311" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 168 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 168 + ], + "type": "text", + "content": "In Figures 3 (c-d), we visualize the average posterior of predicted and perturbed predictions. Given that ImageNet-C10/C100 follow long-tail distributions as illustrated in appendix 9, the predicted posteriors also exhibit long-tailed shapes as expected. However, the perturbations significantly increase the confidence of tail classes, distorting the true inter-class relationships. This distortion of class relationships also makes stealing more difficult for the attacker." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 177, + 141, + 190 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 177, + 141, + 190 + ], + "spans": [ + { + "bbox": [ + 47, + 177, + 141, + 190 + ], + "type": "text", + "content": "5.4. Ablation Study" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 195, + 290, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 195, + 290, + 327 + ], + "spans": [ + { + "bbox": [ + 46, + 195, + 290, + 327 + ], + "type": "text", + "content": "Effectiveness of optimized transition matrix. To validate the effectiveness of the optimized transition matrix, we compare the defense performance between it and the uniform transition matrix with an identical noise rate. The results in Table 3 show that the optimized transition matrix can further reduce the attacker's accuracy while maintaining the defender's accuracy. This demonstrates that the optimized matrix provides improved defense over the uniform matrix. Overall, this shows that capturing the inter-class relationships during optimization contributes to generating a matrix that is more adversarial against the attacker." + } + ] + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 48, + 337, + 287, + 395 + ], + "blocks": [ + { + "bbox": [ + 48, + 337, + 287, + 395 + ], + "lines": [ + { + "bbox": [ + 48, + 337, + 287, + 395 + ], + "spans": [ + { + "bbox": [ + 48, + 337, + 287, + 395 + ], + "type": "table", + "html": "
Attack ScenariosOptimized matrixUniform matrix
DefenderAttackerDefenderAttacker
CIFAR100→CIFAR1093.87±0.1432.41±7.5494.81±0.0671.43±2.26
CIFAR10→CIFAR10074.48±0.2515.65±0.7274.75±0.0735.67±3.44
Caltech256→CUB20080.45±0.1318.25±0.4380.89±0.2230.15±0.21
", + "image_path": "7b6234df0fda0c1a7a0bf5ff0856c9a6691c79776081dea9e7e05979aa21db03.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 46, + 398, + 288, + 431 + ], + "lines": [ + { + "bbox": [ + 46, + 398, + 288, + 431 + ], + "spans": [ + { + "bbox": [ + 46, + 398, + 288, + 431 + ], + "type": "text", + "content": "Table 3. Defender and attacker accuracy " + }, + { + "bbox": [ + 46, + 398, + 288, + 431 + ], + "type": "inline_equation", + "content": "(\\%)" + }, + { + "bbox": [ + 46, + 398, + 288, + 431 + ], + "type": "text", + "content": " with optimized and uniform transition matrix under attack scenarios for CIFAR10, CIFAR100 and CUB200 evaluation datasets." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 445, + 288, + 625 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 445, + 288, + 625 + ], + "spans": [ + { + "bbox": [ + 46, + 445, + 288, + 625 + ], + "type": "text", + "content": "Effects of the number of queries. We also verify the effectiveness of our defense against attacks with varying number of queries in Figure 5. The EMMA-top1 result is depicted as a reference, which denotes the victim model only outputting the predicted label rather than the full posterior. As formalized in Theorem 1, when the victim model assigns full confidence to the true label, the attacker's cloning task degrades to learning from noisy labels. Thus, EMMA-top1 signifies the optimal defense under the same transition matrix. EMMA defense consistently forces degradation of the attacker accuracy across all query set sizes, remaining close to the EMMA-top1. As the query increases, the attacker model obtained from the victim model suffers more accuracy drops, revealing that our defense induces more significant detrimental impacts when the attacker uses more queries." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 630, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 630, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 630, + 288, + 713 + ], + "type": "text", + "content": "Transferability across attack scenarios. To explore the transition matrix's transferable defense capabilities, we test it against an attack type different from the one it was optimized for. As shown in Figure 6, we use the matrix optimized under DA attack to defend against KL attack, denoted as \"DA " + }, + { + "bbox": [ + 46, + 630, + 288, + 713 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 46, + 630, + 288, + 713 + ], + "type": "text", + "content": " KL\". Compared to a matrix specifically optimized for KL attack, denoted as \"KL " + }, + { + "bbox": [ + 46, + 630, + 288, + 713 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 46, + 630, + 288, + 713 + ], + "type": "text", + "content": " KL\", it still maintained" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 310, + 72, + 424, + 146 + ], + "blocks": [ + { + "bbox": [ + 310, + 72, + 424, + 146 + ], + "lines": [ + { + "bbox": [ + 310, + 72, + 424, + 146 + ], + "spans": [ + { + "bbox": [ + 310, + 72, + 424, + 146 + ], + "type": "image", + "image_path": "ed201633e1a9d3cebbb1470398daf17ce8a00e5ea90ce7abce80b21605326866.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 426, + 72, + 541, + 146 + ], + "blocks": [ + { + "bbox": [ + 426, + 72, + 541, + 146 + ], + "lines": [ + { + "bbox": [ + 426, + 72, + 541, + 146 + ], + "spans": [ + { + "bbox": [ + 426, + 72, + 541, + 146 + ], + "type": "image", + "image_path": "9028e23d26bd00004128021321af4a80f66f06940570ffcbe45544bf452a7585.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 309, + 149, + 424, + 222 + ], + "blocks": [ + { + "bbox": [ + 309, + 149, + 424, + 222 + ], + "lines": [ + { + "bbox": [ + 309, + 149, + 424, + 222 + ], + "spans": [ + { + "bbox": [ + 309, + 149, + 424, + 222 + ], + "type": "image", + "image_path": "b3db8ce82356725ea57631298e973ca3391674ef2bbc731709eec883175269c5.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 335, + 224, + 516, + 236 + ], + "lines": [ + { + "bbox": [ + 335, + 224, + 516, + 236 + ], + "spans": [ + { + "bbox": [ + 335, + 224, + 516, + 236 + ], + "type": "text", + "content": "No defense EMMA EMMA-top1" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 347, + 243, + 504, + 255 + ], + "lines": [ + { + "bbox": [ + 347, + 243, + 504, + 255 + ], + "spans": [ + { + "bbox": [ + 347, + 243, + 504, + 255 + ], + "type": "text", + "content": "Figure 5. Effects of the number of queries." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 426, + 149, + 541, + 222 + ], + "blocks": [ + { + "bbox": [ + 426, + 149, + 541, + 222 + ], + "lines": [ + { + "bbox": [ + 426, + 149, + 541, + 222 + ], + "spans": [ + { + "bbox": [ + 426, + 149, + 541, + 222 + ], + "type": "image", + "image_path": "68222c69d426003e06e4838b38ac9bc999b9b7b32a9a33692a19320a5d9c0b34.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 263, + 545, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 263, + 545, + 300 + ], + "spans": [ + { + "bbox": [ + 304, + 263, + 545, + 300 + ], + "type": "text", + "content": "considerable defensive efficacy, exhibiting the optimized transition matrix's robust transferable defense capabilities across different attack types." + } + ] + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 318, + 308, + 426, + 415 + ], + "blocks": [ + { + "bbox": [ + 318, + 308, + 426, + 415 + ], + "lines": [ + { + "bbox": [ + 318, + 308, + 426, + 415 + ], + "spans": [ + { + "bbox": [ + 318, + 308, + 426, + 415 + ], + "type": "image", + "image_path": "293ba659bc74a9176cb97b1b15b3f5c9abe3d9f56361f9a94d345cd9f17369d0.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 336, + 420, + 514, + 431 + ], + "lines": [ + { + "bbox": [ + 336, + 420, + 514, + 431 + ], + "spans": [ + { + "bbox": [ + 336, + 420, + 514, + 431 + ], + "type": "text", + "content": "Figure 6. Transferability across attack scenarios." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 430, + 308, + 531, + 415 + ], + "blocks": [ + { + "bbox": [ + 430, + 308, + 531, + 415 + ], + "lines": [ + { + "bbox": [ + 430, + 308, + 531, + 415 + ], + "spans": [ + { + "bbox": [ + 430, + 308, + 531, + 415 + ], + "type": "image", + "image_path": "7709f982bd19a6cdf67251b1665ad488ccef7430bbad1b683ea188d84c45cc6c.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 439, + 545, + 463 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 439, + 545, + 463 + ], + "spans": [ + { + "bbox": [ + 306, + 439, + 545, + 463 + ], + "type": "text", + "content": "Due to the space limitation, additional experiments of our method are provided in Appendix." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 306, + 474, + 378, + 487 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 474, + 378, + 487 + ], + "spans": [ + { + "bbox": [ + 306, + 474, + 378, + 487 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 494, + 547, + 638 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 494, + 547, + 638 + ], + "spans": [ + { + "bbox": [ + 304, + 494, + 547, + 638 + ], + "type": "text", + "content": "In this work, we revisited the noise transition matrix in a novel way by applying it to model stealing defense for the first time. This deteriorates the attacker's cloning process to a level equivalent to training on noisy labels, hampering attacker model generalization. To optimize the transition matrix, we proposed a novel bi-level optimization training framework that incorporates fidelity training for the victim model and adversarial training for the surrogate model. Experimental results under various attack scenarios showed the effectiveness and robustness of our proposed defense method compared to current state-of-the-art methods, achieving strong defense with minimal utility trade-off." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 306, + 649, + 415, + 662 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 649, + 415, + 662 + ], + "spans": [ + { + "bbox": [ + 306, + 649, + 415, + 662 + ], + "type": "text", + "content": "7. Acknowledgments" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 669, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 669, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 669, + 547, + 715 + ], + "type": "text", + "content": "The authors wish to thank the anonymous reviewers for their helpful comments and suggestions. This work was supported by the National Science Foundation of China (62225602), and the Big Data Computing Center of Southeast University." + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "24312" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 94, + 107, + 106 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 94, + 107, + 106 + ], + "spans": [ + { + "bbox": [ + 48, + 94, + 107, + 106 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 114, + 288, + 714 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 53, + 114, + 287, + 158 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 114, + 287, + 158 + ], + "spans": [ + { + "bbox": [ + 53, + 114, + 287, + 158 + ], + "type": "text", + "content": "[1] Yossi Adi, Carsten Baum, Moustapha Cisse, Benny Pinkas, and Joseph Keshet. Turning your weakness into a strength: Watermarking deep neural networks by backdoorsing. In USENIX Security Symposium, pages 1615-1631, 2018. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 159, + 288, + 224 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 159, + 288, + 224 + ], + "spans": [ + { + "bbox": [ + 53, + 159, + 288, + 224 + ], + "type": "text", + "content": "[2] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. In International Conference on Advanced Neural Information Processing Systems, pages 1877-1901, 2020. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 225, + 288, + 269 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 225, + 288, + 269 + ], + "spans": [ + { + "bbox": [ + 53, + 225, + 288, + 269 + ], + "type": "text", + "content": "[3] Yanjiao Chen, Rui Guan, Xueluan Gong, Jianshuo Dong, and Meng Xue. D-dae: Defense-penetrating model extraction attacks. In IEEE Symposium on Security and Privacy, pages 382-399, 2023. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 270, + 288, + 313 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 270, + 288, + 313 + ], + "spans": [ + { + "bbox": [ + 53, + 270, + 288, + 313 + ], + "type": "text", + "content": "[4] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In IEEE/CVF Computer Vision and Pattern Recognition Conference, pages 248-255, 2009. 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 314, + 288, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 314, + 288, + 358 + ], + "spans": [ + { + "bbox": [ + 53, + 314, + 288, + 358 + ], + "type": "text", + "content": "[5] Adam Dziedzic, Muhammad Ahmad Kaleem, Yu Shen Lu, and Nicolas Papernot. Increasing the cost of model extraction with calibrated proof of work. In International Conference on Learning Representations, 2022. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 358, + 288, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 358, + 288, + 403 + ], + "spans": [ + { + "bbox": [ + 53, + 358, + 288, + 403 + ], + "type": "text", + "content": "[6] Matt Fredrikson, Somesh Jha, and Thomas Ristenpart. Model inversion attacks that exploit confidence information and basic countermeasures. In ACM SIGSAC conference on computer and communications security, pages 1322-1333, 2015. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 403, + 288, + 458 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 403, + 288, + 458 + ], + "spans": [ + { + "bbox": [ + 53, + 403, + 288, + 458 + ], + "type": "text", + "content": "[7] Karan Ganju, Qi Wang, Wei Yang, Carl A Gunter, and Nikita Borisov. Property inference attacks on fully connected neural networks using permutation invariant representations. In ACM SIGSAC conference on computer and communications security, pages 619-633, 2018. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 458, + 288, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 458, + 288, + 480 + ], + "spans": [ + { + "bbox": [ + 53, + 458, + 288, + 480 + ], + "type": "text", + "content": "[8] Gregory Griffin, Alex Holub, and Pietro Perona. Caltech-256 object category dataset. 2007. 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 481, + 288, + 534 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 481, + 288, + 534 + ], + "spans": [ + { + "bbox": [ + 53, + 481, + 288, + 534 + ], + "type": "text", + "content": "[9] Jun Guo, Aishan Liu, Xingyu Zheng, Siyuan Liang, Yisong Xiao, Yichao Wu, and Xianglong Liu. Isolation and induction: Training robust deep neural networks against model stealing attacks. In ACM International Conference on Multimedia, 2023. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 536, + 288, + 581 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 536, + 288, + 581 + ], + "spans": [ + { + "bbox": [ + 48, + 536, + 288, + 581 + ], + "type": "text", + "content": "[10] Bo Han, Quanming Yao, Tongliang Liu, Gang Niu, Ivor W Tsang, James T Kwok, and Masashi Sugiyama. A survey of label-noise representation learning: Past, present and future. arXiv preprint arXiv:2011.04406, 2020. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 581, + 288, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 581, + 288, + 635 + ], + "spans": [ + { + "bbox": [ + 48, + 581, + 288, + 635 + ], + "type": "text", + "content": "[11] Bo Han, Quanming Yao, Xingrui Yu, Gang Niu, Miao Xu, Weihua Hu, Ivor W. Tsang, and Masashi Sugiyama. Robust training of deep neural networks with extremely noisy labels. In International Conference on Advanced Neural Information Processing Systems, page 4, 2020. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 636, + 288, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 636, + 288, + 669 + ], + "spans": [ + { + "bbox": [ + 48, + 636, + 288, + 669 + ], + "type": "text", + "content": "[12] Geoffrey Hinton, Oriol Vinyals, and Jeff Dean. Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531, 2015. 2, 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 670, + 288, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 670, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 48, + 670, + 288, + 714 + ], + "type": "text", + "content": "[13] Surgan Jandial, Yash Khasbage, Arghya Pal, Vineeth N Balasubramanian, and Balaji Krishnamurthy. Distilling the undistillable: Learning from a nasty teacher. In European Conference on Computer Vision, pages 587-603, 2022. 2" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 714 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 307, + 73, + 546, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 546, + 106 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 546, + 106 + ], + "type": "text", + "content": "[14] Robert G Jeroslow. The polynomial hierarchy and a simple model for competitive analysis. Mathematical programming, pages 146-164, 1985. 5" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 107, + 547, + 152 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 107, + 547, + 152 + ], + "spans": [ + { + "bbox": [ + 307, + 107, + 547, + 152 + ], + "type": "text", + "content": "[15] Hengrui Jia, Christopher A Choquette-Choo, Varun Chandrasekaran, and Nicolas Papernot. Entangled watermarks as a defense against model extraction. In USENIX Security Symposium, pages 1937-1954, 2021. 2, 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 152, + 547, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 152, + 547, + 207 + ], + "spans": [ + { + "bbox": [ + 307, + 152, + 547, + 207 + ], + "type": "text", + "content": "[16] Hengrui Jia, Mohammad Yaghini, Christopher A Choquette-Choo, Natalie Dullerud, Anvith Thudi, Varun Chandrasekaran, and Nicolas Papernot. Proof-of-learning: Definitions and practice. In IEEE Symposium on Security and Privacy, pages 1039-1056, 2021. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 207, + 546, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 207, + 546, + 252 + ], + "spans": [ + { + "bbox": [ + 307, + 207, + 546, + 252 + ], + "type": "text", + "content": "[17] Wenbo Jiang, Hongwei Li, Guowen Xu, Tianwei Zhang, and Rongxing Lu. A comprehensive defense framework against model extraction attacks. IEEE Transactions on Dependable and Secure Computing, 2023. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 253, + 546, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 253, + 546, + 297 + ], + "spans": [ + { + "bbox": [ + 307, + 253, + 546, + 297 + ], + "type": "text", + "content": "[18] Ishan Jindal, Matthew Nokleby, and Xuewen Chen. Learning deep networks from noisy labels with dropout regularization. In IEEE International Conference on Data Mining, pages 967-972, 2016. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 297, + 546, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 297, + 546, + 342 + ], + "spans": [ + { + "bbox": [ + 307, + 297, + 546, + 342 + ], + "type": "text", + "content": "[19] Mika Juuti, Sebastian Szyller, Samuel Marchal, and N Asokan. Prada: protecting against dnn model stealing attacks. In IEEE European Symposium on Security and Privacy, pages 512-527, 2019. 1, 2, 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 342, + 547, + 388 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 342, + 547, + 388 + ], + "spans": [ + { + "bbox": [ + 307, + 342, + 547, + 388 + ], + "type": "text", + "content": "[20] Sanjay Kariyappa and Moinuddin K Qureshi. Defending against model stealing attacks with adaptive misinformation. In IEEE/CVF Computer Vision and Pattern Recognition Conference, pages 770-778, 2020. 1, 2, 3, 5, 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 388, + 546, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 388, + 546, + 431 + ], + "spans": [ + { + "bbox": [ + 307, + 388, + 546, + 431 + ], + "type": "text", + "content": "[21] Sanjay Kariyappa, Atul Prakash, and Moinuddin K Qureshi. Protecting dnns from theft using an ensemble of diverse models. In International Conference on Learning Representations, 2020. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 432, + 546, + 466 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 432, + 546, + 466 + ], + "spans": [ + { + "bbox": [ + 307, + 432, + 546, + 466 + ], + "type": "text", + "content": "[22] Sanjay Kariyappa, Atul Prakash, and Moinuddin K Qureshi. Maze: Data-free model stealing attack using zeroth-order gradient estimation. In CVPR, 2021. 1, 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 466, + 546, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 466, + 546, + 510 + ], + "spans": [ + { + "bbox": [ + 307, + 466, + 546, + 510 + ], + "type": "text", + "content": "[23] Manish Kesarwani, Bhaskar Mukholy, Vijay Arya, and Sameep Mehta. Model extraction warning in mlaas paradigm. In Annual Computer Security Applications Conference, pages 371-380, 2018. 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 511, + 546, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 511, + 546, + 566 + ], + "spans": [ + { + "bbox": [ + 307, + 511, + 546, + 566 + ], + "type": "text", + "content": "[24] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In International Conference on Computer Vision, 2023. 1" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 567, + 546, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 567, + 546, + 612 + ], + "spans": [ + { + "bbox": [ + 307, + 567, + 546, + 612 + ], + "type": "text", + "content": "[25] Kalpesh Krishna, Gaurav Singh Tomar, Ankur P Parikh, Nicolas Papernot, and Mohit Iyyer. Thieves on sesame street! model extraction of bert-based apis. In International Conference on Learning Representations, 2020. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 612, + 546, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 612, + 546, + 635 + ], + "spans": [ + { + "bbox": [ + 307, + 612, + 546, + 635 + ], + "type": "text", + "content": "[26] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009. 6" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 635, + 546, + 680 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 635, + 546, + 680 + ], + "spans": [ + { + "bbox": [ + 307, + 635, + 546, + 680 + ], + "type": "text", + "content": "[27] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. ImageNet classification with deep convolutional neural networks. In International Conference on Advanced Neural Information Processing Systems, pages 1106-1114, 2012. 1" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 680, + 546, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 680, + 546, + 714 + ], + "spans": [ + { + "bbox": [ + 307, + 680, + 546, + 714 + ], + "type": "text", + "content": "[28] Souvik Kundu, Qirui Sun, Yao Fu, Massoud Pedram, and Peter Beerel. Analyzing the confidentiality of undistinguishable teachers in knowledge distillation. pages 9181-9192, 2021. 2" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "24313" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 288, + 127 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 288, + 127 + ], + "type": "text", + "content": "[29] Jeonghyun Lee, Sungmin Han, and Sangkyun Lee. Model stealing defense against exploiting information leak through the interpretation of deep neural nets. In International Joint Conference on Artificial Intelligence, pages 710-716, 2022. 1, 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 129, + 288, + 174 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 129, + 288, + 174 + ], + "spans": [ + { + "bbox": [ + 48, + 129, + 288, + 174 + ], + "type": "text", + "content": "[30] Taesung Lee, Benjamin Edwards, Ian Molloy, and Dong Su. Defending against neural network model stealing attacks using deceptive perturbations. In IEEE Security and Privacy Workshops, pages 43-49, 2019. 2, 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 175, + 288, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 175, + 288, + 239 + ], + "spans": [ + { + "bbox": [ + 48, + 175, + 288, + 239 + ], + "type": "text", + "content": "[31] Simin Li, Shuning Zhang, Gujun Chen, Dong Wang, Pu Feng, Jiakai Wang, Aishan Liu, Xin Yi, and Xianglong Liu. Towards benchmarking and assessing visual naturalness of physical world adversarial attacks. In IEEE/CVF Computer Vision and Pattern Recognition Conference, pages 12324–12333, 2023. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 241, + 288, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 241, + 288, + 285 + ], + "spans": [ + { + "bbox": [ + 48, + 241, + 288, + 285 + ], + "type": "text", + "content": "[32] Yiming Li, Linghui Zhu, Xiaojun Jia, Yong Jiang, Shu-Tao Xia, and Xiaochun Cao. Defending against model stealing via verifying embedded external features. In Association for the Advance of Artificial Intelligence, pages 1464-1472, 2022. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 286, + 288, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 286, + 288, + 341 + ], + "spans": [ + { + "bbox": [ + 48, + 286, + 288, + 341 + ], + "type": "text", + "content": "[33] Shunchang Liu, Jiakai Wang, Aishan Liu, Yingwei Li, Yijie Gao, Xianglong Liu, and Dacheng Tao. Harnessing perceptual adversarial patches for crowd counting. In ACM SIGSAC Conference on Computer and Communications Security, pages 2055-2069, 2022. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 342, + 288, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 342, + 288, + 376 + ], + "spans": [ + { + "bbox": [ + 48, + 342, + 288, + 376 + ], + "type": "text", + "content": "[34] Haoyu Ma, Tianlong Chen, Ting-Kuei Hu, Chenyu You, Xiaohui Xie, and Zhangyang Wang. Undistillable: Making a nasty teacher that cannot teach students. In ICLR, 2021. 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 376, + 288, + 421 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 376, + 288, + 421 + ], + "spans": [ + { + "bbox": [ + 48, + 376, + 288, + 421 + ], + "type": "text", + "content": "[35] Ke Ma, Qianqian Xu, Jinshan Zeng, Xiaochun Cao, and Qingming Huang. Poisoning attack against estimating from pairwise comparisons. IEEE Trans. Pattern Anal. Mach. Intell., pages 6393-6408, 2021. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 422, + 288, + 476 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 422, + 288, + 476 + ], + "spans": [ + { + "bbox": [ + 48, + 422, + 288, + 476 + ], + "type": "text", + "content": "[36] Ke Ma, Qianqian Xu, Jinshan Zeng, Guorong Li, Xiaochun Cao, and Qingming Huang. A tale of hodgerank and spectral method: Target attack against rank aggregation is the fixed point of adversarial game. IEEE Trans. Pattern Anal. Mach. Intell., pages 4090-4108, 2022. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 478, + 288, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 478, + 288, + 521 + ], + "spans": [ + { + "bbox": [ + 48, + 478, + 288, + 521 + ], + "type": "text", + "content": "[37] Pratyush Maini, Mohammad Yaghini, and Nicolas Papernot. Dataset inference: Ownership resolution in machine learning. In International Conference on Learning Representations, 2021. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 522, + 288, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 522, + 288, + 567 + ], + "spans": [ + { + "bbox": [ + 48, + 522, + 288, + 567 + ], + "type": "text", + "content": "[38] Mantas Mazeika, Bo Li, and David Forsyth. How to steer your adversary: Targeted and efficient model stealing defenses with gradient redirection. In International Conference on Machine Learning, pages 15241-15254, 2022. 2, 3, 6, 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 568, + 288, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 568, + 288, + 612 + ], + "spans": [ + { + "bbox": [ + 48, + 568, + 288, + 612 + ], + "type": "text", + "content": "[39] Tribhuvanesh Orekondy, Bernt Schiele, and Mario Fritz. Knockoff nets: Stealing functionality of black-box models. In IEEE/CVF Computer Vision and Pattern Recognition Conference, pages 4954-4963, 2019. 1, 2, 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 613, + 288, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 613, + 288, + 657 + ], + "spans": [ + { + "bbox": [ + 48, + 613, + 288, + 657 + ], + "type": "text", + "content": "[40] Tribhuvanesh Orekondy, Bernt Schiele, and Mario Fritz. Prediction poisoning: Towards defenses against dnn model stealing attacks. In International Conference on Learning Representations, 2020. 2, 3, 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 658, + 288, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 288, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 288, + 712 + ], + "type": "text", + "content": "[41] Soham Pal, Yash Gupta, Aditya Shukla, Aditya Kanade, Shirish Shevade, and Vinod Ganapathy. Activethief: Model extraction using active learning and unannotated public data. In Association for the Advance of Artificial Intelligence, pages 865–872, 2020. 2" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 712 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 307, + 73, + 547, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 547, + 127 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 547, + 127 + ], + "type": "text", + "content": "[42] Nicolas Papernot, Patrick McDaniel, Ian Goodfellow, Somesh Jha, Z Berkay Celik, and Ananthram Swami. Practical black-box attacks against machine learning. In ACM on Asia conference on computer and communications security, pages 506-519, 2017. 1, 2, 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 129, + 547, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 129, + 547, + 185 + ], + "spans": [ + { + "bbox": [ + 307, + 129, + 547, + 185 + ], + "type": "text", + "content": "[43] Giorgio Patrini, Alessandro Rozza, Aditya Krishna Menon, Richard Nock, and Lizhen Qu. Making deep neural networks robust to label noise: A loss correction approach. In IEEE/CVF Computer Vision and Pattern Recognition Conference, pages 1944-1952, 2017. 2, 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 186, + 547, + 219 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 186, + 547, + 219 + ], + "spans": [ + { + "bbox": [ + 307, + 186, + 547, + 219 + ], + "type": "text", + "content": "[44] Sunandini Sanyal, Sravanti Addepalli, and R Venkatesh Babu. Towards data-free model stealing in a hard label setting. In CVPR, 2022. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 220, + 545, + 243 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 220, + 545, + 243 + ], + "spans": [ + { + "bbox": [ + 307, + 220, + 545, + 243 + ], + "type": "text", + "content": "[45] Zeyang Sha et al. Can't steal? cont-steal! contrastive stealing attacks against image encoders. In CVPR, 2023. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 245, + 545, + 278 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 245, + 545, + 278 + ], + "spans": [ + { + "bbox": [ + 307, + 245, + 545, + 278 + ], + "type": "text", + "content": "[46] Yu Shi, Dong-Dong Wu, Xin Geng, and Min-Ling Zhang. Robust representation learning for unreliable partial label learning. arXiv preprint arXiv:2308.16718, 2023. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 279, + 545, + 324 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 279, + 545, + 324 + ], + "spans": [ + { + "bbox": [ + 307, + 279, + 545, + 324 + ], + "type": "text", + "content": "[47] Reza Shokri, Marco Stronati, Congzheng Song, and Vitaly Shmatikov. Membership inference attacks against machine learning models. In IEEE symposium on security and privacy, pages 3-18, 2017. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 325, + 545, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 325, + 545, + 358 + ], + "spans": [ + { + "bbox": [ + 307, + 325, + 545, + 358 + ], + "type": "text", + "content": "[48] Sainbayar Sukhbaatar, Joan Bruna, Manohar Paluri, Lubomir Bourdev, and Rob Fergus. Training convolutional networks with noisy labels. arXiv preprint arXiv:1406.2080, 2014. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 360, + 545, + 404 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 360, + 545, + 404 + ], + "spans": [ + { + "bbox": [ + 307, + 360, + 545, + 404 + ], + "type": "text", + "content": "[49] Christian Szegedy, Wojciech Zaremba, Ilya Sutskever, Joan Bruna, Dumitru Erhan, Ian Goodfellow, and Rob Fergus. Intriguing properties of neural networks. In International Conference on Learning Representations, 2013. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 406, + 545, + 450 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 406, + 545, + 450 + ], + "spans": [ + { + "bbox": [ + 307, + 406, + 545, + 450 + ], + "type": "text", + "content": "[50] Sebastian Szyller, Buse Gul Atli, Samuel Marchal, and N Asokan. Dawn: Dynamic adversarial watermarking of neural networks. In ACM International Conference on Multimedia, pages 4417-4425, 2021. 1, 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 452, + 545, + 495 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 452, + 545, + 495 + ], + "spans": [ + { + "bbox": [ + 307, + 452, + 545, + 495 + ], + "type": "text", + "content": "[51] Florian Tramér, Fan Zhang, Ari Juels, Michael K Reiter, and Thomas Ristenpart. Stealing machine learning models via prediction {APIs}. In USENIX security symposium, pages 601-618, 2016. 1, 2, 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 497, + 545, + 552 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 497, + 545, + 552 + ], + "spans": [ + { + "bbox": [ + 307, + 497, + 545, + 552 + ], + "type": "text", + "content": "[52] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In International Conference on Advanced Neural Information Processing Systems, 2017. 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 554, + 545, + 576 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 554, + 545, + 576 + ], + "spans": [ + { + "bbox": [ + 307, + 554, + 545, + 576 + ], + "type": "text", + "content": "[53] M Mitchell Waldrop et al. No drivers required. Nature, page 20, 2015. 1" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 578, + 545, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 578, + 545, + 611 + ], + "spans": [ + { + "bbox": [ + 307, + 578, + 545, + 611 + ], + "type": "text", + "content": "[54] Binghui Wang and Neil Zhenqiang Gong. Stealing hyperparameters in machine learning. In IEEE symposium on security and privacy, pages 36-52, 2018. 1" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 613, + 545, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 613, + 545, + 667 + ], + "spans": [ + { + "bbox": [ + 307, + 613, + 545, + 667 + ], + "type": "text", + "content": "[55] Jiakai Wang, Aishan Liu, Zixin Yin, Shunchang Liu, Shiyu Tang, and Xianglong Liu. Dual attention suppression attack: Generate adversarial camouflage in physical world. In IEEE/CVF Computer Vision and Pattern Recognition Conference, pages 8565-8574, 2021. 1" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 669, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 669, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 307, + 669, + 545, + 712 + ], + "type": "text", + "content": "[56] Zhen Wang, Guosheng Hu, and Qinghua Hu. Training noise-robust deep neural networks via meta-learning. In IEEE/CVF Computer Vision and Pattern Recognition Conference, pages 4524–4533, 2020. 3" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "24314" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 487 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 288, + 105 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 288, + 105 + ], + "type": "text", + "content": "[57] Peter Welinder, Steve Branson, Takeshi Mita, Catherine Wah, Florian Schroff, Serge Belongie, and Pietro Perona. Caltech-ucsd birds 200. 2010. 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 107, + 288, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 107, + 288, + 150 + ], + "spans": [ + { + "bbox": [ + 49, + 107, + 288, + 150 + ], + "type": "text", + "content": "[58] Hongwei Wen, Jingyi Cui, Hanyuan Hang, Jiabin Liu, Yisen Wang, and Zhouchen Lin. Leveraged weighted loss for partial label learning. In International Conference on Machine Learning, pages 11091-11100, 2021. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 152, + 288, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 152, + 288, + 195 + ], + "spans": [ + { + "bbox": [ + 49, + 152, + 288, + 195 + ], + "type": "text", + "content": "[59] Dong-Dong Wu, Deng-Bao Wang, and Min-Ling Zhang. Revisiting consistency regularization for deep partial label learning. In International Conference on Machine Learning, pages 24212-24225, 2022. 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 197, + 288, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 197, + 288, + 251 + ], + "spans": [ + { + "bbox": [ + 49, + 197, + 288, + 251 + ], + "type": "text", + "content": "[60] Xiaobo Xia, Tongliang Liu, Nannan Wang, Bo Han, Chen Gong, Gang Niu, and Masashi Sugiyama. Are anchor points really indispensable in label-noise learning? In International Conference on Advanced Neural Information Processing Systems, pages 6835-6846, 2019. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 49, + 253, + 288, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 253, + 288, + 297 + ], + "spans": [ + { + "bbox": [ + 49, + 253, + 288, + 297 + ], + "type": "text", + "content": "[61] Hao Yang, You-Zhi Jin, Zi-Yin Li, Deng-Bao Wang, Xin Geng, and Min-Ling Zhang. Learning from noisy labels via dynamic loss thresholding. IEEE Transactions on Knowledge and Data Engineering, 2023. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 49, + 297, + 288, + 351 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 297, + 288, + 351 + ], + "spans": [ + { + "bbox": [ + 49, + 297, + 288, + 351 + ], + "type": "text", + "content": "[62] Honggang Yu, Kaichen Yang, Teng Zhang, Yun-Yun Tsai, Tsung-Yi Ho, and Yier Jin. Cloudleak: Large-scale deep learning models stealing through adversarial examples. In Network and Distributed System Security Symposium, 2020. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 49, + 353, + 288, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 353, + 288, + 397 + ], + "spans": [ + { + "bbox": [ + 49, + 353, + 288, + 397 + ], + "type": "text", + "content": "[63] Chiyuan Zhang, Samy Bengio, Moritz Hardt, Benjamin Recht, and Oriol Vinyals. Understanding deep learning (still) requires rethinking generalization. Communications of the ACM, pages 107-115, 2021. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 49, + 399, + 288, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 399, + 288, + 442 + ], + "spans": [ + { + "bbox": [ + 49, + 399, + 288, + 442 + ], + "type": "text", + "content": "[64] Jiliang Zhang, Shuang Peng, Yansong Gao, Zhi Zhang, and Qinghui Hong. Apmsa: adversarial perturbation against model stealing attacks. IEEE Transactions on Information Forensics and Security, pages 1667-1679, 2023. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 49, + 443, + 288, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 443, + 288, + 487 + ], + "spans": [ + { + "bbox": [ + 49, + 443, + 288, + 487 + ], + "type": "text", + "content": "[65] Guoqing Zheng, Ahmed Hassan Awadallah, and Susan Dumas. Meta label correction for noisy label learning. In Association for the Advance of Artificial Intelligence, pages 11053-11061, 2021. 3" + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "24315" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Efficient Multi-scale Network with Learnable Discrete Wavelet Transform for Blind Motion Deblurring/4af0ea1e-f85d-4e0b-ab29-04d268e0386c_content_list.json b/2024/Efficient Multi-scale Network with Learnable Discrete Wavelet Transform for Blind Motion Deblurring/4af0ea1e-f85d-4e0b-ab29-04d268e0386c_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..cef15837a0eca4e3e59df84aeda82df4bac83f11 --- /dev/null +++ b/2024/Efficient Multi-scale Network with Learnable Discrete Wavelet Transform for Blind Motion Deblurring/4af0ea1e-f85d-4e0b-ab29-04d268e0386c_content_list.json @@ -0,0 +1,2026 @@ +[ + { + "type": "text", + "text": "Efficient Multi-scale Network with Learnable Discrete Wavelet Transform for Blind Motion Deblurring", + "text_level": 1, + "bbox": [ + 112, + 128, + 856, + 175 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Xin Gao\\*1,2 Tianheng Qiu\\*2,3,4 Xinyu Zhang†2 Hanlin Bai\\*1 Kang Liu\\*1 Xuan Huang†4 Hu Wei\\*4 Guoying Zhang†1 Huaping Liu\\*2 *Equal Contribution † Corresponding Author", + "bbox": [ + 84, + 181, + 901, + 236 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ China University of Mining & Technology-Beijing $^{2}$ Tsinghua University $^{3}$ University of Science and Technology of China $^{4}$ Hefei Institutes of Physical Science, Chinese Academy of Sciences", + "bbox": [ + 196, + 241, + 797, + 297 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 316, + 313, + 333 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Coarse-to-fine schemes are widely used in traditional single-image motion deblur; however, in the context of deep learning, existing multi-scale algorithms not only require the use of complex modules for feature fusion of low-scale RGB images and deep semantics, but also manually generate low-resolution pairs of images that do not have sufficient confidence. In this work, we propose a multi-scale network based on single-input and multiple-outputs(SIMO) for motion deblurring. This simplifies the complexity of algorithms based on a coarse-to-fine scheme. To alleviate restoration defects impacting detail information brought about by using a multi-scale architecture, we combine the characteristics of real-world blurring trajectories with a learnable wavelet transform module to focus on the directional continuity and frequency features of the step-by-step transitions between blurred images to sharp images. In conclusion, we propose a multi-scale network with a learnable discrete wavelet transform (MLWNet), which exhibits state-of-the-art performance on multiple real-world deblurred datasets, in terms of both subjective and objective quality as well as computational efficiency. Our code is available on https://github.com/thqiu0419/MLWNet.", + "bbox": [ + 75, + 348, + 473, + 681 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 709, + 209, + 724 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Most current top-performing single-image blind deblurring algorithms are based on DNNs, which can be structurally categorized into single-scale [2, 10, 12, 28, 29, 42] and multi-scale approaches [3, 9, 13, 27, 37]. Compared to single-scale methods, multi-scale methods are built on the idea of moving from coarse-to-fine, and thus they decompose the challenging single-image blind deblurring problem into easier-to-solve sub-problems to restore the blurred image step-by-step.", + "bbox": [ + 75, + 733, + 468, + 869 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Earlier DNNs exploiting the coarse-to-fine concept usually employ estimators of different scales to produce re", + "bbox": [ + 76, + 869, + 468, + 900 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/00c418dd730d494de65b5951a619cb9246b6202b63a1494cf6284118d1f7e695.jpg", + "image_caption": [ + "Figure 1. Performance comparison on the RealBlur-J [23] test dataset in terms of PSNR and GMACs. Our proposed MLWNet achieves superiority in comparison with other state-of-the-arts." + ], + "image_footnote": [], + "bbox": [ + 498, + 314, + 893, + 535 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "stored images gradually [20, 21, 27]. However, not only is the extracted semantic information not sufficiently representative, but the use of the same complexity at different scales may lead to redundant computations; More advanced multi-scale algorithms [3] use a global encoder-decoder, which significantly reduces the algorithm's runtime, multiple upsampling and downsampling leads to insufficient restoration of detailed information, and therefore such algorithms require the introduction of additional fusion modules to encode input images for fusion with deeper semantics. In addition, existing multi-scale algorithms use multiple-input and multiple-output (MIMO) architectures, which require the introduction of manually constructed downsampled image pairs, and usually employ simple interpolation algorithms to generate low-resolution images for the sake of efficiency, which is obviously not reliable.", + "bbox": [ + 496, + 583, + 892, + 824 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We note that the gradual insertion of images used by existing coarse-to-fine algorithms is redundant. Referring to numerous algorithms [6, 8, 15, 17, 26, 31] that are widely used in downstream tasks, the network starts with a single input and also extracts features at different scales. These", + "bbox": [ + 496, + 825, + 893, + 898 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 44 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "2733", + "bbox": [ + 480, + 924, + 514, + 936 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "features can be aggregated by simple feature fusion to ultimately obtain competitive results. That is, a DNN itself possesses the ability to learn effective features at different scales. This inspired us to design a multi-scale architecture that retains an original image as input and sequentially restores images at each scale in output stage. This is not only more theoretically sound, but also eliminates the time required to generate pairs of images of different resolutions, as well as the huge complexity increase brought about by fusion modules between RGB domain and deep features.", + "bbox": [ + 75, + 90, + 472, + 242 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The coarse-to-fine algorithm has another inherent defect. During progressive restoration, the solution of the upper-level problem takes the solution of the lower-level as initialization [9]. Although this reduces the difficulty of solving the upper-level problem, due to the smaller resolution of lower-level spatial, the features transmitted upwards are semantically precise but spatially ambiguous, thus the restoration ability of multi-scale network in spatial details is limited. Hence, there is an urgent need to improve the quality of restoration of high-frequency detail part. A simple approach is to introduce a frequency domain transform as an alternative to a spatial domain transform. This would provide the algorithm with a choice and direct its attention to different frequencies. In this paper, we consider using the discrete wavelet transform for the following reasons:", + "bbox": [ + 75, + 244, + 472, + 470 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. Many recent state-of-the-art deblurring algorithms [10, 19, 42] have introduced the discrete Fourier transform (DFT) as a frequency prior, which provides information that helps the algorithm to identify and select high- and low-frequency components that need to be preserved during restoration. Compared to DFT, the discrete wavelet transform (DWT) is better suited to deal with images containing more abrupt signals [5].", + "2. Realistic blur and synthetic blur have significant distributional differences [24]. In the real world, due to the short exposure time of a camera, realistic blur has a specific directionality [29], i.e., the blur trajectory is regionally continuous; conversely, synthetic blur has an unnatural and discontinuous trajectory. To fully utilize the potential deblurring guidance brought by this trajectory continuity, we use 2D-DWT to reveal blur directionality, distinguish changes in the blur signal along different directions, and provide the algorithm with a reliable basis for deblurring through adaptive learning." + ], + "bbox": [ + 76, + 474, + 468, + 762 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To make 2D-DWT fit the data distribution and feature layer space more closely, we implemented 2D-DWT with an adaptive data distribution by using group convolution, transferring feature space from the spatial domain into the wavelet domain, generating sub-signals with different frequency features and different directional features, and then constructing wavelet losses for them in order to constrain them by self-supervision. Experimental results demonstrate that the proposed method has advanced performance in", + "bbox": [ + 75, + 763, + 470, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "terms of accuracy and efficiency (Fig. 1).", + "bbox": [ + 498, + 90, + 772, + 106 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In summary, our contributions are as follows:", + "bbox": [ + 517, + 107, + 818, + 119 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose a single-input and multiple-output(SIMO) multi-scale baseline for progressive image deblur, which reduces the complexity of existing multi-scale deblurring algorithms and improves the overall efficiency of image restoration networks.", + "- We construct a learnable discrete wavelet transform node(LWN) to provide reliable frequency and direction selection for the algorithm, which promotes the algorithm's restoration of the high-frequency components of edges and details.", + "- For the network to work better, reasonable multi-scale loss is proposed to guide pixel-by-pixel and scale-by-scale restoration. We also created reasonable self-supervised losses for the learnable wavelet transform to limit the learning direction of the wavelet kernel.", + "- We demonstrate the effectiveness of our algorithm under several motion blur datasets, especially real datasets, and obtain highly competitive results." + ], + "bbox": [ + 500, + 121, + 890, + 393 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 407, + 640, + 422 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Single-scale Deblurring Algorithm. Image deblurring have developed rapidly in recent years [2, 10, 11, 14, 39, 40]. To preserve richer image details, Kupyn et al. [11] viewed deblurring as a special case of image-to-image conversion, and for the first time utilized a GAN to recover images with richer details. To provide a high-quality and simple baseline, Chen et al. [2] proposed a nonlinear activation-free network that obtained excellent results. Zamir et al. [38] proposed a transformer with an encoder-decoder architecture that can be applied to higher-resolution images and utilizes cross-channel attention. Kong et al. [10] utilized a domain-based self-attention solver to reduce the transformer complexity and suppress artefacts. However, such single-scale algorithms estimate complex restoration problems directly and may require the design of restorers of higher complexity, which is suboptimal in terms of efficiency [9].", + "bbox": [ + 496, + 431, + 890, + 688 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Multi-scale Deblurring Algorithm. Multi-scale motion deblur methods aim to achieve progressive recovery using a multi-stage or multi-sub-network approach. The multiscale algorithm previously used in the field of image deblurring is based on MIMO [3, 13, 20, 21, 27]. Recently, Cho et al. [3] proposed MIMO-UNet, which employs a multi-output single decoder as well as a multi-input single encoder to simulate a multi-scale architecture consisting of stacked networks, thereby greatly simplifying complexity. Zamir et al. [37] designed a multi-stage progressive image recovery network to simplify the information flow between multiple sub-networks. As shown in Sec. 3.1, our proposed multiscale approach further simplifies the structure of multi-scale networks and better balances the accuracy and complexity.", + "bbox": [ + 496, + 688, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2734", + "bbox": [ + 480, + 924, + 514, + 936 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/b39c5e4e8a33272b3c101751994372a21ae473760c6f9291006c75259bca3ecf.jpg", + "image_caption": [ + "Figure 2. The overall architecture of the proposed MLWNet, the SEB is a simple module designed with reference [2], the WFB and WHB apply the LWN that implements the learnable 2D-DWT. In training phase, supervised learning is performed using $\\mathcal{L}_{multi}$ and self-supervised restraint of the wavelet kernel is performed using $\\mathcal{L}_{wavelet}$ . In testing phase, only the highest scale restored images is output." + ], + "image_footnote": [], + "bbox": [ + 78, + 88, + 885, + 358 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Application of Frequency in Deblurring. Frequency is an important property of images, in recent years frequency domain has been introduced into numerous DNNs to be solved for various tasks [10, 22, 25, 32, 35, 41]. Mao et al. [19] proposed DeepRFT, which introduced a residual module based on the Fourier transform into an advanced algorithm. Kong et al. [10] proposed FFTformer, which introduced the Fourier transform into a feed-forward network to effectively utilize different frequency information. Zou et al. [42] proposed SDWNet, which utilized frequency domain information as a complement to spatial domain information. Different from these methods, we introduce a learnable 2D-DWT capable of adapting to data distribution and feature space, more suitable not only for dealing with digital images rich in mutation signals, but also for dealing with real blurring.", + "bbox": [ + 75, + 412, + 472, + 640 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Proposed Method", + "text_level": 1, + "bbox": [ + 76, + 655, + 250, + 672 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our proposed MLWNet(Fig. 2) aims to explore an efficient multi-scale architectural approach to achieve high-quality blind deblurring of a single image. First, we designed a novel and highly scalable SIMO multi-scale baseline to address the performance bottleneck faced by partially multiscale networks. It takes a single image as an input and gradually generates a series of sharp images from the bottom up. Then, we propose a learnable wavelet transform node (LWN) for image deblurring, and it enhances the proposed algorithm's ability to restore detail information.", + "bbox": [ + 75, + 681, + 468, + 833 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Multi-scale Baseline", + "text_level": 1, + "bbox": [ + 76, + 845, + 267, + 859 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our multi-scale network maintains an encoder-decoder architecture, but retains the original image with the highest", + "bbox": [ + 75, + 869, + 470, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "resolution as input, and restores sharp images of various scales sequentially in the output stage. In this way, we eliminate the complexity of the fusion module when dealing RGB images of different scales and with deep semantics.", + "bbox": [ + 496, + 412, + 892, + 472 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "For ease of understanding, we will introduce here the Encoder phase, multi-scale semantic Fusion phase (hereafter referred to as Fusion phase), and Decoder phase sequentially. As shown in Fig. 2, the Encoder stage consists of several gray Simple Encoder Blocks(SEB), where information flows top-down for feature extraction; the Fusion stage consists of several green Wavelet Fusion Blocks(WFB), where information flows bottom-up for semantic fusion at different scales; and the Decoder stage consists of several purple Wavelet Head Blocks(WHB), and the bottom-up flow of information enables gradual restoration. Compared with WHB, SEB replaces LWN with a $1 \\times 1$ conv for channel scaling, and a $3 \\times 3$ depth-conv for feature extraction.", + "bbox": [ + 496, + 474, + 892, + 671 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The Encoder stage is responsible for full feature extraction, downsampling feature maps once after each block, and then passing the feature maps to the Fusion stage or the Decoder stage as needed, whose outputs $E_{out}^{i}$ to block of the $i - th$ layer can be denoted as:", + "bbox": [ + 496, + 672, + 890, + 748 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nE _ {o u t} ^ {i} = \\left\\{ \\begin{array}{l l} \\phi_ {i} (e m b e d (x)), & i = i _ {m a x} \\\\ \\phi_ {i} (E _ {o u t} ^ {i - 1}), & o t h e r w i s e \\end{array} \\right. \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 553, + 762, + 890, + 797 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\phi_{i}$ represent the block of corresponding layer of Encoder. The Fusion stage adapts and fuses information from different scales of semantics produced in Encoder to generate intermediate representations with deep semantics and shallow details. The Fusion stage's outputs $F_{out}^{i}$ to the block of $i - th$ layer is expressed as in Eq. 2, where $\\delta_{i}$", + "bbox": [ + 496, + 809, + 892, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "2735", + "bbox": [ + 480, + 925, + 514, + 936 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "represents the block of the corresponding layer of Fusion.", + "bbox": [ + 76, + 90, + 460, + 107 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nF _ {o u t} ^ {i} = \\left\\{ \\begin{array}{l l} \\delta_ {i} \\left(E _ {o u t} ^ {i} + E _ {o u t} ^ {i - 1}\\right), & i = i _ {\\text {m i n}} \\\\ \\delta_ {i} \\left(F _ {o u t} ^ {i - 1} + E _ {o u t} ^ {i}\\right), & \\text {o t h e r w i s e} \\end{array} \\right. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 122, + 114, + 468, + 150 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The Decoder stage utilizes the information passed by both the Encoder and Fusion stages to progressively upsample the feature maps and generating pre-output feature maps at each scale separately. Output $D_{out}^{i}$ to the block of the $i - th$ layer is denoted as in Eq. 3. Here $\\xi_{i}$ represents the block of the corresponding layer of the Decoder.", + "bbox": [ + 76, + 157, + 470, + 248 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nD _ {o u t} ^ {i} = \\left\\{ \\begin{array}{l l} \\xi_ {i} \\left(E _ {o u t} ^ {i}\\right), & i = i _ {m i n} \\\\ \\xi_ {i} \\left(D _ {o u t} ^ {i - 1} + F _ {o u t} ^ {i}\\right), & i > i _ {m i n} \\text {a n d} i < i _ {m a x} \\\\ \\xi_ {i} \\left(D _ {o u t} ^ {i - 1} + F _ {o u t} ^ {i - 1} + E _ {o u t} ^ {i}\\right), & i = i _ {m a x} \\end{array} \\right.\n$$\n", + "text_format": "latex", + "bbox": [ + 76, + 258, + 460, + 315 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Each block in the Decoder is followed by a $3 \\times 3$ convolution to generate a restored image at each scale. To improve inference efficiency, all but the highest-scale outputs are generated in the training phase to compute the multiscale loss. We note that this design also have a similar role with auxiliary head [36] in facilitating the model's learning of the recovered uniform patterns.", + "bbox": [ + 76, + 323, + 468, + 429 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/c17db24264ae3d97787ae53f7a4da5398aabfa3989e2231c56335c5c045ed4cb.jpg", + "image_caption": [ + "Figure 3. (a) The process of learnable 2D-wavelet convolution. (b) The construction process of the $N \\times N$ 2D-wavelet kernel." + ], + "image_footnote": [], + "bbox": [ + 86, + 439, + 460, + 705 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Learnable Discrete Wavelet Transform", + "text_level": 1, + "bbox": [ + 76, + 756, + 411, + 771 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To alleviate the defects of multi-scale architectures in the recovery of detail information, we design LWN and constructed the WFB and WHB based on it. Unlike the Fourier transform, the wavelet transform has adaptive time-frequency resolution, and performs excellently on digital images rich in mutation signal. Its use in conjunction with DNNs has gained increasing widespread attention [7, 16, 18, 34].", + "bbox": [ + 75, + 780, + 468, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For ease of understanding, the 1D-DWT case is given here first. For the input discrete signal $t$ , given the wavelet function $\\psi_{j,k}(t) = 2^{\\frac{j}{2}}\\psi (2^{j}t - k)$ for scaling factor $j$ , time factor $k$ , and scale function $\\phi_{j,k}(t) = 2^{\\frac{j}{2}}\\phi (2^{j}t - k)$ , the decomposition of the input signal $t$ at $j_0$ is given by:", + "bbox": [ + 496, + 90, + 892, + 172 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nf (t) = \\sum_ {j > j _ {0}} \\sum_ {k} d _ {j, k} \\psi_ {j, k} (t) + \\sum_ {k} c _ {j _ {0}, k} \\phi_ {j _ {0}, k} (t) \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 527, + 181, + 892, + 215 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this process, $d_{j,k} = \\langle f(t),\\psi_{j,k}(t)\\rangle$ represents the detail coefficients (i.e., high-frequency components), $c_{j0,k} = \\langle f(t),\\phi_{j0,k}(t)\\rangle$ represents the approximation coefficients (i.e., low-frequency components) so that the input signal can be disassembled step by step into a rich wavelet domain signal. In order to use wavelet transform in DNN, we introduce the analysis vectors $\\vec{a}_0[k] = \\left\\langle \\frac{1}{\\sqrt{2}}\\phi \\left(\\frac{t}{2}\\right),\\phi (t - k)\\right\\rangle ,$ $\\vec{a}_{1}[k] = \\left\\langle \\frac{1}{\\sqrt{2}}\\psi \\left(\\frac{t}{2}\\right),\\phi (t - k)\\right\\rangle$ to represent the high and low frequency filters respectively, then:", + "bbox": [ + 496, + 228, + 893, + 380 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nc _ {j + 1, p} = \\sum_ {k} \\vec {a} _ {0} [ k - 2 p ] c _ {j, k}\n$$\n", + "text_format": "latex", + "bbox": [ + 599, + 391, + 790, + 419 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nd _ {j + 1, p} = \\sum_ {k} ^ {\\kappa} \\vec {a} _ {1} [ k - 2 p ] c _ {j, k} \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 601, + 417, + 890, + 457 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "According to Eq. 4, the decomposition of the original signal over the wavelet basis can be viewed as a recursive convolution of the original signal with a specific filter using a step size of 2. Similarly, the inverse wavelet transform can be realized by transposed convolution of two synthetic vectors $\\vec{s}_0$ and $\\vec{s}_1$ . We extend the above situation to 2D by efficiently extracting axial high-frequency information using 2D discrete wavelet transform (2D-DWT), and designing learnable methods to make it adaptive to the data distribution and feature layers. Fig. 3(b) presents the construction process of the forward wavelet convolution kernel in 2D, which can be expressed as:", + "bbox": [ + 496, + 469, + 892, + 650 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {F} _ {l l} = \\vec {a _ {0}} \\times \\vec {a _ {0}} ^ {T}, \\mathcal {F} _ {l h} = \\vec {a _ {0}} \\times \\vec {a _ {1}} ^ {T}\n$$\n", + "text_format": "latex", + "bbox": [ + 581, + 661, + 808, + 679 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {F} _ {h l} = \\vec {a _ {1}} \\times \\vec {a _ {0}} ^ {T}, \\mathcal {F} _ {h h} = \\vec {a _ {1}} \\times \\vec {a _ {1}} ^ {T} \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 580, + 683, + 890, + 700 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {K} _ {w} = \\operatorname {c a t} \\left(\\mathcal {F} _ {l l}, \\mathcal {F} _ {l h}, \\mathcal {F} _ {h l}, \\mathcal {F} _ {h h}\\right)\n$$\n", + "text_format": "latex", + "bbox": [ + 581, + 705, + 784, + 720 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We set $\\vec{a_0}$ and $\\vec{a_1}$ as learnable filters; $\\mathcal{F}_{ll}$ , $\\mathcal{F}_{lh}$ , $\\mathcal{F}_{hl}$ , and $\\mathcal{F}_{hh}$ are low-frequency, horizontal high-frequency, vertical high-frequency, and diagonal high-frequency convolution operators obtained from vector outer products; The wavelet kernel $\\kappa_w$ is spliced from above convolution operators.", + "bbox": [ + 496, + 733, + 892, + 809 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Then we implement learnable wavelet transform as a group convolution like Fig. 3(a), given a set of input feature maps $\\mathcal{X}_{in} \\in (C, H, W)$ , its projection in wavelet domain $\\mathcal{X}_{out} \\in \\left(4C, \\frac{H}{2}, \\frac{W}{2}\\right)$ (low-frequency, horizontal high-frequency, vertical high-frequency, and diagonal high-frequency component) is generated through the wavelet", + "bbox": [ + 496, + 810, + 893, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "2736", + "bbox": [ + 480, + 924, + 514, + 936 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "convolution. Similarly, the construction method of the inverse kernel and the forward propagation can be easily deduced, since they are inverse processes of each other.", + "bbox": [ + 75, + 90, + 468, + 136 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Then, a natural question can then be posed: how to ensure the correctness of the wavelet kernel learning and ensure that it does not degrade into a general group convolution and cause performance degradation?", + "bbox": [ + 75, + 136, + 468, + 196 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "A common practice is to introduce the principle of perfect reconstruction [16, 34] to constrain adaptive wavelets, for a complex $z \\in \\mathbb{C}$ , given a filter $\\vec{x}$ that applies to the $z$ -transform, its $z$ -transform can be expressed as $X(z) = \\sum_{n \\in \\mathbb{Z}} \\vec{x}(n) z^{-n}$ . Then $\\vec{a}_0, \\vec{a}_1, \\vec{s}_0, \\vec{s}_1$ can be obtained as their corresponding $z$ -transforms $A_0, A_1, S_0, S_1$ , which must then be satisfied if perfect reconstruction is desired:", + "bbox": [ + 75, + 196, + 468, + 303 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nA _ {0} (- z) S _ {0} (z) + A _ {1} (- z) S _ {1} (z) = 0 \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 153, + 315, + 468, + 332 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nA _ {0} (z) S _ {0} (z) + A _ {1} (z) S _ {1} (z) = 2 \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 166, + 358, + 468, + 375 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Eq. 6 is known as aliasing cancellation condition, which is used to the cancellation of the aliasing effects arising from the downsampling. We will give a detailed derivation of Eq. 6 and Eq. 7 in the Appendix.", + "bbox": [ + 75, + 382, + 468, + 443 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "After wavelet convolution, to achieve frequency restoration in the wavelet domain, the input $\\mathcal{X}$ wavelet domain component was separated into a separate dim, we use depth-wise convolution with an expansion factor of $r$ for wavelet domain feature extraction and transformation after the wavelet positive transform, and $1\\times 1$ convolution for channel expansion and scaling, and finally a learnable wavelet inverse transform is used to reduce the wavelet domain feature maps to the spatial domain for output, which constitutes the LWN. We follow the rules of [2] and design structurally similar Wavelet Head Block (WHB) and Wavelet Fusion Block (WFB), both with LWN as an important base module, but used for semantic aggregation and multiscale output at different scales, respectively; thus, WHB adds the recovery convolution branch after WFB.", + "bbox": [ + 75, + 443, + 468, + 670 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Loss", + "text_level": 1, + "bbox": [ + 76, + 679, + 148, + 694 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Multi-scale Loss. Follow PSNR loss [2], we propose multiscale loss function as the main loss of the algorithm for calculating the pixel difference between restored image and GT at each scale:", + "bbox": [ + 75, + 702, + 468, + 763 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {m u l t i}} (x, y) = \\sum_ {k = 1} ^ {K} w _ {k} \\times \\mathcal {L} _ {\\text {p s n r}} \\left(x _ {k}, y _ {k}\\right) \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 135, + 772, + 468, + 814 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $x, y$ denote the output and GT, respectively, $k$ denotes the downsampling level, $w_{k}$ represents the weights under the corresponding scale. Our design of $w_{k}$ is based on the simple finding that we cannot produce absolutely accurate downsampled clear images, and that the accuracy gap", + "bbox": [ + 75, + 824, + 468, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "is progressively enlarged with decreasing scales. To ensure that lower scale output layers do not negatively affect the final output, we empirically set the loss weight $w_{k}$ to $\\frac{1}{k}$ for the corresponding k-scale.", + "bbox": [ + 496, + 90, + 890, + 151 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Wavelet Loss. Both Eq. 6 and Eq. 7 can be easily converted losses optimized toward a minimum of 0, here we use the mean square error. Then, based on the convolvable nature of the $z$ transform, it is possible to construct convolutionally equivalent substitutions for polynomial multiplications for the perfect reconstruction condition:", + "bbox": [ + 496, + 151, + 890, + 242 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} L _ {w a v e l e t} (\\theta_ {i}) = \\left(\\sum_ {k} ^ {N - 1} \\left(\\langle a _ {0}, s _ {0} \\rangle_ {k} + \\langle a _ {1}, s _ {1} \\rangle_ {k}\\right) - \\hat {V} _ {\\lfloor \\frac {N}{2} \\rfloor}\\right) ^ {2} \\\\ = \\left(\\sum_ {k} ^ {N - 1} \\left(\\left\\langle (- 1) ^ {k} a _ {0}, s _ {0} \\right\\rangle_ {k} + \\left\\langle (- 1) ^ {k} a _ {1}, s _ {1} \\right\\rangle_ {k}\\right)\\right) ^ {2} \\tag {9} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 500, + 252, + 924, + 356 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $k$ denotes the position of the filter when it performs the convolution, $\\hat{V}$ is a vector with a center position value of two, and $\\theta_{i}$ denotes the constructed filter", + "bbox": [ + 496, + 357, + 890, + 401 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Overall Loss. Based on the multi-scale loss and wavelet loss, the loss function used in this study is:", + "bbox": [ + 496, + 402, + 890, + 431 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {t o t a l}} (x, y) = \\mathcal {L} _ {\\text {w a v e l e t}} \\left(\\theta_ {i}\\right) + \\mathcal {L} _ {\\text {m u l t i}} (x, y) \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 535, + 444, + 890, + 460 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experimental Results", + "text_level": 1, + "bbox": [ + 498, + 472, + 705, + 488 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Datasets and Implementation Details", + "text_level": 1, + "bbox": [ + 498, + 497, + 818, + 513 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We evaluate our method on the realistic datasets Real-Blur [23] and RSBur [24], as well as the synthetic dataset GoPro [20], they are all composed of blurry-sharp image pairs. During training we use AdamW optimizer $(\\alpha = 0.9$ and $\\beta = 0.9)$ for a total of $600\\mathrm{K}$ iterations. The initial value of the learning rate is $10^{-3}$ and is updated with cosine annealing schedule. The patch size is set to $256\\times 256$ pixels and we followed the training strategy of [28, 38]. For data augmentation, we only use random flips and rotations. More experimental results can be found in the Appendix.", + "bbox": [ + 496, + 520, + 890, + 672 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Comparison with State-of-the-Art Methods", + "text_level": 1, + "bbox": [ + 498, + 681, + 867, + 696 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We have compared our method with several advanced deblurring methods and use the PSNR and SSIM to evaluate the quality of restored images.", + "bbox": [ + 496, + 704, + 890, + 750 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "**Evaluations on the RealBlur dataset.** Since our proposed method is driven by real-world deblurring, we first conduct comparisons on RealBlur [23]. The quantitative analysis results are shown in Tab. 1, the PSNR value of our method is 0.91dB and 0.49dB higher than state-of-the-art GRL on the RealBlur-J and RealBlur-R datasets respectively. Compared to other multi-scale architectures, our method has fewer model computational and running time while the performance is better. Fig. 4 shows visual comparison on RealBlur-J dataset, our method obtains restored text with", + "bbox": [ + 496, + 750, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "2737", + "bbox": [ + 480, + 925, + 514, + 936 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/215bdb7cdeb1601782a476847e66c891bd797b8de11ef519a88050281aeb185b.jpg", + "image_caption": [ + "Blurry image" + ], + "image_footnote": [], + "bbox": [ + 80, + 88, + 310, + 181 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/f9f64d806f80acd40cacf204f51bef984d3f170b07eba1240a1638ff150c106c.jpg", + "image_caption": [ + "DeblurGAN-v2" + ], + "image_footnote": [], + "bbox": [ + 312, + 88, + 424, + 181 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/f308de45c37be92ad81de0c16181cfe61669b994af9a9e6b40a08c2b9926d208.jpg", + "image_caption": [ + "SRN" + ], + "image_footnote": [], + "bbox": [ + 428, + 88, + 542, + 181 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/717ebaf86742e707eed5b75d11d5d6117c6f0b90dd9a2492ca358dadd843c627.jpg", + "image_caption": [ + "MIMO-UNet+" + ], + "image_footnote": [], + "bbox": [ + 545, + 88, + 658, + 181 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/16cd1a14bbe99a59e5358466b6f42f140bcd8f6bf54d673e7c4d748a255b8253.jpg", + "image_caption": [ + "MPRNet" + ], + "image_footnote": [], + "bbox": [ + 661, + 88, + 774, + 181 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/e6f347886090fcb55653e1237bbf23fe901583ab80271d6d3c655228a5af0c24.jpg", + "image_caption": [ + "BANet" + ], + "image_footnote": [], + "bbox": [ + 777, + 88, + 890, + 181 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/3daa0d4338f14c248d325bd2036e8037fa6b5037d2e3e76130dfcfab43938f88.jpg", + "image_caption": [ + "Blurry" + ], + "image_footnote": [], + "bbox": [ + 80, + 199, + 194, + 292 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/57b0eb1ced75505fc522cc0f50c80a5e1966062e2ba733ce7ab57b1efcd8d607.jpg", + "image_caption": [ + "GT" + ], + "image_footnote": [], + "bbox": [ + 196, + 199, + 310, + 292 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/31dc3d903701c53a32ed339747e77585c5d667cc143e09e70832695e88bcc0d9.jpg", + "image_caption": [ + "MSSNet" + ], + "image_footnote": [], + "bbox": [ + 312, + 199, + 424, + 292 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/35e24b695b5265712dd11e27b6eeb0cd771f581ee3696a63e5337195c88d06f4.jpg", + "image_caption": [ + "Stripformer", + "Figure 4. Visual comparisons on the RealBlur-J dataset [23]. The proposed method generates an image with clearer characters." + ], + "image_footnote": [], + "bbox": [ + 428, + 199, + 542, + 292 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/200900a46d8bfd7a0cc96759283ccb127c4c9378eaf6d1cd57f57ae3fc93d4e8.jpg", + "image_caption": [ + "GRL" + ], + "image_footnote": [], + "bbox": [ + 544, + 199, + 656, + 292 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/37eaff3d66cc31ceebaa60029b2c839c63842d72a257f26c2eac1f5a6833f587.jpg", + "image_caption": [ + "FFTformer" + ], + "image_footnote": [], + "bbox": [ + 660, + 199, + 774, + 292 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/46b70589e14259f5b5eb2ab46222bacc8ccb610982af6f74ab8c8988e938aa3b.jpg", + "image_caption": [ + "Ours" + ], + "image_footnote": [], + "bbox": [ + 777, + 199, + 890, + 292 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/ba7114fd2f35369deadc9e1087a81a66ff7c4d528640200c63e49226cdc83f98.jpg", + "table_caption": [], + "table_footnote": [ + "Table 1. Quantitative evaluations on the RealBlur dataset [23]. The experimental results were trained under the corresponding datasets respectively, and average runtime is tested on $256 \\times 256$ patches." + ], + "table_body": "
MethodRealBlur-JRealBlur-RAvg. runtime
PSNRSSIMPSNRSSIM
DeblurGAN-v2 [12]29.690.87036.440.9350.04s
SRN [27]31.380.90938.650.9650.07s
MPRNet [37]31.760.92239.310.9720.09s
SDWNet [42]30.730.89638.210.9630.04s
MIMO-UNet+ [3]31.920.919--0.02s
MIMO-UNet++ [3]32.050.921---
DeepRFT+ [19]32.190.93139.840.9720.09s
BANet [29]32.000.92339.550.9710.06s
BANet+ [29]32.420.92939.900.9720.12s
Stripformer [28]32.480.92939.840.9740.04s
MSSNet [9]32.100.92839.760.9720.06s
MSDI-Net [13]32.350.923--0.06s
MAXIM-3S [30]32.840.93539.450.962-
FFTformer [10]32.620.93340.110.9730.13s
GRL-B [14]32.820.93240.200.9741.28s
MLWNet-S33.020.933--0.04s
MLWNet-B33.840.94140.690.9760.05s
", + "bbox": [ + 78, + 335, + 470, + 645 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "richer edge details, while achieving the best contrast, sharpness, brightness, and structural details of the image.", + "bbox": [ + 75, + 700, + 468, + 729 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluations on the RSBlur dataset. We further conduct experiments on the latest real-world deblurring dataset RSBlur [24] and follow the protocol of this dataset for fair comparison. Tab. 2 summarizes the quantitative evaluation results of our method with advanced algorithms. The proposed MLWNet achieved the highest PSNR and SSIM, which were 34.94 and 0.880 respectively. In addition, the model we trained on RSBlur dataset also achieved the best results on RealBlur-J dataset. We show some visual comparisons in Fig. 5. We note that our method outperforms other methods in low-light blurry scenes, proving that the", + "bbox": [ + 75, + 734, + 468, + 900 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/7450463824183b9db68bbd13cb5738bdd69d5e93fc599a78406688fb2f5be409.jpg", + "table_caption": [], + "table_footnote": [ + "Table 2. Quantitative evaluations trained on the RSBlur dataset [24], the RealBlur-J dataset was used for testing only." + ], + "table_body": "
MethodRSBlurRealBlur-J
PSNRSSIMPSNRSSIM
SRN [27]32.530.84029.860.886
MIMO-Unet [3]32.730.84629.530.876
MIMO-Unet+ [3]33.370.85629.990.889
MPRNet [37]33.610.86130.460.899
Restormer [38]33.690.86330.480.891
Uformer-B [33]33.980.86630.370.899
SFNet [4]34.350.87230.260.897
MLWNet-B34.940.88030.530.905
", + "bbox": [ + 503, + 335, + 888, + 508 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/efa659ce56bac1ee88871224700e8a6dde43806a34e6a4ffab7d0ab0ee3e4087.jpg", + "table_caption": [], + "table_footnote": [ + "Table 3. Quantitative evaluation for generalizability shows the results of models trained on the RealBlur-J dataset and tested on the RSBlur dataset, MACs are measured on $256 \\times 256$ patches." + ], + "table_body": "
Method[23]→ [24]MACs(G)
PSNRSSIM
DeblurGAN-v2 [12]30.150.76642.0
MPRNet [37]29.560.785760.8
MIMO-UNet+ [3]29.690.792154.4
BANet [29]30.190.806263.9
BANet+ [29]30.240.809588.7
MSSNet [9]29.860.806154.0
FFTformer [10]29.700.787131.8
MLWNet-B30.910.818108.2
", + "bbox": [ + 506, + 549, + 887, + 722 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "proposed method makes textures and structures clearer.", + "bbox": [ + 500, + 777, + 864, + 791 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In addition, we evaluate the RSBlur dataset using models trained only on RealBlur-J to fairly compare the generalization of methods to real scenarios. As shown in Tab. 3, our method produced results with highest PSNR value of 30.91. The results of Tab. 2 and Tab. 3 show that our model has a better generalization ability.", + "bbox": [ + 496, + 792, + 890, + 883 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluations on the GoPro dataset. We conduct a exten", + "bbox": [ + 500, + 885, + 890, + 898 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "2738", + "bbox": [ + 482, + 926, + 514, + 936 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/0b6d3a6ad6bc0f669e5be3cda7b9bc2baab1203c222d8fb672c8cd84d1741279.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 81, + 90, + 348, + 188 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/90971285365ea3258b0f209a1d8299550b1f51c18a78e938ea98f12aca3782f2.jpg", + "image_caption": [ + "SRN" + ], + "image_footnote": [], + "bbox": [ + 351, + 89, + 483, + 188 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/fbe3bf2dcd531d68e504db1dafedbc5efef65c39f6f4c29b9d25bc2d4b7d1eeb.jpg", + "image_caption": [ + "MIMO-UNet" + ], + "image_footnote": [], + "bbox": [ + 486, + 90, + 617, + 188 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b611528098ba55aa72842822a6b13408c852eb2b2a3848380bf3d4c8ff6a1b75.jpg", + "image_caption": [ + "MIMO-UNet+" + ], + "image_footnote": [], + "bbox": [ + 620, + 90, + 751, + 188 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c6b608e12729bb2f6f17075e0d0397d50b43986036af164a3c423af3961e9665.jpg", + "image_caption": [ + "MPRNet" + ], + "image_footnote": [], + "bbox": [ + 756, + 90, + 888, + 188 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/cc83602cb1a7a0ebeb2c8740b78c76780d2882845493c2ba1ba497bf789056a3.jpg", + "image_caption": [ + "Blurry image", + "Blurry" + ], + "image_footnote": [], + "bbox": [ + 80, + 200, + 212, + 299 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/3e004f48ac48ba037fd91aad6041bf4840a2f29198106f857808621c42006dce.jpg", + "image_caption": [ + "GT" + ], + "image_footnote": [], + "bbox": [ + 215, + 200, + 348, + 299 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f00e7c52ed013110c1f65c797c5d0fc44342c733b5e52ffde9afd24b8afbab65.jpg", + "image_caption": [ + "Restormer" + ], + "image_footnote": [], + "bbox": [ + 351, + 200, + 483, + 299 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/668ce462ff79c03992ca1d403b1948c1f2a2d43f0151e5f4608da957b497597b.jpg", + "image_caption": [ + "Uformer-B" + ], + "image_footnote": [], + "bbox": [ + 486, + 200, + 617, + 299 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/585ba272724aba34360d1598872f15fecfe33df5104a6e966bfa337dbfc3d71b.jpg", + "image_caption": [ + "SFNet" + ], + "image_footnote": [], + "bbox": [ + 622, + 200, + 754, + 299 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/6003de8f5bcef6e4b2c2101766cff97d59f4a10adfbf822687d48424ccca6bfc.jpg", + "image_caption": [ + "Ours" + ], + "image_footnote": [], + "bbox": [ + 756, + 200, + 888, + 299 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/acac45ca631bc8e8518d5dd1bbb4ed01f503a3d81280e8fffdd5f9cf8d8e5bfb.jpg", + "image_caption": [ + "Figure 5. Visual comparisons on the RSBur dataset [24]. The deblurring performance of the proposed method in low-light is impressive, The recovery of characters and texture structures far exceeds other advanced methods.", + "Blurry image" + ], + "image_footnote": [], + "bbox": [ + 81, + 340, + 308, + 421 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/0768e11c2fe4ca5753d8a7ea2eaaaa63f9f065e636de54179cb6e118e3f91a7a.jpg", + "image_caption": [ + "MIMO-UNet" + ], + "image_footnote": [], + "bbox": [ + 313, + 340, + 424, + 420 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ec338d67bd377e6bb9d5ea8da1f1a0d773a1e82d5b909405d39fb112daa4d5ef.jpg", + "image_caption": [ + "MIMO-UNet+" + ], + "image_footnote": [], + "bbox": [ + 428, + 340, + 542, + 420 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/301062017a4633defa98faa1e007decee4f9f7e594c78ae00ec8abe8df9c0053.jpg", + "image_caption": [ + "MPRNet" + ], + "image_footnote": [], + "bbox": [ + 545, + 340, + 656, + 420 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/eb2837ebefc59f58cd28d1439d3df4eecbaac650ecac73a34adc751830681732.jpg", + "image_caption": [ + "NAFNet" + ], + "image_footnote": [], + "bbox": [ + 660, + 340, + 772, + 420 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/091cc93979440846cfef3b1c91fede03162da15eea65e86a7a7a3d649a0ada82.jpg", + "image_caption": [ + "MAXIM" + ], + "image_footnote": [], + "bbox": [ + 776, + 340, + 888, + 420 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/31dbd953817bb7c2f23ef591a3a38d007ff62e3b66e95fa2d11b5fa6dcb78acf.jpg", + "image_caption": [ + "Blurry", + "Figure 6. Visual comparisons on the GoPro dataset [20]. Our method better preserves texture information without sharpening." + ], + "image_footnote": [], + "bbox": [ + 80, + 438, + 192, + 518 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/12d8c3b7d37de45745154f0618c99448f1e480beb030c83d63ff7d57cfc8084c.jpg", + "image_caption": [ + "GT" + ], + "image_footnote": [], + "bbox": [ + 196, + 438, + 308, + 518 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/47959d8fc10e347beebe3d17f3b7ede9f64fdf474e8f6d1284af07cca1daf510.jpg", + "image_caption": [ + "Stripformer" + ], + "image_footnote": [], + "bbox": [ + 313, + 438, + 424, + 518 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/5a8102137173b287f0fd68320bffa4493735e681c26ca8f2093c1b205e20fe56.jpg", + "image_caption": [ + "Restormer" + ], + "image_footnote": [], + "bbox": [ + 428, + 438, + 542, + 518 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/62a8eeac1b07563f9ea4a80110438401e91ced256a726e01301094ee97714830.jpg", + "image_caption": [ + "FFTformer" + ], + "image_footnote": [], + "bbox": [ + 545, + 438, + 656, + 518 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/bd3a83710a26c343281ae0c5c97ff0cfc0a78c13e971244036d2b7cd8b960bb2.jpg", + "image_caption": [ + "GRL" + ], + "image_footnote": [], + "bbox": [ + 661, + 438, + 772, + 518 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/53feea5edbd48f545a144cef2f5990d2b2bf004e2376d5d40d5b00464aff7989.jpg", + "image_caption": [ + "Ours" + ], + "image_footnote": [], + "bbox": [ + 777, + 438, + 888, + 518 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "sive comparison with advanced algorithms on GoPro [20]. Tab. 4 shows the quantitative evaluation results. The gap between our method and the state-of-the-art FFTformer is only 0.001 in SSIM value, while the running time is reduced by $60\\%$ . We show a visual comparison on the GoPro dataset in Fig. 6. We can see that our method recovers blurred textures well without sharpening. For the reasons why the proposed MLWNet does not achieve optimal performance in synthetic blurry, we conducted a detailed analysis in Sec. 5.", + "bbox": [ + 75, + 559, + 468, + 696 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5. Analysis and Discussion", + "text_level": 1, + "bbox": [ + 76, + 709, + 300, + 724 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In this section, we provide a more in-depth analysis and show the contribution of each component of the proposed method. We perform 20w iterations on the RealBlur-J dataset for abtation with a batch size of 8 to train our method using a version of the model with a width of 32, and the results are shown in Tab. 6.", + "bbox": [ + 75, + 734, + 468, + 823 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Single-scale vs. Multi-scale. Given that each node of the baseline network is composed of SEB, we have deformed the multi-scale and single-scale according to the architecture. Tab. 5 shows that the multi-scale architecture improves the PSNR and SSIM values to varying degrees, proving that", + "bbox": [ + 75, + 825, + 468, + 900 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/9372afe43d9b399ebe031c9906f4a4a0479063edd24913fadbbaef85a6d70264.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodGOPROAvg.rruntime
PSNRSSIM
DeblurGAN-v2 [12]29.550.9340.04s
SRN [27]30.260.9340.07s
DMPHN [39]31.200.9450.21s
SDWNet [42]31.260.9660.04s
MPRNet [37]32.660.9590.09s
MIMO-UNet+ [3]32.450.9570.02s
DeepRFT+ [19]33.230.9630.09s
MAXIM-3S [30]32.860.961-
Stripformer [28]33.080.9620.04s
MSDI-net [13]33.280.9640.06s
Restormer [38]33.570.9660.08s
NAFNet [2]33.690.9670.04s
FFTformer [10]34.210.9690.13s
GRL-B [14]33.930.9681.28s
MLWNet-B33.830.9680.05s
", + "bbox": [ + 501, + 556, + 890, + 835 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 4. Quantitative evaluations trained and tested on the GoPro dataset [20]. Our proposed MLWNet obtains competitive results with a combination of time efficiency and accuracy.", + "bbox": [ + 500, + 838, + 890, + 878 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "2739", + "bbox": [ + 482, + 926, + 514, + 936 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "the SIMO strategy we employ helps to eliminate motion blur to some extent. As shown in Sec. 3.1, we use SIMO for coarse-to-fine restoration, and only use multiple outputs to calculate multi-scale losses during training.", + "bbox": [ + 75, + 90, + 468, + 151 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "WFB and WHB. We default to using WFB and WHB via $\\mathcal{L}_{wavelet}$ statute in this section, we also proved the effectiveness of LWN, because LWN is a subpart of the first two. After applying them, the PSNR index improved by $0.25\\mathrm{dB}$ compared to the multi-scale baseline, which shows that LWN and its extension modules can easily help multiscale algorithms improve detail recovery capabilities. Fig. 7 shows the four types of feature maps generated after forward learnable wavelet convolution. We can see that the high and low frequency information of the input feature map are mixed together, and after wavelet group convolution, the network exerts different attention on the feature information in different directions or frequencies.", + "bbox": [ + 75, + 151, + 470, + 347 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effectiveness of $\\mathcal{L}_{wavelet}$ . By adding unreduced WFB and WHB to the baseline without using wavelet loss, the performance improvement of the baseline is minimal and does not break the bottleneck of multiscale detail recovery, which suggests that the wavelet convolution degenerates into a general group convolution that only serves to deepen the network, proving that merely deepening the network is ineffective for the restoration of detail information.", + "bbox": [ + 75, + 347, + 468, + 467 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/cd02f1cd3b4839bc6c7318b37781946a0100a9e52e798190ac2065cee10c3818.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodSISOMIMOSIMO
PSNR/SSIM32.29/0.92432.19/0.92832.37/0.929
MACs(G)19.2421.8319.29
", + "bbox": [ + 78, + 470, + 470, + 518 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/a951afaa55d1cd6c5714cbec9b6eb2e0bb02dd70b3a9940f5a92fd9c8257229f.jpg", + "table_caption": [ + "Table 5. Comparison in various input and output modes." + ], + "table_footnote": [], + "table_body": "
SIMOWFBWHBLwaveletPSNRSSIMMACs(G)
✓32.370.92919.29
✓✓✓32.400.92828.21
✓✓✓32.490.92825.28
✓✓✓32.570.92922.22
✓✓✓✓32.620.93128.21
", + "bbox": [ + 78, + 536, + 480, + 641 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/58b3cc4ebc77ac4abc8cad372c4923708f469c568b1480f2a614e68a40c43774.jpg", + "image_caption": [ + "Figure 7. Feature maps representing high-and low-frequency components generated after learnable wavelet convolution. Zoom in on the screen for the best view." + ], + "image_footnote": [], + "bbox": [ + 78, + 686, + 468, + 763 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Limitations and analysis. It is observed from Tab. 1-3 that our method fails to achieve the same expected high performance with synthetic blur as it does with realistic blur. We try to analyze the reasons for this phenomenon via dataset comparison and experimental results. First, the average frame synthesis method used in the synthetic blur", + "bbox": [ + 75, + 809, + 468, + 898 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "dataset leads to unnatural discontinuous blur trajectories, thereby introducing strange high-frequency information interference (see Fig. 8). Second, synthetic data tends to produce an unnatural mixture of high and low frequencies. It is easy to mix with the average color area at the edge of the texture, causing the confusion of high-low frequency information.", + "bbox": [ + 496, + 90, + 890, + 196 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/4479fc93893bb8334181aacfbef23656b02524fd32475217e26cdf6f96467f9a.jpg", + "image_caption": [ + "(a) Realistic Blur" + ], + "image_footnote": [], + "bbox": [ + 503, + 205, + 694, + 334 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/02bae2280ed7d33a0ca02349aadccfa0ecd089ceb702a3479a3e441fca2e5f15.jpg", + "image_caption": [ + "(b) Synthetic Blur", + "Figure 8. The difference between realistic blur (a) and synthetic blur (b). In the green box, the synthetic blur appears with color averaging resulting in high and low frequency confusion, and in the blue box has unnatural discontinuous trajectories." + ], + "image_footnote": [], + "bbox": [ + 697, + 205, + 890, + 334 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/4460013dfa5080df37ad365c0b55f39bae236b6b324939f1633d260e6b63f6b5.jpg", + "table_caption": [ + "Table 6. Ablation study on components of the proposed MLWNet. We set the baseline network to use SEB in its entirety, and models that do not use SIMO will represent single scales using SISO." + ], + "table_footnote": [], + "table_body": "
Noise LevelL1L2L3L4L5
GoPro34.8134.6633.7633.1932.63
RealBlur-J33.9233.8133.9733.9333.54
", + "bbox": [ + 504, + 410, + 888, + 470 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 7. Performance comparison at different noise difference levels, where L3 contains the noise difference mean.", + "bbox": [ + 496, + 473, + 888, + 500 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Finally, there is a large difference in noise levels between synthetic blur and real blur. We follow [1] to calculate the noise level of each testset, and divided the noise difference between clear images and blurred images of GoPro and RealBlur-J into 5 levels for separate testing. As shown in Tab. 7, we note that as the noise difference increases, our method maintains good performance on RealBlur-J, while the performance drops fast on GoPro. This shows that our method is robust to the noise of real blur, and that there are differences in the noise between synthetic blur and real blur.", + "bbox": [ + 496, + 501, + 890, + 652 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 664, + 617, + 679 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we propose a SIMO-based multi-scale architecture to achieve efficient motion deblurring. We have developed a learnable discrete wavelet transform module, which not only improves the algorithm's ability to recover details, but is also more applicable to the real world. In addition, we construct a reasonable multi-scale loss to guide the recovery of blurred images pixel by pixel and scale by scale, and constrain the learning direction of the wavelet kernel with self-supervised loss to achieve better image deblurring. Through extensive experiments on multiple real and synthetic datasets, we demonstrate that it outperforms existing state-of-the-art methods in terms of quality and efficiency of image restoration, especially with optimal deblurring performance and generalization in real scenes.", + "bbox": [ + 496, + 688, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "2740", + "bbox": [ + 480, + 926, + 514, + 936 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Guangyong Chen, Fengyuan Zhu, and Pheng Ann Heng. An efficient statistical method for image noise level estimation. In Proceedings of the IEEE International Conference on Computer Vision, pages 477-485, 2015. 8", + "[2] Liangyu Chen, Xiaojie Chu, Xiangyu Zhang, and Jian Sun. Simple baselines for image restoration. In European Conference on Computer Vision, pages 17-33. Springer, 2022. 1, 2, 3, 5, 7", + "[3] Sung-Jin Cho, Seo-Won Ji, Jun-Pyo Hong, Seung-Won Jung, and Sung-Jea Ko. Rethinking coarse-to-fine approach in single image deblurring. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4641-4650, 2021. 1, 2, 6, 7", + "[4] Yuning Cui, Yi Tao, Zhenshan Bing, Wenqi Ren, Xinwei Gao, Xiaochun Cao, Kai Huang, and Alois Knoll. Selective frequency network for image restoration. In International Conference on Learning Representations, ICLR, 2023. 6", + "[5] Ingrid Daubechies. The wavelet transform, time-frequency localization and signal analysis. IEEE transactions on information theory, 36(5):961-1005, 1990. 2", + "[6] Golnaz Ghiasi, Tsung-Yi Lin, and Quoc V Le. Nas-fpn: Learning scalable feature pyramid architecture for object detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7036-7045, 2019. 1", + "[7] Wooseok Ha, Chandan Singh, Francois Lanusse, Srigokul Upadhyayula, and Bin Yu. Adaptive wavelet distillation from neural networks through interpretations. Advances in Neural Information Processing Systems, 34:20669-20682, 2021. 4", + "[8] Bumsoo Kim, Jonghwan Mun, Kyoung-Woon On, Minchul Shin, Junhyun Lee, and Eun-Sol Kim. Mstr: Multi-scale transformer for end-to-end human-object interaction detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19578-19587, 2022. 1", + "[9] Kiyeon Kim, Seungyong Lee, and Sunghyun Cho. Mssnet: Multi-scale-stage network for single image deblurring. In European Conference on Computer Vision, pages 524-539. Springer, 2022. 1, 2, 6", + "[10] Lingshun Kong, Jiangxin Dong, Jianjun Ge, Mingqiang Li, and Jinshan Pan. Efficient frequency domain-based transformers for high-quality image deblurring. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5886-5895, 2023. 1, 2, 3, 6, 7", + "[11] Orest Kupyn, Volodymyr Budzan, Mykola Mykhailych, Dmytro Mishkin, and Jiří Matas. Deblurgan: Blind motion deblurring using conditional adversarial networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 8183-8192, 2018. 2", + "[12] Orest Kupyn, Tetiana Martyniuk, Junru Wu, and Zhangyang Wang. Deblurgan-v2: Deblurring (orders-of-magnitude) faster and better. In Proceedings of the IEEE/CVF international conference on computer vision, pages 8878-8887, 2019. 1, 6, 7", + "[13] Dasong Li, Yi Zhang, Ka Chun Cheung, Xiaogang Wang, Hongwei Qin, and Hongsheng Li. Learning degradation rep" + ], + "bbox": [ + 78, + 114, + 468, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "resentations for image deblurring. In European Conference on Computer Vision, pages 736-753. Springer, 2022. 1, 2, 6, 7", + "[14] Yawei Li, Yuchen Fan, Xiaoyu Xiang, Denis Demandolx, Rakesh Ranjan, Radu Timofte, and Luc Van Gool. Efficient and explicit modelling of image hierarchies for image restoration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18278-18289, 2023. 2, 6, 7", + "[15] Tsung-Yi Lin, Piotr Dólar, Ross Girshick, Kaiming He, Bharath Hariharan, and Serge Belongie. Feature pyramid networks for object detection. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2117-2125, 2017. 1", + "[16] Pengju Liu, Hongzhi Zhang, Wei Lian, and Wangmeng Zuo. Multi-level wavelet convolutional neural networks. IEEE Access, 7:74973-74985, 2019. 4, 5", + "[17] Shu Liu, Lu Qi, Haifang Qin, Jianping Shi, and Jiaya Jia. Path aggregation network for instance segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 8759-8768, 2018. 1", + "[18] Wei Liu, Qiong Yan, and Yuzhi Zhao. Densely self-guided wavelet network for image denoising. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, pages 432-433, 2020. 4", + "[19] X Mao, Y Liu, W Shen, Q Li, and Y Wang. Deep residual fourier transformation for single image deblurring. arxiv 2021. arXiv preprint arXiv:2111.11745, 2021. 2, 3, 6, 7", + "[20] Seungjun Nah, Tae Hyun Kim, and Kyoung Mu Lee. Deep multi-scale convolutional neural network for dynamic scene deblurring. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3883-3891, 2017. 1, 2, 5, 7", + "[21] Dongwon Park, Dong Un Kang, Jisoo Kim, and Se Young Chun. Multi-temporal recurrent neural networks for progressive non-uniform single image deblurring with incremental temporal training. In European Conference on Computer Vision, pages 327-343. Springer, 2020. 1, 2", + "[22] Zequn Qin, Pengyi Zhang, Fei Wu, and Xi Li. Fcanet: Frequency channel attention networks. In Proceedings of the IEEE/CVF international conference on computer vision, pages 783-792, 2021. 3", + "[23] Jaesung Rim, Haeyun Lee, Jucheol Won, and Sunghyun Cho. Real-world blur dataset for learning and benchmarking deblurring algorithms. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXV 16, pages 184-201. Springer, 2020. 1, 5, 6", + "[24] Jaesung Rim, Geonung Kim, Jungeon Kim, Junyong Lee, Seungyong Lee, and Sunghyun Cho. Realistic blur synthesis for learning image deblurring. In European conference on computer vision, pages 487-503. Springer, 2022. 2, 5, 6, 7", + "[25] Roman Suvorov, Elizaveta Logacheva, Anton Mashikhin, Anastasia Remizova, Arsenii Ashukha, Aleksei Silvestrov, Naejin Kong, Harshith Goka, Kiwoong Park, and Victor Lempitsky. Resolution-robust large mask inpainting with fourier convolutions. In Proceedings of the IEEE/CVF winter" + ], + "bbox": [ + 503, + 92, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "2741", + "bbox": [ + 482, + 926, + 513, + 936 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "conference on applications of computer vision, pages 2149-2159, 2022. 3", + "[26] Mingxing Tan, Ruoming Pang, and Quoc V Le. Efficientdet: Scalable and efficient object detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10781-10790, 2020. 1", + "[27] Xin Tao, Hongyun Gao, Xiaoyong Shen, Jue Wang, and Jiaya Jia. Scale-recurrent network for deep image deblurring. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 8174-8182, 2018. 1, 2, 6, 7", + "[28] Fu-Jen Tsai, Yan-Tsung Peng, Yen-Yu Lin, Chung-Chi Tsai, and Chia-Wen Lin. Stripformer: Strip transformer for fast image deblurring. In European Conference on Computer Vision, pages 146-162. Springer, 2022. 1, 5, 6, 7", + "[29] Fu-Jen Tsai, Yan-Tsung Peng, Chung-Chi Tsai, Yen-Yu Lin, and Chia-Wen Lin. Banet: a blur-aware attention network for dynamic scene deblurring. IEEE Transactions on Image Processing, 31:6789-6799, 2022. 1, 2, 6", + "[30] Zhengzhong Tu, Hossein Talebi, Han Zhang, Feng Yang, Peyman Milanfar, Alan Bovik, and Yinxiao Li. Maxim: Multi-axis mlp for image processing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5769-5780, 2022. 6, 7", + "[31] Chien-Yao Wang, Alexey Bochkovskiy, and Hong-Yuan Mark Liao. Yolov7: Trainable bag-of-freebies sets new state-of-the-art for real-time object detectors. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7464-7475, 2023. 1", + "[32] Luyuan Wang and Yankui Sun. Image classification using convolutional neural network with wavelet domain inputs. IET Image Processing, 16(8):2037-2048, 2022. 3", + "[33] Zhendong Wang, Xiaodong Cun, Jianmin Bao, Wengang Zhou, Jianzhuang Liu, and Houqiang Li. Uformer: A general u-shaped transformer for image restoration. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 17683-17693, 2022. 6", + "[34] Moritz Wolter and Jochen Garcke. Adaptive wavelet pooling for convolutional neural networks. In International Conference on Artificial Intelligence and Statistics, pages 1936-1944. PMLR, 2021. 4, 5", + "[35] Yanchao Yang and Stefano Soatto. Fda: Fourier domain adaptation for semantic segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4085-4095, 2020. 3", + "[36] Changqian Yu, Changxin Gao, Jingbo Wang, Gang Yu, Chunhua Shen, and Nong Sang. Bisenet v2: Bilateral network with guided aggregation for real-time semantic segmentation. International Journal of Computer Vision, 129: 3051-3068, 2021. 4", + "[37] Syed Waqas Zamir, Aditya Arora, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, Ming-Hsuan Yang, and Ling Shao. Multi-stage progressive image restoration. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14821-14831, 2021. 1, 2, 6, 7", + "[38] Syed Waqas Zamir, Aditya Arora, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, and Ming-Hsuan Yang." + ], + "bbox": [ + 78, + 92, + 470, + 902 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Restormer: Efficient transformer for high-resolution image restoration. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5728-5739, 2022. 2, 5, 6, 7", + "[39] Hongguang Zhang, Yuchao Dai, Hongdong Li, and Piotr Koniusz. Deep stacked hierarchical multi-patch network for image deblurring. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5978-5986, 2019. 2, 7", + "[40] Kaihao Zhang, Wenhan Luo, Yiran Zhong, Lin Ma, Bjorn Stenger, Wei Liu, and Hongdong Li. Deblurring by realistic blurring. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2737-2746, 2020. 2", + "[41] Yijie Zhong, Bo Li, Lv Tang, Senyun Kuang, Shuang Wu, and Shouhong Ding. Detecting camouflaged object in frequency domain. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4504-4513, 2022. 3", + "[42] Wenbin Zou, Mingchao Jiang, Yunchen Zhang, Liang Chen, Zhiyong Lu, and Yi Wu. Sdwnet: A straight dilated network with wavelet transformation for image deblurring. In Proceedings of the IEEE/CVF international conference on computer vision, pages 1895-1904, 2021. 1, 2, 3, 6, 7" + ], + "bbox": [ + 501, + 92, + 890, + 429 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "2742", + "bbox": [ + 482, + 926, + 514, + 936 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2024/Efficient Multi-scale Network with Learnable Discrete Wavelet Transform for Blind Motion Deblurring/4af0ea1e-f85d-4e0b-ab29-04d268e0386c_model.json b/2024/Efficient Multi-scale Network with Learnable Discrete Wavelet Transform for Blind Motion Deblurring/4af0ea1e-f85d-4e0b-ab29-04d268e0386c_model.json new file mode 100644 index 0000000000000000000000000000000000000000..2c27097a129820fd80be4ef4177fbb2e2e7acd55 --- /dev/null +++ b/2024/Efficient Multi-scale Network with Learnable Discrete Wavelet Transform for Blind Motion Deblurring/4af0ea1e-f85d-4e0b-ab29-04d268e0386c_model.json @@ -0,0 +1,2882 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.0, + 0.812, + 0.045 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.113, + 0.13, + 0.857, + 0.176 + ], + "angle": 0, + "content": "Efficient Multi-scale Network with Learnable Discrete Wavelet Transform for Blind Motion Deblurring" + }, + { + "type": "text", + "bbox": [ + 0.086, + 0.182, + 0.903, + 0.237 + ], + "angle": 0, + "content": "Xin Gao\\*1,2 Tianheng Qiu\\*2,3,4 Xinyu Zhang†2 Hanlin Bai\\*1 Kang Liu\\*1 Xuan Huang†4 Hu Wei\\*4 Guoying Zhang†1 Huaping Liu\\*2 *Equal Contribution † Corresponding Author" + }, + { + "type": "text", + "bbox": [ + 0.197, + 0.242, + 0.798, + 0.298 + ], + "angle": 0, + "content": "\\(^{1}\\)China University of Mining & Technology-Beijing \\(^{2}\\)Tsinghua University \\(^{3}\\)University of Science and Technology of China \\(^{4}\\)Hefei Institutes of Physical Science, Chinese Academy of Sciences" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.317, + 0.314, + 0.334 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.349, + 0.474, + 0.683 + ], + "angle": 0, + "content": "Coarse-to-fine schemes are widely used in traditional single-image motion deblur; however, in the context of deep learning, existing multi-scale algorithms not only require the use of complex modules for feature fusion of low-scale RGB images and deep semantics, but also manually generate low-resolution pairs of images that do not have sufficient confidence. In this work, we propose a multi-scale network based on single-input and multiple-outputs(SIMO) for motion deblurring. This simplifies the complexity of algorithms based on a coarse-to-fine scheme. To alleviate restoration defects impacting detail information brought about by using a multi-scale architecture, we combine the characteristics of real-world blurring trajectories with a learnable wavelet transform module to focus on the directional continuity and frequency features of the step-by-step transitions between blurred images to sharp images. In conclusion, we propose a multi-scale network with a learnable discrete wavelet transform (MLWNet), which exhibits state-of-the-art performance on multiple real-world deblurred datasets, in terms of both subjective and objective quality as well as computational efficiency. Our code is available on https://github.com/thqiu0419/MLWNet." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.71, + 0.21, + 0.725 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.734, + 0.47, + 0.87 + ], + "angle": 0, + "content": "Most current top-performing single-image blind deblurring algorithms are based on DNNs, which can be structurally categorized into single-scale [2, 10, 12, 28, 29, 42] and multi-scale approaches [3, 9, 13, 27, 37]. Compared to single-scale methods, multi-scale methods are built on the idea of moving from coarse-to-fine, and thus they decompose the challenging single-image blind deblurring problem into easier-to-solve sub-problems to restore the blurred image step-by-step." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.47, + 0.901 + ], + "angle": 0, + "content": "Earlier DNNs exploiting the coarse-to-fine concept usually employ estimators of different scales to produce re" + }, + { + "type": "image", + "bbox": [ + 0.499, + 0.315, + 0.895, + 0.536 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.537, + 0.895, + 0.579 + ], + "angle": 0, + "content": "Figure 1. Performance comparison on the RealBlur-J [23] test dataset in terms of PSNR and GMACs. Our proposed MLWNet achieves superiority in comparison with other state-of-the-arts." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.584, + 0.893, + 0.825 + ], + "angle": 0, + "content": "stored images gradually [20, 21, 27]. However, not only is the extracted semantic information not sufficiently representative, but the use of the same complexity at different scales may lead to redundant computations; More advanced multi-scale algorithms [3] use a global encoder-decoder, which significantly reduces the algorithm's runtime, multiple upsampling and downsampling leads to insufficient restoration of detailed information, and therefore such algorithms require the introduction of additional fusion modules to encode input images for fusion with deeper semantics. In addition, existing multi-scale algorithms use multiple-input and multiple-output (MIMO) architectures, which require the introduction of manually constructed downsampled image pairs, and usually employ simple interpolation algorithms to generate low-resolution images for the sake of efficiency, which is obviously not reliable." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.826, + 0.894, + 0.9 + ], + "angle": 0, + "content": "We note that the gradual insertion of images used by existing coarse-to-fine algorithms is redundant. Referring to numerous algorithms [6, 8, 15, 17, 26, 31] that are widely used in downstream tasks, the network starts with a single input and also extracts features at different scales. These" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.925, + 0.516, + 0.937 + ], + "angle": 0, + "content": "2733" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.243 + ], + "angle": 0, + "content": "features can be aggregated by simple feature fusion to ultimately obtain competitive results. That is, a DNN itself possesses the ability to learn effective features at different scales. This inspired us to design a multi-scale architecture that retains an original image as input and sequentially restores images at each scale in output stage. This is not only more theoretically sound, but also eliminates the time required to generate pairs of images of different resolutions, as well as the huge complexity increase brought about by fusion modules between RGB domain and deep features." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.246, + 0.473, + 0.472 + ], + "angle": 0, + "content": "The coarse-to-fine algorithm has another inherent defect. During progressive restoration, the solution of the upper-level problem takes the solution of the lower-level as initialization [9]. Although this reduces the difficulty of solving the upper-level problem, due to the smaller resolution of lower-level spatial, the features transmitted upwards are semantically precise but spatially ambiguous, thus the restoration ability of multi-scale network in spatial details is limited. Hence, there is an urgent need to improve the quality of restoration of high-frequency detail part. A simple approach is to introduce a frequency domain transform as an alternative to a spatial domain transform. This would provide the algorithm with a choice and direct its attention to different frequencies. In this paper, we consider using the discrete wavelet transform for the following reasons:" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.476, + 0.47, + 0.596 + ], + "angle": 0, + "content": "1. Many recent state-of-the-art deblurring algorithms [10, 19, 42] have introduced the discrete Fourier transform (DFT) as a frequency prior, which provides information that helps the algorithm to identify and select high- and low-frequency components that need to be preserved during restoration. Compared to DFT, the discrete wavelet transform (DWT) is better suited to deal with images containing more abrupt signals [5]." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.596, + 0.47, + 0.763 + ], + "angle": 0, + "content": "2. Realistic blur and synthetic blur have significant distributional differences [24]. In the real world, due to the short exposure time of a camera, realistic blur has a specific directionality [29], i.e., the blur trajectory is regionally continuous; conversely, synthetic blur has an unnatural and discontinuous trajectory. To fully utilize the potential deblurring guidance brought by this trajectory continuity, we use 2D-DWT to reveal blur directionality, distinguish changes in the blur signal along different directions, and provide the algorithm with a reliable basis for deblurring through adaptive learning." + }, + { + "type": "list", + "bbox": [ + 0.078, + 0.476, + 0.47, + 0.763 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.471, + 0.902 + ], + "angle": 0, + "content": "To make 2D-DWT fit the data distribution and feature layer space more closely, we implemented 2D-DWT with an adaptive data distribution by using group convolution, transferring feature space from the spatial domain into the wavelet domain, generating sub-signals with different frequency features and different directional features, and then constructing wavelet losses for them in order to constrain them by self-supervision. Experimental results demonstrate that the proposed method has advanced performance in" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.773, + 0.107 + ], + "angle": 0, + "content": "terms of accuracy and efficiency (Fig. 1)." + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.108, + 0.82, + 0.121 + ], + "angle": 0, + "content": "In summary, our contributions are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.122, + 0.892, + 0.196 + ], + "angle": 0, + "content": "- We propose a single-input and multiple-output(SIMO) multi-scale baseline for progressive image deblur, which reduces the complexity of existing multi-scale deblurring algorithms and improves the overall efficiency of image restoration networks." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.198, + 0.892, + 0.272 + ], + "angle": 0, + "content": "- We construct a learnable discrete wavelet transform node(LWN) to provide reliable frequency and direction selection for the algorithm, which promotes the algorithm's restoration of the high-frequency components of edges and details." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.273, + 0.892, + 0.348 + ], + "angle": 0, + "content": "- For the network to work better, reasonable multi-scale loss is proposed to guide pixel-by-pixel and scale-by-scale restoration. We also created reasonable self-supervised losses for the learnable wavelet transform to limit the learning direction of the wavelet kernel." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.349, + 0.892, + 0.394 + ], + "angle": 0, + "content": "- We demonstrate the effectiveness of our algorithm under several motion blur datasets, especially real datasets, and obtain highly competitive results." + }, + { + "type": "list", + "bbox": [ + 0.5, + 0.122, + 0.892, + 0.394 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.408, + 0.642, + 0.423 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.433, + 0.892, + 0.689 + ], + "angle": 0, + "content": "Single-scale Deblurring Algorithm. Image deblurring have developed rapidly in recent years [2, 10, 11, 14, 39, 40]. To preserve richer image details, Kupyn et al. [11] viewed deblurring as a special case of image-to-image conversion, and for the first time utilized a GAN to recover images with richer details. To provide a high-quality and simple baseline, Chen et al. [2] proposed a nonlinear activation-free network that obtained excellent results. Zamir et al. [38] proposed a transformer with an encoder-decoder architecture that can be applied to higher-resolution images and utilizes cross-channel attention. Kong et al. [10] utilized a domain-based self-attention solver to reduce the transformer complexity and suppress artefacts. However, such single-scale algorithms estimate complex restoration problems directly and may require the design of restorers of higher complexity, which is suboptimal in terms of efficiency [9]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.689, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Multi-scale Deblurring Algorithm. Multi-scale motion deblur methods aim to achieve progressive recovery using a multi-stage or multi-sub-network approach. The multiscale algorithm previously used in the field of image deblurring is based on MIMO [3, 13, 20, 21, 27]. Recently, Cho et al. [3] proposed MIMO-UNet, which employs a multi-output single decoder as well as a multi-input single encoder to simulate a multi-scale architecture consisting of stacked networks, thereby greatly simplifying complexity. Zamir et al. [37] designed a multi-stage progressive image recovery network to simplify the information flow between multiple sub-networks. As shown in Sec. 3.1, our proposed multiscale approach further simplifies the structure of multi-scale networks and better balances the accuracy and complexity." + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.925, + 0.516, + 0.937 + ], + "angle": 0, + "content": "2734" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.08, + 0.089, + 0.887, + 0.359 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.365, + 0.893, + 0.409 + ], + "angle": 0, + "content": "Figure 2. The overall architecture of the proposed MLWNet, the SEB is a simple module designed with reference [2], the WFB and WHB apply the LWN that implements the learnable 2D-DWT. In training phase, supervised learning is performed using \\(\\mathcal{L}_{multi}\\) and self-supervised restraint of the wavelet kernel is performed using \\(\\mathcal{L}_{wavelet}\\). In testing phase, only the highest scale restored images is output." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.413, + 0.473, + 0.641 + ], + "angle": 0, + "content": "Application of Frequency in Deblurring. Frequency is an important property of images, in recent years frequency domain has been introduced into numerous DNNs to be solved for various tasks [10, 22, 25, 32, 35, 41]. Mao et al. [19] proposed DeepRFT, which introduced a residual module based on the Fourier transform into an advanced algorithm. Kong et al. [10] proposed FFTformer, which introduced the Fourier transform into a feed-forward network to effectively utilize different frequency information. Zou et al. [42] proposed SDWNet, which utilized frequency domain information as a complement to spatial domain information. Different from these methods, we introduce a learnable 2D-DWT capable of adapting to data distribution and feature space, more suitable not only for dealing with digital images rich in mutation signals, but also for dealing with real blurring." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.656, + 0.251, + 0.673 + ], + "angle": 0, + "content": "3. Proposed Method" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.682, + 0.47, + 0.834 + ], + "angle": 0, + "content": "Our proposed MLWNet(Fig. 2) aims to explore an efficient multi-scale architectural approach to achieve high-quality blind deblurring of a single image. First, we designed a novel and highly scalable SIMO multi-scale baseline to address the performance bottleneck faced by partially multiscale networks. It takes a single image as an input and gradually generates a series of sharp images from the bottom up. Then, we propose a learnable wavelet transform node (LWN) for image deblurring, and it enhances the proposed algorithm's ability to restore detail information." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.846, + 0.269, + 0.861 + ], + "angle": 0, + "content": "3.1. Multi-scale Baseline" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.871, + 0.471, + 0.901 + ], + "angle": 0, + "content": "Our multi-scale network maintains an encoder-decoder architecture, but retains the original image with the highest" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.413, + 0.893, + 0.473 + ], + "angle": 0, + "content": "resolution as input, and restores sharp images of various scales sequentially in the output stage. In this way, we eliminate the complexity of the fusion module when dealing RGB images of different scales and with deep semantics." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.475, + 0.893, + 0.672 + ], + "angle": 0, + "content": "For ease of understanding, we will introduce here the Encoder phase, multi-scale semantic Fusion phase (hereafter referred to as Fusion phase), and Decoder phase sequentially. As shown in Fig. 2, the Encoder stage consists of several gray Simple Encoder Blocks(SEB), where information flows top-down for feature extraction; the Fusion stage consists of several green Wavelet Fusion Blocks(WFB), where information flows bottom-up for semantic fusion at different scales; and the Decoder stage consists of several purple Wavelet Head Blocks(WHB), and the bottom-up flow of information enables gradual restoration. Compared with WHB, SEB replaces LWN with a \\(1 \\times 1\\) conv for channel scaling, and a \\(3 \\times 3\\) depth-conv for feature extraction." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.673, + 0.892, + 0.749 + ], + "angle": 0, + "content": "The Encoder stage is responsible for full feature extraction, downsampling feature maps once after each block, and then passing the feature maps to the Fusion stage or the Decoder stage as needed, whose outputs \\( E_{out}^{i} \\) to block of the \\( i - th \\) layer can be denoted as:" + }, + { + "type": "equation", + "bbox": [ + 0.555, + 0.763, + 0.892, + 0.798 + ], + "angle": 0, + "content": "\\[\nE _ {o u t} ^ {i} = \\left\\{ \\begin{array}{l l} \\phi_ {i} (e m b e d (x)), & i = i _ {m a x} \\\\ \\phi_ {i} (E _ {o u t} ^ {i - 1}), & o t h e r w i s e \\end{array} \\right. \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.81, + 0.893, + 0.903 + ], + "angle": 0, + "content": "where \\(\\phi_{i}\\) represent the block of corresponding layer of Encoder. The Fusion stage adapts and fuses information from different scales of semantics produced in Encoder to generate intermediate representations with deep semantics and shallow details. The Fusion stage's outputs \\(F_{out}^{i}\\) to the block of \\(i - th\\) layer is expressed as in Eq. 2, where \\(\\delta_{i}\\)" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.926, + 0.516, + 0.938 + ], + "angle": 0, + "content": "2735" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.461, + 0.108 + ], + "angle": 0, + "content": "represents the block of the corresponding layer of Fusion." + }, + { + "type": "equation", + "bbox": [ + 0.123, + 0.115, + 0.47, + 0.151 + ], + "angle": 0, + "content": "\\[\nF _ {o u t} ^ {i} = \\left\\{ \\begin{array}{l l} \\delta_ {i} \\left(E _ {o u t} ^ {i} + E _ {o u t} ^ {i - 1}\\right), & i = i _ {\\text {m i n}} \\\\ \\delta_ {i} \\left(F _ {o u t} ^ {i - 1} + E _ {o u t} ^ {i}\\right), & \\text {o t h e r w i s e} \\end{array} \\right. \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.159, + 0.471, + 0.249 + ], + "angle": 0, + "content": "The Decoder stage utilizes the information passed by both the Encoder and Fusion stages to progressively upsample the feature maps and generating pre-output feature maps at each scale separately. Output \\( D_{out}^{i} \\) to the block of the \\( i - th \\) layer is denoted as in Eq. 3. Here \\( \\xi_{i} \\) represents the block of the corresponding layer of the Decoder." + }, + { + "type": "equation", + "bbox": [ + 0.078, + 0.259, + 0.462, + 0.316 + ], + "angle": 0, + "content": "\\[\nD _ {o u t} ^ {i} = \\left\\{ \\begin{array}{l l} \\xi_ {i} \\left(E _ {o u t} ^ {i}\\right), & i = i _ {m i n} \\\\ \\xi_ {i} \\left(D _ {o u t} ^ {i - 1} + F _ {o u t} ^ {i}\\right), & i > i _ {m i n} \\text {a n d} i < i _ {m a x} \\\\ \\xi_ {i} \\left(D _ {o u t} ^ {i - 1} + F _ {o u t} ^ {i - 1} + E _ {o u t} ^ {i}\\right), & i = i _ {m a x} \\end{array} \\right.\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.324, + 0.47, + 0.43 + ], + "angle": 0, + "content": "Each block in the Decoder is followed by a \\(3 \\times 3\\) convolution to generate a restored image at each scale. To improve inference efficiency, all but the highest-scale outputs are generated in the training phase to compute the multiscale loss. We note that this design also have a similar role with auxiliary head [36] in facilitating the model's learning of the recovered uniform patterns." + }, + { + "type": "image", + "bbox": [ + 0.088, + 0.44, + 0.462, + 0.707 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.715, + 0.468, + 0.743 + ], + "angle": 0, + "content": "Figure 3. (a) The process of learnable 2D-wavelet convolution. (b) The construction process of the \\( N \\times N \\) 2D-wavelet kernel." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.757, + 0.413, + 0.772 + ], + "angle": 0, + "content": "3.2. Learnable Discrete Wavelet Transform" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.781, + 0.469, + 0.901 + ], + "angle": 0, + "content": "To alleviate the defects of multi-scale architectures in the recovery of detail information, we design LWN and constructed the WFB and WHB based on it. Unlike the Fourier transform, the wavelet transform has adaptive time-frequency resolution, and performs excellently on digital images rich in mutation signal. Its use in conjunction with DNNs has gained increasing widespread attention [7, 16, 18, 34]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.893, + 0.173 + ], + "angle": 0, + "content": "For ease of understanding, the 1D-DWT case is given here first. For the input discrete signal \\( t \\), given the wavelet function \\( \\psi_{j,k}(t) = 2^{\\frac{j}{2}}\\psi (2^{j}t - k) \\) for scaling factor \\( j \\), time factor \\( k \\), and scale function \\( \\phi_{j,k}(t) = 2^{\\frac{j}{2}}\\phi (2^{j}t - k) \\), the decomposition of the input signal \\( t \\) at \\( j_0 \\) is given by:" + }, + { + "type": "equation", + "bbox": [ + 0.528, + 0.183, + 0.893, + 0.217 + ], + "angle": 0, + "content": "\\[\nf (t) = \\sum_ {j > j _ {0}} \\sum_ {k} d _ {j, k} \\psi_ {j, k} (t) + \\sum_ {k} c _ {j _ {0}, k} \\phi_ {j _ {0}, k} (t) \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.229, + 0.895, + 0.381 + ], + "angle": 0, + "content": "In this process, \\(d_{j,k} = \\langle f(t),\\psi_{j,k}(t)\\rangle\\) represents the detail coefficients (i.e., high-frequency components), \\(c_{j0,k} = \\langle f(t),\\phi_{j0,k}(t)\\rangle\\) represents the approximation coefficients (i.e., low-frequency components) so that the input signal can be disassembled step by step into a rich wavelet domain signal. In order to use wavelet transform in DNN, we introduce the analysis vectors \\(\\vec{a}_0[k] = \\left\\langle \\frac{1}{\\sqrt{2}}\\phi \\left(\\frac{t}{2}\\right),\\phi (t - k)\\right\\rangle ,\\) \\(\\vec{a}_{1}[k] = \\left\\langle \\frac{1}{\\sqrt{2}}\\psi \\left(\\frac{t}{2}\\right),\\phi (t - k)\\right\\rangle\\) to represent the high and low frequency filters respectively, then:" + }, + { + "type": "equation", + "bbox": [ + 0.6, + 0.392, + 0.791, + 0.42 + ], + "angle": 0, + "content": "\\[\nc _ {j + 1, p} = \\sum_ {k} \\vec {a} _ {0} [ k - 2 p ] c _ {j, k}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.602, + 0.418, + 0.891, + 0.458 + ], + "angle": 0, + "content": "\\[\nd _ {j + 1, p} = \\sum_ {k} ^ {\\kappa} \\vec {a} _ {1} [ k - 2 p ] c _ {j, k} \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.47, + 0.893, + 0.651 + ], + "angle": 0, + "content": "According to Eq. 4, the decomposition of the original signal over the wavelet basis can be viewed as a recursive convolution of the original signal with a specific filter using a step size of 2. Similarly, the inverse wavelet transform can be realized by transposed convolution of two synthetic vectors \\(\\vec{s}_0\\) and \\(\\vec{s}_1\\). We extend the above situation to 2D by efficiently extracting axial high-frequency information using 2D discrete wavelet transform (2D-DWT), and designing learnable methods to make it adaptive to the data distribution and feature layers. Fig. 3(b) presents the construction process of the forward wavelet convolution kernel in 2D, which can be expressed as:" + }, + { + "type": "equation", + "bbox": [ + 0.582, + 0.662, + 0.809, + 0.68 + ], + "angle": 0, + "content": "\\[\n\\mathcal {F} _ {l l} = \\vec {a _ {0}} \\times \\vec {a _ {0}} ^ {T}, \\mathcal {F} _ {l h} = \\vec {a _ {0}} \\times \\vec {a _ {1}} ^ {T}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.581, + 0.684, + 0.891, + 0.702 + ], + "angle": 0, + "content": "\\[\n\\mathcal {F} _ {h l} = \\vec {a _ {1}} \\times \\vec {a _ {0}} ^ {T}, \\mathcal {F} _ {h h} = \\vec {a _ {1}} \\times \\vec {a _ {1}} ^ {T} \\tag {5}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.582, + 0.706, + 0.785, + 0.722 + ], + "angle": 0, + "content": "\\[\n\\mathcal {K} _ {w} = \\operatorname {c a t} \\left(\\mathcal {F} _ {l l}, \\mathcal {F} _ {l h}, \\mathcal {F} _ {h l}, \\mathcal {F} _ {h h}\\right)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.734, + 0.893, + 0.81 + ], + "angle": 0, + "content": "We set \\(\\vec{a_0}\\) and \\(\\vec{a_1}\\) as learnable filters; \\(\\mathcal{F}_{ll}\\), \\(\\mathcal{F}_{lh}\\), \\(\\mathcal{F}_{hl}\\), and \\(\\mathcal{F}_{hh}\\) are low-frequency, horizontal high-frequency, vertical high-frequency, and diagonal high-frequency convolution operators obtained from vector outer products; The wavelet kernel \\(\\kappa_w\\) is spliced from above convolution operators." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.811, + 0.894, + 0.901 + ], + "angle": 0, + "content": "Then we implement learnable wavelet transform as a group convolution like Fig. 3(a), given a set of input feature maps \\(\\mathcal{X}_{in} \\in (C, H, W)\\), its projection in wavelet domain \\(\\mathcal{X}_{out} \\in \\left(4C, \\frac{H}{2}, \\frac{W}{2}\\right)\\) (low-frequency, horizontal high-frequency, vertical high-frequency, and diagonal high-frequency component) is generated through the wavelet" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.925, + 0.516, + 0.937 + ], + "angle": 0, + "content": "2736" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.137 + ], + "angle": 0, + "content": "convolution. Similarly, the construction method of the inverse kernel and the forward propagation can be easily deduced, since they are inverse processes of each other." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.137, + 0.47, + 0.197 + ], + "angle": 0, + "content": "Then, a natural question can then be posed: how to ensure the correctness of the wavelet kernel learning and ensure that it does not degrade into a general group convolution and cause performance degradation?" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.198, + 0.469, + 0.304 + ], + "angle": 0, + "content": "A common practice is to introduce the principle of perfect reconstruction [16, 34] to constrain adaptive wavelets, for a complex \\( z \\in \\mathbb{C} \\), given a filter \\( \\vec{x} \\) that applies to the \\( z \\)-transform, its \\( z \\)-transform can be expressed as \\( X(z) = \\sum_{n \\in \\mathbb{Z}} \\vec{x}(n) z^{-n} \\). Then \\( \\vec{a}_0, \\vec{a}_1, \\vec{s}_0, \\vec{s}_1 \\) can be obtained as their corresponding \\( z \\)-transforms \\( A_0, A_1, S_0, S_1 \\), which must then be satisfied if perfect reconstruction is desired:" + }, + { + "type": "equation", + "bbox": [ + 0.154, + 0.316, + 0.47, + 0.333 + ], + "angle": 0, + "content": "\\[\nA _ {0} (- z) S _ {0} (z) + A _ {1} (- z) S _ {1} (z) = 0 \\tag {6}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.167, + 0.359, + 0.469, + 0.376 + ], + "angle": 0, + "content": "\\[\nA _ {0} (z) S _ {0} (z) + A _ {1} (z) S _ {1} (z) = 2 \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.383, + 0.469, + 0.444 + ], + "angle": 0, + "content": "Eq. 6 is known as aliasing cancellation condition, which is used to the cancellation of the aliasing effects arising from the downsampling. We will give a detailed derivation of Eq. 6 and Eq. 7 in the Appendix." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.444, + 0.469, + 0.671 + ], + "angle": 0, + "content": "After wavelet convolution, to achieve frequency restoration in the wavelet domain, the input \\(\\mathcal{X}\\) wavelet domain component was separated into a separate dim, we use depth-wise convolution with an expansion factor of \\(r\\) for wavelet domain feature extraction and transformation after the wavelet positive transform, and \\(1\\times 1\\) convolution for channel expansion and scaling, and finally a learnable wavelet inverse transform is used to reduce the wavelet domain feature maps to the spatial domain for output, which constitutes the LWN. We follow the rules of [2] and design structurally similar Wavelet Head Block (WHB) and Wavelet Fusion Block (WFB), both with LWN as an important base module, but used for semantic aggregation and multiscale output at different scales, respectively; thus, WHB adds the recovery convolution branch after WFB." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.68, + 0.15, + 0.695 + ], + "angle": 0, + "content": "3.3. Loss" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.703, + 0.469, + 0.764 + ], + "angle": 0, + "content": "Multi-scale Loss. Follow PSNR loss [2], we propose multiscale loss function as the main loss of the algorithm for calculating the pixel difference between restored image and GT at each scale:" + }, + { + "type": "equation", + "bbox": [ + 0.137, + 0.773, + 0.469, + 0.815 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {m u l t i}} (x, y) = \\sum_ {k = 1} ^ {K} w _ {k} \\times \\mathcal {L} _ {\\text {p s n r}} \\left(x _ {k}, y _ {k}\\right) \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.469, + 0.902 + ], + "angle": 0, + "content": "where \\( x, y \\) denote the output and GT, respectively, \\( k \\) denotes the downsampling level, \\( w_{k} \\) represents the weights under the corresponding scale. Our design of \\( w_{k} \\) is based on the simple finding that we cannot produce absolutely accurate downsampled clear images, and that the accuracy gap" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.152 + ], + "angle": 0, + "content": "is progressively enlarged with decreasing scales. To ensure that lower scale output layers do not negatively affect the final output, we empirically set the loss weight \\( w_{k} \\) to \\( \\frac{1}{k} \\) for the corresponding k-scale." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.152, + 0.892, + 0.243 + ], + "angle": 0, + "content": "Wavelet Loss. Both Eq. 6 and Eq. 7 can be easily converted losses optimized toward a minimum of 0, here we use the mean square error. Then, based on the convolvable nature of the \\( z \\) transform, it is possible to construct convolutionally equivalent substitutions for polynomial multiplications for the perfect reconstruction condition:" + }, + { + "type": "equation", + "bbox": [ + 0.5, + 0.253, + 0.926, + 0.357 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} L _ {w a v e l e t} (\\theta_ {i}) = \\left(\\sum_ {k} ^ {N - 1} \\left(\\langle a _ {0}, s _ {0} \\rangle_ {k} + \\langle a _ {1}, s _ {1} \\rangle_ {k}\\right) - \\hat {V} _ {\\lfloor \\frac {N}{2} \\rfloor}\\right) ^ {2} \\\\ = \\left(\\sum_ {k} ^ {N - 1} \\left(\\left\\langle (- 1) ^ {k} a _ {0}, s _ {0} \\right\\rangle_ {k} + \\left\\langle (- 1) ^ {k} a _ {1}, s _ {1} \\right\\rangle_ {k}\\right)\\right) ^ {2} \\tag {9} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.358, + 0.892, + 0.402 + ], + "angle": 0, + "content": "where \\(k\\) denotes the position of the filter when it performs the convolution, \\(\\hat{V}\\) is a vector with a center position value of two, and \\(\\theta_{i}\\) denotes the constructed filter" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.403, + 0.892, + 0.433 + ], + "angle": 0, + "content": "Overall Loss. Based on the multi-scale loss and wavelet loss, the loss function used in this study is:" + }, + { + "type": "equation", + "bbox": [ + 0.537, + 0.445, + 0.892, + 0.461 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {t o t a l}} (x, y) = \\mathcal {L} _ {\\text {w a v e l e t}} \\left(\\theta_ {i}\\right) + \\mathcal {L} _ {\\text {m u l t i}} (x, y) \\tag {10}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.473, + 0.706, + 0.489 + ], + "angle": 0, + "content": "4. Experimental Results" + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.498, + 0.819, + 0.514 + ], + "angle": 0, + "content": "4.1. Datasets and Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.521, + 0.892, + 0.673 + ], + "angle": 0, + "content": "We evaluate our method on the realistic datasets Real-Blur [23] and RSBur [24], as well as the synthetic dataset GoPro [20], they are all composed of blurry-sharp image pairs. During training we use AdamW optimizer \\((\\alpha = 0.9\\) and \\(\\beta = 0.9)\\) for a total of \\(600\\mathrm{K}\\) iterations. The initial value of the learning rate is \\(10^{-3}\\) and is updated with cosine annealing schedule. The patch size is set to \\(256\\times 256\\) pixels and we followed the training strategy of [28, 38]. For data augmentation, we only use random flips and rotations. More experimental results can be found in the Appendix." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.682, + 0.868, + 0.698 + ], + "angle": 0, + "content": "4.2. Comparison with State-of-the-Art Methods" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.705, + 0.891, + 0.75 + ], + "angle": 0, + "content": "We have compared our method with several advanced deblurring methods and use the PSNR and SSIM to evaluate the quality of restored images." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.892, + 0.901 + ], + "angle": 0, + "content": "**Evaluations on the RealBlur dataset.** Since our proposed method is driven by real-world deblurring, we first conduct comparisons on RealBlur [23]. The quantitative analysis results are shown in Tab. 1, the PSNR value of our method is 0.91dB and 0.49dB higher than state-of-the-art GRL on the RealBlur-J and RealBlur-R datasets respectively. Compared to other multi-scale architectures, our method has fewer model computational and running time while the performance is better. Fig. 4 shows visual comparison on RealBlur-J dataset, our method obtains restored text with" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.926, + 0.516, + 0.938 + ], + "angle": 0, + "content": "2737" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.081, + 0.089, + 0.311, + 0.183 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.155, + 0.185, + 0.232, + 0.197 + ], + "angle": 0, + "content": "Blurry image" + }, + { + "type": "image", + "bbox": [ + 0.313, + 0.089, + 0.426, + 0.183 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.326, + 0.185, + 0.416, + 0.196 + ], + "angle": 0, + "content": "DeblurGAN-v2" + }, + { + "type": "image", + "bbox": [ + 0.429, + 0.089, + 0.544, + 0.183 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.472, + 0.185, + 0.501, + 0.196 + ], + "angle": 0, + "content": "SRN" + }, + { + "type": "image", + "bbox": [ + 0.546, + 0.089, + 0.659, + 0.183 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.562, + 0.185, + 0.646, + 0.196 + ], + "angle": 0, + "content": "MIMO-UNet+" + }, + { + "type": "image", + "bbox": [ + 0.662, + 0.089, + 0.775, + 0.183 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.696, + 0.185, + 0.747, + 0.196 + ], + "angle": 0, + "content": "MPRNet" + }, + { + "type": "image", + "bbox": [ + 0.778, + 0.089, + 0.892, + 0.183 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.816, + 0.185, + 0.857, + 0.196 + ], + "angle": 0, + "content": "BANet" + }, + { + "type": "image", + "bbox": [ + 0.081, + 0.2, + 0.195, + 0.293 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.117, + 0.295, + 0.157, + 0.307 + ], + "angle": 0, + "content": "Blurry" + }, + { + "type": "image", + "bbox": [ + 0.197, + 0.2, + 0.311, + 0.293 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.243, + 0.295, + 0.263, + 0.306 + ], + "angle": 0, + "content": "GT" + }, + { + "type": "image", + "bbox": [ + 0.313, + 0.2, + 0.426, + 0.293 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.349, + 0.295, + 0.399, + 0.306 + ], + "angle": 0, + "content": "MSSNet" + }, + { + "type": "image", + "bbox": [ + 0.429, + 0.2, + 0.543, + 0.293 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.451, + 0.295, + 0.519, + 0.308 + ], + "angle": 0, + "content": "Stripformer" + }, + { + "type": "image", + "bbox": [ + 0.545, + 0.2, + 0.658, + 0.293 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.589, + 0.295, + 0.618, + 0.306 + ], + "angle": 0, + "content": "GRL" + }, + { + "type": "image", + "bbox": [ + 0.661, + 0.2, + 0.775, + 0.293 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.687, + 0.295, + 0.752, + 0.306 + ], + "angle": 0, + "content": "FFTformer" + }, + { + "type": "image", + "bbox": [ + 0.778, + 0.2, + 0.892, + 0.293 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.822, + 0.295, + 0.853, + 0.306 + ], + "angle": 0, + "content": "Ours" + }, + { + "type": "image_caption", + "bbox": [ + 0.11, + 0.309, + 0.858, + 0.323 + ], + "angle": 0, + "content": "Figure 4. Visual comparisons on the RealBlur-J dataset [23]. The proposed method generates an image with clearer characters." + }, + { + "type": "table", + "bbox": [ + 0.08, + 0.337, + 0.471, + 0.646 + ], + "angle": 0, + "content": "
MethodRealBlur-JRealBlur-RAvg. runtime
PSNRSSIMPSNRSSIM
DeblurGAN-v2 [12]29.690.87036.440.9350.04s
SRN [27]31.380.90938.650.9650.07s
MPRNet [37]31.760.92239.310.9720.09s
SDWNet [42]30.730.89638.210.9630.04s
MIMO-UNet+ [3]31.920.919--0.02s
MIMO-UNet++ [3]32.050.921---
DeepRFT+ [19]32.190.93139.840.9720.09s
BANet [29]32.000.92339.550.9710.06s
BANet+ [29]32.420.92939.900.9720.12s
Stripformer [28]32.480.92939.840.9740.04s
MSSNet [9]32.100.92839.760.9720.06s
MSDI-Net [13]32.350.923--0.06s
MAXIM-3S [30]32.840.93539.450.962-
FFTformer [10]32.620.93340.110.9730.13s
GRL-B [14]32.820.93240.200.9741.28s
MLWNet-S33.020.933--0.04s
MLWNet-B33.840.94140.690.9760.05s
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.076, + 0.648, + 0.47, + 0.69 + ], + "angle": 0, + "content": "Table 1. Quantitative evaluations on the RealBlur dataset [23]. The experimental results were trained under the corresponding datasets respectively, and average runtime is tested on \\(256 \\times 256\\) patches." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.701, + 0.469, + 0.731 + ], + "angle": 0, + "content": "richer edge details, while achieving the best contrast, sharpness, brightness, and structural details of the image." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.735, + 0.47, + 0.901 + ], + "angle": 0, + "content": "Evaluations on the RSBlur dataset. We further conduct experiments on the latest real-world deblurring dataset RSBlur [24] and follow the protocol of this dataset for fair comparison. Tab. 2 summarizes the quantitative evaluation results of our method with advanced algorithms. The proposed MLWNet achieved the highest PSNR and SSIM, which were 34.94 and 0.880 respectively. In addition, the model we trained on RSBlur dataset also achieved the best results on RealBlur-J dataset. We show some visual comparisons in Fig. 5. We note that our method outperforms other methods in low-light blurry scenes, proving that the" + }, + { + "type": "table", + "bbox": [ + 0.504, + 0.337, + 0.89, + 0.51 + ], + "angle": 0, + "content": "
MethodRSBlurRealBlur-J
PSNRSSIMPSNRSSIM
SRN [27]32.530.84029.860.886
MIMO-Unet [3]32.730.84629.530.876
MIMO-Unet+ [3]33.370.85629.990.889
MPRNet [37]33.610.86130.460.899
Restormer [38]33.690.86330.480.891
Uformer-B [33]33.980.86630.370.899
SFNet [4]34.350.87230.260.897
MLWNet-B34.940.88030.530.905
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.5, + 0.512, + 0.892, + 0.54 + ], + "angle": 0, + "content": "Table 2. Quantitative evaluations trained on the RSBlur dataset [24], the RealBlur-J dataset was used for testing only." + }, + { + "type": "table", + "bbox": [ + 0.507, + 0.55, + 0.888, + 0.723 + ], + "angle": 0, + "content": "
Method[23]→ [24]MACs(G)
PSNRSSIM
DeblurGAN-v2 [12]30.150.76642.0
MPRNet [37]29.560.785760.8
MIMO-UNet+ [3]29.690.792154.4
BANet [29]30.190.806263.9
BANet+ [29]30.240.809588.7
MSSNet [9]29.860.806154.0
FFTformer [10]29.700.787131.8
MLWNet-B30.910.818108.2
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.5, + 0.726, + 0.892, + 0.767 + ], + "angle": 0, + "content": "Table 3. Quantitative evaluation for generalizability shows the results of models trained on the RealBlur-J dataset and tested on the RSBlur dataset, MACs are measured on \\(256 \\times 256\\) patches." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.778, + 0.865, + 0.792 + ], + "angle": 0, + "content": "proposed method makes textures and structures clearer." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.794, + 0.892, + 0.884 + ], + "angle": 0, + "content": "In addition, we evaluate the RSBlur dataset using models trained only on RealBlur-J to fairly compare the generalization of methods to real scenarios. As shown in Tab. 3, our method produced results with highest PSNR value of 30.91. The results of Tab. 2 and Tab. 3 show that our model has a better generalization ability." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.886, + 0.891, + 0.9 + ], + "angle": 0, + "content": "Evaluations on the GoPro dataset. We conduct a exten" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.927, + 0.516, + 0.938 + ], + "angle": 0, + "content": "2738" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.091, + 0.349, + 0.189 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.177, + 0.19, + 0.253, + 0.201 + ], + "angle": 0, + "content": "Blurry image" + }, + { + "type": "image", + "bbox": [ + 0.352, + 0.09, + 0.484, + 0.189 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.405, + 0.19, + 0.432, + 0.199 + ], + "angle": 0, + "content": "SRN" + }, + { + "type": "image", + "bbox": [ + 0.487, + 0.091, + 0.619, + 0.189 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.516, + 0.19, + 0.591, + 0.199 + ], + "angle": 0, + "content": "MIMO-UNet" + }, + { + "type": "image", + "bbox": [ + 0.622, + 0.091, + 0.753, + 0.189 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.651, + 0.19, + 0.733, + 0.199 + ], + "angle": 0, + "content": "MIMO-UNet+" + }, + { + "type": "image", + "bbox": [ + 0.757, + 0.091, + 0.89, + 0.189 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.803, + 0.19, + 0.853, + 0.2 + ], + "angle": 0, + "content": "MPRNet" + }, + { + "type": "image", + "bbox": [ + 0.081, + 0.202, + 0.213, + 0.3 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.126, + 0.302, + 0.164, + 0.313 + ], + "angle": 0, + "content": "Blurry" + }, + { + "type": "image", + "bbox": [ + 0.217, + 0.202, + 0.349, + 0.3 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.273, + 0.302, + 0.292, + 0.312 + ], + "angle": 0, + "content": "GT" + }, + { + "type": "image", + "bbox": [ + 0.352, + 0.202, + 0.484, + 0.3 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.394, + 0.302, + 0.453, + 0.312 + ], + "angle": 0, + "content": "Restormer" + }, + { + "type": "image", + "bbox": [ + 0.487, + 0.202, + 0.619, + 0.3 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.528, + 0.302, + 0.59, + 0.312 + ], + "angle": 0, + "content": "Uformer-B" + }, + { + "type": "image", + "bbox": [ + 0.623, + 0.202, + 0.755, + 0.3 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.674, + 0.302, + 0.71, + 0.312 + ], + "angle": 0, + "content": "SFNet" + }, + { + "type": "image", + "bbox": [ + 0.758, + 0.202, + 0.89, + 0.3 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.804, + 0.302, + 0.833, + 0.312 + ], + "angle": 0, + "content": "Ours" + }, + { + "type": "image_caption", + "bbox": [ + 0.078, + 0.315, + 0.89, + 0.34 + ], + "angle": 0, + "content": "Figure 5. Visual comparisons on the RSBur dataset [24]. The deblurring performance of the proposed method in low-light is impressive, The recovery of characters and texture structures far exceeds other advanced methods." + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.342, + 0.309, + 0.422 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.155, + 0.424, + 0.232, + 0.436 + ], + "angle": 0, + "content": "Blurry image" + }, + { + "type": "image", + "bbox": [ + 0.314, + 0.342, + 0.425, + 0.421 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.331, + 0.424, + 0.408, + 0.435 + ], + "angle": 0, + "content": "MIMO-UNet" + }, + { + "type": "image", + "bbox": [ + 0.429, + 0.342, + 0.543, + 0.421 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.444, + 0.424, + 0.529, + 0.435 + ], + "angle": 0, + "content": "MIMO-UNet+" + }, + { + "type": "image", + "bbox": [ + 0.546, + 0.342, + 0.657, + 0.421 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.577, + 0.424, + 0.631, + 0.435 + ], + "angle": 0, + "content": "MPRNet" + }, + { + "type": "image", + "bbox": [ + 0.661, + 0.342, + 0.774, + 0.421 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.693, + 0.424, + 0.744, + 0.435 + ], + "angle": 0, + "content": "NAFNet" + }, + { + "type": "image", + "bbox": [ + 0.777, + 0.342, + 0.889, + 0.421 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.807, + 0.424, + 0.859, + 0.435 + ], + "angle": 0, + "content": "MAXIM" + }, + { + "type": "image", + "bbox": [ + 0.081, + 0.439, + 0.194, + 0.52 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.117, + 0.522, + 0.158, + 0.535 + ], + "angle": 0, + "content": "Blurry" + }, + { + "type": "image", + "bbox": [ + 0.197, + 0.439, + 0.31, + 0.52 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.245, + 0.522, + 0.264, + 0.532 + ], + "angle": 0, + "content": "GT" + }, + { + "type": "image", + "bbox": [ + 0.315, + 0.439, + 0.425, + 0.52 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.336, + 0.522, + 0.403, + 0.534 + ], + "angle": 0, + "content": "Stripformer" + }, + { + "type": "image", + "bbox": [ + 0.429, + 0.439, + 0.543, + 0.52 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.456, + 0.522, + 0.516, + 0.533 + ], + "angle": 0, + "content": "Restormer" + }, + { + "type": "image", + "bbox": [ + 0.546, + 0.439, + 0.658, + 0.52 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.572, + 0.522, + 0.637, + 0.532 + ], + "angle": 0, + "content": "FFTformer" + }, + { + "type": "image", + "bbox": [ + 0.662, + 0.439, + 0.774, + 0.52 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.704, + 0.522, + 0.735, + 0.532 + ], + "angle": 0, + "content": "GRL" + }, + { + "type": "image", + "bbox": [ + 0.778, + 0.439, + 0.89, + 0.52 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.822, + 0.522, + 0.852, + 0.532 + ], + "angle": 0, + "content": "Ours" + }, + { + "type": "image_caption", + "bbox": [ + 0.113, + 0.536, + 0.855, + 0.55 + ], + "angle": 0, + "content": "Figure 6. Visual comparisons on the GoPro dataset [20]. Our method better preserves texture information without sharpening." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.56, + 0.469, + 0.697 + ], + "angle": 0, + "content": "sive comparison with advanced algorithms on GoPro [20]. Tab. 4 shows the quantitative evaluation results. The gap between our method and the state-of-the-art FFTformer is only 0.001 in SSIM value, while the running time is reduced by \\(60\\%\\). We show a visual comparison on the GoPro dataset in Fig. 6. We can see that our method recovers blurred textures well without sharpening. For the reasons why the proposed MLWNet does not achieve optimal performance in synthetic blurry, we conducted a detailed analysis in Sec. 5." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.71, + 0.302, + 0.726 + ], + "angle": 0, + "content": "5. Analysis and Discussion" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.735, + 0.469, + 0.824 + ], + "angle": 0, + "content": "In this section, we provide a more in-depth analysis and show the contribution of each component of the proposed method. We perform 20w iterations on the RealBlur-J dataset for abtation with a batch size of 8 to train our method using a version of the model with a width of 32, and the results are shown in Tab. 6." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.826, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Single-scale vs. Multi-scale. Given that each node of the baseline network is composed of SEB, we have deformed the multi-scale and single-scale according to the architecture. Tab. 5 shows that the multi-scale architecture improves the PSNR and SSIM values to varying degrees, proving that" + }, + { + "type": "table", + "bbox": [ + 0.503, + 0.558, + 0.892, + 0.837 + ], + "angle": 0, + "content": "
MethodGOPROAvg.rruntime
PSNRSSIM
DeblurGAN-v2 [12]29.550.9340.04s
SRN [27]30.260.9340.07s
DMPHN [39]31.200.9450.21s
SDWNet [42]31.260.9660.04s
MPRNet [37]32.660.9590.09s
MIMO-UNet+ [3]32.450.9570.02s
DeepRFT+ [19]33.230.9630.09s
MAXIM-3S [30]32.860.961-
Stripformer [28]33.080.9620.04s
MSDI-net [13]33.280.9640.06s
Restormer [38]33.570.9660.08s
NAFNet [2]33.690.9670.04s
FFTformer [10]34.210.9690.13s
GRL-B [14]33.930.9681.28s
MLWNet-B33.830.9680.05s
" + }, + { + "type": "table_caption", + "bbox": [ + 0.5, + 0.839, + 0.892, + 0.88 + ], + "angle": 0, + "content": "Table 4. Quantitative evaluations trained and tested on the GoPro dataset [20]. Our proposed MLWNet obtains competitive results with a combination of time efficiency and accuracy." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.927, + 0.516, + 0.938 + ], + "angle": 0, + "content": "2739" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.152 + ], + "angle": 0, + "content": "the SIMO strategy we employ helps to eliminate motion blur to some extent. As shown in Sec. 3.1, we use SIMO for coarse-to-fine restoration, and only use multiple outputs to calculate multi-scale losses during training." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.152, + 0.471, + 0.348 + ], + "angle": 0, + "content": "WFB and WHB. We default to using WFB and WHB via \\(\\mathcal{L}_{wavelet}\\) statute in this section, we also proved the effectiveness of LWN, because LWN is a subpart of the first two. After applying them, the PSNR index improved by \\(0.25\\mathrm{dB}\\) compared to the multi-scale baseline, which shows that LWN and its extension modules can easily help multiscale algorithms improve detail recovery capabilities. Fig. 7 shows the four types of feature maps generated after forward learnable wavelet convolution. We can see that the high and low frequency information of the input feature map are mixed together, and after wavelet group convolution, the network exerts different attention on the feature information in different directions or frequencies." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.348, + 0.47, + 0.468 + ], + "angle": 0, + "content": "Effectiveness of \\(\\mathcal{L}_{wavelet}\\). By adding unreduced WFB and WHB to the baseline without using wavelet loss, the performance improvement of the baseline is minimal and does not break the bottleneck of multiscale detail recovery, which suggests that the wavelet convolution degenerates into a general group convolution that only serves to deepen the network, proving that merely deepening the network is ineffective for the restoration of detail information." + }, + { + "type": "table", + "bbox": [ + 0.08, + 0.472, + 0.472, + 0.519 + ], + "angle": 0, + "content": "
MethodSISOMIMOSIMO
PSNR/SSIM32.29/0.92432.19/0.92832.37/0.929
MACs(G)19.2421.8319.29
" + }, + { + "type": "table_caption", + "bbox": [ + 0.105, + 0.521, + 0.44, + 0.535 + ], + "angle": 0, + "content": "Table 5. Comparison in various input and output modes." + }, + { + "type": "table", + "bbox": [ + 0.08, + 0.537, + 0.481, + 0.642 + ], + "angle": 0, + "content": "
SIMOWFBWHBLwaveletPSNRSSIMMACs(G)
✓32.370.92919.29
✓✓✓32.400.92828.21
✓✓✓32.490.92825.28
✓✓✓32.570.92922.22
✓✓✓✓32.620.93128.21
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.644, + 0.469, + 0.686 + ], + "angle": 0, + "content": "Table 6. Ablation study on components of the proposed MLWNet. We set the baseline network to use SEB in its entirety, and models that do not use SIMO will represent single scales using SISO." + }, + { + "type": "image", + "bbox": [ + 0.08, + 0.687, + 0.469, + 0.765 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.768, + 0.469, + 0.808 + ], + "angle": 0, + "content": "Figure 7. Feature maps representing high-and low-frequency components generated after learnable wavelet convolution. Zoom in on the screen for the best view." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.81, + 0.469, + 0.9 + ], + "angle": 0, + "content": "Limitations and analysis. It is observed from Tab. 1-3 that our method fails to achieve the same expected high performance with synthetic blur as it does with realistic blur. We try to analyze the reasons for this phenomenon via dataset comparison and experimental results. First, the average frame synthesis method used in the synthetic blur" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.198 + ], + "angle": 0, + "content": "dataset leads to unnatural discontinuous blur trajectories, thereby introducing strange high-frequency information interference (see Fig. 8). Second, synthetic data tends to produce an unnatural mixture of high and low frequencies. It is easy to mix with the average color area at the edge of the texture, causing the confusion of high-low frequency information." + }, + { + "type": "image", + "bbox": [ + 0.504, + 0.206, + 0.695, + 0.335 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.551, + 0.337, + 0.642, + 0.348 + ], + "angle": 0, + "content": "(a) Realistic Blur" + }, + { + "type": "image", + "bbox": [ + 0.699, + 0.206, + 0.891, + 0.335 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.741, + 0.337, + 0.836, + 0.348 + ], + "angle": 0, + "content": "(b) Synthetic Blur" + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.355, + 0.892, + 0.41 + ], + "angle": 0, + "content": "Figure 8. The difference between realistic blur (a) and synthetic blur (b). In the green box, the synthetic blur appears with color averaging resulting in high and low frequency confusion, and in the blue box has unnatural discontinuous trajectories." + }, + { + "type": "table", + "bbox": [ + 0.505, + 0.411, + 0.889, + 0.472 + ], + "angle": 0, + "content": "
Noise LevelL1L2L3L4L5
GoPro34.8134.6633.7633.1932.63
RealBlur-J33.9233.8133.9733.9333.54
" + }, + { + "type": "table_caption", + "bbox": [ + 0.498, + 0.474, + 0.89, + 0.5 + ], + "angle": 0, + "content": "Table 7. Performance comparison at different noise difference levels, where L3 contains the noise difference mean." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.502, + 0.892, + 0.653 + ], + "angle": 0, + "content": "Finally, there is a large difference in noise levels between synthetic blur and real blur. We follow [1] to calculate the noise level of each testset, and divided the noise difference between clear images and blurred images of GoPro and RealBlur-J into 5 levels for separate testing. As shown in Tab. 7, we note that as the noise difference increases, our method maintains good performance on RealBlur-J, while the performance drops fast on GoPro. This shows that our method is robust to the noise of real blur, and that there are differences in the noise between synthetic blur and real blur." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.665, + 0.619, + 0.68 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.689, + 0.892, + 0.901 + ], + "angle": 0, + "content": "In this paper, we propose a SIMO-based multi-scale architecture to achieve efficient motion deblurring. We have developed a learnable discrete wavelet transform module, which not only improves the algorithm's ability to recover details, but is also more applicable to the real world. In addition, we construct a reasonable multi-scale loss to guide the recovery of blurred images pixel by pixel and scale by scale, and constrain the learning direction of the wavelet kernel with self-supervised loss to achieve better image deblurring. Through extensive experiments on multiple real and synthetic datasets, we demonstrate that it outperforms existing state-of-the-art methods in terms of quality and efficiency of image restoration, especially with optimal deblurring performance and generalization in real scenes." + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.927, + 0.516, + 0.938 + ], + "angle": 0, + "content": "2740" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.105 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.171 + ], + "angle": 0, + "content": "[1] Guangyong Chen, Fengyuan Zhu, and Pheng Ann Heng. An efficient statistical method for image noise level estimation. In Proceedings of the IEEE International Conference on Computer Vision, pages 477-485, 2015. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.172, + 0.47, + 0.226 + ], + "angle": 0, + "content": "[2] Liangyu Chen, Xiaojie Chu, Xiangyu Zhang, and Jian Sun. Simple baselines for image restoration. In European Conference on Computer Vision, pages 17-33. Springer, 2022. 1, 2, 3, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.228, + 0.47, + 0.296 + ], + "angle": 0, + "content": "[3] Sung-Jin Cho, Seo-Won Ji, Jun-Pyo Hong, Seung-Won Jung, and Sung-Jea Ko. Rethinking coarse-to-fine approach in single image deblurring. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4641-4650, 2021. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.298, + 0.47, + 0.354 + ], + "angle": 0, + "content": "[4] Yuning Cui, Yi Tao, Zhenshan Bing, Wenqi Ren, Xinwei Gao, Xiaochun Cao, Kai Huang, and Alois Knoll. Selective frequency network for image restoration. In International Conference on Learning Representations, ICLR, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.355, + 0.469, + 0.396 + ], + "angle": 0, + "content": "[5] Ingrid Daubechies. The wavelet transform, time-frequency localization and signal analysis. IEEE transactions on information theory, 36(5):961-1005, 1990. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.398, + 0.469, + 0.464 + ], + "angle": 0, + "content": "[6] Golnaz Ghiasi, Tsung-Yi Lin, and Quoc V Le. Nas-fpn: Learning scalable feature pyramid architecture for object detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7036-7045, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.467, + 0.469, + 0.522 + ], + "angle": 0, + "content": "[7] Wooseok Ha, Chandan Singh, Francois Lanusse, Srigokul Upadhyayula, and Bin Yu. Adaptive wavelet distillation from neural networks through interpretations. Advances in Neural Information Processing Systems, 34:20669-20682, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.524, + 0.469, + 0.605 + ], + "angle": 0, + "content": "[8] Bumsoo Kim, Jonghwan Mun, Kyoung-Woon On, Minchul Shin, Junhyun Lee, and Eun-Sol Kim. Mstr: Multi-scale transformer for end-to-end human-object interaction detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19578-19587, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.607, + 0.469, + 0.662 + ], + "angle": 0, + "content": "[9] Kiyeon Kim, Seungyong Lee, and Sunghyun Cho. Mssnet: Multi-scale-stage network for single image deblurring. In European Conference on Computer Vision, pages 524-539. Springer, 2022. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.664, + 0.469, + 0.732 + ], + "angle": 0, + "content": "[10] Lingshun Kong, Jiangxin Dong, Jianjun Ge, Mingqiang Li, and Jinshan Pan. Efficient frequency domain-based transformers for high-quality image deblurring. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5886-5895, 2023. 1, 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.734, + 0.469, + 0.803 + ], + "angle": 0, + "content": "[11] Orest Kupyn, Volodymyr Budzan, Mykola Mykhailych, Dmytro Mishkin, and Jiří Matas. Deblurgan: Blind motion deblurring using conditional adversarial networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 8183-8192, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.804, + 0.469, + 0.871 + ], + "angle": 0, + "content": "[12] Orest Kupyn, Tetiana Martyniuk, Junru Wu, and Zhangyang Wang. Deblurgan-v2: Deblurring (orders-of-magnitude) faster and better. In Proceedings of the IEEE/CVF international conference on computer vision, pages 8878-8887, 2019. 1, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.469, + 0.902 + ], + "angle": 0, + "content": "[13] Dasong Li, Yi Zhang, Ka Chun Cheung, Xiaogang Wang, Hongwei Qin, and Hongsheng Li. Learning degradation rep" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.47, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.892, + 0.133 + ], + "angle": 0, + "content": "resentations for image deblurring. In European Conference on Computer Vision, pages 736-753. Springer, 2022. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.136, + 0.892, + 0.219 + ], + "angle": 0, + "content": "[14] Yawei Li, Yuchen Fan, Xiaoyu Xiang, Denis Demandolx, Rakesh Ranjan, Radu Timofte, and Luc Van Gool. Efficient and explicit modelling of image hierarchies for image restoration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18278-18289, 2023. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.221, + 0.892, + 0.29 + ], + "angle": 0, + "content": "[15] Tsung-Yi Lin, Piotr Dólar, Ross Girshick, Kaiming He, Bharath Hariharan, and Serge Belongie. Feature pyramid networks for object detection. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2117-2125, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.292, + 0.892, + 0.332 + ], + "angle": 0, + "content": "[16] Pengju Liu, Hongzhi Zhang, Wei Lian, and Wangmeng Zuo. Multi-level wavelet convolutional neural networks. IEEE Access, 7:74973-74985, 2019. 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.334, + 0.892, + 0.39 + ], + "angle": 0, + "content": "[17] Shu Liu, Lu Qi, Haifang Qin, Jianping Shi, and Jiaya Jia. Path aggregation network for instance segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 8759-8768, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.392, + 0.892, + 0.447 + ], + "angle": 0, + "content": "[18] Wei Liu, Qiong Yan, and Yuzhi Zhao. Densely self-guided wavelet network for image denoising. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, pages 432-433, 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.449, + 0.892, + 0.49 + ], + "angle": 0, + "content": "[19] X Mao, Y Liu, W Shen, Q Li, and Y Wang. Deep residual fourier transformation for single image deblurring. arxiv 2021. arXiv preprint arXiv:2111.11745, 2021. 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.492, + 0.892, + 0.56 + ], + "angle": 0, + "content": "[20] Seungjun Nah, Tae Hyun Kim, and Kyoung Mu Lee. Deep multi-scale convolutional neural network for dynamic scene deblurring. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3883-3891, 2017. 1, 2, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.563, + 0.892, + 0.631 + ], + "angle": 0, + "content": "[21] Dongwon Park, Dong Un Kang, Jisoo Kim, and Se Young Chun. Multi-temporal recurrent neural networks for progressive non-uniform single image deblurring with incremental temporal training. In European Conference on Computer Vision, pages 327-343. Springer, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.633, + 0.892, + 0.689 + ], + "angle": 0, + "content": "[22] Zequn Qin, Pengyi Zhang, Fei Wu, and Xi Li. Fcanet: Frequency channel attention networks. In Proceedings of the IEEE/CVF international conference on computer vision, pages 783-792, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.691, + 0.892, + 0.773 + ], + "angle": 0, + "content": "[23] Jaesung Rim, Haeyun Lee, Jucheol Won, and Sunghyun Cho. Real-world blur dataset for learning and benchmarking deblurring algorithms. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXV 16, pages 184-201. Springer, 2020. 1, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.775, + 0.892, + 0.831 + ], + "angle": 0, + "content": "[24] Jaesung Rim, Geonung Kim, Jungeon Kim, Junyong Lee, Seungyong Lee, and Sunghyun Cho. Realistic blur synthesis for learning image deblurring. In European conference on computer vision, pages 487-503. Springer, 2022. 2, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.832, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[25] Roman Suvorov, Elizaveta Logacheva, Anton Mashikhin, Anastasia Remizova, Arsenii Ashukha, Aleksei Silvestrov, Naejin Kong, Harshith Goka, Kiwoong Park, and Victor Lempitsky. Resolution-robust large mask inpainting with fourier convolutions. In Proceedings of the IEEE/CVF winter" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.927, + 0.514, + 0.938 + ], + "angle": 0, + "content": "2741" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.11, + 0.093, + 0.47, + 0.12 + ], + "angle": 0, + "content": "conference on applications of computer vision, pages 2149-2159, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.122, + 0.472, + 0.177 + ], + "angle": 0, + "content": "[26] Mingxing Tan, Ruoming Pang, and Quoc V Le. Efficientdet: Scalable and efficient object detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10781-10790, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.179, + 0.47, + 0.234 + ], + "angle": 0, + "content": "[27] Xin Tao, Hongyun Gao, Xiaoyong Shen, Jue Wang, and Jiaya Jia. Scale-recurrent network for deep image deblurring. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 8174-8182, 2018. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.236, + 0.47, + 0.291 + ], + "angle": 0, + "content": "[28] Fu-Jen Tsai, Yan-Tsung Peng, Yen-Yu Lin, Chung-Chi Tsai, and Chia-Wen Lin. Stripformer: Strip transformer for fast image deblurring. In European Conference on Computer Vision, pages 146-162. Springer, 2022. 1, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.293, + 0.47, + 0.347 + ], + "angle": 0, + "content": "[29] Fu-Jen Tsai, Yan-Tsung Peng, Chung-Chi Tsai, Yen-Yu Lin, and Chia-Wen Lin. Banet: a blur-aware attention network for dynamic scene deblurring. IEEE Transactions on Image Processing, 31:6789-6799, 2022. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.35, + 0.47, + 0.419 + ], + "angle": 0, + "content": "[30] Zhengzhong Tu, Hossein Talebi, Han Zhang, Feng Yang, Peyman Milanfar, Alan Bovik, and Yinxiao Li. Maxim: Multi-axis mlp for image processing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5769-5780, 2022. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.421, + 0.47, + 0.489 + ], + "angle": 0, + "content": "[31] Chien-Yao Wang, Alexey Bochkovskiy, and Hong-Yuan Mark Liao. Yolov7: Trainable bag-of-freebies sets new state-of-the-art for real-time object detectors. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7464-7475, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.491, + 0.469, + 0.533 + ], + "angle": 0, + "content": "[32] Luyuan Wang and Yankui Sun. Image classification using convolutional neural network with wavelet domain inputs. IET Image Processing, 16(8):2037-2048, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.535, + 0.469, + 0.603 + ], + "angle": 0, + "content": "[33] Zhendong Wang, Xiaodong Cun, Jianmin Bao, Wengang Zhou, Jianzhuang Liu, and Houqiang Li. Uformer: A general u-shaped transformer for image restoration. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 17683-17693, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.605, + 0.469, + 0.659 + ], + "angle": 0, + "content": "[34] Moritz Wolter and Jochen Garcke. Adaptive wavelet pooling for convolutional neural networks. In International Conference on Artificial Intelligence and Statistics, pages 1936-1944. PMLR, 2021. 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.662, + 0.47, + 0.716 + ], + "angle": 0, + "content": "[35] Yanchao Yang and Stefano Soatto. Fda: Fourier domain adaptation for semantic segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4085-4095, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.719, + 0.469, + 0.786 + ], + "angle": 0, + "content": "[36] Changqian Yu, Changxin Gao, Jingbo Wang, Gang Yu, Chunhua Shen, and Nong Sang. Bisenet v2: Bilateral network with guided aggregation for real-time semantic segmentation. International Journal of Computer Vision, 129: 3051-3068, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.789, + 0.469, + 0.87 + ], + "angle": 0, + "content": "[37] Syed Waqas Zamir, Aditya Arora, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, Ming-Hsuan Yang, and Ling Shao. Multi-stage progressive image restoration. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14821-14831, 2021. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.469, + 0.903 + ], + "angle": 0, + "content": "[38] Syed Waqas Zamir, Aditya Arora, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, and Ming-Hsuan Yang." + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.093, + 0.472, + 0.903 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.892, + 0.147 + ], + "angle": 0, + "content": "Restormer: Efficient transformer for high-resolution image restoration. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5728-5739, 2022. 2, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.15, + 0.892, + 0.218 + ], + "angle": 0, + "content": "[39] Hongguang Zhang, Yuchao Dai, Hongdong Li, and Piotr Koniusz. Deep stacked hierarchical multi-patch network for image deblurring. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5978-5986, 2019. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.221, + 0.892, + 0.288 + ], + "angle": 0, + "content": "[40] Kaihao Zhang, Wenhan Luo, Yiran Zhong, Lin Ma, Bjorn Stenger, Wei Liu, and Hongdong Li. Deblurring by realistic blurring. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2737-2746, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.291, + 0.892, + 0.358 + ], + "angle": 0, + "content": "[41] Yijie Zhong, Bo Li, Lv Tang, Senyun Kuang, Shuang Wu, and Shouhong Ding. Detecting camouflaged object in frequency domain. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4504-4513, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.361, + 0.892, + 0.43 + ], + "angle": 0, + "content": "[42] Wenbin Zou, Mingchao Jiang, Yunchen Zhang, Liang Chen, Zhiyong Lu, and Yi Wu. Sdwnet: A straight dilated network with wavelet transformation for image deblurring. In Proceedings of the IEEE/CVF international conference on computer vision, pages 1895-1904, 2021. 1, 2, 3, 6, 7" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.43 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.927, + 0.516, + 0.938 + ], + "angle": 0, + "content": "2742" + } + ] +] \ No newline at end of file diff --git a/2024/Efficient Multi-scale Network with Learnable Discrete Wavelet Transform for Blind Motion Deblurring/4af0ea1e-f85d-4e0b-ab29-04d268e0386c_origin.pdf b/2024/Efficient Multi-scale Network with Learnable Discrete Wavelet Transform for Blind Motion Deblurring/4af0ea1e-f85d-4e0b-ab29-04d268e0386c_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..d6be7020058bfca0b74b14fe8f0a5be45bd19499 --- /dev/null +++ b/2024/Efficient Multi-scale Network with Learnable Discrete Wavelet Transform for Blind Motion Deblurring/4af0ea1e-f85d-4e0b-ab29-04d268e0386c_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c72ff28eb4c7a39b2c5193dde71a02bf26c768bfced2b609e994dec21ca60477 +size 9898989 diff --git a/2024/Efficient Multi-scale Network with Learnable Discrete Wavelet Transform for Blind Motion Deblurring/full.md b/2024/Efficient Multi-scale Network with Learnable Discrete Wavelet Transform for Blind Motion Deblurring/full.md new file mode 100644 index 0000000000000000000000000000000000000000..929da0d43336d6276b32df1f341b485c8f4ccb07 --- /dev/null +++ b/2024/Efficient Multi-scale Network with Learnable Discrete Wavelet Transform for Blind Motion Deblurring/full.md @@ -0,0 +1,419 @@ +# Efficient Multi-scale Network with Learnable Discrete Wavelet Transform for Blind Motion Deblurring + +Xin Gao\*1,2 Tianheng Qiu\*2,3,4 Xinyu Zhang†2 Hanlin Bai\*1 Kang Liu\*1 Xuan Huang†4 Hu Wei\*4 Guoying Zhang†1 Huaping Liu\*2 *Equal Contribution † Corresponding Author + +$^{1}$ China University of Mining & Technology-Beijing $^{2}$ Tsinghua University $^{3}$ University of Science and Technology of China $^{4}$ Hefei Institutes of Physical Science, Chinese Academy of Sciences + +# Abstract + +Coarse-to-fine schemes are widely used in traditional single-image motion deblur; however, in the context of deep learning, existing multi-scale algorithms not only require the use of complex modules for feature fusion of low-scale RGB images and deep semantics, but also manually generate low-resolution pairs of images that do not have sufficient confidence. In this work, we propose a multi-scale network based on single-input and multiple-outputs(SIMO) for motion deblurring. This simplifies the complexity of algorithms based on a coarse-to-fine scheme. To alleviate restoration defects impacting detail information brought about by using a multi-scale architecture, we combine the characteristics of real-world blurring trajectories with a learnable wavelet transform module to focus on the directional continuity and frequency features of the step-by-step transitions between blurred images to sharp images. In conclusion, we propose a multi-scale network with a learnable discrete wavelet transform (MLWNet), which exhibits state-of-the-art performance on multiple real-world deblurred datasets, in terms of both subjective and objective quality as well as computational efficiency. Our code is available on https://github.com/thqiu0419/MLWNet. + +# 1. Introduction + +Most current top-performing single-image blind deblurring algorithms are based on DNNs, which can be structurally categorized into single-scale [2, 10, 12, 28, 29, 42] and multi-scale approaches [3, 9, 13, 27, 37]. Compared to single-scale methods, multi-scale methods are built on the idea of moving from coarse-to-fine, and thus they decompose the challenging single-image blind deblurring problem into easier-to-solve sub-problems to restore the blurred image step-by-step. + +Earlier DNNs exploiting the coarse-to-fine concept usually employ estimators of different scales to produce re + +![](images/00c418dd730d494de65b5951a619cb9246b6202b63a1494cf6284118d1f7e695.jpg) +Figure 1. Performance comparison on the RealBlur-J [23] test dataset in terms of PSNR and GMACs. Our proposed MLWNet achieves superiority in comparison with other state-of-the-arts. + +stored images gradually [20, 21, 27]. However, not only is the extracted semantic information not sufficiently representative, but the use of the same complexity at different scales may lead to redundant computations; More advanced multi-scale algorithms [3] use a global encoder-decoder, which significantly reduces the algorithm's runtime, multiple upsampling and downsampling leads to insufficient restoration of detailed information, and therefore such algorithms require the introduction of additional fusion modules to encode input images for fusion with deeper semantics. In addition, existing multi-scale algorithms use multiple-input and multiple-output (MIMO) architectures, which require the introduction of manually constructed downsampled image pairs, and usually employ simple interpolation algorithms to generate low-resolution images for the sake of efficiency, which is obviously not reliable. + +We note that the gradual insertion of images used by existing coarse-to-fine algorithms is redundant. Referring to numerous algorithms [6, 8, 15, 17, 26, 31] that are widely used in downstream tasks, the network starts with a single input and also extracts features at different scales. These + +features can be aggregated by simple feature fusion to ultimately obtain competitive results. That is, a DNN itself possesses the ability to learn effective features at different scales. This inspired us to design a multi-scale architecture that retains an original image as input and sequentially restores images at each scale in output stage. This is not only more theoretically sound, but also eliminates the time required to generate pairs of images of different resolutions, as well as the huge complexity increase brought about by fusion modules between RGB domain and deep features. + +The coarse-to-fine algorithm has another inherent defect. During progressive restoration, the solution of the upper-level problem takes the solution of the lower-level as initialization [9]. Although this reduces the difficulty of solving the upper-level problem, due to the smaller resolution of lower-level spatial, the features transmitted upwards are semantically precise but spatially ambiguous, thus the restoration ability of multi-scale network in spatial details is limited. Hence, there is an urgent need to improve the quality of restoration of high-frequency detail part. A simple approach is to introduce a frequency domain transform as an alternative to a spatial domain transform. This would provide the algorithm with a choice and direct its attention to different frequencies. In this paper, we consider using the discrete wavelet transform for the following reasons: + +1. Many recent state-of-the-art deblurring algorithms [10, 19, 42] have introduced the discrete Fourier transform (DFT) as a frequency prior, which provides information that helps the algorithm to identify and select high- and low-frequency components that need to be preserved during restoration. Compared to DFT, the discrete wavelet transform (DWT) is better suited to deal with images containing more abrupt signals [5]. +2. Realistic blur and synthetic blur have significant distributional differences [24]. In the real world, due to the short exposure time of a camera, realistic blur has a specific directionality [29], i.e., the blur trajectory is regionally continuous; conversely, synthetic blur has an unnatural and discontinuous trajectory. To fully utilize the potential deblurring guidance brought by this trajectory continuity, we use 2D-DWT to reveal blur directionality, distinguish changes in the blur signal along different directions, and provide the algorithm with a reliable basis for deblurring through adaptive learning. + +To make 2D-DWT fit the data distribution and feature layer space more closely, we implemented 2D-DWT with an adaptive data distribution by using group convolution, transferring feature space from the spatial domain into the wavelet domain, generating sub-signals with different frequency features and different directional features, and then constructing wavelet losses for them in order to constrain them by self-supervision. Experimental results demonstrate that the proposed method has advanced performance in + +terms of accuracy and efficiency (Fig. 1). + +In summary, our contributions are as follows: + +- We propose a single-input and multiple-output(SIMO) multi-scale baseline for progressive image deblur, which reduces the complexity of existing multi-scale deblurring algorithms and improves the overall efficiency of image restoration networks. +- We construct a learnable discrete wavelet transform node(LWN) to provide reliable frequency and direction selection for the algorithm, which promotes the algorithm's restoration of the high-frequency components of edges and details. +- For the network to work better, reasonable multi-scale loss is proposed to guide pixel-by-pixel and scale-by-scale restoration. We also created reasonable self-supervised losses for the learnable wavelet transform to limit the learning direction of the wavelet kernel. +- We demonstrate the effectiveness of our algorithm under several motion blur datasets, especially real datasets, and obtain highly competitive results. + +# 2. Related Work + +Single-scale Deblurring Algorithm. Image deblurring have developed rapidly in recent years [2, 10, 11, 14, 39, 40]. To preserve richer image details, Kupyn et al. [11] viewed deblurring as a special case of image-to-image conversion, and for the first time utilized a GAN to recover images with richer details. To provide a high-quality and simple baseline, Chen et al. [2] proposed a nonlinear activation-free network that obtained excellent results. Zamir et al. [38] proposed a transformer with an encoder-decoder architecture that can be applied to higher-resolution images and utilizes cross-channel attention. Kong et al. [10] utilized a domain-based self-attention solver to reduce the transformer complexity and suppress artefacts. However, such single-scale algorithms estimate complex restoration problems directly and may require the design of restorers of higher complexity, which is suboptimal in terms of efficiency [9]. + +Multi-scale Deblurring Algorithm. Multi-scale motion deblur methods aim to achieve progressive recovery using a multi-stage or multi-sub-network approach. The multiscale algorithm previously used in the field of image deblurring is based on MIMO [3, 13, 20, 21, 27]. Recently, Cho et al. [3] proposed MIMO-UNet, which employs a multi-output single decoder as well as a multi-input single encoder to simulate a multi-scale architecture consisting of stacked networks, thereby greatly simplifying complexity. Zamir et al. [37] designed a multi-stage progressive image recovery network to simplify the information flow between multiple sub-networks. As shown in Sec. 3.1, our proposed multiscale approach further simplifies the structure of multi-scale networks and better balances the accuracy and complexity. + +![](images/b39c5e4e8a33272b3c101751994372a21ae473760c6f9291006c75259bca3ecf.jpg) +Figure 2. The overall architecture of the proposed MLWNet, the SEB is a simple module designed with reference [2], the WFB and WHB apply the LWN that implements the learnable 2D-DWT. In training phase, supervised learning is performed using $\mathcal{L}_{multi}$ and self-supervised restraint of the wavelet kernel is performed using $\mathcal{L}_{wavelet}$ . In testing phase, only the highest scale restored images is output. + +Application of Frequency in Deblurring. Frequency is an important property of images, in recent years frequency domain has been introduced into numerous DNNs to be solved for various tasks [10, 22, 25, 32, 35, 41]. Mao et al. [19] proposed DeepRFT, which introduced a residual module based on the Fourier transform into an advanced algorithm. Kong et al. [10] proposed FFTformer, which introduced the Fourier transform into a feed-forward network to effectively utilize different frequency information. Zou et al. [42] proposed SDWNet, which utilized frequency domain information as a complement to spatial domain information. Different from these methods, we introduce a learnable 2D-DWT capable of adapting to data distribution and feature space, more suitable not only for dealing with digital images rich in mutation signals, but also for dealing with real blurring. + +# 3. Proposed Method + +Our proposed MLWNet(Fig. 2) aims to explore an efficient multi-scale architectural approach to achieve high-quality blind deblurring of a single image. First, we designed a novel and highly scalable SIMO multi-scale baseline to address the performance bottleneck faced by partially multiscale networks. It takes a single image as an input and gradually generates a series of sharp images from the bottom up. Then, we propose a learnable wavelet transform node (LWN) for image deblurring, and it enhances the proposed algorithm's ability to restore detail information. + +# 3.1. Multi-scale Baseline + +Our multi-scale network maintains an encoder-decoder architecture, but retains the original image with the highest + +resolution as input, and restores sharp images of various scales sequentially in the output stage. In this way, we eliminate the complexity of the fusion module when dealing RGB images of different scales and with deep semantics. + +For ease of understanding, we will introduce here the Encoder phase, multi-scale semantic Fusion phase (hereafter referred to as Fusion phase), and Decoder phase sequentially. As shown in Fig. 2, the Encoder stage consists of several gray Simple Encoder Blocks(SEB), where information flows top-down for feature extraction; the Fusion stage consists of several green Wavelet Fusion Blocks(WFB), where information flows bottom-up for semantic fusion at different scales; and the Decoder stage consists of several purple Wavelet Head Blocks(WHB), and the bottom-up flow of information enables gradual restoration. Compared with WHB, SEB replaces LWN with a $1 \times 1$ conv for channel scaling, and a $3 \times 3$ depth-conv for feature extraction. + +The Encoder stage is responsible for full feature extraction, downsampling feature maps once after each block, and then passing the feature maps to the Fusion stage or the Decoder stage as needed, whose outputs $E_{out}^{i}$ to block of the $i - th$ layer can be denoted as: + +$$ +E _ {o u t} ^ {i} = \left\{ \begin{array}{l l} \phi_ {i} (e m b e d (x)), & i = i _ {m a x} \\ \phi_ {i} (E _ {o u t} ^ {i - 1}), & o t h e r w i s e \end{array} \right. \tag {1} +$$ + +where $\phi_{i}$ represent the block of corresponding layer of Encoder. The Fusion stage adapts and fuses information from different scales of semantics produced in Encoder to generate intermediate representations with deep semantics and shallow details. The Fusion stage's outputs $F_{out}^{i}$ to the block of $i - th$ layer is expressed as in Eq. 2, where $\delta_{i}$ + +represents the block of the corresponding layer of Fusion. + +$$ +F _ {o u t} ^ {i} = \left\{ \begin{array}{l l} \delta_ {i} \left(E _ {o u t} ^ {i} + E _ {o u t} ^ {i - 1}\right), & i = i _ {\text {m i n}} \\ \delta_ {i} \left(F _ {o u t} ^ {i - 1} + E _ {o u t} ^ {i}\right), & \text {o t h e r w i s e} \end{array} \right. \tag {2} +$$ + +The Decoder stage utilizes the information passed by both the Encoder and Fusion stages to progressively upsample the feature maps and generating pre-output feature maps at each scale separately. Output $D_{out}^{i}$ to the block of the $i - th$ layer is denoted as in Eq. 3. Here $\xi_{i}$ represents the block of the corresponding layer of the Decoder. + +$$ +D _ {o u t} ^ {i} = \left\{ \begin{array}{l l} \xi_ {i} \left(E _ {o u t} ^ {i}\right), & i = i _ {m i n} \\ \xi_ {i} \left(D _ {o u t} ^ {i - 1} + F _ {o u t} ^ {i}\right), & i > i _ {m i n} \text {a n d} i < i _ {m a x} \\ \xi_ {i} \left(D _ {o u t} ^ {i - 1} + F _ {o u t} ^ {i - 1} + E _ {o u t} ^ {i}\right), & i = i _ {m a x} \end{array} \right. +$$ + +Each block in the Decoder is followed by a $3 \times 3$ convolution to generate a restored image at each scale. To improve inference efficiency, all but the highest-scale outputs are generated in the training phase to compute the multiscale loss. We note that this design also have a similar role with auxiliary head [36] in facilitating the model's learning of the recovered uniform patterns. + +![](images/c17db24264ae3d97787ae53f7a4da5398aabfa3989e2231c56335c5c045ed4cb.jpg) +Figure 3. (a) The process of learnable 2D-wavelet convolution. (b) The construction process of the $N \times N$ 2D-wavelet kernel. + +# 3.2. Learnable Discrete Wavelet Transform + +To alleviate the defects of multi-scale architectures in the recovery of detail information, we design LWN and constructed the WFB and WHB based on it. Unlike the Fourier transform, the wavelet transform has adaptive time-frequency resolution, and performs excellently on digital images rich in mutation signal. Its use in conjunction with DNNs has gained increasing widespread attention [7, 16, 18, 34]. + +For ease of understanding, the 1D-DWT case is given here first. For the input discrete signal $t$ , given the wavelet function $\psi_{j,k}(t) = 2^{\frac{j}{2}}\psi (2^{j}t - k)$ for scaling factor $j$ , time factor $k$ , and scale function $\phi_{j,k}(t) = 2^{\frac{j}{2}}\phi (2^{j}t - k)$ , the decomposition of the input signal $t$ at $j_0$ is given by: + +$$ +f (t) = \sum_ {j > j _ {0}} \sum_ {k} d _ {j, k} \psi_ {j, k} (t) + \sum_ {k} c _ {j _ {0}, k} \phi_ {j _ {0}, k} (t) \tag {3} +$$ + +In this process, $d_{j,k} = \langle f(t),\psi_{j,k}(t)\rangle$ represents the detail coefficients (i.e., high-frequency components), $c_{j0,k} = \langle f(t),\phi_{j0,k}(t)\rangle$ represents the approximation coefficients (i.e., low-frequency components) so that the input signal can be disassembled step by step into a rich wavelet domain signal. In order to use wavelet transform in DNN, we introduce the analysis vectors $\vec{a}_0[k] = \left\langle \frac{1}{\sqrt{2}}\phi \left(\frac{t}{2}\right),\phi (t - k)\right\rangle ,$ $\vec{a}_{1}[k] = \left\langle \frac{1}{\sqrt{2}}\psi \left(\frac{t}{2}\right),\phi (t - k)\right\rangle$ to represent the high and low frequency filters respectively, then: + +$$ +c _ {j + 1, p} = \sum_ {k} \vec {a} _ {0} [ k - 2 p ] c _ {j, k} +$$ + +$$ +d _ {j + 1, p} = \sum_ {k} ^ {\kappa} \vec {a} _ {1} [ k - 2 p ] c _ {j, k} \tag {4} +$$ + +According to Eq. 4, the decomposition of the original signal over the wavelet basis can be viewed as a recursive convolution of the original signal with a specific filter using a step size of 2. Similarly, the inverse wavelet transform can be realized by transposed convolution of two synthetic vectors $\vec{s}_0$ and $\vec{s}_1$ . We extend the above situation to 2D by efficiently extracting axial high-frequency information using 2D discrete wavelet transform (2D-DWT), and designing learnable methods to make it adaptive to the data distribution and feature layers. Fig. 3(b) presents the construction process of the forward wavelet convolution kernel in 2D, which can be expressed as: + +$$ +\mathcal {F} _ {l l} = \vec {a _ {0}} \times \vec {a _ {0}} ^ {T}, \mathcal {F} _ {l h} = \vec {a _ {0}} \times \vec {a _ {1}} ^ {T} +$$ + +$$ +\mathcal {F} _ {h l} = \vec {a _ {1}} \times \vec {a _ {0}} ^ {T}, \mathcal {F} _ {h h} = \vec {a _ {1}} \times \vec {a _ {1}} ^ {T} \tag {5} +$$ + +$$ +\mathcal {K} _ {w} = \operatorname {c a t} \left(\mathcal {F} _ {l l}, \mathcal {F} _ {l h}, \mathcal {F} _ {h l}, \mathcal {F} _ {h h}\right) +$$ + +We set $\vec{a_0}$ and $\vec{a_1}$ as learnable filters; $\mathcal{F}_{ll}$ , $\mathcal{F}_{lh}$ , $\mathcal{F}_{hl}$ , and $\mathcal{F}_{hh}$ are low-frequency, horizontal high-frequency, vertical high-frequency, and diagonal high-frequency convolution operators obtained from vector outer products; The wavelet kernel $\kappa_w$ is spliced from above convolution operators. + +Then we implement learnable wavelet transform as a group convolution like Fig. 3(a), given a set of input feature maps $\mathcal{X}_{in} \in (C, H, W)$ , its projection in wavelet domain $\mathcal{X}_{out} \in \left(4C, \frac{H}{2}, \frac{W}{2}\right)$ (low-frequency, horizontal high-frequency, vertical high-frequency, and diagonal high-frequency component) is generated through the wavelet + +convolution. Similarly, the construction method of the inverse kernel and the forward propagation can be easily deduced, since they are inverse processes of each other. + +Then, a natural question can then be posed: how to ensure the correctness of the wavelet kernel learning and ensure that it does not degrade into a general group convolution and cause performance degradation? + +A common practice is to introduce the principle of perfect reconstruction [16, 34] to constrain adaptive wavelets, for a complex $z \in \mathbb{C}$ , given a filter $\vec{x}$ that applies to the $z$ -transform, its $z$ -transform can be expressed as $X(z) = \sum_{n \in \mathbb{Z}} \vec{x}(n) z^{-n}$ . Then $\vec{a}_0, \vec{a}_1, \vec{s}_0, \vec{s}_1$ can be obtained as their corresponding $z$ -transforms $A_0, A_1, S_0, S_1$ , which must then be satisfied if perfect reconstruction is desired: + +$$ +A _ {0} (- z) S _ {0} (z) + A _ {1} (- z) S _ {1} (z) = 0 \tag {6} +$$ + +$$ +A _ {0} (z) S _ {0} (z) + A _ {1} (z) S _ {1} (z) = 2 \tag {7} +$$ + +Eq. 6 is known as aliasing cancellation condition, which is used to the cancellation of the aliasing effects arising from the downsampling. We will give a detailed derivation of Eq. 6 and Eq. 7 in the Appendix. + +After wavelet convolution, to achieve frequency restoration in the wavelet domain, the input $\mathcal{X}$ wavelet domain component was separated into a separate dim, we use depth-wise convolution with an expansion factor of $r$ for wavelet domain feature extraction and transformation after the wavelet positive transform, and $1\times 1$ convolution for channel expansion and scaling, and finally a learnable wavelet inverse transform is used to reduce the wavelet domain feature maps to the spatial domain for output, which constitutes the LWN. We follow the rules of [2] and design structurally similar Wavelet Head Block (WHB) and Wavelet Fusion Block (WFB), both with LWN as an important base module, but used for semantic aggregation and multiscale output at different scales, respectively; thus, WHB adds the recovery convolution branch after WFB. + +# 3.3. Loss + +Multi-scale Loss. Follow PSNR loss [2], we propose multiscale loss function as the main loss of the algorithm for calculating the pixel difference between restored image and GT at each scale: + +$$ +\mathcal {L} _ {\text {m u l t i}} (x, y) = \sum_ {k = 1} ^ {K} w _ {k} \times \mathcal {L} _ {\text {p s n r}} \left(x _ {k}, y _ {k}\right) \tag {8} +$$ + +where $x, y$ denote the output and GT, respectively, $k$ denotes the downsampling level, $w_{k}$ represents the weights under the corresponding scale. Our design of $w_{k}$ is based on the simple finding that we cannot produce absolutely accurate downsampled clear images, and that the accuracy gap + +is progressively enlarged with decreasing scales. To ensure that lower scale output layers do not negatively affect the final output, we empirically set the loss weight $w_{k}$ to $\frac{1}{k}$ for the corresponding k-scale. + +Wavelet Loss. Both Eq. 6 and Eq. 7 can be easily converted losses optimized toward a minimum of 0, here we use the mean square error. Then, based on the convolvable nature of the $z$ transform, it is possible to construct convolutionally equivalent substitutions for polynomial multiplications for the perfect reconstruction condition: + +$$ +\begin{array}{l} L _ {w a v e l e t} (\theta_ {i}) = \left(\sum_ {k} ^ {N - 1} \left(\langle a _ {0}, s _ {0} \rangle_ {k} + \langle a _ {1}, s _ {1} \rangle_ {k}\right) - \hat {V} _ {\lfloor \frac {N}{2} \rfloor}\right) ^ {2} \\ = \left(\sum_ {k} ^ {N - 1} \left(\left\langle (- 1) ^ {k} a _ {0}, s _ {0} \right\rangle_ {k} + \left\langle (- 1) ^ {k} a _ {1}, s _ {1} \right\rangle_ {k}\right)\right) ^ {2} \tag {9} \\ \end{array} +$$ + +where $k$ denotes the position of the filter when it performs the convolution, $\hat{V}$ is a vector with a center position value of two, and $\theta_{i}$ denotes the constructed filter + +Overall Loss. Based on the multi-scale loss and wavelet loss, the loss function used in this study is: + +$$ +\mathcal {L} _ {\text {t o t a l}} (x, y) = \mathcal {L} _ {\text {w a v e l e t}} \left(\theta_ {i}\right) + \mathcal {L} _ {\text {m u l t i}} (x, y) \tag {10} +$$ + +# 4. Experimental Results + +# 4.1. Datasets and Implementation Details + +We evaluate our method on the realistic datasets Real-Blur [23] and RSBur [24], as well as the synthetic dataset GoPro [20], they are all composed of blurry-sharp image pairs. During training we use AdamW optimizer $(\alpha = 0.9$ and $\beta = 0.9)$ for a total of $600\mathrm{K}$ iterations. The initial value of the learning rate is $10^{-3}$ and is updated with cosine annealing schedule. The patch size is set to $256\times 256$ pixels and we followed the training strategy of [28, 38]. For data augmentation, we only use random flips and rotations. More experimental results can be found in the Appendix. + +# 4.2. Comparison with State-of-the-Art Methods + +We have compared our method with several advanced deblurring methods and use the PSNR and SSIM to evaluate the quality of restored images. + +**Evaluations on the RealBlur dataset.** Since our proposed method is driven by real-world deblurring, we first conduct comparisons on RealBlur [23]. The quantitative analysis results are shown in Tab. 1, the PSNR value of our method is 0.91dB and 0.49dB higher than state-of-the-art GRL on the RealBlur-J and RealBlur-R datasets respectively. Compared to other multi-scale architectures, our method has fewer model computational and running time while the performance is better. Fig. 4 shows visual comparison on RealBlur-J dataset, our method obtains restored text with + +![](images/215bdb7cdeb1601782a476847e66c891bd797b8de11ef519a88050281aeb185b.jpg) +Blurry image + +![](images/f9f64d806f80acd40cacf204f51bef984d3f170b07eba1240a1638ff150c106c.jpg) +DeblurGAN-v2 + +![](images/f308de45c37be92ad81de0c16181cfe61669b994af9a9e6b40a08c2b9926d208.jpg) +SRN + +![](images/717ebaf86742e707eed5b75d11d5d6117c6f0b90dd9a2492ca358dadd843c627.jpg) +MIMO-UNet+ + +![](images/16cd1a14bbe99a59e5358466b6f42f140bcd8f6bf54d673e7c4d748a255b8253.jpg) +MPRNet + +![](images/e6f347886090fcb55653e1237bbf23fe901583ab80271d6d3c655228a5af0c24.jpg) +BANet + +![](images/3daa0d4338f14c248d325bd2036e8037fa6b5037d2e3e76130dfcfab43938f88.jpg) +Blurry + +![](images/57b0eb1ced75505fc522cc0f50c80a5e1966062e2ba733ce7ab57b1efcd8d607.jpg) +GT + +![](images/31dc3d903701c53a32ed339747e77585c5d667cc143e09e70832695e88bcc0d9.jpg) +MSSNet + +![](images/35e24b695b5265712dd11e27b6eeb0cd771f581ee3696a63e5337195c88d06f4.jpg) +Stripformer +Figure 4. Visual comparisons on the RealBlur-J dataset [23]. The proposed method generates an image with clearer characters. + +![](images/200900a46d8bfd7a0cc96759283ccb127c4c9378eaf6d1cd57f57ae3fc93d4e8.jpg) +GRL + +![](images/37eaff3d66cc31ceebaa60029b2c839c63842d72a257f26c2eac1f5a6833f587.jpg) +FFTformer + +![](images/46b70589e14259f5b5eb2ab46222bacc8ccb610982af6f74ab8c8988e938aa3b.jpg) +Ours + +
MethodRealBlur-JRealBlur-RAvg. runtime
PSNRSSIMPSNRSSIM
DeblurGAN-v2 [12]29.690.87036.440.9350.04s
SRN [27]31.380.90938.650.9650.07s
MPRNet [37]31.760.92239.310.9720.09s
SDWNet [42]30.730.89638.210.9630.04s
MIMO-UNet+ [3]31.920.919--0.02s
MIMO-UNet++ [3]32.050.921---
DeepRFT+ [19]32.190.93139.840.9720.09s
BANet [29]32.000.92339.550.9710.06s
BANet+ [29]32.420.92939.900.9720.12s
Stripformer [28]32.480.92939.840.9740.04s
MSSNet [9]32.100.92839.760.9720.06s
MSDI-Net [13]32.350.923--0.06s
MAXIM-3S [30]32.840.93539.450.962-
FFTformer [10]32.620.93340.110.9730.13s
GRL-B [14]32.820.93240.200.9741.28s
MLWNet-S33.020.933--0.04s
MLWNet-B33.840.94140.690.9760.05s
+ +Table 1. Quantitative evaluations on the RealBlur dataset [23]. The experimental results were trained under the corresponding datasets respectively, and average runtime is tested on $256 \times 256$ patches. + +richer edge details, while achieving the best contrast, sharpness, brightness, and structural details of the image. + +Evaluations on the RSBlur dataset. We further conduct experiments on the latest real-world deblurring dataset RSBlur [24] and follow the protocol of this dataset for fair comparison. Tab. 2 summarizes the quantitative evaluation results of our method with advanced algorithms. The proposed MLWNet achieved the highest PSNR and SSIM, which were 34.94 and 0.880 respectively. In addition, the model we trained on RSBlur dataset also achieved the best results on RealBlur-J dataset. We show some visual comparisons in Fig. 5. We note that our method outperforms other methods in low-light blurry scenes, proving that the + +
MethodRSBlurRealBlur-J
PSNRSSIMPSNRSSIM
SRN [27]32.530.84029.860.886
MIMO-Unet [3]32.730.84629.530.876
MIMO-Unet+ [3]33.370.85629.990.889
MPRNet [37]33.610.86130.460.899
Restormer [38]33.690.86330.480.891
Uformer-B [33]33.980.86630.370.899
SFNet [4]34.350.87230.260.897
MLWNet-B34.940.88030.530.905
+ +Table 2. Quantitative evaluations trained on the RSBlur dataset [24], the RealBlur-J dataset was used for testing only. + +
Method[23]→ [24]MACs(G)
PSNRSSIM
DeblurGAN-v2 [12]30.150.76642.0
MPRNet [37]29.560.785760.8
MIMO-UNet+ [3]29.690.792154.4
BANet [29]30.190.806263.9
BANet+ [29]30.240.809588.7
MSSNet [9]29.860.806154.0
FFTformer [10]29.700.787131.8
MLWNet-B30.910.818108.2
+ +Table 3. Quantitative evaluation for generalizability shows the results of models trained on the RealBlur-J dataset and tested on the RSBlur dataset, MACs are measured on $256 \times 256$ patches. + +proposed method makes textures and structures clearer. + +In addition, we evaluate the RSBlur dataset using models trained only on RealBlur-J to fairly compare the generalization of methods to real scenarios. As shown in Tab. 3, our method produced results with highest PSNR value of 30.91. The results of Tab. 2 and Tab. 3 show that our model has a better generalization ability. + +Evaluations on the GoPro dataset. We conduct a exten + +![](images/0b6d3a6ad6bc0f669e5be3cda7b9bc2baab1203c222d8fb672c8cd84d1741279.jpg) + +![](images/90971285365ea3258b0f209a1d8299550b1f51c18a78e938ea98f12aca3782f2.jpg) +SRN + +![](images/fbe3bf2dcd531d68e504db1dafedbc5efef65c39f6f4c29b9d25bc2d4b7d1eeb.jpg) +MIMO-UNet + +![](images/b611528098ba55aa72842822a6b13408c852eb2b2a3848380bf3d4c8ff6a1b75.jpg) +MIMO-UNet+ + +![](images/c6b608e12729bb2f6f17075e0d0397d50b43986036af164a3c423af3961e9665.jpg) +MPRNet + +![](images/cc83602cb1a7a0ebeb2c8740b78c76780d2882845493c2ba1ba497bf789056a3.jpg) +Blurry image +Blurry + +![](images/3e004f48ac48ba037fd91aad6041bf4840a2f29198106f857808621c42006dce.jpg) +GT + +![](images/f00e7c52ed013110c1f65c797c5d0fc44342c733b5e52ffde9afd24b8afbab65.jpg) +Restormer + +![](images/668ce462ff79c03992ca1d403b1948c1f2a2d43f0151e5f4608da957b497597b.jpg) +Uformer-B + +![](images/585ba272724aba34360d1598872f15fecfe33df5104a6e966bfa337dbfc3d71b.jpg) +SFNet + +![](images/6003de8f5bcef6e4b2c2101766cff97d59f4a10adfbf822687d48424ccca6bfc.jpg) +Ours + +![](images/acac45ca631bc8e8518d5dd1bbb4ed01f503a3d81280e8fffdd5f9cf8d8e5bfb.jpg) +Figure 5. Visual comparisons on the RSBur dataset [24]. The deblurring performance of the proposed method in low-light is impressive, The recovery of characters and texture structures far exceeds other advanced methods. +Blurry image + +![](images/0768e11c2fe4ca5753d8a7ea2eaaaa63f9f065e636de54179cb6e118e3f91a7a.jpg) +MIMO-UNet + +![](images/ec338d67bd377e6bb9d5ea8da1f1a0d773a1e82d5b909405d39fb112daa4d5ef.jpg) +MIMO-UNet+ + +![](images/301062017a4633defa98faa1e007decee4f9f7e594c78ae00ec8abe8df9c0053.jpg) +MPRNet + +![](images/eb2837ebefc59f58cd28d1439d3df4eecbaac650ecac73a34adc751830681732.jpg) +NAFNet + +![](images/091cc93979440846cfef3b1c91fede03162da15eea65e86a7a7a3d649a0ada82.jpg) +MAXIM + +![](images/31dbd953817bb7c2f23ef591a3a38d007ff62e3b66e95fa2d11b5fa6dcb78acf.jpg) +Blurry +Figure 6. Visual comparisons on the GoPro dataset [20]. Our method better preserves texture information without sharpening. + +![](images/12d8c3b7d37de45745154f0618c99448f1e480beb030c83d63ff7d57cfc8084c.jpg) +GT + +![](images/47959d8fc10e347beebe3d17f3b7ede9f64fdf474e8f6d1284af07cca1daf510.jpg) +Stripformer + +![](images/5a8102137173b287f0fd68320bffa4493735e681c26ca8f2093c1b205e20fe56.jpg) +Restormer + +![](images/62a8eeac1b07563f9ea4a80110438401e91ced256a726e01301094ee97714830.jpg) +FFTformer + +![](images/bd3a83710a26c343281ae0c5c97ff0cfc0a78c13e971244036d2b7cd8b960bb2.jpg) +GRL + +![](images/53feea5edbd48f545a144cef2f5990d2b2bf004e2376d5d40d5b00464aff7989.jpg) +Ours + +sive comparison with advanced algorithms on GoPro [20]. Tab. 4 shows the quantitative evaluation results. The gap between our method and the state-of-the-art FFTformer is only 0.001 in SSIM value, while the running time is reduced by $60\%$ . We show a visual comparison on the GoPro dataset in Fig. 6. We can see that our method recovers blurred textures well without sharpening. For the reasons why the proposed MLWNet does not achieve optimal performance in synthetic blurry, we conducted a detailed analysis in Sec. 5. + +# 5. Analysis and Discussion + +In this section, we provide a more in-depth analysis and show the contribution of each component of the proposed method. We perform 20w iterations on the RealBlur-J dataset for abtation with a batch size of 8 to train our method using a version of the model with a width of 32, and the results are shown in Tab. 6. + +Single-scale vs. Multi-scale. Given that each node of the baseline network is composed of SEB, we have deformed the multi-scale and single-scale according to the architecture. Tab. 5 shows that the multi-scale architecture improves the PSNR and SSIM values to varying degrees, proving that + +
MethodGOPROAvg.rruntime
PSNRSSIM
DeblurGAN-v2 [12]29.550.9340.04s
SRN [27]30.260.9340.07s
DMPHN [39]31.200.9450.21s
SDWNet [42]31.260.9660.04s
MPRNet [37]32.660.9590.09s
MIMO-UNet+ [3]32.450.9570.02s
DeepRFT+ [19]33.230.9630.09s
MAXIM-3S [30]32.860.961-
Stripformer [28]33.080.9620.04s
MSDI-net [13]33.280.9640.06s
Restormer [38]33.570.9660.08s
NAFNet [2]33.690.9670.04s
FFTformer [10]34.210.9690.13s
GRL-B [14]33.930.9681.28s
MLWNet-B33.830.9680.05s
+ +Table 4. Quantitative evaluations trained and tested on the GoPro dataset [20]. Our proposed MLWNet obtains competitive results with a combination of time efficiency and accuracy. + +the SIMO strategy we employ helps to eliminate motion blur to some extent. As shown in Sec. 3.1, we use SIMO for coarse-to-fine restoration, and only use multiple outputs to calculate multi-scale losses during training. + +WFB and WHB. We default to using WFB and WHB via $\mathcal{L}_{wavelet}$ statute in this section, we also proved the effectiveness of LWN, because LWN is a subpart of the first two. After applying them, the PSNR index improved by $0.25\mathrm{dB}$ compared to the multi-scale baseline, which shows that LWN and its extension modules can easily help multiscale algorithms improve detail recovery capabilities. Fig. 7 shows the four types of feature maps generated after forward learnable wavelet convolution. We can see that the high and low frequency information of the input feature map are mixed together, and after wavelet group convolution, the network exerts different attention on the feature information in different directions or frequencies. + +Effectiveness of $\mathcal{L}_{wavelet}$ . By adding unreduced WFB and WHB to the baseline without using wavelet loss, the performance improvement of the baseline is minimal and does not break the bottleneck of multiscale detail recovery, which suggests that the wavelet convolution degenerates into a general group convolution that only serves to deepen the network, proving that merely deepening the network is ineffective for the restoration of detail information. + +
MethodSISOMIMOSIMO
PSNR/SSIM32.29/0.92432.19/0.92832.37/0.929
MACs(G)19.2421.8319.29
+ +Table 5. Comparison in various input and output modes. + +
SIMOWFBWHBLwaveletPSNRSSIMMACs(G)
✓32.370.92919.29
✓✓✓32.400.92828.21
✓✓✓32.490.92825.28
✓✓✓32.570.92922.22
✓✓✓✓32.620.93128.21
+ +![](images/58b3cc4ebc77ac4abc8cad372c4923708f469c568b1480f2a614e68a40c43774.jpg) +Figure 7. Feature maps representing high-and low-frequency components generated after learnable wavelet convolution. Zoom in on the screen for the best view. + +Limitations and analysis. It is observed from Tab. 1-3 that our method fails to achieve the same expected high performance with synthetic blur as it does with realistic blur. We try to analyze the reasons for this phenomenon via dataset comparison and experimental results. First, the average frame synthesis method used in the synthetic blur + +dataset leads to unnatural discontinuous blur trajectories, thereby introducing strange high-frequency information interference (see Fig. 8). Second, synthetic data tends to produce an unnatural mixture of high and low frequencies. It is easy to mix with the average color area at the edge of the texture, causing the confusion of high-low frequency information. + +![](images/4479fc93893bb8334181aacfbef23656b02524fd32475217e26cdf6f96467f9a.jpg) +(a) Realistic Blur + +![](images/02bae2280ed7d33a0ca02349aadccfa0ecd089ceb702a3479a3e441fca2e5f15.jpg) +(b) Synthetic Blur +Figure 8. The difference between realistic blur (a) and synthetic blur (b). In the green box, the synthetic blur appears with color averaging resulting in high and low frequency confusion, and in the blue box has unnatural discontinuous trajectories. + +Table 6. Ablation study on components of the proposed MLWNet. We set the baseline network to use SEB in its entirety, and models that do not use SIMO will represent single scales using SISO. + +
Noise LevelL1L2L3L4L5
GoPro34.8134.6633.7633.1932.63
RealBlur-J33.9233.8133.9733.9333.54
+ +Table 7. Performance comparison at different noise difference levels, where L3 contains the noise difference mean. + +Finally, there is a large difference in noise levels between synthetic blur and real blur. We follow [1] to calculate the noise level of each testset, and divided the noise difference between clear images and blurred images of GoPro and RealBlur-J into 5 levels for separate testing. As shown in Tab. 7, we note that as the noise difference increases, our method maintains good performance on RealBlur-J, while the performance drops fast on GoPro. This shows that our method is robust to the noise of real blur, and that there are differences in the noise between synthetic blur and real blur. + +# 6. Conclusion + +In this paper, we propose a SIMO-based multi-scale architecture to achieve efficient motion deblurring. We have developed a learnable discrete wavelet transform module, which not only improves the algorithm's ability to recover details, but is also more applicable to the real world. In addition, we construct a reasonable multi-scale loss to guide the recovery of blurred images pixel by pixel and scale by scale, and constrain the learning direction of the wavelet kernel with self-supervised loss to achieve better image deblurring. Through extensive experiments on multiple real and synthetic datasets, we demonstrate that it outperforms existing state-of-the-art methods in terms of quality and efficiency of image restoration, especially with optimal deblurring performance and generalization in real scenes. + +# References + +[1] Guangyong Chen, Fengyuan Zhu, and Pheng Ann Heng. An efficient statistical method for image noise level estimation. In Proceedings of the IEEE International Conference on Computer Vision, pages 477-485, 2015. 8 +[2] Liangyu Chen, Xiaojie Chu, Xiangyu Zhang, and Jian Sun. Simple baselines for image restoration. In European Conference on Computer Vision, pages 17-33. Springer, 2022. 1, 2, 3, 5, 7 +[3] Sung-Jin Cho, Seo-Won Ji, Jun-Pyo Hong, Seung-Won Jung, and Sung-Jea Ko. Rethinking coarse-to-fine approach in single image deblurring. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4641-4650, 2021. 1, 2, 6, 7 +[4] Yuning Cui, Yi Tao, Zhenshan Bing, Wenqi Ren, Xinwei Gao, Xiaochun Cao, Kai Huang, and Alois Knoll. Selective frequency network for image restoration. In International Conference on Learning Representations, ICLR, 2023. 6 +[5] Ingrid Daubechies. The wavelet transform, time-frequency localization and signal analysis. IEEE transactions on information theory, 36(5):961-1005, 1990. 2 +[6] Golnaz Ghiasi, Tsung-Yi Lin, and Quoc V Le. Nas-fpn: Learning scalable feature pyramid architecture for object detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7036-7045, 2019. 1 +[7] Wooseok Ha, Chandan Singh, Francois Lanusse, Srigokul Upadhyayula, and Bin Yu. Adaptive wavelet distillation from neural networks through interpretations. Advances in Neural Information Processing Systems, 34:20669-20682, 2021. 4 +[8] Bumsoo Kim, Jonghwan Mun, Kyoung-Woon On, Minchul Shin, Junhyun Lee, and Eun-Sol Kim. Mstr: Multi-scale transformer for end-to-end human-object interaction detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19578-19587, 2022. 1 +[9] Kiyeon Kim, Seungyong Lee, and Sunghyun Cho. Mssnet: Multi-scale-stage network for single image deblurring. In European Conference on Computer Vision, pages 524-539. Springer, 2022. 1, 2, 6 +[10] Lingshun Kong, Jiangxin Dong, Jianjun Ge, Mingqiang Li, and Jinshan Pan. Efficient frequency domain-based transformers for high-quality image deblurring. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5886-5895, 2023. 1, 2, 3, 6, 7 +[11] Orest Kupyn, Volodymyr Budzan, Mykola Mykhailych, Dmytro Mishkin, and Jiří Matas. Deblurgan: Blind motion deblurring using conditional adversarial networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 8183-8192, 2018. 2 +[12] Orest Kupyn, Tetiana Martyniuk, Junru Wu, and Zhangyang Wang. Deblurgan-v2: Deblurring (orders-of-magnitude) faster and better. In Proceedings of the IEEE/CVF international conference on computer vision, pages 8878-8887, 2019. 1, 6, 7 +[13] Dasong Li, Yi Zhang, Ka Chun Cheung, Xiaogang Wang, Hongwei Qin, and Hongsheng Li. Learning degradation rep + +resentations for image deblurring. In European Conference on Computer Vision, pages 736-753. Springer, 2022. 1, 2, 6, 7 +[14] Yawei Li, Yuchen Fan, Xiaoyu Xiang, Denis Demandolx, Rakesh Ranjan, Radu Timofte, and Luc Van Gool. Efficient and explicit modelling of image hierarchies for image restoration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18278-18289, 2023. 2, 6, 7 +[15] Tsung-Yi Lin, Piotr Dólar, Ross Girshick, Kaiming He, Bharath Hariharan, and Serge Belongie. Feature pyramid networks for object detection. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2117-2125, 2017. 1 +[16] Pengju Liu, Hongzhi Zhang, Wei Lian, and Wangmeng Zuo. Multi-level wavelet convolutional neural networks. IEEE Access, 7:74973-74985, 2019. 4, 5 +[17] Shu Liu, Lu Qi, Haifang Qin, Jianping Shi, and Jiaya Jia. Path aggregation network for instance segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 8759-8768, 2018. 1 +[18] Wei Liu, Qiong Yan, and Yuzhi Zhao. Densely self-guided wavelet network for image denoising. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, pages 432-433, 2020. 4 +[19] X Mao, Y Liu, W Shen, Q Li, and Y Wang. Deep residual fourier transformation for single image deblurring. arxiv 2021. arXiv preprint arXiv:2111.11745, 2021. 2, 3, 6, 7 +[20] Seungjun Nah, Tae Hyun Kim, and Kyoung Mu Lee. Deep multi-scale convolutional neural network for dynamic scene deblurring. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3883-3891, 2017. 1, 2, 5, 7 +[21] Dongwon Park, Dong Un Kang, Jisoo Kim, and Se Young Chun. Multi-temporal recurrent neural networks for progressive non-uniform single image deblurring with incremental temporal training. In European Conference on Computer Vision, pages 327-343. Springer, 2020. 1, 2 +[22] Zequn Qin, Pengyi Zhang, Fei Wu, and Xi Li. Fcanet: Frequency channel attention networks. In Proceedings of the IEEE/CVF international conference on computer vision, pages 783-792, 2021. 3 +[23] Jaesung Rim, Haeyun Lee, Jucheol Won, and Sunghyun Cho. Real-world blur dataset for learning and benchmarking deblurring algorithms. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXV 16, pages 184-201. Springer, 2020. 1, 5, 6 +[24] Jaesung Rim, Geonung Kim, Jungeon Kim, Junyong Lee, Seungyong Lee, and Sunghyun Cho. Realistic blur synthesis for learning image deblurring. In European conference on computer vision, pages 487-503. Springer, 2022. 2, 5, 6, 7 +[25] Roman Suvorov, Elizaveta Logacheva, Anton Mashikhin, Anastasia Remizova, Arsenii Ashukha, Aleksei Silvestrov, Naejin Kong, Harshith Goka, Kiwoong Park, and Victor Lempitsky. Resolution-robust large mask inpainting with fourier convolutions. In Proceedings of the IEEE/CVF winter + +conference on applications of computer vision, pages 2149-2159, 2022. 3 +[26] Mingxing Tan, Ruoming Pang, and Quoc V Le. Efficientdet: Scalable and efficient object detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10781-10790, 2020. 1 +[27] Xin Tao, Hongyun Gao, Xiaoyong Shen, Jue Wang, and Jiaya Jia. Scale-recurrent network for deep image deblurring. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 8174-8182, 2018. 1, 2, 6, 7 +[28] Fu-Jen Tsai, Yan-Tsung Peng, Yen-Yu Lin, Chung-Chi Tsai, and Chia-Wen Lin. Stripformer: Strip transformer for fast image deblurring. In European Conference on Computer Vision, pages 146-162. Springer, 2022. 1, 5, 6, 7 +[29] Fu-Jen Tsai, Yan-Tsung Peng, Chung-Chi Tsai, Yen-Yu Lin, and Chia-Wen Lin. Banet: a blur-aware attention network for dynamic scene deblurring. IEEE Transactions on Image Processing, 31:6789-6799, 2022. 1, 2, 6 +[30] Zhengzhong Tu, Hossein Talebi, Han Zhang, Feng Yang, Peyman Milanfar, Alan Bovik, and Yinxiao Li. Maxim: Multi-axis mlp for image processing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5769-5780, 2022. 6, 7 +[31] Chien-Yao Wang, Alexey Bochkovskiy, and Hong-Yuan Mark Liao. Yolov7: Trainable bag-of-freebies sets new state-of-the-art for real-time object detectors. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7464-7475, 2023. 1 +[32] Luyuan Wang and Yankui Sun. Image classification using convolutional neural network with wavelet domain inputs. IET Image Processing, 16(8):2037-2048, 2022. 3 +[33] Zhendong Wang, Xiaodong Cun, Jianmin Bao, Wengang Zhou, Jianzhuang Liu, and Houqiang Li. Uformer: A general u-shaped transformer for image restoration. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 17683-17693, 2022. 6 +[34] Moritz Wolter and Jochen Garcke. Adaptive wavelet pooling for convolutional neural networks. In International Conference on Artificial Intelligence and Statistics, pages 1936-1944. PMLR, 2021. 4, 5 +[35] Yanchao Yang and Stefano Soatto. Fda: Fourier domain adaptation for semantic segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4085-4095, 2020. 3 +[36] Changqian Yu, Changxin Gao, Jingbo Wang, Gang Yu, Chunhua Shen, and Nong Sang. Bisenet v2: Bilateral network with guided aggregation for real-time semantic segmentation. International Journal of Computer Vision, 129: 3051-3068, 2021. 4 +[37] Syed Waqas Zamir, Aditya Arora, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, Ming-Hsuan Yang, and Ling Shao. Multi-stage progressive image restoration. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14821-14831, 2021. 1, 2, 6, 7 +[38] Syed Waqas Zamir, Aditya Arora, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, and Ming-Hsuan Yang. + +Restormer: Efficient transformer for high-resolution image restoration. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5728-5739, 2022. 2, 5, 6, 7 +[39] Hongguang Zhang, Yuchao Dai, Hongdong Li, and Piotr Koniusz. Deep stacked hierarchical multi-patch network for image deblurring. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5978-5986, 2019. 2, 7 +[40] Kaihao Zhang, Wenhan Luo, Yiran Zhong, Lin Ma, Bjorn Stenger, Wei Liu, and Hongdong Li. Deblurring by realistic blurring. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2737-2746, 2020. 2 +[41] Yijie Zhong, Bo Li, Lv Tang, Senyun Kuang, Shuang Wu, and Shouhong Ding. Detecting camouflaged object in frequency domain. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4504-4513, 2022. 3 +[42] Wenbin Zou, Mingchao Jiang, Yunchen Zhang, Liang Chen, Zhiyong Lu, and Yi Wu. Sdwnet: A straight dilated network with wavelet transformation for image deblurring. In Proceedings of the IEEE/CVF international conference on computer vision, pages 1895-1904, 2021. 1, 2, 3, 6, 7 \ No newline at end of file diff --git a/2024/Efficient Multi-scale Network with Learnable Discrete Wavelet Transform for Blind Motion Deblurring/images.zip b/2024/Efficient Multi-scale Network with Learnable Discrete Wavelet Transform for Blind Motion Deblurring/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..cdb7604f9ddde2113014e48b4cccc8ec442228c2 --- /dev/null +++ b/2024/Efficient Multi-scale Network with Learnable Discrete Wavelet Transform for Blind Motion Deblurring/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b158fa4dfda514ff453ceca6e590feb4e4c39e5eb36fff2f7e00cffe612765e9 +size 775343 diff --git a/2024/Efficient Multi-scale Network with Learnable Discrete Wavelet Transform for Blind Motion Deblurring/layout.json b/2024/Efficient Multi-scale Network with Learnable Discrete Wavelet Transform for Blind Motion Deblurring/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..8a6974efb6da01ef7acb24a8d83a2e687a2306e2 --- /dev/null +++ b/2024/Efficient Multi-scale Network with Learnable Discrete Wavelet Transform for Blind Motion Deblurring/layout.json @@ -0,0 +1,10633 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 69, + 102, + 524, + 139 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 102, + 524, + 139 + ], + "spans": [ + { + "bbox": [ + 69, + 102, + 524, + 139 + ], + "type": "text", + "content": "Efficient Multi-scale Network with Learnable Discrete Wavelet Transform for Blind Motion Deblurring" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 52, + 144, + 552, + 187 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 52, + 144, + 552, + 187 + ], + "spans": [ + { + "bbox": [ + 52, + 144, + 552, + 187 + ], + "type": "text", + "content": "Xin Gao\\*1,2 Tianheng Qiu\\*2,3,4 Xinyu Zhang†2 Hanlin Bai\\*1 Kang Liu\\*1 Xuan Huang†4 Hu Wei\\*4 Guoying Zhang†1 Huaping Liu\\*2 *Equal Contribution † Corresponding Author" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 120, + 191, + 488, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 120, + 191, + 488, + 236 + ], + "spans": [ + { + "bbox": [ + 120, + 191, + 488, + 236 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 120, + 191, + 488, + 236 + ], + "type": "text", + "content": "China University of Mining & Technology-Beijing " + }, + { + "bbox": [ + 120, + 191, + 488, + 236 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 120, + 191, + 488, + 236 + ], + "type": "text", + "content": "Tsinghua University " + }, + { + "bbox": [ + 120, + 191, + 488, + 236 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 120, + 191, + 488, + 236 + ], + "type": "text", + "content": "University of Science and Technology of China " + }, + { + "bbox": [ + 120, + 191, + 488, + 236 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 120, + 191, + 488, + 236 + ], + "type": "text", + "content": "Hefei Institutes of Physical Science, Chinese Academy of Sciences" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 143, + 251, + 192, + 264 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 251, + 192, + 264 + ], + "spans": [ + { + "bbox": [ + 143, + 251, + 192, + 264 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 276, + 290, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 276, + 290, + 540 + ], + "spans": [ + { + "bbox": [ + 46, + 276, + 290, + 540 + ], + "type": "text", + "content": "Coarse-to-fine schemes are widely used in traditional single-image motion deblur; however, in the context of deep learning, existing multi-scale algorithms not only require the use of complex modules for feature fusion of low-scale RGB images and deep semantics, but also manually generate low-resolution pairs of images that do not have sufficient confidence. In this work, we propose a multi-scale network based on single-input and multiple-outputs(SIMO) for motion deblurring. This simplifies the complexity of algorithms based on a coarse-to-fine scheme. To alleviate restoration defects impacting detail information brought about by using a multi-scale architecture, we combine the characteristics of real-world blurring trajectories with a learnable wavelet transform module to focus on the directional continuity and frequency features of the step-by-step transitions between blurred images to sharp images. In conclusion, we propose a multi-scale network with a learnable discrete wavelet transform (MLWNet), which exhibits state-of-the-art performance on multiple real-world deblurred datasets, in terms of both subjective and objective quality as well as computational efficiency. Our code is available on https://github.com/thqiu0419/MLWNet." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 562, + 128, + 574 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 562, + 128, + 574 + ], + "spans": [ + { + "bbox": [ + 47, + 562, + 128, + 574 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 581, + 287, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 581, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 46, + 581, + 287, + 689 + ], + "type": "text", + "content": "Most current top-performing single-image blind deblurring algorithms are based on DNNs, which can be structurally categorized into single-scale [2, 10, 12, 28, 29, 42] and multi-scale approaches [3, 9, 13, 27, 37]. Compared to single-scale methods, multi-scale methods are built on the idea of moving from coarse-to-fine, and thus they decompose the challenging single-image blind deblurring problem into easier-to-solve sub-problems to restore the blurred image step-by-step." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "content": "Earlier DNNs exploiting the coarse-to-fine concept usually employ estimators of different scales to produce re" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 305, + 249, + 547, + 424 + ], + "blocks": [ + { + "bbox": [ + 305, + 249, + 547, + 424 + ], + "lines": [ + { + "bbox": [ + 305, + 249, + 547, + 424 + ], + "spans": [ + { + "bbox": [ + 305, + 249, + 547, + 424 + ], + "type": "image", + "image_path": "00c418dd730d494de65b5951a619cb9246b6202b63a1494cf6284118d1f7e695.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 425, + 547, + 458 + ], + "lines": [ + { + "bbox": [ + 305, + 425, + 547, + 458 + ], + "spans": [ + { + "bbox": [ + 305, + 425, + 547, + 458 + ], + "type": "text", + "content": "Figure 1. Performance comparison on the RealBlur-J [23] test dataset in terms of PSNR and GMACs. Our proposed MLWNet achieves superiority in comparison with other state-of-the-arts." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 462, + 546, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 462, + 546, + 653 + ], + "spans": [ + { + "bbox": [ + 304, + 462, + 546, + 653 + ], + "type": "text", + "content": "stored images gradually [20, 21, 27]. However, not only is the extracted semantic information not sufficiently representative, but the use of the same complexity at different scales may lead to redundant computations; More advanced multi-scale algorithms [3] use a global encoder-decoder, which significantly reduces the algorithm's runtime, multiple upsampling and downsampling leads to insufficient restoration of detailed information, and therefore such algorithms require the introduction of additional fusion modules to encode input images for fusion with deeper semantics. In addition, existing multi-scale algorithms use multiple-input and multiple-output (MIMO) architectures, which require the introduction of manually constructed downsampled image pairs, and usually employ simple interpolation algorithms to generate low-resolution images for the sake of efficiency, which is obviously not reliable." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 654, + 547, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 547, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 547, + 712 + ], + "type": "text", + "content": "We note that the gradual insertion of images used by existing coarse-to-fine algorithms is redundant. Referring to numerous algorithms [6, 8, 15, 17, 26, 31] that are widely used in downstream tasks, the network starts with a single input and also extracts features at different scales. These" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 294, + 732, + 315, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 732, + 315, + 742 + ], + "spans": [ + { + "bbox": [ + 294, + 732, + 315, + 742 + ], + "type": "text", + "content": "2733" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 192 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 192 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 192 + ], + "type": "text", + "content": "features can be aggregated by simple feature fusion to ultimately obtain competitive results. That is, a DNN itself possesses the ability to learn effective features at different scales. This inspired us to design a multi-scale architecture that retains an original image as input and sequentially restores images at each scale in output stage. This is not only more theoretically sound, but also eliminates the time required to generate pairs of images of different resolutions, as well as the huge complexity increase brought about by fusion modules between RGB domain and deep features." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 194, + 289, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 194, + 289, + 373 + ], + "spans": [ + { + "bbox": [ + 46, + 194, + 289, + 373 + ], + "type": "text", + "content": "The coarse-to-fine algorithm has another inherent defect. During progressive restoration, the solution of the upper-level problem takes the solution of the lower-level as initialization [9]. Although this reduces the difficulty of solving the upper-level problem, due to the smaller resolution of lower-level spatial, the features transmitted upwards are semantically precise but spatially ambiguous, thus the restoration ability of multi-scale network in spatial details is limited. Hence, there is an urgent need to improve the quality of restoration of high-frequency detail part. A simple approach is to introduce a frequency domain transform as an alternative to a spatial domain transform. This would provide the algorithm with a choice and direct its attention to different frequencies. In this paper, we consider using the discrete wavelet transform for the following reasons:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 376, + 287, + 604 + ], + "type": "list", + "angle": 0, + "index": 4, + "blocks": [ + { + "bbox": [ + 47, + 376, + 287, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 376, + 287, + 472 + ], + "spans": [ + { + "bbox": [ + 47, + 376, + 287, + 472 + ], + "type": "text", + "content": "1. Many recent state-of-the-art deblurring algorithms [10, 19, 42] have introduced the discrete Fourier transform (DFT) as a frequency prior, which provides information that helps the algorithm to identify and select high- and low-frequency components that need to be preserved during restoration. Compared to DFT, the discrete wavelet transform (DWT) is better suited to deal with images containing more abrupt signals [5]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 472, + 287, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 472, + 287, + 604 + ], + "spans": [ + { + "bbox": [ + 47, + 472, + 287, + 604 + ], + "type": "text", + "content": "2. Realistic blur and synthetic blur have significant distributional differences [24]. In the real world, due to the short exposure time of a camera, realistic blur has a specific directionality [29], i.e., the blur trajectory is regionally continuous; conversely, synthetic blur has an unnatural and discontinuous trajectory. To fully utilize the potential deblurring guidance brought by this trajectory continuity, we use 2D-DWT to reveal blur directionality, distinguish changes in the blur signal along different directions, and provide the algorithm with a reliable basis for deblurring through adaptive learning." + } + ] + } + ], + "index": 3 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 46, + 605, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 288, + 714 + ], + "type": "text", + "content": "To make 2D-DWT fit the data distribution and feature layer space more closely, we implemented 2D-DWT with an adaptive data distribution by using group convolution, transferring feature space from the spatial domain into the wavelet domain, generating sub-signals with different frequency features and different directional features, and then constructing wavelet losses for them in order to constrain them by self-supervision. Experimental results demonstrate that the proposed method has advanced performance in" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 305, + 72, + 473, + 84 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 473, + 84 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 473, + 84 + ], + "type": "text", + "content": "terms of accuracy and efficiency (Fig. 1)." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 317, + 85, + 501, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 85, + 501, + 95 + ], + "spans": [ + { + "bbox": [ + 317, + 85, + 501, + 95 + ], + "type": "text", + "content": "In summary, our contributions are as follows:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 96, + 545, + 312 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 306, + 96, + 545, + 155 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 96, + 545, + 155 + ], + "spans": [ + { + "bbox": [ + 306, + 96, + 545, + 155 + ], + "type": "text", + "content": "- We propose a single-input and multiple-output(SIMO) multi-scale baseline for progressive image deblur, which reduces the complexity of existing multi-scale deblurring algorithms and improves the overall efficiency of image restoration networks." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 156, + 545, + 215 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 156, + 545, + 215 + ], + "spans": [ + { + "bbox": [ + 306, + 156, + 545, + 215 + ], + "type": "text", + "content": "- We construct a learnable discrete wavelet transform node(LWN) to provide reliable frequency and direction selection for the algorithm, which promotes the algorithm's restoration of the high-frequency components of edges and details." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 216, + 545, + 275 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 216, + 545, + 275 + ], + "spans": [ + { + "bbox": [ + 306, + 216, + 545, + 275 + ], + "type": "text", + "content": "- For the network to work better, reasonable multi-scale loss is proposed to guide pixel-by-pixel and scale-by-scale restoration. We also created reasonable self-supervised losses for the learnable wavelet transform to limit the learning direction of the wavelet kernel." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 276, + 545, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 276, + 545, + 312 + ], + "spans": [ + { + "bbox": [ + 306, + 276, + 545, + 312 + ], + "type": "text", + "content": "- We demonstrate the effectiveness of our algorithm under several motion blur datasets, especially real datasets, and obtain highly competitive results." + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 323, + 392, + 335 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 323, + 392, + 335 + ], + "spans": [ + { + "bbox": [ + 306, + 323, + 392, + 335 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 342, + 545, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 342, + 545, + 545 + ], + "spans": [ + { + "bbox": [ + 304, + 342, + 545, + 545 + ], + "type": "text", + "content": "Single-scale Deblurring Algorithm. Image deblurring have developed rapidly in recent years [2, 10, 11, 14, 39, 40]. To preserve richer image details, Kupyn et al. [11] viewed deblurring as a special case of image-to-image conversion, and for the first time utilized a GAN to recover images with richer details. To provide a high-quality and simple baseline, Chen et al. [2] proposed a nonlinear activation-free network that obtained excellent results. Zamir et al. [38] proposed a transformer with an encoder-decoder architecture that can be applied to higher-resolution images and utilizes cross-channel attention. Kong et al. [10] utilized a domain-based self-attention solver to reduce the transformer complexity and suppress artefacts. However, such single-scale algorithms estimate complex restoration problems directly and may require the design of restorers of higher complexity, which is suboptimal in terms of efficiency [9]." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 545, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 545, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 545, + 545, + 713 + ], + "type": "text", + "content": "Multi-scale Deblurring Algorithm. Multi-scale motion deblur methods aim to achieve progressive recovery using a multi-stage or multi-sub-network approach. The multiscale algorithm previously used in the field of image deblurring is based on MIMO [3, 13, 20, 21, 27]. Recently, Cho et al. [3] proposed MIMO-UNet, which employs a multi-output single decoder as well as a multi-input single encoder to simulate a multi-scale architecture consisting of stacked networks, thereby greatly simplifying complexity. Zamir et al. [37] designed a multi-stage progressive image recovery network to simplify the information flow between multiple sub-networks. As shown in Sec. 3.1, our proposed multiscale approach further simplifies the structure of multi-scale networks and better balances the accuracy and complexity." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 732, + 315, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 732, + 315, + 742 + ], + "spans": [ + { + "bbox": [ + 294, + 732, + 315, + 742 + ], + "type": "text", + "content": "2734" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 48, + 70, + 542, + 284 + ], + "blocks": [ + { + "bbox": [ + 48, + 70, + 542, + 284 + ], + "lines": [ + { + "bbox": [ + 48, + 70, + 542, + 284 + ], + "spans": [ + { + "bbox": [ + 48, + 70, + 542, + 284 + ], + "type": "image", + "image_path": "b39c5e4e8a33272b3c101751994372a21ae473760c6f9291006c75259bca3ecf.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 289, + 546, + 323 + ], + "lines": [ + { + "bbox": [ + 46, + 289, + 546, + 323 + ], + "spans": [ + { + "bbox": [ + 46, + 289, + 546, + 323 + ], + "type": "text", + "content": "Figure 2. The overall architecture of the proposed MLWNet, the SEB is a simple module designed with reference [2], the WFB and WHB apply the LWN that implements the learnable 2D-DWT. In training phase, supervised learning is performed using " + }, + { + "bbox": [ + 46, + 289, + 546, + 323 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{multi}" + }, + { + "bbox": [ + 46, + 289, + 546, + 323 + ], + "type": "text", + "content": " and self-supervised restraint of the wavelet kernel is performed using " + }, + { + "bbox": [ + 46, + 289, + 546, + 323 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{wavelet}" + }, + { + "bbox": [ + 46, + 289, + 546, + 323 + ], + "type": "text", + "content": ". In testing phase, only the highest scale restored images is output." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 327, + 289, + 507 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 327, + 289, + 507 + ], + "spans": [ + { + "bbox": [ + 46, + 327, + 289, + 507 + ], + "type": "text", + "content": "Application of Frequency in Deblurring. Frequency is an important property of images, in recent years frequency domain has been introduced into numerous DNNs to be solved for various tasks [10, 22, 25, 32, 35, 41]. Mao et al. [19] proposed DeepRFT, which introduced a residual module based on the Fourier transform into an advanced algorithm. Kong et al. [10] proposed FFTformer, which introduced the Fourier transform into a feed-forward network to effectively utilize different frequency information. Zou et al. [42] proposed SDWNet, which utilized frequency domain information as a complement to spatial domain information. Different from these methods, we introduce a learnable 2D-DWT capable of adapting to data distribution and feature space, more suitable not only for dealing with digital images rich in mutation signals, but also for dealing with real blurring." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 519, + 153, + 533 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 519, + 153, + 533 + ], + "spans": [ + { + "bbox": [ + 47, + 519, + 153, + 533 + ], + "type": "text", + "content": "3. Proposed Method" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 540, + 287, + 660 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 540, + 287, + 660 + ], + "spans": [ + { + "bbox": [ + 46, + 540, + 287, + 660 + ], + "type": "text", + "content": "Our proposed MLWNet(Fig. 2) aims to explore an efficient multi-scale architectural approach to achieve high-quality blind deblurring of a single image. First, we designed a novel and highly scalable SIMO multi-scale baseline to address the performance bottleneck faced by partially multiscale networks. It takes a single image as an input and gradually generates a series of sharp images from the bottom up. Then, we propose a learnable wavelet transform node (LWN) for image deblurring, and it enhances the proposed algorithm's ability to restore detail information." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 670, + 164, + 681 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 670, + 164, + 681 + ], + "spans": [ + { + "bbox": [ + 47, + 670, + 164, + 681 + ], + "type": "text", + "content": "3.1. Multi-scale Baseline" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 689, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 689, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 689, + 288, + 713 + ], + "type": "text", + "content": "Our multi-scale network maintains an encoder-decoder architecture, but retains the original image with the highest" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 327, + 546, + 374 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 327, + 546, + 374 + ], + "spans": [ + { + "bbox": [ + 304, + 327, + 546, + 374 + ], + "type": "text", + "content": "resolution as input, and restores sharp images of various scales sequentially in the output stage. In this way, we eliminate the complexity of the fusion module when dealing RGB images of different scales and with deep semantics." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 376, + 546, + 532 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 376, + 546, + 532 + ], + "spans": [ + { + "bbox": [ + 304, + 376, + 546, + 532 + ], + "type": "text", + "content": "For ease of understanding, we will introduce here the Encoder phase, multi-scale semantic Fusion phase (hereafter referred to as Fusion phase), and Decoder phase sequentially. As shown in Fig. 2, the Encoder stage consists of several gray Simple Encoder Blocks(SEB), where information flows top-down for feature extraction; the Fusion stage consists of several green Wavelet Fusion Blocks(WFB), where information flows bottom-up for semantic fusion at different scales; and the Decoder stage consists of several purple Wavelet Head Blocks(WHB), and the bottom-up flow of information enables gradual restoration. Compared with WHB, SEB replaces LWN with a " + }, + { + "bbox": [ + 304, + 376, + 546, + 532 + ], + "type": "inline_equation", + "content": "1 \\times 1" + }, + { + "bbox": [ + 304, + 376, + 546, + 532 + ], + "type": "text", + "content": " conv for channel scaling, and a " + }, + { + "bbox": [ + 304, + 376, + 546, + 532 + ], + "type": "inline_equation", + "content": "3 \\times 3" + }, + { + "bbox": [ + 304, + 376, + 546, + 532 + ], + "type": "text", + "content": " depth-conv for feature extraction." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 533, + 545, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 533, + 545, + 593 + ], + "spans": [ + { + "bbox": [ + 304, + 533, + 545, + 593 + ], + "type": "text", + "content": "The Encoder stage is responsible for full feature extraction, downsampling feature maps once after each block, and then passing the feature maps to the Fusion stage or the Decoder stage as needed, whose outputs " + }, + { + "bbox": [ + 304, + 533, + 545, + 593 + ], + "type": "inline_equation", + "content": "E_{out}^{i}" + }, + { + "bbox": [ + 304, + 533, + 545, + 593 + ], + "type": "text", + "content": " to block of the " + }, + { + "bbox": [ + 304, + 533, + 545, + 593 + ], + "type": "inline_equation", + "content": "i - th" + }, + { + "bbox": [ + 304, + 533, + 545, + 593 + ], + "type": "text", + "content": " layer can be denoted as:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 339, + 604, + 545, + 632 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 339, + 604, + 545, + 632 + ], + "spans": [ + { + "bbox": [ + 339, + 604, + 545, + 632 + ], + "type": "interline_equation", + "content": "E _ {o u t} ^ {i} = \\left\\{ \\begin{array}{l l} \\phi_ {i} (e m b e d (x)), & i = i _ {m a x} \\\\ \\phi_ {i} (E _ {o u t} ^ {i - 1}), & o t h e r w i s e \\end{array} \\right. \\tag {1}", + "image_path": "e306b12c6a1859ac5bba094dc361ad142fd0322b2586ccd7ce976f82cbbee8fa.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 641, + 546, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 641, + 546, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 641, + 546, + 715 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 641, + 546, + 715 + ], + "type": "inline_equation", + "content": "\\phi_{i}" + }, + { + "bbox": [ + 304, + 641, + 546, + 715 + ], + "type": "text", + "content": " represent the block of corresponding layer of Encoder. The Fusion stage adapts and fuses information from different scales of semantics produced in Encoder to generate intermediate representations with deep semantics and shallow details. The Fusion stage's outputs " + }, + { + "bbox": [ + 304, + 641, + 546, + 715 + ], + "type": "inline_equation", + "content": "F_{out}^{i}" + }, + { + "bbox": [ + 304, + 641, + 546, + 715 + ], + "type": "text", + "content": " to the block of " + }, + { + "bbox": [ + 304, + 641, + 546, + 715 + ], + "type": "inline_equation", + "content": "i - th" + }, + { + "bbox": [ + 304, + 641, + 546, + 715 + ], + "type": "text", + "content": " layer is expressed as in Eq. 2, where " + }, + { + "bbox": [ + 304, + 641, + 546, + 715 + ], + "type": "inline_equation", + "content": "\\delta_{i}" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 733, + 315, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 733, + 315, + 742 + ], + "spans": [ + { + "bbox": [ + 294, + 733, + 315, + 742 + ], + "type": "text", + "content": "2735" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 282, + 85 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 282, + 85 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 282, + 85 + ], + "type": "text", + "content": "represents the block of the corresponding layer of Fusion." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 75, + 91, + 287, + 119 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 91, + 287, + 119 + ], + "spans": [ + { + "bbox": [ + 75, + 91, + 287, + 119 + ], + "type": "interline_equation", + "content": "F _ {o u t} ^ {i} = \\left\\{ \\begin{array}{l l} \\delta_ {i} \\left(E _ {o u t} ^ {i} + E _ {o u t} ^ {i - 1}\\right), & i = i _ {\\text {m i n}} \\\\ \\delta_ {i} \\left(F _ {o u t} ^ {i - 1} + E _ {o u t} ^ {i}\\right), & \\text {o t h e r w i s e} \\end{array} \\right. \\tag {2}", + "image_path": "6abfdda06ef194a977d504d7595fed5fa8ab107cee778265aa7e03a913cd694a.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 125, + 288, + 197 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 125, + 288, + 197 + ], + "spans": [ + { + "bbox": [ + 47, + 125, + 288, + 197 + ], + "type": "text", + "content": "The Decoder stage utilizes the information passed by both the Encoder and Fusion stages to progressively upsample the feature maps and generating pre-output feature maps at each scale separately. Output " + }, + { + "bbox": [ + 47, + 125, + 288, + 197 + ], + "type": "inline_equation", + "content": "D_{out}^{i}" + }, + { + "bbox": [ + 47, + 125, + 288, + 197 + ], + "type": "text", + "content": " to the block of the " + }, + { + "bbox": [ + 47, + 125, + 288, + 197 + ], + "type": "inline_equation", + "content": "i - th" + }, + { + "bbox": [ + 47, + 125, + 288, + 197 + ], + "type": "text", + "content": " layer is denoted as in Eq. 3. Here " + }, + { + "bbox": [ + 47, + 125, + 288, + 197 + ], + "type": "inline_equation", + "content": "\\xi_{i}" + }, + { + "bbox": [ + 47, + 125, + 288, + 197 + ], + "type": "text", + "content": " represents the block of the corresponding layer of the Decoder." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 205, + 282, + 250 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 205, + 282, + 250 + ], + "spans": [ + { + "bbox": [ + 47, + 205, + 282, + 250 + ], + "type": "interline_equation", + "content": "D _ {o u t} ^ {i} = \\left\\{ \\begin{array}{l l} \\xi_ {i} \\left(E _ {o u t} ^ {i}\\right), & i = i _ {m i n} \\\\ \\xi_ {i} \\left(D _ {o u t} ^ {i - 1} + F _ {o u t} ^ {i}\\right), & i > i _ {m i n} \\text {a n d} i < i _ {m a x} \\\\ \\xi_ {i} \\left(D _ {o u t} ^ {i - 1} + F _ {o u t} ^ {i - 1} + E _ {o u t} ^ {i}\\right), & i = i _ {m a x} \\end{array} \\right.", + "image_path": "8c7d90d9eb572e87e4ba6b25f9a7fadc3312a997b8148fc588c01f0678c87b17.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 256, + 287, + 340 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 256, + 287, + 340 + ], + "spans": [ + { + "bbox": [ + 47, + 256, + 287, + 340 + ], + "type": "text", + "content": "Each block in the Decoder is followed by a " + }, + { + "bbox": [ + 47, + 256, + 287, + 340 + ], + "type": "inline_equation", + "content": "3 \\times 3" + }, + { + "bbox": [ + 47, + 256, + 287, + 340 + ], + "type": "text", + "content": " convolution to generate a restored image at each scale. To improve inference efficiency, all but the highest-scale outputs are generated in the training phase to compute the multiscale loss. We note that this design also have a similar role with auxiliary head [36] in facilitating the model's learning of the recovered uniform patterns." + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 53, + 348, + 282, + 559 + ], + "blocks": [ + { + "bbox": [ + 53, + 348, + 282, + 559 + ], + "lines": [ + { + "bbox": [ + 53, + 348, + 282, + 559 + ], + "spans": [ + { + "bbox": [ + 53, + 348, + 282, + 559 + ], + "type": "image", + "image_path": "c17db24264ae3d97787ae53f7a4da5398aabfa3989e2231c56335c5c045ed4cb.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 47, + 566, + 286, + 588 + ], + "lines": [ + { + "bbox": [ + 47, + 566, + 286, + 588 + ], + "spans": [ + { + "bbox": [ + 47, + 566, + 286, + 588 + ], + "type": "text", + "content": "Figure 3. (a) The process of learnable 2D-wavelet convolution. (b) The construction process of the " + }, + { + "bbox": [ + 47, + 566, + 286, + 588 + ], + "type": "inline_equation", + "content": "N \\times N" + }, + { + "bbox": [ + 47, + 566, + 286, + 588 + ], + "type": "text", + "content": " 2D-wavelet kernel." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 599, + 252, + 611 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 599, + 252, + 611 + ], + "spans": [ + { + "bbox": [ + 47, + 599, + 252, + 611 + ], + "type": "text", + "content": "3.2. Learnable Discrete Wavelet Transform" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 618, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 618, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 618, + 287, + 713 + ], + "type": "text", + "content": "To alleviate the defects of multi-scale architectures in the recovery of detail information, we design LWN and constructed the WFB and WHB based on it. Unlike the Fourier transform, the wavelet transform has adaptive time-frequency resolution, and performs excellently on digital images rich in mutation signal. Its use in conjunction with DNNs has gained increasing widespread attention [7, 16, 18, 34]." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 72, + 546, + 137 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 546, + 137 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 546, + 137 + ], + "type": "text", + "content": "For ease of understanding, the 1D-DWT case is given here first. For the input discrete signal " + }, + { + "bbox": [ + 304, + 72, + 546, + 137 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 72, + 546, + 137 + ], + "type": "text", + "content": ", given the wavelet function " + }, + { + "bbox": [ + 304, + 72, + 546, + 137 + ], + "type": "inline_equation", + "content": "\\psi_{j,k}(t) = 2^{\\frac{j}{2}}\\psi (2^{j}t - k)" + }, + { + "bbox": [ + 304, + 72, + 546, + 137 + ], + "type": "text", + "content": " for scaling factor " + }, + { + "bbox": [ + 304, + 72, + 546, + 137 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 304, + 72, + 546, + 137 + ], + "type": "text", + "content": ", time factor " + }, + { + "bbox": [ + 304, + 72, + 546, + 137 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 72, + 546, + 137 + ], + "type": "text", + "content": ", and scale function " + }, + { + "bbox": [ + 304, + 72, + 546, + 137 + ], + "type": "inline_equation", + "content": "\\phi_{j,k}(t) = 2^{\\frac{j}{2}}\\phi (2^{j}t - k)" + }, + { + "bbox": [ + 304, + 72, + 546, + 137 + ], + "type": "text", + "content": ", the decomposition of the input signal " + }, + { + "bbox": [ + 304, + 72, + 546, + 137 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 72, + 546, + 137 + ], + "type": "text", + "content": " at " + }, + { + "bbox": [ + 304, + 72, + 546, + 137 + ], + "type": "inline_equation", + "content": "j_0" + }, + { + "bbox": [ + 304, + 72, + 546, + 137 + ], + "type": "text", + "content": " is given by:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 323, + 144, + 546, + 171 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 323, + 144, + 546, + 171 + ], + "spans": [ + { + "bbox": [ + 323, + 144, + 546, + 171 + ], + "type": "interline_equation", + "content": "f (t) = \\sum_ {j > j _ {0}} \\sum_ {k} d _ {j, k} \\psi_ {j, k} (t) + \\sum_ {k} c _ {j _ {0}, k} \\phi_ {j _ {0}, k} (t) \\tag {3}", + "image_path": "157611728a9eaf5ec27f74f69ca2f59cbdf5a7b26ae06ebeec05f689dfc42383.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 181, + 547, + 301 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 181, + 547, + 301 + ], + "spans": [ + { + "bbox": [ + 304, + 181, + 547, + 301 + ], + "type": "text", + "content": "In this process, " + }, + { + "bbox": [ + 304, + 181, + 547, + 301 + ], + "type": "inline_equation", + "content": "d_{j,k} = \\langle f(t),\\psi_{j,k}(t)\\rangle" + }, + { + "bbox": [ + 304, + 181, + 547, + 301 + ], + "type": "text", + "content": " represents the detail coefficients (i.e., high-frequency components), " + }, + { + "bbox": [ + 304, + 181, + 547, + 301 + ], + "type": "inline_equation", + "content": "c_{j0,k} = \\langle f(t),\\phi_{j0,k}(t)\\rangle" + }, + { + "bbox": [ + 304, + 181, + 547, + 301 + ], + "type": "text", + "content": " represents the approximation coefficients (i.e., low-frequency components) so that the input signal can be disassembled step by step into a rich wavelet domain signal. In order to use wavelet transform in DNN, we introduce the analysis vectors " + }, + { + "bbox": [ + 304, + 181, + 547, + 301 + ], + "type": "inline_equation", + "content": "\\vec{a}_0[k] = \\left\\langle \\frac{1}{\\sqrt{2}}\\phi \\left(\\frac{t}{2}\\right),\\phi (t - k)\\right\\rangle ," + }, + { + "bbox": [ + 304, + 181, + 547, + 301 + ], + "type": "inline_equation", + "content": "\\vec{a}_{1}[k] = \\left\\langle \\frac{1}{\\sqrt{2}}\\psi \\left(\\frac{t}{2}\\right),\\phi (t - k)\\right\\rangle" + }, + { + "bbox": [ + 304, + 181, + 547, + 301 + ], + "type": "text", + "content": " to represent the high and low frequency filters respectively, then:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 367, + 310, + 484, + 332 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 367, + 310, + 484, + 332 + ], + "spans": [ + { + "bbox": [ + 367, + 310, + 484, + 332 + ], + "type": "interline_equation", + "content": "c _ {j + 1, p} = \\sum_ {k} \\vec {a} _ {0} [ k - 2 p ] c _ {j, k}", + "image_path": "0d6ae453e61a36ed8b8c0b7798c95b36569dd72e0eaee6fb48340ddc3fc52da9.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 368, + 331, + 545, + 362 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 368, + 331, + 545, + 362 + ], + "spans": [ + { + "bbox": [ + 368, + 331, + 545, + 362 + ], + "type": "interline_equation", + "content": "d _ {j + 1, p} = \\sum_ {k} ^ {\\kappa} \\vec {a} _ {1} [ k - 2 p ] c _ {j, k} \\tag {4}", + "image_path": "41e9f81cefa32552879d67fc10193b4bb852e39f3c9a8be847b7c349abdceb6f.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 372, + 546, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 372, + 546, + 515 + ], + "spans": [ + { + "bbox": [ + 304, + 372, + 546, + 515 + ], + "type": "text", + "content": "According to Eq. 4, the decomposition of the original signal over the wavelet basis can be viewed as a recursive convolution of the original signal with a specific filter using a step size of 2. Similarly, the inverse wavelet transform can be realized by transposed convolution of two synthetic vectors " + }, + { + "bbox": [ + 304, + 372, + 546, + 515 + ], + "type": "inline_equation", + "content": "\\vec{s}_0" + }, + { + "bbox": [ + 304, + 372, + 546, + 515 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 372, + 546, + 515 + ], + "type": "inline_equation", + "content": "\\vec{s}_1" + }, + { + "bbox": [ + 304, + 372, + 546, + 515 + ], + "type": "text", + "content": ". We extend the above situation to 2D by efficiently extracting axial high-frequency information using 2D discrete wavelet transform (2D-DWT), and designing learnable methods to make it adaptive to the data distribution and feature layers. Fig. 3(b) presents the construction process of the forward wavelet convolution kernel in 2D, which can be expressed as:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 356, + 524, + 495, + 538 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 356, + 524, + 495, + 538 + ], + "spans": [ + { + "bbox": [ + 356, + 524, + 495, + 538 + ], + "type": "interline_equation", + "content": "\\mathcal {F} _ {l l} = \\vec {a _ {0}} \\times \\vec {a _ {0}} ^ {T}, \\mathcal {F} _ {l h} = \\vec {a _ {0}} \\times \\vec {a _ {1}} ^ {T}", + "image_path": "5a90b8358db1787709ad9c57ebf45e0c6114ee5425cce154186082128daf8a59.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 355, + 541, + 545, + 555 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 355, + 541, + 545, + 555 + ], + "spans": [ + { + "bbox": [ + 355, + 541, + 545, + 555 + ], + "type": "interline_equation", + "content": "\\mathcal {F} _ {h l} = \\vec {a _ {1}} \\times \\vec {a _ {0}} ^ {T}, \\mathcal {F} _ {h h} = \\vec {a _ {1}} \\times \\vec {a _ {1}} ^ {T} \\tag {5}", + "image_path": "9eeb040967c18f0d486a0f0b2753b192a913749480d70bcea24fc206a0a77292.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 356, + 559, + 480, + 571 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 356, + 559, + 480, + 571 + ], + "spans": [ + { + "bbox": [ + 356, + 559, + 480, + 571 + ], + "type": "interline_equation", + "content": "\\mathcal {K} _ {w} = \\operatorname {c a t} \\left(\\mathcal {F} _ {l l}, \\mathcal {F} _ {l h}, \\mathcal {F} _ {h l}, \\mathcal {F} _ {h h}\\right)", + "image_path": "953bddc582b6210ea0cf7b0c2ab72967dcb2fc2ce3812c9292edff722ac7dee3.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 581, + 546, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 581, + 546, + 641 + ], + "spans": [ + { + "bbox": [ + 304, + 581, + 546, + 641 + ], + "type": "text", + "content": "We set " + }, + { + "bbox": [ + 304, + 581, + 546, + 641 + ], + "type": "inline_equation", + "content": "\\vec{a_0}" + }, + { + "bbox": [ + 304, + 581, + 546, + 641 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 581, + 546, + 641 + ], + "type": "inline_equation", + "content": "\\vec{a_1}" + }, + { + "bbox": [ + 304, + 581, + 546, + 641 + ], + "type": "text", + "content": " as learnable filters; " + }, + { + "bbox": [ + 304, + 581, + 546, + 641 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{ll}" + }, + { + "bbox": [ + 304, + 581, + 546, + 641 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 581, + 546, + 641 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{lh}" + }, + { + "bbox": [ + 304, + 581, + 546, + 641 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 581, + 546, + 641 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{hl}" + }, + { + "bbox": [ + 304, + 581, + 546, + 641 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 581, + 546, + 641 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{hh}" + }, + { + "bbox": [ + 304, + 581, + 546, + 641 + ], + "type": "text", + "content": " are low-frequency, horizontal high-frequency, vertical high-frequency, and diagonal high-frequency convolution operators obtained from vector outer products; The wavelet kernel " + }, + { + "bbox": [ + 304, + 581, + 546, + 641 + ], + "type": "inline_equation", + "content": "\\kappa_w" + }, + { + "bbox": [ + 304, + 581, + 546, + 641 + ], + "type": "text", + "content": " is spliced from above convolution operators." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 642, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 642, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 642, + 547, + 713 + ], + "type": "text", + "content": "Then we implement learnable wavelet transform as a group convolution like Fig. 3(a), given a set of input feature maps " + }, + { + "bbox": [ + 304, + 642, + 547, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{X}_{in} \\in (C, H, W)" + }, + { + "bbox": [ + 304, + 642, + 547, + 713 + ], + "type": "text", + "content": ", its projection in wavelet domain " + }, + { + "bbox": [ + 304, + 642, + 547, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{X}_{out} \\in \\left(4C, \\frac{H}{2}, \\frac{W}{2}\\right)" + }, + { + "bbox": [ + 304, + 642, + 547, + 713 + ], + "type": "text", + "content": " (low-frequency, horizontal high-frequency, vertical high-frequency, and diagonal high-frequency component) is generated through the wavelet" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 732, + 315, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 732, + 315, + 742 + ], + "spans": [ + { + "bbox": [ + 294, + 732, + 315, + 742 + ], + "type": "text", + "content": "2736" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 108 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 108 + ], + "type": "text", + "content": "convolution. Similarly, the construction method of the inverse kernel and the forward propagation can be easily deduced, since they are inverse processes of each other." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 108, + 287, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 108, + 287, + 156 + ], + "spans": [ + { + "bbox": [ + 46, + 108, + 287, + 156 + ], + "type": "text", + "content": "Then, a natural question can then be posed: how to ensure the correctness of the wavelet kernel learning and ensure that it does not degrade into a general group convolution and cause performance degradation?" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 156, + 287, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 156, + 287, + 240 + ], + "spans": [ + { + "bbox": [ + 46, + 156, + 287, + 240 + ], + "type": "text", + "content": "A common practice is to introduce the principle of perfect reconstruction [16, 34] to constrain adaptive wavelets, for a complex " + }, + { + "bbox": [ + 46, + 156, + 287, + 240 + ], + "type": "inline_equation", + "content": "z \\in \\mathbb{C}" + }, + { + "bbox": [ + 46, + 156, + 287, + 240 + ], + "type": "text", + "content": ", given a filter " + }, + { + "bbox": [ + 46, + 156, + 287, + 240 + ], + "type": "inline_equation", + "content": "\\vec{x}" + }, + { + "bbox": [ + 46, + 156, + 287, + 240 + ], + "type": "text", + "content": " that applies to the " + }, + { + "bbox": [ + 46, + 156, + 287, + 240 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 46, + 156, + 287, + 240 + ], + "type": "text", + "content": "-transform, its " + }, + { + "bbox": [ + 46, + 156, + 287, + 240 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 46, + 156, + 287, + 240 + ], + "type": "text", + "content": "-transform can be expressed as " + }, + { + "bbox": [ + 46, + 156, + 287, + 240 + ], + "type": "inline_equation", + "content": "X(z) = \\sum_{n \\in \\mathbb{Z}} \\vec{x}(n) z^{-n}" + }, + { + "bbox": [ + 46, + 156, + 287, + 240 + ], + "type": "text", + "content": ". Then " + }, + { + "bbox": [ + 46, + 156, + 287, + 240 + ], + "type": "inline_equation", + "content": "\\vec{a}_0, \\vec{a}_1, \\vec{s}_0, \\vec{s}_1" + }, + { + "bbox": [ + 46, + 156, + 287, + 240 + ], + "type": "text", + "content": " can be obtained as their corresponding " + }, + { + "bbox": [ + 46, + 156, + 287, + 240 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 46, + 156, + 287, + 240 + ], + "type": "text", + "content": "-transforms " + }, + { + "bbox": [ + 46, + 156, + 287, + 240 + ], + "type": "inline_equation", + "content": "A_0, A_1, S_0, S_1" + }, + { + "bbox": [ + 46, + 156, + 287, + 240 + ], + "type": "text", + "content": ", which must then be satisfied if perfect reconstruction is desired:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 94, + 250, + 287, + 263 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 94, + 250, + 287, + 263 + ], + "spans": [ + { + "bbox": [ + 94, + 250, + 287, + 263 + ], + "type": "interline_equation", + "content": "A _ {0} (- z) S _ {0} (z) + A _ {1} (- z) S _ {1} (z) = 0 \\tag {6}", + "image_path": "741eb8aa00d92d6c946b4dd5e1ac80b7b2e66db4c3796c35d4e8dbb19a85dc61.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 102, + 284, + 287, + 297 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 102, + 284, + 287, + 297 + ], + "spans": [ + { + "bbox": [ + 102, + 284, + 287, + 297 + ], + "type": "interline_equation", + "content": "A _ {0} (z) S _ {0} (z) + A _ {1} (z) S _ {1} (z) = 2 \\tag {7}", + "image_path": "843209025724e2a9da2bfa219233fad796f3e58e9c46e2a6aca46a1c92648d50.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 303, + 287, + 351 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 303, + 287, + 351 + ], + "spans": [ + { + "bbox": [ + 46, + 303, + 287, + 351 + ], + "type": "text", + "content": "Eq. 6 is known as aliasing cancellation condition, which is used to the cancellation of the aliasing effects arising from the downsampling. We will give a detailed derivation of Eq. 6 and Eq. 7 in the Appendix." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 351, + 287, + 531 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 351, + 287, + 531 + ], + "spans": [ + { + "bbox": [ + 46, + 351, + 287, + 531 + ], + "type": "text", + "content": "After wavelet convolution, to achieve frequency restoration in the wavelet domain, the input " + }, + { + "bbox": [ + 46, + 351, + 287, + 531 + ], + "type": "inline_equation", + "content": "\\mathcal{X}" + }, + { + "bbox": [ + 46, + 351, + 287, + 531 + ], + "type": "text", + "content": " wavelet domain component was separated into a separate dim, we use depth-wise convolution with an expansion factor of " + }, + { + "bbox": [ + 46, + 351, + 287, + 531 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 46, + 351, + 287, + 531 + ], + "type": "text", + "content": " for wavelet domain feature extraction and transformation after the wavelet positive transform, and " + }, + { + "bbox": [ + 46, + 351, + 287, + 531 + ], + "type": "inline_equation", + "content": "1\\times 1" + }, + { + "bbox": [ + 46, + 351, + 287, + 531 + ], + "type": "text", + "content": " convolution for channel expansion and scaling, and finally a learnable wavelet inverse transform is used to reduce the wavelet domain feature maps to the spatial domain for output, which constitutes the LWN. We follow the rules of [2] and design structurally similar Wavelet Head Block (WHB) and Wavelet Fusion Block (WFB), both with LWN as an important base module, but used for semantic aggregation and multiscale output at different scales, respectively; thus, WHB adds the recovery convolution branch after WFB." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 538, + 91, + 550 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 538, + 91, + 550 + ], + "spans": [ + { + "bbox": [ + 47, + 538, + 91, + 550 + ], + "type": "text", + "content": "3.3. Loss" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 556, + 287, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 556, + 287, + 605 + ], + "spans": [ + { + "bbox": [ + 46, + 556, + 287, + 605 + ], + "type": "text", + "content": "Multi-scale Loss. Follow PSNR loss [2], we propose multiscale loss function as the main loss of the algorithm for calculating the pixel difference between restored image and GT at each scale:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 83, + 612, + 287, + 645 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 83, + 612, + 287, + 645 + ], + "spans": [ + { + "bbox": [ + 83, + 612, + 287, + 645 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {m u l t i}} (x, y) = \\sum_ {k = 1} ^ {K} w _ {k} \\times \\mathcal {L} _ {\\text {p s n r}} \\left(x _ {k}, y _ {k}\\right) \\tag {8}", + "image_path": "ac385b1edcef887e3d424a664fd9793b47e993f7a58912d1b0ec16fa75165a78.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 653, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 287, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 653, + 287, + 714 + ], + "type": "inline_equation", + "content": "x, y" + }, + { + "bbox": [ + 46, + 653, + 287, + 714 + ], + "type": "text", + "content": " denote the output and GT, respectively, " + }, + { + "bbox": [ + 46, + 653, + 287, + 714 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 653, + 287, + 714 + ], + "type": "text", + "content": " denotes the downsampling level, " + }, + { + "bbox": [ + 46, + 653, + 287, + 714 + ], + "type": "inline_equation", + "content": "w_{k}" + }, + { + "bbox": [ + 46, + 653, + 287, + 714 + ], + "type": "text", + "content": " represents the weights under the corresponding scale. Our design of " + }, + { + "bbox": [ + 46, + 653, + 287, + 714 + ], + "type": "inline_equation", + "content": "w_{k}" + }, + { + "bbox": [ + 46, + 653, + 287, + 714 + ], + "type": "text", + "content": " is based on the simple finding that we cannot produce absolutely accurate downsampled clear images, and that the accuracy gap" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "text", + "content": "is progressively enlarged with decreasing scales. To ensure that lower scale output layers do not negatively affect the final output, we empirically set the loss weight " + }, + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "inline_equation", + "content": "w_{k}" + }, + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "inline_equation", + "content": "\\frac{1}{k}" + }, + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "text", + "content": " for the corresponding k-scale." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 120, + 545, + 192 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 120, + 545, + 192 + ], + "spans": [ + { + "bbox": [ + 304, + 120, + 545, + 192 + ], + "type": "text", + "content": "Wavelet Loss. Both Eq. 6 and Eq. 7 can be easily converted losses optimized toward a minimum of 0, here we use the mean square error. Then, based on the convolvable nature of the " + }, + { + "bbox": [ + 304, + 120, + 545, + 192 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 304, + 120, + 545, + 192 + ], + "type": "text", + "content": " transform, it is possible to construct convolutionally equivalent substitutions for polynomial multiplications for the perfect reconstruction condition:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 200, + 566, + 282 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 200, + 566, + 282 + ], + "spans": [ + { + "bbox": [ + 306, + 200, + 566, + 282 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} L _ {w a v e l e t} (\\theta_ {i}) = \\left(\\sum_ {k} ^ {N - 1} \\left(\\langle a _ {0}, s _ {0} \\rangle_ {k} + \\langle a _ {1}, s _ {1} \\rangle_ {k}\\right) - \\hat {V} _ {\\lfloor \\frac {N}{2} \\rfloor}\\right) ^ {2} \\\\ = \\left(\\sum_ {k} ^ {N - 1} \\left(\\left\\langle (- 1) ^ {k} a _ {0}, s _ {0} \\right\\rangle_ {k} + \\left\\langle (- 1) ^ {k} a _ {1}, s _ {1} \\right\\rangle_ {k}\\right)\\right) ^ {2} \\tag {9} \\\\ \\end{array}", + "image_path": "ad6183c13fb1daa5b801274bc5e36562c1ba264f4309694b3a859e95b467e7c2.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 283, + 545, + 318 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 283, + 545, + 318 + ], + "spans": [ + { + "bbox": [ + 304, + 283, + 545, + 318 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 283, + 545, + 318 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 283, + 545, + 318 + ], + "type": "text", + "content": " denotes the position of the filter when it performs the convolution, " + }, + { + "bbox": [ + 304, + 283, + 545, + 318 + ], + "type": "inline_equation", + "content": "\\hat{V}" + }, + { + "bbox": [ + 304, + 283, + 545, + 318 + ], + "type": "text", + "content": " is a vector with a center position value of two, and " + }, + { + "bbox": [ + 304, + 283, + 545, + 318 + ], + "type": "inline_equation", + "content": "\\theta_{i}" + }, + { + "bbox": [ + 304, + 283, + 545, + 318 + ], + "type": "text", + "content": " denotes the constructed filter" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 319, + 545, + 342 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 319, + 545, + 342 + ], + "spans": [ + { + "bbox": [ + 304, + 319, + 545, + 342 + ], + "type": "text", + "content": "Overall Loss. Based on the multi-scale loss and wavelet loss, the loss function used in this study is:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 328, + 352, + 545, + 365 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 328, + 352, + 545, + 365 + ], + "spans": [ + { + "bbox": [ + 328, + 352, + 545, + 365 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {t o t a l}} (x, y) = \\mathcal {L} _ {\\text {w a v e l e t}} \\left(\\theta_ {i}\\right) + \\mathcal {L} _ {\\text {m u l t i}} (x, y) \\tag {10}", + "image_path": "ee10098572c0c8f062edbf2d6e0de9d4bfcc667f5bbfa89b3ed7e74ce9bca653.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 305, + 374, + 432, + 387 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 374, + 432, + 387 + ], + "spans": [ + { + "bbox": [ + 305, + 374, + 432, + 387 + ], + "type": "text", + "content": "4. Experimental Results" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 305, + 394, + 501, + 407 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 394, + 501, + 407 + ], + "spans": [ + { + "bbox": [ + 305, + 394, + 501, + 407 + ], + "type": "text", + "content": "4.1. Datasets and Implementation Details" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 412, + 545, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 412, + 545, + 533 + ], + "spans": [ + { + "bbox": [ + 304, + 412, + 545, + 533 + ], + "type": "text", + "content": "We evaluate our method on the realistic datasets Real-Blur [23] and RSBur [24], as well as the synthetic dataset GoPro [20], they are all composed of blurry-sharp image pairs. During training we use AdamW optimizer " + }, + { + "bbox": [ + 304, + 412, + 545, + 533 + ], + "type": "inline_equation", + "content": "(\\alpha = 0.9" + }, + { + "bbox": [ + 304, + 412, + 545, + 533 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 412, + 545, + 533 + ], + "type": "inline_equation", + "content": "\\beta = 0.9)" + }, + { + "bbox": [ + 304, + 412, + 545, + 533 + ], + "type": "text", + "content": " for a total of " + }, + { + "bbox": [ + 304, + 412, + 545, + 533 + ], + "type": "inline_equation", + "content": "600\\mathrm{K}" + }, + { + "bbox": [ + 304, + 412, + 545, + 533 + ], + "type": "text", + "content": " iterations. The initial value of the learning rate is " + }, + { + "bbox": [ + 304, + 412, + 545, + 533 + ], + "type": "inline_equation", + "content": "10^{-3}" + }, + { + "bbox": [ + 304, + 412, + 545, + 533 + ], + "type": "text", + "content": " and is updated with cosine annealing schedule. The patch size is set to " + }, + { + "bbox": [ + 304, + 412, + 545, + 533 + ], + "type": "inline_equation", + "content": "256\\times 256" + }, + { + "bbox": [ + 304, + 412, + 545, + 533 + ], + "type": "text", + "content": " pixels and we followed the training strategy of [28, 38]. For data augmentation, we only use random flips and rotations. More experimental results can be found in the Appendix." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 305, + 540, + 531, + 552 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 540, + 531, + 552 + ], + "spans": [ + { + "bbox": [ + 305, + 540, + 531, + 552 + ], + "type": "text", + "content": "4.2. Comparison with State-of-the-Art Methods" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 558, + 545, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 558, + 545, + 594 + ], + "spans": [ + { + "bbox": [ + 304, + 558, + 545, + 594 + ], + "type": "text", + "content": "We have compared our method with several advanced deblurring methods and use the PSNR and SSIM to evaluate the quality of restored images." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "text", + "content": "**Evaluations on the RealBlur dataset.** Since our proposed method is driven by real-world deblurring, we first conduct comparisons on RealBlur [23]. The quantitative analysis results are shown in Tab. 1, the PSNR value of our method is 0.91dB and 0.49dB higher than state-of-the-art GRL on the RealBlur-J and RealBlur-R datasets respectively. Compared to other multi-scale architectures, our method has fewer model computational and running time while the performance is better. Fig. 4 shows visual comparison on RealBlur-J dataset, our method obtains restored text with" + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 733, + 315, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 733, + 315, + 742 + ], + "spans": [ + { + "bbox": [ + 294, + 733, + 315, + 742 + ], + "type": "text", + "content": "2737" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 49, + 70, + 190, + 144 + ], + "blocks": [ + { + "bbox": [ + 49, + 70, + 190, + 144 + ], + "lines": [ + { + "bbox": [ + 49, + 70, + 190, + 144 + ], + "spans": [ + { + "bbox": [ + 49, + 70, + 190, + 144 + ], + "type": "image", + "image_path": "215bdb7cdeb1601782a476847e66c891bd797b8de11ef519a88050281aeb185b.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 94, + 146, + 141, + 156 + ], + "lines": [ + { + "bbox": [ + 94, + 146, + 141, + 156 + ], + "spans": [ + { + "bbox": [ + 94, + 146, + 141, + 156 + ], + "type": "text", + "content": "Blurry image" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 191, + 70, + 260, + 144 + ], + "blocks": [ + { + "bbox": [ + 191, + 70, + 260, + 144 + ], + "lines": [ + { + "bbox": [ + 191, + 70, + 260, + 144 + ], + "spans": [ + { + "bbox": [ + 191, + 70, + 260, + 144 + ], + "type": "image", + "image_path": "f9f64d806f80acd40cacf204f51bef984d3f170b07eba1240a1638ff150c106c.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 199, + 146, + 254, + 155 + ], + "lines": [ + { + "bbox": [ + 199, + 146, + 254, + 155 + ], + "spans": [ + { + "bbox": [ + 199, + 146, + 254, + 155 + ], + "type": "text", + "content": "DeblurGAN-v2" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 262, + 70, + 332, + 144 + ], + "blocks": [ + { + "bbox": [ + 262, + 70, + 332, + 144 + ], + "lines": [ + { + "bbox": [ + 262, + 70, + 332, + 144 + ], + "spans": [ + { + "bbox": [ + 262, + 70, + 332, + 144 + ], + "type": "image", + "image_path": "f308de45c37be92ad81de0c16181cfe61669b994af9a9e6b40a08c2b9926d208.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 288, + 146, + 306, + 155 + ], + "lines": [ + { + "bbox": [ + 288, + 146, + 306, + 155 + ], + "spans": [ + { + "bbox": [ + 288, + 146, + 306, + 155 + ], + "type": "text", + "content": "SRN" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 334, + 70, + 403, + 144 + ], + "blocks": [ + { + "bbox": [ + 334, + 70, + 403, + 144 + ], + "lines": [ + { + "bbox": [ + 334, + 70, + 403, + 144 + ], + "spans": [ + { + "bbox": [ + 334, + 70, + 403, + 144 + ], + "type": "image", + "image_path": "717ebaf86742e707eed5b75d11d5d6117c6f0b90dd9a2492ca358dadd843c627.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 343, + 146, + 395, + 155 + ], + "lines": [ + { + "bbox": [ + 343, + 146, + 395, + 155 + ], + "spans": [ + { + "bbox": [ + 343, + 146, + 395, + 155 + ], + "type": "text", + "content": "MIMO-UNet+" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 405, + 70, + 474, + 144 + ], + "blocks": [ + { + "bbox": [ + 405, + 70, + 474, + 144 + ], + "lines": [ + { + "bbox": [ + 405, + 70, + 474, + 144 + ], + "spans": [ + { + "bbox": [ + 405, + 70, + 474, + 144 + ], + "type": "image", + "image_path": "16cd1a14bbe99a59e5358466b6f42f140bcd8f6bf54d673e7c4d748a255b8253.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 425, + 146, + 457, + 155 + ], + "lines": [ + { + "bbox": [ + 425, + 146, + 457, + 155 + ], + "spans": [ + { + "bbox": [ + 425, + 146, + 457, + 155 + ], + "type": "text", + "content": "MPRNet" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 476, + 70, + 545, + 144 + ], + "blocks": [ + { + "bbox": [ + 476, + 70, + 545, + 144 + ], + "lines": [ + { + "bbox": [ + 476, + 70, + 545, + 144 + ], + "spans": [ + { + "bbox": [ + 476, + 70, + 545, + 144 + ], + "type": "image", + "image_path": "e6f347886090fcb55653e1237bbf23fe901583ab80271d6d3c655228a5af0c24.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 499, + 146, + 524, + 155 + ], + "lines": [ + { + "bbox": [ + 499, + 146, + 524, + 155 + ], + "spans": [ + { + "bbox": [ + 499, + 146, + 524, + 155 + ], + "type": "text", + "content": "BANet" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 49, + 158, + 119, + 232 + ], + "blocks": [ + { + "bbox": [ + 49, + 158, + 119, + 232 + ], + "lines": [ + { + "bbox": [ + 49, + 158, + 119, + 232 + ], + "spans": [ + { + "bbox": [ + 49, + 158, + 119, + 232 + ], + "type": "image", + "image_path": "3daa0d4338f14c248d325bd2036e8037fa6b5037d2e3e76130dfcfab43938f88.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 71, + 233, + 96, + 243 + ], + "lines": [ + { + "bbox": [ + 71, + 233, + 96, + 243 + ], + "spans": [ + { + "bbox": [ + 71, + 233, + 96, + 243 + ], + "type": "text", + "content": "Blurry" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 120, + 158, + 190, + 232 + ], + "blocks": [ + { + "bbox": [ + 120, + 158, + 190, + 232 + ], + "lines": [ + { + "bbox": [ + 120, + 158, + 190, + 232 + ], + "spans": [ + { + "bbox": [ + 120, + 158, + 190, + 232 + ], + "type": "image", + "image_path": "57b0eb1ced75505fc522cc0f50c80a5e1966062e2ba733ce7ab57b1efcd8d607.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 148, + 233, + 160, + 242 + ], + "lines": [ + { + "bbox": [ + 148, + 233, + 160, + 242 + ], + "spans": [ + { + "bbox": [ + 148, + 233, + 160, + 242 + ], + "type": "text", + "content": "GT" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 191, + 158, + 260, + 232 + ], + "blocks": [ + { + "bbox": [ + 191, + 158, + 260, + 232 + ], + "lines": [ + { + "bbox": [ + 191, + 158, + 260, + 232 + ], + "spans": [ + { + "bbox": [ + 191, + 158, + 260, + 232 + ], + "type": "image", + "image_path": "31dc3d903701c53a32ed339747e77585c5d667cc143e09e70832695e88bcc0d9.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 213, + 233, + 244, + 242 + ], + "lines": [ + { + "bbox": [ + 213, + 233, + 244, + 242 + ], + "spans": [ + { + "bbox": [ + 213, + 233, + 244, + 242 + ], + "type": "text", + "content": "MSSNet" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 262, + 158, + 332, + 232 + ], + "blocks": [ + { + "bbox": [ + 262, + 158, + 332, + 232 + ], + "lines": [ + { + "bbox": [ + 262, + 158, + 332, + 232 + ], + "spans": [ + { + "bbox": [ + 262, + 158, + 332, + 232 + ], + "type": "image", + "image_path": "35e24b695b5265712dd11e27b6eeb0cd771f581ee3696a63e5337195c88d06f4.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 276, + 233, + 317, + 243 + ], + "lines": [ + { + "bbox": [ + 276, + 233, + 317, + 243 + ], + "spans": [ + { + "bbox": [ + 276, + 233, + 317, + 243 + ], + "type": "text", + "content": "Stripformer" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 67, + 244, + 525, + 255 + ], + "lines": [ + { + "bbox": [ + 67, + 244, + 525, + 255 + ], + "spans": [ + { + "bbox": [ + 67, + 244, + 525, + 255 + ], + "type": "text", + "content": "Figure 4. Visual comparisons on the RealBlur-J dataset [23]. The proposed method generates an image with clearer characters." + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 333, + 158, + 402, + 232 + ], + "blocks": [ + { + "bbox": [ + 333, + 158, + 402, + 232 + ], + "lines": [ + { + "bbox": [ + 333, + 158, + 402, + 232 + ], + "spans": [ + { + "bbox": [ + 333, + 158, + 402, + 232 + ], + "type": "image", + "image_path": "200900a46d8bfd7a0cc96759283ccb127c4c9378eaf6d1cd57f57ae3fc93d4e8.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 360, + 233, + 378, + 242 + ], + "lines": [ + { + "bbox": [ + 360, + 233, + 378, + 242 + ], + "spans": [ + { + "bbox": [ + 360, + 233, + 378, + 242 + ], + "type": "text", + "content": "GRL" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 404, + 158, + 474, + 232 + ], + "blocks": [ + { + "bbox": [ + 404, + 158, + 474, + 232 + ], + "lines": [ + { + "bbox": [ + 404, + 158, + 474, + 232 + ], + "spans": [ + { + "bbox": [ + 404, + 158, + 474, + 232 + ], + "type": "image", + "image_path": "37eaff3d66cc31ceebaa60029b2c839c63842d72a257f26c2eac1f5a6833f587.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 420, + 233, + 460, + 242 + ], + "lines": [ + { + "bbox": [ + 420, + 233, + 460, + 242 + ], + "spans": [ + { + "bbox": [ + 420, + 233, + 460, + 242 + ], + "type": "text", + "content": "FFTformer" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 476, + 158, + 545, + 232 + ], + "blocks": [ + { + "bbox": [ + 476, + 158, + 545, + 232 + ], + "lines": [ + { + "bbox": [ + 476, + 158, + 545, + 232 + ], + "spans": [ + { + "bbox": [ + 476, + 158, + 545, + 232 + ], + "type": "image", + "image_path": "46b70589e14259f5b5eb2ab46222bacc8ccb610982af6f74ab8c8988e938aa3b.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 503, + 233, + 522, + 242 + ], + "lines": [ + { + "bbox": [ + 503, + 233, + 522, + 242 + ], + "spans": [ + { + "bbox": [ + 503, + 233, + 522, + 242 + ], + "type": "text", + "content": "Ours" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + } + ], + "index": 24 + }, + { + "type": "table", + "bbox": [ + 48, + 266, + 288, + 511 + ], + "blocks": [ + { + "bbox": [ + 48, + 266, + 288, + 511 + ], + "lines": [ + { + "bbox": [ + 48, + 266, + 288, + 511 + ], + "spans": [ + { + "bbox": [ + 48, + 266, + 288, + 511 + ], + "type": "table", + "html": "
MethodRealBlur-JRealBlur-RAvg. runtime
PSNRSSIMPSNRSSIM
DeblurGAN-v2 [12]29.690.87036.440.9350.04s
SRN [27]31.380.90938.650.9650.07s
MPRNet [37]31.760.92239.310.9720.09s
SDWNet [42]30.730.89638.210.9630.04s
MIMO-UNet+ [3]31.920.919--0.02s
MIMO-UNet++ [3]32.050.921---
DeepRFT+ [19]32.190.93139.840.9720.09s
BANet [29]32.000.92339.550.9710.06s
BANet+ [29]32.420.92939.900.9720.12s
Stripformer [28]32.480.92939.840.9740.04s
MSSNet [9]32.100.92839.760.9720.06s
MSDI-Net [13]32.350.923--0.06s
MAXIM-3S [30]32.840.93539.450.962-
FFTformer [10]32.620.93340.110.9730.13s
GRL-B [14]32.820.93240.200.9741.28s
MLWNet-S33.020.933--0.04s
MLWNet-B33.840.94140.690.9760.05s
", + "image_path": "ba7114fd2f35369deadc9e1087a81a66ff7c4d528640200c63e49226cdc83f98.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 46, + 513, + 287, + 546 + ], + "lines": [ + { + "bbox": [ + 46, + 513, + 287, + 546 + ], + "spans": [ + { + "bbox": [ + 46, + 513, + 287, + 546 + ], + "type": "text", + "content": "Table 1. Quantitative evaluations on the RealBlur dataset [23]. The experimental results were trained under the corresponding datasets respectively, and average runtime is tested on " + }, + { + "bbox": [ + 46, + 513, + 287, + 546 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 46, + 513, + 287, + 546 + ], + "type": "text", + "content": " patches." + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 27 + }, + { + "bbox": [ + 46, + 555, + 287, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 555, + 287, + 578 + ], + "spans": [ + { + "bbox": [ + 46, + 555, + 287, + 578 + ], + "type": "text", + "content": "richer edge details, while achieving the best contrast, sharpness, brightness, and structural details of the image." + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 46, + 582, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 582, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 582, + 287, + 713 + ], + "type": "text", + "content": "Evaluations on the RSBlur dataset. We further conduct experiments on the latest real-world deblurring dataset RSBlur [24] and follow the protocol of this dataset for fair comparison. Tab. 2 summarizes the quantitative evaluation results of our method with advanced algorithms. The proposed MLWNet achieved the highest PSNR and SSIM, which were 34.94 and 0.880 respectively. In addition, the model we trained on RSBlur dataset also achieved the best results on RealBlur-J dataset. We show some visual comparisons in Fig. 5. We note that our method outperforms other methods in low-light blurry scenes, proving that the" + } + ] + } + ], + "index": 30 + }, + { + "type": "table", + "bbox": [ + 308, + 266, + 544, + 403 + ], + "blocks": [ + { + "bbox": [ + 308, + 266, + 544, + 403 + ], + "lines": [ + { + "bbox": [ + 308, + 266, + 544, + 403 + ], + "spans": [ + { + "bbox": [ + 308, + 266, + 544, + 403 + ], + "type": "table", + "html": "
MethodRSBlurRealBlur-J
PSNRSSIMPSNRSSIM
SRN [27]32.530.84029.860.886
MIMO-Unet [3]32.730.84629.530.876
MIMO-Unet+ [3]33.370.85629.990.889
MPRNet [37]33.610.86130.460.899
Restormer [38]33.690.86330.480.891
Uformer-B [33]33.980.86630.370.899
SFNet [4]34.350.87230.260.897
MLWNet-B34.940.88030.530.905
", + "image_path": "7450463824183b9db68bbd13cb5738bdd69d5e93fc599a78406688fb2f5be409.jpg" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 306, + 405, + 545, + 427 + ], + "lines": [ + { + "bbox": [ + 306, + 405, + 545, + 427 + ], + "spans": [ + { + "bbox": [ + 306, + 405, + 545, + 427 + ], + "type": "text", + "content": "Table 2. Quantitative evaluations trained on the RSBlur dataset [24], the RealBlur-J dataset was used for testing only." + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 31 + }, + { + "type": "table", + "bbox": [ + 310, + 435, + 543, + 572 + ], + "blocks": [ + { + "bbox": [ + 310, + 435, + 543, + 572 + ], + "lines": [ + { + "bbox": [ + 310, + 435, + 543, + 572 + ], + "spans": [ + { + "bbox": [ + 310, + 435, + 543, + 572 + ], + "type": "table", + "html": "
Method[23]→ [24]MACs(G)
PSNRSSIM
DeblurGAN-v2 [12]30.150.76642.0
MPRNet [37]29.560.785760.8
MIMO-UNet+ [3]29.690.792154.4
BANet [29]30.190.806263.9
BANet+ [29]30.240.809588.7
MSSNet [9]29.860.806154.0
FFTformer [10]29.700.787131.8
MLWNet-B30.910.818108.2
", + "image_path": "efa659ce56bac1ee88871224700e8a6dde43806a34e6a4ffab7d0ab0ee3e4087.jpg" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 306, + 574, + 545, + 607 + ], + "lines": [ + { + "bbox": [ + 306, + 574, + 545, + 607 + ], + "spans": [ + { + "bbox": [ + 306, + 574, + 545, + 607 + ], + "type": "text", + "content": "Table 3. Quantitative evaluation for generalizability shows the results of models trained on the RealBlur-J dataset and tested on the RSBlur dataset, MACs are measured on " + }, + { + "bbox": [ + 306, + 574, + 545, + 607 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 306, + 574, + 545, + 607 + ], + "type": "text", + "content": " patches." + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 33 + }, + { + "bbox": [ + 306, + 616, + 529, + 627 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 616, + 529, + 627 + ], + "spans": [ + { + "bbox": [ + 306, + 616, + 529, + 627 + ], + "type": "text", + "content": "proposed method makes textures and structures clearer." + } + ] + } + ], + "index": 35 + }, + { + "bbox": [ + 304, + 628, + 545, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 628, + 545, + 700 + ], + "spans": [ + { + "bbox": [ + 304, + 628, + 545, + 700 + ], + "type": "text", + "content": "In addition, we evaluate the RSBlur dataset using models trained only on RealBlur-J to fairly compare the generalization of methods to real scenarios. As shown in Tab. 3, our method produced results with highest PSNR value of 30.91. The results of Tab. 2 and Tab. 3 show that our model has a better generalization ability." + } + ] + } + ], + "index": 36 + }, + { + "bbox": [ + 306, + 701, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 701, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 306, + 701, + 545, + 712 + ], + "type": "text", + "content": "Evaluations on the GoPro dataset. We conduct a exten" + } + ] + } + ], + "index": 37 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 734, + 315, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 734, + 315, + 742 + ], + "spans": [ + { + "bbox": [ + 295, + 734, + 315, + 742 + ], + "type": "text", + "content": "2738" + } + ] + } + ], + "index": 38 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 72, + 213, + 149 + ], + "blocks": [ + { + "bbox": [ + 50, + 72, + 213, + 149 + ], + "lines": [ + { + "bbox": [ + 50, + 72, + 213, + 149 + ], + "spans": [ + { + "bbox": [ + 50, + 72, + 213, + 149 + ], + "type": "image", + "image_path": "0b6d3a6ad6bc0f669e5be3cda7b9bc2baab1203c222d8fb672c8cd84d1741279.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 215, + 71, + 296, + 149 + ], + "blocks": [ + { + "bbox": [ + 215, + 71, + 296, + 149 + ], + "lines": [ + { + "bbox": [ + 215, + 71, + 296, + 149 + ], + "spans": [ + { + "bbox": [ + 215, + 71, + 296, + 149 + ], + "type": "image", + "image_path": "90971285365ea3258b0f209a1d8299550b1f51c18a78e938ea98f12aca3782f2.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 247, + 150, + 264, + 157 + ], + "lines": [ + { + "bbox": [ + 247, + 150, + 264, + 157 + ], + "spans": [ + { + "bbox": [ + 247, + 150, + 264, + 157 + ], + "type": "text", + "content": "SRN" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 298, + 72, + 378, + 149 + ], + "blocks": [ + { + "bbox": [ + 298, + 72, + 378, + 149 + ], + "lines": [ + { + "bbox": [ + 298, + 72, + 378, + 149 + ], + "spans": [ + { + "bbox": [ + 298, + 72, + 378, + 149 + ], + "type": "image", + "image_path": "fbe3bf2dcd531d68e504db1dafedbc5efef65c39f6f4c29b9d25bc2d4b7d1eeb.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 315, + 150, + 361, + 157 + ], + "lines": [ + { + "bbox": [ + 315, + 150, + 361, + 157 + ], + "spans": [ + { + "bbox": [ + 315, + 150, + 361, + 157 + ], + "type": "text", + "content": "MIMO-UNet" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 380, + 72, + 460, + 149 + ], + "blocks": [ + { + "bbox": [ + 380, + 72, + 460, + 149 + ], + "lines": [ + { + "bbox": [ + 380, + 72, + 460, + 149 + ], + "spans": [ + { + "bbox": [ + 380, + 72, + 460, + 149 + ], + "type": "image", + "image_path": "b611528098ba55aa72842822a6b13408c852eb2b2a3848380bf3d4c8ff6a1b75.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 398, + 150, + 448, + 157 + ], + "lines": [ + { + "bbox": [ + 398, + 150, + 448, + 157 + ], + "spans": [ + { + "bbox": [ + 398, + 150, + 448, + 157 + ], + "type": "text", + "content": "MIMO-UNet+" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 463, + 72, + 544, + 149 + ], + "blocks": [ + { + "bbox": [ + 463, + 72, + 544, + 149 + ], + "lines": [ + { + "bbox": [ + 463, + 72, + 544, + 149 + ], + "spans": [ + { + "bbox": [ + 463, + 72, + 544, + 149 + ], + "type": "image", + "image_path": "c6b608e12729bb2f6f17075e0d0397d50b43986036af164a3c423af3961e9665.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 491, + 150, + 522, + 158 + ], + "lines": [ + { + "bbox": [ + 491, + 150, + 522, + 158 + ], + "spans": [ + { + "bbox": [ + 491, + 150, + 522, + 158 + ], + "type": "text", + "content": "MPRNet" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 49, + 159, + 130, + 237 + ], + "blocks": [ + { + "bbox": [ + 108, + 150, + 154, + 159 + ], + "lines": [ + { + "bbox": [ + 108, + 150, + 154, + 159 + ], + "spans": [ + { + "bbox": [ + 108, + 150, + 154, + 159 + ], + "type": "text", + "content": "Blurry image" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 49, + 159, + 130, + 237 + ], + "lines": [ + { + "bbox": [ + 49, + 159, + 130, + 237 + ], + "spans": [ + { + "bbox": [ + 49, + 159, + 130, + 237 + ], + "type": "image", + "image_path": "cc83602cb1a7a0ebeb2c8740b78c76780d2882845493c2ba1ba497bf789056a3.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 77, + 239, + 100, + 247 + ], + "lines": [ + { + "bbox": [ + 77, + 239, + 100, + 247 + ], + "spans": [ + { + "bbox": [ + 77, + 239, + 100, + 247 + ], + "type": "text", + "content": "Blurry" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 132, + 159, + 213, + 237 + ], + "blocks": [ + { + "bbox": [ + 132, + 159, + 213, + 237 + ], + "lines": [ + { + "bbox": [ + 132, + 159, + 213, + 237 + ], + "spans": [ + { + "bbox": [ + 132, + 159, + 213, + 237 + ], + "type": "image", + "image_path": "3e004f48ac48ba037fd91aad6041bf4840a2f29198106f857808621c42006dce.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 167, + 239, + 178, + 247 + ], + "lines": [ + { + "bbox": [ + 167, + 239, + 178, + 247 + ], + "spans": [ + { + "bbox": [ + 167, + 239, + 178, + 247 + ], + "type": "text", + "content": "GT" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 215, + 159, + 296, + 237 + ], + "blocks": [ + { + "bbox": [ + 215, + 159, + 296, + 237 + ], + "lines": [ + { + "bbox": [ + 215, + 159, + 296, + 237 + ], + "spans": [ + { + "bbox": [ + 215, + 159, + 296, + 237 + ], + "type": "image", + "image_path": "f00e7c52ed013110c1f65c797c5d0fc44342c733b5e52ffde9afd24b8afbab65.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 241, + 239, + 277, + 247 + ], + "lines": [ + { + "bbox": [ + 241, + 239, + 277, + 247 + ], + "spans": [ + { + "bbox": [ + 241, + 239, + 277, + 247 + ], + "type": "text", + "content": "Restormer" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 298, + 159, + 378, + 237 + ], + "blocks": [ + { + "bbox": [ + 298, + 159, + 378, + 237 + ], + "lines": [ + { + "bbox": [ + 298, + 159, + 378, + 237 + ], + "spans": [ + { + "bbox": [ + 298, + 159, + 378, + 237 + ], + "type": "image", + "image_path": "668ce462ff79c03992ca1d403b1948c1f2a2d43f0151e5f4608da957b497597b.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 323, + 239, + 361, + 247 + ], + "lines": [ + { + "bbox": [ + 323, + 239, + 361, + 247 + ], + "spans": [ + { + "bbox": [ + 323, + 239, + 361, + 247 + ], + "type": "text", + "content": "Uformer-B" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 381, + 159, + 462, + 237 + ], + "blocks": [ + { + "bbox": [ + 381, + 159, + 462, + 237 + ], + "lines": [ + { + "bbox": [ + 381, + 159, + 462, + 237 + ], + "spans": [ + { + "bbox": [ + 381, + 159, + 462, + 237 + ], + "type": "image", + "image_path": "585ba272724aba34360d1598872f15fecfe33df5104a6e966bfa337dbfc3d71b.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 412, + 239, + 434, + 247 + ], + "lines": [ + { + "bbox": [ + 412, + 239, + 434, + 247 + ], + "spans": [ + { + "bbox": [ + 412, + 239, + 434, + 247 + ], + "type": "text", + "content": "SFNet" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 463, + 159, + 544, + 237 + ], + "blocks": [ + { + "bbox": [ + 463, + 159, + 544, + 237 + ], + "lines": [ + { + "bbox": [ + 463, + 159, + 544, + 237 + ], + "spans": [ + { + "bbox": [ + 463, + 159, + 544, + 237 + ], + "type": "image", + "image_path": "6003de8f5bcef6e4b2c2101766cff97d59f4a10adfbf822687d48424ccca6bfc.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 492, + 239, + 509, + 247 + ], + "lines": [ + { + "bbox": [ + 492, + 239, + 509, + 247 + ], + "spans": [ + { + "bbox": [ + 492, + 239, + 509, + 247 + ], + "type": "text", + "content": "Ours" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 50, + 270, + 189, + 334 + ], + "blocks": [ + { + "bbox": [ + 47, + 249, + 544, + 269 + ], + "lines": [ + { + "bbox": [ + 47, + 249, + 544, + 269 + ], + "spans": [ + { + "bbox": [ + 47, + 249, + 544, + 269 + ], + "type": "text", + "content": "Figure 5. Visual comparisons on the RSBur dataset [24]. The deblurring performance of the proposed method in low-light is impressive, The recovery of characters and texture structures far exceeds other advanced methods." + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 50, + 270, + 189, + 334 + ], + "lines": [ + { + "bbox": [ + 50, + 270, + 189, + 334 + ], + "spans": [ + { + "bbox": [ + 50, + 270, + 189, + 334 + ], + "type": "image", + "image_path": "acac45ca631bc8e8518d5dd1bbb4ed01f503a3d81280e8fffdd5f9cf8d8e5bfb.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 94, + 335, + 141, + 345 + ], + "lines": [ + { + "bbox": [ + 94, + 335, + 141, + 345 + ], + "spans": [ + { + "bbox": [ + 94, + 335, + 141, + 345 + ], + "type": "text", + "content": "Blurry image" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 192, + 270, + 260, + 333 + ], + "blocks": [ + { + "bbox": [ + 192, + 270, + 260, + 333 + ], + "lines": [ + { + "bbox": [ + 192, + 270, + 260, + 333 + ], + "spans": [ + { + "bbox": [ + 192, + 270, + 260, + 333 + ], + "type": "image", + "image_path": "0768e11c2fe4ca5753d8a7ea2eaaaa63f9f065e636de54179cb6e118e3f91a7a.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 202, + 335, + 249, + 344 + ], + "lines": [ + { + "bbox": [ + 202, + 335, + 249, + 344 + ], + "spans": [ + { + "bbox": [ + 202, + 335, + 249, + 344 + ], + "type": "text", + "content": "MIMO-UNet" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_caption" + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 262, + 270, + 332, + 333 + ], + "blocks": [ + { + "bbox": [ + 262, + 270, + 332, + 333 + ], + "lines": [ + { + "bbox": [ + 262, + 270, + 332, + 333 + ], + "spans": [ + { + "bbox": [ + 262, + 270, + 332, + 333 + ], + "type": "image", + "image_path": "ec338d67bd377e6bb9d5ea8da1f1a0d773a1e82d5b909405d39fb112daa4d5ef.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 271, + 335, + 323, + 344 + ], + "lines": [ + { + "bbox": [ + 271, + 335, + 323, + 344 + ], + "spans": [ + { + "bbox": [ + 271, + 335, + 323, + 344 + ], + "type": "text", + "content": "MIMO-UNet+" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_caption" + } + ], + "index": 27 + }, + { + "type": "image", + "bbox": [ + 334, + 270, + 402, + 333 + ], + "blocks": [ + { + "bbox": [ + 334, + 270, + 402, + 333 + ], + "lines": [ + { + "bbox": [ + 334, + 270, + 402, + 333 + ], + "spans": [ + { + "bbox": [ + 334, + 270, + 402, + 333 + ], + "type": "image", + "image_path": "301062017a4633defa98faa1e007decee4f9f7e594c78ae00ec8abe8df9c0053.jpg" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 353, + 335, + 386, + 344 + ], + "lines": [ + { + "bbox": [ + 353, + 335, + 386, + 344 + ], + "spans": [ + { + "bbox": [ + 353, + 335, + 386, + 344 + ], + "type": "text", + "content": "MPRNet" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_caption" + } + ], + "index": 29 + }, + { + "type": "image", + "bbox": [ + 404, + 270, + 473, + 333 + ], + "blocks": [ + { + "bbox": [ + 404, + 270, + 473, + 333 + ], + "lines": [ + { + "bbox": [ + 404, + 270, + 473, + 333 + ], + "spans": [ + { + "bbox": [ + 404, + 270, + 473, + 333 + ], + "type": "image", + "image_path": "eb2837ebefc59f58cd28d1439d3df4eecbaac650ecac73a34adc751830681732.jpg" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 424, + 335, + 455, + 344 + ], + "lines": [ + { + "bbox": [ + 424, + 335, + 455, + 344 + ], + "spans": [ + { + "bbox": [ + 424, + 335, + 455, + 344 + ], + "type": "text", + "content": "NAFNet" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_caption" + } + ], + "index": 31 + }, + { + "type": "image", + "bbox": [ + 475, + 270, + 544, + 333 + ], + "blocks": [ + { + "bbox": [ + 475, + 270, + 544, + 333 + ], + "lines": [ + { + "bbox": [ + 475, + 270, + 544, + 333 + ], + "spans": [ + { + "bbox": [ + 475, + 270, + 544, + 333 + ], + "type": "image", + "image_path": "091cc93979440846cfef3b1c91fede03162da15eea65e86a7a7a3d649a0ada82.jpg" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 493, + 335, + 525, + 344 + ], + "lines": [ + { + "bbox": [ + 493, + 335, + 525, + 344 + ], + "spans": [ + { + "bbox": [ + 493, + 335, + 525, + 344 + ], + "type": "text", + "content": "MAXIM" + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_caption" + } + ], + "index": 33 + }, + { + "type": "image", + "bbox": [ + 49, + 347, + 118, + 411 + ], + "blocks": [ + { + "bbox": [ + 49, + 347, + 118, + 411 + ], + "lines": [ + { + "bbox": [ + 49, + 347, + 118, + 411 + ], + "spans": [ + { + "bbox": [ + 49, + 347, + 118, + 411 + ], + "type": "image", + "image_path": "31dbd953817bb7c2f23ef591a3a38d007ff62e3b66e95fa2d11b5fa6dcb78acf.jpg" + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 71, + 413, + 96, + 423 + ], + "lines": [ + { + "bbox": [ + 71, + 413, + 96, + 423 + ], + "spans": [ + { + "bbox": [ + 71, + 413, + 96, + 423 + ], + "type": "text", + "content": "Blurry" + } + ] + } + ], + "index": 36, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 69, + 424, + 523, + 435 + ], + "lines": [ + { + "bbox": [ + 69, + 424, + 523, + 435 + ], + "spans": [ + { + "bbox": [ + 69, + 424, + 523, + 435 + ], + "type": "text", + "content": "Figure 6. Visual comparisons on the GoPro dataset [20]. Our method better preserves texture information without sharpening." + } + ] + } + ], + "index": 49, + "angle": 0, + "type": "image_caption" + } + ], + "index": 35 + }, + { + "type": "image", + "bbox": [ + 120, + 347, + 189, + 411 + ], + "blocks": [ + { + "bbox": [ + 120, + 347, + 189, + 411 + ], + "lines": [ + { + "bbox": [ + 120, + 347, + 189, + 411 + ], + "spans": [ + { + "bbox": [ + 120, + 347, + 189, + 411 + ], + "type": "image", + "image_path": "12d8c3b7d37de45745154f0618c99448f1e480beb030c83d63ff7d57cfc8084c.jpg" + } + ] + } + ], + "index": 37, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 149, + 413, + 161, + 421 + ], + "lines": [ + { + "bbox": [ + 149, + 413, + 161, + 421 + ], + "spans": [ + { + "bbox": [ + 149, + 413, + 161, + 421 + ], + "type": "text", + "content": "GT" + } + ] + } + ], + "index": 38, + "angle": 0, + "type": "image_caption" + } + ], + "index": 37 + }, + { + "type": "image", + "bbox": [ + 192, + 347, + 260, + 411 + ], + "blocks": [ + { + "bbox": [ + 192, + 347, + 260, + 411 + ], + "lines": [ + { + "bbox": [ + 192, + 347, + 260, + 411 + ], + "spans": [ + { + "bbox": [ + 192, + 347, + 260, + 411 + ], + "type": "image", + "image_path": "47959d8fc10e347beebe3d17f3b7ede9f64fdf474e8f6d1284af07cca1daf510.jpg" + } + ] + } + ], + "index": 39, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 205, + 413, + 246, + 422 + ], + "lines": [ + { + "bbox": [ + 205, + 413, + 246, + 422 + ], + "spans": [ + { + "bbox": [ + 205, + 413, + 246, + 422 + ], + "type": "text", + "content": "Stripformer" + } + ] + } + ], + "index": 40, + "angle": 0, + "type": "image_caption" + } + ], + "index": 39 + }, + { + "type": "image", + "bbox": [ + 262, + 347, + 332, + 411 + ], + "blocks": [ + { + "bbox": [ + 262, + 347, + 332, + 411 + ], + "lines": [ + { + "bbox": [ + 262, + 347, + 332, + 411 + ], + "spans": [ + { + "bbox": [ + 262, + 347, + 332, + 411 + ], + "type": "image", + "image_path": "5a8102137173b287f0fd68320bffa4493735e681c26ca8f2093c1b205e20fe56.jpg" + } + ] + } + ], + "index": 41, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 279, + 413, + 315, + 422 + ], + "lines": [ + { + "bbox": [ + 279, + 413, + 315, + 422 + ], + "spans": [ + { + "bbox": [ + 279, + 413, + 315, + 422 + ], + "type": "text", + "content": "Restormer" + } + ] + } + ], + "index": 42, + "angle": 0, + "type": "image_caption" + } + ], + "index": 41 + }, + { + "type": "image", + "bbox": [ + 334, + 347, + 402, + 411 + ], + "blocks": [ + { + "bbox": [ + 334, + 347, + 402, + 411 + ], + "lines": [ + { + "bbox": [ + 334, + 347, + 402, + 411 + ], + "spans": [ + { + "bbox": [ + 334, + 347, + 402, + 411 + ], + "type": "image", + "image_path": "62a8eeac1b07563f9ea4a80110438401e91ced256a726e01301094ee97714830.jpg" + } + ] + } + ], + "index": 43, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 350, + 413, + 389, + 421 + ], + "lines": [ + { + "bbox": [ + 350, + 413, + 389, + 421 + ], + "spans": [ + { + "bbox": [ + 350, + 413, + 389, + 421 + ], + "type": "text", + "content": "FFTformer" + } + ] + } + ], + "index": 44, + "angle": 0, + "type": "image_caption" + } + ], + "index": 43 + }, + { + "type": "image", + "bbox": [ + 405, + 347, + 473, + 411 + ], + "blocks": [ + { + "bbox": [ + 405, + 347, + 473, + 411 + ], + "lines": [ + { + "bbox": [ + 405, + 347, + 473, + 411 + ], + "spans": [ + { + "bbox": [ + 405, + 347, + 473, + 411 + ], + "type": "image", + "image_path": "bd3a83710a26c343281ae0c5c97ff0cfc0a78c13e971244036d2b7cd8b960bb2.jpg" + } + ] + } + ], + "index": 45, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 430, + 413, + 449, + 421 + ], + "lines": [ + { + "bbox": [ + 430, + 413, + 449, + 421 + ], + "spans": [ + { + "bbox": [ + 430, + 413, + 449, + 421 + ], + "type": "text", + "content": "GRL" + } + ] + } + ], + "index": 46, + "angle": 0, + "type": "image_caption" + } + ], + "index": 45 + }, + { + "type": "image", + "bbox": [ + 476, + 347, + 544, + 411 + ], + "blocks": [ + { + "bbox": [ + 476, + 347, + 544, + 411 + ], + "lines": [ + { + "bbox": [ + 476, + 347, + 544, + 411 + ], + "spans": [ + { + "bbox": [ + 476, + 347, + 544, + 411 + ], + "type": "image", + "image_path": "53feea5edbd48f545a144cef2f5990d2b2bf004e2376d5d40d5b00464aff7989.jpg" + } + ] + } + ], + "index": 47, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 503, + 413, + 521, + 421 + ], + "lines": [ + { + "bbox": [ + 503, + 413, + 521, + 421 + ], + "spans": [ + { + "bbox": [ + 503, + 413, + 521, + 421 + ], + "type": "text", + "content": "Ours" + } + ] + } + ], + "index": 48, + "angle": 0, + "type": "image_caption" + } + ], + "index": 47 + }, + { + "bbox": [ + 46, + 443, + 287, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 443, + 287, + 552 + ], + "spans": [ + { + "bbox": [ + 46, + 443, + 287, + 552 + ], + "type": "text", + "content": "sive comparison with advanced algorithms on GoPro [20]. Tab. 4 shows the quantitative evaluation results. The gap between our method and the state-of-the-art FFTformer is only 0.001 in SSIM value, while the running time is reduced by " + }, + { + "bbox": [ + 46, + 443, + 287, + 552 + ], + "type": "inline_equation", + "content": "60\\%" + }, + { + "bbox": [ + 46, + 443, + 287, + 552 + ], + "type": "text", + "content": ". We show a visual comparison on the GoPro dataset in Fig. 6. We can see that our method recovers blurred textures well without sharpening. For the reasons why the proposed MLWNet does not achieve optimal performance in synthetic blurry, we conducted a detailed analysis in Sec. 5." + } + ] + } + ], + "index": 50 + }, + { + "bbox": [ + 47, + 562, + 184, + 574 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 562, + 184, + 574 + ], + "spans": [ + { + "bbox": [ + 47, + 562, + 184, + 574 + ], + "type": "text", + "content": "5. Analysis and Discussion" + } + ] + } + ], + "index": 51 + }, + { + "bbox": [ + 46, + 582, + 287, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 582, + 287, + 652 + ], + "spans": [ + { + "bbox": [ + 46, + 582, + 287, + 652 + ], + "type": "text", + "content": "In this section, we provide a more in-depth analysis and show the contribution of each component of the proposed method. We perform 20w iterations on the RealBlur-J dataset for abtation with a batch size of 8 to train our method using a version of the model with a width of 32, and the results are shown in Tab. 6." + } + ] + } + ], + "index": 52 + }, + { + "bbox": [ + 46, + 654, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 654, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 654, + 287, + 713 + ], + "type": "text", + "content": "Single-scale vs. Multi-scale. Given that each node of the baseline network is composed of SEB, we have deformed the multi-scale and single-scale according to the architecture. Tab. 5 shows that the multi-scale architecture improves the PSNR and SSIM values to varying degrees, proving that" + } + ] + } + ], + "index": 53 + }, + { + "type": "table", + "bbox": [ + 307, + 441, + 545, + 662 + ], + "blocks": [ + { + "bbox": [ + 307, + 441, + 545, + 662 + ], + "lines": [ + { + "bbox": [ + 307, + 441, + 545, + 662 + ], + "spans": [ + { + "bbox": [ + 307, + 441, + 545, + 662 + ], + "type": "table", + "html": "
MethodGOPROAvg.rruntime
PSNRSSIM
DeblurGAN-v2 [12]29.550.9340.04s
SRN [27]30.260.9340.07s
DMPHN [39]31.200.9450.21s
SDWNet [42]31.260.9660.04s
MPRNet [37]32.660.9590.09s
MIMO-UNet+ [3]32.450.9570.02s
DeepRFT+ [19]33.230.9630.09s
MAXIM-3S [30]32.860.961-
Stripformer [28]33.080.9620.04s
MSDI-net [13]33.280.9640.06s
Restormer [38]33.570.9660.08s
NAFNet [2]33.690.9670.04s
FFTformer [10]34.210.9690.13s
GRL-B [14]33.930.9681.28s
MLWNet-B33.830.9680.05s
", + "image_path": "9372afe43d9b399ebe031c9906f4a4a0479063edd24913fadbbaef85a6d70264.jpg" + } + ] + } + ], + "index": 54, + "angle": 0, + "type": "table_body" + } + ], + "index": 54 + }, + { + "bbox": [ + 306, + 664, + 545, + 696 + ], + "lines": [ + { + "bbox": [ + 306, + 664, + 545, + 696 + ], + "spans": [ + { + "bbox": [ + 306, + 664, + 545, + 696 + ], + "type": "text", + "content": "Table 4. Quantitative evaluations trained and tested on the GoPro dataset [20]. Our proposed MLWNet obtains competitive results with a combination of time efficiency and accuracy." + } + ] + } + ], + "index": 55, + "angle": 0, + "type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 734, + 315, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 734, + 315, + 742 + ], + "spans": [ + { + "bbox": [ + 295, + 734, + 315, + 742 + ], + "type": "text", + "content": "2739" + } + ] + } + ], + "index": 56 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "text", + "content": "the SIMO strategy we employ helps to eliminate motion blur to some extent. As shown in Sec. 3.1, we use SIMO for coarse-to-fine restoration, and only use multiple outputs to calculate multi-scale losses during training." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 120, + 288, + 275 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 120, + 288, + 275 + ], + "spans": [ + { + "bbox": [ + 46, + 120, + 288, + 275 + ], + "type": "text", + "content": "WFB and WHB. We default to using WFB and WHB via " + }, + { + "bbox": [ + 46, + 120, + 288, + 275 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{wavelet}" + }, + { + "bbox": [ + 46, + 120, + 288, + 275 + ], + "type": "text", + "content": " statute in this section, we also proved the effectiveness of LWN, because LWN is a subpart of the first two. After applying them, the PSNR index improved by " + }, + { + "bbox": [ + 46, + 120, + 288, + 275 + ], + "type": "inline_equation", + "content": "0.25\\mathrm{dB}" + }, + { + "bbox": [ + 46, + 120, + 288, + 275 + ], + "type": "text", + "content": " compared to the multi-scale baseline, which shows that LWN and its extension modules can easily help multiscale algorithms improve detail recovery capabilities. Fig. 7 shows the four types of feature maps generated after forward learnable wavelet convolution. We can see that the high and low frequency information of the input feature map are mixed together, and after wavelet group convolution, the network exerts different attention on the feature information in different directions or frequencies." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 275, + 287, + 370 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 275, + 287, + 370 + ], + "spans": [ + { + "bbox": [ + 46, + 275, + 287, + 370 + ], + "type": "text", + "content": "Effectiveness of " + }, + { + "bbox": [ + 46, + 275, + 287, + 370 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{wavelet}" + }, + { + "bbox": [ + 46, + 275, + 287, + 370 + ], + "type": "text", + "content": ". By adding unreduced WFB and WHB to the baseline without using wavelet loss, the performance improvement of the baseline is minimal and does not break the bottleneck of multiscale detail recovery, which suggests that the wavelet convolution degenerates into a general group convolution that only serves to deepen the network, proving that merely deepening the network is ineffective for the restoration of detail information." + } + ] + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 48, + 373, + 288, + 411 + ], + "blocks": [ + { + "bbox": [ + 48, + 373, + 288, + 411 + ], + "lines": [ + { + "bbox": [ + 48, + 373, + 288, + 411 + ], + "spans": [ + { + "bbox": [ + 48, + 373, + 288, + 411 + ], + "type": "table", + "html": "
MethodSISOMIMOSIMO
PSNR/SSIM32.29/0.92432.19/0.92832.37/0.929
MACs(G)19.2421.8319.29
", + "image_path": "cd02f1cd3b4839bc6c7318b37781946a0100a9e52e798190ac2065cee10c3818.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 48, + 425, + 294, + 508 + ], + "blocks": [ + { + "bbox": [ + 64, + 412, + 269, + 423 + ], + "lines": [ + { + "bbox": [ + 64, + 412, + 269, + 423 + ], + "spans": [ + { + "bbox": [ + 64, + 412, + 269, + 423 + ], + "type": "text", + "content": "Table 5. Comparison in various input and output modes." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 48, + 425, + 294, + 508 + ], + "lines": [ + { + "bbox": [ + 48, + 425, + 294, + 508 + ], + "spans": [ + { + "bbox": [ + 48, + 425, + 294, + 508 + ], + "type": "table", + "html": "
SIMOWFBWHBLwaveletPSNRSSIMMACs(G)
✓32.370.92919.29
✓✓✓32.400.92828.21
✓✓✓32.490.92825.28
✓✓✓32.570.92922.22
✓✓✓✓32.620.93128.21
", + "image_path": "a951afaa55d1cd6c5714cbec9b6eb2e0bb02dd70b3a9940f5a92fd9c8257229f.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 48, + 544, + 287, + 605 + ], + "blocks": [ + { + "bbox": [ + 48, + 544, + 287, + 605 + ], + "lines": [ + { + "bbox": [ + 48, + 544, + 287, + 605 + ], + "spans": [ + { + "bbox": [ + 48, + 544, + 287, + 605 + ], + "type": "image", + "image_path": "58b3cc4ebc77ac4abc8cad372c4923708f469c568b1480f2a614e68a40c43774.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 608, + 287, + 639 + ], + "lines": [ + { + "bbox": [ + 46, + 608, + 287, + 639 + ], + "spans": [ + { + "bbox": [ + 46, + 608, + 287, + 639 + ], + "type": "text", + "content": "Figure 7. Feature maps representing high-and low-frequency components generated after learnable wavelet convolution. Zoom in on the screen for the best view." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 641, + 287, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 641, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 46, + 641, + 287, + 712 + ], + "type": "text", + "content": "Limitations and analysis. It is observed from Tab. 1-3 that our method fails to achieve the same expected high performance with synthetic blur as it does with realistic blur. We try to analyze the reasons for this phenomenon via dataset comparison and experimental results. First, the average frame synthesis method used in the synthetic blur" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "text", + "content": "dataset leads to unnatural discontinuous blur trajectories, thereby introducing strange high-frequency information interference (see Fig. 8). Second, synthetic data tends to produce an unnatural mixture of high and low frequencies. It is easy to mix with the average color area at the edge of the texture, causing the confusion of high-low frequency information." + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 308, + 163, + 425, + 265 + ], + "blocks": [ + { + "bbox": [ + 308, + 163, + 425, + 265 + ], + "lines": [ + { + "bbox": [ + 308, + 163, + 425, + 265 + ], + "spans": [ + { + "bbox": [ + 308, + 163, + 425, + 265 + ], + "type": "image", + "image_path": "4479fc93893bb8334181aacfbef23656b02524fd32475217e26cdf6f96467f9a.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 337, + 266, + 392, + 275 + ], + "lines": [ + { + "bbox": [ + 337, + 266, + 392, + 275 + ], + "spans": [ + { + "bbox": [ + 337, + 266, + 392, + 275 + ], + "type": "text", + "content": "(a) Realistic Blur" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 427, + 163, + 545, + 265 + ], + "blocks": [ + { + "bbox": [ + 427, + 163, + 545, + 265 + ], + "lines": [ + { + "bbox": [ + 427, + 163, + 545, + 265 + ], + "spans": [ + { + "bbox": [ + 427, + 163, + 545, + 265 + ], + "type": "image", + "image_path": "02bae2280ed7d33a0ca02349aadccfa0ecd089ceb702a3479a3e441fca2e5f15.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 453, + 266, + 511, + 275 + ], + "lines": [ + { + "bbox": [ + 453, + 266, + 511, + 275 + ], + "spans": [ + { + "bbox": [ + 453, + 266, + 511, + 275 + ], + "type": "text", + "content": "(b) Synthetic Blur" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 304, + 281, + 545, + 324 + ], + "lines": [ + { + "bbox": [ + 304, + 281, + 545, + 324 + ], + "spans": [ + { + "bbox": [ + 304, + 281, + 545, + 324 + ], + "type": "text", + "content": "Figure 8. The difference between realistic blur (a) and synthetic blur (b). In the green box, the synthetic blur appears with color averaging resulting in high and low frequency confusion, and in the blue box has unnatural discontinuous trajectories." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "table", + "bbox": [ + 309, + 325, + 544, + 373 + ], + "blocks": [ + { + "bbox": [ + 47, + 510, + 287, + 543 + ], + "lines": [ + { + "bbox": [ + 47, + 510, + 287, + 543 + ], + "spans": [ + { + "bbox": [ + 47, + 510, + 287, + 543 + ], + "type": "text", + "content": "Table 6. Ablation study on components of the proposed MLWNet. We set the baseline network to use SEB in its entirety, and models that do not use SIMO will represent single scales using SISO." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 309, + 325, + 544, + 373 + ], + "lines": [ + { + "bbox": [ + 309, + 325, + 544, + 373 + ], + "spans": [ + { + "bbox": [ + 309, + 325, + 544, + 373 + ], + "type": "table", + "html": "
Noise LevelL1L2L3L4L5
GoPro34.8134.6633.7633.1932.63
RealBlur-J33.9233.8133.9733.9333.54
", + "image_path": "4460013dfa5080df37ad365c0b55f39bae236b6b324939f1633d260e6b63f6b5.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "table_body" + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 375, + 544, + 396 + ], + "lines": [ + { + "bbox": [ + 304, + 375, + 544, + 396 + ], + "spans": [ + { + "bbox": [ + 304, + 375, + 544, + 396 + ], + "type": "text", + "content": "Table 7. Performance comparison at different noise difference levels, where L3 contains the noise difference mean." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 397, + 545, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 397, + 545, + 517 + ], + "spans": [ + { + "bbox": [ + 304, + 397, + 545, + 517 + ], + "type": "text", + "content": "Finally, there is a large difference in noise levels between synthetic blur and real blur. We follow [1] to calculate the noise level of each testset, and divided the noise difference between clear images and blurred images of GoPro and RealBlur-J into 5 levels for separate testing. As shown in Tab. 7, we note that as the noise difference increases, our method maintains good performance on RealBlur-J, while the performance drops fast on GoPro. This shows that our method is robust to the noise of real blur, and that there are differences in the noise between synthetic blur and real blur." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 306, + 526, + 378, + 538 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 526, + 378, + 538 + ], + "spans": [ + { + "bbox": [ + 306, + 526, + 378, + 538 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 545, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 545, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 545, + 545, + 713 + ], + "type": "text", + "content": "In this paper, we propose a SIMO-based multi-scale architecture to achieve efficient motion deblurring. We have developed a learnable discrete wavelet transform module, which not only improves the algorithm's ability to recover details, but is also more applicable to the real world. In addition, we construct a reasonable multi-scale loss to guide the recovery of blurred images pixel by pixel and scale by scale, and constrain the learning direction of the wavelet kernel with self-supervised loss to achieve better image deblurring. Through extensive experiments on multiple real and synthetic datasets, we demonstrate that it outperforms existing state-of-the-art methods in terms of quality and efficiency of image restoration, especially with optimal deblurring performance and generalization in real scenes." + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 734, + 315, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 734, + 315, + 742 + ], + "spans": [ + { + "bbox": [ + 294, + 734, + 315, + 742 + ], + "type": "text", + "content": "2740" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 287, + 714 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 135 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 135 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 135 + ], + "type": "text", + "content": "[1] Guangyong Chen, Fengyuan Zhu, and Pheng Ann Heng. An efficient statistical method for image noise level estimation. In Proceedings of the IEEE International Conference on Computer Vision, pages 477-485, 2015. 8" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 136, + 287, + 178 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 136, + 287, + 178 + ], + "spans": [ + { + "bbox": [ + 53, + 136, + 287, + 178 + ], + "type": "text", + "content": "[2] Liangyu Chen, Xiaojie Chu, Xiangyu Zhang, and Jian Sun. Simple baselines for image restoration. In European Conference on Computer Vision, pages 17-33. Springer, 2022. 1, 2, 3, 5, 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 180, + 287, + 234 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 180, + 287, + 234 + ], + "spans": [ + { + "bbox": [ + 53, + 180, + 287, + 234 + ], + "type": "text", + "content": "[3] Sung-Jin Cho, Seo-Won Ji, Jun-Pyo Hong, Seung-Won Jung, and Sung-Jea Ko. Rethinking coarse-to-fine approach in single image deblurring. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4641-4650, 2021. 1, 2, 6, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 236, + 287, + 280 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 236, + 287, + 280 + ], + "spans": [ + { + "bbox": [ + 53, + 236, + 287, + 280 + ], + "type": "text", + "content": "[4] Yuning Cui, Yi Tao, Zhenshan Bing, Wenqi Ren, Xinwei Gao, Xiaochun Cao, Kai Huang, and Alois Knoll. Selective frequency network for image restoration. In International Conference on Learning Representations, ICLR, 2023. 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 281, + 287, + 313 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 281, + 287, + 313 + ], + "spans": [ + { + "bbox": [ + 53, + 281, + 287, + 313 + ], + "type": "text", + "content": "[5] Ingrid Daubechies. The wavelet transform, time-frequency localization and signal analysis. IEEE transactions on information theory, 36(5):961-1005, 1990. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 315, + 287, + 367 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 315, + 287, + 367 + ], + "spans": [ + { + "bbox": [ + 53, + 315, + 287, + 367 + ], + "type": "text", + "content": "[6] Golnaz Ghiasi, Tsung-Yi Lin, and Quoc V Le. Nas-fpn: Learning scalable feature pyramid architecture for object detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7036-7045, 2019. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 369, + 287, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 369, + 287, + 413 + ], + "spans": [ + { + "bbox": [ + 53, + 369, + 287, + 413 + ], + "type": "text", + "content": "[7] Wooseok Ha, Chandan Singh, Francois Lanusse, Srigokul Upadhyayula, and Bin Yu. Adaptive wavelet distillation from neural networks through interpretations. Advances in Neural Information Processing Systems, 34:20669-20682, 2021. 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 415, + 287, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 415, + 287, + 479 + ], + "spans": [ + { + "bbox": [ + 53, + 415, + 287, + 479 + ], + "type": "text", + "content": "[8] Bumsoo Kim, Jonghwan Mun, Kyoung-Woon On, Minchul Shin, Junhyun Lee, and Eun-Sol Kim. Mstr: Multi-scale transformer for end-to-end human-object interaction detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19578-19587, 2022. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 480, + 287, + 524 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 480, + 287, + 524 + ], + "spans": [ + { + "bbox": [ + 53, + 480, + 287, + 524 + ], + "type": "text", + "content": "[9] Kiyeon Kim, Seungyong Lee, and Sunghyun Cho. Mssnet: Multi-scale-stage network for single image deblurring. In European Conference on Computer Vision, pages 524-539. Springer, 2022. 1, 2, 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 525, + 287, + 579 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 525, + 287, + 579 + ], + "spans": [ + { + "bbox": [ + 48, + 525, + 287, + 579 + ], + "type": "text", + "content": "[10] Lingshun Kong, Jiangxin Dong, Jianjun Ge, Mingqiang Li, and Jinshan Pan. Efficient frequency domain-based transformers for high-quality image deblurring. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5886-5895, 2023. 1, 2, 3, 6, 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 581, + 287, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 581, + 287, + 635 + ], + "spans": [ + { + "bbox": [ + 48, + 581, + 287, + 635 + ], + "type": "text", + "content": "[11] Orest Kupyn, Volodymyr Budzan, Mykola Mykhailych, Dmytro Mishkin, and Jiří Matas. Deblurgan: Blind motion deblurring using conditional adversarial networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 8183-8192, 2018. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 636, + 287, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 636, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 636, + 287, + 689 + ], + "type": "text", + "content": "[12] Orest Kupyn, Tetiana Martyniuk, Junru Wu, and Zhangyang Wang. Deblurgan-v2: Deblurring (orders-of-magnitude) faster and better. In Proceedings of the IEEE/CVF international conference on computer vision, pages 8878-8887, 2019. 1, 6, 7" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 691, + 287, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 287, + 714 + ], + "type": "text", + "content": "[13] Dasong Li, Yi Zhang, Ka Chun Cheung, Xiaogang Wang, Hongwei Qin, and Hongsheng Li. Learning degradation rep" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 327, + 73, + 545, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 545, + 105 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 545, + 105 + ], + "type": "text", + "content": "resentations for image deblurring. In European Conference on Computer Vision, pages 736-753. Springer, 2022. 1, 2, 6, 7" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 107, + 545, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 107, + 545, + 173 + ], + "spans": [ + { + "bbox": [ + 308, + 107, + 545, + 173 + ], + "type": "text", + "content": "[14] Yawei Li, Yuchen Fan, Xiaoyu Xiang, Denis Demandolx, Rakesh Ranjan, Radu Timofte, and Luc Van Gool. Efficient and explicit modelling of image hierarchies for image restoration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18278-18289, 2023. 2, 6, 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 175, + 545, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 175, + 545, + 229 + ], + "spans": [ + { + "bbox": [ + 308, + 175, + 545, + 229 + ], + "type": "text", + "content": "[15] Tsung-Yi Lin, Piotr Dólar, Ross Girshick, Kaiming He, Bharath Hariharan, and Serge Belongie. Feature pyramid networks for object detection. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2117-2125, 2017. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 231, + 545, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 231, + 545, + 262 + ], + "spans": [ + { + "bbox": [ + 308, + 231, + 545, + 262 + ], + "type": "text", + "content": "[16] Pengju Liu, Hongzhi Zhang, Wei Lian, and Wangmeng Zuo. Multi-level wavelet convolutional neural networks. IEEE Access, 7:74973-74985, 2019. 4, 5" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 264, + 545, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 264, + 545, + 308 + ], + "spans": [ + { + "bbox": [ + 308, + 264, + 545, + 308 + ], + "type": "text", + "content": "[17] Shu Liu, Lu Qi, Haifang Qin, Jianping Shi, and Jiaya Jia. Path aggregation network for instance segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 8759-8768, 2018. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 310, + 545, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 310, + 545, + 354 + ], + "spans": [ + { + "bbox": [ + 308, + 310, + 545, + 354 + ], + "type": "text", + "content": "[18] Wei Liu, Qiong Yan, and Yuzhi Zhao. Densely self-guided wavelet network for image denoising. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, pages 432-433, 2020. 4" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 355, + 545, + 388 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 355, + 545, + 388 + ], + "spans": [ + { + "bbox": [ + 308, + 355, + 545, + 388 + ], + "type": "text", + "content": "[19] X Mao, Y Liu, W Shen, Q Li, and Y Wang. Deep residual fourier transformation for single image deblurring. arxiv 2021. arXiv preprint arXiv:2111.11745, 2021. 2, 3, 6, 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 389, + 545, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 389, + 545, + 443 + ], + "spans": [ + { + "bbox": [ + 308, + 389, + 545, + 443 + ], + "type": "text", + "content": "[20] Seungjun Nah, Tae Hyun Kim, and Kyoung Mu Lee. Deep multi-scale convolutional neural network for dynamic scene deblurring. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3883-3891, 2017. 1, 2, 5, 7" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 445, + 545, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 445, + 545, + 499 + ], + "spans": [ + { + "bbox": [ + 308, + 445, + 545, + 499 + ], + "type": "text", + "content": "[21] Dongwon Park, Dong Un Kang, Jisoo Kim, and Se Young Chun. Multi-temporal recurrent neural networks for progressive non-uniform single image deblurring with incremental temporal training. In European Conference on Computer Vision, pages 327-343. Springer, 2020. 1, 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 501, + 545, + 545 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 501, + 545, + 545 + ], + "spans": [ + { + "bbox": [ + 308, + 501, + 545, + 545 + ], + "type": "text", + "content": "[22] Zequn Qin, Pengyi Zhang, Fei Wu, and Xi Li. Fcanet: Frequency channel attention networks. In Proceedings of the IEEE/CVF international conference on computer vision, pages 783-792, 2021. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 547, + 545, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 547, + 545, + 612 + ], + "spans": [ + { + "bbox": [ + 308, + 547, + 545, + 612 + ], + "type": "text", + "content": "[23] Jaesung Rim, Haeyun Lee, Jucheol Won, and Sunghyun Cho. Real-world blur dataset for learning and benchmarking deblurring algorithms. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXV 16, pages 184-201. Springer, 2020. 1, 5, 6" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 613, + 545, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 613, + 545, + 658 + ], + "spans": [ + { + "bbox": [ + 308, + 613, + 545, + 658 + ], + "type": "text", + "content": "[24] Jaesung Rim, Geonung Kim, Jungeon Kim, Junyong Lee, Seungyong Lee, and Sunghyun Cho. Realistic blur synthesis for learning image deblurring. In European conference on computer vision, pages 487-503. Springer, 2022. 2, 5, 6, 7" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 658, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 658, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 658, + 545, + 713 + ], + "type": "text", + "content": "[25] Roman Suvorov, Elizaveta Logacheva, Anton Mashikhin, Anastasia Remizova, Arsenii Ashukha, Aleksei Silvestrov, Naejin Kong, Harshith Goka, Kiwoong Park, and Victor Lempitsky. Resolution-robust large mask inpainting with fourier convolutions. In Proceedings of the IEEE/CVF winter" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 734, + 314, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 734, + 314, + 742 + ], + "spans": [ + { + "bbox": [ + 295, + 734, + 314, + 742 + ], + "type": "text", + "content": "2741" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 73, + 288, + 715 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 67, + 73, + 287, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 73, + 287, + 95 + ], + "spans": [ + { + "bbox": [ + 67, + 73, + 287, + 95 + ], + "type": "text", + "content": "conference on applications of computer vision, pages 2149-2159, 2022. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 96, + 288, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 96, + 288, + 140 + ], + "spans": [ + { + "bbox": [ + 48, + 96, + 288, + 140 + ], + "type": "text", + "content": "[26] Mingxing Tan, Ruoming Pang, and Quoc V Le. Efficientdet: Scalable and efficient object detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10781-10790, 2020. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 141, + 287, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 141, + 287, + 185 + ], + "spans": [ + { + "bbox": [ + 48, + 141, + 287, + 185 + ], + "type": "text", + "content": "[27] Xin Tao, Hongyun Gao, Xiaoyong Shen, Jue Wang, and Jiaya Jia. Scale-recurrent network for deep image deblurring. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 8174-8182, 2018. 1, 2, 6, 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 186, + 287, + 230 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 186, + 287, + 230 + ], + "spans": [ + { + "bbox": [ + 48, + 186, + 287, + 230 + ], + "type": "text", + "content": "[28] Fu-Jen Tsai, Yan-Tsung Peng, Yen-Yu Lin, Chung-Chi Tsai, and Chia-Wen Lin. Stripformer: Strip transformer for fast image deblurring. In European Conference on Computer Vision, pages 146-162. Springer, 2022. 1, 5, 6, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 232, + 287, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 232, + 287, + 274 + ], + "spans": [ + { + "bbox": [ + 48, + 232, + 287, + 274 + ], + "type": "text", + "content": "[29] Fu-Jen Tsai, Yan-Tsung Peng, Chung-Chi Tsai, Yen-Yu Lin, and Chia-Wen Lin. Banet: a blur-aware attention network for dynamic scene deblurring. IEEE Transactions on Image Processing, 31:6789-6799, 2022. 1, 2, 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 277, + 287, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 277, + 287, + 331 + ], + "spans": [ + { + "bbox": [ + 48, + 277, + 287, + 331 + ], + "type": "text", + "content": "[30] Zhengzhong Tu, Hossein Talebi, Han Zhang, Feng Yang, Peyman Milanfar, Alan Bovik, and Yinxiao Li. Maxim: Multi-axis mlp for image processing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5769-5780, 2022. 6, 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 333, + 287, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 333, + 287, + 387 + ], + "spans": [ + { + "bbox": [ + 48, + 333, + 287, + 387 + ], + "type": "text", + "content": "[31] Chien-Yao Wang, Alexey Bochkovskiy, and Hong-Yuan Mark Liao. Yolov7: Trainable bag-of-freebies sets new state-of-the-art for real-time object detectors. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7464-7475, 2023. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 388, + 287, + 422 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 388, + 287, + 422 + ], + "spans": [ + { + "bbox": [ + 48, + 388, + 287, + 422 + ], + "type": "text", + "content": "[32] Luyuan Wang and Yankui Sun. Image classification using convolutional neural network with wavelet domain inputs. IET Image Processing, 16(8):2037-2048, 2022. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 423, + 287, + 477 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 423, + 287, + 477 + ], + "spans": [ + { + "bbox": [ + 48, + 423, + 287, + 477 + ], + "type": "text", + "content": "[33] Zhendong Wang, Xiaodong Cun, Jianmin Bao, Wengang Zhou, Jianzhuang Liu, and Houqiang Li. Uformer: A general u-shaped transformer for image restoration. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 17683-17693, 2022. 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 479, + 287, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 479, + 287, + 521 + ], + "spans": [ + { + "bbox": [ + 48, + 479, + 287, + 521 + ], + "type": "text", + "content": "[34] Moritz Wolter and Jochen Garcke. Adaptive wavelet pooling for convolutional neural networks. In International Conference on Artificial Intelligence and Statistics, pages 1936-1944. PMLR, 2021. 4, 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 524, + 287, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 524, + 287, + 567 + ], + "spans": [ + { + "bbox": [ + 48, + 524, + 287, + 567 + ], + "type": "text", + "content": "[35] Yanchao Yang and Stefano Soatto. Fda: Fourier domain adaptation for semantic segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4085-4095, 2020. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 569, + 287, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 569, + 287, + 622 + ], + "spans": [ + { + "bbox": [ + 48, + 569, + 287, + 622 + ], + "type": "text", + "content": "[36] Changqian Yu, Changxin Gao, Jingbo Wang, Gang Yu, Chunhua Shen, and Nong Sang. Bisenet v2: Bilateral network with guided aggregation for real-time semantic segmentation. International Journal of Computer Vision, 129: 3051-3068, 2021. 4" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 624, + 287, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 624, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 624, + 287, + 689 + ], + "type": "text", + "content": "[37] Syed Waqas Zamir, Aditya Arora, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, Ming-Hsuan Yang, and Ling Shao. Multi-stage progressive image restoration. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14821-14831, 2021. 1, 2, 6, 7" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 691, + 287, + 715 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 287, + 715 + ], + "type": "text", + "content": "[38] Syed Waqas Zamir, Aditya Arora, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, and Ming-Hsuan Yang." + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 340 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 327, + 73, + 545, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 545, + 116 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 545, + 116 + ], + "type": "text", + "content": "Restormer: Efficient transformer for high-resolution image restoration. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5728-5739, 2022. 2, 5, 6, 7" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 118, + 545, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 545, + 172 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 545, + 172 + ], + "type": "text", + "content": "[39] Hongguang Zhang, Yuchao Dai, Hongdong Li, and Piotr Koniusz. Deep stacked hierarchical multi-patch network for image deblurring. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5978-5986, 2019. 2, 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 175, + 545, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 175, + 545, + 228 + ], + "spans": [ + { + "bbox": [ + 307, + 175, + 545, + 228 + ], + "type": "text", + "content": "[40] Kaihao Zhang, Wenhan Luo, Yiran Zhong, Lin Ma, Bjorn Stenger, Wei Liu, and Hongdong Li. Deblurring by realistic blurring. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2737-2746, 2020. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 230, + 545, + 283 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 230, + 545, + 283 + ], + "spans": [ + { + "bbox": [ + 307, + 230, + 545, + 283 + ], + "type": "text", + "content": "[41] Yijie Zhong, Bo Li, Lv Tang, Senyun Kuang, Shuang Wu, and Shouhong Ding. Detecting camouflaged object in frequency domain. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4504-4513, 2022. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 285, + 545, + 340 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 285, + 545, + 340 + ], + "spans": [ + { + "bbox": [ + 308, + 285, + 545, + 340 + ], + "type": "text", + "content": "[42] Wenbin Zou, Mingchao Jiang, Yunchen Zhang, Liang Chen, Zhiyong Lu, and Yi Wu. Sdwnet: A straight dilated network with wavelet transformation for image deblurring. In Proceedings of the IEEE/CVF international conference on computer vision, pages 1895-1904, 2021. 1, 2, 3, 6, 7" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 734, + 315, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 734, + 315, + 742 + ], + "spans": [ + { + "bbox": [ + 295, + 734, + 315, + 742 + ], + "type": "text", + "content": "2742" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Efficient Multitask Dense Predictor via Binarization/b91bc222-326c-4762-aabe-433ec8965e13_content_list.json b/2024/Efficient Multitask Dense Predictor via Binarization/b91bc222-326c-4762-aabe-433ec8965e13_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..906cd3286de59d8c4d4ffd994db04f9a34846eeb --- /dev/null +++ b/2024/Efficient Multitask Dense Predictor via Binarization/b91bc222-326c-4762-aabe-433ec8965e13_content_list.json @@ -0,0 +1,1471 @@ +[ + { + "type": "text", + "text": "Efficient Multitask Dense Predictor via Binarization", + "text_level": 1, + "bbox": [ + 220, + 130, + 750, + 150 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yuzhang Shang $^{1}$ , Dan Xu $^{2}$ , Gaowen Liu $^{3}$ , Ramana Rao Kompella $^{3}$ , Yan Yan $^{1,\\dagger}$ $^{1}$ Illinois Institute of Technology, $^{2}$ HKUST, $^{3}$ Cisco Research", + "bbox": [ + 158, + 179, + 815, + 217 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "yzshawn@outlook.com, danxu@cse.ust.hk, {ga", + "bbox": [ + 112, + 219, + 486, + 234 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 268, + 313, + 284 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Multi-task learning for dense prediction has emerged as a pivotal area in computer vision, enabling simultaneous processing of diverse yet interrelated pixel-wise prediction tasks. However, the substantial computational demands of state-of-the-art (SoTA) models often limit their widespread deployment. This paper addresses this challenge by introducing network binarization to compress resource-intensive multi-task dense predictors. Specifically, our goal is to significantly accelerate multi-task dense prediction models via Binary Neural Networks (BNNs) while maintaining and even improving model performance at the same time. To reach this goal, we propose a Binary Multi-task Dense Predictor, Bi-MTPD, and several variants of Bi-MTPD, in which a multi-task dense predictor is constructed via specified binarized modules. Our systematic analysis of this predictor reveals that performance drop from binarization is primarily caused by severe information degradation. To address this issue, we introduce a deep information bottleneck layer that enforces representations for downstream tasks satisfying Gaussian distribution in forward propagation. Moreover, we introduce a knowledge distillation mechanism to correct the direction of information flow in backward propagation. Intriguingly, one variant of Bi-MTPD outperforms full-precision (FP) multi-task dense prediction SoTAs, ARTC [2] (CNN-based) and InvPT [50] (ViT-Based). This result indicates that Bi-MTPD is not merely a naive trade-off between performance and efficiency, but is rather a benefit of the redundant information flow thanks to the multi-task architecture. Code is available at BiMTDP.", + "bbox": [ + 75, + 301, + 473, + 753 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 786, + 209, + 801 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "There is a growing trend in the computer vision community where dense prediction tasks are processed in a multi-task learning manner, such as semantic segmentation, monocular depth estimation, and human parsing [28, 46, 47, 49].", + "bbox": [ + 75, + 811, + 468, + 875 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/398fc688ef3c7dda29b672f931df9b0b9dfd838f6876b178b5f98f00a1f4f3c8.jpg", + "image_caption": [ + "Figure 1. (Left) A conceptual illustration of binary dense predictions in a multi-task manner. In contrast to approaching a series of relevant tasks individually, the multitask model benefits from information supplementation among different tasks via cross-talk structures, but the cumbersome cross-talk modules also add additional computational burden. (Right) Performance summary on NYUD-v2. X-axis and Y-axis denote the performance on depth estimation (lower is better) and segmentation (higher is better), respectively. Size of dots denotes FLOPs. ATRC [2] and InvPT [50] are previous CNN-based and ViT-based SoTAs, respectively." + ], + "image_footnote": [], + "bbox": [ + 504, + 268, + 709, + 393 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/48aa45f0018c86e6c50f5d5054f6e6cd1b9a034290ae3bc61e4395b1df103f7c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 714, + 268, + 890, + 397 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Benefited from the information supplementation mechanism via cross-talk structures in the multi-task models, the overall performance for the series of dense prediction tasks has been greatly improved [17] (see Fig. 1). However, the computational demands of State-of-the-Art (SoTA) multitask dense prediction models, which process multiple complex pixel-wise tasks concurrently, are substantial. This high computational requirement limits their application in resource-constrained environments like autonomous driving, robotics, and virtual reality. Our goal is to optimize these heavy SoTA models for edge devices, balancing speed with performance.", + "bbox": [ + 496, + 551, + 893, + 733 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Several strategies for neural network compression have been explored, including pruning [12], network quantization [15, 37, 40] and knowledge distillation [14]. Notably, network binarization, a form of quantization, minimizes weights and activations to $\\pm 1$ , enabling the replacement of computationally intensive inner-product operations in full-precision networks with more efficient xnor-bitcount operations in Binary Neural Networks (BNNs) [15]. Binarization theoretically reduces memory costs by $32\\times$ and increases inference speed by $64\\times$ , making BNNs suitable for edge-device.", + "bbox": [ + 496, + 734, + 895, + 898 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "†Corresponding author", + "bbox": [ + 96, + 886, + 220, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "15899", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "While BNNs have shown impressive results in image classification, achieving nearly full-precision ResNet-level accuracy [25], their application has largely been limited to small-scale models, overlooking other computationally intensive computer vision tasks [15, 23, 25, 32, 39]. Extending BNNs to larger models should be the next step. However, this expansion has been hindered by issues such as overfitting [19, 38, 39] and information degradation [30]. Techniques effective in full-precision models, like label smoothing [34], dropout [42], and mixup [52] have less effect on BNNs [18, 38, 39]. Furthermore, SoTA multitask dense prediction tasks often require deep and complex models, equipped with multi-modality fusion structures [2, 47, 50], exacerbating the challenges in implementing binarization effectively.", + "bbox": [ + 76, + 90, + 472, + 318 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The primary barrier to applying binarization in multitask dense prediction tasks is the significant degradation of information flow in deep models [15, 30, 32], leading to reduced performance. To address this issue, we first propose a Binary Multitask Dense Predictor (Bi-MTPD) baseline, where a multi-task dense predictor is formulated via binarized modules. Based on a thorough review of this baseline, we conclude that the binarization operation destroys the information flows in multi-task models, and thus representations for downstream tasks are not informative compared with their full-precision counterparts. To tackle this problem, we update Bi-MTPD with additional information flow calibration mechanisms in two directions. First, we implement variational information bottleneck enforcing the embeddings to follow Gaussian distribution with sparsity in forward propagation, in order to filter out the task-irrelevant factors. Second, we leverage the existing FP models via feature-based knowledge distillation to calibrate the gradient of the binary network in backward pass.", + "bbox": [ + 76, + 321, + 472, + 609 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The benefits of Bi-MTPD can be analyzed from two orthogonal perspectives. On one hand, from the perspective of network binarization, the accomplishment of bridging binarization with the multitask dense prediction framework testifies that Bi-MTPD can effectively supplement information, and consequently improve the performance of the individual binary models. On the other hand, from the perspective of multitask dense prediction task, accelerating those cumbersome models is profitable to design more effective and efficient cross-talk modules in it, as shown in Fig. 1. Since existing dense prediction models have severe limitation in modelling the cross-talk modules due to their heavy utilization of convolution operation, it is critical for the multitask dense predictions to learn interactions and inference covering various scopes of the multitask context via the cross-talk mechanism [2, 28, 47, 49, 50, 53]. Intriguingly, a variant of Bi-MTPD outperforms SoTA approach ATRC [2] by $4\\%$ over the segmentation task while remaining $43\\%$ faster in speed, implying that our proposed method", + "bbox": [ + 76, + 613, + 472, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "is not a naive trade-off between performance and efficiency. By empirically investigating this \"free lunch\" achievement, we conclude that the win-win outcome is benefited from our designed information supplementation mechanism which strengthens the representation ability of the binary model.", + "bbox": [ + 496, + 90, + 893, + 167 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 180, + 640, + 196 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Multitask Dense Prediction. Multi-Task Learning (MTL) methods can be generally categorized into two main paradigms in terms of the way where model learns shared representations: hard and soft parameter sharing. Hard parameter sharing characterizes architectures which typically share the first hidden representations among the tasks while branching to independent task-specific representations at a later stage. Most approaches split to task-specific heads at a single branch point [5, 17, 21, 36]. However, such naive branching can be sub-optimal, raising interest in mechanisms that allow for finely branched architectures [26, 45]. As a result, in soft parameter sharing, each task is assigned its own set of parameters and a feature sharing mechanism realize the cross-talk as demonstrated in Fig 1. The following works devise the cross-talk mechanisms focusing on the locations in the network where information or features are exchanged or shared between tasks. Apart from the locations, the feature sharing modules are also widely studied. For example, feature fusing can be introduced along the entire network depth [9, 28]; PAD-Net [49] uses multi-modal distillation to enhance task-specific predictions, in which information flow from each source to target task is regulated with a sigmoid-activated gate function; and MTI-Net [47] combines the multi-modal distillation module of PAD-Net with a multi-scale refinement scheme to facilitate cross-task talk at multiple scales.", + "bbox": [ + 496, + 205, + 893, + 598 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Although increasing the number of cross-talk modules intuitively benefits the overall performance of the models, computational cost is often an obstacle. To handle this issue, ATRC [2] introduces NAS [55] to automatically design an efficient information fusing modules. From the perspective of the efficient representation cross-talk, our proposed models with the binarization module can be interpreted as a new pathway to feature fusing within a notably lower inference speed level.", + "bbox": [ + 496, + 598, + 893, + 734 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Neural Network Binarization. As pioneers, [15] use the sign function to quantize weights and activations to $\\pm 1$ , initiating the trends of studies of network binarization. To tackle the vanishing gradient issue induced by the bina- rization operations, the straight-through estimator (STE) [1] is introduced for the gradient approximation. Rooted in this archetype, considerable studies contribute to improving the performance of BNNs, particularly on ImageNet. For example, [23] propose Bi-Real introducing double residual connections with FP downsampling layers to mitigate the excessive gradient vanishing issue caused by binariza", + "bbox": [ + 496, + 734, + 895, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "15900", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "tion, and consequently demonstrate that delicately ing additional connections within BNNs benefits t dient propagation. [13] design a proxy matrix as of the latent parameter space to guide the alignmen weights with different bits by recovering the smooth BNNs. In summary, a large number of methods h tended the boundary of the network binarization w curacy over classification (e.g., ReActNet [25] within comparable FLOPs of binary ResNet-18 achieves $65.9\\%$ Top-1 accuracy on ImageNet, while full-precision version is only $70.5\\%$ ).", + "bbox": [ + 75, + 90, + 467, + 256 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "However, most of those works validate their effectiveness over classification with relatively small architectures (mostly ResNet18 and ResNet34). Meanwhile, the network-based models for dense prediction tasks are bigger and deeper than those toy models, as the information flow in networks is severely degraded. Consequently, directly implementing existing binarization methods can not achieve supposed success. To mitigate the degradation, we propose to binarize those dense prediction models in a multitask way.", + "bbox": [ + 75, + 258, + 468, + 409 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Multitask Network Binarization", + "text_level": 1, + "bbox": [ + 76, + 426, + 370, + 441 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Binary Neural Network", + "text_level": 1, + "bbox": [ + 76, + 450, + 294, + 467 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To begin with, we briefly review the general idea of binary neural networks (BNNs) in [7, 15]. Here, we only elaborate the speedup mechanism and the degradation of information flow of the binarization. We define a full-precision (FP) neural network with $K$ layers, $f(\\mathbf{x}) = (\\mathbf{W}^{K}\\times \\sigma \\times \\mathbf{W}^{K - 1}\\dots \\sigma \\times \\mathbf{W}^{1})(\\mathbf{x})$ , where $\\mathbf{x}$ is the input sample and $\\mathbf{W}^k:\\mathbb{R}^{d_{k - 1}}\\longmapsto \\mathbb{R}^{d_k}(k = 1,\\ldots ,K)$ stands for the weight matrix connecting the $(k - 1)$ -th and the $k$ -th layer, with $d_{k - 1}$ and $d_{k}$ representing the sizes of the input and output of the $k$ -th network layer, respectively. The $\\sigma (\\cdot)$ function performs element-wise activation for the feature maps.", + "bbox": [ + 75, + 476, + 467, + 642 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "BNNs vary from FP neural networks in terms of the forward operation and the backward gradient approximation. Specifically, in the forward propagation, the BNN maintains FP latent weights $\\mathbf{W}_F$ for gradient updates, and the $k$ -th weight matrix $\\mathbf{W}_F^k$ is binarized into $\\pm 1$ , obtaining the binary weight matrix $\\mathbf{W}_B^k$ via the binarize function $\\text{sign}(\\cdot)$ , i.e. $\\mathbf{W}_B^k = \\text{sign}(\\mathbf{W}_F^k)$ . Then the intermediate activation map (full-precision) of the $k$ -th layer is produced by $\\mathbf{A}_F^k = \\mathbf{W}_B^k\\mathbf{A}_B^{k-1}$ . The same quantization method is used to binarize the full-precision activation map as $\\mathbf{A}_B^k = \\text{sign}(\\mathbf{A}_F^k)$ , and the whole forward pass of binarization is performed by iterating this process for $L$ times, as shown in Fig. 2. For BNNs, the weights and activations are 1-bit, by which the network is accelerated 32 times in terms of memory cost. Importantly, inference of BNN is accelerated 64 times, as the FP multiplication in FP networks is replaced by Xnor-Bitcount in BNNs.", + "bbox": [ + 75, + 643, + 467, + 898 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/115f534f2a2a4159b6b9673aa0cf3cb76dcc96ce61aea7c3c47ae62c5b39f535.jpg", + "image_caption": [ + "Figure 2. A general illustration of the forward propagation of the $k$ -th layer in the BNN." + ], + "image_footnote": [], + "bbox": [ + 506, + 90, + 883, + 122 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In the backward propagation, the main challenge is that the pervasive sign functions are theoretically nondifferentiable, and thus extremely destroy the information flow via gradient the propagation. To address this problem, a large number of researchers [31] widely exploit the straight-through estimator (STE) [1] to numerically approximate the derivative of the whole BNN, i.e.", + "bbox": [ + 496, + 175, + 890, + 280 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\text {B a c k w a r d :} \\quad \\frac {\\partial \\mathcal {L}}{\\partial x} = \\left\\{ \\begin{array}{l l} \\frac {\\partial \\mathcal {L}}{\\partial s g n (x)} & | x | \\leq 1 \\\\ 0 & | x | > 1. \\end{array} \\right. \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 547, + 290, + 890, + 325 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "It is worth noting that we do not implement the aforementioned vanilla approximation method in practice, while we utilize the prevalent Bi-Real [23] and IR-Net [32] to gradually approximate the sign function, which have been proven to be better estimation approaches [29, 30].", + "bbox": [ + 496, + 337, + 890, + 412 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Even though numerous methods have been proposed to eliminate the deterioration of the information flow induced by the binarization, the deterioration is still inevitable due to the severe accuracy loss of weights, activations and gradients [29, 30]. Consequently, binarization destroys the performance of the complicated computer vision models.", + "bbox": [ + 496, + 412, + 890, + 503 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Multitask Dense Predictor", + "text_level": 1, + "bbox": [ + 500, + 513, + 736, + 527 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "After deploying binarization techniques in the models for dense prediction tasks, the performance of the binarized models is unacceptable, as shown in Fig. 1 and the binary single result in Table. 3. Since the architectures of those SoTA dense prediction models are relatively heavier and deeper (e.g., HRNet-48 or ResNet-101 with a task-specific head [27]) than the ones for classification (e.g., ResNet-18 with a fully-connected layer as the classification head). Moreover, the information passing in the binary models via back-propagation, especially in deep models, is notoriously inefficient [23].", + "bbox": [ + 496, + 537, + 890, + 703 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Dense prediction tasks can mutually supplement information, e.g., surface normals and depth can directly be derived from each other, which can be modeled as the regularization of each other [47]. The relevancy among dense prediction tasks is worth being utilized to improve the overall performance of models. For example, before the deep learning era, pioneering work [11] utilizes RGB-D images with depth information to predict scene semantics to improve the quality of the prediction. In the deep learning era, recent attention-based multitask learning methods [2, 47, 49] explicitly and implicitly distill information from other tasks as a complementary signal to improve the targeted task performance. Briefly, the above-mentioned methods are achieved", + "bbox": [ + 496, + 704, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "15901", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/c46068f6572247f241dda6efc770bb397ad2576d027021aec506d6cb7d7123e7.jpg", + "image_caption": [ + "Figure 3. (Left) The illustration of the baseline multitask framework. (Middle) The designed MMD modules for binary representations. Importantly, the MMD module can pass information among different predictions, acting as a cross-talk mechanism. (Right) As all fundamental modules in Bi-MTPD baseline are binarized, inferences can be performed by complete Bool operations, which are very computationally cheap." + ], + "image_footnote": [], + "bbox": [ + 119, + 89, + 477, + 258 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/ec62c60358444a2011064bc45b84c0d67f5441695e4a243e3db3214a07ab84f0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 478, + 90, + 704, + 257 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/be1a3fe3fe8e77b12fda57e37b951d594ec50bea252f9243058c41c7408beb04.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 707, + 90, + 848, + 257 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "by combining an existing backbone network for initial task predictions with a subsequent decoding process, as shown in Fig. 3 (Left).", + "bbox": [ + 75, + 337, + 470, + 383 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Specifically, the shared features extracted by the backbone network are then processed by a set of task-specific heads, which produce a series of initial predictions for $T$ tasks, i.e. $\\{Y_i^k\\} (k = 1,\\dots ,T)$ (the backbone and the task-specific heads are referred as the front-end of the network [47]). Transforming and binarizing $Y_{i}^{t}$ into the form of a 1-bit feature map, we obtain a set of corresponding binary feature maps of the scene, i.e. $\\{F_{B,i}^{t}\\} (t = 1,\\dots ,T)$ which are more task-aware than the shared binary features of the backbone network. The information from these task-specific feature representations is then fused via a multimodal distillation via binarized attention mechanism before making the final task predictions. As previous work featured, our method is also task-saleable. Especially, it is possible that some tasks are only predicted in the front-end of the network (initial prediction). The initial tasks are also called auxiliary tasks since they serve as proxies in order to improve the performance of the final tasks, as shown in Fig. 3.", + "bbox": [ + 75, + 398, + 472, + 688 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Multi-Modal Distillation (MMD) via Binarized Attention Mechanism. The multi-modal distillation module is the key in the multi-task dense prediction model. Specifically, we utilize the attention mechanism for guiding the information passing between the binary feature maps generated from different modalities for different tasks. Since the passed information flow is not always helpful, the attention can act as a gate function to control the flow, in other words, to make the network automatically learn to focus or to ignore information from other binary features [47, 49]. Including the binarization operations, we can formalize the MMD via binarized attention as follows. While passing information to the $k$ -th task, we first obtain a binarized attenuation matrix $\\mathbf{A}$ , and then use it to compute the attention matrix $\\mathbf{B}$ .", + "bbox": [ + 75, + 703, + 472, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "tion map $\\mathbf{A}_{B,i}^{k}$ , i.e.,", + "bbox": [ + 500, + 337, + 633, + 354 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {A} _ {B, i} ^ {k} \\leftarrow b o o l \\left(\\mathbf {W} _ {B} ^ {k} \\otimes \\mathbf {F} _ {B, i} ^ {k}\\right) \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 596, + 363, + 890, + 383 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathbf{W}_B^k$ is the parameters of the binarized convolution layer, $\\mathbf{F}_{B,i}^{k}$ is the binary feature map of the initial prediction, and $\\otimes$ denotes convolution operation. Then the information is passed with the attention map controlled as follows:", + "bbox": [ + 498, + 392, + 892, + 454 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {F} _ {B, i} ^ {o, k} \\leftarrow s i g n \\left[ \\mathbf {F} _ {B, i} ^ {k} + \\sum_ {t = 1, t \\neq k} ^ {T} \\mathbf {A} _ {B, i} ^ {k} \\odot \\left(\\mathbf {W} _ {B, t} \\otimes \\mathbf {F} _ {B, i} ^ {t}\\right) \\right] \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 501, + 460, + 890, + 523 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\odot$ element-wise multiplication. The general demonstration of the distillation process is presented in Fig. 3 Left. The output binary feature map $\\mathbf{F}_{B,i}^{o,k}$ is then used by the head for the corresponding $t$ -th task in Fig. 3 Right. By using the task-specific distillation activations, the network can preserve more information for each task [2, 47, 49], which especially benefits the BNNs where the deteriorated information flow mainly induce the performance drop.", + "bbox": [ + 496, + 523, + 890, + 643 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "On the other hand, multitask dense prediction models benefit from the network binarization in terms of performance. Although these multitask models have achieved a promising performance, they are still limited by the nature of convolution-based distillation modules that are heavily used in a multi-scale way, which model critical spatial and task-related interactions in relatively local perceptive fields [50]. Theoretically, more distillation modules in different network nodes can contribute to model performance, yet we cannot unrestrictedly add distillation modules to the existing model due to the computational limitation. Fortunately, with the saved computational cost of the binary networks, we can implement additional distillation modules in our model.", + "bbox": [ + 496, + 643, + 892, + 853 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Binary Baseline for Multitask Dense Prediction, Bi-MTPD. To obtain dense predictions with BNNs under a multitask learning framework, we practically binarize", + "bbox": [ + 498, + 854, + 893, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "15902", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/704a8c0fde1af054bc8734d52148f028127260aab29887a139d76e5ba492c01b.jpg", + "image_caption": [ + "Figure 4. The pipeline of Bi-MTPD. We introduce a VIB layer after the backbone network to filter-out the nuisance factors which may lead to model overfitting issue in the forward propagation. In addition, we deploy a feature-based knowledge distillation mechanism to guide the optimization direction in the backward propagation." + ], + "image_footnote": [], + "bbox": [ + 122, + 88, + 851, + 289 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "the MTI-Net [47] as the binary baseline. Specifically, the main modules in the full-precision MTI-Net, including backbone, heads, and multi-modal distillation module, are replaced with binary modules (both weights and activations are 1-bit). We call this baseline as Bi-MTPD.", + "bbox": [ + 75, + 353, + 470, + 429 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Information Flow Supplementation", + "text_level": 1, + "bbox": [ + 76, + 439, + 382, + 455 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Although we build a fully binarized baseline Bi-MTPD for multitask dense predictions and train the pipeline with common techniques, the performance is still of major concern. The baseline suffers an immense information degradation as the nuisance factors are over-fitted in the forward propagation and optimization directions severely mismatch in the backward propagation. To solve these problems, in this section, we further propose the variant of Bi-MTPD, Bi-MTPD-F. Specifically, we introduce a variational information bottleneck (VIB) layer after the output of the shared binary backbone to precisely enforces the feature extractor to preserve the minimal sufficient information of the input data. As well, we deploy the feature-based knowledge distillation to guided the optimization direction. We present more details in the following section.", + "bbox": [ + 75, + 462, + 468, + 688 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Variational Information Bottleneck for Filter-Out Nuisance Factors. Obtaining the initial binary representations of input images by the shared backbone, we need to train a series of targeted heads to split them. A straightforward strategy is to feed these representations into the following MMD module. However, the binarized representations lack homogenization leading to model overfitting issue [48]. Therefore, the need to regularize the binarized representations, while the regularization would not to contaminate the information flow in the representations. Fortunately, the information bottleneck (IB) principle directly relates to compression with the best hypothesis that the data misfit and the model complexity should simultaneously be minimized [43, 48].", + "bbox": [ + 75, + 688, + 470, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As the VIB could effectively capture the relevant parts and filter out the irrelevant ones from inputs, we design a novel VIB-based layer after the backbone. In particular, it explicitly enforces the feature extractor to preserve the minimal sufficient information of the input data. In other words, it can help ensure the information flow flexibly to learn clean representation for the targeted tasks. The objective function of our VIB-based classification can be formulated as a term of information loss, written as follows:", + "bbox": [ + 496, + 353, + 890, + 489 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {v i b} = K L \\left[ p (\\mathbf {Z} \\mid \\mathbf {A} _ {\\mathbf {B}}), r (\\mathbf {Z}) \\right], \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 591, + 497, + 890, + 515 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathbf{A_B}$ is the input binary backbone representation, $\\mathbf{B}$ is the latent representation variable, $p(\\mathbf{Z} \\mid \\mathbf{A_B})$ is a multivariate Gaussian distribution, and $r(\\mathbf{Z})$ is a standard normal distribution. Generally, the latter is a regularization term controlling how much information of the input is filtered out. A more detailed discussion about the VIB for binarized models filtering out irrelevant information is in Supplemental Materials.", + "bbox": [ + 496, + 522, + 890, + 641 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Feature-based Knowledge Distillation for Guiding the Direction of Information Flow. Distillation is a common and essential optimization approach to alleviate the performance drop of quantized models on ultra-low bit-width settings, which can be flexibly deployed for any architectures to utilize the knowledge of a full-precision teacher model [3, 16, 30, 51]. The usual practice is to distill the activations in a layerwise manner from the full-precision teacher to the quantized counterparts, i.e., $\\mathbf{F}_{B,l}$ and $\\mathbf{F}_{FP,l}$ ( $l = 1,\\dots ,L$ , where $L$ represents the number of network layers), respectively. We use the mean squared errors (MSE) as the distance function to measure the difference between corresponding features student and teacher. The knowledge distillation loss can be written as follows:", + "bbox": [ + 496, + 643, + 890, + 854 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {k d} = \\sum_ {l = 1} ^ {L} M S E \\left(\\mathbf {F} _ {B, l}, \\mathbf {F} _ {F P, l}\\right). \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 591, + 862, + 890, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "15903", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/35b40e833845c6eb5484737bc94a8f79a914bbedb6eb12b163c2469013f560ca.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 150, + 93, + 267, + 162 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/ec1b8852c1b5b6e652375cc3a6736ddf029d063f02b3b7a2df3159aace136216.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 272, + 95, + 385, + 162 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/8d37ebe5c81add2e10eea6b05065059d823414a3212c146b0e90546bde46e950.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 392, + 95, + 504, + 161 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/bdbf2de8bf8bcf343b6dbb45b7d8b7510b6314e5d771ce7e89a9f7ba659691b4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 509, + 95, + 651, + 190 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/2556a9996ba5f9f78730d228893d9915a55b28310bb0fedcaaac4c082dc91750.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 653, + 90, + 820, + 190 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/33a22209f064b38db6842f805e4707670e679a262f438c6048c7e26505649294.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 151, + 162, + 267, + 222 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/731f2bff1ade657246b93d31a804d5958fbdf747f756e721f828d62e42f26fc3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 274, + 162, + 385, + 222 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/34b2b0b67743437999c41f3481ad5610bbf23bbf772adcb585c78af5700a95b7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 392, + 191, + 504, + 222 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/793d6f89f9c5f2952b2ff237f9b9d16c1a470a682e3ba1a2cdc0c2339fddb773.jpg", + "image_caption": [ + "(b)" + ], + "image_footnote": [], + "bbox": [ + 519, + 191, + 651, + 287 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/4d10f587d7454ee29ecbdb5cf637a6411fc2bfb89f4d885715672ce042247d1f.jpg", + "image_caption": [ + "(c)" + ], + "image_footnote": [], + "bbox": [ + 660, + 191, + 820, + 292 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/6b1d80a1dcc2a0df3695da83eda095de313fd402398b4da4a2432c5406a03612.jpg", + "image_caption": [ + "Figure 5. (a) Grad-cam visualization of feature maps of different multitask dense prediction methods. (b) t-SNE visualization of learned features of all 20 classes on Pascal-Context. (c) Centered Kernel Alignment analyzing the information flow within different networks." + ], + "image_footnote": [], + "bbox": [ + 151, + 227, + 267, + 289 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/965881496a0205c0e2bae25158d2bcdf03d3f2d14f35bcc7666fe8c9edebc9b7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 274, + 227, + 385, + 303 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/c8aa5d41e3c53683f7cbac1eaeacb18e62e14887a1fda2317d957f88c3b22840.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 392, + 227, + 504, + 301 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.4. Counter-Intuitive Results of Bi-MTDP-A", + "text_level": 1, + "bbox": [ + 76, + 358, + 426, + 373 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Intuitively, implementing binarization on FP network inevitably induces representations degradation, as the gradient of the sign function cannot be perfectly estimated [1]. Thus, binarized models are impossible to outperform their full-precision counterpart models. However, Bi-MTPD-C, a variant of Bi-MTPD (i.e., full-precision backbone with only binarized multi-modal distillation) outperforms its fully FP version. Specifically, just binarizing the multi-modal distillation can simultaneously accelerate the model by $\\sim 39\\%$ and improve the mIoU for segmentation by $\\sim 4\\%$ , as shown in Table 1. This result demonstrates that our method is not a naive trade-off between model performance and efficiency but a powerful tool for boosting multitask dense predictors. This exciting 'free-lunch' achievement is even a bit of counter-intuitive. We speculate the reasons are that i) binarization on MMD can filter out task-irrelevant information; ii) and thus the information flow in the network is more effective. To testify this speculation, we conduct a series of experiments in two aspects, the representation ability of Bi-MTPD-C and information flow supplementation within the network.", + "bbox": [ + 75, + 383, + 472, + 700 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Qualitative Study of Learned Features with", + "text_level": 1, + "bbox": [ + 75, + 704, + 468, + 718 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Bi-MTPD To investigate the representation ability of Bi-MTPD-C and its FP counterparts, we visualize i) the feature maps behind the Binarized Multi-Modal Distillation (MMD) module in 2-D space via the t-SNE [44] algorithm, and ii) the regions where the network considers important via the Grad-Cam algorithm [35]. The results are shown in Fig. 5. It is clear that binarized model, Bi-MTPD-C is able to filter the irrelevant information out via the binarized attention module (see Fig. 5 (a)), and thus helps learn more discriminative features (see Fig. 5 (b)) resulting in higher quantitative results. Overall, the generated spatial feature maps for segmentation are better. The enhanced", + "bbox": [ + 75, + 719, + 468, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "representative ability can contribute to higher quantitative results.", + "bbox": [ + 496, + 359, + 890, + 387 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Analysis of Information Flow Supplementation within Network via Centered Kernel Alignment. Analyzing distributional information flow within layers of neural networks is challenging because outputs of layers are distributed across a large number of neurons. Centered kernel alignment (CKA) [6, 22, 33] can address these challenges, by quantitatively comparing activations within or across networks. Specifically, for a network feed by $m$ samples, CKA algorithm takes $\\mathbf{X} \\in \\mathbb{R}^{m \\times p_1}$ and $\\mathbf{Y} \\in \\mathbb{R}^{m \\times p_2}$ as inputs which are output activations of two layers (with $p_1$ and $p_2$ neurons respectively). Letting $\\mathbf{K} \\triangleq \\mathbf{X}\\mathbf{X}^{\\top}$ and $\\mathbf{L} \\triangleq \\mathbf{Y}\\mathbf{Y}^{\\top}$ denote the Gram matrices for the two layers CKA computes:", + "bbox": [ + 496, + 388, + 892, + 585 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {C K A} (\\mathbf {K}, \\mathbf {L}) = \\frac {\\operatorname {H S I C} (\\mathbf {K} , \\mathbf {L})}{\\sqrt {\\operatorname {H S I C} (\\mathbf {K} , \\mathbf {K}) \\operatorname {H S I C} (\\mathbf {L} , \\mathbf {L})}}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 542, + 595, + 890, + 632 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where HSIC is the Hilbert-Schmidt independence criterion [10]. Given the centering matrix $\\mathbf{H} = \\mathbf{I}_n - \\frac{1}{n}\\mathbf{1}\\mathbf{1}^\\top$ and the centered Gram matrices $\\mathbf{K}' = \\mathbf{HKH}$ and $\\mathbf{L}' = \\mathbf{HLH}$ , HSIC $= \\frac{\\operatorname{vec}(\\mathbf{K}')\\operatorname{vec}(\\mathbf{L}')}{(m - 1)^2}$ , the similarity between these centered Gram matrices. Importantly, CKA is invariant to orthogonal transformation of representations (including permutation of neurons), and the normalization term ensures invariance to isotropic scaling. These properties enable meaningful comparison and analysis of neural network hidden representations.", + "bbox": [ + 496, + 641, + 890, + 794 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Therefore, we introduce CKA to study the information flow in the multitask dense prediction models. In the heatmap, the lighter the dot, the more similar the two corresponding layers. Higher similar score between two layers' output representations means those two layers share more information. The results are presented in Fig. 5 (c), we can see that the similar scores among front layers and back lay", + "bbox": [ + 496, + 795, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "15904", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/a5a778d97b261d86c991104ef702b0f450459ae7106149d998cf91808fc3226b.jpg", + "table_caption": [ + "Table 1. Results on NYUD-v2. Bi-MTPD-C: only implementing the binarization in the multi-modal distillation module, Bi-MTPD-F: fully-binarized model." + ], + "table_footnote": [], + "table_body": "
Model\nMethodBackboneSemSeg ↑Depth ↓Normal ↓Bound ↑float32\nParams (M)binary\nParams (M)FLOPs\n(G)
Cross [28]36.340.62920.8876.38241.460338.09
PAP [53]36.720.61720.8276.42189.100256.86
PSD [54]36.690.62520.8776.42224.670315.60
PAD [49]CNN36.610.62720.8576.38170.980230.91
MTI [47]45.700.53720.2777.86144.870212.98
ATRC [2]45.870.54020.0977.34180.000249.24
MTI + Bi-MTPD-F39.200.61221.0476.866.45138.4218.07
MTI + Bi-MTPD-C47.710.53020.0677.3690.2854.59130.65
InvPT [50]ViT-B50.300.53619.0077.60154.790244.71
InvPT + Bi-MTPD-C51.20(0.90↑)0.528(0.08↓)19.50(0.50↑)77.68(0.08↑)127.9226.87183.81
InvPT [50]ViT-L53.560.51819.0478.10239.220331.60
InvPT + Bi-MTPD-C54.86(1.30↑)0.515(0.03↓)19.50(0.46↑)78.20(0.10↑)212.3426.87301.88
", + "bbox": [ + 102, + 121, + 867, + 297 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "ers in Bi-MTPD-C is much higher than the ones in MTI-Net [47]. This indicates that Bi-MTPD-C is able to supplement information flow within network, and thus boost the model performance.", + "bbox": [ + 75, + 301, + 470, + 364 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 76, + 375, + 209, + 393 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In this section, we conducted comprehensive experiments to evaluate our proposed method on two datasets for dense prediction tasks: PASCAL Context [8] and NYUD-v2 [41]. We first describe the implementation details of Bi-MTPD, and then compare our method with SoTA binary neural networks in the task of object detection to demonstrate superiority of the proposed method. Finally, we validate the effectiveness of information bottleneck and feature-based knowledge distillation by a series of ablative studies.", + "bbox": [ + 75, + 400, + 468, + 537 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.1. Datasets, Evaluation, Implementation Details", + "text_level": 1, + "bbox": [ + 76, + 545, + 460, + 561 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Datasets. PASCAL-Content is a popular dataset for dense prediction tasks. We use the split from PASCAL-Context which has annotations for semantic segmentation, human part segmentation, semantic edge detection, surface normals prediction and saliency detection. Note that some annotations were distilled by [27] using pre-trained SoTA models [4]. NYUD-v2 contains various indoor scenes such as offices and living rooms with 795 training and 654 testing images. It provides different dense labels, including semantic segmentation, monocular depth estimation, surface normal estimation and object boundary detection.", + "bbox": [ + 75, + 568, + 468, + 733 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Evaluation. Semantic segmentation (Semseg) and human parsing (Parsing) are evaluated with mean Intersection over Union (mIoU); monocular depth estimation (Depth) is evaluated with Root Mean Square Error (RMSE); surface normal estimation (Normal) is evaluated by the mean error (mErr) of predicted angles; saliency detection (Saliency) is evaluated with maximal F-measure (maxF); object boundary detection (Boundary) is evaluated with the optimal-dataset-scale F-measure (odsF). To evaluate the model efficiency w.r.t. memory cost and inference speed, we adopt the number of parameters and FLOPs for a single round of", + "bbox": [ + 75, + 734, + 470, + 901 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "the model inferring an input image.", + "bbox": [ + 498, + 301, + 735, + 316 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Implementation details. We build our approach on the most prevalent backbone architecture, i.e., HRNet as previous SoTA methods [2, 47]. The task-specific heads are also implemented as two basic residual blocks, i.e., binarized BasicBlock and binarized Bottleneck with additional binary shortcuts as BiReal-Net [23]. We use $\\ell_1$ loss for depth estimation and cross-entropy loss for semantic segmentation on NYUD-v2. As in the prior work, the edge detection task is trained with a positive weighted wpos = 0:95 binary cross-entropy loss. We do not adopt a particular loss weighing strategy on NYUD-v2, but simply sum the losses together. On PASCAL, we reuse the training setup from [47] to facilitate a fair comparison. We reuse the loss weights from there. The initial task predictions in the front-end of the network use the same loss weighing as the final task predictions. We refer to the supplementary material for further implementation details. Importantly, we use Adam optimizer [20] in training, but with different learning rates for binary parameters (1e-5) and FP parameters (1e-4), as Adam with a larger learning rate for binary parameters can lead to better training results [24]. Note that our project is based on the codebases for MTI-Net [47] and more details can be found in the codes in the Supplemental Materials.", + "bbox": [ + 496, + 318, + 890, + 667 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. Comparison with State-of-the-Art", + "text_level": 1, + "bbox": [ + 498, + 679, + 797, + 695 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Tabs. 1 and 2 present a comparative analysis of the proposed Binary Multitask Dense Predictor (Bi-MTPD) with current state-of-the-art models on the NYUD-v2 and PASCAL-Context datasets. This comparison includes notable CNN-based methods such as PAD-Net [49], ASTMT [27], MTI-Net [47], and ATRC [2], among others. Bi-MTPD demonstrates exceptional performance, outperforming other models in 6 of the 9 evaluated metrics, particularly in complex scene understanding tasks like Semantic Segmentation and Parsing. Remarkably, on the NYUD-v2 benchmark, Bi-MTPD surpasses the previously best-performing CNN-based method (ATRC) by a margin of +1.8 (mIoU) in Semantic Segmentation, while requiring", + "bbox": [ + 496, + 703, + 890, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "15905", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/aa770149850673ced9116faa9f56c6416ad6ac285ff9d36e4ef45c030613619c.jpg", + "table_caption": [ + "Table 2. Results on PASCAL-VOC. Bi-MTPD-C: only implementing the binarization in the multi-modal distillation module, Bi-MTPD-F: fully-binarized model." + ], + "table_footnote": [], + "table_body": "
Model\nMethodBackboneSemSeg ↑Parsing ↑Saliency ↑Normal ↓Bound ↑float32\nParams (M)binary\nParams (M)FLOPs\n(G)
ASTMT [27]68.0061.1065.7014.7072.40364.720501.27
PAD [49]53.6059.6065.8015.3072.50231.800289.46
MTI [47]61.7060.1884.7814.2370.80218.560280.12
ATRC [2]62.6959.4284.7014.2070.96241.450310.19
ATRC-A [2]63.6062.2383.9114.3070.86249.870320.57
ATRC-B [2]67.6762.9382.2914.2472.42280.010383.21
MTI [47] + Bi-MTPD-F48.1056.2864.4214.2976.8913.67204.8931.38
MTI [47] + Bi-MTPD-C62.98(1.28↑)60.44(0.26↑)83.56(1.22↓)14.31(0.08↑)71.28(0.26↑)153.7164.85194.51
InvPT[50]ViT-B77.5066.8383.6514.6373.00176.350274.68
InvPT[50] + Bi-MTPD-C76.84 (0.66↓)67.10(0.27↑)84.97 (1.32↑)13.69(0.94↓)73.04 (0.04↑)154.6821.67220.76
InvPT[50]ViT-L79.0367.6184.8114.1573.00422.930425.37
InvPT[50] + Bi-MTPD-C79.83 (0.80↑)68.17(0.56↑)
", + "bbox": [ + 81, + 121, + 916, + 291 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "only $62\\%$ of the storage space for weights and $56\\%$ of the computational FLOPs.", + "bbox": [ + 75, + 299, + 455, + 329 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Furthermore, the application of Bi-MTPD to the ViT-based state-of-the-art method, InvPT [50], shows cases Bi-MTPD's ability to enhance model performance while also improving efficiency. This demonstrates the broad applicability and generalization capability of Bi-MTPD across different architectural frameworks.", + "bbox": [ + 75, + 330, + 455, + 419 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "For a qualitative assessment, Figs. 7 and 6 displays prediction examples from various models. These examples illustrate that Bi-MTPD-C not only competes with but occasionally surpasses the state-of-the-art ATRC in qualitative performance.", + "bbox": [ + 75, + 421, + 455, + 496 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.3. Ablative Studies", + "text_level": 1, + "bbox": [ + 76, + 508, + 238, + 523 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Tab. 3 shows a series of ablative studies of Bi-MTPD-C and Bi-MTPD-F with an HRNet-48 backbone on NYUD-v2. We verify how different components of our model contribute to the multi-task improvements. In summary, every designed module positively impacts the overall model. We would like to highlight the main intuition of our method that binarized dense prediction models in a multitask manner largely outperform the binarized single models through Bi-MTPD-F w.o. VIB & KD vs. Bi-Single.", + "bbox": [ + 75, + 532, + 468, + 670 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/d470da1c6387531fdc7b6fa968787abb6bc97bdfb4181758f03ae3845d584f25.jpg", + "table_caption": [ + "Table 3. Ablative studies on NYUD-v2." + ], + "table_footnote": [], + "table_body": "
ModelSemSeg ↑Depth ↓Normal ↓Bound ↑
Bi-MTPD-C47.710.53020.0677.36
Bi-MTPD-C w.o. binary MMD45.700.53720.2777.86
Bi-MTPD-F39.200.61221.0476.86
Bi-MTPD-F w.o. VIB35.040.65222.3173.88
Bi-MTPD-F w.o. KD36.200.64021.9974.86
Bi-MTPD-F w.o. VIB & KD33.990.66723.2471.30
Bi-Single16.200.87229.3669.24
", + "bbox": [ + 84, + 696, + 468, + 801 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 76, + 814, + 194, + 829 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we significantly accelerate cumbersome dense prediction models, in which BNNs for relevant tasks are modeled and optimized under a multitask framework to supplement degraded information caused by binarization op", + "bbox": [ + 75, + 839, + 468, + 902 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/a3ab1f3b0cc7f333ec50d7394c192dbfda0f09f8fb5740b2042e0acaec76f5d2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 501, + 297, + 890, + 426 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/810f13c9eee9f584993405b74e8ed3200030253d88bc5f2f9e8968eb887be7be.jpg", + "image_caption": [ + "Figure 7. Qualitative comparison with MTI-Net on PASCAL-Context." + ], + "image_footnote": [], + "bbox": [ + 501, + 460, + 890, + 623 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "erations. Based on this binary baseline, we further introduce variational information bottleneck and feature-based knowledge distillation to supplement information flow. Experiment results show that our method significantly accelerates existing SoTA methods with comparably small performance drop over the mainstream dense prediction tasks on PASCAL VOC and NYUD-v2. Intriguingly, Bi-MTPD not only reaches SoTA w.r.t. performance but also saves computational costs, compared with SoTA method ARTC [2].", + "bbox": [ + 496, + 664, + 890, + 800 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments: This research is partially supported by NSF IIS-2309073, ECCS-212352101 and Cisco unrestricted gift. This article solely reflects the opinions and conclusions of its authors and not the funding agencies.", + "bbox": [ + 496, + 800, + 890, + 861 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "15906", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Yoshua Bengio, Nicholas Léonard, and Aaron Courville. Estimating or propagating gradients through stochastic neurons for conditional computation. arXiv:1308.3432, 2013. 2, 3, 6", + "[2] David Brüggemann, Menelaos Kanakis, Anton Obukhov, Stamatios Georgoulis, and Luc Van Gool. Exploring relational context for multi-task dense prediction. In ICCV, 2021. 1, 2, 3, 4, 7, 8", + "[3] Adrian Bulat, Brais Martinez, and Georgios Tzimiropoulos. High-capacity expert binary networks. In ICLR, 2020. 5", + "[4] Liang-Chieh Chen, Yukun Zhu, George Papandreou, Florian Schroff, and Hartwig Adam. Encoder-decoder with atrous separable convolution for semantic image segmentation. In ECCV, 2018. 7", + "[5] Zhao Chen, Vijay Badrinarayanan, Chen-Yu Lee, and Andrew Rabinovich. Gradnorm: Gradient normalization for adaptive loss balancing in deep multitask networks. In ICML, 2018. 2", + "[6] Corinna Cortes, Mehryar Mohri, and Afshin Rostamizadeh. Algorithms for learning kernels based on centered alignment. JMLR, 2012. 6", + "[7] Matthieu Courbariaux, Yoshua Bengio, and Jean-Pierre David. Binaryconnect: Training deep neural networks with binary weights during propagations. In NeurIPS, 2016. 3", + "[8] Mark Everingham, Luc Van Gool, Christopher KI Williams, John Winn, and Andrew Zisserman. The pascal visual object classes (voc) challenge. IJCV, 2010. 7", + "[9] Yuan Gao, Jiayi Ma, Mingbo Zhao, Wei Liu, and Alan L Yuille. Nddr-cnn: Layerwise feature fusing in multi-task cnns by neural discriminative dimensionality reduction. In CVPR, 2019. 2", + "[10] Arthur Gretton, Kenji Fukumizu, Choon Teo, Le Song, Bernhard Scholkopf, and Alex Smola. A kernel statistical test of independence. In NeurIPS, 2007. 6", + "[11] Saurabh Gupta, Ross Girshick, Pablo Arbeláez, and Jitendra Malik. Learning rich features from rgb-d images for object detection and segmentation. In ECCV, 2014. 3", + "[12] Song Han, Huizi Mao, and William J Dally. Deep compression: Compressing deep neural networks with pruning, trained quantization and huffman coding. In ICLR, 2016. 1", + "[13] Xiangyu He, Zitao Mo, Ke Cheng, Weixiang Xu, Qinghao Hu, Peisong Wang, Qingshan Liu, and Jian Cheng. Proxybnn: Learning binarized neural networks via proxy matrices. In CVPR, 2020. 3", + "[14] Geoffrey Hinton, Oriol Vinyls, and Jeff Dean. Distilling the knowledge in a neural network. In NeurIPS, 2014. 1", + "[15] Itay Hubara, Matthieu Courbariaux, Daniel Soudry, Ran ElYaniv, and Yoshua Bengio. Binarized neural networks. In NeurIPS, 2016. 1, 2, 3", + "[16] Xiaoqi Jiao, Yichun Yin, Lifeng Shang, Xin Jiang, Xiao Chen, Linlin Li, Fang Wang, and Qun Liu. Tinybert: Distilling bert for natural language understanding. In EMNLP, 2020. 5", + "[17] Alex Kendall, Yarin Gal, and Roberto Cipolla. Multi-task learning using uncertainty to weigh losses for scene geometry and semantics. In CVPR, 2018. 1, 2", + "[18] Hyungjun Kim, Kyungsu Kim, Jinseok Kim, and Jae-Joon Kim. Binary duo: Reducing gradient mismatch in binary activation network by coupling binary activations. In ICLR," + ], + "bbox": [ + 78, + 114, + 468, + 896 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "2020. 2", + "[19] Hyungjun Kim, Jihoon Park, Changhun Lee, and Jae-Joon Kim. Improving accuracy of binary neural networks using unbalanced activation distribution. In CVPR, 2021. 2", + "[20] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In ICLR, 2014. 7", + "[21] Iasonas Kokkinos. Ethernet: Training a universal convolutional neural network for low-, mid-, and high-level vision using diverse datasets and limited memory. In CVPR, 2017. 2", + "[22] Simon Kornblith, Mohammad Norouzi, Honglak Lee, and Geoffrey Hinton. Similarity of neural network representations revisited. In ICML, 2019. 6", + "[23] Zechun Liu, Wenhan Luo, Baoyuan Wu, Xin Yang, Wei Liu, and Kwang-Ting Cheng. Bi-real net: Binarizing deep network towards real-network performance. IJCV, 2020. 2, 3, 7", + "[24] Zechun Liu, Zhiqiang Shen, Shichao Li, Koen Helwegen, Dong Huang, and Kwang-Ting Cheng. How do adam and training strategies help bnns optimization. In ICML, 2021. 7", + "[25] Zechun Liu, Zhiqiang Shen, Marios Savvides, and Kwang-Ting Cheng. Reactnet: Towards precise binary neural network with generalized activation functions. In ECCV, 2020. 2, 3", + "[26] Yongxi Lu, Abhishek Kumar, Shuangfei Zhai, Yu Cheng, Tara Javidi, and Rogerio Feris. Fully-adaptive feature sharing in multi-task networks with applications in person attribute classification. In CVPR, 2017. 2", + "[27] Kevis-Kokitsi Maninis, Ilija Radosavovic, and Iasonas Kokkinos. Attentive single-tasking of multiple tasks. In CVPR, 2019. 3, 7, 8", + "[28] Ishan Misra, Abhinav Shrivastava, Abhinav Gupta, and Martial Hebert. Cross-stitch networks for multi-task learning. In CVPR, 2016. 1, 2, 7", + "[29] Haotong Qin, Zhongang Cai, Mingyuan Zhang, Yifu Ding, Haiyu Zhao, Shuai Yi, Xianglong Liu, and Hao Su. Bipoint-net: Binary neural network for point clouds. In ICLR, 2021. 3", + "[30] Haotong Qin, Yifu Ding, Mingyuan Zhang, YAN Qinghua, Aishan Liu, Qingqing Dang, Ziwei Liu, and Xianglong Liu. Bibert: Accurate fully binarized bert. In ICLR, 2022. 2, 3, 5", + "[31] Haotong Qin, Ruihao Gong, Xianglong Liu, Xiao Bai, Jingkuan Song, and Nicu Sebe. Binary neural networks: A survey. PR, 2020. 3", + "[32] Haotong Qin, Ruihao Gong, Xianglong Liu, Mingzhu Shen, Ziran Wei, Fengwei Yu, and Jingkuan Song. Forward and backward information retention for accurate binary neural networks. In CVPR, 2020. 2, 3", + "[33] Maithra Raghu, Thomas Unterthiner, Simon Kornblith, Chiyuan Zhang, and Alexey Dosovitskiy. Do vision transformers see like convolutional neural networks? In NeurIPS, 2021. 6", + "[34] Lothar Reichel and Qiang Ye. Simple square smoothing regularization operators. Electronic Transactions on Numerical Analysis, 33:63, 2009. 2", + "[35] Ramprasaath R Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. Grad-cam: Visual explanations from deep networks via gradient-based localization. In ICCV, 2017. 6", + "[36] Ozan Sener and Vladlen Koltun. Multi-task learning as" + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "15907", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "multi-objective optimization. In NeurIPS, 2018. 2", + "[37] Yuzhang Shang, Bingxin Xu, Gaowen Liu, Ramana Rao Kompella, and Yan Yan. Causal-dfq: Causality guided data-free network quantization. In ICCV, 2023. 1", + "[38] Yuzhang Shang, Dan Xu, Bin Duan, Ziliang Zong, Liqiang Nie, and Yan Yan. Lipschitz continuity retained binary neural network. In ECCV, 2022. 2", + "[39] Yuzhang Shang, Dan Xu, Ziliang Zong, and Yan Yan. Network binarization via contrastive learning. In ECCV, 2022. 2", + "[40] Yuzhang Shang, Zhihang Yuan, Bin Xie, Bingzhe Wu, and Yan Yan. Post-training quantization on diffusion models. In CVPR, pages 1972-1981, 2023. 1", + "[41] Nathan Silberman, Derek Hoiem, Pushmeet Kohli, and Rob Fergus. Indoor segmentation and support inference from rgbd images. In ECCV, 2012. 7", + "[42] Nitish Srivastava, Geoffrey Hinton, Alex Krizhevsky, Ilya Sutskever, and Ruslan Salakhutdinov. Dropout: a simple way to prevent neural networks from overfitting. JMLR, 2014. 2", + "[43] Naftali Tishby, Fernando C Pereira, and William Bialek. The information bottleneck method. arXiv preprint physics/0004057, 2000. 5", + "[44] Laurens Van der Maaten and Geoffrey Hinton. Visualizing data using t-sne. JMLR, 2008. 6", + "[45] Simon Vandenhende, Stamatios Georgoulis, Bert De Brabandere, and Luc Van Gool. Branched multi-task networks: deciding what layers to share. In BMVC, 2020. 2", + "[46] S. Vandenhende, S. Georgoulis, W. Van Gansbeke, M. Proesmans, D. Dai, and L. Van Gool. Multi-task learning for dense prediction tasks: A survey. TPAMI, 2021. 1", + "[47] Simon Vandenhende, Stamatios Georgoulis, and Luc Van Gool. Mti-net: Multi-scale task interaction networks for multi-task learning. In CVPR, 2020. 1, 2, 3, 4, 5, 7, 8", + "[48] Ziwei Wang, Ziyi Wu, Jiwen Lu, and Jie Zhou. Bidet: An efficient binarized object detector. In CVPR, 2020. 5", + "[49] Dan Xu, Wanli Ouyang, Xiaogang Wang, and Nicu Sebe. Pad-net: Multi-tasks guided prediction-and-distillation network for simultaneous depth estimation and scene parsing. In CVPR, 2018, 1, 2, 3, 4, 7, 8", + "[50] Hanrong Ye and Dan Xu. Inverted pyramid multi-task transformer for dense scene understanding. arXiv preprint arXiv:2203.07997, 2022. 1, 2, 4, 7, 8", + "[51] Shang Yuzhang, Duan Bin, Zong Ziliang, Nie Liqiang, and Yan Yan. Lipschitz continuity guided knowledge distillation. In ICCV, 2021. 5", + "[52] Hongyi Zhang, Moustapha Cisse, Yann N Dauphin, and David Lopez-Paz. mixup: Beyond empirical risk minimization. In ICLR, 2018. 2", + "[53] Zhenyu Zhang, Zhen Cui, Chunyan Xu, Yan Yan, Nicu Sebe, and Jian Yang. Pattern-affinitive propagation across depth, surface normal and semantic segmentation. In CVPR, 2019. 2, 7", + "[54] Ling Zhou, Zhen Cui, Chunyan Xu, Zhenyu Zhang, Chaoqun Wang, Tong Zhang, and Jian Yang. Pattern-structure diffusion for multi-task learning. In CVPR, 2020. 7", + "[55] Barret Zoph and Quoc V Le. Neural architecture search with reinforcement learning. arXiv preprint arXiv:1611.01578, 2016.2" + ], + "bbox": [ + 78, + 92, + 468, + 869 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "15908", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2024/Efficient Multitask Dense Predictor via Binarization/b91bc222-326c-4762-aabe-433ec8965e13_model.json b/2024/Efficient Multitask Dense Predictor via Binarization/b91bc222-326c-4762-aabe-433ec8965e13_model.json new file mode 100644 index 0000000000000000000000000000000000000000..9a0852d33cbc8d2916e04774cc7880d94524a790 --- /dev/null +++ b/2024/Efficient Multitask Dense Predictor via Binarization/b91bc222-326c-4762-aabe-433ec8965e13_model.json @@ -0,0 +1,2079 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.222, + 0.131, + 0.75, + 0.151 + ], + "angle": 0, + "content": "Efficient Multitask Dense Predictor via Binarization" + }, + { + "type": "text", + "bbox": [ + 0.16, + 0.18, + 0.816, + 0.218 + ], + "angle": 0, + "content": "Yuzhang Shang\\(^{1}\\), Dan Xu\\(^{2}\\), Gaowen Liu\\(^{3}\\), Ramana Rao Kompella\\(^{3}\\), Yan Yan\\(^{1,\\dagger}\\) \n\\(^{1}\\)Illinois Institute of Technology, \\(^{2}\\)HKUST, \\(^{3}\\)Cisco Research" + }, + { + "type": "text", + "bbox": [ + 0.114, + 0.22, + 0.488, + 0.235 + ], + "angle": 0, + "content": "yzshawn@outlook.com, danxu@cse.ust.hk, {ga" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.269, + 0.314, + 0.285 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.302, + 0.474, + 0.755 + ], + "angle": 0, + "content": "Multi-task learning for dense prediction has emerged as a pivotal area in computer vision, enabling simultaneous processing of diverse yet interrelated pixel-wise prediction tasks. However, the substantial computational demands of state-of-the-art (SoTA) models often limit their widespread deployment. This paper addresses this challenge by introducing network binarization to compress resource-intensive multi-task dense predictors. Specifically, our goal is to significantly accelerate multi-task dense prediction models via Binary Neural Networks (BNNs) while maintaining and even improving model performance at the same time. To reach this goal, we propose a Binary Multi-task Dense Predictor, Bi-MTPD, and several variants of Bi-MTPD, in which a multi-task dense predictor is constructed via specified binarized modules. Our systematic analysis of this predictor reveals that performance drop from binarization is primarily caused by severe information degradation. To address this issue, we introduce a deep information bottleneck layer that enforces representations for downstream tasks satisfying Gaussian distribution in forward propagation. Moreover, we introduce a knowledge distillation mechanism to correct the direction of information flow in backward propagation. Intriguingly, one variant of Bi-MTPD outperforms full-precision (FP) multi-task dense prediction SoTAs, ARTC [2] (CNN-based) and InvPT [50] (ViT-Based). This result indicates that Bi-MTPD is not merely a naive trade-off between performance and efficiency, but is rather a benefit of the redundant information flow thanks to the multi-task architecture. Code is available at BiMTDP." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.787, + 0.21, + 0.803 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.813, + 0.47, + 0.875 + ], + "angle": 0, + "content": "There is a growing trend in the computer vision community where dense prediction tasks are processed in a multi-task learning manner, such as semantic segmentation, monocular depth estimation, and human parsing [28, 46, 47, 49]." + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.269, + 0.71, + 0.395 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.715, + 0.269, + 0.891, + 0.398 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.401, + 0.895, + 0.542 + ], + "angle": 0, + "content": "Figure 1. (Left) A conceptual illustration of binary dense predictions in a multi-task manner. In contrast to approaching a series of relevant tasks individually, the multitask model benefits from information supplementation among different tasks via cross-talk structures, but the cumbersome cross-talk modules also add additional computational burden. (Right) Performance summary on NYUD-v2. X-axis and Y-axis denote the performance on depth estimation (lower is better) and segmentation (higher is better), respectively. Size of dots denotes FLOPs. ATRC [2] and InvPT [50] are previous CNN-based and ViT-based SoTAs, respectively." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.552, + 0.895, + 0.734 + ], + "angle": 0, + "content": "Benefited from the information supplementation mechanism via cross-talk structures in the multi-task models, the overall performance for the series of dense prediction tasks has been greatly improved [17] (see Fig. 1). However, the computational demands of State-of-the-Art (SoTA) multitask dense prediction models, which process multiple complex pixel-wise tasks concurrently, are substantial. This high computational requirement limits their application in resource-constrained environments like autonomous driving, robotics, and virtual reality. Our goal is to optimize these heavy SoTA models for edge devices, balancing speed with performance." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.735, + 0.896, + 0.9 + ], + "angle": 0, + "content": "Several strategies for neural network compression have been explored, including pruning [12], network quantization [15, 37, 40] and knowledge distillation [14]. Notably, network binarization, a form of quantization, minimizes weights and activations to \\(\\pm 1\\), enabling the replacement of computationally intensive inner-product operations in full-precision networks with more efficient xnor-bitcount operations in Binary Neural Networks (BNNs) [15]. Binarization theoretically reduces memory costs by \\(32\\times\\) and increases inference speed by \\(64\\times\\), making BNNs suitable for edge-device." + }, + { + "type": "page_footnote", + "bbox": [ + 0.097, + 0.887, + 0.221, + 0.9 + ], + "angle": 0, + "content": "†Corresponding author" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "15899" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.473, + 0.319 + ], + "angle": 0, + "content": "While BNNs have shown impressive results in image classification, achieving nearly full-precision ResNet-level accuracy [25], their application has largely been limited to small-scale models, overlooking other computationally intensive computer vision tasks [15, 23, 25, 32, 39]. Extending BNNs to larger models should be the next step. However, this expansion has been hindered by issues such as overfitting [19, 38, 39] and information degradation [30]. Techniques effective in full-precision models, like label smoothing [34], dropout [42], and mixup [52] have less effect on BNNs [18, 38, 39]. Furthermore, SoTA multitask dense prediction tasks often require deep and complex models, equipped with multi-modality fusion structures [2, 47, 50], exacerbating the challenges in implementing binarization effectively." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.323, + 0.473, + 0.61 + ], + "angle": 0, + "content": "The primary barrier to applying binarization in multitask dense prediction tasks is the significant degradation of information flow in deep models [15, 30, 32], leading to reduced performance. To address this issue, we first propose a Binary Multitask Dense Predictor (Bi-MTPD) baseline, where a multi-task dense predictor is formulated via binarized modules. Based on a thorough review of this baseline, we conclude that the binarization operation destroys the information flows in multi-task models, and thus representations for downstream tasks are not informative compared with their full-precision counterparts. To tackle this problem, we update Bi-MTPD with additional information flow calibration mechanisms in two directions. First, we implement variational information bottleneck enforcing the embeddings to follow Gaussian distribution with sparsity in forward propagation, in order to filter out the task-irrelevant factors. Second, we leverage the existing FP models via feature-based knowledge distillation to calibrate the gradient of the binary network in backward pass." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.614, + 0.473, + 0.903 + ], + "angle": 0, + "content": "The benefits of Bi-MTPD can be analyzed from two orthogonal perspectives. On one hand, from the perspective of network binarization, the accomplishment of bridging binarization with the multitask dense prediction framework testifies that Bi-MTPD can effectively supplement information, and consequently improve the performance of the individual binary models. On the other hand, from the perspective of multitask dense prediction task, accelerating those cumbersome models is profitable to design more effective and efficient cross-talk modules in it, as shown in Fig. 1. Since existing dense prediction models have severe limitation in modelling the cross-talk modules due to their heavy utilization of convolution operation, it is critical for the multitask dense predictions to learn interactions and inference covering various scopes of the multitask context via the cross-talk mechanism [2, 28, 47, 49, 50, 53]. Intriguingly, a variant of Bi-MTPD outperforms SoTA approach ATRC [2] by \\(4\\%\\) over the segmentation task while remaining \\(43\\%\\) faster in speed, implying that our proposed method" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.169 + ], + "angle": 0, + "content": "is not a naive trade-off between performance and efficiency. By empirically investigating this \"free lunch\" achievement, we conclude that the win-win outcome is benefited from our designed information supplementation mechanism which strengthens the representation ability of the binary model." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.181, + 0.642, + 0.197 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.206, + 0.895, + 0.599 + ], + "angle": 0, + "content": "Multitask Dense Prediction. Multi-Task Learning (MTL) methods can be generally categorized into two main paradigms in terms of the way where model learns shared representations: hard and soft parameter sharing. Hard parameter sharing characterizes architectures which typically share the first hidden representations among the tasks while branching to independent task-specific representations at a later stage. Most approaches split to task-specific heads at a single branch point [5, 17, 21, 36]. However, such naive branching can be sub-optimal, raising interest in mechanisms that allow for finely branched architectures [26, 45]. As a result, in soft parameter sharing, each task is assigned its own set of parameters and a feature sharing mechanism realize the cross-talk as demonstrated in Fig 1. The following works devise the cross-talk mechanisms focusing on the locations in the network where information or features are exchanged or shared between tasks. Apart from the locations, the feature sharing modules are also widely studied. For example, feature fusing can be introduced along the entire network depth [9, 28]; PAD-Net [49] uses multi-modal distillation to enhance task-specific predictions, in which information flow from each source to target task is regulated with a sigmoid-activated gate function; and MTI-Net [47] combines the multi-modal distillation module of PAD-Net with a multi-scale refinement scheme to facilitate cross-task talk at multiple scales." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.599, + 0.895, + 0.735 + ], + "angle": 0, + "content": "Although increasing the number of cross-talk modules intuitively benefits the overall performance of the models, computational cost is often an obstacle. To handle this issue, ATRC [2] introduces NAS [55] to automatically design an efficient information fusing modules. From the perspective of the efficient representation cross-talk, our proposed models with the binarization module can be interpreted as a new pathway to feature fusing within a notably lower inference speed level." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.735, + 0.897, + 0.903 + ], + "angle": 0, + "content": "Neural Network Binarization. As pioneers, [15] use the sign function to quantize weights and activations to \\(\\pm 1\\), initiating the trends of studies of network binarization. To tackle the vanishing gradient issue induced by the bina- rization operations, the straight-through estimator (STE) [1] is introduced for the gradient approximation. Rooted in this archetype, considerable studies contribute to improving the performance of BNNs, particularly on ImageNet. For example, [23] propose Bi-Real introducing double residual connections with FP downsampling layers to mitigate the excessive gradient vanishing issue caused by binariza" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "15900" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.468, + 0.257 + ], + "angle": 0, + "content": "tion, and consequently demonstrate that delicately ing additional connections within BNNs benefits t dient propagation. [13] design a proxy matrix as of the latent parameter space to guide the alignmen weights with different bits by recovering the smooth BNNs. In summary, a large number of methods h tended the boundary of the network binarization w curacy over classification (e.g., ReActNet [25] within comparable FLOPs of binary ResNet-18 achieves \\(65.9\\%\\) Top-1 accuracy on ImageNet, while full-precision version is only \\(70.5\\%\\))." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.259, + 0.47, + 0.41 + ], + "angle": 0, + "content": "However, most of those works validate their effectiveness over classification with relatively small architectures (mostly ResNet18 and ResNet34). Meanwhile, the network-based models for dense prediction tasks are bigger and deeper than those toy models, as the information flow in networks is severely degraded. Consequently, directly implementing existing binarization methods can not achieve supposed success. To mitigate the degradation, we propose to binarize those dense prediction models in a multitask way." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.427, + 0.371, + 0.442 + ], + "angle": 0, + "content": "3. Multitask Network Binarization" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.452, + 0.295, + 0.468 + ], + "angle": 0, + "content": "3.1. Binary Neural Network" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.477, + 0.468, + 0.643 + ], + "angle": 0, + "content": "To begin with, we briefly review the general idea of binary neural networks (BNNs) in [7, 15]. Here, we only elaborate the speedup mechanism and the degradation of information flow of the binarization. We define a full-precision (FP) neural network with \\(K\\) layers, \\(f(\\mathbf{x}) = (\\mathbf{W}^{K}\\times \\sigma \\times \\mathbf{W}^{K - 1}\\dots \\sigma \\times \\mathbf{W}^{1})(\\mathbf{x})\\), where \\(\\mathbf{x}\\) is the input sample and \\(\\mathbf{W}^k:\\mathbb{R}^{d_{k - 1}}\\longmapsto \\mathbb{R}^{d_k}(k = 1,\\ldots ,K)\\) stands for the weight matrix connecting the \\((k - 1)\\)-th and the \\(k\\)-th layer, with \\(d_{k - 1}\\) and \\(d_{k}\\) representing the sizes of the input and output of the \\(k\\)-th network layer, respectively. The \\(\\sigma (\\cdot)\\) function performs element-wise activation for the feature maps." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.644, + 0.468, + 0.9 + ], + "angle": 0, + "content": "BNNs vary from FP neural networks in terms of the forward operation and the backward gradient approximation. Specifically, in the forward propagation, the BNN maintains FP latent weights \\(\\mathbf{W}_F\\) for gradient updates, and the \\(k\\)-th weight matrix \\(\\mathbf{W}_F^k\\) is binarized into \\(\\pm 1\\), obtaining the binary weight matrix \\(\\mathbf{W}_B^k\\) via the binarize function \\(\\text{sign}(\\cdot)\\), i.e. \\(\\mathbf{W}_B^k = \\text{sign}(\\mathbf{W}_F^k)\\). Then the intermediate activation map (full-precision) of the \\(k\\)-th layer is produced by \\(\\mathbf{A}_F^k = \\mathbf{W}_B^k\\mathbf{A}_B^{k-1}\\). The same quantization method is used to binarize the full-precision activation map as \\(\\mathbf{A}_B^k = \\text{sign}(\\mathbf{A}_F^k)\\), and the whole forward pass of binarization is performed by iterating this process for \\(L\\) times, as shown in Fig. 2. For BNNs, the weights and activations are 1-bit, by which the network is accelerated 32 times in terms of memory cost. Importantly, inference of BNN is accelerated 64 times, as the FP multiplication in FP networks is replaced by Xnor-Bitcount in BNNs." + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.091, + 0.885, + 0.123 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.129, + 0.892, + 0.158 + ], + "angle": 0, + "content": "Figure 2. A general illustration of the forward propagation of the \\(k\\)-th layer in the BNN." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.176, + 0.892, + 0.281 + ], + "angle": 0, + "content": "In the backward propagation, the main challenge is that the pervasive sign functions are theoretically nondifferentiable, and thus extremely destroy the information flow via gradient the propagation. To address this problem, a large number of researchers [31] widely exploit the straight-through estimator (STE) [1] to numerically approximate the derivative of the whole BNN, i.e." + }, + { + "type": "equation", + "bbox": [ + 0.548, + 0.291, + 0.892, + 0.327 + ], + "angle": 0, + "content": "\\[\n\\text {B a c k w a r d :} \\quad \\frac {\\partial \\mathcal {L}}{\\partial x} = \\left\\{ \\begin{array}{l l} \\frac {\\partial \\mathcal {L}}{\\partial s g n (x)} & | x | \\leq 1 \\\\ 0 & | x | > 1. \\end{array} \\right. \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.338, + 0.892, + 0.413 + ], + "angle": 0, + "content": "It is worth noting that we do not implement the aforementioned vanilla approximation method in practice, while we utilize the prevalent Bi-Real [23] and IR-Net [32] to gradually approximate the sign function, which have been proven to be better estimation approaches [29, 30]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.414, + 0.892, + 0.504 + ], + "angle": 0, + "content": "Even though numerous methods have been proposed to eliminate the deterioration of the information flow induced by the binarization, the deterioration is still inevitable due to the severe accuracy loss of weights, activations and gradients [29, 30]. Consequently, binarization destroys the performance of the complicated computer vision models." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.515, + 0.738, + 0.529 + ], + "angle": 0, + "content": "3.2. Multitask Dense Predictor" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.539, + 0.892, + 0.704 + ], + "angle": 0, + "content": "After deploying binarization techniques in the models for dense prediction tasks, the performance of the binarized models is unacceptable, as shown in Fig. 1 and the binary single result in Table. 3. Since the architectures of those SoTA dense prediction models are relatively heavier and deeper (e.g., HRNet-48 or ResNet-101 with a task-specific head [27]) than the ones for classification (e.g., ResNet-18 with a fully-connected layer as the classification head). Moreover, the information passing in the binary models via back-propagation, especially in deep models, is notoriously inefficient [23]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.705, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Dense prediction tasks can mutually supplement information, e.g., surface normals and depth can directly be derived from each other, which can be modeled as the regularization of each other [47]. The relevancy among dense prediction tasks is worth being utilized to improve the overall performance of models. For example, before the deep learning era, pioneering work [11] utilizes RGB-D images with depth information to predict scene semantics to improve the quality of the prediction. In the deep learning era, recent attention-based multitask learning methods [2, 47, 49] explicitly and implicitly distill information from other tasks as a complementary signal to improve the targeted task performance. Briefly, the above-mentioned methods are achieved" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "15901" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.12, + 0.09, + 0.478, + 0.259 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.48, + 0.091, + 0.705, + 0.258 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.708, + 0.091, + 0.849, + 0.258 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.265, + 0.895, + 0.322 + ], + "angle": 0, + "content": "Figure 3. (Left) The illustration of the baseline multitask framework. (Middle) The designed MMD modules for binary representations. Importantly, the MMD module can pass information among different predictions, acting as a cross-talk mechanism. (Right) As all fundamental modules in Bi-MTPD baseline are binarized, inferences can be performed by complete Bool operations, which are very computationally cheap." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.338, + 0.471, + 0.384 + ], + "angle": 0, + "content": "by combining an existing backbone network for initial task predictions with a subsequent decoding process, as shown in Fig. 3 (Left)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.4, + 0.473, + 0.689 + ], + "angle": 0, + "content": "Specifically, the shared features extracted by the backbone network are then processed by a set of task-specific heads, which produce a series of initial predictions for \\( T \\) tasks, i.e. \\( \\{Y_i^k\\} (k = 1,\\dots ,T) \\) (the backbone and the task-specific heads are referred as the front-end of the network [47]). Transforming and binarizing \\( Y_{i}^{t} \\) into the form of a 1-bit feature map, we obtain a set of corresponding binary feature maps of the scene, i.e. \\( \\{F_{B,i}^{t}\\} (t = 1,\\dots ,T) \\) which are more task-aware than the shared binary features of the backbone network. The information from these task-specific feature representations is then fused via a multimodal distillation via binarized attention mechanism before making the final task predictions. As previous work featured, our method is also task-saleable. Especially, it is possible that some tasks are only predicted in the front-end of the network (initial prediction). The initial tasks are also called auxiliary tasks since they serve as proxies in order to improve the performance of the final tasks, as shown in Fig. 3." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.704, + 0.473, + 0.901 + ], + "angle": 0, + "content": "Multi-Modal Distillation (MMD) via Binarized Attention Mechanism. The multi-modal distillation module is the key in the multi-task dense prediction model. Specifically, we utilize the attention mechanism for guiding the information passing between the binary feature maps generated from different modalities for different tasks. Since the passed information flow is not always helpful, the attention can act as a gate function to control the flow, in other words, to make the network automatically learn to focus or to ignore information from other binary features [47, 49]. Including the binarization operations, we can formalize the MMD via binarized attention as follows. While passing information to the \\(k\\)-th task, we first obtain a binarized attenuation matrix \\(\\mathbf{A}\\), and then use it to compute the attention matrix \\(\\mathbf{B}\\)." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.338, + 0.634, + 0.355 + ], + "angle": 0, + "content": "tion map \\(\\mathbf{A}_{B,i}^{k}\\) , i.e.," + }, + { + "type": "equation", + "bbox": [ + 0.598, + 0.364, + 0.892, + 0.384 + ], + "angle": 0, + "content": "\\[\n\\mathbf {A} _ {B, i} ^ {k} \\leftarrow b o o l \\left(\\mathbf {W} _ {B} ^ {k} \\otimes \\mathbf {F} _ {B, i} ^ {k}\\right) \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.393, + 0.893, + 0.455 + ], + "angle": 0, + "content": "where \\(\\mathbf{W}_B^k\\) is the parameters of the binarized convolution layer, \\(\\mathbf{F}_{B,i}^{k}\\) is the binary feature map of the initial prediction, and \\(\\otimes\\) denotes convolution operation. Then the information is passed with the attention map controlled as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.503, + 0.462, + 0.892, + 0.524 + ], + "angle": 0, + "content": "\\[\n\\mathbf {F} _ {B, i} ^ {o, k} \\leftarrow s i g n \\left[ \\mathbf {F} _ {B, i} ^ {k} + \\sum_ {t = 1, t \\neq k} ^ {T} \\mathbf {A} _ {B, i} ^ {k} \\odot \\left(\\mathbf {W} _ {B, t} \\otimes \\mathbf {F} _ {B, i} ^ {t}\\right) \\right] \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.524, + 0.892, + 0.644 + ], + "angle": 0, + "content": "where \\(\\odot\\) element-wise multiplication. The general demonstration of the distillation process is presented in Fig. 3 Left. The output binary feature map \\(\\mathbf{F}_{B,i}^{o,k}\\) is then used by the head for the corresponding \\(t\\)-th task in Fig. 3 Right. By using the task-specific distillation activations, the network can preserve more information for each task [2, 47, 49], which especially benefits the BNNs where the deteriorated information flow mainly induce the performance drop." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.645, + 0.893, + 0.854 + ], + "angle": 0, + "content": "On the other hand, multitask dense prediction models benefit from the network binarization in terms of performance. Although these multitask models have achieved a promising performance, they are still limited by the nature of convolution-based distillation modules that are heavily used in a multi-scale way, which model critical spatial and task-related interactions in relatively local perceptive fields [50]. Theoretically, more distillation modules in different network nodes can contribute to model performance, yet we cannot unrestrictedly add distillation modules to the existing model due to the computational limitation. Fortunately, with the saved computational cost of the binary networks, we can implement additional distillation modules in our model." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.856, + 0.894, + 0.903 + ], + "angle": 0, + "content": "Binary Baseline for Multitask Dense Prediction, Bi-MTPD. To obtain dense predictions with BNNs under a multitask learning framework, we practically binarize" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "15902" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.123, + 0.089, + 0.852, + 0.29 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.295, + 0.895, + 0.339 + ], + "angle": 0, + "content": "Figure 4. The pipeline of Bi-MTPD. We introduce a VIB layer after the backbone network to filter-out the nuisance factors which may lead to model overfitting issue in the forward propagation. In addition, we deploy a feature-based knowledge distillation mechanism to guide the optimization direction in the backward propagation." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.354, + 0.471, + 0.43 + ], + "angle": 0, + "content": "the MTI-Net [47] as the binary baseline. Specifically, the main modules in the full-precision MTI-Net, including backbone, heads, and multi-modal distillation module, are replaced with binary modules (both weights and activations are 1-bit). We call this baseline as Bi-MTPD." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.44, + 0.383, + 0.456 + ], + "angle": 0, + "content": "3.3. Information Flow Supplementation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.463, + 0.47, + 0.689 + ], + "angle": 0, + "content": "Although we build a fully binarized baseline Bi-MTPD for multitask dense predictions and train the pipeline with common techniques, the performance is still of major concern. The baseline suffers an immense information degradation as the nuisance factors are over-fitted in the forward propagation and optimization directions severely mismatch in the backward propagation. To solve these problems, in this section, we further propose the variant of Bi-MTPD, Bi-MTPD-F. Specifically, we introduce a variational information bottleneck (VIB) layer after the output of the shared binary backbone to precisely enforces the feature extractor to preserve the minimal sufficient information of the input data. As well, we deploy the feature-based knowledge distillation to guided the optimization direction. We present more details in the following section." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.689, + 0.471, + 0.901 + ], + "angle": 0, + "content": "Variational Information Bottleneck for Filter-Out Nuisance Factors. Obtaining the initial binary representations of input images by the shared backbone, we need to train a series of targeted heads to split them. A straightforward strategy is to feed these representations into the following MMD module. However, the binarized representations lack homogenization leading to model overfitting issue [48]. Therefore, the need to regularize the binarized representations, while the regularization would not to contaminate the information flow in the representations. Fortunately, the information bottleneck (IB) principle directly relates to compression with the best hypothesis that the data misfit and the model complexity should simultaneously be minimized [43, 48]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.354, + 0.892, + 0.49 + ], + "angle": 0, + "content": "As the VIB could effectively capture the relevant parts and filter out the irrelevant ones from inputs, we design a novel VIB-based layer after the backbone. In particular, it explicitly enforces the feature extractor to preserve the minimal sufficient information of the input data. In other words, it can help ensure the information flow flexibly to learn clean representation for the targeted tasks. The objective function of our VIB-based classification can be formulated as a term of information loss, written as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.592, + 0.498, + 0.892, + 0.516 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {v i b} = K L \\left[ p (\\mathbf {Z} \\mid \\mathbf {A} _ {\\mathbf {B}}), r (\\mathbf {Z}) \\right], \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.523, + 0.892, + 0.642 + ], + "angle": 0, + "content": "where \\(\\mathbf{A_B}\\) is the input binary backbone representation, \\(\\mathbf{B}\\) is the latent representation variable, \\(p(\\mathbf{Z} \\mid \\mathbf{A_B})\\) is a multivariate Gaussian distribution, and \\(r(\\mathbf{Z})\\) is a standard normal distribution. Generally, the latter is a regularization term controlling how much information of the input is filtered out. A more detailed discussion about the VIB for binarized models filtering out irrelevant information is in Supplemental Materials." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.644, + 0.892, + 0.855 + ], + "angle": 0, + "content": "Feature-based Knowledge Distillation for Guiding the Direction of Information Flow. Distillation is a common and essential optimization approach to alleviate the performance drop of quantized models on ultra-low bit-width settings, which can be flexibly deployed for any architectures to utilize the knowledge of a full-precision teacher model [3, 16, 30, 51]. The usual practice is to distill the activations in a layerwise manner from the full-precision teacher to the quantized counterparts, i.e., \\(\\mathbf{F}_{B,l}\\) and \\(\\mathbf{F}_{FP,l}\\) (\\(l = 1,\\dots ,L\\), where \\(L\\) represents the number of network layers), respectively. We use the mean squared errors (MSE) as the distance function to measure the difference between corresponding features student and teacher. The knowledge distillation loss can be written as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.592, + 0.863, + 0.892, + 0.904 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {k d} = \\sum_ {l = 1} ^ {L} M S E \\left(\\mathbf {F} _ {B, l}, \\mathbf {F} _ {F P, l}\\right). \\tag {5}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15903" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.151, + 0.094, + 0.268, + 0.163 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.274, + 0.096, + 0.387, + 0.163 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.393, + 0.097, + 0.505, + 0.162 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.51, + 0.096, + 0.653, + 0.191 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.655, + 0.091, + 0.821, + 0.191 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.152, + 0.164, + 0.269, + 0.223 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.276, + 0.164, + 0.387, + 0.223 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.393, + 0.192, + 0.505, + 0.223 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.52, + 0.192, + 0.653, + 0.289 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.661, + 0.192, + 0.821, + 0.293 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.152, + 0.228, + 0.269, + 0.29 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.276, + 0.228, + 0.386, + 0.304 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.393, + 0.228, + 0.505, + 0.303 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.576, + 0.303, + 0.599, + 0.319 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image_caption", + "bbox": [ + 0.733, + 0.303, + 0.753, + 0.319 + ], + "angle": 0, + "content": "(c)" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.323, + 0.894, + 0.352 + ], + "angle": 0, + "content": "Figure 5. (a) Grad-cam visualization of feature maps of different multitask dense prediction methods. (b) t-SNE visualization of learned features of all 20 classes on Pascal-Context. (c) Centered Kernel Alignment analyzing the information flow within different networks." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.359, + 0.427, + 0.374 + ], + "angle": 0, + "content": "3.4. Counter-Intuitive Results of Bi-MTDP-A" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.385, + 0.473, + 0.702 + ], + "angle": 0, + "content": "Intuitively, implementing binarization on FP network inevitably induces representations degradation, as the gradient of the sign function cannot be perfectly estimated [1]. Thus, binarized models are impossible to outperform their full-precision counterpart models. However, Bi-MTPD-C, a variant of Bi-MTPD (i.e., full-precision backbone with only binarized multi-modal distillation) outperforms its fully FP version. Specifically, just binarizing the multi-modal distillation can simultaneously accelerate the model by \\(\\sim 39\\%\\) and improve the mIoU for segmentation by \\(\\sim 4\\%\\), as shown in Table 1. This result demonstrates that our method is not a naive trade-off between model performance and efficiency but a powerful tool for boosting multitask dense predictors. This exciting 'free-lunch' achievement is even a bit of counter-intuitive. We speculate the reasons are that i) binarization on MMD can filter out task-irrelevant information; ii) and thus the information flow in the network is more effective. To testify this speculation, we conduct a series of experiments in two aspects, the representation ability of Bi-MTPD-C and information flow supplementation within the network." + }, + { + "type": "title", + "bbox": [ + 0.076, + 0.705, + 0.47, + 0.719 + ], + "angle": 0, + "content": "Qualitative Study of Learned Features with" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.72, + 0.47, + 0.901 + ], + "angle": 0, + "content": "Bi-MTPD To investigate the representation ability of Bi-MTPD-C and its FP counterparts, we visualize i) the feature maps behind the Binarized Multi-Modal Distillation (MMD) module in 2-D space via the t-SNE [44] algorithm, and ii) the regions where the network considers important via the Grad-Cam algorithm [35]. The results are shown in Fig. 5. It is clear that binarized model, Bi-MTPD-C is able to filter the irrelevant information out via the binarized attention module (see Fig. 5 (a)), and thus helps learn more discriminative features (see Fig. 5 (b)) resulting in higher quantitative results. Overall, the generated spatial feature maps for segmentation are better. The enhanced" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.36, + 0.892, + 0.388 + ], + "angle": 0, + "content": "representative ability can contribute to higher quantitative results." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.39, + 0.893, + 0.587 + ], + "angle": 0, + "content": "Analysis of Information Flow Supplementation within Network via Centered Kernel Alignment. Analyzing distributional information flow within layers of neural networks is challenging because outputs of layers are distributed across a large number of neurons. Centered kernel alignment (CKA) [6, 22, 33] can address these challenges, by quantitatively comparing activations within or across networks. Specifically, for a network feed by \\( m \\) samples, CKA algorithm takes \\( \\mathbf{X} \\in \\mathbb{R}^{m \\times p_1} \\) and \\( \\mathbf{Y} \\in \\mathbb{R}^{m \\times p_2} \\) as inputs which are output activations of two layers (with \\( p_1 \\) and \\( p_2 \\) neurons respectively). Letting \\( \\mathbf{K} \\triangleq \\mathbf{X}\\mathbf{X}^{\\top} \\) and \\( \\mathbf{L} \\triangleq \\mathbf{Y}\\mathbf{Y}^{\\top} \\) denote the Gram matrices for the two layers CKA computes:" + }, + { + "type": "equation", + "bbox": [ + 0.543, + 0.596, + 0.892, + 0.633 + ], + "angle": 0, + "content": "\\[\n\\operatorname {C K A} (\\mathbf {K}, \\mathbf {L}) = \\frac {\\operatorname {H S I C} (\\mathbf {K} , \\mathbf {L})}{\\sqrt {\\operatorname {H S I C} (\\mathbf {K} , \\mathbf {K}) \\operatorname {H S I C} (\\mathbf {L} , \\mathbf {L})}}, \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.642, + 0.892, + 0.795 + ], + "angle": 0, + "content": "where HSIC is the Hilbert-Schmidt independence criterion [10]. Given the centering matrix \\(\\mathbf{H} = \\mathbf{I}_n - \\frac{1}{n}\\mathbf{1}\\mathbf{1}^\\top\\) and the centered Gram matrices \\(\\mathbf{K}' = \\mathbf{HKH}\\) and \\(\\mathbf{L}' = \\mathbf{HLH}\\), HSIC \\(= \\frac{\\operatorname{vec}(\\mathbf{K}')\\operatorname{vec}(\\mathbf{L}')}{(m - 1)^2}\\), the similarity between these centered Gram matrices. Importantly, CKA is invariant to orthogonal transformation of representations (including permutation of neurons), and the normalization term ensures invariance to isotropic scaling. These properties enable meaningful comparison and analysis of neural network hidden representations." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.796, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Therefore, we introduce CKA to study the information flow in the multitask dense prediction models. In the heatmap, the lighter the dot, the more similar the two corresponding layers. Higher similar score between two layers' output representations means those two layers share more information. The results are presented in Fig. 5 (c), we can see that the similar scores among front layers and back lay" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15904" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.089, + 0.892, + 0.118 + ], + "angle": 0, + "content": "Table 1. Results on NYUD-v2. Bi-MTPD-C: only implementing the binarization in the multi-modal distillation module, Bi-MTPD-F: fully-binarized model." + }, + { + "type": "table", + "bbox": [ + 0.104, + 0.122, + 0.868, + 0.299 + ], + "angle": 0, + "content": "
Model\nMethodBackboneSemSeg ↑Depth ↓Normal ↓Bound ↑float32\nParams (M)binary\nParams (M)FLOPs\n(G)
Cross [28]36.340.62920.8876.38241.460338.09
PAP [53]36.720.61720.8276.42189.100256.86
PSD [54]36.690.62520.8776.42224.670315.60
PAD [49]CNN36.610.62720.8576.38170.980230.91
MTI [47]45.700.53720.2777.86144.870212.98
ATRC [2]45.870.54020.0977.34180.000249.24
MTI + Bi-MTPD-F39.200.61221.0476.866.45138.4218.07
MTI + Bi-MTPD-C47.710.53020.0677.3690.2854.59130.65
InvPT [50]ViT-B50.300.53619.0077.60154.790244.71
InvPT + Bi-MTPD-C51.20(0.90↑)0.528(0.08↓)19.50(0.50↑)77.68(0.08↑)127.9226.87183.81
InvPT [50]ViT-L53.560.51819.0478.10239.220331.60
InvPT + Bi-MTPD-C54.86(1.30↑)0.515(0.03↓)19.50(0.46↑)78.20(0.10↑)212.3426.87301.88
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.303, + 0.471, + 0.365 + ], + "angle": 0, + "content": "ers in Bi-MTPD-C is much higher than the ones in MTI-Net [47]. This indicates that Bi-MTPD-C is able to supplement information flow within network, and thus boost the model performance." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.376, + 0.21, + 0.394 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.401, + 0.47, + 0.538 + ], + "angle": 0, + "content": "In this section, we conducted comprehensive experiments to evaluate our proposed method on two datasets for dense prediction tasks: PASCAL Context [8] and NYUD-v2 [41]. We first describe the implementation details of Bi-MTPD, and then compare our method with SoTA binary neural networks in the task of object detection to demonstrate superiority of the proposed method. Finally, we validate the effectiveness of information bottleneck and feature-based knowledge distillation by a series of ablative studies." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.546, + 0.462, + 0.563 + ], + "angle": 0, + "content": "4.1. Datasets, Evaluation, Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.569, + 0.47, + 0.734 + ], + "angle": 0, + "content": "Datasets. PASCAL-Content is a popular dataset for dense prediction tasks. We use the split from PASCAL-Context which has annotations for semantic segmentation, human part segmentation, semantic edge detection, surface normals prediction and saliency detection. Note that some annotations were distilled by [27] using pre-trained SoTA models [4]. NYUD-v2 contains various indoor scenes such as offices and living rooms with 795 training and 654 testing images. It provides different dense labels, including semantic segmentation, monocular depth estimation, surface normal estimation and object boundary detection." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.735, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Evaluation. Semantic segmentation (Semseg) and human parsing (Parsing) are evaluated with mean Intersection over Union (mIoU); monocular depth estimation (Depth) is evaluated with Root Mean Square Error (RMSE); surface normal estimation (Normal) is evaluated by the mean error (mErr) of predicted angles; saliency detection (Saliency) is evaluated with maximal F-measure (maxF); object boundary detection (Boundary) is evaluated with the optimal-dataset-scale F-measure (odsF). To evaluate the model efficiency w.r.t. memory cost and inference speed, we adopt the number of parameters and FLOPs for a single round of" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.303, + 0.736, + 0.318 + ], + "angle": 0, + "content": "the model inferring an input image." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.319, + 0.892, + 0.668 + ], + "angle": 0, + "content": "Implementation details. We build our approach on the most prevalent backbone architecture, i.e., HRNet as previous SoTA methods [2, 47]. The task-specific heads are also implemented as two basic residual blocks, i.e., binarized BasicBlock and binarized Bottleneck with additional binary shortcuts as BiReal-Net [23]. We use \\(\\ell_1\\) loss for depth estimation and cross-entropy loss for semantic segmentation on NYUD-v2. As in the prior work, the edge detection task is trained with a positive weighted wpos = 0:95 binary cross-entropy loss. We do not adopt a particular loss weighing strategy on NYUD-v2, but simply sum the losses together. On PASCAL, we reuse the training setup from [47] to facilitate a fair comparison. We reuse the loss weights from there. The initial task predictions in the front-end of the network use the same loss weighing as the final task predictions. We refer to the supplementary material for further implementation details. Importantly, we use Adam optimizer [20] in training, but with different learning rates for binary parameters (1e-5) and FP parameters (1e-4), as Adam with a larger learning rate for binary parameters can lead to better training results [24]. Note that our project is based on the codebases for MTI-Net [47] and more details can be found in the codes in the Supplemental Materials." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.68, + 0.798, + 0.696 + ], + "angle": 0, + "content": "4.2. Comparison with State-of-the-Art" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.704, + 0.892, + 0.903 + ], + "angle": 0, + "content": "Tabs. 1 and 2 present a comparative analysis of the proposed Binary Multitask Dense Predictor (Bi-MTPD) with current state-of-the-art models on the NYUD-v2 and PASCAL-Context datasets. This comparison includes notable CNN-based methods such as PAD-Net [49], ASTMT [27], MTI-Net [47], and ATRC [2], among others. Bi-MTPD demonstrates exceptional performance, outperforming other models in 6 of the 9 evaluated metrics, particularly in complex scene understanding tasks like Semantic Segmentation and Parsing. Remarkably, on the NYUD-v2 benchmark, Bi-MTPD surpasses the previously best-performing CNN-based method (ATRC) by a margin of +1.8 (mIoU) in Semantic Segmentation, while requiring" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15905" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.089, + 0.892, + 0.117 + ], + "angle": 0, + "content": "Table 2. Results on PASCAL-VOC. Bi-MTPD-C: only implementing the binarization in the multi-modal distillation module, Bi-MTPD-F: fully-binarized model." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.122, + 0.917, + 0.292 + ], + "angle": 0, + "content": "
Model\nMethodBackboneSemSeg ↑Parsing ↑Saliency ↑Normal ↓Bound ↑float32\nParams (M)binary\nParams (M)FLOPs\n(G)
ASTMT [27]68.0061.1065.7014.7072.40364.720501.27
PAD [49]53.6059.6065.8015.3072.50231.800289.46
MTI [47]61.7060.1884.7814.2370.80218.560280.12
ATRC [2]62.6959.4284.7014.2070.96241.450310.19
ATRC-A [2]63.6062.2383.9114.3070.86249.870320.57
ATRC-B [2]67.6762.9382.2914.2472.42280.010383.21
MTI [47] + Bi-MTPD-F48.1056.2864.4214.2976.8913.67204.8931.38
MTI [47] + Bi-MTPD-C62.98(1.28↑)60.44(0.26↑)83.56(1.22↓)14.31(0.08↑)71.28(0.26↑)153.7164.85194.51
InvPT[50]ViT-B77.5066.8383.6514.6373.00176.350274.68
InvPT[50] + Bi-MTPD-C76.84 (0.66↓)67.10(0.27↑)84.97 (1.32↑)13.69(0.94↓)73.04 (0.04↑)154.6821.67220.76
InvPT[50]ViT-L79.0367.6184.8114.1573.00422.930425.37
InvPT[50] + Bi-MTPD-C79.83 (0.80↑)68.17(0.56↑)
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.3, + 0.457, + 0.33 + ], + "angle": 0, + "content": "only \\(62\\%\\) of the storage space for weights and \\(56\\%\\) of the computational FLOPs." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.331, + 0.457, + 0.42 + ], + "angle": 0, + "content": "Furthermore, the application of Bi-MTPD to the ViT-based state-of-the-art method, InvPT [50], shows cases Bi-MTPD's ability to enhance model performance while also improving efficiency. This demonstrates the broad applicability and generalization capability of Bi-MTPD across different architectural frameworks." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.422, + 0.457, + 0.497 + ], + "angle": 0, + "content": "For a qualitative assessment, Figs. 7 and 6 displays prediction examples from various models. These examples illustrate that Bi-MTPD-C not only competes with but occasionally surpasses the state-of-the-art ATRC in qualitative performance." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.51, + 0.239, + 0.525 + ], + "angle": 0, + "content": "4.3. Ablative Studies" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.534, + 0.47, + 0.671 + ], + "angle": 0, + "content": "Tab. 3 shows a series of ablative studies of Bi-MTPD-C and Bi-MTPD-F with an HRNet-48 backbone on NYUD-v2. We verify how different components of our model contribute to the multi-task improvements. In summary, every designed module positively impacts the overall model. We would like to highlight the main intuition of our method that binarized dense prediction models in a multitask manner largely outperform the binarized single models through Bi-MTPD-F w.o. VIB & KD vs. Bi-Single." + }, + { + "type": "table_caption", + "bbox": [ + 0.159, + 0.68, + 0.396, + 0.692 + ], + "angle": 0, + "content": "Table 3. Ablative studies on NYUD-v2." + }, + { + "type": "table", + "bbox": [ + 0.086, + 0.697, + 0.47, + 0.802 + ], + "angle": 0, + "content": "
ModelSemSeg ↑Depth ↓Normal ↓Bound ↑
Bi-MTPD-C47.710.53020.0677.36
Bi-MTPD-C w.o. binary MMD45.700.53720.2777.86
Bi-MTPD-F39.200.61221.0476.86
Bi-MTPD-F w.o. VIB35.040.65222.3173.88
Bi-MTPD-F w.o. KD36.200.64021.9974.86
Bi-MTPD-F w.o. VIB & KD33.990.66723.2471.30
Bi-Single16.200.87229.3669.24
" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.815, + 0.195, + 0.83 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.469, + 0.903 + ], + "angle": 0, + "content": "In this paper, we significantly accelerate cumbersome dense prediction models, in which BNNs for relevant tasks are modeled and optimized under a multitask framework to supplement degraded information caused by binarization op" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.298, + 0.892, + 0.427 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.502, + 0.461, + 0.892, + 0.624 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.628, + 0.892, + 0.656 + ], + "angle": 0, + "content": "Figure 7. Qualitative comparison with MTI-Net on PASCAL-Context." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.665, + 0.892, + 0.801 + ], + "angle": 0, + "content": "erations. Based on this binary baseline, we further introduce variational information bottleneck and feature-based knowledge distillation to supplement information flow. Experiment results show that our method significantly accelerates existing SoTA methods with comparably small performance drop over the mainstream dense prediction tasks on PASCAL VOC and NYUD-v2. Intriguingly, Bi-MTPD not only reaches SoTA w.r.t. performance but also saves computational costs, compared with SoTA method ARTC [2]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.801, + 0.892, + 0.862 + ], + "angle": 0, + "content": "Acknowledgments: This research is partially supported by NSF IIS-2309073, ECCS-212352101 and Cisco unrestricted gift. This article solely reflects the opinions and conclusions of its authors and not the funding agencies." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15906" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.156 + ], + "angle": 0, + "content": "[1] Yoshua Bengio, Nicholas Léonard, and Aaron Courville. Estimating or propagating gradients through stochastic neurons for conditional computation. arXiv:1308.3432, 2013. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.157, + 0.47, + 0.21 + ], + "angle": 0, + "content": "[2] David Brüggemann, Menelaos Kanakis, Anton Obukhov, Stamatios Georgoulis, and Luc Van Gool. Exploring relational context for multi-task dense prediction. In ICCV, 2021. 1, 2, 3, 4, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.21, + 0.469, + 0.237 + ], + "angle": 0, + "content": "[3] Adrian Bulat, Brais Martinez, and Georgios Tzimiropoulos. High-capacity expert binary networks. In ICLR, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.238, + 0.469, + 0.29 + ], + "angle": 0, + "content": "[4] Liang-Chieh Chen, Yukun Zhu, George Papandreou, Florian Schroff, and Hartwig Adam. Encoder-decoder with atrous separable convolution for semantic image segmentation. In ECCV, 2018. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.291, + 0.469, + 0.344 + ], + "angle": 0, + "content": "[5] Zhao Chen, Vijay Badrinarayanan, Chen-Yu Lee, and Andrew Rabinovich. Gradnorm: Gradient normalization for adaptive loss balancing in deep multitask networks. In ICML, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.345, + 0.468, + 0.384 + ], + "angle": 0, + "content": "[6] Corinna Cortes, Mehryar Mohri, and Afshin Rostamizadeh. Algorithms for learning kernels based on centered alignment. JMLR, 2012. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.385, + 0.469, + 0.426 + ], + "angle": 0, + "content": "[7] Matthieu Courbariaux, Yoshua Bengio, and Jean-Pierre David. Binaryconnect: Training deep neural networks with binary weights during propagations. In NeurIPS, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.426, + 0.469, + 0.466 + ], + "angle": 0, + "content": "[8] Mark Everingham, Luc Van Gool, Christopher KI Williams, John Winn, and Andrew Zisserman. The pascal visual object classes (voc) challenge. IJCV, 2010. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.466, + 0.469, + 0.519 + ], + "angle": 0, + "content": "[9] Yuan Gao, Jiayi Ma, Mingbo Zhao, Wei Liu, and Alan L Yuille. Nddr-cnn: Layerwise feature fusing in multi-task cnns by neural discriminative dimensionality reduction. In CVPR, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.52, + 0.469, + 0.56 + ], + "angle": 0, + "content": "[10] Arthur Gretton, Kenji Fukumizu, Choon Teo, Le Song, Bernhard Scholkopf, and Alex Smola. A kernel statistical test of independence. In NeurIPS, 2007. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.56, + 0.469, + 0.6 + ], + "angle": 0, + "content": "[11] Saurabh Gupta, Ross Girshick, Pablo Arbeláez, and Jitendra Malik. Learning rich features from rgb-d images for object detection and segmentation. In ECCV, 2014. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.601, + 0.469, + 0.64 + ], + "angle": 0, + "content": "[12] Song Han, Huizi Mao, and William J Dally. Deep compression: Compressing deep neural networks with pruning, trained quantization and huffman coding. In ICLR, 2016. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.64, + 0.469, + 0.693 + ], + "angle": 0, + "content": "[13] Xiangyu He, Zitao Mo, Ke Cheng, Weixiang Xu, Qinghao Hu, Peisong Wang, Qingshan Liu, and Jian Cheng. Proxybnn: Learning binarized neural networks via proxy matrices. In CVPR, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.693, + 0.469, + 0.722 + ], + "angle": 0, + "content": "[14] Geoffrey Hinton, Oriol Vinyls, and Jeff Dean. Distilling the knowledge in a neural network. In NeurIPS, 2014. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.722, + 0.469, + 0.761 + ], + "angle": 0, + "content": "[15] Itay Hubara, Matthieu Courbariaux, Daniel Soudry, Ran ElYaniv, and Yoshua Bengio. Binarized neural networks. In NeurIPS, 2016. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.761, + 0.469, + 0.814 + ], + "angle": 0, + "content": "[16] Xiaoqi Jiao, Yichun Yin, Lifeng Shang, Xin Jiang, Xiao Chen, Linlin Li, Fang Wang, and Qun Liu. Tinybert: Distilling bert for natural language understanding. In EMNLP, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.814, + 0.469, + 0.855 + ], + "angle": 0, + "content": "[17] Alex Kendall, Yarin Gal, and Roberto Cipolla. Multi-task learning using uncertainty to weigh losses for scene geometry and semantics. In CVPR, 2018. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.855, + 0.469, + 0.897 + ], + "angle": 0, + "content": "[18] Hyungjun Kim, Kyungsu Kim, Jinseok Kim, and Jae-Joon Kim. Binary duo: Reducing gradient mismatch in binary activation network by coupling binary activations. In ICLR," + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.47, + 0.897 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.534, + 0.093, + 0.583, + 0.105 + ], + "angle": 0, + "content": "2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.106, + 0.892, + 0.146 + ], + "angle": 0, + "content": "[19] Hyungjun Kim, Jihoon Park, Changhun Lee, and Jae-Joon Kim. Improving accuracy of binary neural networks using unbalanced activation distribution. In CVPR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.147, + 0.892, + 0.172 + ], + "angle": 0, + "content": "[20] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In ICLR, 2014. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.172, + 0.892, + 0.226 + ], + "angle": 0, + "content": "[21] Iasonas Kokkinos. Ethernet: Training a universal convolutional neural network for low-, mid-, and high-level vision using diverse datasets and limited memory. In CVPR, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.227, + 0.892, + 0.267 + ], + "angle": 0, + "content": "[22] Simon Kornblith, Mohammad Norouzi, Honglak Lee, and Geoffrey Hinton. Similarity of neural network representations revisited. In ICML, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.267, + 0.892, + 0.319 + ], + "angle": 0, + "content": "[23] Zechun Liu, Wenhan Luo, Baoyuan Wu, Xin Yang, Wei Liu, and Kwang-Ting Cheng. Bi-real net: Binarizing deep network towards real-network performance. IJCV, 2020. 2, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.321, + 0.892, + 0.361 + ], + "angle": 0, + "content": "[24] Zechun Liu, Zhiqiang Shen, Shichao Li, Koen Helwegen, Dong Huang, and Kwang-Ting Cheng. How do adam and training strategies help bnns optimization. In ICML, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.361, + 0.892, + 0.415 + ], + "angle": 0, + "content": "[25] Zechun Liu, Zhiqiang Shen, Marios Savvides, and Kwang-Ting Cheng. Reactnet: Towards precise binary neural network with generalized activation functions. In ECCV, 2020. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.415, + 0.892, + 0.469 + ], + "angle": 0, + "content": "[26] Yongxi Lu, Abhishek Kumar, Shuangfei Zhai, Yu Cheng, Tara Javidi, and Rogerio Feris. Fully-adaptive feature sharing in multi-task networks with applications in person attribute classification. In CVPR, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.469, + 0.892, + 0.509 + ], + "angle": 0, + "content": "[27] Kevis-Kokitsi Maninis, Ilija Radosavovic, and Iasonas Kokkinos. Attentive single-tasking of multiple tasks. In CVPR, 2019. 3, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.51, + 0.892, + 0.55 + ], + "angle": 0, + "content": "[28] Ishan Misra, Abhinav Shrivastava, Abhinav Gupta, and Martial Hebert. Cross-stitch networks for multi-task learning. In CVPR, 2016. 1, 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.55, + 0.892, + 0.602 + ], + "angle": 0, + "content": "[29] Haotong Qin, Zhongang Cai, Mingyuan Zhang, Yifu Ding, Haiyu Zhao, Shuai Yi, Xianglong Liu, and Hao Su. Bipoint-net: Binary neural network for point clouds. In ICLR, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.602, + 0.892, + 0.644 + ], + "angle": 0, + "content": "[30] Haotong Qin, Yifu Ding, Mingyuan Zhang, YAN Qinghua, Aishan Liu, Qingqing Dang, Ziwei Liu, and Xianglong Liu. Bibert: Accurate fully binarized bert. In ICLR, 2022. 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.644, + 0.892, + 0.685 + ], + "angle": 0, + "content": "[31] Haotong Qin, Ruihao Gong, Xianglong Liu, Xiao Bai, Jingkuan Song, and Nicu Sebe. Binary neural networks: A survey. PR, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.684, + 0.892, + 0.738 + ], + "angle": 0, + "content": "[32] Haotong Qin, Ruihao Gong, Xianglong Liu, Mingzhu Shen, Ziran Wei, Fengwei Yu, and Jingkuan Song. Forward and backward information retention for accurate binary neural networks. In CVPR, 2020. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.738, + 0.892, + 0.792 + ], + "angle": 0, + "content": "[33] Maithra Raghu, Thomas Unterthiner, Simon Kornblith, Chiyuan Zhang, and Alexey Dosovitskiy. Do vision transformers see like convolutional neural networks? In NeurIPS, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.792, + 0.892, + 0.833 + ], + "angle": 0, + "content": "[34] Lothar Reichel and Qiang Ye. Simple square smoothing regularization operators. Electronic Transactions on Numerical Analysis, 33:63, 2009. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.833, + 0.892, + 0.887 + ], + "angle": 0, + "content": "[35] Ramprasaath R Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. Grad-cam: Visual explanations from deep networks via gradient-based localization. In ICCV, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.887, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[36] Ozan Sener and Vladlen Koltun. Multi-task learning as" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15907" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.111, + 0.093, + 0.411, + 0.106 + ], + "angle": 0, + "content": "multi-objective optimization. In NeurIPS, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.107, + 0.469, + 0.146 + ], + "angle": 0, + "content": "[37] Yuzhang Shang, Bingxin Xu, Gaowen Liu, Ramana Rao Kompella, and Yan Yan. Causal-dfq: Causality guided data-free network quantization. In ICCV, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.147, + 0.469, + 0.186 + ], + "angle": 0, + "content": "[38] Yuzhang Shang, Dan Xu, Bin Duan, Ziliang Zong, Liqiang Nie, and Yan Yan. Lipschitz continuity retained binary neural network. In ECCV, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.187, + 0.469, + 0.225 + ], + "angle": 0, + "content": "[39] Yuzhang Shang, Dan Xu, Ziliang Zong, and Yan Yan. Network binarization via contrastive learning. In ECCV, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.227, + 0.469, + 0.267 + ], + "angle": 0, + "content": "[40] Yuzhang Shang, Zhihang Yuan, Bin Xie, Bingzhe Wu, and Yan Yan. Post-training quantization on diffusion models. In CVPR, pages 1972-1981, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.268, + 0.469, + 0.307 + ], + "angle": 0, + "content": "[41] Nathan Silberman, Derek Hoiem, Pushmeet Kohli, and Rob Fergus. Indoor segmentation and support inference from rgbd images. In ECCV, 2012. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.308, + 0.469, + 0.348 + ], + "angle": 0, + "content": "[42] Nitish Srivastava, Geoffrey Hinton, Alex Krizhevsky, Ilya Sutskever, and Ruslan Salakhutdinov. Dropout: a simple way to prevent neural networks from overfitting. JMLR, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.349, + 0.469, + 0.388 + ], + "angle": 0, + "content": "[43] Naftali Tishby, Fernando C Pereira, and William Bialek. The information bottleneck method. arXiv preprint physics/0004057, 2000. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.389, + 0.469, + 0.414 + ], + "angle": 0, + "content": "[44] Laurens Van der Maaten and Geoffrey Hinton. Visualizing data using t-sne. JMLR, 2008. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.415, + 0.469, + 0.453 + ], + "angle": 0, + "content": "[45] Simon Vandenhende, Stamatios Georgoulis, Bert De Brabandere, and Luc Van Gool. Branched multi-task networks: deciding what layers to share. In BMVC, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.455, + 0.469, + 0.494 + ], + "angle": 0, + "content": "[46] S. Vandenhende, S. Georgoulis, W. Van Gansbeke, M. Proesmans, D. Dai, and L. Van Gool. Multi-task learning for dense prediction tasks: A survey. TPAMI, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.495, + 0.469, + 0.535 + ], + "angle": 0, + "content": "[47] Simon Vandenhende, Stamatios Georgoulis, and Luc Van Gool. Mti-net: Multi-scale task interaction networks for multi-task learning. In CVPR, 2020. 1, 2, 3, 4, 5, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.536, + 0.469, + 0.561 + ], + "angle": 0, + "content": "[48] Ziwei Wang, Ziyi Wu, Jiwen Lu, and Jie Zhou. Bidet: An efficient binarized object detector. In CVPR, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.562, + 0.469, + 0.613 + ], + "angle": 0, + "content": "[49] Dan Xu, Wanli Ouyang, Xiaogang Wang, and Nicu Sebe. Pad-net: Multi-tasks guided prediction-and-distillation network for simultaneous depth estimation and scene parsing. In CVPR, 2018, 1, 2, 3, 4, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.615, + 0.469, + 0.655 + ], + "angle": 0, + "content": "[50] Hanrong Ye and Dan Xu. Inverted pyramid multi-task transformer for dense scene understanding. arXiv preprint arXiv:2203.07997, 2022. 1, 2, 4, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.657, + 0.469, + 0.694 + ], + "angle": 0, + "content": "[51] Shang Yuzhang, Duan Bin, Zong Ziliang, Nie Liqiang, and Yan Yan. Lipschitz continuity guided knowledge distillation. In ICCV, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.696, + 0.469, + 0.735 + ], + "angle": 0, + "content": "[52] Hongyi Zhang, Moustapha Cisse, Yann N Dauphin, and David Lopez-Paz. mixup: Beyond empirical risk minimization. In ICLR, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.737, + 0.469, + 0.788 + ], + "angle": 0, + "content": "[53] Zhenyu Zhang, Zhen Cui, Chunyan Xu, Yan Yan, Nicu Sebe, and Jian Yang. Pattern-affinitive propagation across depth, surface normal and semantic segmentation. In CVPR, 2019. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.79, + 0.469, + 0.83 + ], + "angle": 0, + "content": "[54] Ling Zhou, Zhen Cui, Chunyan Xu, Zhenyu Zhang, Chaoqun Wang, Tong Zhang, and Jian Yang. Pattern-structure diffusion for multi-task learning. In CVPR, 2020. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.831, + 0.469, + 0.87 + ], + "angle": 0, + "content": "[55] Barret Zoph and Quoc V Le. Neural architecture search with reinforcement learning. arXiv preprint arXiv:1611.01578, 2016.2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.093, + 0.469, + 0.87 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.956 + ], + "angle": 0, + "content": "15908" + } + ] +] \ No newline at end of file diff --git a/2024/Efficient Multitask Dense Predictor via Binarization/b91bc222-326c-4762-aabe-433ec8965e13_origin.pdf b/2024/Efficient Multitask Dense Predictor via Binarization/b91bc222-326c-4762-aabe-433ec8965e13_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..27070bf32d62dde7077b229bfc2b48ba4c82173e --- /dev/null +++ b/2024/Efficient Multitask Dense Predictor via Binarization/b91bc222-326c-4762-aabe-433ec8965e13_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6951265cdac98827ce71b335e49dce90f9b09773a5bd5a111cb5b8ddaff88f7a +size 2753548 diff --git a/2024/Efficient Multitask Dense Predictor via Binarization/full.md b/2024/Efficient Multitask Dense Predictor via Binarization/full.md new file mode 100644 index 0000000000000000000000000000000000000000..f11136b6480c95698a318054eacdda247ba89c45 --- /dev/null +++ b/2024/Efficient Multitask Dense Predictor via Binarization/full.md @@ -0,0 +1,299 @@ +# Efficient Multitask Dense Predictor via Binarization + +Yuzhang Shang $^{1}$ , Dan Xu $^{2}$ , Gaowen Liu $^{3}$ , Ramana Rao Kompella $^{3}$ , Yan Yan $^{1,\dagger}$ $^{1}$ Illinois Institute of Technology, $^{2}$ HKUST, $^{3}$ Cisco Research + +yzshawn@outlook.com, danxu@cse.ust.hk, {ga + +# Abstract + +Multi-task learning for dense prediction has emerged as a pivotal area in computer vision, enabling simultaneous processing of diverse yet interrelated pixel-wise prediction tasks. However, the substantial computational demands of state-of-the-art (SoTA) models often limit their widespread deployment. This paper addresses this challenge by introducing network binarization to compress resource-intensive multi-task dense predictors. Specifically, our goal is to significantly accelerate multi-task dense prediction models via Binary Neural Networks (BNNs) while maintaining and even improving model performance at the same time. To reach this goal, we propose a Binary Multi-task Dense Predictor, Bi-MTPD, and several variants of Bi-MTPD, in which a multi-task dense predictor is constructed via specified binarized modules. Our systematic analysis of this predictor reveals that performance drop from binarization is primarily caused by severe information degradation. To address this issue, we introduce a deep information bottleneck layer that enforces representations for downstream tasks satisfying Gaussian distribution in forward propagation. Moreover, we introduce a knowledge distillation mechanism to correct the direction of information flow in backward propagation. Intriguingly, one variant of Bi-MTPD outperforms full-precision (FP) multi-task dense prediction SoTAs, ARTC [2] (CNN-based) and InvPT [50] (ViT-Based). This result indicates that Bi-MTPD is not merely a naive trade-off between performance and efficiency, but is rather a benefit of the redundant information flow thanks to the multi-task architecture. Code is available at BiMTDP. + +# 1. Introduction + +There is a growing trend in the computer vision community where dense prediction tasks are processed in a multi-task learning manner, such as semantic segmentation, monocular depth estimation, and human parsing [28, 46, 47, 49]. + +![](images/398fc688ef3c7dda29b672f931df9b0b9dfd838f6876b178b5f98f00a1f4f3c8.jpg) +Figure 1. (Left) A conceptual illustration of binary dense predictions in a multi-task manner. In contrast to approaching a series of relevant tasks individually, the multitask model benefits from information supplementation among different tasks via cross-talk structures, but the cumbersome cross-talk modules also add additional computational burden. (Right) Performance summary on NYUD-v2. X-axis and Y-axis denote the performance on depth estimation (lower is better) and segmentation (higher is better), respectively. Size of dots denotes FLOPs. ATRC [2] and InvPT [50] are previous CNN-based and ViT-based SoTAs, respectively. + +![](images/48aa45f0018c86e6c50f5d5054f6e6cd1b9a034290ae3bc61e4395b1df103f7c.jpg) + +Benefited from the information supplementation mechanism via cross-talk structures in the multi-task models, the overall performance for the series of dense prediction tasks has been greatly improved [17] (see Fig. 1). However, the computational demands of State-of-the-Art (SoTA) multitask dense prediction models, which process multiple complex pixel-wise tasks concurrently, are substantial. This high computational requirement limits their application in resource-constrained environments like autonomous driving, robotics, and virtual reality. Our goal is to optimize these heavy SoTA models for edge devices, balancing speed with performance. + +Several strategies for neural network compression have been explored, including pruning [12], network quantization [15, 37, 40] and knowledge distillation [14]. Notably, network binarization, a form of quantization, minimizes weights and activations to $\pm 1$ , enabling the replacement of computationally intensive inner-product operations in full-precision networks with more efficient xnor-bitcount operations in Binary Neural Networks (BNNs) [15]. Binarization theoretically reduces memory costs by $32\times$ and increases inference speed by $64\times$ , making BNNs suitable for edge-device. + +While BNNs have shown impressive results in image classification, achieving nearly full-precision ResNet-level accuracy [25], their application has largely been limited to small-scale models, overlooking other computationally intensive computer vision tasks [15, 23, 25, 32, 39]. Extending BNNs to larger models should be the next step. However, this expansion has been hindered by issues such as overfitting [19, 38, 39] and information degradation [30]. Techniques effective in full-precision models, like label smoothing [34], dropout [42], and mixup [52] have less effect on BNNs [18, 38, 39]. Furthermore, SoTA multitask dense prediction tasks often require deep and complex models, equipped with multi-modality fusion structures [2, 47, 50], exacerbating the challenges in implementing binarization effectively. + +The primary barrier to applying binarization in multitask dense prediction tasks is the significant degradation of information flow in deep models [15, 30, 32], leading to reduced performance. To address this issue, we first propose a Binary Multitask Dense Predictor (Bi-MTPD) baseline, where a multi-task dense predictor is formulated via binarized modules. Based on a thorough review of this baseline, we conclude that the binarization operation destroys the information flows in multi-task models, and thus representations for downstream tasks are not informative compared with their full-precision counterparts. To tackle this problem, we update Bi-MTPD with additional information flow calibration mechanisms in two directions. First, we implement variational information bottleneck enforcing the embeddings to follow Gaussian distribution with sparsity in forward propagation, in order to filter out the task-irrelevant factors. Second, we leverage the existing FP models via feature-based knowledge distillation to calibrate the gradient of the binary network in backward pass. + +The benefits of Bi-MTPD can be analyzed from two orthogonal perspectives. On one hand, from the perspective of network binarization, the accomplishment of bridging binarization with the multitask dense prediction framework testifies that Bi-MTPD can effectively supplement information, and consequently improve the performance of the individual binary models. On the other hand, from the perspective of multitask dense prediction task, accelerating those cumbersome models is profitable to design more effective and efficient cross-talk modules in it, as shown in Fig. 1. Since existing dense prediction models have severe limitation in modelling the cross-talk modules due to their heavy utilization of convolution operation, it is critical for the multitask dense predictions to learn interactions and inference covering various scopes of the multitask context via the cross-talk mechanism [2, 28, 47, 49, 50, 53]. Intriguingly, a variant of Bi-MTPD outperforms SoTA approach ATRC [2] by $4\%$ over the segmentation task while remaining $43\%$ faster in speed, implying that our proposed method + +is not a naive trade-off between performance and efficiency. By empirically investigating this "free lunch" achievement, we conclude that the win-win outcome is benefited from our designed information supplementation mechanism which strengthens the representation ability of the binary model. + +# 2. Related Work + +Multitask Dense Prediction. Multi-Task Learning (MTL) methods can be generally categorized into two main paradigms in terms of the way where model learns shared representations: hard and soft parameter sharing. Hard parameter sharing characterizes architectures which typically share the first hidden representations among the tasks while branching to independent task-specific representations at a later stage. Most approaches split to task-specific heads at a single branch point [5, 17, 21, 36]. However, such naive branching can be sub-optimal, raising interest in mechanisms that allow for finely branched architectures [26, 45]. As a result, in soft parameter sharing, each task is assigned its own set of parameters and a feature sharing mechanism realize the cross-talk as demonstrated in Fig 1. The following works devise the cross-talk mechanisms focusing on the locations in the network where information or features are exchanged or shared between tasks. Apart from the locations, the feature sharing modules are also widely studied. For example, feature fusing can be introduced along the entire network depth [9, 28]; PAD-Net [49] uses multi-modal distillation to enhance task-specific predictions, in which information flow from each source to target task is regulated with a sigmoid-activated gate function; and MTI-Net [47] combines the multi-modal distillation module of PAD-Net with a multi-scale refinement scheme to facilitate cross-task talk at multiple scales. + +Although increasing the number of cross-talk modules intuitively benefits the overall performance of the models, computational cost is often an obstacle. To handle this issue, ATRC [2] introduces NAS [55] to automatically design an efficient information fusing modules. From the perspective of the efficient representation cross-talk, our proposed models with the binarization module can be interpreted as a new pathway to feature fusing within a notably lower inference speed level. + +Neural Network Binarization. As pioneers, [15] use the sign function to quantize weights and activations to $\pm 1$ , initiating the trends of studies of network binarization. To tackle the vanishing gradient issue induced by the bina- rization operations, the straight-through estimator (STE) [1] is introduced for the gradient approximation. Rooted in this archetype, considerable studies contribute to improving the performance of BNNs, particularly on ImageNet. For example, [23] propose Bi-Real introducing double residual connections with FP downsampling layers to mitigate the excessive gradient vanishing issue caused by binariza + +tion, and consequently demonstrate that delicately ing additional connections within BNNs benefits t dient propagation. [13] design a proxy matrix as of the latent parameter space to guide the alignmen weights with different bits by recovering the smooth BNNs. In summary, a large number of methods h tended the boundary of the network binarization w curacy over classification (e.g., ReActNet [25] within comparable FLOPs of binary ResNet-18 achieves $65.9\%$ Top-1 accuracy on ImageNet, while full-precision version is only $70.5\%$ ). + +However, most of those works validate their effectiveness over classification with relatively small architectures (mostly ResNet18 and ResNet34). Meanwhile, the network-based models for dense prediction tasks are bigger and deeper than those toy models, as the information flow in networks is severely degraded. Consequently, directly implementing existing binarization methods can not achieve supposed success. To mitigate the degradation, we propose to binarize those dense prediction models in a multitask way. + +# 3. Multitask Network Binarization + +# 3.1. Binary Neural Network + +To begin with, we briefly review the general idea of binary neural networks (BNNs) in [7, 15]. Here, we only elaborate the speedup mechanism and the degradation of information flow of the binarization. We define a full-precision (FP) neural network with $K$ layers, $f(\mathbf{x}) = (\mathbf{W}^{K}\times \sigma \times \mathbf{W}^{K - 1}\dots \sigma \times \mathbf{W}^{1})(\mathbf{x})$ , where $\mathbf{x}$ is the input sample and $\mathbf{W}^k:\mathbb{R}^{d_{k - 1}}\longmapsto \mathbb{R}^{d_k}(k = 1,\ldots ,K)$ stands for the weight matrix connecting the $(k - 1)$ -th and the $k$ -th layer, with $d_{k - 1}$ and $d_{k}$ representing the sizes of the input and output of the $k$ -th network layer, respectively. The $\sigma (\cdot)$ function performs element-wise activation for the feature maps. + +BNNs vary from FP neural networks in terms of the forward operation and the backward gradient approximation. Specifically, in the forward propagation, the BNN maintains FP latent weights $\mathbf{W}_F$ for gradient updates, and the $k$ -th weight matrix $\mathbf{W}_F^k$ is binarized into $\pm 1$ , obtaining the binary weight matrix $\mathbf{W}_B^k$ via the binarize function $\text{sign}(\cdot)$ , i.e. $\mathbf{W}_B^k = \text{sign}(\mathbf{W}_F^k)$ . Then the intermediate activation map (full-precision) of the $k$ -th layer is produced by $\mathbf{A}_F^k = \mathbf{W}_B^k\mathbf{A}_B^{k-1}$ . The same quantization method is used to binarize the full-precision activation map as $\mathbf{A}_B^k = \text{sign}(\mathbf{A}_F^k)$ , and the whole forward pass of binarization is performed by iterating this process for $L$ times, as shown in Fig. 2. For BNNs, the weights and activations are 1-bit, by which the network is accelerated 32 times in terms of memory cost. Importantly, inference of BNN is accelerated 64 times, as the FP multiplication in FP networks is replaced by Xnor-Bitcount in BNNs. + +![](images/115f534f2a2a4159b6b9673aa0cf3cb76dcc96ce61aea7c3c47ae62c5b39f535.jpg) +Figure 2. A general illustration of the forward propagation of the $k$ -th layer in the BNN. + +In the backward propagation, the main challenge is that the pervasive sign functions are theoretically nondifferentiable, and thus extremely destroy the information flow via gradient the propagation. To address this problem, a large number of researchers [31] widely exploit the straight-through estimator (STE) [1] to numerically approximate the derivative of the whole BNN, i.e. + +$$ +\text {B a c k w a r d :} \quad \frac {\partial \mathcal {L}}{\partial x} = \left\{ \begin{array}{l l} \frac {\partial \mathcal {L}}{\partial s g n (x)} & | x | \leq 1 \\ 0 & | x | > 1. \end{array} \right. \tag {1} +$$ + +It is worth noting that we do not implement the aforementioned vanilla approximation method in practice, while we utilize the prevalent Bi-Real [23] and IR-Net [32] to gradually approximate the sign function, which have been proven to be better estimation approaches [29, 30]. + +Even though numerous methods have been proposed to eliminate the deterioration of the information flow induced by the binarization, the deterioration is still inevitable due to the severe accuracy loss of weights, activations and gradients [29, 30]. Consequently, binarization destroys the performance of the complicated computer vision models. + +# 3.2. Multitask Dense Predictor + +After deploying binarization techniques in the models for dense prediction tasks, the performance of the binarized models is unacceptable, as shown in Fig. 1 and the binary single result in Table. 3. Since the architectures of those SoTA dense prediction models are relatively heavier and deeper (e.g., HRNet-48 or ResNet-101 with a task-specific head [27]) than the ones for classification (e.g., ResNet-18 with a fully-connected layer as the classification head). Moreover, the information passing in the binary models via back-propagation, especially in deep models, is notoriously inefficient [23]. + +Dense prediction tasks can mutually supplement information, e.g., surface normals and depth can directly be derived from each other, which can be modeled as the regularization of each other [47]. The relevancy among dense prediction tasks is worth being utilized to improve the overall performance of models. For example, before the deep learning era, pioneering work [11] utilizes RGB-D images with depth information to predict scene semantics to improve the quality of the prediction. In the deep learning era, recent attention-based multitask learning methods [2, 47, 49] explicitly and implicitly distill information from other tasks as a complementary signal to improve the targeted task performance. Briefly, the above-mentioned methods are achieved + +![](images/c46068f6572247f241dda6efc770bb397ad2576d027021aec506d6cb7d7123e7.jpg) +Figure 3. (Left) The illustration of the baseline multitask framework. (Middle) The designed MMD modules for binary representations. Importantly, the MMD module can pass information among different predictions, acting as a cross-talk mechanism. (Right) As all fundamental modules in Bi-MTPD baseline are binarized, inferences can be performed by complete Bool operations, which are very computationally cheap. + +![](images/ec62c60358444a2011064bc45b84c0d67f5441695e4a243e3db3214a07ab84f0.jpg) + +![](images/be1a3fe3fe8e77b12fda57e37b951d594ec50bea252f9243058c41c7408beb04.jpg) + +by combining an existing backbone network for initial task predictions with a subsequent decoding process, as shown in Fig. 3 (Left). + +Specifically, the shared features extracted by the backbone network are then processed by a set of task-specific heads, which produce a series of initial predictions for $T$ tasks, i.e. $\{Y_i^k\} (k = 1,\dots ,T)$ (the backbone and the task-specific heads are referred as the front-end of the network [47]). Transforming and binarizing $Y_{i}^{t}$ into the form of a 1-bit feature map, we obtain a set of corresponding binary feature maps of the scene, i.e. $\{F_{B,i}^{t}\} (t = 1,\dots ,T)$ which are more task-aware than the shared binary features of the backbone network. The information from these task-specific feature representations is then fused via a multimodal distillation via binarized attention mechanism before making the final task predictions. As previous work featured, our method is also task-saleable. Especially, it is possible that some tasks are only predicted in the front-end of the network (initial prediction). The initial tasks are also called auxiliary tasks since they serve as proxies in order to improve the performance of the final tasks, as shown in Fig. 3. + +Multi-Modal Distillation (MMD) via Binarized Attention Mechanism. The multi-modal distillation module is the key in the multi-task dense prediction model. Specifically, we utilize the attention mechanism for guiding the information passing between the binary feature maps generated from different modalities for different tasks. Since the passed information flow is not always helpful, the attention can act as a gate function to control the flow, in other words, to make the network automatically learn to focus or to ignore information from other binary features [47, 49]. Including the binarization operations, we can formalize the MMD via binarized attention as follows. While passing information to the $k$ -th task, we first obtain a binarized attenuation matrix $\mathbf{A}$ , and then use it to compute the attention matrix $\mathbf{B}$ . + +tion map $\mathbf{A}_{B,i}^{k}$ , i.e., + +$$ +\mathbf {A} _ {B, i} ^ {k} \leftarrow b o o l \left(\mathbf {W} _ {B} ^ {k} \otimes \mathbf {F} _ {B, i} ^ {k}\right) \tag {2} +$$ + +where $\mathbf{W}_B^k$ is the parameters of the binarized convolution layer, $\mathbf{F}_{B,i}^{k}$ is the binary feature map of the initial prediction, and $\otimes$ denotes convolution operation. Then the information is passed with the attention map controlled as follows: + +$$ +\mathbf {F} _ {B, i} ^ {o, k} \leftarrow s i g n \left[ \mathbf {F} _ {B, i} ^ {k} + \sum_ {t = 1, t \neq k} ^ {T} \mathbf {A} _ {B, i} ^ {k} \odot \left(\mathbf {W} _ {B, t} \otimes \mathbf {F} _ {B, i} ^ {t}\right) \right] \tag {3} +$$ + +where $\odot$ element-wise multiplication. The general demonstration of the distillation process is presented in Fig. 3 Left. The output binary feature map $\mathbf{F}_{B,i}^{o,k}$ is then used by the head for the corresponding $t$ -th task in Fig. 3 Right. By using the task-specific distillation activations, the network can preserve more information for each task [2, 47, 49], which especially benefits the BNNs where the deteriorated information flow mainly induce the performance drop. + +On the other hand, multitask dense prediction models benefit from the network binarization in terms of performance. Although these multitask models have achieved a promising performance, they are still limited by the nature of convolution-based distillation modules that are heavily used in a multi-scale way, which model critical spatial and task-related interactions in relatively local perceptive fields [50]. Theoretically, more distillation modules in different network nodes can contribute to model performance, yet we cannot unrestrictedly add distillation modules to the existing model due to the computational limitation. Fortunately, with the saved computational cost of the binary networks, we can implement additional distillation modules in our model. + +Binary Baseline for Multitask Dense Prediction, Bi-MTPD. To obtain dense predictions with BNNs under a multitask learning framework, we practically binarize + +![](images/704a8c0fde1af054bc8734d52148f028127260aab29887a139d76e5ba492c01b.jpg) +Figure 4. The pipeline of Bi-MTPD. We introduce a VIB layer after the backbone network to filter-out the nuisance factors which may lead to model overfitting issue in the forward propagation. In addition, we deploy a feature-based knowledge distillation mechanism to guide the optimization direction in the backward propagation. + +the MTI-Net [47] as the binary baseline. Specifically, the main modules in the full-precision MTI-Net, including backbone, heads, and multi-modal distillation module, are replaced with binary modules (both weights and activations are 1-bit). We call this baseline as Bi-MTPD. + +# 3.3. Information Flow Supplementation + +Although we build a fully binarized baseline Bi-MTPD for multitask dense predictions and train the pipeline with common techniques, the performance is still of major concern. The baseline suffers an immense information degradation as the nuisance factors are over-fitted in the forward propagation and optimization directions severely mismatch in the backward propagation. To solve these problems, in this section, we further propose the variant of Bi-MTPD, Bi-MTPD-F. Specifically, we introduce a variational information bottleneck (VIB) layer after the output of the shared binary backbone to precisely enforces the feature extractor to preserve the minimal sufficient information of the input data. As well, we deploy the feature-based knowledge distillation to guided the optimization direction. We present more details in the following section. + +Variational Information Bottleneck for Filter-Out Nuisance Factors. Obtaining the initial binary representations of input images by the shared backbone, we need to train a series of targeted heads to split them. A straightforward strategy is to feed these representations into the following MMD module. However, the binarized representations lack homogenization leading to model overfitting issue [48]. Therefore, the need to regularize the binarized representations, while the regularization would not to contaminate the information flow in the representations. Fortunately, the information bottleneck (IB) principle directly relates to compression with the best hypothesis that the data misfit and the model complexity should simultaneously be minimized [43, 48]. + +As the VIB could effectively capture the relevant parts and filter out the irrelevant ones from inputs, we design a novel VIB-based layer after the backbone. In particular, it explicitly enforces the feature extractor to preserve the minimal sufficient information of the input data. In other words, it can help ensure the information flow flexibly to learn clean representation for the targeted tasks. The objective function of our VIB-based classification can be formulated as a term of information loss, written as follows: + +$$ +\mathcal {L} _ {v i b} = K L \left[ p (\mathbf {Z} \mid \mathbf {A} _ {\mathbf {B}}), r (\mathbf {Z}) \right], \tag {4} +$$ + +where $\mathbf{A_B}$ is the input binary backbone representation, $\mathbf{B}$ is the latent representation variable, $p(\mathbf{Z} \mid \mathbf{A_B})$ is a multivariate Gaussian distribution, and $r(\mathbf{Z})$ is a standard normal distribution. Generally, the latter is a regularization term controlling how much information of the input is filtered out. A more detailed discussion about the VIB for binarized models filtering out irrelevant information is in Supplemental Materials. + +Feature-based Knowledge Distillation for Guiding the Direction of Information Flow. Distillation is a common and essential optimization approach to alleviate the performance drop of quantized models on ultra-low bit-width settings, which can be flexibly deployed for any architectures to utilize the knowledge of a full-precision teacher model [3, 16, 30, 51]. The usual practice is to distill the activations in a layerwise manner from the full-precision teacher to the quantized counterparts, i.e., $\mathbf{F}_{B,l}$ and $\mathbf{F}_{FP,l}$ ( $l = 1,\dots ,L$ , where $L$ represents the number of network layers), respectively. We use the mean squared errors (MSE) as the distance function to measure the difference between corresponding features student and teacher. The knowledge distillation loss can be written as follows: + +$$ +\mathcal {L} _ {k d} = \sum_ {l = 1} ^ {L} M S E \left(\mathbf {F} _ {B, l}, \mathbf {F} _ {F P, l}\right). \tag {5} +$$ + +![](images/35b40e833845c6eb5484737bc94a8f79a914bbedb6eb12b163c2469013f560ca.jpg) + +![](images/ec1b8852c1b5b6e652375cc3a6736ddf029d063f02b3b7a2df3159aace136216.jpg) + +![](images/8d37ebe5c81add2e10eea6b05065059d823414a3212c146b0e90546bde46e950.jpg) + +![](images/bdbf2de8bf8bcf343b6dbb45b7d8b7510b6314e5d771ce7e89a9f7ba659691b4.jpg) + +![](images/2556a9996ba5f9f78730d228893d9915a55b28310bb0fedcaaac4c082dc91750.jpg) + +![](images/33a22209f064b38db6842f805e4707670e679a262f438c6048c7e26505649294.jpg) + +![](images/731f2bff1ade657246b93d31a804d5958fbdf747f756e721f828d62e42f26fc3.jpg) + +![](images/34b2b0b67743437999c41f3481ad5610bbf23bbf772adcb585c78af5700a95b7.jpg) + +![](images/793d6f89f9c5f2952b2ff237f9b9d16c1a470a682e3ba1a2cdc0c2339fddb773.jpg) +(b) + +![](images/4d10f587d7454ee29ecbdb5cf637a6411fc2bfb89f4d885715672ce042247d1f.jpg) +(c) + +![](images/6b1d80a1dcc2a0df3695da83eda095de313fd402398b4da4a2432c5406a03612.jpg) +Figure 5. (a) Grad-cam visualization of feature maps of different multitask dense prediction methods. (b) t-SNE visualization of learned features of all 20 classes on Pascal-Context. (c) Centered Kernel Alignment analyzing the information flow within different networks. + +![](images/965881496a0205c0e2bae25158d2bcdf03d3f2d14f35bcc7666fe8c9edebc9b7.jpg) + +![](images/c8aa5d41e3c53683f7cbac1eaeacb18e62e14887a1fda2317d957f88c3b22840.jpg) + +# 3.4. Counter-Intuitive Results of Bi-MTDP-A + +Intuitively, implementing binarization on FP network inevitably induces representations degradation, as the gradient of the sign function cannot be perfectly estimated [1]. Thus, binarized models are impossible to outperform their full-precision counterpart models. However, Bi-MTPD-C, a variant of Bi-MTPD (i.e., full-precision backbone with only binarized multi-modal distillation) outperforms its fully FP version. Specifically, just binarizing the multi-modal distillation can simultaneously accelerate the model by $\sim 39\%$ and improve the mIoU for segmentation by $\sim 4\%$ , as shown in Table 1. This result demonstrates that our method is not a naive trade-off between model performance and efficiency but a powerful tool for boosting multitask dense predictors. This exciting 'free-lunch' achievement is even a bit of counter-intuitive. We speculate the reasons are that i) binarization on MMD can filter out task-irrelevant information; ii) and thus the information flow in the network is more effective. To testify this speculation, we conduct a series of experiments in two aspects, the representation ability of Bi-MTPD-C and information flow supplementation within the network. + +# Qualitative Study of Learned Features with + +Bi-MTPD To investigate the representation ability of Bi-MTPD-C and its FP counterparts, we visualize i) the feature maps behind the Binarized Multi-Modal Distillation (MMD) module in 2-D space via the t-SNE [44] algorithm, and ii) the regions where the network considers important via the Grad-Cam algorithm [35]. The results are shown in Fig. 5. It is clear that binarized model, Bi-MTPD-C is able to filter the irrelevant information out via the binarized attention module (see Fig. 5 (a)), and thus helps learn more discriminative features (see Fig. 5 (b)) resulting in higher quantitative results. Overall, the generated spatial feature maps for segmentation are better. The enhanced + +representative ability can contribute to higher quantitative results. + +Analysis of Information Flow Supplementation within Network via Centered Kernel Alignment. Analyzing distributional information flow within layers of neural networks is challenging because outputs of layers are distributed across a large number of neurons. Centered kernel alignment (CKA) [6, 22, 33] can address these challenges, by quantitatively comparing activations within or across networks. Specifically, for a network feed by $m$ samples, CKA algorithm takes $\mathbf{X} \in \mathbb{R}^{m \times p_1}$ and $\mathbf{Y} \in \mathbb{R}^{m \times p_2}$ as inputs which are output activations of two layers (with $p_1$ and $p_2$ neurons respectively). Letting $\mathbf{K} \triangleq \mathbf{X}\mathbf{X}^{\top}$ and $\mathbf{L} \triangleq \mathbf{Y}\mathbf{Y}^{\top}$ denote the Gram matrices for the two layers CKA computes: + +$$ +\operatorname {C K A} (\mathbf {K}, \mathbf {L}) = \frac {\operatorname {H S I C} (\mathbf {K} , \mathbf {L})}{\sqrt {\operatorname {H S I C} (\mathbf {K} , \mathbf {K}) \operatorname {H S I C} (\mathbf {L} , \mathbf {L})}}, \tag {6} +$$ + +where HSIC is the Hilbert-Schmidt independence criterion [10]. Given the centering matrix $\mathbf{H} = \mathbf{I}_n - \frac{1}{n}\mathbf{1}\mathbf{1}^\top$ and the centered Gram matrices $\mathbf{K}' = \mathbf{HKH}$ and $\mathbf{L}' = \mathbf{HLH}$ , HSIC $= \frac{\operatorname{vec}(\mathbf{K}')\operatorname{vec}(\mathbf{L}')}{(m - 1)^2}$ , the similarity between these centered Gram matrices. Importantly, CKA is invariant to orthogonal transformation of representations (including permutation of neurons), and the normalization term ensures invariance to isotropic scaling. These properties enable meaningful comparison and analysis of neural network hidden representations. + +Therefore, we introduce CKA to study the information flow in the multitask dense prediction models. In the heatmap, the lighter the dot, the more similar the two corresponding layers. Higher similar score between two layers' output representations means those two layers share more information. The results are presented in Fig. 5 (c), we can see that the similar scores among front layers and back lay + +Table 1. Results on NYUD-v2. Bi-MTPD-C: only implementing the binarization in the multi-modal distillation module, Bi-MTPD-F: fully-binarized model. + +
Model +MethodBackboneSemSeg ↑Depth ↓Normal ↓Bound ↑float32 +Params (M)binary +Params (M)FLOPs +(G)
Cross [28]36.340.62920.8876.38241.460338.09
PAP [53]36.720.61720.8276.42189.100256.86
PSD [54]36.690.62520.8776.42224.670315.60
PAD [49]CNN36.610.62720.8576.38170.980230.91
MTI [47]45.700.53720.2777.86144.870212.98
ATRC [2]45.870.54020.0977.34180.000249.24
MTI + Bi-MTPD-F39.200.61221.0476.866.45138.4218.07
MTI + Bi-MTPD-C47.710.53020.0677.3690.2854.59130.65
InvPT [50]ViT-B50.300.53619.0077.60154.790244.71
InvPT + Bi-MTPD-C51.20(0.90↑)0.528(0.08↓)19.50(0.50↑)77.68(0.08↑)127.9226.87183.81
InvPT [50]ViT-L53.560.51819.0478.10239.220331.60
InvPT + Bi-MTPD-C54.86(1.30↑)0.515(0.03↓)19.50(0.46↑)78.20(0.10↑)212.3426.87301.88
+ +ers in Bi-MTPD-C is much higher than the ones in MTI-Net [47]. This indicates that Bi-MTPD-C is able to supplement information flow within network, and thus boost the model performance. + +# 4. Experiments + +In this section, we conducted comprehensive experiments to evaluate our proposed method on two datasets for dense prediction tasks: PASCAL Context [8] and NYUD-v2 [41]. We first describe the implementation details of Bi-MTPD, and then compare our method with SoTA binary neural networks in the task of object detection to demonstrate superiority of the proposed method. Finally, we validate the effectiveness of information bottleneck and feature-based knowledge distillation by a series of ablative studies. + +# 4.1. Datasets, Evaluation, Implementation Details + +Datasets. PASCAL-Content is a popular dataset for dense prediction tasks. We use the split from PASCAL-Context which has annotations for semantic segmentation, human part segmentation, semantic edge detection, surface normals prediction and saliency detection. Note that some annotations were distilled by [27] using pre-trained SoTA models [4]. NYUD-v2 contains various indoor scenes such as offices and living rooms with 795 training and 654 testing images. It provides different dense labels, including semantic segmentation, monocular depth estimation, surface normal estimation and object boundary detection. + +Evaluation. Semantic segmentation (Semseg) and human parsing (Parsing) are evaluated with mean Intersection over Union (mIoU); monocular depth estimation (Depth) is evaluated with Root Mean Square Error (RMSE); surface normal estimation (Normal) is evaluated by the mean error (mErr) of predicted angles; saliency detection (Saliency) is evaluated with maximal F-measure (maxF); object boundary detection (Boundary) is evaluated with the optimal-dataset-scale F-measure (odsF). To evaluate the model efficiency w.r.t. memory cost and inference speed, we adopt the number of parameters and FLOPs for a single round of + +the model inferring an input image. + +Implementation details. We build our approach on the most prevalent backbone architecture, i.e., HRNet as previous SoTA methods [2, 47]. The task-specific heads are also implemented as two basic residual blocks, i.e., binarized BasicBlock and binarized Bottleneck with additional binary shortcuts as BiReal-Net [23]. We use $\ell_1$ loss for depth estimation and cross-entropy loss for semantic segmentation on NYUD-v2. As in the prior work, the edge detection task is trained with a positive weighted wpos = 0:95 binary cross-entropy loss. We do not adopt a particular loss weighing strategy on NYUD-v2, but simply sum the losses together. On PASCAL, we reuse the training setup from [47] to facilitate a fair comparison. We reuse the loss weights from there. The initial task predictions in the front-end of the network use the same loss weighing as the final task predictions. We refer to the supplementary material for further implementation details. Importantly, we use Adam optimizer [20] in training, but with different learning rates for binary parameters (1e-5) and FP parameters (1e-4), as Adam with a larger learning rate for binary parameters can lead to better training results [24]. Note that our project is based on the codebases for MTI-Net [47] and more details can be found in the codes in the Supplemental Materials. + +# 4.2. Comparison with State-of-the-Art + +Tabs. 1 and 2 present a comparative analysis of the proposed Binary Multitask Dense Predictor (Bi-MTPD) with current state-of-the-art models on the NYUD-v2 and PASCAL-Context datasets. This comparison includes notable CNN-based methods such as PAD-Net [49], ASTMT [27], MTI-Net [47], and ATRC [2], among others. Bi-MTPD demonstrates exceptional performance, outperforming other models in 6 of the 9 evaluated metrics, particularly in complex scene understanding tasks like Semantic Segmentation and Parsing. Remarkably, on the NYUD-v2 benchmark, Bi-MTPD surpasses the previously best-performing CNN-based method (ATRC) by a margin of +1.8 (mIoU) in Semantic Segmentation, while requiring + +Table 2. Results on PASCAL-VOC. Bi-MTPD-C: only implementing the binarization in the multi-modal distillation module, Bi-MTPD-F: fully-binarized model. + +
Model +MethodBackboneSemSeg ↑Parsing ↑Saliency ↑Normal ↓Bound ↑float32 +Params (M)binary +Params (M)FLOPs +(G)
ASTMT [27]68.0061.1065.7014.7072.40364.720501.27
PAD [49]53.6059.6065.8015.3072.50231.800289.46
MTI [47]61.7060.1884.7814.2370.80218.560280.12
ATRC [2]62.6959.4284.7014.2070.96241.450310.19
ATRC-A [2]63.6062.2383.9114.3070.86249.870320.57
ATRC-B [2]67.6762.9382.2914.2472.42280.010383.21
MTI [47] + Bi-MTPD-F48.1056.2864.4214.2976.8913.67204.8931.38
MTI [47] + Bi-MTPD-C62.98(1.28↑)60.44(0.26↑)83.56(1.22↓)14.31(0.08↑)71.28(0.26↑)153.7164.85194.51
InvPT[50]ViT-B77.5066.8383.6514.6373.00176.350274.68
InvPT[50] + Bi-MTPD-C76.84 (0.66↓)67.10(0.27↑)84.97 (1.32↑)13.69(0.94↓)73.04 (0.04↑)154.6821.67220.76
InvPT[50]ViT-L79.0367.6184.8114.1573.00422.930425.37
InvPT[50] + Bi-MTPD-C79.83 (0.80↑)68.17(0.56↑)
+ +only $62\%$ of the storage space for weights and $56\%$ of the computational FLOPs. + +Furthermore, the application of Bi-MTPD to the ViT-based state-of-the-art method, InvPT [50], shows cases Bi-MTPD's ability to enhance model performance while also improving efficiency. This demonstrates the broad applicability and generalization capability of Bi-MTPD across different architectural frameworks. + +For a qualitative assessment, Figs. 7 and 6 displays prediction examples from various models. These examples illustrate that Bi-MTPD-C not only competes with but occasionally surpasses the state-of-the-art ATRC in qualitative performance. + +# 4.3. Ablative Studies + +Tab. 3 shows a series of ablative studies of Bi-MTPD-C and Bi-MTPD-F with an HRNet-48 backbone on NYUD-v2. We verify how different components of our model contribute to the multi-task improvements. In summary, every designed module positively impacts the overall model. We would like to highlight the main intuition of our method that binarized dense prediction models in a multitask manner largely outperform the binarized single models through Bi-MTPD-F w.o. VIB & KD vs. Bi-Single. + +Table 3. Ablative studies on NYUD-v2. + +
ModelSemSeg ↑Depth ↓Normal ↓Bound ↑
Bi-MTPD-C47.710.53020.0677.36
Bi-MTPD-C w.o. binary MMD45.700.53720.2777.86
Bi-MTPD-F39.200.61221.0476.86
Bi-MTPD-F w.o. VIB35.040.65222.3173.88
Bi-MTPD-F w.o. KD36.200.64021.9974.86
Bi-MTPD-F w.o. VIB & KD33.990.66723.2471.30
Bi-Single16.200.87229.3669.24
+ +# 5. Conclusion + +In this paper, we significantly accelerate cumbersome dense prediction models, in which BNNs for relevant tasks are modeled and optimized under a multitask framework to supplement degraded information caused by binarization op + +![](images/a3ab1f3b0cc7f333ec50d7394c192dbfda0f09f8fb5740b2042e0acaec76f5d2.jpg) + +![](images/810f13c9eee9f584993405b74e8ed3200030253d88bc5f2f9e8968eb887be7be.jpg) +Figure 7. Qualitative comparison with MTI-Net on PASCAL-Context. + +erations. Based on this binary baseline, we further introduce variational information bottleneck and feature-based knowledge distillation to supplement information flow. Experiment results show that our method significantly accelerates existing SoTA methods with comparably small performance drop over the mainstream dense prediction tasks on PASCAL VOC and NYUD-v2. Intriguingly, Bi-MTPD not only reaches SoTA w.r.t. performance but also saves computational costs, compared with SoTA method ARTC [2]. + +Acknowledgments: This research is partially supported by NSF IIS-2309073, ECCS-212352101 and Cisco unrestricted gift. This article solely reflects the opinions and conclusions of its authors and not the funding agencies. + +# References + +[1] Yoshua Bengio, Nicholas Léonard, and Aaron Courville. Estimating or propagating gradients through stochastic neurons for conditional computation. arXiv:1308.3432, 2013. 2, 3, 6 +[2] David Brüggemann, Menelaos Kanakis, Anton Obukhov, Stamatios Georgoulis, and Luc Van Gool. Exploring relational context for multi-task dense prediction. In ICCV, 2021. 1, 2, 3, 4, 7, 8 +[3] Adrian Bulat, Brais Martinez, and Georgios Tzimiropoulos. High-capacity expert binary networks. In ICLR, 2020. 5 +[4] Liang-Chieh Chen, Yukun Zhu, George Papandreou, Florian Schroff, and Hartwig Adam. Encoder-decoder with atrous separable convolution for semantic image segmentation. In ECCV, 2018. 7 +[5] Zhao Chen, Vijay Badrinarayanan, Chen-Yu Lee, and Andrew Rabinovich. Gradnorm: Gradient normalization for adaptive loss balancing in deep multitask networks. In ICML, 2018. 2 +[6] Corinna Cortes, Mehryar Mohri, and Afshin Rostamizadeh. Algorithms for learning kernels based on centered alignment. JMLR, 2012. 6 +[7] Matthieu Courbariaux, Yoshua Bengio, and Jean-Pierre David. Binaryconnect: Training deep neural networks with binary weights during propagations. In NeurIPS, 2016. 3 +[8] Mark Everingham, Luc Van Gool, Christopher KI Williams, John Winn, and Andrew Zisserman. The pascal visual object classes (voc) challenge. IJCV, 2010. 7 +[9] Yuan Gao, Jiayi Ma, Mingbo Zhao, Wei Liu, and Alan L Yuille. Nddr-cnn: Layerwise feature fusing in multi-task cnns by neural discriminative dimensionality reduction. In CVPR, 2019. 2 +[10] Arthur Gretton, Kenji Fukumizu, Choon Teo, Le Song, Bernhard Scholkopf, and Alex Smola. A kernel statistical test of independence. In NeurIPS, 2007. 6 +[11] Saurabh Gupta, Ross Girshick, Pablo Arbeláez, and Jitendra Malik. Learning rich features from rgb-d images for object detection and segmentation. In ECCV, 2014. 3 +[12] Song Han, Huizi Mao, and William J Dally. Deep compression: Compressing deep neural networks with pruning, trained quantization and huffman coding. In ICLR, 2016. 1 +[13] Xiangyu He, Zitao Mo, Ke Cheng, Weixiang Xu, Qinghao Hu, Peisong Wang, Qingshan Liu, and Jian Cheng. Proxybnn: Learning binarized neural networks via proxy matrices. In CVPR, 2020. 3 +[14] Geoffrey Hinton, Oriol Vinyls, and Jeff Dean. Distilling the knowledge in a neural network. In NeurIPS, 2014. 1 +[15] Itay Hubara, Matthieu Courbariaux, Daniel Soudry, Ran ElYaniv, and Yoshua Bengio. Binarized neural networks. In NeurIPS, 2016. 1, 2, 3 +[16] Xiaoqi Jiao, Yichun Yin, Lifeng Shang, Xin Jiang, Xiao Chen, Linlin Li, Fang Wang, and Qun Liu. Tinybert: Distilling bert for natural language understanding. In EMNLP, 2020. 5 +[17] Alex Kendall, Yarin Gal, and Roberto Cipolla. Multi-task learning using uncertainty to weigh losses for scene geometry and semantics. In CVPR, 2018. 1, 2 +[18] Hyungjun Kim, Kyungsu Kim, Jinseok Kim, and Jae-Joon Kim. Binary duo: Reducing gradient mismatch in binary activation network by coupling binary activations. In ICLR, + +2020. 2 +[19] Hyungjun Kim, Jihoon Park, Changhun Lee, and Jae-Joon Kim. Improving accuracy of binary neural networks using unbalanced activation distribution. In CVPR, 2021. 2 +[20] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In ICLR, 2014. 7 +[21] Iasonas Kokkinos. Ethernet: Training a universal convolutional neural network for low-, mid-, and high-level vision using diverse datasets and limited memory. In CVPR, 2017. 2 +[22] Simon Kornblith, Mohammad Norouzi, Honglak Lee, and Geoffrey Hinton. Similarity of neural network representations revisited. In ICML, 2019. 6 +[23] Zechun Liu, Wenhan Luo, Baoyuan Wu, Xin Yang, Wei Liu, and Kwang-Ting Cheng. Bi-real net: Binarizing deep network towards real-network performance. IJCV, 2020. 2, 3, 7 +[24] Zechun Liu, Zhiqiang Shen, Shichao Li, Koen Helwegen, Dong Huang, and Kwang-Ting Cheng. How do adam and training strategies help bnns optimization. In ICML, 2021. 7 +[25] Zechun Liu, Zhiqiang Shen, Marios Savvides, and Kwang-Ting Cheng. Reactnet: Towards precise binary neural network with generalized activation functions. In ECCV, 2020. 2, 3 +[26] Yongxi Lu, Abhishek Kumar, Shuangfei Zhai, Yu Cheng, Tara Javidi, and Rogerio Feris. Fully-adaptive feature sharing in multi-task networks with applications in person attribute classification. In CVPR, 2017. 2 +[27] Kevis-Kokitsi Maninis, Ilija Radosavovic, and Iasonas Kokkinos. Attentive single-tasking of multiple tasks. In CVPR, 2019. 3, 7, 8 +[28] Ishan Misra, Abhinav Shrivastava, Abhinav Gupta, and Martial Hebert. Cross-stitch networks for multi-task learning. In CVPR, 2016. 1, 2, 7 +[29] Haotong Qin, Zhongang Cai, Mingyuan Zhang, Yifu Ding, Haiyu Zhao, Shuai Yi, Xianglong Liu, and Hao Su. Bipoint-net: Binary neural network for point clouds. In ICLR, 2021. 3 +[30] Haotong Qin, Yifu Ding, Mingyuan Zhang, YAN Qinghua, Aishan Liu, Qingqing Dang, Ziwei Liu, and Xianglong Liu. Bibert: Accurate fully binarized bert. In ICLR, 2022. 2, 3, 5 +[31] Haotong Qin, Ruihao Gong, Xianglong Liu, Xiao Bai, Jingkuan Song, and Nicu Sebe. Binary neural networks: A survey. PR, 2020. 3 +[32] Haotong Qin, Ruihao Gong, Xianglong Liu, Mingzhu Shen, Ziran Wei, Fengwei Yu, and Jingkuan Song. Forward and backward information retention for accurate binary neural networks. In CVPR, 2020. 2, 3 +[33] Maithra Raghu, Thomas Unterthiner, Simon Kornblith, Chiyuan Zhang, and Alexey Dosovitskiy. Do vision transformers see like convolutional neural networks? In NeurIPS, 2021. 6 +[34] Lothar Reichel and Qiang Ye. Simple square smoothing regularization operators. Electronic Transactions on Numerical Analysis, 33:63, 2009. 2 +[35] Ramprasaath R Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. Grad-cam: Visual explanations from deep networks via gradient-based localization. In ICCV, 2017. 6 +[36] Ozan Sener and Vladlen Koltun. Multi-task learning as + +multi-objective optimization. In NeurIPS, 2018. 2 +[37] Yuzhang Shang, Bingxin Xu, Gaowen Liu, Ramana Rao Kompella, and Yan Yan. Causal-dfq: Causality guided data-free network quantization. In ICCV, 2023. 1 +[38] Yuzhang Shang, Dan Xu, Bin Duan, Ziliang Zong, Liqiang Nie, and Yan Yan. Lipschitz continuity retained binary neural network. In ECCV, 2022. 2 +[39] Yuzhang Shang, Dan Xu, Ziliang Zong, and Yan Yan. Network binarization via contrastive learning. In ECCV, 2022. 2 +[40] Yuzhang Shang, Zhihang Yuan, Bin Xie, Bingzhe Wu, and Yan Yan. Post-training quantization on diffusion models. In CVPR, pages 1972-1981, 2023. 1 +[41] Nathan Silberman, Derek Hoiem, Pushmeet Kohli, and Rob Fergus. Indoor segmentation and support inference from rgbd images. In ECCV, 2012. 7 +[42] Nitish Srivastava, Geoffrey Hinton, Alex Krizhevsky, Ilya Sutskever, and Ruslan Salakhutdinov. Dropout: a simple way to prevent neural networks from overfitting. JMLR, 2014. 2 +[43] Naftali Tishby, Fernando C Pereira, and William Bialek. The information bottleneck method. arXiv preprint physics/0004057, 2000. 5 +[44] Laurens Van der Maaten and Geoffrey Hinton. Visualizing data using t-sne. JMLR, 2008. 6 +[45] Simon Vandenhende, Stamatios Georgoulis, Bert De Brabandere, and Luc Van Gool. Branched multi-task networks: deciding what layers to share. In BMVC, 2020. 2 +[46] S. Vandenhende, S. Georgoulis, W. Van Gansbeke, M. Proesmans, D. Dai, and L. Van Gool. Multi-task learning for dense prediction tasks: A survey. TPAMI, 2021. 1 +[47] Simon Vandenhende, Stamatios Georgoulis, and Luc Van Gool. Mti-net: Multi-scale task interaction networks for multi-task learning. In CVPR, 2020. 1, 2, 3, 4, 5, 7, 8 +[48] Ziwei Wang, Ziyi Wu, Jiwen Lu, and Jie Zhou. Bidet: An efficient binarized object detector. In CVPR, 2020. 5 +[49] Dan Xu, Wanli Ouyang, Xiaogang Wang, and Nicu Sebe. Pad-net: Multi-tasks guided prediction-and-distillation network for simultaneous depth estimation and scene parsing. In CVPR, 2018, 1, 2, 3, 4, 7, 8 +[50] Hanrong Ye and Dan Xu. Inverted pyramid multi-task transformer for dense scene understanding. arXiv preprint arXiv:2203.07997, 2022. 1, 2, 4, 7, 8 +[51] Shang Yuzhang, Duan Bin, Zong Ziliang, Nie Liqiang, and Yan Yan. Lipschitz continuity guided knowledge distillation. In ICCV, 2021. 5 +[52] Hongyi Zhang, Moustapha Cisse, Yann N Dauphin, and David Lopez-Paz. mixup: Beyond empirical risk minimization. In ICLR, 2018. 2 +[53] Zhenyu Zhang, Zhen Cui, Chunyan Xu, Yan Yan, Nicu Sebe, and Jian Yang. Pattern-affinitive propagation across depth, surface normal and semantic segmentation. In CVPR, 2019. 2, 7 +[54] Ling Zhou, Zhen Cui, Chunyan Xu, Zhenyu Zhang, Chaoqun Wang, Tong Zhang, and Jian Yang. Pattern-structure diffusion for multi-task learning. In CVPR, 2020. 7 +[55] Barret Zoph and Quoc V Le. Neural architecture search with reinforcement learning. arXiv preprint arXiv:1611.01578, 2016.2 \ No newline at end of file diff --git a/2024/Efficient Multitask Dense Predictor via Binarization/images.zip b/2024/Efficient Multitask Dense Predictor via Binarization/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..af5aede2ef2ece371fdfc78ed21cc2c8495ef2d6 --- /dev/null +++ b/2024/Efficient Multitask Dense Predictor via Binarization/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c1ae58aec389f3e183f0bd36363def302e808b0ed66a8e6b754416c0ef7cc2d5 +size 608730 diff --git a/2024/Efficient Multitask Dense Predictor via Binarization/layout.json b/2024/Efficient Multitask Dense Predictor via Binarization/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..3f246506cb98146d4bcf9c097e3874648b6ed4b0 --- /dev/null +++ b/2024/Efficient Multitask Dense Predictor via Binarization/layout.json @@ -0,0 +1,8049 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 135, + 103, + 459, + 119 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 103, + 459, + 119 + ], + "spans": [ + { + "bbox": [ + 135, + 103, + 459, + 119 + ], + "type": "text", + "content": "Efficient Multitask Dense Predictor via Binarization" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 97, + 142, + 499, + 172 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 97, + 142, + 499, + 172 + ], + "spans": [ + { + "bbox": [ + 97, + 142, + 499, + 172 + ], + "type": "text", + "content": "Yuzhang Shang" + }, + { + "bbox": [ + 97, + 142, + 499, + 172 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 97, + 142, + 499, + 172 + ], + "type": "text", + "content": ", Dan Xu" + }, + { + "bbox": [ + 97, + 142, + 499, + 172 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 97, + 142, + 499, + 172 + ], + "type": "text", + "content": ", Gaowen Liu" + }, + { + "bbox": [ + 97, + 142, + 499, + 172 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 97, + 142, + 499, + 172 + ], + "type": "text", + "content": ", Ramana Rao Kompella" + }, + { + "bbox": [ + 97, + 142, + 499, + 172 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 97, + 142, + 499, + 172 + ], + "type": "text", + "content": ", Yan Yan" + }, + { + "bbox": [ + 97, + 142, + 499, + 172 + ], + "type": "inline_equation", + "content": "^{1,\\dagger}" + }, + { + "bbox": [ + 97, + 142, + 499, + 172 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 97, + 142, + 499, + 172 + ], + "type": "text", + "content": "Illinois Institute of Technology, " + }, + { + "bbox": [ + 97, + 142, + 499, + 172 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 97, + 142, + 499, + 172 + ], + "type": "text", + "content": "HKUST, " + }, + { + "bbox": [ + 97, + 142, + 499, + 172 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 97, + 142, + 499, + 172 + ], + "type": "text", + "content": "Cisco Research" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 69, + 174, + 298, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 174, + 298, + 186 + ], + "spans": [ + { + "bbox": [ + 69, + 174, + 298, + 186 + ], + "type": "text", + "content": "yzshawn@outlook.com, danxu@cse.ust.hk, {ga" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "spans": [ + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 239, + 290, + 597 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 239, + 290, + 597 + ], + "spans": [ + { + "bbox": [ + 46, + 239, + 290, + 597 + ], + "type": "text", + "content": "Multi-task learning for dense prediction has emerged as a pivotal area in computer vision, enabling simultaneous processing of diverse yet interrelated pixel-wise prediction tasks. However, the substantial computational demands of state-of-the-art (SoTA) models often limit their widespread deployment. This paper addresses this challenge by introducing network binarization to compress resource-intensive multi-task dense predictors. Specifically, our goal is to significantly accelerate multi-task dense prediction models via Binary Neural Networks (BNNs) while maintaining and even improving model performance at the same time. To reach this goal, we propose a Binary Multi-task Dense Predictor, Bi-MTPD, and several variants of Bi-MTPD, in which a multi-task dense predictor is constructed via specified binarized modules. Our systematic analysis of this predictor reveals that performance drop from binarization is primarily caused by severe information degradation. To address this issue, we introduce a deep information bottleneck layer that enforces representations for downstream tasks satisfying Gaussian distribution in forward propagation. Moreover, we introduce a knowledge distillation mechanism to correct the direction of information flow in backward propagation. Intriguingly, one variant of Bi-MTPD outperforms full-precision (FP) multi-task dense prediction SoTAs, ARTC [2] (CNN-based) and InvPT [50] (ViT-Based). This result indicates that Bi-MTPD is not merely a naive trade-off between performance and efficiency, but is rather a benefit of the redundant information flow thanks to the multi-task architecture. Code is available at BiMTDP." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 623, + 128, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 623, + 128, + 635 + ], + "spans": [ + { + "bbox": [ + 47, + 623, + 128, + 635 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 643, + 287, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 643, + 287, + 693 + ], + "spans": [ + { + "bbox": [ + 46, + 643, + 287, + 693 + ], + "type": "text", + "content": "There is a growing trend in the computer vision community where dense prediction tasks are processed in a multi-task learning manner, such as semantic segmentation, monocular depth estimation, and human parsing [28, 46, 47, 49]." + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 309, + 213, + 434, + 312 + ], + "blocks": [ + { + "bbox": [ + 309, + 213, + 434, + 312 + ], + "lines": [ + { + "bbox": [ + 309, + 213, + 434, + 312 + ], + "spans": [ + { + "bbox": [ + 309, + 213, + 434, + 312 + ], + "type": "image", + "image_path": "398fc688ef3c7dda29b672f931df9b0b9dfd838f6876b178b5f98f00a1f4f3c8.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 317, + 547, + 429 + ], + "lines": [ + { + "bbox": [ + 304, + 317, + 547, + 429 + ], + "spans": [ + { + "bbox": [ + 304, + 317, + 547, + 429 + ], + "type": "text", + "content": "Figure 1. (Left) A conceptual illustration of binary dense predictions in a multi-task manner. In contrast to approaching a series of relevant tasks individually, the multitask model benefits from information supplementation among different tasks via cross-talk structures, but the cumbersome cross-talk modules also add additional computational burden. (Right) Performance summary on NYUD-v2. X-axis and Y-axis denote the performance on depth estimation (lower is better) and segmentation (higher is better), respectively. Size of dots denotes FLOPs. ATRC [2] and InvPT [50] are previous CNN-based and ViT-based SoTAs, respectively." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 437, + 213, + 545, + 315 + ], + "blocks": [ + { + "bbox": [ + 437, + 213, + 545, + 315 + ], + "lines": [ + { + "bbox": [ + 437, + 213, + 545, + 315 + ], + "spans": [ + { + "bbox": [ + 437, + 213, + 545, + 315 + ], + "type": "image", + "image_path": "48aa45f0018c86e6c50f5d5054f6e6cd1b9a034290ae3bc61e4395b1df103f7c.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 437, + 547, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 437, + 547, + 581 + ], + "spans": [ + { + "bbox": [ + 304, + 437, + 547, + 581 + ], + "type": "text", + "content": "Benefited from the information supplementation mechanism via cross-talk structures in the multi-task models, the overall performance for the series of dense prediction tasks has been greatly improved [17] (see Fig. 1). However, the computational demands of State-of-the-Art (SoTA) multitask dense prediction models, which process multiple complex pixel-wise tasks concurrently, are substantial. This high computational requirement limits their application in resource-constrained environments like autonomous driving, robotics, and virtual reality. Our goal is to optimize these heavy SoTA models for edge devices, balancing speed with performance." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 582, + 548, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 582, + 548, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 582, + 548, + 712 + ], + "type": "text", + "content": "Several strategies for neural network compression have been explored, including pruning [12], network quantization [15, 37, 40] and knowledge distillation [14]. Notably, network binarization, a form of quantization, minimizes weights and activations to " + }, + { + "bbox": [ + 304, + 582, + 548, + 712 + ], + "type": "inline_equation", + "content": "\\pm 1" + }, + { + "bbox": [ + 304, + 582, + 548, + 712 + ], + "type": "text", + "content": ", enabling the replacement of computationally intensive inner-product operations in full-precision networks with more efficient xnor-bitcount operations in Binary Neural Networks (BNNs) [15]. Binarization theoretically reduces memory costs by " + }, + { + "bbox": [ + 304, + 582, + 548, + 712 + ], + "type": "inline_equation", + "content": "32\\times" + }, + { + "bbox": [ + 304, + 582, + 548, + 712 + ], + "type": "text", + "content": " and increases inference speed by " + }, + { + "bbox": [ + 304, + 582, + 548, + 712 + ], + "type": "inline_equation", + "content": "64\\times" + }, + { + "bbox": [ + 304, + 582, + 548, + 712 + ], + "type": "text", + "content": ", making BNNs suitable for edge-device." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 59, + 702, + 135, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 702, + 135, + 712 + ], + "spans": [ + { + "bbox": [ + 59, + 702, + 135, + 712 + ], + "type": "text", + "content": "†Corresponding author" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "15899" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 289, + 252 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 289, + 252 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 289, + 252 + ], + "type": "text", + "content": "While BNNs have shown impressive results in image classification, achieving nearly full-precision ResNet-level accuracy [25], their application has largely been limited to small-scale models, overlooking other computationally intensive computer vision tasks [15, 23, 25, 32, 39]. Extending BNNs to larger models should be the next step. However, this expansion has been hindered by issues such as overfitting [19, 38, 39] and information degradation [30]. Techniques effective in full-precision models, like label smoothing [34], dropout [42], and mixup [52] have less effect on BNNs [18, 38, 39]. Furthermore, SoTA multitask dense prediction tasks often require deep and complex models, equipped with multi-modality fusion structures [2, 47, 50], exacerbating the challenges in implementing binarization effectively." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 255, + 289, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 255, + 289, + 483 + ], + "spans": [ + { + "bbox": [ + 47, + 255, + 289, + 483 + ], + "type": "text", + "content": "The primary barrier to applying binarization in multitask dense prediction tasks is the significant degradation of information flow in deep models [15, 30, 32], leading to reduced performance. To address this issue, we first propose a Binary Multitask Dense Predictor (Bi-MTPD) baseline, where a multi-task dense predictor is formulated via binarized modules. Based on a thorough review of this baseline, we conclude that the binarization operation destroys the information flows in multi-task models, and thus representations for downstream tasks are not informative compared with their full-precision counterparts. To tackle this problem, we update Bi-MTPD with additional information flow calibration mechanisms in two directions. First, we implement variational information bottleneck enforcing the embeddings to follow Gaussian distribution with sparsity in forward propagation, in order to filter out the task-irrelevant factors. Second, we leverage the existing FP models via feature-based knowledge distillation to calibrate the gradient of the binary network in backward pass." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 486, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 486, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 47, + 486, + 289, + 715 + ], + "type": "text", + "content": "The benefits of Bi-MTPD can be analyzed from two orthogonal perspectives. On one hand, from the perspective of network binarization, the accomplishment of bridging binarization with the multitask dense prediction framework testifies that Bi-MTPD can effectively supplement information, and consequently improve the performance of the individual binary models. On the other hand, from the perspective of multitask dense prediction task, accelerating those cumbersome models is profitable to design more effective and efficient cross-talk modules in it, as shown in Fig. 1. Since existing dense prediction models have severe limitation in modelling the cross-talk modules due to their heavy utilization of convolution operation, it is critical for the multitask dense predictions to learn interactions and inference covering various scopes of the multitask context via the cross-talk mechanism [2, 28, 47, 49, 50, 53]. Intriguingly, a variant of Bi-MTPD outperforms SoTA approach ATRC [2] by " + }, + { + "bbox": [ + 47, + 486, + 289, + 715 + ], + "type": "inline_equation", + "content": "4\\%" + }, + { + "bbox": [ + 47, + 486, + 289, + 715 + ], + "type": "text", + "content": " over the segmentation task while remaining " + }, + { + "bbox": [ + 47, + 486, + 289, + 715 + ], + "type": "inline_equation", + "content": "43\\%" + }, + { + "bbox": [ + 47, + 486, + 289, + 715 + ], + "type": "text", + "content": " faster in speed, implying that our proposed method" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 304, + 72, + 547, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 133 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 133 + ], + "type": "text", + "content": "is not a naive trade-off between performance and efficiency. By empirically investigating this \"free lunch\" achievement, we conclude that the win-win outcome is benefited from our designed information supplementation mechanism which strengthens the representation ability of the binary model." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 306, + 143, + 392, + 156 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 143, + 392, + 156 + ], + "spans": [ + { + "bbox": [ + 306, + 143, + 392, + 156 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 163, + 547, + 474 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 163, + 547, + 474 + ], + "spans": [ + { + "bbox": [ + 304, + 163, + 547, + 474 + ], + "type": "text", + "content": "Multitask Dense Prediction. Multi-Task Learning (MTL) methods can be generally categorized into two main paradigms in terms of the way where model learns shared representations: hard and soft parameter sharing. Hard parameter sharing characterizes architectures which typically share the first hidden representations among the tasks while branching to independent task-specific representations at a later stage. Most approaches split to task-specific heads at a single branch point [5, 17, 21, 36]. However, such naive branching can be sub-optimal, raising interest in mechanisms that allow for finely branched architectures [26, 45]. As a result, in soft parameter sharing, each task is assigned its own set of parameters and a feature sharing mechanism realize the cross-talk as demonstrated in Fig 1. The following works devise the cross-talk mechanisms focusing on the locations in the network where information or features are exchanged or shared between tasks. Apart from the locations, the feature sharing modules are also widely studied. For example, feature fusing can be introduced along the entire network depth [9, 28]; PAD-Net [49] uses multi-modal distillation to enhance task-specific predictions, in which information flow from each source to target task is regulated with a sigmoid-activated gate function; and MTI-Net [47] combines the multi-modal distillation module of PAD-Net with a multi-scale refinement scheme to facilitate cross-task talk at multiple scales." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 474, + 547, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 474, + 547, + 582 + ], + "spans": [ + { + "bbox": [ + 304, + 474, + 547, + 582 + ], + "type": "text", + "content": "Although increasing the number of cross-talk modules intuitively benefits the overall performance of the models, computational cost is often an obstacle. To handle this issue, ATRC [2] introduces NAS [55] to automatically design an efficient information fusing modules. From the perspective of the efficient representation cross-talk, our proposed models with the binarization module can be interpreted as a new pathway to feature fusing within a notably lower inference speed level." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 582, + 548, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 582, + 548, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 582, + 548, + 715 + ], + "type": "text", + "content": "Neural Network Binarization. As pioneers, [15] use the sign function to quantize weights and activations to " + }, + { + "bbox": [ + 304, + 582, + 548, + 715 + ], + "type": "inline_equation", + "content": "\\pm 1" + }, + { + "bbox": [ + 304, + 582, + 548, + 715 + ], + "type": "text", + "content": ", initiating the trends of studies of network binarization. To tackle the vanishing gradient issue induced by the bina- rization operations, the straight-through estimator (STE) [1] is introduced for the gradient approximation. Rooted in this archetype, considerable studies contribute to improving the performance of BNNs, particularly on ImageNet. For example, [23] propose Bi-Real introducing double residual connections with FP downsampling layers to mitigate the excessive gradient vanishing issue caused by binariza" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "15900" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 286, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 286, + 203 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 286, + 203 + ], + "type": "text", + "content": "tion, and consequently demonstrate that delicately ing additional connections within BNNs benefits t dient propagation. [13] design a proxy matrix as of the latent parameter space to guide the alignmen weights with different bits by recovering the smooth BNNs. In summary, a large number of methods h tended the boundary of the network binarization w curacy over classification (e.g., ReActNet [25] within comparable FLOPs of binary ResNet-18 achieves " + }, + { + "bbox": [ + 46, + 72, + 286, + 203 + ], + "type": "inline_equation", + "content": "65.9\\%" + }, + { + "bbox": [ + 46, + 72, + 286, + 203 + ], + "type": "text", + "content": " Top-1 accuracy on ImageNet, while full-precision version is only " + }, + { + "bbox": [ + 46, + 72, + 286, + 203 + ], + "type": "inline_equation", + "content": "70.5\\%" + }, + { + "bbox": [ + 46, + 72, + 286, + 203 + ], + "type": "text", + "content": ")." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 205, + 287, + 324 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 205, + 287, + 324 + ], + "spans": [ + { + "bbox": [ + 46, + 205, + 287, + 324 + ], + "type": "text", + "content": "However, most of those works validate their effectiveness over classification with relatively small architectures (mostly ResNet18 and ResNet34). Meanwhile, the network-based models for dense prediction tasks are bigger and deeper than those toy models, as the information flow in networks is severely degraded. Consequently, directly implementing existing binarization methods can not achieve supposed success. To mitigate the degradation, we propose to binarize those dense prediction models in a multitask way." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 338, + 227, + 350 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 338, + 227, + 350 + ], + "spans": [ + { + "bbox": [ + 47, + 338, + 227, + 350 + ], + "type": "text", + "content": "3. Multitask Network Binarization" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 357, + 180, + 370 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 357, + 180, + 370 + ], + "spans": [ + { + "bbox": [ + 47, + 357, + 180, + 370 + ], + "type": "text", + "content": "3.1. Binary Neural Network" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 377, + 286, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 377, + 286, + 509 + ], + "spans": [ + { + "bbox": [ + 46, + 377, + 286, + 509 + ], + "type": "text", + "content": "To begin with, we briefly review the general idea of binary neural networks (BNNs) in [7, 15]. Here, we only elaborate the speedup mechanism and the degradation of information flow of the binarization. We define a full-precision (FP) neural network with " + }, + { + "bbox": [ + 46, + 377, + 286, + 509 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 46, + 377, + 286, + 509 + ], + "type": "text", + "content": " layers, " + }, + { + "bbox": [ + 46, + 377, + 286, + 509 + ], + "type": "inline_equation", + "content": "f(\\mathbf{x}) = (\\mathbf{W}^{K}\\times \\sigma \\times \\mathbf{W}^{K - 1}\\dots \\sigma \\times \\mathbf{W}^{1})(\\mathbf{x})" + }, + { + "bbox": [ + 46, + 377, + 286, + 509 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 377, + 286, + 509 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 46, + 377, + 286, + 509 + ], + "type": "text", + "content": " is the input sample and " + }, + { + "bbox": [ + 46, + 377, + 286, + 509 + ], + "type": "inline_equation", + "content": "\\mathbf{W}^k:\\mathbb{R}^{d_{k - 1}}\\longmapsto \\mathbb{R}^{d_k}(k = 1,\\ldots ,K)" + }, + { + "bbox": [ + 46, + 377, + 286, + 509 + ], + "type": "text", + "content": " stands for the weight matrix connecting the " + }, + { + "bbox": [ + 46, + 377, + 286, + 509 + ], + "type": "inline_equation", + "content": "(k - 1)" + }, + { + "bbox": [ + 46, + 377, + 286, + 509 + ], + "type": "text", + "content": "-th and the " + }, + { + "bbox": [ + 46, + 377, + 286, + 509 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 377, + 286, + 509 + ], + "type": "text", + "content": "-th layer, with " + }, + { + "bbox": [ + 46, + 377, + 286, + 509 + ], + "type": "inline_equation", + "content": "d_{k - 1}" + }, + { + "bbox": [ + 46, + 377, + 286, + 509 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 377, + 286, + 509 + ], + "type": "inline_equation", + "content": "d_{k}" + }, + { + "bbox": [ + 46, + 377, + 286, + 509 + ], + "type": "text", + "content": " representing the sizes of the input and output of the " + }, + { + "bbox": [ + 46, + 377, + 286, + 509 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 377, + 286, + 509 + ], + "type": "text", + "content": "-th network layer, respectively. The " + }, + { + "bbox": [ + 46, + 377, + 286, + 509 + ], + "type": "inline_equation", + "content": "\\sigma (\\cdot)" + }, + { + "bbox": [ + 46, + 377, + 286, + 509 + ], + "type": "text", + "content": " function performs element-wise activation for the feature maps." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 510, + 286, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 510, + 286, + 712 + ], + "spans": [ + { + "bbox": [ + 46, + 510, + 286, + 712 + ], + "type": "text", + "content": "BNNs vary from FP neural networks in terms of the forward operation and the backward gradient approximation. Specifically, in the forward propagation, the BNN maintains FP latent weights " + }, + { + "bbox": [ + 46, + 510, + 286, + 712 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_F" + }, + { + "bbox": [ + 46, + 510, + 286, + 712 + ], + "type": "text", + "content": " for gradient updates, and the " + }, + { + "bbox": [ + 46, + 510, + 286, + 712 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 510, + 286, + 712 + ], + "type": "text", + "content": "-th weight matrix " + }, + { + "bbox": [ + 46, + 510, + 286, + 712 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_F^k" + }, + { + "bbox": [ + 46, + 510, + 286, + 712 + ], + "type": "text", + "content": " is binarized into " + }, + { + "bbox": [ + 46, + 510, + 286, + 712 + ], + "type": "inline_equation", + "content": "\\pm 1" + }, + { + "bbox": [ + 46, + 510, + 286, + 712 + ], + "type": "text", + "content": ", obtaining the binary weight matrix " + }, + { + "bbox": [ + 46, + 510, + 286, + 712 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_B^k" + }, + { + "bbox": [ + 46, + 510, + 286, + 712 + ], + "type": "text", + "content": " via the binarize function " + }, + { + "bbox": [ + 46, + 510, + 286, + 712 + ], + "type": "inline_equation", + "content": "\\text{sign}(\\cdot)" + }, + { + "bbox": [ + 46, + 510, + 286, + 712 + ], + "type": "text", + "content": ", i.e. " + }, + { + "bbox": [ + 46, + 510, + 286, + 712 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_B^k = \\text{sign}(\\mathbf{W}_F^k)" + }, + { + "bbox": [ + 46, + 510, + 286, + 712 + ], + "type": "text", + "content": ". Then the intermediate activation map (full-precision) of the " + }, + { + "bbox": [ + 46, + 510, + 286, + 712 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 510, + 286, + 712 + ], + "type": "text", + "content": "-th layer is produced by " + }, + { + "bbox": [ + 46, + 510, + 286, + 712 + ], + "type": "inline_equation", + "content": "\\mathbf{A}_F^k = \\mathbf{W}_B^k\\mathbf{A}_B^{k-1}" + }, + { + "bbox": [ + 46, + 510, + 286, + 712 + ], + "type": "text", + "content": ". The same quantization method is used to binarize the full-precision activation map as " + }, + { + "bbox": [ + 46, + 510, + 286, + 712 + ], + "type": "inline_equation", + "content": "\\mathbf{A}_B^k = \\text{sign}(\\mathbf{A}_F^k)" + }, + { + "bbox": [ + 46, + 510, + 286, + 712 + ], + "type": "text", + "content": ", and the whole forward pass of binarization is performed by iterating this process for " + }, + { + "bbox": [ + 46, + 510, + 286, + 712 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 46, + 510, + 286, + 712 + ], + "type": "text", + "content": " times, as shown in Fig. 2. For BNNs, the weights and activations are 1-bit, by which the network is accelerated 32 times in terms of memory cost. Importantly, inference of BNN is accelerated 64 times, as the FP multiplication in FP networks is replaced by Xnor-Bitcount in BNNs." + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 310, + 72, + 541, + 97 + ], + "blocks": [ + { + "bbox": [ + 310, + 72, + 541, + 97 + ], + "lines": [ + { + "bbox": [ + 310, + 72, + 541, + 97 + ], + "spans": [ + { + "bbox": [ + 310, + 72, + 541, + 97 + ], + "type": "image", + "image_path": "115f534f2a2a4159b6b9673aa0cf3cb76dcc96ce61aea7c3c47ae62c5b39f535.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 102, + 545, + 125 + ], + "lines": [ + { + "bbox": [ + 305, + 102, + 545, + 125 + ], + "spans": [ + { + "bbox": [ + 305, + 102, + 545, + 125 + ], + "type": "text", + "content": "Figure 2. A general illustration of the forward propagation of the " + }, + { + "bbox": [ + 305, + 102, + 545, + 125 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 305, + 102, + 545, + 125 + ], + "type": "text", + "content": "-th layer in the BNN." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 139, + 545, + 222 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 139, + 545, + 222 + ], + "spans": [ + { + "bbox": [ + 304, + 139, + 545, + 222 + ], + "type": "text", + "content": "In the backward propagation, the main challenge is that the pervasive sign functions are theoretically nondifferentiable, and thus extremely destroy the information flow via gradient the propagation. To address this problem, a large number of researchers [31] widely exploit the straight-through estimator (STE) [1] to numerically approximate the derivative of the whole BNN, i.e." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 335, + 230, + 545, + 258 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 230, + 545, + 258 + ], + "spans": [ + { + "bbox": [ + 335, + 230, + 545, + 258 + ], + "type": "interline_equation", + "content": "\\text {B a c k w a r d :} \\quad \\frac {\\partial \\mathcal {L}}{\\partial x} = \\left\\{ \\begin{array}{l l} \\frac {\\partial \\mathcal {L}}{\\partial s g n (x)} & | x | \\leq 1 \\\\ 0 & | x | > 1. \\end{array} \\right. \\tag {1}", + "image_path": "2ccc52e278d8a351b67a51461befaedffea5a2332710f5d909eb22983eb44d1e.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 267, + 545, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 267, + 545, + 327 + ], + "spans": [ + { + "bbox": [ + 304, + 267, + 545, + 327 + ], + "type": "text", + "content": "It is worth noting that we do not implement the aforementioned vanilla approximation method in practice, while we utilize the prevalent Bi-Real [23] and IR-Net [32] to gradually approximate the sign function, which have been proven to be better estimation approaches [29, 30]." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 327, + 545, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 327, + 545, + 399 + ], + "spans": [ + { + "bbox": [ + 304, + 327, + 545, + 399 + ], + "type": "text", + "content": "Even though numerous methods have been proposed to eliminate the deterioration of the information flow induced by the binarization, the deterioration is still inevitable due to the severe accuracy loss of weights, activations and gradients [29, 30]. Consequently, binarization destroys the performance of the complicated computer vision models." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 407, + 451, + 418 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 407, + 451, + 418 + ], + "spans": [ + { + "bbox": [ + 306, + 407, + 451, + 418 + ], + "type": "text", + "content": "3.2. Multitask Dense Predictor" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 426, + 545, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 426, + 545, + 557 + ], + "spans": [ + { + "bbox": [ + 304, + 426, + 545, + 557 + ], + "type": "text", + "content": "After deploying binarization techniques in the models for dense prediction tasks, the performance of the binarized models is unacceptable, as shown in Fig. 1 and the binary single result in Table. 3. Since the architectures of those SoTA dense prediction models are relatively heavier and deeper (e.g., HRNet-48 or ResNet-101 with a task-specific head [27]) than the ones for classification (e.g., ResNet-18 with a fully-connected layer as the classification head). Moreover, the information passing in the binary models via back-propagation, especially in deep models, is notoriously inefficient [23]." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 558, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 558, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 558, + 545, + 713 + ], + "type": "text", + "content": "Dense prediction tasks can mutually supplement information, e.g., surface normals and depth can directly be derived from each other, which can be modeled as the regularization of each other [47]. The relevancy among dense prediction tasks is worth being utilized to improve the overall performance of models. For example, before the deep learning era, pioneering work [11] utilizes RGB-D images with depth information to predict scene semantics to improve the quality of the prediction. In the deep learning era, recent attention-based multitask learning methods [2, 47, 49] explicitly and implicitly distill information from other tasks as a complementary signal to improve the targeted task performance. Briefly, the above-mentioned methods are achieved" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15901" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 73, + 71, + 292, + 205 + ], + "blocks": [ + { + "bbox": [ + 73, + 71, + 292, + 205 + ], + "lines": [ + { + "bbox": [ + 73, + 71, + 292, + 205 + ], + "spans": [ + { + "bbox": [ + 73, + 71, + 292, + 205 + ], + "type": "image", + "image_path": "c46068f6572247f241dda6efc770bb397ad2576d027021aec506d6cb7d7123e7.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 209, + 547, + 255 + ], + "lines": [ + { + "bbox": [ + 46, + 209, + 547, + 255 + ], + "spans": [ + { + "bbox": [ + 46, + 209, + 547, + 255 + ], + "type": "text", + "content": "Figure 3. (Left) The illustration of the baseline multitask framework. (Middle) The designed MMD modules for binary representations. Importantly, the MMD module can pass information among different predictions, acting as a cross-talk mechanism. (Right) As all fundamental modules in Bi-MTPD baseline are binarized, inferences can be performed by complete Bool operations, which are very computationally cheap." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 293, + 72, + 431, + 204 + ], + "blocks": [ + { + "bbox": [ + 293, + 72, + 431, + 204 + ], + "lines": [ + { + "bbox": [ + 293, + 72, + 431, + 204 + ], + "spans": [ + { + "bbox": [ + 293, + 72, + 431, + 204 + ], + "type": "image", + "image_path": "ec62c60358444a2011064bc45b84c0d67f5441695e4a243e3db3214a07ab84f0.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 433, + 72, + 519, + 204 + ], + "blocks": [ + { + "bbox": [ + 433, + 72, + 519, + 204 + ], + "lines": [ + { + "bbox": [ + 433, + 72, + 519, + 204 + ], + "spans": [ + { + "bbox": [ + 433, + 72, + 519, + 204 + ], + "type": "image", + "image_path": "be1a3fe3fe8e77b12fda57e37b951d594ec50bea252f9243058c41c7408beb04.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 267, + 288, + 304 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 267, + 288, + 304 + ], + "spans": [ + { + "bbox": [ + 46, + 267, + 288, + 304 + ], + "type": "text", + "content": "by combining an existing backbone network for initial task predictions with a subsequent decoding process, as shown in Fig. 3 (Left)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 316, + 289, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 316, + 289, + 545 + ], + "spans": [ + { + "bbox": [ + 46, + 316, + 289, + 545 + ], + "type": "text", + "content": "Specifically, the shared features extracted by the backbone network are then processed by a set of task-specific heads, which produce a series of initial predictions for " + }, + { + "bbox": [ + 46, + 316, + 289, + 545 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 46, + 316, + 289, + 545 + ], + "type": "text", + "content": " tasks, i.e. " + }, + { + "bbox": [ + 46, + 316, + 289, + 545 + ], + "type": "inline_equation", + "content": "\\{Y_i^k\\} (k = 1,\\dots ,T)" + }, + { + "bbox": [ + 46, + 316, + 289, + 545 + ], + "type": "text", + "content": " (the backbone and the task-specific heads are referred as the front-end of the network [47]). Transforming and binarizing " + }, + { + "bbox": [ + 46, + 316, + 289, + 545 + ], + "type": "inline_equation", + "content": "Y_{i}^{t}" + }, + { + "bbox": [ + 46, + 316, + 289, + 545 + ], + "type": "text", + "content": " into the form of a 1-bit feature map, we obtain a set of corresponding binary feature maps of the scene, i.e. " + }, + { + "bbox": [ + 46, + 316, + 289, + 545 + ], + "type": "inline_equation", + "content": "\\{F_{B,i}^{t}\\} (t = 1,\\dots ,T)" + }, + { + "bbox": [ + 46, + 316, + 289, + 545 + ], + "type": "text", + "content": " which are more task-aware than the shared binary features of the backbone network. The information from these task-specific feature representations is then fused via a multimodal distillation via binarized attention mechanism before making the final task predictions. As previous work featured, our method is also task-saleable. Especially, it is possible that some tasks are only predicted in the front-end of the network (initial prediction). The initial tasks are also called auxiliary tasks since they serve as proxies in order to improve the performance of the final tasks, as shown in Fig. 3." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 557, + 289, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 557, + 289, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 557, + 289, + 713 + ], + "type": "text", + "content": "Multi-Modal Distillation (MMD) via Binarized Attention Mechanism. The multi-modal distillation module is the key in the multi-task dense prediction model. Specifically, we utilize the attention mechanism for guiding the information passing between the binary feature maps generated from different modalities for different tasks. Since the passed information flow is not always helpful, the attention can act as a gate function to control the flow, in other words, to make the network automatically learn to focus or to ignore information from other binary features [47, 49]. Including the binarization operations, we can formalize the MMD via binarized attention as follows. While passing information to the " + }, + { + "bbox": [ + 46, + 557, + 289, + 713 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 557, + 289, + 713 + ], + "type": "text", + "content": "-th task, we first obtain a binarized attenuation matrix " + }, + { + "bbox": [ + 46, + 557, + 289, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{A}" + }, + { + "bbox": [ + 46, + 557, + 289, + 713 + ], + "type": "text", + "content": ", and then use it to compute the attention matrix " + }, + { + "bbox": [ + 46, + 557, + 289, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{B}" + }, + { + "bbox": [ + 46, + 557, + 289, + 713 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 267, + 388, + 281 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 267, + 388, + 281 + ], + "spans": [ + { + "bbox": [ + 306, + 267, + 388, + 281 + ], + "type": "text", + "content": "tion map " + }, + { + "bbox": [ + 306, + 267, + 388, + 281 + ], + "type": "inline_equation", + "content": "\\mathbf{A}_{B,i}^{k}" + }, + { + "bbox": [ + 306, + 267, + 388, + 281 + ], + "type": "text", + "content": " , i.e.," + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 365, + 288, + 545, + 304 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 365, + 288, + 545, + 304 + ], + "spans": [ + { + "bbox": [ + 365, + 288, + 545, + 304 + ], + "type": "interline_equation", + "content": "\\mathbf {A} _ {B, i} ^ {k} \\leftarrow b o o l \\left(\\mathbf {W} _ {B} ^ {k} \\otimes \\mathbf {F} _ {B, i} ^ {k}\\right) \\tag {2}", + "image_path": "4b495289dc32e4ba0bfaa67780aa4bf138c8ec10b7c366530f910d3d10ed42db.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 311, + 546, + 360 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 311, + 546, + 360 + ], + "spans": [ + { + "bbox": [ + 305, + 311, + 546, + 360 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 311, + 546, + 360 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_B^k" + }, + { + "bbox": [ + 305, + 311, + 546, + 360 + ], + "type": "text", + "content": " is the parameters of the binarized convolution layer, " + }, + { + "bbox": [ + 305, + 311, + 546, + 360 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{B,i}^{k}" + }, + { + "bbox": [ + 305, + 311, + 546, + 360 + ], + "type": "text", + "content": " is the binary feature map of the initial prediction, and " + }, + { + "bbox": [ + 305, + 311, + 546, + 360 + ], + "type": "inline_equation", + "content": "\\otimes" + }, + { + "bbox": [ + 305, + 311, + 546, + 360 + ], + "type": "text", + "content": " denotes convolution operation. Then the information is passed with the attention map controlled as follows:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 307, + 365, + 545, + 415 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 365, + 545, + 415 + ], + "spans": [ + { + "bbox": [ + 307, + 365, + 545, + 415 + ], + "type": "interline_equation", + "content": "\\mathbf {F} _ {B, i} ^ {o, k} \\leftarrow s i g n \\left[ \\mathbf {F} _ {B, i} ^ {k} + \\sum_ {t = 1, t \\neq k} ^ {T} \\mathbf {A} _ {B, i} ^ {k} \\odot \\left(\\mathbf {W} _ {B, t} \\otimes \\mathbf {F} _ {B, i} ^ {t}\\right) \\right] \\tag {3}", + "image_path": "b0b68662dba89eba99b53645270fa483a455f7714f327fe4cc71fa9f9e4c4fd1.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 415, + 545, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 415, + 545, + 510 + ], + "spans": [ + { + "bbox": [ + 304, + 415, + 545, + 510 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 415, + 545, + 510 + ], + "type": "inline_equation", + "content": "\\odot" + }, + { + "bbox": [ + 304, + 415, + 545, + 510 + ], + "type": "text", + "content": " element-wise multiplication. The general demonstration of the distillation process is presented in Fig. 3 Left. The output binary feature map " + }, + { + "bbox": [ + 304, + 415, + 545, + 510 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{B,i}^{o,k}" + }, + { + "bbox": [ + 304, + 415, + 545, + 510 + ], + "type": "text", + "content": " is then used by the head for the corresponding " + }, + { + "bbox": [ + 304, + 415, + 545, + 510 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 415, + 545, + 510 + ], + "type": "text", + "content": "-th task in Fig. 3 Right. By using the task-specific distillation activations, the network can preserve more information for each task [2, 47, 49], which especially benefits the BNNs where the deteriorated information flow mainly induce the performance drop." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 510, + 546, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 510, + 546, + 676 + ], + "spans": [ + { + "bbox": [ + 304, + 510, + 546, + 676 + ], + "type": "text", + "content": "On the other hand, multitask dense prediction models benefit from the network binarization in terms of performance. Although these multitask models have achieved a promising performance, they are still limited by the nature of convolution-based distillation modules that are heavily used in a multi-scale way, which model critical spatial and task-related interactions in relatively local perceptive fields [50]. Theoretically, more distillation modules in different network nodes can contribute to model performance, yet we cannot unrestrictedly add distillation modules to the existing model due to the computational limitation. Fortunately, with the saved computational cost of the binary networks, we can implement additional distillation modules in our model." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 677, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 677, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 305, + 677, + 547, + 715 + ], + "type": "text", + "content": "Binary Baseline for Multitask Dense Prediction, Bi-MTPD. To obtain dense predictions with BNNs under a multitask learning framework, we practically binarize" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "15902" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 75, + 70, + 521, + 229 + ], + "blocks": [ + { + "bbox": [ + 75, + 70, + 521, + 229 + ], + "lines": [ + { + "bbox": [ + 75, + 70, + 521, + 229 + ], + "spans": [ + { + "bbox": [ + 75, + 70, + 521, + 229 + ], + "type": "image", + "image_path": "704a8c0fde1af054bc8734d52148f028127260aab29887a139d76e5ba492c01b.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 233, + 547, + 268 + ], + "lines": [ + { + "bbox": [ + 46, + 233, + 547, + 268 + ], + "spans": [ + { + "bbox": [ + 46, + 233, + 547, + 268 + ], + "type": "text", + "content": "Figure 4. The pipeline of Bi-MTPD. We introduce a VIB layer after the backbone network to filter-out the nuisance factors which may lead to model overfitting issue in the forward propagation. In addition, we deploy a feature-based knowledge distillation mechanism to guide the optimization direction in the backward propagation." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 280, + 288, + 340 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 280, + 288, + 340 + ], + "spans": [ + { + "bbox": [ + 46, + 280, + 288, + 340 + ], + "type": "text", + "content": "the MTI-Net [47] as the binary baseline. Specifically, the main modules in the full-precision MTI-Net, including backbone, heads, and multi-modal distillation module, are replaced with binary modules (both weights and activations are 1-bit). We call this baseline as Bi-MTPD." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 348, + 234, + 361 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 348, + 234, + 361 + ], + "spans": [ + { + "bbox": [ + 47, + 348, + 234, + 361 + ], + "type": "text", + "content": "3.3. Information Flow Supplementation" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 366, + 287, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 366, + 287, + 545 + ], + "spans": [ + { + "bbox": [ + 46, + 366, + 287, + 545 + ], + "type": "text", + "content": "Although we build a fully binarized baseline Bi-MTPD for multitask dense predictions and train the pipeline with common techniques, the performance is still of major concern. The baseline suffers an immense information degradation as the nuisance factors are over-fitted in the forward propagation and optimization directions severely mismatch in the backward propagation. To solve these problems, in this section, we further propose the variant of Bi-MTPD, Bi-MTPD-F. Specifically, we introduce a variational information bottleneck (VIB) layer after the output of the shared binary backbone to precisely enforces the feature extractor to preserve the minimal sufficient information of the input data. As well, we deploy the feature-based knowledge distillation to guided the optimization direction. We present more details in the following section." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 545, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 545, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 545, + 288, + 713 + ], + "type": "text", + "content": "Variational Information Bottleneck for Filter-Out Nuisance Factors. Obtaining the initial binary representations of input images by the shared backbone, we need to train a series of targeted heads to split them. A straightforward strategy is to feed these representations into the following MMD module. However, the binarized representations lack homogenization leading to model overfitting issue [48]. Therefore, the need to regularize the binarized representations, while the regularization would not to contaminate the information flow in the representations. Fortunately, the information bottleneck (IB) principle directly relates to compression with the best hypothesis that the data misfit and the model complexity should simultaneously be minimized [43, 48]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 280, + 545, + 388 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 280, + 545, + 388 + ], + "spans": [ + { + "bbox": [ + 304, + 280, + 545, + 388 + ], + "type": "text", + "content": "As the VIB could effectively capture the relevant parts and filter out the irrelevant ones from inputs, we design a novel VIB-based layer after the backbone. In particular, it explicitly enforces the feature extractor to preserve the minimal sufficient information of the input data. In other words, it can help ensure the information flow flexibly to learn clean representation for the targeted tasks. The objective function of our VIB-based classification can be formulated as a term of information loss, written as follows:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 362, + 394, + 545, + 408 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 362, + 394, + 545, + 408 + ], + "spans": [ + { + "bbox": [ + 362, + 394, + 545, + 408 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {v i b} = K L \\left[ p (\\mathbf {Z} \\mid \\mathbf {A} _ {\\mathbf {B}}), r (\\mathbf {Z}) \\right], \\tag {4}", + "image_path": "02b4b6811a7cd2bc7d4da116f217dbb02d07cf61ad35b1ed0c50b3f12c854c9f.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 414, + 545, + 508 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 414, + 545, + 508 + ], + "spans": [ + { + "bbox": [ + 304, + 414, + 545, + 508 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 414, + 545, + 508 + ], + "type": "inline_equation", + "content": "\\mathbf{A_B}" + }, + { + "bbox": [ + 304, + 414, + 545, + 508 + ], + "type": "text", + "content": " is the input binary backbone representation, " + }, + { + "bbox": [ + 304, + 414, + 545, + 508 + ], + "type": "inline_equation", + "content": "\\mathbf{B}" + }, + { + "bbox": [ + 304, + 414, + 545, + 508 + ], + "type": "text", + "content": " is the latent representation variable, " + }, + { + "bbox": [ + 304, + 414, + 545, + 508 + ], + "type": "inline_equation", + "content": "p(\\mathbf{Z} \\mid \\mathbf{A_B})" + }, + { + "bbox": [ + 304, + 414, + 545, + 508 + ], + "type": "text", + "content": " is a multivariate Gaussian distribution, and " + }, + { + "bbox": [ + 304, + 414, + 545, + 508 + ], + "type": "inline_equation", + "content": "r(\\mathbf{Z})" + }, + { + "bbox": [ + 304, + 414, + 545, + 508 + ], + "type": "text", + "content": " is a standard normal distribution. Generally, the latter is a regularization term controlling how much information of the input is filtered out. A more detailed discussion about the VIB for binarized models filtering out irrelevant information is in Supplemental Materials." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 510, + 545, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 510, + 545, + 677 + ], + "spans": [ + { + "bbox": [ + 304, + 510, + 545, + 677 + ], + "type": "text", + "content": "Feature-based Knowledge Distillation for Guiding the Direction of Information Flow. Distillation is a common and essential optimization approach to alleviate the performance drop of quantized models on ultra-low bit-width settings, which can be flexibly deployed for any architectures to utilize the knowledge of a full-precision teacher model [3, 16, 30, 51]. The usual practice is to distill the activations in a layerwise manner from the full-precision teacher to the quantized counterparts, i.e., " + }, + { + "bbox": [ + 304, + 510, + 545, + 677 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{B,l}" + }, + { + "bbox": [ + 304, + 510, + 545, + 677 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 510, + 545, + 677 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{FP,l}" + }, + { + "bbox": [ + 304, + 510, + 545, + 677 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 304, + 510, + 545, + 677 + ], + "type": "inline_equation", + "content": "l = 1,\\dots ,L" + }, + { + "bbox": [ + 304, + 510, + 545, + 677 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 510, + 545, + 677 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 304, + 510, + 545, + 677 + ], + "type": "text", + "content": " represents the number of network layers), respectively. We use the mean squared errors (MSE) as the distance function to measure the difference between corresponding features student and teacher. The knowledge distillation loss can be written as follows:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 362, + 683, + 545, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 362, + 683, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 362, + 683, + 545, + 715 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {k d} = \\sum_ {l = 1} ^ {L} M S E \\left(\\mathbf {F} _ {B, l}, \\mathbf {F} _ {F P, l}\\right). \\tag {5}", + "image_path": "541494ec2c4f3c233f767edeaa96628bf42bfadebf3bb074c29088b4f8e1c40c.jpg" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15903" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 92, + 74, + 164, + 129 + ], + "blocks": [ + { + "bbox": [ + 92, + 74, + 164, + 129 + ], + "lines": [ + { + "bbox": [ + 92, + 74, + 164, + 129 + ], + "spans": [ + { + "bbox": [ + 92, + 74, + 164, + 129 + ], + "type": "image", + "image_path": "35b40e833845c6eb5484737bc94a8f79a914bbedb6eb12b163c2469013f560ca.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 167, + 76, + 236, + 129 + ], + "blocks": [ + { + "bbox": [ + 167, + 76, + 236, + 129 + ], + "lines": [ + { + "bbox": [ + 167, + 76, + 236, + 129 + ], + "spans": [ + { + "bbox": [ + 167, + 76, + 236, + 129 + ], + "type": "image", + "image_path": "ec1b8852c1b5b6e652375cc3a6736ddf029d063f02b3b7a2df3159aace136216.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 240, + 76, + 309, + 128 + ], + "blocks": [ + { + "bbox": [ + 240, + 76, + 309, + 128 + ], + "lines": [ + { + "bbox": [ + 240, + 76, + 309, + 128 + ], + "spans": [ + { + "bbox": [ + 240, + 76, + 309, + 128 + ], + "type": "image", + "image_path": "8d37ebe5c81add2e10eea6b05065059d823414a3212c146b0e90546bde46e950.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 312, + 76, + 399, + 151 + ], + "blocks": [ + { + "bbox": [ + 312, + 76, + 399, + 151 + ], + "lines": [ + { + "bbox": [ + 312, + 76, + 399, + 151 + ], + "spans": [ + { + "bbox": [ + 312, + 76, + 399, + 151 + ], + "type": "image", + "image_path": "bdbf2de8bf8bcf343b6dbb45b7d8b7510b6314e5d771ce7e89a9f7ba659691b4.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 400, + 72, + 502, + 151 + ], + "blocks": [ + { + "bbox": [ + 400, + 72, + 502, + 151 + ], + "lines": [ + { + "bbox": [ + 400, + 72, + 502, + 151 + ], + "spans": [ + { + "bbox": [ + 400, + 72, + 502, + 151 + ], + "type": "image", + "image_path": "2556a9996ba5f9f78730d228893d9915a55b28310bb0fedcaaac4c082dc91750.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 93, + 129, + 164, + 176 + ], + "blocks": [ + { + "bbox": [ + 93, + 129, + 164, + 176 + ], + "lines": [ + { + "bbox": [ + 93, + 129, + 164, + 176 + ], + "spans": [ + { + "bbox": [ + 93, + 129, + 164, + 176 + ], + "type": "image", + "image_path": "33a22209f064b38db6842f805e4707670e679a262f438c6048c7e26505649294.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 168, + 129, + 236, + 176 + ], + "blocks": [ + { + "bbox": [ + 168, + 129, + 236, + 176 + ], + "lines": [ + { + "bbox": [ + 168, + 129, + 236, + 176 + ], + "spans": [ + { + "bbox": [ + 168, + 129, + 236, + 176 + ], + "type": "image", + "image_path": "731f2bff1ade657246b93d31a804d5958fbdf747f756e721f828d62e42f26fc3.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 240, + 152, + 309, + 176 + ], + "blocks": [ + { + "bbox": [ + 240, + 152, + 309, + 176 + ], + "lines": [ + { + "bbox": [ + 240, + 152, + 309, + 176 + ], + "spans": [ + { + "bbox": [ + 240, + 152, + 309, + 176 + ], + "type": "image", + "image_path": "34b2b0b67743437999c41f3481ad5610bbf23bbf772adcb585c78af5700a95b7.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 318, + 152, + 399, + 228 + ], + "blocks": [ + { + "bbox": [ + 318, + 152, + 399, + 228 + ], + "lines": [ + { + "bbox": [ + 318, + 152, + 399, + 228 + ], + "spans": [ + { + "bbox": [ + 318, + 152, + 399, + 228 + ], + "type": "image", + "image_path": "793d6f89f9c5f2952b2ff237f9b9d16c1a470a682e3ba1a2cdc0c2339fddb773.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 352, + 239, + 366, + 252 + ], + "lines": [ + { + "bbox": [ + 352, + 239, + 366, + 252 + ], + "spans": [ + { + "bbox": [ + 352, + 239, + 366, + 252 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 404, + 152, + 502, + 232 + ], + "blocks": [ + { + "bbox": [ + 404, + 152, + 502, + 232 + ], + "lines": [ + { + "bbox": [ + 404, + 152, + 502, + 232 + ], + "spans": [ + { + "bbox": [ + 404, + 152, + 502, + 232 + ], + "type": "image", + "image_path": "4d10f587d7454ee29ecbdb5cf637a6411fc2bfb89f4d885715672ce042247d1f.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 448, + 239, + 460, + 252 + ], + "lines": [ + { + "bbox": [ + 448, + 239, + 460, + 252 + ], + "spans": [ + { + "bbox": [ + 448, + 239, + 460, + 252 + ], + "type": "text", + "content": "(c)" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 93, + 180, + 164, + 229 + ], + "blocks": [ + { + "bbox": [ + 93, + 180, + 164, + 229 + ], + "lines": [ + { + "bbox": [ + 93, + 180, + 164, + 229 + ], + "spans": [ + { + "bbox": [ + 93, + 180, + 164, + 229 + ], + "type": "image", + "image_path": "6b1d80a1dcc2a0df3695da83eda095de313fd402398b4da4a2432c5406a03612.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 255, + 547, + 278 + ], + "lines": [ + { + "bbox": [ + 46, + 255, + 547, + 278 + ], + "spans": [ + { + "bbox": [ + 46, + 255, + 547, + 278 + ], + "type": "text", + "content": "Figure 5. (a) Grad-cam visualization of feature maps of different multitask dense prediction methods. (b) t-SNE visualization of learned features of all 20 classes on Pascal-Context. (c) Centered Kernel Alignment analyzing the information flow within different networks." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 168, + 180, + 236, + 240 + ], + "blocks": [ + { + "bbox": [ + 168, + 180, + 236, + 240 + ], + "lines": [ + { + "bbox": [ + 168, + 180, + 236, + 240 + ], + "spans": [ + { + "bbox": [ + 168, + 180, + 236, + 240 + ], + "type": "image", + "image_path": "965881496a0205c0e2bae25158d2bcdf03d3f2d14f35bcc7666fe8c9edebc9b7.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 240, + 180, + 309, + 239 + ], + "blocks": [ + { + "bbox": [ + 240, + 180, + 309, + 239 + ], + "lines": [ + { + "bbox": [ + 240, + 180, + 309, + 239 + ], + "spans": [ + { + "bbox": [ + 240, + 180, + 309, + 239 + ], + "type": "image", + "image_path": "c8aa5d41e3c53683f7cbac1eaeacb18e62e14887a1fda2317d957f88c3b22840.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "bbox": [ + 47, + 284, + 261, + 296 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 284, + 261, + 296 + ], + "spans": [ + { + "bbox": [ + 47, + 284, + 261, + 296 + ], + "type": "text", + "content": "3.4. Counter-Intuitive Results of Bi-MTDP-A" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 46, + 304, + 289, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 304, + 289, + 555 + ], + "spans": [ + { + "bbox": [ + 46, + 304, + 289, + 555 + ], + "type": "text", + "content": "Intuitively, implementing binarization on FP network inevitably induces representations degradation, as the gradient of the sign function cannot be perfectly estimated [1]. Thus, binarized models are impossible to outperform their full-precision counterpart models. However, Bi-MTPD-C, a variant of Bi-MTPD (i.e., full-precision backbone with only binarized multi-modal distillation) outperforms its fully FP version. Specifically, just binarizing the multi-modal distillation can simultaneously accelerate the model by " + }, + { + "bbox": [ + 46, + 304, + 289, + 555 + ], + "type": "inline_equation", + "content": "\\sim 39\\%" + }, + { + "bbox": [ + 46, + 304, + 289, + 555 + ], + "type": "text", + "content": " and improve the mIoU for segmentation by " + }, + { + "bbox": [ + 46, + 304, + 289, + 555 + ], + "type": "inline_equation", + "content": "\\sim 4\\%" + }, + { + "bbox": [ + 46, + 304, + 289, + 555 + ], + "type": "text", + "content": ", as shown in Table 1. This result demonstrates that our method is not a naive trade-off between model performance and efficiency but a powerful tool for boosting multitask dense predictors. This exciting 'free-lunch' achievement is even a bit of counter-intuitive. We speculate the reasons are that i) binarization on MMD can filter out task-irrelevant information; ii) and thus the information flow in the network is more effective. To testify this speculation, we conduct a series of experiments in two aspects, the representation ability of Bi-MTPD-C and information flow supplementation within the network." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 46, + 558, + 287, + 569 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 558, + 287, + 569 + ], + "spans": [ + { + "bbox": [ + 46, + 558, + 287, + 569 + ], + "type": "text", + "content": "Qualitative Study of Learned Features with" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 46, + 570, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 570, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 570, + 287, + 713 + ], + "type": "text", + "content": "Bi-MTPD To investigate the representation ability of Bi-MTPD-C and its FP counterparts, we visualize i) the feature maps behind the Binarized Multi-Modal Distillation (MMD) module in 2-D space via the t-SNE [44] algorithm, and ii) the regions where the network considers important via the Grad-Cam algorithm [35]. The results are shown in Fig. 5. It is clear that binarized model, Bi-MTPD-C is able to filter the irrelevant information out via the binarized attention module (see Fig. 5 (a)), and thus helps learn more discriminative features (see Fig. 5 (b)) resulting in higher quantitative results. Overall, the generated spatial feature maps for segmentation are better. The enhanced" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 285, + 545, + 307 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 285, + 545, + 307 + ], + "spans": [ + { + "bbox": [ + 304, + 285, + 545, + 307 + ], + "type": "text", + "content": "representative ability can contribute to higher quantitative results." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 308, + 546, + 464 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 308, + 546, + 464 + ], + "spans": [ + { + "bbox": [ + 304, + 308, + 546, + 464 + ], + "type": "text", + "content": "Analysis of Information Flow Supplementation within Network via Centered Kernel Alignment. Analyzing distributional information flow within layers of neural networks is challenging because outputs of layers are distributed across a large number of neurons. Centered kernel alignment (CKA) [6, 22, 33] can address these challenges, by quantitatively comparing activations within or across networks. Specifically, for a network feed by " + }, + { + "bbox": [ + 304, + 308, + 546, + 464 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 304, + 308, + 546, + 464 + ], + "type": "text", + "content": " samples, CKA algorithm takes " + }, + { + "bbox": [ + 304, + 308, + 546, + 464 + ], + "type": "inline_equation", + "content": "\\mathbf{X} \\in \\mathbb{R}^{m \\times p_1}" + }, + { + "bbox": [ + 304, + 308, + 546, + 464 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 308, + 546, + 464 + ], + "type": "inline_equation", + "content": "\\mathbf{Y} \\in \\mathbb{R}^{m \\times p_2}" + }, + { + "bbox": [ + 304, + 308, + 546, + 464 + ], + "type": "text", + "content": " as inputs which are output activations of two layers (with " + }, + { + "bbox": [ + 304, + 308, + 546, + 464 + ], + "type": "inline_equation", + "content": "p_1" + }, + { + "bbox": [ + 304, + 308, + 546, + 464 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 308, + 546, + 464 + ], + "type": "inline_equation", + "content": "p_2" + }, + { + "bbox": [ + 304, + 308, + 546, + 464 + ], + "type": "text", + "content": " neurons respectively). Letting " + }, + { + "bbox": [ + 304, + 308, + 546, + 464 + ], + "type": "inline_equation", + "content": "\\mathbf{K} \\triangleq \\mathbf{X}\\mathbf{X}^{\\top}" + }, + { + "bbox": [ + 304, + 308, + 546, + 464 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 308, + 546, + 464 + ], + "type": "inline_equation", + "content": "\\mathbf{L} \\triangleq \\mathbf{Y}\\mathbf{Y}^{\\top}" + }, + { + "bbox": [ + 304, + 308, + 546, + 464 + ], + "type": "text", + "content": " denote the Gram matrices for the two layers CKA computes:" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 332, + 472, + 545, + 501 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 332, + 472, + 545, + 501 + ], + "spans": [ + { + "bbox": [ + 332, + 472, + 545, + 501 + ], + "type": "interline_equation", + "content": "\\operatorname {C K A} (\\mathbf {K}, \\mathbf {L}) = \\frac {\\operatorname {H S I C} (\\mathbf {K} , \\mathbf {L})}{\\sqrt {\\operatorname {H S I C} (\\mathbf {K} , \\mathbf {K}) \\operatorname {H S I C} (\\mathbf {L} , \\mathbf {L})}}, \\tag {6}", + "image_path": "c10455978faec9ad830cf91b57625c2171fad3b0acab86e6fb4da42b3dc5c95f.jpg" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 304, + 508, + 545, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 508, + 545, + 629 + ], + "spans": [ + { + "bbox": [ + 304, + 508, + 545, + 629 + ], + "type": "text", + "content": "where HSIC is the Hilbert-Schmidt independence criterion [10]. Given the centering matrix " + }, + { + "bbox": [ + 304, + 508, + 545, + 629 + ], + "type": "inline_equation", + "content": "\\mathbf{H} = \\mathbf{I}_n - \\frac{1}{n}\\mathbf{1}\\mathbf{1}^\\top" + }, + { + "bbox": [ + 304, + 508, + 545, + 629 + ], + "type": "text", + "content": " and the centered Gram matrices " + }, + { + "bbox": [ + 304, + 508, + 545, + 629 + ], + "type": "inline_equation", + "content": "\\mathbf{K}' = \\mathbf{HKH}" + }, + { + "bbox": [ + 304, + 508, + 545, + 629 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 508, + 545, + 629 + ], + "type": "inline_equation", + "content": "\\mathbf{L}' = \\mathbf{HLH}" + }, + { + "bbox": [ + 304, + 508, + 545, + 629 + ], + "type": "text", + "content": ", HSIC " + }, + { + "bbox": [ + 304, + 508, + 545, + 629 + ], + "type": "inline_equation", + "content": "= \\frac{\\operatorname{vec}(\\mathbf{K}')\\operatorname{vec}(\\mathbf{L}')}{(m - 1)^2}" + }, + { + "bbox": [ + 304, + 508, + 545, + 629 + ], + "type": "text", + "content": ", the similarity between these centered Gram matrices. Importantly, CKA is invariant to orthogonal transformation of representations (including permutation of neurons), and the normalization term ensures invariance to isotropic scaling. These properties enable meaningful comparison and analysis of neural network hidden representations." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "content": "Therefore, we introduce CKA to study the information flow in the multitask dense prediction models. In the heatmap, the lighter the dot, the more similar the two corresponding layers. Higher similar score between two layers' output representations means those two layers share more information. The results are presented in Fig. 5 (c), we can see that the similar scores among front layers and back lay" + } + ] + } + ], + "index": 24 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15904" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 63, + 96, + 531, + 236 + ], + "blocks": [ + { + "bbox": [ + 47, + 70, + 545, + 93 + ], + "lines": [ + { + "bbox": [ + 47, + 70, + 545, + 93 + ], + "spans": [ + { + "bbox": [ + 47, + 70, + 545, + 93 + ], + "type": "text", + "content": "Table 1. Results on NYUD-v2. Bi-MTPD-C: only implementing the binarization in the multi-modal distillation module, Bi-MTPD-F: fully-binarized model." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 63, + 96, + 531, + 236 + ], + "lines": [ + { + "bbox": [ + 63, + 96, + 531, + 236 + ], + "spans": [ + { + "bbox": [ + 63, + 96, + 531, + 236 + ], + "type": "table", + "html": "
Model\nMethodBackboneSemSeg ↑Depth ↓Normal ↓Bound ↑float32\nParams (M)binary\nParams (M)FLOPs\n(G)
Cross [28]36.340.62920.8876.38241.460338.09
PAP [53]36.720.61720.8276.42189.100256.86
PSD [54]36.690.62520.8776.42224.670315.60
PAD [49]CNN36.610.62720.8576.38170.980230.91
MTI [47]45.700.53720.2777.86144.870212.98
ATRC [2]45.870.54020.0977.34180.000249.24
MTI + Bi-MTPD-F39.200.61221.0476.866.45138.4218.07
MTI + Bi-MTPD-C47.710.53020.0677.3690.2854.59130.65
InvPT [50]ViT-B50.300.53619.0077.60154.790244.71
InvPT + Bi-MTPD-C51.20(0.90↑)0.528(0.08↓)19.50(0.50↑)77.68(0.08↑)127.9226.87183.81
InvPT [50]ViT-L53.560.51819.0478.10239.220331.60
InvPT + Bi-MTPD-C54.86(1.30↑)0.515(0.03↓)19.50(0.46↑)78.20(0.10↑)212.3426.87301.88
", + "image_path": "a5a778d97b261d86c991104ef702b0f450459ae7106149d998cf91808fc3226b.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 239, + 288, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 239, + 288, + 289 + ], + "spans": [ + { + "bbox": [ + 46, + 239, + 288, + 289 + ], + "type": "text", + "content": "ers in Bi-MTPD-C is much higher than the ones in MTI-Net [47]. This indicates that Bi-MTPD-C is able to supplement information flow within network, and thus boost the model performance." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 297, + 128, + 312 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 297, + 128, + 312 + ], + "spans": [ + { + "bbox": [ + 47, + 297, + 128, + 312 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 317, + 287, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 317, + 287, + 426 + ], + "spans": [ + { + "bbox": [ + 46, + 317, + 287, + 426 + ], + "type": "text", + "content": "In this section, we conducted comprehensive experiments to evaluate our proposed method on two datasets for dense prediction tasks: PASCAL Context [8] and NYUD-v2 [41]. We first describe the implementation details of Bi-MTPD, and then compare our method with SoTA binary neural networks in the task of object detection to demonstrate superiority of the proposed method. Finally, we validate the effectiveness of information bottleneck and feature-based knowledge distillation by a series of ablative studies." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 432, + 282, + 445 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 432, + 282, + 445 + ], + "spans": [ + { + "bbox": [ + 47, + 432, + 282, + 445 + ], + "type": "text", + "content": "4.1. Datasets, Evaluation, Implementation Details" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 450, + 287, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 450, + 287, + 581 + ], + "spans": [ + { + "bbox": [ + 46, + 450, + 287, + 581 + ], + "type": "text", + "content": "Datasets. PASCAL-Content is a popular dataset for dense prediction tasks. We use the split from PASCAL-Context which has annotations for semantic segmentation, human part segmentation, semantic edge detection, surface normals prediction and saliency detection. Note that some annotations were distilled by [27] using pre-trained SoTA models [4]. NYUD-v2 contains various indoor scenes such as offices and living rooms with 795 training and 654 testing images. It provides different dense labels, including semantic segmentation, monocular depth estimation, surface normal estimation and object boundary detection." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 582, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 582, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 582, + 288, + 714 + ], + "type": "text", + "content": "Evaluation. Semantic segmentation (Semseg) and human parsing (Parsing) are evaluated with mean Intersection over Union (mIoU); monocular depth estimation (Depth) is evaluated with Root Mean Square Error (RMSE); surface normal estimation (Normal) is evaluated by the mean error (mErr) of predicted angles; saliency detection (Saliency) is evaluated with maximal F-measure (maxF); object boundary detection (Boundary) is evaluated with the optimal-dataset-scale F-measure (odsF). To evaluate the model efficiency w.r.t. memory cost and inference speed, we adopt the number of parameters and FLOPs for a single round of" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 239, + 450, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 239, + 450, + 251 + ], + "spans": [ + { + "bbox": [ + 305, + 239, + 450, + 251 + ], + "type": "text", + "content": "the model inferring an input image." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 252, + 545, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 252, + 545, + 529 + ], + "spans": [ + { + "bbox": [ + 304, + 252, + 545, + 529 + ], + "type": "text", + "content": "Implementation details. We build our approach on the most prevalent backbone architecture, i.e., HRNet as previous SoTA methods [2, 47]. The task-specific heads are also implemented as two basic residual blocks, i.e., binarized BasicBlock and binarized Bottleneck with additional binary shortcuts as BiReal-Net [23]. We use " + }, + { + "bbox": [ + 304, + 252, + 545, + 529 + ], + "type": "inline_equation", + "content": "\\ell_1" + }, + { + "bbox": [ + 304, + 252, + 545, + 529 + ], + "type": "text", + "content": " loss for depth estimation and cross-entropy loss for semantic segmentation on NYUD-v2. As in the prior work, the edge detection task is trained with a positive weighted wpos = 0:95 binary cross-entropy loss. We do not adopt a particular loss weighing strategy on NYUD-v2, but simply sum the losses together. On PASCAL, we reuse the training setup from [47] to facilitate a fair comparison. We reuse the loss weights from there. The initial task predictions in the front-end of the network use the same loss weighing as the final task predictions. We refer to the supplementary material for further implementation details. Importantly, we use Adam optimizer [20] in training, but with different learning rates for binary parameters (1e-5) and FP parameters (1e-4), as Adam with a larger learning rate for binary parameters can lead to better training results [24]. Note that our project is based on the codebases for MTI-Net [47] and more details can be found in the codes in the Supplemental Materials." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 538, + 488, + 551 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 538, + 488, + 551 + ], + "spans": [ + { + "bbox": [ + 305, + 538, + 488, + 551 + ], + "type": "text", + "content": "4.2. Comparison with State-of-the-Art" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 557, + 545, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 557, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 557, + 545, + 715 + ], + "type": "text", + "content": "Tabs. 1 and 2 present a comparative analysis of the proposed Binary Multitask Dense Predictor (Bi-MTPD) with current state-of-the-art models on the NYUD-v2 and PASCAL-Context datasets. This comparison includes notable CNN-based methods such as PAD-Net [49], ASTMT [27], MTI-Net [47], and ATRC [2], among others. Bi-MTPD demonstrates exceptional performance, outperforming other models in 6 of the 9 evaluated metrics, particularly in complex scene understanding tasks like Semantic Segmentation and Parsing. Remarkably, on the NYUD-v2 benchmark, Bi-MTPD surpasses the previously best-performing CNN-based method (ATRC) by a margin of +1.8 (mIoU) in Semantic Segmentation, while requiring" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15905" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 50, + 96, + 561, + 231 + ], + "blocks": [ + { + "bbox": [ + 47, + 70, + 545, + 92 + ], + "lines": [ + { + "bbox": [ + 47, + 70, + 545, + 92 + ], + "spans": [ + { + "bbox": [ + 47, + 70, + 545, + 92 + ], + "type": "text", + "content": "Table 2. Results on PASCAL-VOC. Bi-MTPD-C: only implementing the binarization in the multi-modal distillation module, Bi-MTPD-F: fully-binarized model." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 50, + 96, + 561, + 231 + ], + "lines": [ + { + "bbox": [ + 50, + 96, + 561, + 231 + ], + "spans": [ + { + "bbox": [ + 50, + 96, + 561, + 231 + ], + "type": "table", + "html": "
Model\nMethodBackboneSemSeg ↑Parsing ↑Saliency ↑Normal ↓Bound ↑float32\nParams (M)binary\nParams (M)FLOPs\n(G)
ASTMT [27]68.0061.1065.7014.7072.40364.720501.27
PAD [49]53.6059.6065.8015.3072.50231.800289.46
MTI [47]61.7060.1884.7814.2370.80218.560280.12
ATRC [2]62.6959.4284.7014.2070.96241.450310.19
ATRC-A [2]63.6062.2383.9114.3070.86249.870320.57
ATRC-B [2]67.6762.9382.2914.2472.42280.010383.21
MTI [47] + Bi-MTPD-F48.1056.2864.4214.2976.8913.67204.8931.38
MTI [47] + Bi-MTPD-C62.98(1.28↑)60.44(0.26↑)83.56(1.22↓)14.31(0.08↑)71.28(0.26↑)153.7164.85194.51
InvPT[50]ViT-B77.5066.8383.6514.6373.00176.350274.68
InvPT[50] + Bi-MTPD-C76.84 (0.66↓)67.10(0.27↑)84.97 (1.32↑)13.69(0.94↓)73.04 (0.04↑)154.6821.67220.76
InvPT[50]ViT-L79.0367.6184.8114.1573.00422.930425.37
InvPT[50] + Bi-MTPD-C79.83 (0.80↑)68.17(0.56↑)
", + "image_path": "aa770149850673ced9116faa9f56c6416ad6ac285ff9d36e4ef45c030613619c.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 237, + 279, + 261 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 237, + 279, + 261 + ], + "spans": [ + { + "bbox": [ + 46, + 237, + 279, + 261 + ], + "type": "text", + "content": "only " + }, + { + "bbox": [ + 46, + 237, + 279, + 261 + ], + "type": "inline_equation", + "content": "62\\%" + }, + { + "bbox": [ + 46, + 237, + 279, + 261 + ], + "type": "text", + "content": " of the storage space for weights and " + }, + { + "bbox": [ + 46, + 237, + 279, + 261 + ], + "type": "inline_equation", + "content": "56\\%" + }, + { + "bbox": [ + 46, + 237, + 279, + 261 + ], + "type": "text", + "content": " of the computational FLOPs." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 262, + 279, + 332 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 262, + 279, + 332 + ], + "spans": [ + { + "bbox": [ + 46, + 262, + 279, + 332 + ], + "type": "text", + "content": "Furthermore, the application of Bi-MTPD to the ViT-based state-of-the-art method, InvPT [50], shows cases Bi-MTPD's ability to enhance model performance while also improving efficiency. This demonstrates the broad applicability and generalization capability of Bi-MTPD across different architectural frameworks." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 334, + 279, + 393 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 334, + 279, + 393 + ], + "spans": [ + { + "bbox": [ + 46, + 334, + 279, + 393 + ], + "type": "text", + "content": "For a qualitative assessment, Figs. 7 and 6 displays prediction examples from various models. These examples illustrate that Bi-MTPD-C not only competes with but occasionally surpasses the state-of-the-art ATRC in qualitative performance." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 403, + 146, + 415 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 403, + 146, + 415 + ], + "spans": [ + { + "bbox": [ + 47, + 403, + 146, + 415 + ], + "type": "text", + "content": "4.3. Ablative Studies" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 422, + 287, + 531 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 422, + 287, + 531 + ], + "spans": [ + { + "bbox": [ + 46, + 422, + 287, + 531 + ], + "type": "text", + "content": "Tab. 3 shows a series of ablative studies of Bi-MTPD-C and Bi-MTPD-F with an HRNet-48 backbone on NYUD-v2. We verify how different components of our model contribute to the multi-task improvements. In summary, every designed module positively impacts the overall model. We would like to highlight the main intuition of our method that binarized dense prediction models in a multitask manner largely outperform the binarized single models through Bi-MTPD-F w.o. VIB & KD vs. Bi-Single." + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 52, + 552, + 287, + 635 + ], + "blocks": [ + { + "bbox": [ + 97, + 538, + 242, + 548 + ], + "lines": [ + { + "bbox": [ + 97, + 538, + 242, + 548 + ], + "spans": [ + { + "bbox": [ + 97, + 538, + 242, + 548 + ], + "type": "text", + "content": "Table 3. Ablative studies on NYUD-v2." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 52, + 552, + 287, + 635 + ], + "lines": [ + { + "bbox": [ + 52, + 552, + 287, + 635 + ], + "spans": [ + { + "bbox": [ + 52, + 552, + 287, + 635 + ], + "type": "table", + "html": "
ModelSemSeg ↑Depth ↓Normal ↓Bound ↑
Bi-MTPD-C47.710.53020.0677.36
Bi-MTPD-C w.o. binary MMD45.700.53720.2777.86
Bi-MTPD-F39.200.61221.0476.86
Bi-MTPD-F w.o. VIB35.040.65222.3173.88
Bi-MTPD-F w.o. KD36.200.64021.9974.86
Bi-MTPD-F w.o. VIB & KD33.990.66723.2471.30
Bi-Single16.200.87229.3669.24
", + "image_path": "d470da1c6387531fdc7b6fa968787abb6bc97bdfb4181758f03ae3845d584f25.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 645, + 119, + 657 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 645, + 119, + 657 + ], + "spans": [ + { + "bbox": [ + 47, + 645, + 119, + 657 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 665, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 287, + 715 + ], + "type": "text", + "content": "In this paper, we significantly accelerate cumbersome dense prediction models, in which BNNs for relevant tasks are modeled and optimized under a multitask framework to supplement degraded information caused by binarization op" + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 307, + 236, + 545, + 338 + ], + "blocks": [ + { + "bbox": [ + 307, + 236, + 545, + 338 + ], + "lines": [ + { + "bbox": [ + 307, + 236, + 545, + 338 + ], + "spans": [ + { + "bbox": [ + 307, + 236, + 545, + 338 + ], + "type": "image", + "image_path": "a3ab1f3b0cc7f333ec50d7394c192dbfda0f09f8fb5740b2042e0acaec76f5d2.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 307, + 365, + 545, + 494 + ], + "blocks": [ + { + "bbox": [ + 307, + 365, + 545, + 494 + ], + "lines": [ + { + "bbox": [ + 307, + 365, + 545, + 494 + ], + "spans": [ + { + "bbox": [ + 307, + 365, + 545, + 494 + ], + "type": "image", + "image_path": "810f13c9eee9f584993405b74e8ed3200030253d88bc5f2f9e8968eb887be7be.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 497, + 545, + 519 + ], + "lines": [ + { + "bbox": [ + 305, + 497, + 545, + 519 + ], + "spans": [ + { + "bbox": [ + 305, + 497, + 545, + 519 + ], + "type": "text", + "content": "Figure 7. Qualitative comparison with MTI-Net on PASCAL-Context." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 526, + 545, + 634 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 526, + 545, + 634 + ], + "spans": [ + { + "bbox": [ + 304, + 526, + 545, + 634 + ], + "type": "text", + "content": "erations. Based on this binary baseline, we further introduce variational information bottleneck and feature-based knowledge distillation to supplement information flow. Experiment results show that our method significantly accelerates existing SoTA methods with comparably small performance drop over the mainstream dense prediction tasks on PASCAL VOC and NYUD-v2. Intriguingly, Bi-MTPD not only reaches SoTA w.r.t. performance but also saves computational costs, compared with SoTA method ARTC [2]." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 634, + 545, + 682 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 634, + 545, + 682 + ], + "spans": [ + { + "bbox": [ + 304, + 634, + 545, + 682 + ], + "type": "text", + "content": "Acknowledgments: This research is partially supported by NSF IIS-2309073, ECCS-212352101 and Cisco unrestricted gift. This article solely reflects the opinions and conclusions of its authors and not the funding agencies." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15906" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 287, + 710 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 123 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 123 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 123 + ], + "type": "text", + "content": "[1] Yoshua Bengio, Nicholas Léonard, and Aaron Courville. Estimating or propagating gradients through stochastic neurons for conditional computation. arXiv:1308.3432, 2013. 2, 3, 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 124, + 287, + 166 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 124, + 287, + 166 + ], + "spans": [ + { + "bbox": [ + 53, + 124, + 287, + 166 + ], + "type": "text", + "content": "[2] David Brüggemann, Menelaos Kanakis, Anton Obukhov, Stamatios Georgoulis, and Luc Van Gool. Exploring relational context for multi-task dense prediction. In ICCV, 2021. 1, 2, 3, 4, 7, 8" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 166, + 287, + 187 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 166, + 287, + 187 + ], + "spans": [ + { + "bbox": [ + 53, + 166, + 287, + 187 + ], + "type": "text", + "content": "[3] Adrian Bulat, Brais Martinez, and Georgios Tzimiropoulos. High-capacity expert binary networks. In ICLR, 2020. 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 188, + 287, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 188, + 287, + 229 + ], + "spans": [ + { + "bbox": [ + 53, + 188, + 287, + 229 + ], + "type": "text", + "content": "[4] Liang-Chieh Chen, Yukun Zhu, George Papandreou, Florian Schroff, and Hartwig Adam. Encoder-decoder with atrous separable convolution for semantic image segmentation. In ECCV, 2018. 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 230, + 287, + 272 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 230, + 287, + 272 + ], + "spans": [ + { + "bbox": [ + 53, + 230, + 287, + 272 + ], + "type": "text", + "content": "[5] Zhao Chen, Vijay Badrinarayanan, Chen-Yu Lee, and Andrew Rabinovich. Gradnorm: Gradient normalization for adaptive loss balancing in deep multitask networks. In ICML, 2018. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 273, + 286, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 273, + 286, + 304 + ], + "spans": [ + { + "bbox": [ + 53, + 273, + 286, + 304 + ], + "type": "text", + "content": "[6] Corinna Cortes, Mehryar Mohri, and Afshin Rostamizadeh. Algorithms for learning kernels based on centered alignment. JMLR, 2012. 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 304, + 287, + 337 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 304, + 287, + 337 + ], + "spans": [ + { + "bbox": [ + 53, + 304, + 287, + 337 + ], + "type": "text", + "content": "[7] Matthieu Courbariaux, Yoshua Bengio, and Jean-Pierre David. Binaryconnect: Training deep neural networks with binary weights during propagations. In NeurIPS, 2016. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 337, + 287, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 337, + 287, + 369 + ], + "spans": [ + { + "bbox": [ + 53, + 337, + 287, + 369 + ], + "type": "text", + "content": "[8] Mark Everingham, Luc Van Gool, Christopher KI Williams, John Winn, and Andrew Zisserman. The pascal visual object classes (voc) challenge. IJCV, 2010. 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 369, + 287, + 411 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 369, + 287, + 411 + ], + "spans": [ + { + "bbox": [ + 53, + 369, + 287, + 411 + ], + "type": "text", + "content": "[9] Yuan Gao, Jiayi Ma, Mingbo Zhao, Wei Liu, and Alan L Yuille. Nddr-cnn: Layerwise feature fusing in multi-task cnns by neural discriminative dimensionality reduction. In CVPR, 2019. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 411, + 287, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 411, + 287, + 443 + ], + "spans": [ + { + "bbox": [ + 48, + 411, + 287, + 443 + ], + "type": "text", + "content": "[10] Arthur Gretton, Kenji Fukumizu, Choon Teo, Le Song, Bernhard Scholkopf, and Alex Smola. A kernel statistical test of independence. In NeurIPS, 2007. 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 443, + 287, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 443, + 287, + 475 + ], + "spans": [ + { + "bbox": [ + 48, + 443, + 287, + 475 + ], + "type": "text", + "content": "[11] Saurabh Gupta, Ross Girshick, Pablo Arbeláez, and Jitendra Malik. Learning rich features from rgb-d images for object detection and segmentation. In ECCV, 2014. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 475, + 287, + 506 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 475, + 287, + 506 + ], + "spans": [ + { + "bbox": [ + 48, + 475, + 287, + 506 + ], + "type": "text", + "content": "[12] Song Han, Huizi Mao, and William J Dally. Deep compression: Compressing deep neural networks with pruning, trained quantization and huffman coding. In ICLR, 2016. 1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 506, + 287, + 548 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 506, + 287, + 548 + ], + "spans": [ + { + "bbox": [ + 48, + 506, + 287, + 548 + ], + "type": "text", + "content": "[13] Xiangyu He, Zitao Mo, Ke Cheng, Weixiang Xu, Qinghao Hu, Peisong Wang, Qingshan Liu, and Jian Cheng. Proxybnn: Learning binarized neural networks via proxy matrices. In CVPR, 2020. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 548, + 287, + 571 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 548, + 287, + 571 + ], + "spans": [ + { + "bbox": [ + 48, + 548, + 287, + 571 + ], + "type": "text", + "content": "[14] Geoffrey Hinton, Oriol Vinyls, and Jeff Dean. Distilling the knowledge in a neural network. In NeurIPS, 2014. 1" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 571, + 287, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 571, + 287, + 602 + ], + "spans": [ + { + "bbox": [ + 48, + 571, + 287, + 602 + ], + "type": "text", + "content": "[15] Itay Hubara, Matthieu Courbariaux, Daniel Soudry, Ran ElYaniv, and Yoshua Bengio. Binarized neural networks. In NeurIPS, 2016. 1, 2, 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 48, + 602, + 287, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 602, + 287, + 644 + ], + "spans": [ + { + "bbox": [ + 48, + 602, + 287, + 644 + ], + "type": "text", + "content": "[16] Xiaoqi Jiao, Yichun Yin, Lifeng Shang, Xin Jiang, Xiao Chen, Linlin Li, Fang Wang, and Qun Liu. Tinybert: Distilling bert for natural language understanding. In EMNLP, 2020. 5" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 48, + 644, + 287, + 677 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 644, + 287, + 677 + ], + "spans": [ + { + "bbox": [ + 48, + 644, + 287, + 677 + ], + "type": "text", + "content": "[17] Alex Kendall, Yarin Gal, and Roberto Cipolla. Multi-task learning using uncertainty to weigh losses for scene geometry and semantics. In CVPR, 2018. 1, 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 48, + 677, + 287, + 710 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 677, + 287, + 710 + ], + "spans": [ + { + "bbox": [ + 48, + 677, + 287, + 710 + ], + "type": "text", + "content": "[18] Hyungjun Kim, Kyungsu Kim, Jinseok Kim, and Jae-Joon Kim. Binary duo: Reducing gradient mismatch in binary activation network by coupling binary activations. In ICLR," + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 39, + "blocks": [ + { + "bbox": [ + 326, + 73, + 356, + 83 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 356, + 83 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 356, + 83 + ], + "type": "text", + "content": "2020. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 83, + 545, + 115 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 83, + 545, + 115 + ], + "spans": [ + { + "bbox": [ + 308, + 83, + 545, + 115 + ], + "type": "text", + "content": "[19] Hyungjun Kim, Jihoon Park, Changhun Lee, and Jae-Joon Kim. Improving accuracy of binary neural networks using unbalanced activation distribution. In CVPR, 2021. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 116, + 545, + 136 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 116, + 545, + 136 + ], + "spans": [ + { + "bbox": [ + 307, + 116, + 545, + 136 + ], + "type": "text", + "content": "[20] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In ICLR, 2014. 7" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 136, + 545, + 178 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 136, + 545, + 178 + ], + "spans": [ + { + "bbox": [ + 307, + 136, + 545, + 178 + ], + "type": "text", + "content": "[21] Iasonas Kokkinos. Ethernet: Training a universal convolutional neural network for low-, mid-, and high-level vision using diverse datasets and limited memory. In CVPR, 2017. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 179, + 545, + 211 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 179, + 545, + 211 + ], + "spans": [ + { + "bbox": [ + 307, + 179, + 545, + 211 + ], + "type": "text", + "content": "[22] Simon Kornblith, Mohammad Norouzi, Honglak Lee, and Geoffrey Hinton. Similarity of neural network representations revisited. In ICML, 2019. 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 211, + 545, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 211, + 545, + 252 + ], + "spans": [ + { + "bbox": [ + 307, + 211, + 545, + 252 + ], + "type": "text", + "content": "[23] Zechun Liu, Wenhan Luo, Baoyuan Wu, Xin Yang, Wei Liu, and Kwang-Ting Cheng. Bi-real net: Binarizing deep network towards real-network performance. IJCV, 2020. 2, 3, 7" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 254, + 545, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 254, + 545, + 285 + ], + "spans": [ + { + "bbox": [ + 307, + 254, + 545, + 285 + ], + "type": "text", + "content": "[24] Zechun Liu, Zhiqiang Shen, Shichao Li, Koen Helwegen, Dong Huang, and Kwang-Ting Cheng. How do adam and training strategies help bnns optimization. In ICML, 2021. 7" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 285, + 545, + 328 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 285, + 545, + 328 + ], + "spans": [ + { + "bbox": [ + 307, + 285, + 545, + 328 + ], + "type": "text", + "content": "[25] Zechun Liu, Zhiqiang Shen, Marios Savvides, and Kwang-Ting Cheng. Reactnet: Towards precise binary neural network with generalized activation functions. In ECCV, 2020. 2, 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 328, + 545, + 371 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 328, + 545, + 371 + ], + "spans": [ + { + "bbox": [ + 307, + 328, + 545, + 371 + ], + "type": "text", + "content": "[26] Yongxi Lu, Abhishek Kumar, Shuangfei Zhai, Yu Cheng, Tara Javidi, and Rogerio Feris. Fully-adaptive feature sharing in multi-task networks with applications in person attribute classification. In CVPR, 2017. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 371, + 545, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 371, + 545, + 403 + ], + "spans": [ + { + "bbox": [ + 307, + 371, + 545, + 403 + ], + "type": "text", + "content": "[27] Kevis-Kokitsi Maninis, Ilija Radosavovic, and Iasonas Kokkinos. Attentive single-tasking of multiple tasks. In CVPR, 2019. 3, 7, 8" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 403, + 545, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 403, + 545, + 435 + ], + "spans": [ + { + "bbox": [ + 307, + 403, + 545, + 435 + ], + "type": "text", + "content": "[28] Ishan Misra, Abhinav Shrivastava, Abhinav Gupta, and Martial Hebert. Cross-stitch networks for multi-task learning. In CVPR, 2016. 1, 2, 7" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 307, + 435, + 545, + 476 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 435, + 545, + 476 + ], + "spans": [ + { + "bbox": [ + 307, + 435, + 545, + 476 + ], + "type": "text", + "content": "[29] Haotong Qin, Zhongang Cai, Mingyuan Zhang, Yifu Ding, Haiyu Zhao, Shuai Yi, Xianglong Liu, and Hao Su. Bipoint-net: Binary neural network for point clouds. In ICLR, 2021. 3" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 307, + 476, + 545, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 476, + 545, + 510 + ], + "spans": [ + { + "bbox": [ + 307, + 476, + 545, + 510 + ], + "type": "text", + "content": "[30] Haotong Qin, Yifu Ding, Mingyuan Zhang, YAN Qinghua, Aishan Liu, Qingqing Dang, Ziwei Liu, and Xianglong Liu. Bibert: Accurate fully binarized bert. In ICLR, 2022. 2, 3, 5" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 307, + 510, + 545, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 510, + 545, + 542 + ], + "spans": [ + { + "bbox": [ + 307, + 510, + 545, + 542 + ], + "type": "text", + "content": "[31] Haotong Qin, Ruihao Gong, Xianglong Liu, Xiao Bai, Jingkuan Song, and Nicu Sebe. Binary neural networks: A survey. PR, 2020. 3" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 307, + 541, + 545, + 584 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 541, + 545, + 584 + ], + "spans": [ + { + "bbox": [ + 307, + 541, + 545, + 584 + ], + "type": "text", + "content": "[32] Haotong Qin, Ruihao Gong, Xianglong Liu, Mingzhu Shen, Ziran Wei, Fengwei Yu, and Jingkuan Song. Forward and backward information retention for accurate binary neural networks. In CVPR, 2020. 2, 3" + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 307, + 584, + 545, + 627 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 584, + 545, + 627 + ], + "spans": [ + { + "bbox": [ + 307, + 584, + 545, + 627 + ], + "type": "text", + "content": "[33] Maithra Raghu, Thomas Unterthiner, Simon Kornblith, Chiyuan Zhang, and Alexey Dosovitskiy. Do vision transformers see like convolutional neural networks? In NeurIPS, 2021. 6" + } + ] + } + ], + "index": 35 + }, + { + "bbox": [ + 307, + 627, + 545, + 659 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 627, + 545, + 659 + ], + "spans": [ + { + "bbox": [ + 307, + 627, + 545, + 659 + ], + "type": "text", + "content": "[34] Lothar Reichel and Qiang Ye. Simple square smoothing regularization operators. Electronic Transactions on Numerical Analysis, 33:63, 2009. 2" + } + ] + } + ], + "index": 36 + }, + { + "bbox": [ + 307, + 659, + 545, + 702 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 659, + 545, + 702 + ], + "spans": [ + { + "bbox": [ + 307, + 659, + 545, + 702 + ], + "type": "text", + "content": "[35] Ramprasaath R Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. Grad-cam: Visual explanations from deep networks via gradient-based localization. In ICCV, 2017. 6" + } + ] + } + ], + "index": 37 + }, + { + "bbox": [ + 307, + 702, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 702, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 702, + 545, + 713 + ], + "type": "text", + "content": "[36] Ozan Sener and Vladlen Koltun. Multi-task learning as" + } + ] + } + ], + "index": 38 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15907" + } + ] + } + ], + "index": 40 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 73, + 287, + 689 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 67, + 73, + 251, + 83 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 73, + 251, + 83 + ], + "spans": [ + { + "bbox": [ + 67, + 73, + 251, + 83 + ], + "type": "text", + "content": "multi-objective optimization. In NeurIPS, 2018. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 84, + 287, + 115 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 84, + 287, + 115 + ], + "spans": [ + { + "bbox": [ + 48, + 84, + 287, + 115 + ], + "type": "text", + "content": "[37] Yuzhang Shang, Bingxin Xu, Gaowen Liu, Ramana Rao Kompella, and Yan Yan. Causal-dfq: Causality guided data-free network quantization. In ICCV, 2023. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 116, + 287, + 147 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 116, + 287, + 147 + ], + "spans": [ + { + "bbox": [ + 49, + 116, + 287, + 147 + ], + "type": "text", + "content": "[38] Yuzhang Shang, Dan Xu, Bin Duan, Ziliang Zong, Liqiang Nie, and Yan Yan. Lipschitz continuity retained binary neural network. In ECCV, 2022. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 148, + 287, + 178 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 148, + 287, + 178 + ], + "spans": [ + { + "bbox": [ + 49, + 148, + 287, + 178 + ], + "type": "text", + "content": "[39] Yuzhang Shang, Dan Xu, Ziliang Zong, and Yan Yan. Network binarization via contrastive learning. In ECCV, 2022. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 49, + 179, + 287, + 211 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 179, + 287, + 211 + ], + "spans": [ + { + "bbox": [ + 49, + 179, + 287, + 211 + ], + "type": "text", + "content": "[40] Yuzhang Shang, Zhihang Yuan, Bin Xie, Bingzhe Wu, and Yan Yan. Post-training quantization on diffusion models. In CVPR, pages 1972-1981, 2023. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 49, + 212, + 287, + 243 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 212, + 287, + 243 + ], + "spans": [ + { + "bbox": [ + 49, + 212, + 287, + 243 + ], + "type": "text", + "content": "[41] Nathan Silberman, Derek Hoiem, Pushmeet Kohli, and Rob Fergus. Indoor segmentation and support inference from rgbd images. In ECCV, 2012. 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 49, + 243, + 287, + 275 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 243, + 287, + 275 + ], + "spans": [ + { + "bbox": [ + 49, + 243, + 287, + 275 + ], + "type": "text", + "content": "[42] Nitish Srivastava, Geoffrey Hinton, Alex Krizhevsky, Ilya Sutskever, and Ruslan Salakhutdinov. Dropout: a simple way to prevent neural networks from overfitting. JMLR, 2014. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 49, + 276, + 287, + 307 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 276, + 287, + 307 + ], + "spans": [ + { + "bbox": [ + 49, + 276, + 287, + 307 + ], + "type": "text", + "content": "[43] Naftali Tishby, Fernando C Pereira, and William Bialek. The information bottleneck method. arXiv preprint physics/0004057, 2000. 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 49, + 308, + 287, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 308, + 287, + 327 + ], + "spans": [ + { + "bbox": [ + 49, + 308, + 287, + 327 + ], + "type": "text", + "content": "[44] Laurens Van der Maaten and Geoffrey Hinton. Visualizing data using t-sne. JMLR, 2008. 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 49, + 328, + 287, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 328, + 287, + 358 + ], + "spans": [ + { + "bbox": [ + 49, + 328, + 287, + 358 + ], + "type": "text", + "content": "[45] Simon Vandenhende, Stamatios Georgoulis, Bert De Brabandere, and Luc Van Gool. Branched multi-task networks: deciding what layers to share. In BMVC, 2020. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 49, + 360, + 287, + 391 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 360, + 287, + 391 + ], + "spans": [ + { + "bbox": [ + 49, + 360, + 287, + 391 + ], + "type": "text", + "content": "[46] S. Vandenhende, S. Georgoulis, W. Van Gansbeke, M. Proesmans, D. Dai, and L. Van Gool. Multi-task learning for dense prediction tasks: A survey. TPAMI, 2021. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 49, + 392, + 287, + 423 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 392, + 287, + 423 + ], + "spans": [ + { + "bbox": [ + 49, + 392, + 287, + 423 + ], + "type": "text", + "content": "[47] Simon Vandenhende, Stamatios Georgoulis, and Luc Van Gool. Mti-net: Multi-scale task interaction networks for multi-task learning. In CVPR, 2020. 1, 2, 3, 4, 5, 7, 8" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 49, + 424, + 287, + 444 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 424, + 287, + 444 + ], + "spans": [ + { + "bbox": [ + 49, + 424, + 287, + 444 + ], + "type": "text", + "content": "[48] Ziwei Wang, Ziyi Wu, Jiwen Lu, and Jie Zhou. Bidet: An efficient binarized object detector. In CVPR, 2020. 5" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 49, + 445, + 287, + 485 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 445, + 287, + 485 + ], + "spans": [ + { + "bbox": [ + 49, + 445, + 287, + 485 + ], + "type": "text", + "content": "[49] Dan Xu, Wanli Ouyang, Xiaogang Wang, and Nicu Sebe. Pad-net: Multi-tasks guided prediction-and-distillation network for simultaneous depth estimation and scene parsing. In CVPR, 2018, 1, 2, 3, 4, 7, 8" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 49, + 487, + 287, + 518 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 487, + 287, + 518 + ], + "spans": [ + { + "bbox": [ + 49, + 487, + 287, + 518 + ], + "type": "text", + "content": "[50] Hanrong Ye and Dan Xu. Inverted pyramid multi-task transformer for dense scene understanding. arXiv preprint arXiv:2203.07997, 2022. 1, 2, 4, 7, 8" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 49, + 520, + 287, + 549 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 520, + 287, + 549 + ], + "spans": [ + { + "bbox": [ + 49, + 520, + 287, + 549 + ], + "type": "text", + "content": "[51] Shang Yuzhang, Duan Bin, Zong Ziliang, Nie Liqiang, and Yan Yan. Lipschitz continuity guided knowledge distillation. In ICCV, 2021. 5" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 49, + 551, + 287, + 582 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 551, + 287, + 582 + ], + "spans": [ + { + "bbox": [ + 49, + 551, + 287, + 582 + ], + "type": "text", + "content": "[52] Hongyi Zhang, Moustapha Cisse, Yann N Dauphin, and David Lopez-Paz. mixup: Beyond empirical risk minimization. In ICLR, 2018. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 49, + 583, + 287, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 583, + 287, + 624 + ], + "spans": [ + { + "bbox": [ + 49, + 583, + 287, + 624 + ], + "type": "text", + "content": "[53] Zhenyu Zhang, Zhen Cui, Chunyan Xu, Yan Yan, Nicu Sebe, and Jian Yang. Pattern-affinitive propagation across depth, surface normal and semantic segmentation. In CVPR, 2019. 2, 7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 49, + 625, + 287, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 625, + 287, + 657 + ], + "spans": [ + { + "bbox": [ + 49, + 625, + 287, + 657 + ], + "type": "text", + "content": "[54] Ling Zhou, Zhen Cui, Chunyan Xu, Zhenyu Zhang, Chaoqun Wang, Tong Zhang, and Jian Yang. Pattern-structure diffusion for multi-task learning. In CVPR, 2020. 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 49, + 658, + 287, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 658, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 49, + 658, + 287, + 689 + ], + "type": "text", + "content": "[55] Barret Zoph and Quoc V Le. Neural architecture search with reinforcement learning. arXiv preprint arXiv:1611.01578, 2016.2" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "15908" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Efficient Privacy-Preserving Visual Localization Using 3D Ray Clouds/61345f4c-613d-4e27-91d1-09fcd6a52a24_content_list.json b/2024/Efficient Privacy-Preserving Visual Localization Using 3D Ray Clouds/61345f4c-613d-4e27-91d1-09fcd6a52a24_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..a938dc592c977b1b43e6d978d76ec2ead90f2d00 --- /dev/null +++ b/2024/Efficient Privacy-Preserving Visual Localization Using 3D Ray Clouds/61345f4c-613d-4e27-91d1-09fcd6a52a24_content_list.json @@ -0,0 +1,2029 @@ +[ + { + "type": "text", + "text": "Efficient Privacy-Preserving Visual Localization Using 3D Ray Clouds", + "text_level": 1, + "bbox": [ + 130, + 130, + 839, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Heejoon Moon1 Chunghwan Lee2 Je Hyeong Hong1,2†", + "bbox": [ + 264, + 179, + 738, + 199 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Dept. Artificial Intelligence, Hanyang University $^{2}$ Dept. Electronic Engineering, Hanyang University", + "bbox": [ + 86, + 198, + 913, + 217 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 236, + 313, + 253 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The recent success in revealing scene details from sparse 3D point clouds obtained via structure-from-motion has raised significant privacy concerns in visual localization. One prominent approach for mitigating this issue is to lift 3D points to 3D lines thereby reducing the effectiveness of the scene inversion attacks, but this comes at the cost of increased algorithmic complexity for camera localization due to weaker geometric constraints induced by line clouds. To overcome this limitation, we propose a new lifting approach called \"ray cloud\", whereby each lifted 3D line intersects at one of two predefined locations, depicting omnidirectional rays from two cameras. This yields two benefits, i) camera localization can now be cast as relative pose estimation between the query image and the calibrated rig of two perspective cameras which can be efficiently solved using a variant of the 5-point algorithm, and ii) the ray cloud introduces erroneous estimations for the density-based inversion attack, degrading the quality of scene recovery. Moreover, we explore possible modifications of the inversion attack to better recover scenes from the ray clouds and propose a ray sampling technique to reduce the effectiveness of the modified attack. Experimental results on two public datasets show real-time localization speed as well as enhanced privacy-preserving capability over the state-of-the-art without overly sacrificing the localization accuracy.", + "bbox": [ + 76, + 270, + 473, + 648 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 664, + 209, + 681 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Visual localization refers to the task of estimating the camera pose w.r.t. a known 3D scene given a query image. It serves a central role for cutting-edge applications involving robotics [18], augmented reality (AR) [3], virtual reality (VR) [6] and simultaneous localization and mapping (SLAM) [11, 29, 33]. To this date, many visual localization algorithms [27, 37] still utilize sparse 3D point clouds constructed via structure-from-motion (SfM) [1, 44, 45, 48] through which the camera pose is estimated from the correspondences between 2D image features and 3D points. This is commonly solved using a perspective-n-point (PnP) solver [12, 20, 37] equipped with the random sample con", + "bbox": [ + 76, + 691, + 468, + 875 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/366bd3567fcbb48e266aa3fc72bbebef5752819862d8dd07d452697953d11b36.jpg", + "image_caption": [ + "(a) Illustration of ray cloud" + ], + "image_footnote": [], + "bbox": [ + 511, + 241, + 681, + 318 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/c7dfd1040ceb28268dd4cf0d37b2dbf182c1696678ef83c6c8dde24ffb6421d3.jpg", + "image_caption": [ + "(b) Single-image localization speeds" + ], + "image_footnote": [], + "bbox": [ + 689, + 241, + 870, + 318 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/db21b4eba16d5d0cd9fc42a7940145b6b5bdaaf47063920ce1ad2c8b5e7f5391.jpg", + "image_caption": [ + "Ground truth", + "(c) Comparison of revealed scene geometries and images", + "Figure 1. Main characteristics of the proposed ray cloud. (a) shows 3D points lifted as 3D rays half of which intersects at center 1 and the other half intersects at center 2. Estimating the camera pose resembles relative pose estimation with respect to a calibrated rig of perspective cameras. As shown in (b), this allows real-time single-image-based localization on two public datasets [22, 55] while preventing scene-revealing attacks [7, 40] as shown in (c)." + ], + "image_footnote": [], + "bbox": [ + 514, + 335, + 612, + 468 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/50f1964bde8163e009f1e75f57c5a4070bac7fe0040adf812122015850fb3c84.jpg", + "image_caption": [ + "Uniform line cloud", + "Paired-Point Lifting (PPL)" + ], + "image_footnote": [], + "bbox": [ + 614, + 335, + 702, + 468 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/048e209533ed2f407f8bf82451cb5f2df28b6a098347615188de72ae8ddc131f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 704, + 335, + 792, + 468 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/8fc928a23ad2fb20720bdf4c923ca6857711f8f30cb9ea18900c33b70e85d60b.jpg", + "image_caption": [ + "Ray cloud (ours)" + ], + "image_footnote": [], + "bbox": [ + 794, + 335, + 883, + 468 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "sensus (RANSAC) meta-algorithm [9, 15, 42] or its deep network alternative such as differentiable RANSAC [4, 5].", + "bbox": [ + 498, + 597, + 890, + 627 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Surprisingly, it was recently shown [8, 10, 13, 30, 40, 57, 58] that a sparse 3D point cloud may contain enough information to reveal image details of the underlying scene. Most prominently, Pittaluga et al. [40] demonstrated a cascaded U-Net model [43] can generate high-fidelity images of the scene just from the 2D projection of sparse 3D points and their SIFT features [28]. This raised significant privacy concerns [39, 50] as anyone with access to the 3D point cloud could potentially search for sensitive details in the private scene by employing this scene inversion network.", + "bbox": [ + 496, + 628, + 892, + 779 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Currently, the mainstream approach for preventing the aforementioned image inversion attack on the sparse 3D point cloud is geometric lifting [26, 46, 50], whereby each 3D point is transformed to a 3D line passing through the original point with aims to conceal the 3D scene geometry (see Fig. 2). While the earlier approach of randomly drawing the line direction (also known as the uniform line cloud) was shown to be susceptible to geometry inversion attack [7],", + "bbox": [ + 496, + 780, + 893, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "† Corresponding author", + "bbox": [ + 101, + 886, + 225, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "9773", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "this was remedied by recent methods utilizing pairing of 3D points [26, 36] to increase ambiguity in the lifted scene representation. Nevertheless, single image-based localization speeds for these algorithms [26, 36] are far from achieving real-time performance, providing motivation for this work.", + "bbox": [ + 75, + 90, + 467, + 166 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we present a new geometric lifting approach called ray cloud with aims to address above issue of slow camera localization runtime exhibited by previous studies. As will be illustrated in Sec. 3, ray cloud is simple to construct and robust to the geometry inversion attack [7]. Most importantly, camera localization using a ray cloud can be formulated as a well-known relative pose estimation problem between the query image and a virtual calibrated rig of cameras, enabling utilization of the efficient 5-point algorithm [35] with known scale [59]. Moreover, by modifying and enhancing the geometry inversion attack [7] in Sec. 4, we explore potential hidden privacy risks associated with ray clouds, and subsequently propose a technique to enhance the privacy protection of ray clouds in Sec. 5.", + "bbox": [ + 75, + 167, + 467, + 377 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The contributions of this work are summarized below:", + "bbox": [ + 94, + 378, + 454, + 392 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- an efficient privacy-preserving visual localization framework based on a new lifting approach called 3D ray cloud,", + "- customization of the geometry inversion attack [7] to yield undesired (privacy leak) corner cases for ray clouds,", + "- a voxel-based ray sampling technique to prevent above privacy leaks by disrupting nearest neighbor statistic and inducing erroneous recovery of the scene points, and", + "- extensive experimental comparison of ray cloud against other line cloud-based approaches on two public datasets, achieving real-time single-image based localization speed for the first time while preserving privacy-preserving capability and maintaining similar localization accuracy." + ], + "bbox": [ + 76, + 393, + 467, + 574 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related work", + "text_level": 1, + "bbox": [ + 76, + 585, + 212, + 599 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Privacy threats from image inversion attack While several prior studies addressed the possibility of revealing images from sparse keypoints [13, 19, 30, 57, 58], it was Pitaluga et al. [40] who posed a real privacy threat for the first time by recovering detailed scene images from a sparse 3D point cloud using a cascaded U-Net [43]. This network called InvSfM takes in the 2D locations of the projected 3D points as well as the corresponding depths and SIFT descriptors, and outputs a synthesized image of the scene.", + "bbox": [ + 75, + 609, + 467, + 744 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "More recently, Song et al. [49] used a PointNet++ architecture to extract deep features and synthesize scene images. Dangwal et al. [10] devised a generic inversion model that can work with various descriptors [2, 28, 53] and proposed selective feature suppression to hide privacy-sensitive regions. Nonetheless, the pretrained InvSfM model still serves as the baseline for qualitative and quantitative analyses in privacy-preserving visual localization [7, 26, 36, 38].", + "bbox": [ + 75, + 747, + 467, + 867 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Privacy-preserving scene representations After the emergence of InvSfM [40], Speciale et al. [50] quickly em", + "bbox": [ + 75, + 869, + 467, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "braced a geometric lifting approach, also known as the uniform line cloud, where each 3D point in the original sparse point cloud is replaced by a 3D line passing through the point with its direction drawn from a uniform distribution on the unit sphere. This scene representation aims to introduce anonymity to the locations of the 3D points thereby preventing any image inversion via InvSfM. The work was later extended by Shibuya et al. [46] to visual SLAM with a prebuilt map provided in the form of uniform line cloud.", + "bbox": [ + 496, + 90, + 890, + 226 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Unfortunately, above methods [46, 50] are now outdated and susceptible to the geometry inversion attack proposed by Chelani et al. [7], which successfully recovers the sparse 3D point cloud from a uniform line cloud (details in the next subsection). To overcome this limitation, Lee et al. [26] proposed to draw 3D lines through random pairs of 3D points, thereby introducing non-uniform line distributions and multiple 3D point candidates for each line both of which are empirically shown to hinder point cloud inversion. Pan et al. [36] proposed to permute the coordinates of random pairs of keypoints to preserve privacy while still enabling localization from multiple axis-aligned lines. Geppert et al. [16] presented a partial localization approach whereby three 1-dimensional partial maps are constructed from the 3D points and separately stored to enhance security. While these approaches can evade the geometry inversion attack, they are not runtime-efficient to allow real-time camera localization from single images. This issue is addressed through use of the ray cloud proposed in this work.", + "bbox": [ + 496, + 227, + 890, + 513 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Other explored approaches include keypoint lifting [14], carving [34] or obfuscation [41], whereby the feature descriptors of the 3D points are transformed to hide keypoint details, but this does not conceal the point cloud geometry.", + "bbox": [ + 496, + 513, + 890, + 575 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Privacy threats for line clouds Chelani et al. [7] presented a density-based inversion attack that can reveal the 3D point geometry from a uniform line cloud [50]. While the posterior probability of the point cloud $\\mathcal{P}$ given a line cloud $\\mathcal{L}$ can be expressed as $P(\\mathcal{P}|\\mathcal{L})\\propto P(\\mathcal{P})P(\\mathcal{L}|\\mathcal{P})$ using Bayes' rule, it is noted that the likelihood $P(\\mathcal{L}|\\mathcal{P})$ is constant for uniform line clouds. Hence, maximizing the posterior amounts to maximizing the marginal prior $P(\\mathcal{P})$ .", + "bbox": [ + 496, + 580, + 890, + 700 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "For maximizing $P(\\mathcal{P})$ , Chelani et al. [7] resorted to an empirical observation that for any two 3D points $\\mathbf{x}_A$ and $\\mathbf{x}_B$ with their corresponding uniformly-drawn lines $\\mathbf{l}_A \\in Gr(1,3)$ and $\\mathbf{l}_B \\in Gr(1,3)$ , $P(\\| \\mathbf{x}_A - \\mathbf{x}_A^\\perp \\|_2 < \\| \\mathbf{x}_A - \\mathbf{x}_B \\|_2) \\approx 0.8$ , where $\\mathbf{x}_A^\\perp$ denotes the point on $\\mathbf{l}_A$ closest to $\\mathbf{l}_B$ . This implies that if $\\mathbf{x}_A$ and $\\mathbf{x}_B$ are close to each other, then $\\mathbf{x}_A^\\perp$ is likely to become a good approximation of $\\mathbf{x}_A$ . Hence, $P(\\mathcal{P})$ can be maximized if the nearest neighboring points of each 3D point are known in advance.", + "bbox": [ + 496, + 702, + 890, + 839 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "As it is non-trivial to find a set of true $k$ -nearest neighboring points in line clouds, Chelani et al. [7] approximates this by the set of $K$ -nearest lines $\\{\\mathbf{l}_K\\}$ for each line $\\mathbf{l}_i$ , expecting some to include the true positives of the $k$ -nearest", + "bbox": [ + 496, + 839, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "9774", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/a7397179caebe7e424db836405385e05a2671ff62b4d64c2b02abc4cbc3a7615.jpg", + "image_caption": [ + "(a) Point cloud" + ], + "image_footnote": [], + "bbox": [ + 86, + 88, + 279, + 172 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/d4b8d335014fe21b9c8c72cfbadff0d8f016075d5650b5b70836abc65f141799.jpg", + "image_caption": [ + "(b) Uniform line cloud [50]" + ], + "image_footnote": [], + "bbox": [ + 287, + 88, + 483, + 172 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/481016dfcf7ad1fdfc014557093e47a2c40c5ab8c519a78c0c4d61f8a7416f00.jpg", + "image_caption": [ + "(c) Paired-Point Lifting (PPL) [26]" + ], + "image_footnote": [], + "bbox": [ + 493, + 88, + 684, + 172 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/140d4d62b1209a412c9ae1a1018164b819389e7169869e78ae463663fb174d3f.jpg", + "image_caption": [ + "(d) Ray cloud (ours)" + ], + "image_footnote": [], + "bbox": [ + 694, + 88, + 885, + 172 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "points when $K > k$ is large enough (we distinguish $K$ for lines and $k$ for points). Bearing this assumption, [7] computes the closest point $\\mathbf{x}_{k}^{\\perp}$ to each of the $K$ -nearest lines for the current line $\\mathbf{l}_i$ , builds a histogram of closest points $\\{\\mathbf{x}_{k}^{\\perp}\\}$ and finds the peak of the histogram via Kuiper's test [23] to estimate the point $\\mathbf{x}_i$ . This attack is effective for uniform line clouds [50], allowing recovered points to be used for scene recovery with InvSfM [40]. However, it is less effective against more recent line-based approaches [26, 36] and ray clouds presented in this work.", + "bbox": [ + 75, + 231, + 468, + 380 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Camera localization using line clouds As mentioned in Sec. 1, camera localization typically involves a RANSAC algorithm [15] coupled with an efficient minimal solver that can run fast over thousands of iterations [25]. For single-image based localization, a standard line cloud introduces 2D points-to-3D line constraints, which are weaker than 2D point-to-3D point constraints provided by the point clouds. Consequently, the minimal solver for line clouds (p6L) requires 6 correspondences between 2D points and 3D lines [50]. Since p6L innately uses the solver for generalized relative pose estimation [52], achieving real-time speed with $p6L$ is challenging due to the large number of possible solutions (up to 64) and high algorithmic complexity [52, 56]. Shibuya et al. [46] achieved real-time localization with line clouds in the SLAM setup, but this requires a video stream as input and construction of a client-side point map. On the other hand, we will show it is possible to perform image-based localization in real-time using ray clouds.", + "bbox": [ + 75, + 392, + 470, + 664 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. 3D Ray cloud", + "text_level": 1, + "bbox": [ + 76, + 681, + 215, + 698 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We now illustrate the process of constructing a 3D ray cloud shown in Fig. 2 and estimating camera pose from ray clouds. We also discuss inversion attack [7] on ray clouds.", + "bbox": [ + 75, + 708, + 468, + 753 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Motivation The concept of lifting 3D lines to intersect at a particular point has been briefly addressed without detailed investigation in [7, 47]. Chelani et al. [7] mentioned that, if all lines were to intersect at a single point, then the geometry inversion attack in Sec. 2 would fail. This is because the histogram of closest points (to other lines) becomes concentrated about the point of intersection, leading to a degenerate solution. Unfortunately, this representation possesses an intrinsic limitation that the camera pose can", + "bbox": [ + 75, + 763, + 468, + 900 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/ad711c4d29a0eb1a16d0bb6d396b996c112abb08b9da1c5cfc9fc5c93e7c7de0.jpg", + "image_caption": [ + "Figure 2. 3D line clouds yielded from different lifting approaches on the Office1 manolis in the Energy Landscape [55] dataset. Note each line (ray) in the ray cloud passes through one of two center points. We show $10\\%$ of the total number of lines for visual comparison.", + "(a) Original 3D points", + "Figure 3. Illustration of the procedure for selecting points of intersection and constructing a 3D ray cloud." + ], + "image_footnote": [], + "bbox": [ + 514, + 229, + 635, + 338 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/608b13162b1499f0fc5964576cdfef2d47e5b5a5d0147e91975ce9425b9ecaeb.jpg", + "image_caption": [ + "(b) Center assignment" + ], + "image_footnote": [], + "bbox": [ + 637, + 229, + 756, + 338 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/f1e809af13ffa8b556d0b403617e6e378601734c90d64b1747bbc490691ec0b8.jpg", + "image_caption": [ + "(c) Ray cloud" + ], + "image_footnote": [], + "bbox": [ + 759, + 229, + 883, + 338 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "only be recovered up to scale as camera localization resembles relative pose estimation [35] between the query camera and the intersecting 3D lines. We aim to overcome this issue by utilizing two intersection points. This can be viewed as adding another camera to the scene, forming a calibrated rig of perspective cameras to resolve the scale ambiguity.", + "bbox": [ + 496, + 391, + 890, + 482 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Construction procedure", + "text_level": 1, + "bbox": [ + 500, + 492, + 718, + 508 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Lifting 3D points to a ray cloud comprises three steps, namely setting two ray centers and assigning 3D points to one of two ray centers. We choose to set the ray centers by applying the $K$ -means clustering algorithm on the point cloud with $K = 2$ , from which the centers of two clusters are set as the ray centers (explained below). Then, we assign half of randomly sampled points from the point cloud to center $1(\\mathbf{c}_1 \\in \\mathbb{R}^3)$ , and the other half to center $2(\\mathbf{c}_2 \\in \\mathbb{R}^3)$ . Note this assignment is not related to the $K$ -means clustering. Finally, each point $\\mathbf{x}_i$ is lifted to a line as $\\mathbf{c}_j + \\lambda (\\mathbf{x}_i - \\mathbf{c}_j)$ for $\\lambda \\in \\mathbb{R}$ , where $j \\in [1,2]$ is the assigned center index for point $i$ . The final output resembles omnidirectional rays from two perspective cameras (see Fig. 3).", + "bbox": [ + 496, + 515, + 890, + 712 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Effect of changing ray center locations If the ray centers are located very close to each other, then they begin to resemble a single ray center, increasing ambiguity in translation scale. On the other hand, if the ray centers are distant from each other and the scene, then the resulting rays are nearly parallel which degrades localization accuracy. We empirically observe that using the centers from $K$ -means clustering usually avoids both extreme cases (see [32]).", + "bbox": [ + 496, + 715, + 890, + 837 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Camera pose estimation", + "text_level": 1, + "bbox": [ + 500, + 847, + 718, + 863 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We now illustrate an efficient localization approach for ray clouds that can benefit from faster inference speed com", + "bbox": [ + 500, + 869, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "9775", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/466228ecdfb17b4e4e8eb72ed74c0210f46a0a4d2c239bf40b2999d82ecc7afc.jpg", + "image_caption": [ + "(a) Standard line cloud $(p6\\mathrm{L})$" + ], + "image_footnote": [], + "bbox": [ + 109, + 90, + 256, + 165 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/b7c13d92d34eeafcf405d7564e5499eeea8139b76eb8e20daf3139d47ee3d644.jpg", + "image_caption": [ + "(b) Ray cloud $(p5 + 1\\mathrm{R})$", + "Figure 4. Illustration of minimal pose estimation problems solved by different line representations. Camera localization using a uniform line cloud can be seen as the special case of generalized relative pose estimation problem [52] involving one perspective camera (query) and one generalized camera (line cloud). On the other hand, camera localization using a ray cloud can be regarded as the case of perspective relative pose estimation problem between the query image and a rig of calibrated cameras (ray cloud). The latter enjoys shorter runtime as shown in Sec. 6 partly due to stronger ray constraints reducing the number of algebraic solutions." + ], + "image_footnote": [], + "bbox": [ + 295, + 90, + 450, + 162 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "pared to other line-based scene representations [26, 36, 50]. This is inspired by the illustration in Fig. 4b that the ray cloud can be viewed as two virtual perspective cameras located at the ray centers $\\mathbf{c}_1 \\in \\mathbb{R}^3$ and $\\mathbf{c}_2 \\in \\mathbb{R}^3$ . Since the rays from these cameras are omnidirectional, we can set their rotation matrices $\\mathsf{R}_1, \\mathsf{R}_2 \\in SO(3)$ to be aligned with world coordinates, i.e. $[\\mathsf{R}_1|\\mathbf{t}_1] = [\\mathsf{I}| - \\mathbf{c}_1]$ and $[\\mathsf{R}_2|\\mathbf{t}_2] = [\\mathsf{I}| - \\mathbf{c}_2]$ , where $\\mathbf{t}_1, \\mathbf{t}_2 \\in \\mathbb{R}^3$ are the camera translations.", + "bbox": [ + 75, + 334, + 467, + 455 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Below, we show that having 5 correspondences between 2D keypoints from the query image and 3D rays from a randomly selected virtual camera and one correspondence between a 2D query keypoint and a 3D ray from a remaining virtual camera allows us to estimate the absolute camera pose using a ray cloud. We denote this adaptation of minimal $5 + 1$ solver [59] as the perspective- $(5 + 1)$ -rays $(p5 + 1\\mathrm{R})$ solver. Similar to previous studies [26, 50], we assume the camera intrinsics $\\mathbf{K} \\in \\mathbb{R}^{3 \\times 3}$ are known.", + "bbox": [ + 75, + 455, + 467, + 590 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "$p5 + 1\\mathbf{R}$ solver We start by estimating the relative pose between the query image and the first virtual camera using the 5-point algorithm [35]. It yields multiple candidates for the relative pose of the query image w.r.t. the first virtual camera defined as $[\\mathtt{R}_{q1}|\\alpha \\hat{\\mathbf{t}}_{q1}]$ , where $\\mathtt{R}_{q1}\\in SO(3)$ is the relative rotation, $\\hat{\\mathbf{t}}_{q1}\\in S^2$ is the normalized relative translation and $\\alpha \\in \\mathbb{R}$ is an unknown scale factor. Subsequently, the absolute pose of query camera is $[\\mathtt{R}_{q1}|\\alpha \\hat{\\mathbf{t}}_{q1} - \\mathtt{R}_{q1}\\mathbf{c}_1]$ .", + "bbox": [ + 75, + 599, + 467, + 720 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In order to disambiguate $\\alpha$ , we utilize the additional correspondence between a normalized query keypoint $\\mathbf{p} := (p_1, p_2)^\\top \\in \\mathbb{R}^2$ and a 3D ray $\\mathbf{l} \\in Gr(1, 3)$ from the second virtual camera. We note the ray from the query camera center passing through $\\mathbf{p}$ would ideally intersect $\\mathbf{l}$ . Expressing both rays in the query-centered coordinates yields", + "bbox": [ + 76, + 720, + 467, + 811 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\lambda_ {1} \\left[ p _ {1}, p _ {2}, 1 \\right] ^ {\\top} = \\lambda_ {2} \\mathrm {R} _ {q 1} \\hat {\\mathbf {n}} + \\alpha \\hat {\\mathbf {t}} _ {q 1} - \\mathrm {R} _ {q 1} \\mathbf {c} _ {1} \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 133, + 823, + 467, + 843 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\hat{\\mathbf{n}}\\in S^2$ is the line direction of 1 in world coordinates and $\\lambda_1,\\lambda_2\\in \\mathbb{R}$ are scale parameters. Since we have 3 equations with 3 unknowns, $\\alpha$ can be determined.", + "bbox": [ + 75, + 854, + 467, + 898 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "While the $p5 + 1\\mathrm{R}$ solver requires 6 correspondences like $p6\\mathrm{L}$ [26, 50], the utilized 5-point algorithm only yields up to 10 solutions [35] that is smaller than 64 solutions from the generalized solver [52] in $p6\\mathrm{L}$ . This significantly reduces the computational complexity for determining the correct configuration. We also enforce the chirality constraint [17] on the query image to discard infeasible solutions.", + "bbox": [ + 496, + 90, + 890, + 196 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Robust estimation We equip above $p5 + 1R$ solver with the LO-RANSAC [31] algorithm for robust camera localization. For selecting the best model candidate, we adopt an MLESAC [54]-type cost function based on the squared sum of epipolar distances in the query image defined as", + "bbox": [ + 496, + 200, + 890, + 276 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\sum_ {(i, j) \\in \\Omega} \\rho \\left(\\frac {\\left([ \\mathbf {p} _ {i} ^ {\\top} , 1 ] \\mathrm {E} _ {j} \\hat {\\mathbf {n}} _ {i}\\right) ^ {2}}{\\left(\\mathbf {e} _ {j 1} ^ {\\top} \\hat {\\mathbf {n}} _ {i}\\right) ^ {2} + \\left(\\mathbf {e} _ {j 2} ^ {\\top} \\hat {\\mathbf {n}} _ {i}\\right) ^ {2}}\\right), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 576, + 286, + 890, + 330 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\rho(s) := \\min(s, \\tau^2)$ is a robust kernel with upper bound of threshold $\\tau^2$ , $\\Omega$ denotes the pairing arrangement of which keypoint $i$ matches with ray $i$ from center $j$ , $\\mathbf{p}_i \\in \\mathbb{R}^2$ is the $i$ -th normalized keypoint in the query image, $\\hat{\\mathbf{n}}_i \\in S^2$ is ray $i$ passing through the ray center $j \\in [1, 2]$ , and $\\mathbf{E}_j := [\\mathbf{e}_{j1}, \\mathbf{e}_{j2}, \\mathbf{e}_{j3}]^\\top$ is the essential matrix between the virtual camera $j$ and the query camera. Since the two virtual cameras share the same rotation and only differ by translation, $\\mathbf{E}_2$ can be easily computed from $\\mathbf{E}_1$ . The best solution from RANSAC is refined by iteratively minimizing (2).", + "bbox": [ + 496, + 343, + 890, + 494 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Geometry inversion attack on ray clouds", + "text_level": 1, + "bbox": [ + 500, + 503, + 846, + 518 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We show that applying the geometry inversion attack [7] naively to a ray cloud yields a degenerate solution. For this purpose, we make use of some prior knowledge regarding the probabilistic viewpoint and algorithmic details of the geometry inversion attack [7] reviewed in Sec. 2.", + "bbox": [ + 496, + 527, + 890, + 602 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Let $\\mathbf{x}_i\\in \\mathbb{R}^3$ be the $i$ -th point and $\\mathbf{l}_i\\in Gr(1,3)$ be its lifted ray. The point cloud is defined as $\\mathcal{P}\\coloneqq \\{\\mathbf{x}_i\\}$ and the ray cloud is defined as $\\mathcal{L}\\coloneqq \\{\\mathbf{l}_i\\}$ . Since ray centers $\\mathcal{C}\\coloneqq \\{\\mathbf{c}_1,\\mathbf{c}_2\\}$ are easily found by inspecting the intersection points, we aim to maximize the conditional posterior of $\\mathcal{P}$ ,", + "bbox": [ + 496, + 602, + 890, + 678 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nP (\\mathcal {P} | \\mathcal {L}, \\mathcal {C}) \\propto P (\\mathcal {L} | \\mathcal {P}, \\mathcal {C}) P (\\mathcal {P} | \\mathcal {C}). \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 583, + 691, + 890, + 708 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Note the conditional likelihood $P(\\mathcal{L}|\\mathcal{P},\\mathcal{C})$ is a non-zero constant as long as each ray intersects the corresponding pair of ray center and 3D point and the probability of each ray being assigned to one of the ray centers is equally likely. As long as each point $\\mathbf{x}_i$ is estimated to be on its lifted ray $\\mathbf{l}_i$ , then maximizing (3) amounts to maximizing $P(\\mathcal{P}|\\mathcal{C})$ .", + "bbox": [ + 496, + 719, + 890, + 809 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "However, since $P(\\mathcal{P}|\\mathcal{C})$ can depend significantly on the ray centers $\\mathcal{C}$ , it seems difficult to yield a point statistic similar to $P(\\mathcal{P})$ in [7]. We implicitly figure out the difference between $P(\\mathcal{P}|\\mathcal{C})$ and the original prior $P(\\mathcal{P})$ by applying the geometry inversion attack [7], which maximizes $P(\\mathcal{P})$ , as-is to the ray cloud. As visualized in Fig. 5a, the estimated", + "bbox": [ + 496, + 810, + 890, + 898 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "9776", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/820beb607c673b3e9c04979fc5db6a32a23db599b489cdbdd4e287dd7aa956fb.jpg", + "image_caption": [ + "(a) Naive neighbor estimation [7]" + ], + "image_footnote": [], + "bbox": [ + 86, + 85, + 271, + 170 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/352a6a3531ff6fd64b52f8d31f5a431e469af2ef26bb992055f026e6e5022559.jpg", + "image_caption": [ + "(b) Modified neighbor estimation", + "Figure 5. Illustration of peak-finding [7] results from (a) naive neighbor-lines estimation [7] and (b) modified neighbor-lines estimation. (a) indicates that the estimated peak will always be located at the center of the current ray. In contrast, our modification only samples the neighborhood of lines using the bundle of rays originating from the opposite center, eliminating the trivial solution." + ], + "image_footnote": [], + "bbox": [ + 277, + 85, + 464, + 170 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "point for each ray becomes the center of the constituent ray, yielding a trivial recovery. This is because, in the process of the geometric inversion, the $K$ -nearest neighboring lines include a bundle of rays originating from the same center, which yields an erroneous peak at the ray center on the histogram of point candidates, hindering general recovery.", + "bbox": [ + 75, + 282, + 468, + 375 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Adapted inversion attack for ray clouds", + "text_level": 1, + "bbox": [ + 76, + 387, + 431, + 404 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this section, we illustrate two main reasons for trivial recovery of 3D points in Sec. 3.3, namely i) lower significance of nearest points for estimating 3D points, and ii) bad estimation of true nearest rays lifted from nearest points.", + "bbox": [ + 75, + 412, + 468, + 473 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "First, the assumption in [7] that the near-80% of candidates which are closest to lines lifted from $k$ -nearest neighboring points yields accurate points does not hold for ray clouds as some of the rays from these $k$ -nearest points will yield estimates at the current ray center. Second, the assumption in [7] that $K$ -nearest lines will overlap enough with the set of rays lifted from the $k$ -nearest points ( $K > k$ ) is weakened as the $K$ -nearest lines will be rays sharing the same center whose original points may be far away from the original point of the current lifted ray.", + "bbox": [ + 75, + 473, + 468, + 625 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In Sec. 4, we attempt to address these issues by considering known ray centers in order to approximately maximize $P(\\mathcal{P}|\\mathcal{C})$ instead of $P(\\mathcal{P})$ . We also aim to address trivial solutions by incorporating the known positions of the ray centers $\\mathcal{C}$ to approximate maximizing over $P(\\mathcal{P}|\\mathcal{C})$ .", + "bbox": [ + 76, + 625, + 468, + 700 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Eliminating the degenerate solution", + "text_level": 1, + "bbox": [ + 76, + 710, + 385, + 726 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The trivial recovery of 3D points (estimated 3D points $\\mathcal{P}$ always located at the two ray centers $\\mathcal{C}$ ) is due to rays intersecting the same ray center always ending up as the nearest neighboring (NN) lines, producing the peak density of closest points at the ray center (see Fig. 5a). To bypass this issue, we exclude the rays intersecting through the current ray when constructing the set of nearest neighboring lines for estimating the peak density, and only select the NN-lines from the other ray center. This can be viewed as incorporating the known locations of the ray centers for recovering the points (see Fig. 5b) to maximize $P(\\mathcal{P}|\\mathcal{C})$ .", + "bbox": [ + 75, + 734, + 468, + 901 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/31dee82cba4c836a92748c426b53ff2f2e76aa06523a1bc5e177aa8f28b34bb9.jpg", + "image_caption": [ + "(a) ULC [50] (oracle)", + "Figure 6. 3D point cloud recovered on different settings of nearest neighbors for the Apt1 living from [55] (black-original points, red-recovered points). In (b) and (c), we have only utilized neighboring rays connected to the opposite ray center to avoid degenerate solutions (in Sec. 4.1). For the oracle cases of (a) and (b), rays from the true nearest points of the original points are used to estimate the points, showing feasible recovery. In contrast, when the nearest neighboring points are approximated by the nearest neighboring rays, it leads to worse recovery as shown in (c). This implies false positive rays, which are close but whose original points are distant, significantly degrade the recovery of points." + ], + "image_footnote": [], + "bbox": [ + 509, + 87, + 632, + 148 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/5729c9489b4c27e7d7f4600c3bf685873b1f69e86d8c7fe9f6ebd38c4c4d959f.jpg", + "image_caption": [ + "(b) Ray cloud (oracle)" + ], + "image_footnote": [], + "bbox": [ + 638, + 88, + 763, + 147 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/a1d4538230a73ddb956fa84df194d8e09ffe729faffaf755b600da0a25970c3f.jpg", + "image_caption": [ + "(c) Ray cloud" + ], + "image_footnote": [], + "bbox": [ + 766, + 89, + 887, + 146 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Fig. 6b shows that, in the oracle case when the ground truth nearest points of the current ray's original 3D point are known, discarding NN-lines intersecting the current ray can avoid degenerate solution and yield more feasible point cloud. At the same time, Fig. 6c shows that the actual point recovery from a ray cloud is worse. We conjecture this is due to the misalignment between the set of nearest neighboring rays and the set of nearest neighboring points for a given point and its lifted ray which is addressed in Sec. 4.2.", + "bbox": [ + 496, + 330, + 890, + 467 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Rejecting rays closely aligned to the baseline", + "text_level": 1, + "bbox": [ + 500, + 479, + 875, + 496 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We first analyze the cause of the bad recovery in Fig. 6c even after applying the trick in Sec. 4.1. We define the rays close to the current ray but whose original points are far away from the original point of the current ray as the false positive rays. By comparing the oracle case (b) and the normal case (c) in Fig. 6, we anticipate that the false positive rays are the culprits for degraded estimation of the points.", + "bbox": [ + 496, + 505, + 890, + 611 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We conjecture two cases where the false positive rays can lead to misleading estimations. First, when there are many rays from the opposite center that are in the proximity of the current ray center, the peak of the histogram may be located near the current ray center. Vice versa, when the current ray passes through the denser area around the opposite center, the peak will be located near the opposite center. These two cases partially cause the 3D points to shrink (see Fig. 6c). Second, when there are many diverse rays in the proximity of the current ray, multiple peaks can be formed from the point candidates (closest points on the ray to neighboring rays) along the current ray. This can arise when the ray centers are embedded in a room-like environment to form congested rays (see [32] for more discussions).", + "bbox": [ + 496, + 612, + 890, + 823 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Out of above cases, we attempt to reduce the shrinking effect of false positive rays by incorporating known $\\mathcal{C}$ and not considering any rays from the opposite center which are within the proximity of the current ray center. This can be effectively achieved by excluding a fraction $(\\beta)$ of rays most", + "bbox": [ + 496, + 825, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "9777", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/f57a7670a21f370da4d522aac9506a6a8398311e743ae7d71fcc43be0fcc437c.jpg", + "image_caption": [ + "(a) Visualization of rejection-sampling of rays near the baseline" + ], + "image_footnote": [], + "bbox": [ + 78, + 90, + 460, + 195 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/521189a74e1679fcf4451972beafe39ac20309e8c7f49929f075ba1c8524b9bf.jpg", + "image_caption": [ + "(b) Peak-fiding [7] w/o rejection", + "Figure 7. Illustration of the effect of line rejection method. For a synthetic room comprising $20\\mathrm{k}$ points, in (b), peak estimation fails to estimate the correct point due to false positive rays. In (c), the rejection excludes false candidates to estimate an accurate point." + ], + "image_footnote": [], + "bbox": [ + 94, + 214, + 267, + 281 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/32ffe51ebfde302520b5f6777252bf730443794d3462c38cf907db5e17a769a8.jpg", + "image_caption": [ + "(c) Peak-finding [7] with rejection" + ], + "image_footnote": [], + "bbox": [ + 284, + 214, + 455, + 281 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "aligned with the baseline when estimating the histogram of point candidates. Fig. 7 illustrates the effect of this ray rejection which can lead to better point estimation.", + "bbox": [ + 75, + 363, + 468, + 407 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We outline some limitations of this approach. First, if the ray lifted from a 3D point is within the region of rejection sampling, then the estimated point can be inaccurate due to the lack of nearby rays to support the estimation process. Second, the method introduces a hyperparameter $(\\beta)$ that needs to be carefully set (see results in Table 2).", + "bbox": [ + 75, + 409, + 468, + 500 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Voxel-based ray sampling", + "text_level": 1, + "bbox": [ + 76, + 512, + 316, + 530 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We now present a new ray sampling technique that can neutralize the modified inversion attack from Sec. 4.", + "bbox": [ + 75, + 537, + 468, + 566 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Our approach is motivated from Fig. 6 that the quality of point recovery depends on finding the true point candidates once the degenerate solution is eliminated. This time, we try to discourage the true nearest neighboring points from being included in the set of $K$ -nearest neighboring lines when lifted to rays with aims to lower the inversion accuracy.", + "bbox": [ + 75, + 568, + 468, + 657 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For the above purpose, we partition the 3D points into voxels and assign one of two ray centers to each voxel such that i) 3D points within the same voxel are lifted to rays intersecting through the assigned ray center and ii) no neighboring voxels share the same ray center (see Fig. 8). As long as the voxels are large enough, many 3D points except those near the voxel boundaries will have most of their respective $k$ -nearest points to be inside the same voxel and subsequently share the same ray center when lifted to rays. Since the rays lifted from these true $k$ -nearest neighboring points inside the same voxel will yield a trivial solution as illustrated in Sec. 4.1, these will be disallowed from being used for peak estimation. Hence, we essentially minimize the overlap between the set of $K$ -nearest lines and the set of $k$ -nearest points which is a necessary condition for successful recovery of points from lines in [7].", + "bbox": [ + 75, + 659, + 468, + 900 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/0a7186f21b53f05692f48bea903a38b8f88983ac77c35b7bec61dad57493b244.jpg", + "image_caption": [ + "Point cloud", + "Figure 8. The procedure of voxel-based ray sampling. After partitioning the point cloud into a voxel grid, the whole grid is divided into two clusters in such a way that neighboring grids belong to different clusters. Then, all 3D points in each grid within the same cluster are paired with one center, while the other cluster is paired with the remaining center for lifting." + ], + "image_footnote": [], + "bbox": [ + 501, + 93, + 617, + 152 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/e9c4f393c95b84e58da45ef0655353fcdff8be0c6ce41c964cef01e49263802d.jpg", + "image_caption": [ + "Voxel-grid" + ], + "image_footnote": [], + "bbox": [ + 620, + 95, + 687, + 150 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/0923c6ea957cd37df3ebb649627c731d1bd44b1353907841e32c662808cc7df4.jpg", + "image_caption": [ + "Grid assignment" + ], + "image_footnote": [], + "bbox": [ + 696, + 94, + 810, + 152 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/c8346fded2e911cfce79ce695ae1de17b31c5a5cc371e3a114c3630eb3fad2f2.jpg", + "image_caption": [ + "Lifting" + ], + "image_footnote": [], + "bbox": [ + 813, + 93, + 890, + 152 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We adopt a 3D checkerboard pattern for assigning ray centers to voxels with aims to achieve stable camera localization accuracy by uniformly distributing rays from different ray centers across the scene. While a larger voxel size will allow more true nearest neighboring points to be discarded and provide better privacy protection, it will also cause the rays to be more non-uniform and lower the camera localization accuracy (see [32]). Currently, the voxel size needs to be manually determined for each dataset as the optimal trade-off point can vary across different scenes.", + "bbox": [ + 496, + 258, + 890, + 409 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "6. Experimental Results", + "text_level": 1, + "bbox": [ + 500, + 422, + 705, + 439 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We present comprehensive experimental results showing the effectiveness of ray cloud in terms of localization accuracy, inference speed, and privacy-preserving performance.", + "bbox": [ + 498, + 448, + 890, + 494 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Datasets We used the same two public datasets as in [7, 26]. The Energy Landscape [55] dataset consists of 12 indoor scenes with multiple sequences of scanned RGB images. The Cambridge [21, 22] dataset comprises 6 distinct outdoor landmarks, containing multiple image frames with labeled 6-DOF camera poses for the camera re-localization. As in [26], we excluded the scenes of Street and Great Court from [22] due to the presence of large geometric outliers. We followed the same experimental protocols from [26, 50] for constructing sparse 3D point clouds using COLMAP [44] and evaluation. Separated queries not included in 3D map construction were used for evaluation.", + "bbox": [ + 496, + 498, + 892, + 679 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation details We used Poselib [24] to utilize the $p5 + 1\\mathrm{R}$ solver and the repository of PPL [26] to run experiments. We report coarse estimation results of inversion in ray cloud and voxel-based ray cloud as the refine estimation [7] diverges due to incorrect initial estimation (see [32]). We used a workstation with an Intel CPU i9-13900K running at 3.0 GHz and NVIDIA RTX 4090 GPUs.", + "bbox": [ + 496, + 684, + 890, + 789 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation metrics For camera localization, we estimated the absolute 6-DOF pose of the query image with given 2D-3D correspondences. We define $\\mathsf{R} \\in SO(3)$ and $\\mathbf{t} \\in \\mathbb{R}^3$ as the ground truth camera rotation and translation respectively, and $\\hat{\\mathbf{R}} \\in SO(3)$ and $\\hat{\\mathbf{t}} \\in \\mathbb{R}^3$ as the estimated rotation and translation respectively. Then, we obtained the estimated pose errors by following [26, 50, 51], where the", + "bbox": [ + 496, + 795, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "9778", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/27c0946544e607ff07613763855963a08fc2f8dfab2d133b1d6461dc7370be83.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
DatasetScene representationMinimal solverPose errorRANSACPrivacy preserving
ΔR↓meanmedianΔt↓meanmediani meanr meant ↓meaneg↑ medianPSNR ↓ meanSSIM ↓ meanMAE ↑ mean
Energy LandscapePoint cloudp3P0.1780.0310.0070.0011199.544-15.760.51633.29
Uniform line cloud [50]p6L0.2720.0980.0100.00312481.12900.04513.450.40342.83
PPL [26]p6L0.2920.1130.0110.00412381.51880.05610.420.35562.61
PPL+ [26]p6L0.2930.1150.0110.00412581.30850.05610.390.35562.70
Ray cloud (ours)p5+1R0.2620.0870.0100.00314680.0360.66910.290.31962.97
Voxel-based ray cloud (ours)p5+1R0.2970.1160.0110.00414480.1560.77010.050.31564.84
CambridgePoint cloudp3P0.7000.1010.1420.0431298.487-14.000.43339.40
Uniform line cloud [50]p6L0.7750.1690.1720.06643671.872890.74512.710.33146.08
PPL [26]p6L0.7750.1830.1810.08126474.491860.97711.210.29655.53
PPL+ [26]p6L0.7640.1830.1780.08128474.111900.94811.160.29755.91
Ray cloud (ours)p5+1R0.8550.2220.2190.10723974.89129.61410.410.26160.98
Voxel-based ray cloud (ours)p5+1R0.8800.2560.2350.12025075.241210.5410.230.25362.56
", + "bbox": [ + 88, + 84, + 879, + 271 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 1. Comparison of localization and privacy-preserving performance with different 3D scene representations ( $\\Delta R$ : rotation error $[\\circ]$ , $\\Delta t$ : translation error [m], i: number of iterations, r: inlier ratio (\\%), t: runtime [ms], $e_g$ : geometric (3D point) error [m]). We applied modified inversion attack from Sec. 4 with 25% line rejection to ray cloud and voxel-based ray cloud with voxel size of 0.5m and 5.0m for Energy Landscape and Cambridge respectively. Bold indicate the best result in each metric among privacy-preserving representations.", + "bbox": [ + 76, + 273, + 893, + 330 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/eb107fcb95284603b7c0af91d1949ca09f0cdb3dd6c1211101ca5468ca390fac.jpg", + "image_caption": [ + "Figure 9. Comparison of revealed scene geometries and images for Apt2 bed [55]. Modified inversion attack is applied to (d) and (e)." + ], + "image_footnote": [], + "bbox": [ + 93, + 333, + 883, + 479 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "rotational error is computed as $\\Delta \\mathsf{R} = \\arccos \\frac{\\mathrm{Tr}(\\hat{\\mathsf{R}}^{\\top}\\mathsf{R}) - 1}{2}$ and the translation error computed as $\\Delta \\mathbf{t} = \\left\\| \\hat{\\mathsf{R}}^{\\top}\\hat{\\mathbf{t}} -\\mathsf{R}^{\\top}\\mathbf{t}\\right\\|_{2}$ .", + "bbox": [ + 75, + 502, + 470, + 539 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To compare the quality of recovered point clouds, we calculated the 3D median point errors between the original and estimated points in meters as in [26]. We also compared the quality of revealed images by feeding recovered point cloud and ground truth camera pose of query images to InvSfM [39] and calculating peak signal-to-noise ratio (PSNR), structural similarity index measure (SSIM), and mean absolute error (MAE) values against the queries.", + "bbox": [ + 75, + 539, + 468, + 659 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "6.1. Camera localization accuracy and runtime", + "text_level": 1, + "bbox": [ + 76, + 666, + 441, + 681 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "As shown in Table 1, our ray cloud family of approaches achieved the fastest inference speed among the line-lifting methods with no significant decline in pose accuracy. Notably, these achieved real-time performance on both datasets, reaching 137 FPS for the ray cloud and 136 FPS for the voxel-based ray cloud on average. As further demonstrated in [32], this is mainly due to the introduction of the $p5 + 1\\mathrm{R}$ solver which is only available for ray clouds.", + "bbox": [ + 75, + 689, + 468, + 810 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The ray cloud showed the lowest localization error on [55] but slightly higher errors than ULC and PPL on [22]. We conjecture this is due to the increased baseline between two ray centers for the outdoor scenes. We also note that the voxel-based ray cloud does not exhibit a significant reduction in pose error compared to the standard ray cloud.", + "bbox": [ + 75, + 810, + 470, + 901 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "6.2. Quality of revealed scene geometry and images", + "text_level": 1, + "bbox": [ + 498, + 503, + 890, + 520 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The quantitative comparisons of scene reconstruction quality are also shown in Table 1. For a fair comparison, we applied our adapted inversion attack in Sec. 4 for ray cloud and voxel-based ray cloud, while we used the original geometry inversion attack [7] with 3 refinement steps for uniform line cloud (ULC), PPL and $\\mathrm{PPL + }$ .We set $\\beta = 25\\%$ for the ray rejection threshold as it shows the best point estimation in Table 2. Results with $\\beta = 50\\%$ are included in [32].", + "bbox": [ + 496, + 527, + 892, + 648 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Recovery of 3D point clouds The median point errors for ray cloud and voxel-based ray cloud are notably larger than those of other methods, even after applying our adjusted density-based attacks. Additionally, to validate the impact of our adapted inversion attack, we compared the geometric error with and without rejection sampling, showing that rejection sampling contributes to error reduction in Table 2. The qualitative result in Fig. 9 shows that the results from ray cloud and voxel-based ray cloud are unrecognizable.", + "bbox": [ + 496, + 652, + 893, + 789 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Degraded quality of scene details Fig. 9 shows the images recovered from the ray cloud and the voxel-based ray cloud are hardly recognizable. In Table. 1, the voxel-based ray cloud yields the lowest PSNR, SSIM and MAE values with the original ray cloud as a runner-up. This indicates large point errors exhibited by ray clouds translate to lower image quality.", + "bbox": [ + 496, + 795, + 893, + 901 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "9779", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/5642d4b78ddbd722cc01f081a25ccb06d659ffb0084eb81f77db70f02137531b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 101, + 90, + 352, + 170 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/5d7d4a00975b879f1f8ee02fcaa94a1077afac9292ecbf3e04a9865961a01d54.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 362, + 90, + 612, + 170 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/27a4ad7e2e4a5677d827d58c60993e60b4b7ee36271ed13907f6b3cb6fbde635.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 622, + 90, + 875, + 170 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/0f673db3068e79c2ce0b4021075ce17d09f9a6bf08e575697d53e0ff84ffdd18.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 94, + 171, + 349, + 250 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/a3155d9364c152845ad8dd9e4cd1ae44ec5237e81da7cd11ab4472044b1ccc2e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 357, + 171, + 612, + 251 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/8434caf5240d2be464914b669eeff13aadb2c4d11235891eb4c97d4dcc3469d6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 620, + 171, + 875, + 251 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/830a66628c4ef3b7b6cb7f91ee2fa9de75fb0382b3152a1d5c2639228662f83d.jpg", + "image_caption": [ + "Figure 10. Average localization performance vs. reconstructed image quality across different levels of the ray density on the Energy Landscape [55] dataset. The tested line densities are $100\\%$ (biggest marker), $50\\%$ , $25\\%$ and $10\\%$ (smallest marker)." + ], + "image_footnote": [], + "bbox": [ + 96, + 252, + 349, + 330 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/758292b2a65000ee87c6e1a9a95cb05288f5927811877e7f27b5ddaaf04bb487.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 359, + 252, + 611, + 330 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/0994ad02842ca40f62e80b2cc7d3d15a8e9ffca8af370b61b178bf4b9c275b20.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 622, + 252, + 875, + 330 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/951fed8a2095f6b00d7a9acb41f206e912174989e6daed63800e8ab80fde55a3.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
DatasetRejection threshold (β)eg[m] medianPSNR meanSSIM meanMAE mean
Energy LandscapeNone1.0279.9590.33066.06
10 (%)0.76110.240.32363.62
25 (%)0.66910.290.31962.97
50 (%)0.74610.320.32062.52
CambridgeNone11.4710.100.27364.06
10 (%)10.2910.350.26761.60
25 (%)9.61410.410.26160.98
50 (%)12.989.9440.24264.70
", + "bbox": [ + 91, + 363, + 452, + 489 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 2. Effect of ray rejection (see Sec. 4.2) on the quality of scene recovery. Rejecting rays close to the baseline of two ray centers overall reduces the point estimation errors. Bold represents the best inversion results amongst different thresholds.", + "bbox": [ + 75, + 489, + 468, + 546 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Additionally, Fig. 10 comparing the pose estimation errors against image reconstruction qualities demonstrate that the ray cloud-based approaches are capable of preserving the privacy especially for dense scenes while significantly decreasing the inference runtime. In contrast, the voxel-based ray cloud shows slightly higher pose estimation errors potentially due to coagulation of rays yielding less uniform distribution of rays across the scene. The fact that the SSIM value increases for lower density ray clouds also needs further investigation (see [32] for qualitative results).", + "bbox": [ + 75, + 556, + 468, + 707 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In Fig. 11, we demonstrate some corner cases revealed through our adapted inversion attack in Sec. 4. Fig. 11d shows the voxel-based ray sampling proposed in Sec. 5 can degrade the quality of scene recovery.", + "bbox": [ + 75, + 708, + 468, + 768 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "7. Conclusion", + "text_level": 1, + "bbox": [ + 76, + 784, + 194, + 799 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We presented a privacy-preserving visual localization framework based on a new scene representation called 3D ray cloud, which is robust to the known geometry inversion attack and achieves an order of magnitude faster localization than previous line-based approaches without significantly compromising the localization accuracy. This is", + "bbox": [ + 75, + 810, + 468, + 900 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/0d983a917d4d9abb839885a7fd993bd44ed790063e17117807e29f89aef6c84c.jpg", + "image_caption": [ + "(a) ground truth" + ], + "image_footnote": [], + "bbox": [ + 514, + 364, + 602, + 498 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/55b9e92788ca61edade378f77089139dc0824f10b29d680aa9582e8823fc2bb7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 607, + 363, + 694, + 431 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/39b2f6884efa3440ead7868be06e85e1090736218f613c692e94517471d8b1c9.jpg", + "image_caption": [ + "(b) w/o rejection" + ], + "image_footnote": [], + "bbox": [ + 607, + 431, + 694, + 500 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/8582d96efe7f038d7a7399a806e7f1c224b75d9e14c03374bbb8d1702b5b70a7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 700, + 364, + 787, + 431 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/29ee5d1d0c067f767159fbe18b77c190e34b72461c5a17fbebb85b54a04f6409.jpg", + "image_caption": [ + "(c) w/ rejection" + ], + "image_footnote": [], + "bbox": [ + 700, + 431, + 787, + 500 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/4b0ddd2170e980e7e7c1ad24511abae142139e0b2ffaed688b1a7ecbd51bb5b7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 794, + 364, + 882, + 431 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/ba40c2e6c34ee9efcd1fe4be39c61baf2c6c8b75f6609ab9fc8f7bc94495f17d.jpg", + "image_caption": [ + "(d) voxel-based" + ], + "image_footnote": [], + "bbox": [ + 794, + 431, + 880, + 500 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Figure 11. Illustration of point-to-image translation [40] results with point clouds from (a) ground truth, (b) recovered from ray cloud without rejection sampling, (c) recovered with rejection sampling, and (d) recovered from voxel-based ray cloud. While (c) shows improved recovery via ray rejection in Sec. 4.2, (d) shows voxel-based ray sampling lowering the reconstruction quality.", + "bbox": [ + 498, + 515, + 890, + 599 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "enabled by the intrinsic geometry of the ray cloud allowing utilization of efficient widely-available perspective relative pose estimation solver. We have also explored potential improvements for the geometry-inversion attack on ray clouds and proposed a respective defense mechanism.", + "bbox": [ + 496, + 604, + 890, + 681 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Limitation Although this work can successfully defend the known inversion attack [7], we note it may be potentially vulnerable to a new type of attack in which omnidirectional rays are projected to a cylindrical or spherical image plane which is then unfolded and directly passed through an inversion network for image reconstruction. While realizing this attack may face its own challenges such as unknown depths and mixed chirality, we believe utilizing more ray centers will provide a defense against such attack by reducing the density of ray projections in each ray center.", + "bbox": [ + 496, + 685, + 890, + 835 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgement This work was supported by the NRF (Natioanl Research Foundation of Korea) grants funded by the Korea government (MSIT) (No. 2022R1C1C1004907). We thank anonymous reviewers for their valuable feedback.", + "bbox": [ + 496, + 839, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "9780", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 174, + 106 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Sameer Agarwal, Yasutaka Furukawa, Noah Snavely, Ian Simon, Brian Curless, Steven M Seitz, and Richard Szeliski. Building Rome in a day. Communications of the ACM, 54(10):105-112, 2011. 1", + "[2] Alexandre Alahi, Raphael Ortiz, and Pierre Vandergheynst. Freak: Fast retina keypoint. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 510-517, 2012. 2", + "[3] Clemens Arth, Daniel Wagner, Manfred Klopschitz, Arnold Irschara, and Dieter Schmalstieg. Wide area localization on mobile phones. In 2009 8th IEEE International Symposium on Mixed and Augmented Reality, pages 73-82. IEEE, 2009. 1", + "[4] Eric Brachmann, Alexander Krull, Sebastian Nowozin, Jamie Shotton, Frank Michel, Stefan Gumhold, and Carsten Rother. DSAC — Differentiable RANSAC for camera localization. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 2492–2500, 2017. 1", + "[5] Eric Brachmann and Carsten Rother. Learning less is more - 6D camera localization via 3D surface regression. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4654-4662, 2018. 1", + "[6] Robert Castle, Georg Klein, and David W Murray. Videorelate localization in multiple maps for wearable augmented reality. In 2008 12th IEEE International Symposium on Wearable Computers, pages 15-22. IEEE, 2008. 1", + "[7] Kunal Chelani, Fredrik Kahl, and Torsten Sattler. How privacy-preserving are line clouds? Recovering scene details from 3D lines. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 15663-15673, 2021. 1, 2, 3, 4, 5, 6, 7, 8", + "[8] Kunal Chelani, Torsten Sattler, Fredrik Kahl, and Zuzana Kukelova. Privacy-preserving representations are not enough: Recovering scene content from camera poses. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13132-13141, 2023. 1", + "[9] Ondrej Chum, Jiri Matas, and Josef Kittler. Locally optimized ransac. In Pattern Recognition: 25th DAGM Symposium, Magdeburg, Germany, September 10-12, 2003. Proceedings 25, pages 236-243. Springer, 2003. 1", + "[10] Deeksha Dangwal, Vincent T Lee, Hyo Jin Kim, Tianwei Shen, Meghan Cowan, Rajvi Shah, Caroline Trippel, Brandon Reagen, Timothy Sherwood, Vasileios Balntas, et al. Mitigating reverse engineering attacks on local feature descriptors. In Proceeding of the British Machine Vision Conference (BMVC), 2021. 1, 2", + "[11] Andrew J. Davison, Ian D. Reid, Nicholas D. Molton, and Olivier Stasse. MonoSLAM: Real-time single camera SLAM. IEEE Transactions on Pattern Analysis and Machine Intelligence, 29(6):1052-1067, 2007. 1", + "[12] Yaqing Ding, Jian Yang, Viktor Larsson, Carl Olsson, and Kalle Åström. Revisiting the p3p problem. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4872-4880, 2023. 1", + "[13] Alexey Dosovitskiy and Thomas Brox. Inverting visual representations with convolutional networks. In Proceedings" + ], + "bbox": [ + 78, + 114, + 470, + 888 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 4829-4837, 2016. 1, 2", + "[14] Mihai Dusmanu, Johannes L. Schonberger, Sudipta N. Sinha, and Marc Pollefeys. Privacy-preserving image features via adversarial affine subspace embeddings. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 14262-14272, 2021. 2", + "[15] Martin A. Fischler and Robert C. Bolles. Random sample consensus: A paradigm for model fitting with applications to image analysis and automated cartography. Communications of the ACM, 24(6):381-395, June 1981. 1, 3", + "[16] Marcel Geppert, Viktor Larsson, Johannes L Schonberger, and Marc Pollefeys. Privacy preserving partial localization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 17337-17347, 2022. 2", + "[17] Richard Hartley. Cheirality. International journal of computer vision, 26(1):41-61, 1998. 4", + "[18] Lionel Heng, Benjamin Choi, Zhaopeng Cui, Marcel Geppert, Sixing Hu, Benson Kuan, Peidong Liu, Rang Nguyen, Ye Chuan Yeo, Andreas Geiger, et al. Project automation: Localization and 3d scene perception for an autonomous vehicle with a multi-camera system. In Proceedings of the IEEE International Conference on Robotics and Automation (ICRA), pages 4695-4702, 2019. 1", + "[19] Hiroharu Kato and Tatsuya Harada. Image reconstruction from bag-of-visual-words. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 955–962, 2014. 2", + "[20] Tong Ke and Stergios I Roumeliotis. An efficient algebraic solution to the perspective-three-point problem. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 7225-7233, 2017. 1", + "[21] Alex Kendall and Roberto Cipolla. Geometric loss functions for camera pose regression with deep learning. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 6555-6564, 2017. 6", + "[22] Alex Kendall, Matthew Grimes, and Roberto Cipolla. PoseNet: A convolutional network for real-time 6-DOF camera relocalization. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), pages 2938-2946, 2015. 1, 6, 7", + "[23] Nicolaas H Kuiper. Tests concerning random points on a circle. In Nederl. Akad. Wetensch. Proc. Ser. A, volume 63, pages 38-47, 1960. 3", + "[24] Viktor Larsson. PoseLib - Minimal Solvers for Camera Pose Estimation. https://github.com/vlarsson/ PoseLib, 2020. Accessed: 2022-10-30. 6", + "[25] Viktor Larsson, Kalle Astrom, and Magnus Oskarsson. Efficient solvers for minimal problems by syzygy-based reduction. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 820-829, 2017. 3", + "[26] Chunghwan Lee, Jaihoon Kim, Chanhyuk Yun, and Je Hyeong Hong. Paired-point lifting for enhanced privacy-preserving visual localization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 17266-17275, 2023. 1, 2, 3, 4," + ], + "bbox": [ + 501, + 92, + 890, + 892 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9781", + "bbox": [ + 482, + 945, + 513, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "6,7", + "[27] Hyon Lim, Sudipta N Sinha, Michael F Cohen, and Matthew Uytendaele. Real-time image-based 6-DOF localization in large-scale environments. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 1043–1050. IEEE, 2012. 1", + "[28] David G. Lowe. Distinctive image features from scale-invariant keypoints. International Journal of Computer Vision, 60(2):91-110, Nov. 2004. 1, 2", + "[29] Simon Lynen, Torsten Sattler, Michael Bosse, Joel A Hesch, Marc Pollefeys, and Roland Siegwart. Get out of my lab: Large-scale, real-time visual-inertial localization. In Robotics: Science and Systems, volume 1, page 1, 2015. 1", + "[30] Aravindh Mahendran and Andrea Vedaldi. Understanding deep image representations by inverting them. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 5188-5196, 2015. 1, 2", + "[31] Anastasiia Mishchuk, Dmytro Mishkin, Filip Radenovic, and Jiri Matas. Working hard to know your neighbor's margins: Local descriptor learning loss. Advances in Neural Information Processing Systems (NeurIPS), 30, 2017. 4", + "[32] Heejoon Moon, Chunghwan Lee, and Je Hyeong Hong. Supplementary document for efficient privacy-preserving visual localization using 3d ray clouds, 2024. 3, 5, 6, 7, 8", + "[33] Raul Mur-Artal, Jose Maria Martinez Montiel, and Juan D Tardos. ORB-SLAM: A versatile and accurate monocular SLAM system. IEEE Transactions on Robotics, 31(5):1147-1163, 2015. 1", + "[34] Tony Ng, Hyo Jin Kim, Vincent T Lee, Daniel DeTone, Tsun-Yi Yang, Tianwei Shen, Eddy Ilg, Vassileios Balntas, Krystian Mikolajczyk, and Chris Sweeney. NinjaDesc: Content-concealing visual descriptors via adversarial learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12797-12807, 2022. 2", + "[35] David Nister. An efficient solution to the five-point relative pose problem. IEEE Transactions on Pattern Analysis and Machine Intelligence, 26(6):756-770, 2004. 2, 3, 4", + "[36] Linfei Pan, Johannes Lutz Schonberger, Viktor Larsson, and Marc Pollefeys. Privacy Preserving Localization via Coordinate Permutations. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), 2023. 2, 3, 4", + "[37] Mikael Persson and Klas Nordberg. Lambda twist: An accurate fast robust perspective three point (p3P) solver. In Proceedings of the European conference on computer vision (ECCV), pages 318-332, 2018. 1", + "[38] Maxime Pietrantoni, Martin Humenberger, Torsten Sattler, and Gabriela Csurka. Segloc: Learning segmentation-based representations for privacy-preserving visual localization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15380-15391, 2023. 2", + "[39] Francesco Pittaluga, Sanjeev Koppal, and Ayan Chakrabarti. Learning privacy preserving encodings through adversarial training. In Proceedings of the IEEE Winter Conference on Applications of Computer Vision (WACV), pages 791-799. IEEE, 2019. 1, 7", + "[40] Francesco Pittaluga, Sanjeev J Koppal, Sing Bing Kang, and Sudipta N Sinha. Revealing scenes by inverting structure from motion reconstructions. In Proceedings of the" + ], + "bbox": [ + 78, + 93, + 468, + 891 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 145-154, 2019. 1, 2, 3, 8", + "[41] Francesco Pittaluga and Bingbing Zhuang. Ldp-feat: Image features with local differential privacy. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 17580–17590, 2023. 2", + "[42] Rahul Raguram, Ondrej Chum, Marc Pollefeys, Jiri Matas, and Jan-Michael Frahm. USAC: A universal framework for random sample consensus. IEEE Transactions on Pattern Analysis and Machine Intelligence, 35(8):2022-2038, 2012. 1", + "[43] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-Net: Convolutional networks for biomedical image segmentation. In Proceedings of the International Conference on Medical Image Computing and Computer-Assisted Intervention (MICCAI), pages 234–241. Springer, 2015. 1, 2", + "[44] Johannes Lutz Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 1, 6", + "[45] Johannes Lutz Schonberger, Enliang Zheng, Marc Pollefeys, and Jan-Michael Frahm. Pixelwise view selection for unstructured multi-view stereo. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part III 14, pages 501-518. Springer, 2016. 1", + "[46] Mikiya Shibuya, Shinya Sumikura, and Ken Sakurada. Privacy preserving visual SLAM. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXII 16, pages 102-118. Springer, 2020. 1, 2, 3", + "[47] Sudipta Narayan Sinha, Pablo Alejandro Speciale, Sing Bing Kang, and Marc Andre Leon Pollefeys. Device pose estimation using 3d line clouds, Mar. 30 2021. US Patent 10,964,053. 3", + "[48] Noah Snavely, Steven M Seitz, and Richard Szeliski. Modeling the world from internet photo collections. International Journal of Computer Vision (IJCV), 80:189-210, 2008. 1", + "[49] Zhenbo Song, Wayne Chen, Dylan Campbell, and Hongdong Li. Deep novel view synthesis from colored 3d point clouds. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXIV 16, pages 1-17. Springer, 2020. 2", + "[50] Pablo Speciale, Johannes L Schonberger, Sing Bing Kang, Sudipta N Sinha, and Marc Pollefeys. Privacy preserving image-based localization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5493-5503, 2019. 1, 2, 3, 4, 5, 6, 7", + "[51] Pablo Speciale, Johannes L Schonberger, Sudipta N Sinha, and Marc Pollefeys. Privacy preserving image queries for camera localization. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 1486-1496, 2019. 6", + "[52] Henrik Stewénius, Magnus Oskarsson, Kalle Aström, and David Nister. Solutions to minimal generalized relative pose problems, 2005. 3, 4", + "[53] Yurun Tian, Xin Yu, Bin Fan, Fuchao Wu, Huub Heijnen, and Vassileios Balntas. SOSNet: Second order similarity regularization for local descriptor learning. In Proceedings of" + ], + "bbox": [ + 501, + 92, + 893, + 892 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "9782", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11016-11025, 2019. 2", + "[54] Philip HS Torr and Andrew Zisserman. Mlesac: A new robust estimator with application to estimating image geometry. Computer vision and image understanding, 78(1):138-156, 2000. 4", + "[55] Julien Valentin, Angela Dai, Matthias Niessner, Pushmeet Kohli, Philip Torr, Shahram Izadi, and Cem Keskin. Learning to navigate the energy landscape. In 2016 Fourth International Conference on 3D Vision (3DV), pages 323-332, 2016. 1, 3, 5, 6, 7, 8", + "[56] Jonathan Ventura, Clemens Arth, and Vincent Lepetit. An efficient minimal solution for multi-camera motion. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), pages 747-755, 2015. 3", + "[57] Carl Vondrick, Aditya Khosla, Tomasz Malisiewicz, and Antonio Torralba. HOGgles: Visualizing object detection features. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), pages 1-8, 2013. 1, 2", + "[58] Philippe Weinzaepfel, Herve Jégou, and Patrick Pérez. Reconstructing an image from its local descriptors. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 337-344, 2011. 1, 2", + "[59] Enliang Zheng and Changchang Wu. Structure from motion using structure-less resection. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), pages 2075-2083, 2015. 2, 4" + ], + "bbox": [ + 78, + 90, + 468, + 457 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "9783", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2024/Efficient Privacy-Preserving Visual Localization Using 3D Ray Clouds/61345f4c-613d-4e27-91d1-09fcd6a52a24_model.json b/2024/Efficient Privacy-Preserving Visual Localization Using 3D Ray Clouds/61345f4c-613d-4e27-91d1-09fcd6a52a24_model.json new file mode 100644 index 0000000000000000000000000000000000000000..b280f8e5cd518098b3e39965df9bd9313ab8933e --- /dev/null +++ b/2024/Efficient Privacy-Preserving Visual Localization Using 3D Ray Clouds/61345f4c-613d-4e27-91d1-09fcd6a52a24_model.json @@ -0,0 +1,2983 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.132, + 0.131, + 0.84, + 0.154 + ], + "angle": 0, + "content": "Efficient Privacy-Preserving Visual Localization Using 3D Ray Clouds" + }, + { + "type": "text", + "bbox": [ + 0.266, + 0.18, + 0.74, + 0.2 + ], + "angle": 0, + "content": "Heejoon Moon1 Chunghwan Lee2 Je Hyeong Hong1,2†" + }, + { + "type": "text", + "bbox": [ + 0.088, + 0.199, + 0.914, + 0.218 + ], + "angle": 0, + "content": "\\(^{1}\\)Dept. Artificial Intelligence, Hanyang University \\(^{2}\\)Dept. Electronic Engineering, Hanyang University" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.237, + 0.314, + 0.254 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.271, + 0.474, + 0.65 + ], + "angle": 0, + "content": "The recent success in revealing scene details from sparse 3D point clouds obtained via structure-from-motion has raised significant privacy concerns in visual localization. One prominent approach for mitigating this issue is to lift 3D points to 3D lines thereby reducing the effectiveness of the scene inversion attacks, but this comes at the cost of increased algorithmic complexity for camera localization due to weaker geometric constraints induced by line clouds. To overcome this limitation, we propose a new lifting approach called \"ray cloud\", whereby each lifted 3D line intersects at one of two predefined locations, depicting omnidirectional rays from two cameras. This yields two benefits, i) camera localization can now be cast as relative pose estimation between the query image and the calibrated rig of two perspective cameras which can be efficiently solved using a variant of the 5-point algorithm, and ii) the ray cloud introduces erroneous estimations for the density-based inversion attack, degrading the quality of scene recovery. Moreover, we explore possible modifications of the inversion attack to better recover scenes from the ray clouds and propose a ray sampling technique to reduce the effectiveness of the modified attack. Experimental results on two public datasets show real-time localization speed as well as enhanced privacy-preserving capability over the state-of-the-art without overly sacrificing the localization accuracy." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.665, + 0.21, + 0.682 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.692, + 0.47, + 0.875 + ], + "angle": 0, + "content": "Visual localization refers to the task of estimating the camera pose w.r.t. a known 3D scene given a query image. It serves a central role for cutting-edge applications involving robotics [18], augmented reality (AR) [3], virtual reality (VR) [6] and simultaneous localization and mapping (SLAM) [11, 29, 33]. To this date, many visual localization algorithms [27, 37] still utilize sparse 3D point clouds constructed via structure-from-motion (SfM) [1, 44, 45, 48] through which the camera pose is estimated from the correspondences between 2D image features and 3D points. This is commonly solved using a perspective-n-point (PnP) solver [12, 20, 37] equipped with the random sample con" + }, + { + "type": "image", + "bbox": [ + 0.513, + 0.242, + 0.682, + 0.319 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.524, + 0.322, + 0.668, + 0.334 + ], + "angle": 0, + "content": "(a) Illustration of ray cloud" + }, + { + "type": "image", + "bbox": [ + 0.69, + 0.242, + 0.871, + 0.319 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.692, + 0.322, + 0.884, + 0.334 + ], + "angle": 0, + "content": "(b) Single-image localization speeds" + }, + { + "type": "image", + "bbox": [ + 0.515, + 0.336, + 0.614, + 0.469 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.545, + 0.469, + 0.594, + 0.477 + ], + "angle": 0, + "content": "Ground truth" + }, + { + "type": "image", + "bbox": [ + 0.615, + 0.336, + 0.704, + 0.469 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.626, + 0.469, + 0.693, + 0.477 + ], + "angle": 0, + "content": "Uniform line cloud" + }, + { + "type": "image", + "bbox": [ + 0.705, + 0.336, + 0.794, + 0.469 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.703, + 0.469, + 0.795, + 0.477 + ], + "angle": 0, + "content": "Paired-Point Lifting (PPL)" + }, + { + "type": "image", + "bbox": [ + 0.795, + 0.336, + 0.884, + 0.469 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.807, + 0.469, + 0.87, + 0.477 + ], + "angle": 0, + "content": "Ray cloud (ours)" + }, + { + "type": "image_caption", + "bbox": [ + 0.546, + 0.481, + 0.848, + 0.493 + ], + "angle": 0, + "content": "(c) Comparison of revealed scene geometries and images" + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.495, + 0.895, + 0.592 + ], + "angle": 0, + "content": "Figure 1. Main characteristics of the proposed ray cloud. (a) shows 3D points lifted as 3D rays half of which intersects at center 1 and the other half intersects at center 2. Estimating the camera pose resembles relative pose estimation with respect to a calibrated rig of perspective cameras. As shown in (b), this allows real-time single-image-based localization on two public datasets [22, 55] while preventing scene-revealing attacks [7, 40] as shown in (c)." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.598, + 0.892, + 0.628 + ], + "angle": 0, + "content": "sensus (RANSAC) meta-algorithm [9, 15, 42] or its deep network alternative such as differentiable RANSAC [4, 5]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.629, + 0.893, + 0.78 + ], + "angle": 0, + "content": "Surprisingly, it was recently shown [8, 10, 13, 30, 40, 57, 58] that a sparse 3D point cloud may contain enough information to reveal image details of the underlying scene. Most prominently, Pittaluga et al. [40] demonstrated a cascaded U-Net model [43] can generate high-fidelity images of the scene just from the 2D projection of sparse 3D points and their SIFT features [28]. This raised significant privacy concerns [39, 50] as anyone with access to the 3D point cloud could potentially search for sensitive details in the private scene by employing this scene inversion network." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.781, + 0.895, + 0.902 + ], + "angle": 0, + "content": "Currently, the mainstream approach for preventing the aforementioned image inversion attack on the sparse 3D point cloud is geometric lifting [26, 46, 50], whereby each 3D point is transformed to a 3D line passing through the original point with aims to conceal the 3D scene geometry (see Fig. 2). While the earlier approach of randomly drawing the line direction (also known as the uniform line cloud) was shown to be susceptible to geometry inversion attack [7]," + }, + { + "type": "page_footnote", + "bbox": [ + 0.102, + 0.887, + 0.227, + 0.901 + ], + "angle": 0, + "content": "† Corresponding author" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9773" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.468, + 0.167 + ], + "angle": 0, + "content": "this was remedied by recent methods utilizing pairing of 3D points [26, 36] to increase ambiguity in the lifted scene representation. Nevertheless, single image-based localization speeds for these algorithms [26, 36] are far from achieving real-time performance, providing motivation for this work." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.168, + 0.468, + 0.378 + ], + "angle": 0, + "content": "In this paper, we present a new geometric lifting approach called ray cloud with aims to address above issue of slow camera localization runtime exhibited by previous studies. As will be illustrated in Sec. 3, ray cloud is simple to construct and robust to the geometry inversion attack [7]. Most importantly, camera localization using a ray cloud can be formulated as a well-known relative pose estimation problem between the query image and a virtual calibrated rig of cameras, enabling utilization of the efficient 5-point algorithm [35] with known scale [59]. Moreover, by modifying and enhancing the geometry inversion attack [7] in Sec. 4, we explore potential hidden privacy risks associated with ray clouds, and subsequently propose a technique to enhance the privacy protection of ray clouds in Sec. 5." + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.379, + 0.455, + 0.393 + ], + "angle": 0, + "content": "The contributions of this work are summarized below:" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.394, + 0.468, + 0.424 + ], + "angle": 0, + "content": "- an efficient privacy-preserving visual localization framework based on a new lifting approach called 3D ray cloud," + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.424, + 0.468, + 0.453 + ], + "angle": 0, + "content": "- customization of the geometry inversion attack [7] to yield undesired (privacy leak) corner cases for ray clouds," + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.455, + 0.468, + 0.499 + ], + "angle": 0, + "content": "- a voxel-based ray sampling technique to prevent above privacy leaks by disrupting nearest neighbor statistic and inducing erroneous recovery of the scene points, and" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.5, + 0.468, + 0.575 + ], + "angle": 0, + "content": "- extensive experimental comparison of ray cloud against other line cloud-based approaches on two public datasets, achieving real-time single-image based localization speed for the first time while preserving privacy-preserving capability and maintaining similar localization accuracy." + }, + { + "type": "list", + "bbox": [ + 0.078, + 0.394, + 0.468, + 0.575 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.586, + 0.214, + 0.601 + ], + "angle": 0, + "content": "2. Related work" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.611, + 0.468, + 0.746 + ], + "angle": 0, + "content": "Privacy threats from image inversion attack While several prior studies addressed the possibility of revealing images from sparse keypoints [13, 19, 30, 57, 58], it was Pitaluga et al. [40] who posed a real privacy threat for the first time by recovering detailed scene images from a sparse 3D point cloud using a cascaded U-Net [43]. This network called InvSfM takes in the 2D locations of the projected 3D points as well as the corresponding depths and SIFT descriptors, and outputs a synthesized image of the scene." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.748, + 0.468, + 0.868 + ], + "angle": 0, + "content": "More recently, Song et al. [49] used a PointNet++ architecture to extract deep features and synthesize scene images. Dangwal et al. [10] devised a generic inversion model that can work with various descriptors [2, 28, 53] and proposed selective feature suppression to hide privacy-sensitive regions. Nonetheless, the pretrained InvSfM model still serves as the baseline for qualitative and quantitative analyses in privacy-preserving visual localization [7, 26, 36, 38]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.871, + 0.468, + 0.901 + ], + "angle": 0, + "content": "Privacy-preserving scene representations After the emergence of InvSfM [40], Speciale et al. [50] quickly em" + }, + { + "type": "text", + "bbox": [ + 0.497, + 0.092, + 0.892, + 0.227 + ], + "angle": 0, + "content": "braced a geometric lifting approach, also known as the uniform line cloud, where each 3D point in the original sparse point cloud is replaced by a 3D line passing through the point with its direction drawn from a uniform distribution on the unit sphere. This scene representation aims to introduce anonymity to the locations of the 3D points thereby preventing any image inversion via InvSfM. The work was later extended by Shibuya et al. [46] to visual SLAM with a prebuilt map provided in the form of uniform line cloud." + }, + { + "type": "text", + "bbox": [ + 0.497, + 0.228, + 0.892, + 0.514 + ], + "angle": 0, + "content": "Unfortunately, above methods [46, 50] are now outdated and susceptible to the geometry inversion attack proposed by Chelani et al. [7], which successfully recovers the sparse 3D point cloud from a uniform line cloud (details in the next subsection). To overcome this limitation, Lee et al. [26] proposed to draw 3D lines through random pairs of 3D points, thereby introducing non-uniform line distributions and multiple 3D point candidates for each line both of which are empirically shown to hinder point cloud inversion. Pan et al. [36] proposed to permute the coordinates of random pairs of keypoints to preserve privacy while still enabling localization from multiple axis-aligned lines. Geppert et al. [16] presented a partial localization approach whereby three 1-dimensional partial maps are constructed from the 3D points and separately stored to enhance security. While these approaches can evade the geometry inversion attack, they are not runtime-efficient to allow real-time camera localization from single images. This issue is addressed through use of the ray cloud proposed in this work." + }, + { + "type": "text", + "bbox": [ + 0.497, + 0.515, + 0.892, + 0.576 + ], + "angle": 0, + "content": "Other explored approaches include keypoint lifting [14], carving [34] or obfuscation [41], whereby the feature descriptors of the 3D points are transformed to hide keypoint details, but this does not conceal the point cloud geometry." + }, + { + "type": "text", + "bbox": [ + 0.497, + 0.581, + 0.892, + 0.702 + ], + "angle": 0, + "content": "Privacy threats for line clouds Chelani et al. [7] presented a density-based inversion attack that can reveal the 3D point geometry from a uniform line cloud [50]. While the posterior probability of the point cloud \\(\\mathcal{P}\\) given a line cloud \\(\\mathcal{L}\\) can be expressed as \\(P(\\mathcal{P}|\\mathcal{L})\\propto P(\\mathcal{P})P(\\mathcal{L}|\\mathcal{P})\\) using Bayes' rule, it is noted that the likelihood \\(P(\\mathcal{L}|\\mathcal{P})\\) is constant for uniform line clouds. Hence, maximizing the posterior amounts to maximizing the marginal prior \\(P(\\mathcal{P})\\)." + }, + { + "type": "text", + "bbox": [ + 0.497, + 0.703, + 0.892, + 0.84 + ], + "angle": 0, + "content": "For maximizing \\( P(\\mathcal{P}) \\), Chelani et al. [7] resorted to an empirical observation that for any two 3D points \\( \\mathbf{x}_A \\) and \\( \\mathbf{x}_B \\) with their corresponding uniformly-drawn lines \\( \\mathbf{l}_A \\in Gr(1,3) \\) and \\( \\mathbf{l}_B \\in Gr(1,3) \\), \\( P(\\| \\mathbf{x}_A - \\mathbf{x}_A^\\perp \\|_2 < \\| \\mathbf{x}_A - \\mathbf{x}_B \\|_2) \\approx 0.8 \\), where \\( \\mathbf{x}_A^\\perp \\) denotes the point on \\( \\mathbf{l}_A \\) closest to \\( \\mathbf{l}_B \\). This implies that if \\( \\mathbf{x}_A \\) and \\( \\mathbf{x}_B \\) are close to each other, then \\( \\mathbf{x}_A^\\perp \\) is likely to become a good approximation of \\( \\mathbf{x}_A \\). Hence, \\( P(\\mathcal{P}) \\) can be maximized if the nearest neighboring points of each 3D point are known in advance." + }, + { + "type": "text", + "bbox": [ + 0.497, + 0.84, + 0.892, + 0.901 + ], + "angle": 0, + "content": "As it is non-trivial to find a set of true \\(k\\)-nearest neighboring points in line clouds, Chelani et al. [7] approximates this by the set of \\(K\\)-nearest lines \\(\\{\\mathbf{l}_K\\}\\) for each line \\(\\mathbf{l}_i\\), expecting some to include the true positives of the \\(k\\)-nearest" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "9774" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.088, + 0.089, + 0.28, + 0.174 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.141, + 0.18, + 0.222, + 0.192 + ], + "angle": 0, + "content": "(a) Point cloud" + }, + { + "type": "image", + "bbox": [ + 0.289, + 0.089, + 0.484, + 0.174 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.311, + 0.18, + 0.458, + 0.193 + ], + "angle": 0, + "content": "(b) Uniform line cloud [50]" + }, + { + "type": "image", + "bbox": [ + 0.495, + 0.089, + 0.685, + 0.174 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.497, + 0.18, + 0.68, + 0.193 + ], + "angle": 0, + "content": "(c) Paired-Point Lifting (PPL) [26]" + }, + { + "type": "image", + "bbox": [ + 0.696, + 0.089, + 0.886, + 0.174 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.735, + 0.18, + 0.843, + 0.193 + ], + "angle": 0, + "content": "(d) Ray cloud (ours)" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.195, + 0.89, + 0.223 + ], + "angle": 0, + "content": "Figure 2. 3D line clouds yielded from different lifting approaches on the Office1 manolis in the Energy Landscape [55] dataset. Note each line (ray) in the ray cloud passes through one of two center points. We show \\(10\\%\\) of the total number of lines for visual comparison." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.232, + 0.47, + 0.381 + ], + "angle": 0, + "content": "points when \\( K > k \\) is large enough (we distinguish \\( K \\) for lines and \\( k \\) for points). Bearing this assumption, [7] computes the closest point \\( \\mathbf{x}_{k}^{\\perp} \\) to each of the \\( K \\)-nearest lines for the current line \\( \\mathbf{l}_i \\), builds a histogram of closest points \\( \\{\\mathbf{x}_{k}^{\\perp}\\} \\) and finds the peak of the histogram via Kuiper's test [23] to estimate the point \\( \\mathbf{x}_i \\). This attack is effective for uniform line clouds [50], allowing recovered points to be used for scene recovery with InvSfM [40]. However, it is less effective against more recent line-based approaches [26, 36] and ray clouds presented in this work." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.393, + 0.471, + 0.665 + ], + "angle": 0, + "content": "Camera localization using line clouds As mentioned in Sec. 1, camera localization typically involves a RANSAC algorithm [15] coupled with an efficient minimal solver that can run fast over thousands of iterations [25]. For single-image based localization, a standard line cloud introduces 2D points-to-3D line constraints, which are weaker than 2D point-to-3D point constraints provided by the point clouds. Consequently, the minimal solver for line clouds (p6L) requires 6 correspondences between 2D points and 3D lines [50]. Since p6L innately uses the solver for generalized relative pose estimation [52], achieving real-time speed with \\( p6L \\) is challenging due to the large number of possible solutions (up to 64) and high algorithmic complexity [52, 56]. Shibuya et al. [46] achieved real-time localization with line clouds in the SLAM setup, but this requires a video stream as input and construction of a client-side point map. On the other hand, we will show it is possible to perform image-based localization in real-time using ray clouds." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.682, + 0.216, + 0.699 + ], + "angle": 0, + "content": "3. 3D Ray cloud" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.709, + 0.469, + 0.755 + ], + "angle": 0, + "content": "We now illustrate the process of constructing a 3D ray cloud shown in Fig. 2 and estimating camera pose from ray clouds. We also discuss inversion attack [7] on ray clouds." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Motivation The concept of lifting 3D lines to intersect at a particular point has been briefly addressed without detailed investigation in [7, 47]. Chelani et al. [7] mentioned that, if all lines were to intersect at a single point, then the geometry inversion attack in Sec. 2 would fail. This is because the histogram of closest points (to other lines) becomes concentrated about the point of intersection, leading to a degenerate solution. Unfortunately, this representation possesses an intrinsic limitation that the camera pose can" + }, + { + "type": "image", + "bbox": [ + 0.515, + 0.23, + 0.637, + 0.339 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.514, + 0.342, + 0.633, + 0.353 + ], + "angle": 0, + "content": "(a) Original 3D points" + }, + { + "type": "image", + "bbox": [ + 0.638, + 0.23, + 0.757, + 0.339 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.638, + 0.342, + 0.756, + 0.353 + ], + "angle": 0, + "content": "(b) Center assignment" + }, + { + "type": "image", + "bbox": [ + 0.76, + 0.23, + 0.884, + 0.339 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.784, + 0.342, + 0.857, + 0.353 + ], + "angle": 0, + "content": "(c) Ray cloud" + }, + { + "type": "image_caption", + "bbox": [ + 0.5, + 0.355, + 0.89, + 0.383 + ], + "angle": 0, + "content": "Figure 3. Illustration of the procedure for selecting points of intersection and constructing a 3D ray cloud." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.392, + 0.892, + 0.483 + ], + "angle": 0, + "content": "only be recovered up to scale as camera localization resembles relative pose estimation [35] between the query camera and the intersecting 3D lines. We aim to overcome this issue by utilizing two intersection points. This can be viewed as adding another camera to the scene, forming a calibrated rig of perspective cameras to resolve the scale ambiguity." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.493, + 0.72, + 0.509 + ], + "angle": 0, + "content": "3.1. Construction procedure" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.516, + 0.892, + 0.713 + ], + "angle": 0, + "content": "Lifting 3D points to a ray cloud comprises three steps, namely setting two ray centers and assigning 3D points to one of two ray centers. We choose to set the ray centers by applying the \\(K\\)-means clustering algorithm on the point cloud with \\(K = 2\\), from which the centers of two clusters are set as the ray centers (explained below). Then, we assign half of randomly sampled points from the point cloud to center \\(1(\\mathbf{c}_1 \\in \\mathbb{R}^3)\\), and the other half to center \\(2(\\mathbf{c}_2 \\in \\mathbb{R}^3)\\). Note this assignment is not related to the \\(K\\)-means clustering. Finally, each point \\(\\mathbf{x}_i\\) is lifted to a line as \\(\\mathbf{c}_j + \\lambda (\\mathbf{x}_i - \\mathbf{c}_j)\\) for \\(\\lambda \\in \\mathbb{R}\\), where \\(j \\in [1,2]\\) is the assigned center index for point \\(i\\). The final output resembles omnidirectional rays from two perspective cameras (see Fig. 3)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.717, + 0.892, + 0.838 + ], + "angle": 0, + "content": "Effect of changing ray center locations If the ray centers are located very close to each other, then they begin to resemble a single ray center, increasing ambiguity in translation scale. On the other hand, if the ray centers are distant from each other and the scene, then the resulting rays are nearly parallel which degrades localization accuracy. We empirically observe that using the centers from \\( K \\)-means clustering usually avoids both extreme cases (see [32])." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.848, + 0.72, + 0.864 + ], + "angle": 0, + "content": "3.2. Camera pose estimation" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.892, + 0.901 + ], + "angle": 0, + "content": "We now illustrate an efficient localization approach for ray clouds that can benefit from faster inference speed com" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9775" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.111, + 0.091, + 0.257, + 0.166 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.107, + 0.169, + 0.262, + 0.181 + ], + "angle": 0, + "content": "(a) Standard line cloud \\((p6\\mathrm{L})\\)" + }, + { + "type": "image", + "bbox": [ + 0.296, + 0.091, + 0.452, + 0.164 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.304, + 0.169, + 0.426, + 0.181 + ], + "angle": 0, + "content": "(b) Ray cloud \\((p5 + 1\\mathrm{R})\\)" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.186, + 0.47, + 0.325 + ], + "angle": 0, + "content": "Figure 4. Illustration of minimal pose estimation problems solved by different line representations. Camera localization using a uniform line cloud can be seen as the special case of generalized relative pose estimation problem [52] involving one perspective camera (query) and one generalized camera (line cloud). On the other hand, camera localization using a ray cloud can be regarded as the case of perspective relative pose estimation problem between the query image and a rig of calibrated cameras (ray cloud). The latter enjoys shorter runtime as shown in Sec. 6 partly due to stronger ray constraints reducing the number of algebraic solutions." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.335, + 0.468, + 0.456 + ], + "angle": 0, + "content": "pared to other line-based scene representations [26, 36, 50]. This is inspired by the illustration in Fig. 4b that the ray cloud can be viewed as two virtual perspective cameras located at the ray centers \\(\\mathbf{c}_1 \\in \\mathbb{R}^3\\) and \\(\\mathbf{c}_2 \\in \\mathbb{R}^3\\). Since the rays from these cameras are omnidirectional, we can set their rotation matrices \\(\\mathsf{R}_1, \\mathsf{R}_2 \\in SO(3)\\) to be aligned with world coordinates, i.e. \\([\\mathsf{R}_1|\\mathbf{t}_1] = [\\mathsf{I}| - \\mathbf{c}_1]\\) and \\([\\mathsf{R}_2|\\mathbf{t}_2] = [\\mathsf{I}| - \\mathbf{c}_2]\\), where \\(\\mathbf{t}_1, \\mathbf{t}_2 \\in \\mathbb{R}^3\\) are the camera translations." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.457, + 0.468, + 0.592 + ], + "angle": 0, + "content": "Below, we show that having 5 correspondences between 2D keypoints from the query image and 3D rays from a randomly selected virtual camera and one correspondence between a 2D query keypoint and a 3D ray from a remaining virtual camera allows us to estimate the absolute camera pose using a ray cloud. We denote this adaptation of minimal \\(5 + 1\\) solver [59] as the perspective-\\((5 + 1)\\)-rays \\((p5 + 1\\mathrm{R})\\) solver. Similar to previous studies [26, 50], we assume the camera intrinsics \\(\\mathbf{K} \\in \\mathbb{R}^{3 \\times 3}\\) are known." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.6, + 0.468, + 0.721 + ], + "angle": 0, + "content": "\\(p5 + 1\\mathbf{R}\\) solver We start by estimating the relative pose between the query image and the first virtual camera using the 5-point algorithm [35]. It yields multiple candidates for the relative pose of the query image w.r.t. the first virtual camera defined as \\([\\mathtt{R}_{q1}|\\alpha \\hat{\\mathbf{t}}_{q1}]\\), where \\(\\mathtt{R}_{q1}\\in SO(3)\\) is the relative rotation, \\(\\hat{\\mathbf{t}}_{q1}\\in S^2\\) is the normalized relative translation and \\(\\alpha \\in \\mathbb{R}\\) is an unknown scale factor. Subsequently, the absolute pose of query camera is \\([\\mathtt{R}_{q1}|\\alpha \\hat{\\mathbf{t}}_{q1} - \\mathtt{R}_{q1}\\mathbf{c}_1]\\)." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.722, + 0.468, + 0.813 + ], + "angle": 0, + "content": "In order to disambiguate \\(\\alpha\\), we utilize the additional correspondence between a normalized query keypoint \\(\\mathbf{p} := (p_1, p_2)^\\top \\in \\mathbb{R}^2\\) and a 3D ray \\(\\mathbf{l} \\in Gr(1, 3)\\) from the second virtual camera. We note the ray from the query camera center passing through \\(\\mathbf{p}\\) would ideally intersect \\(\\mathbf{l}\\). Expressing both rays in the query-centered coordinates yields" + }, + { + "type": "equation", + "bbox": [ + 0.135, + 0.824, + 0.468, + 0.844 + ], + "angle": 0, + "content": "\\[\n\\lambda_ {1} \\left[ p _ {1}, p _ {2}, 1 \\right] ^ {\\top} = \\lambda_ {2} \\mathrm {R} _ {q 1} \\hat {\\mathbf {n}} + \\alpha \\hat {\\mathbf {t}} _ {q 1} - \\mathrm {R} _ {q 1} \\mathbf {c} _ {1} \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.855, + 0.468, + 0.9 + ], + "angle": 0, + "content": "where \\(\\hat{\\mathbf{n}}\\in S^2\\) is the line direction of 1 in world coordinates and \\(\\lambda_1,\\lambda_2\\in \\mathbb{R}\\) are scale parameters. Since we have 3 equations with 3 unknowns, \\(\\alpha\\) can be determined." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.198 + ], + "angle": 0, + "content": "While the \\( p5 + 1\\mathrm{R} \\) solver requires 6 correspondences like \\( p6\\mathrm{L} \\) [26, 50], the utilized 5-point algorithm only yields up to 10 solutions [35] that is smaller than 64 solutions from the generalized solver [52] in \\( p6\\mathrm{L} \\). This significantly reduces the computational complexity for determining the correct configuration. We also enforce the chirality constraint [17] on the query image to discard infeasible solutions." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.201, + 0.892, + 0.277 + ], + "angle": 0, + "content": "Robust estimation We equip above \\( p5 + 1R \\) solver with the LO-RANSAC [31] algorithm for robust camera localization. For selecting the best model candidate, we adopt an MLESAC [54]-type cost function based on the squared sum of epipolar distances in the query image defined as" + }, + { + "type": "equation", + "bbox": [ + 0.578, + 0.287, + 0.891, + 0.331 + ], + "angle": 0, + "content": "\\[\n\\sum_ {(i, j) \\in \\Omega} \\rho \\left(\\frac {\\left([ \\mathbf {p} _ {i} ^ {\\top} , 1 ] \\mathrm {E} _ {j} \\hat {\\mathbf {n}} _ {i}\\right) ^ {2}}{\\left(\\mathbf {e} _ {j 1} ^ {\\top} \\hat {\\mathbf {n}} _ {i}\\right) ^ {2} + \\left(\\mathbf {e} _ {j 2} ^ {\\top} \\hat {\\mathbf {n}} _ {i}\\right) ^ {2}}\\right), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.344, + 0.892, + 0.495 + ], + "angle": 0, + "content": "where \\(\\rho(s) := \\min(s, \\tau^2)\\) is a robust kernel with upper bound of threshold \\(\\tau^2\\), \\(\\Omega\\) denotes the pairing arrangement of which keypoint \\(i\\) matches with ray \\(i\\) from center \\(j\\), \\(\\mathbf{p}_i \\in \\mathbb{R}^2\\) is the \\(i\\)-th normalized keypoint in the query image, \\(\\hat{\\mathbf{n}}_i \\in S^2\\) is ray \\(i\\) passing through the ray center \\(j \\in [1, 2]\\), and \\(\\mathbf{E}_j := [\\mathbf{e}_{j1}, \\mathbf{e}_{j2}, \\mathbf{e}_{j3}]^\\top\\) is the essential matrix between the virtual camera \\(j\\) and the query camera. Since the two virtual cameras share the same rotation and only differ by translation, \\(\\mathbf{E}_2\\) can be easily computed from \\(\\mathbf{E}_1\\). The best solution from RANSAC is refined by iteratively minimizing (2)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.505, + 0.847, + 0.52 + ], + "angle": 0, + "content": "3.3. Geometry inversion attack on ray clouds" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.528, + 0.891, + 0.603 + ], + "angle": 0, + "content": "We show that applying the geometry inversion attack [7] naively to a ray cloud yields a degenerate solution. For this purpose, we make use of some prior knowledge regarding the probabilistic viewpoint and algorithmic details of the geometry inversion attack [7] reviewed in Sec. 2." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.603, + 0.892, + 0.679 + ], + "angle": 0, + "content": "Let \\(\\mathbf{x}_i\\in \\mathbb{R}^3\\) be the \\(i\\)-th point and \\(\\mathbf{l}_i\\in Gr(1,3)\\) be its lifted ray. The point cloud is defined as \\(\\mathcal{P}\\coloneqq \\{\\mathbf{x}_i\\}\\) and the ray cloud is defined as \\(\\mathcal{L}\\coloneqq \\{\\mathbf{l}_i\\}\\). Since ray centers \\(\\mathcal{C}\\coloneqq \\{\\mathbf{c}_1,\\mathbf{c}_2\\}\\) are easily found by inspecting the intersection points, we aim to maximize the conditional posterior of \\(\\mathcal{P}\\)," + }, + { + "type": "equation", + "bbox": [ + 0.584, + 0.692, + 0.891, + 0.709 + ], + "angle": 0, + "content": "\\[\nP (\\mathcal {P} | \\mathcal {L}, \\mathcal {C}) \\propto P (\\mathcal {L} | \\mathcal {P}, \\mathcal {C}) P (\\mathcal {P} | \\mathcal {C}). \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.892, + 0.81 + ], + "angle": 0, + "content": "Note the conditional likelihood \\( P(\\mathcal{L}|\\mathcal{P},\\mathcal{C}) \\) is a non-zero constant as long as each ray intersects the corresponding pair of ray center and 3D point and the probability of each ray being assigned to one of the ray centers is equally likely. As long as each point \\( \\mathbf{x}_i \\) is estimated to be on its lifted ray \\( \\mathbf{l}_i \\), then maximizing (3) amounts to maximizing \\( P(\\mathcal{P}|\\mathcal{C}) \\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.811, + 0.892, + 0.9 + ], + "angle": 0, + "content": "However, since \\( P(\\mathcal{P}|\\mathcal{C}) \\) can depend significantly on the ray centers \\( \\mathcal{C} \\), it seems difficult to yield a point statistic similar to \\( P(\\mathcal{P}) \\) in [7]. We implicitly figure out the difference between \\( P(\\mathcal{P}|\\mathcal{C}) \\) and the original prior \\( P(\\mathcal{P}) \\) by applying the geometry inversion attack [7], which maximizes \\( P(\\mathcal{P}) \\), as-is to the ray cloud. As visualized in Fig. 5a, the estimated" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "9776" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.087, + 0.087, + 0.272, + 0.171 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.172, + 0.266, + 0.185 + ], + "angle": 0, + "content": "(a) Naive neighbor estimation [7]" + }, + { + "type": "image", + "bbox": [ + 0.278, + 0.087, + 0.465, + 0.171 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.28, + 0.172, + 0.456, + 0.185 + ], + "angle": 0, + "content": "(b) Modified neighbor estimation" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.189, + 0.47, + 0.274 + ], + "angle": 0, + "content": "Figure 5. Illustration of peak-finding [7] results from (a) naive neighbor-lines estimation [7] and (b) modified neighbor-lines estimation. (a) indicates that the estimated peak will always be located at the center of the current ray. In contrast, our modification only samples the neighborhood of lines using the bundle of rays originating from the opposite center, eliminating the trivial solution." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.283, + 0.469, + 0.375 + ], + "angle": 0, + "content": "point for each ray becomes the center of the constituent ray, yielding a trivial recovery. This is because, in the process of the geometric inversion, the \\(K\\)-nearest neighboring lines include a bundle of rays originating from the same center, which yields an erroneous peak at the ray center on the histogram of point candidates, hindering general recovery." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.388, + 0.433, + 0.405 + ], + "angle": 0, + "content": "4. Adapted inversion attack for ray clouds" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.414, + 0.469, + 0.474 + ], + "angle": 0, + "content": "In this section, we illustrate two main reasons for trivial recovery of 3D points in Sec. 3.3, namely i) lower significance of nearest points for estimating 3D points, and ii) bad estimation of true nearest rays lifted from nearest points." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.474, + 0.469, + 0.625 + ], + "angle": 0, + "content": "First, the assumption in [7] that the near-80% of candidates which are closest to lines lifted from \\( k \\)-nearest neighboring points yields accurate points does not hold for ray clouds as some of the rays from these \\( k \\)-nearest points will yield estimates at the current ray center. Second, the assumption in [7] that \\( K \\)-nearest lines will overlap enough with the set of rays lifted from the \\( k \\)-nearest points (\\( K > k \\)) is weakened as the \\( K \\)-nearest lines will be rays sharing the same center whose original points may be far away from the original point of the current lifted ray." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.626, + 0.469, + 0.702 + ], + "angle": 0, + "content": "In Sec. 4, we attempt to address these issues by considering known ray centers in order to approximately maximize \\( P(\\mathcal{P}|\\mathcal{C}) \\) instead of \\( P(\\mathcal{P}) \\). We also aim to address trivial solutions by incorporating the known positions of the ray centers \\( \\mathcal{C} \\) to approximate maximizing over \\( P(\\mathcal{P}|\\mathcal{C}) \\)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.711, + 0.387, + 0.727 + ], + "angle": 0, + "content": "4.1. Eliminating the degenerate solution" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.735, + 0.469, + 0.902 + ], + "angle": 0, + "content": "The trivial recovery of 3D points (estimated 3D points \\(\\mathcal{P}\\) always located at the two ray centers \\(\\mathcal{C}\\)) is due to rays intersecting the same ray center always ending up as the nearest neighboring (NN) lines, producing the peak density of closest points at the ray center (see Fig. 5a). To bypass this issue, we exclude the rays intersecting through the current ray when constructing the set of nearest neighboring lines for estimating the peak density, and only select the NN-lines from the other ray center. This can be viewed as incorporating the known locations of the ray centers for recovering the points (see Fig. 5b) to maximize \\(P(\\mathcal{P}|\\mathcal{C})\\)." + }, + { + "type": "image", + "bbox": [ + 0.511, + 0.088, + 0.633, + 0.149 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.151, + 0.629, + 0.164 + ], + "angle": 0, + "content": "(a) ULC [50] (oracle)" + }, + { + "type": "image", + "bbox": [ + 0.64, + 0.089, + 0.764, + 0.148 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.641, + 0.15, + 0.758, + 0.163 + ], + "angle": 0, + "content": "(b) Ray cloud (oracle)" + }, + { + "type": "image", + "bbox": [ + 0.767, + 0.09, + 0.888, + 0.147 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.789, + 0.15, + 0.862, + 0.163 + ], + "angle": 0, + "content": "(c) Ray cloud" + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.168, + 0.892, + 0.321 + ], + "angle": 0, + "content": "Figure 6. 3D point cloud recovered on different settings of nearest neighbors for the Apt1 living from [55] (black-original points, red-recovered points). In (b) and (c), we have only utilized neighboring rays connected to the opposite ray center to avoid degenerate solutions (in Sec. 4.1). For the oracle cases of (a) and (b), rays from the true nearest points of the original points are used to estimate the points, showing feasible recovery. In contrast, when the nearest neighboring points are approximated by the nearest neighboring rays, it leads to worse recovery as shown in (c). This implies false positive rays, which are close but whose original points are distant, significantly degrade the recovery of points." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.332, + 0.892, + 0.468 + ], + "angle": 0, + "content": "Fig. 6b shows that, in the oracle case when the ground truth nearest points of the current ray's original 3D point are known, discarding NN-lines intersecting the current ray can avoid degenerate solution and yield more feasible point cloud. At the same time, Fig. 6c shows that the actual point recovery from a ray cloud is worse. We conjecture this is due to the misalignment between the set of nearest neighboring rays and the set of nearest neighboring points for a given point and its lifted ray which is addressed in Sec. 4.2." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.481, + 0.877, + 0.497 + ], + "angle": 0, + "content": "4.2. Rejecting rays closely aligned to the baseline" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.506, + 0.891, + 0.612 + ], + "angle": 0, + "content": "We first analyze the cause of the bad recovery in Fig. 6c even after applying the trick in Sec. 4.1. We define the rays close to the current ray but whose original points are far away from the original point of the current ray as the false positive rays. By comparing the oracle case (b) and the normal case (c) in Fig. 6, we anticipate that the false positive rays are the culprits for degraded estimation of the points." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.613, + 0.892, + 0.824 + ], + "angle": 0, + "content": "We conjecture two cases where the false positive rays can lead to misleading estimations. First, when there are many rays from the opposite center that are in the proximity of the current ray center, the peak of the histogram may be located near the current ray center. Vice versa, when the current ray passes through the denser area around the opposite center, the peak will be located near the opposite center. These two cases partially cause the 3D points to shrink (see Fig. 6c). Second, when there are many diverse rays in the proximity of the current ray, multiple peaks can be formed from the point candidates (closest points on the ray to neighboring rays) along the current ray. This can arise when the ray centers are embedded in a room-like environment to form congested rays (see [32] for more discussions)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.826, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Out of above cases, we attempt to reduce the shrinking effect of false positive rays by incorporating known \\(\\mathcal{C}\\) and not considering any rays from the opposite center which are within the proximity of the current ray center. This can be effectively achieved by excluding a fraction \\((\\beta)\\) of rays most" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9777" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.08, + 0.092, + 0.462, + 0.196 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.105, + 0.2, + 0.442, + 0.213 + ], + "angle": 0, + "content": "(a) Visualization of rejection-sampling of rays near the baseline" + }, + { + "type": "image", + "bbox": [ + 0.096, + 0.215, + 0.268, + 0.282 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.285, + 0.215, + 0.457, + 0.282 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.093, + 0.287, + 0.265, + 0.3 + ], + "angle": 0, + "content": "(b) Peak-fiding [7] w/o rejection" + }, + { + "type": "image_caption", + "bbox": [ + 0.279, + 0.287, + 0.46, + 0.3 + ], + "angle": 0, + "content": "(c) Peak-finding [7] with rejection" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.301, + 0.469, + 0.357 + ], + "angle": 0, + "content": "Figure 7. Illustration of the effect of line rejection method. For a synthetic room comprising \\(20\\mathrm{k}\\) points, in (b), peak estimation fails to estimate the correct point due to false positive rays. In (c), the rejection excludes false candidates to estimate an accurate point." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.364, + 0.47, + 0.409 + ], + "angle": 0, + "content": "aligned with the baseline when estimating the histogram of point candidates. Fig. 7 illustrates the effect of this ray rejection which can lead to better point estimation." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.41, + 0.469, + 0.501 + ], + "angle": 0, + "content": "We outline some limitations of this approach. First, if the ray lifted from a 3D point is within the region of rejection sampling, then the estimated point can be inaccurate due to the lack of nearby rays to support the estimation process. Second, the method introduces a hyperparameter \\((\\beta)\\) that needs to be carefully set (see results in Table 2)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.513, + 0.317, + 0.531 + ], + "angle": 0, + "content": "5. Voxel-based ray sampling" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.539, + 0.469, + 0.568 + ], + "angle": 0, + "content": "We now present a new ray sampling technique that can neutralize the modified inversion attack from Sec. 4." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.569, + 0.469, + 0.659 + ], + "angle": 0, + "content": "Our approach is motivated from Fig. 6 that the quality of point recovery depends on finding the true point candidates once the degenerate solution is eliminated. This time, we try to discourage the true nearest neighboring points from being included in the set of \\( K \\)-nearest neighboring lines when lifted to rays with aims to lower the inversion accuracy." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.66, + 0.47, + 0.901 + ], + "angle": 0, + "content": "For the above purpose, we partition the 3D points into voxels and assign one of two ray centers to each voxel such that i) 3D points within the same voxel are lifted to rays intersecting through the assigned ray center and ii) no neighboring voxels share the same ray center (see Fig. 8). As long as the voxels are large enough, many 3D points except those near the voxel boundaries will have most of their respective \\( k \\)-nearest points to be inside the same voxel and subsequently share the same ray center when lifted to rays. Since the rays lifted from these true \\( k \\)-nearest neighboring points inside the same voxel will yield a trivial solution as illustrated in Sec. 4.1, these will be disallowed from being used for peak estimation. Hence, we essentially minimize the overlap between the set of \\( K \\)-nearest lines and the set of \\( k \\)-nearest points which is a necessary condition for successful recovery of points from lines in [7]." + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.094, + 0.619, + 0.153 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.534, + 0.153, + 0.587, + 0.163 + ], + "angle": 0, + "content": "Point cloud" + }, + { + "type": "image", + "bbox": [ + 0.622, + 0.096, + 0.688, + 0.151 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.632, + 0.153, + 0.682, + 0.164 + ], + "angle": 0, + "content": "Voxel-grid" + }, + { + "type": "image", + "bbox": [ + 0.697, + 0.095, + 0.812, + 0.153 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.716, + 0.153, + 0.795, + 0.164 + ], + "angle": 0, + "content": "Grid assignment" + }, + { + "type": "image", + "bbox": [ + 0.815, + 0.094, + 0.891, + 0.153 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.839, + 0.153, + 0.869, + 0.164 + ], + "angle": 0, + "content": "Lifting" + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.166, + 0.892, + 0.249 + ], + "angle": 0, + "content": "Figure 8. The procedure of voxel-based ray sampling. After partitioning the point cloud into a voxel grid, the whole grid is divided into two clusters in such a way that neighboring grids belong to different clusters. Then, all 3D points in each grid within the same cluster are paired with one center, while the other cluster is paired with the remaining center for lifting." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.259, + 0.892, + 0.41 + ], + "angle": 0, + "content": "We adopt a 3D checkerboard pattern for assigning ray centers to voxels with aims to achieve stable camera localization accuracy by uniformly distributing rays from different ray centers across the scene. While a larger voxel size will allow more true nearest neighboring points to be discarded and provide better privacy protection, it will also cause the rays to be more non-uniform and lower the camera localization accuracy (see [32]). Currently, the voxel size needs to be manually determined for each dataset as the optimal trade-off point can vary across different scenes." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.424, + 0.706, + 0.44 + ], + "angle": 0, + "content": "6. Experimental Results" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.449, + 0.892, + 0.495 + ], + "angle": 0, + "content": "We present comprehensive experimental results showing the effectiveness of ray cloud in terms of localization accuracy, inference speed, and privacy-preserving performance." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.499, + 0.893, + 0.68 + ], + "angle": 0, + "content": "Datasets We used the same two public datasets as in [7, 26]. The Energy Landscape [55] dataset consists of 12 indoor scenes with multiple sequences of scanned RGB images. The Cambridge [21, 22] dataset comprises 6 distinct outdoor landmarks, containing multiple image frames with labeled 6-DOF camera poses for the camera re-localization. As in [26], we excluded the scenes of Street and Great Court from [22] due to the presence of large geometric outliers. We followed the same experimental protocols from [26, 50] for constructing sparse 3D point clouds using COLMAP [44] and evaluation. Separated queries not included in 3D map construction were used for evaluation." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.685, + 0.892, + 0.79 + ], + "angle": 0, + "content": "Implementation details We used Poselib [24] to utilize the \\( p5 + 1\\mathrm{R} \\) solver and the repository of PPL [26] to run experiments. We report coarse estimation results of inversion in ray cloud and voxel-based ray cloud as the refine estimation [7] diverges due to incorrect initial estimation (see [32]). We used a workstation with an Intel CPU i9-13900K running at 3.0 GHz and NVIDIA RTX 4090 GPUs." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.796, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Evaluation metrics For camera localization, we estimated the absolute 6-DOF pose of the query image with given 2D-3D correspondences. We define \\(\\mathsf{R} \\in SO(3)\\) and \\(\\mathbf{t} \\in \\mathbb{R}^3\\) as the ground truth camera rotation and translation respectively, and \\(\\hat{\\mathbf{R}} \\in SO(3)\\) and \\(\\hat{\\mathbf{t}} \\in \\mathbb{R}^3\\) as the estimated rotation and translation respectively. Then, we obtained the estimated pose errors by following [26, 50, 51], where the" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9778" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.089, + 0.085, + 0.88, + 0.272 + ], + "angle": 0, + "content": "
DatasetScene representationMinimal solverPose errorRANSACPrivacy preserving
ΔR↓meanmedianΔt↓meanmediani meanr meant ↓meaneg↑ medianPSNR ↓ meanSSIM ↓ meanMAE ↑ mean
Energy LandscapePoint cloudp3P0.1780.0310.0070.0011199.544-15.760.51633.29
Uniform line cloud [50]p6L0.2720.0980.0100.00312481.12900.04513.450.40342.83
PPL [26]p6L0.2920.1130.0110.00412381.51880.05610.420.35562.61
PPL+ [26]p6L0.2930.1150.0110.00412581.30850.05610.390.35562.70
Ray cloud (ours)p5+1R0.2620.0870.0100.00314680.0360.66910.290.31962.97
Voxel-based ray cloud (ours)p5+1R0.2970.1160.0110.00414480.1560.77010.050.31564.84
CambridgePoint cloudp3P0.7000.1010.1420.0431298.487-14.000.43339.40
Uniform line cloud [50]p6L0.7750.1690.1720.06643671.872890.74512.710.33146.08
PPL [26]p6L0.7750.1830.1810.08126474.491860.97711.210.29655.53
PPL+ [26]p6L0.7640.1830.1780.08128474.111900.94811.160.29755.91
Ray cloud (ours)p5+1R0.8550.2220.2190.10723974.89129.61410.410.26160.98
Voxel-based ray cloud (ours)p5+1R0.8800.2560.2350.12025075.241210.5410.230.25362.56
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.275, + 0.894, + 0.332 + ], + "angle": 0, + "content": "Table 1. Comparison of localization and privacy-preserving performance with different 3D scene representations (\\(\\Delta R\\): rotation error \\([\\circ]\\), \\(\\Delta t\\): translation error [m], i: number of iterations, r: inlier ratio (\\%), t: runtime [ms], \\(e_g\\): geometric (3D point) error [m]). We applied modified inversion attack from Sec. 4 with 25% line rejection to ray cloud and voxel-based ray cloud with voxel size of 0.5m and 5.0m for Energy Landscape and Cambridge respectively. Bold indicate the best result in each metric among privacy-preserving representations." + }, + { + "type": "image", + "bbox": [ + 0.094, + 0.334, + 0.884, + 0.481 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.091, + 0.482, + 0.877, + 0.496 + ], + "angle": 0, + "content": "Figure 9. Comparison of revealed scene geometries and images for Apt2 bed [55]. Modified inversion attack is applied to (d) and (e)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.503, + 0.471, + 0.54 + ], + "angle": 0, + "content": "rotational error is computed as \\(\\Delta \\mathsf{R} = \\arccos \\frac{\\mathrm{Tr}(\\hat{\\mathsf{R}}^{\\top}\\mathsf{R}) - 1}{2}\\) and the translation error computed as \\(\\Delta \\mathbf{t} = \\left\\| \\hat{\\mathsf{R}}^{\\top}\\hat{\\mathbf{t}} -\\mathsf{R}^{\\top}\\mathbf{t}\\right\\|_{2}\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.54, + 0.47, + 0.66 + ], + "angle": 0, + "content": "To compare the quality of recovered point clouds, we calculated the 3D median point errors between the original and estimated points in meters as in [26]. We also compared the quality of revealed images by feeding recovered point cloud and ground truth camera pose of query images to InvSfM [39] and calculating peak signal-to-noise ratio (PSNR), structural similarity index measure (SSIM), and mean absolute error (MAE) values against the queries." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.667, + 0.442, + 0.683 + ], + "angle": 0, + "content": "6.1. Camera localization accuracy and runtime" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.69, + 0.47, + 0.811 + ], + "angle": 0, + "content": "As shown in Table 1, our ray cloud family of approaches achieved the fastest inference speed among the line-lifting methods with no significant decline in pose accuracy. Notably, these achieved real-time performance on both datasets, reaching 137 FPS for the ray cloud and 136 FPS for the voxel-based ray cloud on average. As further demonstrated in [32], this is mainly due to the introduction of the \\( p5 + 1\\mathrm{R} \\) solver which is only available for ray clouds." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.811, + 0.471, + 0.902 + ], + "angle": 0, + "content": "The ray cloud showed the lowest localization error on [55] but slightly higher errors than ULC and PPL on [22]. We conjecture this is due to the increased baseline between two ray centers for the outdoor scenes. We also note that the voxel-based ray cloud does not exhibit a significant reduction in pose error compared to the standard ray cloud." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.504, + 0.892, + 0.521 + ], + "angle": 0, + "content": "6.2. Quality of revealed scene geometry and images" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.528, + 0.893, + 0.65 + ], + "angle": 0, + "content": "The quantitative comparisons of scene reconstruction quality are also shown in Table 1. For a fair comparison, we applied our adapted inversion attack in Sec. 4 for ray cloud and voxel-based ray cloud, while we used the original geometry inversion attack [7] with 3 refinement steps for uniform line cloud (ULC), PPL and \\(\\mathrm{PPL + }\\) .We set \\(\\beta = 25\\%\\) for the ray rejection threshold as it shows the best point estimation in Table 2. Results with \\(\\beta = 50\\%\\) are included in [32]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.654, + 0.894, + 0.79 + ], + "angle": 0, + "content": "Recovery of 3D point clouds The median point errors for ray cloud and voxel-based ray cloud are notably larger than those of other methods, even after applying our adjusted density-based attacks. Additionally, to validate the impact of our adapted inversion attack, we compared the geometric error with and without rejection sampling, showing that rejection sampling contributes to error reduction in Table 2. The qualitative result in Fig. 9 shows that the results from ray cloud and voxel-based ray cloud are unrecognizable." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.796, + 0.894, + 0.902 + ], + "angle": 0, + "content": "Degraded quality of scene details Fig. 9 shows the images recovered from the ray cloud and the voxel-based ray cloud are hardly recognizable. In Table. 1, the voxel-based ray cloud yields the lowest PSNR, SSIM and MAE values with the original ray cloud as a runner-up. This indicates large point errors exhibited by ray clouds translate to lower image quality." + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9779" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.102, + 0.092, + 0.353, + 0.171 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.364, + 0.092, + 0.613, + 0.171 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.624, + 0.092, + 0.876, + 0.171 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.096, + 0.172, + 0.351, + 0.251 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.358, + 0.172, + 0.613, + 0.252 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.621, + 0.172, + 0.876, + 0.252 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.098, + 0.253, + 0.351, + 0.331 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.361, + 0.253, + 0.612, + 0.331 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.623, + 0.253, + 0.876, + 0.331 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.332, + 0.892, + 0.358 + ], + "angle": 0, + "content": "Figure 10. Average localization performance vs. reconstructed image quality across different levels of the ray density on the Energy Landscape [55] dataset. The tested line densities are \\(100\\%\\) (biggest marker), \\(50\\%\\), \\(25\\%\\) and \\(10\\%\\) (smallest marker)." + }, + { + "type": "table", + "bbox": [ + 0.092, + 0.364, + 0.453, + 0.49 + ], + "angle": 0, + "content": "
DatasetRejection threshold (β)eg[m] medianPSNR meanSSIM meanMAE mean
Energy LandscapeNone1.0279.9590.33066.06
10 (%)0.76110.240.32363.62
25 (%)0.66910.290.31962.97
50 (%)0.74610.320.32062.52
CambridgeNone11.4710.100.27364.06
10 (%)10.2910.350.26761.60
25 (%)9.61410.410.26160.98
50 (%)12.989.9440.24264.70
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.491, + 0.47, + 0.547 + ], + "angle": 0, + "content": "Table 2. Effect of ray rejection (see Sec. 4.2) on the quality of scene recovery. Rejecting rays close to the baseline of two ray centers overall reduces the point estimation errors. Bold represents the best inversion results amongst different thresholds." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.557, + 0.469, + 0.708 + ], + "angle": 0, + "content": "Additionally, Fig. 10 comparing the pose estimation errors against image reconstruction qualities demonstrate that the ray cloud-based approaches are capable of preserving the privacy especially for dense scenes while significantly decreasing the inference runtime. In contrast, the voxel-based ray cloud shows slightly higher pose estimation errors potentially due to coagulation of rays yielding less uniform distribution of rays across the scene. The fact that the SSIM value increases for lower density ray clouds also needs further investigation (see [32] for qualitative results)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.709, + 0.47, + 0.769 + ], + "angle": 0, + "content": "In Fig. 11, we demonstrate some corner cases revealed through our adapted inversion attack in Sec. 4. Fig. 11d shows the voxel-based ray sampling proposed in Sec. 5 can degrade the quality of scene recovery." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.785, + 0.196, + 0.8 + ], + "angle": 0, + "content": "7. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.811, + 0.47, + 0.901 + ], + "angle": 0, + "content": "We presented a privacy-preserving visual localization framework based on a new scene representation called 3D ray cloud, which is robust to the known geometry inversion attack and achieves an order of magnitude faster localization than previous line-based approaches without significantly compromising the localization accuracy. This is" + }, + { + "type": "image", + "bbox": [ + 0.516, + 0.365, + 0.603, + 0.499 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.515, + 0.502, + 0.6, + 0.513 + ], + "angle": 0, + "content": "(a) ground truth" + }, + { + "type": "image", + "bbox": [ + 0.609, + 0.364, + 0.696, + 0.432 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.609, + 0.432, + 0.696, + 0.5 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.607, + 0.501, + 0.695, + 0.513 + ], + "angle": 0, + "content": "(b) w/o rejection" + }, + { + "type": "image", + "bbox": [ + 0.701, + 0.365, + 0.789, + 0.432 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.701, + 0.432, + 0.789, + 0.5 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.701, + 0.501, + 0.785, + 0.513 + ], + "angle": 0, + "content": "(c) w/ rejection" + }, + { + "type": "image", + "bbox": [ + 0.795, + 0.365, + 0.883, + 0.432 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.795, + 0.432, + 0.882, + 0.5 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.795, + 0.501, + 0.877, + 0.513 + ], + "angle": 0, + "content": "(d) voxel-based" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.516, + 0.892, + 0.6 + ], + "angle": 0, + "content": "Figure 11. Illustration of point-to-image translation [40] results with point clouds from (a) ground truth, (b) recovered from ray cloud without rejection sampling, (c) recovered with rejection sampling, and (d) recovered from voxel-based ray cloud. While (c) shows improved recovery via ray rejection in Sec. 4.2, (d) shows voxel-based ray sampling lowering the reconstruction quality." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.606, + 0.892, + 0.682 + ], + "angle": 0, + "content": "enabled by the intrinsic geometry of the ray cloud allowing utilization of efficient widely-available perspective relative pose estimation solver. We have also explored potential improvements for the geometry-inversion attack on ray clouds and proposed a respective defense mechanism." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.686, + 0.892, + 0.837 + ], + "angle": 0, + "content": "Limitation Although this work can successfully defend the known inversion attack [7], we note it may be potentially vulnerable to a new type of attack in which omnidirectional rays are projected to a cylindrical or spherical image plane which is then unfolded and directly passed through an inversion network for image reconstruction. While realizing this attack may face its own challenges such as unknown depths and mixed chirality, we believe utilizing more ray centers will provide a defense against such attack by reducing the density of ray projections in each ray center." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Acknowledgement This work was supported by the NRF (Natioanl Research Foundation of Korea) grants funded by the Korea government (MSIT) (No. 2022R1C1C1004907). We thank anonymous reviewers for their valuable feedback." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "9780" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.079, + 0.09, + 0.176, + 0.107 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.169 + ], + "angle": 0, + "content": "[1] Sameer Agarwal, Yasutaka Furukawa, Noah Snavely, Ian Simon, Brian Curless, Steven M Seitz, and Richard Szeliski. Building Rome in a day. Communications of the ACM, 54(10):105-112, 2011. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.17, + 0.471, + 0.224 + ], + "angle": 0, + "content": "[2] Alexandre Alahi, Raphael Ortiz, and Pierre Vandergheynst. Freak: Fast retina keypoint. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 510-517, 2012. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.224, + 0.471, + 0.292 + ], + "angle": 0, + "content": "[3] Clemens Arth, Daniel Wagner, Manfred Klopschitz, Arnold Irschara, and Dieter Schmalstieg. Wide area localization on mobile phones. In 2009 8th IEEE International Symposium on Mixed and Augmented Reality, pages 73-82. IEEE, 2009. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.292, + 0.471, + 0.373 + ], + "angle": 0, + "content": "[4] Eric Brachmann, Alexander Krull, Sebastian Nowozin, Jamie Shotton, Frank Michel, Stefan Gumhold, and Carsten Rother. DSAC — Differentiable RANSAC for camera localization. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 2492–2500, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.374, + 0.471, + 0.429 + ], + "angle": 0, + "content": "[5] Eric Brachmann and Carsten Rother. Learning less is more - 6D camera localization via 3D surface regression. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4654-4662, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.429, + 0.471, + 0.482 + ], + "angle": 0, + "content": "[6] Robert Castle, Georg Klein, and David W Murray. Videorelate localization in multiple maps for wearable augmented reality. In 2008 12th IEEE International Symposium on Wearable Computers, pages 15-22. IEEE, 2008. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.482, + 0.471, + 0.55 + ], + "angle": 0, + "content": "[7] Kunal Chelani, Fredrik Kahl, and Torsten Sattler. How privacy-preserving are line clouds? Recovering scene details from 3D lines. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 15663-15673, 2021. 1, 2, 3, 4, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.55, + 0.471, + 0.619 + ], + "angle": 0, + "content": "[8] Kunal Chelani, Torsten Sattler, Fredrik Kahl, and Zuzana Kukelova. Privacy-preserving representations are not enough: Recovering scene content from camera poses. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13132-13141, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.619, + 0.471, + 0.673 + ], + "angle": 0, + "content": "[9] Ondrej Chum, Jiri Matas, and Josef Kittler. Locally optimized ransac. In Pattern Recognition: 25th DAGM Symposium, Magdeburg, Germany, September 10-12, 2003. Proceedings 25, pages 236-243. Springer, 2003. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.673, + 0.471, + 0.755 + ], + "angle": 0, + "content": "[10] Deeksha Dangwal, Vincent T Lee, Hyo Jin Kim, Tianwei Shen, Meghan Cowan, Rajvi Shah, Caroline Trippel, Brandon Reagen, Timothy Sherwood, Vasileios Balntas, et al. Mitigating reverse engineering attacks on local feature descriptors. In Proceeding of the British Machine Vision Conference (BMVC), 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.755, + 0.471, + 0.808 + ], + "angle": 0, + "content": "[11] Andrew J. Davison, Ian D. Reid, Nicholas D. Molton, and Olivier Stasse. MonoSLAM: Real-time single camera SLAM. IEEE Transactions on Pattern Analysis and Machine Intelligence, 29(6):1052-1067, 2007. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.808, + 0.471, + 0.863 + ], + "angle": 0, + "content": "[12] Yaqing Ding, Jian Yang, Viktor Larsson, Carl Olsson, and Kalle Åström. Revisiting the p3p problem. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4872-4880, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.863, + 0.471, + 0.89 + ], + "angle": 0, + "content": "[13] Alexey Dosovitskiy and Thomas Brox. Inverting visual representations with convolutional networks. In Proceedings" + }, + { + "type": "list", + "bbox": [ + 0.079, + 0.115, + 0.471, + 0.89 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 4829-4837, 2016. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.12, + 0.892, + 0.2 + ], + "angle": 0, + "content": "[14] Mihai Dusmanu, Johannes L. Schonberger, Sudipta N. Sinha, and Marc Pollefeys. Privacy-preserving image features via adversarial affine subspace embeddings. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 14262-14272, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.201, + 0.892, + 0.256 + ], + "angle": 0, + "content": "[15] Martin A. Fischler and Robert C. Bolles. Random sample consensus: A paradigm for model fitting with applications to image analysis and automated cartography. Communications of the ACM, 24(6):381-395, June 1981. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.256, + 0.892, + 0.323 + ], + "angle": 0, + "content": "[16] Marcel Geppert, Viktor Larsson, Johannes L Schonberger, and Marc Pollefeys. Privacy preserving partial localization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 17337-17347, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.323, + 0.892, + 0.35 + ], + "angle": 0, + "content": "[17] Richard Hartley. Cheirality. International journal of computer vision, 26(1):41-61, 1998. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.35, + 0.892, + 0.446 + ], + "angle": 0, + "content": "[18] Lionel Heng, Benjamin Choi, Zhaopeng Cui, Marcel Geppert, Sixing Hu, Benson Kuan, Peidong Liu, Rang Nguyen, Ye Chuan Yeo, Andreas Geiger, et al. Project automation: Localization and 3d scene perception for an autonomous vehicle with a multi-camera system. In Proceedings of the IEEE International Conference on Robotics and Automation (ICRA), pages 4695-4702, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.446, + 0.892, + 0.5 + ], + "angle": 0, + "content": "[19] Hiroharu Kato and Tatsuya Harada. Image reconstruction from bag-of-visual-words. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 955–962, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.5, + 0.892, + 0.554 + ], + "angle": 0, + "content": "[20] Tong Ke and Stergios I Roumeliotis. An efficient algebraic solution to the perspective-three-point problem. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 7225-7233, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.554, + 0.892, + 0.608 + ], + "angle": 0, + "content": "[21] Alex Kendall and Roberto Cipolla. Geometric loss functions for camera pose regression with deep learning. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 6555-6564, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.608, + 0.892, + 0.675 + ], + "angle": 0, + "content": "[22] Alex Kendall, Matthew Grimes, and Roberto Cipolla. PoseNet: A convolutional network for real-time 6-DOF camera relocalization. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), pages 2938-2946, 2015. 1, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.676, + 0.892, + 0.717 + ], + "angle": 0, + "content": "[23] Nicolaas H Kuiper. Tests concerning random points on a circle. In Nederl. Akad. Wetensch. Proc. Ser. A, volume 63, pages 38-47, 1960. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.717, + 0.892, + 0.756 + ], + "angle": 0, + "content": "[24] Viktor Larsson. PoseLib - Minimal Solvers for Camera Pose Estimation. https://github.com/vlarsson/ PoseLib, 2020. Accessed: 2022-10-30. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.756, + 0.892, + 0.824 + ], + "angle": 0, + "content": "[25] Viktor Larsson, Kalle Astrom, and Magnus Oskarsson. Efficient solvers for minimal problems by syzygy-based reduction. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 820-829, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.824, + 0.892, + 0.893 + ], + "angle": 0, + "content": "[26] Chunghwan Lee, Jaihoon Kim, Chanhyuk Yun, and Je Hyeong Hong. Paired-point lifting for enhanced privacy-preserving visual localization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 17266-17275, 2023. 1, 2, 3, 4," + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.893 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.514, + 0.957 + ], + "angle": 0, + "content": "9781" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.112, + 0.094, + 0.137, + 0.105 + ], + "angle": 0, + "content": "6,7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.106, + 0.47, + 0.174 + ], + "angle": 0, + "content": "[27] Hyon Lim, Sudipta N Sinha, Michael F Cohen, and Matthew Uytendaele. Real-time image-based 6-DOF localization in large-scale environments. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 1043–1050. IEEE, 2012. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.174, + 0.47, + 0.214 + ], + "angle": 0, + "content": "[28] David G. Lowe. Distinctive image features from scale-invariant keypoints. International Journal of Computer Vision, 60(2):91-110, Nov. 2004. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.214, + 0.47, + 0.269 + ], + "angle": 0, + "content": "[29] Simon Lynen, Torsten Sattler, Michael Bosse, Joel A Hesch, Marc Pollefeys, and Roland Siegwart. Get out of my lab: Large-scale, real-time visual-inertial localization. In Robotics: Science and Systems, volume 1, page 1, 2015. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.269, + 0.47, + 0.322 + ], + "angle": 0, + "content": "[30] Aravindh Mahendran and Andrea Vedaldi. Understanding deep image representations by inverting them. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 5188-5196, 2015. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.322, + 0.47, + 0.377 + ], + "angle": 0, + "content": "[31] Anastasiia Mishchuk, Dmytro Mishkin, Filip Radenovic, and Jiri Matas. Working hard to know your neighbor's margins: Local descriptor learning loss. Advances in Neural Information Processing Systems (NeurIPS), 30, 2017. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.377, + 0.47, + 0.417 + ], + "angle": 0, + "content": "[32] Heejoon Moon, Chunghwan Lee, and Je Hyeong Hong. Supplementary document for efficient privacy-preserving visual localization using 3d ray clouds, 2024. 3, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.417, + 0.47, + 0.47 + ], + "angle": 0, + "content": "[33] Raul Mur-Artal, Jose Maria Martinez Montiel, and Juan D Tardos. ORB-SLAM: A versatile and accurate monocular SLAM system. IEEE Transactions on Robotics, 31(5):1147-1163, 2015. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.471, + 0.47, + 0.566 + ], + "angle": 0, + "content": "[34] Tony Ng, Hyo Jin Kim, Vincent T Lee, Daniel DeTone, Tsun-Yi Yang, Tianwei Shen, Eddy Ilg, Vassileios Balntas, Krystian Mikolajczyk, and Chris Sweeney. NinjaDesc: Content-concealing visual descriptors via adversarial learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12797-12807, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.566, + 0.47, + 0.607 + ], + "angle": 0, + "content": "[35] David Nister. An efficient solution to the five-point relative pose problem. IEEE Transactions on Pattern Analysis and Machine Intelligence, 26(6):756-770, 2004. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.607, + 0.47, + 0.661 + ], + "angle": 0, + "content": "[36] Linfei Pan, Johannes Lutz Schonberger, Viktor Larsson, and Marc Pollefeys. Privacy Preserving Localization via Coordinate Permutations. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), 2023. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.661, + 0.47, + 0.715 + ], + "angle": 0, + "content": "[37] Mikael Persson and Klas Nordberg. Lambda twist: An accurate fast robust perspective three point (p3P) solver. In Proceedings of the European conference on computer vision (ECCV), pages 318-332, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.715, + 0.47, + 0.783 + ], + "angle": 0, + "content": "[38] Maxime Pietrantoni, Martin Humenberger, Torsten Sattler, and Gabriela Csurka. Segloc: Learning segmentation-based representations for privacy-preserving visual localization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15380-15391, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.783, + 0.47, + 0.85 + ], + "angle": 0, + "content": "[39] Francesco Pittaluga, Sanjeev Koppal, and Ayan Chakrabarti. Learning privacy preserving encodings through adversarial training. In Proceedings of the IEEE Winter Conference on Applications of Computer Vision (WACV), pages 791-799. IEEE, 2019. 1, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.85, + 0.47, + 0.892 + ], + "angle": 0, + "content": "[40] Francesco Pittaluga, Sanjeev J Koppal, Sing Bing Kang, and Sudipta N Sinha. Revealing scenes by inverting structure from motion reconstructions. In Proceedings of the" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.094, + 0.47, + 0.892 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 145-154, 2019. 1, 2, 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.12, + 0.892, + 0.174 + ], + "angle": 0, + "content": "[41] Francesco Pittaluga and Bingbing Zhuang. Ldp-feat: Image features with local differential privacy. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 17580–17590, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.174, + 0.892, + 0.241 + ], + "angle": 0, + "content": "[42] Rahul Raguram, Ondrej Chum, Marc Pollefeys, Jiri Matas, and Jan-Michael Frahm. USAC: A universal framework for random sample consensus. IEEE Transactions on Pattern Analysis and Machine Intelligence, 35(8):2022-2038, 2012. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.241, + 0.892, + 0.31 + ], + "angle": 0, + "content": "[43] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-Net: Convolutional networks for biomedical image segmentation. In Proceedings of the International Conference on Medical Image Computing and Computer-Assisted Intervention (MICCAI), pages 234–241. Springer, 2015. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.31, + 0.892, + 0.364 + ], + "angle": 0, + "content": "[44] Johannes Lutz Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.364, + 0.892, + 0.446 + ], + "angle": 0, + "content": "[45] Johannes Lutz Schonberger, Enliang Zheng, Marc Pollefeys, and Jan-Michael Frahm. Pixelwise view selection for unstructured multi-view stereo. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part III 14, pages 501-518. Springer, 2016. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.446, + 0.892, + 0.514 + ], + "angle": 0, + "content": "[46] Mikiya Shibuya, Shinya Sumikura, and Ken Sakurada. Privacy preserving visual SLAM. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXII 16, pages 102-118. Springer, 2020. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.514, + 0.892, + 0.568 + ], + "angle": 0, + "content": "[47] Sudipta Narayan Sinha, Pablo Alejandro Speciale, Sing Bing Kang, and Marc Andre Leon Pollefeys. Device pose estimation using 3d line clouds, Mar. 30 2021. US Patent 10,964,053. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.568, + 0.892, + 0.609 + ], + "angle": 0, + "content": "[48] Noah Snavely, Steven M Seitz, and Richard Szeliski. Modeling the world from internet photo collections. International Journal of Computer Vision (IJCV), 80:189-210, 2008. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.608, + 0.892, + 0.677 + ], + "angle": 0, + "content": "[49] Zhenbo Song, Wayne Chen, Dylan Campbell, and Hongdong Li. Deep novel view synthesis from colored 3d point clouds. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXIV 16, pages 1-17. Springer, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.676, + 0.892, + 0.744 + ], + "angle": 0, + "content": "[50] Pablo Speciale, Johannes L Schonberger, Sing Bing Kang, Sudipta N Sinha, and Marc Pollefeys. Privacy preserving image-based localization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5493-5503, 2019. 1, 2, 3, 4, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.744, + 0.892, + 0.811 + ], + "angle": 0, + "content": "[51] Pablo Speciale, Johannes L Schonberger, Sudipta N Sinha, and Marc Pollefeys. Privacy preserving image queries for camera localization. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 1486-1496, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.811, + 0.892, + 0.852 + ], + "angle": 0, + "content": "[52] Henrik Stewénius, Magnus Oskarsson, Kalle Aström, and David Nister. Solutions to minimal generalized relative pose problems, 2005. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.852, + 0.894, + 0.893 + ], + "angle": 0, + "content": "[53] Yurun Tian, Xin Yu, Bin Fan, Fuchao Wu, Huub Heijnen, and Vassileios Balntas. SOSNet: Second order similarity regularization for local descriptor learning. In Proceedings of" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.893 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9782" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.092, + 0.469, + 0.12 + ], + "angle": 0, + "content": "the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11016-11025, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.12, + 0.47, + 0.173 + ], + "angle": 0, + "content": "[54] Philip HS Torr and Andrew Zisserman. Mlesac: A new robust estimator with application to estimating image geometry. Computer vision and image understanding, 78(1):138-156, 2000. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.174, + 0.469, + 0.242 + ], + "angle": 0, + "content": "[55] Julien Valentin, Angela Dai, Matthias Niessner, Pushmeet Kohli, Philip Torr, Shahram Izadi, and Cem Keskin. Learning to navigate the energy landscape. In 2016 Fourth International Conference on 3D Vision (3DV), pages 323-332, 2016. 1, 3, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.242, + 0.469, + 0.296 + ], + "angle": 0, + "content": "[56] Jonathan Ventura, Clemens Arth, and Vincent Lepetit. An efficient minimal solution for multi-camera motion. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), pages 747-755, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.296, + 0.469, + 0.35 + ], + "angle": 0, + "content": "[57] Carl Vondrick, Aditya Khosla, Tomasz Malisiewicz, and Antonio Torralba. HOGgles: Visualizing object detection features. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), pages 1-8, 2013. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.35, + 0.469, + 0.404 + ], + "angle": 0, + "content": "[58] Philippe Weinzaepfel, Herve Jégou, and Patrick Pérez. Reconstructing an image from its local descriptors. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 337-344, 2011. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.404, + 0.469, + 0.458 + ], + "angle": 0, + "content": "[59] Enliang Zheng and Changchang Wu. Structure from motion using structure-less resection. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), pages 2075-2083, 2015. 2, 4" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.458 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.515, + 0.956 + ], + "angle": 0, + "content": "9783" + } + ] +] \ No newline at end of file diff --git a/2024/Efficient Privacy-Preserving Visual Localization Using 3D Ray Clouds/61345f4c-613d-4e27-91d1-09fcd6a52a24_origin.pdf b/2024/Efficient Privacy-Preserving Visual Localization Using 3D Ray Clouds/61345f4c-613d-4e27-91d1-09fcd6a52a24_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..c584e86adb29307fbec441d199fbdb9fa02f3073 --- /dev/null +++ b/2024/Efficient Privacy-Preserving Visual Localization Using 3D Ray Clouds/61345f4c-613d-4e27-91d1-09fcd6a52a24_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4a984119f3320258f02d67759d5e702daeb416e50a98a0f67f3d02304aa18e06 +size 6410214 diff --git a/2024/Efficient Privacy-Preserving Visual Localization Using 3D Ray Clouds/full.md b/2024/Efficient Privacy-Preserving Visual Localization Using 3D Ray Clouds/full.md new file mode 100644 index 0000000000000000000000000000000000000000..a8b01c782ff9f5740bda7832f7c172fc5d5ec6ba --- /dev/null +++ b/2024/Efficient Privacy-Preserving Visual Localization Using 3D Ray Clouds/full.md @@ -0,0 +1,400 @@ +# Efficient Privacy-Preserving Visual Localization Using 3D Ray Clouds + +Heejoon Moon1 Chunghwan Lee2 Je Hyeong Hong1,2† + +$^{1}$ Dept. Artificial Intelligence, Hanyang University $^{2}$ Dept. Electronic Engineering, Hanyang University + +# Abstract + +The recent success in revealing scene details from sparse 3D point clouds obtained via structure-from-motion has raised significant privacy concerns in visual localization. One prominent approach for mitigating this issue is to lift 3D points to 3D lines thereby reducing the effectiveness of the scene inversion attacks, but this comes at the cost of increased algorithmic complexity for camera localization due to weaker geometric constraints induced by line clouds. To overcome this limitation, we propose a new lifting approach called "ray cloud", whereby each lifted 3D line intersects at one of two predefined locations, depicting omnidirectional rays from two cameras. This yields two benefits, i) camera localization can now be cast as relative pose estimation between the query image and the calibrated rig of two perspective cameras which can be efficiently solved using a variant of the 5-point algorithm, and ii) the ray cloud introduces erroneous estimations for the density-based inversion attack, degrading the quality of scene recovery. Moreover, we explore possible modifications of the inversion attack to better recover scenes from the ray clouds and propose a ray sampling technique to reduce the effectiveness of the modified attack. Experimental results on two public datasets show real-time localization speed as well as enhanced privacy-preserving capability over the state-of-the-art without overly sacrificing the localization accuracy. + +# 1. Introduction + +Visual localization refers to the task of estimating the camera pose w.r.t. a known 3D scene given a query image. It serves a central role for cutting-edge applications involving robotics [18], augmented reality (AR) [3], virtual reality (VR) [6] and simultaneous localization and mapping (SLAM) [11, 29, 33]. To this date, many visual localization algorithms [27, 37] still utilize sparse 3D point clouds constructed via structure-from-motion (SfM) [1, 44, 45, 48] through which the camera pose is estimated from the correspondences between 2D image features and 3D points. This is commonly solved using a perspective-n-point (PnP) solver [12, 20, 37] equipped with the random sample con + +![](images/366bd3567fcbb48e266aa3fc72bbebef5752819862d8dd07d452697953d11b36.jpg) +(a) Illustration of ray cloud + +![](images/c7dfd1040ceb28268dd4cf0d37b2dbf182c1696678ef83c6c8dde24ffb6421d3.jpg) +(b) Single-image localization speeds + +![](images/db21b4eba16d5d0cd9fc42a7940145b6b5bdaaf47063920ce1ad2c8b5e7f5391.jpg) +Ground truth +(c) Comparison of revealed scene geometries and images +Figure 1. Main characteristics of the proposed ray cloud. (a) shows 3D points lifted as 3D rays half of which intersects at center 1 and the other half intersects at center 2. Estimating the camera pose resembles relative pose estimation with respect to a calibrated rig of perspective cameras. As shown in (b), this allows real-time single-image-based localization on two public datasets [22, 55] while preventing scene-revealing attacks [7, 40] as shown in (c). + +![](images/50f1964bde8163e009f1e75f57c5a4070bac7fe0040adf812122015850fb3c84.jpg) +Uniform line cloud +Paired-Point Lifting (PPL) + +![](images/048e209533ed2f407f8bf82451cb5f2df28b6a098347615188de72ae8ddc131f.jpg) + +![](images/8fc928a23ad2fb20720bdf4c923ca6857711f8f30cb9ea18900c33b70e85d60b.jpg) +Ray cloud (ours) + +sensus (RANSAC) meta-algorithm [9, 15, 42] or its deep network alternative such as differentiable RANSAC [4, 5]. + +Surprisingly, it was recently shown [8, 10, 13, 30, 40, 57, 58] that a sparse 3D point cloud may contain enough information to reveal image details of the underlying scene. Most prominently, Pittaluga et al. [40] demonstrated a cascaded U-Net model [43] can generate high-fidelity images of the scene just from the 2D projection of sparse 3D points and their SIFT features [28]. This raised significant privacy concerns [39, 50] as anyone with access to the 3D point cloud could potentially search for sensitive details in the private scene by employing this scene inversion network. + +Currently, the mainstream approach for preventing the aforementioned image inversion attack on the sparse 3D point cloud is geometric lifting [26, 46, 50], whereby each 3D point is transformed to a 3D line passing through the original point with aims to conceal the 3D scene geometry (see Fig. 2). While the earlier approach of randomly drawing the line direction (also known as the uniform line cloud) was shown to be susceptible to geometry inversion attack [7], + +this was remedied by recent methods utilizing pairing of 3D points [26, 36] to increase ambiguity in the lifted scene representation. Nevertheless, single image-based localization speeds for these algorithms [26, 36] are far from achieving real-time performance, providing motivation for this work. + +In this paper, we present a new geometric lifting approach called ray cloud with aims to address above issue of slow camera localization runtime exhibited by previous studies. As will be illustrated in Sec. 3, ray cloud is simple to construct and robust to the geometry inversion attack [7]. Most importantly, camera localization using a ray cloud can be formulated as a well-known relative pose estimation problem between the query image and a virtual calibrated rig of cameras, enabling utilization of the efficient 5-point algorithm [35] with known scale [59]. Moreover, by modifying and enhancing the geometry inversion attack [7] in Sec. 4, we explore potential hidden privacy risks associated with ray clouds, and subsequently propose a technique to enhance the privacy protection of ray clouds in Sec. 5. + +The contributions of this work are summarized below: + +- an efficient privacy-preserving visual localization framework based on a new lifting approach called 3D ray cloud, +- customization of the geometry inversion attack [7] to yield undesired (privacy leak) corner cases for ray clouds, +- a voxel-based ray sampling technique to prevent above privacy leaks by disrupting nearest neighbor statistic and inducing erroneous recovery of the scene points, and +- extensive experimental comparison of ray cloud against other line cloud-based approaches on two public datasets, achieving real-time single-image based localization speed for the first time while preserving privacy-preserving capability and maintaining similar localization accuracy. + +# 2. Related work + +Privacy threats from image inversion attack While several prior studies addressed the possibility of revealing images from sparse keypoints [13, 19, 30, 57, 58], it was Pitaluga et al. [40] who posed a real privacy threat for the first time by recovering detailed scene images from a sparse 3D point cloud using a cascaded U-Net [43]. This network called InvSfM takes in the 2D locations of the projected 3D points as well as the corresponding depths and SIFT descriptors, and outputs a synthesized image of the scene. + +More recently, Song et al. [49] used a PointNet++ architecture to extract deep features and synthesize scene images. Dangwal et al. [10] devised a generic inversion model that can work with various descriptors [2, 28, 53] and proposed selective feature suppression to hide privacy-sensitive regions. Nonetheless, the pretrained InvSfM model still serves as the baseline for qualitative and quantitative analyses in privacy-preserving visual localization [7, 26, 36, 38]. + +Privacy-preserving scene representations After the emergence of InvSfM [40], Speciale et al. [50] quickly em + +braced a geometric lifting approach, also known as the uniform line cloud, where each 3D point in the original sparse point cloud is replaced by a 3D line passing through the point with its direction drawn from a uniform distribution on the unit sphere. This scene representation aims to introduce anonymity to the locations of the 3D points thereby preventing any image inversion via InvSfM. The work was later extended by Shibuya et al. [46] to visual SLAM with a prebuilt map provided in the form of uniform line cloud. + +Unfortunately, above methods [46, 50] are now outdated and susceptible to the geometry inversion attack proposed by Chelani et al. [7], which successfully recovers the sparse 3D point cloud from a uniform line cloud (details in the next subsection). To overcome this limitation, Lee et al. [26] proposed to draw 3D lines through random pairs of 3D points, thereby introducing non-uniform line distributions and multiple 3D point candidates for each line both of which are empirically shown to hinder point cloud inversion. Pan et al. [36] proposed to permute the coordinates of random pairs of keypoints to preserve privacy while still enabling localization from multiple axis-aligned lines. Geppert et al. [16] presented a partial localization approach whereby three 1-dimensional partial maps are constructed from the 3D points and separately stored to enhance security. While these approaches can evade the geometry inversion attack, they are not runtime-efficient to allow real-time camera localization from single images. This issue is addressed through use of the ray cloud proposed in this work. + +Other explored approaches include keypoint lifting [14], carving [34] or obfuscation [41], whereby the feature descriptors of the 3D points are transformed to hide keypoint details, but this does not conceal the point cloud geometry. + +Privacy threats for line clouds Chelani et al. [7] presented a density-based inversion attack that can reveal the 3D point geometry from a uniform line cloud [50]. While the posterior probability of the point cloud $\mathcal{P}$ given a line cloud $\mathcal{L}$ can be expressed as $P(\mathcal{P}|\mathcal{L})\propto P(\mathcal{P})P(\mathcal{L}|\mathcal{P})$ using Bayes' rule, it is noted that the likelihood $P(\mathcal{L}|\mathcal{P})$ is constant for uniform line clouds. Hence, maximizing the posterior amounts to maximizing the marginal prior $P(\mathcal{P})$ . + +For maximizing $P(\mathcal{P})$ , Chelani et al. [7] resorted to an empirical observation that for any two 3D points $\mathbf{x}_A$ and $\mathbf{x}_B$ with their corresponding uniformly-drawn lines $\mathbf{l}_A \in Gr(1,3)$ and $\mathbf{l}_B \in Gr(1,3)$ , $P(\| \mathbf{x}_A - \mathbf{x}_A^\perp \|_2 < \| \mathbf{x}_A - \mathbf{x}_B \|_2) \approx 0.8$ , where $\mathbf{x}_A^\perp$ denotes the point on $\mathbf{l}_A$ closest to $\mathbf{l}_B$ . This implies that if $\mathbf{x}_A$ and $\mathbf{x}_B$ are close to each other, then $\mathbf{x}_A^\perp$ is likely to become a good approximation of $\mathbf{x}_A$ . Hence, $P(\mathcal{P})$ can be maximized if the nearest neighboring points of each 3D point are known in advance. + +As it is non-trivial to find a set of true $k$ -nearest neighboring points in line clouds, Chelani et al. [7] approximates this by the set of $K$ -nearest lines $\{\mathbf{l}_K\}$ for each line $\mathbf{l}_i$ , expecting some to include the true positives of the $k$ -nearest + +![](images/a7397179caebe7e424db836405385e05a2671ff62b4d64c2b02abc4cbc3a7615.jpg) +(a) Point cloud + +![](images/d4b8d335014fe21b9c8c72cfbadff0d8f016075d5650b5b70836abc65f141799.jpg) +(b) Uniform line cloud [50] + +![](images/481016dfcf7ad1fdfc014557093e47a2c40c5ab8c519a78c0c4d61f8a7416f00.jpg) +(c) Paired-Point Lifting (PPL) [26] + +![](images/140d4d62b1209a412c9ae1a1018164b819389e7169869e78ae463663fb174d3f.jpg) +(d) Ray cloud (ours) + +points when $K > k$ is large enough (we distinguish $K$ for lines and $k$ for points). Bearing this assumption, [7] computes the closest point $\mathbf{x}_{k}^{\perp}$ to each of the $K$ -nearest lines for the current line $\mathbf{l}_i$ , builds a histogram of closest points $\{\mathbf{x}_{k}^{\perp}\}$ and finds the peak of the histogram via Kuiper's test [23] to estimate the point $\mathbf{x}_i$ . This attack is effective for uniform line clouds [50], allowing recovered points to be used for scene recovery with InvSfM [40]. However, it is less effective against more recent line-based approaches [26, 36] and ray clouds presented in this work. + +Camera localization using line clouds As mentioned in Sec. 1, camera localization typically involves a RANSAC algorithm [15] coupled with an efficient minimal solver that can run fast over thousands of iterations [25]. For single-image based localization, a standard line cloud introduces 2D points-to-3D line constraints, which are weaker than 2D point-to-3D point constraints provided by the point clouds. Consequently, the minimal solver for line clouds (p6L) requires 6 correspondences between 2D points and 3D lines [50]. Since p6L innately uses the solver for generalized relative pose estimation [52], achieving real-time speed with $p6L$ is challenging due to the large number of possible solutions (up to 64) and high algorithmic complexity [52, 56]. Shibuya et al. [46] achieved real-time localization with line clouds in the SLAM setup, but this requires a video stream as input and construction of a client-side point map. On the other hand, we will show it is possible to perform image-based localization in real-time using ray clouds. + +# 3. 3D Ray cloud + +We now illustrate the process of constructing a 3D ray cloud shown in Fig. 2 and estimating camera pose from ray clouds. We also discuss inversion attack [7] on ray clouds. + +Motivation The concept of lifting 3D lines to intersect at a particular point has been briefly addressed without detailed investigation in [7, 47]. Chelani et al. [7] mentioned that, if all lines were to intersect at a single point, then the geometry inversion attack in Sec. 2 would fail. This is because the histogram of closest points (to other lines) becomes concentrated about the point of intersection, leading to a degenerate solution. Unfortunately, this representation possesses an intrinsic limitation that the camera pose can + +![](images/ad711c4d29a0eb1a16d0bb6d396b996c112abb08b9da1c5cfc9fc5c93e7c7de0.jpg) +Figure 2. 3D line clouds yielded from different lifting approaches on the Office1 manolis in the Energy Landscape [55] dataset. Note each line (ray) in the ray cloud passes through one of two center points. We show $10\%$ of the total number of lines for visual comparison. +(a) Original 3D points +Figure 3. Illustration of the procedure for selecting points of intersection and constructing a 3D ray cloud. + +![](images/608b13162b1499f0fc5964576cdfef2d47e5b5a5d0147e91975ce9425b9ecaeb.jpg) +(b) Center assignment + +![](images/f1e809af13ffa8b556d0b403617e6e378601734c90d64b1747bbc490691ec0b8.jpg) +(c) Ray cloud + +only be recovered up to scale as camera localization resembles relative pose estimation [35] between the query camera and the intersecting 3D lines. We aim to overcome this issue by utilizing two intersection points. This can be viewed as adding another camera to the scene, forming a calibrated rig of perspective cameras to resolve the scale ambiguity. + +# 3.1. Construction procedure + +Lifting 3D points to a ray cloud comprises three steps, namely setting two ray centers and assigning 3D points to one of two ray centers. We choose to set the ray centers by applying the $K$ -means clustering algorithm on the point cloud with $K = 2$ , from which the centers of two clusters are set as the ray centers (explained below). Then, we assign half of randomly sampled points from the point cloud to center $1(\mathbf{c}_1 \in \mathbb{R}^3)$ , and the other half to center $2(\mathbf{c}_2 \in \mathbb{R}^3)$ . Note this assignment is not related to the $K$ -means clustering. Finally, each point $\mathbf{x}_i$ is lifted to a line as $\mathbf{c}_j + \lambda (\mathbf{x}_i - \mathbf{c}_j)$ for $\lambda \in \mathbb{R}$ , where $j \in [1,2]$ is the assigned center index for point $i$ . The final output resembles omnidirectional rays from two perspective cameras (see Fig. 3). + +Effect of changing ray center locations If the ray centers are located very close to each other, then they begin to resemble a single ray center, increasing ambiguity in translation scale. On the other hand, if the ray centers are distant from each other and the scene, then the resulting rays are nearly parallel which degrades localization accuracy. We empirically observe that using the centers from $K$ -means clustering usually avoids both extreme cases (see [32]). + +# 3.2. Camera pose estimation + +We now illustrate an efficient localization approach for ray clouds that can benefit from faster inference speed com + +![](images/466228ecdfb17b4e4e8eb72ed74c0210f46a0a4d2c239bf40b2999d82ecc7afc.jpg) +(a) Standard line cloud $(p6\mathrm{L})$ + +![](images/b7c13d92d34eeafcf405d7564e5499eeea8139b76eb8e20daf3139d47ee3d644.jpg) +(b) Ray cloud $(p5 + 1\mathrm{R})$ +Figure 4. Illustration of minimal pose estimation problems solved by different line representations. Camera localization using a uniform line cloud can be seen as the special case of generalized relative pose estimation problem [52] involving one perspective camera (query) and one generalized camera (line cloud). On the other hand, camera localization using a ray cloud can be regarded as the case of perspective relative pose estimation problem between the query image and a rig of calibrated cameras (ray cloud). The latter enjoys shorter runtime as shown in Sec. 6 partly due to stronger ray constraints reducing the number of algebraic solutions. + +pared to other line-based scene representations [26, 36, 50]. This is inspired by the illustration in Fig. 4b that the ray cloud can be viewed as two virtual perspective cameras located at the ray centers $\mathbf{c}_1 \in \mathbb{R}^3$ and $\mathbf{c}_2 \in \mathbb{R}^3$ . Since the rays from these cameras are omnidirectional, we can set their rotation matrices $\mathsf{R}_1, \mathsf{R}_2 \in SO(3)$ to be aligned with world coordinates, i.e. $[\mathsf{R}_1|\mathbf{t}_1] = [\mathsf{I}| - \mathbf{c}_1]$ and $[\mathsf{R}_2|\mathbf{t}_2] = [\mathsf{I}| - \mathbf{c}_2]$ , where $\mathbf{t}_1, \mathbf{t}_2 \in \mathbb{R}^3$ are the camera translations. + +Below, we show that having 5 correspondences between 2D keypoints from the query image and 3D rays from a randomly selected virtual camera and one correspondence between a 2D query keypoint and a 3D ray from a remaining virtual camera allows us to estimate the absolute camera pose using a ray cloud. We denote this adaptation of minimal $5 + 1$ solver [59] as the perspective- $(5 + 1)$ -rays $(p5 + 1\mathrm{R})$ solver. Similar to previous studies [26, 50], we assume the camera intrinsics $\mathbf{K} \in \mathbb{R}^{3 \times 3}$ are known. + +$p5 + 1\mathbf{R}$ solver We start by estimating the relative pose between the query image and the first virtual camera using the 5-point algorithm [35]. It yields multiple candidates for the relative pose of the query image w.r.t. the first virtual camera defined as $[\mathtt{R}_{q1}|\alpha \hat{\mathbf{t}}_{q1}]$ , where $\mathtt{R}_{q1}\in SO(3)$ is the relative rotation, $\hat{\mathbf{t}}_{q1}\in S^2$ is the normalized relative translation and $\alpha \in \mathbb{R}$ is an unknown scale factor. Subsequently, the absolute pose of query camera is $[\mathtt{R}_{q1}|\alpha \hat{\mathbf{t}}_{q1} - \mathtt{R}_{q1}\mathbf{c}_1]$ . + +In order to disambiguate $\alpha$ , we utilize the additional correspondence between a normalized query keypoint $\mathbf{p} := (p_1, p_2)^\top \in \mathbb{R}^2$ and a 3D ray $\mathbf{l} \in Gr(1, 3)$ from the second virtual camera. We note the ray from the query camera center passing through $\mathbf{p}$ would ideally intersect $\mathbf{l}$ . Expressing both rays in the query-centered coordinates yields + +$$ +\lambda_ {1} \left[ p _ {1}, p _ {2}, 1 \right] ^ {\top} = \lambda_ {2} \mathrm {R} _ {q 1} \hat {\mathbf {n}} + \alpha \hat {\mathbf {t}} _ {q 1} - \mathrm {R} _ {q 1} \mathbf {c} _ {1} \tag {1} +$$ + +where $\hat{\mathbf{n}}\in S^2$ is the line direction of 1 in world coordinates and $\lambda_1,\lambda_2\in \mathbb{R}$ are scale parameters. Since we have 3 equations with 3 unknowns, $\alpha$ can be determined. + +While the $p5 + 1\mathrm{R}$ solver requires 6 correspondences like $p6\mathrm{L}$ [26, 50], the utilized 5-point algorithm only yields up to 10 solutions [35] that is smaller than 64 solutions from the generalized solver [52] in $p6\mathrm{L}$ . This significantly reduces the computational complexity for determining the correct configuration. We also enforce the chirality constraint [17] on the query image to discard infeasible solutions. + +Robust estimation We equip above $p5 + 1R$ solver with the LO-RANSAC [31] algorithm for robust camera localization. For selecting the best model candidate, we adopt an MLESAC [54]-type cost function based on the squared sum of epipolar distances in the query image defined as + +$$ +\sum_ {(i, j) \in \Omega} \rho \left(\frac {\left([ \mathbf {p} _ {i} ^ {\top} , 1 ] \mathrm {E} _ {j} \hat {\mathbf {n}} _ {i}\right) ^ {2}}{\left(\mathbf {e} _ {j 1} ^ {\top} \hat {\mathbf {n}} _ {i}\right) ^ {2} + \left(\mathbf {e} _ {j 2} ^ {\top} \hat {\mathbf {n}} _ {i}\right) ^ {2}}\right), \tag {2} +$$ + +where $\rho(s) := \min(s, \tau^2)$ is a robust kernel with upper bound of threshold $\tau^2$ , $\Omega$ denotes the pairing arrangement of which keypoint $i$ matches with ray $i$ from center $j$ , $\mathbf{p}_i \in \mathbb{R}^2$ is the $i$ -th normalized keypoint in the query image, $\hat{\mathbf{n}}_i \in S^2$ is ray $i$ passing through the ray center $j \in [1, 2]$ , and $\mathbf{E}_j := [\mathbf{e}_{j1}, \mathbf{e}_{j2}, \mathbf{e}_{j3}]^\top$ is the essential matrix between the virtual camera $j$ and the query camera. Since the two virtual cameras share the same rotation and only differ by translation, $\mathbf{E}_2$ can be easily computed from $\mathbf{E}_1$ . The best solution from RANSAC is refined by iteratively minimizing (2). + +# 3.3. Geometry inversion attack on ray clouds + +We show that applying the geometry inversion attack [7] naively to a ray cloud yields a degenerate solution. For this purpose, we make use of some prior knowledge regarding the probabilistic viewpoint and algorithmic details of the geometry inversion attack [7] reviewed in Sec. 2. + +Let $\mathbf{x}_i\in \mathbb{R}^3$ be the $i$ -th point and $\mathbf{l}_i\in Gr(1,3)$ be its lifted ray. The point cloud is defined as $\mathcal{P}\coloneqq \{\mathbf{x}_i\}$ and the ray cloud is defined as $\mathcal{L}\coloneqq \{\mathbf{l}_i\}$ . Since ray centers $\mathcal{C}\coloneqq \{\mathbf{c}_1,\mathbf{c}_2\}$ are easily found by inspecting the intersection points, we aim to maximize the conditional posterior of $\mathcal{P}$ , + +$$ +P (\mathcal {P} | \mathcal {L}, \mathcal {C}) \propto P (\mathcal {L} | \mathcal {P}, \mathcal {C}) P (\mathcal {P} | \mathcal {C}). \tag {3} +$$ + +Note the conditional likelihood $P(\mathcal{L}|\mathcal{P},\mathcal{C})$ is a non-zero constant as long as each ray intersects the corresponding pair of ray center and 3D point and the probability of each ray being assigned to one of the ray centers is equally likely. As long as each point $\mathbf{x}_i$ is estimated to be on its lifted ray $\mathbf{l}_i$ , then maximizing (3) amounts to maximizing $P(\mathcal{P}|\mathcal{C})$ . + +However, since $P(\mathcal{P}|\mathcal{C})$ can depend significantly on the ray centers $\mathcal{C}$ , it seems difficult to yield a point statistic similar to $P(\mathcal{P})$ in [7]. We implicitly figure out the difference between $P(\mathcal{P}|\mathcal{C})$ and the original prior $P(\mathcal{P})$ by applying the geometry inversion attack [7], which maximizes $P(\mathcal{P})$ , as-is to the ray cloud. As visualized in Fig. 5a, the estimated + +![](images/820beb607c673b3e9c04979fc5db6a32a23db599b489cdbdd4e287dd7aa956fb.jpg) +(a) Naive neighbor estimation [7] + +![](images/352a6a3531ff6fd64b52f8d31f5a431e469af2ef26bb992055f026e6e5022559.jpg) +(b) Modified neighbor estimation +Figure 5. Illustration of peak-finding [7] results from (a) naive neighbor-lines estimation [7] and (b) modified neighbor-lines estimation. (a) indicates that the estimated peak will always be located at the center of the current ray. In contrast, our modification only samples the neighborhood of lines using the bundle of rays originating from the opposite center, eliminating the trivial solution. + +point for each ray becomes the center of the constituent ray, yielding a trivial recovery. This is because, in the process of the geometric inversion, the $K$ -nearest neighboring lines include a bundle of rays originating from the same center, which yields an erroneous peak at the ray center on the histogram of point candidates, hindering general recovery. + +# 4. Adapted inversion attack for ray clouds + +In this section, we illustrate two main reasons for trivial recovery of 3D points in Sec. 3.3, namely i) lower significance of nearest points for estimating 3D points, and ii) bad estimation of true nearest rays lifted from nearest points. + +First, the assumption in [7] that the near-80% of candidates which are closest to lines lifted from $k$ -nearest neighboring points yields accurate points does not hold for ray clouds as some of the rays from these $k$ -nearest points will yield estimates at the current ray center. Second, the assumption in [7] that $K$ -nearest lines will overlap enough with the set of rays lifted from the $k$ -nearest points ( $K > k$ ) is weakened as the $K$ -nearest lines will be rays sharing the same center whose original points may be far away from the original point of the current lifted ray. + +In Sec. 4, we attempt to address these issues by considering known ray centers in order to approximately maximize $P(\mathcal{P}|\mathcal{C})$ instead of $P(\mathcal{P})$ . We also aim to address trivial solutions by incorporating the known positions of the ray centers $\mathcal{C}$ to approximate maximizing over $P(\mathcal{P}|\mathcal{C})$ . + +# 4.1. Eliminating the degenerate solution + +The trivial recovery of 3D points (estimated 3D points $\mathcal{P}$ always located at the two ray centers $\mathcal{C}$ ) is due to rays intersecting the same ray center always ending up as the nearest neighboring (NN) lines, producing the peak density of closest points at the ray center (see Fig. 5a). To bypass this issue, we exclude the rays intersecting through the current ray when constructing the set of nearest neighboring lines for estimating the peak density, and only select the NN-lines from the other ray center. This can be viewed as incorporating the known locations of the ray centers for recovering the points (see Fig. 5b) to maximize $P(\mathcal{P}|\mathcal{C})$ . + +![](images/31dee82cba4c836a92748c426b53ff2f2e76aa06523a1bc5e177aa8f28b34bb9.jpg) +(a) ULC [50] (oracle) +Figure 6. 3D point cloud recovered on different settings of nearest neighbors for the Apt1 living from [55] (black-original points, red-recovered points). In (b) and (c), we have only utilized neighboring rays connected to the opposite ray center to avoid degenerate solutions (in Sec. 4.1). For the oracle cases of (a) and (b), rays from the true nearest points of the original points are used to estimate the points, showing feasible recovery. In contrast, when the nearest neighboring points are approximated by the nearest neighboring rays, it leads to worse recovery as shown in (c). This implies false positive rays, which are close but whose original points are distant, significantly degrade the recovery of points. + +![](images/5729c9489b4c27e7d7f4600c3bf685873b1f69e86d8c7fe9f6ebd38c4c4d959f.jpg) +(b) Ray cloud (oracle) + +![](images/a1d4538230a73ddb956fa84df194d8e09ffe729faffaf755b600da0a25970c3f.jpg) +(c) Ray cloud + +Fig. 6b shows that, in the oracle case when the ground truth nearest points of the current ray's original 3D point are known, discarding NN-lines intersecting the current ray can avoid degenerate solution and yield more feasible point cloud. At the same time, Fig. 6c shows that the actual point recovery from a ray cloud is worse. We conjecture this is due to the misalignment between the set of nearest neighboring rays and the set of nearest neighboring points for a given point and its lifted ray which is addressed in Sec. 4.2. + +# 4.2. Rejecting rays closely aligned to the baseline + +We first analyze the cause of the bad recovery in Fig. 6c even after applying the trick in Sec. 4.1. We define the rays close to the current ray but whose original points are far away from the original point of the current ray as the false positive rays. By comparing the oracle case (b) and the normal case (c) in Fig. 6, we anticipate that the false positive rays are the culprits for degraded estimation of the points. + +We conjecture two cases where the false positive rays can lead to misleading estimations. First, when there are many rays from the opposite center that are in the proximity of the current ray center, the peak of the histogram may be located near the current ray center. Vice versa, when the current ray passes through the denser area around the opposite center, the peak will be located near the opposite center. These two cases partially cause the 3D points to shrink (see Fig. 6c). Second, when there are many diverse rays in the proximity of the current ray, multiple peaks can be formed from the point candidates (closest points on the ray to neighboring rays) along the current ray. This can arise when the ray centers are embedded in a room-like environment to form congested rays (see [32] for more discussions). + +Out of above cases, we attempt to reduce the shrinking effect of false positive rays by incorporating known $\mathcal{C}$ and not considering any rays from the opposite center which are within the proximity of the current ray center. This can be effectively achieved by excluding a fraction $(\beta)$ of rays most + +![](images/f57a7670a21f370da4d522aac9506a6a8398311e743ae7d71fcc43be0fcc437c.jpg) +(a) Visualization of rejection-sampling of rays near the baseline + +![](images/521189a74e1679fcf4451972beafe39ac20309e8c7f49929f075ba1c8524b9bf.jpg) +(b) Peak-fiding [7] w/o rejection +Figure 7. Illustration of the effect of line rejection method. For a synthetic room comprising $20\mathrm{k}$ points, in (b), peak estimation fails to estimate the correct point due to false positive rays. In (c), the rejection excludes false candidates to estimate an accurate point. + +![](images/32ffe51ebfde302520b5f6777252bf730443794d3462c38cf907db5e17a769a8.jpg) +(c) Peak-finding [7] with rejection + +aligned with the baseline when estimating the histogram of point candidates. Fig. 7 illustrates the effect of this ray rejection which can lead to better point estimation. + +We outline some limitations of this approach. First, if the ray lifted from a 3D point is within the region of rejection sampling, then the estimated point can be inaccurate due to the lack of nearby rays to support the estimation process. Second, the method introduces a hyperparameter $(\beta)$ that needs to be carefully set (see results in Table 2). + +# 5. Voxel-based ray sampling + +We now present a new ray sampling technique that can neutralize the modified inversion attack from Sec. 4. + +Our approach is motivated from Fig. 6 that the quality of point recovery depends on finding the true point candidates once the degenerate solution is eliminated. This time, we try to discourage the true nearest neighboring points from being included in the set of $K$ -nearest neighboring lines when lifted to rays with aims to lower the inversion accuracy. + +For the above purpose, we partition the 3D points into voxels and assign one of two ray centers to each voxel such that i) 3D points within the same voxel are lifted to rays intersecting through the assigned ray center and ii) no neighboring voxels share the same ray center (see Fig. 8). As long as the voxels are large enough, many 3D points except those near the voxel boundaries will have most of their respective $k$ -nearest points to be inside the same voxel and subsequently share the same ray center when lifted to rays. Since the rays lifted from these true $k$ -nearest neighboring points inside the same voxel will yield a trivial solution as illustrated in Sec. 4.1, these will be disallowed from being used for peak estimation. Hence, we essentially minimize the overlap between the set of $K$ -nearest lines and the set of $k$ -nearest points which is a necessary condition for successful recovery of points from lines in [7]. + +![](images/0a7186f21b53f05692f48bea903a38b8f88983ac77c35b7bec61dad57493b244.jpg) +Point cloud +Figure 8. The procedure of voxel-based ray sampling. After partitioning the point cloud into a voxel grid, the whole grid is divided into two clusters in such a way that neighboring grids belong to different clusters. Then, all 3D points in each grid within the same cluster are paired with one center, while the other cluster is paired with the remaining center for lifting. + +![](images/e9c4f393c95b84e58da45ef0655353fcdff8be0c6ce41c964cef01e49263802d.jpg) +Voxel-grid + +![](images/0923c6ea957cd37df3ebb649627c731d1bd44b1353907841e32c662808cc7df4.jpg) +Grid assignment + +![](images/c8346fded2e911cfce79ce695ae1de17b31c5a5cc371e3a114c3630eb3fad2f2.jpg) +Lifting + +We adopt a 3D checkerboard pattern for assigning ray centers to voxels with aims to achieve stable camera localization accuracy by uniformly distributing rays from different ray centers across the scene. While a larger voxel size will allow more true nearest neighboring points to be discarded and provide better privacy protection, it will also cause the rays to be more non-uniform and lower the camera localization accuracy (see [32]). Currently, the voxel size needs to be manually determined for each dataset as the optimal trade-off point can vary across different scenes. + +# 6. Experimental Results + +We present comprehensive experimental results showing the effectiveness of ray cloud in terms of localization accuracy, inference speed, and privacy-preserving performance. + +Datasets We used the same two public datasets as in [7, 26]. The Energy Landscape [55] dataset consists of 12 indoor scenes with multiple sequences of scanned RGB images. The Cambridge [21, 22] dataset comprises 6 distinct outdoor landmarks, containing multiple image frames with labeled 6-DOF camera poses for the camera re-localization. As in [26], we excluded the scenes of Street and Great Court from [22] due to the presence of large geometric outliers. We followed the same experimental protocols from [26, 50] for constructing sparse 3D point clouds using COLMAP [44] and evaluation. Separated queries not included in 3D map construction were used for evaluation. + +Implementation details We used Poselib [24] to utilize the $p5 + 1\mathrm{R}$ solver and the repository of PPL [26] to run experiments. We report coarse estimation results of inversion in ray cloud and voxel-based ray cloud as the refine estimation [7] diverges due to incorrect initial estimation (see [32]). We used a workstation with an Intel CPU i9-13900K running at 3.0 GHz and NVIDIA RTX 4090 GPUs. + +Evaluation metrics For camera localization, we estimated the absolute 6-DOF pose of the query image with given 2D-3D correspondences. We define $\mathsf{R} \in SO(3)$ and $\mathbf{t} \in \mathbb{R}^3$ as the ground truth camera rotation and translation respectively, and $\hat{\mathbf{R}} \in SO(3)$ and $\hat{\mathbf{t}} \in \mathbb{R}^3$ as the estimated rotation and translation respectively. Then, we obtained the estimated pose errors by following [26, 50, 51], where the + +
DatasetScene representationMinimal solverPose errorRANSACPrivacy preserving
ΔR↓meanmedianΔt↓meanmediani meanr meant ↓meaneg↑ medianPSNR ↓ meanSSIM ↓ meanMAE ↑ mean
Energy LandscapePoint cloudp3P0.1780.0310.0070.0011199.544-15.760.51633.29
Uniform line cloud [50]p6L0.2720.0980.0100.00312481.12900.04513.450.40342.83
PPL [26]p6L0.2920.1130.0110.00412381.51880.05610.420.35562.61
PPL+ [26]p6L0.2930.1150.0110.00412581.30850.05610.390.35562.70
Ray cloud (ours)p5+1R0.2620.0870.0100.00314680.0360.66910.290.31962.97
Voxel-based ray cloud (ours)p5+1R0.2970.1160.0110.00414480.1560.77010.050.31564.84
CambridgePoint cloudp3P0.7000.1010.1420.0431298.487-14.000.43339.40
Uniform line cloud [50]p6L0.7750.1690.1720.06643671.872890.74512.710.33146.08
PPL [26]p6L0.7750.1830.1810.08126474.491860.97711.210.29655.53
PPL+ [26]p6L0.7640.1830.1780.08128474.111900.94811.160.29755.91
Ray cloud (ours)p5+1R0.8550.2220.2190.10723974.89129.61410.410.26160.98
Voxel-based ray cloud (ours)p5+1R0.8800.2560.2350.12025075.241210.5410.230.25362.56
+ +Table 1. Comparison of localization and privacy-preserving performance with different 3D scene representations ( $\Delta R$ : rotation error $[\circ]$ , $\Delta t$ : translation error [m], i: number of iterations, r: inlier ratio (\%), t: runtime [ms], $e_g$ : geometric (3D point) error [m]). We applied modified inversion attack from Sec. 4 with 25% line rejection to ray cloud and voxel-based ray cloud with voxel size of 0.5m and 5.0m for Energy Landscape and Cambridge respectively. Bold indicate the best result in each metric among privacy-preserving representations. + +![](images/eb107fcb95284603b7c0af91d1949ca09f0cdb3dd6c1211101ca5468ca390fac.jpg) +Figure 9. Comparison of revealed scene geometries and images for Apt2 bed [55]. Modified inversion attack is applied to (d) and (e). + +rotational error is computed as $\Delta \mathsf{R} = \arccos \frac{\mathrm{Tr}(\hat{\mathsf{R}}^{\top}\mathsf{R}) - 1}{2}$ and the translation error computed as $\Delta \mathbf{t} = \left\| \hat{\mathsf{R}}^{\top}\hat{\mathbf{t}} -\mathsf{R}^{\top}\mathbf{t}\right\|_{2}$ . + +To compare the quality of recovered point clouds, we calculated the 3D median point errors between the original and estimated points in meters as in [26]. We also compared the quality of revealed images by feeding recovered point cloud and ground truth camera pose of query images to InvSfM [39] and calculating peak signal-to-noise ratio (PSNR), structural similarity index measure (SSIM), and mean absolute error (MAE) values against the queries. + +# 6.1. Camera localization accuracy and runtime + +As shown in Table 1, our ray cloud family of approaches achieved the fastest inference speed among the line-lifting methods with no significant decline in pose accuracy. Notably, these achieved real-time performance on both datasets, reaching 137 FPS for the ray cloud and 136 FPS for the voxel-based ray cloud on average. As further demonstrated in [32], this is mainly due to the introduction of the $p5 + 1\mathrm{R}$ solver which is only available for ray clouds. + +The ray cloud showed the lowest localization error on [55] but slightly higher errors than ULC and PPL on [22]. We conjecture this is due to the increased baseline between two ray centers for the outdoor scenes. We also note that the voxel-based ray cloud does not exhibit a significant reduction in pose error compared to the standard ray cloud. + +# 6.2. Quality of revealed scene geometry and images + +The quantitative comparisons of scene reconstruction quality are also shown in Table 1. For a fair comparison, we applied our adapted inversion attack in Sec. 4 for ray cloud and voxel-based ray cloud, while we used the original geometry inversion attack [7] with 3 refinement steps for uniform line cloud (ULC), PPL and $\mathrm{PPL + }$ .We set $\beta = 25\%$ for the ray rejection threshold as it shows the best point estimation in Table 2. Results with $\beta = 50\%$ are included in [32]. + +Recovery of 3D point clouds The median point errors for ray cloud and voxel-based ray cloud are notably larger than those of other methods, even after applying our adjusted density-based attacks. Additionally, to validate the impact of our adapted inversion attack, we compared the geometric error with and without rejection sampling, showing that rejection sampling contributes to error reduction in Table 2. The qualitative result in Fig. 9 shows that the results from ray cloud and voxel-based ray cloud are unrecognizable. + +Degraded quality of scene details Fig. 9 shows the images recovered from the ray cloud and the voxel-based ray cloud are hardly recognizable. In Table. 1, the voxel-based ray cloud yields the lowest PSNR, SSIM and MAE values with the original ray cloud as a runner-up. This indicates large point errors exhibited by ray clouds translate to lower image quality. + +![](images/5642d4b78ddbd722cc01f081a25ccb06d659ffb0084eb81f77db70f02137531b.jpg) + +![](images/5d7d4a00975b879f1f8ee02fcaa94a1077afac9292ecbf3e04a9865961a01d54.jpg) + +![](images/27a4ad7e2e4a5677d827d58c60993e60b4b7ee36271ed13907f6b3cb6fbde635.jpg) + +![](images/0f673db3068e79c2ce0b4021075ce17d09f9a6bf08e575697d53e0ff84ffdd18.jpg) + +![](images/a3155d9364c152845ad8dd9e4cd1ae44ec5237e81da7cd11ab4472044b1ccc2e.jpg) + +![](images/8434caf5240d2be464914b669eeff13aadb2c4d11235891eb4c97d4dcc3469d6.jpg) + +![](images/830a66628c4ef3b7b6cb7f91ee2fa9de75fb0382b3152a1d5c2639228662f83d.jpg) +Figure 10. Average localization performance vs. reconstructed image quality across different levels of the ray density on the Energy Landscape [55] dataset. The tested line densities are $100\%$ (biggest marker), $50\%$ , $25\%$ and $10\%$ (smallest marker). + +![](images/758292b2a65000ee87c6e1a9a95cb05288f5927811877e7f27b5ddaaf04bb487.jpg) + +![](images/0994ad02842ca40f62e80b2cc7d3d15a8e9ffca8af370b61b178bf4b9c275b20.jpg) + +
DatasetRejection threshold (β)eg[m] medianPSNR meanSSIM meanMAE mean
Energy LandscapeNone1.0279.9590.33066.06
10 (%)0.76110.240.32363.62
25 (%)0.66910.290.31962.97
50 (%)0.74610.320.32062.52
CambridgeNone11.4710.100.27364.06
10 (%)10.2910.350.26761.60
25 (%)9.61410.410.26160.98
50 (%)12.989.9440.24264.70
+ +Table 2. Effect of ray rejection (see Sec. 4.2) on the quality of scene recovery. Rejecting rays close to the baseline of two ray centers overall reduces the point estimation errors. Bold represents the best inversion results amongst different thresholds. + +Additionally, Fig. 10 comparing the pose estimation errors against image reconstruction qualities demonstrate that the ray cloud-based approaches are capable of preserving the privacy especially for dense scenes while significantly decreasing the inference runtime. In contrast, the voxel-based ray cloud shows slightly higher pose estimation errors potentially due to coagulation of rays yielding less uniform distribution of rays across the scene. The fact that the SSIM value increases for lower density ray clouds also needs further investigation (see [32] for qualitative results). + +In Fig. 11, we demonstrate some corner cases revealed through our adapted inversion attack in Sec. 4. Fig. 11d shows the voxel-based ray sampling proposed in Sec. 5 can degrade the quality of scene recovery. + +# 7. Conclusion + +We presented a privacy-preserving visual localization framework based on a new scene representation called 3D ray cloud, which is robust to the known geometry inversion attack and achieves an order of magnitude faster localization than previous line-based approaches without significantly compromising the localization accuracy. This is + +![](images/0d983a917d4d9abb839885a7fd993bd44ed790063e17117807e29f89aef6c84c.jpg) +(a) ground truth + +![](images/55b9e92788ca61edade378f77089139dc0824f10b29d680aa9582e8823fc2bb7.jpg) + +![](images/39b2f6884efa3440ead7868be06e85e1090736218f613c692e94517471d8b1c9.jpg) +(b) w/o rejection + +![](images/8582d96efe7f038d7a7399a806e7f1c224b75d9e14c03374bbb8d1702b5b70a7.jpg) + +![](images/29ee5d1d0c067f767159fbe18b77c190e34b72461c5a17fbebb85b54a04f6409.jpg) +(c) w/ rejection + +![](images/4b0ddd2170e980e7e7c1ad24511abae142139e0b2ffaed688b1a7ecbd51bb5b7.jpg) + +![](images/ba40c2e6c34ee9efcd1fe4be39c61baf2c6c8b75f6609ab9fc8f7bc94495f17d.jpg) +(d) voxel-based + +Figure 11. Illustration of point-to-image translation [40] results with point clouds from (a) ground truth, (b) recovered from ray cloud without rejection sampling, (c) recovered with rejection sampling, and (d) recovered from voxel-based ray cloud. While (c) shows improved recovery via ray rejection in Sec. 4.2, (d) shows voxel-based ray sampling lowering the reconstruction quality. + +enabled by the intrinsic geometry of the ray cloud allowing utilization of efficient widely-available perspective relative pose estimation solver. We have also explored potential improvements for the geometry-inversion attack on ray clouds and proposed a respective defense mechanism. + +Limitation Although this work can successfully defend the known inversion attack [7], we note it may be potentially vulnerable to a new type of attack in which omnidirectional rays are projected to a cylindrical or spherical image plane which is then unfolded and directly passed through an inversion network for image reconstruction. While realizing this attack may face its own challenges such as unknown depths and mixed chirality, we believe utilizing more ray centers will provide a defense against such attack by reducing the density of ray projections in each ray center. + +Acknowledgement This work was supported by the NRF (Natioanl Research Foundation of Korea) grants funded by the Korea government (MSIT) (No. 2022R1C1C1004907). We thank anonymous reviewers for their valuable feedback. + +# References + +[1] Sameer Agarwal, Yasutaka Furukawa, Noah Snavely, Ian Simon, Brian Curless, Steven M Seitz, and Richard Szeliski. Building Rome in a day. Communications of the ACM, 54(10):105-112, 2011. 1 +[2] Alexandre Alahi, Raphael Ortiz, and Pierre Vandergheynst. Freak: Fast retina keypoint. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 510-517, 2012. 2 +[3] Clemens Arth, Daniel Wagner, Manfred Klopschitz, Arnold Irschara, and Dieter Schmalstieg. Wide area localization on mobile phones. In 2009 8th IEEE International Symposium on Mixed and Augmented Reality, pages 73-82. IEEE, 2009. 1 +[4] Eric Brachmann, Alexander Krull, Sebastian Nowozin, Jamie Shotton, Frank Michel, Stefan Gumhold, and Carsten Rother. DSAC — Differentiable RANSAC for camera localization. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 2492–2500, 2017. 1 +[5] Eric Brachmann and Carsten Rother. Learning less is more - 6D camera localization via 3D surface regression. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4654-4662, 2018. 1 +[6] Robert Castle, Georg Klein, and David W Murray. Videorelate localization in multiple maps for wearable augmented reality. In 2008 12th IEEE International Symposium on Wearable Computers, pages 15-22. IEEE, 2008. 1 +[7] Kunal Chelani, Fredrik Kahl, and Torsten Sattler. How privacy-preserving are line clouds? Recovering scene details from 3D lines. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 15663-15673, 2021. 1, 2, 3, 4, 5, 6, 7, 8 +[8] Kunal Chelani, Torsten Sattler, Fredrik Kahl, and Zuzana Kukelova. Privacy-preserving representations are not enough: Recovering scene content from camera poses. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13132-13141, 2023. 1 +[9] Ondrej Chum, Jiri Matas, and Josef Kittler. Locally optimized ransac. In Pattern Recognition: 25th DAGM Symposium, Magdeburg, Germany, September 10-12, 2003. Proceedings 25, pages 236-243. Springer, 2003. 1 +[10] Deeksha Dangwal, Vincent T Lee, Hyo Jin Kim, Tianwei Shen, Meghan Cowan, Rajvi Shah, Caroline Trippel, Brandon Reagen, Timothy Sherwood, Vasileios Balntas, et al. Mitigating reverse engineering attacks on local feature descriptors. In Proceeding of the British Machine Vision Conference (BMVC), 2021. 1, 2 +[11] Andrew J. Davison, Ian D. Reid, Nicholas D. Molton, and Olivier Stasse. MonoSLAM: Real-time single camera SLAM. IEEE Transactions on Pattern Analysis and Machine Intelligence, 29(6):1052-1067, 2007. 1 +[12] Yaqing Ding, Jian Yang, Viktor Larsson, Carl Olsson, and Kalle Åström. Revisiting the p3p problem. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4872-4880, 2023. 1 +[13] Alexey Dosovitskiy and Thomas Brox. Inverting visual representations with convolutional networks. In Proceedings + +of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 4829-4837, 2016. 1, 2 +[14] Mihai Dusmanu, Johannes L. Schonberger, Sudipta N. Sinha, and Marc Pollefeys. Privacy-preserving image features via adversarial affine subspace embeddings. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 14262-14272, 2021. 2 +[15] Martin A. Fischler and Robert C. Bolles. Random sample consensus: A paradigm for model fitting with applications to image analysis and automated cartography. Communications of the ACM, 24(6):381-395, June 1981. 1, 3 +[16] Marcel Geppert, Viktor Larsson, Johannes L Schonberger, and Marc Pollefeys. Privacy preserving partial localization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 17337-17347, 2022. 2 +[17] Richard Hartley. Cheirality. International journal of computer vision, 26(1):41-61, 1998. 4 +[18] Lionel Heng, Benjamin Choi, Zhaopeng Cui, Marcel Geppert, Sixing Hu, Benson Kuan, Peidong Liu, Rang Nguyen, Ye Chuan Yeo, Andreas Geiger, et al. Project automation: Localization and 3d scene perception for an autonomous vehicle with a multi-camera system. In Proceedings of the IEEE International Conference on Robotics and Automation (ICRA), pages 4695-4702, 2019. 1 +[19] Hiroharu Kato and Tatsuya Harada. Image reconstruction from bag-of-visual-words. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 955–962, 2014. 2 +[20] Tong Ke and Stergios I Roumeliotis. An efficient algebraic solution to the perspective-three-point problem. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 7225-7233, 2017. 1 +[21] Alex Kendall and Roberto Cipolla. Geometric loss functions for camera pose regression with deep learning. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 6555-6564, 2017. 6 +[22] Alex Kendall, Matthew Grimes, and Roberto Cipolla. PoseNet: A convolutional network for real-time 6-DOF camera relocalization. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), pages 2938-2946, 2015. 1, 6, 7 +[23] Nicolaas H Kuiper. Tests concerning random points on a circle. In Nederl. Akad. Wetensch. Proc. Ser. A, volume 63, pages 38-47, 1960. 3 +[24] Viktor Larsson. PoseLib - Minimal Solvers for Camera Pose Estimation. https://github.com/vlarsson/ PoseLib, 2020. Accessed: 2022-10-30. 6 +[25] Viktor Larsson, Kalle Astrom, and Magnus Oskarsson. Efficient solvers for minimal problems by syzygy-based reduction. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 820-829, 2017. 3 +[26] Chunghwan Lee, Jaihoon Kim, Chanhyuk Yun, and Je Hyeong Hong. Paired-point lifting for enhanced privacy-preserving visual localization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 17266-17275, 2023. 1, 2, 3, 4, + +6,7 +[27] Hyon Lim, Sudipta N Sinha, Michael F Cohen, and Matthew Uytendaele. Real-time image-based 6-DOF localization in large-scale environments. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 1043–1050. IEEE, 2012. 1 +[28] David G. Lowe. Distinctive image features from scale-invariant keypoints. International Journal of Computer Vision, 60(2):91-110, Nov. 2004. 1, 2 +[29] Simon Lynen, Torsten Sattler, Michael Bosse, Joel A Hesch, Marc Pollefeys, and Roland Siegwart. Get out of my lab: Large-scale, real-time visual-inertial localization. In Robotics: Science and Systems, volume 1, page 1, 2015. 1 +[30] Aravindh Mahendran and Andrea Vedaldi. Understanding deep image representations by inverting them. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 5188-5196, 2015. 1, 2 +[31] Anastasiia Mishchuk, Dmytro Mishkin, Filip Radenovic, and Jiri Matas. Working hard to know your neighbor's margins: Local descriptor learning loss. Advances in Neural Information Processing Systems (NeurIPS), 30, 2017. 4 +[32] Heejoon Moon, Chunghwan Lee, and Je Hyeong Hong. Supplementary document for efficient privacy-preserving visual localization using 3d ray clouds, 2024. 3, 5, 6, 7, 8 +[33] Raul Mur-Artal, Jose Maria Martinez Montiel, and Juan D Tardos. ORB-SLAM: A versatile and accurate monocular SLAM system. IEEE Transactions on Robotics, 31(5):1147-1163, 2015. 1 +[34] Tony Ng, Hyo Jin Kim, Vincent T Lee, Daniel DeTone, Tsun-Yi Yang, Tianwei Shen, Eddy Ilg, Vassileios Balntas, Krystian Mikolajczyk, and Chris Sweeney. NinjaDesc: Content-concealing visual descriptors via adversarial learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12797-12807, 2022. 2 +[35] David Nister. An efficient solution to the five-point relative pose problem. IEEE Transactions on Pattern Analysis and Machine Intelligence, 26(6):756-770, 2004. 2, 3, 4 +[36] Linfei Pan, Johannes Lutz Schonberger, Viktor Larsson, and Marc Pollefeys. Privacy Preserving Localization via Coordinate Permutations. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), 2023. 2, 3, 4 +[37] Mikael Persson and Klas Nordberg. Lambda twist: An accurate fast robust perspective three point (p3P) solver. In Proceedings of the European conference on computer vision (ECCV), pages 318-332, 2018. 1 +[38] Maxime Pietrantoni, Martin Humenberger, Torsten Sattler, and Gabriela Csurka. Segloc: Learning segmentation-based representations for privacy-preserving visual localization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15380-15391, 2023. 2 +[39] Francesco Pittaluga, Sanjeev Koppal, and Ayan Chakrabarti. Learning privacy preserving encodings through adversarial training. In Proceedings of the IEEE Winter Conference on Applications of Computer Vision (WACV), pages 791-799. IEEE, 2019. 1, 7 +[40] Francesco Pittaluga, Sanjeev J Koppal, Sing Bing Kang, and Sudipta N Sinha. Revealing scenes by inverting structure from motion reconstructions. In Proceedings of the + +IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 145-154, 2019. 1, 2, 3, 8 +[41] Francesco Pittaluga and Bingbing Zhuang. Ldp-feat: Image features with local differential privacy. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 17580–17590, 2023. 2 +[42] Rahul Raguram, Ondrej Chum, Marc Pollefeys, Jiri Matas, and Jan-Michael Frahm. USAC: A universal framework for random sample consensus. IEEE Transactions on Pattern Analysis and Machine Intelligence, 35(8):2022-2038, 2012. 1 +[43] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-Net: Convolutional networks for biomedical image segmentation. In Proceedings of the International Conference on Medical Image Computing and Computer-Assisted Intervention (MICCAI), pages 234–241. Springer, 2015. 1, 2 +[44] Johannes Lutz Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 1, 6 +[45] Johannes Lutz Schonberger, Enliang Zheng, Marc Pollefeys, and Jan-Michael Frahm. Pixelwise view selection for unstructured multi-view stereo. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part III 14, pages 501-518. Springer, 2016. 1 +[46] Mikiya Shibuya, Shinya Sumikura, and Ken Sakurada. Privacy preserving visual SLAM. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXII 16, pages 102-118. Springer, 2020. 1, 2, 3 +[47] Sudipta Narayan Sinha, Pablo Alejandro Speciale, Sing Bing Kang, and Marc Andre Leon Pollefeys. Device pose estimation using 3d line clouds, Mar. 30 2021. US Patent 10,964,053. 3 +[48] Noah Snavely, Steven M Seitz, and Richard Szeliski. Modeling the world from internet photo collections. International Journal of Computer Vision (IJCV), 80:189-210, 2008. 1 +[49] Zhenbo Song, Wayne Chen, Dylan Campbell, and Hongdong Li. Deep novel view synthesis from colored 3d point clouds. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXIV 16, pages 1-17. Springer, 2020. 2 +[50] Pablo Speciale, Johannes L Schonberger, Sing Bing Kang, Sudipta N Sinha, and Marc Pollefeys. Privacy preserving image-based localization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5493-5503, 2019. 1, 2, 3, 4, 5, 6, 7 +[51] Pablo Speciale, Johannes L Schonberger, Sudipta N Sinha, and Marc Pollefeys. Privacy preserving image queries for camera localization. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 1486-1496, 2019. 6 +[52] Henrik Stewénius, Magnus Oskarsson, Kalle Aström, and David Nister. Solutions to minimal generalized relative pose problems, 2005. 3, 4 +[53] Yurun Tian, Xin Yu, Bin Fan, Fuchao Wu, Huub Heijnen, and Vassileios Balntas. SOSNet: Second order similarity regularization for local descriptor learning. In Proceedings of + +the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11016-11025, 2019. 2 +[54] Philip HS Torr and Andrew Zisserman. Mlesac: A new robust estimator with application to estimating image geometry. Computer vision and image understanding, 78(1):138-156, 2000. 4 +[55] Julien Valentin, Angela Dai, Matthias Niessner, Pushmeet Kohli, Philip Torr, Shahram Izadi, and Cem Keskin. Learning to navigate the energy landscape. In 2016 Fourth International Conference on 3D Vision (3DV), pages 323-332, 2016. 1, 3, 5, 6, 7, 8 +[56] Jonathan Ventura, Clemens Arth, and Vincent Lepetit. An efficient minimal solution for multi-camera motion. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), pages 747-755, 2015. 3 +[57] Carl Vondrick, Aditya Khosla, Tomasz Malisiewicz, and Antonio Torralba. HOGgles: Visualizing object detection features. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), pages 1-8, 2013. 1, 2 +[58] Philippe Weinzaepfel, Herve Jégou, and Patrick Pérez. Reconstructing an image from its local descriptors. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 337-344, 2011. 1, 2 +[59] Enliang Zheng and Changchang Wu. Structure from motion using structure-less resection. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), pages 2075-2083, 2015. 2, 4 \ No newline at end of file diff --git a/2024/Efficient Privacy-Preserving Visual Localization Using 3D Ray Clouds/images.zip b/2024/Efficient Privacy-Preserving Visual Localization Using 3D Ray Clouds/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..491be191b825d6e3ee50579e8fcf921345720c2b --- /dev/null +++ b/2024/Efficient Privacy-Preserving Visual Localization Using 3D Ray Clouds/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ae65fb7ccb6be02e1db9ac3e5723d62050a11874b2086f2733f53badf14af544 +size 640773 diff --git a/2024/Efficient Privacy-Preserving Visual Localization Using 3D Ray Clouds/layout.json b/2024/Efficient Privacy-Preserving Visual Localization Using 3D Ray Clouds/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..6da841191a1962ff8a53b486cf66a6ebfc1532e6 --- /dev/null +++ b/2024/Efficient Privacy-Preserving Visual Localization Using 3D Ray Clouds/layout.json @@ -0,0 +1,12427 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 80, + 103, + 514, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 80, + 103, + 514, + 121 + ], + "spans": [ + { + "bbox": [ + 80, + 103, + 514, + 121 + ], + "type": "text", + "content": "Efficient Privacy-Preserving Visual Localization Using 3D Ray Clouds" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 162, + 142, + 452, + 158 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 162, + 142, + 452, + 158 + ], + "spans": [ + { + "bbox": [ + 162, + 142, + 452, + 158 + ], + "type": "text", + "content": "Heejoon Moon1 Chunghwan Lee2 Je Hyeong Hong1,2†" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 157, + 559, + 172 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 157, + 559, + 172 + ], + "spans": [ + { + "bbox": [ + 53, + 157, + 559, + 172 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 53, + 157, + 559, + 172 + ], + "type": "text", + "content": "Dept. Artificial Intelligence, Hanyang University " + }, + { + "bbox": [ + 53, + 157, + 559, + 172 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 53, + 157, + 559, + 172 + ], + "type": "text", + "content": "Dept. Electronic Engineering, Hanyang University" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 143, + 187, + 192, + 201 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 187, + 192, + 201 + ], + "spans": [ + { + "bbox": [ + 143, + 187, + 192, + 201 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 214, + 290, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 214, + 290, + 514 + ], + "spans": [ + { + "bbox": [ + 47, + 214, + 290, + 514 + ], + "type": "text", + "content": "The recent success in revealing scene details from sparse 3D point clouds obtained via structure-from-motion has raised significant privacy concerns in visual localization. One prominent approach for mitigating this issue is to lift 3D points to 3D lines thereby reducing the effectiveness of the scene inversion attacks, but this comes at the cost of increased algorithmic complexity for camera localization due to weaker geometric constraints induced by line clouds. To overcome this limitation, we propose a new lifting approach called \"ray cloud\", whereby each lifted 3D line intersects at one of two predefined locations, depicting omnidirectional rays from two cameras. This yields two benefits, i) camera localization can now be cast as relative pose estimation between the query image and the calibrated rig of two perspective cameras which can be efficiently solved using a variant of the 5-point algorithm, and ii) the ray cloud introduces erroneous estimations for the density-based inversion attack, degrading the quality of scene recovery. Moreover, we explore possible modifications of the inversion attack to better recover scenes from the ray clouds and propose a ray sampling technique to reduce the effectiveness of the modified attack. Experimental results on two public datasets show real-time localization speed as well as enhanced privacy-preserving capability over the state-of-the-art without overly sacrificing the localization accuracy." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 526, + 128, + 540 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 526, + 128, + 540 + ], + "spans": [ + { + "bbox": [ + 47, + 526, + 128, + 540 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 548, + 287, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 548, + 287, + 693 + ], + "spans": [ + { + "bbox": [ + 47, + 548, + 287, + 693 + ], + "type": "text", + "content": "Visual localization refers to the task of estimating the camera pose w.r.t. a known 3D scene given a query image. It serves a central role for cutting-edge applications involving robotics [18], augmented reality (AR) [3], virtual reality (VR) [6] and simultaneous localization and mapping (SLAM) [11, 29, 33]. To this date, many visual localization algorithms [27, 37] still utilize sparse 3D point clouds constructed via structure-from-motion (SfM) [1, 44, 45, 48] through which the camera pose is estimated from the correspondences between 2D image features and 3D points. This is commonly solved using a perspective-n-point (PnP) solver [12, 20, 37] equipped with the random sample con" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 313, + 191, + 417, + 252 + ], + "blocks": [ + { + "bbox": [ + 313, + 191, + 417, + 252 + ], + "lines": [ + { + "bbox": [ + 313, + 191, + 417, + 252 + ], + "spans": [ + { + "bbox": [ + 313, + 191, + 417, + 252 + ], + "type": "image", + "image_path": "366bd3567fcbb48e266aa3fc72bbebef5752819862d8dd07d452697953d11b36.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 320, + 255, + 408, + 264 + ], + "lines": [ + { + "bbox": [ + 320, + 255, + 408, + 264 + ], + "spans": [ + { + "bbox": [ + 320, + 255, + 408, + 264 + ], + "type": "text", + "content": "(a) Illustration of ray cloud" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 422, + 191, + 533, + 252 + ], + "blocks": [ + { + "bbox": [ + 422, + 191, + 533, + 252 + ], + "lines": [ + { + "bbox": [ + 422, + 191, + 533, + 252 + ], + "spans": [ + { + "bbox": [ + 422, + 191, + 533, + 252 + ], + "type": "image", + "image_path": "c7dfd1040ceb28268dd4cf0d37b2dbf182c1696678ef83c6c8dde24ffb6421d3.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 423, + 255, + 541, + 264 + ], + "lines": [ + { + "bbox": [ + 423, + 255, + 541, + 264 + ], + "spans": [ + { + "bbox": [ + 423, + 255, + 541, + 264 + ], + "type": "text", + "content": "(b) Single-image localization speeds" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 315, + 266, + 375, + 371 + ], + "blocks": [ + { + "bbox": [ + 315, + 266, + 375, + 371 + ], + "lines": [ + { + "bbox": [ + 315, + 266, + 375, + 371 + ], + "spans": [ + { + "bbox": [ + 315, + 266, + 375, + 371 + ], + "type": "image", + "image_path": "db21b4eba16d5d0cd9fc42a7940145b6b5bdaaf47063920ce1ad2c8b5e7f5391.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 333, + 371, + 363, + 377 + ], + "lines": [ + { + "bbox": [ + 333, + 371, + 363, + 377 + ], + "spans": [ + { + "bbox": [ + 333, + 371, + 363, + 377 + ], + "type": "text", + "content": "Ground truth" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 334, + 380, + 518, + 390 + ], + "lines": [ + { + "bbox": [ + 334, + 380, + 518, + 390 + ], + "spans": [ + { + "bbox": [ + 334, + 380, + 518, + 390 + ], + "type": "text", + "content": "(c) Comparison of revealed scene geometries and images" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 305, + 392, + 547, + 468 + ], + "lines": [ + { + "bbox": [ + 305, + 392, + 547, + 468 + ], + "spans": [ + { + "bbox": [ + 305, + 392, + 547, + 468 + ], + "type": "text", + "content": "Figure 1. Main characteristics of the proposed ray cloud. (a) shows 3D points lifted as 3D rays half of which intersects at center 1 and the other half intersects at center 2. Estimating the camera pose resembles relative pose estimation with respect to a calibrated rig of perspective cameras. As shown in (b), this allows real-time single-image-based localization on two public datasets [22, 55] while preventing scene-revealing attacks [7, 40] as shown in (c)." + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 376, + 266, + 430, + 371 + ], + "blocks": [ + { + "bbox": [ + 376, + 266, + 430, + 371 + ], + "lines": [ + { + "bbox": [ + 376, + 266, + 430, + 371 + ], + "spans": [ + { + "bbox": [ + 376, + 266, + 430, + 371 + ], + "type": "image", + "image_path": "50f1964bde8163e009f1e75f57c5a4070bac7fe0040adf812122015850fb3c84.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 383, + 371, + 424, + 377 + ], + "lines": [ + { + "bbox": [ + 383, + 371, + 424, + 377 + ], + "spans": [ + { + "bbox": [ + 383, + 371, + 424, + 377 + ], + "type": "text", + "content": "Uniform line cloud" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 430, + 371, + 486, + 377 + ], + "lines": [ + { + "bbox": [ + 430, + 371, + 486, + 377 + ], + "spans": [ + { + "bbox": [ + 430, + 371, + 486, + 377 + ], + "type": "text", + "content": "Paired-Point Lifting (PPL)" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 431, + 266, + 485, + 371 + ], + "blocks": [ + { + "bbox": [ + 431, + 266, + 485, + 371 + ], + "lines": [ + { + "bbox": [ + 431, + 266, + 485, + 371 + ], + "spans": [ + { + "bbox": [ + 431, + 266, + 485, + 371 + ], + "type": "image", + "image_path": "048e209533ed2f407f8bf82451cb5f2df28b6a098347615188de72ae8ddc131f.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 486, + 266, + 541, + 371 + ], + "blocks": [ + { + "bbox": [ + 486, + 266, + 541, + 371 + ], + "lines": [ + { + "bbox": [ + 486, + 266, + 541, + 371 + ], + "spans": [ + { + "bbox": [ + 486, + 266, + 541, + 371 + ], + "type": "image", + "image_path": "8fc928a23ad2fb20720bdf4c923ca6857711f8f30cb9ea18900c33b70e85d60b.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 493, + 371, + 532, + 377 + ], + "lines": [ + { + "bbox": [ + 493, + 371, + 532, + 377 + ], + "spans": [ + { + "bbox": [ + 493, + 371, + 532, + 377 + ], + "type": "text", + "content": "Ray cloud (ours)" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "bbox": [ + 305, + 473, + 545, + 497 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 473, + 545, + 497 + ], + "spans": [ + { + "bbox": [ + 305, + 473, + 545, + 497 + ], + "type": "text", + "content": "sensus (RANSAC) meta-algorithm [9, 15, 42] or its deep network alternative such as differentiable RANSAC [4, 5]." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 304, + 498, + 546, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 498, + 546, + 617 + ], + "spans": [ + { + "bbox": [ + 304, + 498, + 546, + 617 + ], + "type": "text", + "content": "Surprisingly, it was recently shown [8, 10, 13, 30, 40, 57, 58] that a sparse 3D point cloud may contain enough information to reveal image details of the underlying scene. Most prominently, Pittaluga et al. [40] demonstrated a cascaded U-Net model [43] can generate high-fidelity images of the scene just from the 2D projection of sparse 3D points and their SIFT features [28]. This raised significant privacy concerns [39, 50] as anyone with access to the 3D point cloud could potentially search for sensitive details in the private scene by employing this scene inversion network." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 304, + 618, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 618, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 618, + 547, + 714 + ], + "type": "text", + "content": "Currently, the mainstream approach for preventing the aforementioned image inversion attack on the sparse 3D point cloud is geometric lifting [26, 46, 50], whereby each 3D point is transformed to a 3D line passing through the original point with aims to conceal the 3D scene geometry (see Fig. 2). While the earlier approach of randomly drawing the line direction (also known as the uniform line cloud) was shown to be susceptible to geometry inversion attack [7]," + } + ] + } + ], + "index": 25 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 62, + 702, + 138, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 702, + 138, + 713 + ], + "spans": [ + { + "bbox": [ + 62, + 702, + 138, + 713 + ], + "type": "text", + "content": "† Corresponding author" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "9773" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 286, + 132 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 286, + 132 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 286, + 132 + ], + "type": "text", + "content": "this was remedied by recent methods utilizing pairing of 3D points [26, 36] to increase ambiguity in the lifted scene representation. Nevertheless, single image-based localization speeds for these algorithms [26, 36] are far from achieving real-time performance, providing motivation for this work." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 133, + 286, + 299 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 133, + 286, + 299 + ], + "spans": [ + { + "bbox": [ + 46, + 133, + 286, + 299 + ], + "type": "text", + "content": "In this paper, we present a new geometric lifting approach called ray cloud with aims to address above issue of slow camera localization runtime exhibited by previous studies. As will be illustrated in Sec. 3, ray cloud is simple to construct and robust to the geometry inversion attack [7]. Most importantly, camera localization using a ray cloud can be formulated as a well-known relative pose estimation problem between the query image and a virtual calibrated rig of cameras, enabling utilization of the efficient 5-point algorithm [35] with known scale [59]. Moreover, by modifying and enhancing the geometry inversion attack [7] in Sec. 4, we explore potential hidden privacy risks associated with ray clouds, and subsequently propose a technique to enhance the privacy protection of ray clouds in Sec. 5." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 300, + 278, + 311 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 300, + 278, + 311 + ], + "spans": [ + { + "bbox": [ + 58, + 300, + 278, + 311 + ], + "type": "text", + "content": "The contributions of this work are summarized below:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 312, + 286, + 455 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 47, + 312, + 286, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 312, + 286, + 335 + ], + "spans": [ + { + "bbox": [ + 47, + 312, + 286, + 335 + ], + "type": "text", + "content": "- an efficient privacy-preserving visual localization framework based on a new lifting approach called 3D ray cloud," + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 335, + 286, + 358 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 335, + 286, + 358 + ], + "spans": [ + { + "bbox": [ + 47, + 335, + 286, + 358 + ], + "type": "text", + "content": "- customization of the geometry inversion attack [7] to yield undesired (privacy leak) corner cases for ray clouds," + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 360, + 286, + 395 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 360, + 286, + 395 + ], + "spans": [ + { + "bbox": [ + 47, + 360, + 286, + 395 + ], + "type": "text", + "content": "- a voxel-based ray sampling technique to prevent above privacy leaks by disrupting nearest neighbor statistic and inducing erroneous recovery of the scene points, and" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 396, + 286, + 455 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 396, + 286, + 455 + ], + "spans": [ + { + "bbox": [ + 47, + 396, + 286, + 455 + ], + "type": "text", + "content": "- extensive experimental comparison of ray cloud against other line cloud-based approaches on two public datasets, achieving real-time single-image based localization speed for the first time while preserving privacy-preserving capability and maintaining similar localization accuracy." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 47, + 464, + 130, + 475 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 464, + 130, + 475 + ], + "spans": [ + { + "bbox": [ + 47, + 464, + 130, + 475 + ], + "type": "text", + "content": "2. Related work" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 483, + 286, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 483, + 286, + 590 + ], + "spans": [ + { + "bbox": [ + 46, + 483, + 286, + 590 + ], + "type": "text", + "content": "Privacy threats from image inversion attack While several prior studies addressed the possibility of revealing images from sparse keypoints [13, 19, 30, 57, 58], it was Pitaluga et al. [40] who posed a real privacy threat for the first time by recovering detailed scene images from a sparse 3D point cloud using a cascaded U-Net [43]. This network called InvSfM takes in the 2D locations of the projected 3D points as well as the corresponding depths and SIFT descriptors, and outputs a synthesized image of the scene." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 592, + 286, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 592, + 286, + 687 + ], + "spans": [ + { + "bbox": [ + 46, + 592, + 286, + 687 + ], + "type": "text", + "content": "More recently, Song et al. [49] used a PointNet++ architecture to extract deep features and synthesize scene images. Dangwal et al. [10] devised a generic inversion model that can work with various descriptors [2, 28, 53] and proposed selective feature suppression to hide privacy-sensitive regions. Nonetheless, the pretrained InvSfM model still serves as the baseline for qualitative and quantitative analyses in privacy-preserving visual localization [7, 26, 36, 38]." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 689, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 689, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 689, + 286, + 713 + ], + "type": "text", + "content": "Privacy-preserving scene representations After the emergence of InvSfM [40], Speciale et al. [50] quickly em" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 72, + 545, + 179 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 179 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 179 + ], + "type": "text", + "content": "braced a geometric lifting approach, also known as the uniform line cloud, where each 3D point in the original sparse point cloud is replaced by a 3D line passing through the point with its direction drawn from a uniform distribution on the unit sphere. This scene representation aims to introduce anonymity to the locations of the 3D points thereby preventing any image inversion via InvSfM. The work was later extended by Shibuya et al. [46] to visual SLAM with a prebuilt map provided in the form of uniform line cloud." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 180, + 545, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 180, + 545, + 407 + ], + "spans": [ + { + "bbox": [ + 304, + 180, + 545, + 407 + ], + "type": "text", + "content": "Unfortunately, above methods [46, 50] are now outdated and susceptible to the geometry inversion attack proposed by Chelani et al. [7], which successfully recovers the sparse 3D point cloud from a uniform line cloud (details in the next subsection). To overcome this limitation, Lee et al. [26] proposed to draw 3D lines through random pairs of 3D points, thereby introducing non-uniform line distributions and multiple 3D point candidates for each line both of which are empirically shown to hinder point cloud inversion. Pan et al. [36] proposed to permute the coordinates of random pairs of keypoints to preserve privacy while still enabling localization from multiple axis-aligned lines. Geppert et al. [16] presented a partial localization approach whereby three 1-dimensional partial maps are constructed from the 3D points and separately stored to enhance security. While these approaches can evade the geometry inversion attack, they are not runtime-efficient to allow real-time camera localization from single images. This issue is addressed through use of the ray cloud proposed in this work." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 407, + 545, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 407, + 545, + 456 + ], + "spans": [ + { + "bbox": [ + 304, + 407, + 545, + 456 + ], + "type": "text", + "content": "Other explored approaches include keypoint lifting [14], carving [34] or obfuscation [41], whereby the feature descriptors of the 3D points are transformed to hide keypoint details, but this does not conceal the point cloud geometry." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 460, + 545, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 460, + 545, + 555 + ], + "spans": [ + { + "bbox": [ + 304, + 460, + 545, + 555 + ], + "type": "text", + "content": "Privacy threats for line clouds Chelani et al. [7] presented a density-based inversion attack that can reveal the 3D point geometry from a uniform line cloud [50]. While the posterior probability of the point cloud " + }, + { + "bbox": [ + 304, + 460, + 545, + 555 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 304, + 460, + 545, + 555 + ], + "type": "text", + "content": " given a line cloud " + }, + { + "bbox": [ + 304, + 460, + 545, + 555 + ], + "type": "inline_equation", + "content": "\\mathcal{L}" + }, + { + "bbox": [ + 304, + 460, + 545, + 555 + ], + "type": "text", + "content": " can be expressed as " + }, + { + "bbox": [ + 304, + 460, + 545, + 555 + ], + "type": "inline_equation", + "content": "P(\\mathcal{P}|\\mathcal{L})\\propto P(\\mathcal{P})P(\\mathcal{L}|\\mathcal{P})" + }, + { + "bbox": [ + 304, + 460, + 545, + 555 + ], + "type": "text", + "content": " using Bayes' rule, it is noted that the likelihood " + }, + { + "bbox": [ + 304, + 460, + 545, + 555 + ], + "type": "inline_equation", + "content": "P(\\mathcal{L}|\\mathcal{P})" + }, + { + "bbox": [ + 304, + 460, + 545, + 555 + ], + "type": "text", + "content": " is constant for uniform line clouds. Hence, maximizing the posterior amounts to maximizing the marginal prior " + }, + { + "bbox": [ + 304, + 460, + 545, + 555 + ], + "type": "inline_equation", + "content": "P(\\mathcal{P})" + }, + { + "bbox": [ + 304, + 460, + 545, + 555 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 556, + 545, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 556, + 545, + 665 + ], + "spans": [ + { + "bbox": [ + 304, + 556, + 545, + 665 + ], + "type": "text", + "content": "For maximizing " + }, + { + "bbox": [ + 304, + 556, + 545, + 665 + ], + "type": "inline_equation", + "content": "P(\\mathcal{P})" + }, + { + "bbox": [ + 304, + 556, + 545, + 665 + ], + "type": "text", + "content": ", Chelani et al. [7] resorted to an empirical observation that for any two 3D points " + }, + { + "bbox": [ + 304, + 556, + 545, + 665 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_A" + }, + { + "bbox": [ + 304, + 556, + 545, + 665 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 556, + 545, + 665 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_B" + }, + { + "bbox": [ + 304, + 556, + 545, + 665 + ], + "type": "text", + "content": " with their corresponding uniformly-drawn lines " + }, + { + "bbox": [ + 304, + 556, + 545, + 665 + ], + "type": "inline_equation", + "content": "\\mathbf{l}_A \\in Gr(1,3)" + }, + { + "bbox": [ + 304, + 556, + 545, + 665 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 556, + 545, + 665 + ], + "type": "inline_equation", + "content": "\\mathbf{l}_B \\in Gr(1,3)" + }, + { + "bbox": [ + 304, + 556, + 545, + 665 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 556, + 545, + 665 + ], + "type": "inline_equation", + "content": "P(\\| \\mathbf{x}_A - \\mathbf{x}_A^\\perp \\|_2 < \\| \\mathbf{x}_A - \\mathbf{x}_B \\|_2) \\approx 0.8" + }, + { + "bbox": [ + 304, + 556, + 545, + 665 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 556, + 545, + 665 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_A^\\perp" + }, + { + "bbox": [ + 304, + 556, + 545, + 665 + ], + "type": "text", + "content": " denotes the point on " + }, + { + "bbox": [ + 304, + 556, + 545, + 665 + ], + "type": "inline_equation", + "content": "\\mathbf{l}_A" + }, + { + "bbox": [ + 304, + 556, + 545, + 665 + ], + "type": "text", + "content": " closest to " + }, + { + "bbox": [ + 304, + 556, + 545, + 665 + ], + "type": "inline_equation", + "content": "\\mathbf{l}_B" + }, + { + "bbox": [ + 304, + 556, + 545, + 665 + ], + "type": "text", + "content": ". This implies that if " + }, + { + "bbox": [ + 304, + 556, + 545, + 665 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_A" + }, + { + "bbox": [ + 304, + 556, + 545, + 665 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 556, + 545, + 665 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_B" + }, + { + "bbox": [ + 304, + 556, + 545, + 665 + ], + "type": "text", + "content": " are close to each other, then " + }, + { + "bbox": [ + 304, + 556, + 545, + 665 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_A^\\perp" + }, + { + "bbox": [ + 304, + 556, + 545, + 665 + ], + "type": "text", + "content": " is likely to become a good approximation of " + }, + { + "bbox": [ + 304, + 556, + 545, + 665 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_A" + }, + { + "bbox": [ + 304, + 556, + 545, + 665 + ], + "type": "text", + "content": ". Hence, " + }, + { + "bbox": [ + 304, + 556, + 545, + 665 + ], + "type": "inline_equation", + "content": "P(\\mathcal{P})" + }, + { + "bbox": [ + 304, + 556, + 545, + 665 + ], + "type": "text", + "content": " can be maximized if the nearest neighboring points of each 3D point are known in advance." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "content": "As it is non-trivial to find a set of true " + }, + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "content": "-nearest neighboring points in line clouds, Chelani et al. [7] approximates this by the set of " + }, + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "content": "-nearest lines " + }, + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{l}_K\\}" + }, + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "content": " for each line " + }, + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{l}_i" + }, + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "content": ", expecting some to include the true positives of the " + }, + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "content": "-nearest" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "9774" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 53, + 70, + 171, + 137 + ], + "blocks": [ + { + "bbox": [ + 53, + 70, + 171, + 137 + ], + "lines": [ + { + "bbox": [ + 53, + 70, + 171, + 137 + ], + "spans": [ + { + "bbox": [ + 53, + 70, + 171, + 137 + ], + "type": "image", + "image_path": "a7397179caebe7e424db836405385e05a2671ff62b4d64c2b02abc4cbc3a7615.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 86, + 142, + 135, + 152 + ], + "lines": [ + { + "bbox": [ + 86, + 142, + 135, + 152 + ], + "spans": [ + { + "bbox": [ + 86, + 142, + 135, + 152 + ], + "type": "text", + "content": "(a) Point cloud" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 176, + 70, + 296, + 137 + ], + "blocks": [ + { + "bbox": [ + 176, + 70, + 296, + 137 + ], + "lines": [ + { + "bbox": [ + 176, + 70, + 296, + 137 + ], + "spans": [ + { + "bbox": [ + 176, + 70, + 296, + 137 + ], + "type": "image", + "image_path": "d4b8d335014fe21b9c8c72cfbadff0d8f016075d5650b5b70836abc65f141799.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 190, + 142, + 280, + 152 + ], + "lines": [ + { + "bbox": [ + 190, + 142, + 280, + 152 + ], + "spans": [ + { + "bbox": [ + 190, + 142, + 280, + 152 + ], + "type": "text", + "content": "(b) Uniform line cloud [50]" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 302, + 70, + 419, + 137 + ], + "blocks": [ + { + "bbox": [ + 302, + 70, + 419, + 137 + ], + "lines": [ + { + "bbox": [ + 302, + 70, + 419, + 137 + ], + "spans": [ + { + "bbox": [ + 302, + 70, + 419, + 137 + ], + "type": "image", + "image_path": "481016dfcf7ad1fdfc014557093e47a2c40c5ab8c519a78c0c4d61f8a7416f00.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 142, + 416, + 152 + ], + "lines": [ + { + "bbox": [ + 304, + 142, + 416, + 152 + ], + "spans": [ + { + "bbox": [ + 304, + 142, + 416, + 152 + ], + "type": "text", + "content": "(c) Paired-Point Lifting (PPL) [26]" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 425, + 70, + 542, + 137 + ], + "blocks": [ + { + "bbox": [ + 425, + 70, + 542, + 137 + ], + "lines": [ + { + "bbox": [ + 425, + 70, + 542, + 137 + ], + "spans": [ + { + "bbox": [ + 425, + 70, + 542, + 137 + ], + "type": "image", + "image_path": "140d4d62b1209a412c9ae1a1018164b819389e7169869e78ae463663fb174d3f.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 449, + 142, + 515, + 152 + ], + "lines": [ + { + "bbox": [ + 449, + 142, + 515, + 152 + ], + "spans": [ + { + "bbox": [ + 449, + 142, + 515, + 152 + ], + "type": "text", + "content": "(d) Ray cloud (ours)" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 183, + 287, + 301 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 183, + 287, + 301 + ], + "spans": [ + { + "bbox": [ + 46, + 183, + 287, + 301 + ], + "type": "text", + "content": "points when " + }, + { + "bbox": [ + 46, + 183, + 287, + 301 + ], + "type": "inline_equation", + "content": "K > k" + }, + { + "bbox": [ + 46, + 183, + 287, + 301 + ], + "type": "text", + "content": " is large enough (we distinguish " + }, + { + "bbox": [ + 46, + 183, + 287, + 301 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 46, + 183, + 287, + 301 + ], + "type": "text", + "content": " for lines and " + }, + { + "bbox": [ + 46, + 183, + 287, + 301 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 183, + 287, + 301 + ], + "type": "text", + "content": " for points). Bearing this assumption, [7] computes the closest point " + }, + { + "bbox": [ + 46, + 183, + 287, + 301 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{k}^{\\perp}" + }, + { + "bbox": [ + 46, + 183, + 287, + 301 + ], + "type": "text", + "content": " to each of the " + }, + { + "bbox": [ + 46, + 183, + 287, + 301 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 46, + 183, + 287, + 301 + ], + "type": "text", + "content": "-nearest lines for the current line " + }, + { + "bbox": [ + 46, + 183, + 287, + 301 + ], + "type": "inline_equation", + "content": "\\mathbf{l}_i" + }, + { + "bbox": [ + 46, + 183, + 287, + 301 + ], + "type": "text", + "content": ", builds a histogram of closest points " + }, + { + "bbox": [ + 46, + 183, + 287, + 301 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{x}_{k}^{\\perp}\\}" + }, + { + "bbox": [ + 46, + 183, + 287, + 301 + ], + "type": "text", + "content": " and finds the peak of the histogram via Kuiper's test [23] to estimate the point " + }, + { + "bbox": [ + 46, + 183, + 287, + 301 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_i" + }, + { + "bbox": [ + 46, + 183, + 287, + 301 + ], + "type": "text", + "content": ". This attack is effective for uniform line clouds [50], allowing recovered points to be used for scene recovery with InvSfM [40]. However, it is less effective against more recent line-based approaches [26, 36] and ray clouds presented in this work." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 311, + 288, + 526 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 311, + 288, + 526 + ], + "spans": [ + { + "bbox": [ + 46, + 311, + 288, + 526 + ], + "type": "text", + "content": "Camera localization using line clouds As mentioned in Sec. 1, camera localization typically involves a RANSAC algorithm [15] coupled with an efficient minimal solver that can run fast over thousands of iterations [25]. For single-image based localization, a standard line cloud introduces 2D points-to-3D line constraints, which are weaker than 2D point-to-3D point constraints provided by the point clouds. Consequently, the minimal solver for line clouds (p6L) requires 6 correspondences between 2D points and 3D lines [50]. Since p6L innately uses the solver for generalized relative pose estimation [52], achieving real-time speed with " + }, + { + "bbox": [ + 46, + 311, + 288, + 526 + ], + "type": "inline_equation", + "content": "p6L" + }, + { + "bbox": [ + 46, + 311, + 288, + 526 + ], + "type": "text", + "content": " is challenging due to the large number of possible solutions (up to 64) and high algorithmic complexity [52, 56]. Shibuya et al. [46] achieved real-time localization with line clouds in the SLAM setup, but this requires a video stream as input and construction of a client-side point map. On the other hand, we will show it is possible to perform image-based localization in real-time using ray clouds." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 540, + 132, + 553 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 540, + 132, + 553 + ], + "spans": [ + { + "bbox": [ + 47, + 540, + 132, + 553 + ], + "type": "text", + "content": "3. 3D Ray cloud" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 561, + 287, + 597 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 561, + 287, + 597 + ], + "spans": [ + { + "bbox": [ + 46, + 561, + 287, + 597 + ], + "type": "text", + "content": "We now illustrate the process of constructing a 3D ray cloud shown in Fig. 2 and estimating camera pose from ray clouds. We also discuss inversion attack [7] on ray clouds." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "type": "text", + "content": "Motivation The concept of lifting 3D lines to intersect at a particular point has been briefly addressed without detailed investigation in [7, 47]. Chelani et al. [7] mentioned that, if all lines were to intersect at a single point, then the geometry inversion attack in Sec. 2 would fail. This is because the histogram of closest points (to other lines) becomes concentrated about the point of intersection, leading to a degenerate solution. Unfortunately, this representation possesses an intrinsic limitation that the camera pose can" + } + ] + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 315, + 182, + 389, + 268 + ], + "blocks": [ + { + "bbox": [ + 46, + 154, + 544, + 176 + ], + "lines": [ + { + "bbox": [ + 46, + 154, + 544, + 176 + ], + "spans": [ + { + "bbox": [ + 46, + 154, + 544, + 176 + ], + "type": "text", + "content": "Figure 2. 3D line clouds yielded from different lifting approaches on the Office1 manolis in the Energy Landscape [55] dataset. Note each line (ray) in the ray cloud passes through one of two center points. We show " + }, + { + "bbox": [ + 46, + 154, + 544, + 176 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 46, + 154, + 544, + 176 + ], + "type": "text", + "content": " of the total number of lines for visual comparison." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 315, + 182, + 389, + 268 + ], + "lines": [ + { + "bbox": [ + 315, + 182, + 389, + 268 + ], + "spans": [ + { + "bbox": [ + 315, + 182, + 389, + 268 + ], + "type": "image", + "image_path": "ad711c4d29a0eb1a16d0bb6d396b996c112abb08b9da1c5cfc9fc5c93e7c7de0.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 314, + 270, + 387, + 279 + ], + "lines": [ + { + "bbox": [ + 314, + 270, + 387, + 279 + ], + "spans": [ + { + "bbox": [ + 314, + 270, + 387, + 279 + ], + "type": "text", + "content": "(a) Original 3D points" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 306, + 281, + 544, + 303 + ], + "lines": [ + { + "bbox": [ + 306, + 281, + 544, + 303 + ], + "spans": [ + { + "bbox": [ + 306, + 281, + 544, + 303 + ], + "type": "text", + "content": "Figure 3. Illustration of the procedure for selecting points of intersection and constructing a 3D ray cloud." + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 390, + 182, + 463, + 268 + ], + "blocks": [ + { + "bbox": [ + 390, + 182, + 463, + 268 + ], + "lines": [ + { + "bbox": [ + 390, + 182, + 463, + 268 + ], + "spans": [ + { + "bbox": [ + 390, + 182, + 463, + 268 + ], + "type": "image", + "image_path": "608b13162b1499f0fc5964576cdfef2d47e5b5a5d0147e91975ce9425b9ecaeb.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 390, + 270, + 462, + 279 + ], + "lines": [ + { + "bbox": [ + 390, + 270, + 462, + 279 + ], + "spans": [ + { + "bbox": [ + 390, + 270, + 462, + 279 + ], + "type": "text", + "content": "(b) Center assignment" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 465, + 182, + 541, + 268 + ], + "blocks": [ + { + "bbox": [ + 465, + 182, + 541, + 268 + ], + "lines": [ + { + "bbox": [ + 465, + 182, + 541, + 268 + ], + "spans": [ + { + "bbox": [ + 465, + 182, + 541, + 268 + ], + "type": "image", + "image_path": "f1e809af13ffa8b556d0b403617e6e378601734c90d64b1747bbc490691ec0b8.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 479, + 270, + 524, + 279 + ], + "lines": [ + { + "bbox": [ + 479, + 270, + 524, + 279 + ], + "spans": [ + { + "bbox": [ + 479, + 270, + 524, + 279 + ], + "type": "text", + "content": "(c) Ray cloud" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 310, + 545, + 382 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 310, + 545, + 382 + ], + "spans": [ + { + "bbox": [ + 304, + 310, + 545, + 382 + ], + "type": "text", + "content": "only be recovered up to scale as camera localization resembles relative pose estimation [35] between the query camera and the intersecting 3D lines. We aim to overcome this issue by utilizing two intersection points. This can be viewed as adding another camera to the scene, forming a calibrated rig of perspective cameras to resolve the scale ambiguity." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 306, + 390, + 440, + 403 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 390, + 440, + 403 + ], + "spans": [ + { + "bbox": [ + 306, + 390, + 440, + 403 + ], + "type": "text", + "content": "3.1. Construction procedure" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 304, + 408, + 545, + 564 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 408, + 545, + 564 + ], + "spans": [ + { + "bbox": [ + 304, + 408, + 545, + 564 + ], + "type": "text", + "content": "Lifting 3D points to a ray cloud comprises three steps, namely setting two ray centers and assigning 3D points to one of two ray centers. We choose to set the ray centers by applying the " + }, + { + "bbox": [ + 304, + 408, + 545, + 564 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 304, + 408, + 545, + 564 + ], + "type": "text", + "content": "-means clustering algorithm on the point cloud with " + }, + { + "bbox": [ + 304, + 408, + 545, + 564 + ], + "type": "inline_equation", + "content": "K = 2" + }, + { + "bbox": [ + 304, + 408, + 545, + 564 + ], + "type": "text", + "content": ", from which the centers of two clusters are set as the ray centers (explained below). Then, we assign half of randomly sampled points from the point cloud to center " + }, + { + "bbox": [ + 304, + 408, + 545, + 564 + ], + "type": "inline_equation", + "content": "1(\\mathbf{c}_1 \\in \\mathbb{R}^3)" + }, + { + "bbox": [ + 304, + 408, + 545, + 564 + ], + "type": "text", + "content": ", and the other half to center " + }, + { + "bbox": [ + 304, + 408, + 545, + 564 + ], + "type": "inline_equation", + "content": "2(\\mathbf{c}_2 \\in \\mathbb{R}^3)" + }, + { + "bbox": [ + 304, + 408, + 545, + 564 + ], + "type": "text", + "content": ". Note this assignment is not related to the " + }, + { + "bbox": [ + 304, + 408, + 545, + 564 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 304, + 408, + 545, + 564 + ], + "type": "text", + "content": "-means clustering. Finally, each point " + }, + { + "bbox": [ + 304, + 408, + 545, + 564 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_i" + }, + { + "bbox": [ + 304, + 408, + 545, + 564 + ], + "type": "text", + "content": " is lifted to a line as " + }, + { + "bbox": [ + 304, + 408, + 545, + 564 + ], + "type": "inline_equation", + "content": "\\mathbf{c}_j + \\lambda (\\mathbf{x}_i - \\mathbf{c}_j)" + }, + { + "bbox": [ + 304, + 408, + 545, + 564 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 304, + 408, + 545, + 564 + ], + "type": "inline_equation", + "content": "\\lambda \\in \\mathbb{R}" + }, + { + "bbox": [ + 304, + 408, + 545, + 564 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 408, + 545, + 564 + ], + "type": "inline_equation", + "content": "j \\in [1,2]" + }, + { + "bbox": [ + 304, + 408, + 545, + 564 + ], + "type": "text", + "content": " is the assigned center index for point " + }, + { + "bbox": [ + 304, + 408, + 545, + 564 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 408, + 545, + 564 + ], + "type": "text", + "content": ". The final output resembles omnidirectional rays from two perspective cameras (see Fig. 3)." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 304, + 567, + 545, + 663 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 567, + 545, + 663 + ], + "spans": [ + { + "bbox": [ + 304, + 567, + 545, + 663 + ], + "type": "text", + "content": "Effect of changing ray center locations If the ray centers are located very close to each other, then they begin to resemble a single ray center, increasing ambiguity in translation scale. On the other hand, if the ray centers are distant from each other and the scene, then the resulting rays are nearly parallel which degrades localization accuracy. We empirically observe that using the centers from " + }, + { + "bbox": [ + 304, + 567, + 545, + 663 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 304, + 567, + 545, + 663 + ], + "type": "text", + "content": "-means clustering usually avoids both extreme cases (see [32])." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 306, + 671, + 440, + 684 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 671, + 440, + 684 + ], + "spans": [ + { + "bbox": [ + 306, + 671, + 440, + 684 + ], + "type": "text", + "content": "3.2. Camera pose estimation" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "content": "We now illustrate an efficient localization approach for ray clouds that can benefit from faster inference speed com" + } + ] + } + ], + "index": 26 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "9775" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 67, + 72, + 157, + 131 + ], + "blocks": [ + { + "bbox": [ + 67, + 72, + 157, + 131 + ], + "lines": [ + { + "bbox": [ + 67, + 72, + 157, + 131 + ], + "spans": [ + { + "bbox": [ + 67, + 72, + 157, + 131 + ], + "type": "image", + "image_path": "466228ecdfb17b4e4e8eb72ed74c0210f46a0a4d2c239bf40b2999d82ecc7afc.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 65, + 133, + 160, + 143 + ], + "lines": [ + { + "bbox": [ + 65, + 133, + 160, + 143 + ], + "spans": [ + { + "bbox": [ + 65, + 133, + 160, + 143 + ], + "type": "text", + "content": "(a) Standard line cloud " + }, + { + "bbox": [ + 65, + 133, + 160, + 143 + ], + "type": "inline_equation", + "content": "(p6\\mathrm{L})" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 181, + 72, + 276, + 129 + ], + "blocks": [ + { + "bbox": [ + 181, + 72, + 276, + 129 + ], + "lines": [ + { + "bbox": [ + 181, + 72, + 276, + 129 + ], + "spans": [ + { + "bbox": [ + 181, + 72, + 276, + 129 + ], + "type": "image", + "image_path": "b7c13d92d34eeafcf405d7564e5499eeea8139b76eb8e20daf3139d47ee3d644.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 186, + 133, + 260, + 143 + ], + "lines": [ + { + "bbox": [ + 186, + 133, + 260, + 143 + ], + "spans": [ + { + "bbox": [ + 186, + 133, + 260, + 143 + ], + "type": "text", + "content": "(b) Ray cloud " + }, + { + "bbox": [ + 186, + 133, + 260, + 143 + ], + "type": "inline_equation", + "content": "(p5 + 1\\mathrm{R})" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 147, + 287, + 257 + ], + "lines": [ + { + "bbox": [ + 46, + 147, + 287, + 257 + ], + "spans": [ + { + "bbox": [ + 46, + 147, + 287, + 257 + ], + "type": "text", + "content": "Figure 4. Illustration of minimal pose estimation problems solved by different line representations. Camera localization using a uniform line cloud can be seen as the special case of generalized relative pose estimation problem [52] involving one perspective camera (query) and one generalized camera (line cloud). On the other hand, camera localization using a ray cloud can be regarded as the case of perspective relative pose estimation problem between the query image and a rig of calibrated cameras (ray cloud). The latter enjoys shorter runtime as shown in Sec. 6 partly due to stronger ray constraints reducing the number of algebraic solutions." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 265, + 286, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 265, + 286, + 361 + ], + "spans": [ + { + "bbox": [ + 46, + 265, + 286, + 361 + ], + "type": "text", + "content": "pared to other line-based scene representations [26, 36, 50]. This is inspired by the illustration in Fig. 4b that the ray cloud can be viewed as two virtual perspective cameras located at the ray centers " + }, + { + "bbox": [ + 46, + 265, + 286, + 361 + ], + "type": "inline_equation", + "content": "\\mathbf{c}_1 \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 46, + 265, + 286, + 361 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 265, + 286, + 361 + ], + "type": "inline_equation", + "content": "\\mathbf{c}_2 \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 46, + 265, + 286, + 361 + ], + "type": "text", + "content": ". Since the rays from these cameras are omnidirectional, we can set their rotation matrices " + }, + { + "bbox": [ + 46, + 265, + 286, + 361 + ], + "type": "inline_equation", + "content": "\\mathsf{R}_1, \\mathsf{R}_2 \\in SO(3)" + }, + { + "bbox": [ + 46, + 265, + 286, + 361 + ], + "type": "text", + "content": " to be aligned with world coordinates, i.e. " + }, + { + "bbox": [ + 46, + 265, + 286, + 361 + ], + "type": "inline_equation", + "content": "[\\mathsf{R}_1|\\mathbf{t}_1] = [\\mathsf{I}| - \\mathbf{c}_1]" + }, + { + "bbox": [ + 46, + 265, + 286, + 361 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 265, + 286, + 361 + ], + "type": "inline_equation", + "content": "[\\mathsf{R}_2|\\mathbf{t}_2] = [\\mathsf{I}| - \\mathbf{c}_2]" + }, + { + "bbox": [ + 46, + 265, + 286, + 361 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 265, + 286, + 361 + ], + "type": "inline_equation", + "content": "\\mathbf{t}_1, \\mathbf{t}_2 \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 46, + 265, + 286, + 361 + ], + "type": "text", + "content": " are the camera translations." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 361, + 286, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 361, + 286, + 468 + ], + "spans": [ + { + "bbox": [ + 46, + 361, + 286, + 468 + ], + "type": "text", + "content": "Below, we show that having 5 correspondences between 2D keypoints from the query image and 3D rays from a randomly selected virtual camera and one correspondence between a 2D query keypoint and a 3D ray from a remaining virtual camera allows us to estimate the absolute camera pose using a ray cloud. We denote this adaptation of minimal " + }, + { + "bbox": [ + 46, + 361, + 286, + 468 + ], + "type": "inline_equation", + "content": "5 + 1" + }, + { + "bbox": [ + 46, + 361, + 286, + 468 + ], + "type": "text", + "content": " solver [59] as the perspective-" + }, + { + "bbox": [ + 46, + 361, + 286, + 468 + ], + "type": "inline_equation", + "content": "(5 + 1)" + }, + { + "bbox": [ + 46, + 361, + 286, + 468 + ], + "type": "text", + "content": "-rays " + }, + { + "bbox": [ + 46, + 361, + 286, + 468 + ], + "type": "inline_equation", + "content": "(p5 + 1\\mathrm{R})" + }, + { + "bbox": [ + 46, + 361, + 286, + 468 + ], + "type": "text", + "content": " solver. Similar to previous studies [26, 50], we assume the camera intrinsics " + }, + { + "bbox": [ + 46, + 361, + 286, + 468 + ], + "type": "inline_equation", + "content": "\\mathbf{K} \\in \\mathbb{R}^{3 \\times 3}" + }, + { + "bbox": [ + 46, + 361, + 286, + 468 + ], + "type": "text", + "content": " are known." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 475, + 286, + 571 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 475, + 286, + 571 + ], + "spans": [ + { + "bbox": [ + 46, + 475, + 286, + 571 + ], + "type": "inline_equation", + "content": "p5 + 1\\mathbf{R}" + }, + { + "bbox": [ + 46, + 475, + 286, + 571 + ], + "type": "text", + "content": " solver We start by estimating the relative pose between the query image and the first virtual camera using the 5-point algorithm [35]. It yields multiple candidates for the relative pose of the query image w.r.t. the first virtual camera defined as " + }, + { + "bbox": [ + 46, + 475, + 286, + 571 + ], + "type": "inline_equation", + "content": "[\\mathtt{R}_{q1}|\\alpha \\hat{\\mathbf{t}}_{q1}]" + }, + { + "bbox": [ + 46, + 475, + 286, + 571 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 475, + 286, + 571 + ], + "type": "inline_equation", + "content": "\\mathtt{R}_{q1}\\in SO(3)" + }, + { + "bbox": [ + 46, + 475, + 286, + 571 + ], + "type": "text", + "content": " is the relative rotation, " + }, + { + "bbox": [ + 46, + 475, + 286, + 571 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{t}}_{q1}\\in S^2" + }, + { + "bbox": [ + 46, + 475, + 286, + 571 + ], + "type": "text", + "content": " is the normalized relative translation and " + }, + { + "bbox": [ + 46, + 475, + 286, + 571 + ], + "type": "inline_equation", + "content": "\\alpha \\in \\mathbb{R}" + }, + { + "bbox": [ + 46, + 475, + 286, + 571 + ], + "type": "text", + "content": " is an unknown scale factor. Subsequently, the absolute pose of query camera is " + }, + { + "bbox": [ + 46, + 475, + 286, + 571 + ], + "type": "inline_equation", + "content": "[\\mathtt{R}_{q1}|\\alpha \\hat{\\mathbf{t}}_{q1} - \\mathtt{R}_{q1}\\mathbf{c}_1]" + }, + { + "bbox": [ + 46, + 475, + 286, + 571 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 571, + 286, + 643 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 571, + 286, + 643 + ], + "spans": [ + { + "bbox": [ + 47, + 571, + 286, + 643 + ], + "type": "text", + "content": "In order to disambiguate " + }, + { + "bbox": [ + 47, + 571, + 286, + 643 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 47, + 571, + 286, + 643 + ], + "type": "text", + "content": ", we utilize the additional correspondence between a normalized query keypoint " + }, + { + "bbox": [ + 47, + 571, + 286, + 643 + ], + "type": "inline_equation", + "content": "\\mathbf{p} := (p_1, p_2)^\\top \\in \\mathbb{R}^2" + }, + { + "bbox": [ + 47, + 571, + 286, + 643 + ], + "type": "text", + "content": " and a 3D ray " + }, + { + "bbox": [ + 47, + 571, + 286, + 643 + ], + "type": "inline_equation", + "content": "\\mathbf{l} \\in Gr(1, 3)" + }, + { + "bbox": [ + 47, + 571, + 286, + 643 + ], + "type": "text", + "content": " from the second virtual camera. We note the ray from the query camera center passing through " + }, + { + "bbox": [ + 47, + 571, + 286, + 643 + ], + "type": "inline_equation", + "content": "\\mathbf{p}" + }, + { + "bbox": [ + 47, + 571, + 286, + 643 + ], + "type": "text", + "content": " would ideally intersect " + }, + { + "bbox": [ + 47, + 571, + 286, + 643 + ], + "type": "inline_equation", + "content": "\\mathbf{l}" + }, + { + "bbox": [ + 47, + 571, + 286, + 643 + ], + "type": "text", + "content": ". Expressing both rays in the query-centered coordinates yields" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 82, + 652, + 286, + 668 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 82, + 652, + 286, + 668 + ], + "spans": [ + { + "bbox": [ + 82, + 652, + 286, + 668 + ], + "type": "interline_equation", + "content": "\\lambda_ {1} \\left[ p _ {1}, p _ {2}, 1 \\right] ^ {\\top} = \\lambda_ {2} \\mathrm {R} _ {q 1} \\hat {\\mathbf {n}} + \\alpha \\hat {\\mathbf {t}} _ {q 1} - \\mathrm {R} _ {q 1} \\mathbf {c} _ {1} \\tag {1}", + "image_path": "471fb778aab568bab221181b540210c8cc1ca2f9a4b0f37e23548bfa38975c41.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 677, + 286, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 286, + 712 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 286, + 712 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 677, + 286, + 712 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{n}}\\in S^2" + }, + { + "bbox": [ + 46, + 677, + 286, + 712 + ], + "type": "text", + "content": " is the line direction of 1 in world coordinates and " + }, + { + "bbox": [ + 46, + 677, + 286, + 712 + ], + "type": "inline_equation", + "content": "\\lambda_1,\\lambda_2\\in \\mathbb{R}" + }, + { + "bbox": [ + 46, + 677, + 286, + 712 + ], + "type": "text", + "content": " are scale parameters. Since we have 3 equations with 3 unknowns, " + }, + { + "bbox": [ + 46, + 677, + 286, + 712 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 46, + 677, + 286, + 712 + ], + "type": "text", + "content": " can be determined." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "text", + "content": "While the " + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "inline_equation", + "content": "p5 + 1\\mathrm{R}" + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "text", + "content": " solver requires 6 correspondences like " + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "inline_equation", + "content": "p6\\mathrm{L}" + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "text", + "content": " [26, 50], the utilized 5-point algorithm only yields up to 10 solutions [35] that is smaller than 64 solutions from the generalized solver [52] in " + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "inline_equation", + "content": "p6\\mathrm{L}" + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "text", + "content": ". This significantly reduces the computational complexity for determining the correct configuration. We also enforce the chirality constraint [17] on the query image to discard infeasible solutions." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 159, + 545, + 219 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 159, + 545, + 219 + ], + "spans": [ + { + "bbox": [ + 304, + 159, + 545, + 219 + ], + "type": "text", + "content": "Robust estimation We equip above " + }, + { + "bbox": [ + 304, + 159, + 545, + 219 + ], + "type": "inline_equation", + "content": "p5 + 1R" + }, + { + "bbox": [ + 304, + 159, + 545, + 219 + ], + "type": "text", + "content": " solver with the LO-RANSAC [31] algorithm for robust camera localization. For selecting the best model candidate, we adopt an MLESAC [54]-type cost function based on the squared sum of epipolar distances in the query image defined as" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 353, + 227, + 545, + 262 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 353, + 227, + 545, + 262 + ], + "spans": [ + { + "bbox": [ + 353, + 227, + 545, + 262 + ], + "type": "interline_equation", + "content": "\\sum_ {(i, j) \\in \\Omega} \\rho \\left(\\frac {\\left([ \\mathbf {p} _ {i} ^ {\\top} , 1 ] \\mathrm {E} _ {j} \\hat {\\mathbf {n}} _ {i}\\right) ^ {2}}{\\left(\\mathbf {e} _ {j 1} ^ {\\top} \\hat {\\mathbf {n}} _ {i}\\right) ^ {2} + \\left(\\mathbf {e} _ {j 2} ^ {\\top} \\hat {\\mathbf {n}} _ {i}\\right) ^ {2}}\\right), \\tag {2}", + "image_path": "6f56202341d3974eab62a4cb9d318b803a6b03841b57adbb9a19433a72cda1dd.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 272, + 545, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 272, + 545, + 392 + ], + "spans": [ + { + "bbox": [ + 304, + 272, + 545, + 392 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 272, + 545, + 392 + ], + "type": "inline_equation", + "content": "\\rho(s) := \\min(s, \\tau^2)" + }, + { + "bbox": [ + 304, + 272, + 545, + 392 + ], + "type": "text", + "content": " is a robust kernel with upper bound of threshold " + }, + { + "bbox": [ + 304, + 272, + 545, + 392 + ], + "type": "inline_equation", + "content": "\\tau^2" + }, + { + "bbox": [ + 304, + 272, + 545, + 392 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 272, + 545, + 392 + ], + "type": "inline_equation", + "content": "\\Omega" + }, + { + "bbox": [ + 304, + 272, + 545, + 392 + ], + "type": "text", + "content": " denotes the pairing arrangement of which keypoint " + }, + { + "bbox": [ + 304, + 272, + 545, + 392 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 272, + 545, + 392 + ], + "type": "text", + "content": " matches with ray " + }, + { + "bbox": [ + 304, + 272, + 545, + 392 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 272, + 545, + 392 + ], + "type": "text", + "content": " from center " + }, + { + "bbox": [ + 304, + 272, + 545, + 392 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 304, + 272, + 545, + 392 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 272, + 545, + 392 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_i \\in \\mathbb{R}^2" + }, + { + "bbox": [ + 304, + 272, + 545, + 392 + ], + "type": "text", + "content": " is the " + }, + { + "bbox": [ + 304, + 272, + 545, + 392 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 272, + 545, + 392 + ], + "type": "text", + "content": "-th normalized keypoint in the query image, " + }, + { + "bbox": [ + 304, + 272, + 545, + 392 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{n}}_i \\in S^2" + }, + { + "bbox": [ + 304, + 272, + 545, + 392 + ], + "type": "text", + "content": " is ray " + }, + { + "bbox": [ + 304, + 272, + 545, + 392 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 272, + 545, + 392 + ], + "type": "text", + "content": " passing through the ray center " + }, + { + "bbox": [ + 304, + 272, + 545, + 392 + ], + "type": "inline_equation", + "content": "j \\in [1, 2]" + }, + { + "bbox": [ + 304, + 272, + 545, + 392 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 272, + 545, + 392 + ], + "type": "inline_equation", + "content": "\\mathbf{E}_j := [\\mathbf{e}_{j1}, \\mathbf{e}_{j2}, \\mathbf{e}_{j3}]^\\top" + }, + { + "bbox": [ + 304, + 272, + 545, + 392 + ], + "type": "text", + "content": " is the essential matrix between the virtual camera " + }, + { + "bbox": [ + 304, + 272, + 545, + 392 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 304, + 272, + 545, + 392 + ], + "type": "text", + "content": " and the query camera. Since the two virtual cameras share the same rotation and only differ by translation, " + }, + { + "bbox": [ + 304, + 272, + 545, + 392 + ], + "type": "inline_equation", + "content": "\\mathbf{E}_2" + }, + { + "bbox": [ + 304, + 272, + 545, + 392 + ], + "type": "text", + "content": " can be easily computed from " + }, + { + "bbox": [ + 304, + 272, + 545, + 392 + ], + "type": "inline_equation", + "content": "\\mathbf{E}_1" + }, + { + "bbox": [ + 304, + 272, + 545, + 392 + ], + "type": "text", + "content": ". The best solution from RANSAC is refined by iteratively minimizing (2)." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 399, + 518, + 411 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 399, + 518, + 411 + ], + "spans": [ + { + "bbox": [ + 306, + 399, + 518, + 411 + ], + "type": "text", + "content": "3.3. Geometry inversion attack on ray clouds" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 418, + 545, + 477 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 418, + 545, + 477 + ], + "spans": [ + { + "bbox": [ + 304, + 418, + 545, + 477 + ], + "type": "text", + "content": "We show that applying the geometry inversion attack [7] naively to a ray cloud yields a degenerate solution. For this purpose, we make use of some prior knowledge regarding the probabilistic viewpoint and algorithmic details of the geometry inversion attack [7] reviewed in Sec. 2." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 477, + 545, + 537 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 477, + 545, + 537 + ], + "spans": [ + { + "bbox": [ + 304, + 477, + 545, + 537 + ], + "type": "text", + "content": "Let " + }, + { + "bbox": [ + 304, + 477, + 545, + 537 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_i\\in \\mathbb{R}^3" + }, + { + "bbox": [ + 304, + 477, + 545, + 537 + ], + "type": "text", + "content": " be the " + }, + { + "bbox": [ + 304, + 477, + 545, + 537 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 477, + 545, + 537 + ], + "type": "text", + "content": "-th point and " + }, + { + "bbox": [ + 304, + 477, + 545, + 537 + ], + "type": "inline_equation", + "content": "\\mathbf{l}_i\\in Gr(1,3)" + }, + { + "bbox": [ + 304, + 477, + 545, + 537 + ], + "type": "text", + "content": " be its lifted ray. The point cloud is defined as " + }, + { + "bbox": [ + 304, + 477, + 545, + 537 + ], + "type": "inline_equation", + "content": "\\mathcal{P}\\coloneqq \\{\\mathbf{x}_i\\}" + }, + { + "bbox": [ + 304, + 477, + 545, + 537 + ], + "type": "text", + "content": " and the ray cloud is defined as " + }, + { + "bbox": [ + 304, + 477, + 545, + 537 + ], + "type": "inline_equation", + "content": "\\mathcal{L}\\coloneqq \\{\\mathbf{l}_i\\}" + }, + { + "bbox": [ + 304, + 477, + 545, + 537 + ], + "type": "text", + "content": ". Since ray centers " + }, + { + "bbox": [ + 304, + 477, + 545, + 537 + ], + "type": "inline_equation", + "content": "\\mathcal{C}\\coloneqq \\{\\mathbf{c}_1,\\mathbf{c}_2\\}" + }, + { + "bbox": [ + 304, + 477, + 545, + 537 + ], + "type": "text", + "content": " are easily found by inspecting the intersection points, we aim to maximize the conditional posterior of " + }, + { + "bbox": [ + 304, + 477, + 545, + 537 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 304, + 477, + 545, + 537 + ], + "type": "text", + "content": "," + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 357, + 548, + 545, + 561 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 357, + 548, + 545, + 561 + ], + "spans": [ + { + "bbox": [ + 357, + 548, + 545, + 561 + ], + "type": "interline_equation", + "content": "P (\\mathcal {P} | \\mathcal {L}, \\mathcal {C}) \\propto P (\\mathcal {L} | \\mathcal {P}, \\mathcal {C}) P (\\mathcal {P} | \\mathcal {C}). \\tag {3}", + "image_path": "e1df63e3a6453ee052e3cddd833701550ddb047523b8bdbbab1dc9cb83d27bca.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 570, + 545, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 545, + 641 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 545, + 641 + ], + "type": "text", + "content": "Note the conditional likelihood " + }, + { + "bbox": [ + 304, + 570, + 545, + 641 + ], + "type": "inline_equation", + "content": "P(\\mathcal{L}|\\mathcal{P},\\mathcal{C})" + }, + { + "bbox": [ + 304, + 570, + 545, + 641 + ], + "type": "text", + "content": " is a non-zero constant as long as each ray intersects the corresponding pair of ray center and 3D point and the probability of each ray being assigned to one of the ray centers is equally likely. As long as each point " + }, + { + "bbox": [ + 304, + 570, + 545, + 641 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_i" + }, + { + "bbox": [ + 304, + 570, + 545, + 641 + ], + "type": "text", + "content": " is estimated to be on its lifted ray " + }, + { + "bbox": [ + 304, + 570, + 545, + 641 + ], + "type": "inline_equation", + "content": "\\mathbf{l}_i" + }, + { + "bbox": [ + 304, + 570, + 545, + 641 + ], + "type": "text", + "content": ", then maximizing (3) amounts to maximizing " + }, + { + "bbox": [ + 304, + 570, + 545, + 641 + ], + "type": "inline_equation", + "content": "P(\\mathcal{P}|\\mathcal{C})" + }, + { + "bbox": [ + 304, + 570, + 545, + 641 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 642, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 642, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 642, + 545, + 712 + ], + "type": "text", + "content": "However, since " + }, + { + "bbox": [ + 304, + 642, + 545, + 712 + ], + "type": "inline_equation", + "content": "P(\\mathcal{P}|\\mathcal{C})" + }, + { + "bbox": [ + 304, + 642, + 545, + 712 + ], + "type": "text", + "content": " can depend significantly on the ray centers " + }, + { + "bbox": [ + 304, + 642, + 545, + 712 + ], + "type": "inline_equation", + "content": "\\mathcal{C}" + }, + { + "bbox": [ + 304, + 642, + 545, + 712 + ], + "type": "text", + "content": ", it seems difficult to yield a point statistic similar to " + }, + { + "bbox": [ + 304, + 642, + 545, + 712 + ], + "type": "inline_equation", + "content": "P(\\mathcal{P})" + }, + { + "bbox": [ + 304, + 642, + 545, + 712 + ], + "type": "text", + "content": " in [7]. We implicitly figure out the difference between " + }, + { + "bbox": [ + 304, + 642, + 545, + 712 + ], + "type": "inline_equation", + "content": "P(\\mathcal{P}|\\mathcal{C})" + }, + { + "bbox": [ + 304, + 642, + 545, + 712 + ], + "type": "text", + "content": " and the original prior " + }, + { + "bbox": [ + 304, + 642, + 545, + 712 + ], + "type": "inline_equation", + "content": "P(\\mathcal{P})" + }, + { + "bbox": [ + 304, + 642, + 545, + 712 + ], + "type": "text", + "content": " by applying the geometry inversion attack [7], which maximizes " + }, + { + "bbox": [ + 304, + 642, + 545, + 712 + ], + "type": "inline_equation", + "content": "P(\\mathcal{P})" + }, + { + "bbox": [ + 304, + 642, + 545, + 712 + ], + "type": "text", + "content": ", as-is to the ray cloud. As visualized in Fig. 5a, the estimated" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "9776" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 53, + 68, + 166, + 135 + ], + "blocks": [ + { + "bbox": [ + 53, + 68, + 166, + 135 + ], + "lines": [ + { + "bbox": [ + 53, + 68, + 166, + 135 + ], + "spans": [ + { + "bbox": [ + 53, + 68, + 166, + 135 + ], + "type": "image", + "image_path": "820beb607c673b3e9c04979fc5db6a32a23db599b489cdbdd4e287dd7aa956fb.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 136, + 162, + 146 + ], + "lines": [ + { + "bbox": [ + 54, + 136, + 162, + 146 + ], + "spans": [ + { + "bbox": [ + 54, + 136, + 162, + 146 + ], + "type": "text", + "content": "(a) Naive neighbor estimation [7]" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 170, + 68, + 284, + 135 + ], + "blocks": [ + { + "bbox": [ + 170, + 68, + 284, + 135 + ], + "lines": [ + { + "bbox": [ + 170, + 68, + 284, + 135 + ], + "spans": [ + { + "bbox": [ + 170, + 68, + 284, + 135 + ], + "type": "image", + "image_path": "352a6a3531ff6fd64b52f8d31f5a431e469af2ef26bb992055f026e6e5022559.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 171, + 136, + 279, + 146 + ], + "lines": [ + { + "bbox": [ + 171, + 136, + 279, + 146 + ], + "spans": [ + { + "bbox": [ + 171, + 136, + 279, + 146 + ], + "type": "text", + "content": "(b) Modified neighbor estimation" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 149, + 287, + 217 + ], + "lines": [ + { + "bbox": [ + 46, + 149, + 287, + 217 + ], + "spans": [ + { + "bbox": [ + 46, + 149, + 287, + 217 + ], + "type": "text", + "content": "Figure 5. Illustration of peak-finding [7] results from (a) naive neighbor-lines estimation [7] and (b) modified neighbor-lines estimation. (a) indicates that the estimated peak will always be located at the center of the current ray. In contrast, our modification only samples the neighborhood of lines using the bundle of rays originating from the opposite center, eliminating the trivial solution." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 224, + 287, + 297 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 224, + 287, + 297 + ], + "spans": [ + { + "bbox": [ + 46, + 224, + 287, + 297 + ], + "type": "text", + "content": "point for each ray becomes the center of the constituent ray, yielding a trivial recovery. This is because, in the process of the geometric inversion, the " + }, + { + "bbox": [ + 46, + 224, + 287, + 297 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 46, + 224, + 287, + 297 + ], + "type": "text", + "content": "-nearest neighboring lines include a bundle of rays originating from the same center, which yields an erroneous peak at the ray center on the histogram of point candidates, hindering general recovery." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 307, + 264, + 320 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 307, + 264, + 320 + ], + "spans": [ + { + "bbox": [ + 47, + 307, + 264, + 320 + ], + "type": "text", + "content": "4. Adapted inversion attack for ray clouds" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 327, + 287, + 375 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 327, + 287, + 375 + ], + "spans": [ + { + "bbox": [ + 46, + 327, + 287, + 375 + ], + "type": "text", + "content": "In this section, we illustrate two main reasons for trivial recovery of 3D points in Sec. 3.3, namely i) lower significance of nearest points for estimating 3D points, and ii) bad estimation of true nearest rays lifted from nearest points." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 375, + 287, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 375, + 287, + 495 + ], + "spans": [ + { + "bbox": [ + 46, + 375, + 287, + 495 + ], + "type": "text", + "content": "First, the assumption in [7] that the near-80% of candidates which are closest to lines lifted from " + }, + { + "bbox": [ + 46, + 375, + 287, + 495 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 375, + 287, + 495 + ], + "type": "text", + "content": "-nearest neighboring points yields accurate points does not hold for ray clouds as some of the rays from these " + }, + { + "bbox": [ + 46, + 375, + 287, + 495 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 375, + 287, + 495 + ], + "type": "text", + "content": "-nearest points will yield estimates at the current ray center. Second, the assumption in [7] that " + }, + { + "bbox": [ + 46, + 375, + 287, + 495 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 46, + 375, + 287, + 495 + ], + "type": "text", + "content": "-nearest lines will overlap enough with the set of rays lifted from the " + }, + { + "bbox": [ + 46, + 375, + 287, + 495 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 375, + 287, + 495 + ], + "type": "text", + "content": "-nearest points (" + }, + { + "bbox": [ + 46, + 375, + 287, + 495 + ], + "type": "inline_equation", + "content": "K > k" + }, + { + "bbox": [ + 46, + 375, + 287, + 495 + ], + "type": "text", + "content": ") is weakened as the " + }, + { + "bbox": [ + 46, + 375, + 287, + 495 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 46, + 375, + 287, + 495 + ], + "type": "text", + "content": "-nearest lines will be rays sharing the same center whose original points may be far away from the original point of the current lifted ray." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 495, + 287, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 495, + 287, + 555 + ], + "spans": [ + { + "bbox": [ + 47, + 495, + 287, + 555 + ], + "type": "text", + "content": "In Sec. 4, we attempt to address these issues by considering known ray centers in order to approximately maximize " + }, + { + "bbox": [ + 47, + 495, + 287, + 555 + ], + "type": "inline_equation", + "content": "P(\\mathcal{P}|\\mathcal{C})" + }, + { + "bbox": [ + 47, + 495, + 287, + 555 + ], + "type": "text", + "content": " instead of " + }, + { + "bbox": [ + 47, + 495, + 287, + 555 + ], + "type": "inline_equation", + "content": "P(\\mathcal{P})" + }, + { + "bbox": [ + 47, + 495, + 287, + 555 + ], + "type": "text", + "content": ". We also aim to address trivial solutions by incorporating the known positions of the ray centers " + }, + { + "bbox": [ + 47, + 495, + 287, + 555 + ], + "type": "inline_equation", + "content": "\\mathcal{C}" + }, + { + "bbox": [ + 47, + 495, + 287, + 555 + ], + "type": "text", + "content": " to approximate maximizing over " + }, + { + "bbox": [ + 47, + 495, + 287, + 555 + ], + "type": "inline_equation", + "content": "P(\\mathcal{P}|\\mathcal{C})" + }, + { + "bbox": [ + 47, + 495, + 287, + 555 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 563, + 236, + 575 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 563, + 236, + 575 + ], + "spans": [ + { + "bbox": [ + 47, + 563, + 236, + 575 + ], + "type": "text", + "content": "4.1. Eliminating the degenerate solution" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 582, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 582, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 582, + 287, + 714 + ], + "type": "text", + "content": "The trivial recovery of 3D points (estimated 3D points " + }, + { + "bbox": [ + 46, + 582, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 46, + 582, + 287, + 714 + ], + "type": "text", + "content": " always located at the two ray centers " + }, + { + "bbox": [ + 46, + 582, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{C}" + }, + { + "bbox": [ + 46, + 582, + 287, + 714 + ], + "type": "text", + "content": ") is due to rays intersecting the same ray center always ending up as the nearest neighboring (NN) lines, producing the peak density of closest points at the ray center (see Fig. 5a). To bypass this issue, we exclude the rays intersecting through the current ray when constructing the set of nearest neighboring lines for estimating the peak density, and only select the NN-lines from the other ray center. This can be viewed as incorporating the known locations of the ray centers for recovering the points (see Fig. 5b) to maximize " + }, + { + "bbox": [ + 46, + 582, + 287, + 714 + ], + "type": "inline_equation", + "content": "P(\\mathcal{P}|\\mathcal{C})" + }, + { + "bbox": [ + 46, + 582, + 287, + 714 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 312, + 69, + 387, + 118 + ], + "blocks": [ + { + "bbox": [ + 312, + 69, + 387, + 118 + ], + "lines": [ + { + "bbox": [ + 312, + 69, + 387, + 118 + ], + "spans": [ + { + "bbox": [ + 312, + 69, + 387, + 118 + ], + "type": "image", + "image_path": "31dee82cba4c836a92748c426b53ff2f2e76aa06523a1bc5e177aa8f28b34bb9.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 119, + 384, + 129 + ], + "lines": [ + { + "bbox": [ + 313, + 119, + 384, + 129 + ], + "spans": [ + { + "bbox": [ + 313, + 119, + 384, + 129 + ], + "type": "text", + "content": "(a) ULC [50] (oracle)" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 304, + 133, + 545, + 254 + ], + "lines": [ + { + "bbox": [ + 304, + 133, + 545, + 254 + ], + "spans": [ + { + "bbox": [ + 304, + 133, + 545, + 254 + ], + "type": "text", + "content": "Figure 6. 3D point cloud recovered on different settings of nearest neighbors for the Apt1 living from [55] (black-original points, red-recovered points). In (b) and (c), we have only utilized neighboring rays connected to the opposite ray center to avoid degenerate solutions (in Sec. 4.1). For the oracle cases of (a) and (b), rays from the true nearest points of the original points are used to estimate the points, showing feasible recovery. In contrast, when the nearest neighboring points are approximated by the nearest neighboring rays, it leads to worse recovery as shown in (c). This implies false positive rays, which are close but whose original points are distant, significantly degrade the recovery of points." + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 391, + 70, + 467, + 117 + ], + "blocks": [ + { + "bbox": [ + 391, + 70, + 467, + 117 + ], + "lines": [ + { + "bbox": [ + 391, + 70, + 467, + 117 + ], + "spans": [ + { + "bbox": [ + 391, + 70, + 467, + 117 + ], + "type": "image", + "image_path": "5729c9489b4c27e7d7f4600c3bf685873b1f69e86d8c7fe9f6ebd38c4c4d959f.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 392, + 118, + 463, + 129 + ], + "lines": [ + { + "bbox": [ + 392, + 118, + 463, + 129 + ], + "spans": [ + { + "bbox": [ + 392, + 118, + 463, + 129 + ], + "type": "text", + "content": "(b) Ray cloud (oracle)" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 469, + 71, + 543, + 116 + ], + "blocks": [ + { + "bbox": [ + 469, + 71, + 543, + 116 + ], + "lines": [ + { + "bbox": [ + 469, + 71, + 543, + 116 + ], + "spans": [ + { + "bbox": [ + 469, + 71, + 543, + 116 + ], + "type": "image", + "image_path": "a1d4538230a73ddb956fa84df194d8e09ffe729faffaf755b600da0a25970c3f.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 482, + 118, + 527, + 129 + ], + "lines": [ + { + "bbox": [ + 482, + 118, + 527, + 129 + ], + "spans": [ + { + "bbox": [ + 482, + 118, + 527, + 129 + ], + "type": "text", + "content": "(c) Ray cloud" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 262, + 545, + 370 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 262, + 545, + 370 + ], + "spans": [ + { + "bbox": [ + 304, + 262, + 545, + 370 + ], + "type": "text", + "content": "Fig. 6b shows that, in the oracle case when the ground truth nearest points of the current ray's original 3D point are known, discarding NN-lines intersecting the current ray can avoid degenerate solution and yield more feasible point cloud. At the same time, Fig. 6c shows that the actual point recovery from a ray cloud is worse. We conjecture this is due to the misalignment between the set of nearest neighboring rays and the set of nearest neighboring points for a given point and its lifted ray which is addressed in Sec. 4.2." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 306, + 380, + 536, + 393 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 380, + 536, + 393 + ], + "spans": [ + { + "bbox": [ + 306, + 380, + 536, + 393 + ], + "type": "text", + "content": "4.2. Rejecting rays closely aligned to the baseline" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 400, + 545, + 484 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 400, + 545, + 484 + ], + "spans": [ + { + "bbox": [ + 304, + 400, + 545, + 484 + ], + "type": "text", + "content": "We first analyze the cause of the bad recovery in Fig. 6c even after applying the trick in Sec. 4.1. We define the rays close to the current ray but whose original points are far away from the original point of the current ray as the false positive rays. By comparing the oracle case (b) and the normal case (c) in Fig. 6, we anticipate that the false positive rays are the culprits for degraded estimation of the points." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 304, + 485, + 545, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 485, + 545, + 652 + ], + "spans": [ + { + "bbox": [ + 304, + 485, + 545, + 652 + ], + "type": "text", + "content": "We conjecture two cases where the false positive rays can lead to misleading estimations. First, when there are many rays from the opposite center that are in the proximity of the current ray center, the peak of the histogram may be located near the current ray center. Vice versa, when the current ray passes through the denser area around the opposite center, the peak will be located near the opposite center. These two cases partially cause the 3D points to shrink (see Fig. 6c). Second, when there are many diverse rays in the proximity of the current ray, multiple peaks can be formed from the point candidates (closest points on the ray to neighboring rays) along the current ray. This can arise when the ray centers are embedded in a room-like environment to form congested rays (see [32] for more discussions)." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "content": "Out of above cases, we attempt to reduce the shrinking effect of false positive rays by incorporating known " + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{C}" + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "content": " and not considering any rays from the opposite center which are within the proximity of the current ray center. This can be effectively achieved by excluding a fraction " + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "inline_equation", + "content": "(\\beta)" + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "content": " of rays most" + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "9777" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 48, + 72, + 282, + 155 + ], + "blocks": [ + { + "bbox": [ + 48, + 72, + 282, + 155 + ], + "lines": [ + { + "bbox": [ + 48, + 72, + 282, + 155 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 282, + 155 + ], + "type": "image", + "image_path": "f57a7670a21f370da4d522aac9506a6a8398311e743ae7d71fcc43be0fcc437c.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 64, + 158, + 270, + 168 + ], + "lines": [ + { + "bbox": [ + 64, + 158, + 270, + 168 + ], + "spans": [ + { + "bbox": [ + 64, + 158, + 270, + 168 + ], + "type": "text", + "content": "(a) Visualization of rejection-sampling of rays near the baseline" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 58, + 170, + 164, + 223 + ], + "blocks": [ + { + "bbox": [ + 58, + 170, + 164, + 223 + ], + "lines": [ + { + "bbox": [ + 58, + 170, + 164, + 223 + ], + "spans": [ + { + "bbox": [ + 58, + 170, + 164, + 223 + ], + "type": "image", + "image_path": "521189a74e1679fcf4451972beafe39ac20309e8c7f49929f075ba1c8524b9bf.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 56, + 227, + 162, + 237 + ], + "lines": [ + { + "bbox": [ + 56, + 227, + 162, + 237 + ], + "spans": [ + { + "bbox": [ + 56, + 227, + 162, + 237 + ], + "type": "text", + "content": "(b) Peak-fiding [7] w/o rejection" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 238, + 287, + 282 + ], + "lines": [ + { + "bbox": [ + 46, + 238, + 287, + 282 + ], + "spans": [ + { + "bbox": [ + 46, + 238, + 287, + 282 + ], + "type": "text", + "content": "Figure 7. Illustration of the effect of line rejection method. For a synthetic room comprising " + }, + { + "bbox": [ + 46, + 238, + 287, + 282 + ], + "type": "inline_equation", + "content": "20\\mathrm{k}" + }, + { + "bbox": [ + 46, + 238, + 287, + 282 + ], + "type": "text", + "content": " points, in (b), peak estimation fails to estimate the correct point due to false positive rays. In (c), the rejection excludes false candidates to estimate an accurate point." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 174, + 170, + 279, + 223 + ], + "blocks": [ + { + "bbox": [ + 174, + 170, + 279, + 223 + ], + "lines": [ + { + "bbox": [ + 174, + 170, + 279, + 223 + ], + "spans": [ + { + "bbox": [ + 174, + 170, + 279, + 223 + ], + "type": "image", + "image_path": "32ffe51ebfde302520b5f6777252bf730443794d3462c38cf907db5e17a769a8.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 170, + 227, + 281, + 237 + ], + "lines": [ + { + "bbox": [ + 170, + 227, + 281, + 237 + ], + "spans": [ + { + "bbox": [ + 170, + 227, + 281, + 237 + ], + "type": "text", + "content": "(c) Peak-finding [7] with rejection" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 288, + 287, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 288, + 287, + 323 + ], + "spans": [ + { + "bbox": [ + 46, + 288, + 287, + 323 + ], + "type": "text", + "content": "aligned with the baseline when estimating the histogram of point candidates. Fig. 7 illustrates the effect of this ray rejection which can lead to better point estimation." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 324, + 287, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 324, + 287, + 396 + ], + "spans": [ + { + "bbox": [ + 46, + 324, + 287, + 396 + ], + "type": "text", + "content": "We outline some limitations of this approach. First, if the ray lifted from a 3D point is within the region of rejection sampling, then the estimated point can be inaccurate due to the lack of nearby rays to support the estimation process. Second, the method introduces a hyperparameter " + }, + { + "bbox": [ + 46, + 324, + 287, + 396 + ], + "type": "inline_equation", + "content": "(\\beta)" + }, + { + "bbox": [ + 46, + 324, + 287, + 396 + ], + "type": "text", + "content": " that needs to be carefully set (see results in Table 2)." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 406, + 194, + 420 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 406, + 194, + 420 + ], + "spans": [ + { + "bbox": [ + 47, + 406, + 194, + 420 + ], + "type": "text", + "content": "5. Voxel-based ray sampling" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 426, + 287, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 426, + 287, + 449 + ], + "spans": [ + { + "bbox": [ + 46, + 426, + 287, + 449 + ], + "type": "text", + "content": "We now present a new ray sampling technique that can neutralize the modified inversion attack from Sec. 4." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 450, + 287, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 450, + 287, + 521 + ], + "spans": [ + { + "bbox": [ + 46, + 450, + 287, + 521 + ], + "type": "text", + "content": "Our approach is motivated from Fig. 6 that the quality of point recovery depends on finding the true point candidates once the degenerate solution is eliminated. This time, we try to discourage the true nearest neighboring points from being included in the set of " + }, + { + "bbox": [ + 46, + 450, + 287, + 521 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 46, + 450, + 287, + 521 + ], + "type": "text", + "content": "-nearest neighboring lines when lifted to rays with aims to lower the inversion accuracy." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 522, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 522, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 522, + 287, + 713 + ], + "type": "text", + "content": "For the above purpose, we partition the 3D points into voxels and assign one of two ray centers to each voxel such that i) 3D points within the same voxel are lifted to rays intersecting through the assigned ray center and ii) no neighboring voxels share the same ray center (see Fig. 8). As long as the voxels are large enough, many 3D points except those near the voxel boundaries will have most of their respective " + }, + { + "bbox": [ + 46, + 522, + 287, + 713 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 522, + 287, + 713 + ], + "type": "text", + "content": "-nearest points to be inside the same voxel and subsequently share the same ray center when lifted to rays. Since the rays lifted from these true " + }, + { + "bbox": [ + 46, + 522, + 287, + 713 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 522, + 287, + 713 + ], + "type": "text", + "content": "-nearest neighboring points inside the same voxel will yield a trivial solution as illustrated in Sec. 4.1, these will be disallowed from being used for peak estimation. Hence, we essentially minimize the overlap between the set of " + }, + { + "bbox": [ + 46, + 522, + 287, + 713 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 46, + 522, + 287, + 713 + ], + "type": "text", + "content": "-nearest lines and the set of " + }, + { + "bbox": [ + 46, + 522, + 287, + 713 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 522, + 287, + 713 + ], + "type": "text", + "content": "-nearest points which is a necessary condition for successful recovery of points from lines in [7]." + } + ] + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 307, + 74, + 378, + 121 + ], + "blocks": [ + { + "bbox": [ + 307, + 74, + 378, + 121 + ], + "lines": [ + { + "bbox": [ + 307, + 74, + 378, + 121 + ], + "spans": [ + { + "bbox": [ + 307, + 74, + 378, + 121 + ], + "type": "image", + "image_path": "0a7186f21b53f05692f48bea903a38b8f88983ac77c35b7bec61dad57493b244.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 326, + 121, + 359, + 129 + ], + "lines": [ + { + "bbox": [ + 326, + 121, + 359, + 129 + ], + "spans": [ + { + "bbox": [ + 326, + 121, + 359, + 129 + ], + "type": "text", + "content": "Point cloud" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 305, + 131, + 545, + 197 + ], + "lines": [ + { + "bbox": [ + 305, + 131, + 545, + 197 + ], + "spans": [ + { + "bbox": [ + 305, + 131, + 545, + 197 + ], + "type": "text", + "content": "Figure 8. The procedure of voxel-based ray sampling. After partitioning the point cloud into a voxel grid, the whole grid is divided into two clusters in such a way that neighboring grids belong to different clusters. Then, all 3D points in each grid within the same cluster are paired with one center, while the other cluster is paired with the remaining center for lifting." + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 380, + 76, + 421, + 119 + ], + "blocks": [ + { + "bbox": [ + 380, + 76, + 421, + 119 + ], + "lines": [ + { + "bbox": [ + 380, + 76, + 421, + 119 + ], + "spans": [ + { + "bbox": [ + 380, + 76, + 421, + 119 + ], + "type": "image", + "image_path": "e9c4f393c95b84e58da45ef0655353fcdff8be0c6ce41c964cef01e49263802d.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 386, + 121, + 417, + 129 + ], + "lines": [ + { + "bbox": [ + 386, + 121, + 417, + 129 + ], + "spans": [ + { + "bbox": [ + 386, + 121, + 417, + 129 + ], + "type": "text", + "content": "Voxel-grid" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 426, + 75, + 496, + 121 + ], + "blocks": [ + { + "bbox": [ + 426, + 75, + 496, + 121 + ], + "lines": [ + { + "bbox": [ + 426, + 75, + 496, + 121 + ], + "spans": [ + { + "bbox": [ + 426, + 75, + 496, + 121 + ], + "type": "image", + "image_path": "0923c6ea957cd37df3ebb649627c731d1bd44b1353907841e32c662808cc7df4.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 438, + 121, + 486, + 129 + ], + "lines": [ + { + "bbox": [ + 438, + 121, + 486, + 129 + ], + "spans": [ + { + "bbox": [ + 438, + 121, + 486, + 129 + ], + "type": "text", + "content": "Grid assignment" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 498, + 74, + 545, + 121 + ], + "blocks": [ + { + "bbox": [ + 498, + 74, + 545, + 121 + ], + "lines": [ + { + "bbox": [ + 498, + 74, + 545, + 121 + ], + "spans": [ + { + "bbox": [ + 498, + 74, + 545, + 121 + ], + "type": "image", + "image_path": "c8346fded2e911cfce79ce695ae1de17b31c5a5cc371e3a114c3630eb3fad2f2.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 513, + 121, + 531, + 129 + ], + "lines": [ + { + "bbox": [ + 513, + 121, + 531, + 129 + ], + "spans": [ + { + "bbox": [ + 513, + 121, + 531, + 129 + ], + "type": "text", + "content": "Lifting" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 205, + 545, + 324 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 205, + 545, + 324 + ], + "spans": [ + { + "bbox": [ + 304, + 205, + 545, + 324 + ], + "type": "text", + "content": "We adopt a 3D checkerboard pattern for assigning ray centers to voxels with aims to achieve stable camera localization accuracy by uniformly distributing rays from different ray centers across the scene. While a larger voxel size will allow more true nearest neighboring points to be discarded and provide better privacy protection, it will also cause the rays to be more non-uniform and lower the camera localization accuracy (see [32]). Currently, the voxel size needs to be manually determined for each dataset as the optimal trade-off point can vary across different scenes." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 306, + 335, + 432, + 348 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 335, + 432, + 348 + ], + "spans": [ + { + "bbox": [ + 306, + 335, + 432, + 348 + ], + "type": "text", + "content": "6. Experimental Results" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 305, + 355, + 545, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 355, + 545, + 392 + ], + "spans": [ + { + "bbox": [ + 305, + 355, + 545, + 392 + ], + "type": "text", + "content": "We present comprehensive experimental results showing the effectiveness of ray cloud in terms of localization accuracy, inference speed, and privacy-preserving performance." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 304, + 395, + 546, + 538 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 395, + 546, + 538 + ], + "spans": [ + { + "bbox": [ + 304, + 395, + 546, + 538 + ], + "type": "text", + "content": "Datasets We used the same two public datasets as in [7, 26]. The Energy Landscape [55] dataset consists of 12 indoor scenes with multiple sequences of scanned RGB images. The Cambridge [21, 22] dataset comprises 6 distinct outdoor landmarks, containing multiple image frames with labeled 6-DOF camera poses for the camera re-localization. As in [26], we excluded the scenes of Street and Great Court from [22] due to the presence of large geometric outliers. We followed the same experimental protocols from [26, 50] for constructing sparse 3D point clouds using COLMAP [44] and evaluation. Separated queries not included in 3D map construction were used for evaluation." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 304, + 542, + 545, + 625 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 542, + 545, + 625 + ], + "spans": [ + { + "bbox": [ + 304, + 542, + 545, + 625 + ], + "type": "text", + "content": "Implementation details We used Poselib [24] to utilize the " + }, + { + "bbox": [ + 304, + 542, + 545, + 625 + ], + "type": "inline_equation", + "content": "p5 + 1\\mathrm{R}" + }, + { + "bbox": [ + 304, + 542, + 545, + 625 + ], + "type": "text", + "content": " solver and the repository of PPL [26] to run experiments. We report coarse estimation results of inversion in ray cloud and voxel-based ray cloud as the refine estimation [7] diverges due to incorrect initial estimation (see [32]). We used a workstation with an Intel CPU i9-13900K running at 3.0 GHz and NVIDIA RTX 4090 GPUs." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "content": "Evaluation metrics For camera localization, we estimated the absolute 6-DOF pose of the query image with given 2D-3D correspondences. We define " + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathsf{R} \\in SO(3)" + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{t} \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "content": " as the ground truth camera rotation and translation respectively, and " + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{R}} \\in SO(3)" + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{t}} \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "content": " as the estimated rotation and translation respectively. Then, we obtained the estimated pose errors by following [26, 50, 51], where the" + } + ] + } + ], + "index": 27 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "9778" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 54, + 67, + 538, + 215 + ], + "blocks": [ + { + "bbox": [ + 54, + 67, + 538, + 215 + ], + "lines": [ + { + "bbox": [ + 54, + 67, + 538, + 215 + ], + "spans": [ + { + "bbox": [ + 54, + 67, + 538, + 215 + ], + "type": "table", + "html": "
DatasetScene representationMinimal solverPose errorRANSACPrivacy preserving
ΔR↓meanmedianΔt↓meanmediani meanr meant ↓meaneg↑ medianPSNR ↓ meanSSIM ↓ meanMAE ↑ mean
Energy LandscapePoint cloudp3P0.1780.0310.0070.0011199.544-15.760.51633.29
Uniform line cloud [50]p6L0.2720.0980.0100.00312481.12900.04513.450.40342.83
PPL [26]p6L0.2920.1130.0110.00412381.51880.05610.420.35562.61
PPL+ [26]p6L0.2930.1150.0110.00412581.30850.05610.390.35562.70
Ray cloud (ours)p5+1R0.2620.0870.0100.00314680.0360.66910.290.31962.97
Voxel-based ray cloud (ours)p5+1R0.2970.1160.0110.00414480.1560.77010.050.31564.84
CambridgePoint cloudp3P0.7000.1010.1420.0431298.487-14.000.43339.40
Uniform line cloud [50]p6L0.7750.1690.1720.06643671.872890.74512.710.33146.08
PPL [26]p6L0.7750.1830.1810.08126474.491860.97711.210.29655.53
PPL+ [26]p6L0.7640.1830.1780.08128474.111900.94811.160.29755.91
Ray cloud (ours)p5+1R0.8550.2220.2190.10723974.89129.61410.410.26160.98
Voxel-based ray cloud (ours)p5+1R0.8800.2560.2350.12025075.241210.5410.230.25362.56
", + "image_path": "27c0946544e607ff07613763855963a08fc2f8dfab2d133b1d6461dc7370be83.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 217, + 547, + 262 + ], + "lines": [ + { + "bbox": [ + 47, + 217, + 547, + 262 + ], + "spans": [ + { + "bbox": [ + 47, + 217, + 547, + 262 + ], + "type": "text", + "content": "Table 1. Comparison of localization and privacy-preserving performance with different 3D scene representations (" + }, + { + "bbox": [ + 47, + 217, + 547, + 262 + ], + "type": "inline_equation", + "content": "\\Delta R" + }, + { + "bbox": [ + 47, + 217, + 547, + 262 + ], + "type": "text", + "content": ": rotation error " + }, + { + "bbox": [ + 47, + 217, + 547, + 262 + ], + "type": "inline_equation", + "content": "[\\circ]" + }, + { + "bbox": [ + 47, + 217, + 547, + 262 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 47, + 217, + 547, + 262 + ], + "type": "inline_equation", + "content": "\\Delta t" + }, + { + "bbox": [ + 47, + 217, + 547, + 262 + ], + "type": "text", + "content": ": translation error [m], i: number of iterations, r: inlier ratio (\\%), t: runtime [ms], " + }, + { + "bbox": [ + 47, + 217, + 547, + 262 + ], + "type": "inline_equation", + "content": "e_g" + }, + { + "bbox": [ + 47, + 217, + 547, + 262 + ], + "type": "text", + "content": ": geometric (3D point) error [m]). We applied modified inversion attack from Sec. 4 with 25% line rejection to ray cloud and voxel-based ray cloud with voxel size of 0.5m and 5.0m for Energy Landscape and Cambridge respectively. Bold indicate the best result in each metric among privacy-preserving representations." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 57, + 264, + 541, + 380 + ], + "blocks": [ + { + "bbox": [ + 57, + 264, + 541, + 380 + ], + "lines": [ + { + "bbox": [ + 57, + 264, + 541, + 380 + ], + "spans": [ + { + "bbox": [ + 57, + 264, + 541, + 380 + ], + "type": "image", + "image_path": "eb107fcb95284603b7c0af91d1949ca09f0cdb3dd6c1211101ca5468ca390fac.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 381, + 536, + 392 + ], + "lines": [ + { + "bbox": [ + 55, + 381, + 536, + 392 + ], + "spans": [ + { + "bbox": [ + 55, + 381, + 536, + 392 + ], + "type": "text", + "content": "Figure 9. Comparison of revealed scene geometries and images for Apt2 bed [55]. Modified inversion attack is applied to (d) and (e)." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 398, + 288, + 427 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 398, + 288, + 427 + ], + "spans": [ + { + "bbox": [ + 46, + 398, + 288, + 427 + ], + "type": "text", + "content": "rotational error is computed as " + }, + { + "bbox": [ + 46, + 398, + 288, + 427 + ], + "type": "inline_equation", + "content": "\\Delta \\mathsf{R} = \\arccos \\frac{\\mathrm{Tr}(\\hat{\\mathsf{R}}^{\\top}\\mathsf{R}) - 1}{2}" + }, + { + "bbox": [ + 46, + 398, + 288, + 427 + ], + "type": "text", + "content": " and the translation error computed as " + }, + { + "bbox": [ + 46, + 398, + 288, + 427 + ], + "type": "inline_equation", + "content": "\\Delta \\mathbf{t} = \\left\\| \\hat{\\mathsf{R}}^{\\top}\\hat{\\mathbf{t}} -\\mathsf{R}^{\\top}\\mathbf{t}\\right\\|_{2}" + }, + { + "bbox": [ + 46, + 398, + 288, + 427 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 427, + 287, + 522 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 427, + 287, + 522 + ], + "spans": [ + { + "bbox": [ + 46, + 427, + 287, + 522 + ], + "type": "text", + "content": "To compare the quality of recovered point clouds, we calculated the 3D median point errors between the original and estimated points in meters as in [26]. We also compared the quality of revealed images by feeding recovered point cloud and ground truth camera pose of query images to InvSfM [39] and calculating peak signal-to-noise ratio (PSNR), structural similarity index measure (SSIM), and mean absolute error (MAE) values against the queries." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 528, + 270, + 540 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 528, + 270, + 540 + ], + "spans": [ + { + "bbox": [ + 47, + 528, + 270, + 540 + ], + "type": "text", + "content": "6.1. Camera localization accuracy and runtime" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 546, + 287, + 642 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 546, + 287, + 642 + ], + "spans": [ + { + "bbox": [ + 46, + 546, + 287, + 642 + ], + "type": "text", + "content": "As shown in Table 1, our ray cloud family of approaches achieved the fastest inference speed among the line-lifting methods with no significant decline in pose accuracy. Notably, these achieved real-time performance on both datasets, reaching 137 FPS for the ray cloud and 136 FPS for the voxel-based ray cloud on average. As further demonstrated in [32], this is mainly due to the introduction of the " + }, + { + "bbox": [ + 46, + 546, + 287, + 642 + ], + "type": "inline_equation", + "content": "p5 + 1\\mathrm{R}" + }, + { + "bbox": [ + 46, + 546, + 287, + 642 + ], + "type": "text", + "content": " solver which is only available for ray clouds." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 642, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 642, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 642, + 288, + 714 + ], + "type": "text", + "content": "The ray cloud showed the lowest localization error on [55] but slightly higher errors than ULC and PPL on [22]. We conjecture this is due to the increased baseline between two ray centers for the outdoor scenes. We also note that the voxel-based ray cloud does not exhibit a significant reduction in pose error compared to the standard ray cloud." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 399, + 545, + 412 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 399, + 545, + 412 + ], + "spans": [ + { + "bbox": [ + 305, + 399, + 545, + 412 + ], + "type": "text", + "content": "6.2. Quality of revealed scene geometry and images" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 418, + 546, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 418, + 546, + 514 + ], + "spans": [ + { + "bbox": [ + 304, + 418, + 546, + 514 + ], + "type": "text", + "content": "The quantitative comparisons of scene reconstruction quality are also shown in Table 1. For a fair comparison, we applied our adapted inversion attack in Sec. 4 for ray cloud and voxel-based ray cloud, while we used the original geometry inversion attack [7] with 3 refinement steps for uniform line cloud (ULC), PPL and " + }, + { + "bbox": [ + 304, + 418, + 546, + 514 + ], + "type": "inline_equation", + "content": "\\mathrm{PPL + }" + }, + { + "bbox": [ + 304, + 418, + 546, + 514 + ], + "type": "text", + "content": " .We set " + }, + { + "bbox": [ + 304, + 418, + 546, + 514 + ], + "type": "inline_equation", + "content": "\\beta = 25\\%" + }, + { + "bbox": [ + 304, + 418, + 546, + 514 + ], + "type": "text", + "content": " for the ray rejection threshold as it shows the best point estimation in Table 2. Results with " + }, + { + "bbox": [ + 304, + 418, + 546, + 514 + ], + "type": "inline_equation", + "content": "\\beta = 50\\%" + }, + { + "bbox": [ + 304, + 418, + 546, + 514 + ], + "type": "text", + "content": " are included in [32]." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 517, + 547, + 625 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 517, + 547, + 625 + ], + "spans": [ + { + "bbox": [ + 304, + 517, + 547, + 625 + ], + "type": "text", + "content": "Recovery of 3D point clouds The median point errors for ray cloud and voxel-based ray cloud are notably larger than those of other methods, even after applying our adjusted density-based attacks. Additionally, to validate the impact of our adapted inversion attack, we compared the geometric error with and without rejection sampling, showing that rejection sampling contributes to error reduction in Table 2. The qualitative result in Fig. 9 shows that the results from ray cloud and voxel-based ray cloud are unrecognizable." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 630, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 630, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 630, + 547, + 714 + ], + "type": "text", + "content": "Degraded quality of scene details Fig. 9 shows the images recovered from the ray cloud and the voxel-based ray cloud are hardly recognizable. In Table. 1, the voxel-based ray cloud yields the lowest PSNR, SSIM and MAE values with the original ray cloud as a runner-up. This indicates large point errors exhibited by ray clouds translate to lower image quality." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "9779" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 62, + 72, + 216, + 135 + ], + "blocks": [ + { + "bbox": [ + 62, + 72, + 216, + 135 + ], + "lines": [ + { + "bbox": [ + 62, + 72, + 216, + 135 + ], + "spans": [ + { + "bbox": [ + 62, + 72, + 216, + 135 + ], + "type": "image", + "image_path": "5642d4b78ddbd722cc01f081a25ccb06d659ffb0084eb81f77db70f02137531b.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 222, + 72, + 375, + 135 + ], + "blocks": [ + { + "bbox": [ + 222, + 72, + 375, + 135 + ], + "lines": [ + { + "bbox": [ + 222, + 72, + 375, + 135 + ], + "spans": [ + { + "bbox": [ + 222, + 72, + 375, + 135 + ], + "type": "image", + "image_path": "5d7d4a00975b879f1f8ee02fcaa94a1077afac9292ecbf3e04a9865961a01d54.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 381, + 72, + 536, + 135 + ], + "blocks": [ + { + "bbox": [ + 381, + 72, + 536, + 135 + ], + "lines": [ + { + "bbox": [ + 381, + 72, + 536, + 135 + ], + "spans": [ + { + "bbox": [ + 381, + 72, + 536, + 135 + ], + "type": "image", + "image_path": "27a4ad7e2e4a5677d827d58c60993e60b4b7ee36271ed13907f6b3cb6fbde635.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 58, + 136, + 214, + 198 + ], + "blocks": [ + { + "bbox": [ + 58, + 136, + 214, + 198 + ], + "lines": [ + { + "bbox": [ + 58, + 136, + 214, + 198 + ], + "spans": [ + { + "bbox": [ + 58, + 136, + 214, + 198 + ], + "type": "image", + "image_path": "0f673db3068e79c2ce0b4021075ce17d09f9a6bf08e575697d53e0ff84ffdd18.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 219, + 136, + 375, + 199 + ], + "blocks": [ + { + "bbox": [ + 219, + 136, + 375, + 199 + ], + "lines": [ + { + "bbox": [ + 219, + 136, + 375, + 199 + ], + "spans": [ + { + "bbox": [ + 219, + 136, + 375, + 199 + ], + "type": "image", + "image_path": "a3155d9364c152845ad8dd9e4cd1ae44ec5237e81da7cd11ab4472044b1ccc2e.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 380, + 136, + 536, + 199 + ], + "blocks": [ + { + "bbox": [ + 380, + 136, + 536, + 199 + ], + "lines": [ + { + "bbox": [ + 380, + 136, + 536, + 199 + ], + "spans": [ + { + "bbox": [ + 380, + 136, + 536, + 199 + ], + "type": "image", + "image_path": "8434caf5240d2be464914b669eeff13aadb2c4d11235891eb4c97d4dcc3469d6.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 59, + 200, + 214, + 262 + ], + "blocks": [ + { + "bbox": [ + 59, + 200, + 214, + 262 + ], + "lines": [ + { + "bbox": [ + 59, + 200, + 214, + 262 + ], + "spans": [ + { + "bbox": [ + 59, + 200, + 214, + 262 + ], + "type": "image", + "image_path": "830a66628c4ef3b7b6cb7f91ee2fa9de75fb0382b3152a1d5c2639228662f83d.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 262, + 545, + 283 + ], + "lines": [ + { + "bbox": [ + 46, + 262, + 545, + 283 + ], + "spans": [ + { + "bbox": [ + 46, + 262, + 545, + 283 + ], + "type": "text", + "content": "Figure 10. Average localization performance vs. reconstructed image quality across different levels of the ray density on the Energy Landscape [55] dataset. The tested line densities are " + }, + { + "bbox": [ + 46, + 262, + 545, + 283 + ], + "type": "inline_equation", + "content": "100\\%" + }, + { + "bbox": [ + 46, + 262, + 545, + 283 + ], + "type": "text", + "content": " (biggest marker), " + }, + { + "bbox": [ + 46, + 262, + 545, + 283 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 46, + 262, + 545, + 283 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 262, + 545, + 283 + ], + "type": "inline_equation", + "content": "25\\%" + }, + { + "bbox": [ + 46, + 262, + 545, + 283 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 262, + 545, + 283 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 46, + 262, + 545, + 283 + ], + "type": "text", + "content": " (smallest marker)." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 220, + 200, + 374, + 262 + ], + "blocks": [ + { + "bbox": [ + 220, + 200, + 374, + 262 + ], + "lines": [ + { + "bbox": [ + 220, + 200, + 374, + 262 + ], + "spans": [ + { + "bbox": [ + 220, + 200, + 374, + 262 + ], + "type": "image", + "image_path": "758292b2a65000ee87c6e1a9a95cb05288f5927811877e7f27b5ddaaf04bb487.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 381, + 200, + 536, + 262 + ], + "blocks": [ + { + "bbox": [ + 381, + 200, + 536, + 262 + ], + "lines": [ + { + "bbox": [ + 381, + 200, + 536, + 262 + ], + "spans": [ + { + "bbox": [ + 381, + 200, + 536, + 262 + ], + "type": "image", + "image_path": "0994ad02842ca40f62e80b2cc7d3d15a8e9ffca8af370b61b178bf4b9c275b20.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 56, + 288, + 277, + 388 + ], + "blocks": [ + { + "bbox": [ + 56, + 288, + 277, + 388 + ], + "lines": [ + { + "bbox": [ + 56, + 288, + 277, + 388 + ], + "spans": [ + { + "bbox": [ + 56, + 288, + 277, + 388 + ], + "type": "table", + "html": "
DatasetRejection threshold (β)eg[m] medianPSNR meanSSIM meanMAE mean
Energy LandscapeNone1.0279.9590.33066.06
10 (%)0.76110.240.32363.62
25 (%)0.66910.290.31962.97
50 (%)0.74610.320.32062.52
CambridgeNone11.4710.100.27364.06
10 (%)10.2910.350.26761.60
25 (%)9.61410.410.26160.98
50 (%)12.989.9440.24264.70
", + "image_path": "951fed8a2095f6b00d7a9acb41f206e912174989e6daed63800e8ab80fde55a3.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 388, + 287, + 433 + ], + "lines": [ + { + "bbox": [ + 46, + 388, + 287, + 433 + ], + "spans": [ + { + "bbox": [ + 46, + 388, + 287, + 433 + ], + "type": "text", + "content": "Table 2. Effect of ray rejection (see Sec. 4.2) on the quality of scene recovery. Rejecting rays close to the baseline of two ray centers overall reduces the point estimation errors. Bold represents the best inversion results amongst different thresholds." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 441, + 287, + 560 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 441, + 287, + 560 + ], + "spans": [ + { + "bbox": [ + 46, + 441, + 287, + 560 + ], + "type": "text", + "content": "Additionally, Fig. 10 comparing the pose estimation errors against image reconstruction qualities demonstrate that the ray cloud-based approaches are capable of preserving the privacy especially for dense scenes while significantly decreasing the inference runtime. In contrast, the voxel-based ray cloud shows slightly higher pose estimation errors potentially due to coagulation of rays yielding less uniform distribution of rays across the scene. The fact that the SSIM value increases for lower density ray clouds also needs further investigation (see [32] for qualitative results)." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 46, + 561, + 287, + 609 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 561, + 287, + 609 + ], + "spans": [ + { + "bbox": [ + 46, + 561, + 287, + 609 + ], + "type": "text", + "content": "In Fig. 11, we demonstrate some corner cases revealed through our adapted inversion attack in Sec. 4. Fig. 11d shows the voxel-based ray sampling proposed in Sec. 5 can degrade the quality of scene recovery." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 47, + 621, + 119, + 633 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 621, + 119, + 633 + ], + "spans": [ + { + "bbox": [ + 47, + 621, + 119, + 633 + ], + "type": "text", + "content": "7. Conclusion" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 46, + 642, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 642, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 642, + 287, + 713 + ], + "type": "text", + "content": "We presented a privacy-preserving visual localization framework based on a new scene representation called 3D ray cloud, which is robust to the known geometry inversion attack and achieves an order of magnitude faster localization than previous line-based approaches without significantly compromising the localization accuracy. This is" + } + ] + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 315, + 289, + 369, + 395 + ], + "blocks": [ + { + "bbox": [ + 315, + 289, + 369, + 395 + ], + "lines": [ + { + "bbox": [ + 315, + 289, + 369, + 395 + ], + "spans": [ + { + "bbox": [ + 315, + 289, + 369, + 395 + ], + "type": "image", + "image_path": "0d983a917d4d9abb839885a7fd993bd44ed790063e17117807e29f89aef6c84c.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 315, + 397, + 367, + 406 + ], + "lines": [ + { + "bbox": [ + 315, + 397, + 367, + 406 + ], + "spans": [ + { + "bbox": [ + 315, + 397, + 367, + 406 + ], + "type": "text", + "content": "(a) ground truth" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 372, + 288, + 425, + 342 + ], + "blocks": [ + { + "bbox": [ + 372, + 288, + 425, + 342 + ], + "lines": [ + { + "bbox": [ + 372, + 288, + 425, + 342 + ], + "spans": [ + { + "bbox": [ + 372, + 288, + 425, + 342 + ], + "type": "image", + "image_path": "55b9e92788ca61edade378f77089139dc0824f10b29d680aa9582e8823fc2bb7.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 372, + 342, + 425, + 396 + ], + "blocks": [ + { + "bbox": [ + 372, + 342, + 425, + 396 + ], + "lines": [ + { + "bbox": [ + 372, + 342, + 425, + 396 + ], + "spans": [ + { + "bbox": [ + 372, + 342, + 425, + 396 + ], + "type": "image", + "image_path": "39b2f6884efa3440ead7868be06e85e1090736218f613c692e94517471d8b1c9.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 371, + 396, + 425, + 406 + ], + "lines": [ + { + "bbox": [ + 371, + 396, + 425, + 406 + ], + "spans": [ + { + "bbox": [ + 371, + 396, + 425, + 406 + ], + "type": "text", + "content": "(b) w/o rejection" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 429, + 289, + 482, + 342 + ], + "blocks": [ + { + "bbox": [ + 429, + 289, + 482, + 342 + ], + "lines": [ + { + "bbox": [ + 429, + 289, + 482, + 342 + ], + "spans": [ + { + "bbox": [ + 429, + 289, + 482, + 342 + ], + "type": "image", + "image_path": "8582d96efe7f038d7a7399a806e7f1c224b75d9e14c03374bbb8d1702b5b70a7.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 429, + 342, + 482, + 396 + ], + "blocks": [ + { + "bbox": [ + 429, + 342, + 482, + 396 + ], + "lines": [ + { + "bbox": [ + 429, + 342, + 482, + 396 + ], + "spans": [ + { + "bbox": [ + 429, + 342, + 482, + 396 + ], + "type": "image", + "image_path": "29ee5d1d0c067f767159fbe18b77c190e34b72461c5a17fbebb85b54a04f6409.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 429, + 396, + 480, + 406 + ], + "lines": [ + { + "bbox": [ + 429, + 396, + 480, + 406 + ], + "spans": [ + { + "bbox": [ + 429, + 396, + 480, + 406 + ], + "type": "text", + "content": "(c) w/ rejection" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 486, + 289, + 540, + 342 + ], + "blocks": [ + { + "bbox": [ + 486, + 289, + 540, + 342 + ], + "lines": [ + { + "bbox": [ + 486, + 289, + 540, + 342 + ], + "spans": [ + { + "bbox": [ + 486, + 289, + 540, + 342 + ], + "type": "image", + "image_path": "4b0ddd2170e980e7e7c1ad24511abae142139e0b2ffaed688b1a7ecbd51bb5b7.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 486, + 342, + 539, + 396 + ], + "blocks": [ + { + "bbox": [ + 486, + 342, + 539, + 396 + ], + "lines": [ + { + "bbox": [ + 486, + 342, + 539, + 396 + ], + "spans": [ + { + "bbox": [ + 486, + 342, + 539, + 396 + ], + "type": "image", + "image_path": "ba40c2e6c34ee9efcd1fe4be39c61baf2c6c8b75f6609ab9fc8f7bc94495f17d.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 486, + 396, + 536, + 406 + ], + "lines": [ + { + "bbox": [ + 486, + 396, + 536, + 406 + ], + "spans": [ + { + "bbox": [ + 486, + 396, + 536, + 406 + ], + "type": "text", + "content": "(d) voxel-based" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_caption" + } + ], + "index": 25 + }, + { + "bbox": [ + 305, + 408, + 545, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 408, + 545, + 475 + ], + "spans": [ + { + "bbox": [ + 305, + 408, + 545, + 475 + ], + "type": "text", + "content": "Figure 11. Illustration of point-to-image translation [40] results with point clouds from (a) ground truth, (b) recovered from ray cloud without rejection sampling, (c) recovered with rejection sampling, and (d) recovered from voxel-based ray cloud. While (c) shows improved recovery via ray rejection in Sec. 4.2, (d) shows voxel-based ray sampling lowering the reconstruction quality." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 304, + 479, + 545, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 479, + 545, + 540 + ], + "spans": [ + { + "bbox": [ + 304, + 479, + 545, + 540 + ], + "type": "text", + "content": "enabled by the intrinsic geometry of the ray cloud allowing utilization of efficient widely-available perspective relative pose estimation solver. We have also explored potential improvements for the geometry-inversion attack on ray clouds and proposed a respective defense mechanism." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 304, + 543, + 545, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 543, + 545, + 662 + ], + "spans": [ + { + "bbox": [ + 304, + 543, + 545, + 662 + ], + "type": "text", + "content": "Limitation Although this work can successfully defend the known inversion attack [7], we note it may be potentially vulnerable to a new type of attack in which omnidirectional rays are projected to a cylindrical or spherical image plane which is then unfolded and directly passed through an inversion network for image reconstruction. While realizing this attack may face its own challenges such as unknown depths and mixed chirality, we believe utilizing more ray centers will provide a defense against such attack by reducing the density of ray projections in each ray center." + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "content": "Acknowledgement This work was supported by the NRF (Natioanl Research Foundation of Korea) grants funded by the Korea government (MSIT) (No. 2022R1C1C1004907). We thank anonymous reviewers for their valuable feedback." + } + ] + } + ], + "index": 30 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "9780" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 704 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 133 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 133 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 133 + ], + "type": "text", + "content": "[1] Sameer Agarwal, Yasutaka Furukawa, Noah Snavely, Ian Simon, Brian Curless, Steven M Seitz, and Richard Szeliski. Building Rome in a day. Communications of the ACM, 54(10):105-112, 2011. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 134, + 288, + 177 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 134, + 288, + 177 + ], + "spans": [ + { + "bbox": [ + 53, + 134, + 288, + 177 + ], + "type": "text", + "content": "[2] Alexandre Alahi, Raphael Ortiz, and Pierre Vandergheynst. Freak: Fast retina keypoint. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 510-517, 2012. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 177, + 288, + 231 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 177, + 288, + 231 + ], + "spans": [ + { + "bbox": [ + 53, + 177, + 288, + 231 + ], + "type": "text", + "content": "[3] Clemens Arth, Daniel Wagner, Manfred Klopschitz, Arnold Irschara, and Dieter Schmalstieg. Wide area localization on mobile phones. In 2009 8th IEEE International Symposium on Mixed and Augmented Reality, pages 73-82. IEEE, 2009. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 231, + 288, + 295 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 231, + 288, + 295 + ], + "spans": [ + { + "bbox": [ + 53, + 231, + 288, + 295 + ], + "type": "text", + "content": "[4] Eric Brachmann, Alexander Krull, Sebastian Nowozin, Jamie Shotton, Frank Michel, Stefan Gumhold, and Carsten Rother. DSAC — Differentiable RANSAC for camera localization. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 2492–2500, 2017. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 296, + 288, + 339 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 296, + 288, + 339 + ], + "spans": [ + { + "bbox": [ + 53, + 296, + 288, + 339 + ], + "type": "text", + "content": "[5] Eric Brachmann and Carsten Rother. Learning less is more - 6D camera localization via 3D surface regression. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4654-4662, 2018. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 339, + 288, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 339, + 288, + 381 + ], + "spans": [ + { + "bbox": [ + 53, + 339, + 288, + 381 + ], + "type": "text", + "content": "[6] Robert Castle, Georg Klein, and David W Murray. Videorelate localization in multiple maps for wearable augmented reality. In 2008 12th IEEE International Symposium on Wearable Computers, pages 15-22. IEEE, 2008. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 381, + 288, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 381, + 288, + 435 + ], + "spans": [ + { + "bbox": [ + 53, + 381, + 288, + 435 + ], + "type": "text", + "content": "[7] Kunal Chelani, Fredrik Kahl, and Torsten Sattler. How privacy-preserving are line clouds? Recovering scene details from 3D lines. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 15663-15673, 2021. 1, 2, 3, 4, 5, 6, 7, 8" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 435, + 288, + 490 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 435, + 288, + 490 + ], + "spans": [ + { + "bbox": [ + 53, + 435, + 288, + 490 + ], + "type": "text", + "content": "[8] Kunal Chelani, Torsten Sattler, Fredrik Kahl, and Zuzana Kukelova. Privacy-preserving representations are not enough: Recovering scene content from camera poses. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13132-13141, 2023. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 490, + 288, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 490, + 288, + 533 + ], + "spans": [ + { + "bbox": [ + 53, + 490, + 288, + 533 + ], + "type": "text", + "content": "[9] Ondrej Chum, Jiri Matas, and Josef Kittler. Locally optimized ransac. In Pattern Recognition: 25th DAGM Symposium, Magdeburg, Germany, September 10-12, 2003. Proceedings 25, pages 236-243. Springer, 2003. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 533, + 288, + 597 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 533, + 288, + 597 + ], + "spans": [ + { + "bbox": [ + 48, + 533, + 288, + 597 + ], + "type": "text", + "content": "[10] Deeksha Dangwal, Vincent T Lee, Hyo Jin Kim, Tianwei Shen, Meghan Cowan, Rajvi Shah, Caroline Trippel, Brandon Reagen, Timothy Sherwood, Vasileios Balntas, et al. Mitigating reverse engineering attacks on local feature descriptors. In Proceeding of the British Machine Vision Conference (BMVC), 2021. 1, 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 597, + 288, + 639 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 597, + 288, + 639 + ], + "spans": [ + { + "bbox": [ + 48, + 597, + 288, + 639 + ], + "type": "text", + "content": "[11] Andrew J. Davison, Ian D. Reid, Nicholas D. Molton, and Olivier Stasse. MonoSLAM: Real-time single camera SLAM. IEEE Transactions on Pattern Analysis and Machine Intelligence, 29(6):1052-1067, 2007. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 639, + 288, + 683 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 639, + 288, + 683 + ], + "spans": [ + { + "bbox": [ + 48, + 639, + 288, + 683 + ], + "type": "text", + "content": "[12] Yaqing Ding, Jian Yang, Viktor Larsson, Carl Olsson, and Kalle Åström. Revisiting the p3p problem. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4872-4880, 2023. 1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 683, + 288, + 704 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 683, + 288, + 704 + ], + "spans": [ + { + "bbox": [ + 48, + 683, + 288, + 704 + ], + "type": "text", + "content": "[13] Alexey Dosovitskiy and Thomas Brox. Inverting visual representations with convolutional networks. In Proceedings" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 707 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "text", + "content": "of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 4829-4837, 2016. 1, 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 95, + 545, + 158 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 95, + 545, + 158 + ], + "spans": [ + { + "bbox": [ + 307, + 95, + 545, + 158 + ], + "type": "text", + "content": "[14] Mihai Dusmanu, Johannes L. Schonberger, Sudipta N. Sinha, and Marc Pollefeys. Privacy-preserving image features via adversarial affine subspace embeddings. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 14262-14272, 2021. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 159, + 545, + 202 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 159, + 545, + 202 + ], + "spans": [ + { + "bbox": [ + 307, + 159, + 545, + 202 + ], + "type": "text", + "content": "[15] Martin A. Fischler and Robert C. Bolles. Random sample consensus: A paradigm for model fitting with applications to image analysis and automated cartography. Communications of the ACM, 24(6):381-395, June 1981. 1, 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 202, + 545, + 255 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 202, + 545, + 255 + ], + "spans": [ + { + "bbox": [ + 307, + 202, + 545, + 255 + ], + "type": "text", + "content": "[16] Marcel Geppert, Viktor Larsson, Johannes L Schonberger, and Marc Pollefeys. Privacy preserving partial localization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 17337-17347, 2022. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 255, + 545, + 277 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 255, + 545, + 277 + ], + "spans": [ + { + "bbox": [ + 307, + 255, + 545, + 277 + ], + "type": "text", + "content": "[17] Richard Hartley. Cheirality. International journal of computer vision, 26(1):41-61, 1998. 4" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 277, + 545, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 277, + 545, + 353 + ], + "spans": [ + { + "bbox": [ + 307, + 277, + 545, + 353 + ], + "type": "text", + "content": "[18] Lionel Heng, Benjamin Choi, Zhaopeng Cui, Marcel Geppert, Sixing Hu, Benson Kuan, Peidong Liu, Rang Nguyen, Ye Chuan Yeo, Andreas Geiger, et al. Project automation: Localization and 3d scene perception for an autonomous vehicle with a multi-camera system. In Proceedings of the IEEE International Conference on Robotics and Automation (ICRA), pages 4695-4702, 2019. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 353, + 545, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 353, + 545, + 396 + ], + "spans": [ + { + "bbox": [ + 307, + 353, + 545, + 396 + ], + "type": "text", + "content": "[19] Hiroharu Kato and Tatsuya Harada. Image reconstruction from bag-of-visual-words. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 955–962, 2014. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 396, + 545, + 438 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 396, + 545, + 438 + ], + "spans": [ + { + "bbox": [ + 307, + 396, + 545, + 438 + ], + "type": "text", + "content": "[20] Tong Ke and Stergios I Roumeliotis. An efficient algebraic solution to the perspective-three-point problem. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 7225-7233, 2017. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 438, + 545, + 481 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 438, + 545, + 481 + ], + "spans": [ + { + "bbox": [ + 307, + 438, + 545, + 481 + ], + "type": "text", + "content": "[21] Alex Kendall and Roberto Cipolla. Geometric loss functions for camera pose regression with deep learning. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 6555-6564, 2017. 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 481, + 545, + 534 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 481, + 545, + 534 + ], + "spans": [ + { + "bbox": [ + 307, + 481, + 545, + 534 + ], + "type": "text", + "content": "[22] Alex Kendall, Matthew Grimes, and Roberto Cipolla. PoseNet: A convolutional network for real-time 6-DOF camera relocalization. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), pages 2938-2946, 2015. 1, 6, 7" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 535, + 545, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 535, + 545, + 567 + ], + "spans": [ + { + "bbox": [ + 307, + 535, + 545, + 567 + ], + "type": "text", + "content": "[23] Nicolaas H Kuiper. Tests concerning random points on a circle. In Nederl. Akad. Wetensch. Proc. Ser. A, volume 63, pages 38-47, 1960. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 567, + 545, + 598 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 567, + 545, + 598 + ], + "spans": [ + { + "bbox": [ + 307, + 567, + 545, + 598 + ], + "type": "text", + "content": "[24] Viktor Larsson. PoseLib - Minimal Solvers for Camera Pose Estimation. https://github.com/vlarsson/ PoseLib, 2020. Accessed: 2022-10-30. 6" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 598, + 545, + 652 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 598, + 545, + 652 + ], + "spans": [ + { + "bbox": [ + 307, + 598, + 545, + 652 + ], + "type": "text", + "content": "[25] Viktor Larsson, Kalle Astrom, and Magnus Oskarsson. Efficient solvers for minimal problems by syzygy-based reduction. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 820-829, 2017. 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 652, + 545, + 707 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 652, + 545, + 707 + ], + "spans": [ + { + "bbox": [ + 307, + 652, + 545, + 707 + ], + "type": "text", + "content": "[26] Chunghwan Lee, Jaihoon Kim, Chanhyuk Yun, and Je Hyeong Hong. Paired-point lifting for enhanced privacy-preserving visual localization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 17266-17275, 2023. 1, 2, 3, 4," + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 314, + 757 + ], + "type": "text", + "content": "9781" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 74, + 287, + 706 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 68, + 74, + 83, + 83 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 74, + 83, + 83 + ], + "spans": [ + { + "bbox": [ + 68, + 74, + 83, + 83 + ], + "type": "text", + "content": "6,7" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 83, + 287, + 137 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 83, + 287, + 137 + ], + "spans": [ + { + "bbox": [ + 48, + 83, + 287, + 137 + ], + "type": "text", + "content": "[27] Hyon Lim, Sudipta N Sinha, Michael F Cohen, and Matthew Uytendaele. Real-time image-based 6-DOF localization in large-scale environments. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 1043–1050. IEEE, 2012. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 137, + 287, + 169 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 137, + 287, + 169 + ], + "spans": [ + { + "bbox": [ + 48, + 137, + 287, + 169 + ], + "type": "text", + "content": "[28] David G. Lowe. Distinctive image features from scale-invariant keypoints. International Journal of Computer Vision, 60(2):91-110, Nov. 2004. 1, 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 169, + 287, + 213 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 169, + 287, + 213 + ], + "spans": [ + { + "bbox": [ + 48, + 169, + 287, + 213 + ], + "type": "text", + "content": "[29] Simon Lynen, Torsten Sattler, Michael Bosse, Joel A Hesch, Marc Pollefeys, and Roland Siegwart. Get out of my lab: Large-scale, real-time visual-inertial localization. In Robotics: Science and Systems, volume 1, page 1, 2015. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 213, + 287, + 255 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 213, + 287, + 255 + ], + "spans": [ + { + "bbox": [ + 48, + 213, + 287, + 255 + ], + "type": "text", + "content": "[30] Aravindh Mahendran and Andrea Vedaldi. Understanding deep image representations by inverting them. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 5188-5196, 2015. 1, 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 255, + 287, + 298 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 255, + 287, + 298 + ], + "spans": [ + { + "bbox": [ + 48, + 255, + 287, + 298 + ], + "type": "text", + "content": "[31] Anastasiia Mishchuk, Dmytro Mishkin, Filip Radenovic, and Jiri Matas. Working hard to know your neighbor's margins: Local descriptor learning loss. Advances in Neural Information Processing Systems (NeurIPS), 30, 2017. 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 298, + 287, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 298, + 287, + 330 + ], + "spans": [ + { + "bbox": [ + 48, + 298, + 287, + 330 + ], + "type": "text", + "content": "[32] Heejoon Moon, Chunghwan Lee, and Je Hyeong Hong. Supplementary document for efficient privacy-preserving visual localization using 3d ray clouds, 2024. 3, 5, 6, 7, 8" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 330, + 287, + 372 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 330, + 287, + 372 + ], + "spans": [ + { + "bbox": [ + 48, + 330, + 287, + 372 + ], + "type": "text", + "content": "[33] Raul Mur-Artal, Jose Maria Martinez Montiel, and Juan D Tardos. ORB-SLAM: A versatile and accurate monocular SLAM system. IEEE Transactions on Robotics, 31(5):1147-1163, 2015. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 373, + 287, + 448 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 373, + 287, + 448 + ], + "spans": [ + { + "bbox": [ + 48, + 373, + 287, + 448 + ], + "type": "text", + "content": "[34] Tony Ng, Hyo Jin Kim, Vincent T Lee, Daniel DeTone, Tsun-Yi Yang, Tianwei Shen, Eddy Ilg, Vassileios Balntas, Krystian Mikolajczyk, and Chris Sweeney. NinjaDesc: Content-concealing visual descriptors via adversarial learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12797-12807, 2022. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 448, + 287, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 448, + 287, + 480 + ], + "spans": [ + { + "bbox": [ + 48, + 448, + 287, + 480 + ], + "type": "text", + "content": "[35] David Nister. An efficient solution to the five-point relative pose problem. IEEE Transactions on Pattern Analysis and Machine Intelligence, 26(6):756-770, 2004. 2, 3, 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 480, + 287, + 523 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 480, + 287, + 523 + ], + "spans": [ + { + "bbox": [ + 48, + 480, + 287, + 523 + ], + "type": "text", + "content": "[36] Linfei Pan, Johannes Lutz Schonberger, Viktor Larsson, and Marc Pollefeys. Privacy Preserving Localization via Coordinate Permutations. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), 2023. 2, 3, 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 523, + 287, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 523, + 287, + 566 + ], + "spans": [ + { + "bbox": [ + 48, + 523, + 287, + 566 + ], + "type": "text", + "content": "[37] Mikael Persson and Klas Nordberg. Lambda twist: An accurate fast robust perspective three point (p3P) solver. In Proceedings of the European conference on computer vision (ECCV), pages 318-332, 2018. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 566, + 287, + 620 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 566, + 287, + 620 + ], + "spans": [ + { + "bbox": [ + 48, + 566, + 287, + 620 + ], + "type": "text", + "content": "[38] Maxime Pietrantoni, Martin Humenberger, Torsten Sattler, and Gabriela Csurka. Segloc: Learning segmentation-based representations for privacy-preserving visual localization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15380-15391, 2023. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 620, + 287, + 673 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 620, + 287, + 673 + ], + "spans": [ + { + "bbox": [ + 48, + 620, + 287, + 673 + ], + "type": "text", + "content": "[39] Francesco Pittaluga, Sanjeev Koppal, and Ayan Chakrabarti. Learning privacy preserving encodings through adversarial training. In Proceedings of the IEEE Winter Conference on Applications of Computer Vision (WACV), pages 791-799. IEEE, 2019. 1, 7" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 673, + 287, + 706 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 673, + 287, + 706 + ], + "spans": [ + { + "bbox": [ + 48, + 673, + 287, + 706 + ], + "type": "text", + "content": "[40] Francesco Pittaluga, Sanjeev J Koppal, Sing Bing Kang, and Sudipta N Sinha. Revealing scenes by inverting structure from motion reconstructions. In Proceedings of the" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 707 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "text", + "content": "IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 145-154, 2019. 1, 2, 3, 8" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 95, + 545, + 137 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 95, + 545, + 137 + ], + "spans": [ + { + "bbox": [ + 307, + 95, + 545, + 137 + ], + "type": "text", + "content": "[41] Francesco Pittaluga and Bingbing Zhuang. Ldp-feat: Image features with local differential privacy. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 17580–17590, 2023. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 137, + 545, + 190 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 137, + 545, + 190 + ], + "spans": [ + { + "bbox": [ + 307, + 137, + 545, + 190 + ], + "type": "text", + "content": "[42] Rahul Raguram, Ondrej Chum, Marc Pollefeys, Jiri Matas, and Jan-Michael Frahm. USAC: A universal framework for random sample consensus. IEEE Transactions on Pattern Analysis and Machine Intelligence, 35(8):2022-2038, 2012. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 190, + 545, + 245 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 190, + 545, + 245 + ], + "spans": [ + { + "bbox": [ + 307, + 190, + 545, + 245 + ], + "type": "text", + "content": "[43] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-Net: Convolutional networks for biomedical image segmentation. In Proceedings of the International Conference on Medical Image Computing and Computer-Assisted Intervention (MICCAI), pages 234–241. Springer, 2015. 1, 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 245, + 545, + 288 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 245, + 545, + 288 + ], + "spans": [ + { + "bbox": [ + 307, + 245, + 545, + 288 + ], + "type": "text", + "content": "[44] Johannes Lutz Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 1, 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 288, + 545, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 288, + 545, + 353 + ], + "spans": [ + { + "bbox": [ + 307, + 288, + 545, + 353 + ], + "type": "text", + "content": "[45] Johannes Lutz Schonberger, Enliang Zheng, Marc Pollefeys, and Jan-Michael Frahm. Pixelwise view selection for unstructured multi-view stereo. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part III 14, pages 501-518. Springer, 2016. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 353, + 545, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 353, + 545, + 407 + ], + "spans": [ + { + "bbox": [ + 307, + 353, + 545, + 407 + ], + "type": "text", + "content": "[46] Mikiya Shibuya, Shinya Sumikura, and Ken Sakurada. Privacy preserving visual SLAM. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXII 16, pages 102-118. Springer, 2020. 1, 2, 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 407, + 545, + 449 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 407, + 545, + 449 + ], + "spans": [ + { + "bbox": [ + 307, + 407, + 545, + 449 + ], + "type": "text", + "content": "[47] Sudipta Narayan Sinha, Pablo Alejandro Speciale, Sing Bing Kang, and Marc Andre Leon Pollefeys. Device pose estimation using 3d line clouds, Mar. 30 2021. US Patent 10,964,053. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 449, + 545, + 482 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 449, + 545, + 482 + ], + "spans": [ + { + "bbox": [ + 307, + 449, + 545, + 482 + ], + "type": "text", + "content": "[48] Noah Snavely, Steven M Seitz, and Richard Szeliski. Modeling the world from internet photo collections. International Journal of Computer Vision (IJCV), 80:189-210, 2008. 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 481, + 545, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 481, + 545, + 536 + ], + "spans": [ + { + "bbox": [ + 307, + 481, + 545, + 536 + ], + "type": "text", + "content": "[49] Zhenbo Song, Wayne Chen, Dylan Campbell, and Hongdong Li. Deep novel view synthesis from colored 3d point clouds. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXIV 16, pages 1-17. Springer, 2020. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 535, + 545, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 535, + 545, + 589 + ], + "spans": [ + { + "bbox": [ + 307, + 535, + 545, + 589 + ], + "type": "text", + "content": "[50] Pablo Speciale, Johannes L Schonberger, Sing Bing Kang, Sudipta N Sinha, and Marc Pollefeys. Privacy preserving image-based localization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5493-5503, 2019. 1, 2, 3, 4, 5, 6, 7" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 589, + 545, + 642 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 589, + 545, + 642 + ], + "spans": [ + { + "bbox": [ + 307, + 589, + 545, + 642 + ], + "type": "text", + "content": "[51] Pablo Speciale, Johannes L Schonberger, Sudipta N Sinha, and Marc Pollefeys. Privacy preserving image queries for camera localization. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 1486-1496, 2019. 6" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 642, + 545, + 674 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 642, + 545, + 674 + ], + "spans": [ + { + "bbox": [ + 307, + 642, + 545, + 674 + ], + "type": "text", + "content": "[52] Henrik Stewénius, Magnus Oskarsson, Kalle Aström, and David Nister. Solutions to minimal generalized relative pose problems, 2005. 3, 4" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 674, + 547, + 707 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 674, + 547, + 707 + ], + "spans": [ + { + "bbox": [ + 307, + 674, + 547, + 707 + ], + "type": "text", + "content": "[53] Yurun Tian, Xin Yu, Bin Fan, Fuchao Wu, Huub Heijnen, and Vassileios Balntas. SOSNet: Second order similarity regularization for local descriptor learning. In Proceedings of" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "9782" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 362 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 66, + 72, + 287, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 72, + 287, + 95 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 287, + 95 + ], + "type": "text", + "content": "the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11016-11025, 2019. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 95, + 287, + 137 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 95, + 287, + 137 + ], + "spans": [ + { + "bbox": [ + 48, + 95, + 287, + 137 + ], + "type": "text", + "content": "[54] Philip HS Torr and Andrew Zisserman. Mlesac: A new robust estimator with application to estimating image geometry. Computer vision and image understanding, 78(1):138-156, 2000. 4" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 137, + 287, + 191 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 137, + 287, + 191 + ], + "spans": [ + { + "bbox": [ + 48, + 137, + 287, + 191 + ], + "type": "text", + "content": "[55] Julien Valentin, Angela Dai, Matthias Niessner, Pushmeet Kohli, Philip Torr, Shahram Izadi, and Cem Keskin. Learning to navigate the energy landscape. In 2016 Fourth International Conference on 3D Vision (3DV), pages 323-332, 2016. 1, 3, 5, 6, 7, 8" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 191, + 287, + 234 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 191, + 287, + 234 + ], + "spans": [ + { + "bbox": [ + 48, + 191, + 287, + 234 + ], + "type": "text", + "content": "[56] Jonathan Ventura, Clemens Arth, and Vincent Lepetit. An efficient minimal solution for multi-camera motion. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), pages 747-755, 2015. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 234, + 287, + 277 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 234, + 287, + 277 + ], + "spans": [ + { + "bbox": [ + 48, + 234, + 287, + 277 + ], + "type": "text", + "content": "[57] Carl Vondrick, Aditya Khosla, Tomasz Malisiewicz, and Antonio Torralba. HOGgles: Visualizing object detection features. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), pages 1-8, 2013. 1, 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 277, + 287, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 277, + 287, + 319 + ], + "spans": [ + { + "bbox": [ + 48, + 277, + 287, + 319 + ], + "type": "text", + "content": "[58] Philippe Weinzaepfel, Herve Jégou, and Patrick Pérez. Reconstructing an image from its local descriptors. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 337-344, 2011. 1, 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 319, + 287, + 362 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 319, + 287, + 362 + ], + "spans": [ + { + "bbox": [ + 48, + 319, + 287, + 362 + ], + "type": "text", + "content": "[59] Enliang Zheng and Changchang Wu. Structure from motion using structure-less resection. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), pages 2075-2083, 2015. 2, 4" + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "9783" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Efficient Scene Recovery Using Luminous Flux Prior/0327d2f4-6c3a-4e74-a5a8-041807da336a_content_list.json b/2024/Efficient Scene Recovery Using Luminous Flux Prior/0327d2f4-6c3a-4e74-a5a8-041807da336a_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..bcb2d434439fa16a86aab859abcbeee49123a323 --- /dev/null +++ b/2024/Efficient Scene Recovery Using Luminous Flux Prior/0327d2f4-6c3a-4e74-a5a8-041807da336a_content_list.json @@ -0,0 +1,1590 @@ +[ + { + "type": "text", + "text": "Efficient Scene Recovery Using Luminous Flux Prior", + "text_level": 1, + "bbox": [ + 220, + 130, + 750, + 151 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Zhongyu Li1, Lei Zhang1* \n1University of Science and Technology of China, Hefei, China \nruin2@mail.ustc.edu.cn, leizh23@ustc.edu.cn,", + "bbox": [ + 238, + 179, + 730, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 268, + 313, + 284 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Scene recovery, the restoration of images degraded by adverse weather conditions, presents significant challenges for existing methods. Physical models, constrained by their inherent assumptions, often fail when these assumptions are not met; Deep learning models are powerful, they are limited by the diversity of their training datasets, leading to poor generalization and high computational demands. To address these limitations, we propose the Luminous Flux Prior (LFP), to recover degraded images under diverse adverse weather without learning. Luminous flux, a physical measure that reflects image brightness, has a rate of change that demonstrates a significant correlation with transmission. Consequently, we leverage this rate of change in luminous flux as prior knowledge to estimate transmission, which in turn assists in image recovery. This approach reduces dependency on physical parameters and enhances adaptability to various weather. Experimental validation under diverse conditions, such as sandstorms, underwater environments, and haze, attests to the robustness of LFP in restoring clear images. With a time complexity of $\\mathcal{O}(N\\log N)$ , LFP enables real-time recovery, making it a suitable for devices with limited computational resources.", + "bbox": [ + 75, + 301, + 472, + 633 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 651, + 209, + 667 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Observational devices encounter significant challenges under adverse lighting conditions, which can be caused by phenomena such as sandstorms, underwater environments, and haze. These conditions lead to image degradation, typified by a loss of detail, contrast distortion, and color bias [13, 50]. Such degradation impairs computer vision tasks, including object detection [31], image segmentation [42], and scene understanding [43]. Therefore, restoring clear images from these degraded observations becomes a critical challenge in the field of computer vision [27].", + "bbox": [ + 75, + 676, + 468, + 828 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Image degradation primarily arises from the scattering and absorption of light [10, 11, 16, 26, 49]. Environmental elements, such as the atmosphere and water, scatter", + "bbox": [ + 76, + 829, + 468, + 875 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "light sources, engendering intricate environmental light and the absorption of parts of the visible light spectrum. This process significantly degrades images by diminishing their brightness. The physical quantity symbolizing the brightness per unit area, referred to as luminous flux, is denoted by $F$ . The fundamental instigator of this degradation is a decrease in the transmission rate $t$ , representing the proportion of light brightness received by the camera.", + "bbox": [ + 496, + 268, + 893, + 390 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The challenge in image restoration is the estimation of transmission $t$ . We bifurcate scene restoration methods into two principal types based on the calculation of $t$ : physical models and data-driven deep learning methods. Physical models [1, 2, 16, 19, 20, 38, 54] make assumptions about the physical parameters or conditions for transmission $t$ to be valid. However, if these assumptions fall short, the restored image suffers to varying extents. Conversely, data-driven deep learning methods extrapolate an estimation model for haze-free images from extensive datasets [3, 23, 40]. Yet, these supervised learning methods exhibit limited efficacy in real-world scenarios due to the scarcity of large-scale, diverse, and fully aligned paired training data. Additionally, deep learning algorithms necessitate complex loss functions and training strategies, resulting in high training costs [48]. Most existing methods rely solely on L1/L2 image reconstruction loss and lack robust regularization, which may induce unsuitable gradient updates and grapple with the uncertainty of ill-posed problems due to a dearth of robust prior information. Both physical models and deep learning methods find it challenging to adapt to a variety of different scenes owing to their inherent limitations [27].", + "bbox": [ + 496, + 391, + 893, + 724 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Typical physical models illustrate the image formation process impacted by scattering [44]. This model can generally be represented by the following equation [36, 50]:", + "bbox": [ + 496, + 724, + 893, + 768 + ], + "page_idx": 0 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {I} (x) = \\mathbf {J} (x) \\cdot t (x) + \\mathbf {A} (1 - t (x)), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 576, + 785, + 890, + 801 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "where $x$ denotes the location information within the image, $\\mathbf{I}(x)$ refers to the observed degraded image, $\\mathbf{J}(x)$ signifies the clear image (scene radiance) we aim to restore, and $\\mathbf{A}$ symbolizes the global ambient light. $t(x)$ describes the proportion of light reaching the camera. To restore a clear image $\\mathbf{J}(x)$ , we need to estimate the transmission $t(x)$ and", + "bbox": [ + 496, + 809, + 893, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 44 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Lei Zhang is the corresponding author.", + "bbox": [ + 94, + 886, + 308, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "2743", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/b44ffb7f4ee619adbbb03f00640e0649b13199f0da7b46db6cfc780785e1448d.jpg", + "image_caption": [ + "(a)" + ], + "image_footnote": [], + "bbox": [ + 84, + 88, + 197, + 255 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/aec446a9c0bf0e24348e389fb95ab6c5673e3e1742f7a9f13e1f205c83c75e0e.jpg", + "image_caption": [ + "(b)" + ], + "image_footnote": [], + "bbox": [ + 199, + 88, + 312, + 256 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/7651c8eca67c9c20a28419ad349cdba7054d9d53d6e7061f734c537dc06b6b19.jpg", + "image_caption": [ + "(c)" + ], + "image_footnote": [], + "bbox": [ + 313, + 88, + 419, + 255 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/810751d257b6c0fd4ad051ee129986f7996ebe0364a978aff362a615bcf425af.jpg", + "image_caption": [ + "(d)", + "Figure 1. Challenging real scene recovery using LFP. The first row depicts degraded images, while the second row presents the restored clear images. These images represent three adverse conditions: haze, sandstorm, and underwater environments. The restored images all exhibit superior visual effects. Notably, the color bias in images (b), (c), and (e) has been rectified; images (c) and (g) are free from halo artifacts; and the brightness in images (a), (b), (c), (d), and (f) has been supplemented and adjusted." + ], + "image_footnote": [], + "bbox": [ + 419, + 88, + 653, + 256 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/79cdd6655033a93ca105738d6b8091ff4b7d72814c8d27a51784f1ee04e05694.jpg", + "image_caption": [ + "(f)" + ], + "image_footnote": [], + "bbox": [ + 653, + 89, + 763, + 255 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/5a2787ff6b6525e0d82e2735e6c0dca5c50fb7e4e3f29c888b02e717dc4735f3.jpg", + "image_caption": [ + "(g)" + ], + "image_footnote": [], + "bbox": [ + 763, + 89, + 880, + 255 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "ambient light $\\mathbf{A}$ . $t(x)$ can be defined as:", + "bbox": [ + 76, + 338, + 344, + 353 + ], + "page_idx": 1 + }, + { + "type": "equation", + "text": "\n$$\nt (x) = e ^ {- \\beta d (x)}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 217, + 359, + 468, + 376 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "where $\\beta$ is the attenuation coefficient, indicating the degree of scattering in the environmental medium, and $d(x)$ designates the scene depth [54].", + "bbox": [ + 75, + 383, + 468, + 429 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The complexity of estimating transmission lies in its dependence on multiple physical parameters [54]. Adverse weather conditions that diminish luminous flux $F$ (detailed in Sec.3) result in image darkening, attributed to decreased $t$ . Utilizing $F$ directly to estimate $t$ is impractical. We propose LFP that employs the rate of change of luminous flux, $\\Delta F$ (detailed in Sec.3), as a prior to simplify the process. In extensive experimental validation, we found $\\Delta F$ correlates significantly with $t$ , thereby transforming $t$ estimation into a $\\Delta F$ computation. As a dimensionless quantity, $\\Delta F$ exhibits wide applicability, underscoring the adaptability of LFP. LFP is computationally efficient and generates visually pleasing images, as demonstrated by real scenes in Fig 1.", + "bbox": [ + 75, + 430, + 468, + 625 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In summary, our contributions are as follows:", + "bbox": [ + 96, + 626, + 395, + 640 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. We introduce a new method to estimate transmission $(t)$ , leveraging the rate of change in luminous flux $(\\Delta F)$ . This approach is widely applicable across different weather and imaging conditions, and does not require any learnable parameters, thus eliminating the need for dataset construction.", + "2. Our algorithm has a low computational complexity, $O(N \\cdot \\log N)$ , where $N$ is the size of a single image. With GPU acceleration, it is exceedingly fast, enhancing computation speeds by 7 to 50 times compared to CPU processing. This makes the algorithm suitable for devices with limited computational power.", + "3. We have transformed the estimation of $t$ into a computation involving $F$ and $\\Delta F$ , eliminates the influence of physical parameters on $t$ . Experimental validation confirms that this method delivers robust performance and can enhance images to a high visual quality." + ], + "bbox": [ + 76, + 641, + 468, + 898 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 500, + 335, + 648, + 351 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Physical Models: Physical models employ image statistical attributes to design priors. For instance, Dark Channel Prior (DCP) [16] utilizes the minimum value in the RGB channels of pixels. Building upon DCP, BCCR [34] introduces a more efficient method for estimating transmission coefficients. Other methods, such as Fast Visibility Restoration (FVR) [51], make assumptions about the atmospheric veil, considering it locally flat, and use median filters instead of minimum filters for transmission estimation. Color Attenuation Prior (CAP) [54] formulates a linear model between depth information and the attenuation coefficient, while ROP [27] proposes an intensity projection strategy for transmission estimation. These prior-based methods perform admirably in specific scene types, but their adaptability to diverse scenes is limited. For example, DCP tends to darken already dark areas (i.e. skies and shadows) when dehazing. To rectify this, LDCP [55] modifies sky regions. Brightness is a crucial factor for accurate transmission estimation [55], enhancing contrast and improving visual quality. Many image enhancement techniques leverage this attribute [12, 45, 52]. These algorithms typically convert RGB images to other color spaces, such as YUV and LAB, to perform color correction [45, 52], for example, HRDCP [45] and CVC [18]. In dealing with sandstorms and underwater environments, color correction methods like white balance, akin to the Gray World Hypothesis [33], are important. The hypothesis posits that the average values of the red, green, and blue channels in a scene should be approximately equal. Considering that sandstorm weather absorbs blue and green light, and underwater environments absorb red light, using color balance as prior knowledge can result in more realistic recovery effects [33].", + "bbox": [ + 496, + 366, + 890, + 851 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Data-Driven Methods: Learning-based methods extrapolate models for haze-free image estimation from extensive datasets. Initial models like Waternet [25], MSCNN [40],", + "bbox": [ + 498, + 854, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2744", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "and AOD-Net [23], equipped with shallow architectures, outperformed previous prior-based methods. Some even incorporated prior knowledge, such as UWCNN [26] and DehazeNet [3]. Lately, most learning-based methods design more intricate and larger parameterized network architectures to enhance the network's expressivity. These include the attention mechanism [5, 7, 39], multi-scale features [6, 15, 29, 30], and even dehazing networks based on Vision Transformer (ViT) [8, 14, 53], such as gUNet [47]. Despite these advancements, the efficacy of these supervised learning methods is limited by their generalization ability in real-world scenes. The scarcity of large-scale, diverse, and fully aligned paired training data poses a significant hurdle. These methods often rely on synthetic data, but procuring synthetic data for sand and underwater environments presents challenges [28, 46]. Deep learning-based algorithms also encounter difficulties in training, requiring complex loss functions and training strategies, which lead to high training costs [48]. Many existing methods only use image reconstruction loss based on L1/L2, lacking robust regularization. This deficiency can result in unsuitable gradient updates and difficulties in grappling with the uncertainty of ill-conditioned problems, underscoring the need for robust prior information.", + "bbox": [ + 76, + 90, + 472, + 454 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 76, + 469, + 171, + 486 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we first validate the correlation between $\\Delta F$ and $t$ from a statistical perspective (Sec.3.1). Next, we derive the analytical form of transmittance using luminous flux and its rate of change (Sec.3.2). Finally, we present the numerical calculation methods for luminous flux and its rate of change, the image recovery formula and analysis of algorithmic complexity (Sec.3.3).", + "bbox": [ + 75, + 496, + 470, + 602 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Luminous Flux Prior Validation", + "text_level": 1, + "bbox": [ + 76, + 614, + 362, + 630 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We validate the correlation between the rate of change in luminous flux, denoted as $\\Delta F$ , and transmission $t$ , utilizing datasets from three disparate weather conditions. These conditions encompass images from haze, sandstorms, and underwater environments, both simulated and real. The correlation between $\\Delta F$ and $t$ remains robust across all conditions, which underscores the utility of $\\Delta F$ as a prior for image enhancement. The datasets we use are as follows:", + "bbox": [ + 75, + 638, + 468, + 758 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. Haze: We utilize the RESIDE dataset [24], which contains 13,990 synthetic indoor and 72,135 synthetic outdoor images.", + "2. Sandstorm: We rely on the Sand-dust Image Reconstruction Benchmark (SIRB) dataset [46], which simulates sandstorm conditions with different dust intensities.", + "3. Underwater: We employ the Underwater Image Enhancement Benchmark Dataset (UIEBD) [26] which includes 890 real underwater environment images together" + ], + "bbox": [ + 76, + 763, + 468, + 900 + ], + "page_idx": 2 + }, + { + "type": "table", + "img_path": "images/2f97fdb62d113b5518f0da3d1d82c2ca4c87986196c3e9263f537c4a6b547c50.jpg", + "table_caption": [ + "Table 1. We apply the KLD to measure the difference between the rate of change of luminance and the transmission map. Here, $N$ symbolizes a truncated standard normal distribution, and $U$ represents a uniform distribution within the range [0, 1]." + ], + "table_footnote": [], + "table_body": "
RESIDESIRBO-HazeUIEBD
ITSOTSLightMediumDense
ΔF0.290.210.130.130.160.120.15
N0.870.640.750.991.660.530.77
U0.950.531.271.681.900.430.64
", + "bbox": [ + 501, + 155, + 888, + 220 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "with reference images free of water distortion, obtained via their Deep Underwater Image Enhancement Network.", + "bbox": [ + 519, + 253, + 890, + 282 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In total, these datasets encompass 90,960 images, comprising our entire sample space with each image representing an individual sample. For an image of size $m \\times n$ , denoted as I, each pixel corresponds to three values: $t$ , $F$ , and $\\Delta F$ . T, F, and $\\Delta \\mathbf{F}$ represent these values in matrix form. Both $\\Delta \\mathbf{F}$ and $\\mathbf{T}$ are of the same size as image I. $\\Delta \\mathbf{F} \\in \\mathbb{R}^{m \\times n \\times c}$ , where $c$ denotes the number of channels in image I. For instance, when I exists in the RGB color space, $c = 3$ ; when I is a grayscale image, $c = 1$ .", + "bbox": [ + 496, + 292, + 890, + 429 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We employ the Kullback-Leibler Divergence (KLD) to authenticate the correlation between $\\Delta \\mathbf{F}$ and $\\mathbf{T}$ , measuring their difference as follows:", + "bbox": [ + 498, + 431, + 890, + 476 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nK L D (\\Delta \\mathbf {F} | | \\mathbf {T}) \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 643, + 496, + 890, + 512 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We generate the transmission map $\\mathbf{T}$ of an image by subtracting a low-quality image from a clear counterpart and taking the absolute value. The resulting image represents the degree of light scattering, with brighter regions indicating severe degradation and darker regions suggesting minor or no degradation. We temporarily disregard the effect of atmospheric light. Subsequently, we compute $\\Delta \\mathbf{F}$ for the degraded image and then verify the correlation between $\\Delta \\mathbf{F}$ and $\\mathbf{T}$ using Eq 3.", + "bbox": [ + 496, + 534, + 890, + 671 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "As presented in Tab 1, the KLD values for $\\mathbf{T}$ when using uniform or Gaussian distributions range between 0.53 to 1.66 and 0.43 to 1.90 respectively. It is important to note that the use of uniform or normal distributions does not facilitate image recovery, thus their KL values can be perceived as the upper bounds. As the KLD values inch closer to these upper bounds, image recovery becomes increasingly challenging. In contrast, the KLD values for $\\Delta F$ fall within the significantly lower range of 0.12 to 0.29, approximately 4 to 10 times less than the aforementioned upper bounds. This signifying relatively low information loss when estimating $\\mathbf{T}$ with $\\Delta \\mathbf{F}$ . Therefore, $\\Delta \\mathbf{F}$ effectively encapsulates the characteristics of $\\mathbf{T}$ , indicating a pronounced correlation between $\\Delta \\mathbf{F}$ and $\\mathbf{T}$ . These findings substantiate the effectiveness of the LFP.", + "bbox": [ + 496, + 674, + 890, + 898 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "2745", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/64e2b6d36d125d3a553138e24ad4df4bc0a5962db75ac08d89189efd59788888.jpg", + "image_caption": [ + "Figure 2. Visualization of the LFP Algorithm Workflow: First, it calculates Luminous Flux $F$ and it's ratio of change $\\Delta F$ within a sliding window across the image. Second, it estimates transmission $t$ using $F$ and $\\Delta F$ . Finally, recover the image using Eq 15." + ], + "image_footnote": [], + "bbox": [ + 83, + 85, + 869, + 270 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Scene Recovery with LFP", + "text_level": 1, + "bbox": [ + 76, + 334, + 310, + 349 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2.1 Analytical Form of Luminous Flux", + "text_level": 1, + "bbox": [ + 76, + 358, + 372, + 372 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The Bouguer-Lambert law [35] offers a comprehensive analytical expression for the measurement of luminous flux. This law utilizes the attenuation coefficient $\\beta$ [37] and the scene depth $d(x)$ [21] as primary factors to accurately evaluate luminous flux Eq 4. In this section, to ensure clarity and simplicity, we have chosen to omit the position information denoted by $x$ from the symbols used.", + "bbox": [ + 75, + 381, + 468, + 487 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nF = \\tilde {F} e ^ {- \\beta d} \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 230, + 500, + 468, + 517 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\tilde{F}$ represents the luminous flux of incident light.", + "bbox": [ + 76, + 523, + 437, + 539 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2.2 Estimating Transmission", + "text_level": 1, + "bbox": [ + 76, + 556, + 303, + 571 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Transmission, a property of the material, is the ratio of $F$ to $\\tilde{F}$ , as depicted in Eq 5:", + "bbox": [ + 76, + 580, + 468, + 611 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nt = F / \\tilde {F}. \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 236, + 619, + 468, + 638 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In line with the Retinex [22] model, we perceive an image as composed of incident and reflected components. Incident light illuminates the object, and the object's reflected light forms the image we observe. The incident light $\\tilde{F}$ can be eliminated using division, as shown in Eq 6:", + "bbox": [ + 76, + 648, + 468, + 723 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nt _ {i j} = \\frac {t _ {i}}{t _ {j}} = \\frac {F _ {i} / \\tilde {F} _ {i}}{F _ {j} / \\tilde {F} _ {j}} = F _ {i} / F _ {j}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 173, + 732, + 468, + 770 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $i$ and $j$ represent two adjacent pixels $x_{i}$ and $x_{j}$ , $\\tilde{F}_{i} = \\tilde{F} j$ , $t_{ij}$ is a dimensionless parameter depicting the luminous flux $F$ ratio between adjacent pixels. This allows analysis of flux changes within the image, crucial for revealing local features and their variations.", + "bbox": [ + 76, + 779, + 468, + 854 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Based on Eq 6, $t_{ij}$ can be further expressed as:", + "bbox": [ + 96, + 857, + 408, + 873 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nt _ {i j} = F _ {i} / F _ {j} = e ^ {- d _ {i} \\cdot \\left(\\beta_ {i} - \\beta_ {j}\\right)}. \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 174, + 883, + 468, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In Eq 7, to comprehend the influence of luminous flux on $\\beta$ , we introduce $\\mathrm{d}\\beta = \\frac{\\beta_i - \\beta_j}{F_i - F_j}$ , where $\\mathrm{d}\\beta$ is a differential with respect to $\\beta$ . By substituting $\\mathrm{d}\\beta$ into Eq 7, we obtain:", + "bbox": [ + 498, + 335, + 890, + 383 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} t _ {i j} = e ^ {- d _ {i} \\left(\\beta_ {i} - \\beta_ {j}\\right)} \\\\ = e ^ {- d _ {i} \\left(\\beta_ {i} - \\beta_ {j}\\right) \\frac {\\Delta F}{\\Delta F}} \\tag {8} \\\\ = e ^ {- d _ {i} \\cdot \\Delta F \\cdot \\mathrm {d} \\beta}. \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 622, + 388, + 890, + 449 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Simultaneously integrating both sides of Eq 8, obtain:", + "bbox": [ + 519, + 459, + 877, + 474 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\int t \\mathrm {d} t = \\int e ^ {- d \\cdot \\Delta F \\cdot \\mathrm {d} \\beta} \\tag {9} \\\\ t = \\sqrt {2 e ^ {- d \\cdot \\beta \\cdot \\Delta F}}. \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 617, + 481, + 890, + 531 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The proof is provided in the appendix.", + "bbox": [ + 500, + 541, + 753, + 555 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For ease of discussion, we consider $\\tilde{\\mathbf{F}}$ as a constant, denoted as $k$ , where $k \\in (0,1]$ . From Eq 4, $\\beta$ can be expressed as in Eq 10:", + "bbox": [ + 496, + 555, + 890, + 599 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\beta = - \\frac {k}{d} \\log F. \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 642, + 607, + 890, + 637 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "By substituting Eq 10 into Eq 9, we obtain the expression for $t$ as follows:", + "bbox": [ + 498, + 642, + 888, + 670 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} t = \\sqrt {2 e ^ {- d \\cdot \\beta \\cdot \\Delta F}} \\\\ = \\sqrt {2 e ^ {\\log F \\cdot k \\cdot \\Delta F}} \\tag {11} \\\\ = \\sqrt {2 F ^ {k \\Delta F}}. \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 630, + 676, + 890, + 738 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In Eq 11, we have successfully eliminated the two physical parameters $d(x)$ and $\\beta$ and derived an analytical form for $t$ . Considering that the exponent part is relatively small compared to the base, Eq 11 can be further simplified, resulting in Eq 12:", + "bbox": [ + 496, + 747, + 890, + 823 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nt \\approx 1 + k \\cdot \\log F \\cdot \\Delta F. \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 614, + 830, + 890, + 847 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In Eq 12, we consolidate all constants from Eq 11 into $k$ , where $k \\in (0,1]$ , as the maximum value of luminous flux is normalized to 1. The proof is provided in the appendix.", + "bbox": [ + 496, + 854, + 890, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "2746", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Proposed algorithm", + "text_level": 1, + "bbox": [ + 76, + 90, + 266, + 107 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The numerical method for computing luminous flux, denoted as $F$ , leverages its physical definition: the ratio of light intensity over a given area to the area itself. We implement a sliding window approach, traversing the entire image. For each window, we calculate the mean intensity and subsequently divide by the window's area to compute the numerical value of the luminous flux for each pixel. In this context, $\\Omega$ represents the sliding window, $\\Omega(\\cdot)$ is the mean value of the samplings within the window, and $S_{\\Omega}$ is the area of the window. Denoted as Eq 13:", + "bbox": [ + 75, + 113, + 472, + 265 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nF = \\sum \\Omega (\\mathbf {I}) / S _ {\\Omega}. \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 207, + 276, + 468, + 297 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For numerical computation, we sample the neighborhoods of each pixel by using sliding window. Consequently, for each pixel, the $\\Delta F$ can be expressed as shown in Eq 14:", + "bbox": [ + 76, + 308, + 468, + 354 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\Delta F = \\frac {\\min \\Omega (F)}{\\max \\Omega (F)} \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 209, + 364, + 468, + 397 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We use the extreme ratio within the sliding window to amplify $\\Delta F$ because the difference between adjacent pixels is close to 0. However, $\\Delta F$ might be affected by noise. To mitigate this, we first truncate $\\Omega(F)$ within the range [0.005, 0.95]. Following this, it may be necessary to truncate the transmission $t$ during subsequent processing to ensure stability and accuracy in our results.", + "bbox": [ + 76, + 402, + 468, + 507 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In recovery process, we define a clear image $\\mathbf{J}$ using an image formulation model from Eq 1. This is expressed as follows in Eq 15:", + "bbox": [ + 76, + 507, + 468, + 554 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {J} (x) = \\frac {\\mathbf {I} (x) + \\mathbf {A} \\omega (t (x) - 1)}{\\operatorname* {m a x} (\\omega t (x) , t _ {0})}. \\tag {15}\n$$\n", + "text_format": "latex", + "bbox": [ + 166, + 563, + 468, + 597 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In Eq 15, $\\omega$ is a constant relaxation parameter which is constrained to be in the range $(0,1]$ , and $t_0$ , set as 0.05, is a lower bound that provides stability during computation.", + "bbox": [ + 76, + 606, + 468, + 651 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The global atmospheric light, represented by $\\mathbf{A}$ , plays a crucial role in our model. We estimate $\\mathbf{A}$ swiftly using a", + "bbox": [ + 76, + 652, + 468, + 681 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Algorithm 1 Pseudo-code for LFP", + "bbox": [ + 501, + 90, + 732, + 104 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1: Setup: constant $k = 0.5$ and window size $\\Omega = 7$", + "2: Input: Degraded image $I$", + "3: while $\\Omega$ in $I$ do", + "4: Compute mean intensity: $I_{\\Omega} = \\sum \\Omega (I) / S_{\\Omega}$", + "5: Compute luminous flux: $F = I_{\\Omega} / S_{\\Omega}$", + "6: Compute rate of change: $\\Delta F = \\frac{\\min\\Omega(F)}{\\max\\Omega(F)}$", + "7: end while", + "8: $t\\approx 1 + k\\cdot \\log F\\cdot \\Delta F$ Compute the transmission", + "9: guided filtering on transmission", + "10: Output: $J = \\frac{I + A \\cdot (t - 1)}{t}$" + ], + "bbox": [ + 504, + 109, + 890, + 309 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "two-step approach. Initially, $\\mathbf{A}$ is estimated using mean filtering and the maximum grayscale value of the original image. This gives us an initial approximation of the global atmospheric light. Subsequently, the image is segmented into small regions, and we select the maximum value from each region to represent the $\\mathbf{A}$ for that particular region. The final estimate of $\\mathbf{A}$ is then calculated as the average of all the $\\mathbf{A}$ values from these small regions. This method allows us to get a more refined and accurate estimate of the global atmospheric light. The complete process of our scene recovery method is thoroughly depicted in Fig 2, and the corresponding pseudo-code is provided in Algorithm 1. In terms of computational complexity, our algorithm starts off with a complexity of $\\mathcal{O}(N)$ as we traverse all pixels in the image using a sliding window. Then, it rises by $\\mathcal{O}(\\log N)$ when we compute the maximum and minimum values for each window. Finally, the total computational complexity for the subsequent steps, which include the calculation of transmission according to Eq 11 and application of a guided filter [17] for smoothing, is $\\mathcal{O}(N\\log N)$ , where $N$ denotes the size of a single image. This complexity analysis demonstrates that our method is efficient and feasible for large scale image recovery tasks.", + "bbox": [ + 496, + 338, + 893, + 686 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/acafea452d1b56b835d5b38a6164235a83588187bba4da843608a48bfabccd7a.jpg", + "image_caption": [ + "Figure 3. Visual comparisons of different dehazing methods on outdoor and indoor images. The algorithms designed for water removal (blue) and sand removal (yellow-brown) fail in haze (grey) environments. Our method effectively restores details in the sky region without introducing artifacts. It also improves the visual quality of indoor images by enhancing brightness while preserving the structure." + ], + "image_footnote": [], + "bbox": [ + 107, + 702, + 867, + 847 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "2747", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 76, + 89, + 209, + 107 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Implementation Details", + "text_level": 1, + "bbox": [ + 76, + 114, + 294, + 131 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Our method relies solely on two parameters: $k$ and the sliding window $\\Omega$ . And uses a non-parametric approach for the estimation of transmission $t$ , enhancing visual results under a variety of visibility conditions. Based on empirical experiences, we have determined the optimal settings to be $k = 0.5$ and $\\Omega = 7$ . All experiments were conducted on a machine with an AMD Ryzen 73750H CPU@2.3GHz and a GeForce RTX 2080 GPU (11GB). The deep learning models were built with the PyTorch, TensorFlow and Matlab frameworks. The test images are publicly available.", + "bbox": [ + 75, + 138, + 470, + 289 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Experimental Settings", + "text_level": 1, + "bbox": [ + 76, + 299, + 284, + 316 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We evaluated the LFP's performance using various methods and datasets, including physical models and deep learning models. Selected datasets spanned varied environments and scattering degrees, including RESIDE [24], SIRB [46], UIEBD [26], Haze4K [32], and RS-Haze [24]. These encompassed synthetic to real-world scenes, ensuring broad evaluation robustness. We used Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity Index Measure (SSIM) metrics for objective and fair evaluations across scenarios and methods.", + "bbox": [ + 75, + 323, + 468, + 473 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Qualitative Comparison", + "text_level": 1, + "bbox": [ + 76, + 484, + 299, + 500 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Image Dehazing. We use RESIDE dataset [24] to validate the efficacy and robustness of LFP, as is shown in Fig 3, by comparing it with $\\mathrm{Dive + }$ and HRDCP. While $\\mathrm{Dive + }$ fails to restore image clarity, HRDCP results in color deviation, likely due to the specific absorption and scattering properties of sand particles. In contrast, our LFP method can generate more natural-looking results with better structures in a more stable manner. This demonstrates that images in hazy and underwater environments are formed based on different optical properties. We then compare the performance of the LFP with different methods like MSCNN, gUNet, and DCP. Learning-based methods such as gUNet and MSCNN are susceptible to dataset bias, which makes it difficult to", + "bbox": [ + 75, + 508, + 470, + 704 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "enhance visual quality. Meanwhile, traditional methods like DCP, which rely on hand-crafted priors, struggle with varied real-world conditions. LFP adeptly overcomes the major obstacle faced by previous dehazing approaches, which is the problematic influence of scene depth. The inherent uncertainty in depth estimation often leads to unstable performance. By confidently deriving the relationship between $\\Delta F$ and $t$ without involving $d$ , LFP successfully avoids the dataset bias that learning-based methods encounter. LFP effectively bypasses complex parameters of transmission and offers reliable solution with robust theoretical guarantees. Our method demonstrates a significantly stronger capability for robust image restoration in hazy weather conditions, resulting in highly satisfactory visual results.", + "bbox": [ + 496, + 90, + 890, + 303 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Underwater Image Enhancement. We validate the efficacy and robustness of LFP for underwater image enhancement through comprehensive experiments, using UIEBD dataset [26]. Results are shown in Fig 4. We first compare it with DCP and HRDCP, which are methodologies designed for different weather conditions. DCP fails completely, and although HRDCP achieves some visual effects, it suffers from noticeable color distortion. These experimental results underscore the necessity for image processing techniques to be specifically designed and optimized for each environment, reinforcing that an algorithm effective in one environment can't simply be applied to another without adaptation. $\\mathrm{Dive + }$ and UWCNN struggle to effectively eliminate the green light prevalent in underwater scenes, leading to undesired color shifts. Although Waternet and ROP perform well in color restoration, they both suffer from dimmed lighting and reduced contrast, thereby compromising the overall visual quality. LFP overcomes the major obstacle faced by previous underwater image enhancement approaches, which is the varying influence of the attenuation coefficient $\\beta$ on the RGB color channels that often results in color bias. By eliminating the attenuation coefficient $\\beta$ during the derivation process, LFP avoids the color bias, rendering it a reliable solution with robust theoretical guarantees. Therefore, LFP method demonstrates competitive color correction and contrast enhancement capabilities.", + "bbox": [ + 496, + 311, + 890, + 705 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/749b5781291848e2186314ec964d16b18c9a9420a7ba8c74a490ece395cc85bf.jpg", + "image_caption": [ + "Figure 4. Comparison of different methods for underwater conditions. The algorithms designed for haze removal (grey) and sand removal (brown) fail in underwater (blue) environments. Our method corrects color shifts, enhance image brightness and improve visibility." + ], + "image_footnote": [], + "bbox": [ + 86, + 724, + 888, + 859 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "2748", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/22b3c6d596c818426b2167537e6aa81c6ff38a2fa3a2cf8520fa08b995c7edcb.jpg", + "image_caption": [ + "Figure 5. Comparison of different methods for sandstorm image enhancement. The algorithms designed for haze removal (grey) and water removal (blue) fail in sandstorm (brown) environments. Our method corrects color shifts, enhances image brightness, and improves contrast, resulting in more vibrant color reproduction. Moreover, LFP produces a more realistic sky area closer to real scenes." + ], + "image_footnote": [], + "bbox": [ + 84, + 87, + 885, + 208 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Visibility Restoration in Sandstorms. We assess our LFP method's effectiveness in restoring visibility during sandstorms, by using SIRB dataset [46]. Results are shown in Fig 5. Initially, we compare LFP's performance with DCP, MSCNN, and $\\mathrm{Dive + }$ . The first two fail in dust and sandstorm mitigation, and $\\mathrm{Dive + }$ introduces noticeable color deviations in sky areas. This emphasizes that hazy and sandstorm images are influenced by distinct optical properties, making dehaze or dewater algorithms unsuitable for sandstorms. Next, we contrast our method with CVC, HRDCP, and ROP. All three significantly enhance visibility in sandstorms. However, CVC lacks contrast, leading to detail loss. HRDCP, despite its visual improvements, exhibits a dim sky and color distortions due to its DCP-based design. ROP performs well but struggles with color distortion, resulting in a cool appearance. LFP overcomes prior challenges by addressing scene depth and attenuation coefficient $\\beta$ . By eliminating hard-to-estimate parameters like atmospheric light and scene depth, LFP avoids common issues, making it robust in sandstorm conditions without relying on assumptions. Thus, our LFP method excels in sandstorm image restoration, producing visually pleasing results without color distortion.", + "bbox": [ + 75, + 268, + 472, + 601 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In summary, the LFP has the ability to restore scenes in a variety of severe weather conditions, a characteristic that is absent in most algorithms. By eliminating the physical parameters that impact transmittance, LFP effectively bypasses the influence of scene depth, thus enhancing visual quality. Moreover, it minimizes color deviations by removing the attenuation coefficient. This approach provides a theoretically and empirically supported a priori method for scene restoration algorithms.", + "bbox": [ + 75, + 602, + 468, + 739 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Quantitative Comparison", + "text_level": 1, + "bbox": [ + 76, + 753, + 310, + 771 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Dehazing Quality We selected DCP, BCCR [34], and FVR [51] as our benchmarks to represent different types of prior-based dehazing methodologies. The rationale behind this selection, instead of opting for deep learning techniques, is multi-fold. Traditional dehazing methods not only possess the capability to restore clear images, but they also notably improve visual effects, thereby enhancing the visual experience. Furthermore, LFP does not contain any learnable pa", + "bbox": [ + 75, + 779, + 472, + 902 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/8852362e7cd509a395e219750f6ccabd1342db20dd28464559bd5e118cfdc7cc.jpg", + "table_caption": [ + "Table 2. This table succinctly presents the comparative results of the LFP method and three other distinct prior methods. The evaluation spans across all datasets and metrics, with LFP consistently achieving superior performance. The highest PSNR and SSIM values, represented by bold text, clearly demonstrate the effectiveness of our LFP method." + ], + "table_footnote": [], + "table_body": "
MethodsRESIDE-INRESIDE-OUTHaze-4KRS-Haze
PSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIM
DCP16.620.81819.130.81514.010.76017.860.734
BCCR16.880.79114.020.77713.780.75618.840.747
FVR15.720.74811.610.66812.650.68816.120.626
LFP19.760.86921.050.86118.290.87318.240.786
", + "bbox": [ + 503, + 354, + 890, + 438 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "rameters, differentiating it from deep learning approaches. As illustrated in Tab 2, demonstrates that LFP outperforms other methods across all four of our test datasets.", + "bbox": [ + 496, + 464, + 890, + 508 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The performance is consistent whether we evaluate in terms of PSNR or SSIM. This robust evidence strongly validates the high generalization ability of the LFP method. Furthermore, the LFP method excels in restoring clear, detailed imagery, as particularly demonstrated by the significant improvement in SSIM scores. For more detailed results on desanding and dewatering, please refer to the appendix.", + "bbox": [ + 496, + 510, + 890, + 617 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Runtime Efficiency We conducted extensive comparisons on both CPU and GPU platforms. On the CPU, we compared LFP with traditional dehazing methods; on the GPU, we chose a variety of deep learning dehazing models, including: MSCNN, AOD-Net [23], GFN [41], GCANet [4], GDNet [30], MSBDN [6], FFA [39], CycleD [9], DehazeF [47], and gUNet-T. In terms of runtime, Tab 3 compares the efficiency on both CPU and GPU. The results show that the LFP method consistently achieves the fastest running speed, surpassing existing algorithms.", + "bbox": [ + 496, + 627, + 890, + 779 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "CPU Performance. As evidenced by the data presented in Tab 3, the LFP model consistently exhibits superior runtime performance across all tested resolutions. With a runtime of merely 0.007 seconds at the lower-end resolution of $360\\mathrm{p}$ , and a still impressive 0.863 seconds at the high-end 4k resolution, these empirical findings underscore the remarkable efficiency of the LFP model when processing large-scale images or performing real-time scene recovery.", + "bbox": [ + 496, + 789, + 890, + 912 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "2749", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/4ed6b6466f7f54abc991f6acf1a650fe7f082c49285f27b8bf92cffabbc5564f.jpg", + "table_caption": [ + "Table 3. Run-time(seconds) performance. Testing the efficiency of different methods on various single image sizes. Models marked with (*) indicate that the results for those models were generated using MATLAB. The best performance is denoted in red, while the second best performance is indicated in blue." + ], + "table_footnote": [], + "table_body": "
CPU Performance
Model360p480p720p1080g2k4k
DCP[16]0.4590.7651.762.6006.7415.11
Retinex[22]0.1490.2860.4780.8042.3675.421
LDCP[55]0.5170.8212.2053.0918.26418.09
CAP[54]0.2180.3100.6661.0212.8666.403
BCCR[34]1.6913.5716.41611.7029.6064.26
FVR[51]0.3013.9028.58912.5834.3576.50
ROP*[27]0.0670.1440.4111.3992.0732.515
LFP0.0070.0170.0730.1510.4070.863
GPU Performance
MSCNN[40]0.1150.1470.2660.4621.1142.481
DehazeNet[3]0.1650.2230.4050.6211.3742.951
AOD-Net[23]0.0760.1060.1750.2910.6961.519
GFN*[41]0.6600.8581.4942.4456.35313.09
GCANet[4]0.4460.6411.1071.8214.2629.266
GDNet[30]0.6440.7421.5032.0274.67510.48
MSBDN[6]0.9521.2122.1943.3358.20317.46
FFA[39]1.6752.2533.9605.97714.2730.61
CycleD[9]0.6070.8101.4282.1585.15511.03
DehazeF[47]0.2890.3870.6771.0322.4755.310
gUNet-T[47]0.1970.2660.4730.7161.7193.669
ROP*[27]0.0470.0510.1130.2090.5320.894
LFP0.0010.0010.0020.0030.0080.019
", + "bbox": [ + 78, + 170, + 480, + 537 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "GPU Performance. Tab 3 shows that LFPsignificantly outperforms all the compared algorithms in processing speed, especially for high-resolution images. For instance, at a resolution of 4K, LFP's average runtime is merely 0.019 seconds, whereas the runtimes of other algorithms range from 0.894 seconds to a staggering 30.609 seconds. This results in a speedup of up to 34 times for LFP, marking its superior computational efficiency.", + "bbox": [ + 75, + 563, + 468, + 686 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.5. Ablation Study on $k$ and $\\Omega$", + "text_level": 1, + "bbox": [ + 76, + 695, + 316, + 710 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Our ablation study further reveals that the changes in the $k$ parameter not only influence the luminous flux rate of change $\\Delta F$ but also directly affect the transmission estimation. A properly tuned $k$ facilitates an accurate estimation of the transmission map, thus effectively reducing the impact of different weather conditions on the image. On the other hand, the size of the sliding window $\\Omega$ directly affects the local variance $\\Delta F$ . A larger $\\Omega$ captures more significant changes in the luminous flux, which is particularly effective in dense haze conditions or when the scene has a greater depth of view. This broadened perspective allows for a more comprehensive understanding of the scene, lead", + "bbox": [ + 75, + 719, + 470, + 901 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/96f687e52cf63cb58d7a015b6f8d6e83b5e1864b8c257177420c6948036ecc05.jpg", + "image_caption": [ + "Figure 6. The x-axis represents $k$ , the y-axis represents $\\Omega$ . When the depth of view is further, or the haze is denser, $\\Omega$ and $k$ need to be increased correspondingly." + ], + "image_footnote": [], + "bbox": [ + 504, + 87, + 893, + 377 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "ing to a higher-quality image after enhancement. Finally, the interplay between $k$ and $\\Omega$ is also crucial. Balancing these two factors allows for a fine-tuned approach to image enhancement, taking into account both the overall brightness and the local variance of the image, this ability to cope with complex conditions adds another layer of robustness to our technique. This balance ensures that the final output image is visually pleasing while maintaining the fidelity of the original scene.", + "bbox": [ + 496, + 445, + 893, + 582 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 594, + 619, + 609 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this work, we propose an efficient scene recovery algorithm to restore clear images under poor visibility conditions. By analyzing the image formation model, we derive an analytical form for transmission that eliminates multiple influencing factors and leverages the correlation with flux change rate as prior knowledge. Our approach achieves state-of-the-art performance while maintaining low computational complexity, enabling real-time processing on resource-limited devices. Extensive experiments validate the effectiveness of the proposed method. The novel integration of prior knowledge and analytical transmission modeling provides opportunities to inform future learning-based methods. Overall, this paper makes significant contributions towards robust scene recovery under challenging imaging conditions.", + "bbox": [ + 496, + 619, + 893, + 832 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Acknowledgments", + "text_level": 1, + "bbox": [ + 500, + 844, + 678, + 861 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This work was supported by the National Natural Science Foundation of China (No. 62336001, No. 62072014)", + "bbox": [ + 500, + 869, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "2750", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 171, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Dana Berman, Shai Avidan, et al. Non-local image dehazing. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1674-1682, 2016. 1", + "[2] Trung Minh Bui and Wonha Kim. Single image dehazing using color ellipsoid prior. IEEE Transactions on Image Processing, 27(2):999-1009, 2017. 1", + "[3] Bolun Cai, Xiangmin Xu, Kui Jia, Chunmei Qing, and Dacheng Tao. Dehazenet: An end-to-end system for single image haze removal. IEEE Transactions on Image Processing, 25(11):5187-5198, 2016. 1, 3, 8", + "[4] Dongdong Chen, Mingming He, Qingnan Fan, Jing Liao, Liheng Zhang, Dongdong Hou, Lu Yuan, and Gang Hua. Gated context aggregation network for image dehazing and deraining. In 2019 IEEE winter conference on applications of computer vision (WACV), pages 1375-1383. IEEE, 2019. 7, 8", + "[5] Qili Deng, Ziling Huang, Chung-Chi Tsai, and Chia-Wen Lin. Hardgan: A haze-aware representation distillation gan for single image dehazing. In European Conference on Computer Vision, pages 722-738. Springer, 2020. 3", + "[6] Hang Dong, Jinshan Pan, Lei Xiang, Zhe Hu, Xinyi Zhang, Fei Wang, and Ming-Hsuan Yang. Multi-scale boosted dehazing network with dense feature fusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2157-2167, 2020. 3, 7, 8", + "[7] Jiangxin Dong and Jinshan Pan. Physics-based feature dehazing networks. In European Conference on Computer Vision, pages 188-204. Springer, 2020. 3", + "[8] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 3", + "[9] Deniz Engin, Anil Genç, and Hazim Kemal Ekenel. Cycle-dehaze: Enhanced cyclegan for single image dehazing. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition Workshops, pages 825-833, 2018. 7, 8", + "[10] Faming Fang, Fang Li, and Tieyong Zeng. Single image dehazing and denoising: a fast variational approach. SIAM Journal on Imaging Sciences, 7(2):969-996, 2014. 1", + "[11] Xueyang Fu, Yue Huang, Delu Zeng, Xiao-Ping Zhang, and Xinghao Ding. A fusion-based enhancing approach for single sandstorm image. In 2014 IEEE 16th international workshop on multimedia signal processing (MMSP), pages 1-5. IEEE, 2014. 1", + "[12] Han Gao, Ping Wei, and Jun Ke. Color enhancement and image defogging in hsi based on retinax model. In 2015 International Conference on Optical Instruments and Technology: Optoelectronic Imaging and Processing Technology, pages 9-16. SPIE, 2015. 2", + "[13] Jie Gui, Xiaofeng Cong, Yuan Cao, Wenqi Ren, Jun Zhang, Jing Zhang, Jieuxin Cao, and Dacheng Tao. A comprehensive survey and taxonomy on single image dehazing based on deep learning. ACM Computing Surveys, 55(13s):1-37, 2023. 1" + ], + "bbox": [ + 78, + 114, + 467, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[14] Chun-Le Guo, Qixin Yan, Saeed Anwar, Runmin Cong, Wenqi Ren, and Chongyi Li. Image dehazing transformer with transmission-aware 3d position embedding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5812-5820, 2022. 3", + "[15] Yecai Guo, Hanyu Li, and Peixian Zhuang. Underwater image enhancement using a multiscale dense generative adversarial network. IEEE Journal of Oceanic Engineering, 45(3):862-870, 2019. 3", + "[16] Kaiming He, Jian Sun, and Xiaou Tang. Single image haze removal using dark channel prior. IEEE transactions on pattern analysis and machine intelligence, 33(12):2341-2353, 2010. 1, 2, 8", + "[17] Kaiming He, Jian Sun, and Xiaou Tang. Guided image filtering. IEEE transactions on pattern analysis and machine intelligence, 35(6):1397-1409, 2012. 5", + "[18] Jong-Ju Jeon, Tae-Hee Park, and Il-Kyu Eom. Sand-dust image enhancement using chromatic variance consistency and gamma correction-based dehazing. Sensors, 22(23):9048, 2022. 2", + "[19] Mingye Ju, Can Ding, Y Jay Guo, and Dengyin Zhang. Idgcp: Image dehazing based on gamma correction prior. IEEE Transactions on Image Processing, 29:3104-3118, 2019. 1", + "[20] Manjit Kaur, Dilbag Singh, Vijay Kumar, and Kehui Sun. Color image dehazing using gradient channel prior and guided 10 filter. Information Sciences, 521:326-342, 2020. 1", + "[21] Johannes Kopf, Boris Neubert, Billy Chen, Michael Cohen, Daniel Cohen-Or, Oliver Deussen, Matt Uytendaele, and Dani Lischinski. Deep photo: Model-based photograph enhancement and viewing. ACM transactions on graphics (TOG), 27(5):1-10, 2008. 4", + "[22] Edwin H Land. The retina theory of color vision. Scientific american, 237(6):108-129, 1977. 4, 8", + "[23] Boyi Li, Xiulian Peng, Zhangyang Wang, Jizheng Xu, and Dan Feng. Aod-net: All-in-one dehazing network. In Proceedings of the IEEE international conference on computer vision, pages 4770-4778, 2017. 1, 3, 7, 8", + "[24] Boyi Li, Wenqi Ren, Dengpan Fu, Dacheng Tao, Dan Feng, Wenjun Zeng, and Zhangyang Wang. Benchmarking single-image dehazing and beyond. IEEE Transactions on Image Processing, 28(1):492-505, 2019. 3, 6", + "[25] Chongyi Li, Chunle Guo, Wenqi Ren, Runmin Cong, Junhui Hou, Sam Kwong, and Dacheng Tao. An underwater image enhancement benchmark dataset and beyond. IEEE Transactions on Image Processing, 29:4376-4389, 2019. 2", + "[26] Chongyi Li, Saeed Anwar, and Fatih Porikli. Underwater scene prior inspired deep underwater image and video enhancement. Pattern Recognition, 98:107038, 2020. 1, 3, 6", + "[27] Jun Liu, Wen Liu, Jianing Sun, and Tieyong Zeng. Rank-one prior: Toward real-time scene recovery. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14802-14810, 2021. 1, 2, 8", + "[28] Risheng Liu, Xin Fan, Ming Zhu, Minjun Hou, and Zhongxuan Luo. Real-world underwater enhancement: Challenges, benchmarks, and solutions under natural light. IEEE transactions on circuits and systems for video technology, 30(12): 4861-4875, 2020. 3" + ], + "bbox": [ + 501, + 92, + 890, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "2751", + "bbox": [ + 482, + 945, + 513, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[29] Xiaodong Liu, Zhi Gao, and Ben M Chen. Mlfcgan: Multilevel feature fusion-based conditional gan for underwater image color correction. IEEE Geoscience and Remote Sensing Letters, 17(9):1488-1492, 2019. 3", + "[30] Xiaohong Liu, Yongrui Ma, Zhihao Shi, and Jun Chen. Griddehazenet: Attention-based multi-scale network for image dehazing. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7314-7323, 2019. 3, 7, 8", + "[31] Yu Liu, Guanlong Zhao, Boyuan Gong, Yang Li, Ritu Raj, Niraj Goel, Satya Kesav, Sandeep Gottimukkala, Zhangyang Wang, Wenqi Ren, et al. Improved techniques for learning to dehaze and beyond: A collective study. arXiv preprint arXiv:1807.00202, 2018. 1", + "[32] Ye Liu, Lei Zhu, Shunda Pei, Huazhu Fu, Jing Qin, Qing Zhang, Liang Wan, and Wei Feng. From synthetic to real: Image dehazing collaborating with unlabeled real data. In Proceedings of the 29th ACM International Conference on Multimedia, pages 50-58, 2021. 6", + "[33] Rastislav Lukac. Single-sensor imaging: methods and applications for digital cameras. CRC Press, 2018. 2", + "[34] Gaofeng Meng, Ying Wang, Jiangyong Duan, Shiming Xiang, and Chunhong Pan. Efficient image dehazing with boundary constraint and contextual regularization. In Proceedings of the IEEE international conference on computer vision, pages 617-624, 2013. 2, 7, 8", + "[35] WE Knowles Middleton. Bouguer, lambert, and the theory of horizontal visibility. Isis, 51(2):145-149, 1960. 4", + "[36] Srinivasa G Narasimhan and Shree K Nayar. Vision and the atmosphere. International journal of computer vision, 48(3): 233-254, 2002. 1", + "[37] Srinivasa G Narasimhan and Shree K Nayar. Interactive (de) weathering of an image using physical models. In IEEE Workshop on color and photometric Methods in computer Vision, page 1. France, 2003. 4", + "[38] Ko Nishino, Louis Kratz, and Stephen Lombardi. Bayesian defogging. International journal of computer vision, 98(3): 263-278, 2012. 1", + "[39] Xu Qin, Zhilin Wang, Yuanchao Bai, Xiaodong Xie, and Huizhu Jia. Ffa-net: Feature fusion attention network for single image dehazing. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 11908-11915, 2020. 3, 7, 8", + "[40] Wenqi Ren, Si Liu, Hua Zhang, Jinshan Pan, Xiaochun Cao, and Ming-Hsuan Yang. Single image dehazing via multiscale convolutional neural networks. In European conference on computer vision, pages 154-169. Springer, 2016. 1, 2, 8", + "[41] Wenqi Ren, Lin Ma, Jiawei Zhang, Jinshan Pan, Xiaochun Cao, Wei Liu, and Ming-Hsuan Yang. Gated fusion network for single image dehazing. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 3253-3261, 2018. 7, 8", + "[42] Wenqi Ren, Jingang Zhang, Xiangyu Xu, Lin Ma, Xiaochun Cao, Gaofeng Meng, and Wei Liu. Deep video dehazing with semantic segmentation. IEEE Transactions on Image Processing, 28(4):1895-1908, 2018. 1", + "[43] Christos Sakaridis, Dengxin Dai, Simon Hecker, and Luc Van Gool. Model adaptation with synthetic and real data for" + ], + "bbox": [ + 78, + 90, + 470, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "semantic dense foggy scene understanding. In Proceedings of the European Conference on Computer Vision (ECCV), pages 687-704, 2018. 1", + "[44] Yoav Y Schechner and Yuval Averbuch. Regularized image recovery in scattering media. IEEE Transactions on Pattern Analysis and Machine Intelligence, 29(9):1655-1660, 2007. 1", + "[45] Zhenghao Shi, Yaning Feng, Minghua Zhao, Erhu Zhang, and Lifeng He. Let you see in sand dust weather: A method based on halo-reduced dark channel prior dehazing for sand-dust image enhancement. *Ieee Access*, 7:116722–116733, 2019. 2", + "[46] Yazhong Si, Fan Yang, Ya Guo, Wei Zhang, and Yipu Yang. A comprehensive benchmark analysis for sand dust image reconstruction. Journal of Visual Communication and Image Representation, 89:103638, 2022. 3, 6, 7", + "[47] Yuda Song, Zhuqing He, Hui Qian, and Xin Du. Vision transformers for single image dehazing. arXiv preprint arXiv:2204.03883, 2022. 3, 7, 8", + "[48] Yuda Song, Yang Zhou, Hui Qian, and Xin Du. Rethinking performance gains in image dehazing networks. arXiv preprint arXiv:2209.11448, 2022. 1, 3", + "[49] Matan Sulami, Itamar Glatzer, Raanan Fattal, and Mike Werman. Automatic recovery of the atmospheric light in hazy images. In 2014 IEEE International Conference on Computational Photography (ICCP), pages 1-11. IEEE, 2014. 1", + "[50] Robby T Tan. Visibility in bad weather from a single image. In 2008 IEEE Conference on Computer Vision and Pattern Recognition, pages 1-8. IEEE, 2008. 1", + "[51] Jean-Philippe Tarel and Nicolas Hautiere. Fast visibility restoration from a single color or gray level image. In 2009 IEEE 12th International Conference on Computer Vision, pages 2201–2208. IEEE, 2009. 2, 7, 8", + "[52] Ting Yan, Liejun Wang, and Jiaxing Wang. Method to enhance degraded image in dust environment. J. Softw., 9(10): 2672-2677, 2014. 2", + "[53] Dong Zhao, Jia Li, Hongyu Li, and Long Xu. Hybrid local-global transformer for image dehazing. arXiv preprint arXiv:2109.07100, 2021. 3", + "[54] Qingsong Zhu, Jiaming Mai, and Ling Shao. A fast single image haze removal algorithm using color attenuation prior. IEEE transactions on image processing, 24(11):3522-3533, 2015. 1, 2, 8", + "[55] Yingying Zhu, Gaoyang Tang, Xiaoyan Zhang, Jianmin Jiang, and Qi Tian. Haze removal method for natural restoration of images with sky. neurocomputing, 275:499-510, 2018. 2, 8" + ], + "bbox": [ + 503, + 92, + 890, + 756 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "2752", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2024/Efficient Scene Recovery Using Luminous Flux Prior/0327d2f4-6c3a-4e74-a5a8-041807da336a_model.json b/2024/Efficient Scene Recovery Using Luminous Flux Prior/0327d2f4-6c3a-4e74-a5a8-041807da336a_model.json new file mode 100644 index 0000000000000000000000000000000000000000..82c7fd9e47c1e8135472f12ac9760f71a7eda72b --- /dev/null +++ b/2024/Efficient Scene Recovery Using Luminous Flux Prior/0327d2f4-6c3a-4e74-a5a8-041807da336a_model.json @@ -0,0 +1,2387 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.0, + 0.812, + 0.045 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.221, + 0.131, + 0.75, + 0.152 + ], + "angle": 0, + "content": "Efficient Scene Recovery Using Luminous Flux Prior" + }, + { + "type": "text", + "bbox": [ + 0.24, + 0.18, + 0.731, + 0.234 + ], + "angle": 0, + "content": "Zhongyu Li1, Lei Zhang1* \n1University of Science and Technology of China, Hefei, China \nruin2@mail.ustc.edu.cn, leizh23@ustc.edu.cn," + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.269, + 0.314, + 0.285 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.302, + 0.473, + 0.635 + ], + "angle": 0, + "content": "Scene recovery, the restoration of images degraded by adverse weather conditions, presents significant challenges for existing methods. Physical models, constrained by their inherent assumptions, often fail when these assumptions are not met; Deep learning models are powerful, they are limited by the diversity of their training datasets, leading to poor generalization and high computational demands. To address these limitations, we propose the Luminous Flux Prior (LFP), to recover degraded images under diverse adverse weather without learning. Luminous flux, a physical measure that reflects image brightness, has a rate of change that demonstrates a significant correlation with transmission. Consequently, we leverage this rate of change in luminous flux as prior knowledge to estimate transmission, which in turn assists in image recovery. This approach reduces dependency on physical parameters and enhances adaptability to various weather. Experimental validation under diverse conditions, such as sandstorms, underwater environments, and haze, attests to the robustness of LFP in restoring clear images. With a time complexity of \\(\\mathcal{O}(N\\log N)\\), LFP enables real-time recovery, making it a suitable for devices with limited computational resources." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.652, + 0.21, + 0.669 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.678, + 0.47, + 0.829 + ], + "angle": 0, + "content": "Observational devices encounter significant challenges under adverse lighting conditions, which can be caused by phenomena such as sandstorms, underwater environments, and haze. These conditions lead to image degradation, typified by a loss of detail, contrast distortion, and color bias [13, 50]. Such degradation impairs computer vision tasks, including object detection [31], image segmentation [42], and scene understanding [43]. Therefore, restoring clear images from these degraded observations becomes a critical challenge in the field of computer vision [27]." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.83, + 0.47, + 0.876 + ], + "angle": 0, + "content": "Image degradation primarily arises from the scattering and absorption of light [10, 11, 16, 26, 49]. Environmental elements, such as the atmosphere and water, scatter" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.27, + 0.895, + 0.391 + ], + "angle": 0, + "content": "light sources, engendering intricate environmental light and the absorption of parts of the visible light spectrum. This process significantly degrades images by diminishing their brightness. The physical quantity symbolizing the brightness per unit area, referred to as luminous flux, is denoted by \\( F \\). The fundamental instigator of this degradation is a decrease in the transmission rate \\( t \\), representing the proportion of light brightness received by the camera." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.392, + 0.895, + 0.725 + ], + "angle": 0, + "content": "The challenge in image restoration is the estimation of transmission \\( t \\). We bifurcate scene restoration methods into two principal types based on the calculation of \\( t \\): physical models and data-driven deep learning methods. Physical models [1, 2, 16, 19, 20, 38, 54] make assumptions about the physical parameters or conditions for transmission \\( t \\) to be valid. However, if these assumptions fall short, the restored image suffers to varying extents. Conversely, data-driven deep learning methods extrapolate an estimation model for haze-free images from extensive datasets [3, 23, 40]. Yet, these supervised learning methods exhibit limited efficacy in real-world scenarios due to the scarcity of large-scale, diverse, and fully aligned paired training data. Additionally, deep learning algorithms necessitate complex loss functions and training strategies, resulting in high training costs [48]. Most existing methods rely solely on L1/L2 image reconstruction loss and lack robust regularization, which may induce unsuitable gradient updates and grapple with the uncertainty of ill-posed problems due to a dearth of robust prior information. Both physical models and deep learning methods find it challenging to adapt to a variety of different scenes owing to their inherent limitations [27]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.725, + 0.895, + 0.77 + ], + "angle": 0, + "content": "Typical physical models illustrate the image formation process impacted by scattering [44]. This model can generally be represented by the following equation [36, 50]:" + }, + { + "type": "equation", + "bbox": [ + 0.577, + 0.786, + 0.891, + 0.803 + ], + "angle": 0, + "content": "\\[\n\\mathbf {I} (x) = \\mathbf {J} (x) \\cdot t (x) + \\mathbf {A} (1 - t (x)), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.81, + 0.895, + 0.902 + ], + "angle": 0, + "content": "where \\( x \\) denotes the location information within the image, \\( \\mathbf{I}(x) \\) refers to the observed degraded image, \\( \\mathbf{J}(x) \\) signifies the clear image (scene radiance) we aim to restore, and \\( \\mathbf{A} \\) symbolizes the global ambient light. \\( t(x) \\) describes the proportion of light reaching the camera. To restore a clear image \\( \\mathbf{J}(x) \\), we need to estimate the transmission \\( t(x) \\) and" + }, + { + "type": "page_footnote", + "bbox": [ + 0.095, + 0.887, + 0.309, + 0.9 + ], + "angle": 0, + "content": "*Lei Zhang is the corresponding author." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2743" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.086, + 0.089, + 0.199, + 0.256 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.134, + 0.257, + 0.151, + 0.268 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.2, + 0.089, + 0.313, + 0.257 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.246, + 0.258, + 0.264, + 0.268 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image", + "bbox": [ + 0.314, + 0.089, + 0.421, + 0.256 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.361, + 0.257, + 0.376, + 0.268 + ], + "angle": 0, + "content": "(c)" + }, + { + "type": "image", + "bbox": [ + 0.421, + 0.089, + 0.655, + 0.257 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.474, + 0.258, + 0.491, + 0.268 + ], + "angle": 0, + "content": "(d)" + }, + { + "type": "image", + "bbox": [ + 0.655, + 0.09, + 0.764, + 0.256 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.701, + 0.258, + 0.716, + 0.268 + ], + "angle": 0, + "content": "(f)" + }, + { + "type": "image", + "bbox": [ + 0.764, + 0.09, + 0.881, + 0.256 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.813, + 0.258, + 0.83, + 0.268 + ], + "angle": 0, + "content": "(g)" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.27, + 0.894, + 0.325 + ], + "angle": 0, + "content": "Figure 1. Challenging real scene recovery using LFP. The first row depicts degraded images, while the second row presents the restored clear images. These images represent three adverse conditions: haze, sandstorm, and underwater environments. The restored images all exhibit superior visual effects. Notably, the color bias in images (b), (c), and (e) has been rectified; images (c) and (g) are free from halo artifacts; and the brightness in images (a), (b), (c), (d), and (f) has been supplemented and adjusted." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.339, + 0.345, + 0.354 + ], + "angle": 0, + "content": "ambient light \\(\\mathbf{A}\\). \\(t(x)\\) can be defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.218, + 0.36, + 0.469, + 0.377 + ], + "angle": 0, + "content": "\\[\nt (x) = e ^ {- \\beta d (x)}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.385, + 0.469, + 0.43 + ], + "angle": 0, + "content": "where \\(\\beta\\) is the attenuation coefficient, indicating the degree of scattering in the environmental medium, and \\(d(x)\\) designates the scene depth [54]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.431, + 0.47, + 0.626 + ], + "angle": 0, + "content": "The complexity of estimating transmission lies in its dependence on multiple physical parameters [54]. Adverse weather conditions that diminish luminous flux \\( F \\) (detailed in Sec.3) result in image darkening, attributed to decreased \\( t \\). Utilizing \\( F \\) directly to estimate \\( t \\) is impractical. We propose LFP that employs the rate of change of luminous flux, \\( \\Delta F \\) (detailed in Sec.3), as a prior to simplify the process. In extensive experimental validation, we found \\( \\Delta F \\) correlates significantly with \\( t \\), thereby transforming \\( t \\) estimation into a \\( \\Delta F \\) computation. As a dimensionless quantity, \\( \\Delta F \\) exhibits wide applicability, underscoring the adaptability of LFP. LFP is computationally efficient and generates visually pleasing images, as demonstrated by real scenes in Fig 1." + }, + { + "type": "text", + "bbox": [ + 0.098, + 0.627, + 0.397, + 0.641 + ], + "angle": 0, + "content": "In summary, our contributions are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.642, + 0.47, + 0.731 + ], + "angle": 0, + "content": "1. We introduce a new method to estimate transmission \\((t)\\), leveraging the rate of change in luminous flux \\((\\Delta F)\\). This approach is widely applicable across different weather and imaging conditions, and does not require any learnable parameters, thus eliminating the need for dataset construction." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.733, + 0.469, + 0.822 + ], + "angle": 0, + "content": "2. Our algorithm has a low computational complexity, \\( O(N \\cdot \\log N) \\), where \\( N \\) is the size of a single image. With GPU acceleration, it is exceedingly fast, enhancing computation speeds by 7 to 50 times compared to CPU processing. This makes the algorithm suitable for devices with limited computational power." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.823, + 0.469, + 0.899 + ], + "angle": 0, + "content": "3. We have transformed the estimation of \\( t \\) into a computation involving \\( F \\) and \\( \\Delta F \\), eliminates the influence of physical parameters on \\( t \\). Experimental validation confirms that this method delivers robust performance and can enhance images to a high visual quality." + }, + { + "type": "list", + "bbox": [ + 0.078, + 0.642, + 0.47, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.337, + 0.649, + 0.352 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.367, + 0.892, + 0.852 + ], + "angle": 0, + "content": "Physical Models: Physical models employ image statistical attributes to design priors. For instance, Dark Channel Prior (DCP) [16] utilizes the minimum value in the RGB channels of pixels. Building upon DCP, BCCR [34] introduces a more efficient method for estimating transmission coefficients. Other methods, such as Fast Visibility Restoration (FVR) [51], make assumptions about the atmospheric veil, considering it locally flat, and use median filters instead of minimum filters for transmission estimation. Color Attenuation Prior (CAP) [54] formulates a linear model between depth information and the attenuation coefficient, while ROP [27] proposes an intensity projection strategy for transmission estimation. These prior-based methods perform admirably in specific scene types, but their adaptability to diverse scenes is limited. For example, DCP tends to darken already dark areas (i.e. skies and shadows) when dehazing. To rectify this, LDCP [55] modifies sky regions. Brightness is a crucial factor for accurate transmission estimation [55], enhancing contrast and improving visual quality. Many image enhancement techniques leverage this attribute [12, 45, 52]. These algorithms typically convert RGB images to other color spaces, such as YUV and LAB, to perform color correction [45, 52], for example, HRDCP [45] and CVC [18]. In dealing with sandstorms and underwater environments, color correction methods like white balance, akin to the Gray World Hypothesis [33], are important. The hypothesis posits that the average values of the red, green, and blue channels in a scene should be approximately equal. Considering that sandstorm weather absorbs blue and green light, and underwater environments absorb red light, using color balance as prior knowledge can result in more realistic recovery effects [33]." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.856, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Data-Driven Methods: Learning-based methods extrapolate models for haze-free image estimation from extensive datasets. Initial models like Waternet [25], MSCNN [40]," + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2744" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.473, + 0.455 + ], + "angle": 0, + "content": "and AOD-Net [23], equipped with shallow architectures, outperformed previous prior-based methods. Some even incorporated prior knowledge, such as UWCNN [26] and DehazeNet [3]. Lately, most learning-based methods design more intricate and larger parameterized network architectures to enhance the network's expressivity. These include the attention mechanism [5, 7, 39], multi-scale features [6, 15, 29, 30], and even dehazing networks based on Vision Transformer (ViT) [8, 14, 53], such as gUNet [47]. Despite these advancements, the efficacy of these supervised learning methods is limited by their generalization ability in real-world scenes. The scarcity of large-scale, diverse, and fully aligned paired training data poses a significant hurdle. These methods often rely on synthetic data, but procuring synthetic data for sand and underwater environments presents challenges [28, 46]. Deep learning-based algorithms also encounter difficulties in training, requiring complex loss functions and training strategies, which lead to high training costs [48]. Many existing methods only use image reconstruction loss based on L1/L2, lacking robust regularization. This deficiency can result in unsuitable gradient updates and difficulties in grappling with the uncertainty of ill-conditioned problems, underscoring the need for robust prior information." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.47, + 0.172, + 0.487 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.497, + 0.472, + 0.603 + ], + "angle": 0, + "content": "In this section, we first validate the correlation between \\(\\Delta F\\) and \\(t\\) from a statistical perspective (Sec.3.1). Next, we derive the analytical form of transmittance using luminous flux and its rate of change (Sec.3.2). Finally, we present the numerical calculation methods for luminous flux and its rate of change, the image recovery formula and analysis of algorithmic complexity (Sec.3.3)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.615, + 0.363, + 0.631 + ], + "angle": 0, + "content": "3.1. Luminous Flux Prior Validation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.64, + 0.47, + 0.76 + ], + "angle": 0, + "content": "We validate the correlation between the rate of change in luminous flux, denoted as \\(\\Delta F\\), and transmission \\(t\\), utilizing datasets from three disparate weather conditions. These conditions encompass images from haze, sandstorms, and underwater environments, both simulated and real. The correlation between \\(\\Delta F\\) and \\(t\\) remains robust across all conditions, which underscores the utility of \\(\\Delta F\\) as a prior for image enhancement. The datasets we use are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.765, + 0.469, + 0.81 + ], + "angle": 0, + "content": "1. Haze: We utilize the RESIDE dataset [24], which contains 13,990 synthetic indoor and 72,135 synthetic outdoor images." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.811, + 0.469, + 0.854 + ], + "angle": 0, + "content": "2. Sandstorm: We rely on the Sand-dust Image Reconstruction Benchmark (SIRB) dataset [46], which simulates sandstorm conditions with different dust intensities." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.856, + 0.469, + 0.901 + ], + "angle": 0, + "content": "3. Underwater: We employ the Underwater Image Enhancement Benchmark Dataset (UIEBD) [26] which includes 890 real underwater environment images together" + }, + { + "type": "list", + "bbox": [ + 0.078, + 0.765, + 0.469, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.089, + 0.892, + 0.146 + ], + "angle": 0, + "content": "Table 1. We apply the KLD to measure the difference between the rate of change of luminance and the transmission map. Here, \\( N \\) symbolizes a truncated standard normal distribution, and \\( U \\) represents a uniform distribution within the range [0, 1]." + }, + { + "type": "table", + "bbox": [ + 0.502, + 0.156, + 0.89, + 0.222 + ], + "angle": 0, + "content": "
RESIDESIRBO-HazeUIEBD
ITSOTSLightMediumDense
ΔF0.290.210.130.130.160.120.15
N0.870.640.750.991.660.530.77
U0.950.531.271.681.900.430.64
" + }, + { + "type": "text", + "bbox": [ + 0.52, + 0.254, + 0.892, + 0.284 + ], + "angle": 0, + "content": "with reference images free of water distortion, obtained via their Deep Underwater Image Enhancement Network." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.293, + 0.892, + 0.43 + ], + "angle": 0, + "content": "In total, these datasets encompass 90,960 images, comprising our entire sample space with each image representing an individual sample. For an image of size \\( m \\times n \\), denoted as I, each pixel corresponds to three values: \\( t \\), \\( F \\), and \\( \\Delta F \\). T, F, and \\( \\Delta \\mathbf{F} \\) represent these values in matrix form. Both \\( \\Delta \\mathbf{F} \\) and \\( \\mathbf{T} \\) are of the same size as image I. \\( \\Delta \\mathbf{F} \\in \\mathbb{R}^{m \\times n \\times c} \\), where \\( c \\) denotes the number of channels in image I. For instance, when I exists in the RGB color space, \\( c = 3 \\); when I is a grayscale image, \\( c = 1 \\)." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.433, + 0.892, + 0.477 + ], + "angle": 0, + "content": "We employ the Kullback-Leibler Divergence (KLD) to authenticate the correlation between \\(\\Delta \\mathbf{F}\\) and \\(\\mathbf{T}\\), measuring their difference as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.644, + 0.497, + 0.892, + 0.513 + ], + "angle": 0, + "content": "\\[\nK L D (\\Delta \\mathbf {F} | | \\mathbf {T}) \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.535, + 0.892, + 0.672 + ], + "angle": 0, + "content": "We generate the transmission map \\(\\mathbf{T}\\) of an image by subtracting a low-quality image from a clear counterpart and taking the absolute value. The resulting image represents the degree of light scattering, with brighter regions indicating severe degradation and darker regions suggesting minor or no degradation. We temporarily disregard the effect of atmospheric light. Subsequently, we compute \\(\\Delta \\mathbf{F}\\) for the degraded image and then verify the correlation between \\(\\Delta \\mathbf{F}\\) and \\(\\mathbf{T}\\) using Eq 3." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.675, + 0.892, + 0.9 + ], + "angle": 0, + "content": "As presented in Tab 1, the KLD values for \\(\\mathbf{T}\\) when using uniform or Gaussian distributions range between 0.53 to 1.66 and 0.43 to 1.90 respectively. It is important to note that the use of uniform or normal distributions does not facilitate image recovery, thus their KL values can be perceived as the upper bounds. As the KLD values inch closer to these upper bounds, image recovery becomes increasingly challenging. In contrast, the KLD values for \\(\\Delta F\\) fall within the significantly lower range of 0.12 to 0.29, approximately 4 to 10 times less than the aforementioned upper bounds. This signifying relatively low information loss when estimating \\(\\mathbf{T}\\) with \\(\\Delta \\mathbf{F}\\). Therefore, \\(\\Delta \\mathbf{F}\\) effectively encapsulates the characteristics of \\(\\mathbf{T}\\), indicating a pronounced correlation between \\(\\Delta \\mathbf{F}\\) and \\(\\mathbf{T}\\). These findings substantiate the effectiveness of the LFP." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2745" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.084, + 0.086, + 0.87, + 0.271 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.281, + 0.893, + 0.311 + ], + "angle": 0, + "content": "Figure 2. Visualization of the LFP Algorithm Workflow: First, it calculates Luminous Flux \\( F \\) and it's ratio of change \\( \\Delta F \\) within a sliding window across the image. Second, it estimates transmission \\( t \\) using \\( F \\) and \\( \\Delta F \\). Finally, recover the image using Eq 15." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.335, + 0.312, + 0.351 + ], + "angle": 0, + "content": "3.2. Scene Recovery with LFP" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.359, + 0.373, + 0.373 + ], + "angle": 0, + "content": "3.2.1 Analytical Form of Luminous Flux" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.382, + 0.47, + 0.488 + ], + "angle": 0, + "content": "The Bouguer-Lambert law [35] offers a comprehensive analytical expression for the measurement of luminous flux. This law utilizes the attenuation coefficient \\(\\beta\\) [37] and the scene depth \\(d(x)\\) [21] as primary factors to accurately evaluate luminous flux Eq 4. In this section, to ensure clarity and simplicity, we have chosen to omit the position information denoted by \\(x\\) from the symbols used." + }, + { + "type": "equation", + "bbox": [ + 0.231, + 0.501, + 0.469, + 0.518 + ], + "angle": 0, + "content": "\\[\nF = \\tilde {F} e ^ {- \\beta d} \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.524, + 0.438, + 0.54 + ], + "angle": 0, + "content": "where \\(\\tilde{F}\\) represents the luminous flux of incident light." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.558, + 0.305, + 0.573 + ], + "angle": 0, + "content": "3.2.2 Estimating Transmission" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.581, + 0.469, + 0.612 + ], + "angle": 0, + "content": "Transmission, a property of the material, is the ratio of \\( F \\) to \\( \\tilde{F} \\), as depicted in Eq 5:" + }, + { + "type": "equation", + "bbox": [ + 0.237, + 0.621, + 0.469, + 0.639 + ], + "angle": 0, + "content": "\\[\nt = F / \\tilde {F}. \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.649, + 0.469, + 0.724 + ], + "angle": 0, + "content": "In line with the Retinex [22] model, we perceive an image as composed of incident and reflected components. Incident light illuminates the object, and the object's reflected light forms the image we observe. The incident light \\(\\tilde{F}\\) can be eliminated using division, as shown in Eq 6:" + }, + { + "type": "equation", + "bbox": [ + 0.174, + 0.733, + 0.469, + 0.771 + ], + "angle": 0, + "content": "\\[\nt _ {i j} = \\frac {t _ {i}}{t _ {j}} = \\frac {F _ {i} / \\tilde {F} _ {i}}{F _ {j} / \\tilde {F} _ {j}} = F _ {i} / F _ {j}, \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.78, + 0.469, + 0.856 + ], + "angle": 0, + "content": "where \\(i\\) and \\(j\\) represent two adjacent pixels \\(x_{i}\\) and \\(x_{j}\\), \\(\\tilde{F}_{i} = \\tilde{F} j\\), \\(t_{ij}\\) is a dimensionless parameter depicting the luminous flux \\(F\\) ratio between adjacent pixels. This allows analysis of flux changes within the image, crucial for revealing local features and their variations." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.858, + 0.409, + 0.874 + ], + "angle": 0, + "content": "Based on Eq 6, \\( t_{ij} \\) can be further expressed as:" + }, + { + "type": "equation", + "bbox": [ + 0.176, + 0.884, + 0.469, + 0.903 + ], + "angle": 0, + "content": "\\[\nt _ {i j} = F _ {i} / F _ {j} = e ^ {- d _ {i} \\cdot \\left(\\beta_ {i} - \\beta_ {j}\\right)}. \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.336, + 0.892, + 0.385 + ], + "angle": 0, + "content": "In Eq 7, to comprehend the influence of luminous flux on \\(\\beta\\), we introduce \\(\\mathrm{d}\\beta = \\frac{\\beta_i - \\beta_j}{F_i - F_j}\\), where \\(\\mathrm{d}\\beta\\) is a differential with respect to \\(\\beta\\). By substituting \\(\\mathrm{d}\\beta\\) into Eq 7, we obtain:" + }, + { + "type": "equation", + "bbox": [ + 0.623, + 0.39, + 0.891, + 0.45 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} t _ {i j} = e ^ {- d _ {i} \\left(\\beta_ {i} - \\beta_ {j}\\right)} \\\\ = e ^ {- d _ {i} \\left(\\beta_ {i} - \\beta_ {j}\\right) \\frac {\\Delta F}{\\Delta F}} \\tag {8} \\\\ = e ^ {- d _ {i} \\cdot \\Delta F \\cdot \\mathrm {d} \\beta}. \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.52, + 0.46, + 0.878, + 0.475 + ], + "angle": 0, + "content": "Simultaneously integrating both sides of Eq 8, obtain:" + }, + { + "type": "equation", + "bbox": [ + 0.618, + 0.482, + 0.891, + 0.532 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\int t \\mathrm {d} t = \\int e ^ {- d \\cdot \\Delta F \\cdot \\mathrm {d} \\beta} \\tag {9} \\\\ t = \\sqrt {2 e ^ {- d \\cdot \\beta \\cdot \\Delta F}}. \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.542, + 0.754, + 0.556 + ], + "angle": 0, + "content": "The proof is provided in the appendix." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.556, + 0.891, + 0.601 + ], + "angle": 0, + "content": "For ease of discussion, we consider \\(\\tilde{\\mathbf{F}}\\) as a constant, denoted as \\(k\\), where \\(k \\in (0,1]\\). From Eq 4, \\(\\beta\\) can be expressed as in Eq 10:" + }, + { + "type": "equation", + "bbox": [ + 0.643, + 0.608, + 0.891, + 0.638 + ], + "angle": 0, + "content": "\\[\n\\beta = - \\frac {k}{d} \\log F. \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.643, + 0.89, + 0.671 + ], + "angle": 0, + "content": "By substituting Eq 10 into Eq 9, we obtain the expression for \\( t \\) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.631, + 0.677, + 0.891, + 0.739 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} t = \\sqrt {2 e ^ {- d \\cdot \\beta \\cdot \\Delta F}} \\\\ = \\sqrt {2 e ^ {\\log F \\cdot k \\cdot \\Delta F}} \\tag {11} \\\\ = \\sqrt {2 F ^ {k \\Delta F}}. \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.748, + 0.892, + 0.824 + ], + "angle": 0, + "content": "In Eq 11, we have successfully eliminated the two physical parameters \\(d(x)\\) and \\(\\beta\\) and derived an analytical form for \\(t\\). Considering that the exponent part is relatively small compared to the base, Eq 11 can be further simplified, resulting in Eq 12:" + }, + { + "type": "equation", + "bbox": [ + 0.616, + 0.832, + 0.891, + 0.848 + ], + "angle": 0, + "content": "\\[\nt \\approx 1 + k \\cdot \\log F \\cdot \\Delta F. \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.892, + 0.902 + ], + "angle": 0, + "content": "In Eq 12, we consolidate all constants from Eq 11 into \\( k \\), where \\( k \\in (0,1] \\), as the maximum value of luminous flux is normalized to 1. The proof is provided in the appendix." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2746" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.078, + 0.091, + 0.267, + 0.108 + ], + "angle": 0, + "content": "3.3. Proposed algorithm" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.114, + 0.473, + 0.266 + ], + "angle": 0, + "content": "The numerical method for computing luminous flux, denoted as \\( F \\), leverages its physical definition: the ratio of light intensity over a given area to the area itself. We implement a sliding window approach, traversing the entire image. For each window, we calculate the mean intensity and subsequently divide by the window's area to compute the numerical value of the luminous flux for each pixel. In this context, \\( \\Omega \\) represents the sliding window, \\( \\Omega(\\cdot) \\) is the mean value of the samplings within the window, and \\( S_{\\Omega} \\) is the area of the window. Denoted as Eq 13:" + }, + { + "type": "equation", + "bbox": [ + 0.208, + 0.277, + 0.47, + 0.299 + ], + "angle": 0, + "content": "\\[\nF = \\sum \\Omega (\\mathbf {I}) / S _ {\\Omega}. \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.309, + 0.47, + 0.356 + ], + "angle": 0, + "content": "For numerical computation, we sample the neighborhoods of each pixel by using sliding window. Consequently, for each pixel, the \\(\\Delta F\\) can be expressed as shown in Eq 14:" + }, + { + "type": "equation", + "bbox": [ + 0.21, + 0.365, + 0.47, + 0.398 + ], + "angle": 0, + "content": "\\[\n\\Delta F = \\frac {\\min \\Omega (F)}{\\max \\Omega (F)} \\tag {14}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.403, + 0.47, + 0.508 + ], + "angle": 0, + "content": "We use the extreme ratio within the sliding window to amplify \\(\\Delta F\\) because the difference between adjacent pixels is close to 0. However, \\(\\Delta F\\) might be affected by noise. To mitigate this, we first truncate \\(\\Omega(F)\\) within the range [0.005, 0.95]. Following this, it may be necessary to truncate the transmission \\(t\\) during subsequent processing to ensure stability and accuracy in our results." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.508, + 0.47, + 0.555 + ], + "angle": 0, + "content": "In recovery process, we define a clear image \\(\\mathbf{J}\\) using an image formulation model from Eq 1. This is expressed as follows in Eq 15:" + }, + { + "type": "equation", + "bbox": [ + 0.168, + 0.564, + 0.47, + 0.598 + ], + "angle": 0, + "content": "\\[\n\\mathbf {J} (x) = \\frac {\\mathbf {I} (x) + \\mathbf {A} \\omega (t (x) - 1)}{\\operatorname* {m a x} (\\omega t (x) , t _ {0})}. \\tag {15}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.607, + 0.47, + 0.652 + ], + "angle": 0, + "content": "In Eq 15, \\(\\omega\\) is a constant relaxation parameter which is constrained to be in the range \\((0,1]\\), and \\(t_0\\), set as 0.05, is a lower bound that provides stability during computation." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.653, + 0.47, + 0.683 + ], + "angle": 0, + "content": "The global atmospheric light, represented by \\(\\mathbf{A}\\), plays a crucial role in our model. We estimate \\(\\mathbf{A}\\) swiftly using a" + }, + { + "type": "code_caption", + "bbox": [ + 0.502, + 0.091, + 0.733, + 0.106 + ], + "angle": 0, + "content": "Algorithm 1 Pseudo-code for LFP" + }, + { + "type": "text", + "bbox": [ + 0.509, + 0.11, + 0.854, + 0.125 + ], + "angle": 0, + "content": "1: Setup: constant \\( k = 0.5 \\) and window size \\( \\Omega = 7 \\)" + }, + { + "type": "text", + "bbox": [ + 0.51, + 0.126, + 0.703, + 0.141 + ], + "angle": 0, + "content": "2: Input: Degraded image \\(I\\)" + }, + { + "type": "text", + "bbox": [ + 0.51, + 0.142, + 0.892, + 0.155 + ], + "angle": 0, + "content": "3: while \\(\\Omega\\) in \\(I\\) do" + }, + { + "type": "text", + "bbox": [ + 0.51, + 0.156, + 0.846, + 0.172 + ], + "angle": 0, + "content": "4: Compute mean intensity: \\( I_{\\Omega} = \\sum \\Omega (I) / S_{\\Omega} \\)" + }, + { + "type": "text", + "bbox": [ + 0.51, + 0.179, + 0.801, + 0.195 + ], + "angle": 0, + "content": "5: Compute luminous flux: \\( F = I_{\\Omega} / S_{\\Omega} \\)" + }, + { + "type": "text", + "bbox": [ + 0.51, + 0.198, + 0.83, + 0.221 + ], + "angle": 0, + "content": "6: Compute rate of change: \\(\\Delta F = \\frac{\\min\\Omega(F)}{\\max\\Omega(F)}\\)" + }, + { + "type": "text", + "bbox": [ + 0.51, + 0.224, + 0.6, + 0.238 + ], + "angle": 0, + "content": "7: end while" + }, + { + "type": "text", + "bbox": [ + 0.51, + 0.246, + 0.892, + 0.262 + ], + "angle": 0, + "content": "8: \\(t\\approx 1 + k\\cdot \\log F\\cdot \\Delta F\\) Compute the transmission" + }, + { + "type": "text", + "bbox": [ + 0.51, + 0.269, + 0.741, + 0.284 + ], + "angle": 0, + "content": "9: guided filtering on transmission" + }, + { + "type": "text", + "bbox": [ + 0.506, + 0.288, + 0.892, + 0.31 + ], + "angle": 0, + "content": "10: Output: \\( J = \\frac{I + A \\cdot (t - 1)}{t} \\)" + }, + { + "type": "list", + "bbox": [ + 0.506, + 0.11, + 0.892, + 0.31 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.339, + 0.895, + 0.687 + ], + "angle": 0, + "content": "two-step approach. Initially, \\(\\mathbf{A}\\) is estimated using mean filtering and the maximum grayscale value of the original image. This gives us an initial approximation of the global atmospheric light. Subsequently, the image is segmented into small regions, and we select the maximum value from each region to represent the \\(\\mathbf{A}\\) for that particular region. The final estimate of \\(\\mathbf{A}\\) is then calculated as the average of all the \\(\\mathbf{A}\\) values from these small regions. This method allows us to get a more refined and accurate estimate of the global atmospheric light. The complete process of our scene recovery method is thoroughly depicted in Fig 2, and the corresponding pseudo-code is provided in Algorithm 1. In terms of computational complexity, our algorithm starts off with a complexity of \\(\\mathcal{O}(N)\\) as we traverse all pixels in the image using a sliding window. Then, it rises by \\(\\mathcal{O}(\\log N)\\) when we compute the maximum and minimum values for each window. Finally, the total computational complexity for the subsequent steps, which include the calculation of transmission according to Eq 11 and application of a guided filter [17] for smoothing, is \\(\\mathcal{O}(N\\log N)\\), where \\(N\\) denotes the size of a single image. This complexity analysis demonstrates that our method is efficient and feasible for large scale image recovery tasks." + }, + { + "type": "image", + "bbox": [ + 0.109, + 0.703, + 0.868, + 0.848 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.855, + 0.894, + 0.898 + ], + "angle": 0, + "content": "Figure 3. Visual comparisons of different dehazing methods on outdoor and indoor images. The algorithms designed for water removal (blue) and sand removal (yellow-brown) fail in haze (grey) environments. Our method effectively restores details in the sky region without introducing artifacts. It also improves the visual quality of indoor images by enhancing brightness while preserving the structure." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2747" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.077, + 0.09, + 0.21, + 0.108 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.115, + 0.295, + 0.132 + ], + "angle": 0, + "content": "4.1. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.139, + 0.471, + 0.29 + ], + "angle": 0, + "content": "Our method relies solely on two parameters: \\( k \\) and the sliding window \\( \\Omega \\). And uses a non-parametric approach for the estimation of transmission \\( t \\), enhancing visual results under a variety of visibility conditions. Based on empirical experiences, we have determined the optimal settings to be \\( k = 0.5 \\) and \\( \\Omega = 7 \\). All experiments were conducted on a machine with an AMD Ryzen 73750H CPU@2.3GHz and a GeForce RTX 2080 GPU (11GB). The deep learning models were built with the PyTorch, TensorFlow and Matlab frameworks. The test images are publicly available." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.3, + 0.285, + 0.317 + ], + "angle": 0, + "content": "4.2. Experimental Settings" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.324, + 0.47, + 0.474 + ], + "angle": 0, + "content": "We evaluated the LFP's performance using various methods and datasets, including physical models and deep learning models. Selected datasets spanned varied environments and scattering degrees, including RESIDE [24], SIRB [46], UIEBD [26], Haze4K [32], and RS-Haze [24]. These encompassed synthetic to real-world scenes, ensuring broad evaluation robustness. We used Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity Index Measure (SSIM) metrics for objective and fair evaluations across scenarios and methods." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.485, + 0.3, + 0.501 + ], + "angle": 0, + "content": "4.3. Qualitative Comparison" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.509, + 0.471, + 0.705 + ], + "angle": 0, + "content": "Image Dehazing. We use RESIDE dataset [24] to validate the efficacy and robustness of LFP, as is shown in Fig 3, by comparing it with \\(\\mathrm{Dive + }\\) and HRDCP. While \\(\\mathrm{Dive + }\\) fails to restore image clarity, HRDCP results in color deviation, likely due to the specific absorption and scattering properties of sand particles. In contrast, our LFP method can generate more natural-looking results with better structures in a more stable manner. This demonstrates that images in hazy and underwater environments are formed based on different optical properties. We then compare the performance of the LFP with different methods like MSCNN, gUNet, and DCP. Learning-based methods such as gUNet and MSCNN are susceptible to dataset bias, which makes it difficult to" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.304 + ], + "angle": 0, + "content": "enhance visual quality. Meanwhile, traditional methods like DCP, which rely on hand-crafted priors, struggle with varied real-world conditions. LFP adeptly overcomes the major obstacle faced by previous dehazing approaches, which is the problematic influence of scene depth. The inherent uncertainty in depth estimation often leads to unstable performance. By confidently deriving the relationship between \\(\\Delta F\\) and \\(t\\) without involving \\(d\\), LFP successfully avoids the dataset bias that learning-based methods encounter. LFP effectively bypasses complex parameters of transmission and offers reliable solution with robust theoretical guarantees. Our method demonstrates a significantly stronger capability for robust image restoration in hazy weather conditions, resulting in highly satisfactory visual results." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.312, + 0.892, + 0.706 + ], + "angle": 0, + "content": "Underwater Image Enhancement. We validate the efficacy and robustness of LFP for underwater image enhancement through comprehensive experiments, using UIEBD dataset [26]. Results are shown in Fig 4. We first compare it with DCP and HRDCP, which are methodologies designed for different weather conditions. DCP fails completely, and although HRDCP achieves some visual effects, it suffers from noticeable color distortion. These experimental results underscore the necessity for image processing techniques to be specifically designed and optimized for each environment, reinforcing that an algorithm effective in one environment can't simply be applied to another without adaptation. \\(\\mathrm{Dive + }\\) and UWCNN struggle to effectively eliminate the green light prevalent in underwater scenes, leading to undesired color shifts. Although Waternet and ROP perform well in color restoration, they both suffer from dimmed lighting and reduced contrast, thereby compromising the overall visual quality. LFP overcomes the major obstacle faced by previous underwater image enhancement approaches, which is the varying influence of the attenuation coefficient \\(\\beta\\) on the RGB color channels that often results in color bias. By eliminating the attenuation coefficient \\(\\beta\\) during the derivation process, LFP avoids the color bias, rendering it a reliable solution with robust theoretical guarantees. Therefore, LFP method demonstrates competitive color correction and contrast enhancement capabilities." + }, + { + "type": "image", + "bbox": [ + 0.088, + 0.725, + 0.89, + 0.86 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.869, + 0.892, + 0.898 + ], + "angle": 0, + "content": "Figure 4. Comparison of different methods for underwater conditions. The algorithms designed for haze removal (grey) and sand removal (brown) fail in underwater (blue) environments. Our method corrects color shifts, enhance image brightness and improve visibility." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2748" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.085, + 0.088, + 0.887, + 0.209 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.212, + 0.895, + 0.255 + ], + "angle": 0, + "content": "Figure 5. Comparison of different methods for sandstorm image enhancement. The algorithms designed for haze removal (grey) and water removal (blue) fail in sandstorm (brown) environments. Our method corrects color shifts, enhances image brightness, and improves contrast, resulting in more vibrant color reproduction. Moreover, LFP produces a more realistic sky area closer to real scenes." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.269, + 0.473, + 0.602 + ], + "angle": 0, + "content": "Visibility Restoration in Sandstorms. We assess our LFP method's effectiveness in restoring visibility during sandstorms, by using SIRB dataset [46]. Results are shown in Fig 5. Initially, we compare LFP's performance with DCP, MSCNN, and \\(\\mathrm{Dive + }\\). The first two fail in dust and sandstorm mitigation, and \\(\\mathrm{Dive + }\\) introduces noticeable color deviations in sky areas. This emphasizes that hazy and sandstorm images are influenced by distinct optical properties, making dehaze or dewater algorithms unsuitable for sandstorms. Next, we contrast our method with CVC, HRDCP, and ROP. All three significantly enhance visibility in sandstorms. However, CVC lacks contrast, leading to detail loss. HRDCP, despite its visual improvements, exhibits a dim sky and color distortions due to its DCP-based design. ROP performs well but struggles with color distortion, resulting in a cool appearance. LFP overcomes prior challenges by addressing scene depth and attenuation coefficient \\(\\beta\\). By eliminating hard-to-estimate parameters like atmospheric light and scene depth, LFP avoids common issues, making it robust in sandstorm conditions without relying on assumptions. Thus, our LFP method excels in sandstorm image restoration, producing visually pleasing results without color distortion." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.603, + 0.47, + 0.741 + ], + "angle": 0, + "content": "In summary, the LFP has the ability to restore scenes in a variety of severe weather conditions, a characteristic that is absent in most algorithms. By eliminating the physical parameters that impact transmittance, LFP effectively bypasses the influence of scene depth, thus enhancing visual quality. Moreover, it minimizes color deviations by removing the attenuation coefficient. This approach provides a theoretically and empirically supported a priori method for scene restoration algorithms." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.755, + 0.312, + 0.772 + ], + "angle": 0, + "content": "4.4. Quantitative Comparison" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.78, + 0.473, + 0.903 + ], + "angle": 0, + "content": "Dehazing Quality We selected DCP, BCCR [34], and FVR [51] as our benchmarks to represent different types of prior-based dehazing methodologies. The rationale behind this selection, instead of opting for deep learning techniques, is multi-fold. Traditional dehazing methods not only possess the capability to restore clear images, but they also notably improve visual effects, thereby enhancing the visual experience. Furthermore, LFP does not contain any learnable pa" + }, + { + "type": "table_caption", + "bbox": [ + 0.498, + 0.268, + 0.895, + 0.351 + ], + "angle": 0, + "content": "Table 2. This table succinctly presents the comparative results of the LFP method and three other distinct prior methods. The evaluation spans across all datasets and metrics, with LFP consistently achieving superior performance. The highest PSNR and SSIM values, represented by bold text, clearly demonstrate the effectiveness of our LFP method." + }, + { + "type": "table", + "bbox": [ + 0.504, + 0.356, + 0.892, + 0.439 + ], + "angle": 0, + "content": "
MethodsRESIDE-INRESIDE-OUTHaze-4KRS-Haze
PSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIM
DCP16.620.81819.130.81514.010.76017.860.734
BCCR16.880.79114.020.77713.780.75618.840.747
FVR15.720.74811.610.66812.650.68816.120.626
LFP19.760.86921.050.86118.290.87318.240.786
" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.465, + 0.892, + 0.51 + ], + "angle": 0, + "content": "rameters, differentiating it from deep learning approaches. As illustrated in Tab 2, demonstrates that LFP outperforms other methods across all four of our test datasets." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.511, + 0.892, + 0.618 + ], + "angle": 0, + "content": "The performance is consistent whether we evaluate in terms of PSNR or SSIM. This robust evidence strongly validates the high generalization ability of the LFP method. Furthermore, the LFP method excels in restoring clear, detailed imagery, as particularly demonstrated by the significant improvement in SSIM scores. For more detailed results on desanding and dewatering, please refer to the appendix." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.628, + 0.892, + 0.78 + ], + "angle": 0, + "content": "Runtime Efficiency We conducted extensive comparisons on both CPU and GPU platforms. On the CPU, we compared LFP with traditional dehazing methods; on the GPU, we chose a variety of deep learning dehazing models, including: MSCNN, AOD-Net [23], GFN [41], GCANet [4], GDNet [30], MSBDN [6], FFA [39], CycleD [9], DehazeF [47], and gUNet-T. In terms of runtime, Tab 3 compares the efficiency on both CPU and GPU. The results show that the LFP method consistently achieves the fastest running speed, surpassing existing algorithms." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.79, + 0.892, + 0.913 + ], + "angle": 0, + "content": "CPU Performance. As evidenced by the data presented in Tab 3, the LFP model consistently exhibits superior runtime performance across all tested resolutions. With a runtime of merely 0.007 seconds at the lower-end resolution of \\(360\\mathrm{p}\\), and a still impressive 0.863 seconds at the high-end 4k resolution, these empirical findings underscore the remarkable efficiency of the LFP model when processing large-scale images or performing real-time scene recovery." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2749" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.089, + 0.472, + 0.16 + ], + "angle": 0, + "content": "Table 3. Run-time(seconds) performance. Testing the efficiency of different methods on various single image sizes. Models marked with (*) indicate that the results for those models were generated using MATLAB. The best performance is denoted in red, while the second best performance is indicated in blue." + }, + { + "type": "table", + "bbox": [ + 0.079, + 0.171, + 0.481, + 0.539 + ], + "angle": 0, + "content": "
CPU Performance
Model360p480p720p1080g2k4k
DCP[16]0.4590.7651.762.6006.7415.11
Retinex[22]0.1490.2860.4780.8042.3675.421
LDCP[55]0.5170.8212.2053.0918.26418.09
CAP[54]0.2180.3100.6661.0212.8666.403
BCCR[34]1.6913.5716.41611.7029.6064.26
FVR[51]0.3013.9028.58912.5834.3576.50
ROP*[27]0.0670.1440.4111.3992.0732.515
LFP0.0070.0170.0730.1510.4070.863
GPU Performance
MSCNN[40]0.1150.1470.2660.4621.1142.481
DehazeNet[3]0.1650.2230.4050.6211.3742.951
AOD-Net[23]0.0760.1060.1750.2910.6961.519
GFN*[41]0.6600.8581.4942.4456.35313.09
GCANet[4]0.4460.6411.1071.8214.2629.266
GDNet[30]0.6440.7421.5032.0274.67510.48
MSBDN[6]0.9521.2122.1943.3358.20317.46
FFA[39]1.6752.2533.9605.97714.2730.61
CycleD[9]0.6070.8101.4282.1585.15511.03
DehazeF[47]0.2890.3870.6771.0322.4755.310
gUNet-T[47]0.1970.2660.4730.7161.7193.669
ROP*[27]0.0470.0510.1130.2090.5320.894
LFP0.0010.0010.0020.0030.0080.019
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.564, + 0.47, + 0.687 + ], + "angle": 0, + "content": "GPU Performance. Tab 3 shows that LFPsignificantly outperforms all the compared algorithms in processing speed, especially for high-resolution images. For instance, at a resolution of 4K, LFP's average runtime is merely 0.019 seconds, whereas the runtimes of other algorithms range from 0.894 seconds to a staggering 30.609 seconds. This results in a speedup of up to 34 times for LFP, marking its superior computational efficiency." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.696, + 0.317, + 0.712 + ], + "angle": 0, + "content": "4.5. Ablation Study on \\(k\\) and \\(\\Omega\\)" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.72, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Our ablation study further reveals that the changes in the \\(k\\) parameter not only influence the luminous flux rate of change \\(\\Delta F\\) but also directly affect the transmission estimation. A properly tuned \\(k\\) facilitates an accurate estimation of the transmission map, thus effectively reducing the impact of different weather conditions on the image. On the other hand, the size of the sliding window \\(\\Omega\\) directly affects the local variance \\(\\Delta F\\). A larger \\(\\Omega\\) captures more significant changes in the luminous flux, which is particularly effective in dense haze conditions or when the scene has a greater depth of view. This broadened perspective allows for a more comprehensive understanding of the scene, lead" + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.088, + 0.895, + 0.378 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.39, + 0.895, + 0.434 + ], + "angle": 0, + "content": "Figure 6. The x-axis represents \\( k \\), the y-axis represents \\( \\Omega \\). When the depth of view is further, or the haze is denser, \\( \\Omega \\) and \\( k \\) need to be increased correspondingly." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.446, + 0.895, + 0.583 + ], + "angle": 0, + "content": "ing to a higher-quality image after enhancement. Finally, the interplay between \\( k \\) and \\( \\Omega \\) is also crucial. Balancing these two factors allows for a fine-tuned approach to image enhancement, taking into account both the overall brightness and the local variance of the image, this ability to cope with complex conditions adds another layer of robustness to our technique. This balance ensures that the final output image is visually pleasing while maintaining the fidelity of the original scene." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.595, + 0.62, + 0.611 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.621, + 0.894, + 0.833 + ], + "angle": 0, + "content": "In this work, we propose an efficient scene recovery algorithm to restore clear images under poor visibility conditions. By analyzing the image formation model, we derive an analytical form for transmission that eliminates multiple influencing factors and leverages the correlation with flux change rate as prior knowledge. Our approach achieves state-of-the-art performance while maintaining low computational complexity, enabling real-time processing on resource-limited devices. Extensive experiments validate the effectiveness of the proposed method. The novel integration of prior knowledge and analytical transmission modeling provides opportunities to inform future learning-based methods. Overall, this paper makes significant contributions towards robust scene recovery under challenging imaging conditions." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.845, + 0.679, + 0.862 + ], + "angle": 0, + "content": "6. Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.892, + 0.901 + ], + "angle": 0, + "content": "This work was supported by the National Natural Science Foundation of China (No. 62336001, No. 62072014)" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2750" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.173, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.468, + 0.156 + ], + "angle": 0, + "content": "[1] Dana Berman, Shai Avidan, et al. Non-local image dehazing. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1674-1682, 2016. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.159, + 0.468, + 0.199 + ], + "angle": 0, + "content": "[2] Trung Minh Bui and Wonha Kim. Single image dehazing using color ellipsoid prior. IEEE Transactions on Image Processing, 27(2):999-1009, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.089, + 0.201, + 0.468, + 0.255 + ], + "angle": 0, + "content": "[3] Bolun Cai, Xiangmin Xu, Kui Jia, Chunmei Qing, and Dacheng Tao. Dehazenet: An end-to-end system for single image haze removal. IEEE Transactions on Image Processing, 25(11):5187-5198, 2016. 1, 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.089, + 0.257, + 0.468, + 0.338 + ], + "angle": 0, + "content": "[4] Dongdong Chen, Mingming He, Qingnan Fan, Jing Liao, Liheng Zhang, Dongdong Hou, Lu Yuan, and Gang Hua. Gated context aggregation network for image dehazing and deraining. In 2019 IEEE winter conference on applications of computer vision (WACV), pages 1375-1383. IEEE, 2019. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.341, + 0.468, + 0.396 + ], + "angle": 0, + "content": "[5] Qili Deng, Ziling Huang, Chung-Chi Tsai, and Chia-Wen Lin. Hardgan: A haze-aware representation distillation gan for single image dehazing. In European Conference on Computer Vision, pages 722-738. Springer, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.398, + 0.468, + 0.465 + ], + "angle": 0, + "content": "[6] Hang Dong, Jinshan Pan, Lei Xiang, Zhe Hu, Xinyi Zhang, Fei Wang, and Ming-Hsuan Yang. Multi-scale boosted dehazing network with dense feature fusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2157-2167, 2020. 3, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.468, + 0.468, + 0.508 + ], + "angle": 0, + "content": "[7] Jiangxin Dong and Jinshan Pan. Physics-based feature dehazing networks. In European Conference on Computer Vision, pages 188-204. Springer, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.51, + 0.468, + 0.591 + ], + "angle": 0, + "content": "[8] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.594, + 0.468, + 0.661 + ], + "angle": 0, + "content": "[9] Deniz Engin, Anil Genç, and Hazim Kemal Ekenel. Cycle-dehaze: Enhanced cyclegan for single image dehazing. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition Workshops, pages 825-833, 2018. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.664, + 0.468, + 0.704 + ], + "angle": 0, + "content": "[10] Faming Fang, Fang Li, and Tieyong Zeng. Single image dehazing and denoising: a fast variational approach. SIAM Journal on Imaging Sciences, 7(2):969-996, 2014. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.706, + 0.468, + 0.773 + ], + "angle": 0, + "content": "[11] Xueyang Fu, Yue Huang, Delu Zeng, Xiao-Ping Zhang, and Xinghao Ding. A fusion-based enhancing approach for single sandstorm image. In 2014 IEEE 16th international workshop on multimedia signal processing (MMSP), pages 1-5. IEEE, 2014. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.776, + 0.468, + 0.844 + ], + "angle": 0, + "content": "[12] Han Gao, Ping Wei, and Jun Ke. Color enhancement and image defogging in hsi based on retinax model. In 2015 International Conference on Optical Instruments and Technology: Optoelectronic Imaging and Processing Technology, pages 9-16. SPIE, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.846, + 0.468, + 0.901 + ], + "angle": 0, + "content": "[13] Jie Gui, Xiaofeng Cong, Yuan Cao, Wenqi Ren, Jun Zhang, Jing Zhang, Jieuxin Cao, and Dacheng Tao. A comprehensive survey and taxonomy on single image dehazing based on deep learning. ACM Computing Surveys, 55(13s):1-37, 2023. 1" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.468, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.161 + ], + "angle": 0, + "content": "[14] Chun-Le Guo, Qixin Yan, Saeed Anwar, Runmin Cong, Wenqi Ren, and Chongyi Li. Image dehazing transformer with transmission-aware 3d position embedding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5812-5820, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.164, + 0.892, + 0.216 + ], + "angle": 0, + "content": "[15] Yecai Guo, Hanyu Li, and Peixian Zhuang. Underwater image enhancement using a multiscale dense generative adversarial network. IEEE Journal of Oceanic Engineering, 45(3):862-870, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.219, + 0.892, + 0.272 + ], + "angle": 0, + "content": "[16] Kaiming He, Jian Sun, and Xiaou Tang. Single image haze removal using dark channel prior. IEEE transactions on pattern analysis and machine intelligence, 33(12):2341-2353, 2010. 1, 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.274, + 0.892, + 0.315 + ], + "angle": 0, + "content": "[17] Kaiming He, Jian Sun, and Xiaou Tang. Guided image filtering. IEEE transactions on pattern analysis and machine intelligence, 35(6):1397-1409, 2012. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.317, + 0.892, + 0.369 + ], + "angle": 0, + "content": "[18] Jong-Ju Jeon, Tae-Hee Park, and Il-Kyu Eom. Sand-dust image enhancement using chromatic variance consistency and gamma correction-based dehazing. Sensors, 22(23):9048, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.372, + 0.892, + 0.412 + ], + "angle": 0, + "content": "[19] Mingye Ju, Can Ding, Y Jay Guo, and Dengyin Zhang. Idgcp: Image dehazing based on gamma correction prior. IEEE Transactions on Image Processing, 29:3104-3118, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.414, + 0.892, + 0.455 + ], + "angle": 0, + "content": "[20] Manjit Kaur, Dilbag Singh, Vijay Kumar, and Kehui Sun. Color image dehazing using gradient channel prior and guided 10 filter. Information Sciences, 521:326-342, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.457, + 0.892, + 0.523 + ], + "angle": 0, + "content": "[21] Johannes Kopf, Boris Neubert, Billy Chen, Michael Cohen, Daniel Cohen-Or, Oliver Deussen, Matt Uytendaele, and Dani Lischinski. Deep photo: Model-based photograph enhancement and viewing. ACM transactions on graphics (TOG), 27(5):1-10, 2008. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.525, + 0.892, + 0.551 + ], + "angle": 0, + "content": "[22] Edwin H Land. The retina theory of color vision. Scientific american, 237(6):108-129, 1977. 4, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.554, + 0.892, + 0.607 + ], + "angle": 0, + "content": "[23] Boyi Li, Xiulian Peng, Zhangyang Wang, Jizheng Xu, and Dan Feng. Aod-net: All-in-one dehazing network. In Proceedings of the IEEE international conference on computer vision, pages 4770-4778, 2017. 1, 3, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.61, + 0.892, + 0.663 + ], + "angle": 0, + "content": "[24] Boyi Li, Wenqi Ren, Dengpan Fu, Dacheng Tao, Dan Feng, Wenjun Zeng, and Zhangyang Wang. Benchmarking single-image dehazing and beyond. IEEE Transactions on Image Processing, 28(1):492-505, 2019. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.665, + 0.892, + 0.719 + ], + "angle": 0, + "content": "[25] Chongyi Li, Chunle Guo, Wenqi Ren, Runmin Cong, Junhui Hou, Sam Kwong, and Dacheng Tao. An underwater image enhancement benchmark dataset and beyond. IEEE Transactions on Image Processing, 29:4376-4389, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.721, + 0.892, + 0.773 + ], + "angle": 0, + "content": "[26] Chongyi Li, Saeed Anwar, and Fatih Porikli. Underwater scene prior inspired deep underwater image and video enhancement. Pattern Recognition, 98:107038, 2020. 1, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.776, + 0.892, + 0.83 + ], + "angle": 0, + "content": "[27] Jun Liu, Wen Liu, Jianing Sun, and Tieyong Zeng. Rank-one prior: Toward real-time scene recovery. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14802-14810, 2021. 1, 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.832, + 0.892, + 0.899 + ], + "angle": 0, + "content": "[28] Risheng Liu, Xin Fan, Ming Zhu, Minjun Hou, and Zhongxuan Luo. Real-world underwater enhancement: Challenges, benchmarks, and solutions under natural light. IEEE transactions on circuits and systems for video technology, 30(12): 4861-4875, 2020. 3" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.514, + 0.957 + ], + "angle": 0, + "content": "2751" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.147 + ], + "angle": 0, + "content": "[29] Xiaodong Liu, Zhi Gao, and Ben M Chen. Mlfcgan: Multilevel feature fusion-based conditional gan for underwater image color correction. IEEE Geoscience and Remote Sensing Letters, 17(9):1488-1492, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.149, + 0.472, + 0.217 + ], + "angle": 0, + "content": "[30] Xiaohong Liu, Yongrui Ma, Zhihao Shi, and Jun Chen. Griddehazenet: Attention-based multi-scale network for image dehazing. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7314-7323, 2019. 3, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.22, + 0.47, + 0.289 + ], + "angle": 0, + "content": "[31] Yu Liu, Guanlong Zhao, Boyuan Gong, Yang Li, Ritu Raj, Niraj Goel, Satya Kesav, Sandeep Gottimukkala, Zhangyang Wang, Wenqi Ren, et al. Improved techniques for learning to dehaze and beyond: A collective study. arXiv preprint arXiv:1807.00202, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.291, + 0.47, + 0.359 + ], + "angle": 0, + "content": "[32] Ye Liu, Lei Zhu, Shunda Pei, Huazhu Fu, Jing Qin, Qing Zhang, Liang Wan, and Wei Feng. From synthetic to real: Image dehazing collaborating with unlabeled real data. In Proceedings of the 29th ACM International Conference on Multimedia, pages 50-58, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.361, + 0.469, + 0.389 + ], + "angle": 0, + "content": "[33] Rastislav Lukac. Single-sensor imaging: methods and applications for digital cameras. CRC Press, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.391, + 0.47, + 0.459 + ], + "angle": 0, + "content": "[34] Gaofeng Meng, Ying Wang, Jiangyong Duan, Shiming Xiang, and Chunhong Pan. Efficient image dehazing with boundary constraint and contextual regularization. In Proceedings of the IEEE international conference on computer vision, pages 617-624, 2013. 2, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.461, + 0.469, + 0.488 + ], + "angle": 0, + "content": "[35] WE Knowles Middleton. Bouguer, lambert, and the theory of horizontal visibility. Isis, 51(2):145-149, 1960. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.49, + 0.469, + 0.53 + ], + "angle": 0, + "content": "[36] Srinivasa G Narasimhan and Shree K Nayar. Vision and the atmosphere. International journal of computer vision, 48(3): 233-254, 2002. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.533, + 0.47, + 0.588 + ], + "angle": 0, + "content": "[37] Srinivasa G Narasimhan and Shree K Nayar. Interactive (de) weathering of an image using physical models. In IEEE Workshop on color and photometric Methods in computer Vision, page 1. France, 2003. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.59, + 0.469, + 0.63 + ], + "angle": 0, + "content": "[38] Ko Nishino, Louis Kratz, and Stephen Lombardi. Bayesian defogging. International journal of computer vision, 98(3): 263-278, 2012. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.633, + 0.469, + 0.687 + ], + "angle": 0, + "content": "[39] Xu Qin, Zhilin Wang, Yuanchao Bai, Xiaodong Xie, and Huizhu Jia. Ffa-net: Feature fusion attention network for single image dehazing. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 11908-11915, 2020. 3, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.689, + 0.469, + 0.745 + ], + "angle": 0, + "content": "[40] Wenqi Ren, Si Liu, Hua Zhang, Jinshan Pan, Xiaochun Cao, and Ming-Hsuan Yang. Single image dehazing via multiscale convolutional neural networks. In European conference on computer vision, pages 154-169. Springer, 2016. 1, 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.747, + 0.469, + 0.814 + ], + "angle": 0, + "content": "[41] Wenqi Ren, Lin Ma, Jiawei Zhang, Jinshan Pan, Xiaochun Cao, Wei Liu, and Ming-Hsuan Yang. Gated fusion network for single image dehazing. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 3253-3261, 2018. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.817, + 0.469, + 0.871 + ], + "angle": 0, + "content": "[42] Wenqi Ren, Jingang Zhang, Xiangyu Xu, Lin Ma, Xiaochun Cao, Gaofeng Meng, and Wei Liu. Deep video dehazing with semantic segmentation. IEEE Transactions on Image Processing, 28(4):1895-1908, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[43] Christos Sakaridis, Dengxin Dai, Simon Hecker, and Luc Van Gool. Model adaptation with synthetic and real data for" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.892, + 0.135 + ], + "angle": 0, + "content": "semantic dense foggy scene understanding. In Proceedings of the European Conference on Computer Vision (ECCV), pages 687-704, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.136, + 0.892, + 0.19 + ], + "angle": 0, + "content": "[44] Yoav Y Schechner and Yuval Averbuch. Regularized image recovery in scattering media. IEEE Transactions on Pattern Analysis and Machine Intelligence, 29(9):1655-1660, 2007. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.193, + 0.892, + 0.26 + ], + "angle": 0, + "content": "[45] Zhenghao Shi, Yaning Feng, Minghua Zhao, Erhu Zhang, and Lifeng He. Let you see in sand dust weather: A method based on halo-reduced dark channel prior dehazing for sand-dust image enhancement. *Ieee Access*, 7:116722–116733, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.263, + 0.892, + 0.318 + ], + "angle": 0, + "content": "[46] Yazhong Si, Fan Yang, Ya Guo, Wei Zhang, and Yipu Yang. A comprehensive benchmark analysis for sand dust image reconstruction. Journal of Visual Communication and Image Representation, 89:103638, 2022. 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.32, + 0.892, + 0.36 + ], + "angle": 0, + "content": "[47] Yuda Song, Zhuqing He, Hui Qian, and Xin Du. Vision transformers for single image dehazing. arXiv preprint arXiv:2204.03883, 2022. 3, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.362, + 0.892, + 0.403 + ], + "angle": 0, + "content": "[48] Yuda Song, Yang Zhou, Hui Qian, and Xin Du. Rethinking performance gains in image dehazing networks. arXiv preprint arXiv:2209.11448, 2022. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.405, + 0.892, + 0.46 + ], + "angle": 0, + "content": "[49] Matan Sulami, Itamar Glatzer, Raanan Fattal, and Mike Werman. Automatic recovery of the atmospheric light in hazy images. In 2014 IEEE International Conference on Computational Photography (ICCP), pages 1-11. IEEE, 2014. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.462, + 0.892, + 0.503 + ], + "angle": 0, + "content": "[50] Robby T Tan. Visibility in bad weather from a single image. In 2008 IEEE Conference on Computer Vision and Pattern Recognition, pages 1-8. IEEE, 2008. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.505, + 0.892, + 0.559 + ], + "angle": 0, + "content": "[51] Jean-Philippe Tarel and Nicolas Hautiere. Fast visibility restoration from a single color or gray level image. In 2009 IEEE 12th International Conference on Computer Vision, pages 2201–2208. IEEE, 2009. 2, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.561, + 0.892, + 0.601 + ], + "angle": 0, + "content": "[52] Ting Yan, Liejun Wang, and Jiaxing Wang. Method to enhance degraded image in dust environment. J. Softw., 9(10): 2672-2677, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.604, + 0.892, + 0.644 + ], + "angle": 0, + "content": "[53] Dong Zhao, Jia Li, Hongyu Li, and Long Xu. Hybrid local-global transformer for image dehazing. arXiv preprint arXiv:2109.07100, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.647, + 0.892, + 0.701 + ], + "angle": 0, + "content": "[54] Qingsong Zhu, Jiaming Mai, and Ling Shao. A fast single image haze removal algorithm using color attenuation prior. IEEE transactions on image processing, 24(11):3522-3533, 2015. 1, 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.703, + 0.892, + 0.757 + ], + "angle": 0, + "content": "[55] Yingying Zhu, Gaoyang Tang, Xiaoyan Zhang, Jianmin Jiang, and Qi Tian. Haze removal method for natural restoration of images with sky. neurocomputing, 275:499-510, 2018. 2, 8" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.892, + 0.757 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2752" + } + ] +] \ No newline at end of file diff --git a/2024/Efficient Scene Recovery Using Luminous Flux Prior/0327d2f4-6c3a-4e74-a5a8-041807da336a_origin.pdf b/2024/Efficient Scene Recovery Using Luminous Flux Prior/0327d2f4-6c3a-4e74-a5a8-041807da336a_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..6a3a0b3c5da79d873cd1a11f125866b4d2c23373 --- /dev/null +++ b/2024/Efficient Scene Recovery Using Luminous Flux Prior/0327d2f4-6c3a-4e74-a5a8-041807da336a_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a2cdfd93f3d4f7ff9e2a658f1ae498ac9f8a7b6f821723631605b620dcd54934 +size 3951353 diff --git a/2024/Efficient Scene Recovery Using Luminous Flux Prior/full.md b/2024/Efficient Scene Recovery Using Luminous Flux Prior/full.md new file mode 100644 index 0000000000000000000000000000000000000000..1c2ef3cdf566e6e3cc59cd20ca19bc330c53037d --- /dev/null +++ b/2024/Efficient Scene Recovery Using Luminous Flux Prior/full.md @@ -0,0 +1,344 @@ +# Efficient Scene Recovery Using Luminous Flux Prior + +Zhongyu Li1, Lei Zhang1* +1University of Science and Technology of China, Hefei, China +ruin2@mail.ustc.edu.cn, leizh23@ustc.edu.cn, + +# Abstract + +Scene recovery, the restoration of images degraded by adverse weather conditions, presents significant challenges for existing methods. Physical models, constrained by their inherent assumptions, often fail when these assumptions are not met; Deep learning models are powerful, they are limited by the diversity of their training datasets, leading to poor generalization and high computational demands. To address these limitations, we propose the Luminous Flux Prior (LFP), to recover degraded images under diverse adverse weather without learning. Luminous flux, a physical measure that reflects image brightness, has a rate of change that demonstrates a significant correlation with transmission. Consequently, we leverage this rate of change in luminous flux as prior knowledge to estimate transmission, which in turn assists in image recovery. This approach reduces dependency on physical parameters and enhances adaptability to various weather. Experimental validation under diverse conditions, such as sandstorms, underwater environments, and haze, attests to the robustness of LFP in restoring clear images. With a time complexity of $\mathcal{O}(N\log N)$ , LFP enables real-time recovery, making it a suitable for devices with limited computational resources. + +# 1. Introduction + +Observational devices encounter significant challenges under adverse lighting conditions, which can be caused by phenomena such as sandstorms, underwater environments, and haze. These conditions lead to image degradation, typified by a loss of detail, contrast distortion, and color bias [13, 50]. Such degradation impairs computer vision tasks, including object detection [31], image segmentation [42], and scene understanding [43]. Therefore, restoring clear images from these degraded observations becomes a critical challenge in the field of computer vision [27]. + +Image degradation primarily arises from the scattering and absorption of light [10, 11, 16, 26, 49]. Environmental elements, such as the atmosphere and water, scatter + +light sources, engendering intricate environmental light and the absorption of parts of the visible light spectrum. This process significantly degrades images by diminishing their brightness. The physical quantity symbolizing the brightness per unit area, referred to as luminous flux, is denoted by $F$ . The fundamental instigator of this degradation is a decrease in the transmission rate $t$ , representing the proportion of light brightness received by the camera. + +The challenge in image restoration is the estimation of transmission $t$ . We bifurcate scene restoration methods into two principal types based on the calculation of $t$ : physical models and data-driven deep learning methods. Physical models [1, 2, 16, 19, 20, 38, 54] make assumptions about the physical parameters or conditions for transmission $t$ to be valid. However, if these assumptions fall short, the restored image suffers to varying extents. Conversely, data-driven deep learning methods extrapolate an estimation model for haze-free images from extensive datasets [3, 23, 40]. Yet, these supervised learning methods exhibit limited efficacy in real-world scenarios due to the scarcity of large-scale, diverse, and fully aligned paired training data. Additionally, deep learning algorithms necessitate complex loss functions and training strategies, resulting in high training costs [48]. Most existing methods rely solely on L1/L2 image reconstruction loss and lack robust regularization, which may induce unsuitable gradient updates and grapple with the uncertainty of ill-posed problems due to a dearth of robust prior information. Both physical models and deep learning methods find it challenging to adapt to a variety of different scenes owing to their inherent limitations [27]. + +Typical physical models illustrate the image formation process impacted by scattering [44]. This model can generally be represented by the following equation [36, 50]: + +$$ +\mathbf {I} (x) = \mathbf {J} (x) \cdot t (x) + \mathbf {A} (1 - t (x)), \tag {1} +$$ + +where $x$ denotes the location information within the image, $\mathbf{I}(x)$ refers to the observed degraded image, $\mathbf{J}(x)$ signifies the clear image (scene radiance) we aim to restore, and $\mathbf{A}$ symbolizes the global ambient light. $t(x)$ describes the proportion of light reaching the camera. To restore a clear image $\mathbf{J}(x)$ , we need to estimate the transmission $t(x)$ and + +![](images/b44ffb7f4ee619adbbb03f00640e0649b13199f0da7b46db6cfc780785e1448d.jpg) +(a) + +![](images/aec446a9c0bf0e24348e389fb95ab6c5673e3e1742f7a9f13e1f205c83c75e0e.jpg) +(b) + +![](images/7651c8eca67c9c20a28419ad349cdba7054d9d53d6e7061f734c537dc06b6b19.jpg) +(c) + +![](images/810751d257b6c0fd4ad051ee129986f7996ebe0364a978aff362a615bcf425af.jpg) +(d) +Figure 1. Challenging real scene recovery using LFP. The first row depicts degraded images, while the second row presents the restored clear images. These images represent three adverse conditions: haze, sandstorm, and underwater environments. The restored images all exhibit superior visual effects. Notably, the color bias in images (b), (c), and (e) has been rectified; images (c) and (g) are free from halo artifacts; and the brightness in images (a), (b), (c), (d), and (f) has been supplemented and adjusted. + +![](images/79cdd6655033a93ca105738d6b8091ff4b7d72814c8d27a51784f1ee04e05694.jpg) +(f) + +![](images/5a2787ff6b6525e0d82e2735e6c0dca5c50fb7e4e3f29c888b02e717dc4735f3.jpg) +(g) + +ambient light $\mathbf{A}$ . $t(x)$ can be defined as: + +$$ +t (x) = e ^ {- \beta d (x)}, \tag {2} +$$ + +where $\beta$ is the attenuation coefficient, indicating the degree of scattering in the environmental medium, and $d(x)$ designates the scene depth [54]. + +The complexity of estimating transmission lies in its dependence on multiple physical parameters [54]. Adverse weather conditions that diminish luminous flux $F$ (detailed in Sec.3) result in image darkening, attributed to decreased $t$ . Utilizing $F$ directly to estimate $t$ is impractical. We propose LFP that employs the rate of change of luminous flux, $\Delta F$ (detailed in Sec.3), as a prior to simplify the process. In extensive experimental validation, we found $\Delta F$ correlates significantly with $t$ , thereby transforming $t$ estimation into a $\Delta F$ computation. As a dimensionless quantity, $\Delta F$ exhibits wide applicability, underscoring the adaptability of LFP. LFP is computationally efficient and generates visually pleasing images, as demonstrated by real scenes in Fig 1. + +In summary, our contributions are as follows: + +1. We introduce a new method to estimate transmission $(t)$ , leveraging the rate of change in luminous flux $(\Delta F)$ . This approach is widely applicable across different weather and imaging conditions, and does not require any learnable parameters, thus eliminating the need for dataset construction. +2. Our algorithm has a low computational complexity, $O(N \cdot \log N)$ , where $N$ is the size of a single image. With GPU acceleration, it is exceedingly fast, enhancing computation speeds by 7 to 50 times compared to CPU processing. This makes the algorithm suitable for devices with limited computational power. +3. We have transformed the estimation of $t$ into a computation involving $F$ and $\Delta F$ , eliminates the influence of physical parameters on $t$ . Experimental validation confirms that this method delivers robust performance and can enhance images to a high visual quality. + +# 2. Related Works + +Physical Models: Physical models employ image statistical attributes to design priors. For instance, Dark Channel Prior (DCP) [16] utilizes the minimum value in the RGB channels of pixels. Building upon DCP, BCCR [34] introduces a more efficient method for estimating transmission coefficients. Other methods, such as Fast Visibility Restoration (FVR) [51], make assumptions about the atmospheric veil, considering it locally flat, and use median filters instead of minimum filters for transmission estimation. Color Attenuation Prior (CAP) [54] formulates a linear model between depth information and the attenuation coefficient, while ROP [27] proposes an intensity projection strategy for transmission estimation. These prior-based methods perform admirably in specific scene types, but their adaptability to diverse scenes is limited. For example, DCP tends to darken already dark areas (i.e. skies and shadows) when dehazing. To rectify this, LDCP [55] modifies sky regions. Brightness is a crucial factor for accurate transmission estimation [55], enhancing contrast and improving visual quality. Many image enhancement techniques leverage this attribute [12, 45, 52]. These algorithms typically convert RGB images to other color spaces, such as YUV and LAB, to perform color correction [45, 52], for example, HRDCP [45] and CVC [18]. In dealing with sandstorms and underwater environments, color correction methods like white balance, akin to the Gray World Hypothesis [33], are important. The hypothesis posits that the average values of the red, green, and blue channels in a scene should be approximately equal. Considering that sandstorm weather absorbs blue and green light, and underwater environments absorb red light, using color balance as prior knowledge can result in more realistic recovery effects [33]. + +Data-Driven Methods: Learning-based methods extrapolate models for haze-free image estimation from extensive datasets. Initial models like Waternet [25], MSCNN [40], + +and AOD-Net [23], equipped with shallow architectures, outperformed previous prior-based methods. Some even incorporated prior knowledge, such as UWCNN [26] and DehazeNet [3]. Lately, most learning-based methods design more intricate and larger parameterized network architectures to enhance the network's expressivity. These include the attention mechanism [5, 7, 39], multi-scale features [6, 15, 29, 30], and even dehazing networks based on Vision Transformer (ViT) [8, 14, 53], such as gUNet [47]. Despite these advancements, the efficacy of these supervised learning methods is limited by their generalization ability in real-world scenes. The scarcity of large-scale, diverse, and fully aligned paired training data poses a significant hurdle. These methods often rely on synthetic data, but procuring synthetic data for sand and underwater environments presents challenges [28, 46]. Deep learning-based algorithms also encounter difficulties in training, requiring complex loss functions and training strategies, which lead to high training costs [48]. Many existing methods only use image reconstruction loss based on L1/L2, lacking robust regularization. This deficiency can result in unsuitable gradient updates and difficulties in grappling with the uncertainty of ill-conditioned problems, underscoring the need for robust prior information. + +# 3. Method + +In this section, we first validate the correlation between $\Delta F$ and $t$ from a statistical perspective (Sec.3.1). Next, we derive the analytical form of transmittance using luminous flux and its rate of change (Sec.3.2). Finally, we present the numerical calculation methods for luminous flux and its rate of change, the image recovery formula and analysis of algorithmic complexity (Sec.3.3). + +# 3.1. Luminous Flux Prior Validation + +We validate the correlation between the rate of change in luminous flux, denoted as $\Delta F$ , and transmission $t$ , utilizing datasets from three disparate weather conditions. These conditions encompass images from haze, sandstorms, and underwater environments, both simulated and real. The correlation between $\Delta F$ and $t$ remains robust across all conditions, which underscores the utility of $\Delta F$ as a prior for image enhancement. The datasets we use are as follows: + +1. Haze: We utilize the RESIDE dataset [24], which contains 13,990 synthetic indoor and 72,135 synthetic outdoor images. +2. Sandstorm: We rely on the Sand-dust Image Reconstruction Benchmark (SIRB) dataset [46], which simulates sandstorm conditions with different dust intensities. +3. Underwater: We employ the Underwater Image Enhancement Benchmark Dataset (UIEBD) [26] which includes 890 real underwater environment images together + +Table 1. We apply the KLD to measure the difference between the rate of change of luminance and the transmission map. Here, $N$ symbolizes a truncated standard normal distribution, and $U$ represents a uniform distribution within the range [0, 1]. + +
RESIDESIRBO-HazeUIEBD
ITSOTSLightMediumDense
ΔF0.290.210.130.130.160.120.15
N0.870.640.750.991.660.530.77
U0.950.531.271.681.900.430.64
+ +with reference images free of water distortion, obtained via their Deep Underwater Image Enhancement Network. + +In total, these datasets encompass 90,960 images, comprising our entire sample space with each image representing an individual sample. For an image of size $m \times n$ , denoted as I, each pixel corresponds to three values: $t$ , $F$ , and $\Delta F$ . T, F, and $\Delta \mathbf{F}$ represent these values in matrix form. Both $\Delta \mathbf{F}$ and $\mathbf{T}$ are of the same size as image I. $\Delta \mathbf{F} \in \mathbb{R}^{m \times n \times c}$ , where $c$ denotes the number of channels in image I. For instance, when I exists in the RGB color space, $c = 3$ ; when I is a grayscale image, $c = 1$ . + +We employ the Kullback-Leibler Divergence (KLD) to authenticate the correlation between $\Delta \mathbf{F}$ and $\mathbf{T}$ , measuring their difference as follows: + +$$ +K L D (\Delta \mathbf {F} | | \mathbf {T}) \tag {3} +$$ + +We generate the transmission map $\mathbf{T}$ of an image by subtracting a low-quality image from a clear counterpart and taking the absolute value. The resulting image represents the degree of light scattering, with brighter regions indicating severe degradation and darker regions suggesting minor or no degradation. We temporarily disregard the effect of atmospheric light. Subsequently, we compute $\Delta \mathbf{F}$ for the degraded image and then verify the correlation between $\Delta \mathbf{F}$ and $\mathbf{T}$ using Eq 3. + +As presented in Tab 1, the KLD values for $\mathbf{T}$ when using uniform or Gaussian distributions range between 0.53 to 1.66 and 0.43 to 1.90 respectively. It is important to note that the use of uniform or normal distributions does not facilitate image recovery, thus their KL values can be perceived as the upper bounds. As the KLD values inch closer to these upper bounds, image recovery becomes increasingly challenging. In contrast, the KLD values for $\Delta F$ fall within the significantly lower range of 0.12 to 0.29, approximately 4 to 10 times less than the aforementioned upper bounds. This signifying relatively low information loss when estimating $\mathbf{T}$ with $\Delta \mathbf{F}$ . Therefore, $\Delta \mathbf{F}$ effectively encapsulates the characteristics of $\mathbf{T}$ , indicating a pronounced correlation between $\Delta \mathbf{F}$ and $\mathbf{T}$ . These findings substantiate the effectiveness of the LFP. + +![](images/64e2b6d36d125d3a553138e24ad4df4bc0a5962db75ac08d89189efd59788888.jpg) +Figure 2. Visualization of the LFP Algorithm Workflow: First, it calculates Luminous Flux $F$ and it's ratio of change $\Delta F$ within a sliding window across the image. Second, it estimates transmission $t$ using $F$ and $\Delta F$ . Finally, recover the image using Eq 15. + +# 3.2. Scene Recovery with LFP + +# 3.2.1 Analytical Form of Luminous Flux + +The Bouguer-Lambert law [35] offers a comprehensive analytical expression for the measurement of luminous flux. This law utilizes the attenuation coefficient $\beta$ [37] and the scene depth $d(x)$ [21] as primary factors to accurately evaluate luminous flux Eq 4. In this section, to ensure clarity and simplicity, we have chosen to omit the position information denoted by $x$ from the symbols used. + +$$ +F = \tilde {F} e ^ {- \beta d} \tag {4} +$$ + +where $\tilde{F}$ represents the luminous flux of incident light. + +# 3.2.2 Estimating Transmission + +Transmission, a property of the material, is the ratio of $F$ to $\tilde{F}$ , as depicted in Eq 5: + +$$ +t = F / \tilde {F}. \tag {5} +$$ + +In line with the Retinex [22] model, we perceive an image as composed of incident and reflected components. Incident light illuminates the object, and the object's reflected light forms the image we observe. The incident light $\tilde{F}$ can be eliminated using division, as shown in Eq 6: + +$$ +t _ {i j} = \frac {t _ {i}}{t _ {j}} = \frac {F _ {i} / \tilde {F} _ {i}}{F _ {j} / \tilde {F} _ {j}} = F _ {i} / F _ {j}, \tag {6} +$$ + +where $i$ and $j$ represent two adjacent pixels $x_{i}$ and $x_{j}$ , $\tilde{F}_{i} = \tilde{F} j$ , $t_{ij}$ is a dimensionless parameter depicting the luminous flux $F$ ratio between adjacent pixels. This allows analysis of flux changes within the image, crucial for revealing local features and their variations. + +Based on Eq 6, $t_{ij}$ can be further expressed as: + +$$ +t _ {i j} = F _ {i} / F _ {j} = e ^ {- d _ {i} \cdot \left(\beta_ {i} - \beta_ {j}\right)}. \tag {7} +$$ + +In Eq 7, to comprehend the influence of luminous flux on $\beta$ , we introduce $\mathrm{d}\beta = \frac{\beta_i - \beta_j}{F_i - F_j}$ , where $\mathrm{d}\beta$ is a differential with respect to $\beta$ . By substituting $\mathrm{d}\beta$ into Eq 7, we obtain: + +$$ +\begin{array}{l} t _ {i j} = e ^ {- d _ {i} \left(\beta_ {i} - \beta_ {j}\right)} \\ = e ^ {- d _ {i} \left(\beta_ {i} - \beta_ {j}\right) \frac {\Delta F}{\Delta F}} \tag {8} \\ = e ^ {- d _ {i} \cdot \Delta F \cdot \mathrm {d} \beta}. \\ \end{array} +$$ + +Simultaneously integrating both sides of Eq 8, obtain: + +$$ +\begin{array}{l} \int t \mathrm {d} t = \int e ^ {- d \cdot \Delta F \cdot \mathrm {d} \beta} \tag {9} \\ t = \sqrt {2 e ^ {- d \cdot \beta \cdot \Delta F}}. \\ \end{array} +$$ + +The proof is provided in the appendix. + +For ease of discussion, we consider $\tilde{\mathbf{F}}$ as a constant, denoted as $k$ , where $k \in (0,1]$ . From Eq 4, $\beta$ can be expressed as in Eq 10: + +$$ +\beta = - \frac {k}{d} \log F. \tag {10} +$$ + +By substituting Eq 10 into Eq 9, we obtain the expression for $t$ as follows: + +$$ +\begin{array}{l} t = \sqrt {2 e ^ {- d \cdot \beta \cdot \Delta F}} \\ = \sqrt {2 e ^ {\log F \cdot k \cdot \Delta F}} \tag {11} \\ = \sqrt {2 F ^ {k \Delta F}}. \\ \end{array} +$$ + +In Eq 11, we have successfully eliminated the two physical parameters $d(x)$ and $\beta$ and derived an analytical form for $t$ . Considering that the exponent part is relatively small compared to the base, Eq 11 can be further simplified, resulting in Eq 12: + +$$ +t \approx 1 + k \cdot \log F \cdot \Delta F. \tag {12} +$$ + +In Eq 12, we consolidate all constants from Eq 11 into $k$ , where $k \in (0,1]$ , as the maximum value of luminous flux is normalized to 1. The proof is provided in the appendix. + +# 3.3. Proposed algorithm + +The numerical method for computing luminous flux, denoted as $F$ , leverages its physical definition: the ratio of light intensity over a given area to the area itself. We implement a sliding window approach, traversing the entire image. For each window, we calculate the mean intensity and subsequently divide by the window's area to compute the numerical value of the luminous flux for each pixel. In this context, $\Omega$ represents the sliding window, $\Omega(\cdot)$ is the mean value of the samplings within the window, and $S_{\Omega}$ is the area of the window. Denoted as Eq 13: + +$$ +F = \sum \Omega (\mathbf {I}) / S _ {\Omega}. \tag {13} +$$ + +For numerical computation, we sample the neighborhoods of each pixel by using sliding window. Consequently, for each pixel, the $\Delta F$ can be expressed as shown in Eq 14: + +$$ +\Delta F = \frac {\min \Omega (F)}{\max \Omega (F)} \tag {14} +$$ + +We use the extreme ratio within the sliding window to amplify $\Delta F$ because the difference between adjacent pixels is close to 0. However, $\Delta F$ might be affected by noise. To mitigate this, we first truncate $\Omega(F)$ within the range [0.005, 0.95]. Following this, it may be necessary to truncate the transmission $t$ during subsequent processing to ensure stability and accuracy in our results. + +In recovery process, we define a clear image $\mathbf{J}$ using an image formulation model from Eq 1. This is expressed as follows in Eq 15: + +$$ +\mathbf {J} (x) = \frac {\mathbf {I} (x) + \mathbf {A} \omega (t (x) - 1)}{\operatorname* {m a x} (\omega t (x) , t _ {0})}. \tag {15} +$$ + +In Eq 15, $\omega$ is a constant relaxation parameter which is constrained to be in the range $(0,1]$ , and $t_0$ , set as 0.05, is a lower bound that provides stability during computation. + +The global atmospheric light, represented by $\mathbf{A}$ , plays a crucial role in our model. We estimate $\mathbf{A}$ swiftly using a + +Algorithm 1 Pseudo-code for LFP + +1: Setup: constant $k = 0.5$ and window size $\Omega = 7$ +2: Input: Degraded image $I$ +3: while $\Omega$ in $I$ do +4: Compute mean intensity: $I_{\Omega} = \sum \Omega (I) / S_{\Omega}$ +5: Compute luminous flux: $F = I_{\Omega} / S_{\Omega}$ +6: Compute rate of change: $\Delta F = \frac{\min\Omega(F)}{\max\Omega(F)}$ +7: end while +8: $t\approx 1 + k\cdot \log F\cdot \Delta F$ Compute the transmission +9: guided filtering on transmission +10: Output: $J = \frac{I + A \cdot (t - 1)}{t}$ + +two-step approach. Initially, $\mathbf{A}$ is estimated using mean filtering and the maximum grayscale value of the original image. This gives us an initial approximation of the global atmospheric light. Subsequently, the image is segmented into small regions, and we select the maximum value from each region to represent the $\mathbf{A}$ for that particular region. The final estimate of $\mathbf{A}$ is then calculated as the average of all the $\mathbf{A}$ values from these small regions. This method allows us to get a more refined and accurate estimate of the global atmospheric light. The complete process of our scene recovery method is thoroughly depicted in Fig 2, and the corresponding pseudo-code is provided in Algorithm 1. In terms of computational complexity, our algorithm starts off with a complexity of $\mathcal{O}(N)$ as we traverse all pixels in the image using a sliding window. Then, it rises by $\mathcal{O}(\log N)$ when we compute the maximum and minimum values for each window. Finally, the total computational complexity for the subsequent steps, which include the calculation of transmission according to Eq 11 and application of a guided filter [17] for smoothing, is $\mathcal{O}(N\log N)$ , where $N$ denotes the size of a single image. This complexity analysis demonstrates that our method is efficient and feasible for large scale image recovery tasks. + +![](images/acafea452d1b56b835d5b38a6164235a83588187bba4da843608a48bfabccd7a.jpg) +Figure 3. Visual comparisons of different dehazing methods on outdoor and indoor images. The algorithms designed for water removal (blue) and sand removal (yellow-brown) fail in haze (grey) environments. Our method effectively restores details in the sky region without introducing artifacts. It also improves the visual quality of indoor images by enhancing brightness while preserving the structure. + +# 4. Experiments + +# 4.1. Implementation Details + +Our method relies solely on two parameters: $k$ and the sliding window $\Omega$ . And uses a non-parametric approach for the estimation of transmission $t$ , enhancing visual results under a variety of visibility conditions. Based on empirical experiences, we have determined the optimal settings to be $k = 0.5$ and $\Omega = 7$ . All experiments were conducted on a machine with an AMD Ryzen 73750H CPU@2.3GHz and a GeForce RTX 2080 GPU (11GB). The deep learning models were built with the PyTorch, TensorFlow and Matlab frameworks. The test images are publicly available. + +# 4.2. Experimental Settings + +We evaluated the LFP's performance using various methods and datasets, including physical models and deep learning models. Selected datasets spanned varied environments and scattering degrees, including RESIDE [24], SIRB [46], UIEBD [26], Haze4K [32], and RS-Haze [24]. These encompassed synthetic to real-world scenes, ensuring broad evaluation robustness. We used Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity Index Measure (SSIM) metrics for objective and fair evaluations across scenarios and methods. + +# 4.3. Qualitative Comparison + +Image Dehazing. We use RESIDE dataset [24] to validate the efficacy and robustness of LFP, as is shown in Fig 3, by comparing it with $\mathrm{Dive + }$ and HRDCP. While $\mathrm{Dive + }$ fails to restore image clarity, HRDCP results in color deviation, likely due to the specific absorption and scattering properties of sand particles. In contrast, our LFP method can generate more natural-looking results with better structures in a more stable manner. This demonstrates that images in hazy and underwater environments are formed based on different optical properties. We then compare the performance of the LFP with different methods like MSCNN, gUNet, and DCP. Learning-based methods such as gUNet and MSCNN are susceptible to dataset bias, which makes it difficult to + +enhance visual quality. Meanwhile, traditional methods like DCP, which rely on hand-crafted priors, struggle with varied real-world conditions. LFP adeptly overcomes the major obstacle faced by previous dehazing approaches, which is the problematic influence of scene depth. The inherent uncertainty in depth estimation often leads to unstable performance. By confidently deriving the relationship between $\Delta F$ and $t$ without involving $d$ , LFP successfully avoids the dataset bias that learning-based methods encounter. LFP effectively bypasses complex parameters of transmission and offers reliable solution with robust theoretical guarantees. Our method demonstrates a significantly stronger capability for robust image restoration in hazy weather conditions, resulting in highly satisfactory visual results. + +Underwater Image Enhancement. We validate the efficacy and robustness of LFP for underwater image enhancement through comprehensive experiments, using UIEBD dataset [26]. Results are shown in Fig 4. We first compare it with DCP and HRDCP, which are methodologies designed for different weather conditions. DCP fails completely, and although HRDCP achieves some visual effects, it suffers from noticeable color distortion. These experimental results underscore the necessity for image processing techniques to be specifically designed and optimized for each environment, reinforcing that an algorithm effective in one environment can't simply be applied to another without adaptation. $\mathrm{Dive + }$ and UWCNN struggle to effectively eliminate the green light prevalent in underwater scenes, leading to undesired color shifts. Although Waternet and ROP perform well in color restoration, they both suffer from dimmed lighting and reduced contrast, thereby compromising the overall visual quality. LFP overcomes the major obstacle faced by previous underwater image enhancement approaches, which is the varying influence of the attenuation coefficient $\beta$ on the RGB color channels that often results in color bias. By eliminating the attenuation coefficient $\beta$ during the derivation process, LFP avoids the color bias, rendering it a reliable solution with robust theoretical guarantees. Therefore, LFP method demonstrates competitive color correction and contrast enhancement capabilities. + +![](images/749b5781291848e2186314ec964d16b18c9a9420a7ba8c74a490ece395cc85bf.jpg) +Figure 4. Comparison of different methods for underwater conditions. The algorithms designed for haze removal (grey) and sand removal (brown) fail in underwater (blue) environments. Our method corrects color shifts, enhance image brightness and improve visibility. + +![](images/22b3c6d596c818426b2167537e6aa81c6ff38a2fa3a2cf8520fa08b995c7edcb.jpg) +Figure 5. Comparison of different methods for sandstorm image enhancement. The algorithms designed for haze removal (grey) and water removal (blue) fail in sandstorm (brown) environments. Our method corrects color shifts, enhances image brightness, and improves contrast, resulting in more vibrant color reproduction. Moreover, LFP produces a more realistic sky area closer to real scenes. + +Visibility Restoration in Sandstorms. We assess our LFP method's effectiveness in restoring visibility during sandstorms, by using SIRB dataset [46]. Results are shown in Fig 5. Initially, we compare LFP's performance with DCP, MSCNN, and $\mathrm{Dive + }$ . The first two fail in dust and sandstorm mitigation, and $\mathrm{Dive + }$ introduces noticeable color deviations in sky areas. This emphasizes that hazy and sandstorm images are influenced by distinct optical properties, making dehaze or dewater algorithms unsuitable for sandstorms. Next, we contrast our method with CVC, HRDCP, and ROP. All three significantly enhance visibility in sandstorms. However, CVC lacks contrast, leading to detail loss. HRDCP, despite its visual improvements, exhibits a dim sky and color distortions due to its DCP-based design. ROP performs well but struggles with color distortion, resulting in a cool appearance. LFP overcomes prior challenges by addressing scene depth and attenuation coefficient $\beta$ . By eliminating hard-to-estimate parameters like atmospheric light and scene depth, LFP avoids common issues, making it robust in sandstorm conditions without relying on assumptions. Thus, our LFP method excels in sandstorm image restoration, producing visually pleasing results without color distortion. + +In summary, the LFP has the ability to restore scenes in a variety of severe weather conditions, a characteristic that is absent in most algorithms. By eliminating the physical parameters that impact transmittance, LFP effectively bypasses the influence of scene depth, thus enhancing visual quality. Moreover, it minimizes color deviations by removing the attenuation coefficient. This approach provides a theoretically and empirically supported a priori method for scene restoration algorithms. + +# 4.4. Quantitative Comparison + +Dehazing Quality We selected DCP, BCCR [34], and FVR [51] as our benchmarks to represent different types of prior-based dehazing methodologies. The rationale behind this selection, instead of opting for deep learning techniques, is multi-fold. Traditional dehazing methods not only possess the capability to restore clear images, but they also notably improve visual effects, thereby enhancing the visual experience. Furthermore, LFP does not contain any learnable pa + +Table 2. This table succinctly presents the comparative results of the LFP method and three other distinct prior methods. The evaluation spans across all datasets and metrics, with LFP consistently achieving superior performance. The highest PSNR and SSIM values, represented by bold text, clearly demonstrate the effectiveness of our LFP method. + +
MethodsRESIDE-INRESIDE-OUTHaze-4KRS-Haze
PSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIM
DCP16.620.81819.130.81514.010.76017.860.734
BCCR16.880.79114.020.77713.780.75618.840.747
FVR15.720.74811.610.66812.650.68816.120.626
LFP19.760.86921.050.86118.290.87318.240.786
+ +rameters, differentiating it from deep learning approaches. As illustrated in Tab 2, demonstrates that LFP outperforms other methods across all four of our test datasets. + +The performance is consistent whether we evaluate in terms of PSNR or SSIM. This robust evidence strongly validates the high generalization ability of the LFP method. Furthermore, the LFP method excels in restoring clear, detailed imagery, as particularly demonstrated by the significant improvement in SSIM scores. For more detailed results on desanding and dewatering, please refer to the appendix. + +Runtime Efficiency We conducted extensive comparisons on both CPU and GPU platforms. On the CPU, we compared LFP with traditional dehazing methods; on the GPU, we chose a variety of deep learning dehazing models, including: MSCNN, AOD-Net [23], GFN [41], GCANet [4], GDNet [30], MSBDN [6], FFA [39], CycleD [9], DehazeF [47], and gUNet-T. In terms of runtime, Tab 3 compares the efficiency on both CPU and GPU. The results show that the LFP method consistently achieves the fastest running speed, surpassing existing algorithms. + +CPU Performance. As evidenced by the data presented in Tab 3, the LFP model consistently exhibits superior runtime performance across all tested resolutions. With a runtime of merely 0.007 seconds at the lower-end resolution of $360\mathrm{p}$ , and a still impressive 0.863 seconds at the high-end 4k resolution, these empirical findings underscore the remarkable efficiency of the LFP model when processing large-scale images or performing real-time scene recovery. + +Table 3. Run-time(seconds) performance. Testing the efficiency of different methods on various single image sizes. Models marked with (*) indicate that the results for those models were generated using MATLAB. The best performance is denoted in red, while the second best performance is indicated in blue. + +
CPU Performance
Model360p480p720p1080g2k4k
DCP[16]0.4590.7651.762.6006.7415.11
Retinex[22]0.1490.2860.4780.8042.3675.421
LDCP[55]0.5170.8212.2053.0918.26418.09
CAP[54]0.2180.3100.6661.0212.8666.403
BCCR[34]1.6913.5716.41611.7029.6064.26
FVR[51]0.3013.9028.58912.5834.3576.50
ROP*[27]0.0670.1440.4111.3992.0732.515
LFP0.0070.0170.0730.1510.4070.863
GPU Performance
MSCNN[40]0.1150.1470.2660.4621.1142.481
DehazeNet[3]0.1650.2230.4050.6211.3742.951
AOD-Net[23]0.0760.1060.1750.2910.6961.519
GFN*[41]0.6600.8581.4942.4456.35313.09
GCANet[4]0.4460.6411.1071.8214.2629.266
GDNet[30]0.6440.7421.5032.0274.67510.48
MSBDN[6]0.9521.2122.1943.3358.20317.46
FFA[39]1.6752.2533.9605.97714.2730.61
CycleD[9]0.6070.8101.4282.1585.15511.03
DehazeF[47]0.2890.3870.6771.0322.4755.310
gUNet-T[47]0.1970.2660.4730.7161.7193.669
ROP*[27]0.0470.0510.1130.2090.5320.894
LFP0.0010.0010.0020.0030.0080.019
+ +GPU Performance. Tab 3 shows that LFPsignificantly outperforms all the compared algorithms in processing speed, especially for high-resolution images. For instance, at a resolution of 4K, LFP's average runtime is merely 0.019 seconds, whereas the runtimes of other algorithms range from 0.894 seconds to a staggering 30.609 seconds. This results in a speedup of up to 34 times for LFP, marking its superior computational efficiency. + +# 4.5. Ablation Study on $k$ and $\Omega$ + +Our ablation study further reveals that the changes in the $k$ parameter not only influence the luminous flux rate of change $\Delta F$ but also directly affect the transmission estimation. A properly tuned $k$ facilitates an accurate estimation of the transmission map, thus effectively reducing the impact of different weather conditions on the image. On the other hand, the size of the sliding window $\Omega$ directly affects the local variance $\Delta F$ . A larger $\Omega$ captures more significant changes in the luminous flux, which is particularly effective in dense haze conditions or when the scene has a greater depth of view. This broadened perspective allows for a more comprehensive understanding of the scene, lead + +![](images/96f687e52cf63cb58d7a015b6f8d6e83b5e1864b8c257177420c6948036ecc05.jpg) +Figure 6. The x-axis represents $k$ , the y-axis represents $\Omega$ . When the depth of view is further, or the haze is denser, $\Omega$ and $k$ need to be increased correspondingly. + +ing to a higher-quality image after enhancement. Finally, the interplay between $k$ and $\Omega$ is also crucial. Balancing these two factors allows for a fine-tuned approach to image enhancement, taking into account both the overall brightness and the local variance of the image, this ability to cope with complex conditions adds another layer of robustness to our technique. This balance ensures that the final output image is visually pleasing while maintaining the fidelity of the original scene. + +# 5. Conclusion + +In this work, we propose an efficient scene recovery algorithm to restore clear images under poor visibility conditions. By analyzing the image formation model, we derive an analytical form for transmission that eliminates multiple influencing factors and leverages the correlation with flux change rate as prior knowledge. Our approach achieves state-of-the-art performance while maintaining low computational complexity, enabling real-time processing on resource-limited devices. Extensive experiments validate the effectiveness of the proposed method. The novel integration of prior knowledge and analytical transmission modeling provides opportunities to inform future learning-based methods. Overall, this paper makes significant contributions towards robust scene recovery under challenging imaging conditions. + +# 6. Acknowledgments + +This work was supported by the National Natural Science Foundation of China (No. 62336001, No. 62072014) + +# References + +[1] Dana Berman, Shai Avidan, et al. Non-local image dehazing. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1674-1682, 2016. 1 +[2] Trung Minh Bui and Wonha Kim. Single image dehazing using color ellipsoid prior. IEEE Transactions on Image Processing, 27(2):999-1009, 2017. 1 +[3] Bolun Cai, Xiangmin Xu, Kui Jia, Chunmei Qing, and Dacheng Tao. Dehazenet: An end-to-end system for single image haze removal. IEEE Transactions on Image Processing, 25(11):5187-5198, 2016. 1, 3, 8 +[4] Dongdong Chen, Mingming He, Qingnan Fan, Jing Liao, Liheng Zhang, Dongdong Hou, Lu Yuan, and Gang Hua. Gated context aggregation network for image dehazing and deraining. In 2019 IEEE winter conference on applications of computer vision (WACV), pages 1375-1383. IEEE, 2019. 7, 8 +[5] Qili Deng, Ziling Huang, Chung-Chi Tsai, and Chia-Wen Lin. Hardgan: A haze-aware representation distillation gan for single image dehazing. In European Conference on Computer Vision, pages 722-738. Springer, 2020. 3 +[6] Hang Dong, Jinshan Pan, Lei Xiang, Zhe Hu, Xinyi Zhang, Fei Wang, and Ming-Hsuan Yang. Multi-scale boosted dehazing network with dense feature fusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2157-2167, 2020. 3, 7, 8 +[7] Jiangxin Dong and Jinshan Pan. Physics-based feature dehazing networks. In European Conference on Computer Vision, pages 188-204. Springer, 2020. 3 +[8] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 3 +[9] Deniz Engin, Anil Genç, and Hazim Kemal Ekenel. Cycle-dehaze: Enhanced cyclegan for single image dehazing. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition Workshops, pages 825-833, 2018. 7, 8 +[10] Faming Fang, Fang Li, and Tieyong Zeng. Single image dehazing and denoising: a fast variational approach. SIAM Journal on Imaging Sciences, 7(2):969-996, 2014. 1 +[11] Xueyang Fu, Yue Huang, Delu Zeng, Xiao-Ping Zhang, and Xinghao Ding. A fusion-based enhancing approach for single sandstorm image. In 2014 IEEE 16th international workshop on multimedia signal processing (MMSP), pages 1-5. IEEE, 2014. 1 +[12] Han Gao, Ping Wei, and Jun Ke. Color enhancement and image defogging in hsi based on retinax model. In 2015 International Conference on Optical Instruments and Technology: Optoelectronic Imaging and Processing Technology, pages 9-16. SPIE, 2015. 2 +[13] Jie Gui, Xiaofeng Cong, Yuan Cao, Wenqi Ren, Jun Zhang, Jing Zhang, Jieuxin Cao, and Dacheng Tao. A comprehensive survey and taxonomy on single image dehazing based on deep learning. ACM Computing Surveys, 55(13s):1-37, 2023. 1 + +[14] Chun-Le Guo, Qixin Yan, Saeed Anwar, Runmin Cong, Wenqi Ren, and Chongyi Li. Image dehazing transformer with transmission-aware 3d position embedding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5812-5820, 2022. 3 +[15] Yecai Guo, Hanyu Li, and Peixian Zhuang. Underwater image enhancement using a multiscale dense generative adversarial network. IEEE Journal of Oceanic Engineering, 45(3):862-870, 2019. 3 +[16] Kaiming He, Jian Sun, and Xiaou Tang. Single image haze removal using dark channel prior. IEEE transactions on pattern analysis and machine intelligence, 33(12):2341-2353, 2010. 1, 2, 8 +[17] Kaiming He, Jian Sun, and Xiaou Tang. Guided image filtering. IEEE transactions on pattern analysis and machine intelligence, 35(6):1397-1409, 2012. 5 +[18] Jong-Ju Jeon, Tae-Hee Park, and Il-Kyu Eom. Sand-dust image enhancement using chromatic variance consistency and gamma correction-based dehazing. Sensors, 22(23):9048, 2022. 2 +[19] Mingye Ju, Can Ding, Y Jay Guo, and Dengyin Zhang. Idgcp: Image dehazing based on gamma correction prior. IEEE Transactions on Image Processing, 29:3104-3118, 2019. 1 +[20] Manjit Kaur, Dilbag Singh, Vijay Kumar, and Kehui Sun. Color image dehazing using gradient channel prior and guided 10 filter. Information Sciences, 521:326-342, 2020. 1 +[21] Johannes Kopf, Boris Neubert, Billy Chen, Michael Cohen, Daniel Cohen-Or, Oliver Deussen, Matt Uytendaele, and Dani Lischinski. Deep photo: Model-based photograph enhancement and viewing. ACM transactions on graphics (TOG), 27(5):1-10, 2008. 4 +[22] Edwin H Land. The retina theory of color vision. Scientific american, 237(6):108-129, 1977. 4, 8 +[23] Boyi Li, Xiulian Peng, Zhangyang Wang, Jizheng Xu, and Dan Feng. Aod-net: All-in-one dehazing network. In Proceedings of the IEEE international conference on computer vision, pages 4770-4778, 2017. 1, 3, 7, 8 +[24] Boyi Li, Wenqi Ren, Dengpan Fu, Dacheng Tao, Dan Feng, Wenjun Zeng, and Zhangyang Wang. Benchmarking single-image dehazing and beyond. IEEE Transactions on Image Processing, 28(1):492-505, 2019. 3, 6 +[25] Chongyi Li, Chunle Guo, Wenqi Ren, Runmin Cong, Junhui Hou, Sam Kwong, and Dacheng Tao. An underwater image enhancement benchmark dataset and beyond. IEEE Transactions on Image Processing, 29:4376-4389, 2019. 2 +[26] Chongyi Li, Saeed Anwar, and Fatih Porikli. Underwater scene prior inspired deep underwater image and video enhancement. Pattern Recognition, 98:107038, 2020. 1, 3, 6 +[27] Jun Liu, Wen Liu, Jianing Sun, and Tieyong Zeng. Rank-one prior: Toward real-time scene recovery. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14802-14810, 2021. 1, 2, 8 +[28] Risheng Liu, Xin Fan, Ming Zhu, Minjun Hou, and Zhongxuan Luo. Real-world underwater enhancement: Challenges, benchmarks, and solutions under natural light. IEEE transactions on circuits and systems for video technology, 30(12): 4861-4875, 2020. 3 + +[29] Xiaodong Liu, Zhi Gao, and Ben M Chen. Mlfcgan: Multilevel feature fusion-based conditional gan for underwater image color correction. IEEE Geoscience and Remote Sensing Letters, 17(9):1488-1492, 2019. 3 +[30] Xiaohong Liu, Yongrui Ma, Zhihao Shi, and Jun Chen. Griddehazenet: Attention-based multi-scale network for image dehazing. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7314-7323, 2019. 3, 7, 8 +[31] Yu Liu, Guanlong Zhao, Boyuan Gong, Yang Li, Ritu Raj, Niraj Goel, Satya Kesav, Sandeep Gottimukkala, Zhangyang Wang, Wenqi Ren, et al. Improved techniques for learning to dehaze and beyond: A collective study. arXiv preprint arXiv:1807.00202, 2018. 1 +[32] Ye Liu, Lei Zhu, Shunda Pei, Huazhu Fu, Jing Qin, Qing Zhang, Liang Wan, and Wei Feng. From synthetic to real: Image dehazing collaborating with unlabeled real data. In Proceedings of the 29th ACM International Conference on Multimedia, pages 50-58, 2021. 6 +[33] Rastislav Lukac. Single-sensor imaging: methods and applications for digital cameras. CRC Press, 2018. 2 +[34] Gaofeng Meng, Ying Wang, Jiangyong Duan, Shiming Xiang, and Chunhong Pan. Efficient image dehazing with boundary constraint and contextual regularization. In Proceedings of the IEEE international conference on computer vision, pages 617-624, 2013. 2, 7, 8 +[35] WE Knowles Middleton. Bouguer, lambert, and the theory of horizontal visibility. Isis, 51(2):145-149, 1960. 4 +[36] Srinivasa G Narasimhan and Shree K Nayar. Vision and the atmosphere. International journal of computer vision, 48(3): 233-254, 2002. 1 +[37] Srinivasa G Narasimhan and Shree K Nayar. Interactive (de) weathering of an image using physical models. In IEEE Workshop on color and photometric Methods in computer Vision, page 1. France, 2003. 4 +[38] Ko Nishino, Louis Kratz, and Stephen Lombardi. Bayesian defogging. International journal of computer vision, 98(3): 263-278, 2012. 1 +[39] Xu Qin, Zhilin Wang, Yuanchao Bai, Xiaodong Xie, and Huizhu Jia. Ffa-net: Feature fusion attention network for single image dehazing. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 11908-11915, 2020. 3, 7, 8 +[40] Wenqi Ren, Si Liu, Hua Zhang, Jinshan Pan, Xiaochun Cao, and Ming-Hsuan Yang. Single image dehazing via multiscale convolutional neural networks. In European conference on computer vision, pages 154-169. Springer, 2016. 1, 2, 8 +[41] Wenqi Ren, Lin Ma, Jiawei Zhang, Jinshan Pan, Xiaochun Cao, Wei Liu, and Ming-Hsuan Yang. Gated fusion network for single image dehazing. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 3253-3261, 2018. 7, 8 +[42] Wenqi Ren, Jingang Zhang, Xiangyu Xu, Lin Ma, Xiaochun Cao, Gaofeng Meng, and Wei Liu. Deep video dehazing with semantic segmentation. IEEE Transactions on Image Processing, 28(4):1895-1908, 2018. 1 +[43] Christos Sakaridis, Dengxin Dai, Simon Hecker, and Luc Van Gool. Model adaptation with synthetic and real data for + +semantic dense foggy scene understanding. In Proceedings of the European Conference on Computer Vision (ECCV), pages 687-704, 2018. 1 +[44] Yoav Y Schechner and Yuval Averbuch. Regularized image recovery in scattering media. IEEE Transactions on Pattern Analysis and Machine Intelligence, 29(9):1655-1660, 2007. 1 +[45] Zhenghao Shi, Yaning Feng, Minghua Zhao, Erhu Zhang, and Lifeng He. Let you see in sand dust weather: A method based on halo-reduced dark channel prior dehazing for sand-dust image enhancement. *Ieee Access*, 7:116722–116733, 2019. 2 +[46] Yazhong Si, Fan Yang, Ya Guo, Wei Zhang, and Yipu Yang. A comprehensive benchmark analysis for sand dust image reconstruction. Journal of Visual Communication and Image Representation, 89:103638, 2022. 3, 6, 7 +[47] Yuda Song, Zhuqing He, Hui Qian, and Xin Du. Vision transformers for single image dehazing. arXiv preprint arXiv:2204.03883, 2022. 3, 7, 8 +[48] Yuda Song, Yang Zhou, Hui Qian, and Xin Du. Rethinking performance gains in image dehazing networks. arXiv preprint arXiv:2209.11448, 2022. 1, 3 +[49] Matan Sulami, Itamar Glatzer, Raanan Fattal, and Mike Werman. Automatic recovery of the atmospheric light in hazy images. In 2014 IEEE International Conference on Computational Photography (ICCP), pages 1-11. IEEE, 2014. 1 +[50] Robby T Tan. Visibility in bad weather from a single image. In 2008 IEEE Conference on Computer Vision and Pattern Recognition, pages 1-8. IEEE, 2008. 1 +[51] Jean-Philippe Tarel and Nicolas Hautiere. Fast visibility restoration from a single color or gray level image. In 2009 IEEE 12th International Conference on Computer Vision, pages 2201–2208. IEEE, 2009. 2, 7, 8 +[52] Ting Yan, Liejun Wang, and Jiaxing Wang. Method to enhance degraded image in dust environment. J. Softw., 9(10): 2672-2677, 2014. 2 +[53] Dong Zhao, Jia Li, Hongyu Li, and Long Xu. Hybrid local-global transformer for image dehazing. arXiv preprint arXiv:2109.07100, 2021. 3 +[54] Qingsong Zhu, Jiaming Mai, and Ling Shao. A fast single image haze removal algorithm using color attenuation prior. IEEE transactions on image processing, 24(11):3522-3533, 2015. 1, 2, 8 +[55] Yingying Zhu, Gaoyang Tang, Xiaoyan Zhang, Jianmin Jiang, and Qi Tian. Haze removal method for natural restoration of images with sky. neurocomputing, 275:499-510, 2018. 2, 8 \ No newline at end of file diff --git a/2024/Efficient Scene Recovery Using Luminous Flux Prior/images.zip b/2024/Efficient Scene Recovery Using Luminous Flux Prior/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..ab48e0e3495287adf8b5ea0f39bcc890f05dc145 --- /dev/null +++ b/2024/Efficient Scene Recovery Using Luminous Flux Prior/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7f1d810890b264b392d09f15afb680a4976a1a1c810d600d8ff06406708faf7d +size 733522 diff --git a/2024/Efficient Scene Recovery Using Luminous Flux Prior/layout.json b/2024/Efficient Scene Recovery Using Luminous Flux Prior/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..a6b26d215d44e9337194c8d361b0980df5aa83c7 --- /dev/null +++ b/2024/Efficient Scene Recovery Using Luminous Flux Prior/layout.json @@ -0,0 +1,10554 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 135, + 103, + 459, + 120 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 103, + 459, + 120 + ], + "spans": [ + { + "bbox": [ + 135, + 103, + 459, + 120 + ], + "type": "text", + "content": "Efficient Scene Recovery Using Luminous Flux Prior" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 146, + 142, + 447, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 142, + 447, + 185 + ], + "spans": [ + { + "bbox": [ + 146, + 142, + 447, + 185 + ], + "type": "text", + "content": "Zhongyu Li1, Lei Zhang1* \n1University of Science and Technology of China, Hefei, China \nruin2@mail.ustc.edu.cn, leizh23@ustc.edu.cn," + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "spans": [ + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 239, + 289, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 239, + 289, + 502 + ], + "spans": [ + { + "bbox": [ + 46, + 239, + 289, + 502 + ], + "type": "text", + "content": "Scene recovery, the restoration of images degraded by adverse weather conditions, presents significant challenges for existing methods. Physical models, constrained by their inherent assumptions, often fail when these assumptions are not met; Deep learning models are powerful, they are limited by the diversity of their training datasets, leading to poor generalization and high computational demands. To address these limitations, we propose the Luminous Flux Prior (LFP), to recover degraded images under diverse adverse weather without learning. Luminous flux, a physical measure that reflects image brightness, has a rate of change that demonstrates a significant correlation with transmission. Consequently, we leverage this rate of change in luminous flux as prior knowledge to estimate transmission, which in turn assists in image recovery. This approach reduces dependency on physical parameters and enhances adaptability to various weather. Experimental validation under diverse conditions, such as sandstorms, underwater environments, and haze, attests to the robustness of LFP in restoring clear images. With a time complexity of " + }, + { + "bbox": [ + 46, + 239, + 289, + 502 + ], + "type": "inline_equation", + "content": "\\mathcal{O}(N\\log N)" + }, + { + "bbox": [ + 46, + 239, + 289, + 502 + ], + "type": "text", + "content": ", LFP enables real-time recovery, making it a suitable for devices with limited computational resources." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 516, + 128, + 529 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 516, + 128, + 529 + ], + "spans": [ + { + "bbox": [ + 47, + 516, + 128, + 529 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 536, + 287, + 656 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 536, + 287, + 656 + ], + "spans": [ + { + "bbox": [ + 46, + 536, + 287, + 656 + ], + "type": "text", + "content": "Observational devices encounter significant challenges under adverse lighting conditions, which can be caused by phenomena such as sandstorms, underwater environments, and haze. These conditions lead to image degradation, typified by a loss of detail, contrast distortion, and color bias [13, 50]. Such degradation impairs computer vision tasks, including object detection [31], image segmentation [42], and scene understanding [43]. Therefore, restoring clear images from these degraded observations becomes a critical challenge in the field of computer vision [27]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 657, + 287, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 657, + 287, + 693 + ], + "spans": [ + { + "bbox": [ + 47, + 657, + 287, + 693 + ], + "type": "text", + "content": "Image degradation primarily arises from the scattering and absorption of light [10, 11, 16, 26, 49]. Environmental elements, such as the atmosphere and water, scatter" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 213, + 547, + 309 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 213, + 547, + 309 + ], + "spans": [ + { + "bbox": [ + 304, + 213, + 547, + 309 + ], + "type": "text", + "content": "light sources, engendering intricate environmental light and the absorption of parts of the visible light spectrum. This process significantly degrades images by diminishing their brightness. The physical quantity symbolizing the brightness per unit area, referred to as luminous flux, is denoted by " + }, + { + "bbox": [ + 304, + 213, + 547, + 309 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 304, + 213, + 547, + 309 + ], + "type": "text", + "content": ". The fundamental instigator of this degradation is a decrease in the transmission rate " + }, + { + "bbox": [ + 304, + 213, + 547, + 309 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 213, + 547, + 309 + ], + "type": "text", + "content": ", representing the proportion of light brightness received by the camera." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 310, + 547, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 310, + 547, + 574 + ], + "spans": [ + { + "bbox": [ + 304, + 310, + 547, + 574 + ], + "type": "text", + "content": "The challenge in image restoration is the estimation of transmission " + }, + { + "bbox": [ + 304, + 310, + 547, + 574 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 310, + 547, + 574 + ], + "type": "text", + "content": ". We bifurcate scene restoration methods into two principal types based on the calculation of " + }, + { + "bbox": [ + 304, + 310, + 547, + 574 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 310, + 547, + 574 + ], + "type": "text", + "content": ": physical models and data-driven deep learning methods. Physical models [1, 2, 16, 19, 20, 38, 54] make assumptions about the physical parameters or conditions for transmission " + }, + { + "bbox": [ + 304, + 310, + 547, + 574 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 310, + 547, + 574 + ], + "type": "text", + "content": " to be valid. However, if these assumptions fall short, the restored image suffers to varying extents. Conversely, data-driven deep learning methods extrapolate an estimation model for haze-free images from extensive datasets [3, 23, 40]. Yet, these supervised learning methods exhibit limited efficacy in real-world scenarios due to the scarcity of large-scale, diverse, and fully aligned paired training data. Additionally, deep learning algorithms necessitate complex loss functions and training strategies, resulting in high training costs [48]. Most existing methods rely solely on L1/L2 image reconstruction loss and lack robust regularization, which may induce unsuitable gradient updates and grapple with the uncertainty of ill-posed problems due to a dearth of robust prior information. Both physical models and deep learning methods find it challenging to adapt to a variety of different scenes owing to their inherent limitations [27]." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 574, + 547, + 609 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 574, + 547, + 609 + ], + "spans": [ + { + "bbox": [ + 304, + 574, + 547, + 609 + ], + "type": "text", + "content": "Typical physical models illustrate the image formation process impacted by scattering [44]. This model can generally be represented by the following equation [36, 50]:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 353, + 622, + 545, + 635 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 353, + 622, + 545, + 635 + ], + "spans": [ + { + "bbox": [ + 353, + 622, + 545, + 635 + ], + "type": "interline_equation", + "content": "\\mathbf {I} (x) = \\mathbf {J} (x) \\cdot t (x) + \\mathbf {A} (1 - t (x)), \\tag {1}", + "image_path": "1ef959b145ae870f2eca0341956fc957b86e7a1296216384a1d1a5e88f559c47.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 641, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 641, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 641, + 547, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 641, + 547, + 714 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 304, + 641, + 547, + 714 + ], + "type": "text", + "content": " denotes the location information within the image, " + }, + { + "bbox": [ + 304, + 641, + 547, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{I}(x)" + }, + { + "bbox": [ + 304, + 641, + 547, + 714 + ], + "type": "text", + "content": " refers to the observed degraded image, " + }, + { + "bbox": [ + 304, + 641, + 547, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{J}(x)" + }, + { + "bbox": [ + 304, + 641, + 547, + 714 + ], + "type": "text", + "content": " signifies the clear image (scene radiance) we aim to restore, and " + }, + { + "bbox": [ + 304, + 641, + 547, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{A}" + }, + { + "bbox": [ + 304, + 641, + 547, + 714 + ], + "type": "text", + "content": " symbolizes the global ambient light. " + }, + { + "bbox": [ + 304, + 641, + 547, + 714 + ], + "type": "inline_equation", + "content": "t(x)" + }, + { + "bbox": [ + 304, + 641, + 547, + 714 + ], + "type": "text", + "content": " describes the proportion of light reaching the camera. To restore a clear image " + }, + { + "bbox": [ + 304, + 641, + 547, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{J}(x)" + }, + { + "bbox": [ + 304, + 641, + 547, + 714 + ], + "type": "text", + "content": ", we need to estimate the transmission " + }, + { + "bbox": [ + 304, + 641, + 547, + 714 + ], + "type": "inline_equation", + "content": "t(x)" + }, + { + "bbox": [ + 304, + 641, + 547, + 714 + ], + "type": "text", + "content": " and" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 702, + 189, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 702, + 189, + 712 + ], + "spans": [ + { + "bbox": [ + 58, + 702, + 189, + 712 + ], + "type": "text", + "content": "*Lei Zhang is the corresponding author." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2743" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 52, + 70, + 121, + 202 + ], + "blocks": [ + { + "bbox": [ + 52, + 70, + 121, + 202 + ], + "lines": [ + { + "bbox": [ + 52, + 70, + 121, + 202 + ], + "spans": [ + { + "bbox": [ + 52, + 70, + 121, + 202 + ], + "type": "image", + "image_path": "b44ffb7f4ee619adbbb03f00640e0649b13199f0da7b46db6cfc780785e1448d.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 82, + 203, + 92, + 212 + ], + "lines": [ + { + "bbox": [ + 82, + 203, + 92, + 212 + ], + "spans": [ + { + "bbox": [ + 82, + 203, + 92, + 212 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 122, + 70, + 191, + 203 + ], + "blocks": [ + { + "bbox": [ + 122, + 70, + 191, + 203 + ], + "lines": [ + { + "bbox": [ + 122, + 70, + 191, + 203 + ], + "spans": [ + { + "bbox": [ + 122, + 70, + 191, + 203 + ], + "type": "image", + "image_path": "aec446a9c0bf0e24348e389fb95ab6c5673e3e1742f7a9f13e1f205c83c75e0e.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 150, + 204, + 161, + 212 + ], + "lines": [ + { + "bbox": [ + 150, + 204, + 161, + 212 + ], + "spans": [ + { + "bbox": [ + 150, + 204, + 161, + 212 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 192, + 70, + 257, + 202 + ], + "blocks": [ + { + "bbox": [ + 192, + 70, + 257, + 202 + ], + "lines": [ + { + "bbox": [ + 192, + 70, + 257, + 202 + ], + "spans": [ + { + "bbox": [ + 192, + 70, + 257, + 202 + ], + "type": "image", + "image_path": "7651c8eca67c9c20a28419ad349cdba7054d9d53d6e7061f734c537dc06b6b19.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 220, + 203, + 230, + 212 + ], + "lines": [ + { + "bbox": [ + 220, + 203, + 230, + 212 + ], + "spans": [ + { + "bbox": [ + 220, + 203, + 230, + 212 + ], + "type": "text", + "content": "(c)" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 257, + 70, + 400, + 203 + ], + "blocks": [ + { + "bbox": [ + 257, + 70, + 400, + 203 + ], + "lines": [ + { + "bbox": [ + 257, + 70, + 400, + 203 + ], + "spans": [ + { + "bbox": [ + 257, + 70, + 400, + 203 + ], + "type": "image", + "image_path": "810751d257b6c0fd4ad051ee129986f7996ebe0364a978aff362a615bcf425af.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 290, + 204, + 300, + 212 + ], + "lines": [ + { + "bbox": [ + 290, + 204, + 300, + 212 + ], + "spans": [ + { + "bbox": [ + 290, + 204, + 300, + 212 + ], + "type": "text", + "content": "(d)" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 213, + 547, + 257 + ], + "lines": [ + { + "bbox": [ + 46, + 213, + 547, + 257 + ], + "spans": [ + { + "bbox": [ + 46, + 213, + 547, + 257 + ], + "type": "text", + "content": "Figure 1. Challenging real scene recovery using LFP. The first row depicts degraded images, while the second row presents the restored clear images. These images represent three adverse conditions: haze, sandstorm, and underwater environments. The restored images all exhibit superior visual effects. Notably, the color bias in images (b), (c), and (e) has been rectified; images (c) and (g) are free from halo artifacts; and the brightness in images (a), (b), (c), (d), and (f) has been supplemented and adjusted." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 400, + 71, + 467, + 202 + ], + "blocks": [ + { + "bbox": [ + 400, + 71, + 467, + 202 + ], + "lines": [ + { + "bbox": [ + 400, + 71, + 467, + 202 + ], + "spans": [ + { + "bbox": [ + 400, + 71, + 467, + 202 + ], + "type": "image", + "image_path": "79cdd6655033a93ca105738d6b8091ff4b7d72814c8d27a51784f1ee04e05694.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 429, + 204, + 438, + 212 + ], + "lines": [ + { + "bbox": [ + 429, + 204, + 438, + 212 + ], + "spans": [ + { + "bbox": [ + 429, + 204, + 438, + 212 + ], + "type": "text", + "content": "(f)" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 467, + 71, + 539, + 202 + ], + "blocks": [ + { + "bbox": [ + 467, + 71, + 539, + 202 + ], + "lines": [ + { + "bbox": [ + 467, + 71, + 539, + 202 + ], + "spans": [ + { + "bbox": [ + 467, + 71, + 539, + 202 + ], + "type": "image", + "image_path": "5a2787ff6b6525e0d82e2735e6c0dca5c50fb7e4e3f29c888b02e717dc4735f3.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 497, + 204, + 507, + 212 + ], + "lines": [ + { + "bbox": [ + 497, + 204, + 507, + 212 + ], + "spans": [ + { + "bbox": [ + 497, + 204, + 507, + 212 + ], + "type": "text", + "content": "(g)" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 268, + 211, + 280 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 268, + 211, + 280 + ], + "spans": [ + { + "bbox": [ + 47, + 268, + 211, + 280 + ], + "type": "text", + "content": "ambient light " + }, + { + "bbox": [ + 47, + 268, + 211, + 280 + ], + "type": "inline_equation", + "content": "\\mathbf{A}" + }, + { + "bbox": [ + 47, + 268, + 211, + 280 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 47, + 268, + 211, + 280 + ], + "type": "inline_equation", + "content": "t(x)" + }, + { + "bbox": [ + 47, + 268, + 211, + 280 + ], + "type": "text", + "content": " can be defined as:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 133, + 285, + 287, + 298 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 285, + 287, + 298 + ], + "spans": [ + { + "bbox": [ + 133, + 285, + 287, + 298 + ], + "type": "interline_equation", + "content": "t (x) = e ^ {- \\beta d (x)}, \\tag {2}", + "image_path": "20dff964f9151c0aec392389aabfbb5837f0e1c9432bb71e67dd2d8f8353edfb.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 46, + 304, + 287, + 340 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 304, + 287, + 340 + ], + "spans": [ + { + "bbox": [ + 46, + 304, + 287, + 340 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 304, + 287, + 340 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 46, + 304, + 287, + 340 + ], + "type": "text", + "content": " is the attenuation coefficient, indicating the degree of scattering in the environmental medium, and " + }, + { + "bbox": [ + 46, + 304, + 287, + 340 + ], + "type": "inline_equation", + "content": "d(x)" + }, + { + "bbox": [ + 46, + 304, + 287, + 340 + ], + "type": "text", + "content": " designates the scene depth [54]." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 46, + 341, + 287, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 341, + 287, + 495 + ], + "spans": [ + { + "bbox": [ + 46, + 341, + 287, + 495 + ], + "type": "text", + "content": "The complexity of estimating transmission lies in its dependence on multiple physical parameters [54]. Adverse weather conditions that diminish luminous flux " + }, + { + "bbox": [ + 46, + 341, + 287, + 495 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 46, + 341, + 287, + 495 + ], + "type": "text", + "content": " (detailed in Sec.3) result in image darkening, attributed to decreased " + }, + { + "bbox": [ + 46, + 341, + 287, + 495 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 341, + 287, + 495 + ], + "type": "text", + "content": ". Utilizing " + }, + { + "bbox": [ + 46, + 341, + 287, + 495 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 46, + 341, + 287, + 495 + ], + "type": "text", + "content": " directly to estimate " + }, + { + "bbox": [ + 46, + 341, + 287, + 495 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 341, + 287, + 495 + ], + "type": "text", + "content": " is impractical. We propose LFP that employs the rate of change of luminous flux, " + }, + { + "bbox": [ + 46, + 341, + 287, + 495 + ], + "type": "inline_equation", + "content": "\\Delta F" + }, + { + "bbox": [ + 46, + 341, + 287, + 495 + ], + "type": "text", + "content": " (detailed in Sec.3), as a prior to simplify the process. In extensive experimental validation, we found " + }, + { + "bbox": [ + 46, + 341, + 287, + 495 + ], + "type": "inline_equation", + "content": "\\Delta F" + }, + { + "bbox": [ + 46, + 341, + 287, + 495 + ], + "type": "text", + "content": " correlates significantly with " + }, + { + "bbox": [ + 46, + 341, + 287, + 495 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 341, + 287, + 495 + ], + "type": "text", + "content": ", thereby transforming " + }, + { + "bbox": [ + 46, + 341, + 287, + 495 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 341, + 287, + 495 + ], + "type": "text", + "content": " estimation into a " + }, + { + "bbox": [ + 46, + 341, + 287, + 495 + ], + "type": "inline_equation", + "content": "\\Delta F" + }, + { + "bbox": [ + 46, + 341, + 287, + 495 + ], + "type": "text", + "content": " computation. As a dimensionless quantity, " + }, + { + "bbox": [ + 46, + 341, + 287, + 495 + ], + "type": "inline_equation", + "content": "\\Delta F" + }, + { + "bbox": [ + 46, + 341, + 287, + 495 + ], + "type": "text", + "content": " exhibits wide applicability, underscoring the adaptability of LFP. LFP is computationally efficient and generates visually pleasing images, as demonstrated by real scenes in Fig 1." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 59, + 496, + 242, + 507 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 496, + 242, + 507 + ], + "spans": [ + { + "bbox": [ + 59, + 496, + 242, + 507 + ], + "type": "text", + "content": "In summary, our contributions are as follows:" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 47, + 508, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 21, + "blocks": [ + { + "bbox": [ + 47, + 508, + 287, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 508, + 287, + 578 + ], + "spans": [ + { + "bbox": [ + 47, + 508, + 287, + 578 + ], + "type": "text", + "content": "1. We introduce a new method to estimate transmission " + }, + { + "bbox": [ + 47, + 508, + 287, + 578 + ], + "type": "inline_equation", + "content": "(t)" + }, + { + "bbox": [ + 47, + 508, + 287, + 578 + ], + "type": "text", + "content": ", leveraging the rate of change in luminous flux " + }, + { + "bbox": [ + 47, + 508, + 287, + 578 + ], + "type": "inline_equation", + "content": "(\\Delta F)" + }, + { + "bbox": [ + 47, + 508, + 287, + 578 + ], + "type": "text", + "content": ". This approach is widely applicable across different weather and imaging conditions, and does not require any learnable parameters, thus eliminating the need for dataset construction." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 47, + 580, + 287, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 580, + 287, + 651 + ], + "spans": [ + { + "bbox": [ + 47, + 580, + 287, + 651 + ], + "type": "text", + "content": "2. Our algorithm has a low computational complexity, " + }, + { + "bbox": [ + 47, + 580, + 287, + 651 + ], + "type": "inline_equation", + "content": "O(N \\cdot \\log N)" + }, + { + "bbox": [ + 47, + 580, + 287, + 651 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 47, + 580, + 287, + 651 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 47, + 580, + 287, + 651 + ], + "type": "text", + "content": " is the size of a single image. With GPU acceleration, it is exceedingly fast, enhancing computation speeds by 7 to 50 times compared to CPU processing. This makes the algorithm suitable for devices with limited computational power." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 47, + 651, + 287, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 651, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 47, + 651, + 287, + 712 + ], + "type": "text", + "content": "3. We have transformed the estimation of " + }, + { + "bbox": [ + 47, + 651, + 287, + 712 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 47, + 651, + 287, + 712 + ], + "type": "text", + "content": " into a computation involving " + }, + { + "bbox": [ + 47, + 651, + 287, + 712 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 47, + 651, + 287, + 712 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 651, + 287, + 712 + ], + "type": "inline_equation", + "content": "\\Delta F" + }, + { + "bbox": [ + 47, + 651, + 287, + 712 + ], + "type": "text", + "content": ", eliminates the influence of physical parameters on " + }, + { + "bbox": [ + 47, + 651, + 287, + 712 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 47, + 651, + 287, + 712 + ], + "type": "text", + "content": ". Experimental validation confirms that this method delivers robust performance and can enhance images to a high visual quality." + } + ] + } + ], + "index": 20 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 266, + 397, + 278 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 266, + 397, + 278 + ], + "spans": [ + { + "bbox": [ + 306, + 266, + 397, + 278 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 304, + 290, + 545, + 674 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 290, + 545, + 674 + ], + "spans": [ + { + "bbox": [ + 304, + 290, + 545, + 674 + ], + "type": "text", + "content": "Physical Models: Physical models employ image statistical attributes to design priors. For instance, Dark Channel Prior (DCP) [16] utilizes the minimum value in the RGB channels of pixels. Building upon DCP, BCCR [34] introduces a more efficient method for estimating transmission coefficients. Other methods, such as Fast Visibility Restoration (FVR) [51], make assumptions about the atmospheric veil, considering it locally flat, and use median filters instead of minimum filters for transmission estimation. Color Attenuation Prior (CAP) [54] formulates a linear model between depth information and the attenuation coefficient, while ROP [27] proposes an intensity projection strategy for transmission estimation. These prior-based methods perform admirably in specific scene types, but their adaptability to diverse scenes is limited. For example, DCP tends to darken already dark areas (i.e. skies and shadows) when dehazing. To rectify this, LDCP [55] modifies sky regions. Brightness is a crucial factor for accurate transmission estimation [55], enhancing contrast and improving visual quality. Many image enhancement techniques leverage this attribute [12, 45, 52]. These algorithms typically convert RGB images to other color spaces, such as YUV and LAB, to perform color correction [45, 52], for example, HRDCP [45] and CVC [18]. In dealing with sandstorms and underwater environments, color correction methods like white balance, akin to the Gray World Hypothesis [33], are important. The hypothesis posits that the average values of the red, green, and blue channels in a scene should be approximately equal. Considering that sandstorm weather absorbs blue and green light, and underwater environments absorb red light, using color balance as prior knowledge can result in more realistic recovery effects [33]." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 305, + 677, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 677, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 677, + 545, + 713 + ], + "type": "text", + "content": "Data-Driven Methods: Learning-based methods extrapolate models for haze-free image estimation from extensive datasets. Initial models like Waternet [25], MSCNN [40]," + } + ] + } + ], + "index": 24 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2744" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 289, + 360 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 289, + 360 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 289, + 360 + ], + "type": "text", + "content": "and AOD-Net [23], equipped with shallow architectures, outperformed previous prior-based methods. Some even incorporated prior knowledge, such as UWCNN [26] and DehazeNet [3]. Lately, most learning-based methods design more intricate and larger parameterized network architectures to enhance the network's expressivity. These include the attention mechanism [5, 7, 39], multi-scale features [6, 15, 29, 30], and even dehazing networks based on Vision Transformer (ViT) [8, 14, 53], such as gUNet [47]. Despite these advancements, the efficacy of these supervised learning methods is limited by their generalization ability in real-world scenes. The scarcity of large-scale, diverse, and fully aligned paired training data poses a significant hurdle. These methods often rely on synthetic data, but procuring synthetic data for sand and underwater environments presents challenges [28, 46]. Deep learning-based algorithms also encounter difficulties in training, requiring complex loss functions and training strategies, which lead to high training costs [48]. Many existing methods only use image reconstruction loss based on L1/L2, lacking robust regularization. This deficiency can result in unsuitable gradient updates and difficulties in grappling with the uncertainty of ill-conditioned problems, underscoring the need for robust prior information." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 372, + 105, + 385 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 372, + 105, + 385 + ], + "spans": [ + { + "bbox": [ + 47, + 372, + 105, + 385 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 393, + 288, + 477 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 393, + 288, + 477 + ], + "spans": [ + { + "bbox": [ + 46, + 393, + 288, + 477 + ], + "type": "text", + "content": "In this section, we first validate the correlation between " + }, + { + "bbox": [ + 46, + 393, + 288, + 477 + ], + "type": "inline_equation", + "content": "\\Delta F" + }, + { + "bbox": [ + 46, + 393, + 288, + 477 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 393, + 288, + 477 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 393, + 288, + 477 + ], + "type": "text", + "content": " from a statistical perspective (Sec.3.1). Next, we derive the analytical form of transmittance using luminous flux and its rate of change (Sec.3.2). Finally, we present the numerical calculation methods for luminous flux and its rate of change, the image recovery formula and analysis of algorithmic complexity (Sec.3.3)." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 487, + 222, + 499 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 487, + 222, + 499 + ], + "spans": [ + { + "bbox": [ + 47, + 487, + 222, + 499 + ], + "type": "text", + "content": "3.1. Luminous Flux Prior Validation" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 506, + 287, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 506, + 287, + 601 + ], + "spans": [ + { + "bbox": [ + 46, + 506, + 287, + 601 + ], + "type": "text", + "content": "We validate the correlation between the rate of change in luminous flux, denoted as " + }, + { + "bbox": [ + 46, + 506, + 287, + 601 + ], + "type": "inline_equation", + "content": "\\Delta F" + }, + { + "bbox": [ + 46, + 506, + 287, + 601 + ], + "type": "text", + "content": ", and transmission " + }, + { + "bbox": [ + 46, + 506, + 287, + 601 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 506, + 287, + 601 + ], + "type": "text", + "content": ", utilizing datasets from three disparate weather conditions. These conditions encompass images from haze, sandstorms, and underwater environments, both simulated and real. The correlation between " + }, + { + "bbox": [ + 46, + 506, + 287, + 601 + ], + "type": "inline_equation", + "content": "\\Delta F" + }, + { + "bbox": [ + 46, + 506, + 287, + 601 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 506, + 287, + 601 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 506, + 287, + 601 + ], + "type": "text", + "content": " remains robust across all conditions, which underscores the utility of " + }, + { + "bbox": [ + 46, + 506, + 287, + 601 + ], + "type": "inline_equation", + "content": "\\Delta F" + }, + { + "bbox": [ + 46, + 506, + 287, + 601 + ], + "type": "text", + "content": " as a prior for image enhancement. The datasets we use are as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 605, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 47, + 605, + 287, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 605, + 287, + 641 + ], + "spans": [ + { + "bbox": [ + 47, + 605, + 287, + 641 + ], + "type": "text", + "content": "1. Haze: We utilize the RESIDE dataset [24], which contains 13,990 synthetic indoor and 72,135 synthetic outdoor images." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 642, + 287, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 642, + 287, + 676 + ], + "spans": [ + { + "bbox": [ + 47, + 642, + 287, + 676 + ], + "type": "text", + "content": "2. Sandstorm: We rely on the Sand-dust Image Reconstruction Benchmark (SIRB) dataset [46], which simulates sandstorm conditions with different dust intensities." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 677, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 677, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 677, + 287, + 713 + ], + "type": "text", + "content": "3. Underwater: We employ the Underwater Image Enhancement Benchmark Dataset (UIEBD) [26] which includes 890 real underwater environment images together" + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "type": "table", + "bbox": [ + 307, + 123, + 544, + 175 + ], + "blocks": [ + { + "bbox": [ + 305, + 70, + 545, + 115 + ], + "lines": [ + { + "bbox": [ + 305, + 70, + 545, + 115 + ], + "spans": [ + { + "bbox": [ + 305, + 70, + 545, + 115 + ], + "type": "text", + "content": "Table 1. We apply the KLD to measure the difference between the rate of change of luminance and the transmission map. Here, " + }, + { + "bbox": [ + 305, + 70, + 545, + 115 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 305, + 70, + 545, + 115 + ], + "type": "text", + "content": " symbolizes a truncated standard normal distribution, and " + }, + { + "bbox": [ + 305, + 70, + 545, + 115 + ], + "type": "inline_equation", + "content": "U" + }, + { + "bbox": [ + 305, + 70, + 545, + 115 + ], + "type": "text", + "content": " represents a uniform distribution within the range [0, 1]." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 307, + 123, + 544, + 175 + ], + "lines": [ + { + "bbox": [ + 307, + 123, + 544, + 175 + ], + "spans": [ + { + "bbox": [ + 307, + 123, + 544, + 175 + ], + "type": "table", + "html": "
RESIDESIRBO-HazeUIEBD
ITSOTSLightMediumDense
ΔF0.290.210.130.130.160.120.15
N0.870.640.750.991.660.530.77
U0.950.531.271.681.900.430.64
", + "image_path": "2f97fdb62d113b5518f0da3d1d82c2ca4c87986196c3e9263f537c4a6b547c50.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 318, + 201, + 545, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 201, + 545, + 224 + ], + "spans": [ + { + "bbox": [ + 318, + 201, + 545, + 224 + ], + "type": "text", + "content": "with reference images free of water distortion, obtained via their Deep Underwater Image Enhancement Network." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 232, + 545, + 340 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 232, + 545, + 340 + ], + "spans": [ + { + "bbox": [ + 304, + 232, + 545, + 340 + ], + "type": "text", + "content": "In total, these datasets encompass 90,960 images, comprising our entire sample space with each image representing an individual sample. For an image of size " + }, + { + "bbox": [ + 304, + 232, + 545, + 340 + ], + "type": "inline_equation", + "content": "m \\times n" + }, + { + "bbox": [ + 304, + 232, + 545, + 340 + ], + "type": "text", + "content": ", denoted as I, each pixel corresponds to three values: " + }, + { + "bbox": [ + 304, + 232, + 545, + 340 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 232, + 545, + 340 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 232, + 545, + 340 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 304, + 232, + 545, + 340 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 232, + 545, + 340 + ], + "type": "inline_equation", + "content": "\\Delta F" + }, + { + "bbox": [ + 304, + 232, + 545, + 340 + ], + "type": "text", + "content": ". T, F, and " + }, + { + "bbox": [ + 304, + 232, + 545, + 340 + ], + "type": "inline_equation", + "content": "\\Delta \\mathbf{F}" + }, + { + "bbox": [ + 304, + 232, + 545, + 340 + ], + "type": "text", + "content": " represent these values in matrix form. Both " + }, + { + "bbox": [ + 304, + 232, + 545, + 340 + ], + "type": "inline_equation", + "content": "\\Delta \\mathbf{F}" + }, + { + "bbox": [ + 304, + 232, + 545, + 340 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 232, + 545, + 340 + ], + "type": "inline_equation", + "content": "\\mathbf{T}" + }, + { + "bbox": [ + 304, + 232, + 545, + 340 + ], + "type": "text", + "content": " are of the same size as image I. " + }, + { + "bbox": [ + 304, + 232, + 545, + 340 + ], + "type": "inline_equation", + "content": "\\Delta \\mathbf{F} \\in \\mathbb{R}^{m \\times n \\times c}" + }, + { + "bbox": [ + 304, + 232, + 545, + 340 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 232, + 545, + 340 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 304, + 232, + 545, + 340 + ], + "type": "text", + "content": " denotes the number of channels in image I. For instance, when I exists in the RGB color space, " + }, + { + "bbox": [ + 304, + 232, + 545, + 340 + ], + "type": "inline_equation", + "content": "c = 3" + }, + { + "bbox": [ + 304, + 232, + 545, + 340 + ], + "type": "text", + "content": "; when I is a grayscale image, " + }, + { + "bbox": [ + 304, + 232, + 545, + 340 + ], + "type": "inline_equation", + "content": "c = 1" + }, + { + "bbox": [ + 304, + 232, + 545, + 340 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 342, + 545, + 377 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 342, + 545, + 377 + ], + "spans": [ + { + "bbox": [ + 305, + 342, + 545, + 377 + ], + "type": "text", + "content": "We employ the Kullback-Leibler Divergence (KLD) to authenticate the correlation between " + }, + { + "bbox": [ + 305, + 342, + 545, + 377 + ], + "type": "inline_equation", + "content": "\\Delta \\mathbf{F}" + }, + { + "bbox": [ + 305, + 342, + 545, + 377 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 342, + 545, + 377 + ], + "type": "inline_equation", + "content": "\\mathbf{T}" + }, + { + "bbox": [ + 305, + 342, + 545, + 377 + ], + "type": "text", + "content": ", measuring their difference as follows:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 394, + 393, + 545, + 406 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 394, + 393, + 545, + 406 + ], + "spans": [ + { + "bbox": [ + 394, + 393, + 545, + 406 + ], + "type": "interline_equation", + "content": "K L D (\\Delta \\mathbf {F} | | \\mathbf {T}) \\tag {3}", + "image_path": "f5992471984dbdcaeb5c4dcf22c1a27fec4b131124c23f16141276fc4be410b8.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 423, + 545, + 532 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 423, + 545, + 532 + ], + "spans": [ + { + "bbox": [ + 304, + 423, + 545, + 532 + ], + "type": "text", + "content": "We generate the transmission map " + }, + { + "bbox": [ + 304, + 423, + 545, + 532 + ], + "type": "inline_equation", + "content": "\\mathbf{T}" + }, + { + "bbox": [ + 304, + 423, + 545, + 532 + ], + "type": "text", + "content": " of an image by subtracting a low-quality image from a clear counterpart and taking the absolute value. The resulting image represents the degree of light scattering, with brighter regions indicating severe degradation and darker regions suggesting minor or no degradation. We temporarily disregard the effect of atmospheric light. Subsequently, we compute " + }, + { + "bbox": [ + 304, + 423, + 545, + 532 + ], + "type": "inline_equation", + "content": "\\Delta \\mathbf{F}" + }, + { + "bbox": [ + 304, + 423, + 545, + 532 + ], + "type": "text", + "content": " for the degraded image and then verify the correlation between " + }, + { + "bbox": [ + 304, + 423, + 545, + 532 + ], + "type": "inline_equation", + "content": "\\Delta \\mathbf{F}" + }, + { + "bbox": [ + 304, + 423, + 545, + 532 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 423, + 545, + 532 + ], + "type": "inline_equation", + "content": "\\mathbf{T}" + }, + { + "bbox": [ + 304, + 423, + 545, + 532 + ], + "type": "text", + "content": " using Eq 3." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 534, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 534, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 534, + 545, + 712 + ], + "type": "text", + "content": "As presented in Tab 1, the KLD values for " + }, + { + "bbox": [ + 304, + 534, + 545, + 712 + ], + "type": "inline_equation", + "content": "\\mathbf{T}" + }, + { + "bbox": [ + 304, + 534, + 545, + 712 + ], + "type": "text", + "content": " when using uniform or Gaussian distributions range between 0.53 to 1.66 and 0.43 to 1.90 respectively. It is important to note that the use of uniform or normal distributions does not facilitate image recovery, thus their KL values can be perceived as the upper bounds. As the KLD values inch closer to these upper bounds, image recovery becomes increasingly challenging. In contrast, the KLD values for " + }, + { + "bbox": [ + 304, + 534, + 545, + 712 + ], + "type": "inline_equation", + "content": "\\Delta F" + }, + { + "bbox": [ + 304, + 534, + 545, + 712 + ], + "type": "text", + "content": " fall within the significantly lower range of 0.12 to 0.29, approximately 4 to 10 times less than the aforementioned upper bounds. This signifying relatively low information loss when estimating " + }, + { + "bbox": [ + 304, + 534, + 545, + 712 + ], + "type": "inline_equation", + "content": "\\mathbf{T}" + }, + { + "bbox": [ + 304, + 534, + 545, + 712 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 304, + 534, + 545, + 712 + ], + "type": "inline_equation", + "content": "\\Delta \\mathbf{F}" + }, + { + "bbox": [ + 304, + 534, + 545, + 712 + ], + "type": "text", + "content": ". Therefore, " + }, + { + "bbox": [ + 304, + 534, + 545, + 712 + ], + "type": "inline_equation", + "content": "\\Delta \\mathbf{F}" + }, + { + "bbox": [ + 304, + 534, + 545, + 712 + ], + "type": "text", + "content": " effectively encapsulates the characteristics of " + }, + { + "bbox": [ + 304, + 534, + 545, + 712 + ], + "type": "inline_equation", + "content": "\\mathbf{T}" + }, + { + "bbox": [ + 304, + 534, + 545, + 712 + ], + "type": "text", + "content": ", indicating a pronounced correlation between " + }, + { + "bbox": [ + 304, + 534, + 545, + 712 + ], + "type": "inline_equation", + "content": "\\Delta \\mathbf{F}" + }, + { + "bbox": [ + 304, + 534, + 545, + 712 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 534, + 545, + 712 + ], + "type": "inline_equation", + "content": "\\mathbf{T}" + }, + { + "bbox": [ + 304, + 534, + 545, + 712 + ], + "type": "text", + "content": ". These findings substantiate the effectiveness of the LFP." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2745" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 51, + 68, + 532, + 214 + ], + "blocks": [ + { + "bbox": [ + 51, + 68, + 532, + 214 + ], + "lines": [ + { + "bbox": [ + 51, + 68, + 532, + 214 + ], + "spans": [ + { + "bbox": [ + 51, + 68, + 532, + 214 + ], + "type": "image", + "image_path": "64e2b6d36d125d3a553138e24ad4df4bc0a5962db75ac08d89189efd59788888.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 222, + 546, + 246 + ], + "lines": [ + { + "bbox": [ + 46, + 222, + 546, + 246 + ], + "spans": [ + { + "bbox": [ + 46, + 222, + 546, + 246 + ], + "type": "text", + "content": "Figure 2. Visualization of the LFP Algorithm Workflow: First, it calculates Luminous Flux " + }, + { + "bbox": [ + 46, + 222, + 546, + 246 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 46, + 222, + 546, + 246 + ], + "type": "text", + "content": " and it's ratio of change " + }, + { + "bbox": [ + 46, + 222, + 546, + 246 + ], + "type": "inline_equation", + "content": "\\Delta F" + }, + { + "bbox": [ + 46, + 222, + 546, + 246 + ], + "type": "text", + "content": " within a sliding window across the image. Second, it estimates transmission " + }, + { + "bbox": [ + 46, + 222, + 546, + 246 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 222, + 546, + 246 + ], + "type": "text", + "content": " using " + }, + { + "bbox": [ + 46, + 222, + 546, + 246 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 46, + 222, + 546, + 246 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 222, + 546, + 246 + ], + "type": "inline_equation", + "content": "\\Delta F" + }, + { + "bbox": [ + 46, + 222, + 546, + 246 + ], + "type": "text", + "content": ". Finally, recover the image using Eq 15." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 265, + 190, + 277 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 265, + 190, + 277 + ], + "spans": [ + { + "bbox": [ + 47, + 265, + 190, + 277 + ], + "type": "text", + "content": "3.2. Scene Recovery with LFP" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 284, + 228, + 295 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 284, + 228, + 295 + ], + "spans": [ + { + "bbox": [ + 47, + 284, + 228, + 295 + ], + "type": "text", + "content": "3.2.1 Analytical Form of Luminous Flux" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 302, + 287, + 386 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 302, + 287, + 386 + ], + "spans": [ + { + "bbox": [ + 46, + 302, + 287, + 386 + ], + "type": "text", + "content": "The Bouguer-Lambert law [35] offers a comprehensive analytical expression for the measurement of luminous flux. This law utilizes the attenuation coefficient " + }, + { + "bbox": [ + 46, + 302, + 287, + 386 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 46, + 302, + 287, + 386 + ], + "type": "text", + "content": " [37] and the scene depth " + }, + { + "bbox": [ + 46, + 302, + 287, + 386 + ], + "type": "inline_equation", + "content": "d(x)" + }, + { + "bbox": [ + 46, + 302, + 287, + 386 + ], + "type": "text", + "content": " [21] as primary factors to accurately evaluate luminous flux Eq 4. In this section, to ensure clarity and simplicity, we have chosen to omit the position information denoted by " + }, + { + "bbox": [ + 46, + 302, + 287, + 386 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 46, + 302, + 287, + 386 + ], + "type": "text", + "content": " from the symbols used." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 141, + 396, + 287, + 410 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 396, + 287, + 410 + ], + "spans": [ + { + "bbox": [ + 141, + 396, + 287, + 410 + ], + "type": "interline_equation", + "content": "F = \\tilde {F} e ^ {- \\beta d} \\tag {4}", + "image_path": "0ca833fa20bcdec582f607f5bc79e55fa99ec9ca6292bf350f400bc7d5542f6b.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 415, + 268, + 427 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 415, + 268, + 427 + ], + "spans": [ + { + "bbox": [ + 47, + 415, + 268, + 427 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 415, + 268, + 427 + ], + "type": "inline_equation", + "content": "\\tilde{F}" + }, + { + "bbox": [ + 47, + 415, + 268, + 427 + ], + "type": "text", + "content": " represents the luminous flux of incident light." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 441, + 186, + 453 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 441, + 186, + 453 + ], + "spans": [ + { + "bbox": [ + 47, + 441, + 186, + 453 + ], + "type": "text", + "content": "3.2.2 Estimating Transmission" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 460, + 287, + 484 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 460, + 287, + 484 + ], + "spans": [ + { + "bbox": [ + 47, + 460, + 287, + 484 + ], + "type": "text", + "content": "Transmission, a property of the material, is the ratio of " + }, + { + "bbox": [ + 47, + 460, + 287, + 484 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 47, + 460, + 287, + 484 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 47, + 460, + 287, + 484 + ], + "type": "inline_equation", + "content": "\\tilde{F}" + }, + { + "bbox": [ + 47, + 460, + 287, + 484 + ], + "type": "text", + "content": ", as depicted in Eq 5:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 145, + 491, + 287, + 506 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 491, + 287, + 506 + ], + "spans": [ + { + "bbox": [ + 145, + 491, + 287, + 506 + ], + "type": "interline_equation", + "content": "t = F / \\tilde {F}. \\tag {5}", + "image_path": "96cef7b93ad0352f135cb287f179a08cba9cf877fef0098e347aa02ed1041623.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 514, + 287, + 573 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 514, + 287, + 573 + ], + "spans": [ + { + "bbox": [ + 47, + 514, + 287, + 573 + ], + "type": "text", + "content": "In line with the Retinex [22] model, we perceive an image as composed of incident and reflected components. Incident light illuminates the object, and the object's reflected light forms the image we observe. The incident light " + }, + { + "bbox": [ + 47, + 514, + 287, + 573 + ], + "type": "inline_equation", + "content": "\\tilde{F}" + }, + { + "bbox": [ + 47, + 514, + 287, + 573 + ], + "type": "text", + "content": " can be eliminated using division, as shown in Eq 6:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 580, + 287, + 610 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 106, + 580, + 287, + 610 + ], + "spans": [ + { + "bbox": [ + 106, + 580, + 287, + 610 + ], + "type": "interline_equation", + "content": "t _ {i j} = \\frac {t _ {i}}{t _ {j}} = \\frac {F _ {i} / \\tilde {F} _ {i}}{F _ {j} / \\tilde {F} _ {j}} = F _ {i} / F _ {j}, \\tag {6}", + "image_path": "d90f8d1724f3216a6eb52117405e5ff149d65835146d3db811ada8aff089c377.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 47, + 617, + 287, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 617, + 287, + 677 + ], + "spans": [ + { + "bbox": [ + 47, + 617, + 287, + 677 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 617, + 287, + 677 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 47, + 617, + 287, + 677 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 617, + 287, + 677 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 47, + 617, + 287, + 677 + ], + "type": "text", + "content": " represent two adjacent pixels " + }, + { + "bbox": [ + 47, + 617, + 287, + 677 + ], + "type": "inline_equation", + "content": "x_{i}" + }, + { + "bbox": [ + 47, + 617, + 287, + 677 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 617, + 287, + 677 + ], + "type": "inline_equation", + "content": "x_{j}" + }, + { + "bbox": [ + 47, + 617, + 287, + 677 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 47, + 617, + 287, + 677 + ], + "type": "inline_equation", + "content": "\\tilde{F}_{i} = \\tilde{F} j" + }, + { + "bbox": [ + 47, + 617, + 287, + 677 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 47, + 617, + 287, + 677 + ], + "type": "inline_equation", + "content": "t_{ij}" + }, + { + "bbox": [ + 47, + 617, + 287, + 677 + ], + "type": "text", + "content": " is a dimensionless parameter depicting the luminous flux " + }, + { + "bbox": [ + 47, + 617, + 287, + 677 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 47, + 617, + 287, + 677 + ], + "type": "text", + "content": " ratio between adjacent pixels. This allows analysis of flux changes within the image, crucial for revealing local features and their variations." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 59, + 679, + 250, + 692 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 679, + 250, + 692 + ], + "spans": [ + { + "bbox": [ + 59, + 679, + 250, + 692 + ], + "type": "text", + "content": "Based on Eq 6, " + }, + { + "bbox": [ + 59, + 679, + 250, + 692 + ], + "type": "inline_equation", + "content": "t_{ij}" + }, + { + "bbox": [ + 59, + 679, + 250, + 692 + ], + "type": "text", + "content": " can be further expressed as:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 107, + 700, + 287, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 107, + 700, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 107, + 700, + 287, + 715 + ], + "type": "interline_equation", + "content": "t _ {i j} = F _ {i} / F _ {j} = e ^ {- d _ {i} \\cdot \\left(\\beta_ {i} - \\beta_ {j}\\right)}. \\tag {7}", + "image_path": "d7c9a2c88dd52d91c3193ecac8e8fa68d513b07afef732883f7049601caafd9b.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 266, + 545, + 304 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 266, + 545, + 304 + ], + "spans": [ + { + "bbox": [ + 305, + 266, + 545, + 304 + ], + "type": "text", + "content": "In Eq 7, to comprehend the influence of luminous flux on " + }, + { + "bbox": [ + 305, + 266, + 545, + 304 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 305, + 266, + 545, + 304 + ], + "type": "text", + "content": ", we introduce " + }, + { + "bbox": [ + 305, + 266, + 545, + 304 + ], + "type": "inline_equation", + "content": "\\mathrm{d}\\beta = \\frac{\\beta_i - \\beta_j}{F_i - F_j}" + }, + { + "bbox": [ + 305, + 266, + 545, + 304 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 305, + 266, + 545, + 304 + ], + "type": "inline_equation", + "content": "\\mathrm{d}\\beta" + }, + { + "bbox": [ + 305, + 266, + 545, + 304 + ], + "type": "text", + "content": " is a differential with respect to " + }, + { + "bbox": [ + 305, + 266, + 545, + 304 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 305, + 266, + 545, + 304 + ], + "type": "text", + "content": ". By substituting " + }, + { + "bbox": [ + 305, + 266, + 545, + 304 + ], + "type": "inline_equation", + "content": "\\mathrm{d}\\beta" + }, + { + "bbox": [ + 305, + 266, + 545, + 304 + ], + "type": "text", + "content": " into Eq 7, we obtain:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 381, + 308, + 545, + 356 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 381, + 308, + 545, + 356 + ], + "spans": [ + { + "bbox": [ + 381, + 308, + 545, + 356 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} t _ {i j} = e ^ {- d _ {i} \\left(\\beta_ {i} - \\beta_ {j}\\right)} \\\\ = e ^ {- d _ {i} \\left(\\beta_ {i} - \\beta_ {j}\\right) \\frac {\\Delta F}{\\Delta F}} \\tag {8} \\\\ = e ^ {- d _ {i} \\cdot \\Delta F \\cdot \\mathrm {d} \\beta}. \\\\ \\end{array}", + "image_path": "49de15bb5bc7bce8963ac8fc93b57b638cb8e6e8cc39e28eb66d621dc532ded5.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 318, + 364, + 537, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 364, + 537, + 376 + ], + "spans": [ + { + "bbox": [ + 318, + 364, + 537, + 376 + ], + "type": "text", + "content": "Simultaneously integrating both sides of Eq 8, obtain:" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 378, + 381, + 545, + 421 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 378, + 381, + 545, + 421 + ], + "spans": [ + { + "bbox": [ + 378, + 381, + 545, + 421 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\int t \\mathrm {d} t = \\int e ^ {- d \\cdot \\Delta F \\cdot \\mathrm {d} \\beta} \\tag {9} \\\\ t = \\sqrt {2 e ^ {- d \\cdot \\beta \\cdot \\Delta F}}. \\\\ \\end{array}", + "image_path": "069c00ca9d078fcc3bde6acd3d5c132f539581ff2f66e7509777152a7467f5a8.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 306, + 429, + 461, + 440 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 429, + 461, + 440 + ], + "spans": [ + { + "bbox": [ + 306, + 429, + 461, + 440 + ], + "type": "text", + "content": "The proof is provided in the appendix." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 440, + 545, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 440, + 545, + 475 + ], + "spans": [ + { + "bbox": [ + 304, + 440, + 545, + 475 + ], + "type": "text", + "content": "For ease of discussion, we consider " + }, + { + "bbox": [ + 304, + 440, + 545, + 475 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{F}}" + }, + { + "bbox": [ + 304, + 440, + 545, + 475 + ], + "type": "text", + "content": " as a constant, denoted as " + }, + { + "bbox": [ + 304, + 440, + 545, + 475 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 440, + 545, + 475 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 440, + 545, + 475 + ], + "type": "inline_equation", + "content": "k \\in (0,1]" + }, + { + "bbox": [ + 304, + 440, + 545, + 475 + ], + "type": "text", + "content": ". From Eq 4, " + }, + { + "bbox": [ + 304, + 440, + 545, + 475 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 304, + 440, + 545, + 475 + ], + "type": "text", + "content": " can be expressed as in Eq 10:" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 393, + 481, + 545, + 505 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 393, + 481, + 545, + 505 + ], + "spans": [ + { + "bbox": [ + 393, + 481, + 545, + 505 + ], + "type": "interline_equation", + "content": "\\beta = - \\frac {k}{d} \\log F. \\tag {10}", + "image_path": "7e761078dc025c45439527f51fa2e1c03bc7e1908dc9b03bff999e9a89234aab.jpg" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 305, + 509, + 544, + 531 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 509, + 544, + 531 + ], + "spans": [ + { + "bbox": [ + 305, + 509, + 544, + 531 + ], + "type": "text", + "content": "By substituting Eq 10 into Eq 9, we obtain the expression for " + }, + { + "bbox": [ + 305, + 509, + 544, + 531 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 305, + 509, + 544, + 531 + ], + "type": "text", + "content": " as follows:" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 386, + 536, + 545, + 585 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 386, + 536, + 545, + 585 + ], + "spans": [ + { + "bbox": [ + 386, + 536, + 545, + 585 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} t = \\sqrt {2 e ^ {- d \\cdot \\beta \\cdot \\Delta F}} \\\\ = \\sqrt {2 e ^ {\\log F \\cdot k \\cdot \\Delta F}} \\tag {11} \\\\ = \\sqrt {2 F ^ {k \\Delta F}}. \\\\ \\end{array}", + "image_path": "168f37443c1d69858f8940e9082e6079f4248285b4ea285c2085241027c97a29.jpg" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 304, + 592, + 545, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 592, + 545, + 652 + ], + "spans": [ + { + "bbox": [ + 304, + 592, + 545, + 652 + ], + "type": "text", + "content": "In Eq 11, we have successfully eliminated the two physical parameters " + }, + { + "bbox": [ + 304, + 592, + 545, + 652 + ], + "type": "inline_equation", + "content": "d(x)" + }, + { + "bbox": [ + 304, + 592, + 545, + 652 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 592, + 545, + 652 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 304, + 592, + 545, + 652 + ], + "type": "text", + "content": " and derived an analytical form for " + }, + { + "bbox": [ + 304, + 592, + 545, + 652 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 592, + 545, + 652 + ], + "type": "text", + "content": ". Considering that the exponent part is relatively small compared to the base, Eq 11 can be further simplified, resulting in Eq 12:" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 376, + 658, + 545, + 671 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 376, + 658, + 545, + 671 + ], + "spans": [ + { + "bbox": [ + 376, + 658, + 545, + 671 + ], + "type": "interline_equation", + "content": "t \\approx 1 + k \\cdot \\log F \\cdot \\Delta F. \\tag {12}", + "image_path": "2254129b37165bc8a5aa55315b80990db6b21cf8aba6a6817fe4aab490f6b6bb.jpg" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 304, + 677, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 545, + 714 + ], + "type": "text", + "content": "In Eq 12, we consolidate all constants from Eq 11 into " + }, + { + "bbox": [ + 304, + 677, + 545, + 714 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 677, + 545, + 714 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 677, + 545, + 714 + ], + "type": "inline_equation", + "content": "k \\in (0,1]" + }, + { + "bbox": [ + 304, + 677, + 545, + 714 + ], + "type": "text", + "content": ", as the maximum value of luminous flux is normalized to 1. The proof is provided in the appendix." + } + ] + } + ], + "index": 26 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2746" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 163, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 163, + 85 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 163, + 85 + ], + "type": "text", + "content": "3.3. Proposed algorithm" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 90, + 289, + 210 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 90, + 289, + 210 + ], + "spans": [ + { + "bbox": [ + 46, + 90, + 289, + 210 + ], + "type": "text", + "content": "The numerical method for computing luminous flux, denoted as " + }, + { + "bbox": [ + 46, + 90, + 289, + 210 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 46, + 90, + 289, + 210 + ], + "type": "text", + "content": ", leverages its physical definition: the ratio of light intensity over a given area to the area itself. We implement a sliding window approach, traversing the entire image. For each window, we calculate the mean intensity and subsequently divide by the window's area to compute the numerical value of the luminous flux for each pixel. In this context, " + }, + { + "bbox": [ + 46, + 90, + 289, + 210 + ], + "type": "inline_equation", + "content": "\\Omega" + }, + { + "bbox": [ + 46, + 90, + 289, + 210 + ], + "type": "text", + "content": " represents the sliding window, " + }, + { + "bbox": [ + 46, + 90, + 289, + 210 + ], + "type": "inline_equation", + "content": "\\Omega(\\cdot)" + }, + { + "bbox": [ + 46, + 90, + 289, + 210 + ], + "type": "text", + "content": " is the mean value of the samplings within the window, and " + }, + { + "bbox": [ + 46, + 90, + 289, + 210 + ], + "type": "inline_equation", + "content": "S_{\\Omega}" + }, + { + "bbox": [ + 46, + 90, + 289, + 210 + ], + "type": "text", + "content": " is the area of the window. Denoted as Eq 13:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 127, + 219, + 287, + 236 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 127, + 219, + 287, + 236 + ], + "spans": [ + { + "bbox": [ + 127, + 219, + 287, + 236 + ], + "type": "interline_equation", + "content": "F = \\sum \\Omega (\\mathbf {I}) / S _ {\\Omega}. \\tag {13}", + "image_path": "aa983396b409924d168f751ca57845f0a5354c7a68de5af58bc7eaf1557733e1.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 244, + 287, + 281 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 244, + 287, + 281 + ], + "spans": [ + { + "bbox": [ + 47, + 244, + 287, + 281 + ], + "type": "text", + "content": "For numerical computation, we sample the neighborhoods of each pixel by using sliding window. Consequently, for each pixel, the " + }, + { + "bbox": [ + 47, + 244, + 287, + 281 + ], + "type": "inline_equation", + "content": "\\Delta F" + }, + { + "bbox": [ + 47, + 244, + 287, + 281 + ], + "type": "text", + "content": " can be expressed as shown in Eq 14:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 128, + 289, + 287, + 315 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 128, + 289, + 287, + 315 + ], + "spans": [ + { + "bbox": [ + 128, + 289, + 287, + 315 + ], + "type": "interline_equation", + "content": "\\Delta F = \\frac {\\min \\Omega (F)}{\\max \\Omega (F)} \\tag {14}", + "image_path": "dc0ddd660906983b5aeab45a5f64abb270695b65f2e6f72ebe1069765421200b.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 319, + 287, + 402 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 319, + 287, + 402 + ], + "spans": [ + { + "bbox": [ + 47, + 319, + 287, + 402 + ], + "type": "text", + "content": "We use the extreme ratio within the sliding window to amplify " + }, + { + "bbox": [ + 47, + 319, + 287, + 402 + ], + "type": "inline_equation", + "content": "\\Delta F" + }, + { + "bbox": [ + 47, + 319, + 287, + 402 + ], + "type": "text", + "content": " because the difference between adjacent pixels is close to 0. However, " + }, + { + "bbox": [ + 47, + 319, + 287, + 402 + ], + "type": "inline_equation", + "content": "\\Delta F" + }, + { + "bbox": [ + 47, + 319, + 287, + 402 + ], + "type": "text", + "content": " might be affected by noise. To mitigate this, we first truncate " + }, + { + "bbox": [ + 47, + 319, + 287, + 402 + ], + "type": "inline_equation", + "content": "\\Omega(F)" + }, + { + "bbox": [ + 47, + 319, + 287, + 402 + ], + "type": "text", + "content": " within the range [0.005, 0.95]. Following this, it may be necessary to truncate the transmission " + }, + { + "bbox": [ + 47, + 319, + 287, + 402 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 47, + 319, + 287, + 402 + ], + "type": "text", + "content": " during subsequent processing to ensure stability and accuracy in our results." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 402, + 287, + 439 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 402, + 287, + 439 + ], + "spans": [ + { + "bbox": [ + 47, + 402, + 287, + 439 + ], + "type": "text", + "content": "In recovery process, we define a clear image " + }, + { + "bbox": [ + 47, + 402, + 287, + 439 + ], + "type": "inline_equation", + "content": "\\mathbf{J}" + }, + { + "bbox": [ + 47, + 402, + 287, + 439 + ], + "type": "text", + "content": " using an image formulation model from Eq 1. This is expressed as follows in Eq 15:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 102, + 446, + 287, + 473 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 102, + 446, + 287, + 473 + ], + "spans": [ + { + "bbox": [ + 102, + 446, + 287, + 473 + ], + "type": "interline_equation", + "content": "\\mathbf {J} (x) = \\frac {\\mathbf {I} (x) + \\mathbf {A} \\omega (t (x) - 1)}{\\operatorname* {m a x} (\\omega t (x) , t _ {0})}. \\tag {15}", + "image_path": "7f5999bc9d8eb106774efe2a321c9fb3fc6856d274405ebac5881dca29f69c32.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 480, + 287, + 516 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 480, + 287, + 516 + ], + "spans": [ + { + "bbox": [ + 47, + 480, + 287, + 516 + ], + "type": "text", + "content": "In Eq 15, " + }, + { + "bbox": [ + 47, + 480, + 287, + 516 + ], + "type": "inline_equation", + "content": "\\omega" + }, + { + "bbox": [ + 47, + 480, + 287, + 516 + ], + "type": "text", + "content": " is a constant relaxation parameter which is constrained to be in the range " + }, + { + "bbox": [ + 47, + 480, + 287, + 516 + ], + "type": "inline_equation", + "content": "(0,1]" + }, + { + "bbox": [ + 47, + 480, + 287, + 516 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 47, + 480, + 287, + 516 + ], + "type": "inline_equation", + "content": "t_0" + }, + { + "bbox": [ + 47, + 480, + 287, + 516 + ], + "type": "text", + "content": ", set as 0.05, is a lower bound that provides stability during computation." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 517, + 287, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 517, + 287, + 540 + ], + "spans": [ + { + "bbox": [ + 47, + 517, + 287, + 540 + ], + "type": "text", + "content": "The global atmospheric light, represented by " + }, + { + "bbox": [ + 47, + 517, + 287, + 540 + ], + "type": "inline_equation", + "content": "\\mathbf{A}" + }, + { + "bbox": [ + 47, + 517, + 287, + 540 + ], + "type": "text", + "content": ", plays a crucial role in our model. We estimate " + }, + { + "bbox": [ + 47, + 517, + 287, + 540 + ], + "type": "inline_equation", + "content": "\\mathbf{A}" + }, + { + "bbox": [ + 47, + 517, + 287, + 540 + ], + "type": "text", + "content": " swiftly using a" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 307, + 72, + 448, + 83 + ], + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 72, + 448, + 83 + ], + "spans": [ + { + "bbox": [ + 307, + 72, + 448, + 83 + ], + "type": "text", + "content": "Algorithm 1 Pseudo-code for LFP" + } + ] + } + ], + "index": 10, + "type": "text" + }, + { + "bbox": [ + 309, + 87, + 545, + 245 + ], + "type": "list", + "angle": 0, + "index": 21, + "blocks": [ + { + "bbox": [ + 311, + 87, + 522, + 99 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 87, + 522, + 99 + ], + "spans": [ + { + "bbox": [ + 311, + 87, + 522, + 99 + ], + "type": "text", + "content": "1: Setup: constant " + }, + { + "bbox": [ + 311, + 87, + 522, + 99 + ], + "type": "inline_equation", + "content": "k = 0.5" + }, + { + "bbox": [ + 311, + 87, + 522, + 99 + ], + "type": "text", + "content": " and window size " + }, + { + "bbox": [ + 311, + 87, + 522, + 99 + ], + "type": "inline_equation", + "content": "\\Omega = 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 312, + 99, + 430, + 111 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 99, + 430, + 111 + ], + "spans": [ + { + "bbox": [ + 312, + 99, + 430, + 111 + ], + "type": "text", + "content": "2: Input: Degraded image " + }, + { + "bbox": [ + 312, + 99, + 430, + 111 + ], + "type": "inline_equation", + "content": "I" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 312, + 112, + 545, + 122 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 112, + 545, + 122 + ], + "spans": [ + { + "bbox": [ + 312, + 112, + 545, + 122 + ], + "type": "text", + "content": "3: while " + }, + { + "bbox": [ + 312, + 112, + 545, + 122 + ], + "type": "inline_equation", + "content": "\\Omega" + }, + { + "bbox": [ + 312, + 112, + 545, + 122 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 312, + 112, + 545, + 122 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 312, + 112, + 545, + 122 + ], + "type": "text", + "content": " do" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 312, + 123, + 517, + 136 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 123, + 517, + 136 + ], + "spans": [ + { + "bbox": [ + 312, + 123, + 517, + 136 + ], + "type": "text", + "content": "4: Compute mean intensity: " + }, + { + "bbox": [ + 312, + 123, + 517, + 136 + ], + "type": "inline_equation", + "content": "I_{\\Omega} = \\sum \\Omega (I) / S_{\\Omega}" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 312, + 141, + 490, + 154 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 141, + 490, + 154 + ], + "spans": [ + { + "bbox": [ + 312, + 141, + 490, + 154 + ], + "type": "text", + "content": "5: Compute luminous flux: " + }, + { + "bbox": [ + 312, + 141, + 490, + 154 + ], + "type": "inline_equation", + "content": "F = I_{\\Omega} / S_{\\Omega}" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 312, + 156, + 507, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 156, + 507, + 175 + ], + "spans": [ + { + "bbox": [ + 312, + 156, + 507, + 175 + ], + "type": "text", + "content": "6: Compute rate of change: " + }, + { + "bbox": [ + 312, + 156, + 507, + 175 + ], + "type": "inline_equation", + "content": "\\Delta F = \\frac{\\min\\Omega(F)}{\\max\\Omega(F)}" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 312, + 177, + 367, + 188 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 177, + 367, + 188 + ], + "spans": [ + { + "bbox": [ + 312, + 177, + 367, + 188 + ], + "type": "text", + "content": "7: end while" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 312, + 194, + 545, + 207 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 194, + 545, + 207 + ], + "spans": [ + { + "bbox": [ + 312, + 194, + 545, + 207 + ], + "type": "text", + "content": "8: " + }, + { + "bbox": [ + 312, + 194, + 545, + 207 + ], + "type": "inline_equation", + "content": "t\\approx 1 + k\\cdot \\log F\\cdot \\Delta F" + }, + { + "bbox": [ + 312, + 194, + 545, + 207 + ], + "type": "text", + "content": " Compute the transmission" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 312, + 213, + 453, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 213, + 453, + 224 + ], + "spans": [ + { + "bbox": [ + 312, + 213, + 453, + 224 + ], + "type": "text", + "content": "9: guided filtering on transmission" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 309, + 228, + 545, + 245 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 309, + 228, + 545, + 245 + ], + "spans": [ + { + "bbox": [ + 309, + 228, + 545, + 245 + ], + "type": "text", + "content": "10: Output: " + }, + { + "bbox": [ + 309, + 228, + 545, + 245 + ], + "type": "inline_equation", + "content": "J = \\frac{I + A \\cdot (t - 1)}{t}" + } + ] + } + ], + "index": 20 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 304, + 268, + 547, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 268, + 547, + 544 + ], + "spans": [ + { + "bbox": [ + 304, + 268, + 547, + 544 + ], + "type": "text", + "content": "two-step approach. Initially, " + }, + { + "bbox": [ + 304, + 268, + 547, + 544 + ], + "type": "inline_equation", + "content": "\\mathbf{A}" + }, + { + "bbox": [ + 304, + 268, + 547, + 544 + ], + "type": "text", + "content": " is estimated using mean filtering and the maximum grayscale value of the original image. This gives us an initial approximation of the global atmospheric light. Subsequently, the image is segmented into small regions, and we select the maximum value from each region to represent the " + }, + { + "bbox": [ + 304, + 268, + 547, + 544 + ], + "type": "inline_equation", + "content": "\\mathbf{A}" + }, + { + "bbox": [ + 304, + 268, + 547, + 544 + ], + "type": "text", + "content": " for that particular region. The final estimate of " + }, + { + "bbox": [ + 304, + 268, + 547, + 544 + ], + "type": "inline_equation", + "content": "\\mathbf{A}" + }, + { + "bbox": [ + 304, + 268, + 547, + 544 + ], + "type": "text", + "content": " is then calculated as the average of all the " + }, + { + "bbox": [ + 304, + 268, + 547, + 544 + ], + "type": "inline_equation", + "content": "\\mathbf{A}" + }, + { + "bbox": [ + 304, + 268, + 547, + 544 + ], + "type": "text", + "content": " values from these small regions. This method allows us to get a more refined and accurate estimate of the global atmospheric light. The complete process of our scene recovery method is thoroughly depicted in Fig 2, and the corresponding pseudo-code is provided in Algorithm 1. In terms of computational complexity, our algorithm starts off with a complexity of " + }, + { + "bbox": [ + 304, + 268, + 547, + 544 + ], + "type": "inline_equation", + "content": "\\mathcal{O}(N)" + }, + { + "bbox": [ + 304, + 268, + 547, + 544 + ], + "type": "text", + "content": " as we traverse all pixels in the image using a sliding window. Then, it rises by " + }, + { + "bbox": [ + 304, + 268, + 547, + 544 + ], + "type": "inline_equation", + "content": "\\mathcal{O}(\\log N)" + }, + { + "bbox": [ + 304, + 268, + 547, + 544 + ], + "type": "text", + "content": " when we compute the maximum and minimum values for each window. Finally, the total computational complexity for the subsequent steps, which include the calculation of transmission according to Eq 11 and application of a guided filter [17] for smoothing, is " + }, + { + "bbox": [ + 304, + 268, + 547, + 544 + ], + "type": "inline_equation", + "content": "\\mathcal{O}(N\\log N)" + }, + { + "bbox": [ + 304, + 268, + 547, + 544 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 268, + 547, + 544 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 268, + 547, + 544 + ], + "type": "text", + "content": " denotes the size of a single image. This complexity analysis demonstrates that our method is efficient and feasible for large scale image recovery tasks." + } + ] + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 66, + 556, + 531, + 671 + ], + "blocks": [ + { + "bbox": [ + 66, + 556, + 531, + 671 + ], + "lines": [ + { + "bbox": [ + 66, + 556, + 531, + 671 + ], + "spans": [ + { + "bbox": [ + 66, + 556, + 531, + 671 + ], + "type": "image", + "image_path": "acafea452d1b56b835d5b38a6164235a83588187bba4da843608a48bfabccd7a.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 677, + 547, + 711 + ], + "lines": [ + { + "bbox": [ + 46, + 677, + 547, + 711 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 547, + 711 + ], + "type": "text", + "content": "Figure 3. Visual comparisons of different dehazing methods on outdoor and indoor images. The algorithms designed for water removal (blue) and sand removal (yellow-brown) fail in haze (grey) environments. Our method effectively restores details in the sky region without introducing artifacts. It also improves the visual quality of indoor images by enhancing brightness while preserving the structure." + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2747" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 71, + 128, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 71, + 128, + 85 + ], + "spans": [ + { + "bbox": [ + 47, + 71, + 128, + 85 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 91, + 180, + 104 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 91, + 180, + 104 + ], + "spans": [ + { + "bbox": [ + 47, + 91, + 180, + 104 + ], + "type": "text", + "content": "4.1. Implementation Details" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 110, + 288, + 229 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 110, + 288, + 229 + ], + "spans": [ + { + "bbox": [ + 46, + 110, + 288, + 229 + ], + "type": "text", + "content": "Our method relies solely on two parameters: " + }, + { + "bbox": [ + 46, + 110, + 288, + 229 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 110, + 288, + 229 + ], + "type": "text", + "content": " and the sliding window " + }, + { + "bbox": [ + 46, + 110, + 288, + 229 + ], + "type": "inline_equation", + "content": "\\Omega" + }, + { + "bbox": [ + 46, + 110, + 288, + 229 + ], + "type": "text", + "content": ". And uses a non-parametric approach for the estimation of transmission " + }, + { + "bbox": [ + 46, + 110, + 288, + 229 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 110, + 288, + 229 + ], + "type": "text", + "content": ", enhancing visual results under a variety of visibility conditions. Based on empirical experiences, we have determined the optimal settings to be " + }, + { + "bbox": [ + 46, + 110, + 288, + 229 + ], + "type": "inline_equation", + "content": "k = 0.5" + }, + { + "bbox": [ + 46, + 110, + 288, + 229 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 110, + 288, + 229 + ], + "type": "inline_equation", + "content": "\\Omega = 7" + }, + { + "bbox": [ + 46, + 110, + 288, + 229 + ], + "type": "text", + "content": ". All experiments were conducted on a machine with an AMD Ryzen 73750H CPU@2.3GHz and a GeForce RTX 2080 GPU (11GB). The deep learning models were built with the PyTorch, TensorFlow and Matlab frameworks. The test images are publicly available." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 237, + 174, + 251 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 237, + 174, + 251 + ], + "spans": [ + { + "bbox": [ + 47, + 237, + 174, + 251 + ], + "type": "text", + "content": "4.2. Experimental Settings" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 256, + 287, + 375 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 256, + 287, + 375 + ], + "spans": [ + { + "bbox": [ + 46, + 256, + 287, + 375 + ], + "type": "text", + "content": "We evaluated the LFP's performance using various methods and datasets, including physical models and deep learning models. Selected datasets spanned varied environments and scattering degrees, including RESIDE [24], SIRB [46], UIEBD [26], Haze4K [32], and RS-Haze [24]. These encompassed synthetic to real-world scenes, ensuring broad evaluation robustness. We used Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity Index Measure (SSIM) metrics for objective and fair evaluations across scenarios and methods." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 384, + 183, + 396 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 384, + 183, + 396 + ], + "spans": [ + { + "bbox": [ + 47, + 384, + 183, + 396 + ], + "type": "text", + "content": "4.3. Qualitative Comparison" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 403, + 288, + 558 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 403, + 288, + 558 + ], + "spans": [ + { + "bbox": [ + 46, + 403, + 288, + 558 + ], + "type": "text", + "content": "Image Dehazing. We use RESIDE dataset [24] to validate the efficacy and robustness of LFP, as is shown in Fig 3, by comparing it with " + }, + { + "bbox": [ + 46, + 403, + 288, + 558 + ], + "type": "inline_equation", + "content": "\\mathrm{Dive + }" + }, + { + "bbox": [ + 46, + 403, + 288, + 558 + ], + "type": "text", + "content": " and HRDCP. While " + }, + { + "bbox": [ + 46, + 403, + 288, + 558 + ], + "type": "inline_equation", + "content": "\\mathrm{Dive + }" + }, + { + "bbox": [ + 46, + 403, + 288, + 558 + ], + "type": "text", + "content": " fails to restore image clarity, HRDCP results in color deviation, likely due to the specific absorption and scattering properties of sand particles. In contrast, our LFP method can generate more natural-looking results with better structures in a more stable manner. This demonstrates that images in hazy and underwater environments are formed based on different optical properties. We then compare the performance of the LFP with different methods like MSCNN, gUNet, and DCP. Learning-based methods such as gUNet and MSCNN are susceptible to dataset bias, which makes it difficult to" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 72, + 545, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 240 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 240 + ], + "type": "text", + "content": "enhance visual quality. Meanwhile, traditional methods like DCP, which rely on hand-crafted priors, struggle with varied real-world conditions. LFP adeptly overcomes the major obstacle faced by previous dehazing approaches, which is the problematic influence of scene depth. The inherent uncertainty in depth estimation often leads to unstable performance. By confidently deriving the relationship between " + }, + { + "bbox": [ + 304, + 72, + 545, + 240 + ], + "type": "inline_equation", + "content": "\\Delta F" + }, + { + "bbox": [ + 304, + 72, + 545, + 240 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 72, + 545, + 240 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 72, + 545, + 240 + ], + "type": "text", + "content": " without involving " + }, + { + "bbox": [ + 304, + 72, + 545, + 240 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 304, + 72, + 545, + 240 + ], + "type": "text", + "content": ", LFP successfully avoids the dataset bias that learning-based methods encounter. LFP effectively bypasses complex parameters of transmission and offers reliable solution with robust theoretical guarantees. Our method demonstrates a significantly stronger capability for robust image restoration in hazy weather conditions, resulting in highly satisfactory visual results." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 247, + 545, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 247, + 545, + 559 + ], + "spans": [ + { + "bbox": [ + 304, + 247, + 545, + 559 + ], + "type": "text", + "content": "Underwater Image Enhancement. We validate the efficacy and robustness of LFP for underwater image enhancement through comprehensive experiments, using UIEBD dataset [26]. Results are shown in Fig 4. We first compare it with DCP and HRDCP, which are methodologies designed for different weather conditions. DCP fails completely, and although HRDCP achieves some visual effects, it suffers from noticeable color distortion. These experimental results underscore the necessity for image processing techniques to be specifically designed and optimized for each environment, reinforcing that an algorithm effective in one environment can't simply be applied to another without adaptation. " + }, + { + "bbox": [ + 304, + 247, + 545, + 559 + ], + "type": "inline_equation", + "content": "\\mathrm{Dive + }" + }, + { + "bbox": [ + 304, + 247, + 545, + 559 + ], + "type": "text", + "content": " and UWCNN struggle to effectively eliminate the green light prevalent in underwater scenes, leading to undesired color shifts. Although Waternet and ROP perform well in color restoration, they both suffer from dimmed lighting and reduced contrast, thereby compromising the overall visual quality. LFP overcomes the major obstacle faced by previous underwater image enhancement approaches, which is the varying influence of the attenuation coefficient " + }, + { + "bbox": [ + 304, + 247, + 545, + 559 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 304, + 247, + 545, + 559 + ], + "type": "text", + "content": " on the RGB color channels that often results in color bias. By eliminating the attenuation coefficient " + }, + { + "bbox": [ + 304, + 247, + 545, + 559 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 304, + 247, + 545, + 559 + ], + "type": "text", + "content": " during the derivation process, LFP avoids the color bias, rendering it a reliable solution with robust theoretical guarantees. Therefore, LFP method demonstrates competitive color correction and contrast enhancement capabilities." + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 53, + 574, + 544, + 681 + ], + "blocks": [ + { + "bbox": [ + 53, + 574, + 544, + 681 + ], + "lines": [ + { + "bbox": [ + 53, + 574, + 544, + 681 + ], + "spans": [ + { + "bbox": [ + 53, + 574, + 544, + 681 + ], + "type": "image", + "image_path": "749b5781291848e2186314ec964d16b18c9a9420a7ba8c74a490ece395cc85bf.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 688, + 545, + 711 + ], + "lines": [ + { + "bbox": [ + 46, + 688, + 545, + 711 + ], + "spans": [ + { + "bbox": [ + 46, + 688, + 545, + 711 + ], + "type": "text", + "content": "Figure 4. Comparison of different methods for underwater conditions. The algorithms designed for haze removal (grey) and sand removal (brown) fail in underwater (blue) environments. Our method corrects color shifts, enhance image brightness and improve visibility." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2748" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 52, + 69, + 542, + 165 + ], + "blocks": [ + { + "bbox": [ + 52, + 69, + 542, + 165 + ], + "lines": [ + { + "bbox": [ + 52, + 69, + 542, + 165 + ], + "spans": [ + { + "bbox": [ + 52, + 69, + 542, + 165 + ], + "type": "image", + "image_path": "22b3c6d596c818426b2167537e6aa81c6ff38a2fa3a2cf8520fa08b995c7edcb.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 167, + 547, + 201 + ], + "lines": [ + { + "bbox": [ + 46, + 167, + 547, + 201 + ], + "spans": [ + { + "bbox": [ + 46, + 167, + 547, + 201 + ], + "type": "text", + "content": "Figure 5. Comparison of different methods for sandstorm image enhancement. The algorithms designed for haze removal (grey) and water removal (blue) fail in sandstorm (brown) environments. Our method corrects color shifts, enhances image brightness, and improves contrast, resulting in more vibrant color reproduction. Moreover, LFP produces a more realistic sky area closer to real scenes." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 213, + 289, + 476 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 213, + 289, + 476 + ], + "spans": [ + { + "bbox": [ + 46, + 213, + 289, + 476 + ], + "type": "text", + "content": "Visibility Restoration in Sandstorms. We assess our LFP method's effectiveness in restoring visibility during sandstorms, by using SIRB dataset [46]. Results are shown in Fig 5. Initially, we compare LFP's performance with DCP, MSCNN, and " + }, + { + "bbox": [ + 46, + 213, + 289, + 476 + ], + "type": "inline_equation", + "content": "\\mathrm{Dive + }" + }, + { + "bbox": [ + 46, + 213, + 289, + 476 + ], + "type": "text", + "content": ". The first two fail in dust and sandstorm mitigation, and " + }, + { + "bbox": [ + 46, + 213, + 289, + 476 + ], + "type": "inline_equation", + "content": "\\mathrm{Dive + }" + }, + { + "bbox": [ + 46, + 213, + 289, + 476 + ], + "type": "text", + "content": " introduces noticeable color deviations in sky areas. This emphasizes that hazy and sandstorm images are influenced by distinct optical properties, making dehaze or dewater algorithms unsuitable for sandstorms. Next, we contrast our method with CVC, HRDCP, and ROP. All three significantly enhance visibility in sandstorms. However, CVC lacks contrast, leading to detail loss. HRDCP, despite its visual improvements, exhibits a dim sky and color distortions due to its DCP-based design. ROP performs well but struggles with color distortion, resulting in a cool appearance. LFP overcomes prior challenges by addressing scene depth and attenuation coefficient " + }, + { + "bbox": [ + 46, + 213, + 289, + 476 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 46, + 213, + 289, + 476 + ], + "type": "text", + "content": ". By eliminating hard-to-estimate parameters like atmospheric light and scene depth, LFP avoids common issues, making it robust in sandstorm conditions without relying on assumptions. Thus, our LFP method excels in sandstorm image restoration, producing visually pleasing results without color distortion." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 477, + 287, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 477, + 287, + 586 + ], + "spans": [ + { + "bbox": [ + 46, + 477, + 287, + 586 + ], + "type": "text", + "content": "In summary, the LFP has the ability to restore scenes in a variety of severe weather conditions, a characteristic that is absent in most algorithms. By eliminating the physical parameters that impact transmittance, LFP effectively bypasses the influence of scene depth, thus enhancing visual quality. Moreover, it minimizes color deviations by removing the attenuation coefficient. This approach provides a theoretically and empirically supported a priori method for scene restoration algorithms." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 597, + 190, + 611 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 597, + 190, + 611 + ], + "spans": [ + { + "bbox": [ + 47, + 597, + 190, + 611 + ], + "type": "text", + "content": "4.4. Quantitative Comparison" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 617, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 617, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 617, + 289, + 715 + ], + "type": "text", + "content": "Dehazing Quality We selected DCP, BCCR [34], and FVR [51] as our benchmarks to represent different types of prior-based dehazing methodologies. The rationale behind this selection, instead of opting for deep learning techniques, is multi-fold. Traditional dehazing methods not only possess the capability to restore clear images, but they also notably improve visual effects, thereby enhancing the visual experience. Furthermore, LFP does not contain any learnable pa" + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 308, + 281, + 545, + 347 + ], + "blocks": [ + { + "bbox": [ + 304, + 212, + 547, + 277 + ], + "lines": [ + { + "bbox": [ + 304, + 212, + 547, + 277 + ], + "spans": [ + { + "bbox": [ + 304, + 212, + 547, + 277 + ], + "type": "text", + "content": "Table 2. This table succinctly presents the comparative results of the LFP method and three other distinct prior methods. The evaluation spans across all datasets and metrics, with LFP consistently achieving superior performance. The highest PSNR and SSIM values, represented by bold text, clearly demonstrate the effectiveness of our LFP method." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 308, + 281, + 545, + 347 + ], + "lines": [ + { + "bbox": [ + 308, + 281, + 545, + 347 + ], + "spans": [ + { + "bbox": [ + 308, + 281, + 545, + 347 + ], + "type": "table", + "html": "
MethodsRESIDE-INRESIDE-OUTHaze-4KRS-Haze
PSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIM
DCP16.620.81819.130.81514.010.76017.860.734
BCCR16.880.79114.020.77713.780.75618.840.747
FVR15.720.74811.610.66812.650.68816.120.626
LFP19.760.86921.050.86118.290.87318.240.786
", + "image_path": "8852362e7cd509a395e219750f6ccabd1342db20dd28464559bd5e118cfdc7cc.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 368, + 545, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 368, + 545, + 403 + ], + "spans": [ + { + "bbox": [ + 304, + 368, + 545, + 403 + ], + "type": "text", + "content": "rameters, differentiating it from deep learning approaches. As illustrated in Tab 2, demonstrates that LFP outperforms other methods across all four of our test datasets." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 404, + 545, + 489 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 404, + 545, + 489 + ], + "spans": [ + { + "bbox": [ + 304, + 404, + 545, + 489 + ], + "type": "text", + "content": "The performance is consistent whether we evaluate in terms of PSNR or SSIM. This robust evidence strongly validates the high generalization ability of the LFP method. Furthermore, the LFP method excels in restoring clear, detailed imagery, as particularly demonstrated by the significant improvement in SSIM scores. For more detailed results on desanding and dewatering, please refer to the appendix." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 497, + 545, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 497, + 545, + 617 + ], + "spans": [ + { + "bbox": [ + 304, + 497, + 545, + 617 + ], + "type": "text", + "content": "Runtime Efficiency We conducted extensive comparisons on both CPU and GPU platforms. On the CPU, we compared LFP with traditional dehazing methods; on the GPU, we chose a variety of deep learning dehazing models, including: MSCNN, AOD-Net [23], GFN [41], GCANet [4], GDNet [30], MSBDN [6], FFA [39], CycleD [9], DehazeF [47], and gUNet-T. In terms of runtime, Tab 3 compares the efficiency on both CPU and GPU. The results show that the LFP method consistently achieves the fastest running speed, surpassing existing algorithms." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 625, + 545, + 723 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 625, + 545, + 723 + ], + "spans": [ + { + "bbox": [ + 304, + 625, + 545, + 723 + ], + "type": "text", + "content": "CPU Performance. As evidenced by the data presented in Tab 3, the LFP model consistently exhibits superior runtime performance across all tested resolutions. With a runtime of merely 0.007 seconds at the lower-end resolution of " + }, + { + "bbox": [ + 304, + 625, + 545, + 723 + ], + "type": "inline_equation", + "content": "360\\mathrm{p}" + }, + { + "bbox": [ + 304, + 625, + 545, + 723 + ], + "type": "text", + "content": ", and a still impressive 0.863 seconds at the high-end 4k resolution, these empirical findings underscore the remarkable efficiency of the LFP model when processing large-scale images or performing real-time scene recovery." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2749" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 48, + 135, + 294, + 426 + ], + "blocks": [ + { + "bbox": [ + 47, + 70, + 288, + 126 + ], + "lines": [ + { + "bbox": [ + 47, + 70, + 288, + 126 + ], + "spans": [ + { + "bbox": [ + 47, + 70, + 288, + 126 + ], + "type": "text", + "content": "Table 3. Run-time(seconds) performance. Testing the efficiency of different methods on various single image sizes. Models marked with (*) indicate that the results for those models were generated using MATLAB. The best performance is denoted in red, while the second best performance is indicated in blue." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 48, + 135, + 294, + 426 + ], + "lines": [ + { + "bbox": [ + 48, + 135, + 294, + 426 + ], + "spans": [ + { + "bbox": [ + 48, + 135, + 294, + 426 + ], + "type": "table", + "html": "
CPU Performance
Model360p480p720p1080g2k4k
DCP[16]0.4590.7651.762.6006.7415.11
Retinex[22]0.1490.2860.4780.8042.3675.421
LDCP[55]0.5170.8212.2053.0918.26418.09
CAP[54]0.2180.3100.6661.0212.8666.403
BCCR[34]1.6913.5716.41611.7029.6064.26
FVR[51]0.3013.9028.58912.5834.3576.50
ROP*[27]0.0670.1440.4111.3992.0732.515
LFP0.0070.0170.0730.1510.4070.863
GPU Performance
MSCNN[40]0.1150.1470.2660.4621.1142.481
DehazeNet[3]0.1650.2230.4050.6211.3742.951
AOD-Net[23]0.0760.1060.1750.2910.6961.519
GFN*[41]0.6600.8581.4942.4456.35313.09
GCANet[4]0.4460.6411.1071.8214.2629.266
GDNet[30]0.6440.7421.5032.0274.67510.48
MSBDN[6]0.9521.2122.1943.3358.20317.46
FFA[39]1.6752.2533.9605.97714.2730.61
CycleD[9]0.6070.8101.4282.1585.15511.03
DehazeF[47]0.2890.3870.6771.0322.4755.310
gUNet-T[47]0.1970.2660.4730.7161.7193.669
ROP*[27]0.0470.0510.1130.2090.5320.894
LFP0.0010.0010.0020.0030.0080.019
", + "image_path": "4ed6b6466f7f54abc991f6acf1a650fe7f082c49285f27b8bf92cffabbc5564f.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 446, + 287, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 446, + 287, + 544 + ], + "spans": [ + { + "bbox": [ + 46, + 446, + 287, + 544 + ], + "type": "text", + "content": "GPU Performance. Tab 3 shows that LFPsignificantly outperforms all the compared algorithms in processing speed, especially for high-resolution images. For instance, at a resolution of 4K, LFP's average runtime is merely 0.019 seconds, whereas the runtimes of other algorithms range from 0.894 seconds to a staggering 30.609 seconds. This results in a speedup of up to 34 times for LFP, marking its superior computational efficiency." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 551, + 194, + 563 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 551, + 194, + 563 + ], + "spans": [ + { + "bbox": [ + 47, + 551, + 194, + 563 + ], + "type": "text", + "content": "4.5. Ablation Study on " + }, + { + "bbox": [ + 47, + 551, + 194, + 563 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 47, + 551, + 194, + 563 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 551, + 194, + 563 + ], + "type": "inline_equation", + "content": "\\Omega" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "type": "text", + "content": "Our ablation study further reveals that the changes in the " + }, + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "type": "text", + "content": " parameter not only influence the luminous flux rate of change " + }, + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\Delta F" + }, + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "type": "text", + "content": " but also directly affect the transmission estimation. A properly tuned " + }, + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "type": "text", + "content": " facilitates an accurate estimation of the transmission map, thus effectively reducing the impact of different weather conditions on the image. On the other hand, the size of the sliding window " + }, + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\Omega" + }, + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "type": "text", + "content": " directly affects the local variance " + }, + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\Delta F" + }, + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "type": "text", + "content": ". A larger " + }, + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\Omega" + }, + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "type": "text", + "content": " captures more significant changes in the luminous flux, which is particularly effective in dense haze conditions or when the scene has a greater depth of view. This broadened perspective allows for a more comprehensive understanding of the scene, lead" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 309, + 69, + 547, + 299 + ], + "blocks": [ + { + "bbox": [ + 309, + 69, + 547, + 299 + ], + "lines": [ + { + "bbox": [ + 309, + 69, + 547, + 299 + ], + "spans": [ + { + "bbox": [ + 309, + 69, + 547, + 299 + ], + "type": "image", + "image_path": "96f687e52cf63cb58d7a015b6f8d6e83b5e1864b8c257177420c6948036ecc05.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 308, + 547, + 343 + ], + "lines": [ + { + "bbox": [ + 304, + 308, + 547, + 343 + ], + "spans": [ + { + "bbox": [ + 304, + 308, + 547, + 343 + ], + "type": "text", + "content": "Figure 6. The x-axis represents " + }, + { + "bbox": [ + 304, + 308, + 547, + 343 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 308, + 547, + 343 + ], + "type": "text", + "content": ", the y-axis represents " + }, + { + "bbox": [ + 304, + 308, + 547, + 343 + ], + "type": "inline_equation", + "content": "\\Omega" + }, + { + "bbox": [ + 304, + 308, + 547, + 343 + ], + "type": "text", + "content": ". When the depth of view is further, or the haze is denser, " + }, + { + "bbox": [ + 304, + 308, + 547, + 343 + ], + "type": "inline_equation", + "content": "\\Omega" + }, + { + "bbox": [ + 304, + 308, + 547, + 343 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 308, + 547, + 343 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 308, + 547, + 343 + ], + "type": "text", + "content": " need to be increased correspondingly." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 353, + 547, + 461 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 353, + 547, + 461 + ], + "spans": [ + { + "bbox": [ + 304, + 353, + 547, + 461 + ], + "type": "text", + "content": "ing to a higher-quality image after enhancement. Finally, the interplay between " + }, + { + "bbox": [ + 304, + 353, + 547, + 461 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 353, + 547, + 461 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 353, + 547, + 461 + ], + "type": "inline_equation", + "content": "\\Omega" + }, + { + "bbox": [ + 304, + 353, + 547, + 461 + ], + "type": "text", + "content": " is also crucial. Balancing these two factors allows for a fine-tuned approach to image enhancement, taking into account both the overall brightness and the local variance of the image, this ability to cope with complex conditions adds another layer of robustness to our technique. This balance ensures that the final output image is visually pleasing while maintaining the fidelity of the original scene." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 471, + 379, + 483 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 471, + 379, + 483 + ], + "spans": [ + { + "bbox": [ + 306, + 471, + 379, + 483 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 491, + 547, + 659 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 491, + 547, + 659 + ], + "spans": [ + { + "bbox": [ + 304, + 491, + 547, + 659 + ], + "type": "text", + "content": "In this work, we propose an efficient scene recovery algorithm to restore clear images under poor visibility conditions. By analyzing the image formation model, we derive an analytical form for transmission that eliminates multiple influencing factors and leverages the correlation with flux change rate as prior knowledge. Our approach achieves state-of-the-art performance while maintaining low computational complexity, enabling real-time processing on resource-limited devices. Extensive experiments validate the effectiveness of the proposed method. The novel integration of prior knowledge and analytical transmission modeling provides opportunities to inform future learning-based methods. Overall, this paper makes significant contributions towards robust scene recovery under challenging imaging conditions." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 669, + 415, + 682 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 669, + 415, + 682 + ], + "spans": [ + { + "bbox": [ + 306, + 669, + 415, + 682 + ], + "type": "text", + "content": "6. Acknowledgments" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "content": "This work was supported by the National Natural Science Foundation of China (No. 62336001, No. 62072014)" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2750" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 105, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 105, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 105, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 286, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 53, + 91, + 286, + 123 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 286, + 123 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 286, + 123 + ], + "type": "text", + "content": "[1] Dana Berman, Shai Avidan, et al. Non-local image dehazing. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1674-1682, 2016. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 125, + 286, + 157 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 125, + 286, + 157 + ], + "spans": [ + { + "bbox": [ + 53, + 125, + 286, + 157 + ], + "type": "text", + "content": "[2] Trung Minh Bui and Wonha Kim. Single image dehazing using color ellipsoid prior. IEEE Transactions on Image Processing, 27(2):999-1009, 2017. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 159, + 286, + 201 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 159, + 286, + 201 + ], + "spans": [ + { + "bbox": [ + 54, + 159, + 286, + 201 + ], + "type": "text", + "content": "[3] Bolun Cai, Xiangmin Xu, Kui Jia, Chunmei Qing, and Dacheng Tao. Dehazenet: An end-to-end system for single image haze removal. IEEE Transactions on Image Processing, 25(11):5187-5198, 2016. 1, 3, 8" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 203, + 286, + 267 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 203, + 286, + 267 + ], + "spans": [ + { + "bbox": [ + 54, + 203, + 286, + 267 + ], + "type": "text", + "content": "[4] Dongdong Chen, Mingming He, Qingnan Fan, Jing Liao, Liheng Zhang, Dongdong Hou, Lu Yuan, and Gang Hua. Gated context aggregation network for image dehazing and deraining. In 2019 IEEE winter conference on applications of computer vision (WACV), pages 1375-1383. IEEE, 2019. 7, 8" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 270, + 286, + 313 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 270, + 286, + 313 + ], + "spans": [ + { + "bbox": [ + 53, + 270, + 286, + 313 + ], + "type": "text", + "content": "[5] Qili Deng, Ziling Huang, Chung-Chi Tsai, and Chia-Wen Lin. Hardgan: A haze-aware representation distillation gan for single image dehazing. In European Conference on Computer Vision, pages 722-738. Springer, 2020. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 315, + 286, + 368 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 315, + 286, + 368 + ], + "spans": [ + { + "bbox": [ + 53, + 315, + 286, + 368 + ], + "type": "text", + "content": "[6] Hang Dong, Jinshan Pan, Lei Xiang, Zhe Hu, Xinyi Zhang, Fei Wang, and Ming-Hsuan Yang. Multi-scale boosted dehazing network with dense feature fusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2157-2167, 2020. 3, 7, 8" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 370, + 286, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 370, + 286, + 402 + ], + "spans": [ + { + "bbox": [ + 53, + 370, + 286, + 402 + ], + "type": "text", + "content": "[7] Jiangxin Dong and Jinshan Pan. Physics-based feature dehazing networks. In European Conference on Computer Vision, pages 188-204. Springer, 2020. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 403, + 286, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 403, + 286, + 468 + ], + "spans": [ + { + "bbox": [ + 53, + 403, + 286, + 468 + ], + "type": "text", + "content": "[8] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 470, + 286, + 523 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 470, + 286, + 523 + ], + "spans": [ + { + "bbox": [ + 53, + 470, + 286, + 523 + ], + "type": "text", + "content": "[9] Deniz Engin, Anil Genç, and Hazim Kemal Ekenel. Cycle-dehaze: Enhanced cyclegan for single image dehazing. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition Workshops, pages 825-833, 2018. 7, 8" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 525, + 286, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 525, + 286, + 557 + ], + "spans": [ + { + "bbox": [ + 48, + 525, + 286, + 557 + ], + "type": "text", + "content": "[10] Faming Fang, Fang Li, and Tieyong Zeng. Single image dehazing and denoising: a fast variational approach. SIAM Journal on Imaging Sciences, 7(2):969-996, 2014. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 559, + 286, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 559, + 286, + 612 + ], + "spans": [ + { + "bbox": [ + 48, + 559, + 286, + 612 + ], + "type": "text", + "content": "[11] Xueyang Fu, Yue Huang, Delu Zeng, Xiao-Ping Zhang, and Xinghao Ding. A fusion-based enhancing approach for single sandstorm image. In 2014 IEEE 16th international workshop on multimedia signal processing (MMSP), pages 1-5. IEEE, 2014. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 614, + 286, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 614, + 286, + 668 + ], + "spans": [ + { + "bbox": [ + 48, + 614, + 286, + 668 + ], + "type": "text", + "content": "[12] Han Gao, Ping Wei, and Jun Ke. Color enhancement and image defogging in hsi based on retinax model. In 2015 International Conference on Optical Instruments and Technology: Optoelectronic Imaging and Processing Technology, pages 9-16. SPIE, 2015. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 670, + 286, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 670, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 670, + 286, + 713 + ], + "type": "text", + "content": "[13] Jie Gui, Xiaofeng Cong, Yuan Cao, Wenqi Ren, Jun Zhang, Jing Zhang, Jieuxin Cao, and Dacheng Tao. A comprehensive survey and taxonomy on single image dehazing based on deep learning. ACM Computing Surveys, 55(13s):1-37, 2023. 1" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 127 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 127 + ], + "type": "text", + "content": "[14] Chun-Le Guo, Qixin Yan, Saeed Anwar, Runmin Cong, Wenqi Ren, and Chongyi Li. Image dehazing transformer with transmission-aware 3d position embedding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5812-5820, 2022. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 129, + 545, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 129, + 545, + 171 + ], + "spans": [ + { + "bbox": [ + 308, + 129, + 545, + 171 + ], + "type": "text", + "content": "[15] Yecai Guo, Hanyu Li, and Peixian Zhuang. Underwater image enhancement using a multiscale dense generative adversarial network. IEEE Journal of Oceanic Engineering, 45(3):862-870, 2019. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 173, + 545, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 173, + 545, + 215 + ], + "spans": [ + { + "bbox": [ + 308, + 173, + 545, + 215 + ], + "type": "text", + "content": "[16] Kaiming He, Jian Sun, and Xiaou Tang. Single image haze removal using dark channel prior. IEEE transactions on pattern analysis and machine intelligence, 33(12):2341-2353, 2010. 1, 2, 8" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 217, + 545, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 217, + 545, + 249 + ], + "spans": [ + { + "bbox": [ + 308, + 217, + 545, + 249 + ], + "type": "text", + "content": "[17] Kaiming He, Jian Sun, and Xiaou Tang. Guided image filtering. IEEE transactions on pattern analysis and machine intelligence, 35(6):1397-1409, 2012. 5" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 251, + 545, + 292 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 251, + 545, + 292 + ], + "spans": [ + { + "bbox": [ + 308, + 251, + 545, + 292 + ], + "type": "text", + "content": "[18] Jong-Ju Jeon, Tae-Hee Park, and Il-Kyu Eom. Sand-dust image enhancement using chromatic variance consistency and gamma correction-based dehazing. Sensors, 22(23):9048, 2022. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 294, + 545, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 294, + 545, + 326 + ], + "spans": [ + { + "bbox": [ + 308, + 294, + 545, + 326 + ], + "type": "text", + "content": "[19] Mingye Ju, Can Ding, Y Jay Guo, and Dengyin Zhang. Idgcp: Image dehazing based on gamma correction prior. IEEE Transactions on Image Processing, 29:3104-3118, 2019. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 327, + 545, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 327, + 545, + 360 + ], + "spans": [ + { + "bbox": [ + 308, + 327, + 545, + 360 + ], + "type": "text", + "content": "[20] Manjit Kaur, Dilbag Singh, Vijay Kumar, and Kehui Sun. Color image dehazing using gradient channel prior and guided 10 filter. Information Sciences, 521:326-342, 2020. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 361, + 545, + 414 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 361, + 545, + 414 + ], + "spans": [ + { + "bbox": [ + 308, + 361, + 545, + 414 + ], + "type": "text", + "content": "[21] Johannes Kopf, Boris Neubert, Billy Chen, Michael Cohen, Daniel Cohen-Or, Oliver Deussen, Matt Uytendaele, and Dani Lischinski. Deep photo: Model-based photograph enhancement and viewing. ACM transactions on graphics (TOG), 27(5):1-10, 2008. 4" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 415, + 545, + 436 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 415, + 545, + 436 + ], + "spans": [ + { + "bbox": [ + 308, + 415, + 545, + 436 + ], + "type": "text", + "content": "[22] Edwin H Land. The retina theory of color vision. Scientific american, 237(6):108-129, 1977. 4, 8" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 438, + 545, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 438, + 545, + 480 + ], + "spans": [ + { + "bbox": [ + 308, + 438, + 545, + 480 + ], + "type": "text", + "content": "[23] Boyi Li, Xiulian Peng, Zhangyang Wang, Jizheng Xu, and Dan Feng. Aod-net: All-in-one dehazing network. In Proceedings of the IEEE international conference on computer vision, pages 4770-4778, 2017. 1, 3, 7, 8" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 483, + 545, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 483, + 545, + 525 + ], + "spans": [ + { + "bbox": [ + 308, + 483, + 545, + 525 + ], + "type": "text", + "content": "[24] Boyi Li, Wenqi Ren, Dengpan Fu, Dacheng Tao, Dan Feng, Wenjun Zeng, and Zhangyang Wang. Benchmarking single-image dehazing and beyond. IEEE Transactions on Image Processing, 28(1):492-505, 2019. 3, 6" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 526, + 545, + 569 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 526, + 545, + 569 + ], + "spans": [ + { + "bbox": [ + 308, + 526, + 545, + 569 + ], + "type": "text", + "content": "[25] Chongyi Li, Chunle Guo, Wenqi Ren, Runmin Cong, Junhui Hou, Sam Kwong, and Dacheng Tao. An underwater image enhancement benchmark dataset and beyond. IEEE Transactions on Image Processing, 29:4376-4389, 2019. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 571, + 545, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 571, + 545, + 612 + ], + "spans": [ + { + "bbox": [ + 308, + 571, + 545, + 612 + ], + "type": "text", + "content": "[26] Chongyi Li, Saeed Anwar, and Fatih Porikli. Underwater scene prior inspired deep underwater image and video enhancement. Pattern Recognition, 98:107038, 2020. 1, 3, 6" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 614, + 545, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 614, + 545, + 657 + ], + "spans": [ + { + "bbox": [ + 308, + 614, + 545, + 657 + ], + "type": "text", + "content": "[27] Jun Liu, Wen Liu, Jianing Sun, and Tieyong Zeng. Rank-one prior: Toward real-time scene recovery. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14802-14810, 2021. 1, 2, 8" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 658, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 658, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 308, + 658, + 545, + 712 + ], + "type": "text", + "content": "[28] Risheng Liu, Xin Fan, Ming Zhu, Minjun Hou, and Zhongxuan Luo. Real-world underwater enhancement: Challenges, benchmarks, and solutions under natural light. IEEE transactions on circuits and systems for video technology, 30(12): 4861-4875, 2020. 3" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 314, + 757 + ], + "type": "text", + "content": "2751" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "text", + "content": "[29] Xiaodong Liu, Zhi Gao, and Ben M Chen. Mlfcgan: Multilevel feature fusion-based conditional gan for underwater image color correction. IEEE Geoscience and Remote Sensing Letters, 17(9):1488-1492, 2019. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 118, + 288, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 118, + 288, + 171 + ], + "spans": [ + { + "bbox": [ + 48, + 118, + 288, + 171 + ], + "type": "text", + "content": "[30] Xiaohong Liu, Yongrui Ma, Zhihao Shi, and Jun Chen. Griddehazenet: Attention-based multi-scale network for image dehazing. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7314-7323, 2019. 3, 7, 8" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 174, + 287, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 174, + 287, + 228 + ], + "spans": [ + { + "bbox": [ + 48, + 174, + 287, + 228 + ], + "type": "text", + "content": "[31] Yu Liu, Guanlong Zhao, Boyuan Gong, Yang Li, Ritu Raj, Niraj Goel, Satya Kesav, Sandeep Gottimukkala, Zhangyang Wang, Wenqi Ren, et al. Improved techniques for learning to dehaze and beyond: A collective study. arXiv preprint arXiv:1807.00202, 2018. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 230, + 287, + 284 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 230, + 287, + 284 + ], + "spans": [ + { + "bbox": [ + 48, + 230, + 287, + 284 + ], + "type": "text", + "content": "[32] Ye Liu, Lei Zhu, Shunda Pei, Huazhu Fu, Jing Qin, Qing Zhang, Liang Wan, and Wei Feng. From synthetic to real: Image dehazing collaborating with unlabeled real data. In Proceedings of the 29th ACM International Conference on Multimedia, pages 50-58, 2021. 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 285, + 287, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 285, + 287, + 308 + ], + "spans": [ + { + "bbox": [ + 48, + 285, + 287, + 308 + ], + "type": "text", + "content": "[33] Rastislav Lukac. Single-sensor imaging: methods and applications for digital cameras. CRC Press, 2018. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 309, + 287, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 309, + 287, + 363 + ], + "spans": [ + { + "bbox": [ + 48, + 309, + 287, + 363 + ], + "type": "text", + "content": "[34] Gaofeng Meng, Ying Wang, Jiangyong Duan, Shiming Xiang, and Chunhong Pan. Efficient image dehazing with boundary constraint and contextual regularization. In Proceedings of the IEEE international conference on computer vision, pages 617-624, 2013. 2, 7, 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 365, + 287, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 365, + 287, + 386 + ], + "spans": [ + { + "bbox": [ + 48, + 365, + 287, + 386 + ], + "type": "text", + "content": "[35] WE Knowles Middleton. Bouguer, lambert, and the theory of horizontal visibility. Isis, 51(2):145-149, 1960. 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 388, + 287, + 419 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 388, + 287, + 419 + ], + "spans": [ + { + "bbox": [ + 48, + 388, + 287, + 419 + ], + "type": "text", + "content": "[36] Srinivasa G Narasimhan and Shree K Nayar. Vision and the atmosphere. International journal of computer vision, 48(3): 233-254, 2002. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 422, + 287, + 465 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 422, + 287, + 465 + ], + "spans": [ + { + "bbox": [ + 48, + 422, + 287, + 465 + ], + "type": "text", + "content": "[37] Srinivasa G Narasimhan and Shree K Nayar. Interactive (de) weathering of an image using physical models. In IEEE Workshop on color and photometric Methods in computer Vision, page 1. France, 2003. 4" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 467, + 287, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 467, + 287, + 498 + ], + "spans": [ + { + "bbox": [ + 48, + 467, + 287, + 498 + ], + "type": "text", + "content": "[38] Ko Nishino, Louis Kratz, and Stephen Lombardi. Bayesian defogging. International journal of computer vision, 98(3): 263-278, 2012. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 501, + 287, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 501, + 287, + 544 + ], + "spans": [ + { + "bbox": [ + 48, + 501, + 287, + 544 + ], + "type": "text", + "content": "[39] Xu Qin, Zhilin Wang, Yuanchao Bai, Xiaodong Xie, and Huizhu Jia. Ffa-net: Feature fusion attention network for single image dehazing. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 11908-11915, 2020. 3, 7, 8" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 545, + 287, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 545, + 287, + 590 + ], + "spans": [ + { + "bbox": [ + 48, + 545, + 287, + 590 + ], + "type": "text", + "content": "[40] Wenqi Ren, Si Liu, Hua Zhang, Jinshan Pan, Xiaochun Cao, and Ming-Hsuan Yang. Single image dehazing via multiscale convolutional neural networks. In European conference on computer vision, pages 154-169. Springer, 2016. 1, 2, 8" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 591, + 287, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 591, + 287, + 644 + ], + "spans": [ + { + "bbox": [ + 48, + 591, + 287, + 644 + ], + "type": "text", + "content": "[41] Wenqi Ren, Lin Ma, Jiawei Zhang, Jinshan Pan, Xiaochun Cao, Wei Liu, and Ming-Hsuan Yang. Gated fusion network for single image dehazing. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 3253-3261, 2018. 7, 8" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 647, + 287, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 647, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 647, + 287, + 689 + ], + "type": "text", + "content": "[42] Wenqi Ren, Jingang Zhang, Xiangyu Xu, Lin Ma, Xiaochun Cao, Gaofeng Meng, and Wei Liu. Deep video dehazing with semantic segmentation. IEEE Transactions on Image Processing, 28(4):1895-1908, 2018. 1" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "type": "text", + "content": "[43] Christos Sakaridis, Dengxin Dai, Simon Hecker, and Luc Van Gool. Model adaptation with synthetic and real data for" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 545, + 599 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 327, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 545, + 106 + ], + "type": "text", + "content": "semantic dense foggy scene understanding. In Proceedings of the European Conference on Computer Vision (ECCV), pages 687-704, 2018. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 107, + 545, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 107, + 545, + 150 + ], + "spans": [ + { + "bbox": [ + 308, + 107, + 545, + 150 + ], + "type": "text", + "content": "[44] Yoav Y Schechner and Yuval Averbuch. Regularized image recovery in scattering media. IEEE Transactions on Pattern Analysis and Machine Intelligence, 29(9):1655-1660, 2007. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 152, + 545, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 152, + 545, + 205 + ], + "spans": [ + { + "bbox": [ + 308, + 152, + 545, + 205 + ], + "type": "text", + "content": "[45] Zhenghao Shi, Yaning Feng, Minghua Zhao, Erhu Zhang, and Lifeng He. Let you see in sand dust weather: A method based on halo-reduced dark channel prior dehazing for sand-dust image enhancement. *Ieee Access*, 7:116722–116733, 2019. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 208, + 545, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 208, + 545, + 251 + ], + "spans": [ + { + "bbox": [ + 308, + 208, + 545, + 251 + ], + "type": "text", + "content": "[46] Yazhong Si, Fan Yang, Ya Guo, Wei Zhang, and Yipu Yang. A comprehensive benchmark analysis for sand dust image reconstruction. Journal of Visual Communication and Image Representation, 89:103638, 2022. 3, 6, 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 253, + 545, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 253, + 545, + 285 + ], + "spans": [ + { + "bbox": [ + 308, + 253, + 545, + 285 + ], + "type": "text", + "content": "[47] Yuda Song, Zhuqing He, Hui Qian, and Xin Du. Vision transformers for single image dehazing. arXiv preprint arXiv:2204.03883, 2022. 3, 7, 8" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 286, + 545, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 286, + 545, + 319 + ], + "spans": [ + { + "bbox": [ + 308, + 286, + 545, + 319 + ], + "type": "text", + "content": "[48] Yuda Song, Yang Zhou, Hui Qian, and Xin Du. Rethinking performance gains in image dehazing networks. arXiv preprint arXiv:2209.11448, 2022. 1, 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 320, + 545, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 320, + 545, + 364 + ], + "spans": [ + { + "bbox": [ + 308, + 320, + 545, + 364 + ], + "type": "text", + "content": "[49] Matan Sulami, Itamar Glatzer, Raanan Fattal, and Mike Werman. Automatic recovery of the atmospheric light in hazy images. In 2014 IEEE International Conference on Computational Photography (ICCP), pages 1-11. IEEE, 2014. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 365, + 545, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 365, + 545, + 398 + ], + "spans": [ + { + "bbox": [ + 308, + 365, + 545, + 398 + ], + "type": "text", + "content": "[50] Robby T Tan. Visibility in bad weather from a single image. In 2008 IEEE Conference on Computer Vision and Pattern Recognition, pages 1-8. IEEE, 2008. 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 399, + 545, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 399, + 545, + 442 + ], + "spans": [ + { + "bbox": [ + 308, + 399, + 545, + 442 + ], + "type": "text", + "content": "[51] Jean-Philippe Tarel and Nicolas Hautiere. Fast visibility restoration from a single color or gray level image. In 2009 IEEE 12th International Conference on Computer Vision, pages 2201–2208. IEEE, 2009. 2, 7, 8" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 444, + 545, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 444, + 545, + 475 + ], + "spans": [ + { + "bbox": [ + 308, + 444, + 545, + 475 + ], + "type": "text", + "content": "[52] Ting Yan, Liejun Wang, and Jiaxing Wang. Method to enhance degraded image in dust environment. J. Softw., 9(10): 2672-2677, 2014. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 478, + 545, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 478, + 545, + 510 + ], + "spans": [ + { + "bbox": [ + 308, + 478, + 545, + 510 + ], + "type": "text", + "content": "[53] Dong Zhao, Jia Li, Hongyu Li, and Long Xu. Hybrid local-global transformer for image dehazing. arXiv preprint arXiv:2109.07100, 2021. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 512, + 545, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 512, + 545, + 555 + ], + "spans": [ + { + "bbox": [ + 308, + 512, + 545, + 555 + ], + "type": "text", + "content": "[54] Qingsong Zhu, Jiaming Mai, and Ling Shao. A fast single image haze removal algorithm using color attenuation prior. IEEE transactions on image processing, 24(11):3522-3533, 2015. 1, 2, 8" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 556, + 545, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 556, + 545, + 599 + ], + "spans": [ + { + "bbox": [ + 308, + 556, + 545, + 599 + ], + "type": "text", + "content": "[55] Yingying Zhu, Gaoyang Tang, Xiaoyan Zhang, Jianmin Jiang, and Qi Tian. Haze removal method for natural restoration of images with sky. neurocomputing, 275:499-510, 2018. 2, 8" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2752" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Efficient Solution of Point-Line Absolute Pose/8561d28f-e735-46cc-9618-c6c206bdafc2_content_list.json b/2024/Efficient Solution of Point-Line Absolute Pose/8561d28f-e735-46cc-9618-c6c206bdafc2_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..a41b22d03191365938609991ef2c076487176e28 --- /dev/null +++ b/2024/Efficient Solution of Point-Line Absolute Pose/8561d28f-e735-46cc-9618-c6c206bdafc2_content_list.json @@ -0,0 +1,2673 @@ +[ + { + "type": "text", + "text": "Efficient Solution of Point-Line Absolute Pose", + "text_level": 1, + "bbox": [ + 254, + 130, + 718, + 150 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Petr Hruby ETH Zürich", + "bbox": [ + 184, + 181, + 282, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Timothy Duff \nUniversity of Washington", + "bbox": [ + 336, + 181, + 542, + 217 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Marc Pollefeys \nETH Zürich, Microsoft", + "bbox": [ + 591, + 181, + 779, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 251, + 313, + 266 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We revisit certain problems of pose estimation based on 3D-2D correspondences between features which may be points or lines. Specifically, we address the two previously-studied minimal problems of estimating camera extrinsics from $p \\in \\{1,2\\}$ point-point correspondences and $l = 3 - p$ line-line correspondences. To the best of our knowledge, all of the previously-known practical solutions to these problems required computing the roots of degree $\\geq 4$ (univariate) polynomials when $p = 2$ , or degree $\\geq 8$ polynomials when $p = 1$ . We describe and implement two elementary solutions which reduce the degrees of the needed polynomials from 4 to 2 and from 8 to 4, respectively. We show experimentally that the resulting solvers are numerically stable and fast: when compared to the previous state-of-the-art, we may obtain nearly an order of magnitude speedup. The code is available at https://github.com/petrhruby97/efficient_absolute", + "bbox": [ + 75, + 282, + 473, + 541 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 565, + 209, + 580 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1.1. Motivation", + "text_level": 1, + "bbox": [ + 76, + 590, + 199, + 604 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The problem of registering images to a known 3D coordinate system plays a crucial role in applications such as visual localization [26], autonomous driving [12], and augmented reality [28], as well as in general paradigms like SLAM [15] and SfM [27]. Robust estimators based on RANSAC [11], or one of its many refinements [3, 24], are among the most successful tools for solving these problems. Such an estimator traditionally relies on a minimal P3P solver [8, 13, 16, 23] to efficiently hypothesize poses from putative matches between 3D and 2D points.", + "bbox": [ + 75, + 613, + 468, + 763 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The literature on P3P and other purely point-based methods for absolute pose estimation is vast. Absolute pose estimation from non-point features such as lines [2, 5, 7, 25, 29, 30], point-line incidences [10], and affine correspondences [28], has received comparatively less attention, but remains an active research area. In particular, solutions relying on both points and lines are of increasing importance, due to the prevalence of both types of feature in man-made environments, as well as several recent advances in the com", + "bbox": [ + 75, + 763, + 470, + 901 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/92879d98ea4ada4118f256e76bcc5e1846b9d6f5e2287fa51eb66b458b2b3d8e.jpg", + "image_caption": [ + "Figure 1. The P2P1L (Top) and P1P2L (Bottom) problems." + ], + "image_footnote": [], + "bbox": [ + 531, + 247, + 867, + 460 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ponents of 3D reconstruction systems responsible for line detection [21], matching [22], and bundle adjustment [19].", + "bbox": [ + 498, + 512, + 890, + 544 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1.2. Contribution", + "text_level": 1, + "bbox": [ + 500, + 556, + 637, + 570 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this paper, we revisit the two minimal problems of absolute pose estimation that combine both points and lines: the Perspective-2-Point-1-Line Problem (or P2P1L, Section 2.1), and the Perspective-1-Point-2-Line Problem (P1P2L, Section 2.2.) See Figure 1 for illustrations.", + "bbox": [ + 496, + 580, + 890, + 656 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In contrast to the purely point or line based minimal problems P3P and P3L, we observe that the existing solutions for the \"mixed cases\" considered here are suboptimal from both a theoretical and practical point of view. Consequently, we develop novel solutions to both problems, which optimally exploit their underlying algebraic structure, and exhibit comparable or better performance than the state of the art on simulated and real data—see Section 3. On the other hand, these solvers are simple to implement, and no knowledge of mathematics beyond elementary algebra is needed to understand them.", + "bbox": [ + 496, + 657, + 892, + 823 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To provide further context for our work, we include the following quotation from [25]: Although we do not theoretically prove that our solutions are of the lowest possible degrees, we believe so because of the following argument. The best existing solutions for pose estimation using three", + "bbox": [ + 496, + 824, + 893, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "21316", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "points and three lines use 4th and 8th degree solutions respectively. Since mixed cases are in the middle, our solutions for (2 points, 1 line) and (1 point, 2 lines) cases use 4th and 8th degree solutions respectively. Recently, it was shown using Galois theory that the solutions that use the lowest possible degrees are the optimal ones [[20]].", + "bbox": [ + 75, + 90, + 468, + 181 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Contrary to the informal reasoning presented above, we claim that the existing solutions to the mixed point-line cases are not optimal. As far as we know, the only support for this claim appearing in the literature prior to our work occurs in [9, §3]. This previous work showed, on the basis of Galois group computation, that the P2P1L and P1P2L problems decompose into simpler subproblems. However, this theoretical observation was not accompanied by a practical solution method for either problem. In this paper, we rectify the situation by devising practical solvers for both problems that incorporate these recent insights.", + "bbox": [ + 75, + 184, + 470, + 349 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To complete our discussion of what makes a solution algebraically optimal, we recall one proposal of such a notion from work of Nister et al. [20]. In this work, a restricted class of algorithms $\\mathcal{P}_n$ is considered for each natural number $n \\geq 1$ : an algorithm in $\\mathcal{P}_n$ consists of a finite sequence of steps, each of which extracts the roots of some polynomial equation $p(x) = 0$ , with either $p(x) = x^m - a$ (where $a \\in \\mathbb{Q}$ and $\\deg(p) = m$ can be arbitrary), or with $\\deg(p) \\leq n$ and the coefficients of $p$ belonging to a field containing $\\mathbb{Q}$ and any previously-computed roots. In this setting, a solution is optimal if it leads to solving a polynomial system of the lowest possible degree. Our proposed solutions immediately establish that P2P1L is $\\mathcal{P}_2$ -solvable and P1P2L is $\\mathcal{P}_4$ -solvable. It should be of little surprise that P2P1L is not $\\mathcal{P}_1$ -solvable; therefore, we may say that our solution to P2P1L is algebraically optimal. On the other hand, the problem P1P2L is also $\\mathcal{P}_3$ -solvable. This is because any quartic equation can be solved by the standard method which reduces the problem to computing the roots of the associated resolvent cubic. Thus, our solution, which computes the roots of a quartic with this same method, is also algebraically optimal in the sense of [20]. The same observation, of course, holds for the classical quartic-based methods for solving P3P. Alternative P3P solvers that directly employ a cubic [8, 23], despite being superior in practical terms, are not distinguished by the complexity classes $\\mathcal{P}_n$ . Table 1 provides a summary of the algebraic complexities, based on the Galois groups computed in [9, §3].", + "bbox": [ + 75, + 352, + 470, + 776 + ], + "page_idx": 1 + }, + { + "type": "table", + "img_path": "images/03b944cedaa670d68124a363866ae8e74ed32c07c8bed66f2db52374c0e03c99.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
problemP3PP2P1LP1P2LP3L
classP3P2P3P8
", + "bbox": [ + 124, + 792, + 424, + 825 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Table 1. For each minimal absolute pose problem with points and line features, the class of polynomial root finding algorithms $\\mathcal{P}_n$ that solve the problem with $n$ as small as possible [20, §2.1].", + "bbox": [ + 75, + 835, + 468, + 878 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "1.3. Related work", + "text_level": 1, + "bbox": [ + 501, + 90, + 638, + 104 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The first solutions to P2P1L and P1P2L were presented by Ramalingam et al. [25], reducing the problems to computing the roots of polynomials of degree 4 and 8 via careful choices of special reference frames in the world and image. Although we obtain polynomials of lower degree, these special reference frames are also an ingredient in our approach.", + "bbox": [ + 496, + 114, + 890, + 205 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In work subsequent to [25], it was observed that both of these problems could be solved using the E3Q3 solver [17]. This is a highly optimized method for computing the points where three quadric surfaces in $\\mathbb{R}^3$ intersect. Typically, there are $2^{3} = 8$ such points, by Bézout's theorem; however, EQ3Q also handles degenerate cases where the number of solutions may drop. In the follow-up work [30], a stabilization scheme for E3Q3 was applied to these problems, and experimentally shown to be more accurate than the solvers in [25]. Efficient solvers for both problems based on this stabilized E3Q3 are implemented in PoseLib [18].", + "bbox": [ + 496, + 205, + 892, + 371 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The idea of using Galois groups to study minimal problems originates from [20]. The main takeaway from this paper is that a problem with $n$ solutions whose Galois group is the full symmetric group $S_{n}$ cannot be solved by an algorithm in $\\mathcal{P}_{n-1}$ . In the other direction, the main takeaway from [9] is that if the Galois group is contained in the wreath product $S_{n_{1}} \\wr S_{n_{2}}$ , where $n = n_{1}n_{2}$ , then the problem can be solved with an algorithm in $\\mathcal{P}_{\\max(n_{1},n_{2})}$ . A recent work using such an insight to guide the more efficient solution of a relative pose estimation problem may be found in [14].", + "bbox": [ + 496, + 371, + 893, + 523 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Minimal solvers", + "text_level": 1, + "bbox": [ + 500, + 537, + 660, + 553 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this section, we introduce our algebraically optimal solutions to the P2P1L and P1P2L problems. These problems are depicted in Figure 1. As revealed by the Galois group computed in [9, §3], the problem P2P1L can be reduced to computing the roots of a quadratic equation, and the problem P1P2L can be reduced to a quartic equation. In Sections 2.1 and 2.2, we turn these insights into explicit solutions to the P2P1L and P12PL problems, respectively. These solutions work generically—they are valid outside of a measure-zero subset of the space of input point-point/line-line correspondences. In Section 2.3, we describe a method that stabilizes the P1P2L solver of Section 2.2 in a common but non-generic case. Finally, Section 2.4 for a discussion when the point-line configuration is coplanar.", + "bbox": [ + 496, + 564, + 890, + 776 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. P2P1L", + "text_level": 1, + "bbox": [ + 500, + 785, + 589, + 800 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Here, we provide the algebraically optimal solution to the P2P1L problem. We mostly follow the notation of [25].", + "bbox": [ + 496, + 809, + 890, + 839 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Problem Statement: Consider $P_1^0, P_2^0 \\in \\mathbb{R}^3$ , two 3D points, and their 2D projections under an unknown calibrated camera. Consider also a 3D line spanned by points $L_3^0, L_4^0 \\in \\mathbb{R}^3$ and its projection. The goal is to recover the", + "bbox": [ + 496, + 839, + 890, + 901 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "21317", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/20445638a11240eeb032dbf9aa20b68a3f5e0ff62a2dfbc3eea5f7ee8d5f1d57.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 107, + 90, + 261, + 200 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/43c16aa4a49f8858a75473230dd5eaa7300de3b17ad8be6e8c7c274b0f96d3ad.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 285, + 90, + 429, + 200 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/7ab54c24980c7d8f3dfccb159e7c7f38c5184844ac9305c9b364214ab8856d71.jpg", + "image_caption": [ + "Figure 2. Special reference frame for P2P1L. Camera frame $\\mathcal{C}_0$ is transformed to $\\mathcal{C}_1$ , world frame $\\mathcal{W}_0$ to $\\mathcal{W}_1$ . See text for details." + ], + "image_footnote": [], + "bbox": [ + 109, + 202, + 254, + 291 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/e709c755edc91e6f86341c74511e1079caf27c378108579d7dc7e3a31ed5ea1e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 287, + 202, + 426, + 290 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "unknown camera matrix, whose center we denote by $C^0$", + "bbox": [ + 76, + 352, + 452, + 368 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our first step is to rigidly transform the input into the special reference frame introduced in [25]. Transforming to this special frame simplifies the equations and helps to reveal the algebraically optimal solution. Figure 2 illustrates the input reference frame $\\mathcal{C}_0 / \\mathcal{W}_0$ (top), and the special frame $\\mathcal{C}_1 / \\mathcal{W}_1$ (bottom), which are related as follows:", + "bbox": [ + 76, + 368, + 468, + 458 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "- In the world frame $\\mathcal{W}_1$ , the 3D points take the form", + "bbox": [ + 76, + 459, + 431, + 473 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nP _ {1} = \\left( \\begin{array}{l l l} 0 & 0 & 0 \\end{array} \\right) ^ {T}, \\quad P _ {2} = \\left( \\begin{array}{l l l} X _ {2} & 0 & 0 \\end{array} \\right) ^ {T}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 132, + 483, + 468, + 506 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "and the 3D line is spanned by two points of the form", + "bbox": [ + 89, + 516, + 439, + 531 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nL _ {3} = \\left( \\begin{array}{l l l} X _ {3} & Y _ {3} & 0 \\end{array} \\right) ^ {T}, \\quad L _ {4} = \\left( \\begin{array}{l l l} X _ {4} & Y _ {4} & Z _ {4} \\end{array} \\right) ^ {T}. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 102, + 540, + 468, + 561 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To find the transformation $\\mathcal{W}_0\\to \\mathcal{W}_1$ , we set", + "bbox": [ + 76, + 571, + 393, + 587 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nP _ {1} = \\left( \\begin{array}{c c c} 0 & 0 & 0 \\end{array} \\right) ^ {T}, \\quad P _ {2} = \\left(\\| P _ {1} ^ {0} - P _ {2} ^ {2} \\| \\quad 0 \\quad 0\\right) ^ {T},\n$$\n", + "text_format": "latex", + "bbox": [ + 104, + 595, + 452, + 618 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} L _ {3} = \\left( \\begin{array}{c c c} X _ {3} & Y _ {3} & 0 \\end{array} \\right) ^ {T}, \\mathrm {w /} X _ {3} = \\left(L _ {3} ^ {0} - P _ {1}\\right) ^ {T} \\frac {P _ {2} - P _ {1}}{\\| P _ {2} - P _ {1} \\|}, \\\\ Y _ {3} = \\left\\| L _ {3} ^ {0} - P _ {1} - X _ {3} \\cdot \\frac {P _ {2} - P _ {1}}{\\| P _ {2} - P _ {1} \\|} \\right\\|. \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 91, + 627, + 468, + 698 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We transform the rays $\\overrightarrow{P_1^0P_2^0}$ and $\\overrightarrow{P_1^0L_3^0}$ to $\\overrightarrow{P_1P_2}$ and $\\overrightarrow{P_1L_3}$ , respectively, via suitable translation and rotation.", + "bbox": [ + 89, + 703, + 468, + 741 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "- In the camera frame, the camera center is fixed at", + "bbox": [ + 76, + 742, + 416, + 755 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nC = \\left( \\begin{array}{l l l} 0 & 0 & - 1 \\end{array} \\right) ^ {T}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 207, + 765, + 468, + 786 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "the 2D points are projections of points of the form", + "bbox": [ + 89, + 797, + 423, + 811 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nD _ {1} = \\left( \\begin{array}{l l l} a _ {1} & b _ {1} & 0 \\end{array} \\right) ^ {T}, \\quad D _ {2} = \\left( \\begin{array}{l l l} a _ {2} & b _ {2} & 0 \\end{array} \\right) ^ {T}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 112, + 821, + 468, + 843 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "and the 2D line is the projection of $\\overline{D_3D_4}$ , where", + "bbox": [ + 89, + 854, + 416, + 871 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nD _ {3} = \\left( \\begin{array}{l l l} 0 & 0 & 0 \\end{array} \\right) ^ {T}, \\quad D _ {4} = \\left( \\begin{array}{l l l} 1 & 0 & 0 \\end{array} \\right) ^ {T}. \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 135, + 881, + 468, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "- To find the transformation $\\mathcal{C}_0\\to \\mathcal{C}_1$ , let $D_3^0$ and $D_4^0$ be the homogeneous coordinates of two distinct points along the given line. Independently of $C^0$ , we may rigidly transform the rays $d_{3}\\coloneqq \\overrightarrow{C^{0}D_{3}^{0}}$ , and $d_4\\coloneqq \\overrightarrow{C^0D_3^0}$ to $\\overrightarrow{CD_3}$ and $\\overrightarrow{CD_4}$ , by suitable rotation and translation.", + "bbox": [ + 500, + 90, + 890, + 172 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Let us now write $R = (R_{ij})_{1 \\leq i,j \\leq 3}$ , $T = (T_i)_{1 \\leq i \\leq 3}$ for the unknown camera pose in this special reference frame. Our projection constraints can then be written as", + "bbox": [ + 498, + 172, + 890, + 218 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\left(D _ {i} - C\\right) \\times \\left(R P _ {i} + T\\right) = 0, \\quad i \\in \\{1, 2 \\}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 553, + 231, + 890, + 247 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\left(D _ {3} \\times D _ {4}\\right) ^ {T} \\left(R L _ {i} + T\\right) = 0, \\quad i \\in \\{3, 4 \\}. \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 553, + 250, + 890, + 268 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "This gives a system of equations in 12 unknowns. However, in what follows, we consider a system of equations in a smaller set of unknowns, namely", + "bbox": [ + 498, + 280, + 890, + 325 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {x} ^ {T} = \\left( \\begin{array}{l l l l l l l l} R _ {1 1} & R _ {2 1} & R _ {3 1} & R _ {2 2} & R _ {2 3} & T _ {1} & T _ {2} & T _ {3} \\end{array} \\right). \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 511, + 335, + 890, + 368 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The entries of $\\mathbf{x}$ are constrained by", + "bbox": [ + 500, + 369, + 733, + 383 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nR _ {1 1} ^ {2} + R _ {2 1} ^ {2} + R _ {3 1} ^ {2} = R _ {2 1} ^ {2} + R _ {2 2} ^ {2} + R _ {2 3} ^ {2} = 1. \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 545, + 395, + 890, + 414 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Thus, the entries of $\\mathbf{x}$ specify a translation vector and a partially-filled rotation matrix, which can be uniquely completed to a rotation matrix using the formulae", + "bbox": [ + 498, + 425, + 890, + 470 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\binom {R _ {1 2}} {R _ {1 3}} = \\left(R _ {2 2} ^ {2} + R _ {2 3} ^ {2}\\right) ^ {- 1} \\cdot \\binom {- R _ {1 1} R _ {2 1} R _ {2 2} + R _ {2 3} R _ {3 1}} {- R _ {1 1} R _ {2 1} R _ {2 3} - R _ {2 2} R _ {3 1}} \\\\ - R _ {2 1} R _ {2 2} R _ {3 1} - R _ {1 1} R _ {2 3} \\\\ - R _ {2 1} R _ {2 3} R _ {3 1} + R _ {1 1} R _ {2 2}) \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 501, + 481, + 892, + 560 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "subject to the genericity condition $R_{22}^2 + R_{23}^2 \\neq 0$ , i.e. when $R$ is not a rotation in the $xz$ -plane. Hence, we focus on recovering the entries of $\\mathbf{x}$ . From 2 out of the 3 redundant constraints from (6) together with (7), we obtain linear constraints on $\\mathbf{x}$ ([25, eq. (4)-(5)]),", + "bbox": [ + 498, + 573, + 890, + 650 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nA \\mathbf {x} = \\mathbf {b}, \\quad \\text {w h e r e}\n$$\n", + "text_format": "latex", + "bbox": [ + 547, + 662, + 669, + 676 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nA = \\left( \\begin{array}{c c c c c c c c} 0 & 0 & 0 & 0 & 0 & - b _ {1} & a _ {1} & 0 \\\\ 0 & 0 & 0 & 0 & 0 & 0 & - 1 & b _ {1} \\\\ - b _ {2} X _ {2} & a _ {2} X _ {2} & 0 & 0 & 0 & - b _ {2} & a _ {2} & 0 \\\\ 0 & - X _ {2} & b _ {2} X _ {2} & 0 & 0 & 0 & - 1 & b _ {2} \\\\ 0 & X _ {3} & 0 & Y _ {3} & 0 & 0 & 1 & 0 \\\\ 0 & X _ {4} & 0 & Y _ {4} & Z _ {4} & 0 & 1 & 0 \\end{array} \\right),\n$$\n", + "text_format": "latex", + "bbox": [ + 557, + 681, + 844, + 739 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {b} ^ {T} = \\left( \\begin{array}{c c c c c c} 0 & - b _ {1} & 0 & - b _ {2} & 0 & 0 \\end{array} \\right). \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 547, + 739, + 890, + 758 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We may use these linear equations to solve for the translation $T$ in terms of $R$ and the problem data as follows:", + "bbox": [ + 500, + 771, + 888, + 800 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nT _ {1} = a _ {1} B, \\quad T _ {2} = b _ {1} B, \\quad T _ {3} = - 1 + B, \\quad \\text {w h e r e}\n$$\n", + "text_format": "latex", + "bbox": [ + 522, + 813, + 866, + 829 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nB = \\frac {X _ {2} \\left(a _ {2} R _ {2 1} - b _ {2} R _ {1 1}\\right)}{b _ {2} a _ {1} - b _ {1} a _ {2}}. \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 529, + 833, + 890, + 864 + ], + "page_idx": 2 + }, + { + "type": "page_footnote", + "text": "1If $R$ is known to be a plane rotation, two generic point-point correspondences suffice to recover $R$ and $T$ .", + "bbox": [ + 500, + 875, + 888, + 898 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "21318", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "(1) Transform data in world and camera frames to the special reference frames satisfying (1), (2), (3), (4), (5).", + "(2) Compute up to 2 solutions in $v$ to the quadratic (16).", + "(3) For each root $v$ in step (2), recover a corresponding value of $u$ using either of the linear expressions in (46).", + "(4) Recover up to 4 solutions in $\\{R_{11}, R_{21}\\}$ using (15).", + "(5) Recover $R$ and $T$ using (13), (12), (10).", + "(6) Reverse the transformation applied in step (1)." + ], + "bbox": [ + 70, + 87, + 468, + 207 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Moreover, from (11) we may use $R_{11}$ and $R_{21}$ , to express the remaining entries of $R$ as", + "bbox": [ + 76, + 260, + 468, + 291 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nR _ {3 1} = \\frac {(a _ {1} - a _ {2}) R _ {2 1} + (b _ {2} - b _ {1}) R _ {1 1})}{b _ {2} a _ {1} - b _ {1} a _ {2}},\n$$\n", + "text_format": "latex", + "bbox": [ + 78, + 296, + 344, + 329 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nR _ {2 2} = \\frac {b _ {1} b _ {2} X _ {2} R _ {1 1} + (b _ {1} a _ {2} X _ {2} - b _ {1} a _ {2} X _ {2} - b _ {2} a _ {1} X _ {3}) R _ {2 1}}{Y _ {3} (b _ {2} a _ {1} - b _ {1} a _ {2})},\n$$\n", + "text_format": "latex", + "bbox": [ + 81, + 330, + 467, + 364 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} R _ {2 3} = \\frac {b _ {1} b _ {2} X _ {2} \\left(Y _ {3} - Y _ {4}\\right) R _ {1 1}}{Y _ {3} Z _ {4} \\left(b _ {2} a _ {1} - b _ {1} a _ {2}\\right)} \\\\ + \\left(\\frac {X _ {3} Y _ {4} - X _ {4} Y _ {3}}{Y _ {3} Z _ {4}} + \\frac {b _ {2} a _ {1} X _ {3} Y _ {4} - b _ {1} a _ {2} X _ {2} Y _ {3}}{Y _ {3} Z _ {4} \\left(b _ {2} a _ {1} - b _ {1} a _ {2}\\right)}\\right) R _ {2 1}. \\tag {13} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 78, + 371, + 468, + 460 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Substituting (13) into (9), we obtain two bivariate quadratic constraints in $R_{11}$ and $R_{21}$ . In matrix form,", + "bbox": [ + 76, + 469, + 468, + 500 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\left( \\begin{array}{l l l} c _ {1} & c _ {2} & c _ {3} \\\\ d _ {1} & d _ {2} & d _ {3} \\end{array} \\right) \\cdot \\left( \\begin{array}{c} R _ {1 1} ^ {2} \\\\ R _ {1 1} R _ {2 1} \\\\ R _ {2 1} ^ {2} \\end{array} \\right) = \\binom {1} {1}, \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 142, + 506, + 468, + 555 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where the coefficients $c_{1}, c_{2}, c_{3}, d_{1}, d_{2}, d_{3}$ are rational functions of the problem data. Applying the change of variables", + "bbox": [ + 76, + 561, + 468, + 592 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nu = R _ {1 1} ^ {2}, \\quad v = R _ {2 1} / R _ {1 1}, \\tag {15}\n$$\n", + "text_format": "latex", + "bbox": [ + 181, + 604, + 468, + 623 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "and subtracting the two equations in (14), we obtain", + "bbox": [ + 76, + 627, + 421, + 642 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\left(c _ {1} - d _ {1}\\right) u + \\left(c _ {2} - d _ {2}\\right) u v + \\left(c _ {3} - d _ {3}\\right) u v ^ {2} = 0.\n$$\n", + "text_format": "latex", + "bbox": [ + 112, + 650, + 431, + 667 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Assuming $u \\neq 0$ , we therefore have the univariate quadratic equation in $v$", + "bbox": [ + 76, + 676, + 468, + 705 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\left(c _ {1} - d _ {1}\\right) + \\left(c _ {2} - d _ {2}\\right) v + \\left(c _ {3} - d _ {3}\\right) v ^ {2} = 0. \\tag {16}\n$$\n", + "text_format": "latex", + "bbox": [ + 112, + 714, + 468, + 731 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "If $v$ is one of the roots of (16), we may recover a corresponding value for $u$ using one of the equations in (14), eg.", + "bbox": [ + 76, + 739, + 468, + 771 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nu = \\left(c _ {1} + c _ {2} v + c _ {3} v ^ {2}\\right) ^ {- 1}, \\quad \\text {o r} \\quad u = \\left(d _ {1} + d _ {2} v + d _ {3} v ^ {2}\\right) ^ {- 1}. \\tag {17}\n$$\n", + "text_format": "latex", + "bbox": [ + 76, + 792, + 468, + 824 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To summarize, we provide the outline of steps for solving the P2P1L absolute pose problem in Figure 3.", + "bbox": [ + 76, + 825, + 467, + 854 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Algebraically, the nontrivial steps in Figure 3 are the second and fourth, which require solving a univariate quadratic equation and computing square roots, respectively.", + "bbox": [ + 76, + 854, + 468, + 901 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/b4abe6b8df011381fda4452e30692a308445f6f91b28742711a701246179f0df.jpg", + "image_caption": [ + "Figure 3. Generic P2P1L solver. See Sec. 2.1 for details." + ], + "image_footnote": [], + "bbox": [ + 524, + 90, + 671, + 199 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/9a0741bac57a0382aa9616fe8857009f1c20205d53c2e9f5227403b71972772b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 702, + 97, + 857, + 199 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/a1a81f7e8cca138e09932e76cff33e20b2d1e8329fc6a6c10f92ee769aaa5f8d.jpg", + "image_caption": [ + "Figure 4. Special reference frame for P1P2L. Camera frame $\\mathcal{C}_0$ is transformed to $\\mathcal{C}_1$ , world frame $\\mathcal{W}_0$ to $\\mathcal{W}_1$ . See text for details." + ], + "image_footnote": [], + "bbox": [ + 527, + 200, + 669, + 290 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/83257912b90f36c02c4353cfc522c5df9718efe3b89b8405f0bac422d13bd179.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 712, + 202, + 849, + 290 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.2.P1P2L", + "text_level": 1, + "bbox": [ + 500, + 354, + 588, + 368 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Here, we provide the algebraically optimal solution to the P1P2L problem. We mostly follow the notation of [25].", + "bbox": [ + 498, + 378, + 890, + 407 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Problem Statement: Let us consider a 3D point $P_1^0 \\in \\mathbb{R}^3$ , and its homogeneous 2D projection $D_1 \\in \\mathbb{R}^3$ under an unknown calibrated camera matrix. Let us also consider two 3D lines—the first spanned by points $L_1, L_2 \\in \\mathbb{R}^3$ , and the second one is spanned by $L_3, L_4 \\in \\mathbb{R}^3$ —and both of their corresponding projections. Our task is to recover the camera matrix from the given point-point correspondence and line-line correspondences.", + "bbox": [ + 498, + 409, + 890, + 530 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Much like the P2P1L solver, our P1P2L solver begins by transforming the input data into a special reference frame, as illustrated in Figure 4. Specifically,", + "bbox": [ + 498, + 531, + 890, + 577 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "- In the world frame, the 3D point takes the form", + "bbox": [ + 500, + 578, + 826, + 592 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nP _ {1} = \\left( \\begin{array}{l l l} 0 & 0 & 0 \\end{array} \\right) ^ {T}. \\tag {18}\n$$\n", + "text_format": "latex", + "bbox": [ + 635, + 604, + 890, + 627 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The first line is spanned by points", + "bbox": [ + 513, + 638, + 738, + 654 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nL _ {1} = \\left( \\begin{array}{l l l} X _ {1} & Y _ {1} & Z _ {1} \\end{array} \\right) ^ {T}, \\quad L _ {2} = \\left( \\begin{array}{l l l} X _ {2} & Y _ {2} & Z _ {2} \\end{array} \\right) ^ {T}, \\tag {19}\n$$\n", + "text_format": "latex", + "bbox": [ + 527, + 666, + 890, + 700 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "and the second line is spanned by points", + "bbox": [ + 513, + 702, + 779, + 715 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nL _ {3} = \\left( \\begin{array}{l l l} X _ {3} & Y _ {3} & Z _ {3} \\end{array} \\right) ^ {T}, \\quad L _ {4} = \\left( \\begin{array}{l l l} X _ {4} & Y _ {4} & Z _ {4} \\end{array} \\right) ^ {T}. \\tag {20}\n$$\n", + "text_format": "latex", + "bbox": [ + 527, + 728, + 890, + 762 + ], + "page_idx": 3 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- For $\\mathcal{W}_0 \\to \\mathcal{W}_1$ , we simply translate $P_1^0$ to $P_1$ .", + "- The world frame may still be rotated freely. We may use the strategy of Section 2.3, which makes the solver stable for a larger class of non-coplanar scenes. However, if $P_{1}, \\overline{L_{1}L_{2}}$ , and $\\overline{L_{3}L_{4}}$ are coplanar, no rotation is recommended.", + "- In the camera frame, the camera center is fixed at" + ], + "bbox": [ + 500, + 763, + 890, + 868 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nC = \\left( \\begin{array}{l l l} 0 & 0 & - 1 \\end{array} \\right) ^ {T}. \\tag {21}\n$$\n", + "text_format": "latex", + "bbox": [ + 630, + 881, + 890, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "21319", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The image point is the projection of a point of the form", + "bbox": [ + 89, + 90, + 455, + 106 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nD _ {1} = \\left( \\begin{array}{l l l} a _ {1} & b _ {1} & 0 \\end{array} \\right) ^ {T}, \\tag {22}\n$$\n", + "text_format": "latex", + "bbox": [ + 204, + 114, + 468, + 137 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "the first line in the image is the projection of a line spanned by points of the form", + "bbox": [ + 89, + 146, + 468, + 176 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nD _ {2} = \\left( \\begin{array}{l l l} a _ {2} & 0 & 0 \\end{array} \\right) ^ {T}, \\quad D _ {3} = \\left( \\begin{array}{l l l} 0 & 0 & 0 \\end{array} \\right) ^ {T}, \\tag {23}\n$$\n", + "text_format": "latex", + "bbox": [ + 117, + 185, + 468, + 208 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "and the second line in the image is the projection of a line spanned by points of the form", + "bbox": [ + 89, + 217, + 468, + 248 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nD _ {4} = \\left( \\begin{array}{l l l} a _ {4} & b _ {4} & 0 \\end{array} \\right) ^ {T}, \\quad D _ {5} = \\left( \\begin{array}{l l l} a _ {5} & b _ {5} & 0 \\end{array} \\right) ^ {T}. \\tag {24}\n$$\n", + "text_format": "latex", + "bbox": [ + 107, + 256, + 468, + 279 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- The homogeneous coordinates of the lines in the camera frame are $n_1 = d_2 \\times d_3$ , $n_2 = d_4 \\times d_5$ , where the rays $d_2 \\neq d_3$ , $d_4 \\neq d_5$ to meet the lines in distinct points.", + "To find the transformation $\\mathcal{C}_0\\to \\mathcal{C}_1$ , we define" + ], + "bbox": [ + 76, + 287, + 468, + 348 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nd _ {1 2} = n _ {1} \\times n _ {2}, \\quad D _ {2} ^ {0} = C ^ {0} + \\frac {d _ {2}}{d _ {2} ^ {T} d _ {1 2}}, \\quad D _ {3} ^ {0} = C ^ {0} + d _ {3},\n$$\n", + "text_format": "latex", + "bbox": [ + 89, + 357, + 468, + 390 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nD _ {2} = \\left( \\begin{array}{c c} \\tan (\\cos^ {- 1} (d _ {2} ^ {T} d _ {1 2})) & 0 \\quad 0 \\end{array} \\right) ^ {T}, \\quad D _ {3} = \\left( \\begin{array}{c c c} 0 & 0 & 0 \\end{array} \\right) ^ {T},\n$$\n", + "text_format": "latex", + "bbox": [ + 89, + 398, + 500, + 421 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "and, independently of $C^0$ , rigidly transform the rays $d_2 \\coloneqq \\overrightarrow{C^0D_2^0}$ and $d_3 \\coloneqq \\overrightarrow{C^0D_3^0}$ to $\\overrightarrow{C^0D_2^0}$ and $\\overrightarrow{CD_3}$ .", + "bbox": [ + 89, + 425, + 468, + 460 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In the special reference frame, we now write the pose as $(R_{ij})_{1\\leq i,j\\leq 3}$ , $(T_i)_{1\\leq i\\leq 3}$ . The projection constraints can be formulated as:", + "bbox": [ + 76, + 460, + 468, + 503 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\left(D _ {1} - C\\right) \\times \\left(R P _ {1} + T\\right) = 0, \\tag {25}\n$$\n", + "text_format": "latex", + "bbox": [ + 173, + 515, + 468, + 532 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\left(D _ {2} \\times D _ {3}\\right) ^ {T} \\left(R L _ {i} + T\\right) = 0, i \\in \\{1, 2 \\}, \\tag {26}\n$$\n", + "text_format": "latex", + "bbox": [ + 138, + 555, + 468, + 574 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\left(D _ {4} \\times D _ {5}\\right) ^ {T} \\left(R L _ {i} + T\\right) = 0, i \\in \\{3, 4 \\}. \\tag {27}\n$$\n", + "text_format": "latex", + "bbox": [ + 138, + 593, + 468, + 611 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Analogously to (11), we have a system of linear equations $A\\mathbf{x} = \\mathbf{b}$ , obtained by picking 2 out of the 3 redundant constraints from (25) together with (26), (27), where now (cf. [25, eq. 7-8])", + "bbox": [ + 76, + 617, + 468, + 676 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nA = \\left( \\begin{array}{c c c c c c c c c} 0 & 0 & 0 & 0 & 0 & 0 & - b _ {1} & a _ {1} & 0 \\\\ 0 & 0 & 0 & 0 & 0 & 0 & 0 & - 1 & b _ {1} \\\\ 0 & 0 & X _ {1} & Y _ {1} & Y _ {1} & Z _ {1} & 0 & 1 & 0 \\\\ 0 & 0 & X _ {2} & Y _ {2} & Y _ {2} & Z _ {2} & 0 & 1 & 0 \\\\ - b _ {4} X _ {3} & - b _ {4} Y _ {3} & a _ {4} X _ {3} & a _ {4} Y _ {3} & a _ {4} Y _ {3} & a _ {4} Z _ {3} & - b _ {4} & a _ {4} & 0 \\\\ - b _ {4} X _ {4} & - b _ {4} Z _ {4} & a _ {4} X _ {4} & a _ {4} Y _ {4} & a _ {4} Y _ {4} & a _ {4} Z _ {4} & - b _ {4} & a _ {4} & 0 \\end{array} \\right),\n$$\n", + "text_format": "latex", + "bbox": [ + 76, + 686, + 431, + 744 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {x} = \\left( \\begin{array}{c c c c c c c c c} R _ {1 1} & R _ {1 2} & R _ {1 3} & R _ {2 1} & R _ {2 2} & R _ {2 3} & T _ {1} & T _ {2} & T _ {3} \\end{array} \\right) ^ {T},\n$$\n", + "text_format": "latex", + "bbox": [ + 78, + 744, + 485, + 766 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {b} = \\left( \\begin{array}{c c c c c c} 0 & - b _ {1} & 0 & 0 & 0 & 0 \\end{array} \\right) ^ {T}.\n$$\n", + "text_format": "latex", + "bbox": [ + 78, + 768, + 297, + 789 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Solving for translation as in (12), we have", + "bbox": [ + 76, + 799, + 356, + 814 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nT _ {1} = \\frac {a _ {1}}{b _ {1}} (- R _ {2 1} X _ {1} - R _ {2 2} Y _ {1} - R _ {2 3} Z _ {1}),\n$$\n", + "text_format": "latex", + "bbox": [ + 104, + 823, + 372, + 852 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nT _ {2} = - R _ {2 1} X _ {1} - R _ {2 2} Y _ {1} - R _ {2 3} Z _ {1}, \\tag {28}\n$$\n", + "text_format": "latex", + "bbox": [ + 106, + 854, + 468, + 871 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nT _ {3} = - 1 + \\frac {1}{b _ {1}} \\left(- R _ {2 1} X _ {1} - R _ {2 2} Y _ {1} - R _ {2 3} Z _ {1}\\right).\n$$\n", + "text_format": "latex", + "bbox": [ + 106, + 875, + 411, + 904 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Furthermore, we express $R_{23}$ from the fourth equation as", + "bbox": [ + 500, + 90, + 879, + 107 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nR _ {2 3} = \\frac {1}{Z _ {1} - Z _ {2}} \\left(R _ {2 1} \\left(X _ {2} - X _ {1}\\right) + R _ {2 2} \\left(Y _ {2} - Y _ {1}\\right)\\right), \\tag {29}\n$$\n", + "text_format": "latex", + "bbox": [ + 509, + 111, + 890, + 141 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "and we solve for $R_{11}, R_{12}$ using the last two equations.", + "bbox": [ + 500, + 145, + 866, + 159 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Now, we can express $R_{23}, T_2, T_1, R_{12}, R_{11}$ as linear combinations of $R_{13}, R_{21}, R_{22}$ in the following form:", + "bbox": [ + 500, + 160, + 890, + 190 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nR _ {2 3} = c _ {1} R _ {2 1} + c _ {2} R _ {2 2},\n$$\n", + "text_format": "latex", + "bbox": [ + 576, + 196, + 733, + 210 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nT _ {2} = c _ {3} R _ {2 1} + c _ {4} R _ {2 2},\n$$\n", + "text_format": "latex", + "bbox": [ + 584, + 215, + 733, + 229 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nT _ {1} = c _ {5} R _ {2 1} + c _ {6} R _ {2 2}, \\tag {30}\n$$\n", + "text_format": "latex", + "bbox": [ + 584, + 234, + 890, + 250 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nR _ {1 2} = c _ {7} R _ {1 3} + c _ {8} R _ {2 1} + c _ {9} R _ {2 2},\n$$\n", + "text_format": "latex", + "bbox": [ + 576, + 253, + 794, + 268 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nR _ {1 1} = c _ {1 0} R _ {1 3} + c _ {1 1} R _ {2 1} + c _ {1 2} R _ {2 2}.\n$$\n", + "text_format": "latex", + "bbox": [ + 578, + 272, + 813, + 286 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The non-linear internal constraints imposed on the elements of $R$ have the form:", + "bbox": [ + 500, + 292, + 890, + 321 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nR _ {1 1} ^ {2} + R _ {1 2} ^ {2} + R _ {1 3} ^ {2} = 1,\n$$\n", + "text_format": "latex", + "bbox": [ + 656, + 324, + 812, + 340 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nR _ {2 1} ^ {2} + R _ {2 2} ^ {2} + R _ {2 3} ^ {2} = 1, \\tag {31}\n$$\n", + "text_format": "latex", + "bbox": [ + 656, + 345, + 890, + 362 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nR _ {1 1} R _ {2 1} + R _ {1 2} R _ {2 2} + R _ {1 3} R _ {2 3} = 0.\n$$\n", + "text_format": "latex", + "bbox": [ + 578, + 366, + 812, + 380 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We substitute (30) into (31), obtaining equations of the form", + "bbox": [ + 500, + 386, + 890, + 402 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\left( \\begin{array}{l l l l l l} d _ {1} & d _ {2} & d _ {3} & d _ {4} & d _ {5} & d _ {6} \\end{array} \\right) \\nu (R) = 1, \\tag {32}\n$$\n", + "text_format": "latex", + "bbox": [ + 562, + 422, + 890, + 439 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\left( \\begin{array}{c c c c c c} 0 & 0 & 0 & d _ {7} & d _ {8} & d _ {9} \\end{array} \\right) \\nu (R) = 1, \\tag {33}\n$$\n", + "text_format": "latex", + "bbox": [ + 584, + 443, + 890, + 459 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\left( \\begin{array}{c c c c c c} 0 & d _ {1 0} & d _ {1 1} & d _ {1 2} & d _ {1 3} & d _ {1 4} \\end{array} \\right) \\nu (R) = 0, \\tag {34}\n$$\n", + "text_format": "latex", + "bbox": [ + 537, + 463, + 890, + 479 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where", + "bbox": [ + 500, + 486, + 544, + 498 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\nu (R) = \\left( \\begin{array}{l l l l l} R _ {1 3} ^ {2} & R _ {1 3} R _ {2 1} & R _ {1 3} R _ {2 2} & R _ {2 1} ^ {2} R _ {2 1} R _ {2 2} & R _ {2 2} ^ {2} \\end{array} \\right) ^ {T}. \\tag {35}\n$$\n", + "text_format": "latex", + "bbox": [ + 500, + 502, + 890, + 537 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We express $R_{13}$ from (34) as", + "bbox": [ + 500, + 537, + 694, + 551 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nR _ {1 3} = - \\frac {d _ {1 2} R _ {2 1} ^ {2} + d _ {1 3} R _ {2 1} R _ {2 2} + d _ {1 4} R _ {2 2} ^ {2}}{d _ {1 0} R _ {2 1} + d _ {1 1} R _ {2 2}}. \\tag {36}\n$$\n", + "text_format": "latex", + "bbox": [ + 539, + 556, + 890, + 589 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Substituting (36) into (32), we obtain", + "bbox": [ + 500, + 593, + 746, + 608 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nd _ {1} \\frac {\\left(d _ {1 2} R _ {2 1} ^ {2} + d _ {1 3} R _ {2 1} R _ {2 2} + d _ {1 4} R _ {2 2} ^ {2}\\right) ^ {2}}{\\left(d _ {1 0} R _ {2 1} + d _ {1 1} R _ {2 2}\\right) ^ {2}} -\n$$\n", + "text_format": "latex", + "bbox": [ + 555, + 612, + 826, + 646 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nd _ {2} \\frac {d _ {1 2} R _ {2 1} ^ {3} + d _ {1 3} R _ {2 1} ^ {2} R _ {2 2} + d _ {1 4} R _ {2 1} R _ {2 2} ^ {2}}{d _ {1 0} R _ {2 1} + d _ {1 1} R _ {2 2}} -\n$$\n", + "text_format": "latex", + "bbox": [ + 557, + 648, + 833, + 681 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nd _ {3} \\frac {d _ {1 2} R _ {2 1} ^ {2} R _ {2 2} + d _ {1 3} R _ {2 1} R _ {2 2} ^ {2} + d _ {1 4} R _ {2 2} ^ {3}}{d _ {1 0} R _ {2 1} + d _ {1 1} R _ {2 2}} +\n$$\n", + "text_format": "latex", + "bbox": [ + 557, + 683, + 833, + 715 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nd _ {4} R _ {2 1} ^ {2} + d _ {5} R _ {2 1} R _ {2 2} + d _ {6} R _ {2 2} ^ {2} = 1. \\tag {37}\n$$\n", + "text_format": "latex", + "bbox": [ + 557, + 718, + 890, + 734 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Subtracting (33) from (37),", + "bbox": [ + 500, + 739, + 681, + 755 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nd _ {1} \\frac {\\left(d _ {1 2} R _ {2 1} ^ {2} + d _ {1 3} R _ {2 1} R _ {2 2} + d _ {1 4} R _ {2 2} ^ {2}\\right) ^ {2}}{\\left(d _ {1 0} R _ {2 1} + d _ {1 1} R _ {2 2}\\right) ^ {2}} -\n$$\n", + "text_format": "latex", + "bbox": [ + 506, + 758, + 777, + 792 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nd _ {2} \\frac {d _ {1 2} R _ {2 1} ^ {3} + d _ {1 3} R _ {2 1} ^ {2} R _ {2 2} + d _ {1 4} R _ {2 1} R _ {2 2} ^ {2}}{d _ {1 0} R _ {2 1} + d _ {1 1} R _ {2 2}} -\n$$\n", + "text_format": "latex", + "bbox": [ + 506, + 795, + 782, + 827 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nd _ {3} \\frac {d _ {1 2} R _ {2 1} ^ {2} R _ {2 2} + d _ {1 3} R _ {2 1} R _ {2 2} ^ {2} + d _ {1 4} R _ {2 2} ^ {3}}{d _ {1 0} R _ {2 1} + d _ {1 1} R _ {2 2}} +\n$$\n", + "text_format": "latex", + "bbox": [ + 506, + 830, + 782, + 862 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\left(d _ {4} - d _ {7}\\right) R _ {2 1} ^ {2} + \\left(d _ {5} - d _ {8}\\right) R _ {2 1} R _ {2 2} + \\left(d _ {6} - d _ {9}\\right) R _ {2 2} ^ {2} = 0. \\tag {38}\n$$\n", + "text_format": "latex", + "bbox": [ + 506, + 864, + 890, + 895 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "21320", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "(1) Transform data in world and camera frames to the special reference frames satisfying (18), (19), (20), (21), (22), (23), (24).", + "(2) Compute up to 4 solutions in $v$ to the quartic (41).", + "(3) For each root $v$ in step (2), recover 2 solutions in $\\{R_{21}, R_{22}\\}$ using $R_{21} = R_{22}v$ and (33).", + "(4) Recover remaining coordinates for each solution using (34),(30).", + "(5) Reverse the transformation applied in step (1)." + ], + "bbox": [ + 68, + 87, + 468, + 223 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/0b3fc4accd39c7cfdb4da1c7c5909a4ea8e8cff9c9a37faba300cce6245b0a78.jpg", + "image_caption": [ + "Figure 6. Stabilization of P1P2L solver. The frame is rotated to align the line $\\overline{L_1L_2}$ with the z-axis. See Sec. 2.3 for details." + ], + "image_footnote": [], + "bbox": [ + 117, + 273, + 254, + 367 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/cd7d9f719a68f865c3772ae71e7acfebff8f70065de201c7b0b0fe42427766d8.jpg", + "image_caption": [ + "Figure 5. Generic P1P2L solver. See Sec. 2.2 for details." + ], + "image_footnote": [], + "bbox": [ + 285, + 272, + 421, + 367 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We then clear denominators in (38), multiplying by $(d_{10}R_{21} + d_{11}R_{22})^2$ to get a polynomial equation", + "bbox": [ + 76, + 434, + 468, + 465 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} d _ {1} \\left(d _ {1 2} R _ {2 1} ^ {2} + d _ {1 3} R _ {2 1} R _ {2 2} + d _ {1 4} R _ {2 2} ^ {2}\\right) ^ {2} - \\\\ \\left(\\left(d _ {2} \\left(d _ {1 2} R _ {2 1} ^ {3} + d _ {1 3} R _ {2 1} ^ {2} R _ {2 2} + d _ {1 4} R _ {2 1} R _ {2 2} ^ {2}\\right) - \\right. \\right. \\\\ \\left. d _ {3} \\left(d _ {1 2} R _ {2 1} ^ {2} R _ {2 2} + d _ {1 3} R _ {2 1} R _ {2 2} ^ {2} + d _ {1 4} R _ {2 2} ^ {3}\\right)\\right) \\cdot \\tag {39} \\\\ \\left(d _ {1 0} R _ {2 1} + d _ {1 1} R _ {2 2}\\right) + \\left(\\left(d _ {4} - d _ {7}\\right) R _ {2 1} ^ {2} + \\right. \\\\ \\left. \\left(d _ {5} - d _ {8}\\right) R _ {2 1} R _ {2 2} + \\left(d _ {6} - d _ {9}\\right) R _ {2 2} ^ {2}\\right) \\Bigg). \\\\ \\left(d _ {1 0} R _ {2 1} + d _ {1 1} R _ {2 2}\\right) = 0. \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 107, + 479, + 468, + 636 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Upon expanding equation (39), we find that it takes the form", + "bbox": [ + 76, + 652, + 468, + 667 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\alpha_ {1} R _ {2 1} ^ {4} + \\alpha_ {2} R _ {2 1} ^ {3} R _ {2 2} + \\alpha_ {3} R _ {2 1} ^ {2} R _ {2 2} ^ {2} + \\alpha_ {4} R _ {2 1} R _ {2 2} ^ {3} + \\alpha_ {5} 2 R _ {2 2} ^ {4} = 0 \\tag {40}\n$$\n", + "text_format": "latex", + "bbox": [ + 76, + 679, + 468, + 710 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Similarly to (16), we divide (40) by $R_{22}^{4}$ and define a new variable $v = \\frac{R_{21}}{R_{22}}$ , and thereby deduce the univariate quartic", + "bbox": [ + 76, + 710, + 468, + 744 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\alpha_ {1} v ^ {4} + \\alpha_ {2} v ^ {3} + \\alpha_ {3} v ^ {2} + \\alpha_ {4} v + \\alpha_ {5} = 0. \\tag {41}\n$$\n", + "text_format": "latex", + "bbox": [ + 140, + 761, + 468, + 779 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "After solving for $v$ , we can solve for $R_{21}, R_{22}$ using (33), and for the other equations using (34), and (30).", + "bbox": [ + 76, + 792, + 467, + 823 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To summarize, we provide the outline of steps for solving the P1P2L absolute pose problem in Figure 5. Algebraically, the nontrivial steps in Figure 5 are second and third, which require solving a univariate quartic equation and computing square roots, respectively.", + "bbox": [ + 75, + 824, + 468, + 900 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/765307c92acc345246e93bff90c2a30594fcdf5293daa32cd6e07501282fcf2e.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
GenericMean RMed. RMean TMed. T
P1P2L no fix0.000501.0e-140.000651.8e-13
P1P2L fix9.0e-094.2e-153.4e-077.0e-14
P2P1L no fix7.1e-111.4e-151.2e-092.1e-14
P2P1L fix3.2e-092.6e-157.4e-084.3e-14
CoplanarMean RMed. RMean TMed. T
P1P2L no fix0.000229.6e-150.000301.75e-13
P1P2L fix1.10.791.20.96
P2P1L no fix2.43.142.23.14
P2P1L fix1.2e-124.0e-157.9e-116.3e-14
", + "bbox": [ + 501, + 87, + 901, + 248 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 2. The mean and median of the rot.(R) and tran.(T) errors over 1e5 noiseless generic (top) and coplanar (bottom) samples. In radians. The best results are marked bold.", + "bbox": [ + 498, + 258, + 890, + 300 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "2.3. Stabilizing the P1P2L Solver", + "text_level": 1, + "bbox": [ + 500, + 327, + 754, + 343 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this section, we outline a method to increase the stability of the P1P2L solver (Sec. 2.2).", + "bbox": [ + 498, + 351, + 890, + 380 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Our proposed P1P2L solution faces a degeneracy when $Z_{1} - Z_{2} = 0$ , since this leads to a division by zero in equation (29). Furthermore, our observations indicate that the result of the solver is unstable if the value of $Z_{1} - Z_{2}$ is close to zero. Note, that this instability is specific to our solver and not inherent to the P1P2L problem.", + "bbox": [ + 498, + 381, + 890, + 470 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "This degeneracy may be interpreted geometrically as follows: the values of $Z_{1}$ and $Z_{2}$ are the last coordinates of the points $L_{1}, L_{2}$ , which span the first 3D line. Therefore, the vector $L_{1} - L_{2}$ represents the direction of the line, and $Z_{1} - Z_{2}$ represents the last coordinate of this direction.", + "bbox": [ + 498, + 472, + 890, + 547 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Since the special reference frame used in Section 2.2 is independent of the rotation of the world frame, we can remove the source of instability by rotating the world frame such that the first line $\\overline{L_1L_2}$ aligns with the z-axis. See Figure 6 for an illustration.", + "bbox": [ + 498, + 547, + 890, + 622 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "2.4. Resolving the coplanar case", + "text_level": 1, + "bbox": [ + 500, + 633, + 748, + 650 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this section, we discuss the performance of the P2P1L (Sec. 2.1) and P1P2L (Sec. 2.2) solvers in the case, where all 3D points and 3D lines are coplanar.", + "bbox": [ + 498, + 657, + 890, + 702 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "P2P1L. The P2P1L solver presented in Sec. 2.1 (P2P1L no fix) is degenerate in coplanar situations, since then $Z_{4} = 0$ (2), causing a division by zero issue in equation (11). To resolve this issue, we propose a modified solver (P2P1L fix), which handles the coplanar case. See SM A for a description of this modified solver.", + "bbox": [ + 498, + 703, + 890, + 792 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "P1P2L. The original version of the P1P2L solver (P1P2L no fix, Sec. 2.2) is able to handle the coplanar case. However, the stabilized P1P2L solver (P1P2L fix, Sec. 2.3) fails due to a degeneracy for coplanar input.", + "bbox": [ + 498, + 794, + 890, + 854 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We conducted a synthetic experiment to evaluate the impact of the proposed stabilization scheme. In this experiment, we used the same setting as in Section 3.1. The re", + "bbox": [ + 498, + 854, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "21321", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/f3ac5acb4ef713bba12a1367b6196e05f9c5c0ef13e64dd9aeee7b2e99403755.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 127, + 89, + 434, + 142 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f62ca8369537d116d1a2718261a301ede8c087a96305d92bf80aed04c791324d.jpg", + "image_caption": [ + "Figure 7. Stability test. Histogram of rotation $(l.)$ , and translation $(r.)$ errors computed over $1e5$ noiseless samples." + ], + "image_footnote": [], + "bbox": [ + 81, + 143, + 460, + 292 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/9bdda44185fe7690ca56b8a57f9ea1006f0c7d211a8526bad9fc666ad01a083f.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodMeanMed.MinMax
P2P1L Ours313.8324.9230.83061.0
P2P1L 3Q31860.61909.91439.110102.1
P2P1L R.+S8897.59491.25805.149984.1
P2P1L R.+L4720.85239.52763.315362.6
P1P2L Ours504.0521.0364.04554.4
P1P2L 3Q31967.12008.21483.912931.1
", + "bbox": [ + 86, + 342, + 459, + 454 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 3. Runtime The mean, median and max values of the runtime. In nanoseconds. Best values for each problem are bold.", + "bbox": [ + 76, + 464, + 467, + 492 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "sults, presented in Table 2 demonstrate that solvers $P2P1L$ no fix and $P1P2L$ fix achieve superior numerical stability in the generic case, while solvers $P2P1L$ fix and $P1P2L$ no fix handle the coplanar case. Since it is simple to detect the coplanar case, we recommend to use solvers $P2P1L$ fix and $P1P2L$ no fix in the coplanar case and solvers $P2P1L$ no fix and $P1P2L$ fix in the generic case.", + "bbox": [ + 75, + 522, + 467, + 628 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3. Experiments", + "text_level": 1, + "bbox": [ + 76, + 645, + 207, + 662 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In this section, we experimentally compare the proposed solvers with the state-of-the-art methods, specifically 3Q3 [17] and Ramalingam [25]. For 3Q3, we employ the publicly available implementation from Poselib. As the implementation of Ramalingam is not publicly accessible, we have created our own implementation. Since it is not explicitly specified which method should be used for solving the linear equations, we compare with two variants: SVD and LU decomposition. All the solvers have been implemented in $\\mathrm{C + + }$ , and the experiments are conducted on a desktop computer with an AMD Ryzen 9 CPU with $3.9\\mathrm{GHz}$ .", + "bbox": [ + 75, + 672, + 467, + 837 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In Sec. 3.1, we provide an analysis of the numerical stability and runtime performance of the solvers on synthetic data. Subsequently, in Section 3.2, we present an evaluation of the solvers within the RANSAC scheme.", + "bbox": [ + 75, + 839, + 467, + 898 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.1. Synthetic experiments", + "text_level": 1, + "bbox": [ + 500, + 90, + 705, + 106 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In this section, we present an analysis of the numerical stability and runtime performance of the minimal solvers on synthetic data.", + "bbox": [ + 498, + 114, + 890, + 157 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We generate instances of each minimal problem (either P2P1L or P1P2L), according to the following procedure:", + "bbox": [ + 498, + 159, + 890, + 190 + ], + "page_idx": 6 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Rotation Matrix $(R)$ : An axis $v \\in \\mathbb{R}^3$ is sampled from the uniform distribution on the unit sphere, and an angle $\\alpha$ is sampled from the normal distribution $\\mathcal{N}(0,1)$ . The rotation matrix $R$ is then constructed using the angle-axis formula, $R = I + \\sin(\\alpha)[v]_{\\times} + (1 - \\cos(\\alpha))[v]_{\\times}^2$ .", + "- Translation $(T)$ : The camera center $C$ is sampled uniformly at random from the unit sphere. The translation vector $T$ is computed as $T = -RC$ .", + "- Point correspondence: A 3D point $X_{i}$ is sampled from the trivariate normal distribution with mean vector $\\mu = [0,0,5]^T$ and standard deviation $\\sigma = 1$ in each component. The corresponding 2D point $D_{i}$ is obtained by projecting $X_{i}$ onto a pinhole camera with the pose $(R,T)$ .", + "- Line correspondence: Two 3D points, $L_{i}$ and $L_{i+1}$ , are sampled as described in the previous bullet-point. These points define the 3D line $L$ . Two 2D points, $D_{i}$ and $D_{i+1}$ , are obtained by sampling two points on the 3D line $L$ and projecting them onto the camera with the pose $(R, T)$ ." + ], + "bbox": [ + 500, + 193, + 890, + 465 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In the case of P2P1L, two points and one line are sampled, while in the case of P1P2L, one point and two lines are sampled. Let $(R_{est},T_{est})$ be the pose obtained by the minimal solver. We measure the rotation error as the angle $\\xi_R = \\arccos ((1 + \\mathrm{trace}(R_{est}^T R)) / 2)$ . We calculate the translation error $\\xi_T$ as $\\xi_T = \\frac{|T_{est} - T|}{|T|}$ . The histograms of rotation errors $(\\xi_R)$ and translation errors $(\\xi_T)$ for all considered solvers are depicted in Figure 7. Furthermore, summary statistics including the median, mean, and maximum errors are provided in Table 4. The results indicate that all minimal solvers are stable. Our algebraically optimal solvers demonstrate superior stability compared to the previous solvers.", + "bbox": [ + 496, + 468, + 890, + 667 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The runtime evaluation for all considered solvers is shown in Table. 3. Our P2P1L solver requires 313.8ns on average, which is about 6x faster than the 3Q3 solver [17]. Similarly, our P1P2L solver requires on average 504.0ns, which is about 4x faster than the 3Q3 solver. These results show a significant speedup achieved by our solvers in comparison to the previous methods.", + "bbox": [ + 496, + 669, + 890, + 775 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.2. Experiments in RANSAC", + "text_level": 1, + "bbox": [ + 500, + 786, + 732, + 801 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Here, we provide an analysis of the minimal solvers within the RANSAC scheme [4], using the Oxford Multi-view Data [1]. This dataset includes a variety of indoor and outdoor scenes, some of which contain both point and line matches. An example of images from this dataset is illustrated in Figure 8.", + "bbox": [ + 496, + 809, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "21322", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/46608815a26be9030d013de8b62756e2bfe2fc546a4f5ba225ed55073d869966.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodMean RMedian RMax RMean TMedian TMax T
P2P1L Ours5.3e-121.4e-151.2e-073.7e-102.1e-142.2e-05
P2P1L 3Q32.8e-056.5e-152.72.0e-051.2e-131.04
P2P1L Ramalingam (SVD)4.9e-091.5e-140.000411.6e-072.7e-130.0099
P2P1L Ramalingam (LU)4.7e-071.3e-140.0402.3e-052.4e-130.10
PP1P2L Ours (fix 2)1.2e-074.4e-150.0102.0e-067.1e-140.13
P1P2L 3Q33.3e-057.2e-152.603.4e-051.5e-131.01
", + "bbox": [ + 155, + 87, + 810, + 199 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/5a89bfd738c035f70daacdbac4faeb7fe21f09b1563c076422fe95f0fd737054.jpg", + "table_caption": [ + "Table 4. The mean, median and max values of the rot.(R) and tran.(T) errors over ${1e5}$ noiseless samples. The best results are marked bold." + ], + "table_footnote": [], + "table_body": "
DatasetP2P1LP1P2L
OUR3Q3Ram. SVDRam. LUOUR3Q3
Model House7.72 (0.85x)9.09 (1x)20.69 (2.28x)17.30 (1.90x)8.06 (0.85x)9.52 (1x)
Corridor11.32 (0.91x)12.46 (1x)26.56 (2.13x)23.54 (1.89x)13.3 (0.90x)14.76 (1x)
Merton I35.93 (0.98x)36.83 (1x)77.96 (2.12x)74.87 (2.03x)28.85 (0.64x)45.41 (1x)
Merton II33.30 (0.97x)34.27 (1x)74.37 (2.17x)71.35 (2.08x)26.7 (0.67x)39.64 (1x)
Merton III24.04 (0.96x)25.07 (1x)52.07 (2.08x)48.97 (1.95x)10.91 (0.37x)29.67 (1x)
Library32.42 (0.97x)33.31 (1x)69.29 (2.08x)66.23 (1.99x)10.04 (0.28x)35.98 (1x)
Wadham39.60 (0.98x)40.51 (1x)86.10 (2.13x)83.00 (2.05x)22.96 (0.45x)50.79 (1x)
Avg. Speed-up0.94x1x2.14x1.99x0.59x1x
", + "bbox": [ + 119, + 229, + 848, + 397 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 5. RANSAC timing, on Oxford Multi-view dataset [1], in milliseconds. Speed-up compared to the 3Q3 in the bracket.", + "bbox": [ + 114, + 402, + 852, + 417 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/1a1467d082b96b6409e6de7ca27d3e1e313ffb3301faa7f70a5f7a861fa0ba8f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 117, + 441, + 263, + 527 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/c0a5efd34e5788c9b8db10104226b00eadb0bafe109346eed28a4f5da64ce227.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 277, + 441, + 424, + 526 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/937618e532ccbafba2a2787bad9aa48be71248dcdeeb2a2423e084d01d652d76.jpg", + "image_caption": [ + "Figure 8. Example images of University Library (top) and Merton College III (bottom), from [1]." + ], + "image_footnote": [], + "bbox": [ + 117, + 527, + 264, + 613 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/7774b5909468adbd6f981ab9a5bb9161a98e5523a8b2af6612a8aaeaecc5f8f2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 279, + 527, + 424, + 613 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We utilize the Poselib [18] implementation of LORANSAC [6]. In this implementation, local optimization is applied both when the new solution surpasses the currently best one and at the end of the entire procedure. The RANSAC parameters used include a maximum number of iterations set to 100000, a minimum number of iterations set to 1000, a success probability of 0.9999, and an inlier threshold of 1 pixel. To ensure a fair comparison between the solvers, we use a fixed random seed.", + "bbox": [ + 75, + 672, + 468, + 808 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Since all the solvers are stable, we do not expect any significant difference in the estimated pose when using different solvers. Therefore, our primary focus is on comparing the runtime of the RANSAC scheme. The runtime comparison is presented in Table 5, which displays the average runtime for each scene in the Oxford multiview dataset and", + "bbox": [ + 75, + 810, + 470, + 898 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "for each considered solver. Additionally, the table includes a speedup value compared to the 3Q3 solver of [18]. The results show that our solvers consistently outperform the 3Q3 [17] and Ramalingam [25] solvers in terms of runtime. However, there is a notable variation in the speed-up among the scenes, which ranges from 0.85 to 0.98 for the P2P1L problem and from 0.28 to 0.90 for the P1P2L solver. This variation can be attributed to the fact that different scenes require varying proportions of time spent by RANSAC on scoring hypothesized models and local optimization. The more time is dedicated to these tasks, the less significant speedup can be reached. In cases with fewer matches or a lower inlier ratio, more substantial speedups are possible.", + "bbox": [ + 496, + 444, + 890, + 640 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The final pose estimation errors in RANSAC are shown in the Supplementary. As expected, the errors are the same for all solvers. This demonstrates that we can achieve improved runtimes without sacrificing any accuracy.", + "bbox": [ + 496, + 641, + 890, + 702 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 712, + 617, + 727 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In revisiting absolute pose from mixed point-line correspondences, we have developed new solvers, which have the merit of being both algebraically optimal and elementary in nature. Moreover, our solvers outperform the state-of-the-art in terms of both runtime and numerical stability. Despite the simplicity of many minimal absolute pose problems, our findings suggest that their continued investigation remains worthwhile. The code is available at https://github.com/petrhruby97/efficientAbsolutely", + "bbox": [ + 496, + 734, + 890, + 869 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements: TD was supported by NSF DMS 2103310.", + "bbox": [ + 500, + 869, + 890, + 898 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "21323", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Multi-view data, oxford visual geometry group. https://www.robots.ox.ac.uk/vgg/data/mview/. Accessed: Nov 8 2023. 7, 8, 1", + "[2] Sérgio Agostinho, João Gomes, and Alessio Del Bue. CvxPnPL: A unified convex solution to the absolute pose estimation problem from point and line correspondences. Journal of Mathematical Imaging and Vision, 65(3):492-512, 2023. 1", + "[3] Daniel Barath and Jiri Matas. Graph-cut RANSAC: local optimization on spatially coherent structures. IEEE Trans. Pattern Anal. Mach. Intell., 44(9):4961-4974, 2022. 1", + "[4] Robert C. Bolles and Martin A. Fischler. A ransac-based approach to model fitting and its application to finding cylinders in range data. In Proceedings of the 7th International Joint Conference on Artificial Intelligence, IJCAI '81, Vancouver, BC, Canada, August 24-28, 1981, pages 637-643. William Kaufmann, 1981. 7", + "[5] Homer H. Chen. Pose determination from line-to-plane correspondences: Existence condition and closed-form solutions. IEEE Trans. Pattern Anal. Mach. Intell., 13(6):530-541, 1991. 1", + "[6] Ondrej Chum, Jiri Matas, and Josef Kittler. Locally optimized RANSAC. In Pattern Recognition, 25th DAGM Symposium, Magdeburg, Germany, September 10-12, 2003, Proceedings, pages 236-243. Springer, 2003. 8", + "[7] Michel Dhome, Marc Richetin, Jean-Thierry Lapresté, and Gérard Rives. Determination of the attitude of 3d objects from a single perspective view. IEEE Trans. Pattern Anal. Mach. Intell., 11(12):1265-1278, 1989. 1", + "[8] Yaqing Ding, Jian Yang, Viktor Larsson, Carl Olsson, and Kalle Åström. Revisiting the P3P problem. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2023, Vancouver, BC, Canada, June 17-24, 2023, pages 4872-4880. IEEE, 2023. 1, 2", + "[9] Timothy Duff, Viktor Korotynskiy, Tomas Pajdla, and Margaret H. Regan. Galois/monodromy groups for decomposing minimal problems in 3d reconstruction. SIAM Journal on Applied Algebra and Geometry, 6(4):740-772, 2022. 2", + "[10] Ricardo Fabbri, Peter J. Giblin, and Benjamin B. Kimia. Camera pose estimation using first-order curve differential geometry. IEEE Trans. Pattern Anal. Mach. Intell., 43(10): 3321-3332, 2021. 1", + "[11] Martin A. Fischler and Robert C. Bolles. Random sample consensus: A paradigm for model fitting with applications to image analysis and automated cartography. Commun. ACM, 24(6):381-395, 1981. 1", + "[12] Christian Hane, Lionel Heng, Gim Hee Lee, Friedrich Fraundorfer, Paul Furgale, Torsten Sattler, and Marc Pollefeys. 3d visual perception for self-driving cars using a multi-camera system: Calibration, mapping, localization, and obstacle detection. Image Vis. Comput., 68:14-27, 2017. 1", + "[13] Robert M. Haralick, Chung-Nan Lee, Karsten Ottenberg, and Michael Nölle. Review and analysis of solutions of the three point perspective pose estimation problem. Int. J. Comput. Vis., 13(3):331-356, 1994. 1" + ], + "bbox": [ + 78, + 114, + 467, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[14] Petr Hruby, Viktor Korotynskiy, Timothy Duff, Luke Oeding, Marc Pollefeys, Tomás Pajdla, and Viktor Larsson. Four-view geometry with unknown radial distortion. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2023, Vancouver, BC, Canada, June 17-24, 2023, pages 8990-9000. IEEE, 2023. 2", + "[15] Xiaomei Hu, Luying Zhu, Ping Wang, Haili Yang, and Xuan Li. Improved ORB-SLAM2 mobile robot vision algorithm based on multiple feature fusion. IEEE Access, 11:100659-100671, 2023. 1", + "[16] Laurent Kneip, Davide Scaramuzza, and Roland Siegwart. A novel parametrization of the perspective-three-point problem for a direct computation of absolute camera position and orientation. In The 24th IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2011, Colorado Springs, CO, USA, 20-25 June 2011, pages 2969-2976. IEEE Computer Society, 2011. 1", + "[17] Zuzana Kukelova, Jan Heller, and Andrew W. Fitzgibbon. Efficient intersection of three quadrics and applications in computer vision. In 2016 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2016, Las Vegas, NV, USA, June 27-30, 2016, pages 1799-1808. IEEE Computer Society, 2016. 2, 7, 8", + "[18] Viktor Larsson and contributors. PoseLib - Minimal Solvers for Camera Pose Estimation, 2020. 2, 8", + "[19] Shaohui Liu, Yifan Yu, Rémi Pautrat, Marc Pollefeys, and Viktor Larsson. 3d line mapping revisited. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2023, Vancouver, BC, Canada, June 17-24, 2023, pages 21445-21455. IEEE, 2023. 1", + "[20] David Nister, Richard I. Hartley, and Henrik Stewenius. Using galois theory to prove structure from motion algorithms are optimal. In 2007 IEEE Computer Society Conference on Computer Vision and Pattern Recognition (CVPR 2007), 18-23 June 2007, Minneapolis, Minnesota, USA. IEEE Computer Society, 2007. 2", + "[21] Rémi Pautrat, Daniel Barath, Viktor Larsson, Martin R. Oswald, and Marc Pollefeys. Deeplsd: Line segment detection and refinement with deep image gradients. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2023, Vancouver, BC, Canada, June 17-24, 2023, pages 17327-17336. IEEE, 2023. 1", + "[22] Rémi Pautrat, Iago Suárez, Yifan Yu, Marc Pollefeys, and Viktor Larsson. Gluestick: Robust image matching by sticking points and lines together. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9706-9716, 2023. 1", + "[23] Mikael Persson and Klas Nordberg. Lambda twist: An accurate fast robust perspective three point (P3P) solver. In Computer Vision - ECCV 2018 - 15th European Conference, Munich, Germany, September 8-14, 2018, Proceedings, Part IV, pages 334-349. Springer, 2018. 1, 2", + "[24] Rahul Raguram, Ondrej Chum, Marc Pollefeys, Jiri Matas, and Jan-Michael Frahm. USAC: A universal framework for random sample consensus. IEEE Trans. Pattern Anal. Mach. Intell., 35(8):2022-2038, 2013. 1", + "[25] Srikumar Ramalingam, Sofien Bouaziz, and Peter F. Sturm. Pose estimation using both points and lines for geo-" + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "21324", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "localization. In IEEE International Conference on Robotics and Automation, ICRA 2011, Shanghai, China, 9-13 May 2011, pages 4716-4723. IEEE, 2011. 1, 2, 3, 4, 5, 7, 8", + "[26] Torsten Sattler, Bastian Leibe, and Leif Kobbelt. Efficient & effective prioritized matching for large-scale image-based localization. IEEE Trans. Pattern Anal. Mach. Intell., 39(9): 1744-1756, 2017. 1", + "[27] Johannes L. Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In 2016 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2016, Las Vegas, NV, USA, June 27-30, 2016, pages 4104-4113. IEEE Computer Society, 2016. 1", + "[28] Jonathan Ventura, Clemens Arth, Gerhard Reitmayr, and Dieter Schmalstieg. Global localization from monocular SLAM on a mobile phone. IEEE Trans. Vis. Comput. Graph., 20(4):531-539, 2014. 1", + "[29] Chi Xu, Lilian Zhang, Li Cheng, and Reinhard Koch. Pose estimation from line correspondences: A complete analysis and a series of solutions. IEEE Trans. Pattern Anal. Mach. Intell., 39(6):1209-1222, 2017. 1", + "[30] Lipu Zhou, Jiamin Ye, and Michael Kaess. A stable algebraic camera pose estimation for minimal configurations of 2d/3d point and line correspondences. In Computer Vision - ACCV 2018 - 14th Asian Conference on Computer Vision, Perth, Australia, December 2-6, 2018, Revised Selected Papers, Part IV, pages 273-288. Springer, 2018. 1, 2" + ], + "bbox": [ + 78, + 90, + 468, + 459 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "21325", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2024/Efficient Solution of Point-Line Absolute Pose/8561d28f-e735-46cc-9618-c6c206bdafc2_model.json b/2024/Efficient Solution of Point-Line Absolute Pose/8561d28f-e735-46cc-9618-c6c206bdafc2_model.json new file mode 100644 index 0000000000000000000000000000000000000000..7a46af8d86bb6c900bad7d9d35db96c763bde9da --- /dev/null +++ b/2024/Efficient Solution of Point-Line Absolute Pose/8561d28f-e735-46cc-9618-c6c206bdafc2_model.json @@ -0,0 +1,3135 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.255, + 0.131, + 0.719, + 0.151 + ], + "angle": 0, + "content": "Efficient Solution of Point-Line Absolute Pose" + }, + { + "type": "text", + "bbox": [ + 0.186, + 0.182, + 0.284, + 0.216 + ], + "angle": 0, + "content": "Petr Hruby ETH Zürich" + }, + { + "type": "text", + "bbox": [ + 0.337, + 0.182, + 0.543, + 0.218 + ], + "angle": 0, + "content": "Timothy Duff \nUniversity of Washington" + }, + { + "type": "text", + "bbox": [ + 0.592, + 0.182, + 0.78, + 0.216 + ], + "angle": 0, + "content": "Marc Pollefeys \nETH Zürich, Microsoft" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.252, + 0.314, + 0.267 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.283, + 0.474, + 0.542 + ], + "angle": 0, + "content": "We revisit certain problems of pose estimation based on 3D-2D correspondences between features which may be points or lines. Specifically, we address the two previously-studied minimal problems of estimating camera extrinsics from \\( p \\in \\{1,2\\} \\) point-point correspondences and \\( l = 3 - p \\) line-line correspondences. To the best of our knowledge, all of the previously-known practical solutions to these problems required computing the roots of degree \\( \\geq 4 \\) (univariate) polynomials when \\( p = 2 \\), or degree \\( \\geq 8 \\) polynomials when \\( p = 1 \\). We describe and implement two elementary solutions which reduce the degrees of the needed polynomials from 4 to 2 and from 8 to 4, respectively. We show experimentally that the resulting solvers are numerically stable and fast: when compared to the previous state-of-the-art, we may obtain nearly an order of magnitude speedup. The code is available at https://github.com/petrhruby97/efficient_absolute" + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.566, + 0.21, + 0.582 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.591, + 0.2, + 0.606 + ], + "angle": 0, + "content": "1.1. Motivation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.614, + 0.47, + 0.765 + ], + "angle": 0, + "content": "The problem of registering images to a known 3D coordinate system plays a crucial role in applications such as visual localization [26], autonomous driving [12], and augmented reality [28], as well as in general paradigms like SLAM [15] and SfM [27]. Robust estimators based on RANSAC [11], or one of its many refinements [3, 24], are among the most successful tools for solving these problems. Such an estimator traditionally relies on a minimal P3P solver [8, 13, 16, 23] to efficiently hypothesize poses from putative matches between 3D and 2D points." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.471, + 0.902 + ], + "angle": 0, + "content": "The literature on P3P and other purely point-based methods for absolute pose estimation is vast. Absolute pose estimation from non-point features such as lines [2, 5, 7, 25, 29, 30], point-line incidences [10], and affine correspondences [28], has received comparatively less attention, but remains an active research area. In particular, solutions relying on both points and lines are of increasing importance, due to the prevalence of both types of feature in man-made environments, as well as several recent advances in the com" + }, + { + "type": "image", + "bbox": [ + 0.532, + 0.248, + 0.868, + 0.462 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.521, + 0.469, + 0.871, + 0.485 + ], + "angle": 0, + "content": "Figure 1. The P2P1L (Top) and P1P2L (Bottom) problems." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.513, + 0.892, + 0.545 + ], + "angle": 0, + "content": "ponents of 3D reconstruction systems responsible for line detection [21], matching [22], and bundle adjustment [19]." + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.557, + 0.638, + 0.571 + ], + "angle": 0, + "content": "1.2. Contribution" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.581, + 0.892, + 0.657 + ], + "angle": 0, + "content": "In this paper, we revisit the two minimal problems of absolute pose estimation that combine both points and lines: the Perspective-2-Point-1-Line Problem (or P2P1L, Section 2.1), and the Perspective-1-Point-2-Line Problem (P1P2L, Section 2.2.) See Figure 1 for illustrations." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.658, + 0.893, + 0.824 + ], + "angle": 0, + "content": "In contrast to the purely point or line based minimal problems P3P and P3L, we observe that the existing solutions for the \"mixed cases\" considered here are suboptimal from both a theoretical and practical point of view. Consequently, we develop novel solutions to both problems, which optimally exploit their underlying algebraic structure, and exhibit comparable or better performance than the state of the art on simulated and real data—see Section 3. On the other hand, these solvers are simple to implement, and no knowledge of mathematics beyond elementary algebra is needed to understand them." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.825, + 0.894, + 0.903 + ], + "angle": 0, + "content": "To provide further context for our work, we include the following quotation from [25]: Although we do not theoretically prove that our solutions are of the lowest possible degrees, we believe so because of the following argument. The best existing solutions for pose estimation using three" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "21316" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.183 + ], + "angle": 0, + "content": "points and three lines use 4th and 8th degree solutions respectively. Since mixed cases are in the middle, our solutions for (2 points, 1 line) and (1 point, 2 lines) cases use 4th and 8th degree solutions respectively. Recently, it was shown using Galois theory that the solutions that use the lowest possible degrees are the optimal ones [[20]]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.185, + 0.471, + 0.351 + ], + "angle": 0, + "content": "Contrary to the informal reasoning presented above, we claim that the existing solutions to the mixed point-line cases are not optimal. As far as we know, the only support for this claim appearing in the literature prior to our work occurs in [9, §3]. This previous work showed, on the basis of Galois group computation, that the P2P1L and P1P2L problems decompose into simpler subproblems. However, this theoretical observation was not accompanied by a practical solution method for either problem. In this paper, we rectify the situation by devising practical solvers for both problems that incorporate these recent insights." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.353, + 0.471, + 0.777 + ], + "angle": 0, + "content": "To complete our discussion of what makes a solution algebraically optimal, we recall one proposal of such a notion from work of Nister et al. [20]. In this work, a restricted class of algorithms \\(\\mathcal{P}_n\\) is considered for each natural number \\(n \\geq 1\\): an algorithm in \\(\\mathcal{P}_n\\) consists of a finite sequence of steps, each of which extracts the roots of some polynomial equation \\(p(x) = 0\\), with either \\(p(x) = x^m - a\\) (where \\(a \\in \\mathbb{Q}\\) and \\(\\deg(p) = m\\) can be arbitrary), or with \\(\\deg(p) \\leq n\\) and the coefficients of \\(p\\) belonging to a field containing \\(\\mathbb{Q}\\) and any previously-computed roots. In this setting, a solution is optimal if it leads to solving a polynomial system of the lowest possible degree. Our proposed solutions immediately establish that P2P1L is \\(\\mathcal{P}_2\\)-solvable and P1P2L is \\(\\mathcal{P}_4\\)-solvable. It should be of little surprise that P2P1L is not \\(\\mathcal{P}_1\\)-solvable; therefore, we may say that our solution to P2P1L is algebraically optimal. On the other hand, the problem P1P2L is also \\(\\mathcal{P}_3\\)-solvable. This is because any quartic equation can be solved by the standard method which reduces the problem to computing the roots of the associated resolvent cubic. Thus, our solution, which computes the roots of a quartic with this same method, is also algebraically optimal in the sense of [20]. The same observation, of course, holds for the classical quartic-based methods for solving P3P. Alternative P3P solvers that directly employ a cubic [8, 23], despite being superior in practical terms, are not distinguished by the complexity classes \\(\\mathcal{P}_n\\). Table 1 provides a summary of the algebraic complexities, based on the Galois groups computed in [9, §3]." + }, + { + "type": "table", + "bbox": [ + 0.125, + 0.793, + 0.426, + 0.827 + ], + "angle": 0, + "content": "
problemP3PP2P1LP1P2LP3L
classP3P2P3P8
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.837, + 0.47, + 0.88 + ], + "angle": 0, + "content": "Table 1. For each minimal absolute pose problem with points and line features, the class of polynomial root finding algorithms \\(\\mathcal{P}_n\\) that solve the problem with \\(n\\) as small as possible [20, §2.1]." + }, + { + "type": "title", + "bbox": [ + 0.502, + 0.091, + 0.64, + 0.105 + ], + "angle": 0, + "content": "1.3. Related work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.115, + 0.892, + 0.206 + ], + "angle": 0, + "content": "The first solutions to P2P1L and P1P2L were presented by Ramalingam et al. [25], reducing the problems to computing the roots of polynomials of degree 4 and 8 via careful choices of special reference frames in the world and image. Although we obtain polynomials of lower degree, these special reference frames are also an ingredient in our approach." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.207, + 0.893, + 0.372 + ], + "angle": 0, + "content": "In work subsequent to [25], it was observed that both of these problems could be solved using the E3Q3 solver [17]. This is a highly optimized method for computing the points where three quadric surfaces in \\(\\mathbb{R}^3\\) intersect. Typically, there are \\(2^{3} = 8\\) such points, by Bézout's theorem; however, EQ3Q also handles degenerate cases where the number of solutions may drop. In the follow-up work [30], a stabilization scheme for E3Q3 was applied to these problems, and experimentally shown to be more accurate than the solvers in [25]. Efficient solvers for both problems based on this stabilized E3Q3 are implemented in PoseLib [18]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.372, + 0.894, + 0.525 + ], + "angle": 0, + "content": "The idea of using Galois groups to study minimal problems originates from [20]. The main takeaway from this paper is that a problem with \\( n \\) solutions whose Galois group is the full symmetric group \\( S_{n} \\) cannot be solved by an algorithm in \\( \\mathcal{P}_{n-1} \\). In the other direction, the main takeaway from [9] is that if the Galois group is contained in the wreath product \\( S_{n_{1}} \\wr S_{n_{2}} \\), where \\( n = n_{1}n_{2} \\), then the problem can be solved with an algorithm in \\( \\mathcal{P}_{\\max(n_{1},n_{2})} \\). A recent work using such an insight to guide the more efficient solution of a relative pose estimation problem may be found in [14]." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.538, + 0.661, + 0.554 + ], + "angle": 0, + "content": "2. Minimal solvers" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.565, + 0.892, + 0.777 + ], + "angle": 0, + "content": "In this section, we introduce our algebraically optimal solutions to the P2P1L and P1P2L problems. These problems are depicted in Figure 1. As revealed by the Galois group computed in [9, §3], the problem P2P1L can be reduced to computing the roots of a quadratic equation, and the problem P1P2L can be reduced to a quartic equation. In Sections 2.1 and 2.2, we turn these insights into explicit solutions to the P2P1L and P12PL problems, respectively. These solutions work generically—they are valid outside of a measure-zero subset of the space of input point-point/line-line correspondences. In Section 2.3, we describe a method that stabilizes the P1P2L solver of Section 2.2 in a common but non-generic case. Finally, Section 2.4 for a discussion when the point-line configuration is coplanar." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.786, + 0.59, + 0.801 + ], + "angle": 0, + "content": "2.1. P2P1L" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.81, + 0.891, + 0.84 + ], + "angle": 0, + "content": "Here, we provide the algebraically optimal solution to the P2P1L problem. We mostly follow the notation of [25]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.892, + 0.902 + ], + "angle": 0, + "content": "Problem Statement: Consider \\( P_1^0, P_2^0 \\in \\mathbb{R}^3 \\), two 3D points, and their 2D projections under an unknown calibrated camera. Consider also a 3D line spanned by points \\( L_3^0, L_4^0 \\in \\mathbb{R}^3 \\) and its projection. The goal is to recover the" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "21317" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.109, + 0.091, + 0.262, + 0.201 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.286, + 0.092, + 0.43, + 0.201 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.111, + 0.203, + 0.256, + 0.292 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.289, + 0.203, + 0.428, + 0.291 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.3, + 0.468, + 0.327 + ], + "angle": 0, + "content": "Figure 2. Special reference frame for P2P1L. Camera frame \\(\\mathcal{C}_0\\) is transformed to \\(\\mathcal{C}_1\\), world frame \\(\\mathcal{W}_0\\) to \\(\\mathcal{W}_1\\). See text for details." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.353, + 0.453, + 0.369 + ], + "angle": 0, + "content": "unknown camera matrix, whose center we denote by \\(C^0\\)" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.369, + 0.469, + 0.459 + ], + "angle": 0, + "content": "Our first step is to rigidly transform the input into the special reference frame introduced in [25]. Transforming to this special frame simplifies the equations and helps to reveal the algebraically optimal solution. Figure 2 illustrates the input reference frame \\(\\mathcal{C}_0 / \\mathcal{W}_0\\) (top), and the special frame \\(\\mathcal{C}_1 / \\mathcal{W}_1\\) (bottom), which are related as follows:" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.46, + 0.433, + 0.474 + ], + "angle": 0, + "content": "- In the world frame \\(\\mathcal{W}_1\\), the 3D points take the form" + }, + { + "type": "equation", + "bbox": [ + 0.133, + 0.484, + 0.469, + 0.507 + ], + "angle": 0, + "content": "\\[\nP _ {1} = \\left( \\begin{array}{l l l} 0 & 0 & 0 \\end{array} \\right) ^ {T}, \\quad P _ {2} = \\left( \\begin{array}{l l l} X _ {2} & 0 & 0 \\end{array} \\right) ^ {T}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.517, + 0.44, + 0.532 + ], + "angle": 0, + "content": "and the 3D line is spanned by two points of the form" + }, + { + "type": "equation", + "bbox": [ + 0.103, + 0.541, + 0.469, + 0.563 + ], + "angle": 0, + "content": "\\[\nL _ {3} = \\left( \\begin{array}{l l l} X _ {3} & Y _ {3} & 0 \\end{array} \\right) ^ {T}, \\quad L _ {4} = \\left( \\begin{array}{l l l} X _ {4} & Y _ {4} & Z _ {4} \\end{array} \\right) ^ {T}. \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.573, + 0.394, + 0.588 + ], + "angle": 0, + "content": "To find the transformation \\(\\mathcal{W}_0\\to \\mathcal{W}_1\\) , we set" + }, + { + "type": "equation", + "bbox": [ + 0.105, + 0.597, + 0.454, + 0.619 + ], + "angle": 0, + "content": "\\[\nP _ {1} = \\left( \\begin{array}{c c c} 0 & 0 & 0 \\end{array} \\right) ^ {T}, \\quad P _ {2} = \\left(\\| P _ {1} ^ {0} - P _ {2} ^ {2} \\| \\quad 0 \\quad 0\\right) ^ {T},\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.092, + 0.628, + 0.47, + 0.699 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} L _ {3} = \\left( \\begin{array}{c c c} X _ {3} & Y _ {3} & 0 \\end{array} \\right) ^ {T}, \\mathrm {w /} X _ {3} = \\left(L _ {3} ^ {0} - P _ {1}\\right) ^ {T} \\frac {P _ {2} - P _ {1}}{\\| P _ {2} - P _ {1} \\|}, \\\\ Y _ {3} = \\left\\| L _ {3} ^ {0} - P _ {1} - X _ {3} \\cdot \\frac {P _ {2} - P _ {1}}{\\| P _ {2} - P _ {1} \\|} \\right\\|. \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.704, + 0.469, + 0.742 + ], + "angle": 0, + "content": "We transform the rays \\(\\overrightarrow{P_1^0P_2^0}\\) and \\(\\overrightarrow{P_1^0L_3^0}\\) to \\(\\overrightarrow{P_1P_2}\\) and \\(\\overrightarrow{P_1L_3}\\), respectively, via suitable translation and rotation." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.743, + 0.418, + 0.756 + ], + "angle": 0, + "content": "- In the camera frame, the camera center is fixed at" + }, + { + "type": "equation", + "bbox": [ + 0.209, + 0.766, + 0.469, + 0.787 + ], + "angle": 0, + "content": "\\[\nC = \\left( \\begin{array}{l l l} 0 & 0 & - 1 \\end{array} \\right) ^ {T}, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.798, + 0.424, + 0.813 + ], + "angle": 0, + "content": "the 2D points are projections of points of the form" + }, + { + "type": "equation", + "bbox": [ + 0.113, + 0.823, + 0.469, + 0.844 + ], + "angle": 0, + "content": "\\[\nD _ {1} = \\left( \\begin{array}{l l l} a _ {1} & b _ {1} & 0 \\end{array} \\right) ^ {T}, \\quad D _ {2} = \\left( \\begin{array}{l l l} a _ {2} & b _ {2} & 0 \\end{array} \\right) ^ {T}, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.856, + 0.418, + 0.872 + ], + "angle": 0, + "content": "and the 2D line is the projection of \\(\\overline{D_3D_4}\\), where" + }, + { + "type": "equation", + "bbox": [ + 0.136, + 0.882, + 0.469, + 0.903 + ], + "angle": 0, + "content": "\\[\nD _ {3} = \\left( \\begin{array}{l l l} 0 & 0 & 0 \\end{array} \\right) ^ {T}, \\quad D _ {4} = \\left( \\begin{array}{l l l} 1 & 0 & 0 \\end{array} \\right) ^ {T}. \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.091, + 0.892, + 0.173 + ], + "angle": 0, + "content": "- To find the transformation \\(\\mathcal{C}_0\\to \\mathcal{C}_1\\) , let \\(D_3^0\\) and \\(D_4^0\\) be the homogeneous coordinates of two distinct points along the given line. Independently of \\(C^0\\) , we may rigidly transform the rays \\(d_{3}\\coloneqq \\overrightarrow{C^{0}D_{3}^{0}}\\) , and \\(d_4\\coloneqq \\overrightarrow{C^0D_3^0}\\) to \\(\\overrightarrow{CD_3}\\) and \\(\\overrightarrow{CD_4}\\) , by suitable rotation and translation." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.174, + 0.892, + 0.219 + ], + "angle": 0, + "content": "Let us now write \\( R = (R_{ij})_{1 \\leq i,j \\leq 3} \\), \\( T = (T_i)_{1 \\leq i \\leq 3} \\) for the unknown camera pose in this special reference frame. Our projection constraints can then be written as" + }, + { + "type": "equation", + "bbox": [ + 0.554, + 0.232, + 0.891, + 0.248 + ], + "angle": 0, + "content": "\\[\n\\left(D _ {i} - C\\right) \\times \\left(R P _ {i} + T\\right) = 0, \\quad i \\in \\{1, 2 \\}, \\tag {6}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.555, + 0.251, + 0.891, + 0.269 + ], + "angle": 0, + "content": "\\[\n\\left(D _ {3} \\times D _ {4}\\right) ^ {T} \\left(R L _ {i} + T\\right) = 0, \\quad i \\in \\{3, 4 \\}. \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.281, + 0.892, + 0.327 + ], + "angle": 0, + "content": "This gives a system of equations in 12 unknowns. However, in what follows, we consider a system of equations in a smaller set of unknowns, namely" + }, + { + "type": "equation", + "bbox": [ + 0.513, + 0.337, + 0.891, + 0.369 + ], + "angle": 0, + "content": "\\[\n\\mathbf {x} ^ {T} = \\left( \\begin{array}{l l l l l l l l} R _ {1 1} & R _ {2 1} & R _ {3 1} & R _ {2 2} & R _ {2 3} & T _ {1} & T _ {2} & T _ {3} \\end{array} \\right). \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.37, + 0.734, + 0.385 + ], + "angle": 0, + "content": "The entries of \\(\\mathbf{x}\\) are constrained by" + }, + { + "type": "equation", + "bbox": [ + 0.547, + 0.396, + 0.891, + 0.415 + ], + "angle": 0, + "content": "\\[\nR _ {1 1} ^ {2} + R _ {2 1} ^ {2} + R _ {3 1} ^ {2} = R _ {2 1} ^ {2} + R _ {2 2} ^ {2} + R _ {2 3} ^ {2} = 1. \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.426, + 0.892, + 0.471 + ], + "angle": 0, + "content": "Thus, the entries of \\(\\mathbf{x}\\) specify a translation vector and a partially-filled rotation matrix, which can be uniquely completed to a rotation matrix using the formulae" + }, + { + "type": "equation", + "bbox": [ + 0.503, + 0.482, + 0.893, + 0.561 + ], + "angle": 0, + "content": "\\[\n\\binom {R _ {1 2}} {R _ {1 3}} = \\left(R _ {2 2} ^ {2} + R _ {2 3} ^ {2}\\right) ^ {- 1} \\cdot \\binom {- R _ {1 1} R _ {2 1} R _ {2 2} + R _ {2 3} R _ {3 1}} {- R _ {1 1} R _ {2 1} R _ {2 3} - R _ {2 2} R _ {3 1}} \\\\ - R _ {2 1} R _ {2 2} R _ {3 1} - R _ {1 1} R _ {2 3} \\\\ - R _ {2 1} R _ {2 3} R _ {3 1} + R _ {1 1} R _ {2 2}) \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.574, + 0.892, + 0.651 + ], + "angle": 0, + "content": "subject to the genericity condition \\( R_{22}^2 + R_{23}^2 \\neq 0 \\), i.e. when \\( R \\) is not a rotation in the \\( xz \\)-plane. Hence, we focus on recovering the entries of \\( \\mathbf{x} \\). From 2 out of the 3 redundant constraints from (6) together with (7), we obtain linear constraints on \\( \\mathbf{x} \\) ([25, eq. (4)-(5)])," + }, + { + "type": "equation", + "bbox": [ + 0.548, + 0.664, + 0.671, + 0.678 + ], + "angle": 0, + "content": "\\[\nA \\mathbf {x} = \\mathbf {b}, \\quad \\text {w h e r e}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.558, + 0.682, + 0.845, + 0.74 + ], + "angle": 0, + "content": "\\[\nA = \\left( \\begin{array}{c c c c c c c c} 0 & 0 & 0 & 0 & 0 & - b _ {1} & a _ {1} & 0 \\\\ 0 & 0 & 0 & 0 & 0 & 0 & - 1 & b _ {1} \\\\ - b _ {2} X _ {2} & a _ {2} X _ {2} & 0 & 0 & 0 & - b _ {2} & a _ {2} & 0 \\\\ 0 & - X _ {2} & b _ {2} X _ {2} & 0 & 0 & 0 & - 1 & b _ {2} \\\\ 0 & X _ {3} & 0 & Y _ {3} & 0 & 0 & 1 & 0 \\\\ 0 & X _ {4} & 0 & Y _ {4} & Z _ {4} & 0 & 1 & 0 \\end{array} \\right),\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.548, + 0.741, + 0.891, + 0.76 + ], + "angle": 0, + "content": "\\[\n\\mathbf {b} ^ {T} = \\left( \\begin{array}{c c c c c c} 0 & - b _ {1} & 0 & - b _ {2} & 0 & 0 \\end{array} \\right). \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.772, + 0.89, + 0.801 + ], + "angle": 0, + "content": "We may use these linear equations to solve for the translation \\(T\\) in terms of \\(R\\) and the problem data as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.524, + 0.814, + 0.867, + 0.83 + ], + "angle": 0, + "content": "\\[\nT _ {1} = a _ {1} B, \\quad T _ {2} = b _ {1} B, \\quad T _ {3} = - 1 + B, \\quad \\text {w h e r e}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.53, + 0.834, + 0.891, + 0.866 + ], + "angle": 0, + "content": "\\[\nB = \\frac {X _ {2} \\left(a _ {2} R _ {2 1} - b _ {2} R _ {1 1}\\right)}{b _ {2} a _ {1} - b _ {1} a _ {2}}. \\tag {12}\n\\]" + }, + { + "type": "page_footnote", + "bbox": [ + 0.5, + 0.875, + 0.89, + 0.9 + ], + "angle": 0, + "content": "1If \\(R\\) is known to be a plane rotation, two generic point-point correspondences suffice to recover \\(R\\) and \\(T\\)." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "21318" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.071, + 0.088, + 0.468, + 0.118 + ], + "angle": 0, + "content": "(1) Transform data in world and camera frames to the special reference frames satisfying (1), (2), (3), (4), (5)." + }, + { + "type": "text", + "bbox": [ + 0.071, + 0.118, + 0.445, + 0.133 + ], + "angle": 0, + "content": "(2) Compute up to 2 solutions in \\(v\\) to the quadratic (16)." + }, + { + "type": "text", + "bbox": [ + 0.071, + 0.133, + 0.469, + 0.163 + ], + "angle": 0, + "content": "(3) For each root \\( v \\) in step (2), recover a corresponding value of \\( u \\) using either of the linear expressions in (46)." + }, + { + "type": "text", + "bbox": [ + 0.071, + 0.163, + 0.44, + 0.179 + ], + "angle": 0, + "content": "(4) Recover up to 4 solutions in \\(\\{R_{11}, R_{21}\\}\\) using (15)." + }, + { + "type": "text", + "bbox": [ + 0.072, + 0.179, + 0.363, + 0.193 + ], + "angle": 0, + "content": "(5) Recover \\(R\\) and \\(T\\) using (13), (12), (10)." + }, + { + "type": "text", + "bbox": [ + 0.071, + 0.193, + 0.405, + 0.208 + ], + "angle": 0, + "content": "(6) Reverse the transformation applied in step (1)." + }, + { + "type": "list", + "bbox": [ + 0.071, + 0.088, + 0.469, + 0.208 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.102, + 0.21, + 0.443, + 0.224 + ], + "angle": 0, + "content": "Figure 3. Generic P2P1L solver. See Sec. 2.1 for details." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.261, + 0.469, + 0.292 + ], + "angle": 0, + "content": "Moreover, from (11) we may use \\( R_{11} \\) and \\( R_{21} \\), to express the remaining entries of \\( R \\) as" + }, + { + "type": "equation", + "bbox": [ + 0.08, + 0.297, + 0.345, + 0.33 + ], + "angle": 0, + "content": "\\[\nR _ {3 1} = \\frac {(a _ {1} - a _ {2}) R _ {2 1} + (b _ {2} - b _ {1}) R _ {1 1})}{b _ {2} a _ {1} - b _ {1} a _ {2}},\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.082, + 0.332, + 0.468, + 0.365 + ], + "angle": 0, + "content": "\\[\nR _ {2 2} = \\frac {b _ {1} b _ {2} X _ {2} R _ {1 1} + (b _ {1} a _ {2} X _ {2} - b _ {1} a _ {2} X _ {2} - b _ {2} a _ {1} X _ {3}) R _ {2 1}}{Y _ {3} (b _ {2} a _ {1} - b _ {1} a _ {2})},\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.08, + 0.372, + 0.469, + 0.461 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} R _ {2 3} = \\frac {b _ {1} b _ {2} X _ {2} \\left(Y _ {3} - Y _ {4}\\right) R _ {1 1}}{Y _ {3} Z _ {4} \\left(b _ {2} a _ {1} - b _ {1} a _ {2}\\right)} \\\\ + \\left(\\frac {X _ {3} Y _ {4} - X _ {4} Y _ {3}}{Y _ {3} Z _ {4}} + \\frac {b _ {2} a _ {1} X _ {3} Y _ {4} - b _ {1} a _ {2} X _ {2} Y _ {3}}{Y _ {3} Z _ {4} \\left(b _ {2} a _ {1} - b _ {1} a _ {2}\\right)}\\right) R _ {2 1}. \\tag {13} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.47, + 0.469, + 0.5 + ], + "angle": 0, + "content": "Substituting (13) into (9), we obtain two bivariate quadratic constraints in \\( R_{11} \\) and \\( R_{21} \\). In matrix form," + }, + { + "type": "equation", + "bbox": [ + 0.143, + 0.507, + 0.469, + 0.556 + ], + "angle": 0, + "content": "\\[\n\\left( \\begin{array}{l l l} c _ {1} & c _ {2} & c _ {3} \\\\ d _ {1} & d _ {2} & d _ {3} \\end{array} \\right) \\cdot \\left( \\begin{array}{c} R _ {1 1} ^ {2} \\\\ R _ {1 1} R _ {2 1} \\\\ R _ {2 1} ^ {2} \\end{array} \\right) = \\binom {1} {1}, \\tag {14}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.563, + 0.469, + 0.593 + ], + "angle": 0, + "content": "where the coefficients \\(c_{1}, c_{2}, c_{3}, d_{1}, d_{2}, d_{3}\\) are rational functions of the problem data. Applying the change of variables" + }, + { + "type": "equation", + "bbox": [ + 0.182, + 0.606, + 0.469, + 0.624 + ], + "angle": 0, + "content": "\\[\nu = R _ {1 1} ^ {2}, \\quad v = R _ {2 1} / R _ {1 1}, \\tag {15}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.628, + 0.422, + 0.643 + ], + "angle": 0, + "content": "and subtracting the two equations in (14), we obtain" + }, + { + "type": "equation", + "bbox": [ + 0.114, + 0.651, + 0.432, + 0.669 + ], + "angle": 0, + "content": "\\[\n\\left(c _ {1} - d _ {1}\\right) u + \\left(c _ {2} - d _ {2}\\right) u v + \\left(c _ {3} - d _ {3}\\right) u v ^ {2} = 0.\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.677, + 0.469, + 0.707 + ], + "angle": 0, + "content": "Assuming \\( u \\neq 0 \\), we therefore have the univariate quadratic equation in \\( v \\)" + }, + { + "type": "equation", + "bbox": [ + 0.114, + 0.715, + 0.469, + 0.732 + ], + "angle": 0, + "content": "\\[\n\\left(c _ {1} - d _ {1}\\right) + \\left(c _ {2} - d _ {2}\\right) v + \\left(c _ {3} - d _ {3}\\right) v ^ {2} = 0. \\tag {16}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.741, + 0.469, + 0.772 + ], + "angle": 0, + "content": "If \\( v \\) is one of the roots of (16), we may recover a corresponding value for \\( u \\) using one of the equations in (14), eg." + }, + { + "type": "equation", + "bbox": [ + 0.078, + 0.793, + 0.469, + 0.825 + ], + "angle": 0, + "content": "\\[\nu = \\left(c _ {1} + c _ {2} v + c _ {3} v ^ {2}\\right) ^ {- 1}, \\quad \\text {o r} \\quad u = \\left(d _ {1} + d _ {2} v + d _ {3} v ^ {2}\\right) ^ {- 1}. \\tag {17}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.826, + 0.468, + 0.856 + ], + "angle": 0, + "content": "To summarize, we provide the outline of steps for solving the P2P1L absolute pose problem in Figure 3." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.856, + 0.469, + 0.902 + ], + "angle": 0, + "content": "Algebraically, the nontrivial steps in Figure 3 are the second and fourth, which require solving a univariate quadratic equation and computing square roots, respectively." + }, + { + "type": "image", + "bbox": [ + 0.525, + 0.092, + 0.673, + 0.2 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.704, + 0.098, + 0.859, + 0.2 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.528, + 0.202, + 0.67, + 0.291 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.713, + 0.203, + 0.85, + 0.291 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.3, + 0.892, + 0.327 + ], + "angle": 0, + "content": "Figure 4. Special reference frame for P1P2L. Camera frame \\(\\mathcal{C}_0\\) is transformed to \\(\\mathcal{C}_1\\), world frame \\(\\mathcal{W}_0\\) to \\(\\mathcal{W}_1\\). See text for details." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.355, + 0.589, + 0.369 + ], + "angle": 0, + "content": "2.2.P1P2L" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.38, + 0.891, + 0.409 + ], + "angle": 0, + "content": "Here, we provide the algebraically optimal solution to the P1P2L problem. We mostly follow the notation of [25]." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.41, + 0.892, + 0.531 + ], + "angle": 0, + "content": "Problem Statement: Let us consider a 3D point \\( P_1^0 \\in \\mathbb{R}^3 \\), and its homogeneous 2D projection \\( D_1 \\in \\mathbb{R}^3 \\) under an unknown calibrated camera matrix. Let us also consider two 3D lines—the first spanned by points \\( L_1, L_2 \\in \\mathbb{R}^3 \\), and the second one is spanned by \\( L_3, L_4 \\in \\mathbb{R}^3 \\)—and both of their corresponding projections. Our task is to recover the camera matrix from the given point-point correspondence and line-line correspondences." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.532, + 0.891, + 0.578 + ], + "angle": 0, + "content": "Much like the P2P1L solver, our P1P2L solver begins by transforming the input data into a special reference frame, as illustrated in Figure 4. Specifically," + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.579, + 0.827, + 0.593 + ], + "angle": 0, + "content": "- In the world frame, the 3D point takes the form" + }, + { + "type": "equation", + "bbox": [ + 0.636, + 0.606, + 0.891, + 0.628 + ], + "angle": 0, + "content": "\\[\nP _ {1} = \\left( \\begin{array}{l l l} 0 & 0 & 0 \\end{array} \\right) ^ {T}. \\tag {18}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.64, + 0.74, + 0.655 + ], + "angle": 0, + "content": "The first line is spanned by points" + }, + { + "type": "equation", + "bbox": [ + 0.529, + 0.667, + 0.891, + 0.702 + ], + "angle": 0, + "content": "\\[\nL _ {1} = \\left( \\begin{array}{l l l} X _ {1} & Y _ {1} & Z _ {1} \\end{array} \\right) ^ {T}, \\quad L _ {2} = \\left( \\begin{array}{l l l} X _ {2} & Y _ {2} & Z _ {2} \\end{array} \\right) ^ {T}, \\tag {19}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.703, + 0.781, + 0.717 + ], + "angle": 0, + "content": "and the second line is spanned by points" + }, + { + "type": "equation", + "bbox": [ + 0.529, + 0.729, + 0.891, + 0.763 + ], + "angle": 0, + "content": "\\[\nL _ {3} = \\left( \\begin{array}{l l l} X _ {3} & Y _ {3} & Z _ {3} \\end{array} \\right) ^ {T}, \\quad L _ {4} = \\left( \\begin{array}{l l l} X _ {4} & Y _ {4} & Z _ {4} \\end{array} \\right) ^ {T}. \\tag {20}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.764, + 0.818, + 0.78 + ], + "angle": 0, + "content": "- For \\(\\mathcal{W}_0 \\to \\mathcal{W}_1\\), we simply translate \\(P_1^0\\) to \\(P_1\\)." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.78, + 0.892, + 0.853 + ], + "angle": 0, + "content": "- The world frame may still be rotated freely. We may use the strategy of Section 2.3, which makes the solver stable for a larger class of non-coplanar scenes. However, if \\( P_{1}, \\overline{L_{1}L_{2}} \\), and \\( \\overline{L_{3}L_{4}} \\) are coplanar, no rotation is recommended." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.856, + 0.841, + 0.869 + ], + "angle": 0, + "content": "- In the camera frame, the camera center is fixed at" + }, + { + "type": "list", + "bbox": [ + 0.5, + 0.764, + 0.892, + 0.869 + ], + "angle": 0, + "content": null + }, + { + "type": "equation", + "bbox": [ + 0.631, + 0.882, + 0.891, + 0.903 + ], + "angle": 0, + "content": "\\[\nC = \\left( \\begin{array}{l l l} 0 & 0 & - 1 \\end{array} \\right) ^ {T}. \\tag {21}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "21319" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.091, + 0.092, + 0.457, + 0.107 + ], + "angle": 0, + "content": "The image point is the projection of a point of the form" + }, + { + "type": "equation", + "bbox": [ + 0.205, + 0.116, + 0.469, + 0.138 + ], + "angle": 0, + "content": "\\[\nD _ {1} = \\left( \\begin{array}{l l l} a _ {1} & b _ {1} & 0 \\end{array} \\right) ^ {T}, \\tag {22}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.147, + 0.47, + 0.178 + ], + "angle": 0, + "content": "the first line in the image is the projection of a line spanned by points of the form" + }, + { + "type": "equation", + "bbox": [ + 0.119, + 0.186, + 0.469, + 0.209 + ], + "angle": 0, + "content": "\\[\nD _ {2} = \\left( \\begin{array}{l l l} a _ {2} & 0 & 0 \\end{array} \\right) ^ {T}, \\quad D _ {3} = \\left( \\begin{array}{l l l} 0 & 0 & 0 \\end{array} \\right) ^ {T}, \\tag {23}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.218, + 0.47, + 0.249 + ], + "angle": 0, + "content": "and the second line in the image is the projection of a line spanned by points of the form" + }, + { + "type": "equation", + "bbox": [ + 0.109, + 0.257, + 0.469, + 0.28 + ], + "angle": 0, + "content": "\\[\nD _ {4} = \\left( \\begin{array}{l l l} a _ {4} & b _ {4} & 0 \\end{array} \\right) ^ {T}, \\quad D _ {5} = \\left( \\begin{array}{l l l} a _ {5} & b _ {5} & 0 \\end{array} \\right) ^ {T}. \\tag {24}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.289, + 0.469, + 0.334 + ], + "angle": 0, + "content": "- The homogeneous coordinates of the lines in the camera frame are \\( n_1 = d_2 \\times d_3 \\), \\( n_2 = d_4 \\times d_5 \\), where the rays \\( d_2 \\neq d_3 \\), \\( d_4 \\neq d_5 \\) to meet the lines in distinct points." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.334, + 0.4, + 0.349 + ], + "angle": 0, + "content": "To find the transformation \\(\\mathcal{C}_0\\to \\mathcal{C}_1\\) , we define" + }, + { + "type": "list", + "bbox": [ + 0.078, + 0.289, + 0.469, + 0.349 + ], + "angle": 0, + "content": null + }, + { + "type": "equation", + "bbox": [ + 0.091, + 0.358, + 0.469, + 0.391 + ], + "angle": 0, + "content": "\\[\nd _ {1 2} = n _ {1} \\times n _ {2}, \\quad D _ {2} ^ {0} = C ^ {0} + \\frac {d _ {2}}{d _ {2} ^ {T} d _ {1 2}}, \\quad D _ {3} ^ {0} = C ^ {0} + d _ {3},\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.091, + 0.4, + 0.501, + 0.422 + ], + "angle": 0, + "content": "\\[\nD _ {2} = \\left( \\begin{array}{c c} \\tan (\\cos^ {- 1} (d _ {2} ^ {T} d _ {1 2})) & 0 \\quad 0 \\end{array} \\right) ^ {T}, \\quad D _ {3} = \\left( \\begin{array}{c c c} 0 & 0 & 0 \\end{array} \\right) ^ {T},\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.426, + 0.469, + 0.461 + ], + "angle": 0, + "content": "and, independently of \\(C^0\\), rigidly transform the rays \\(d_2 \\coloneqq \\overrightarrow{C^0D_2^0}\\) and \\(d_3 \\coloneqq \\overrightarrow{C^0D_3^0}\\) to \\(\\overrightarrow{C^0D_2^0}\\) and \\(\\overrightarrow{CD_3}\\)." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.461, + 0.469, + 0.504 + ], + "angle": 0, + "content": "In the special reference frame, we now write the pose as \\((R_{ij})_{1\\leq i,j\\leq 3}\\), \\((T_i)_{1\\leq i\\leq 3}\\). The projection constraints can be formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.174, + 0.516, + 0.469, + 0.533 + ], + "angle": 0, + "content": "\\[\n\\left(D _ {1} - C\\right) \\times \\left(R P _ {1} + T\\right) = 0, \\tag {25}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.139, + 0.556, + 0.469, + 0.575 + ], + "angle": 0, + "content": "\\[\n\\left(D _ {2} \\times D _ {3}\\right) ^ {T} \\left(R L _ {i} + T\\right) = 0, i \\in \\{1, 2 \\}, \\tag {26}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.139, + 0.594, + 0.469, + 0.612 + ], + "angle": 0, + "content": "\\[\n\\left(D _ {4} \\times D _ {5}\\right) ^ {T} \\left(R L _ {i} + T\\right) = 0, i \\in \\{3, 4 \\}. \\tag {27}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.618, + 0.469, + 0.678 + ], + "angle": 0, + "content": "Analogously to (11), we have a system of linear equations \\( A\\mathbf{x} = \\mathbf{b} \\), obtained by picking 2 out of the 3 redundant constraints from (25) together with (26), (27), where now (cf. [25, eq. 7-8])" + }, + { + "type": "equation", + "bbox": [ + 0.078, + 0.687, + 0.432, + 0.745 + ], + "angle": 0, + "content": "\\[\nA = \\left( \\begin{array}{c c c c c c c c c} 0 & 0 & 0 & 0 & 0 & 0 & - b _ {1} & a _ {1} & 0 \\\\ 0 & 0 & 0 & 0 & 0 & 0 & 0 & - 1 & b _ {1} \\\\ 0 & 0 & X _ {1} & Y _ {1} & Y _ {1} & Z _ {1} & 0 & 1 & 0 \\\\ 0 & 0 & X _ {2} & Y _ {2} & Y _ {2} & Z _ {2} & 0 & 1 & 0 \\\\ - b _ {4} X _ {3} & - b _ {4} Y _ {3} & a _ {4} X _ {3} & a _ {4} Y _ {3} & a _ {4} Y _ {3} & a _ {4} Z _ {3} & - b _ {4} & a _ {4} & 0 \\\\ - b _ {4} X _ {4} & - b _ {4} Z _ {4} & a _ {4} X _ {4} & a _ {4} Y _ {4} & a _ {4} Y _ {4} & a _ {4} Z _ {4} & - b _ {4} & a _ {4} & 0 \\end{array} \\right),\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.08, + 0.746, + 0.486, + 0.767 + ], + "angle": 0, + "content": "\\[\n\\mathbf {x} = \\left( \\begin{array}{c c c c c c c c c} R _ {1 1} & R _ {1 2} & R _ {1 3} & R _ {2 1} & R _ {2 2} & R _ {2 3} & T _ {1} & T _ {2} & T _ {3} \\end{array} \\right) ^ {T},\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.08, + 0.769, + 0.298, + 0.79 + ], + "angle": 0, + "content": "\\[\n\\mathbf {b} = \\left( \\begin{array}{c c c c c c} 0 & - b _ {1} & 0 & 0 & 0 & 0 \\end{array} \\right) ^ {T}.\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.8, + 0.357, + 0.815 + ], + "angle": 0, + "content": "Solving for translation as in (12), we have" + }, + { + "type": "equation", + "bbox": [ + 0.106, + 0.824, + 0.374, + 0.853 + ], + "angle": 0, + "content": "\\[\nT _ {1} = \\frac {a _ {1}}{b _ {1}} (- R _ {2 1} X _ {1} - R _ {2 2} Y _ {1} - R _ {2 3} Z _ {1}),\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.107, + 0.856, + 0.469, + 0.872 + ], + "angle": 0, + "content": "\\[\nT _ {2} = - R _ {2 1} X _ {1} - R _ {2 2} Y _ {1} - R _ {2 3} Z _ {1}, \\tag {28}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.107, + 0.875, + 0.412, + 0.905 + ], + "angle": 0, + "content": "\\[\nT _ {3} = - 1 + \\frac {1}{b _ {1}} \\left(- R _ {2 1} X _ {1} - R _ {2 2} Y _ {1} - R _ {2 3} Z _ {1}\\right).\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.092, + 0.88, + 0.108 + ], + "angle": 0, + "content": "Furthermore, we express \\( R_{23} \\) from the fourth equation as" + }, + { + "type": "equation", + "bbox": [ + 0.511, + 0.112, + 0.892, + 0.142 + ], + "angle": 0, + "content": "\\[\nR _ {2 3} = \\frac {1}{Z _ {1} - Z _ {2}} \\left(R _ {2 1} \\left(X _ {2} - X _ {1}\\right) + R _ {2 2} \\left(Y _ {2} - Y _ {1}\\right)\\right), \\tag {29}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.146, + 0.867, + 0.16 + ], + "angle": 0, + "content": "and we solve for \\( R_{11}, R_{12} \\) using the last two equations." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.161, + 0.892, + 0.191 + ], + "angle": 0, + "content": "Now, we can express \\( R_{23}, T_2, T_1, R_{12}, R_{11} \\) as linear combinations of \\( R_{13}, R_{21}, R_{22} \\) in the following form:" + }, + { + "type": "equation", + "bbox": [ + 0.577, + 0.197, + 0.735, + 0.212 + ], + "angle": 0, + "content": "\\[\nR _ {2 3} = c _ {1} R _ {2 1} + c _ {2} R _ {2 2},\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.586, + 0.217, + 0.735, + 0.231 + ], + "angle": 0, + "content": "\\[\nT _ {2} = c _ {3} R _ {2 1} + c _ {4} R _ {2 2},\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.586, + 0.235, + 0.891, + 0.25 + ], + "angle": 0, + "content": "\\[\nT _ {1} = c _ {5} R _ {2 1} + c _ {6} R _ {2 2}, \\tag {30}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.577, + 0.254, + 0.795, + 0.269 + ], + "angle": 0, + "content": "\\[\nR _ {1 2} = c _ {7} R _ {1 3} + c _ {8} R _ {2 1} + c _ {9} R _ {2 2},\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.579, + 0.273, + 0.815, + 0.287 + ], + "angle": 0, + "content": "\\[\nR _ {1 1} = c _ {1 0} R _ {1 3} + c _ {1 1} R _ {2 1} + c _ {1 2} R _ {2 2}.\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.294, + 0.891, + 0.322 + ], + "angle": 0, + "content": "The non-linear internal constraints imposed on the elements of \\( R \\) have the form:" + }, + { + "type": "equation", + "bbox": [ + 0.658, + 0.325, + 0.813, + 0.342 + ], + "angle": 0, + "content": "\\[\nR _ {1 1} ^ {2} + R _ {1 2} ^ {2} + R _ {1 3} ^ {2} = 1,\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.658, + 0.346, + 0.891, + 0.363 + ], + "angle": 0, + "content": "\\[\nR _ {2 1} ^ {2} + R _ {2 2} ^ {2} + R _ {2 3} ^ {2} = 1, \\tag {31}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.58, + 0.367, + 0.813, + 0.381 + ], + "angle": 0, + "content": "\\[\nR _ {1 1} R _ {2 1} + R _ {1 2} R _ {2 2} + R _ {1 3} R _ {2 3} = 0.\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.387, + 0.891, + 0.403 + ], + "angle": 0, + "content": "We substitute (30) into (31), obtaining equations of the form" + }, + { + "type": "equation", + "bbox": [ + 0.563, + 0.423, + 0.891, + 0.44 + ], + "angle": 0, + "content": "\\[\n\\left( \\begin{array}{l l l l l l} d _ {1} & d _ {2} & d _ {3} & d _ {4} & d _ {5} & d _ {6} \\end{array} \\right) \\nu (R) = 1, \\tag {32}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.586, + 0.444, + 0.891, + 0.46 + ], + "angle": 0, + "content": "\\[\n\\left( \\begin{array}{c c c c c c} 0 & 0 & 0 & d _ {7} & d _ {8} & d _ {9} \\end{array} \\right) \\nu (R) = 1, \\tag {33}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.539, + 0.464, + 0.891, + 0.481 + ], + "angle": 0, + "content": "\\[\n\\left( \\begin{array}{c c c c c c} 0 & d _ {1 0} & d _ {1 1} & d _ {1 2} & d _ {1 3} & d _ {1 4} \\end{array} \\right) \\nu (R) = 0, \\tag {34}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.487, + 0.545, + 0.499 + ], + "angle": 0, + "content": "where" + }, + { + "type": "equation", + "bbox": [ + 0.5, + 0.503, + 0.892, + 0.538 + ], + "angle": 0, + "content": "\\[\n\\nu (R) = \\left( \\begin{array}{l l l l l} R _ {1 3} ^ {2} & R _ {1 3} R _ {2 1} & R _ {1 3} R _ {2 2} & R _ {2 1} ^ {2} R _ {2 1} R _ {2 2} & R _ {2 2} ^ {2} \\end{array} \\right) ^ {T}. \\tag {35}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.539, + 0.695, + 0.553 + ], + "angle": 0, + "content": "We express \\(R_{13}\\) from (34) as" + }, + { + "type": "equation", + "bbox": [ + 0.54, + 0.557, + 0.891, + 0.59 + ], + "angle": 0, + "content": "\\[\nR _ {1 3} = - \\frac {d _ {1 2} R _ {2 1} ^ {2} + d _ {1 3} R _ {2 1} R _ {2 2} + d _ {1 4} R _ {2 2} ^ {2}}{d _ {1 0} R _ {2 1} + d _ {1 1} R _ {2 2}}. \\tag {36}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.594, + 0.748, + 0.609 + ], + "angle": 0, + "content": "Substituting (36) into (32), we obtain" + }, + { + "type": "equation", + "bbox": [ + 0.557, + 0.613, + 0.828, + 0.647 + ], + "angle": 0, + "content": "\\[\nd _ {1} \\frac {\\left(d _ {1 2} R _ {2 1} ^ {2} + d _ {1 3} R _ {2 1} R _ {2 2} + d _ {1 4} R _ {2 2} ^ {2}\\right) ^ {2}}{\\left(d _ {1 0} R _ {2 1} + d _ {1 1} R _ {2 2}\\right) ^ {2}} -\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.558, + 0.649, + 0.834, + 0.682 + ], + "angle": 0, + "content": "\\[\nd _ {2} \\frac {d _ {1 2} R _ {2 1} ^ {3} + d _ {1 3} R _ {2 1} ^ {2} R _ {2 2} + d _ {1 4} R _ {2 1} R _ {2 2} ^ {2}}{d _ {1 0} R _ {2 1} + d _ {1 1} R _ {2 2}} -\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.558, + 0.684, + 0.834, + 0.716 + ], + "angle": 0, + "content": "\\[\nd _ {3} \\frac {d _ {1 2} R _ {2 1} ^ {2} R _ {2 2} + d _ {1 3} R _ {2 1} R _ {2 2} ^ {2} + d _ {1 4} R _ {2 2} ^ {3}}{d _ {1 0} R _ {2 1} + d _ {1 1} R _ {2 2}} +\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.558, + 0.719, + 0.891, + 0.735 + ], + "angle": 0, + "content": "\\[\nd _ {4} R _ {2 1} ^ {2} + d _ {5} R _ {2 1} R _ {2 2} + d _ {6} R _ {2 2} ^ {2} = 1. \\tag {37}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.741, + 0.682, + 0.756 + ], + "angle": 0, + "content": "Subtracting (33) from (37)," + }, + { + "type": "equation", + "bbox": [ + 0.508, + 0.759, + 0.779, + 0.794 + ], + "angle": 0, + "content": "\\[\nd _ {1} \\frac {\\left(d _ {1 2} R _ {2 1} ^ {2} + d _ {1 3} R _ {2 1} R _ {2 2} + d _ {1 4} R _ {2 2} ^ {2}\\right) ^ {2}}{\\left(d _ {1 0} R _ {2 1} + d _ {1 1} R _ {2 2}\\right) ^ {2}} -\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.508, + 0.796, + 0.784, + 0.828 + ], + "angle": 0, + "content": "\\[\nd _ {2} \\frac {d _ {1 2} R _ {2 1} ^ {3} + d _ {1 3} R _ {2 1} ^ {2} R _ {2 2} + d _ {1 4} R _ {2 1} R _ {2 2} ^ {2}}{d _ {1 0} R _ {2 1} + d _ {1 1} R _ {2 2}} -\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.508, + 0.831, + 0.784, + 0.863 + ], + "angle": 0, + "content": "\\[\nd _ {3} \\frac {d _ {1 2} R _ {2 1} ^ {2} R _ {2 2} + d _ {1 3} R _ {2 1} R _ {2 2} ^ {2} + d _ {1 4} R _ {2 2} ^ {3}}{d _ {1 0} R _ {2 1} + d _ {1 1} R _ {2 2}} +\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.508, + 0.865, + 0.891, + 0.896 + ], + "angle": 0, + "content": "\\[\n\\left(d _ {4} - d _ {7}\\right) R _ {2 1} ^ {2} + \\left(d _ {5} - d _ {8}\\right) R _ {2 1} R _ {2 2} + \\left(d _ {6} - d _ {9}\\right) R _ {2 2} ^ {2} = 0. \\tag {38}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "21320" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.07, + 0.088, + 0.468, + 0.132 + ], + "angle": 0, + "content": "(1) Transform data in world and camera frames to the special reference frames satisfying (18), (19), (20), (21), (22), (23), (24)." + }, + { + "type": "text", + "bbox": [ + 0.071, + 0.133, + 0.43, + 0.148 + ], + "angle": 0, + "content": "(2) Compute up to 4 solutions in \\(v\\) to the quartic (41)." + }, + { + "type": "text", + "bbox": [ + 0.071, + 0.148, + 0.469, + 0.178 + ], + "angle": 0, + "content": "(3) For each root \\( v \\) in step (2), recover 2 solutions in \\( \\{R_{21}, R_{22}\\} \\) using \\( R_{21} = R_{22}v \\) and (33)." + }, + { + "type": "text", + "bbox": [ + 0.071, + 0.179, + 0.469, + 0.208 + ], + "angle": 0, + "content": "(4) Recover remaining coordinates for each solution using (34),(30)." + }, + { + "type": "text", + "bbox": [ + 0.071, + 0.209, + 0.405, + 0.224 + ], + "angle": 0, + "content": "(5) Reverse the transformation applied in step (1)." + }, + { + "type": "list", + "bbox": [ + 0.07, + 0.088, + 0.469, + 0.224 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.102, + 0.225, + 0.443, + 0.239 + ], + "angle": 0, + "content": "Figure 5. Generic P1P2L solver. See Sec. 2.2 for details." + }, + { + "type": "image", + "bbox": [ + 0.119, + 0.274, + 0.255, + 0.368 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.286, + 0.273, + 0.422, + 0.368 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.376, + 0.469, + 0.404 + ], + "angle": 0, + "content": "Figure 6. Stabilization of P1P2L solver. The frame is rotated to align the line \\(\\overline{L_1L_2}\\) with the z-axis. See Sec. 2.3 for details." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.435, + 0.469, + 0.467 + ], + "angle": 0, + "content": "We then clear denominators in (38), multiplying by \\((d_{10}R_{21} + d_{11}R_{22})^2\\) to get a polynomial equation" + }, + { + "type": "equation", + "bbox": [ + 0.109, + 0.48, + 0.469, + 0.637 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} d _ {1} \\left(d _ {1 2} R _ {2 1} ^ {2} + d _ {1 3} R _ {2 1} R _ {2 2} + d _ {1 4} R _ {2 2} ^ {2}\\right) ^ {2} - \\\\ \\left(\\left(d _ {2} \\left(d _ {1 2} R _ {2 1} ^ {3} + d _ {1 3} R _ {2 1} ^ {2} R _ {2 2} + d _ {1 4} R _ {2 1} R _ {2 2} ^ {2}\\right) - \\right. \\right. \\\\ \\left. d _ {3} \\left(d _ {1 2} R _ {2 1} ^ {2} R _ {2 2} + d _ {1 3} R _ {2 1} R _ {2 2} ^ {2} + d _ {1 4} R _ {2 2} ^ {3}\\right)\\right) \\cdot \\tag {39} \\\\ \\left(d _ {1 0} R _ {2 1} + d _ {1 1} R _ {2 2}\\right) + \\left(\\left(d _ {4} - d _ {7}\\right) R _ {2 1} ^ {2} + \\right. \\\\ \\left. \\left(d _ {5} - d _ {8}\\right) R _ {2 1} R _ {2 2} + \\left(d _ {6} - d _ {9}\\right) R _ {2 2} ^ {2}\\right) \\Bigg). \\\\ \\left(d _ {1 0} R _ {2 1} + d _ {1 1} R _ {2 2}\\right) = 0. \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.653, + 0.469, + 0.669 + ], + "angle": 0, + "content": "Upon expanding equation (39), we find that it takes the form" + }, + { + "type": "equation", + "bbox": [ + 0.077, + 0.68, + 0.469, + 0.711 + ], + "angle": 0, + "content": "\\[\n\\alpha_ {1} R _ {2 1} ^ {4} + \\alpha_ {2} R _ {2 1} ^ {3} R _ {2 2} + \\alpha_ {3} R _ {2 1} ^ {2} R _ {2 2} ^ {2} + \\alpha_ {4} R _ {2 1} R _ {2 2} ^ {3} + \\alpha_ {5} 2 R _ {2 2} ^ {4} = 0 \\tag {40}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.712, + 0.469, + 0.745 + ], + "angle": 0, + "content": "Similarly to (16), we divide (40) by \\( R_{22}^{4} \\) and define a new variable \\( v = \\frac{R_{21}}{R_{22}} \\), and thereby deduce the univariate quartic" + }, + { + "type": "equation", + "bbox": [ + 0.141, + 0.762, + 0.469, + 0.78 + ], + "angle": 0, + "content": "\\[\n\\alpha_ {1} v ^ {4} + \\alpha_ {2} v ^ {3} + \\alpha_ {3} v ^ {2} + \\alpha_ {4} v + \\alpha_ {5} = 0. \\tag {41}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.793, + 0.468, + 0.824 + ], + "angle": 0, + "content": "After solving for \\( v \\), we can solve for \\( R_{21}, R_{22} \\) using (33), and for the other equations using (34), and (30)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.469, + 0.901 + ], + "angle": 0, + "content": "To summarize, we provide the outline of steps for solving the P1P2L absolute pose problem in Figure 5. Algebraically, the nontrivial steps in Figure 5 are second and third, which require solving a univariate quartic equation and computing square roots, respectively." + }, + { + "type": "table", + "bbox": [ + 0.502, + 0.088, + 0.903, + 0.249 + ], + "angle": 0, + "content": "
GenericMean RMed. RMean TMed. T
P1P2L no fix0.000501.0e-140.000651.8e-13
P1P2L fix9.0e-094.2e-153.4e-077.0e-14
P2P1L no fix7.1e-111.4e-151.2e-092.1e-14
P2P1L fix3.2e-092.6e-157.4e-084.3e-14
CoplanarMean RMed. RMean TMed. T
P1P2L no fix0.000229.6e-150.000301.75e-13
P1P2L fix1.10.791.20.96
P2P1L no fix2.43.142.23.14
P2P1L fix1.2e-124.0e-157.9e-116.3e-14
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.259, + 0.892, + 0.301 + ], + "angle": 0, + "content": "Table 2. The mean and median of the rot.(R) and tran.(T) errors over 1e5 noiseless generic (top) and coplanar (bottom) samples. In radians. The best results are marked bold." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.328, + 0.756, + 0.344 + ], + "angle": 0, + "content": "2.3. Stabilizing the P1P2L Solver" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.352, + 0.891, + 0.381 + ], + "angle": 0, + "content": "In this section, we outline a method to increase the stability of the P1P2L solver (Sec. 2.2)." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.382, + 0.892, + 0.472 + ], + "angle": 0, + "content": "Our proposed P1P2L solution faces a degeneracy when \\( Z_{1} - Z_{2} = 0 \\), since this leads to a division by zero in equation (29). Furthermore, our observations indicate that the result of the solver is unstable if the value of \\( Z_{1} - Z_{2} \\) is close to zero. Note, that this instability is specific to our solver and not inherent to the P1P2L problem." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.473, + 0.892, + 0.548 + ], + "angle": 0, + "content": "This degeneracy may be interpreted geometrically as follows: the values of \\( Z_{1} \\) and \\( Z_{2} \\) are the last coordinates of the points \\( L_{1}, L_{2} \\), which span the first 3D line. Therefore, the vector \\( L_{1} - L_{2} \\) represents the direction of the line, and \\( Z_{1} - Z_{2} \\) represents the last coordinate of this direction." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.549, + 0.892, + 0.623 + ], + "angle": 0, + "content": "Since the special reference frame used in Section 2.2 is independent of the rotation of the world frame, we can remove the source of instability by rotating the world frame such that the first line \\(\\overline{L_1L_2}\\) aligns with the z-axis. See Figure 6 for an illustration." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.635, + 0.749, + 0.651 + ], + "angle": 0, + "content": "2.4. Resolving the coplanar case" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.658, + 0.892, + 0.703 + ], + "angle": 0, + "content": "In this section, we discuss the performance of the P2P1L (Sec. 2.1) and P1P2L (Sec. 2.2) solvers in the case, where all 3D points and 3D lines are coplanar." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.704, + 0.892, + 0.793 + ], + "angle": 0, + "content": "P2P1L. The P2P1L solver presented in Sec. 2.1 (P2P1L no fix) is degenerate in coplanar situations, since then \\( Z_{4} = 0 \\) (2), causing a division by zero issue in equation (11). To resolve this issue, we propose a modified solver (P2P1L fix), which handles the coplanar case. See SM A for a description of this modified solver." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.795, + 0.892, + 0.855 + ], + "angle": 0, + "content": "P1P2L. The original version of the P1P2L solver (P1P2L no fix, Sec. 2.2) is able to handle the coplanar case. However, the stabilized P1P2L solver (P1P2L fix, Sec. 2.3) fails due to a degeneracy for coplanar input." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.856, + 0.892, + 0.901 + ], + "angle": 0, + "content": "We conducted a synthetic experiment to evaluate the impact of the proposed stabilization scheme. In this experiment, we used the same setting as in Section 3.1. The re" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "21321" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.129, + 0.09, + 0.435, + 0.143 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.083, + 0.144, + 0.462, + 0.294 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.299, + 0.468, + 0.327 + ], + "angle": 0, + "content": "Figure 7. Stability test. Histogram of rotation \\((l.)\\), and translation \\((r.)\\) errors computed over \\(1e5\\) noiseless samples." + }, + { + "type": "table", + "bbox": [ + 0.088, + 0.343, + 0.46, + 0.455 + ], + "angle": 0, + "content": "
MethodMeanMed.MinMax
P2P1L Ours313.8324.9230.83061.0
P2P1L 3Q31860.61909.91439.110102.1
P2P1L R.+S8897.59491.25805.149984.1
P2P1L R.+L4720.85239.52763.315362.6
P1P2L Ours504.0521.0364.04554.4
P1P2L 3Q31967.12008.21483.912931.1
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.465, + 0.468, + 0.493 + ], + "angle": 0, + "content": "Table 3. Runtime The mean, median and max values of the runtime. In nanoseconds. Best values for each problem are bold." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.523, + 0.468, + 0.629 + ], + "angle": 0, + "content": "sults, presented in Table 2 demonstrate that solvers \\( P2P1L \\) no fix and \\( P1P2L \\) fix achieve superior numerical stability in the generic case, while solvers \\( P2P1L \\) fix and \\( P1P2L \\) no fix handle the coplanar case. Since it is simple to detect the coplanar case, we recommend to use solvers \\( P2P1L \\) fix and \\( P1P2L \\) no fix in the coplanar case and solvers \\( P2P1L \\) no fix and \\( P1P2L \\) fix in the generic case." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.646, + 0.209, + 0.663 + ], + "angle": 0, + "content": "3. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.673, + 0.468, + 0.838 + ], + "angle": 0, + "content": "In this section, we experimentally compare the proposed solvers with the state-of-the-art methods, specifically 3Q3 [17] and Ramalingam [25]. For 3Q3, we employ the publicly available implementation from Poselib. As the implementation of Ramalingam is not publicly accessible, we have created our own implementation. Since it is not explicitly specified which method should be used for solving the linear equations, we compare with two variants: SVD and LU decomposition. All the solvers have been implemented in \\(\\mathrm{C + + }\\), and the experiments are conducted on a desktop computer with an AMD Ryzen 9 CPU with \\(3.9\\mathrm{GHz}\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.468, + 0.9 + ], + "angle": 0, + "content": "In Sec. 3.1, we provide an analysis of the numerical stability and runtime performance of the solvers on synthetic data. Subsequently, in Section 3.2, we present an evaluation of the solvers within the RANSAC scheme." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.091, + 0.707, + 0.107 + ], + "angle": 0, + "content": "3.1. Synthetic experiments" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.115, + 0.892, + 0.159 + ], + "angle": 0, + "content": "In this section, we present an analysis of the numerical stability and runtime performance of the minimal solvers on synthetic data." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.16, + 0.892, + 0.191 + ], + "angle": 0, + "content": "We generate instances of each minimal problem (either P2P1L or P1P2L), according to the following procedure:" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.194, + 0.892, + 0.27 + ], + "angle": 0, + "content": "- Rotation Matrix \\((R)\\): An axis \\(v \\in \\mathbb{R}^3\\) is sampled from the uniform distribution on the unit sphere, and an angle \\(\\alpha\\) is sampled from the normal distribution \\(\\mathcal{N}(0,1)\\). The rotation matrix \\(R\\) is then constructed using the angle-axis formula, \\(R = I + \\sin(\\alpha)[v]_{\\times} + (1 - \\cos(\\alpha))[v]_{\\times}^2\\)." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.27, + 0.892, + 0.314 + ], + "angle": 0, + "content": "- Translation \\((T)\\): The camera center \\(C\\) is sampled uniformly at random from the unit sphere. The translation vector \\(T\\) is computed as \\(T = -RC\\)." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.315, + 0.892, + 0.39 + ], + "angle": 0, + "content": "- Point correspondence: A 3D point \\( X_{i} \\) is sampled from the trivariate normal distribution with mean vector \\( \\mu = [0,0,5]^T \\) and standard deviation \\( \\sigma = 1 \\) in each component. The corresponding 2D point \\( D_{i} \\) is obtained by projecting \\( X_{i} \\) onto a pinhole camera with the pose \\( (R,T) \\)." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.391, + 0.892, + 0.466 + ], + "angle": 0, + "content": "- Line correspondence: Two 3D points, \\( L_{i} \\) and \\( L_{i+1} \\), are sampled as described in the previous bullet-point. These points define the 3D line \\( L \\). Two 2D points, \\( D_{i} \\) and \\( D_{i+1} \\), are obtained by sampling two points on the 3D line \\( L \\) and projecting them onto the camera with the pose \\( (R, T) \\)." + }, + { + "type": "list", + "bbox": [ + 0.5, + 0.194, + 0.892, + 0.466 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.469, + 0.892, + 0.668 + ], + "angle": 0, + "content": "In the case of P2P1L, two points and one line are sampled, while in the case of P1P2L, one point and two lines are sampled. Let \\((R_{est},T_{est})\\) be the pose obtained by the minimal solver. We measure the rotation error as the angle \\(\\xi_R = \\arccos ((1 + \\mathrm{trace}(R_{est}^T R)) / 2)\\). We calculate the translation error \\(\\xi_T\\) as \\(\\xi_T = \\frac{|T_{est} - T|}{|T|}\\). The histograms of rotation errors \\((\\xi_R)\\) and translation errors \\((\\xi_T)\\) for all considered solvers are depicted in Figure 7. Furthermore, summary statistics including the median, mean, and maximum errors are provided in Table 4. The results indicate that all minimal solvers are stable. Our algebraically optimal solvers demonstrate superior stability compared to the previous solvers." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.67, + 0.892, + 0.776 + ], + "angle": 0, + "content": "The runtime evaluation for all considered solvers is shown in Table. 3. Our P2P1L solver requires 313.8ns on average, which is about 6x faster than the 3Q3 solver [17]. Similarly, our P1P2L solver requires on average 504.0ns, which is about 4x faster than the 3Q3 solver. These results show a significant speedup achieved by our solvers in comparison to the previous methods." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.787, + 0.733, + 0.802 + ], + "angle": 0, + "content": "3.2. Experiments in RANSAC" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.81, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Here, we provide an analysis of the minimal solvers within the RANSAC scheme [4], using the Oxford Multi-view Data [1]. This dataset includes a variety of indoor and outdoor scenes, some of which contain both point and line matches. An example of images from this dataset is illustrated in Figure 8." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "21322" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.156, + 0.088, + 0.812, + 0.2 + ], + "angle": 0, + "content": "
MethodMean RMedian RMax RMean TMedian TMax T
P2P1L Ours5.3e-121.4e-151.2e-073.7e-102.1e-142.2e-05
P2P1L 3Q32.8e-056.5e-152.72.0e-051.2e-131.04
P2P1L Ramalingam (SVD)4.9e-091.5e-140.000411.6e-072.7e-130.0099
P2P1L Ramalingam (LU)4.7e-071.3e-140.0402.3e-052.4e-130.10
PP1P2L Ours (fix 2)1.2e-074.4e-150.0102.0e-067.1e-140.13
P1P2L 3Q33.3e-057.2e-152.603.4e-051.5e-131.01
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.202, + 0.892, + 0.217 + ], + "angle": 0, + "content": "Table 4. The mean, median and max values of the rot.(R) and tran.(T) errors over \\( {1e5} \\) noiseless samples. The best results are marked bold." + }, + { + "type": "table", + "bbox": [ + 0.12, + 0.23, + 0.849, + 0.398 + ], + "angle": 0, + "content": "
DatasetP2P1LP1P2L
OUR3Q3Ram. SVDRam. LUOUR3Q3
Model House7.72 (0.85x)9.09 (1x)20.69 (2.28x)17.30 (1.90x)8.06 (0.85x)9.52 (1x)
Corridor11.32 (0.91x)12.46 (1x)26.56 (2.13x)23.54 (1.89x)13.3 (0.90x)14.76 (1x)
Merton I35.93 (0.98x)36.83 (1x)77.96 (2.12x)74.87 (2.03x)28.85 (0.64x)45.41 (1x)
Merton II33.30 (0.97x)34.27 (1x)74.37 (2.17x)71.35 (2.08x)26.7 (0.67x)39.64 (1x)
Merton III24.04 (0.96x)25.07 (1x)52.07 (2.08x)48.97 (1.95x)10.91 (0.37x)29.67 (1x)
Library32.42 (0.97x)33.31 (1x)69.29 (2.08x)66.23 (1.99x)10.04 (0.28x)35.98 (1x)
Wadham39.60 (0.98x)40.51 (1x)86.10 (2.13x)83.00 (2.05x)22.96 (0.45x)50.79 (1x)
Avg. Speed-up0.94x1x2.14x1.99x0.59x1x
" + }, + { + "type": "table_caption", + "bbox": [ + 0.116, + 0.404, + 0.854, + 0.418 + ], + "angle": 0, + "content": "Table 5. RANSAC timing, on Oxford Multi-view dataset [1], in milliseconds. Speed-up compared to the 3Q3 in the bracket." + }, + { + "type": "image", + "bbox": [ + 0.119, + 0.442, + 0.264, + 0.528 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.279, + 0.442, + 0.425, + 0.527 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.119, + 0.529, + 0.265, + 0.614 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.28, + 0.528, + 0.425, + 0.614 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.618, + 0.47, + 0.647 + ], + "angle": 0, + "content": "Figure 8. Example images of University Library (top) and Merton College III (bottom), from [1]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.674, + 0.47, + 0.809 + ], + "angle": 0, + "content": "We utilize the Poselib [18] implementation of LORANSAC [6]. In this implementation, local optimization is applied both when the new solution surpasses the currently best one and at the end of the entire procedure. The RANSAC parameters used include a maximum number of iterations set to 100000, a minimum number of iterations set to 1000, a success probability of 0.9999, and an inlier threshold of 1 pixel. To ensure a fair comparison between the solvers, we use a fixed random seed." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.811, + 0.471, + 0.9 + ], + "angle": 0, + "content": "Since all the solvers are stable, we do not expect any significant difference in the estimated pose when using different solvers. Therefore, our primary focus is on comparing the runtime of the RANSAC scheme. The runtime comparison is presented in Table 5, which displays the average runtime for each scene in the Oxford multiview dataset and" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.445, + 0.892, + 0.641 + ], + "angle": 0, + "content": "for each considered solver. Additionally, the table includes a speedup value compared to the 3Q3 solver of [18]. The results show that our solvers consistently outperform the 3Q3 [17] and Ramalingam [25] solvers in terms of runtime. However, there is a notable variation in the speed-up among the scenes, which ranges from 0.85 to 0.98 for the P2P1L problem and from 0.28 to 0.90 for the P1P2L solver. This variation can be attributed to the fact that different scenes require varying proportions of time spent by RANSAC on scoring hypothesized models and local optimization. The more time is dedicated to these tasks, the less significant speedup can be reached. In cases with fewer matches or a lower inlier ratio, more substantial speedups are possible." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.642, + 0.892, + 0.703 + ], + "angle": 0, + "content": "The final pose estimation errors in RANSAC are shown in the Supplementary. As expected, the errors are the same for all solvers. This demonstrates that we can achieve improved runtimes without sacrificing any accuracy." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.713, + 0.619, + 0.728 + ], + "angle": 0, + "content": "4. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.735, + 0.892, + 0.87 + ], + "angle": 0, + "content": "In revisiting absolute pose from mixed point-line correspondences, we have developed new solvers, which have the merit of being both algebraically optimal and elementary in nature. Moreover, our solvers outperform the state-of-the-art in terms of both runtime and numerical stability. Despite the simplicity of many minimal absolute pose problems, our findings suggest that their continued investigation remains worthwhile. The code is available at https://github.com/petrhruby97/efficientAbsolutely" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.892, + 0.9 + ], + "angle": 0, + "content": "Acknowledgements: TD was supported by NSF DMS 2103310." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "21323" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.105 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.468, + 0.156 + ], + "angle": 0, + "content": "[1] Multi-view data, oxford visual geometry group. https://www.robots.ox.ac.uk/vgg/data/mview/. Accessed: Nov 8 2023. 7, 8, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.159, + 0.468, + 0.226 + ], + "angle": 0, + "content": "[2] Sérgio Agostinho, João Gomes, and Alessio Del Bue. CvxPnPL: A unified convex solution to the absolute pose estimation problem from point and line correspondences. Journal of Mathematical Imaging and Vision, 65(3):492-512, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.23, + 0.468, + 0.271 + ], + "angle": 0, + "content": "[3] Daniel Barath and Jiri Matas. Graph-cut RANSAC: local optimization on spatially coherent structures. IEEE Trans. Pattern Anal. Mach. Intell., 44(9):4961-4974, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.274, + 0.468, + 0.355 + ], + "angle": 0, + "content": "[4] Robert C. Bolles and Martin A. Fischler. A ransac-based approach to model fitting and its application to finding cylinders in range data. In Proceedings of the 7th International Joint Conference on Artificial Intelligence, IJCAI '81, Vancouver, BC, Canada, August 24-28, 1981, pages 637-643. William Kaufmann, 1981. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.359, + 0.468, + 0.413 + ], + "angle": 0, + "content": "[5] Homer H. Chen. Pose determination from line-to-plane correspondences: Existence condition and closed-form solutions. IEEE Trans. Pattern Anal. Mach. Intell., 13(6):530-541, 1991. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.416, + 0.468, + 0.471 + ], + "angle": 0, + "content": "[6] Ondrej Chum, Jiri Matas, and Josef Kittler. Locally optimized RANSAC. In Pattern Recognition, 25th DAGM Symposium, Magdeburg, Germany, September 10-12, 2003, Proceedings, pages 236-243. Springer, 2003. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.474, + 0.468, + 0.528 + ], + "angle": 0, + "content": "[7] Michel Dhome, Marc Richetin, Jean-Thierry Lapresté, and Gérard Rives. Determination of the attitude of 3d objects from a single perspective view. IEEE Trans. Pattern Anal. Mach. Intell., 11(12):1265-1278, 1989. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.531, + 0.468, + 0.599 + ], + "angle": 0, + "content": "[8] Yaqing Ding, Jian Yang, Viktor Larsson, Carl Olsson, and Kalle Åström. Revisiting the P3P problem. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2023, Vancouver, BC, Canada, June 17-24, 2023, pages 4872-4880. IEEE, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.602, + 0.468, + 0.657 + ], + "angle": 0, + "content": "[9] Timothy Duff, Viktor Korotynskiy, Tomas Pajdla, and Margaret H. Regan. Galois/monodromy groups for decomposing minimal problems in 3d reconstruction. SIAM Journal on Applied Algebra and Geometry, 6(4):740-772, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.66, + 0.468, + 0.714 + ], + "angle": 0, + "content": "[10] Ricardo Fabbri, Peter J. Giblin, and Benjamin B. Kimia. Camera pose estimation using first-order curve differential geometry. IEEE Trans. Pattern Anal. Mach. Intell., 43(10): 3321-3332, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.717, + 0.468, + 0.771 + ], + "angle": 0, + "content": "[11] Martin A. Fischler and Robert C. Bolles. Random sample consensus: A paradigm for model fitting with applications to image analysis and automated cartography. Commun. ACM, 24(6):381-395, 1981. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.774, + 0.468, + 0.843 + ], + "angle": 0, + "content": "[12] Christian Hane, Lionel Heng, Gim Hee Lee, Friedrich Fraundorfer, Paul Furgale, Torsten Sattler, and Marc Pollefeys. 3d visual perception for self-driving cars using a multi-camera system: Calibration, mapping, localization, and obstacle detection. Image Vis. Comput., 68:14-27, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.846, + 0.468, + 0.9 + ], + "angle": 0, + "content": "[13] Robert M. Haralick, Chung-Nan Lee, Karsten Ottenberg, and Michael Nölle. Review and analysis of solutions of the three point perspective pose estimation problem. Int. J. Comput. Vis., 13(3):331-356, 1994. 1" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.468, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.175 + ], + "angle": 0, + "content": "[14] Petr Hruby, Viktor Korotynskiy, Timothy Duff, Luke Oeding, Marc Pollefeys, Tomás Pajdla, and Viktor Larsson. Four-view geometry with unknown radial distortion. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2023, Vancouver, BC, Canada, June 17-24, 2023, pages 8990-9000. IEEE, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.177, + 0.892, + 0.23 + ], + "angle": 0, + "content": "[15] Xiaomei Hu, Luying Zhu, Ping Wang, Haili Yang, and Xuan Li. Improved ORB-SLAM2 mobile robot vision algorithm based on multiple feature fusion. IEEE Access, 11:100659-100671, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.233, + 0.892, + 0.328 + ], + "angle": 0, + "content": "[16] Laurent Kneip, Davide Scaramuzza, and Roland Siegwart. A novel parametrization of the perspective-three-point problem for a direct computation of absolute camera position and orientation. In The 24th IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2011, Colorado Springs, CO, USA, 20-25 June 2011, pages 2969-2976. IEEE Computer Society, 2011. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.33, + 0.892, + 0.412 + ], + "angle": 0, + "content": "[17] Zuzana Kukelova, Jan Heller, and Andrew W. Fitzgibbon. Efficient intersection of three quadrics and applications in computer vision. In 2016 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2016, Las Vegas, NV, USA, June 27-30, 2016, pages 1799-1808. IEEE Computer Society, 2016. 2, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.414, + 0.891, + 0.439 + ], + "angle": 0, + "content": "[18] Viktor Larsson and contributors. PoseLib - Minimal Solvers for Camera Pose Estimation, 2020. 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.442, + 0.892, + 0.509 + ], + "angle": 0, + "content": "[19] Shaohui Liu, Yifan Yu, Rémi Pautrat, Marc Pollefeys, and Viktor Larsson. 3d line mapping revisited. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2023, Vancouver, BC, Canada, June 17-24, 2023, pages 21445-21455. IEEE, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.511, + 0.892, + 0.593 + ], + "angle": 0, + "content": "[20] David Nister, Richard I. Hartley, and Henrik Stewenius. Using galois theory to prove structure from motion algorithms are optimal. In 2007 IEEE Computer Society Conference on Computer Vision and Pattern Recognition (CVPR 2007), 18-23 June 2007, Minneapolis, Minnesota, USA. IEEE Computer Society, 2007. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.595, + 0.892, + 0.677 + ], + "angle": 0, + "content": "[21] Rémi Pautrat, Daniel Barath, Viktor Larsson, Martin R. Oswald, and Marc Pollefeys. Deeplsd: Line segment detection and refinement with deep image gradients. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2023, Vancouver, BC, Canada, June 17-24, 2023, pages 17327-17336. IEEE, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.679, + 0.892, + 0.746 + ], + "angle": 0, + "content": "[22] Rémi Pautrat, Iago Suárez, Yifan Yu, Marc Pollefeys, and Viktor Larsson. Gluestick: Robust image matching by sticking points and lines together. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9706-9716, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.748, + 0.892, + 0.816 + ], + "angle": 0, + "content": "[23] Mikael Persson and Klas Nordberg. Lambda twist: An accurate fast robust perspective three point (P3P) solver. In Computer Vision - ECCV 2018 - 15th European Conference, Munich, Germany, September 8-14, 2018, Proceedings, Part IV, pages 334-349. Springer, 2018. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.818, + 0.892, + 0.871 + ], + "angle": 0, + "content": "[24] Rahul Raguram, Ondrej Chum, Marc Pollefeys, Jiri Matas, and Jan-Michael Frahm. USAC: A universal framework for random sample consensus. IEEE Trans. Pattern Anal. Mach. Intell., 35(8):2022-2038, 2013. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.874, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[25] Srikumar Ramalingam, Sofien Bouaziz, and Peter F. Sturm. Pose estimation using both points and lines for geo-" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "21324" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.092, + 0.47, + 0.134 + ], + "angle": 0, + "content": "localization. In IEEE International Conference on Robotics and Automation, ICRA 2011, Shanghai, China, 9-13 May 2011, pages 4716-4723. IEEE, 2011. 1, 2, 3, 4, 5, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.136, + 0.47, + 0.19 + ], + "angle": 0, + "content": "[26] Torsten Sattler, Bastian Leibe, and Leif Kobbelt. Efficient & effective prioritized matching for large-scale image-based localization. IEEE Trans. Pattern Anal. Mach. Intell., 39(9): 1744-1756, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.192, + 0.47, + 0.262 + ], + "angle": 0, + "content": "[27] Johannes L. Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In 2016 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2016, Las Vegas, NV, USA, June 27-30, 2016, pages 4104-4113. IEEE Computer Society, 2016. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.263, + 0.47, + 0.318 + ], + "angle": 0, + "content": "[28] Jonathan Ventura, Clemens Arth, Gerhard Reitmayr, and Dieter Schmalstieg. Global localization from monocular SLAM on a mobile phone. IEEE Trans. Vis. Comput. Graph., 20(4):531-539, 2014. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.32, + 0.47, + 0.374 + ], + "angle": 0, + "content": "[29] Chi Xu, Lilian Zhang, Li Cheng, and Reinhard Koch. Pose estimation from line correspondences: A complete analysis and a series of solutions. IEEE Trans. Pattern Anal. Mach. Intell., 39(6):1209-1222, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.376, + 0.47, + 0.46 + ], + "angle": 0, + "content": "[30] Lipu Zhou, Jiamin Ye, and Michael Kaess. A stable algebraic camera pose estimation for minimal configurations of 2d/3d point and line correspondences. In Computer Vision - ACCV 2018 - 14th Asian Conference on Computer Vision, Perth, Australia, December 2-6, 2018, Revised Selected Papers, Part IV, pages 273-288. Springer, 2018. 1, 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.46 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "21325" + } + ] +] \ No newline at end of file diff --git a/2024/Efficient Solution of Point-Line Absolute Pose/8561d28f-e735-46cc-9618-c6c206bdafc2_origin.pdf b/2024/Efficient Solution of Point-Line Absolute Pose/8561d28f-e735-46cc-9618-c6c206bdafc2_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..d768eeebd2f1a99924a4a9f401aa2f8f716dd30c --- /dev/null +++ b/2024/Efficient Solution of Point-Line Absolute Pose/8561d28f-e735-46cc-9618-c6c206bdafc2_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:34f3e1332419e700c893e7686286cb3b848e233b5688a4c61266bce22b27feb2 +size 2334882 diff --git a/2024/Efficient Solution of Point-Line Absolute Pose/full.md b/2024/Efficient Solution of Point-Line Absolute Pose/full.md new file mode 100644 index 0000000000000000000000000000000000000000..d3220b44e5564ecb98ea0fe8db40866b849c27c0 --- /dev/null +++ b/2024/Efficient Solution of Point-Line Absolute Pose/full.md @@ -0,0 +1,610 @@ +# Efficient Solution of Point-Line Absolute Pose + +Petr Hruby ETH Zürich + +Timothy Duff +University of Washington + +Marc Pollefeys +ETH Zürich, Microsoft + +# Abstract + +We revisit certain problems of pose estimation based on 3D-2D correspondences between features which may be points or lines. Specifically, we address the two previously-studied minimal problems of estimating camera extrinsics from $p \in \{1,2\}$ point-point correspondences and $l = 3 - p$ line-line correspondences. To the best of our knowledge, all of the previously-known practical solutions to these problems required computing the roots of degree $\geq 4$ (univariate) polynomials when $p = 2$ , or degree $\geq 8$ polynomials when $p = 1$ . We describe and implement two elementary solutions which reduce the degrees of the needed polynomials from 4 to 2 and from 8 to 4, respectively. We show experimentally that the resulting solvers are numerically stable and fast: when compared to the previous state-of-the-art, we may obtain nearly an order of magnitude speedup. The code is available at https://github.com/petrhruby97/efficient_absolute + +# 1. Introduction + +# 1.1. Motivation + +The problem of registering images to a known 3D coordinate system plays a crucial role in applications such as visual localization [26], autonomous driving [12], and augmented reality [28], as well as in general paradigms like SLAM [15] and SfM [27]. Robust estimators based on RANSAC [11], or one of its many refinements [3, 24], are among the most successful tools for solving these problems. Such an estimator traditionally relies on a minimal P3P solver [8, 13, 16, 23] to efficiently hypothesize poses from putative matches between 3D and 2D points. + +The literature on P3P and other purely point-based methods for absolute pose estimation is vast. Absolute pose estimation from non-point features such as lines [2, 5, 7, 25, 29, 30], point-line incidences [10], and affine correspondences [28], has received comparatively less attention, but remains an active research area. In particular, solutions relying on both points and lines are of increasing importance, due to the prevalence of both types of feature in man-made environments, as well as several recent advances in the com + +![](images/92879d98ea4ada4118f256e76bcc5e1846b9d6f5e2287fa51eb66b458b2b3d8e.jpg) +Figure 1. The P2P1L (Top) and P1P2L (Bottom) problems. + +ponents of 3D reconstruction systems responsible for line detection [21], matching [22], and bundle adjustment [19]. + +# 1.2. Contribution + +In this paper, we revisit the two minimal problems of absolute pose estimation that combine both points and lines: the Perspective-2-Point-1-Line Problem (or P2P1L, Section 2.1), and the Perspective-1-Point-2-Line Problem (P1P2L, Section 2.2.) See Figure 1 for illustrations. + +In contrast to the purely point or line based minimal problems P3P and P3L, we observe that the existing solutions for the "mixed cases" considered here are suboptimal from both a theoretical and practical point of view. Consequently, we develop novel solutions to both problems, which optimally exploit their underlying algebraic structure, and exhibit comparable or better performance than the state of the art on simulated and real data—see Section 3. On the other hand, these solvers are simple to implement, and no knowledge of mathematics beyond elementary algebra is needed to understand them. + +To provide further context for our work, we include the following quotation from [25]: Although we do not theoretically prove that our solutions are of the lowest possible degrees, we believe so because of the following argument. The best existing solutions for pose estimation using three + +points and three lines use 4th and 8th degree solutions respectively. Since mixed cases are in the middle, our solutions for (2 points, 1 line) and (1 point, 2 lines) cases use 4th and 8th degree solutions respectively. Recently, it was shown using Galois theory that the solutions that use the lowest possible degrees are the optimal ones [[20]]. + +Contrary to the informal reasoning presented above, we claim that the existing solutions to the mixed point-line cases are not optimal. As far as we know, the only support for this claim appearing in the literature prior to our work occurs in [9, §3]. This previous work showed, on the basis of Galois group computation, that the P2P1L and P1P2L problems decompose into simpler subproblems. However, this theoretical observation was not accompanied by a practical solution method for either problem. In this paper, we rectify the situation by devising practical solvers for both problems that incorporate these recent insights. + +To complete our discussion of what makes a solution algebraically optimal, we recall one proposal of such a notion from work of Nister et al. [20]. In this work, a restricted class of algorithms $\mathcal{P}_n$ is considered for each natural number $n \geq 1$ : an algorithm in $\mathcal{P}_n$ consists of a finite sequence of steps, each of which extracts the roots of some polynomial equation $p(x) = 0$ , with either $p(x) = x^m - a$ (where $a \in \mathbb{Q}$ and $\deg(p) = m$ can be arbitrary), or with $\deg(p) \leq n$ and the coefficients of $p$ belonging to a field containing $\mathbb{Q}$ and any previously-computed roots. In this setting, a solution is optimal if it leads to solving a polynomial system of the lowest possible degree. Our proposed solutions immediately establish that P2P1L is $\mathcal{P}_2$ -solvable and P1P2L is $\mathcal{P}_4$ -solvable. It should be of little surprise that P2P1L is not $\mathcal{P}_1$ -solvable; therefore, we may say that our solution to P2P1L is algebraically optimal. On the other hand, the problem P1P2L is also $\mathcal{P}_3$ -solvable. This is because any quartic equation can be solved by the standard method which reduces the problem to computing the roots of the associated resolvent cubic. Thus, our solution, which computes the roots of a quartic with this same method, is also algebraically optimal in the sense of [20]. The same observation, of course, holds for the classical quartic-based methods for solving P3P. Alternative P3P solvers that directly employ a cubic [8, 23], despite being superior in practical terms, are not distinguished by the complexity classes $\mathcal{P}_n$ . Table 1 provides a summary of the algebraic complexities, based on the Galois groups computed in [9, §3]. + +
problemP3PP2P1LP1P2LP3L
classP3P2P3P8
+ +Table 1. For each minimal absolute pose problem with points and line features, the class of polynomial root finding algorithms $\mathcal{P}_n$ that solve the problem with $n$ as small as possible [20, §2.1]. + +# 1.3. Related work + +The first solutions to P2P1L and P1P2L were presented by Ramalingam et al. [25], reducing the problems to computing the roots of polynomials of degree 4 and 8 via careful choices of special reference frames in the world and image. Although we obtain polynomials of lower degree, these special reference frames are also an ingredient in our approach. + +In work subsequent to [25], it was observed that both of these problems could be solved using the E3Q3 solver [17]. This is a highly optimized method for computing the points where three quadric surfaces in $\mathbb{R}^3$ intersect. Typically, there are $2^{3} = 8$ such points, by Bézout's theorem; however, EQ3Q also handles degenerate cases where the number of solutions may drop. In the follow-up work [30], a stabilization scheme for E3Q3 was applied to these problems, and experimentally shown to be more accurate than the solvers in [25]. Efficient solvers for both problems based on this stabilized E3Q3 are implemented in PoseLib [18]. + +The idea of using Galois groups to study minimal problems originates from [20]. The main takeaway from this paper is that a problem with $n$ solutions whose Galois group is the full symmetric group $S_{n}$ cannot be solved by an algorithm in $\mathcal{P}_{n-1}$ . In the other direction, the main takeaway from [9] is that if the Galois group is contained in the wreath product $S_{n_{1}} \wr S_{n_{2}}$ , where $n = n_{1}n_{2}$ , then the problem can be solved with an algorithm in $\mathcal{P}_{\max(n_{1},n_{2})}$ . A recent work using such an insight to guide the more efficient solution of a relative pose estimation problem may be found in [14]. + +# 2. Minimal solvers + +In this section, we introduce our algebraically optimal solutions to the P2P1L and P1P2L problems. These problems are depicted in Figure 1. As revealed by the Galois group computed in [9, §3], the problem P2P1L can be reduced to computing the roots of a quadratic equation, and the problem P1P2L can be reduced to a quartic equation. In Sections 2.1 and 2.2, we turn these insights into explicit solutions to the P2P1L and P12PL problems, respectively. These solutions work generically—they are valid outside of a measure-zero subset of the space of input point-point/line-line correspondences. In Section 2.3, we describe a method that stabilizes the P1P2L solver of Section 2.2 in a common but non-generic case. Finally, Section 2.4 for a discussion when the point-line configuration is coplanar. + +# 2.1. P2P1L + +Here, we provide the algebraically optimal solution to the P2P1L problem. We mostly follow the notation of [25]. + +Problem Statement: Consider $P_1^0, P_2^0 \in \mathbb{R}^3$ , two 3D points, and their 2D projections under an unknown calibrated camera. Consider also a 3D line spanned by points $L_3^0, L_4^0 \in \mathbb{R}^3$ and its projection. The goal is to recover the + +![](images/20445638a11240eeb032dbf9aa20b68a3f5e0ff62a2dfbc3eea5f7ee8d5f1d57.jpg) + +![](images/43c16aa4a49f8858a75473230dd5eaa7300de3b17ad8be6e8c7c274b0f96d3ad.jpg) + +![](images/7ab54c24980c7d8f3dfccb159e7c7f38c5184844ac9305c9b364214ab8856d71.jpg) +Figure 2. Special reference frame for P2P1L. Camera frame $\mathcal{C}_0$ is transformed to $\mathcal{C}_1$ , world frame $\mathcal{W}_0$ to $\mathcal{W}_1$ . See text for details. + +![](images/e709c755edc91e6f86341c74511e1079caf27c378108579d7dc7e3a31ed5ea1e.jpg) + +unknown camera matrix, whose center we denote by $C^0$ + +Our first step is to rigidly transform the input into the special reference frame introduced in [25]. Transforming to this special frame simplifies the equations and helps to reveal the algebraically optimal solution. Figure 2 illustrates the input reference frame $\mathcal{C}_0 / \mathcal{W}_0$ (top), and the special frame $\mathcal{C}_1 / \mathcal{W}_1$ (bottom), which are related as follows: + +- In the world frame $\mathcal{W}_1$ , the 3D points take the form + +$$ +P _ {1} = \left( \begin{array}{l l l} 0 & 0 & 0 \end{array} \right) ^ {T}, \quad P _ {2} = \left( \begin{array}{l l l} X _ {2} & 0 & 0 \end{array} \right) ^ {T}, \tag {1} +$$ + +and the 3D line is spanned by two points of the form + +$$ +L _ {3} = \left( \begin{array}{l l l} X _ {3} & Y _ {3} & 0 \end{array} \right) ^ {T}, \quad L _ {4} = \left( \begin{array}{l l l} X _ {4} & Y _ {4} & Z _ {4} \end{array} \right) ^ {T}. \tag {2} +$$ + +To find the transformation $\mathcal{W}_0\to \mathcal{W}_1$ , we set + +$$ +P _ {1} = \left( \begin{array}{c c c} 0 & 0 & 0 \end{array} \right) ^ {T}, \quad P _ {2} = \left(\| P _ {1} ^ {0} - P _ {2} ^ {2} \| \quad 0 \quad 0\right) ^ {T}, +$$ + +$$ +\begin{array}{l} L _ {3} = \left( \begin{array}{c c c} X _ {3} & Y _ {3} & 0 \end{array} \right) ^ {T}, \mathrm {w /} X _ {3} = \left(L _ {3} ^ {0} - P _ {1}\right) ^ {T} \frac {P _ {2} - P _ {1}}{\| P _ {2} - P _ {1} \|}, \\ Y _ {3} = \left\| L _ {3} ^ {0} - P _ {1} - X _ {3} \cdot \frac {P _ {2} - P _ {1}}{\| P _ {2} - P _ {1} \|} \right\|. \\ \end{array} +$$ + +We transform the rays $\overrightarrow{P_1^0P_2^0}$ and $\overrightarrow{P_1^0L_3^0}$ to $\overrightarrow{P_1P_2}$ and $\overrightarrow{P_1L_3}$ , respectively, via suitable translation and rotation. + +- In the camera frame, the camera center is fixed at + +$$ +C = \left( \begin{array}{l l l} 0 & 0 & - 1 \end{array} \right) ^ {T}, \tag {3} +$$ + +the 2D points are projections of points of the form + +$$ +D _ {1} = \left( \begin{array}{l l l} a _ {1} & b _ {1} & 0 \end{array} \right) ^ {T}, \quad D _ {2} = \left( \begin{array}{l l l} a _ {2} & b _ {2} & 0 \end{array} \right) ^ {T}, \tag {4} +$$ + +and the 2D line is the projection of $\overline{D_3D_4}$ , where + +$$ +D _ {3} = \left( \begin{array}{l l l} 0 & 0 & 0 \end{array} \right) ^ {T}, \quad D _ {4} = \left( \begin{array}{l l l} 1 & 0 & 0 \end{array} \right) ^ {T}. \tag {5} +$$ + +- To find the transformation $\mathcal{C}_0\to \mathcal{C}_1$ , let $D_3^0$ and $D_4^0$ be the homogeneous coordinates of two distinct points along the given line. Independently of $C^0$ , we may rigidly transform the rays $d_{3}\coloneqq \overrightarrow{C^{0}D_{3}^{0}}$ , and $d_4\coloneqq \overrightarrow{C^0D_3^0}$ to $\overrightarrow{CD_3}$ and $\overrightarrow{CD_4}$ , by suitable rotation and translation. + +Let us now write $R = (R_{ij})_{1 \leq i,j \leq 3}$ , $T = (T_i)_{1 \leq i \leq 3}$ for the unknown camera pose in this special reference frame. Our projection constraints can then be written as + +$$ +\left(D _ {i} - C\right) \times \left(R P _ {i} + T\right) = 0, \quad i \in \{1, 2 \}, \tag {6} +$$ + +$$ +\left(D _ {3} \times D _ {4}\right) ^ {T} \left(R L _ {i} + T\right) = 0, \quad i \in \{3, 4 \}. \tag {7} +$$ + +This gives a system of equations in 12 unknowns. However, in what follows, we consider a system of equations in a smaller set of unknowns, namely + +$$ +\mathbf {x} ^ {T} = \left( \begin{array}{l l l l l l l l} R _ {1 1} & R _ {2 1} & R _ {3 1} & R _ {2 2} & R _ {2 3} & T _ {1} & T _ {2} & T _ {3} \end{array} \right). \tag {8} +$$ + +The entries of $\mathbf{x}$ are constrained by + +$$ +R _ {1 1} ^ {2} + R _ {2 1} ^ {2} + R _ {3 1} ^ {2} = R _ {2 1} ^ {2} + R _ {2 2} ^ {2} + R _ {2 3} ^ {2} = 1. \tag {9} +$$ + +Thus, the entries of $\mathbf{x}$ specify a translation vector and a partially-filled rotation matrix, which can be uniquely completed to a rotation matrix using the formulae + +$$ +\binom {R _ {1 2}} {R _ {1 3}} = \left(R _ {2 2} ^ {2} + R _ {2 3} ^ {2}\right) ^ {- 1} \cdot \binom {- R _ {1 1} R _ {2 1} R _ {2 2} + R _ {2 3} R _ {3 1}} {- R _ {1 1} R _ {2 1} R _ {2 3} - R _ {2 2} R _ {3 1}} \\ - R _ {2 1} R _ {2 2} R _ {3 1} - R _ {1 1} R _ {2 3} \\ - R _ {2 1} R _ {2 3} R _ {3 1} + R _ {1 1} R _ {2 2}) \tag {10} +$$ + +subject to the genericity condition $R_{22}^2 + R_{23}^2 \neq 0$ , i.e. when $R$ is not a rotation in the $xz$ -plane. Hence, we focus on recovering the entries of $\mathbf{x}$ . From 2 out of the 3 redundant constraints from (6) together with (7), we obtain linear constraints on $\mathbf{x}$ ([25, eq. (4)-(5)]), + +$$ +A \mathbf {x} = \mathbf {b}, \quad \text {w h e r e} +$$ + +$$ +A = \left( \begin{array}{c c c c c c c c} 0 & 0 & 0 & 0 & 0 & - b _ {1} & a _ {1} & 0 \\ 0 & 0 & 0 & 0 & 0 & 0 & - 1 & b _ {1} \\ - b _ {2} X _ {2} & a _ {2} X _ {2} & 0 & 0 & 0 & - b _ {2} & a _ {2} & 0 \\ 0 & - X _ {2} & b _ {2} X _ {2} & 0 & 0 & 0 & - 1 & b _ {2} \\ 0 & X _ {3} & 0 & Y _ {3} & 0 & 0 & 1 & 0 \\ 0 & X _ {4} & 0 & Y _ {4} & Z _ {4} & 0 & 1 & 0 \end{array} \right), +$$ + +$$ +\mathbf {b} ^ {T} = \left( \begin{array}{c c c c c c} 0 & - b _ {1} & 0 & - b _ {2} & 0 & 0 \end{array} \right). \tag {11} +$$ + +We may use these linear equations to solve for the translation $T$ in terms of $R$ and the problem data as follows: + +$$ +T _ {1} = a _ {1} B, \quad T _ {2} = b _ {1} B, \quad T _ {3} = - 1 + B, \quad \text {w h e r e} +$$ + +$$ +B = \frac {X _ {2} \left(a _ {2} R _ {2 1} - b _ {2} R _ {1 1}\right)}{b _ {2} a _ {1} - b _ {1} a _ {2}}. \tag {12} +$$ + +(1) Transform data in world and camera frames to the special reference frames satisfying (1), (2), (3), (4), (5). +(2) Compute up to 2 solutions in $v$ to the quadratic (16). +(3) For each root $v$ in step (2), recover a corresponding value of $u$ using either of the linear expressions in (46). +(4) Recover up to 4 solutions in $\{R_{11}, R_{21}\}$ using (15). +(5) Recover $R$ and $T$ using (13), (12), (10). +(6) Reverse the transformation applied in step (1). + +Moreover, from (11) we may use $R_{11}$ and $R_{21}$ , to express the remaining entries of $R$ as + +$$ +R _ {3 1} = \frac {(a _ {1} - a _ {2}) R _ {2 1} + (b _ {2} - b _ {1}) R _ {1 1})}{b _ {2} a _ {1} - b _ {1} a _ {2}}, +$$ + +$$ +R _ {2 2} = \frac {b _ {1} b _ {2} X _ {2} R _ {1 1} + (b _ {1} a _ {2} X _ {2} - b _ {1} a _ {2} X _ {2} - b _ {2} a _ {1} X _ {3}) R _ {2 1}}{Y _ {3} (b _ {2} a _ {1} - b _ {1} a _ {2})}, +$$ + +$$ +\begin{array}{l} R _ {2 3} = \frac {b _ {1} b _ {2} X _ {2} \left(Y _ {3} - Y _ {4}\right) R _ {1 1}}{Y _ {3} Z _ {4} \left(b _ {2} a _ {1} - b _ {1} a _ {2}\right)} \\ + \left(\frac {X _ {3} Y _ {4} - X _ {4} Y _ {3}}{Y _ {3} Z _ {4}} + \frac {b _ {2} a _ {1} X _ {3} Y _ {4} - b _ {1} a _ {2} X _ {2} Y _ {3}}{Y _ {3} Z _ {4} \left(b _ {2} a _ {1} - b _ {1} a _ {2}\right)}\right) R _ {2 1}. \tag {13} \\ \end{array} +$$ + +Substituting (13) into (9), we obtain two bivariate quadratic constraints in $R_{11}$ and $R_{21}$ . In matrix form, + +$$ +\left( \begin{array}{l l l} c _ {1} & c _ {2} & c _ {3} \\ d _ {1} & d _ {2} & d _ {3} \end{array} \right) \cdot \left( \begin{array}{c} R _ {1 1} ^ {2} \\ R _ {1 1} R _ {2 1} \\ R _ {2 1} ^ {2} \end{array} \right) = \binom {1} {1}, \tag {14} +$$ + +where the coefficients $c_{1}, c_{2}, c_{3}, d_{1}, d_{2}, d_{3}$ are rational functions of the problem data. Applying the change of variables + +$$ +u = R _ {1 1} ^ {2}, \quad v = R _ {2 1} / R _ {1 1}, \tag {15} +$$ + +and subtracting the two equations in (14), we obtain + +$$ +\left(c _ {1} - d _ {1}\right) u + \left(c _ {2} - d _ {2}\right) u v + \left(c _ {3} - d _ {3}\right) u v ^ {2} = 0. +$$ + +Assuming $u \neq 0$ , we therefore have the univariate quadratic equation in $v$ + +$$ +\left(c _ {1} - d _ {1}\right) + \left(c _ {2} - d _ {2}\right) v + \left(c _ {3} - d _ {3}\right) v ^ {2} = 0. \tag {16} +$$ + +If $v$ is one of the roots of (16), we may recover a corresponding value for $u$ using one of the equations in (14), eg. + +$$ +u = \left(c _ {1} + c _ {2} v + c _ {3} v ^ {2}\right) ^ {- 1}, \quad \text {o r} \quad u = \left(d _ {1} + d _ {2} v + d _ {3} v ^ {2}\right) ^ {- 1}. \tag {17} +$$ + +To summarize, we provide the outline of steps for solving the P2P1L absolute pose problem in Figure 3. + +Algebraically, the nontrivial steps in Figure 3 are the second and fourth, which require solving a univariate quadratic equation and computing square roots, respectively. + +![](images/b4abe6b8df011381fda4452e30692a308445f6f91b28742711a701246179f0df.jpg) +Figure 3. Generic P2P1L solver. See Sec. 2.1 for details. + +![](images/9a0741bac57a0382aa9616fe8857009f1c20205d53c2e9f5227403b71972772b.jpg) + +![](images/a1a81f7e8cca138e09932e76cff33e20b2d1e8329fc6a6c10f92ee769aaa5f8d.jpg) +Figure 4. Special reference frame for P1P2L. Camera frame $\mathcal{C}_0$ is transformed to $\mathcal{C}_1$ , world frame $\mathcal{W}_0$ to $\mathcal{W}_1$ . See text for details. + +![](images/83257912b90f36c02c4353cfc522c5df9718efe3b89b8405f0bac422d13bd179.jpg) + +# 2.2.P1P2L + +Here, we provide the algebraically optimal solution to the P1P2L problem. We mostly follow the notation of [25]. + +Problem Statement: Let us consider a 3D point $P_1^0 \in \mathbb{R}^3$ , and its homogeneous 2D projection $D_1 \in \mathbb{R}^3$ under an unknown calibrated camera matrix. Let us also consider two 3D lines—the first spanned by points $L_1, L_2 \in \mathbb{R}^3$ , and the second one is spanned by $L_3, L_4 \in \mathbb{R}^3$ —and both of their corresponding projections. Our task is to recover the camera matrix from the given point-point correspondence and line-line correspondences. + +Much like the P2P1L solver, our P1P2L solver begins by transforming the input data into a special reference frame, as illustrated in Figure 4. Specifically, + +- In the world frame, the 3D point takes the form + +$$ +P _ {1} = \left( \begin{array}{l l l} 0 & 0 & 0 \end{array} \right) ^ {T}. \tag {18} +$$ + +The first line is spanned by points + +$$ +L _ {1} = \left( \begin{array}{l l l} X _ {1} & Y _ {1} & Z _ {1} \end{array} \right) ^ {T}, \quad L _ {2} = \left( \begin{array}{l l l} X _ {2} & Y _ {2} & Z _ {2} \end{array} \right) ^ {T}, \tag {19} +$$ + +and the second line is spanned by points + +$$ +L _ {3} = \left( \begin{array}{l l l} X _ {3} & Y _ {3} & Z _ {3} \end{array} \right) ^ {T}, \quad L _ {4} = \left( \begin{array}{l l l} X _ {4} & Y _ {4} & Z _ {4} \end{array} \right) ^ {T}. \tag {20} +$$ + +- For $\mathcal{W}_0 \to \mathcal{W}_1$ , we simply translate $P_1^0$ to $P_1$ . +- The world frame may still be rotated freely. We may use the strategy of Section 2.3, which makes the solver stable for a larger class of non-coplanar scenes. However, if $P_{1}, \overline{L_{1}L_{2}}$ , and $\overline{L_{3}L_{4}}$ are coplanar, no rotation is recommended. +- In the camera frame, the camera center is fixed at + +$$ +C = \left( \begin{array}{l l l} 0 & 0 & - 1 \end{array} \right) ^ {T}. \tag {21} +$$ + +The image point is the projection of a point of the form + +$$ +D _ {1} = \left( \begin{array}{l l l} a _ {1} & b _ {1} & 0 \end{array} \right) ^ {T}, \tag {22} +$$ + +the first line in the image is the projection of a line spanned by points of the form + +$$ +D _ {2} = \left( \begin{array}{l l l} a _ {2} & 0 & 0 \end{array} \right) ^ {T}, \quad D _ {3} = \left( \begin{array}{l l l} 0 & 0 & 0 \end{array} \right) ^ {T}, \tag {23} +$$ + +and the second line in the image is the projection of a line spanned by points of the form + +$$ +D _ {4} = \left( \begin{array}{l l l} a _ {4} & b _ {4} & 0 \end{array} \right) ^ {T}, \quad D _ {5} = \left( \begin{array}{l l l} a _ {5} & b _ {5} & 0 \end{array} \right) ^ {T}. \tag {24} +$$ + +- The homogeneous coordinates of the lines in the camera frame are $n_1 = d_2 \times d_3$ , $n_2 = d_4 \times d_5$ , where the rays $d_2 \neq d_3$ , $d_4 \neq d_5$ to meet the lines in distinct points. +To find the transformation $\mathcal{C}_0\to \mathcal{C}_1$ , we define + +$$ +d _ {1 2} = n _ {1} \times n _ {2}, \quad D _ {2} ^ {0} = C ^ {0} + \frac {d _ {2}}{d _ {2} ^ {T} d _ {1 2}}, \quad D _ {3} ^ {0} = C ^ {0} + d _ {3}, +$$ + +$$ +D _ {2} = \left( \begin{array}{c c} \tan (\cos^ {- 1} (d _ {2} ^ {T} d _ {1 2})) & 0 \quad 0 \end{array} \right) ^ {T}, \quad D _ {3} = \left( \begin{array}{c c c} 0 & 0 & 0 \end{array} \right) ^ {T}, +$$ + +and, independently of $C^0$ , rigidly transform the rays $d_2 \coloneqq \overrightarrow{C^0D_2^0}$ and $d_3 \coloneqq \overrightarrow{C^0D_3^0}$ to $\overrightarrow{C^0D_2^0}$ and $\overrightarrow{CD_3}$ . + +In the special reference frame, we now write the pose as $(R_{ij})_{1\leq i,j\leq 3}$ , $(T_i)_{1\leq i\leq 3}$ . The projection constraints can be formulated as: + +$$ +\left(D _ {1} - C\right) \times \left(R P _ {1} + T\right) = 0, \tag {25} +$$ + +$$ +\left(D _ {2} \times D _ {3}\right) ^ {T} \left(R L _ {i} + T\right) = 0, i \in \{1, 2 \}, \tag {26} +$$ + +$$ +\left(D _ {4} \times D _ {5}\right) ^ {T} \left(R L _ {i} + T\right) = 0, i \in \{3, 4 \}. \tag {27} +$$ + +Analogously to (11), we have a system of linear equations $A\mathbf{x} = \mathbf{b}$ , obtained by picking 2 out of the 3 redundant constraints from (25) together with (26), (27), where now (cf. [25, eq. 7-8]) + +$$ +A = \left( \begin{array}{c c c c c c c c c} 0 & 0 & 0 & 0 & 0 & 0 & - b _ {1} & a _ {1} & 0 \\ 0 & 0 & 0 & 0 & 0 & 0 & 0 & - 1 & b _ {1} \\ 0 & 0 & X _ {1} & Y _ {1} & Y _ {1} & Z _ {1} & 0 & 1 & 0 \\ 0 & 0 & X _ {2} & Y _ {2} & Y _ {2} & Z _ {2} & 0 & 1 & 0 \\ - b _ {4} X _ {3} & - b _ {4} Y _ {3} & a _ {4} X _ {3} & a _ {4} Y _ {3} & a _ {4} Y _ {3} & a _ {4} Z _ {3} & - b _ {4} & a _ {4} & 0 \\ - b _ {4} X _ {4} & - b _ {4} Z _ {4} & a _ {4} X _ {4} & a _ {4} Y _ {4} & a _ {4} Y _ {4} & a _ {4} Z _ {4} & - b _ {4} & a _ {4} & 0 \end{array} \right), +$$ + +$$ +\mathbf {x} = \left( \begin{array}{c c c c c c c c c} R _ {1 1} & R _ {1 2} & R _ {1 3} & R _ {2 1} & R _ {2 2} & R _ {2 3} & T _ {1} & T _ {2} & T _ {3} \end{array} \right) ^ {T}, +$$ + +$$ +\mathbf {b} = \left( \begin{array}{c c c c c c} 0 & - b _ {1} & 0 & 0 & 0 & 0 \end{array} \right) ^ {T}. +$$ + +Solving for translation as in (12), we have + +$$ +T _ {1} = \frac {a _ {1}}{b _ {1}} (- R _ {2 1} X _ {1} - R _ {2 2} Y _ {1} - R _ {2 3} Z _ {1}), +$$ + +$$ +T _ {2} = - R _ {2 1} X _ {1} - R _ {2 2} Y _ {1} - R _ {2 3} Z _ {1}, \tag {28} +$$ + +$$ +T _ {3} = - 1 + \frac {1}{b _ {1}} \left(- R _ {2 1} X _ {1} - R _ {2 2} Y _ {1} - R _ {2 3} Z _ {1}\right). +$$ + +Furthermore, we express $R_{23}$ from the fourth equation as + +$$ +R _ {2 3} = \frac {1}{Z _ {1} - Z _ {2}} \left(R _ {2 1} \left(X _ {2} - X _ {1}\right) + R _ {2 2} \left(Y _ {2} - Y _ {1}\right)\right), \tag {29} +$$ + +and we solve for $R_{11}, R_{12}$ using the last two equations. + +Now, we can express $R_{23}, T_2, T_1, R_{12}, R_{11}$ as linear combinations of $R_{13}, R_{21}, R_{22}$ in the following form: + +$$ +R _ {2 3} = c _ {1} R _ {2 1} + c _ {2} R _ {2 2}, +$$ + +$$ +T _ {2} = c _ {3} R _ {2 1} + c _ {4} R _ {2 2}, +$$ + +$$ +T _ {1} = c _ {5} R _ {2 1} + c _ {6} R _ {2 2}, \tag {30} +$$ + +$$ +R _ {1 2} = c _ {7} R _ {1 3} + c _ {8} R _ {2 1} + c _ {9} R _ {2 2}, +$$ + +$$ +R _ {1 1} = c _ {1 0} R _ {1 3} + c _ {1 1} R _ {2 1} + c _ {1 2} R _ {2 2}. +$$ + +The non-linear internal constraints imposed on the elements of $R$ have the form: + +$$ +R _ {1 1} ^ {2} + R _ {1 2} ^ {2} + R _ {1 3} ^ {2} = 1, +$$ + +$$ +R _ {2 1} ^ {2} + R _ {2 2} ^ {2} + R _ {2 3} ^ {2} = 1, \tag {31} +$$ + +$$ +R _ {1 1} R _ {2 1} + R _ {1 2} R _ {2 2} + R _ {1 3} R _ {2 3} = 0. +$$ + +We substitute (30) into (31), obtaining equations of the form + +$$ +\left( \begin{array}{l l l l l l} d _ {1} & d _ {2} & d _ {3} & d _ {4} & d _ {5} & d _ {6} \end{array} \right) \nu (R) = 1, \tag {32} +$$ + +$$ +\left( \begin{array}{c c c c c c} 0 & 0 & 0 & d _ {7} & d _ {8} & d _ {9} \end{array} \right) \nu (R) = 1, \tag {33} +$$ + +$$ +\left( \begin{array}{c c c c c c} 0 & d _ {1 0} & d _ {1 1} & d _ {1 2} & d _ {1 3} & d _ {1 4} \end{array} \right) \nu (R) = 0, \tag {34} +$$ + +where + +$$ +\nu (R) = \left( \begin{array}{l l l l l} R _ {1 3} ^ {2} & R _ {1 3} R _ {2 1} & R _ {1 3} R _ {2 2} & R _ {2 1} ^ {2} R _ {2 1} R _ {2 2} & R _ {2 2} ^ {2} \end{array} \right) ^ {T}. \tag {35} +$$ + +We express $R_{13}$ from (34) as + +$$ +R _ {1 3} = - \frac {d _ {1 2} R _ {2 1} ^ {2} + d _ {1 3} R _ {2 1} R _ {2 2} + d _ {1 4} R _ {2 2} ^ {2}}{d _ {1 0} R _ {2 1} + d _ {1 1} R _ {2 2}}. \tag {36} +$$ + +Substituting (36) into (32), we obtain + +$$ +d _ {1} \frac {\left(d _ {1 2} R _ {2 1} ^ {2} + d _ {1 3} R _ {2 1} R _ {2 2} + d _ {1 4} R _ {2 2} ^ {2}\right) ^ {2}}{\left(d _ {1 0} R _ {2 1} + d _ {1 1} R _ {2 2}\right) ^ {2}} - +$$ + +$$ +d _ {2} \frac {d _ {1 2} R _ {2 1} ^ {3} + d _ {1 3} R _ {2 1} ^ {2} R _ {2 2} + d _ {1 4} R _ {2 1} R _ {2 2} ^ {2}}{d _ {1 0} R _ {2 1} + d _ {1 1} R _ {2 2}} - +$$ + +$$ +d _ {3} \frac {d _ {1 2} R _ {2 1} ^ {2} R _ {2 2} + d _ {1 3} R _ {2 1} R _ {2 2} ^ {2} + d _ {1 4} R _ {2 2} ^ {3}}{d _ {1 0} R _ {2 1} + d _ {1 1} R _ {2 2}} + +$$ + +$$ +d _ {4} R _ {2 1} ^ {2} + d _ {5} R _ {2 1} R _ {2 2} + d _ {6} R _ {2 2} ^ {2} = 1. \tag {37} +$$ + +Subtracting (33) from (37), + +$$ +d _ {1} \frac {\left(d _ {1 2} R _ {2 1} ^ {2} + d _ {1 3} R _ {2 1} R _ {2 2} + d _ {1 4} R _ {2 2} ^ {2}\right) ^ {2}}{\left(d _ {1 0} R _ {2 1} + d _ {1 1} R _ {2 2}\right) ^ {2}} - +$$ + +$$ +d _ {2} \frac {d _ {1 2} R _ {2 1} ^ {3} + d _ {1 3} R _ {2 1} ^ {2} R _ {2 2} + d _ {1 4} R _ {2 1} R _ {2 2} ^ {2}}{d _ {1 0} R _ {2 1} + d _ {1 1} R _ {2 2}} - +$$ + +$$ +d _ {3} \frac {d _ {1 2} R _ {2 1} ^ {2} R _ {2 2} + d _ {1 3} R _ {2 1} R _ {2 2} ^ {2} + d _ {1 4} R _ {2 2} ^ {3}}{d _ {1 0} R _ {2 1} + d _ {1 1} R _ {2 2}} + +$$ + +$$ +\left(d _ {4} - d _ {7}\right) R _ {2 1} ^ {2} + \left(d _ {5} - d _ {8}\right) R _ {2 1} R _ {2 2} + \left(d _ {6} - d _ {9}\right) R _ {2 2} ^ {2} = 0. \tag {38} +$$ + +(1) Transform data in world and camera frames to the special reference frames satisfying (18), (19), (20), (21), (22), (23), (24). +(2) Compute up to 4 solutions in $v$ to the quartic (41). +(3) For each root $v$ in step (2), recover 2 solutions in $\{R_{21}, R_{22}\}$ using $R_{21} = R_{22}v$ and (33). +(4) Recover remaining coordinates for each solution using (34),(30). +(5) Reverse the transformation applied in step (1). + +![](images/0b3fc4accd39c7cfdb4da1c7c5909a4ea8e8cff9c9a37faba300cce6245b0a78.jpg) +Figure 6. Stabilization of P1P2L solver. The frame is rotated to align the line $\overline{L_1L_2}$ with the z-axis. See Sec. 2.3 for details. + +![](images/cd7d9f719a68f865c3772ae71e7acfebff8f70065de201c7b0b0fe42427766d8.jpg) +Figure 5. Generic P1P2L solver. See Sec. 2.2 for details. + +We then clear denominators in (38), multiplying by $(d_{10}R_{21} + d_{11}R_{22})^2$ to get a polynomial equation + +$$ +\begin{array}{l} d _ {1} \left(d _ {1 2} R _ {2 1} ^ {2} + d _ {1 3} R _ {2 1} R _ {2 2} + d _ {1 4} R _ {2 2} ^ {2}\right) ^ {2} - \\ \left(\left(d _ {2} \left(d _ {1 2} R _ {2 1} ^ {3} + d _ {1 3} R _ {2 1} ^ {2} R _ {2 2} + d _ {1 4} R _ {2 1} R _ {2 2} ^ {2}\right) - \right. \right. \\ \left. d _ {3} \left(d _ {1 2} R _ {2 1} ^ {2} R _ {2 2} + d _ {1 3} R _ {2 1} R _ {2 2} ^ {2} + d _ {1 4} R _ {2 2} ^ {3}\right)\right) \cdot \tag {39} \\ \left(d _ {1 0} R _ {2 1} + d _ {1 1} R _ {2 2}\right) + \left(\left(d _ {4} - d _ {7}\right) R _ {2 1} ^ {2} + \right. \\ \left. \left(d _ {5} - d _ {8}\right) R _ {2 1} R _ {2 2} + \left(d _ {6} - d _ {9}\right) R _ {2 2} ^ {2}\right) \Bigg). \\ \left(d _ {1 0} R _ {2 1} + d _ {1 1} R _ {2 2}\right) = 0. \\ \end{array} +$$ + +Upon expanding equation (39), we find that it takes the form + +$$ +\alpha_ {1} R _ {2 1} ^ {4} + \alpha_ {2} R _ {2 1} ^ {3} R _ {2 2} + \alpha_ {3} R _ {2 1} ^ {2} R _ {2 2} ^ {2} + \alpha_ {4} R _ {2 1} R _ {2 2} ^ {3} + \alpha_ {5} 2 R _ {2 2} ^ {4} = 0 \tag {40} +$$ + +Similarly to (16), we divide (40) by $R_{22}^{4}$ and define a new variable $v = \frac{R_{21}}{R_{22}}$ , and thereby deduce the univariate quartic + +$$ +\alpha_ {1} v ^ {4} + \alpha_ {2} v ^ {3} + \alpha_ {3} v ^ {2} + \alpha_ {4} v + \alpha_ {5} = 0. \tag {41} +$$ + +After solving for $v$ , we can solve for $R_{21}, R_{22}$ using (33), and for the other equations using (34), and (30). + +To summarize, we provide the outline of steps for solving the P1P2L absolute pose problem in Figure 5. Algebraically, the nontrivial steps in Figure 5 are second and third, which require solving a univariate quartic equation and computing square roots, respectively. + +
GenericMean RMed. RMean TMed. T
P1P2L no fix0.000501.0e-140.000651.8e-13
P1P2L fix9.0e-094.2e-153.4e-077.0e-14
P2P1L no fix7.1e-111.4e-151.2e-092.1e-14
P2P1L fix3.2e-092.6e-157.4e-084.3e-14
CoplanarMean RMed. RMean TMed. T
P1P2L no fix0.000229.6e-150.000301.75e-13
P1P2L fix1.10.791.20.96
P2P1L no fix2.43.142.23.14
P2P1L fix1.2e-124.0e-157.9e-116.3e-14
+ +Table 2. The mean and median of the rot.(R) and tran.(T) errors over 1e5 noiseless generic (top) and coplanar (bottom) samples. In radians. The best results are marked bold. + +# 2.3. Stabilizing the P1P2L Solver + +In this section, we outline a method to increase the stability of the P1P2L solver (Sec. 2.2). + +Our proposed P1P2L solution faces a degeneracy when $Z_{1} - Z_{2} = 0$ , since this leads to a division by zero in equation (29). Furthermore, our observations indicate that the result of the solver is unstable if the value of $Z_{1} - Z_{2}$ is close to zero. Note, that this instability is specific to our solver and not inherent to the P1P2L problem. + +This degeneracy may be interpreted geometrically as follows: the values of $Z_{1}$ and $Z_{2}$ are the last coordinates of the points $L_{1}, L_{2}$ , which span the first 3D line. Therefore, the vector $L_{1} - L_{2}$ represents the direction of the line, and $Z_{1} - Z_{2}$ represents the last coordinate of this direction. + +Since the special reference frame used in Section 2.2 is independent of the rotation of the world frame, we can remove the source of instability by rotating the world frame such that the first line $\overline{L_1L_2}$ aligns with the z-axis. See Figure 6 for an illustration. + +# 2.4. Resolving the coplanar case + +In this section, we discuss the performance of the P2P1L (Sec. 2.1) and P1P2L (Sec. 2.2) solvers in the case, where all 3D points and 3D lines are coplanar. + +P2P1L. The P2P1L solver presented in Sec. 2.1 (P2P1L no fix) is degenerate in coplanar situations, since then $Z_{4} = 0$ (2), causing a division by zero issue in equation (11). To resolve this issue, we propose a modified solver (P2P1L fix), which handles the coplanar case. See SM A for a description of this modified solver. + +P1P2L. The original version of the P1P2L solver (P1P2L no fix, Sec. 2.2) is able to handle the coplanar case. However, the stabilized P1P2L solver (P1P2L fix, Sec. 2.3) fails due to a degeneracy for coplanar input. + +We conducted a synthetic experiment to evaluate the impact of the proposed stabilization scheme. In this experiment, we used the same setting as in Section 3.1. The re + +![](images/f3ac5acb4ef713bba12a1367b6196e05f9c5c0ef13e64dd9aeee7b2e99403755.jpg) + +![](images/f62ca8369537d116d1a2718261a301ede8c087a96305d92bf80aed04c791324d.jpg) +Figure 7. Stability test. Histogram of rotation $(l.)$ , and translation $(r.)$ errors computed over $1e5$ noiseless samples. + +
MethodMeanMed.MinMax
P2P1L Ours313.8324.9230.83061.0
P2P1L 3Q31860.61909.91439.110102.1
P2P1L R.+S8897.59491.25805.149984.1
P2P1L R.+L4720.85239.52763.315362.6
P1P2L Ours504.0521.0364.04554.4
P1P2L 3Q31967.12008.21483.912931.1
+ +Table 3. Runtime The mean, median and max values of the runtime. In nanoseconds. Best values for each problem are bold. + +sults, presented in Table 2 demonstrate that solvers $P2P1L$ no fix and $P1P2L$ fix achieve superior numerical stability in the generic case, while solvers $P2P1L$ fix and $P1P2L$ no fix handle the coplanar case. Since it is simple to detect the coplanar case, we recommend to use solvers $P2P1L$ fix and $P1P2L$ no fix in the coplanar case and solvers $P2P1L$ no fix and $P1P2L$ fix in the generic case. + +# 3. Experiments + +In this section, we experimentally compare the proposed solvers with the state-of-the-art methods, specifically 3Q3 [17] and Ramalingam [25]. For 3Q3, we employ the publicly available implementation from Poselib. As the implementation of Ramalingam is not publicly accessible, we have created our own implementation. Since it is not explicitly specified which method should be used for solving the linear equations, we compare with two variants: SVD and LU decomposition. All the solvers have been implemented in $\mathrm{C + + }$ , and the experiments are conducted on a desktop computer with an AMD Ryzen 9 CPU with $3.9\mathrm{GHz}$ . + +In Sec. 3.1, we provide an analysis of the numerical stability and runtime performance of the solvers on synthetic data. Subsequently, in Section 3.2, we present an evaluation of the solvers within the RANSAC scheme. + +# 3.1. Synthetic experiments + +In this section, we present an analysis of the numerical stability and runtime performance of the minimal solvers on synthetic data. + +We generate instances of each minimal problem (either P2P1L or P1P2L), according to the following procedure: + +- Rotation Matrix $(R)$ : An axis $v \in \mathbb{R}^3$ is sampled from the uniform distribution on the unit sphere, and an angle $\alpha$ is sampled from the normal distribution $\mathcal{N}(0,1)$ . The rotation matrix $R$ is then constructed using the angle-axis formula, $R = I + \sin(\alpha)[v]_{\times} + (1 - \cos(\alpha))[v]_{\times}^2$ . +- Translation $(T)$ : The camera center $C$ is sampled uniformly at random from the unit sphere. The translation vector $T$ is computed as $T = -RC$ . +- Point correspondence: A 3D point $X_{i}$ is sampled from the trivariate normal distribution with mean vector $\mu = [0,0,5]^T$ and standard deviation $\sigma = 1$ in each component. The corresponding 2D point $D_{i}$ is obtained by projecting $X_{i}$ onto a pinhole camera with the pose $(R,T)$ . +- Line correspondence: Two 3D points, $L_{i}$ and $L_{i+1}$ , are sampled as described in the previous bullet-point. These points define the 3D line $L$ . Two 2D points, $D_{i}$ and $D_{i+1}$ , are obtained by sampling two points on the 3D line $L$ and projecting them onto the camera with the pose $(R, T)$ . + +In the case of P2P1L, two points and one line are sampled, while in the case of P1P2L, one point and two lines are sampled. Let $(R_{est},T_{est})$ be the pose obtained by the minimal solver. We measure the rotation error as the angle $\xi_R = \arccos ((1 + \mathrm{trace}(R_{est}^T R)) / 2)$ . We calculate the translation error $\xi_T$ as $\xi_T = \frac{|T_{est} - T|}{|T|}$ . The histograms of rotation errors $(\xi_R)$ and translation errors $(\xi_T)$ for all considered solvers are depicted in Figure 7. Furthermore, summary statistics including the median, mean, and maximum errors are provided in Table 4. The results indicate that all minimal solvers are stable. Our algebraically optimal solvers demonstrate superior stability compared to the previous solvers. + +The runtime evaluation for all considered solvers is shown in Table. 3. Our P2P1L solver requires 313.8ns on average, which is about 6x faster than the 3Q3 solver [17]. Similarly, our P1P2L solver requires on average 504.0ns, which is about 4x faster than the 3Q3 solver. These results show a significant speedup achieved by our solvers in comparison to the previous methods. + +# 3.2. Experiments in RANSAC + +Here, we provide an analysis of the minimal solvers within the RANSAC scheme [4], using the Oxford Multi-view Data [1]. This dataset includes a variety of indoor and outdoor scenes, some of which contain both point and line matches. An example of images from this dataset is illustrated in Figure 8. + +
MethodMean RMedian RMax RMean TMedian TMax T
P2P1L Ours5.3e-121.4e-151.2e-073.7e-102.1e-142.2e-05
P2P1L 3Q32.8e-056.5e-152.72.0e-051.2e-131.04
P2P1L Ramalingam (SVD)4.9e-091.5e-140.000411.6e-072.7e-130.0099
P2P1L Ramalingam (LU)4.7e-071.3e-140.0402.3e-052.4e-130.10
PP1P2L Ours (fix 2)1.2e-074.4e-150.0102.0e-067.1e-140.13
P1P2L 3Q33.3e-057.2e-152.603.4e-051.5e-131.01
+ +Table 4. The mean, median and max values of the rot.(R) and tran.(T) errors over ${1e5}$ noiseless samples. The best results are marked bold. + +
DatasetP2P1LP1P2L
OUR3Q3Ram. SVDRam. LUOUR3Q3
Model House7.72 (0.85x)9.09 (1x)20.69 (2.28x)17.30 (1.90x)8.06 (0.85x)9.52 (1x)
Corridor11.32 (0.91x)12.46 (1x)26.56 (2.13x)23.54 (1.89x)13.3 (0.90x)14.76 (1x)
Merton I35.93 (0.98x)36.83 (1x)77.96 (2.12x)74.87 (2.03x)28.85 (0.64x)45.41 (1x)
Merton II33.30 (0.97x)34.27 (1x)74.37 (2.17x)71.35 (2.08x)26.7 (0.67x)39.64 (1x)
Merton III24.04 (0.96x)25.07 (1x)52.07 (2.08x)48.97 (1.95x)10.91 (0.37x)29.67 (1x)
Library32.42 (0.97x)33.31 (1x)69.29 (2.08x)66.23 (1.99x)10.04 (0.28x)35.98 (1x)
Wadham39.60 (0.98x)40.51 (1x)86.10 (2.13x)83.00 (2.05x)22.96 (0.45x)50.79 (1x)
Avg. Speed-up0.94x1x2.14x1.99x0.59x1x
+ +Table 5. RANSAC timing, on Oxford Multi-view dataset [1], in milliseconds. Speed-up compared to the 3Q3 in the bracket. + +![](images/1a1467d082b96b6409e6de7ca27d3e1e313ffb3301faa7f70a5f7a861fa0ba8f.jpg) + +![](images/c0a5efd34e5788c9b8db10104226b00eadb0bafe109346eed28a4f5da64ce227.jpg) + +![](images/937618e532ccbafba2a2787bad9aa48be71248dcdeeb2a2423e084d01d652d76.jpg) +Figure 8. Example images of University Library (top) and Merton College III (bottom), from [1]. + +![](images/7774b5909468adbd6f981ab9a5bb9161a98e5523a8b2af6612a8aaeaecc5f8f2.jpg) + +We utilize the Poselib [18] implementation of LORANSAC [6]. In this implementation, local optimization is applied both when the new solution surpasses the currently best one and at the end of the entire procedure. The RANSAC parameters used include a maximum number of iterations set to 100000, a minimum number of iterations set to 1000, a success probability of 0.9999, and an inlier threshold of 1 pixel. To ensure a fair comparison between the solvers, we use a fixed random seed. + +Since all the solvers are stable, we do not expect any significant difference in the estimated pose when using different solvers. Therefore, our primary focus is on comparing the runtime of the RANSAC scheme. The runtime comparison is presented in Table 5, which displays the average runtime for each scene in the Oxford multiview dataset and + +for each considered solver. Additionally, the table includes a speedup value compared to the 3Q3 solver of [18]. The results show that our solvers consistently outperform the 3Q3 [17] and Ramalingam [25] solvers in terms of runtime. However, there is a notable variation in the speed-up among the scenes, which ranges from 0.85 to 0.98 for the P2P1L problem and from 0.28 to 0.90 for the P1P2L solver. This variation can be attributed to the fact that different scenes require varying proportions of time spent by RANSAC on scoring hypothesized models and local optimization. The more time is dedicated to these tasks, the less significant speedup can be reached. In cases with fewer matches or a lower inlier ratio, more substantial speedups are possible. + +The final pose estimation errors in RANSAC are shown in the Supplementary. As expected, the errors are the same for all solvers. This demonstrates that we can achieve improved runtimes without sacrificing any accuracy. + +# 4. Conclusion + +In revisiting absolute pose from mixed point-line correspondences, we have developed new solvers, which have the merit of being both algebraically optimal and elementary in nature. Moreover, our solvers outperform the state-of-the-art in terms of both runtime and numerical stability. Despite the simplicity of many minimal absolute pose problems, our findings suggest that their continued investigation remains worthwhile. The code is available at https://github.com/petrhruby97/efficientAbsolutely + +Acknowledgements: TD was supported by NSF DMS 2103310. + +# References + +[1] Multi-view data, oxford visual geometry group. https://www.robots.ox.ac.uk/vgg/data/mview/. Accessed: Nov 8 2023. 7, 8, 1 +[2] Sérgio Agostinho, João Gomes, and Alessio Del Bue. CvxPnPL: A unified convex solution to the absolute pose estimation problem from point and line correspondences. Journal of Mathematical Imaging and Vision, 65(3):492-512, 2023. 1 +[3] Daniel Barath and Jiri Matas. Graph-cut RANSAC: local optimization on spatially coherent structures. IEEE Trans. Pattern Anal. Mach. Intell., 44(9):4961-4974, 2022. 1 +[4] Robert C. Bolles and Martin A. Fischler. A ransac-based approach to model fitting and its application to finding cylinders in range data. In Proceedings of the 7th International Joint Conference on Artificial Intelligence, IJCAI '81, Vancouver, BC, Canada, August 24-28, 1981, pages 637-643. William Kaufmann, 1981. 7 +[5] Homer H. Chen. Pose determination from line-to-plane correspondences: Existence condition and closed-form solutions. IEEE Trans. Pattern Anal. Mach. Intell., 13(6):530-541, 1991. 1 +[6] Ondrej Chum, Jiri Matas, and Josef Kittler. Locally optimized RANSAC. In Pattern Recognition, 25th DAGM Symposium, Magdeburg, Germany, September 10-12, 2003, Proceedings, pages 236-243. Springer, 2003. 8 +[7] Michel Dhome, Marc Richetin, Jean-Thierry Lapresté, and Gérard Rives. Determination of the attitude of 3d objects from a single perspective view. IEEE Trans. Pattern Anal. Mach. Intell., 11(12):1265-1278, 1989. 1 +[8] Yaqing Ding, Jian Yang, Viktor Larsson, Carl Olsson, and Kalle Åström. Revisiting the P3P problem. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2023, Vancouver, BC, Canada, June 17-24, 2023, pages 4872-4880. IEEE, 2023. 1, 2 +[9] Timothy Duff, Viktor Korotynskiy, Tomas Pajdla, and Margaret H. Regan. Galois/monodromy groups for decomposing minimal problems in 3d reconstruction. SIAM Journal on Applied Algebra and Geometry, 6(4):740-772, 2022. 2 +[10] Ricardo Fabbri, Peter J. Giblin, and Benjamin B. Kimia. Camera pose estimation using first-order curve differential geometry. IEEE Trans. Pattern Anal. Mach. Intell., 43(10): 3321-3332, 2021. 1 +[11] Martin A. Fischler and Robert C. Bolles. Random sample consensus: A paradigm for model fitting with applications to image analysis and automated cartography. Commun. ACM, 24(6):381-395, 1981. 1 +[12] Christian Hane, Lionel Heng, Gim Hee Lee, Friedrich Fraundorfer, Paul Furgale, Torsten Sattler, and Marc Pollefeys. 3d visual perception for self-driving cars using a multi-camera system: Calibration, mapping, localization, and obstacle detection. Image Vis. Comput., 68:14-27, 2017. 1 +[13] Robert M. Haralick, Chung-Nan Lee, Karsten Ottenberg, and Michael Nölle. Review and analysis of solutions of the three point perspective pose estimation problem. Int. J. Comput. Vis., 13(3):331-356, 1994. 1 + +[14] Petr Hruby, Viktor Korotynskiy, Timothy Duff, Luke Oeding, Marc Pollefeys, Tomás Pajdla, and Viktor Larsson. Four-view geometry with unknown radial distortion. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2023, Vancouver, BC, Canada, June 17-24, 2023, pages 8990-9000. IEEE, 2023. 2 +[15] Xiaomei Hu, Luying Zhu, Ping Wang, Haili Yang, and Xuan Li. Improved ORB-SLAM2 mobile robot vision algorithm based on multiple feature fusion. IEEE Access, 11:100659-100671, 2023. 1 +[16] Laurent Kneip, Davide Scaramuzza, and Roland Siegwart. A novel parametrization of the perspective-three-point problem for a direct computation of absolute camera position and orientation. In The 24th IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2011, Colorado Springs, CO, USA, 20-25 June 2011, pages 2969-2976. IEEE Computer Society, 2011. 1 +[17] Zuzana Kukelova, Jan Heller, and Andrew W. Fitzgibbon. Efficient intersection of three quadrics and applications in computer vision. In 2016 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2016, Las Vegas, NV, USA, June 27-30, 2016, pages 1799-1808. IEEE Computer Society, 2016. 2, 7, 8 +[18] Viktor Larsson and contributors. PoseLib - Minimal Solvers for Camera Pose Estimation, 2020. 2, 8 +[19] Shaohui Liu, Yifan Yu, Rémi Pautrat, Marc Pollefeys, and Viktor Larsson. 3d line mapping revisited. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2023, Vancouver, BC, Canada, June 17-24, 2023, pages 21445-21455. IEEE, 2023. 1 +[20] David Nister, Richard I. Hartley, and Henrik Stewenius. Using galois theory to prove structure from motion algorithms are optimal. In 2007 IEEE Computer Society Conference on Computer Vision and Pattern Recognition (CVPR 2007), 18-23 June 2007, Minneapolis, Minnesota, USA. IEEE Computer Society, 2007. 2 +[21] Rémi Pautrat, Daniel Barath, Viktor Larsson, Martin R. Oswald, and Marc Pollefeys. Deeplsd: Line segment detection and refinement with deep image gradients. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2023, Vancouver, BC, Canada, June 17-24, 2023, pages 17327-17336. IEEE, 2023. 1 +[22] Rémi Pautrat, Iago Suárez, Yifan Yu, Marc Pollefeys, and Viktor Larsson. Gluestick: Robust image matching by sticking points and lines together. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9706-9716, 2023. 1 +[23] Mikael Persson and Klas Nordberg. Lambda twist: An accurate fast robust perspective three point (P3P) solver. In Computer Vision - ECCV 2018 - 15th European Conference, Munich, Germany, September 8-14, 2018, Proceedings, Part IV, pages 334-349. Springer, 2018. 1, 2 +[24] Rahul Raguram, Ondrej Chum, Marc Pollefeys, Jiri Matas, and Jan-Michael Frahm. USAC: A universal framework for random sample consensus. IEEE Trans. Pattern Anal. Mach. Intell., 35(8):2022-2038, 2013. 1 +[25] Srikumar Ramalingam, Sofien Bouaziz, and Peter F. Sturm. Pose estimation using both points and lines for geo- + +localization. In IEEE International Conference on Robotics and Automation, ICRA 2011, Shanghai, China, 9-13 May 2011, pages 4716-4723. IEEE, 2011. 1, 2, 3, 4, 5, 7, 8 +[26] Torsten Sattler, Bastian Leibe, and Leif Kobbelt. Efficient & effective prioritized matching for large-scale image-based localization. IEEE Trans. Pattern Anal. Mach. Intell., 39(9): 1744-1756, 2017. 1 +[27] Johannes L. Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In 2016 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2016, Las Vegas, NV, USA, June 27-30, 2016, pages 4104-4113. IEEE Computer Society, 2016. 1 +[28] Jonathan Ventura, Clemens Arth, Gerhard Reitmayr, and Dieter Schmalstieg. Global localization from monocular SLAM on a mobile phone. IEEE Trans. Vis. Comput. Graph., 20(4):531-539, 2014. 1 +[29] Chi Xu, Lilian Zhang, Li Cheng, and Reinhard Koch. Pose estimation from line correspondences: A complete analysis and a series of solutions. IEEE Trans. Pattern Anal. Mach. Intell., 39(6):1209-1222, 2017. 1 +[30] Lipu Zhou, Jiamin Ye, and Michael Kaess. A stable algebraic camera pose estimation for minimal configurations of 2d/3d point and line correspondences. In Computer Vision - ACCV 2018 - 14th Asian Conference on Computer Vision, Perth, Australia, December 2-6, 2018, Revised Selected Papers, Part IV, pages 273-288. Springer, 2018. 1, 2 \ No newline at end of file diff --git a/2024/Efficient Solution of Point-Line Absolute Pose/images.zip b/2024/Efficient Solution of Point-Line Absolute Pose/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..deccad0a17304a96633fb980d2a1c851f503eefc --- /dev/null +++ b/2024/Efficient Solution of Point-Line Absolute Pose/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2449cd932de53f2f58fd2c038f5f238ef08492137e775a28f69ea9d45d32c2a8 +size 824308 diff --git a/2024/Efficient Solution of Point-Line Absolute Pose/layout.json b/2024/Efficient Solution of Point-Line Absolute Pose/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..4ae99ffdaeef3590046f0bb7f6d743c75328813f --- /dev/null +++ b/2024/Efficient Solution of Point-Line Absolute Pose/layout.json @@ -0,0 +1,13200 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 156, + 103, + 440, + 119 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 156, + 103, + 440, + 119 + ], + "spans": [ + { + "bbox": [ + 156, + 103, + 440, + 119 + ], + "type": "text", + "content": "Efficient Solution of Point-Line Absolute Pose" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 113, + 144, + 173, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 113, + 144, + 173, + 171 + ], + "spans": [ + { + "bbox": [ + 113, + 144, + 173, + 171 + ], + "type": "text", + "content": "Petr Hruby ETH Zürich" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 206, + 144, + 332, + 172 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 206, + 144, + 332, + 172 + ], + "spans": [ + { + "bbox": [ + 206, + 144, + 332, + 172 + ], + "type": "text", + "content": "Timothy Duff \nUniversity of Washington" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 362, + 144, + 477, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 362, + 144, + 477, + 171 + ], + "spans": [ + { + "bbox": [ + 362, + 144, + 477, + 171 + ], + "type": "text", + "content": "Marc Pollefeys \nETH Zürich, Microsoft" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 143, + 199, + 192, + 211 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 199, + 192, + 211 + ], + "spans": [ + { + "bbox": [ + 143, + 199, + 192, + 211 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 224, + 290, + 429 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 224, + 290, + 429 + ], + "spans": [ + { + "bbox": [ + 46, + 224, + 290, + 429 + ], + "type": "text", + "content": "We revisit certain problems of pose estimation based on 3D-2D correspondences between features which may be points or lines. Specifically, we address the two previously-studied minimal problems of estimating camera extrinsics from " + }, + { + "bbox": [ + 46, + 224, + 290, + 429 + ], + "type": "inline_equation", + "content": "p \\in \\{1,2\\}" + }, + { + "bbox": [ + 46, + 224, + 290, + 429 + ], + "type": "text", + "content": " point-point correspondences and " + }, + { + "bbox": [ + 46, + 224, + 290, + 429 + ], + "type": "inline_equation", + "content": "l = 3 - p" + }, + { + "bbox": [ + 46, + 224, + 290, + 429 + ], + "type": "text", + "content": " line-line correspondences. To the best of our knowledge, all of the previously-known practical solutions to these problems required computing the roots of degree " + }, + { + "bbox": [ + 46, + 224, + 290, + 429 + ], + "type": "inline_equation", + "content": "\\geq 4" + }, + { + "bbox": [ + 46, + 224, + 290, + 429 + ], + "type": "text", + "content": " (univariate) polynomials when " + }, + { + "bbox": [ + 46, + 224, + 290, + 429 + ], + "type": "inline_equation", + "content": "p = 2" + }, + { + "bbox": [ + 46, + 224, + 290, + 429 + ], + "type": "text", + "content": ", or degree " + }, + { + "bbox": [ + 46, + 224, + 290, + 429 + ], + "type": "inline_equation", + "content": "\\geq 8" + }, + { + "bbox": [ + 46, + 224, + 290, + 429 + ], + "type": "text", + "content": " polynomials when " + }, + { + "bbox": [ + 46, + 224, + 290, + 429 + ], + "type": "inline_equation", + "content": "p = 1" + }, + { + "bbox": [ + 46, + 224, + 290, + 429 + ], + "type": "text", + "content": ". We describe and implement two elementary solutions which reduce the degrees of the needed polynomials from 4 to 2 and from 8 to 4, respectively. We show experimentally that the resulting solvers are numerically stable and fast: when compared to the previous state-of-the-art, we may obtain nearly an order of magnitude speedup. The code is available at https://github.com/petrhruby97/efficient_absolute" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 448, + 128, + 460 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 448, + 128, + 460 + ], + "spans": [ + { + "bbox": [ + 47, + 448, + 128, + 460 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 468, + 122, + 479 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 468, + 122, + 479 + ], + "spans": [ + { + "bbox": [ + 47, + 468, + 122, + 479 + ], + "type": "text", + "content": "1.1. Motivation" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 486, + 287, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 486, + 287, + 605 + ], + "spans": [ + { + "bbox": [ + 46, + 486, + 287, + 605 + ], + "type": "text", + "content": "The problem of registering images to a known 3D coordinate system plays a crucial role in applications such as visual localization [26], autonomous driving [12], and augmented reality [28], as well as in general paradigms like SLAM [15] and SfM [27]. Robust estimators based on RANSAC [11], or one of its many refinements [3, 24], are among the most successful tools for solving these problems. Such an estimator traditionally relies on a minimal P3P solver [8, 13, 16, 23] to efficiently hypothesize poses from putative matches between 3D and 2D points." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 605, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 288, + 714 + ], + "type": "text", + "content": "The literature on P3P and other purely point-based methods for absolute pose estimation is vast. Absolute pose estimation from non-point features such as lines [2, 5, 7, 25, 29, 30], point-line incidences [10], and affine correspondences [28], has received comparatively less attention, but remains an active research area. In particular, solutions relying on both points and lines are of increasing importance, due to the prevalence of both types of feature in man-made environments, as well as several recent advances in the com" + } + ] + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 325, + 196, + 531, + 365 + ], + "blocks": [ + { + "bbox": [ + 325, + 196, + 531, + 365 + ], + "lines": [ + { + "bbox": [ + 325, + 196, + 531, + 365 + ], + "spans": [ + { + "bbox": [ + 325, + 196, + 531, + 365 + ], + "type": "image", + "image_path": "92879d98ea4ada4118f256e76bcc5e1846b9d6f5e2287fa51eb66b458b2b3d8e.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 318, + 371, + 533, + 384 + ], + "lines": [ + { + "bbox": [ + 318, + 371, + 533, + 384 + ], + "spans": [ + { + "bbox": [ + 318, + 371, + 533, + 384 + ], + "type": "text", + "content": "Figure 1. The P2P1L (Top) and P1P2L (Bottom) problems." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 406, + 545, + 431 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 406, + 545, + 431 + ], + "spans": [ + { + "bbox": [ + 305, + 406, + 545, + 431 + ], + "type": "text", + "content": "ponents of 3D reconstruction systems responsible for line detection [21], matching [22], and bundle adjustment [19]." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 441, + 390, + 452 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 441, + 390, + 452 + ], + "spans": [ + { + "bbox": [ + 306, + 441, + 390, + 452 + ], + "type": "text", + "content": "1.2. Contribution" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 460, + 545, + 520 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 460, + 545, + 520 + ], + "spans": [ + { + "bbox": [ + 304, + 460, + 545, + 520 + ], + "type": "text", + "content": "In this paper, we revisit the two minimal problems of absolute pose estimation that combine both points and lines: the Perspective-2-Point-1-Line Problem (or P2P1L, Section 2.1), and the Perspective-1-Point-2-Line Problem (P1P2L, Section 2.2.) See Figure 1 for illustrations." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 521, + 546, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 521, + 546, + 652 + ], + "spans": [ + { + "bbox": [ + 304, + 521, + 546, + 652 + ], + "type": "text", + "content": "In contrast to the purely point or line based minimal problems P3P and P3L, we observe that the existing solutions for the \"mixed cases\" considered here are suboptimal from both a theoretical and practical point of view. Consequently, we develop novel solutions to both problems, which optimally exploit their underlying algebraic structure, and exhibit comparable or better performance than the state of the art on simulated and real data—see Section 3. On the other hand, these solvers are simple to implement, and no knowledge of mathematics beyond elementary algebra is needed to understand them." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 653, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 653, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 653, + 547, + 715 + ], + "type": "text", + "content": "To provide further context for our work, we include the following quotation from [25]: Although we do not theoretically prove that our solutions are of the lowest possible degrees, we believe so because of the following argument. The best existing solutions for pose estimation using three" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "21316" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 144 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 144 + ], + "type": "text", + "content": "points and three lines use 4th and 8th degree solutions respectively. Since mixed cases are in the middle, our solutions for (2 points, 1 line) and (1 point, 2 lines) cases use 4th and 8th degree solutions respectively. Recently, it was shown using Galois theory that the solutions that use the lowest possible degrees are the optimal ones [[20]]." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 146, + 288, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 146, + 288, + 277 + ], + "spans": [ + { + "bbox": [ + 46, + 146, + 288, + 277 + ], + "type": "text", + "content": "Contrary to the informal reasoning presented above, we claim that the existing solutions to the mixed point-line cases are not optimal. As far as we know, the only support for this claim appearing in the literature prior to our work occurs in [9, §3]. This previous work showed, on the basis of Galois group computation, that the P2P1L and P1P2L problems decompose into simpler subproblems. However, this theoretical observation was not accompanied by a practical solution method for either problem. In this paper, we rectify the situation by devising practical solvers for both problems that incorporate these recent insights." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 279, + 288, + 615 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 279, + 288, + 615 + ], + "spans": [ + { + "bbox": [ + 46, + 279, + 288, + 615 + ], + "type": "text", + "content": "To complete our discussion of what makes a solution algebraically optimal, we recall one proposal of such a notion from work of Nister et al. [20]. In this work, a restricted class of algorithms " + }, + { + "bbox": [ + 46, + 279, + 288, + 615 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_n" + }, + { + "bbox": [ + 46, + 279, + 288, + 615 + ], + "type": "text", + "content": " is considered for each natural number " + }, + { + "bbox": [ + 46, + 279, + 288, + 615 + ], + "type": "inline_equation", + "content": "n \\geq 1" + }, + { + "bbox": [ + 46, + 279, + 288, + 615 + ], + "type": "text", + "content": ": an algorithm in " + }, + { + "bbox": [ + 46, + 279, + 288, + 615 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_n" + }, + { + "bbox": [ + 46, + 279, + 288, + 615 + ], + "type": "text", + "content": " consists of a finite sequence of steps, each of which extracts the roots of some polynomial equation " + }, + { + "bbox": [ + 46, + 279, + 288, + 615 + ], + "type": "inline_equation", + "content": "p(x) = 0" + }, + { + "bbox": [ + 46, + 279, + 288, + 615 + ], + "type": "text", + "content": ", with either " + }, + { + "bbox": [ + 46, + 279, + 288, + 615 + ], + "type": "inline_equation", + "content": "p(x) = x^m - a" + }, + { + "bbox": [ + 46, + 279, + 288, + 615 + ], + "type": "text", + "content": " (where " + }, + { + "bbox": [ + 46, + 279, + 288, + 615 + ], + "type": "inline_equation", + "content": "a \\in \\mathbb{Q}" + }, + { + "bbox": [ + 46, + 279, + 288, + 615 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 279, + 288, + 615 + ], + "type": "inline_equation", + "content": "\\deg(p) = m" + }, + { + "bbox": [ + 46, + 279, + 288, + 615 + ], + "type": "text", + "content": " can be arbitrary), or with " + }, + { + "bbox": [ + 46, + 279, + 288, + 615 + ], + "type": "inline_equation", + "content": "\\deg(p) \\leq n" + }, + { + "bbox": [ + 46, + 279, + 288, + 615 + ], + "type": "text", + "content": " and the coefficients of " + }, + { + "bbox": [ + 46, + 279, + 288, + 615 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 46, + 279, + 288, + 615 + ], + "type": "text", + "content": " belonging to a field containing " + }, + { + "bbox": [ + 46, + 279, + 288, + 615 + ], + "type": "inline_equation", + "content": "\\mathbb{Q}" + }, + { + "bbox": [ + 46, + 279, + 288, + 615 + ], + "type": "text", + "content": " and any previously-computed roots. In this setting, a solution is optimal if it leads to solving a polynomial system of the lowest possible degree. Our proposed solutions immediately establish that P2P1L is " + }, + { + "bbox": [ + 46, + 279, + 288, + 615 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_2" + }, + { + "bbox": [ + 46, + 279, + 288, + 615 + ], + "type": "text", + "content": "-solvable and P1P2L is " + }, + { + "bbox": [ + 46, + 279, + 288, + 615 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_4" + }, + { + "bbox": [ + 46, + 279, + 288, + 615 + ], + "type": "text", + "content": "-solvable. It should be of little surprise that P2P1L is not " + }, + { + "bbox": [ + 46, + 279, + 288, + 615 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_1" + }, + { + "bbox": [ + 46, + 279, + 288, + 615 + ], + "type": "text", + "content": "-solvable; therefore, we may say that our solution to P2P1L is algebraically optimal. On the other hand, the problem P1P2L is also " + }, + { + "bbox": [ + 46, + 279, + 288, + 615 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_3" + }, + { + "bbox": [ + 46, + 279, + 288, + 615 + ], + "type": "text", + "content": "-solvable. This is because any quartic equation can be solved by the standard method which reduces the problem to computing the roots of the associated resolvent cubic. Thus, our solution, which computes the roots of a quartic with this same method, is also algebraically optimal in the sense of [20]. The same observation, of course, holds for the classical quartic-based methods for solving P3P. Alternative P3P solvers that directly employ a cubic [8, 23], despite being superior in practical terms, are not distinguished by the complexity classes " + }, + { + "bbox": [ + 46, + 279, + 288, + 615 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_n" + }, + { + "bbox": [ + 46, + 279, + 288, + 615 + ], + "type": "text", + "content": ". Table 1 provides a summary of the algebraic complexities, based on the Galois groups computed in [9, §3]." + } + ] + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 76, + 628, + 260, + 654 + ], + "blocks": [ + { + "bbox": [ + 76, + 628, + 260, + 654 + ], + "lines": [ + { + "bbox": [ + 76, + 628, + 260, + 654 + ], + "spans": [ + { + "bbox": [ + 76, + 628, + 260, + 654 + ], + "type": "table", + "html": "
problemP3PP2P1LP1P2LP3L
classP3P2P3P8
", + "image_path": "03b944cedaa670d68124a363866ae8e74ed32c07c8bed66f2db52374c0e03c99.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 662, + 287, + 696 + ], + "lines": [ + { + "bbox": [ + 46, + 662, + 287, + 696 + ], + "spans": [ + { + "bbox": [ + 46, + 662, + 287, + 696 + ], + "type": "text", + "content": "Table 1. For each minimal absolute pose problem with points and line features, the class of polynomial root finding algorithms " + }, + { + "bbox": [ + 46, + 662, + 287, + 696 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_n" + }, + { + "bbox": [ + 46, + 662, + 287, + 696 + ], + "type": "text", + "content": " that solve the problem with " + }, + { + "bbox": [ + 46, + 662, + 287, + 696 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 46, + 662, + 287, + 696 + ], + "type": "text", + "content": " as small as possible [20, §2.1]." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 307, + 72, + 391, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 72, + 391, + 83 + ], + "spans": [ + { + "bbox": [ + 307, + 72, + 391, + 83 + ], + "type": "text", + "content": "1.3. Related work" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 91, + 545, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 91, + 545, + 163 + ], + "spans": [ + { + "bbox": [ + 304, + 91, + 545, + 163 + ], + "type": "text", + "content": "The first solutions to P2P1L and P1P2L were presented by Ramalingam et al. [25], reducing the problems to computing the roots of polynomials of degree 4 and 8 via careful choices of special reference frames in the world and image. Although we obtain polynomials of lower degree, these special reference frames are also an ingredient in our approach." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 163, + 546, + 294 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 163, + 546, + 294 + ], + "spans": [ + { + "bbox": [ + 304, + 163, + 546, + 294 + ], + "type": "text", + "content": "In work subsequent to [25], it was observed that both of these problems could be solved using the E3Q3 solver [17]. This is a highly optimized method for computing the points where three quadric surfaces in " + }, + { + "bbox": [ + 304, + 163, + 546, + 294 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^3" + }, + { + "bbox": [ + 304, + 163, + 546, + 294 + ], + "type": "text", + "content": " intersect. Typically, there are " + }, + { + "bbox": [ + 304, + 163, + 546, + 294 + ], + "type": "inline_equation", + "content": "2^{3} = 8" + }, + { + "bbox": [ + 304, + 163, + 546, + 294 + ], + "type": "text", + "content": " such points, by Bézout's theorem; however, EQ3Q also handles degenerate cases where the number of solutions may drop. In the follow-up work [30], a stabilization scheme for E3Q3 was applied to these problems, and experimentally shown to be more accurate than the solvers in [25]. Efficient solvers for both problems based on this stabilized E3Q3 are implemented in PoseLib [18]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 294, + 547, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 294, + 547, + 415 + ], + "spans": [ + { + "bbox": [ + 304, + 294, + 547, + 415 + ], + "type": "text", + "content": "The idea of using Galois groups to study minimal problems originates from [20]. The main takeaway from this paper is that a problem with " + }, + { + "bbox": [ + 304, + 294, + 547, + 415 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 304, + 294, + 547, + 415 + ], + "type": "text", + "content": " solutions whose Galois group is the full symmetric group " + }, + { + "bbox": [ + 304, + 294, + 547, + 415 + ], + "type": "inline_equation", + "content": "S_{n}" + }, + { + "bbox": [ + 304, + 294, + 547, + 415 + ], + "type": "text", + "content": " cannot be solved by an algorithm in " + }, + { + "bbox": [ + 304, + 294, + 547, + 415 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_{n-1}" + }, + { + "bbox": [ + 304, + 294, + 547, + 415 + ], + "type": "text", + "content": ". In the other direction, the main takeaway from [9] is that if the Galois group is contained in the wreath product " + }, + { + "bbox": [ + 304, + 294, + 547, + 415 + ], + "type": "inline_equation", + "content": "S_{n_{1}} \\wr S_{n_{2}}" + }, + { + "bbox": [ + 304, + 294, + 547, + 415 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 294, + 547, + 415 + ], + "type": "inline_equation", + "content": "n = n_{1}n_{2}" + }, + { + "bbox": [ + 304, + 294, + 547, + 415 + ], + "type": "text", + "content": ", then the problem can be solved with an algorithm in " + }, + { + "bbox": [ + 304, + 294, + 547, + 415 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_{\\max(n_{1},n_{2})}" + }, + { + "bbox": [ + 304, + 294, + 547, + 415 + ], + "type": "text", + "content": ". A recent work using such an insight to guide the more efficient solution of a relative pose estimation problem may be found in [14]." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 426, + 404, + 438 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 426, + 404, + 438 + ], + "spans": [ + { + "bbox": [ + 306, + 426, + 404, + 438 + ], + "type": "text", + "content": "2. Minimal solvers" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 447, + 545, + 615 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 447, + 545, + 615 + ], + "spans": [ + { + "bbox": [ + 304, + 447, + 545, + 615 + ], + "type": "text", + "content": "In this section, we introduce our algebraically optimal solutions to the P2P1L and P1P2L problems. These problems are depicted in Figure 1. As revealed by the Galois group computed in [9, §3], the problem P2P1L can be reduced to computing the roots of a quadratic equation, and the problem P1P2L can be reduced to a quartic equation. In Sections 2.1 and 2.2, we turn these insights into explicit solutions to the P2P1L and P12PL problems, respectively. These solutions work generically—they are valid outside of a measure-zero subset of the space of input point-point/line-line correspondences. In Section 2.3, we describe a method that stabilizes the P1P2L solver of Section 2.2 in a common but non-generic case. Finally, Section 2.4 for a discussion when the point-line configuration is coplanar." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 622, + 361, + 634 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 622, + 361, + 634 + ], + "spans": [ + { + "bbox": [ + 306, + 622, + 361, + 634 + ], + "type": "text", + "content": "2.1. P2P1L" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 641, + 545, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 641, + 545, + 665 + ], + "spans": [ + { + "bbox": [ + 304, + 641, + 545, + 665 + ], + "type": "text", + "content": "Here, we provide the algebraically optimal solution to the P2P1L problem. We mostly follow the notation of [25]." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 665, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 545, + 714 + ], + "type": "text", + "content": "Problem Statement: Consider " + }, + { + "bbox": [ + 304, + 665, + 545, + 714 + ], + "type": "inline_equation", + "content": "P_1^0, P_2^0 \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 304, + 665, + 545, + 714 + ], + "type": "text", + "content": ", two 3D points, and their 2D projections under an unknown calibrated camera. Consider also a 3D line spanned by points " + }, + { + "bbox": [ + 304, + 665, + 545, + 714 + ], + "type": "inline_equation", + "content": "L_3^0, L_4^0 \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 304, + 665, + 545, + 714 + ], + "type": "text", + "content": " and its projection. The goal is to recover the" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "21317" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 66, + 72, + 160, + 159 + ], + "blocks": [ + { + "bbox": [ + 66, + 72, + 160, + 159 + ], + "lines": [ + { + "bbox": [ + 66, + 72, + 160, + 159 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 160, + 159 + ], + "type": "image", + "image_path": "20445638a11240eeb032dbf9aa20b68a3f5e0ff62a2dfbc3eea5f7ee8d5f1d57.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 175, + 72, + 263, + 159 + ], + "blocks": [ + { + "bbox": [ + 175, + 72, + 263, + 159 + ], + "lines": [ + { + "bbox": [ + 175, + 72, + 263, + 159 + ], + "spans": [ + { + "bbox": [ + 175, + 72, + 263, + 159 + ], + "type": "image", + "image_path": "43c16aa4a49f8858a75473230dd5eaa7300de3b17ad8be6e8c7c274b0f96d3ad.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 67, + 160, + 156, + 231 + ], + "blocks": [ + { + "bbox": [ + 67, + 160, + 156, + 231 + ], + "lines": [ + { + "bbox": [ + 67, + 160, + 156, + 231 + ], + "spans": [ + { + "bbox": [ + 67, + 160, + 156, + 231 + ], + "type": "image", + "image_path": "7ab54c24980c7d8f3dfccb159e7c7f38c5184844ac9305c9b364214ab8856d71.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 47, + 237, + 286, + 258 + ], + "lines": [ + { + "bbox": [ + 47, + 237, + 286, + 258 + ], + "spans": [ + { + "bbox": [ + 47, + 237, + 286, + 258 + ], + "type": "text", + "content": "Figure 2. Special reference frame for P2P1L. Camera frame " + }, + { + "bbox": [ + 47, + 237, + 286, + 258 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_0" + }, + { + "bbox": [ + 47, + 237, + 286, + 258 + ], + "type": "text", + "content": " is transformed to " + }, + { + "bbox": [ + 47, + 237, + 286, + 258 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_1" + }, + { + "bbox": [ + 47, + 237, + 286, + 258 + ], + "type": "text", + "content": ", world frame " + }, + { + "bbox": [ + 47, + 237, + 286, + 258 + ], + "type": "inline_equation", + "content": "\\mathcal{W}_0" + }, + { + "bbox": [ + 47, + 237, + 286, + 258 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 47, + 237, + 286, + 258 + ], + "type": "inline_equation", + "content": "\\mathcal{W}_1" + }, + { + "bbox": [ + 47, + 237, + 286, + 258 + ], + "type": "text", + "content": ". See text for details." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 176, + 160, + 261, + 230 + ], + "blocks": [ + { + "bbox": [ + 176, + 160, + 261, + 230 + ], + "lines": [ + { + "bbox": [ + 176, + 160, + 261, + 230 + ], + "spans": [ + { + "bbox": [ + 176, + 160, + 261, + 230 + ], + "type": "image", + "image_path": "e709c755edc91e6f86341c74511e1079caf27c378108579d7dc7e3a31ed5ea1e.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 279, + 277, + 292 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 279, + 277, + 292 + ], + "spans": [ + { + "bbox": [ + 47, + 279, + 277, + 292 + ], + "type": "text", + "content": "unknown camera matrix, whose center we denote by " + }, + { + "bbox": [ + 47, + 279, + 277, + 292 + ], + "type": "inline_equation", + "content": "C^0" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 292, + 287, + 363 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 292, + 287, + 363 + ], + "spans": [ + { + "bbox": [ + 47, + 292, + 287, + 363 + ], + "type": "text", + "content": "Our first step is to rigidly transform the input into the special reference frame introduced in [25]. Transforming to this special frame simplifies the equations and helps to reveal the algebraically optimal solution. Figure 2 illustrates the input reference frame " + }, + { + "bbox": [ + 47, + 292, + 287, + 363 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_0 / \\mathcal{W}_0" + }, + { + "bbox": [ + 47, + 292, + 287, + 363 + ], + "type": "text", + "content": " (top), and the special frame " + }, + { + "bbox": [ + 47, + 292, + 287, + 363 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_1 / \\mathcal{W}_1" + }, + { + "bbox": [ + 47, + 292, + 287, + 363 + ], + "type": "text", + "content": " (bottom), which are related as follows:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 364, + 264, + 375 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 364, + 264, + 375 + ], + "spans": [ + { + "bbox": [ + 47, + 364, + 264, + 375 + ], + "type": "text", + "content": "- In the world frame " + }, + { + "bbox": [ + 47, + 364, + 264, + 375 + ], + "type": "inline_equation", + "content": "\\mathcal{W}_1" + }, + { + "bbox": [ + 47, + 364, + 264, + 375 + ], + "type": "text", + "content": ", the 3D points take the form" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 81, + 383, + 287, + 401 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 81, + 383, + 287, + 401 + ], + "spans": [ + { + "bbox": [ + 81, + 383, + 287, + 401 + ], + "type": "interline_equation", + "content": "P _ {1} = \\left( \\begin{array}{l l l} 0 & 0 & 0 \\end{array} \\right) ^ {T}, \\quad P _ {2} = \\left( \\begin{array}{l l l} X _ {2} & 0 & 0 \\end{array} \\right) ^ {T}, \\tag {1}", + "image_path": "89c869a636904322e2c3ba7a8ab998ec6083679ed9422bc2ba05cf9eaaa13513.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 409, + 269, + 421 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 409, + 269, + 421 + ], + "spans": [ + { + "bbox": [ + 55, + 409, + 269, + 421 + ], + "type": "text", + "content": "and the 3D line is spanned by two points of the form" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 63, + 428, + 287, + 445 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 428, + 287, + 445 + ], + "spans": [ + { + "bbox": [ + 63, + 428, + 287, + 445 + ], + "type": "interline_equation", + "content": "L _ {3} = \\left( \\begin{array}{l l l} X _ {3} & Y _ {3} & 0 \\end{array} \\right) ^ {T}, \\quad L _ {4} = \\left( \\begin{array}{l l l} X _ {4} & Y _ {4} & Z _ {4} \\end{array} \\right) ^ {T}. \\tag {2}", + "image_path": "4c17d8fb09de3750fe3f983872223d4d9b25905f57d95a6aeb8fc79b82af0119.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 453, + 241, + 465 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 453, + 241, + 465 + ], + "spans": [ + { + "bbox": [ + 47, + 453, + 241, + 465 + ], + "type": "text", + "content": "To find the transformation " + }, + { + "bbox": [ + 47, + 453, + 241, + 465 + ], + "type": "inline_equation", + "content": "\\mathcal{W}_0\\to \\mathcal{W}_1" + }, + { + "bbox": [ + 47, + 453, + 241, + 465 + ], + "type": "text", + "content": " , we set" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 64, + 472, + 277, + 490 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 472, + 277, + 490 + ], + "spans": [ + { + "bbox": [ + 64, + 472, + 277, + 490 + ], + "type": "interline_equation", + "content": "P _ {1} = \\left( \\begin{array}{c c c} 0 & 0 & 0 \\end{array} \\right) ^ {T}, \\quad P _ {2} = \\left(\\| P _ {1} ^ {0} - P _ {2} ^ {2} \\| \\quad 0 \\quad 0\\right) ^ {T},", + "image_path": "f3373300ad506b3003ebac68ae3c2766bf08112c6f570b2e2336ba9b1dd52439.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 497, + 287, + 553 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 497, + 287, + 553 + ], + "spans": [ + { + "bbox": [ + 56, + 497, + 287, + 553 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} L _ {3} = \\left( \\begin{array}{c c c} X _ {3} & Y _ {3} & 0 \\end{array} \\right) ^ {T}, \\mathrm {w /} X _ {3} = \\left(L _ {3} ^ {0} - P _ {1}\\right) ^ {T} \\frac {P _ {2} - P _ {1}}{\\| P _ {2} - P _ {1} \\|}, \\\\ Y _ {3} = \\left\\| L _ {3} ^ {0} - P _ {1} - X _ {3} \\cdot \\frac {P _ {2} - P _ {1}}{\\| P _ {2} - P _ {1} \\|} \\right\\|. \\\\ \\end{array}", + "image_path": "0d775f28d3721086114ad4a48459a3a8b7ae36cc903c5fcf47a3c9607114606b.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 55, + 557, + 287, + 587 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 557, + 287, + 587 + ], + "spans": [ + { + "bbox": [ + 55, + 557, + 287, + 587 + ], + "type": "text", + "content": "We transform the rays " + }, + { + "bbox": [ + 55, + 557, + 287, + 587 + ], + "type": "inline_equation", + "content": "\\overrightarrow{P_1^0P_2^0}" + }, + { + "bbox": [ + 55, + 557, + 287, + 587 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 557, + 287, + 587 + ], + "type": "inline_equation", + "content": "\\overrightarrow{P_1^0L_3^0}" + }, + { + "bbox": [ + 55, + 557, + 287, + 587 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 55, + 557, + 287, + 587 + ], + "type": "inline_equation", + "content": "\\overrightarrow{P_1P_2}" + }, + { + "bbox": [ + 55, + 557, + 287, + 587 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 557, + 287, + 587 + ], + "type": "inline_equation", + "content": "\\overrightarrow{P_1L_3}" + }, + { + "bbox": [ + 55, + 557, + 287, + 587 + ], + "type": "text", + "content": ", respectively, via suitable translation and rotation." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 47, + 588, + 255, + 598 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 588, + 255, + 598 + ], + "spans": [ + { + "bbox": [ + 47, + 588, + 255, + 598 + ], + "type": "text", + "content": "- In the camera frame, the camera center is fixed at" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 127, + 606, + 287, + 623 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 127, + 606, + 287, + 623 + ], + "spans": [ + { + "bbox": [ + 127, + 606, + 287, + 623 + ], + "type": "interline_equation", + "content": "C = \\left( \\begin{array}{l l l} 0 & 0 & - 1 \\end{array} \\right) ^ {T}, \\tag {3}", + "image_path": "1fb087a8a412668c42b2360f40d11ed035ba4b2560be69a76aa25968535b0bc1.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 55, + 632, + 259, + 643 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 632, + 259, + 643 + ], + "spans": [ + { + "bbox": [ + 55, + 632, + 259, + 643 + ], + "type": "text", + "content": "the 2D points are projections of points of the form" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 69, + 651, + 287, + 668 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 651, + 287, + 668 + ], + "spans": [ + { + "bbox": [ + 69, + 651, + 287, + 668 + ], + "type": "interline_equation", + "content": "D _ {1} = \\left( \\begin{array}{l l l} a _ {1} & b _ {1} & 0 \\end{array} \\right) ^ {T}, \\quad D _ {2} = \\left( \\begin{array}{l l l} a _ {2} & b _ {2} & 0 \\end{array} \\right) ^ {T}, \\tag {4}", + "image_path": "563299138e6444a42f6986a6d30a4fe0b78d3582352b3c4596b0c35b75ee0c9f.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 55, + 677, + 255, + 690 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 255, + 690 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 255, + 690 + ], + "type": "text", + "content": "and the 2D line is the projection of " + }, + { + "bbox": [ + 55, + 677, + 255, + 690 + ], + "type": "inline_equation", + "content": "\\overline{D_3D_4}" + }, + { + "bbox": [ + 55, + 677, + 255, + 690 + ], + "type": "text", + "content": ", where" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 83, + 698, + 287, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 83, + 698, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 83, + 698, + 287, + 715 + ], + "type": "interline_equation", + "content": "D _ {3} = \\left( \\begin{array}{l l l} 0 & 0 & 0 \\end{array} \\right) ^ {T}, \\quad D _ {4} = \\left( \\begin{array}{l l l} 1 & 0 & 0 \\end{array} \\right) ^ {T}. \\tag {5}", + "image_path": "fe73678c3b94e747d4e25f7da0c27e2151e904d5e2a77b7a9c069abd3061e145.jpg" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 306, + 72, + 545, + 137 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 545, + 137 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 545, + 137 + ], + "type": "text", + "content": "- To find the transformation " + }, + { + "bbox": [ + 306, + 72, + 545, + 137 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_0\\to \\mathcal{C}_1" + }, + { + "bbox": [ + 306, + 72, + 545, + 137 + ], + "type": "text", + "content": " , let " + }, + { + "bbox": [ + 306, + 72, + 545, + 137 + ], + "type": "inline_equation", + "content": "D_3^0" + }, + { + "bbox": [ + 306, + 72, + 545, + 137 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 306, + 72, + 545, + 137 + ], + "type": "inline_equation", + "content": "D_4^0" + }, + { + "bbox": [ + 306, + 72, + 545, + 137 + ], + "type": "text", + "content": " be the homogeneous coordinates of two distinct points along the given line. Independently of " + }, + { + "bbox": [ + 306, + 72, + 545, + 137 + ], + "type": "inline_equation", + "content": "C^0" + }, + { + "bbox": [ + 306, + 72, + 545, + 137 + ], + "type": "text", + "content": " , we may rigidly transform the rays " + }, + { + "bbox": [ + 306, + 72, + 545, + 137 + ], + "type": "inline_equation", + "content": "d_{3}\\coloneqq \\overrightarrow{C^{0}D_{3}^{0}}" + }, + { + "bbox": [ + 306, + 72, + 545, + 137 + ], + "type": "text", + "content": " , and " + }, + { + "bbox": [ + 306, + 72, + 545, + 137 + ], + "type": "inline_equation", + "content": "d_4\\coloneqq \\overrightarrow{C^0D_3^0}" + }, + { + "bbox": [ + 306, + 72, + 545, + 137 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 306, + 72, + 545, + 137 + ], + "type": "inline_equation", + "content": "\\overrightarrow{CD_3}" + }, + { + "bbox": [ + 306, + 72, + 545, + 137 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 306, + 72, + 545, + 137 + ], + "type": "inline_equation", + "content": "\\overrightarrow{CD_4}" + }, + { + "bbox": [ + 306, + 72, + 545, + 137 + ], + "type": "text", + "content": " , by suitable rotation and translation." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 305, + 137, + 545, + 173 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 137, + 545, + 173 + ], + "spans": [ + { + "bbox": [ + 305, + 137, + 545, + 173 + ], + "type": "text", + "content": "Let us now write " + }, + { + "bbox": [ + 305, + 137, + 545, + 173 + ], + "type": "inline_equation", + "content": "R = (R_{ij})_{1 \\leq i,j \\leq 3}" + }, + { + "bbox": [ + 305, + 137, + 545, + 173 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 305, + 137, + 545, + 173 + ], + "type": "inline_equation", + "content": "T = (T_i)_{1 \\leq i \\leq 3}" + }, + { + "bbox": [ + 305, + 137, + 545, + 173 + ], + "type": "text", + "content": " for the unknown camera pose in this special reference frame. Our projection constraints can then be written as" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 339, + 183, + 545, + 196 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 339, + 183, + 545, + 196 + ], + "spans": [ + { + "bbox": [ + 339, + 183, + 545, + 196 + ], + "type": "interline_equation", + "content": "\\left(D _ {i} - C\\right) \\times \\left(R P _ {i} + T\\right) = 0, \\quad i \\in \\{1, 2 \\}, \\tag {6}", + "image_path": "b515c9649a164e186924c0d6058fbd120db4a12d7ed1dcea9d6360f5a12d75bb.jpg" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 339, + 198, + 545, + 213 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 339, + 198, + 545, + 213 + ], + "spans": [ + { + "bbox": [ + 339, + 198, + 545, + 213 + ], + "type": "interline_equation", + "content": "\\left(D _ {3} \\times D _ {4}\\right) ^ {T} \\left(R L _ {i} + T\\right) = 0, \\quad i \\in \\{3, 4 \\}. \\tag {7}", + "image_path": "bafd580663476afd39e797472fb8ff432993daf655862846e631c13d15734754.jpg" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 305, + 222, + 545, + 258 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 222, + 545, + 258 + ], + "spans": [ + { + "bbox": [ + 305, + 222, + 545, + 258 + ], + "type": "text", + "content": "This gives a system of equations in 12 unknowns. However, in what follows, we consider a system of equations in a smaller set of unknowns, namely" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 313, + 266, + 545, + 292 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 266, + 545, + 292 + ], + "spans": [ + { + "bbox": [ + 313, + 266, + 545, + 292 + ], + "type": "interline_equation", + "content": "\\mathbf {x} ^ {T} = \\left( \\begin{array}{l l l l l l l l} R _ {1 1} & R _ {2 1} & R _ {3 1} & R _ {2 2} & R _ {2 3} & T _ {1} & T _ {2} & T _ {3} \\end{array} \\right). \\tag {8}", + "image_path": "c304a18dbae25f2c80993c4788aa14fb8d3eae76cab1a1495a9da8dfd70d9b92.jpg" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 306, + 293, + 449, + 304 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 293, + 449, + 304 + ], + "spans": [ + { + "bbox": [ + 306, + 293, + 449, + 304 + ], + "type": "text", + "content": "The entries of " + }, + { + "bbox": [ + 306, + 293, + 449, + 304 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 306, + 293, + 449, + 304 + ], + "type": "text", + "content": " are constrained by" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 334, + 313, + 545, + 328 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 334, + 313, + 545, + 328 + ], + "spans": [ + { + "bbox": [ + 334, + 313, + 545, + 328 + ], + "type": "interline_equation", + "content": "R _ {1 1} ^ {2} + R _ {2 1} ^ {2} + R _ {3 1} ^ {2} = R _ {2 1} ^ {2} + R _ {2 2} ^ {2} + R _ {2 3} ^ {2} = 1. \\tag {9}", + "image_path": "c9d53d4898744e1a6f6f0b7fa32a54e258988a3962440ae299d411786a40b4b0.jpg" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 305, + 337, + 545, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 337, + 545, + 373 + ], + "spans": [ + { + "bbox": [ + 305, + 337, + 545, + 373 + ], + "type": "text", + "content": "Thus, the entries of " + }, + { + "bbox": [ + 305, + 337, + 545, + 373 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 305, + 337, + 545, + 373 + ], + "type": "text", + "content": " specify a translation vector and a partially-filled rotation matrix, which can be uniquely completed to a rotation matrix using the formulae" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 381, + 546, + 444 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 381, + 546, + 444 + ], + "spans": [ + { + "bbox": [ + 307, + 381, + 546, + 444 + ], + "type": "interline_equation", + "content": "\\binom {R _ {1 2}} {R _ {1 3}} = \\left(R _ {2 2} ^ {2} + R _ {2 3} ^ {2}\\right) ^ {- 1} \\cdot \\binom {- R _ {1 1} R _ {2 1} R _ {2 2} + R _ {2 3} R _ {3 1}} {- R _ {1 1} R _ {2 1} R _ {2 3} - R _ {2 2} R _ {3 1}} \\\\ - R _ {2 1} R _ {2 2} R _ {3 1} - R _ {1 1} R _ {2 3} \\\\ - R _ {2 1} R _ {2 3} R _ {3 1} + R _ {1 1} R _ {2 2}) \\tag {10}", + "image_path": "226450659272033658c188579a7a3c58bc655c0138c40d1cbf90c934ba69db5a.jpg" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 305, + 454, + 545, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 454, + 545, + 515 + ], + "spans": [ + { + "bbox": [ + 305, + 454, + 545, + 515 + ], + "type": "text", + "content": "subject to the genericity condition " + }, + { + "bbox": [ + 305, + 454, + 545, + 515 + ], + "type": "inline_equation", + "content": "R_{22}^2 + R_{23}^2 \\neq 0" + }, + { + "bbox": [ + 305, + 454, + 545, + 515 + ], + "type": "text", + "content": ", i.e. when " + }, + { + "bbox": [ + 305, + 454, + 545, + 515 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 305, + 454, + 545, + 515 + ], + "type": "text", + "content": " is not a rotation in the " + }, + { + "bbox": [ + 305, + 454, + 545, + 515 + ], + "type": "inline_equation", + "content": "xz" + }, + { + "bbox": [ + 305, + 454, + 545, + 515 + ], + "type": "text", + "content": "-plane. Hence, we focus on recovering the entries of " + }, + { + "bbox": [ + 305, + 454, + 545, + 515 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 305, + 454, + 545, + 515 + ], + "type": "text", + "content": ". From 2 out of the 3 redundant constraints from (6) together with (7), we obtain linear constraints on " + }, + { + "bbox": [ + 305, + 454, + 545, + 515 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 305, + 454, + 545, + 515 + ], + "type": "text", + "content": " ([25, eq. (4)-(5)])," + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 335, + 525, + 410, + 536 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 525, + 410, + 536 + ], + "spans": [ + { + "bbox": [ + 335, + 525, + 410, + 536 + ], + "type": "interline_equation", + "content": "A \\mathbf {x} = \\mathbf {b}, \\quad \\text {w h e r e}", + "image_path": "a0e3bc4597959ced927a87c75677bc8a6033a4a6d33ba14cd100cd5d91b9a4a5.jpg" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 341, + 540, + 517, + 586 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 341, + 540, + 517, + 586 + ], + "spans": [ + { + "bbox": [ + 341, + 540, + 517, + 586 + ], + "type": "interline_equation", + "content": "A = \\left( \\begin{array}{c c c c c c c c} 0 & 0 & 0 & 0 & 0 & - b _ {1} & a _ {1} & 0 \\\\ 0 & 0 & 0 & 0 & 0 & 0 & - 1 & b _ {1} \\\\ - b _ {2} X _ {2} & a _ {2} X _ {2} & 0 & 0 & 0 & - b _ {2} & a _ {2} & 0 \\\\ 0 & - X _ {2} & b _ {2} X _ {2} & 0 & 0 & 0 & - 1 & b _ {2} \\\\ 0 & X _ {3} & 0 & Y _ {3} & 0 & 0 & 1 & 0 \\\\ 0 & X _ {4} & 0 & Y _ {4} & Z _ {4} & 0 & 1 & 0 \\end{array} \\right),", + "image_path": "8221a147d2bfa1c049098d1b89a184c899c0d1a70514d73aab50fe1a26290ada.jpg" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 335, + 586, + 545, + 601 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 586, + 545, + 601 + ], + "spans": [ + { + "bbox": [ + 335, + 586, + 545, + 601 + ], + "type": "interline_equation", + "content": "\\mathbf {b} ^ {T} = \\left( \\begin{array}{c c c c c c} 0 & - b _ {1} & 0 & - b _ {2} & 0 & 0 \\end{array} \\right). \\tag {11}", + "image_path": "491df2a1bcd0d3c31e289518f0257c5553fdf42fdb847122a2b7a51911949537.jpg" + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 306, + 611, + 544, + 634 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 611, + 544, + 634 + ], + "spans": [ + { + "bbox": [ + 306, + 611, + 544, + 634 + ], + "type": "text", + "content": "We may use these linear equations to solve for the translation " + }, + { + "bbox": [ + 306, + 611, + 544, + 634 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 306, + 611, + 544, + 634 + ], + "type": "text", + "content": " in terms of " + }, + { + "bbox": [ + 306, + 611, + 544, + 634 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 306, + 611, + 544, + 634 + ], + "type": "text", + "content": " and the problem data as follows:" + } + ] + } + ], + "index": 35 + }, + { + "bbox": [ + 320, + 644, + 530, + 657 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 644, + 530, + 657 + ], + "spans": [ + { + "bbox": [ + 320, + 644, + 530, + 657 + ], + "type": "interline_equation", + "content": "T _ {1} = a _ {1} B, \\quad T _ {2} = b _ {1} B, \\quad T _ {3} = - 1 + B, \\quad \\text {w h e r e}", + "image_path": "24ab172d3bf4160266785fc88c7728433e12ddc71e4193bc70781fcc4167ea55.jpg" + } + ] + } + ], + "index": 36 + }, + { + "bbox": [ + 324, + 660, + 545, + 685 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 324, + 660, + 545, + 685 + ], + "spans": [ + { + "bbox": [ + 324, + 660, + 545, + 685 + ], + "type": "interline_equation", + "content": "B = \\frac {X _ {2} \\left(a _ {2} R _ {2 1} - b _ {2} R _ {1 1}\\right)}{b _ {2} a _ {1} - b _ {1} a _ {2}}. \\tag {12}", + "image_path": "3e63ecd22ea0f6315342f401ed6e739308f96a009faa8d1b9ccd9d39df3d8ae6.jpg" + } + ] + } + ], + "index": 37 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 306, + 693, + 544, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 693, + 544, + 712 + ], + "spans": [ + { + "bbox": [ + 306, + 693, + 544, + 712 + ], + "type": "text", + "content": "1If " + }, + { + "bbox": [ + 306, + 693, + 544, + 712 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 306, + 693, + 544, + 712 + ], + "type": "text", + "content": " is known to be a plane rotation, two generic point-point correspondences suffice to recover " + }, + { + "bbox": [ + 306, + 693, + 544, + 712 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 306, + 693, + 544, + 712 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 306, + 693, + 544, + 712 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 306, + 693, + 544, + 712 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 38 + }, + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "21318" + } + ] + } + ], + "index": 39 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 43, + 69, + 287, + 164 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 43, + 69, + 286, + 93 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 43, + 69, + 286, + 93 + ], + "spans": [ + { + "bbox": [ + 43, + 69, + 286, + 93 + ], + "type": "text", + "content": "(1) Transform data in world and camera frames to the special reference frames satisfying (1), (2), (3), (4), (5)." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 43, + 93, + 272, + 105 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 43, + 93, + 272, + 105 + ], + "spans": [ + { + "bbox": [ + 43, + 93, + 272, + 105 + ], + "type": "text", + "content": "(2) Compute up to 2 solutions in " + }, + { + "bbox": [ + 43, + 93, + 272, + 105 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 43, + 93, + 272, + 105 + ], + "type": "text", + "content": " to the quadratic (16)." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 43, + 105, + 287, + 129 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 43, + 105, + 287, + 129 + ], + "spans": [ + { + "bbox": [ + 43, + 105, + 287, + 129 + ], + "type": "text", + "content": "(3) For each root " + }, + { + "bbox": [ + 43, + 105, + 287, + 129 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 43, + 105, + 287, + 129 + ], + "type": "text", + "content": " in step (2), recover a corresponding value of " + }, + { + "bbox": [ + 43, + 105, + 287, + 129 + ], + "type": "inline_equation", + "content": "u" + }, + { + "bbox": [ + 43, + 105, + 287, + 129 + ], + "type": "text", + "content": " using either of the linear expressions in (46)." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 43, + 129, + 269, + 141 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 43, + 129, + 269, + 141 + ], + "spans": [ + { + "bbox": [ + 43, + 129, + 269, + 141 + ], + "type": "text", + "content": "(4) Recover up to 4 solutions in " + }, + { + "bbox": [ + 43, + 129, + 269, + 141 + ], + "type": "inline_equation", + "content": "\\{R_{11}, R_{21}\\}" + }, + { + "bbox": [ + 43, + 129, + 269, + 141 + ], + "type": "text", + "content": " using (15)." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 44, + 141, + 222, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 44, + 141, + 222, + 152 + ], + "spans": [ + { + "bbox": [ + 44, + 141, + 222, + 152 + ], + "type": "text", + "content": "(5) Recover " + }, + { + "bbox": [ + 44, + 141, + 222, + 152 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 44, + 141, + 222, + 152 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 44, + 141, + 222, + 152 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 44, + 141, + 222, + 152 + ], + "type": "text", + "content": " using (13), (12), (10)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 43, + 152, + 247, + 164 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 43, + 152, + 247, + 164 + ], + "spans": [ + { + "bbox": [ + 43, + 152, + 247, + 164 + ], + "type": "text", + "content": "(6) Reverse the transformation applied in step (1)." + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 47, + 206, + 287, + 231 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 206, + 287, + 231 + ], + "spans": [ + { + "bbox": [ + 47, + 206, + 287, + 231 + ], + "type": "text", + "content": "Moreover, from (11) we may use " + }, + { + "bbox": [ + 47, + 206, + 287, + 231 + ], + "type": "inline_equation", + "content": "R_{11}" + }, + { + "bbox": [ + 47, + 206, + 287, + 231 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 206, + 287, + 231 + ], + "type": "inline_equation", + "content": "R_{21}" + }, + { + "bbox": [ + 47, + 206, + 287, + 231 + ], + "type": "text", + "content": ", to express the remaining entries of " + }, + { + "bbox": [ + 47, + 206, + 287, + 231 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 47, + 206, + 287, + 231 + ], + "type": "text", + "content": " as" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 235, + 211, + 261 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 235, + 211, + 261 + ], + "spans": [ + { + "bbox": [ + 48, + 235, + 211, + 261 + ], + "type": "interline_equation", + "content": "R _ {3 1} = \\frac {(a _ {1} - a _ {2}) R _ {2 1} + (b _ {2} - b _ {1}) R _ {1 1})}{b _ {2} a _ {1} - b _ {1} a _ {2}},", + "image_path": "6c6618c8e6738d724388d30b4294036e07a33cbe377acca114b4649c5b84763c.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 50, + 262, + 286, + 289 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 50, + 262, + 286, + 289 + ], + "spans": [ + { + "bbox": [ + 50, + 262, + 286, + 289 + ], + "type": "interline_equation", + "content": "R _ {2 2} = \\frac {b _ {1} b _ {2} X _ {2} R _ {1 1} + (b _ {1} a _ {2} X _ {2} - b _ {1} a _ {2} X _ {2} - b _ {2} a _ {1} X _ {3}) R _ {2 1}}{Y _ {3} (b _ {2} a _ {1} - b _ {1} a _ {2})},", + "image_path": "4097b1ddd147bb35e0dbd6aa69e9135eb2b1d0d911f8a708e8d12b173a52cd86.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 294, + 287, + 365 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 294, + 287, + 365 + ], + "spans": [ + { + "bbox": [ + 48, + 294, + 287, + 365 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} R _ {2 3} = \\frac {b _ {1} b _ {2} X _ {2} \\left(Y _ {3} - Y _ {4}\\right) R _ {1 1}}{Y _ {3} Z _ {4} \\left(b _ {2} a _ {1} - b _ {1} a _ {2}\\right)} \\\\ + \\left(\\frac {X _ {3} Y _ {4} - X _ {4} Y _ {3}}{Y _ {3} Z _ {4}} + \\frac {b _ {2} a _ {1} X _ {3} Y _ {4} - b _ {1} a _ {2} X _ {2} Y _ {3}}{Y _ {3} Z _ {4} \\left(b _ {2} a _ {1} - b _ {1} a _ {2}\\right)}\\right) R _ {2 1}. \\tag {13} \\\\ \\end{array}", + "image_path": "de1ec16dcdfed26c0c89ec7b74db14340ba412566acfcc18d475b5aebba761a1.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 47, + 372, + 287, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 372, + 287, + 396 + ], + "spans": [ + { + "bbox": [ + 47, + 372, + 287, + 396 + ], + "type": "text", + "content": "Substituting (13) into (9), we obtain two bivariate quadratic constraints in " + }, + { + "bbox": [ + 47, + 372, + 287, + 396 + ], + "type": "inline_equation", + "content": "R_{11}" + }, + { + "bbox": [ + 47, + 372, + 287, + 396 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 372, + 287, + 396 + ], + "type": "inline_equation", + "content": "R_{21}" + }, + { + "bbox": [ + 47, + 372, + 287, + 396 + ], + "type": "text", + "content": ". In matrix form," + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 87, + 401, + 287, + 440 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 87, + 401, + 287, + 440 + ], + "spans": [ + { + "bbox": [ + 87, + 401, + 287, + 440 + ], + "type": "interline_equation", + "content": "\\left( \\begin{array}{l l l} c _ {1} & c _ {2} & c _ {3} \\\\ d _ {1} & d _ {2} & d _ {3} \\end{array} \\right) \\cdot \\left( \\begin{array}{c} R _ {1 1} ^ {2} \\\\ R _ {1 1} R _ {2 1} \\\\ R _ {2 1} ^ {2} \\end{array} \\right) = \\binom {1} {1}, \\tag {14}", + "image_path": "951bbd87d9213c47f4475bfed09135cfc9a97c530f9f3c4ce279f7c50215d0ce.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 47, + 445, + 287, + 469 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 445, + 287, + 469 + ], + "spans": [ + { + "bbox": [ + 47, + 445, + 287, + 469 + ], + "type": "text", + "content": "where the coefficients " + }, + { + "bbox": [ + 47, + 445, + 287, + 469 + ], + "type": "inline_equation", + "content": "c_{1}, c_{2}, c_{3}, d_{1}, d_{2}, d_{3}" + }, + { + "bbox": [ + 47, + 445, + 287, + 469 + ], + "type": "text", + "content": " are rational functions of the problem data. Applying the change of variables" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 111, + 479, + 287, + 494 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 111, + 479, + 287, + 494 + ], + "spans": [ + { + "bbox": [ + 111, + 479, + 287, + 494 + ], + "type": "interline_equation", + "content": "u = R _ {1 1} ^ {2}, \\quad v = R _ {2 1} / R _ {1 1}, \\tag {15}", + "image_path": "b09528f1215b87bc75928f0fd5838a48752eedc5673fa6d3e472edc5ae89a01c.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 47, + 497, + 258, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 497, + 258, + 509 + ], + "spans": [ + { + "bbox": [ + 47, + 497, + 258, + 509 + ], + "type": "text", + "content": "and subtracting the two equations in (14), we obtain" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 69, + 515, + 264, + 529 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 515, + 264, + 529 + ], + "spans": [ + { + "bbox": [ + 69, + 515, + 264, + 529 + ], + "type": "interline_equation", + "content": "\\left(c _ {1} - d _ {1}\\right) u + \\left(c _ {2} - d _ {2}\\right) u v + \\left(c _ {3} - d _ {3}\\right) u v ^ {2} = 0.", + "image_path": "3a8ac4a051d79a7b5e05257488f11fe68ef959d62c1e02bd965d18f5ad10acf5.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 47, + 536, + 287, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 536, + 287, + 559 + ], + "spans": [ + { + "bbox": [ + 47, + 536, + 287, + 559 + ], + "type": "text", + "content": "Assuming " + }, + { + "bbox": [ + 47, + 536, + 287, + 559 + ], + "type": "inline_equation", + "content": "u \\neq 0" + }, + { + "bbox": [ + 47, + 536, + 287, + 559 + ], + "type": "text", + "content": ", we therefore have the univariate quadratic equation in " + }, + { + "bbox": [ + 47, + 536, + 287, + 559 + ], + "type": "inline_equation", + "content": "v" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 69, + 566, + 287, + 579 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 566, + 287, + 579 + ], + "spans": [ + { + "bbox": [ + 69, + 566, + 287, + 579 + ], + "type": "interline_equation", + "content": "\\left(c _ {1} - d _ {1}\\right) + \\left(c _ {2} - d _ {2}\\right) v + \\left(c _ {3} - d _ {3}\\right) v ^ {2} = 0. \\tag {16}", + "image_path": "1f3c7dccd1cf27cf38921165a7c2800530daa551d36c85ac0a11475c4abd1ae2.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 47, + 586, + 287, + 611 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 586, + 287, + 611 + ], + "spans": [ + { + "bbox": [ + 47, + 586, + 287, + 611 + ], + "type": "text", + "content": "If " + }, + { + "bbox": [ + 47, + 586, + 287, + 611 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 47, + 586, + 287, + 611 + ], + "type": "text", + "content": " is one of the roots of (16), we may recover a corresponding value for " + }, + { + "bbox": [ + 47, + 586, + 287, + 611 + ], + "type": "inline_equation", + "content": "u" + }, + { + "bbox": [ + 47, + 586, + 287, + 611 + ], + "type": "text", + "content": " using one of the equations in (14), eg." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 47, + 628, + 287, + 653 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 628, + 287, + 653 + ], + "spans": [ + { + "bbox": [ + 47, + 628, + 287, + 653 + ], + "type": "interline_equation", + "content": "u = \\left(c _ {1} + c _ {2} v + c _ {3} v ^ {2}\\right) ^ {- 1}, \\quad \\text {o r} \\quad u = \\left(d _ {1} + d _ {2} v + d _ {3} v ^ {2}\\right) ^ {- 1}. \\tag {17}", + "image_path": "72bb82037fcc184a60186e467b7460625562143927f9d3ef555c99afcc7692a7.jpg" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 47, + 654, + 286, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 654, + 286, + 677 + ], + "spans": [ + { + "bbox": [ + 47, + 654, + 286, + 677 + ], + "type": "text", + "content": "To summarize, we provide the outline of steps for solving the P2P1L absolute pose problem in Figure 3." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 47, + 677, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 677, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 677, + 287, + 714 + ], + "type": "text", + "content": "Algebraically, the nontrivial steps in Figure 3 are the second and fourth, which require solving a univariate quadratic equation and computing square roots, respectively." + } + ] + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 321, + 72, + 411, + 158 + ], + "blocks": [ + { + "bbox": [ + 62, + 166, + 271, + 177 + ], + "lines": [ + { + "bbox": [ + 62, + 166, + 271, + 177 + ], + "spans": [ + { + "bbox": [ + 62, + 166, + 271, + 177 + ], + "type": "text", + "content": "Figure 3. Generic P2P1L solver. See Sec. 2.1 for details." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 321, + 72, + 411, + 158 + ], + "lines": [ + { + "bbox": [ + 321, + 72, + 411, + 158 + ], + "spans": [ + { + "bbox": [ + 321, + 72, + 411, + 158 + ], + "type": "image", + "image_path": "b4abe6b8df011381fda4452e30692a308445f6f91b28742711a701246179f0df.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 430, + 77, + 525, + 158 + ], + "blocks": [ + { + "bbox": [ + 430, + 77, + 525, + 158 + ], + "lines": [ + { + "bbox": [ + 430, + 77, + 525, + 158 + ], + "spans": [ + { + "bbox": [ + 430, + 77, + 525, + 158 + ], + "type": "image", + "image_path": "9a0741bac57a0382aa9616fe8857009f1c20205d53c2e9f5227403b71972772b.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 323, + 159, + 410, + 230 + ], + "blocks": [ + { + "bbox": [ + 323, + 159, + 410, + 230 + ], + "lines": [ + { + "bbox": [ + 323, + 159, + 410, + 230 + ], + "spans": [ + { + "bbox": [ + 323, + 159, + 410, + 230 + ], + "type": "image", + "image_path": "a1a81f7e8cca138e09932e76cff33e20b2d1e8329fc6a6c10f92ee769aaa5f8d.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 237, + 545, + 258 + ], + "lines": [ + { + "bbox": [ + 305, + 237, + 545, + 258 + ], + "spans": [ + { + "bbox": [ + 305, + 237, + 545, + 258 + ], + "type": "text", + "content": "Figure 4. Special reference frame for P1P2L. Camera frame " + }, + { + "bbox": [ + 305, + 237, + 545, + 258 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_0" + }, + { + "bbox": [ + 305, + 237, + 545, + 258 + ], + "type": "text", + "content": " is transformed to " + }, + { + "bbox": [ + 305, + 237, + 545, + 258 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_1" + }, + { + "bbox": [ + 305, + 237, + 545, + 258 + ], + "type": "text", + "content": ", world frame " + }, + { + "bbox": [ + 305, + 237, + 545, + 258 + ], + "type": "inline_equation", + "content": "\\mathcal{W}_0" + }, + { + "bbox": [ + 305, + 237, + 545, + 258 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 305, + 237, + 545, + 258 + ], + "type": "inline_equation", + "content": "\\mathcal{W}_1" + }, + { + "bbox": [ + 305, + 237, + 545, + 258 + ], + "type": "text", + "content": ". See text for details." + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_caption" + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 436, + 160, + 520, + 230 + ], + "blocks": [ + { + "bbox": [ + 436, + 160, + 520, + 230 + ], + "lines": [ + { + "bbox": [ + 436, + 160, + 520, + 230 + ], + "spans": [ + { + "bbox": [ + 436, + 160, + 520, + 230 + ], + "type": "image", + "image_path": "83257912b90f36c02c4353cfc522c5df9718efe3b89b8405f0bac422d13bd179.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_body" + } + ], + "index": 27 + }, + { + "bbox": [ + 306, + 281, + 360, + 292 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 281, + 360, + 292 + ], + "spans": [ + { + "bbox": [ + 306, + 281, + 360, + 292 + ], + "type": "text", + "content": "2.2.P1P2L" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 305, + 300, + 545, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 300, + 545, + 323 + ], + "spans": [ + { + "bbox": [ + 305, + 300, + 545, + 323 + ], + "type": "text", + "content": "Here, we provide the algebraically optimal solution to the P1P2L problem. We mostly follow the notation of [25]." + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 305, + 324, + 545, + 420 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 324, + 545, + 420 + ], + "spans": [ + { + "bbox": [ + 305, + 324, + 545, + 420 + ], + "type": "text", + "content": "Problem Statement: Let us consider a 3D point " + }, + { + "bbox": [ + 305, + 324, + 545, + 420 + ], + "type": "inline_equation", + "content": "P_1^0 \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 305, + 324, + 545, + 420 + ], + "type": "text", + "content": ", and its homogeneous 2D projection " + }, + { + "bbox": [ + 305, + 324, + 545, + 420 + ], + "type": "inline_equation", + "content": "D_1 \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 305, + 324, + 545, + 420 + ], + "type": "text", + "content": " under an unknown calibrated camera matrix. Let us also consider two 3D lines—the first spanned by points " + }, + { + "bbox": [ + 305, + 324, + 545, + 420 + ], + "type": "inline_equation", + "content": "L_1, L_2 \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 305, + 324, + 545, + 420 + ], + "type": "text", + "content": ", and the second one is spanned by " + }, + { + "bbox": [ + 305, + 324, + 545, + 420 + ], + "type": "inline_equation", + "content": "L_3, L_4 \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 305, + 324, + 545, + 420 + ], + "type": "text", + "content": "—and both of their corresponding projections. Our task is to recover the camera matrix from the given point-point correspondence and line-line correspondences." + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 305, + 421, + 545, + 457 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 421, + 545, + 457 + ], + "spans": [ + { + "bbox": [ + 305, + 421, + 545, + 457 + ], + "type": "text", + "content": "Much like the P2P1L solver, our P1P2L solver begins by transforming the input data into a special reference frame, as illustrated in Figure 4. Specifically," + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 306, + 458, + 506, + 469 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 458, + 506, + 469 + ], + "spans": [ + { + "bbox": [ + 306, + 458, + 506, + 469 + ], + "type": "text", + "content": "- In the world frame, the 3D point takes the form" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 389, + 479, + 545, + 497 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 389, + 479, + 545, + 497 + ], + "spans": [ + { + "bbox": [ + 389, + 479, + 545, + 497 + ], + "type": "interline_equation", + "content": "P _ {1} = \\left( \\begin{array}{l l l} 0 & 0 & 0 \\end{array} \\right) ^ {T}. \\tag {18}", + "image_path": "31e87cd3ed6e1173a858078263f3583d1a3f172472a65476741a084dc12fc510.jpg" + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 314, + 506, + 452, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 506, + 452, + 518 + ], + "spans": [ + { + "bbox": [ + 314, + 506, + 452, + 518 + ], + "type": "text", + "content": "The first line is spanned by points" + } + ] + } + ], + "index": 35 + }, + { + "bbox": [ + 323, + 528, + 545, + 555 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 323, + 528, + 545, + 555 + ], + "spans": [ + { + "bbox": [ + 323, + 528, + 545, + 555 + ], + "type": "interline_equation", + "content": "L _ {1} = \\left( \\begin{array}{l l l} X _ {1} & Y _ {1} & Z _ {1} \\end{array} \\right) ^ {T}, \\quad L _ {2} = \\left( \\begin{array}{l l l} X _ {2} & Y _ {2} & Z _ {2} \\end{array} \\right) ^ {T}, \\tag {19}", + "image_path": "9e9b17b48a3f586555560285a0a5d290c66919d2788731ce916da2e45d3f6dc3.jpg" + } + ] + } + ], + "index": 36 + }, + { + "bbox": [ + 314, + 556, + 477, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 556, + 477, + 567 + ], + "spans": [ + { + "bbox": [ + 314, + 556, + 477, + 567 + ], + "type": "text", + "content": "and the second line is spanned by points" + } + ] + } + ], + "index": 37 + }, + { + "bbox": [ + 323, + 577, + 545, + 604 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 323, + 577, + 545, + 604 + ], + "spans": [ + { + "bbox": [ + 323, + 577, + 545, + 604 + ], + "type": "interline_equation", + "content": "L _ {3} = \\left( \\begin{array}{l l l} X _ {3} & Y _ {3} & Z _ {3} \\end{array} \\right) ^ {T}, \\quad L _ {4} = \\left( \\begin{array}{l l l} X _ {4} & Y _ {4} & Z _ {4} \\end{array} \\right) ^ {T}. \\tag {20}", + "image_path": "b37d2cf8fbd3fdb74c2cc71fb9a0b0d712afb421fc06f5db7cdb4827072b360a.jpg" + } + ] + } + ], + "index": 38 + }, + { + "bbox": [ + 306, + 605, + 545, + 688 + ], + "type": "list", + "angle": 0, + "index": 42, + "blocks": [ + { + "bbox": [ + 306, + 605, + 500, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 605, + 500, + 617 + ], + "spans": [ + { + "bbox": [ + 306, + 605, + 500, + 617 + ], + "type": "text", + "content": "- For " + }, + { + "bbox": [ + 306, + 605, + 500, + 617 + ], + "type": "inline_equation", + "content": "\\mathcal{W}_0 \\to \\mathcal{W}_1" + }, + { + "bbox": [ + 306, + 605, + 500, + 617 + ], + "type": "text", + "content": ", we simply translate " + }, + { + "bbox": [ + 306, + 605, + 500, + 617 + ], + "type": "inline_equation", + "content": "P_1^0" + }, + { + "bbox": [ + 306, + 605, + 500, + 617 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 306, + 605, + 500, + 617 + ], + "type": "inline_equation", + "content": "P_1" + }, + { + "bbox": [ + 306, + 605, + 500, + 617 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 39 + }, + { + "bbox": [ + 306, + 617, + 545, + 675 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 617, + 545, + 675 + ], + "spans": [ + { + "bbox": [ + 306, + 617, + 545, + 675 + ], + "type": "text", + "content": "- The world frame may still be rotated freely. We may use the strategy of Section 2.3, which makes the solver stable for a larger class of non-coplanar scenes. However, if " + }, + { + "bbox": [ + 306, + 617, + 545, + 675 + ], + "type": "inline_equation", + "content": "P_{1}, \\overline{L_{1}L_{2}}" + }, + { + "bbox": [ + 306, + 617, + 545, + 675 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 306, + 617, + 545, + 675 + ], + "type": "inline_equation", + "content": "\\overline{L_{3}L_{4}}" + }, + { + "bbox": [ + 306, + 617, + 545, + 675 + ], + "type": "text", + "content": " are coplanar, no rotation is recommended." + } + ] + } + ], + "index": 40 + }, + { + "bbox": [ + 306, + 677, + 514, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 677, + 514, + 688 + ], + "spans": [ + { + "bbox": [ + 306, + 677, + 514, + 688 + ], + "type": "text", + "content": "- In the camera frame, the camera center is fixed at" + } + ] + } + ], + "index": 41 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 386, + 698, + 545, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 386, + 698, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 386, + 698, + 545, + 715 + ], + "type": "interline_equation", + "content": "C = \\left( \\begin{array}{l l l} 0 & 0 & - 1 \\end{array} \\right) ^ {T}. \\tag {21}", + "image_path": "e9b2e6e7509b49a337170384d924f674a1ebeaf8b22604d7d9a4aa912efb812f.jpg" + } + ] + } + ], + "index": 43 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "21319" + } + ] + } + ], + "index": 44 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 279, + 84 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 279, + 84 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 279, + 84 + ], + "type": "text", + "content": "The image point is the projection of a point of the form" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 125, + 91, + 287, + 109 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 125, + 91, + 287, + 109 + ], + "spans": [ + { + "bbox": [ + 125, + 91, + 287, + 109 + ], + "type": "interline_equation", + "content": "D _ {1} = \\left( \\begin{array}{l l l} a _ {1} & b _ {1} & 0 \\end{array} \\right) ^ {T}, \\tag {22}", + "image_path": "dc9e3331b0ed9db6de4b0f93e2d9c438441c5fb29571efda41559406944548f1.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 116, + 287, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 116, + 287, + 140 + ], + "spans": [ + { + "bbox": [ + 55, + 116, + 287, + 140 + ], + "type": "text", + "content": "the first line in the image is the projection of a line spanned by points of the form" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 72, + 147, + 287, + 165 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 72, + 147, + 287, + 165 + ], + "spans": [ + { + "bbox": [ + 72, + 147, + 287, + 165 + ], + "type": "interline_equation", + "content": "D _ {2} = \\left( \\begin{array}{l l l} a _ {2} & 0 & 0 \\end{array} \\right) ^ {T}, \\quad D _ {3} = \\left( \\begin{array}{l l l} 0 & 0 & 0 \\end{array} \\right) ^ {T}, \\tag {23}", + "image_path": "8e26680f943e3a58822bf7c096b567ae3922fa05edbcde1ae4c23c9294ff2374.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 172, + 287, + 197 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 172, + 287, + 197 + ], + "spans": [ + { + "bbox": [ + 55, + 172, + 287, + 197 + ], + "type": "text", + "content": "and the second line in the image is the projection of a line spanned by points of the form" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 66, + 203, + 287, + 221 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 203, + 287, + 221 + ], + "spans": [ + { + "bbox": [ + 66, + 203, + 287, + 221 + ], + "type": "interline_equation", + "content": "D _ {4} = \\left( \\begin{array}{l l l} a _ {4} & b _ {4} & 0 \\end{array} \\right) ^ {T}, \\quad D _ {5} = \\left( \\begin{array}{l l l} a _ {5} & b _ {5} & 0 \\end{array} \\right) ^ {T}. \\tag {24}", + "image_path": "56f0a9454334eafcf6f0c65d96750e6193517be847c4bba20aa1db0f35b1d137.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 228, + 287, + 276 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 47, + 228, + 287, + 264 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 228, + 287, + 264 + ], + "spans": [ + { + "bbox": [ + 47, + 228, + 287, + 264 + ], + "type": "text", + "content": "- The homogeneous coordinates of the lines in the camera frame are " + }, + { + "bbox": [ + 47, + 228, + 287, + 264 + ], + "type": "inline_equation", + "content": "n_1 = d_2 \\times d_3" + }, + { + "bbox": [ + 47, + 228, + 287, + 264 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 47, + 228, + 287, + 264 + ], + "type": "inline_equation", + "content": "n_2 = d_4 \\times d_5" + }, + { + "bbox": [ + 47, + 228, + 287, + 264 + ], + "type": "text", + "content": ", where the rays " + }, + { + "bbox": [ + 47, + 228, + 287, + 264 + ], + "type": "inline_equation", + "content": "d_2 \\neq d_3" + }, + { + "bbox": [ + 47, + 228, + 287, + 264 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 47, + 228, + 287, + 264 + ], + "type": "inline_equation", + "content": "d_4 \\neq d_5" + }, + { + "bbox": [ + 47, + 228, + 287, + 264 + ], + "type": "text", + "content": " to meet the lines in distinct points." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 264, + 244, + 276 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 264, + 244, + 276 + ], + "spans": [ + { + "bbox": [ + 47, + 264, + 244, + 276 + ], + "type": "text", + "content": "To find the transformation " + }, + { + "bbox": [ + 47, + 264, + 244, + 276 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_0\\to \\mathcal{C}_1" + }, + { + "bbox": [ + 47, + 264, + 244, + 276 + ], + "type": "text", + "content": " , we define" + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 283, + 287, + 309 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 283, + 287, + 309 + ], + "spans": [ + { + "bbox": [ + 55, + 283, + 287, + 309 + ], + "type": "interline_equation", + "content": "d _ {1 2} = n _ {1} \\times n _ {2}, \\quad D _ {2} ^ {0} = C ^ {0} + \\frac {d _ {2}}{d _ {2} ^ {T} d _ {1 2}}, \\quad D _ {3} ^ {0} = C ^ {0} + d _ {3},", + "image_path": "b5d18c939421480c0af4f45e6f1b6dcfdc30930b8b8a16c3569f4ae8ac49ea51.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 316, + 306, + 334 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 316, + 306, + 334 + ], + "spans": [ + { + "bbox": [ + 55, + 316, + 306, + 334 + ], + "type": "interline_equation", + "content": "D _ {2} = \\left( \\begin{array}{c c} \\tan (\\cos^ {- 1} (d _ {2} ^ {T} d _ {1 2})) & 0 \\quad 0 \\end{array} \\right) ^ {T}, \\quad D _ {3} = \\left( \\begin{array}{c c c} 0 & 0 & 0 \\end{array} \\right) ^ {T},", + "image_path": "8cec95aaafa7d8d4c92cc0db2b34980a010d7b38c2f5e67de19cd0bb2c518543.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 337, + 287, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 337, + 287, + 365 + ], + "spans": [ + { + "bbox": [ + 55, + 337, + 287, + 365 + ], + "type": "text", + "content": "and, independently of " + }, + { + "bbox": [ + 55, + 337, + 287, + 365 + ], + "type": "inline_equation", + "content": "C^0" + }, + { + "bbox": [ + 55, + 337, + 287, + 365 + ], + "type": "text", + "content": ", rigidly transform the rays " + }, + { + "bbox": [ + 55, + 337, + 287, + 365 + ], + "type": "inline_equation", + "content": "d_2 \\coloneqq \\overrightarrow{C^0D_2^0}" + }, + { + "bbox": [ + 55, + 337, + 287, + 365 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 337, + 287, + 365 + ], + "type": "inline_equation", + "content": "d_3 \\coloneqq \\overrightarrow{C^0D_3^0}" + }, + { + "bbox": [ + 55, + 337, + 287, + 365 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 55, + 337, + 287, + 365 + ], + "type": "inline_equation", + "content": "\\overrightarrow{C^0D_2^0}" + }, + { + "bbox": [ + 55, + 337, + 287, + 365 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 337, + 287, + 365 + ], + "type": "inline_equation", + "content": "\\overrightarrow{CD_3}" + }, + { + "bbox": [ + 55, + 337, + 287, + 365 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 47, + 365, + 287, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 365, + 287, + 399 + ], + "spans": [ + { + "bbox": [ + 47, + 365, + 287, + 399 + ], + "type": "text", + "content": "In the special reference frame, we now write the pose as " + }, + { + "bbox": [ + 47, + 365, + 287, + 399 + ], + "type": "inline_equation", + "content": "(R_{ij})_{1\\leq i,j\\leq 3}" + }, + { + "bbox": [ + 47, + 365, + 287, + 399 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 47, + 365, + 287, + 399 + ], + "type": "inline_equation", + "content": "(T_i)_{1\\leq i\\leq 3}" + }, + { + "bbox": [ + 47, + 365, + 287, + 399 + ], + "type": "text", + "content": ". The projection constraints can be formulated as:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 408, + 287, + 422 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 106, + 408, + 287, + 422 + ], + "spans": [ + { + "bbox": [ + 106, + 408, + 287, + 422 + ], + "type": "interline_equation", + "content": "\\left(D _ {1} - C\\right) \\times \\left(R P _ {1} + T\\right) = 0, \\tag {25}", + "image_path": "5854d3072393f8cba297a502ecf6c8778fef478c2491ae7db8e5c0f1eda1d0cc.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 85, + 440, + 287, + 455 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 85, + 440, + 287, + 455 + ], + "spans": [ + { + "bbox": [ + 85, + 440, + 287, + 455 + ], + "type": "interline_equation", + "content": "\\left(D _ {2} \\times D _ {3}\\right) ^ {T} \\left(R L _ {i} + T\\right) = 0, i \\in \\{1, 2 \\}, \\tag {26}", + "image_path": "cb49a8464c7e1c06de30cea198a1076274af32a2f1c3d721c25e94b207f3d284.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 85, + 470, + 287, + 484 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 85, + 470, + 287, + 484 + ], + "spans": [ + { + "bbox": [ + 85, + 470, + 287, + 484 + ], + "type": "interline_equation", + "content": "\\left(D _ {4} \\times D _ {5}\\right) ^ {T} \\left(R L _ {i} + T\\right) = 0, i \\in \\{3, 4 \\}. \\tag {27}", + "image_path": "d28df0219247a30d84e1a812e09dd46ac484770b21aba7852d3ba19ee4b91659.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 47, + 489, + 287, + 536 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 489, + 287, + 536 + ], + "spans": [ + { + "bbox": [ + 47, + 489, + 287, + 536 + ], + "type": "text", + "content": "Analogously to (11), we have a system of linear equations " + }, + { + "bbox": [ + 47, + 489, + 287, + 536 + ], + "type": "inline_equation", + "content": "A\\mathbf{x} = \\mathbf{b}" + }, + { + "bbox": [ + 47, + 489, + 287, + 536 + ], + "type": "text", + "content": ", obtained by picking 2 out of the 3 redundant constraints from (25) together with (26), (27), where now (cf. [25, eq. 7-8])" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 47, + 544, + 264, + 590 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 544, + 264, + 590 + ], + "spans": [ + { + "bbox": [ + 47, + 544, + 264, + 590 + ], + "type": "interline_equation", + "content": "A = \\left( \\begin{array}{c c c c c c c c c} 0 & 0 & 0 & 0 & 0 & 0 & - b _ {1} & a _ {1} & 0 \\\\ 0 & 0 & 0 & 0 & 0 & 0 & 0 & - 1 & b _ {1} \\\\ 0 & 0 & X _ {1} & Y _ {1} & Y _ {1} & Z _ {1} & 0 & 1 & 0 \\\\ 0 & 0 & X _ {2} & Y _ {2} & Y _ {2} & Z _ {2} & 0 & 1 & 0 \\\\ - b _ {4} X _ {3} & - b _ {4} Y _ {3} & a _ {4} X _ {3} & a _ {4} Y _ {3} & a _ {4} Y _ {3} & a _ {4} Z _ {3} & - b _ {4} & a _ {4} & 0 \\\\ - b _ {4} X _ {4} & - b _ {4} Z _ {4} & a _ {4} X _ {4} & a _ {4} Y _ {4} & a _ {4} Y _ {4} & a _ {4} Z _ {4} & - b _ {4} & a _ {4} & 0 \\end{array} \\right),", + "image_path": "5b3e01006cccb132270bba6a1b892e4326d7783d86664575f38a7439349b9743.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 48, + 590, + 297, + 607 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 590, + 297, + 607 + ], + "spans": [ + { + "bbox": [ + 48, + 590, + 297, + 607 + ], + "type": "interline_equation", + "content": "\\mathbf {x} = \\left( \\begin{array}{c c c c c c c c c} R _ {1 1} & R _ {1 2} & R _ {1 3} & R _ {2 1} & R _ {2 2} & R _ {2 3} & T _ {1} & T _ {2} & T _ {3} \\end{array} \\right) ^ {T},", + "image_path": "530ae3c85f76f14d6a2e051fcc75ac1bbd49cb74e74d1fdf6d6de6053ce8e51e.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 48, + 609, + 182, + 625 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 609, + 182, + 625 + ], + "spans": [ + { + "bbox": [ + 48, + 609, + 182, + 625 + ], + "type": "interline_equation", + "content": "\\mathbf {b} = \\left( \\begin{array}{c c c c c c} 0 & - b _ {1} & 0 & 0 & 0 & 0 \\end{array} \\right) ^ {T}.", + "image_path": "92428c17c1158da970e6fff4b36a478e69b5279d615f148d4e5226655126af79.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 47, + 633, + 218, + 645 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 633, + 218, + 645 + ], + "spans": [ + { + "bbox": [ + 47, + 633, + 218, + 645 + ], + "type": "text", + "content": "Solving for translation as in (12), we have" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 64, + 652, + 228, + 675 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 652, + 228, + 675 + ], + "spans": [ + { + "bbox": [ + 64, + 652, + 228, + 675 + ], + "type": "interline_equation", + "content": "T _ {1} = \\frac {a _ {1}}{b _ {1}} (- R _ {2 1} X _ {1} - R _ {2 2} Y _ {1} - R _ {2 3} Z _ {1}),", + "image_path": "8648162c728c5a0c0ccc21049babe7c77bdb3fdd07d8d2f0e8240ba151cf3337.jpg" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 65, + 677, + 287, + 690 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 677, + 287, + 690 + ], + "spans": [ + { + "bbox": [ + 65, + 677, + 287, + 690 + ], + "type": "interline_equation", + "content": "T _ {2} = - R _ {2 1} X _ {1} - R _ {2 2} Y _ {1} - R _ {2 3} Z _ {1}, \\tag {28}", + "image_path": "2b534f658fc4893c429b7341d05facd64267cbe25db37d6dd86cbdc19ea851d8.jpg" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 65, + 693, + 252, + 716 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 693, + 252, + 716 + ], + "spans": [ + { + "bbox": [ + 65, + 693, + 252, + 716 + ], + "type": "interline_equation", + "content": "T _ {3} = - 1 + \\frac {1}{b _ {1}} \\left(- R _ {2 1} X _ {1} - R _ {2 2} Y _ {1} - R _ {2 3} Z _ {1}\\right).", + "image_path": "fa9e4dcc4eac51b0f2e6391b0c4cddcae3c5193d2dc3816c4276833539b8f724.jpg" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 306, + 72, + 538, + 85 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 538, + 85 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 538, + 85 + ], + "type": "text", + "content": "Furthermore, we express " + }, + { + "bbox": [ + 306, + 72, + 538, + 85 + ], + "type": "inline_equation", + "content": "R_{23}" + }, + { + "bbox": [ + 306, + 72, + 538, + 85 + ], + "type": "text", + "content": " from the fourth equation as" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 312, + 88, + 545, + 112 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 88, + 545, + 112 + ], + "spans": [ + { + "bbox": [ + 312, + 88, + 545, + 112 + ], + "type": "interline_equation", + "content": "R _ {2 3} = \\frac {1}{Z _ {1} - Z _ {2}} \\left(R _ {2 1} \\left(X _ {2} - X _ {1}\\right) + R _ {2 2} \\left(Y _ {2} - Y _ {1}\\right)\\right), \\tag {29}", + "image_path": "a0434669a5a069afb6dc695e43cb81bbe35e2d83be9a376c8adb3784586f785e.jpg" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 306, + 115, + 530, + 126 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 115, + 530, + 126 + ], + "spans": [ + { + "bbox": [ + 306, + 115, + 530, + 126 + ], + "type": "text", + "content": "and we solve for " + }, + { + "bbox": [ + 306, + 115, + 530, + 126 + ], + "type": "inline_equation", + "content": "R_{11}, R_{12}" + }, + { + "bbox": [ + 306, + 115, + 530, + 126 + ], + "type": "text", + "content": " using the last two equations." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 306, + 127, + 545, + 151 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 127, + 545, + 151 + ], + "spans": [ + { + "bbox": [ + 306, + 127, + 545, + 151 + ], + "type": "text", + "content": "Now, we can express " + }, + { + "bbox": [ + 306, + 127, + 545, + 151 + ], + "type": "inline_equation", + "content": "R_{23}, T_2, T_1, R_{12}, R_{11}" + }, + { + "bbox": [ + 306, + 127, + 545, + 151 + ], + "type": "text", + "content": " as linear combinations of " + }, + { + "bbox": [ + 306, + 127, + 545, + 151 + ], + "type": "inline_equation", + "content": "R_{13}, R_{21}, R_{22}" + }, + { + "bbox": [ + 306, + 127, + 545, + 151 + ], + "type": "text", + "content": " in the following form:" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 353, + 156, + 449, + 167 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 353, + 156, + 449, + 167 + ], + "spans": [ + { + "bbox": [ + 353, + 156, + 449, + 167 + ], + "type": "interline_equation", + "content": "R _ {2 3} = c _ {1} R _ {2 1} + c _ {2} R _ {2 2},", + "image_path": "b1b40ac3c50d2f90fcbd18843c63e1f85b123df01b49e5471a005c8b65d49d3f.jpg" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 358, + 171, + 449, + 182 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 358, + 171, + 449, + 182 + ], + "spans": [ + { + "bbox": [ + 358, + 171, + 449, + 182 + ], + "type": "interline_equation", + "content": "T _ {2} = c _ {3} R _ {2 1} + c _ {4} R _ {2 2},", + "image_path": "e1ddba8c75f966ebb386d8ead99e1c98de282696da5b8c7135f76b53946e7fe3.jpg" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 358, + 186, + 545, + 198 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 358, + 186, + 545, + 198 + ], + "spans": [ + { + "bbox": [ + 358, + 186, + 545, + 198 + ], + "type": "interline_equation", + "content": "T _ {1} = c _ {5} R _ {2 1} + c _ {6} R _ {2 2}, \\tag {30}", + "image_path": "713ebd6f30574ce3ed024cac9813691194d7389e9ce613f7e9469bfd008b657e.jpg" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 353, + 201, + 486, + 213 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 353, + 201, + 486, + 213 + ], + "spans": [ + { + "bbox": [ + 353, + 201, + 486, + 213 + ], + "type": "interline_equation", + "content": "R _ {1 2} = c _ {7} R _ {1 3} + c _ {8} R _ {2 1} + c _ {9} R _ {2 2},", + "image_path": "296dbee7e475ad98e28fb3beb2d8155352ffe8c5a8a9a4aca9c2bd30375b697a.jpg" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 354, + 216, + 498, + 227 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 354, + 216, + 498, + 227 + ], + "spans": [ + { + "bbox": [ + 354, + 216, + 498, + 227 + ], + "type": "interline_equation", + "content": "R _ {1 1} = c _ {1 0} R _ {1 3} + c _ {1 1} R _ {2 1} + c _ {1 2} R _ {2 2}.", + "image_path": "918d3ecccbb9440d38ae5c6af4cc978e0d4539484c62ead0bde6370f3e607813.jpg" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 306, + 232, + 545, + 255 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 232, + 545, + 255 + ], + "spans": [ + { + "bbox": [ + 306, + 232, + 545, + 255 + ], + "type": "text", + "content": "The non-linear internal constraints imposed on the elements of " + }, + { + "bbox": [ + 306, + 232, + 545, + 255 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 306, + 232, + 545, + 255 + ], + "type": "text", + "content": " have the form:" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 402, + 257, + 497, + 270 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 402, + 257, + 497, + 270 + ], + "spans": [ + { + "bbox": [ + 402, + 257, + 497, + 270 + ], + "type": "interline_equation", + "content": "R _ {1 1} ^ {2} + R _ {1 2} ^ {2} + R _ {1 3} ^ {2} = 1,", + "image_path": "98978c52a55d91bf48e5b98318ca5a825c5b0cb9acec4485237ad5d72c5abe2e.jpg" + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 402, + 274, + 545, + 287 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 402, + 274, + 545, + 287 + ], + "spans": [ + { + "bbox": [ + 402, + 274, + 545, + 287 + ], + "type": "interline_equation", + "content": "R _ {2 1} ^ {2} + R _ {2 2} ^ {2} + R _ {2 3} ^ {2} = 1, \\tag {31}", + "image_path": "96b5d8191da5c8c9852e4e7bf10e208d5aa4c2f9f09beeb35ba5a1e003695f83.jpg" + } + ] + } + ], + "index": 35 + }, + { + "bbox": [ + 354, + 290, + 497, + 301 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 354, + 290, + 497, + 301 + ], + "spans": [ + { + "bbox": [ + 354, + 290, + 497, + 301 + ], + "type": "interline_equation", + "content": "R _ {1 1} R _ {2 1} + R _ {1 2} R _ {2 2} + R _ {1 3} R _ {2 3} = 0.", + "image_path": "231f83871a567aafba501b9f91828153f5d0261c0c344e9ca186c0698d0e0e0d.jpg" + } + ] + } + ], + "index": 36 + }, + { + "bbox": [ + 306, + 306, + 545, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 306, + 545, + 319 + ], + "spans": [ + { + "bbox": [ + 306, + 306, + 545, + 319 + ], + "type": "text", + "content": "We substitute (30) into (31), obtaining equations of the form" + } + ] + } + ], + "index": 37 + }, + { + "bbox": [ + 344, + 335, + 545, + 348 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 344, + 335, + 545, + 348 + ], + "spans": [ + { + "bbox": [ + 344, + 335, + 545, + 348 + ], + "type": "interline_equation", + "content": "\\left( \\begin{array}{l l l l l l} d _ {1} & d _ {2} & d _ {3} & d _ {4} & d _ {5} & d _ {6} \\end{array} \\right) \\nu (R) = 1, \\tag {32}", + "image_path": "bdd9f967c4e3a6b8260f195582272271281a61ffb0def07c3a4ae79b3ef6b308.jpg" + } + ] + } + ], + "index": 38 + }, + { + "bbox": [ + 358, + 351, + 545, + 364 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 358, + 351, + 545, + 364 + ], + "spans": [ + { + "bbox": [ + 358, + 351, + 545, + 364 + ], + "type": "interline_equation", + "content": "\\left( \\begin{array}{c c c c c c} 0 & 0 & 0 & d _ {7} & d _ {8} & d _ {9} \\end{array} \\right) \\nu (R) = 1, \\tag {33}", + "image_path": "7916c2a242c904a891c9d14c2099cd2a5eb15a0c93aa841fa51fe8c6d50e7417.jpg" + } + ] + } + ], + "index": 39 + }, + { + "bbox": [ + 329, + 367, + 545, + 380 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 329, + 367, + 545, + 380 + ], + "spans": [ + { + "bbox": [ + 329, + 367, + 545, + 380 + ], + "type": "interline_equation", + "content": "\\left( \\begin{array}{c c c c c c} 0 & d _ {1 0} & d _ {1 1} & d _ {1 2} & d _ {1 3} & d _ {1 4} \\end{array} \\right) \\nu (R) = 0, \\tag {34}", + "image_path": "4bf8244de7ecc1002314289cb240142e82f0fe44753f66435f70198e43aaaf39.jpg" + } + ] + } + ], + "index": 40 + }, + { + "bbox": [ + 306, + 385, + 333, + 395 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 385, + 333, + 395 + ], + "spans": [ + { + "bbox": [ + 306, + 385, + 333, + 395 + ], + "type": "text", + "content": "where" + } + ] + } + ], + "index": 41 + }, + { + "bbox": [ + 306, + 398, + 545, + 426 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 398, + 545, + 426 + ], + "spans": [ + { + "bbox": [ + 306, + 398, + 545, + 426 + ], + "type": "interline_equation", + "content": "\\nu (R) = \\left( \\begin{array}{l l l l l} R _ {1 3} ^ {2} & R _ {1 3} R _ {2 1} & R _ {1 3} R _ {2 2} & R _ {2 1} ^ {2} R _ {2 1} R _ {2 2} & R _ {2 2} ^ {2} \\end{array} \\right) ^ {T}. \\tag {35}", + "image_path": "f36100f5df898d955a1d566303e5d7e2dd2f0532cb4cfe5ce75dbeefe884f358.jpg" + } + ] + } + ], + "index": 42 + }, + { + "bbox": [ + 306, + 426, + 425, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 426, + 425, + 437 + ], + "spans": [ + { + "bbox": [ + 306, + 426, + 425, + 437 + ], + "type": "text", + "content": "We express " + }, + { + "bbox": [ + 306, + 426, + 425, + 437 + ], + "type": "inline_equation", + "content": "R_{13}" + }, + { + "bbox": [ + 306, + 426, + 425, + 437 + ], + "type": "text", + "content": " from (34) as" + } + ] + } + ], + "index": 43 + }, + { + "bbox": [ + 330, + 441, + 545, + 467 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 330, + 441, + 545, + 467 + ], + "spans": [ + { + "bbox": [ + 330, + 441, + 545, + 467 + ], + "type": "interline_equation", + "content": "R _ {1 3} = - \\frac {d _ {1 2} R _ {2 1} ^ {2} + d _ {1 3} R _ {2 1} R _ {2 2} + d _ {1 4} R _ {2 2} ^ {2}}{d _ {1 0} R _ {2 1} + d _ {1 1} R _ {2 2}}. \\tag {36}", + "image_path": "8c881b9369e5d89a1d196e0638b17e01b2e009b01a171391ebb9554f2a449181.jpg" + } + ] + } + ], + "index": 44 + }, + { + "bbox": [ + 306, + 470, + 457, + 482 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 470, + 457, + 482 + ], + "spans": [ + { + "bbox": [ + 306, + 470, + 457, + 482 + ], + "type": "text", + "content": "Substituting (36) into (32), we obtain" + } + ] + } + ], + "index": 45 + }, + { + "bbox": [ + 340, + 485, + 506, + 512 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 340, + 485, + 506, + 512 + ], + "spans": [ + { + "bbox": [ + 340, + 485, + 506, + 512 + ], + "type": "interline_equation", + "content": "d _ {1} \\frac {\\left(d _ {1 2} R _ {2 1} ^ {2} + d _ {1 3} R _ {2 1} R _ {2 2} + d _ {1 4} R _ {2 2} ^ {2}\\right) ^ {2}}{\\left(d _ {1 0} R _ {2 1} + d _ {1 1} R _ {2 2}\\right) ^ {2}} -", + "image_path": "52f479c54ec7d20da48b200f9621de0dd38b5f0b5230b1b2e4a1030de0433e97.jpg" + } + ] + } + ], + "index": 46 + }, + { + "bbox": [ + 341, + 514, + 510, + 540 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 341, + 514, + 510, + 540 + ], + "spans": [ + { + "bbox": [ + 341, + 514, + 510, + 540 + ], + "type": "interline_equation", + "content": "d _ {2} \\frac {d _ {1 2} R _ {2 1} ^ {3} + d _ {1 3} R _ {2 1} ^ {2} R _ {2 2} + d _ {1 4} R _ {2 1} R _ {2 2} ^ {2}}{d _ {1 0} R _ {2 1} + d _ {1 1} R _ {2 2}} -", + "image_path": "777b85aab1fe65f66da3a70c62d3df8847a81f60aa4ed2e694f9f20485c4a497.jpg" + } + ] + } + ], + "index": 47 + }, + { + "bbox": [ + 341, + 541, + 510, + 567 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 341, + 541, + 510, + 567 + ], + "spans": [ + { + "bbox": [ + 341, + 541, + 510, + 567 + ], + "type": "interline_equation", + "content": "d _ {3} \\frac {d _ {1 2} R _ {2 1} ^ {2} R _ {2 2} + d _ {1 3} R _ {2 1} R _ {2 2} ^ {2} + d _ {1 4} R _ {2 2} ^ {3}}{d _ {1 0} R _ {2 1} + d _ {1 1} R _ {2 2}} +", + "image_path": "c87fd803c692309876080e82687a775e06561618f17018efccdb429a98e4f04e.jpg" + } + ] + } + ], + "index": 48 + }, + { + "bbox": [ + 341, + 569, + 545, + 582 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 341, + 569, + 545, + 582 + ], + "spans": [ + { + "bbox": [ + 341, + 569, + 545, + 582 + ], + "type": "interline_equation", + "content": "d _ {4} R _ {2 1} ^ {2} + d _ {5} R _ {2 1} R _ {2 2} + d _ {6} R _ {2 2} ^ {2} = 1. \\tag {37}", + "image_path": "40a043cc574b8b0de4a848cc9b94cf193586e993d80e97cf483e3495b3a9df1a.jpg" + } + ] + } + ], + "index": 49 + }, + { + "bbox": [ + 306, + 586, + 417, + 598 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 586, + 417, + 598 + ], + "spans": [ + { + "bbox": [ + 306, + 586, + 417, + 598 + ], + "type": "text", + "content": "Subtracting (33) from (37)," + } + ] + } + ], + "index": 50 + }, + { + "bbox": [ + 310, + 601, + 476, + 628 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 310, + 601, + 476, + 628 + ], + "spans": [ + { + "bbox": [ + 310, + 601, + 476, + 628 + ], + "type": "interline_equation", + "content": "d _ {1} \\frac {\\left(d _ {1 2} R _ {2 1} ^ {2} + d _ {1 3} R _ {2 1} R _ {2 2} + d _ {1 4} R _ {2 2} ^ {2}\\right) ^ {2}}{\\left(d _ {1 0} R _ {2 1} + d _ {1 1} R _ {2 2}\\right) ^ {2}} -", + "image_path": "e43ce766bd14a134a87a5b01f93501f143c86305147a3cb8c79b739469595f6b.jpg" + } + ] + } + ], + "index": 51 + }, + { + "bbox": [ + 310, + 630, + 479, + 655 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 310, + 630, + 479, + 655 + ], + "spans": [ + { + "bbox": [ + 310, + 630, + 479, + 655 + ], + "type": "interline_equation", + "content": "d _ {2} \\frac {d _ {1 2} R _ {2 1} ^ {3} + d _ {1 3} R _ {2 1} ^ {2} R _ {2 2} + d _ {1 4} R _ {2 1} R _ {2 2} ^ {2}}{d _ {1 0} R _ {2 1} + d _ {1 1} R _ {2 2}} -", + "image_path": "885c98a9d2213c718c630492aaaec3680bef0f8edb36f3196e8cbb4ca9a5ca70.jpg" + } + ] + } + ], + "index": 52 + }, + { + "bbox": [ + 310, + 658, + 479, + 683 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 310, + 658, + 479, + 683 + ], + "spans": [ + { + "bbox": [ + 310, + 658, + 479, + 683 + ], + "type": "interline_equation", + "content": "d _ {3} \\frac {d _ {1 2} R _ {2 1} ^ {2} R _ {2 2} + d _ {1 3} R _ {2 1} R _ {2 2} ^ {2} + d _ {1 4} R _ {2 2} ^ {3}}{d _ {1 0} R _ {2 1} + d _ {1 1} R _ {2 2}} +", + "image_path": "20648536fe6b889700223059811035d262285bda2d0d111d667d96f1a645a1af.jpg" + } + ] + } + ], + "index": 53 + }, + { + "bbox": [ + 310, + 685, + 545, + 709 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 310, + 685, + 545, + 709 + ], + "spans": [ + { + "bbox": [ + 310, + 685, + 545, + 709 + ], + "type": "interline_equation", + "content": "\\left(d _ {4} - d _ {7}\\right) R _ {2 1} ^ {2} + \\left(d _ {5} - d _ {8}\\right) R _ {2 1} R _ {2 2} + \\left(d _ {6} - d _ {9}\\right) R _ {2 2} ^ {2} = 0. \\tag {38}", + "image_path": "989fc16d7f2a6f404c13ebe217030cf15fd2e8c7207f5be79c09d97a2c9f82ee.jpg" + } + ] + } + ], + "index": 54 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "21320" + } + ] + } + ], + "index": 55 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 42, + 69, + 287, + 177 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 42, + 69, + 286, + 104 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 42, + 69, + 286, + 104 + ], + "spans": [ + { + "bbox": [ + 42, + 69, + 286, + 104 + ], + "type": "text", + "content": "(1) Transform data in world and camera frames to the special reference frames satisfying (18), (19), (20), (21), (22), (23), (24)." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 43, + 105, + 263, + 117 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 43, + 105, + 263, + 117 + ], + "spans": [ + { + "bbox": [ + 43, + 105, + 263, + 117 + ], + "type": "text", + "content": "(2) Compute up to 4 solutions in " + }, + { + "bbox": [ + 43, + 105, + 263, + 117 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 43, + 105, + 263, + 117 + ], + "type": "text", + "content": " to the quartic (41)." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 43, + 117, + 287, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 43, + 117, + 287, + 140 + ], + "spans": [ + { + "bbox": [ + 43, + 117, + 287, + 140 + ], + "type": "text", + "content": "(3) For each root " + }, + { + "bbox": [ + 43, + 117, + 287, + 140 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 43, + 117, + 287, + 140 + ], + "type": "text", + "content": " in step (2), recover 2 solutions in " + }, + { + "bbox": [ + 43, + 117, + 287, + 140 + ], + "type": "inline_equation", + "content": "\\{R_{21}, R_{22}\\}" + }, + { + "bbox": [ + 43, + 117, + 287, + 140 + ], + "type": "text", + "content": " using " + }, + { + "bbox": [ + 43, + 117, + 287, + 140 + ], + "type": "inline_equation", + "content": "R_{21} = R_{22}v" + }, + { + "bbox": [ + 43, + 117, + 287, + 140 + ], + "type": "text", + "content": " and (33)." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 43, + 141, + 287, + 164 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 43, + 141, + 287, + 164 + ], + "spans": [ + { + "bbox": [ + 43, + 141, + 287, + 164 + ], + "type": "text", + "content": "(4) Recover remaining coordinates for each solution using (34),(30)." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 43, + 165, + 247, + 177 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 43, + 165, + 247, + 177 + ], + "spans": [ + { + "bbox": [ + 43, + 165, + 247, + 177 + ], + "type": "text", + "content": "(5) Reverse the transformation applied in step (1)." + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "text" + }, + { + "type": "image", + "bbox": [ + 72, + 217, + 156, + 291 + ], + "blocks": [ + { + "bbox": [ + 72, + 217, + 156, + 291 + ], + "lines": [ + { + "bbox": [ + 72, + 217, + 156, + 291 + ], + "spans": [ + { + "bbox": [ + 72, + 217, + 156, + 291 + ], + "type": "image", + "image_path": "0b3fc4accd39c7cfdb4da1c7c5909a4ea8e8cff9c9a37faba300cce6245b0a78.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 47, + 297, + 287, + 319 + ], + "lines": [ + { + "bbox": [ + 47, + 297, + 287, + 319 + ], + "spans": [ + { + "bbox": [ + 47, + 297, + 287, + 319 + ], + "type": "text", + "content": "Figure 6. Stabilization of P1P2L solver. The frame is rotated to align the line " + }, + { + "bbox": [ + 47, + 297, + 287, + 319 + ], + "type": "inline_equation", + "content": "\\overline{L_1L_2}" + }, + { + "bbox": [ + 47, + 297, + 287, + 319 + ], + "type": "text", + "content": " with the z-axis. See Sec. 2.3 for details." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 175, + 216, + 258, + 291 + ], + "blocks": [ + { + "bbox": [ + 62, + 178, + 271, + 189 + ], + "lines": [ + { + "bbox": [ + 62, + 178, + 271, + 189 + ], + "spans": [ + { + "bbox": [ + 62, + 178, + 271, + 189 + ], + "type": "text", + "content": "Figure 5. Generic P1P2L solver. See Sec. 2.2 for details." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 175, + 216, + 258, + 291 + ], + "lines": [ + { + "bbox": [ + 175, + 216, + 258, + 291 + ], + "spans": [ + { + "bbox": [ + 175, + 216, + 258, + 291 + ], + "type": "image", + "image_path": "cd7d9f719a68f865c3772ae71e7acfebff8f70065de201c7b0b0fe42427766d8.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 344, + 287, + 369 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 344, + 287, + 369 + ], + "spans": [ + { + "bbox": [ + 47, + 344, + 287, + 369 + ], + "type": "text", + "content": "We then clear denominators in (38), multiplying by " + }, + { + "bbox": [ + 47, + 344, + 287, + 369 + ], + "type": "inline_equation", + "content": "(d_{10}R_{21} + d_{11}R_{22})^2" + }, + { + "bbox": [ + 47, + 344, + 287, + 369 + ], + "type": "text", + "content": " to get a polynomial equation" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 66, + 380, + 287, + 504 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 380, + 287, + 504 + ], + "spans": [ + { + "bbox": [ + 66, + 380, + 287, + 504 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} d _ {1} \\left(d _ {1 2} R _ {2 1} ^ {2} + d _ {1 3} R _ {2 1} R _ {2 2} + d _ {1 4} R _ {2 2} ^ {2}\\right) ^ {2} - \\\\ \\left(\\left(d _ {2} \\left(d _ {1 2} R _ {2 1} ^ {3} + d _ {1 3} R _ {2 1} ^ {2} R _ {2 2} + d _ {1 4} R _ {2 1} R _ {2 2} ^ {2}\\right) - \\right. \\right. \\\\ \\left. d _ {3} \\left(d _ {1 2} R _ {2 1} ^ {2} R _ {2 2} + d _ {1 3} R _ {2 1} R _ {2 2} ^ {2} + d _ {1 4} R _ {2 2} ^ {3}\\right)\\right) \\cdot \\tag {39} \\\\ \\left(d _ {1 0} R _ {2 1} + d _ {1 1} R _ {2 2}\\right) + \\left(\\left(d _ {4} - d _ {7}\\right) R _ {2 1} ^ {2} + \\right. \\\\ \\left. \\left(d _ {5} - d _ {8}\\right) R _ {2 1} R _ {2 2} + \\left(d _ {6} - d _ {9}\\right) R _ {2 2} ^ {2}\\right) \\Bigg). \\\\ \\left(d _ {1 0} R _ {2 1} + d _ {1 1} R _ {2 2}\\right) = 0. \\\\ \\end{array}", + "image_path": "25a20c4e5445e6d836cd427e0b36629796e331bae89154b375e0d5c2cbfec276.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 47, + 517, + 287, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 517, + 287, + 529 + ], + "spans": [ + { + "bbox": [ + 47, + 517, + 287, + 529 + ], + "type": "text", + "content": "Upon expanding equation (39), we find that it takes the form" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 47, + 538, + 287, + 563 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 538, + 287, + 563 + ], + "spans": [ + { + "bbox": [ + 47, + 538, + 287, + 563 + ], + "type": "interline_equation", + "content": "\\alpha_ {1} R _ {2 1} ^ {4} + \\alpha_ {2} R _ {2 1} ^ {3} R _ {2 2} + \\alpha_ {3} R _ {2 1} ^ {2} R _ {2 2} ^ {2} + \\alpha_ {4} R _ {2 1} R _ {2 2} ^ {3} + \\alpha_ {5} 2 R _ {2 2} ^ {4} = 0 \\tag {40}", + "image_path": "25276a5ace45c215561723f9b935a1dbe5ebb85d6feef601febe9bf125eb1775.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 47, + 563, + 287, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 563, + 287, + 590 + ], + "spans": [ + { + "bbox": [ + 47, + 563, + 287, + 590 + ], + "type": "text", + "content": "Similarly to (16), we divide (40) by " + }, + { + "bbox": [ + 47, + 563, + 287, + 590 + ], + "type": "inline_equation", + "content": "R_{22}^{4}" + }, + { + "bbox": [ + 47, + 563, + 287, + 590 + ], + "type": "text", + "content": " and define a new variable " + }, + { + "bbox": [ + 47, + 563, + 287, + 590 + ], + "type": "inline_equation", + "content": "v = \\frac{R_{21}}{R_{22}}" + }, + { + "bbox": [ + 47, + 563, + 287, + 590 + ], + "type": "text", + "content": ", and thereby deduce the univariate quartic" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 86, + 603, + 287, + 617 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 86, + 603, + 287, + 617 + ], + "spans": [ + { + "bbox": [ + 86, + 603, + 287, + 617 + ], + "type": "interline_equation", + "content": "\\alpha_ {1} v ^ {4} + \\alpha_ {2} v ^ {3} + \\alpha_ {3} v ^ {2} + \\alpha_ {4} v + \\alpha_ {5} = 0. \\tag {41}", + "image_path": "7e9b1081bf30c7d32cee9a0d81e8371295fbb9edcd543cbaa3e15a1b16337ff6.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 47, + 628, + 286, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 628, + 286, + 652 + ], + "spans": [ + { + "bbox": [ + 47, + 628, + 286, + 652 + ], + "type": "text", + "content": "After solving for " + }, + { + "bbox": [ + 47, + 628, + 286, + 652 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 47, + 628, + 286, + 652 + ], + "type": "text", + "content": ", we can solve for " + }, + { + "bbox": [ + 47, + 628, + 286, + 652 + ], + "type": "inline_equation", + "content": "R_{21}, R_{22}" + }, + { + "bbox": [ + 47, + 628, + 286, + 652 + ], + "type": "text", + "content": " using (33), and for the other equations using (34), and (30)." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 46, + 653, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 287, + 713 + ], + "type": "text", + "content": "To summarize, we provide the outline of steps for solving the P1P2L absolute pose problem in Figure 5. Algebraically, the nontrivial steps in Figure 5 are second and third, which require solving a univariate quartic equation and computing square roots, respectively." + } + ] + } + ], + "index": 17 + }, + { + "type": "table", + "bbox": [ + 307, + 69, + 552, + 197 + ], + "blocks": [ + { + "bbox": [ + 307, + 69, + 552, + 197 + ], + "lines": [ + { + "bbox": [ + 307, + 69, + 552, + 197 + ], + "spans": [ + { + "bbox": [ + 307, + 69, + 552, + 197 + ], + "type": "table", + "html": "
GenericMean RMed. RMean TMed. T
P1P2L no fix0.000501.0e-140.000651.8e-13
P1P2L fix9.0e-094.2e-153.4e-077.0e-14
P2P1L no fix7.1e-111.4e-151.2e-092.1e-14
P2P1L fix3.2e-092.6e-157.4e-084.3e-14
CoplanarMean RMed. RMean TMed. T
P1P2L no fix0.000229.6e-150.000301.75e-13
P1P2L fix1.10.791.20.96
P2P1L no fix2.43.142.23.14
P2P1L fix1.2e-124.0e-157.9e-116.3e-14
", + "image_path": "765307c92acc345246e93bff90c2a30594fcdf5293daa32cd6e07501282fcf2e.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "table_body" + } + ], + "index": 18 + }, + { + "bbox": [ + 305, + 205, + 545, + 238 + ], + "lines": [ + { + "bbox": [ + 305, + 205, + 545, + 238 + ], + "spans": [ + { + "bbox": [ + 305, + 205, + 545, + 238 + ], + "type": "text", + "content": "Table 2. The mean and median of the rot.(R) and tran.(T) errors over 1e5 noiseless generic (top) and coplanar (bottom) samples. In radians. The best results are marked bold." + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 306, + 259, + 462, + 272 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 259, + 462, + 272 + ], + "spans": [ + { + "bbox": [ + 306, + 259, + 462, + 272 + ], + "type": "text", + "content": "2.3. Stabilizing the P1P2L Solver" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 305, + 278, + 545, + 301 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 278, + 545, + 301 + ], + "spans": [ + { + "bbox": [ + 305, + 278, + 545, + 301 + ], + "type": "text", + "content": "In this section, we outline a method to increase the stability of the P1P2L solver (Sec. 2.2)." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 305, + 302, + 545, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 302, + 545, + 373 + ], + "spans": [ + { + "bbox": [ + 305, + 302, + 545, + 373 + ], + "type": "text", + "content": "Our proposed P1P2L solution faces a degeneracy when " + }, + { + "bbox": [ + 305, + 302, + 545, + 373 + ], + "type": "inline_equation", + "content": "Z_{1} - Z_{2} = 0" + }, + { + "bbox": [ + 305, + 302, + 545, + 373 + ], + "type": "text", + "content": ", since this leads to a division by zero in equation (29). Furthermore, our observations indicate that the result of the solver is unstable if the value of " + }, + { + "bbox": [ + 305, + 302, + 545, + 373 + ], + "type": "inline_equation", + "content": "Z_{1} - Z_{2}" + }, + { + "bbox": [ + 305, + 302, + 545, + 373 + ], + "type": "text", + "content": " is close to zero. Note, that this instability is specific to our solver and not inherent to the P1P2L problem." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 305, + 374, + 545, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 374, + 545, + 434 + ], + "spans": [ + { + "bbox": [ + 305, + 374, + 545, + 434 + ], + "type": "text", + "content": "This degeneracy may be interpreted geometrically as follows: the values of " + }, + { + "bbox": [ + 305, + 374, + 545, + 434 + ], + "type": "inline_equation", + "content": "Z_{1}" + }, + { + "bbox": [ + 305, + 374, + 545, + 434 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 374, + 545, + 434 + ], + "type": "inline_equation", + "content": "Z_{2}" + }, + { + "bbox": [ + 305, + 374, + 545, + 434 + ], + "type": "text", + "content": " are the last coordinates of the points " + }, + { + "bbox": [ + 305, + 374, + 545, + 434 + ], + "type": "inline_equation", + "content": "L_{1}, L_{2}" + }, + { + "bbox": [ + 305, + 374, + 545, + 434 + ], + "type": "text", + "content": ", which span the first 3D line. Therefore, the vector " + }, + { + "bbox": [ + 305, + 374, + 545, + 434 + ], + "type": "inline_equation", + "content": "L_{1} - L_{2}" + }, + { + "bbox": [ + 305, + 374, + 545, + 434 + ], + "type": "text", + "content": " represents the direction of the line, and " + }, + { + "bbox": [ + 305, + 374, + 545, + 434 + ], + "type": "inline_equation", + "content": "Z_{1} - Z_{2}" + }, + { + "bbox": [ + 305, + 374, + 545, + 434 + ], + "type": "text", + "content": " represents the last coordinate of this direction." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 305, + 434, + 545, + 493 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 434, + 545, + 493 + ], + "spans": [ + { + "bbox": [ + 305, + 434, + 545, + 493 + ], + "type": "text", + "content": "Since the special reference frame used in Section 2.2 is independent of the rotation of the world frame, we can remove the source of instability by rotating the world frame such that the first line " + }, + { + "bbox": [ + 305, + 434, + 545, + 493 + ], + "type": "inline_equation", + "content": "\\overline{L_1L_2}" + }, + { + "bbox": [ + 305, + 434, + 545, + 493 + ], + "type": "text", + "content": " aligns with the z-axis. See Figure 6 for an illustration." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 306, + 502, + 458, + 515 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 502, + 458, + 515 + ], + "spans": [ + { + "bbox": [ + 306, + 502, + 458, + 515 + ], + "type": "text", + "content": "2.4. Resolving the coplanar case" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 305, + 521, + 545, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 521, + 545, + 556 + ], + "spans": [ + { + "bbox": [ + 305, + 521, + 545, + 556 + ], + "type": "text", + "content": "In this section, we discuss the performance of the P2P1L (Sec. 2.1) and P1P2L (Sec. 2.2) solvers in the case, where all 3D points and 3D lines are coplanar." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 305, + 557, + 545, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 557, + 545, + 628 + ], + "spans": [ + { + "bbox": [ + 305, + 557, + 545, + 628 + ], + "type": "text", + "content": "P2P1L. The P2P1L solver presented in Sec. 2.1 (P2P1L no fix) is degenerate in coplanar situations, since then " + }, + { + "bbox": [ + 305, + 557, + 545, + 628 + ], + "type": "inline_equation", + "content": "Z_{4} = 0" + }, + { + "bbox": [ + 305, + 557, + 545, + 628 + ], + "type": "text", + "content": " (2), causing a division by zero issue in equation (11). To resolve this issue, we propose a modified solver (P2P1L fix), which handles the coplanar case. See SM A for a description of this modified solver." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 305, + 629, + 545, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 629, + 545, + 677 + ], + "spans": [ + { + "bbox": [ + 305, + 629, + 545, + 677 + ], + "type": "text", + "content": "P1P2L. The original version of the P1P2L solver (P1P2L no fix, Sec. 2.2) is able to handle the coplanar case. However, the stabilized P1P2L solver (P1P2L fix, Sec. 2.3) fails due to a degeneracy for coplanar input." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 305, + 677, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 677, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 677, + 545, + 713 + ], + "type": "text", + "content": "We conducted a synthetic experiment to evaluate the impact of the proposed stabilization scheme. In this experiment, we used the same setting as in Section 3.1. The re" + } + ] + } + ], + "index": 29 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "21321" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 78, + 71, + 266, + 113 + ], + "blocks": [ + { + "bbox": [ + 78, + 71, + 266, + 113 + ], + "lines": [ + { + "bbox": [ + 78, + 71, + 266, + 113 + ], + "spans": [ + { + "bbox": [ + 78, + 71, + 266, + 113 + ], + "type": "image", + "image_path": "f3ac5acb4ef713bba12a1367b6196e05f9c5c0ef13e64dd9aeee7b2e99403755.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 50, + 114, + 282, + 232 + ], + "blocks": [ + { + "bbox": [ + 50, + 114, + 282, + 232 + ], + "lines": [ + { + "bbox": [ + 50, + 114, + 282, + 232 + ], + "spans": [ + { + "bbox": [ + 50, + 114, + 282, + 232 + ], + "type": "image", + "image_path": "f62ca8369537d116d1a2718261a301ede8c087a96305d92bf80aed04c791324d.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 47, + 236, + 286, + 258 + ], + "lines": [ + { + "bbox": [ + 47, + 236, + 286, + 258 + ], + "spans": [ + { + "bbox": [ + 47, + 236, + 286, + 258 + ], + "type": "text", + "content": "Figure 7. Stability test. Histogram of rotation " + }, + { + "bbox": [ + 47, + 236, + 286, + 258 + ], + "type": "inline_equation", + "content": "(l.)" + }, + { + "bbox": [ + 47, + 236, + 286, + 258 + ], + "type": "text", + "content": ", and translation " + }, + { + "bbox": [ + 47, + 236, + 286, + 258 + ], + "type": "inline_equation", + "content": "(r.)" + }, + { + "bbox": [ + 47, + 236, + 286, + 258 + ], + "type": "text", + "content": " errors computed over " + }, + { + "bbox": [ + 47, + 236, + 286, + 258 + ], + "type": "inline_equation", + "content": "1e5" + }, + { + "bbox": [ + 47, + 236, + 286, + 258 + ], + "type": "text", + "content": " noiseless samples." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 53, + 271, + 281, + 360 + ], + "blocks": [ + { + "bbox": [ + 53, + 271, + 281, + 360 + ], + "lines": [ + { + "bbox": [ + 53, + 271, + 281, + 360 + ], + "spans": [ + { + "bbox": [ + 53, + 271, + 281, + 360 + ], + "type": "table", + "html": "
MethodMeanMed.MinMax
P2P1L Ours313.8324.9230.83061.0
P2P1L 3Q31860.61909.91439.110102.1
P2P1L R.+S8897.59491.25805.149984.1
P2P1L R.+L4720.85239.52763.315362.6
P1P2L Ours504.0521.0364.04554.4
P1P2L 3Q31967.12008.21483.912931.1
", + "image_path": "9bdda44185fe7690ca56b8a57f9ea1006f0c7d211a8526bad9fc666ad01a083f.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 368, + 286, + 390 + ], + "lines": [ + { + "bbox": [ + 47, + 368, + 286, + 390 + ], + "spans": [ + { + "bbox": [ + 47, + 368, + 286, + 390 + ], + "type": "text", + "content": "Table 3. Runtime The mean, median and max values of the runtime. In nanoseconds. Best values for each problem are bold." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 414, + 286, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 414, + 286, + 498 + ], + "spans": [ + { + "bbox": [ + 46, + 414, + 286, + 498 + ], + "type": "text", + "content": "sults, presented in Table 2 demonstrate that solvers " + }, + { + "bbox": [ + 46, + 414, + 286, + 498 + ], + "type": "inline_equation", + "content": "P2P1L" + }, + { + "bbox": [ + 46, + 414, + 286, + 498 + ], + "type": "text", + "content": " no fix and " + }, + { + "bbox": [ + 46, + 414, + 286, + 498 + ], + "type": "inline_equation", + "content": "P1P2L" + }, + { + "bbox": [ + 46, + 414, + 286, + 498 + ], + "type": "text", + "content": " fix achieve superior numerical stability in the generic case, while solvers " + }, + { + "bbox": [ + 46, + 414, + 286, + 498 + ], + "type": "inline_equation", + "content": "P2P1L" + }, + { + "bbox": [ + 46, + 414, + 286, + 498 + ], + "type": "text", + "content": " fix and " + }, + { + "bbox": [ + 46, + 414, + 286, + 498 + ], + "type": "inline_equation", + "content": "P1P2L" + }, + { + "bbox": [ + 46, + 414, + 286, + 498 + ], + "type": "text", + "content": " no fix handle the coplanar case. Since it is simple to detect the coplanar case, we recommend to use solvers " + }, + { + "bbox": [ + 46, + 414, + 286, + 498 + ], + "type": "inline_equation", + "content": "P2P1L" + }, + { + "bbox": [ + 46, + 414, + 286, + 498 + ], + "type": "text", + "content": " fix and " + }, + { + "bbox": [ + 46, + 414, + 286, + 498 + ], + "type": "inline_equation", + "content": "P1P2L" + }, + { + "bbox": [ + 46, + 414, + 286, + 498 + ], + "type": "text", + "content": " no fix in the coplanar case and solvers " + }, + { + "bbox": [ + 46, + 414, + 286, + 498 + ], + "type": "inline_equation", + "content": "P2P1L" + }, + { + "bbox": [ + 46, + 414, + 286, + 498 + ], + "type": "text", + "content": " no fix and " + }, + { + "bbox": [ + 46, + 414, + 286, + 498 + ], + "type": "inline_equation", + "content": "P1P2L" + }, + { + "bbox": [ + 46, + 414, + 286, + 498 + ], + "type": "text", + "content": " fix in the generic case." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 511, + 127, + 525 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 511, + 127, + 525 + ], + "spans": [ + { + "bbox": [ + 47, + 511, + 127, + 525 + ], + "type": "text", + "content": "3. Experiments" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 533, + 286, + 663 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 533, + 286, + 663 + ], + "spans": [ + { + "bbox": [ + 46, + 533, + 286, + 663 + ], + "type": "text", + "content": "In this section, we experimentally compare the proposed solvers with the state-of-the-art methods, specifically 3Q3 [17] and Ramalingam [25]. For 3Q3, we employ the publicly available implementation from Poselib. As the implementation of Ramalingam is not publicly accessible, we have created our own implementation. Since it is not explicitly specified which method should be used for solving the linear equations, we compare with two variants: SVD and LU decomposition. All the solvers have been implemented in " + }, + { + "bbox": [ + 46, + 533, + 286, + 663 + ], + "type": "inline_equation", + "content": "\\mathrm{C + + }" + }, + { + "bbox": [ + 46, + 533, + 286, + 663 + ], + "type": "text", + "content": ", and the experiments are conducted on a desktop computer with an AMD Ryzen 9 CPU with " + }, + { + "bbox": [ + 46, + 533, + 286, + 663 + ], + "type": "inline_equation", + "content": "3.9\\mathrm{GHz}" + }, + { + "bbox": [ + 46, + 533, + 286, + 663 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 665, + 286, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 286, + 712 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 286, + 712 + ], + "type": "text", + "content": "In Sec. 3.1, we provide an analysis of the numerical stability and runtime performance of the solvers on synthetic data. Subsequently, in Section 3.2, we present an evaluation of the solvers within the RANSAC scheme." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 72, + 432, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 432, + 84 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 432, + 84 + ], + "type": "text", + "content": "3.1. Synthetic experiments" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 91, + 545, + 125 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 91, + 545, + 125 + ], + "spans": [ + { + "bbox": [ + 305, + 91, + 545, + 125 + ], + "type": "text", + "content": "In this section, we present an analysis of the numerical stability and runtime performance of the minimal solvers on synthetic data." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 126, + 545, + 151 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 126, + 545, + 151 + ], + "spans": [ + { + "bbox": [ + 305, + 126, + 545, + 151 + ], + "type": "text", + "content": "We generate instances of each minimal problem (either P2P1L or P1P2L), according to the following procedure:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 153, + 545, + 369 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 306, + 153, + 545, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 153, + 545, + 213 + ], + "spans": [ + { + "bbox": [ + 306, + 153, + 545, + 213 + ], + "type": "text", + "content": "- Rotation Matrix " + }, + { + "bbox": [ + 306, + 153, + 545, + 213 + ], + "type": "inline_equation", + "content": "(R)" + }, + { + "bbox": [ + 306, + 153, + 545, + 213 + ], + "type": "text", + "content": ": An axis " + }, + { + "bbox": [ + 306, + 153, + 545, + 213 + ], + "type": "inline_equation", + "content": "v \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 306, + 153, + 545, + 213 + ], + "type": "text", + "content": " is sampled from the uniform distribution on the unit sphere, and an angle " + }, + { + "bbox": [ + 306, + 153, + 545, + 213 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 306, + 153, + 545, + 213 + ], + "type": "text", + "content": " is sampled from the normal distribution " + }, + { + "bbox": [ + 306, + 153, + 545, + 213 + ], + "type": "inline_equation", + "content": "\\mathcal{N}(0,1)" + }, + { + "bbox": [ + 306, + 153, + 545, + 213 + ], + "type": "text", + "content": ". The rotation matrix " + }, + { + "bbox": [ + 306, + 153, + 545, + 213 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 306, + 153, + 545, + 213 + ], + "type": "text", + "content": " is then constructed using the angle-axis formula, " + }, + { + "bbox": [ + 306, + 153, + 545, + 213 + ], + "type": "inline_equation", + "content": "R = I + \\sin(\\alpha)[v]_{\\times} + (1 - \\cos(\\alpha))[v]_{\\times}^2" + }, + { + "bbox": [ + 306, + 153, + 545, + 213 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 213, + 545, + 248 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 213, + 545, + 248 + ], + "spans": [ + { + "bbox": [ + 306, + 213, + 545, + 248 + ], + "type": "text", + "content": "- Translation " + }, + { + "bbox": [ + 306, + 213, + 545, + 248 + ], + "type": "inline_equation", + "content": "(T)" + }, + { + "bbox": [ + 306, + 213, + 545, + 248 + ], + "type": "text", + "content": ": The camera center " + }, + { + "bbox": [ + 306, + 213, + 545, + 248 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 306, + 213, + 545, + 248 + ], + "type": "text", + "content": " is sampled uniformly at random from the unit sphere. The translation vector " + }, + { + "bbox": [ + 306, + 213, + 545, + 248 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 306, + 213, + 545, + 248 + ], + "type": "text", + "content": " is computed as " + }, + { + "bbox": [ + 306, + 213, + 545, + 248 + ], + "type": "inline_equation", + "content": "T = -RC" + }, + { + "bbox": [ + 306, + 213, + 545, + 248 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 249, + 545, + 308 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 249, + 545, + 308 + ], + "spans": [ + { + "bbox": [ + 306, + 249, + 545, + 308 + ], + "type": "text", + "content": "- Point correspondence: A 3D point " + }, + { + "bbox": [ + 306, + 249, + 545, + 308 + ], + "type": "inline_equation", + "content": "X_{i}" + }, + { + "bbox": [ + 306, + 249, + 545, + 308 + ], + "type": "text", + "content": " is sampled from the trivariate normal distribution with mean vector " + }, + { + "bbox": [ + 306, + 249, + 545, + 308 + ], + "type": "inline_equation", + "content": "\\mu = [0,0,5]^T" + }, + { + "bbox": [ + 306, + 249, + 545, + 308 + ], + "type": "text", + "content": " and standard deviation " + }, + { + "bbox": [ + 306, + 249, + 545, + 308 + ], + "type": "inline_equation", + "content": "\\sigma = 1" + }, + { + "bbox": [ + 306, + 249, + 545, + 308 + ], + "type": "text", + "content": " in each component. The corresponding 2D point " + }, + { + "bbox": [ + 306, + 249, + 545, + 308 + ], + "type": "inline_equation", + "content": "D_{i}" + }, + { + "bbox": [ + 306, + 249, + 545, + 308 + ], + "type": "text", + "content": " is obtained by projecting " + }, + { + "bbox": [ + 306, + 249, + 545, + 308 + ], + "type": "inline_equation", + "content": "X_{i}" + }, + { + "bbox": [ + 306, + 249, + 545, + 308 + ], + "type": "text", + "content": " onto a pinhole camera with the pose " + }, + { + "bbox": [ + 306, + 249, + 545, + 308 + ], + "type": "inline_equation", + "content": "(R,T)" + }, + { + "bbox": [ + 306, + 249, + 545, + 308 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 309, + 545, + 369 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 309, + 545, + 369 + ], + "spans": [ + { + "bbox": [ + 306, + 309, + 545, + 369 + ], + "type": "text", + "content": "- Line correspondence: Two 3D points, " + }, + { + "bbox": [ + 306, + 309, + 545, + 369 + ], + "type": "inline_equation", + "content": "L_{i}" + }, + { + "bbox": [ + 306, + 309, + 545, + 369 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 306, + 309, + 545, + 369 + ], + "type": "inline_equation", + "content": "L_{i+1}" + }, + { + "bbox": [ + 306, + 309, + 545, + 369 + ], + "type": "text", + "content": ", are sampled as described in the previous bullet-point. These points define the 3D line " + }, + { + "bbox": [ + 306, + 309, + 545, + 369 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 306, + 309, + 545, + 369 + ], + "type": "text", + "content": ". Two 2D points, " + }, + { + "bbox": [ + 306, + 309, + 545, + 369 + ], + "type": "inline_equation", + "content": "D_{i}" + }, + { + "bbox": [ + 306, + 309, + 545, + 369 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 306, + 309, + 545, + 369 + ], + "type": "inline_equation", + "content": "D_{i+1}" + }, + { + "bbox": [ + 306, + 309, + 545, + 369 + ], + "type": "text", + "content": ", are obtained by sampling two points on the 3D line " + }, + { + "bbox": [ + 306, + 309, + 545, + 369 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 306, + 309, + 545, + 369 + ], + "type": "text", + "content": " and projecting them onto the camera with the pose " + }, + { + "bbox": [ + 306, + 309, + 545, + 369 + ], + "type": "inline_equation", + "content": "(R, T)" + }, + { + "bbox": [ + 306, + 309, + 545, + 369 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 304, + 371, + 545, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 371, + 545, + 529 + ], + "spans": [ + { + "bbox": [ + 304, + 371, + 545, + 529 + ], + "type": "text", + "content": "In the case of P2P1L, two points and one line are sampled, while in the case of P1P2L, one point and two lines are sampled. Let " + }, + { + "bbox": [ + 304, + 371, + 545, + 529 + ], + "type": "inline_equation", + "content": "(R_{est},T_{est})" + }, + { + "bbox": [ + 304, + 371, + 545, + 529 + ], + "type": "text", + "content": " be the pose obtained by the minimal solver. We measure the rotation error as the angle " + }, + { + "bbox": [ + 304, + 371, + 545, + 529 + ], + "type": "inline_equation", + "content": "\\xi_R = \\arccos ((1 + \\mathrm{trace}(R_{est}^T R)) / 2)" + }, + { + "bbox": [ + 304, + 371, + 545, + 529 + ], + "type": "text", + "content": ". We calculate the translation error " + }, + { + "bbox": [ + 304, + 371, + 545, + 529 + ], + "type": "inline_equation", + "content": "\\xi_T" + }, + { + "bbox": [ + 304, + 371, + 545, + 529 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 304, + 371, + 545, + 529 + ], + "type": "inline_equation", + "content": "\\xi_T = \\frac{|T_{est} - T|}{|T|}" + }, + { + "bbox": [ + 304, + 371, + 545, + 529 + ], + "type": "text", + "content": ". The histograms of rotation errors " + }, + { + "bbox": [ + 304, + 371, + 545, + 529 + ], + "type": "inline_equation", + "content": "(\\xi_R)" + }, + { + "bbox": [ + 304, + 371, + 545, + 529 + ], + "type": "text", + "content": " and translation errors " + }, + { + "bbox": [ + 304, + 371, + 545, + 529 + ], + "type": "inline_equation", + "content": "(\\xi_T)" + }, + { + "bbox": [ + 304, + 371, + 545, + 529 + ], + "type": "text", + "content": " for all considered solvers are depicted in Figure 7. Furthermore, summary statistics including the median, mean, and maximum errors are provided in Table 4. The results indicate that all minimal solvers are stable. Our algebraically optimal solvers demonstrate superior stability compared to the previous solvers." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 530, + 545, + 614 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 530, + 545, + 614 + ], + "spans": [ + { + "bbox": [ + 304, + 530, + 545, + 614 + ], + "type": "text", + "content": "The runtime evaluation for all considered solvers is shown in Table. 3. Our P2P1L solver requires 313.8ns on average, which is about 6x faster than the 3Q3 solver [17]. Similarly, our P1P2L solver requires on average 504.0ns, which is about 4x faster than the 3Q3 solver. These results show a significant speedup achieved by our solvers in comparison to the previous methods." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 306, + 623, + 448, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 623, + 448, + 635 + ], + "spans": [ + { + "bbox": [ + 306, + 623, + 448, + 635 + ], + "type": "text", + "content": "3.2. Experiments in RANSAC" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "type": "text", + "content": "Here, we provide an analysis of the minimal solvers within the RANSAC scheme [4], using the Oxford Multi-view Data [1]. This dataset includes a variety of indoor and outdoor scenes, some of which contain both point and line matches. An example of images from this dataset is illustrated in Figure 8." + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "21322" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 95, + 69, + 496, + 158 + ], + "blocks": [ + { + "bbox": [ + 95, + 69, + 496, + 158 + ], + "lines": [ + { + "bbox": [ + 95, + 69, + 496, + 158 + ], + "spans": [ + { + "bbox": [ + 95, + 69, + 496, + 158 + ], + "type": "table", + "html": "
MethodMean RMedian RMax RMean TMedian TMax T
P2P1L Ours5.3e-121.4e-151.2e-073.7e-102.1e-142.2e-05
P2P1L 3Q32.8e-056.5e-152.72.0e-051.2e-131.04
P2P1L Ramalingam (SVD)4.9e-091.5e-140.000411.6e-072.7e-130.0099
P2P1L Ramalingam (LU)4.7e-071.3e-140.0402.3e-052.4e-130.10
PP1P2L Ours (fix 2)1.2e-074.4e-150.0102.0e-067.1e-140.13
P1P2L 3Q33.3e-057.2e-152.603.4e-051.5e-131.01
", + "image_path": "46608815a26be9030d013de8b62756e2bfe2fc546a4f5ba225ed55073d869966.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 73, + 182, + 519, + 315 + ], + "blocks": [ + { + "bbox": [ + 47, + 159, + 545, + 171 + ], + "lines": [ + { + "bbox": [ + 47, + 159, + 545, + 171 + ], + "spans": [ + { + "bbox": [ + 47, + 159, + 545, + 171 + ], + "type": "text", + "content": "Table 4. The mean, median and max values of the rot.(R) and tran.(T) errors over " + }, + { + "bbox": [ + 47, + 159, + 545, + 171 + ], + "type": "inline_equation", + "content": "{1e5}" + }, + { + "bbox": [ + 47, + 159, + 545, + 171 + ], + "type": "text", + "content": " noiseless samples. The best results are marked bold." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 73, + 182, + 519, + 315 + ], + "lines": [ + { + "bbox": [ + 73, + 182, + 519, + 315 + ], + "spans": [ + { + "bbox": [ + 73, + 182, + 519, + 315 + ], + "type": "table", + "html": "
DatasetP2P1LP1P2L
OUR3Q3Ram. SVDRam. LUOUR3Q3
Model House7.72 (0.85x)9.09 (1x)20.69 (2.28x)17.30 (1.90x)8.06 (0.85x)9.52 (1x)
Corridor11.32 (0.91x)12.46 (1x)26.56 (2.13x)23.54 (1.89x)13.3 (0.90x)14.76 (1x)
Merton I35.93 (0.98x)36.83 (1x)77.96 (2.12x)74.87 (2.03x)28.85 (0.64x)45.41 (1x)
Merton II33.30 (0.97x)34.27 (1x)74.37 (2.17x)71.35 (2.08x)26.7 (0.67x)39.64 (1x)
Merton III24.04 (0.96x)25.07 (1x)52.07 (2.08x)48.97 (1.95x)10.91 (0.37x)29.67 (1x)
Library32.42 (0.97x)33.31 (1x)69.29 (2.08x)66.23 (1.99x)10.04 (0.28x)35.98 (1x)
Wadham39.60 (0.98x)40.51 (1x)86.10 (2.13x)83.00 (2.05x)22.96 (0.45x)50.79 (1x)
Avg. Speed-up0.94x1x2.14x1.99x0.59x1x
", + "image_path": "5a89bfd738c035f70daacdbac4faeb7fe21f09b1563c076422fe95f0fd737054.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 70, + 319, + 522, + 331 + ], + "lines": [ + { + "bbox": [ + 70, + 319, + 522, + 331 + ], + "spans": [ + { + "bbox": [ + 70, + 319, + 522, + 331 + ], + "type": "text", + "content": "Table 5. RANSAC timing, on Oxford Multi-view dataset [1], in milliseconds. Speed-up compared to the 3Q3 in the bracket." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 72, + 350, + 161, + 418 + ], + "blocks": [ + { + "bbox": [ + 72, + 350, + 161, + 418 + ], + "lines": [ + { + "bbox": [ + 72, + 350, + 161, + 418 + ], + "spans": [ + { + "bbox": [ + 72, + 350, + 161, + 418 + ], + "type": "image", + "image_path": "1a1467d082b96b6409e6de7ca27d3e1e313ffb3301faa7f70a5f7a861fa0ba8f.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 170, + 350, + 260, + 417 + ], + "blocks": [ + { + "bbox": [ + 170, + 350, + 260, + 417 + ], + "lines": [ + { + "bbox": [ + 170, + 350, + 260, + 417 + ], + "spans": [ + { + "bbox": [ + 170, + 350, + 260, + 417 + ], + "type": "image", + "image_path": "c0a5efd34e5788c9b8db10104226b00eadb0bafe109346eed28a4f5da64ce227.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 72, + 418, + 162, + 486 + ], + "blocks": [ + { + "bbox": [ + 72, + 418, + 162, + 486 + ], + "lines": [ + { + "bbox": [ + 72, + 418, + 162, + 486 + ], + "spans": [ + { + "bbox": [ + 72, + 418, + 162, + 486 + ], + "type": "image", + "image_path": "937618e532ccbafba2a2787bad9aa48be71248dcdeeb2a2423e084d01d652d76.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 47, + 489, + 287, + 512 + ], + "lines": [ + { + "bbox": [ + 47, + 489, + 287, + 512 + ], + "spans": [ + { + "bbox": [ + 47, + 489, + 287, + 512 + ], + "type": "text", + "content": "Figure 8. Example images of University Library (top) and Merton College III (bottom), from [1]." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 171, + 418, + 260, + 486 + ], + "blocks": [ + { + "bbox": [ + 171, + 418, + 260, + 486 + ], + "lines": [ + { + "bbox": [ + 171, + 418, + 260, + 486 + ], + "spans": [ + { + "bbox": [ + 171, + 418, + 260, + 486 + ], + "type": "image", + "image_path": "7774b5909468adbd6f981ab9a5bb9161a98e5523a8b2af6612a8aaeaecc5f8f2.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 533, + 287, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 533, + 287, + 640 + ], + "spans": [ + { + "bbox": [ + 46, + 533, + 287, + 640 + ], + "type": "text", + "content": "We utilize the Poselib [18] implementation of LORANSAC [6]. In this implementation, local optimization is applied both when the new solution surpasses the currently best one and at the end of the entire procedure. The RANSAC parameters used include a maximum number of iterations set to 100000, a minimum number of iterations set to 1000, a success probability of 0.9999, and an inlier threshold of 1 pixel. To ensure a fair comparison between the solvers, we use a fixed random seed." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 642, + 288, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 642, + 288, + 712 + ], + "spans": [ + { + "bbox": [ + 46, + 642, + 288, + 712 + ], + "type": "text", + "content": "Since all the solvers are stable, we do not expect any significant difference in the estimated pose when using different solvers. Therefore, our primary focus is on comparing the runtime of the RANSAC scheme. The runtime comparison is presented in Table 5, which displays the average runtime for each scene in the Oxford multiview dataset and" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 352, + 545, + 507 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 352, + 545, + 507 + ], + "spans": [ + { + "bbox": [ + 304, + 352, + 545, + 507 + ], + "type": "text", + "content": "for each considered solver. Additionally, the table includes a speedup value compared to the 3Q3 solver of [18]. The results show that our solvers consistently outperform the 3Q3 [17] and Ramalingam [25] solvers in terms of runtime. However, there is a notable variation in the speed-up among the scenes, which ranges from 0.85 to 0.98 for the P2P1L problem and from 0.28 to 0.90 for the P1P2L solver. This variation can be attributed to the fact that different scenes require varying proportions of time spent by RANSAC on scoring hypothesized models and local optimization. The more time is dedicated to these tasks, the less significant speedup can be reached. In cases with fewer matches or a lower inlier ratio, more substantial speedups are possible." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 508, + 545, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 508, + 545, + 556 + ], + "spans": [ + { + "bbox": [ + 304, + 508, + 545, + 556 + ], + "type": "text", + "content": "The final pose estimation errors in RANSAC are shown in the Supplementary. As expected, the errors are the same for all solvers. This demonstrates that we can achieve improved runtimes without sacrificing any accuracy." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 564, + 378, + 576 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 564, + 378, + 576 + ], + "spans": [ + { + "bbox": [ + 306, + 564, + 378, + 576 + ], + "type": "text", + "content": "4. Conclusion" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 582, + 545, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 582, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 304, + 582, + 545, + 689 + ], + "type": "text", + "content": "In revisiting absolute pose from mixed point-line correspondences, we have developed new solvers, which have the merit of being both algebraically optimal and elementary in nature. Moreover, our solvers outperform the state-of-the-art in terms of both runtime and numerical stability. Despite the simplicity of many minimal absolute pose problems, our findings suggest that their continued investigation remains worthwhile. The code is available at https://github.com/petrhruby97/efficientAbsolutely" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 689, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 545, + 712 + ], + "type": "text", + "content": "Acknowledgements: TD was supported by NSF DMS 2103310." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "21323" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 286, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 53, + 91, + 286, + 123 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 286, + 123 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 286, + 123 + ], + "type": "text", + "content": "[1] Multi-view data, oxford visual geometry group. https://www.robots.ox.ac.uk/vgg/data/mview/. Accessed: Nov 8 2023. 7, 8, 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 125, + 286, + 178 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 125, + 286, + 178 + ], + "spans": [ + { + "bbox": [ + 53, + 125, + 286, + 178 + ], + "type": "text", + "content": "[2] Sérgio Agostinho, João Gomes, and Alessio Del Bue. CvxPnPL: A unified convex solution to the absolute pose estimation problem from point and line correspondences. Journal of Mathematical Imaging and Vision, 65(3):492-512, 2023. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 182, + 286, + 214 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 182, + 286, + 214 + ], + "spans": [ + { + "bbox": [ + 53, + 182, + 286, + 214 + ], + "type": "text", + "content": "[3] Daniel Barath and Jiri Matas. Graph-cut RANSAC: local optimization on spatially coherent structures. IEEE Trans. Pattern Anal. Mach. Intell., 44(9):4961-4974, 2022. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 217, + 286, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 217, + 286, + 281 + ], + "spans": [ + { + "bbox": [ + 53, + 217, + 286, + 281 + ], + "type": "text", + "content": "[4] Robert C. Bolles and Martin A. Fischler. A ransac-based approach to model fitting and its application to finding cylinders in range data. In Proceedings of the 7th International Joint Conference on Artificial Intelligence, IJCAI '81, Vancouver, BC, Canada, August 24-28, 1981, pages 637-643. William Kaufmann, 1981. 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 284, + 286, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 284, + 286, + 327 + ], + "spans": [ + { + "bbox": [ + 53, + 284, + 286, + 327 + ], + "type": "text", + "content": "[5] Homer H. Chen. Pose determination from line-to-plane correspondences: Existence condition and closed-form solutions. IEEE Trans. Pattern Anal. Mach. Intell., 13(6):530-541, 1991. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 329, + 286, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 329, + 286, + 373 + ], + "spans": [ + { + "bbox": [ + 53, + 329, + 286, + 373 + ], + "type": "text", + "content": "[6] Ondrej Chum, Jiri Matas, and Josef Kittler. Locally optimized RANSAC. In Pattern Recognition, 25th DAGM Symposium, Magdeburg, Germany, September 10-12, 2003, Proceedings, pages 236-243. Springer, 2003. 8" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 375, + 286, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 375, + 286, + 418 + ], + "spans": [ + { + "bbox": [ + 53, + 375, + 286, + 418 + ], + "type": "text", + "content": "[7] Michel Dhome, Marc Richetin, Jean-Thierry Lapresté, and Gérard Rives. Determination of the attitude of 3d objects from a single perspective view. IEEE Trans. Pattern Anal. Mach. Intell., 11(12):1265-1278, 1989. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 420, + 286, + 474 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 420, + 286, + 474 + ], + "spans": [ + { + "bbox": [ + 53, + 420, + 286, + 474 + ], + "type": "text", + "content": "[8] Yaqing Ding, Jian Yang, Viktor Larsson, Carl Olsson, and Kalle Åström. Revisiting the P3P problem. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2023, Vancouver, BC, Canada, June 17-24, 2023, pages 4872-4880. IEEE, 2023. 1, 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 476, + 286, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 476, + 286, + 520 + ], + "spans": [ + { + "bbox": [ + 53, + 476, + 286, + 520 + ], + "type": "text", + "content": "[9] Timothy Duff, Viktor Korotynskiy, Tomas Pajdla, and Margaret H. Regan. Galois/monodromy groups for decomposing minimal problems in 3d reconstruction. SIAM Journal on Applied Algebra and Geometry, 6(4):740-772, 2022. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 522, + 286, + 565 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 522, + 286, + 565 + ], + "spans": [ + { + "bbox": [ + 48, + 522, + 286, + 565 + ], + "type": "text", + "content": "[10] Ricardo Fabbri, Peter J. Giblin, and Benjamin B. Kimia. Camera pose estimation using first-order curve differential geometry. IEEE Trans. Pattern Anal. Mach. Intell., 43(10): 3321-3332, 2021. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 567, + 286, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 567, + 286, + 610 + ], + "spans": [ + { + "bbox": [ + 48, + 567, + 286, + 610 + ], + "type": "text", + "content": "[11] Martin A. Fischler and Robert C. Bolles. Random sample consensus: A paradigm for model fitting with applications to image analysis and automated cartography. Commun. ACM, 24(6):381-395, 1981. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 613, + 286, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 613, + 286, + 667 + ], + "spans": [ + { + "bbox": [ + 48, + 613, + 286, + 667 + ], + "type": "text", + "content": "[12] Christian Hane, Lionel Heng, Gim Hee Lee, Friedrich Fraundorfer, Paul Furgale, Torsten Sattler, and Marc Pollefeys. 3d visual perception for self-driving cars using a multi-camera system: Calibration, mapping, localization, and obstacle detection. Image Vis. Comput., 68:14-27, 2017. 1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 670, + 286, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 670, + 286, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 670, + 286, + 712 + ], + "type": "text", + "content": "[13] Robert M. Haralick, Chung-Nan Lee, Karsten Ottenberg, and Michael Nölle. Review and analysis of solutions of the three point perspective pose estimation problem. Int. J. Comput. Vis., 13(3):331-356, 1994. 1" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 138 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 138 + ], + "type": "text", + "content": "[14] Petr Hruby, Viktor Korotynskiy, Timothy Duff, Luke Oeding, Marc Pollefeys, Tomás Pajdla, and Viktor Larsson. Four-view geometry with unknown radial distortion. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2023, Vancouver, BC, Canada, June 17-24, 2023, pages 8990-9000. IEEE, 2023. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 140, + 545, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 140, + 545, + 182 + ], + "spans": [ + { + "bbox": [ + 307, + 140, + 545, + 182 + ], + "type": "text", + "content": "[15] Xiaomei Hu, Luying Zhu, Ping Wang, Haili Yang, and Xuan Li. Improved ORB-SLAM2 mobile robot vision algorithm based on multiple feature fusion. IEEE Access, 11:100659-100671, 2023. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 184, + 545, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 184, + 545, + 259 + ], + "spans": [ + { + "bbox": [ + 308, + 184, + 545, + 259 + ], + "type": "text", + "content": "[16] Laurent Kneip, Davide Scaramuzza, and Roland Siegwart. A novel parametrization of the perspective-three-point problem for a direct computation of absolute camera position and orientation. In The 24th IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2011, Colorado Springs, CO, USA, 20-25 June 2011, pages 2969-2976. IEEE Computer Society, 2011. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 261, + 545, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 261, + 545, + 326 + ], + "spans": [ + { + "bbox": [ + 308, + 261, + 545, + 326 + ], + "type": "text", + "content": "[17] Zuzana Kukelova, Jan Heller, and Andrew W. Fitzgibbon. Efficient intersection of three quadrics and applications in computer vision. In 2016 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2016, Las Vegas, NV, USA, June 27-30, 2016, pages 1799-1808. IEEE Computer Society, 2016. 2, 7, 8" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 327, + 545, + 347 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 327, + 545, + 347 + ], + "spans": [ + { + "bbox": [ + 308, + 327, + 545, + 347 + ], + "type": "text", + "content": "[18] Viktor Larsson and contributors. PoseLib - Minimal Solvers for Camera Pose Estimation, 2020. 2, 8" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 350, + 545, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 350, + 545, + 403 + ], + "spans": [ + { + "bbox": [ + 308, + 350, + 545, + 403 + ], + "type": "text", + "content": "[19] Shaohui Liu, Yifan Yu, Rémi Pautrat, Marc Pollefeys, and Viktor Larsson. 3d line mapping revisited. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2023, Vancouver, BC, Canada, June 17-24, 2023, pages 21445-21455. IEEE, 2023. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 404, + 545, + 469 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 404, + 545, + 469 + ], + "spans": [ + { + "bbox": [ + 308, + 404, + 545, + 469 + ], + "type": "text", + "content": "[20] David Nister, Richard I. Hartley, and Henrik Stewenius. Using galois theory to prove structure from motion algorithms are optimal. In 2007 IEEE Computer Society Conference on Computer Vision and Pattern Recognition (CVPR 2007), 18-23 June 2007, Minneapolis, Minnesota, USA. IEEE Computer Society, 2007. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 471, + 545, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 471, + 545, + 536 + ], + "spans": [ + { + "bbox": [ + 308, + 471, + 545, + 536 + ], + "type": "text", + "content": "[21] Rémi Pautrat, Daniel Barath, Viktor Larsson, Martin R. Oswald, and Marc Pollefeys. Deeplsd: Line segment detection and refinement with deep image gradients. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2023, Vancouver, BC, Canada, June 17-24, 2023, pages 17327-17336. IEEE, 2023. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 537, + 545, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 537, + 545, + 590 + ], + "spans": [ + { + "bbox": [ + 308, + 537, + 545, + 590 + ], + "type": "text", + "content": "[22] Rémi Pautrat, Iago Suárez, Yifan Yu, Marc Pollefeys, and Viktor Larsson. Gluestick: Robust image matching by sticking points and lines together. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9706-9716, 2023. 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 592, + 545, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 592, + 545, + 646 + ], + "spans": [ + { + "bbox": [ + 308, + 592, + 545, + 646 + ], + "type": "text", + "content": "[23] Mikael Persson and Klas Nordberg. Lambda twist: An accurate fast robust perspective three point (P3P) solver. In Computer Vision - ECCV 2018 - 15th European Conference, Munich, Germany, September 8-14, 2018, Proceedings, Part IV, pages 334-349. Springer, 2018. 1, 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 647, + 545, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 647, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 308, + 647, + 545, + 689 + ], + "type": "text", + "content": "[24] Rahul Raguram, Ondrej Chum, Marc Pollefeys, Jiri Matas, and Jan-Michael Frahm. USAC: A universal framework for random sample consensus. IEEE Trans. Pattern Anal. Mach. Intell., 35(8):2022-2038, 2013. 1" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 692, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 692, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 692, + 545, + 713 + ], + "type": "text", + "content": "[25] Srikumar Ramalingam, Sofien Bouaziz, and Peter F. Sturm. Pose estimation using both points and lines for geo-" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "21324" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 364 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 66, + 72, + 287, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 72, + 287, + 106 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 287, + 106 + ], + "type": "text", + "content": "localization. In IEEE International Conference on Robotics and Automation, ICRA 2011, Shanghai, China, 9-13 May 2011, pages 4716-4723. IEEE, 2011. 1, 2, 3, 4, 5, 7, 8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 107, + 287, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 107, + 287, + 150 + ], + "spans": [ + { + "bbox": [ + 48, + 107, + 287, + 150 + ], + "type": "text", + "content": "[26] Torsten Sattler, Bastian Leibe, and Leif Kobbelt. Efficient & effective prioritized matching for large-scale image-based localization. IEEE Trans. Pattern Anal. Mach. Intell., 39(9): 1744-1756, 2017. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 152, + 287, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 152, + 287, + 207 + ], + "spans": [ + { + "bbox": [ + 48, + 152, + 287, + 207 + ], + "type": "text", + "content": "[27] Johannes L. Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In 2016 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2016, Las Vegas, NV, USA, June 27-30, 2016, pages 4104-4113. IEEE Computer Society, 2016. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 208, + 287, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 208, + 287, + 251 + ], + "spans": [ + { + "bbox": [ + 48, + 208, + 287, + 251 + ], + "type": "text", + "content": "[28] Jonathan Ventura, Clemens Arth, Gerhard Reitmayr, and Dieter Schmalstieg. Global localization from monocular SLAM on a mobile phone. IEEE Trans. Vis. Comput. Graph., 20(4):531-539, 2014. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 253, + 287, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 253, + 287, + 296 + ], + "spans": [ + { + "bbox": [ + 48, + 253, + 287, + 296 + ], + "type": "text", + "content": "[29] Chi Xu, Lilian Zhang, Li Cheng, and Reinhard Koch. Pose estimation from line correspondences: A complete analysis and a series of solutions. IEEE Trans. Pattern Anal. Mach. Intell., 39(6):1209-1222, 2017. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 297, + 287, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 297, + 287, + 364 + ], + "spans": [ + { + "bbox": [ + 48, + 297, + 287, + 364 + ], + "type": "text", + "content": "[30] Lipu Zhou, Jiamin Ye, and Michael Kaess. A stable algebraic camera pose estimation for minimal configurations of 2d/3d point and line correspondences. In Computer Vision - ACCV 2018 - 14th Asian Conference on Computer Vision, Perth, Australia, December 2-6, 2018, Revised Selected Papers, Part IV, pages 273-288. Springer, 2018. 1, 2" + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "21325" + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Efficient Stitchable Task Adaptation/03126c3f-71ba-4532-921c-1c95d792e515_content_list.json b/2024/Efficient Stitchable Task Adaptation/03126c3f-71ba-4532-921c-1c95d792e515_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..e917b438365f2afcdec2ded0be6efdb1a445c841 --- /dev/null +++ b/2024/Efficient Stitchable Task Adaptation/03126c3f-71ba-4532-921c-1c95d792e515_content_list.json @@ -0,0 +1,1637 @@ +[ + { + "type": "text", + "text": "Efficient Stitchable Task Adaptation", + "text_level": 1, + "bbox": [ + 300, + 130, + 669, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Haoyu He Zizheng Pan Jing Liu Jianfei Cai Bohan Zhuang", + "bbox": [ + 212, + 180, + 763, + 200 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ZIP Lab, Monash University, Australia", + "bbox": [ + 326, + 205, + 637, + 223 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 257, + 313, + 273 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The paradigm of pre-training and fine-tuning has laid the foundation for deploying deep learning models. However, most fine-tuning methods are designed to meet a specific resource budget. Recently, considering diverse deployment scenarios with various resource budgets, SN-Net [51] is introduced to quickly obtain numerous new networks (stitches) from the pre-trained models (anchors) in a model family via model stitching. Although promising, SN-Net confronts new challenges when adapting it to new target domains, including huge memory and storage requirements and a long and sub-optimal multistage adaptation process. In this work, we present a novel framework, Efficient Stitchable Task Adaptation (ESTA), to efficiently produce a palette of fine-tuned models that adhere to diverse resource constraints. Specifically, we first tailor parameter-efficient fine-tuning to share low-rank updates among the stitches while maintaining independent bias terms. In this way, we largely reduce fine-tuning memory burdens and mitigate the interference among stitches that arises in task adaptation. Furthermore, we streamline a simple yet effective one-stage deployment pipeline, which estimates the important stitches to deploy with training-time gradient statistics. By assigning higher sampling probabilities to important stitches, we also get a boosted Pareto frontier. Extensive experiments on 25 downstream visual recognition tasks demonstrate that our ESTA is capable of generating stitches with smooth accuracy-efficiency trade-offs and surpasses the direct SN-Net adaptation by remarkable margins with significantly lower training time and fewer trainable parameters. Furthermore, we demonstrate the flexibility and scalability of our ESTA framework by stitching LLMs from LLaMA family, obtaining chatbot stitches of assorted sizes1.", + "bbox": [ + 75, + 289, + 472, + 773 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 800, + 209, + 815 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The paradigm of pre-training and fine-tuning has underpinned modern applications in both vision and lan", + "bbox": [ + 75, + 825, + 468, + 854 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "guage. With off-the-shelf models pre-trained on large-scale datasets, the de-facto choice is vanilla full fine-tuning, which tunes all the model parameters with the downstream data. To reduce memory footprint and avoid overfitting, an emerging trend is to tune a small proportion of the model parameters while freezing the majority ones with Parameter-Efficient Fine-Tuning (PEFT) [27, 28, 31]. However, both full fine-tuning and PEFT target an exclusive specific resource budget for each downstream task, while in reality, we often need to deploy multiple models simultaneously to meet various resource demands. This makes us ponder: What is an effective and efficient way to obtain a palette of fine-tuned models meeting different resource constraints?", + "bbox": [ + 500, + 258, + 893, + 470 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "A natural approach is to compress a well-trained large model into numerous smaller ones [11, 21, 29, 36]. However, the computational cost grows linearly with the number of deployment scenarios. Following the once-for-all network [8, 20], a few works first pre-train a weight-sharing over-parameterized supernet, then adapt the supernet to the downstream tasks. Although promising, training supernets at scale with large datasets requires prohibitive computational resources, e.g., thousands of GPU training hours, which is practically infeasible.", + "bbox": [ + 496, + 474, + 893, + 626 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "On the other hand, there are many open-source large models [5, 23, 64] from communities such as Hugging-Face [72] that are pre-trained on large-scale datasets and ready to be downloaded. Inspired by the success of model stitching [4, 12, 38], recently, Stitchable Neural Networks (SN-Net) [51] has been proposed to stitch pre-trained models (anchors) of the same family to quickly obtain a set of candidate new networks (stitches) with different accuracy-efficiency trade-offs of a wide FLOPs range. However, the scope of SN-Net is limited to the pre-training classification task on the same source domain [56]. When directly adapting the standard approach of SN-Net to a target domain, it faces new challenges. Specifically, the straightforward way to employ SN-Net in task adaptation typically requires three stages: first adapting anchors individually to the target domain, followed by stitching the adapted anchors, and finally evaluating all stitches to find and deploy the ones on the Pareto Frontier. Such a three-stage adaptation pro", + "bbox": [ + 496, + 628, + 895, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "†Corresponding author. E-mail: bohan.zhuang@gmail.com", + "bbox": [ + 94, + 862, + 421, + 875 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "1Source code will be released at https://github.com/ziplab/Stitched_LLaMA.", + "bbox": [ + 80, + 875, + 468, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "28555", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "table", + "img_path": "images/f7a9d9a85b8f7e89ffaed77ac90fa81aa88e8ccc66740e8c713227c8b4e7efd5.jpg", + "table_caption": [ + "Table 1. Comparison of frameworks for obtaining a palette of networks to meet different hardware efficiency constraints. Supernet-based methods NAT [47] and TOFA [35] necessitate a supernet training stage plus long deployment GPU hours and are restricted to small FLOPs ranges. SN-Net [51] employs model stitching to achieve a wide FLOPs range with significantly reduced deployment GPU hours. Our ESTA framework further reduces the deployment GPU hours, fine-tuning memory, and trainable parameters, and is scalable to produce LLM stitches with billions of parameters (Section 5.2). The deployment GPU hours and fine-tuning memory for both SN-Net and ESTA are measured on a single NVIDIA GeForce RTX 3090 GPU when stitching ViT-Ti/S/B anchors [16] with batch size 64. * indicates data is inferred from the training recipe of TOFA's supernet architecture FBNetV3 [14]." + ], + "table_footnote": [], + "table_body": "
MethodAdapting supernetDeployment GPU hoursFine-tuning memoryTrainable parameter #FLOPs range (G)LLM friendly
NAT [47]X>1,000->10M[0.2, 0.6]X
TOFA [35]✓>1,000*->10M*[0.2, 2.5]X
SN-Net [51]X19.313,235M124.2M[1.3, 17.6]X
ESTA (ours)X5.09,685M4.6M[1.3, 17.6]✓
", + "bbox": [ + 81, + 191, + 890, + 270 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "cess is expensive and sub-optimal. Moreover, adapting SN-Net typically needs to load and optimize all parameters of multiple models, leading to a daunting training-time memory cost. For instance, stitching the two smallest models in the LLaMA family [64, 64] with SN-Net under standard training settings (32-bit floating point parameters and Adam optimizer on a single GPU) requires more than 80G Video RAM that cannot fit into high-end GPUs such as NVIDIA A100. SN-Net also requires saving a separate instance of the model family for each task, making the storage grow linearly with the number of deployed applications.", + "bbox": [ + 75, + 276, + 472, + 444 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Therefore, in this paper, we introduce a novel Efficient Stitchable Task Adaptation (ESTA) framework. Specifically, built upon SN-Net [51], ESTA introduces two main designs to overcome the aforementioned issues. First, to reduce the memory and storage costs, we tailor a Parameter-efficient Stitch fine-Tuning (PST) method, which incorporates the representative sparse fine-tuning technique LoRA [28] to keep the anchors and stitching layer weights frozen, while approximating their updates with trainable low-rank decomposition matrices. In addition to stitch-agnostic LoRA modules, we further introduce stitch-specific bias terms to alleviate the conflict among different stitches. Second, we propose a simple one-stage deployment pipeline to simultaneously adapt and stitch the pretrained anchors to the target domain. Moreover, we propose to estimate and accumulate an importance score for each stitch via the saliency pruning criterion [37]. With the importance scores, we develop a novel stitch-sampling method to assign important stitches with higher sampling probabilities. More importantly, after fine-tuning, the stitch importance scores can be directly used to infer the best ones at the Pareto frontier for deployment. Table 1 compares different efficient frameworks for diverse deployments.", + "bbox": [ + 75, + 445, + 472, + 792 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Overall, our paper has the following major contributions. 1) This is a pioneering work to investigate the problem of efficient fine-tuning of SN-Net for task adaptation. Our solution ESTA is a successful endeavor that delivers dozens of ready-to-deploy vision models up to 17.6G FLOPs in 5 GPU hours as well as multiple chatbot models with billions of parameters by stitching LLaMA mod", + "bbox": [ + 75, + 795, + 470, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "els [64, 65] for the instruction-following task. 2) We devise a parameter-efficient stitch fine-tuning method that incorporates trainable stitch-agnostic LoRA modules and stitch-specific bias terms, which largely reduces the fine-tuning memory footprints while alleviating the interference issue among stitches. Moreover, we streamline a simple one-stage deployment pipeline with a novel task-specific stitch sampling strategy that greatly reduces the deployment time and improves Pareto frontiers. 3) Extensive experiments conducted on 25 downstream visual recognition tasks demonstrate that when stitching ViT-Ti/S/B [16], our ESTA achieves remarkable performance improvements compared to the direct SN-Net adaptation counterpart while utilizing significantly fewer trainable parameters.", + "bbox": [ + 496, + 276, + 893, + 489 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 508, + 640, + 523 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Model stitching. Model stitching [4, 38] targets connecting the bottom layers of one network to the top layers of another with a stitching layer. Following [38], model stitching is first employed to measure the similarities for the inner representations learned by deep neural networks [4, 12, 25]. Previous works [4, 12] observe that trained networks with different initializations can be effectively stitched without incurring significant performance drop. Based on this observation, DeRy [76] stitches the blocks dissected from different pre-trained models and assembles a new model based on their quantified similarity for better performance. Herdt et al. [24] propose interpreting the inner representations of a deep network by stitching it with a pre-trained GAN generator. Recently, from the perspective of flexible deployment, SN-Net [51] stitches models of different sizes within a model family to cheaply produce numerous new networks with diverse accuracy and efficiency trade-offs in the pretraining task. In contrast to [51], our work specifically targets obtaining a palette of fine-tuned models for the task adaptation setting. To do so, we make key redesigns including a parameter-efficient stitch fine-tuning method and a simple one-stage deployment pipeline to overcome the efficiency challenges in SN-Net adaptation.", + "bbox": [ + 496, + 535, + 893, + 883 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Parameter-efficient fine-tuning. Parameter-efficient fine", + "bbox": [ + 500, + 885, + 890, + 901 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "28556", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "tuning [27, 28, 31, 40] is a powerful alternative to vanilla full tine-tuning, which updates only a small number of parameters while freezing the majority ones. Freezing the majority of parameters allows us to optimize storage and reduces the burden on training GPU memory, as there is no need to store their gradients or other training time statistics. Recent research freezes the vast majority of parameters while fine-tuning either the parameters that are inherited in the backbone [77, 80] or additionally added, in the form of learnable prompts [31, 39], low-rank bottleneck layers [27], learnable scaling and shifting factors [41], and separated small networks [61]. Meanwhile, PEFT has been employed in numerous downstream applications. For instance, while freezing the most parameters, Polyhistor [45] learns a hyper-network to generate adapter weights [27] for multi-task adaptation, LLaMA-Adapter [83] inserts learnable prompts into Transformer layers for the instruction-following task, and ControlNet [82] employs the zero-initialized convolutions to fine-tune diffusion generative models [55]. However, these works target only one exclusive resource budget and are not scalable to diverse deployment scenarios. Unlike other methods, our ESTA is crafted to yield multiple fine-tuned models with diverse capacities while jointly incorporating stitch-specific and stitch-agnostic lightweight trainable parameters.", + "bbox": [ + 76, + 90, + 472, + 468 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Fine-tuning with multiple models. The presence of more large-scale models has unleashed the potential to utilize multiple models instead of a single one during fine-tuning. One line of work selects the best pre-trained model in the model zoo to fine-tune. To identify the best model, prior studies quantize the transferability of the pre-trained models ahead of fine-tuning by estimating their accuracy on the downstream tasks [3, 17, 33] or the generalization capability to mitigate the domain gaps [6, 49, 66, 78]. Another way is to do model selection using an efficient online learning regime [74]. Considering the diversity among the pre-trained models, research efforts have been devoted to effectively exploit their knowledge with feature aggregation [42, 57, 58], model merging [13, 30, 48, 73], or a mixture-of-experts architecture [59]. In contrast to previous works that merely target better performance, our work employs off-the-shelf pre-trained model families to produce plentiful new models for diverse deployment requirements.", + "bbox": [ + 76, + 469, + 472, + 743 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Preliminaries", + "text_level": 1, + "bbox": [ + 76, + 758, + 215, + 773 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Model Stitching", + "text_level": 1, + "bbox": [ + 76, + 784, + 236, + 801 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Considering an $L$ -layer feed-forward artificial neural network $f_{\\theta_1}:\\mathcal{X}\\to \\mathcal{Y}$ parameterized by $\\theta_{1}$ , that maps any input from the input space $\\mathcal{X}$ to the output space $\\mathcal{V}$ , $f_{\\theta_1}$ can be denoted as a composition of functions that $f_{\\theta_1} = f_L\\circ \\dots \\circ f_1$ , where $\\circ$ denotes the function composition. In model stitching, $f_{\\theta_1}$ can be split up at the $l$ -th layer into two", + "bbox": [ + 76, + 809, + 470, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "portions of functions where $l\\in [1,L - 1]$ .Given any input $\\pmb {X}\\subseteq \\mathcal{X}$ , the first portion of the head functions maps $\\pmb{X}$ to activations $\\pmb {X}_l$ at layer $l$ , i.e., $H_{\\theta_1,l}(\\pmb {X}) = f_l\\circ \\dots \\circ f_1 = \\pmb {X}_l$ The second portion of the tail functions maps $\\pmb {X}_l$ to the final output, i.e., $T_{\\theta_1,l}(\\pmb {X}_l) = f_L\\circ \\dots \\circ f_{l + 1}$", + "bbox": [ + 496, + 90, + 890, + 167 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given another pre-trained $M$ -layer artificial neural network $f_{\\theta_2}$ parameterized by $\\theta_{2}$ that is split up at the $m$ -th layer that $m\\in [2,M]$ , we can then employ a stitching layer $S:\\mathcal{A}_{\\theta_1,l}\\to \\mathcal{A}_{\\theta_2,m}$ to map between the two activation spaces at layers $l$ and $m$ . In this case, we obtain a new network parameterized by $\\phi$ and the stitching layer parameters, i.e.,", + "bbox": [ + 496, + 167, + 890, + 271 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nF _ {\\phi , S} (\\boldsymbol {X}) = T _ {\\theta_ {2}, m} \\circ S \\circ H _ {\\theta_ {1}, l} (\\boldsymbol {X}), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 576, + 272, + 890, + 287 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\phi$ consists of partial parameters from both $\\theta_{1}$ and $\\theta_{2}$ .", + "bbox": [ + 496, + 287, + 890, + 301 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Stitchable Neural Network", + "text_level": 1, + "bbox": [ + 498, + 309, + 741, + 324 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In Stitchable Neural Network (SN-Net) [51], let a pretrained model family of size $Z$ be $\\mathcal{Z} = \\{f_{\\theta_z}\\}_{z=1}^Z$ , where $\\theta_z$ is the model parameter of the $z$ -th model. The goal is to derive additional $N$ candidate new networks $\\mathcal{N} = \\{F_{\\phi_n, S_n}\\}_{n=1}^N$ to adapt to various resource constraints. To do so, SN-Net first selects a pair of pre-trained models (anchors) from $\\mathcal{Z}$ and then stitches them with Eq. (1) at different layer indexes to get $N$ stitches. Then, SN-Net samples the stitches randomly and jointly optimizes them. An overview of SN-Net is depicted in Figure 1 (a). Since the anchors vary in scale, the newly assembled stitches have diverse performances and complexities. Importantly, SN-Net gives practical principles to design the space of $\\mathcal{N}$ that one should 1) stitch a pair of nearest anchors in terms of model complexity in a model family; 2) stitch the head of a faster and smaller anchor to the tail of a larger and slower anchor. Unless specified otherwise, we adopt these as the default experimental settings in this paper.", + "bbox": [ + 496, + 332, + 893, + 604 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4. Method", + "text_level": 1, + "bbox": [ + 498, + 616, + 591, + 632 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we introduce our ESTA framework, which consists of two major components: parameter-efficient stitch fine-tuning (PST) and a simple one-stage deployment pipeline. PST is tailored to cheaply fine-tune a palette of stitches (Section 4.1 and Figure 1 (b)). The one-stage deployment pipeline aims to simultaneously save deployment time and improve adaptation performance (Section 4.2).", + "bbox": [ + 496, + 642, + 893, + 750 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4.1. Parameter-efficient Stitch Fine-tuning", + "text_level": 1, + "bbox": [ + 498, + 756, + 830, + 773 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To address the problems of large storage and memory consumption when adapting SN-Net to downstream tasks with full fine-tuning, we introduce a parameter-efficient stitch fine-tuning (PST) method as depicted in Figure 1 (b). Our basic idea is to adapt the representative sparse fine-tuning method LoRA [28] for SN-Net fine-tuning.", + "bbox": [ + 496, + 780, + 890, + 869 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "LoRA on Transformer layers. Specifically, let any pretrained weight matrix in the multi-head self-attention layer", + "bbox": [ + 496, + 869, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "28557", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/f58364be157338b8aede459082270502387115ac7876c1765f57901f865d8003.jpg", + "image_caption": [ + "(a)" + ], + "image_footnote": [], + "bbox": [ + 83, + 93, + 302, + 196 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/e73f4f8d58c28c8099c1a7173896f46f9c4ebe670937baed455fa23fa39f713d.jpg", + "image_caption": [ + "(b)" + ], + "image_footnote": [], + "bbox": [ + 313, + 93, + 625, + 198 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/73e269c68c0a28a129cb02ef51ae34e723c6d46556b82f43ab9cbab09c2b5e86.jpg", + "image_caption": [ + "(c)" + ], + "image_footnote": [], + "bbox": [ + 635, + 102, + 880, + 198 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/61edbb8eb4bb41ca5c949343fd8c80b8b6d8584515dc95f17e9a0378ea3f20db.jpg", + "image_caption": [ + "Figure 1. (a) Illustration of Stretchable Neural Network [51]. With two anchors from the same model family, SN-Net connects the early layers of the smaller one to the latter layers of the larger one with stitching layers to obtain a set of new networks with different performance-efficiency trade-offs, e.g., the path in Blue. (b) Overview of our PST method tailored for fine-tuning a palette of stitches, which integrates stitch-agnostic LoRA modules with stitch-specific bias terms, aiming to promote diverse representations among stitches while maintaining low trainable parameters. (c) Overview of our task-specific stitch sampling. We estimate the importance scores of the stitches with a scoring function $Q(\\cdot, \\cdot)$ and accumulate them as global statistics with moving averages. For a resource constraint $\\gamma$ , we sample with a categorical distribution $\\pi(\\mathcal{N}_{\\gamma})$ that is parameterized by the normalized importance scores so as to assign the important stitches with higher sampling probabilities. After fine-tuning, we directly deploy the stitches with the highest scores to avoid the costly evaluation stage." + ], + "image_footnote": [], + "bbox": [ + 78, + 329, + 272, + 445 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/51758cda33ca580a38fa388b6ee0a71a8833b34eccb6413244410d7595ac2eae.jpg", + "image_caption": [ + "Figure 2. Distribution of pair-wise gradient angles among stitches when updating shared weights at fine-tuning iteration 600. We highlight angle $90^{\\circ}$ with a dashed red line. For simplicity, we show the gradient angles among the combined query, key, and value projection matrices for a total of 32 stitches when stitching ViT-Ti and ViT-S anchors. Generally, the gradient angles are larger in the target domain Stanford Cars [18] than in the source domain ImageNet-1k [56]." + ], + "image_footnote": [], + "bbox": [ + 272, + 329, + 460, + 445 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "be $W \\in \\mathbb{R}^{d \\times k}$ , we freeze $W$ and insert trainable low-rank decomposition matrices $W_{\\mathrm{down}} \\in \\mathbb{R}^{d \\times r}$ and $W_{\\mathrm{up}} \\in \\mathbb{R}^{r \\times k}$ , where $r$ is the pre-defined rank that $r \\ll \\min(d, k)$ . In this way, the updated version of $W$ can be formulated as", + "bbox": [ + 75, + 561, + 468, + 625 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {W} ^ {\\prime} \\leftarrow \\boldsymbol {W} + \\boldsymbol {W} _ {\\text {d o w n}} \\boldsymbol {W} _ {\\text {u p}}. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 181, + 633, + 468, + 651 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We follow [28] to respectively use Gaussian and zero initializations for $W_{\\mathrm{up}}$ and $W_{\\mathrm{down}}$ , so that $W_{\\mathrm{down}}W_{\\mathrm{up}}$ is zero at the beginning of fine-tuning.", + "bbox": [ + 75, + 657, + 468, + 703 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "LoRA on stitching layers. We further extend the low-rank weight update to stitching layers. In SN-Net, a stitching layer is designed as a full-rank transformation matrix. Without loss of generality, we show the low-rank update of stitching layer $S_{n}$ that stitches $f_{\\theta_1}$ and $f_{\\theta_2}$ at layers $l$ and $m$ , respectively. We first follow the initialization of stitching layers in [12, 51] to align the activations at layers $l$ and $m - 1$ . Let the activations be $H_{\\theta_1,l}(\\mathbf{X}) \\in \\mathbb{R}^{b \\times d_1}$ and $H_{\\theta_2,m - 1}(\\mathbf{X}) \\in \\mathbb{R}^{b \\times d_2}$ with sequence length $b$ and feature dimensions $d_1$ and $d_2$ . The stitching layer is parameterized by a transformation matrix $M \\in \\mathbb{R}^{d_1 \\times d_2}$ , which is initialized by solving a least squares problem, i.e., $M = H_{\\theta_1,l}(X)^{\\dagger}H_{\\theta_2,m - 1}(X)$ , where $H_{\\theta_1,l}(X)^{\\dagger}$ is the", + "bbox": [ + 75, + 704, + 470, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Moore-Penrose pseudoinverse of $H_{\\theta_1,l}(\\mathbf{X})$ . We find that such initialization already has impressive representational capacity on the target domain and they can be updated with low-rank decomposition without hurting the performance too much. Accordingly, we update $M$ similar to updating $W$ in Eq. (2) by approximating its update with two learnable low-rank matrices with the same initializations.", + "bbox": [ + 496, + 330, + 890, + 436 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Although the low-rank updates significantly enhance parameter efficiency, the low-rank essence [2] largely limits the network capacity. Particularly, in task adaptation, we often observe conflicting updates on the shared weights among the stitches. Figure 2 gives an example, where we stitch pre-trained anchors on ImageNet-1k [56] and Stanford Cars [18] and visualize the distribution of pair-wise gradient angles among different stitches when updating shared weights. We find that stitches agree less and conflict more (i.e., with more large angles) on the target domain Stanford Cars compared to the source domain ImageNet-1k. Stitch-specific bias. To address the above issue, we further introduce stitch-specific bias terms as depicted in Figure 1 (b). Particularly, for a linear layer parameterized by $W$ and a bias term $b$ , we add bias term $b^s$ when optimizing the $s$ -th stitch, for which the output activation becomes $WX + b + b^s$ . We employ a set of distinct bias terms at different layers for each stitch. In this way, the stitches are encouraged to learn distinct feature representations for different resource requirements. To restrict the number of trainable parameters, following [28], we only apply PST to self-attention layers while freezing feed-forward layers.", + "bbox": [ + 496, + 439, + 892, + 773 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.2. One-stage Deployment Pipeline", + "text_level": 1, + "bbox": [ + 498, + 785, + 777, + 801 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Simultaneously adapt and stitch anchors. As aforementioned, given a model family $\\mathcal{Z}$ pre-trained in the source domain, directly adapting SN-Net to a target domain typically involves three stages. First, adapting each anchor to the target domain $\\mathcal{D}$ by solving $\\theta_z^* = \\mathrm{argmin}_{\\theta_z} \\mathcal{L}(\\theta_z; \\mathcal{D})$ . In the second stage, at each training iteration, SN-Net sam", + "bbox": [ + 496, + 809, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "28558", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "plies a stitch and optimizes it with objective $\\mathcal{L}(\\phi_n^*, S_n; \\mathcal{D})$ , where $\\phi_n^*$ is the set of SN-Net parameters for stitch $F_{\\phi_n, S_n}$ which has been optimized once in the first stage. In the final stage, all optimized stitches need to be evaluated to identify the ones on the Pareto frontier for diverse resource constraints. Such a three-stage approach is expensive and sub-optimal (see more discussions in the supplementary material). Thus, we propose to adapt and stitch the anchors simultaneously within one stage, by directly optimizing $\\mathcal{L}(\\phi_n, S_n; \\mathcal{D})$ for each sampled stitch. Moreover, we introduce a novel task-specific stitch-sampling method to allow the promising stitches more likely to be sampled, based on a stitch importance score. After fine-tuning, the stitches with the highest scores are naturally selected for deployment without the need for the above final stage evaluation.", + "bbox": [ + 76, + 90, + 472, + 316 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Task-specific stitch sampling. SN-Net assumes that all stitches are equally important and performs random (uniform) sampling during training. However, as recognized by prior research [43, 70] that random sampling causes a gap between training and deployment as only the best stitches on the Pareto frontier are actually deployed. Moreover, since the importance of the pre-trained weights varies in different downstream tasks [19, 22, 75], we argue that the performance of stitches that are parameterized by these weights is also different across these tasks.", + "bbox": [ + 76, + 316, + 472, + 467 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To this end, we propose to assign the important stitches that are likely to be on the Pareto frontier with higher sampling probabilities during fine-tuning to ensure that they are optimized sufficiently. Specifically, we first estimate the importance of each stitch with a data-dependent saliency pruning metric SNIP [37], which measures the importance with the first-order gradient information with barely any extra computational cost. When the $n$ -th stitch is sampled, we can get its importance score $Q(F_{\\phi, S_n}, X)$ with the scoring function $Q(\\cdot, \\cdot)$ [37]. To obtain robust importance scores for sampling, we accumulate scores with moving average during training, i.e.,", + "bbox": [ + 75, + 468, + 472, + 650 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nq _ {n} ^ {t} \\leftarrow \\eta q _ {n} ^ {t - 1} + (1 - \\eta) Q \\left(F _ {\\phi_ {n}, S _ {n}}, \\boldsymbol {X}\\right), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 143, + 660, + 468, + 680 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $q_{n}^{t}$ and $q_{n}^{t - 1}$ are the importance score at the $t$ -th and $(t - 1)$ -th iteration, respectively, and $\\eta \\in [0,1)$ is the momentum coefficient. In this way, we can get stable importance scores. After a warm-up period that employs uniform sampling to accumulate the scores, we assign the important stitches with higher sampling probabilities for the rest of the fine-tuning epochs. To do so, we uniformly divide the resource constraint range of our stitches into several intervals and sample an interval $\\gamma$ with uniform sampling. Accordingly, we can obtain a subset of stitches $\\mathcal{N}_{\\gamma}$ whose resource constraints belong to $\\gamma$ and their corresponding importance scores $\\mathcal{Q}_{\\gamma}$ . We can then define a categorical distribution based on the normalized importance scores, i.e., $\\pi (\\mathcal{N}_{\\gamma}) = \\mathrm{softmax}(\\mathcal{Q}_{\\gamma})$ and finally get the sampled stitch", + "bbox": [ + 76, + 688, + 472, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "$F_{\\phi_n,S_n}\\sim \\pi (\\mathcal{N}_\\gamma)$ . We show that our task-specific stitch sampling improves the performance in Section 5.3.", + "bbox": [ + 498, + 90, + 890, + 119 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Additionally, with the sampled stitch, SN-Net employs knowledge distillation with a RegNetY-160 [54] pre-trained teacher to improve its performance. However, pre-trained teachers on downstream tasks are often unavailable. Therefore, in each training iteration, we always sample and train the largest anchor (teacher) and transfer its knowledge to the sampled stitch (student) similar to inplace distillation [79]. In practice, we employ the hard-label distillation [26, 63].", + "bbox": [ + 496, + 121, + 893, + 243 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Reuse stitch importance scores for deployment. After fine-tuning, with the set of accumulated importance scores $\\mathcal{Q}_{\\gamma}$ for all stitches, we directly use them to select the stitches with the highest scores to deploy, which eliminates the need for the costly evaluation of the final stage of the straightforward SN-Net adaptation solution. We empirically observe that the deployed stitches are mostly on the Pareto frontier in Section 5.3, suggesting our estimation of the important stitches is accurate. Furthermore, in the supplementary material, we show that the important stitches differ across different tasks, proving the need for task-specific deployment strategies, as highlighted in [19, 22, 75]. It's worth noting that although the SNIP metric has been widely employed as a one-shot metric to efficiently estimate network performance [1, 69], we extend it to estimate stitch importance for better training and selection.", + "bbox": [ + 496, + 244, + 893, + 484 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 500, + 502, + 633, + 518 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.1. Visual Recognition", + "text_level": 1, + "bbox": [ + 500, + 527, + 679, + 542 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We evaluate the effectiveness of our method on a total of 25 downstream visual recognition tasks, including fine-grained visual classification (FGVC) benchmark, common visual classification benchmark CIFAR-100 [34], and VTAB-1k [81] benchmark. FGVC benchmark contains NABirds [67], CUB-200-2011 [68], Stanford Cars [18], Stanford Dogs [32], and Oxford Flowers [50] tasks and the VTAB-1k benchmark includes 19 tasks in different domains, each of which has 800 training samples. For visual recognition experiments, we stitch ImageNet-21k pretrained ViT-Ti/S/B models [16] from [60].", + "bbox": [ + 496, + 551, + 890, + 717 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Implementation details. We employ the stitching space and settings (kernel size as 2 and stride as 1) in [51] for downstream visual recognition tasks. We uniformly divide the resource constraints supported by the stitches into around 15 intervals and deploy one stitch within each interval. We set the hyper-parameter $r$ for the rank of weight updating in the self-attention layers to be 32, 16, and 8 for ViT-Ti, ViT-S, and ViT-B, respectively. We also set $r = d_1 // 4$ universally for all stitching layers and $\\eta$ in Eq. (3) as 0.9 by grid search. We follow [31] to fine-tune 100 epochs on each task and select the other hyper-parameters and augmentation methods. We include more implementation details in", + "bbox": [ + 496, + 719, + 893, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "28559", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/e307fd95fed9c167bc222d554736d56ea0f914ca7e6f482f6bbdce418c4110b5.jpg", + "image_caption": [ + "Figure 3. Performance comparisons with SN-Net [51] for adapting ViT-Ti/S/B pre-trained on ImageNet-22k [15] to Stanford Cars [18], CUB-200-2011 [68], Stanford Dogs [32], and NABirds [67]. We denote individually fine-tuned anchors as yellow stars. We also show the number of trainable parameters." + ], + "image_footnote": [], + "bbox": [ + 80, + 89, + 284, + 203 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/1af46de912baf31fd1ba0b6c8e60c6943b6add1837011fa259f6feca9af4f84d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 282, + 89, + 486, + 203 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/2d1d320775be21db95f2bb6155c0f27ee7d00e8a46405e38fd095248054e2753.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 486, + 89, + 687, + 203 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/964957dde80b4e1b42c30c4ae4610a4a08c5518734d116b17c0b1eb272edf2da.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 687, + 89, + 890, + 203 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/68d6af674840a83cd087b0f24427380108f82ef6b28e3f28334f9c7e03424e0b.jpg", + "image_caption": [ + "Figure 4. Performance comparisons with SN-Net [51] for adapting ViT-Ti/S/B pre-trained on ImageNet-22k [15] to VTAB-1k [81] and CIFAR-100 [34]. We denote individually fine-tuned anchors as yellow stars and also show the number of trainable parameters." + ], + "image_footnote": [], + "bbox": [ + 78, + 252, + 272, + 359 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/671fb773bb232eec83b70da8eb196089159085b58ba9bc8f8fbdeda9aa8451c4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 272, + 252, + 467, + 359 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/2d4ce083b240f7e2038c00ec3d779a6e476150d743dc7b051c645c2922cf38c6.jpg", + "table_caption": [ + "Table 2. Relative response quality to Alpaca-LoRA 7B [62] on Vicuna Bench [10]. Our Stitched LLaMA successfully interpolates the answer quality between Alpaca-LoRA 7B and 13B." + ], + "table_footnote": [], + "table_body": "
Alpaca-LoRAStitched LLaMA
Parameter # (B)6.713.06.97.811.512.513.0
Quality (%)100123113114115120123
", + "bbox": [ + 81, + 468, + 465, + 513 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "the supplementary material.", + "bbox": [ + 76, + 522, + 263, + 536 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Main results. We compare the performance of our ESTA with SN-Net and the individual anchors fine-tuned with LoRA [28] on various datasets. The FLOPs-accuracy curves are visualized in Figures 3 and 4. We report the averaged results over the three task groups on VTAB-1k following [31]. Overall, the stitches obtained by our ESTA exhibit smooth FLOPs-accuracy trade-off curves consistently across all datasets. Notably, the curve for our ESTA outperforms SN-Net by significant margins with much fewer trainable parameters. We show that the inferior results of SN-Net on downstream tasks are led by the sub-optimal designs of fine-tuning all model parameters and a three-stage adaptation pipeline in Section 5.3.", + "bbox": [ + 75, + 537, + 468, + 733 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We also observe that the stitches even outperform the anchors with comparable or lower FLOPs in many cases. For instance, stitches outperform individually fine-tuned ViT-S/B anchors by clear margins on NABirds, CUB-200-2011, Stanford Cars, and Stanford Dogs datasets with comparable computational complexity. This phenomenon is consistent with the observations made in [52, 79]. We conjecture that weight sharing among the stitches serves as a strong regularization to improve their generalization. We show more results on few-shot learning and stitching Convnets [46] in the supplementary material.", + "bbox": [ + 75, + 734, + 470, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.2. Instruction-following Task", + "text_level": 1, + "bbox": [ + 498, + 252, + 740, + 268 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We also evaluate the instruction-following capability when stitching the LLaMA family [64] with our ESTA framework. We adapt and fine-tune LLaMA-7B/13B on the 52K instruction-following dataset from [62] to obtain chatbot stitches of varying sizes. The 52K instruction-following data is generated from 175 human-written instruction-output pairs with [71]. Each sample in the dataset is formulated by a task description, the context of the task, and the answer that is generated by GPT-3.5 (text-davinci-003) [7].", + "bbox": [ + 496, + 276, + 890, + 412 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation details. We employ the stitching space and settings (kernel size as 1 and stride as 1) in [51] for the instruction-following task and introduce 40 stitches. We uniformly divide the resource constraints supported by the stitches into 6 intervals and deploy one stitch within each interval. We set $\\eta$ in Eq. (3) as 0.9. The rank $r$ is set to 16 and $d_1 // 4$ for LoRA modules in self-attention and stitching layers, respectively. We employ Adam optimizer and fine-tune with batch size 128 for 10 epochs with gradient accumulation. During the generation stage, we employ top- $p$ sampling to sample from the top subword candidates with the probability $p = 0.75$ , conduct the beam search, and set the beam size as 4.", + "bbox": [ + 496, + 412, + 892, + 608 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Results. We term the stitches obtained by our ESTA framework as Stitched LLaMA and compare them with Alpaca-LoRA 7B and 13B that respectively freeze the LLaMA 7B and 13B backbones and update the self-attention weights with LoRA modules [28]. For quantitative results reported in Table 2, we observe that similar to the visual recognition tasks, our Stitched LLaMA with different numbers of parameters generates answers with plausible quality in between Alpaca-LoRA 7B and 13B. Figure 5 gives qualitative comparisons. Our Stitched LLaMAs produce reasonable responses, suggesting the potential of our ESTA for adapting large language models (LLMs) for more downstream language tasks. More examples can be found in the supplementary material.", + "bbox": [ + 496, + 609, + 892, + 821 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.3. Ablation Study", + "text_level": 1, + "bbox": [ + 500, + 832, + 653, + 849 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Effect of parameter-efficient stitch fine-tuning. We investigate the effectiveness of our PST on five FGVC tasks and visualize the averaged results in Figure 6 (a) and (b).", + "bbox": [ + 498, + 854, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "28560", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Instruction: Tell me about alpacas.", + "text_level": 1, + "bbox": [ + 86, + 93, + 259, + 104 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f04c44941f3c8d7da59b4f13ce2725b9edc715daee04f06880f6c899b845ccc5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 83, + 109, + 120, + 136 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Alpaca-LoRA 7B (0.035 s/token): Alpacas are members of the camelid family, which also includes llamas and camels.", + "bbox": [ + 122, + 114, + 406, + 136 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "They are native to the Andes Mountains in South America, where they have been domesticated for thousands of years. Alpacas are known for their soft, luxurious fiber, which is used to make clothing, blankets, and other textiles. They are also raised for their meat, which is considered a delicacy in some parts of the world. Alpacas are gentle, intelligent animals, and they are prized for their docile nature and ability to thrive in a wide range of climates.", + "bbox": [ + 84, + 137, + 410, + 213 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/6ba465d987693d004666b0c529a660dad07275b93b25b12df950b2502ffcce14.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 419, + 109, + 457, + 136 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Stitched LLaMA 8B (0.037 s/token): Alpacas are herbivorous mammals that belong to the family Camelidae, which also includes camels and camelids. They are native to South American continent and are commonly found in Peru, Argentina, Bolivia, and Chile. They are smaller than their cousin, the camel, and are typically about 80-90 inches tall fully grown. They are known for their soft, shimmering, and light-weight fibers, which tend to make clothing, carpets, and other textile products. Alpacas are herbivores, and catets consists mainly of grasses, leaves, and straw. They are social animals and live in a few up to 100 alpacas, led by a dominant male. Alpacas are known for their calm and mature, and they are often used as livestock for their soft, shimmering fibers.", + "bbox": [ + 460, + 114, + 877, + 213 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a8086e6c950405e98b75fa0f6c7d5934260f4851d4b3c8e5cf2c1905e96d242b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 83, + 219, + 120, + 244 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Stitched LLaMA 11B (0.041 s/token): Alpacas are native to South America and are members of the camelid family. They are known for their soft, fluffy coats and their gentle nature. Alpacas are herbivores, meaning they feed on grass and other vegetation. They have a gestation period of approximately 11 days and usually give birth to a single baby, although twins and triplets are not uncommon. Alpacas are social animals and live in herds, led by a dominant they communicate with each other using a variety of vocalizations, including whistles, grunts, and barks. Alpacas are intelligent animals and are known for curiosity and playful nature. Alpacas have been domesticated for thousands of years and are used for a variety of purposes, including as pets, for their fiber, for their meat. Alpaca fiber is prized for its softness and warmth, making it a popular material for clothing and blankets. Alpaca meat is also considered a key in some parts of the world. Overall, alpacas are gentle, intelligent animals that have been domesticated for thousands of years. They are valued for their meat, and companionship.", + "bbox": [ + 122, + 224, + 869, + 313 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Instruction: Can you help me write a resignation letter to my current employer, while leaving on good terms and expressing gratitude for the opportunities provided?", + "text_level": 1, + "bbox": [ + 84, + 323, + 857, + 334 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/411f36edf167bbf47494194987b87d052b2e34c5c29b7e7aa860be3dbd9ee5f9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 84, + 339, + 117, + 364 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Alpaca-LoRA 7B (0.035 s-token): Dear [Name]", + "bbox": [ + 120, + 342, + 344, + 354 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "of Employer], I am writing to inform you that I am resigning from my position as [Job Title] with [Company Name]. It has been a pleasure working with you and I am grateful for the opportunities you have provided me. I wish you and the company all the best in the future. Sincerely, [Your Name]", + "bbox": [ + 84, + 354, + 351, + 420 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/5931474e9cf30c70a33e7ce73682c1bf912f17e153c4880262b3e013fa17f1d2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 374, + 339, + 411, + 364 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Stitched LLaMA 8B (0.037 s/token): Dear [Employer], I am writing to inform you that I am resigning from my position at [Company Name] with a heavy heart. I have enjoyed my time with you and the team, and I am grateful for the opportunities that have been provided to me. I like to express my sincere appreciation for the training, support, and opportunities that have provided to me during my time at [Company Name]. I have learned a great deal and have grown really and professionally. Please accept my best wishes for the continued success of [Company and the team. I wish you all the best in the future. Sincerely, [Your Name]", + "bbox": [ + 413, + 342, + 877, + 420 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d4f12cfe8a5f3a5d8c794e044a81734dca23d4486d52cdb0b046f85ecc524e72.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 83, + 425, + 117, + 452 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Sti 118 (0.041 s/ token): Dear [Employer], I am writing this letter to inform you of my decision to resign from my position at [Company]. It has been a great pleasure to work for [Company] and I have learned a lot during my time here. I would like to express my gratitude for the opportunities that I seen provided to me and for the support and guidance that I have received from my colleagues and supervisors. I have grown both professionally and finally during my time at [Company] and I am grateful for the experiences I have had. I am leaving on good terms and would be happy to provide a positive chance for [Company] in the future. If there is anything I can do to assist with the transition, please do not hesitate to reach out. Thank you for the opportunity to offer [Company] and I wish you all the best in your future endeavors. Sincerely, [Your Name]", + "bbox": [ + 120, + 431, + 880, + 500 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/1783f314e4bf19760d94fd88dbe429f7b73f0bf028fd5c5f79aeaf4afe4b574b.jpg", + "image_caption": [ + "Figure 5. Instruction-following comparison between Stitched LLaMA obtained by our ESTA and the Alpaca-LoRA 7B fine-tuned with LoRA [28]." + ], + "image_footnote": [], + "bbox": [ + 78, + 546, + 285, + 669 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/8740576bf0900c0f058d9529d2da21aa3f2b5b820d1e3c70ad40df89d84429ba.jpg", + "image_caption": [ + "Figure 6. Ablation studies with results averaged over five FGVC tasks. (a) Effect of our stitch-agnostic LoRA modules. \"Full FT\" represents full fine-tuning of all the model parameters. \"PST w/ full SL\" employs fully fine-tuned stitching layers. We also show the number of trainable parameters. (b) Effect of our stitch-specific bias terms. \"SSB\" and \"scaled r\" represent stitch-specific bias terms and scaling the rank hyperparameter to reach 4.6M trainable parameters, respectively. (c) Effect of our task-specific stitch sampling. \"Uniform, all\" and \"Uniform, best\" represent all the stitches fine-tuned with uniform sampling and their best ones in each resource constraint interval on the Pareto frontier, respectively. (d) Effect of our strategy to simultaneously adapt and stitch anchors. Our strategy \"Adapt-and-stitch\" takes 100 epochs for fine-tuning. In contrast, \"Adapt-then-stitch\", as a straightforward approach to apply SN-Net, first individually adapts each anchor for a total of 300 epochs, and then fine-tunes SN-Net for another 100 epochs." + ], + "image_footnote": [], + "bbox": [ + 282, + 547, + 485, + 670 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/498598ad00fe8f261880fe20e83c7953d2cb06ed289b570d59ad9748e964b2bd.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 485, + 547, + 687, + 670 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ce9fa356fb864e0fedd2391c575e610e23e69370aad23b33fd2095ad337a34c4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 687, + 547, + 888, + 670 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To recall, our PST keeps the majority of the model parameters frozen while introducing trainable stitch-agonistic low-rank decomposition matrices and stitch-specific bias terms. In Figure 6 (a), we observe that when employing LoRA to the multi-head self-attention layers, there are solid performance gains on all FLOPs ranges with significantly fewer trainable parameters, echoing the observations in [31, 41].", + "bbox": [ + 75, + 792, + 470, + 898 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "It is suggested that low-rank weight updates is a powerful alternative to full fine-tuning for stitching and adapting anchors and we speculate that low-rank weight update avoids overfitting under limited downstream data. We further employ low-rank weight updates to the stitching layers and save 2.4M more trainable parameters with a slight overall performance drop. We conjecture that the stitching layers", + "bbox": [ + 496, + 792, + 893, + 898 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "28561", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "for large pre-trained models are also in low intrinsic dimensions [2] and least-squares initialization [51] already provide a good initialization for them.", + "bbox": [ + 75, + 90, + 468, + 136 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Figure 6 (b) shows that employing stitch-specific bias terms achieves better performance with affordable 1.4M extra trainable parameters. We also include the baseline that simply scales the low-rank dimension $r$ to reach the same number of trainable parameters as our PST. This baseline is inferior to our PST. We speculate that the stitch-specific bias terms enable flexible adjustments of feature representations for different stitches, leading to enhanced performance.", + "bbox": [ + 75, + 138, + 467, + 258 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effect of task-specific stitch sampling. We investigate the effectiveness of our sampling strategy (introduced in Section 4.2) on the five FGVC tasks. The averaged results are visualized in Figure 6 (c). We compare our task-specific stitch sampling with the uniform one that is adopted in [51]. We empirically find that for uniform sampling, there is a clear performance gap between the best stitches that will be deployed and the others that will be dropped after training, suggesting the need for concentrating on the important stitches during training. To this end, our task-specific sampling strategy performs better than uniform sampling, indicating that our sampling strategy selects accurate task-specific important stitches. Most importantly, our strategy circumvents the need for the costly evaluation stage in the direct solution of adapting SN-Net for downstream tasks.", + "bbox": [ + 75, + 261, + 467, + 487 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effect of simultaneously adapting and stitching anchors. We study the effect of simultaneously adapting and stitching anchors strategy (introduced in Section 4.2) on five FGVC tasks. The averaged results are visualized in Figure 6 (d). We compare our approach with the direct SN-Net adaptation that first adapts each anchor individually for a total of 300 epochs before stitching them under the exact setting as ESTA. It can be seen that the direct SN-Net adaptation yields poor results in task adaptation. Fine-tuned individual anchors, especially the larger ones, are likely to overfit the limited downstream data [44, 53]. Accordingly, we speculate that the adapted anchor weights serve as a sub-optimal initialization for the stitches in the context of model stitching. On the other hand, our approach achieves better performance and greatly saves the fine-tuning cost.", + "bbox": [ + 75, + 489, + 467, + 715 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effect of the PEFT technique choices. We compare the effect of different PEFT technique choices and visualize the averaged results on the five FGVC tasks in Figure 7 left. We experiment to employ other PEFT techniques Adapter [27] and Adaptformer [9]. We follow their default setting to respectively insert bottleneck structures sequentially and in parallel to the feed-forward layers and fix the number of trainable parameters the same as ours for a fair comparison. We observe that the overall performance for different PEFT techniques is similar, suggesting that our ESTA framework is compatible with more PEFT technique choices. Since low-rank weight updates of LoRA [28] can be merged into", + "bbox": [ + 75, + 719, + 467, + 900 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/eb0215283bf4cc3a8215ec628dc353308860aebd9127fbbc2c15e565555c2864.jpg", + "image_caption": [ + "Figure 7. Left: effect of different PEFT technique choices. Right: effect of the rank hyper-parameter $r$ in LoRA modules of self-attention layers, for which the three values represent the ranks in ViT-Ti, ViT-S, and ViT-B, respectively. The results are averaged on five FGVC tasks." + ], + "image_footnote": [], + "bbox": [ + 501, + 88, + 694, + 198 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/f73035690bfa05c8498872e9d57e73f9fe6c8aa13500c1b2ab00781a7f797b43.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 694, + 88, + 890, + 198 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "the backbone after training without extra inference computational cost, we employ LoRA by default.", + "bbox": [ + 496, + 299, + 890, + 329 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effect of the low-rank hyper-parameter $r$ . We investigate the effect of different $r$ which controls the rank when updating weights in self-attention layers on the five FGVC tasks. The averaged results on the five FGVC tasks are visualized in Figure 7 right. We find that setting $r$ to be different values barely has any effect on the overall performance, suggesting our ESTA framework is robust to the rank hyper-parameter $r$ . Since setting $r$ to be 32, 16, and 8 for low-rank updating the self-attention weights in ViT-Ti, ViT-S, and ViT-B has slightly higher averaged performance among the stitches, we make it our default setting.", + "bbox": [ + 496, + 334, + 890, + 500 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion and Future Work", + "text_level": 1, + "bbox": [ + 500, + 532, + 767, + 547 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we have introduced a novel task adaptation framework to cheaply obtain a palette of fine-tuned networks via model stitching, supporting diverse efficiency-performance tradeoffs at runtime. Specifically, built on SN-Net, we have tailored a parameter-efficient stitch finetuning method, which learns lightweight stitch-agnostic LoRA modules and stitch-specific bias terms while keeping the majority of the parameters frozen. Our design significantly reduces fine-tuning memory and storage costs for downstream task adaptation. Moreover, we have devised a task-specific stitch sampling strategy to assign higher sampling probability to the important stitches during training, which simultaneously improves the Pareto frontiers and avoids a costly evaluation stage. Extensive experiments on 25 downstream visual recognition tasks and the instruction-following task have demonstrated the effectiveness of our proposed framework.", + "bbox": [ + 496, + 561, + 890, + 819 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Limitations. Due to the constraints of computational resources, our experiments are limited to visual recognition and instruction-tuning tasks. In the future, we will explore adapting pre-trained model families to dense prediction and multi-modal tasks.", + "bbox": [ + 496, + 825, + 890, + 898 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "28562", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Mohamed S. Abdelfattah, Abhinav Mehrotra, Lukasz Dudziak, and Nicholas D. Lane. Zero-cost proxies for lightweight nas. In ICLR, 2021. 5", + "[2] Armen Aghajanyan, Luke Zettlemoyer, and Sonal Gupta. Intrinsic dimensionality explains the effectiveness of language model fine-tuning. In ACL, 2021. 4, 8", + "[3] Haitham Bou Ammar, Eric Eaton, Matthew E Taylor, Decebal Constantin Mocanu, Kurt Driessens, Gerhard Weiss, and Karl Tuyls. An automated measure of mdp similarity for transfer in reinforcement learning. In AAAIw, 2014. 3", + "[4] Yamini Bansal, Preetum Nakkiran, and Boaz Barak. Revisiting model stitching to compare neural representations. NeurIPS, 34:225-236, 2021. 1, 2", + "[5] Hangbo Bao, Li Dong, Songhao Piao, and Furu Wei. Beit: Bert pre-training of image transformers. In ICLR, 2022. 1", + "[6] Daniel Bolya, Rohit Mittapalli, and Judy Hoffman. Scalable diverse model selection for accessible transfer learning. NeurIPS, 34:19301-19312, 2021. 3", + "[7] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. *NeurIPS*, 33:1877–1901, 2020. 6", + "[8] Han Cai, Chuang Gan, Tianzhe Wang, Zhekai Zhang, and Song Han. Once-for-all: Train one network and specialize it for efficient deployment. In ICLR, 2020. 1", + "[9] Shoufa Chen, Chongjian Ge, Zhan Tong, Jiangliu Wang, Yibing Song, Jue Wang, and Ping Luo. Adaptformer: Adapting vision transformers for scalable visual recognition. NeurIPS, 2022. 8", + "[10] Wei-Lin Chiang, Zhuohan Li, Zi Lin, Ying Sheng, Zhang-hao Wu, Hao Zhang, Lianmin Zheng, Siyuan Zhuang, Yong-hao Zhuang, Joseph E. Gonzalez, Ion Stoica, and Eric P. Xing. Vicuna: An open-source chatbot impressing gpt-4 with $90\\%$ * chatgpt quality, 2023. 6", + "[11] Adam Coates and Andrew Y Ng. The importance of encoding versus training with sparse coding and vector quantization. In ICML, pages 921-928, 2011. 1", + "[12] Adrián Csiszárik, Péter Körösí-Szabó, Akos Matszangosz, Gergely Papp, and Daniel Varga. Similarity and matching of neural network representations. NeurIPS, 34:5656-5668, 2021. 1, 2, 4", + "[13] Dong Dai and Tong Zhang. Greedy model averaging. NeurIPS, 24, 2011. 3", + "[14] Xiaoliang Dai, Alvin Wan, Peizhao Zhang, Bichen Wu, Zijian He, Zhen Wei, Kan Chen, Yuandong Tian, Matthew Yu, Peter Vajda, et al. Fbnetv3: Joint architecture-recipe search using predictor pretraining. In CVPR, pages 16276–16285, 2021. 2", + "[15] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR, pages 248-255, 2009. 6", + "[16] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is" + ], + "bbox": [ + 78, + 114, + 470, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2021. 2, 5", + "[17] Eric Eaton, Marie Desjardins, and Terran Lane. Modeling transfer relationships between learning tasks for improved inductive transfer. In ECML PKDD, pages 317-332. Springer, 2008. 3", + "[18] Timnit Gebru, Jonathan Krause, Yilun Wang, Duyun Chen, Jia Deng, and Li Fei-Fei. Fine-grained car detection for visual census estimation. In AAAI, 2017. 4, 5, 6", + "[19] Yunhui Guo, Honghui Shi, Abhishek Kumar, Kristen Grauman, Tajana Rosing, and Rogerio Feris. Spottune: transfer learning through adaptive fine-tuning. In CVPR, pages 4805-4814, 2019. 5", + "[20] Zichao Guo, Xiangyu Zhang, Haoyuan Mu, Wen Heng, Zechun Liu, Yichen Wei, and Jian Sun. Single path one-shot neural architecture search with uniform sampling. In ECCV, pages 544-560. Springer, 2020. 1", + "[21] Song Han, Jeff Pool, John Tran, and William Dally. Learning both weights and connections for efficient neural network. NeurIPS, 28, 2015. 1", + "[22] Haoyu He, Jianfei Cai, Jing Zhang, Dacheng Tao, and Bohan Zhuang. Sensitivity-aware visual parameter-efficient tuning. In ICCV, 2023. 5", + "[23] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollar, and Ross Girshick. Masked autoencoders are scalable vision learners. In CVPR, pages 16000-16009, 2022. 1", + "[24] Rudolf Herdt, Maximilian Schmidt, Daniel Otero Baguer, Jean Le'Clerc Arrastia, and Peter Maass. Model stitching and visualization how gan generators can invert networks in real-time. arXiv preprint arXiv:2302.02181, 2023. 2", + "[25] Adriano Hernandez, Rumen Dangovski, Peter Y. Lu, and Marin Soljacic. Model stitching: Looking for functional similarity between representations. NeurIPSW, 2022. 2", + "[26] Geoffrey Hinton, Oriol Vinyals, and Jeff Dean. Distilling the knowledge in a neural network. NeurIPSW, 2014. 5", + "[27] Neil Houlsby, Andrei Giurgiu, Stanislaw Jastrzebski, Bruna Morrone, Quentin De Laroussilhe, Andrea Gesmundo, Mona Attariyan, and Sylvain Gelly. Parameter-efficient transfer learning for nlp. In ICML, pages 2790-2799, 2019. 1, 3, 8", + "[28] Edward J Hu, yelong shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. LoRA: Low-rank adaptation of large language models. In ICLR, 2022. 1, 2, 3, 4, 6, 7, 8", + "[29] Itay Hubara, Matthieu Courbariaux, Daniel Soudry, Ran El-Yaniv, and Yoshua Bengio. Quantized neural networks: Training neural networks with low precision weights and activations. JMLR, 18(1):6869-6898, 2017. 1", + "[30] Pavel Izmailov, Dmitrii Podoprikhin, Timur Garipov, Dmitry Vetrov, and Andrew Gordon Wilson. Averaging weights leads to wider optima and better generalization. In UAI, 2018. 3", + "[31] Menglin Jia, Luming Tang, Bor-Chun Chen, Claire Cardie, Serge Belongie, Bharath Hariharan, and Ser-Nam Lim. Visual prompt tuning. In ECCV, 2022. 1, 3, 5, 6, 7", + "[32] Aditya Khosla, Nityananda Jayadevaprakash, Bangpeng Yao, and Li Fei-Fei. Novel dataset for fine-grained image categorization. In CVPRW, 2011. 5, 6" + ], + "bbox": [ + 503, + 92, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "28563", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[33] Simon Kornblith, Jonathon Shlens, and Quoc V Le. Do better imagenet models transfer better? In CVPR, pages 2661-2671, 2019. 3", + "[34] Alex Krizhevsky. Learning multiple layers of features from tiny images. Technical report, 2009. 5, 6", + "[35] Achintya Kundu, Laura Wynter, Rhui Dih Lee, and Luis Angel Bathen. Tofa: Transfer-once-for-all. arXiv preprint arXiv:2303.15485, 2023. 2", + "[36] Yann LeCun, John Denker, and Sara Solla. Optimal brain damage. NeurIPS, 2, 1989. 1", + "[37] Namhoon Lee, Thalaiyasingam Ajanthan, and Philip Torr. Snip: Single-shot network pruning based on connection sensitivity. In ICLR, 2019. 2, 5", + "[38] Karel Lenc and Andrea Vedaldi. Understanding image representations by measuring their equivariance and equivalence. In CVPR, pages 991-999, 2015. 1, 2", + "[39] Brian Lester, Rami Al-Rfou, and Noah Constant. The power of scale for parameter-efficient prompt tuning. In EMNLP, 2021. 3", + "[40] Xiang Lisa Li and Percy Liang. Prefix-tuning: Optimizing continuous prompts for generation. In ACL, 2021. 3", + "[41] Dongze Lian, Daquan Zhou, Jiashi Feng, and Xinchao Wang. Scaling & shifting your features: A new baseline for efficient model tuning. NeurIPS, 2022. 3, 7", + "[42] Iou-Jen Liu, Jian Peng, and Alexander G Schwing. Knowledge flow: Improve upon your teachers. In ICLR, 2019. 3", + "[43] Jing Liu, Jianfei Cai, and Bohan Zhuang. Focusformer: Focusing on what we need via architecture sampler. arXiv preprint arXiv:2208.10861, 2022.5", + "[44] Pengfei Liu, Weizhe Yuan, Jinlan Fu, Zhengbao Jiang, Hiroaki Hayashi, and Graham Neubig. Pre-train, prompt, and predict: A systematic survey of prompting methods in natural language processing. ACM Comput. Surv., 2023. 8", + "[45] Yen-Cheng Liu, Chih-Yao Ma, Junjiao Tian, Zijian He, and Zsolt Kira. Polyhistor: Parameter-efficient multi-task adaptation for dense vision tasks. NeurIPS, 2022. 3", + "[46] Zhuang Liu, Hanzi Mao, Chao-Yuan Wu, Christoph Feichtenhofer, Trevor Darrell, and Saining Xie. A convnet for the 2020s. In CVPR, pages 11976-11986, 2022. 6", + "[47] Zhichao Lu, Gautam Sreekumar, Erik Goodman, Wolfgang Banzhaf, Kalyanmoy Deb, and Vishnu Naresh Boddeti. Neural architecture transfer. TPAMI, 43(9):2971-2989, 2021. 2", + "[48] Michael S Matena and Colin A Raffel. Merging models with fisher-weighted averaging. NeurIPS, 35:17703-17716, 2022. 3", + "[49] Cuong Nguyen, Tal Hassner, Matthias Seeger, and Cedric Archambeau. Leep: A new measure to evaluate transferability of learned representations. In ICML, pages 7294-7305. PMLR, 2020. 3", + "[50] Maria-Elena Nilsback and Andrew Zisserman. Automated flower classification over a large number of classes. In ICVGIP, pages 722-729. IEEE, 2008. 5", + "[51] Zizheng Pan, Jianfei Cai, and Bohan Zhuang. Stretchable neural networks. In CVPR, 2023. 1, 2, 3, 4, 5, 6, 8", + "[52] Hieu Pham, Melody Guan, Barret Zoph, Quoc Le, and Jeff Dean. Efficient neural architecture search via parameters sharing. In ICML, pages 4095-4104. PMLR, 2018. 6" + ], + "bbox": [ + 78, + 90, + 470, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[53] Jo Pleted and Tom Gedeon. Deep transfer learning for image classification: a survey. arXiv preprint arXiv:2205.09904, 2022. 8", + "[54] Ilija Radosavovic, Raj Prateek Kosaraju, Ross Girshick, Kaiming He, and Piotr Dólar. Designing network design spaces. In CVPR, pages 10428-10436, 2020. 5", + "[55] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, pages 10684-10695, 2022. 3", + "[56] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, et al. Imagenet large scale visual recognition challenge. IJCV, 115:211-252, 2015. 1, 4", + "[57] Andrei A Rusu, Neil C Rabinowitz, Guillaume Desjardins, Hubert Soyer, James Kirkpatrick, Koray Kavukcuoglu, Razvan Pascanu, and Raia Hadsell. Progressive neural networks. arXiv preprint arXiv:1606.04671, 2016. 3", + "[58] Yang Shu, Zhi Kou, Zhangjie Cao, Jianmin Wang, and Mingsheng Long. Zoo-tuning: Adaptive transfer from a zoo of models. In ICML, pages 9626-9637. PMLR, 2021. 3", + "[59] Yang Shu, Zhangjie Cao, Ziyang Zhang, Jianmin Wang, and Mingsheng Long. Hub-pathway: Transfer learning from a hub of pre-trained models. NeurIPS, 2022. 3", + "[60] Andreas Steiner, Alexander Kolesnikov, Xiaohua Zhai, Ross Wightman, Jakob Uszkoreit, and Lucas Beyer. How to train your vit? data, augmentation, and regularization in vision transformers. TMLR, 2022. 5", + "[61] Yi-Lin Sung, Jaemin Cho, and Mohit Bansal. Lst: Ladder side-tuning for parameter and memory efficient transfer learning. NeurIPS, 2022. 3", + "[62] Rohan Taori, Ishaan Gulrajani, Tianyi Zhang, Yann Dubois, Xuechen Li, Carlos Guestrin, Percy Liang, and Tatsunori B. Hashimoto. Stanford alpaca: An instruction-following llama model. https://github.com/tatsu-lab/stanford_alpaca, 2023. 6", + "[63] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Hervé Jégou. Training data-efficient image transformers & distillation through attention. In ICML, pages 10347-10357. PMLR, 2021. 5", + "[64] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 1, 2, 6", + "[65] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288, 2023. 2", + "[66] Anh T Tran, Cuong V Nguyen, and Tal Hassner. Transferability and hardness of supervised classification tasks. In ICCV, pages 1395-1405, 2019. 3", + "[67] Grant Van Horn, Steve Branson, Ryan Farrell, Scott Haber, Jessie Barry, Panos Ipeirotis, Pietro Perona, and Serge Belongie. Building a bird recognition app and large scale" + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "28564", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "dataset with citizen scientists: The fine print in fine-grained dataset collection. In CVPR, pages 595-604, 2015. 5, 6", + "[68] Catherine Wah, Steve Branson, Peter Welinder, Pietro Perona, and Serge Belongie. The caltech-ucsd birds-200-2011 dataset. Tech. Rep. CNS-TR-2011-001, California Institute of Technology, 2011. 5, 6", + "[69] Chaoqi Wang, Guodong Zhang, and Roger Grosse. Picking winning tickets before training by preserving gradient flow. In ICLR, 2020. 5", + "[70] Dilin Wang, Meng Li, Chengyue Gong, and Vikas Chandra. Attentivenas: Improving neural architecture search via attentive sampling. In CVPR, pages 6418-6427, 2021. 5", + "[71] Yizhong Wang, Yeganeh Kordi, Swaroop Mishra, Alisa Liu, Noah A Smith, Daniel Khashabi, and Hannaneh Hajishirzi. Self-instruct: Aligning language model with self generated instructions. arXiv preprint arXiv:2212.10560, 2022. 6", + "[72] Thomas Wolf, Lysandre Debut, Victor Sanh, Julien Chau-mond, Clement Delangue, Anthony Moi, Pierrick Cistac, Tim Rault, Rémi Louf, Morgan Funtowicz, et al. Transformers: State-of-the-art natural language processing. In EMNLP, pages 38-45, 2020. 1", + "[73] Mitchell Wortsman, Gabriel Ilharco, Samir Ya Gadre, Rebecca Roelofs, Raphael Gontijo-Lopes, Ari S Morcos, Hongseok Namkoong, Ali Farhadi, Yair Carmon, Simon Kornblith, et al. Model soups: averaging weights of multiple fine-tuned models improves accuracy without increasing inference time. In ICML, pages 23965-23998. PMLR, 2022. 3", + "[74] Hang Xu, Ning Kang, Gengwei Zhang, Chuanlong Xie, Xiaodan Liang, and Zhenguo Li. Nasoa: Towards faster task-oriented online fine-tuning with a zoo of models. In ICCV, pages 5097-5106, 2021. 3", + "[75] Runxin Xu, Fuli Luo, Zhiyuan Zhang, Chuanqi Tan, Baobao Chang, Songfang Huang, and Fei Huang. Raise a child in large language model: Towards effective and generalizable fine-tuning. In EMNLP, 2021. 5", + "[76] Xingyi Yang, Daquan Zhou, Songhua Liu, Jingwen Ye, and Xinchao Wang. Deep model reassembly. NeurIPS, 35: 25739-25753, 2022. 2", + "[77] Jason Yosinski, Jeff Clune, Yoshua Bengio, and Hod Lipson. How transferable are features in deep neural networks? NeurIPS, 27, 2014. 3", + "[78] Kaichao You, Yong Liu, Jianmin Wang, and Mingsheng Long. Logme: Practical assessment of pre-trained models for transfer learning. In ICML, pages 12133-12143. PMLR, 2021. 3", + "[79] Jiahui Yu and Thomas S Huang. Universally slimmable networks and improved training techniques. In ICCV, pages 1803-1811, 2019. 5, 6", + "[80] Elad Ben Zaken, Yoav Goldberg, and Shauli Ravfogel. Bitfit: Simple parameter-efficient fine-tuning for transformer-based masked language-models. In ACL, pages 1-9, 2022. 3", + "[81] Xiaohua Zhai, Joan Puigcerver, Alexander Kolesnikov, Pierre Ruyssen, Carlos Riquelme, Mario Lucic, Josip Djolonga, Andre Susano Pinto, Maxim Neumann, Alexey Dosovitskiy, et al. A large-scale study of representation learning with the visual task adaptation benchmark. arXiv preprint arXiv:1910.04867, 2019. 5, 6" + ], + "bbox": [ + 78, + 90, + 468, + 898 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[82] Lvmin Zhang and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In ICCV, 2023. 3", + "[83] Renrui Zhang, Jiaming Han, Aojun Zhou, Xiangfei Hu, Shilin Yan, Pan Lu, Hongsheng Li, Peng Gao, and Yu Qiao. Llama-adapter: Efficient fine-tuning of language models with zero-init attention. arXiv preprint arXiv:2303.16199, 2023. 3" + ], + "bbox": [ + 501, + 92, + 890, + 188 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "28565", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2024/Efficient Stitchable Task Adaptation/03126c3f-71ba-4532-921c-1c95d792e515_model.json b/2024/Efficient Stitchable Task Adaptation/03126c3f-71ba-4532-921c-1c95d792e515_model.json new file mode 100644 index 0000000000000000000000000000000000000000..e43f5ea37c3a06f2477410cc6515cec22d5c1dd9 --- /dev/null +++ b/2024/Efficient Stitchable Task Adaptation/03126c3f-71ba-4532-921c-1c95d792e515_model.json @@ -0,0 +1,2543 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.302, + 0.131, + 0.67, + 0.154 + ], + "angle": 0, + "content": "Efficient Stitchable Task Adaptation" + }, + { + "type": "text", + "bbox": [ + 0.213, + 0.181, + 0.764, + 0.201 + ], + "angle": 0, + "content": "Haoyu He Zizheng Pan Jing Liu Jianfei Cai Bohan Zhuang" + }, + { + "type": "text", + "bbox": [ + 0.328, + 0.206, + 0.638, + 0.224 + ], + "angle": 0, + "content": "ZIP Lab, Monash University, Australia" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.258, + 0.314, + 0.274 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.29, + 0.473, + 0.775 + ], + "angle": 0, + "content": "The paradigm of pre-training and fine-tuning has laid the foundation for deploying deep learning models. However, most fine-tuning methods are designed to meet a specific resource budget. Recently, considering diverse deployment scenarios with various resource budgets, SN-Net [51] is introduced to quickly obtain numerous new networks (stitches) from the pre-trained models (anchors) in a model family via model stitching. Although promising, SN-Net confronts new challenges when adapting it to new target domains, including huge memory and storage requirements and a long and sub-optimal multistage adaptation process. In this work, we present a novel framework, Efficient Stitchable Task Adaptation (ESTA), to efficiently produce a palette of fine-tuned models that adhere to diverse resource constraints. Specifically, we first tailor parameter-efficient fine-tuning to share low-rank updates among the stitches while maintaining independent bias terms. In this way, we largely reduce fine-tuning memory burdens and mitigate the interference among stitches that arises in task adaptation. Furthermore, we streamline a simple yet effective one-stage deployment pipeline, which estimates the important stitches to deploy with training-time gradient statistics. By assigning higher sampling probabilities to important stitches, we also get a boosted Pareto frontier. Extensive experiments on 25 downstream visual recognition tasks demonstrate that our ESTA is capable of generating stitches with smooth accuracy-efficiency trade-offs and surpasses the direct SN-Net adaptation by remarkable margins with significantly lower training time and fewer trainable parameters. Furthermore, we demonstrate the flexibility and scalability of our ESTA framework by stitching LLMs from LLaMA family, obtaining chatbot stitches of assorted sizes1." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.801, + 0.21, + 0.816 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.826, + 0.47, + 0.856 + ], + "angle": 0, + "content": "The paradigm of pre-training and fine-tuning has underpinned modern applications in both vision and lan" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.26, + 0.895, + 0.472 + ], + "angle": 0, + "content": "guage. With off-the-shelf models pre-trained on large-scale datasets, the de-facto choice is vanilla full fine-tuning, which tunes all the model parameters with the downstream data. To reduce memory footprint and avoid overfitting, an emerging trend is to tune a small proportion of the model parameters while freezing the majority ones with Parameter-Efficient Fine-Tuning (PEFT) [27, 28, 31]. However, both full fine-tuning and PEFT target an exclusive specific resource budget for each downstream task, while in reality, we often need to deploy multiple models simultaneously to meet various resource demands. This makes us ponder: What is an effective and efficient way to obtain a palette of fine-tuned models meeting different resource constraints?" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.475, + 0.895, + 0.627 + ], + "angle": 0, + "content": "A natural approach is to compress a well-trained large model into numerous smaller ones [11, 21, 29, 36]. However, the computational cost grows linearly with the number of deployment scenarios. Following the once-for-all network [8, 20], a few works first pre-train a weight-sharing over-parameterized supernet, then adapt the supernet to the downstream tasks. Although promising, training supernets at scale with large datasets requires prohibitive computational resources, e.g., thousands of GPU training hours, which is practically infeasible." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.629, + 0.896, + 0.902 + ], + "angle": 0, + "content": "On the other hand, there are many open-source large models [5, 23, 64] from communities such as Hugging-Face [72] that are pre-trained on large-scale datasets and ready to be downloaded. Inspired by the success of model stitching [4, 12, 38], recently, Stitchable Neural Networks (SN-Net) [51] has been proposed to stitch pre-trained models (anchors) of the same family to quickly obtain a set of candidate new networks (stitches) with different accuracy-efficiency trade-offs of a wide FLOPs range. However, the scope of SN-Net is limited to the pre-training classification task on the same source domain [56]. When directly adapting the standard approach of SN-Net to a target domain, it faces new challenges. Specifically, the straightforward way to employ SN-Net in task adaptation typically requires three stages: first adapting anchors individually to the target domain, followed by stitching the adapted anchors, and finally evaluating all stitches to find and deploy the ones on the Pareto Frontier. Such a three-stage adaptation pro" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.863, + 0.422, + 0.876 + ], + "angle": 0, + "content": "†Corresponding author. E-mail: bohan.zhuang@gmail.com" + }, + { + "type": "page_footnote", + "bbox": [ + 0.081, + 0.876, + 0.469, + 0.901 + ], + "angle": 0, + "content": "1Source code will be released at https://github.com/ziplab/Stitched_LLaMA." + }, + { + "type": "list", + "bbox": [ + 0.081, + 0.863, + 0.469, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "28555" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.089, + 0.895, + 0.188 + ], + "angle": 0, + "content": "Table 1. Comparison of frameworks for obtaining a palette of networks to meet different hardware efficiency constraints. Supernet-based methods NAT [47] and TOFA [35] necessitate a supernet training stage plus long deployment GPU hours and are restricted to small FLOPs ranges. SN-Net [51] employs model stitching to achieve a wide FLOPs range with significantly reduced deployment GPU hours. Our ESTA framework further reduces the deployment GPU hours, fine-tuning memory, and trainable parameters, and is scalable to produce LLM stitches with billions of parameters (Section 5.2). The deployment GPU hours and fine-tuning memory for both SN-Net and ESTA are measured on a single NVIDIA GeForce RTX 3090 GPU when stitching ViT-Ti/S/B anchors [16] with batch size 64. * indicates data is inferred from the training recipe of TOFA's supernet architecture FBNetV3 [14]." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.192, + 0.891, + 0.271 + ], + "angle": 0, + "content": "
MethodAdapting supernetDeployment GPU hoursFine-tuning memoryTrainable parameter #FLOPs range (G)LLM friendly
NAT [47]X>1,000->10M[0.2, 0.6]X
TOFA [35]✓>1,000*->10M*[0.2, 2.5]X
SN-Net [51]X19.313,235M124.2M[1.3, 17.6]X
ESTA (ours)X5.09,685M4.6M[1.3, 17.6]✓
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.277, + 0.473, + 0.445 + ], + "angle": 0, + "content": "cess is expensive and sub-optimal. Moreover, adapting SN-Net typically needs to load and optimize all parameters of multiple models, leading to a daunting training-time memory cost. For instance, stitching the two smallest models in the LLaMA family [64, 64] with SN-Net under standard training settings (32-bit floating point parameters and Adam optimizer on a single GPU) requires more than 80G Video RAM that cannot fit into high-end GPUs such as NVIDIA A100. SN-Net also requires saving a separate instance of the model family for each task, making the storage grow linearly with the number of deployed applications." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.446, + 0.473, + 0.793 + ], + "angle": 0, + "content": "Therefore, in this paper, we introduce a novel Efficient Stitchable Task Adaptation (ESTA) framework. Specifically, built upon SN-Net [51], ESTA introduces two main designs to overcome the aforementioned issues. First, to reduce the memory and storage costs, we tailor a Parameter-efficient Stitch fine-Tuning (PST) method, which incorporates the representative sparse fine-tuning technique LoRA [28] to keep the anchors and stitching layer weights frozen, while approximating their updates with trainable low-rank decomposition matrices. In addition to stitch-agnostic LoRA modules, we further introduce stitch-specific bias terms to alleviate the conflict among different stitches. Second, we propose a simple one-stage deployment pipeline to simultaneously adapt and stitch the pretrained anchors to the target domain. Moreover, we propose to estimate and accumulate an importance score for each stitch via the saliency pruning criterion [37]. With the importance scores, we develop a novel stitch-sampling method to assign important stitches with higher sampling probabilities. More importantly, after fine-tuning, the stitch importance scores can be directly used to infer the best ones at the Pareto frontier for deployment. Table 1 compares different efficient frameworks for diverse deployments." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.796, + 0.472, + 0.903 + ], + "angle": 0, + "content": "Overall, our paper has the following major contributions. 1) This is a pioneering work to investigate the problem of efficient fine-tuning of SN-Net for task adaptation. Our solution ESTA is a successful endeavor that delivers dozens of ready-to-deploy vision models up to 17.6G FLOPs in 5 GPU hours as well as multiple chatbot models with billions of parameters by stitching LLaMA mod" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.277, + 0.895, + 0.49 + ], + "angle": 0, + "content": "els [64, 65] for the instruction-following task. 2) We devise a parameter-efficient stitch fine-tuning method that incorporates trainable stitch-agnostic LoRA modules and stitch-specific bias terms, which largely reduces the fine-tuning memory footprints while alleviating the interference issue among stitches. Moreover, we streamline a simple one-stage deployment pipeline with a novel task-specific stitch sampling strategy that greatly reduces the deployment time and improves Pareto frontiers. 3) Extensive experiments conducted on 25 downstream visual recognition tasks demonstrate that when stitching ViT-Ti/S/B [16], our ESTA achieves remarkable performance improvements compared to the direct SN-Net adaptation counterpart while utilizing significantly fewer trainable parameters." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.509, + 0.642, + 0.525 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.536, + 0.895, + 0.884 + ], + "angle": 0, + "content": "Model stitching. Model stitching [4, 38] targets connecting the bottom layers of one network to the top layers of another with a stitching layer. Following [38], model stitching is first employed to measure the similarities for the inner representations learned by deep neural networks [4, 12, 25]. Previous works [4, 12] observe that trained networks with different initializations can be effectively stitched without incurring significant performance drop. Based on this observation, DeRy [76] stitches the blocks dissected from different pre-trained models and assembles a new model based on their quantified similarity for better performance. Herdt et al. [24] propose interpreting the inner representations of a deep network by stitching it with a pre-trained GAN generator. Recently, from the perspective of flexible deployment, SN-Net [51] stitches models of different sizes within a model family to cheaply produce numerous new networks with diverse accuracy and efficiency trade-offs in the pretraining task. In contrast to [51], our work specifically targets obtaining a palette of fine-tuned models for the task adaptation setting. To do so, we make key redesigns including a parameter-efficient stitch fine-tuning method and a simple one-stage deployment pipeline to overcome the efficiency challenges in SN-Net adaptation." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.886, + 0.892, + 0.902 + ], + "angle": 0, + "content": "Parameter-efficient fine-tuning. Parameter-efficient fine" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.946, + 0.521, + 0.957 + ], + "angle": 0, + "content": "28556" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.473, + 0.469 + ], + "angle": 0, + "content": "tuning [27, 28, 31, 40] is a powerful alternative to vanilla full tine-tuning, which updates only a small number of parameters while freezing the majority ones. Freezing the majority of parameters allows us to optimize storage and reduces the burden on training GPU memory, as there is no need to store their gradients or other training time statistics. Recent research freezes the vast majority of parameters while fine-tuning either the parameters that are inherited in the backbone [77, 80] or additionally added, in the form of learnable prompts [31, 39], low-rank bottleneck layers [27], learnable scaling and shifting factors [41], and separated small networks [61]. Meanwhile, PEFT has been employed in numerous downstream applications. For instance, while freezing the most parameters, Polyhistor [45] learns a hyper-network to generate adapter weights [27] for multi-task adaptation, LLaMA-Adapter [83] inserts learnable prompts into Transformer layers for the instruction-following task, and ControlNet [82] employs the zero-initialized convolutions to fine-tune diffusion generative models [55]. However, these works target only one exclusive resource budget and are not scalable to diverse deployment scenarios. Unlike other methods, our ESTA is crafted to yield multiple fine-tuned models with diverse capacities while jointly incorporating stitch-specific and stitch-agnostic lightweight trainable parameters." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.47, + 0.473, + 0.744 + ], + "angle": 0, + "content": "Fine-tuning with multiple models. The presence of more large-scale models has unleashed the potential to utilize multiple models instead of a single one during fine-tuning. One line of work selects the best pre-trained model in the model zoo to fine-tune. To identify the best model, prior studies quantize the transferability of the pre-trained models ahead of fine-tuning by estimating their accuracy on the downstream tasks [3, 17, 33] or the generalization capability to mitigate the domain gaps [6, 49, 66, 78]. Another way is to do model selection using an efficient online learning regime [74]. Considering the diversity among the pre-trained models, research efforts have been devoted to effectively exploit their knowledge with feature aggregation [42, 57, 58], model merging [13, 30, 48, 73], or a mixture-of-experts architecture [59]. In contrast to previous works that merely target better performance, our work employs off-the-shelf pre-trained model families to produce plentiful new models for diverse deployment requirements." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.76, + 0.217, + 0.775 + ], + "angle": 0, + "content": "3. Preliminaries" + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.785, + 0.238, + 0.802 + ], + "angle": 0, + "content": "3.1. Model Stitching" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.81, + 0.472, + 0.903 + ], + "angle": 0, + "content": "Considering an \\(L\\)-layer feed-forward artificial neural network \\(f_{\\theta_1}:\\mathcal{X}\\to \\mathcal{Y}\\) parameterized by \\(\\theta_{1}\\), that maps any input from the input space \\(\\mathcal{X}\\) to the output space \\(\\mathcal{V}\\), \\(f_{\\theta_1}\\) can be denoted as a composition of functions that \\(f_{\\theta_1} = f_L\\circ \\dots \\circ f_1\\), where \\(\\circ\\) denotes the function composition. In model stitching, \\(f_{\\theta_1}\\) can be split up at the \\(l\\)-th layer into two" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.168 + ], + "angle": 0, + "content": "portions of functions where \\(l\\in [1,L - 1]\\) .Given any input \\(\\pmb {X}\\subseteq \\mathcal{X}\\) , the first portion of the head functions maps \\(\\pmb{X}\\) to activations \\(\\pmb {X}_l\\) at layer \\(l\\) , i.e., \\(H_{\\theta_1,l}(\\pmb {X}) = f_l\\circ \\dots \\circ f_1 = \\pmb {X}_l\\) The second portion of the tail functions maps \\(\\pmb {X}_l\\) to the final output, i.e., \\(T_{\\theta_1,l}(\\pmb {X}_l) = f_L\\circ \\dots \\circ f_{l + 1}\\)" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.168, + 0.892, + 0.272 + ], + "angle": 0, + "content": "Given another pre-trained \\(M\\)-layer artificial neural network \\(f_{\\theta_2}\\) parameterized by \\(\\theta_{2}\\) that is split up at the \\(m\\)-th layer that \\(m\\in [2,M]\\), we can then employ a stitching layer \\(S:\\mathcal{A}_{\\theta_1,l}\\to \\mathcal{A}_{\\theta_2,m}\\) to map between the two activation spaces at layers \\(l\\) and \\(m\\). In this case, we obtain a new network parameterized by \\(\\phi\\) and the stitching layer parameters, i.e.," + }, + { + "type": "equation", + "bbox": [ + 0.578, + 0.273, + 0.892, + 0.288 + ], + "angle": 0, + "content": "\\[\nF _ {\\phi , S} (\\boldsymbol {X}) = T _ {\\theta_ {2}, m} \\circ S \\circ H _ {\\theta_ {1}, l} (\\boldsymbol {X}), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.288, + 0.892, + 0.303 + ], + "angle": 0, + "content": "where \\(\\phi\\) consists of partial parameters from both \\(\\theta_{1}\\) and \\(\\theta_{2}\\)." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.31, + 0.743, + 0.325 + ], + "angle": 0, + "content": "3.2. Stitchable Neural Network" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.333, + 0.895, + 0.606 + ], + "angle": 0, + "content": "In Stitchable Neural Network (SN-Net) [51], let a pretrained model family of size \\( Z \\) be \\( \\mathcal{Z} = \\{f_{\\theta_z}\\}_{z=1}^Z \\), where \\( \\theta_z \\) is the model parameter of the \\( z \\)-th model. The goal is to derive additional \\( N \\) candidate new networks \\( \\mathcal{N} = \\{F_{\\phi_n, S_n}\\}_{n=1}^N \\) to adapt to various resource constraints. To do so, SN-Net first selects a pair of pre-trained models (anchors) from \\( \\mathcal{Z} \\) and then stitches them with Eq. (1) at different layer indexes to get \\( N \\) stitches. Then, SN-Net samples the stitches randomly and jointly optimizes them. An overview of SN-Net is depicted in Figure 1 (a). Since the anchors vary in scale, the newly assembled stitches have diverse performances and complexities. Importantly, SN-Net gives practical principles to design the space of \\( \\mathcal{N} \\) that one should 1) stitch a pair of nearest anchors in terms of model complexity in a model family; 2) stitch the head of a faster and smaller anchor to the tail of a larger and slower anchor. Unless specified otherwise, we adopt these as the default experimental settings in this paper." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.617, + 0.593, + 0.633 + ], + "angle": 0, + "content": "4. Method" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.643, + 0.894, + 0.75 + ], + "angle": 0, + "content": "In this section, we introduce our ESTA framework, which consists of two major components: parameter-efficient stitch fine-tuning (PST) and a simple one-stage deployment pipeline. PST is tailored to cheaply fine-tune a palette of stitches (Section 4.1 and Figure 1 (b)). The one-stage deployment pipeline aims to simultaneously save deployment time and improve adaptation performance (Section 4.2)." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.757, + 0.831, + 0.774 + ], + "angle": 0, + "content": "4.1. Parameter-efficient Stitch Fine-tuning" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.781, + 0.892, + 0.87 + ], + "angle": 0, + "content": "To address the problems of large storage and memory consumption when adapting SN-Net to downstream tasks with full fine-tuning, we introduce a parameter-efficient stitch fine-tuning (PST) method as depicted in Figure 1 (b). Our basic idea is to adapt the representative sparse fine-tuning method LoRA [28] for SN-Net fine-tuning." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.871, + 0.892, + 0.901 + ], + "angle": 0, + "content": "LoRA on Transformer layers. Specifically, let any pretrained weight matrix in the multi-head self-attention layer" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "28557" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.084, + 0.094, + 0.303, + 0.197 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.18, + 0.2, + 0.196, + 0.21 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.315, + 0.094, + 0.627, + 0.199 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.463, + 0.2, + 0.479, + 0.21 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image", + "bbox": [ + 0.636, + 0.103, + 0.882, + 0.199 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.764, + 0.2, + 0.779, + 0.21 + ], + "angle": 0, + "content": "(c)" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.213, + 0.894, + 0.325 + ], + "angle": 0, + "content": "Figure 1. (a) Illustration of Stretchable Neural Network [51]. With two anchors from the same model family, SN-Net connects the early layers of the smaller one to the latter layers of the larger one with stitching layers to obtain a set of new networks with different performance-efficiency trade-offs, e.g., the path in Blue. (b) Overview of our PST method tailored for fine-tuning a palette of stitches, which integrates stitch-agnostic LoRA modules with stitch-specific bias terms, aiming to promote diverse representations among stitches while maintaining low trainable parameters. (c) Overview of our task-specific stitch sampling. We estimate the importance scores of the stitches with a scoring function \\( Q(\\cdot, \\cdot) \\) and accumulate them as global statistics with moving averages. For a resource constraint \\( \\gamma \\), we sample with a categorical distribution \\( \\pi(\\mathcal{N}_{\\gamma}) \\) that is parameterized by the normalized importance scores so as to assign the important stitches with higher sampling probabilities. After fine-tuning, we directly deploy the stitches with the highest scores to avoid the costly evaluation stage." + }, + { + "type": "image", + "bbox": [ + 0.08, + 0.33, + 0.273, + 0.446 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.273, + 0.33, + 0.461, + 0.446 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.447, + 0.47, + 0.558 + ], + "angle": 0, + "content": "Figure 2. Distribution of pair-wise gradient angles among stitches when updating shared weights at fine-tuning iteration 600. We highlight angle \\(90^{\\circ}\\) with a dashed red line. For simplicity, we show the gradient angles among the combined query, key, and value projection matrices for a total of 32 stitches when stitching ViT-Ti and ViT-S anchors. Generally, the gradient angles are larger in the target domain Stanford Cars [18] than in the source domain ImageNet-1k [56]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.563, + 0.47, + 0.625 + ], + "angle": 0, + "content": "be \\(W \\in \\mathbb{R}^{d \\times k}\\), we freeze \\(W\\) and insert trainable low-rank decomposition matrices \\(W_{\\mathrm{down}} \\in \\mathbb{R}^{d \\times r}\\) and \\(W_{\\mathrm{up}} \\in \\mathbb{R}^{r \\times k}\\), where \\(r\\) is the pre-defined rank that \\(r \\ll \\min(d, k)\\). In this way, the updated version of \\(W\\) can be formulated as" + }, + { + "type": "equation", + "bbox": [ + 0.182, + 0.634, + 0.469, + 0.652 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {W} ^ {\\prime} \\leftarrow \\boldsymbol {W} + \\boldsymbol {W} _ {\\text {d o w n}} \\boldsymbol {W} _ {\\text {u p}}. \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.659, + 0.47, + 0.704 + ], + "angle": 0, + "content": "We follow [28] to respectively use Gaussian and zero initializations for \\( W_{\\mathrm{up}} \\) and \\( W_{\\mathrm{down}} \\), so that \\( W_{\\mathrm{down}}W_{\\mathrm{up}} \\) is zero at the beginning of fine-tuning." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.705, + 0.471, + 0.903 + ], + "angle": 0, + "content": "LoRA on stitching layers. We further extend the low-rank weight update to stitching layers. In SN-Net, a stitching layer is designed as a full-rank transformation matrix. Without loss of generality, we show the low-rank update of stitching layer \\( S_{n} \\) that stitches \\( f_{\\theta_1} \\) and \\( f_{\\theta_2} \\) at layers \\( l \\) and \\( m \\), respectively. We first follow the initialization of stitching layers in [12, 51] to align the activations at layers \\( l \\) and \\( m - 1 \\). Let the activations be \\( H_{\\theta_1,l}(\\mathbf{X}) \\in \\mathbb{R}^{b \\times d_1} \\) and \\( H_{\\theta_2,m - 1}(\\mathbf{X}) \\in \\mathbb{R}^{b \\times d_2} \\) with sequence length \\( b \\) and feature dimensions \\( d_1 \\) and \\( d_2 \\). The stitching layer is parameterized by a transformation matrix \\( M \\in \\mathbb{R}^{d_1 \\times d_2} \\), which is initialized by solving a least squares problem, i.e., \\( M = H_{\\theta_1,l}(X)^{\\dagger}H_{\\theta_2,m - 1}(X) \\), where \\( H_{\\theta_1,l}(X)^{\\dagger} \\) is the" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.332, + 0.892, + 0.437 + ], + "angle": 0, + "content": "Moore-Penrose pseudoinverse of \\( H_{\\theta_1,l}(\\mathbf{X}) \\). We find that such initialization already has impressive representational capacity on the target domain and they can be updated with low-rank decomposition without hurting the performance too much. Accordingly, we update \\( M \\) similar to updating \\( W \\) in Eq. (2) by approximating its update with two learnable low-rank matrices with the same initializations." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.44, + 0.893, + 0.775 + ], + "angle": 0, + "content": "Although the low-rank updates significantly enhance parameter efficiency, the low-rank essence [2] largely limits the network capacity. Particularly, in task adaptation, we often observe conflicting updates on the shared weights among the stitches. Figure 2 gives an example, where we stitch pre-trained anchors on ImageNet-1k [56] and Stanford Cars [18] and visualize the distribution of pair-wise gradient angles among different stitches when updating shared weights. We find that stitches agree less and conflict more (i.e., with more large angles) on the target domain Stanford Cars compared to the source domain ImageNet-1k. Stitch-specific bias. To address the above issue, we further introduce stitch-specific bias terms as depicted in Figure 1 (b). Particularly, for a linear layer parameterized by \\( W \\) and a bias term \\( b \\), we add bias term \\( b^s \\) when optimizing the \\( s \\)-th stitch, for which the output activation becomes \\( WX + b + b^s \\). We employ a set of distinct bias terms at different layers for each stitch. In this way, the stitches are encouraged to learn distinct feature representations for different resource requirements. To restrict the number of trainable parameters, following [28], we only apply PST to self-attention layers while freezing feed-forward layers." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.786, + 0.779, + 0.803 + ], + "angle": 0, + "content": "4.2. One-stage Deployment Pipeline" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.81, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Simultaneously adapt and stitch anchors. As aforementioned, given a model family \\(\\mathcal{Z}\\) pre-trained in the source domain, directly adapting SN-Net to a target domain typically involves three stages. First, adapting each anchor to the target domain \\(\\mathcal{D}\\) by solving \\(\\theta_z^* = \\mathrm{argmin}_{\\theta_z} \\mathcal{L}(\\theta_z; \\mathcal{D})\\). In the second stage, at each training iteration, SN-Net sam" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "28558" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.091, + 0.473, + 0.317 + ], + "angle": 0, + "content": "plies a stitch and optimizes it with objective \\(\\mathcal{L}(\\phi_n^*, S_n; \\mathcal{D})\\), where \\(\\phi_n^*\\) is the set of SN-Net parameters for stitch \\(F_{\\phi_n, S_n}\\) which has been optimized once in the first stage. In the final stage, all optimized stitches need to be evaluated to identify the ones on the Pareto frontier for diverse resource constraints. Such a three-stage approach is expensive and sub-optimal (see more discussions in the supplementary material). Thus, we propose to adapt and stitch the anchors simultaneously within one stage, by directly optimizing \\(\\mathcal{L}(\\phi_n, S_n; \\mathcal{D})\\) for each sampled stitch. Moreover, we introduce a novel task-specific stitch-sampling method to allow the promising stitches more likely to be sampled, based on a stitch importance score. After fine-tuning, the stitches with the highest scores are naturally selected for deployment without the need for the above final stage evaluation." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.318, + 0.473, + 0.468 + ], + "angle": 0, + "content": "Task-specific stitch sampling. SN-Net assumes that all stitches are equally important and performs random (uniform) sampling during training. However, as recognized by prior research [43, 70] that random sampling causes a gap between training and deployment as only the best stitches on the Pareto frontier are actually deployed. Moreover, since the importance of the pre-trained weights varies in different downstream tasks [19, 22, 75], we argue that the performance of stitches that are parameterized by these weights is also different across these tasks." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.469, + 0.473, + 0.651 + ], + "angle": 0, + "content": "To this end, we propose to assign the important stitches that are likely to be on the Pareto frontier with higher sampling probabilities during fine-tuning to ensure that they are optimized sufficiently. Specifically, we first estimate the importance of each stitch with a data-dependent saliency pruning metric SNIP [37], which measures the importance with the first-order gradient information with barely any extra computational cost. When the \\( n \\)-th stitch is sampled, we can get its importance score \\( Q(F_{\\phi, S_n}, X) \\) with the scoring function \\( Q(\\cdot, \\cdot) \\) [37]. To obtain robust importance scores for sampling, we accumulate scores with moving average during training, i.e.," + }, + { + "type": "equation", + "bbox": [ + 0.144, + 0.661, + 0.47, + 0.681 + ], + "angle": 0, + "content": "\\[\nq _ {n} ^ {t} \\leftarrow \\eta q _ {n} ^ {t - 1} + (1 - \\eta) Q \\left(F _ {\\phi_ {n}, S _ {n}}, \\boldsymbol {X}\\right), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.689, + 0.473, + 0.902 + ], + "angle": 0, + "content": "where \\( q_{n}^{t} \\) and \\( q_{n}^{t - 1} \\) are the importance score at the \\( t \\)-th and \\( (t - 1) \\)-th iteration, respectively, and \\( \\eta \\in [0,1) \\) is the momentum coefficient. In this way, we can get stable importance scores. After a warm-up period that employs uniform sampling to accumulate the scores, we assign the important stitches with higher sampling probabilities for the rest of the fine-tuning epochs. To do so, we uniformly divide the resource constraint range of our stitches into several intervals and sample an interval \\( \\gamma \\) with uniform sampling. Accordingly, we can obtain a subset of stitches \\( \\mathcal{N}_{\\gamma} \\) whose resource constraints belong to \\( \\gamma \\) and their corresponding importance scores \\( \\mathcal{Q}_{\\gamma} \\). We can then define a categorical distribution based on the normalized importance scores, i.e., \\( \\pi (\\mathcal{N}_{\\gamma}) = \\mathrm{softmax}(\\mathcal{Q}_{\\gamma}) \\) and finally get the sampled stitch" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.091, + 0.892, + 0.121 + ], + "angle": 0, + "content": "\\(F_{\\phi_n,S_n}\\sim \\pi (\\mathcal{N}_\\gamma)\\). We show that our task-specific stitch sampling improves the performance in Section 5.3." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.122, + 0.894, + 0.244 + ], + "angle": 0, + "content": "Additionally, with the sampled stitch, SN-Net employs knowledge distillation with a RegNetY-160 [54] pre-trained teacher to improve its performance. However, pre-trained teachers on downstream tasks are often unavailable. Therefore, in each training iteration, we always sample and train the largest anchor (teacher) and transfer its knowledge to the sampled stitch (student) similar to inplace distillation [79]. In practice, we employ the hard-label distillation [26, 63]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.245, + 0.895, + 0.486 + ], + "angle": 0, + "content": "Reuse stitch importance scores for deployment. After fine-tuning, with the set of accumulated importance scores \\( \\mathcal{Q}_{\\gamma} \\) for all stitches, we directly use them to select the stitches with the highest scores to deploy, which eliminates the need for the costly evaluation of the final stage of the straightforward SN-Net adaptation solution. We empirically observe that the deployed stitches are mostly on the Pareto frontier in Section 5.3, suggesting our estimation of the important stitches is accurate. Furthermore, in the supplementary material, we show that the important stitches differ across different tasks, proving the need for task-specific deployment strategies, as highlighted in [19, 22, 75]. It's worth noting that although the SNIP metric has been widely employed as a one-shot metric to efficiently estimate network performance [1, 69], we extend it to estimate stitch importance for better training and selection." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.503, + 0.634, + 0.52 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.528, + 0.681, + 0.544 + ], + "angle": 0, + "content": "5.1. Visual Recognition" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.553, + 0.892, + 0.718 + ], + "angle": 0, + "content": "We evaluate the effectiveness of our method on a total of 25 downstream visual recognition tasks, including fine-grained visual classification (FGVC) benchmark, common visual classification benchmark CIFAR-100 [34], and VTAB-1k [81] benchmark. FGVC benchmark contains NABirds [67], CUB-200-2011 [68], Stanford Cars [18], Stanford Dogs [32], and Oxford Flowers [50] tasks and the VTAB-1k benchmark includes 19 tasks in different domains, each of which has 800 training samples. For visual recognition experiments, we stitch ImageNet-21k pretrained ViT-Ti/S/B models [16] from [60]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.895, + 0.901 + ], + "angle": 0, + "content": "Implementation details. We employ the stitching space and settings (kernel size as 2 and stride as 1) in [51] for downstream visual recognition tasks. We uniformly divide the resource constraints supported by the stitches into around 15 intervals and deploy one stitch within each interval. We set the hyper-parameter \\( r \\) for the rank of weight updating in the self-attention layers to be 32, 16, and 8 for ViT-Ti, ViT-S, and ViT-B, respectively. We also set \\( r = d_1 // 4 \\) universally for all stitching layers and \\( \\eta \\) in Eq. (3) as 0.9 by grid search. We follow [31] to fine-tune 100 epochs on each task and select the other hyper-parameters and augmentation methods. We include more implementation details in" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "28559" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.081, + 0.09, + 0.285, + 0.204 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.284, + 0.09, + 0.487, + 0.204 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.487, + 0.09, + 0.688, + 0.204 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.688, + 0.09, + 0.892, + 0.204 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.205, + 0.894, + 0.247 + ], + "angle": 0, + "content": "Figure 3. Performance comparisons with SN-Net [51] for adapting ViT-Ti/S/B pre-trained on ImageNet-22k [15] to Stanford Cars [18], CUB-200-2011 [68], Stanford Dogs [32], and NABirds [67]. We denote individually fine-tuned anchors as yellow stars. We also show the number of trainable parameters." + }, + { + "type": "image", + "bbox": [ + 0.08, + 0.253, + 0.273, + 0.361 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.274, + 0.253, + 0.468, + 0.361 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.362, + 0.47, + 0.418 + ], + "angle": 0, + "content": "Figure 4. Performance comparisons with SN-Net [51] for adapting ViT-Ti/S/B pre-trained on ImageNet-22k [15] to VTAB-1k [81] and CIFAR-100 [34]. We denote individually fine-tuned anchors as yellow stars and also show the number of trainable parameters." + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.426, + 0.47, + 0.468 + ], + "angle": 0, + "content": "Table 2. Relative response quality to Alpaca-LoRA 7B [62] on Vicuna Bench [10]. Our Stitched LLaMA successfully interpolates the answer quality between Alpaca-LoRA 7B and 13B." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.469, + 0.466, + 0.515 + ], + "angle": 0, + "content": "
Alpaca-LoRAStitched LLaMA
Parameter # (B)6.713.06.97.811.512.513.0
Quality (%)100123113114115120123
" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.523, + 0.264, + 0.537 + ], + "angle": 0, + "content": "the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.538, + 0.47, + 0.734 + ], + "angle": 0, + "content": "Main results. We compare the performance of our ESTA with SN-Net and the individual anchors fine-tuned with LoRA [28] on various datasets. The FLOPs-accuracy curves are visualized in Figures 3 and 4. We report the averaged results over the three task groups on VTAB-1k following [31]. Overall, the stitches obtained by our ESTA exhibit smooth FLOPs-accuracy trade-off curves consistently across all datasets. Notably, the curve for our ESTA outperforms SN-Net by significant margins with much fewer trainable parameters. We show that the inferior results of SN-Net on downstream tasks are led by the sub-optimal designs of fine-tuning all model parameters and a three-stage adaptation pipeline in Section 5.3." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.735, + 0.471, + 0.902 + ], + "angle": 0, + "content": "We also observe that the stitches even outperform the anchors with comparable or lower FLOPs in many cases. For instance, stitches outperform individually fine-tuned ViT-S/B anchors by clear margins on NABirds, CUB-200-2011, Stanford Cars, and Stanford Dogs datasets with comparable computational complexity. This phenomenon is consistent with the observations made in [52, 79]. We conjecture that weight sharing among the stitches serves as a strong regularization to improve their generalization. We show more results on few-shot learning and stitching Convnets [46] in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.253, + 0.741, + 0.27 + ], + "angle": 0, + "content": "5.2. Instruction-following Task" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.277, + 0.892, + 0.413 + ], + "angle": 0, + "content": "We also evaluate the instruction-following capability when stitching the LLaMA family [64] with our ESTA framework. We adapt and fine-tune LLaMA-7B/13B on the 52K instruction-following dataset from [62] to obtain chatbot stitches of varying sizes. The 52K instruction-following data is generated from 175 human-written instruction-output pairs with [71]. Each sample in the dataset is formulated by a task description, the context of the task, and the answer that is generated by GPT-3.5 (text-davinci-003) [7]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.414, + 0.893, + 0.609 + ], + "angle": 0, + "content": "Implementation details. We employ the stitching space and settings (kernel size as 1 and stride as 1) in [51] for the instruction-following task and introduce 40 stitches. We uniformly divide the resource constraints supported by the stitches into 6 intervals and deploy one stitch within each interval. We set \\(\\eta\\) in Eq. (3) as 0.9. The rank \\(r\\) is set to 16 and \\(d_1 // 4\\) for LoRA modules in self-attention and stitching layers, respectively. We employ Adam optimizer and fine-tune with batch size 128 for 10 epochs with gradient accumulation. During the generation stage, we employ top-\\(p\\) sampling to sample from the top subword candidates with the probability \\(p = 0.75\\), conduct the beam search, and set the beam size as 4." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.611, + 0.893, + 0.822 + ], + "angle": 0, + "content": "Results. We term the stitches obtained by our ESTA framework as Stitched LLaMA and compare them with Alpaca-LoRA 7B and 13B that respectively freeze the LLaMA 7B and 13B backbones and update the self-attention weights with LoRA modules [28]. For quantitative results reported in Table 2, we observe that similar to the visual recognition tasks, our Stitched LLaMA with different numbers of parameters generates answers with plausible quality in between Alpaca-LoRA 7B and 13B. Figure 5 gives qualitative comparisons. Our Stitched LLaMAs produce reasonable responses, suggesting the potential of our ESTA for adapting large language models (LLMs) for more downstream language tasks. More examples can be found in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.833, + 0.655, + 0.85 + ], + "angle": 0, + "content": "5.3. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.856, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Effect of parameter-efficient stitch fine-tuning. We investigate the effectiveness of our PST on five FGVC tasks and visualize the averaged results in Figure 6 (a) and (b)." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "28560" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.087, + 0.094, + 0.26, + 0.105 + ], + "angle": 0, + "content": "Instruction: Tell me about alpacas." + }, + { + "type": "image", + "bbox": [ + 0.084, + 0.11, + 0.121, + 0.137 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.124, + 0.115, + 0.408, + 0.137 + ], + "angle": 0, + "content": "Alpaca-LoRA 7B (0.035 s/token): Alpacas are members of the camelid family, which also includes llamas and camels." + }, + { + "type": "text", + "bbox": [ + 0.086, + 0.138, + 0.411, + 0.214 + ], + "angle": 0, + "content": "They are native to the Andes Mountains in South America, where they have been domesticated for thousands of years. Alpacas are known for their soft, luxurious fiber, which is used to make clothing, blankets, and other textiles. They are also raised for their meat, which is considered a delicacy in some parts of the world. Alpacas are gentle, intelligent animals, and they are prized for their docile nature and ability to thrive in a wide range of climates." + }, + { + "type": "image", + "bbox": [ + 0.421, + 0.11, + 0.458, + 0.137 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.462, + 0.115, + 0.878, + 0.214 + ], + "angle": 0, + "content": "Stitched LLaMA 8B (0.037 s/token): Alpacas are herbivorous mammals that belong to the family Camelidae, which also includes camels and camelids. They are native to South American continent and are commonly found in Peru, Argentina, Bolivia, and Chile. They are smaller than their cousin, the camel, and are typically about 80-90 inches tall fully grown. They are known for their soft, shimmering, and light-weight fibers, which tend to make clothing, carpets, and other textile products. Alpacas are herbivores, and catets consists mainly of grasses, leaves, and straw. They are social animals and live in a few up to 100 alpacas, led by a dominant male. Alpacas are known for their calm and mature, and they are often used as livestock for their soft, shimmering fibers." + }, + { + "type": "image", + "bbox": [ + 0.084, + 0.22, + 0.121, + 0.246 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.124, + 0.225, + 0.87, + 0.314 + ], + "angle": 0, + "content": "Stitched LLaMA 11B (0.041 s/token): Alpacas are native to South America and are members of the camelid family. They are known for their soft, fluffy coats and their gentle nature. Alpacas are herbivores, meaning they feed on grass and other vegetation. They have a gestation period of approximately 11 days and usually give birth to a single baby, although twins and triplets are not uncommon. Alpacas are social animals and live in herds, led by a dominant they communicate with each other using a variety of vocalizations, including whistles, grunts, and barks. Alpacas are intelligent animals and are known for curiosity and playful nature. Alpacas have been domesticated for thousands of years and are used for a variety of purposes, including as pets, for their fiber, for their meat. Alpaca fiber is prized for its softness and warmth, making it a popular material for clothing and blankets. Alpaca meat is also considered a key in some parts of the world. Overall, alpacas are gentle, intelligent animals that have been domesticated for thousands of years. They are valued for their meat, and companionship." + }, + { + "type": "title", + "bbox": [ + 0.086, + 0.324, + 0.858, + 0.335 + ], + "angle": 0, + "content": "Instruction: Can you help me write a resignation letter to my current employer, while leaving on good terms and expressing gratitude for the opportunities provided?" + }, + { + "type": "image", + "bbox": [ + 0.085, + 0.34, + 0.119, + 0.365 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.121, + 0.343, + 0.346, + 0.355 + ], + "angle": 0, + "content": "Alpaca-LoRA 7B (0.035 s-token): Dear [Name]" + }, + { + "type": "text", + "bbox": [ + 0.086, + 0.355, + 0.352, + 0.421 + ], + "angle": 0, + "content": "of Employer], I am writing to inform you that I am resigning from my position as [Job Title] with [Company Name]. It has been a pleasure working with you and I am grateful for the opportunities you have provided me. I wish you and the company all the best in the future. Sincerely, [Your Name]" + }, + { + "type": "image", + "bbox": [ + 0.375, + 0.34, + 0.412, + 0.366 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.414, + 0.343, + 0.878, + 0.421 + ], + "angle": 0, + "content": "Stitched LLaMA 8B (0.037 s/token): Dear [Employer], I am writing to inform you that I am resigning from my position at [Company Name] with a heavy heart. I have enjoyed my time with you and the team, and I am grateful for the opportunities that have been provided to me. I like to express my sincere appreciation for the training, support, and opportunities that have provided to me during my time at [Company Name]. I have learned a great deal and have grown really and professionally. Please accept my best wishes for the continued success of [Company and the team. I wish you all the best in the future. Sincerely, [Your Name]" + }, + { + "type": "image", + "bbox": [ + 0.084, + 0.426, + 0.119, + 0.453 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.121, + 0.433, + 0.882, + 0.5 + ], + "angle": 0, + "content": "Sti 118 (0.041 s/ token): Dear [Employer], I am writing this letter to inform you of my decision to resign from my position at [Company]. It has been a great pleasure to work for [Company] and I have learned a lot during my time here. I would like to express my gratitude for the opportunities that I seen provided to me and for the support and guidance that I have received from my colleagues and supervisors. I have grown both professionally and finally during my time at [Company] and I am grateful for the experiences I have had. I am leaving on good terms and would be happy to provide a positive chance for [Company] in the future. If there is anything I can do to assist with the transition, please do not hesitate to reach out. Thank you for the opportunity to offer [Company] and I wish you all the best in your future endeavors. Sincerely, [Your Name]" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.511, + 0.893, + 0.539 + ], + "angle": 0, + "content": "Figure 5. Instruction-following comparison between Stitched LLaMA obtained by our ESTA and the Alpaca-LoRA 7B fine-tuned with LoRA [28]." + }, + { + "type": "image", + "bbox": [ + 0.08, + 0.547, + 0.287, + 0.67 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.284, + 0.548, + 0.486, + 0.671 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.486, + 0.548, + 0.688, + 0.671 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.688, + 0.548, + 0.89, + 0.671 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.075, + 0.674, + 0.893, + 0.786 + ], + "angle": 0, + "content": "Figure 6. Ablation studies with results averaged over five FGVC tasks. (a) Effect of our stitch-agnostic LoRA modules. \"Full FT\" represents full fine-tuning of all the model parameters. \"PST w/ full SL\" employs fully fine-tuned stitching layers. We also show the number of trainable parameters. (b) Effect of our stitch-specific bias terms. \"SSB\" and \"scaled r\" represent stitch-specific bias terms and scaling the rank hyperparameter to reach 4.6M trainable parameters, respectively. (c) Effect of our task-specific stitch sampling. \"Uniform, all\" and \"Uniform, best\" represent all the stitches fine-tuned with uniform sampling and their best ones in each resource constraint interval on the Pareto frontier, respectively. (d) Effect of our strategy to simultaneously adapt and stitch anchors. Our strategy \"Adapt-and-stitch\" takes 100 epochs for fine-tuning. In contrast, \"Adapt-then-stitch\", as a straightforward approach to apply SN-Net, first individually adapts each anchor for a total of 300 epochs, and then fine-tunes SN-Net for another 100 epochs." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.793, + 0.471, + 0.9 + ], + "angle": 0, + "content": "To recall, our PST keeps the majority of the model parameters frozen while introducing trainable stitch-agonistic low-rank decomposition matrices and stitch-specific bias terms. In Figure 6 (a), we observe that when employing LoRA to the multi-head self-attention layers, there are solid performance gains on all FLOPs ranges with significantly fewer trainable parameters, echoing the observations in [31, 41]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.793, + 0.895, + 0.9 + ], + "angle": 0, + "content": "It is suggested that low-rank weight updates is a powerful alternative to full fine-tuning for stitching and adapting anchors and we speculate that low-rank weight update avoids overfitting under limited downstream data. We further employ low-rank weight updates to the stitching layers and save 2.4M more trainable parameters with a slight overall performance drop. We conjecture that the stitching layers" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "28561" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.137 + ], + "angle": 0, + "content": "for large pre-trained models are also in low intrinsic dimensions [2] and least-squares initialization [51] already provide a good initialization for them." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.139, + 0.468, + 0.26 + ], + "angle": 0, + "content": "Figure 6 (b) shows that employing stitch-specific bias terms achieves better performance with affordable 1.4M extra trainable parameters. We also include the baseline that simply scales the low-rank dimension \\( r \\) to reach the same number of trainable parameters as our PST. This baseline is inferior to our PST. We speculate that the stitch-specific bias terms enable flexible adjustments of feature representations for different stitches, leading to enhanced performance." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.262, + 0.468, + 0.488 + ], + "angle": 0, + "content": "Effect of task-specific stitch sampling. We investigate the effectiveness of our sampling strategy (introduced in Section 4.2) on the five FGVC tasks. The averaged results are visualized in Figure 6 (c). We compare our task-specific stitch sampling with the uniform one that is adopted in [51]. We empirically find that for uniform sampling, there is a clear performance gap between the best stitches that will be deployed and the others that will be dropped after training, suggesting the need for concentrating on the important stitches during training. To this end, our task-specific sampling strategy performs better than uniform sampling, indicating that our sampling strategy selects accurate task-specific important stitches. Most importantly, our strategy circumvents the need for the costly evaluation stage in the direct solution of adapting SN-Net for downstream tasks." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.491, + 0.468, + 0.717 + ], + "angle": 0, + "content": "Effect of simultaneously adapting and stitching anchors. We study the effect of simultaneously adapting and stitching anchors strategy (introduced in Section 4.2) on five FGVC tasks. The averaged results are visualized in Figure 6 (d). We compare our approach with the direct SN-Net adaptation that first adapts each anchor individually for a total of 300 epochs before stitching them under the exact setting as ESTA. It can be seen that the direct SN-Net adaptation yields poor results in task adaptation. Fine-tuned individual anchors, especially the larger ones, are likely to overfit the limited downstream data [44, 53]. Accordingly, we speculate that the adapted anchor weights serve as a sub-optimal initialization for the stitches in the context of model stitching. On the other hand, our approach achieves better performance and greatly saves the fine-tuning cost." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.72, + 0.468, + 0.901 + ], + "angle": 0, + "content": "Effect of the PEFT technique choices. We compare the effect of different PEFT technique choices and visualize the averaged results on the five FGVC tasks in Figure 7 left. We experiment to employ other PEFT techniques Adapter [27] and Adaptformer [9]. We follow their default setting to respectively insert bottleneck structures sequentially and in parallel to the feed-forward layers and fix the number of trainable parameters the same as ours for a fair comparison. We observe that the overall performance for different PEFT techniques is similar, suggesting that our ESTA framework is compatible with more PEFT technique choices. Since low-rank weight updates of LoRA [28] can be merged into" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.089, + 0.695, + 0.199 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.696, + 0.089, + 0.891, + 0.199 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.21, + 0.892, + 0.28 + ], + "angle": 0, + "content": "Figure 7. Left: effect of different PEFT technique choices. Right: effect of the rank hyper-parameter \\( r \\) in LoRA modules of self-attention layers, for which the three values represent the ranks in ViT-Ti, ViT-S, and ViT-B, respectively. The results are averaged on five FGVC tasks." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.3, + 0.891, + 0.33 + ], + "angle": 0, + "content": "the backbone after training without extra inference computational cost, we employ LoRA by default." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.335, + 0.892, + 0.501 + ], + "angle": 0, + "content": "Effect of the low-rank hyper-parameter \\( r \\). We investigate the effect of different \\( r \\) which controls the rank when updating weights in self-attention layers on the five FGVC tasks. The averaged results on the five FGVC tasks are visualized in Figure 7 right. We find that setting \\( r \\) to be different values barely has any effect on the overall performance, suggesting our ESTA framework is robust to the rank hyper-parameter \\( r \\). Since setting \\( r \\) to be 32, 16, and 8 for low-rank updating the self-attention weights in ViT-Ti, ViT-S, and ViT-B has slightly higher averaged performance among the stitches, we make it our default setting." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.533, + 0.768, + 0.548 + ], + "angle": 0, + "content": "6. Conclusion and Future Work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.563, + 0.892, + 0.82 + ], + "angle": 0, + "content": "In this paper, we have introduced a novel task adaptation framework to cheaply obtain a palette of fine-tuned networks via model stitching, supporting diverse efficiency-performance tradeoffs at runtime. Specifically, built on SN-Net, we have tailored a parameter-efficient stitch finetuning method, which learns lightweight stitch-agnostic LoRA modules and stitch-specific bias terms while keeping the majority of the parameters frozen. Our design significantly reduces fine-tuning memory and storage costs for downstream task adaptation. Moreover, we have devised a task-specific stitch sampling strategy to assign higher sampling probability to the important stitches during training, which simultaneously improves the Pareto frontiers and avoids a costly evaluation stage. Extensive experiments on 25 downstream visual recognition tasks and the instruction-following task have demonstrated the effectiveness of our proposed framework." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.826, + 0.892, + 0.9 + ], + "angle": 0, + "content": "Limitations. Due to the constraints of computational resources, our experiments are limited to visual recognition and instruction-tuning tasks. In the future, we will explore adapting pre-trained model families to dense prediction and multi-modal tasks." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "28562" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.157 + ], + "angle": 0, + "content": "[1] Mohamed S. Abdelfattah, Abhinav Mehrotra, Lukasz Dudziak, and Nicholas D. Lane. Zero-cost proxies for lightweight nas. In ICLR, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.157, + 0.47, + 0.199 + ], + "angle": 0, + "content": "[2] Armen Aghajanyan, Luke Zettlemoyer, and Sonal Gupta. Intrinsic dimensionality explains the effectiveness of language model fine-tuning. In ACL, 2021. 4, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.2, + 0.471, + 0.255 + ], + "angle": 0, + "content": "[3] Haitham Bou Ammar, Eric Eaton, Matthew E Taylor, Decebal Constantin Mocanu, Kurt Driessens, Gerhard Weiss, and Karl Tuyls. An automated measure of mdp similarity for transfer in reinforcement learning. In AAAIw, 2014. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.256, + 0.47, + 0.297 + ], + "angle": 0, + "content": "[4] Yamini Bansal, Preetum Nakkiran, and Boaz Barak. Revisiting model stitching to compare neural representations. NeurIPS, 34:225-236, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.298, + 0.468, + 0.327 + ], + "angle": 0, + "content": "[5] Hangbo Bao, Li Dong, Songhao Piao, and Furu Wei. Beit: Bert pre-training of image transformers. In ICLR, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.327, + 0.468, + 0.368 + ], + "angle": 0, + "content": "[6] Daniel Bolya, Rohit Mittapalli, and Judy Hoffman. Scalable diverse model selection for accessible transfer learning. NeurIPS, 34:19301-19312, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.369, + 0.468, + 0.436 + ], + "angle": 0, + "content": "[7] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. *NeurIPS*, 33:1877–1901, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.438, + 0.468, + 0.48 + ], + "angle": 0, + "content": "[8] Han Cai, Chuang Gan, Tianzhe Wang, Zhekai Zhang, and Song Han. Once-for-all: Train one network and specialize it for efficient deployment. In ICLR, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.481, + 0.468, + 0.535 + ], + "angle": 0, + "content": "[9] Shoufa Chen, Chongjian Ge, Zhan Tong, Jiangliu Wang, Yibing Song, Jue Wang, and Ping Luo. Adaptformer: Adapting vision transformers for scalable visual recognition. NeurIPS, 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.537, + 0.468, + 0.606 + ], + "angle": 0, + "content": "[10] Wei-Lin Chiang, Zhuohan Li, Zi Lin, Ying Sheng, Zhang-hao Wu, Hao Zhang, Lianmin Zheng, Siyuan Zhuang, Yong-hao Zhuang, Joseph E. Gonzalez, Ion Stoica, and Eric P. Xing. Vicuna: An open-source chatbot impressing gpt-4 with \\(90\\%\\) * chatgpt quality, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.607, + 0.468, + 0.648 + ], + "angle": 0, + "content": "[11] Adam Coates and Andrew Y Ng. The importance of encoding versus training with sparse coding and vector quantization. In ICML, pages 921-928, 2011. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.649, + 0.468, + 0.704 + ], + "angle": 0, + "content": "[12] Adrián Csiszárik, Péter Körösí-Szabó, Akos Matszangosz, Gergely Papp, and Daniel Varga. Similarity and matching of neural network representations. NeurIPS, 34:5656-5668, 2021. 1, 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.705, + 0.468, + 0.733 + ], + "angle": 0, + "content": "[13] Dong Dai and Tong Zhang. Greedy model averaging. NeurIPS, 24, 2011. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.734, + 0.468, + 0.801 + ], + "angle": 0, + "content": "[14] Xiaoliang Dai, Alvin Wan, Peizhao Zhang, Bichen Wu, Zijian He, Zhen Wei, Kan Chen, Yuandong Tian, Matthew Yu, Peter Vajda, et al. Fbnetv3: Joint architecture-recipe search using predictor pretraining. In CVPR, pages 16276–16285, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.803, + 0.468, + 0.845 + ], + "angle": 0, + "content": "[15] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR, pages 248-255, 2009. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.846, + 0.468, + 0.901 + ], + "angle": 0, + "content": "[16] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.471, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2021. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.121, + 0.892, + 0.176 + ], + "angle": 0, + "content": "[17] Eric Eaton, Marie Desjardins, and Terran Lane. Modeling transfer relationships between learning tasks for improved inductive transfer. In ECML PKDD, pages 317-332. Springer, 2008. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.177, + 0.892, + 0.218 + ], + "angle": 0, + "content": "[18] Timnit Gebru, Jonathan Krause, Yilun Wang, Duyun Chen, Jia Deng, and Li Fei-Fei. Fine-grained car detection for visual census estimation. In AAAI, 2017. 4, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.219, + 0.892, + 0.272 + ], + "angle": 0, + "content": "[19] Yunhui Guo, Honghui Shi, Abhishek Kumar, Kristen Grauman, Tajana Rosing, and Rogerio Feris. Spottune: transfer learning through adaptive fine-tuning. In CVPR, pages 4805-4814, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.273, + 0.892, + 0.329 + ], + "angle": 0, + "content": "[20] Zichao Guo, Xiangyu Zhang, Haoyuan Mu, Wen Heng, Zechun Liu, Yichen Wei, and Jian Sun. Single path one-shot neural architecture search with uniform sampling. In ECCV, pages 544-560. Springer, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.33, + 0.892, + 0.37 + ], + "angle": 0, + "content": "[21] Song Han, Jeff Pool, John Tran, and William Dally. Learning both weights and connections for efficient neural network. NeurIPS, 28, 2015. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.371, + 0.892, + 0.412 + ], + "angle": 0, + "content": "[22] Haoyu He, Jianfei Cai, Jing Zhang, Dacheng Tao, and Bohan Zhuang. Sensitivity-aware visual parameter-efficient tuning. In ICCV, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.413, + 0.892, + 0.455 + ], + "angle": 0, + "content": "[23] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollar, and Ross Girshick. Masked autoencoders are scalable vision learners. In CVPR, pages 16000-16009, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.456, + 0.892, + 0.511 + ], + "angle": 0, + "content": "[24] Rudolf Herdt, Maximilian Schmidt, Daniel Otero Baguer, Jean Le'Clerc Arrastia, and Peter Maass. Model stitching and visualization how gan generators can invert networks in real-time. arXiv preprint arXiv:2302.02181, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.511, + 0.892, + 0.553 + ], + "angle": 0, + "content": "[25] Adriano Hernandez, Rumen Dangovski, Peter Y. Lu, and Marin Soljacic. Model stitching: Looking for functional similarity between representations. NeurIPSW, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.554, + 0.892, + 0.58 + ], + "angle": 0, + "content": "[26] Geoffrey Hinton, Oriol Vinyals, and Jeff Dean. Distilling the knowledge in a neural network. NeurIPSW, 2014. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.581, + 0.892, + 0.649 + ], + "angle": 0, + "content": "[27] Neil Houlsby, Andrei Giurgiu, Stanislaw Jastrzebski, Bruna Morrone, Quentin De Laroussilhe, Andrea Gesmundo, Mona Attariyan, and Sylvain Gelly. Parameter-efficient transfer learning for nlp. In ICML, pages 2790-2799, 2019. 1, 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.65, + 0.892, + 0.705 + ], + "angle": 0, + "content": "[28] Edward J Hu, yelong shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. LoRA: Low-rank adaptation of large language models. In ICLR, 2022. 1, 2, 3, 4, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.706, + 0.892, + 0.761 + ], + "angle": 0, + "content": "[29] Itay Hubara, Matthieu Courbariaux, Daniel Soudry, Ran El-Yaniv, and Yoshua Bengio. Quantized neural networks: Training neural networks with low precision weights and activations. JMLR, 18(1):6869-6898, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.762, + 0.892, + 0.815 + ], + "angle": 0, + "content": "[30] Pavel Izmailov, Dmitrii Podoprikhin, Timur Garipov, Dmitry Vetrov, and Andrew Gordon Wilson. Averaging weights leads to wider optima and better generalization. In UAI, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.817, + 0.892, + 0.859 + ], + "angle": 0, + "content": "[31] Menglin Jia, Luming Tang, Bor-Chun Chen, Claire Cardie, Serge Belongie, Bharath Hariharan, and Ser-Nam Lim. Visual prompt tuning. In ECCV, 2022. 1, 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.86, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[32] Aditya Khosla, Nityananda Jayadevaprakash, Bangpeng Yao, and Li Fei-Fei. Novel dataset for fine-grained image categorization. In CVPRW, 2011. 5, 6" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "28563" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.133 + ], + "angle": 0, + "content": "[33] Simon Kornblith, Jonathon Shlens, and Quoc V Le. Do better imagenet models transfer better? In CVPR, pages 2661-2671, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.134, + 0.47, + 0.163 + ], + "angle": 0, + "content": "[34] Alex Krizhevsky. Learning multiple layers of features from tiny images. Technical report, 2009. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.164, + 0.471, + 0.205 + ], + "angle": 0, + "content": "[35] Achintya Kundu, Laura Wynter, Rhui Dih Lee, and Luis Angel Bathen. Tofa: Transfer-once-for-all. arXiv preprint arXiv:2303.15485, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.207, + 0.471, + 0.234 + ], + "angle": 0, + "content": "[36] Yann LeCun, John Denker, and Sara Solla. Optimal brain damage. NeurIPS, 2, 1989. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.235, + 0.471, + 0.276 + ], + "angle": 0, + "content": "[37] Namhoon Lee, Thalaiyasingam Ajanthan, and Philip Torr. Snip: Single-shot network pruning based on connection sensitivity. In ICLR, 2019. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.278, + 0.471, + 0.32 + ], + "angle": 0, + "content": "[38] Karel Lenc and Andrea Vedaldi. Understanding image representations by measuring their equivariance and equivalence. In CVPR, pages 991-999, 2015. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.321, + 0.471, + 0.36 + ], + "angle": 0, + "content": "[39] Brian Lester, Rami Al-Rfou, and Noah Constant. The power of scale for parameter-efficient prompt tuning. In EMNLP, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.362, + 0.471, + 0.39 + ], + "angle": 0, + "content": "[40] Xiang Lisa Li and Percy Liang. Prefix-tuning: Optimizing continuous prompts for generation. In ACL, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.391, + 0.471, + 0.433 + ], + "angle": 0, + "content": "[41] Dongze Lian, Daquan Zhou, Jiashi Feng, and Xinchao Wang. Scaling & shifting your features: A new baseline for efficient model tuning. NeurIPS, 2022. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.434, + 0.471, + 0.462 + ], + "angle": 0, + "content": "[42] Iou-Jen Liu, Jian Peng, and Alexander G Schwing. Knowledge flow: Improve upon your teachers. In ICLR, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.463, + 0.471, + 0.504 + ], + "angle": 0, + "content": "[43] Jing Liu, Jianfei Cai, and Bohan Zhuang. Focusformer: Focusing on what we need via architecture sampler. arXiv preprint arXiv:2208.10861, 2022.5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.505, + 0.471, + 0.56 + ], + "angle": 0, + "content": "[44] Pengfei Liu, Weizhe Yuan, Jinlan Fu, Zhengbao Jiang, Hiroaki Hayashi, and Graham Neubig. Pre-train, prompt, and predict: A systematic survey of prompting methods in natural language processing. ACM Comput. Surv., 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.561, + 0.471, + 0.603 + ], + "angle": 0, + "content": "[45] Yen-Cheng Liu, Chih-Yao Ma, Junjiao Tian, Zijian He, and Zsolt Kira. Polyhistor: Parameter-efficient multi-task adaptation for dense vision tasks. NeurIPS, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.604, + 0.471, + 0.646 + ], + "angle": 0, + "content": "[46] Zhuang Liu, Hanzi Mao, Chao-Yuan Wu, Christoph Feichtenhofer, Trevor Darrell, and Saining Xie. A convnet for the 2020s. In CVPR, pages 11976-11986, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.647, + 0.471, + 0.688 + ], + "angle": 0, + "content": "[47] Zhichao Lu, Gautam Sreekumar, Erik Goodman, Wolfgang Banzhaf, Kalyanmoy Deb, and Vishnu Naresh Boddeti. Neural architecture transfer. TPAMI, 43(9):2971-2989, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.689, + 0.471, + 0.73 + ], + "angle": 0, + "content": "[48] Michael S Matena and Colin A Raffel. Merging models with fisher-weighted averaging. NeurIPS, 35:17703-17716, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.731, + 0.471, + 0.786 + ], + "angle": 0, + "content": "[49] Cuong Nguyen, Tal Hassner, Matthias Seeger, and Cedric Archambeau. Leep: A new measure to evaluate transferability of learned representations. In ICML, pages 7294-7305. PMLR, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.788, + 0.471, + 0.83 + ], + "angle": 0, + "content": "[50] Maria-Elena Nilsback and Andrew Zisserman. Automated flower classification over a large number of classes. In ICVGIP, pages 722-729. IEEE, 2008. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.831, + 0.471, + 0.858 + ], + "angle": 0, + "content": "[51] Zizheng Pan, Jianfei Cai, and Bohan Zhuang. Stretchable neural networks. In CVPR, 2023. 1, 2, 3, 4, 5, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.859, + 0.471, + 0.901 + ], + "angle": 0, + "content": "[52] Hieu Pham, Melody Guan, Barret Zoph, Quoc Le, and Jeff Dean. Efficient neural architecture search via parameters sharing. In ICML, pages 4095-4104. PMLR, 2018. 6" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.471, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.134 + ], + "angle": 0, + "content": "[53] Jo Pleted and Tom Gedeon. Deep transfer learning for image classification: a survey. arXiv preprint arXiv:2205.09904, 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.136, + 0.892, + 0.177 + ], + "angle": 0, + "content": "[54] Ilija Radosavovic, Raj Prateek Kosaraju, Ross Girshick, Kaiming He, and Piotr Dólar. Designing network design spaces. In CVPR, pages 10428-10436, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.179, + 0.892, + 0.233 + ], + "angle": 0, + "content": "[55] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, pages 10684-10695, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.235, + 0.892, + 0.303 + ], + "angle": 0, + "content": "[56] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, et al. Imagenet large scale visual recognition challenge. IJCV, 115:211-252, 2015. 1, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.305, + 0.892, + 0.36 + ], + "angle": 0, + "content": "[57] Andrei A Rusu, Neil C Rabinowitz, Guillaume Desjardins, Hubert Soyer, James Kirkpatrick, Koray Kavukcuoglu, Razvan Pascanu, and Raia Hadsell. Progressive neural networks. arXiv preprint arXiv:1606.04671, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.362, + 0.892, + 0.404 + ], + "angle": 0, + "content": "[58] Yang Shu, Zhi Kou, Zhangjie Cao, Jianmin Wang, and Mingsheng Long. Zoo-tuning: Adaptive transfer from a zoo of models. In ICML, pages 9626-9637. PMLR, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.405, + 0.892, + 0.447 + ], + "angle": 0, + "content": "[59] Yang Shu, Zhangjie Cao, Ziyang Zhang, Jianmin Wang, and Mingsheng Long. Hub-pathway: Transfer learning from a hub of pre-trained models. NeurIPS, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.449, + 0.892, + 0.503 + ], + "angle": 0, + "content": "[60] Andreas Steiner, Alexander Kolesnikov, Xiaohua Zhai, Ross Wightman, Jakob Uszkoreit, and Lucas Beyer. How to train your vit? data, augmentation, and regularization in vision transformers. TMLR, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.505, + 0.892, + 0.546 + ], + "angle": 0, + "content": "[61] Yi-Lin Sung, Jaemin Cho, and Mohit Bansal. Lst: Ladder side-tuning for parameter and memory efficient transfer learning. NeurIPS, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.548, + 0.892, + 0.617 + ], + "angle": 0, + "content": "[62] Rohan Taori, Ishaan Gulrajani, Tianyi Zhang, Yann Dubois, Xuechen Li, Carlos Guestrin, Percy Liang, and Tatsunori B. Hashimoto. Stanford alpaca: An instruction-following llama model. https://github.com/tatsu-lab/stanford_alpaca, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.618, + 0.892, + 0.674 + ], + "angle": 0, + "content": "[63] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Hervé Jégou. Training data-efficient image transformers & distillation through attention. In ICML, pages 10347-10357. PMLR, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.675, + 0.892, + 0.745 + ], + "angle": 0, + "content": "[64] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.746, + 0.892, + 0.815 + ], + "angle": 0, + "content": "[65] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.817, + 0.892, + 0.858 + ], + "angle": 0, + "content": "[66] Anh T Tran, Cuong V Nguyen, and Tal Hassner. Transferability and hardness of supervised classification tasks. In ICCV, pages 1395-1405, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.859, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[67] Grant Van Horn, Steve Branson, Ryan Farrell, Scott Haber, Jessie Barry, Panos Ipeirotis, Pietro Perona, and Serge Belongie. Building a bird recognition app and large scale" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "28564" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.092, + 0.469, + 0.12 + ], + "angle": 0, + "content": "dataset with citizen scientists: The fine print in fine-grained dataset collection. In CVPR, pages 595-604, 2015. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.121, + 0.469, + 0.176 + ], + "angle": 0, + "content": "[68] Catherine Wah, Steve Branson, Peter Welinder, Pietro Perona, and Serge Belongie. The caltech-ucsd birds-200-2011 dataset. Tech. Rep. CNS-TR-2011-001, California Institute of Technology, 2011. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.177, + 0.469, + 0.217 + ], + "angle": 0, + "content": "[69] Chaoqi Wang, Guodong Zhang, and Roger Grosse. Picking winning tickets before training by preserving gradient flow. In ICLR, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.218, + 0.469, + 0.26 + ], + "angle": 0, + "content": "[70] Dilin Wang, Meng Li, Chengyue Gong, and Vikas Chandra. Attentivenas: Improving neural architecture search via attentive sampling. In CVPR, pages 6418-6427, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.261, + 0.469, + 0.316 + ], + "angle": 0, + "content": "[71] Yizhong Wang, Yeganeh Kordi, Swaroop Mishra, Alisa Liu, Noah A Smith, Daniel Khashabi, and Hannaneh Hajishirzi. Self-instruct: Aligning language model with self generated instructions. arXiv preprint arXiv:2212.10560, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.317, + 0.469, + 0.385 + ], + "angle": 0, + "content": "[72] Thomas Wolf, Lysandre Debut, Victor Sanh, Julien Chau-mond, Clement Delangue, Anthony Moi, Pierrick Cistac, Tim Rault, Rémi Louf, Morgan Funtowicz, et al. Transformers: State-of-the-art natural language processing. In EMNLP, pages 38-45, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.386, + 0.469, + 0.481 + ], + "angle": 0, + "content": "[73] Mitchell Wortsman, Gabriel Ilharco, Samir Ya Gadre, Rebecca Roelofs, Raphael Gontijo-Lopes, Ari S Morcos, Hongseok Namkoong, Ali Farhadi, Yair Carmon, Simon Kornblith, et al. Model soups: averaging weights of multiple fine-tuned models improves accuracy without increasing inference time. In ICML, pages 23965-23998. PMLR, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.483, + 0.469, + 0.538 + ], + "angle": 0, + "content": "[74] Hang Xu, Ning Kang, Gengwei Zhang, Chuanlong Xie, Xiaodan Liang, and Zhenguo Li. Nasoa: Towards faster task-oriented online fine-tuning with a zoo of models. In ICCV, pages 5097-5106, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.539, + 0.469, + 0.594 + ], + "angle": 0, + "content": "[75] Runxin Xu, Fuli Luo, Zhiyuan Zhang, Chuanqi Tan, Baobao Chang, Songfang Huang, and Fei Huang. Raise a child in large language model: Towards effective and generalizable fine-tuning. In EMNLP, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.595, + 0.469, + 0.636 + ], + "angle": 0, + "content": "[76] Xingyi Yang, Daquan Zhou, Songhua Liu, Jingwen Ye, and Xinchao Wang. Deep model reassembly. NeurIPS, 35: 25739-25753, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.637, + 0.469, + 0.677 + ], + "angle": 0, + "content": "[77] Jason Yosinski, Jeff Clune, Yoshua Bengio, and Hod Lipson. How transferable are features in deep neural networks? NeurIPS, 27, 2014. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.679, + 0.469, + 0.733 + ], + "angle": 0, + "content": "[78] Kaichao You, Yong Liu, Jianmin Wang, and Mingsheng Long. Logme: Practical assessment of pre-trained models for transfer learning. In ICML, pages 12133-12143. PMLR, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.734, + 0.469, + 0.774 + ], + "angle": 0, + "content": "[79] Jiahui Yu and Thomas S Huang. Universally slimmable networks and improved training techniques. In ICCV, pages 1803-1811, 2019. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.776, + 0.469, + 0.817 + ], + "angle": 0, + "content": "[80] Elad Ben Zaken, Yoav Goldberg, and Shauli Ravfogel. Bitfit: Simple parameter-efficient fine-tuning for transformer-based masked language-models. In ACL, pages 1-9, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.818, + 0.469, + 0.9 + ], + "angle": 0, + "content": "[81] Xiaohua Zhai, Joan Puigcerver, Alexander Kolesnikov, Pierre Ruyssen, Carlos Riquelme, Mario Lucic, Josip Djolonga, Andre Susano Pinto, Maxim Neumann, Alexey Dosovitskiy, et al. A large-scale study of representation learning with the visual task adaptation benchmark. arXiv preprint arXiv:1910.04867, 2019. 5, 6" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.469, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "[82] Lvmin Zhang and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In ICCV, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.121, + 0.892, + 0.189 + ], + "angle": 0, + "content": "[83] Renrui Zhang, Jiaming Han, Aojun Zhou, Xiangfei Hu, Shilin Yan, Pan Lu, Hongsheng Li, Peng Gao, and Yu Qiao. Llama-adapter: Efficient fine-tuning of language models with zero-init attention. arXiv preprint arXiv:2303.16199, 2023. 3" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.189 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.518, + 0.957 + ], + "angle": 0, + "content": "28565" + } + ] +] \ No newline at end of file diff --git a/2024/Efficient Stitchable Task Adaptation/03126c3f-71ba-4532-921c-1c95d792e515_origin.pdf b/2024/Efficient Stitchable Task Adaptation/03126c3f-71ba-4532-921c-1c95d792e515_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..4cdd6c025e7285e4c05f00fd57de8d741e064d61 --- /dev/null +++ b/2024/Efficient Stitchable Task Adaptation/03126c3f-71ba-4532-921c-1c95d792e515_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e5b889ac7304a265c31525e5b539b9258cbd6a15f8b872fa7b30c4058eea6256 +size 1051628 diff --git a/2024/Efficient Stitchable Task Adaptation/full.md b/2024/Efficient Stitchable Task Adaptation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..3e95841c30212739b2188f4ed9826bf594e16c19 --- /dev/null +++ b/2024/Efficient Stitchable Task Adaptation/full.md @@ -0,0 +1,330 @@ +# Efficient Stitchable Task Adaptation + +Haoyu He Zizheng Pan Jing Liu Jianfei Cai Bohan Zhuang + +ZIP Lab, Monash University, Australia + +# Abstract + +The paradigm of pre-training and fine-tuning has laid the foundation for deploying deep learning models. However, most fine-tuning methods are designed to meet a specific resource budget. Recently, considering diverse deployment scenarios with various resource budgets, SN-Net [51] is introduced to quickly obtain numerous new networks (stitches) from the pre-trained models (anchors) in a model family via model stitching. Although promising, SN-Net confronts new challenges when adapting it to new target domains, including huge memory and storage requirements and a long and sub-optimal multistage adaptation process. In this work, we present a novel framework, Efficient Stitchable Task Adaptation (ESTA), to efficiently produce a palette of fine-tuned models that adhere to diverse resource constraints. Specifically, we first tailor parameter-efficient fine-tuning to share low-rank updates among the stitches while maintaining independent bias terms. In this way, we largely reduce fine-tuning memory burdens and mitigate the interference among stitches that arises in task adaptation. Furthermore, we streamline a simple yet effective one-stage deployment pipeline, which estimates the important stitches to deploy with training-time gradient statistics. By assigning higher sampling probabilities to important stitches, we also get a boosted Pareto frontier. Extensive experiments on 25 downstream visual recognition tasks demonstrate that our ESTA is capable of generating stitches with smooth accuracy-efficiency trade-offs and surpasses the direct SN-Net adaptation by remarkable margins with significantly lower training time and fewer trainable parameters. Furthermore, we demonstrate the flexibility and scalability of our ESTA framework by stitching LLMs from LLaMA family, obtaining chatbot stitches of assorted sizes1. + +# 1. Introduction + +The paradigm of pre-training and fine-tuning has underpinned modern applications in both vision and lan + +guage. With off-the-shelf models pre-trained on large-scale datasets, the de-facto choice is vanilla full fine-tuning, which tunes all the model parameters with the downstream data. To reduce memory footprint and avoid overfitting, an emerging trend is to tune a small proportion of the model parameters while freezing the majority ones with Parameter-Efficient Fine-Tuning (PEFT) [27, 28, 31]. However, both full fine-tuning and PEFT target an exclusive specific resource budget for each downstream task, while in reality, we often need to deploy multiple models simultaneously to meet various resource demands. This makes us ponder: What is an effective and efficient way to obtain a palette of fine-tuned models meeting different resource constraints? + +A natural approach is to compress a well-trained large model into numerous smaller ones [11, 21, 29, 36]. However, the computational cost grows linearly with the number of deployment scenarios. Following the once-for-all network [8, 20], a few works first pre-train a weight-sharing over-parameterized supernet, then adapt the supernet to the downstream tasks. Although promising, training supernets at scale with large datasets requires prohibitive computational resources, e.g., thousands of GPU training hours, which is practically infeasible. + +On the other hand, there are many open-source large models [5, 23, 64] from communities such as Hugging-Face [72] that are pre-trained on large-scale datasets and ready to be downloaded. Inspired by the success of model stitching [4, 12, 38], recently, Stitchable Neural Networks (SN-Net) [51] has been proposed to stitch pre-trained models (anchors) of the same family to quickly obtain a set of candidate new networks (stitches) with different accuracy-efficiency trade-offs of a wide FLOPs range. However, the scope of SN-Net is limited to the pre-training classification task on the same source domain [56]. When directly adapting the standard approach of SN-Net to a target domain, it faces new challenges. Specifically, the straightforward way to employ SN-Net in task adaptation typically requires three stages: first adapting anchors individually to the target domain, followed by stitching the adapted anchors, and finally evaluating all stitches to find and deploy the ones on the Pareto Frontier. Such a three-stage adaptation pro + +Table 1. Comparison of frameworks for obtaining a palette of networks to meet different hardware efficiency constraints. Supernet-based methods NAT [47] and TOFA [35] necessitate a supernet training stage plus long deployment GPU hours and are restricted to small FLOPs ranges. SN-Net [51] employs model stitching to achieve a wide FLOPs range with significantly reduced deployment GPU hours. Our ESTA framework further reduces the deployment GPU hours, fine-tuning memory, and trainable parameters, and is scalable to produce LLM stitches with billions of parameters (Section 5.2). The deployment GPU hours and fine-tuning memory for both SN-Net and ESTA are measured on a single NVIDIA GeForce RTX 3090 GPU when stitching ViT-Ti/S/B anchors [16] with batch size 64. * indicates data is inferred from the training recipe of TOFA's supernet architecture FBNetV3 [14]. + +
MethodAdapting supernetDeployment GPU hoursFine-tuning memoryTrainable parameter #FLOPs range (G)LLM friendly
NAT [47]X>1,000->10M[0.2, 0.6]X
TOFA [35]✓>1,000*->10M*[0.2, 2.5]X
SN-Net [51]X19.313,235M124.2M[1.3, 17.6]X
ESTA (ours)X5.09,685M4.6M[1.3, 17.6]✓
+ +cess is expensive and sub-optimal. Moreover, adapting SN-Net typically needs to load and optimize all parameters of multiple models, leading to a daunting training-time memory cost. For instance, stitching the two smallest models in the LLaMA family [64, 64] with SN-Net under standard training settings (32-bit floating point parameters and Adam optimizer on a single GPU) requires more than 80G Video RAM that cannot fit into high-end GPUs such as NVIDIA A100. SN-Net also requires saving a separate instance of the model family for each task, making the storage grow linearly with the number of deployed applications. + +Therefore, in this paper, we introduce a novel Efficient Stitchable Task Adaptation (ESTA) framework. Specifically, built upon SN-Net [51], ESTA introduces two main designs to overcome the aforementioned issues. First, to reduce the memory and storage costs, we tailor a Parameter-efficient Stitch fine-Tuning (PST) method, which incorporates the representative sparse fine-tuning technique LoRA [28] to keep the anchors and stitching layer weights frozen, while approximating their updates with trainable low-rank decomposition matrices. In addition to stitch-agnostic LoRA modules, we further introduce stitch-specific bias terms to alleviate the conflict among different stitches. Second, we propose a simple one-stage deployment pipeline to simultaneously adapt and stitch the pretrained anchors to the target domain. Moreover, we propose to estimate and accumulate an importance score for each stitch via the saliency pruning criterion [37]. With the importance scores, we develop a novel stitch-sampling method to assign important stitches with higher sampling probabilities. More importantly, after fine-tuning, the stitch importance scores can be directly used to infer the best ones at the Pareto frontier for deployment. Table 1 compares different efficient frameworks for diverse deployments. + +Overall, our paper has the following major contributions. 1) This is a pioneering work to investigate the problem of efficient fine-tuning of SN-Net for task adaptation. Our solution ESTA is a successful endeavor that delivers dozens of ready-to-deploy vision models up to 17.6G FLOPs in 5 GPU hours as well as multiple chatbot models with billions of parameters by stitching LLaMA mod + +els [64, 65] for the instruction-following task. 2) We devise a parameter-efficient stitch fine-tuning method that incorporates trainable stitch-agnostic LoRA modules and stitch-specific bias terms, which largely reduces the fine-tuning memory footprints while alleviating the interference issue among stitches. Moreover, we streamline a simple one-stage deployment pipeline with a novel task-specific stitch sampling strategy that greatly reduces the deployment time and improves Pareto frontiers. 3) Extensive experiments conducted on 25 downstream visual recognition tasks demonstrate that when stitching ViT-Ti/S/B [16], our ESTA achieves remarkable performance improvements compared to the direct SN-Net adaptation counterpart while utilizing significantly fewer trainable parameters. + +# 2. Related Work + +Model stitching. Model stitching [4, 38] targets connecting the bottom layers of one network to the top layers of another with a stitching layer. Following [38], model stitching is first employed to measure the similarities for the inner representations learned by deep neural networks [4, 12, 25]. Previous works [4, 12] observe that trained networks with different initializations can be effectively stitched without incurring significant performance drop. Based on this observation, DeRy [76] stitches the blocks dissected from different pre-trained models and assembles a new model based on their quantified similarity for better performance. Herdt et al. [24] propose interpreting the inner representations of a deep network by stitching it with a pre-trained GAN generator. Recently, from the perspective of flexible deployment, SN-Net [51] stitches models of different sizes within a model family to cheaply produce numerous new networks with diverse accuracy and efficiency trade-offs in the pretraining task. In contrast to [51], our work specifically targets obtaining a palette of fine-tuned models for the task adaptation setting. To do so, we make key redesigns including a parameter-efficient stitch fine-tuning method and a simple one-stage deployment pipeline to overcome the efficiency challenges in SN-Net adaptation. + +Parameter-efficient fine-tuning. Parameter-efficient fine + +tuning [27, 28, 31, 40] is a powerful alternative to vanilla full tine-tuning, which updates only a small number of parameters while freezing the majority ones. Freezing the majority of parameters allows us to optimize storage and reduces the burden on training GPU memory, as there is no need to store their gradients or other training time statistics. Recent research freezes the vast majority of parameters while fine-tuning either the parameters that are inherited in the backbone [77, 80] or additionally added, in the form of learnable prompts [31, 39], low-rank bottleneck layers [27], learnable scaling and shifting factors [41], and separated small networks [61]. Meanwhile, PEFT has been employed in numerous downstream applications. For instance, while freezing the most parameters, Polyhistor [45] learns a hyper-network to generate adapter weights [27] for multi-task adaptation, LLaMA-Adapter [83] inserts learnable prompts into Transformer layers for the instruction-following task, and ControlNet [82] employs the zero-initialized convolutions to fine-tune diffusion generative models [55]. However, these works target only one exclusive resource budget and are not scalable to diverse deployment scenarios. Unlike other methods, our ESTA is crafted to yield multiple fine-tuned models with diverse capacities while jointly incorporating stitch-specific and stitch-agnostic lightweight trainable parameters. + +Fine-tuning with multiple models. The presence of more large-scale models has unleashed the potential to utilize multiple models instead of a single one during fine-tuning. One line of work selects the best pre-trained model in the model zoo to fine-tune. To identify the best model, prior studies quantize the transferability of the pre-trained models ahead of fine-tuning by estimating their accuracy on the downstream tasks [3, 17, 33] or the generalization capability to mitigate the domain gaps [6, 49, 66, 78]. Another way is to do model selection using an efficient online learning regime [74]. Considering the diversity among the pre-trained models, research efforts have been devoted to effectively exploit their knowledge with feature aggregation [42, 57, 58], model merging [13, 30, 48, 73], or a mixture-of-experts architecture [59]. In contrast to previous works that merely target better performance, our work employs off-the-shelf pre-trained model families to produce plentiful new models for diverse deployment requirements. + +# 3. Preliminaries + +# 3.1. Model Stitching + +Considering an $L$ -layer feed-forward artificial neural network $f_{\theta_1}:\mathcal{X}\to \mathcal{Y}$ parameterized by $\theta_{1}$ , that maps any input from the input space $\mathcal{X}$ to the output space $\mathcal{V}$ , $f_{\theta_1}$ can be denoted as a composition of functions that $f_{\theta_1} = f_L\circ \dots \circ f_1$ , where $\circ$ denotes the function composition. In model stitching, $f_{\theta_1}$ can be split up at the $l$ -th layer into two + +portions of functions where $l\in [1,L - 1]$ .Given any input $\pmb {X}\subseteq \mathcal{X}$ , the first portion of the head functions maps $\pmb{X}$ to activations $\pmb {X}_l$ at layer $l$ , i.e., $H_{\theta_1,l}(\pmb {X}) = f_l\circ \dots \circ f_1 = \pmb {X}_l$ The second portion of the tail functions maps $\pmb {X}_l$ to the final output, i.e., $T_{\theta_1,l}(\pmb {X}_l) = f_L\circ \dots \circ f_{l + 1}$ + +Given another pre-trained $M$ -layer artificial neural network $f_{\theta_2}$ parameterized by $\theta_{2}$ that is split up at the $m$ -th layer that $m\in [2,M]$ , we can then employ a stitching layer $S:\mathcal{A}_{\theta_1,l}\to \mathcal{A}_{\theta_2,m}$ to map between the two activation spaces at layers $l$ and $m$ . In this case, we obtain a new network parameterized by $\phi$ and the stitching layer parameters, i.e., + +$$ +F _ {\phi , S} (\boldsymbol {X}) = T _ {\theta_ {2}, m} \circ S \circ H _ {\theta_ {1}, l} (\boldsymbol {X}), \tag {1} +$$ + +where $\phi$ consists of partial parameters from both $\theta_{1}$ and $\theta_{2}$ . + +# 3.2. Stitchable Neural Network + +In Stitchable Neural Network (SN-Net) [51], let a pretrained model family of size $Z$ be $\mathcal{Z} = \{f_{\theta_z}\}_{z=1}^Z$ , where $\theta_z$ is the model parameter of the $z$ -th model. The goal is to derive additional $N$ candidate new networks $\mathcal{N} = \{F_{\phi_n, S_n}\}_{n=1}^N$ to adapt to various resource constraints. To do so, SN-Net first selects a pair of pre-trained models (anchors) from $\mathcal{Z}$ and then stitches them with Eq. (1) at different layer indexes to get $N$ stitches. Then, SN-Net samples the stitches randomly and jointly optimizes them. An overview of SN-Net is depicted in Figure 1 (a). Since the anchors vary in scale, the newly assembled stitches have diverse performances and complexities. Importantly, SN-Net gives practical principles to design the space of $\mathcal{N}$ that one should 1) stitch a pair of nearest anchors in terms of model complexity in a model family; 2) stitch the head of a faster and smaller anchor to the tail of a larger and slower anchor. Unless specified otherwise, we adopt these as the default experimental settings in this paper. + +# 4. Method + +In this section, we introduce our ESTA framework, which consists of two major components: parameter-efficient stitch fine-tuning (PST) and a simple one-stage deployment pipeline. PST is tailored to cheaply fine-tune a palette of stitches (Section 4.1 and Figure 1 (b)). The one-stage deployment pipeline aims to simultaneously save deployment time and improve adaptation performance (Section 4.2). + +# 4.1. Parameter-efficient Stitch Fine-tuning + +To address the problems of large storage and memory consumption when adapting SN-Net to downstream tasks with full fine-tuning, we introduce a parameter-efficient stitch fine-tuning (PST) method as depicted in Figure 1 (b). Our basic idea is to adapt the representative sparse fine-tuning method LoRA [28] for SN-Net fine-tuning. + +LoRA on Transformer layers. Specifically, let any pretrained weight matrix in the multi-head self-attention layer + +![](images/f58364be157338b8aede459082270502387115ac7876c1765f57901f865d8003.jpg) +(a) + +![](images/e73f4f8d58c28c8099c1a7173896f46f9c4ebe670937baed455fa23fa39f713d.jpg) +(b) + +![](images/73e269c68c0a28a129cb02ef51ae34e723c6d46556b82f43ab9cbab09c2b5e86.jpg) +(c) + +![](images/61edbb8eb4bb41ca5c949343fd8c80b8b6d8584515dc95f17e9a0378ea3f20db.jpg) +Figure 1. (a) Illustration of Stretchable Neural Network [51]. With two anchors from the same model family, SN-Net connects the early layers of the smaller one to the latter layers of the larger one with stitching layers to obtain a set of new networks with different performance-efficiency trade-offs, e.g., the path in Blue. (b) Overview of our PST method tailored for fine-tuning a palette of stitches, which integrates stitch-agnostic LoRA modules with stitch-specific bias terms, aiming to promote diverse representations among stitches while maintaining low trainable parameters. (c) Overview of our task-specific stitch sampling. We estimate the importance scores of the stitches with a scoring function $Q(\cdot, \cdot)$ and accumulate them as global statistics with moving averages. For a resource constraint $\gamma$ , we sample with a categorical distribution $\pi(\mathcal{N}_{\gamma})$ that is parameterized by the normalized importance scores so as to assign the important stitches with higher sampling probabilities. After fine-tuning, we directly deploy the stitches with the highest scores to avoid the costly evaluation stage. + +![](images/51758cda33ca580a38fa388b6ee0a71a8833b34eccb6413244410d7595ac2eae.jpg) +Figure 2. Distribution of pair-wise gradient angles among stitches when updating shared weights at fine-tuning iteration 600. We highlight angle $90^{\circ}$ with a dashed red line. For simplicity, we show the gradient angles among the combined query, key, and value projection matrices for a total of 32 stitches when stitching ViT-Ti and ViT-S anchors. Generally, the gradient angles are larger in the target domain Stanford Cars [18] than in the source domain ImageNet-1k [56]. + +be $W \in \mathbb{R}^{d \times k}$ , we freeze $W$ and insert trainable low-rank decomposition matrices $W_{\mathrm{down}} \in \mathbb{R}^{d \times r}$ and $W_{\mathrm{up}} \in \mathbb{R}^{r \times k}$ , where $r$ is the pre-defined rank that $r \ll \min(d, k)$ . In this way, the updated version of $W$ can be formulated as + +$$ +\boldsymbol {W} ^ {\prime} \leftarrow \boldsymbol {W} + \boldsymbol {W} _ {\text {d o w n}} \boldsymbol {W} _ {\text {u p}}. \tag {2} +$$ + +We follow [28] to respectively use Gaussian and zero initializations for $W_{\mathrm{up}}$ and $W_{\mathrm{down}}$ , so that $W_{\mathrm{down}}W_{\mathrm{up}}$ is zero at the beginning of fine-tuning. + +LoRA on stitching layers. We further extend the low-rank weight update to stitching layers. In SN-Net, a stitching layer is designed as a full-rank transformation matrix. Without loss of generality, we show the low-rank update of stitching layer $S_{n}$ that stitches $f_{\theta_1}$ and $f_{\theta_2}$ at layers $l$ and $m$ , respectively. We first follow the initialization of stitching layers in [12, 51] to align the activations at layers $l$ and $m - 1$ . Let the activations be $H_{\theta_1,l}(\mathbf{X}) \in \mathbb{R}^{b \times d_1}$ and $H_{\theta_2,m - 1}(\mathbf{X}) \in \mathbb{R}^{b \times d_2}$ with sequence length $b$ and feature dimensions $d_1$ and $d_2$ . The stitching layer is parameterized by a transformation matrix $M \in \mathbb{R}^{d_1 \times d_2}$ , which is initialized by solving a least squares problem, i.e., $M = H_{\theta_1,l}(X)^{\dagger}H_{\theta_2,m - 1}(X)$ , where $H_{\theta_1,l}(X)^{\dagger}$ is the + +Moore-Penrose pseudoinverse of $H_{\theta_1,l}(\mathbf{X})$ . We find that such initialization already has impressive representational capacity on the target domain and they can be updated with low-rank decomposition without hurting the performance too much. Accordingly, we update $M$ similar to updating $W$ in Eq. (2) by approximating its update with two learnable low-rank matrices with the same initializations. + +Although the low-rank updates significantly enhance parameter efficiency, the low-rank essence [2] largely limits the network capacity. Particularly, in task adaptation, we often observe conflicting updates on the shared weights among the stitches. Figure 2 gives an example, where we stitch pre-trained anchors on ImageNet-1k [56] and Stanford Cars [18] and visualize the distribution of pair-wise gradient angles among different stitches when updating shared weights. We find that stitches agree less and conflict more (i.e., with more large angles) on the target domain Stanford Cars compared to the source domain ImageNet-1k. Stitch-specific bias. To address the above issue, we further introduce stitch-specific bias terms as depicted in Figure 1 (b). Particularly, for a linear layer parameterized by $W$ and a bias term $b$ , we add bias term $b^s$ when optimizing the $s$ -th stitch, for which the output activation becomes $WX + b + b^s$ . We employ a set of distinct bias terms at different layers for each stitch. In this way, the stitches are encouraged to learn distinct feature representations for different resource requirements. To restrict the number of trainable parameters, following [28], we only apply PST to self-attention layers while freezing feed-forward layers. + +# 4.2. One-stage Deployment Pipeline + +Simultaneously adapt and stitch anchors. As aforementioned, given a model family $\mathcal{Z}$ pre-trained in the source domain, directly adapting SN-Net to a target domain typically involves three stages. First, adapting each anchor to the target domain $\mathcal{D}$ by solving $\theta_z^* = \mathrm{argmin}_{\theta_z} \mathcal{L}(\theta_z; \mathcal{D})$ . In the second stage, at each training iteration, SN-Net sam + +plies a stitch and optimizes it with objective $\mathcal{L}(\phi_n^*, S_n; \mathcal{D})$ , where $\phi_n^*$ is the set of SN-Net parameters for stitch $F_{\phi_n, S_n}$ which has been optimized once in the first stage. In the final stage, all optimized stitches need to be evaluated to identify the ones on the Pareto frontier for diverse resource constraints. Such a three-stage approach is expensive and sub-optimal (see more discussions in the supplementary material). Thus, we propose to adapt and stitch the anchors simultaneously within one stage, by directly optimizing $\mathcal{L}(\phi_n, S_n; \mathcal{D})$ for each sampled stitch. Moreover, we introduce a novel task-specific stitch-sampling method to allow the promising stitches more likely to be sampled, based on a stitch importance score. After fine-tuning, the stitches with the highest scores are naturally selected for deployment without the need for the above final stage evaluation. + +Task-specific stitch sampling. SN-Net assumes that all stitches are equally important and performs random (uniform) sampling during training. However, as recognized by prior research [43, 70] that random sampling causes a gap between training and deployment as only the best stitches on the Pareto frontier are actually deployed. Moreover, since the importance of the pre-trained weights varies in different downstream tasks [19, 22, 75], we argue that the performance of stitches that are parameterized by these weights is also different across these tasks. + +To this end, we propose to assign the important stitches that are likely to be on the Pareto frontier with higher sampling probabilities during fine-tuning to ensure that they are optimized sufficiently. Specifically, we first estimate the importance of each stitch with a data-dependent saliency pruning metric SNIP [37], which measures the importance with the first-order gradient information with barely any extra computational cost. When the $n$ -th stitch is sampled, we can get its importance score $Q(F_{\phi, S_n}, X)$ with the scoring function $Q(\cdot, \cdot)$ [37]. To obtain robust importance scores for sampling, we accumulate scores with moving average during training, i.e., + +$$ +q _ {n} ^ {t} \leftarrow \eta q _ {n} ^ {t - 1} + (1 - \eta) Q \left(F _ {\phi_ {n}, S _ {n}}, \boldsymbol {X}\right), \tag {3} +$$ + +where $q_{n}^{t}$ and $q_{n}^{t - 1}$ are the importance score at the $t$ -th and $(t - 1)$ -th iteration, respectively, and $\eta \in [0,1)$ is the momentum coefficient. In this way, we can get stable importance scores. After a warm-up period that employs uniform sampling to accumulate the scores, we assign the important stitches with higher sampling probabilities for the rest of the fine-tuning epochs. To do so, we uniformly divide the resource constraint range of our stitches into several intervals and sample an interval $\gamma$ with uniform sampling. Accordingly, we can obtain a subset of stitches $\mathcal{N}_{\gamma}$ whose resource constraints belong to $\gamma$ and their corresponding importance scores $\mathcal{Q}_{\gamma}$ . We can then define a categorical distribution based on the normalized importance scores, i.e., $\pi (\mathcal{N}_{\gamma}) = \mathrm{softmax}(\mathcal{Q}_{\gamma})$ and finally get the sampled stitch + +$F_{\phi_n,S_n}\sim \pi (\mathcal{N}_\gamma)$ . We show that our task-specific stitch sampling improves the performance in Section 5.3. + +Additionally, with the sampled stitch, SN-Net employs knowledge distillation with a RegNetY-160 [54] pre-trained teacher to improve its performance. However, pre-trained teachers on downstream tasks are often unavailable. Therefore, in each training iteration, we always sample and train the largest anchor (teacher) and transfer its knowledge to the sampled stitch (student) similar to inplace distillation [79]. In practice, we employ the hard-label distillation [26, 63]. + +Reuse stitch importance scores for deployment. After fine-tuning, with the set of accumulated importance scores $\mathcal{Q}_{\gamma}$ for all stitches, we directly use them to select the stitches with the highest scores to deploy, which eliminates the need for the costly evaluation of the final stage of the straightforward SN-Net adaptation solution. We empirically observe that the deployed stitches are mostly on the Pareto frontier in Section 5.3, suggesting our estimation of the important stitches is accurate. Furthermore, in the supplementary material, we show that the important stitches differ across different tasks, proving the need for task-specific deployment strategies, as highlighted in [19, 22, 75]. It's worth noting that although the SNIP metric has been widely employed as a one-shot metric to efficiently estimate network performance [1, 69], we extend it to estimate stitch importance for better training and selection. + +# 5. Experiments + +# 5.1. Visual Recognition + +We evaluate the effectiveness of our method on a total of 25 downstream visual recognition tasks, including fine-grained visual classification (FGVC) benchmark, common visual classification benchmark CIFAR-100 [34], and VTAB-1k [81] benchmark. FGVC benchmark contains NABirds [67], CUB-200-2011 [68], Stanford Cars [18], Stanford Dogs [32], and Oxford Flowers [50] tasks and the VTAB-1k benchmark includes 19 tasks in different domains, each of which has 800 training samples. For visual recognition experiments, we stitch ImageNet-21k pretrained ViT-Ti/S/B models [16] from [60]. + +Implementation details. We employ the stitching space and settings (kernel size as 2 and stride as 1) in [51] for downstream visual recognition tasks. We uniformly divide the resource constraints supported by the stitches into around 15 intervals and deploy one stitch within each interval. We set the hyper-parameter $r$ for the rank of weight updating in the self-attention layers to be 32, 16, and 8 for ViT-Ti, ViT-S, and ViT-B, respectively. We also set $r = d_1 // 4$ universally for all stitching layers and $\eta$ in Eq. (3) as 0.9 by grid search. We follow [31] to fine-tune 100 epochs on each task and select the other hyper-parameters and augmentation methods. We include more implementation details in + +![](images/e307fd95fed9c167bc222d554736d56ea0f914ca7e6f482f6bbdce418c4110b5.jpg) +Figure 3. Performance comparisons with SN-Net [51] for adapting ViT-Ti/S/B pre-trained on ImageNet-22k [15] to Stanford Cars [18], CUB-200-2011 [68], Stanford Dogs [32], and NABirds [67]. We denote individually fine-tuned anchors as yellow stars. We also show the number of trainable parameters. + +![](images/1af46de912baf31fd1ba0b6c8e60c6943b6add1837011fa259f6feca9af4f84d.jpg) + +![](images/2d1d320775be21db95f2bb6155c0f27ee7d00e8a46405e38fd095248054e2753.jpg) + +![](images/964957dde80b4e1b42c30c4ae4610a4a08c5518734d116b17c0b1eb272edf2da.jpg) + +![](images/68d6af674840a83cd087b0f24427380108f82ef6b28e3f28334f9c7e03424e0b.jpg) +Figure 4. Performance comparisons with SN-Net [51] for adapting ViT-Ti/S/B pre-trained on ImageNet-22k [15] to VTAB-1k [81] and CIFAR-100 [34]. We denote individually fine-tuned anchors as yellow stars and also show the number of trainable parameters. + +![](images/671fb773bb232eec83b70da8eb196089159085b58ba9bc8f8fbdeda9aa8451c4.jpg) + +Table 2. Relative response quality to Alpaca-LoRA 7B [62] on Vicuna Bench [10]. Our Stitched LLaMA successfully interpolates the answer quality between Alpaca-LoRA 7B and 13B. + +
Alpaca-LoRAStitched LLaMA
Parameter # (B)6.713.06.97.811.512.513.0
Quality (%)100123113114115120123
+ +the supplementary material. + +Main results. We compare the performance of our ESTA with SN-Net and the individual anchors fine-tuned with LoRA [28] on various datasets. The FLOPs-accuracy curves are visualized in Figures 3 and 4. We report the averaged results over the three task groups on VTAB-1k following [31]. Overall, the stitches obtained by our ESTA exhibit smooth FLOPs-accuracy trade-off curves consistently across all datasets. Notably, the curve for our ESTA outperforms SN-Net by significant margins with much fewer trainable parameters. We show that the inferior results of SN-Net on downstream tasks are led by the sub-optimal designs of fine-tuning all model parameters and a three-stage adaptation pipeline in Section 5.3. + +We also observe that the stitches even outperform the anchors with comparable or lower FLOPs in many cases. For instance, stitches outperform individually fine-tuned ViT-S/B anchors by clear margins on NABirds, CUB-200-2011, Stanford Cars, and Stanford Dogs datasets with comparable computational complexity. This phenomenon is consistent with the observations made in [52, 79]. We conjecture that weight sharing among the stitches serves as a strong regularization to improve their generalization. We show more results on few-shot learning and stitching Convnets [46] in the supplementary material. + +# 5.2. Instruction-following Task + +We also evaluate the instruction-following capability when stitching the LLaMA family [64] with our ESTA framework. We adapt and fine-tune LLaMA-7B/13B on the 52K instruction-following dataset from [62] to obtain chatbot stitches of varying sizes. The 52K instruction-following data is generated from 175 human-written instruction-output pairs with [71]. Each sample in the dataset is formulated by a task description, the context of the task, and the answer that is generated by GPT-3.5 (text-davinci-003) [7]. + +Implementation details. We employ the stitching space and settings (kernel size as 1 and stride as 1) in [51] for the instruction-following task and introduce 40 stitches. We uniformly divide the resource constraints supported by the stitches into 6 intervals and deploy one stitch within each interval. We set $\eta$ in Eq. (3) as 0.9. The rank $r$ is set to 16 and $d_1 // 4$ for LoRA modules in self-attention and stitching layers, respectively. We employ Adam optimizer and fine-tune with batch size 128 for 10 epochs with gradient accumulation. During the generation stage, we employ top- $p$ sampling to sample from the top subword candidates with the probability $p = 0.75$ , conduct the beam search, and set the beam size as 4. + +Results. We term the stitches obtained by our ESTA framework as Stitched LLaMA and compare them with Alpaca-LoRA 7B and 13B that respectively freeze the LLaMA 7B and 13B backbones and update the self-attention weights with LoRA modules [28]. For quantitative results reported in Table 2, we observe that similar to the visual recognition tasks, our Stitched LLaMA with different numbers of parameters generates answers with plausible quality in between Alpaca-LoRA 7B and 13B. Figure 5 gives qualitative comparisons. Our Stitched LLaMAs produce reasonable responses, suggesting the potential of our ESTA for adapting large language models (LLMs) for more downstream language tasks. More examples can be found in the supplementary material. + +# 5.3. Ablation Study + +Effect of parameter-efficient stitch fine-tuning. We investigate the effectiveness of our PST on five FGVC tasks and visualize the averaged results in Figure 6 (a) and (b). + +# Instruction: Tell me about alpacas. + +![](images/f04c44941f3c8d7da59b4f13ce2725b9edc715daee04f06880f6c899b845ccc5.jpg) + +Alpaca-LoRA 7B (0.035 s/token): Alpacas are members of the camelid family, which also includes llamas and camels. + +They are native to the Andes Mountains in South America, where they have been domesticated for thousands of years. Alpacas are known for their soft, luxurious fiber, which is used to make clothing, blankets, and other textiles. They are also raised for their meat, which is considered a delicacy in some parts of the world. Alpacas are gentle, intelligent animals, and they are prized for their docile nature and ability to thrive in a wide range of climates. + +![](images/6ba465d987693d004666b0c529a660dad07275b93b25b12df950b2502ffcce14.jpg) + +Stitched LLaMA 8B (0.037 s/token): Alpacas are herbivorous mammals that belong to the family Camelidae, which also includes camels and camelids. They are native to South American continent and are commonly found in Peru, Argentina, Bolivia, and Chile. They are smaller than their cousin, the camel, and are typically about 80-90 inches tall fully grown. They are known for their soft, shimmering, and light-weight fibers, which tend to make clothing, carpets, and other textile products. Alpacas are herbivores, and catets consists mainly of grasses, leaves, and straw. They are social animals and live in a few up to 100 alpacas, led by a dominant male. Alpacas are known for their calm and mature, and they are often used as livestock for their soft, shimmering fibers. + +![](images/a8086e6c950405e98b75fa0f6c7d5934260f4851d4b3c8e5cf2c1905e96d242b.jpg) + +Stitched LLaMA 11B (0.041 s/token): Alpacas are native to South America and are members of the camelid family. They are known for their soft, fluffy coats and their gentle nature. Alpacas are herbivores, meaning they feed on grass and other vegetation. They have a gestation period of approximately 11 days and usually give birth to a single baby, although twins and triplets are not uncommon. Alpacas are social animals and live in herds, led by a dominant they communicate with each other using a variety of vocalizations, including whistles, grunts, and barks. Alpacas are intelligent animals and are known for curiosity and playful nature. Alpacas have been domesticated for thousands of years and are used for a variety of purposes, including as pets, for their fiber, for their meat. Alpaca fiber is prized for its softness and warmth, making it a popular material for clothing and blankets. Alpaca meat is also considered a key in some parts of the world. Overall, alpacas are gentle, intelligent animals that have been domesticated for thousands of years. They are valued for their meat, and companionship. + +# Instruction: Can you help me write a resignation letter to my current employer, while leaving on good terms and expressing gratitude for the opportunities provided? + +![](images/411f36edf167bbf47494194987b87d052b2e34c5c29b7e7aa860be3dbd9ee5f9.jpg) + +Alpaca-LoRA 7B (0.035 s-token): Dear [Name] + +of Employer], I am writing to inform you that I am resigning from my position as [Job Title] with [Company Name]. It has been a pleasure working with you and I am grateful for the opportunities you have provided me. I wish you and the company all the best in the future. Sincerely, [Your Name] + +![](images/5931474e9cf30c70a33e7ce73682c1bf912f17e153c4880262b3e013fa17f1d2.jpg) + +Stitched LLaMA 8B (0.037 s/token): Dear [Employer], I am writing to inform you that I am resigning from my position at [Company Name] with a heavy heart. I have enjoyed my time with you and the team, and I am grateful for the opportunities that have been provided to me. I like to express my sincere appreciation for the training, support, and opportunities that have provided to me during my time at [Company Name]. I have learned a great deal and have grown really and professionally. Please accept my best wishes for the continued success of [Company and the team. I wish you all the best in the future. Sincerely, [Your Name] + +![](images/d4f12cfe8a5f3a5d8c794e044a81734dca23d4486d52cdb0b046f85ecc524e72.jpg) + +Sti 118 (0.041 s/ token): Dear [Employer], I am writing this letter to inform you of my decision to resign from my position at [Company]. It has been a great pleasure to work for [Company] and I have learned a lot during my time here. I would like to express my gratitude for the opportunities that I seen provided to me and for the support and guidance that I have received from my colleagues and supervisors. I have grown both professionally and finally during my time at [Company] and I am grateful for the experiences I have had. I am leaving on good terms and would be happy to provide a positive chance for [Company] in the future. If there is anything I can do to assist with the transition, please do not hesitate to reach out. Thank you for the opportunity to offer [Company] and I wish you all the best in your future endeavors. Sincerely, [Your Name] + +![](images/1783f314e4bf19760d94fd88dbe429f7b73f0bf028fd5c5f79aeaf4afe4b574b.jpg) +Figure 5. Instruction-following comparison between Stitched LLaMA obtained by our ESTA and the Alpaca-LoRA 7B fine-tuned with LoRA [28]. + +![](images/8740576bf0900c0f058d9529d2da21aa3f2b5b820d1e3c70ad40df89d84429ba.jpg) +Figure 6. Ablation studies with results averaged over five FGVC tasks. (a) Effect of our stitch-agnostic LoRA modules. "Full FT" represents full fine-tuning of all the model parameters. "PST w/ full SL" employs fully fine-tuned stitching layers. We also show the number of trainable parameters. (b) Effect of our stitch-specific bias terms. "SSB" and "scaled r" represent stitch-specific bias terms and scaling the rank hyperparameter to reach 4.6M trainable parameters, respectively. (c) Effect of our task-specific stitch sampling. "Uniform, all" and "Uniform, best" represent all the stitches fine-tuned with uniform sampling and their best ones in each resource constraint interval on the Pareto frontier, respectively. (d) Effect of our strategy to simultaneously adapt and stitch anchors. Our strategy "Adapt-and-stitch" takes 100 epochs for fine-tuning. In contrast, "Adapt-then-stitch", as a straightforward approach to apply SN-Net, first individually adapts each anchor for a total of 300 epochs, and then fine-tunes SN-Net for another 100 epochs. + +![](images/498598ad00fe8f261880fe20e83c7953d2cb06ed289b570d59ad9748e964b2bd.jpg) + +![](images/ce9fa356fb864e0fedd2391c575e610e23e69370aad23b33fd2095ad337a34c4.jpg) + +To recall, our PST keeps the majority of the model parameters frozen while introducing trainable stitch-agonistic low-rank decomposition matrices and stitch-specific bias terms. In Figure 6 (a), we observe that when employing LoRA to the multi-head self-attention layers, there are solid performance gains on all FLOPs ranges with significantly fewer trainable parameters, echoing the observations in [31, 41]. + +It is suggested that low-rank weight updates is a powerful alternative to full fine-tuning for stitching and adapting anchors and we speculate that low-rank weight update avoids overfitting under limited downstream data. We further employ low-rank weight updates to the stitching layers and save 2.4M more trainable parameters with a slight overall performance drop. We conjecture that the stitching layers + +for large pre-trained models are also in low intrinsic dimensions [2] and least-squares initialization [51] already provide a good initialization for them. + +Figure 6 (b) shows that employing stitch-specific bias terms achieves better performance with affordable 1.4M extra trainable parameters. We also include the baseline that simply scales the low-rank dimension $r$ to reach the same number of trainable parameters as our PST. This baseline is inferior to our PST. We speculate that the stitch-specific bias terms enable flexible adjustments of feature representations for different stitches, leading to enhanced performance. + +Effect of task-specific stitch sampling. We investigate the effectiveness of our sampling strategy (introduced in Section 4.2) on the five FGVC tasks. The averaged results are visualized in Figure 6 (c). We compare our task-specific stitch sampling with the uniform one that is adopted in [51]. We empirically find that for uniform sampling, there is a clear performance gap between the best stitches that will be deployed and the others that will be dropped after training, suggesting the need for concentrating on the important stitches during training. To this end, our task-specific sampling strategy performs better than uniform sampling, indicating that our sampling strategy selects accurate task-specific important stitches. Most importantly, our strategy circumvents the need for the costly evaluation stage in the direct solution of adapting SN-Net for downstream tasks. + +Effect of simultaneously adapting and stitching anchors. We study the effect of simultaneously adapting and stitching anchors strategy (introduced in Section 4.2) on five FGVC tasks. The averaged results are visualized in Figure 6 (d). We compare our approach with the direct SN-Net adaptation that first adapts each anchor individually for a total of 300 epochs before stitching them under the exact setting as ESTA. It can be seen that the direct SN-Net adaptation yields poor results in task adaptation. Fine-tuned individual anchors, especially the larger ones, are likely to overfit the limited downstream data [44, 53]. Accordingly, we speculate that the adapted anchor weights serve as a sub-optimal initialization for the stitches in the context of model stitching. On the other hand, our approach achieves better performance and greatly saves the fine-tuning cost. + +Effect of the PEFT technique choices. We compare the effect of different PEFT technique choices and visualize the averaged results on the five FGVC tasks in Figure 7 left. We experiment to employ other PEFT techniques Adapter [27] and Adaptformer [9]. We follow their default setting to respectively insert bottleneck structures sequentially and in parallel to the feed-forward layers and fix the number of trainable parameters the same as ours for a fair comparison. We observe that the overall performance for different PEFT techniques is similar, suggesting that our ESTA framework is compatible with more PEFT technique choices. Since low-rank weight updates of LoRA [28] can be merged into + +![](images/eb0215283bf4cc3a8215ec628dc353308860aebd9127fbbc2c15e565555c2864.jpg) +Figure 7. Left: effect of different PEFT technique choices. Right: effect of the rank hyper-parameter $r$ in LoRA modules of self-attention layers, for which the three values represent the ranks in ViT-Ti, ViT-S, and ViT-B, respectively. The results are averaged on five FGVC tasks. + +![](images/f73035690bfa05c8498872e9d57e73f9fe6c8aa13500c1b2ab00781a7f797b43.jpg) + +the backbone after training without extra inference computational cost, we employ LoRA by default. + +Effect of the low-rank hyper-parameter $r$ . We investigate the effect of different $r$ which controls the rank when updating weights in self-attention layers on the five FGVC tasks. The averaged results on the five FGVC tasks are visualized in Figure 7 right. We find that setting $r$ to be different values barely has any effect on the overall performance, suggesting our ESTA framework is robust to the rank hyper-parameter $r$ . Since setting $r$ to be 32, 16, and 8 for low-rank updating the self-attention weights in ViT-Ti, ViT-S, and ViT-B has slightly higher averaged performance among the stitches, we make it our default setting. + +# 6. Conclusion and Future Work + +In this paper, we have introduced a novel task adaptation framework to cheaply obtain a palette of fine-tuned networks via model stitching, supporting diverse efficiency-performance tradeoffs at runtime. Specifically, built on SN-Net, we have tailored a parameter-efficient stitch finetuning method, which learns lightweight stitch-agnostic LoRA modules and stitch-specific bias terms while keeping the majority of the parameters frozen. Our design significantly reduces fine-tuning memory and storage costs for downstream task adaptation. Moreover, we have devised a task-specific stitch sampling strategy to assign higher sampling probability to the important stitches during training, which simultaneously improves the Pareto frontiers and avoids a costly evaluation stage. Extensive experiments on 25 downstream visual recognition tasks and the instruction-following task have demonstrated the effectiveness of our proposed framework. + +Limitations. Due to the constraints of computational resources, our experiments are limited to visual recognition and instruction-tuning tasks. In the future, we will explore adapting pre-trained model families to dense prediction and multi-modal tasks. + +# References + +[1] Mohamed S. Abdelfattah, Abhinav Mehrotra, Lukasz Dudziak, and Nicholas D. Lane. Zero-cost proxies for lightweight nas. In ICLR, 2021. 5 +[2] Armen Aghajanyan, Luke Zettlemoyer, and Sonal Gupta. Intrinsic dimensionality explains the effectiveness of language model fine-tuning. In ACL, 2021. 4, 8 +[3] Haitham Bou Ammar, Eric Eaton, Matthew E Taylor, Decebal Constantin Mocanu, Kurt Driessens, Gerhard Weiss, and Karl Tuyls. An automated measure of mdp similarity for transfer in reinforcement learning. In AAAIw, 2014. 3 +[4] Yamini Bansal, Preetum Nakkiran, and Boaz Barak. Revisiting model stitching to compare neural representations. NeurIPS, 34:225-236, 2021. 1, 2 +[5] Hangbo Bao, Li Dong, Songhao Piao, and Furu Wei. Beit: Bert pre-training of image transformers. In ICLR, 2022. 1 +[6] Daniel Bolya, Rohit Mittapalli, and Judy Hoffman. Scalable diverse model selection for accessible transfer learning. NeurIPS, 34:19301-19312, 2021. 3 +[7] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. *NeurIPS*, 33:1877–1901, 2020. 6 +[8] Han Cai, Chuang Gan, Tianzhe Wang, Zhekai Zhang, and Song Han. Once-for-all: Train one network and specialize it for efficient deployment. In ICLR, 2020. 1 +[9] Shoufa Chen, Chongjian Ge, Zhan Tong, Jiangliu Wang, Yibing Song, Jue Wang, and Ping Luo. Adaptformer: Adapting vision transformers for scalable visual recognition. NeurIPS, 2022. 8 +[10] Wei-Lin Chiang, Zhuohan Li, Zi Lin, Ying Sheng, Zhang-hao Wu, Hao Zhang, Lianmin Zheng, Siyuan Zhuang, Yong-hao Zhuang, Joseph E. Gonzalez, Ion Stoica, and Eric P. Xing. Vicuna: An open-source chatbot impressing gpt-4 with $90\%$ * chatgpt quality, 2023. 6 +[11] Adam Coates and Andrew Y Ng. The importance of encoding versus training with sparse coding and vector quantization. In ICML, pages 921-928, 2011. 1 +[12] Adrián Csiszárik, Péter Körösí-Szabó, Akos Matszangosz, Gergely Papp, and Daniel Varga. Similarity and matching of neural network representations. NeurIPS, 34:5656-5668, 2021. 1, 2, 4 +[13] Dong Dai and Tong Zhang. Greedy model averaging. NeurIPS, 24, 2011. 3 +[14] Xiaoliang Dai, Alvin Wan, Peizhao Zhang, Bichen Wu, Zijian He, Zhen Wei, Kan Chen, Yuandong Tian, Matthew Yu, Peter Vajda, et al. Fbnetv3: Joint architecture-recipe search using predictor pretraining. In CVPR, pages 16276–16285, 2021. 2 +[15] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR, pages 248-255, 2009. 6 +[16] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is + +worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2021. 2, 5 +[17] Eric Eaton, Marie Desjardins, and Terran Lane. Modeling transfer relationships between learning tasks for improved inductive transfer. In ECML PKDD, pages 317-332. Springer, 2008. 3 +[18] Timnit Gebru, Jonathan Krause, Yilun Wang, Duyun Chen, Jia Deng, and Li Fei-Fei. Fine-grained car detection for visual census estimation. In AAAI, 2017. 4, 5, 6 +[19] Yunhui Guo, Honghui Shi, Abhishek Kumar, Kristen Grauman, Tajana Rosing, and Rogerio Feris. Spottune: transfer learning through adaptive fine-tuning. In CVPR, pages 4805-4814, 2019. 5 +[20] Zichao Guo, Xiangyu Zhang, Haoyuan Mu, Wen Heng, Zechun Liu, Yichen Wei, and Jian Sun. Single path one-shot neural architecture search with uniform sampling. In ECCV, pages 544-560. Springer, 2020. 1 +[21] Song Han, Jeff Pool, John Tran, and William Dally. Learning both weights and connections for efficient neural network. NeurIPS, 28, 2015. 1 +[22] Haoyu He, Jianfei Cai, Jing Zhang, Dacheng Tao, and Bohan Zhuang. Sensitivity-aware visual parameter-efficient tuning. In ICCV, 2023. 5 +[23] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollar, and Ross Girshick. Masked autoencoders are scalable vision learners. In CVPR, pages 16000-16009, 2022. 1 +[24] Rudolf Herdt, Maximilian Schmidt, Daniel Otero Baguer, Jean Le'Clerc Arrastia, and Peter Maass. Model stitching and visualization how gan generators can invert networks in real-time. arXiv preprint arXiv:2302.02181, 2023. 2 +[25] Adriano Hernandez, Rumen Dangovski, Peter Y. Lu, and Marin Soljacic. Model stitching: Looking for functional similarity between representations. NeurIPSW, 2022. 2 +[26] Geoffrey Hinton, Oriol Vinyals, and Jeff Dean. Distilling the knowledge in a neural network. NeurIPSW, 2014. 5 +[27] Neil Houlsby, Andrei Giurgiu, Stanislaw Jastrzebski, Bruna Morrone, Quentin De Laroussilhe, Andrea Gesmundo, Mona Attariyan, and Sylvain Gelly. Parameter-efficient transfer learning for nlp. In ICML, pages 2790-2799, 2019. 1, 3, 8 +[28] Edward J Hu, yelong shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. LoRA: Low-rank adaptation of large language models. In ICLR, 2022. 1, 2, 3, 4, 6, 7, 8 +[29] Itay Hubara, Matthieu Courbariaux, Daniel Soudry, Ran El-Yaniv, and Yoshua Bengio. Quantized neural networks: Training neural networks with low precision weights and activations. JMLR, 18(1):6869-6898, 2017. 1 +[30] Pavel Izmailov, Dmitrii Podoprikhin, Timur Garipov, Dmitry Vetrov, and Andrew Gordon Wilson. Averaging weights leads to wider optima and better generalization. In UAI, 2018. 3 +[31] Menglin Jia, Luming Tang, Bor-Chun Chen, Claire Cardie, Serge Belongie, Bharath Hariharan, and Ser-Nam Lim. Visual prompt tuning. In ECCV, 2022. 1, 3, 5, 6, 7 +[32] Aditya Khosla, Nityananda Jayadevaprakash, Bangpeng Yao, and Li Fei-Fei. Novel dataset for fine-grained image categorization. In CVPRW, 2011. 5, 6 + +[33] Simon Kornblith, Jonathon Shlens, and Quoc V Le. Do better imagenet models transfer better? In CVPR, pages 2661-2671, 2019. 3 +[34] Alex Krizhevsky. Learning multiple layers of features from tiny images. Technical report, 2009. 5, 6 +[35] Achintya Kundu, Laura Wynter, Rhui Dih Lee, and Luis Angel Bathen. Tofa: Transfer-once-for-all. arXiv preprint arXiv:2303.15485, 2023. 2 +[36] Yann LeCun, John Denker, and Sara Solla. Optimal brain damage. NeurIPS, 2, 1989. 1 +[37] Namhoon Lee, Thalaiyasingam Ajanthan, and Philip Torr. Snip: Single-shot network pruning based on connection sensitivity. In ICLR, 2019. 2, 5 +[38] Karel Lenc and Andrea Vedaldi. Understanding image representations by measuring their equivariance and equivalence. In CVPR, pages 991-999, 2015. 1, 2 +[39] Brian Lester, Rami Al-Rfou, and Noah Constant. The power of scale for parameter-efficient prompt tuning. In EMNLP, 2021. 3 +[40] Xiang Lisa Li and Percy Liang. Prefix-tuning: Optimizing continuous prompts for generation. In ACL, 2021. 3 +[41] Dongze Lian, Daquan Zhou, Jiashi Feng, and Xinchao Wang. Scaling & shifting your features: A new baseline for efficient model tuning. NeurIPS, 2022. 3, 7 +[42] Iou-Jen Liu, Jian Peng, and Alexander G Schwing. Knowledge flow: Improve upon your teachers. In ICLR, 2019. 3 +[43] Jing Liu, Jianfei Cai, and Bohan Zhuang. Focusformer: Focusing on what we need via architecture sampler. arXiv preprint arXiv:2208.10861, 2022.5 +[44] Pengfei Liu, Weizhe Yuan, Jinlan Fu, Zhengbao Jiang, Hiroaki Hayashi, and Graham Neubig. Pre-train, prompt, and predict: A systematic survey of prompting methods in natural language processing. ACM Comput. Surv., 2023. 8 +[45] Yen-Cheng Liu, Chih-Yao Ma, Junjiao Tian, Zijian He, and Zsolt Kira. Polyhistor: Parameter-efficient multi-task adaptation for dense vision tasks. NeurIPS, 2022. 3 +[46] Zhuang Liu, Hanzi Mao, Chao-Yuan Wu, Christoph Feichtenhofer, Trevor Darrell, and Saining Xie. A convnet for the 2020s. In CVPR, pages 11976-11986, 2022. 6 +[47] Zhichao Lu, Gautam Sreekumar, Erik Goodman, Wolfgang Banzhaf, Kalyanmoy Deb, and Vishnu Naresh Boddeti. Neural architecture transfer. TPAMI, 43(9):2971-2989, 2021. 2 +[48] Michael S Matena and Colin A Raffel. Merging models with fisher-weighted averaging. NeurIPS, 35:17703-17716, 2022. 3 +[49] Cuong Nguyen, Tal Hassner, Matthias Seeger, and Cedric Archambeau. Leep: A new measure to evaluate transferability of learned representations. In ICML, pages 7294-7305. PMLR, 2020. 3 +[50] Maria-Elena Nilsback and Andrew Zisserman. Automated flower classification over a large number of classes. In ICVGIP, pages 722-729. IEEE, 2008. 5 +[51] Zizheng Pan, Jianfei Cai, and Bohan Zhuang. Stretchable neural networks. In CVPR, 2023. 1, 2, 3, 4, 5, 6, 8 +[52] Hieu Pham, Melody Guan, Barret Zoph, Quoc Le, and Jeff Dean. Efficient neural architecture search via parameters sharing. In ICML, pages 4095-4104. PMLR, 2018. 6 + +[53] Jo Pleted and Tom Gedeon. Deep transfer learning for image classification: a survey. arXiv preprint arXiv:2205.09904, 2022. 8 +[54] Ilija Radosavovic, Raj Prateek Kosaraju, Ross Girshick, Kaiming He, and Piotr Dólar. Designing network design spaces. In CVPR, pages 10428-10436, 2020. 5 +[55] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, pages 10684-10695, 2022. 3 +[56] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, et al. Imagenet large scale visual recognition challenge. IJCV, 115:211-252, 2015. 1, 4 +[57] Andrei A Rusu, Neil C Rabinowitz, Guillaume Desjardins, Hubert Soyer, James Kirkpatrick, Koray Kavukcuoglu, Razvan Pascanu, and Raia Hadsell. Progressive neural networks. arXiv preprint arXiv:1606.04671, 2016. 3 +[58] Yang Shu, Zhi Kou, Zhangjie Cao, Jianmin Wang, and Mingsheng Long. Zoo-tuning: Adaptive transfer from a zoo of models. In ICML, pages 9626-9637. PMLR, 2021. 3 +[59] Yang Shu, Zhangjie Cao, Ziyang Zhang, Jianmin Wang, and Mingsheng Long. Hub-pathway: Transfer learning from a hub of pre-trained models. NeurIPS, 2022. 3 +[60] Andreas Steiner, Alexander Kolesnikov, Xiaohua Zhai, Ross Wightman, Jakob Uszkoreit, and Lucas Beyer. How to train your vit? data, augmentation, and regularization in vision transformers. TMLR, 2022. 5 +[61] Yi-Lin Sung, Jaemin Cho, and Mohit Bansal. Lst: Ladder side-tuning for parameter and memory efficient transfer learning. NeurIPS, 2022. 3 +[62] Rohan Taori, Ishaan Gulrajani, Tianyi Zhang, Yann Dubois, Xuechen Li, Carlos Guestrin, Percy Liang, and Tatsunori B. Hashimoto. Stanford alpaca: An instruction-following llama model. https://github.com/tatsu-lab/stanford_alpaca, 2023. 6 +[63] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Hervé Jégou. Training data-efficient image transformers & distillation through attention. In ICML, pages 10347-10357. PMLR, 2021. 5 +[64] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 1, 2, 6 +[65] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288, 2023. 2 +[66] Anh T Tran, Cuong V Nguyen, and Tal Hassner. Transferability and hardness of supervised classification tasks. In ICCV, pages 1395-1405, 2019. 3 +[67] Grant Van Horn, Steve Branson, Ryan Farrell, Scott Haber, Jessie Barry, Panos Ipeirotis, Pietro Perona, and Serge Belongie. Building a bird recognition app and large scale + +dataset with citizen scientists: The fine print in fine-grained dataset collection. In CVPR, pages 595-604, 2015. 5, 6 +[68] Catherine Wah, Steve Branson, Peter Welinder, Pietro Perona, and Serge Belongie. The caltech-ucsd birds-200-2011 dataset. Tech. Rep. CNS-TR-2011-001, California Institute of Technology, 2011. 5, 6 +[69] Chaoqi Wang, Guodong Zhang, and Roger Grosse. Picking winning tickets before training by preserving gradient flow. In ICLR, 2020. 5 +[70] Dilin Wang, Meng Li, Chengyue Gong, and Vikas Chandra. Attentivenas: Improving neural architecture search via attentive sampling. In CVPR, pages 6418-6427, 2021. 5 +[71] Yizhong Wang, Yeganeh Kordi, Swaroop Mishra, Alisa Liu, Noah A Smith, Daniel Khashabi, and Hannaneh Hajishirzi. Self-instruct: Aligning language model with self generated instructions. arXiv preprint arXiv:2212.10560, 2022. 6 +[72] Thomas Wolf, Lysandre Debut, Victor Sanh, Julien Chau-mond, Clement Delangue, Anthony Moi, Pierrick Cistac, Tim Rault, Rémi Louf, Morgan Funtowicz, et al. Transformers: State-of-the-art natural language processing. In EMNLP, pages 38-45, 2020. 1 +[73] Mitchell Wortsman, Gabriel Ilharco, Samir Ya Gadre, Rebecca Roelofs, Raphael Gontijo-Lopes, Ari S Morcos, Hongseok Namkoong, Ali Farhadi, Yair Carmon, Simon Kornblith, et al. Model soups: averaging weights of multiple fine-tuned models improves accuracy without increasing inference time. In ICML, pages 23965-23998. PMLR, 2022. 3 +[74] Hang Xu, Ning Kang, Gengwei Zhang, Chuanlong Xie, Xiaodan Liang, and Zhenguo Li. Nasoa: Towards faster task-oriented online fine-tuning with a zoo of models. In ICCV, pages 5097-5106, 2021. 3 +[75] Runxin Xu, Fuli Luo, Zhiyuan Zhang, Chuanqi Tan, Baobao Chang, Songfang Huang, and Fei Huang. Raise a child in large language model: Towards effective and generalizable fine-tuning. In EMNLP, 2021. 5 +[76] Xingyi Yang, Daquan Zhou, Songhua Liu, Jingwen Ye, and Xinchao Wang. Deep model reassembly. NeurIPS, 35: 25739-25753, 2022. 2 +[77] Jason Yosinski, Jeff Clune, Yoshua Bengio, and Hod Lipson. How transferable are features in deep neural networks? NeurIPS, 27, 2014. 3 +[78] Kaichao You, Yong Liu, Jianmin Wang, and Mingsheng Long. Logme: Practical assessment of pre-trained models for transfer learning. In ICML, pages 12133-12143. PMLR, 2021. 3 +[79] Jiahui Yu and Thomas S Huang. Universally slimmable networks and improved training techniques. In ICCV, pages 1803-1811, 2019. 5, 6 +[80] Elad Ben Zaken, Yoav Goldberg, and Shauli Ravfogel. Bitfit: Simple parameter-efficient fine-tuning for transformer-based masked language-models. In ACL, pages 1-9, 2022. 3 +[81] Xiaohua Zhai, Joan Puigcerver, Alexander Kolesnikov, Pierre Ruyssen, Carlos Riquelme, Mario Lucic, Josip Djolonga, Andre Susano Pinto, Maxim Neumann, Alexey Dosovitskiy, et al. A large-scale study of representation learning with the visual task adaptation benchmark. arXiv preprint arXiv:1910.04867, 2019. 5, 6 + +[82] Lvmin Zhang and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In ICCV, 2023. 3 +[83] Renrui Zhang, Jiaming Han, Aojun Zhou, Xiangfei Hu, Shilin Yan, Pan Lu, Hongsheng Li, Peng Gao, and Yu Qiao. Llama-adapter: Efficient fine-tuning of language models with zero-init attention. arXiv preprint arXiv:2303.16199, 2023. 3 \ No newline at end of file diff --git a/2024/Efficient Stitchable Task Adaptation/images.zip b/2024/Efficient Stitchable Task Adaptation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..05c86eb46445a0ca5bf9ae4c7f036aab907114a1 --- /dev/null +++ b/2024/Efficient Stitchable Task Adaptation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6bd828d206f91a7a9676f46cdba9507bcb7a354dde8940dcb59038127d099531 +size 286191 diff --git a/2024/Efficient Stitchable Task Adaptation/layout.json b/2024/Efficient Stitchable Task Adaptation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..6ec7e98e1935b56e5518515d85c61281c0170e18 --- /dev/null +++ b/2024/Efficient Stitchable Task Adaptation/layout.json @@ -0,0 +1,10102 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 184, + 103, + 410, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 184, + 103, + 410, + 121 + ], + "spans": [ + { + "bbox": [ + 184, + 103, + 410, + 121 + ], + "type": "text", + "content": "Efficient Stitchable Task Adaptation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 143, + 467, + 159 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 143, + 467, + 159 + ], + "spans": [ + { + "bbox": [ + 130, + 143, + 467, + 159 + ], + "type": "text", + "content": "Haoyu He Zizheng Pan Jing Liu Jianfei Cai Bohan Zhuang" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 200, + 163, + 390, + 177 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 200, + 163, + 390, + 177 + ], + "spans": [ + { + "bbox": [ + 200, + 163, + 390, + 177 + ], + "type": "text", + "content": "ZIP Lab, Monash University, Australia" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 143, + 204, + 192, + 217 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 204, + 192, + 217 + ], + "spans": [ + { + "bbox": [ + 143, + 204, + 192, + 217 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 229, + 289, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 229, + 289, + 613 + ], + "spans": [ + { + "bbox": [ + 46, + 229, + 289, + 613 + ], + "type": "text", + "content": "The paradigm of pre-training and fine-tuning has laid the foundation for deploying deep learning models. However, most fine-tuning methods are designed to meet a specific resource budget. Recently, considering diverse deployment scenarios with various resource budgets, SN-Net [51] is introduced to quickly obtain numerous new networks (stitches) from the pre-trained models (anchors) in a model family via model stitching. Although promising, SN-Net confronts new challenges when adapting it to new target domains, including huge memory and storage requirements and a long and sub-optimal multistage adaptation process. In this work, we present a novel framework, Efficient Stitchable Task Adaptation (ESTA), to efficiently produce a palette of fine-tuned models that adhere to diverse resource constraints. Specifically, we first tailor parameter-efficient fine-tuning to share low-rank updates among the stitches while maintaining independent bias terms. In this way, we largely reduce fine-tuning memory burdens and mitigate the interference among stitches that arises in task adaptation. Furthermore, we streamline a simple yet effective one-stage deployment pipeline, which estimates the important stitches to deploy with training-time gradient statistics. By assigning higher sampling probabilities to important stitches, we also get a boosted Pareto frontier. Extensive experiments on 25 downstream visual recognition tasks demonstrate that our ESTA is capable of generating stitches with smooth accuracy-efficiency trade-offs and surpasses the direct SN-Net adaptation by remarkable margins with significantly lower training time and fewer trainable parameters. Furthermore, we demonstrate the flexibility and scalability of our ESTA framework by stitching LLMs from LLaMA family, obtaining chatbot stitches of assorted sizes1." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 634, + 128, + 646 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 634, + 128, + 646 + ], + "spans": [ + { + "bbox": [ + 47, + 634, + 128, + 646 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 654, + 287, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 654, + 287, + 677 + ], + "spans": [ + { + "bbox": [ + 46, + 654, + 287, + 677 + ], + "type": "text", + "content": "The paradigm of pre-training and fine-tuning has underpinned modern applications in both vision and lan" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 205, + 547, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 205, + 547, + 373 + ], + "spans": [ + { + "bbox": [ + 306, + 205, + 547, + 373 + ], + "type": "text", + "content": "guage. With off-the-shelf models pre-trained on large-scale datasets, the de-facto choice is vanilla full fine-tuning, which tunes all the model parameters with the downstream data. To reduce memory footprint and avoid overfitting, an emerging trend is to tune a small proportion of the model parameters while freezing the majority ones with Parameter-Efficient Fine-Tuning (PEFT) [27, 28, 31]. However, both full fine-tuning and PEFT target an exclusive specific resource budget for each downstream task, while in reality, we often need to deploy multiple models simultaneously to meet various resource demands. This makes us ponder: What is an effective and efficient way to obtain a palette of fine-tuned models meeting different resource constraints?" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 376, + 547, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 376, + 547, + 496 + ], + "spans": [ + { + "bbox": [ + 304, + 376, + 547, + 496 + ], + "type": "text", + "content": "A natural approach is to compress a well-trained large model into numerous smaller ones [11, 21, 29, 36]. However, the computational cost grows linearly with the number of deployment scenarios. Following the once-for-all network [8, 20], a few works first pre-train a weight-sharing over-parameterized supernet, then adapt the supernet to the downstream tasks. Although promising, training supernets at scale with large datasets requires prohibitive computational resources, e.g., thousands of GPU training hours, which is practically infeasible." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 498, + 548, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 498, + 548, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 498, + 548, + 714 + ], + "type": "text", + "content": "On the other hand, there are many open-source large models [5, 23, 64] from communities such as Hugging-Face [72] that are pre-trained on large-scale datasets and ready to be downloaded. Inspired by the success of model stitching [4, 12, 38], recently, Stitchable Neural Networks (SN-Net) [51] has been proposed to stitch pre-trained models (anchors) of the same family to quickly obtain a set of candidate new networks (stitches) with different accuracy-efficiency trade-offs of a wide FLOPs range. However, the scope of SN-Net is limited to the pre-training classification task on the same source domain [56]. When directly adapting the standard approach of SN-Net to a target domain, it faces new challenges. Specifically, the straightforward way to employ SN-Net in task adaptation typically requires three stages: first adapting anchors individually to the target domain, followed by stitching the adapted anchors, and finally evaluating all stitches to find and deploy the ones on the Pareto Frontier. Such a three-stage adaptation pro" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 683, + 258, + 693 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 683, + 258, + 693 + ], + "spans": [ + { + "bbox": [ + 58, + 683, + 258, + 693 + ], + "type": "text", + "content": "†Corresponding author. E-mail: bohan.zhuang@gmail.com" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 49, + 693, + 287, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 693, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 49, + 693, + 287, + 713 + ], + "type": "text", + "content": "1Source code will be released at https://github.com/ziplab/Stitched_LLaMA." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "28555" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 50, + 152, + 545, + 214 + ], + "blocks": [ + { + "bbox": [ + 46, + 70, + 547, + 148 + ], + "lines": [ + { + "bbox": [ + 46, + 70, + 547, + 148 + ], + "spans": [ + { + "bbox": [ + 46, + 70, + 547, + 148 + ], + "type": "text", + "content": "Table 1. Comparison of frameworks for obtaining a palette of networks to meet different hardware efficiency constraints. Supernet-based methods NAT [47] and TOFA [35] necessitate a supernet training stage plus long deployment GPU hours and are restricted to small FLOPs ranges. SN-Net [51] employs model stitching to achieve a wide FLOPs range with significantly reduced deployment GPU hours. Our ESTA framework further reduces the deployment GPU hours, fine-tuning memory, and trainable parameters, and is scalable to produce LLM stitches with billions of parameters (Section 5.2). The deployment GPU hours and fine-tuning memory for both SN-Net and ESTA are measured on a single NVIDIA GeForce RTX 3090 GPU when stitching ViT-Ti/S/B anchors [16] with batch size 64. * indicates data is inferred from the training recipe of TOFA's supernet architecture FBNetV3 [14]." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 50, + 152, + 545, + 214 + ], + "lines": [ + { + "bbox": [ + 50, + 152, + 545, + 214 + ], + "spans": [ + { + "bbox": [ + 50, + 152, + 545, + 214 + ], + "type": "table", + "html": "
MethodAdapting supernetDeployment GPU hoursFine-tuning memoryTrainable parameter #FLOPs range (G)LLM friendly
NAT [47]X>1,000->10M[0.2, 0.6]X
TOFA [35]✓>1,000*->10M*[0.2, 2.5]X
SN-Net [51]X19.313,235M124.2M[1.3, 17.6]X
ESTA (ours)X5.09,685M4.6M[1.3, 17.6]✓
", + "image_path": "f7a9d9a85b8f7e89ffaed77ac90fa81aa88e8ccc66740e8c713227c8b4e7efd5.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 219, + 289, + 352 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 219, + 289, + 352 + ], + "spans": [ + { + "bbox": [ + 46, + 219, + 289, + 352 + ], + "type": "text", + "content": "cess is expensive and sub-optimal. Moreover, adapting SN-Net typically needs to load and optimize all parameters of multiple models, leading to a daunting training-time memory cost. For instance, stitching the two smallest models in the LLaMA family [64, 64] with SN-Net under standard training settings (32-bit floating point parameters and Adam optimizer on a single GPU) requires more than 80G Video RAM that cannot fit into high-end GPUs such as NVIDIA A100. SN-Net also requires saving a separate instance of the model family for each task, making the storage grow linearly with the number of deployed applications." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 353, + 289, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 353, + 289, + 628 + ], + "spans": [ + { + "bbox": [ + 46, + 353, + 289, + 628 + ], + "type": "text", + "content": "Therefore, in this paper, we introduce a novel Efficient Stitchable Task Adaptation (ESTA) framework. Specifically, built upon SN-Net [51], ESTA introduces two main designs to overcome the aforementioned issues. First, to reduce the memory and storage costs, we tailor a Parameter-efficient Stitch fine-Tuning (PST) method, which incorporates the representative sparse fine-tuning technique LoRA [28] to keep the anchors and stitching layer weights frozen, while approximating their updates with trainable low-rank decomposition matrices. In addition to stitch-agnostic LoRA modules, we further introduce stitch-specific bias terms to alleviate the conflict among different stitches. Second, we propose a simple one-stage deployment pipeline to simultaneously adapt and stitch the pretrained anchors to the target domain. Moreover, we propose to estimate and accumulate an importance score for each stitch via the saliency pruning criterion [37]. With the importance scores, we develop a novel stitch-sampling method to assign important stitches with higher sampling probabilities. More importantly, after fine-tuning, the stitch importance scores can be directly used to infer the best ones at the Pareto frontier for deployment. Table 1 compares different efficient frameworks for diverse deployments." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 630, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 630, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 630, + 288, + 715 + ], + "type": "text", + "content": "Overall, our paper has the following major contributions. 1) This is a pioneering work to investigate the problem of efficient fine-tuning of SN-Net for task adaptation. Our solution ESTA is a successful endeavor that delivers dozens of ready-to-deploy vision models up to 17.6G FLOPs in 5 GPU hours as well as multiple chatbot models with billions of parameters by stitching LLaMA mod" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 219, + 547, + 388 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 219, + 547, + 388 + ], + "spans": [ + { + "bbox": [ + 304, + 219, + 547, + 388 + ], + "type": "text", + "content": "els [64, 65] for the instruction-following task. 2) We devise a parameter-efficient stitch fine-tuning method that incorporates trainable stitch-agnostic LoRA modules and stitch-specific bias terms, which largely reduces the fine-tuning memory footprints while alleviating the interference issue among stitches. Moreover, we streamline a simple one-stage deployment pipeline with a novel task-specific stitch sampling strategy that greatly reduces the deployment time and improves Pareto frontiers. 3) Extensive experiments conducted on 25 downstream visual recognition tasks demonstrate that when stitching ViT-Ti/S/B [16], our ESTA achieves remarkable performance improvements compared to the direct SN-Net adaptation counterpart while utilizing significantly fewer trainable parameters." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 403, + 392, + 415 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 403, + 392, + 415 + ], + "spans": [ + { + "bbox": [ + 306, + 403, + 392, + 415 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 424, + 547, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 424, + 547, + 700 + ], + "spans": [ + { + "bbox": [ + 304, + 424, + 547, + 700 + ], + "type": "text", + "content": "Model stitching. Model stitching [4, 38] targets connecting the bottom layers of one network to the top layers of another with a stitching layer. Following [38], model stitching is first employed to measure the similarities for the inner representations learned by deep neural networks [4, 12, 25]. Previous works [4, 12] observe that trained networks with different initializations can be effectively stitched without incurring significant performance drop. Based on this observation, DeRy [76] stitches the blocks dissected from different pre-trained models and assembles a new model based on their quantified similarity for better performance. Herdt et al. [24] propose interpreting the inner representations of a deep network by stitching it with a pre-trained GAN generator. Recently, from the perspective of flexible deployment, SN-Net [51] stitches models of different sizes within a model family to cheaply produce numerous new networks with diverse accuracy and efficiency trade-offs in the pretraining task. In contrast to [51], our work specifically targets obtaining a palette of fine-tuned models for the task adaptation setting. To do so, we make key redesigns including a parameter-efficient stitch fine-tuning method and a simple one-stage deployment pipeline to overcome the efficiency challenges in SN-Net adaptation." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 701, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 701, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 306, + 701, + 545, + 714 + ], + "type": "text", + "content": "Parameter-efficient fine-tuning. Parameter-efficient fine" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "28556" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 289, + 371 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 289, + 371 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 289, + 371 + ], + "type": "text", + "content": "tuning [27, 28, 31, 40] is a powerful alternative to vanilla full tine-tuning, which updates only a small number of parameters while freezing the majority ones. Freezing the majority of parameters allows us to optimize storage and reduces the burden on training GPU memory, as there is no need to store their gradients or other training time statistics. Recent research freezes the vast majority of parameters while fine-tuning either the parameters that are inherited in the backbone [77, 80] or additionally added, in the form of learnable prompts [31, 39], low-rank bottleneck layers [27], learnable scaling and shifting factors [41], and separated small networks [61]. Meanwhile, PEFT has been employed in numerous downstream applications. For instance, while freezing the most parameters, Polyhistor [45] learns a hyper-network to generate adapter weights [27] for multi-task adaptation, LLaMA-Adapter [83] inserts learnable prompts into Transformer layers for the instruction-following task, and ControlNet [82] employs the zero-initialized convolutions to fine-tune diffusion generative models [55]. However, these works target only one exclusive resource budget and are not scalable to diverse deployment scenarios. Unlike other methods, our ESTA is crafted to yield multiple fine-tuned models with diverse capacities while jointly incorporating stitch-specific and stitch-agnostic lightweight trainable parameters." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 372, + 289, + 589 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 372, + 289, + 589 + ], + "spans": [ + { + "bbox": [ + 47, + 372, + 289, + 589 + ], + "type": "text", + "content": "Fine-tuning with multiple models. The presence of more large-scale models has unleashed the potential to utilize multiple models instead of a single one during fine-tuning. One line of work selects the best pre-trained model in the model zoo to fine-tune. To identify the best model, prior studies quantize the transferability of the pre-trained models ahead of fine-tuning by estimating their accuracy on the downstream tasks [3, 17, 33] or the generalization capability to mitigate the domain gaps [6, 49, 66, 78]. Another way is to do model selection using an efficient online learning regime [74]. Considering the diversity among the pre-trained models, research efforts have been devoted to effectively exploit their knowledge with feature aggregation [42, 57, 58], model merging [13, 30, 48, 73], or a mixture-of-experts architecture [59]. In contrast to previous works that merely target better performance, our work employs off-the-shelf pre-trained model families to produce plentiful new models for diverse deployment requirements." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 601, + 132, + 613 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 601, + 132, + 613 + ], + "spans": [ + { + "bbox": [ + 47, + 601, + 132, + 613 + ], + "type": "text", + "content": "3. Preliminaries" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 621, + 145, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 621, + 145, + 635 + ], + "spans": [ + { + "bbox": [ + 47, + 621, + 145, + 635 + ], + "type": "text", + "content": "3.1. Model Stitching" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 641, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 641, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 47, + 641, + 288, + 715 + ], + "type": "text", + "content": "Considering an " + }, + { + "bbox": [ + 47, + 641, + 288, + 715 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 47, + 641, + 288, + 715 + ], + "type": "text", + "content": "-layer feed-forward artificial neural network " + }, + { + "bbox": [ + 47, + 641, + 288, + 715 + ], + "type": "inline_equation", + "content": "f_{\\theta_1}:\\mathcal{X}\\to \\mathcal{Y}" + }, + { + "bbox": [ + 47, + 641, + 288, + 715 + ], + "type": "text", + "content": " parameterized by " + }, + { + "bbox": [ + 47, + 641, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\theta_{1}" + }, + { + "bbox": [ + 47, + 641, + 288, + 715 + ], + "type": "text", + "content": ", that maps any input from the input space " + }, + { + "bbox": [ + 47, + 641, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{X}" + }, + { + "bbox": [ + 47, + 641, + 288, + 715 + ], + "type": "text", + "content": " to the output space " + }, + { + "bbox": [ + 47, + 641, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{V}" + }, + { + "bbox": [ + 47, + 641, + 288, + 715 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 47, + 641, + 288, + 715 + ], + "type": "inline_equation", + "content": "f_{\\theta_1}" + }, + { + "bbox": [ + 47, + 641, + 288, + 715 + ], + "type": "text", + "content": " can be denoted as a composition of functions that " + }, + { + "bbox": [ + 47, + 641, + 288, + 715 + ], + "type": "inline_equation", + "content": "f_{\\theta_1} = f_L\\circ \\dots \\circ f_1" + }, + { + "bbox": [ + 47, + 641, + 288, + 715 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 47, + 641, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\circ" + }, + { + "bbox": [ + 47, + 641, + 288, + 715 + ], + "type": "text", + "content": " denotes the function composition. In model stitching, " + }, + { + "bbox": [ + 47, + 641, + 288, + 715 + ], + "type": "inline_equation", + "content": "f_{\\theta_1}" + }, + { + "bbox": [ + 47, + 641, + 288, + 715 + ], + "type": "text", + "content": " can be split up at the " + }, + { + "bbox": [ + 47, + 641, + 288, + 715 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 47, + 641, + 288, + 715 + ], + "type": "text", + "content": "-th layer into two" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "text", + "content": "portions of functions where " + }, + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "inline_equation", + "content": "l\\in [1,L - 1]" + }, + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "text", + "content": " .Given any input " + }, + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "inline_equation", + "content": "\\pmb {X}\\subseteq \\mathcal{X}" + }, + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "text", + "content": " , the first portion of the head functions maps " + }, + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "inline_equation", + "content": "\\pmb{X}" + }, + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "text", + "content": " to activations " + }, + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "inline_equation", + "content": "\\pmb {X}_l" + }, + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "text", + "content": " at layer " + }, + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "text", + "content": " , i.e., " + }, + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "inline_equation", + "content": "H_{\\theta_1,l}(\\pmb {X}) = f_l\\circ \\dots \\circ f_1 = \\pmb {X}_l" + }, + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "text", + "content": " The second portion of the tail functions maps " + }, + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "inline_equation", + "content": "\\pmb {X}_l" + }, + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "text", + "content": " to the final output, i.e., " + }, + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "inline_equation", + "content": "T_{\\theta_1,l}(\\pmb {X}_l) = f_L\\circ \\dots \\circ f_{l + 1}" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 133, + 545, + 215 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 133, + 545, + 215 + ], + "spans": [ + { + "bbox": [ + 304, + 133, + 545, + 215 + ], + "type": "text", + "content": "Given another pre-trained " + }, + { + "bbox": [ + 304, + 133, + 545, + 215 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 304, + 133, + 545, + 215 + ], + "type": "text", + "content": "-layer artificial neural network " + }, + { + "bbox": [ + 304, + 133, + 545, + 215 + ], + "type": "inline_equation", + "content": "f_{\\theta_2}" + }, + { + "bbox": [ + 304, + 133, + 545, + 215 + ], + "type": "text", + "content": " parameterized by " + }, + { + "bbox": [ + 304, + 133, + 545, + 215 + ], + "type": "inline_equation", + "content": "\\theta_{2}" + }, + { + "bbox": [ + 304, + 133, + 545, + 215 + ], + "type": "text", + "content": " that is split up at the " + }, + { + "bbox": [ + 304, + 133, + 545, + 215 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 304, + 133, + 545, + 215 + ], + "type": "text", + "content": "-th layer that " + }, + { + "bbox": [ + 304, + 133, + 545, + 215 + ], + "type": "inline_equation", + "content": "m\\in [2,M]" + }, + { + "bbox": [ + 304, + 133, + 545, + 215 + ], + "type": "text", + "content": ", we can then employ a stitching layer " + }, + { + "bbox": [ + 304, + 133, + 545, + 215 + ], + "type": "inline_equation", + "content": "S:\\mathcal{A}_{\\theta_1,l}\\to \\mathcal{A}_{\\theta_2,m}" + }, + { + "bbox": [ + 304, + 133, + 545, + 215 + ], + "type": "text", + "content": " to map between the two activation spaces at layers " + }, + { + "bbox": [ + 304, + 133, + 545, + 215 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 304, + 133, + 545, + 215 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 133, + 545, + 215 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 304, + 133, + 545, + 215 + ], + "type": "text", + "content": ". In this case, we obtain a new network parameterized by " + }, + { + "bbox": [ + 304, + 133, + 545, + 215 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 304, + 133, + 545, + 215 + ], + "type": "text", + "content": " and the stitching layer parameters, i.e.," + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 353, + 216, + 545, + 228 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 353, + 216, + 545, + 228 + ], + "spans": [ + { + "bbox": [ + 353, + 216, + 545, + 228 + ], + "type": "interline_equation", + "content": "F _ {\\phi , S} (\\boldsymbol {X}) = T _ {\\theta_ {2}, m} \\circ S \\circ H _ {\\theta_ {1}, l} (\\boldsymbol {X}), \\tag {1}", + "image_path": "34e38c3a0b2b2d0aee69d4d05edd7b8e30b5a2acf4a57f2e09494e6de0780a75.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 228, + 545, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 228, + 545, + 239 + ], + "spans": [ + { + "bbox": [ + 304, + 228, + 545, + 239 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 228, + 545, + 239 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 304, + 228, + 545, + 239 + ], + "type": "text", + "content": " consists of partial parameters from both " + }, + { + "bbox": [ + 304, + 228, + 545, + 239 + ], + "type": "inline_equation", + "content": "\\theta_{1}" + }, + { + "bbox": [ + 304, + 228, + 545, + 239 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 228, + 545, + 239 + ], + "type": "inline_equation", + "content": "\\theta_{2}" + }, + { + "bbox": [ + 304, + 228, + 545, + 239 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 245, + 454, + 257 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 245, + 454, + 257 + ], + "spans": [ + { + "bbox": [ + 305, + 245, + 454, + 257 + ], + "type": "text", + "content": "3.2. Stitchable Neural Network" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 263, + 547, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 263, + 547, + 479 + ], + "spans": [ + { + "bbox": [ + 304, + 263, + 547, + 479 + ], + "type": "text", + "content": "In Stitchable Neural Network (SN-Net) [51], let a pretrained model family of size " + }, + { + "bbox": [ + 304, + 263, + 547, + 479 + ], + "type": "inline_equation", + "content": "Z" + }, + { + "bbox": [ + 304, + 263, + 547, + 479 + ], + "type": "text", + "content": " be " + }, + { + "bbox": [ + 304, + 263, + 547, + 479 + ], + "type": "inline_equation", + "content": "\\mathcal{Z} = \\{f_{\\theta_z}\\}_{z=1}^Z" + }, + { + "bbox": [ + 304, + 263, + 547, + 479 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 263, + 547, + 479 + ], + "type": "inline_equation", + "content": "\\theta_z" + }, + { + "bbox": [ + 304, + 263, + 547, + 479 + ], + "type": "text", + "content": " is the model parameter of the " + }, + { + "bbox": [ + 304, + 263, + 547, + 479 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 304, + 263, + 547, + 479 + ], + "type": "text", + "content": "-th model. The goal is to derive additional " + }, + { + "bbox": [ + 304, + 263, + 547, + 479 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 263, + 547, + 479 + ], + "type": "text", + "content": " candidate new networks " + }, + { + "bbox": [ + 304, + 263, + 547, + 479 + ], + "type": "inline_equation", + "content": "\\mathcal{N} = \\{F_{\\phi_n, S_n}\\}_{n=1}^N" + }, + { + "bbox": [ + 304, + 263, + 547, + 479 + ], + "type": "text", + "content": " to adapt to various resource constraints. To do so, SN-Net first selects a pair of pre-trained models (anchors) from " + }, + { + "bbox": [ + 304, + 263, + 547, + 479 + ], + "type": "inline_equation", + "content": "\\mathcal{Z}" + }, + { + "bbox": [ + 304, + 263, + 547, + 479 + ], + "type": "text", + "content": " and then stitches them with Eq. (1) at different layer indexes to get " + }, + { + "bbox": [ + 304, + 263, + 547, + 479 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 263, + 547, + 479 + ], + "type": "text", + "content": " stitches. Then, SN-Net samples the stitches randomly and jointly optimizes them. An overview of SN-Net is depicted in Figure 1 (a). Since the anchors vary in scale, the newly assembled stitches have diverse performances and complexities. Importantly, SN-Net gives practical principles to design the space of " + }, + { + "bbox": [ + 304, + 263, + 547, + 479 + ], + "type": "inline_equation", + "content": "\\mathcal{N}" + }, + { + "bbox": [ + 304, + 263, + 547, + 479 + ], + "type": "text", + "content": " that one should 1) stitch a pair of nearest anchors in terms of model complexity in a model family; 2) stitch the head of a faster and smaller anchor to the tail of a larger and slower anchor. Unless specified otherwise, we adopt these as the default experimental settings in this paper." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 488, + 362, + 501 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 488, + 362, + 501 + ], + "spans": [ + { + "bbox": [ + 305, + 488, + 362, + 501 + ], + "type": "text", + "content": "4. Method" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 509, + 547, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 509, + 547, + 594 + ], + "spans": [ + { + "bbox": [ + 304, + 509, + 547, + 594 + ], + "type": "text", + "content": "In this section, we introduce our ESTA framework, which consists of two major components: parameter-efficient stitch fine-tuning (PST) and a simple one-stage deployment pipeline. PST is tailored to cheaply fine-tune a palette of stitches (Section 4.1 and Figure 1 (b)). The one-stage deployment pipeline aims to simultaneously save deployment time and improve adaptation performance (Section 4.2)." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 599, + 508, + 613 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 599, + 508, + 613 + ], + "spans": [ + { + "bbox": [ + 305, + 599, + 508, + 613 + ], + "type": "text", + "content": "4.1. Parameter-efficient Stitch Fine-tuning" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 618, + 545, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 618, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 304, + 618, + 545, + 689 + ], + "type": "text", + "content": "To address the problems of large storage and memory consumption when adapting SN-Net to downstream tasks with full fine-tuning, we introduce a parameter-efficient stitch fine-tuning (PST) method as depicted in Figure 1 (b). Our basic idea is to adapt the representative sparse fine-tuning method LoRA [28] for SN-Net fine-tuning." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 689, + 545, + 713 + ], + "type": "text", + "content": "LoRA on Transformer layers. Specifically, let any pretrained weight matrix in the multi-head self-attention layer" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "28557" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 51, + 74, + 185, + 156 + ], + "blocks": [ + { + "bbox": [ + 51, + 74, + 185, + 156 + ], + "lines": [ + { + "bbox": [ + 51, + 74, + 185, + 156 + ], + "spans": [ + { + "bbox": [ + 51, + 74, + 185, + 156 + ], + "type": "image", + "image_path": "f58364be157338b8aede459082270502387115ac7876c1765f57901f865d8003.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 110, + 158, + 119, + 166 + ], + "lines": [ + { + "bbox": [ + 110, + 158, + 119, + 166 + ], + "spans": [ + { + "bbox": [ + 110, + 158, + 119, + 166 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 192, + 74, + 383, + 157 + ], + "blocks": [ + { + "bbox": [ + 192, + 74, + 383, + 157 + ], + "lines": [ + { + "bbox": [ + 192, + 74, + 383, + 157 + ], + "spans": [ + { + "bbox": [ + 192, + 74, + 383, + 157 + ], + "type": "image", + "image_path": "e73f4f8d58c28c8099c1a7173896f46f9c4ebe670937baed455fa23fa39f713d.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 283, + 158, + 293, + 166 + ], + "lines": [ + { + "bbox": [ + 283, + 158, + 293, + 166 + ], + "spans": [ + { + "bbox": [ + 283, + 158, + 293, + 166 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 389, + 81, + 539, + 157 + ], + "blocks": [ + { + "bbox": [ + 389, + 81, + 539, + 157 + ], + "lines": [ + { + "bbox": [ + 389, + 81, + 539, + 157 + ], + "spans": [ + { + "bbox": [ + 389, + 81, + 539, + 157 + ], + "type": "image", + "image_path": "73e269c68c0a28a129cb02ef51ae34e723c6d46556b82f43ab9cbab09c2b5e86.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 467, + 158, + 476, + 166 + ], + "lines": [ + { + "bbox": [ + 467, + 158, + 476, + 166 + ], + "spans": [ + { + "bbox": [ + 467, + 158, + 476, + 166 + ], + "type": "text", + "content": "(c)" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 48, + 261, + 167, + 353 + ], + "blocks": [ + { + "bbox": [ + 46, + 168, + 547, + 257 + ], + "lines": [ + { + "bbox": [ + 46, + 168, + 547, + 257 + ], + "spans": [ + { + "bbox": [ + 46, + 168, + 547, + 257 + ], + "type": "text", + "content": "Figure 1. (a) Illustration of Stretchable Neural Network [51]. With two anchors from the same model family, SN-Net connects the early layers of the smaller one to the latter layers of the larger one with stitching layers to obtain a set of new networks with different performance-efficiency trade-offs, e.g., the path in Blue. (b) Overview of our PST method tailored for fine-tuning a palette of stitches, which integrates stitch-agnostic LoRA modules with stitch-specific bias terms, aiming to promote diverse representations among stitches while maintaining low trainable parameters. (c) Overview of our task-specific stitch sampling. We estimate the importance scores of the stitches with a scoring function " + }, + { + "bbox": [ + 46, + 168, + 547, + 257 + ], + "type": "inline_equation", + "content": "Q(\\cdot, \\cdot)" + }, + { + "bbox": [ + 46, + 168, + 547, + 257 + ], + "type": "text", + "content": " and accumulate them as global statistics with moving averages. For a resource constraint " + }, + { + "bbox": [ + 46, + 168, + 547, + 257 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 46, + 168, + 547, + 257 + ], + "type": "text", + "content": ", we sample with a categorical distribution " + }, + { + "bbox": [ + 46, + 168, + 547, + 257 + ], + "type": "inline_equation", + "content": "\\pi(\\mathcal{N}_{\\gamma})" + }, + { + "bbox": [ + 46, + 168, + 547, + 257 + ], + "type": "text", + "content": " that is parameterized by the normalized importance scores so as to assign the important stitches with higher sampling probabilities. After fine-tuning, we directly deploy the stitches with the highest scores to avoid the costly evaluation stage." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 48, + 261, + 167, + 353 + ], + "lines": [ + { + "bbox": [ + 48, + 261, + 167, + 353 + ], + "spans": [ + { + "bbox": [ + 48, + 261, + 167, + 353 + ], + "type": "image", + "image_path": "61edbb8eb4bb41ca5c949343fd8c80b8b6d8584515dc95f17e9a0378ea3f20db.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 167, + 261, + 282, + 353 + ], + "blocks": [ + { + "bbox": [ + 167, + 261, + 282, + 353 + ], + "lines": [ + { + "bbox": [ + 167, + 261, + 282, + 353 + ], + "spans": [ + { + "bbox": [ + 167, + 261, + 282, + 353 + ], + "type": "image", + "image_path": "51758cda33ca580a38fa388b6ee0a71a8833b34eccb6413244410d7595ac2eae.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 354, + 287, + 441 + ], + "lines": [ + { + "bbox": [ + 46, + 354, + 287, + 441 + ], + "spans": [ + { + "bbox": [ + 46, + 354, + 287, + 441 + ], + "type": "text", + "content": "Figure 2. Distribution of pair-wise gradient angles among stitches when updating shared weights at fine-tuning iteration 600. We highlight angle " + }, + { + "bbox": [ + 46, + 354, + 287, + 441 + ], + "type": "inline_equation", + "content": "90^{\\circ}" + }, + { + "bbox": [ + 46, + 354, + 287, + 441 + ], + "type": "text", + "content": " with a dashed red line. For simplicity, we show the gradient angles among the combined query, key, and value projection matrices for a total of 32 stitches when stitching ViT-Ti and ViT-S anchors. Generally, the gradient angles are larger in the target domain Stanford Cars [18] than in the source domain ImageNet-1k [56]." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 445, + 287, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 445, + 287, + 495 + ], + "spans": [ + { + "bbox": [ + 46, + 445, + 287, + 495 + ], + "type": "text", + "content": "be " + }, + { + "bbox": [ + 46, + 445, + 287, + 495 + ], + "type": "inline_equation", + "content": "W \\in \\mathbb{R}^{d \\times k}" + }, + { + "bbox": [ + 46, + 445, + 287, + 495 + ], + "type": "text", + "content": ", we freeze " + }, + { + "bbox": [ + 46, + 445, + 287, + 495 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 46, + 445, + 287, + 495 + ], + "type": "text", + "content": " and insert trainable low-rank decomposition matrices " + }, + { + "bbox": [ + 46, + 445, + 287, + 495 + ], + "type": "inline_equation", + "content": "W_{\\mathrm{down}} \\in \\mathbb{R}^{d \\times r}" + }, + { + "bbox": [ + 46, + 445, + 287, + 495 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 445, + 287, + 495 + ], + "type": "inline_equation", + "content": "W_{\\mathrm{up}} \\in \\mathbb{R}^{r \\times k}" + }, + { + "bbox": [ + 46, + 445, + 287, + 495 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 445, + 287, + 495 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 46, + 445, + 287, + 495 + ], + "type": "text", + "content": " is the pre-defined rank that " + }, + { + "bbox": [ + 46, + 445, + 287, + 495 + ], + "type": "inline_equation", + "content": "r \\ll \\min(d, k)" + }, + { + "bbox": [ + 46, + 445, + 287, + 495 + ], + "type": "text", + "content": ". In this way, the updated version of " + }, + { + "bbox": [ + 46, + 445, + 287, + 495 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 46, + 445, + 287, + 495 + ], + "type": "text", + "content": " can be formulated as" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 111, + 502, + 287, + 516 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 111, + 502, + 287, + 516 + ], + "spans": [ + { + "bbox": [ + 111, + 502, + 287, + 516 + ], + "type": "interline_equation", + "content": "\\boldsymbol {W} ^ {\\prime} \\leftarrow \\boldsymbol {W} + \\boldsymbol {W} _ {\\text {d o w n}} \\boldsymbol {W} _ {\\text {u p}}. \\tag {2}", + "image_path": "916559f3c3d8f317937ad491fbfe0290ef03845724a867b7958b826fe7c09a6e.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 521, + 287, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 521, + 287, + 557 + ], + "spans": [ + { + "bbox": [ + 46, + 521, + 287, + 557 + ], + "type": "text", + "content": "We follow [28] to respectively use Gaussian and zero initializations for " + }, + { + "bbox": [ + 46, + 521, + 287, + 557 + ], + "type": "inline_equation", + "content": "W_{\\mathrm{up}}" + }, + { + "bbox": [ + 46, + 521, + 287, + 557 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 521, + 287, + 557 + ], + "type": "inline_equation", + "content": "W_{\\mathrm{down}}" + }, + { + "bbox": [ + 46, + 521, + 287, + 557 + ], + "type": "text", + "content": ", so that " + }, + { + "bbox": [ + 46, + 521, + 287, + 557 + ], + "type": "inline_equation", + "content": "W_{\\mathrm{down}}W_{\\mathrm{up}}" + }, + { + "bbox": [ + 46, + 521, + 287, + 557 + ], + "type": "text", + "content": " is zero at the beginning of fine-tuning." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 46, + 558, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 558, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 558, + 288, + 715 + ], + "type": "text", + "content": "LoRA on stitching layers. We further extend the low-rank weight update to stitching layers. In SN-Net, a stitching layer is designed as a full-rank transformation matrix. Without loss of generality, we show the low-rank update of stitching layer " + }, + { + "bbox": [ + 46, + 558, + 288, + 715 + ], + "type": "inline_equation", + "content": "S_{n}" + }, + { + "bbox": [ + 46, + 558, + 288, + 715 + ], + "type": "text", + "content": " that stitches " + }, + { + "bbox": [ + 46, + 558, + 288, + 715 + ], + "type": "inline_equation", + "content": "f_{\\theta_1}" + }, + { + "bbox": [ + 46, + 558, + 288, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 558, + 288, + 715 + ], + "type": "inline_equation", + "content": "f_{\\theta_2}" + }, + { + "bbox": [ + 46, + 558, + 288, + 715 + ], + "type": "text", + "content": " at layers " + }, + { + "bbox": [ + 46, + 558, + 288, + 715 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 46, + 558, + 288, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 558, + 288, + 715 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 46, + 558, + 288, + 715 + ], + "type": "text", + "content": ", respectively. We first follow the initialization of stitching layers in [12, 51] to align the activations at layers " + }, + { + "bbox": [ + 46, + 558, + 288, + 715 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 46, + 558, + 288, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 558, + 288, + 715 + ], + "type": "inline_equation", + "content": "m - 1" + }, + { + "bbox": [ + 46, + 558, + 288, + 715 + ], + "type": "text", + "content": ". Let the activations be " + }, + { + "bbox": [ + 46, + 558, + 288, + 715 + ], + "type": "inline_equation", + "content": "H_{\\theta_1,l}(\\mathbf{X}) \\in \\mathbb{R}^{b \\times d_1}" + }, + { + "bbox": [ + 46, + 558, + 288, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 558, + 288, + 715 + ], + "type": "inline_equation", + "content": "H_{\\theta_2,m - 1}(\\mathbf{X}) \\in \\mathbb{R}^{b \\times d_2}" + }, + { + "bbox": [ + 46, + 558, + 288, + 715 + ], + "type": "text", + "content": " with sequence length " + }, + { + "bbox": [ + 46, + 558, + 288, + 715 + ], + "type": "inline_equation", + "content": "b" + }, + { + "bbox": [ + 46, + 558, + 288, + 715 + ], + "type": "text", + "content": " and feature dimensions " + }, + { + "bbox": [ + 46, + 558, + 288, + 715 + ], + "type": "inline_equation", + "content": "d_1" + }, + { + "bbox": [ + 46, + 558, + 288, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 558, + 288, + 715 + ], + "type": "inline_equation", + "content": "d_2" + }, + { + "bbox": [ + 46, + 558, + 288, + 715 + ], + "type": "text", + "content": ". The stitching layer is parameterized by a transformation matrix " + }, + { + "bbox": [ + 46, + 558, + 288, + 715 + ], + "type": "inline_equation", + "content": "M \\in \\mathbb{R}^{d_1 \\times d_2}" + }, + { + "bbox": [ + 46, + 558, + 288, + 715 + ], + "type": "text", + "content": ", which is initialized by solving a least squares problem, i.e., " + }, + { + "bbox": [ + 46, + 558, + 288, + 715 + ], + "type": "inline_equation", + "content": "M = H_{\\theta_1,l}(X)^{\\dagger}H_{\\theta_2,m - 1}(X)" + }, + { + "bbox": [ + 46, + 558, + 288, + 715 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 558, + 288, + 715 + ], + "type": "inline_equation", + "content": "H_{\\theta_1,l}(X)^{\\dagger}" + }, + { + "bbox": [ + 46, + 558, + 288, + 715 + ], + "type": "text", + "content": " is the" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 262, + 545, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 262, + 545, + 346 + ], + "spans": [ + { + "bbox": [ + 304, + 262, + 545, + 346 + ], + "type": "text", + "content": "Moore-Penrose pseudoinverse of " + }, + { + "bbox": [ + 304, + 262, + 545, + 346 + ], + "type": "inline_equation", + "content": "H_{\\theta_1,l}(\\mathbf{X})" + }, + { + "bbox": [ + 304, + 262, + 545, + 346 + ], + "type": "text", + "content": ". We find that such initialization already has impressive representational capacity on the target domain and they can be updated with low-rank decomposition without hurting the performance too much. Accordingly, we update " + }, + { + "bbox": [ + 304, + 262, + 545, + 346 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 304, + 262, + 545, + 346 + ], + "type": "text", + "content": " similar to updating " + }, + { + "bbox": [ + 304, + 262, + 545, + 346 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 304, + 262, + 545, + 346 + ], + "type": "text", + "content": " in Eq. (2) by approximating its update with two learnable low-rank matrices with the same initializations." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 348, + 546, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 348, + 546, + 613 + ], + "spans": [ + { + "bbox": [ + 304, + 348, + 546, + 613 + ], + "type": "text", + "content": "Although the low-rank updates significantly enhance parameter efficiency, the low-rank essence [2] largely limits the network capacity. Particularly, in task adaptation, we often observe conflicting updates on the shared weights among the stitches. Figure 2 gives an example, where we stitch pre-trained anchors on ImageNet-1k [56] and Stanford Cars [18] and visualize the distribution of pair-wise gradient angles among different stitches when updating shared weights. We find that stitches agree less and conflict more (i.e., with more large angles) on the target domain Stanford Cars compared to the source domain ImageNet-1k. Stitch-specific bias. To address the above issue, we further introduce stitch-specific bias terms as depicted in Figure 1 (b). Particularly, for a linear layer parameterized by " + }, + { + "bbox": [ + 304, + 348, + 546, + 613 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 304, + 348, + 546, + 613 + ], + "type": "text", + "content": " and a bias term " + }, + { + "bbox": [ + 304, + 348, + 546, + 613 + ], + "type": "inline_equation", + "content": "b" + }, + { + "bbox": [ + 304, + 348, + 546, + 613 + ], + "type": "text", + "content": ", we add bias term " + }, + { + "bbox": [ + 304, + 348, + 546, + 613 + ], + "type": "inline_equation", + "content": "b^s" + }, + { + "bbox": [ + 304, + 348, + 546, + 613 + ], + "type": "text", + "content": " when optimizing the " + }, + { + "bbox": [ + 304, + 348, + 546, + 613 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 304, + 348, + 546, + 613 + ], + "type": "text", + "content": "-th stitch, for which the output activation becomes " + }, + { + "bbox": [ + 304, + 348, + 546, + 613 + ], + "type": "inline_equation", + "content": "WX + b + b^s" + }, + { + "bbox": [ + 304, + 348, + 546, + 613 + ], + "type": "text", + "content": ". We employ a set of distinct bias terms at different layers for each stitch. In this way, the stitches are encouraged to learn distinct feature representations for different resource requirements. To restrict the number of trainable parameters, following [28], we only apply PST to self-attention layers while freezing feed-forward layers." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 305, + 622, + 476, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 622, + 476, + 635 + ], + "spans": [ + { + "bbox": [ + 305, + 622, + 476, + 635 + ], + "type": "text", + "content": "4.2. One-stage Deployment Pipeline" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "type": "text", + "content": "Simultaneously adapt and stitch anchors. As aforementioned, given a model family " + }, + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{Z}" + }, + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "type": "text", + "content": " pre-trained in the source domain, directly adapting SN-Net to a target domain typically involves three stages. First, adapting each anchor to the target domain " + }, + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "type": "text", + "content": " by solving " + }, + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\theta_z^* = \\mathrm{argmin}_{\\theta_z} \\mathcal{L}(\\theta_z; \\mathcal{D})" + }, + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "type": "text", + "content": ". In the second stage, at each training iteration, SN-Net sam" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "28558" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 289, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 289, + 251 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 289, + 251 + ], + "type": "text", + "content": "plies a stitch and optimizes it with objective " + }, + { + "bbox": [ + 47, + 72, + 289, + 251 + ], + "type": "inline_equation", + "content": "\\mathcal{L}(\\phi_n^*, S_n; \\mathcal{D})" + }, + { + "bbox": [ + 47, + 72, + 289, + 251 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 47, + 72, + 289, + 251 + ], + "type": "inline_equation", + "content": "\\phi_n^*" + }, + { + "bbox": [ + 47, + 72, + 289, + 251 + ], + "type": "text", + "content": " is the set of SN-Net parameters for stitch " + }, + { + "bbox": [ + 47, + 72, + 289, + 251 + ], + "type": "inline_equation", + "content": "F_{\\phi_n, S_n}" + }, + { + "bbox": [ + 47, + 72, + 289, + 251 + ], + "type": "text", + "content": " which has been optimized once in the first stage. In the final stage, all optimized stitches need to be evaluated to identify the ones on the Pareto frontier for diverse resource constraints. Such a three-stage approach is expensive and sub-optimal (see more discussions in the supplementary material). Thus, we propose to adapt and stitch the anchors simultaneously within one stage, by directly optimizing " + }, + { + "bbox": [ + 47, + 72, + 289, + 251 + ], + "type": "inline_equation", + "content": "\\mathcal{L}(\\phi_n, S_n; \\mathcal{D})" + }, + { + "bbox": [ + 47, + 72, + 289, + 251 + ], + "type": "text", + "content": " for each sampled stitch. Moreover, we introduce a novel task-specific stitch-sampling method to allow the promising stitches more likely to be sampled, based on a stitch importance score. After fine-tuning, the stitches with the highest scores are naturally selected for deployment without the need for the above final stage evaluation." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 251, + 289, + 370 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 251, + 289, + 370 + ], + "spans": [ + { + "bbox": [ + 47, + 251, + 289, + 370 + ], + "type": "text", + "content": "Task-specific stitch sampling. SN-Net assumes that all stitches are equally important and performs random (uniform) sampling during training. However, as recognized by prior research [43, 70] that random sampling causes a gap between training and deployment as only the best stitches on the Pareto frontier are actually deployed. Moreover, since the importance of the pre-trained weights varies in different downstream tasks [19, 22, 75], we argue that the performance of stitches that are parameterized by these weights is also different across these tasks." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 371, + 289, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 371, + 289, + 515 + ], + "spans": [ + { + "bbox": [ + 46, + 371, + 289, + 515 + ], + "type": "text", + "content": "To this end, we propose to assign the important stitches that are likely to be on the Pareto frontier with higher sampling probabilities during fine-tuning to ensure that they are optimized sufficiently. Specifically, we first estimate the importance of each stitch with a data-dependent saliency pruning metric SNIP [37], which measures the importance with the first-order gradient information with barely any extra computational cost. When the " + }, + { + "bbox": [ + 46, + 371, + 289, + 515 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 46, + 371, + 289, + 515 + ], + "type": "text", + "content": "-th stitch is sampled, we can get its importance score " + }, + { + "bbox": [ + 46, + 371, + 289, + 515 + ], + "type": "inline_equation", + "content": "Q(F_{\\phi, S_n}, X)" + }, + { + "bbox": [ + 46, + 371, + 289, + 515 + ], + "type": "text", + "content": " with the scoring function " + }, + { + "bbox": [ + 46, + 371, + 289, + 515 + ], + "type": "inline_equation", + "content": "Q(\\cdot, \\cdot)" + }, + { + "bbox": [ + 46, + 371, + 289, + 515 + ], + "type": "text", + "content": " [37]. To obtain robust importance scores for sampling, we accumulate scores with moving average during training, i.e.," + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 88, + 523, + 287, + 539 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 88, + 523, + 287, + 539 + ], + "spans": [ + { + "bbox": [ + 88, + 523, + 287, + 539 + ], + "type": "interline_equation", + "content": "q _ {n} ^ {t} \\leftarrow \\eta q _ {n} ^ {t - 1} + (1 - \\eta) Q \\left(F _ {\\phi_ {n}, S _ {n}}, \\boldsymbol {X}\\right), \\tag {3}", + "image_path": "2bd80fbc2c8f731be14de6d5313605fbd16bf324efd190333189ed6e06c904a1.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 545, + 289, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 545, + 289, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 545, + 289, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 545, + 289, + 714 + ], + "type": "inline_equation", + "content": "q_{n}^{t}" + }, + { + "bbox": [ + 47, + 545, + 289, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 545, + 289, + 714 + ], + "type": "inline_equation", + "content": "q_{n}^{t - 1}" + }, + { + "bbox": [ + 47, + 545, + 289, + 714 + ], + "type": "text", + "content": " are the importance score at the " + }, + { + "bbox": [ + 47, + 545, + 289, + 714 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 47, + 545, + 289, + 714 + ], + "type": "text", + "content": "-th and " + }, + { + "bbox": [ + 47, + 545, + 289, + 714 + ], + "type": "inline_equation", + "content": "(t - 1)" + }, + { + "bbox": [ + 47, + 545, + 289, + 714 + ], + "type": "text", + "content": "-th iteration, respectively, and " + }, + { + "bbox": [ + 47, + 545, + 289, + 714 + ], + "type": "inline_equation", + "content": "\\eta \\in [0,1)" + }, + { + "bbox": [ + 47, + 545, + 289, + 714 + ], + "type": "text", + "content": " is the momentum coefficient. In this way, we can get stable importance scores. After a warm-up period that employs uniform sampling to accumulate the scores, we assign the important stitches with higher sampling probabilities for the rest of the fine-tuning epochs. To do so, we uniformly divide the resource constraint range of our stitches into several intervals and sample an interval " + }, + { + "bbox": [ + 47, + 545, + 289, + 714 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 47, + 545, + 289, + 714 + ], + "type": "text", + "content": " with uniform sampling. Accordingly, we can obtain a subset of stitches " + }, + { + "bbox": [ + 47, + 545, + 289, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{N}_{\\gamma}" + }, + { + "bbox": [ + 47, + 545, + 289, + 714 + ], + "type": "text", + "content": " whose resource constraints belong to " + }, + { + "bbox": [ + 47, + 545, + 289, + 714 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 47, + 545, + 289, + 714 + ], + "type": "text", + "content": " and their corresponding importance scores " + }, + { + "bbox": [ + 47, + 545, + 289, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{Q}_{\\gamma}" + }, + { + "bbox": [ + 47, + 545, + 289, + 714 + ], + "type": "text", + "content": ". We can then define a categorical distribution based on the normalized importance scores, i.e., " + }, + { + "bbox": [ + 47, + 545, + 289, + 714 + ], + "type": "inline_equation", + "content": "\\pi (\\mathcal{N}_{\\gamma}) = \\mathrm{softmax}(\\mathcal{Q}_{\\gamma})" + }, + { + "bbox": [ + 47, + 545, + 289, + 714 + ], + "type": "text", + "content": " and finally get the sampled stitch" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 305, + 72, + 545, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 95 + ], + "type": "inline_equation", + "content": "F_{\\phi_n,S_n}\\sim \\pi (\\mathcal{N}_\\gamma)" + }, + { + "bbox": [ + 305, + 72, + 545, + 95 + ], + "type": "text", + "content": ". We show that our task-specific stitch sampling improves the performance in Section 5.3." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 96, + 547, + 193 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 96, + 547, + 193 + ], + "spans": [ + { + "bbox": [ + 304, + 96, + 547, + 193 + ], + "type": "text", + "content": "Additionally, with the sampled stitch, SN-Net employs knowledge distillation with a RegNetY-160 [54] pre-trained teacher to improve its performance. However, pre-trained teachers on downstream tasks are often unavailable. Therefore, in each training iteration, we always sample and train the largest anchor (teacher) and transfer its knowledge to the sampled stitch (student) similar to inplace distillation [79]. In practice, we employ the hard-label distillation [26, 63]." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 194, + 547, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 194, + 547, + 384 + ], + "spans": [ + { + "bbox": [ + 304, + 194, + 547, + 384 + ], + "type": "text", + "content": "Reuse stitch importance scores for deployment. After fine-tuning, with the set of accumulated importance scores " + }, + { + "bbox": [ + 304, + 194, + 547, + 384 + ], + "type": "inline_equation", + "content": "\\mathcal{Q}_{\\gamma}" + }, + { + "bbox": [ + 304, + 194, + 547, + 384 + ], + "type": "text", + "content": " for all stitches, we directly use them to select the stitches with the highest scores to deploy, which eliminates the need for the costly evaluation of the final stage of the straightforward SN-Net adaptation solution. We empirically observe that the deployed stitches are mostly on the Pareto frontier in Section 5.3, suggesting our estimation of the important stitches is accurate. Furthermore, in the supplementary material, we show that the important stitches differ across different tasks, proving the need for task-specific deployment strategies, as highlighted in [19, 22, 75]. It's worth noting that although the SNIP metric has been widely employed as a one-shot metric to efficiently estimate network performance [1, 69], we extend it to estimate stitch importance for better training and selection." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 398, + 388, + 411 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 398, + 388, + 411 + ], + "spans": [ + { + "bbox": [ + 306, + 398, + 388, + 411 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 418, + 416, + 430 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 418, + 416, + 430 + ], + "spans": [ + { + "bbox": [ + 306, + 418, + 416, + 430 + ], + "type": "text", + "content": "5.1. Visual Recognition" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 437, + 545, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 437, + 545, + 568 + ], + "spans": [ + { + "bbox": [ + 304, + 437, + 545, + 568 + ], + "type": "text", + "content": "We evaluate the effectiveness of our method on a total of 25 downstream visual recognition tasks, including fine-grained visual classification (FGVC) benchmark, common visual classification benchmark CIFAR-100 [34], and VTAB-1k [81] benchmark. FGVC benchmark contains NABirds [67], CUB-200-2011 [68], Stanford Cars [18], Stanford Dogs [32], and Oxford Flowers [50] tasks and the VTAB-1k benchmark includes 19 tasks in different domains, each of which has 800 training samples. For visual recognition experiments, we stitch ImageNet-21k pretrained ViT-Ti/S/B models [16] from [60]." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 570, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 547, + 713 + ], + "type": "text", + "content": "Implementation details. We employ the stitching space and settings (kernel size as 2 and stride as 1) in [51] for downstream visual recognition tasks. We uniformly divide the resource constraints supported by the stitches into around 15 intervals and deploy one stitch within each interval. We set the hyper-parameter " + }, + { + "bbox": [ + 304, + 570, + 547, + 713 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 304, + 570, + 547, + 713 + ], + "type": "text", + "content": " for the rank of weight updating in the self-attention layers to be 32, 16, and 8 for ViT-Ti, ViT-S, and ViT-B, respectively. We also set " + }, + { + "bbox": [ + 304, + 570, + 547, + 713 + ], + "type": "inline_equation", + "content": "r = d_1 // 4" + }, + { + "bbox": [ + 304, + 570, + 547, + 713 + ], + "type": "text", + "content": " universally for all stitching layers and " + }, + { + "bbox": [ + 304, + 570, + 547, + 713 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 304, + 570, + 547, + 713 + ], + "type": "text", + "content": " in Eq. (3) as 0.9 by grid search. We follow [31] to fine-tune 100 epochs on each task and select the other hyper-parameters and augmentation methods. We include more implementation details in" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "28559" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 49, + 71, + 174, + 161 + ], + "blocks": [ + { + "bbox": [ + 49, + 71, + 174, + 161 + ], + "lines": [ + { + "bbox": [ + 49, + 71, + 174, + 161 + ], + "spans": [ + { + "bbox": [ + 49, + 71, + 174, + 161 + ], + "type": "image", + "image_path": "e307fd95fed9c167bc222d554736d56ea0f914ca7e6f482f6bbdce418c4110b5.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 162, + 547, + 195 + ], + "lines": [ + { + "bbox": [ + 46, + 162, + 547, + 195 + ], + "spans": [ + { + "bbox": [ + 46, + 162, + 547, + 195 + ], + "type": "text", + "content": "Figure 3. Performance comparisons with SN-Net [51] for adapting ViT-Ti/S/B pre-trained on ImageNet-22k [15] to Stanford Cars [18], CUB-200-2011 [68], Stanford Dogs [32], and NABirds [67]. We denote individually fine-tuned anchors as yellow stars. We also show the number of trainable parameters." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 173, + 71, + 298, + 161 + ], + "blocks": [ + { + "bbox": [ + 173, + 71, + 298, + 161 + ], + "lines": [ + { + "bbox": [ + 173, + 71, + 298, + 161 + ], + "spans": [ + { + "bbox": [ + 173, + 71, + 298, + 161 + ], + "type": "image", + "image_path": "1af46de912baf31fd1ba0b6c8e60c6943b6add1837011fa259f6feca9af4f84d.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 298, + 71, + 421, + 161 + ], + "blocks": [ + { + "bbox": [ + 298, + 71, + 421, + 161 + ], + "lines": [ + { + "bbox": [ + 298, + 71, + 421, + 161 + ], + "spans": [ + { + "bbox": [ + 298, + 71, + 421, + 161 + ], + "type": "image", + "image_path": "2d1d320775be21db95f2bb6155c0f27ee7d00e8a46405e38fd095248054e2753.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 421, + 71, + 545, + 161 + ], + "blocks": [ + { + "bbox": [ + 421, + 71, + 545, + 161 + ], + "lines": [ + { + "bbox": [ + 421, + 71, + 545, + 161 + ], + "spans": [ + { + "bbox": [ + 421, + 71, + 545, + 161 + ], + "type": "image", + "image_path": "964957dde80b4e1b42c30c4ae4610a4a08c5518734d116b17c0b1eb272edf2da.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 48, + 200, + 167, + 285 + ], + "blocks": [ + { + "bbox": [ + 48, + 200, + 167, + 285 + ], + "lines": [ + { + "bbox": [ + 48, + 200, + 167, + 285 + ], + "spans": [ + { + "bbox": [ + 48, + 200, + 167, + 285 + ], + "type": "image", + "image_path": "68d6af674840a83cd087b0f24427380108f82ef6b28e3f28334f9c7e03424e0b.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 286, + 287, + 331 + ], + "lines": [ + { + "bbox": [ + 46, + 286, + 287, + 331 + ], + "spans": [ + { + "bbox": [ + 46, + 286, + 287, + 331 + ], + "type": "text", + "content": "Figure 4. Performance comparisons with SN-Net [51] for adapting ViT-Ti/S/B pre-trained on ImageNet-22k [15] to VTAB-1k [81] and CIFAR-100 [34]. We denote individually fine-tuned anchors as yellow stars and also show the number of trainable parameters." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 167, + 200, + 286, + 285 + ], + "blocks": [ + { + "bbox": [ + 167, + 200, + 286, + 285 + ], + "lines": [ + { + "bbox": [ + 167, + 200, + 286, + 285 + ], + "spans": [ + { + "bbox": [ + 167, + 200, + 286, + 285 + ], + "type": "image", + "image_path": "671fb773bb232eec83b70da8eb196089159085b58ba9bc8f8fbdeda9aa8451c4.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 50, + 371, + 285, + 407 + ], + "blocks": [ + { + "bbox": [ + 46, + 337, + 287, + 370 + ], + "lines": [ + { + "bbox": [ + 46, + 337, + 287, + 370 + ], + "spans": [ + { + "bbox": [ + 46, + 337, + 287, + 370 + ], + "type": "text", + "content": "Table 2. Relative response quality to Alpaca-LoRA 7B [62] on Vicuna Bench [10]. Our Stitched LLaMA successfully interpolates the answer quality between Alpaca-LoRA 7B and 13B." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 50, + 371, + 285, + 407 + ], + "lines": [ + { + "bbox": [ + 50, + 371, + 285, + 407 + ], + "spans": [ + { + "bbox": [ + 50, + 371, + 285, + 407 + ], + "type": "table", + "html": "
Alpaca-LoRAStitched LLaMA
Parameter # (B)6.713.06.97.811.512.513.0
Quality (%)100123113114115120123
", + "image_path": "2d4ce083b240f7e2038c00ec3d779a6e476150d743dc7b051c645c2922cf38c6.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 414, + 161, + 425 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 414, + 161, + 425 + ], + "spans": [ + { + "bbox": [ + 47, + 414, + 161, + 425 + ], + "type": "text", + "content": "the supplementary material." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 426, + 287, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 426, + 287, + 581 + ], + "spans": [ + { + "bbox": [ + 46, + 426, + 287, + 581 + ], + "type": "text", + "content": "Main results. We compare the performance of our ESTA with SN-Net and the individual anchors fine-tuned with LoRA [28] on various datasets. The FLOPs-accuracy curves are visualized in Figures 3 and 4. We report the averaged results over the three task groups on VTAB-1k following [31]. Overall, the stitches obtained by our ESTA exhibit smooth FLOPs-accuracy trade-off curves consistently across all datasets. Notably, the curve for our ESTA outperforms SN-Net by significant margins with much fewer trainable parameters. We show that the inferior results of SN-Net on downstream tasks are led by the sub-optimal designs of fine-tuning all model parameters and a three-stage adaptation pipeline in Section 5.3." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 582, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 582, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 582, + 288, + 714 + ], + "type": "text", + "content": "We also observe that the stitches even outperform the anchors with comparable or lower FLOPs in many cases. For instance, stitches outperform individually fine-tuned ViT-S/B anchors by clear margins on NABirds, CUB-200-2011, Stanford Cars, and Stanford Dogs datasets with comparable computational complexity. This phenomenon is consistent with the observations made in [52, 79]. We conjecture that weight sharing among the stitches serves as a strong regularization to improve their generalization. We show more results on few-shot learning and stitching Convnets [46] in the supplementary material." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 200, + 453, + 213 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 200, + 453, + 213 + ], + "spans": [ + { + "bbox": [ + 305, + 200, + 453, + 213 + ], + "type": "text", + "content": "5.2. Instruction-following Task" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 219, + 545, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 219, + 545, + 327 + ], + "spans": [ + { + "bbox": [ + 304, + 219, + 545, + 327 + ], + "type": "text", + "content": "We also evaluate the instruction-following capability when stitching the LLaMA family [64] with our ESTA framework. We adapt and fine-tune LLaMA-7B/13B on the 52K instruction-following dataset from [62] to obtain chatbot stitches of varying sizes. The 52K instruction-following data is generated from 175 human-written instruction-output pairs with [71]. Each sample in the dataset is formulated by a task description, the context of the task, and the answer that is generated by GPT-3.5 (text-davinci-003) [7]." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 327, + 546, + 482 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 327, + 546, + 482 + ], + "spans": [ + { + "bbox": [ + 304, + 327, + 546, + 482 + ], + "type": "text", + "content": "Implementation details. We employ the stitching space and settings (kernel size as 1 and stride as 1) in [51] for the instruction-following task and introduce 40 stitches. We uniformly divide the resource constraints supported by the stitches into 6 intervals and deploy one stitch within each interval. We set " + }, + { + "bbox": [ + 304, + 327, + 546, + 482 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 304, + 327, + 546, + 482 + ], + "type": "text", + "content": " in Eq. (3) as 0.9. The rank " + }, + { + "bbox": [ + 304, + 327, + 546, + 482 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 304, + 327, + 546, + 482 + ], + "type": "text", + "content": " is set to 16 and " + }, + { + "bbox": [ + 304, + 327, + 546, + 482 + ], + "type": "inline_equation", + "content": "d_1 // 4" + }, + { + "bbox": [ + 304, + 327, + 546, + 482 + ], + "type": "text", + "content": " for LoRA modules in self-attention and stitching layers, respectively. We employ Adam optimizer and fine-tune with batch size 128 for 10 epochs with gradient accumulation. During the generation stage, we employ top-" + }, + { + "bbox": [ + 304, + 327, + 546, + 482 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 304, + 327, + 546, + 482 + ], + "type": "text", + "content": " sampling to sample from the top subword candidates with the probability " + }, + { + "bbox": [ + 304, + 327, + 546, + 482 + ], + "type": "inline_equation", + "content": "p = 0.75" + }, + { + "bbox": [ + 304, + 327, + 546, + 482 + ], + "type": "text", + "content": ", conduct the beam search, and set the beam size as 4." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 483, + 546, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 483, + 546, + 651 + ], + "spans": [ + { + "bbox": [ + 304, + 483, + 546, + 651 + ], + "type": "text", + "content": "Results. We term the stitches obtained by our ESTA framework as Stitched LLaMA and compare them with Alpaca-LoRA 7B and 13B that respectively freeze the LLaMA 7B and 13B backbones and update the self-attention weights with LoRA modules [28]. For quantitative results reported in Table 2, we observe that similar to the visual recognition tasks, our Stitched LLaMA with different numbers of parameters generates answers with plausible quality in between Alpaca-LoRA 7B and 13B. Figure 5 gives qualitative comparisons. Our Stitched LLaMAs produce reasonable responses, suggesting the potential of our ESTA for adapting large language models (LLMs) for more downstream language tasks. More examples can be found in the supplementary material." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 659, + 400, + 673 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 659, + 400, + 673 + ], + "spans": [ + { + "bbox": [ + 306, + 659, + 400, + 673 + ], + "type": "text", + "content": "5.3. Ablation Study" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 305, + 677, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 677, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 677, + 545, + 713 + ], + "type": "text", + "content": "Effect of parameter-efficient stitch fine-tuning. We investigate the effectiveness of our PST on five FGVC tasks and visualize the averaged results in Figure 6 (a) and (b)." + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "28560" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 53, + 74, + 159, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 74, + 159, + 83 + ], + "spans": [ + { + "bbox": [ + 53, + 74, + 159, + 83 + ], + "type": "text", + "content": "Instruction: Tell me about alpacas." + } + ] + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 51, + 87, + 74, + 108 + ], + "blocks": [ + { + "bbox": [ + 51, + 87, + 74, + 108 + ], + "lines": [ + { + "bbox": [ + 51, + 87, + 74, + 108 + ], + "spans": [ + { + "bbox": [ + 51, + 87, + 74, + 108 + ], + "type": "image", + "image_path": "f04c44941f3c8d7da59b4f13ce2725b9edc715daee04f06880f6c899b845ccc5.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 75, + 91, + 249, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 91, + 249, + 108 + ], + "spans": [ + { + "bbox": [ + 75, + 91, + 249, + 108 + ], + "type": "text", + "content": "Alpaca-LoRA 7B (0.035 s/token): Alpacas are members of the camelid family, which also includes llamas and camels." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 52, + 109, + 251, + 169 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 52, + 109, + 251, + 169 + ], + "spans": [ + { + "bbox": [ + 52, + 109, + 251, + 169 + ], + "type": "text", + "content": "They are native to the Andes Mountains in South America, where they have been domesticated for thousands of years. Alpacas are known for their soft, luxurious fiber, which is used to make clothing, blankets, and other textiles. They are also raised for their meat, which is considered a delicacy in some parts of the world. Alpacas are gentle, intelligent animals, and they are prized for their docile nature and ability to thrive in a wide range of climates." + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 257, + 87, + 280, + 108 + ], + "blocks": [ + { + "bbox": [ + 257, + 87, + 280, + 108 + ], + "lines": [ + { + "bbox": [ + 257, + 87, + 280, + 108 + ], + "spans": [ + { + "bbox": [ + 257, + 87, + 280, + 108 + ], + "type": "image", + "image_path": "6ba465d987693d004666b0c529a660dad07275b93b25b12df950b2502ffcce14.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 282, + 91, + 537, + 169 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 282, + 91, + 537, + 169 + ], + "spans": [ + { + "bbox": [ + 282, + 91, + 537, + 169 + ], + "type": "text", + "content": "Stitched LLaMA 8B (0.037 s/token): Alpacas are herbivorous mammals that belong to the family Camelidae, which also includes camels and camelids. They are native to South American continent and are commonly found in Peru, Argentina, Bolivia, and Chile. They are smaller than their cousin, the camel, and are typically about 80-90 inches tall fully grown. They are known for their soft, shimmering, and light-weight fibers, which tend to make clothing, carpets, and other textile products. Alpacas are herbivores, and catets consists mainly of grasses, leaves, and straw. They are social animals and live in a few up to 100 alpacas, led by a dominant male. Alpacas are known for their calm and mature, and they are often used as livestock for their soft, shimmering fibers." + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 51, + 174, + 74, + 194 + ], + "blocks": [ + { + "bbox": [ + 51, + 174, + 74, + 194 + ], + "lines": [ + { + "bbox": [ + 51, + 174, + 74, + 194 + ], + "spans": [ + { + "bbox": [ + 51, + 174, + 74, + 194 + ], + "type": "image", + "image_path": "a8086e6c950405e98b75fa0f6c7d5934260f4851d4b3c8e5cf2c1905e96d242b.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 75, + 178, + 532, + 248 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 178, + 532, + 248 + ], + "spans": [ + { + "bbox": [ + 75, + 178, + 532, + 248 + ], + "type": "text", + "content": "Stitched LLaMA 11B (0.041 s/token): Alpacas are native to South America and are members of the camelid family. They are known for their soft, fluffy coats and their gentle nature. Alpacas are herbivores, meaning they feed on grass and other vegetation. They have a gestation period of approximately 11 days and usually give birth to a single baby, although twins and triplets are not uncommon. Alpacas are social animals and live in herds, led by a dominant they communicate with each other using a variety of vocalizations, including whistles, grunts, and barks. Alpacas are intelligent animals and are known for curiosity and playful nature. Alpacas have been domesticated for thousands of years and are used for a variety of purposes, including as pets, for their fiber, for their meat. Alpaca fiber is prized for its softness and warmth, making it a popular material for clothing and blankets. Alpaca meat is also considered a key in some parts of the world. Overall, alpacas are gentle, intelligent animals that have been domesticated for thousands of years. They are valued for their meat, and companionship." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 52, + 256, + 525, + 265 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 52, + 256, + 525, + 265 + ], + "spans": [ + { + "bbox": [ + 52, + 256, + 525, + 265 + ], + "type": "text", + "content": "Instruction: Can you help me write a resignation letter to my current employer, while leaving on good terms and expressing gratitude for the opportunities provided?" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 52, + 269, + 72, + 289 + ], + "blocks": [ + { + "bbox": [ + 52, + 269, + 72, + 289 + ], + "lines": [ + { + "bbox": [ + 52, + 269, + 72, + 289 + ], + "spans": [ + { + "bbox": [ + 52, + 269, + 72, + 289 + ], + "type": "image", + "image_path": "411f36edf167bbf47494194987b87d052b2e34c5c29b7e7aa860be3dbd9ee5f9.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 74, + 271, + 211, + 281 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 74, + 271, + 211, + 281 + ], + "spans": [ + { + "bbox": [ + 74, + 271, + 211, + 281 + ], + "type": "text", + "content": "Alpaca-LoRA 7B (0.035 s-token): Dear [Name]" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 52, + 281, + 215, + 333 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 52, + 281, + 215, + 333 + ], + "spans": [ + { + "bbox": [ + 52, + 281, + 215, + 333 + ], + "type": "text", + "content": "of Employer], I am writing to inform you that I am resigning from my position as [Job Title] with [Company Name]. It has been a pleasure working with you and I am grateful for the opportunities you have provided me. I wish you and the company all the best in the future. Sincerely, [Your Name]" + } + ] + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 229, + 269, + 252, + 289 + ], + "blocks": [ + { + "bbox": [ + 229, + 269, + 252, + 289 + ], + "lines": [ + { + "bbox": [ + 229, + 269, + 252, + 289 + ], + "spans": [ + { + "bbox": [ + 229, + 269, + 252, + 289 + ], + "type": "image", + "image_path": "5931474e9cf30c70a33e7ce73682c1bf912f17e153c4880262b3e013fa17f1d2.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "bbox": [ + 253, + 271, + 537, + 333 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 253, + 271, + 537, + 333 + ], + "spans": [ + { + "bbox": [ + 253, + 271, + 537, + 333 + ], + "type": "text", + "content": "Stitched LLaMA 8B (0.037 s/token): Dear [Employer], I am writing to inform you that I am resigning from my position at [Company Name] with a heavy heart. I have enjoyed my time with you and the team, and I am grateful for the opportunities that have been provided to me. I like to express my sincere appreciation for the training, support, and opportunities that have provided to me during my time at [Company Name]. I have learned a great deal and have grown really and professionally. Please accept my best wishes for the continued success of [Company and the team. I wish you all the best in the future. Sincerely, [Your Name]" + } + ] + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 51, + 337, + 72, + 358 + ], + "blocks": [ + { + "bbox": [ + 51, + 337, + 72, + 358 + ], + "lines": [ + { + "bbox": [ + 51, + 337, + 72, + 358 + ], + "spans": [ + { + "bbox": [ + 51, + 337, + 72, + 358 + ], + "type": "image", + "image_path": "d4f12cfe8a5f3a5d8c794e044a81734dca23d4486d52cdb0b046f85ecc524e72.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "bbox": [ + 74, + 342, + 539, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 74, + 342, + 539, + 396 + ], + "spans": [ + { + "bbox": [ + 74, + 342, + 539, + 396 + ], + "type": "text", + "content": "Sti 118 (0.041 s/ token): Dear [Employer], I am writing this letter to inform you of my decision to resign from my position at [Company]. It has been a great pleasure to work for [Company] and I have learned a lot during my time here. I would like to express my gratitude for the opportunities that I seen provided to me and for the support and guidance that I have received from my colleagues and supervisors. I have grown both professionally and finally during my time at [Company] and I am grateful for the experiences I have had. I am leaving on good terms and would be happy to provide a positive chance for [Company] in the future. If there is anything I can do to assist with the transition, please do not hesitate to reach out. Thank you for the opportunity to offer [Company] and I wish you all the best in your future endeavors. Sincerely, [Your Name]" + } + ] + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 48, + 433, + 175, + 530 + ], + "blocks": [ + { + "bbox": [ + 46, + 404, + 546, + 426 + ], + "lines": [ + { + "bbox": [ + 46, + 404, + 546, + 426 + ], + "spans": [ + { + "bbox": [ + 46, + 404, + 546, + 426 + ], + "type": "text", + "content": "Figure 5. Instruction-following comparison between Stitched LLaMA obtained by our ESTA and the Alpaca-LoRA 7B fine-tuned with LoRA [28]." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 48, + 433, + 175, + 530 + ], + "lines": [ + { + "bbox": [ + 48, + 433, + 175, + 530 + ], + "spans": [ + { + "bbox": [ + 48, + 433, + 175, + 530 + ], + "type": "image", + "image_path": "1783f314e4bf19760d94fd88dbe429f7b73f0bf028fd5c5f79aeaf4afe4b574b.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 173, + 434, + 297, + 531 + ], + "blocks": [ + { + "bbox": [ + 173, + 434, + 297, + 531 + ], + "lines": [ + { + "bbox": [ + 173, + 434, + 297, + 531 + ], + "spans": [ + { + "bbox": [ + 173, + 434, + 297, + 531 + ], + "type": "image", + "image_path": "8740576bf0900c0f058d9529d2da21aa3f2b5b820d1e3c70ad40df89d84429ba.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 45, + 533, + 546, + 622 + ], + "lines": [ + { + "bbox": [ + 45, + 533, + 546, + 622 + ], + "spans": [ + { + "bbox": [ + 45, + 533, + 546, + 622 + ], + "type": "text", + "content": "Figure 6. Ablation studies with results averaged over five FGVC tasks. (a) Effect of our stitch-agnostic LoRA modules. \"Full FT\" represents full fine-tuning of all the model parameters. \"PST w/ full SL\" employs fully fine-tuned stitching layers. We also show the number of trainable parameters. (b) Effect of our stitch-specific bias terms. \"SSB\" and \"scaled r\" represent stitch-specific bias terms and scaling the rank hyperparameter to reach 4.6M trainable parameters, respectively. (c) Effect of our task-specific stitch sampling. \"Uniform, all\" and \"Uniform, best\" represent all the stitches fine-tuned with uniform sampling and their best ones in each resource constraint interval on the Pareto frontier, respectively. (d) Effect of our strategy to simultaneously adapt and stitch anchors. Our strategy \"Adapt-and-stitch\" takes 100 epochs for fine-tuning. In contrast, \"Adapt-then-stitch\", as a straightforward approach to apply SN-Net, first individually adapts each anchor for a total of 300 epochs, and then fine-tunes SN-Net for another 100 epochs." + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 297, + 434, + 421, + 531 + ], + "blocks": [ + { + "bbox": [ + 297, + 434, + 421, + 531 + ], + "lines": [ + { + "bbox": [ + 297, + 434, + 421, + 531 + ], + "spans": [ + { + "bbox": [ + 297, + 434, + 421, + 531 + ], + "type": "image", + "image_path": "498598ad00fe8f261880fe20e83c7953d2cb06ed289b570d59ad9748e964b2bd.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 421, + 434, + 544, + 531 + ], + "blocks": [ + { + "bbox": [ + 421, + 434, + 544, + 531 + ], + "lines": [ + { + "bbox": [ + 421, + 434, + 544, + 531 + ], + "spans": [ + { + "bbox": [ + 421, + 434, + 544, + 531 + ], + "type": "image", + "image_path": "ce9fa356fb864e0fedd2391c575e610e23e69370aad23b33fd2095ad337a34c4.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + } + ], + "index": 20 + }, + { + "bbox": [ + 46, + 628, + 288, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 628, + 288, + 712 + ], + "spans": [ + { + "bbox": [ + 46, + 628, + 288, + 712 + ], + "type": "text", + "content": "To recall, our PST keeps the majority of the model parameters frozen while introducing trainable stitch-agonistic low-rank decomposition matrices and stitch-specific bias terms. In Figure 6 (a), we observe that when employing LoRA to the multi-head self-attention layers, there are solid performance gains on all FLOPs ranges with significantly fewer trainable parameters, echoing the observations in [31, 41]." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 304, + 628, + 547, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 628, + 547, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 628, + 547, + 712 + ], + "type": "text", + "content": "It is suggested that low-rank weight updates is a powerful alternative to full fine-tuning for stitching and adapting anchors and we speculate that low-rank weight update avoids overfitting under limited downstream data. We further employ low-rank weight updates to the stitching layers and save 2.4M more trainable parameters with a slight overall performance drop. We conjecture that the stitching layers" + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "28561" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 108 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 108 + ], + "type": "text", + "content": "for large pre-trained models are also in low intrinsic dimensions [2] and least-squares initialization [51] already provide a good initialization for them." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 110, + 286, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 110, + 286, + 205 + ], + "spans": [ + { + "bbox": [ + 46, + 110, + 286, + 205 + ], + "type": "text", + "content": "Figure 6 (b) shows that employing stitch-specific bias terms achieves better performance with affordable 1.4M extra trainable parameters. We also include the baseline that simply scales the low-rank dimension " + }, + { + "bbox": [ + 46, + 110, + 286, + 205 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 46, + 110, + 286, + 205 + ], + "type": "text", + "content": " to reach the same number of trainable parameters as our PST. This baseline is inferior to our PST. We speculate that the stitch-specific bias terms enable flexible adjustments of feature representations for different stitches, leading to enhanced performance." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 207, + 286, + 386 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 207, + 286, + 386 + ], + "spans": [ + { + "bbox": [ + 46, + 207, + 286, + 386 + ], + "type": "text", + "content": "Effect of task-specific stitch sampling. We investigate the effectiveness of our sampling strategy (introduced in Section 4.2) on the five FGVC tasks. The averaged results are visualized in Figure 6 (c). We compare our task-specific stitch sampling with the uniform one that is adopted in [51]. We empirically find that for uniform sampling, there is a clear performance gap between the best stitches that will be deployed and the others that will be dropped after training, suggesting the need for concentrating on the important stitches during training. To this end, our task-specific sampling strategy performs better than uniform sampling, indicating that our sampling strategy selects accurate task-specific important stitches. Most importantly, our strategy circumvents the need for the costly evaluation stage in the direct solution of adapting SN-Net for downstream tasks." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 388, + 286, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 388, + 286, + 567 + ], + "spans": [ + { + "bbox": [ + 46, + 388, + 286, + 567 + ], + "type": "text", + "content": "Effect of simultaneously adapting and stitching anchors. We study the effect of simultaneously adapting and stitching anchors strategy (introduced in Section 4.2) on five FGVC tasks. The averaged results are visualized in Figure 6 (d). We compare our approach with the direct SN-Net adaptation that first adapts each anchor individually for a total of 300 epochs before stitching them under the exact setting as ESTA. It can be seen that the direct SN-Net adaptation yields poor results in task adaptation. Fine-tuned individual anchors, especially the larger ones, are likely to overfit the limited downstream data [44, 53]. Accordingly, we speculate that the adapted anchor weights serve as a sub-optimal initialization for the stitches in the context of model stitching. On the other hand, our approach achieves better performance and greatly saves the fine-tuning cost." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 570, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 570, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 570, + 286, + 713 + ], + "type": "text", + "content": "Effect of the PEFT technique choices. We compare the effect of different PEFT technique choices and visualize the averaged results on the five FGVC tasks in Figure 7 left. We experiment to employ other PEFT techniques Adapter [27] and Adaptformer [9]. We follow their default setting to respectively insert bottleneck structures sequentially and in parallel to the feed-forward layers and fix the number of trainable parameters the same as ours for a fair comparison. We observe that the overall performance for different PEFT techniques is similar, suggesting that our ESTA framework is compatible with more PEFT technique choices. Since low-rank weight updates of LoRA [28] can be merged into" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 307, + 70, + 425, + 157 + ], + "blocks": [ + { + "bbox": [ + 307, + 70, + 425, + 157 + ], + "lines": [ + { + "bbox": [ + 307, + 70, + 425, + 157 + ], + "spans": [ + { + "bbox": [ + 307, + 70, + 425, + 157 + ], + "type": "image", + "image_path": "eb0215283bf4cc3a8215ec628dc353308860aebd9127fbbc2c15e565555c2864.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 166, + 545, + 221 + ], + "lines": [ + { + "bbox": [ + 304, + 166, + 545, + 221 + ], + "spans": [ + { + "bbox": [ + 304, + 166, + 545, + 221 + ], + "type": "text", + "content": "Figure 7. Left: effect of different PEFT technique choices. Right: effect of the rank hyper-parameter " + }, + { + "bbox": [ + 304, + 166, + 545, + 221 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 304, + 166, + 545, + 221 + ], + "type": "text", + "content": " in LoRA modules of self-attention layers, for which the three values represent the ranks in ViT-Ti, ViT-S, and ViT-B, respectively. The results are averaged on five FGVC tasks." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 425, + 70, + 545, + 157 + ], + "blocks": [ + { + "bbox": [ + 425, + 70, + 545, + 157 + ], + "lines": [ + { + "bbox": [ + 425, + 70, + 545, + 157 + ], + "spans": [ + { + "bbox": [ + 425, + 70, + 545, + 157 + ], + "type": "image", + "image_path": "f73035690bfa05c8498872e9d57e73f9fe6c8aa13500c1b2ab00781a7f797b43.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 237, + 545, + 261 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 237, + 545, + 261 + ], + "spans": [ + { + "bbox": [ + 304, + 237, + 545, + 261 + ], + "type": "text", + "content": "the backbone after training without extra inference computational cost, we employ LoRA by default." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 265, + 545, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 265, + 545, + 396 + ], + "spans": [ + { + "bbox": [ + 304, + 265, + 545, + 396 + ], + "type": "text", + "content": "Effect of the low-rank hyper-parameter " + }, + { + "bbox": [ + 304, + 265, + 545, + 396 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 304, + 265, + 545, + 396 + ], + "type": "text", + "content": ". We investigate the effect of different " + }, + { + "bbox": [ + 304, + 265, + 545, + 396 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 304, + 265, + 545, + 396 + ], + "type": "text", + "content": " which controls the rank when updating weights in self-attention layers on the five FGVC tasks. The averaged results on the five FGVC tasks are visualized in Figure 7 right. We find that setting " + }, + { + "bbox": [ + 304, + 265, + 545, + 396 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 304, + 265, + 545, + 396 + ], + "type": "text", + "content": " to be different values barely has any effect on the overall performance, suggesting our ESTA framework is robust to the rank hyper-parameter " + }, + { + "bbox": [ + 304, + 265, + 545, + 396 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 304, + 265, + 545, + 396 + ], + "type": "text", + "content": ". Since setting " + }, + { + "bbox": [ + 304, + 265, + 545, + 396 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 304, + 265, + 545, + 396 + ], + "type": "text", + "content": " to be 32, 16, and 8 for low-rank updating the self-attention weights in ViT-Ti, ViT-S, and ViT-B has slightly higher averaged performance among the stitches, we make it our default setting." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 422, + 470, + 434 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 422, + 470, + 434 + ], + "spans": [ + { + "bbox": [ + 306, + 422, + 470, + 434 + ], + "type": "text", + "content": "6. Conclusion and Future Work" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 445, + 545, + 649 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 445, + 545, + 649 + ], + "spans": [ + { + "bbox": [ + 304, + 445, + 545, + 649 + ], + "type": "text", + "content": "In this paper, we have introduced a novel task adaptation framework to cheaply obtain a palette of fine-tuned networks via model stitching, supporting diverse efficiency-performance tradeoffs at runtime. Specifically, built on SN-Net, we have tailored a parameter-efficient stitch finetuning method, which learns lightweight stitch-agnostic LoRA modules and stitch-specific bias terms while keeping the majority of the parameters frozen. Our design significantly reduces fine-tuning memory and storage costs for downstream task adaptation. Moreover, we have devised a task-specific stitch sampling strategy to assign higher sampling probability to the important stitches during training, which simultaneously improves the Pareto frontiers and avoids a costly evaluation stage. Extensive experiments on 25 downstream visual recognition tasks and the instruction-following task have demonstrated the effectiveness of our proposed framework." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 654, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 545, + 712 + ], + "type": "text", + "content": "Limitations. Due to the constraints of computational resources, our experiments are limited to visual recognition and instruction-tuning tasks. In the future, we will explore adapting pre-trained model families to dense prediction and multi-modal tasks." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "28562" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 124 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 124 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 124 + ], + "type": "text", + "content": "[1] Mohamed S. Abdelfattah, Abhinav Mehrotra, Lukasz Dudziak, and Nicholas D. Lane. Zero-cost proxies for lightweight nas. In ICLR, 2021. 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 124, + 287, + 157 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 124, + 287, + 157 + ], + "spans": [ + { + "bbox": [ + 53, + 124, + 287, + 157 + ], + "type": "text", + "content": "[2] Armen Aghajanyan, Luke Zettlemoyer, and Sonal Gupta. Intrinsic dimensionality explains the effectiveness of language model fine-tuning. In ACL, 2021. 4, 8" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 158, + 288, + 201 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 158, + 288, + 201 + ], + "spans": [ + { + "bbox": [ + 53, + 158, + 288, + 201 + ], + "type": "text", + "content": "[3] Haitham Bou Ammar, Eric Eaton, Matthew E Taylor, Decebal Constantin Mocanu, Kurt Driessens, Gerhard Weiss, and Karl Tuyls. An automated measure of mdp similarity for transfer in reinforcement learning. In AAAIw, 2014. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 202, + 287, + 235 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 202, + 287, + 235 + ], + "spans": [ + { + "bbox": [ + 53, + 202, + 287, + 235 + ], + "type": "text", + "content": "[4] Yamini Bansal, Preetum Nakkiran, and Boaz Barak. Revisiting model stitching to compare neural representations. NeurIPS, 34:225-236, 2021. 1, 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 236, + 286, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 236, + 286, + 258 + ], + "spans": [ + { + "bbox": [ + 53, + 236, + 286, + 258 + ], + "type": "text", + "content": "[5] Hangbo Bao, Li Dong, Songhao Piao, and Furu Wei. Beit: Bert pre-training of image transformers. In ICLR, 2022. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 258, + 286, + 291 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 258, + 286, + 291 + ], + "spans": [ + { + "bbox": [ + 53, + 258, + 286, + 291 + ], + "type": "text", + "content": "[6] Daniel Bolya, Rohit Mittapalli, and Judy Hoffman. Scalable diverse model selection for accessible transfer learning. NeurIPS, 34:19301-19312, 2021. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 292, + 286, + 345 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 292, + 286, + 345 + ], + "spans": [ + { + "bbox": [ + 53, + 292, + 286, + 345 + ], + "type": "text", + "content": "[7] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. *NeurIPS*, 33:1877–1901, 2020. 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 346, + 286, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 346, + 286, + 380 + ], + "spans": [ + { + "bbox": [ + 53, + 346, + 286, + 380 + ], + "type": "text", + "content": "[8] Han Cai, Chuang Gan, Tianzhe Wang, Zhekai Zhang, and Song Han. Once-for-all: Train one network and specialize it for efficient deployment. In ICLR, 2020. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 380, + 286, + 423 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 380, + 286, + 423 + ], + "spans": [ + { + "bbox": [ + 53, + 380, + 286, + 423 + ], + "type": "text", + "content": "[9] Shoufa Chen, Chongjian Ge, Zhan Tong, Jiangliu Wang, Yibing Song, Jue Wang, and Ping Luo. Adaptformer: Adapting vision transformers for scalable visual recognition. NeurIPS, 2022. 8" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 425, + 286, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 425, + 286, + 479 + ], + "spans": [ + { + "bbox": [ + 48, + 425, + 286, + 479 + ], + "type": "text", + "content": "[10] Wei-Lin Chiang, Zhuohan Li, Zi Lin, Ying Sheng, Zhang-hao Wu, Hao Zhang, Lianmin Zheng, Siyuan Zhuang, Yong-hao Zhuang, Joseph E. Gonzalez, Ion Stoica, and Eric P. Xing. Vicuna: An open-source chatbot impressing gpt-4 with " + }, + { + "bbox": [ + 48, + 425, + 286, + 479 + ], + "type": "inline_equation", + "content": "90\\%" + }, + { + "bbox": [ + 48, + 425, + 286, + 479 + ], + "type": "text", + "content": " * chatgpt quality, 2023. 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 480, + 286, + 513 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 480, + 286, + 513 + ], + "spans": [ + { + "bbox": [ + 48, + 480, + 286, + 513 + ], + "type": "text", + "content": "[11] Adam Coates and Andrew Y Ng. The importance of encoding versus training with sparse coding and vector quantization. In ICML, pages 921-928, 2011. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 514, + 286, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 514, + 286, + 557 + ], + "spans": [ + { + "bbox": [ + 48, + 514, + 286, + 557 + ], + "type": "text", + "content": "[12] Adrián Csiszárik, Péter Körösí-Szabó, Akos Matszangosz, Gergely Papp, and Daniel Varga. Similarity and matching of neural network representations. NeurIPS, 34:5656-5668, 2021. 1, 2, 4" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 558, + 286, + 580 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 558, + 286, + 580 + ], + "spans": [ + { + "bbox": [ + 48, + 558, + 286, + 580 + ], + "type": "text", + "content": "[13] Dong Dai and Tong Zhang. Greedy model averaging. NeurIPS, 24, 2011. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 581, + 286, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 581, + 286, + 634 + ], + "spans": [ + { + "bbox": [ + 48, + 581, + 286, + 634 + ], + "type": "text", + "content": "[14] Xiaoliang Dai, Alvin Wan, Peizhao Zhang, Bichen Wu, Zijian He, Zhen Wei, Kan Chen, Yuandong Tian, Matthew Yu, Peter Vajda, et al. Fbnetv3: Joint architecture-recipe search using predictor pretraining. In CVPR, pages 16276–16285, 2021. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 635, + 286, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 635, + 286, + 669 + ], + "spans": [ + { + "bbox": [ + 48, + 635, + 286, + 669 + ], + "type": "text", + "content": "[15] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR, pages 248-255, 2009. 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 48, + 670, + 286, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 670, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 670, + 286, + 713 + ], + "type": "text", + "content": "[16] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 35, + "blocks": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "type": "text", + "content": "worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2021. 2, 5" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 95, + 545, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 95, + 545, + 139 + ], + "spans": [ + { + "bbox": [ + 308, + 95, + 545, + 139 + ], + "type": "text", + "content": "[17] Eric Eaton, Marie Desjardins, and Terran Lane. Modeling transfer relationships between learning tasks for improved inductive transfer. In ECML PKDD, pages 317-332. Springer, 2008. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 140, + 545, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 140, + 545, + 172 + ], + "spans": [ + { + "bbox": [ + 308, + 140, + 545, + 172 + ], + "type": "text", + "content": "[18] Timnit Gebru, Jonathan Krause, Yilun Wang, Duyun Chen, Jia Deng, and Li Fei-Fei. Fine-grained car detection for visual census estimation. In AAAI, 2017. 4, 5, 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 173, + 545, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 173, + 545, + 215 + ], + "spans": [ + { + "bbox": [ + 308, + 173, + 545, + 215 + ], + "type": "text", + "content": "[19] Yunhui Guo, Honghui Shi, Abhishek Kumar, Kristen Grauman, Tajana Rosing, and Rogerio Feris. Spottune: transfer learning through adaptive fine-tuning. In CVPR, pages 4805-4814, 2019. 5" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 216, + 545, + 260 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 216, + 545, + 260 + ], + "spans": [ + { + "bbox": [ + 308, + 216, + 545, + 260 + ], + "type": "text", + "content": "[20] Zichao Guo, Xiangyu Zhang, Haoyuan Mu, Wen Heng, Zechun Liu, Yichen Wei, and Jian Sun. Single path one-shot neural architecture search with uniform sampling. In ECCV, pages 544-560. Springer, 2020. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 261, + 545, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 261, + 545, + 293 + ], + "spans": [ + { + "bbox": [ + 308, + 261, + 545, + 293 + ], + "type": "text", + "content": "[21] Song Han, Jeff Pool, John Tran, and William Dally. Learning both weights and connections for efficient neural network. NeurIPS, 28, 2015. 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 293, + 545, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 293, + 545, + 326 + ], + "spans": [ + { + "bbox": [ + 308, + 293, + 545, + 326 + ], + "type": "text", + "content": "[22] Haoyu He, Jianfei Cai, Jing Zhang, Dacheng Tao, and Bohan Zhuang. Sensitivity-aware visual parameter-efficient tuning. In ICCV, 2023. 5" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 327, + 545, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 327, + 545, + 360 + ], + "spans": [ + { + "bbox": [ + 308, + 327, + 545, + 360 + ], + "type": "text", + "content": "[23] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollar, and Ross Girshick. Masked autoencoders are scalable vision learners. In CVPR, pages 16000-16009, 2022. 1" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 361, + 545, + 404 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 361, + 545, + 404 + ], + "spans": [ + { + "bbox": [ + 308, + 361, + 545, + 404 + ], + "type": "text", + "content": "[24] Rudolf Herdt, Maximilian Schmidt, Daniel Otero Baguer, Jean Le'Clerc Arrastia, and Peter Maass. Model stitching and visualization how gan generators can invert networks in real-time. arXiv preprint arXiv:2302.02181, 2023. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 404, + 545, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 404, + 545, + 437 + ], + "spans": [ + { + "bbox": [ + 308, + 404, + 545, + 437 + ], + "type": "text", + "content": "[25] Adriano Hernandez, Rumen Dangovski, Peter Y. Lu, and Marin Soljacic. Model stitching: Looking for functional similarity between representations. NeurIPSW, 2022. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 438, + 545, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 438, + 545, + 459 + ], + "spans": [ + { + "bbox": [ + 308, + 438, + 545, + 459 + ], + "type": "text", + "content": "[26] Geoffrey Hinton, Oriol Vinyals, and Jeff Dean. Distilling the knowledge in a neural network. NeurIPSW, 2014. 5" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 460, + 545, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 460, + 545, + 514 + ], + "spans": [ + { + "bbox": [ + 308, + 460, + 545, + 514 + ], + "type": "text", + "content": "[27] Neil Houlsby, Andrei Giurgiu, Stanislaw Jastrzebski, Bruna Morrone, Quentin De Laroussilhe, Andrea Gesmundo, Mona Attariyan, and Sylvain Gelly. Parameter-efficient transfer learning for nlp. In ICML, pages 2790-2799, 2019. 1, 3, 8" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 308, + 514, + 545, + 558 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 514, + 545, + 558 + ], + "spans": [ + { + "bbox": [ + 308, + 514, + 545, + 558 + ], + "type": "text", + "content": "[28] Edward J Hu, yelong shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. LoRA: Low-rank adaptation of large language models. In ICLR, 2022. 1, 2, 3, 4, 6, 7, 8" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 308, + 559, + 545, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 559, + 545, + 602 + ], + "spans": [ + { + "bbox": [ + 308, + 559, + 545, + 602 + ], + "type": "text", + "content": "[29] Itay Hubara, Matthieu Courbariaux, Daniel Soudry, Ran El-Yaniv, and Yoshua Bengio. Quantized neural networks: Training neural networks with low precision weights and activations. JMLR, 18(1):6869-6898, 2017. 1" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 308, + 603, + 545, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 603, + 545, + 645 + ], + "spans": [ + { + "bbox": [ + 308, + 603, + 545, + 645 + ], + "type": "text", + "content": "[30] Pavel Izmailov, Dmitrii Podoprikhin, Timur Garipov, Dmitry Vetrov, and Andrew Gordon Wilson. Averaging weights leads to wider optima and better generalization. In UAI, 2018. 3" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 308, + 647, + 545, + 680 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 647, + 545, + 680 + ], + "spans": [ + { + "bbox": [ + 308, + 647, + 545, + 680 + ], + "type": "text", + "content": "[31] Menglin Jia, Luming Tang, Bor-Chun Chen, Claire Cardie, Serge Belongie, Bharath Hariharan, and Ser-Nam Lim. Visual prompt tuning. In ECCV, 2022. 1, 3, 5, 6, 7" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 308, + 681, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 681, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 681, + 545, + 713 + ], + "type": "text", + "content": "[32] Aditya Khosla, Nityananda Jayadevaprakash, Bangpeng Yao, and Li Fei-Fei. Novel dataset for fine-grained image categorization. In CVPRW, 2011. 5, 6" + } + ] + } + ], + "index": 34 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "28563" + } + ] + } + ], + "index": 36 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 105 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 105 + ], + "type": "text", + "content": "[33] Simon Kornblith, Jonathon Shlens, and Quoc V Le. Do better imagenet models transfer better? In CVPR, pages 2661-2671, 2019. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 106, + 287, + 129 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 106, + 287, + 129 + ], + "spans": [ + { + "bbox": [ + 49, + 106, + 287, + 129 + ], + "type": "text", + "content": "[34] Alex Krizhevsky. Learning multiple layers of features from tiny images. Technical report, 2009. 5, 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 129, + 288, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 129, + 288, + 162 + ], + "spans": [ + { + "bbox": [ + 48, + 129, + 288, + 162 + ], + "type": "text", + "content": "[35] Achintya Kundu, Laura Wynter, Rhui Dih Lee, and Luis Angel Bathen. Tofa: Transfer-once-for-all. arXiv preprint arXiv:2303.15485, 2023. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 163, + 288, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 163, + 288, + 185 + ], + "spans": [ + { + "bbox": [ + 48, + 163, + 288, + 185 + ], + "type": "text", + "content": "[36] Yann LeCun, John Denker, and Sara Solla. Optimal brain damage. NeurIPS, 2, 1989. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 186, + 288, + 218 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 186, + 288, + 218 + ], + "spans": [ + { + "bbox": [ + 48, + 186, + 288, + 218 + ], + "type": "text", + "content": "[37] Namhoon Lee, Thalaiyasingam Ajanthan, and Philip Torr. Snip: Single-shot network pruning based on connection sensitivity. In ICLR, 2019. 2, 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 220, + 288, + 253 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 220, + 288, + 253 + ], + "spans": [ + { + "bbox": [ + 48, + 220, + 288, + 253 + ], + "type": "text", + "content": "[38] Karel Lenc and Andrea Vedaldi. Understanding image representations by measuring their equivariance and equivalence. In CVPR, pages 991-999, 2015. 1, 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 254, + 288, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 254, + 288, + 285 + ], + "spans": [ + { + "bbox": [ + 48, + 254, + 288, + 285 + ], + "type": "text", + "content": "[39] Brian Lester, Rami Al-Rfou, and Noah Constant. The power of scale for parameter-efficient prompt tuning. In EMNLP, 2021. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 286, + 288, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 286, + 288, + 308 + ], + "spans": [ + { + "bbox": [ + 48, + 286, + 288, + 308 + ], + "type": "text", + "content": "[40] Xiang Lisa Li and Percy Liang. Prefix-tuning: Optimizing continuous prompts for generation. In ACL, 2021. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 309, + 288, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 309, + 288, + 342 + ], + "spans": [ + { + "bbox": [ + 48, + 309, + 288, + 342 + ], + "type": "text", + "content": "[41] Dongze Lian, Daquan Zhou, Jiashi Feng, and Xinchao Wang. Scaling & shifting your features: A new baseline for efficient model tuning. NeurIPS, 2022. 3, 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 343, + 288, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 343, + 288, + 365 + ], + "spans": [ + { + "bbox": [ + 48, + 343, + 288, + 365 + ], + "type": "text", + "content": "[42] Iou-Jen Liu, Jian Peng, and Alexander G Schwing. Knowledge flow: Improve upon your teachers. In ICLR, 2019. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 366, + 288, + 399 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 366, + 288, + 399 + ], + "spans": [ + { + "bbox": [ + 48, + 366, + 288, + 399 + ], + "type": "text", + "content": "[43] Jing Liu, Jianfei Cai, and Bohan Zhuang. Focusformer: Focusing on what we need via architecture sampler. arXiv preprint arXiv:2208.10861, 2022.5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 399, + 288, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 399, + 288, + 443 + ], + "spans": [ + { + "bbox": [ + 48, + 399, + 288, + 443 + ], + "type": "text", + "content": "[44] Pengfei Liu, Weizhe Yuan, Jinlan Fu, Zhengbao Jiang, Hiroaki Hayashi, and Graham Neubig. Pre-train, prompt, and predict: A systematic survey of prompting methods in natural language processing. ACM Comput. Surv., 2023. 8" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 444, + 288, + 477 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 444, + 288, + 477 + ], + "spans": [ + { + "bbox": [ + 48, + 444, + 288, + 477 + ], + "type": "text", + "content": "[45] Yen-Cheng Liu, Chih-Yao Ma, Junjiao Tian, Zijian He, and Zsolt Kira. Polyhistor: Parameter-efficient multi-task adaptation for dense vision tasks. NeurIPS, 2022. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 478, + 288, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 478, + 288, + 511 + ], + "spans": [ + { + "bbox": [ + 48, + 478, + 288, + 511 + ], + "type": "text", + "content": "[46] Zhuang Liu, Hanzi Mao, Chao-Yuan Wu, Christoph Feichtenhofer, Trevor Darrell, and Saining Xie. A convnet for the 2020s. In CVPR, pages 11976-11986, 2022. 6" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 512, + 288, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 512, + 288, + 544 + ], + "spans": [ + { + "bbox": [ + 48, + 512, + 288, + 544 + ], + "type": "text", + "content": "[47] Zhichao Lu, Gautam Sreekumar, Erik Goodman, Wolfgang Banzhaf, Kalyanmoy Deb, and Vishnu Naresh Boddeti. Neural architecture transfer. TPAMI, 43(9):2971-2989, 2021. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 545, + 288, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 545, + 288, + 578 + ], + "spans": [ + { + "bbox": [ + 48, + 545, + 288, + 578 + ], + "type": "text", + "content": "[48] Michael S Matena and Colin A Raffel. Merging models with fisher-weighted averaging. NeurIPS, 35:17703-17716, 2022. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 48, + 578, + 288, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 578, + 288, + 622 + ], + "spans": [ + { + "bbox": [ + 48, + 578, + 288, + 622 + ], + "type": "text", + "content": "[49] Cuong Nguyen, Tal Hassner, Matthias Seeger, and Cedric Archambeau. Leep: A new measure to evaluate transferability of learned representations. In ICML, pages 7294-7305. PMLR, 2020. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 48, + 624, + 288, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 624, + 288, + 657 + ], + "spans": [ + { + "bbox": [ + 48, + 624, + 288, + 657 + ], + "type": "text", + "content": "[50] Maria-Elena Nilsback and Andrew Zisserman. Automated flower classification over a large number of classes. In ICVGIP, pages 722-729. IEEE, 2008. 5" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 48, + 658, + 288, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 288, + 679 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 288, + 679 + ], + "type": "text", + "content": "[51] Zizheng Pan, Jianfei Cai, and Bohan Zhuang. Stretchable neural networks. In CVPR, 2023. 1, 2, 3, 4, 5, 6, 8" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 48, + 680, + 288, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 680, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 680, + 288, + 713 + ], + "type": "text", + "content": "[52] Hieu Pham, Melody Guan, Barret Zoph, Quoc Le, and Jeff Dean. Efficient neural architecture search via parameters sharing. In ICML, pages 4095-4104. PMLR, 2018. 6" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 36, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "text", + "content": "[53] Jo Pleted and Tom Gedeon. Deep transfer learning for image classification: a survey. arXiv preprint arXiv:2205.09904, 2022. 8" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 107, + 545, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 107, + 545, + 140 + ], + "spans": [ + { + "bbox": [ + 307, + 107, + 545, + 140 + ], + "type": "text", + "content": "[54] Ilija Radosavovic, Raj Prateek Kosaraju, Ross Girshick, Kaiming He, and Piotr Dólar. Designing network design spaces. In CVPR, pages 10428-10436, 2020. 5" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 141, + 545, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 141, + 545, + 184 + ], + "spans": [ + { + "bbox": [ + 307, + 141, + 545, + 184 + ], + "type": "text", + "content": "[55] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, pages 10684-10695, 2022. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 186, + 545, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 186, + 545, + 239 + ], + "spans": [ + { + "bbox": [ + 307, + 186, + 545, + 239 + ], + "type": "text", + "content": "[56] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, et al. Imagenet large scale visual recognition challenge. IJCV, 115:211-252, 2015. 1, 4" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 241, + 545, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 241, + 545, + 285 + ], + "spans": [ + { + "bbox": [ + 307, + 241, + 545, + 285 + ], + "type": "text", + "content": "[57] Andrei A Rusu, Neil C Rabinowitz, Guillaume Desjardins, Hubert Soyer, James Kirkpatrick, Koray Kavukcuoglu, Razvan Pascanu, and Raia Hadsell. Progressive neural networks. arXiv preprint arXiv:1606.04671, 2016. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 286, + 545, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 286, + 545, + 319 + ], + "spans": [ + { + "bbox": [ + 307, + 286, + 545, + 319 + ], + "type": "text", + "content": "[58] Yang Shu, Zhi Kou, Zhangjie Cao, Jianmin Wang, and Mingsheng Long. Zoo-tuning: Adaptive transfer from a zoo of models. In ICML, pages 9626-9637. PMLR, 2021. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 320, + 545, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 320, + 545, + 354 + ], + "spans": [ + { + "bbox": [ + 307, + 320, + 545, + 354 + ], + "type": "text", + "content": "[59] Yang Shu, Zhangjie Cao, Ziyang Zhang, Jianmin Wang, and Mingsheng Long. Hub-pathway: Transfer learning from a hub of pre-trained models. NeurIPS, 2022. 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 355, + 545, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 355, + 545, + 398 + ], + "spans": [ + { + "bbox": [ + 307, + 355, + 545, + 398 + ], + "type": "text", + "content": "[60] Andreas Steiner, Alexander Kolesnikov, Xiaohua Zhai, Ross Wightman, Jakob Uszkoreit, and Lucas Beyer. How to train your vit? data, augmentation, and regularization in vision transformers. TMLR, 2022. 5" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 399, + 545, + 432 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 399, + 545, + 432 + ], + "spans": [ + { + "bbox": [ + 307, + 399, + 545, + 432 + ], + "type": "text", + "content": "[61] Yi-Lin Sung, Jaemin Cho, and Mohit Bansal. Lst: Ladder side-tuning for parameter and memory efficient transfer learning. NeurIPS, 2022. 3" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 434, + 545, + 488 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 434, + 545, + 488 + ], + "spans": [ + { + "bbox": [ + 307, + 434, + 545, + 488 + ], + "type": "text", + "content": "[62] Rohan Taori, Ishaan Gulrajani, Tianyi Zhang, Yann Dubois, Xuechen Li, Carlos Guestrin, Percy Liang, and Tatsunori B. Hashimoto. Stanford alpaca: An instruction-following llama model. https://github.com/tatsu-lab/stanford_alpaca, 2023. 6" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 307, + 489, + 545, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 489, + 545, + 533 + ], + "spans": [ + { + "bbox": [ + 307, + 489, + 545, + 533 + ], + "type": "text", + "content": "[63] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Hervé Jégou. Training data-efficient image transformers & distillation through attention. In ICML, pages 10347-10357. PMLR, 2021. 5" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 307, + 534, + 545, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 534, + 545, + 590 + ], + "spans": [ + { + "bbox": [ + 307, + 534, + 545, + 590 + ], + "type": "text", + "content": "[64] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 1, 2, 6" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 307, + 590, + 545, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 590, + 545, + 645 + ], + "spans": [ + { + "bbox": [ + 307, + 590, + 545, + 645 + ], + "type": "text", + "content": "[65] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288, 2023. 2" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 307, + 647, + 545, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 647, + 545, + 679 + ], + "spans": [ + { + "bbox": [ + 307, + 647, + 545, + 679 + ], + "type": "text", + "content": "[66] Anh T Tran, Cuong V Nguyen, and Tal Hassner. Transferability and hardness of supervised classification tasks. In ICCV, pages 1395-1405, 2019. 3" + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 307, + 680, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 680, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 680, + 545, + 713 + ], + "type": "text", + "content": "[67] Grant Van Horn, Steve Branson, Ryan Farrell, Scott Haber, Jessie Barry, Panos Ipeirotis, Pietro Perona, and Serge Belongie. Building a bird recognition app and large scale" + } + ] + } + ], + "index": 35 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "28564" + } + ] + } + ], + "index": 37 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 66, + 72, + 287, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 72, + 287, + 95 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 287, + 95 + ], + "type": "text", + "content": "dataset with citizen scientists: The fine print in fine-grained dataset collection. In CVPR, pages 595-604, 2015. 5, 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 95, + 287, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 95, + 287, + 139 + ], + "spans": [ + { + "bbox": [ + 48, + 95, + 287, + 139 + ], + "type": "text", + "content": "[68] Catherine Wah, Steve Branson, Peter Welinder, Pietro Perona, and Serge Belongie. The caltech-ucsd birds-200-2011 dataset. Tech. Rep. CNS-TR-2011-001, California Institute of Technology, 2011. 5, 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 140, + 287, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 140, + 287, + 171 + ], + "spans": [ + { + "bbox": [ + 48, + 140, + 287, + 171 + ], + "type": "text", + "content": "[69] Chaoqi Wang, Guodong Zhang, and Roger Grosse. Picking winning tickets before training by preserving gradient flow. In ICLR, 2020. 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 172, + 287, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 172, + 287, + 205 + ], + "spans": [ + { + "bbox": [ + 48, + 172, + 287, + 205 + ], + "type": "text", + "content": "[70] Dilin Wang, Meng Li, Chengyue Gong, and Vikas Chandra. Attentivenas: Improving neural architecture search via attentive sampling. In CVPR, pages 6418-6427, 2021. 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 206, + 287, + 250 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 206, + 287, + 250 + ], + "spans": [ + { + "bbox": [ + 48, + 206, + 287, + 250 + ], + "type": "text", + "content": "[71] Yizhong Wang, Yeganeh Kordi, Swaroop Mishra, Alisa Liu, Noah A Smith, Daniel Khashabi, and Hannaneh Hajishirzi. Self-instruct: Aligning language model with self generated instructions. arXiv preprint arXiv:2212.10560, 2022. 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 251, + 287, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 251, + 287, + 304 + ], + "spans": [ + { + "bbox": [ + 48, + 251, + 287, + 304 + ], + "type": "text", + "content": "[72] Thomas Wolf, Lysandre Debut, Victor Sanh, Julien Chau-mond, Clement Delangue, Anthony Moi, Pierrick Cistac, Tim Rault, Rémi Louf, Morgan Funtowicz, et al. Transformers: State-of-the-art natural language processing. In EMNLP, pages 38-45, 2020. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 305, + 287, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 305, + 287, + 380 + ], + "spans": [ + { + "bbox": [ + 48, + 305, + 287, + 380 + ], + "type": "text", + "content": "[73] Mitchell Wortsman, Gabriel Ilharco, Samir Ya Gadre, Rebecca Roelofs, Raphael Gontijo-Lopes, Ari S Morcos, Hongseok Namkoong, Ali Farhadi, Yair Carmon, Simon Kornblith, et al. Model soups: averaging weights of multiple fine-tuned models improves accuracy without increasing inference time. In ICML, pages 23965-23998. PMLR, 2022. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 382, + 287, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 382, + 287, + 426 + ], + "spans": [ + { + "bbox": [ + 48, + 382, + 287, + 426 + ], + "type": "text", + "content": "[74] Hang Xu, Ning Kang, Gengwei Zhang, Chuanlong Xie, Xiaodan Liang, and Zhenguo Li. Nasoa: Towards faster task-oriented online fine-tuning with a zoo of models. In ICCV, pages 5097-5106, 2021. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 426, + 287, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 426, + 287, + 470 + ], + "spans": [ + { + "bbox": [ + 48, + 426, + 287, + 470 + ], + "type": "text", + "content": "[75] Runxin Xu, Fuli Luo, Zhiyuan Zhang, Chuanqi Tan, Baobao Chang, Songfang Huang, and Fei Huang. Raise a child in large language model: Towards effective and generalizable fine-tuning. In EMNLP, 2021. 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 471, + 287, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 471, + 287, + 503 + ], + "spans": [ + { + "bbox": [ + 48, + 471, + 287, + 503 + ], + "type": "text", + "content": "[76] Xingyi Yang, Daquan Zhou, Songhua Liu, Jingwen Ye, and Xinchao Wang. Deep model reassembly. NeurIPS, 35: 25739-25753, 2022. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 504, + 287, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 504, + 287, + 536 + ], + "spans": [ + { + "bbox": [ + 48, + 504, + 287, + 536 + ], + "type": "text", + "content": "[77] Jason Yosinski, Jeff Clune, Yoshua Bengio, and Hod Lipson. How transferable are features in deep neural networks? NeurIPS, 27, 2014. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 537, + 287, + 580 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 537, + 287, + 580 + ], + "spans": [ + { + "bbox": [ + 48, + 537, + 287, + 580 + ], + "type": "text", + "content": "[78] Kaichao You, Yong Liu, Jianmin Wang, and Mingsheng Long. Logme: Practical assessment of pre-trained models for transfer learning. In ICML, pages 12133-12143. PMLR, 2021. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 581, + 287, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 581, + 287, + 613 + ], + "spans": [ + { + "bbox": [ + 48, + 581, + 287, + 613 + ], + "type": "text", + "content": "[79] Jiahui Yu and Thomas S Huang. Universally slimmable networks and improved training techniques. In ICCV, pages 1803-1811, 2019. 5, 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 614, + 287, + 647 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 614, + 287, + 647 + ], + "spans": [ + { + "bbox": [ + 48, + 614, + 287, + 647 + ], + "type": "text", + "content": "[80] Elad Ben Zaken, Yoav Goldberg, and Shauli Ravfogel. Bitfit: Simple parameter-efficient fine-tuning for transformer-based masked language-models. In ACL, pages 1-9, 2022. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 647, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 647, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 647, + 287, + 712 + ], + "type": "text", + "content": "[81] Xiaohua Zhai, Joan Puigcerver, Alexander Kolesnikov, Pierre Ruyssen, Carlos Riquelme, Mario Lucic, Josip Djolonga, Andre Susano Pinto, Maxim Neumann, Alexey Dosovitskiy, et al. A large-scale study of representation learning with the visual task adaptation benchmark. arXiv preprint arXiv:1910.04867, 2019. 5, 6" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 149 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 95 + ], + "type": "text", + "content": "[82] Lvmin Zhang and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In ICCV, 2023. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 95, + 545, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 95, + 545, + 149 + ], + "spans": [ + { + "bbox": [ + 307, + 95, + 545, + 149 + ], + "type": "text", + "content": "[83] Renrui Zhang, Jiaming Han, Aojun Zhou, Xiangfei Hu, Shilin Yan, Pan Lu, Hongsheng Li, Peng Gao, and Yu Qiao. Llama-adapter: Efficient fine-tuning of language models with zero-init attention. arXiv preprint arXiv:2303.16199, 2023. 3" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "28565" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Efficient Test-Time Adaptation of Vision-Language Models/f86a8c24-7089-4003-88ca-fa1a5c70fe2d_content_list.json b/2024/Efficient Test-Time Adaptation of Vision-Language Models/f86a8c24-7089-4003-88ca-fa1a5c70fe2d_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..505f91bc54af8801fb251db9b463032cde87f17f --- /dev/null +++ b/2024/Efficient Test-Time Adaptation of Vision-Language Models/f86a8c24-7089-4003-88ca-fa1a5c70fe2d_content_list.json @@ -0,0 +1,1258 @@ +[ + { + "type": "text", + "text": "Efficient Test-Time Adaptation of Vision-Language Models", + "text_level": 1, + "bbox": [ + 187, + 130, + 782, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Adilbek Karmanov $^{1*}$ Dayan Guan $^{2*}$ Shijian Lu $^{1,2\\dagger}$ Abdulmotaleb El Saddik $^{1,3}$ Eric Xing $^{1,4}$", + "bbox": [ + 96, + 178, + 867, + 200 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Mohamed bin Zayed University of Artificial Intelligence $^{2}$ Nanyang Technological University $^{3}$ University of Ottawa $^{4}$ Carnegie Mellon University", + "bbox": [ + 104, + 200, + 864, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 271, + 313, + 287 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Test-time adaptation with pre-trained vision-language models has attracted increasing attention for tackling distribution shifts during the test time. Though prior studies have achieved very promising performance, they involve intensive computation which is severely unaligned with test-time adaptation. We design TDA, a training-free dynamic adapter that enables effective and efficient test-time adaptation with vision-language models. TDA works with a lightweight key-value cache that maintains a dynamic queue with few-shot pseudo labels as values and the corresponding test-sample features as keys. Leveraging the key-value cache, TDA allows adapting to test data gradually via progressive pseudo label refinement which is super-efficient without incurring any backpropagation. In addition, we introduce negative pseudo labeling that alleviates the adverse impact of pseudo label noises by assigning pseudo labels to certain negative classes when the model is uncertain about its pseudo label predictions. Extensive experiments over two benchmarks demonstrate TDA's superior effectiveness and efficiency as compared with the state-of-the-art. The code has been released in https://kdiaaa.github.io/tda/.", + "bbox": [ + 75, + 304, + 473, + 636 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 667, + 209, + 684 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recent advances in vision-language models [19, 31, 43] have opened a new door for integrating human language into various computer vision tasks. Take CLIP [31] as an example. It can enable zero-shot image classification by leveraging a shared embedding space that is learnt from web-scale image-text pairs. Within this shared space, images can be directly recognized by matching their features with the text embeddings of CLIP classes. At the other end, CLIP often faces challenges while handling various specific downstream images, especially when the downstream images have clear domain and distribution shifts as compared", + "bbox": [ + 75, + 695, + 468, + 863 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/1d0a13a9b4fa3f9ddd9c9eee1a89ba333532057ee01bec39e001bf87a863ba0b.jpg", + "image_caption": [ + "(a) Test-time Prompt Tuning [9, 35]" + ], + "image_footnote": [], + "bbox": [ + 501, + 270, + 890, + 429 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/b84cafaa5da2b0e13490f6052a78e44341a30fea0f1aa56e779523647d7ada74.jpg", + "image_caption": [ + "(b) Training-free Dynamic Adapter (Ours)", + "Figure 1. Comparison of our proposed Training-free Dynamic Adapter (TDA) with Test-time Prompt Tuning TPT [35] and its enhancement DiffTPT [9]: both TPT and DiffTPT require significant computational resources to optimize the learnable prompt via backpropagation; TDA is a dynamic cache that is training-free without any backpropagation, making it efficient for test-time adaptation in various real-world scenarios." + ], + "image_footnote": [], + "bbox": [ + 501, + 455, + 890, + 641 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "with the CLIP training images.", + "bbox": [ + 500, + 805, + 705, + 821 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Several recent studies [2, 3, 40] introduce a new paradigm called test-time adaptation for mitigating the domain shift. The idea of test-time adaptation is well aligned with real-world scenarios where a model needs to adapt to new environments quickly. Despite its great research and", + "bbox": [ + 496, + 824, + 893, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Equal Contribution.", + "bbox": [ + 94, + 875, + 209, + 887 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "† Corresponding Author.", + "bbox": [ + 96, + 887, + 225, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "14162", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "application values, test-time adaptation of vision-language models has been largely neglected. Recently, Test-time Prompt Tuning, as introduced in TPT [35] and its enhancement DiffTPT [9], attempts to adapt vision-language models by learning domain-specific prompts from test data. As illustrated in Fig. 1 (a), both TPT and DiffTPT train a learnable prompt for each test sample by feeding its augmented views into the CLIP model for generating predictions and minimizing the marginal entropy of confident predictions. Despite its decent performance, the prompt optimization in both TPT [50] and DiffTPT [9] is computationally intensive which hinders its applications in various real-world scenarios.", + "bbox": [ + 75, + 90, + 472, + 287 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We design a training-free dynamic adapter (TDA) that allows efficient and effective test-time adaptation of vision-language models without requiring any backpropagation in test time. As Fig. 1 (b) shows, TDA constructs a lightweight Dynamic Adapter that keeps a dynamic queue with the pseudo labels of a stream of test samples as values and the corresponding CLIP-extracted features as keys. TDA has two desirable features that make its test-time adaptation highly applicable in real-world scenarios. First, TDA is highly effective, as it improves the quality of pseudo labels via progressive incorporation of test predictions of lower entropy [11, 40]. Second, TDA is very efficient as the key-value cache is non-parametric and requires no backpropagation during testing. Beyond that, TDA cache is lightweight due to the few-shot setup and it can be computed with simple matrix multiplications [12, 21, 29, 47].", + "bbox": [ + 75, + 289, + 472, + 532 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Note that the performance of TDA depends heavily on the pseudo labels of unlabelled test samples which are often noisy with prediction errors. Inspired by the idea of negative learning [22, 23, 33], we introduce negative pseudo labeling to reduce the impact of noisy estimated labels. Traditional pseudo labeling methods identify the presence of particular classes in unlabeled data, which may result in erroneous pseudo labels being assigned when comparable high probabilities are observed. In contrast, our designed negative pseudo labeling determines the absence of certain classes and can provide more accurate pseudo labels as the probabilities of these complementary classes are very low. Concretely, we construct an additional TDA cache that stores negative pseudo labels to complement the positive TDA cache. By combining positive and negative caches, TDA is more tolerant to noisy pseudo labels and can better generalize to testing data. Extensive experiments over two widely adopted test-time adaptation benchmarks show that TDA outperforms the state-of-the-art by large margins while significantly reducing the testing time from over 12 hours to 16 minutes on the ImageNet dataset.", + "bbox": [ + 75, + 534, + 472, + 852 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In summary, the contributions of this work are threefold. First, we design a training-free dynamic adapter (TDA) that can achieve test-time adaptation of vision-language models", + "bbox": [ + 76, + 854, + 470, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "efficiently and effectively. To the best of our knowledge, this is the first work that investigates the efficiency issue of test-time adaptation of vision-language models. Second, we introduce negative pseudo labeling to alleviate the adverse impact of pseudo label noises which makes TDA more robust to pseudo label noises and generalizable to testing data. Third, we evaluate TDA extensively over two benchmarks, and experiments show that TDA achieves superior accuracy and efficiency compared with the state-of-the-art.", + "bbox": [ + 496, + 90, + 893, + 227 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 242, + 640, + 258 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Vision-language models [6, 17, 19, 31, 43] have demonstrated significant potential in learning semantic representations effectively by undergoing extensive training on image-text data. CLIP [31] stands out among these models for its ability to establish links between visual and textual representations, which enables it to achieve impressive zero-shot results on various downstream tasks. To enhance the transfer learning capability of the CLIP model in the downstream classification tasks, researchers have proposed integrating language prompt learners such as CoOp [50] and CoCoOp [49], as well as vision adapters such as CLIP-Adapter [10], Tip-Adapter [47], CaFo [48], TaskRes [44] and GraphAdapter [25]. Although these methods have shown considerable performance improvements, they typically require a large amount of training data in downstream tasks, making them less practical for real-world scenarios. On other hand, this work focuses on a new paradigm named test-time adaptation without accessing the original training data.", + "bbox": [ + 496, + 268, + 893, + 554 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Test-time adaptation refers to the process of adapting models to testing data that may have distributional differences from the training data. It is particularly beneficial for real-world applications that require models to be deployed in diverse environments, such as autonomous driving in various weather conditions, medical diagnosis in different hospitals, and etc. Several recent works utilize each batch of testing samples to update partial weights [18, 37, 38], normalization statistics [34], or a combination of both [40, 45]. To avoid updating models with multiple testing samples, MEMO [46] proposes enforcing the invariant predictions from different augmentations of each sample in the testing data stream. TPT [35] tackles the same challenge with vision-language models by fine-tuning a learnable prompt with each testing sample. DiffTPT [9] innovates test-time prompt tuning by leveraging pre-trained diffusion models to augment the diversity of test data samples used in TPT. Although TPT [35] and DiffTPT [9] are effective in addressing test-time adaptation of vision-language models, prompt learning is computationally expensive and time-consuming. This paper aims to mitigate the computational efficiency challenges of TPT and DiffTPT through the introduction of", + "bbox": [ + 496, + 568, + 893, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "14163", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "a cache model.", + "bbox": [ + 76, + 90, + 179, + 104 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Cache model benefits the adaptation techniques by providing efficient inference and non-parametric processing without requiring parameter updates. Unbounded Cache [12] and PSMM [27] have shown promising results in the text generation task by storing a large amount of the training dataset to capture long-term dependencies, however, this approach poses a challenge of memory efficiency during the test-time adaptation. A different technique [20] was proposed to mitigate this limitation by reducing the size of the cache memory through the search for application. Alternatively, Tip-Adapter [47] solves the cache memory problem by only storing few-shot samples per class to create a cache model for vision-language models. Inspired by this work, our dynamic adapter uses the same architecture for the test-time adaptation setting, where there is no access to the source data, and testing samples can only be accessed one by one. To address the lack of access to source data during testing, our adapter collects the most reliable test samples and their pseudo labels to form cache model.", + "bbox": [ + 76, + 118, + 472, + 405 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 76, + 417, + 169, + 434 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Preliminaries", + "text_level": 1, + "bbox": [ + 76, + 443, + 217, + 457 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "CLIP [31] is a vision-language model that performs a proxy task of predicting the correct pairings of text and image. Consider an $N$ -class classification problem where the CLIP's objective in the zero-shot setting is to match images with their most relevant textual descriptions using an image encoder $E_{v}$ and a text encoder $E_{t}$ . To obtain the textual descriptions, $N$ -class names are concatenated with handcrafted prompts and then mapped into the $c$ -channeled text embeddings $\\mathbf{W}_{c}$ using the text encoder $E_{t}$ .", + "bbox": [ + 76, + 465, + 468, + 603 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "TPT [35] focuses on test-time adaptation of CLIP. In TPT, a prompt tuning method is proposed to learn an adaptive prompt $\\mathbf{p}_c$ using individual test samples. A set of augmentation functions $\\mathcal{A}$ is used to generate $n$ randomly augmented views $\\tilde{x}_{\\mathrm{test}} = \\mathcal{A}_n(x_{\\mathrm{test}})$ of a test sample $x_{\\mathrm{test}}$ . The objective of TPT is to reduce variation in the model's predictions across different augmentations $\\tilde{f}_{\\mathrm{test}} = E_v(\\tilde{x}_{\\mathrm{test}})$ by minimizing the marginal entropy among the outputs of the augmented views. Furthermore, TPT also includes confidence selection to discard noisy augmentations that could result in ambiguous model predictions, as shown in Figure 1a. This is achieved by filtering out augmented views with high-entropy predictions as:", + "bbox": [ + 76, + 614, + 472, + 811 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nP _ {\\mathrm {T P T}} \\left(\\tilde {f} _ {\\text {t e s t}}\\right) = \\frac {1}{\\rho n} \\sum_ {i = 1} ^ {n} \\mathbb {1} \\left[ \\mathrm {H} \\left(\\tilde {f} _ {i} \\mathbf {p} _ {c} ^ {T}\\right) \\leq \\tau \\right] \\tilde {f} _ {i} \\mathbf {p} _ {c} ^ {T}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 119, + 821, + 468, + 861 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mathrm{H}$ is the self-entropy function of the softmax logits predictions, $\\tilde{f}_i\\mathbf{p}_c^T$ is the class probabilities vector of size", + "bbox": [ + 76, + 869, + 470, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "$N$ generated from the given $i$ -th augmented view of the test image, and the parameter $\\tau$ determines that only $\\rho$ -percentile of confident samples with entropy values below this threshold can be selected out of $n$ augmented views.", + "bbox": [ + 498, + 90, + 890, + 151 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Tip-Adapter [47] provides a training-free solution that uses a key-value cache model and integrates knowledge from the pre-trained CLIP model with few-shot labeled samples. The cache model is created using a set of $k$ -shot labeled samples $x_{k}$ from $N$ classes and their corresponding ground-truth labels $y_{N}$ . It can be conceptualized as a linear two-layer model, where the first layer contains train image features $\\mathbf{F}_{\\mathrm{train}} = E_v(x_k)$ and the second layer consists of one-hot vectors $\\mathbf{L}_{\\mathrm{train}}$ encoded from the labels $y_{N}$ . Given test image features $f_{\\mathrm{test}}$ generated from the CLIP's image encoder $E_{v}$ , the prediction from the cache model can be calculated as follows:", + "bbox": [ + 498, + 164, + 892, + 344 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nP _ {\\mathrm {c a c h e}} \\left(f _ {\\mathrm {t e s t}}\\right) = A \\left(f _ {\\mathrm {t e s t}} \\mathbf {F} _ {\\mathrm {t r a i n}} ^ {T}\\right) \\mathbf {L} _ {\\mathrm {t r a i n}}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 570, + 354, + 890, + 375 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $A(z) = \\alpha \\exp(-\\beta (1 - z))$ is an adaptation function within a weighting factor $\\alpha$ and a sharpness ratio $\\beta$ . During inference, the prediction of Tip-Adapter is computed by combining the pre-trained CLIP model and the cache model as: $P_{\\mathrm{TA}}(f_{\\mathrm{test}}) = P_{\\mathrm{cache}}(f_{\\mathrm{test}}) + f_{\\mathrm{test}} \\mathbf{W}_c^T$ .", + "bbox": [ + 498, + 383, + 890, + 462 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Training-free Dynamic Adapter", + "text_level": 1, + "bbox": [ + 500, + 469, + 779, + 486 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "During testing, pre-trained vision-language models like CLIP may encounter distribution shifts that degrade the classification performance. To address this issue, existing test-time prompt tuning methods train a learnable prompt by enforcing consistency across different image augmentations during testing. However, it requires a large number of augmentation operations on each test image and computationally-heavy optimization steps to learn the prompt, limiting its applicability in various real-world settings.", + "bbox": [ + 496, + 492, + 890, + 643 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this paper, our motivation is to design an efficient method for test-time adaptation of the pre-trained vision-language models like CLIP. Inspired by the concept of TipAdapter, we propose a training-free dynamic adapter (TDA) to enable efficient and effective test-time adaptation with CLIP. As shown in Figure 2, TDA includes two lightweight key-value caches, where each cache stores a dynamic queue of few-shot test features as keys and the corresponding pseudo labels as values. The first cache is intended for positive learning and it dynamically updates key-value pairs with high-confidence predictions to improve the accuracy. The second cache is designed for negative learning and it aims to address the adverse effects of noisy pseudo labels by introducing negative pseudo labeling to identify class absence rather than presence. By combining the positive cache and negative cache, the proposed TDA can achieve superior performance in terms of both speed and accuracy.", + "bbox": [ + 496, + 643, + 892, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "14164", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/c29d8b07ccb6929f5a98a2ce188137beb7875f7b6c69de9ed5f29a2420f6b6b4.jpg", + "image_caption": [ + "Figure 2. Overview of the proposed Training-free Dynamic Adapter (TDA). TDA constructs and updates two key-value caches to store the knowledge of a stream of test samples, and uses the two caches to generate positive and negative predictions which are combined with CLIP predictions to produce the final prediction. Specifically, the CLIP predictions are generated by performing the dot product between the image features generated by CLIP's image encoder $E_{v}$ and the text embeddings generated by CLIP's text encoder $E_{t}$ , using the hand-crafted prompt and class names. The two key-value caches are updated by gradually incorporating the test features and their corresponding pseudo labels calculated from CLIP's predictions, based on prediction entropy and cache capacity." + ], + "image_footnote": [], + "bbox": [ + 83, + 85, + 893, + 315 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our goal is to conduct test-time adaptation by gathering adequate knowledge from the testing data stream and improving the predictions through the adaptation of the image features. To accomplish this goal, we create positive and negative caches that capture particular characteristics of the testing data stream. In the upcoming parts, we will present the process of collecting data for each cache and define the conditions in which we can utilize cache information to adapt features and enhance model predictions.", + "bbox": [ + 75, + 440, + 472, + 577 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Positive Cache. The positive cache in TDA is a key-value cache, in which keys and values are represented as a dynamic queue. It aims to collect high-quality few-shot pseudo labels $\\hat{\\mathbf{L}}_{\\mathfrak{p}}$ as positive values and the corresponding features $\\mathbf{Q}_{\\mathfrak{p}}$ as keys. The key-value cache is initially empty and then accumulates a sufficient number of key-value pairs during the test-time adaptation. To maintain high-quality pseudo labels, TDA progressively incorporates test predictions with lower entropy while limiting shot capacity1 in the positive cache. Different from a normal queue like a FIFO queue with a fixed size, the dynamic queue in our method expands in size during testing. Besides, the dynamic queue operates similarly to a priority queue, using entropy as the criterion for prioritization. Note that, each class has its own queue to maintain the order and correct data structure of each class in the cache.", + "bbox": [ + 75, + 599, + 468, + 840 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Given a pre-trained CLIP model that consists of a text encoder $E_{t}$ and an image encoder $E_{v}$ , $E_{t}$ processes class", + "bbox": [ + 76, + 842, + 468, + 875 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "names with pre-defined prompts to generate $c$ -channeled text embeddings $\\mathbf{W}_c$ and $E_v$ processes each test image $x_{\\mathrm{test}}$ to produce image features $f_{\\mathrm{test}}$ . To build the positive cache, TDA first generates a pseudo label $\\hat{l}$ , a one-hot encoded vector of a categorical distribution, for each test sample $x_{\\mathrm{test}}$ by applying the softmax function to the prediction $f_{\\mathrm{test}} \\mathbf{W}_c^T$ . We establish two conditions to ascertain whether and how to include the pseudo label $\\hat{l}$ and its corresponding image features $f_{\\mathrm{test}}$ into the positive cache. The first condition is defined as: if the shot number (the number of collected pairs per class) of $\\hat{\\mathbf{L}}_{\\mathbf{p}}$ is less than the maximum shot capacity (the maximum capacity number of pairs per class) $k$ , TDA will add $\\hat{l}$ and $f_{\\mathrm{test}}$ as a new value and a new key to $\\hat{\\mathbf{L}}_{\\mathbf{p}}$ and $\\mathbf{Q}_{\\mathbf{p}}$ respectively. Meanwhile, the second condition is defined as: if the shot number of $\\hat{\\mathbf{L}}_{\\mathbf{p}}$ has reached the maximum shot capacity $k$ , TDA will replace an 'uncertain' key-value pair $\\{\\mathbf{q}^{ent}, \\hat{l}_p^{ent}\\}$ with $\\{f_{\\mathrm{test}}, \\hat{l}\\}$ when $\\mathrm{H}(f_{\\mathrm{test}} \\mathbf{W}_c^T) < \\mathrm{H}(\\mathbf{q}^{ent} \\mathbf{W}_c^T)$ . Here, $\\mathrm{H}$ denotes the entropy function and the term 'uncertain' indicates that the entropy of a particular key-value pair is the highest compared to the entropy of all other key-value pairs of the same class in the cache model. By applying these two conditions, TDA can gradually integrate test predictions with lower entropy while controlling the shot capacity, which helps to ensure the collection of high-quality pseudo labels in positive cache.", + "bbox": [ + 500, + 440, + 890, + 835 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "During test-time adaptation, the positive cache can quickly retrieve relevant information by treating the image features $f_{\\mathrm{test}}$ generated from a test example as a query and searching the stored key-value pairs $\\{\\mathbf{Q}_{\\mathbf{p}},\\hat{\\mathbf{L}}_{\\mathbf{p}}\\}$ for match-", + "bbox": [ + 498, + 839, + 893, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_footnote", + "text": "1The shot capacity refers to the maximum number of pairs per class.", + "bbox": [ + 94, + 886, + 457, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "14165", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "ing information. The adapted predictions using the positive cache can then be obtained as:", + "bbox": [ + 75, + 90, + 468, + 119 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nP _ {\\mathrm {p o s}} \\left(f _ {\\mathrm {t e s t}}\\right) = A \\left(f _ {\\mathrm {t e s t}} \\mathbf {Q} _ {\\mathbf {p}} ^ {\\mathbf {T}}\\right) \\hat {\\mathbf {L}} _ {\\mathbf {p}}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 169, + 127, + 468, + 147 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $A$ is the adaptation function defined in Tip-Adapter.", + "bbox": [ + 76, + 152, + 464, + 167 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Negative Cache. Similar to the positive cache in our TDA, the negative cache is also a dynamic queue structure with negative keys and negative values denoted as $\\mathbf{Q}_{\\mathbf{n}}$ and $\\hat{\\mathbf{L}}_{\\mathbf{n}}$ , respectively. It aims to gather CLIP-generated image features to $\\mathbf{Q}_{\\mathbf{n}}$ and the corresponding negative pseudo labels to $\\hat{\\mathbf{L}}_{\\mathbf{n}}$ . Unlike the pseudo labels in the positive cache, the negative pseudo labels are obtained by applying negative mask on the class probabilities as:", + "bbox": [ + 75, + 184, + 468, + 305 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathbf {L}} _ {\\mathbf {n}} = - \\mathbb {1} [ p _ {l} < P (\\mathbf {Q} _ {\\mathbf {n}}) ], \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 189, + 310, + 468, + 330 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where higher probabilities than $p_l$ are selected as negative pseudo labels from uncertain predictions and the uncertainty is measured by the entropy of predictions. Here, $p_l$ represents a threshold in negative pseudo labeling, and $\\hat{\\mathbf{L}}_{\\mathbf{n}}$ denotes a negative pseudo label, which is a vector whose elements larger than $p_l$ have a value -1 and otherwise 0. Different from existing negative learning methods [22, 23, 33] that select negative labels from all noisy labels, TDA selects negative pseudo labels from uncertain predictions to avoid bias to the data with certain predictions.", + "bbox": [ + 75, + 335, + 468, + 487 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "When constructing the negative cache, the testing feature $f_{\\mathrm{test}}$ will be included in negative cache if it satisfies the condition $\\gamma (f_{\\mathrm{test}})$ : the entropy of the prediction is in the specified interval between $\\tau_{l}$ and $\\tau_{h}$ :", + "bbox": [ + 76, + 487, + 468, + 547 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\gamma \\left(f _ {\\text {t e s t}}\\right): \\tau_ {l} < \\mathrm {H} \\left(f _ {\\text {t e s t}} \\mathbf {W} _ {c} ^ {T}\\right) < \\tau_ {h}. \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 156, + 554, + 468, + 571 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "This condition is designed to mitigate the risk of prediction errors due to high entropy or be biased to certain predictions (characterized by very low entropy), by incorporating test samples that exhibit a moderate degree of prediction uncertainty. Once the $\\gamma(f_{\\mathrm{test}})$ check is completed, the remaining steps for collecting uncertain samples in the negative cache follow the same two conditions designed in the positive cache. Similar to the positive cache, TDA also limits the shot capacity $\\tilde{k}$ in the negative cache.", + "bbox": [ + 75, + 578, + 468, + 714 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "During test-time adaption, the testing features $f_{\\mathrm{test}}$ can be quickly adapted to target domains by retrieving the knowledge from $\\{\\mathbf{Q}_{\\mathbf{n}}, \\hat{\\mathbf{L}}_{\\mathbf{n}}\\}$ in the negative cache and the adapted prediction can be obtained as:", + "bbox": [ + 75, + 714, + 468, + 773 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nP _ {\\text {n e g}} \\left(f _ {\\text {t e s t}}\\right) = - A \\left(f _ {\\text {t e s t}} \\mathbf {Q} _ {\\mathbf {n}} ^ {\\mathbf {T}}\\right) \\hat {\\mathbf {L}} _ {\\mathbf {n}}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 163, + 780, + 468, + 799 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $A$ is the adaptation function defined in Tip-Adapter. The predictions of TDA can be formulated by combining the negative cache, the positive cache and the pre-trained CLIP model together as follows:", + "bbox": [ + 75, + 805, + 468, + 864 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nP _ {\\mathrm {T D A}} \\left(f _ {\\text {t e s t}}\\right) = f _ {\\text {t e s t}} \\mathbf {W} _ {c} ^ {T} + P _ {\\text {p o s}} \\left(f _ {\\text {t e s t}}\\right) + P _ {\\text {n e g}} \\left(f _ {\\text {t e s t}}\\right). \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 84, + 872, + 468, + 891 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Relationship with TPT and Tip-Adapter", + "text_level": 1, + "bbox": [ + 500, + 90, + 844, + 107 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Both TPT and TDA are designed to handle the challenge of adapting models to test data that have distributional discrepancies from the training data. TPT trains a learnable prompt $\\mathbf{p}_c$ in Eq. (1) for each test sample with a large number of augmentations and such training process requires backpropagation and is computationally intensive. In contrast, our TDA is training-free as both positive cache $\\{\\mathbf{Q_p},\\hat{\\mathbf{L}}_{\\mathbf{p}}\\}$ and negative cache $\\{\\mathbf{Q_n},\\hat{\\mathbf{L}}_{\\mathbf{n}}\\}$ are non-parametric, which is super-efficient without incurring any backpropagation.", + "bbox": [ + 496, + 119, + 890, + 258 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our TDA employs a cache model that shares similarities with the Tip-Adapter, where features and labels are stored as key-value pairs in a memory cache. One notable distinction between the Tip-Adapter and TDA lies in the type of cache used. Specifically, the Tip-Adapter relies on a static cache $\\{\\mathbf{F}_{\\mathrm{train}},\\mathbf{L}_{\\mathrm{train}}\\}$ because it is designed for a supervised adaptation setting, where the ground-truth labels $\\mathbf{L}_{\\mathrm{train}}$ are predetermined and readily available. In contrast, our TDA introduces a new dynamic cache $\\{\\mathbf{Q_p},\\hat{\\mathbf{L}}_p\\}$ designed for a test-time adaptation setting, where the pseudo labels $\\hat{\\mathbf{L}}_{\\mathbf{p}}$ are generated on-the-fly from a stream of test samples. Furthermore, TDA incorporates a novel negative cache model $\\{\\mathbf{Q_n},\\hat{\\mathbf{L}}_n\\}$ that enhances testing predictions by utilizing indirect knowledge that a test image does not belong to certain negative classes. By combining positive and negative caches, our proposed TDA is more robust to noisy pseudo labels and can generalize well to testing data.", + "bbox": [ + 496, + 263, + 890, + 521 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 500, + 553, + 632, + 570 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Experimental Setup", + "text_level": 1, + "bbox": [ + 500, + 584, + 689, + 599 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Benchmarks. We conducted main experiments on two benchmarks: out-of-distribution (OOD) benchmark and cross-domain benchmark, both applied in the previous work [35] that adapts vision-language models in test time. The OOD benchmark serves as a measure of the robustness of our approach by involving assessment on 4 out-of-distribution datasets derived from ImageNet [5]: ImageNet-A [16], ImageNet-V2 [32], ImageNet-R [15], and ImageNet-S [41]. This benchmark is specifically designed to evaluate a model's capacity to generalize to new and unseen data. The cross-domain benchmark, on the other hand, is involved to evaluate the model's performance across 10 diverse image classification datasets, each from a distinct domain with different classes: Aircraft [26], Caltech101 [8], Cars [24], DTD [4], EuroSAT [14], Flower102 [28], Food101 [1], Pets [30], SUN397 [42], and UCF101 [36]. This benchmark provides a comprehensive evaluation of the model's adaptability during test time across various class spaces.", + "bbox": [ + 496, + 613, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "14166", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/5c6859d6292bfbc1c78f9c699710b10037fe858960d85f6f24cb7861db37ee53.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodImageNetImageNet-AImageNet-V2ImageNet-RImageNet-SAverageOOD Average
CLIP-ResNet-5059.8123.2452.9160.7235.4846.4343.09
CoOp63.3323.0655.4056.6034.6746.6142.43
CoCoOp62.8123.3255.7257.7434.4846.8142.82
Tip-Adapter62.0323.1353.9760.3535.7447.0443.30
TPT60.7426.6754.7059.1135.0947.2643.89
DiffTPT60.8031.0655.8058.8037.1048.7145.69
TDA (Ours)61.3530.2955.5462.5838.1249.5846.63
CLIP-ViT-B/1668.3449.8961.8877.6548.2461.2059.42
CoOp71.5149.7164.2075.2147.9961.7259.28
CoCoOp71.0250.6364.0776.1848.7562.1359.91
Tip-Adapter70.7551.0463.4177.7648.8862.3760.27
TPT68.9854.7763.4577.0647.9462.4460.81
DiffTPT70.3055.6865.1075.0046.8062.2860.52
TDA (Ours)69.5160.1164.6780.2450.5465.0163.89
", + "bbox": [ + 81, + 88, + 888, + 340 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/b82f94048bf44351ef676b5370ed3b21bcbcd8bf85bb3a3513f45de9f9dfeeef.jpg", + "table_caption": [ + "Table 1. Results on the OOD Benchmark. Our TDA is compared with several state-of-the-art methods designed for vision-language models: the baseline method CLIP, three train-time adaptation methods (i.e., CoOp, CoCoOp, and Tip-Adapter), and two test-time adaptation methods (i.e., TPT and DiffTPT). All the compared methods are built upon CLIP-ResNet-50 or CLIP-ViT-B/16 baselines. The two evaluation metrics Average and OOD Average are calculated by taking the mean accuracy across all five datasets and four OOD datasets excluding ImageNet. The results of CLIP, CoOp, CoCoOp, and TPT are obtained from the TPT paper, the results of DiffTPT are obtained from the DiffTPT paper, while the results of Tip-Adapter are reproduced using the official codes." + ], + "table_footnote": [], + "table_body": "
MethodTesting TimeAccuracyGain
CLIP-ResNet-5012min59.810
TPT12h 50min60.74+0.93
DiffTPT34h 45min60.80+0.99
TDA (Ours)16min61.35+1.54
", + "bbox": [ + 84, + 460, + 460, + 551 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 2. Comparisons of our TDA with CLIP-ResNet-50, TPT, and DiffTPT in terms of efficiency (Testing Time) and effectiveness (Accuracy). The final column shows the accuracy gain relative to the baseline CLIP. Note that the testing time of DiffTPT does not include the duration required for the image generation process with pre-trained diffusion models, which is an additional time-consuming factor during the testing phase.", + "bbox": [ + 75, + 561, + 468, + 659 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation details. All the models in our experiments are built upon the pre-trained CLIP model [31] that consists of an image encoder and a text encoder. The image encoder can be either ResNet [13] or Vision Transformer [7], while the text encoder is Transformer [39]. Test-time adaptation is set for single-image scenarios, using a batch size of 1. We conduct a search for all our hyperparameters using a single ImageNet validation set. The threshold $p_l$ for negative pseudo-labeling in Eq 4 is set as 0.03. The upper and lower thresholds $[\\tau_l, \\tau_h]$ for testing feature selection in Eq 5 are set as [0.2, 0.5]. Once searched, these hyperparameters are fixed and evaluated across various new datasets. To avoid incurring backpropagation when using learnable prompt, we follow [31] to use hand-crafted", + "bbox": [ + 75, + 688, + 470, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "prompts. We use top-1 accuracy $(\\%)$ , a standard classification criterion, as our evaluation metric. All the experiments are conducted using a single NVIDIA Quadro RTX 6000 GPU.", + "bbox": [ + 496, + 462, + 890, + 521 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Comparisons with State-of-the-art", + "text_level": 1, + "bbox": [ + 498, + 530, + 802, + 546 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this section, we compare our proposed TDA with several state-of-the-art methods, including CLIP [31], three train-time adaptation methods, i.e., CoOp [50], CoCoOp [49], and Tip-Adapter [47], as well as two existing test-time adaptation methods TPT [35] and DiffTPT [9], all of which are designed for vision-language models. Specifically, CLIP is evaluated using an ensemble of 80 handcrafted prompts as in [31]. All train-time adaptation methods are trained on the ImageNet train set with 16 shots per class and tested on other datasets as in [35]. CoOp utilizes a learnable context module of size 4 that is fine-tuned for specific downstream tasks, while CoCoOp is the generalized version of CoOp with the input-conditional token for image features. Tip-Adapter employs the optimal hyperparameters obtained from the ImageNet validation set during evaluation. We would like to note that Tip-Adapter is unable to handle new classes during testing, limiting its implementation to OOD benchmark evaluation where the training classes encompass all the testing classes. Different from train-time adaptation methods, test-time adaptation methods (i.e., TPT, DiffTPT, and our TDA) do not utilize the ImageNet train set. Instead, they are fine-tuned with target datasets using a stream of unlabeled test samples. Follow", + "bbox": [ + 496, + 553, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "14167", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/b579dcdc8cca83c65012405cf7013f9f57878ef8dd7b9a145bf00fb726155893.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodAircraftCaltech101CarsDTDEuroSATFlower102Food101PetsSUN397UCF101Average
CLIP-ResNet-5016.1187.2655.8940.3725.7962.7774.8282.9760.8559.4856.63
CoOp15.1286.5355.3237.2926.2061.5575.5987.0058.1559.0556.18
CoCoOp14.6187.3856.2238.5328.7365.5776.2088.3959.6157.1057.23
TPT17.5887.0258.4640.8428.3362.6974.8884.4961.4660.8257.66
DiffTPT17.6086.8960.7140.7241.0463.5379.2183.4062.7262.6759.85
TDA (Ours)17.6189.7057.7843.7442.1168.7477.7586.1862.5364.1861.03
CLIP-ViT-B/1623.2293.5566.1145.0450.4266.9982.8686.9265.6365.1664.59
CoOp18.4793.7064.5141.9246.3968.7185.3089.1464.1566.5563.88
CoCoOp22.2993.7964.9045.4539.2370.8583.9790.4666.8968.4464.63
TPT24.7894.1666.8747.7542.4468.9884.6787.7965.5068.0465.10
DiffTPT25.6092.4967.0147.0043.1370.1087.2388.2265.7462.6765.47
TDA (Ours)23.9194.2467.2847.4058.0071.4286.1488.6367.6270.6667.53
", + "bbox": [ + 81, + 88, + 888, + 292 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 3. Results on the Cross-Domain Benchmark. Our TDA is compared with several state-of-the-art methods designed for vision-language models: the baseline method CLIP, two train-time adaptation methods (i.e., CoOp and CoCoOp), and two test-time adaptation methods (i.e., TPT and DiffTPT). Note that Tip-Adapter is unable to be evaluated on the Cross-Domain Benchmark as it cannot handle new classes during testing. The evaluation metric Average is calculated by taking the mean accuracy across all ten datasets. The results of CLIP, CoOp, CoCoOp, and TPT are obtained from the TPT paper, while the results of DiffTPT are obtained from the DiffTPT paper.", + "bbox": [ + 78, + 301, + 892, + 373 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "ing TPT and DiffTPT, we compare TDA with the state-of-the-art over two public benchmarks: OOD benchmark and cross-domain benchmark.", + "bbox": [ + 80, + 411, + 467, + 455 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Results on the OOD Benchmark. We first compare TDA with state-of-the-art methods over the OOD benchmark. Table 1 presents the experimental results, highlighting the superior performance of the proposed TDA compared to both TPT and DiffTPT across various OOD datasets derived from ImageNet. Specifically, TDA outperforms TPT on both ResNet-50 and ViT-B/16 architectures, improving OOD accuracy by $2.74\\%$ and $3.08\\%$ on average, respectively. Furthermore, compared to DiffTPT, TDA exhibits an average accuracy improvement of $0.94\\%$ and $3.37\\%$ in the OOD benchmark for ResNet-50 and ViT-B/16, respectively. These results validate the effectiveness of TDA in enhancing test-time adaptation performance on various OOD test datasets.", + "bbox": [ + 80, + 477, + 467, + 686 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In order to provide a more comprehensive evaluation of our proposed method's efficiency and effectiveness, we compared it with the baseline CLIP-ResNet-50 and two existing test-time adaptation methods (i.e., TPT and DiffTPT). This comparison encompasses both testing time and testing accuracy, and the corresponding results are shown in Table 2. This evaluation is performed on the ImageNet validation dataset, which consists of 50,000 images, using a single NVIDIA Quadro RTX 6000 GPU. When compared to the baseline CLIP-ResNet-50, the proposed TDA demonstrates a significant improvement in testing accuracy $(+1.54\\%)$ , with only a minimal sacrifice in testing efficiency (requiring an additional 4min). In comparison to TPT and DiffTPT, the proposed TDA demonstrates not only superior", + "bbox": [ + 80, + 688, + 467, + 898 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "testing accuracy but also significantly improved efficiency. It reduces the testing time dramatically from $12\\mathrm{h}50\\mathrm{min}$ by TPT and even more from $34\\mathrm{h}45\\mathrm{min}$ by DiffTPT, down to just 16 minutes. Without including the image generation time, DiffTPT consumes clearly more test time than TPT as it involves a time-consuming multi-step prompt updating process whereas TPT requires a single step only. The experimental results strongly validate the effectiveness and efficiency of our proposed method, establishing its suitability for real-world applications.", + "bbox": [ + 503, + 411, + 890, + 561 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We then compare TDA with state-of-the-art methods over cross-domain benchmark. The results, presented in Table 3, demonstrate that TDA not only surpasses the performance of the TPT method but also shows a significant advantage over its improvement method DiffTPT. Specifically, when utilizing CLIP-ResNet-50 and CLIP-ViT-B/16 as the backbone, TDA achieves an improvement in average accuracy over TPT by $3.37\\%$ and $2.43\\%$ , respectively. These improvements, along with a $1.18\\%$ and $2.06\\%$ gain over DiffTPT for the respective backbones, further verify the effectiveness of TDA in adapting to diverse class datasets during test time. This attribute holds significant value for vision-language models such as CLIP, as it enables them to classify arbitrary classes in image classification without the need for additional training.", + "bbox": [ + 503, + 564, + 890, + 789 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Ablation Studies", + "text_level": 1, + "bbox": [ + 503, + 801, + 661, + 816 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In this section, we perform ablation studies to examine the effectiveness of our designs. All the ablation studies are conducted over the ImageNet dataset, where TDA can achieve an accuracy of $61.35\\%$ under default settings. TDA consists of a Positive Cache and a Negative Cache, which", + "bbox": [ + 503, + 825, + 890, + 898 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "14168", + "bbox": [ + 480, + 946, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/83ab17c90bae8b8ba66f047b4294bcf5a80889fa31294d1162de42bad164366d.jpg", + "image_caption": [ + "Figure 3. Ablation studies on two cache designs in TDA: Positive Cache and Negative Cache. All the models are built upon the baseline model CLIP-ResNet-50." + ], + "image_footnote": [], + "bbox": [ + 81, + 90, + 467, + 316 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "perform positive and negative pseudo labeling within our designed dynamic adapter, respectively. We first assess the efficacy of the two cache designs in TDA. As shown in Figure 3, both Positive Cache and Negative Cache significantly surpass the baseline model CLIP, demonstrating that test-time adaptation can be improved by introducing a dynamic adapter with either positive pseudo labeling or negative pseudo labeling. Besides, the two cache designs in TDA can complement each other as TDA (i.e., the combination of the two designs) clearly outperforms either Positive Cache or Negative Cache on the challenging ImageNet dataset. Moreover, we extend our ablation studies to the cross-domain benchmark, and results show that the positive cache achieved $60.38\\%$ accuracy, the negative cache $60.11\\%$ , while their combination yielded a $61.03\\%$ accuracy, thereby highlighting the significance of each type of cache in enhancing TDA's performance.", + "bbox": [ + 73, + 398, + 468, + 656 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We proceed to perform parameter studies on the shot capacity, which refers to the maximum number of key-value pairs per class, in both Positive Cache and Negative Cache models. These studies aims to find the optimal balance between the diversity and accuracy of the key-value pairs. Figure 4 shows that the performance of both cache models is significantly affected when the shot capacity is either too low or too high. We find that the shot capacity is set as 3 for the Positive Cache and 2 for the Negative Cache yields the best performance. This is because an appropriate shot capacity ensures both high-quality pseudo labels (paired values) and diverse image features (paired keys) in the cache models. The negative cache is sensitive to shot numbers due to its role in storing probabilities for various negative pseudo labels, each representing a class to be excluded from the model. Contrary to the intuition that a larger negative", + "bbox": [ + 75, + 659, + 470, + 901 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/0003c7a8fe921051a3ee6ea5ce1be61aef18c4e9274ecd6118067ace10d5425a.jpg", + "image_caption": [ + "Figure 4. Parameter studies on the Shot Capacity in Positive Cache and Negative Cache." + ], + "image_footnote": [], + "bbox": [ + 504, + 90, + 890, + 321 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "cache might be beneficial, a larger negative cache leads to more high-entropy, noisier pseudo labels, as highlighted in self-training studies like [11], thereby lowering the performance. Conversely, the positive cache, which stores single and high-confidence prediction, is less affected by variations in shot capacity, thereby maintaining consistent accuracy across different shot capacities. To facilitate practical applications, the shot capacity settings are fixed and directly applied to new datasets without the need for additional parameter adjustments.", + "bbox": [ + 496, + 383, + 893, + 536 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 547, + 619, + 565 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this work, we have presented TDA, a dynamic adapter for efficient and effective test-time adaptation of vision-language models. The proposed method employs a key-value cache, which maintains a dynamic queue with test-sample features as keys and corresponding few-shot pseudo labels as values, allowing for gradual adaptation to test data through progressive pseudo label improvement. Moreover, TDA introduces a negative cache to mitigate the undesirable effects of noisy pseudo labels by assigning negative pseudo labels to certain classes when the model is uncertain about its predictions. The results of extensive experiments over two benchmarks demonstrate that TDA outperforms state-of-the-art test-time adaptation methods while significantly reducing testing time. This work contributes to the research and application values of test-time adaptation and presents a promising solution to the efficiency issue of test-time adaptation of vision-language models.", + "bbox": [ + 496, + 573, + 893, + 830 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgement", + "text_level": 1, + "bbox": [ + 500, + 844, + 660, + 861 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This study is supported under the Mohamed bin Zayed University of Artificial Intelligence.", + "bbox": [ + 500, + 869, + 890, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "14169", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 174, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Lukas Bossard, Matthieu Guillaumin, and Luc Van Gool. Food-101 – mining discriminative components with random forests. In European Conference on Computer Vision, 2014. 5", + "[2] Malik Boudiaf, Romain Mueller, Ismail Ben Ayed, and Luca Bertinetto. Parameter-free online test-time adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8344-8353, 2022. 1", + "[3] Dian Chen, Dequan Wang, Trevor Darrell, and Sayna Ebrahimi. Contrastive test-time adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 295-305, 2022. 1", + "[4] M. Cimpoi, S. Maji, I. Kokkinos, S. Mohamed, , and A. Vedaldi. Describing textures in the wild. In Proceedings of the IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2014. 5", + "[5] J. Deng, W. Dong, R. Socher, L.-J. Li, K. Li, and L. Fei-Fei. ImageNet: A Large-Scale Hierarchical Image Database. In CVPR09, 2009. 5", + "[6] Karan Desai and Justin Johnson. VirTex: Learning Visual Representations from Textual Annotations. In CVPR, 2021. 2", + "[7] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 6", + "[8] Li Fei-Fei, Rob Fergus, and Pietro Perona. Learning generative visual models from few training examples: An incremental bayesian approach tested on 101 object categories. In CVPR Workshops, 2004. 5", + "[9] Chun-Mei Feng, Kai Yu, Yong Liu, Salman Khan, and Wangmeng Zuo. Diverse data augmentation with diffusions for effective test-time prompt tuning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2704-2714, 2023. 1, 2, 6", + "[10] Peng Gao, Shijie Geng, Renrui Zhang, Teli Ma, Rongyao Fang, Yongfeng Zhang, Hongsheng Li, and Yu Qiao. Clip-adapter: Better vision-language models with feature adapters. arXiv preprint arXiv:2110.04544, 2021. 2", + "[11] Yves Grandvalet and Yoshua Bengio. Semi-supervised learning by entropy minimization. Advances in neural information processing systems, 17, 2004. 2, 8", + "[12] Edouard Grave, Moustapha M Cisse, and Armand Joulin. Unbounded cache model for online language modeling with open vocabulary. Advances in neural information processing systems, 30, 2017. 2, 3", + "[13] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, 2016. 6", + "[14] Patrick Helber, Benjamin Bischke, Andreas Dengel, and Damian Borth. Eurosat: A novel dataset and deep learning benchmark for land use and land cover classification. IEEE J. Sel. Top. Appl. Earth Obs. Remote. Sens., 2019. 5" + ], + "bbox": [ + 78, + 114, + 470, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[15] Dan Hendrycks, Steven Basart, Norman Mu, Saurav Kadavath, Frank Wang, Evan Dorundo, Rahul Desai, Tyler Zhu, Samyak Parajuli, Mike Guo, Dawn Song, Jacob Steinhardt, and Justin Gilmer. The many faces of robustness: A critical analysis of out-of-distribution generalization. In ICCV, 2021. 5", + "[16] Dan Hendrycks, Kevin Zhao, Steven Basart, Jacob Steinhardt, and Dawn Song. Natural adversarial examples. In CVPR, pages 15262-15271, 2021. 5", + "[17] Xiaowei Hu, Zhe Gan, Jianfeng Wang, Zhengyuan Yang, Zicheng Liu, Yumao Lu, and Lijuan Wang. Scaling up vision-language pre-training for image captioning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17980-17989, 2022. 2", + "[18] Yusuke Iwasawa and Yutaka Matsuo. Test-time classifier adjustment module for model-agnostic domain generalization. Advances in Neural Information Processing Systems, 34:2427-2440, 2021. 2", + "[19] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In International Conference on Machine Learning, pages 4904-4916. PMLR, 2021. 1, 2", + "[20] Jeff Johnson, Matthijs Douze, and Hervé Jégou. Billionscale similarity search with gpus. IEEE Transactions on Big Data, 7(3):535-547, 2019. 3", + "[21] Urvashi Khandelwal, Omer Levy, Dan Jurafsky, Luke Zettlemoyer, and Mike Lewis. Generalization through memorization: Nearest neighbor language models. In International Conference on Learning Representations, 2020. 2", + "[22] Youngdong Kim, Junho Yim, Juseung Yun, and Junmo Kim. NInl: Negative learning for noisy labels. In Proceedings of the IEEE/CVF international conference on computer vision, pages 101-110, 2019. 2, 5", + "[23] Youngdong Kim, Juseung Yun, Hyounguk Shon, and Junmo Kim. Joint negative and positive learning for noisy labels. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9442-9451, 2021. 2, 5", + "[24] Jonathan Krause, Michael Stark, Jia Deng, and Li Fei-Fei. 3d object representations for fine-grained categorization. In ICCV Workshops, 2013. 5", + "[25] Xin Li, Dongze Lian, Zhihe Lu, Jiawang Bai, Zhibo Chen, and Xinchao Wang. Graphadapter: Tuning vision-language models with dual knowledge graph. Advances in Neural Information Processing Systems, 36, 2024. 2", + "[26] S. Maji, J. Kannala, E. Rahtu, M. Blaschko, and A. Vedaldi. Fine-grained visual classification of aircraft. Technical report, 2013. 5", + "[27] Stephen Merity, Caiming Xiong, James Bradbury, and Richard Socher. Pointer sentinel mixture models. arXiv preprint arXiv:1609.07843, 2016. 3", + "[28] Maria-Elena Nilsback and Andrew Zisserman. Automated flower classification over a large number of classes. In Indian Conference on Computer Vision, Graphics and Image Processing, 2008. 5" + ], + "bbox": [ + 501, + 92, + 893, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "14170", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[29] Emin Orhan. A simple cache model for image recognition. Advances in Neural Information Processing Systems, 31, 2018. 2", + "[30] Omkar M. Parkhi, Andrea Vedaldi, Andrew Zisserman, and C. V. Jawahar. Cats and dogs. In CVPR, 2012. 5", + "[31] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In ICML, 2021. 1, 2, 3, 6", + "[32] Benjamin Recht, Rebecca Roelofs, Ludwig Schmidt, and Vaishaal Shankar. Do imagenet classifiers generalize to imageNet? In ICML, 2019. 5", + "[33] Mamshad Nayeem Rizve, Kevin Duarte, Yogesh S Rawat, and Mubarak Shah. In defense of pseudo-labeling: An uncertainty-aware pseudo-label selection framework for semi-supervised learning. arXiv preprint arXiv:2101.06329, 2021. 2, 5", + "[34] Steffen Schneider, Evgenia Rusak, Luisa Eck, Oliver Bringmann, Wieland Brendel, and Matthias Bethge. Improving robustness against common corruptions by covariate shift adaptation. Advances in Neural Information Processing Systems, 33:11539-11551, 2020. 2", + "[35] Manli Shu, Weili Nie, De-An Huang, Zhiding Yu, Tom Goldstein, Anima Anandkumar, and Chaowei Xiao. Test-time prompt tuning for zero-shot generalization in vision-language models. In Advances in Neural Information Processing Systems, 2022. 1, 2, 3, 5, 6", + "[36] Khurram Soomro, Amir Roshan Zamir, and Mubarak Shah. UCF101: A dataset of 101 human actions classes from videos in the wild. CoRR, abs/1212.0402, 2012. 5", + "[37] Yu Sun, Xiaolong Wang, Zhuang Liu, John Miller, Alexei A. Efros, and Moritz Hardt. Test-time training with self-supervision for generalization under distribution shifts. In ICML, 2020. 2", + "[38] Thomas Varsavsky, Maurizio Orbes-Arteaga, Carole H Sudre, Mark S Graham, Parashkev Nachev, and M Jorge Cardoso. Test-time unsupervised domain adaptation. In Medical Image Computing and Computer Assisted Intervention—MICCAI 2020: 23rd International Conference, Lima, Peru, October 4–8, 2020, Proceedings, Part I 23, pages 428–436. Springer, 2020. 2", + "[39] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 6", + "[40] Dequan Wang, Evan Shelhamer, Shaoteng Liu, Bruno A. Olshausen, and Trevor Darrell. Tent: Fully test-time adaptation by entropy minimization. In ICLR, 2021. 1, 2", + "[41] Haohan Wang, Songwei Ge, Zachary Lipton, and Eric P Xing. Learning robust global representations by penalizing local predictive power. In NeurIPS, 2019. 5", + "[42] J. Xiao, J. Hays, K. A. Ehinger, A. Oliva, and A. Torralba. Sun database: Large-scale scene recognition from abbey to zoo. In 2010 IEEE Computer Society Conference on Computer Vision and Pattern Recognition, 2010. 5" + ], + "bbox": [ + 78, + 90, + 468, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[43] Jinyu Yang, Jiali Duan, Son Tran, Yi Xu, Sampath Chanda, Liquun Chen, Belinda Zeng, Trishul Chilimbi, and Junzhou Huang. Vision-language pre-training with triple contrastive learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15671-15680, 2022. 1, 2", + "[44] Tao Yu, Zhihe Lu, Xin Jin, Zhibo Chen, and Xinchao Wang. Task residual for tuning vision-language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10899-10909, 2023. 2", + "[45] Marvin Zhang, Henrik Marklund, Nikita Dhawan, Abhishek Gupta, Sergey Levine, and Chelsea Finn. Adaptive risk minimization: Learning to adapt to domain shift. Advances in Neural Information Processing Systems, 34:23664-23678, 2021. 2", + "[46] Marvin Mengxin Zhang, Sergey Levine, and Chelsea Finn. MEMO: Test time robustness via adaptation and augmentation. In Advances in Neural Information Processing Systems, 2022. 2", + "[47] Renrui Zhang, Wei Zhang, Rongyao Fang, Peng Gao, Kunchang Li, Jifeng Dai, Yu Qiao, and Hongsheng Li. Tip-adapter: Training-free adaption of clip for few-shot classification. In European Conference on Computer Vision, pages 493-510. Springer, 2022. 2, 3, 6", + "[48] Renrui Zhang, Xiangfei Hu, Bohao Li, Siyuan Huang, Hanqiu Deng, Yu Qiao, Peng Gao, and Hongsheng Li. Prompt, generate, then cache: Cascade of foundation models makes strong few-shot learners. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15211-15222, 2023. 2", + "[49] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Conditional prompt learning for vision-language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16816-16825, 2022. 2, 6", + "[50] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Learning to prompt for vision-language models. International Journal of Computer Vision, 130(9):2337-2348, 2022. 2, 6" + ], + "bbox": [ + 501, + 92, + 890, + 638 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "14171", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2024/Efficient Test-Time Adaptation of Vision-Language Models/f86a8c24-7089-4003-88ca-fa1a5c70fe2d_model.json b/2024/Efficient Test-Time Adaptation of Vision-Language Models/f86a8c24-7089-4003-88ca-fa1a5c70fe2d_model.json new file mode 100644 index 0000000000000000000000000000000000000000..0b982bfd3fe83b0a581ad7887672ef9ce8eb52a1 --- /dev/null +++ b/2024/Efficient Test-Time Adaptation of Vision-Language Models/f86a8c24-7089-4003-88ca-fa1a5c70fe2d_model.json @@ -0,0 +1,1804 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.188, + 0.131, + 0.784, + 0.154 + ], + "angle": 0, + "content": "Efficient Test-Time Adaptation of Vision-Language Models" + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.179, + 0.868, + 0.201 + ], + "angle": 0, + "content": "Adilbek Karmanov\\(^{1*}\\) Dayan Guan\\(^{2*}\\) Shijian Lu\\(^{1,2\\dagger}\\) Abdulmotaleb El Saddik\\(^{1,3}\\) Eric Xing\\(^{1,4}\\)" + }, + { + "type": "text", + "bbox": [ + 0.105, + 0.201, + 0.865, + 0.24 + ], + "angle": 0, + "content": "\\(^{1}\\)Mohamed bin Zayed University of Artificial Intelligence \\(^{2}\\)Nanyang Technological University \\(^{3}\\)University of Ottawa \\(^{4}\\)Carnegie Mellon University" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.272, + 0.314, + 0.289 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.305, + 0.474, + 0.637 + ], + "angle": 0, + "content": "Test-time adaptation with pre-trained vision-language models has attracted increasing attention for tackling distribution shifts during the test time. Though prior studies have achieved very promising performance, they involve intensive computation which is severely unaligned with test-time adaptation. We design TDA, a training-free dynamic adapter that enables effective and efficient test-time adaptation with vision-language models. TDA works with a lightweight key-value cache that maintains a dynamic queue with few-shot pseudo labels as values and the corresponding test-sample features as keys. Leveraging the key-value cache, TDA allows adapting to test data gradually via progressive pseudo label refinement which is super-efficient without incurring any backpropagation. In addition, we introduce negative pseudo labeling that alleviates the adverse impact of pseudo label noises by assigning pseudo labels to certain negative classes when the model is uncertain about its pseudo label predictions. Extensive experiments over two benchmarks demonstrate TDA's superior effectiveness and efficiency as compared with the state-of-the-art. The code has been released in https://kdiaaa.github.io/tda/." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.669, + 0.21, + 0.685 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.696, + 0.47, + 0.864 + ], + "angle": 0, + "content": "Recent advances in vision-language models [19, 31, 43] have opened a new door for integrating human language into various computer vision tasks. Take CLIP [31] as an example. It can enable zero-shot image classification by leveraging a shared embedding space that is learnt from web-scale image-text pairs. Within this shared space, images can be directly recognized by matching their features with the text embeddings of CLIP classes. At the other end, CLIP often faces challenges while handling various specific downstream images, especially when the downstream images have clear domain and distribution shifts as compared" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.271, + 0.891, + 0.43 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.602, + 0.438, + 0.793, + 0.452 + ], + "angle": 0, + "content": "(a) Test-time Prompt Tuning [9, 35]" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.457, + 0.891, + 0.642 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.584, + 0.648, + 0.81, + 0.661 + ], + "angle": 0, + "content": "(b) Training-free Dynamic Adapter (Ours)" + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.673, + 0.894, + 0.77 + ], + "angle": 0, + "content": "Figure 1. Comparison of our proposed Training-free Dynamic Adapter (TDA) with Test-time Prompt Tuning TPT [35] and its enhancement DiffTPT [9]: both TPT and DiffTPT require significant computational resources to optimize the learnable prompt via backpropagation; TDA is a dynamic cache that is training-free without any backpropagation, making it efficient for test-time adaptation in various real-world scenarios." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.806, + 0.707, + 0.822 + ], + "angle": 0, + "content": "with the CLIP training images." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.825, + 0.894, + 0.903 + ], + "angle": 0, + "content": "Several recent studies [2, 3, 40] introduce a new paradigm called test-time adaptation for mitigating the domain shift. The idea of test-time adaptation is well aligned with real-world scenarios where a model needs to adapt to new environments quickly. Despite its great research and" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.875, + 0.21, + 0.888 + ], + "angle": 0, + "content": "*Equal Contribution." + }, + { + "type": "page_footnote", + "bbox": [ + 0.098, + 0.888, + 0.227, + 0.9 + ], + "angle": 0, + "content": "† Corresponding Author." + }, + { + "type": "list", + "bbox": [ + 0.096, + 0.875, + 0.227, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "14162" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.288 + ], + "angle": 0, + "content": "application values, test-time adaptation of vision-language models has been largely neglected. Recently, Test-time Prompt Tuning, as introduced in TPT [35] and its enhancement DiffTPT [9], attempts to adapt vision-language models by learning domain-specific prompts from test data. As illustrated in Fig. 1 (a), both TPT and DiffTPT train a learnable prompt for each test sample by feeding its augmented views into the CLIP model for generating predictions and minimizing the marginal entropy of confident predictions. Despite its decent performance, the prompt optimization in both TPT [50] and DiffTPT [9] is computationally intensive which hinders its applications in various real-world scenarios." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.29, + 0.473, + 0.534 + ], + "angle": 0, + "content": "We design a training-free dynamic adapter (TDA) that allows efficient and effective test-time adaptation of vision-language models without requiring any backpropagation in test time. As Fig. 1 (b) shows, TDA constructs a lightweight Dynamic Adapter that keeps a dynamic queue with the pseudo labels of a stream of test samples as values and the corresponding CLIP-extracted features as keys. TDA has two desirable features that make its test-time adaptation highly applicable in real-world scenarios. First, TDA is highly effective, as it improves the quality of pseudo labels via progressive incorporation of test predictions of lower entropy [11, 40]. Second, TDA is very efficient as the key-value cache is non-parametric and requires no backpropagation during testing. Beyond that, TDA cache is lightweight due to the few-shot setup and it can be computed with simple matrix multiplications [12, 21, 29, 47]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.535, + 0.473, + 0.853 + ], + "angle": 0, + "content": "Note that the performance of TDA depends heavily on the pseudo labels of unlabelled test samples which are often noisy with prediction errors. Inspired by the idea of negative learning [22, 23, 33], we introduce negative pseudo labeling to reduce the impact of noisy estimated labels. Traditional pseudo labeling methods identify the presence of particular classes in unlabeled data, which may result in erroneous pseudo labels being assigned when comparable high probabilities are observed. In contrast, our designed negative pseudo labeling determines the absence of certain classes and can provide more accurate pseudo labels as the probabilities of these complementary classes are very low. Concretely, we construct an additional TDA cache that stores negative pseudo labels to complement the positive TDA cache. By combining positive and negative caches, TDA is more tolerant to noisy pseudo labels and can better generalize to testing data. Extensive experiments over two widely adopted test-time adaptation benchmarks show that TDA outperforms the state-of-the-art by large margins while significantly reducing the testing time from over 12 hours to 16 minutes on the ImageNet dataset." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.856, + 0.472, + 0.903 + ], + "angle": 0, + "content": "In summary, the contributions of this work are threefold. First, we design a training-free dynamic adapter (TDA) that can achieve test-time adaptation of vision-language models" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.228 + ], + "angle": 0, + "content": "efficiently and effectively. To the best of our knowledge, this is the first work that investigates the efficiency issue of test-time adaptation of vision-language models. Second, we introduce negative pseudo labeling to alleviate the adverse impact of pseudo label noises which makes TDA more robust to pseudo label noises and generalizable to testing data. Third, we evaluate TDA extensively over two benchmarks, and experiments show that TDA achieves superior accuracy and efficiency compared with the state-of-the-art." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.243, + 0.642, + 0.259 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.269, + 0.895, + 0.555 + ], + "angle": 0, + "content": "Vision-language models [6, 17, 19, 31, 43] have demonstrated significant potential in learning semantic representations effectively by undergoing extensive training on image-text data. CLIP [31] stands out among these models for its ability to establish links between visual and textual representations, which enables it to achieve impressive zero-shot results on various downstream tasks. To enhance the transfer learning capability of the CLIP model in the downstream classification tasks, researchers have proposed integrating language prompt learners such as CoOp [50] and CoCoOp [49], as well as vision adapters such as CLIP-Adapter [10], Tip-Adapter [47], CaFo [48], TaskRes [44] and GraphAdapter [25]. Although these methods have shown considerable performance improvements, they typically require a large amount of training data in downstream tasks, making them less practical for real-world scenarios. On other hand, this work focuses on a new paradigm named test-time adaptation without accessing the original training data." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.569, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Test-time adaptation refers to the process of adapting models to testing data that may have distributional differences from the training data. It is particularly beneficial for real-world applications that require models to be deployed in diverse environments, such as autonomous driving in various weather conditions, medical diagnosis in different hospitals, and etc. Several recent works utilize each batch of testing samples to update partial weights [18, 37, 38], normalization statistics [34], or a combination of both [40, 45]. To avoid updating models with multiple testing samples, MEMO [46] proposes enforcing the invariant predictions from different augmentations of each sample in the testing data stream. TPT [35] tackles the same challenge with vision-language models by fine-tuning a learnable prompt with each testing sample. DiffTPT [9] innovates test-time prompt tuning by leveraging pre-trained diffusion models to augment the diversity of test data samples used in TPT. Although TPT [35] and DiffTPT [9] are effective in addressing test-time adaptation of vision-language models, prompt learning is computationally expensive and time-consuming. This paper aims to mitigate the computational efficiency challenges of TPT and DiffTPT through the introduction of" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "14163" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.078, + 0.092, + 0.181, + 0.106 + ], + "angle": 0, + "content": "a cache model." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.119, + 0.473, + 0.406 + ], + "angle": 0, + "content": "Cache model benefits the adaptation techniques by providing efficient inference and non-parametric processing without requiring parameter updates. Unbounded Cache [12] and PSMM [27] have shown promising results in the text generation task by storing a large amount of the training dataset to capture long-term dependencies, however, this approach poses a challenge of memory efficiency during the test-time adaptation. A different technique [20] was proposed to mitigate this limitation by reducing the size of the cache memory through the search for application. Alternatively, Tip-Adapter [47] solves the cache memory problem by only storing few-shot samples per class to create a cache model for vision-language models. Inspired by this work, our dynamic adapter uses the same architecture for the test-time adaptation setting, where there is no access to the source data, and testing samples can only be accessed one by one. To address the lack of access to source data during testing, our adapter collects the most reliable test samples and their pseudo labels to form cache model." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.419, + 0.17, + 0.435 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.444, + 0.218, + 0.458 + ], + "angle": 0, + "content": "3.1. Preliminaries" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.467, + 0.47, + 0.604 + ], + "angle": 0, + "content": "CLIP [31] is a vision-language model that performs a proxy task of predicting the correct pairings of text and image. Consider an \\(N\\)-class classification problem where the CLIP's objective in the zero-shot setting is to match images with their most relevant textual descriptions using an image encoder \\(E_{v}\\) and a text encoder \\(E_{t}\\). To obtain the textual descriptions, \\(N\\)-class names are concatenated with handcrafted prompts and then mapped into the \\(c\\)-channeled text embeddings \\(\\mathbf{W}_{c}\\) using the text encoder \\(E_{t}\\)." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.615, + 0.473, + 0.812 + ], + "angle": 0, + "content": "TPT [35] focuses on test-time adaptation of CLIP. In TPT, a prompt tuning method is proposed to learn an adaptive prompt \\(\\mathbf{p}_c\\) using individual test samples. A set of augmentation functions \\(\\mathcal{A}\\) is used to generate \\(n\\) randomly augmented views \\(\\tilde{x}_{\\mathrm{test}} = \\mathcal{A}_n(x_{\\mathrm{test}})\\) of a test sample \\(x_{\\mathrm{test}}\\). The objective of TPT is to reduce variation in the model's predictions across different augmentations \\(\\tilde{f}_{\\mathrm{test}} = E_v(\\tilde{x}_{\\mathrm{test}})\\) by minimizing the marginal entropy among the outputs of the augmented views. Furthermore, TPT also includes confidence selection to discard noisy augmentations that could result in ambiguous model predictions, as shown in Figure 1a. This is achieved by filtering out augmented views with high-entropy predictions as:" + }, + { + "type": "equation", + "bbox": [ + 0.12, + 0.822, + 0.47, + 0.862 + ], + "angle": 0, + "content": "\\[\nP _ {\\mathrm {T P T}} \\left(\\tilde {f} _ {\\text {t e s t}}\\right) = \\frac {1}{\\rho n} \\sum_ {i = 1} ^ {n} \\mathbb {1} \\left[ \\mathrm {H} \\left(\\tilde {f} _ {i} \\mathbf {p} _ {c} ^ {T}\\right) \\leq \\tau \\right] \\tilde {f} _ {i} \\mathbf {p} _ {c} ^ {T}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.472, + 0.903 + ], + "angle": 0, + "content": "where \\(\\mathrm{H}\\) is the self-entropy function of the softmax logits predictions, \\(\\tilde{f}_i\\mathbf{p}_c^T\\) is the class probabilities vector of size" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.892, + 0.152 + ], + "angle": 0, + "content": "\\(N\\) generated from the given \\(i\\)-th augmented view of the test image, and the parameter \\(\\tau\\) determines that only \\(\\rho\\)-percentile of confident samples with entropy values below this threshold can be selected out of \\(n\\) augmented views." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.165, + 0.893, + 0.345 + ], + "angle": 0, + "content": "Tip-Adapter [47] provides a training-free solution that uses a key-value cache model and integrates knowledge from the pre-trained CLIP model with few-shot labeled samples. The cache model is created using a set of \\( k \\)-shot labeled samples \\( x_{k} \\) from \\( N \\) classes and their corresponding ground-truth labels \\( y_{N} \\). It can be conceptualized as a linear two-layer model, where the first layer contains train image features \\( \\mathbf{F}_{\\mathrm{train}} = E_v(x_k) \\) and the second layer consists of one-hot vectors \\( \\mathbf{L}_{\\mathrm{train}} \\) encoded from the labels \\( y_{N} \\). Given test image features \\( f_{\\mathrm{test}} \\) generated from the CLIP's image encoder \\( E_{v} \\), the prediction from the cache model can be calculated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.571, + 0.356, + 0.892, + 0.375 + ], + "angle": 0, + "content": "\\[\nP _ {\\mathrm {c a c h e}} \\left(f _ {\\mathrm {t e s t}}\\right) = A \\left(f _ {\\mathrm {t e s t}} \\mathbf {F} _ {\\mathrm {t r a i n}} ^ {T}\\right) \\mathbf {L} _ {\\mathrm {t r a i n}}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.385, + 0.892, + 0.463 + ], + "angle": 0, + "content": "where \\( A(z) = \\alpha \\exp(-\\beta (1 - z)) \\) is an adaptation function within a weighting factor \\( \\alpha \\) and a sharpness ratio \\( \\beta \\). During inference, the prediction of Tip-Adapter is computed by combining the pre-trained CLIP model and the cache model as: \\( P_{\\mathrm{TA}}(f_{\\mathrm{test}}) = P_{\\mathrm{cache}}(f_{\\mathrm{test}}) + f_{\\mathrm{test}} \\mathbf{W}_c^T \\)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.47, + 0.78, + 0.487 + ], + "angle": 0, + "content": "3.2. Training-free Dynamic Adapter" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.493, + 0.892, + 0.644 + ], + "angle": 0, + "content": "During testing, pre-trained vision-language models like CLIP may encounter distribution shifts that degrade the classification performance. To address this issue, existing test-time prompt tuning methods train a learnable prompt by enforcing consistency across different image augmentations during testing. However, it requires a large number of augmentation operations on each test image and computationally-heavy optimization steps to learn the prompt, limiting its applicability in various real-world settings." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.645, + 0.893, + 0.902 + ], + "angle": 0, + "content": "In this paper, our motivation is to design an efficient method for test-time adaptation of the pre-trained vision-language models like CLIP. Inspired by the concept of TipAdapter, we propose a training-free dynamic adapter (TDA) to enable efficient and effective test-time adaptation with CLIP. As shown in Figure 2, TDA includes two lightweight key-value caches, where each cache stores a dynamic queue of few-shot test features as keys and the corresponding pseudo labels as values. The first cache is intended for positive learning and it dynamically updates key-value pairs with high-confidence predictions to improve the accuracy. The second cache is designed for negative learning and it aims to address the adverse effects of noisy pseudo labels by introducing negative pseudo labeling to identify class absence rather than presence. By combining the positive cache and negative cache, the proposed TDA can achieve superior performance in terms of both speed and accuracy." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "14164" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.084, + 0.087, + 0.895, + 0.316 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.33, + 0.895, + 0.416 + ], + "angle": 0, + "content": "Figure 2. Overview of the proposed Training-free Dynamic Adapter (TDA). TDA constructs and updates two key-value caches to store the knowledge of a stream of test samples, and uses the two caches to generate positive and negative predictions which are combined with CLIP predictions to produce the final prediction. Specifically, the CLIP predictions are generated by performing the dot product between the image features generated by CLIP's image encoder \\( E_{v} \\) and the text embeddings generated by CLIP's text encoder \\( E_{t} \\), using the hand-crafted prompt and class names. The two key-value caches are updated by gradually incorporating the test features and their corresponding pseudo labels calculated from CLIP's predictions, based on prediction entropy and cache capacity." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.441, + 0.473, + 0.578 + ], + "angle": 0, + "content": "Our goal is to conduct test-time adaptation by gathering adequate knowledge from the testing data stream and improving the predictions through the adaptation of the image features. To accomplish this goal, we create positive and negative caches that capture particular characteristics of the testing data stream. In the upcoming parts, we will present the process of collecting data for each cache and define the conditions in which we can utilize cache information to adapt features and enhance model predictions." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.601, + 0.47, + 0.842 + ], + "angle": 0, + "content": "Positive Cache. The positive cache in TDA is a key-value cache, in which keys and values are represented as a dynamic queue. It aims to collect high-quality few-shot pseudo labels \\(\\hat{\\mathbf{L}}_{\\mathfrak{p}}\\) as positive values and the corresponding features \\(\\mathbf{Q}_{\\mathfrak{p}}\\) as keys. The key-value cache is initially empty and then accumulates a sufficient number of key-value pairs during the test-time adaptation. To maintain high-quality pseudo labels, TDA progressively incorporates test predictions with lower entropy while limiting shot capacity1 in the positive cache. Different from a normal queue like a FIFO queue with a fixed size, the dynamic queue in our method expands in size during testing. Besides, the dynamic queue operates similarly to a priority queue, using entropy as the criterion for prioritization. Note that, each class has its own queue to maintain the order and correct data structure of each class in the cache." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.843, + 0.47, + 0.875 + ], + "angle": 0, + "content": "Given a pre-trained CLIP model that consists of a text encoder \\( E_{t} \\) and an image encoder \\( E_{v} \\), \\( E_{t} \\) processes class" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.441, + 0.892, + 0.837 + ], + "angle": 0, + "content": "names with pre-defined prompts to generate \\( c \\)-channeled text embeddings \\( \\mathbf{W}_c \\) and \\( E_v \\) processes each test image \\( x_{\\mathrm{test}} \\) to produce image features \\( f_{\\mathrm{test}} \\). To build the positive cache, TDA first generates a pseudo label \\( \\hat{l} \\), a one-hot encoded vector of a categorical distribution, for each test sample \\( x_{\\mathrm{test}} \\) by applying the softmax function to the prediction \\( f_{\\mathrm{test}} \\mathbf{W}_c^T \\). We establish two conditions to ascertain whether and how to include the pseudo label \\( \\hat{l} \\) and its corresponding image features \\( f_{\\mathrm{test}} \\) into the positive cache. The first condition is defined as: if the shot number (the number of collected pairs per class) of \\( \\hat{\\mathbf{L}}_{\\mathbf{p}} \\) is less than the maximum shot capacity (the maximum capacity number of pairs per class) \\( k \\), TDA will add \\( \\hat{l} \\) and \\( f_{\\mathrm{test}} \\) as a new value and a new key to \\( \\hat{\\mathbf{L}}_{\\mathbf{p}} \\) and \\( \\mathbf{Q}_{\\mathbf{p}} \\) respectively. Meanwhile, the second condition is defined as: if the shot number of \\( \\hat{\\mathbf{L}}_{\\mathbf{p}} \\) has reached the maximum shot capacity \\( k \\), TDA will replace an 'uncertain' key-value pair \\( \\{\\mathbf{q}^{ent}, \\hat{l}_p^{ent}\\} \\) with \\( \\{f_{\\mathrm{test}}, \\hat{l}\\} \\) when \\( \\mathrm{H}(f_{\\mathrm{test}} \\mathbf{W}_c^T) < \\mathrm{H}(\\mathbf{q}^{ent} \\mathbf{W}_c^T) \\). Here, \\( \\mathrm{H} \\) denotes the entropy function and the term 'uncertain' indicates that the entropy of a particular key-value pair is the highest compared to the entropy of all other key-value pairs of the same class in the cache model. By applying these two conditions, TDA can gradually integrate test predictions with lower entropy while controlling the shot capacity, which helps to ensure the collection of high-quality pseudo labels in positive cache." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.84, + 0.894, + 0.903 + ], + "angle": 0, + "content": "During test-time adaptation, the positive cache can quickly retrieve relevant information by treating the image features \\( f_{\\mathrm{test}} \\) generated from a test example as a query and searching the stored key-value pairs \\( \\{\\mathbf{Q}_{\\mathbf{p}},\\hat{\\mathbf{L}}_{\\mathbf{p}}\\} \\) for match-" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.887, + 0.458, + 0.902 + ], + "angle": 0, + "content": "1The shot capacity refers to the maximum number of pairs per class." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "14165" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.121 + ], + "angle": 0, + "content": "ing information. The adapted predictions using the positive cache can then be obtained as:" + }, + { + "type": "equation", + "bbox": [ + 0.171, + 0.128, + 0.469, + 0.148 + ], + "angle": 0, + "content": "\\[\nP _ {\\mathrm {p o s}} \\left(f _ {\\mathrm {t e s t}}\\right) = A \\left(f _ {\\mathrm {t e s t}} \\mathbf {Q} _ {\\mathbf {p}} ^ {\\mathbf {T}}\\right) \\hat {\\mathbf {L}} _ {\\mathbf {p}}, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.153, + 0.465, + 0.168 + ], + "angle": 0, + "content": "where \\(A\\) is the adaptation function defined in Tip-Adapter." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.185, + 0.469, + 0.306 + ], + "angle": 0, + "content": "Negative Cache. Similar to the positive cache in our TDA, the negative cache is also a dynamic queue structure with negative keys and negative values denoted as \\(\\mathbf{Q}_{\\mathbf{n}}\\) and \\(\\hat{\\mathbf{L}}_{\\mathbf{n}}\\), respectively. It aims to gather CLIP-generated image features to \\(\\mathbf{Q}_{\\mathbf{n}}\\) and the corresponding negative pseudo labels to \\(\\hat{\\mathbf{L}}_{\\mathbf{n}}\\). Unlike the pseudo labels in the positive cache, the negative pseudo labels are obtained by applying negative mask on the class probabilities as:" + }, + { + "type": "equation", + "bbox": [ + 0.19, + 0.311, + 0.469, + 0.331 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathbf {L}} _ {\\mathbf {n}} = - \\mathbb {1} [ p _ {l} < P (\\mathbf {Q} _ {\\mathbf {n}}) ], \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.337, + 0.469, + 0.488 + ], + "angle": 0, + "content": "where higher probabilities than \\( p_l \\) are selected as negative pseudo labels from uncertain predictions and the uncertainty is measured by the entropy of predictions. Here, \\( p_l \\) represents a threshold in negative pseudo labeling, and \\( \\hat{\\mathbf{L}}_{\\mathbf{n}} \\) denotes a negative pseudo label, which is a vector whose elements larger than \\( p_l \\) have a value -1 and otherwise 0. Different from existing negative learning methods [22, 23, 33] that select negative labels from all noisy labels, TDA selects negative pseudo labels from uncertain predictions to avoid bias to the data with certain predictions." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.488, + 0.469, + 0.549 + ], + "angle": 0, + "content": "When constructing the negative cache, the testing feature \\( f_{\\mathrm{test}} \\) will be included in negative cache if it satisfies the condition \\( \\gamma (f_{\\mathrm{test}}) \\): the entropy of the prediction is in the specified interval between \\( \\tau_{l} \\) and \\( \\tau_{h} \\):" + }, + { + "type": "equation", + "bbox": [ + 0.157, + 0.555, + 0.469, + 0.573 + ], + "angle": 0, + "content": "\\[\n\\gamma \\left(f _ {\\text {t e s t}}\\right): \\tau_ {l} < \\mathrm {H} \\left(f _ {\\text {t e s t}} \\mathbf {W} _ {c} ^ {T}\\right) < \\tau_ {h}. \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.579, + 0.469, + 0.715 + ], + "angle": 0, + "content": "This condition is designed to mitigate the risk of prediction errors due to high entropy or be biased to certain predictions (characterized by very low entropy), by incorporating test samples that exhibit a moderate degree of prediction uncertainty. Once the \\(\\gamma(f_{\\mathrm{test}})\\) check is completed, the remaining steps for collecting uncertain samples in the negative cache follow the same two conditions designed in the positive cache. Similar to the positive cache, TDA also limits the shot capacity \\(\\tilde{k}\\) in the negative cache." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.715, + 0.469, + 0.775 + ], + "angle": 0, + "content": "During test-time adaption, the testing features \\( f_{\\mathrm{test}} \\) can be quickly adapted to target domains by retrieving the knowledge from \\( \\{\\mathbf{Q}_{\\mathbf{n}}, \\hat{\\mathbf{L}}_{\\mathbf{n}}\\} \\) in the negative cache and the adapted prediction can be obtained as:" + }, + { + "type": "equation", + "bbox": [ + 0.165, + 0.781, + 0.469, + 0.8 + ], + "angle": 0, + "content": "\\[\nP _ {\\text {n e g}} \\left(f _ {\\text {t e s t}}\\right) = - A \\left(f _ {\\text {t e s t}} \\mathbf {Q} _ {\\mathbf {n}} ^ {\\mathbf {T}}\\right) \\hat {\\mathbf {L}} _ {\\mathbf {n}}, \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.806, + 0.469, + 0.866 + ], + "angle": 0, + "content": "where \\( A \\) is the adaptation function defined in Tip-Adapter. The predictions of TDA can be formulated by combining the negative cache, the positive cache and the pre-trained CLIP model together as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.086, + 0.873, + 0.469, + 0.892 + ], + "angle": 0, + "content": "\\[\nP _ {\\mathrm {T D A}} \\left(f _ {\\text {t e s t}}\\right) = f _ {\\text {t e s t}} \\mathbf {W} _ {c} ^ {T} + P _ {\\text {p o s}} \\left(f _ {\\text {t e s t}}\\right) + P _ {\\text {n e g}} \\left(f _ {\\text {t e s t}}\\right). \\tag {7}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.091, + 0.845, + 0.108 + ], + "angle": 0, + "content": "3.3. Relationship with TPT and Tip-Adapter" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.12, + 0.892, + 0.259 + ], + "angle": 0, + "content": "Both TPT and TDA are designed to handle the challenge of adapting models to test data that have distributional discrepancies from the training data. TPT trains a learnable prompt \\(\\mathbf{p}_c\\) in Eq. (1) for each test sample with a large number of augmentations and such training process requires backpropagation and is computationally intensive. In contrast, our TDA is training-free as both positive cache \\(\\{\\mathbf{Q_p},\\hat{\\mathbf{L}}_{\\mathbf{p}}\\}\\) and negative cache \\(\\{\\mathbf{Q_n},\\hat{\\mathbf{L}}_{\\mathbf{n}}\\}\\) are non-parametric, which is super-efficient without incurring any backpropagation." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.264, + 0.892, + 0.522 + ], + "angle": 0, + "content": "Our TDA employs a cache model that shares similarities with the Tip-Adapter, where features and labels are stored as key-value pairs in a memory cache. One notable distinction between the Tip-Adapter and TDA lies in the type of cache used. Specifically, the Tip-Adapter relies on a static cache \\(\\{\\mathbf{F}_{\\mathrm{train}},\\mathbf{L}_{\\mathrm{train}}\\}\\) because it is designed for a supervised adaptation setting, where the ground-truth labels \\(\\mathbf{L}_{\\mathrm{train}}\\) are predetermined and readily available. In contrast, our TDA introduces a new dynamic cache \\(\\{\\mathbf{Q_p},\\hat{\\mathbf{L}}_p\\}\\) designed for a test-time adaptation setting, where the pseudo labels \\(\\hat{\\mathbf{L}}_{\\mathbf{p}}\\) are generated on-the-fly from a stream of test samples. Furthermore, TDA incorporates a novel negative cache model \\(\\{\\mathbf{Q_n},\\hat{\\mathbf{L}}_n\\}\\) that enhances testing predictions by utilizing indirect knowledge that a test image does not belong to certain negative classes. By combining positive and negative caches, our proposed TDA is more robust to noisy pseudo labels and can generalize well to testing data." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.554, + 0.633, + 0.571 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.585, + 0.691, + 0.601 + ], + "angle": 0, + "content": "4.1. Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.614, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Benchmarks. We conducted main experiments on two benchmarks: out-of-distribution (OOD) benchmark and cross-domain benchmark, both applied in the previous work [35] that adapts vision-language models in test time. The OOD benchmark serves as a measure of the robustness of our approach by involving assessment on 4 out-of-distribution datasets derived from ImageNet [5]: ImageNet-A [16], ImageNet-V2 [32], ImageNet-R [15], and ImageNet-S [41]. This benchmark is specifically designed to evaluate a model's capacity to generalize to new and unseen data. The cross-domain benchmark, on the other hand, is involved to evaluate the model's performance across 10 diverse image classification datasets, each from a distinct domain with different classes: Aircraft [26], Caltech101 [8], Cars [24], DTD [4], EuroSAT [14], Flower102 [28], Food101 [1], Pets [30], SUN397 [42], and UCF101 [36]. This benchmark provides a comprehensive evaluation of the model's adaptability during test time across various class spaces." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "14166" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.082, + 0.089, + 0.89, + 0.342 + ], + "angle": 0, + "content": "
MethodImageNetImageNet-AImageNet-V2ImageNet-RImageNet-SAverageOOD Average
CLIP-ResNet-5059.8123.2452.9160.7235.4846.4343.09
CoOp63.3323.0655.4056.6034.6746.6142.43
CoCoOp62.8123.3255.7257.7434.4846.8142.82
Tip-Adapter62.0323.1353.9760.3535.7447.0443.30
TPT60.7426.6754.7059.1135.0947.2643.89
DiffTPT60.8031.0655.8058.8037.1048.7145.69
TDA (Ours)61.3530.2955.5462.5838.1249.5846.63
CLIP-ViT-B/1668.3449.8961.8877.6548.2461.2059.42
CoOp71.5149.7164.2075.2147.9961.7259.28
CoCoOp71.0250.6364.0776.1848.7562.1359.91
Tip-Adapter70.7551.0463.4177.7648.8862.3760.27
TPT68.9854.7763.4577.0647.9462.4460.81
DiffTPT70.3055.6865.1075.0046.8062.2860.52
TDA (Ours)69.5160.1164.6780.2450.5465.0163.89
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.352, + 0.893, + 0.437 + ], + "angle": 0, + "content": "Table 1. Results on the OOD Benchmark. Our TDA is compared with several state-of-the-art methods designed for vision-language models: the baseline method CLIP, three train-time adaptation methods (i.e., CoOp, CoCoOp, and Tip-Adapter), and two test-time adaptation methods (i.e., TPT and DiffTPT). All the compared methods are built upon CLIP-ResNet-50 or CLIP-ViT-B/16 baselines. The two evaluation metrics Average and OOD Average are calculated by taking the mean accuracy across all five datasets and four OOD datasets excluding ImageNet. The results of CLIP, CoOp, CoCoOp, and TPT are obtained from the TPT paper, the results of DiffTPT are obtained from the DiffTPT paper, while the results of Tip-Adapter are reproduced using the official codes." + }, + { + "type": "table", + "bbox": [ + 0.086, + 0.461, + 0.462, + 0.552 + ], + "angle": 0, + "content": "
MethodTesting TimeAccuracyGain
CLIP-ResNet-5012min59.810
TPT12h 50min60.74+0.93
DiffTPT34h 45min60.80+0.99
TDA (Ours)16min61.35+1.54
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.562, + 0.47, + 0.66 + ], + "angle": 0, + "content": "Table 2. Comparisons of our TDA with CLIP-ResNet-50, TPT, and DiffTPT in terms of efficiency (Testing Time) and effectiveness (Accuracy). The final column shows the accuracy gain relative to the baseline CLIP. Note that the testing time of DiffTPT does not include the duration required for the image generation process with pre-trained diffusion models, which is an additional time-consuming factor during the testing phase." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.689, + 0.471, + 0.901 + ], + "angle": 0, + "content": "Implementation details. All the models in our experiments are built upon the pre-trained CLIP model [31] that consists of an image encoder and a text encoder. The image encoder can be either ResNet [13] or Vision Transformer [7], while the text encoder is Transformer [39]. Test-time adaptation is set for single-image scenarios, using a batch size of 1. We conduct a search for all our hyperparameters using a single ImageNet validation set. The threshold \\( p_l \\) for negative pseudo-labeling in Eq 4 is set as 0.03. The upper and lower thresholds \\( [\\tau_l, \\tau_h] \\) for testing feature selection in Eq 5 are set as [0.2, 0.5]. Once searched, these hyperparameters are fixed and evaluated across various new datasets. To avoid incurring backpropagation when using learnable prompt, we follow [31] to use hand-crafted" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.463, + 0.892, + 0.522 + ], + "angle": 0, + "content": "prompts. We use top-1 accuracy \\((\\%)\\), a standard classification criterion, as our evaluation metric. All the experiments are conducted using a single NVIDIA Quadro RTX 6000 GPU." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.531, + 0.803, + 0.547 + ], + "angle": 0, + "content": "4.2. Comparisons with State-of-the-art" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.554, + 0.892, + 0.901 + ], + "angle": 0, + "content": "In this section, we compare our proposed TDA with several state-of-the-art methods, including CLIP [31], three train-time adaptation methods, i.e., CoOp [50], CoCoOp [49], and Tip-Adapter [47], as well as two existing test-time adaptation methods TPT [35] and DiffTPT [9], all of which are designed for vision-language models. Specifically, CLIP is evaluated using an ensemble of 80 handcrafted prompts as in [31]. All train-time adaptation methods are trained on the ImageNet train set with 16 shots per class and tested on other datasets as in [35]. CoOp utilizes a learnable context module of size 4 that is fine-tuned for specific downstream tasks, while CoCoOp is the generalized version of CoOp with the input-conditional token for image features. Tip-Adapter employs the optimal hyperparameters obtained from the ImageNet validation set during evaluation. We would like to note that Tip-Adapter is unable to handle new classes during testing, limiting its implementation to OOD benchmark evaluation where the training classes encompass all the testing classes. Different from train-time adaptation methods, test-time adaptation methods (i.e., TPT, DiffTPT, and our TDA) do not utilize the ImageNet train set. Instead, they are fine-tuned with target datasets using a stream of unlabeled test samples. Follow" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "14167" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.082, + 0.089, + 0.89, + 0.293 + ], + "angle": 0, + "content": "
MethodAircraftCaltech101CarsDTDEuroSATFlower102Food101PetsSUN397UCF101Average
CLIP-ResNet-5016.1187.2655.8940.3725.7962.7774.8282.9760.8559.4856.63
CoOp15.1286.5355.3237.2926.2061.5575.5987.0058.1559.0556.18
CoCoOp14.6187.3856.2238.5328.7365.5776.2088.3959.6157.1057.23
TPT17.5887.0258.4640.8428.3362.6974.8884.4961.4660.8257.66
DiffTPT17.6086.8960.7140.7241.0463.5379.2183.4062.7262.6759.85
TDA (Ours)17.6189.7057.7843.7442.1168.7477.7586.1862.5364.1861.03
CLIP-ViT-B/1623.2293.5566.1145.0450.4266.9982.8686.9265.6365.1664.59
CoOp18.4793.7064.5141.9246.3968.7185.3089.1464.1566.5563.88
CoCoOp22.2993.7964.9045.4539.2370.8583.9790.4666.8968.4464.63
TPT24.7894.1666.8747.7542.4468.9884.6787.7965.5068.0465.10
DiffTPT25.6092.4967.0147.0043.1370.1087.2388.2265.7462.6765.47
TDA (Ours)23.9194.2467.2847.4058.0071.4286.1488.6367.6270.6667.53
" + }, + { + "type": "table_caption", + "bbox": [ + 0.08, + 0.303, + 0.893, + 0.374 + ], + "angle": 0, + "content": "Table 3. Results on the Cross-Domain Benchmark. Our TDA is compared with several state-of-the-art methods designed for vision-language models: the baseline method CLIP, two train-time adaptation methods (i.e., CoOp and CoCoOp), and two test-time adaptation methods (i.e., TPT and DiffTPT). Note that Tip-Adapter is unable to be evaluated on the Cross-Domain Benchmark as it cannot handle new classes during testing. The evaluation metric Average is calculated by taking the mean accuracy across all ten datasets. The results of CLIP, CoOp, CoCoOp, and TPT are obtained from the TPT paper, while the results of DiffTPT are obtained from the DiffTPT paper." + }, + { + "type": "text", + "bbox": [ + 0.081, + 0.412, + 0.468, + 0.457 + ], + "angle": 0, + "content": "ing TPT and DiffTPT, we compare TDA with the state-of-the-art over two public benchmarks: OOD benchmark and cross-domain benchmark." + }, + { + "type": "text", + "bbox": [ + 0.081, + 0.478, + 0.468, + 0.688 + ], + "angle": 0, + "content": "Results on the OOD Benchmark. We first compare TDA with state-of-the-art methods over the OOD benchmark. Table 1 presents the experimental results, highlighting the superior performance of the proposed TDA compared to both TPT and DiffTPT across various OOD datasets derived from ImageNet. Specifically, TDA outperforms TPT on both ResNet-50 and ViT-B/16 architectures, improving OOD accuracy by \\(2.74\\%\\) and \\(3.08\\%\\) on average, respectively. Furthermore, compared to DiffTPT, TDA exhibits an average accuracy improvement of \\(0.94\\%\\) and \\(3.37\\%\\) in the OOD benchmark for ResNet-50 and ViT-B/16, respectively. These results validate the effectiveness of TDA in enhancing test-time adaptation performance on various OOD test datasets." + }, + { + "type": "text", + "bbox": [ + 0.081, + 0.689, + 0.468, + 0.9 + ], + "angle": 0, + "content": "In order to provide a more comprehensive evaluation of our proposed method's efficiency and effectiveness, we compared it with the baseline CLIP-ResNet-50 and two existing test-time adaptation methods (i.e., TPT and DiffTPT). This comparison encompasses both testing time and testing accuracy, and the corresponding results are shown in Table 2. This evaluation is performed on the ImageNet validation dataset, which consists of 50,000 images, using a single NVIDIA Quadro RTX 6000 GPU. When compared to the baseline CLIP-ResNet-50, the proposed TDA demonstrates a significant improvement in testing accuracy \\((+1.54\\%)\\), with only a minimal sacrifice in testing efficiency (requiring an additional 4min). In comparison to TPT and DiffTPT, the proposed TDA demonstrates not only superior" + }, + { + "type": "text", + "bbox": [ + 0.504, + 0.412, + 0.891, + 0.563 + ], + "angle": 0, + "content": "testing accuracy but also significantly improved efficiency. It reduces the testing time dramatically from \\(12\\mathrm{h}50\\mathrm{min}\\) by TPT and even more from \\(34\\mathrm{h}45\\mathrm{min}\\) by DiffTPT, down to just 16 minutes. Without including the image generation time, DiffTPT consumes clearly more test time than TPT as it involves a time-consuming multi-step prompt updating process whereas TPT requires a single step only. The experimental results strongly validate the effectiveness and efficiency of our proposed method, establishing its suitability for real-world applications." + }, + { + "type": "text", + "bbox": [ + 0.504, + 0.565, + 0.891, + 0.79 + ], + "angle": 0, + "content": "We then compare TDA with state-of-the-art methods over cross-domain benchmark. The results, presented in Table 3, demonstrate that TDA not only surpasses the performance of the TPT method but also shows a significant advantage over its improvement method DiffTPT. Specifically, when utilizing CLIP-ResNet-50 and CLIP-ViT-B/16 as the backbone, TDA achieves an improvement in average accuracy over TPT by \\(3.37\\%\\) and \\(2.43\\%\\), respectively. These improvements, along with a \\(1.18\\%\\) and \\(2.06\\%\\) gain over DiffTPT for the respective backbones, further verify the effectiveness of TDA in adapting to diverse class datasets during test time. This attribute holds significant value for vision-language models such as CLIP, as it enables them to classify arbitrary classes in image classification without the need for additional training." + }, + { + "type": "title", + "bbox": [ + 0.504, + 0.802, + 0.663, + 0.817 + ], + "angle": 0, + "content": "4.3. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.504, + 0.826, + 0.891, + 0.899 + ], + "angle": 0, + "content": "In this section, we perform ablation studies to examine the effectiveness of our designs. All the ablation studies are conducted over the ImageNet dataset, where TDA can achieve an accuracy of \\(61.35\\%\\) under default settings. TDA consists of a Positive Cache and a Negative Cache, which" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.947, + 0.518, + 0.957 + ], + "angle": 0, + "content": "14168" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.092, + 0.468, + 0.318 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.326, + 0.472, + 0.368 + ], + "angle": 0, + "content": "Figure 3. Ablation studies on two cache designs in TDA: Positive Cache and Negative Cache. All the models are built upon the baseline model CLIP-ResNet-50." + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.4, + 0.47, + 0.657 + ], + "angle": 0, + "content": "perform positive and negative pseudo labeling within our designed dynamic adapter, respectively. We first assess the efficacy of the two cache designs in TDA. As shown in Figure 3, both Positive Cache and Negative Cache significantly surpass the baseline model CLIP, demonstrating that test-time adaptation can be improved by introducing a dynamic adapter with either positive pseudo labeling or negative pseudo labeling. Besides, the two cache designs in TDA can complement each other as TDA (i.e., the combination of the two designs) clearly outperforms either Positive Cache or Negative Cache on the challenging ImageNet dataset. Moreover, we extend our ablation studies to the cross-domain benchmark, and results show that the positive cache achieved \\(60.38\\%\\) accuracy, the negative cache \\(60.11\\%\\), while their combination yielded a \\(61.03\\%\\) accuracy, thereby highlighting the significance of each type of cache in enhancing TDA's performance." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.66, + 0.471, + 0.902 + ], + "angle": 0, + "content": "We proceed to perform parameter studies on the shot capacity, which refers to the maximum number of key-value pairs per class, in both Positive Cache and Negative Cache models. These studies aims to find the optimal balance between the diversity and accuracy of the key-value pairs. Figure 4 shows that the performance of both cache models is significantly affected when the shot capacity is either too low or too high. We find that the shot capacity is set as 3 for the Positive Cache and 2 for the Negative Cache yields the best performance. This is because an appropriate shot capacity ensures both high-quality pseudo labels (paired values) and diverse image features (paired keys) in the cache models. The negative cache is sensitive to shot numbers due to its role in storing probabilities for various negative pseudo labels, each representing a class to be excluded from the model. Contrary to the intuition that a larger negative" + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.092, + 0.891, + 0.322 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.329, + 0.895, + 0.358 + ], + "angle": 0, + "content": "Figure 4. Parameter studies on the Shot Capacity in Positive Cache and Negative Cache." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.384, + 0.895, + 0.537 + ], + "angle": 0, + "content": "cache might be beneficial, a larger negative cache leads to more high-entropy, noisier pseudo labels, as highlighted in self-training studies like [11], thereby lowering the performance. Conversely, the positive cache, which stores single and high-confidence prediction, is less affected by variations in shot capacity, thereby maintaining consistent accuracy across different shot capacities. To facilitate practical applications, the shot capacity settings are fixed and directly applied to new datasets without the need for additional parameter adjustments." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.549, + 0.62, + 0.566 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.574, + 0.895, + 0.832 + ], + "angle": 0, + "content": "In this work, we have presented TDA, a dynamic adapter for efficient and effective test-time adaptation of vision-language models. The proposed method employs a key-value cache, which maintains a dynamic queue with test-sample features as keys and corresponding few-shot pseudo labels as values, allowing for gradual adaptation to test data through progressive pseudo label improvement. Moreover, TDA introduces a negative cache to mitigate the undesirable effects of noisy pseudo labels by assigning negative pseudo labels to certain classes when the model is uncertain about its predictions. The results of extensive experiments over two benchmarks demonstrate that TDA outperforms state-of-the-art test-time adaptation methods while significantly reducing testing time. This work contributes to the research and application values of test-time adaptation and presents a promising solution to the efficiency issue of test-time adaptation of vision-language models." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.845, + 0.661, + 0.862 + ], + "angle": 0, + "content": "Acknowledgement" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.87, + 0.892, + 0.903 + ], + "angle": 0, + "content": "This study is supported under the Mohamed bin Zayed University of Artificial Intelligence." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "14169" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.175, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.169 + ], + "angle": 0, + "content": "[1] Lukas Bossard, Matthieu Guillaumin, and Luc Van Gool. Food-101 – mining discriminative components with random forests. In European Conference on Computer Vision, 2014. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.173, + 0.472, + 0.227 + ], + "angle": 0, + "content": "[2] Malik Boudiaf, Romain Mueller, Ismail Ben Ayed, and Luca Bertinetto. Parameter-free online test-time adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8344-8353, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.23, + 0.47, + 0.285 + ], + "angle": 0, + "content": "[3] Dian Chen, Dequan Wang, Trevor Darrell, and Sayna Ebrahimi. Contrastive test-time adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 295-305, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.287, + 0.471, + 0.342 + ], + "angle": 0, + "content": "[4] M. Cimpoi, S. Maji, I. Kokkinos, S. Mohamed, , and A. Vedaldi. Describing textures in the wild. In Proceedings of the IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2014. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.345, + 0.47, + 0.384 + ], + "angle": 0, + "content": "[5] J. Deng, W. Dong, R. Socher, L.-J. Li, K. Li, and L. Fei-Fei. ImageNet: A Large-Scale Hierarchical Image Database. In CVPR09, 2009. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.388, + 0.47, + 0.427 + ], + "angle": 0, + "content": "[6] Karan Desai and Justin Johnson. VirTex: Learning Visual Representations from Textual Annotations. In CVPR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.431, + 0.47, + 0.513 + ], + "angle": 0, + "content": "[7] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.516, + 0.47, + 0.571 + ], + "angle": 0, + "content": "[8] Li Fei-Fei, Rob Fergus, and Pietro Perona. Learning generative visual models from few training examples: An incremental bayesian approach tested on 101 object categories. In CVPR Workshops, 2004. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.574, + 0.47, + 0.642 + ], + "angle": 0, + "content": "[9] Chun-Mei Feng, Kai Yu, Yong Liu, Salman Khan, and Wangmeng Zuo. Diverse data augmentation with diffusions for effective test-time prompt tuning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2704-2714, 2023. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.645, + 0.469, + 0.699 + ], + "angle": 0, + "content": "[10] Peng Gao, Shijie Geng, Renrui Zhang, Teli Ma, Rongyao Fang, Yongfeng Zhang, Hongsheng Li, and Yu Qiao. Clip-adapter: Better vision-language models with feature adapters. arXiv preprint arXiv:2110.04544, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.702, + 0.469, + 0.742 + ], + "angle": 0, + "content": "[11] Yves Grandvalet and Yoshua Bengio. Semi-supervised learning by entropy minimization. Advances in neural information processing systems, 17, 2004. 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.745, + 0.469, + 0.8 + ], + "angle": 0, + "content": "[12] Edouard Grave, Moustapha M Cisse, and Armand Joulin. Unbounded cache model for online language modeling with open vocabulary. Advances in neural information processing systems, 30, 2017. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.802, + 0.469, + 0.841 + ], + "angle": 0, + "content": "[13] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, 2016. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.846, + 0.469, + 0.9 + ], + "angle": 0, + "content": "[14] Patrick Helber, Benjamin Bischke, Andreas Dengel, and Damian Borth. Eurosat: A novel dataset and deep learning benchmark for land use and land cover classification. IEEE J. Sel. Top. Appl. Earth Obs. Remote. Sens., 2019. 5" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.472, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.174 + ], + "angle": 0, + "content": "[15] Dan Hendrycks, Steven Basart, Norman Mu, Saurav Kadavath, Frank Wang, Evan Dorundo, Rahul Desai, Tyler Zhu, Samyak Parajuli, Mike Guo, Dawn Song, Jacob Steinhardt, and Justin Gilmer. The many faces of robustness: A critical analysis of out-of-distribution generalization. In ICCV, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.178, + 0.892, + 0.218 + ], + "angle": 0, + "content": "[16] Dan Hendrycks, Kevin Zhao, Steven Basart, Jacob Steinhardt, and Dawn Song. Natural adversarial examples. In CVPR, pages 15262-15271, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.221, + 0.892, + 0.289 + ], + "angle": 0, + "content": "[17] Xiaowei Hu, Zhe Gan, Jianfeng Wang, Zhengyuan Yang, Zicheng Liu, Yumao Lu, and Lijuan Wang. Scaling up vision-language pre-training for image captioning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17980-17989, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.291, + 0.892, + 0.345 + ], + "angle": 0, + "content": "[18] Yusuke Iwasawa and Yutaka Matsuo. Test-time classifier adjustment module for model-agnostic domain generalization. Advances in Neural Information Processing Systems, 34:2427-2440, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.348, + 0.892, + 0.429 + ], + "angle": 0, + "content": "[19] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In International Conference on Machine Learning, pages 4904-4916. PMLR, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.433, + 0.892, + 0.472 + ], + "angle": 0, + "content": "[20] Jeff Johnson, Matthijs Douze, and Hervé Jégou. Billionscale similarity search with gpus. IEEE Transactions on Big Data, 7(3):535-547, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.476, + 0.892, + 0.531 + ], + "angle": 0, + "content": "[21] Urvashi Khandelwal, Omer Levy, Dan Jurafsky, Luke Zettlemoyer, and Mike Lewis. Generalization through memorization: Nearest neighbor language models. In International Conference on Learning Representations, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.533, + 0.894, + 0.587 + ], + "angle": 0, + "content": "[22] Youngdong Kim, Junho Yim, Juseung Yun, and Junmo Kim. NInl: Negative learning for noisy labels. In Proceedings of the IEEE/CVF international conference on computer vision, pages 101-110, 2019. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.589, + 0.892, + 0.656 + ], + "angle": 0, + "content": "[23] Youngdong Kim, Juseung Yun, Hyounguk Shon, and Junmo Kim. Joint negative and positive learning for noisy labels. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9442-9451, 2021. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.66, + 0.892, + 0.701 + ], + "angle": 0, + "content": "[24] Jonathan Krause, Michael Stark, Jia Deng, and Li Fei-Fei. 3d object representations for fine-grained categorization. In ICCV Workshops, 2013. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.703, + 0.892, + 0.758 + ], + "angle": 0, + "content": "[25] Xin Li, Dongze Lian, Zhihe Lu, Jiawang Bai, Zhibo Chen, and Xinchao Wang. Graphadapter: Tuning vision-language models with dual knowledge graph. Advances in Neural Information Processing Systems, 36, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.76, + 0.892, + 0.8 + ], + "angle": 0, + "content": "[26] S. Maji, J. Kannala, E. Rahtu, M. Blaschko, and A. Vedaldi. Fine-grained visual classification of aircraft. Technical report, 2013. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.802, + 0.892, + 0.843 + ], + "angle": 0, + "content": "[27] Stephen Merity, Caiming Xiong, James Bradbury, and Richard Socher. Pointer sentinel mixture models. arXiv preprint arXiv:1609.07843, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.846, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[28] Maria-Elena Nilsback and Andrew Zisserman. Automated flower classification over a large number of classes. In Indian Conference on Computer Vision, Graphics and Image Processing, 2008. 5" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "14170" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.469, + 0.133 + ], + "angle": 0, + "content": "[29] Emin Orhan. A simple cache model for image recognition. Advances in Neural Information Processing Systems, 31, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.136, + 0.469, + 0.163 + ], + "angle": 0, + "content": "[30] Omkar M. Parkhi, Andrea Vedaldi, Andrew Zisserman, and C. V. Jawahar. Cats and dogs. In CVPR, 2012. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.165, + 0.469, + 0.247 + ], + "angle": 0, + "content": "[31] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In ICML, 2021. 1, 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.25, + 0.469, + 0.29 + ], + "angle": 0, + "content": "[32] Benjamin Recht, Rebecca Roelofs, Ludwig Schmidt, and Vaishaal Shankar. Do imagenet classifiers generalize to imageNet? In ICML, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.293, + 0.469, + 0.36 + ], + "angle": 0, + "content": "[33] Mamshad Nayeem Rizve, Kevin Duarte, Yogesh S Rawat, and Mubarak Shah. In defense of pseudo-labeling: An uncertainty-aware pseudo-label selection framework for semi-supervised learning. arXiv preprint arXiv:2101.06329, 2021. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.363, + 0.469, + 0.431 + ], + "angle": 0, + "content": "[34] Steffen Schneider, Evgenia Rusak, Luisa Eck, Oliver Bringmann, Wieland Brendel, and Matthias Bethge. Improving robustness against common corruptions by covariate shift adaptation. Advances in Neural Information Processing Systems, 33:11539-11551, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.434, + 0.469, + 0.503 + ], + "angle": 0, + "content": "[35] Manli Shu, Weili Nie, De-An Huang, Zhiding Yu, Tom Goldstein, Anima Anandkumar, and Chaowei Xiao. Test-time prompt tuning for zero-shot generalization in vision-language models. In Advances in Neural Information Processing Systems, 2022. 1, 2, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.505, + 0.469, + 0.545 + ], + "angle": 0, + "content": "[36] Khurram Soomro, Amir Roshan Zamir, and Mubarak Shah. UCF101: A dataset of 101 human actions classes from videos in the wild. CoRR, abs/1212.0402, 2012. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.548, + 0.469, + 0.602 + ], + "angle": 0, + "content": "[37] Yu Sun, Xiaolong Wang, Zhuang Liu, John Miller, Alexei A. Efros, and Moritz Hardt. Test-time training with self-supervision for generalization under distribution shifts. In ICML, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.605, + 0.469, + 0.701 + ], + "angle": 0, + "content": "[38] Thomas Varsavsky, Maurizio Orbes-Arteaga, Carole H Sudre, Mark S Graham, Parashkev Nachev, and M Jorge Cardoso. Test-time unsupervised domain adaptation. In Medical Image Computing and Computer Assisted Intervention—MICCAI 2020: 23rd International Conference, Lima, Peru, October 4–8, 2020, Proceedings, Part I 23, pages 428–436. Springer, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.704, + 0.469, + 0.758 + ], + "angle": 0, + "content": "[39] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.76, + 0.469, + 0.801 + ], + "angle": 0, + "content": "[40] Dequan Wang, Evan Shelhamer, Shaoteng Liu, Bruno A. Olshausen, and Trevor Darrell. Tent: Fully test-time adaptation by entropy minimization. In ICLR, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.804, + 0.469, + 0.844 + ], + "angle": 0, + "content": "[41] Haohan Wang, Songwei Ge, Zachary Lipton, and Eric P Xing. Learning robust global representations by penalizing local predictive power. In NeurIPS, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.847, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[42] J. Xiao, J. Hays, K. A. Ehinger, A. Oliva, and A. Torralba. Sun database: Large-scale scene recognition from abbey to zoo. In 2010 IEEE Computer Society Conference on Computer Vision and Pattern Recognition, 2010. 5" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.469, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.175 + ], + "angle": 0, + "content": "[43] Jinyu Yang, Jiali Duan, Son Tran, Yi Xu, Sampath Chanda, Liquun Chen, Belinda Zeng, Trishul Chilimbi, and Junzhou Huang. Vision-language pre-training with triple contrastive learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15671-15680, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.178, + 0.892, + 0.232 + ], + "angle": 0, + "content": "[44] Tao Yu, Zhihe Lu, Xin Jin, Zhibo Chen, and Xinchao Wang. Task residual for tuning vision-language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10899-10909, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.234, + 0.892, + 0.301 + ], + "angle": 0, + "content": "[45] Marvin Zhang, Henrik Marklund, Nikita Dhawan, Abhishek Gupta, Sergey Levine, and Chelsea Finn. Adaptive risk minimization: Learning to adapt to domain shift. Advances in Neural Information Processing Systems, 34:23664-23678, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.305, + 0.892, + 0.357 + ], + "angle": 0, + "content": "[46] Marvin Mengxin Zhang, Sergey Levine, and Chelsea Finn. MEMO: Test time robustness via adaptation and augmentation. In Advances in Neural Information Processing Systems, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.361, + 0.892, + 0.43 + ], + "angle": 0, + "content": "[47] Renrui Zhang, Wei Zhang, Rongyao Fang, Peng Gao, Kunchang Li, Jifeng Dai, Yu Qiao, and Hongsheng Li. Tip-adapter: Training-free adaption of clip for few-shot classification. In European Conference on Computer Vision, pages 493-510. Springer, 2022. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.432, + 0.892, + 0.514 + ], + "angle": 0, + "content": "[48] Renrui Zhang, Xiangfei Hu, Bohao Li, Siyuan Huang, Hanqiu Deng, Yu Qiao, Peng Gao, and Hongsheng Li. Prompt, generate, then cache: Cascade of foundation models makes strong few-shot learners. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15211-15222, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.517, + 0.892, + 0.584 + ], + "angle": 0, + "content": "[49] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Conditional prompt learning for vision-language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16816-16825, 2022. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.586, + 0.892, + 0.64 + ], + "angle": 0, + "content": "[50] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Learning to prompt for vision-language models. International Journal of Computer Vision, 130(9):2337-2348, 2022. 2, 6" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.64 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "14171" + } + ] +] \ No newline at end of file diff --git a/2024/Efficient Test-Time Adaptation of Vision-Language Models/f86a8c24-7089-4003-88ca-fa1a5c70fe2d_origin.pdf b/2024/Efficient Test-Time Adaptation of Vision-Language Models/f86a8c24-7089-4003-88ca-fa1a5c70fe2d_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..a7e40772251ad2008e2a294caa1a6dd287292a61 --- /dev/null +++ b/2024/Efficient Test-Time Adaptation of Vision-Language Models/f86a8c24-7089-4003-88ca-fa1a5c70fe2d_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1f3f902ae565a3e059c8c1b7d7869c57b85bc399b0cbe437604aa307a1654ef9 +size 1124422 diff --git a/2024/Efficient Test-Time Adaptation of Vision-Language Models/full.md b/2024/Efficient Test-Time Adaptation of Vision-Language Models/full.md new file mode 100644 index 0000000000000000000000000000000000000000..3e5141310266e99eaa9afdb16412d5d397393505 --- /dev/null +++ b/2024/Efficient Test-Time Adaptation of Vision-Language Models/full.md @@ -0,0 +1,245 @@ +# Efficient Test-Time Adaptation of Vision-Language Models + +Adilbek Karmanov $^{1*}$ Dayan Guan $^{2*}$ Shijian Lu $^{1,2\dagger}$ Abdulmotaleb El Saddik $^{1,3}$ Eric Xing $^{1,4}$ + +$^{1}$ Mohamed bin Zayed University of Artificial Intelligence $^{2}$ Nanyang Technological University $^{3}$ University of Ottawa $^{4}$ Carnegie Mellon University + +# Abstract + +Test-time adaptation with pre-trained vision-language models has attracted increasing attention for tackling distribution shifts during the test time. Though prior studies have achieved very promising performance, they involve intensive computation which is severely unaligned with test-time adaptation. We design TDA, a training-free dynamic adapter that enables effective and efficient test-time adaptation with vision-language models. TDA works with a lightweight key-value cache that maintains a dynamic queue with few-shot pseudo labels as values and the corresponding test-sample features as keys. Leveraging the key-value cache, TDA allows adapting to test data gradually via progressive pseudo label refinement which is super-efficient without incurring any backpropagation. In addition, we introduce negative pseudo labeling that alleviates the adverse impact of pseudo label noises by assigning pseudo labels to certain negative classes when the model is uncertain about its pseudo label predictions. Extensive experiments over two benchmarks demonstrate TDA's superior effectiveness and efficiency as compared with the state-of-the-art. The code has been released in https://kdiaaa.github.io/tda/. + +# 1. Introduction + +Recent advances in vision-language models [19, 31, 43] have opened a new door for integrating human language into various computer vision tasks. Take CLIP [31] as an example. It can enable zero-shot image classification by leveraging a shared embedding space that is learnt from web-scale image-text pairs. Within this shared space, images can be directly recognized by matching their features with the text embeddings of CLIP classes. At the other end, CLIP often faces challenges while handling various specific downstream images, especially when the downstream images have clear domain and distribution shifts as compared + +![](images/1d0a13a9b4fa3f9ddd9c9eee1a89ba333532057ee01bec39e001bf87a863ba0b.jpg) +(a) Test-time Prompt Tuning [9, 35] + +![](images/b84cafaa5da2b0e13490f6052a78e44341a30fea0f1aa56e779523647d7ada74.jpg) +(b) Training-free Dynamic Adapter (Ours) +Figure 1. Comparison of our proposed Training-free Dynamic Adapter (TDA) with Test-time Prompt Tuning TPT [35] and its enhancement DiffTPT [9]: both TPT and DiffTPT require significant computational resources to optimize the learnable prompt via backpropagation; TDA is a dynamic cache that is training-free without any backpropagation, making it efficient for test-time adaptation in various real-world scenarios. + +with the CLIP training images. + +Several recent studies [2, 3, 40] introduce a new paradigm called test-time adaptation for mitigating the domain shift. The idea of test-time adaptation is well aligned with real-world scenarios where a model needs to adapt to new environments quickly. Despite its great research and + +application values, test-time adaptation of vision-language models has been largely neglected. Recently, Test-time Prompt Tuning, as introduced in TPT [35] and its enhancement DiffTPT [9], attempts to adapt vision-language models by learning domain-specific prompts from test data. As illustrated in Fig. 1 (a), both TPT and DiffTPT train a learnable prompt for each test sample by feeding its augmented views into the CLIP model for generating predictions and minimizing the marginal entropy of confident predictions. Despite its decent performance, the prompt optimization in both TPT [50] and DiffTPT [9] is computationally intensive which hinders its applications in various real-world scenarios. + +We design a training-free dynamic adapter (TDA) that allows efficient and effective test-time adaptation of vision-language models without requiring any backpropagation in test time. As Fig. 1 (b) shows, TDA constructs a lightweight Dynamic Adapter that keeps a dynamic queue with the pseudo labels of a stream of test samples as values and the corresponding CLIP-extracted features as keys. TDA has two desirable features that make its test-time adaptation highly applicable in real-world scenarios. First, TDA is highly effective, as it improves the quality of pseudo labels via progressive incorporation of test predictions of lower entropy [11, 40]. Second, TDA is very efficient as the key-value cache is non-parametric and requires no backpropagation during testing. Beyond that, TDA cache is lightweight due to the few-shot setup and it can be computed with simple matrix multiplications [12, 21, 29, 47]. + +Note that the performance of TDA depends heavily on the pseudo labels of unlabelled test samples which are often noisy with prediction errors. Inspired by the idea of negative learning [22, 23, 33], we introduce negative pseudo labeling to reduce the impact of noisy estimated labels. Traditional pseudo labeling methods identify the presence of particular classes in unlabeled data, which may result in erroneous pseudo labels being assigned when comparable high probabilities are observed. In contrast, our designed negative pseudo labeling determines the absence of certain classes and can provide more accurate pseudo labels as the probabilities of these complementary classes are very low. Concretely, we construct an additional TDA cache that stores negative pseudo labels to complement the positive TDA cache. By combining positive and negative caches, TDA is more tolerant to noisy pseudo labels and can better generalize to testing data. Extensive experiments over two widely adopted test-time adaptation benchmarks show that TDA outperforms the state-of-the-art by large margins while significantly reducing the testing time from over 12 hours to 16 minutes on the ImageNet dataset. + +In summary, the contributions of this work are threefold. First, we design a training-free dynamic adapter (TDA) that can achieve test-time adaptation of vision-language models + +efficiently and effectively. To the best of our knowledge, this is the first work that investigates the efficiency issue of test-time adaptation of vision-language models. Second, we introduce negative pseudo labeling to alleviate the adverse impact of pseudo label noises which makes TDA more robust to pseudo label noises and generalizable to testing data. Third, we evaluate TDA extensively over two benchmarks, and experiments show that TDA achieves superior accuracy and efficiency compared with the state-of-the-art. + +# 2. Related Work + +Vision-language models [6, 17, 19, 31, 43] have demonstrated significant potential in learning semantic representations effectively by undergoing extensive training on image-text data. CLIP [31] stands out among these models for its ability to establish links between visual and textual representations, which enables it to achieve impressive zero-shot results on various downstream tasks. To enhance the transfer learning capability of the CLIP model in the downstream classification tasks, researchers have proposed integrating language prompt learners such as CoOp [50] and CoCoOp [49], as well as vision adapters such as CLIP-Adapter [10], Tip-Adapter [47], CaFo [48], TaskRes [44] and GraphAdapter [25]. Although these methods have shown considerable performance improvements, they typically require a large amount of training data in downstream tasks, making them less practical for real-world scenarios. On other hand, this work focuses on a new paradigm named test-time adaptation without accessing the original training data. + +Test-time adaptation refers to the process of adapting models to testing data that may have distributional differences from the training data. It is particularly beneficial for real-world applications that require models to be deployed in diverse environments, such as autonomous driving in various weather conditions, medical diagnosis in different hospitals, and etc. Several recent works utilize each batch of testing samples to update partial weights [18, 37, 38], normalization statistics [34], or a combination of both [40, 45]. To avoid updating models with multiple testing samples, MEMO [46] proposes enforcing the invariant predictions from different augmentations of each sample in the testing data stream. TPT [35] tackles the same challenge with vision-language models by fine-tuning a learnable prompt with each testing sample. DiffTPT [9] innovates test-time prompt tuning by leveraging pre-trained diffusion models to augment the diversity of test data samples used in TPT. Although TPT [35] and DiffTPT [9] are effective in addressing test-time adaptation of vision-language models, prompt learning is computationally expensive and time-consuming. This paper aims to mitigate the computational efficiency challenges of TPT and DiffTPT through the introduction of + +a cache model. + +Cache model benefits the adaptation techniques by providing efficient inference and non-parametric processing without requiring parameter updates. Unbounded Cache [12] and PSMM [27] have shown promising results in the text generation task by storing a large amount of the training dataset to capture long-term dependencies, however, this approach poses a challenge of memory efficiency during the test-time adaptation. A different technique [20] was proposed to mitigate this limitation by reducing the size of the cache memory through the search for application. Alternatively, Tip-Adapter [47] solves the cache memory problem by only storing few-shot samples per class to create a cache model for vision-language models. Inspired by this work, our dynamic adapter uses the same architecture for the test-time adaptation setting, where there is no access to the source data, and testing samples can only be accessed one by one. To address the lack of access to source data during testing, our adapter collects the most reliable test samples and their pseudo labels to form cache model. + +# 3. Method + +# 3.1. Preliminaries + +CLIP [31] is a vision-language model that performs a proxy task of predicting the correct pairings of text and image. Consider an $N$ -class classification problem where the CLIP's objective in the zero-shot setting is to match images with their most relevant textual descriptions using an image encoder $E_{v}$ and a text encoder $E_{t}$ . To obtain the textual descriptions, $N$ -class names are concatenated with handcrafted prompts and then mapped into the $c$ -channeled text embeddings $\mathbf{W}_{c}$ using the text encoder $E_{t}$ . + +TPT [35] focuses on test-time adaptation of CLIP. In TPT, a prompt tuning method is proposed to learn an adaptive prompt $\mathbf{p}_c$ using individual test samples. A set of augmentation functions $\mathcal{A}$ is used to generate $n$ randomly augmented views $\tilde{x}_{\mathrm{test}} = \mathcal{A}_n(x_{\mathrm{test}})$ of a test sample $x_{\mathrm{test}}$ . The objective of TPT is to reduce variation in the model's predictions across different augmentations $\tilde{f}_{\mathrm{test}} = E_v(\tilde{x}_{\mathrm{test}})$ by minimizing the marginal entropy among the outputs of the augmented views. Furthermore, TPT also includes confidence selection to discard noisy augmentations that could result in ambiguous model predictions, as shown in Figure 1a. This is achieved by filtering out augmented views with high-entropy predictions as: + +$$ +P _ {\mathrm {T P T}} \left(\tilde {f} _ {\text {t e s t}}\right) = \frac {1}{\rho n} \sum_ {i = 1} ^ {n} \mathbb {1} \left[ \mathrm {H} \left(\tilde {f} _ {i} \mathbf {p} _ {c} ^ {T}\right) \leq \tau \right] \tilde {f} _ {i} \mathbf {p} _ {c} ^ {T}, \tag {1} +$$ + +where $\mathrm{H}$ is the self-entropy function of the softmax logits predictions, $\tilde{f}_i\mathbf{p}_c^T$ is the class probabilities vector of size + +$N$ generated from the given $i$ -th augmented view of the test image, and the parameter $\tau$ determines that only $\rho$ -percentile of confident samples with entropy values below this threshold can be selected out of $n$ augmented views. + +Tip-Adapter [47] provides a training-free solution that uses a key-value cache model and integrates knowledge from the pre-trained CLIP model with few-shot labeled samples. The cache model is created using a set of $k$ -shot labeled samples $x_{k}$ from $N$ classes and their corresponding ground-truth labels $y_{N}$ . It can be conceptualized as a linear two-layer model, where the first layer contains train image features $\mathbf{F}_{\mathrm{train}} = E_v(x_k)$ and the second layer consists of one-hot vectors $\mathbf{L}_{\mathrm{train}}$ encoded from the labels $y_{N}$ . Given test image features $f_{\mathrm{test}}$ generated from the CLIP's image encoder $E_{v}$ , the prediction from the cache model can be calculated as follows: + +$$ +P _ {\mathrm {c a c h e}} \left(f _ {\mathrm {t e s t}}\right) = A \left(f _ {\mathrm {t e s t}} \mathbf {F} _ {\mathrm {t r a i n}} ^ {T}\right) \mathbf {L} _ {\mathrm {t r a i n}}, \tag {2} +$$ + +where $A(z) = \alpha \exp(-\beta (1 - z))$ is an adaptation function within a weighting factor $\alpha$ and a sharpness ratio $\beta$ . During inference, the prediction of Tip-Adapter is computed by combining the pre-trained CLIP model and the cache model as: $P_{\mathrm{TA}}(f_{\mathrm{test}}) = P_{\mathrm{cache}}(f_{\mathrm{test}}) + f_{\mathrm{test}} \mathbf{W}_c^T$ . + +# 3.2. Training-free Dynamic Adapter + +During testing, pre-trained vision-language models like CLIP may encounter distribution shifts that degrade the classification performance. To address this issue, existing test-time prompt tuning methods train a learnable prompt by enforcing consistency across different image augmentations during testing. However, it requires a large number of augmentation operations on each test image and computationally-heavy optimization steps to learn the prompt, limiting its applicability in various real-world settings. + +In this paper, our motivation is to design an efficient method for test-time adaptation of the pre-trained vision-language models like CLIP. Inspired by the concept of TipAdapter, we propose a training-free dynamic adapter (TDA) to enable efficient and effective test-time adaptation with CLIP. As shown in Figure 2, TDA includes two lightweight key-value caches, where each cache stores a dynamic queue of few-shot test features as keys and the corresponding pseudo labels as values. The first cache is intended for positive learning and it dynamically updates key-value pairs with high-confidence predictions to improve the accuracy. The second cache is designed for negative learning and it aims to address the adverse effects of noisy pseudo labels by introducing negative pseudo labeling to identify class absence rather than presence. By combining the positive cache and negative cache, the proposed TDA can achieve superior performance in terms of both speed and accuracy. + +![](images/c29d8b07ccb6929f5a98a2ce188137beb7875f7b6c69de9ed5f29a2420f6b6b4.jpg) +Figure 2. Overview of the proposed Training-free Dynamic Adapter (TDA). TDA constructs and updates two key-value caches to store the knowledge of a stream of test samples, and uses the two caches to generate positive and negative predictions which are combined with CLIP predictions to produce the final prediction. Specifically, the CLIP predictions are generated by performing the dot product between the image features generated by CLIP's image encoder $E_{v}$ and the text embeddings generated by CLIP's text encoder $E_{t}$ , using the hand-crafted prompt and class names. The two key-value caches are updated by gradually incorporating the test features and their corresponding pseudo labels calculated from CLIP's predictions, based on prediction entropy and cache capacity. + +Our goal is to conduct test-time adaptation by gathering adequate knowledge from the testing data stream and improving the predictions through the adaptation of the image features. To accomplish this goal, we create positive and negative caches that capture particular characteristics of the testing data stream. In the upcoming parts, we will present the process of collecting data for each cache and define the conditions in which we can utilize cache information to adapt features and enhance model predictions. + +Positive Cache. The positive cache in TDA is a key-value cache, in which keys and values are represented as a dynamic queue. It aims to collect high-quality few-shot pseudo labels $\hat{\mathbf{L}}_{\mathfrak{p}}$ as positive values and the corresponding features $\mathbf{Q}_{\mathfrak{p}}$ as keys. The key-value cache is initially empty and then accumulates a sufficient number of key-value pairs during the test-time adaptation. To maintain high-quality pseudo labels, TDA progressively incorporates test predictions with lower entropy while limiting shot capacity1 in the positive cache. Different from a normal queue like a FIFO queue with a fixed size, the dynamic queue in our method expands in size during testing. Besides, the dynamic queue operates similarly to a priority queue, using entropy as the criterion for prioritization. Note that, each class has its own queue to maintain the order and correct data structure of each class in the cache. + +Given a pre-trained CLIP model that consists of a text encoder $E_{t}$ and an image encoder $E_{v}$ , $E_{t}$ processes class + +names with pre-defined prompts to generate $c$ -channeled text embeddings $\mathbf{W}_c$ and $E_v$ processes each test image $x_{\mathrm{test}}$ to produce image features $f_{\mathrm{test}}$ . To build the positive cache, TDA first generates a pseudo label $\hat{l}$ , a one-hot encoded vector of a categorical distribution, for each test sample $x_{\mathrm{test}}$ by applying the softmax function to the prediction $f_{\mathrm{test}} \mathbf{W}_c^T$ . We establish two conditions to ascertain whether and how to include the pseudo label $\hat{l}$ and its corresponding image features $f_{\mathrm{test}}$ into the positive cache. The first condition is defined as: if the shot number (the number of collected pairs per class) of $\hat{\mathbf{L}}_{\mathbf{p}}$ is less than the maximum shot capacity (the maximum capacity number of pairs per class) $k$ , TDA will add $\hat{l}$ and $f_{\mathrm{test}}$ as a new value and a new key to $\hat{\mathbf{L}}_{\mathbf{p}}$ and $\mathbf{Q}_{\mathbf{p}}$ respectively. Meanwhile, the second condition is defined as: if the shot number of $\hat{\mathbf{L}}_{\mathbf{p}}$ has reached the maximum shot capacity $k$ , TDA will replace an 'uncertain' key-value pair $\{\mathbf{q}^{ent}, \hat{l}_p^{ent}\}$ with $\{f_{\mathrm{test}}, \hat{l}\}$ when $\mathrm{H}(f_{\mathrm{test}} \mathbf{W}_c^T) < \mathrm{H}(\mathbf{q}^{ent} \mathbf{W}_c^T)$ . Here, $\mathrm{H}$ denotes the entropy function and the term 'uncertain' indicates that the entropy of a particular key-value pair is the highest compared to the entropy of all other key-value pairs of the same class in the cache model. By applying these two conditions, TDA can gradually integrate test predictions with lower entropy while controlling the shot capacity, which helps to ensure the collection of high-quality pseudo labels in positive cache. + +During test-time adaptation, the positive cache can quickly retrieve relevant information by treating the image features $f_{\mathrm{test}}$ generated from a test example as a query and searching the stored key-value pairs $\{\mathbf{Q}_{\mathbf{p}},\hat{\mathbf{L}}_{\mathbf{p}}\}$ for match- + +ing information. The adapted predictions using the positive cache can then be obtained as: + +$$ +P _ {\mathrm {p o s}} \left(f _ {\mathrm {t e s t}}\right) = A \left(f _ {\mathrm {t e s t}} \mathbf {Q} _ {\mathbf {p}} ^ {\mathbf {T}}\right) \hat {\mathbf {L}} _ {\mathbf {p}}, \tag {3} +$$ + +where $A$ is the adaptation function defined in Tip-Adapter. + +Negative Cache. Similar to the positive cache in our TDA, the negative cache is also a dynamic queue structure with negative keys and negative values denoted as $\mathbf{Q}_{\mathbf{n}}$ and $\hat{\mathbf{L}}_{\mathbf{n}}$ , respectively. It aims to gather CLIP-generated image features to $\mathbf{Q}_{\mathbf{n}}$ and the corresponding negative pseudo labels to $\hat{\mathbf{L}}_{\mathbf{n}}$ . Unlike the pseudo labels in the positive cache, the negative pseudo labels are obtained by applying negative mask on the class probabilities as: + +$$ +\hat {\mathbf {L}} _ {\mathbf {n}} = - \mathbb {1} [ p _ {l} < P (\mathbf {Q} _ {\mathbf {n}}) ], \tag {4} +$$ + +where higher probabilities than $p_l$ are selected as negative pseudo labels from uncertain predictions and the uncertainty is measured by the entropy of predictions. Here, $p_l$ represents a threshold in negative pseudo labeling, and $\hat{\mathbf{L}}_{\mathbf{n}}$ denotes a negative pseudo label, which is a vector whose elements larger than $p_l$ have a value -1 and otherwise 0. Different from existing negative learning methods [22, 23, 33] that select negative labels from all noisy labels, TDA selects negative pseudo labels from uncertain predictions to avoid bias to the data with certain predictions. + +When constructing the negative cache, the testing feature $f_{\mathrm{test}}$ will be included in negative cache if it satisfies the condition $\gamma (f_{\mathrm{test}})$ : the entropy of the prediction is in the specified interval between $\tau_{l}$ and $\tau_{h}$ : + +$$ +\gamma \left(f _ {\text {t e s t}}\right): \tau_ {l} < \mathrm {H} \left(f _ {\text {t e s t}} \mathbf {W} _ {c} ^ {T}\right) < \tau_ {h}. \tag {5} +$$ + +This condition is designed to mitigate the risk of prediction errors due to high entropy or be biased to certain predictions (characterized by very low entropy), by incorporating test samples that exhibit a moderate degree of prediction uncertainty. Once the $\gamma(f_{\mathrm{test}})$ check is completed, the remaining steps for collecting uncertain samples in the negative cache follow the same two conditions designed in the positive cache. Similar to the positive cache, TDA also limits the shot capacity $\tilde{k}$ in the negative cache. + +During test-time adaption, the testing features $f_{\mathrm{test}}$ can be quickly adapted to target domains by retrieving the knowledge from $\{\mathbf{Q}_{\mathbf{n}}, \hat{\mathbf{L}}_{\mathbf{n}}\}$ in the negative cache and the adapted prediction can be obtained as: + +$$ +P _ {\text {n e g}} \left(f _ {\text {t e s t}}\right) = - A \left(f _ {\text {t e s t}} \mathbf {Q} _ {\mathbf {n}} ^ {\mathbf {T}}\right) \hat {\mathbf {L}} _ {\mathbf {n}}, \tag {6} +$$ + +where $A$ is the adaptation function defined in Tip-Adapter. The predictions of TDA can be formulated by combining the negative cache, the positive cache and the pre-trained CLIP model together as follows: + +$$ +P _ {\mathrm {T D A}} \left(f _ {\text {t e s t}}\right) = f _ {\text {t e s t}} \mathbf {W} _ {c} ^ {T} + P _ {\text {p o s}} \left(f _ {\text {t e s t}}\right) + P _ {\text {n e g}} \left(f _ {\text {t e s t}}\right). \tag {7} +$$ + +# 3.3. Relationship with TPT and Tip-Adapter + +Both TPT and TDA are designed to handle the challenge of adapting models to test data that have distributional discrepancies from the training data. TPT trains a learnable prompt $\mathbf{p}_c$ in Eq. (1) for each test sample with a large number of augmentations and such training process requires backpropagation and is computationally intensive. In contrast, our TDA is training-free as both positive cache $\{\mathbf{Q_p},\hat{\mathbf{L}}_{\mathbf{p}}\}$ and negative cache $\{\mathbf{Q_n},\hat{\mathbf{L}}_{\mathbf{n}}\}$ are non-parametric, which is super-efficient without incurring any backpropagation. + +Our TDA employs a cache model that shares similarities with the Tip-Adapter, where features and labels are stored as key-value pairs in a memory cache. One notable distinction between the Tip-Adapter and TDA lies in the type of cache used. Specifically, the Tip-Adapter relies on a static cache $\{\mathbf{F}_{\mathrm{train}},\mathbf{L}_{\mathrm{train}}\}$ because it is designed for a supervised adaptation setting, where the ground-truth labels $\mathbf{L}_{\mathrm{train}}$ are predetermined and readily available. In contrast, our TDA introduces a new dynamic cache $\{\mathbf{Q_p},\hat{\mathbf{L}}_p\}$ designed for a test-time adaptation setting, where the pseudo labels $\hat{\mathbf{L}}_{\mathbf{p}}$ are generated on-the-fly from a stream of test samples. Furthermore, TDA incorporates a novel negative cache model $\{\mathbf{Q_n},\hat{\mathbf{L}}_n\}$ that enhances testing predictions by utilizing indirect knowledge that a test image does not belong to certain negative classes. By combining positive and negative caches, our proposed TDA is more robust to noisy pseudo labels and can generalize well to testing data. + +# 4. Experiments + +# 4.1. Experimental Setup + +Benchmarks. We conducted main experiments on two benchmarks: out-of-distribution (OOD) benchmark and cross-domain benchmark, both applied in the previous work [35] that adapts vision-language models in test time. The OOD benchmark serves as a measure of the robustness of our approach by involving assessment on 4 out-of-distribution datasets derived from ImageNet [5]: ImageNet-A [16], ImageNet-V2 [32], ImageNet-R [15], and ImageNet-S [41]. This benchmark is specifically designed to evaluate a model's capacity to generalize to new and unseen data. The cross-domain benchmark, on the other hand, is involved to evaluate the model's performance across 10 diverse image classification datasets, each from a distinct domain with different classes: Aircraft [26], Caltech101 [8], Cars [24], DTD [4], EuroSAT [14], Flower102 [28], Food101 [1], Pets [30], SUN397 [42], and UCF101 [36]. This benchmark provides a comprehensive evaluation of the model's adaptability during test time across various class spaces. + +
MethodImageNetImageNet-AImageNet-V2ImageNet-RImageNet-SAverageOOD Average
CLIP-ResNet-5059.8123.2452.9160.7235.4846.4343.09
CoOp63.3323.0655.4056.6034.6746.6142.43
CoCoOp62.8123.3255.7257.7434.4846.8142.82
Tip-Adapter62.0323.1353.9760.3535.7447.0443.30
TPT60.7426.6754.7059.1135.0947.2643.89
DiffTPT60.8031.0655.8058.8037.1048.7145.69
TDA (Ours)61.3530.2955.5462.5838.1249.5846.63
CLIP-ViT-B/1668.3449.8961.8877.6548.2461.2059.42
CoOp71.5149.7164.2075.2147.9961.7259.28
CoCoOp71.0250.6364.0776.1848.7562.1359.91
Tip-Adapter70.7551.0463.4177.7648.8862.3760.27
TPT68.9854.7763.4577.0647.9462.4460.81
DiffTPT70.3055.6865.1075.0046.8062.2860.52
TDA (Ours)69.5160.1164.6780.2450.5465.0163.89
+ +Table 1. Results on the OOD Benchmark. Our TDA is compared with several state-of-the-art methods designed for vision-language models: the baseline method CLIP, three train-time adaptation methods (i.e., CoOp, CoCoOp, and Tip-Adapter), and two test-time adaptation methods (i.e., TPT and DiffTPT). All the compared methods are built upon CLIP-ResNet-50 or CLIP-ViT-B/16 baselines. The two evaluation metrics Average and OOD Average are calculated by taking the mean accuracy across all five datasets and four OOD datasets excluding ImageNet. The results of CLIP, CoOp, CoCoOp, and TPT are obtained from the TPT paper, the results of DiffTPT are obtained from the DiffTPT paper, while the results of Tip-Adapter are reproduced using the official codes. + +
MethodTesting TimeAccuracyGain
CLIP-ResNet-5012min59.810
TPT12h 50min60.74+0.93
DiffTPT34h 45min60.80+0.99
TDA (Ours)16min61.35+1.54
+ +Table 2. Comparisons of our TDA with CLIP-ResNet-50, TPT, and DiffTPT in terms of efficiency (Testing Time) and effectiveness (Accuracy). The final column shows the accuracy gain relative to the baseline CLIP. Note that the testing time of DiffTPT does not include the duration required for the image generation process with pre-trained diffusion models, which is an additional time-consuming factor during the testing phase. + +Implementation details. All the models in our experiments are built upon the pre-trained CLIP model [31] that consists of an image encoder and a text encoder. The image encoder can be either ResNet [13] or Vision Transformer [7], while the text encoder is Transformer [39]. Test-time adaptation is set for single-image scenarios, using a batch size of 1. We conduct a search for all our hyperparameters using a single ImageNet validation set. The threshold $p_l$ for negative pseudo-labeling in Eq 4 is set as 0.03. The upper and lower thresholds $[\tau_l, \tau_h]$ for testing feature selection in Eq 5 are set as [0.2, 0.5]. Once searched, these hyperparameters are fixed and evaluated across various new datasets. To avoid incurring backpropagation when using learnable prompt, we follow [31] to use hand-crafted + +prompts. We use top-1 accuracy $(\%)$ , a standard classification criterion, as our evaluation metric. All the experiments are conducted using a single NVIDIA Quadro RTX 6000 GPU. + +# 4.2. Comparisons with State-of-the-art + +In this section, we compare our proposed TDA with several state-of-the-art methods, including CLIP [31], three train-time adaptation methods, i.e., CoOp [50], CoCoOp [49], and Tip-Adapter [47], as well as two existing test-time adaptation methods TPT [35] and DiffTPT [9], all of which are designed for vision-language models. Specifically, CLIP is evaluated using an ensemble of 80 handcrafted prompts as in [31]. All train-time adaptation methods are trained on the ImageNet train set with 16 shots per class and tested on other datasets as in [35]. CoOp utilizes a learnable context module of size 4 that is fine-tuned for specific downstream tasks, while CoCoOp is the generalized version of CoOp with the input-conditional token for image features. Tip-Adapter employs the optimal hyperparameters obtained from the ImageNet validation set during evaluation. We would like to note that Tip-Adapter is unable to handle new classes during testing, limiting its implementation to OOD benchmark evaluation where the training classes encompass all the testing classes. Different from train-time adaptation methods, test-time adaptation methods (i.e., TPT, DiffTPT, and our TDA) do not utilize the ImageNet train set. Instead, they are fine-tuned with target datasets using a stream of unlabeled test samples. Follow + +
MethodAircraftCaltech101CarsDTDEuroSATFlower102Food101PetsSUN397UCF101Average
CLIP-ResNet-5016.1187.2655.8940.3725.7962.7774.8282.9760.8559.4856.63
CoOp15.1286.5355.3237.2926.2061.5575.5987.0058.1559.0556.18
CoCoOp14.6187.3856.2238.5328.7365.5776.2088.3959.6157.1057.23
TPT17.5887.0258.4640.8428.3362.6974.8884.4961.4660.8257.66
DiffTPT17.6086.8960.7140.7241.0463.5379.2183.4062.7262.6759.85
TDA (Ours)17.6189.7057.7843.7442.1168.7477.7586.1862.5364.1861.03
CLIP-ViT-B/1623.2293.5566.1145.0450.4266.9982.8686.9265.6365.1664.59
CoOp18.4793.7064.5141.9246.3968.7185.3089.1464.1566.5563.88
CoCoOp22.2993.7964.9045.4539.2370.8583.9790.4666.8968.4464.63
TPT24.7894.1666.8747.7542.4468.9884.6787.7965.5068.0465.10
DiffTPT25.6092.4967.0147.0043.1370.1087.2388.2265.7462.6765.47
TDA (Ours)23.9194.2467.2847.4058.0071.4286.1488.6367.6270.6667.53
+ +Table 3. Results on the Cross-Domain Benchmark. Our TDA is compared with several state-of-the-art methods designed for vision-language models: the baseline method CLIP, two train-time adaptation methods (i.e., CoOp and CoCoOp), and two test-time adaptation methods (i.e., TPT and DiffTPT). Note that Tip-Adapter is unable to be evaluated on the Cross-Domain Benchmark as it cannot handle new classes during testing. The evaluation metric Average is calculated by taking the mean accuracy across all ten datasets. The results of CLIP, CoOp, CoCoOp, and TPT are obtained from the TPT paper, while the results of DiffTPT are obtained from the DiffTPT paper. + +ing TPT and DiffTPT, we compare TDA with the state-of-the-art over two public benchmarks: OOD benchmark and cross-domain benchmark. + +Results on the OOD Benchmark. We first compare TDA with state-of-the-art methods over the OOD benchmark. Table 1 presents the experimental results, highlighting the superior performance of the proposed TDA compared to both TPT and DiffTPT across various OOD datasets derived from ImageNet. Specifically, TDA outperforms TPT on both ResNet-50 and ViT-B/16 architectures, improving OOD accuracy by $2.74\%$ and $3.08\%$ on average, respectively. Furthermore, compared to DiffTPT, TDA exhibits an average accuracy improvement of $0.94\%$ and $3.37\%$ in the OOD benchmark for ResNet-50 and ViT-B/16, respectively. These results validate the effectiveness of TDA in enhancing test-time adaptation performance on various OOD test datasets. + +In order to provide a more comprehensive evaluation of our proposed method's efficiency and effectiveness, we compared it with the baseline CLIP-ResNet-50 and two existing test-time adaptation methods (i.e., TPT and DiffTPT). This comparison encompasses both testing time and testing accuracy, and the corresponding results are shown in Table 2. This evaluation is performed on the ImageNet validation dataset, which consists of 50,000 images, using a single NVIDIA Quadro RTX 6000 GPU. When compared to the baseline CLIP-ResNet-50, the proposed TDA demonstrates a significant improvement in testing accuracy $(+1.54\%)$ , with only a minimal sacrifice in testing efficiency (requiring an additional 4min). In comparison to TPT and DiffTPT, the proposed TDA demonstrates not only superior + +testing accuracy but also significantly improved efficiency. It reduces the testing time dramatically from $12\mathrm{h}50\mathrm{min}$ by TPT and even more from $34\mathrm{h}45\mathrm{min}$ by DiffTPT, down to just 16 minutes. Without including the image generation time, DiffTPT consumes clearly more test time than TPT as it involves a time-consuming multi-step prompt updating process whereas TPT requires a single step only. The experimental results strongly validate the effectiveness and efficiency of our proposed method, establishing its suitability for real-world applications. + +We then compare TDA with state-of-the-art methods over cross-domain benchmark. The results, presented in Table 3, demonstrate that TDA not only surpasses the performance of the TPT method but also shows a significant advantage over its improvement method DiffTPT. Specifically, when utilizing CLIP-ResNet-50 and CLIP-ViT-B/16 as the backbone, TDA achieves an improvement in average accuracy over TPT by $3.37\%$ and $2.43\%$ , respectively. These improvements, along with a $1.18\%$ and $2.06\%$ gain over DiffTPT for the respective backbones, further verify the effectiveness of TDA in adapting to diverse class datasets during test time. This attribute holds significant value for vision-language models such as CLIP, as it enables them to classify arbitrary classes in image classification without the need for additional training. + +# 4.3. Ablation Studies + +In this section, we perform ablation studies to examine the effectiveness of our designs. All the ablation studies are conducted over the ImageNet dataset, where TDA can achieve an accuracy of $61.35\%$ under default settings. TDA consists of a Positive Cache and a Negative Cache, which + +![](images/83ab17c90bae8b8ba66f047b4294bcf5a80889fa31294d1162de42bad164366d.jpg) +Figure 3. Ablation studies on two cache designs in TDA: Positive Cache and Negative Cache. All the models are built upon the baseline model CLIP-ResNet-50. + +perform positive and negative pseudo labeling within our designed dynamic adapter, respectively. We first assess the efficacy of the two cache designs in TDA. As shown in Figure 3, both Positive Cache and Negative Cache significantly surpass the baseline model CLIP, demonstrating that test-time adaptation can be improved by introducing a dynamic adapter with either positive pseudo labeling or negative pseudo labeling. Besides, the two cache designs in TDA can complement each other as TDA (i.e., the combination of the two designs) clearly outperforms either Positive Cache or Negative Cache on the challenging ImageNet dataset. Moreover, we extend our ablation studies to the cross-domain benchmark, and results show that the positive cache achieved $60.38\%$ accuracy, the negative cache $60.11\%$ , while their combination yielded a $61.03\%$ accuracy, thereby highlighting the significance of each type of cache in enhancing TDA's performance. + +We proceed to perform parameter studies on the shot capacity, which refers to the maximum number of key-value pairs per class, in both Positive Cache and Negative Cache models. These studies aims to find the optimal balance between the diversity and accuracy of the key-value pairs. Figure 4 shows that the performance of both cache models is significantly affected when the shot capacity is either too low or too high. We find that the shot capacity is set as 3 for the Positive Cache and 2 for the Negative Cache yields the best performance. This is because an appropriate shot capacity ensures both high-quality pseudo labels (paired values) and diverse image features (paired keys) in the cache models. The negative cache is sensitive to shot numbers due to its role in storing probabilities for various negative pseudo labels, each representing a class to be excluded from the model. Contrary to the intuition that a larger negative + +![](images/0003c7a8fe921051a3ee6ea5ce1be61aef18c4e9274ecd6118067ace10d5425a.jpg) +Figure 4. Parameter studies on the Shot Capacity in Positive Cache and Negative Cache. + +cache might be beneficial, a larger negative cache leads to more high-entropy, noisier pseudo labels, as highlighted in self-training studies like [11], thereby lowering the performance. Conversely, the positive cache, which stores single and high-confidence prediction, is less affected by variations in shot capacity, thereby maintaining consistent accuracy across different shot capacities. To facilitate practical applications, the shot capacity settings are fixed and directly applied to new datasets without the need for additional parameter adjustments. + +# 5. Conclusion + +In this work, we have presented TDA, a dynamic adapter for efficient and effective test-time adaptation of vision-language models. The proposed method employs a key-value cache, which maintains a dynamic queue with test-sample features as keys and corresponding few-shot pseudo labels as values, allowing for gradual adaptation to test data through progressive pseudo label improvement. Moreover, TDA introduces a negative cache to mitigate the undesirable effects of noisy pseudo labels by assigning negative pseudo labels to certain classes when the model is uncertain about its predictions. The results of extensive experiments over two benchmarks demonstrate that TDA outperforms state-of-the-art test-time adaptation methods while significantly reducing testing time. This work contributes to the research and application values of test-time adaptation and presents a promising solution to the efficiency issue of test-time adaptation of vision-language models. + +# Acknowledgement + +This study is supported under the Mohamed bin Zayed University of Artificial Intelligence. + +# References + +[1] Lukas Bossard, Matthieu Guillaumin, and Luc Van Gool. Food-101 – mining discriminative components with random forests. In European Conference on Computer Vision, 2014. 5 +[2] Malik Boudiaf, Romain Mueller, Ismail Ben Ayed, and Luca Bertinetto. Parameter-free online test-time adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8344-8353, 2022. 1 +[3] Dian Chen, Dequan Wang, Trevor Darrell, and Sayna Ebrahimi. Contrastive test-time adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 295-305, 2022. 1 +[4] M. Cimpoi, S. Maji, I. Kokkinos, S. Mohamed, , and A. Vedaldi. Describing textures in the wild. In Proceedings of the IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2014. 5 +[5] J. Deng, W. Dong, R. Socher, L.-J. Li, K. Li, and L. Fei-Fei. ImageNet: A Large-Scale Hierarchical Image Database. In CVPR09, 2009. 5 +[6] Karan Desai and Justin Johnson. VirTex: Learning Visual Representations from Textual Annotations. In CVPR, 2021. 2 +[7] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 6 +[8] Li Fei-Fei, Rob Fergus, and Pietro Perona. Learning generative visual models from few training examples: An incremental bayesian approach tested on 101 object categories. In CVPR Workshops, 2004. 5 +[9] Chun-Mei Feng, Kai Yu, Yong Liu, Salman Khan, and Wangmeng Zuo. Diverse data augmentation with diffusions for effective test-time prompt tuning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2704-2714, 2023. 1, 2, 6 +[10] Peng Gao, Shijie Geng, Renrui Zhang, Teli Ma, Rongyao Fang, Yongfeng Zhang, Hongsheng Li, and Yu Qiao. Clip-adapter: Better vision-language models with feature adapters. arXiv preprint arXiv:2110.04544, 2021. 2 +[11] Yves Grandvalet and Yoshua Bengio. Semi-supervised learning by entropy minimization. Advances in neural information processing systems, 17, 2004. 2, 8 +[12] Edouard Grave, Moustapha M Cisse, and Armand Joulin. Unbounded cache model for online language modeling with open vocabulary. Advances in neural information processing systems, 30, 2017. 2, 3 +[13] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, 2016. 6 +[14] Patrick Helber, Benjamin Bischke, Andreas Dengel, and Damian Borth. Eurosat: A novel dataset and deep learning benchmark for land use and land cover classification. IEEE J. Sel. Top. Appl. Earth Obs. Remote. Sens., 2019. 5 + +[15] Dan Hendrycks, Steven Basart, Norman Mu, Saurav Kadavath, Frank Wang, Evan Dorundo, Rahul Desai, Tyler Zhu, Samyak Parajuli, Mike Guo, Dawn Song, Jacob Steinhardt, and Justin Gilmer. The many faces of robustness: A critical analysis of out-of-distribution generalization. In ICCV, 2021. 5 +[16] Dan Hendrycks, Kevin Zhao, Steven Basart, Jacob Steinhardt, and Dawn Song. Natural adversarial examples. In CVPR, pages 15262-15271, 2021. 5 +[17] Xiaowei Hu, Zhe Gan, Jianfeng Wang, Zhengyuan Yang, Zicheng Liu, Yumao Lu, and Lijuan Wang. Scaling up vision-language pre-training for image captioning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17980-17989, 2022. 2 +[18] Yusuke Iwasawa and Yutaka Matsuo. Test-time classifier adjustment module for model-agnostic domain generalization. Advances in Neural Information Processing Systems, 34:2427-2440, 2021. 2 +[19] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In International Conference on Machine Learning, pages 4904-4916. PMLR, 2021. 1, 2 +[20] Jeff Johnson, Matthijs Douze, and Hervé Jégou. Billionscale similarity search with gpus. IEEE Transactions on Big Data, 7(3):535-547, 2019. 3 +[21] Urvashi Khandelwal, Omer Levy, Dan Jurafsky, Luke Zettlemoyer, and Mike Lewis. Generalization through memorization: Nearest neighbor language models. In International Conference on Learning Representations, 2020. 2 +[22] Youngdong Kim, Junho Yim, Juseung Yun, and Junmo Kim. NInl: Negative learning for noisy labels. In Proceedings of the IEEE/CVF international conference on computer vision, pages 101-110, 2019. 2, 5 +[23] Youngdong Kim, Juseung Yun, Hyounguk Shon, and Junmo Kim. Joint negative and positive learning for noisy labels. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9442-9451, 2021. 2, 5 +[24] Jonathan Krause, Michael Stark, Jia Deng, and Li Fei-Fei. 3d object representations for fine-grained categorization. In ICCV Workshops, 2013. 5 +[25] Xin Li, Dongze Lian, Zhihe Lu, Jiawang Bai, Zhibo Chen, and Xinchao Wang. Graphadapter: Tuning vision-language models with dual knowledge graph. Advances in Neural Information Processing Systems, 36, 2024. 2 +[26] S. Maji, J. Kannala, E. Rahtu, M. Blaschko, and A. Vedaldi. Fine-grained visual classification of aircraft. Technical report, 2013. 5 +[27] Stephen Merity, Caiming Xiong, James Bradbury, and Richard Socher. Pointer sentinel mixture models. arXiv preprint arXiv:1609.07843, 2016. 3 +[28] Maria-Elena Nilsback and Andrew Zisserman. Automated flower classification over a large number of classes. In Indian Conference on Computer Vision, Graphics and Image Processing, 2008. 5 + +[29] Emin Orhan. A simple cache model for image recognition. Advances in Neural Information Processing Systems, 31, 2018. 2 +[30] Omkar M. Parkhi, Andrea Vedaldi, Andrew Zisserman, and C. V. Jawahar. Cats and dogs. In CVPR, 2012. 5 +[31] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In ICML, 2021. 1, 2, 3, 6 +[32] Benjamin Recht, Rebecca Roelofs, Ludwig Schmidt, and Vaishaal Shankar. Do imagenet classifiers generalize to imageNet? In ICML, 2019. 5 +[33] Mamshad Nayeem Rizve, Kevin Duarte, Yogesh S Rawat, and Mubarak Shah. In defense of pseudo-labeling: An uncertainty-aware pseudo-label selection framework for semi-supervised learning. arXiv preprint arXiv:2101.06329, 2021. 2, 5 +[34] Steffen Schneider, Evgenia Rusak, Luisa Eck, Oliver Bringmann, Wieland Brendel, and Matthias Bethge. Improving robustness against common corruptions by covariate shift adaptation. Advances in Neural Information Processing Systems, 33:11539-11551, 2020. 2 +[35] Manli Shu, Weili Nie, De-An Huang, Zhiding Yu, Tom Goldstein, Anima Anandkumar, and Chaowei Xiao. Test-time prompt tuning for zero-shot generalization in vision-language models. In Advances in Neural Information Processing Systems, 2022. 1, 2, 3, 5, 6 +[36] Khurram Soomro, Amir Roshan Zamir, and Mubarak Shah. UCF101: A dataset of 101 human actions classes from videos in the wild. CoRR, abs/1212.0402, 2012. 5 +[37] Yu Sun, Xiaolong Wang, Zhuang Liu, John Miller, Alexei A. Efros, and Moritz Hardt. Test-time training with self-supervision for generalization under distribution shifts. In ICML, 2020. 2 +[38] Thomas Varsavsky, Maurizio Orbes-Arteaga, Carole H Sudre, Mark S Graham, Parashkev Nachev, and M Jorge Cardoso. Test-time unsupervised domain adaptation. In Medical Image Computing and Computer Assisted Intervention—MICCAI 2020: 23rd International Conference, Lima, Peru, October 4–8, 2020, Proceedings, Part I 23, pages 428–436. Springer, 2020. 2 +[39] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 6 +[40] Dequan Wang, Evan Shelhamer, Shaoteng Liu, Bruno A. Olshausen, and Trevor Darrell. Tent: Fully test-time adaptation by entropy minimization. In ICLR, 2021. 1, 2 +[41] Haohan Wang, Songwei Ge, Zachary Lipton, and Eric P Xing. Learning robust global representations by penalizing local predictive power. In NeurIPS, 2019. 5 +[42] J. Xiao, J. Hays, K. A. Ehinger, A. Oliva, and A. Torralba. Sun database: Large-scale scene recognition from abbey to zoo. In 2010 IEEE Computer Society Conference on Computer Vision and Pattern Recognition, 2010. 5 + +[43] Jinyu Yang, Jiali Duan, Son Tran, Yi Xu, Sampath Chanda, Liquun Chen, Belinda Zeng, Trishul Chilimbi, and Junzhou Huang. Vision-language pre-training with triple contrastive learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15671-15680, 2022. 1, 2 +[44] Tao Yu, Zhihe Lu, Xin Jin, Zhibo Chen, and Xinchao Wang. Task residual for tuning vision-language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10899-10909, 2023. 2 +[45] Marvin Zhang, Henrik Marklund, Nikita Dhawan, Abhishek Gupta, Sergey Levine, and Chelsea Finn. Adaptive risk minimization: Learning to adapt to domain shift. Advances in Neural Information Processing Systems, 34:23664-23678, 2021. 2 +[46] Marvin Mengxin Zhang, Sergey Levine, and Chelsea Finn. MEMO: Test time robustness via adaptation and augmentation. In Advances in Neural Information Processing Systems, 2022. 2 +[47] Renrui Zhang, Wei Zhang, Rongyao Fang, Peng Gao, Kunchang Li, Jifeng Dai, Yu Qiao, and Hongsheng Li. Tip-adapter: Training-free adaption of clip for few-shot classification. In European Conference on Computer Vision, pages 493-510. Springer, 2022. 2, 3, 6 +[48] Renrui Zhang, Xiangfei Hu, Bohao Li, Siyuan Huang, Hanqiu Deng, Yu Qiao, Peng Gao, and Hongsheng Li. Prompt, generate, then cache: Cascade of foundation models makes strong few-shot learners. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15211-15222, 2023. 2 +[49] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Conditional prompt learning for vision-language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16816-16825, 2022. 2, 6 +[50] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Learning to prompt for vision-language models. International Journal of Computer Vision, 130(9):2337-2348, 2022. 2, 6 \ No newline at end of file diff --git a/2024/Efficient Test-Time Adaptation of Vision-Language Models/images.zip b/2024/Efficient Test-Time Adaptation of Vision-Language Models/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..05be08864b88858c9615e33638ca4770d7b8eacd --- /dev/null +++ b/2024/Efficient Test-Time Adaptation of Vision-Language Models/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9ac5434ebc4f47d8d1b0ad1ead19d95f159d0e5eb3f561501ec9d9f6735ebd40 +size 448092 diff --git a/2024/Efficient Test-Time Adaptation of Vision-Language Models/layout.json b/2024/Efficient Test-Time Adaptation of Vision-Language Models/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..154ab079f2ddd9b5a4f10e87cebcdd8a372c7bbe --- /dev/null +++ b/2024/Efficient Test-Time Adaptation of Vision-Language Models/layout.json @@ -0,0 +1,7769 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 115, + 103, + 479, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 115, + 103, + 479, + 121 + ], + "spans": [ + { + "bbox": [ + 115, + 103, + 479, + 121 + ], + "type": "text", + "content": "Efficient Test-Time Adaptation of Vision-Language Models" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 59, + 141, + 531, + 159 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 141, + 531, + 159 + ], + "spans": [ + { + "bbox": [ + 59, + 141, + 531, + 159 + ], + "type": "text", + "content": "Adilbek Karmanov" + }, + { + "bbox": [ + 59, + 141, + 531, + 159 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 59, + 141, + 531, + 159 + ], + "type": "text", + "content": " Dayan Guan" + }, + { + "bbox": [ + 59, + 141, + 531, + 159 + ], + "type": "inline_equation", + "content": "^{2*}" + }, + { + "bbox": [ + 59, + 141, + 531, + 159 + ], + "type": "text", + "content": " Shijian Lu" + }, + { + "bbox": [ + 59, + 141, + 531, + 159 + ], + "type": "inline_equation", + "content": "^{1,2\\dagger}" + }, + { + "bbox": [ + 59, + 141, + 531, + 159 + ], + "type": "text", + "content": " Abdulmotaleb El Saddik" + }, + { + "bbox": [ + 59, + 141, + 531, + 159 + ], + "type": "inline_equation", + "content": "^{1,3}" + }, + { + "bbox": [ + 59, + 141, + 531, + 159 + ], + "type": "text", + "content": " Eric Xing" + }, + { + "bbox": [ + 59, + 141, + 531, + 159 + ], + "type": "inline_equation", + "content": "^{1,4}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 64, + 159, + 529, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 159, + 529, + 190 + ], + "spans": [ + { + "bbox": [ + 64, + 159, + 529, + 190 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 64, + 159, + 529, + 190 + ], + "type": "text", + "content": "Mohamed bin Zayed University of Artificial Intelligence " + }, + { + "bbox": [ + 64, + 159, + 529, + 190 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 64, + 159, + 529, + 190 + ], + "type": "text", + "content": "Nanyang Technological University " + }, + { + "bbox": [ + 64, + 159, + 529, + 190 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 64, + 159, + 529, + 190 + ], + "type": "text", + "content": "University of Ottawa " + }, + { + "bbox": [ + 64, + 159, + 529, + 190 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 64, + 159, + 529, + 190 + ], + "type": "text", + "content": "Carnegie Mellon University" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 143, + 215, + 192, + 228 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 215, + 192, + 228 + ], + "spans": [ + { + "bbox": [ + 143, + 215, + 192, + 228 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 241, + 290, + 504 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 241, + 290, + 504 + ], + "spans": [ + { + "bbox": [ + 46, + 241, + 290, + 504 + ], + "type": "text", + "content": "Test-time adaptation with pre-trained vision-language models has attracted increasing attention for tackling distribution shifts during the test time. Though prior studies have achieved very promising performance, they involve intensive computation which is severely unaligned with test-time adaptation. We design TDA, a training-free dynamic adapter that enables effective and efficient test-time adaptation with vision-language models. TDA works with a lightweight key-value cache that maintains a dynamic queue with few-shot pseudo labels as values and the corresponding test-sample features as keys. Leveraging the key-value cache, TDA allows adapting to test data gradually via progressive pseudo label refinement which is super-efficient without incurring any backpropagation. In addition, we introduce negative pseudo labeling that alleviates the adverse impact of pseudo label noises by assigning pseudo labels to certain negative classes when the model is uncertain about its pseudo label predictions. Extensive experiments over two benchmarks demonstrate TDA's superior effectiveness and efficiency as compared with the state-of-the-art. The code has been released in https://kdiaaa.github.io/tda/." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 529, + 128, + 542 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 529, + 128, + 542 + ], + "spans": [ + { + "bbox": [ + 47, + 529, + 128, + 542 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 551, + 287, + 684 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 551, + 287, + 684 + ], + "spans": [ + { + "bbox": [ + 46, + 551, + 287, + 684 + ], + "type": "text", + "content": "Recent advances in vision-language models [19, 31, 43] have opened a new door for integrating human language into various computer vision tasks. Take CLIP [31] as an example. It can enable zero-shot image classification by leveraging a shared embedding space that is learnt from web-scale image-text pairs. Within this shared space, images can be directly recognized by matching their features with the text embeddings of CLIP classes. At the other end, CLIP often faces challenges while handling various specific downstream images, especially when the downstream images have clear domain and distribution shifts as compared" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 307, + 214, + 545, + 340 + ], + "blocks": [ + { + "bbox": [ + 307, + 214, + 545, + 340 + ], + "lines": [ + { + "bbox": [ + 307, + 214, + 545, + 340 + ], + "spans": [ + { + "bbox": [ + 307, + 214, + 545, + 340 + ], + "type": "image", + "image_path": "1d0a13a9b4fa3f9ddd9c9eee1a89ba333532057ee01bec39e001bf87a863ba0b.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 368, + 346, + 485, + 357 + ], + "lines": [ + { + "bbox": [ + 368, + 346, + 485, + 357 + ], + "spans": [ + { + "bbox": [ + 368, + 346, + 485, + 357 + ], + "type": "text", + "content": "(a) Test-time Prompt Tuning [9, 35]" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 307, + 361, + 545, + 508 + ], + "blocks": [ + { + "bbox": [ + 307, + 361, + 545, + 508 + ], + "lines": [ + { + "bbox": [ + 307, + 361, + 545, + 508 + ], + "spans": [ + { + "bbox": [ + 307, + 361, + 545, + 508 + ], + "type": "image", + "image_path": "b84cafaa5da2b0e13490f6052a78e44341a30fea0f1aa56e779523647d7ada74.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 357, + 513, + 495, + 523 + ], + "lines": [ + { + "bbox": [ + 357, + 513, + 495, + 523 + ], + "spans": [ + { + "bbox": [ + 357, + 513, + 495, + 523 + ], + "type": "text", + "content": "(b) Training-free Dynamic Adapter (Ours)" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 304, + 533, + 547, + 609 + ], + "lines": [ + { + "bbox": [ + 304, + 533, + 547, + 609 + ], + "spans": [ + { + "bbox": [ + 304, + 533, + 547, + 609 + ], + "type": "text", + "content": "Figure 1. Comparison of our proposed Training-free Dynamic Adapter (TDA) with Test-time Prompt Tuning TPT [35] and its enhancement DiffTPT [9]: both TPT and DiffTPT require significant computational resources to optimize the learnable prompt via backpropagation; TDA is a dynamic cache that is training-free without any backpropagation, making it efficient for test-time adaptation in various real-world scenarios." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 638, + 432, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 638, + 432, + 651 + ], + "spans": [ + { + "bbox": [ + 306, + 638, + 432, + 651 + ], + "type": "text", + "content": "with the CLIP training images." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 653, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 653, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 653, + 547, + 715 + ], + "type": "text", + "content": "Several recent studies [2, 3, 40] introduce a new paradigm called test-time adaptation for mitigating the domain shift. The idea of test-time adaptation is well aligned with real-world scenarios where a model needs to adapt to new environments quickly. Despite its great research and" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 693, + 128, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 693, + 128, + 703 + ], + "spans": [ + { + "bbox": [ + 58, + 693, + 128, + 703 + ], + "type": "text", + "content": "*Equal Contribution." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 59, + 703, + 138, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 703, + 138, + 712 + ], + "spans": [ + { + "bbox": [ + 59, + 703, + 138, + 712 + ], + "type": "text", + "content": "† Corresponding Author." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "14162" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 228 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 228 + ], + "type": "text", + "content": "application values, test-time adaptation of vision-language models has been largely neglected. Recently, Test-time Prompt Tuning, as introduced in TPT [35] and its enhancement DiffTPT [9], attempts to adapt vision-language models by learning domain-specific prompts from test data. As illustrated in Fig. 1 (a), both TPT and DiffTPT train a learnable prompt for each test sample by feeding its augmented views into the CLIP model for generating predictions and minimizing the marginal entropy of confident predictions. Despite its decent performance, the prompt optimization in both TPT [50] and DiffTPT [9] is computationally intensive which hinders its applications in various real-world scenarios." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 229, + 289, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 229, + 289, + 422 + ], + "spans": [ + { + "bbox": [ + 46, + 229, + 289, + 422 + ], + "type": "text", + "content": "We design a training-free dynamic adapter (TDA) that allows efficient and effective test-time adaptation of vision-language models without requiring any backpropagation in test time. As Fig. 1 (b) shows, TDA constructs a lightweight Dynamic Adapter that keeps a dynamic queue with the pseudo labels of a stream of test samples as values and the corresponding CLIP-extracted features as keys. TDA has two desirable features that make its test-time adaptation highly applicable in real-world scenarios. First, TDA is highly effective, as it improves the quality of pseudo labels via progressive incorporation of test predictions of lower entropy [11, 40]. Second, TDA is very efficient as the key-value cache is non-parametric and requires no backpropagation during testing. Beyond that, TDA cache is lightweight due to the few-shot setup and it can be computed with simple matrix multiplications [12, 21, 29, 47]." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 423, + 289, + 675 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 423, + 289, + 675 + ], + "spans": [ + { + "bbox": [ + 46, + 423, + 289, + 675 + ], + "type": "text", + "content": "Note that the performance of TDA depends heavily on the pseudo labels of unlabelled test samples which are often noisy with prediction errors. Inspired by the idea of negative learning [22, 23, 33], we introduce negative pseudo labeling to reduce the impact of noisy estimated labels. Traditional pseudo labeling methods identify the presence of particular classes in unlabeled data, which may result in erroneous pseudo labels being assigned when comparable high probabilities are observed. In contrast, our designed negative pseudo labeling determines the absence of certain classes and can provide more accurate pseudo labels as the probabilities of these complementary classes are very low. Concretely, we construct an additional TDA cache that stores negative pseudo labels to complement the positive TDA cache. By combining positive and negative caches, TDA is more tolerant to noisy pseudo labels and can better generalize to testing data. Extensive experiments over two widely adopted test-time adaptation benchmarks show that TDA outperforms the state-of-the-art by large margins while significantly reducing the testing time from over 12 hours to 16 minutes on the ImageNet dataset." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 677, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 677, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 47, + 677, + 288, + 715 + ], + "type": "text", + "content": "In summary, the contributions of this work are threefold. First, we design a training-free dynamic adapter (TDA) that can achieve test-time adaptation of vision-language models" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 304, + 72, + 547, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 180 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 180 + ], + "type": "text", + "content": "efficiently and effectively. To the best of our knowledge, this is the first work that investigates the efficiency issue of test-time adaptation of vision-language models. Second, we introduce negative pseudo labeling to alleviate the adverse impact of pseudo label noises which makes TDA more robust to pseudo label noises and generalizable to testing data. Third, we evaluate TDA extensively over two benchmarks, and experiments show that TDA achieves superior accuracy and efficiency compared with the state-of-the-art." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 306, + 192, + 392, + 205 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 192, + 392, + 205 + ], + "spans": [ + { + "bbox": [ + 306, + 192, + 392, + 205 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 213, + 547, + 439 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 213, + 547, + 439 + ], + "spans": [ + { + "bbox": [ + 304, + 213, + 547, + 439 + ], + "type": "text", + "content": "Vision-language models [6, 17, 19, 31, 43] have demonstrated significant potential in learning semantic representations effectively by undergoing extensive training on image-text data. CLIP [31] stands out among these models for its ability to establish links between visual and textual representations, which enables it to achieve impressive zero-shot results on various downstream tasks. To enhance the transfer learning capability of the CLIP model in the downstream classification tasks, researchers have proposed integrating language prompt learners such as CoOp [50] and CoCoOp [49], as well as vision adapters such as CLIP-Adapter [10], Tip-Adapter [47], CaFo [48], TaskRes [44] and GraphAdapter [25]. Although these methods have shown considerable performance improvements, they typically require a large amount of training data in downstream tasks, making them less practical for real-world scenarios. On other hand, this work focuses on a new paradigm named test-time adaptation without accessing the original training data." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 450, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 450, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 450, + 547, + 715 + ], + "type": "text", + "content": "Test-time adaptation refers to the process of adapting models to testing data that may have distributional differences from the training data. It is particularly beneficial for real-world applications that require models to be deployed in diverse environments, such as autonomous driving in various weather conditions, medical diagnosis in different hospitals, and etc. Several recent works utilize each batch of testing samples to update partial weights [18, 37, 38], normalization statistics [34], or a combination of both [40, 45]. To avoid updating models with multiple testing samples, MEMO [46] proposes enforcing the invariant predictions from different augmentations of each sample in the testing data stream. TPT [35] tackles the same challenge with vision-language models by fine-tuning a learnable prompt with each testing sample. DiffTPT [9] innovates test-time prompt tuning by leveraging pre-trained diffusion models to augment the diversity of test data samples used in TPT. Although TPT [35] and DiffTPT [9] are effective in addressing test-time adaptation of vision-language models, prompt learning is computationally expensive and time-consuming. This paper aims to mitigate the computational efficiency challenges of TPT and DiffTPT through the introduction of" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "14163" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 110, + 83 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 110, + 83 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 110, + 83 + ], + "type": "text", + "content": "a cache model." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 94, + 289, + 321 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 94, + 289, + 321 + ], + "spans": [ + { + "bbox": [ + 47, + 94, + 289, + 321 + ], + "type": "text", + "content": "Cache model benefits the adaptation techniques by providing efficient inference and non-parametric processing without requiring parameter updates. Unbounded Cache [12] and PSMM [27] have shown promising results in the text generation task by storing a large amount of the training dataset to capture long-term dependencies, however, this approach poses a challenge of memory efficiency during the test-time adaptation. A different technique [20] was proposed to mitigate this limitation by reducing the size of the cache memory through the search for application. Alternatively, Tip-Adapter [47] solves the cache memory problem by only storing few-shot samples per class to create a cache model for vision-language models. Inspired by this work, our dynamic adapter uses the same architecture for the test-time adaptation setting, where there is no access to the source data, and testing samples can only be accessed one by one. To address the lack of access to source data during testing, our adapter collects the most reliable test samples and their pseudo labels to form cache model." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 331, + 104, + 344 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 331, + 104, + 344 + ], + "spans": [ + { + "bbox": [ + 47, + 331, + 104, + 344 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 351, + 133, + 362 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 351, + 133, + 362 + ], + "spans": [ + { + "bbox": [ + 47, + 351, + 133, + 362 + ], + "type": "text", + "content": "3.1. Preliminaries" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 369, + 287, + 478 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 369, + 287, + 478 + ], + "spans": [ + { + "bbox": [ + 47, + 369, + 287, + 478 + ], + "type": "text", + "content": "CLIP [31] is a vision-language model that performs a proxy task of predicting the correct pairings of text and image. Consider an " + }, + { + "bbox": [ + 47, + 369, + 287, + 478 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 47, + 369, + 287, + 478 + ], + "type": "text", + "content": "-class classification problem where the CLIP's objective in the zero-shot setting is to match images with their most relevant textual descriptions using an image encoder " + }, + { + "bbox": [ + 47, + 369, + 287, + 478 + ], + "type": "inline_equation", + "content": "E_{v}" + }, + { + "bbox": [ + 47, + 369, + 287, + 478 + ], + "type": "text", + "content": " and a text encoder " + }, + { + "bbox": [ + 47, + 369, + 287, + 478 + ], + "type": "inline_equation", + "content": "E_{t}" + }, + { + "bbox": [ + 47, + 369, + 287, + 478 + ], + "type": "text", + "content": ". To obtain the textual descriptions, " + }, + { + "bbox": [ + 47, + 369, + 287, + 478 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 47, + 369, + 287, + 478 + ], + "type": "text", + "content": "-class names are concatenated with handcrafted prompts and then mapped into the " + }, + { + "bbox": [ + 47, + 369, + 287, + 478 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 47, + 369, + 287, + 478 + ], + "type": "text", + "content": "-channeled text embeddings " + }, + { + "bbox": [ + 47, + 369, + 287, + 478 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_{c}" + }, + { + "bbox": [ + 47, + 369, + 287, + 478 + ], + "type": "text", + "content": " using the text encoder " + }, + { + "bbox": [ + 47, + 369, + 287, + 478 + ], + "type": "inline_equation", + "content": "E_{t}" + }, + { + "bbox": [ + 47, + 369, + 287, + 478 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 487, + 289, + 643 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 487, + 289, + 643 + ], + "spans": [ + { + "bbox": [ + 47, + 487, + 289, + 643 + ], + "type": "text", + "content": "TPT [35] focuses on test-time adaptation of CLIP. In TPT, a prompt tuning method is proposed to learn an adaptive prompt " + }, + { + "bbox": [ + 47, + 487, + 289, + 643 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_c" + }, + { + "bbox": [ + 47, + 487, + 289, + 643 + ], + "type": "text", + "content": " using individual test samples. A set of augmentation functions " + }, + { + "bbox": [ + 47, + 487, + 289, + 643 + ], + "type": "inline_equation", + "content": "\\mathcal{A}" + }, + { + "bbox": [ + 47, + 487, + 289, + 643 + ], + "type": "text", + "content": " is used to generate " + }, + { + "bbox": [ + 47, + 487, + 289, + 643 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 47, + 487, + 289, + 643 + ], + "type": "text", + "content": " randomly augmented views " + }, + { + "bbox": [ + 47, + 487, + 289, + 643 + ], + "type": "inline_equation", + "content": "\\tilde{x}_{\\mathrm{test}} = \\mathcal{A}_n(x_{\\mathrm{test}})" + }, + { + "bbox": [ + 47, + 487, + 289, + 643 + ], + "type": "text", + "content": " of a test sample " + }, + { + "bbox": [ + 47, + 487, + 289, + 643 + ], + "type": "inline_equation", + "content": "x_{\\mathrm{test}}" + }, + { + "bbox": [ + 47, + 487, + 289, + 643 + ], + "type": "text", + "content": ". The objective of TPT is to reduce variation in the model's predictions across different augmentations " + }, + { + "bbox": [ + 47, + 487, + 289, + 643 + ], + "type": "inline_equation", + "content": "\\tilde{f}_{\\mathrm{test}} = E_v(\\tilde{x}_{\\mathrm{test}})" + }, + { + "bbox": [ + 47, + 487, + 289, + 643 + ], + "type": "text", + "content": " by minimizing the marginal entropy among the outputs of the augmented views. Furthermore, TPT also includes confidence selection to discard noisy augmentations that could result in ambiguous model predictions, as shown in Figure 1a. This is achieved by filtering out augmented views with high-entropy predictions as:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 73, + 651, + 287, + 682 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 73, + 651, + 287, + 682 + ], + "spans": [ + { + "bbox": [ + 73, + 651, + 287, + 682 + ], + "type": "interline_equation", + "content": "P _ {\\mathrm {T P T}} \\left(\\tilde {f} _ {\\text {t e s t}}\\right) = \\frac {1}{\\rho n} \\sum_ {i = 1} ^ {n} \\mathbb {1} \\left[ \\mathrm {H} \\left(\\tilde {f} _ {i} \\mathbf {p} _ {c} ^ {T}\\right) \\leq \\tau \\right] \\tilde {f} _ {i} \\mathbf {p} _ {c} ^ {T}, \\tag {1}", + "image_path": "1ad7ae0880970c890f067e23775e74584f519d257c4e184a0f1491511e490acb.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 689, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 288, + 715 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 689, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\mathrm{H}" + }, + { + "bbox": [ + 47, + 689, + 288, + 715 + ], + "type": "text", + "content": " is the self-entropy function of the softmax logits predictions, " + }, + { + "bbox": [ + 47, + 689, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\tilde{f}_i\\mathbf{p}_c^T" + }, + { + "bbox": [ + 47, + 689, + 288, + 715 + ], + "type": "text", + "content": " is the class probabilities vector of size" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 72, + 545, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 120 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 120 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 305, + 72, + 545, + 120 + ], + "type": "text", + "content": " generated from the given " + }, + { + "bbox": [ + 305, + 72, + 545, + 120 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 305, + 72, + 545, + 120 + ], + "type": "text", + "content": "-th augmented view of the test image, and the parameter " + }, + { + "bbox": [ + 305, + 72, + 545, + 120 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 305, + 72, + 545, + 120 + ], + "type": "text", + "content": " determines that only " + }, + { + "bbox": [ + 305, + 72, + 545, + 120 + ], + "type": "inline_equation", + "content": "\\rho" + }, + { + "bbox": [ + 305, + 72, + 545, + 120 + ], + "type": "text", + "content": "-percentile of confident samples with entropy values below this threshold can be selected out of " + }, + { + "bbox": [ + 305, + 72, + 545, + 120 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 305, + 72, + 545, + 120 + ], + "type": "text", + "content": " augmented views." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 130, + 546, + 273 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 130, + 546, + 273 + ], + "spans": [ + { + "bbox": [ + 305, + 130, + 546, + 273 + ], + "type": "text", + "content": "Tip-Adapter [47] provides a training-free solution that uses a key-value cache model and integrates knowledge from the pre-trained CLIP model with few-shot labeled samples. The cache model is created using a set of " + }, + { + "bbox": [ + 305, + 130, + 546, + 273 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 305, + 130, + 546, + 273 + ], + "type": "text", + "content": "-shot labeled samples " + }, + { + "bbox": [ + 305, + 130, + 546, + 273 + ], + "type": "inline_equation", + "content": "x_{k}" + }, + { + "bbox": [ + 305, + 130, + 546, + 273 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 305, + 130, + 546, + 273 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 305, + 130, + 546, + 273 + ], + "type": "text", + "content": " classes and their corresponding ground-truth labels " + }, + { + "bbox": [ + 305, + 130, + 546, + 273 + ], + "type": "inline_equation", + "content": "y_{N}" + }, + { + "bbox": [ + 305, + 130, + 546, + 273 + ], + "type": "text", + "content": ". It can be conceptualized as a linear two-layer model, where the first layer contains train image features " + }, + { + "bbox": [ + 305, + 130, + 546, + 273 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{\\mathrm{train}} = E_v(x_k)" + }, + { + "bbox": [ + 305, + 130, + 546, + 273 + ], + "type": "text", + "content": " and the second layer consists of one-hot vectors " + }, + { + "bbox": [ + 305, + 130, + 546, + 273 + ], + "type": "inline_equation", + "content": "\\mathbf{L}_{\\mathrm{train}}" + }, + { + "bbox": [ + 305, + 130, + 546, + 273 + ], + "type": "text", + "content": " encoded from the labels " + }, + { + "bbox": [ + 305, + 130, + 546, + 273 + ], + "type": "inline_equation", + "content": "y_{N}" + }, + { + "bbox": [ + 305, + 130, + 546, + 273 + ], + "type": "text", + "content": ". Given test image features " + }, + { + "bbox": [ + 305, + 130, + 546, + 273 + ], + "type": "inline_equation", + "content": "f_{\\mathrm{test}}" + }, + { + "bbox": [ + 305, + 130, + 546, + 273 + ], + "type": "text", + "content": " generated from the CLIP's image encoder " + }, + { + "bbox": [ + 305, + 130, + 546, + 273 + ], + "type": "inline_equation", + "content": "E_{v}" + }, + { + "bbox": [ + 305, + 130, + 546, + 273 + ], + "type": "text", + "content": ", the prediction from the cache model can be calculated as follows:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 349, + 281, + 545, + 297 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 349, + 281, + 545, + 297 + ], + "spans": [ + { + "bbox": [ + 349, + 281, + 545, + 297 + ], + "type": "interline_equation", + "content": "P _ {\\mathrm {c a c h e}} \\left(f _ {\\mathrm {t e s t}}\\right) = A \\left(f _ {\\mathrm {t e s t}} \\mathbf {F} _ {\\mathrm {t r a i n}} ^ {T}\\right) \\mathbf {L} _ {\\mathrm {t r a i n}}, \\tag {2}", + "image_path": "9e883337902b2ebd630e8d76acba1bc969ff124476504b2ed14378e251e6b287.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 304, + 545, + 366 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 304, + 545, + 366 + ], + "spans": [ + { + "bbox": [ + 305, + 304, + 545, + 366 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 304, + 545, + 366 + ], + "type": "inline_equation", + "content": "A(z) = \\alpha \\exp(-\\beta (1 - z))" + }, + { + "bbox": [ + 305, + 304, + 545, + 366 + ], + "type": "text", + "content": " is an adaptation function within a weighting factor " + }, + { + "bbox": [ + 305, + 304, + 545, + 366 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 305, + 304, + 545, + 366 + ], + "type": "text", + "content": " and a sharpness ratio " + }, + { + "bbox": [ + 305, + 304, + 545, + 366 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 305, + 304, + 545, + 366 + ], + "type": "text", + "content": ". During inference, the prediction of Tip-Adapter is computed by combining the pre-trained CLIP model and the cache model as: " + }, + { + "bbox": [ + 305, + 304, + 545, + 366 + ], + "type": "inline_equation", + "content": "P_{\\mathrm{TA}}(f_{\\mathrm{test}}) = P_{\\mathrm{cache}}(f_{\\mathrm{test}}) + f_{\\mathrm{test}} \\mathbf{W}_c^T" + }, + { + "bbox": [ + 305, + 304, + 545, + 366 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 372, + 477, + 385 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 372, + 477, + 385 + ], + "spans": [ + { + "bbox": [ + 306, + 372, + 477, + 385 + ], + "type": "text", + "content": "3.2. Training-free Dynamic Adapter" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 390, + 545, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 390, + 545, + 510 + ], + "spans": [ + { + "bbox": [ + 304, + 390, + 545, + 510 + ], + "type": "text", + "content": "During testing, pre-trained vision-language models like CLIP may encounter distribution shifts that degrade the classification performance. To address this issue, existing test-time prompt tuning methods train a learnable prompt by enforcing consistency across different image augmentations during testing. However, it requires a large number of augmentation operations on each test image and computationally-heavy optimization steps to learn the prompt, limiting its applicability in various real-world settings." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 510, + 546, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 510, + 546, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 510, + 546, + 714 + ], + "type": "text", + "content": "In this paper, our motivation is to design an efficient method for test-time adaptation of the pre-trained vision-language models like CLIP. Inspired by the concept of TipAdapter, we propose a training-free dynamic adapter (TDA) to enable efficient and effective test-time adaptation with CLIP. As shown in Figure 2, TDA includes two lightweight key-value caches, where each cache stores a dynamic queue of few-shot test features as keys and the corresponding pseudo labels as values. The first cache is intended for positive learning and it dynamically updates key-value pairs with high-confidence predictions to improve the accuracy. The second cache is designed for negative learning and it aims to address the adverse effects of noisy pseudo labels by introducing negative pseudo labeling to identify class absence rather than presence. By combining the positive cache and negative cache, the proposed TDA can achieve superior performance in terms of both speed and accuracy." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "14164" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 51, + 68, + 547, + 250 + ], + "blocks": [ + { + "bbox": [ + 51, + 68, + 547, + 250 + ], + "lines": [ + { + "bbox": [ + 51, + 68, + 547, + 250 + ], + "spans": [ + { + "bbox": [ + 51, + 68, + 547, + 250 + ], + "type": "image", + "image_path": "c29d8b07ccb6929f5a98a2ce188137beb7875f7b6c69de9ed5f29a2420f6b6b4.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 261, + 547, + 329 + ], + "lines": [ + { + "bbox": [ + 46, + 261, + 547, + 329 + ], + "spans": [ + { + "bbox": [ + 46, + 261, + 547, + 329 + ], + "type": "text", + "content": "Figure 2. Overview of the proposed Training-free Dynamic Adapter (TDA). TDA constructs and updates two key-value caches to store the knowledge of a stream of test samples, and uses the two caches to generate positive and negative predictions which are combined with CLIP predictions to produce the final prediction. Specifically, the CLIP predictions are generated by performing the dot product between the image features generated by CLIP's image encoder " + }, + { + "bbox": [ + 46, + 261, + 547, + 329 + ], + "type": "inline_equation", + "content": "E_{v}" + }, + { + "bbox": [ + 46, + 261, + 547, + 329 + ], + "type": "text", + "content": " and the text embeddings generated by CLIP's text encoder " + }, + { + "bbox": [ + 46, + 261, + 547, + 329 + ], + "type": "inline_equation", + "content": "E_{t}" + }, + { + "bbox": [ + 46, + 261, + 547, + 329 + ], + "type": "text", + "content": ", using the hand-crafted prompt and class names. The two key-value caches are updated by gradually incorporating the test features and their corresponding pseudo labels calculated from CLIP's predictions, based on prediction entropy and cache capacity." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 349, + 289, + 457 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 349, + 289, + 457 + ], + "spans": [ + { + "bbox": [ + 46, + 349, + 289, + 457 + ], + "type": "text", + "content": "Our goal is to conduct test-time adaptation by gathering adequate knowledge from the testing data stream and improving the predictions through the adaptation of the image features. To accomplish this goal, we create positive and negative caches that capture particular characteristics of the testing data stream. In the upcoming parts, we will present the process of collecting data for each cache and define the conditions in which we can utilize cache information to adapt features and enhance model predictions." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 475, + 287, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 475, + 287, + 666 + ], + "spans": [ + { + "bbox": [ + 46, + 475, + 287, + 666 + ], + "type": "text", + "content": "Positive Cache. The positive cache in TDA is a key-value cache, in which keys and values are represented as a dynamic queue. It aims to collect high-quality few-shot pseudo labels " + }, + { + "bbox": [ + 46, + 475, + 287, + 666 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{L}}_{\\mathfrak{p}}" + }, + { + "bbox": [ + 46, + 475, + 287, + 666 + ], + "type": "text", + "content": " as positive values and the corresponding features " + }, + { + "bbox": [ + 46, + 475, + 287, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}_{\\mathfrak{p}}" + }, + { + "bbox": [ + 46, + 475, + 287, + 666 + ], + "type": "text", + "content": " as keys. The key-value cache is initially empty and then accumulates a sufficient number of key-value pairs during the test-time adaptation. To maintain high-quality pseudo labels, TDA progressively incorporates test predictions with lower entropy while limiting shot capacity1 in the positive cache. Different from a normal queue like a FIFO queue with a fixed size, the dynamic queue in our method expands in size during testing. Besides, the dynamic queue operates similarly to a priority queue, using entropy as the criterion for prioritization. Note that, each class has its own queue to maintain the order and correct data structure of each class in the cache." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 667, + 287, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 667, + 287, + 693 + ], + "spans": [ + { + "bbox": [ + 47, + 667, + 287, + 693 + ], + "type": "text", + "content": "Given a pre-trained CLIP model that consists of a text encoder " + }, + { + "bbox": [ + 47, + 667, + 287, + 693 + ], + "type": "inline_equation", + "content": "E_{t}" + }, + { + "bbox": [ + 47, + 667, + 287, + 693 + ], + "type": "text", + "content": " and an image encoder " + }, + { + "bbox": [ + 47, + 667, + 287, + 693 + ], + "type": "inline_equation", + "content": "E_{v}" + }, + { + "bbox": [ + 47, + 667, + 287, + 693 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 47, + 667, + 287, + 693 + ], + "type": "inline_equation", + "content": "E_{t}" + }, + { + "bbox": [ + 47, + 667, + 287, + 693 + ], + "type": "text", + "content": " processes class" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 306, + 349, + 545, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 349, + 545, + 662 + ], + "spans": [ + { + "bbox": [ + 306, + 349, + 545, + 662 + ], + "type": "text", + "content": "names with pre-defined prompts to generate " + }, + { + "bbox": [ + 306, + 349, + 545, + 662 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 306, + 349, + 545, + 662 + ], + "type": "text", + "content": "-channeled text embeddings " + }, + { + "bbox": [ + 306, + 349, + 545, + 662 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_c" + }, + { + "bbox": [ + 306, + 349, + 545, + 662 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 306, + 349, + 545, + 662 + ], + "type": "inline_equation", + "content": "E_v" + }, + { + "bbox": [ + 306, + 349, + 545, + 662 + ], + "type": "text", + "content": " processes each test image " + }, + { + "bbox": [ + 306, + 349, + 545, + 662 + ], + "type": "inline_equation", + "content": "x_{\\mathrm{test}}" + }, + { + "bbox": [ + 306, + 349, + 545, + 662 + ], + "type": "text", + "content": " to produce image features " + }, + { + "bbox": [ + 306, + 349, + 545, + 662 + ], + "type": "inline_equation", + "content": "f_{\\mathrm{test}}" + }, + { + "bbox": [ + 306, + 349, + 545, + 662 + ], + "type": "text", + "content": ". To build the positive cache, TDA first generates a pseudo label " + }, + { + "bbox": [ + 306, + 349, + 545, + 662 + ], + "type": "inline_equation", + "content": "\\hat{l}" + }, + { + "bbox": [ + 306, + 349, + 545, + 662 + ], + "type": "text", + "content": ", a one-hot encoded vector of a categorical distribution, for each test sample " + }, + { + "bbox": [ + 306, + 349, + 545, + 662 + ], + "type": "inline_equation", + "content": "x_{\\mathrm{test}}" + }, + { + "bbox": [ + 306, + 349, + 545, + 662 + ], + "type": "text", + "content": " by applying the softmax function to the prediction " + }, + { + "bbox": [ + 306, + 349, + 545, + 662 + ], + "type": "inline_equation", + "content": "f_{\\mathrm{test}} \\mathbf{W}_c^T" + }, + { + "bbox": [ + 306, + 349, + 545, + 662 + ], + "type": "text", + "content": ". We establish two conditions to ascertain whether and how to include the pseudo label " + }, + { + "bbox": [ + 306, + 349, + 545, + 662 + ], + "type": "inline_equation", + "content": "\\hat{l}" + }, + { + "bbox": [ + 306, + 349, + 545, + 662 + ], + "type": "text", + "content": " and its corresponding image features " + }, + { + "bbox": [ + 306, + 349, + 545, + 662 + ], + "type": "inline_equation", + "content": "f_{\\mathrm{test}}" + }, + { + "bbox": [ + 306, + 349, + 545, + 662 + ], + "type": "text", + "content": " into the positive cache. The first condition is defined as: if the shot number (the number of collected pairs per class) of " + }, + { + "bbox": [ + 306, + 349, + 545, + 662 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{L}}_{\\mathbf{p}}" + }, + { + "bbox": [ + 306, + 349, + 545, + 662 + ], + "type": "text", + "content": " is less than the maximum shot capacity (the maximum capacity number of pairs per class) " + }, + { + "bbox": [ + 306, + 349, + 545, + 662 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 306, + 349, + 545, + 662 + ], + "type": "text", + "content": ", TDA will add " + }, + { + "bbox": [ + 306, + 349, + 545, + 662 + ], + "type": "inline_equation", + "content": "\\hat{l}" + }, + { + "bbox": [ + 306, + 349, + 545, + 662 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 306, + 349, + 545, + 662 + ], + "type": "inline_equation", + "content": "f_{\\mathrm{test}}" + }, + { + "bbox": [ + 306, + 349, + 545, + 662 + ], + "type": "text", + "content": " as a new value and a new key to " + }, + { + "bbox": [ + 306, + 349, + 545, + 662 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{L}}_{\\mathbf{p}}" + }, + { + "bbox": [ + 306, + 349, + 545, + 662 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 306, + 349, + 545, + 662 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}_{\\mathbf{p}}" + }, + { + "bbox": [ + 306, + 349, + 545, + 662 + ], + "type": "text", + "content": " respectively. Meanwhile, the second condition is defined as: if the shot number of " + }, + { + "bbox": [ + 306, + 349, + 545, + 662 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{L}}_{\\mathbf{p}}" + }, + { + "bbox": [ + 306, + 349, + 545, + 662 + ], + "type": "text", + "content": " has reached the maximum shot capacity " + }, + { + "bbox": [ + 306, + 349, + 545, + 662 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 306, + 349, + 545, + 662 + ], + "type": "text", + "content": ", TDA will replace an 'uncertain' key-value pair " + }, + { + "bbox": [ + 306, + 349, + 545, + 662 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{q}^{ent}, \\hat{l}_p^{ent}\\}" + }, + { + "bbox": [ + 306, + 349, + 545, + 662 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 306, + 349, + 545, + 662 + ], + "type": "inline_equation", + "content": "\\{f_{\\mathrm{test}}, \\hat{l}\\}" + }, + { + "bbox": [ + 306, + 349, + 545, + 662 + ], + "type": "text", + "content": " when " + }, + { + "bbox": [ + 306, + 349, + 545, + 662 + ], + "type": "inline_equation", + "content": "\\mathrm{H}(f_{\\mathrm{test}} \\mathbf{W}_c^T) < \\mathrm{H}(\\mathbf{q}^{ent} \\mathbf{W}_c^T)" + }, + { + "bbox": [ + 306, + 349, + 545, + 662 + ], + "type": "text", + "content": ". Here, " + }, + { + "bbox": [ + 306, + 349, + 545, + 662 + ], + "type": "inline_equation", + "content": "\\mathrm{H}" + }, + { + "bbox": [ + 306, + 349, + 545, + 662 + ], + "type": "text", + "content": " denotes the entropy function and the term 'uncertain' indicates that the entropy of a particular key-value pair is the highest compared to the entropy of all other key-value pairs of the same class in the cache model. By applying these two conditions, TDA can gradually integrate test predictions with lower entropy while controlling the shot capacity, which helps to ensure the collection of high-quality pseudo labels in positive cache." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 305, + 665, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 665, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 305, + 665, + 547, + 715 + ], + "type": "text", + "content": "During test-time adaptation, the positive cache can quickly retrieve relevant information by treating the image features " + }, + { + "bbox": [ + 305, + 665, + 547, + 715 + ], + "type": "inline_equation", + "content": "f_{\\mathrm{test}}" + }, + { + "bbox": [ + 305, + 665, + 547, + 715 + ], + "type": "text", + "content": " generated from a test example as a query and searching the stored key-value pairs " + }, + { + "bbox": [ + 305, + 665, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{Q}_{\\mathbf{p}},\\hat{\\mathbf{L}}_{\\mathbf{p}}\\}" + }, + { + "bbox": [ + 305, + 665, + 547, + 715 + ], + "type": "text", + "content": " for match-" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 58, + 702, + 280, + 714 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 702, + 280, + 714 + ], + "spans": [ + { + "bbox": [ + 58, + 702, + 280, + 714 + ], + "type": "text", + "content": "1The shot capacity refers to the maximum number of pairs per class." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14165" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 95 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 95 + ], + "type": "text", + "content": "ing information. The adapted predictions using the positive cache can then be obtained as:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 104, + 101, + 287, + 117 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 104, + 101, + 287, + 117 + ], + "spans": [ + { + "bbox": [ + 104, + 101, + 287, + 117 + ], + "type": "interline_equation", + "content": "P _ {\\mathrm {p o s}} \\left(f _ {\\mathrm {t e s t}}\\right) = A \\left(f _ {\\mathrm {t e s t}} \\mathbf {Q} _ {\\mathbf {p}} ^ {\\mathbf {T}}\\right) \\hat {\\mathbf {L}} _ {\\mathbf {p}}, \\tag {3}", + "image_path": "3a584015414426d5304565c85714e3c5ec012880d642f45c9c18fb721ea1ef79.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 121, + 284, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 121, + 284, + 133 + ], + "spans": [ + { + "bbox": [ + 47, + 121, + 284, + 133 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 121, + 284, + 133 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 47, + 121, + 284, + 133 + ], + "type": "text", + "content": " is the adaptation function defined in Tip-Adapter." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 146, + 287, + 242 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 146, + 287, + 242 + ], + "spans": [ + { + "bbox": [ + 46, + 146, + 287, + 242 + ], + "type": "text", + "content": "Negative Cache. Similar to the positive cache in our TDA, the negative cache is also a dynamic queue structure with negative keys and negative values denoted as " + }, + { + "bbox": [ + 46, + 146, + 287, + 242 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}_{\\mathbf{n}}" + }, + { + "bbox": [ + 46, + 146, + 287, + 242 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 146, + 287, + 242 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{L}}_{\\mathbf{n}}" + }, + { + "bbox": [ + 46, + 146, + 287, + 242 + ], + "type": "text", + "content": ", respectively. It aims to gather CLIP-generated image features to " + }, + { + "bbox": [ + 46, + 146, + 287, + 242 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}_{\\mathbf{n}}" + }, + { + "bbox": [ + 46, + 146, + 287, + 242 + ], + "type": "text", + "content": " and the corresponding negative pseudo labels to " + }, + { + "bbox": [ + 46, + 146, + 287, + 242 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{L}}_{\\mathbf{n}}" + }, + { + "bbox": [ + 46, + 146, + 287, + 242 + ], + "type": "text", + "content": ". Unlike the pseudo labels in the positive cache, the negative pseudo labels are obtained by applying negative mask on the class probabilities as:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 116, + 246, + 287, + 262 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 116, + 246, + 287, + 262 + ], + "spans": [ + { + "bbox": [ + 116, + 246, + 287, + 262 + ], + "type": "interline_equation", + "content": "\\hat {\\mathbf {L}} _ {\\mathbf {n}} = - \\mathbb {1} [ p _ {l} < P (\\mathbf {Q} _ {\\mathbf {n}}) ], \\tag {4}", + "image_path": "5077f450023f8b4da66a2f0c2e8ba5d248457ad96d71753573d01cbb23f3230e.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 266, + 287, + 386 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 266, + 287, + 386 + ], + "spans": [ + { + "bbox": [ + 46, + 266, + 287, + 386 + ], + "type": "text", + "content": "where higher probabilities than " + }, + { + "bbox": [ + 46, + 266, + 287, + 386 + ], + "type": "inline_equation", + "content": "p_l" + }, + { + "bbox": [ + 46, + 266, + 287, + 386 + ], + "type": "text", + "content": " are selected as negative pseudo labels from uncertain predictions and the uncertainty is measured by the entropy of predictions. Here, " + }, + { + "bbox": [ + 46, + 266, + 287, + 386 + ], + "type": "inline_equation", + "content": "p_l" + }, + { + "bbox": [ + 46, + 266, + 287, + 386 + ], + "type": "text", + "content": " represents a threshold in negative pseudo labeling, and " + }, + { + "bbox": [ + 46, + 266, + 287, + 386 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{L}}_{\\mathbf{n}}" + }, + { + "bbox": [ + 46, + 266, + 287, + 386 + ], + "type": "text", + "content": " denotes a negative pseudo label, which is a vector whose elements larger than " + }, + { + "bbox": [ + 46, + 266, + 287, + 386 + ], + "type": "inline_equation", + "content": "p_l" + }, + { + "bbox": [ + 46, + 266, + 287, + 386 + ], + "type": "text", + "content": " have a value -1 and otherwise 0. Different from existing negative learning methods [22, 23, 33] that select negative labels from all noisy labels, TDA selects negative pseudo labels from uncertain predictions to avoid bias to the data with certain predictions." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 386, + 287, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 386, + 287, + 434 + ], + "spans": [ + { + "bbox": [ + 47, + 386, + 287, + 434 + ], + "type": "text", + "content": "When constructing the negative cache, the testing feature " + }, + { + "bbox": [ + 47, + 386, + 287, + 434 + ], + "type": "inline_equation", + "content": "f_{\\mathrm{test}}" + }, + { + "bbox": [ + 47, + 386, + 287, + 434 + ], + "type": "text", + "content": " will be included in negative cache if it satisfies the condition " + }, + { + "bbox": [ + 47, + 386, + 287, + 434 + ], + "type": "inline_equation", + "content": "\\gamma (f_{\\mathrm{test}})" + }, + { + "bbox": [ + 47, + 386, + 287, + 434 + ], + "type": "text", + "content": ": the entropy of the prediction is in the specified interval between " + }, + { + "bbox": [ + 47, + 386, + 287, + 434 + ], + "type": "inline_equation", + "content": "\\tau_{l}" + }, + { + "bbox": [ + 47, + 386, + 287, + 434 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 386, + 287, + 434 + ], + "type": "inline_equation", + "content": "\\tau_{h}" + }, + { + "bbox": [ + 47, + 386, + 287, + 434 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 96, + 439, + 287, + 453 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 96, + 439, + 287, + 453 + ], + "spans": [ + { + "bbox": [ + 96, + 439, + 287, + 453 + ], + "type": "interline_equation", + "content": "\\gamma \\left(f _ {\\text {t e s t}}\\right): \\tau_ {l} < \\mathrm {H} \\left(f _ {\\text {t e s t}} \\mathbf {W} _ {c} ^ {T}\\right) < \\tau_ {h}. \\tag {5}", + "image_path": "4dba60a2054e10175e9daf02d399a4aaf89e0b5306e6237d6377518f313e9873.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 458, + 287, + 566 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 458, + 287, + 566 + ], + "spans": [ + { + "bbox": [ + 46, + 458, + 287, + 566 + ], + "type": "text", + "content": "This condition is designed to mitigate the risk of prediction errors due to high entropy or be biased to certain predictions (characterized by very low entropy), by incorporating test samples that exhibit a moderate degree of prediction uncertainty. Once the " + }, + { + "bbox": [ + 46, + 458, + 287, + 566 + ], + "type": "inline_equation", + "content": "\\gamma(f_{\\mathrm{test}})" + }, + { + "bbox": [ + 46, + 458, + 287, + 566 + ], + "type": "text", + "content": " check is completed, the remaining steps for collecting uncertain samples in the negative cache follow the same two conditions designed in the positive cache. Similar to the positive cache, TDA also limits the shot capacity " + }, + { + "bbox": [ + 46, + 458, + 287, + 566 + ], + "type": "inline_equation", + "content": "\\tilde{k}" + }, + { + "bbox": [ + 46, + 458, + 287, + 566 + ], + "type": "text", + "content": " in the negative cache." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 566, + 287, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 566, + 287, + 613 + ], + "spans": [ + { + "bbox": [ + 46, + 566, + 287, + 613 + ], + "type": "text", + "content": "During test-time adaption, the testing features " + }, + { + "bbox": [ + 46, + 566, + 287, + 613 + ], + "type": "inline_equation", + "content": "f_{\\mathrm{test}}" + }, + { + "bbox": [ + 46, + 566, + 287, + 613 + ], + "type": "text", + "content": " can be quickly adapted to target domains by retrieving the knowledge from " + }, + { + "bbox": [ + 46, + 566, + 287, + 613 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{Q}_{\\mathbf{n}}, \\hat{\\mathbf{L}}_{\\mathbf{n}}\\}" + }, + { + "bbox": [ + 46, + 566, + 287, + 613 + ], + "type": "text", + "content": " in the negative cache and the adapted prediction can be obtained as:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 100, + 618, + 287, + 633 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 100, + 618, + 287, + 633 + ], + "spans": [ + { + "bbox": [ + 100, + 618, + 287, + 633 + ], + "type": "interline_equation", + "content": "P _ {\\text {n e g}} \\left(f _ {\\text {t e s t}}\\right) = - A \\left(f _ {\\text {t e s t}} \\mathbf {Q} _ {\\mathbf {n}} ^ {\\mathbf {T}}\\right) \\hat {\\mathbf {L}} _ {\\mathbf {n}}, \\tag {6}", + "image_path": "bc4a9dab21872b63c95c24c76698c22722f5019b267863684572b4f52a0c749a.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 638, + 287, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 638, + 287, + 685 + ], + "spans": [ + { + "bbox": [ + 46, + 638, + 287, + 685 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 638, + 287, + 685 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 46, + 638, + 287, + 685 + ], + "type": "text", + "content": " is the adaptation function defined in Tip-Adapter. The predictions of TDA can be formulated by combining the negative cache, the positive cache and the pre-trained CLIP model together as follows:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 52, + 691, + 287, + 706 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 52, + 691, + 287, + 706 + ], + "spans": [ + { + "bbox": [ + 52, + 691, + 287, + 706 + ], + "type": "interline_equation", + "content": "P _ {\\mathrm {T D A}} \\left(f _ {\\text {t e s t}}\\right) = f _ {\\text {t e s t}} \\mathbf {W} _ {c} ^ {T} + P _ {\\text {p o s}} \\left(f _ {\\text {t e s t}}\\right) + P _ {\\text {n e g}} \\left(f _ {\\text {t e s t}}\\right). \\tag {7}", + "image_path": "820701cc08e7d45cfa14b6423a734f5a02d58cb29652f8d9594019ec9f9da006.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 72, + 517, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 517, + 85 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 517, + 85 + ], + "type": "text", + "content": "3.3. Relationship with TPT and Tip-Adapter" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 95, + 545, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 95, + 545, + 205 + ], + "spans": [ + { + "bbox": [ + 304, + 95, + 545, + 205 + ], + "type": "text", + "content": "Both TPT and TDA are designed to handle the challenge of adapting models to test data that have distributional discrepancies from the training data. TPT trains a learnable prompt " + }, + { + "bbox": [ + 304, + 95, + 545, + 205 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_c" + }, + { + "bbox": [ + 304, + 95, + 545, + 205 + ], + "type": "text", + "content": " in Eq. (1) for each test sample with a large number of augmentations and such training process requires backpropagation and is computationally intensive. In contrast, our TDA is training-free as both positive cache " + }, + { + "bbox": [ + 304, + 95, + 545, + 205 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{Q_p},\\hat{\\mathbf{L}}_{\\mathbf{p}}\\}" + }, + { + "bbox": [ + 304, + 95, + 545, + 205 + ], + "type": "text", + "content": " and negative cache " + }, + { + "bbox": [ + 304, + 95, + 545, + 205 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{Q_n},\\hat{\\mathbf{L}}_{\\mathbf{n}}\\}" + }, + { + "bbox": [ + 304, + 95, + 545, + 205 + ], + "type": "text", + "content": " are non-parametric, which is super-efficient without incurring any backpropagation." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 209, + 545, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 209, + 545, + 413 + ], + "spans": [ + { + "bbox": [ + 304, + 209, + 545, + 413 + ], + "type": "text", + "content": "Our TDA employs a cache model that shares similarities with the Tip-Adapter, where features and labels are stored as key-value pairs in a memory cache. One notable distinction between the Tip-Adapter and TDA lies in the type of cache used. Specifically, the Tip-Adapter relies on a static cache " + }, + { + "bbox": [ + 304, + 209, + 545, + 413 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{F}_{\\mathrm{train}},\\mathbf{L}_{\\mathrm{train}}\\}" + }, + { + "bbox": [ + 304, + 209, + 545, + 413 + ], + "type": "text", + "content": " because it is designed for a supervised adaptation setting, where the ground-truth labels " + }, + { + "bbox": [ + 304, + 209, + 545, + 413 + ], + "type": "inline_equation", + "content": "\\mathbf{L}_{\\mathrm{train}}" + }, + { + "bbox": [ + 304, + 209, + 545, + 413 + ], + "type": "text", + "content": " are predetermined and readily available. In contrast, our TDA introduces a new dynamic cache " + }, + { + "bbox": [ + 304, + 209, + 545, + 413 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{Q_p},\\hat{\\mathbf{L}}_p\\}" + }, + { + "bbox": [ + 304, + 209, + 545, + 413 + ], + "type": "text", + "content": " designed for a test-time adaptation setting, where the pseudo labels " + }, + { + "bbox": [ + 304, + 209, + 545, + 413 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{L}}_{\\mathbf{p}}" + }, + { + "bbox": [ + 304, + 209, + 545, + 413 + ], + "type": "text", + "content": " are generated on-the-fly from a stream of test samples. Furthermore, TDA incorporates a novel negative cache model " + }, + { + "bbox": [ + 304, + 209, + 545, + 413 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{Q_n},\\hat{\\mathbf{L}}_n\\}" + }, + { + "bbox": [ + 304, + 209, + 545, + 413 + ], + "type": "text", + "content": " that enhances testing predictions by utilizing indirect knowledge that a test image does not belong to certain negative classes. By combining positive and negative caches, our proposed TDA is more robust to noisy pseudo labels and can generalize well to testing data." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 438, + 387, + 452 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 438, + 387, + 452 + ], + "spans": [ + { + "bbox": [ + 306, + 438, + 387, + 452 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 463, + 422, + 475 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 463, + 422, + 475 + ], + "spans": [ + { + "bbox": [ + 306, + 463, + 422, + 475 + ], + "type": "text", + "content": "4.1. Experimental Setup" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 486, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 486, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 486, + 545, + 713 + ], + "type": "text", + "content": "Benchmarks. We conducted main experiments on two benchmarks: out-of-distribution (OOD) benchmark and cross-domain benchmark, both applied in the previous work [35] that adapts vision-language models in test time. The OOD benchmark serves as a measure of the robustness of our approach by involving assessment on 4 out-of-distribution datasets derived from ImageNet [5]: ImageNet-A [16], ImageNet-V2 [32], ImageNet-R [15], and ImageNet-S [41]. This benchmark is specifically designed to evaluate a model's capacity to generalize to new and unseen data. The cross-domain benchmark, on the other hand, is involved to evaluate the model's performance across 10 diverse image classification datasets, each from a distinct domain with different classes: Aircraft [26], Caltech101 [8], Cars [24], DTD [4], EuroSAT [14], Flower102 [28], Food101 [1], Pets [30], SUN397 [42], and UCF101 [36]. This benchmark provides a comprehensive evaluation of the model's adaptability during test time across various class spaces." + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14166" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 50, + 70, + 544, + 270 + ], + "blocks": [ + { + "bbox": [ + 50, + 70, + 544, + 270 + ], + "lines": [ + { + "bbox": [ + 50, + 70, + 544, + 270 + ], + "spans": [ + { + "bbox": [ + 50, + 70, + 544, + 270 + ], + "type": "table", + "html": "
MethodImageNetImageNet-AImageNet-V2ImageNet-RImageNet-SAverageOOD Average
CLIP-ResNet-5059.8123.2452.9160.7235.4846.4343.09
CoOp63.3323.0655.4056.6034.6746.6142.43
CoCoOp62.8123.3255.7257.7434.4846.8142.82
Tip-Adapter62.0323.1353.9760.3535.7447.0443.30
TPT60.7426.6754.7059.1135.0947.2643.89
DiffTPT60.8031.0655.8058.8037.1048.7145.69
TDA (Ours)61.3530.2955.5462.5838.1249.5846.63
CLIP-ViT-B/1668.3449.8961.8877.6548.2461.2059.42
CoOp71.5149.7164.2075.2147.9961.7259.28
CoCoOp71.0250.6364.0776.1848.7562.1359.91
Tip-Adapter70.7551.0463.4177.7648.8862.3760.27
TPT68.9854.7763.4577.0647.9462.4460.81
DiffTPT70.3055.6865.1075.0046.8062.2860.52
TDA (Ours)69.5160.1164.6780.2450.5465.0163.89
", + "image_path": "5c6859d6292bfbc1c78f9c699710b10037fe858960d85f6f24cb7861db37ee53.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 52, + 365, + 282, + 437 + ], + "blocks": [ + { + "bbox": [ + 46, + 278, + 546, + 346 + ], + "lines": [ + { + "bbox": [ + 46, + 278, + 546, + 346 + ], + "spans": [ + { + "bbox": [ + 46, + 278, + 546, + 346 + ], + "type": "text", + "content": "Table 1. Results on the OOD Benchmark. Our TDA is compared with several state-of-the-art methods designed for vision-language models: the baseline method CLIP, three train-time adaptation methods (i.e., CoOp, CoCoOp, and Tip-Adapter), and two test-time adaptation methods (i.e., TPT and DiffTPT). All the compared methods are built upon CLIP-ResNet-50 or CLIP-ViT-B/16 baselines. The two evaluation metrics Average and OOD Average are calculated by taking the mean accuracy across all five datasets and four OOD datasets excluding ImageNet. The results of CLIP, CoOp, CoCoOp, and TPT are obtained from the TPT paper, the results of DiffTPT are obtained from the DiffTPT paper, while the results of Tip-Adapter are reproduced using the official codes." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 52, + 365, + 282, + 437 + ], + "lines": [ + { + "bbox": [ + 52, + 365, + 282, + 437 + ], + "spans": [ + { + "bbox": [ + 52, + 365, + 282, + 437 + ], + "type": "table", + "html": "
MethodTesting TimeAccuracyGain
CLIP-ResNet-5012min59.810
TPT12h 50min60.74+0.93
DiffTPT34h 45min60.80+0.99
TDA (Ours)16min61.35+1.54
", + "image_path": "b82f94048bf44351ef676b5370ed3b21bcbcd8bf85bb3a3513f45de9f9dfeeef.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 445, + 287, + 522 + ], + "lines": [ + { + "bbox": [ + 46, + 445, + 287, + 522 + ], + "spans": [ + { + "bbox": [ + 46, + 445, + 287, + 522 + ], + "type": "text", + "content": "Table 2. Comparisons of our TDA with CLIP-ResNet-50, TPT, and DiffTPT in terms of efficiency (Testing Time) and effectiveness (Accuracy). The final column shows the accuracy gain relative to the baseline CLIP. Note that the testing time of DiffTPT does not include the duration required for the image generation process with pre-trained diffusion models, which is an additional time-consuming factor during the testing phase." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 545, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 545, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 545, + 288, + 713 + ], + "type": "text", + "content": "Implementation details. All the models in our experiments are built upon the pre-trained CLIP model [31] that consists of an image encoder and a text encoder. The image encoder can be either ResNet [13] or Vision Transformer [7], while the text encoder is Transformer [39]. Test-time adaptation is set for single-image scenarios, using a batch size of 1. We conduct a search for all our hyperparameters using a single ImageNet validation set. The threshold " + }, + { + "bbox": [ + 46, + 545, + 288, + 713 + ], + "type": "inline_equation", + "content": "p_l" + }, + { + "bbox": [ + 46, + 545, + 288, + 713 + ], + "type": "text", + "content": " for negative pseudo-labeling in Eq 4 is set as 0.03. The upper and lower thresholds " + }, + { + "bbox": [ + 46, + 545, + 288, + 713 + ], + "type": "inline_equation", + "content": "[\\tau_l, \\tau_h]" + }, + { + "bbox": [ + 46, + 545, + 288, + 713 + ], + "type": "text", + "content": " for testing feature selection in Eq 5 are set as [0.2, 0.5]. Once searched, these hyperparameters are fixed and evaluated across various new datasets. To avoid incurring backpropagation when using learnable prompt, we follow [31] to use hand-crafted" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 366, + 545, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 366, + 545, + 413 + ], + "spans": [ + { + "bbox": [ + 304, + 366, + 545, + 413 + ], + "type": "text", + "content": "prompts. We use top-1 accuracy " + }, + { + "bbox": [ + 304, + 366, + 545, + 413 + ], + "type": "inline_equation", + "content": "(\\%)" + }, + { + "bbox": [ + 304, + 366, + 545, + 413 + ], + "type": "text", + "content": ", a standard classification criterion, as our evaluation metric. All the experiments are conducted using a single NVIDIA Quadro RTX 6000 GPU." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 305, + 420, + 491, + 433 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 420, + 491, + 433 + ], + "spans": [ + { + "bbox": [ + 305, + 420, + 491, + 433 + ], + "type": "text", + "content": "4.2. Comparisons with State-of-the-art" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 438, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 438, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 438, + 545, + 713 + ], + "type": "text", + "content": "In this section, we compare our proposed TDA with several state-of-the-art methods, including CLIP [31], three train-time adaptation methods, i.e., CoOp [50], CoCoOp [49], and Tip-Adapter [47], as well as two existing test-time adaptation methods TPT [35] and DiffTPT [9], all of which are designed for vision-language models. Specifically, CLIP is evaluated using an ensemble of 80 handcrafted prompts as in [31]. All train-time adaptation methods are trained on the ImageNet train set with 16 shots per class and tested on other datasets as in [35]. CoOp utilizes a learnable context module of size 4 that is fine-tuned for specific downstream tasks, while CoCoOp is the generalized version of CoOp with the input-conditional token for image features. Tip-Adapter employs the optimal hyperparameters obtained from the ImageNet validation set during evaluation. We would like to note that Tip-Adapter is unable to handle new classes during testing, limiting its implementation to OOD benchmark evaluation where the training classes encompass all the testing classes. Different from train-time adaptation methods, test-time adaptation methods (i.e., TPT, DiffTPT, and our TDA) do not utilize the ImageNet train set. Instead, they are fine-tuned with target datasets using a stream of unlabeled test samples. Follow" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "14167" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 50, + 70, + 544, + 232 + ], + "blocks": [ + { + "bbox": [ + 50, + 70, + 544, + 232 + ], + "lines": [ + { + "bbox": [ + 50, + 70, + 544, + 232 + ], + "spans": [ + { + "bbox": [ + 50, + 70, + 544, + 232 + ], + "type": "table", + "html": "
MethodAircraftCaltech101CarsDTDEuroSATFlower102Food101PetsSUN397UCF101Average
CLIP-ResNet-5016.1187.2655.8940.3725.7962.7774.8282.9760.8559.4856.63
CoOp15.1286.5355.3237.2926.2061.5575.5987.0058.1559.0556.18
CoCoOp14.6187.3856.2238.5328.7365.5776.2088.3959.6157.1057.23
TPT17.5887.0258.4640.8428.3362.6974.8884.4961.4660.8257.66
DiffTPT17.6086.8960.7140.7241.0463.5379.2183.4062.7262.6759.85
TDA (Ours)17.6189.7057.7843.7442.1168.7477.7586.1862.5364.1861.03
CLIP-ViT-B/1623.2293.5566.1145.0450.4266.9982.8686.9265.6365.1664.59
CoOp18.4793.7064.5141.9246.3968.7185.3089.1464.1566.5563.88
CoCoOp22.2993.7964.9045.4539.2370.8583.9790.4666.8968.4464.63
TPT24.7894.1666.8747.7542.4468.9884.6787.7965.5068.0465.10
DiffTPT25.6092.4967.0147.0043.1370.1087.2388.2265.7462.6765.47
TDA (Ours)23.9194.2467.2847.4058.0071.4286.1488.6367.6270.6667.53
", + "image_path": "b579dcdc8cca83c65012405cf7013f9f57878ef8dd7b9a145bf00fb726155893.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 239, + 546, + 296 + ], + "lines": [ + { + "bbox": [ + 48, + 239, + 546, + 296 + ], + "spans": [ + { + "bbox": [ + 48, + 239, + 546, + 296 + ], + "type": "text", + "content": "Table 3. Results on the Cross-Domain Benchmark. Our TDA is compared with several state-of-the-art methods designed for vision-language models: the baseline method CLIP, two train-time adaptation methods (i.e., CoOp and CoCoOp), and two test-time adaptation methods (i.e., TPT and DiffTPT). Note that Tip-Adapter is unable to be evaluated on the Cross-Domain Benchmark as it cannot handle new classes during testing. The evaluation metric Average is calculated by taking the mean accuracy across all ten datasets. The results of CLIP, CoOp, CoCoOp, and TPT are obtained from the TPT paper, while the results of DiffTPT are obtained from the DiffTPT paper." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 49, + 326, + 286, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 326, + 286, + 361 + ], + "spans": [ + { + "bbox": [ + 49, + 326, + 286, + 361 + ], + "type": "text", + "content": "ing TPT and DiffTPT, we compare TDA with the state-of-the-art over two public benchmarks: OOD benchmark and cross-domain benchmark." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 378, + 286, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 378, + 286, + 544 + ], + "spans": [ + { + "bbox": [ + 49, + 378, + 286, + 544 + ], + "type": "text", + "content": "Results on the OOD Benchmark. We first compare TDA with state-of-the-art methods over the OOD benchmark. Table 1 presents the experimental results, highlighting the superior performance of the proposed TDA compared to both TPT and DiffTPT across various OOD datasets derived from ImageNet. Specifically, TDA outperforms TPT on both ResNet-50 and ViT-B/16 architectures, improving OOD accuracy by " + }, + { + "bbox": [ + 49, + 378, + 286, + 544 + ], + "type": "inline_equation", + "content": "2.74\\%" + }, + { + "bbox": [ + 49, + 378, + 286, + 544 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 49, + 378, + 286, + 544 + ], + "type": "inline_equation", + "content": "3.08\\%" + }, + { + "bbox": [ + 49, + 378, + 286, + 544 + ], + "type": "text", + "content": " on average, respectively. Furthermore, compared to DiffTPT, TDA exhibits an average accuracy improvement of " + }, + { + "bbox": [ + 49, + 378, + 286, + 544 + ], + "type": "inline_equation", + "content": "0.94\\%" + }, + { + "bbox": [ + 49, + 378, + 286, + 544 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 49, + 378, + 286, + 544 + ], + "type": "inline_equation", + "content": "3.37\\%" + }, + { + "bbox": [ + 49, + 378, + 286, + 544 + ], + "type": "text", + "content": " in the OOD benchmark for ResNet-50 and ViT-B/16, respectively. These results validate the effectiveness of TDA in enhancing test-time adaptation performance on various OOD test datasets." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 49, + 545, + 286, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 545, + 286, + 712 + ], + "spans": [ + { + "bbox": [ + 49, + 545, + 286, + 712 + ], + "type": "text", + "content": "In order to provide a more comprehensive evaluation of our proposed method's efficiency and effectiveness, we compared it with the baseline CLIP-ResNet-50 and two existing test-time adaptation methods (i.e., TPT and DiffTPT). This comparison encompasses both testing time and testing accuracy, and the corresponding results are shown in Table 2. This evaluation is performed on the ImageNet validation dataset, which consists of 50,000 images, using a single NVIDIA Quadro RTX 6000 GPU. When compared to the baseline CLIP-ResNet-50, the proposed TDA demonstrates a significant improvement in testing accuracy " + }, + { + "bbox": [ + 49, + 545, + 286, + 712 + ], + "type": "inline_equation", + "content": "(+1.54\\%)" + }, + { + "bbox": [ + 49, + 545, + 286, + 712 + ], + "type": "text", + "content": ", with only a minimal sacrifice in testing efficiency (requiring an additional 4min). In comparison to TPT and DiffTPT, the proposed TDA demonstrates not only superior" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 308, + 326, + 545, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 326, + 545, + 445 + ], + "spans": [ + { + "bbox": [ + 308, + 326, + 545, + 445 + ], + "type": "text", + "content": "testing accuracy but also significantly improved efficiency. It reduces the testing time dramatically from " + }, + { + "bbox": [ + 308, + 326, + 545, + 445 + ], + "type": "inline_equation", + "content": "12\\mathrm{h}50\\mathrm{min}" + }, + { + "bbox": [ + 308, + 326, + 545, + 445 + ], + "type": "text", + "content": " by TPT and even more from " + }, + { + "bbox": [ + 308, + 326, + 545, + 445 + ], + "type": "inline_equation", + "content": "34\\mathrm{h}45\\mathrm{min}" + }, + { + "bbox": [ + 308, + 326, + 545, + 445 + ], + "type": "text", + "content": " by DiffTPT, down to just 16 minutes. Without including the image generation time, DiffTPT consumes clearly more test time than TPT as it involves a time-consuming multi-step prompt updating process whereas TPT requires a single step only. The experimental results strongly validate the effectiveness and efficiency of our proposed method, establishing its suitability for real-world applications." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 308, + 447, + 545, + 625 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 447, + 545, + 625 + ], + "spans": [ + { + "bbox": [ + 308, + 447, + 545, + 625 + ], + "type": "text", + "content": "We then compare TDA with state-of-the-art methods over cross-domain benchmark. The results, presented in Table 3, demonstrate that TDA not only surpasses the performance of the TPT method but also shows a significant advantage over its improvement method DiffTPT. Specifically, when utilizing CLIP-ResNet-50 and CLIP-ViT-B/16 as the backbone, TDA achieves an improvement in average accuracy over TPT by " + }, + { + "bbox": [ + 308, + 447, + 545, + 625 + ], + "type": "inline_equation", + "content": "3.37\\%" + }, + { + "bbox": [ + 308, + 447, + 545, + 625 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 308, + 447, + 545, + 625 + ], + "type": "inline_equation", + "content": "2.43\\%" + }, + { + "bbox": [ + 308, + 447, + 545, + 625 + ], + "type": "text", + "content": ", respectively. These improvements, along with a " + }, + { + "bbox": [ + 308, + 447, + 545, + 625 + ], + "type": "inline_equation", + "content": "1.18\\%" + }, + { + "bbox": [ + 308, + 447, + 545, + 625 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 308, + 447, + 545, + 625 + ], + "type": "inline_equation", + "content": "2.06\\%" + }, + { + "bbox": [ + 308, + 447, + 545, + 625 + ], + "type": "text", + "content": " gain over DiffTPT for the respective backbones, further verify the effectiveness of TDA in adapting to diverse class datasets during test time. This attribute holds significant value for vision-language models such as CLIP, as it enables them to classify arbitrary classes in image classification without the need for additional training." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 308, + 635, + 405, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 635, + 405, + 647 + ], + "spans": [ + { + "bbox": [ + 308, + 635, + 405, + 647 + ], + "type": "text", + "content": "4.3. Ablation Studies" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 308, + 654, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 654, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 308, + 654, + 545, + 712 + ], + "type": "text", + "content": "In this section, we perform ablation studies to examine the effectiveness of our designs. All the ablation studies are conducted over the ImageNet dataset, where TDA can achieve an accuracy of " + }, + { + "bbox": [ + 308, + 654, + 545, + 712 + ], + "type": "inline_equation", + "content": "61.35\\%" + }, + { + "bbox": [ + 308, + 654, + 545, + 712 + ], + "type": "text", + "content": " under default settings. TDA consists of a Positive Cache and a Negative Cache, which" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 750, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 750, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 750, + 317, + 757 + ], + "type": "text", + "content": "14168" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 72, + 286, + 251 + ], + "blocks": [ + { + "bbox": [ + 50, + 72, + 286, + 251 + ], + "lines": [ + { + "bbox": [ + 50, + 72, + 286, + 251 + ], + "spans": [ + { + "bbox": [ + 50, + 72, + 286, + 251 + ], + "type": "image", + "image_path": "83ab17c90bae8b8ba66f047b4294bcf5a80889fa31294d1162de42bad164366d.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 258, + 288, + 291 + ], + "lines": [ + { + "bbox": [ + 46, + 258, + 288, + 291 + ], + "spans": [ + { + "bbox": [ + 46, + 258, + 288, + 291 + ], + "type": "text", + "content": "Figure 3. Ablation studies on two cache designs in TDA: Positive Cache and Negative Cache. All the models are built upon the baseline model CLIP-ResNet-50." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 45, + 316, + 287, + 520 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 316, + 287, + 520 + ], + "spans": [ + { + "bbox": [ + 45, + 316, + 287, + 520 + ], + "type": "text", + "content": "perform positive and negative pseudo labeling within our designed dynamic adapter, respectively. We first assess the efficacy of the two cache designs in TDA. As shown in Figure 3, both Positive Cache and Negative Cache significantly surpass the baseline model CLIP, demonstrating that test-time adaptation can be improved by introducing a dynamic adapter with either positive pseudo labeling or negative pseudo labeling. Besides, the two cache designs in TDA can complement each other as TDA (i.e., the combination of the two designs) clearly outperforms either Positive Cache or Negative Cache on the challenging ImageNet dataset. Moreover, we extend our ablation studies to the cross-domain benchmark, and results show that the positive cache achieved " + }, + { + "bbox": [ + 45, + 316, + 287, + 520 + ], + "type": "inline_equation", + "content": "60.38\\%" + }, + { + "bbox": [ + 45, + 316, + 287, + 520 + ], + "type": "text", + "content": " accuracy, the negative cache " + }, + { + "bbox": [ + 45, + 316, + 287, + 520 + ], + "type": "inline_equation", + "content": "60.11\\%" + }, + { + "bbox": [ + 45, + 316, + 287, + 520 + ], + "type": "text", + "content": ", while their combination yielded a " + }, + { + "bbox": [ + 45, + 316, + 287, + 520 + ], + "type": "inline_equation", + "content": "61.03\\%" + }, + { + "bbox": [ + 45, + 316, + 287, + 520 + ], + "type": "text", + "content": " accuracy, thereby highlighting the significance of each type of cache in enhancing TDA's performance." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 522, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 522, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 522, + 288, + 714 + ], + "type": "text", + "content": "We proceed to perform parameter studies on the shot capacity, which refers to the maximum number of key-value pairs per class, in both Positive Cache and Negative Cache models. These studies aims to find the optimal balance between the diversity and accuracy of the key-value pairs. Figure 4 shows that the performance of both cache models is significantly affected when the shot capacity is either too low or too high. We find that the shot capacity is set as 3 for the Positive Cache and 2 for the Negative Cache yields the best performance. This is because an appropriate shot capacity ensures both high-quality pseudo labels (paired values) and diverse image features (paired keys) in the cache models. The negative cache is sensitive to shot numbers due to its role in storing probabilities for various negative pseudo labels, each representing a class to be excluded from the model. Contrary to the intuition that a larger negative" + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 309, + 72, + 545, + 255 + ], + "blocks": [ + { + "bbox": [ + 309, + 72, + 545, + 255 + ], + "lines": [ + { + "bbox": [ + 309, + 72, + 545, + 255 + ], + "spans": [ + { + "bbox": [ + 309, + 72, + 545, + 255 + ], + "type": "image", + "image_path": "0003c7a8fe921051a3ee6ea5ce1be61aef18c4e9274ecd6118067ace10d5425a.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 260, + 547, + 283 + ], + "lines": [ + { + "bbox": [ + 305, + 260, + 547, + 283 + ], + "spans": [ + { + "bbox": [ + 305, + 260, + 547, + 283 + ], + "type": "text", + "content": "Figure 4. Parameter studies on the Shot Capacity in Positive Cache and Negative Cache." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 304, + 547, + 425 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 304, + 547, + 425 + ], + "spans": [ + { + "bbox": [ + 304, + 304, + 547, + 425 + ], + "type": "text", + "content": "cache might be beneficial, a larger negative cache leads to more high-entropy, noisier pseudo labels, as highlighted in self-training studies like [11], thereby lowering the performance. Conversely, the positive cache, which stores single and high-confidence prediction, is less affected by variations in shot capacity, thereby maintaining consistent accuracy across different shot capacities. To facilitate practical applications, the shot capacity settings are fixed and directly applied to new datasets without the need for additional parameter adjustments." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 434, + 379, + 448 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 434, + 379, + 448 + ], + "spans": [ + { + "bbox": [ + 306, + 434, + 379, + 448 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 454, + 547, + 658 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 454, + 547, + 658 + ], + "spans": [ + { + "bbox": [ + 304, + 454, + 547, + 658 + ], + "type": "text", + "content": "In this work, we have presented TDA, a dynamic adapter for efficient and effective test-time adaptation of vision-language models. The proposed method employs a key-value cache, which maintains a dynamic queue with test-sample features as keys and corresponding few-shot pseudo labels as values, allowing for gradual adaptation to test data through progressive pseudo label improvement. Moreover, TDA introduces a negative cache to mitigate the undesirable effects of noisy pseudo labels by assigning negative pseudo labels to certain classes when the model is uncertain about its predictions. The results of extensive experiments over two benchmarks demonstrate that TDA outperforms state-of-the-art test-time adaptation methods while significantly reducing testing time. This work contributes to the research and application values of test-time adaptation and presents a promising solution to the efficiency issue of test-time adaptation of vision-language models." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 669, + 404, + 682 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 669, + 404, + 682 + ], + "spans": [ + { + "bbox": [ + 306, + 669, + 404, + 682 + ], + "type": "text", + "content": "Acknowledgement" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 689, + 545, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 545, + 715 + ], + "type": "text", + "content": "This study is supported under the Mohamed bin Zayed University of Artificial Intelligence." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "14169" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 133 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 133 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 133 + ], + "type": "text", + "content": "[1] Lukas Bossard, Matthieu Guillaumin, and Luc Van Gool. Food-101 – mining discriminative components with random forests. In European Conference on Computer Vision, 2014. 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 137, + 288, + 179 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 137, + 288, + 179 + ], + "spans": [ + { + "bbox": [ + 53, + 137, + 288, + 179 + ], + "type": "text", + "content": "[2] Malik Boudiaf, Romain Mueller, Ismail Ben Ayed, and Luca Bertinetto. Parameter-free online test-time adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8344-8353, 2022. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 182, + 287, + 225 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 182, + 287, + 225 + ], + "spans": [ + { + "bbox": [ + 53, + 182, + 287, + 225 + ], + "type": "text", + "content": "[3] Dian Chen, Dequan Wang, Trevor Darrell, and Sayna Ebrahimi. Contrastive test-time adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 295-305, 2022. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 227, + 288, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 227, + 288, + 270 + ], + "spans": [ + { + "bbox": [ + 53, + 227, + 288, + 270 + ], + "type": "text", + "content": "[4] M. Cimpoi, S. Maji, I. Kokkinos, S. Mohamed, , and A. Vedaldi. Describing textures in the wild. In Proceedings of the IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2014. 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 273, + 287, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 273, + 287, + 304 + ], + "spans": [ + { + "bbox": [ + 53, + 273, + 287, + 304 + ], + "type": "text", + "content": "[5] J. Deng, W. Dong, R. Socher, L.-J. Li, K. Li, and L. Fei-Fei. ImageNet: A Large-Scale Hierarchical Image Database. In CVPR09, 2009. 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 307, + 287, + 338 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 307, + 287, + 338 + ], + "spans": [ + { + "bbox": [ + 53, + 307, + 287, + 338 + ], + "type": "text", + "content": "[6] Karan Desai and Justin Johnson. VirTex: Learning Visual Representations from Textual Annotations. In CVPR, 2021. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 341, + 287, + 406 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 341, + 287, + 406 + ], + "spans": [ + { + "bbox": [ + 53, + 341, + 287, + 406 + ], + "type": "text", + "content": "[7] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 408, + 287, + 452 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 408, + 287, + 452 + ], + "spans": [ + { + "bbox": [ + 53, + 408, + 287, + 452 + ], + "type": "text", + "content": "[8] Li Fei-Fei, Rob Fergus, and Pietro Perona. Learning generative visual models from few training examples: An incremental bayesian approach tested on 101 object categories. In CVPR Workshops, 2004. 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 454, + 287, + 508 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 454, + 287, + 508 + ], + "spans": [ + { + "bbox": [ + 53, + 454, + 287, + 508 + ], + "type": "text", + "content": "[9] Chun-Mei Feng, Kai Yu, Yong Liu, Salman Khan, and Wangmeng Zuo. Diverse data augmentation with diffusions for effective test-time prompt tuning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2704-2714, 2023. 1, 2, 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 510, + 287, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 510, + 287, + 553 + ], + "spans": [ + { + "bbox": [ + 48, + 510, + 287, + 553 + ], + "type": "text", + "content": "[10] Peng Gao, Shijie Geng, Renrui Zhang, Teli Ma, Rongyao Fang, Yongfeng Zhang, Hongsheng Li, and Yu Qiao. Clip-adapter: Better vision-language models with feature adapters. arXiv preprint arXiv:2110.04544, 2021. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 555, + 287, + 587 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 555, + 287, + 587 + ], + "spans": [ + { + "bbox": [ + 48, + 555, + 287, + 587 + ], + "type": "text", + "content": "[11] Yves Grandvalet and Yoshua Bengio. Semi-supervised learning by entropy minimization. Advances in neural information processing systems, 17, 2004. 2, 8" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 590, + 287, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 590, + 287, + 633 + ], + "spans": [ + { + "bbox": [ + 48, + 590, + 287, + 633 + ], + "type": "text", + "content": "[12] Edouard Grave, Moustapha M Cisse, and Armand Joulin. Unbounded cache model for online language modeling with open vocabulary. Advances in neural information processing systems, 30, 2017. 2, 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 635, + 287, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 635, + 287, + 666 + ], + "spans": [ + { + "bbox": [ + 48, + 635, + 287, + 666 + ], + "type": "text", + "content": "[13] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, 2016. 6" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "type": "text", + "content": "[14] Patrick Helber, Benjamin Bischke, Andreas Dengel, and Damian Borth. Eurosat: A novel dataset and deep learning benchmark for land use and land cover classification. IEEE J. Sel. Top. Appl. Earth Obs. Remote. Sens., 2019. 5" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 712 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 137 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 137 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 137 + ], + "type": "text", + "content": "[15] Dan Hendrycks, Steven Basart, Norman Mu, Saurav Kadavath, Frank Wang, Evan Dorundo, Rahul Desai, Tyler Zhu, Samyak Parajuli, Mike Guo, Dawn Song, Jacob Steinhardt, and Justin Gilmer. The many faces of robustness: A critical analysis of out-of-distribution generalization. In ICCV, 2021. 5" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 140, + 545, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 140, + 545, + 172 + ], + "spans": [ + { + "bbox": [ + 307, + 140, + 545, + 172 + ], + "type": "text", + "content": "[16] Dan Hendrycks, Kevin Zhao, Steven Basart, Jacob Steinhardt, and Dawn Song. Natural adversarial examples. In CVPR, pages 15262-15271, 2021. 5" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 175, + 545, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 175, + 545, + 228 + ], + "spans": [ + { + "bbox": [ + 307, + 175, + 545, + 228 + ], + "type": "text", + "content": "[17] Xiaowei Hu, Zhe Gan, Jianfeng Wang, Zhengyuan Yang, Zicheng Liu, Yumao Lu, and Lijuan Wang. Scaling up vision-language pre-training for image captioning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17980-17989, 2022. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 230, + 545, + 273 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 230, + 545, + 273 + ], + "spans": [ + { + "bbox": [ + 307, + 230, + 545, + 273 + ], + "type": "text", + "content": "[18] Yusuke Iwasawa and Yutaka Matsuo. Test-time classifier adjustment module for model-agnostic domain generalization. Advances in Neural Information Processing Systems, 34:2427-2440, 2021. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 275, + 545, + 339 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 275, + 545, + 339 + ], + "spans": [ + { + "bbox": [ + 308, + 275, + 545, + 339 + ], + "type": "text", + "content": "[19] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In International Conference on Machine Learning, pages 4904-4916. PMLR, 2021. 1, 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 342, + 545, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 342, + 545, + 373 + ], + "spans": [ + { + "bbox": [ + 307, + 342, + 545, + 373 + ], + "type": "text", + "content": "[20] Jeff Johnson, Matthijs Douze, and Hervé Jégou. Billionscale similarity search with gpus. IEEE Transactions on Big Data, 7(3):535-547, 2019. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 376, + 545, + 420 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 376, + 545, + 420 + ], + "spans": [ + { + "bbox": [ + 307, + 376, + 545, + 420 + ], + "type": "text", + "content": "[21] Urvashi Khandelwal, Omer Levy, Dan Jurafsky, Luke Zettlemoyer, and Mike Lewis. Generalization through memorization: Nearest neighbor language models. In International Conference on Learning Representations, 2020. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 422, + 547, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 422, + 547, + 464 + ], + "spans": [ + { + "bbox": [ + 307, + 422, + 547, + 464 + ], + "type": "text", + "content": "[22] Youngdong Kim, Junho Yim, Juseung Yun, and Junmo Kim. NInl: Negative learning for noisy labels. In Proceedings of the IEEE/CVF international conference on computer vision, pages 101-110, 2019. 2, 5" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 466, + 545, + 519 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 466, + 545, + 519 + ], + "spans": [ + { + "bbox": [ + 307, + 466, + 545, + 519 + ], + "type": "text", + "content": "[23] Youngdong Kim, Juseung Yun, Hyounguk Shon, and Junmo Kim. Joint negative and positive learning for noisy labels. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9442-9451, 2021. 2, 5" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 522, + 545, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 522, + 545, + 555 + ], + "spans": [ + { + "bbox": [ + 307, + 522, + 545, + 555 + ], + "type": "text", + "content": "[24] Jonathan Krause, Michael Stark, Jia Deng, and Li Fei-Fei. 3d object representations for fine-grained categorization. In ICCV Workshops, 2013. 5" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 556, + 545, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 556, + 545, + 600 + ], + "spans": [ + { + "bbox": [ + 307, + 556, + 545, + 600 + ], + "type": "text", + "content": "[25] Xin Li, Dongze Lian, Zhihe Lu, Jiawang Bai, Zhibo Chen, and Xinchao Wang. Graphadapter: Tuning vision-language models with dual knowledge graph. Advances in Neural Information Processing Systems, 36, 2024. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 601, + 545, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 601, + 545, + 633 + ], + "spans": [ + { + "bbox": [ + 307, + 601, + 545, + 633 + ], + "type": "text", + "content": "[26] S. Maji, J. Kannala, E. Rahtu, M. Blaschko, and A. Vedaldi. Fine-grained visual classification of aircraft. Technical report, 2013. 5" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 635, + 545, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 635, + 545, + 667 + ], + "spans": [ + { + "bbox": [ + 307, + 635, + 545, + 667 + ], + "type": "text", + "content": "[27] Stephen Merity, Caiming Xiong, James Bradbury, and Richard Socher. Pointer sentinel mixture models. arXiv preprint arXiv:1609.07843, 2016. 3" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 670, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 670, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 307, + 670, + 545, + 712 + ], + "type": "text", + "content": "[28] Maria-Elena Nilsback and Andrew Zisserman. Automated flower classification over a large number of classes. In Indian Conference on Computer Vision, Graphics and Image Processing, 2008. 5" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "14170" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 105 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 105 + ], + "type": "text", + "content": "[29] Emin Orhan. A simple cache model for image recognition. Advances in Neural Information Processing Systems, 31, 2018. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 107, + 287, + 129 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 107, + 287, + 129 + ], + "spans": [ + { + "bbox": [ + 49, + 107, + 287, + 129 + ], + "type": "text", + "content": "[30] Omkar M. Parkhi, Andrea Vedaldi, Andrew Zisserman, and C. V. Jawahar. Cats and dogs. In CVPR, 2012. 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 130, + 287, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 130, + 287, + 195 + ], + "spans": [ + { + "bbox": [ + 49, + 130, + 287, + 195 + ], + "type": "text", + "content": "[31] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In ICML, 2021. 1, 2, 3, 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 198, + 287, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 198, + 287, + 229 + ], + "spans": [ + { + "bbox": [ + 49, + 198, + 287, + 229 + ], + "type": "text", + "content": "[32] Benjamin Recht, Rebecca Roelofs, Ludwig Schmidt, and Vaishaal Shankar. Do imagenet classifiers generalize to imageNet? In ICML, 2019. 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 49, + 232, + 287, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 232, + 287, + 285 + ], + "spans": [ + { + "bbox": [ + 49, + 232, + 287, + 285 + ], + "type": "text", + "content": "[33] Mamshad Nayeem Rizve, Kevin Duarte, Yogesh S Rawat, and Mubarak Shah. In defense of pseudo-labeling: An uncertainty-aware pseudo-label selection framework for semi-supervised learning. arXiv preprint arXiv:2101.06329, 2021. 2, 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 49, + 287, + 287, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 287, + 287, + 341 + ], + "spans": [ + { + "bbox": [ + 49, + 287, + 287, + 341 + ], + "type": "text", + "content": "[34] Steffen Schneider, Evgenia Rusak, Luisa Eck, Oliver Bringmann, Wieland Brendel, and Matthias Bethge. Improving robustness against common corruptions by covariate shift adaptation. Advances in Neural Information Processing Systems, 33:11539-11551, 2020. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 49, + 343, + 287, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 343, + 287, + 398 + ], + "spans": [ + { + "bbox": [ + 49, + 343, + 287, + 398 + ], + "type": "text", + "content": "[35] Manli Shu, Weili Nie, De-An Huang, Zhiding Yu, Tom Goldstein, Anima Anandkumar, and Chaowei Xiao. Test-time prompt tuning for zero-shot generalization in vision-language models. In Advances in Neural Information Processing Systems, 2022. 1, 2, 3, 5, 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 49, + 399, + 287, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 399, + 287, + 431 + ], + "spans": [ + { + "bbox": [ + 49, + 399, + 287, + 431 + ], + "type": "text", + "content": "[36] Khurram Soomro, Amir Roshan Zamir, and Mubarak Shah. UCF101: A dataset of 101 human actions classes from videos in the wild. CoRR, abs/1212.0402, 2012. 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 49, + 434, + 287, + 476 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 434, + 287, + 476 + ], + "spans": [ + { + "bbox": [ + 49, + 434, + 287, + 476 + ], + "type": "text", + "content": "[37] Yu Sun, Xiaolong Wang, Zhuang Liu, John Miller, Alexei A. Efros, and Moritz Hardt. Test-time training with self-supervision for generalization under distribution shifts. In ICML, 2020. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 49, + 479, + 287, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 479, + 287, + 555 + ], + "spans": [ + { + "bbox": [ + 49, + 479, + 287, + 555 + ], + "type": "text", + "content": "[38] Thomas Varsavsky, Maurizio Orbes-Arteaga, Carole H Sudre, Mark S Graham, Parashkev Nachev, and M Jorge Cardoso. Test-time unsupervised domain adaptation. In Medical Image Computing and Computer Assisted Intervention—MICCAI 2020: 23rd International Conference, Lima, Peru, October 4–8, 2020, Proceedings, Part I 23, pages 428–436. Springer, 2020. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 49, + 557, + 287, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 557, + 287, + 600 + ], + "spans": [ + { + "bbox": [ + 49, + 557, + 287, + 600 + ], + "type": "text", + "content": "[39] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 49, + 601, + 287, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 601, + 287, + 634 + ], + "spans": [ + { + "bbox": [ + 49, + 601, + 287, + 634 + ], + "type": "text", + "content": "[40] Dequan Wang, Evan Shelhamer, Shaoteng Liu, Bruno A. Olshausen, and Trevor Darrell. Tent: Fully test-time adaptation by entropy minimization. In ICLR, 2021. 1, 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 49, + 636, + 287, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 636, + 287, + 668 + ], + "spans": [ + { + "bbox": [ + 49, + 636, + 287, + 668 + ], + "type": "text", + "content": "[41] Haohan Wang, Songwei Ge, Zachary Lipton, and Eric P Xing. Learning robust global representations by penalizing local predictive power. In NeurIPS, 2019. 5" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 49, + 670, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 670, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 49, + 670, + 287, + 713 + ], + "type": "text", + "content": "[42] J. Xiao, J. Hays, K. A. Ehinger, A. Oliva, and A. Torralba. Sun database: Large-scale scene recognition from abbey to zoo. In 2010 IEEE Computer Society Conference on Computer Vision and Pattern Recognition, 2010. 5" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 506 + ], + "type": "list", + "angle": 0, + "index": 23, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 138 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 138 + ], + "type": "text", + "content": "[43] Jinyu Yang, Jiali Duan, Son Tran, Yi Xu, Sampath Chanda, Liquun Chen, Belinda Zeng, Trishul Chilimbi, and Junzhou Huang. Vision-language pre-training with triple contrastive learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15671-15680, 2022. 1, 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 140, + 545, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 140, + 545, + 183 + ], + "spans": [ + { + "bbox": [ + 308, + 140, + 545, + 183 + ], + "type": "text", + "content": "[44] Tao Yu, Zhihe Lu, Xin Jin, Zhibo Chen, and Xinchao Wang. Task residual for tuning vision-language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10899-10909, 2023. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 185, + 545, + 238 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 185, + 545, + 238 + ], + "spans": [ + { + "bbox": [ + 308, + 185, + 545, + 238 + ], + "type": "text", + "content": "[45] Marvin Zhang, Henrik Marklund, Nikita Dhawan, Abhishek Gupta, Sergey Levine, and Chelsea Finn. Adaptive risk minimization: Learning to adapt to domain shift. Advances in Neural Information Processing Systems, 34:23664-23678, 2021. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 241, + 545, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 241, + 545, + 282 + ], + "spans": [ + { + "bbox": [ + 308, + 241, + 545, + 282 + ], + "type": "text", + "content": "[46] Marvin Mengxin Zhang, Sergey Levine, and Chelsea Finn. MEMO: Test time robustness via adaptation and augmentation. In Advances in Neural Information Processing Systems, 2022. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 285, + 545, + 340 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 285, + 545, + 340 + ], + "spans": [ + { + "bbox": [ + 308, + 285, + 545, + 340 + ], + "type": "text", + "content": "[47] Renrui Zhang, Wei Zhang, Rongyao Fang, Peng Gao, Kunchang Li, Jifeng Dai, Yu Qiao, and Hongsheng Li. Tip-adapter: Training-free adaption of clip for few-shot classification. In European Conference on Computer Vision, pages 493-510. Springer, 2022. 2, 3, 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 342, + 545, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 342, + 545, + 407 + ], + "spans": [ + { + "bbox": [ + 308, + 342, + 545, + 407 + ], + "type": "text", + "content": "[48] Renrui Zhang, Xiangfei Hu, Bohao Li, Siyuan Huang, Hanqiu Deng, Yu Qiao, Peng Gao, and Hongsheng Li. Prompt, generate, then cache: Cascade of foundation models makes strong few-shot learners. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15211-15222, 2023. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 409, + 545, + 462 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 409, + 545, + 462 + ], + "spans": [ + { + "bbox": [ + 308, + 409, + 545, + 462 + ], + "type": "text", + "content": "[49] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Conditional prompt learning for vision-language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16816-16825, 2022. 2, 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 464, + 545, + 506 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 464, + 545, + 506 + ], + "spans": [ + { + "bbox": [ + 308, + 464, + 545, + 506 + ], + "type": "text", + "content": "[50] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Learning to prompt for vision-language models. International Journal of Computer Vision, 130(9):2337-2348, 2022. 2, 6" + } + ] + } + ], + "index": 22 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14171" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Efficient Vision-Language Pre-training by Cluster Masking/24f9f648-410c-4f99-8615-dc6d0380efad_content_list.json b/2024/Efficient Vision-Language Pre-training by Cluster Masking/24f9f648-410c-4f99-8615-dc6d0380efad_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..e88eff33d48df0f014e342a4b23ca95325b15f01 --- /dev/null +++ b/2024/Efficient Vision-Language Pre-training by Cluster Masking/24f9f648-410c-4f99-8615-dc6d0380efad_content_list.json @@ -0,0 +1,1666 @@ +[ + { + "type": "text", + "text": "Efficient Vision-Language Pre-training by Cluster Masking", + "text_level": 1, + "bbox": [ + 186, + 130, + 784, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Zihao Wei* Zixuan Pan* Andrew Owens University of Michigan", + "bbox": [ + 290, + 180, + 676, + 217 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 251, + 313, + 266 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We propose a simple strategy for masking image patches during visual-language contrastive learning that improves the quality of the learned representations and the training speed. During each iteration of training, we randomly mask clusters of visually similar image patches, as measured by their raw pixel intensities. This provides an extra learning signal, beyond the contrastive training itself, since it forces a model to predict words for masked visual structures solely from context. It also speeds up training by reducing the amount of data used in each image. We evaluate the effectiveness of our model by pre-training on a number of benchmarks, finding that it outperforms other masking strategies, such as FLIP, on the quality of the learned representation.", + "bbox": [ + 75, + 282, + 473, + 479 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 506, + 209, + 522 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Images contain a great deal of redundant information, making it challenging to efficiently learn representations from them at scale. Recent work has addressed this problem by masking image patches during vision-language contrastive learning [15, 33, 36, 70]. One simple approach is to drop a large fraction of the patches at random, making training more efficient by reducing the computational cost and memory usage in each training iteration [36]. An alternative strategy is to mask sets of semantically related patches [15, 33, 70], such as those that belong to the same object. This forces the learned model to predict words that describes missing scene structures from context, improving the learned representation. However, this approach requires a separate mechanism to group together semantically related patches, which adds considerable complexity to the learning procedure and is computationally expensive.", + "bbox": [ + 75, + 531, + 468, + 773 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We propose a simple masking strategy for multimodal contrastive learning that avoids these shortcomings. During training, we mask random clusters of patches (Fig. 1). For this clustering, we use the patches' raw RGB values as the feature representation. Our approach takes advantage of the fact that simple measures of visual similarity can often capture coherent visual structures, such as object parts [18, 53],", + "bbox": [ + 75, + 773, + 470, + 878 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/359733b6359ec5b4e02180bb3be237448bf176d9397519aab975778441c107b8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 506, + 247, + 645, + 340 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/2a783606c85da704f7a2c1320a7e5e9c347d142920272f838a1a6f82847a212f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 508, + 340, + 645, + 433 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/5a7767e3e2c64ce15fbba58836cf176d4666fc6fd50671da9645db1ea73e47b7.jpg", + "image_caption": [ + "(b) A burgundy low-rider show-quality, hot rod Chevy truck." + ], + "image_footnote": [], + "bbox": [ + 508, + 434, + 645, + 525 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/f106417e32da587fc2930d19502b6f7d9fdc598d5006e26d85da1d3c38b8cb35.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 647, + 250, + 764, + 340 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/a03b9ce5d8b77083b3c26dd5a910242b0795d1566742e624387e88eb7a3e6ce2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 647, + 340, + 764, + 433 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/dba2b09c764298c9bb464f9f691664adbfb61dc23c0e10abc8f5f41ef30ce8eb.jpg", + "image_caption": [ + "(a) A red and gold \npainted fire \nhydrant on the street." + ], + "image_footnote": [], + "bbox": [ + 647, + 434, + 764, + 525 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/b37a70da0e179ab31adf7cc3cca7eb91e6366ac96b02c29dd354c1243f163d34.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 767, + 250, + 885, + 340 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/13423bc72e420b736b579cce42c50d583de23038efda2d74de1536338b6d0661.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 767, + 340, + 885, + 433 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/3285c2ad9cb1ca563e5525586e7b3e1a1bc9bd417b2b83ffa38dea7527290c77.jpg", + "image_caption": [ + "(c) Two surf boars on a beach near the water.", + "Figure 1. Cluster masking. We mask random clusters of visually similar image patches when training contrastive vision-language models (bottom). This masking strategy distinguishes our approach from methods that independently mask image patches for efficiency [36] (middle), while providing a similar improvement in training speed. It provides an extra learning signal, since it forces a model to predict words for missing scene structures solely from context." + ], + "image_footnote": [], + "bbox": [ + 767, + 434, + 885, + 525 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "especially when clusters are sampled randomly (Fig. 1). Our approach thus leads to more efficient training, like approaches that independently drop patches [36], while improving the learned representation via context prediction.", + "bbox": [ + 496, + 700, + 890, + 762 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We take inspiration from masked region classification, a pre-training task widely used in vision-language models [9, 56, 57]. These models extract object features, then predict object labels for the randomly masked out regions. Our masking approach provides a similar training signal, since meaningful labels are included in the image caption. For example, as shown in Figure 1(a), the model is tasked with associating the words \"fire hydrant\" with an image even with the hydrant itself is mostly masked out.", + "bbox": [ + 496, + 763, + 892, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Equal contribution. Author order was determined by a coin flip.", + "bbox": [ + 94, + 886, + 437, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "26815", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We train our model on Conceptual 12M dataset [4] and evaluate our learned representation on a number of downstream tasks. These tasks include the zero-shot classification and linear probing on ImageNet [11], text and image retrieval on MS-COCO [38], and the SUGARCREPE language composition benchmark [25]. In our experiments, our model outperforms FLIP [36] and CLIP [49] on downstream performance, while having efficiency comparable with FLIP. We also show that the performance can further be improved by using the model's learned feature embedding during clustering.", + "bbox": [ + 75, + 90, + 472, + 257 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 76, + 268, + 218, + 284 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Contrastive Vision-Language Pre-training. Vision-Language Pre-training (VLP) focuses on establishing connections between images or their components and human-interpretable language. This field initially evolved from transferring supervised learning models, which incorporated object detection modules to generate fine-grained visual labels [9, 56, 57]. Subsequently, there was a shift towards large-scale learning using noisy web data, moving away from reliance on fine-grained labels [1, 27, 34, 50, 69, 72, 73]. A significant development in this domain was CLIP [50], which applied contrastive learning techniques [6, 21] to train models to associate correct image-text pairs and dissociate incorrect ones. CLIP scaled contrastive visual-language models significantly beyond previous work, enabling strong feature learning and zero-shot performance. However, further scaling significantly increases the pre-training demands, requiring larger datasets and batch sizes.", + "bbox": [ + 75, + 294, + 472, + 565 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In response to these challenges, recent research has explored incorporating masking into images to reduce training time and allow for more samples per batch [15, 20, 36, 70]. Methods such as MaskClip [15], FLIP [36], and VIOLET [20] have implemented random masking strategies. Yet, it has been noted that random masking may not be as effective on relatively small datasets [35, 70]. To address this, ACLIP [70] introduced a method of masking tokens with low cross-attention scores with text. However, this approach necessitates two forward passes to generate the attention map and requires additional computational modules [41]. In our work, we aim to avoid these limitations, proposing an effective masking method that is based on a patch's raw RGB values.", + "bbox": [ + 75, + 566, + 472, + 777 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Masked Image Modeling. In the field of language modeling, the effectiveness of models that learn to reconstruct corrupted inputs for generating robust features has been recognized [29, 39]. This approach, known as Mask Language Modeling (MLM), has been adapted in the realm of image processing as Mask Image Modeling (MIM). MIM techniques involve reconstructing either image patches or", + "bbox": [ + 75, + 794, + 470, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "their features [2, 5, 8, 14, 22, 63-65, 68]. The pioneering work in BEIT [2] introduced the reconstruction of discrete tokens, akin to VQ-VAE [59], using block-wise masking. This method demonstrated results on par with contrastive learning and self-distillation methods [3, 7] during model fine-tuning. Later approaches include PeCo's [14] novel visual codebook learning method and BEIT V2's [47] integration of self-distillation methods, using a teacher-student backbone and feature-level KL divergence loss [58]. Further exploration in this field has led to the use of natural image signals as reconstruction targets, moving away from learned features. Examples include SimMIM [68], which reconstructs pure RGB values, MaskFeat [64], introducing reconstruction of the Histogram of Oriented Gradients (HOG) features, and MAE [22], which reconstructs pixel-normalized RGB values. Our work draws inspiration from these studies, particularly in using pixel-normalized RGB values to compute patch similarities, arguing for a more effective distribution of patch features.", + "bbox": [ + 496, + 90, + 893, + 378 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Masking Strategies in MIM. Parallel investigations have focused on masking strategies in MIMs [19, 28, 33, 37, 54, 66, 67]. Early works like BEIT and its successors used block masking, while others such as SimMIM, MaskFeat, and MAE applied random patch-wise masking. Attention-based masking strategies have also been explored, typically using attention maps from vision transformers. MST [37] masks less essential parts with low attention scores, using a reconstruction loss approach. In contrast, AttnMask [28] masks highly attentive patches and applies self-distillation loss. These methods involve simultaneous updates of attention maps and masks during training. A potential limitation of this approach is that insufficiently trained attention maps may not capture structured features effectively. SemMAE [33], starting with iBot features [75], adopts an easy-to-hard masking strategy, starting with masking parts within clusters and gradually expanding to entire clusters. Wilf et al. [66] introduce a unique entity-reinforced language model for masking objects in video frames. However, the reliance on pre-trained features or extracting attention maps can be computationally intensive. Evolved Part Masking [19] proposes using EM algorithm on attention maps to get a clustering before performing SemMAE style masking. Our approach also adopts a cluster based masking strategy in vision-language pre-training, enabling faster pre-training without requiring additional modifications to the model.", + "bbox": [ + 496, + 395, + 893, + 787 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 500, + 800, + 591, + 814 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We propose a cluster based masking strategy for contrastive vision-language pre-training, focusing on masking random clusters with visually similar semantics. Our method selects random anchor patches as cluster centers and computes pairwise patch distances to form clusters. These clus", + "bbox": [ + 496, + 825, + 893, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "26816", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/ff16a277160bab84a9c9d544c6dc9b3baafd381cc5fd6ec67322211670a6659f.jpg", + "image_caption": [ + "Anchor Patch" + ], + "image_footnote": [], + "bbox": [ + 86, + 88, + 205, + 180 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/613d61cce0bbf808ae881e59cca840a596aa12f4ccc087eeaa28f2858cd613c0.jpg", + "image_caption": [ + "Distance Heatmap" + ], + "image_footnote": [], + "bbox": [ + 210, + 88, + 328, + 180 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/9144a605330311de07b51888b727f52f0012e9134aa8ccb6289442802c452134.jpg", + "image_caption": [ + "Masked Image", + "Figure 2. Choosing clusters. The process begins by randomly selecting anchor patches from the image. We then calculate the pairwise distances among all patches. Clusters formed within a distance threshold are masked out. We show cluster obtained from a single anchor patch." + ], + "image_footnote": [], + "bbox": [ + 331, + 88, + 450, + 180 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "ters are then masked entirely. To enhance accuracy in cluster formation, we introduce an adaptive layer for refining the distance matrix. Additionally, attention masks and a hard patch cutoff are used to ensure uniform input sizes are consistent in batches for auto differentiation.", + "bbox": [ + 75, + 284, + 468, + 358 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Contrastive Vision-Language Pre-training", + "text_level": 1, + "bbox": [ + 76, + 372, + 434, + 388 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our approach builds on contrastive vision-language pretraining methods, such as CLIP [49]. We use contrastive learning to align embeddings of matching text-image pairs and separate those of non-matching pairs. This process is steered by two symmetric InfoNCE losses [43]: the vision-to-language loss $\\mathcal{L}_{v\\rightarrow l}$ and its counterpart, the language-to-vision loss $\\mathcal{L}_{l\\rightarrow v}$ . The vision-to-language loss is defined as:", + "bbox": [ + 75, + 396, + 468, + 492 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {v \\rightarrow l} = - \\log \\frac {\\exp (\\sin (I , T) / \\tau)}{\\sum_ {j = 1} ^ {N} \\exp (\\sin (I , T _ {j}) / \\tau)}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 143, + 494, + 468, + 527 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $I$ and $T$ are the embeddings for the image and text respectively, sim denotes the similarity function (we use a dot product), and $\\tau$ is a temperature parameter. Similarly, $\\mathcal{L}_{l\\to v}$ is formulated by normalizing the loss using a batch of $N$ image examples $\\{I_j\\}_{i = 1}^N$ instead of text examples $\\{T_j\\}_{i = 1}^N$ .", + "bbox": [ + 75, + 530, + 468, + 606 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Cluster Masking", + "text_level": 1, + "bbox": [ + 76, + 618, + 241, + 633 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We introduce a masking strategy that drops out random clusters. While one option would be to use an off-the-shelf clustering method, such as K-Means [40], we choose instead to use a simple and efficient method that results in a random clustering each training iteration (Figure 2). Our approach resembles a single iteration of K-Means, and works by selecting a set of exemplar patches, which each define a cluster. In our experiments, we also evaluate masking clusters obtained using K-Means as an alternative approach.", + "bbox": [ + 75, + 642, + 468, + 792 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We split an input $H\\times W$ image into patches, following [16]. We then compute the pairwise cosine similarity between every pair of normalized patches, which we use as a distance function $d(\\mathbf{x},\\mathbf{y})$ . We choose a small subset (less than $5\\%$ ) of these patches at random to act as cluster centers. For each of these selected anchor patches, we define a cluster consisting of patches that lie within a distance", + "bbox": [ + 75, + 794, + 468, + 898 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "$r$ . The cluster for an exemplar patch $\\mathbf{x}$ is represented by: $\\mathrm{S_x} = \\{\\mathbf{y} \\mid d(\\mathbf{x}, \\mathbf{y}) \\leq r\\}$ for image patches.", + "bbox": [ + 498, + 90, + 888, + 121 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "All patches within a cluster are masked out. The distance threshold $r$ is automatically searched before training according to an average masking ratio. We provide a simplified pseudocode of the masking strategy in Algorithm 1.", + "bbox": [ + 496, + 122, + 888, + 181 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Clustering Embedding Features. Another variant of patch feature is the combination of pure RGB values and patch embedding layer features from transformers [16]. When computing similarity scores, we integrate these two measures into a weighted sum, where the weight of each measure is determined by:", + "bbox": [ + 496, + 200, + 890, + 284 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nd (\\mathbf {x}, \\mathbf {y}) = \\alpha \\cdot d _ {r g b} (\\mathbf {x}, \\mathbf {y}) + (1 - \\alpha) \\cdot d _ {e m b} (\\mathbf {x}, \\mathbf {y}), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 544, + 287, + 890, + 304 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mathbf{x}$ and $\\mathbf{y}$ represent two patches, $d_{rgb}$ is the cosine similarity based on pure RGB values, and $d_{emb}$ is the cosine similarity based on the transformer's embedding features. The weight parameter $\\alpha$ linearly increases from 0 to 1 during training.", + "bbox": [ + 496, + 306, + 890, + 381 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The embedding layer is calculated before the patches enter the transformer, thus we could reuse the patch embeddings in the transformer without computing them twice. Using the embedding layer is advantageous because it incorporates positional encodings [60]. This integration potentially introduces spatial constraints, which we believe can further enhance our masking strategy.", + "bbox": [ + 496, + 382, + 890, + 488 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Handling Batched Inputs. Deep learning libraries, like PyTorch [46], typically process batched inputs of uniform size. However, in our method, the mask ratio would vary across different images, leading to fluctuations in the number of patches. To accelerate the process, we introduce a minimum mask ratio threshold $\\beta$ for each image. If the calculated mask ratio for an image doesn't meet this predefined threshold, we proceed to randomly drop patches until the desired ratio is achieved. Conversely, for images with patch counts less than the threshold, we use attention masks to avoid masked parts engaging in attention calculation [13].", + "bbox": [ + 496, + 508, + 890, + 675 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 500, + 689, + 632, + 705 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We present a comprehensive evaluation of our proposed algorithm to exhibit the performance, robustness, scalability, and efficiency of our framework.", + "bbox": [ + 496, + 715, + 888, + 760 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4.1. Implementation Details", + "text_level": 1, + "bbox": [ + 500, + 771, + 715, + 786 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Datasets and Training Details. We train our model using the Conceptual 12M (CC12M) dataset[4], containing 12 million unique image-text pairs, for pre-training our vision-language models. We use ViT-B/16 as backbone for image encoder. The text encoder is a 12-layer transformer, equipped with 8 multi-head attention units and 512-dimensional embeddings. Input images are processed at a", + "bbox": [ + 496, + 794, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "26817", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "code", + "sub_type": "code", + "code_caption": [ + "Algorithm 1 Pseudocode of cluster based masking in a PyTorch-like style." + ], + "code_body": "# img: the image for mask\n# mask_ratio: the ratio for chosing anchor patches\n# r: the threshold for computing cluster\ndef generate_mask(img, mask_ratio, r):\n # Make image into shape (N,L,C).\n # N: number of samples per batch\n # L: number of patches\n # C: size of feature dimension\n img = patchify(img)\n #Normalize each patch\n img = (img - img.mean(dim=-1)) / img.std(dim=-1)\n #Compute pairwise cosine similarity matrix\n x = img / img(norm(dim=-1))\n distance = bmm(x, x.transpose(-2, -1))\n #Generate a boolean masking matrix of shape (N, L)\n init_mask = random_batch_indicies(mask_ratio)\n init_mask = init_mask[:, :, None].float()\n #Get the cluster-based mask candidates = (distance * init_mask) ≥ r\n cluster_mask = candidates.sum(1) ≥ 1\n #Mask both the anchor patches and clusters\n mask = init_mask or cluster_mask\n return mask", + "guess_lang": "python", + "bbox": [ + 76, + 130, + 460, + 377 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "resolution of $224 \\times 224$ , and text inputs are adjusted to 77 tokens, either by truncation or padding. A class token is transformed into a 512-dimensional feature embedding via a multi-layer perceptron (MLP). For optimization, we use the AdamW optimizer with a learning rate of $5 \\times 10^{-4}$ , $\\beta_{1} = 0.9$ , and $\\beta_{2} = 0.98$ . We use a batch size of 256 per GPU, and train using 8 NVIDIA A40 GPUs.", + "bbox": [ + 75, + 404, + 468, + 508 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our method comes in three variants: K-Means, RGB and Embedding. The RGB model clusters based on raw image patches, while the embedding model integrates patch embedding features with RGB for clustering. In the K-Means variant of the model, we mask out half of the clusters randomly. The model constructs 12 clusters and runs for a maximum of 10 iterations. For both RGB and embedding models, we set an average masking ratio of $50\\%$ , following the recommendations of FLIP [36] for optimal masking ratios. In the RGB approach, we use a $50\\%$ cutoff for Ours- $\\mathsf{RGB}_{0.5}$ and $30\\%$ for Ours- $\\mathsf{RGB}_{0.3}$ , whereas the Ours-Embedding model uses a $30\\%$ cutoff. Also, the RGB model selects anchor patches at a $3\\%$ ratio, compared to a $5\\%$ ratio in the embedding model.", + "bbox": [ + 75, + 511, + 470, + 723 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Baselines. In our study, we establish baselines using three models: CLIP, FLIP, and $\\mathrm{FLIP}_{\\mathrm{Attn}}$ , each trained from scratch on the CC12M dataset. These baseline models are derived from the open-source implementation of CLIP, known as OpenClip [10, 26, 48, 52]. For both FLIP and $\\mathrm{FLIP}_{\\mathrm{Attn}}$ , we implement a patch dropout ratio of $50\\%$ . Specifically, FLIP uses a random dropout approach, whereas $\\mathrm{FLIP}_{\\mathrm{Attn}}$ adopts an attention-based masking strategy inspired by ACLIP [70]. This strategy involves processing the image through the encoder and then averaging", + "bbox": [ + 75, + 750, + 470, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "across attention heads in the final transformer block to determine attention scores. Patches that receive the highest attention in relation to the [CLS] token are retained.", + "bbox": [ + 496, + 90, + 890, + 136 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To ensure a fair comparison among these methods, we keep the number of patches consistent the same as ours within a single batch, which means for FLIP and $\\mathrm{FLIP}_{\\mathrm{Attn}}$ we apply a batch size of 256 for one GPU and for CLIP, we use 128 instead. Additionally, we apply a scaling law on learning rate across different models.", + "bbox": [ + 496, + 136, + 890, + 226 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Evaluation Details. Our models are tested across various benchmarks to ensure its robustness and effectiveness. We conduct zero-shot image-to-text and text-to-image retrieval tasks on COCO [38] and Flickr [71], assessing its performance meticulously. Further more, we evaluate the models' image representation quality by reporting both the zero-shot classification and linear probing performance on three mainstream datasets: ImageNet [11], CIFAR-10, and CIFAR-100 [30]. Zero-shot results of some other datasets like ImageNet variants [23, 24, 51, 55, 62], Caltech101 [17], Flowers [42] and Pets [61], are also reported to verify the method's robustness.", + "bbox": [ + 496, + 244, + 890, + 425 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For these tasks, our approach adheres strictly to the implementation used in the CLIP benchmark, ensuring consistency and reliability in the evaluation process. Furthermore, we assess the effectiveness of our methodology on language composition tasks using SUGARCREPE [25]. This evaluation aims to determine its adaptability and efficiency across various contexts, including object, attribute, and relation manipulations. Within the SUGARCREPE framework, models are tasked with identifying the correct caption that accurately describes an image, distinguishing it from a closely related but incorrect text hard negatives. The hard negatives is characterized by minor composition differences from the accurate caption.", + "bbox": [ + 496, + 426, + 890, + 623 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.2. Main Results", + "text_level": 1, + "bbox": [ + 500, + 632, + 637, + 646 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Visualization of Clusters. Figure 3 offers a visual depiction of our cluster based masking technique as outlined in the methodology section. For this illustration, we randomly select a number of image-text pairs from the COCO validation set and apply our masking method to the pure RGB data of the images. The visualization is showing the masking result of the two-stages. In the first stage, a subset of patches (5%) is randomly selected as anchor patches from the pool of all image patches, which are annotated with the red boxes. In the second stage, we visualize the masked clusters that are calculated based on the similarity matrix, where each cluster is represented by a distinct color.", + "bbox": [ + 496, + 655, + 890, + 837 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Zero-shot Retrieval Results. In our investigation into the model's understanding of the relationship between visual and linguistic representations, we conduct zero-shot", + "bbox": [ + 496, + 854, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "26818", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/ffb62e24ea35d217bcad8b3d1a79f6c5a8dcc52634fd02c98c98a24181891b88.jpg", + "image_caption": [ + "Figure 3. Visualization of cluster masks. Different colors represent distinct clusters formed by the similarity matrix calculated from the chosen anchor patches." + ], + "image_footnote": [], + "bbox": [ + 166, + 87, + 803, + 383 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "retrieval tests on several leading retrieval benchmarks. The results, detailed in Table 1, provide insights into the performance of our approach against others, particularly in the context of Image2Text and Text2Image's recall precision at top1(R1), top5(R5) and top10(R10) metrics.", + "bbox": [ + 75, + 424, + 470, + 500 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In the evaluation on the MS-COCO [38], Flickr8k, and Flickr30k [71] datasets, our model outperforms both the baselines in most parts. Notably, in the Image-to-Text tasks, our model performs best in most datasets, with the exception of a slight performance decrease compared to $\\mathrm{FLIP}_{\\mathrm{Attn}}$ on the MS-COCO dataset. We attribute this success to our training strategy, which prioritizes primary clusters and minimizes the influence of noise. Furthermore, we observe that methods combining RGB information with token embeddings outperform those relying solely on RGB. We hypothesize that this is because the embedding layer, which contains slightly higher-level information.", + "bbox": [ + 75, + 500, + 468, + 681 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "When comparing FLIP to CLIP, FLIP's performance is noticeably weaker, even with large batch sizes. We suspect that FLIP's sub-optimal results in our experimental settings may not fully exploit its strengths. This aligns with findings from other studies, such as Yang et al.'s research on ACLIP [70], which also noted FLIP's limitations. We observe that using attention scores for masking can improve performance compared to purely random masking. However, random masking still falls short of our cluster based masking or even the original CLIP method in some benchmarks.", + "bbox": [ + 75, + 681, + 470, + 848 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Results on Zero-shot Classification and Linear Probing. We evaluate our model on several widely recognized clas", + "bbox": [ + 76, + 869, + 468, + 902 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "sification benchmarks. The zero-shot classification results are presented in Table 2, while the linear probing results can be found in Table 3. For better evaluating the time spent on training, we normalize all method's training time by the CLIP's training time, which is considered as $1 \\times$ .", + "bbox": [ + 496, + 424, + 890, + 500 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "When comparing our model's performance to CLIP (i.e., no masking), our model demonstrates superior results on the majority of test cases, showcasing an average improvement of $+2.1\\%$ , with about $+36\\%$ speeding up. In comparison to the FLIP strategy, which has a similar training duration, our model has an improvement of $+5.5\\%$ . In comparison to the $\\mathrm{FLIP}_{\\mathrm{Attn}}$ , our model does not need the attention map for guidance, which gives a much fast training speed, while having a performance of $+2.6\\%$ on average.", + "bbox": [ + 496, + 500, + 892, + 636 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Out of 12 datasets on the zero-shot classification benchmark, our RGB and embedding model achieves the top performance on 11 of them. In particular, it obtains strong performance on the ImageNet variants: ImageNet-A [24], ImageNet-O [55], ImageNet-R [55], and ImageNet-S [62], which often contain challenging and diverse images. The RGB version of our method also significantly outperforms FLIP and surpasses the CLIP model, especially on ImageNet and its variants, which demonstrates the effectiveness of our method with even the natural guidance.", + "bbox": [ + 496, + 637, + 892, + 787 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The linear probing results further suggest the effectiveness of our method. Our models achieve $+1.8\\%$ accuracy on ImageNet, $+3.1\\%$ on CIFAR-10, and $+4.2\\%$ on CIFAR-100.", + "bbox": [ + 496, + 787, + 890, + 847 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Language Composition. A potential drawback of our method might be understanding compositions of concepts", + "bbox": [ + 500, + 869, + 890, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "26819", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/a441aa6772f24158658fd26eed403897ad6cc599cd1cb1af2bbcaff7b57c0f9f.jpg", + "table_caption": [ + "text $\\rightarrow$ image", + "image $\\rightarrow$ text" + ], + "table_footnote": [], + "table_body": "
MS-COCOFlickr8kFlickr30kMS-COCOFlickr8kFlickr30k
R1R5R10R1R5R10R1R5R10R1R5R10R1R5R10R1R5R10
CLIP [49]34.6061.9872.7255.7081.6089.9058.5083.8089.1023.4947.8059.6640.5468.9080.2043.1870.4480.40
FLIP [36]32.6259.1470.6455.0080.9088.9053.8080.8088.5022.5646.0858.0940.3268.1078.6441.5267.9077.46
FLIPAttn [70]33.6660.1871.0253.7080.0987.9955.2981.4087.6023.8948.3360.0440.5868.8878.8643.0670.1078.82
Ours-Kmeans33.6861.1471.9755.1083.3090.9055.4082.0088.9022.6046.9358.8440.1069.8679.9141.3268.5077.58
Ours-RGB0.532.8260.2071.4052.1081.2089.5054.9081.2088.0022.9747.2959.2140.8468.7279.1441.8070.2079.42
Ours-RGB0.335.8761.5072.3454.7081.4090.7057.6083.9090.3023.6547.5459.2540.9068.2279.0042.9269.9679.12
Ours-Embedding34.2661.9673.3057.0082.7090.1055.8084.2089.6023.7748.1859.7642.0069.4079.6443.3070.9280.16
", + "bbox": [ + 81, + 99, + 890, + 208 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/2f5a32d315b8107d57d969b86dc3ee02ecd3beb0bf5ff60d8ba7ae28f603fef2.jpg", + "table_caption": [ + "Table 1. Zero-shot retrieval results. We evaluate on MS-COCO [38], Flickr8k and Flickr30k datasets [71], where the Recall@1 (R1), Recall@5 (R5), and Recall@10 (R10) are reported." + ], + "table_footnote": [], + "table_body": "
MethodN.T.IN-1K [11]IN-A [24]IN-O [55]IN-R [23]IN-S [62]INv2 [51]MNIST [12]Cal101 [17]CIFAR-10 [31]CIFAR-100 [31]Flowers [42]Pets [61]Average
CLIP [49]1.00×36.18.038.447.624.930.711.773.557.725.026.053.936.1
FLIP [36]0.53×34.47.139.541.420.129.510.470.452.824.525.346.033.5
FLIPAtm [70]1.06×35.28.139.445.123.730.19.473.561.627.125.751.235.8
Ours-Kmeans0.70×35.57.238.343.022.230.19.769.961.425.325.652.135.0
Ours-RGB0.50.54×36.07.739.845.323.830.511.272.263.627.226.155.036.5
Ours-RGB0.30.64×36.68.839.945.924.931.89.472.363.126.325.457.336.8
Ours-Embedding0.64×36.38.139.647.925.430.711.073.770.732.028.455.438.2
", + "bbox": [ + 81, + 247, + 890, + 325 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/e63bd7224db14a50f70b2f5a6fcd7a8b606e2a3d77515bdf73ab3aab757608b8.jpg", + "table_caption": [ + "Table 2. Zero-shot classification result. We evaluate popular datasets using clip-benchmark [32]. The training time is normalized according to the CLIP's training time. Here N.T. represents normalized time against and IN represents ImageNet." + ], + "table_footnote": [], + "table_body": "
MethodCIFAR-10CIFAR-100IN-1K
CLIP88.067.462.3
FLIP85.965.561.3
\\(FLIP_{Attn}\\)86.466.162.0
Ours-Kmeans88.069.162.2
Ours-\\(RGB_{0.5}\\)86.766.262.5
Ours-\\(RGB_{0.3}\\)88.668.763.1
Ours-Embedding89.069.762.7
", + "bbox": [ + 109, + 371, + 434, + 486 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 3. Linear probing result. All methods are trained for 10 epochs at learning rate of 1e-3. For CIFAR-10 and CIFAR-100, we use a batch size of 64 and for ImageNet-1k the batch size is 1024.", + "bbox": [ + 76, + 489, + 470, + 545 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "in language. As we mask out clusters, there is a risk that the model may increasingly adopt bag-of-words tendencies [74], which could impede its ability to learn the relationships between objects. For example, if an image is captioned with \"dog on grass\", the grass may be masked for a large portion in our model as they are highly similar to each other. This will make learning the relation \"on\" difficult. Therefore, we apply SUGARCREPE [25] benchmarks to test the model's ability to understand language compositions. SUGARCREPE benchmarks assess this by generating negative captions through manipulations like adding, swapping, or replacing concepts in sentences, followed by text retrieval tests to evaluate the model's accuracy in selecting the correct answer. From the our test results, which shown in Table 4, our model yields comparable results in Relation tests and demonstrates a significant enhancement in Object and Attribution tests, with an average improvement of $+3.9\\%$ and $+3.0\\%$ respectively, compared to FLIP. This improvement may stem from the masking of entire objects, which simplifies the challenge of contrastive learning by reducing ambiguity. This clarity facilitates the model's learning of relationships, a crucial factor for composition", + "bbox": [ + 75, + 568, + 468, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "understanding.", + "bbox": [ + 500, + 375, + 601, + 390 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Qualitative Comparison of Masking Strategies Our method outperforms the random masking strategy by preserving more semantic content in the unmasked image patches, a comparison showcased in Figure 1. The advantage of our technique is further explored by the captioning experiment detailed in Figure 4, wherein two sets of images, each masked differently, are fed into a captioner, GPT-4 [44, 45]. The captioner is prompted to generate MS-COCO-style captions for the unobscured sections. When comparing these captions to the standard references, it becomes clear that our cluster based masking not only retains key elements but also the interrelations among them. For instance, our approach accurately enables the captioning system to identify an airplane in the first example and to describe the baseball player's action in the second, while the random masking strategy failed to achieve this clarity. These results indicate that our masking method provides a more detailed comprehension of the image.", + "bbox": [ + 496, + 411, + 893, + 684 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Ablation Study", + "text_level": 1, + "bbox": [ + 500, + 695, + 653, + 710 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Ablation on Anchor Patch Ratio. In our study, we conducted an ablation on finding the optimal proportion of patches to serve as anchor patches. The results of this ablation are summarized in Figure 5. We use zero-shot learning results on the ImageNet-1k dataset as the benchmark for assessing the quality of the representations learned by our model. Additionally, we calibrate the threshold for each experiment to ensure that the average final masking ratio was maintained at $50\\%$ .", + "bbox": [ + 496, + 718, + 893, + 853 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Our findings indicate that a smaller proportion of anchor patches tends to yield superior performance. We hypothesize that this improvement is due to the decreased ran", + "bbox": [ + 496, + 854, + 893, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "26820", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/94becf52994dde290257ee681f5a227d97c3d7b85a40f4ef0f4b6cff543930ab.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
REPLACESWAPADDAverage
ObjectAttributeRelationObjectAttributeObjectAttributeObjectAttributeRelation
CLIP85.7779.1864.5161.7858.7174.2468.3573.7168.7564.51
FLIP84.0775.8866.0060.1661.5671.6763.1571.9766.8666.00
\\( FLIP_{Attn-0.5} \\)86.6275.5063.2252.4463.0671.6566.7671.5768.3963.22
\\( FLIP_{Attn-0.3} \\)86.0775.0062.2360.5759.1674.6868.6472.8263.4762.23
Ours-Kmeans84.5076.9063.0962.2061.8671.7765.4673.6667.6063.09
Ours-RGB0.586.8673.4760.2459.3563.3673.3366.7673.1868.4260.24
Ours-RGB0.386.1375.1364.6566.6763.3674.9271.2475.9169.9164.65
", + "bbox": [ + 158, + 88, + 803, + 209 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/3dc13961a49aaf88cf054ec05906767b78d76027460b0e26d4ec608c7a544209.jpg", + "image_caption": [ + "Figure 4. Generated caption from visible patches. We process the masked images through GPT-4 [44, 45] to create captions for the unmasked segments." + ], + "image_footnote": [], + "bbox": [ + 122, + 260, + 419, + 481 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "domness in the selection of anchor patches, which in turn enhances the clustering performance by providing a more stable set of reference points for the model to learn from. However, the masking ratio cannot be too small, as the similarity threshold will become too small, which may reduce the clustering quality.", + "bbox": [ + 75, + 537, + 468, + 630 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/bc1e7ca87859ffff6a472a75dd482e1122d51f255d7a45d6e5c1b5e5f6842664.jpg", + "image_caption": [ + "Figure 5. Effect of anchor patch ratio. All the final masking ratio is tuned to be $50\\%$ ." + ], + "image_footnote": [], + "bbox": [ + 150, + 638, + 401, + 806 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Ablation on Minimum Mask Ratio. We further demonstrate the capability of our method by setting the minimum mask ratio $\\beta$ the same as the average masking ratio of FLIP,", + "bbox": [ + 75, + 854, + 472, + 902 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "as shown in Table 1, 2, and Table 5. For our method, the cutoff ratio denotes the minimum mask ratio applied and the true visible patch ratio is shown in visible ratio. For the FLIP counterpart, it maintains a consistent mask ratio across all images. The results indicate that our method not only matches the speed of FLIP but also surpasses FLIP in zero-shot ImageNet-1K classification accuracy with a $+1.6\\%$ improvement even by seeing fewer patches. The attention based masking method with less mask ratio achieves similar performance as ours but the speed is much slower.", + "bbox": [ + 496, + 263, + 890, + 414 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "These findings suggest that cluster based masking serves as an effective denoising technique for the dataset. A reason for this enhanced performance is that we could easily mask out typically irrelevant areas, such as uniformly colored backgrounds, which are less informative and oftentimes do not correspond to any word in the caption. This targeted approach enables the model to focus on more meaningful content within the images.", + "bbox": [ + 496, + 415, + 892, + 535 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Additionally, our findings reveal an improved feature learning by the model when a smaller random masking is applied. By reducing the cutoff ratio from $50\\%$ to $30\\%$ , we observed a $1\\%$ enhancement in classification accuracy. Thus, there is some trade-off between the model's performance and speed. Despite this, our model with larger masking cut off still remains significantly faster compared to attention-based masking or the original CLIP method.", + "bbox": [ + 496, + 536, + 890, + 656 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/8875e639aa6eb36016bf7038a7755f2d687a783f3203f36562f53089656ae065.jpg", + "table_caption": [ + "Table 4. Language composition test result. This table presents the performance of models on the SUGARCREPE evaluation, which involves replacing, swapping, or adding atomic concepts such as objects, attributes, and relations in a sentence to create mismatched captions." + ], + "table_footnote": [], + "table_body": "
MethodβVisible RatioNormalized TimeIN-1K
FLIP50%50%0.84×34.4
FLIP30%70%1.00×35.4
\\( \\text{FLIP}_{\\text{attn}} \\)50%50%1.73×35.2
\\( \\text{FLIP}_{\\text{attn}} \\)30%70%1.97×36.6
Ours-RGB50%43%0.84×36.0
Ours-RGB30%50%1.00×36.6
", + "bbox": [ + 519, + 669, + 867, + 762 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 5. Ablation on minimum mask ratio $\\beta$ . Comparison of various methods against different minimum masking ratios. The zero-shot ImageNet-1k classification results are used as metric. The time is normalized to ours RGB model with $\\beta = 30\\%$", + "bbox": [ + 496, + 765, + 890, + 821 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Ablation on Pixel Normalization. In our experiments, we incorporate pixel normalization (making each patch mean zero and unit standard deviation 1) into the process of computing the similarity matrix for images. This yields", + "bbox": [ + 496, + 839, + 890, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "26821", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "a performance improvement of $+1.1\\%$ , as shown in the results presented in Table 6a. The underlying rationale for this enhancement is attributed to the standardization of image patches. By using pixel normalization, we focus on the relative intensity of pixels, thereby diminishing the impact of lighting variations among different images.", + "bbox": [ + 75, + 90, + 468, + 181 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This normalization process is particularly beneficial in scenarios where the dynamic range of pixel values varies significantly across different patches. By scaling the patches to a common range, pixel-norm mitigates the risk of disproportionate influence from patches with higher intensity values. Consequently, this leads to a more balanced and equitable comparison among patches, enhancing the model's ability to discern and quantify similarities more effectively.", + "bbox": [ + 75, + 183, + 470, + 319 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/40a72b5747e7c21c6c0ed77b63fb6c10010e1bb956ce4c91530cfab6e24d8c9d.jpg", + "table_caption": [], + "table_footnote": [ + "(a) Ablation on pixel normalization (P.N.)." + ], + "table_body": "
MethodIN-1K
w/o P.N.35.5
w/ P.N.36.6
", + "bbox": [ + 101, + 339, + 228, + 393 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/f0557ff8784f130e5391a155aa1993945a474ce7e9356b91c8fe26babfe55aea.jpg", + "table_caption": [], + "table_footnote": [ + "(b) Ablation on polynomial coefficient $k$ ." + ], + "table_body": "
kIN-1K
0.536.1
136.3
235.9
", + "bbox": [ + 325, + 337, + 426, + 407 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 6. Ablation Study. Table 6a presents an ablation study on the application of pixel normalization when calculating the similarity matrix for clustering. Table 6b explores the effects of varying the polynomial coefficient $k$ , which adjusts the adaptive rate used when combining RGB and embedding features.", + "bbox": [ + 75, + 439, + 468, + 508 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effect of Features used in Clustering. In Tables 1 and 2, embedding-based methods surpass those dependent solely on RGB data, especially in image-to-text retrieval tasks. One reason for this may be the fact that the embedding model has access to the positional encoding, whereas the RGB-based model solely uses the appearance of each patch. Figure 6 qualitatively shows this advantage: while the RGB-only approach masks extra areas (such as hair or shadows in the first scenario; a laptop and phone in the second) due to color similarities, the embedding-based method masks more complete object parts.", + "bbox": [ + 75, + 541, + 468, + 708 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Ablation on Adaptive Rate. In our approach, we interpolate between using RGB features and the patch embedding layer feature using a coefficient, $\\alpha$ , which varies with each epoch. Denoting the current running epoch as $\\mathrm{E}_c$ and the number of total training epochs as $\\mathrm{E}_t$ , this coefficient is defined as $\\alpha = \\left(\\frac{\\mathrm{E}_c}{\\mathrm{E}_t}\\right)^k$ . In addition to the linear method where $k = 1$ , we explore other polynomial coefficients $k$ for this combination, as summarized in Table 6b. Our findings suggest that the linear combination is most effective, likely due to its smooth transition.", + "bbox": [ + 75, + 739, + 468, + 900 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/6f890ac5bbe7189e0459b33fbd8f0ed0ca1fa0d1000e709c2d97871524b35656.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 532, + 88, + 638, + 169 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/a7f85c1ab0faae754ef71dce2346d9271fd6ba8d51d634e490103229533caa0b.jpg", + "image_caption": [ + "Anchor Patch" + ], + "image_footnote": [], + "bbox": [ + 532, + 170, + 637, + 250 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/76aa7a07f50c691bb505c6ecbae409d764261161137570a5e351185e6b817c8b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 640, + 88, + 743, + 169 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/bf955542c4782b2e87ed9108e711e2cfd9c2d78f1e1a76652191879ef9246003.jpg", + "image_caption": [ + "RGB-based" + ], + "image_footnote": [], + "bbox": [ + 640, + 170, + 743, + 250 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/8595e0bf6a2d1eb5fa2dc2b499614faddf8d0e339798f9a342bff11879838a78.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 748, + 88, + 849, + 169 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/61ccb72de9683aeb0e65a285cc2d1732ef07d7682d11d6b316ced20c19c3e435.jpg", + "image_caption": [ + "Embedding-based", + "Figure 6. Representation used in clustering. We illustrate the distinctions between using pure RGB and ViT embeddings to compute the similarity matrix used in masking patches." + ], + "image_footnote": [], + "bbox": [ + 748, + 170, + 849, + 250 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4. Limitations", + "text_level": 1, + "bbox": [ + 500, + 319, + 624, + 333 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Our methodology uses a uniform threshold for all images, a strategy that, while effective, may not be the most optimal. Future research could explore the implementation of individualized thresholds for each image, potentially leading to a more intelligent and adaptive masking process.", + "bbox": [ + 496, + 342, + 890, + 417 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "All of our approaches use the popular backbone architecture ViT-B/16 [16] and are trained solely on the CC12M dataset [4]. Expanding the scope of the experiments could offer additional insights.", + "bbox": [ + 496, + 419, + 890, + 479 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 492, + 617, + 507 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In our study, we introduce a novel cluster based masking strategy designed for vision-language pret-raining. Using either pure RGB values or shallow features from the patch embedding layer, our method effectively clusters image patches, maintaining essential visual semantics. We then randomly mask out these clusters, enabling efficient training. Our approach demonstrates success across various downstream evaluation tasks, including both pure image-based tasks such as image classification and multimodal tasks like image-text retrieval and language composition tests. We believe our work marks a considerable progression in this domain and anticipate that it will stimulate further research into optimizing masking strategies for similar applications.", + "bbox": [ + 496, + 517, + 890, + 729 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Author Contributions All authors contributed to designing projects, launching experiments, and writing the paper. Zixuan Pan focused on algorithm optimization; Zihao Wei focused on code design; Andrew Owens supervised this project, offered feedback, and assisted in writing the paper.", + "bbox": [ + 496, + 746, + 890, + 823 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements This research is supported by a Sony Research Award. We are grateful to Jeongsoo Park, Chao Feng, Yiming Dou, Daniel Geng, Ziyang Chen, and Ayush Shrivastava for their valuable suggestions and discussion.", + "bbox": [ + 496, + 839, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "26822", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. Advances in Neural Information Processing Systems, 35:23716-23736, 2022. 2", + "[2] Hangbo Bao, Li Dong, Songhao Piao, and Furu Wei. Beit: Bert pre-training of image transformers. arXiv preprint arXiv:2106.08254, 2021. 2", + "[3] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9650-9660, 2021. 2", + "[4] Soravit Changpinyo, Piyush Sharma, Nan Ding, and Radu Soricut. Conceptual 12m: Pushing web-scale image-text pretraining to recognize long-tail visual concepts. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3558-3568, 2021. 2, 3, 8", + "[5] Mark Chen, Alec Radford, Rewon Child, Jeffrey Wu, Heewoo Jun, David Luan, and Ilya Sutskever. Generative pretraining from pixels. In International Conference on Machine Learning, pages 1691-1703. PMLR, 2020. 2", + "[6] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International conference on machine learning, pages 1597-1607. PMLR, 2020. 2", + "[7] X. Chen, S. Xie, and K. He. An empirical study of training self-supervised vision transformers. In 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 9620-9629, 2021. 2", + "[8] Xiaokang Chen, Mingyu Ding, Xiaodi Wang, Ying Xin, Shentong Mo, Yunhao Wang, Shumin Han, Ping Luo, Gang Zeng, and Jingdong Wang. Context autoencoder for self-supervised representation learning. arXiv preprint arXiv:2202.03026, 2022. 2", + "[9] Yen-Chun Chen, Linjie Li, Licheng Yu, Ahmed El Kholy, Faisal Ahmed, Zhe Gan, Yu Cheng, and Jingjing Liu. Uniter: Universal image-text representation learning. In European conference on computer vision, pages 104-120. Springer, 2020. 1, 2", + "[10] Mehdi Cherti, Romain Beaumont, Ross Wightman, Mitchell Wortsman, Gabriel Ilharco, Cade Gordon, Christoph Schuhmann, Ludwig Schmidt, and Jenia Jitsev. Reproducible scaling laws for contrastive language-image learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2818-2829, 2023. 4", + "[11] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE Conference on Computer Vision and Pattern Recognition, pages 248–255, 2009. 2, 4, 6", + "[12] Li Deng. The mnist database of handwritten digit images for machine learning research. IEEE Signal Processing Magazine, 29(6):141-142, 2012. 6" + ], + "bbox": [ + 78, + 114, + 470, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[13] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. 2018. 3", + "[14] Xiaoyi Dong, Jianmin Bao, Ting Zhang, Dongdong Chen, Weiming Zhang, Lu Yuan, Dong Chen, Fang Wen, Nenghai Yu, and Baining Guo. Peco: Perceptual codebook for bert pre-training of vision transformers. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 552-560, 2023. 2", + "[15] Xiaoyi Dong, Jianmin Bao, Yinglin Zheng, Ting Zhang, Dongdong Chen, Hao Yang, Ming Zeng, Weiming Zhang, Lu Yuan, Dong Chen, Fang Wen, and Nenghai Yu. Maskclip: Masked self-distillation advances contrastive language-image pretraining, 2023. 1, 2", + "[16] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. ICLR, 2021. 3, 8", + "[17] Li Fei-Fei, R. Fergus, and P. Perona. One-shot learning of object categories. IEEE Transactions on Pattern Analysis and Machine Intelligence, 28(4):594–611, 2006. 4, 6", + "[18] Pedro F Felzenszwalb and Daniel P Huttenlocher. Efficient graph-based image segmentation. International journal of computer vision, 59:167-181, 2004. 1", + "[19] Zhanzhou Feng and Shiliang Zhang. Evolved part masking for self-supervised learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10386-10395, 2023. 2", + "[20] Tsu-Jui Fu, Linjie Li, Zhe Gan, Kevin Lin, William Yang Wang, Lijuan Wang, and Zicheng Liu. Violet: End-to-end video-language transformers with masked visual-token modeling. arXiv preprint arXiv:2111.12681, 2021. 2", + "[21] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9729-9738, 2020. 2", + "[22] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollar, and Ross Girshick. Masked autoencoders are scalable vision learners. arXiv preprint arXiv:2111.06377, 2021. 2", + "[23] Dan Hendrycks, Steven Basart, Norman Mu, Saurav Kadavath, Frank Wang, Evan Dorundo, Rahul Desai, Tyler Zhu, Samyak Parajuli, Mike Guo, Dawn Song, Jacob Steinhardt, and Justin Gilmer. The many faces of robustness: A critical analysis of out-of-distribution generalization. ICCV, 2021. 4, 6", + "[24] Dan Hendrycks, Kevin Zhao, Steven Basart, Jacob Steinhardt, and Dawn Song. Natural adversarial examples. CVPR, 2021. 4, 5, 6", + "[25] Cheng-Yu Hsieh, Jieyu Zhang, Zixian Ma, Aniruddha Kembhavi, and Ranjay Krishna. Sugarcrepe: Fixing hackable benchmarks for vision-language compositionality. arXiv preprint arXiv:2306.14610, 2023. 2, 4, 6", + "[26] Gabriel Ilharco, Mitchell Wortsman, Ross Wightman, Cade Gordon, Nicholas Carlini, Rohan Taori, Achal Dave," + ], + "bbox": [ + 501, + 92, + 893, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "26823", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Vaishaal Shankar, Hongseok Namkoong, John Miller, Hannaneh Hajishirzi, Ali Farhadi, and Ludwig Schmidt. Openclip, 2021. If you use this software, please cite it as below. 4", + "[27] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In International conference on machine learning, pages 4904-4916. PMLR, 2021. 2", + "[28] Ioannis Kakogeorgiou, Spyros Gidaris, Bill Psomas, Yannis Avrithis, Andrei Bursuc, Konstantinos Karantzalos, and Nikos Komodakis. What to hide from your students: Attention-guided masked image modeling. In Computer Vision - ECCV 2022, pages 300-318. Springer Nature Switzerland, 2022. 2", + "[29] Jacob Devlin Ming-Wei Chang Kenton and Lee Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of naacL-HLT, page 2, 2019. 2", + "[30] Alex Krizhevsky. Learning multiple layers of features from tiny images. Technical report, 2009. 4", + "[31] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009. 6", + "[32] LAION-AI. Clip benchmark. https://github.com/LAION-AI/CLIP_benchmark, 2023.6", + "[33] Gang Li, Heliang Zheng, Daqing Liu, Chaoyue Wang, Bing Su, and Changwen Zheng. Semmae: Semantic-guided masking for learning masked autoencoders. Advances in Neural Information Processing Systems, 35:14290-14302, 2022. 1, 2", + "[34] Junnan Li, Ramprasaath Selvaraju, Akhilesh Gotmare, Shafiq Joty, Caiming Xiong, and Steven Chu Hong Hoi. Align before fuse: Vision and language representation learning with momentum distillation. Advances in neural information processing systems, 34:9694-9705, 2021. 2", + "[35] Xianhang Li, Zeyu Wang, and Cihang Xie. Clipa-v2: Scaling clip training with 81.12", + "[36] Yanghao Li, Haoqi Fan, Ronghang Hu, Christoph Feichtenhofer, and Kaiming He. Scaling language-image pre-training via masking, 2023. 1, 2, 4, 6", + "[37] Zhaowen Li, Zhiyang Chen, Fan Yang, Wei Li, Yousong Zhu, Chaoyang Zhao, Rui Deng, Liwei Wu, Rui Zhao, Ming Tang, et al. Mst: Masked self-supervised transformer for visual representation. Advances in Neural Information Processing Systems, 34:13165-13176, 2021. 2", + "[38] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dollar, and Larry Zitnick. Microsoft coco: Common objects in context. In ECCV. European Conference on Computer Vision, 2014. 2, 4, 5, 6", + "[39] Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. Roberta: A robustly optimized bert pretraining approach. arXiv preprint arXiv:1907.11692, 2019. 2", + "[40] Stuart Lloyd. Least squares quantization in lcm. IEEE transactions on information theory, 28(2):129-137, 1982. 3" + ], + "bbox": [ + 78, + 90, + 470, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[41] Norman Mu, Alexander Kirillov, David Wagner, and Saining Xie. Slip: Self-supervision meets language-image pretraining, 2021. 2", + "[42] Maria-Elena Nilsback and Andrew Zisserman. Automated flower classification over a large number of classes. In Proceedings of the 2008 Sixth Indian Conference on Computer Vision, Graphics & Image Processing, page 722-729, USA, 2008. IEEE Computer Society. 4, 6", + "[43] Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748, 2018. 3", + "[44] OpenAI. Gpt-4 technical report, 2023. 6, 7", + "[45] Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, John Schulman, Jacob Hilton, Fraser Kelton, Luke Miller, Maddie Simens, Amanda Askell, Peter Welinder, Paul F Christiano, Jan Leike, and Ryan Lowe. Training language models to follow instructions with human feedback. In Advances in Neural Information Processing Systems, pages 27730-27744. Curran Associates, Inc., 2022. 6, 7", + "[46] Adam Paszke, Sam Gross, Soumith Chintala, Gregory Chanan, Edward Yang, Zachary DeVito, Zeming Lin, Alban Desmaison, Luca Antiga, and Adam Lerer. Automatic differentiation in pytorch. 2017. 3", + "[47] Zhiliang Peng, Li Dong, Hangbo Bao, Qixiang Ye, and Furu Wei. Beit v2: Masked image modeling with vector-quantized visual tokenizers. arXiv preprint arXiv:2208.06366, 2022. 2", + "[48] Alec Radford, Jong Wook Kim, Chris Hallacy, A. Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In ICML, 2021. 4", + "[49] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 2, 3, 6", + "[50] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 2", + "[51] Benjamin Recht, Rebecca Roelofs, Ludwig Schmidt, and Vaishaal Shankar. DoImagenet classifiers generalize to imageNet?, 2019. 4, 6", + "[52] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade W Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, Patrick Schramowski, Srivatsa R Kundurthy, Katherine Crowson, Ludwig Schmidt, Robert Kaczmarczyk, and Jenia Jitsev. LAION-5b: An open large-scale dataset for training next generation image-text models. In Thirty-sixth Conference on Neural Information Processing Systems Datasets and Benchmarks Track, 2022. 4" + ], + "bbox": [ + 501, + 92, + 893, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "26824", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[53] Jianbo Shi and Jitendra Malik. Normalized cuts and image segmentation. IEEE Transactions on pattern analysis and machine intelligence, 22(8):888-905, 2000. 1", + "[54] Yuge Shi, N Siddharth, Philip Torr, and Adam R Kosiorek. Adversarial masking for self-supervised learning. In International Conference on Machine Learning, pages 20026-20040. PMLR, 2022. 2", + "[55] Anugya Srivastava, Shriya Jain, and Mugdha Thigle. Out of distribution detection onImagenet-o, 2022. 4, 5, 6", + "[56] Weijie Su, Xizhou Zhu, Yue Cao, Bin Li, Lewei Lu, Furu Wei, and Jifeng Dai. Vl-bert: Pre-training of generic visual-linguistic representations. arXiv preprint arXiv:1908.08530, 2019. 1, 2", + "[57] Hao Tan and Mohit Bansal. Lxmert: Learning cross-modality encoder representations from transformers. arXiv preprint arXiv:1908.07490, 2019. 1, 2", + "[58] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Hervé Jégou. Training data-efficient image transformers & distillation through attention. In International conference on machine learning, pages 10347-10357. PMLR, 2021. 2", + "[59] Aaron Van Den Oord, Oriol Vinyals, et al. Neural discrete representation learning. Advances in neural information processing systems, 30, 2017. 2", + "[60] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. 2017. 3", + "[61] Andrea Vedaldi. Cats and dogs. In Proceedings of the 2012 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), page 3498-3505, USA, 2012. IEEE Computer Society. 4, 6", + "[62] Haohan Wang, Songwei Ge, Zachary Lipton, and Eric P Xing. Learning robust global representations by penalizing local predictive power. In Advances in Neural Information Processing Systems, pages 10506-10518, 2019. 4, 5, 6", + "[63] Yiqing Wang, Zihan Li, Jieru Mei, Zihao Wei, Li Liu, Chen Wang, Shengtian Sang, Alan Yuille, Cihang Xie, and Yuyin Zhou. Swinnmm: Masked multi-view with swim transformers for 3d medical image segmentation. In MICCAI, 2023. 2", + "[64] Chen Wei, Haoqi Fan, Saining Xie, Chao-Yuan Wu, Alan Yuille, and Christoph Feichtenhofer. Masked feature prediction for self-supervised visual pre-training. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14668-14678, 2022. 2", + "[65] Zihao Wei, Chen Wei, Jieru Mei, Zeyu Wang, Xianhang Li, Huiyu Wang, Alan Yuille, Yuyin Zhou, and Cihang Xie. MAE are secretly efficient learners, 2023. 2", + "[66] Alex Wilf, Syeda Nahida Akter, Leena Mathur, Paul Pu Liang, Sheryl Mathew, Mengrou Shou, Eric Nyberg, and Louis-Philippe Morency. Difference-masking: Choosing what to mask in continued pretraining. arXiv preprint arXiv:2305.14577, 2023. 2", + "[67] Jiahao Xie, Wei Li, Xiaohang Zhan, Ziwei Liu, Yew Soon Ong, and Chen Change Loy. Masked frequency modeling for self-supervised visual pre-training. arXiv preprint arXiv:2206.07706, 2022. 2" + ], + "bbox": [ + 78, + 90, + 470, + 898 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[68] Zhenda Xie, Zheng Zhang, Yue Cao, Yutong Lin, Jianmin Bao, Zhuliang Yao, Qi Dai, and Han Hu. Simmim: A simple framework for masked image modeling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9653-9663, 2022. 2", + "[69] Fengyu Yang, Chao Feng, Ziyang Chen, Hyoungseob Park, Daniel Wang, Yiming Dou, Ziyao Zeng, Xien Chen, Rit Gangopadhyay, Andrew Owens, et al. Binding touch to everything: Learning unified multimodal tactile representations. arXiv preprint arXiv:2401.18084, 2024. 2", + "[70] Yifan Yang, Weiquan Huang, Yixuan Wei, Houwen Peng, Xinyang Jiang, Huiqiang Jiang, Fangyun Wei, Yin Wang, Han Hu, Lili Qiu, and Yuqing Yang. Attentive mask clip, 2022. 1, 2, 4, 5, 6", + "[71] Peter Young, Alice Lai, Micah Hodosh, and Julia Hockenmaier. From image descriptions to visual denotations: New similarity metrics for semantic inference over event descriptions. Transactions of the Association for Computational Linguistics, 2:67-78, 2014. 4, 5, 6", + "[72] Jiahui Yu, Zirui Wang, Vijay Vasudevan, Legg Yeung, Mojtaba Seyedhosseini, and Yonghui Wu. Coca: Contrastive captioners are image-text foundation models. arXiv preprint arXiv:2205.01917, 2022. 2", + "[73] Lu Yuan, Dongdong Chen, Yi-Ling Chen, Noel Codella, Xiyang Dai, Jianfeng Gao, Houdong Hu, Xuedong Huang, Boxin Li, Chunyuan Li, et al. Florence: A new foundation model for computer vision. arXiv preprint arXiv:2111.11432, 2021. 2", + "[74] Mert Yuksekgonul, Federico Bianchi, Pratyusha Kalluri, Dan Jurafsky, and James Zou. When and why vision-language models behave like bags-of-words, and what to do about it? In The Eleventh International Conference on Learning Representations, 2022. 6", + "[75] Jinghao Zhou, Chen Wei, Huiyu Wang, Wei Shen, Cihang Xie, Alan Yuille, and Tao Kong. ibot: Image bert pre-training with online tokenizer. arXiv preprint arXiv:2111.07832, 2021. 2" + ], + "bbox": [ + 501, + 92, + 893, + 612 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "26825", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2024/Efficient Vision-Language Pre-training by Cluster Masking/24f9f648-410c-4f99-8615-dc6d0380efad_model.json b/2024/Efficient Vision-Language Pre-training by Cluster Masking/24f9f648-410c-4f99-8615-dc6d0380efad_model.json new file mode 100644 index 0000000000000000000000000000000000000000..03dadd43fce9453410432938820e2c9152c14aef --- /dev/null +++ b/2024/Efficient Vision-Language Pre-training by Cluster Masking/24f9f648-410c-4f99-8615-dc6d0380efad_model.json @@ -0,0 +1,2565 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.187, + 0.131, + 0.785, + 0.154 + ], + "angle": 0, + "content": "Efficient Vision-Language Pre-training by Cluster Masking" + }, + { + "type": "text", + "bbox": [ + 0.292, + 0.181, + 0.677, + 0.218 + ], + "angle": 0, + "content": "Zihao Wei* Zixuan Pan* Andrew Owens University of Michigan" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.252, + 0.314, + 0.267 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.283, + 0.474, + 0.481 + ], + "angle": 0, + "content": "We propose a simple strategy for masking image patches during visual-language contrastive learning that improves the quality of the learned representations and the training speed. During each iteration of training, we randomly mask clusters of visually similar image patches, as measured by their raw pixel intensities. This provides an extra learning signal, beyond the contrastive training itself, since it forces a model to predict words for masked visual structures solely from context. It also speeds up training by reducing the amount of data used in each image. We evaluate the effectiveness of our model by pre-training on a number of benchmarks, finding that it outperforms other masking strategies, such as FLIP, on the quality of the learned representation." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.507, + 0.21, + 0.523 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.532, + 0.47, + 0.775 + ], + "angle": 0, + "content": "Images contain a great deal of redundant information, making it challenging to efficiently learn representations from them at scale. Recent work has addressed this problem by masking image patches during vision-language contrastive learning [15, 33, 36, 70]. One simple approach is to drop a large fraction of the patches at random, making training more efficient by reducing the computational cost and memory usage in each training iteration [36]. An alternative strategy is to mask sets of semantically related patches [15, 33, 70], such as those that belong to the same object. This forces the learned model to predict words that describes missing scene structures from context, improving the learned representation. However, this approach requires a separate mechanism to group together semantically related patches, which adds considerable complexity to the learning procedure and is computationally expensive." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.774, + 0.471, + 0.88 + ], + "angle": 0, + "content": "We propose a simple masking strategy for multimodal contrastive learning that avoids these shortcomings. During training, we mask random clusters of patches (Fig. 1). For this clustering, we use the patches' raw RGB values as the feature representation. Our approach takes advantage of the fact that simple measures of visual similarity can often capture coherent visual structures, such as object parts [18, 53]," + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.248, + 0.646, + 0.342 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.509, + 0.342, + 0.646, + 0.434 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.509, + 0.435, + 0.646, + 0.526 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.509, + 0.532, + 0.645, + 0.569 + ], + "angle": 0, + "content": "(a) A red and gold \npainted fire \nhydrant on the street." + }, + { + "type": "image", + "bbox": [ + 0.648, + 0.25, + 0.766, + 0.342 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.648, + 0.342, + 0.766, + 0.434 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.648, + 0.435, + 0.766, + 0.526 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.649, + 0.529, + 0.765, + 0.57 + ], + "angle": 0, + "content": "(b) A burgundy low-rider show-quality, hot rod Chevy truck." + }, + { + "type": "image", + "bbox": [ + 0.768, + 0.25, + 0.886, + 0.342 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.768, + 0.342, + 0.886, + 0.434 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.768, + 0.435, + 0.886, + 0.526 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.784, + 0.529, + 0.882, + 0.569 + ], + "angle": 0, + "content": "(c) Two surf boars on a beach near the water." + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.576, + 0.892, + 0.686 + ], + "angle": 0, + "content": "Figure 1. Cluster masking. We mask random clusters of visually similar image patches when training contrastive vision-language models (bottom). This masking strategy distinguishes our approach from methods that independently mask image patches for efficiency [36] (middle), while providing a similar improvement in training speed. It provides an extra learning signal, since it forces a model to predict words for missing scene structures solely from context." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.702, + 0.892, + 0.763 + ], + "angle": 0, + "content": "especially when clusters are sampled randomly (Fig. 1). Our approach thus leads to more efficient training, like approaches that independently drop patches [36], while improving the learned representation via context prediction." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.893, + 0.901 + ], + "angle": 0, + "content": "We take inspiration from masked region classification, a pre-training task widely used in vision-language models [9, 56, 57]. These models extract object features, then predict object labels for the randomly masked out regions. Our masking approach provides a similar training signal, since meaningful labels are included in the image caption. For example, as shown in Figure 1(a), the model is tasked with associating the words \"fire hydrant\" with an image even with the hydrant itself is mostly masked out." + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.887, + 0.438, + 0.901 + ], + "angle": 0, + "content": "*Equal contribution. Author order was determined by a coin flip." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "26815" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.258 + ], + "angle": 0, + "content": "We train our model on Conceptual 12M dataset [4] and evaluate our learned representation on a number of downstream tasks. These tasks include the zero-shot classification and linear probing on ImageNet [11], text and image retrieval on MS-COCO [38], and the SUGARCREPE language composition benchmark [25]. In our experiments, our model outperforms FLIP [36] and CLIP [49] on downstream performance, while having efficiency comparable with FLIP. We also show that the performance can further be improved by using the model's learned feature embedding during clustering." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.269, + 0.22, + 0.285 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.295, + 0.473, + 0.566 + ], + "angle": 0, + "content": "Contrastive Vision-Language Pre-training. Vision-Language Pre-training (VLP) focuses on establishing connections between images or their components and human-interpretable language. This field initially evolved from transferring supervised learning models, which incorporated object detection modules to generate fine-grained visual labels [9, 56, 57]. Subsequently, there was a shift towards large-scale learning using noisy web data, moving away from reliance on fine-grained labels [1, 27, 34, 50, 69, 72, 73]. A significant development in this domain was CLIP [50], which applied contrastive learning techniques [6, 21] to train models to associate correct image-text pairs and dissociate incorrect ones. CLIP scaled contrastive visual-language models significantly beyond previous work, enabling strong feature learning and zero-shot performance. However, further scaling significantly increases the pre-training demands, requiring larger datasets and batch sizes." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.567, + 0.473, + 0.779 + ], + "angle": 0, + "content": "In response to these challenges, recent research has explored incorporating masking into images to reduce training time and allow for more samples per batch [15, 20, 36, 70]. Methods such as MaskClip [15], FLIP [36], and VIOLET [20] have implemented random masking strategies. Yet, it has been noted that random masking may not be as effective on relatively small datasets [35, 70]. To address this, ACLIP [70] introduced a method of masking tokens with low cross-attention scores with text. However, this approach necessitates two forward passes to generate the attention map and requires additional computational modules [41]. In our work, we aim to avoid these limitations, proposing an effective masking method that is based on a patch's raw RGB values." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.795, + 0.472, + 0.903 + ], + "angle": 0, + "content": "Masked Image Modeling. In the field of language modeling, the effectiveness of models that learn to reconstruct corrupted inputs for generating robust features has been recognized [29, 39]. This approach, known as Mask Language Modeling (MLM), has been adapted in the realm of image processing as Mask Image Modeling (MIM). MIM techniques involve reconstructing either image patches or" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.38 + ], + "angle": 0, + "content": "their features [2, 5, 8, 14, 22, 63-65, 68]. The pioneering work in BEIT [2] introduced the reconstruction of discrete tokens, akin to VQ-VAE [59], using block-wise masking. This method demonstrated results on par with contrastive learning and self-distillation methods [3, 7] during model fine-tuning. Later approaches include PeCo's [14] novel visual codebook learning method and BEIT V2's [47] integration of self-distillation methods, using a teacher-student backbone and feature-level KL divergence loss [58]. Further exploration in this field has led to the use of natural image signals as reconstruction targets, moving away from learned features. Examples include SimMIM [68], which reconstructs pure RGB values, MaskFeat [64], introducing reconstruction of the Histogram of Oriented Gradients (HOG) features, and MAE [22], which reconstructs pixel-normalized RGB values. Our work draws inspiration from these studies, particularly in using pixel-normalized RGB values to compute patch similarities, arguing for a more effective distribution of patch features." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.396, + 0.895, + 0.789 + ], + "angle": 0, + "content": "Masking Strategies in MIM. Parallel investigations have focused on masking strategies in MIMs [19, 28, 33, 37, 54, 66, 67]. Early works like BEIT and its successors used block masking, while others such as SimMIM, MaskFeat, and MAE applied random patch-wise masking. Attention-based masking strategies have also been explored, typically using attention maps from vision transformers. MST [37] masks less essential parts with low attention scores, using a reconstruction loss approach. In contrast, AttnMask [28] masks highly attentive patches and applies self-distillation loss. These methods involve simultaneous updates of attention maps and masks during training. A potential limitation of this approach is that insufficiently trained attention maps may not capture structured features effectively. SemMAE [33], starting with iBot features [75], adopts an easy-to-hard masking strategy, starting with masking parts within clusters and gradually expanding to entire clusters. Wilf et al. [66] introduce a unique entity-reinforced language model for masking objects in video frames. However, the reliance on pre-trained features or extracting attention maps can be computationally intensive. Evolved Part Masking [19] proposes using EM algorithm on attention maps to get a clustering before performing SemMAE style masking. Our approach also adopts a cluster based masking strategy in vision-language pre-training, enabling faster pre-training without requiring additional modifications to the model." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.801, + 0.593, + 0.815 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.826, + 0.895, + 0.903 + ], + "angle": 0, + "content": "We propose a cluster based masking strategy for contrastive vision-language pre-training, focusing on masking random clusters with visually similar semantics. Our method selects random anchor patches as cluster centers and computes pairwise patch distances to form clusters. These clus" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "26816" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.088, + 0.089, + 0.207, + 0.181 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.106, + 0.185, + 0.188, + 0.197 + ], + "angle": 0, + "content": "Anchor Patch" + }, + { + "type": "image", + "bbox": [ + 0.211, + 0.089, + 0.329, + 0.181 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.185, + 0.325, + 0.198 + ], + "angle": 0, + "content": "Distance Heatmap" + }, + { + "type": "image", + "bbox": [ + 0.333, + 0.089, + 0.452, + 0.181 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.348, + 0.185, + 0.437, + 0.198 + ], + "angle": 0, + "content": "Masked Image" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.204, + 0.47, + 0.273 + ], + "angle": 0, + "content": "Figure 2. Choosing clusters. The process begins by randomly selecting anchor patches from the image. We then calculate the pairwise distances among all patches. Clusters formed within a distance threshold are masked out. We show cluster obtained from a single anchor patch." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.285, + 0.469, + 0.359 + ], + "angle": 0, + "content": "ters are then masked entirely. To enhance accuracy in cluster formation, we introduce an adaptive layer for refining the distance matrix. Additionally, attention masks and a hard patch cutoff are used to ensure uniform input sizes are consistent in batches for auto differentiation." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.373, + 0.436, + 0.389 + ], + "angle": 0, + "content": "3.1. Contrastive Vision-Language Pre-training" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.397, + 0.469, + 0.493 + ], + "angle": 0, + "content": "Our approach builds on contrastive vision-language pretraining methods, such as CLIP [49]. We use contrastive learning to align embeddings of matching text-image pairs and separate those of non-matching pairs. This process is steered by two symmetric InfoNCE losses [43]: the vision-to-language loss \\(\\mathcal{L}_{v\\rightarrow l}\\) and its counterpart, the language-to-vision loss \\(\\mathcal{L}_{l\\rightarrow v}\\). The vision-to-language loss is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.145, + 0.496, + 0.469, + 0.529 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {v \\rightarrow l} = - \\log \\frac {\\exp (\\sin (I , T) / \\tau)}{\\sum_ {j = 1} ^ {N} \\exp (\\sin (I , T _ {j}) / \\tau)}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.531, + 0.469, + 0.607 + ], + "angle": 0, + "content": "where \\(I\\) and \\(T\\) are the embeddings for the image and text respectively, sim denotes the similarity function (we use a dot product), and \\(\\tau\\) is a temperature parameter. Similarly, \\(\\mathcal{L}_{l\\to v}\\) is formulated by normalizing the loss using a batch of \\(N\\) image examples \\(\\{I_j\\}_{i = 1}^N\\) instead of text examples \\(\\{T_j\\}_{i = 1}^N\\)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.619, + 0.242, + 0.635 + ], + "angle": 0, + "content": "3.2. Cluster Masking" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.643, + 0.469, + 0.793 + ], + "angle": 0, + "content": "We introduce a masking strategy that drops out random clusters. While one option would be to use an off-the-shelf clustering method, such as K-Means [40], we choose instead to use a simple and efficient method that results in a random clustering each training iteration (Figure 2). Our approach resembles a single iteration of K-Means, and works by selecting a set of exemplar patches, which each define a cluster. In our experiments, we also evaluate masking clusters obtained using K-Means as an alternative approach." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.795, + 0.469, + 0.9 + ], + "angle": 0, + "content": "We split an input \\(H\\times W\\) image into patches, following [16]. We then compute the pairwise cosine similarity between every pair of normalized patches, which we use as a distance function \\(d(\\mathbf{x},\\mathbf{y})\\). We choose a small subset (less than \\(5\\%\\)) of these patches at random to act as cluster centers. For each of these selected anchor patches, we define a cluster consisting of patches that lie within a distance" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.89, + 0.122 + ], + "angle": 0, + "content": "\\(r\\). The cluster for an exemplar patch \\(\\mathbf{x}\\) is represented by: \\(\\mathrm{S_x} = \\{\\mathbf{y} \\mid d(\\mathbf{x}, \\mathbf{y}) \\leq r\\}\\) for image patches." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.123, + 0.89, + 0.183 + ], + "angle": 0, + "content": "All patches within a cluster are masked out. The distance threshold \\( r \\) is automatically searched before training according to an average masking ratio. We provide a simplified pseudocode of the masking strategy in Algorithm 1." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.202, + 0.892, + 0.285 + ], + "angle": 0, + "content": "Clustering Embedding Features. Another variant of patch feature is the combination of pure RGB values and patch embedding layer features from transformers [16]. When computing similarity scores, we integrate these two measures into a weighted sum, where the weight of each measure is determined by:" + }, + { + "type": "equation", + "bbox": [ + 0.545, + 0.289, + 0.891, + 0.305 + ], + "angle": 0, + "content": "\\[\nd (\\mathbf {x}, \\mathbf {y}) = \\alpha \\cdot d _ {r g b} (\\mathbf {x}, \\mathbf {y}) + (1 - \\alpha) \\cdot d _ {e m b} (\\mathbf {x}, \\mathbf {y}), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.307, + 0.892, + 0.382 + ], + "angle": 0, + "content": "where \\(\\mathbf{x}\\) and \\(\\mathbf{y}\\) represent two patches, \\(d_{rgb}\\) is the cosine similarity based on pure RGB values, and \\(d_{emb}\\) is the cosine similarity based on the transformer's embedding features. The weight parameter \\(\\alpha\\) linearly increases from 0 to 1 during training." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.383, + 0.892, + 0.489 + ], + "angle": 0, + "content": "The embedding layer is calculated before the patches enter the transformer, thus we could reuse the patch embeddings in the transformer without computing them twice. Using the embedding layer is advantageous because it incorporates positional encodings [60]. This integration potentially introduces spatial constraints, which we believe can further enhance our masking strategy." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.51, + 0.892, + 0.676 + ], + "angle": 0, + "content": "Handling Batched Inputs. Deep learning libraries, like PyTorch [46], typically process batched inputs of uniform size. However, in our method, the mask ratio would vary across different images, leading to fluctuations in the number of patches. To accelerate the process, we introduce a minimum mask ratio threshold \\(\\beta\\) for each image. If the calculated mask ratio for an image doesn't meet this predefined threshold, we proceed to randomly drop patches until the desired ratio is achieved. Conversely, for images with patch counts less than the threshold, we use attention masks to avoid masked parts engaging in attention calculation [13]." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.69, + 0.633, + 0.707 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.716, + 0.89, + 0.761 + ], + "angle": 0, + "content": "We present a comprehensive evaluation of our proposed algorithm to exhibit the performance, robustness, scalability, and efficiency of our framework." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.772, + 0.717, + 0.787 + ], + "angle": 0, + "content": "4.1. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.795, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Datasets and Training Details. We train our model using the Conceptual 12M (CC12M) dataset[4], containing 12 million unique image-text pairs, for pre-training our vision-language models. We use ViT-B/16 as backbone for image encoder. The text encoder is a 12-layer transformer, equipped with 8 multi-head attention units and 512-dimensional embeddings. Input images are processed at a" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "26817" + } + ], + [ + { + "type": "code_caption", + "bbox": [ + 0.078, + 0.091, + 0.47, + 0.122 + ], + "angle": 0, + "content": "Algorithm 1 Pseudocode of cluster based masking in a PyTorch-like style." + }, + { + "type": "code", + "bbox": [ + 0.078, + 0.131, + 0.462, + 0.378 + ], + "angle": 0, + "content": "# img: the image for mask\n# mask_ratio: the ratio for chosing anchor patches\n# r: the threshold for computing cluster\ndef generate_mask(img, mask_ratio, r):\n # Make image into shape (N,L,C).\n # N: number of samples per batch\n # L: number of patches\n # C: size of feature dimension\n img = patchify(img)\n #Normalize each patch\n img = (img - img.mean(dim=-1)) / img.std(dim=-1)\n #Compute pairwise cosine similarity matrix\n x = img / img(norm(dim=-1))\n distance = bmm(x, x.transpose(-2, -1))\n #Generate a boolean masking matrix of shape (N, L)\n init_mask = random_batch_indicies(mask_ratio)\n init_mask = init_mask[:, :, None].float()\n #Get the cluster-based mask candidates = (distance * init_mask) ≥ r\n cluster_mask = candidates.sum(1) ≥ 1\n #Mask both the anchor patches and clusters\n mask = init_mask or cluster_mask\n return mask" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.405, + 0.47, + 0.51 + ], + "angle": 0, + "content": "resolution of \\(224 \\times 224\\), and text inputs are adjusted to 77 tokens, either by truncation or padding. A class token is transformed into a 512-dimensional feature embedding via a multi-layer perceptron (MLP). For optimization, we use the AdamW optimizer with a learning rate of \\(5 \\times 10^{-4}\\), \\(\\beta_{1} = 0.9\\), and \\(\\beta_{2} = 0.98\\). We use a batch size of 256 per GPU, and train using 8 NVIDIA A40 GPUs." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.512, + 0.471, + 0.724 + ], + "angle": 0, + "content": "Our method comes in three variants: K-Means, RGB and Embedding. The RGB model clusters based on raw image patches, while the embedding model integrates patch embedding features with RGB for clustering. In the K-Means variant of the model, we mask out half of the clusters randomly. The model constructs 12 clusters and runs for a maximum of 10 iterations. For both RGB and embedding models, we set an average masking ratio of \\(50\\%\\), following the recommendations of FLIP [36] for optimal masking ratios. In the RGB approach, we use a \\(50\\%\\) cutoff for Ours-\\(\\mathsf{RGB}_{0.5}\\) and \\(30\\%\\) for Ours-\\(\\mathsf{RGB}_{0.3}\\), whereas the Ours-Embedding model uses a \\(30\\%\\) cutoff. Also, the RGB model selects anchor patches at a \\(3\\%\\) ratio, compared to a \\(5\\%\\) ratio in the embedding model." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Baselines. In our study, we establish baselines using three models: CLIP, FLIP, and \\(\\mathrm{FLIP}_{\\mathrm{Attn}}\\), each trained from scratch on the CC12M dataset. These baseline models are derived from the open-source implementation of CLIP, known as OpenClip [10, 26, 48, 52]. For both FLIP and \\(\\mathrm{FLIP}_{\\mathrm{Attn}}\\), we implement a patch dropout ratio of \\(50\\%\\). Specifically, FLIP uses a random dropout approach, whereas \\(\\mathrm{FLIP}_{\\mathrm{Attn}}\\) adopts an attention-based masking strategy inspired by ACLIP [70]. This strategy involves processing the image through the encoder and then averaging" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.137 + ], + "angle": 0, + "content": "across attention heads in the final transformer block to determine attention scores. Patches that receive the highest attention in relation to the [CLS] token are retained." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.137, + 0.892, + 0.227 + ], + "angle": 0, + "content": "To ensure a fair comparison among these methods, we keep the number of patches consistent the same as ours within a single batch, which means for FLIP and \\(\\mathrm{FLIP}_{\\mathrm{Attn}}\\) we apply a batch size of 256 for one GPU and for CLIP, we use 128 instead. Additionally, we apply a scaling law on learning rate across different models." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.246, + 0.892, + 0.426 + ], + "angle": 0, + "content": "Evaluation Details. Our models are tested across various benchmarks to ensure its robustness and effectiveness. We conduct zero-shot image-to-text and text-to-image retrieval tasks on COCO [38] and Flickr [71], assessing its performance meticulously. Further more, we evaluate the models' image representation quality by reporting both the zero-shot classification and linear probing performance on three mainstream datasets: ImageNet [11], CIFAR-10, and CIFAR-100 [30]. Zero-shot results of some other datasets like ImageNet variants [23, 24, 51, 55, 62], Caltech101 [17], Flowers [42] and Pets [61], are also reported to verify the method's robustness." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.428, + 0.892, + 0.624 + ], + "angle": 0, + "content": "For these tasks, our approach adheres strictly to the implementation used in the CLIP benchmark, ensuring consistency and reliability in the evaluation process. Furthermore, we assess the effectiveness of our methodology on language composition tasks using SUGARCREPE [25]. This evaluation aims to determine its adaptability and efficiency across various contexts, including object, attribute, and relation manipulations. Within the SUGARCREPE framework, models are tasked with identifying the correct caption that accurately describes an image, distinguishing it from a closely related but incorrect text hard negatives. The hard negatives is characterized by minor composition differences from the accurate caption." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.633, + 0.638, + 0.647 + ], + "angle": 0, + "content": "4.2. Main Results" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.656, + 0.892, + 0.838 + ], + "angle": 0, + "content": "Visualization of Clusters. Figure 3 offers a visual depiction of our cluster based masking technique as outlined in the methodology section. For this illustration, we randomly select a number of image-text pairs from the COCO validation set and apply our masking method to the pure RGB data of the images. The visualization is showing the masking result of the two-stages. In the first stage, a subset of patches (5%) is randomly selected as anchor patches from the pool of all image patches, which are annotated with the red boxes. In the second stage, we visualize the masked clusters that are calculated based on the similarity matrix, where each cluster is represented by a distinct color." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Zero-shot Retrieval Results. In our investigation into the model's understanding of the relationship between visual and linguistic representations, we conduct zero-shot" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "26818" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.167, + 0.088, + 0.805, + 0.384 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.388, + 0.893, + 0.418 + ], + "angle": 0, + "content": "Figure 3. Visualization of cluster masks. Different colors represent distinct clusters formed by the similarity matrix calculated from the chosen anchor patches." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.425, + 0.471, + 0.501 + ], + "angle": 0, + "content": "retrieval tests on several leading retrieval benchmarks. The results, detailed in Table 1, provide insights into the performance of our approach against others, particularly in the context of Image2Text and Text2Image's recall precision at top1(R1), top5(R5) and top10(R10) metrics." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.501, + 0.47, + 0.682 + ], + "angle": 0, + "content": "In the evaluation on the MS-COCO [38], Flickr8k, and Flickr30k [71] datasets, our model outperforms both the baselines in most parts. Notably, in the Image-to-Text tasks, our model performs best in most datasets, with the exception of a slight performance decrease compared to \\(\\mathrm{FLIP}_{\\mathrm{Attn}}\\) on the MS-COCO dataset. We attribute this success to our training strategy, which prioritizes primary clusters and minimizes the influence of noise. Furthermore, we observe that methods combining RGB information with token embeddings outperform those relying solely on RGB. We hypothesize that this is because the embedding layer, which contains slightly higher-level information." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.683, + 0.471, + 0.849 + ], + "angle": 0, + "content": "When comparing FLIP to CLIP, FLIP's performance is noticeably weaker, even with large batch sizes. We suspect that FLIP's sub-optimal results in our experimental settings may not fully exploit its strengths. This aligns with findings from other studies, such as Yang et al.'s research on ACLIP [70], which also noted FLIP's limitations. We observe that using attention scores for masking can improve performance compared to purely random masking. However, random masking still falls short of our cluster based masking or even the original CLIP method in some benchmarks." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.47, + 0.903 + ], + "angle": 0, + "content": "Results on Zero-shot Classification and Linear Probing. We evaluate our model on several widely recognized clas" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.425, + 0.892, + 0.501 + ], + "angle": 0, + "content": "sification benchmarks. The zero-shot classification results are presented in Table 2, while the linear probing results can be found in Table 3. For better evaluating the time spent on training, we normalize all method's training time by the CLIP's training time, which is considered as \\(1 \\times\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.501, + 0.893, + 0.637 + ], + "angle": 0, + "content": "When comparing our model's performance to CLIP (i.e., no masking), our model demonstrates superior results on the majority of test cases, showcasing an average improvement of \\(+2.1\\%\\), with about \\(+36\\%\\) speeding up. In comparison to the FLIP strategy, which has a similar training duration, our model has an improvement of \\(+5.5\\%\\). In comparison to the \\(\\mathrm{FLIP}_{\\mathrm{Attn}}\\), our model does not need the attention map for guidance, which gives a much fast training speed, while having a performance of \\(+2.6\\%\\) on average." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.638, + 0.893, + 0.789 + ], + "angle": 0, + "content": "Out of 12 datasets on the zero-shot classification benchmark, our RGB and embedding model achieves the top performance on 11 of them. In particular, it obtains strong performance on the ImageNet variants: ImageNet-A [24], ImageNet-O [55], ImageNet-R [55], and ImageNet-S [62], which often contain challenging and diverse images. The RGB version of our method also significantly outperforms FLIP and surpasses the CLIP model, especially on ImageNet and its variants, which demonstrates the effectiveness of our method with even the natural guidance." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.789, + 0.892, + 0.848 + ], + "angle": 0, + "content": "The linear probing results further suggest the effectiveness of our method. Our models achieve \\(+1.8\\%\\) accuracy on ImageNet, \\(+3.1\\%\\) on CIFAR-10, and \\(+4.2\\%\\) on CIFAR-100." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.892, + 0.903 + ], + "angle": 0, + "content": "Language Composition. A potential drawback of our method might be understanding compositions of concepts" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "26819" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.322, + 0.091, + 0.387, + 0.1 + ], + "angle": 0, + "content": "text \\(\\rightarrow\\) image" + }, + { + "type": "table_caption", + "bbox": [ + 0.678, + 0.091, + 0.743, + 0.1 + ], + "angle": 0, + "content": "image \\(\\rightarrow\\) text" + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.1, + 0.891, + 0.209 + ], + "angle": 0, + "content": "
MS-COCOFlickr8kFlickr30kMS-COCOFlickr8kFlickr30k
R1R5R10R1R5R10R1R5R10R1R5R10R1R5R10R1R5R10
CLIP [49]34.6061.9872.7255.7081.6089.9058.5083.8089.1023.4947.8059.6640.5468.9080.2043.1870.4480.40
FLIP [36]32.6259.1470.6455.0080.9088.9053.8080.8088.5022.5646.0858.0940.3268.1078.6441.5267.9077.46
FLIPAttn [70]33.6660.1871.0253.7080.0987.9955.2981.4087.6023.8948.3360.0440.5868.8878.8643.0670.1078.82
Ours-Kmeans33.6861.1471.9755.1083.3090.9055.4082.0088.9022.6046.9358.8440.1069.8679.9141.3268.5077.58
Ours-RGB0.532.8260.2071.4052.1081.2089.5054.9081.2088.0022.9747.2959.2140.8468.7279.1441.8070.2079.42
Ours-RGB0.335.8761.5072.3454.7081.4090.7057.6083.9090.3023.6547.5459.2540.9068.2279.0042.9269.9679.12
Ours-Embedding34.2661.9673.3057.0082.7090.1055.8084.2089.6023.7748.1859.7642.0069.4079.6443.3070.9280.16
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.213, + 0.893, + 0.242 + ], + "angle": 0, + "content": "Table 1. Zero-shot retrieval results. We evaluate on MS-COCO [38], Flickr8k and Flickr30k datasets [71], where the Recall@1 (R1), Recall@5 (R5), and Recall@10 (R10) are reported." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.248, + 0.891, + 0.327 + ], + "angle": 0, + "content": "
MethodN.T.IN-1K [11]IN-A [24]IN-O [55]IN-R [23]IN-S [62]INv2 [51]MNIST [12]Cal101 [17]CIFAR-10 [31]CIFAR-100 [31]Flowers [42]Pets [61]Average
CLIP [49]1.00×36.18.038.447.624.930.711.773.557.725.026.053.936.1
FLIP [36]0.53×34.47.139.541.420.129.510.470.452.824.525.346.033.5
FLIPAtm [70]1.06×35.28.139.445.123.730.19.473.561.627.125.751.235.8
Ours-Kmeans0.70×35.57.238.343.022.230.19.769.961.425.325.652.135.0
Ours-RGB0.50.54×36.07.739.845.323.830.511.272.263.627.226.155.036.5
Ours-RGB0.30.64×36.68.839.945.924.931.89.472.363.126.325.457.336.8
Ours-Embedding0.64×36.38.139.647.925.430.711.073.770.732.028.455.438.2
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.331, + 0.893, + 0.36 + ], + "angle": 0, + "content": "Table 2. Zero-shot classification result. We evaluate popular datasets using clip-benchmark [32]. The training time is normalized according to the CLIP's training time. Here N.T. represents normalized time against and IN represents ImageNet." + }, + { + "type": "table", + "bbox": [ + 0.11, + 0.372, + 0.435, + 0.487 + ], + "angle": 0, + "content": "
MethodCIFAR-10CIFAR-100IN-1K
CLIP88.067.462.3
FLIP85.965.561.3
\\(FLIP_{Attn}\\)86.466.162.0
Ours-Kmeans88.069.162.2
Ours-\\(RGB_{0.5}\\)86.766.262.5
Ours-\\(RGB_{0.3}\\)88.668.763.1
Ours-Embedding89.069.762.7
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.491, + 0.472, + 0.546 + ], + "angle": 0, + "content": "Table 3. Linear probing result. All methods are trained for 10 epochs at learning rate of 1e-3. For CIFAR-10 and CIFAR-100, we use a batch size of 64 and for ImageNet-1k the batch size is 1024." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.569, + 0.47, + 0.901 + ], + "angle": 0, + "content": "in language. As we mask out clusters, there is a risk that the model may increasingly adopt bag-of-words tendencies [74], which could impede its ability to learn the relationships between objects. For example, if an image is captioned with \"dog on grass\", the grass may be masked for a large portion in our model as they are highly similar to each other. This will make learning the relation \"on\" difficult. Therefore, we apply SUGARCREPE [25] benchmarks to test the model's ability to understand language compositions. SUGARCREPE benchmarks assess this by generating negative captions through manipulations like adding, swapping, or replacing concepts in sentences, followed by text retrieval tests to evaluate the model's accuracy in selecting the correct answer. From the our test results, which shown in Table 4, our model yields comparable results in Relation tests and demonstrates a significant enhancement in Object and Attribution tests, with an average improvement of \\(+3.9\\%\\) and \\(+3.0\\%\\) respectively, compared to FLIP. This improvement may stem from the masking of entire objects, which simplifies the challenge of contrastive learning by reducing ambiguity. This clarity facilitates the model's learning of relationships, a crucial factor for composition" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.376, + 0.602, + 0.391 + ], + "angle": 0, + "content": "understanding." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.412, + 0.895, + 0.685 + ], + "angle": 0, + "content": "Qualitative Comparison of Masking Strategies Our method outperforms the random masking strategy by preserving more semantic content in the unmasked image patches, a comparison showcased in Figure 1. The advantage of our technique is further explored by the captioning experiment detailed in Figure 4, wherein two sets of images, each masked differently, are fed into a captioner, GPT-4 [44, 45]. The captioner is prompted to generate MS-COCO-style captions for the unobscured sections. When comparing these captions to the standard references, it becomes clear that our cluster based masking not only retains key elements but also the interrelations among them. For instance, our approach accurately enables the captioning system to identify an airplane in the first example and to describe the baseball player's action in the second, while the random masking strategy failed to achieve this clarity. These results indicate that our masking method provides a more detailed comprehension of the image." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.696, + 0.655, + 0.712 + ], + "angle": 0, + "content": "4.3. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.719, + 0.894, + 0.854 + ], + "angle": 0, + "content": "Ablation on Anchor Patch Ratio. In our study, we conducted an ablation on finding the optimal proportion of patches to serve as anchor patches. The results of this ablation are summarized in Figure 5. We use zero-shot learning results on the ImageNet-1k dataset as the benchmark for assessing the quality of the representations learned by our model. Additionally, we calibrate the threshold for each experiment to ensure that the average final masking ratio was maintained at \\(50\\%\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.894, + 0.901 + ], + "angle": 0, + "content": "Our findings indicate that a smaller proportion of anchor patches tends to yield superior performance. We hypothesize that this improvement is due to the decreased ran" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "26820" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.16, + 0.089, + 0.805, + 0.21 + ], + "angle": 0, + "content": "
REPLACESWAPADDAverage
ObjectAttributeRelationObjectAttributeObjectAttributeObjectAttributeRelation
CLIP85.7779.1864.5161.7858.7174.2468.3573.7168.7564.51
FLIP84.0775.8866.0060.1661.5671.6763.1571.9766.8666.00
\\( FLIP_{Attn-0.5} \\)86.6275.5063.2252.4463.0671.6566.7671.5768.3963.22
\\( FLIP_{Attn-0.3} \\)86.0775.0062.2360.5759.1674.6868.6472.8263.4762.23
Ours-Kmeans84.5076.9063.0962.2061.8671.7765.4673.6667.6063.09
Ours-RGB0.586.8673.4760.2459.3563.3673.3366.7673.1868.4260.24
Ours-RGB0.386.1375.1364.6566.6763.3674.9271.2475.9169.9164.65
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.213, + 0.895, + 0.255 + ], + "angle": 0, + "content": "Table 4. Language composition test result. This table presents the performance of models on the SUGARCREPE evaluation, which involves replacing, swapping, or adding atomic concepts such as objects, attributes, and relations in a sentence to create mismatched captions." + }, + { + "type": "image", + "bbox": [ + 0.123, + 0.261, + 0.42, + 0.482 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.489, + 0.472, + 0.532 + ], + "angle": 0, + "content": "Figure 4. Generated caption from visible patches. We process the masked images through GPT-4 [44, 45] to create captions for the unmasked segments." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.539, + 0.47, + 0.631 + ], + "angle": 0, + "content": "domness in the selection of anchor patches, which in turn enhances the clustering performance by providing a more stable set of reference points for the model to learn from. However, the masking ratio cannot be too small, as the similarity threshold will become too small, which may reduce the clustering quality." + }, + { + "type": "image", + "bbox": [ + 0.151, + 0.64, + 0.402, + 0.807 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.811, + 0.47, + 0.839 + ], + "angle": 0, + "content": "Figure 5. Effect of anchor patch ratio. All the final masking ratio is tuned to be \\(50\\%\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.473, + 0.903 + ], + "angle": 0, + "content": "Ablation on Minimum Mask Ratio. We further demonstrate the capability of our method by setting the minimum mask ratio \\(\\beta\\) the same as the average masking ratio of FLIP," + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.264, + 0.892, + 0.415 + ], + "angle": 0, + "content": "as shown in Table 1, 2, and Table 5. For our method, the cutoff ratio denotes the minimum mask ratio applied and the true visible patch ratio is shown in visible ratio. For the FLIP counterpart, it maintains a consistent mask ratio across all images. The results indicate that our method not only matches the speed of FLIP but also surpasses FLIP in zero-shot ImageNet-1K classification accuracy with a \\(+1.6\\%\\) improvement even by seeing fewer patches. The attention based masking method with less mask ratio achieves similar performance as ours but the speed is much slower." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.416, + 0.893, + 0.536 + ], + "angle": 0, + "content": "These findings suggest that cluster based masking serves as an effective denoising technique for the dataset. A reason for this enhanced performance is that we could easily mask out typically irrelevant areas, such as uniformly colored backgrounds, which are less informative and oftentimes do not correspond to any word in the caption. This targeted approach enables the model to focus on more meaningful content within the images." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.537, + 0.892, + 0.657 + ], + "angle": 0, + "content": "Additionally, our findings reveal an improved feature learning by the model when a smaller random masking is applied. By reducing the cutoff ratio from \\(50\\%\\) to \\(30\\%\\), we observed a \\(1\\%\\) enhancement in classification accuracy. Thus, there is some trade-off between the model's performance and speed. Despite this, our model with larger masking cut off still remains significantly faster compared to attention-based masking or the original CLIP method." + }, + { + "type": "table", + "bbox": [ + 0.521, + 0.67, + 0.868, + 0.763 + ], + "angle": 0, + "content": "
MethodβVisible RatioNormalized TimeIN-1K
FLIP50%50%0.84×34.4
FLIP30%70%1.00×35.4
\\( \\text{FLIP}_{\\text{attn}} \\)50%50%1.73×35.2
\\( \\text{FLIP}_{\\text{attn}} \\)30%70%1.97×36.6
Ours-RGB50%43%0.84×36.0
Ours-RGB30%50%1.00×36.6
" + }, + { + "type": "table_caption", + "bbox": [ + 0.498, + 0.766, + 0.892, + 0.822 + ], + "angle": 0, + "content": "Table 5. Ablation on minimum mask ratio \\( \\beta \\). Comparison of various methods against different minimum masking ratios. The zero-shot ImageNet-1k classification results are used as metric. The time is normalized to ours RGB model with \\( \\beta = 30\\% \\)" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.892, + 0.903 + ], + "angle": 0, + "content": "Ablation on Pixel Normalization. In our experiments, we incorporate pixel normalization (making each patch mean zero and unit standard deviation 1) into the process of computing the similarity matrix for images. This yields" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "26821" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.182 + ], + "angle": 0, + "content": "a performance improvement of \\(+1.1\\%\\), as shown in the results presented in Table 6a. The underlying rationale for this enhancement is attributed to the standardization of image patches. By using pixel normalization, we focus on the relative intensity of pixels, thereby diminishing the impact of lighting variations among different images." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.184, + 0.471, + 0.32 + ], + "angle": 0, + "content": "This normalization process is particularly beneficial in scenarios where the dynamic range of pixel values varies significantly across different patches. By scaling the patches to a common range, pixel-norm mitigates the risk of disproportionate influence from patches with higher intensity values. Consequently, this leads to a more balanced and equitable comparison among patches, enhancing the model's ability to discern and quantify similarities more effectively." + }, + { + "type": "table", + "bbox": [ + 0.102, + 0.34, + 0.23, + 0.395 + ], + "angle": 0, + "content": "
MethodIN-1K
w/o P.N.35.5
w/ P.N.36.6
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.078, + 0.41, + 0.255, + 0.434 + ], + "angle": 0, + "content": "(a) Ablation on pixel normalization (P.N.)." + }, + { + "type": "table", + "bbox": [ + 0.326, + 0.338, + 0.428, + 0.408 + ], + "angle": 0, + "content": "
kIN-1K
0.536.1
136.3
235.9
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.287, + 0.411, + 0.465, + 0.433 + ], + "angle": 0, + "content": "(b) Ablation on polynomial coefficient \\( k \\)." + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.44, + 0.469, + 0.509 + ], + "angle": 0, + "content": "Table 6. Ablation Study. Table 6a presents an ablation study on the application of pixel normalization when calculating the similarity matrix for clustering. Table 6b explores the effects of varying the polynomial coefficient \\( k \\), which adjusts the adaptive rate used when combining RGB and embedding features." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.542, + 0.47, + 0.709 + ], + "angle": 0, + "content": "Effect of Features used in Clustering. In Tables 1 and 2, embedding-based methods surpass those dependent solely on RGB data, especially in image-to-text retrieval tasks. One reason for this may be the fact that the embedding model has access to the positional encoding, whereas the RGB-based model solely uses the appearance of each patch. Figure 6 qualitatively shows this advantage: while the RGB-only approach masks extra areas (such as hair or shadows in the first scenario; a laptop and phone in the second) due to color similarities, the embedding-based method masks more complete object parts." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.74, + 0.47, + 0.901 + ], + "angle": 0, + "content": "Ablation on Adaptive Rate. In our approach, we interpolate between using RGB features and the patch embedding layer feature using a coefficient, \\(\\alpha\\), which varies with each epoch. Denoting the current running epoch as \\(\\mathrm{E}_c\\) and the number of total training epochs as \\(\\mathrm{E}_t\\), this coefficient is defined as \\(\\alpha = \\left(\\frac{\\mathrm{E}_c}{\\mathrm{E}_t}\\right)^k\\). In addition to the linear method where \\(k = 1\\), we explore other polynomial coefficients \\(k\\) for this combination, as summarized in Table 6b. Our findings suggest that the linear combination is most effective, likely due to its smooth transition." + }, + { + "type": "image", + "bbox": [ + 0.534, + 0.089, + 0.639, + 0.17 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.534, + 0.171, + 0.638, + 0.251 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.545, + 0.254, + 0.625, + 0.265 + ], + "angle": 0, + "content": "Anchor Patch" + }, + { + "type": "image", + "bbox": [ + 0.641, + 0.089, + 0.744, + 0.17 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.641, + 0.171, + 0.744, + 0.251 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.657, + 0.254, + 0.726, + 0.265 + ], + "angle": 0, + "content": "RGB-based" + }, + { + "type": "image", + "bbox": [ + 0.749, + 0.089, + 0.851, + 0.17 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.749, + 0.171, + 0.851, + 0.251 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.747, + 0.254, + 0.851, + 0.266 + ], + "angle": 0, + "content": "Embedding-based" + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.272, + 0.892, + 0.314 + ], + "angle": 0, + "content": "Figure 6. Representation used in clustering. We illustrate the distinctions between using pure RGB and ViT embeddings to compute the similarity matrix used in masking patches." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.32, + 0.625, + 0.334 + ], + "angle": 0, + "content": "4.4. Limitations" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.343, + 0.892, + 0.419 + ], + "angle": 0, + "content": "Our methodology uses a uniform threshold for all images, a strategy that, while effective, may not be the most optimal. Future research could explore the implementation of individualized thresholds for each image, potentially leading to a more intelligent and adaptive masking process." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.42, + 0.892, + 0.48 + ], + "angle": 0, + "content": "All of our approaches use the popular backbone architecture ViT-B/16 [16] and are trained solely on the CC12M dataset [4]. Expanding the scope of the experiments could offer additional insights." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.493, + 0.619, + 0.508 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.518, + 0.892, + 0.73 + ], + "angle": 0, + "content": "In our study, we introduce a novel cluster based masking strategy designed for vision-language pret-raining. Using either pure RGB values or shallow features from the patch embedding layer, our method effectively clusters image patches, maintaining essential visual semantics. We then randomly mask out these clusters, enabling efficient training. Our approach demonstrates success across various downstream evaluation tasks, including both pure image-based tasks such as image classification and multimodal tasks like image-text retrieval and language composition tests. We believe our work marks a considerable progression in this domain and anticipate that it will stimulate further research into optimizing masking strategies for similar applications." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.747, + 0.892, + 0.824 + ], + "angle": 0, + "content": "Author Contributions All authors contributed to designing projects, launching experiments, and writing the paper. Zixuan Pan focused on algorithm optimization; Zihao Wei focused on code design; Andrew Owens supervised this project, offered feedback, and assisted in writing the paper." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Acknowledgements This research is supported by a Sony Research Award. We are grateful to Jeongsoo Park, Chao Feng, Yiming Dou, Daniel Geng, Ziyang Chen, and Ayush Shrivastava for their valuable suggestions and discussion." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "26822" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.198 + ], + "angle": 0, + "content": "[1] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. Advances in Neural Information Processing Systems, 35:23716-23736, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.201, + 0.472, + 0.242 + ], + "angle": 0, + "content": "[2] Hangbo Bao, Li Dong, Songhao Piao, and Furu Wei. Beit: Bert pre-training of image transformers. arXiv preprint arXiv:2106.08254, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.245, + 0.47, + 0.314 + ], + "angle": 0, + "content": "[3] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9650-9660, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.316, + 0.47, + 0.398 + ], + "angle": 0, + "content": "[4] Soravit Changpinyo, Piyush Sharma, Nan Ding, and Radu Soricut. Conceptual 12m: Pushing web-scale image-text pretraining to recognize long-tail visual concepts. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3558-3568, 2021. 2, 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.401, + 0.47, + 0.457 + ], + "angle": 0, + "content": "[5] Mark Chen, Alec Radford, Rewon Child, Jeffrey Wu, Heewoo Jun, David Luan, and Ilya Sutskever. Generative pretraining from pixels. In International Conference on Machine Learning, pages 1691-1703. PMLR, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.459, + 0.47, + 0.514 + ], + "angle": 0, + "content": "[6] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International conference on machine learning, pages 1597-1607. PMLR, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.517, + 0.47, + 0.571 + ], + "angle": 0, + "content": "[7] X. Chen, S. Xie, and K. He. An empirical study of training self-supervised vision transformers. In 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 9620-9629, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.574, + 0.47, + 0.643 + ], + "angle": 0, + "content": "[8] Xiaokang Chen, Mingyu Ding, Xiaodi Wang, Ying Xin, Shentong Mo, Yunhao Wang, Shumin Han, Ping Luo, Gang Zeng, and Jingdong Wang. Context autoencoder for self-supervised representation learning. arXiv preprint arXiv:2202.03026, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.646, + 0.47, + 0.714 + ], + "angle": 0, + "content": "[9] Yen-Chun Chen, Linjie Li, Licheng Yu, Ahmed El Kholy, Faisal Ahmed, Zhe Gan, Yu Cheng, and Jingjing Liu. Uniter: Universal image-text representation learning. In European conference on computer vision, pages 104-120. Springer, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.717, + 0.47, + 0.8 + ], + "angle": 0, + "content": "[10] Mehdi Cherti, Romain Beaumont, Ross Wightman, Mitchell Wortsman, Gabriel Ilharco, Cade Gordon, Christoph Schuhmann, Ludwig Schmidt, and Jenia Jitsev. Reproducible scaling laws for contrastive language-image learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2818-2829, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.802, + 0.47, + 0.857 + ], + "angle": 0, + "content": "[11] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE Conference on Computer Vision and Pattern Recognition, pages 248–255, 2009. 2, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.859, + 0.47, + 0.901 + ], + "angle": 0, + "content": "[12] Li Deng. The mnist database of handwritten digit images for machine learning research. IEEE Signal Processing Magazine, 29(6):141-142, 2012. 6" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.135 + ], + "angle": 0, + "content": "[13] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.137, + 0.894, + 0.218 + ], + "angle": 0, + "content": "[14] Xiaoyi Dong, Jianmin Bao, Ting Zhang, Dongdong Chen, Weiming Zhang, Lu Yuan, Dong Chen, Fang Wen, Nenghai Yu, and Baining Guo. Peco: Perceptual codebook for bert pre-training of vision transformers. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 552-560, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.221, + 0.892, + 0.29 + ], + "angle": 0, + "content": "[15] Xiaoyi Dong, Jianmin Bao, Yinglin Zheng, Ting Zhang, Dongdong Chen, Hao Yang, Ming Zeng, Weiming Zhang, Lu Yuan, Dong Chen, Fang Wen, and Nenghai Yu. Maskclip: Masked self-distillation advances contrastive language-image pretraining, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.292, + 0.892, + 0.373 + ], + "angle": 0, + "content": "[16] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. ICLR, 2021. 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.376, + 0.892, + 0.417 + ], + "angle": 0, + "content": "[17] Li Fei-Fei, R. Fergus, and P. Perona. One-shot learning of object categories. IEEE Transactions on Pattern Analysis and Machine Intelligence, 28(4):594–611, 2006. 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.419, + 0.894, + 0.46 + ], + "angle": 0, + "content": "[18] Pedro F Felzenszwalb and Daniel P Huttenlocher. Efficient graph-based image segmentation. International journal of computer vision, 59:167-181, 2004. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.462, + 0.892, + 0.517 + ], + "angle": 0, + "content": "[19] Zhanzhou Feng and Shiliang Zhang. Evolved part masking for self-supervised learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10386-10395, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.519, + 0.892, + 0.574 + ], + "angle": 0, + "content": "[20] Tsu-Jui Fu, Linjie Li, Zhe Gan, Kevin Lin, William Yang Wang, Lijuan Wang, and Zicheng Liu. Violet: End-to-end video-language transformers with masked visual-token modeling. arXiv preprint arXiv:2111.12681, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.576, + 0.892, + 0.643 + ], + "angle": 0, + "content": "[21] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9729-9738, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.646, + 0.892, + 0.687 + ], + "angle": 0, + "content": "[22] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollar, and Ross Girshick. Masked autoencoders are scalable vision learners. arXiv preprint arXiv:2111.06377, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.689, + 0.892, + 0.771 + ], + "angle": 0, + "content": "[23] Dan Hendrycks, Steven Basart, Norman Mu, Saurav Kadavath, Frank Wang, Evan Dorundo, Rahul Desai, Tyler Zhu, Samyak Parajuli, Mike Guo, Dawn Song, Jacob Steinhardt, and Justin Gilmer. The many faces of robustness: A critical analysis of out-of-distribution generalization. ICCV, 2021. 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.774, + 0.892, + 0.814 + ], + "angle": 0, + "content": "[24] Dan Hendrycks, Kevin Zhao, Steven Basart, Jacob Steinhardt, and Dawn Song. Natural adversarial examples. CVPR, 2021. 4, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.817, + 0.892, + 0.872 + ], + "angle": 0, + "content": "[25] Cheng-Yu Hsieh, Jieyu Zhang, Zixian Ma, Aniruddha Kembhavi, and Ranjay Krishna. Sugarcrepe: Fixing hackable benchmarks for vision-language compositionality. arXiv preprint arXiv:2306.14610, 2023. 2, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.874, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[26] Gabriel Ilharco, Mitchell Wortsman, Ross Wightman, Cade Gordon, Nicholas Carlini, Rohan Taori, Achal Dave," + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "26823" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.092, + 0.47, + 0.146 + ], + "angle": 0, + "content": "Vaishaal Shankar, Hongseok Namkoong, John Miller, Hannaneh Hajishirzi, Ali Farhadi, and Ludwig Schmidt. Openclip, 2021. If you use this software, please cite it as below. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.149, + 0.471, + 0.232 + ], + "angle": 0, + "content": "[27] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In International conference on machine learning, pages 4904-4916. PMLR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.234, + 0.471, + 0.316 + ], + "angle": 0, + "content": "[28] Ioannis Kakogeorgiou, Spyros Gidaris, Bill Psomas, Yannis Avrithis, Andrei Bursuc, Konstantinos Karantzalos, and Nikos Komodakis. What to hide from your students: Attention-guided masked image modeling. In Computer Vision - ECCV 2022, pages 300-318. Springer Nature Switzerland, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.318, + 0.472, + 0.374 + ], + "angle": 0, + "content": "[29] Jacob Devlin Ming-Wei Chang Kenton and Lee Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of naacL-HLT, page 2, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.375, + 0.471, + 0.403 + ], + "angle": 0, + "content": "[30] Alex Krizhevsky. Learning multiple layers of features from tiny images. Technical report, 2009. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.404, + 0.471, + 0.432 + ], + "angle": 0, + "content": "[31] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.434, + 0.469, + 0.46 + ], + "angle": 0, + "content": "[32] LAION-AI. Clip benchmark. https://github.com/LAION-AI/CLIP_benchmark, 2023.6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.462, + 0.47, + 0.53 + ], + "angle": 0, + "content": "[33] Gang Li, Heliang Zheng, Daqing Liu, Chaoyue Wang, Bing Su, and Changwen Zheng. Semmae: Semantic-guided masking for learning masked autoencoders. Advances in Neural Information Processing Systems, 35:14290-14302, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.533, + 0.47, + 0.602 + ], + "angle": 0, + "content": "[34] Junnan Li, Ramprasaath Selvaraju, Akhilesh Gotmare, Shafiq Joty, Caiming Xiong, and Steven Chu Hong Hoi. Align before fuse: Vision and language representation learning with momentum distillation. Advances in neural information processing systems, 34:9694-9705, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.603, + 0.47, + 0.631 + ], + "angle": 0, + "content": "[35] Xianhang Li, Zeyu Wang, and Cihang Xie. Clipa-v2: Scaling clip training with 81.12" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.633, + 0.47, + 0.673 + ], + "angle": 0, + "content": "[36] Yanghao Li, Haoqi Fan, Ronghang Hu, Christoph Feichtenhofer, and Kaiming He. Scaling language-image pre-training via masking, 2023. 1, 2, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.675, + 0.47, + 0.744 + ], + "angle": 0, + "content": "[37] Zhaowen Li, Zhiyang Chen, Fan Yang, Wei Li, Yousong Zhu, Chaoyang Zhao, Rui Deng, Liwei Wu, Rui Zhao, Ming Tang, et al. Mst: Masked self-supervised transformer for visual representation. Advances in Neural Information Processing Systems, 34:13165-13176, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.746, + 0.469, + 0.802 + ], + "angle": 0, + "content": "[38] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dollar, and Larry Zitnick. Microsoft coco: Common objects in context. In ECCV. European Conference on Computer Vision, 2014. 2, 4, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.803, + 0.47, + 0.87 + ], + "angle": 0, + "content": "[39] Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. Roberta: A robustly optimized bert pretraining approach. arXiv preprint arXiv:1907.11692, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.872, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[40] Stuart Lloyd. Least squares quantization in lcm. IEEE transactions on information theory, 28(2):129-137, 1982. 3" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.135 + ], + "angle": 0, + "content": "[41] Norman Mu, Alexander Kirillov, David Wagner, and Saining Xie. Slip: Self-supervision meets language-image pretraining, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.137, + 0.894, + 0.207 + ], + "angle": 0, + "content": "[42] Maria-Elena Nilsback and Andrew Zisserman. Automated flower classification over a large number of classes. In Proceedings of the 2008 Sixth Indian Conference on Computer Vision, Graphics & Image Processing, page 722-729, USA, 2008. IEEE Computer Society. 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.208, + 0.892, + 0.248 + ], + "angle": 0, + "content": "[43] Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.25, + 0.791, + 0.264 + ], + "angle": 0, + "content": "[44] OpenAI. Gpt-4 technical report, 2023. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.266, + 0.892, + 0.39 + ], + "angle": 0, + "content": "[45] Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, John Schulman, Jacob Hilton, Fraser Kelton, Luke Miller, Maddie Simens, Amanda Askell, Peter Welinder, Paul F Christiano, Jan Leike, and Ryan Lowe. Training language models to follow instructions with human feedback. In Advances in Neural Information Processing Systems, pages 27730-27744. Curran Associates, Inc., 2022. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.392, + 0.892, + 0.448 + ], + "angle": 0, + "content": "[46] Adam Paszke, Sam Gross, Soumith Chintala, Gregory Chanan, Edward Yang, Zachary DeVito, Zeming Lin, Alban Desmaison, Luca Antiga, and Adam Lerer. Automatic differentiation in pytorch. 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.449, + 0.892, + 0.491 + ], + "angle": 0, + "content": "[47] Zhiliang Peng, Li Dong, Hangbo Bao, Qixiang Ye, and Furu Wei. Beit v2: Masked image modeling with vector-quantized visual tokenizers. arXiv preprint arXiv:2208.06366, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.493, + 0.892, + 0.562 + ], + "angle": 0, + "content": "[48] Alec Radford, Jong Wook Kim, Chris Hallacy, A. Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In ICML, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.564, + 0.892, + 0.646 + ], + "angle": 0, + "content": "[49] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.648, + 0.892, + 0.731 + ], + "angle": 0, + "content": "[50] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.733, + 0.892, + 0.774 + ], + "angle": 0, + "content": "[51] Benjamin Recht, Rebecca Roelofs, Ludwig Schmidt, and Vaishaal Shankar. DoImagenet classifiers generalize to imageNet?, 2019. 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.776, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[52] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade W Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, Patrick Schramowski, Srivatsa R Kundurthy, Katherine Crowson, Ludwig Schmidt, Robert Kaczmarczyk, and Jenia Jitsev. LAION-5b: An open large-scale dataset for training next generation image-text models. In Thirty-sixth Conference on Neural Information Processing Systems Datasets and Benchmarks Track, 2022. 4" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "26824" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.134 + ], + "angle": 0, + "content": "[53] Jianbo Shi and Jitendra Malik. Normalized cuts and image segmentation. IEEE Transactions on pattern analysis and machine intelligence, 22(8):888-905, 2000. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.136, + 0.47, + 0.191 + ], + "angle": 0, + "content": "[54] Yuge Shi, N Siddharth, Philip Torr, and Adam R Kosiorek. Adversarial masking for self-supervised learning. In International Conference on Machine Learning, pages 20026-20040. PMLR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.194, + 0.471, + 0.223 + ], + "angle": 0, + "content": "[55] Anugya Srivastava, Shriya Jain, and Mugdha Thigle. Out of distribution detection onImagenet-o, 2022. 4, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.224, + 0.47, + 0.279 + ], + "angle": 0, + "content": "[56] Weijie Su, Xizhou Zhu, Yue Cao, Bin Li, Lewei Lu, Furu Wei, and Jifeng Dai. Vl-bert: Pre-training of generic visual-linguistic representations. arXiv preprint arXiv:1908.08530, 2019. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.282, + 0.47, + 0.324 + ], + "angle": 0, + "content": "[57] Hao Tan and Mohit Bansal. Lxmert: Learning cross-modality encoder representations from transformers. arXiv preprint arXiv:1908.07490, 2019. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.326, + 0.47, + 0.395 + ], + "angle": 0, + "content": "[58] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Hervé Jégou. Training data-efficient image transformers & distillation through attention. In International conference on machine learning, pages 10347-10357. PMLR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.398, + 0.47, + 0.439 + ], + "angle": 0, + "content": "[59] Aaron Van Den Oord, Oriol Vinyals, et al. Neural discrete representation learning. Advances in neural information processing systems, 30, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.442, + 0.47, + 0.483 + ], + "angle": 0, + "content": "[60] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.485, + 0.47, + 0.541 + ], + "angle": 0, + "content": "[61] Andrea Vedaldi. Cats and dogs. In Proceedings of the 2012 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), page 3498-3505, USA, 2012. IEEE Computer Society. 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.543, + 0.47, + 0.599 + ], + "angle": 0, + "content": "[62] Haohan Wang, Songwei Ge, Zachary Lipton, and Eric P Xing. Learning robust global representations by penalizing local predictive power. In Advances in Neural Information Processing Systems, pages 10506-10518, 2019. 4, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.601, + 0.47, + 0.657 + ], + "angle": 0, + "content": "[63] Yiqing Wang, Zihan Li, Jieru Mei, Zihao Wei, Li Liu, Chen Wang, Shengtian Sang, Alan Yuille, Cihang Xie, and Yuyin Zhou. Swinnmm: Masked multi-view with swim transformers for 3d medical image segmentation. In MICCAI, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.659, + 0.47, + 0.728 + ], + "angle": 0, + "content": "[64] Chen Wei, Haoqi Fan, Saining Xie, Chao-Yuan Wu, Alan Yuille, and Christoph Feichtenhofer. Masked feature prediction for self-supervised visual pre-training. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14668-14678, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.73, + 0.47, + 0.771 + ], + "angle": 0, + "content": "[65] Zihao Wei, Chen Wei, Jieru Mei, Zeyu Wang, Xianhang Li, Huiyu Wang, Alan Yuille, Yuyin Zhou, and Cihang Xie. MAE are secretly efficient learners, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.774, + 0.47, + 0.843 + ], + "angle": 0, + "content": "[66] Alex Wilf, Syeda Nahida Akter, Leena Mathur, Paul Pu Liang, Sheryl Mathew, Mengrou Shou, Eric Nyberg, and Louis-Philippe Morency. Difference-masking: Choosing what to mask in continued pretraining. arXiv preprint arXiv:2305.14577, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.846, + 0.47, + 0.9 + ], + "angle": 0, + "content": "[67] Jiahao Xie, Wei Li, Xiaohang Zhan, Ziwei Liu, Yew Soon Ong, and Chen Change Loy. Masked frequency modeling for self-supervised visual pre-training. arXiv preprint arXiv:2206.07706, 2022. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.471, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.162 + ], + "angle": 0, + "content": "[68] Zhenda Xie, Zheng Zhang, Yue Cao, Yutong Lin, Jianmin Bao, Zhuliang Yao, Qi Dai, and Han Hu. Simmim: A simple framework for masked image modeling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9653-9663, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.164, + 0.893, + 0.233 + ], + "angle": 0, + "content": "[69] Fengyu Yang, Chao Feng, Ziyang Chen, Hyoungseob Park, Daniel Wang, Yiming Dou, Ziyao Zeng, Xien Chen, Rit Gangopadhyay, Andrew Owens, et al. Binding touch to everything: Learning unified multimodal tactile representations. arXiv preprint arXiv:2401.18084, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.234, + 0.893, + 0.289 + ], + "angle": 0, + "content": "[70] Yifan Yang, Weiquan Huang, Yixuan Wei, Houwen Peng, Xinyang Jiang, Huiqiang Jiang, Fangyun Wei, Yin Wang, Han Hu, Lili Qiu, and Yuqing Yang. Attentive mask clip, 2022. 1, 2, 4, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.291, + 0.893, + 0.359 + ], + "angle": 0, + "content": "[71] Peter Young, Alice Lai, Micah Hodosh, and Julia Hockenmaier. From image descriptions to visual denotations: New similarity metrics for semantic inference over event descriptions. Transactions of the Association for Computational Linguistics, 2:67-78, 2014. 4, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.361, + 0.893, + 0.415 + ], + "angle": 0, + "content": "[72] Jiahui Yu, Zirui Wang, Vijay Vasudevan, Legg Yeung, Mojtaba Seyedhosseini, and Yonghui Wu. Coca: Contrastive captioners are image-text foundation models. arXiv preprint arXiv:2205.01917, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.418, + 0.893, + 0.486 + ], + "angle": 0, + "content": "[73] Lu Yuan, Dongdong Chen, Yi-Ling Chen, Noel Codella, Xiyang Dai, Jianfeng Gao, Houdong Hu, Xuedong Huang, Boxin Li, Chunyuan Li, et al. Florence: A new foundation model for computer vision. arXiv preprint arXiv:2111.11432, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.488, + 0.893, + 0.557 + ], + "angle": 0, + "content": "[74] Mert Yuksekgonul, Federico Bianchi, Pratyusha Kalluri, Dan Jurafsky, and James Zou. When and why vision-language models behave like bags-of-words, and what to do about it? In The Eleventh International Conference on Learning Representations, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.558, + 0.893, + 0.613 + ], + "angle": 0, + "content": "[75] Jinghao Zhou, Chen Wei, Huiyu Wang, Wei Shen, Cihang Xie, Alan Yuille, and Tao Kong. ibot: Image bert pre-training with online tokenizer. arXiv preprint arXiv:2111.07832, 2021. 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.613 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "26825" + } + ] +] \ No newline at end of file diff --git a/2024/Efficient Vision-Language Pre-training by Cluster Masking/24f9f648-410c-4f99-8615-dc6d0380efad_origin.pdf b/2024/Efficient Vision-Language Pre-training by Cluster Masking/24f9f648-410c-4f99-8615-dc6d0380efad_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..144d52ab09ac2653eab2cf06bbfa4f6899197b4a --- /dev/null +++ b/2024/Efficient Vision-Language Pre-training by Cluster Masking/24f9f648-410c-4f99-8615-dc6d0380efad_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:47c28d03b14ae4934c0e5cde1fd68f66dd5f029f58b406478118e68ea257ca06 +size 6501535 diff --git a/2024/Efficient Vision-Language Pre-training by Cluster Masking/full.md b/2024/Efficient Vision-Language Pre-training by Cluster Masking/full.md new file mode 100644 index 0000000000000000000000000000000000000000..dc2b51e379444d40ea2ef5b0a9bd1bf9328d9142 --- /dev/null +++ b/2024/Efficient Vision-Language Pre-training by Cluster Masking/full.md @@ -0,0 +1,363 @@ +# Efficient Vision-Language Pre-training by Cluster Masking + +Zihao Wei* Zixuan Pan* Andrew Owens University of Michigan + +# Abstract + +We propose a simple strategy for masking image patches during visual-language contrastive learning that improves the quality of the learned representations and the training speed. During each iteration of training, we randomly mask clusters of visually similar image patches, as measured by their raw pixel intensities. This provides an extra learning signal, beyond the contrastive training itself, since it forces a model to predict words for masked visual structures solely from context. It also speeds up training by reducing the amount of data used in each image. We evaluate the effectiveness of our model by pre-training on a number of benchmarks, finding that it outperforms other masking strategies, such as FLIP, on the quality of the learned representation. + +# 1. Introduction + +Images contain a great deal of redundant information, making it challenging to efficiently learn representations from them at scale. Recent work has addressed this problem by masking image patches during vision-language contrastive learning [15, 33, 36, 70]. One simple approach is to drop a large fraction of the patches at random, making training more efficient by reducing the computational cost and memory usage in each training iteration [36]. An alternative strategy is to mask sets of semantically related patches [15, 33, 70], such as those that belong to the same object. This forces the learned model to predict words that describes missing scene structures from context, improving the learned representation. However, this approach requires a separate mechanism to group together semantically related patches, which adds considerable complexity to the learning procedure and is computationally expensive. + +We propose a simple masking strategy for multimodal contrastive learning that avoids these shortcomings. During training, we mask random clusters of patches (Fig. 1). For this clustering, we use the patches' raw RGB values as the feature representation. Our approach takes advantage of the fact that simple measures of visual similarity can often capture coherent visual structures, such as object parts [18, 53], + +![](images/359733b6359ec5b4e02180bb3be237448bf176d9397519aab975778441c107b8.jpg) + +![](images/2a783606c85da704f7a2c1320a7e5e9c347d142920272f838a1a6f82847a212f.jpg) + +![](images/5a7767e3e2c64ce15fbba58836cf176d4666fc6fd50671da9645db1ea73e47b7.jpg) +(b) A burgundy low-rider show-quality, hot rod Chevy truck. + +![](images/f106417e32da587fc2930d19502b6f7d9fdc598d5006e26d85da1d3c38b8cb35.jpg) + +![](images/a03b9ce5d8b77083b3c26dd5a910242b0795d1566742e624387e88eb7a3e6ce2.jpg) + +![](images/dba2b09c764298c9bb464f9f691664adbfb61dc23c0e10abc8f5f41ef30ce8eb.jpg) +(a) A red and gold +painted fire +hydrant on the street. + +![](images/b37a70da0e179ab31adf7cc3cca7eb91e6366ac96b02c29dd354c1243f163d34.jpg) + +![](images/13423bc72e420b736b579cce42c50d583de23038efda2d74de1536338b6d0661.jpg) + +![](images/3285c2ad9cb1ca563e5525586e7b3e1a1bc9bd417b2b83ffa38dea7527290c77.jpg) +(c) Two surf boars on a beach near the water. +Figure 1. Cluster masking. We mask random clusters of visually similar image patches when training contrastive vision-language models (bottom). This masking strategy distinguishes our approach from methods that independently mask image patches for efficiency [36] (middle), while providing a similar improvement in training speed. It provides an extra learning signal, since it forces a model to predict words for missing scene structures solely from context. + +especially when clusters are sampled randomly (Fig. 1). Our approach thus leads to more efficient training, like approaches that independently drop patches [36], while improving the learned representation via context prediction. + +We take inspiration from masked region classification, a pre-training task widely used in vision-language models [9, 56, 57]. These models extract object features, then predict object labels for the randomly masked out regions. Our masking approach provides a similar training signal, since meaningful labels are included in the image caption. For example, as shown in Figure 1(a), the model is tasked with associating the words "fire hydrant" with an image even with the hydrant itself is mostly masked out. + +We train our model on Conceptual 12M dataset [4] and evaluate our learned representation on a number of downstream tasks. These tasks include the zero-shot classification and linear probing on ImageNet [11], text and image retrieval on MS-COCO [38], and the SUGARCREPE language composition benchmark [25]. In our experiments, our model outperforms FLIP [36] and CLIP [49] on downstream performance, while having efficiency comparable with FLIP. We also show that the performance can further be improved by using the model's learned feature embedding during clustering. + +# 2. Related Work + +Contrastive Vision-Language Pre-training. Vision-Language Pre-training (VLP) focuses on establishing connections between images or their components and human-interpretable language. This field initially evolved from transferring supervised learning models, which incorporated object detection modules to generate fine-grained visual labels [9, 56, 57]. Subsequently, there was a shift towards large-scale learning using noisy web data, moving away from reliance on fine-grained labels [1, 27, 34, 50, 69, 72, 73]. A significant development in this domain was CLIP [50], which applied contrastive learning techniques [6, 21] to train models to associate correct image-text pairs and dissociate incorrect ones. CLIP scaled contrastive visual-language models significantly beyond previous work, enabling strong feature learning and zero-shot performance. However, further scaling significantly increases the pre-training demands, requiring larger datasets and batch sizes. + +In response to these challenges, recent research has explored incorporating masking into images to reduce training time and allow for more samples per batch [15, 20, 36, 70]. Methods such as MaskClip [15], FLIP [36], and VIOLET [20] have implemented random masking strategies. Yet, it has been noted that random masking may not be as effective on relatively small datasets [35, 70]. To address this, ACLIP [70] introduced a method of masking tokens with low cross-attention scores with text. However, this approach necessitates two forward passes to generate the attention map and requires additional computational modules [41]. In our work, we aim to avoid these limitations, proposing an effective masking method that is based on a patch's raw RGB values. + +Masked Image Modeling. In the field of language modeling, the effectiveness of models that learn to reconstruct corrupted inputs for generating robust features has been recognized [29, 39]. This approach, known as Mask Language Modeling (MLM), has been adapted in the realm of image processing as Mask Image Modeling (MIM). MIM techniques involve reconstructing either image patches or + +their features [2, 5, 8, 14, 22, 63-65, 68]. The pioneering work in BEIT [2] introduced the reconstruction of discrete tokens, akin to VQ-VAE [59], using block-wise masking. This method demonstrated results on par with contrastive learning and self-distillation methods [3, 7] during model fine-tuning. Later approaches include PeCo's [14] novel visual codebook learning method and BEIT V2's [47] integration of self-distillation methods, using a teacher-student backbone and feature-level KL divergence loss [58]. Further exploration in this field has led to the use of natural image signals as reconstruction targets, moving away from learned features. Examples include SimMIM [68], which reconstructs pure RGB values, MaskFeat [64], introducing reconstruction of the Histogram of Oriented Gradients (HOG) features, and MAE [22], which reconstructs pixel-normalized RGB values. Our work draws inspiration from these studies, particularly in using pixel-normalized RGB values to compute patch similarities, arguing for a more effective distribution of patch features. + +Masking Strategies in MIM. Parallel investigations have focused on masking strategies in MIMs [19, 28, 33, 37, 54, 66, 67]. Early works like BEIT and its successors used block masking, while others such as SimMIM, MaskFeat, and MAE applied random patch-wise masking. Attention-based masking strategies have also been explored, typically using attention maps from vision transformers. MST [37] masks less essential parts with low attention scores, using a reconstruction loss approach. In contrast, AttnMask [28] masks highly attentive patches and applies self-distillation loss. These methods involve simultaneous updates of attention maps and masks during training. A potential limitation of this approach is that insufficiently trained attention maps may not capture structured features effectively. SemMAE [33], starting with iBot features [75], adopts an easy-to-hard masking strategy, starting with masking parts within clusters and gradually expanding to entire clusters. Wilf et al. [66] introduce a unique entity-reinforced language model for masking objects in video frames. However, the reliance on pre-trained features or extracting attention maps can be computationally intensive. Evolved Part Masking [19] proposes using EM algorithm on attention maps to get a clustering before performing SemMAE style masking. Our approach also adopts a cluster based masking strategy in vision-language pre-training, enabling faster pre-training without requiring additional modifications to the model. + +# 3. Method + +We propose a cluster based masking strategy for contrastive vision-language pre-training, focusing on masking random clusters with visually similar semantics. Our method selects random anchor patches as cluster centers and computes pairwise patch distances to form clusters. These clus + +![](images/ff16a277160bab84a9c9d544c6dc9b3baafd381cc5fd6ec67322211670a6659f.jpg) +Anchor Patch + +![](images/613d61cce0bbf808ae881e59cca840a596aa12f4ccc087eeaa28f2858cd613c0.jpg) +Distance Heatmap + +![](images/9144a605330311de07b51888b727f52f0012e9134aa8ccb6289442802c452134.jpg) +Masked Image +Figure 2. Choosing clusters. The process begins by randomly selecting anchor patches from the image. We then calculate the pairwise distances among all patches. Clusters formed within a distance threshold are masked out. We show cluster obtained from a single anchor patch. + +ters are then masked entirely. To enhance accuracy in cluster formation, we introduce an adaptive layer for refining the distance matrix. Additionally, attention masks and a hard patch cutoff are used to ensure uniform input sizes are consistent in batches for auto differentiation. + +# 3.1. Contrastive Vision-Language Pre-training + +Our approach builds on contrastive vision-language pretraining methods, such as CLIP [49]. We use contrastive learning to align embeddings of matching text-image pairs and separate those of non-matching pairs. This process is steered by two symmetric InfoNCE losses [43]: the vision-to-language loss $\mathcal{L}_{v\rightarrow l}$ and its counterpart, the language-to-vision loss $\mathcal{L}_{l\rightarrow v}$ . The vision-to-language loss is defined as: + +$$ +\mathcal {L} _ {v \rightarrow l} = - \log \frac {\exp (\sin (I , T) / \tau)}{\sum_ {j = 1} ^ {N} \exp (\sin (I , T _ {j}) / \tau)}, \tag {1} +$$ + +where $I$ and $T$ are the embeddings for the image and text respectively, sim denotes the similarity function (we use a dot product), and $\tau$ is a temperature parameter. Similarly, $\mathcal{L}_{l\to v}$ is formulated by normalizing the loss using a batch of $N$ image examples $\{I_j\}_{i = 1}^N$ instead of text examples $\{T_j\}_{i = 1}^N$ . + +# 3.2. Cluster Masking + +We introduce a masking strategy that drops out random clusters. While one option would be to use an off-the-shelf clustering method, such as K-Means [40], we choose instead to use a simple and efficient method that results in a random clustering each training iteration (Figure 2). Our approach resembles a single iteration of K-Means, and works by selecting a set of exemplar patches, which each define a cluster. In our experiments, we also evaluate masking clusters obtained using K-Means as an alternative approach. + +We split an input $H\times W$ image into patches, following [16]. We then compute the pairwise cosine similarity between every pair of normalized patches, which we use as a distance function $d(\mathbf{x},\mathbf{y})$ . We choose a small subset (less than $5\%$ ) of these patches at random to act as cluster centers. For each of these selected anchor patches, we define a cluster consisting of patches that lie within a distance + +$r$ . The cluster for an exemplar patch $\mathbf{x}$ is represented by: $\mathrm{S_x} = \{\mathbf{y} \mid d(\mathbf{x}, \mathbf{y}) \leq r\}$ for image patches. + +All patches within a cluster are masked out. The distance threshold $r$ is automatically searched before training according to an average masking ratio. We provide a simplified pseudocode of the masking strategy in Algorithm 1. + +Clustering Embedding Features. Another variant of patch feature is the combination of pure RGB values and patch embedding layer features from transformers [16]. When computing similarity scores, we integrate these two measures into a weighted sum, where the weight of each measure is determined by: + +$$ +d (\mathbf {x}, \mathbf {y}) = \alpha \cdot d _ {r g b} (\mathbf {x}, \mathbf {y}) + (1 - \alpha) \cdot d _ {e m b} (\mathbf {x}, \mathbf {y}), \tag {2} +$$ + +where $\mathbf{x}$ and $\mathbf{y}$ represent two patches, $d_{rgb}$ is the cosine similarity based on pure RGB values, and $d_{emb}$ is the cosine similarity based on the transformer's embedding features. The weight parameter $\alpha$ linearly increases from 0 to 1 during training. + +The embedding layer is calculated before the patches enter the transformer, thus we could reuse the patch embeddings in the transformer without computing them twice. Using the embedding layer is advantageous because it incorporates positional encodings [60]. This integration potentially introduces spatial constraints, which we believe can further enhance our masking strategy. + +Handling Batched Inputs. Deep learning libraries, like PyTorch [46], typically process batched inputs of uniform size. However, in our method, the mask ratio would vary across different images, leading to fluctuations in the number of patches. To accelerate the process, we introduce a minimum mask ratio threshold $\beta$ for each image. If the calculated mask ratio for an image doesn't meet this predefined threshold, we proceed to randomly drop patches until the desired ratio is achieved. Conversely, for images with patch counts less than the threshold, we use attention masks to avoid masked parts engaging in attention calculation [13]. + +# 4. Experiments + +We present a comprehensive evaluation of our proposed algorithm to exhibit the performance, robustness, scalability, and efficiency of our framework. + +# 4.1. Implementation Details + +Datasets and Training Details. We train our model using the Conceptual 12M (CC12M) dataset[4], containing 12 million unique image-text pairs, for pre-training our vision-language models. We use ViT-B/16 as backbone for image encoder. The text encoder is a 12-layer transformer, equipped with 8 multi-head attention units and 512-dimensional embeddings. Input images are processed at a + +Algorithm 1 Pseudocode of cluster based masking in a PyTorch-like style. +```python +# img: the image for mask +# mask_ratio: the ratio for chosing anchor patches +# r: the threshold for computing cluster +def generate_mask(img, mask_ratio, r): + # Make image into shape (N,L,C). + # N: number of samples per batch + # L: number of patches + # C: size of feature dimension + img = patchify(img) + #Normalize each patch + img = (img - img.mean(dim=-1)) / img.std(dim=-1) + #Compute pairwise cosine similarity matrix + x = img / img(norm(dim=-1)) + distance = bmm(x, x.transpose(-2, -1)) + #Generate a boolean masking matrix of shape (N, L) + init_mask = random_batch_indicies(mask_ratio) + init_mask = init_mask[:, :, None].float() + #Get the cluster-based mask candidates = (distance * init_mask) ≥ r + cluster_mask = candidates.sum(1) ≥ 1 + #Mask both the anchor patches and clusters + mask = init_mask or cluster_mask + return mask +``` + +resolution of $224 \times 224$ , and text inputs are adjusted to 77 tokens, either by truncation or padding. A class token is transformed into a 512-dimensional feature embedding via a multi-layer perceptron (MLP). For optimization, we use the AdamW optimizer with a learning rate of $5 \times 10^{-4}$ , $\beta_{1} = 0.9$ , and $\beta_{2} = 0.98$ . We use a batch size of 256 per GPU, and train using 8 NVIDIA A40 GPUs. + +Our method comes in three variants: K-Means, RGB and Embedding. The RGB model clusters based on raw image patches, while the embedding model integrates patch embedding features with RGB for clustering. In the K-Means variant of the model, we mask out half of the clusters randomly. The model constructs 12 clusters and runs for a maximum of 10 iterations. For both RGB and embedding models, we set an average masking ratio of $50\%$ , following the recommendations of FLIP [36] for optimal masking ratios. In the RGB approach, we use a $50\%$ cutoff for Ours- $\mathsf{RGB}_{0.5}$ and $30\%$ for Ours- $\mathsf{RGB}_{0.3}$ , whereas the Ours-Embedding model uses a $30\%$ cutoff. Also, the RGB model selects anchor patches at a $3\%$ ratio, compared to a $5\%$ ratio in the embedding model. + +Baselines. In our study, we establish baselines using three models: CLIP, FLIP, and $\mathrm{FLIP}_{\mathrm{Attn}}$ , each trained from scratch on the CC12M dataset. These baseline models are derived from the open-source implementation of CLIP, known as OpenClip [10, 26, 48, 52]. For both FLIP and $\mathrm{FLIP}_{\mathrm{Attn}}$ , we implement a patch dropout ratio of $50\%$ . Specifically, FLIP uses a random dropout approach, whereas $\mathrm{FLIP}_{\mathrm{Attn}}$ adopts an attention-based masking strategy inspired by ACLIP [70]. This strategy involves processing the image through the encoder and then averaging + +across attention heads in the final transformer block to determine attention scores. Patches that receive the highest attention in relation to the [CLS] token are retained. + +To ensure a fair comparison among these methods, we keep the number of patches consistent the same as ours within a single batch, which means for FLIP and $\mathrm{FLIP}_{\mathrm{Attn}}$ we apply a batch size of 256 for one GPU and for CLIP, we use 128 instead. Additionally, we apply a scaling law on learning rate across different models. + +Evaluation Details. Our models are tested across various benchmarks to ensure its robustness and effectiveness. We conduct zero-shot image-to-text and text-to-image retrieval tasks on COCO [38] and Flickr [71], assessing its performance meticulously. Further more, we evaluate the models' image representation quality by reporting both the zero-shot classification and linear probing performance on three mainstream datasets: ImageNet [11], CIFAR-10, and CIFAR-100 [30]. Zero-shot results of some other datasets like ImageNet variants [23, 24, 51, 55, 62], Caltech101 [17], Flowers [42] and Pets [61], are also reported to verify the method's robustness. + +For these tasks, our approach adheres strictly to the implementation used in the CLIP benchmark, ensuring consistency and reliability in the evaluation process. Furthermore, we assess the effectiveness of our methodology on language composition tasks using SUGARCREPE [25]. This evaluation aims to determine its adaptability and efficiency across various contexts, including object, attribute, and relation manipulations. Within the SUGARCREPE framework, models are tasked with identifying the correct caption that accurately describes an image, distinguishing it from a closely related but incorrect text hard negatives. The hard negatives is characterized by minor composition differences from the accurate caption. + +# 4.2. Main Results + +Visualization of Clusters. Figure 3 offers a visual depiction of our cluster based masking technique as outlined in the methodology section. For this illustration, we randomly select a number of image-text pairs from the COCO validation set and apply our masking method to the pure RGB data of the images. The visualization is showing the masking result of the two-stages. In the first stage, a subset of patches (5%) is randomly selected as anchor patches from the pool of all image patches, which are annotated with the red boxes. In the second stage, we visualize the masked clusters that are calculated based on the similarity matrix, where each cluster is represented by a distinct color. + +Zero-shot Retrieval Results. In our investigation into the model's understanding of the relationship between visual and linguistic representations, we conduct zero-shot + +![](images/ffb62e24ea35d217bcad8b3d1a79f6c5a8dcc52634fd02c98c98a24181891b88.jpg) +Figure 3. Visualization of cluster masks. Different colors represent distinct clusters formed by the similarity matrix calculated from the chosen anchor patches. + +retrieval tests on several leading retrieval benchmarks. The results, detailed in Table 1, provide insights into the performance of our approach against others, particularly in the context of Image2Text and Text2Image's recall precision at top1(R1), top5(R5) and top10(R10) metrics. + +In the evaluation on the MS-COCO [38], Flickr8k, and Flickr30k [71] datasets, our model outperforms both the baselines in most parts. Notably, in the Image-to-Text tasks, our model performs best in most datasets, with the exception of a slight performance decrease compared to $\mathrm{FLIP}_{\mathrm{Attn}}$ on the MS-COCO dataset. We attribute this success to our training strategy, which prioritizes primary clusters and minimizes the influence of noise. Furthermore, we observe that methods combining RGB information with token embeddings outperform those relying solely on RGB. We hypothesize that this is because the embedding layer, which contains slightly higher-level information. + +When comparing FLIP to CLIP, FLIP's performance is noticeably weaker, even with large batch sizes. We suspect that FLIP's sub-optimal results in our experimental settings may not fully exploit its strengths. This aligns with findings from other studies, such as Yang et al.'s research on ACLIP [70], which also noted FLIP's limitations. We observe that using attention scores for masking can improve performance compared to purely random masking. However, random masking still falls short of our cluster based masking or even the original CLIP method in some benchmarks. + +Results on Zero-shot Classification and Linear Probing. We evaluate our model on several widely recognized clas + +sification benchmarks. The zero-shot classification results are presented in Table 2, while the linear probing results can be found in Table 3. For better evaluating the time spent on training, we normalize all method's training time by the CLIP's training time, which is considered as $1 \times$ . + +When comparing our model's performance to CLIP (i.e., no masking), our model demonstrates superior results on the majority of test cases, showcasing an average improvement of $+2.1\%$ , with about $+36\%$ speeding up. In comparison to the FLIP strategy, which has a similar training duration, our model has an improvement of $+5.5\%$ . In comparison to the $\mathrm{FLIP}_{\mathrm{Attn}}$ , our model does not need the attention map for guidance, which gives a much fast training speed, while having a performance of $+2.6\%$ on average. + +Out of 12 datasets on the zero-shot classification benchmark, our RGB and embedding model achieves the top performance on 11 of them. In particular, it obtains strong performance on the ImageNet variants: ImageNet-A [24], ImageNet-O [55], ImageNet-R [55], and ImageNet-S [62], which often contain challenging and diverse images. The RGB version of our method also significantly outperforms FLIP and surpasses the CLIP model, especially on ImageNet and its variants, which demonstrates the effectiveness of our method with even the natural guidance. + +The linear probing results further suggest the effectiveness of our method. Our models achieve $+1.8\%$ accuracy on ImageNet, $+3.1\%$ on CIFAR-10, and $+4.2\%$ on CIFAR-100. + +Language Composition. A potential drawback of our method might be understanding compositions of concepts + +text $\rightarrow$ image +image $\rightarrow$ text + +
MS-COCOFlickr8kFlickr30kMS-COCOFlickr8kFlickr30k
R1R5R10R1R5R10R1R5R10R1R5R10R1R5R10R1R5R10
CLIP [49]34.6061.9872.7255.7081.6089.9058.5083.8089.1023.4947.8059.6640.5468.9080.2043.1870.4480.40
FLIP [36]32.6259.1470.6455.0080.9088.9053.8080.8088.5022.5646.0858.0940.3268.1078.6441.5267.9077.46
FLIPAttn [70]33.6660.1871.0253.7080.0987.9955.2981.4087.6023.8948.3360.0440.5868.8878.8643.0670.1078.82
Ours-Kmeans33.6861.1471.9755.1083.3090.9055.4082.0088.9022.6046.9358.8440.1069.8679.9141.3268.5077.58
Ours-RGB0.532.8260.2071.4052.1081.2089.5054.9081.2088.0022.9747.2959.2140.8468.7279.1441.8070.2079.42
Ours-RGB0.335.8761.5072.3454.7081.4090.7057.6083.9090.3023.6547.5459.2540.9068.2279.0042.9269.9679.12
Ours-Embedding34.2661.9673.3057.0082.7090.1055.8084.2089.6023.7748.1859.7642.0069.4079.6443.3070.9280.16
+ +Table 1. Zero-shot retrieval results. We evaluate on MS-COCO [38], Flickr8k and Flickr30k datasets [71], where the Recall@1 (R1), Recall@5 (R5), and Recall@10 (R10) are reported. + +
MethodN.T.IN-1K [11]IN-A [24]IN-O [55]IN-R [23]IN-S [62]INv2 [51]MNIST [12]Cal101 [17]CIFAR-10 [31]CIFAR-100 [31]Flowers [42]Pets [61]Average
CLIP [49]1.00×36.18.038.447.624.930.711.773.557.725.026.053.936.1
FLIP [36]0.53×34.47.139.541.420.129.510.470.452.824.525.346.033.5
FLIPAtm [70]1.06×35.28.139.445.123.730.19.473.561.627.125.751.235.8
Ours-Kmeans0.70×35.57.238.343.022.230.19.769.961.425.325.652.135.0
Ours-RGB0.50.54×36.07.739.845.323.830.511.272.263.627.226.155.036.5
Ours-RGB0.30.64×36.68.839.945.924.931.89.472.363.126.325.457.336.8
Ours-Embedding0.64×36.38.139.647.925.430.711.073.770.732.028.455.438.2
+ +Table 2. Zero-shot classification result. We evaluate popular datasets using clip-benchmark [32]. The training time is normalized according to the CLIP's training time. Here N.T. represents normalized time against and IN represents ImageNet. + +
MethodCIFAR-10CIFAR-100IN-1K
CLIP88.067.462.3
FLIP85.965.561.3
\(FLIP_{Attn}\)86.466.162.0
Ours-Kmeans88.069.162.2
Ours-\(RGB_{0.5}\)86.766.262.5
Ours-\(RGB_{0.3}\)88.668.763.1
Ours-Embedding89.069.762.7
+ +Table 3. Linear probing result. All methods are trained for 10 epochs at learning rate of 1e-3. For CIFAR-10 and CIFAR-100, we use a batch size of 64 and for ImageNet-1k the batch size is 1024. + +in language. As we mask out clusters, there is a risk that the model may increasingly adopt bag-of-words tendencies [74], which could impede its ability to learn the relationships between objects. For example, if an image is captioned with "dog on grass", the grass may be masked for a large portion in our model as they are highly similar to each other. This will make learning the relation "on" difficult. Therefore, we apply SUGARCREPE [25] benchmarks to test the model's ability to understand language compositions. SUGARCREPE benchmarks assess this by generating negative captions through manipulations like adding, swapping, or replacing concepts in sentences, followed by text retrieval tests to evaluate the model's accuracy in selecting the correct answer. From the our test results, which shown in Table 4, our model yields comparable results in Relation tests and demonstrates a significant enhancement in Object and Attribution tests, with an average improvement of $+3.9\%$ and $+3.0\%$ respectively, compared to FLIP. This improvement may stem from the masking of entire objects, which simplifies the challenge of contrastive learning by reducing ambiguity. This clarity facilitates the model's learning of relationships, a crucial factor for composition + +understanding. + +Qualitative Comparison of Masking Strategies Our method outperforms the random masking strategy by preserving more semantic content in the unmasked image patches, a comparison showcased in Figure 1. The advantage of our technique is further explored by the captioning experiment detailed in Figure 4, wherein two sets of images, each masked differently, are fed into a captioner, GPT-4 [44, 45]. The captioner is prompted to generate MS-COCO-style captions for the unobscured sections. When comparing these captions to the standard references, it becomes clear that our cluster based masking not only retains key elements but also the interrelations among them. For instance, our approach accurately enables the captioning system to identify an airplane in the first example and to describe the baseball player's action in the second, while the random masking strategy failed to achieve this clarity. These results indicate that our masking method provides a more detailed comprehension of the image. + +# 4.3. Ablation Study + +Ablation on Anchor Patch Ratio. In our study, we conducted an ablation on finding the optimal proportion of patches to serve as anchor patches. The results of this ablation are summarized in Figure 5. We use zero-shot learning results on the ImageNet-1k dataset as the benchmark for assessing the quality of the representations learned by our model. Additionally, we calibrate the threshold for each experiment to ensure that the average final masking ratio was maintained at $50\%$ . + +Our findings indicate that a smaller proportion of anchor patches tends to yield superior performance. We hypothesize that this improvement is due to the decreased ran + +
REPLACESWAPADDAverage
ObjectAttributeRelationObjectAttributeObjectAttributeObjectAttributeRelation
CLIP85.7779.1864.5161.7858.7174.2468.3573.7168.7564.51
FLIP84.0775.8866.0060.1661.5671.6763.1571.9766.8666.00
\( FLIP_{Attn-0.5} \)86.6275.5063.2252.4463.0671.6566.7671.5768.3963.22
\( FLIP_{Attn-0.3} \)86.0775.0062.2360.5759.1674.6868.6472.8263.4762.23
Ours-Kmeans84.5076.9063.0962.2061.8671.7765.4673.6667.6063.09
Ours-RGB0.586.8673.4760.2459.3563.3673.3366.7673.1868.4260.24
Ours-RGB0.386.1375.1364.6566.6763.3674.9271.2475.9169.9164.65
+ +![](images/3dc13961a49aaf88cf054ec05906767b78d76027460b0e26d4ec608c7a544209.jpg) +Figure 4. Generated caption from visible patches. We process the masked images through GPT-4 [44, 45] to create captions for the unmasked segments. + +domness in the selection of anchor patches, which in turn enhances the clustering performance by providing a more stable set of reference points for the model to learn from. However, the masking ratio cannot be too small, as the similarity threshold will become too small, which may reduce the clustering quality. + +![](images/bc1e7ca87859ffff6a472a75dd482e1122d51f255d7a45d6e5c1b5e5f6842664.jpg) +Figure 5. Effect of anchor patch ratio. All the final masking ratio is tuned to be $50\%$ . + +Ablation on Minimum Mask Ratio. We further demonstrate the capability of our method by setting the minimum mask ratio $\beta$ the same as the average masking ratio of FLIP, + +as shown in Table 1, 2, and Table 5. For our method, the cutoff ratio denotes the minimum mask ratio applied and the true visible patch ratio is shown in visible ratio. For the FLIP counterpart, it maintains a consistent mask ratio across all images. The results indicate that our method not only matches the speed of FLIP but also surpasses FLIP in zero-shot ImageNet-1K classification accuracy with a $+1.6\%$ improvement even by seeing fewer patches. The attention based masking method with less mask ratio achieves similar performance as ours but the speed is much slower. + +These findings suggest that cluster based masking serves as an effective denoising technique for the dataset. A reason for this enhanced performance is that we could easily mask out typically irrelevant areas, such as uniformly colored backgrounds, which are less informative and oftentimes do not correspond to any word in the caption. This targeted approach enables the model to focus on more meaningful content within the images. + +Additionally, our findings reveal an improved feature learning by the model when a smaller random masking is applied. By reducing the cutoff ratio from $50\%$ to $30\%$ , we observed a $1\%$ enhancement in classification accuracy. Thus, there is some trade-off between the model's performance and speed. Despite this, our model with larger masking cut off still remains significantly faster compared to attention-based masking or the original CLIP method. + +Table 4. Language composition test result. This table presents the performance of models on the SUGARCREPE evaluation, which involves replacing, swapping, or adding atomic concepts such as objects, attributes, and relations in a sentence to create mismatched captions. + +
MethodβVisible RatioNormalized TimeIN-1K
FLIP50%50%0.84×34.4
FLIP30%70%1.00×35.4
\( \text{FLIP}_{\text{attn}} \)50%50%1.73×35.2
\( \text{FLIP}_{\text{attn}} \)30%70%1.97×36.6
Ours-RGB50%43%0.84×36.0
Ours-RGB30%50%1.00×36.6
+ +Table 5. Ablation on minimum mask ratio $\beta$ . Comparison of various methods against different minimum masking ratios. The zero-shot ImageNet-1k classification results are used as metric. The time is normalized to ours RGB model with $\beta = 30\%$ + +Ablation on Pixel Normalization. In our experiments, we incorporate pixel normalization (making each patch mean zero and unit standard deviation 1) into the process of computing the similarity matrix for images. This yields + +a performance improvement of $+1.1\%$ , as shown in the results presented in Table 6a. The underlying rationale for this enhancement is attributed to the standardization of image patches. By using pixel normalization, we focus on the relative intensity of pixels, thereby diminishing the impact of lighting variations among different images. + +This normalization process is particularly beneficial in scenarios where the dynamic range of pixel values varies significantly across different patches. By scaling the patches to a common range, pixel-norm mitigates the risk of disproportionate influence from patches with higher intensity values. Consequently, this leads to a more balanced and equitable comparison among patches, enhancing the model's ability to discern and quantify similarities more effectively. + +
MethodIN-1K
w/o P.N.35.5
w/ P.N.36.6
+ +(a) Ablation on pixel normalization (P.N.). + +
kIN-1K
0.536.1
136.3
235.9
+ +(b) Ablation on polynomial coefficient $k$ . + +Table 6. Ablation Study. Table 6a presents an ablation study on the application of pixel normalization when calculating the similarity matrix for clustering. Table 6b explores the effects of varying the polynomial coefficient $k$ , which adjusts the adaptive rate used when combining RGB and embedding features. + +Effect of Features used in Clustering. In Tables 1 and 2, embedding-based methods surpass those dependent solely on RGB data, especially in image-to-text retrieval tasks. One reason for this may be the fact that the embedding model has access to the positional encoding, whereas the RGB-based model solely uses the appearance of each patch. Figure 6 qualitatively shows this advantage: while the RGB-only approach masks extra areas (such as hair or shadows in the first scenario; a laptop and phone in the second) due to color similarities, the embedding-based method masks more complete object parts. + +Ablation on Adaptive Rate. In our approach, we interpolate between using RGB features and the patch embedding layer feature using a coefficient, $\alpha$ , which varies with each epoch. Denoting the current running epoch as $\mathrm{E}_c$ and the number of total training epochs as $\mathrm{E}_t$ , this coefficient is defined as $\alpha = \left(\frac{\mathrm{E}_c}{\mathrm{E}_t}\right)^k$ . In addition to the linear method where $k = 1$ , we explore other polynomial coefficients $k$ for this combination, as summarized in Table 6b. Our findings suggest that the linear combination is most effective, likely due to its smooth transition. + +![](images/6f890ac5bbe7189e0459b33fbd8f0ed0ca1fa0d1000e709c2d97871524b35656.jpg) + +![](images/a7f85c1ab0faae754ef71dce2346d9271fd6ba8d51d634e490103229533caa0b.jpg) +Anchor Patch + +![](images/76aa7a07f50c691bb505c6ecbae409d764261161137570a5e351185e6b817c8b.jpg) + +![](images/bf955542c4782b2e87ed9108e711e2cfd9c2d78f1e1a76652191879ef9246003.jpg) +RGB-based + +![](images/8595e0bf6a2d1eb5fa2dc2b499614faddf8d0e339798f9a342bff11879838a78.jpg) + +![](images/61ccb72de9683aeb0e65a285cc2d1732ef07d7682d11d6b316ced20c19c3e435.jpg) +Embedding-based +Figure 6. Representation used in clustering. We illustrate the distinctions between using pure RGB and ViT embeddings to compute the similarity matrix used in masking patches. + +# 4.4. Limitations + +Our methodology uses a uniform threshold for all images, a strategy that, while effective, may not be the most optimal. Future research could explore the implementation of individualized thresholds for each image, potentially leading to a more intelligent and adaptive masking process. + +All of our approaches use the popular backbone architecture ViT-B/16 [16] and are trained solely on the CC12M dataset [4]. Expanding the scope of the experiments could offer additional insights. + +# 5. Conclusion + +In our study, we introduce a novel cluster based masking strategy designed for vision-language pret-raining. Using either pure RGB values or shallow features from the patch embedding layer, our method effectively clusters image patches, maintaining essential visual semantics. We then randomly mask out these clusters, enabling efficient training. Our approach demonstrates success across various downstream evaluation tasks, including both pure image-based tasks such as image classification and multimodal tasks like image-text retrieval and language composition tests. We believe our work marks a considerable progression in this domain and anticipate that it will stimulate further research into optimizing masking strategies for similar applications. + +Author Contributions All authors contributed to designing projects, launching experiments, and writing the paper. Zixuan Pan focused on algorithm optimization; Zihao Wei focused on code design; Andrew Owens supervised this project, offered feedback, and assisted in writing the paper. + +Acknowledgements This research is supported by a Sony Research Award. We are grateful to Jeongsoo Park, Chao Feng, Yiming Dou, Daniel Geng, Ziyang Chen, and Ayush Shrivastava for their valuable suggestions and discussion. + +# References + +[1] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. Advances in Neural Information Processing Systems, 35:23716-23736, 2022. 2 +[2] Hangbo Bao, Li Dong, Songhao Piao, and Furu Wei. Beit: Bert pre-training of image transformers. arXiv preprint arXiv:2106.08254, 2021. 2 +[3] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9650-9660, 2021. 2 +[4] Soravit Changpinyo, Piyush Sharma, Nan Ding, and Radu Soricut. Conceptual 12m: Pushing web-scale image-text pretraining to recognize long-tail visual concepts. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3558-3568, 2021. 2, 3, 8 +[5] Mark Chen, Alec Radford, Rewon Child, Jeffrey Wu, Heewoo Jun, David Luan, and Ilya Sutskever. Generative pretraining from pixels. In International Conference on Machine Learning, pages 1691-1703. PMLR, 2020. 2 +[6] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International conference on machine learning, pages 1597-1607. PMLR, 2020. 2 +[7] X. Chen, S. Xie, and K. He. An empirical study of training self-supervised vision transformers. In 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 9620-9629, 2021. 2 +[8] Xiaokang Chen, Mingyu Ding, Xiaodi Wang, Ying Xin, Shentong Mo, Yunhao Wang, Shumin Han, Ping Luo, Gang Zeng, and Jingdong Wang. Context autoencoder for self-supervised representation learning. arXiv preprint arXiv:2202.03026, 2022. 2 +[9] Yen-Chun Chen, Linjie Li, Licheng Yu, Ahmed El Kholy, Faisal Ahmed, Zhe Gan, Yu Cheng, and Jingjing Liu. Uniter: Universal image-text representation learning. In European conference on computer vision, pages 104-120. Springer, 2020. 1, 2 +[10] Mehdi Cherti, Romain Beaumont, Ross Wightman, Mitchell Wortsman, Gabriel Ilharco, Cade Gordon, Christoph Schuhmann, Ludwig Schmidt, and Jenia Jitsev. Reproducible scaling laws for contrastive language-image learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2818-2829, 2023. 4 +[11] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE Conference on Computer Vision and Pattern Recognition, pages 248–255, 2009. 2, 4, 6 +[12] Li Deng. The mnist database of handwritten digit images for machine learning research. IEEE Signal Processing Magazine, 29(6):141-142, 2012. 6 + +[13] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. 2018. 3 +[14] Xiaoyi Dong, Jianmin Bao, Ting Zhang, Dongdong Chen, Weiming Zhang, Lu Yuan, Dong Chen, Fang Wen, Nenghai Yu, and Baining Guo. Peco: Perceptual codebook for bert pre-training of vision transformers. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 552-560, 2023. 2 +[15] Xiaoyi Dong, Jianmin Bao, Yinglin Zheng, Ting Zhang, Dongdong Chen, Hao Yang, Ming Zeng, Weiming Zhang, Lu Yuan, Dong Chen, Fang Wen, and Nenghai Yu. Maskclip: Masked self-distillation advances contrastive language-image pretraining, 2023. 1, 2 +[16] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. ICLR, 2021. 3, 8 +[17] Li Fei-Fei, R. Fergus, and P. Perona. One-shot learning of object categories. IEEE Transactions on Pattern Analysis and Machine Intelligence, 28(4):594–611, 2006. 4, 6 +[18] Pedro F Felzenszwalb and Daniel P Huttenlocher. Efficient graph-based image segmentation. International journal of computer vision, 59:167-181, 2004. 1 +[19] Zhanzhou Feng and Shiliang Zhang. Evolved part masking for self-supervised learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10386-10395, 2023. 2 +[20] Tsu-Jui Fu, Linjie Li, Zhe Gan, Kevin Lin, William Yang Wang, Lijuan Wang, and Zicheng Liu. Violet: End-to-end video-language transformers with masked visual-token modeling. arXiv preprint arXiv:2111.12681, 2021. 2 +[21] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9729-9738, 2020. 2 +[22] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollar, and Ross Girshick. Masked autoencoders are scalable vision learners. arXiv preprint arXiv:2111.06377, 2021. 2 +[23] Dan Hendrycks, Steven Basart, Norman Mu, Saurav Kadavath, Frank Wang, Evan Dorundo, Rahul Desai, Tyler Zhu, Samyak Parajuli, Mike Guo, Dawn Song, Jacob Steinhardt, and Justin Gilmer. The many faces of robustness: A critical analysis of out-of-distribution generalization. ICCV, 2021. 4, 6 +[24] Dan Hendrycks, Kevin Zhao, Steven Basart, Jacob Steinhardt, and Dawn Song. Natural adversarial examples. CVPR, 2021. 4, 5, 6 +[25] Cheng-Yu Hsieh, Jieyu Zhang, Zixian Ma, Aniruddha Kembhavi, and Ranjay Krishna. Sugarcrepe: Fixing hackable benchmarks for vision-language compositionality. arXiv preprint arXiv:2306.14610, 2023. 2, 4, 6 +[26] Gabriel Ilharco, Mitchell Wortsman, Ross Wightman, Cade Gordon, Nicholas Carlini, Rohan Taori, Achal Dave, + +Vaishaal Shankar, Hongseok Namkoong, John Miller, Hannaneh Hajishirzi, Ali Farhadi, and Ludwig Schmidt. Openclip, 2021. If you use this software, please cite it as below. 4 +[27] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In International conference on machine learning, pages 4904-4916. PMLR, 2021. 2 +[28] Ioannis Kakogeorgiou, Spyros Gidaris, Bill Psomas, Yannis Avrithis, Andrei Bursuc, Konstantinos Karantzalos, and Nikos Komodakis. What to hide from your students: Attention-guided masked image modeling. In Computer Vision - ECCV 2022, pages 300-318. Springer Nature Switzerland, 2022. 2 +[29] Jacob Devlin Ming-Wei Chang Kenton and Lee Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of naacL-HLT, page 2, 2019. 2 +[30] Alex Krizhevsky. Learning multiple layers of features from tiny images. Technical report, 2009. 4 +[31] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009. 6 +[32] LAION-AI. Clip benchmark. https://github.com/LAION-AI/CLIP_benchmark, 2023.6 +[33] Gang Li, Heliang Zheng, Daqing Liu, Chaoyue Wang, Bing Su, and Changwen Zheng. Semmae: Semantic-guided masking for learning masked autoencoders. Advances in Neural Information Processing Systems, 35:14290-14302, 2022. 1, 2 +[34] Junnan Li, Ramprasaath Selvaraju, Akhilesh Gotmare, Shafiq Joty, Caiming Xiong, and Steven Chu Hong Hoi. Align before fuse: Vision and language representation learning with momentum distillation. Advances in neural information processing systems, 34:9694-9705, 2021. 2 +[35] Xianhang Li, Zeyu Wang, and Cihang Xie. Clipa-v2: Scaling clip training with 81.12 +[36] Yanghao Li, Haoqi Fan, Ronghang Hu, Christoph Feichtenhofer, and Kaiming He. Scaling language-image pre-training via masking, 2023. 1, 2, 4, 6 +[37] Zhaowen Li, Zhiyang Chen, Fan Yang, Wei Li, Yousong Zhu, Chaoyang Zhao, Rui Deng, Liwei Wu, Rui Zhao, Ming Tang, et al. Mst: Masked self-supervised transformer for visual representation. Advances in Neural Information Processing Systems, 34:13165-13176, 2021. 2 +[38] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dollar, and Larry Zitnick. Microsoft coco: Common objects in context. In ECCV. European Conference on Computer Vision, 2014. 2, 4, 5, 6 +[39] Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. Roberta: A robustly optimized bert pretraining approach. arXiv preprint arXiv:1907.11692, 2019. 2 +[40] Stuart Lloyd. Least squares quantization in lcm. IEEE transactions on information theory, 28(2):129-137, 1982. 3 + +[41] Norman Mu, Alexander Kirillov, David Wagner, and Saining Xie. Slip: Self-supervision meets language-image pretraining, 2021. 2 +[42] Maria-Elena Nilsback and Andrew Zisserman. Automated flower classification over a large number of classes. In Proceedings of the 2008 Sixth Indian Conference on Computer Vision, Graphics & Image Processing, page 722-729, USA, 2008. IEEE Computer Society. 4, 6 +[43] Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748, 2018. 3 +[44] OpenAI. Gpt-4 technical report, 2023. 6, 7 +[45] Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, John Schulman, Jacob Hilton, Fraser Kelton, Luke Miller, Maddie Simens, Amanda Askell, Peter Welinder, Paul F Christiano, Jan Leike, and Ryan Lowe. Training language models to follow instructions with human feedback. In Advances in Neural Information Processing Systems, pages 27730-27744. Curran Associates, Inc., 2022. 6, 7 +[46] Adam Paszke, Sam Gross, Soumith Chintala, Gregory Chanan, Edward Yang, Zachary DeVito, Zeming Lin, Alban Desmaison, Luca Antiga, and Adam Lerer. Automatic differentiation in pytorch. 2017. 3 +[47] Zhiliang Peng, Li Dong, Hangbo Bao, Qixiang Ye, and Furu Wei. Beit v2: Masked image modeling with vector-quantized visual tokenizers. arXiv preprint arXiv:2208.06366, 2022. 2 +[48] Alec Radford, Jong Wook Kim, Chris Hallacy, A. Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In ICML, 2021. 4 +[49] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 2, 3, 6 +[50] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 2 +[51] Benjamin Recht, Rebecca Roelofs, Ludwig Schmidt, and Vaishaal Shankar. DoImagenet classifiers generalize to imageNet?, 2019. 4, 6 +[52] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade W Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, Patrick Schramowski, Srivatsa R Kundurthy, Katherine Crowson, Ludwig Schmidt, Robert Kaczmarczyk, and Jenia Jitsev. LAION-5b: An open large-scale dataset for training next generation image-text models. In Thirty-sixth Conference on Neural Information Processing Systems Datasets and Benchmarks Track, 2022. 4 + +[53] Jianbo Shi and Jitendra Malik. Normalized cuts and image segmentation. IEEE Transactions on pattern analysis and machine intelligence, 22(8):888-905, 2000. 1 +[54] Yuge Shi, N Siddharth, Philip Torr, and Adam R Kosiorek. Adversarial masking for self-supervised learning. In International Conference on Machine Learning, pages 20026-20040. PMLR, 2022. 2 +[55] Anugya Srivastava, Shriya Jain, and Mugdha Thigle. Out of distribution detection onImagenet-o, 2022. 4, 5, 6 +[56] Weijie Su, Xizhou Zhu, Yue Cao, Bin Li, Lewei Lu, Furu Wei, and Jifeng Dai. Vl-bert: Pre-training of generic visual-linguistic representations. arXiv preprint arXiv:1908.08530, 2019. 1, 2 +[57] Hao Tan and Mohit Bansal. Lxmert: Learning cross-modality encoder representations from transformers. arXiv preprint arXiv:1908.07490, 2019. 1, 2 +[58] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Hervé Jégou. Training data-efficient image transformers & distillation through attention. In International conference on machine learning, pages 10347-10357. PMLR, 2021. 2 +[59] Aaron Van Den Oord, Oriol Vinyals, et al. Neural discrete representation learning. Advances in neural information processing systems, 30, 2017. 2 +[60] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. 2017. 3 +[61] Andrea Vedaldi. Cats and dogs. In Proceedings of the 2012 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), page 3498-3505, USA, 2012. IEEE Computer Society. 4, 6 +[62] Haohan Wang, Songwei Ge, Zachary Lipton, and Eric P Xing. Learning robust global representations by penalizing local predictive power. In Advances in Neural Information Processing Systems, pages 10506-10518, 2019. 4, 5, 6 +[63] Yiqing Wang, Zihan Li, Jieru Mei, Zihao Wei, Li Liu, Chen Wang, Shengtian Sang, Alan Yuille, Cihang Xie, and Yuyin Zhou. Swinnmm: Masked multi-view with swim transformers for 3d medical image segmentation. In MICCAI, 2023. 2 +[64] Chen Wei, Haoqi Fan, Saining Xie, Chao-Yuan Wu, Alan Yuille, and Christoph Feichtenhofer. Masked feature prediction for self-supervised visual pre-training. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14668-14678, 2022. 2 +[65] Zihao Wei, Chen Wei, Jieru Mei, Zeyu Wang, Xianhang Li, Huiyu Wang, Alan Yuille, Yuyin Zhou, and Cihang Xie. MAE are secretly efficient learners, 2023. 2 +[66] Alex Wilf, Syeda Nahida Akter, Leena Mathur, Paul Pu Liang, Sheryl Mathew, Mengrou Shou, Eric Nyberg, and Louis-Philippe Morency. Difference-masking: Choosing what to mask in continued pretraining. arXiv preprint arXiv:2305.14577, 2023. 2 +[67] Jiahao Xie, Wei Li, Xiaohang Zhan, Ziwei Liu, Yew Soon Ong, and Chen Change Loy. Masked frequency modeling for self-supervised visual pre-training. arXiv preprint arXiv:2206.07706, 2022. 2 + +[68] Zhenda Xie, Zheng Zhang, Yue Cao, Yutong Lin, Jianmin Bao, Zhuliang Yao, Qi Dai, and Han Hu. Simmim: A simple framework for masked image modeling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9653-9663, 2022. 2 +[69] Fengyu Yang, Chao Feng, Ziyang Chen, Hyoungseob Park, Daniel Wang, Yiming Dou, Ziyao Zeng, Xien Chen, Rit Gangopadhyay, Andrew Owens, et al. Binding touch to everything: Learning unified multimodal tactile representations. arXiv preprint arXiv:2401.18084, 2024. 2 +[70] Yifan Yang, Weiquan Huang, Yixuan Wei, Houwen Peng, Xinyang Jiang, Huiqiang Jiang, Fangyun Wei, Yin Wang, Han Hu, Lili Qiu, and Yuqing Yang. Attentive mask clip, 2022. 1, 2, 4, 5, 6 +[71] Peter Young, Alice Lai, Micah Hodosh, and Julia Hockenmaier. From image descriptions to visual denotations: New similarity metrics for semantic inference over event descriptions. Transactions of the Association for Computational Linguistics, 2:67-78, 2014. 4, 5, 6 +[72] Jiahui Yu, Zirui Wang, Vijay Vasudevan, Legg Yeung, Mojtaba Seyedhosseini, and Yonghui Wu. Coca: Contrastive captioners are image-text foundation models. arXiv preprint arXiv:2205.01917, 2022. 2 +[73] Lu Yuan, Dongdong Chen, Yi-Ling Chen, Noel Codella, Xiyang Dai, Jianfeng Gao, Houdong Hu, Xuedong Huang, Boxin Li, Chunyuan Li, et al. Florence: A new foundation model for computer vision. arXiv preprint arXiv:2111.11432, 2021. 2 +[74] Mert Yuksekgonul, Federico Bianchi, Pratyusha Kalluri, Dan Jurafsky, and James Zou. When and why vision-language models behave like bags-of-words, and what to do about it? In The Eleventh International Conference on Learning Representations, 2022. 6 +[75] Jinghao Zhou, Chen Wei, Huiyu Wang, Wei Shen, Cihang Xie, Alan Yuille, and Tao Kong. ibot: Image bert pre-training with online tokenizer. arXiv preprint arXiv:2111.07832, 2021. 2 \ No newline at end of file diff --git a/2024/Efficient Vision-Language Pre-training by Cluster Masking/images.zip b/2024/Efficient Vision-Language Pre-training by Cluster Masking/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..d8cf251d89315edd1c78a7e9d6804753186a9a20 --- /dev/null +++ b/2024/Efficient Vision-Language Pre-training by Cluster Masking/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:00a9b8bdde41e7e9fe19c03885dfa132f38c260f6c183363fca4cb0d1caa87e2 +size 589466 diff --git a/2024/Efficient Vision-Language Pre-training by Cluster Masking/layout.json b/2024/Efficient Vision-Language Pre-training by Cluster Masking/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..3858d2349b3e0594b2dc7ee554bc3cb8d61c7fd2 --- /dev/null +++ b/2024/Efficient Vision-Language Pre-training by Cluster Masking/layout.json @@ -0,0 +1,9395 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 114, + 103, + 480, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 114, + 103, + 480, + 121 + ], + "spans": [ + { + "bbox": [ + 114, + 103, + 480, + 121 + ], + "type": "text", + "content": "Efficient Vision-Language Pre-training by Cluster Masking" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 178, + 143, + 414, + 172 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 178, + 143, + 414, + 172 + ], + "spans": [ + { + "bbox": [ + 178, + 143, + 414, + 172 + ], + "type": "text", + "content": "Zihao Wei* Zixuan Pan* Andrew Owens University of Michigan" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 143, + 199, + 192, + 211 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 199, + 192, + 211 + ], + "spans": [ + { + "bbox": [ + 143, + 199, + 192, + 211 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 224, + 290, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 224, + 290, + 380 + ], + "spans": [ + { + "bbox": [ + 46, + 224, + 290, + 380 + ], + "type": "text", + "content": "We propose a simple strategy for masking image patches during visual-language contrastive learning that improves the quality of the learned representations and the training speed. During each iteration of training, we randomly mask clusters of visually similar image patches, as measured by their raw pixel intensities. This provides an extra learning signal, beyond the contrastive training itself, since it forces a model to predict words for masked visual structures solely from context. It also speeds up training by reducing the amount of data used in each image. We evaluate the effectiveness of our model by pre-training on a number of benchmarks, finding that it outperforms other masking strategies, such as FLIP, on the quality of the learned representation." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 401, + 128, + 414 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 401, + 128, + 414 + ], + "spans": [ + { + "bbox": [ + 47, + 401, + 128, + 414 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 421, + 287, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 421, + 287, + 613 + ], + "spans": [ + { + "bbox": [ + 46, + 421, + 287, + 613 + ], + "type": "text", + "content": "Images contain a great deal of redundant information, making it challenging to efficiently learn representations from them at scale. Recent work has addressed this problem by masking image patches during vision-language contrastive learning [15, 33, 36, 70]. One simple approach is to drop a large fraction of the patches at random, making training more efficient by reducing the computational cost and memory usage in each training iteration [36]. An alternative strategy is to mask sets of semantically related patches [15, 33, 70], such as those that belong to the same object. This forces the learned model to predict words that describes missing scene structures from context, improving the learned representation. However, this approach requires a separate mechanism to group together semantically related patches, which adds considerable complexity to the learning procedure and is computationally expensive." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 613, + 288, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 613, + 288, + 696 + ], + "spans": [ + { + "bbox": [ + 46, + 613, + 288, + 696 + ], + "type": "text", + "content": "We propose a simple masking strategy for multimodal contrastive learning that avoids these shortcomings. During training, we mask random clusters of patches (Fig. 1). For this clustering, we use the patches' raw RGB values as the feature representation. Our approach takes advantage of the fact that simple measures of visual similarity can often capture coherent visual structures, such as object parts [18, 53]," + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 310, + 196, + 395, + 270 + ], + "blocks": [ + { + "bbox": [ + 310, + 196, + 395, + 270 + ], + "lines": [ + { + "bbox": [ + 310, + 196, + 395, + 270 + ], + "spans": [ + { + "bbox": [ + 310, + 196, + 395, + 270 + ], + "type": "image", + "image_path": "359733b6359ec5b4e02180bb3be237448bf176d9397519aab975778441c107b8.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 311, + 270, + 395, + 343 + ], + "blocks": [ + { + "bbox": [ + 311, + 270, + 395, + 343 + ], + "lines": [ + { + "bbox": [ + 311, + 270, + 395, + 343 + ], + "spans": [ + { + "bbox": [ + 311, + 270, + 395, + 343 + ], + "type": "image", + "image_path": "2a783606c85da704f7a2c1320a7e5e9c347d142920272f838a1a6f82847a212f.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 311, + 344, + 395, + 416 + ], + "blocks": [ + { + "bbox": [ + 311, + 344, + 395, + 416 + ], + "lines": [ + { + "bbox": [ + 311, + 344, + 395, + 416 + ], + "spans": [ + { + "bbox": [ + 311, + 344, + 395, + 416 + ], + "type": "image", + "image_path": "5a7767e3e2c64ce15fbba58836cf176d4666fc6fd50671da9645db1ea73e47b7.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 397, + 418, + 468, + 451 + ], + "lines": [ + { + "bbox": [ + 397, + 418, + 468, + 451 + ], + "spans": [ + { + "bbox": [ + 397, + 418, + 468, + 451 + ], + "type": "text", + "content": "(b) A burgundy low-rider show-quality, hot rod Chevy truck." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 396, + 198, + 468, + 270 + ], + "blocks": [ + { + "bbox": [ + 396, + 198, + 468, + 270 + ], + "lines": [ + { + "bbox": [ + 396, + 198, + 468, + 270 + ], + "spans": [ + { + "bbox": [ + 396, + 198, + 468, + 270 + ], + "type": "image", + "image_path": "f106417e32da587fc2930d19502b6f7d9fdc598d5006e26d85da1d3c38b8cb35.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 396, + 270, + 468, + 343 + ], + "blocks": [ + { + "bbox": [ + 396, + 270, + 468, + 343 + ], + "lines": [ + { + "bbox": [ + 396, + 270, + 468, + 343 + ], + "spans": [ + { + "bbox": [ + 396, + 270, + 468, + 343 + ], + "type": "image", + "image_path": "a03b9ce5d8b77083b3c26dd5a910242b0795d1566742e624387e88eb7a3e6ce2.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 396, + 344, + 468, + 416 + ], + "blocks": [ + { + "bbox": [ + 311, + 421, + 394, + 450 + ], + "lines": [ + { + "bbox": [ + 311, + 421, + 394, + 450 + ], + "spans": [ + { + "bbox": [ + 311, + 421, + 394, + 450 + ], + "type": "text", + "content": "(a) A red and gold \npainted fire \nhydrant on the street." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 396, + 344, + 468, + 416 + ], + "lines": [ + { + "bbox": [ + 396, + 344, + 468, + 416 + ], + "spans": [ + { + "bbox": [ + 396, + 344, + 468, + 416 + ], + "type": "image", + "image_path": "dba2b09c764298c9bb464f9f691664adbfb61dc23c0e10abc8f5f41ef30ce8eb.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 470, + 198, + 542, + 270 + ], + "blocks": [ + { + "bbox": [ + 470, + 198, + 542, + 270 + ], + "lines": [ + { + "bbox": [ + 470, + 198, + 542, + 270 + ], + "spans": [ + { + "bbox": [ + 470, + 198, + 542, + 270 + ], + "type": "image", + "image_path": "b37a70da0e179ab31adf7cc3cca7eb91e6366ac96b02c29dd354c1243f163d34.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 470, + 270, + 542, + 343 + ], + "blocks": [ + { + "bbox": [ + 470, + 270, + 542, + 343 + ], + "lines": [ + { + "bbox": [ + 470, + 270, + 542, + 343 + ], + "spans": [ + { + "bbox": [ + 470, + 270, + 542, + 343 + ], + "type": "image", + "image_path": "13423bc72e420b736b579cce42c50d583de23038efda2d74de1536338b6d0661.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 470, + 344, + 542, + 416 + ], + "blocks": [ + { + "bbox": [ + 470, + 344, + 542, + 416 + ], + "lines": [ + { + "bbox": [ + 470, + 344, + 542, + 416 + ], + "spans": [ + { + "bbox": [ + 470, + 344, + 542, + 416 + ], + "type": "image", + "image_path": "3285c2ad9cb1ca563e5525586e7b3e1a1bc9bd417b2b83ffa38dea7527290c77.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 479, + 418, + 539, + 450 + ], + "lines": [ + { + "bbox": [ + 479, + 418, + 539, + 450 + ], + "spans": [ + { + "bbox": [ + 479, + 418, + 539, + 450 + ], + "type": "text", + "content": "(c) Two surf boars on a beach near the water." + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 305, + 456, + 545, + 543 + ], + "lines": [ + { + "bbox": [ + 305, + 456, + 545, + 543 + ], + "spans": [ + { + "bbox": [ + 305, + 456, + 545, + 543 + ], + "type": "text", + "content": "Figure 1. Cluster masking. We mask random clusters of visually similar image patches when training contrastive vision-language models (bottom). This masking strategy distinguishes our approach from methods that independently mask image patches for efficiency [36] (middle), while providing a similar improvement in training speed. It provides an extra learning signal, since it forces a model to predict words for missing scene structures solely from context." + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 555, + 545, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 555, + 545, + 604 + ], + "spans": [ + { + "bbox": [ + 304, + 555, + 545, + 604 + ], + "type": "text", + "content": "especially when clusters are sampled randomly (Fig. 1). Our approach thus leads to more efficient training, like approaches that independently drop patches [36], while improving the learned representation via context prediction." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 304, + 605, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 546, + 713 + ], + "type": "text", + "content": "We take inspiration from masked region classification, a pre-training task widely used in vision-language models [9, 56, 57]. These models extract object features, then predict object labels for the randomly masked out regions. Our masking approach provides a similar training signal, since meaningful labels are included in the image caption. For example, as shown in Figure 1(a), the model is tasked with associating the words \"fire hydrant\" with an image even with the hydrant itself is mostly masked out." + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 702, + 268, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 702, + 268, + 713 + ], + "spans": [ + { + "bbox": [ + 58, + 702, + 268, + 713 + ], + "type": "text", + "content": "*Equal contribution. Author order was determined by a coin flip." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "26815" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 204 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 204 + ], + "type": "text", + "content": "We train our model on Conceptual 12M dataset [4] and evaluate our learned representation on a number of downstream tasks. These tasks include the zero-shot classification and linear probing on ImageNet [11], text and image retrieval on MS-COCO [38], and the SUGARCREPE language composition benchmark [25]. In our experiments, our model outperforms FLIP [36] and CLIP [49] on downstream performance, while having efficiency comparable with FLIP. We also show that the performance can further be improved by using the model's learned feature embedding during clustering." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 213, + 134, + 225 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 213, + 134, + 225 + ], + "spans": [ + { + "bbox": [ + 47, + 213, + 134, + 225 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 233, + 289, + 448 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 233, + 289, + 448 + ], + "spans": [ + { + "bbox": [ + 46, + 233, + 289, + 448 + ], + "type": "text", + "content": "Contrastive Vision-Language Pre-training. Vision-Language Pre-training (VLP) focuses on establishing connections between images or their components and human-interpretable language. This field initially evolved from transferring supervised learning models, which incorporated object detection modules to generate fine-grained visual labels [9, 56, 57]. Subsequently, there was a shift towards large-scale learning using noisy web data, moving away from reliance on fine-grained labels [1, 27, 34, 50, 69, 72, 73]. A significant development in this domain was CLIP [50], which applied contrastive learning techniques [6, 21] to train models to associate correct image-text pairs and dissociate incorrect ones. CLIP scaled contrastive visual-language models significantly beyond previous work, enabling strong feature learning and zero-shot performance. However, further scaling significantly increases the pre-training demands, requiring larger datasets and batch sizes." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 449, + 289, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 449, + 289, + 616 + ], + "spans": [ + { + "bbox": [ + 46, + 449, + 289, + 616 + ], + "type": "text", + "content": "In response to these challenges, recent research has explored incorporating masking into images to reduce training time and allow for more samples per batch [15, 20, 36, 70]. Methods such as MaskClip [15], FLIP [36], and VIOLET [20] have implemented random masking strategies. Yet, it has been noted that random masking may not be as effective on relatively small datasets [35, 70]. To address this, ACLIP [70] introduced a method of masking tokens with low cross-attention scores with text. However, this approach necessitates two forward passes to generate the attention map and requires additional computational modules [41]. In our work, we aim to avoid these limitations, proposing an effective masking method that is based on a patch's raw RGB values." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "type": "text", + "content": "Masked Image Modeling. In the field of language modeling, the effectiveness of models that learn to reconstruct corrupted inputs for generating robust features has been recognized [29, 39]. This approach, known as Mask Language Modeling (MLM), has been adapted in the realm of image processing as Mask Image Modeling (MIM). MIM techniques involve reconstructing either image patches or" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 72, + 547, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 300 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 300 + ], + "type": "text", + "content": "their features [2, 5, 8, 14, 22, 63-65, 68]. The pioneering work in BEIT [2] introduced the reconstruction of discrete tokens, akin to VQ-VAE [59], using block-wise masking. This method demonstrated results on par with contrastive learning and self-distillation methods [3, 7] during model fine-tuning. Later approaches include PeCo's [14] novel visual codebook learning method and BEIT V2's [47] integration of self-distillation methods, using a teacher-student backbone and feature-level KL divergence loss [58]. Further exploration in this field has led to the use of natural image signals as reconstruction targets, moving away from learned features. Examples include SimMIM [68], which reconstructs pure RGB values, MaskFeat [64], introducing reconstruction of the Histogram of Oriented Gradients (HOG) features, and MAE [22], which reconstructs pixel-normalized RGB values. Our work draws inspiration from these studies, particularly in using pixel-normalized RGB values to compute patch similarities, arguing for a more effective distribution of patch features." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 313, + 547, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 313, + 547, + 624 + ], + "spans": [ + { + "bbox": [ + 304, + 313, + 547, + 624 + ], + "type": "text", + "content": "Masking Strategies in MIM. Parallel investigations have focused on masking strategies in MIMs [19, 28, 33, 37, 54, 66, 67]. Early works like BEIT and its successors used block masking, while others such as SimMIM, MaskFeat, and MAE applied random patch-wise masking. Attention-based masking strategies have also been explored, typically using attention maps from vision transformers. MST [37] masks less essential parts with low attention scores, using a reconstruction loss approach. In contrast, AttnMask [28] masks highly attentive patches and applies self-distillation loss. These methods involve simultaneous updates of attention maps and masks during training. A potential limitation of this approach is that insufficiently trained attention maps may not capture structured features effectively. SemMAE [33], starting with iBot features [75], adopts an easy-to-hard masking strategy, starting with masking parts within clusters and gradually expanding to entire clusters. Wilf et al. [66] introduce a unique entity-reinforced language model for masking objects in video frames. However, the reliance on pre-trained features or extracting attention maps can be computationally intensive. Evolved Part Masking [19] proposes using EM algorithm on attention maps to get a clustering before performing SemMAE style masking. Our approach also adopts a cluster based masking strategy in vision-language pre-training, enabling faster pre-training without requiring additional modifications to the model." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 634, + 362, + 645 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 634, + 362, + 645 + ], + "spans": [ + { + "bbox": [ + 306, + 634, + 362, + 645 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 654, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 547, + 715 + ], + "type": "text", + "content": "We propose a cluster based masking strategy for contrastive vision-language pre-training, focusing on masking random clusters with visually similar semantics. Our method selects random anchor patches as cluster centers and computes pairwise patch distances to form clusters. These clus" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "26816" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 53, + 70, + 126, + 143 + ], + "blocks": [ + { + "bbox": [ + 53, + 70, + 126, + 143 + ], + "lines": [ + { + "bbox": [ + 53, + 70, + 126, + 143 + ], + "spans": [ + { + "bbox": [ + 53, + 70, + 126, + 143 + ], + "type": "image", + "image_path": "ff16a277160bab84a9c9d544c6dc9b3baafd381cc5fd6ec67322211670a6659f.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 64, + 146, + 115, + 156 + ], + "lines": [ + { + "bbox": [ + 64, + 146, + 115, + 156 + ], + "spans": [ + { + "bbox": [ + 64, + 146, + 115, + 156 + ], + "type": "text", + "content": "Anchor Patch" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 129, + 70, + 201, + 143 + ], + "blocks": [ + { + "bbox": [ + 129, + 70, + 201, + 143 + ], + "lines": [ + { + "bbox": [ + 129, + 70, + 201, + 143 + ], + "spans": [ + { + "bbox": [ + 129, + 70, + 201, + 143 + ], + "type": "image", + "image_path": "613d61cce0bbf808ae881e59cca840a596aa12f4ccc087eeaa28f2858cd613c0.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 146, + 198, + 156 + ], + "lines": [ + { + "bbox": [ + 131, + 146, + 198, + 156 + ], + "spans": [ + { + "bbox": [ + 131, + 146, + 198, + 156 + ], + "type": "text", + "content": "Distance Heatmap" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 203, + 70, + 276, + 143 + ], + "blocks": [ + { + "bbox": [ + 203, + 70, + 276, + 143 + ], + "lines": [ + { + "bbox": [ + 203, + 70, + 276, + 143 + ], + "spans": [ + { + "bbox": [ + 203, + 70, + 276, + 143 + ], + "type": "image", + "image_path": "9144a605330311de07b51888b727f52f0012e9134aa8ccb6289442802c452134.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 212, + 146, + 267, + 156 + ], + "lines": [ + { + "bbox": [ + 212, + 146, + 267, + 156 + ], + "spans": [ + { + "bbox": [ + 212, + 146, + 267, + 156 + ], + "type": "text", + "content": "Masked Image" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 161, + 287, + 216 + ], + "lines": [ + { + "bbox": [ + 46, + 161, + 287, + 216 + ], + "spans": [ + { + "bbox": [ + 46, + 161, + 287, + 216 + ], + "type": "text", + "content": "Figure 2. Choosing clusters. The process begins by randomly selecting anchor patches from the image. We then calculate the pairwise distances among all patches. Clusters formed within a distance threshold are masked out. We show cluster obtained from a single anchor patch." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 225, + 287, + 284 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 225, + 287, + 284 + ], + "spans": [ + { + "bbox": [ + 46, + 225, + 287, + 284 + ], + "type": "text", + "content": "ters are then masked entirely. To enhance accuracy in cluster formation, we introduce an adaptive layer for refining the distance matrix. Additionally, attention masks and a hard patch cutoff are used to ensure uniform input sizes are consistent in batches for auto differentiation." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 295, + 266, + 308 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 295, + 266, + 308 + ], + "spans": [ + { + "bbox": [ + 47, + 295, + 266, + 308 + ], + "type": "text", + "content": "3.1. Contrastive Vision-Language Pre-training" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 314, + 287, + 390 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 314, + 287, + 390 + ], + "spans": [ + { + "bbox": [ + 46, + 314, + 287, + 390 + ], + "type": "text", + "content": "Our approach builds on contrastive vision-language pretraining methods, such as CLIP [49]. We use contrastive learning to align embeddings of matching text-image pairs and separate those of non-matching pairs. This process is steered by two symmetric InfoNCE losses [43]: the vision-to-language loss " + }, + { + "bbox": [ + 46, + 314, + 287, + 390 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{v\\rightarrow l}" + }, + { + "bbox": [ + 46, + 314, + 287, + 390 + ], + "type": "text", + "content": " and its counterpart, the language-to-vision loss " + }, + { + "bbox": [ + 46, + 314, + 287, + 390 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{l\\rightarrow v}" + }, + { + "bbox": [ + 46, + 314, + 287, + 390 + ], + "type": "text", + "content": ". The vision-to-language loss is defined as:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 88, + 392, + 287, + 418 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 88, + 392, + 287, + 418 + ], + "spans": [ + { + "bbox": [ + 88, + 392, + 287, + 418 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {v \\rightarrow l} = - \\log \\frac {\\exp (\\sin (I , T) / \\tau)}{\\sum_ {j = 1} ^ {N} \\exp (\\sin (I , T _ {j}) / \\tau)}, \\tag {1}", + "image_path": "db5b2f1f4961a1c08fdfa1748c15efd86b8795823d6053292d4342efd488618b.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 420, + 287, + 480 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 420, + 287, + 480 + ], + "spans": [ + { + "bbox": [ + 46, + 420, + 287, + 480 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 420, + 287, + 480 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 46, + 420, + 287, + 480 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 420, + 287, + 480 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 46, + 420, + 287, + 480 + ], + "type": "text", + "content": " are the embeddings for the image and text respectively, sim denotes the similarity function (we use a dot product), and " + }, + { + "bbox": [ + 46, + 420, + 287, + 480 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 46, + 420, + 287, + 480 + ], + "type": "text", + "content": " is a temperature parameter. Similarly, " + }, + { + "bbox": [ + 46, + 420, + 287, + 480 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{l\\to v}" + }, + { + "bbox": [ + 46, + 420, + 287, + 480 + ], + "type": "text", + "content": " is formulated by normalizing the loss using a batch of " + }, + { + "bbox": [ + 46, + 420, + 287, + 480 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 420, + 287, + 480 + ], + "type": "text", + "content": " image examples " + }, + { + "bbox": [ + 46, + 420, + 287, + 480 + ], + "type": "inline_equation", + "content": "\\{I_j\\}_{i = 1}^N" + }, + { + "bbox": [ + 46, + 420, + 287, + 480 + ], + "type": "text", + "content": " instead of text examples " + }, + { + "bbox": [ + 46, + 420, + 287, + 480 + ], + "type": "inline_equation", + "content": "\\{T_j\\}_{i = 1}^N" + }, + { + "bbox": [ + 46, + 420, + 287, + 480 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 47, + 490, + 148, + 502 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 490, + 148, + 502 + ], + "spans": [ + { + "bbox": [ + 47, + 490, + 148, + 502 + ], + "type": "text", + "content": "3.2. Cluster Masking" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 46, + 509, + 287, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 509, + 287, + 628 + ], + "spans": [ + { + "bbox": [ + 46, + 509, + 287, + 628 + ], + "type": "text", + "content": "We introduce a masking strategy that drops out random clusters. While one option would be to use an off-the-shelf clustering method, such as K-Means [40], we choose instead to use a simple and efficient method that results in a random clustering each training iteration (Figure 2). Our approach resembles a single iteration of K-Means, and works by selecting a set of exemplar patches, which each define a cluster. In our experiments, we also evaluate masking clusters obtained using K-Means as an alternative approach." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 46, + 629, + 287, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 629, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 46, + 629, + 287, + 712 + ], + "type": "text", + "content": "We split an input " + }, + { + "bbox": [ + 46, + 629, + 287, + 712 + ], + "type": "inline_equation", + "content": "H\\times W" + }, + { + "bbox": [ + 46, + 629, + 287, + 712 + ], + "type": "text", + "content": " image into patches, following [16]. We then compute the pairwise cosine similarity between every pair of normalized patches, which we use as a distance function " + }, + { + "bbox": [ + 46, + 629, + 287, + 712 + ], + "type": "inline_equation", + "content": "d(\\mathbf{x},\\mathbf{y})" + }, + { + "bbox": [ + 46, + 629, + 287, + 712 + ], + "type": "text", + "content": ". We choose a small subset (less than " + }, + { + "bbox": [ + 46, + 629, + 287, + 712 + ], + "type": "inline_equation", + "content": "5\\%" + }, + { + "bbox": [ + 46, + 629, + 287, + 712 + ], + "type": "text", + "content": ") of these patches at random to act as cluster centers. For each of these selected anchor patches, we define a cluster consisting of patches that lie within a distance" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 72, + 544, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 544, + 96 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 544, + 96 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 305, + 72, + 544, + 96 + ], + "type": "text", + "content": ". The cluster for an exemplar patch " + }, + { + "bbox": [ + 305, + 72, + 544, + 96 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 305, + 72, + 544, + 96 + ], + "type": "text", + "content": " is represented by: " + }, + { + "bbox": [ + 305, + 72, + 544, + 96 + ], + "type": "inline_equation", + "content": "\\mathrm{S_x} = \\{\\mathbf{y} \\mid d(\\mathbf{x}, \\mathbf{y}) \\leq r\\}" + }, + { + "bbox": [ + 305, + 72, + 544, + 96 + ], + "type": "text", + "content": " for image patches." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 97, + 544, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 97, + 544, + 144 + ], + "spans": [ + { + "bbox": [ + 304, + 97, + 544, + 144 + ], + "type": "text", + "content": "All patches within a cluster are masked out. The distance threshold " + }, + { + "bbox": [ + 304, + 97, + 544, + 144 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 304, + 97, + 544, + 144 + ], + "type": "text", + "content": " is automatically searched before training according to an average masking ratio. We provide a simplified pseudocode of the masking strategy in Algorithm 1." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 159, + 545, + 225 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 159, + 545, + 225 + ], + "spans": [ + { + "bbox": [ + 304, + 159, + 545, + 225 + ], + "type": "text", + "content": "Clustering Embedding Features. Another variant of patch feature is the combination of pure RGB values and patch embedding layer features from transformers [16]. When computing similarity scores, we integrate these two measures into a weighted sum, where the weight of each measure is determined by:" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 333, + 228, + 545, + 241 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 228, + 545, + 241 + ], + "spans": [ + { + "bbox": [ + 333, + 228, + 545, + 241 + ], + "type": "interline_equation", + "content": "d (\\mathbf {x}, \\mathbf {y}) = \\alpha \\cdot d _ {r g b} (\\mathbf {x}, \\mathbf {y}) + (1 - \\alpha) \\cdot d _ {e m b} (\\mathbf {x}, \\mathbf {y}), \\tag {2}", + "image_path": "8ca2e84f4b05e023dc877f700671084875fca2972613301f2fed68a6fd8dbdb4.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 243, + 545, + 302 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 243, + 545, + 302 + ], + "spans": [ + { + "bbox": [ + 304, + 243, + 545, + 302 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 243, + 545, + 302 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 304, + 243, + 545, + 302 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 243, + 545, + 302 + ], + "type": "inline_equation", + "content": "\\mathbf{y}" + }, + { + "bbox": [ + 304, + 243, + 545, + 302 + ], + "type": "text", + "content": " represent two patches, " + }, + { + "bbox": [ + 304, + 243, + 545, + 302 + ], + "type": "inline_equation", + "content": "d_{rgb}" + }, + { + "bbox": [ + 304, + 243, + 545, + 302 + ], + "type": "text", + "content": " is the cosine similarity based on pure RGB values, and " + }, + { + "bbox": [ + 304, + 243, + 545, + 302 + ], + "type": "inline_equation", + "content": "d_{emb}" + }, + { + "bbox": [ + 304, + 243, + 545, + 302 + ], + "type": "text", + "content": " is the cosine similarity based on the transformer's embedding features. The weight parameter " + }, + { + "bbox": [ + 304, + 243, + 545, + 302 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 304, + 243, + 545, + 302 + ], + "type": "text", + "content": " linearly increases from 0 to 1 during training." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 303, + 545, + 387 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 303, + 545, + 387 + ], + "spans": [ + { + "bbox": [ + 304, + 303, + 545, + 387 + ], + "type": "text", + "content": "The embedding layer is calculated before the patches enter the transformer, thus we could reuse the patch embeddings in the transformer without computing them twice. Using the embedding layer is advantageous because it incorporates positional encodings [60]. This integration potentially introduces spatial constraints, which we believe can further enhance our masking strategy." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 403, + 545, + 535 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 403, + 545, + 535 + ], + "spans": [ + { + "bbox": [ + 304, + 403, + 545, + 535 + ], + "type": "text", + "content": "Handling Batched Inputs. Deep learning libraries, like PyTorch [46], typically process batched inputs of uniform size. However, in our method, the mask ratio would vary across different images, leading to fluctuations in the number of patches. To accelerate the process, we introduce a minimum mask ratio threshold " + }, + { + "bbox": [ + 304, + 403, + 545, + 535 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 304, + 403, + 545, + 535 + ], + "type": "text", + "content": " for each image. If the calculated mask ratio for an image doesn't meet this predefined threshold, we proceed to randomly drop patches until the desired ratio is achieved. Conversely, for images with patch counts less than the threshold, we use attention masks to avoid masked parts engaging in attention calculation [13]." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 306, + 546, + 387, + 559 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 546, + 387, + 559 + ], + "spans": [ + { + "bbox": [ + 306, + 546, + 387, + 559 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 304, + 567, + 544, + 602 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 567, + 544, + 602 + ], + "spans": [ + { + "bbox": [ + 304, + 567, + 544, + 602 + ], + "type": "text", + "content": "We present a comprehensive evaluation of our proposed algorithm to exhibit the performance, robustness, scalability, and efficiency of our framework." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 306, + 611, + 438, + 623 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 611, + 438, + 623 + ], + "spans": [ + { + "bbox": [ + 306, + 611, + 438, + 623 + ], + "type": "text", + "content": "4.1. Implementation Details" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "text", + "content": "Datasets and Training Details. We train our model using the Conceptual 12M (CC12M) dataset[4], containing 12 million unique image-text pairs, for pre-training our vision-language models. We use ViT-B/16 as backbone for image encoder. The text encoder is a 12-layer transformer, equipped with 8 multi-head attention units and 512-dimensional embeddings. Input images are processed at a" + } + ] + } + ], + "index": 25 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "26817" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "code", + "bbox": [ + 47, + 103, + 282, + 299 + ], + "blocks": [ + { + "bbox": [ + 47, + 72, + 287, + 96 + ], + "lines": [ + { + "bbox": [ + 47, + 72, + 287, + 96 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 287, + 96 + ], + "type": "text", + "content": "Algorithm 1 Pseudocode of cluster based masking in a PyTorch-like style." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "code_caption" + }, + { + "bbox": [ + 47, + 103, + 282, + 299 + ], + "lines": [ + { + "bbox": [ + 47, + 103, + 282, + 299 + ], + "spans": [ + { + "bbox": [ + 47, + 103, + 282, + 299 + ], + "type": "text", + "content": "# img: the image for mask\n# mask_ratio: the ratio for chosing anchor patches\n# r: the threshold for computing cluster\ndef generate_mask(img, mask_ratio, r):\n # Make image into shape (N,L,C).\n # N: number of samples per batch\n # L: number of patches\n # C: size of feature dimension\n img = patchify(img)\n #Normalize each patch\n img = (img - img.mean(dim=-1)) / img.std(dim=-1)\n #Compute pairwise cosine similarity matrix\n x = img / img(norm(dim=-1))\n distance = bmm(x, x.transpose(-2, -1))\n #Generate a boolean masking matrix of shape (N, L)\n init_mask = random_batch_indicies(mask_ratio)\n init_mask = init_mask[:, :, None].float()\n #Get the cluster-based mask candidates = (distance * init_mask) ≥ r\n cluster_mask = candidates.sum(1) ≥ 1\n #Mask both the anchor patches and clusters\n mask = init_mask or cluster_mask\n return mask" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "code_body" + } + ], + "index": 1, + "sub_type": "code", + "guess_lang": "python" + }, + { + "bbox": [ + 46, + 320, + 287, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 320, + 287, + 403 + ], + "spans": [ + { + "bbox": [ + 46, + 320, + 287, + 403 + ], + "type": "text", + "content": "resolution of " + }, + { + "bbox": [ + 46, + 320, + 287, + 403 + ], + "type": "inline_equation", + "content": "224 \\times 224" + }, + { + "bbox": [ + 46, + 320, + 287, + 403 + ], + "type": "text", + "content": ", and text inputs are adjusted to 77 tokens, either by truncation or padding. A class token is transformed into a 512-dimensional feature embedding via a multi-layer perceptron (MLP). For optimization, we use the AdamW optimizer with a learning rate of " + }, + { + "bbox": [ + 46, + 320, + 287, + 403 + ], + "type": "inline_equation", + "content": "5 \\times 10^{-4}" + }, + { + "bbox": [ + 46, + 320, + 287, + 403 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 320, + 287, + 403 + ], + "type": "inline_equation", + "content": "\\beta_{1} = 0.9" + }, + { + "bbox": [ + 46, + 320, + 287, + 403 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 320, + 287, + 403 + ], + "type": "inline_equation", + "content": "\\beta_{2} = 0.98" + }, + { + "bbox": [ + 46, + 320, + 287, + 403 + ], + "type": "text", + "content": ". We use a batch size of 256 per GPU, and train using 8 NVIDIA A40 GPUs." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 405, + 288, + 573 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 405, + 288, + 573 + ], + "spans": [ + { + "bbox": [ + 46, + 405, + 288, + 573 + ], + "type": "text", + "content": "Our method comes in three variants: K-Means, RGB and Embedding. The RGB model clusters based on raw image patches, while the embedding model integrates patch embedding features with RGB for clustering. In the K-Means variant of the model, we mask out half of the clusters randomly. The model constructs 12 clusters and runs for a maximum of 10 iterations. For both RGB and embedding models, we set an average masking ratio of " + }, + { + "bbox": [ + 46, + 405, + 288, + 573 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 46, + 405, + 288, + 573 + ], + "type": "text", + "content": ", following the recommendations of FLIP [36] for optimal masking ratios. In the RGB approach, we use a " + }, + { + "bbox": [ + 46, + 405, + 288, + 573 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 46, + 405, + 288, + 573 + ], + "type": "text", + "content": " cutoff for Ours-" + }, + { + "bbox": [ + 46, + 405, + 288, + 573 + ], + "type": "inline_equation", + "content": "\\mathsf{RGB}_{0.5}" + }, + { + "bbox": [ + 46, + 405, + 288, + 573 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 405, + 288, + 573 + ], + "type": "inline_equation", + "content": "30\\%" + }, + { + "bbox": [ + 46, + 405, + 288, + 573 + ], + "type": "text", + "content": " for Ours-" + }, + { + "bbox": [ + 46, + 405, + 288, + 573 + ], + "type": "inline_equation", + "content": "\\mathsf{RGB}_{0.3}" + }, + { + "bbox": [ + 46, + 405, + 288, + 573 + ], + "type": "text", + "content": ", whereas the Ours-Embedding model uses a " + }, + { + "bbox": [ + 46, + 405, + 288, + 573 + ], + "type": "inline_equation", + "content": "30\\%" + }, + { + "bbox": [ + 46, + 405, + 288, + 573 + ], + "type": "text", + "content": " cutoff. Also, the RGB model selects anchor patches at a " + }, + { + "bbox": [ + 46, + 405, + 288, + 573 + ], + "type": "inline_equation", + "content": "3\\%" + }, + { + "bbox": [ + 46, + 405, + 288, + 573 + ], + "type": "text", + "content": " ratio, compared to a " + }, + { + "bbox": [ + 46, + 405, + 288, + 573 + ], + "type": "inline_equation", + "content": "5\\%" + }, + { + "bbox": [ + 46, + 405, + 288, + 573 + ], + "type": "text", + "content": " ratio in the embedding model." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "text", + "content": "Baselines. In our study, we establish baselines using three models: CLIP, FLIP, and " + }, + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\mathrm{FLIP}_{\\mathrm{Attn}}" + }, + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "text", + "content": ", each trained from scratch on the CC12M dataset. These baseline models are derived from the open-source implementation of CLIP, known as OpenClip [10, 26, 48, 52]. For both FLIP and " + }, + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\mathrm{FLIP}_{\\mathrm{Attn}}" + }, + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "text", + "content": ", we implement a patch dropout ratio of " + }, + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "text", + "content": ". Specifically, FLIP uses a random dropout approach, whereas " + }, + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\mathrm{FLIP}_{\\mathrm{Attn}}" + }, + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "text", + "content": " adopts an attention-based masking strategy inspired by ACLIP [70]. This strategy involves processing the image through the encoder and then averaging" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 72, + 545, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 108 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 108 + ], + "type": "text", + "content": "across attention heads in the final transformer block to determine attention scores. Patches that receive the highest attention in relation to the [CLS] token are retained." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 108, + 545, + 179 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 108, + 545, + 179 + ], + "spans": [ + { + "bbox": [ + 304, + 108, + 545, + 179 + ], + "type": "text", + "content": "To ensure a fair comparison among these methods, we keep the number of patches consistent the same as ours within a single batch, which means for FLIP and " + }, + { + "bbox": [ + 304, + 108, + 545, + 179 + ], + "type": "inline_equation", + "content": "\\mathrm{FLIP}_{\\mathrm{Attn}}" + }, + { + "bbox": [ + 304, + 108, + 545, + 179 + ], + "type": "text", + "content": " we apply a batch size of 256 for one GPU and for CLIP, we use 128 instead. Additionally, we apply a scaling law on learning rate across different models." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 194, + 545, + 337 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 194, + 545, + 337 + ], + "spans": [ + { + "bbox": [ + 304, + 194, + 545, + 337 + ], + "type": "text", + "content": "Evaluation Details. Our models are tested across various benchmarks to ensure its robustness and effectiveness. We conduct zero-shot image-to-text and text-to-image retrieval tasks on COCO [38] and Flickr [71], assessing its performance meticulously. Further more, we evaluate the models' image representation quality by reporting both the zero-shot classification and linear probing performance on three mainstream datasets: ImageNet [11], CIFAR-10, and CIFAR-100 [30]. Zero-shot results of some other datasets like ImageNet variants [23, 24, 51, 55, 62], Caltech101 [17], Flowers [42] and Pets [61], are also reported to verify the method's robustness." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 338, + 545, + 494 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 338, + 545, + 494 + ], + "spans": [ + { + "bbox": [ + 304, + 338, + 545, + 494 + ], + "type": "text", + "content": "For these tasks, our approach adheres strictly to the implementation used in the CLIP benchmark, ensuring consistency and reliability in the evaluation process. Furthermore, we assess the effectiveness of our methodology on language composition tasks using SUGARCREPE [25]. This evaluation aims to determine its adaptability and efficiency across various contexts, including object, attribute, and relation manipulations. Within the SUGARCREPE framework, models are tasked with identifying the correct caption that accurately describes an image, distinguishing it from a closely related but incorrect text hard negatives. The hard negatives is characterized by minor composition differences from the accurate caption." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 501, + 390, + 512 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 501, + 390, + 512 + ], + "spans": [ + { + "bbox": [ + 306, + 501, + 390, + 512 + ], + "type": "text", + "content": "4.2. Main Results" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 519, + 545, + 663 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 519, + 545, + 663 + ], + "spans": [ + { + "bbox": [ + 304, + 519, + 545, + 663 + ], + "type": "text", + "content": "Visualization of Clusters. Figure 3 offers a visual depiction of our cluster based masking technique as outlined in the methodology section. For this illustration, we randomly select a number of image-text pairs from the COCO validation set and apply our masking method to the pure RGB data of the images. The visualization is showing the masking result of the two-stages. In the first stage, a subset of patches (5%) is randomly selected as anchor patches from the pool of all image patches, which are annotated with the red boxes. In the second stage, we visualize the masked clusters that are calculated based on the similarity matrix, where each cluster is represented by a distinct color." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "content": "Zero-shot Retrieval Results. In our investigation into the model's understanding of the relationship between visual and linguistic representations, we conduct zero-shot" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "26818" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 102, + 69, + 492, + 304 + ], + "blocks": [ + { + "bbox": [ + 102, + 69, + 492, + 304 + ], + "lines": [ + { + "bbox": [ + 102, + 69, + 492, + 304 + ], + "spans": [ + { + "bbox": [ + 102, + 69, + 492, + 304 + ], + "type": "image", + "image_path": "ffb62e24ea35d217bcad8b3d1a79f6c5a8dcc52634fd02c98c98a24181891b88.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 307, + 546, + 331 + ], + "lines": [ + { + "bbox": [ + 46, + 307, + 546, + 331 + ], + "spans": [ + { + "bbox": [ + 46, + 307, + 546, + 331 + ], + "type": "text", + "content": "Figure 3. Visualization of cluster masks. Different colors represent distinct clusters formed by the similarity matrix calculated from the chosen anchor patches." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 336, + 288, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 336, + 288, + 396 + ], + "spans": [ + { + "bbox": [ + 46, + 336, + 288, + 396 + ], + "type": "text", + "content": "retrieval tests on several leading retrieval benchmarks. The results, detailed in Table 1, provide insights into the performance of our approach against others, particularly in the context of Image2Text and Text2Image's recall precision at top1(R1), top5(R5) and top10(R10) metrics." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 396, + 287, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 396, + 287, + 540 + ], + "spans": [ + { + "bbox": [ + 46, + 396, + 287, + 540 + ], + "type": "text", + "content": "In the evaluation on the MS-COCO [38], Flickr8k, and Flickr30k [71] datasets, our model outperforms both the baselines in most parts. Notably, in the Image-to-Text tasks, our model performs best in most datasets, with the exception of a slight performance decrease compared to " + }, + { + "bbox": [ + 46, + 396, + 287, + 540 + ], + "type": "inline_equation", + "content": "\\mathrm{FLIP}_{\\mathrm{Attn}}" + }, + { + "bbox": [ + 46, + 396, + 287, + 540 + ], + "type": "text", + "content": " on the MS-COCO dataset. We attribute this success to our training strategy, which prioritizes primary clusters and minimizes the influence of noise. Furthermore, we observe that methods combining RGB information with token embeddings outperform those relying solely on RGB. We hypothesize that this is because the embedding layer, which contains slightly higher-level information." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 540, + 288, + 672 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 540, + 288, + 672 + ], + "spans": [ + { + "bbox": [ + 46, + 540, + 288, + 672 + ], + "type": "text", + "content": "When comparing FLIP to CLIP, FLIP's performance is noticeably weaker, even with large batch sizes. We suspect that FLIP's sub-optimal results in our experimental settings may not fully exploit its strengths. This aligns with findings from other studies, such as Yang et al.'s research on ACLIP [70], which also noted FLIP's limitations. We observe that using attention scores for masking can improve performance compared to purely random masking. However, random masking still falls short of our cluster based masking or even the original CLIP method in some benchmarks." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 689, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 287, + 715 + ], + "type": "text", + "content": "Results on Zero-shot Classification and Linear Probing. We evaluate our model on several widely recognized clas" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 336, + 545, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 336, + 545, + 396 + ], + "spans": [ + { + "bbox": [ + 304, + 336, + 545, + 396 + ], + "type": "text", + "content": "sification benchmarks. The zero-shot classification results are presented in Table 2, while the linear probing results can be found in Table 3. For better evaluating the time spent on training, we normalize all method's training time by the CLIP's training time, which is considered as " + }, + { + "bbox": [ + 304, + 336, + 545, + 396 + ], + "type": "inline_equation", + "content": "1 \\times" + }, + { + "bbox": [ + 304, + 336, + 545, + 396 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 396, + 546, + 504 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 396, + 546, + 504 + ], + "spans": [ + { + "bbox": [ + 304, + 396, + 546, + 504 + ], + "type": "text", + "content": "When comparing our model's performance to CLIP (i.e., no masking), our model demonstrates superior results on the majority of test cases, showcasing an average improvement of " + }, + { + "bbox": [ + 304, + 396, + 546, + 504 + ], + "type": "inline_equation", + "content": "+2.1\\%" + }, + { + "bbox": [ + 304, + 396, + 546, + 504 + ], + "type": "text", + "content": ", with about " + }, + { + "bbox": [ + 304, + 396, + 546, + 504 + ], + "type": "inline_equation", + "content": "+36\\%" + }, + { + "bbox": [ + 304, + 396, + 546, + 504 + ], + "type": "text", + "content": " speeding up. In comparison to the FLIP strategy, which has a similar training duration, our model has an improvement of " + }, + { + "bbox": [ + 304, + 396, + 546, + 504 + ], + "type": "inline_equation", + "content": "+5.5\\%" + }, + { + "bbox": [ + 304, + 396, + 546, + 504 + ], + "type": "text", + "content": ". In comparison to the " + }, + { + "bbox": [ + 304, + 396, + 546, + 504 + ], + "type": "inline_equation", + "content": "\\mathrm{FLIP}_{\\mathrm{Attn}}" + }, + { + "bbox": [ + 304, + 396, + 546, + 504 + ], + "type": "text", + "content": ", our model does not need the attention map for guidance, which gives a much fast training speed, while having a performance of " + }, + { + "bbox": [ + 304, + 396, + 546, + 504 + ], + "type": "inline_equation", + "content": "+2.6\\%" + }, + { + "bbox": [ + 304, + 396, + 546, + 504 + ], + "type": "text", + "content": " on average." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 505, + 546, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 505, + 546, + 624 + ], + "spans": [ + { + "bbox": [ + 304, + 505, + 546, + 624 + ], + "type": "text", + "content": "Out of 12 datasets on the zero-shot classification benchmark, our RGB and embedding model achieves the top performance on 11 of them. In particular, it obtains strong performance on the ImageNet variants: ImageNet-A [24], ImageNet-O [55], ImageNet-R [55], and ImageNet-S [62], which often contain challenging and diverse images. The RGB version of our method also significantly outperforms FLIP and surpasses the CLIP model, especially on ImageNet and its variants, which demonstrates the effectiveness of our method with even the natural guidance." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 624, + 545, + 671 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 624, + 545, + 671 + ], + "spans": [ + { + "bbox": [ + 304, + 624, + 545, + 671 + ], + "type": "text", + "content": "The linear probing results further suggest the effectiveness of our method. Our models achieve " + }, + { + "bbox": [ + 304, + 624, + 545, + 671 + ], + "type": "inline_equation", + "content": "+1.8\\%" + }, + { + "bbox": [ + 304, + 624, + 545, + 671 + ], + "type": "text", + "content": " accuracy on ImageNet, " + }, + { + "bbox": [ + 304, + 624, + 545, + 671 + ], + "type": "inline_equation", + "content": "+3.1\\%" + }, + { + "bbox": [ + 304, + 624, + 545, + 671 + ], + "type": "text", + "content": " on CIFAR-10, and " + }, + { + "bbox": [ + 304, + 624, + 545, + 671 + ], + "type": "inline_equation", + "content": "+4.2\\%" + }, + { + "bbox": [ + 304, + 624, + 545, + 671 + ], + "type": "text", + "content": " on CIFAR-100." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 689, + 545, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 545, + 715 + ], + "type": "text", + "content": "Language Composition. A potential drawback of our method might be understanding compositions of concepts" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "26819" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 50, + 79, + 545, + 165 + ], + "blocks": [ + { + "bbox": [ + 197, + 72, + 236, + 79 + ], + "lines": [ + { + "bbox": [ + 197, + 72, + 236, + 79 + ], + "spans": [ + { + "bbox": [ + 197, + 72, + 236, + 79 + ], + "type": "text", + "content": "text " + }, + { + "bbox": [ + 197, + 72, + 236, + 79 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 197, + 72, + 236, + 79 + ], + "type": "text", + "content": " image" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 414, + 72, + 454, + 79 + ], + "lines": [ + { + "bbox": [ + 414, + 72, + 454, + 79 + ], + "spans": [ + { + "bbox": [ + 414, + 72, + 454, + 79 + ], + "type": "text", + "content": "image " + }, + { + "bbox": [ + 414, + 72, + 454, + 79 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 414, + 72, + 454, + 79 + ], + "type": "text", + "content": " text" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 50, + 79, + 545, + 165 + ], + "lines": [ + { + "bbox": [ + 50, + 79, + 545, + 165 + ], + "spans": [ + { + "bbox": [ + 50, + 79, + 545, + 165 + ], + "type": "table", + "html": "
MS-COCOFlickr8kFlickr30kMS-COCOFlickr8kFlickr30k
R1R5R10R1R5R10R1R5R10R1R5R10R1R5R10R1R5R10
CLIP [49]34.6061.9872.7255.7081.6089.9058.5083.8089.1023.4947.8059.6640.5468.9080.2043.1870.4480.40
FLIP [36]32.6259.1470.6455.0080.9088.9053.8080.8088.5022.5646.0858.0940.3268.1078.6441.5267.9077.46
FLIPAttn [70]33.6660.1871.0253.7080.0987.9955.2981.4087.6023.8948.3360.0440.5868.8878.8643.0670.1078.82
Ours-Kmeans33.6861.1471.9755.1083.3090.9055.4082.0088.9022.6046.9358.8440.1069.8679.9141.3268.5077.58
Ours-RGB0.532.8260.2071.4052.1081.2089.5054.9081.2088.0022.9747.2959.2140.8468.7279.1441.8070.2079.42
Ours-RGB0.335.8761.5072.3454.7081.4090.7057.6083.9090.3023.6547.5459.2540.9068.2279.0042.9269.9679.12
Ours-Embedding34.2661.9673.3057.0082.7090.1055.8084.2089.6023.7748.1859.7642.0069.4079.6443.3070.9280.16
", + "image_path": "a441aa6772f24158658fd26eed403897ad6cc599cd1cb1af2bbcaff7b57c0f9f.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 50, + 196, + 545, + 258 + ], + "blocks": [ + { + "bbox": [ + 47, + 168, + 546, + 191 + ], + "lines": [ + { + "bbox": [ + 47, + 168, + 546, + 191 + ], + "spans": [ + { + "bbox": [ + 47, + 168, + 546, + 191 + ], + "type": "text", + "content": "Table 1. Zero-shot retrieval results. We evaluate on MS-COCO [38], Flickr8k and Flickr30k datasets [71], where the Recall@1 (R1), Recall@5 (R5), and Recall@10 (R10) are reported." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 50, + 196, + 545, + 258 + ], + "lines": [ + { + "bbox": [ + 50, + 196, + 545, + 258 + ], + "spans": [ + { + "bbox": [ + 50, + 196, + 545, + 258 + ], + "type": "table", + "html": "
MethodN.T.IN-1K [11]IN-A [24]IN-O [55]IN-R [23]IN-S [62]INv2 [51]MNIST [12]Cal101 [17]CIFAR-10 [31]CIFAR-100 [31]Flowers [42]Pets [61]Average
CLIP [49]1.00×36.18.038.447.624.930.711.773.557.725.026.053.936.1
FLIP [36]0.53×34.47.139.541.420.129.510.470.452.824.525.346.033.5
FLIPAtm [70]1.06×35.28.139.445.123.730.19.473.561.627.125.751.235.8
Ours-Kmeans0.70×35.57.238.343.022.230.19.769.961.425.325.652.135.0
Ours-RGB0.50.54×36.07.739.845.323.830.511.272.263.627.226.155.036.5
Ours-RGB0.30.64×36.68.839.945.924.931.89.472.363.126.325.457.336.8
Ours-Embedding0.64×36.38.139.647.925.430.711.073.770.732.028.455.438.2
", + "image_path": "2f5a32d315b8107d57d969b86dc3ee02ecd3beb0bf5ff60d8ba7ae28f603fef2.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 67, + 294, + 266, + 385 + ], + "blocks": [ + { + "bbox": [ + 47, + 262, + 546, + 285 + ], + "lines": [ + { + "bbox": [ + 47, + 262, + 546, + 285 + ], + "spans": [ + { + "bbox": [ + 47, + 262, + 546, + 285 + ], + "type": "text", + "content": "Table 2. Zero-shot classification result. We evaluate popular datasets using clip-benchmark [32]. The training time is normalized according to the CLIP's training time. Here N.T. represents normalized time against and IN represents ImageNet." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 67, + 294, + 266, + 385 + ], + "lines": [ + { + "bbox": [ + 67, + 294, + 266, + 385 + ], + "spans": [ + { + "bbox": [ + 67, + 294, + 266, + 385 + ], + "type": "table", + "html": "
MethodCIFAR-10CIFAR-100IN-1K
CLIP88.067.462.3
FLIP85.965.561.3
\\(FLIP_{Attn}\\)86.466.162.0
Ours-Kmeans88.069.162.2
Ours-\\(RGB_{0.5}\\)86.766.262.5
Ours-\\(RGB_{0.3}\\)88.668.763.1
Ours-Embedding89.069.762.7
", + "image_path": "e63bd7224db14a50f70b2f5a6fcd7a8b606e2a3d77515bdf73ab3aab757608b8.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 388, + 288, + 432 + ], + "lines": [ + { + "bbox": [ + 47, + 388, + 288, + 432 + ], + "spans": [ + { + "bbox": [ + 47, + 388, + 288, + 432 + ], + "type": "text", + "content": "Table 3. Linear probing result. All methods are trained for 10 epochs at learning rate of 1e-3. For CIFAR-10 and CIFAR-100, we use a batch size of 64 and for ImageNet-1k the batch size is 1024." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 450, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 450, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 450, + 287, + 713 + ], + "type": "text", + "content": "in language. As we mask out clusters, there is a risk that the model may increasingly adopt bag-of-words tendencies [74], which could impede its ability to learn the relationships between objects. For example, if an image is captioned with \"dog on grass\", the grass may be masked for a large portion in our model as they are highly similar to each other. This will make learning the relation \"on\" difficult. Therefore, we apply SUGARCREPE [25] benchmarks to test the model's ability to understand language compositions. SUGARCREPE benchmarks assess this by generating negative captions through manipulations like adding, swapping, or replacing concepts in sentences, followed by text retrieval tests to evaluate the model's accuracy in selecting the correct answer. From the our test results, which shown in Table 4, our model yields comparable results in Relation tests and demonstrates a significant enhancement in Object and Attribution tests, with an average improvement of " + }, + { + "bbox": [ + 46, + 450, + 287, + 713 + ], + "type": "inline_equation", + "content": "+3.9\\%" + }, + { + "bbox": [ + 46, + 450, + 287, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 450, + 287, + 713 + ], + "type": "inline_equation", + "content": "+3.0\\%" + }, + { + "bbox": [ + 46, + 450, + 287, + 713 + ], + "type": "text", + "content": " respectively, compared to FLIP. This improvement may stem from the masking of entire objects, which simplifies the challenge of contrastive learning by reducing ambiguity. This clarity facilitates the model's learning of relationships, a crucial factor for composition" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 297, + 368, + 309 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 297, + 368, + 309 + ], + "spans": [ + { + "bbox": [ + 306, + 297, + 368, + 309 + ], + "type": "text", + "content": "understanding." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 326, + 547, + 542 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 326, + 547, + 542 + ], + "spans": [ + { + "bbox": [ + 304, + 326, + 547, + 542 + ], + "type": "text", + "content": "Qualitative Comparison of Masking Strategies Our method outperforms the random masking strategy by preserving more semantic content in the unmasked image patches, a comparison showcased in Figure 1. The advantage of our technique is further explored by the captioning experiment detailed in Figure 4, wherein two sets of images, each masked differently, are fed into a captioner, GPT-4 [44, 45]. The captioner is prompted to generate MS-COCO-style captions for the unobscured sections. When comparing these captions to the standard references, it becomes clear that our cluster based masking not only retains key elements but also the interrelations among them. For instance, our approach accurately enables the captioning system to identify an airplane in the first example and to describe the baseball player's action in the second, while the random masking strategy failed to achieve this clarity. These results indicate that our masking method provides a more detailed comprehension of the image." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 551, + 400, + 563 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 551, + 400, + 563 + ], + "spans": [ + { + "bbox": [ + 306, + 551, + 400, + 563 + ], + "type": "text", + "content": "4.3. Ablation Study" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 569, + 547, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 569, + 547, + 676 + ], + "spans": [ + { + "bbox": [ + 304, + 569, + 547, + 676 + ], + "type": "text", + "content": "Ablation on Anchor Patch Ratio. In our study, we conducted an ablation on finding the optimal proportion of patches to serve as anchor patches. The results of this ablation are summarized in Figure 5. We use zero-shot learning results on the ImageNet-1k dataset as the benchmark for assessing the quality of the representations learned by our model. Additionally, we calibrate the threshold for each experiment to ensure that the average final masking ratio was maintained at " + }, + { + "bbox": [ + 304, + 569, + 547, + 676 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 304, + 569, + 547, + 676 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 677, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 547, + 713 + ], + "type": "text", + "content": "Our findings indicate that a smaller proportion of anchor patches tends to yield superior performance. We hypothesize that this improvement is due to the decreased ran" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "26820" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 97, + 70, + 492, + 166 + ], + "blocks": [ + { + "bbox": [ + 97, + 70, + 492, + 166 + ], + "lines": [ + { + "bbox": [ + 97, + 70, + 492, + 166 + ], + "spans": [ + { + "bbox": [ + 97, + 70, + 492, + 166 + ], + "type": "table", + "html": "
REPLACESWAPADDAverage
ObjectAttributeRelationObjectAttributeObjectAttributeObjectAttributeRelation
CLIP85.7779.1864.5161.7858.7174.2468.3573.7168.7564.51
FLIP84.0775.8866.0060.1661.5671.6763.1571.9766.8666.00
\\( FLIP_{Attn-0.5} \\)86.6275.5063.2252.4463.0671.6566.7671.5768.3963.22
\\( FLIP_{Attn-0.3} \\)86.0775.0062.2360.5759.1674.6868.6472.8263.4762.23
Ours-Kmeans84.5076.9063.0962.2061.8671.7765.4673.6667.6063.09
Ours-RGB0.586.8673.4760.2459.3563.3673.3366.7673.1868.4260.24
Ours-RGB0.386.1375.1364.6566.6763.3674.9271.2475.9169.9164.65
", + "image_path": "94becf52994dde290257ee681f5a227d97c3d7b85a40f4ef0f4b6cff543930ab.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 75, + 206, + 257, + 381 + ], + "blocks": [ + { + "bbox": [ + 75, + 206, + 257, + 381 + ], + "lines": [ + { + "bbox": [ + 75, + 206, + 257, + 381 + ], + "spans": [ + { + "bbox": [ + 75, + 206, + 257, + 381 + ], + "type": "image", + "image_path": "3dc13961a49aaf88cf054ec05906767b78d76027460b0e26d4ec608c7a544209.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 387, + 288, + 421 + ], + "lines": [ + { + "bbox": [ + 46, + 387, + 288, + 421 + ], + "spans": [ + { + "bbox": [ + 46, + 387, + 288, + 421 + ], + "type": "text", + "content": "Figure 4. Generated caption from visible patches. We process the masked images through GPT-4 [44, 45] to create captions for the unmasked segments." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 426, + 287, + 499 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 426, + 287, + 499 + ], + "spans": [ + { + "bbox": [ + 46, + 426, + 287, + 499 + ], + "type": "text", + "content": "domness in the selection of anchor patches, which in turn enhances the clustering performance by providing a more stable set of reference points for the model to learn from. However, the masking ratio cannot be too small, as the similarity threshold will become too small, which may reduce the clustering quality." + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 92, + 506, + 246, + 639 + ], + "blocks": [ + { + "bbox": [ + 92, + 506, + 246, + 639 + ], + "lines": [ + { + "bbox": [ + 92, + 506, + 246, + 639 + ], + "spans": [ + { + "bbox": [ + 92, + 506, + 246, + 639 + ], + "type": "image", + "image_path": "bc1e7ca87859ffff6a472a75dd482e1122d51f255d7a45d6e5c1b5e5f6842664.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 642, + 287, + 664 + ], + "lines": [ + { + "bbox": [ + 46, + 642, + 287, + 664 + ], + "spans": [ + { + "bbox": [ + 46, + 642, + 287, + 664 + ], + "type": "text", + "content": "Figure 5. Effect of anchor patch ratio. All the final masking ratio is tuned to be " + }, + { + "bbox": [ + 46, + 642, + 287, + 664 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 46, + 642, + 287, + 664 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 677, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 289, + 715 + ], + "type": "text", + "content": "Ablation on Minimum Mask Ratio. We further demonstrate the capability of our method by setting the minimum mask ratio " + }, + { + "bbox": [ + 46, + 677, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 46, + 677, + 289, + 715 + ], + "type": "text", + "content": " the same as the average masking ratio of FLIP," + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 209, + 545, + 328 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 209, + 545, + 328 + ], + "spans": [ + { + "bbox": [ + 304, + 209, + 545, + 328 + ], + "type": "text", + "content": "as shown in Table 1, 2, and Table 5. For our method, the cutoff ratio denotes the minimum mask ratio applied and the true visible patch ratio is shown in visible ratio. For the FLIP counterpart, it maintains a consistent mask ratio across all images. The results indicate that our method not only matches the speed of FLIP but also surpasses FLIP in zero-shot ImageNet-1K classification accuracy with a " + }, + { + "bbox": [ + 304, + 209, + 545, + 328 + ], + "type": "inline_equation", + "content": "+1.6\\%" + }, + { + "bbox": [ + 304, + 209, + 545, + 328 + ], + "type": "text", + "content": " improvement even by seeing fewer patches. The attention based masking method with less mask ratio achieves similar performance as ours but the speed is much slower." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 329, + 546, + 424 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 329, + 546, + 424 + ], + "spans": [ + { + "bbox": [ + 304, + 329, + 546, + 424 + ], + "type": "text", + "content": "These findings suggest that cluster based masking serves as an effective denoising technique for the dataset. A reason for this enhanced performance is that we could easily mask out typically irrelevant areas, such as uniformly colored backgrounds, which are less informative and oftentimes do not correspond to any word in the caption. This targeted approach enables the model to focus on more meaningful content within the images." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 425, + 545, + 520 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 425, + 545, + 520 + ], + "spans": [ + { + "bbox": [ + 304, + 425, + 545, + 520 + ], + "type": "text", + "content": "Additionally, our findings reveal an improved feature learning by the model when a smaller random masking is applied. By reducing the cutoff ratio from " + }, + { + "bbox": [ + 304, + 425, + 545, + 520 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 304, + 425, + 545, + 520 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 304, + 425, + 545, + 520 + ], + "type": "inline_equation", + "content": "30\\%" + }, + { + "bbox": [ + 304, + 425, + 545, + 520 + ], + "type": "text", + "content": ", we observed a " + }, + { + "bbox": [ + 304, + 425, + 545, + 520 + ], + "type": "inline_equation", + "content": "1\\%" + }, + { + "bbox": [ + 304, + 425, + 545, + 520 + ], + "type": "text", + "content": " enhancement in classification accuracy. Thus, there is some trade-off between the model's performance and speed. Despite this, our model with larger masking cut off still remains significantly faster compared to attention-based masking or the original CLIP method." + } + ] + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 318, + 530, + 531, + 604 + ], + "blocks": [ + { + "bbox": [ + 46, + 168, + 547, + 201 + ], + "lines": [ + { + "bbox": [ + 46, + 168, + 547, + 201 + ], + "spans": [ + { + "bbox": [ + 46, + 168, + 547, + 201 + ], + "type": "text", + "content": "Table 4. Language composition test result. This table presents the performance of models on the SUGARCREPE evaluation, which involves replacing, swapping, or adding atomic concepts such as objects, attributes, and relations in a sentence to create mismatched captions." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 318, + 530, + 531, + 604 + ], + "lines": [ + { + "bbox": [ + 318, + 530, + 531, + 604 + ], + "spans": [ + { + "bbox": [ + 318, + 530, + 531, + 604 + ], + "type": "table", + "html": "
MethodβVisible RatioNormalized TimeIN-1K
FLIP50%50%0.84×34.4
FLIP30%70%1.00×35.4
\\( \\text{FLIP}_{\\text{attn}} \\)50%50%1.73×35.2
\\( \\text{FLIP}_{\\text{attn}} \\)30%70%1.97×36.6
Ours-RGB50%43%0.84×36.0
Ours-RGB30%50%1.00×36.6
", + "image_path": "8875e639aa6eb36016bf7038a7755f2d687a783f3203f36562f53089656ae065.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 606, + 545, + 651 + ], + "lines": [ + { + "bbox": [ + 304, + 606, + 545, + 651 + ], + "spans": [ + { + "bbox": [ + 304, + 606, + 545, + 651 + ], + "type": "text", + "content": "Table 5. Ablation on minimum mask ratio " + }, + { + "bbox": [ + 304, + 606, + 545, + 651 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 304, + 606, + 545, + 651 + ], + "type": "text", + "content": ". Comparison of various methods against different minimum masking ratios. The zero-shot ImageNet-1k classification results are used as metric. The time is normalized to ours RGB model with " + }, + { + "bbox": [ + 304, + 606, + 545, + 651 + ], + "type": "inline_equation", + "content": "\\beta = 30\\%" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 665, + 545, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 545, + 715 + ], + "type": "text", + "content": "Ablation on Pixel Normalization. In our experiments, we incorporate pixel normalization (making each patch mean zero and unit standard deviation 1) into the process of computing the similarity matrix for images. This yields" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "26821" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 144 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 144 + ], + "type": "text", + "content": "a performance improvement of " + }, + { + "bbox": [ + 46, + 72, + 287, + 144 + ], + "type": "inline_equation", + "content": "+1.1\\%" + }, + { + "bbox": [ + 46, + 72, + 287, + 144 + ], + "type": "text", + "content": ", as shown in the results presented in Table 6a. The underlying rationale for this enhancement is attributed to the standardization of image patches. By using pixel normalization, we focus on the relative intensity of pixels, thereby diminishing the impact of lighting variations among different images." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 145, + 288, + 253 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 145, + 288, + 253 + ], + "spans": [ + { + "bbox": [ + 46, + 145, + 288, + 253 + ], + "type": "text", + "content": "This normalization process is particularly beneficial in scenarios where the dynamic range of pixel values varies significantly across different patches. By scaling the patches to a common range, pixel-norm mitigates the risk of disproportionate influence from patches with higher intensity values. Consequently, this leads to a more balanced and equitable comparison among patches, enhancing the model's ability to discern and quantify similarities more effectively." + } + ] + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 62, + 269, + 140, + 312 + ], + "blocks": [ + { + "bbox": [ + 62, + 269, + 140, + 312 + ], + "lines": [ + { + "bbox": [ + 62, + 269, + 140, + 312 + ], + "spans": [ + { + "bbox": [ + 62, + 269, + 140, + 312 + ], + "type": "table", + "html": "
MethodIN-1K
w/o P.N.35.5
w/ P.N.36.6
", + "image_path": "40a72b5747e7c21c6c0ed77b63fb6c10010e1bb956ce4c91530cfab6e24d8c9d.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 47, + 324, + 156, + 343 + ], + "lines": [ + { + "bbox": [ + 47, + 324, + 156, + 343 + ], + "spans": [ + { + "bbox": [ + 47, + 324, + 156, + 343 + ], + "type": "text", + "content": "(a) Ablation on pixel normalization (P.N.)." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 199, + 267, + 261, + 323 + ], + "blocks": [ + { + "bbox": [ + 199, + 267, + 261, + 323 + ], + "lines": [ + { + "bbox": [ + 199, + 267, + 261, + 323 + ], + "spans": [ + { + "bbox": [ + 199, + 267, + 261, + 323 + ], + "type": "table", + "html": "
kIN-1K
0.536.1
136.3
235.9
", + "image_path": "f0557ff8784f130e5391a155aa1993945a474ce7e9356b91c8fe26babfe55aea.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 175, + 325, + 284, + 342 + ], + "lines": [ + { + "bbox": [ + 175, + 325, + 284, + 342 + ], + "spans": [ + { + "bbox": [ + 175, + 325, + 284, + 342 + ], + "type": "text", + "content": "(b) Ablation on polynomial coefficient " + }, + { + "bbox": [ + 175, + 325, + 284, + 342 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 175, + 325, + 284, + 342 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 348, + 287, + 403 + ], + "lines": [ + { + "bbox": [ + 46, + 348, + 287, + 403 + ], + "spans": [ + { + "bbox": [ + 46, + 348, + 287, + 403 + ], + "type": "text", + "content": "Table 6. Ablation Study. Table 6a presents an ablation study on the application of pixel normalization when calculating the similarity matrix for clustering. Table 6b explores the effects of varying the polynomial coefficient " + }, + { + "bbox": [ + 46, + 348, + 287, + 403 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 348, + 287, + 403 + ], + "type": "text", + "content": ", which adjusts the adaptive rate used when combining RGB and embedding features." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 429, + 287, + 561 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 429, + 287, + 561 + ], + "spans": [ + { + "bbox": [ + 46, + 429, + 287, + 561 + ], + "type": "text", + "content": "Effect of Features used in Clustering. In Tables 1 and 2, embedding-based methods surpass those dependent solely on RGB data, especially in image-to-text retrieval tasks. One reason for this may be the fact that the embedding model has access to the positional encoding, whereas the RGB-based model solely uses the appearance of each patch. Figure 6 qualitatively shows this advantage: while the RGB-only approach masks extra areas (such as hair or shadows in the first scenario; a laptop and phone in the second) due to color similarities, the embedding-based method masks more complete object parts." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 586, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 586, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 586, + 287, + 713 + ], + "type": "text", + "content": "Ablation on Adaptive Rate. In our approach, we interpolate between using RGB features and the patch embedding layer feature using a coefficient, " + }, + { + "bbox": [ + 46, + 586, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 46, + 586, + 287, + 713 + ], + "type": "text", + "content": ", which varies with each epoch. Denoting the current running epoch as " + }, + { + "bbox": [ + 46, + 586, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{E}_c" + }, + { + "bbox": [ + 46, + 586, + 287, + 713 + ], + "type": "text", + "content": " and the number of total training epochs as " + }, + { + "bbox": [ + 46, + 586, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{E}_t" + }, + { + "bbox": [ + 46, + 586, + 287, + 713 + ], + "type": "text", + "content": ", this coefficient is defined as " + }, + { + "bbox": [ + 46, + 586, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\alpha = \\left(\\frac{\\mathrm{E}_c}{\\mathrm{E}_t}\\right)^k" + }, + { + "bbox": [ + 46, + 586, + 287, + 713 + ], + "type": "text", + "content": ". In addition to the linear method where " + }, + { + "bbox": [ + 46, + 586, + 287, + 713 + ], + "type": "inline_equation", + "content": "k = 1" + }, + { + "bbox": [ + 46, + 586, + 287, + 713 + ], + "type": "text", + "content": ", we explore other polynomial coefficients " + }, + { + "bbox": [ + 46, + 586, + 287, + 713 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 586, + 287, + 713 + ], + "type": "text", + "content": " for this combination, as summarized in Table 6b. Our findings suggest that the linear combination is most effective, likely due to its smooth transition." + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 326, + 70, + 391, + 134 + ], + "blocks": [ + { + "bbox": [ + 326, + 70, + 391, + 134 + ], + "lines": [ + { + "bbox": [ + 326, + 70, + 391, + 134 + ], + "spans": [ + { + "bbox": [ + 326, + 70, + 391, + 134 + ], + "type": "image", + "image_path": "6f890ac5bbe7189e0459b33fbd8f0ed0ca1fa0d1000e709c2d97871524b35656.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 326, + 135, + 390, + 198 + ], + "blocks": [ + { + "bbox": [ + 326, + 135, + 390, + 198 + ], + "lines": [ + { + "bbox": [ + 326, + 135, + 390, + 198 + ], + "spans": [ + { + "bbox": [ + 326, + 135, + 390, + 198 + ], + "type": "image", + "image_path": "a7f85c1ab0faae754ef71dce2346d9271fd6ba8d51d634e490103229533caa0b.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 333, + 201, + 382, + 209 + ], + "lines": [ + { + "bbox": [ + 333, + 201, + 382, + 209 + ], + "spans": [ + { + "bbox": [ + 333, + 201, + 382, + 209 + ], + "type": "text", + "content": "Anchor Patch" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 392, + 70, + 455, + 134 + ], + "blocks": [ + { + "bbox": [ + 392, + 70, + 455, + 134 + ], + "lines": [ + { + "bbox": [ + 392, + 70, + 455, + 134 + ], + "spans": [ + { + "bbox": [ + 392, + 70, + 455, + 134 + ], + "type": "image", + "image_path": "76aa7a07f50c691bb505c6ecbae409d764261161137570a5e351185e6b817c8b.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 392, + 135, + 455, + 198 + ], + "blocks": [ + { + "bbox": [ + 392, + 135, + 455, + 198 + ], + "lines": [ + { + "bbox": [ + 392, + 135, + 455, + 198 + ], + "spans": [ + { + "bbox": [ + 392, + 135, + 455, + 198 + ], + "type": "image", + "image_path": "bf955542c4782b2e87ed9108e711e2cfd9c2d78f1e1a76652191879ef9246003.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 402, + 201, + 444, + 209 + ], + "lines": [ + { + "bbox": [ + 402, + 201, + 444, + 209 + ], + "spans": [ + { + "bbox": [ + 402, + 201, + 444, + 209 + ], + "type": "text", + "content": "RGB-based" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 458, + 70, + 520, + 134 + ], + "blocks": [ + { + "bbox": [ + 458, + 70, + 520, + 134 + ], + "lines": [ + { + "bbox": [ + 458, + 70, + 520, + 134 + ], + "spans": [ + { + "bbox": [ + 458, + 70, + 520, + 134 + ], + "type": "image", + "image_path": "8595e0bf6a2d1eb5fa2dc2b499614faddf8d0e339798f9a342bff11879838a78.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 458, + 135, + 520, + 198 + ], + "blocks": [ + { + "bbox": [ + 458, + 135, + 520, + 198 + ], + "lines": [ + { + "bbox": [ + 458, + 135, + 520, + 198 + ], + "spans": [ + { + "bbox": [ + 458, + 135, + 520, + 198 + ], + "type": "image", + "image_path": "61ccb72de9683aeb0e65a285cc2d1732ef07d7682d11d6b316ced20c19c3e435.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 457, + 201, + 520, + 210 + ], + "lines": [ + { + "bbox": [ + 457, + 201, + 520, + 210 + ], + "spans": [ + { + "bbox": [ + 457, + 201, + 520, + 210 + ], + "type": "text", + "content": "Embedding-based" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 305, + 215, + 545, + 248 + ], + "lines": [ + { + "bbox": [ + 305, + 215, + 545, + 248 + ], + "spans": [ + { + "bbox": [ + 305, + 215, + 545, + 248 + ], + "type": "text", + "content": "Figure 6. Representation used in clustering. We illustrate the distinctions between using pure RGB and ViT embeddings to compute the similarity matrix used in masking patches." + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 253, + 382, + 264 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 253, + 382, + 264 + ], + "spans": [ + { + "bbox": [ + 306, + 253, + 382, + 264 + ], + "type": "text", + "content": "4.4. Limitations" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 271, + 545, + 331 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 271, + 545, + 331 + ], + "spans": [ + { + "bbox": [ + 304, + 271, + 545, + 331 + ], + "type": "text", + "content": "Our methodology uses a uniform threshold for all images, a strategy that, while effective, may not be the most optimal. Future research could explore the implementation of individualized thresholds for each image, potentially leading to a more intelligent and adaptive masking process." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 332, + 545, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 332, + 545, + 380 + ], + "spans": [ + { + "bbox": [ + 304, + 332, + 545, + 380 + ], + "type": "text", + "content": "All of our approaches use the popular backbone architecture ViT-B/16 [16] and are trained solely on the CC12M dataset [4]. Expanding the scope of the experiments could offer additional insights." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 306, + 390, + 378, + 402 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 390, + 378, + 402 + ], + "spans": [ + { + "bbox": [ + 306, + 390, + 378, + 402 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 304, + 410, + 545, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 410, + 545, + 578 + ], + "spans": [ + { + "bbox": [ + 304, + 410, + 545, + 578 + ], + "type": "text", + "content": "In our study, we introduce a novel cluster based masking strategy designed for vision-language pret-raining. Using either pure RGB values or shallow features from the patch embedding layer, our method effectively clusters image patches, maintaining essential visual semantics. We then randomly mask out these clusters, enabling efficient training. Our approach demonstrates success across various downstream evaluation tasks, including both pure image-based tasks such as image classification and multimodal tasks like image-text retrieval and language composition tests. We believe our work marks a considerable progression in this domain and anticipate that it will stimulate further research into optimizing masking strategies for similar applications." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 304, + 591, + 545, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 591, + 545, + 652 + ], + "spans": [ + { + "bbox": [ + 304, + 591, + 545, + 652 + ], + "type": "text", + "content": "Author Contributions All authors contributed to designing projects, launching experiments, and writing the paper. Zixuan Pan focused on algorithm optimization; Zihao Wei focused on code design; Andrew Owens supervised this project, offered feedback, and assisted in writing the paper." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "content": "Acknowledgements This research is supported by a Sony Research Award. We are grateful to Jeongsoo Park, Chao Feng, Yiming Dou, Daniel Geng, Ziyang Chen, and Ayush Shrivastava for their valuable suggestions and discussion." + } + ] + } + ], + "index": 25 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "26822" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 156 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 156 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 156 + ], + "type": "text", + "content": "[1] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. Advances in Neural Information Processing Systems, 35:23716-23736, 2022. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 159, + 288, + 191 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 159, + 288, + 191 + ], + "spans": [ + { + "bbox": [ + 53, + 159, + 288, + 191 + ], + "type": "text", + "content": "[2] Hangbo Bao, Li Dong, Songhao Piao, and Furu Wei. Beit: Bert pre-training of image transformers. arXiv preprint arXiv:2106.08254, 2021. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 194, + 287, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 194, + 287, + 248 + ], + "spans": [ + { + "bbox": [ + 53, + 194, + 287, + 248 + ], + "type": "text", + "content": "[3] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9650-9660, 2021. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 250, + 287, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 250, + 287, + 315 + ], + "spans": [ + { + "bbox": [ + 53, + 250, + 287, + 315 + ], + "type": "text", + "content": "[4] Soravit Changpinyo, Piyush Sharma, Nan Ding, and Radu Soricut. Conceptual 12m: Pushing web-scale image-text pretraining to recognize long-tail visual concepts. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3558-3568, 2021. 2, 3, 8" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 317, + 287, + 361 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 317, + 287, + 361 + ], + "spans": [ + { + "bbox": [ + 53, + 317, + 287, + 361 + ], + "type": "text", + "content": "[5] Mark Chen, Alec Radford, Rewon Child, Jeffrey Wu, Heewoo Jun, David Luan, and Ilya Sutskever. Generative pretraining from pixels. In International Conference on Machine Learning, pages 1691-1703. PMLR, 2020. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 363, + 287, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 363, + 287, + 407 + ], + "spans": [ + { + "bbox": [ + 53, + 363, + 287, + 407 + ], + "type": "text", + "content": "[6] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International conference on machine learning, pages 1597-1607. PMLR, 2020. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 409, + 287, + 452 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 409, + 287, + 452 + ], + "spans": [ + { + "bbox": [ + 53, + 409, + 287, + 452 + ], + "type": "text", + "content": "[7] X. Chen, S. Xie, and K. He. An empirical study of training self-supervised vision transformers. In 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 9620-9629, 2021. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 454, + 287, + 509 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 454, + 287, + 509 + ], + "spans": [ + { + "bbox": [ + 53, + 454, + 287, + 509 + ], + "type": "text", + "content": "[8] Xiaokang Chen, Mingyu Ding, Xiaodi Wang, Ying Xin, Shentong Mo, Yunhao Wang, Shumin Han, Ping Luo, Gang Zeng, and Jingdong Wang. Context autoencoder for self-supervised representation learning. arXiv preprint arXiv:2202.03026, 2022. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 511, + 287, + 565 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 511, + 287, + 565 + ], + "spans": [ + { + "bbox": [ + 53, + 511, + 287, + 565 + ], + "type": "text", + "content": "[9] Yen-Chun Chen, Linjie Li, Licheng Yu, Ahmed El Kholy, Faisal Ahmed, Zhe Gan, Yu Cheng, and Jingjing Liu. Uniter: Universal image-text representation learning. In European conference on computer vision, pages 104-120. Springer, 2020. 1, 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 567, + 287, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 567, + 287, + 633 + ], + "spans": [ + { + "bbox": [ + 48, + 567, + 287, + 633 + ], + "type": "text", + "content": "[10] Mehdi Cherti, Romain Beaumont, Ross Wightman, Mitchell Wortsman, Gabriel Ilharco, Cade Gordon, Christoph Schuhmann, Ludwig Schmidt, and Jenia Jitsev. Reproducible scaling laws for contrastive language-image learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2818-2829, 2023. 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 635, + 287, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 635, + 287, + 678 + ], + "spans": [ + { + "bbox": [ + 48, + 635, + 287, + 678 + ], + "type": "text", + "content": "[11] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE Conference on Computer Vision and Pattern Recognition, pages 248–255, 2009. 2, 4, 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "type": "text", + "content": "[12] Li Deng. The mnist database of handwritten digit images for machine learning research. IEEE Signal Processing Magazine, 29(6):141-142, 2012. 6" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "text", + "content": "[13] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. 2018. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 108, + 547, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 108, + 547, + 172 + ], + "spans": [ + { + "bbox": [ + 307, + 108, + 547, + 172 + ], + "type": "text", + "content": "[14] Xiaoyi Dong, Jianmin Bao, Ting Zhang, Dongdong Chen, Weiming Zhang, Lu Yuan, Dong Chen, Fang Wen, Nenghai Yu, and Baining Guo. Peco: Perceptual codebook for bert pre-training of vision transformers. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 552-560, 2023. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 175, + 545, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 175, + 545, + 229 + ], + "spans": [ + { + "bbox": [ + 307, + 175, + 545, + 229 + ], + "type": "text", + "content": "[15] Xiaoyi Dong, Jianmin Bao, Yinglin Zheng, Ting Zhang, Dongdong Chen, Hao Yang, Ming Zeng, Weiming Zhang, Lu Yuan, Dong Chen, Fang Wen, and Nenghai Yu. Maskclip: Masked self-distillation advances contrastive language-image pretraining, 2023. 1, 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 231, + 545, + 295 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 231, + 545, + 295 + ], + "spans": [ + { + "bbox": [ + 308, + 231, + 545, + 295 + ], + "type": "text", + "content": "[16] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. ICLR, 2021. 3, 8" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 297, + 545, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 297, + 545, + 330 + ], + "spans": [ + { + "bbox": [ + 308, + 297, + 545, + 330 + ], + "type": "text", + "content": "[17] Li Fei-Fei, R. Fergus, and P. Perona. One-shot learning of object categories. IEEE Transactions on Pattern Analysis and Machine Intelligence, 28(4):594–611, 2006. 4, 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 331, + 547, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 331, + 547, + 364 + ], + "spans": [ + { + "bbox": [ + 308, + 331, + 547, + 364 + ], + "type": "text", + "content": "[18] Pedro F Felzenszwalb and Daniel P Huttenlocher. Efficient graph-based image segmentation. International journal of computer vision, 59:167-181, 2004. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 365, + 545, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 365, + 545, + 409 + ], + "spans": [ + { + "bbox": [ + 308, + 365, + 545, + 409 + ], + "type": "text", + "content": "[19] Zhanzhou Feng and Shiliang Zhang. Evolved part masking for self-supervised learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10386-10395, 2023. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 411, + 545, + 454 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 411, + 545, + 454 + ], + "spans": [ + { + "bbox": [ + 308, + 411, + 545, + 454 + ], + "type": "text", + "content": "[20] Tsu-Jui Fu, Linjie Li, Zhe Gan, Kevin Lin, William Yang Wang, Lijuan Wang, and Zicheng Liu. Violet: End-to-end video-language transformers with masked visual-token modeling. arXiv preprint arXiv:2111.12681, 2021. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 456, + 545, + 509 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 456, + 545, + 509 + ], + "spans": [ + { + "bbox": [ + 308, + 456, + 545, + 509 + ], + "type": "text", + "content": "[21] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9729-9738, 2020. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 511, + 545, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 511, + 545, + 544 + ], + "spans": [ + { + "bbox": [ + 308, + 511, + 545, + 544 + ], + "type": "text", + "content": "[22] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollar, and Ross Girshick. Masked autoencoders are scalable vision learners. arXiv preprint arXiv:2111.06377, 2021. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 545, + 545, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 545, + 545, + 610 + ], + "spans": [ + { + "bbox": [ + 308, + 545, + 545, + 610 + ], + "type": "text", + "content": "[23] Dan Hendrycks, Steven Basart, Norman Mu, Saurav Kadavath, Frank Wang, Evan Dorundo, Rahul Desai, Tyler Zhu, Samyak Parajuli, Mike Guo, Dawn Song, Jacob Steinhardt, and Justin Gilmer. The many faces of robustness: A critical analysis of out-of-distribution generalization. ICCV, 2021. 4, 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 613, + 545, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 613, + 545, + 644 + ], + "spans": [ + { + "bbox": [ + 308, + 613, + 545, + 644 + ], + "type": "text", + "content": "[24] Dan Hendrycks, Kevin Zhao, Steven Basart, Jacob Steinhardt, and Dawn Song. Natural adversarial examples. CVPR, 2021. 4, 5, 6" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 647, + 545, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 647, + 545, + 690 + ], + "spans": [ + { + "bbox": [ + 308, + 647, + 545, + 690 + ], + "type": "text", + "content": "[25] Cheng-Yu Hsieh, Jieyu Zhang, Zixian Ma, Aniruddha Kembhavi, and Ranjay Krishna. Sugarcrepe: Fixing hackable benchmarks for vision-language compositionality. arXiv preprint arXiv:2306.14610, 2023. 2, 4, 6" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 692, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 692, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 692, + 545, + 713 + ], + "type": "text", + "content": "[26] Gabriel Ilharco, Mitchell Wortsman, Ross Wightman, Cade Gordon, Nicholas Carlini, Rohan Taori, Achal Dave," + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "26823" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 66, + 72, + 287, + 115 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 72, + 287, + 115 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 287, + 115 + ], + "type": "text", + "content": "Vaishaal Shankar, Hongseok Namkoong, John Miller, Hannaneh Hajishirzi, Ali Farhadi, and Ludwig Schmidt. Openclip, 2021. If you use this software, please cite it as below. 4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 118, + 288, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 118, + 288, + 183 + ], + "spans": [ + { + "bbox": [ + 48, + 118, + 288, + 183 + ], + "type": "text", + "content": "[27] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In International conference on machine learning, pages 4904-4916. PMLR, 2021. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 185, + 288, + 250 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 185, + 288, + 250 + ], + "spans": [ + { + "bbox": [ + 48, + 185, + 288, + 250 + ], + "type": "text", + "content": "[28] Ioannis Kakogeorgiou, Spyros Gidaris, Bill Psomas, Yannis Avrithis, Andrei Bursuc, Konstantinos Karantzalos, and Nikos Komodakis. What to hide from your students: Attention-guided masked image modeling. In Computer Vision - ECCV 2022, pages 300-318. Springer Nature Switzerland, 2022. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 251, + 288, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 251, + 288, + 296 + ], + "spans": [ + { + "bbox": [ + 48, + 251, + 288, + 296 + ], + "type": "text", + "content": "[29] Jacob Devlin Ming-Wei Chang Kenton and Lee Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of naacL-HLT, page 2, 2019. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 297, + 288, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 297, + 288, + 319 + ], + "spans": [ + { + "bbox": [ + 48, + 297, + 288, + 319 + ], + "type": "text", + "content": "[30] Alex Krizhevsky. Learning multiple layers of features from tiny images. Technical report, 2009. 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 319, + 288, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 319, + 288, + 342 + ], + "spans": [ + { + "bbox": [ + 48, + 319, + 288, + 342 + ], + "type": "text", + "content": "[31] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009. 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 343, + 287, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 343, + 287, + 364 + ], + "spans": [ + { + "bbox": [ + 48, + 343, + 287, + 364 + ], + "type": "text", + "content": "[32] LAION-AI. Clip benchmark. https://github.com/LAION-AI/CLIP_benchmark, 2023.6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 365, + 287, + 419 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 365, + 287, + 419 + ], + "spans": [ + { + "bbox": [ + 48, + 365, + 287, + 419 + ], + "type": "text", + "content": "[33] Gang Li, Heliang Zheng, Daqing Liu, Chaoyue Wang, Bing Su, and Changwen Zheng. Semmae: Semantic-guided masking for learning masked autoencoders. Advances in Neural Information Processing Systems, 35:14290-14302, 2022. 1, 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 422, + 287, + 476 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 422, + 287, + 476 + ], + "spans": [ + { + "bbox": [ + 48, + 422, + 287, + 476 + ], + "type": "text", + "content": "[34] Junnan Li, Ramprasaath Selvaraju, Akhilesh Gotmare, Shafiq Joty, Caiming Xiong, and Steven Chu Hong Hoi. Align before fuse: Vision and language representation learning with momentum distillation. Advances in neural information processing systems, 34:9694-9705, 2021. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 477, + 287, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 477, + 287, + 499 + ], + "spans": [ + { + "bbox": [ + 48, + 477, + 287, + 499 + ], + "type": "text", + "content": "[35] Xianhang Li, Zeyu Wang, and Cihang Xie. Clipa-v2: Scaling clip training with 81.12" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 501, + 287, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 501, + 287, + 533 + ], + "spans": [ + { + "bbox": [ + 48, + 501, + 287, + 533 + ], + "type": "text", + "content": "[36] Yanghao Li, Haoqi Fan, Ronghang Hu, Christoph Feichtenhofer, and Kaiming He. Scaling language-image pre-training via masking, 2023. 1, 2, 4, 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 534, + 287, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 534, + 287, + 589 + ], + "spans": [ + { + "bbox": [ + 48, + 534, + 287, + 589 + ], + "type": "text", + "content": "[37] Zhaowen Li, Zhiyang Chen, Fan Yang, Wei Li, Yousong Zhu, Chaoyang Zhao, Rui Deng, Liwei Wu, Rui Zhao, Ming Tang, et al. Mst: Masked self-supervised transformer for visual representation. Advances in Neural Information Processing Systems, 34:13165-13176, 2021. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 590, + 287, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 590, + 287, + 635 + ], + "spans": [ + { + "bbox": [ + 48, + 590, + 287, + 635 + ], + "type": "text", + "content": "[38] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dollar, and Larry Zitnick. Microsoft coco: Common objects in context. In ECCV. European Conference on Computer Vision, 2014. 2, 4, 5, 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 635, + 287, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 635, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 635, + 287, + 689 + ], + "type": "text", + "content": "[39] Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. Roberta: A robustly optimized bert pretraining approach. arXiv preprint arXiv:1907.11692, 2019. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 690, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 690, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 690, + 287, + 713 + ], + "type": "text", + "content": "[40] Stuart Lloyd. Least squares quantization in lcm. IEEE transactions on information theory, 28(2):129-137, 1982. 3" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "text", + "content": "[41] Norman Mu, Alexander Kirillov, David Wagner, and Saining Xie. Slip: Self-supervision meets language-image pretraining, 2021. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 108, + 547, + 163 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 108, + 547, + 163 + ], + "spans": [ + { + "bbox": [ + 307, + 108, + 547, + 163 + ], + "type": "text", + "content": "[42] Maria-Elena Nilsback and Andrew Zisserman. Automated flower classification over a large number of classes. In Proceedings of the 2008 Sixth Indian Conference on Computer Vision, Graphics & Image Processing, page 722-729, USA, 2008. IEEE Computer Society. 4, 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 164, + 545, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 164, + 545, + 196 + ], + "spans": [ + { + "bbox": [ + 307, + 164, + 545, + 196 + ], + "type": "text", + "content": "[43] Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748, 2018. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 198, + 484, + 209 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 198, + 484, + 209 + ], + "spans": [ + { + "bbox": [ + 307, + 198, + 484, + 209 + ], + "type": "text", + "content": "[44] OpenAI. Gpt-4 technical report, 2023. 6, 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 210, + 545, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 210, + 545, + 308 + ], + "spans": [ + { + "bbox": [ + 308, + 210, + 545, + 308 + ], + "type": "text", + "content": "[45] Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, John Schulman, Jacob Hilton, Fraser Kelton, Luke Miller, Maddie Simens, Amanda Askell, Peter Welinder, Paul F Christiano, Jan Leike, and Ryan Lowe. Training language models to follow instructions with human feedback. In Advances in Neural Information Processing Systems, pages 27730-27744. Curran Associates, Inc., 2022. 6, 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 310, + 545, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 310, + 545, + 354 + ], + "spans": [ + { + "bbox": [ + 308, + 310, + 545, + 354 + ], + "type": "text", + "content": "[46] Adam Paszke, Sam Gross, Soumith Chintala, Gregory Chanan, Edward Yang, Zachary DeVito, Zeming Lin, Alban Desmaison, Luca Antiga, and Adam Lerer. Automatic differentiation in pytorch. 2017. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 355, + 545, + 388 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 355, + 545, + 388 + ], + "spans": [ + { + "bbox": [ + 308, + 355, + 545, + 388 + ], + "type": "text", + "content": "[47] Zhiliang Peng, Li Dong, Hangbo Bao, Qixiang Ye, and Furu Wei. Beit v2: Masked image modeling with vector-quantized visual tokenizers. arXiv preprint arXiv:2208.06366, 2022. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 390, + 545, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 390, + 545, + 445 + ], + "spans": [ + { + "bbox": [ + 308, + 390, + 545, + 445 + ], + "type": "text", + "content": "[48] Alec Radford, Jong Wook Kim, Chris Hallacy, A. Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In ICML, 2021. 4" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 446, + 545, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 446, + 545, + 511 + ], + "spans": [ + { + "bbox": [ + 308, + 446, + 545, + 511 + ], + "type": "text", + "content": "[49] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 2, 3, 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 513, + 545, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 513, + 545, + 578 + ], + "spans": [ + { + "bbox": [ + 308, + 513, + 545, + 578 + ], + "type": "text", + "content": "[50] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 580, + 545, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 580, + 545, + 613 + ], + "spans": [ + { + "bbox": [ + 308, + 580, + 545, + 613 + ], + "type": "text", + "content": "[51] Benjamin Recht, Rebecca Roelofs, Ludwig Schmidt, and Vaishaal Shankar. DoImagenet classifiers generalize to imageNet?, 2019. 4, 6" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 614, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 614, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 614, + 545, + 713 + ], + "type": "text", + "content": "[52] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade W Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, Patrick Schramowski, Srivatsa R Kundurthy, Katherine Crowson, Ludwig Schmidt, Robert Kaczmarczyk, and Jenia Jitsev. LAION-5b: An open large-scale dataset for training next generation image-text models. In Thirty-sixth Conference on Neural Information Processing Systems Datasets and Benchmarks Track, 2022. 4" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "26824" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "type": "text", + "content": "[53] Jianbo Shi and Jitendra Malik. Normalized cuts and image segmentation. IEEE Transactions on pattern analysis and machine intelligence, 22(8):888-905, 2000. 1" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 107, + 287, + 151 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 107, + 287, + 151 + ], + "spans": [ + { + "bbox": [ + 49, + 107, + 287, + 151 + ], + "type": "text", + "content": "[54] Yuge Shi, N Siddharth, Philip Torr, and Adam R Kosiorek. Adversarial masking for self-supervised learning. In International Conference on Machine Learning, pages 20026-20040. PMLR, 2022. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 153, + 288, + 176 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 153, + 288, + 176 + ], + "spans": [ + { + "bbox": [ + 49, + 153, + 288, + 176 + ], + "type": "text", + "content": "[55] Anugya Srivastava, Shriya Jain, and Mugdha Thigle. Out of distribution detection onImagenet-o, 2022. 4, 5, 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 177, + 287, + 220 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 177, + 287, + 220 + ], + "spans": [ + { + "bbox": [ + 49, + 177, + 287, + 220 + ], + "type": "text", + "content": "[56] Weijie Su, Xizhou Zhu, Yue Cao, Bin Li, Lewei Lu, Furu Wei, and Jifeng Dai. Vl-bert: Pre-training of generic visual-linguistic representations. arXiv preprint arXiv:1908.08530, 2019. 1, 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 49, + 223, + 287, + 256 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 223, + 287, + 256 + ], + "spans": [ + { + "bbox": [ + 49, + 223, + 287, + 256 + ], + "type": "text", + "content": "[57] Hao Tan and Mohit Bansal. Lxmert: Learning cross-modality encoder representations from transformers. arXiv preprint arXiv:1908.07490, 2019. 1, 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 49, + 258, + 287, + 312 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 258, + 287, + 312 + ], + "spans": [ + { + "bbox": [ + 49, + 258, + 287, + 312 + ], + "type": "text", + "content": "[58] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Hervé Jégou. Training data-efficient image transformers & distillation through attention. In International conference on machine learning, pages 10347-10357. PMLR, 2021. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 49, + 315, + 287, + 347 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 315, + 287, + 347 + ], + "spans": [ + { + "bbox": [ + 49, + 315, + 287, + 347 + ], + "type": "text", + "content": "[59] Aaron Van Den Oord, Oriol Vinyals, et al. Neural discrete representation learning. Advances in neural information processing systems, 30, 2017. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 49, + 350, + 287, + 382 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 350, + 287, + 382 + ], + "spans": [ + { + "bbox": [ + 49, + 350, + 287, + 382 + ], + "type": "text", + "content": "[60] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. 2017. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 49, + 384, + 287, + 428 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 384, + 287, + 428 + ], + "spans": [ + { + "bbox": [ + 49, + 384, + 287, + 428 + ], + "type": "text", + "content": "[61] Andrea Vedaldi. Cats and dogs. In Proceedings of the 2012 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), page 3498-3505, USA, 2012. IEEE Computer Society. 4, 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 49, + 430, + 287, + 474 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 430, + 287, + 474 + ], + "spans": [ + { + "bbox": [ + 49, + 430, + 287, + 474 + ], + "type": "text", + "content": "[62] Haohan Wang, Songwei Ge, Zachary Lipton, and Eric P Xing. Learning robust global representations by penalizing local predictive power. In Advances in Neural Information Processing Systems, pages 10506-10518, 2019. 4, 5, 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 49, + 475, + 287, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 475, + 287, + 520 + ], + "spans": [ + { + "bbox": [ + 49, + 475, + 287, + 520 + ], + "type": "text", + "content": "[63] Yiqing Wang, Zihan Li, Jieru Mei, Zihao Wei, Li Liu, Chen Wang, Shengtian Sang, Alan Yuille, Cihang Xie, and Yuyin Zhou. Swinnmm: Masked multi-view with swim transformers for 3d medical image segmentation. In MICCAI, 2023. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 49, + 521, + 287, + 576 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 521, + 287, + 576 + ], + "spans": [ + { + "bbox": [ + 49, + 521, + 287, + 576 + ], + "type": "text", + "content": "[64] Chen Wei, Haoqi Fan, Saining Xie, Chao-Yuan Wu, Alan Yuille, and Christoph Feichtenhofer. Masked feature prediction for self-supervised visual pre-training. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14668-14678, 2022. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 49, + 578, + 287, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 578, + 287, + 610 + ], + "spans": [ + { + "bbox": [ + 49, + 578, + 287, + 610 + ], + "type": "text", + "content": "[65] Zihao Wei, Chen Wei, Jieru Mei, Zeyu Wang, Xianhang Li, Huiyu Wang, Alan Yuille, Yuyin Zhou, and Cihang Xie. MAE are secretly efficient learners, 2023. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 49, + 613, + 287, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 613, + 287, + 667 + ], + "spans": [ + { + "bbox": [ + 49, + 613, + 287, + 667 + ], + "type": "text", + "content": "[66] Alex Wilf, Syeda Nahida Akter, Leena Mathur, Paul Pu Liang, Sheryl Mathew, Mengrou Shou, Eric Nyberg, and Louis-Philippe Morency. Difference-masking: Choosing what to mask in continued pretraining. arXiv preprint arXiv:2305.14577, 2023. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 49, + 670, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 670, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 49, + 670, + 287, + 712 + ], + "type": "text", + "content": "[67] Jiahao Xie, Wei Li, Xiaohang Zhan, Ziwei Liu, Yew Soon Ong, and Chen Change Loy. Masked frequency modeling for self-supervised visual pre-training. arXiv preprint arXiv:2206.07706, 2022. 2" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 485 + ], + "type": "list", + "angle": 0, + "index": 24, + "blocks": [ + { + "bbox": [ + 307, + 73, + 547, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 547, + 128 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 547, + 128 + ], + "type": "text", + "content": "[68] Zhenda Xie, Zheng Zhang, Yue Cao, Yutong Lin, Jianmin Bao, Zhuliang Yao, Qi Dai, and Han Hu. Simmim: A simple framework for masked image modeling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9653-9663, 2022. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 129, + 546, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 129, + 546, + 184 + ], + "spans": [ + { + "bbox": [ + 307, + 129, + 546, + 184 + ], + "type": "text", + "content": "[69] Fengyu Yang, Chao Feng, Ziyang Chen, Hyoungseob Park, Daniel Wang, Yiming Dou, Ziyao Zeng, Xien Chen, Rit Gangopadhyay, Andrew Owens, et al. Binding touch to everything: Learning unified multimodal tactile representations. arXiv preprint arXiv:2401.18084, 2024. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 185, + 546, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 185, + 546, + 228 + ], + "spans": [ + { + "bbox": [ + 308, + 185, + 546, + 228 + ], + "type": "text", + "content": "[70] Yifan Yang, Weiquan Huang, Yixuan Wei, Houwen Peng, Xinyang Jiang, Huiqiang Jiang, Fangyun Wei, Yin Wang, Han Hu, Lili Qiu, and Yuqing Yang. Attentive mask clip, 2022. 1, 2, 4, 5, 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 230, + 546, + 284 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 230, + 546, + 284 + ], + "spans": [ + { + "bbox": [ + 308, + 230, + 546, + 284 + ], + "type": "text", + "content": "[71] Peter Young, Alice Lai, Micah Hodosh, and Julia Hockenmaier. From image descriptions to visual denotations: New similarity metrics for semantic inference over event descriptions. Transactions of the Association for Computational Linguistics, 2:67-78, 2014. 4, 5, 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 285, + 546, + 328 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 285, + 546, + 328 + ], + "spans": [ + { + "bbox": [ + 308, + 285, + 546, + 328 + ], + "type": "text", + "content": "[72] Jiahui Yu, Zirui Wang, Vijay Vasudevan, Legg Yeung, Mojtaba Seyedhosseini, and Yonghui Wu. Coca: Contrastive captioners are image-text foundation models. arXiv preprint arXiv:2205.01917, 2022. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 331, + 546, + 384 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 331, + 546, + 384 + ], + "spans": [ + { + "bbox": [ + 308, + 331, + 546, + 384 + ], + "type": "text", + "content": "[73] Lu Yuan, Dongdong Chen, Yi-Ling Chen, Noel Codella, Xiyang Dai, Jianfeng Gao, Houdong Hu, Xuedong Huang, Boxin Li, Chunyuan Li, et al. Florence: A new foundation model for computer vision. arXiv preprint arXiv:2111.11432, 2021. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 386, + 546, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 386, + 546, + 441 + ], + "spans": [ + { + "bbox": [ + 308, + 386, + 546, + 441 + ], + "type": "text", + "content": "[74] Mert Yuksekgonul, Federico Bianchi, Pratyusha Kalluri, Dan Jurafsky, and James Zou. When and why vision-language models behave like bags-of-words, and what to do about it? In The Eleventh International Conference on Learning Representations, 2022. 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 441, + 546, + 485 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 441, + 546, + 485 + ], + "spans": [ + { + "bbox": [ + 308, + 441, + 546, + 485 + ], + "type": "text", + "content": "[75] Jinghao Zhou, Chen Wei, Huiyu Wang, Wei Shen, Cihang Xie, Alan Yuille, and Tao Kong. ibot: Image bert pre-training with online tokenizer. arXiv preprint arXiv:2111.07832, 2021. 2" + } + ] + } + ], + "index": 23 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "26825" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Efficient and Effective Weakly-Supervised Action Segmentation via Action-Transition-Aware Boundary Alignment/1ae3fdc0-77c3-4dc9-87e2-e3b0568d2efa_content_list.json b/2024/Efficient and Effective Weakly-Supervised Action Segmentation via Action-Transition-Aware Boundary Alignment/1ae3fdc0-77c3-4dc9-87e2-e3b0568d2efa_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..6dd77ed9438961e285c25e7082f808ba59e4df77 --- /dev/null +++ b/2024/Efficient and Effective Weakly-Supervised Action Segmentation via Action-Transition-Aware Boundary Alignment/1ae3fdc0-77c3-4dc9-87e2-e3b0568d2efa_content_list.json @@ -0,0 +1,1619 @@ +[ + { + "type": "text", + "text": "Efficient and Effective Weakly-Supervised Action Segmentation via Action-Transition-Aware Boundary Alignment", + "text_level": 1, + "bbox": [ + 145, + 130, + 823, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Angchi Xu $^{1}$ , Wei-Shi Zheng $^{1,2,3,*}$", + "bbox": [ + 344, + 202, + 624, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ School of Computer Science and Engineering, Sun Yat-sen University, China", + "bbox": [ + 176, + 220, + 792, + 238 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ Key Laboratory of Machine Intelligence and Advanced Computing, Ministry of Education, China", + "bbox": [ + 98, + 239, + 872, + 257 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3Guangdong Province Key Laboratory of Information Security Technology, Guangzhou, China", + "bbox": [ + 111, + 257, + 857, + 273 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "xuangch@mail2.sysu.edu.cn, wszheng@ieee.org", + "bbox": [ + 292, + 277, + 676, + 292 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 325, + 313, + 343 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Weakly-supervised action segmentation is a task of learning to partition a long video into several action segments, where training videos are only accompanied by transcripts (ordered list of actions). Most of existing methods need to infer pseudo segmentation for training by serial alignment between all frames and the transcript, which is time-consuming and hard to be parallelized while training. In this work, we aim to escape from this inefficient alignment with massive but redundant frames, and instead to directly localize a few action transitions for pseudo segmentation generation, where a transition refers to the change from an action segment to its next adjacent one in the transcript. As the true transitions are submerged in noisy boundaries due to intra-segment visual variation, we propose a novel Action-Transition-Aware Boundary Alignment (ATBA) framework to efficiently and effectively filter out noisy boundaries and detect transitions. In addition, to boost the semantic learning in the case that noise is inevitably present in the pseudo segmentation, we also introduce video-level losses to utilize the trusted video-level supervision. Extensive experiments show the effectiveness of our approach on both performance and training speed.", + "bbox": [ + 75, + 358, + 473, + 691 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 720, + 209, + 736 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Action segmentation aims to partition a long untrimmed video into several segments and classify each segment into an action category [8, 12, 13, 15, 16, 29, 32, 35, 38, 44]. It is an important yet challenging task for instructional or procedural video understanding. Although fully-supervised action segmentation (FSAS) methods [1, 6, 13, 14, 27] have achieved great progress, they require frame-wise dense an-", + "bbox": [ + 75, + 746, + 468, + 853 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/8009ded21ae803606f935635d485eedc87f6f083ad130d5b86a91cd05b54ebe0.jpg", + "image_caption": [ + "Figure 1. Comparison of performance and training time of WSAS methods on the Breakfast dataset. MoF-The main metric of the task, the higher the better. *-Alignment-free methods. Our ATBA achieves the best performance with a very short training time." + ], + "image_footnote": [], + "bbox": [ + 500, + 325, + 890, + 527 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "notation, which is labor-intensive and time-consuming to collect. As a result, many works [5, 9, 25, 28, 34, 38, 46] explore the weakly-supervised action segmentation (WSAS) only requiring the transcript annotation, which refers to the ordered list of actions occurring in the video without their start and end times. The transcripts are less costly to obtain and can be accessed directly from video narrations or other meta data [2, 9, 18, 28, 30, 33, 47].", + "bbox": [ + 496, + 625, + 892, + 744 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Most of previous WSAS methods have to infer the pseudo segmentation (pseudo frame-wise labels) for training via a sequence alignment process between the video and given transcript, such as Viterbi [22, 23, 25, 28, 33, 34] or Dynamic Time Warping (DTW) [4, 5]. These alignment algorithms are usually designed in a recursive form which needs to be performed serially frame-by-frame and hard to be parallelized, resulting in very slow training process.", + "bbox": [ + 496, + 747, + 893, + 868 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this work, we argue that the frame-by-frame alignment is NOT necessary, since the pseudo segmentation is", + "bbox": [ + 500, + 869, + 893, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Corresponding author.", + "bbox": [ + 94, + 862, + 222, + 875 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "1Code is available at https://github.com/iSEE-", + "bbox": [ + 96, + 875, + 467, + 888 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "Laboratory/CVPR24_ATBA.", + "bbox": [ + 78, + 888, + 256, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "18253", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/2f08921554523255d4b8e94954c42c301787ca2957e1de5eef3b6567101e30a9.jpg", + "image_caption": [ + "Figure 2. The necessity of proposed ATBA. The example is P54-webcam01-P54-coffee in Breakfast dataset. GT-The ground-truth segmentation. C.A.Bdy.-Only class-agnostic boundary detection is applied (Exp.1 of Table 3). Acc.-The accuracy of pseudo segmentation. In the video clip around the \"star\" point, the coffee pot undergoes a change from being picked up to tilted pouring within the segment \"Pour Coffee\", and this noisy visual change is incorrectly detected. In addition, although two boundaries are correctly detected by the \"C.A.Bdy.\" (diamonds), they correspond to incorrect transitions due to one false positive error (star), resulting in complete dislocation of segments within the dashed box. Best viewed in color." + ], + "image_footnote": [], + "bbox": [ + 78, + 89, + 467, + 209 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "fundamentally determined by the locations of a small number of action transitions (i.e., the change from an action segment to its next adjacent action segment in the transcript). Hence, the pseudo segmentation generation can be viewed as a transition detection problem, implying the way to more efficient designs. Intuitively, action transitions are often accompanied by significant visual changes, and there are already many approaches that can detect class-agnostic action boundaries based on these changes [12, 20]. However, due to the intra-segment visual variation and sub-segments under finer granularity, there are numerous noisy boundaries not corresponding to any transitions. Moreover, as the class-agnostic way cannot guarantee correct correspondence between the boundaries and transitions, even slight errors in the detection can result in severe deviation (Fig. 2).", + "bbox": [ + 75, + 415, + 468, + 642 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To overcome the above noisy boundary issue, we propose an efficient and effective framework for WSAS, termed Action-Transition-Aware Boundary Alignment (ATBA), which directly detects the transitions for faster and effective pseudo segmentation generation. To tolerate the noisy boundaries, the ATBA generates more class-agnostic boundaries than the number of transitions as candidates, and then determines a subset from candidates that optimally matches all desired transitions via a drop-allowed alignment algorithm. Furthermore, to fortify the semantic learning under the inevitable noise in pseudo segmentation, we also introduce video-level losses to make use of the trusted video-level supervision. Our ATBA is efficient, because the number of generated candidates will be proportional to the length of the transcript, and therefore the complexity of alignment is now independent of the very long video length. Moreover, other computations", + "bbox": [ + 75, + 643, + 470, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "required by ATBA, i.e., measuring how likely a frame is to be a boundary and how likely a candidate corresponds to a desired transition, are both built on a convolution-like algorithm inspired by [20], which can be parallelized on GPUs efficiently.", + "bbox": [ + 496, + 90, + 890, + 166 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "For inference, we directly adopt the results from the trained frame-wise classifier, without the need for any alignment processing with retrieved or predicted transcript (the ground-truth transcript is not available during inference) like previous WSAS methods [4, 5, 25, 28, 34, 38], which also improves the inference efficiency.", + "bbox": [ + 496, + 167, + 890, + 256 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In summary, our contributions are as follows. (1) We propose to directly localize the action transitions for efficient pseudo segmentation generation during training, without the need of time-consuming frame-by-frame alignment. (2) For robustness to noisy boundaries, we propose a novel ATBA framework to effectively determine boundaries corresponding to each transition. Video-level losses are also introduced to regularize the semantic learning involving the unavoidable noise in the pseudo segmentation. Experiments are conducted on three popular datasets to evaluate our approach: Breakfast [21], Hollywood Extended [2] and CrossTask [47]. Our ATBA achieves state-of-the-art or comparable results with one of the fastest training speed (Fig. 1), demonstrating the effectiveness of ours.", + "bbox": [ + 496, + 257, + 892, + 469 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 482, + 640, + 498 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Weakly-supervised action segmentation methods learn to partition a video into several action segments from training videos only annotated by transcripts [2, 4, 5, 9, 18, 22, 23, 25, 28, 33, 34, 38, 39, 46]. Despite different optimization objectives, most of them generate the pseudo segmentation for training by solving alignment objectives between two sequences (video and transcript) via Connectionist Temporal Classification (CTC) [18], Viterbi [22, 23, 25, 28, 33, 34] or Dynamic Time Warping (DTW) [4, 5].", + "bbox": [ + 496, + 508, + 890, + 643 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Specifically, [18] proposes an extended version of CTC to evaluate all valid alignments between videos and transcripts, which additionally takes the visual similarities of frames into account. Inspired by speech recognition, [22, 23, 33] all use the Hidden Markov Model (HMM) to model the relationship between videos and actions. [9] always generates uniform segmentation but iteratively adjusts the boundaries by inserting repeating actions into the transcript. [34] proposes an alignment objective based on explicit context and length models, which can be solved by Viterbi, and the solved optimal alignment would serve as pseudo labels to train the frame-wise classifier. [25] and [28] both focus on novel learning objectives, but still require the optimal pseudo segmentation produced by Viterbi. [4, 5] both learn from the contrast of aligning the video to the ground-truth transcript and negative transcripts, where the alignment is performed by DTW. Whether using CTC, Viterbi or DTW,", + "bbox": [ + 496, + 643, + 892, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "18254", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/098625b2798000cf20009f8d545e6607bc334dc4819785ae7eacb5cdbf259f1c.jpg", + "image_caption": [ + "Figure 3. The overall framework. We propose an Action-Transition-Aware Boundary Alignment (ATBA) framework, which takes the class-agnostic boundary pattern and action transition pattern together into account to efficiently generate pseudo labels. The trusted video-level supervision is also utilized to further enhance the performance." + ], + "image_footnote": [], + "bbox": [ + 84, + 92, + 890, + 345 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "the above approaches except [9] require frame-by-frame serial calculation, which are inefficient.", + "bbox": [ + 75, + 424, + 467, + 453 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Recently, some efficient methods are proposed with alignment-free design. [38] learns from the mutual consistency between two forms of a segmentation (i.e., framewise classification and category/length pairs). [29] proposes a loss to enforce the output order of any two actions to be consistent with the transcript. In this work, we also propose an efficient framework with different technical roadmap, and our performance is better at the same level of training speed.", + "bbox": [ + 75, + 455, + 468, + 590 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Approach", + "text_level": 1, + "bbox": [ + 76, + 604, + 184, + 622 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Problem Statement", + "text_level": 1, + "bbox": [ + 76, + 630, + 261, + 643 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Action segmentation is a task of partitioning a video into several temporal segments with action labels, which is equivalent to predicting the action categories of each frame. Formally, given a sequence of $T$ $d$ -dimensional frame-wise features $\\mathbf{X} = [\\pmb{x}_1,\\dots,\\pmb{x}_T] \\in \\mathbb{R}^{T\\times d}$ for a video with $T$ frames, the goal is to predict a sequence of actions $\\hat{\\mathcal{Y}} = [\\hat{y}_1,\\dots,\\hat{y}_T]$ , where $\\hat{y}_t \\in \\mathcal{C}$ and $\\mathcal{C} = \\{1,2,\\dots,|\\mathcal{C}|\\}$ is the set of action categories across the dataset (including the background). Under the setting of WSAS, the frame-wise ground-truth $\\mathcal{Y} = [y_1,\\dots,y_T]$ is NOT available during training. Instead, the ordered list of actions called transcript $\\mathcal{A} = [a_1,\\dots,a_M]$ is provided (including the background segments), where $a_{m} \\in \\mathcal{C}$ and $M$ is the total number of action segments in the video. The action transitions of $\\mathcal{A}$", + "bbox": [ + 75, + 652, + 468, + 864 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "are naturally formulated as $\\mathcal{R} = \\{(a_r,a_{r + 1})\\}_{r = 1}^{M - 1}$", + "bbox": [ + 500, + 422, + 836, + 441 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Overview", + "text_level": 1, + "bbox": [ + 500, + 452, + 609, + 465 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our proposed framework is illustrated in Fig. 3. At first, the input sequence $\\mathbf{X}$ is further encoded by a temporal network to generate more task-relevant representations $\\mathbf{X}' \\in \\mathbb{R}^{T \\times d'}$ (Sec. 3.3), then a classifier shared along the temporal axis followed by a category softmax activation will predict the frame-wise class probabilities $\\mathbf{P} = [\\mathbf{p}_1, \\dots, \\mathbf{p}_T] \\in \\mathbb{R}^{T \\times |C|}$ from $\\mathbf{X}'$ . After that, the Action-Transition-Aware Boundary Alignment (ATBA) module takes $\\mathbf{P}$ and the transcript $\\mathcal{A}$ as input to infer the pseudo frame-wise labels $\\widetilde{\\mathcal{Y}} = [\\widetilde{y}_1, \\dots, \\widetilde{y}_T]$ , where $\\widetilde{y}_t \\in C$ (Sec. 3.4). Finally, $\\widetilde{\\mathcal{Y}}$ is used back to supervise $\\mathbf{P}$ by a standard cross entropy:", + "bbox": [ + 496, + 476, + 890, + 643 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {c l s}} = - \\frac {1}{T} \\sum_ {t = 1} ^ {T} \\sum_ {c = 1} ^ {| c |} \\mathbb {I} (\\widetilde {y} _ {t} = c) \\log \\boldsymbol {P} _ {t, c}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 566, + 656, + 890, + 699 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mathbb{I}(\\cdot)$ is the indicator function which returns 1 if the condition is satisfied and 0 otherwise. Note that the pseudo frame-wise labels are inferred at each iteration for a batch of data and used immediately for current training at the same iteration. The additional video-level losses are stated in Sec. 3.5, and the training and inference processes are described in Sec. 3.6.", + "bbox": [ + 496, + 712, + 890, + 816 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.3. Temporal Network", + "text_level": 1, + "bbox": [ + 500, + 830, + 679, + 845 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We employ a slightly modified pre-norm [42] Transformer [40] encoder with learnable positional embeddings as the temporal network for feature learning. Following [11, 45],", + "bbox": [ + 496, + 854, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_footnote", + "text": "2POC [29] is a set-supervised method but can be extended to transcript supervision naturally, so we cite its corresponding results for comparison.", + "bbox": [ + 76, + 875, + 467, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "18255", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/44df138bd8e2149a40074b7c9f1f5ee9a54d566ce63806497cc63237ce56f3b2.jpg", + "image_caption": [ + "(a)" + ], + "image_footnote": [], + "bbox": [ + 117, + 88, + 256, + 194 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/d0e9d0643e18099af8f651bbd6efe538db1c372cd8e8b0c0965054bb73432cc9.jpg", + "image_caption": [ + "Figure 4. (a) A $7 \\times 7$ template for class-agnostic boundary scoring. \n(b) A $2 \\times 7$ template for action transition scoring." + ], + "image_footnote": [], + "bbox": [ + 292, + 117, + 429, + 151 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "the vanilla full self-attention is replaced with a pyramid hierarchical local attention to better adapt to the action segmentation task (see supplemental material for more details).", + "bbox": [ + 76, + 273, + 467, + 319 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.4. Action-Transition-Aware Boundary Alignment", + "text_level": 1, + "bbox": [ + 76, + 327, + 467, + 343 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The ATBA is the core component of our framework. It generates the pseudo frame-wise labels $\\widetilde{\\mathcal{V}}$ by inferring the boundaries corresponding to $M - 1$ action transitions from the predicted class probabilities $\\pmb{P}$ . Once they are found, $\\widetilde{\\mathcal{V}}$ can be naturally generated by assigning the action labels of $\\mathcal{A}$ one-by-one into the intervals between these boundaries.", + "bbox": [ + 76, + 349, + 468, + 440 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Briefly, the ATBA first generates a set of candidate boundaries via a class-agnostic way, and then finds $M - 1$ points from this set that optimally match all transitions of $\\mathcal{A}$ via a dynamic programming (DP) algorithm. We describe the details in the following.", + "bbox": [ + 75, + 441, + 468, + 516 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "- Class-Agnostic Boundary Scoring. Firstly, the ATBA calculates the class-agnostic boundary scores $\\mathcal{V}^{\\mathrm{b}} = [v_1^{\\mathrm{b}},\\dots,v_T^{\\mathrm{b}}]$ for each timestamp. We employ a pattern-matching-based scoring method proposed by a generic event boundary detection approach termed UBoCo [20], which considers the pattern of each frame's neighborhood. Specifically, for timestamp $t$ , a pairwise similarity matrix $\\Gamma^{(t)}\\in \\mathbb{R}^{w^{\\mathrm{b}}\\times w^{\\mathrm{b}}}$ is calculated within a local window with size $w^{\\mathrm{b}}$ centered at $t$ , whose $(i,j)$ -entry represents the class-agnostic similarity between $i$ -th and $j$ -th frames inside the window, with values ranging from -1 to 1 (see the computational details in supplementary material). Clearly, if $t$ is a boundary, the frame feature should change dramatically at $t$ and keep stable elsewhere, so its $\\Gamma^{(t)}$ should show the pattern of that the values in the upper-left and lower-right areas are close to 1 and otherwise close to -1. Hence a template $\\Omega^{\\mathrm{b}}\\in \\mathbb{R}^{w^{\\mathrm{b}}\\times w^{\\mathrm{b}}}$ like Fig. 4(a) is designed to capture this pattern and output the class-agnostic boundary score for each $t$ by a correlation operation:", + "bbox": [ + 76, + 517, + 468, + 806 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nv _ {t} ^ {\\mathrm {b}} = \\frac {1}{w ^ {\\mathrm {b}} \\times w ^ {\\mathrm {b}}} \\sum_ {i = 1} ^ {w ^ {\\mathrm {b}}} \\sum_ {j = 1} ^ {w ^ {\\mathrm {b}}} \\Omega_ {i, j} ^ {\\mathrm {b}} \\Gamma_ {i, j} ^ {(t)}. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 161, + 816, + 468, + 859 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "- Candidate Boundary Selection. After calculating $\\nu^{\\mathrm{b}}$ , we select a set of $K$ candidate boundaries $\\tilde{\\mathcal{B}} = \\{b_k\\}_{k=1}^K$ , where", + "bbox": [ + 76, + 869, + 468, + 901 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/7fa9cab4368f949836b38db60133b5627548cd5a104037341dbdf1c7b6ee3418.jpg", + "image_caption": [ + "(b)", + "Figure 5. Illustration of the alignment between action transitions $\\mathcal{R}$ and candidate boundaries $\\tilde{\\mathcal{B}}$ . Blue circles are aligned and gray ones are dropped. (a) A valid alignment. (b) An invalid alignment. The red dashed arrow violates the ordering consistency. Best viewed in color." + ], + "image_footnote": [], + "bbox": [ + 540, + 89, + 856, + 174 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "$1 < b_{1} < \\ldots < b_{K} \\leq T$ and $K > M - 1$ . The selection is performed by a simple greedy strategy with non maximum suppression inspired by [12], i.e., each time we select one timestamp with current highest score $v_{t}^{\\mathrm{b}}$ , and invalidate its neighborhood to avoid selecting multiple timestamps corresponding to one same boundary, until the number of selected timestamps reach an upper bound or all remaining timestamps are invalid. The radius of the invalid interval is set adaptively to $\\frac{\\mu T}{M}$ , where $\\mu \\in [0,1]$ is a hyper-parameter, and in practice, the upper bound for $K$ is set to $\\lambda(M - 1)$ , where $\\lambda \\in \\mathbb{N}_{+}$ is also a hyper-parameter.", + "bbox": [ + 496, + 281, + 890, + 446 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "- Action Transition Scoring. As mentioned in Sec. 1, the class-agnostic scores $\\mathcal{V}^{\\mathrm{b}}$ are not enough to detect action transitions. To this end, we then calculate an action transition score matrix $V^{\\mathrm{a}}\\in \\mathbb{R}^{K\\times (M - 1)}$ , where $V_{k,r}^{\\mathrm{a}}$ measures the possibility that the $k$ -th candidate boundary corresponds to the $r$ -th transition, i.e., separates the $r$ -th and $(r + 1)$ -th action segments. This matrix is also calculated via pattern matching. Clearly, if candidate $b_{k}$ corresponds to the $r$ -th transition, the classifier's activation for class $a_{r}$ should drop sharply after $b_{k}$ , while rise for class $a_{r + 1}$ . Hence, a template $\\Omega^{\\mathrm{a}}\\in \\mathbb{R}^{2\\times w^{\\mathrm{a}}}$ with temporal size $w^{\\mathrm{a}}$ like Fig. 4(b) is employed to detect this pattern around $b_{k}$ :", + "bbox": [ + 496, + 448, + 890, + 628 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {V} _ {k, r} ^ {\\mathrm {a}} = \\frac {1}{2 w ^ {\\mathrm {a}}} \\sum_ {i = 1} ^ {2} \\sum_ {j = 1} ^ {w ^ {\\mathrm {a}}} \\boldsymbol {\\Omega} _ {i, j} ^ {\\mathrm {a}} \\boldsymbol {P} _ {\\text {i n d} ^ {\\mathrm {a}} \\left(b _ {k}, j\\right), a _ {r + i - 1}}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 547, + 640, + 890, + 683 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {i n d} ^ {\\mathrm {a}} \\left(b _ {k}, j\\right) = b _ {k} - \\left\\lfloor \\frac {w ^ {\\mathrm {a}}}{2} \\right\\rfloor + j - 1,\n$$\n", + "text_format": "latex", + "bbox": [ + 581, + 686, + 808, + 715 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathrm{ind}^{\\mathrm{a}}(b_k,j)$ is the index transform from the index $j$ of the local window centered at $b_{k}$ to the global timestamp index. Finally, the class-agnostic scores $\\mathcal{V}^{\\mathrm{b}}$ are added back to $\\pmb{V}^{\\mathrm{a}}$ to produce the final score matrix $\\pmb {V}\\in \\mathbb{R}^{K\\times (M - 1)}$ :", + "bbox": [ + 496, + 724, + 890, + 785 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {V} _ {k, r} = \\boldsymbol {V} _ {k, r} ^ {\\mathrm {a}} + v _ {b _ {k}} ^ {\\mathrm {b}}. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 625, + 796, + 890, + 814 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The above equation means that the class-agnostic boundary score of the $k$ -th boundary $v_{b_k}^{\\mathrm{b}}$ is added to all transition scores corresponding to the $k$ -th boundary $\\mathbf{V}_{k,r}^{\\mathrm{a}}, \\forall r$ .", + "bbox": [ + 496, + 825, + 890, + 869 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "- Action Transition Alignment. The final step is to find exact $M - 1$ optimal boundaries from the candidates $\\widetilde{B}$ based", + "bbox": [ + 496, + 869, + 890, + 898 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "18256", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/fb7848cf7d882894564cc54bbca57f5d3b3fa5e915f0363ff36afdc90de45a4f.jpg", + "image_caption": [ + "Figure 6. An example of action transition alignment between $\\widetilde{\\mathcal{B}}$ of length 5 and $\\mathcal{R}$ of length 3. (a) The cost matrix $\\Delta$ . (b) The cumulative cost matrix $D$ . Blue areas are directly initialized. Yellow areas correspond to optimal alignment. Red arrows are allowed path directions. (c) The optimal alignment. Best viewed in color." + ], + "image_footnote": [], + "bbox": [ + 76, + 88, + 468, + 166 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "on the score matrix $V$ . It is equivalent to seek an one-to-one alignment with lowest cost between the transitions $\\mathcal{R}$ and candidates $\\widetilde{\\mathcal{B}}$ while requiring $K - M + 1$ candidates to be dropped (Fig. 5). Formally, if we denote the optimal aligned boundary set as $\\mathcal{B} = \\{b_{k_r}\\}_{r=1}^{M-1}$ , where $b_{k_r}$ corresponds to the $r$ -th transition and $1 \\leq k_1 < \\ldots < k_{M-1} \\leq K$ . It should satisfy:", + "bbox": [ + 75, + 271, + 470, + 377 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {B} = \\arg \\min _ {\\mathcal {B} ^ {\\prime}} \\psi \\left(\\mathcal {B} ^ {\\prime}\\right), \\quad \\psi \\left(\\mathcal {B} ^ {\\prime}\\right) = - \\sum_ {r = 1} ^ {M - 1} \\boldsymbol {V} _ {k _ {r} ^ {\\prime}, r}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 116, + 383, + 468, + 425 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\psi (\\mathcal{B}^{\\prime})$ is the cost function of an alignment and $\\mathcal{B}'$ is any feasible aligned boundary set such as $\\mathcal{B}$ .", + "bbox": [ + 75, + 431, + 468, + 462 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To solve Eq. (5), we propose a DP algorithm, of which an example is shown in Fig. 6. To allow some candidates in $\\widetilde{B}$ to be dropped, inspired by [36], we first expand the transition sequence $\\mathcal{R}$ by inserting the empty symbol $\\phi$ interleaved with transition symbols, obtaining $\\mathcal{R}' = [\\phi, (a_1, a_2), \\phi, (a_2, a_3), \\phi, \\dots, \\phi, (a_{M-1}, a_M), \\phi]$ . Now the candidate boundary matched to $\\phi$ will be discarded. Then, a cost matrix $\\Delta \\in \\mathbb{R}^{K \\times (2(M-1) + 1)}$ is constructed (Fig. 6(a)), whose $(k, r')$ -th entry is the cost of aligning the $b_k$ with the $r'$ -th symbol in $\\mathcal{R}'$ . For even number of $r'$ (transition symbols), it is clear that $\\Delta_{k,r'} = -V_{k,r'/2}$ from Eq. (5). Odd $r'$ means dropping the candidate, and the cost is set to 0.", + "bbox": [ + 75, + 463, + 468, + 643 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "A valid alignment is represented by a path from the top-left to bottom-right of $\\Delta$ with constraint on directions. Specifically, the allowed start positions are $(1,1)$ and $(1,2)$ , respectively corresponding to two cases where $b_{1}$ is dropped and $b_{1}$ is matched with the first transition. Similarly, there are two allowed end positions: $(K,2(M - 1))$ and $(K,2(M - 1) + 1)$ . Then for the position $(k,r^{\\prime})$ on the path, if $r^{\\prime}$ is odd, its previous position is either $(k - 1,r^{\\prime})$ or $(k - 1,r^{\\prime} - 1)$ . It means that if $b_{k}$ is dropped, the $b_{k - 1}$ is also dropped or matched with the previous transition. On the other hand, the previous position is either $(k - 1,r^{\\prime} - 1)$ or $(k - 1,r^{\\prime} - 2)$ if $r^{\\prime}$ is even, with the similar meaning. These allowed directions are indicated by red arrows in Fig. 6(b).", + "bbox": [ + 75, + 645, + 470, + 839 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The DP algorithm generally finds the optimal solution by computing a cumulative cost matrix $\\pmb{D}$ with the same shape as $\\pmb{\\Delta}$ (Fig. 6(b)), where $D_{k,r'}$ represents the minimum cumulative cost of all valid paths ending at $(k, r')$ . It", + "bbox": [ + 75, + 839, + 470, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "is computed by the following recursive equation, considering paths coming from valid directions aforementioned:", + "bbox": [ + 498, + 90, + 890, + 121 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {D} _ {k, r ^ {\\prime}} = \\boldsymbol {\\Delta} _ {k, r ^ {\\prime}} + \\left\\{ \\begin{array}{l} \\min \\left(\\boldsymbol {D} _ {k - 1, r ^ {\\prime}}, \\boldsymbol {D} _ {k - 1, r ^ {\\prime} - 1}\\right), r ^ {\\prime} \\text {i s o d d}, \\\\ \\min \\left(\\boldsymbol {D} _ {k - 1, r ^ {\\prime} - 1}, \\boldsymbol {D} _ {k - 1, r ^ {\\prime} - 2}\\right), r ^ {\\prime} \\text {i s e v e n}. \\end{array} \\right. \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 500, + 132, + 890, + 180 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The above equation is not applicable to the first row and the first two columns of $D$ . They are directly initialized before the computation (blue areas of Fig. 6(b)). We set the cost of invalid positions to $\\infty$ (inf) and others are copied from $\\Delta$ .", + "bbox": [ + 496, + 181, + 890, + 241 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The cumulative cost of the optimal alignment is $\\min (D_{K,2(M - 1)},D_{K,2(M - 1) + 1})$ , The complete alignment, i.e., the optimal boundary set $\\mathcal{B}$ can be obtained by starting from the corresponding position of the above min operation and backtracking (Fig. 6(b/c)). More details can be found in the supplementary material.", + "bbox": [ + 496, + 242, + 890, + 330 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "- Complexity. Clearly, the class-agnostic boundary scoring and action transition scoring can be implemented by the unfold operator and matrix operations, thus can be parallelized on GPUs. The action transition alignment must be performed serially with a time complexity of $O(KM)$ or $O(\\lambda M^2)$ . Compared with the previous alignment methods such as Viterbi ( $O(T^2M)$ [25, 28, 34]) and DTW ( $O(TM)$ ), it decreases from a function involving video length $T$ to one of only transcript length $M$ ( $M << T$ ).", + "bbox": [ + 496, + 332, + 890, + 468 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.5. Video-Level Losses", + "text_level": 1, + "bbox": [ + 500, + 477, + 683, + 491 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Inevitably, the inferred pseudo segmentation contains some degree of noise, which can be unexpectedly fit by the network. To improve the semantic robustness, we propose to jointly train a video-level multi-label classification task that predicts whether each action appears in the video, which is precisely indicated by the transcript. Specifically, inspired by the common practice of Transformer [7, 10, 19, 24, 43], the input sequence $\\mathbf{X}$ is augmented by a set of $|\\mathcal{C}|$ learnable class tokens which are fed into the same network and responsible for predicting action occurrence. The outputs, denoted as $\\{e_c'\\}_{c=1}^{|\\mathcal{C}|}$ , are then used to predict the action occurrence probability via $\\xi_c = \\sigma(\\boldsymbol{w}_c^T \\boldsymbol{e}_c' + \\epsilon_c)$ for each category $c$ , where $\\boldsymbol{w}_c$ and $\\epsilon_c$ are the weight vector and bias scalar for class $c$ in the classifier shared with frame-wise classification. $\\sigma(\\cdot)$ is the sigmoid activation. These predictions are supervised by the binary cross entropy:", + "bbox": [ + 496, + 500, + 890, + 744 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {v i d}} = - \\frac {1}{| \\mathcal {C} |} \\sum_ {c = 1} ^ {| \\mathcal {C} |} \\left[ y _ {c} ^ {\\mathrm {v i d}} \\log \\xi_ {c} + \\left(1 - y _ {c} ^ {\\mathrm {v i d}}\\right) \\log \\left(1 - \\xi_ {c}\\right) \\right], \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 506, + 755, + 890, + 797 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $y_c^{\\mathrm{vid}} \\in \\{0,1\\}$ is the binary label indicating whether action $c$ appears in the video. Through the shared network and the interactions between tokens, the semantics learned from video-level training can benefit the main task.", + "bbox": [ + 496, + 809, + 890, + 869 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Moreover, we treat the class tokens with global information as the prototypes of each action and attempt to align the", + "bbox": [ + 498, + 869, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "18257", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "frame features to them for semantic consistency. Specifically, we first calculate the feature centroid of each category $c$ that appears in the video by averaging the output frame features with the corresponding pseudo labels:", + "bbox": [ + 75, + 90, + 468, + 151 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\overline {{\\boldsymbol {x}}} _ {c} = \\frac {\\sum_ {t = 1} ^ {T} \\mathbb {I} \\left(\\widetilde {y} _ {t} = c\\right) \\boldsymbol {x} _ {t} ^ {\\prime}}{\\sum_ {t = 1} ^ {T} \\mathbb {I} \\left(\\widetilde {y} _ {t} = c\\right)}, c \\in \\operatorname {S e t} (\\mathcal {A}), \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 142, + 162, + 468, + 202 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where the $\\operatorname{Set}(\\cdot)$ operator converts a list into an unordered set and removes duplicates. These centroids will move towards the corresponding $e_c'$ under the guidance of a global-local contrastive loss in the form of InfoNCE [17, 31]:", + "bbox": [ + 76, + 212, + 468, + 273 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {g l c}} = - \\frac {1}{| \\operatorname {S e t} (\\mathcal {A}) |} \\sum_ {c \\in \\operatorname {S e t} (\\mathcal {A})} \\log \\frac {\\exp \\left(\\left\\langle \\overline {{\\boldsymbol {x}}} _ {c} \\right\\rangle^ {T} \\left\\langle \\boldsymbol {e} _ {c} ^ {\\prime} \\right\\rangle / \\tau\\right)}{\\sum_ {c ^ {\\prime} = 1} ^ {| \\mathcal {C} |} \\exp \\left(\\left\\langle \\overline {{\\boldsymbol {x}}} _ {c} \\right\\rangle^ {T} \\left\\langle \\boldsymbol {e} _ {c ^ {\\prime}} ^ {\\prime} \\right\\rangle / \\tau\\right)}, \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 78, + 282, + 468, + 335 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\langle \\pmb{x}\\rangle = \\pmb{x} / \\| \\pmb{x}\\| _2$ is the $l_{2}$ -normalization operator, and $\\tau$ is the temperature hyper-parameter.", + "bbox": [ + 76, + 335, + 468, + 367 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.6. Training and Inference", + "text_level": 1, + "bbox": [ + 76, + 377, + 290, + 393 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "- Training. Since the pseudo labeling always requires a relatively good initialization, we utilize a two-stage training strategy. The first stage only uses the reliable labels, i.e., the video-level labels, to pretrain the network. Hence the loss function is $\\mathcal{L}_{\\mathrm{I}} = \\mathcal{L}_{\\mathrm{vid}}$ . The loss for the second stage further includes the frame-wise losses on both classification and representation using pseudo labels: $\\mathcal{L}_{\\mathrm{II}} = \\alpha \\mathcal{L}_{\\mathrm{vid}} + \\beta \\mathcal{L}_{\\mathrm{cls}} + \\gamma \\mathcal{L}_{\\mathrm{glc}}$ , where $\\alpha$ , $\\beta$ and $\\gamma$ are all hyperparameters.", + "bbox": [ + 75, + 400, + 468, + 535 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "- Inference. For inference, we obtain the action label $\\hat{y}_t$ for frame $t$ directly from the frame-wise class probabilities: $\\hat{y}_t = \\arg \\max_c P_{t,c}$ . Note that we do not require any alignment processing during inference, which is performed by previous WSAS methods for smoother predictions. However, since the ground-truth transcript is not available during inference, these methods either run alignment with every transcript from the training set [4, 5, 25, 28, 34], or predict the transcript via another model [38], which makes the pipeline more time-consuming and less practical.", + "bbox": [ + 75, + 536, + 468, + 686 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 76, + 700, + 207, + 717 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Experimental Setup", + "text_level": 1, + "bbox": [ + 76, + 726, + 266, + 742 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "- Datasets. We perform experiments on three datasets. The Breakfast [21] dataset contains 1712 videos of breakfast cooking with 48 different actions. On average, each video has 6.8 segments and $7.3\\%$ frames are background. The Hollywood Extended [2] dataset contains 937 videos taken from movies with 16 categories of daily actions such as walk or sit. On average, each video has 5.9 segments and $60.9\\%$ frames are background. The CrossTask [47] dataset contains videos from 18 primary tasks. Following [28], 14 cooking-related tasks are selected, which have 2552 videos", + "bbox": [ + 75, + 750, + 468, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "and 80 action categories. On average, each video has 14.4 segments and $74.8\\%$ frames are background. For Breakfast, we use the released 4 training/test splits and report the average. For Hollywood, we perform a 10-fold cross-validation. For CrossTask, we use the released training/test split. These evaluation protocols are consistent with previous methods.", + "bbox": [ + 496, + 90, + 890, + 181 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "- Metrics. To evaluate our method, we use 4 standard metrics, following [5, 9, 25, 28]. (1) The Mean-over-Frames (MoF) is the percentage of frames whose labels are correctly predicted. (2) The Mean-over-Frames without Background (MoF-Bg) is the MoF over non-background frames. It is more suitable than MoF for the datasets with high background rate such as Hollywood and CrossTask. (3) The Intersection-over-Union (IoU) is defined as $|I \\cap I^{*}| / |I \\cup I^{*}|$ while (4) the Intersection-over-Detection (IoD) is $|I \\cap I^{*}| / |I|$ , where $I^{*}$ and $I$ are the ground-truth (GT) segment and the predicted segment with the same class, respectively. For each GT segment, the highest IoU/IoD with one predicted segment are preserved, and the average of all GT segments is reported. Note that the definition of the IoU/IoD in [28] is different from other works [5, 9, 25], and will be indicated with special symbols when used.", + "bbox": [ + 496, + 186, + 892, + 429 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "- Input Features. As the input features $X$ for Breakfast and Hollywood, we use the 2048-dimensional RGB+flow I3D features [3] adopted by MuCon [38] and most FSAS methods [13, 26, 27, 45], while some WSAS methods [5, 25, 28, 34] still use the iDT features [41]. Since recent studies [37, 38] have found that they perform worse with more advanced I3D features, we still report the performance with the iDT features for them following [38]. For CrossTask, the officially released 3200-dimensional features are adopted, but we do not perform dimension reduction via PCA as with [28]. For GPU memory efficiency, we perform a $10 \\times$ temporal downsampling for Breakfast and $5 \\times$ for Hollywood. During inference, the output is upsampled to match the original video/ground-truth length.", + "bbox": [ + 496, + 435, + 890, + 647 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "- Implementation Details. We adopt a 6-layer Transformer [40] with single-head self-attention as the temporal network, and the latent dimension $d'$ is 256. For the ATBA, we set $w^{\\mathrm{b}} = 7$ and $\\lambda = 4$ for all datasets. $w^{\\mathrm{a}}$ is set to 31 for Breakfast/CrossTask and 23 for Hollywood. $\\mu$ is set to 0.3 for Breakfast/Hollywood and 0.1 for CrossTask. The loss weights $\\alpha$ , $\\beta$ are set to 1.0 and $\\gamma$ is 0.1. $\\tau$ in Eq. (9) is set to 0.2. Besides, to alleviate the issue that the predictions on Hollywood/CrossTask are overly dominated by background, we lower the sample weights of pseudo background frames to 0.8 in Eq. (1) for these datasets.", + "bbox": [ + 496, + 652, + 890, + 819 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "- Multiple Runs. Due to the alternating nature of learning from weak supervision, there are often fluctuations in the results of WSAS methods [28, 38]. Hence, following [38], we report the average and standard deviation over 5 runs with different random seeds for better evaluation.", + "bbox": [ + 496, + 825, + 890, + 898 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "18258", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/41814da1ae32ae998c1ac57b07da48c4df1e644f9fa04426456d0fd2d022bc97.jpg", + "table_caption": [], + "table_footnote": [ + "Table 2. Comparison of accuracy, training and inference time on the Breakfast. The training time is measured as the entire training duration on the split 1, and the inference time is measured as the average time for inferring a video from the test set of the split 1. Tr.A.-The alignment algorithm adopted during training (X-No Alignment. V-Viterbi. D-DTW. B-Boundary Alignment). *-Measured by [38]. ‡-Measured by us. Best results are in bold, second best are underlined." + ], + "table_body": "
Breakfast
MethodMoF±stdMoF-Bg±stdIoU±stdIoD±std
HMM+RNN [33]33.3---
[33]+Length [23]36.7---
TCFPN+ISBA [9]38.4/36.4±1.0*38.424.240.6
NN-Viterbi [34]43.0/39.7±2.4*---
D3TW [4]45.7---
CDFL [25]50.2/48.1±2.5*48.033.745.4
DPDTW [5]50.8-35.645.1
TASL [28]♠47.8-35.2†46.1†
MuCon [38]♠48.5±1.850.3*40.9*54.0*
POC [29]♠45.7-38.3†-
AdaAct [46]51.248.336.346.4
ATBA♠(Ours)53.9±1.254.4±1.241.1±0.761.7±1.1
39.5±0.8†55.9±1.0†
", + "bbox": [ + 81, + 88, + 468, + 276 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/a3520456853db04fe1dc8361182fe79f9aa7b018408fb900b27605c9bb4d2975.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Hollywood Extended
MethodMoF±stdMoF-Bg±stdIoU±stdIoD±std
HMM+RNN [33]--11.9-
[33]+Length [23]--12.3-
TCFPN+ISBA [9]28.734.512.618.3
D3TW [4]33.6---
CDFL [25]45.040.619.525.8
DPDTW [5]55.625.6#33.243.3
TASL [28]42.1#27.2#23.3†#33.0†#
MuCon [38]♠-41.613.9*-
ATBA♠(Ours)47.7±2.540.2±1.630.9±1.655.8±0.8
28.5±1.6†44.9±0.5†
", + "bbox": [ + 81, + 276, + 468, + 431 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/80a8a7ea6a98f581dff8a7609fc21da2f25b1d66a33d57b8ec598f3fe890459d.jpg", + "table_caption": [], + "table_footnote": [ + "Table 1. Comparisons of ours with other WSAS methods on three datasets. std is the standard deviation of multiple runs (if any). $\\dagger$ The metric is computed by the definition of [28]. *-Results are reported by other works. Please refer to the supplementary material for detailed sources. $\\sharp$ -Results are obtained by us via rerunning the open sources (The TASL [28] does not follow the common 10-fold evaluation protocol for Hollywood so we re-produce the results). $\\spadesuit$ The reported results are the average of multiple runs. Best results are in bold, second best are underlined." + ], + "table_body": "
CrossTask
MethodMoF±stdMoF-Bg±stdIoU±stdIoD±std
NN-Viterbi [34]♠26.5*-10.7†*24.0†*
CDFL [25]♠31.9*-11.5†*23.8†*
TASL [28]♠40.727.4#14.5†25.1†
POC [29]♠42.817.6#15.6†-
ATBA♠(Ours)50.6±1.331.3±0.720.9±0.444.6±0.7
15.7±0.3†24.6±0.4†
", + "bbox": [ + 81, + 431, + 468, + 540 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. Comparison with the State-of-the-Art", + "text_level": 1, + "bbox": [ + 76, + 705, + 401, + 722 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "- Performance. In Table 1, we compare our proposed method with previous WSAS methods. Our ATBA achieves state-of-the-art (SOTA) by a clear margin $(+2.7\\% \\mathrm{MoF})$ on the Breakfast [21], demonstrating the effectiveness of focusing on action transitions. Moreover, comparing the standard deviation with other methods, it can be found that our method is also relatively stable.", + "bbox": [ + 75, + 731, + 468, + 838 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Our ATBA also achieves comparable performance on the Hollywood [2]. The reason why our method does not show significant advantage is probably because this dataset is collected from movies and so contains many shot changes, re", + "bbox": [ + 75, + 839, + 468, + 900 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/df7ef5c07ee0a06de86f38495c759067dc90227f32906b13a8fb3df009bbac74.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodTr.A.MoFTraining (Hours)Inference (Seconds)
TCFPN+ISBA [9]✘33.312.75*0.01*
NN-Viterbi [34]V43.011.23*56.25*
CDFL [25]V50.266.73*62.37*
DPDTW [5]D50.831.02‡0.69‡
TASL [28]V47.824.66‡54.99‡
MuCon [38]✘48.54.573.03
POC [29]✘45.72.28‡0.01‡
ATBA (Ours)B53.93.450.01
", + "bbox": [ + 524, + 89, + 867, + 234 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "sulting in more noisy boundaries. Note that although the DPDTW [5] achieves significantly high MoF, this metric is severely biased in case of highly imbalanced categories (60.9% frames are background). We report the MoF-Bg metric for it, on which it performs poorly, proving that it cannot recognize real actions very well. In contrast, the performance of our method is more balanced.", + "bbox": [ + 496, + 397, + 890, + 502 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "For the more difficult dataset CrossTask [47] with the most segments and highest background rate, our approach also outperforms previous methods whether or not the metric involves background ( $+7.8\\%$ MoF and $+3.9\\%$ MoF-Bg).", + "bbox": [ + 496, + 508, + 890, + 570 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "- Efficiency. Besides of the performance, we also compare the training time of our approach with previous WSAS methods to show the efficiency of ours. Following [38], we train our model on an Nvidia GeForce GTX 1080Ti GPU, and the training time is measured as the wall time over the whole training phase, during which any irrelevant operations such as intermediate evaluation and saving checkpoints are removed. As all the WSAS methods directly load pre-computed features, the time measurement also does not include the time to extract features from raw videos. As shown in Table 2, our ATBA achieves the best performance with the second shortest training time, which demonstrates the effectiveness of our design. Specifically, the training speed of ours is on average 10 times (varying from 3 to 20) faster than methods performing frame-by-frame alignment [5, 25, 28, 34]. Compared to the alignment-free methods with comparable training speed [29, 38], ours achieves better performance.", + "bbox": [ + 496, + 575, + 890, + 848 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We also compare the time to infer a single test video. As mentioned in Sec. 3.6, ours does not require any alignment processing, which makes it the fastest during inference.", + "bbox": [ + 496, + 854, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "18259", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/c0f7687e3d233059dd00f6231e24f504067130fe3adf66526d75cf12db4ef53f.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Exp.\\( \\gamma^b \\)Cost Mat. \\( V^a \\)P.L.MoFIoUIoD
1✓38.632.323.152.5
2✓✓51.741.330.247.8
3✓○✓67.954.041.1
", + "bbox": [ + 120, + 88, + 426, + 170 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.3. Ablation Studies", + "text_level": 1, + "bbox": [ + 76, + 306, + 240, + 321 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "- Effect of ATBA. In Table 3, we conducted an in-depth evaluation of our proposed ATBA. From Exp.1, the pseudo label quality and the evaluation performance are both very poor when only using the class-agnostic boundary scores $\\nu^{\\mathrm{b}}$ to directly select $M - 1$ boundaries for transitions via the greedy strategy stated in Sec. 3.4, showing that it is critical to take the class-specific transition pattern into account as with our design to suppress the noisy boundaries for more precise pseudo segmentation. Fig. 2 provides an intuitive example. In addition, comparing Exp.2 and 3, it's better to involve the class-agnostic boundary scores in the action transition alignment. We think it is because that some candidates can have very low boundary scores (unlikely to be a boundary) as the candidate selection only depends on ranking, and when the transition patterns are not discriminative at the beginning of training, these candidates may be unexpectedly aligned if the boundary scores are not involved.", + "bbox": [ + 75, + 330, + 468, + 587 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "- Effect of Video-Level Losses. In Table 4, we evaluate the $\\mathcal{L}_{\\mathrm{vid}}$ and $\\mathcal{L}_{\\mathrm{glc}}$ . Comparing Exp.1 and 2, the model with $\\mathcal{L}_{\\mathrm{vid}}$ achieves higher performance ( $+6.5\\%$ MoF) at the same level of pseudo label accuracy, demonstrating that the video-level supervision can promote the learning of precise action semantics. In Exp.3, the explicit representation alignment $(\\mathcal{L}_{\\mathrm{glc}})$ further improves the quality of frame features and thus the performance ( $+1.6\\%$ MoF). Moreover, Exp.4 shows that whether or not to share the classifier between frame-/video-level classification has little impact.", + "bbox": [ + 75, + 588, + 468, + 739 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We also provide the analysis on the effect of important hyper-parameters in the supplementary material, including $w^{\\mathrm{b}}$ , $w^{\\mathrm{a}}$ , $\\mu$ and $\\lambda$ in the ATBA.", + "bbox": [ + 75, + 741, + 468, + 787 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4. Qualitative Results", + "text_level": 1, + "bbox": [ + 76, + 800, + 259, + 816 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We show the qualitative inference result compared with two recent open source WSAS methods, i.e., MuCon [38] and TASL [28], in Fig. 7. Our ATBA achieves significantly more accurate results in this challenging video with many segments. The result of TASL shows order reversal (\"Crake", + "bbox": [ + 75, + 825, + 468, + 900 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/8dfefe78dd654143252e7bdf1cf17fc804d871cb3ead2d6c9573e623b2aaeb7d.jpg", + "table_caption": [ + "Table 3. Ablation studies of ATBA on the Breakfast. Exp.-Different experiment configurations. Cost Mat.-Different choices for the cost matrix in action transition alignment. o-According to Eq. (4), $V$ contains $V^{\\mathrm{a}}$ . P.L.-The accuracy of pseudo labels during training. Exp.1 means that only the class-agnostic boundary detection is adopted to localize action transitions. Exp.3 is our default setting. Best results are in bold." + ], + "table_footnote": [], + "table_body": "
Exp.\\( \\mathcal{L}_{\\text{vid}} \\)\\( \\mathcal{L}_{\\text{glc}} \\)D.Cls.P.L.MoFIoUIoD
167.545.940.561.9
2✓66.652.438.861.4
3✓✓67.954.041.162.3
4✓✓✓68.153.740.161.7
", + "bbox": [ + 532, + 89, + 859, + 172 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 4. Ablation studies of video-level losses on the Breakfast. Exp.-Different experiment configurations. D.Cls.-Different classifiers for frame-wise prediction and action occurrence prediction. P.L.-The accuracy of pseudo labels during training. Exp.3 is our default setting. Best results are in bold.", + "bbox": [ + 498, + 183, + 890, + 253 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/e1977212b6e5ed40c5dde94ef8a5582bd3e4c1ac887b4ba97c5238537188ea79.jpg", + "image_caption": [ + "Figure 7. Qualitative results on the Breakfast. The example test video is P15-stereo01-P15-pancake. We compare our inference result with two recent methods. Best viewed in color." + ], + "image_footnote": [], + "bbox": [ + 503, + 270, + 890, + 424 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Egg\" and \"Spoon Flour\") and hallucination (\"Butter Pan\"). The MuCon predicts the correct action ordering, but the result deviates severely. In contrast, ours successfully predicts an accurate segmentation, indicating the action semantics are well learned. More qualitative results are provided in supplementary material.", + "bbox": [ + 498, + 505, + 890, + 597 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 614, + 617, + 631 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this work, we propose to directly localize action transitions for efficient pseudo segmentation generation in the WSAS task, thus avoiding the time-consuming frame-by-frame alignment. Due to the presence of noisy boundaries, a novel Action-Transition-Aware Boundary Alignment (ATBA) framework is proposed to efficiently and effectively filter out noise and detect transitions. Moreover, we also design some video-level losses to utilize video-level supervision to improve the semantic robustness. Extensive experiments show the effectiveness of our ATBA.", + "bbox": [ + 496, + 642, + 890, + 792 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Acknowledgment", + "text_level": 1, + "bbox": [ + 500, + 813, + 669, + 830 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This work was supported partially by the NSFC (U21A20471), National Key Research and Development Program of China (2023YFA1008503), Guangdong NSF Project (No. 2023B1515040025, 2020B1515120085).", + "bbox": [ + 498, + 839, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "18260", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Emad Bahrami, Gianpiero Francesca, and Juergen Gall. How much temporal long-term context is needed for action segmentation? arXiv preprint arXiv:2308.11358, 2023. 1", + "[2] Piotr Bojanowski, Rémi Lajugie, Francis Bach, Ivan Laptev, Jean Ponce, Cordelia Schmid, and Josef Sivic. Weakly supervised action labeling in videos under ordering constraints. In Proceedings of the European Conference on Computer Vision, pages 628-643. Springer, 2014. 1, 2, 6, 7", + "[3] Joao Carreira and Andrew Zisserman. Quo vadis, action recognition? a new model and the kinetics dataset. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 6299-6308, 2017. 6", + "[4] Chien-Yi Chang, De-An Huang, Yanan Sui, Li Fei-Fei, and Juan Carlos Niebles. D3tw: Discriminative differentiable dynamic time warping for weakly supervised action alignment and segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3546–3555, 2019. 1, 2, 6, 7", + "[5] Xiaobin Chang, Frederick Tung, and Greg Mori. Learning discriminative prototypes with dynamic time warping. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8395-8404, 2021. 1, 2, 6, 7", + "[6] Min-Hung Chen, Baopu Li, Yingze Bao, Ghassan Al-Regib, and Zsolt Kira. Action segmentation with joint self-supervised temporal domain adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9454–9463, 2020. 1", + "[7] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018. 5", + "[8] Guodong Ding and Angela Yao. Leveraging action affinity and continuity for semi-supervised temporal action segmentation. In European Conference on Computer Vision, pages 17-32. Springer, 2022. 1", + "[9] Li Ding and Chenliang Xu. Weakly-supervised action segmentation with iterative soft boundary assignment. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6508-6516, 2018. 1, 2, 3, 6, 7", + "[10] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020.5", + "[11] Dazhao Du, Bing Su, Yu Li, Zhongang Qi, Lingyu Si, and Ying Shan. Efficient u-transformer with boundary-aware loss for action segmentation. arXiv preprint arXiv:2205.13425, 2022. 3", + "[12] Zexing Du, Xue Wang, Guoqing Zhou, and Qing Wang. Fast and unsupervised action boundary detection for action segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3323-3332, 2022. 1, 2, 4" + ], + "bbox": [ + 78, + 114, + 468, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[13] Yazan Abu Farha and Jurgen Gall. Ms-tcn: Multi-stage temporal convolutional network for action segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3575-3584, 2019. 1, 6", + "[14] Shang-Hua Gao, Qi Han, Zhong-Yu Li, Pai Peng, Liang Wang, and Ming-Ming Cheng. Global2local: Efficient structure search for video action segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16805-16814, 2021. 1", + "[15] Reza Ghoddoosian, Isht Dwivedi, Nakul Agarwal, Chiho Choi, and Behzad Dariush. Weakly-supervised online action segmentation in multi-view instructional videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13780-13790, 2022. 1", + "[16] Reza Ghoddoosian, Isht Dwivedi, Nakul Agarwal, and Behzad Dariush. Weakly-supervised action segmentation and unseen error detection in anomalous instructional videos. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10128-10138, 2023. 1", + "[17] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9729-9738, 2020. 6", + "[18] De-An Huang, Li Fei-Fei, and Juan Carlos Niebles. Connectionist temporal modeling for weakly supervised action labeling. In Proceedings of the European Conference on Computer Vision, pages 137–153. Springer, 2016. 1, 2", + "[19] Zi-Hang Jiang, Qibin Hou, Li Yuan, Daquan Zhou, Yujun Shi, Xiaojie Jin, Anran Wang, and Jiashi Feng. All tokens matter: Token labeling for training better vision transformers. Advances in neural information processing systems, 34: 18590-18602, 2021. 5", + "[20] Hyolim Kang, Jinwoo Kim, Taehyun Kim, and Seon Joo Kim. Uboco: Unsupervised boundary contrastive learning for generic event boundary detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20073-20082, 2022. 2, 4", + "[21] Hilde Kuehne, Ali Arslan, and Thomas Serre. The language of actions: Recovering the syntax and semantics of goal-directed human activities. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 780-787, 2014. 2, 6, 7", + "[22] Hilde Kuehne, Alexander Richard, and Juergen Gall. Weakly supervised learning of actions from transcripts. Computer Vision and Image Understanding, 163:78-89, 2017. 1, 2", + "[23] Hilde Kuehne, Alexander Richard, and Juergen Gall. A hybrid rnn-hmm approach for weakly supervised temporal action segmentation. IEEE transactions on pattern analysis and machine intelligence, 42(4):765-779, 2018. 1, 2, 7", + "[24] Jack Lanchantin, Tianlu Wang, Vicente Ordonez, and Yanjun Qi. General multi-label image classification with transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16478-16488, 2021. 5", + "[25] Jun Li, Peng Lei, and Sinisa Todorovic. Weakly supervised energy-based learning for action segmentation. In Proceed-" + ], + "bbox": [ + 501, + 92, + 893, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "18261", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "ings of the IEEE/CVF International Conference on Computer Vision, pages 6243-6251, 2019. 1, 2, 5, 6, 7", + "[26] Shi-Jie Li, Yazan AbuFarha, Yun Liu, Ming-Ming Cheng, and Juergen Gall. Ms-tcn++: Multi-stage temporal convolutional network for action segmentation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2020. 6", + "[27] Daochang Liu, Qiyue Li, AnhDung Dinh, Tingting Jiang, Mubarak Shah, and Chang Xu. Diffusion action segmentation. arXiv preprint arXiv:2303.17959, 2023. 1, 6", + "[28] Zijia Lu and Ehsan Elhamifar. Weakly-supervised action segmentation and alignment via transcript-aware union-of-subspaces learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8085-8095, 2021. 1, 2, 5, 6, 7, 8", + "[29] Zijia Lu and Ehsan Elhamifar. Set-supervised action learning in procedural task videos via pairwise order consistency. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19903-19913, 2022. 1, 3, 7", + "[30] Antoine Miech, Dimitri Zhukov, Jean-Baptiste Alayrac, Makarand Tapaswi, Ivan Laptev, and Josef Sivic. Howto100m: Learning a text-video embedding by watching hundred million narrated video clips. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2630-2640, 2019. 1", + "[31] Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748, 2018. 6", + "[32] Rahul Rahaman, Dipika Singhania, Alexandre Thiery, and Angela Yao. A generalized and robust framework for timestamp supervision in temporal action segmentation. In Proceedings of the European Conference on Computer Vision, pages 279–296. Springer, 2022. 1", + "[33] Alexander Richard, Hilde Kuehne, and Juergen Gall. Weakly supervised action learning with rn based fine-to-coarse modeling. In Proceedings of the IEEE conference on Computer Vision and Pattern Recognition, pages 754-763, 2017. 1, 2, 7", + "[34] Alexander Richard, Hilde Kuehne, Ahsan Iqbal, and Juergen Gall. Neuralnetwork-viterbi: A framework for weakly supervised video learning. In Proceedings of the IEEE conference on Computer Vision and Pattern Recognition, pages 7386-7395, 2018. 1, 2, 5, 6, 7", + "[35] Yuhan Shen and Ehsan Elhamifar. Semi-weakly-supervised learning of complex actions from instructional task videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3344–3354, 2022. 1", + "[36] Yuhan Shen, Lu Wang, and Ehsan Elhamifar. Learning to segment actions from visual and language instructions via differentiable weak sequence alignment. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10156-10165, 2021. 5", + "[37] Yaser Souri, Alexander Richard, Luca Minciullo, and Juergen Gall. On evaluating weakly supervised action segmentation methods. arXiv preprint arXiv:2005.09743, 2020. 6", + "[38] Yaser Souri, Mohsen Fayyaz, Luca Minciullo, Gianpiero Francesca, and Juergen Gall. Fast weakly supervised action" + ], + "bbox": [ + 78, + 90, + 468, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "segmentation using mutual consistency. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(10):6196-6208, 2021. 1, 2, 3, 6, 7, 8", + "[39] Yaser Souri, Yazan Abu Farha, Fabien Despinoy, Gianpiero Francesca, and Juergen Gall. Fif: Fast inference approximation for action segmentation. In Proceedings of the DAGM German Conference on Pattern Recognition, pages 282-296. Springer, 2022. 2", + "[40] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 3, 6", + "[41] Heng Wang and Cordelia Schmid. Action recognition with improved trajectories. In Proceedings of the IEEE international conference on computer vision, pages 3551-3558, 2013. 6", + "[42] Ruibin Xiong, Yunchang Yang, Di He, Kai Zheng, Shuxin Zheng, Chen Xing, Huishuai Zhang, Yanyan Lan, Liwei Wang, and Tieyan Liu. On layer normalization in the transformer architecture. In International Conference on Machine Learning, pages 10524-10533. PMLR, 2020. 3", + "[43] Lian Xu, Wanli Ouyang, Mohammed Bennamoun, Farid Boussaid, and Dan Xu. Multi-class token transformer for weakly supervised semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4310-4319, 2022. 5", + "[44] Di Yang, Yaohui Wang, Antitza Dantcheva, Quan Kong, Lorenzo Garattoni, Gianpiero Francesca, and Francois Bremond. Lac-latent action composition for skeleton-based action segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13679-13690, 2023. 1", + "[45] Fangqiu Yi, Hongyu Wen, and Tingting Jiang. Asformer: Transformer for action segmentation. arXiv preprint arXiv:2110.08568, 2021. 3, 6", + "[46] Runzhong Zhang, Suchen Wang, Yueqi Duan, Yansong Tang, Yue Zhang, and Yap-Peng Tan. Hoi-aware adaptive network for weakly-supervised action segmentation. In Proceedings of the Thirty-Second International Joint Conference on Artificial Intelligence, pages 1722–1730, 2023. 1, 2, 7", + "[47] Dimitri Zhukov, Jean-Baptiste Alayrac, Ramadan Gokberk Cinbis, David Fouhey, Ivan Laptev, and Josef Sivic. Cross-task weakly supervised learning from instructional videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3537-3545, 2019. 1, 2, 6, 7" + ], + "bbox": [ + 501, + 92, + 890, + 752 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "18262", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2024/Efficient and Effective Weakly-Supervised Action Segmentation via Action-Transition-Aware Boundary Alignment/1ae3fdc0-77c3-4dc9-87e2-e3b0568d2efa_model.json b/2024/Efficient and Effective Weakly-Supervised Action Segmentation via Action-Transition-Aware Boundary Alignment/1ae3fdc0-77c3-4dc9-87e2-e3b0568d2efa_model.json new file mode 100644 index 0000000000000000000000000000000000000000..d1b6744b37f4d6726fb8a6e6edb644ea4496e9cf --- /dev/null +++ b/2024/Efficient and Effective Weakly-Supervised Action Segmentation via Action-Transition-Aware Boundary Alignment/1ae3fdc0-77c3-4dc9-87e2-e3b0568d2efa_model.json @@ -0,0 +1,2178 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.147, + 0.131, + 0.825, + 0.177 + ], + "angle": 0, + "content": "Efficient and Effective Weakly-Supervised Action Segmentation via Action-Transition-Aware Boundary Alignment" + }, + { + "type": "text", + "bbox": [ + 0.346, + 0.203, + 0.625, + 0.221 + ], + "angle": 0, + "content": "Angchi Xu\\(^{1}\\), Wei-Shi Zheng\\(^{1,2,3,*}\\)" + }, + { + "type": "text", + "bbox": [ + 0.178, + 0.221, + 0.793, + 0.239 + ], + "angle": 0, + "content": "\\(^{1}\\)School of Computer Science and Engineering, Sun Yat-sen University, China" + }, + { + "type": "text", + "bbox": [ + 0.099, + 0.24, + 0.873, + 0.258 + ], + "angle": 0, + "content": "\\(^{2}\\)Key Laboratory of Machine Intelligence and Advanced Computing, Ministry of Education, China" + }, + { + "type": "text", + "bbox": [ + 0.112, + 0.258, + 0.858, + 0.275 + ], + "angle": 0, + "content": "3Guangdong Province Key Laboratory of Information Security Technology, Guangzhou, China" + }, + { + "type": "text", + "bbox": [ + 0.293, + 0.278, + 0.678, + 0.293 + ], + "angle": 0, + "content": "xuangch@mail2.sysu.edu.cn, wszheng@ieee.org" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.327, + 0.314, + 0.344 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.359, + 0.474, + 0.692 + ], + "angle": 0, + "content": "Weakly-supervised action segmentation is a task of learning to partition a long video into several action segments, where training videos are only accompanied by transcripts (ordered list of actions). Most of existing methods need to infer pseudo segmentation for training by serial alignment between all frames and the transcript, which is time-consuming and hard to be parallelized while training. In this work, we aim to escape from this inefficient alignment with massive but redundant frames, and instead to directly localize a few action transitions for pseudo segmentation generation, where a transition refers to the change from an action segment to its next adjacent one in the transcript. As the true transitions are submerged in noisy boundaries due to intra-segment visual variation, we propose a novel Action-Transition-Aware Boundary Alignment (ATBA) framework to efficiently and effectively filter out noisy boundaries and detect transitions. In addition, to boost the semantic learning in the case that noise is inevitably present in the pseudo segmentation, we also introduce video-level losses to utilize the trusted video-level supervision. Extensive experiments show the effectiveness of our approach on both performance and training speed." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.721, + 0.21, + 0.737 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.747, + 0.47, + 0.854 + ], + "angle": 0, + "content": "Action segmentation aims to partition a long untrimmed video into several segments and classify each segment into an action category [8, 12, 13, 15, 16, 29, 32, 35, 38, 44]. It is an important yet challenging task for instructional or procedural video understanding. Although fully-supervised action segmentation (FSAS) methods [1, 6, 13, 14, 27] have achieved great progress, they require frame-wise dense an-" + }, + { + "type": "image", + "bbox": [ + 0.501, + 0.326, + 0.891, + 0.528 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.539, + 0.894, + 0.596 + ], + "angle": 0, + "content": "Figure 1. Comparison of performance and training time of WSAS methods on the Breakfast dataset. MoF-The main metric of the task, the higher the better. *-Alignment-free methods. Our ATBA achieves the best performance with a very short training time." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.625, + 0.893, + 0.746 + ], + "angle": 0, + "content": "notation, which is labor-intensive and time-consuming to collect. As a result, many works [5, 9, 25, 28, 34, 38, 46] explore the weakly-supervised action segmentation (WSAS) only requiring the transcript annotation, which refers to the ordered list of actions occurring in the video without their start and end times. The transcripts are less costly to obtain and can be accessed directly from video narrations or other meta data [2, 9, 18, 28, 30, 33, 47]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.748, + 0.895, + 0.869 + ], + "angle": 0, + "content": "Most of previous WSAS methods have to infer the pseudo segmentation (pseudo frame-wise labels) for training via a sequence alignment process between the video and given transcript, such as Viterbi [22, 23, 25, 28, 33, 34] or Dynamic Time Warping (DTW) [4, 5]. These alignment algorithms are usually designed in a recursive form which needs to be performed serially frame-by-frame and hard to be parallelized, resulting in very slow training process." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.894, + 0.902 + ], + "angle": 0, + "content": "In this work, we argue that the frame-by-frame alignment is NOT necessary, since the pseudo segmentation is" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.863, + 0.223, + 0.876 + ], + "angle": 0, + "content": "* Corresponding author." + }, + { + "type": "page_footnote", + "bbox": [ + 0.097, + 0.876, + 0.468, + 0.889 + ], + "angle": 0, + "content": "1Code is available at https://github.com/iSEE-" + }, + { + "type": "page_footnote", + "bbox": [ + 0.08, + 0.889, + 0.258, + 0.9 + ], + "angle": 0, + "content": "Laboratory/CVPR24_ATBA." + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.863, + 0.468, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18253" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.08, + 0.09, + 0.468, + 0.21 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.22, + 0.473, + 0.388 + ], + "angle": 0, + "content": "Figure 2. The necessity of proposed ATBA. The example is P54-webcam01-P54-coffee in Breakfast dataset. GT-The ground-truth segmentation. C.A.Bdy.-Only class-agnostic boundary detection is applied (Exp.1 of Table 3). Acc.-The accuracy of pseudo segmentation. In the video clip around the \"star\" point, the coffee pot undergoes a change from being picked up to tilted pouring within the segment \"Pour Coffee\", and this noisy visual change is incorrectly detected. In addition, although two boundaries are correctly detected by the \"C.A.Bdy.\" (diamonds), they correspond to incorrect transitions due to one false positive error (star), resulting in complete dislocation of segments within the dashed box. Best viewed in color." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.416, + 0.47, + 0.643 + ], + "angle": 0, + "content": "fundamentally determined by the locations of a small number of action transitions (i.e., the change from an action segment to its next adjacent action segment in the transcript). Hence, the pseudo segmentation generation can be viewed as a transition detection problem, implying the way to more efficient designs. Intuitively, action transitions are often accompanied by significant visual changes, and there are already many approaches that can detect class-agnostic action boundaries based on these changes [12, 20]. However, due to the intra-segment visual variation and sub-segments under finer granularity, there are numerous noisy boundaries not corresponding to any transitions. Moreover, as the class-agnostic way cannot guarantee correct correspondence between the boundaries and transitions, even slight errors in the detection can result in severe deviation (Fig. 2)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.644, + 0.471, + 0.902 + ], + "angle": 0, + "content": "To overcome the above noisy boundary issue, we propose an efficient and effective framework for WSAS, termed Action-Transition-Aware Boundary Alignment (ATBA), which directly detects the transitions for faster and effective pseudo segmentation generation. To tolerate the noisy boundaries, the ATBA generates more class-agnostic boundaries than the number of transitions as candidates, and then determines a subset from candidates that optimally matches all desired transitions via a drop-allowed alignment algorithm. Furthermore, to fortify the semantic learning under the inevitable noise in pseudo segmentation, we also introduce video-level losses to make use of the trusted video-level supervision. Our ATBA is efficient, because the number of generated candidates will be proportional to the length of the transcript, and therefore the complexity of alignment is now independent of the very long video length. Moreover, other computations" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.167 + ], + "angle": 0, + "content": "required by ATBA, i.e., measuring how likely a frame is to be a boundary and how likely a candidate corresponds to a desired transition, are both built on a convolution-like algorithm inspired by [20], which can be parallelized on GPUs efficiently." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.168, + 0.892, + 0.257 + ], + "angle": 0, + "content": "For inference, we directly adopt the results from the trained frame-wise classifier, without the need for any alignment processing with retrieved or predicted transcript (the ground-truth transcript is not available during inference) like previous WSAS methods [4, 5, 25, 28, 34, 38], which also improves the inference efficiency." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.258, + 0.893, + 0.47 + ], + "angle": 0, + "content": "In summary, our contributions are as follows. (1) We propose to directly localize the action transitions for efficient pseudo segmentation generation during training, without the need of time-consuming frame-by-frame alignment. (2) For robustness to noisy boundaries, we propose a novel ATBA framework to effectively determine boundaries corresponding to each transition. Video-level losses are also introduced to regularize the semantic learning involving the unavoidable noise in the pseudo segmentation. Experiments are conducted on three popular datasets to evaluate our approach: Breakfast [21], Hollywood Extended [2] and CrossTask [47]. Our ATBA achieves state-of-the-art or comparable results with one of the fastest training speed (Fig. 1), demonstrating the effectiveness of ours." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.483, + 0.642, + 0.499 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.509, + 0.892, + 0.644 + ], + "angle": 0, + "content": "Weakly-supervised action segmentation methods learn to partition a video into several action segments from training videos only annotated by transcripts [2, 4, 5, 9, 18, 22, 23, 25, 28, 33, 34, 38, 39, 46]. Despite different optimization objectives, most of them generate the pseudo segmentation for training by solving alignment objectives between two sequences (video and transcript) via Connectionist Temporal Classification (CTC) [18], Viterbi [22, 23, 25, 28, 33, 34] or Dynamic Time Warping (DTW) [4, 5]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.644, + 0.893, + 0.901 + ], + "angle": 0, + "content": "Specifically, [18] proposes an extended version of CTC to evaluate all valid alignments between videos and transcripts, which additionally takes the visual similarities of frames into account. Inspired by speech recognition, [22, 23, 33] all use the Hidden Markov Model (HMM) to model the relationship between videos and actions. [9] always generates uniform segmentation but iteratively adjusts the boundaries by inserting repeating actions into the transcript. [34] proposes an alignment objective based on explicit context and length models, which can be solved by Viterbi, and the solved optimal alignment would serve as pseudo labels to train the frame-wise classifier. [25] and [28] both focus on novel learning objectives, but still require the optimal pseudo segmentation produced by Viterbi. [4, 5] both learn from the contrast of aligning the video to the ground-truth transcript and negative transcripts, where the alignment is performed by DTW. Whether using CTC, Viterbi or DTW," + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "18254" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.086, + 0.093, + 0.891, + 0.346 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.356, + 0.893, + 0.399 + ], + "angle": 0, + "content": "Figure 3. The overall framework. We propose an Action-Transition-Aware Boundary Alignment (ATBA) framework, which takes the class-agnostic boundary pattern and action transition pattern together into account to efficiently generate pseudo labels. The trusted video-level supervision is also utilized to further enhance the performance." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.425, + 0.468, + 0.454 + ], + "angle": 0, + "content": "the above approaches except [9] require frame-by-frame serial calculation, which are inefficient." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.456, + 0.47, + 0.592 + ], + "angle": 0, + "content": "Recently, some efficient methods are proposed with alignment-free design. [38] learns from the mutual consistency between two forms of a segmentation (i.e., framewise classification and category/length pairs). [29] proposes a loss to enforce the output order of any two actions to be consistent with the transcript. In this work, we also propose an efficient framework with different technical roadmap, and our performance is better at the same level of training speed." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.606, + 0.185, + 0.623 + ], + "angle": 0, + "content": "3. Approach" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.631, + 0.262, + 0.645 + ], + "angle": 0, + "content": "3.1. Problem Statement" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.654, + 0.47, + 0.866 + ], + "angle": 0, + "content": "Action segmentation is a task of partitioning a video into several temporal segments with action labels, which is equivalent to predicting the action categories of each frame. Formally, given a sequence of \\( T \\) \\( d \\)-dimensional frame-wise features \\( \\mathbf{X} = [\\pmb{x}_1,\\dots,\\pmb{x}_T] \\in \\mathbb{R}^{T\\times d} \\) for a video with \\( T \\) frames, the goal is to predict a sequence of actions \\( \\hat{\\mathcal{Y}} = [\\hat{y}_1,\\dots,\\hat{y}_T] \\), where \\( \\hat{y}_t \\in \\mathcal{C} \\) and \\( \\mathcal{C} = \\{1,2,\\dots,|\\mathcal{C}|\\} \\) is the set of action categories across the dataset (including the background). Under the setting of WSAS, the frame-wise ground-truth \\( \\mathcal{Y} = [y_1,\\dots,y_T] \\) is NOT available during training. Instead, the ordered list of actions called transcript \\( \\mathcal{A} = [a_1,\\dots,a_M] \\) is provided (including the background segments), where \\( a_{m} \\in \\mathcal{C} \\) and \\( M \\) is the total number of action segments in the video. The action transitions of \\( \\mathcal{A} \\)" + }, + { + "type": "page_footnote", + "bbox": [ + 0.077, + 0.875, + 0.468, + 0.901 + ], + "angle": 0, + "content": "2POC [29] is a set-supervised method but can be extended to transcript supervision naturally, so we cite its corresponding results for comparison." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.423, + 0.838, + 0.442 + ], + "angle": 0, + "content": "are naturally formulated as \\(\\mathcal{R} = \\{(a_r,a_{r + 1})\\}_{r = 1}^{M - 1}\\)" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.453, + 0.61, + 0.467 + ], + "angle": 0, + "content": "3.2. Overview" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.477, + 0.892, + 0.644 + ], + "angle": 0, + "content": "Our proposed framework is illustrated in Fig. 3. At first, the input sequence \\( \\mathbf{X} \\) is further encoded by a temporal network to generate more task-relevant representations \\( \\mathbf{X}' \\in \\mathbb{R}^{T \\times d'} \\) (Sec. 3.3), then a classifier shared along the temporal axis followed by a category softmax activation will predict the frame-wise class probabilities \\( \\mathbf{P} = [\\mathbf{p}_1, \\dots, \\mathbf{p}_T] \\in \\mathbb{R}^{T \\times |C|} \\) from \\( \\mathbf{X}' \\). After that, the Action-Transition-Aware Boundary Alignment (ATBA) module takes \\( \\mathbf{P} \\) and the transcript \\( \\mathcal{A} \\) as input to infer the pseudo frame-wise labels \\( \\widetilde{\\mathcal{Y}} = [\\widetilde{y}_1, \\dots, \\widetilde{y}_T] \\), where \\( \\widetilde{y}_t \\in C \\) (Sec. 3.4). Finally, \\( \\widetilde{\\mathcal{Y}} \\) is used back to supervise \\( \\mathbf{P} \\) by a standard cross entropy:" + }, + { + "type": "equation", + "bbox": [ + 0.567, + 0.657, + 0.891, + 0.7 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {c l s}} = - \\frac {1}{T} \\sum_ {t = 1} ^ {T} \\sum_ {c = 1} ^ {| c |} \\mathbb {I} (\\widetilde {y} _ {t} = c) \\log \\boldsymbol {P} _ {t, c}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.713, + 0.892, + 0.817 + ], + "angle": 0, + "content": "where \\(\\mathbb{I}(\\cdot)\\) is the indicator function which returns 1 if the condition is satisfied and 0 otherwise. Note that the pseudo frame-wise labels are inferred at each iteration for a batch of data and used immediately for current training at the same iteration. The additional video-level losses are stated in Sec. 3.5, and the training and inference processes are described in Sec. 3.6." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.831, + 0.681, + 0.847 + ], + "angle": 0, + "content": "3.3. Temporal Network" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.892, + 0.901 + ], + "angle": 0, + "content": "We employ a slightly modified pre-norm [42] Transformer [40] encoder with learnable positional embeddings as the temporal network for feature learning. Following [11, 45]," + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "18255" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.118, + 0.089, + 0.257, + 0.195 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.179, + 0.197, + 0.196, + 0.21 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.293, + 0.118, + 0.431, + 0.152 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.352, + 0.197, + 0.371, + 0.21 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.221, + 0.468, + 0.249 + ], + "angle": 0, + "content": "Figure 4. (a) A \\(7 \\times 7\\) template for class-agnostic boundary scoring. \n(b) A \\(2 \\times 7\\) template for action transition scoring." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.274, + 0.468, + 0.32 + ], + "angle": 0, + "content": "the vanilla full self-attention is replaced with a pyramid hierarchical local attention to better adapt to the action segmentation task (see supplemental material for more details)." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.328, + 0.468, + 0.344 + ], + "angle": 0, + "content": "3.4. Action-Transition-Aware Boundary Alignment" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.351, + 0.469, + 0.441 + ], + "angle": 0, + "content": "The ATBA is the core component of our framework. It generates the pseudo frame-wise labels \\(\\widetilde{\\mathcal{V}}\\) by inferring the boundaries corresponding to \\(M - 1\\) action transitions from the predicted class probabilities \\(\\pmb{P}\\). Once they are found, \\(\\widetilde{\\mathcal{V}}\\) can be naturally generated by assigning the action labels of \\(\\mathcal{A}\\) one-by-one into the intervals between these boundaries." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.442, + 0.469, + 0.517 + ], + "angle": 0, + "content": "Briefly, the ATBA first generates a set of candidate boundaries via a class-agnostic way, and then finds \\( M - 1 \\) points from this set that optimally match all transitions of \\( \\mathcal{A} \\) via a dynamic programming (DP) algorithm. We describe the details in the following." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.518, + 0.469, + 0.807 + ], + "angle": 0, + "content": "- Class-Agnostic Boundary Scoring. Firstly, the ATBA calculates the class-agnostic boundary scores \\(\\mathcal{V}^{\\mathrm{b}} = [v_1^{\\mathrm{b}},\\dots,v_T^{\\mathrm{b}}]\\) for each timestamp. We employ a pattern-matching-based scoring method proposed by a generic event boundary detection approach termed UBoCo [20], which considers the pattern of each frame's neighborhood. Specifically, for timestamp \\(t\\), a pairwise similarity matrix \\(\\Gamma^{(t)}\\in \\mathbb{R}^{w^{\\mathrm{b}}\\times w^{\\mathrm{b}}}\\) is calculated within a local window with size \\(w^{\\mathrm{b}}\\) centered at \\(t\\), whose \\((i,j)\\)-entry represents the class-agnostic similarity between \\(i\\)-th and \\(j\\)-th frames inside the window, with values ranging from -1 to 1 (see the computational details in supplementary material). Clearly, if \\(t\\) is a boundary, the frame feature should change dramatically at \\(t\\) and keep stable elsewhere, so its \\(\\Gamma^{(t)}\\) should show the pattern of that the values in the upper-left and lower-right areas are close to 1 and otherwise close to -1. Hence a template \\(\\Omega^{\\mathrm{b}}\\in \\mathbb{R}^{w^{\\mathrm{b}}\\times w^{\\mathrm{b}}}\\) like Fig. 4(a) is designed to capture this pattern and output the class-agnostic boundary score for each \\(t\\) by a correlation operation:" + }, + { + "type": "equation", + "bbox": [ + 0.163, + 0.817, + 0.469, + 0.86 + ], + "angle": 0, + "content": "\\[\nv _ {t} ^ {\\mathrm {b}} = \\frac {1}{w ^ {\\mathrm {b}} \\times w ^ {\\mathrm {b}}} \\sum_ {i = 1} ^ {w ^ {\\mathrm {b}}} \\sum_ {j = 1} ^ {w ^ {\\mathrm {b}}} \\Omega_ {i, j} ^ {\\mathrm {b}} \\Gamma_ {i, j} ^ {(t)}. \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.469, + 0.902 + ], + "angle": 0, + "content": "- Candidate Boundary Selection. After calculating \\(\\nu^{\\mathrm{b}}\\), we select a set of \\(K\\) candidate boundaries \\(\\tilde{\\mathcal{B}} = \\{b_k\\}_{k=1}^K\\), where" + }, + { + "type": "image", + "bbox": [ + 0.541, + 0.09, + 0.857, + 0.175 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.186, + 0.892, + 0.255 + ], + "angle": 0, + "content": "Figure 5. Illustration of the alignment between action transitions \\(\\mathcal{R}\\) and candidate boundaries \\(\\tilde{\\mathcal{B}}\\). Blue circles are aligned and gray ones are dropped. (a) A valid alignment. (b) An invalid alignment. The red dashed arrow violates the ordering consistency. Best viewed in color." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.282, + 0.892, + 0.448 + ], + "angle": 0, + "content": "\\(1 < b_{1} < \\ldots < b_{K} \\leq T\\) and \\(K > M - 1\\). The selection is performed by a simple greedy strategy with non maximum suppression inspired by [12], i.e., each time we select one timestamp with current highest score \\(v_{t}^{\\mathrm{b}}\\), and invalidate its neighborhood to avoid selecting multiple timestamps corresponding to one same boundary, until the number of selected timestamps reach an upper bound or all remaining timestamps are invalid. The radius of the invalid interval is set adaptively to \\(\\frac{\\mu T}{M}\\), where \\(\\mu \\in [0,1]\\) is a hyper-parameter, and in practice, the upper bound for \\(K\\) is set to \\(\\lambda(M - 1)\\), where \\(\\lambda \\in \\mathbb{N}_{+}\\) is also a hyper-parameter." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.449, + 0.892, + 0.629 + ], + "angle": 0, + "content": "- Action Transition Scoring. As mentioned in Sec. 1, the class-agnostic scores \\(\\mathcal{V}^{\\mathrm{b}}\\) are not enough to detect action transitions. To this end, we then calculate an action transition score matrix \\(V^{\\mathrm{a}}\\in \\mathbb{R}^{K\\times (M - 1)}\\), where \\(V_{k,r}^{\\mathrm{a}}\\) measures the possibility that the \\(k\\)-th candidate boundary corresponds to the \\(r\\)-th transition, i.e., separates the \\(r\\)-th and \\((r + 1)\\)-th action segments. This matrix is also calculated via pattern matching. Clearly, if candidate \\(b_{k}\\) corresponds to the \\(r\\)-th transition, the classifier's activation for class \\(a_{r}\\) should drop sharply after \\(b_{k}\\), while rise for class \\(a_{r + 1}\\). Hence, a template \\(\\Omega^{\\mathrm{a}}\\in \\mathbb{R}^{2\\times w^{\\mathrm{a}}}\\) with temporal size \\(w^{\\mathrm{a}}\\) like Fig. 4(b) is employed to detect this pattern around \\(b_{k}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.549, + 0.641, + 0.891, + 0.684 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {V} _ {k, r} ^ {\\mathrm {a}} = \\frac {1}{2 w ^ {\\mathrm {a}}} \\sum_ {i = 1} ^ {2} \\sum_ {j = 1} ^ {w ^ {\\mathrm {a}}} \\boldsymbol {\\Omega} _ {i, j} ^ {\\mathrm {a}} \\boldsymbol {P} _ {\\text {i n d} ^ {\\mathrm {a}} \\left(b _ {k}, j\\right), a _ {r + i - 1}}, \\tag {3}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.582, + 0.687, + 0.809, + 0.716 + ], + "angle": 0, + "content": "\\[\n\\operatorname {i n d} ^ {\\mathrm {a}} \\left(b _ {k}, j\\right) = b _ {k} - \\left\\lfloor \\frac {w ^ {\\mathrm {a}}}{2} \\right\\rfloor + j - 1,\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.725, + 0.892, + 0.786 + ], + "angle": 0, + "content": "where \\(\\mathrm{ind}^{\\mathrm{a}}(b_k,j)\\) is the index transform from the index \\(j\\) of the local window centered at \\(b_{k}\\) to the global timestamp index. Finally, the class-agnostic scores \\(\\mathcal{V}^{\\mathrm{b}}\\) are added back to \\(\\pmb{V}^{\\mathrm{a}}\\) to produce the final score matrix \\(\\pmb {V}\\in \\mathbb{R}^{K\\times (M - 1)}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.626, + 0.797, + 0.891, + 0.815 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {V} _ {k, r} = \\boldsymbol {V} _ {k, r} ^ {\\mathrm {a}} + v _ {b _ {k}} ^ {\\mathrm {b}}. \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.826, + 0.891, + 0.871 + ], + "angle": 0, + "content": "The above equation means that the class-agnostic boundary score of the \\(k\\)-th boundary \\(v_{b_k}^{\\mathrm{b}}\\) is added to all transition scores corresponding to the \\(k\\)-th boundary \\(\\mathbf{V}_{k,r}^{\\mathrm{a}}, \\forall r\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.871, + 0.892, + 0.9 + ], + "angle": 0, + "content": "- Action Transition Alignment. The final step is to find exact \\( M - 1 \\) optimal boundaries from the candidates \\( \\widetilde{B} \\) based" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18256" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.078, + 0.089, + 0.47, + 0.167 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.179, + 0.471, + 0.249 + ], + "angle": 0, + "content": "Figure 6. An example of action transition alignment between \\(\\widetilde{\\mathcal{B}}\\) of length 5 and \\(\\mathcal{R}\\) of length 3. (a) The cost matrix \\(\\Delta\\). (b) The cumulative cost matrix \\(D\\). Blue areas are directly initialized. Yellow areas correspond to optimal alignment. Red arrows are allowed path directions. (c) The optimal alignment. Best viewed in color." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.272, + 0.471, + 0.378 + ], + "angle": 0, + "content": "on the score matrix \\(V\\). It is equivalent to seek an one-to-one alignment with lowest cost between the transitions \\(\\mathcal{R}\\) and candidates \\(\\widetilde{\\mathcal{B}}\\) while requiring \\(K - M + 1\\) candidates to be dropped (Fig. 5). Formally, if we denote the optimal aligned boundary set as \\(\\mathcal{B} = \\{b_{k_r}\\}_{r=1}^{M-1}\\), where \\(b_{k_r}\\) corresponds to the \\(r\\)-th transition and \\(1 \\leq k_1 < \\ldots < k_{M-1} \\leq K\\). It should satisfy:" + }, + { + "type": "equation", + "bbox": [ + 0.117, + 0.385, + 0.47, + 0.426 + ], + "angle": 0, + "content": "\\[\n\\mathcal {B} = \\arg \\min _ {\\mathcal {B} ^ {\\prime}} \\psi \\left(\\mathcal {B} ^ {\\prime}\\right), \\quad \\psi \\left(\\mathcal {B} ^ {\\prime}\\right) = - \\sum_ {r = 1} ^ {M - 1} \\boldsymbol {V} _ {k _ {r} ^ {\\prime}, r}, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.433, + 0.47, + 0.463 + ], + "angle": 0, + "content": "where \\(\\psi (\\mathcal{B}^{\\prime})\\) is the cost function of an alignment and \\(\\mathcal{B}'\\) is any feasible aligned boundary set such as \\(\\mathcal{B}\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.464, + 0.47, + 0.644 + ], + "angle": 0, + "content": "To solve Eq. (5), we propose a DP algorithm, of which an example is shown in Fig. 6. To allow some candidates in \\(\\widetilde{B}\\) to be dropped, inspired by [36], we first expand the transition sequence \\(\\mathcal{R}\\) by inserting the empty symbol \\(\\phi\\) interleaved with transition symbols, obtaining \\(\\mathcal{R}' = [\\phi, (a_1, a_2), \\phi, (a_2, a_3), \\phi, \\dots, \\phi, (a_{M-1}, a_M), \\phi]\\). Now the candidate boundary matched to \\(\\phi\\) will be discarded. Then, a cost matrix \\(\\Delta \\in \\mathbb{R}^{K \\times (2(M-1) + 1)}\\) is constructed (Fig. 6(a)), whose \\((k, r')\\)-th entry is the cost of aligning the \\(b_k\\) with the \\(r'\\)-th symbol in \\(\\mathcal{R}'\\). For even number of \\(r'\\) (transition symbols), it is clear that \\(\\Delta_{k,r'} = -V_{k,r'/2}\\) from Eq. (5). Odd \\(r'\\) means dropping the candidate, and the cost is set to 0." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.646, + 0.471, + 0.84 + ], + "angle": 0, + "content": "A valid alignment is represented by a path from the top-left to bottom-right of \\(\\Delta\\) with constraint on directions. Specifically, the allowed start positions are \\((1,1)\\) and \\((1,2)\\), respectively corresponding to two cases where \\(b_{1}\\) is dropped and \\(b_{1}\\) is matched with the first transition. Similarly, there are two allowed end positions: \\((K,2(M - 1))\\) and \\((K,2(M - 1) + 1)\\). Then for the position \\((k,r^{\\prime})\\) on the path, if \\(r^{\\prime}\\) is odd, its previous position is either \\((k - 1,r^{\\prime})\\) or \\((k - 1,r^{\\prime} - 1)\\). It means that if \\(b_{k}\\) is dropped, the \\(b_{k - 1}\\) is also dropped or matched with the previous transition. On the other hand, the previous position is either \\((k - 1,r^{\\prime} - 1)\\) or \\((k - 1,r^{\\prime} - 2)\\) if \\(r^{\\prime}\\) is even, with the similar meaning. These allowed directions are indicated by red arrows in Fig. 6(b)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.471, + 0.902 + ], + "angle": 0, + "content": "The DP algorithm generally finds the optimal solution by computing a cumulative cost matrix \\( \\pmb{D} \\) with the same shape as \\( \\pmb{\\Delta} \\) (Fig. 6(b)), where \\( D_{k,r'} \\) represents the minimum cumulative cost of all valid paths ending at \\( (k, r') \\). It" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.892, + 0.122 + ], + "angle": 0, + "content": "is computed by the following recursive equation, considering paths coming from valid directions aforementioned:" + }, + { + "type": "equation", + "bbox": [ + 0.5, + 0.133, + 0.892, + 0.181 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {D} _ {k, r ^ {\\prime}} = \\boldsymbol {\\Delta} _ {k, r ^ {\\prime}} + \\left\\{ \\begin{array}{l} \\min \\left(\\boldsymbol {D} _ {k - 1, r ^ {\\prime}}, \\boldsymbol {D} _ {k - 1, r ^ {\\prime} - 1}\\right), r ^ {\\prime} \\text {i s o d d}, \\\\ \\min \\left(\\boldsymbol {D} _ {k - 1, r ^ {\\prime} - 1}, \\boldsymbol {D} _ {k - 1, r ^ {\\prime} - 2}\\right), r ^ {\\prime} \\text {i s e v e n}. \\end{array} \\right. \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.182, + 0.892, + 0.242 + ], + "angle": 0, + "content": "The above equation is not applicable to the first row and the first two columns of \\( D \\). They are directly initialized before the computation (blue areas of Fig. 6(b)). We set the cost of invalid positions to \\( \\infty \\) (inf) and others are copied from \\( \\Delta \\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.243, + 0.892, + 0.332 + ], + "angle": 0, + "content": "The cumulative cost of the optimal alignment is \\(\\min (D_{K,2(M - 1)},D_{K,2(M - 1) + 1})\\), The complete alignment, i.e., the optimal boundary set \\(\\mathcal{B}\\) can be obtained by starting from the corresponding position of the above min operation and backtracking (Fig. 6(b/c)). More details can be found in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.333, + 0.892, + 0.469 + ], + "angle": 0, + "content": "- Complexity. Clearly, the class-agnostic boundary scoring and action transition scoring can be implemented by the unfold operator and matrix operations, thus can be parallelized on GPUs. The action transition alignment must be performed serially with a time complexity of \\( O(KM) \\) or \\( O(\\lambda M^2) \\). Compared with the previous alignment methods such as Viterbi (\\( O(T^2M) \\) [25, 28, 34]) and DTW (\\( O(TM) \\)), it decreases from a function involving video length \\( T \\) to one of only transcript length \\( M \\) (\\( M << T \\))." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.478, + 0.684, + 0.492 + ], + "angle": 0, + "content": "3.5. Video-Level Losses" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.501, + 0.892, + 0.745 + ], + "angle": 0, + "content": "Inevitably, the inferred pseudo segmentation contains some degree of noise, which can be unexpectedly fit by the network. To improve the semantic robustness, we propose to jointly train a video-level multi-label classification task that predicts whether each action appears in the video, which is precisely indicated by the transcript. Specifically, inspired by the common practice of Transformer [7, 10, 19, 24, 43], the input sequence \\( \\mathbf{X} \\) is augmented by a set of \\( |\\mathcal{C}| \\) learnable class tokens which are fed into the same network and responsible for predicting action occurrence. The outputs, denoted as \\( \\{e_c'\\}_{c=1}^{|\\mathcal{C}|} \\), are then used to predict the action occurrence probability via \\( \\xi_c = \\sigma(\\boldsymbol{w}_c^T \\boldsymbol{e}_c' + \\epsilon_c) \\) for each category \\( c \\), where \\( \\boldsymbol{w}_c \\) and \\( \\epsilon_c \\) are the weight vector and bias scalar for class \\( c \\) in the classifier shared with frame-wise classification. \\( \\sigma(\\cdot) \\) is the sigmoid activation. These predictions are supervised by the binary cross entropy:" + }, + { + "type": "equation", + "bbox": [ + 0.508, + 0.756, + 0.892, + 0.799 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {v i d}} = - \\frac {1}{| \\mathcal {C} |} \\sum_ {c = 1} ^ {| \\mathcal {C} |} \\left[ y _ {c} ^ {\\mathrm {v i d}} \\log \\xi_ {c} + \\left(1 - y _ {c} ^ {\\mathrm {v i d}}\\right) \\log \\left(1 - \\xi_ {c}\\right) \\right], \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.81, + 0.892, + 0.87 + ], + "angle": 0, + "content": "where \\( y_c^{\\mathrm{vid}} \\in \\{0,1\\} \\) is the binary label indicating whether action \\( c \\) appears in the video. Through the shared network and the interactions between tokens, the semantics learned from video-level training can benefit the main task." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.871, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Moreover, we treat the class tokens with global information as the prototypes of each action and attempt to align the" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18257" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.152 + ], + "angle": 0, + "content": "frame features to them for semantic consistency. Specifically, we first calculate the feature centroid of each category \\( c \\) that appears in the video by averaging the output frame features with the corresponding pseudo labels:" + }, + { + "type": "equation", + "bbox": [ + 0.143, + 0.163, + 0.469, + 0.203 + ], + "angle": 0, + "content": "\\[\n\\overline {{\\boldsymbol {x}}} _ {c} = \\frac {\\sum_ {t = 1} ^ {T} \\mathbb {I} \\left(\\widetilde {y} _ {t} = c\\right) \\boldsymbol {x} _ {t} ^ {\\prime}}{\\sum_ {t = 1} ^ {T} \\mathbb {I} \\left(\\widetilde {y} _ {t} = c\\right)}, c \\in \\operatorname {S e t} (\\mathcal {A}), \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.213, + 0.47, + 0.274 + ], + "angle": 0, + "content": "where the \\(\\operatorname{Set}(\\cdot)\\) operator converts a list into an unordered set and removes duplicates. These centroids will move towards the corresponding \\(e_c'\\) under the guidance of a global-local contrastive loss in the form of InfoNCE [17, 31]:" + }, + { + "type": "equation", + "bbox": [ + 0.079, + 0.284, + 0.469, + 0.337 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {g l c}} = - \\frac {1}{| \\operatorname {S e t} (\\mathcal {A}) |} \\sum_ {c \\in \\operatorname {S e t} (\\mathcal {A})} \\log \\frac {\\exp \\left(\\left\\langle \\overline {{\\boldsymbol {x}}} _ {c} \\right\\rangle^ {T} \\left\\langle \\boldsymbol {e} _ {c} ^ {\\prime} \\right\\rangle / \\tau\\right)}{\\sum_ {c ^ {\\prime} = 1} ^ {| \\mathcal {C} |} \\exp \\left(\\left\\langle \\overline {{\\boldsymbol {x}}} _ {c} \\right\\rangle^ {T} \\left\\langle \\boldsymbol {e} _ {c ^ {\\prime}} ^ {\\prime} \\right\\rangle / \\tau\\right)}, \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.337, + 0.469, + 0.368 + ], + "angle": 0, + "content": "where \\(\\langle \\pmb{x}\\rangle = \\pmb{x} / \\| \\pmb{x}\\| _2\\) is the \\(l_{2}\\)-normalization operator, and \\(\\tau\\) is the temperature hyper-parameter." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.378, + 0.291, + 0.394 + ], + "angle": 0, + "content": "3.6. Training and Inference" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.401, + 0.469, + 0.536 + ], + "angle": 0, + "content": "- Training. Since the pseudo labeling always requires a relatively good initialization, we utilize a two-stage training strategy. The first stage only uses the reliable labels, i.e., the video-level labels, to pretrain the network. Hence the loss function is \\(\\mathcal{L}_{\\mathrm{I}} = \\mathcal{L}_{\\mathrm{vid}}\\). The loss for the second stage further includes the frame-wise losses on both classification and representation using pseudo labels: \\(\\mathcal{L}_{\\mathrm{II}} = \\alpha \\mathcal{L}_{\\mathrm{vid}} + \\beta \\mathcal{L}_{\\mathrm{cls}} + \\gamma \\mathcal{L}_{\\mathrm{glc}}\\), where \\(\\alpha\\), \\(\\beta\\) and \\(\\gamma\\) are all hyperparameters." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.537, + 0.469, + 0.688 + ], + "angle": 0, + "content": "- Inference. For inference, we obtain the action label \\(\\hat{y}_t\\) for frame \\(t\\) directly from the frame-wise class probabilities: \\(\\hat{y}_t = \\arg \\max_c P_{t,c}\\). Note that we do not require any alignment processing during inference, which is performed by previous WSAS methods for smoother predictions. However, since the ground-truth transcript is not available during inference, these methods either run alignment with every transcript from the training set [4, 5, 25, 28, 34], or predict the transcript via another model [38], which makes the pipeline more time-consuming and less practical." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.702, + 0.209, + 0.718 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.727, + 0.267, + 0.743 + ], + "angle": 0, + "content": "4.1. Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.469, + 0.901 + ], + "angle": 0, + "content": "- Datasets. We perform experiments on three datasets. The Breakfast [21] dataset contains 1712 videos of breakfast cooking with 48 different actions. On average, each video has 6.8 segments and \\(7.3\\%\\) frames are background. The Hollywood Extended [2] dataset contains 937 videos taken from movies with 16 categories of daily actions such as walk or sit. On average, each video has 5.9 segments and \\(60.9\\%\\) frames are background. The CrossTask [47] dataset contains videos from 18 primary tasks. Following [28], 14 cooking-related tasks are selected, which have 2552 videos" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.182 + ], + "angle": 0, + "content": "and 80 action categories. On average, each video has 14.4 segments and \\(74.8\\%\\) frames are background. For Breakfast, we use the released 4 training/test splits and report the average. For Hollywood, we perform a 10-fold cross-validation. For CrossTask, we use the released training/test split. These evaluation protocols are consistent with previous methods." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.188, + 0.893, + 0.43 + ], + "angle": 0, + "content": "- Metrics. To evaluate our method, we use 4 standard metrics, following [5, 9, 25, 28]. (1) The Mean-over-Frames (MoF) is the percentage of frames whose labels are correctly predicted. (2) The Mean-over-Frames without Background (MoF-Bg) is the MoF over non-background frames. It is more suitable than MoF for the datasets with high background rate such as Hollywood and CrossTask. (3) The Intersection-over-Union (IoU) is defined as \\( |I \\cap I^{*}| / |I \\cup I^{*}| \\) while (4) the Intersection-over-Detection (IoD) is \\( |I \\cap I^{*}| / |I| \\), where \\( I^{*} \\) and \\( I \\) are the ground-truth (GT) segment and the predicted segment with the same class, respectively. For each GT segment, the highest IoU/IoD with one predicted segment are preserved, and the average of all GT segments is reported. Note that the definition of the IoU/IoD in [28] is different from other works [5, 9, 25], and will be indicated with special symbols when used." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.436, + 0.892, + 0.648 + ], + "angle": 0, + "content": "- Input Features. As the input features \\( X \\) for Breakfast and Hollywood, we use the 2048-dimensional RGB+flow I3D features [3] adopted by MuCon [38] and most FSAS methods [13, 26, 27, 45], while some WSAS methods [5, 25, 28, 34] still use the iDT features [41]. Since recent studies [37, 38] have found that they perform worse with more advanced I3D features, we still report the performance with the iDT features for them following [38]. For CrossTask, the officially released 3200-dimensional features are adopted, but we do not perform dimension reduction via PCA as with [28]. For GPU memory efficiency, we perform a \\( 10 \\times \\) temporal downsampling for Breakfast and \\( 5 \\times \\) for Hollywood. During inference, the output is upsampled to match the original video/ground-truth length." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.653, + 0.892, + 0.82 + ], + "angle": 0, + "content": "- Implementation Details. We adopt a 6-layer Transformer [40] with single-head self-attention as the temporal network, and the latent dimension \\( d' \\) is 256. For the ATBA, we set \\( w^{\\mathrm{b}} = 7 \\) and \\( \\lambda = 4 \\) for all datasets. \\( w^{\\mathrm{a}} \\) is set to 31 for Breakfast/CrossTask and 23 for Hollywood. \\( \\mu \\) is set to 0.3 for Breakfast/Hollywood and 0.1 for CrossTask. The loss weights \\( \\alpha \\), \\( \\beta \\) are set to 1.0 and \\( \\gamma \\) is 0.1. \\( \\tau \\) in Eq. (9) is set to 0.2. Besides, to alleviate the issue that the predictions on Hollywood/CrossTask are overly dominated by background, we lower the sample weights of pseudo background frames to 0.8 in Eq. (1) for these datasets." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.826, + 0.892, + 0.9 + ], + "angle": 0, + "content": "- Multiple Runs. Due to the alternating nature of learning from weak supervision, there are often fluctuations in the results of WSAS methods [28, 38]. Hence, following [38], we report the average and standard deviation over 5 runs with different random seeds for better evaluation." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18258" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.082, + 0.089, + 0.47, + 0.277 + ], + "angle": 0, + "content": "
Breakfast
MethodMoF±stdMoF-Bg±stdIoU±stdIoD±std
HMM+RNN [33]33.3---
[33]+Length [23]36.7---
TCFPN+ISBA [9]38.4/36.4±1.0*38.424.240.6
NN-Viterbi [34]43.0/39.7±2.4*---
D3TW [4]45.7---
CDFL [25]50.2/48.1±2.5*48.033.745.4
DPDTW [5]50.8-35.645.1
TASL [28]♠47.8-35.2†46.1†
MuCon [38]♠48.5±1.850.3*40.9*54.0*
POC [29]♠45.7-38.3†-
AdaAct [46]51.248.336.346.4
ATBA♠(Ours)53.9±1.254.4±1.241.1±0.761.7±1.1
39.5±0.8†55.9±1.0†
" + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.277, + 0.47, + 0.432 + ], + "angle": 0, + "content": "
Hollywood Extended
MethodMoF±stdMoF-Bg±stdIoU±stdIoD±std
HMM+RNN [33]--11.9-
[33]+Length [23]--12.3-
TCFPN+ISBA [9]28.734.512.618.3
D3TW [4]33.6---
CDFL [25]45.040.619.525.8
DPDTW [5]55.625.6#33.243.3
TASL [28]42.1#27.2#23.3†#33.0†#
MuCon [38]♠-41.613.9*-
ATBA♠(Ours)47.7±2.540.2±1.630.9±1.655.8±0.8
28.5±1.6†44.9±0.5†
" + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.432, + 0.47, + 0.541 + ], + "angle": 0, + "content": "
CrossTask
MethodMoF±stdMoF-Bg±stdIoU±stdIoD±std
NN-Viterbi [34]♠26.5*-10.7†*24.0†*
CDFL [25]♠31.9*-11.5†*23.8†*
TASL [28]♠40.727.4#14.5†25.1†
POC [29]♠42.817.6#15.6†-
ATBA♠(Ours)50.6±1.331.3±0.720.9±0.444.6±0.7
15.7±0.3†24.6±0.4†
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.076, + 0.551, + 0.47, + 0.676 + ], + "angle": 0, + "content": "Table 1. Comparisons of ours with other WSAS methods on three datasets. std is the standard deviation of multiple runs (if any). \\(\\dagger\\) The metric is computed by the definition of [28]. *-Results are reported by other works. Please refer to the supplementary material for detailed sources. \\(\\sharp\\)-Results are obtained by us via rerunning the open sources (The TASL [28] does not follow the common 10-fold evaluation protocol for Hollywood so we re-produce the results). \\(\\spadesuit\\) The reported results are the average of multiple runs. Best results are in bold, second best are underlined." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.707, + 0.403, + 0.723 + ], + "angle": 0, + "content": "4.2. Comparison with the State-of-the-Art" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.732, + 0.469, + 0.839 + ], + "angle": 0, + "content": "- Performance. In Table 1, we compare our proposed method with previous WSAS methods. Our ATBA achieves state-of-the-art (SOTA) by a clear margin \\((+2.7\\% \\mathrm{MoF})\\) on the Breakfast [21], demonstrating the effectiveness of focusing on action transitions. Moreover, comparing the standard deviation with other methods, it can be found that our method is also relatively stable." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Our ATBA also achieves comparable performance on the Hollywood [2]. The reason why our method does not show significant advantage is probably because this dataset is collected from movies and so contains many shot changes, re" + }, + { + "type": "table", + "bbox": [ + 0.525, + 0.09, + 0.869, + 0.235 + ], + "angle": 0, + "content": "
MethodTr.A.MoFTraining (Hours)Inference (Seconds)
TCFPN+ISBA [9]✘33.312.75*0.01*
NN-Viterbi [34]V43.011.23*56.25*
CDFL [25]V50.266.73*62.37*
DPDTW [5]D50.831.02‡0.69‡
TASL [28]V47.824.66‡54.99‡
MuCon [38]✘48.54.573.03
POC [29]✘45.72.28‡0.01‡
ATBA (Ours)B53.93.450.01
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.498, + 0.246, + 0.892, + 0.356 + ], + "angle": 0, + "content": "Table 2. Comparison of accuracy, training and inference time on the Breakfast. The training time is measured as the entire training duration on the split 1, and the inference time is measured as the average time for inferring a video from the test set of the split 1. Tr.A.-The alignment algorithm adopted during training (X-No Alignment. V-Viterbi. D-DTW. B-Boundary Alignment). *-Measured by [38]. ‡-Measured by us. Best results are in bold, second best are underlined." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.398, + 0.892, + 0.503 + ], + "angle": 0, + "content": "sulting in more noisy boundaries. Note that although the DPDTW [5] achieves significantly high MoF, this metric is severely biased in case of highly imbalanced categories (60.9% frames are background). We report the MoF-Bg metric for it, on which it performs poorly, proving that it cannot recognize real actions very well. In contrast, the performance of our method is more balanced." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.51, + 0.892, + 0.571 + ], + "angle": 0, + "content": "For the more difficult dataset CrossTask [47] with the most segments and highest background rate, our approach also outperforms previous methods whether or not the metric involves background (\\(+7.8\\%\\) MoF and \\(+3.9\\%\\) MoF-Bg)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.577, + 0.892, + 0.849 + ], + "angle": 0, + "content": "- Efficiency. Besides of the performance, we also compare the training time of our approach with previous WSAS methods to show the efficiency of ours. Following [38], we train our model on an Nvidia GeForce GTX 1080Ti GPU, and the training time is measured as the wall time over the whole training phase, during which any irrelevant operations such as intermediate evaluation and saving checkpoints are removed. As all the WSAS methods directly load pre-computed features, the time measurement also does not include the time to extract features from raw videos. As shown in Table 2, our ATBA achieves the best performance with the second shortest training time, which demonstrates the effectiveness of our design. Specifically, the training speed of ours is on average 10 times (varying from 3 to 20) faster than methods performing frame-by-frame alignment [5, 25, 28, 34]. Compared to the alignment-free methods with comparable training speed [29, 38], ours achieves better performance." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.892, + 0.901 + ], + "angle": 0, + "content": "We also compare the time to infer a single test video. As mentioned in Sec. 3.6, ours does not require any alignment processing, which makes it the fastest during inference." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18259" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.121, + 0.089, + 0.428, + 0.171 + ], + "angle": 0, + "content": "
Exp.\\( \\gamma^b \\)Cost Mat. \\( V^a \\)P.L.MoFIoUIoD
1✓38.632.323.152.5
2✓✓51.741.330.247.8
3✓○✓67.954.041.1
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.181, + 0.471, + 0.278 + ], + "angle": 0, + "content": "Table 3. Ablation studies of ATBA on the Breakfast. Exp.-Different experiment configurations. Cost Mat.-Different choices for the cost matrix in action transition alignment. o-According to Eq. (4), \\( V \\) contains \\( V^{\\mathrm{a}} \\). P.L.-The accuracy of pseudo labels during training. Exp.1 means that only the class-agnostic boundary detection is adopted to localize action transitions. Exp.3 is our default setting. Best results are in bold." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.307, + 0.241, + 0.322 + ], + "angle": 0, + "content": "4.3. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.332, + 0.47, + 0.588 + ], + "angle": 0, + "content": "- Effect of ATBA. In Table 3, we conducted an in-depth evaluation of our proposed ATBA. From Exp.1, the pseudo label quality and the evaluation performance are both very poor when only using the class-agnostic boundary scores \\(\\nu^{\\mathrm{b}}\\) to directly select \\(M - 1\\) boundaries for transitions via the greedy strategy stated in Sec. 3.4, showing that it is critical to take the class-specific transition pattern into account as with our design to suppress the noisy boundaries for more precise pseudo segmentation. Fig. 2 provides an intuitive example. In addition, comparing Exp.2 and 3, it's better to involve the class-agnostic boundary scores in the action transition alignment. We think it is because that some candidates can have very low boundary scores (unlikely to be a boundary) as the candidate selection only depends on ranking, and when the transition patterns are not discriminative at the beginning of training, these candidates may be unexpectedly aligned if the boundary scores are not involved." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.589, + 0.47, + 0.741 + ], + "angle": 0, + "content": "- Effect of Video-Level Losses. In Table 4, we evaluate the \\(\\mathcal{L}_{\\mathrm{vid}}\\) and \\(\\mathcal{L}_{\\mathrm{glc}}\\). Comparing Exp.1 and 2, the model with \\(\\mathcal{L}_{\\mathrm{vid}}\\) achieves higher performance (\\(+6.5\\%\\) MoF) at the same level of pseudo label accuracy, demonstrating that the video-level supervision can promote the learning of precise action semantics. In Exp.3, the explicit representation alignment \\((\\mathcal{L}_{\\mathrm{glc}})\\) further improves the quality of frame features and thus the performance (\\(+1.6\\%\\) MoF). Moreover, Exp.4 shows that whether or not to share the classifier between frame-/video-level classification has little impact." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.742, + 0.469, + 0.788 + ], + "angle": 0, + "content": "We also provide the analysis on the effect of important hyper-parameters in the supplementary material, including \\( w^{\\mathrm{b}} \\), \\( w^{\\mathrm{a}} \\), \\( \\mu \\) and \\( \\lambda \\) in the ATBA." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.801, + 0.26, + 0.817 + ], + "angle": 0, + "content": "4.4. Qualitative Results" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.826, + 0.469, + 0.901 + ], + "angle": 0, + "content": "We show the qualitative inference result compared with two recent open source WSAS methods, i.e., MuCon [38] and TASL [28], in Fig. 7. Our ATBA achieves significantly more accurate results in this challenging video with many segments. The result of TASL shows order reversal (\"Crake" + }, + { + "type": "table", + "bbox": [ + 0.534, + 0.09, + 0.86, + 0.173 + ], + "angle": 0, + "content": "
Exp.\\( \\mathcal{L}_{\\text{vid}} \\)\\( \\mathcal{L}_{\\text{glc}} \\)D.Cls.P.L.MoFIoUIoD
167.545.940.561.9
2✓66.652.438.861.4
3✓✓67.954.041.162.3
4✓✓✓68.153.740.161.7
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.184, + 0.892, + 0.254 + ], + "angle": 0, + "content": "Table 4. Ablation studies of video-level losses on the Breakfast. Exp.-Different experiment configurations. D.Cls.-Different classifiers for frame-wise prediction and action occurrence prediction. P.L.-The accuracy of pseudo labels during training. Exp.3 is our default setting. Best results are in bold." + }, + { + "type": "image", + "bbox": [ + 0.504, + 0.271, + 0.891, + 0.425 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.434, + 0.892, + 0.476 + ], + "angle": 0, + "content": "Figure 7. Qualitative results on the Breakfast. The example test video is P15-stereo01-P15-pancake. We compare our inference result with two recent methods. Best viewed in color." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.506, + 0.892, + 0.598 + ], + "angle": 0, + "content": "Egg\" and \"Spoon Flour\") and hallucination (\"Butter Pan\"). The MuCon predicts the correct action ordering, but the result deviates severely. In contrast, ours successfully predicts an accurate segmentation, indicating the action semantics are well learned. More qualitative results are provided in supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.616, + 0.619, + 0.632 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.643, + 0.892, + 0.794 + ], + "angle": 0, + "content": "In this work, we propose to directly localize action transitions for efficient pseudo segmentation generation in the WSAS task, thus avoiding the time-consuming frame-by-frame alignment. Due to the presence of noisy boundaries, a novel Action-Transition-Aware Boundary Alignment (ATBA) framework is proposed to efficiently and effectively filter out noise and detect transitions. Moreover, we also design some video-level losses to utilize video-level supervision to improve the semantic robustness. Extensive experiments show the effectiveness of our ATBA." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.814, + 0.671, + 0.831 + ], + "angle": 0, + "content": "6. Acknowledgment" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.84, + 0.892, + 0.901 + ], + "angle": 0, + "content": "This work was supported partially by the NSFC (U21A20471), National Key Research and Development Program of China (2023YFA1008503), Guangdong NSF Project (No. 2023B1515040025, 2020B1515120085)." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18260" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.468, + 0.157 + ], + "angle": 0, + "content": "[1] Emad Bahrami, Gianpiero Francesca, and Juergen Gall. How much temporal long-term context is needed for action segmentation? arXiv preprint arXiv:2308.11358, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.159, + 0.468, + 0.228 + ], + "angle": 0, + "content": "[2] Piotr Bojanowski, Rémi Lajugie, Francis Bach, Ivan Laptev, Jean Ponce, Cordelia Schmid, and Josef Sivic. Weakly supervised action labeling in videos under ordering constraints. In Proceedings of the European Conference on Computer Vision, pages 628-643. Springer, 2014. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.231, + 0.468, + 0.286 + ], + "angle": 0, + "content": "[3] Joao Carreira and Andrew Zisserman. Quo vadis, action recognition? a new model and the kinetics dataset. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 6299-6308, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.288, + 0.468, + 0.37 + ], + "angle": 0, + "content": "[4] Chien-Yi Chang, De-An Huang, Yanan Sui, Li Fei-Fei, and Juan Carlos Niebles. D3tw: Discriminative differentiable dynamic time warping for weakly supervised action alignment and segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3546–3555, 2019. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.373, + 0.468, + 0.441 + ], + "angle": 0, + "content": "[5] Xiaobin Chang, Frederick Tung, and Greg Mori. Learning discriminative prototypes with dynamic time warping. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8395-8404, 2021. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.445, + 0.47, + 0.514 + ], + "angle": 0, + "content": "[6] Min-Hung Chen, Baopu Li, Yingze Bao, Ghassan Al-Regib, and Zsolt Kira. Action segmentation with joint self-supervised temporal domain adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9454–9463, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.516, + 0.468, + 0.571 + ], + "angle": 0, + "content": "[7] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.574, + 0.468, + 0.629 + ], + "angle": 0, + "content": "[8] Guodong Ding and Angela Yao. Leveraging action affinity and continuity for semi-supervised temporal action segmentation. In European Conference on Computer Vision, pages 17-32. Springer, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.631, + 0.468, + 0.687 + ], + "angle": 0, + "content": "[9] Li Ding and Chenliang Xu. Weakly-supervised action segmentation with iterative soft boundary assignment. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6508-6516, 2018. 1, 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.689, + 0.468, + 0.771 + ], + "angle": 0, + "content": "[10] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020.5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.774, + 0.468, + 0.829 + ], + "angle": 0, + "content": "[11] Dazhao Du, Bing Su, Yu Li, Zhongang Qi, Lingyu Si, and Ying Shan. Efficient u-transformer with boundary-aware loss for action segmentation. arXiv preprint arXiv:2205.13425, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.832, + 0.468, + 0.9 + ], + "angle": 0, + "content": "[12] Zexing Du, Xue Wang, Guoqing Zhou, and Qing Wang. Fast and unsupervised action boundary detection for action segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3323-3332, 2022. 1, 2, 4" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.47, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.148 + ], + "angle": 0, + "content": "[13] Yazan Abu Farha and Jurgen Gall. Ms-tcn: Multi-stage temporal convolutional network for action segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3575-3584, 2019. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.15, + 0.894, + 0.219 + ], + "angle": 0, + "content": "[14] Shang-Hua Gao, Qi Han, Zhong-Yu Li, Pai Peng, Liang Wang, and Ming-Ming Cheng. Global2local: Efficient structure search for video action segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16805-16814, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.221, + 0.892, + 0.29 + ], + "angle": 0, + "content": "[15] Reza Ghoddoosian, Isht Dwivedi, Nakul Agarwal, Chiho Choi, and Behzad Dariush. Weakly-supervised online action segmentation in multi-view instructional videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13780-13790, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.292, + 0.892, + 0.36 + ], + "angle": 0, + "content": "[16] Reza Ghoddoosian, Isht Dwivedi, Nakul Agarwal, and Behzad Dariush. Weakly-supervised action segmentation and unseen error detection in anomalous instructional videos. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10128-10138, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.362, + 0.892, + 0.431 + ], + "angle": 0, + "content": "[17] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9729-9738, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.433, + 0.892, + 0.489 + ], + "angle": 0, + "content": "[18] De-An Huang, Li Fei-Fei, and Juan Carlos Niebles. Connectionist temporal modeling for weakly supervised action labeling. In Proceedings of the European Conference on Computer Vision, pages 137–153. Springer, 2016. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.49, + 0.892, + 0.558 + ], + "angle": 0, + "content": "[19] Zi-Hang Jiang, Qibin Hou, Li Yuan, Daquan Zhou, Yujun Shi, Xiaojie Jin, Anran Wang, and Jiashi Feng. All tokens matter: Token labeling for training better vision transformers. Advances in neural information processing systems, 34: 18590-18602, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.561, + 0.892, + 0.63 + ], + "angle": 0, + "content": "[20] Hyolim Kang, Jinwoo Kim, Taehyun Kim, and Seon Joo Kim. Uboco: Unsupervised boundary contrastive learning for generic event boundary detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20073-20082, 2022. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.632, + 0.892, + 0.7 + ], + "angle": 0, + "content": "[21] Hilde Kuehne, Ali Arslan, and Thomas Serre. The language of actions: Recovering the syntax and semantics of goal-directed human activities. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 780-787, 2014. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.702, + 0.892, + 0.744 + ], + "angle": 0, + "content": "[22] Hilde Kuehne, Alexander Richard, and Juergen Gall. Weakly supervised learning of actions from transcripts. Computer Vision and Image Understanding, 163:78-89, 2017. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.746, + 0.892, + 0.801 + ], + "angle": 0, + "content": "[23] Hilde Kuehne, Alexander Richard, and Juergen Gall. A hybrid rnn-hmm approach for weakly supervised temporal action segmentation. IEEE transactions on pattern analysis and machine intelligence, 42(4):765-779, 2018. 1, 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.803, + 0.892, + 0.87 + ], + "angle": 0, + "content": "[24] Jack Lanchantin, Tianlu Wang, Vicente Ordonez, and Yanjun Qi. General multi-label image classification with transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16478-16488, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.873, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[25] Jun Li, Peng Lei, and Sinisa Todorovic. Weakly supervised energy-based learning for action segmentation. In Proceed-" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.518, + 0.957 + ], + "angle": 0, + "content": "18261" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.092, + 0.468, + 0.12 + ], + "angle": 0, + "content": "ings of the IEEE/CVF International Conference on Computer Vision, pages 6243-6251, 2019. 1, 2, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.122, + 0.469, + 0.177 + ], + "angle": 0, + "content": "[26] Shi-Jie Li, Yazan AbuFarha, Yun Liu, Ming-Ming Cheng, and Juergen Gall. Ms-tcn++: Multi-stage temporal convolutional network for action segmentation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.179, + 0.469, + 0.22 + ], + "angle": 0, + "content": "[27] Daochang Liu, Qiyue Li, AnhDung Dinh, Tingting Jiang, Mubarak Shah, and Chang Xu. Diffusion action segmentation. arXiv preprint arXiv:2303.17959, 2023. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.222, + 0.469, + 0.29 + ], + "angle": 0, + "content": "[28] Zijia Lu and Ehsan Elhamifar. Weakly-supervised action segmentation and alignment via transcript-aware union-of-subspaces learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8085-8095, 2021. 1, 2, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.293, + 0.469, + 0.36 + ], + "angle": 0, + "content": "[29] Zijia Lu and Ehsan Elhamifar. Set-supervised action learning in procedural task videos via pairwise order consistency. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19903-19913, 2022. 1, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.363, + 0.469, + 0.446 + ], + "angle": 0, + "content": "[30] Antoine Miech, Dimitri Zhukov, Jean-Baptiste Alayrac, Makarand Tapaswi, Ivan Laptev, and Josef Sivic. Howto100m: Learning a text-video embedding by watching hundred million narrated video clips. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2630-2640, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.448, + 0.469, + 0.489 + ], + "angle": 0, + "content": "[31] Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.491, + 0.469, + 0.56 + ], + "angle": 0, + "content": "[32] Rahul Rahaman, Dipika Singhania, Alexandre Thiery, and Angela Yao. A generalized and robust framework for timestamp supervision in temporal action segmentation. In Proceedings of the European Conference on Computer Vision, pages 279–296. Springer, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.562, + 0.469, + 0.629 + ], + "angle": 0, + "content": "[33] Alexander Richard, Hilde Kuehne, and Juergen Gall. Weakly supervised action learning with rn based fine-to-coarse modeling. In Proceedings of the IEEE conference on Computer Vision and Pattern Recognition, pages 754-763, 2017. 1, 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.632, + 0.469, + 0.7 + ], + "angle": 0, + "content": "[34] Alexander Richard, Hilde Kuehne, Ahsan Iqbal, and Juergen Gall. Neuralnetwork-viterbi: A framework for weakly supervised video learning. In Proceedings of the IEEE conference on Computer Vision and Pattern Recognition, pages 7386-7395, 2018. 1, 2, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.703, + 0.469, + 0.758 + ], + "angle": 0, + "content": "[35] Yuhan Shen and Ehsan Elhamifar. Semi-weakly-supervised learning of complex actions from instructional task videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3344–3354, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.76, + 0.469, + 0.829 + ], + "angle": 0, + "content": "[36] Yuhan Shen, Lu Wang, and Ehsan Elhamifar. Learning to segment actions from visual and language instructions via differentiable weak sequence alignment. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10156-10165, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.831, + 0.469, + 0.871 + ], + "angle": 0, + "content": "[37] Yaser Souri, Alexander Richard, Luca Minciullo, and Juergen Gall. On evaluating weakly supervised action segmentation methods. arXiv preprint arXiv:2005.09743, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.874, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[38] Yaser Souri, Mohsen Fayyaz, Luca Minciullo, Gianpiero Francesca, and Juergen Gall. Fast weakly supervised action" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.469, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.892, + 0.134 + ], + "angle": 0, + "content": "segmentation using mutual consistency. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(10):6196-6208, 2021. 1, 2, 3, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.136, + 0.892, + 0.205 + ], + "angle": 0, + "content": "[39] Yaser Souri, Yazan Abu Farha, Fabien Despinoy, Gianpiero Francesca, and Juergen Gall. Fif: Fast inference approximation for action segmentation. In Proceedings of the DAGM German Conference on Pattern Recognition, pages 282-296. Springer, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.207, + 0.892, + 0.261 + ], + "angle": 0, + "content": "[40] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.263, + 0.892, + 0.316 + ], + "angle": 0, + "content": "[41] Heng Wang and Cordelia Schmid. Action recognition with improved trajectories. In Proceedings of the IEEE international conference on computer vision, pages 3551-3558, 2013. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.319, + 0.892, + 0.388 + ], + "angle": 0, + "content": "[42] Ruibin Xiong, Yunchang Yang, Di He, Kai Zheng, Shuxin Zheng, Chen Xing, Huishuai Zhang, Yanyan Lan, Liwei Wang, and Tieyan Liu. On layer normalization in the transformer architecture. In International Conference on Machine Learning, pages 10524-10533. PMLR, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.39, + 0.892, + 0.459 + ], + "angle": 0, + "content": "[43] Lian Xu, Wanli Ouyang, Mohammed Bennamoun, Farid Boussaid, and Dan Xu. Multi-class token transformer for weakly supervised semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4310-4319, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.461, + 0.892, + 0.542 + ], + "angle": 0, + "content": "[44] Di Yang, Yaohui Wang, Antitza Dantcheva, Quan Kong, Lorenzo Garattoni, Gianpiero Francesca, and Francois Bremond. Lac-latent action composition for skeleton-based action segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13679-13690, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.545, + 0.892, + 0.585 + ], + "angle": 0, + "content": "[45] Fangqiu Yi, Hongyu Wen, and Tingting Jiang. Asformer: Transformer for action segmentation. arXiv preprint arXiv:2110.08568, 2021. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.587, + 0.892, + 0.669 + ], + "angle": 0, + "content": "[46] Runzhong Zhang, Suchen Wang, Yueqi Duan, Yansong Tang, Yue Zhang, and Yap-Peng Tan. Hoi-aware adaptive network for weakly-supervised action segmentation. In Proceedings of the Thirty-Second International Joint Conference on Artificial Intelligence, pages 1722–1730, 2023. 1, 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.672, + 0.892, + 0.753 + ], + "angle": 0, + "content": "[47] Dimitri Zhukov, Jean-Baptiste Alayrac, Ramadan Gokberk Cinbis, David Fouhey, Ivan Laptev, and Josef Sivic. Cross-task weakly supervised learning from instructional videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3537-3545, 2019. 1, 2, 6, 7" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.753 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "18262" + } + ] +] \ No newline at end of file diff --git a/2024/Efficient and Effective Weakly-Supervised Action Segmentation via Action-Transition-Aware Boundary Alignment/1ae3fdc0-77c3-4dc9-87e2-e3b0568d2efa_origin.pdf b/2024/Efficient and Effective Weakly-Supervised Action Segmentation via Action-Transition-Aware Boundary Alignment/1ae3fdc0-77c3-4dc9-87e2-e3b0568d2efa_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..7035aa59355815451839e0af89a2b116a95280e3 --- /dev/null +++ b/2024/Efficient and Effective Weakly-Supervised Action Segmentation via Action-Transition-Aware Boundary Alignment/1ae3fdc0-77c3-4dc9-87e2-e3b0568d2efa_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:847bf22c7517eadd2a771f1a57a0c76dc90cfdd1fe883184c0b9791948f8e7ee +size 827470 diff --git a/2024/Efficient and Effective Weakly-Supervised Action Segmentation via Action-Transition-Aware Boundary Alignment/full.md b/2024/Efficient and Effective Weakly-Supervised Action Segmentation via Action-Transition-Aware Boundary Alignment/full.md new file mode 100644 index 0000000000000000000000000000000000000000..478ffcfec6cc2d6d897db08e93c0bcaee345b8e1 --- /dev/null +++ b/2024/Efficient and Effective Weakly-Supervised Action Segmentation via Action-Transition-Aware Boundary Alignment/full.md @@ -0,0 +1,316 @@ +# Efficient and Effective Weakly-Supervised Action Segmentation via Action-Transition-Aware Boundary Alignment + +Angchi Xu $^{1}$ , Wei-Shi Zheng $^{1,2,3,*}$ + +$^{1}$ School of Computer Science and Engineering, Sun Yat-sen University, China + +$^{2}$ Key Laboratory of Machine Intelligence and Advanced Computing, Ministry of Education, China + +3Guangdong Province Key Laboratory of Information Security Technology, Guangzhou, China + +xuangch@mail2.sysu.edu.cn, wszheng@ieee.org + +# Abstract + +Weakly-supervised action segmentation is a task of learning to partition a long video into several action segments, where training videos are only accompanied by transcripts (ordered list of actions). Most of existing methods need to infer pseudo segmentation for training by serial alignment between all frames and the transcript, which is time-consuming and hard to be parallelized while training. In this work, we aim to escape from this inefficient alignment with massive but redundant frames, and instead to directly localize a few action transitions for pseudo segmentation generation, where a transition refers to the change from an action segment to its next adjacent one in the transcript. As the true transitions are submerged in noisy boundaries due to intra-segment visual variation, we propose a novel Action-Transition-Aware Boundary Alignment (ATBA) framework to efficiently and effectively filter out noisy boundaries and detect transitions. In addition, to boost the semantic learning in the case that noise is inevitably present in the pseudo segmentation, we also introduce video-level losses to utilize the trusted video-level supervision. Extensive experiments show the effectiveness of our approach on both performance and training speed. + +# 1. Introduction + +Action segmentation aims to partition a long untrimmed video into several segments and classify each segment into an action category [8, 12, 13, 15, 16, 29, 32, 35, 38, 44]. It is an important yet challenging task for instructional or procedural video understanding. Although fully-supervised action segmentation (FSAS) methods [1, 6, 13, 14, 27] have achieved great progress, they require frame-wise dense an- + +![](images/8009ded21ae803606f935635d485eedc87f6f083ad130d5b86a91cd05b54ebe0.jpg) +Figure 1. Comparison of performance and training time of WSAS methods on the Breakfast dataset. MoF-The main metric of the task, the higher the better. *-Alignment-free methods. Our ATBA achieves the best performance with a very short training time. + +notation, which is labor-intensive and time-consuming to collect. As a result, many works [5, 9, 25, 28, 34, 38, 46] explore the weakly-supervised action segmentation (WSAS) only requiring the transcript annotation, which refers to the ordered list of actions occurring in the video without their start and end times. The transcripts are less costly to obtain and can be accessed directly from video narrations or other meta data [2, 9, 18, 28, 30, 33, 47]. + +Most of previous WSAS methods have to infer the pseudo segmentation (pseudo frame-wise labels) for training via a sequence alignment process between the video and given transcript, such as Viterbi [22, 23, 25, 28, 33, 34] or Dynamic Time Warping (DTW) [4, 5]. These alignment algorithms are usually designed in a recursive form which needs to be performed serially frame-by-frame and hard to be parallelized, resulting in very slow training process. + +In this work, we argue that the frame-by-frame alignment is NOT necessary, since the pseudo segmentation is + +![](images/2f08921554523255d4b8e94954c42c301787ca2957e1de5eef3b6567101e30a9.jpg) +Figure 2. The necessity of proposed ATBA. The example is P54-webcam01-P54-coffee in Breakfast dataset. GT-The ground-truth segmentation. C.A.Bdy.-Only class-agnostic boundary detection is applied (Exp.1 of Table 3). Acc.-The accuracy of pseudo segmentation. In the video clip around the "star" point, the coffee pot undergoes a change from being picked up to tilted pouring within the segment "Pour Coffee", and this noisy visual change is incorrectly detected. In addition, although two boundaries are correctly detected by the "C.A.Bdy." (diamonds), they correspond to incorrect transitions due to one false positive error (star), resulting in complete dislocation of segments within the dashed box. Best viewed in color. + +fundamentally determined by the locations of a small number of action transitions (i.e., the change from an action segment to its next adjacent action segment in the transcript). Hence, the pseudo segmentation generation can be viewed as a transition detection problem, implying the way to more efficient designs. Intuitively, action transitions are often accompanied by significant visual changes, and there are already many approaches that can detect class-agnostic action boundaries based on these changes [12, 20]. However, due to the intra-segment visual variation and sub-segments under finer granularity, there are numerous noisy boundaries not corresponding to any transitions. Moreover, as the class-agnostic way cannot guarantee correct correspondence between the boundaries and transitions, even slight errors in the detection can result in severe deviation (Fig. 2). + +To overcome the above noisy boundary issue, we propose an efficient and effective framework for WSAS, termed Action-Transition-Aware Boundary Alignment (ATBA), which directly detects the transitions for faster and effective pseudo segmentation generation. To tolerate the noisy boundaries, the ATBA generates more class-agnostic boundaries than the number of transitions as candidates, and then determines a subset from candidates that optimally matches all desired transitions via a drop-allowed alignment algorithm. Furthermore, to fortify the semantic learning under the inevitable noise in pseudo segmentation, we also introduce video-level losses to make use of the trusted video-level supervision. Our ATBA is efficient, because the number of generated candidates will be proportional to the length of the transcript, and therefore the complexity of alignment is now independent of the very long video length. Moreover, other computations + +required by ATBA, i.e., measuring how likely a frame is to be a boundary and how likely a candidate corresponds to a desired transition, are both built on a convolution-like algorithm inspired by [20], which can be parallelized on GPUs efficiently. + +For inference, we directly adopt the results from the trained frame-wise classifier, without the need for any alignment processing with retrieved or predicted transcript (the ground-truth transcript is not available during inference) like previous WSAS methods [4, 5, 25, 28, 34, 38], which also improves the inference efficiency. + +In summary, our contributions are as follows. (1) We propose to directly localize the action transitions for efficient pseudo segmentation generation during training, without the need of time-consuming frame-by-frame alignment. (2) For robustness to noisy boundaries, we propose a novel ATBA framework to effectively determine boundaries corresponding to each transition. Video-level losses are also introduced to regularize the semantic learning involving the unavoidable noise in the pseudo segmentation. Experiments are conducted on three popular datasets to evaluate our approach: Breakfast [21], Hollywood Extended [2] and CrossTask [47]. Our ATBA achieves state-of-the-art or comparable results with one of the fastest training speed (Fig. 1), demonstrating the effectiveness of ours. + +# 2. Related Work + +Weakly-supervised action segmentation methods learn to partition a video into several action segments from training videos only annotated by transcripts [2, 4, 5, 9, 18, 22, 23, 25, 28, 33, 34, 38, 39, 46]. Despite different optimization objectives, most of them generate the pseudo segmentation for training by solving alignment objectives between two sequences (video and transcript) via Connectionist Temporal Classification (CTC) [18], Viterbi [22, 23, 25, 28, 33, 34] or Dynamic Time Warping (DTW) [4, 5]. + +Specifically, [18] proposes an extended version of CTC to evaluate all valid alignments between videos and transcripts, which additionally takes the visual similarities of frames into account. Inspired by speech recognition, [22, 23, 33] all use the Hidden Markov Model (HMM) to model the relationship between videos and actions. [9] always generates uniform segmentation but iteratively adjusts the boundaries by inserting repeating actions into the transcript. [34] proposes an alignment objective based on explicit context and length models, which can be solved by Viterbi, and the solved optimal alignment would serve as pseudo labels to train the frame-wise classifier. [25] and [28] both focus on novel learning objectives, but still require the optimal pseudo segmentation produced by Viterbi. [4, 5] both learn from the contrast of aligning the video to the ground-truth transcript and negative transcripts, where the alignment is performed by DTW. Whether using CTC, Viterbi or DTW, + +![](images/098625b2798000cf20009f8d545e6607bc334dc4819785ae7eacb5cdbf259f1c.jpg) +Figure 3. The overall framework. We propose an Action-Transition-Aware Boundary Alignment (ATBA) framework, which takes the class-agnostic boundary pattern and action transition pattern together into account to efficiently generate pseudo labels. The trusted video-level supervision is also utilized to further enhance the performance. + +the above approaches except [9] require frame-by-frame serial calculation, which are inefficient. + +Recently, some efficient methods are proposed with alignment-free design. [38] learns from the mutual consistency between two forms of a segmentation (i.e., framewise classification and category/length pairs). [29] proposes a loss to enforce the output order of any two actions to be consistent with the transcript. In this work, we also propose an efficient framework with different technical roadmap, and our performance is better at the same level of training speed. + +# 3. Approach + +# 3.1. Problem Statement + +Action segmentation is a task of partitioning a video into several temporal segments with action labels, which is equivalent to predicting the action categories of each frame. Formally, given a sequence of $T$ $d$ -dimensional frame-wise features $\mathbf{X} = [\pmb{x}_1,\dots,\pmb{x}_T] \in \mathbb{R}^{T\times d}$ for a video with $T$ frames, the goal is to predict a sequence of actions $\hat{\mathcal{Y}} = [\hat{y}_1,\dots,\hat{y}_T]$ , where $\hat{y}_t \in \mathcal{C}$ and $\mathcal{C} = \{1,2,\dots,|\mathcal{C}|\}$ is the set of action categories across the dataset (including the background). Under the setting of WSAS, the frame-wise ground-truth $\mathcal{Y} = [y_1,\dots,y_T]$ is NOT available during training. Instead, the ordered list of actions called transcript $\mathcal{A} = [a_1,\dots,a_M]$ is provided (including the background segments), where $a_{m} \in \mathcal{C}$ and $M$ is the total number of action segments in the video. The action transitions of $\mathcal{A}$ + +are naturally formulated as $\mathcal{R} = \{(a_r,a_{r + 1})\}_{r = 1}^{M - 1}$ + +# 3.2. Overview + +Our proposed framework is illustrated in Fig. 3. At first, the input sequence $\mathbf{X}$ is further encoded by a temporal network to generate more task-relevant representations $\mathbf{X}' \in \mathbb{R}^{T \times d'}$ (Sec. 3.3), then a classifier shared along the temporal axis followed by a category softmax activation will predict the frame-wise class probabilities $\mathbf{P} = [\mathbf{p}_1, \dots, \mathbf{p}_T] \in \mathbb{R}^{T \times |C|}$ from $\mathbf{X}'$ . After that, the Action-Transition-Aware Boundary Alignment (ATBA) module takes $\mathbf{P}$ and the transcript $\mathcal{A}$ as input to infer the pseudo frame-wise labels $\widetilde{\mathcal{Y}} = [\widetilde{y}_1, \dots, \widetilde{y}_T]$ , where $\widetilde{y}_t \in C$ (Sec. 3.4). Finally, $\widetilde{\mathcal{Y}}$ is used back to supervise $\mathbf{P}$ by a standard cross entropy: + +$$ +\mathcal {L} _ {\mathrm {c l s}} = - \frac {1}{T} \sum_ {t = 1} ^ {T} \sum_ {c = 1} ^ {| c |} \mathbb {I} (\widetilde {y} _ {t} = c) \log \boldsymbol {P} _ {t, c}, \tag {1} +$$ + +where $\mathbb{I}(\cdot)$ is the indicator function which returns 1 if the condition is satisfied and 0 otherwise. Note that the pseudo frame-wise labels are inferred at each iteration for a batch of data and used immediately for current training at the same iteration. The additional video-level losses are stated in Sec. 3.5, and the training and inference processes are described in Sec. 3.6. + +# 3.3. Temporal Network + +We employ a slightly modified pre-norm [42] Transformer [40] encoder with learnable positional embeddings as the temporal network for feature learning. Following [11, 45], + +![](images/44df138bd8e2149a40074b7c9f1f5ee9a54d566ce63806497cc63237ce56f3b2.jpg) +(a) + +![](images/d0e9d0643e18099af8f651bbd6efe538db1c372cd8e8b0c0965054bb73432cc9.jpg) +Figure 4. (a) A $7 \times 7$ template for class-agnostic boundary scoring. +(b) A $2 \times 7$ template for action transition scoring. + +the vanilla full self-attention is replaced with a pyramid hierarchical local attention to better adapt to the action segmentation task (see supplemental material for more details). + +# 3.4. Action-Transition-Aware Boundary Alignment + +The ATBA is the core component of our framework. It generates the pseudo frame-wise labels $\widetilde{\mathcal{V}}$ by inferring the boundaries corresponding to $M - 1$ action transitions from the predicted class probabilities $\pmb{P}$ . Once they are found, $\widetilde{\mathcal{V}}$ can be naturally generated by assigning the action labels of $\mathcal{A}$ one-by-one into the intervals between these boundaries. + +Briefly, the ATBA first generates a set of candidate boundaries via a class-agnostic way, and then finds $M - 1$ points from this set that optimally match all transitions of $\mathcal{A}$ via a dynamic programming (DP) algorithm. We describe the details in the following. + +- Class-Agnostic Boundary Scoring. Firstly, the ATBA calculates the class-agnostic boundary scores $\mathcal{V}^{\mathrm{b}} = [v_1^{\mathrm{b}},\dots,v_T^{\mathrm{b}}]$ for each timestamp. We employ a pattern-matching-based scoring method proposed by a generic event boundary detection approach termed UBoCo [20], which considers the pattern of each frame's neighborhood. Specifically, for timestamp $t$ , a pairwise similarity matrix $\Gamma^{(t)}\in \mathbb{R}^{w^{\mathrm{b}}\times w^{\mathrm{b}}}$ is calculated within a local window with size $w^{\mathrm{b}}$ centered at $t$ , whose $(i,j)$ -entry represents the class-agnostic similarity between $i$ -th and $j$ -th frames inside the window, with values ranging from -1 to 1 (see the computational details in supplementary material). Clearly, if $t$ is a boundary, the frame feature should change dramatically at $t$ and keep stable elsewhere, so its $\Gamma^{(t)}$ should show the pattern of that the values in the upper-left and lower-right areas are close to 1 and otherwise close to -1. Hence a template $\Omega^{\mathrm{b}}\in \mathbb{R}^{w^{\mathrm{b}}\times w^{\mathrm{b}}}$ like Fig. 4(a) is designed to capture this pattern and output the class-agnostic boundary score for each $t$ by a correlation operation: + +$$ +v _ {t} ^ {\mathrm {b}} = \frac {1}{w ^ {\mathrm {b}} \times w ^ {\mathrm {b}}} \sum_ {i = 1} ^ {w ^ {\mathrm {b}}} \sum_ {j = 1} ^ {w ^ {\mathrm {b}}} \Omega_ {i, j} ^ {\mathrm {b}} \Gamma_ {i, j} ^ {(t)}. \tag {2} +$$ + +- Candidate Boundary Selection. After calculating $\nu^{\mathrm{b}}$ , we select a set of $K$ candidate boundaries $\tilde{\mathcal{B}} = \{b_k\}_{k=1}^K$ , where + +![](images/7fa9cab4368f949836b38db60133b5627548cd5a104037341dbdf1c7b6ee3418.jpg) +(b) +Figure 5. Illustration of the alignment between action transitions $\mathcal{R}$ and candidate boundaries $\tilde{\mathcal{B}}$ . Blue circles are aligned and gray ones are dropped. (a) A valid alignment. (b) An invalid alignment. The red dashed arrow violates the ordering consistency. Best viewed in color. + +$1 < b_{1} < \ldots < b_{K} \leq T$ and $K > M - 1$ . The selection is performed by a simple greedy strategy with non maximum suppression inspired by [12], i.e., each time we select one timestamp with current highest score $v_{t}^{\mathrm{b}}$ , and invalidate its neighborhood to avoid selecting multiple timestamps corresponding to one same boundary, until the number of selected timestamps reach an upper bound or all remaining timestamps are invalid. The radius of the invalid interval is set adaptively to $\frac{\mu T}{M}$ , where $\mu \in [0,1]$ is a hyper-parameter, and in practice, the upper bound for $K$ is set to $\lambda(M - 1)$ , where $\lambda \in \mathbb{N}_{+}$ is also a hyper-parameter. + +- Action Transition Scoring. As mentioned in Sec. 1, the class-agnostic scores $\mathcal{V}^{\mathrm{b}}$ are not enough to detect action transitions. To this end, we then calculate an action transition score matrix $V^{\mathrm{a}}\in \mathbb{R}^{K\times (M - 1)}$ , where $V_{k,r}^{\mathrm{a}}$ measures the possibility that the $k$ -th candidate boundary corresponds to the $r$ -th transition, i.e., separates the $r$ -th and $(r + 1)$ -th action segments. This matrix is also calculated via pattern matching. Clearly, if candidate $b_{k}$ corresponds to the $r$ -th transition, the classifier's activation for class $a_{r}$ should drop sharply after $b_{k}$ , while rise for class $a_{r + 1}$ . Hence, a template $\Omega^{\mathrm{a}}\in \mathbb{R}^{2\times w^{\mathrm{a}}}$ with temporal size $w^{\mathrm{a}}$ like Fig. 4(b) is employed to detect this pattern around $b_{k}$ : + +$$ +\boldsymbol {V} _ {k, r} ^ {\mathrm {a}} = \frac {1}{2 w ^ {\mathrm {a}}} \sum_ {i = 1} ^ {2} \sum_ {j = 1} ^ {w ^ {\mathrm {a}}} \boldsymbol {\Omega} _ {i, j} ^ {\mathrm {a}} \boldsymbol {P} _ {\text {i n d} ^ {\mathrm {a}} \left(b _ {k}, j\right), a _ {r + i - 1}}, \tag {3} +$$ + +$$ +\operatorname {i n d} ^ {\mathrm {a}} \left(b _ {k}, j\right) = b _ {k} - \left\lfloor \frac {w ^ {\mathrm {a}}}{2} \right\rfloor + j - 1, +$$ + +where $\mathrm{ind}^{\mathrm{a}}(b_k,j)$ is the index transform from the index $j$ of the local window centered at $b_{k}$ to the global timestamp index. Finally, the class-agnostic scores $\mathcal{V}^{\mathrm{b}}$ are added back to $\pmb{V}^{\mathrm{a}}$ to produce the final score matrix $\pmb {V}\in \mathbb{R}^{K\times (M - 1)}$ : + +$$ +\boldsymbol {V} _ {k, r} = \boldsymbol {V} _ {k, r} ^ {\mathrm {a}} + v _ {b _ {k}} ^ {\mathrm {b}}. \tag {4} +$$ + +The above equation means that the class-agnostic boundary score of the $k$ -th boundary $v_{b_k}^{\mathrm{b}}$ is added to all transition scores corresponding to the $k$ -th boundary $\mathbf{V}_{k,r}^{\mathrm{a}}, \forall r$ . + +- Action Transition Alignment. The final step is to find exact $M - 1$ optimal boundaries from the candidates $\widetilde{B}$ based + +![](images/fb7848cf7d882894564cc54bbca57f5d3b3fa5e915f0363ff36afdc90de45a4f.jpg) +Figure 6. An example of action transition alignment between $\widetilde{\mathcal{B}}$ of length 5 and $\mathcal{R}$ of length 3. (a) The cost matrix $\Delta$ . (b) The cumulative cost matrix $D$ . Blue areas are directly initialized. Yellow areas correspond to optimal alignment. Red arrows are allowed path directions. (c) The optimal alignment. Best viewed in color. + +on the score matrix $V$ . It is equivalent to seek an one-to-one alignment with lowest cost between the transitions $\mathcal{R}$ and candidates $\widetilde{\mathcal{B}}$ while requiring $K - M + 1$ candidates to be dropped (Fig. 5). Formally, if we denote the optimal aligned boundary set as $\mathcal{B} = \{b_{k_r}\}_{r=1}^{M-1}$ , where $b_{k_r}$ corresponds to the $r$ -th transition and $1 \leq k_1 < \ldots < k_{M-1} \leq K$ . It should satisfy: + +$$ +\mathcal {B} = \arg \min _ {\mathcal {B} ^ {\prime}} \psi \left(\mathcal {B} ^ {\prime}\right), \quad \psi \left(\mathcal {B} ^ {\prime}\right) = - \sum_ {r = 1} ^ {M - 1} \boldsymbol {V} _ {k _ {r} ^ {\prime}, r}, \tag {5} +$$ + +where $\psi (\mathcal{B}^{\prime})$ is the cost function of an alignment and $\mathcal{B}'$ is any feasible aligned boundary set such as $\mathcal{B}$ . + +To solve Eq. (5), we propose a DP algorithm, of which an example is shown in Fig. 6. To allow some candidates in $\widetilde{B}$ to be dropped, inspired by [36], we first expand the transition sequence $\mathcal{R}$ by inserting the empty symbol $\phi$ interleaved with transition symbols, obtaining $\mathcal{R}' = [\phi, (a_1, a_2), \phi, (a_2, a_3), \phi, \dots, \phi, (a_{M-1}, a_M), \phi]$ . Now the candidate boundary matched to $\phi$ will be discarded. Then, a cost matrix $\Delta \in \mathbb{R}^{K \times (2(M-1) + 1)}$ is constructed (Fig. 6(a)), whose $(k, r')$ -th entry is the cost of aligning the $b_k$ with the $r'$ -th symbol in $\mathcal{R}'$ . For even number of $r'$ (transition symbols), it is clear that $\Delta_{k,r'} = -V_{k,r'/2}$ from Eq. (5). Odd $r'$ means dropping the candidate, and the cost is set to 0. + +A valid alignment is represented by a path from the top-left to bottom-right of $\Delta$ with constraint on directions. Specifically, the allowed start positions are $(1,1)$ and $(1,2)$ , respectively corresponding to two cases where $b_{1}$ is dropped and $b_{1}$ is matched with the first transition. Similarly, there are two allowed end positions: $(K,2(M - 1))$ and $(K,2(M - 1) + 1)$ . Then for the position $(k,r^{\prime})$ on the path, if $r^{\prime}$ is odd, its previous position is either $(k - 1,r^{\prime})$ or $(k - 1,r^{\prime} - 1)$ . It means that if $b_{k}$ is dropped, the $b_{k - 1}$ is also dropped or matched with the previous transition. On the other hand, the previous position is either $(k - 1,r^{\prime} - 1)$ or $(k - 1,r^{\prime} - 2)$ if $r^{\prime}$ is even, with the similar meaning. These allowed directions are indicated by red arrows in Fig. 6(b). + +The DP algorithm generally finds the optimal solution by computing a cumulative cost matrix $\pmb{D}$ with the same shape as $\pmb{\Delta}$ (Fig. 6(b)), where $D_{k,r'}$ represents the minimum cumulative cost of all valid paths ending at $(k, r')$ . It + +is computed by the following recursive equation, considering paths coming from valid directions aforementioned: + +$$ +\boldsymbol {D} _ {k, r ^ {\prime}} = \boldsymbol {\Delta} _ {k, r ^ {\prime}} + \left\{ \begin{array}{l} \min \left(\boldsymbol {D} _ {k - 1, r ^ {\prime}}, \boldsymbol {D} _ {k - 1, r ^ {\prime} - 1}\right), r ^ {\prime} \text {i s o d d}, \\ \min \left(\boldsymbol {D} _ {k - 1, r ^ {\prime} - 1}, \boldsymbol {D} _ {k - 1, r ^ {\prime} - 2}\right), r ^ {\prime} \text {i s e v e n}. \end{array} \right. \tag {6} +$$ + +The above equation is not applicable to the first row and the first two columns of $D$ . They are directly initialized before the computation (blue areas of Fig. 6(b)). We set the cost of invalid positions to $\infty$ (inf) and others are copied from $\Delta$ . + +The cumulative cost of the optimal alignment is $\min (D_{K,2(M - 1)},D_{K,2(M - 1) + 1})$ , The complete alignment, i.e., the optimal boundary set $\mathcal{B}$ can be obtained by starting from the corresponding position of the above min operation and backtracking (Fig. 6(b/c)). More details can be found in the supplementary material. + +- Complexity. Clearly, the class-agnostic boundary scoring and action transition scoring can be implemented by the unfold operator and matrix operations, thus can be parallelized on GPUs. The action transition alignment must be performed serially with a time complexity of $O(KM)$ or $O(\lambda M^2)$ . Compared with the previous alignment methods such as Viterbi ( $O(T^2M)$ [25, 28, 34]) and DTW ( $O(TM)$ ), it decreases from a function involving video length $T$ to one of only transcript length $M$ ( $M << T$ ). + +# 3.5. Video-Level Losses + +Inevitably, the inferred pseudo segmentation contains some degree of noise, which can be unexpectedly fit by the network. To improve the semantic robustness, we propose to jointly train a video-level multi-label classification task that predicts whether each action appears in the video, which is precisely indicated by the transcript. Specifically, inspired by the common practice of Transformer [7, 10, 19, 24, 43], the input sequence $\mathbf{X}$ is augmented by a set of $|\mathcal{C}|$ learnable class tokens which are fed into the same network and responsible for predicting action occurrence. The outputs, denoted as $\{e_c'\}_{c=1}^{|\mathcal{C}|}$ , are then used to predict the action occurrence probability via $\xi_c = \sigma(\boldsymbol{w}_c^T \boldsymbol{e}_c' + \epsilon_c)$ for each category $c$ , where $\boldsymbol{w}_c$ and $\epsilon_c$ are the weight vector and bias scalar for class $c$ in the classifier shared with frame-wise classification. $\sigma(\cdot)$ is the sigmoid activation. These predictions are supervised by the binary cross entropy: + +$$ +\mathcal {L} _ {\mathrm {v i d}} = - \frac {1}{| \mathcal {C} |} \sum_ {c = 1} ^ {| \mathcal {C} |} \left[ y _ {c} ^ {\mathrm {v i d}} \log \xi_ {c} + \left(1 - y _ {c} ^ {\mathrm {v i d}}\right) \log \left(1 - \xi_ {c}\right) \right], \tag {7} +$$ + +where $y_c^{\mathrm{vid}} \in \{0,1\}$ is the binary label indicating whether action $c$ appears in the video. Through the shared network and the interactions between tokens, the semantics learned from video-level training can benefit the main task. + +Moreover, we treat the class tokens with global information as the prototypes of each action and attempt to align the + +frame features to them for semantic consistency. Specifically, we first calculate the feature centroid of each category $c$ that appears in the video by averaging the output frame features with the corresponding pseudo labels: + +$$ +\overline {{\boldsymbol {x}}} _ {c} = \frac {\sum_ {t = 1} ^ {T} \mathbb {I} \left(\widetilde {y} _ {t} = c\right) \boldsymbol {x} _ {t} ^ {\prime}}{\sum_ {t = 1} ^ {T} \mathbb {I} \left(\widetilde {y} _ {t} = c\right)}, c \in \operatorname {S e t} (\mathcal {A}), \tag {8} +$$ + +where the $\operatorname{Set}(\cdot)$ operator converts a list into an unordered set and removes duplicates. These centroids will move towards the corresponding $e_c'$ under the guidance of a global-local contrastive loss in the form of InfoNCE [17, 31]: + +$$ +\mathcal {L} _ {\mathrm {g l c}} = - \frac {1}{| \operatorname {S e t} (\mathcal {A}) |} \sum_ {c \in \operatorname {S e t} (\mathcal {A})} \log \frac {\exp \left(\left\langle \overline {{\boldsymbol {x}}} _ {c} \right\rangle^ {T} \left\langle \boldsymbol {e} _ {c} ^ {\prime} \right\rangle / \tau\right)}{\sum_ {c ^ {\prime} = 1} ^ {| \mathcal {C} |} \exp \left(\left\langle \overline {{\boldsymbol {x}}} _ {c} \right\rangle^ {T} \left\langle \boldsymbol {e} _ {c ^ {\prime}} ^ {\prime} \right\rangle / \tau\right)}, \tag {9} +$$ + +where $\langle \pmb{x}\rangle = \pmb{x} / \| \pmb{x}\| _2$ is the $l_{2}$ -normalization operator, and $\tau$ is the temperature hyper-parameter. + +# 3.6. Training and Inference + +- Training. Since the pseudo labeling always requires a relatively good initialization, we utilize a two-stage training strategy. The first stage only uses the reliable labels, i.e., the video-level labels, to pretrain the network. Hence the loss function is $\mathcal{L}_{\mathrm{I}} = \mathcal{L}_{\mathrm{vid}}$ . The loss for the second stage further includes the frame-wise losses on both classification and representation using pseudo labels: $\mathcal{L}_{\mathrm{II}} = \alpha \mathcal{L}_{\mathrm{vid}} + \beta \mathcal{L}_{\mathrm{cls}} + \gamma \mathcal{L}_{\mathrm{glc}}$ , where $\alpha$ , $\beta$ and $\gamma$ are all hyperparameters. + +- Inference. For inference, we obtain the action label $\hat{y}_t$ for frame $t$ directly from the frame-wise class probabilities: $\hat{y}_t = \arg \max_c P_{t,c}$ . Note that we do not require any alignment processing during inference, which is performed by previous WSAS methods for smoother predictions. However, since the ground-truth transcript is not available during inference, these methods either run alignment with every transcript from the training set [4, 5, 25, 28, 34], or predict the transcript via another model [38], which makes the pipeline more time-consuming and less practical. + +# 4. Experiments + +# 4.1. Experimental Setup + +- Datasets. We perform experiments on three datasets. The Breakfast [21] dataset contains 1712 videos of breakfast cooking with 48 different actions. On average, each video has 6.8 segments and $7.3\%$ frames are background. The Hollywood Extended [2] dataset contains 937 videos taken from movies with 16 categories of daily actions such as walk or sit. On average, each video has 5.9 segments and $60.9\%$ frames are background. The CrossTask [47] dataset contains videos from 18 primary tasks. Following [28], 14 cooking-related tasks are selected, which have 2552 videos + +and 80 action categories. On average, each video has 14.4 segments and $74.8\%$ frames are background. For Breakfast, we use the released 4 training/test splits and report the average. For Hollywood, we perform a 10-fold cross-validation. For CrossTask, we use the released training/test split. These evaluation protocols are consistent with previous methods. + +- Metrics. To evaluate our method, we use 4 standard metrics, following [5, 9, 25, 28]. (1) The Mean-over-Frames (MoF) is the percentage of frames whose labels are correctly predicted. (2) The Mean-over-Frames without Background (MoF-Bg) is the MoF over non-background frames. It is more suitable than MoF for the datasets with high background rate such as Hollywood and CrossTask. (3) The Intersection-over-Union (IoU) is defined as $|I \cap I^{*}| / |I \cup I^{*}|$ while (4) the Intersection-over-Detection (IoD) is $|I \cap I^{*}| / |I|$ , where $I^{*}$ and $I$ are the ground-truth (GT) segment and the predicted segment with the same class, respectively. For each GT segment, the highest IoU/IoD with one predicted segment are preserved, and the average of all GT segments is reported. Note that the definition of the IoU/IoD in [28] is different from other works [5, 9, 25], and will be indicated with special symbols when used. + +- Input Features. As the input features $X$ for Breakfast and Hollywood, we use the 2048-dimensional RGB+flow I3D features [3] adopted by MuCon [38] and most FSAS methods [13, 26, 27, 45], while some WSAS methods [5, 25, 28, 34] still use the iDT features [41]. Since recent studies [37, 38] have found that they perform worse with more advanced I3D features, we still report the performance with the iDT features for them following [38]. For CrossTask, the officially released 3200-dimensional features are adopted, but we do not perform dimension reduction via PCA as with [28]. For GPU memory efficiency, we perform a $10 \times$ temporal downsampling for Breakfast and $5 \times$ for Hollywood. During inference, the output is upsampled to match the original video/ground-truth length. + +- Implementation Details. We adopt a 6-layer Transformer [40] with single-head self-attention as the temporal network, and the latent dimension $d'$ is 256. For the ATBA, we set $w^{\mathrm{b}} = 7$ and $\lambda = 4$ for all datasets. $w^{\mathrm{a}}$ is set to 31 for Breakfast/CrossTask and 23 for Hollywood. $\mu$ is set to 0.3 for Breakfast/Hollywood and 0.1 for CrossTask. The loss weights $\alpha$ , $\beta$ are set to 1.0 and $\gamma$ is 0.1. $\tau$ in Eq. (9) is set to 0.2. Besides, to alleviate the issue that the predictions on Hollywood/CrossTask are overly dominated by background, we lower the sample weights of pseudo background frames to 0.8 in Eq. (1) for these datasets. + +- Multiple Runs. Due to the alternating nature of learning from weak supervision, there are often fluctuations in the results of WSAS methods [28, 38]. Hence, following [38], we report the average and standard deviation over 5 runs with different random seeds for better evaluation. + +
Breakfast
MethodMoF±stdMoF-Bg±stdIoU±stdIoD±std
HMM+RNN [33]33.3---
[33]+Length [23]36.7---
TCFPN+ISBA [9]38.4/36.4±1.0*38.424.240.6
NN-Viterbi [34]43.0/39.7±2.4*---
D3TW [4]45.7---
CDFL [25]50.2/48.1±2.5*48.033.745.4
DPDTW [5]50.8-35.645.1
TASL [28]♠47.8-35.2†46.1†
MuCon [38]♠48.5±1.850.3*40.9*54.0*
POC [29]♠45.7-38.3†-
AdaAct [46]51.248.336.346.4
ATBA♠(Ours)53.9±1.254.4±1.241.1±0.761.7±1.1
39.5±0.8†55.9±1.0†
+ +Table 2. Comparison of accuracy, training and inference time on the Breakfast. The training time is measured as the entire training duration on the split 1, and the inference time is measured as the average time for inferring a video from the test set of the split 1. Tr.A.-The alignment algorithm adopted during training (X-No Alignment. V-Viterbi. D-DTW. B-Boundary Alignment). *-Measured by [38]. ‡-Measured by us. Best results are in bold, second best are underlined. + +
Hollywood Extended
MethodMoF±stdMoF-Bg±stdIoU±stdIoD±std
HMM+RNN [33]--11.9-
[33]+Length [23]--12.3-
TCFPN+ISBA [9]28.734.512.618.3
D3TW [4]33.6---
CDFL [25]45.040.619.525.8
DPDTW [5]55.625.6#33.243.3
TASL [28]42.1#27.2#23.3†#33.0†#
MuCon [38]♠-41.613.9*-
ATBA♠(Ours)47.7±2.540.2±1.630.9±1.655.8±0.8
28.5±1.6†44.9±0.5†
+ +
CrossTask
MethodMoF±stdMoF-Bg±stdIoU±stdIoD±std
NN-Viterbi [34]♠26.5*-10.7†*24.0†*
CDFL [25]♠31.9*-11.5†*23.8†*
TASL [28]♠40.727.4#14.5†25.1†
POC [29]♠42.817.6#15.6†-
ATBA♠(Ours)50.6±1.331.3±0.720.9±0.444.6±0.7
15.7±0.3†24.6±0.4†
+ +Table 1. Comparisons of ours with other WSAS methods on three datasets. std is the standard deviation of multiple runs (if any). $\dagger$ The metric is computed by the definition of [28]. *-Results are reported by other works. Please refer to the supplementary material for detailed sources. $\sharp$ -Results are obtained by us via rerunning the open sources (The TASL [28] does not follow the common 10-fold evaluation protocol for Hollywood so we re-produce the results). $\spadesuit$ The reported results are the average of multiple runs. Best results are in bold, second best are underlined. + +# 4.2. Comparison with the State-of-the-Art + +- Performance. In Table 1, we compare our proposed method with previous WSAS methods. Our ATBA achieves state-of-the-art (SOTA) by a clear margin $(+2.7\% \mathrm{MoF})$ on the Breakfast [21], demonstrating the effectiveness of focusing on action transitions. Moreover, comparing the standard deviation with other methods, it can be found that our method is also relatively stable. + +Our ATBA also achieves comparable performance on the Hollywood [2]. The reason why our method does not show significant advantage is probably because this dataset is collected from movies and so contains many shot changes, re + +
MethodTr.A.MoFTraining (Hours)Inference (Seconds)
TCFPN+ISBA [9]✘33.312.75*0.01*
NN-Viterbi [34]V43.011.23*56.25*
CDFL [25]V50.266.73*62.37*
DPDTW [5]D50.831.02‡0.69‡
TASL [28]V47.824.66‡54.99‡
MuCon [38]✘48.54.573.03
POC [29]✘45.72.28‡0.01‡
ATBA (Ours)B53.93.450.01
+ +sulting in more noisy boundaries. Note that although the DPDTW [5] achieves significantly high MoF, this metric is severely biased in case of highly imbalanced categories (60.9% frames are background). We report the MoF-Bg metric for it, on which it performs poorly, proving that it cannot recognize real actions very well. In contrast, the performance of our method is more balanced. + +For the more difficult dataset CrossTask [47] with the most segments and highest background rate, our approach also outperforms previous methods whether or not the metric involves background ( $+7.8\%$ MoF and $+3.9\%$ MoF-Bg). + +- Efficiency. Besides of the performance, we also compare the training time of our approach with previous WSAS methods to show the efficiency of ours. Following [38], we train our model on an Nvidia GeForce GTX 1080Ti GPU, and the training time is measured as the wall time over the whole training phase, during which any irrelevant operations such as intermediate evaluation and saving checkpoints are removed. As all the WSAS methods directly load pre-computed features, the time measurement also does not include the time to extract features from raw videos. As shown in Table 2, our ATBA achieves the best performance with the second shortest training time, which demonstrates the effectiveness of our design. Specifically, the training speed of ours is on average 10 times (varying from 3 to 20) faster than methods performing frame-by-frame alignment [5, 25, 28, 34]. Compared to the alignment-free methods with comparable training speed [29, 38], ours achieves better performance. + +We also compare the time to infer a single test video. As mentioned in Sec. 3.6, ours does not require any alignment processing, which makes it the fastest during inference. + +
Exp.\( \gamma^b \)Cost Mat. \( V^a \)P.L.MoFIoUIoD
1✓38.632.323.152.5
2✓✓51.741.330.247.8
3✓○✓67.954.041.1
+ +# 4.3. Ablation Studies + +- Effect of ATBA. In Table 3, we conducted an in-depth evaluation of our proposed ATBA. From Exp.1, the pseudo label quality and the evaluation performance are both very poor when only using the class-agnostic boundary scores $\nu^{\mathrm{b}}$ to directly select $M - 1$ boundaries for transitions via the greedy strategy stated in Sec. 3.4, showing that it is critical to take the class-specific transition pattern into account as with our design to suppress the noisy boundaries for more precise pseudo segmentation. Fig. 2 provides an intuitive example. In addition, comparing Exp.2 and 3, it's better to involve the class-agnostic boundary scores in the action transition alignment. We think it is because that some candidates can have very low boundary scores (unlikely to be a boundary) as the candidate selection only depends on ranking, and when the transition patterns are not discriminative at the beginning of training, these candidates may be unexpectedly aligned if the boundary scores are not involved. + +- Effect of Video-Level Losses. In Table 4, we evaluate the $\mathcal{L}_{\mathrm{vid}}$ and $\mathcal{L}_{\mathrm{glc}}$ . Comparing Exp.1 and 2, the model with $\mathcal{L}_{\mathrm{vid}}$ achieves higher performance ( $+6.5\%$ MoF) at the same level of pseudo label accuracy, demonstrating that the video-level supervision can promote the learning of precise action semantics. In Exp.3, the explicit representation alignment $(\mathcal{L}_{\mathrm{glc}})$ further improves the quality of frame features and thus the performance ( $+1.6\%$ MoF). Moreover, Exp.4 shows that whether or not to share the classifier between frame-/video-level classification has little impact. + +We also provide the analysis on the effect of important hyper-parameters in the supplementary material, including $w^{\mathrm{b}}$ , $w^{\mathrm{a}}$ , $\mu$ and $\lambda$ in the ATBA. + +# 4.4. Qualitative Results + +We show the qualitative inference result compared with two recent open source WSAS methods, i.e., MuCon [38] and TASL [28], in Fig. 7. Our ATBA achieves significantly more accurate results in this challenging video with many segments. The result of TASL shows order reversal ("Crake + +Table 3. Ablation studies of ATBA on the Breakfast. Exp.-Different experiment configurations. Cost Mat.-Different choices for the cost matrix in action transition alignment. o-According to Eq. (4), $V$ contains $V^{\mathrm{a}}$ . P.L.-The accuracy of pseudo labels during training. Exp.1 means that only the class-agnostic boundary detection is adopted to localize action transitions. Exp.3 is our default setting. Best results are in bold. + +
Exp.\( \mathcal{L}_{\text{vid}} \)\( \mathcal{L}_{\text{glc}} \)D.Cls.P.L.MoFIoUIoD
167.545.940.561.9
2✓66.652.438.861.4
3✓✓67.954.041.162.3
4✓✓✓68.153.740.161.7
+ +Table 4. Ablation studies of video-level losses on the Breakfast. Exp.-Different experiment configurations. D.Cls.-Different classifiers for frame-wise prediction and action occurrence prediction. P.L.-The accuracy of pseudo labels during training. Exp.3 is our default setting. Best results are in bold. + +![](images/e1977212b6e5ed40c5dde94ef8a5582bd3e4c1ac887b4ba97c5238537188ea79.jpg) +Figure 7. Qualitative results on the Breakfast. The example test video is P15-stereo01-P15-pancake. We compare our inference result with two recent methods. Best viewed in color. + +Egg" and "Spoon Flour") and hallucination ("Butter Pan"). The MuCon predicts the correct action ordering, but the result deviates severely. In contrast, ours successfully predicts an accurate segmentation, indicating the action semantics are well learned. More qualitative results are provided in supplementary material. + +# 5. Conclusion + +In this work, we propose to directly localize action transitions for efficient pseudo segmentation generation in the WSAS task, thus avoiding the time-consuming frame-by-frame alignment. Due to the presence of noisy boundaries, a novel Action-Transition-Aware Boundary Alignment (ATBA) framework is proposed to efficiently and effectively filter out noise and detect transitions. Moreover, we also design some video-level losses to utilize video-level supervision to improve the semantic robustness. Extensive experiments show the effectiveness of our ATBA. + +# 6. Acknowledgment + +This work was supported partially by the NSFC (U21A20471), National Key Research and Development Program of China (2023YFA1008503), Guangdong NSF Project (No. 2023B1515040025, 2020B1515120085). + +# References + +[1] Emad Bahrami, Gianpiero Francesca, and Juergen Gall. How much temporal long-term context is needed for action segmentation? arXiv preprint arXiv:2308.11358, 2023. 1 +[2] Piotr Bojanowski, Rémi Lajugie, Francis Bach, Ivan Laptev, Jean Ponce, Cordelia Schmid, and Josef Sivic. Weakly supervised action labeling in videos under ordering constraints. In Proceedings of the European Conference on Computer Vision, pages 628-643. Springer, 2014. 1, 2, 6, 7 +[3] Joao Carreira and Andrew Zisserman. Quo vadis, action recognition? a new model and the kinetics dataset. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 6299-6308, 2017. 6 +[4] Chien-Yi Chang, De-An Huang, Yanan Sui, Li Fei-Fei, and Juan Carlos Niebles. D3tw: Discriminative differentiable dynamic time warping for weakly supervised action alignment and segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3546–3555, 2019. 1, 2, 6, 7 +[5] Xiaobin Chang, Frederick Tung, and Greg Mori. Learning discriminative prototypes with dynamic time warping. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8395-8404, 2021. 1, 2, 6, 7 +[6] Min-Hung Chen, Baopu Li, Yingze Bao, Ghassan Al-Regib, and Zsolt Kira. Action segmentation with joint self-supervised temporal domain adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9454–9463, 2020. 1 +[7] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018. 5 +[8] Guodong Ding and Angela Yao. Leveraging action affinity and continuity for semi-supervised temporal action segmentation. In European Conference on Computer Vision, pages 17-32. Springer, 2022. 1 +[9] Li Ding and Chenliang Xu. Weakly-supervised action segmentation with iterative soft boundary assignment. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6508-6516, 2018. 1, 2, 3, 6, 7 +[10] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020.5 +[11] Dazhao Du, Bing Su, Yu Li, Zhongang Qi, Lingyu Si, and Ying Shan. Efficient u-transformer with boundary-aware loss for action segmentation. arXiv preprint arXiv:2205.13425, 2022. 3 +[12] Zexing Du, Xue Wang, Guoqing Zhou, and Qing Wang. Fast and unsupervised action boundary detection for action segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3323-3332, 2022. 1, 2, 4 + +[13] Yazan Abu Farha and Jurgen Gall. Ms-tcn: Multi-stage temporal convolutional network for action segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3575-3584, 2019. 1, 6 +[14] Shang-Hua Gao, Qi Han, Zhong-Yu Li, Pai Peng, Liang Wang, and Ming-Ming Cheng. Global2local: Efficient structure search for video action segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16805-16814, 2021. 1 +[15] Reza Ghoddoosian, Isht Dwivedi, Nakul Agarwal, Chiho Choi, and Behzad Dariush. Weakly-supervised online action segmentation in multi-view instructional videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13780-13790, 2022. 1 +[16] Reza Ghoddoosian, Isht Dwivedi, Nakul Agarwal, and Behzad Dariush. Weakly-supervised action segmentation and unseen error detection in anomalous instructional videos. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10128-10138, 2023. 1 +[17] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9729-9738, 2020. 6 +[18] De-An Huang, Li Fei-Fei, and Juan Carlos Niebles. Connectionist temporal modeling for weakly supervised action labeling. In Proceedings of the European Conference on Computer Vision, pages 137–153. Springer, 2016. 1, 2 +[19] Zi-Hang Jiang, Qibin Hou, Li Yuan, Daquan Zhou, Yujun Shi, Xiaojie Jin, Anran Wang, and Jiashi Feng. All tokens matter: Token labeling for training better vision transformers. Advances in neural information processing systems, 34: 18590-18602, 2021. 5 +[20] Hyolim Kang, Jinwoo Kim, Taehyun Kim, and Seon Joo Kim. Uboco: Unsupervised boundary contrastive learning for generic event boundary detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20073-20082, 2022. 2, 4 +[21] Hilde Kuehne, Ali Arslan, and Thomas Serre. The language of actions: Recovering the syntax and semantics of goal-directed human activities. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 780-787, 2014. 2, 6, 7 +[22] Hilde Kuehne, Alexander Richard, and Juergen Gall. Weakly supervised learning of actions from transcripts. Computer Vision and Image Understanding, 163:78-89, 2017. 1, 2 +[23] Hilde Kuehne, Alexander Richard, and Juergen Gall. A hybrid rnn-hmm approach for weakly supervised temporal action segmentation. IEEE transactions on pattern analysis and machine intelligence, 42(4):765-779, 2018. 1, 2, 7 +[24] Jack Lanchantin, Tianlu Wang, Vicente Ordonez, and Yanjun Qi. General multi-label image classification with transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16478-16488, 2021. 5 +[25] Jun Li, Peng Lei, and Sinisa Todorovic. Weakly supervised energy-based learning for action segmentation. In Proceed- + +ings of the IEEE/CVF International Conference on Computer Vision, pages 6243-6251, 2019. 1, 2, 5, 6, 7 +[26] Shi-Jie Li, Yazan AbuFarha, Yun Liu, Ming-Ming Cheng, and Juergen Gall. Ms-tcn++: Multi-stage temporal convolutional network for action segmentation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2020. 6 +[27] Daochang Liu, Qiyue Li, AnhDung Dinh, Tingting Jiang, Mubarak Shah, and Chang Xu. Diffusion action segmentation. arXiv preprint arXiv:2303.17959, 2023. 1, 6 +[28] Zijia Lu and Ehsan Elhamifar. Weakly-supervised action segmentation and alignment via transcript-aware union-of-subspaces learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8085-8095, 2021. 1, 2, 5, 6, 7, 8 +[29] Zijia Lu and Ehsan Elhamifar. Set-supervised action learning in procedural task videos via pairwise order consistency. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19903-19913, 2022. 1, 3, 7 +[30] Antoine Miech, Dimitri Zhukov, Jean-Baptiste Alayrac, Makarand Tapaswi, Ivan Laptev, and Josef Sivic. Howto100m: Learning a text-video embedding by watching hundred million narrated video clips. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2630-2640, 2019. 1 +[31] Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748, 2018. 6 +[32] Rahul Rahaman, Dipika Singhania, Alexandre Thiery, and Angela Yao. A generalized and robust framework for timestamp supervision in temporal action segmentation. In Proceedings of the European Conference on Computer Vision, pages 279–296. Springer, 2022. 1 +[33] Alexander Richard, Hilde Kuehne, and Juergen Gall. Weakly supervised action learning with rn based fine-to-coarse modeling. In Proceedings of the IEEE conference on Computer Vision and Pattern Recognition, pages 754-763, 2017. 1, 2, 7 +[34] Alexander Richard, Hilde Kuehne, Ahsan Iqbal, and Juergen Gall. Neuralnetwork-viterbi: A framework for weakly supervised video learning. In Proceedings of the IEEE conference on Computer Vision and Pattern Recognition, pages 7386-7395, 2018. 1, 2, 5, 6, 7 +[35] Yuhan Shen and Ehsan Elhamifar. Semi-weakly-supervised learning of complex actions from instructional task videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3344–3354, 2022. 1 +[36] Yuhan Shen, Lu Wang, and Ehsan Elhamifar. Learning to segment actions from visual and language instructions via differentiable weak sequence alignment. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10156-10165, 2021. 5 +[37] Yaser Souri, Alexander Richard, Luca Minciullo, and Juergen Gall. On evaluating weakly supervised action segmentation methods. arXiv preprint arXiv:2005.09743, 2020. 6 +[38] Yaser Souri, Mohsen Fayyaz, Luca Minciullo, Gianpiero Francesca, and Juergen Gall. Fast weakly supervised action + +segmentation using mutual consistency. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(10):6196-6208, 2021. 1, 2, 3, 6, 7, 8 +[39] Yaser Souri, Yazan Abu Farha, Fabien Despinoy, Gianpiero Francesca, and Juergen Gall. Fif: Fast inference approximation for action segmentation. In Proceedings of the DAGM German Conference on Pattern Recognition, pages 282-296. Springer, 2022. 2 +[40] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 3, 6 +[41] Heng Wang and Cordelia Schmid. Action recognition with improved trajectories. In Proceedings of the IEEE international conference on computer vision, pages 3551-3558, 2013. 6 +[42] Ruibin Xiong, Yunchang Yang, Di He, Kai Zheng, Shuxin Zheng, Chen Xing, Huishuai Zhang, Yanyan Lan, Liwei Wang, and Tieyan Liu. On layer normalization in the transformer architecture. In International Conference on Machine Learning, pages 10524-10533. PMLR, 2020. 3 +[43] Lian Xu, Wanli Ouyang, Mohammed Bennamoun, Farid Boussaid, and Dan Xu. Multi-class token transformer for weakly supervised semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4310-4319, 2022. 5 +[44] Di Yang, Yaohui Wang, Antitza Dantcheva, Quan Kong, Lorenzo Garattoni, Gianpiero Francesca, and Francois Bremond. Lac-latent action composition for skeleton-based action segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13679-13690, 2023. 1 +[45] Fangqiu Yi, Hongyu Wen, and Tingting Jiang. Asformer: Transformer for action segmentation. arXiv preprint arXiv:2110.08568, 2021. 3, 6 +[46] Runzhong Zhang, Suchen Wang, Yueqi Duan, Yansong Tang, Yue Zhang, and Yap-Peng Tan. Hoi-aware adaptive network for weakly-supervised action segmentation. In Proceedings of the Thirty-Second International Joint Conference on Artificial Intelligence, pages 1722–1730, 2023. 1, 2, 7 +[47] Dimitri Zhukov, Jean-Baptiste Alayrac, Ramadan Gokberk Cinbis, David Fouhey, Ivan Laptev, and Josef Sivic. Cross-task weakly supervised learning from instructional videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3537-3545, 2019. 1, 2, 6, 7 \ No newline at end of file diff --git a/2024/Efficient and Effective Weakly-Supervised Action Segmentation via Action-Transition-Aware Boundary Alignment/images.zip b/2024/Efficient and Effective Weakly-Supervised Action Segmentation via Action-Transition-Aware Boundary Alignment/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..bc132f624762a2ccc6c35906f35655912fcd0a8b --- /dev/null +++ b/2024/Efficient and Effective Weakly-Supervised Action Segmentation via Action-Transition-Aware Boundary Alignment/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a30a206cc952cfdc9baa1a9c34951ac7eb4a6776a845c264e2049943bdd5b339 +size 468031 diff --git a/2024/Efficient and Effective Weakly-Supervised Action Segmentation via Action-Transition-Aware Boundary Alignment/layout.json b/2024/Efficient and Effective Weakly-Supervised Action Segmentation via Action-Transition-Aware Boundary Alignment/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..d002e6aff4fdfa4ee5051209a71afafc1794e05b --- /dev/null +++ b/2024/Efficient and Effective Weakly-Supervised Action Segmentation via Action-Transition-Aware Boundary Alignment/layout.json @@ -0,0 +1,10869 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 89, + 103, + 504, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 103, + 504, + 140 + ], + "spans": [ + { + "bbox": [ + 89, + 103, + 504, + 140 + ], + "type": "text", + "content": "Efficient and Effective Weakly-Supervised Action Segmentation via Action-Transition-Aware Boundary Alignment" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 211, + 160, + 382, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 211, + 160, + 382, + 175 + ], + "spans": [ + { + "bbox": [ + 211, + 160, + 382, + 175 + ], + "type": "text", + "content": "Angchi Xu" + }, + { + "bbox": [ + 211, + 160, + 382, + 175 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 211, + 160, + 382, + 175 + ], + "type": "text", + "content": ", Wei-Shi Zheng" + }, + { + "bbox": [ + 211, + 160, + 382, + 175 + ], + "type": "inline_equation", + "content": "^{1,2,3,*}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 108, + 175, + 485, + 189 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 108, + 175, + 485, + 189 + ], + "spans": [ + { + "bbox": [ + 108, + 175, + 485, + 189 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 108, + 175, + 485, + 189 + ], + "type": "text", + "content": "School of Computer Science and Engineering, Sun Yat-sen University, China" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 60, + 190, + 534, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 60, + 190, + 534, + 204 + ], + "spans": [ + { + "bbox": [ + 60, + 190, + 534, + 204 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 60, + 190, + 534, + 204 + ], + "type": "text", + "content": "Key Laboratory of Machine Intelligence and Advanced Computing, Ministry of Education, China" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 68, + 204, + 525, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 204, + 525, + 217 + ], + "spans": [ + { + "bbox": [ + 68, + 204, + 525, + 217 + ], + "type": "text", + "content": "3Guangdong Province Key Laboratory of Information Security Technology, Guangzhou, China" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 179, + 220, + 414, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 179, + 220, + 414, + 232 + ], + "spans": [ + { + "bbox": [ + 179, + 220, + 414, + 232 + ], + "type": "text", + "content": "xuangch@mail2.sysu.edu.cn, wszheng@ieee.org" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 143, + 258, + 192, + 272 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 258, + 192, + 272 + ], + "spans": [ + { + "bbox": [ + 143, + 258, + 192, + 272 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 284, + 290, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 284, + 290, + 548 + ], + "spans": [ + { + "bbox": [ + 46, + 284, + 290, + 548 + ], + "type": "text", + "content": "Weakly-supervised action segmentation is a task of learning to partition a long video into several action segments, where training videos are only accompanied by transcripts (ordered list of actions). Most of existing methods need to infer pseudo segmentation for training by serial alignment between all frames and the transcript, which is time-consuming and hard to be parallelized while training. In this work, we aim to escape from this inefficient alignment with massive but redundant frames, and instead to directly localize a few action transitions for pseudo segmentation generation, where a transition refers to the change from an action segment to its next adjacent one in the transcript. As the true transitions are submerged in noisy boundaries due to intra-segment visual variation, we propose a novel Action-Transition-Aware Boundary Alignment (ATBA) framework to efficiently and effectively filter out noisy boundaries and detect transitions. In addition, to boost the semantic learning in the case that noise is inevitably present in the pseudo segmentation, we also introduce video-level losses to utilize the trusted video-level supervision. Extensive experiments show the effectiveness of our approach on both performance and training speed." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 571, + 128, + 583 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 571, + 128, + 583 + ], + "spans": [ + { + "bbox": [ + 47, + 571, + 128, + 583 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 591, + 287, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 591, + 287, + 676 + ], + "spans": [ + { + "bbox": [ + 46, + 591, + 287, + 676 + ], + "type": "text", + "content": "Action segmentation aims to partition a long untrimmed video into several segments and classify each segment into an action category [8, 12, 13, 15, 16, 29, 32, 35, 38, 44]. It is an important yet challenging task for instructional or procedural video understanding. Although fully-supervised action segmentation (FSAS) methods [1, 6, 13, 14, 27] have achieved great progress, they require frame-wise dense an-" + } + ] + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 306, + 258, + 545, + 418 + ], + "blocks": [ + { + "bbox": [ + 306, + 258, + 545, + 418 + ], + "lines": [ + { + "bbox": [ + 306, + 258, + 545, + 418 + ], + "spans": [ + { + "bbox": [ + 306, + 258, + 545, + 418 + ], + "type": "image", + "image_path": "8009ded21ae803606f935635d485eedc87f6f083ad130d5b86a91cd05b54ebe0.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 426, + 547, + 472 + ], + "lines": [ + { + "bbox": [ + 305, + 426, + 547, + 472 + ], + "spans": [ + { + "bbox": [ + 305, + 426, + 547, + 472 + ], + "type": "text", + "content": "Figure 1. Comparison of performance and training time of WSAS methods on the Breakfast dataset. MoF-The main metric of the task, the higher the better. *-Alignment-free methods. Our ATBA achieves the best performance with a very short training time." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 495, + 546, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 495, + 546, + 590 + ], + "spans": [ + { + "bbox": [ + 304, + 495, + 546, + 590 + ], + "type": "text", + "content": "notation, which is labor-intensive and time-consuming to collect. As a result, many works [5, 9, 25, 28, 34, 38, 46] explore the weakly-supervised action segmentation (WSAS) only requiring the transcript annotation, which refers to the ordered list of actions occurring in the video without their start and end times. The transcripts are less costly to obtain and can be accessed directly from video narrations or other meta data [2, 9, 18, 28, 30, 33, 47]." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 592, + 547, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 592, + 547, + 688 + ], + "spans": [ + { + "bbox": [ + 304, + 592, + 547, + 688 + ], + "type": "text", + "content": "Most of previous WSAS methods have to infer the pseudo segmentation (pseudo frame-wise labels) for training via a sequence alignment process between the video and given transcript, such as Viterbi [22, 23, 25, 28, 33, 34] or Dynamic Time Warping (DTW) [4, 5]. These alignment algorithms are usually designed in a recursive form which needs to be performed serially frame-by-frame and hard to be parallelized, resulting in very slow training process." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 689, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 547, + 714 + ], + "type": "text", + "content": "In this work, we argue that the frame-by-frame alignment is NOT necessary, since the pseudo segmentation is" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 683, + 136, + 693 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 683, + 136, + 693 + ], + "spans": [ + { + "bbox": [ + 58, + 683, + 136, + 693 + ], + "type": "text", + "content": "* Corresponding author." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 59, + 693, + 286, + 704 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 693, + 286, + 704 + ], + "spans": [ + { + "bbox": [ + 59, + 693, + 286, + 704 + ], + "type": "text", + "content": "1Code is available at https://github.com/iSEE-" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 48, + 704, + 157, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 704, + 157, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 704, + 157, + 712 + ], + "type": "text", + "content": "Laboratory/CVPR24_ATBA." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18253" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 48, + 71, + 286, + 166 + ], + "blocks": [ + { + "bbox": [ + 48, + 71, + 286, + 166 + ], + "lines": [ + { + "bbox": [ + 48, + 71, + 286, + 166 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 286, + 166 + ], + "type": "image", + "image_path": "2f08921554523255d4b8e94954c42c301787ca2957e1de5eef3b6567101e30a9.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 174, + 289, + 307 + ], + "lines": [ + { + "bbox": [ + 46, + 174, + 289, + 307 + ], + "spans": [ + { + "bbox": [ + 46, + 174, + 289, + 307 + ], + "type": "text", + "content": "Figure 2. The necessity of proposed ATBA. The example is P54-webcam01-P54-coffee in Breakfast dataset. GT-The ground-truth segmentation. C.A.Bdy.-Only class-agnostic boundary detection is applied (Exp.1 of Table 3). Acc.-The accuracy of pseudo segmentation. In the video clip around the \"star\" point, the coffee pot undergoes a change from being picked up to tilted pouring within the segment \"Pour Coffee\", and this noisy visual change is incorrectly detected. In addition, although two boundaries are correctly detected by the \"C.A.Bdy.\" (diamonds), they correspond to incorrect transitions due to one false positive error (star), resulting in complete dislocation of segments within the dashed box. Best viewed in color." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 329, + 287, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 329, + 287, + 509 + ], + "spans": [ + { + "bbox": [ + 46, + 329, + 287, + 509 + ], + "type": "text", + "content": "fundamentally determined by the locations of a small number of action transitions (i.e., the change from an action segment to its next adjacent action segment in the transcript). Hence, the pseudo segmentation generation can be viewed as a transition detection problem, implying the way to more efficient designs. Intuitively, action transitions are often accompanied by significant visual changes, and there are already many approaches that can detect class-agnostic action boundaries based on these changes [12, 20]. However, due to the intra-segment visual variation and sub-segments under finer granularity, there are numerous noisy boundaries not corresponding to any transitions. Moreover, as the class-agnostic way cannot guarantee correct correspondence between the boundaries and transitions, even slight errors in the detection can result in severe deviation (Fig. 2)." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 510, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 510, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 510, + 288, + 714 + ], + "type": "text", + "content": "To overcome the above noisy boundary issue, we propose an efficient and effective framework for WSAS, termed Action-Transition-Aware Boundary Alignment (ATBA), which directly detects the transitions for faster and effective pseudo segmentation generation. To tolerate the noisy boundaries, the ATBA generates more class-agnostic boundaries than the number of transitions as candidates, and then determines a subset from candidates that optimally matches all desired transitions via a drop-allowed alignment algorithm. Furthermore, to fortify the semantic learning under the inevitable noise in pseudo segmentation, we also introduce video-level losses to make use of the trusted video-level supervision. Our ATBA is efficient, because the number of generated candidates will be proportional to the length of the transcript, and therefore the complexity of alignment is now independent of the very long video length. Moreover, other computations" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 304, + 72, + 545, + 132 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 132 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 132 + ], + "type": "text", + "content": "required by ATBA, i.e., measuring how likely a frame is to be a boundary and how likely a candidate corresponds to a desired transition, are both built on a convolution-like algorithm inspired by [20], which can be parallelized on GPUs efficiently." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 133, + 545, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 133, + 545, + 203 + ], + "spans": [ + { + "bbox": [ + 304, + 133, + 545, + 203 + ], + "type": "text", + "content": "For inference, we directly adopt the results from the trained frame-wise classifier, without the need for any alignment processing with retrieved or predicted transcript (the ground-truth transcript is not available during inference) like previous WSAS methods [4, 5, 25, 28, 34, 38], which also improves the inference efficiency." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 204, + 546, + 372 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 204, + 546, + 372 + ], + "spans": [ + { + "bbox": [ + 304, + 204, + 546, + 372 + ], + "type": "text", + "content": "In summary, our contributions are as follows. (1) We propose to directly localize the action transitions for efficient pseudo segmentation generation during training, without the need of time-consuming frame-by-frame alignment. (2) For robustness to noisy boundaries, we propose a novel ATBA framework to effectively determine boundaries corresponding to each transition. Video-level losses are also introduced to regularize the semantic learning involving the unavoidable noise in the pseudo segmentation. Experiments are conducted on three popular datasets to evaluate our approach: Breakfast [21], Hollywood Extended [2] and CrossTask [47]. Our ATBA achieves state-of-the-art or comparable results with one of the fastest training speed (Fig. 1), demonstrating the effectiveness of ours." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 382, + 392, + 395 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 382, + 392, + 395 + ], + "spans": [ + { + "bbox": [ + 306, + 382, + 392, + 395 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 403, + 545, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 403, + 545, + 510 + ], + "spans": [ + { + "bbox": [ + 304, + 403, + 545, + 510 + ], + "type": "text", + "content": "Weakly-supervised action segmentation methods learn to partition a video into several action segments from training videos only annotated by transcripts [2, 4, 5, 9, 18, 22, 23, 25, 28, 33, 34, 38, 39, 46]. Despite different optimization objectives, most of them generate the pseudo segmentation for training by solving alignment objectives between two sequences (video and transcript) via Connectionist Temporal Classification (CTC) [18], Viterbi [22, 23, 25, 28, 33, 34] or Dynamic Time Warping (DTW) [4, 5]." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 510, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 510, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 510, + 546, + 713 + ], + "type": "text", + "content": "Specifically, [18] proposes an extended version of CTC to evaluate all valid alignments between videos and transcripts, which additionally takes the visual similarities of frames into account. Inspired by speech recognition, [22, 23, 33] all use the Hidden Markov Model (HMM) to model the relationship between videos and actions. [9] always generates uniform segmentation but iteratively adjusts the boundaries by inserting repeating actions into the transcript. [34] proposes an alignment objective based on explicit context and length models, which can be solved by Viterbi, and the solved optimal alignment would serve as pseudo labels to train the frame-wise classifier. [25] and [28] both focus on novel learning objectives, but still require the optimal pseudo segmentation produced by Viterbi. [4, 5] both learn from the contrast of aligning the video to the ground-truth transcript and negative transcripts, where the alignment is performed by DTW. Whether using CTC, Viterbi or DTW," + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "18254" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 52, + 73, + 545, + 274 + ], + "blocks": [ + { + "bbox": [ + 52, + 73, + 545, + 274 + ], + "lines": [ + { + "bbox": [ + 52, + 73, + 545, + 274 + ], + "spans": [ + { + "bbox": [ + 52, + 73, + 545, + 274 + ], + "type": "image", + "image_path": "098625b2798000cf20009f8d545e6607bc334dc4819785ae7eacb5cdbf259f1c.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 281, + 546, + 316 + ], + "lines": [ + { + "bbox": [ + 46, + 281, + 546, + 316 + ], + "spans": [ + { + "bbox": [ + 46, + 281, + 546, + 316 + ], + "type": "text", + "content": "Figure 3. The overall framework. We propose an Action-Transition-Aware Boundary Alignment (ATBA) framework, which takes the class-agnostic boundary pattern and action transition pattern together into account to efficiently generate pseudo labels. The trusted video-level supervision is also utilized to further enhance the performance." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 336, + 286, + 359 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 336, + 286, + 359 + ], + "spans": [ + { + "bbox": [ + 46, + 336, + 286, + 359 + ], + "type": "text", + "content": "the above approaches except [9] require frame-by-frame serial calculation, which are inefficient." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 361, + 287, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 361, + 287, + 468 + ], + "spans": [ + { + "bbox": [ + 46, + 361, + 287, + 468 + ], + "type": "text", + "content": "Recently, some efficient methods are proposed with alignment-free design. [38] learns from the mutual consistency between two forms of a segmentation (i.e., framewise classification and category/length pairs). [29] proposes a loss to enforce the output order of any two actions to be consistent with the transcript. In this work, we also propose an efficient framework with different technical roadmap, and our performance is better at the same level of training speed." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 479, + 113, + 493 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 479, + 113, + 493 + ], + "spans": [ + { + "bbox": [ + 47, + 479, + 113, + 493 + ], + "type": "text", + "content": "3. Approach" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 499, + 160, + 510 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 499, + 160, + 510 + ], + "spans": [ + { + "bbox": [ + 47, + 499, + 160, + 510 + ], + "type": "text", + "content": "3.1. Problem Statement" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 517, + 287, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 517, + 287, + 685 + ], + "spans": [ + { + "bbox": [ + 46, + 517, + 287, + 685 + ], + "type": "text", + "content": "Action segmentation is a task of partitioning a video into several temporal segments with action labels, which is equivalent to predicting the action categories of each frame. Formally, given a sequence of " + }, + { + "bbox": [ + 46, + 517, + 287, + 685 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 46, + 517, + 287, + 685 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 46, + 517, + 287, + 685 + ], + "type": "text", + "content": "-dimensional frame-wise features " + }, + { + "bbox": [ + 46, + 517, + 287, + 685 + ], + "type": "inline_equation", + "content": "\\mathbf{X} = [\\pmb{x}_1,\\dots,\\pmb{x}_T] \\in \\mathbb{R}^{T\\times d}" + }, + { + "bbox": [ + 46, + 517, + 287, + 685 + ], + "type": "text", + "content": " for a video with " + }, + { + "bbox": [ + 46, + 517, + 287, + 685 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 46, + 517, + 287, + 685 + ], + "type": "text", + "content": " frames, the goal is to predict a sequence of actions " + }, + { + "bbox": [ + 46, + 517, + 287, + 685 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{Y}} = [\\hat{y}_1,\\dots,\\hat{y}_T]" + }, + { + "bbox": [ + 46, + 517, + 287, + 685 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 517, + 287, + 685 + ], + "type": "inline_equation", + "content": "\\hat{y}_t \\in \\mathcal{C}" + }, + { + "bbox": [ + 46, + 517, + 287, + 685 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 517, + 287, + 685 + ], + "type": "inline_equation", + "content": "\\mathcal{C} = \\{1,2,\\dots,|\\mathcal{C}|\\}" + }, + { + "bbox": [ + 46, + 517, + 287, + 685 + ], + "type": "text", + "content": " is the set of action categories across the dataset (including the background). Under the setting of WSAS, the frame-wise ground-truth " + }, + { + "bbox": [ + 46, + 517, + 287, + 685 + ], + "type": "inline_equation", + "content": "\\mathcal{Y} = [y_1,\\dots,y_T]" + }, + { + "bbox": [ + 46, + 517, + 287, + 685 + ], + "type": "text", + "content": " is NOT available during training. Instead, the ordered list of actions called transcript " + }, + { + "bbox": [ + 46, + 517, + 287, + 685 + ], + "type": "inline_equation", + "content": "\\mathcal{A} = [a_1,\\dots,a_M]" + }, + { + "bbox": [ + 46, + 517, + 287, + 685 + ], + "type": "text", + "content": " is provided (including the background segments), where " + }, + { + "bbox": [ + 46, + 517, + 287, + 685 + ], + "type": "inline_equation", + "content": "a_{m} \\in \\mathcal{C}" + }, + { + "bbox": [ + 46, + 517, + 287, + 685 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 517, + 287, + 685 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 46, + 517, + 287, + 685 + ], + "type": "text", + "content": " is the total number of action segments in the video. The action transitions of " + }, + { + "bbox": [ + 46, + 517, + 287, + 685 + ], + "type": "inline_equation", + "content": "\\mathcal{A}" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 335, + 512, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 335, + 512, + 350 + ], + "spans": [ + { + "bbox": [ + 306, + 335, + 512, + 350 + ], + "type": "text", + "content": "are naturally formulated as " + }, + { + "bbox": [ + 306, + 335, + 512, + 350 + ], + "type": "inline_equation", + "content": "\\mathcal{R} = \\{(a_r,a_{r + 1})\\}_{r = 1}^{M - 1}" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 358, + 373, + 369 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 358, + 373, + 369 + ], + "spans": [ + { + "bbox": [ + 306, + 358, + 373, + 369 + ], + "type": "text", + "content": "3.2. Overview" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 377, + 545, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 377, + 545, + 510 + ], + "spans": [ + { + "bbox": [ + 304, + 377, + 545, + 510 + ], + "type": "text", + "content": "Our proposed framework is illustrated in Fig. 3. At first, the input sequence " + }, + { + "bbox": [ + 304, + 377, + 545, + 510 + ], + "type": "inline_equation", + "content": "\\mathbf{X}" + }, + { + "bbox": [ + 304, + 377, + 545, + 510 + ], + "type": "text", + "content": " is further encoded by a temporal network to generate more task-relevant representations " + }, + { + "bbox": [ + 304, + 377, + 545, + 510 + ], + "type": "inline_equation", + "content": "\\mathbf{X}' \\in \\mathbb{R}^{T \\times d'}" + }, + { + "bbox": [ + 304, + 377, + 545, + 510 + ], + "type": "text", + "content": " (Sec. 3.3), then a classifier shared along the temporal axis followed by a category softmax activation will predict the frame-wise class probabilities " + }, + { + "bbox": [ + 304, + 377, + 545, + 510 + ], + "type": "inline_equation", + "content": "\\mathbf{P} = [\\mathbf{p}_1, \\dots, \\mathbf{p}_T] \\in \\mathbb{R}^{T \\times |C|}" + }, + { + "bbox": [ + 304, + 377, + 545, + 510 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 304, + 377, + 545, + 510 + ], + "type": "inline_equation", + "content": "\\mathbf{X}'" + }, + { + "bbox": [ + 304, + 377, + 545, + 510 + ], + "type": "text", + "content": ". After that, the Action-Transition-Aware Boundary Alignment (ATBA) module takes " + }, + { + "bbox": [ + 304, + 377, + 545, + 510 + ], + "type": "inline_equation", + "content": "\\mathbf{P}" + }, + { + "bbox": [ + 304, + 377, + 545, + 510 + ], + "type": "text", + "content": " and the transcript " + }, + { + "bbox": [ + 304, + 377, + 545, + 510 + ], + "type": "inline_equation", + "content": "\\mathcal{A}" + }, + { + "bbox": [ + 304, + 377, + 545, + 510 + ], + "type": "text", + "content": " as input to infer the pseudo frame-wise labels " + }, + { + "bbox": [ + 304, + 377, + 545, + 510 + ], + "type": "inline_equation", + "content": "\\widetilde{\\mathcal{Y}} = [\\widetilde{y}_1, \\dots, \\widetilde{y}_T]" + }, + { + "bbox": [ + 304, + 377, + 545, + 510 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 377, + 545, + 510 + ], + "type": "inline_equation", + "content": "\\widetilde{y}_t \\in C" + }, + { + "bbox": [ + 304, + 377, + 545, + 510 + ], + "type": "text", + "content": " (Sec. 3.4). Finally, " + }, + { + "bbox": [ + 304, + 377, + 545, + 510 + ], + "type": "inline_equation", + "content": "\\widetilde{\\mathcal{Y}}" + }, + { + "bbox": [ + 304, + 377, + 545, + 510 + ], + "type": "text", + "content": " is used back to supervise " + }, + { + "bbox": [ + 304, + 377, + 545, + 510 + ], + "type": "inline_equation", + "content": "\\mathbf{P}" + }, + { + "bbox": [ + 304, + 377, + 545, + 510 + ], + "type": "text", + "content": " by a standard cross entropy:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 347, + 520, + 545, + 554 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 347, + 520, + 545, + 554 + ], + "spans": [ + { + "bbox": [ + 347, + 520, + 545, + 554 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {c l s}} = - \\frac {1}{T} \\sum_ {t = 1} ^ {T} \\sum_ {c = 1} ^ {| c |} \\mathbb {I} (\\widetilde {y} _ {t} = c) \\log \\boldsymbol {P} _ {t, c}, \\tag {1}", + "image_path": "17ce06956cdd3995c74070aad587820236fe8f2e465125022f01e547b81cc95f.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 564, + 545, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 564, + 545, + 647 + ], + "spans": [ + { + "bbox": [ + 304, + 564, + 545, + 647 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 564, + 545, + 647 + ], + "type": "inline_equation", + "content": "\\mathbb{I}(\\cdot)" + }, + { + "bbox": [ + 304, + 564, + 545, + 647 + ], + "type": "text", + "content": " is the indicator function which returns 1 if the condition is satisfied and 0 otherwise. Note that the pseudo frame-wise labels are inferred at each iteration for a batch of data and used immediately for current training at the same iteration. The additional video-level losses are stated in Sec. 3.5, and the training and inference processes are described in Sec. 3.6." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 658, + 416, + 670 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 658, + 416, + 670 + ], + "spans": [ + { + "bbox": [ + 306, + 658, + 416, + 670 + ], + "type": "text", + "content": "3.3. Temporal Network" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "content": "We employ a slightly modified pre-norm [42] Transformer [40] encoder with learnable positional embeddings as the temporal network for feature learning. Following [11, 45]," + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 47, + 693, + 286, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 693, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 693, + 286, + 713 + ], + "type": "text", + "content": "2POC [29] is a set-supervised method but can be extended to transcript supervision naturally, so we cite its corresponding results for comparison." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18255" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 72, + 70, + 157, + 154 + ], + "blocks": [ + { + "bbox": [ + 72, + 70, + 157, + 154 + ], + "lines": [ + { + "bbox": [ + 72, + 70, + 157, + 154 + ], + "spans": [ + { + "bbox": [ + 72, + 70, + 157, + 154 + ], + "type": "image", + "image_path": "44df138bd8e2149a40074b7c9f1f5ee9a54d566ce63806497cc63237ce56f3b2.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 109, + 156, + 119, + 166 + ], + "lines": [ + { + "bbox": [ + 109, + 156, + 119, + 166 + ], + "spans": [ + { + "bbox": [ + 109, + 156, + 119, + 166 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 179, + 93, + 263, + 120 + ], + "blocks": [ + { + "bbox": [ + 179, + 93, + 263, + 120 + ], + "lines": [ + { + "bbox": [ + 179, + 93, + 263, + 120 + ], + "spans": [ + { + "bbox": [ + 179, + 93, + 263, + 120 + ], + "type": "image", + "image_path": "d0e9d0643e18099af8f651bbd6efe538db1c372cd8e8b0c0965054bb73432cc9.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 47, + 175, + 286, + 197 + ], + "lines": [ + { + "bbox": [ + 47, + 175, + 286, + 197 + ], + "spans": [ + { + "bbox": [ + 47, + 175, + 286, + 197 + ], + "type": "text", + "content": "Figure 4. (a) A " + }, + { + "bbox": [ + 47, + 175, + 286, + 197 + ], + "type": "inline_equation", + "content": "7 \\times 7" + }, + { + "bbox": [ + 47, + 175, + 286, + 197 + ], + "type": "text", + "content": " template for class-agnostic boundary scoring. \n(b) A " + }, + { + "bbox": [ + 47, + 175, + 286, + 197 + ], + "type": "inline_equation", + "content": "2 \\times 7" + }, + { + "bbox": [ + 47, + 175, + 286, + 197 + ], + "type": "text", + "content": " template for action transition scoring." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 217, + 286, + 253 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 217, + 286, + 253 + ], + "spans": [ + { + "bbox": [ + 47, + 217, + 286, + 253 + ], + "type": "text", + "content": "the vanilla full self-attention is replaced with a pyramid hierarchical local attention to better adapt to the action segmentation task (see supplemental material for more details)." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 259, + 286, + 272 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 259, + 286, + 272 + ], + "spans": [ + { + "bbox": [ + 47, + 259, + 286, + 272 + ], + "type": "text", + "content": "3.4. Action-Transition-Aware Boundary Alignment" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 277, + 287, + 349 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 277, + 287, + 349 + ], + "spans": [ + { + "bbox": [ + 47, + 277, + 287, + 349 + ], + "type": "text", + "content": "The ATBA is the core component of our framework. It generates the pseudo frame-wise labels " + }, + { + "bbox": [ + 47, + 277, + 287, + 349 + ], + "type": "inline_equation", + "content": "\\widetilde{\\mathcal{V}}" + }, + { + "bbox": [ + 47, + 277, + 287, + 349 + ], + "type": "text", + "content": " by inferring the boundaries corresponding to " + }, + { + "bbox": [ + 47, + 277, + 287, + 349 + ], + "type": "inline_equation", + "content": "M - 1" + }, + { + "bbox": [ + 47, + 277, + 287, + 349 + ], + "type": "text", + "content": " action transitions from the predicted class probabilities " + }, + { + "bbox": [ + 47, + 277, + 287, + 349 + ], + "type": "inline_equation", + "content": "\\pmb{P}" + }, + { + "bbox": [ + 47, + 277, + 287, + 349 + ], + "type": "text", + "content": ". Once they are found, " + }, + { + "bbox": [ + 47, + 277, + 287, + 349 + ], + "type": "inline_equation", + "content": "\\widetilde{\\mathcal{V}}" + }, + { + "bbox": [ + 47, + 277, + 287, + 349 + ], + "type": "text", + "content": " can be naturally generated by assigning the action labels of " + }, + { + "bbox": [ + 47, + 277, + 287, + 349 + ], + "type": "inline_equation", + "content": "\\mathcal{A}" + }, + { + "bbox": [ + 47, + 277, + 287, + 349 + ], + "type": "text", + "content": " one-by-one into the intervals between these boundaries." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 350, + 287, + 409 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 350, + 287, + 409 + ], + "spans": [ + { + "bbox": [ + 46, + 350, + 287, + 409 + ], + "type": "text", + "content": "Briefly, the ATBA first generates a set of candidate boundaries via a class-agnostic way, and then finds " + }, + { + "bbox": [ + 46, + 350, + 287, + 409 + ], + "type": "inline_equation", + "content": "M - 1" + }, + { + "bbox": [ + 46, + 350, + 287, + 409 + ], + "type": "text", + "content": " points from this set that optimally match all transitions of " + }, + { + "bbox": [ + 46, + 350, + 287, + 409 + ], + "type": "inline_equation", + "content": "\\mathcal{A}" + }, + { + "bbox": [ + 46, + 350, + 287, + 409 + ], + "type": "text", + "content": " via a dynamic programming (DP) algorithm. We describe the details in the following." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 410, + 287, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 410, + 287, + 639 + ], + "spans": [ + { + "bbox": [ + 47, + 410, + 287, + 639 + ], + "type": "text", + "content": "- Class-Agnostic Boundary Scoring. Firstly, the ATBA calculates the class-agnostic boundary scores " + }, + { + "bbox": [ + 47, + 410, + 287, + 639 + ], + "type": "inline_equation", + "content": "\\mathcal{V}^{\\mathrm{b}} = [v_1^{\\mathrm{b}},\\dots,v_T^{\\mathrm{b}}]" + }, + { + "bbox": [ + 47, + 410, + 287, + 639 + ], + "type": "text", + "content": " for each timestamp. We employ a pattern-matching-based scoring method proposed by a generic event boundary detection approach termed UBoCo [20], which considers the pattern of each frame's neighborhood. Specifically, for timestamp " + }, + { + "bbox": [ + 47, + 410, + 287, + 639 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 47, + 410, + 287, + 639 + ], + "type": "text", + "content": ", a pairwise similarity matrix " + }, + { + "bbox": [ + 47, + 410, + 287, + 639 + ], + "type": "inline_equation", + "content": "\\Gamma^{(t)}\\in \\mathbb{R}^{w^{\\mathrm{b}}\\times w^{\\mathrm{b}}}" + }, + { + "bbox": [ + 47, + 410, + 287, + 639 + ], + "type": "text", + "content": " is calculated within a local window with size " + }, + { + "bbox": [ + 47, + 410, + 287, + 639 + ], + "type": "inline_equation", + "content": "w^{\\mathrm{b}}" + }, + { + "bbox": [ + 47, + 410, + 287, + 639 + ], + "type": "text", + "content": " centered at " + }, + { + "bbox": [ + 47, + 410, + 287, + 639 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 47, + 410, + 287, + 639 + ], + "type": "text", + "content": ", whose " + }, + { + "bbox": [ + 47, + 410, + 287, + 639 + ], + "type": "inline_equation", + "content": "(i,j)" + }, + { + "bbox": [ + 47, + 410, + 287, + 639 + ], + "type": "text", + "content": "-entry represents the class-agnostic similarity between " + }, + { + "bbox": [ + 47, + 410, + 287, + 639 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 47, + 410, + 287, + 639 + ], + "type": "text", + "content": "-th and " + }, + { + "bbox": [ + 47, + 410, + 287, + 639 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 47, + 410, + 287, + 639 + ], + "type": "text", + "content": "-th frames inside the window, with values ranging from -1 to 1 (see the computational details in supplementary material). Clearly, if " + }, + { + "bbox": [ + 47, + 410, + 287, + 639 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 47, + 410, + 287, + 639 + ], + "type": "text", + "content": " is a boundary, the frame feature should change dramatically at " + }, + { + "bbox": [ + 47, + 410, + 287, + 639 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 47, + 410, + 287, + 639 + ], + "type": "text", + "content": " and keep stable elsewhere, so its " + }, + { + "bbox": [ + 47, + 410, + 287, + 639 + ], + "type": "inline_equation", + "content": "\\Gamma^{(t)}" + }, + { + "bbox": [ + 47, + 410, + 287, + 639 + ], + "type": "text", + "content": " should show the pattern of that the values in the upper-left and lower-right areas are close to 1 and otherwise close to -1. Hence a template " + }, + { + "bbox": [ + 47, + 410, + 287, + 639 + ], + "type": "inline_equation", + "content": "\\Omega^{\\mathrm{b}}\\in \\mathbb{R}^{w^{\\mathrm{b}}\\times w^{\\mathrm{b}}}" + }, + { + "bbox": [ + 47, + 410, + 287, + 639 + ], + "type": "text", + "content": " like Fig. 4(a) is designed to capture this pattern and output the class-agnostic boundary score for each " + }, + { + "bbox": [ + 47, + 410, + 287, + 639 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 47, + 410, + 287, + 639 + ], + "type": "text", + "content": " by a correlation operation:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 99, + 647, + 287, + 681 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 99, + 647, + 287, + 681 + ], + "spans": [ + { + "bbox": [ + 99, + 647, + 287, + 681 + ], + "type": "interline_equation", + "content": "v _ {t} ^ {\\mathrm {b}} = \\frac {1}{w ^ {\\mathrm {b}} \\times w ^ {\\mathrm {b}}} \\sum_ {i = 1} ^ {w ^ {\\mathrm {b}}} \\sum_ {j = 1} ^ {w ^ {\\mathrm {b}}} \\Omega_ {i, j} ^ {\\mathrm {b}} \\Gamma_ {i, j} ^ {(t)}. \\tag {2}", + "image_path": "827c9fdd7dd0dd81e98ba445a3790f5f8c6483d396b41a9c1ce88661cc381b91.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "type": "text", + "content": "- Candidate Boundary Selection. After calculating " + }, + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\nu^{\\mathrm{b}}" + }, + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "type": "text", + "content": ", we select a set of " + }, + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "type": "text", + "content": " candidate boundaries " + }, + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathcal{B}} = \\{b_k\\}_{k=1}^K" + }, + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "type": "text", + "content": ", where" + } + ] + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 331, + 71, + 524, + 138 + ], + "blocks": [ + { + "bbox": [ + 215, + 156, + 227, + 166 + ], + "lines": [ + { + "bbox": [ + 215, + 156, + 227, + 166 + ], + "spans": [ + { + "bbox": [ + 215, + 156, + 227, + 166 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 331, + 71, + 524, + 138 + ], + "lines": [ + { + "bbox": [ + 331, + 71, + 524, + 138 + ], + "spans": [ + { + "bbox": [ + 331, + 71, + 524, + 138 + ], + "type": "image", + "image_path": "7fa9cab4368f949836b38db60133b5627548cd5a104037341dbdf1c7b6ee3418.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 147, + 545, + 201 + ], + "lines": [ + { + "bbox": [ + 305, + 147, + 545, + 201 + ], + "spans": [ + { + "bbox": [ + 305, + 147, + 545, + 201 + ], + "type": "text", + "content": "Figure 5. Illustration of the alignment between action transitions " + }, + { + "bbox": [ + 305, + 147, + 545, + 201 + ], + "type": "inline_equation", + "content": "\\mathcal{R}" + }, + { + "bbox": [ + 305, + 147, + 545, + 201 + ], + "type": "text", + "content": " and candidate boundaries " + }, + { + "bbox": [ + 305, + 147, + 545, + 201 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathcal{B}}" + }, + { + "bbox": [ + 305, + 147, + 545, + 201 + ], + "type": "text", + "content": ". Blue circles are aligned and gray ones are dropped. (a) A valid alignment. (b) An invalid alignment. The red dashed arrow violates the ordering consistency. Best viewed in color." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 223, + 545, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 223, + 545, + 354 + ], + "spans": [ + { + "bbox": [ + 304, + 223, + 545, + 354 + ], + "type": "inline_equation", + "content": "1 < b_{1} < \\ldots < b_{K} \\leq T" + }, + { + "bbox": [ + 304, + 223, + 545, + 354 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 223, + 545, + 354 + ], + "type": "inline_equation", + "content": "K > M - 1" + }, + { + "bbox": [ + 304, + 223, + 545, + 354 + ], + "type": "text", + "content": ". The selection is performed by a simple greedy strategy with non maximum suppression inspired by [12], i.e., each time we select one timestamp with current highest score " + }, + { + "bbox": [ + 304, + 223, + 545, + 354 + ], + "type": "inline_equation", + "content": "v_{t}^{\\mathrm{b}}" + }, + { + "bbox": [ + 304, + 223, + 545, + 354 + ], + "type": "text", + "content": ", and invalidate its neighborhood to avoid selecting multiple timestamps corresponding to one same boundary, until the number of selected timestamps reach an upper bound or all remaining timestamps are invalid. The radius of the invalid interval is set adaptively to " + }, + { + "bbox": [ + 304, + 223, + 545, + 354 + ], + "type": "inline_equation", + "content": "\\frac{\\mu T}{M}" + }, + { + "bbox": [ + 304, + 223, + 545, + 354 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 223, + 545, + 354 + ], + "type": "inline_equation", + "content": "\\mu \\in [0,1]" + }, + { + "bbox": [ + 304, + 223, + 545, + 354 + ], + "type": "text", + "content": " is a hyper-parameter, and in practice, the upper bound for " + }, + { + "bbox": [ + 304, + 223, + 545, + 354 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 304, + 223, + 545, + 354 + ], + "type": "text", + "content": " is set to " + }, + { + "bbox": [ + 304, + 223, + 545, + 354 + ], + "type": "inline_equation", + "content": "\\lambda(M - 1)" + }, + { + "bbox": [ + 304, + 223, + 545, + 354 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 223, + 545, + 354 + ], + "type": "inline_equation", + "content": "\\lambda \\in \\mathbb{N}_{+}" + }, + { + "bbox": [ + 304, + 223, + 545, + 354 + ], + "type": "text", + "content": " is also a hyper-parameter." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 355, + 545, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 355, + 545, + 498 + ], + "spans": [ + { + "bbox": [ + 304, + 355, + 545, + 498 + ], + "type": "text", + "content": "- Action Transition Scoring. As mentioned in Sec. 1, the class-agnostic scores " + }, + { + "bbox": [ + 304, + 355, + 545, + 498 + ], + "type": "inline_equation", + "content": "\\mathcal{V}^{\\mathrm{b}}" + }, + { + "bbox": [ + 304, + 355, + 545, + 498 + ], + "type": "text", + "content": " are not enough to detect action transitions. To this end, we then calculate an action transition score matrix " + }, + { + "bbox": [ + 304, + 355, + 545, + 498 + ], + "type": "inline_equation", + "content": "V^{\\mathrm{a}}\\in \\mathbb{R}^{K\\times (M - 1)}" + }, + { + "bbox": [ + 304, + 355, + 545, + 498 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 355, + 545, + 498 + ], + "type": "inline_equation", + "content": "V_{k,r}^{\\mathrm{a}}" + }, + { + "bbox": [ + 304, + 355, + 545, + 498 + ], + "type": "text", + "content": " measures the possibility that the " + }, + { + "bbox": [ + 304, + 355, + 545, + 498 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 355, + 545, + 498 + ], + "type": "text", + "content": "-th candidate boundary corresponds to the " + }, + { + "bbox": [ + 304, + 355, + 545, + 498 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 304, + 355, + 545, + 498 + ], + "type": "text", + "content": "-th transition, i.e., separates the " + }, + { + "bbox": [ + 304, + 355, + 545, + 498 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 304, + 355, + 545, + 498 + ], + "type": "text", + "content": "-th and " + }, + { + "bbox": [ + 304, + 355, + 545, + 498 + ], + "type": "inline_equation", + "content": "(r + 1)" + }, + { + "bbox": [ + 304, + 355, + 545, + 498 + ], + "type": "text", + "content": "-th action segments. This matrix is also calculated via pattern matching. Clearly, if candidate " + }, + { + "bbox": [ + 304, + 355, + 545, + 498 + ], + "type": "inline_equation", + "content": "b_{k}" + }, + { + "bbox": [ + 304, + 355, + 545, + 498 + ], + "type": "text", + "content": " corresponds to the " + }, + { + "bbox": [ + 304, + 355, + 545, + 498 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 304, + 355, + 545, + 498 + ], + "type": "text", + "content": "-th transition, the classifier's activation for class " + }, + { + "bbox": [ + 304, + 355, + 545, + 498 + ], + "type": "inline_equation", + "content": "a_{r}" + }, + { + "bbox": [ + 304, + 355, + 545, + 498 + ], + "type": "text", + "content": " should drop sharply after " + }, + { + "bbox": [ + 304, + 355, + 545, + 498 + ], + "type": "inline_equation", + "content": "b_{k}" + }, + { + "bbox": [ + 304, + 355, + 545, + 498 + ], + "type": "text", + "content": ", while rise for class " + }, + { + "bbox": [ + 304, + 355, + 545, + 498 + ], + "type": "inline_equation", + "content": "a_{r + 1}" + }, + { + "bbox": [ + 304, + 355, + 545, + 498 + ], + "type": "text", + "content": ". Hence, a template " + }, + { + "bbox": [ + 304, + 355, + 545, + 498 + ], + "type": "inline_equation", + "content": "\\Omega^{\\mathrm{a}}\\in \\mathbb{R}^{2\\times w^{\\mathrm{a}}}" + }, + { + "bbox": [ + 304, + 355, + 545, + 498 + ], + "type": "text", + "content": " with temporal size " + }, + { + "bbox": [ + 304, + 355, + 545, + 498 + ], + "type": "inline_equation", + "content": "w^{\\mathrm{a}}" + }, + { + "bbox": [ + 304, + 355, + 545, + 498 + ], + "type": "text", + "content": " like Fig. 4(b) is employed to detect this pattern around " + }, + { + "bbox": [ + 304, + 355, + 545, + 498 + ], + "type": "inline_equation", + "content": "b_{k}" + }, + { + "bbox": [ + 304, + 355, + 545, + 498 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 335, + 507, + 545, + 541 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 507, + 545, + 541 + ], + "spans": [ + { + "bbox": [ + 335, + 507, + 545, + 541 + ], + "type": "interline_equation", + "content": "\\boldsymbol {V} _ {k, r} ^ {\\mathrm {a}} = \\frac {1}{2 w ^ {\\mathrm {a}}} \\sum_ {i = 1} ^ {2} \\sum_ {j = 1} ^ {w ^ {\\mathrm {a}}} \\boldsymbol {\\Omega} _ {i, j} ^ {\\mathrm {a}} \\boldsymbol {P} _ {\\text {i n d} ^ {\\mathrm {a}} \\left(b _ {k}, j\\right), a _ {r + i - 1}}, \\tag {3}", + "image_path": "02991cfdcdb64d26ece5b147021c53f2258f482d72fd751c55f6f8229fc8c793.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 356, + 544, + 495, + 567 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 356, + 544, + 495, + 567 + ], + "spans": [ + { + "bbox": [ + 356, + 544, + 495, + 567 + ], + "type": "interline_equation", + "content": "\\operatorname {i n d} ^ {\\mathrm {a}} \\left(b _ {k}, j\\right) = b _ {k} - \\left\\lfloor \\frac {w ^ {\\mathrm {a}}}{2} \\right\\rfloor + j - 1,", + "image_path": "bb0b093b45127f549b192a6582ef789e989a8d21bbc89e698d8921157bb5ca13.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 574, + 545, + 622 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 574, + 545, + 622 + ], + "spans": [ + { + "bbox": [ + 304, + 574, + 545, + 622 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 574, + 545, + 622 + ], + "type": "inline_equation", + "content": "\\mathrm{ind}^{\\mathrm{a}}(b_k,j)" + }, + { + "bbox": [ + 304, + 574, + 545, + 622 + ], + "type": "text", + "content": " is the index transform from the index " + }, + { + "bbox": [ + 304, + 574, + 545, + 622 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 304, + 574, + 545, + 622 + ], + "type": "text", + "content": " of the local window centered at " + }, + { + "bbox": [ + 304, + 574, + 545, + 622 + ], + "type": "inline_equation", + "content": "b_{k}" + }, + { + "bbox": [ + 304, + 574, + 545, + 622 + ], + "type": "text", + "content": " to the global timestamp index. Finally, the class-agnostic scores " + }, + { + "bbox": [ + 304, + 574, + 545, + 622 + ], + "type": "inline_equation", + "content": "\\mathcal{V}^{\\mathrm{b}}" + }, + { + "bbox": [ + 304, + 574, + 545, + 622 + ], + "type": "text", + "content": " are added back to " + }, + { + "bbox": [ + 304, + 574, + 545, + 622 + ], + "type": "inline_equation", + "content": "\\pmb{V}^{\\mathrm{a}}" + }, + { + "bbox": [ + 304, + 574, + 545, + 622 + ], + "type": "text", + "content": " to produce the final score matrix " + }, + { + "bbox": [ + 304, + 574, + 545, + 622 + ], + "type": "inline_equation", + "content": "\\pmb {V}\\in \\mathbb{R}^{K\\times (M - 1)}" + }, + { + "bbox": [ + 304, + 574, + 545, + 622 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 383, + 631, + 545, + 645 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 383, + 631, + 545, + 645 + ], + "spans": [ + { + "bbox": [ + 383, + 631, + 545, + 645 + ], + "type": "interline_equation", + "content": "\\boldsymbol {V} _ {k, r} = \\boldsymbol {V} _ {k, r} ^ {\\mathrm {a}} + v _ {b _ {k}} ^ {\\mathrm {b}}. \\tag {4}", + "image_path": "86758fbea5dce3794667e948ea530e2f32ff873e0bfcdf187318cb4a77c85f11.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 654, + 545, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 545, + 689 + ], + "type": "text", + "content": "The above equation means that the class-agnostic boundary score of the " + }, + { + "bbox": [ + 304, + 654, + 545, + 689 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 654, + 545, + 689 + ], + "type": "text", + "content": "-th boundary " + }, + { + "bbox": [ + 304, + 654, + 545, + 689 + ], + "type": "inline_equation", + "content": "v_{b_k}^{\\mathrm{b}}" + }, + { + "bbox": [ + 304, + 654, + 545, + 689 + ], + "type": "text", + "content": " is added to all transition scores corresponding to the " + }, + { + "bbox": [ + 304, + 654, + 545, + 689 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 654, + 545, + 689 + ], + "type": "text", + "content": "-th boundary " + }, + { + "bbox": [ + 304, + 654, + 545, + 689 + ], + "type": "inline_equation", + "content": "\\mathbf{V}_{k,r}^{\\mathrm{a}}, \\forall r" + }, + { + "bbox": [ + 304, + 654, + 545, + 689 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 689, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 689, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 689, + 545, + 712 + ], + "type": "text", + "content": "- Action Transition Alignment. The final step is to find exact " + }, + { + "bbox": [ + 304, + 689, + 545, + 712 + ], + "type": "inline_equation", + "content": "M - 1" + }, + { + "bbox": [ + 304, + 689, + 545, + 712 + ], + "type": "text", + "content": " optimal boundaries from the candidates " + }, + { + "bbox": [ + 304, + 689, + 545, + 712 + ], + "type": "inline_equation", + "content": "\\widetilde{B}" + }, + { + "bbox": [ + 304, + 689, + 545, + 712 + ], + "type": "text", + "content": " based" + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18256" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 47, + 70, + 287, + 132 + ], + "blocks": [ + { + "bbox": [ + 47, + 70, + 287, + 132 + ], + "lines": [ + { + "bbox": [ + 47, + 70, + 287, + 132 + ], + "spans": [ + { + "bbox": [ + 47, + 70, + 287, + 132 + ], + "type": "image", + "image_path": "fb7848cf7d882894564cc54bbca57f5d3b3fa5e915f0363ff36afdc90de45a4f.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 141, + 288, + 197 + ], + "lines": [ + { + "bbox": [ + 46, + 141, + 288, + 197 + ], + "spans": [ + { + "bbox": [ + 46, + 141, + 288, + 197 + ], + "type": "text", + "content": "Figure 6. An example of action transition alignment between " + }, + { + "bbox": [ + 46, + 141, + 288, + 197 + ], + "type": "inline_equation", + "content": "\\widetilde{\\mathcal{B}}" + }, + { + "bbox": [ + 46, + 141, + 288, + 197 + ], + "type": "text", + "content": " of length 5 and " + }, + { + "bbox": [ + 46, + 141, + 288, + 197 + ], + "type": "inline_equation", + "content": "\\mathcal{R}" + }, + { + "bbox": [ + 46, + 141, + 288, + 197 + ], + "type": "text", + "content": " of length 3. (a) The cost matrix " + }, + { + "bbox": [ + 46, + 141, + 288, + 197 + ], + "type": "inline_equation", + "content": "\\Delta" + }, + { + "bbox": [ + 46, + 141, + 288, + 197 + ], + "type": "text", + "content": ". (b) The cumulative cost matrix " + }, + { + "bbox": [ + 46, + 141, + 288, + 197 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 46, + 141, + 288, + 197 + ], + "type": "text", + "content": ". Blue areas are directly initialized. Yellow areas correspond to optimal alignment. Red arrows are allowed path directions. (c) The optimal alignment. Best viewed in color." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 215, + 288, + 299 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 215, + 288, + 299 + ], + "spans": [ + { + "bbox": [ + 46, + 215, + 288, + 299 + ], + "type": "text", + "content": "on the score matrix " + }, + { + "bbox": [ + 46, + 215, + 288, + 299 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 46, + 215, + 288, + 299 + ], + "type": "text", + "content": ". It is equivalent to seek an one-to-one alignment with lowest cost between the transitions " + }, + { + "bbox": [ + 46, + 215, + 288, + 299 + ], + "type": "inline_equation", + "content": "\\mathcal{R}" + }, + { + "bbox": [ + 46, + 215, + 288, + 299 + ], + "type": "text", + "content": " and candidates " + }, + { + "bbox": [ + 46, + 215, + 288, + 299 + ], + "type": "inline_equation", + "content": "\\widetilde{\\mathcal{B}}" + }, + { + "bbox": [ + 46, + 215, + 288, + 299 + ], + "type": "text", + "content": " while requiring " + }, + { + "bbox": [ + 46, + 215, + 288, + 299 + ], + "type": "inline_equation", + "content": "K - M + 1" + }, + { + "bbox": [ + 46, + 215, + 288, + 299 + ], + "type": "text", + "content": " candidates to be dropped (Fig. 5). Formally, if we denote the optimal aligned boundary set as " + }, + { + "bbox": [ + 46, + 215, + 288, + 299 + ], + "type": "inline_equation", + "content": "\\mathcal{B} = \\{b_{k_r}\\}_{r=1}^{M-1}" + }, + { + "bbox": [ + 46, + 215, + 288, + 299 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 215, + 288, + 299 + ], + "type": "inline_equation", + "content": "b_{k_r}" + }, + { + "bbox": [ + 46, + 215, + 288, + 299 + ], + "type": "text", + "content": " corresponds to the " + }, + { + "bbox": [ + 46, + 215, + 288, + 299 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 46, + 215, + 288, + 299 + ], + "type": "text", + "content": "-th transition and " + }, + { + "bbox": [ + 46, + 215, + 288, + 299 + ], + "type": "inline_equation", + "content": "1 \\leq k_1 < \\ldots < k_{M-1} \\leq K" + }, + { + "bbox": [ + 46, + 215, + 288, + 299 + ], + "type": "text", + "content": ". It should satisfy:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 71, + 304, + 287, + 337 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 71, + 304, + 287, + 337 + ], + "spans": [ + { + "bbox": [ + 71, + 304, + 287, + 337 + ], + "type": "interline_equation", + "content": "\\mathcal {B} = \\arg \\min _ {\\mathcal {B} ^ {\\prime}} \\psi \\left(\\mathcal {B} ^ {\\prime}\\right), \\quad \\psi \\left(\\mathcal {B} ^ {\\prime}\\right) = - \\sum_ {r = 1} ^ {M - 1} \\boldsymbol {V} _ {k _ {r} ^ {\\prime}, r}, \\tag {5}", + "image_path": "83f3ae3c9c2fcf3c1c79e7e21d19c26bb185d69ab3cad1eef23b6e9e070cc346.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 342, + 287, + 366 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 342, + 287, + 366 + ], + "spans": [ + { + "bbox": [ + 46, + 342, + 287, + 366 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 342, + 287, + 366 + ], + "type": "inline_equation", + "content": "\\psi (\\mathcal{B}^{\\prime})" + }, + { + "bbox": [ + 46, + 342, + 287, + 366 + ], + "type": "text", + "content": " is the cost function of an alignment and " + }, + { + "bbox": [ + 46, + 342, + 287, + 366 + ], + "type": "inline_equation", + "content": "\\mathcal{B}'" + }, + { + "bbox": [ + 46, + 342, + 287, + 366 + ], + "type": "text", + "content": " is any feasible aligned boundary set such as " + }, + { + "bbox": [ + 46, + 342, + 287, + 366 + ], + "type": "inline_equation", + "content": "\\mathcal{B}" + }, + { + "bbox": [ + 46, + 342, + 287, + 366 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 367, + 287, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 367, + 287, + 510 + ], + "spans": [ + { + "bbox": [ + 46, + 367, + 287, + 510 + ], + "type": "text", + "content": "To solve Eq. (5), we propose a DP algorithm, of which an example is shown in Fig. 6. To allow some candidates in " + }, + { + "bbox": [ + 46, + 367, + 287, + 510 + ], + "type": "inline_equation", + "content": "\\widetilde{B}" + }, + { + "bbox": [ + 46, + 367, + 287, + 510 + ], + "type": "text", + "content": " to be dropped, inspired by [36], we first expand the transition sequence " + }, + { + "bbox": [ + 46, + 367, + 287, + 510 + ], + "type": "inline_equation", + "content": "\\mathcal{R}" + }, + { + "bbox": [ + 46, + 367, + 287, + 510 + ], + "type": "text", + "content": " by inserting the empty symbol " + }, + { + "bbox": [ + 46, + 367, + 287, + 510 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 46, + 367, + 287, + 510 + ], + "type": "text", + "content": " interleaved with transition symbols, obtaining " + }, + { + "bbox": [ + 46, + 367, + 287, + 510 + ], + "type": "inline_equation", + "content": "\\mathcal{R}' = [\\phi, (a_1, a_2), \\phi, (a_2, a_3), \\phi, \\dots, \\phi, (a_{M-1}, a_M), \\phi]" + }, + { + "bbox": [ + 46, + 367, + 287, + 510 + ], + "type": "text", + "content": ". Now the candidate boundary matched to " + }, + { + "bbox": [ + 46, + 367, + 287, + 510 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 46, + 367, + 287, + 510 + ], + "type": "text", + "content": " will be discarded. Then, a cost matrix " + }, + { + "bbox": [ + 46, + 367, + 287, + 510 + ], + "type": "inline_equation", + "content": "\\Delta \\in \\mathbb{R}^{K \\times (2(M-1) + 1)}" + }, + { + "bbox": [ + 46, + 367, + 287, + 510 + ], + "type": "text", + "content": " is constructed (Fig. 6(a)), whose " + }, + { + "bbox": [ + 46, + 367, + 287, + 510 + ], + "type": "inline_equation", + "content": "(k, r')" + }, + { + "bbox": [ + 46, + 367, + 287, + 510 + ], + "type": "text", + "content": "-th entry is the cost of aligning the " + }, + { + "bbox": [ + 46, + 367, + 287, + 510 + ], + "type": "inline_equation", + "content": "b_k" + }, + { + "bbox": [ + 46, + 367, + 287, + 510 + ], + "type": "text", + "content": " with the " + }, + { + "bbox": [ + 46, + 367, + 287, + 510 + ], + "type": "inline_equation", + "content": "r'" + }, + { + "bbox": [ + 46, + 367, + 287, + 510 + ], + "type": "text", + "content": "-th symbol in " + }, + { + "bbox": [ + 46, + 367, + 287, + 510 + ], + "type": "inline_equation", + "content": "\\mathcal{R}'" + }, + { + "bbox": [ + 46, + 367, + 287, + 510 + ], + "type": "text", + "content": ". For even number of " + }, + { + "bbox": [ + 46, + 367, + 287, + 510 + ], + "type": "inline_equation", + "content": "r'" + }, + { + "bbox": [ + 46, + 367, + 287, + 510 + ], + "type": "text", + "content": " (transition symbols), it is clear that " + }, + { + "bbox": [ + 46, + 367, + 287, + 510 + ], + "type": "inline_equation", + "content": "\\Delta_{k,r'} = -V_{k,r'/2}" + }, + { + "bbox": [ + 46, + 367, + 287, + 510 + ], + "type": "text", + "content": " from Eq. (5). Odd " + }, + { + "bbox": [ + 46, + 367, + 287, + 510 + ], + "type": "inline_equation", + "content": "r'" + }, + { + "bbox": [ + 46, + 367, + 287, + 510 + ], + "type": "text", + "content": " means dropping the candidate, and the cost is set to 0." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 511, + 288, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 511, + 288, + 665 + ], + "spans": [ + { + "bbox": [ + 46, + 511, + 288, + 665 + ], + "type": "text", + "content": "A valid alignment is represented by a path from the top-left to bottom-right of " + }, + { + "bbox": [ + 46, + 511, + 288, + 665 + ], + "type": "inline_equation", + "content": "\\Delta" + }, + { + "bbox": [ + 46, + 511, + 288, + 665 + ], + "type": "text", + "content": " with constraint on directions. Specifically, the allowed start positions are " + }, + { + "bbox": [ + 46, + 511, + 288, + 665 + ], + "type": "inline_equation", + "content": "(1,1)" + }, + { + "bbox": [ + 46, + 511, + 288, + 665 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 511, + 288, + 665 + ], + "type": "inline_equation", + "content": "(1,2)" + }, + { + "bbox": [ + 46, + 511, + 288, + 665 + ], + "type": "text", + "content": ", respectively corresponding to two cases where " + }, + { + "bbox": [ + 46, + 511, + 288, + 665 + ], + "type": "inline_equation", + "content": "b_{1}" + }, + { + "bbox": [ + 46, + 511, + 288, + 665 + ], + "type": "text", + "content": " is dropped and " + }, + { + "bbox": [ + 46, + 511, + 288, + 665 + ], + "type": "inline_equation", + "content": "b_{1}" + }, + { + "bbox": [ + 46, + 511, + 288, + 665 + ], + "type": "text", + "content": " is matched with the first transition. Similarly, there are two allowed end positions: " + }, + { + "bbox": [ + 46, + 511, + 288, + 665 + ], + "type": "inline_equation", + "content": "(K,2(M - 1))" + }, + { + "bbox": [ + 46, + 511, + 288, + 665 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 511, + 288, + 665 + ], + "type": "inline_equation", + "content": "(K,2(M - 1) + 1)" + }, + { + "bbox": [ + 46, + 511, + 288, + 665 + ], + "type": "text", + "content": ". Then for the position " + }, + { + "bbox": [ + 46, + 511, + 288, + 665 + ], + "type": "inline_equation", + "content": "(k,r^{\\prime})" + }, + { + "bbox": [ + 46, + 511, + 288, + 665 + ], + "type": "text", + "content": " on the path, if " + }, + { + "bbox": [ + 46, + 511, + 288, + 665 + ], + "type": "inline_equation", + "content": "r^{\\prime}" + }, + { + "bbox": [ + 46, + 511, + 288, + 665 + ], + "type": "text", + "content": " is odd, its previous position is either " + }, + { + "bbox": [ + 46, + 511, + 288, + 665 + ], + "type": "inline_equation", + "content": "(k - 1,r^{\\prime})" + }, + { + "bbox": [ + 46, + 511, + 288, + 665 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 46, + 511, + 288, + 665 + ], + "type": "inline_equation", + "content": "(k - 1,r^{\\prime} - 1)" + }, + { + "bbox": [ + 46, + 511, + 288, + 665 + ], + "type": "text", + "content": ". It means that if " + }, + { + "bbox": [ + 46, + 511, + 288, + 665 + ], + "type": "inline_equation", + "content": "b_{k}" + }, + { + "bbox": [ + 46, + 511, + 288, + 665 + ], + "type": "text", + "content": " is dropped, the " + }, + { + "bbox": [ + 46, + 511, + 288, + 665 + ], + "type": "inline_equation", + "content": "b_{k - 1}" + }, + { + "bbox": [ + 46, + 511, + 288, + 665 + ], + "type": "text", + "content": " is also dropped or matched with the previous transition. On the other hand, the previous position is either " + }, + { + "bbox": [ + 46, + 511, + 288, + 665 + ], + "type": "inline_equation", + "content": "(k - 1,r^{\\prime} - 1)" + }, + { + "bbox": [ + 46, + 511, + 288, + 665 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 46, + 511, + 288, + 665 + ], + "type": "inline_equation", + "content": "(k - 1,r^{\\prime} - 2)" + }, + { + "bbox": [ + 46, + 511, + 288, + 665 + ], + "type": "text", + "content": " if " + }, + { + "bbox": [ + 46, + 511, + 288, + 665 + ], + "type": "inline_equation", + "content": "r^{\\prime}" + }, + { + "bbox": [ + 46, + 511, + 288, + 665 + ], + "type": "text", + "content": " is even, with the similar meaning. These allowed directions are indicated by red arrows in Fig. 6(b)." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 665, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 288, + 714 + ], + "type": "text", + "content": "The DP algorithm generally finds the optimal solution by computing a cumulative cost matrix " + }, + { + "bbox": [ + 46, + 665, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\pmb{D}" + }, + { + "bbox": [ + 46, + 665, + 288, + 714 + ], + "type": "text", + "content": " with the same shape as " + }, + { + "bbox": [ + 46, + 665, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\pmb{\\Delta}" + }, + { + "bbox": [ + 46, + 665, + 288, + 714 + ], + "type": "text", + "content": " (Fig. 6(b)), where " + }, + { + "bbox": [ + 46, + 665, + 288, + 714 + ], + "type": "inline_equation", + "content": "D_{k,r'}" + }, + { + "bbox": [ + 46, + 665, + 288, + 714 + ], + "type": "text", + "content": " represents the minimum cumulative cost of all valid paths ending at " + }, + { + "bbox": [ + 46, + 665, + 288, + 714 + ], + "type": "inline_equation", + "content": "(k, r')" + }, + { + "bbox": [ + 46, + 665, + 288, + 714 + ], + "type": "text", + "content": ". It" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "type": "text", + "content": "is computed by the following recursive equation, considering paths coming from valid directions aforementioned:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 105, + 545, + 143 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 105, + 545, + 143 + ], + "spans": [ + { + "bbox": [ + 306, + 105, + 545, + 143 + ], + "type": "interline_equation", + "content": "\\boldsymbol {D} _ {k, r ^ {\\prime}} = \\boldsymbol {\\Delta} _ {k, r ^ {\\prime}} + \\left\\{ \\begin{array}{l} \\min \\left(\\boldsymbol {D} _ {k - 1, r ^ {\\prime}}, \\boldsymbol {D} _ {k - 1, r ^ {\\prime} - 1}\\right), r ^ {\\prime} \\text {i s o d d}, \\\\ \\min \\left(\\boldsymbol {D} _ {k - 1, r ^ {\\prime} - 1}, \\boldsymbol {D} _ {k - 1, r ^ {\\prime} - 2}\\right), r ^ {\\prime} \\text {i s e v e n}. \\end{array} \\right. \\tag {6}", + "image_path": "16b011e2a5299811978ccba494d02e2a65124299c69e965da343e3c0f1a62cf9.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 144, + 545, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 144, + 545, + 191 + ], + "spans": [ + { + "bbox": [ + 304, + 144, + 545, + 191 + ], + "type": "text", + "content": "The above equation is not applicable to the first row and the first two columns of " + }, + { + "bbox": [ + 304, + 144, + 545, + 191 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 304, + 144, + 545, + 191 + ], + "type": "text", + "content": ". They are directly initialized before the computation (blue areas of Fig. 6(b)). We set the cost of invalid positions to " + }, + { + "bbox": [ + 304, + 144, + 545, + 191 + ], + "type": "inline_equation", + "content": "\\infty" + }, + { + "bbox": [ + 304, + 144, + 545, + 191 + ], + "type": "text", + "content": " (inf) and others are copied from " + }, + { + "bbox": [ + 304, + 144, + 545, + 191 + ], + "type": "inline_equation", + "content": "\\Delta" + }, + { + "bbox": [ + 304, + 144, + 545, + 191 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 192, + 545, + 262 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 192, + 545, + 262 + ], + "spans": [ + { + "bbox": [ + 304, + 192, + 545, + 262 + ], + "type": "text", + "content": "The cumulative cost of the optimal alignment is " + }, + { + "bbox": [ + 304, + 192, + 545, + 262 + ], + "type": "inline_equation", + "content": "\\min (D_{K,2(M - 1)},D_{K,2(M - 1) + 1})" + }, + { + "bbox": [ + 304, + 192, + 545, + 262 + ], + "type": "text", + "content": ", The complete alignment, i.e., the optimal boundary set " + }, + { + "bbox": [ + 304, + 192, + 545, + 262 + ], + "type": "inline_equation", + "content": "\\mathcal{B}" + }, + { + "bbox": [ + 304, + 192, + 545, + 262 + ], + "type": "text", + "content": " can be obtained by starting from the corresponding position of the above min operation and backtracking (Fig. 6(b/c)). More details can be found in the supplementary material." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 263, + 545, + 371 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 263, + 545, + 371 + ], + "spans": [ + { + "bbox": [ + 304, + 263, + 545, + 371 + ], + "type": "text", + "content": "- Complexity. Clearly, the class-agnostic boundary scoring and action transition scoring can be implemented by the unfold operator and matrix operations, thus can be parallelized on GPUs. The action transition alignment must be performed serially with a time complexity of " + }, + { + "bbox": [ + 304, + 263, + 545, + 371 + ], + "type": "inline_equation", + "content": "O(KM)" + }, + { + "bbox": [ + 304, + 263, + 545, + 371 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 304, + 263, + 545, + 371 + ], + "type": "inline_equation", + "content": "O(\\lambda M^2)" + }, + { + "bbox": [ + 304, + 263, + 545, + 371 + ], + "type": "text", + "content": ". Compared with the previous alignment methods such as Viterbi (" + }, + { + "bbox": [ + 304, + 263, + 545, + 371 + ], + "type": "inline_equation", + "content": "O(T^2M)" + }, + { + "bbox": [ + 304, + 263, + 545, + 371 + ], + "type": "text", + "content": " [25, 28, 34]) and DTW (" + }, + { + "bbox": [ + 304, + 263, + 545, + 371 + ], + "type": "inline_equation", + "content": "O(TM)" + }, + { + "bbox": [ + 304, + 263, + 545, + 371 + ], + "type": "text", + "content": "), it decreases from a function involving video length " + }, + { + "bbox": [ + 304, + 263, + 545, + 371 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 304, + 263, + 545, + 371 + ], + "type": "text", + "content": " to one of only transcript length " + }, + { + "bbox": [ + 304, + 263, + 545, + 371 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 304, + 263, + 545, + 371 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 304, + 263, + 545, + 371 + ], + "type": "inline_equation", + "content": "M << T" + }, + { + "bbox": [ + 304, + 263, + 545, + 371 + ], + "type": "text", + "content": ")." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 378, + 418, + 389 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 378, + 418, + 389 + ], + "spans": [ + { + "bbox": [ + 306, + 378, + 418, + 389 + ], + "type": "text", + "content": "3.5. Video-Level Losses" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 396, + 545, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 396, + 545, + 590 + ], + "spans": [ + { + "bbox": [ + 304, + 396, + 545, + 590 + ], + "type": "text", + "content": "Inevitably, the inferred pseudo segmentation contains some degree of noise, which can be unexpectedly fit by the network. To improve the semantic robustness, we propose to jointly train a video-level multi-label classification task that predicts whether each action appears in the video, which is precisely indicated by the transcript. Specifically, inspired by the common practice of Transformer [7, 10, 19, 24, 43], the input sequence " + }, + { + "bbox": [ + 304, + 396, + 545, + 590 + ], + "type": "inline_equation", + "content": "\\mathbf{X}" + }, + { + "bbox": [ + 304, + 396, + 545, + 590 + ], + "type": "text", + "content": " is augmented by a set of " + }, + { + "bbox": [ + 304, + 396, + 545, + 590 + ], + "type": "inline_equation", + "content": "|\\mathcal{C}|" + }, + { + "bbox": [ + 304, + 396, + 545, + 590 + ], + "type": "text", + "content": " learnable class tokens which are fed into the same network and responsible for predicting action occurrence. The outputs, denoted as " + }, + { + "bbox": [ + 304, + 396, + 545, + 590 + ], + "type": "inline_equation", + "content": "\\{e_c'\\}_{c=1}^{|\\mathcal{C}|}" + }, + { + "bbox": [ + 304, + 396, + 545, + 590 + ], + "type": "text", + "content": ", are then used to predict the action occurrence probability via " + }, + { + "bbox": [ + 304, + 396, + 545, + 590 + ], + "type": "inline_equation", + "content": "\\xi_c = \\sigma(\\boldsymbol{w}_c^T \\boldsymbol{e}_c' + \\epsilon_c)" + }, + { + "bbox": [ + 304, + 396, + 545, + 590 + ], + "type": "text", + "content": " for each category " + }, + { + "bbox": [ + 304, + 396, + 545, + 590 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 304, + 396, + 545, + 590 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 396, + 545, + 590 + ], + "type": "inline_equation", + "content": "\\boldsymbol{w}_c" + }, + { + "bbox": [ + 304, + 396, + 545, + 590 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 396, + 545, + 590 + ], + "type": "inline_equation", + "content": "\\epsilon_c" + }, + { + "bbox": [ + 304, + 396, + 545, + 590 + ], + "type": "text", + "content": " are the weight vector and bias scalar for class " + }, + { + "bbox": [ + 304, + 396, + 545, + 590 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 304, + 396, + 545, + 590 + ], + "type": "text", + "content": " in the classifier shared with frame-wise classification. " + }, + { + "bbox": [ + 304, + 396, + 545, + 590 + ], + "type": "inline_equation", + "content": "\\sigma(\\cdot)" + }, + { + "bbox": [ + 304, + 396, + 545, + 590 + ], + "type": "text", + "content": " is the sigmoid activation. These predictions are supervised by the binary cross entropy:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 310, + 598, + 545, + 632 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 310, + 598, + 545, + 632 + ], + "spans": [ + { + "bbox": [ + 310, + 598, + 545, + 632 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {v i d}} = - \\frac {1}{| \\mathcal {C} |} \\sum_ {c = 1} ^ {| \\mathcal {C} |} \\left[ y _ {c} ^ {\\mathrm {v i d}} \\log \\xi_ {c} + \\left(1 - y _ {c} ^ {\\mathrm {v i d}}\\right) \\log \\left(1 - \\xi_ {c}\\right) \\right], \\tag {7}", + "image_path": "e4feeb8a34c68977dff73137f763a1e80d359152330682fd5adbb9e6efcb4ce9.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 641, + 545, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 641, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 304, + 641, + 545, + 689 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 641, + 545, + 689 + ], + "type": "inline_equation", + "content": "y_c^{\\mathrm{vid}} \\in \\{0,1\\}" + }, + { + "bbox": [ + 304, + 641, + 545, + 689 + ], + "type": "text", + "content": " is the binary label indicating whether action " + }, + { + "bbox": [ + 304, + 641, + 545, + 689 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 304, + 641, + 545, + 689 + ], + "type": "text", + "content": " appears in the video. Through the shared network and the interactions between tokens, the semantics learned from video-level training can benefit the main task." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "text", + "content": "Moreover, we treat the class tokens with global information as the prototypes of each action and attempt to align the" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18257" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "text", + "content": "frame features to them for semantic consistency. Specifically, we first calculate the feature centroid of each category " + }, + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "text", + "content": " that appears in the video by averaging the output frame features with the corresponding pseudo labels:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 87, + 129, + 287, + 160 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 87, + 129, + 287, + 160 + ], + "spans": [ + { + "bbox": [ + 87, + 129, + 287, + 160 + ], + "type": "interline_equation", + "content": "\\overline {{\\boldsymbol {x}}} _ {c} = \\frac {\\sum_ {t = 1} ^ {T} \\mathbb {I} \\left(\\widetilde {y} _ {t} = c\\right) \\boldsymbol {x} _ {t} ^ {\\prime}}{\\sum_ {t = 1} ^ {T} \\mathbb {I} \\left(\\widetilde {y} _ {t} = c\\right)}, c \\in \\operatorname {S e t} (\\mathcal {A}), \\tag {8}", + "image_path": "cd192bde32dad9501496c61b87917291001c04fd45540e97355ee09e6c4339a1.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 168, + 287, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 168, + 287, + 217 + ], + "spans": [ + { + "bbox": [ + 47, + 168, + 287, + 217 + ], + "type": "text", + "content": "where the " + }, + { + "bbox": [ + 47, + 168, + 287, + 217 + ], + "type": "inline_equation", + "content": "\\operatorname{Set}(\\cdot)" + }, + { + "bbox": [ + 47, + 168, + 287, + 217 + ], + "type": "text", + "content": " operator converts a list into an unordered set and removes duplicates. These centroids will move towards the corresponding " + }, + { + "bbox": [ + 47, + 168, + 287, + 217 + ], + "type": "inline_equation", + "content": "e_c'" + }, + { + "bbox": [ + 47, + 168, + 287, + 217 + ], + "type": "text", + "content": " under the guidance of a global-local contrastive loss in the form of InfoNCE [17, 31]:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 224, + 287, + 266 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 224, + 287, + 266 + ], + "spans": [ + { + "bbox": [ + 48, + 224, + 287, + 266 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {g l c}} = - \\frac {1}{| \\operatorname {S e t} (\\mathcal {A}) |} \\sum_ {c \\in \\operatorname {S e t} (\\mathcal {A})} \\log \\frac {\\exp \\left(\\left\\langle \\overline {{\\boldsymbol {x}}} _ {c} \\right\\rangle^ {T} \\left\\langle \\boldsymbol {e} _ {c} ^ {\\prime} \\right\\rangle / \\tau\\right)}{\\sum_ {c ^ {\\prime} = 1} ^ {| \\mathcal {C} |} \\exp \\left(\\left\\langle \\overline {{\\boldsymbol {x}}} _ {c} \\right\\rangle^ {T} \\left\\langle \\boldsymbol {e} _ {c ^ {\\prime}} ^ {\\prime} \\right\\rangle / \\tau\\right)}, \\tag {9}", + "image_path": "012e7ae791576386290e8d2fa6c89b3279355e0459e0ea8eec514f3e911a8b56.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 266, + 287, + 291 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 266, + 287, + 291 + ], + "spans": [ + { + "bbox": [ + 47, + 266, + 287, + 291 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 266, + 287, + 291 + ], + "type": "inline_equation", + "content": "\\langle \\pmb{x}\\rangle = \\pmb{x} / \\| \\pmb{x}\\| _2" + }, + { + "bbox": [ + 47, + 266, + 287, + 291 + ], + "type": "text", + "content": " is the " + }, + { + "bbox": [ + 47, + 266, + 287, + 291 + ], + "type": "inline_equation", + "content": "l_{2}" + }, + { + "bbox": [ + 47, + 266, + 287, + 291 + ], + "type": "text", + "content": "-normalization operator, and " + }, + { + "bbox": [ + 47, + 266, + 287, + 291 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 47, + 266, + 287, + 291 + ], + "type": "text", + "content": " is the temperature hyper-parameter." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 299, + 178, + 312 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 299, + 178, + 312 + ], + "spans": [ + { + "bbox": [ + 47, + 299, + 178, + 312 + ], + "type": "text", + "content": "3.6. Training and Inference" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 317, + 287, + 424 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 317, + 287, + 424 + ], + "spans": [ + { + "bbox": [ + 46, + 317, + 287, + 424 + ], + "type": "text", + "content": "- Training. Since the pseudo labeling always requires a relatively good initialization, we utilize a two-stage training strategy. The first stage only uses the reliable labels, i.e., the video-level labels, to pretrain the network. Hence the loss function is " + }, + { + "bbox": [ + 46, + 317, + 287, + 424 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{I}} = \\mathcal{L}_{\\mathrm{vid}}" + }, + { + "bbox": [ + 46, + 317, + 287, + 424 + ], + "type": "text", + "content": ". The loss for the second stage further includes the frame-wise losses on both classification and representation using pseudo labels: " + }, + { + "bbox": [ + 46, + 317, + 287, + 424 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{II}} = \\alpha \\mathcal{L}_{\\mathrm{vid}} + \\beta \\mathcal{L}_{\\mathrm{cls}} + \\gamma \\mathcal{L}_{\\mathrm{glc}}" + }, + { + "bbox": [ + 46, + 317, + 287, + 424 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 317, + 287, + 424 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 46, + 317, + 287, + 424 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 317, + 287, + 424 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 46, + 317, + 287, + 424 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 317, + 287, + 424 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 46, + 317, + 287, + 424 + ], + "type": "text", + "content": " are all hyperparameters." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 425, + 287, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 425, + 287, + 544 + ], + "spans": [ + { + "bbox": [ + 46, + 425, + 287, + 544 + ], + "type": "text", + "content": "- Inference. For inference, we obtain the action label " + }, + { + "bbox": [ + 46, + 425, + 287, + 544 + ], + "type": "inline_equation", + "content": "\\hat{y}_t" + }, + { + "bbox": [ + 46, + 425, + 287, + 544 + ], + "type": "text", + "content": " for frame " + }, + { + "bbox": [ + 46, + 425, + 287, + 544 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 425, + 287, + 544 + ], + "type": "text", + "content": " directly from the frame-wise class probabilities: " + }, + { + "bbox": [ + 46, + 425, + 287, + 544 + ], + "type": "inline_equation", + "content": "\\hat{y}_t = \\arg \\max_c P_{t,c}" + }, + { + "bbox": [ + 46, + 425, + 287, + 544 + ], + "type": "text", + "content": ". Note that we do not require any alignment processing during inference, which is performed by previous WSAS methods for smoother predictions. However, since the ground-truth transcript is not available during inference, these methods either run alignment with every transcript from the training set [4, 5, 25, 28, 34], or predict the transcript via another model [38], which makes the pipeline more time-consuming and less practical." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 555, + 127, + 568 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 555, + 127, + 568 + ], + "spans": [ + { + "bbox": [ + 47, + 555, + 127, + 568 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 575, + 163, + 588 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 575, + 163, + 588 + ], + "spans": [ + { + "bbox": [ + 47, + 575, + 163, + 588 + ], + "type": "text", + "content": "4.1. Experimental Setup" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 594, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 287, + 713 + ], + "type": "text", + "content": "- Datasets. We perform experiments on three datasets. The Breakfast [21] dataset contains 1712 videos of breakfast cooking with 48 different actions. On average, each video has 6.8 segments and " + }, + { + "bbox": [ + 46, + 594, + 287, + 713 + ], + "type": "inline_equation", + "content": "7.3\\%" + }, + { + "bbox": [ + 46, + 594, + 287, + 713 + ], + "type": "text", + "content": " frames are background. The Hollywood Extended [2] dataset contains 937 videos taken from movies with 16 categories of daily actions such as walk or sit. On average, each video has 5.9 segments and " + }, + { + "bbox": [ + 46, + 594, + 287, + 713 + ], + "type": "inline_equation", + "content": "60.9\\%" + }, + { + "bbox": [ + 46, + 594, + 287, + 713 + ], + "type": "text", + "content": " frames are background. The CrossTask [47] dataset contains videos from 18 primary tasks. Following [28], 14 cooking-related tasks are selected, which have 2552 videos" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "text", + "content": "and 80 action categories. On average, each video has 14.4 segments and " + }, + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "inline_equation", + "content": "74.8\\%" + }, + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "text", + "content": " frames are background. For Breakfast, we use the released 4 training/test splits and report the average. For Hollywood, we perform a 10-fold cross-validation. For CrossTask, we use the released training/test split. These evaluation protocols are consistent with previous methods." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 148, + 546, + 340 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 148, + 546, + 340 + ], + "spans": [ + { + "bbox": [ + 304, + 148, + 546, + 340 + ], + "type": "text", + "content": "- Metrics. To evaluate our method, we use 4 standard metrics, following [5, 9, 25, 28]. (1) The Mean-over-Frames (MoF) is the percentage of frames whose labels are correctly predicted. (2) The Mean-over-Frames without Background (MoF-Bg) is the MoF over non-background frames. It is more suitable than MoF for the datasets with high background rate such as Hollywood and CrossTask. (3) The Intersection-over-Union (IoU) is defined as " + }, + { + "bbox": [ + 304, + 148, + 546, + 340 + ], + "type": "inline_equation", + "content": "|I \\cap I^{*}| / |I \\cup I^{*}|" + }, + { + "bbox": [ + 304, + 148, + 546, + 340 + ], + "type": "text", + "content": " while (4) the Intersection-over-Detection (IoD) is " + }, + { + "bbox": [ + 304, + 148, + 546, + 340 + ], + "type": "inline_equation", + "content": "|I \\cap I^{*}| / |I|" + }, + { + "bbox": [ + 304, + 148, + 546, + 340 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 148, + 546, + 340 + ], + "type": "inline_equation", + "content": "I^{*}" + }, + { + "bbox": [ + 304, + 148, + 546, + 340 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 148, + 546, + 340 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 304, + 148, + 546, + 340 + ], + "type": "text", + "content": " are the ground-truth (GT) segment and the predicted segment with the same class, respectively. For each GT segment, the highest IoU/IoD with one predicted segment are preserved, and the average of all GT segments is reported. Note that the definition of the IoU/IoD in [28] is different from other works [5, 9, 25], and will be indicated with special symbols when used." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 345, + 545, + 513 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 345, + 545, + 513 + ], + "spans": [ + { + "bbox": [ + 304, + 345, + 545, + 513 + ], + "type": "text", + "content": "- Input Features. As the input features " + }, + { + "bbox": [ + 304, + 345, + 545, + 513 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 304, + 345, + 545, + 513 + ], + "type": "text", + "content": " for Breakfast and Hollywood, we use the 2048-dimensional RGB+flow I3D features [3] adopted by MuCon [38] and most FSAS methods [13, 26, 27, 45], while some WSAS methods [5, 25, 28, 34] still use the iDT features [41]. Since recent studies [37, 38] have found that they perform worse with more advanced I3D features, we still report the performance with the iDT features for them following [38]. For CrossTask, the officially released 3200-dimensional features are adopted, but we do not perform dimension reduction via PCA as with [28]. For GPU memory efficiency, we perform a " + }, + { + "bbox": [ + 304, + 345, + 545, + 513 + ], + "type": "inline_equation", + "content": "10 \\times" + }, + { + "bbox": [ + 304, + 345, + 545, + 513 + ], + "type": "text", + "content": " temporal downsampling for Breakfast and " + }, + { + "bbox": [ + 304, + 345, + 545, + 513 + ], + "type": "inline_equation", + "content": "5 \\times" + }, + { + "bbox": [ + 304, + 345, + 545, + 513 + ], + "type": "text", + "content": " for Hollywood. During inference, the output is upsampled to match the original video/ground-truth length." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 517, + 545, + 649 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 517, + 545, + 649 + ], + "spans": [ + { + "bbox": [ + 304, + 517, + 545, + 649 + ], + "type": "text", + "content": "- Implementation Details. We adopt a 6-layer Transformer [40] with single-head self-attention as the temporal network, and the latent dimension " + }, + { + "bbox": [ + 304, + 517, + 545, + 649 + ], + "type": "inline_equation", + "content": "d'" + }, + { + "bbox": [ + 304, + 517, + 545, + 649 + ], + "type": "text", + "content": " is 256. For the ATBA, we set " + }, + { + "bbox": [ + 304, + 517, + 545, + 649 + ], + "type": "inline_equation", + "content": "w^{\\mathrm{b}} = 7" + }, + { + "bbox": [ + 304, + 517, + 545, + 649 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 517, + 545, + 649 + ], + "type": "inline_equation", + "content": "\\lambda = 4" + }, + { + "bbox": [ + 304, + 517, + 545, + 649 + ], + "type": "text", + "content": " for all datasets. " + }, + { + "bbox": [ + 304, + 517, + 545, + 649 + ], + "type": "inline_equation", + "content": "w^{\\mathrm{a}}" + }, + { + "bbox": [ + 304, + 517, + 545, + 649 + ], + "type": "text", + "content": " is set to 31 for Breakfast/CrossTask and 23 for Hollywood. " + }, + { + "bbox": [ + 304, + 517, + 545, + 649 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 304, + 517, + 545, + 649 + ], + "type": "text", + "content": " is set to 0.3 for Breakfast/Hollywood and 0.1 for CrossTask. The loss weights " + }, + { + "bbox": [ + 304, + 517, + 545, + 649 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 304, + 517, + 545, + 649 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 517, + 545, + 649 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 304, + 517, + 545, + 649 + ], + "type": "text", + "content": " are set to 1.0 and " + }, + { + "bbox": [ + 304, + 517, + 545, + 649 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 304, + 517, + 545, + 649 + ], + "type": "text", + "content": " is 0.1. " + }, + { + "bbox": [ + 304, + 517, + 545, + 649 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 304, + 517, + 545, + 649 + ], + "type": "text", + "content": " in Eq. (9) is set to 0.2. Besides, to alleviate the issue that the predictions on Hollywood/CrossTask are overly dominated by background, we lower the sample weights of pseudo background frames to 0.8 in Eq. (1) for these datasets." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 654, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 545, + 712 + ], + "type": "text", + "content": "- Multiple Runs. Due to the alternating nature of learning from weak supervision, there are often fluctuations in the results of WSAS methods [28, 38]. Hence, following [38], we report the average and standard deviation over 5 runs with different random seeds for better evaluation." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18258" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 50, + 70, + 287, + 219 + ], + "blocks": [ + { + "bbox": [ + 50, + 70, + 287, + 219 + ], + "lines": [ + { + "bbox": [ + 50, + 70, + 287, + 219 + ], + "spans": [ + { + "bbox": [ + 50, + 70, + 287, + 219 + ], + "type": "table", + "html": "
Breakfast
MethodMoF±stdMoF-Bg±stdIoU±stdIoD±std
HMM+RNN [33]33.3---
[33]+Length [23]36.7---
TCFPN+ISBA [9]38.4/36.4±1.0*38.424.240.6
NN-Viterbi [34]43.0/39.7±2.4*---
D3TW [4]45.7---
CDFL [25]50.2/48.1±2.5*48.033.745.4
DPDTW [5]50.8-35.645.1
TASL [28]♠47.8-35.2†46.1†
MuCon [38]♠48.5±1.850.3*40.9*54.0*
POC [29]♠45.7-38.3†-
AdaAct [46]51.248.336.346.4
ATBA♠(Ours)53.9±1.254.4±1.241.1±0.761.7±1.1
39.5±0.8†55.9±1.0†
", + "image_path": "41814da1ae32ae998c1ac57b07da48c4df1e644f9fa04426456d0fd2d022bc97.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 304, + 194, + 545, + 281 + ], + "lines": [ + { + "bbox": [ + 304, + 194, + 545, + 281 + ], + "spans": [ + { + "bbox": [ + 304, + 194, + 545, + 281 + ], + "type": "text", + "content": "Table 2. Comparison of accuracy, training and inference time on the Breakfast. The training time is measured as the entire training duration on the split 1, and the inference time is measured as the average time for inferring a video from the test set of the split 1. Tr.A.-The alignment algorithm adopted during training (X-No Alignment. V-Viterbi. D-DTW. B-Boundary Alignment). *-Measured by [38]. ‡-Measured by us. Best results are in bold, second best are underlined." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 50, + 219, + 287, + 342 + ], + "blocks": [ + { + "bbox": [ + 50, + 219, + 287, + 342 + ], + "lines": [ + { + "bbox": [ + 50, + 219, + 287, + 342 + ], + "spans": [ + { + "bbox": [ + 50, + 219, + 287, + 342 + ], + "type": "table", + "html": "
Hollywood Extended
MethodMoF±stdMoF-Bg±stdIoU±stdIoD±std
HMM+RNN [33]--11.9-
[33]+Length [23]--12.3-
TCFPN+ISBA [9]28.734.512.618.3
D3TW [4]33.6---
CDFL [25]45.040.619.525.8
DPDTW [5]55.625.6#33.243.3
TASL [28]42.1#27.2#23.3†#33.0†#
MuCon [38]♠-41.613.9*-
ATBA♠(Ours)47.7±2.540.2±1.630.9±1.655.8±0.8
28.5±1.6†44.9±0.5†
", + "image_path": "a3520456853db04fe1dc8361182fe79f9aa7b018408fb900b27605c9bb4d2975.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 50, + 342, + 287, + 428 + ], + "blocks": [ + { + "bbox": [ + 50, + 342, + 287, + 428 + ], + "lines": [ + { + "bbox": [ + 50, + 342, + 287, + 428 + ], + "spans": [ + { + "bbox": [ + 50, + 342, + 287, + 428 + ], + "type": "table", + "html": "
CrossTask
MethodMoF±stdMoF-Bg±stdIoU±stdIoD±std
NN-Viterbi [34]♠26.5*-10.7†*24.0†*
CDFL [25]♠31.9*-11.5†*23.8†*
TASL [28]♠40.727.4#14.5†25.1†
POC [29]♠42.817.6#15.6†-
ATBA♠(Ours)50.6±1.331.3±0.720.9±0.444.6±0.7
15.7±0.3†24.6±0.4†
", + "image_path": "80a8a7ea6a98f581dff8a7609fc21da2f25b1d66a33d57b8ec598f3fe890459d.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 46, + 436, + 287, + 535 + ], + "lines": [ + { + "bbox": [ + 46, + 436, + 287, + 535 + ], + "spans": [ + { + "bbox": [ + 46, + 436, + 287, + 535 + ], + "type": "text", + "content": "Table 1. Comparisons of ours with other WSAS methods on three datasets. std is the standard deviation of multiple runs (if any). " + }, + { + "bbox": [ + 46, + 436, + 287, + 535 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 46, + 436, + 287, + 535 + ], + "type": "text", + "content": " The metric is computed by the definition of [28]. *-Results are reported by other works. Please refer to the supplementary material for detailed sources. " + }, + { + "bbox": [ + 46, + 436, + 287, + 535 + ], + "type": "inline_equation", + "content": "\\sharp" + }, + { + "bbox": [ + 46, + 436, + 287, + 535 + ], + "type": "text", + "content": "-Results are obtained by us via rerunning the open sources (The TASL [28] does not follow the common 10-fold evaluation protocol for Hollywood so we re-produce the results). " + }, + { + "bbox": [ + 46, + 436, + 287, + 535 + ], + "type": "inline_equation", + "content": "\\spadesuit" + }, + { + "bbox": [ + 46, + 436, + 287, + 535 + ], + "type": "text", + "content": " The reported results are the average of multiple runs. Best results are in bold, second best are underlined." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 559, + 246, + 572 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 559, + 246, + 572 + ], + "spans": [ + { + "bbox": [ + 47, + 559, + 246, + 572 + ], + "type": "text", + "content": "4.2. Comparison with the State-of-the-Art" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 579, + 287, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 579, + 287, + 664 + ], + "spans": [ + { + "bbox": [ + 46, + 579, + 287, + 664 + ], + "type": "text", + "content": "- Performance. In Table 1, we compare our proposed method with previous WSAS methods. Our ATBA achieves state-of-the-art (SOTA) by a clear margin " + }, + { + "bbox": [ + 46, + 579, + 287, + 664 + ], + "type": "inline_equation", + "content": "(+2.7\\% \\mathrm{MoF})" + }, + { + "bbox": [ + 46, + 579, + 287, + 664 + ], + "type": "text", + "content": " on the Breakfast [21], demonstrating the effectiveness of focusing on action transitions. Moreover, comparing the standard deviation with other methods, it can be found that our method is also relatively stable." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 665, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 287, + 713 + ], + "type": "text", + "content": "Our ATBA also achieves comparable performance on the Hollywood [2]. The reason why our method does not show significant advantage is probably because this dataset is collected from movies and so contains many shot changes, re" + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 321, + 71, + 531, + 186 + ], + "blocks": [ + { + "bbox": [ + 321, + 71, + 531, + 186 + ], + "lines": [ + { + "bbox": [ + 321, + 71, + 531, + 186 + ], + "spans": [ + { + "bbox": [ + 321, + 71, + 531, + 186 + ], + "type": "table", + "html": "
MethodTr.A.MoFTraining (Hours)Inference (Seconds)
TCFPN+ISBA [9]✘33.312.75*0.01*
NN-Viterbi [34]V43.011.23*56.25*
CDFL [25]V50.266.73*62.37*
DPDTW [5]D50.831.02‡0.69‡
TASL [28]V47.824.66‡54.99‡
MuCon [38]✘48.54.573.03
POC [29]✘45.72.28‡0.01‡
ATBA (Ours)B53.93.450.01
", + "image_path": "df7ef5c07ee0a06de86f38495c759067dc90227f32906b13a8fb3df009bbac74.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 315, + 545, + 398 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 315, + 545, + 398 + ], + "spans": [ + { + "bbox": [ + 304, + 315, + 545, + 398 + ], + "type": "text", + "content": "sulting in more noisy boundaries. Note that although the DPDTW [5] achieves significantly high MoF, this metric is severely biased in case of highly imbalanced categories (60.9% frames are background). We report the MoF-Bg metric for it, on which it performs poorly, proving that it cannot recognize real actions very well. In contrast, the performance of our method is more balanced." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 403, + 545, + 452 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 403, + 545, + 452 + ], + "spans": [ + { + "bbox": [ + 304, + 403, + 545, + 452 + ], + "type": "text", + "content": "For the more difficult dataset CrossTask [47] with the most segments and highest background rate, our approach also outperforms previous methods whether or not the metric involves background (" + }, + { + "bbox": [ + 304, + 403, + 545, + 452 + ], + "type": "inline_equation", + "content": "+7.8\\%" + }, + { + "bbox": [ + 304, + 403, + 545, + 452 + ], + "type": "text", + "content": " MoF and " + }, + { + "bbox": [ + 304, + 403, + 545, + 452 + ], + "type": "inline_equation", + "content": "+3.9\\%" + }, + { + "bbox": [ + 304, + 403, + 545, + 452 + ], + "type": "text", + "content": " MoF-Bg)." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 456, + 545, + 672 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 456, + 545, + 672 + ], + "spans": [ + { + "bbox": [ + 304, + 456, + 545, + 672 + ], + "type": "text", + "content": "- Efficiency. Besides of the performance, we also compare the training time of our approach with previous WSAS methods to show the efficiency of ours. Following [38], we train our model on an Nvidia GeForce GTX 1080Ti GPU, and the training time is measured as the wall time over the whole training phase, during which any irrelevant operations such as intermediate evaluation and saving checkpoints are removed. As all the WSAS methods directly load pre-computed features, the time measurement also does not include the time to extract features from raw videos. As shown in Table 2, our ATBA achieves the best performance with the second shortest training time, which demonstrates the effectiveness of our design. Specifically, the training speed of ours is on average 10 times (varying from 3 to 20) faster than methods performing frame-by-frame alignment [5, 25, 28, 34]. Compared to the alignment-free methods with comparable training speed [29, 38], ours achieves better performance." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "content": "We also compare the time to infer a single test video. As mentioned in Sec. 3.6, ours does not require any alignment processing, which makes it the fastest during inference." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18259" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 74, + 70, + 261, + 135 + ], + "blocks": [ + { + "bbox": [ + 74, + 70, + 261, + 135 + ], + "lines": [ + { + "bbox": [ + 74, + 70, + 261, + 135 + ], + "spans": [ + { + "bbox": [ + 74, + 70, + 261, + 135 + ], + "type": "table", + "html": "
Exp.\\( \\gamma^b \\)Cost Mat. \\( V^a \\)P.L.MoFIoUIoD
1✓38.632.323.152.5
2✓✓51.741.330.247.8
3✓○✓67.954.041.1
", + "image_path": "c0f7687e3d233059dd00f6231e24f504067130fe3adf66526d75cf12db4ef53f.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 243, + 147, + 255 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 243, + 147, + 255 + ], + "spans": [ + { + "bbox": [ + 47, + 243, + 147, + 255 + ], + "type": "text", + "content": "4.3. Ablation Studies" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 262, + 287, + 465 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 262, + 287, + 465 + ], + "spans": [ + { + "bbox": [ + 46, + 262, + 287, + 465 + ], + "type": "text", + "content": "- Effect of ATBA. In Table 3, we conducted an in-depth evaluation of our proposed ATBA. From Exp.1, the pseudo label quality and the evaluation performance are both very poor when only using the class-agnostic boundary scores " + }, + { + "bbox": [ + 46, + 262, + 287, + 465 + ], + "type": "inline_equation", + "content": "\\nu^{\\mathrm{b}}" + }, + { + "bbox": [ + 46, + 262, + 287, + 465 + ], + "type": "text", + "content": " to directly select " + }, + { + "bbox": [ + 46, + 262, + 287, + 465 + ], + "type": "inline_equation", + "content": "M - 1" + }, + { + "bbox": [ + 46, + 262, + 287, + 465 + ], + "type": "text", + "content": " boundaries for transitions via the greedy strategy stated in Sec. 3.4, showing that it is critical to take the class-specific transition pattern into account as with our design to suppress the noisy boundaries for more precise pseudo segmentation. Fig. 2 provides an intuitive example. In addition, comparing Exp.2 and 3, it's better to involve the class-agnostic boundary scores in the action transition alignment. We think it is because that some candidates can have very low boundary scores (unlikely to be a boundary) as the candidate selection only depends on ranking, and when the transition patterns are not discriminative at the beginning of training, these candidates may be unexpectedly aligned if the boundary scores are not involved." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 466, + 287, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 466, + 287, + 586 + ], + "spans": [ + { + "bbox": [ + 46, + 466, + 287, + 586 + ], + "type": "text", + "content": "- Effect of Video-Level Losses. In Table 4, we evaluate the " + }, + { + "bbox": [ + 46, + 466, + 287, + 586 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{vid}}" + }, + { + "bbox": [ + 46, + 466, + 287, + 586 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 466, + 287, + 586 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{glc}}" + }, + { + "bbox": [ + 46, + 466, + 287, + 586 + ], + "type": "text", + "content": ". Comparing Exp.1 and 2, the model with " + }, + { + "bbox": [ + 46, + 466, + 287, + 586 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{vid}}" + }, + { + "bbox": [ + 46, + 466, + 287, + 586 + ], + "type": "text", + "content": " achieves higher performance (" + }, + { + "bbox": [ + 46, + 466, + 287, + 586 + ], + "type": "inline_equation", + "content": "+6.5\\%" + }, + { + "bbox": [ + 46, + 466, + 287, + 586 + ], + "type": "text", + "content": " MoF) at the same level of pseudo label accuracy, demonstrating that the video-level supervision can promote the learning of precise action semantics. In Exp.3, the explicit representation alignment " + }, + { + "bbox": [ + 46, + 466, + 287, + 586 + ], + "type": "inline_equation", + "content": "(\\mathcal{L}_{\\mathrm{glc}})" + }, + { + "bbox": [ + 46, + 466, + 287, + 586 + ], + "type": "text", + "content": " further improves the quality of frame features and thus the performance (" + }, + { + "bbox": [ + 46, + 466, + 287, + 586 + ], + "type": "inline_equation", + "content": "+1.6\\%" + }, + { + "bbox": [ + 46, + 466, + 287, + 586 + ], + "type": "text", + "content": " MoF). Moreover, Exp.4 shows that whether or not to share the classifier between frame-/video-level classification has little impact." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 587, + 287, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 587, + 287, + 624 + ], + "spans": [ + { + "bbox": [ + 46, + 587, + 287, + 624 + ], + "type": "text", + "content": "We also provide the analysis on the effect of important hyper-parameters in the supplementary material, including " + }, + { + "bbox": [ + 46, + 587, + 287, + 624 + ], + "type": "inline_equation", + "content": "w^{\\mathrm{b}}" + }, + { + "bbox": [ + 46, + 587, + 287, + 624 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 587, + 287, + 624 + ], + "type": "inline_equation", + "content": "w^{\\mathrm{a}}" + }, + { + "bbox": [ + 46, + 587, + 287, + 624 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 587, + 287, + 624 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 46, + 587, + 287, + 624 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 587, + 287, + 624 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 46, + 587, + 287, + 624 + ], + "type": "text", + "content": " in the ATBA." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 634, + 159, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 634, + 159, + 647 + ], + "spans": [ + { + "bbox": [ + 47, + 634, + 159, + 647 + ], + "type": "text", + "content": "4.4. Qualitative Results" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 654, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 654, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 654, + 287, + 713 + ], + "type": "text", + "content": "We show the qualitative inference result compared with two recent open source WSAS methods, i.e., MuCon [38] and TASL [28], in Fig. 7. Our ATBA achieves significantly more accurate results in this challenging video with many segments. The result of TASL shows order reversal (\"Crake" + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 326, + 71, + 526, + 137 + ], + "blocks": [ + { + "bbox": [ + 46, + 143, + 288, + 220 + ], + "lines": [ + { + "bbox": [ + 46, + 143, + 288, + 220 + ], + "spans": [ + { + "bbox": [ + 46, + 143, + 288, + 220 + ], + "type": "text", + "content": "Table 3. Ablation studies of ATBA on the Breakfast. Exp.-Different experiment configurations. Cost Mat.-Different choices for the cost matrix in action transition alignment. o-According to Eq. (4), " + }, + { + "bbox": [ + 46, + 143, + 288, + 220 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 46, + 143, + 288, + 220 + ], + "type": "text", + "content": " contains " + }, + { + "bbox": [ + 46, + 143, + 288, + 220 + ], + "type": "inline_equation", + "content": "V^{\\mathrm{a}}" + }, + { + "bbox": [ + 46, + 143, + 288, + 220 + ], + "type": "text", + "content": ". P.L.-The accuracy of pseudo labels during training. Exp.1 means that only the class-agnostic boundary detection is adopted to localize action transitions. Exp.3 is our default setting. Best results are in bold." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 326, + 71, + 526, + 137 + ], + "lines": [ + { + "bbox": [ + 326, + 71, + 526, + 137 + ], + "spans": [ + { + "bbox": [ + 326, + 71, + 526, + 137 + ], + "type": "table", + "html": "
Exp.\\( \\mathcal{L}_{\\text{vid}} \\)\\( \\mathcal{L}_{\\text{glc}} \\)D.Cls.P.L.MoFIoUIoD
167.545.940.561.9
2✓66.652.438.861.4
3✓✓67.954.041.162.3
4✓✓✓68.153.740.161.7
", + "image_path": "8dfefe78dd654143252e7bdf1cf17fc804d871cb3ead2d6c9573e623b2aaeb7d.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 145, + 545, + 201 + ], + "lines": [ + { + "bbox": [ + 305, + 145, + 545, + 201 + ], + "spans": [ + { + "bbox": [ + 305, + 145, + 545, + 201 + ], + "type": "text", + "content": "Table 4. Ablation studies of video-level losses on the Breakfast. Exp.-Different experiment configurations. D.Cls.-Different classifiers for frame-wise prediction and action occurrence prediction. P.L.-The accuracy of pseudo labels during training. Exp.3 is our default setting. Best results are in bold." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 308, + 214, + 545, + 336 + ], + "blocks": [ + { + "bbox": [ + 308, + 214, + 545, + 336 + ], + "lines": [ + { + "bbox": [ + 308, + 214, + 545, + 336 + ], + "spans": [ + { + "bbox": [ + 308, + 214, + 545, + 336 + ], + "type": "image", + "image_path": "e1977212b6e5ed40c5dde94ef8a5582bd3e4c1ac887b4ba97c5238537188ea79.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 343, + 545, + 376 + ], + "lines": [ + { + "bbox": [ + 305, + 343, + 545, + 376 + ], + "spans": [ + { + "bbox": [ + 305, + 343, + 545, + 376 + ], + "type": "text", + "content": "Figure 7. Qualitative results on the Breakfast. The example test video is P15-stereo01-P15-pancake. We compare our inference result with two recent methods. Best viewed in color." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 400, + 545, + 473 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 400, + 545, + 473 + ], + "spans": [ + { + "bbox": [ + 305, + 400, + 545, + 473 + ], + "type": "text", + "content": "Egg\" and \"Spoon Flour\") and hallucination (\"Butter Pan\"). The MuCon predicts the correct action ordering, but the result deviates severely. In contrast, ours successfully predicts an accurate segmentation, indicating the action semantics are well learned. More qualitative results are provided in supplementary material." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 487, + 378, + 500 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 487, + 378, + 500 + ], + "spans": [ + { + "bbox": [ + 306, + 487, + 378, + 500 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 509, + 545, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 509, + 545, + 628 + ], + "spans": [ + { + "bbox": [ + 304, + 509, + 545, + 628 + ], + "type": "text", + "content": "In this work, we propose to directly localize action transitions for efficient pseudo segmentation generation in the WSAS task, thus avoiding the time-consuming frame-by-frame alignment. Due to the presence of noisy boundaries, a novel Action-Transition-Aware Boundary Alignment (ATBA) framework is proposed to efficiently and effectively filter out noise and detect transitions. Moreover, we also design some video-level losses to utilize video-level supervision to improve the semantic robustness. Extensive experiments show the effectiveness of our ATBA." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 644, + 410, + 658 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 644, + 410, + 658 + ], + "spans": [ + { + "bbox": [ + 306, + 644, + 410, + 658 + ], + "type": "text", + "content": "6. Acknowledgment" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 305, + 665, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 665, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 665, + 545, + 713 + ], + "type": "text", + "content": "This work was supported partially by the NSFC (U21A20471), National Key Research and Development Program of China (2023YFA1008503), Guangdong NSF Project (No. 2023B1515040025, 2020B1515120085)." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18260" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 53, + 91, + 286, + 124 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 286, + 124 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 286, + 124 + ], + "type": "text", + "content": "[1] Emad Bahrami, Gianpiero Francesca, and Juergen Gall. How much temporal long-term context is needed for action segmentation? arXiv preprint arXiv:2308.11358, 2023. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 125, + 286, + 180 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 125, + 286, + 180 + ], + "spans": [ + { + "bbox": [ + 53, + 125, + 286, + 180 + ], + "type": "text", + "content": "[2] Piotr Bojanowski, Rémi Lajugie, Francis Bach, Ivan Laptev, Jean Ponce, Cordelia Schmid, and Josef Sivic. Weakly supervised action labeling in videos under ordering constraints. In Proceedings of the European Conference on Computer Vision, pages 628-643. Springer, 2014. 1, 2, 6, 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 182, + 286, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 182, + 286, + 226 + ], + "spans": [ + { + "bbox": [ + 53, + 182, + 286, + 226 + ], + "type": "text", + "content": "[3] Joao Carreira and Andrew Zisserman. Quo vadis, action recognition? a new model and the kinetics dataset. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 6299-6308, 2017. 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 228, + 286, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 228, + 286, + 293 + ], + "spans": [ + { + "bbox": [ + 53, + 228, + 286, + 293 + ], + "type": "text", + "content": "[4] Chien-Yi Chang, De-An Huang, Yanan Sui, Li Fei-Fei, and Juan Carlos Niebles. D3tw: Discriminative differentiable dynamic time warping for weakly supervised action alignment and segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3546–3555, 2019. 1, 2, 6, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 295, + 286, + 349 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 295, + 286, + 349 + ], + "spans": [ + { + "bbox": [ + 53, + 295, + 286, + 349 + ], + "type": "text", + "content": "[5] Xiaobin Chang, Frederick Tung, and Greg Mori. Learning discriminative prototypes with dynamic time warping. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8395-8404, 2021. 1, 2, 6, 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 352, + 287, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 352, + 287, + 407 + ], + "spans": [ + { + "bbox": [ + 53, + 352, + 287, + 407 + ], + "type": "text", + "content": "[6] Min-Hung Chen, Baopu Li, Yingze Bao, Ghassan Al-Regib, and Zsolt Kira. Action segmentation with joint self-supervised temporal domain adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9454–9463, 2020. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 408, + 286, + 452 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 408, + 286, + 452 + ], + "spans": [ + { + "bbox": [ + 53, + 408, + 286, + 452 + ], + "type": "text", + "content": "[7] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018. 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 454, + 286, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 454, + 286, + 498 + ], + "spans": [ + { + "bbox": [ + 53, + 454, + 286, + 498 + ], + "type": "text", + "content": "[8] Guodong Ding and Angela Yao. Leveraging action affinity and continuity for semi-supervised temporal action segmentation. In European Conference on Computer Vision, pages 17-32. Springer, 2022. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 499, + 286, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 499, + 286, + 544 + ], + "spans": [ + { + "bbox": [ + 53, + 499, + 286, + 544 + ], + "type": "text", + "content": "[9] Li Ding and Chenliang Xu. Weakly-supervised action segmentation with iterative soft boundary assignment. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6508-6516, 2018. 1, 2, 3, 6, 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 545, + 286, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 545, + 286, + 610 + ], + "spans": [ + { + "bbox": [ + 48, + 545, + 286, + 610 + ], + "type": "text", + "content": "[10] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020.5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 613, + 286, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 613, + 286, + 656 + ], + "spans": [ + { + "bbox": [ + 48, + 613, + 286, + 656 + ], + "type": "text", + "content": "[11] Dazhao Du, Bing Su, Yu Li, Zhongang Qi, Lingyu Si, and Ying Shan. Efficient u-transformer with boundary-aware loss for action segmentation. arXiv preprint arXiv:2205.13425, 2022. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 658, + 286, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 286, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 286, + 712 + ], + "type": "text", + "content": "[12] Zexing Du, Xue Wang, Guoqing Zhou, and Qing Wang. Fast and unsupervised action boundary detection for action segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3323-3332, 2022. 1, 2, 4" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 713 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 117 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 117 + ], + "type": "text", + "content": "[13] Yazan Abu Farha and Jurgen Gall. Ms-tcn: Multi-stage temporal convolutional network for action segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3575-3584, 2019. 1, 6" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 118, + 547, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 547, + 173 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 547, + 173 + ], + "type": "text", + "content": "[14] Shang-Hua Gao, Qi Han, Zhong-Yu Li, Pai Peng, Liang Wang, and Ming-Ming Cheng. Global2local: Efficient structure search for video action segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16805-16814, 2021. 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 175, + 545, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 175, + 545, + 229 + ], + "spans": [ + { + "bbox": [ + 307, + 175, + 545, + 229 + ], + "type": "text", + "content": "[15] Reza Ghoddoosian, Isht Dwivedi, Nakul Agarwal, Chiho Choi, and Behzad Dariush. Weakly-supervised online action segmentation in multi-view instructional videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13780-13790, 2022. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 231, + 545, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 231, + 545, + 285 + ], + "spans": [ + { + "bbox": [ + 308, + 231, + 545, + 285 + ], + "type": "text", + "content": "[16] Reza Ghoddoosian, Isht Dwivedi, Nakul Agarwal, and Behzad Dariush. Weakly-supervised action segmentation and unseen error detection in anomalous instructional videos. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10128-10138, 2023. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 286, + 545, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 286, + 545, + 341 + ], + "spans": [ + { + "bbox": [ + 308, + 286, + 545, + 341 + ], + "type": "text", + "content": "[17] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9729-9738, 2020. 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 342, + 545, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 342, + 545, + 387 + ], + "spans": [ + { + "bbox": [ + 308, + 342, + 545, + 387 + ], + "type": "text", + "content": "[18] De-An Huang, Li Fei-Fei, and Juan Carlos Niebles. Connectionist temporal modeling for weakly supervised action labeling. In Proceedings of the European Conference on Computer Vision, pages 137–153. Springer, 2016. 1, 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 388, + 545, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 388, + 545, + 441 + ], + "spans": [ + { + "bbox": [ + 308, + 388, + 545, + 441 + ], + "type": "text", + "content": "[19] Zi-Hang Jiang, Qibin Hou, Li Yuan, Daquan Zhou, Yujun Shi, Xiaojie Jin, Anran Wang, and Jiashi Feng. All tokens matter: Token labeling for training better vision transformers. Advances in neural information processing systems, 34: 18590-18602, 2021. 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 444, + 545, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 444, + 545, + 498 + ], + "spans": [ + { + "bbox": [ + 308, + 444, + 545, + 498 + ], + "type": "text", + "content": "[20] Hyolim Kang, Jinwoo Kim, Taehyun Kim, and Seon Joo Kim. Uboco: Unsupervised boundary contrastive learning for generic event boundary detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20073-20082, 2022. 2, 4" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 500, + 545, + 554 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 500, + 545, + 554 + ], + "spans": [ + { + "bbox": [ + 308, + 500, + 545, + 554 + ], + "type": "text", + "content": "[21] Hilde Kuehne, Ali Arslan, and Thomas Serre. The language of actions: Recovering the syntax and semantics of goal-directed human activities. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 780-787, 2014. 2, 6, 7" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 555, + 545, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 555, + 545, + 589 + ], + "spans": [ + { + "bbox": [ + 308, + 555, + 545, + 589 + ], + "type": "text", + "content": "[22] Hilde Kuehne, Alexander Richard, and Juergen Gall. Weakly supervised learning of actions from transcripts. Computer Vision and Image Understanding, 163:78-89, 2017. 1, 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 590, + 545, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 590, + 545, + 634 + ], + "spans": [ + { + "bbox": [ + 308, + 590, + 545, + 634 + ], + "type": "text", + "content": "[23] Hilde Kuehne, Alexander Richard, and Juergen Gall. A hybrid rnn-hmm approach for weakly supervised temporal action segmentation. IEEE transactions on pattern analysis and machine intelligence, 42(4):765-779, 2018. 1, 2, 7" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 635, + 545, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 635, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 308, + 635, + 545, + 689 + ], + "type": "text", + "content": "[24] Jack Lanchantin, Tianlu Wang, Vicente Ordonez, and Yanjun Qi. General multi-label image classification with transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16478-16488, 2021. 5" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 691, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 691, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 691, + 545, + 713 + ], + "type": "text", + "content": "[25] Jun Li, Peng Lei, and Sinisa Todorovic. Weakly supervised energy-based learning for action segmentation. In Proceed-" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "18261" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 66, + 72, + 286, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 72, + 286, + 95 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 286, + 95 + ], + "type": "text", + "content": "ings of the IEEE/CVF International Conference on Computer Vision, pages 6243-6251, 2019. 1, 2, 5, 6, 7" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 96, + 287, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 96, + 287, + 140 + ], + "spans": [ + { + "bbox": [ + 48, + 96, + 287, + 140 + ], + "type": "text", + "content": "[26] Shi-Jie Li, Yazan AbuFarha, Yun Liu, Ming-Ming Cheng, and Juergen Gall. Ms-tcn++: Multi-stage temporal convolutional network for action segmentation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2020. 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 141, + 287, + 174 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 141, + 287, + 174 + ], + "spans": [ + { + "bbox": [ + 48, + 141, + 287, + 174 + ], + "type": "text", + "content": "[27] Daochang Liu, Qiyue Li, AnhDung Dinh, Tingting Jiang, Mubarak Shah, and Chang Xu. Diffusion action segmentation. arXiv preprint arXiv:2303.17959, 2023. 1, 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 175, + 287, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 175, + 287, + 229 + ], + "spans": [ + { + "bbox": [ + 48, + 175, + 287, + 229 + ], + "type": "text", + "content": "[28] Zijia Lu and Ehsan Elhamifar. Weakly-supervised action segmentation and alignment via transcript-aware union-of-subspaces learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8085-8095, 2021. 1, 2, 5, 6, 7, 8" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 232, + 287, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 232, + 287, + 285 + ], + "spans": [ + { + "bbox": [ + 48, + 232, + 287, + 285 + ], + "type": "text", + "content": "[29] Zijia Lu and Ehsan Elhamifar. Set-supervised action learning in procedural task videos via pairwise order consistency. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19903-19913, 2022. 1, 3, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 287, + 287, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 287, + 287, + 353 + ], + "spans": [ + { + "bbox": [ + 48, + 287, + 287, + 353 + ], + "type": "text", + "content": "[30] Antoine Miech, Dimitri Zhukov, Jean-Baptiste Alayrac, Makarand Tapaswi, Ivan Laptev, and Josef Sivic. Howto100m: Learning a text-video embedding by watching hundred million narrated video clips. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2630-2640, 2019. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 354, + 287, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 354, + 287, + 387 + ], + "spans": [ + { + "bbox": [ + 48, + 354, + 287, + 387 + ], + "type": "text", + "content": "[31] Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748, 2018. 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 388, + 287, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 388, + 287, + 443 + ], + "spans": [ + { + "bbox": [ + 48, + 388, + 287, + 443 + ], + "type": "text", + "content": "[32] Rahul Rahaman, Dipika Singhania, Alexandre Thiery, and Angela Yao. A generalized and robust framework for timestamp supervision in temporal action segmentation. In Proceedings of the European Conference on Computer Vision, pages 279–296. Springer, 2022. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 445, + 287, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 445, + 287, + 498 + ], + "spans": [ + { + "bbox": [ + 48, + 445, + 287, + 498 + ], + "type": "text", + "content": "[33] Alexander Richard, Hilde Kuehne, and Juergen Gall. Weakly supervised action learning with rn based fine-to-coarse modeling. In Proceedings of the IEEE conference on Computer Vision and Pattern Recognition, pages 754-763, 2017. 1, 2, 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 500, + 287, + 554 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 500, + 287, + 554 + ], + "spans": [ + { + "bbox": [ + 48, + 500, + 287, + 554 + ], + "type": "text", + "content": "[34] Alexander Richard, Hilde Kuehne, Ahsan Iqbal, and Juergen Gall. Neuralnetwork-viterbi: A framework for weakly supervised video learning. In Proceedings of the IEEE conference on Computer Vision and Pattern Recognition, pages 7386-7395, 2018. 1, 2, 5, 6, 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 556, + 287, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 556, + 287, + 600 + ], + "spans": [ + { + "bbox": [ + 48, + 556, + 287, + 600 + ], + "type": "text", + "content": "[35] Yuhan Shen and Ehsan Elhamifar. Semi-weakly-supervised learning of complex actions from instructional task videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3344–3354, 2022. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 601, + 287, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 601, + 287, + 656 + ], + "spans": [ + { + "bbox": [ + 48, + 601, + 287, + 656 + ], + "type": "text", + "content": "[36] Yuhan Shen, Lu Wang, and Ehsan Elhamifar. Learning to segment actions from visual and language instructions via differentiable weak sequence alignment. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10156-10165, 2021. 5" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 658, + 287, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 287, + 689 + ], + "type": "text", + "content": "[37] Yaser Souri, Alexander Richard, Luca Minciullo, and Juergen Gall. On evaluating weakly supervised action segmentation methods. arXiv preprint arXiv:2005.09743, 2020. 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 692, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 692, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 692, + 287, + 713 + ], + "type": "text", + "content": "[38] Yaser Souri, Mohsen Fayyaz, Luca Minciullo, Gianpiero Francesca, and Juergen Gall. Fast weakly supervised action" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 596 + ], + "type": "list", + "angle": 0, + "index": 25, + "blocks": [ + { + "bbox": [ + 327, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 545, + 106 + ], + "type": "text", + "content": "segmentation using mutual consistency. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(10):6196-6208, 2021. 1, 2, 3, 6, 7, 8" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 107, + 545, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 107, + 545, + 162 + ], + "spans": [ + { + "bbox": [ + 307, + 107, + 545, + 162 + ], + "type": "text", + "content": "[39] Yaser Souri, Yazan Abu Farha, Fabien Despinoy, Gianpiero Francesca, and Juergen Gall. Fif: Fast inference approximation for action segmentation. In Proceedings of the DAGM German Conference on Pattern Recognition, pages 282-296. Springer, 2022. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 163, + 545, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 163, + 545, + 206 + ], + "spans": [ + { + "bbox": [ + 307, + 163, + 545, + 206 + ], + "type": "text", + "content": "[40] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 3, 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 208, + 545, + 250 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 208, + 545, + 250 + ], + "spans": [ + { + "bbox": [ + 307, + 208, + 545, + 250 + ], + "type": "text", + "content": "[41] Heng Wang and Cordelia Schmid. Action recognition with improved trajectories. In Proceedings of the IEEE international conference on computer vision, pages 3551-3558, 2013. 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 252, + 545, + 307 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 252, + 545, + 307 + ], + "spans": [ + { + "bbox": [ + 307, + 252, + 545, + 307 + ], + "type": "text", + "content": "[42] Ruibin Xiong, Yunchang Yang, Di He, Kai Zheng, Shuxin Zheng, Chen Xing, Huishuai Zhang, Yanyan Lan, Liwei Wang, and Tieyan Liu. On layer normalization in the transformer architecture. In International Conference on Machine Learning, pages 10524-10533. PMLR, 2020. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 308, + 545, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 308, + 545, + 363 + ], + "spans": [ + { + "bbox": [ + 307, + 308, + 545, + 363 + ], + "type": "text", + "content": "[43] Lian Xu, Wanli Ouyang, Mohammed Bennamoun, Farid Boussaid, and Dan Xu. Multi-class token transformer for weakly supervised semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4310-4319, 2022. 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 365, + 545, + 429 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 365, + 545, + 429 + ], + "spans": [ + { + "bbox": [ + 307, + 365, + 545, + 429 + ], + "type": "text", + "content": "[44] Di Yang, Yaohui Wang, Antitza Dantcheva, Quan Kong, Lorenzo Garattoni, Gianpiero Francesca, and Francois Bremond. Lac-latent action composition for skeleton-based action segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13679-13690, 2023. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 431, + 545, + 463 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 431, + 545, + 463 + ], + "spans": [ + { + "bbox": [ + 307, + 431, + 545, + 463 + ], + "type": "text", + "content": "[45] Fangqiu Yi, Hongyu Wen, and Tingting Jiang. Asformer: Transformer for action segmentation. arXiv preprint arXiv:2110.08568, 2021. 3, 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 464, + 545, + 529 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 464, + 545, + 529 + ], + "spans": [ + { + "bbox": [ + 307, + 464, + 545, + 529 + ], + "type": "text", + "content": "[46] Runzhong Zhang, Suchen Wang, Yueqi Duan, Yansong Tang, Yue Zhang, and Yap-Peng Tan. Hoi-aware adaptive network for weakly-supervised action segmentation. In Proceedings of the Thirty-Second International Joint Conference on Artificial Intelligence, pages 1722–1730, 2023. 1, 2, 7" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 532, + 545, + 596 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 532, + 545, + 596 + ], + "spans": [ + { + "bbox": [ + 307, + 532, + 545, + 596 + ], + "type": "text", + "content": "[47] Dimitri Zhukov, Jean-Baptiste Alayrac, Ramadan Gokberk Cinbis, David Fouhey, Ivan Laptev, and Josef Sivic. Cross-task weakly supervised learning from instructional videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3537-3545, 2019. 1, 2, 6, 7" + } + ] + } + ], + "index": 24 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "18262" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/EfficientDreamer_ High-Fidelity and Robust 3D Creation via Orthogonal-view Diffusion Priors/6a261e68-4855-488d-9a23-298bfd3be3c9_content_list.json b/2024/EfficientDreamer_ High-Fidelity and Robust 3D Creation via Orthogonal-view Diffusion Priors/6a261e68-4855-488d-9a23-298bfd3be3c9_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..a869a40add205a2b4a8d10c77b62ae94a2d61030 --- /dev/null +++ b/2024/EfficientDreamer_ High-Fidelity and Robust 3D Creation via Orthogonal-view Diffusion Priors/6a261e68-4855-488d-9a23-298bfd3be3c9_content_list.json @@ -0,0 +1,1295 @@ +[ + { + "type": "text", + "text": "EfficientDreamer: High-Fidelity and Stable 3D Creation via Orthogonal-view Diffusion Priors", + "text_level": 1, + "bbox": [ + 200, + 130, + 772, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Zhipeng Hu $^{1\\dagger}$ , Minda Zhao $^{1\\dagger}$ , Chaoyi Zhao $^{1}$ , Xinyue Liang $^{1}$ , Lincheng Li $^{1*}$ , Zeng Zhao $^{1}$ , Changjie Fan $^{1}$ , Xiaowei Zhou $^{2}$ , Xin Yu $^{3}$", + "bbox": [ + 181, + 202, + 785, + 238 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ NetEase Fuxi AI Lab, $^{2}$ State Key Lab of CAD&CG, Zhejiang University $^{3}$ University of Queensland", + "bbox": [ + 88, + 239, + 880, + 257 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 291, + 313, + 306 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "While image diffusion models have made significant progress in text-driven 3D content creation, they often fail to accurately capture the intended meaning of text prompts, especially for view information. This limitation leads to the Janus problem, where multi-faced 3D models are generated under the guidance of such diffusion models. In this paper, we propose a robust high-quality 3D content generation pipeline by exploiting orthogonal-view image guidance. First, we introduce a novel 2D diffusion model that generates an image consisting of four orthogonal-view subimages based on the given text prompt. Then, the 3D content is created using this diffusion model. Notably, the generated orthogonal-view image provides strong geometric structure priors and thus improves 3D consistency. As a result, it effectively resolves the Janus problem and significantly enhances the quality of 3D content creation. Additionally, we present a 3D synthesis fusion network that can further improve the details of the generated 3D contents. Both quantitative and qualitative evaluations demonstrate that our method surpasses previous text-to-3D techniques. Project page: https://efficientdreamer.github.io.", + "bbox": [ + 75, + 323, + 473, + 642 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 659, + 209, + 674 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The field of generating photorealistic 2D images from simple text prompts is experiencing rapid growth, and recent advancements are largely attributed to the utilization of diffusion models. These advancements have led to the creation of models capable of producing images of exceptional quality [23-25]. In contrast, generating high-quality 3D content from text prompts poses significantly greater challenges due to the much larger output space and the requirement for 3D consistency. Moreover, the scarcity of adequately large training pairs containing both text and 3D models further handicaps the development of effective models.", + "bbox": [ + 75, + 683, + 468, + 849 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Early methods [9, 12, 19] explore zero-shot text-guided", + "bbox": [ + 94, + 849, + 468, + 866 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/642a679807abd0ee12d8d68e522d94108fcbc6fbd7fbee65788d630494604d1a.jpg", + "image_caption": [ + "Figure 1. The given prompt is A pig wearing a backpack, high quality. Left: The pre-trained Stable Diffusion model struggles to accurately generate images based on specific view instructions, thereby encountering the Janus problem in the text-driven 3D generation. Right: EfficientDreamer leverages our newly introduced orthogonal-view diffusion model, enabling the generation of 3D consistent images depicting the same scene from multiple orthogonal viewpoints." + ], + "image_footnote": [], + "bbox": [ + 501, + 290, + 890, + 510 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3D content creation by utilizing guidance from CLIP [22]. These approaches optimize the underlying 3D representations, such as NeRFs and meshes, to achieve high text-image alignment scores for all 2D renderings. However, these methods tend to generate 3D shapes with poor geometry and unsatisfactory appearance. To address these limitations, DreamFusion [21] showcases impressive capabilities in text-to-3D synthesis by leveraging a powerful pretrained text-to-image diffusion model [24] as a strong image prior. A NeRF is trained to represent 3D models using novel Score Distillation Sampling (SDS) gradients [21] and view-dependent prompts. Subsequent works, such as TextMesh[30], Magic3D [14], and ProlificDreamer [33], further improve the quality of the generated 3D contents.", + "bbox": [ + 496, + 654, + 892, + 867 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "While previous methods have demonstrated impressive results, a significant drawback is the Janus problem, which", + "bbox": [ + 500, + 869, + 893, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "$\\dagger$ Equal contribution", + "bbox": [ + 94, + 875, + 202, + 887 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Corresponding author", + "bbox": [ + 96, + 887, + 218, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "4949", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "remains a major challenge. According to recent studies [1, 16], the primary cause of the Janus problem lies in the inability of pre-trained 2D diffusion models to accurately interpret the view instructions specified in the text prompts. Even though explicit view instructions are given, the diffusion model may still generate images that do not align with these instructions. Since the stability of the 3D content creation process relies heavily on accurate guidance from the diffusion model, imprecise guidance can introduce instability and compromise the quality of the generated content.", + "bbox": [ + 75, + 90, + 472, + 243 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To cope with these weaknesses, we propose EfficientDreamer, which leverages a viewpoint-aware diffusion prior to create high-quality and consistent 3D contents. Our approach introduces a novel image diffusion model called orthogonal-view diffusion model, which can generate a composite image comprising four sub-images that are mutually consistent. These sub-images, within a composite image, can be derived from arbitrary viewpoints while adhering to orthogonal views. To achieve this, we maintain the architecture design of the 2D image diffusion model while adapting it to orthogonal-view image generation. We render the 3D models in a large-scale open-source 3D dataset called Objaverse [4], from orthogonal views, and assemble them to fine-tune the pre-trained 2D diffusion model. This process ensures the geometric consistency of our orthogonal-view diffusion model while preserving the capacity of 2D diffusion models to generate detailed and diverse high-quality objects that may not be present in the 3D datasets.", + "bbox": [ + 75, + 244, + 472, + 531 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "By leveraging our innovative orthogonal-view image diffusion model, we can effectively tackle the challenge of the multi-face Janus problem while generating consistent 3D models. Specifically, during each training step, we render the 3D representation (like Neus [32] or DMTet [28]) from different orthogonal viewpoints in a differentiable manner. These rendered images are then assembled into a composite image, which is further utilized to optimize the 3D representation via orthogonal-view score distillation. While our orthogonal-view diffusion model, trained with the Objaverse 3D dataset, provides crucial geometric structure priors for 3D content creation, it is important to note that the scale of the 3D dataset is considerably smaller compared to the text-image training pairs. Additionally, the 3D assets in this dataset typically consist of common objects from the real world, having relatively simple topology and texture.", + "bbox": [ + 75, + 535, + 472, + 777 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Relying solely on such an orthogonal-view diffusion prior may lead to potential artifacts in the local appearance of the 3D contents. To address this issue, we also incorporate the original pre-trained 2D diffusion prior in the optimization of the 3D representation. In particular, we introduce a 3D synthesis fusion network which integrates the orthogonal-view diffusion model with a pre-trained 2D diffusion model. We also present a dynamic 3D synthesis strat", + "bbox": [ + 75, + 780, + 472, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "egy to balance the guidance of the orthogonal-view diffusion prior and pre-trained 2D diffusion prior during the SDS optimization process. Initially, our focus is on generating 3D consistent geometric structures, where the high weight assigned to the orthogonal-view diffusion prior helps efficiently represent the geometry and alleviate the Janus problem. As the 3D objects become more complete, we gradually reduce the weight of orthogonal-view diffusion prior and increase the guidance from the original 2D diffusion prior. This approach allows for the generation of local texture details and refinement of local appearance.", + "bbox": [ + 496, + 90, + 893, + 257 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Overall, our contributions are summarized as follows:", + "bbox": [ + 517, + 258, + 874, + 272 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose EfficientDreamer, a novel approach for high-fidelity and robust 3D creation using orthogonal-view diffusion prior. This method effectively tackles the Janus problem and significantly enhances the stability of the generation process.", + "- We introduce a novel image diffusion model that generates a composite image composed of four sub-images from orthogonal viewpoints. These orthogonal-view images serve as essential geometric structure priors, enabling the generation of 3D consistent geometric structures.", + "- We present a novel text-driven 3D creation pipeline based on the orthogonal-view SDS loss. We introduce a 3D synthesis fusion network that combines the newly introduced orthogonal-view diffusion prior with the original 2D diffusion prior, aiming to ensure high-quality 3D creations." + ], + "bbox": [ + 500, + 273, + 890, + 515 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 532, + 640, + 549 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. 3D Reconstruction with Neural Fields", + "text_level": 1, + "bbox": [ + 500, + 559, + 823, + 574 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Traditional Multi-View Stereo (MVS) methods reconstruct 3D point clouds from predicted depth maps [6, 26, 34, 42]. Benefiting from deep learning, supervised MVS methods [7, 31, 35-37] have shown impressive performance on benchmarks but require training on specific datasets [10, 38]. These methods often struggle in less constrained scenarios and cannot easily integrate with other learning-based systems. On the other hand, in recent times, there have been significant advancements in neural fields that have demonstrated remarkable results across various tasks. As a famous neural volume rendering method, NeRF [5, 18, 20, 40] combines classical volumetric rendering with implicit function to render high-quality 2D images. However, directly extracting a mesh from a NeRF representation is non-trivial [29]. Recent works have focused on improving the geometric network while establishing connections between density-based and surface-based representations. VolSDF [39] models the volume density as Laplace Cumulative Distribution Function applied to an SDF representation. Similar to VolSDF, NeuS [32] transforms the SDF field to the accumulated transmittance for volume rendering with the", + "bbox": [ + 496, + 583, + 893, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "4950", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Logistic Cumulative Distribution Function and it is emphasized that the weight of the rendering should peak at the first intersection point from the outside to the inside. To this end, we adopt the NeuS representation for the convenience of mesh extraction.", + "bbox": [ + 75, + 90, + 470, + 165 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. Text-to-image Generation", + "text_level": 1, + "bbox": [ + 76, + 176, + 310, + 193 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In recent years, with the availability of extremely large datasets of image-text pairs [27], significant progress has been made in text-to-image generation with diffusion models. A typical method is the Stable Diffusion model, which samples from a lower-resolution latent space and decodes latent into high-resolution images [24]. Such sampling operation in the latent greatly improves the efficiency of image generation. Another kind of method utilizes a cascade of super-resolution models [2, 25]. These methods first generate a low-resolution image from a given text prompt and then enlarge the images with several super-resolution models. Recent works design special neural network structures to control diffusion models by adding extra conditions [41], it can learn task-specific input conditions and produce customized images. Due to the strong generative ability of pretrained diffusion models, they are convenient to enable text-to-3D mesh synthesis.", + "bbox": [ + 75, + 200, + 470, + 455 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.3. Text-to-3D Generation", + "text_level": 1, + "bbox": [ + 76, + 468, + 287, + 483 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In recent years, text-to-3D has received significant attention due to the desire to create high-quality 3D content from simple semantics, such as text prompts. Early works attempt to use a CLIP objective to supervise the generation. CLIPMesh [19] deforms a 3D sphere using a CLIP loss to obtain a 3D mesh that fits the input prompt. Text2Mesh [17] stylizes a 3D mesh by predicting color and local geometric details which conform to a target text prompt. Dreamfield [9] optimizes a NeRF from many camera views so that rendered images score highly with a target caption according to a pre-trained CLIP model. However, these methods tend to generate poor 3D shapes with unsatisfactory geometry as well as appearance.", + "bbox": [ + 75, + 492, + 468, + 688 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To overcome such limitations, DreamFusion [21] employs text-to-3D synthesis by utilizing a powerful pretrained text-to-image diffusion model as a strong image prior. They optimize NeRF with their proposed Score Distillation Sampling (SDS) loss. Magic3D [14] extends such a method to a two-stage coarse-to-fine process, which utilizes DMTet for mesh representation. Fantasy3D [3] disentangles geometry and appearance into a two-stage optimization and introduces the spatially varying Bidirectional Reflectance Distribution Function (BRDF) into the text-to-3D task. TextMesh [30] replaces NeRF with VolSDF for more accurate mesh expression. It proposes a novel multi-view consistent and mesh-conditioned re-texturing, enabling the generation of a photorealistic 3D mesh model.", + "bbox": [ + 75, + 688, + 470, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "More recently, ProlificDreamer [33] proposes Variational Score Distillation (VSD), a principled particle-based variational framework to promote the quality of created 3D models.", + "bbox": [ + 496, + 90, + 890, + 150 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Methodology", + "text_level": 1, + "bbox": [ + 498, + 164, + 635, + 181 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given a text prompt $y$ , our goal is to create high-quality 3D mesh representations with photorealistic texture. To achieve this, we propose a novel orthogonal-view diffusion model, denoted as $\\phi_{ov}$ , which generates composite images consisting of four orthogonal-view sub-images. Subsequently, we design a 3D synthesis fusion network that combines the orthogonal-view diffusion model with a pre-trained diffusion model, denoted as $\\phi_{pre}$ , to enhance the quality of the generated 3D content.", + "bbox": [ + 496, + 188, + 890, + 325 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Orthogonal-view Diffusion Model", + "text_level": 1, + "bbox": [ + 498, + 333, + 795, + 349 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our objective is to generate a composite image in response to a given text prompt $y$ using a fine-tuned diffusion model. The composite image will consist of four sub-images from orthogonal viewpoints, arranged in a $2 \\times 2$ grid.", + "bbox": [ + 496, + 357, + 890, + 417 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Common text-to-image diffusion models are trained on large-scale text-image pairs, enabling them to describe objects from various viewpoints. However, these models do not explicitly align corresponding parts of the object across viewpoints. Moreover, they tend to generate images of objects in canonical poses. Hence, it is essential to introduce a novel diffusion model that can generate representations of the same object from diverse viewpoints while maintaining 3D consistency. In this paper, we propose that showcasing the object from four orthogonal views is suitable as it provides essential geometric structure priors.", + "bbox": [ + 496, + 417, + 890, + 583 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To obtain such a specific image diffusion model, we use the recently released Objaverse [4] dataset for fine-tuning the commonly-used diffusion model, i.e., the Stable Diffusion model [24]. This dataset contains $800\\mathrm{K} + 3\\mathrm{D}$ models created by $100\\mathrm{K}+$ artists. The descriptions for each 3D model can be found in the Cap3D [15] dataset, which is processed with BLIP2 [13] and GPT4. We first filter some improper 3D models, like complex scene representation, textureless 3D models, and point clouds, and approximately $420\\mathrm{K}$ high-quality 3D models are left. We normalize all assets into a unit cube $[-0.5, 0.5]^3$ . The radius (distance away from the center) is set at 1.8, with a field of view set at 35 degrees. For each 3D model, we uniformly render 12 images in the range of azimuth angle $\\xi_{cam} \\in (0^\\circ, 360^\\circ)$ and elevation angle list $\\delta_{cam} = [0^\\circ, 15^\\circ, 30^\\circ, 45^\\circ]$ . Thus, for each 3D asset, we render 48 RGBA images in $512 \\times 512$ resolution in total while its corresponding text prompt is denoted as $y$ . We use the Stable Diffusion 2.1 as the original latent diffusion architecture with an encoder $\\mathcal{E}$ , a denoiser U-Net $\\epsilon_{\\theta}$ , and a decoder $\\mathcal{D}$ . We utilize $\\mathcal{E}$ and $\\mathcal{D}$ with pre-trained weight and only fine-tune $\\epsilon_{\\theta}$ for generating composite im", + "bbox": [ + 496, + 584, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "4951", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/da7ecf985dc177f07d465a5ef1ca0fa96a757db29bec895f15e87cb9d2cf416d.jpg", + "image_caption": [ + "Figure 2. The overview of EfficientDreamer involves two key steps. Firstly, we train an orthogonal-view diffusion model by rendering images from the Objaverse dataset. Secondly, we optimize the 3D scene representation by leveraging both the newly introduced orthogonal-view diffusion model and the pre-trained text-to-image diffusion model. To ensure high-fidelity and robust 3D creation, we employ a dynamic 3D synthesis strategy." + ], + "image_footnote": [], + "bbox": [ + 84, + 87, + 890, + 415 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "ages from orthogonal views. As illustrated in Fig. 2(a), for each training step, we first sample an elevation angle from the elevation angle list and then select four render images in this elevation angle, the rendered images are required to be 90 degrees apart from each other in azimuth angle. Then these four images are tiled on a $2 \\times 2$ grid in a clockwise rotation. To match the diffusion model, we resize this composite image into $512 \\times 512$ resolution, which is denoted as $x$ . At the diffusion time step $t \\sim [1, 1000]$ , let $c(y)$ be the embedding of $y$ , and $z_t$ be the noisy latent image by adding noise $\\epsilon$ to a clean latent image $z$ , we then solve for the following objective to fine-tune the model:", + "bbox": [ + 75, + 494, + 472, + 676 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\underset {\\theta} {\\min } \\mathbb {E} _ {z \\sim \\mathcal {E} (x), t, \\epsilon \\sim \\mathcal {N} (0, 1)} | | \\epsilon - \\epsilon_ {\\theta} \\left(z _ {t}, t, c (y)\\right) | | ^ {2}. \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 119, + 685, + 468, + 710 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We find the trained diffusion model is capable of learning a generic mechanism to explicitly align the correspondence of an identical object from different viewpoints, even for objects never appearing in the training dataset. The comparison between the orthogonal-view diffusion model and the pre-trained diffusion model is depicted in Fig. 3.", + "bbox": [ + 75, + 717, + 470, + 809 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Text-to-3D via the 3D Synthesis Fusion Network", + "text_level": 1, + "bbox": [ + 76, + 816, + 470, + 834 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Once the orthogonal-view diffusion model is trained, we can generate our initial 3D model by training a neural distance field using the score distillation sampling strategy. The scene is represented as a differential render $g(\\psi)$ , where", + "bbox": [ + 75, + 839, + 470, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "$\\psi$ denotes the learnable parameters. Given a randomly sampled camera pose with azimuth angle $\\xi_{cam}$ , elevation angle $\\delta_{cam}$ and radius $r$ , we fix $\\delta_{cam}$ and $r$ and sample four camera poses by extending $\\xi_{cam}$ to $[\\xi_{cam}, \\xi_{cam} + 90^{\\circ}, \\xi_{cam} + 180^{\\circ}, \\xi_{cam} + 270^{\\circ}]$ . Then we render images with these camera poses, which are denoted as $I_0$ , $I_1$ , $I_2$ , $I_3$ . We also tile them on a $2 \\times 2$ grid in a clockwise rotation and resize the composite image into $512 \\times 512$ resolution, which is denoted as $\\hat{I}$ . Then we sample random normal noise according to time step $t$ and add it to $\\hat{I}$ . The noisy images $\\tilde{I}_t$ , together with text embedding $c(y)$ are fed to our orthogonal-view diffusion model $\\phi_{ov}$ , which attempts to predict the noise $\\epsilon$ . The denoiser U-Net of $\\phi_{ov}$ is denoted as $\\epsilon_{\\phi_{ov}}$ and the score function of noise estimation is denoted as $\\epsilon_{\\phi_{ov}}(\\tilde{I}_t; t, c(y))$ . This score function guides the direction of the gradient for updating the scene parameters, and the gradient is calculated by orthogonal-view Score Distillation Sampling (SDS):", + "bbox": [ + 496, + 494, + 893, + 767 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\nabla_ {\\psi} \\mathcal {L} _ {S D S} ^ {o v} = \\mathbb {E} _ {t, \\epsilon} [ \\omega (t) (\\epsilon_ {\\phi_ {o v}} (\\tilde {I} _ {t}; t, c (y)) - \\epsilon) \\frac {\\partial \\hat {I}}{\\partial \\psi} ], \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 521, + 777, + 890, + 813 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\omega (t)$ denotes a weighting function.", + "bbox": [ + 498, + 824, + 774, + 839 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Motivated by [8] and to enhance the stability of the training procedure, we decrease the time step (t) as training progresses, resulting in a gradual decline in the degree of added noise. As our training procedure supervises the 3D creation", + "bbox": [ + 496, + 839, + 893, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4952", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/293360eb425ebb78f77f12e8571fd6b49974451128e890cc5110a50daedaee30.jpg", + "image_caption": [ + "Figure 3. Comparison between The orthogonal-view diffusion model and the pre-trained Stable Diffusion model with additional viewpoint instructions. The pre-trained Stable Diffusion model struggles to generate images based on specific view instructions, while the orthogonal-view diffusion model can generate composite images from orthogonal views." + ], + "image_footnote": [], + "bbox": [ + 83, + 90, + 467, + 426 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "from four orthogonal views at each time step, the neural distance field can receive semantic structure guidance from different viewpoints. This fundamentally eliminates the possibility of creating multi-face problems.", + "bbox": [ + 75, + 564, + 468, + 625 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our orthogonal-view diffusion model, trained on the Objaverse 3D dataset, provides essential geometric structure priors for 3D creation. However, it is important to note that the scale of the 3D dataset is relatively small compared to the text-image training pairs. Additionally, the included 3D assets mainly consist of simple objects with uncomplicated topology and texture. Therefore, relying solely on the orthogonal-view diffusion prior may introduce potential artifacts in the local appearance of the generated 3D contents.", + "bbox": [ + 75, + 633, + 468, + 772 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To resolve such a problem, we introduce a 3D synthesis fusion network which integrates the proposed orthogonal-view model with the original text-to-image diffusion model to guide the 3D model generation with a 3D synthesis dynamic strategy. As shown in Fig. 2(b), let $\\epsilon_{\\phi_{pre}}$ denote the denoiser U-Net of the pre-trained Stable Diffusion model, and rearrange the original rendered images $I_0, I_1, I_2, I_3$ as an image batch $\\hat{\\mathbb{I}}$ . $\\tilde{\\mathbb{I}}_t$ denotes the noisy result of $\\hat{\\mathbb{I}}$ . The", + "bbox": [ + 75, + 780, + 470, + 902 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "corresponding SDS loss is formulated as:", + "bbox": [ + 498, + 90, + 774, + 107 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\nabla_ {\\psi} \\mathcal {L} _ {S D S} ^ {p r e} = \\mathbb {E} _ {t, \\epsilon} [ \\omega (t) (\\epsilon_ {\\phi_ {p r e}} (\\tilde {\\mathbb {I}} _ {t}; t, c (y)) - \\epsilon) \\frac {\\partial \\hat {\\mathbb {I}}}{\\partial \\psi} ]. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 519, + 116, + 890, + 150 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We propose a new prior loss for the novel view supervision to combine both two priors:", + "bbox": [ + 498, + 160, + 890, + 190 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\nabla_ {\\psi} \\mathcal {L} _ {S D S} = (1 - (\\frac {l}{L}) ^ {\\lambda}) \\nabla_ {\\psi} \\mathcal {L} _ {S D S} ^ {o v} + (\\frac {l}{L}) ^ {\\lambda} \\nabla_ {\\psi} \\mathcal {L} _ {S D S} ^ {p r e}, (4)\n$$\n", + "text_format": "latex", + "bbox": [ + 508, + 200, + 890, + 231 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $l$ is the current iteration while $L$ is the total iterations. $\\lambda$ is a hyperparameter to determine the strength of two diffusion priors, which is set as 1 in our experiments.", + "bbox": [ + 498, + 238, + 890, + 282 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Initially, we prioritize generating 3D content with consistent geometric structures by emphasizing the orthogonal-view diffusion prior. As the generation progresses, we gradually shift the focus towards incorporating guidance from the original 2D diffusion prior to enhance local texture details and refine the overall appearance.", + "bbox": [ + 498, + 284, + 890, + 375 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Implementation Details", + "text_level": 1, + "bbox": [ + 500, + 383, + 718, + 400 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Orthogonal-view Diffusion Model Training: Our model is implemented on the PyTorch platform with 8 NVIDIA Tesla A100 GPUs. We train the model utilizing AdamW optimizer with a learning rate of 1e-5. The whole training procedure takes 200K iterations with a batch size of 256, which takes about 5 days.", + "bbox": [ + 498, + 406, + 890, + 496 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Text-to-3D Generation: We employ a two-stage coarse-to-fine optimization process to generate high-quality 3D mesh representations with photorealistic textures from text prompts. In the coarse stage, we utilize NeuS [32] as the 3D scene representation, as density-based representations such as NeRF are not well-suited for extracting 3D geometry and obtaining meshes [39]. The rendering resolution is set at $64 \\times 64$ , and the 3D representation is optimized for 5000 steps with 3D Synthesis Fusion Network. In the refinement stage, we employ DMTet [28] as the scene representation, enabling efficient rendering of textured meshes with differentiable rasterization at $512 \\times 512$ resolutions. Additionally, we utilize SDS and VSD guidance [33] to optimize the geometry and texture for another 5000 steps.", + "bbox": [ + 496, + 497, + 892, + 709 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 500, + 722, + 633, + 739 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this section, we conduct extensive experiments to evaluate the effectiveness of our proposed method for text-to-3D content creation. We present qualitative and quantitative results, compare our method with state-of-the-art methods, perform a user study, and analyze the effectiveness of our proposed techniques through an ablation study.", + "bbox": [ + 496, + 747, + 890, + 839 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Comparison with State-of-the-arts", + "text_level": 1, + "bbox": [ + 500, + 847, + 802, + 864 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We present comprehensive experiments to evaluate the effectiveness of our proposed method for text-to-3D content", + "bbox": [ + 500, + 869, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "4953", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/62c7d2d651bd7a7cc77d5e883faf542aabd95fb07bbc7e8d792e89b227a74194.jpg", + "image_caption": [ + "Figure 4. Comparison with other text-to-3D methods. We render each 3D model from two views. Our method outperforms other techniques by generating more high-fidelity 3D models without encountering the Janus problem." + ], + "image_footnote": [], + "bbox": [ + 81, + 90, + 883, + 420 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "creation. The compared methods include three crucial Text-to-3D methods, i.e., DreamFusion [21], Magic3D [14] and TextMesh [30]. All these methods are implemented with a unified framework for 3D content creation*. As DreamFusion and TextMesh generate 3D contents with volume rendering of $64 \\times 64$ resolution, the generated 3D models may suffer from over-smoothness and lack of details. However, Magic3D can generate higher-quality 3D shapes with improved geometry and texture. Since our method also utilizes a coarse-to-fine two-stage optimization, we are able to create high-quality 3D mesh models. Furthermore, all of the compared methods may be confronted with the Janus problem as they rely on open-source pre-trained diffusion priors that do not enforce valid 3D consistency constraints. Our method can resolve the multi-face problem with our newly introduced orthogonal-view diffusion prior. Please refer to the supplementary material for more comparison results.", + "bbox": [ + 75, + 473, + 472, + 731 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We also evaluate our method using the CLIP score [22] and FID [11] metrics. The CLIP score measures how well images rendered from the generated objects correlate with the provided input text prompt, while FID evaluates the quality and photorealistic appearance of the generated shapes. We render four images with fixed camera poses for each 3D model. The results are depicted in Table 1, which demonstrates that our methods achieve better performance than the state-of-the-art methods.", + "bbox": [ + 75, + 733, + 468, + 869 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1. Comparison with state-of-the-arts. Comparing our method against the state-of-the-arts using CLIP score and FID. $\\downarrow$ indicates the lower the better, and $\\uparrow$ indicates the higher the better.", + "bbox": [ + 500, + 472, + 892, + 513 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/922cd00208ed3484b412cd3c50b5b84dceed2a7ce26e5d7b63c6b8367aead305.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodsCLIP ↑FID ↓
DreamFusion [21]28.40374.44
Magic3D [14]29.15310.57
TextMesh [30]27.65305.77
Ours30.33284.98
", + "bbox": [ + 565, + 523, + 818, + 612 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Comparison with Perp-Neg", + "text_level": 1, + "bbox": [ + 500, + 631, + 746, + 647 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In addition, we conduct a comparison between our method and another approach, Perp-Neg [1], which is designed to mitigate the Janus problem. We integrate Perp-Neg into DreamFusion and evaluate its performance against our method. The results, shown in Figure 5, reveal that Perp-Neg can address the Janus problem to some extent. For instance, it is capable of generating 3D models with only one head and face. However, these results may exhibit unreasonable details such as three ears or five legs. In contrast, our method, which incorporates guidance from orthogonal views simultaneously, enables us to avoid such issues.", + "bbox": [ + 496, + 654, + 890, + 821 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Coarse-to-fine Two-stage Optimization Results", + "text_level": 1, + "bbox": [ + 500, + 830, + 890, + 848 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In Figure 6, we demonstrate the 3D contents generated through the coarse and fine stages of our optimization procedure. In the coarse stage, we achieve precise 3D represent", + "bbox": [ + 496, + 854, + 892, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_footnote", + "text": "* https://github.com/threestudio-project/threestudio", + "bbox": [ + 94, + 886, + 364, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "4954", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/a76a699429cdd41e563810680f4b6402b1206c7f2f25abf8b56a1d6bd454b971.jpg", + "image_caption": [ + "Figure 5. Comparison between the Perp-Neg method and our approach: Our method offers a more comprehensive solution to the Janus problem compared to Perp-Neg." + ], + "image_footnote": [], + "bbox": [ + 76, + 88, + 472, + 281 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/5917dccc5fd9a428838b4dcc961c904a54667557b0eb35b3607c47b1c0a175a0.jpg", + "image_caption": [ + "Figure 6. Coarse-to-fine two-stage optimization results. In the coarse stage, we get semantically reasonable 3D scene representation. In the fine stage, we improve the visual quality of generated 3D assets." + ], + "image_footnote": [], + "bbox": [ + 81, + 367, + 472, + 636 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "tations for the given text prompts while effectively addressing the challenging Janus problem. This is accomplished with the guidance of the orthogonal-view diffusion model, which provides essential geometric structure priors for 3D creation. In the fine stage, we initialize the mesh representation with the results obtained from the coarse stage and proceed to optimize the geometry using SDS guidance. Additionally, we leverage VSD guidance to enhance the texture, ensuring that the generated 3D model is visually appealing and realistic. As a result, this process enables the generation of significantly higher-quality 3D models.", + "bbox": [ + 75, + 734, + 472, + 902 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/f8502a21869d43ca15f8027257789acb1dacbed6587066ae8c59313c64c480f2.jpg", + "table_caption": [ + "Table 2. Results of our user study. The rating score represents the average rating results for different methods, while the preference indicates the ratio of user preference for the results of each method with higher quality." + ], + "table_footnote": [], + "table_body": "
MethodsRating Score ↑Preference(%) ↑
DreamFusion [21]2.094.55
Magic3D [14]2.446.27
TextMesh [30]1.734.91
Ours3.7484.27
", + "bbox": [ + 521, + 155, + 859, + 242 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. User Study", + "text_level": 1, + "bbox": [ + 500, + 256, + 622, + 273 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We conduct a user study with Fuxi Youling Crowdsourcing \\* for quantifying the subjective quality. Concretely, we evaluate the study on 22 prompts. At each evaluation time, the participants are shown the created 3D models by DreamFusion [21], Magic3D [14], TextMesh[30] and our method, and asked to rate them on a scale of 1 to 4. In total, we collect 1100 responses from 50 participants. The average rating scores for the different methods are shown in Table 2. Notably, our method receives the highest average rating, indicating that it produces the most visually high-fidelity 3D models. Furthermore, we calculate the preference ratios for each method based on the quality of the results. The analysis reveals that $84.27\\%$ of the users consider the results of our method to be of higher quality than those generated by all competing methods. This finding further validates the effectiveness of our method in generating visually highfidelity 3D models.", + "bbox": [ + 496, + 280, + 893, + 537 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.5. Ablation Study", + "text_level": 1, + "bbox": [ + 500, + 547, + 653, + 565 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Impact of the 3D synthesis Fusion Network: We introduce a 3D synthesis fusion network with the dynamic 3D synthesis strategy. Here we provide analysis and experiments to show its effectiveness. Given several text prompts, we employ 3D creation with different diffusion model guidance. The corresponding results are shown in Fig. 7. The first two columns show the results generated without the orthogonal-view diffusion model, which is equivalent to $\\lambda = 0$ . Obvious Janus problems and inaccurate 3D representation can be found for the first two prompts. As for the creation of a cottage, we get similar shapes from different viewpoints. The next two columns demonstrate the result guided without the pre-trained diffusion model, which is equivalent to $\\lambda = \\infty$ . While the orthogonal-view diffusion model can provide comprehensive structure constraints for 3D generation and restrain multi-face issues, we can see local distortion and unreasonable roughness. The last two columns demonstrate the results with our 3D synthesis fusion network, which means $\\lambda = 1$ . We get complete and smooth meshes and eliminate the possible Janus problem.", + "bbox": [ + 496, + 571, + 893, + 876 + ], + "page_idx": 6 + }, + { + "type": "page_footnote", + "text": "* https://fuxi.163.com/solution/data", + "bbox": [ + 517, + 886, + 705, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "4955", + "bbox": [ + 482, + 944, + 514, + 957 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/3c1a79e6fb43a75a12aaff8ef95cc67180e5f8914ef5989464d27fdf57852e19.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 83, + 88, + 460, + 304 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/4cd282259356fea7311a49bde6920368471e53357565a9722651718ec2e0f6ef.jpg", + "image_caption": [ + "Figure 7. Ablation study on the impact of the 3D synthesis fusion network." + ], + "image_footnote": [], + "bbox": [ + 83, + 306, + 460, + 401 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Quantitative results are depicted in Table 3, which demonstrate the effectiveness of the 3D synthesis fusion network. Impact of Varying Number of View Supervision: Here we present an ablation study on the effects of varying the number of view supervisions on our orthogonal-view diffusion model. For two-view supervision, we randomly mask the opposite viewpoints in the rendered images during training, resulting in the optimization of the 3D representation using only two views. In the case of three-view supervision, we randomly mask one viewpoint, allowing the optimization of the 3D representation using three views. The corresponding results are shown in Figure 8. Generally, supervising the 3D creation with our orthogonal-view diffusion model can mitigate Janus problems. However, creating 3D models with two-view supervision may still suffer from slight multi-face problems, such as two barrels in a tank or multiple faces in a squirrel and a horse. With three-view supervision, these problems are alleviated to some extent. Nevertheless, there are still local distortions in generated 3D models, such as unexpected sags or bulges. These issues can be addressed by using the intact four-view supervision of our orthogonal-view diffusion model. Quantitative results are also presented in Table 3, illustrating the impact of reducing the number of view supervision signals during training. It is evident that there is a decrease in performance observed in this scenario. This highlights the significance of optimizing a 3D representation using a set of four orthogonal-view images, as it provides sufficient geometric structure priors for ensuring consistent 3D content creation.", + "bbox": [ + 75, + 446, + 472, + 900 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/4ef702f17ba7668fb5bcbfc6673da021817437f796bb128d9ed16090ef46e3cb.jpg", + "image_caption": [ + "Figure 8. Ablation study on the impact of varying number of view supervision." + ], + "image_footnote": [], + "bbox": [ + 511, + 90, + 885, + 392 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/ba05ab6dd04e92a9bad7a6d9b7fcd25879214cd3c4fabd288ca9d982fd7c3d6b.jpg", + "table_caption": [ + "Table 3. Ablation study results using CLIP and FID. $\\downarrow$ indicates the lower the better, and $\\uparrow$ indicates the higher the better." + ], + "table_footnote": [], + "table_body": "
Impact of the 3D synthesis Fusion Network
Metricsw/o Orthogonal-vieww/o Pre-trainedFull
CLIP ↑31.1332.0532.55
FID ↓372.86369.73342.62
Impact of Varying Number of View Supervision
MetricsTwo-viewThree-viewFour-view
CLIP ↑30.9230.4831.47
FID ↓399.58373.04345.72
", + "bbox": [ + 501, + 489, + 905, + 609 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 633, + 617, + 648 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we present EfficientDreamer, a novel approach for high-fidelity and robust 3D creation using orthogonal-view diffusion priors. We introduce a novel diffusion model that can create composite images comprising sub-images from different orthogonal viewpoints. These composite images provide important geometric structure priors for 3D model generation by showcasing an object from multiple perspectives. We enhance the text-to-3D generation framework by leveraging the orthogonal-view diffusion model. Additionally, we introduce a 3D synthesis fusion network that integrates the orthogonal-view diffusion prior with the pre-trained diffusion prior. We also introduce a dynamic 3D synthesis strategy to balance the guidance of these two diffusion models during the SDS optimization. Our method effectively addresses the challenging Janus problem while ensuring the fidelity of the generated 3D models.", + "bbox": [ + 496, + 664, + 893, + 898 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "4956", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Mohammadreza Armandpour, Huangjie Zheng, Ali Sadeghian, Amir Sadeghian, and Mingyuan Zhou. Re-imagine the Negative Prompt Algorithm: Transform 2D Diffusion into 3D, alleviate Janus problem and Beyond. arXiv preprint arXiv:2304.04968, 2023. 2, 6", + "[2] Yogesh Balaji, Seungjun Nah, Xun Huang, Arash Vahdat, Jiaming Song, Karsten Kreis, Miika Aittala, Timo Aila, Samuli Laine, Bryan Catanzaro, et al. edifi: Text-to-image diffusion models with an ensemble of expert denoisers. arXiv preprint arXiv:2211.01324, 2022. 3", + "[3] Rui Chen, Yongwei Chen, Ningxin Jiao, and Kui Jia. Fantasia3d: Disentangling geometry and appearance for high-quality text-to-3d content creation. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 22246–22256, 2023. 3", + "[4] Matt Deitke, Dustin Schwenk, Jordi Salvador, Luca Weihs, Oscar Michel, Eli VanderBilt, Ludwig Schmidt, Kiana Ehsani, Aniruddha Kembhavi, and Ali Farhadi. Objverse: A universe of annotated 3d objects. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13142-13153, 2023. 2, 3", + "[5] Sara Fridovich-Keil, Alex Yu, Matthew Tancik, Qinhong Chen, Benjamin Recht, and Angjoo Kanazawa. Plenoxels: Radiance fields without neural networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5501–5510, 2022. 2", + "[6] Silvano Galliani, Katrin Lasinger, and Konrad Schindler. Massively parallel multiview stereopsis by surface normal diffusion. In Proceedings of the IEEE International Conference on Computer Vision, pages 873-881, 2015. 2", + "[7] Xiaodong Gu, Zhiwen Fan, Siyu Zhu, Zuozhuo Dai, Feitong Tan, and Ping Tan. Cascade cost volume for high-resolution multi-view stereo and stereo matching. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2495-2504, 2020. 2", + "[8] Yukun Huang, Jianan Wang, Yukai Shi, Xianbiao Qi, Zheng-Jun Zha, and Lei Zhang. Dreamtime: An improved optimization strategy for text-to-3d content creation. arXiv preprint arXiv:2306.12422, 2023. 4", + "[9] Ajay Jain, Ben Mildenhall, Jonathan T Barron, Pieter Abbeel, and Ben Poole. Zero-shot text-guided object generation with dream fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 867-876, 2022. 1, 3", + "[10] Rasmus Jensen, Anders Dahl, George Vogiatzis, Engin Tola, and Henrik Aanæs. Large scale multi-view stereopsis evaluation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 406-413, 2014. 2", + "[11] Tuomas Kynkänniemi, Tero Karras, Miika Aittala, Timo Aila, and Jaakko Lehtinen. The role of imagenet classes in fr\\echet inception distance. arXiv preprint arXiv:2203.06026, 2022. 6", + "[12] Han-Hung Lee and Angel X Chang. Understanding pure clip guidance for voxel grid nerf models. arXiv preprint arXiv:2209.15172, 2022. 1" + ], + "bbox": [ + 78, + 114, + 470, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[13] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. arXiv preprint arXiv:2301.12597, 2023. 3", + "[14] Chen-Hsuan Lin, Jun Gao, Luming Tang, Towaki Takikawa, Xiaohui Zeng, Xun Huang, Karsten Kreis, Sanja Fidler, Ming-Yu Liu, and Tsung-Yi Lin. Magic3d: High-resolution text-to-3d content creation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 300–309, 2023. 1, 3, 6, 7", + "[15] Tiange Luo, Chris Rockwell, Honglak Lee, and Justin Johnson. Scalable 3d captioning with pretrained models. arXiv preprint arXiv:2306.07279, 2023. 3", + "[16] Gal Metzer, Elad Richardson, Or Patashnik, Raja Giryes, and Daniel Cohen-Or. Latent-nerf for shape-guided generation of 3d shapes and textures. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12663–12673, 2023. 2", + "[17] Oscar Michel, Roi Bar-On, Richard Liu, Sagie Benaim, and Rana Hanocka. Text2mesh: Text-driven neural stylization for meshes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13492-13502, 2022. 3", + "[18] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 2", + "[19] Nasir Mohammad Khalid, Tianhao Xie, Eugene Belilovsky, and Tiberiu Popa. Clip-mesh: Generating textured meshes from text using pretrained image-text models. In SIGGRAPH Asia 2022 conference papers, pages 1-8, 2022. 1, 3", + "[20] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM Transactions on Graphics (ToG), 41(4):1-15, 2022. 2", + "[21] Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. In ICLR, 2023. 1, 3, 6, 7", + "[22] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 1, 6", + "[23] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022. 1", + "[24] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 1, 3", + "[25] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans," + ], + "bbox": [ + 501, + 92, + 893, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "4957", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in Neural Information Processing Systems, 35:36479-36494, 2022. 1, 3", + "[26] Johannes L Schonberger, Enliang Zheng, Jan-Michael Frahm, and Marc Pollefeys. Pixelwise view selection for unstructured multi-view stereo. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part III 14, pages 501-518. Springer, 2016. 2", + "[27] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. Advances in Neural Information Processing Systems, 35:25278-25294, 2022. 3", + "[28] Tianchang Shen, Jun Gao, Kangxue Yin, Ming-Yu Liu, and Sanja Fidler. Deep marching tetrahedra: a hybrid representation for high-resolution 3d shape synthesis. Advances in Neural Information Processing Systems, 34:6087-6101, 2021. 2, 5", + "[29] Jiaxiang Tang, Hang Zhou, Xiaokang Chen, Tianshu Hu, Errui Ding, Jingdong Wang, and Gang Zeng. Delicate textured mesh recovery from nerf via adaptive surface refinement. arXiv preprint arXiv:2303.02091, 2023. 2", + "[30] Christina Tsalicoglou, Fabian Manhardt, Alessio Tonioni, Michael Niemeyer, and Federico Tombari. Textmesh: Generation of realistic 3d meshes from text prompts. 2024. 1, 3, 6, 7", + "[31] Fangjinhua Wang, Silvano Galliani, Christoph Vogel, Pablo Speciale, and Marc Pollefeys. Patchmatchnet: Learned multi-view patchmatch stereo. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14194-14203, 2021. 2", + "[32] Peng Wang, Lingjie Liu, Yuan Liu, Christian Theobalt, Taku Komura, and Wenping Wang. Neus: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. arXiv preprint arXiv:2106.10689, 2021. 2, 5", + "[33] Zhengyi Wang, Cheng Lu, Yikai Wang, Fan Bao, Chongxuan Li, Hang Su, and Jun Zhu. Prolificdreamer: High-fidelity and diverse text-to-3d generation with variational score distillation. arXiv preprint arXiv:2305.16213, 2023. 1, 3, 5", + "[34] Qingshan Xu and Wenbing Tao. Multi-scale geometric consistency guided multi-view stereo. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5483-5492, 2019. 2", + "[35] Jiayu Yang, Jose M Alvarez, and Miaomiao Liu. Nonparametric depth distribution modelling based depth inference for multi-view stereo. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8626-8634, 2022. 2", + "[36] Yao Yao, Zixin Luo, Shiwei Li, Tian Fang, and Long Quan. Mvsnet: Depth inference for unstructured multi-view stereo. In Proceedings of the European conference on computer vision (ECCV), pages 767-783, 2018.", + "[37] Yao Yao, Zixin Luo, Shiwei Li, Tianwei Shen, Tian Fang, and Long Quan. Recurrent mvsnet for high-resolution multi-view stereo depth inference. In Proceedings of" + ], + "bbox": [ + 78, + 90, + 470, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "the IEEE/CVF conference on computer vision and pattern recognition, pages 5525-5534, 2019. 2", + "[38] Yao Yao, Zixin Luo, Shiwei Li, Jingyang Zhang, Yufan Ren, Lei Zhou, Tian Fang, and Long Quan. Blendedmvs: A large-scale dataset for generalized multi-view stereo networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1790-1799, 2020. 2", + "[39] Lior Yariv, Jiatao Gu, Yoni Kasten, and Yaron Lipman. Volume rendering of neural implicit surfaces. Advances in Neural Information Processing Systems, 34:4805-4815, 2021. 2, 5", + "[40] Kai Zhang, Gernot Riegler, Noah Snavely, and Vladlen Koltun. Nerf++: Analyzing and improving neural radiance fields. arXiv preprint arXiv:2010.07492, 2020. 2", + "[41] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 3836-3847, 2023. 3", + "[42] Enliang Zheng, Enrique Dunn, Vladimir Jojic, and Jan-Michael Frahm. Patchmatch based joint view selection and depthmap estimation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 1510-1517, 2014. 2" + ], + "bbox": [ + 501, + 92, + 890, + 415 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "4958", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2024/EfficientDreamer_ High-Fidelity and Robust 3D Creation via Orthogonal-view Diffusion Priors/6a261e68-4855-488d-9a23-298bfd3be3c9_model.json b/2024/EfficientDreamer_ High-Fidelity and Robust 3D Creation via Orthogonal-view Diffusion Priors/6a261e68-4855-488d-9a23-298bfd3be3c9_model.json new file mode 100644 index 0000000000000000000000000000000000000000..4af740c52ae637defd7e4b75ec25a176a54993e5 --- /dev/null +++ b/2024/EfficientDreamer_ High-Fidelity and Robust 3D Creation via Orthogonal-view Diffusion Priors/6a261e68-4855-488d-9a23-298bfd3be3c9_model.json @@ -0,0 +1,1826 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.202, + 0.131, + 0.773, + 0.177 + ], + "angle": 0, + "content": "EfficientDreamer: High-Fidelity and Stable 3D Creation via Orthogonal-view Diffusion Priors" + }, + { + "type": "text", + "bbox": [ + 0.183, + 0.203, + 0.786, + 0.239 + ], + "angle": 0, + "content": "Zhipeng Hu\\(^{1\\dagger}\\), Minda Zhao\\(^{1\\dagger}\\), Chaoyi Zhao\\(^{1}\\), Xinyue Liang\\(^{1}\\), Lincheng Li\\(^{1*}\\), Zeng Zhao\\(^{1}\\), Changjie Fan\\(^{1}\\), Xiaowei Zhou\\(^{2}\\), Xin Yu\\(^{3}\\)" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.24, + 0.882, + 0.258 + ], + "angle": 0, + "content": "\\(^{1}\\)NetEase Fuxi AI Lab, \\(^{2}\\)State Key Lab of CAD&CG, Zhejiang University \\(^{3}\\)University of Queensland" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.292, + 0.314, + 0.308 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.324, + 0.474, + 0.643 + ], + "angle": 0, + "content": "While image diffusion models have made significant progress in text-driven 3D content creation, they often fail to accurately capture the intended meaning of text prompts, especially for view information. This limitation leads to the Janus problem, where multi-faced 3D models are generated under the guidance of such diffusion models. In this paper, we propose a robust high-quality 3D content generation pipeline by exploiting orthogonal-view image guidance. First, we introduce a novel 2D diffusion model that generates an image consisting of four orthogonal-view subimages based on the given text prompt. Then, the 3D content is created using this diffusion model. Notably, the generated orthogonal-view image provides strong geometric structure priors and thus improves 3D consistency. As a result, it effectively resolves the Janus problem and significantly enhances the quality of 3D content creation. Additionally, we present a 3D synthesis fusion network that can further improve the details of the generated 3D contents. Both quantitative and qualitative evaluations demonstrate that our method surpasses previous text-to-3D techniques. Project page: https://efficientdreamer.github.io." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.66, + 0.21, + 0.675 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.684, + 0.47, + 0.85 + ], + "angle": 0, + "content": "The field of generating photorealistic 2D images from simple text prompts is experiencing rapid growth, and recent advancements are largely attributed to the utilization of diffusion models. These advancements have led to the creation of models capable of producing images of exceptional quality [23-25]. In contrast, generating high-quality 3D content from text prompts poses significantly greater challenges due to the much larger output space and the requirement for 3D consistency. Moreover, the scarcity of adequately large training pairs containing both text and 3D models further handicaps the development of effective models." + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.851, + 0.47, + 0.867 + ], + "angle": 0, + "content": "Early methods [9, 12, 19] explore zero-shot text-guided" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.291, + 0.892, + 0.511 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.522, + 0.895, + 0.634 + ], + "angle": 0, + "content": "Figure 1. The given prompt is A pig wearing a backpack, high quality. Left: The pre-trained Stable Diffusion model struggles to accurately generate images based on specific view instructions, thereby encountering the Janus problem in the text-driven 3D generation. Right: EfficientDreamer leverages our newly introduced orthogonal-view diffusion model, enabling the generation of 3D consistent images depicting the same scene from multiple orthogonal viewpoints." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.655, + 0.893, + 0.868 + ], + "angle": 0, + "content": "3D content creation by utilizing guidance from CLIP [22]. These approaches optimize the underlying 3D representations, such as NeRFs and meshes, to achieve high text-image alignment scores for all 2D renderings. However, these methods tend to generate 3D shapes with poor geometry and unsatisfactory appearance. To address these limitations, DreamFusion [21] showcases impressive capabilities in text-to-3D synthesis by leveraging a powerful pretrained text-to-image diffusion model [24] as a strong image prior. A NeRF is trained to represent 3D models using novel Score Distillation Sampling (SDS) gradients [21] and view-dependent prompts. Subsequent works, such as TextMesh[30], Magic3D [14], and ProlificDreamer [33], further improve the quality of the generated 3D contents." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.894, + 0.902 + ], + "angle": 0, + "content": "While previous methods have demonstrated impressive results, a significant drawback is the Janus problem, which" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.875, + 0.204, + 0.888 + ], + "angle": 0, + "content": "\\(\\dagger\\) Equal contribution" + }, + { + "type": "page_footnote", + "bbox": [ + 0.098, + 0.888, + 0.22, + 0.9 + ], + "angle": 0, + "content": "* Corresponding author" + }, + { + "type": "list", + "bbox": [ + 0.096, + 0.875, + 0.22, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4949" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.244 + ], + "angle": 0, + "content": "remains a major challenge. According to recent studies [1, 16], the primary cause of the Janus problem lies in the inability of pre-trained 2D diffusion models to accurately interpret the view instructions specified in the text prompts. Even though explicit view instructions are given, the diffusion model may still generate images that do not align with these instructions. Since the stability of the 3D content creation process relies heavily on accurate guidance from the diffusion model, imprecise guidance can introduce instability and compromise the quality of the generated content." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.246, + 0.473, + 0.532 + ], + "angle": 0, + "content": "To cope with these weaknesses, we propose EfficientDreamer, which leverages a viewpoint-aware diffusion prior to create high-quality and consistent 3D contents. Our approach introduces a novel image diffusion model called orthogonal-view diffusion model, which can generate a composite image comprising four sub-images that are mutually consistent. These sub-images, within a composite image, can be derived from arbitrary viewpoints while adhering to orthogonal views. To achieve this, we maintain the architecture design of the 2D image diffusion model while adapting it to orthogonal-view image generation. We render the 3D models in a large-scale open-source 3D dataset called Objaverse [4], from orthogonal views, and assemble them to fine-tune the pre-trained 2D diffusion model. This process ensures the geometric consistency of our orthogonal-view diffusion model while preserving the capacity of 2D diffusion models to generate detailed and diverse high-quality objects that may not be present in the 3D datasets." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.536, + 0.473, + 0.779 + ], + "angle": 0, + "content": "By leveraging our innovative orthogonal-view image diffusion model, we can effectively tackle the challenge of the multi-face Janus problem while generating consistent 3D models. Specifically, during each training step, we render the 3D representation (like Neus [32] or DMTet [28]) from different orthogonal viewpoints in a differentiable manner. These rendered images are then assembled into a composite image, which is further utilized to optimize the 3D representation via orthogonal-view score distillation. While our orthogonal-view diffusion model, trained with the Objaverse 3D dataset, provides crucial geometric structure priors for 3D content creation, it is important to note that the scale of the 3D dataset is considerably smaller compared to the text-image training pairs. Additionally, the 3D assets in this dataset typically consist of common objects from the real world, having relatively simple topology and texture." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.781, + 0.473, + 0.903 + ], + "angle": 0, + "content": "Relying solely on such an orthogonal-view diffusion prior may lead to potential artifacts in the local appearance of the 3D contents. To address this issue, we also incorporate the original pre-trained 2D diffusion prior in the optimization of the 3D representation. In particular, we introduce a 3D synthesis fusion network which integrates the orthogonal-view diffusion model with a pre-trained 2D diffusion model. We also present a dynamic 3D synthesis strat" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.258 + ], + "angle": 0, + "content": "egy to balance the guidance of the orthogonal-view diffusion prior and pre-trained 2D diffusion prior during the SDS optimization process. Initially, our focus is on generating 3D consistent geometric structures, where the high weight assigned to the orthogonal-view diffusion prior helps efficiently represent the geometry and alleviate the Janus problem. As the 3D objects become more complete, we gradually reduce the weight of orthogonal-view diffusion prior and increase the guidance from the original 2D diffusion prior. This approach allows for the generation of local texture details and refinement of local appearance." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.26, + 0.875, + 0.273 + ], + "angle": 0, + "content": "Overall, our contributions are summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.275, + 0.891, + 0.35 + ], + "angle": 0, + "content": "- We propose EfficientDreamer, a novel approach for high-fidelity and robust 3D creation using orthogonal-view diffusion prior. This method effectively tackles the Janus problem and significantly enhances the stability of the generation process." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.351, + 0.892, + 0.439 + ], + "angle": 0, + "content": "- We introduce a novel image diffusion model that generates a composite image composed of four sub-images from orthogonal viewpoints. These orthogonal-view images serve as essential geometric structure priors, enabling the generation of 3D consistent geometric structures." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.442, + 0.892, + 0.516 + ], + "angle": 0, + "content": "- We present a novel text-driven 3D creation pipeline based on the orthogonal-view SDS loss. We introduce a 3D synthesis fusion network that combines the newly introduced orthogonal-view diffusion prior with the original 2D diffusion prior, aiming to ensure high-quality 3D creations." + }, + { + "type": "list", + "bbox": [ + 0.5, + 0.275, + 0.892, + 0.516 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.534, + 0.642, + 0.55 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.56, + 0.825, + 0.575 + ], + "angle": 0, + "content": "2.1. 3D Reconstruction with Neural Fields" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.584, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Traditional Multi-View Stereo (MVS) methods reconstruct 3D point clouds from predicted depth maps [6, 26, 34, 42]. Benefiting from deep learning, supervised MVS methods [7, 31, 35-37] have shown impressive performance on benchmarks but require training on specific datasets [10, 38]. These methods often struggle in less constrained scenarios and cannot easily integrate with other learning-based systems. On the other hand, in recent times, there have been significant advancements in neural fields that have demonstrated remarkable results across various tasks. As a famous neural volume rendering method, NeRF [5, 18, 20, 40] combines classical volumetric rendering with implicit function to render high-quality 2D images. However, directly extracting a mesh from a NeRF representation is non-trivial [29]. Recent works have focused on improving the geometric network while establishing connections between density-based and surface-based representations. VolSDF [39] models the volume density as Laplace Cumulative Distribution Function applied to an SDF representation. Similar to VolSDF, NeuS [32] transforms the SDF field to the accumulated transmittance for volume rendering with the" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.957 + ], + "angle": 0, + "content": "4950" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.471, + 0.166 + ], + "angle": 0, + "content": "Logistic Cumulative Distribution Function and it is emphasized that the weight of the rendering should peak at the first intersection point from the outside to the inside. To this end, we adopt the NeuS representation for the convenience of mesh extraction." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.178, + 0.312, + 0.194 + ], + "angle": 0, + "content": "2.2. Text-to-image Generation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.202, + 0.471, + 0.457 + ], + "angle": 0, + "content": "In recent years, with the availability of extremely large datasets of image-text pairs [27], significant progress has been made in text-to-image generation with diffusion models. A typical method is the Stable Diffusion model, which samples from a lower-resolution latent space and decodes latent into high-resolution images [24]. Such sampling operation in the latent greatly improves the efficiency of image generation. Another kind of method utilizes a cascade of super-resolution models [2, 25]. These methods first generate a low-resolution image from a given text prompt and then enlarge the images with several super-resolution models. Recent works design special neural network structures to control diffusion models by adding extra conditions [41], it can learn task-specific input conditions and produce customized images. Due to the strong generative ability of pretrained diffusion models, they are convenient to enable text-to-3D mesh synthesis." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.469, + 0.288, + 0.484 + ], + "angle": 0, + "content": "2.3. Text-to-3D Generation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.493, + 0.47, + 0.689 + ], + "angle": 0, + "content": "In recent years, text-to-3D has received significant attention due to the desire to create high-quality 3D content from simple semantics, such as text prompts. Early works attempt to use a CLIP objective to supervise the generation. CLIPMesh [19] deforms a 3D sphere using a CLIP loss to obtain a 3D mesh that fits the input prompt. Text2Mesh [17] stylizes a 3D mesh by predicting color and local geometric details which conform to a target text prompt. Dreamfield [9] optimizes a NeRF from many camera views so that rendered images score highly with a target caption according to a pre-trained CLIP model. However, these methods tend to generate poor 3D shapes with unsatisfactory geometry as well as appearance." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.689, + 0.471, + 0.901 + ], + "angle": 0, + "content": "To overcome such limitations, DreamFusion [21] employs text-to-3D synthesis by utilizing a powerful pretrained text-to-image diffusion model as a strong image prior. They optimize NeRF with their proposed Score Distillation Sampling (SDS) loss. Magic3D [14] extends such a method to a two-stage coarse-to-fine process, which utilizes DMTet for mesh representation. Fantasy3D [3] disentangles geometry and appearance into a two-stage optimization and introduces the spatially varying Bidirectional Reflectance Distribution Function (BRDF) into the text-to-3D task. TextMesh [30] replaces NeRF with VolSDF for more accurate mesh expression. It proposes a novel multi-view consistent and mesh-conditioned re-texturing, enabling the generation of a photorealistic 3D mesh model." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.151 + ], + "angle": 0, + "content": "More recently, ProlificDreamer [33] proposes Variational Score Distillation (VSD), a principled particle-based variational framework to promote the quality of created 3D models." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.165, + 0.637, + 0.183 + ], + "angle": 0, + "content": "3. Methodology" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.189, + 0.892, + 0.326 + ], + "angle": 0, + "content": "Given a text prompt \\( y \\), our goal is to create high-quality 3D mesh representations with photorealistic texture. To achieve this, we propose a novel orthogonal-view diffusion model, denoted as \\( \\phi_{ov} \\), which generates composite images consisting of four orthogonal-view sub-images. Subsequently, we design a 3D synthesis fusion network that combines the orthogonal-view diffusion model with a pre-trained diffusion model, denoted as \\( \\phi_{pre} \\), to enhance the quality of the generated 3D content." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.334, + 0.796, + 0.35 + ], + "angle": 0, + "content": "3.1. Orthogonal-view Diffusion Model" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.358, + 0.892, + 0.418 + ], + "angle": 0, + "content": "Our objective is to generate a composite image in response to a given text prompt \\( y \\) using a fine-tuned diffusion model. The composite image will consist of four sub-images from orthogonal viewpoints, arranged in a \\( 2 \\times 2 \\) grid." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.419, + 0.892, + 0.584 + ], + "angle": 0, + "content": "Common text-to-image diffusion models are trained on large-scale text-image pairs, enabling them to describe objects from various viewpoints. However, these models do not explicitly align corresponding parts of the object across viewpoints. Moreover, they tend to generate images of objects in canonical poses. Hence, it is essential to introduce a novel diffusion model that can generate representations of the same object from diverse viewpoints while maintaining 3D consistency. In this paper, we propose that showcasing the object from four orthogonal views is suitable as it provides essential geometric structure priors." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.585, + 0.892, + 0.901 + ], + "angle": 0, + "content": "To obtain such a specific image diffusion model, we use the recently released Objaverse [4] dataset for fine-tuning the commonly-used diffusion model, i.e., the Stable Diffusion model [24]. This dataset contains \\(800\\mathrm{K} + 3\\mathrm{D}\\) models created by \\(100\\mathrm{K}+\\) artists. The descriptions for each 3D model can be found in the Cap3D [15] dataset, which is processed with BLIP2 [13] and GPT4. We first filter some improper 3D models, like complex scene representation, textureless 3D models, and point clouds, and approximately \\(420\\mathrm{K}\\) high-quality 3D models are left. We normalize all assets into a unit cube \\([-0.5, 0.5]^3\\). The radius (distance away from the center) is set at 1.8, with a field of view set at 35 degrees. For each 3D model, we uniformly render 12 images in the range of azimuth angle \\(\\xi_{cam} \\in (0^\\circ, 360^\\circ)\\) and elevation angle list \\(\\delta_{cam} = [0^\\circ, 15^\\circ, 30^\\circ, 45^\\circ]\\). Thus, for each 3D asset, we render 48 RGBA images in \\(512 \\times 512\\) resolution in total while its corresponding text prompt is denoted as \\(y\\). We use the Stable Diffusion 2.1 as the original latent diffusion architecture with an encoder \\(\\mathcal{E}\\), a denoiser U-Net \\(\\epsilon_{\\theta}\\), and a decoder \\(\\mathcal{D}\\). We utilize \\(\\mathcal{E}\\) and \\(\\mathcal{D}\\) with pre-trained weight and only fine-tune \\(\\epsilon_{\\theta}\\) for generating composite im" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "4951" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.086, + 0.088, + 0.891, + 0.416 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.425, + 0.896, + 0.482 + ], + "angle": 0, + "content": "Figure 2. The overview of EfficientDreamer involves two key steps. Firstly, we train an orthogonal-view diffusion model by rendering images from the Objaverse dataset. Secondly, we optimize the 3D scene representation by leveraging both the newly introduced orthogonal-view diffusion model and the pre-trained text-to-image diffusion model. To ensure high-fidelity and robust 3D creation, we employ a dynamic 3D synthesis strategy." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.495, + 0.473, + 0.678 + ], + "angle": 0, + "content": "ages from orthogonal views. As illustrated in Fig. 2(a), for each training step, we first sample an elevation angle from the elevation angle list and then select four render images in this elevation angle, the rendered images are required to be 90 degrees apart from each other in azimuth angle. Then these four images are tiled on a \\(2 \\times 2\\) grid in a clockwise rotation. To match the diffusion model, we resize this composite image into \\(512 \\times 512\\) resolution, which is denoted as \\(x\\). At the diffusion time step \\(t \\sim [1, 1000]\\), let \\(c(y)\\) be the embedding of \\(y\\), and \\(z_t\\) be the noisy latent image by adding noise \\(\\epsilon\\) to a clean latent image \\(z\\), we then solve for the following objective to fine-tune the model:" + }, + { + "type": "equation", + "bbox": [ + 0.12, + 0.686, + 0.47, + 0.711 + ], + "angle": 0, + "content": "\\[\n\\underset {\\theta} {\\min } \\mathbb {E} _ {z \\sim \\mathcal {E} (x), t, \\epsilon \\sim \\mathcal {N} (0, 1)} | | \\epsilon - \\epsilon_ {\\theta} \\left(z _ {t}, t, c (y)\\right) | | ^ {2}. \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.718, + 0.471, + 0.81 + ], + "angle": 0, + "content": "We find the trained diffusion model is capable of learning a generic mechanism to explicitly align the correspondence of an identical object from different viewpoints, even for objects never appearing in the training dataset. The comparison between the orthogonal-view diffusion model and the pre-trained diffusion model is depicted in Fig. 3." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.818, + 0.471, + 0.835 + ], + "angle": 0, + "content": "3.2. Text-to-3D via the 3D Synthesis Fusion Network" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Once the orthogonal-view diffusion model is trained, we can generate our initial 3D model by training a neural distance field using the score distillation sampling strategy. The scene is represented as a differential render \\( g(\\psi) \\), where" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.495, + 0.895, + 0.768 + ], + "angle": 0, + "content": "\\(\\psi\\) denotes the learnable parameters. Given a randomly sampled camera pose with azimuth angle \\(\\xi_{cam}\\), elevation angle \\(\\delta_{cam}\\) and radius \\(r\\), we fix \\(\\delta_{cam}\\) and \\(r\\) and sample four camera poses by extending \\(\\xi_{cam}\\) to \\([\\xi_{cam}, \\xi_{cam} + 90^{\\circ}, \\xi_{cam} + 180^{\\circ}, \\xi_{cam} + 270^{\\circ}]\\). Then we render images with these camera poses, which are denoted as \\(I_0\\), \\(I_1\\), \\(I_2\\), \\(I_3\\). We also tile them on a \\(2 \\times 2\\) grid in a clockwise rotation and resize the composite image into \\(512 \\times 512\\) resolution, which is denoted as \\(\\hat{I}\\). Then we sample random normal noise according to time step \\(t\\) and add it to \\(\\hat{I}\\). The noisy images \\(\\tilde{I}_t\\), together with text embedding \\(c(y)\\) are fed to our orthogonal-view diffusion model \\(\\phi_{ov}\\), which attempts to predict the noise \\(\\epsilon\\). The denoiser U-Net of \\(\\phi_{ov}\\) is denoted as \\(\\epsilon_{\\phi_{ov}}\\) and the score function of noise estimation is denoted as \\(\\epsilon_{\\phi_{ov}}(\\tilde{I}_t; t, c(y))\\). This score function guides the direction of the gradient for updating the scene parameters, and the gradient is calculated by orthogonal-view Score Distillation Sampling (SDS):" + }, + { + "type": "equation", + "bbox": [ + 0.522, + 0.779, + 0.892, + 0.814 + ], + "angle": 0, + "content": "\\[\n\\nabla_ {\\psi} \\mathcal {L} _ {S D S} ^ {o v} = \\mathbb {E} _ {t, \\epsilon} [ \\omega (t) (\\epsilon_ {\\phi_ {o v}} (\\tilde {I} _ {t}; t, c (y)) - \\epsilon) \\frac {\\partial \\hat {I}}{\\partial \\psi} ], \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.825, + 0.776, + 0.84 + ], + "angle": 0, + "content": "where \\(\\omega (t)\\) denotes a weighting function." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.894, + 0.901 + ], + "angle": 0, + "content": "Motivated by [8] and to enhance the stability of the training procedure, we decrease the time step (t) as training progresses, resulting in a gradual decline in the degree of added noise. As our training procedure supervises the 3D creation" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4952" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.084, + 0.092, + 0.468, + 0.427 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.435, + 0.47, + 0.52 + ], + "angle": 0, + "content": "Figure 3. Comparison between The orthogonal-view diffusion model and the pre-trained Stable Diffusion model with additional viewpoint instructions. The pre-trained Stable Diffusion model struggles to generate images based on specific view instructions, while the orthogonal-view diffusion model can generate composite images from orthogonal views." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.565, + 0.47, + 0.626 + ], + "angle": 0, + "content": "from four orthogonal views at each time step, the neural distance field can receive semantic structure guidance from different viewpoints. This fundamentally eliminates the possibility of creating multi-face problems." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.635, + 0.47, + 0.773 + ], + "angle": 0, + "content": "Our orthogonal-view diffusion model, trained on the Objaverse 3D dataset, provides essential geometric structure priors for 3D creation. However, it is important to note that the scale of the 3D dataset is relatively small compared to the text-image training pairs. Additionally, the included 3D assets mainly consist of simple objects with uncomplicated topology and texture. Therefore, relying solely on the orthogonal-view diffusion prior may introduce potential artifacts in the local appearance of the generated 3D contents." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.781, + 0.471, + 0.903 + ], + "angle": 0, + "content": "To resolve such a problem, we introduce a 3D synthesis fusion network which integrates the proposed orthogonal-view model with the original text-to-image diffusion model to guide the 3D model generation with a 3D synthesis dynamic strategy. As shown in Fig. 2(b), let \\(\\epsilon_{\\phi_{pre}}\\) denote the denoiser U-Net of the pre-trained Stable Diffusion model, and rearrange the original rendered images \\(I_0, I_1, I_2, I_3\\) as an image batch \\(\\hat{\\mathbb{I}}\\). \\(\\tilde{\\mathbb{I}}_t\\) denotes the noisy result of \\(\\hat{\\mathbb{I}}\\). The" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.775, + 0.108 + ], + "angle": 0, + "content": "corresponding SDS loss is formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.521, + 0.117, + 0.892, + 0.151 + ], + "angle": 0, + "content": "\\[\n\\nabla_ {\\psi} \\mathcal {L} _ {S D S} ^ {p r e} = \\mathbb {E} _ {t, \\epsilon} [ \\omega (t) (\\epsilon_ {\\phi_ {p r e}} (\\tilde {\\mathbb {I}} _ {t}; t, c (y)) - \\epsilon) \\frac {\\partial \\hat {\\mathbb {I}}}{\\partial \\psi} ]. \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.161, + 0.892, + 0.191 + ], + "angle": 0, + "content": "We propose a new prior loss for the novel view supervision to combine both two priors:" + }, + { + "type": "equation", + "bbox": [ + 0.509, + 0.201, + 0.892, + 0.232 + ], + "angle": 0, + "content": "\\[\n\\nabla_ {\\psi} \\mathcal {L} _ {S D S} = (1 - (\\frac {l}{L}) ^ {\\lambda}) \\nabla_ {\\psi} \\mathcal {L} _ {S D S} ^ {o v} + (\\frac {l}{L}) ^ {\\lambda} \\nabla_ {\\psi} \\mathcal {L} _ {S D S} ^ {p r e}, (4)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.239, + 0.892, + 0.284 + ], + "angle": 0, + "content": "where \\(l\\) is the current iteration while \\(L\\) is the total iterations. \\(\\lambda\\) is a hyperparameter to determine the strength of two diffusion priors, which is set as 1 in our experiments." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.285, + 0.892, + 0.376 + ], + "angle": 0, + "content": "Initially, we prioritize generating 3D content with consistent geometric structures by emphasizing the orthogonal-view diffusion prior. As the generation progresses, we gradually shift the focus towards incorporating guidance from the original 2D diffusion prior to enhance local texture details and refine the overall appearance." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.384, + 0.719, + 0.401 + ], + "angle": 0, + "content": "3.3. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.407, + 0.892, + 0.497 + ], + "angle": 0, + "content": "Orthogonal-view Diffusion Model Training: Our model is implemented on the PyTorch platform with 8 NVIDIA Tesla A100 GPUs. We train the model utilizing AdamW optimizer with a learning rate of 1e-5. The whole training procedure takes 200K iterations with a batch size of 256, which takes about 5 days." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.498, + 0.893, + 0.71 + ], + "angle": 0, + "content": "Text-to-3D Generation: We employ a two-stage coarse-to-fine optimization process to generate high-quality 3D mesh representations with photorealistic textures from text prompts. In the coarse stage, we utilize NeuS [32] as the 3D scene representation, as density-based representations such as NeRF are not well-suited for extracting 3D geometry and obtaining meshes [39]. The rendering resolution is set at \\(64 \\times 64\\), and the 3D representation is optimized for 5000 steps with 3D Synthesis Fusion Network. In the refinement stage, we employ DMTet [28] as the scene representation, enabling efficient rendering of textured meshes with differentiable rasterization at \\(512 \\times 512\\) resolutions. Additionally, we utilize SDS and VSD guidance [33] to optimize the geometry and texture for another 5000 steps." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.723, + 0.634, + 0.74 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.748, + 0.892, + 0.84 + ], + "angle": 0, + "content": "In this section, we conduct extensive experiments to evaluate the effectiveness of our proposed method for text-to-3D content creation. We present qualitative and quantitative results, compare our method with state-of-the-art methods, perform a user study, and analyze the effectiveness of our proposed techniques through an ablation study." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.848, + 0.803, + 0.865 + ], + "angle": 0, + "content": "4.1. Comparison with State-of-the-arts" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.892, + 0.901 + ], + "angle": 0, + "content": "We present comprehensive experiments to evaluate the effectiveness of our proposed method for text-to-3D content" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4953" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.083, + 0.091, + 0.885, + 0.421 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.432, + 0.893, + 0.462 + ], + "angle": 0, + "content": "Figure 4. Comparison with other text-to-3D methods. We render each 3D model from two views. Our method outperforms other techniques by generating more high-fidelity 3D models without encountering the Janus problem." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.474, + 0.473, + 0.732 + ], + "angle": 0, + "content": "creation. The compared methods include three crucial Text-to-3D methods, i.e., DreamFusion [21], Magic3D [14] and TextMesh [30]. All these methods are implemented with a unified framework for 3D content creation*. As DreamFusion and TextMesh generate 3D contents with volume rendering of \\(64 \\times 64\\) resolution, the generated 3D models may suffer from over-smoothness and lack of details. However, Magic3D can generate higher-quality 3D shapes with improved geometry and texture. Since our method also utilizes a coarse-to-fine two-stage optimization, we are able to create high-quality 3D mesh models. Furthermore, all of the compared methods may be confronted with the Janus problem as they rely on open-source pre-trained diffusion priors that do not enforce valid 3D consistency constraints. Our method can resolve the multi-face problem with our newly introduced orthogonal-view diffusion prior. Please refer to the supplementary material for more comparison results." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.734, + 0.47, + 0.87 + ], + "angle": 0, + "content": "We also evaluate our method using the CLIP score [22] and FID [11] metrics. The CLIP score measures how well images rendered from the generated objects correlate with the provided input text prompt, while FID evaluates the quality and photorealistic appearance of the generated shapes. We render four images with fixed camera poses for each 3D model. The results are depicted in Table 1, which demonstrates that our methods achieve better performance than the state-of-the-art methods." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.473, + 0.893, + 0.515 + ], + "angle": 0, + "content": "Table 1. Comparison with state-of-the-arts. Comparing our method against the state-of-the-arts using CLIP score and FID. \\( \\downarrow \\) indicates the lower the better, and \\( \\uparrow \\) indicates the higher the better." + }, + { + "type": "table", + "bbox": [ + 0.566, + 0.525, + 0.819, + 0.613 + ], + "angle": 0, + "content": "
MethodsCLIP ↑FID ↓
DreamFusion [21]28.40374.44
Magic3D [14]29.15310.57
TextMesh [30]27.65305.77
Ours30.33284.98
" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.632, + 0.747, + 0.648 + ], + "angle": 0, + "content": "4.2. Comparison with Perp-Neg" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.655, + 0.892, + 0.822 + ], + "angle": 0, + "content": "In addition, we conduct a comparison between our method and another approach, Perp-Neg [1], which is designed to mitigate the Janus problem. We integrate Perp-Neg into DreamFusion and evaluate its performance against our method. The results, shown in Figure 5, reveal that Perp-Neg can address the Janus problem to some extent. For instance, it is capable of generating 3D models with only one head and face. However, these results may exhibit unreasonable details such as three ears or five legs. In contrast, our method, which incorporates guidance from orthogonal views simultaneously, enables us to avoid such issues." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.832, + 0.892, + 0.849 + ], + "angle": 0, + "content": "4.3. Coarse-to-fine Two-stage Optimization Results" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.893, + 0.901 + ], + "angle": 0, + "content": "In Figure 6, we demonstrate the 3D contents generated through the coarse and fine stages of our optimization procedure. In the coarse stage, we achieve precise 3D represent" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.887, + 0.365, + 0.901 + ], + "angle": 0, + "content": "* https://github.com/threestudio-project/threestudio" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4954" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.078, + 0.089, + 0.473, + 0.282 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.29, + 0.471, + 0.334 + ], + "angle": 0, + "content": "Figure 5. Comparison between the Perp-Neg method and our approach: Our method offers a more comprehensive solution to the Janus problem compared to Perp-Neg." + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.368, + 0.473, + 0.637 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.642, + 0.471, + 0.698 + ], + "angle": 0, + "content": "Figure 6. Coarse-to-fine two-stage optimization results. In the coarse stage, we get semantically reasonable 3D scene representation. In the fine stage, we improve the visual quality of generated 3D assets." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.735, + 0.473, + 0.903 + ], + "angle": 0, + "content": "tations for the given text prompts while effectively addressing the challenging Janus problem. This is accomplished with the guidance of the orthogonal-view diffusion model, which provides essential geometric structure priors for 3D creation. In the fine stage, we initialize the mesh representation with the results obtained from the coarse stage and proceed to optimize the geometry using SDS guidance. Additionally, we leverage VSD guidance to enhance the texture, ensuring that the generated 3D model is visually appealing and realistic. As a result, this process enables the generation of significantly higher-quality 3D models." + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.089, + 0.895, + 0.148 + ], + "angle": 0, + "content": "Table 2. Results of our user study. The rating score represents the average rating results for different methods, while the preference indicates the ratio of user preference for the results of each method with higher quality." + }, + { + "type": "table", + "bbox": [ + 0.522, + 0.156, + 0.86, + 0.243 + ], + "angle": 0, + "content": "
MethodsRating Score ↑Preference(%) ↑
DreamFusion [21]2.094.55
Magic3D [14]2.446.27
TextMesh [30]1.734.91
Ours3.7484.27
" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.257, + 0.623, + 0.274 + ], + "angle": 0, + "content": "4.4. User Study" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.281, + 0.895, + 0.538 + ], + "angle": 0, + "content": "We conduct a user study with Fuxi Youling Crowdsourcing \\* for quantifying the subjective quality. Concretely, we evaluate the study on 22 prompts. At each evaluation time, the participants are shown the created 3D models by DreamFusion [21], Magic3D [14], TextMesh[30] and our method, and asked to rate them on a scale of 1 to 4. In total, we collect 1100 responses from 50 participants. The average rating scores for the different methods are shown in Table 2. Notably, our method receives the highest average rating, indicating that it produces the most visually high-fidelity 3D models. Furthermore, we calculate the preference ratios for each method based on the quality of the results. The analysis reveals that \\(84.27\\%\\) of the users consider the results of our method to be of higher quality than those generated by all competing methods. This finding further validates the effectiveness of our method in generating visually highfidelity 3D models." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.549, + 0.655, + 0.566 + ], + "angle": 0, + "content": "4.5. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.573, + 0.895, + 0.877 + ], + "angle": 0, + "content": "Impact of the 3D synthesis Fusion Network: We introduce a 3D synthesis fusion network with the dynamic 3D synthesis strategy. Here we provide analysis and experiments to show its effectiveness. Given several text prompts, we employ 3D creation with different diffusion model guidance. The corresponding results are shown in Fig. 7. The first two columns show the results generated without the orthogonal-view diffusion model, which is equivalent to \\(\\lambda = 0\\). Obvious Janus problems and inaccurate 3D representation can be found for the first two prompts. As for the creation of a cottage, we get similar shapes from different viewpoints. The next two columns demonstrate the result guided without the pre-trained diffusion model, which is equivalent to \\(\\lambda = \\infty\\). While the orthogonal-view diffusion model can provide comprehensive structure constraints for 3D generation and restrain multi-face issues, we can see local distortion and unreasonable roughness. The last two columns demonstrate the results with our 3D synthesis fusion network, which means \\(\\lambda = 1\\). We get complete and smooth meshes and eliminate the possible Janus problem." + }, + { + "type": "page_footnote", + "bbox": [ + 0.518, + 0.887, + 0.707, + 0.901 + ], + "angle": 0, + "content": "* https://fuxi.163.com/solution/data" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.958 + ], + "angle": 0, + "content": "4955" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.084, + 0.089, + 0.462, + 0.305 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.084, + 0.307, + 0.462, + 0.402 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.405, + 0.47, + 0.432 + ], + "angle": 0, + "content": "Figure 7. Ablation study on the impact of the 3D synthesis fusion network." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.448, + 0.473, + 0.901 + ], + "angle": 0, + "content": "Quantitative results are depicted in Table 3, which demonstrate the effectiveness of the 3D synthesis fusion network. Impact of Varying Number of View Supervision: Here we present an ablation study on the effects of varying the number of view supervisions on our orthogonal-view diffusion model. For two-view supervision, we randomly mask the opposite viewpoints in the rendered images during training, resulting in the optimization of the 3D representation using only two views. In the case of three-view supervision, we randomly mask one viewpoint, allowing the optimization of the 3D representation using three views. The corresponding results are shown in Figure 8. Generally, supervising the 3D creation with our orthogonal-view diffusion model can mitigate Janus problems. However, creating 3D models with two-view supervision may still suffer from slight multi-face problems, such as two barrels in a tank or multiple faces in a squirrel and a horse. With three-view supervision, these problems are alleviated to some extent. Nevertheless, there are still local distortions in generated 3D models, such as unexpected sags or bulges. These issues can be addressed by using the intact four-view supervision of our orthogonal-view diffusion model. Quantitative results are also presented in Table 3, illustrating the impact of reducing the number of view supervision signals during training. It is evident that there is a decrease in performance observed in this scenario. This highlights the significance of optimizing a 3D representation using a set of four orthogonal-view images, as it provides sufficient geometric structure priors for ensuring consistent 3D content creation." + }, + { + "type": "image", + "bbox": [ + 0.513, + 0.092, + 0.887, + 0.393 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.5, + 0.398, + 0.892, + 0.426 + ], + "angle": 0, + "content": "Figure 8. Ablation study on the impact of varying number of view supervision." + }, + { + "type": "table_caption", + "bbox": [ + 0.5, + 0.451, + 0.892, + 0.48 + ], + "angle": 0, + "content": "Table 3. Ablation study results using CLIP and FID. \\( \\downarrow \\) indicates the lower the better, and \\( \\uparrow \\) indicates the higher the better." + }, + { + "type": "table", + "bbox": [ + 0.502, + 0.49, + 0.906, + 0.611 + ], + "angle": 0, + "content": "
Impact of the 3D synthesis Fusion Network
Metricsw/o Orthogonal-vieww/o Pre-trainedFull
CLIP ↑31.1332.0532.55
FID ↓372.86369.73342.62
Impact of Varying Number of View Supervision
MetricsTwo-viewThree-viewFour-view
CLIP ↑30.9230.4831.47
FID ↓399.58373.04345.72
" + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.635, + 0.619, + 0.65 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.665, + 0.895, + 0.9 + ], + "angle": 0, + "content": "In this paper, we present EfficientDreamer, a novel approach for high-fidelity and robust 3D creation using orthogonal-view diffusion priors. We introduce a novel diffusion model that can create composite images comprising sub-images from different orthogonal viewpoints. These composite images provide important geometric structure priors for 3D model generation by showcasing an object from multiple perspectives. We enhance the text-to-3D generation framework by leveraging the orthogonal-view diffusion model. Additionally, we introduce a 3D synthesis fusion network that integrates the orthogonal-view diffusion prior with the pre-trained diffusion prior. We also introduce a dynamic 3D synthesis strategy to balance the guidance of these two diffusion models during the SDS optimization. Our method effectively addresses the challenging Janus problem while ensuring the fidelity of the generated 3D models." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4956" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.184 + ], + "angle": 0, + "content": "[1] Mohammadreza Armandpour, Huangjie Zheng, Ali Sadeghian, Amir Sadeghian, and Mingyuan Zhou. Re-imagine the Negative Prompt Algorithm: Transform 2D Diffusion into 3D, alleviate Janus problem and Beyond. arXiv preprint arXiv:2304.04968, 2023. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.187, + 0.472, + 0.257 + ], + "angle": 0, + "content": "[2] Yogesh Balaji, Seungjun Nah, Xun Huang, Arash Vahdat, Jiaming Song, Karsten Kreis, Miika Aittala, Timo Aila, Samuli Laine, Bryan Catanzaro, et al. edifi: Text-to-image diffusion models with an ensemble of expert denoisers. arXiv preprint arXiv:2211.01324, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.259, + 0.472, + 0.327 + ], + "angle": 0, + "content": "[3] Rui Chen, Yongwei Chen, Ningxin Jiao, and Kui Jia. Fantasia3d: Disentangling geometry and appearance for high-quality text-to-3d content creation. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 22246–22256, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.33, + 0.472, + 0.413 + ], + "angle": 0, + "content": "[4] Matt Deitke, Dustin Schwenk, Jordi Salvador, Luca Weihs, Oscar Michel, Eli VanderBilt, Ludwig Schmidt, Kiana Ehsani, Aniruddha Kembhavi, and Ali Farhadi. Objverse: A universe of annotated 3d objects. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13142-13153, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.415, + 0.472, + 0.485 + ], + "angle": 0, + "content": "[5] Sara Fridovich-Keil, Alex Yu, Matthew Tancik, Qinhong Chen, Benjamin Recht, and Angjoo Kanazawa. Plenoxels: Radiance fields without neural networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5501–5510, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.486, + 0.472, + 0.543 + ], + "angle": 0, + "content": "[6] Silvano Galliani, Katrin Lasinger, and Konrad Schindler. Massively parallel multiview stereopsis by surface normal diffusion. In Proceedings of the IEEE International Conference on Computer Vision, pages 873-881, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.545, + 0.472, + 0.613 + ], + "angle": 0, + "content": "[7] Xiaodong Gu, Zhiwen Fan, Siyu Zhu, Zuozhuo Dai, Feitong Tan, and Ping Tan. Cascade cost volume for high-resolution multi-view stereo and stereo matching. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2495-2504, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.615, + 0.472, + 0.671 + ], + "angle": 0, + "content": "[8] Yukun Huang, Jianan Wang, Yukai Shi, Xianbiao Qi, Zheng-Jun Zha, and Lei Zhang. Dreamtime: An improved optimization strategy for text-to-3d content creation. arXiv preprint arXiv:2306.12422, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.673, + 0.472, + 0.741 + ], + "angle": 0, + "content": "[9] Ajay Jain, Ben Mildenhall, Jonathan T Barron, Pieter Abbeel, and Ben Poole. Zero-shot text-guided object generation with dream fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 867-876, 2022. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.744, + 0.47, + 0.8 + ], + "angle": 0, + "content": "[10] Rasmus Jensen, Anders Dahl, George Vogiatzis, Engin Tola, and Henrik Aanæs. Large scale multi-view stereopsis evaluation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 406-413, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.802, + 0.47, + 0.857 + ], + "angle": 0, + "content": "[11] Tuomas Kynkänniemi, Tero Karras, Miika Aittala, Timo Aila, and Jaakko Lehtinen. The role of imagenet classes in fr\\echet inception distance. arXiv preprint arXiv:2203.06026, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.859, + 0.47, + 0.901 + ], + "angle": 0, + "content": "[12] Han-Hung Lee and Angel X Chang. Understanding pure clip guidance for voxel grid nerf models. arXiv preprint arXiv:2209.15172, 2022. 1" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.148 + ], + "angle": 0, + "content": "[13] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. arXiv preprint arXiv:2301.12597, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.15, + 0.894, + 0.233 + ], + "angle": 0, + "content": "[14] Chen-Hsuan Lin, Jun Gao, Luming Tang, Towaki Takikawa, Xiaohui Zeng, Xun Huang, Karsten Kreis, Sanja Fidler, Ming-Yu Liu, and Tsung-Yi Lin. Magic3d: High-resolution text-to-3d content creation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 300–309, 2023. 1, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.235, + 0.892, + 0.276 + ], + "angle": 0, + "content": "[15] Tiange Luo, Chris Rockwell, Honglak Lee, and Justin Johnson. Scalable 3d captioning with pretrained models. arXiv preprint arXiv:2306.07279, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.278, + 0.892, + 0.348 + ], + "angle": 0, + "content": "[16] Gal Metzer, Elad Richardson, Or Patashnik, Raja Giryes, and Daniel Cohen-Or. Latent-nerf for shape-guided generation of 3d shapes and textures. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12663–12673, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.349, + 0.892, + 0.417 + ], + "angle": 0, + "content": "[17] Oscar Michel, Roi Bar-On, Richard Liu, Sagie Benaim, and Rana Hanocka. Text2mesh: Text-driven neural stylization for meshes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13492-13502, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.419, + 0.892, + 0.487 + ], + "angle": 0, + "content": "[18] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.49, + 0.892, + 0.546 + ], + "angle": 0, + "content": "[19] Nasir Mohammad Khalid, Tianhao Xie, Eugene Belilovsky, and Tiberiu Popa. Clip-mesh: Generating textured meshes from text using pretrained image-text models. In SIGGRAPH Asia 2022 conference papers, pages 1-8, 2022. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.547, + 0.892, + 0.603 + ], + "angle": 0, + "content": "[20] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM Transactions on Graphics (ToG), 41(4):1-15, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.604, + 0.892, + 0.645 + ], + "angle": 0, + "content": "[21] Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. In ICLR, 2023. 1, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.647, + 0.892, + 0.73 + ], + "angle": 0, + "content": "[22] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.732, + 0.892, + 0.786 + ], + "angle": 0, + "content": "[23] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.789, + 0.892, + 0.858 + ], + "angle": 0, + "content": "[24] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.86, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[25] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans," + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4957" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.092, + 0.47, + 0.134 + ], + "angle": 0, + "content": "et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in Neural Information Processing Systems, 35:36479-36494, 2022. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.136, + 0.471, + 0.219 + ], + "angle": 0, + "content": "[26] Johannes L Schonberger, Enliang Zheng, Jan-Michael Frahm, and Marc Pollefeys. Pixelwise view selection for unstructured multi-view stereo. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part III 14, pages 501-518. Springer, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.221, + 0.471, + 0.303 + ], + "angle": 0, + "content": "[27] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. Advances in Neural Information Processing Systems, 35:25278-25294, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.305, + 0.471, + 0.374 + ], + "angle": 0, + "content": "[28] Tianchang Shen, Jun Gao, Kangxue Yin, Ming-Yu Liu, and Sanja Fidler. Deep marching tetrahedra: a hybrid representation for high-resolution 3d shape synthesis. Advances in Neural Information Processing Systems, 34:6087-6101, 2021. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.376, + 0.471, + 0.432 + ], + "angle": 0, + "content": "[29] Jiaxiang Tang, Hang Zhou, Xiaokang Chen, Tianshu Hu, Errui Ding, Jingdong Wang, and Gang Zeng. Delicate textured mesh recovery from nerf via adaptive surface refinement. arXiv preprint arXiv:2303.02091, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.433, + 0.471, + 0.487 + ], + "angle": 0, + "content": "[30] Christina Tsalicoglou, Fabian Manhardt, Alessio Tonioni, Michael Niemeyer, and Federico Tombari. Textmesh: Generation of realistic 3d meshes from text prompts. 2024. 1, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.49, + 0.471, + 0.559 + ], + "angle": 0, + "content": "[31] Fangjinhua Wang, Silvano Galliani, Christoph Vogel, Pablo Speciale, and Marc Pollefeys. Patchmatchnet: Learned multi-view patchmatch stereo. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14194-14203, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.561, + 0.471, + 0.616 + ], + "angle": 0, + "content": "[32] Peng Wang, Lingjie Liu, Yuan Liu, Christian Theobalt, Taku Komura, and Wenping Wang. Neus: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. arXiv preprint arXiv:2106.10689, 2021. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.618, + 0.471, + 0.673 + ], + "angle": 0, + "content": "[33] Zhengyi Wang, Cheng Lu, Yikai Wang, Fan Bao, Chongxuan Li, Hang Su, and Jun Zhu. Prolificdreamer: High-fidelity and diverse text-to-3d generation with variational score distillation. arXiv preprint arXiv:2305.16213, 2023. 1, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.675, + 0.471, + 0.731 + ], + "angle": 0, + "content": "[34] Qingshan Xu and Wenbing Tao. Multi-scale geometric consistency guided multi-view stereo. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5483-5492, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.732, + 0.471, + 0.801 + ], + "angle": 0, + "content": "[35] Jiayu Yang, Jose M Alvarez, and Miaomiao Liu. Nonparametric depth distribution modelling based depth inference for multi-view stereo. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8626-8634, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.803, + 0.471, + 0.858 + ], + "angle": 0, + "content": "[36] Yao Yao, Zixin Luo, Shiwei Li, Tian Fang, and Long Quan. Mvsnet: Depth inference for unstructured multi-view stereo. In Proceedings of the European conference on computer vision (ECCV), pages 767-783, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.86, + 0.471, + 0.901 + ], + "angle": 0, + "content": "[37] Yao Yao, Zixin Luo, Shiwei Li, Tianwei Shen, Tian Fang, and Long Quan. Recurrent mvsnet for high-resolution multi-view stereo depth inference. In Proceedings of" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.471, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "the IEEE/CVF conference on computer vision and pattern recognition, pages 5525-5534, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.122, + 0.892, + 0.19 + ], + "angle": 0, + "content": "[38] Yao Yao, Zixin Luo, Shiwei Li, Jingyang Zhang, Yufan Ren, Lei Zhou, Tian Fang, and Long Quan. Blendedmvs: A large-scale dataset for generalized multi-view stereo networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1790-1799, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.192, + 0.892, + 0.246 + ], + "angle": 0, + "content": "[39] Lior Yariv, Jiatao Gu, Yoni Kasten, and Yaron Lipman. Volume rendering of neural implicit surfaces. Advances in Neural Information Processing Systems, 34:4805-4815, 2021. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.249, + 0.892, + 0.29 + ], + "angle": 0, + "content": "[40] Kai Zhang, Gernot Riegler, Noah Snavely, and Vladlen Koltun. Nerf++: Analyzing and improving neural radiance fields. arXiv preprint arXiv:2010.07492, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.292, + 0.892, + 0.347 + ], + "angle": 0, + "content": "[41] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 3836-3847, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.349, + 0.892, + 0.416 + ], + "angle": 0, + "content": "[42] Enliang Zheng, Enrique Dunn, Vladimir Jojic, and Jan-Michael Frahm. Patchmatch based joint view selection and depthmap estimation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 1510-1517, 2014. 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.416 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "4958" + } + ] +] \ No newline at end of file diff --git a/2024/EfficientDreamer_ High-Fidelity and Robust 3D Creation via Orthogonal-view Diffusion Priors/6a261e68-4855-488d-9a23-298bfd3be3c9_origin.pdf b/2024/EfficientDreamer_ High-Fidelity and Robust 3D Creation via Orthogonal-view Diffusion Priors/6a261e68-4855-488d-9a23-298bfd3be3c9_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..f7dea6bfae85711068ac2025f32a1bb2c7075df4 --- /dev/null +++ b/2024/EfficientDreamer_ High-Fidelity and Robust 3D Creation via Orthogonal-view Diffusion Priors/6a261e68-4855-488d-9a23-298bfd3be3c9_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:34429ccb09db11eeef8db2ca2dd1dc17e2e32076d6bb103b3f1b9d668885a512 +size 3408174 diff --git a/2024/EfficientDreamer_ High-Fidelity and Robust 3D Creation via Orthogonal-view Diffusion Priors/full.md b/2024/EfficientDreamer_ High-Fidelity and Robust 3D Creation via Orthogonal-view Diffusion Priors/full.md new file mode 100644 index 0000000000000000000000000000000000000000..2e6f5af434281de66dd2a14e424cc38f55e0aa3a --- /dev/null +++ b/2024/EfficientDreamer_ High-Fidelity and Robust 3D Creation via Orthogonal-view Diffusion Priors/full.md @@ -0,0 +1,241 @@ +# EfficientDreamer: High-Fidelity and Stable 3D Creation via Orthogonal-view Diffusion Priors + +Zhipeng Hu $^{1\dagger}$ , Minda Zhao $^{1\dagger}$ , Chaoyi Zhao $^{1}$ , Xinyue Liang $^{1}$ , Lincheng Li $^{1*}$ , Zeng Zhao $^{1}$ , Changjie Fan $^{1}$ , Xiaowei Zhou $^{2}$ , Xin Yu $^{3}$ + +$^{1}$ NetEase Fuxi AI Lab, $^{2}$ State Key Lab of CAD&CG, Zhejiang University $^{3}$ University of Queensland + +# Abstract + +While image diffusion models have made significant progress in text-driven 3D content creation, they often fail to accurately capture the intended meaning of text prompts, especially for view information. This limitation leads to the Janus problem, where multi-faced 3D models are generated under the guidance of such diffusion models. In this paper, we propose a robust high-quality 3D content generation pipeline by exploiting orthogonal-view image guidance. First, we introduce a novel 2D diffusion model that generates an image consisting of four orthogonal-view subimages based on the given text prompt. Then, the 3D content is created using this diffusion model. Notably, the generated orthogonal-view image provides strong geometric structure priors and thus improves 3D consistency. As a result, it effectively resolves the Janus problem and significantly enhances the quality of 3D content creation. Additionally, we present a 3D synthesis fusion network that can further improve the details of the generated 3D contents. Both quantitative and qualitative evaluations demonstrate that our method surpasses previous text-to-3D techniques. Project page: https://efficientdreamer.github.io. + +# 1. Introduction + +The field of generating photorealistic 2D images from simple text prompts is experiencing rapid growth, and recent advancements are largely attributed to the utilization of diffusion models. These advancements have led to the creation of models capable of producing images of exceptional quality [23-25]. In contrast, generating high-quality 3D content from text prompts poses significantly greater challenges due to the much larger output space and the requirement for 3D consistency. Moreover, the scarcity of adequately large training pairs containing both text and 3D models further handicaps the development of effective models. + +Early methods [9, 12, 19] explore zero-shot text-guided + +![](images/642a679807abd0ee12d8d68e522d94108fcbc6fbd7fbee65788d630494604d1a.jpg) +Figure 1. The given prompt is A pig wearing a backpack, high quality. Left: The pre-trained Stable Diffusion model struggles to accurately generate images based on specific view instructions, thereby encountering the Janus problem in the text-driven 3D generation. Right: EfficientDreamer leverages our newly introduced orthogonal-view diffusion model, enabling the generation of 3D consistent images depicting the same scene from multiple orthogonal viewpoints. + +3D content creation by utilizing guidance from CLIP [22]. These approaches optimize the underlying 3D representations, such as NeRFs and meshes, to achieve high text-image alignment scores for all 2D renderings. However, these methods tend to generate 3D shapes with poor geometry and unsatisfactory appearance. To address these limitations, DreamFusion [21] showcases impressive capabilities in text-to-3D synthesis by leveraging a powerful pretrained text-to-image diffusion model [24] as a strong image prior. A NeRF is trained to represent 3D models using novel Score Distillation Sampling (SDS) gradients [21] and view-dependent prompts. Subsequent works, such as TextMesh[30], Magic3D [14], and ProlificDreamer [33], further improve the quality of the generated 3D contents. + +While previous methods have demonstrated impressive results, a significant drawback is the Janus problem, which + +remains a major challenge. According to recent studies [1, 16], the primary cause of the Janus problem lies in the inability of pre-trained 2D diffusion models to accurately interpret the view instructions specified in the text prompts. Even though explicit view instructions are given, the diffusion model may still generate images that do not align with these instructions. Since the stability of the 3D content creation process relies heavily on accurate guidance from the diffusion model, imprecise guidance can introduce instability and compromise the quality of the generated content. + +To cope with these weaknesses, we propose EfficientDreamer, which leverages a viewpoint-aware diffusion prior to create high-quality and consistent 3D contents. Our approach introduces a novel image diffusion model called orthogonal-view diffusion model, which can generate a composite image comprising four sub-images that are mutually consistent. These sub-images, within a composite image, can be derived from arbitrary viewpoints while adhering to orthogonal views. To achieve this, we maintain the architecture design of the 2D image diffusion model while adapting it to orthogonal-view image generation. We render the 3D models in a large-scale open-source 3D dataset called Objaverse [4], from orthogonal views, and assemble them to fine-tune the pre-trained 2D diffusion model. This process ensures the geometric consistency of our orthogonal-view diffusion model while preserving the capacity of 2D diffusion models to generate detailed and diverse high-quality objects that may not be present in the 3D datasets. + +By leveraging our innovative orthogonal-view image diffusion model, we can effectively tackle the challenge of the multi-face Janus problem while generating consistent 3D models. Specifically, during each training step, we render the 3D representation (like Neus [32] or DMTet [28]) from different orthogonal viewpoints in a differentiable manner. These rendered images are then assembled into a composite image, which is further utilized to optimize the 3D representation via orthogonal-view score distillation. While our orthogonal-view diffusion model, trained with the Objaverse 3D dataset, provides crucial geometric structure priors for 3D content creation, it is important to note that the scale of the 3D dataset is considerably smaller compared to the text-image training pairs. Additionally, the 3D assets in this dataset typically consist of common objects from the real world, having relatively simple topology and texture. + +Relying solely on such an orthogonal-view diffusion prior may lead to potential artifacts in the local appearance of the 3D contents. To address this issue, we also incorporate the original pre-trained 2D diffusion prior in the optimization of the 3D representation. In particular, we introduce a 3D synthesis fusion network which integrates the orthogonal-view diffusion model with a pre-trained 2D diffusion model. We also present a dynamic 3D synthesis strat + +egy to balance the guidance of the orthogonal-view diffusion prior and pre-trained 2D diffusion prior during the SDS optimization process. Initially, our focus is on generating 3D consistent geometric structures, where the high weight assigned to the orthogonal-view diffusion prior helps efficiently represent the geometry and alleviate the Janus problem. As the 3D objects become more complete, we gradually reduce the weight of orthogonal-view diffusion prior and increase the guidance from the original 2D diffusion prior. This approach allows for the generation of local texture details and refinement of local appearance. + +Overall, our contributions are summarized as follows: + +- We propose EfficientDreamer, a novel approach for high-fidelity and robust 3D creation using orthogonal-view diffusion prior. This method effectively tackles the Janus problem and significantly enhances the stability of the generation process. +- We introduce a novel image diffusion model that generates a composite image composed of four sub-images from orthogonal viewpoints. These orthogonal-view images serve as essential geometric structure priors, enabling the generation of 3D consistent geometric structures. +- We present a novel text-driven 3D creation pipeline based on the orthogonal-view SDS loss. We introduce a 3D synthesis fusion network that combines the newly introduced orthogonal-view diffusion prior with the original 2D diffusion prior, aiming to ensure high-quality 3D creations. + +# 2. Related Work + +# 2.1. 3D Reconstruction with Neural Fields + +Traditional Multi-View Stereo (MVS) methods reconstruct 3D point clouds from predicted depth maps [6, 26, 34, 42]. Benefiting from deep learning, supervised MVS methods [7, 31, 35-37] have shown impressive performance on benchmarks but require training on specific datasets [10, 38]. These methods often struggle in less constrained scenarios and cannot easily integrate with other learning-based systems. On the other hand, in recent times, there have been significant advancements in neural fields that have demonstrated remarkable results across various tasks. As a famous neural volume rendering method, NeRF [5, 18, 20, 40] combines classical volumetric rendering with implicit function to render high-quality 2D images. However, directly extracting a mesh from a NeRF representation is non-trivial [29]. Recent works have focused on improving the geometric network while establishing connections between density-based and surface-based representations. VolSDF [39] models the volume density as Laplace Cumulative Distribution Function applied to an SDF representation. Similar to VolSDF, NeuS [32] transforms the SDF field to the accumulated transmittance for volume rendering with the + +Logistic Cumulative Distribution Function and it is emphasized that the weight of the rendering should peak at the first intersection point from the outside to the inside. To this end, we adopt the NeuS representation for the convenience of mesh extraction. + +# 2.2. Text-to-image Generation + +In recent years, with the availability of extremely large datasets of image-text pairs [27], significant progress has been made in text-to-image generation with diffusion models. A typical method is the Stable Diffusion model, which samples from a lower-resolution latent space and decodes latent into high-resolution images [24]. Such sampling operation in the latent greatly improves the efficiency of image generation. Another kind of method utilizes a cascade of super-resolution models [2, 25]. These methods first generate a low-resolution image from a given text prompt and then enlarge the images with several super-resolution models. Recent works design special neural network structures to control diffusion models by adding extra conditions [41], it can learn task-specific input conditions and produce customized images. Due to the strong generative ability of pretrained diffusion models, they are convenient to enable text-to-3D mesh synthesis. + +# 2.3. Text-to-3D Generation + +In recent years, text-to-3D has received significant attention due to the desire to create high-quality 3D content from simple semantics, such as text prompts. Early works attempt to use a CLIP objective to supervise the generation. CLIPMesh [19] deforms a 3D sphere using a CLIP loss to obtain a 3D mesh that fits the input prompt. Text2Mesh [17] stylizes a 3D mesh by predicting color and local geometric details which conform to a target text prompt. Dreamfield [9] optimizes a NeRF from many camera views so that rendered images score highly with a target caption according to a pre-trained CLIP model. However, these methods tend to generate poor 3D shapes with unsatisfactory geometry as well as appearance. + +To overcome such limitations, DreamFusion [21] employs text-to-3D synthesis by utilizing a powerful pretrained text-to-image diffusion model as a strong image prior. They optimize NeRF with their proposed Score Distillation Sampling (SDS) loss. Magic3D [14] extends such a method to a two-stage coarse-to-fine process, which utilizes DMTet for mesh representation. Fantasy3D [3] disentangles geometry and appearance into a two-stage optimization and introduces the spatially varying Bidirectional Reflectance Distribution Function (BRDF) into the text-to-3D task. TextMesh [30] replaces NeRF with VolSDF for more accurate mesh expression. It proposes a novel multi-view consistent and mesh-conditioned re-texturing, enabling the generation of a photorealistic 3D mesh model. + +More recently, ProlificDreamer [33] proposes Variational Score Distillation (VSD), a principled particle-based variational framework to promote the quality of created 3D models. + +# 3. Methodology + +Given a text prompt $y$ , our goal is to create high-quality 3D mesh representations with photorealistic texture. To achieve this, we propose a novel orthogonal-view diffusion model, denoted as $\phi_{ov}$ , which generates composite images consisting of four orthogonal-view sub-images. Subsequently, we design a 3D synthesis fusion network that combines the orthogonal-view diffusion model with a pre-trained diffusion model, denoted as $\phi_{pre}$ , to enhance the quality of the generated 3D content. + +# 3.1. Orthogonal-view Diffusion Model + +Our objective is to generate a composite image in response to a given text prompt $y$ using a fine-tuned diffusion model. The composite image will consist of four sub-images from orthogonal viewpoints, arranged in a $2 \times 2$ grid. + +Common text-to-image diffusion models are trained on large-scale text-image pairs, enabling them to describe objects from various viewpoints. However, these models do not explicitly align corresponding parts of the object across viewpoints. Moreover, they tend to generate images of objects in canonical poses. Hence, it is essential to introduce a novel diffusion model that can generate representations of the same object from diverse viewpoints while maintaining 3D consistency. In this paper, we propose that showcasing the object from four orthogonal views is suitable as it provides essential geometric structure priors. + +To obtain such a specific image diffusion model, we use the recently released Objaverse [4] dataset for fine-tuning the commonly-used diffusion model, i.e., the Stable Diffusion model [24]. This dataset contains $800\mathrm{K} + 3\mathrm{D}$ models created by $100\mathrm{K}+$ artists. The descriptions for each 3D model can be found in the Cap3D [15] dataset, which is processed with BLIP2 [13] and GPT4. We first filter some improper 3D models, like complex scene representation, textureless 3D models, and point clouds, and approximately $420\mathrm{K}$ high-quality 3D models are left. We normalize all assets into a unit cube $[-0.5, 0.5]^3$ . The radius (distance away from the center) is set at 1.8, with a field of view set at 35 degrees. For each 3D model, we uniformly render 12 images in the range of azimuth angle $\xi_{cam} \in (0^\circ, 360^\circ)$ and elevation angle list $\delta_{cam} = [0^\circ, 15^\circ, 30^\circ, 45^\circ]$ . Thus, for each 3D asset, we render 48 RGBA images in $512 \times 512$ resolution in total while its corresponding text prompt is denoted as $y$ . We use the Stable Diffusion 2.1 as the original latent diffusion architecture with an encoder $\mathcal{E}$ , a denoiser U-Net $\epsilon_{\theta}$ , and a decoder $\mathcal{D}$ . We utilize $\mathcal{E}$ and $\mathcal{D}$ with pre-trained weight and only fine-tune $\epsilon_{\theta}$ for generating composite im + +![](images/da7ecf985dc177f07d465a5ef1ca0fa96a757db29bec895f15e87cb9d2cf416d.jpg) +Figure 2. The overview of EfficientDreamer involves two key steps. Firstly, we train an orthogonal-view diffusion model by rendering images from the Objaverse dataset. Secondly, we optimize the 3D scene representation by leveraging both the newly introduced orthogonal-view diffusion model and the pre-trained text-to-image diffusion model. To ensure high-fidelity and robust 3D creation, we employ a dynamic 3D synthesis strategy. + +ages from orthogonal views. As illustrated in Fig. 2(a), for each training step, we first sample an elevation angle from the elevation angle list and then select four render images in this elevation angle, the rendered images are required to be 90 degrees apart from each other in azimuth angle. Then these four images are tiled on a $2 \times 2$ grid in a clockwise rotation. To match the diffusion model, we resize this composite image into $512 \times 512$ resolution, which is denoted as $x$ . At the diffusion time step $t \sim [1, 1000]$ , let $c(y)$ be the embedding of $y$ , and $z_t$ be the noisy latent image by adding noise $\epsilon$ to a clean latent image $z$ , we then solve for the following objective to fine-tune the model: + +$$ +\underset {\theta} {\min } \mathbb {E} _ {z \sim \mathcal {E} (x), t, \epsilon \sim \mathcal {N} (0, 1)} | | \epsilon - \epsilon_ {\theta} \left(z _ {t}, t, c (y)\right) | | ^ {2}. \tag {1} +$$ + +We find the trained diffusion model is capable of learning a generic mechanism to explicitly align the correspondence of an identical object from different viewpoints, even for objects never appearing in the training dataset. The comparison between the orthogonal-view diffusion model and the pre-trained diffusion model is depicted in Fig. 3. + +# 3.2. Text-to-3D via the 3D Synthesis Fusion Network + +Once the orthogonal-view diffusion model is trained, we can generate our initial 3D model by training a neural distance field using the score distillation sampling strategy. The scene is represented as a differential render $g(\psi)$ , where + +$\psi$ denotes the learnable parameters. Given a randomly sampled camera pose with azimuth angle $\xi_{cam}$ , elevation angle $\delta_{cam}$ and radius $r$ , we fix $\delta_{cam}$ and $r$ and sample four camera poses by extending $\xi_{cam}$ to $[\xi_{cam}, \xi_{cam} + 90^{\circ}, \xi_{cam} + 180^{\circ}, \xi_{cam} + 270^{\circ}]$ . Then we render images with these camera poses, which are denoted as $I_0$ , $I_1$ , $I_2$ , $I_3$ . We also tile them on a $2 \times 2$ grid in a clockwise rotation and resize the composite image into $512 \times 512$ resolution, which is denoted as $\hat{I}$ . Then we sample random normal noise according to time step $t$ and add it to $\hat{I}$ . The noisy images $\tilde{I}_t$ , together with text embedding $c(y)$ are fed to our orthogonal-view diffusion model $\phi_{ov}$ , which attempts to predict the noise $\epsilon$ . The denoiser U-Net of $\phi_{ov}$ is denoted as $\epsilon_{\phi_{ov}}$ and the score function of noise estimation is denoted as $\epsilon_{\phi_{ov}}(\tilde{I}_t; t, c(y))$ . This score function guides the direction of the gradient for updating the scene parameters, and the gradient is calculated by orthogonal-view Score Distillation Sampling (SDS): + +$$ +\nabla_ {\psi} \mathcal {L} _ {S D S} ^ {o v} = \mathbb {E} _ {t, \epsilon} [ \omega (t) (\epsilon_ {\phi_ {o v}} (\tilde {I} _ {t}; t, c (y)) - \epsilon) \frac {\partial \hat {I}}{\partial \psi} ], \tag {2} +$$ + +where $\omega (t)$ denotes a weighting function. + +Motivated by [8] and to enhance the stability of the training procedure, we decrease the time step (t) as training progresses, resulting in a gradual decline in the degree of added noise. As our training procedure supervises the 3D creation + +![](images/293360eb425ebb78f77f12e8571fd6b49974451128e890cc5110a50daedaee30.jpg) +Figure 3. Comparison between The orthogonal-view diffusion model and the pre-trained Stable Diffusion model with additional viewpoint instructions. The pre-trained Stable Diffusion model struggles to generate images based on specific view instructions, while the orthogonal-view diffusion model can generate composite images from orthogonal views. + +from four orthogonal views at each time step, the neural distance field can receive semantic structure guidance from different viewpoints. This fundamentally eliminates the possibility of creating multi-face problems. + +Our orthogonal-view diffusion model, trained on the Objaverse 3D dataset, provides essential geometric structure priors for 3D creation. However, it is important to note that the scale of the 3D dataset is relatively small compared to the text-image training pairs. Additionally, the included 3D assets mainly consist of simple objects with uncomplicated topology and texture. Therefore, relying solely on the orthogonal-view diffusion prior may introduce potential artifacts in the local appearance of the generated 3D contents. + +To resolve such a problem, we introduce a 3D synthesis fusion network which integrates the proposed orthogonal-view model with the original text-to-image diffusion model to guide the 3D model generation with a 3D synthesis dynamic strategy. As shown in Fig. 2(b), let $\epsilon_{\phi_{pre}}$ denote the denoiser U-Net of the pre-trained Stable Diffusion model, and rearrange the original rendered images $I_0, I_1, I_2, I_3$ as an image batch $\hat{\mathbb{I}}$ . $\tilde{\mathbb{I}}_t$ denotes the noisy result of $\hat{\mathbb{I}}$ . The + +corresponding SDS loss is formulated as: + +$$ +\nabla_ {\psi} \mathcal {L} _ {S D S} ^ {p r e} = \mathbb {E} _ {t, \epsilon} [ \omega (t) (\epsilon_ {\phi_ {p r e}} (\tilde {\mathbb {I}} _ {t}; t, c (y)) - \epsilon) \frac {\partial \hat {\mathbb {I}}}{\partial \psi} ]. \tag {3} +$$ + +We propose a new prior loss for the novel view supervision to combine both two priors: + +$$ +\nabla_ {\psi} \mathcal {L} _ {S D S} = (1 - (\frac {l}{L}) ^ {\lambda}) \nabla_ {\psi} \mathcal {L} _ {S D S} ^ {o v} + (\frac {l}{L}) ^ {\lambda} \nabla_ {\psi} \mathcal {L} _ {S D S} ^ {p r e}, (4) +$$ + +where $l$ is the current iteration while $L$ is the total iterations. $\lambda$ is a hyperparameter to determine the strength of two diffusion priors, which is set as 1 in our experiments. + +Initially, we prioritize generating 3D content with consistent geometric structures by emphasizing the orthogonal-view diffusion prior. As the generation progresses, we gradually shift the focus towards incorporating guidance from the original 2D diffusion prior to enhance local texture details and refine the overall appearance. + +# 3.3. Implementation Details + +Orthogonal-view Diffusion Model Training: Our model is implemented on the PyTorch platform with 8 NVIDIA Tesla A100 GPUs. We train the model utilizing AdamW optimizer with a learning rate of 1e-5. The whole training procedure takes 200K iterations with a batch size of 256, which takes about 5 days. + +Text-to-3D Generation: We employ a two-stage coarse-to-fine optimization process to generate high-quality 3D mesh representations with photorealistic textures from text prompts. In the coarse stage, we utilize NeuS [32] as the 3D scene representation, as density-based representations such as NeRF are not well-suited for extracting 3D geometry and obtaining meshes [39]. The rendering resolution is set at $64 \times 64$ , and the 3D representation is optimized for 5000 steps with 3D Synthesis Fusion Network. In the refinement stage, we employ DMTet [28] as the scene representation, enabling efficient rendering of textured meshes with differentiable rasterization at $512 \times 512$ resolutions. Additionally, we utilize SDS and VSD guidance [33] to optimize the geometry and texture for another 5000 steps. + +# 4. Experiments + +In this section, we conduct extensive experiments to evaluate the effectiveness of our proposed method for text-to-3D content creation. We present qualitative and quantitative results, compare our method with state-of-the-art methods, perform a user study, and analyze the effectiveness of our proposed techniques through an ablation study. + +# 4.1. Comparison with State-of-the-arts + +We present comprehensive experiments to evaluate the effectiveness of our proposed method for text-to-3D content + +![](images/62c7d2d651bd7a7cc77d5e883faf542aabd95fb07bbc7e8d792e89b227a74194.jpg) +Figure 4. Comparison with other text-to-3D methods. We render each 3D model from two views. Our method outperforms other techniques by generating more high-fidelity 3D models without encountering the Janus problem. + +creation. The compared methods include three crucial Text-to-3D methods, i.e., DreamFusion [21], Magic3D [14] and TextMesh [30]. All these methods are implemented with a unified framework for 3D content creation*. As DreamFusion and TextMesh generate 3D contents with volume rendering of $64 \times 64$ resolution, the generated 3D models may suffer from over-smoothness and lack of details. However, Magic3D can generate higher-quality 3D shapes with improved geometry and texture. Since our method also utilizes a coarse-to-fine two-stage optimization, we are able to create high-quality 3D mesh models. Furthermore, all of the compared methods may be confronted with the Janus problem as they rely on open-source pre-trained diffusion priors that do not enforce valid 3D consistency constraints. Our method can resolve the multi-face problem with our newly introduced orthogonal-view diffusion prior. Please refer to the supplementary material for more comparison results. + +We also evaluate our method using the CLIP score [22] and FID [11] metrics. The CLIP score measures how well images rendered from the generated objects correlate with the provided input text prompt, while FID evaluates the quality and photorealistic appearance of the generated shapes. We render four images with fixed camera poses for each 3D model. The results are depicted in Table 1, which demonstrates that our methods achieve better performance than the state-of-the-art methods. + +Table 1. Comparison with state-of-the-arts. Comparing our method against the state-of-the-arts using CLIP score and FID. $\downarrow$ indicates the lower the better, and $\uparrow$ indicates the higher the better. + +
MethodsCLIP ↑FID ↓
DreamFusion [21]28.40374.44
Magic3D [14]29.15310.57
TextMesh [30]27.65305.77
Ours30.33284.98
+ +# 4.2. Comparison with Perp-Neg + +In addition, we conduct a comparison between our method and another approach, Perp-Neg [1], which is designed to mitigate the Janus problem. We integrate Perp-Neg into DreamFusion and evaluate its performance against our method. The results, shown in Figure 5, reveal that Perp-Neg can address the Janus problem to some extent. For instance, it is capable of generating 3D models with only one head and face. However, these results may exhibit unreasonable details such as three ears or five legs. In contrast, our method, which incorporates guidance from orthogonal views simultaneously, enables us to avoid such issues. + +# 4.3. Coarse-to-fine Two-stage Optimization Results + +In Figure 6, we demonstrate the 3D contents generated through the coarse and fine stages of our optimization procedure. In the coarse stage, we achieve precise 3D represent + +![](images/a76a699429cdd41e563810680f4b6402b1206c7f2f25abf8b56a1d6bd454b971.jpg) +Figure 5. Comparison between the Perp-Neg method and our approach: Our method offers a more comprehensive solution to the Janus problem compared to Perp-Neg. + +![](images/5917dccc5fd9a428838b4dcc961c904a54667557b0eb35b3607c47b1c0a175a0.jpg) +Figure 6. Coarse-to-fine two-stage optimization results. In the coarse stage, we get semantically reasonable 3D scene representation. In the fine stage, we improve the visual quality of generated 3D assets. + +tations for the given text prompts while effectively addressing the challenging Janus problem. This is accomplished with the guidance of the orthogonal-view diffusion model, which provides essential geometric structure priors for 3D creation. In the fine stage, we initialize the mesh representation with the results obtained from the coarse stage and proceed to optimize the geometry using SDS guidance. Additionally, we leverage VSD guidance to enhance the texture, ensuring that the generated 3D model is visually appealing and realistic. As a result, this process enables the generation of significantly higher-quality 3D models. + +Table 2. Results of our user study. The rating score represents the average rating results for different methods, while the preference indicates the ratio of user preference for the results of each method with higher quality. + +
MethodsRating Score ↑Preference(%) ↑
DreamFusion [21]2.094.55
Magic3D [14]2.446.27
TextMesh [30]1.734.91
Ours3.7484.27
+ +# 4.4. User Study + +We conduct a user study with Fuxi Youling Crowdsourcing \* for quantifying the subjective quality. Concretely, we evaluate the study on 22 prompts. At each evaluation time, the participants are shown the created 3D models by DreamFusion [21], Magic3D [14], TextMesh[30] and our method, and asked to rate them on a scale of 1 to 4. In total, we collect 1100 responses from 50 participants. The average rating scores for the different methods are shown in Table 2. Notably, our method receives the highest average rating, indicating that it produces the most visually high-fidelity 3D models. Furthermore, we calculate the preference ratios for each method based on the quality of the results. The analysis reveals that $84.27\%$ of the users consider the results of our method to be of higher quality than those generated by all competing methods. This finding further validates the effectiveness of our method in generating visually highfidelity 3D models. + +# 4.5. Ablation Study + +Impact of the 3D synthesis Fusion Network: We introduce a 3D synthesis fusion network with the dynamic 3D synthesis strategy. Here we provide analysis and experiments to show its effectiveness. Given several text prompts, we employ 3D creation with different diffusion model guidance. The corresponding results are shown in Fig. 7. The first two columns show the results generated without the orthogonal-view diffusion model, which is equivalent to $\lambda = 0$ . Obvious Janus problems and inaccurate 3D representation can be found for the first two prompts. As for the creation of a cottage, we get similar shapes from different viewpoints. The next two columns demonstrate the result guided without the pre-trained diffusion model, which is equivalent to $\lambda = \infty$ . While the orthogonal-view diffusion model can provide comprehensive structure constraints for 3D generation and restrain multi-face issues, we can see local distortion and unreasonable roughness. The last two columns demonstrate the results with our 3D synthesis fusion network, which means $\lambda = 1$ . We get complete and smooth meshes and eliminate the possible Janus problem. + +![](images/3c1a79e6fb43a75a12aaff8ef95cc67180e5f8914ef5989464d27fdf57852e19.jpg) + +![](images/4cd282259356fea7311a49bde6920368471e53357565a9722651718ec2e0f6ef.jpg) +Figure 7. Ablation study on the impact of the 3D synthesis fusion network. + +Quantitative results are depicted in Table 3, which demonstrate the effectiveness of the 3D synthesis fusion network. Impact of Varying Number of View Supervision: Here we present an ablation study on the effects of varying the number of view supervisions on our orthogonal-view diffusion model. For two-view supervision, we randomly mask the opposite viewpoints in the rendered images during training, resulting in the optimization of the 3D representation using only two views. In the case of three-view supervision, we randomly mask one viewpoint, allowing the optimization of the 3D representation using three views. The corresponding results are shown in Figure 8. Generally, supervising the 3D creation with our orthogonal-view diffusion model can mitigate Janus problems. However, creating 3D models with two-view supervision may still suffer from slight multi-face problems, such as two barrels in a tank or multiple faces in a squirrel and a horse. With three-view supervision, these problems are alleviated to some extent. Nevertheless, there are still local distortions in generated 3D models, such as unexpected sags or bulges. These issues can be addressed by using the intact four-view supervision of our orthogonal-view diffusion model. Quantitative results are also presented in Table 3, illustrating the impact of reducing the number of view supervision signals during training. It is evident that there is a decrease in performance observed in this scenario. This highlights the significance of optimizing a 3D representation using a set of four orthogonal-view images, as it provides sufficient geometric structure priors for ensuring consistent 3D content creation. + +![](images/4ef702f17ba7668fb5bcbfc6673da021817437f796bb128d9ed16090ef46e3cb.jpg) +Figure 8. Ablation study on the impact of varying number of view supervision. + +Table 3. Ablation study results using CLIP and FID. $\downarrow$ indicates the lower the better, and $\uparrow$ indicates the higher the better. + +
Impact of the 3D synthesis Fusion Network
Metricsw/o Orthogonal-vieww/o Pre-trainedFull
CLIP ↑31.1332.0532.55
FID ↓372.86369.73342.62
Impact of Varying Number of View Supervision
MetricsTwo-viewThree-viewFour-view
CLIP ↑30.9230.4831.47
FID ↓399.58373.04345.72
+ +# 5. Conclusion + +In this paper, we present EfficientDreamer, a novel approach for high-fidelity and robust 3D creation using orthogonal-view diffusion priors. We introduce a novel diffusion model that can create composite images comprising sub-images from different orthogonal viewpoints. These composite images provide important geometric structure priors for 3D model generation by showcasing an object from multiple perspectives. We enhance the text-to-3D generation framework by leveraging the orthogonal-view diffusion model. Additionally, we introduce a 3D synthesis fusion network that integrates the orthogonal-view diffusion prior with the pre-trained diffusion prior. We also introduce a dynamic 3D synthesis strategy to balance the guidance of these two diffusion models during the SDS optimization. Our method effectively addresses the challenging Janus problem while ensuring the fidelity of the generated 3D models. + +# References + +[1] Mohammadreza Armandpour, Huangjie Zheng, Ali Sadeghian, Amir Sadeghian, and Mingyuan Zhou. Re-imagine the Negative Prompt Algorithm: Transform 2D Diffusion into 3D, alleviate Janus problem and Beyond. arXiv preprint arXiv:2304.04968, 2023. 2, 6 +[2] Yogesh Balaji, Seungjun Nah, Xun Huang, Arash Vahdat, Jiaming Song, Karsten Kreis, Miika Aittala, Timo Aila, Samuli Laine, Bryan Catanzaro, et al. edifi: Text-to-image diffusion models with an ensemble of expert denoisers. arXiv preprint arXiv:2211.01324, 2022. 3 +[3] Rui Chen, Yongwei Chen, Ningxin Jiao, and Kui Jia. Fantasia3d: Disentangling geometry and appearance for high-quality text-to-3d content creation. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 22246–22256, 2023. 3 +[4] Matt Deitke, Dustin Schwenk, Jordi Salvador, Luca Weihs, Oscar Michel, Eli VanderBilt, Ludwig Schmidt, Kiana Ehsani, Aniruddha Kembhavi, and Ali Farhadi. Objverse: A universe of annotated 3d objects. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13142-13153, 2023. 2, 3 +[5] Sara Fridovich-Keil, Alex Yu, Matthew Tancik, Qinhong Chen, Benjamin Recht, and Angjoo Kanazawa. Plenoxels: Radiance fields without neural networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5501–5510, 2022. 2 +[6] Silvano Galliani, Katrin Lasinger, and Konrad Schindler. Massively parallel multiview stereopsis by surface normal diffusion. In Proceedings of the IEEE International Conference on Computer Vision, pages 873-881, 2015. 2 +[7] Xiaodong Gu, Zhiwen Fan, Siyu Zhu, Zuozhuo Dai, Feitong Tan, and Ping Tan. Cascade cost volume for high-resolution multi-view stereo and stereo matching. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2495-2504, 2020. 2 +[8] Yukun Huang, Jianan Wang, Yukai Shi, Xianbiao Qi, Zheng-Jun Zha, and Lei Zhang. Dreamtime: An improved optimization strategy for text-to-3d content creation. arXiv preprint arXiv:2306.12422, 2023. 4 +[9] Ajay Jain, Ben Mildenhall, Jonathan T Barron, Pieter Abbeel, and Ben Poole. Zero-shot text-guided object generation with dream fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 867-876, 2022. 1, 3 +[10] Rasmus Jensen, Anders Dahl, George Vogiatzis, Engin Tola, and Henrik Aanæs. Large scale multi-view stereopsis evaluation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 406-413, 2014. 2 +[11] Tuomas Kynkänniemi, Tero Karras, Miika Aittala, Timo Aila, and Jaakko Lehtinen. The role of imagenet classes in fr\echet inception distance. arXiv preprint arXiv:2203.06026, 2022. 6 +[12] Han-Hung Lee and Angel X Chang. Understanding pure clip guidance for voxel grid nerf models. arXiv preprint arXiv:2209.15172, 2022. 1 + +[13] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. arXiv preprint arXiv:2301.12597, 2023. 3 +[14] Chen-Hsuan Lin, Jun Gao, Luming Tang, Towaki Takikawa, Xiaohui Zeng, Xun Huang, Karsten Kreis, Sanja Fidler, Ming-Yu Liu, and Tsung-Yi Lin. Magic3d: High-resolution text-to-3d content creation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 300–309, 2023. 1, 3, 6, 7 +[15] Tiange Luo, Chris Rockwell, Honglak Lee, and Justin Johnson. Scalable 3d captioning with pretrained models. arXiv preprint arXiv:2306.07279, 2023. 3 +[16] Gal Metzer, Elad Richardson, Or Patashnik, Raja Giryes, and Daniel Cohen-Or. Latent-nerf for shape-guided generation of 3d shapes and textures. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12663–12673, 2023. 2 +[17] Oscar Michel, Roi Bar-On, Richard Liu, Sagie Benaim, and Rana Hanocka. Text2mesh: Text-driven neural stylization for meshes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13492-13502, 2022. 3 +[18] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 2 +[19] Nasir Mohammad Khalid, Tianhao Xie, Eugene Belilovsky, and Tiberiu Popa. Clip-mesh: Generating textured meshes from text using pretrained image-text models. In SIGGRAPH Asia 2022 conference papers, pages 1-8, 2022. 1, 3 +[20] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM Transactions on Graphics (ToG), 41(4):1-15, 2022. 2 +[21] Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. In ICLR, 2023. 1, 3, 6, 7 +[22] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 1, 6 +[23] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022. 1 +[24] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 1, 3 +[25] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, + +et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in Neural Information Processing Systems, 35:36479-36494, 2022. 1, 3 +[26] Johannes L Schonberger, Enliang Zheng, Jan-Michael Frahm, and Marc Pollefeys. Pixelwise view selection for unstructured multi-view stereo. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part III 14, pages 501-518. Springer, 2016. 2 +[27] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. Advances in Neural Information Processing Systems, 35:25278-25294, 2022. 3 +[28] Tianchang Shen, Jun Gao, Kangxue Yin, Ming-Yu Liu, and Sanja Fidler. Deep marching tetrahedra: a hybrid representation for high-resolution 3d shape synthesis. Advances in Neural Information Processing Systems, 34:6087-6101, 2021. 2, 5 +[29] Jiaxiang Tang, Hang Zhou, Xiaokang Chen, Tianshu Hu, Errui Ding, Jingdong Wang, and Gang Zeng. Delicate textured mesh recovery from nerf via adaptive surface refinement. arXiv preprint arXiv:2303.02091, 2023. 2 +[30] Christina Tsalicoglou, Fabian Manhardt, Alessio Tonioni, Michael Niemeyer, and Federico Tombari. Textmesh: Generation of realistic 3d meshes from text prompts. 2024. 1, 3, 6, 7 +[31] Fangjinhua Wang, Silvano Galliani, Christoph Vogel, Pablo Speciale, and Marc Pollefeys. Patchmatchnet: Learned multi-view patchmatch stereo. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14194-14203, 2021. 2 +[32] Peng Wang, Lingjie Liu, Yuan Liu, Christian Theobalt, Taku Komura, and Wenping Wang. Neus: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. arXiv preprint arXiv:2106.10689, 2021. 2, 5 +[33] Zhengyi Wang, Cheng Lu, Yikai Wang, Fan Bao, Chongxuan Li, Hang Su, and Jun Zhu. Prolificdreamer: High-fidelity and diverse text-to-3d generation with variational score distillation. arXiv preprint arXiv:2305.16213, 2023. 1, 3, 5 +[34] Qingshan Xu and Wenbing Tao. Multi-scale geometric consistency guided multi-view stereo. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5483-5492, 2019. 2 +[35] Jiayu Yang, Jose M Alvarez, and Miaomiao Liu. Nonparametric depth distribution modelling based depth inference for multi-view stereo. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8626-8634, 2022. 2 +[36] Yao Yao, Zixin Luo, Shiwei Li, Tian Fang, and Long Quan. Mvsnet: Depth inference for unstructured multi-view stereo. In Proceedings of the European conference on computer vision (ECCV), pages 767-783, 2018. +[37] Yao Yao, Zixin Luo, Shiwei Li, Tianwei Shen, Tian Fang, and Long Quan. Recurrent mvsnet for high-resolution multi-view stereo depth inference. In Proceedings of + +the IEEE/CVF conference on computer vision and pattern recognition, pages 5525-5534, 2019. 2 +[38] Yao Yao, Zixin Luo, Shiwei Li, Jingyang Zhang, Yufan Ren, Lei Zhou, Tian Fang, and Long Quan. Blendedmvs: A large-scale dataset for generalized multi-view stereo networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1790-1799, 2020. 2 +[39] Lior Yariv, Jiatao Gu, Yoni Kasten, and Yaron Lipman. Volume rendering of neural implicit surfaces. Advances in Neural Information Processing Systems, 34:4805-4815, 2021. 2, 5 +[40] Kai Zhang, Gernot Riegler, Noah Snavely, and Vladlen Koltun. Nerf++: Analyzing and improving neural radiance fields. arXiv preprint arXiv:2010.07492, 2020. 2 +[41] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 3836-3847, 2023. 3 +[42] Enliang Zheng, Enrique Dunn, Vladimir Jojic, and Jan-Michael Frahm. Patchmatch based joint view selection and depthmap estimation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 1510-1517, 2014. 2 \ No newline at end of file diff --git a/2024/EfficientDreamer_ High-Fidelity and Robust 3D Creation via Orthogonal-view Diffusion Priors/images.zip b/2024/EfficientDreamer_ High-Fidelity and Robust 3D Creation via Orthogonal-view Diffusion Priors/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..336c8f51bddaf436216f5ba4c25dc9657e3f9078 --- /dev/null +++ b/2024/EfficientDreamer_ High-Fidelity and Robust 3D Creation via Orthogonal-view Diffusion Priors/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:564fdfd0383f66b2c0cd8df8171d1129d686386768e7d577367796380e04c9ee +size 709009 diff --git a/2024/EfficientDreamer_ High-Fidelity and Robust 3D Creation via Orthogonal-view Diffusion Priors/layout.json b/2024/EfficientDreamer_ High-Fidelity and Robust 3D Creation via Orthogonal-view Diffusion Priors/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..8475bf1be137c3ebc5853a5c3a080b32f4a3af0e --- /dev/null +++ b/2024/EfficientDreamer_ High-Fidelity and Robust 3D Creation via Orthogonal-view Diffusion Priors/layout.json @@ -0,0 +1,7231 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 123, + 103, + 473, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 123, + 103, + 473, + 140 + ], + "spans": [ + { + "bbox": [ + 123, + 103, + 473, + 140 + ], + "type": "text", + "content": "EfficientDreamer: High-Fidelity and Stable 3D Creation via Orthogonal-view Diffusion Priors" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 111, + 160, + 481, + 189 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 111, + 160, + 481, + 189 + ], + "spans": [ + { + "bbox": [ + 111, + 160, + 481, + 189 + ], + "type": "text", + "content": "Zhipeng Hu" + }, + { + "bbox": [ + 111, + 160, + 481, + 189 + ], + "type": "inline_equation", + "content": "^{1\\dagger}" + }, + { + "bbox": [ + 111, + 160, + 481, + 189 + ], + "type": "text", + "content": ", Minda Zhao" + }, + { + "bbox": [ + 111, + 160, + 481, + 189 + ], + "type": "inline_equation", + "content": "^{1\\dagger}" + }, + { + "bbox": [ + 111, + 160, + 481, + 189 + ], + "type": "text", + "content": ", Chaoyi Zhao" + }, + { + "bbox": [ + 111, + 160, + 481, + 189 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 111, + 160, + 481, + 189 + ], + "type": "text", + "content": ", Xinyue Liang" + }, + { + "bbox": [ + 111, + 160, + 481, + 189 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 111, + 160, + 481, + 189 + ], + "type": "text", + "content": ", Lincheng Li" + }, + { + "bbox": [ + 111, + 160, + 481, + 189 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 111, + 160, + 481, + 189 + ], + "type": "text", + "content": ", Zeng Zhao" + }, + { + "bbox": [ + 111, + 160, + 481, + 189 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 111, + 160, + 481, + 189 + ], + "type": "text", + "content": ", Changjie Fan" + }, + { + "bbox": [ + 111, + 160, + 481, + 189 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 111, + 160, + 481, + 189 + ], + "type": "text", + "content": ", Xiaowei Zhou" + }, + { + "bbox": [ + 111, + 160, + 481, + 189 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 111, + 160, + 481, + 189 + ], + "type": "text", + "content": ", Xin Yu" + }, + { + "bbox": [ + 111, + 160, + 481, + 189 + ], + "type": "inline_equation", + "content": "^{3}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 190, + 539, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 190, + 539, + 204 + ], + "spans": [ + { + "bbox": [ + 54, + 190, + 539, + 204 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 54, + 190, + 539, + 204 + ], + "type": "text", + "content": "NetEase Fuxi AI Lab, " + }, + { + "bbox": [ + 54, + 190, + 539, + 204 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 54, + 190, + 539, + 204 + ], + "type": "text", + "content": "State Key Lab of CAD&CG, Zhejiang University " + }, + { + "bbox": [ + 54, + 190, + 539, + 204 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 54, + 190, + 539, + 204 + ], + "type": "text", + "content": "University of Queensland" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 143, + 231, + 192, + 243 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 231, + 192, + 243 + ], + "spans": [ + { + "bbox": [ + 143, + 231, + 192, + 243 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 256, + 290, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 256, + 290, + 509 + ], + "spans": [ + { + "bbox": [ + 46, + 256, + 290, + 509 + ], + "type": "text", + "content": "While image diffusion models have made significant progress in text-driven 3D content creation, they often fail to accurately capture the intended meaning of text prompts, especially for view information. This limitation leads to the Janus problem, where multi-faced 3D models are generated under the guidance of such diffusion models. In this paper, we propose a robust high-quality 3D content generation pipeline by exploiting orthogonal-view image guidance. First, we introduce a novel 2D diffusion model that generates an image consisting of four orthogonal-view subimages based on the given text prompt. Then, the 3D content is created using this diffusion model. Notably, the generated orthogonal-view image provides strong geometric structure priors and thus improves 3D consistency. As a result, it effectively resolves the Janus problem and significantly enhances the quality of 3D content creation. Additionally, we present a 3D synthesis fusion network that can further improve the details of the generated 3D contents. Both quantitative and qualitative evaluations demonstrate that our method surpasses previous text-to-3D techniques. Project page: https://efficientdreamer.github.io." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 522, + 128, + 534 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 522, + 128, + 534 + ], + "spans": [ + { + "bbox": [ + 47, + 522, + 128, + 534 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 541, + 287, + 673 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 541, + 287, + 673 + ], + "spans": [ + { + "bbox": [ + 46, + 541, + 287, + 673 + ], + "type": "text", + "content": "The field of generating photorealistic 2D images from simple text prompts is experiencing rapid growth, and recent advancements are largely attributed to the utilization of diffusion models. These advancements have led to the creation of models capable of producing images of exceptional quality [23-25]. In contrast, generating high-quality 3D content from text prompts poses significantly greater challenges due to the much larger output space and the requirement for 3D consistency. Moreover, the scarcity of adequately large training pairs containing both text and 3D models further handicaps the development of effective models." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 58, + 673, + 287, + 686 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 673, + 287, + 686 + ], + "spans": [ + { + "bbox": [ + 58, + 673, + 287, + 686 + ], + "type": "text", + "content": "Early methods [9, 12, 19] explore zero-shot text-guided" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 307, + 230, + 545, + 404 + ], + "blocks": [ + { + "bbox": [ + 307, + 230, + 545, + 404 + ], + "lines": [ + { + "bbox": [ + 307, + 230, + 545, + 404 + ], + "spans": [ + { + "bbox": [ + 307, + 230, + 545, + 404 + ], + "type": "image", + "image_path": "642a679807abd0ee12d8d68e522d94108fcbc6fbd7fbee65788d630494604d1a.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 413, + 547, + 502 + ], + "lines": [ + { + "bbox": [ + 304, + 413, + 547, + 502 + ], + "spans": [ + { + "bbox": [ + 304, + 413, + 547, + 502 + ], + "type": "text", + "content": "Figure 1. The given prompt is A pig wearing a backpack, high quality. Left: The pre-trained Stable Diffusion model struggles to accurately generate images based on specific view instructions, thereby encountering the Janus problem in the text-driven 3D generation. Right: EfficientDreamer leverages our newly introduced orthogonal-view diffusion model, enabling the generation of 3D consistent images depicting the same scene from multiple orthogonal viewpoints." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 518, + 546, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 518, + 546, + 687 + ], + "spans": [ + { + "bbox": [ + 304, + 518, + 546, + 687 + ], + "type": "text", + "content": "3D content creation by utilizing guidance from CLIP [22]. These approaches optimize the underlying 3D representations, such as NeRFs and meshes, to achieve high text-image alignment scores for all 2D renderings. However, these methods tend to generate 3D shapes with poor geometry and unsatisfactory appearance. To address these limitations, DreamFusion [21] showcases impressive capabilities in text-to-3D synthesis by leveraging a powerful pretrained text-to-image diffusion model [24] as a strong image prior. A NeRF is trained to represent 3D models using novel Score Distillation Sampling (SDS) gradients [21] and view-dependent prompts. Subsequent works, such as TextMesh[30], Magic3D [14], and ProlificDreamer [33], further improve the quality of the generated 3D contents." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 689, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 547, + 714 + ], + "type": "text", + "content": "While previous methods have demonstrated impressive results, a significant drawback is the Janus problem, which" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 693, + 124, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 693, + 124, + 703 + ], + "spans": [ + { + "bbox": [ + 58, + 693, + 124, + 703 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 58, + 693, + 124, + 703 + ], + "type": "text", + "content": " Equal contribution" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 59, + 703, + 134, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 703, + 134, + 712 + ], + "spans": [ + { + "bbox": [ + 59, + 703, + 134, + 712 + ], + "type": "text", + "content": "* Corresponding author" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4949" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 193 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 193 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 193 + ], + "type": "text", + "content": "remains a major challenge. According to recent studies [1, 16], the primary cause of the Janus problem lies in the inability of pre-trained 2D diffusion models to accurately interpret the view instructions specified in the text prompts. Even though explicit view instructions are given, the diffusion model may still generate images that do not align with these instructions. Since the stability of the 3D content creation process relies heavily on accurate guidance from the diffusion model, imprecise guidance can introduce instability and compromise the quality of the generated content." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 194, + 289, + 421 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 194, + 289, + 421 + ], + "spans": [ + { + "bbox": [ + 46, + 194, + 289, + 421 + ], + "type": "text", + "content": "To cope with these weaknesses, we propose EfficientDreamer, which leverages a viewpoint-aware diffusion prior to create high-quality and consistent 3D contents. Our approach introduces a novel image diffusion model called orthogonal-view diffusion model, which can generate a composite image comprising four sub-images that are mutually consistent. These sub-images, within a composite image, can be derived from arbitrary viewpoints while adhering to orthogonal views. To achieve this, we maintain the architecture design of the 2D image diffusion model while adapting it to orthogonal-view image generation. We render the 3D models in a large-scale open-source 3D dataset called Objaverse [4], from orthogonal views, and assemble them to fine-tune the pre-trained 2D diffusion model. This process ensures the geometric consistency of our orthogonal-view diffusion model while preserving the capacity of 2D diffusion models to generate detailed and diverse high-quality objects that may not be present in the 3D datasets." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 424, + 289, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 424, + 289, + 616 + ], + "spans": [ + { + "bbox": [ + 46, + 424, + 289, + 616 + ], + "type": "text", + "content": "By leveraging our innovative orthogonal-view image diffusion model, we can effectively tackle the challenge of the multi-face Janus problem while generating consistent 3D models. Specifically, during each training step, we render the 3D representation (like Neus [32] or DMTet [28]) from different orthogonal viewpoints in a differentiable manner. These rendered images are then assembled into a composite image, which is further utilized to optimize the 3D representation via orthogonal-view score distillation. While our orthogonal-view diffusion model, trained with the Objaverse 3D dataset, provides crucial geometric structure priors for 3D content creation, it is important to note that the scale of the 3D dataset is considerably smaller compared to the text-image training pairs. Additionally, the 3D assets in this dataset typically consist of common objects from the real world, having relatively simple topology and texture." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 618, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 618, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 618, + 289, + 715 + ], + "type": "text", + "content": "Relying solely on such an orthogonal-view diffusion prior may lead to potential artifacts in the local appearance of the 3D contents. To address this issue, we also incorporate the original pre-trained 2D diffusion prior in the optimization of the 3D representation. In particular, we introduce a 3D synthesis fusion network which integrates the orthogonal-view diffusion model with a pre-trained 2D diffusion model. We also present a dynamic 3D synthesis strat" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 304, + 72, + 547, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 204 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 204 + ], + "type": "text", + "content": "egy to balance the guidance of the orthogonal-view diffusion prior and pre-trained 2D diffusion prior during the SDS optimization process. Initially, our focus is on generating 3D consistent geometric structures, where the high weight assigned to the orthogonal-view diffusion prior helps efficiently represent the geometry and alleviate the Janus problem. As the 3D objects become more complete, we gradually reduce the weight of orthogonal-view diffusion prior and increase the guidance from the original 2D diffusion prior. This approach allows for the generation of local texture details and refinement of local appearance." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 317, + 205, + 535, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 205, + 535, + 216 + ], + "spans": [ + { + "bbox": [ + 317, + 205, + 535, + 216 + ], + "type": "text", + "content": "Overall, our contributions are summarized as follows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 217, + 545, + 408 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 306, + 217, + 545, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 217, + 545, + 277 + ], + "spans": [ + { + "bbox": [ + 306, + 217, + 545, + 277 + ], + "type": "text", + "content": "- We propose EfficientDreamer, a novel approach for high-fidelity and robust 3D creation using orthogonal-view diffusion prior. This method effectively tackles the Janus problem and significantly enhances the stability of the generation process." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 277, + 545, + 347 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 277, + 545, + 347 + ], + "spans": [ + { + "bbox": [ + 306, + 277, + 545, + 347 + ], + "type": "text", + "content": "- We introduce a novel image diffusion model that generates a composite image composed of four sub-images from orthogonal viewpoints. These orthogonal-view images serve as essential geometric structure priors, enabling the generation of 3D consistent geometric structures." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 350, + 545, + 408 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 350, + 545, + 408 + ], + "spans": [ + { + "bbox": [ + 306, + 350, + 545, + 408 + ], + "type": "text", + "content": "- We present a novel text-driven 3D creation pipeline based on the orthogonal-view SDS loss. We introduce a 3D synthesis fusion network that combines the newly introduced orthogonal-view diffusion prior with the original 2D diffusion prior, aiming to ensure high-quality 3D creations." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 422, + 392, + 435 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 422, + 392, + 435 + ], + "spans": [ + { + "bbox": [ + 306, + 422, + 392, + 435 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 443, + 504, + 455 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 443, + 504, + 455 + ], + "spans": [ + { + "bbox": [ + 306, + 443, + 504, + 455 + ], + "type": "text", + "content": "2.1. 3D Reconstruction with Neural Fields" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 462, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 462, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 462, + 547, + 715 + ], + "type": "text", + "content": "Traditional Multi-View Stereo (MVS) methods reconstruct 3D point clouds from predicted depth maps [6, 26, 34, 42]. Benefiting from deep learning, supervised MVS methods [7, 31, 35-37] have shown impressive performance on benchmarks but require training on specific datasets [10, 38]. These methods often struggle in less constrained scenarios and cannot easily integrate with other learning-based systems. On the other hand, in recent times, there have been significant advancements in neural fields that have demonstrated remarkable results across various tasks. As a famous neural volume rendering method, NeRF [5, 18, 20, 40] combines classical volumetric rendering with implicit function to render high-quality 2D images. However, directly extracting a mesh from a NeRF representation is non-trivial [29]. Recent works have focused on improving the geometric network while establishing connections between density-based and surface-based representations. VolSDF [39] models the volume density as Laplace Cumulative Distribution Function applied to an SDF representation. Similar to VolSDF, NeuS [32] transforms the SDF field to the accumulated transmittance for volume rendering with the" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "4950" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 288, + 131 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 288, + 131 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 288, + 131 + ], + "type": "text", + "content": "Logistic Cumulative Distribution Function and it is emphasized that the weight of the rendering should peak at the first intersection point from the outside to the inside. To this end, we adopt the NeuS representation for the convenience of mesh extraction." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 140, + 190, + 153 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 140, + 190, + 153 + ], + "spans": [ + { + "bbox": [ + 47, + 140, + 190, + 153 + ], + "type": "text", + "content": "2.2. Text-to-image Generation" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 159, + 288, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 159, + 288, + 361 + ], + "spans": [ + { + "bbox": [ + 46, + 159, + 288, + 361 + ], + "type": "text", + "content": "In recent years, with the availability of extremely large datasets of image-text pairs [27], significant progress has been made in text-to-image generation with diffusion models. A typical method is the Stable Diffusion model, which samples from a lower-resolution latent space and decodes latent into high-resolution images [24]. Such sampling operation in the latent greatly improves the efficiency of image generation. Another kind of method utilizes a cascade of super-resolution models [2, 25]. These methods first generate a low-resolution image from a given text prompt and then enlarge the images with several super-resolution models. Recent works design special neural network structures to control diffusion models by adding extra conditions [41], it can learn task-specific input conditions and produce customized images. Due to the strong generative ability of pretrained diffusion models, they are convenient to enable text-to-3D mesh synthesis." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 371, + 176, + 383 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 371, + 176, + 383 + ], + "spans": [ + { + "bbox": [ + 47, + 371, + 176, + 383 + ], + "type": "text", + "content": "2.3. Text-to-3D Generation" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 390, + 287, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 390, + 287, + 545 + ], + "spans": [ + { + "bbox": [ + 46, + 390, + 287, + 545 + ], + "type": "text", + "content": "In recent years, text-to-3D has received significant attention due to the desire to create high-quality 3D content from simple semantics, such as text prompts. Early works attempt to use a CLIP objective to supervise the generation. CLIPMesh [19] deforms a 3D sphere using a CLIP loss to obtain a 3D mesh that fits the input prompt. Text2Mesh [17] stylizes a 3D mesh by predicting color and local geometric details which conform to a target text prompt. Dreamfield [9] optimizes a NeRF from many camera views so that rendered images score highly with a target caption according to a pre-trained CLIP model. However, these methods tend to generate poor 3D shapes with unsatisfactory geometry as well as appearance." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 545, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 545, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 545, + 288, + 713 + ], + "type": "text", + "content": "To overcome such limitations, DreamFusion [21] employs text-to-3D synthesis by utilizing a powerful pretrained text-to-image diffusion model as a strong image prior. They optimize NeRF with their proposed Score Distillation Sampling (SDS) loss. Magic3D [14] extends such a method to a two-stage coarse-to-fine process, which utilizes DMTet for mesh representation. Fantasy3D [3] disentangles geometry and appearance into a two-stage optimization and introduces the spatially varying Bidirectional Reflectance Distribution Function (BRDF) into the text-to-3D task. TextMesh [30] replaces NeRF with VolSDF for more accurate mesh expression. It proposes a novel multi-view consistent and mesh-conditioned re-texturing, enabling the generation of a photorealistic 3D mesh model." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 72, + 545, + 119 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 119 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 119 + ], + "type": "text", + "content": "More recently, ProlificDreamer [33] proposes Variational Score Distillation (VSD), a principled particle-based variational framework to promote the quality of created 3D models." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 130, + 389, + 144 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 130, + 389, + 144 + ], + "spans": [ + { + "bbox": [ + 305, + 130, + 389, + 144 + ], + "type": "text", + "content": "3. Methodology" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 149, + 545, + 258 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 149, + 545, + 258 + ], + "spans": [ + { + "bbox": [ + 304, + 149, + 545, + 258 + ], + "type": "text", + "content": "Given a text prompt " + }, + { + "bbox": [ + 304, + 149, + 545, + 258 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 304, + 149, + 545, + 258 + ], + "type": "text", + "content": ", our goal is to create high-quality 3D mesh representations with photorealistic texture. To achieve this, we propose a novel orthogonal-view diffusion model, denoted as " + }, + { + "bbox": [ + 304, + 149, + 545, + 258 + ], + "type": "inline_equation", + "content": "\\phi_{ov}" + }, + { + "bbox": [ + 304, + 149, + 545, + 258 + ], + "type": "text", + "content": ", which generates composite images consisting of four orthogonal-view sub-images. Subsequently, we design a 3D synthesis fusion network that combines the orthogonal-view diffusion model with a pre-trained diffusion model, denoted as " + }, + { + "bbox": [ + 304, + 149, + 545, + 258 + ], + "type": "inline_equation", + "content": "\\phi_{pre}" + }, + { + "bbox": [ + 304, + 149, + 545, + 258 + ], + "type": "text", + "content": ", to enhance the quality of the generated 3D content." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 264, + 487, + 277 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 264, + 487, + 277 + ], + "spans": [ + { + "bbox": [ + 305, + 264, + 487, + 277 + ], + "type": "text", + "content": "3.1. Orthogonal-view Diffusion Model" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 283, + 545, + 331 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 283, + 545, + 331 + ], + "spans": [ + { + "bbox": [ + 304, + 283, + 545, + 331 + ], + "type": "text", + "content": "Our objective is to generate a composite image in response to a given text prompt " + }, + { + "bbox": [ + 304, + 283, + 545, + 331 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 304, + 283, + 545, + 331 + ], + "type": "text", + "content": " using a fine-tuned diffusion model. The composite image will consist of four sub-images from orthogonal viewpoints, arranged in a " + }, + { + "bbox": [ + 304, + 283, + 545, + 331 + ], + "type": "inline_equation", + "content": "2 \\times 2" + }, + { + "bbox": [ + 304, + 283, + 545, + 331 + ], + "type": "text", + "content": " grid." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 331, + 545, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 331, + 545, + 462 + ], + "spans": [ + { + "bbox": [ + 304, + 331, + 545, + 462 + ], + "type": "text", + "content": "Common text-to-image diffusion models are trained on large-scale text-image pairs, enabling them to describe objects from various viewpoints. However, these models do not explicitly align corresponding parts of the object across viewpoints. Moreover, they tend to generate images of objects in canonical poses. Hence, it is essential to introduce a novel diffusion model that can generate representations of the same object from diverse viewpoints while maintaining 3D consistency. In this paper, we propose that showcasing the object from four orthogonal views is suitable as it provides essential geometric structure priors." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 463, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 463, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 463, + 545, + 713 + ], + "type": "text", + "content": "To obtain such a specific image diffusion model, we use the recently released Objaverse [4] dataset for fine-tuning the commonly-used diffusion model, i.e., the Stable Diffusion model [24]. This dataset contains " + }, + { + "bbox": [ + 304, + 463, + 545, + 713 + ], + "type": "inline_equation", + "content": "800\\mathrm{K} + 3\\mathrm{D}" + }, + { + "bbox": [ + 304, + 463, + 545, + 713 + ], + "type": "text", + "content": " models created by " + }, + { + "bbox": [ + 304, + 463, + 545, + 713 + ], + "type": "inline_equation", + "content": "100\\mathrm{K}+" + }, + { + "bbox": [ + 304, + 463, + 545, + 713 + ], + "type": "text", + "content": " artists. The descriptions for each 3D model can be found in the Cap3D [15] dataset, which is processed with BLIP2 [13] and GPT4. We first filter some improper 3D models, like complex scene representation, textureless 3D models, and point clouds, and approximately " + }, + { + "bbox": [ + 304, + 463, + 545, + 713 + ], + "type": "inline_equation", + "content": "420\\mathrm{K}" + }, + { + "bbox": [ + 304, + 463, + 545, + 713 + ], + "type": "text", + "content": " high-quality 3D models are left. We normalize all assets into a unit cube " + }, + { + "bbox": [ + 304, + 463, + 545, + 713 + ], + "type": "inline_equation", + "content": "[-0.5, 0.5]^3" + }, + { + "bbox": [ + 304, + 463, + 545, + 713 + ], + "type": "text", + "content": ". The radius (distance away from the center) is set at 1.8, with a field of view set at 35 degrees. For each 3D model, we uniformly render 12 images in the range of azimuth angle " + }, + { + "bbox": [ + 304, + 463, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\xi_{cam} \\in (0^\\circ, 360^\\circ)" + }, + { + "bbox": [ + 304, + 463, + 545, + 713 + ], + "type": "text", + "content": " and elevation angle list " + }, + { + "bbox": [ + 304, + 463, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\delta_{cam} = [0^\\circ, 15^\\circ, 30^\\circ, 45^\\circ]" + }, + { + "bbox": [ + 304, + 463, + 545, + 713 + ], + "type": "text", + "content": ". Thus, for each 3D asset, we render 48 RGBA images in " + }, + { + "bbox": [ + 304, + 463, + 545, + 713 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 304, + 463, + 545, + 713 + ], + "type": "text", + "content": " resolution in total while its corresponding text prompt is denoted as " + }, + { + "bbox": [ + 304, + 463, + 545, + 713 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 304, + 463, + 545, + 713 + ], + "type": "text", + "content": ". We use the Stable Diffusion 2.1 as the original latent diffusion architecture with an encoder " + }, + { + "bbox": [ + 304, + 463, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{E}" + }, + { + "bbox": [ + 304, + 463, + 545, + 713 + ], + "type": "text", + "content": ", a denoiser U-Net " + }, + { + "bbox": [ + 304, + 463, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}" + }, + { + "bbox": [ + 304, + 463, + 545, + 713 + ], + "type": "text", + "content": ", and a decoder " + }, + { + "bbox": [ + 304, + 463, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 304, + 463, + 545, + 713 + ], + "type": "text", + "content": ". We utilize " + }, + { + "bbox": [ + 304, + 463, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{E}" + }, + { + "bbox": [ + 304, + 463, + 545, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 463, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 304, + 463, + 545, + 713 + ], + "type": "text", + "content": " with pre-trained weight and only fine-tune " + }, + { + "bbox": [ + 304, + 463, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}" + }, + { + "bbox": [ + 304, + 463, + 545, + 713 + ], + "type": "text", + "content": " for generating composite im" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "4951" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 52, + 69, + 545, + 329 + ], + "blocks": [ + { + "bbox": [ + 52, + 69, + 545, + 329 + ], + "lines": [ + { + "bbox": [ + 52, + 69, + 545, + 329 + ], + "spans": [ + { + "bbox": [ + 52, + 69, + 545, + 329 + ], + "type": "image", + "image_path": "da7ecf985dc177f07d465a5ef1ca0fa96a757db29bec895f15e87cb9d2cf416d.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 336, + 548, + 381 + ], + "lines": [ + { + "bbox": [ + 46, + 336, + 548, + 381 + ], + "spans": [ + { + "bbox": [ + 46, + 336, + 548, + 381 + ], + "type": "text", + "content": "Figure 2. The overview of EfficientDreamer involves two key steps. Firstly, we train an orthogonal-view diffusion model by rendering images from the Objaverse dataset. Secondly, we optimize the 3D scene representation by leveraging both the newly introduced orthogonal-view diffusion model and the pre-trained text-to-image diffusion model. To ensure high-fidelity and robust 3D creation, we employ a dynamic 3D synthesis strategy." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 392, + 289, + 536 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 392, + 289, + 536 + ], + "spans": [ + { + "bbox": [ + 46, + 392, + 289, + 536 + ], + "type": "text", + "content": "ages from orthogonal views. As illustrated in Fig. 2(a), for each training step, we first sample an elevation angle from the elevation angle list and then select four render images in this elevation angle, the rendered images are required to be 90 degrees apart from each other in azimuth angle. Then these four images are tiled on a " + }, + { + "bbox": [ + 46, + 392, + 289, + 536 + ], + "type": "inline_equation", + "content": "2 \\times 2" + }, + { + "bbox": [ + 46, + 392, + 289, + 536 + ], + "type": "text", + "content": " grid in a clockwise rotation. To match the diffusion model, we resize this composite image into " + }, + { + "bbox": [ + 46, + 392, + 289, + 536 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 46, + 392, + 289, + 536 + ], + "type": "text", + "content": " resolution, which is denoted as " + }, + { + "bbox": [ + 46, + 392, + 289, + 536 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 46, + 392, + 289, + 536 + ], + "type": "text", + "content": ". At the diffusion time step " + }, + { + "bbox": [ + 46, + 392, + 289, + 536 + ], + "type": "inline_equation", + "content": "t \\sim [1, 1000]" + }, + { + "bbox": [ + 46, + 392, + 289, + 536 + ], + "type": "text", + "content": ", let " + }, + { + "bbox": [ + 46, + 392, + 289, + 536 + ], + "type": "inline_equation", + "content": "c(y)" + }, + { + "bbox": [ + 46, + 392, + 289, + 536 + ], + "type": "text", + "content": " be the embedding of " + }, + { + "bbox": [ + 46, + 392, + 289, + 536 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 46, + 392, + 289, + 536 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 392, + 289, + 536 + ], + "type": "inline_equation", + "content": "z_t" + }, + { + "bbox": [ + 46, + 392, + 289, + 536 + ], + "type": "text", + "content": " be the noisy latent image by adding noise " + }, + { + "bbox": [ + 46, + 392, + 289, + 536 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 46, + 392, + 289, + 536 + ], + "type": "text", + "content": " to a clean latent image " + }, + { + "bbox": [ + 46, + 392, + 289, + 536 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 46, + 392, + 289, + 536 + ], + "type": "text", + "content": ", we then solve for the following objective to fine-tune the model:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 73, + 543, + 287, + 563 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 73, + 543, + 287, + 563 + ], + "spans": [ + { + "bbox": [ + 73, + 543, + 287, + 563 + ], + "type": "interline_equation", + "content": "\\underset {\\theta} {\\min } \\mathbb {E} _ {z \\sim \\mathcal {E} (x), t, \\epsilon \\sim \\mathcal {N} (0, 1)} | | \\epsilon - \\epsilon_ {\\theta} \\left(z _ {t}, t, c (y)\\right) | | ^ {2}. \\tag {1}", + "image_path": "58b746d0f20ee91f01da4fa2da9f12e566e978fcf9f308b20e02686b5fef78d8.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 568, + 288, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 568, + 288, + 641 + ], + "spans": [ + { + "bbox": [ + 46, + 568, + 288, + 641 + ], + "type": "text", + "content": "We find the trained diffusion model is capable of learning a generic mechanism to explicitly align the correspondence of an identical object from different viewpoints, even for objects never appearing in the training dataset. The comparison between the orthogonal-view diffusion model and the pre-trained diffusion model is depicted in Fig. 3." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 647, + 288, + 661 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 647, + 288, + 661 + ], + "spans": [ + { + "bbox": [ + 47, + 647, + 288, + 661 + ], + "type": "text", + "content": "3.2. Text-to-3D via the 3D Synthesis Fusion Network" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "type": "text", + "content": "Once the orthogonal-view diffusion model is trained, we can generate our initial 3D model by training a neural distance field using the score distillation sampling strategy. The scene is represented as a differential render " + }, + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "type": "inline_equation", + "content": "g(\\psi)" + }, + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "type": "text", + "content": ", where" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 392, + 547, + 608 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 392, + 547, + 608 + ], + "spans": [ + { + "bbox": [ + 304, + 392, + 547, + 608 + ], + "type": "inline_equation", + "content": "\\psi" + }, + { + "bbox": [ + 304, + 392, + 547, + 608 + ], + "type": "text", + "content": " denotes the learnable parameters. Given a randomly sampled camera pose with azimuth angle " + }, + { + "bbox": [ + 304, + 392, + 547, + 608 + ], + "type": "inline_equation", + "content": "\\xi_{cam}" + }, + { + "bbox": [ + 304, + 392, + 547, + 608 + ], + "type": "text", + "content": ", elevation angle " + }, + { + "bbox": [ + 304, + 392, + 547, + 608 + ], + "type": "inline_equation", + "content": "\\delta_{cam}" + }, + { + "bbox": [ + 304, + 392, + 547, + 608 + ], + "type": "text", + "content": " and radius " + }, + { + "bbox": [ + 304, + 392, + 547, + 608 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 304, + 392, + 547, + 608 + ], + "type": "text", + "content": ", we fix " + }, + { + "bbox": [ + 304, + 392, + 547, + 608 + ], + "type": "inline_equation", + "content": "\\delta_{cam}" + }, + { + "bbox": [ + 304, + 392, + 547, + 608 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 392, + 547, + 608 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 304, + 392, + 547, + 608 + ], + "type": "text", + "content": " and sample four camera poses by extending " + }, + { + "bbox": [ + 304, + 392, + 547, + 608 + ], + "type": "inline_equation", + "content": "\\xi_{cam}" + }, + { + "bbox": [ + 304, + 392, + 547, + 608 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 304, + 392, + 547, + 608 + ], + "type": "inline_equation", + "content": "[\\xi_{cam}, \\xi_{cam} + 90^{\\circ}, \\xi_{cam} + 180^{\\circ}, \\xi_{cam} + 270^{\\circ}]" + }, + { + "bbox": [ + 304, + 392, + 547, + 608 + ], + "type": "text", + "content": ". Then we render images with these camera poses, which are denoted as " + }, + { + "bbox": [ + 304, + 392, + 547, + 608 + ], + "type": "inline_equation", + "content": "I_0" + }, + { + "bbox": [ + 304, + 392, + 547, + 608 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 392, + 547, + 608 + ], + "type": "inline_equation", + "content": "I_1" + }, + { + "bbox": [ + 304, + 392, + 547, + 608 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 392, + 547, + 608 + ], + "type": "inline_equation", + "content": "I_2" + }, + { + "bbox": [ + 304, + 392, + 547, + 608 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 392, + 547, + 608 + ], + "type": "inline_equation", + "content": "I_3" + }, + { + "bbox": [ + 304, + 392, + 547, + 608 + ], + "type": "text", + "content": ". We also tile them on a " + }, + { + "bbox": [ + 304, + 392, + 547, + 608 + ], + "type": "inline_equation", + "content": "2 \\times 2" + }, + { + "bbox": [ + 304, + 392, + 547, + 608 + ], + "type": "text", + "content": " grid in a clockwise rotation and resize the composite image into " + }, + { + "bbox": [ + 304, + 392, + 547, + 608 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 304, + 392, + 547, + 608 + ], + "type": "text", + "content": " resolution, which is denoted as " + }, + { + "bbox": [ + 304, + 392, + 547, + 608 + ], + "type": "inline_equation", + "content": "\\hat{I}" + }, + { + "bbox": [ + 304, + 392, + 547, + 608 + ], + "type": "text", + "content": ". Then we sample random normal noise according to time step " + }, + { + "bbox": [ + 304, + 392, + 547, + 608 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 392, + 547, + 608 + ], + "type": "text", + "content": " and add it to " + }, + { + "bbox": [ + 304, + 392, + 547, + 608 + ], + "type": "inline_equation", + "content": "\\hat{I}" + }, + { + "bbox": [ + 304, + 392, + 547, + 608 + ], + "type": "text", + "content": ". The noisy images " + }, + { + "bbox": [ + 304, + 392, + 547, + 608 + ], + "type": "inline_equation", + "content": "\\tilde{I}_t" + }, + { + "bbox": [ + 304, + 392, + 547, + 608 + ], + "type": "text", + "content": ", together with text embedding " + }, + { + "bbox": [ + 304, + 392, + 547, + 608 + ], + "type": "inline_equation", + "content": "c(y)" + }, + { + "bbox": [ + 304, + 392, + 547, + 608 + ], + "type": "text", + "content": " are fed to our orthogonal-view diffusion model " + }, + { + "bbox": [ + 304, + 392, + 547, + 608 + ], + "type": "inline_equation", + "content": "\\phi_{ov}" + }, + { + "bbox": [ + 304, + 392, + 547, + 608 + ], + "type": "text", + "content": ", which attempts to predict the noise " + }, + { + "bbox": [ + 304, + 392, + 547, + 608 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 304, + 392, + 547, + 608 + ], + "type": "text", + "content": ". The denoiser U-Net of " + }, + { + "bbox": [ + 304, + 392, + 547, + 608 + ], + "type": "inline_equation", + "content": "\\phi_{ov}" + }, + { + "bbox": [ + 304, + 392, + 547, + 608 + ], + "type": "text", + "content": " is denoted as " + }, + { + "bbox": [ + 304, + 392, + 547, + 608 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\phi_{ov}}" + }, + { + "bbox": [ + 304, + 392, + 547, + 608 + ], + "type": "text", + "content": " and the score function of noise estimation is denoted as " + }, + { + "bbox": [ + 304, + 392, + 547, + 608 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\phi_{ov}}(\\tilde{I}_t; t, c(y))" + }, + { + "bbox": [ + 304, + 392, + 547, + 608 + ], + "type": "text", + "content": ". This score function guides the direction of the gradient for updating the scene parameters, and the gradient is calculated by orthogonal-view Score Distillation Sampling (SDS):" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 319, + 616, + 545, + 644 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 616, + 545, + 644 + ], + "spans": [ + { + "bbox": [ + 319, + 616, + 545, + 644 + ], + "type": "interline_equation", + "content": "\\nabla_ {\\psi} \\mathcal {L} _ {S D S} ^ {o v} = \\mathbb {E} _ {t, \\epsilon} [ \\omega (t) (\\epsilon_ {\\phi_ {o v}} (\\tilde {I} _ {t}; t, c (y)) - \\epsilon) \\frac {\\partial \\hat {I}}{\\partial \\psi} ], \\tag {2}", + "image_path": "0b77a72c5f9b971ba3878ad9cad25f2bcd7c86490bfa20cfb9df39925e95f897.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 653, + 474, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 653, + 474, + 665 + ], + "spans": [ + { + "bbox": [ + 305, + 653, + 474, + 665 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 653, + 474, + 665 + ], + "type": "inline_equation", + "content": "\\omega (t)" + }, + { + "bbox": [ + 305, + 653, + 474, + 665 + ], + "type": "text", + "content": " denotes a weighting function." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "type": "text", + "content": "Motivated by [8] and to enhance the stability of the training procedure, we decrease the time step (t) as training progresses, resulting in a gradual decline in the degree of added noise. As our training procedure supervises the 3D creation" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4952" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 51, + 72, + 286, + 338 + ], + "blocks": [ + { + "bbox": [ + 51, + 72, + 286, + 338 + ], + "lines": [ + { + "bbox": [ + 51, + 72, + 286, + 338 + ], + "spans": [ + { + "bbox": [ + 51, + 72, + 286, + 338 + ], + "type": "image", + "image_path": "293360eb425ebb78f77f12e8571fd6b49974451128e890cc5110a50daedaee30.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 344, + 287, + 411 + ], + "lines": [ + { + "bbox": [ + 46, + 344, + 287, + 411 + ], + "spans": [ + { + "bbox": [ + 46, + 344, + 287, + 411 + ], + "type": "text", + "content": "Figure 3. Comparison between The orthogonal-view diffusion model and the pre-trained Stable Diffusion model with additional viewpoint instructions. The pre-trained Stable Diffusion model struggles to generate images based on specific view instructions, while the orthogonal-view diffusion model can generate composite images from orthogonal views." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 447, + 287, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 447, + 287, + 495 + ], + "spans": [ + { + "bbox": [ + 46, + 447, + 287, + 495 + ], + "type": "text", + "content": "from four orthogonal views at each time step, the neural distance field can receive semantic structure guidance from different viewpoints. This fundamentally eliminates the possibility of creating multi-face problems." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 502, + 287, + 612 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 502, + 287, + 612 + ], + "spans": [ + { + "bbox": [ + 46, + 502, + 287, + 612 + ], + "type": "text", + "content": "Our orthogonal-view diffusion model, trained on the Objaverse 3D dataset, provides essential geometric structure priors for 3D creation. However, it is important to note that the scale of the 3D dataset is relatively small compared to the text-image training pairs. Additionally, the included 3D assets mainly consist of simple objects with uncomplicated topology and texture. Therefore, relying solely on the orthogonal-view diffusion prior may introduce potential artifacts in the local appearance of the generated 3D contents." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 618, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 618, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 618, + 288, + 715 + ], + "type": "text", + "content": "To resolve such a problem, we introduce a 3D synthesis fusion network which integrates the proposed orthogonal-view model with the original text-to-image diffusion model to guide the 3D model generation with a 3D synthesis dynamic strategy. As shown in Fig. 2(b), let " + }, + { + "bbox": [ + 46, + 618, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\phi_{pre}}" + }, + { + "bbox": [ + 46, + 618, + 288, + 715 + ], + "type": "text", + "content": " denote the denoiser U-Net of the pre-trained Stable Diffusion model, and rearrange the original rendered images " + }, + { + "bbox": [ + 46, + 618, + 288, + 715 + ], + "type": "inline_equation", + "content": "I_0, I_1, I_2, I_3" + }, + { + "bbox": [ + 46, + 618, + 288, + 715 + ], + "type": "text", + "content": " as an image batch " + }, + { + "bbox": [ + 46, + 618, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbb{I}}" + }, + { + "bbox": [ + 46, + 618, + 288, + 715 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 46, + 618, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbb{I}}_t" + }, + { + "bbox": [ + 46, + 618, + 288, + 715 + ], + "type": "text", + "content": " denotes the noisy result of " + }, + { + "bbox": [ + 46, + 618, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbb{I}}" + }, + { + "bbox": [ + 46, + 618, + 288, + 715 + ], + "type": "text", + "content": ". The" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 305, + 72, + 474, + 85 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 474, + 85 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 474, + 85 + ], + "type": "text", + "content": "corresponding SDS loss is formulated as:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 318, + 92, + 545, + 119 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 92, + 545, + 119 + ], + "spans": [ + { + "bbox": [ + 318, + 92, + 545, + 119 + ], + "type": "interline_equation", + "content": "\\nabla_ {\\psi} \\mathcal {L} _ {S D S} ^ {p r e} = \\mathbb {E} _ {t, \\epsilon} [ \\omega (t) (\\epsilon_ {\\phi_ {p r e}} (\\tilde {\\mathbb {I}} _ {t}; t, c (y)) - \\epsilon) \\frac {\\partial \\hat {\\mathbb {I}}}{\\partial \\psi} ]. \\tag {3}", + "image_path": "268badf8832a914a2195d3adf84a36c1233f138f5d29141fc7c309b9faa06cab.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 127, + 545, + 151 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 127, + 545, + 151 + ], + "spans": [ + { + "bbox": [ + 305, + 127, + 545, + 151 + ], + "type": "text", + "content": "We propose a new prior loss for the novel view supervision to combine both two priors:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 311, + 159, + 545, + 183 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 159, + 545, + 183 + ], + "spans": [ + { + "bbox": [ + 311, + 159, + 545, + 183 + ], + "type": "interline_equation", + "content": "\\nabla_ {\\psi} \\mathcal {L} _ {S D S} = (1 - (\\frac {l}{L}) ^ {\\lambda}) \\nabla_ {\\psi} \\mathcal {L} _ {S D S} ^ {o v} + (\\frac {l}{L}) ^ {\\lambda} \\nabla_ {\\psi} \\mathcal {L} _ {S D S} ^ {p r e}, (4)", + "image_path": "33a2e04ff5928723010ef56175ea88fd0422fabe9f3ccc64163090bdaade59b2.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 189, + 545, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 189, + 545, + 224 + ], + "spans": [ + { + "bbox": [ + 305, + 189, + 545, + 224 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 189, + 545, + 224 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 305, + 189, + 545, + 224 + ], + "type": "text", + "content": " is the current iteration while " + }, + { + "bbox": [ + 305, + 189, + 545, + 224 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 305, + 189, + 545, + 224 + ], + "type": "text", + "content": " is the total iterations. " + }, + { + "bbox": [ + 305, + 189, + 545, + 224 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 305, + 189, + 545, + 224 + ], + "type": "text", + "content": " is a hyperparameter to determine the strength of two diffusion priors, which is set as 1 in our experiments." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 225, + 545, + 297 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 225, + 545, + 297 + ], + "spans": [ + { + "bbox": [ + 305, + 225, + 545, + 297 + ], + "type": "text", + "content": "Initially, we prioritize generating 3D content with consistent geometric structures by emphasizing the orthogonal-view diffusion prior. As the generation progresses, we gradually shift the focus towards incorporating guidance from the original 2D diffusion prior to enhance local texture details and refine the overall appearance." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 304, + 440, + 317 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 304, + 440, + 317 + ], + "spans": [ + { + "bbox": [ + 306, + 304, + 440, + 317 + ], + "type": "text", + "content": "3.3. Implementation Details" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 322, + 545, + 393 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 322, + 545, + 393 + ], + "spans": [ + { + "bbox": [ + 305, + 322, + 545, + 393 + ], + "type": "text", + "content": "Orthogonal-view Diffusion Model Training: Our model is implemented on the PyTorch platform with 8 NVIDIA Tesla A100 GPUs. We train the model utilizing AdamW optimizer with a learning rate of 1e-5. The whole training procedure takes 200K iterations with a batch size of 256, which takes about 5 days." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 394, + 546, + 562 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 394, + 546, + 562 + ], + "spans": [ + { + "bbox": [ + 304, + 394, + 546, + 562 + ], + "type": "text", + "content": "Text-to-3D Generation: We employ a two-stage coarse-to-fine optimization process to generate high-quality 3D mesh representations with photorealistic textures from text prompts. In the coarse stage, we utilize NeuS [32] as the 3D scene representation, as density-based representations such as NeRF are not well-suited for extracting 3D geometry and obtaining meshes [39]. The rendering resolution is set at " + }, + { + "bbox": [ + 304, + 394, + 546, + 562 + ], + "type": "inline_equation", + "content": "64 \\times 64" + }, + { + "bbox": [ + 304, + 394, + 546, + 562 + ], + "type": "text", + "content": ", and the 3D representation is optimized for 5000 steps with 3D Synthesis Fusion Network. In the refinement stage, we employ DMTet [28] as the scene representation, enabling efficient rendering of textured meshes with differentiable rasterization at " + }, + { + "bbox": [ + 304, + 394, + 546, + 562 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 304, + 394, + 546, + 562 + ], + "type": "text", + "content": " resolutions. Additionally, we utilize SDS and VSD guidance [33] to optimize the geometry and texture for another 5000 steps." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 572, + 388, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 572, + 388, + 586 + ], + "spans": [ + { + "bbox": [ + 306, + 572, + 388, + 586 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 592, + 545, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 592, + 545, + 665 + ], + "spans": [ + { + "bbox": [ + 304, + 592, + 545, + 665 + ], + "type": "text", + "content": "In this section, we conduct extensive experiments to evaluate the effectiveness of our proposed method for text-to-3D content creation. We present qualitative and quantitative results, compare our method with state-of-the-art methods, perform a user study, and analyze the effectiveness of our proposed techniques through an ablation study." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 671, + 491, + 685 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 671, + 491, + 685 + ], + "spans": [ + { + "bbox": [ + 306, + 671, + 491, + 685 + ], + "type": "text", + "content": "4.1. Comparison with State-of-the-arts" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "content": "We present comprehensive experiments to evaluate the effectiveness of our proposed method for text-to-3D content" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4953" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 72, + 541, + 333 + ], + "blocks": [ + { + "bbox": [ + 50, + 72, + 541, + 333 + ], + "lines": [ + { + "bbox": [ + 50, + 72, + 541, + 333 + ], + "spans": [ + { + "bbox": [ + 50, + 72, + 541, + 333 + ], + "type": "image", + "image_path": "62c7d2d651bd7a7cc77d5e883faf542aabd95fb07bbc7e8d792e89b227a74194.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 342, + 546, + 365 + ], + "lines": [ + { + "bbox": [ + 46, + 342, + 546, + 365 + ], + "spans": [ + { + "bbox": [ + 46, + 342, + 546, + 365 + ], + "type": "text", + "content": "Figure 4. Comparison with other text-to-3D methods. We render each 3D model from two views. Our method outperforms other techniques by generating more high-fidelity 3D models without encountering the Janus problem." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 375, + 289, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 375, + 289, + 579 + ], + "spans": [ + { + "bbox": [ + 46, + 375, + 289, + 579 + ], + "type": "text", + "content": "creation. The compared methods include three crucial Text-to-3D methods, i.e., DreamFusion [21], Magic3D [14] and TextMesh [30]. All these methods are implemented with a unified framework for 3D content creation*. As DreamFusion and TextMesh generate 3D contents with volume rendering of " + }, + { + "bbox": [ + 46, + 375, + 289, + 579 + ], + "type": "inline_equation", + "content": "64 \\times 64" + }, + { + "bbox": [ + 46, + 375, + 289, + 579 + ], + "type": "text", + "content": " resolution, the generated 3D models may suffer from over-smoothness and lack of details. However, Magic3D can generate higher-quality 3D shapes with improved geometry and texture. Since our method also utilizes a coarse-to-fine two-stage optimization, we are able to create high-quality 3D mesh models. Furthermore, all of the compared methods may be confronted with the Janus problem as they rely on open-source pre-trained diffusion priors that do not enforce valid 3D consistency constraints. Our method can resolve the multi-face problem with our newly introduced orthogonal-view diffusion prior. Please refer to the supplementary material for more comparison results." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 581, + 287, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 581, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 46, + 581, + 287, + 689 + ], + "type": "text", + "content": "We also evaluate our method using the CLIP score [22] and FID [11] metrics. The CLIP score measures how well images rendered from the generated objects correlate with the provided input text prompt, while FID evaluates the quality and photorealistic appearance of the generated shapes. We render four images with fixed camera poses for each 3D model. The results are depicted in Table 1, which demonstrates that our methods achieve better performance than the state-of-the-art methods." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 306, + 374, + 546, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 374, + 546, + 407 + ], + "spans": [ + { + "bbox": [ + 306, + 374, + 546, + 407 + ], + "type": "text", + "content": "Table 1. Comparison with state-of-the-arts. Comparing our method against the state-of-the-arts using CLIP score and FID. " + }, + { + "bbox": [ + 306, + 374, + 546, + 407 + ], + "type": "inline_equation", + "content": "\\downarrow" + }, + { + "bbox": [ + 306, + 374, + 546, + 407 + ], + "type": "text", + "content": " indicates the lower the better, and " + }, + { + "bbox": [ + 306, + 374, + 546, + 407 + ], + "type": "inline_equation", + "content": "\\uparrow" + }, + { + "bbox": [ + 306, + 374, + 546, + 407 + ], + "type": "text", + "content": " indicates the higher the better." + } + ] + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 346, + 415, + 501, + 485 + ], + "blocks": [ + { + "bbox": [ + 346, + 415, + 501, + 485 + ], + "lines": [ + { + "bbox": [ + 346, + 415, + 501, + 485 + ], + "spans": [ + { + "bbox": [ + 346, + 415, + 501, + 485 + ], + "type": "table", + "html": "
MethodsCLIP ↑FID ↓
DreamFusion [21]28.40374.44
Magic3D [14]29.15310.57
TextMesh [30]27.65305.77
Ours30.33284.98
", + "image_path": "922cd00208ed3484b412cd3c50b5b84dceed2a7ce26e5d7b63c6b8367aead305.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 500, + 457, + 513 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 500, + 457, + 513 + ], + "spans": [ + { + "bbox": [ + 306, + 500, + 457, + 513 + ], + "type": "text", + "content": "4.2. Comparison with Perp-Neg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 518, + 545, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 518, + 545, + 651 + ], + "spans": [ + { + "bbox": [ + 304, + 518, + 545, + 651 + ], + "type": "text", + "content": "In addition, we conduct a comparison between our method and another approach, Perp-Neg [1], which is designed to mitigate the Janus problem. We integrate Perp-Neg into DreamFusion and evaluate its performance against our method. The results, shown in Figure 5, reveal that Perp-Neg can address the Janus problem to some extent. For instance, it is capable of generating 3D models with only one head and face. However, these results may exhibit unreasonable details such as three ears or five legs. In contrast, our method, which incorporates guidance from orthogonal views simultaneously, enables us to avoid such issues." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 658, + 545, + 672 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 658, + 545, + 672 + ], + "spans": [ + { + "bbox": [ + 306, + 658, + 545, + 672 + ], + "type": "text", + "content": "4.3. Coarse-to-fine Two-stage Optimization Results" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 677, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 546, + 713 + ], + "type": "text", + "content": "In Figure 6, we demonstrate the 3D contents generated through the coarse and fine stages of our optimization procedure. In the coarse stage, we achieve precise 3D represent" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 58, + 702, + 223, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 702, + 223, + 713 + ], + "spans": [ + { + "bbox": [ + 58, + 702, + 223, + 713 + ], + "type": "text", + "content": "* https://github.com/threestudio-project/threestudio" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4954" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 47, + 70, + 289, + 223 + ], + "blocks": [ + { + "bbox": [ + 47, + 70, + 289, + 223 + ], + "lines": [ + { + "bbox": [ + 47, + 70, + 289, + 223 + ], + "spans": [ + { + "bbox": [ + 47, + 70, + 289, + 223 + ], + "type": "image", + "image_path": "a76a699429cdd41e563810680f4b6402b1206c7f2f25abf8b56a1d6bd454b971.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 229, + 288, + 264 + ], + "lines": [ + { + "bbox": [ + 46, + 229, + 288, + 264 + ], + "spans": [ + { + "bbox": [ + 46, + 229, + 288, + 264 + ], + "type": "text", + "content": "Figure 5. Comparison between the Perp-Neg method and our approach: Our method offers a more comprehensive solution to the Janus problem compared to Perp-Neg." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 50, + 291, + 289, + 504 + ], + "blocks": [ + { + "bbox": [ + 50, + 291, + 289, + 504 + ], + "lines": [ + { + "bbox": [ + 50, + 291, + 289, + 504 + ], + "spans": [ + { + "bbox": [ + 50, + 291, + 289, + 504 + ], + "type": "image", + "image_path": "5917dccc5fd9a428838b4dcc961c904a54667557b0eb35b3607c47b1c0a175a0.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 508, + 288, + 552 + ], + "lines": [ + { + "bbox": [ + 46, + 508, + 288, + 552 + ], + "spans": [ + { + "bbox": [ + 46, + 508, + 288, + 552 + ], + "type": "text", + "content": "Figure 6. Coarse-to-fine two-stage optimization results. In the coarse stage, we get semantically reasonable 3D scene representation. In the fine stage, we improve the visual quality of generated 3D assets." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "text", + "content": "tations for the given text prompts while effectively addressing the challenging Janus problem. This is accomplished with the guidance of the orthogonal-view diffusion model, which provides essential geometric structure priors for 3D creation. In the fine stage, we initialize the mesh representation with the results obtained from the coarse stage and proceed to optimize the geometry using SDS guidance. Additionally, we leverage VSD guidance to enhance the texture, ensuring that the generated 3D model is visually appealing and realistic. As a result, this process enables the generation of significantly higher-quality 3D models." + } + ] + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 319, + 123, + 526, + 192 + ], + "blocks": [ + { + "bbox": [ + 305, + 70, + 547, + 117 + ], + "lines": [ + { + "bbox": [ + 305, + 70, + 547, + 117 + ], + "spans": [ + { + "bbox": [ + 305, + 70, + 547, + 117 + ], + "type": "text", + "content": "Table 2. Results of our user study. The rating score represents the average rating results for different methods, while the preference indicates the ratio of user preference for the results of each method with higher quality." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 319, + 123, + 526, + 192 + ], + "lines": [ + { + "bbox": [ + 319, + 123, + 526, + 192 + ], + "spans": [ + { + "bbox": [ + 319, + 123, + 526, + 192 + ], + "type": "table", + "html": "
MethodsRating Score ↑Preference(%) ↑
DreamFusion [21]2.094.55
Magic3D [14]2.446.27
TextMesh [30]1.734.91
Ours3.7484.27
", + "image_path": "f8502a21869d43ca15f8027257789acb1dacbed6587066ae8c59313c64c480f2.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 203, + 381, + 217 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 203, + 381, + 217 + ], + "spans": [ + { + "bbox": [ + 306, + 203, + 381, + 217 + ], + "type": "text", + "content": "4.4. User Study" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 222, + 547, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 222, + 547, + 426 + ], + "spans": [ + { + "bbox": [ + 304, + 222, + 547, + 426 + ], + "type": "text", + "content": "We conduct a user study with Fuxi Youling Crowdsourcing \\* for quantifying the subjective quality. Concretely, we evaluate the study on 22 prompts. At each evaluation time, the participants are shown the created 3D models by DreamFusion [21], Magic3D [14], TextMesh[30] and our method, and asked to rate them on a scale of 1 to 4. In total, we collect 1100 responses from 50 participants. The average rating scores for the different methods are shown in Table 2. Notably, our method receives the highest average rating, indicating that it produces the most visually high-fidelity 3D models. Furthermore, we calculate the preference ratios for each method based on the quality of the results. The analysis reveals that " + }, + { + "bbox": [ + 304, + 222, + 547, + 426 + ], + "type": "inline_equation", + "content": "84.27\\%" + }, + { + "bbox": [ + 304, + 222, + 547, + 426 + ], + "type": "text", + "content": " of the users consider the results of our method to be of higher quality than those generated by all competing methods. This finding further validates the effectiveness of our method in generating visually highfidelity 3D models." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 434, + 400, + 448 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 434, + 400, + 448 + ], + "spans": [ + { + "bbox": [ + 306, + 434, + 400, + 448 + ], + "type": "text", + "content": "4.5. Ablation Study" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 453, + 547, + 694 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 453, + 547, + 694 + ], + "spans": [ + { + "bbox": [ + 304, + 453, + 547, + 694 + ], + "type": "text", + "content": "Impact of the 3D synthesis Fusion Network: We introduce a 3D synthesis fusion network with the dynamic 3D synthesis strategy. Here we provide analysis and experiments to show its effectiveness. Given several text prompts, we employ 3D creation with different diffusion model guidance. The corresponding results are shown in Fig. 7. The first two columns show the results generated without the orthogonal-view diffusion model, which is equivalent to " + }, + { + "bbox": [ + 304, + 453, + 547, + 694 + ], + "type": "inline_equation", + "content": "\\lambda = 0" + }, + { + "bbox": [ + 304, + 453, + 547, + 694 + ], + "type": "text", + "content": ". Obvious Janus problems and inaccurate 3D representation can be found for the first two prompts. As for the creation of a cottage, we get similar shapes from different viewpoints. The next two columns demonstrate the result guided without the pre-trained diffusion model, which is equivalent to " + }, + { + "bbox": [ + 304, + 453, + 547, + 694 + ], + "type": "inline_equation", + "content": "\\lambda = \\infty" + }, + { + "bbox": [ + 304, + 453, + 547, + 694 + ], + "type": "text", + "content": ". While the orthogonal-view diffusion model can provide comprehensive structure constraints for 3D generation and restrain multi-face issues, we can see local distortion and unreasonable roughness. The last two columns demonstrate the results with our 3D synthesis fusion network, which means " + }, + { + "bbox": [ + 304, + 453, + 547, + 694 + ], + "type": "inline_equation", + "content": "\\lambda = 1" + }, + { + "bbox": [ + 304, + 453, + 547, + 694 + ], + "type": "text", + "content": ". We get complete and smooth meshes and eliminate the possible Janus problem." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 317, + 702, + 432, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 702, + 432, + 713 + ], + "spans": [ + { + "bbox": [ + 317, + 702, + 432, + 713 + ], + "type": "text", + "content": "* https://fuxi.163.com/solution/data" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 295, + 748, + 315, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 758 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 758 + ], + "type": "text", + "content": "4955" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 51, + 70, + 282, + 241 + ], + "blocks": [ + { + "bbox": [ + 51, + 70, + 282, + 241 + ], + "lines": [ + { + "bbox": [ + 51, + 70, + 282, + 241 + ], + "spans": [ + { + "bbox": [ + 51, + 70, + 282, + 241 + ], + "type": "image", + "image_path": "3c1a79e6fb43a75a12aaff8ef95cc67180e5f8914ef5989464d27fdf57852e19.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 51, + 243, + 282, + 318 + ], + "blocks": [ + { + "bbox": [ + 51, + 243, + 282, + 318 + ], + "lines": [ + { + "bbox": [ + 51, + 243, + 282, + 318 + ], + "spans": [ + { + "bbox": [ + 51, + 243, + 282, + 318 + ], + "type": "image", + "image_path": "4cd282259356fea7311a49bde6920368471e53357565a9722651718ec2e0f6ef.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 320, + 287, + 342 + ], + "lines": [ + { + "bbox": [ + 46, + 320, + 287, + 342 + ], + "spans": [ + { + "bbox": [ + 46, + 320, + 287, + 342 + ], + "type": "text", + "content": "Figure 7. Ablation study on the impact of the 3D synthesis fusion network." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 354, + 289, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 354, + 289, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 354, + 289, + 713 + ], + "type": "text", + "content": "Quantitative results are depicted in Table 3, which demonstrate the effectiveness of the 3D synthesis fusion network. Impact of Varying Number of View Supervision: Here we present an ablation study on the effects of varying the number of view supervisions on our orthogonal-view diffusion model. For two-view supervision, we randomly mask the opposite viewpoints in the rendered images during training, resulting in the optimization of the 3D representation using only two views. In the case of three-view supervision, we randomly mask one viewpoint, allowing the optimization of the 3D representation using three views. The corresponding results are shown in Figure 8. Generally, supervising the 3D creation with our orthogonal-view diffusion model can mitigate Janus problems. However, creating 3D models with two-view supervision may still suffer from slight multi-face problems, such as two barrels in a tank or multiple faces in a squirrel and a horse. With three-view supervision, these problems are alleviated to some extent. Nevertheless, there are still local distortions in generated 3D models, such as unexpected sags or bulges. These issues can be addressed by using the intact four-view supervision of our orthogonal-view diffusion model. Quantitative results are also presented in Table 3, illustrating the impact of reducing the number of view supervision signals during training. It is evident that there is a decrease in performance observed in this scenario. This highlights the significance of optimizing a 3D representation using a set of four orthogonal-view images, as it provides sufficient geometric structure priors for ensuring consistent 3D content creation." + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 313, + 72, + 542, + 311 + ], + "blocks": [ + { + "bbox": [ + 313, + 72, + 542, + 311 + ], + "lines": [ + { + "bbox": [ + 313, + 72, + 542, + 311 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 542, + 311 + ], + "type": "image", + "image_path": "4ef702f17ba7668fb5bcbfc6673da021817437f796bb128d9ed16090ef46e3cb.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 306, + 315, + 545, + 337 + ], + "lines": [ + { + "bbox": [ + 306, + 315, + 545, + 337 + ], + "spans": [ + { + "bbox": [ + 306, + 315, + 545, + 337 + ], + "type": "text", + "content": "Figure 8. Ablation study on the impact of varying number of view supervision." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 307, + 388, + 554, + 483 + ], + "blocks": [ + { + "bbox": [ + 306, + 357, + 545, + 380 + ], + "lines": [ + { + "bbox": [ + 306, + 357, + 545, + 380 + ], + "spans": [ + { + "bbox": [ + 306, + 357, + 545, + 380 + ], + "type": "text", + "content": "Table 3. Ablation study results using CLIP and FID. " + }, + { + "bbox": [ + 306, + 357, + 545, + 380 + ], + "type": "inline_equation", + "content": "\\downarrow" + }, + { + "bbox": [ + 306, + 357, + 545, + 380 + ], + "type": "text", + "content": " indicates the lower the better, and " + }, + { + "bbox": [ + 306, + 357, + 545, + 380 + ], + "type": "inline_equation", + "content": "\\uparrow" + }, + { + "bbox": [ + 306, + 357, + 545, + 380 + ], + "type": "text", + "content": " indicates the higher the better." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 307, + 388, + 554, + 483 + ], + "lines": [ + { + "bbox": [ + 307, + 388, + 554, + 483 + ], + "spans": [ + { + "bbox": [ + 307, + 388, + 554, + 483 + ], + "type": "table", + "html": "
Impact of the 3D synthesis Fusion Network
Metricsw/o Orthogonal-vieww/o Pre-trainedFull
CLIP ↑31.1332.0532.55
FID ↓372.86369.73342.62
Impact of Varying Number of View Supervision
MetricsTwo-viewThree-viewFour-view
CLIP ↑30.9230.4831.47
FID ↓399.58373.04345.72
", + "image_path": "ba05ab6dd04e92a9bad7a6d9b7fcd25879214cd3c4fabd288ca9d982fd7c3d6b.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 502, + 378, + 514 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 502, + 378, + 514 + ], + "spans": [ + { + "bbox": [ + 306, + 502, + 378, + 514 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 526, + 547, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 526, + 547, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 526, + 547, + 712 + ], + "type": "text", + "content": "In this paper, we present EfficientDreamer, a novel approach for high-fidelity and robust 3D creation using orthogonal-view diffusion priors. We introduce a novel diffusion model that can create composite images comprising sub-images from different orthogonal viewpoints. These composite images provide important geometric structure priors for 3D model generation by showcasing an object from multiple perspectives. We enhance the text-to-3D generation framework by leveraging the orthogonal-view diffusion model. Additionally, we introduce a 3D synthesis fusion network that integrates the orthogonal-view diffusion prior with the pre-trained diffusion prior. We also introduce a dynamic 3D synthesis strategy to balance the guidance of these two diffusion models during the SDS optimization. Our method effectively addresses the challenging Janus problem while ensuring the fidelity of the generated 3D models." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4956" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "type": "text", + "content": "[1] Mohammadreza Armandpour, Huangjie Zheng, Ali Sadeghian, Amir Sadeghian, and Mingyuan Zhou. Re-imagine the Negative Prompt Algorithm: Transform 2D Diffusion into 3D, alleviate Janus problem and Beyond. arXiv preprint arXiv:2304.04968, 2023. 2, 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 148, + 288, + 203 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 148, + 288, + 203 + ], + "spans": [ + { + "bbox": [ + 53, + 148, + 288, + 203 + ], + "type": "text", + "content": "[2] Yogesh Balaji, Seungjun Nah, Xun Huang, Arash Vahdat, Jiaming Song, Karsten Kreis, Miika Aittala, Timo Aila, Samuli Laine, Bryan Catanzaro, et al. edifi: Text-to-image diffusion models with an ensemble of expert denoisers. arXiv preprint arXiv:2211.01324, 2022. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 205, + 288, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 205, + 288, + 258 + ], + "spans": [ + { + "bbox": [ + 53, + 205, + 288, + 258 + ], + "type": "text", + "content": "[3] Rui Chen, Yongwei Chen, Ningxin Jiao, and Kui Jia. Fantasia3d: Disentangling geometry and appearance for high-quality text-to-3d content creation. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 22246–22256, 2023. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 261, + 288, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 261, + 288, + 327 + ], + "spans": [ + { + "bbox": [ + 53, + 261, + 288, + 327 + ], + "type": "text", + "content": "[4] Matt Deitke, Dustin Schwenk, Jordi Salvador, Luca Weihs, Oscar Michel, Eli VanderBilt, Ludwig Schmidt, Kiana Ehsani, Aniruddha Kembhavi, and Ali Farhadi. Objverse: A universe of annotated 3d objects. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13142-13153, 2023. 2, 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 328, + 288, + 384 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 328, + 288, + 384 + ], + "spans": [ + { + "bbox": [ + 53, + 328, + 288, + 384 + ], + "type": "text", + "content": "[5] Sara Fridovich-Keil, Alex Yu, Matthew Tancik, Qinhong Chen, Benjamin Recht, and Angjoo Kanazawa. Plenoxels: Radiance fields without neural networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5501–5510, 2022. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 384, + 288, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 384, + 288, + 430 + ], + "spans": [ + { + "bbox": [ + 53, + 384, + 288, + 430 + ], + "type": "text", + "content": "[6] Silvano Galliani, Katrin Lasinger, and Konrad Schindler. Massively parallel multiview stereopsis by surface normal diffusion. In Proceedings of the IEEE International Conference on Computer Vision, pages 873-881, 2015. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 431, + 288, + 485 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 431, + 288, + 485 + ], + "spans": [ + { + "bbox": [ + 53, + 431, + 288, + 485 + ], + "type": "text", + "content": "[7] Xiaodong Gu, Zhiwen Fan, Siyu Zhu, Zuozhuo Dai, Feitong Tan, and Ping Tan. Cascade cost volume for high-resolution multi-view stereo and stereo matching. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2495-2504, 2020. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 487, + 288, + 531 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 487, + 288, + 531 + ], + "spans": [ + { + "bbox": [ + 53, + 487, + 288, + 531 + ], + "type": "text", + "content": "[8] Yukun Huang, Jianan Wang, Yukai Shi, Xianbiao Qi, Zheng-Jun Zha, and Lei Zhang. Dreamtime: An improved optimization strategy for text-to-3d content creation. arXiv preprint arXiv:2306.12422, 2023. 4" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 533, + 288, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 533, + 288, + 586 + ], + "spans": [ + { + "bbox": [ + 53, + 533, + 288, + 586 + ], + "type": "text", + "content": "[9] Ajay Jain, Ben Mildenhall, Jonathan T Barron, Pieter Abbeel, and Ben Poole. Zero-shot text-guided object generation with dream fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 867-876, 2022. 1, 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 589, + 287, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 589, + 287, + 633 + ], + "spans": [ + { + "bbox": [ + 48, + 589, + 287, + 633 + ], + "type": "text", + "content": "[10] Rasmus Jensen, Anders Dahl, George Vogiatzis, Engin Tola, and Henrik Aanæs. Large scale multi-view stereopsis evaluation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 406-413, 2014. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 635, + 287, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 635, + 287, + 678 + ], + "spans": [ + { + "bbox": [ + 48, + 635, + 287, + 678 + ], + "type": "text", + "content": "[11] Tuomas Kynkänniemi, Tero Karras, Miika Aittala, Timo Aila, and Jaakko Lehtinen. The role of imagenet classes in fr\\echet inception distance. arXiv preprint arXiv:2203.06026, 2022. 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "type": "text", + "content": "[12] Han-Hung Lee and Angel X Chang. Understanding pure clip guidance for voxel grid nerf models. arXiv preprint arXiv:2209.15172, 2022. 1" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 713 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 117 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 117 + ], + "type": "text", + "content": "[13] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. arXiv preprint arXiv:2301.12597, 2023. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 118, + 547, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 547, + 184 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 547, + 184 + ], + "type": "text", + "content": "[14] Chen-Hsuan Lin, Jun Gao, Luming Tang, Towaki Takikawa, Xiaohui Zeng, Xun Huang, Karsten Kreis, Sanja Fidler, Ming-Yu Liu, and Tsung-Yi Lin. Magic3d: High-resolution text-to-3d content creation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 300–309, 2023. 1, 3, 6, 7" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 186, + 545, + 218 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 186, + 545, + 218 + ], + "spans": [ + { + "bbox": [ + 307, + 186, + 545, + 218 + ], + "type": "text", + "content": "[15] Tiange Luo, Chris Rockwell, Honglak Lee, and Justin Johnson. Scalable 3d captioning with pretrained models. arXiv preprint arXiv:2306.07279, 2023. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 220, + 545, + 275 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 220, + 545, + 275 + ], + "spans": [ + { + "bbox": [ + 307, + 220, + 545, + 275 + ], + "type": "text", + "content": "[16] Gal Metzer, Elad Richardson, Or Patashnik, Raja Giryes, and Daniel Cohen-Or. Latent-nerf for shape-guided generation of 3d shapes and textures. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12663–12673, 2023. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 276, + 545, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 276, + 545, + 330 + ], + "spans": [ + { + "bbox": [ + 307, + 276, + 545, + 330 + ], + "type": "text", + "content": "[17] Oscar Michel, Roi Bar-On, Richard Liu, Sagie Benaim, and Rana Hanocka. Text2mesh: Text-driven neural stylization for meshes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13492-13502, 2022. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 331, + 545, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 331, + 545, + 385 + ], + "spans": [ + { + "bbox": [ + 307, + 331, + 545, + 385 + ], + "type": "text", + "content": "[18] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 388, + 545, + 432 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 388, + 545, + 432 + ], + "spans": [ + { + "bbox": [ + 307, + 388, + 545, + 432 + ], + "type": "text", + "content": "[19] Nasir Mohammad Khalid, Tianhao Xie, Eugene Belilovsky, and Tiberiu Popa. Clip-mesh: Generating textured meshes from text using pretrained image-text models. In SIGGRAPH Asia 2022 conference papers, pages 1-8, 2022. 1, 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 433, + 545, + 477 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 433, + 545, + 477 + ], + "spans": [ + { + "bbox": [ + 307, + 433, + 545, + 477 + ], + "type": "text", + "content": "[20] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM Transactions on Graphics (ToG), 41(4):1-15, 2022. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 478, + 545, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 478, + 545, + 510 + ], + "spans": [ + { + "bbox": [ + 307, + 478, + 545, + 510 + ], + "type": "text", + "content": "[21] Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. In ICLR, 2023. 1, 3, 6, 7" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 512, + 545, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 512, + 545, + 578 + ], + "spans": [ + { + "bbox": [ + 307, + 512, + 545, + 578 + ], + "type": "text", + "content": "[22] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 1, 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 579, + 545, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 579, + 545, + 622 + ], + "spans": [ + { + "bbox": [ + 307, + 579, + 545, + 622 + ], + "type": "text", + "content": "[23] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022. 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 624, + 545, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 624, + 545, + 679 + ], + "spans": [ + { + "bbox": [ + 307, + 624, + 545, + 679 + ], + "type": "text", + "content": "[24] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 1, 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 681, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 681, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 681, + 545, + 713 + ], + "type": "text", + "content": "[25] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans," + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4957" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 66, + 72, + 287, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 72, + 287, + 106 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 287, + 106 + ], + "type": "text", + "content": "et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in Neural Information Processing Systems, 35:36479-36494, 2022. 1, 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 107, + 288, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 107, + 288, + 173 + ], + "spans": [ + { + "bbox": [ + 48, + 107, + 288, + 173 + ], + "type": "text", + "content": "[26] Johannes L Schonberger, Enliang Zheng, Jan-Michael Frahm, and Marc Pollefeys. Pixelwise view selection for unstructured multi-view stereo. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part III 14, pages 501-518. Springer, 2016. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 175, + 288, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 175, + 288, + 239 + ], + "spans": [ + { + "bbox": [ + 48, + 175, + 288, + 239 + ], + "type": "text", + "content": "[27] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. Advances in Neural Information Processing Systems, 35:25278-25294, 2022. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 241, + 288, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 241, + 288, + 296 + ], + "spans": [ + { + "bbox": [ + 48, + 241, + 288, + 296 + ], + "type": "text", + "content": "[28] Tianchang Shen, Jun Gao, Kangxue Yin, Ming-Yu Liu, and Sanja Fidler. Deep marching tetrahedra: a hybrid representation for high-resolution 3d shape synthesis. Advances in Neural Information Processing Systems, 34:6087-6101, 2021. 2, 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 297, + 288, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 297, + 288, + 342 + ], + "spans": [ + { + "bbox": [ + 48, + 297, + 288, + 342 + ], + "type": "text", + "content": "[29] Jiaxiang Tang, Hang Zhou, Xiaokang Chen, Tianshu Hu, Errui Ding, Jingdong Wang, and Gang Zeng. Delicate textured mesh recovery from nerf via adaptive surface refinement. arXiv preprint arXiv:2303.02091, 2023. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 342, + 288, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 342, + 288, + 385 + ], + "spans": [ + { + "bbox": [ + 48, + 342, + 288, + 385 + ], + "type": "text", + "content": "[30] Christina Tsalicoglou, Fabian Manhardt, Alessio Tonioni, Michael Niemeyer, and Federico Tombari. Textmesh: Generation of realistic 3d meshes from text prompts. 2024. 1, 3, 6, 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 388, + 288, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 388, + 288, + 442 + ], + "spans": [ + { + "bbox": [ + 48, + 388, + 288, + 442 + ], + "type": "text", + "content": "[31] Fangjinhua Wang, Silvano Galliani, Christoph Vogel, Pablo Speciale, and Marc Pollefeys. Patchmatchnet: Learned multi-view patchmatch stereo. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14194-14203, 2021. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 444, + 288, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 444, + 288, + 487 + ], + "spans": [ + { + "bbox": [ + 48, + 444, + 288, + 487 + ], + "type": "text", + "content": "[32] Peng Wang, Lingjie Liu, Yuan Liu, Christian Theobalt, Taku Komura, and Wenping Wang. Neus: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. arXiv preprint arXiv:2106.10689, 2021. 2, 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 489, + 288, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 489, + 288, + 533 + ], + "spans": [ + { + "bbox": [ + 48, + 489, + 288, + 533 + ], + "type": "text", + "content": "[33] Zhengyi Wang, Cheng Lu, Yikai Wang, Fan Bao, Chongxuan Li, Hang Su, and Jun Zhu. Prolificdreamer: High-fidelity and diverse text-to-3d generation with variational score distillation. arXiv preprint arXiv:2305.16213, 2023. 1, 3, 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 534, + 288, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 534, + 288, + 578 + ], + "spans": [ + { + "bbox": [ + 48, + 534, + 288, + 578 + ], + "type": "text", + "content": "[34] Qingshan Xu and Wenbing Tao. Multi-scale geometric consistency guided multi-view stereo. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5483-5492, 2019. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 579, + 288, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 579, + 288, + 634 + ], + "spans": [ + { + "bbox": [ + 48, + 579, + 288, + 634 + ], + "type": "text", + "content": "[35] Jiayu Yang, Jose M Alvarez, and Miaomiao Liu. Nonparametric depth distribution modelling based depth inference for multi-view stereo. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8626-8634, 2022. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 635, + 288, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 635, + 288, + 679 + ], + "spans": [ + { + "bbox": [ + 48, + 635, + 288, + 679 + ], + "type": "text", + "content": "[36] Yao Yao, Zixin Luo, Shiwei Li, Tian Fang, and Long Quan. Mvsnet: Depth inference for unstructured multi-view stereo. In Proceedings of the European conference on computer vision (ECCV), pages 767-783, 2018." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 681, + 288, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 681, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 681, + 288, + 713 + ], + "type": "text", + "content": "[37] Yao Yao, Zixin Luo, Shiwei Li, Tianwei Shen, Tian Fang, and Long Quan. Recurrent mvsnet for high-resolution multi-view stereo depth inference. In Proceedings of" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 329 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "type": "text", + "content": "the IEEE/CVF conference on computer vision and pattern recognition, pages 5525-5534, 2019. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 96, + 545, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 96, + 545, + 150 + ], + "spans": [ + { + "bbox": [ + 307, + 96, + 545, + 150 + ], + "type": "text", + "content": "[38] Yao Yao, Zixin Luo, Shiwei Li, Jingyang Zhang, Yufan Ren, Lei Zhou, Tian Fang, and Long Quan. Blendedmvs: A large-scale dataset for generalized multi-view stereo networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1790-1799, 2020. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 152, + 545, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 152, + 545, + 194 + ], + "spans": [ + { + "bbox": [ + 307, + 152, + 545, + 194 + ], + "type": "text", + "content": "[39] Lior Yariv, Jiatao Gu, Yoni Kasten, and Yaron Lipman. Volume rendering of neural implicit surfaces. Advances in Neural Information Processing Systems, 34:4805-4815, 2021. 2, 5" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 197, + 545, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 197, + 545, + 229 + ], + "spans": [ + { + "bbox": [ + 307, + 197, + 545, + 229 + ], + "type": "text", + "content": "[40] Kai Zhang, Gernot Riegler, Noah Snavely, and Vladlen Koltun. Nerf++: Analyzing and improving neural radiance fields. arXiv preprint arXiv:2010.07492, 2020. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 231, + 545, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 231, + 545, + 274 + ], + "spans": [ + { + "bbox": [ + 307, + 231, + 545, + 274 + ], + "type": "text", + "content": "[41] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 3836-3847, 2023. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 276, + 545, + 329 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 276, + 545, + 329 + ], + "spans": [ + { + "bbox": [ + 307, + 276, + 545, + 329 + ], + "type": "text", + "content": "[42] Enliang Zheng, Enrique Dunn, Vladimir Jojic, and Jan-Michael Frahm. Patchmatch based joint view selection and depthmap estimation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 1510-1517, 2014. 2" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4958" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/EfficientSAM_ Leveraged Masked Image Pretraining for Efficient Segment Anything/f1e649e9-e5d4-48b3-8018-af1f58065ff2_content_list.json b/2024/EfficientSAM_ Leveraged Masked Image Pretraining for Efficient Segment Anything/f1e649e9-e5d4-48b3-8018-af1f58065ff2_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..bc1308f7b8034434a56e1952f6d31a5bd7723c64 --- /dev/null +++ b/2024/EfficientSAM_ Leveraged Masked Image Pretraining for Efficient Segment Anything/f1e649e9-e5d4-48b3-8018-af1f58065ff2_content_list.json @@ -0,0 +1,1561 @@ +[ + { + "type": "text", + "text": "EfficientSAM: Leveraged Masked Image Pretraining for Efficient Segment Anything", + "text_level": 1, + "bbox": [ + 107, + 128, + 861, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yunyang Xiong, Bala Varadarajan*, Lemeng Wu*, Xiaoyu Xiang, Fanyi Xiao, Chenchen Zhu, Xiaoliang Dai, Dilin Wang, Fei Sun, Forrest Iandola, Raghuraman Krishnamoorthi, Vikas Chandra Meta AI Research", + "bbox": [ + 94, + 203, + 872, + 255 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "https://yformer.github.io/efficient-sam/", + "bbox": [ + 297, + 258, + 656, + 273 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 308, + 313, + 325 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Segment Anything Model (SAM) has emerged as a powerful tool for numerous vision applications. A key component that drives the impressive performance for zero-shot transfer and high versatility is a super large Transformer model trained on the extensive high-quality SA-1B dataset. While beneficial, the huge computation cost of SAM model has limited its applications to wider real-world applications. To address this limitation, we propose EfficientSAMs, lightweight SAM models that exhibits decent performance with largely reduced complexity. Our idea is based on leveraging masked image pretraining, SAMI, which learns to reconstruct features from SAM image encoder for effective visual representation learning. Further, we take SAMI-pretrained light-weight image encoders and mask decoder to build EfficientSAMs, and finetune the models on SA-1B for segment anything task. We perform evaluations on multiple vision tasks including image classification, object detection, instance segmentation, and semantic segmentation, and find that our proposed pretraining method, SAMI, consistently outperforms other masked image pretraining methods. On segment anything task such as zero-shot instance segmentation, our EfficientSAMs with SAMI-pretrained lightweight image encoders perform favorably with a significant gain (e.g., $\\sim 4$ AP on COCO/LVIS) over other fast SAM models. Our EfficientSAM code and models are available at here.", + "bbox": [ + 75, + 340, + 473, + 720 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 746, + 209, + 762 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Segment Anything Model (SAM) [31] has been very successful in the vision field, achieving state-of-the-art performance in a variety of image segmentation tasks such as zero-shot edge detection[1, 31], zero-shot object proposal generation[31, 54], and zero-shot instance segmentation[31], and many other real-world applications[24, 37, 41, 50-52]. The key feature of SAM is a prompt-based Vision Trans", + "bbox": [ + 75, + 771, + 470, + 878 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/92e92a677f7a1fde0e71de11ef6055bfa06f71d67f0d8eadb98663cb69831ba3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 506, + 311, + 888, + 378 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/7a852e0935c384ff68ee367b16efca629b1a790931f382b6c3f3626e5cbe12bb.jpg", + "image_caption": [ + "Figure 1. The comparative analysis result. (Top) The overview of EfficientSAM model by taking a well-pretrained light-weight image encoder for instance segmentation with largely reduced complexity. (Bottom) Throughput/Parameter/Performance comparison of EfficientSAM, MobileSAM, FastSAM, and SAM for zero-shot instance segmentation on COCO. We benchmark throughput (image per second) of all models on a single NVIDIA A100 with one box prompt. The input image resolution is $1024 \\times 1024$ . Our EfficientSAMs outperform MobileSAM and FastSAM by a large margin, $\\sim 4$ AP, with comparable complexity. Our EfficientSAM-S reduces the inference time of SAM by $\\sim 20x$ and the parameter size by $\\sim 20x$ with a small performance drop, 44.4 AP vs 46.5 AP." + ], + "image_footnote": [], + "bbox": [ + 509, + 387, + 880, + 542 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "former (ViT)[19]model trained on a large-scale visual dataset with more than 1B masks from 11M images, SA-1B[31], which allows segmenting any object on a given image. This ability of Segment Anything makes SAM a foundation model in vision and enables its applications even beyond vision.", + "bbox": [ + 496, + 748, + 893, + 824 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Despite the forgoing advantages, the model of SAM turns out to be a major efficiency bottleneck for practical deployment since the architecture of SAM, especially, the image encoder (e.g., ViT-H) is very expensive. Note that ViT-H image encoder in SAM has 632M parameters while the", + "bbox": [ + 496, + 825, + 895, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Joint second author.", + "bbox": [ + 94, + 887, + 209, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "16111", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "prompt-based decoder only take 3.87M parameters. As a result, it leads to high computation and memory costs when using SAM to perform segment anything tasks in practice, which makes it challenging for real-time applications.", + "bbox": [ + 75, + 90, + 470, + 151 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address this challenge, several recent works have proposed strategies that avoid incurring the huge cost when applying SAM for the prompt-based instance segmentation. For example, [68] suggests distilling the knowledge from the default ViT-H image encoder to a tiny ViT image encoder. In [71], the computation cost can be reduced with a real-time CNN-based architecture for Segment Anything task.", + "bbox": [ + 75, + 152, + 470, + 258 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we propose using a well-pretrained lightweight ViT image encoder (e.g., ViT-Tiny/-Small[53]) to reduce the complexity of SAM while maintain decent performance. Our method, SAM-leveraged masked image pertraining (SAMI), produces desired pretrained lightweight ViT backbones for segment anything task. This is achieved by leveraging the celebrated MAE[26] pretraining method with SAM model to obtain high-quality pretrained ViT encoders. Specifically, our SAMI makes use of SAM encoder, ViT-H, to generate feature embedding and train a masked image model with lightweight encoders to reconstruct features from ViT-H of SAM instead of image patches. This leads to generalized ViT backbones, which can be used for downstream tasks such as image classification, object detection, and segment anything. Then we finetune the pretrained lightweight encoders with SAM decoders for segment anything task[31].", + "bbox": [ + 75, + 258, + 472, + 517 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To evaluate our method, we consider a transfer learning setting for masked image pretraining, where models are first pretrained with a reconstructive loss on ImageNet with image resolution $224 \\times 224$ , and then finetuned on target tasks using supervised data. Our SAMI learns lightweight encoders that generalize well. With SAMI pretraining, we can train models like ViT-Tiny/-Small/-Base on ImageNet-1K with improved generalization performance. For a ViT-Small model, we achieve $82.7\\%$ top-1 accuracy when finetuned on ImageNet-1K with 100 epochs, which outperforms other state-of-the-art image pretraining baselines. We also finetune our pretrained models on object detection, instance segmentation, and semantic segmentation. Across all these tasks, our pretraining method achieves better results than other pretraining baselines, and more importantly, we observe significant gains for small models. Further, we evaluate our models on Segment Anything task. On zero-shot instance segmentation, our model performs well compared to recent lightweight SAM methods, including FastSAM, by a margin of 4.1 /5.2 AP on COCO/LVIS.", + "bbox": [ + 75, + 517, + 470, + 818 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our main contribution can be summarized as follows:", + "bbox": [ + 96, + 820, + 450, + 834 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "- We propose a SAM-leveraged masked image pretraining framework called SAMI, which trains the model to reconstruct features from SAM ViT-H image encoder. We show that this can substantially improve the", + "bbox": [ + 96, + 839, + 470, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "performance of image masked pretraining method.", + "bbox": [ + 532, + 90, + 870, + 106 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We demonstrate that SAMI-pretrained backbones can generalize well to many tasks including image classification, object detection, and semantic segmentation.", + "- We deliver EfficientSAMs, light-weight SAM models with state-of-the-art quality-efficiency trade-offs (Fig. 1), which is complementary to SAM for practical deployment." + ], + "bbox": [ + 519, + 106, + 893, + 212 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 223, + 640, + 239 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We briefly review relevant works on segment anything model, vision transformers, knowledge distillation, and masked image pretraining.", + "bbox": [ + 498, + 250, + 893, + 295 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Segment Anything Model", + "text_level": 1, + "bbox": [ + 500, + 301, + 732, + 318 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "SAM[31] has been considered as a milestone vision foundation model, which can segment any object in the image based on interaction prompts. SAM has shown remarkable zero-shot transfer performance and high versatility for many vision tasks including a variety of segmentation application[7, 8, 10, 17], in-painting[67], image restoration[29], image editing[21], image shadow removal[69], object tracking[14, 65], and 3D object reconstruction[49]. There are many other works attempting to generalize SAM to real-world scenarios, including medical image segmentation[41], camouflaged object detection[51], transparent object detection[24], concept-based explanation[50], semantic communication[52], and helping people with visual impairments[37]. Due to its wide real-world applications, practical deployment of SAM has also gained increasing attention. Several recent works including [68, 71] have proposed strategies to reduce the computation costs of SAM. FastSAM[68] develops a CNN-based architecture, YOLOv8-seg[30], to segment all objects in an image for efficiency improvement. MobileSAM[71] presents a decoupled distillation for obtaining a lightweight image encoder of SAM. Our work focuses on dealing with this efficiency issue for practical deployment of SAM.", + "bbox": [ + 496, + 325, + 893, + 657 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Vision Transformers", + "text_level": 1, + "bbox": [ + 500, + 665, + 692, + 680 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "ViTs [19] have achieved impressive performance in vision applications[5, 20, 26, 34, 39, 44]. ViTs demonstrate the advantages of and generalization over their CNN counterparts[26]. There are also a number of works on efficient ViTs for deployment. Smaller ViTs such as ViT-Small/DeiT-Small and ViT-Tiny/DeiT-Tiny are introduced in [53] for complementing ViT-Huge, ViT-Large, and ViT-Base in [19]. Motivated by the ability of convolution to capture local information with reduced parameters and computation costs, MobileViT[42] explore combining ViT with convolutions, which outperforms light-weight CNN models such as MobileNet-v2/v3[32, 48] with better task-level generalization properties and reduced memory size and computation cost. This trick has been used in many follow-up works", + "bbox": [ + 496, + 688, + 893, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "16112", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "including LeViT[22], EfficientFormer[35], Next-ViT[33], Tiny-ViT[61], Castling-ViT[66], EfficientViT[38]. This line of progress for designing efficient ViTs is orthogonal to our EfficientSAM work towards building efficient SAM.", + "bbox": [ + 75, + 90, + 470, + 151 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.3. Knowledge Distillation", + "text_level": 1, + "bbox": [ + 76, + 162, + 289, + 178 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Knowledge distillation (KD) is a technique to improve the performance of deep learning models without changing their architectures. [27] is a pioneering work to distill the dark knowledge from a larger teacher model to a smaller student model. The learning of a student model is supervised by the hard labels and the soft labels from a teacher model. This practice is followed by multiple works which aim to make better use of soft labels to transfer more knowledge. In [64], the distillation method decouples representation learning and classification. Decoupled knowledge distillation[70] separates the classical KD loss into two parts, target class knowledge distillation and non-target class knowledge distillation, which improves the effectiveness and flexibility of knowledge transfer. Another line of work is to transfer knowledge from intermediate features. FitNet [47] is a pioneering work by distilling the semantic information from the teacher model's intermediate feature directly. In [60], a self-supervised teaching assistant (SSTA) is introduced to guide the learning of a ViT-based student model with a supervised teacher together. [2] studies the potential of knowledge distillation from pre-training MAE model by aligning the intermediate features between the larger MAE teacher model and smaller MAE student model.", + "bbox": [ + 75, + 186, + 472, + 532 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.4. Masked Image Pretraining", + "text_level": 1, + "bbox": [ + 76, + 544, + 320, + 561 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Self-supervised pretraining approaches [6] have attracted significant attention in computer vision. One line of work is contrastive learning methods[9, 11, 57, 62], which learn augmentation in-variance by imposing high similarity between different augmented views of a given image. While the learned representation shows good properties such as high linear separability, contrastive learning methods relies on strong augmentation and negative sampling. Another interesting line of work is masked image modeling (MIM), which helps models learn meaningful representations by reconstructing masked image patches. The MIM pioneering works focus on using denoising autoencoders[56] and context encoders[43] to train vision Transformer with masked prediction objectives. There are various promising works on using MIM for self-supervised image pretraining. BEiT[3] is the first one that adopts MIM for ViT pretraining to predict visual tokens. In BEiTv2[44], a semantic-rich image tokenizer is utilized for a better reconstruction target. In MaskFeat[59], reconstructing the local gradient features generated from HOG descriptor leads to effective visual pretraining. In SimMIM[63] and MAE[26], directly reconstructing the pixel values of the masked image patches achieves effec", + "bbox": [ + 75, + 568, + 472, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "tive visual representation learning. There are MAE-based follow-up works that use large teacher models to guide MAE pretraining[2, 28, 60]. Our work is built on MAE and finds that leveraging MAE to reconstruct the features from SAM image encoder enables the pretraining to be highly effective.", + "bbox": [ + 496, + 90, + 893, + 167 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Approach", + "text_level": 1, + "bbox": [ + 500, + 181, + 609, + 199 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Preliminary", + "text_level": 1, + "bbox": [ + 500, + 207, + 629, + 223 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Masked Autoencoders. Masked Autoencoders (MAE) model has two components, an encoder and a decoder. Both encoder and decoder are built on Transformer layers[55]. MAE takes image tokens, i.e., non-overlapping patches from the input images, as input. These input tokens are grouped to unmasked tokens and masked tokens with a given masking ratio. The unmasked tokens will be kept for encoder to extract features and the masked tokens will be set as the learning targets of the MAE decoder that need to be reconstructed during self-supervised learning (MIM). MAE[26] adopts a high mask ratio (e.g., $75\\%$ ), which prevents information leakage (e.g., simply extrapolating masked pixels based on the neighbors) in the pretraining stage.", + "bbox": [ + 496, + 229, + 893, + 428 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. SAM-Leveraged Masked Image Pretraining", + "text_level": 1, + "bbox": [ + 498, + 438, + 872, + 454 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We now adapt MAE framework to obtain efficient image encoders for segment anything model. Motivated by the high versatility of SAM [31], we explore latent features from SAM image encoder as the reconstruction target to leverage MAE. Our method emphasizes transferring the knowledge embedded in SAM. Fig. 2 (top) illustrates an overview of the proposed SAM-leveraged masked image pretraining, SAMI. The encoder transforms the unmasked tokens into latent feature representation and the decoder reconstructs the representation of the masked tokens aided by the output feature embedding from the encoder. The representation learning of reconstruction is guided by latent features from SAM.", + "bbox": [ + 496, + 460, + 893, + 643 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Cross-Attention Decoder. With the supervision of SAM features, we observe that only masked tokens need to be reconstructed via decoder while the encoder's output can serve as anchors during the reconstruction. In the cross-attention decoder, queries come from masked tokens, and keys and values derive from both unmasked features from encoder and masked features. We merge the output features of masked tokens from cross-attention decoder and the output features of unmasked tokens from encoder for the MAE output embedding. Then, this combined features will be reordered to the original positions of input image tokens for the final MAE outputs.", + "bbox": [ + 496, + 643, + 893, + 824 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Linear Projection Head. We obtain the image output through our encoder and cross-attention decoder. Then we feed such features into a small project head for aligning the features from SAM image encoder. For simplicity, we just use a linear projection head to address the feature dimension", + "bbox": [ + 496, + 825, + 893, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "16113", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/4bda53427016eea619c4cdbf7e4079c640260ea75673c08a5f7cb216b9e50126.jpg", + "image_caption": [ + "Figure 2. The overview of EfficientSAM framework. Our proposed EfficientSAM contains two stages: SAMI pretraining (top) on ImageNet and SAM finetuning (bottom) on SA-1B. For SAMI pretraining, the masked autoencoder takes the feature embeddings from SAM image encoder as the reconstruction target. After SAMI pretraining, the decoder is discarded and the light-weight encoder is served as the image encoder of EfficientSAM for finetuning on SA-1B." + ], + "image_footnote": [], + "bbox": [ + 109, + 95, + 861, + 354 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "mismatch between the output of SAM image encoder and MAE.", + "bbox": [ + 75, + 446, + 470, + 474 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Reconstruction Loss. At each training iteration, SAMI consists of a feedforward feature extraction from SAM image encoder, and a feedforward and a backpropagation procedure of MAE. The outputs from SAM image encoder and MAE linear projection head are compared to compute the reconstruction loss.", + "bbox": [ + 75, + 477, + 470, + 566 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Let us denote the SAM image encoder as $f^{\\mathrm{sam}}$ , and the encoder and decoder of MAE as $g^{e}$ with weights $W_{e}$ and $g^{d}$ with weights $W_{d}$ , linear projection head as $h^{\\theta}$ with weights $W_{\\theta}$ respectively. Assume the input tokens are denoted as $\\{\\mathbf{x}_i\\}_{i=1}^N$ , where $N$ is the number of tokens. The input tokens are randomly grouped into the unmasked tokens, $\\{\\mathbf{x}_i\\}_{i \\in \\mathcal{U}}$ , the masked tokens $\\{\\mathbf{x}_i\\}_{i \\in \\mathcal{M}}$ with a given masked ratio. Let the feature reordering operator be $\\phi$ , and the merging operator be $\\oplus$ .", + "bbox": [ + 75, + 566, + 470, + 700 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our target features from SAM image encoder can be written as $f^{\\mathrm{sam}}(\\mathbf{x}) = f^{\\mathrm{sam}}(\\{\\mathbf{x}_i\\}_{i=1}^N)$ , the output from MAE encoder is $g^{e}(\\{\\mathbf{x}_{i}\\}_{i\\in \\mathcal{U}})$ , the decoder output is $g^{d}(\\{\\mathbf{x}_{i}\\}_{i\\in \\mathcal{M}})$ . The output from linear projection head is $f^{h}(\\mathbf{x}) = h^{\\theta}(\\phi(g^{e}\\{\\mathbf{x}_{i}\\}_{i\\in \\mathcal{U}} \\oplus g^{d}\\{\\mathbf{x}_{i}\\}_{i\\in \\mathcal{M}}))$ . Therefore, our target reconstruction loss can be formulated as,", + "bbox": [ + 75, + 703, + 470, + 792 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nL _ {W _ {e}, W _ {d}, W _ {\\theta}} = \\frac {1}{N} \\cdot \\sum_ {j = 1} ^ {N} \\left\\| f ^ {\\operatorname {s a m}} (\\mathbf {x}) - f ^ {h} (\\mathbf {x}) \\right\\| ^ {2}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 120, + 803, + 470, + 844 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $N$ is the number of input tokens, $\\| \\cdot \\|$ denotes a norm. We use $\\ell_2$ norm for reconstruction loss in our experiments. By minimizing the reconstruction loss, $L_{W_e,W_d,W_\\theta}$ ,", + "bbox": [ + 75, + 854, + 472, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "our encoder $g^{e}$ is optimized to serve as an image backbone to extract features as SAM image encoder. Our encoder, decoder, and linear projection head are optimized to learn context modeling ability from SAM image encoder. Optimizing the reconstruction loss on all tokens transfer the knowledge embedded in SAM.", + "bbox": [ + 496, + 446, + 893, + 536 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "SAMI for EfficientSAM. After pretraining, our encoder extract feature representations for various vision tasks and the decoder is discarded. In particular, to build efficient SAM models for the segmentation anything task, we take the SAMI-pretrained lightweight encoder such as ViT-Tiny and ViT-Small as the image encoder and the default mask decoder of SAM for our EfficientSAM, as illustrated in Fig. 2 (bottom). We finetune our EfficientSAM models on SA-1B dataset for the segment anything task. The overview of our EfficientSAM framework is illustrated in Fig. 2.", + "bbox": [ + 496, + 537, + 895, + 689 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 500, + 700, + 633, + 718 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1. Experimental Settings", + "text_level": 1, + "bbox": [ + 500, + 726, + 705, + 742 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Pretraining Datasets. Our masked image pretraining method, SAMI, is conducted on ImageNet-1K training set with 1.2M images. Following masked image pretraining [26], we do not use the label information. We use the SAM ViT-H image encoders from [31] to generate reconstruction features when pretraining our ViT models, ViT-Tiny, ViT-Small, and ViT-Base.", + "bbox": [ + 496, + 750, + 893, + 853 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Pretraining Implementation Details. Our ViT models are pretrained with a mean squared error (MSE) loss for reconstruction. We use a batch size of 4096, AdamW optimizer", + "bbox": [ + 496, + 854, + 893, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "16114", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "[40] with learning rate $2.4\\mathrm{e} - 3$ , $\\beta_{1} = 0.9$ , $\\beta_{2} = 0.95$ , weight decay 0.05, linear learning rate warm-up over the first 40 epochs, cosine learning rate decay to update our models. We only adopt random resize crop to $224\\times 224$ resolution, random horizontal flip, and normalization for data augmentation. The mask ratio is set to $75\\%$ and the decoder contains 8 Transformer blocks with 512 dimensions as in [26]. We pretrain SAMI for 400 epochs using PyTorch framework on V100 machines. For reference, 1600-epoch pretraining is required for MAE[26].", + "bbox": [ + 75, + 90, + 470, + 242 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Downstream Tasks/Datasets/Models. Tasks and Datasets. We first consider three benchmarking datasets and several representative vision tasks to demonstrate the superiority of the proposed SAMI, including image classification on ImageNet dataset [16] with 1.2 million training and 50K validation images; Object detection and instance segmentation on COCO dataset [36] with 118K training and 5K validation images; Semantic segmentation on ADE20K dataset [72] with 20K/2K/3K images for training, validation, and testing, respectively. Then, we consider segment anything task to further show the advantages of our proposed SAMI. We finetune our pretrained lightweight image encoders for SAM on SA-1B dataset [31] with more than 1B masks from 11M high-resolution images, and test interactive instance segmentation and zero-shot instance segmentation ability of our EfficientSAMs on COCO and LVIS [23]. Models. We discard the decoder of SAMI while keeping the encoder as backbone to extract features for different tasks as in MAE[26]. We apply our well-pretrained ViT backbones for different tasks including ViTs for the classification, ViTDet [34] for the detection and instance segmentation, Mask2former [13] for the semantic segmentation task, and SAM for segment anything.", + "bbox": [ + 75, + 246, + 472, + 595 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Finetuning Settings. For the classification task, we use the AdamW optimizer with $\\beta_{1} = 0.9$ , $\\beta_{2} = 0.999$ , weight decay 0.05 to finetune ViTs for 100 epochs using 32 V100 GPUs, with each GPU having a batch size of 32. The initial learning rate is 1e-3 with first 5 epochs for linear warm-up and decays to zero by a cosine learning rate scheduler. We set the layer-wise learning rate decay factor to 0.75 for ViT-Small and ViT-Base. We do not apply layer-wise learning rate decay for ViT-Tiny. For data augmentation, we adopt RandAugment [15] and set label smoothing to 0.1, mixup to 0.8. For the detection and instance segmentation task, We follow the ViTDet [34] framework by adapting ViT backbones to a simple feature pyramid, for object detection and instance segmentation. We adopt AdamW optimizer with momentum $\\beta_{1} = 0.9$ , $\\beta_{2} = 0.999$ and weight decay 0.1 to train models on COCO. All models are trained on 64 V100 GPUs for 100 epochs with each GPU having 1 batch size with image resolution $1024 \\times 1024$ . The initial learning rate is $2e - 4$ , linearly warmed up for the first 10 epochs, and decayed to 0 by a cosine learning rate schedule. Models are trained", + "bbox": [ + 75, + 598, + 472, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "for 100 epochs. For the segmentation task, Our pretrained ViT models serve as the backbone of Mask2former [13], which is finetuned together with the segmentation layers on ADE20K. We adopt the AdamW optimizer with $\\beta_{1} = 0.9$ , $\\beta_{2} = 0.999$ , a mini-batch size of 16, a weight decay of 0.05, and an initial learning rate of 2e-4. The learning rate is decayed to 0 by a poly learning rate schedule. A learning rate multiplier is set to 0.1 for the backbone. The input image resolution is $512 \\times 512$ . Models are trained for 160K iterations using 8 V100 GPUs. For the segmentation anything task, Following [31], we take our pretrained lightweight ViT models, ViT-Tiny and ViT-Small, as the image encoder of SAM framework and finetune the encoder and decoder together of our EfficientSAM on SA-1B dataset for 5 epochs. We use the AdamW optimizer with a momentum, $(\\beta_{1} = 0.9$ , $\\beta_{2} = 0.999)$ , a mini-batch size of 128, and a initial lreanning rate of $4e - 4$ . The learning rate is decayed to 0 by a linear learning rate schedule. We set weight decay to 0.1. We do not apply data augmentation. The input image resolution is $1024 \\times 1024$ . Our EfficientSAMs are trained on 64 A100 GPUs with 40GB GPU memory.", + "bbox": [ + 496, + 90, + 893, + 407 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Baselines and Evaluation Metrics. Baselines. For the classification task, we compare the performance of ViT backbones from different pretraining/distillation methods including MAE[26], SSTA[60], DMAE[2], BEiT[3], CAE[12], DINO[6], iBOT[73], DeiT[53], etc. For the detection and instance semantic task, and semantic segmentation task, we also compare with several pretrained ViT backbones for ViTDet[34] and Mask2former[13]. For the segment every thing task, we compare with SAM[31], FastSAM[71], and MobileSAM[68]. Evaluation Metrics. We evaluate our method and all baselines in terms of accuracy. Specifically, the accuracy metrics refer to top-1 accuracy for the classification task; $\\mathrm{AP^{box}}$ , $\\mathrm{AP^{mask}}$ , for the detection and instance segmentation task (AP: average precision); mIoU, for the semantic segmentation task (mIoU: mean intersection over union); mIoU, AP, $\\mathrm{APS}$ , $\\mathrm{AP^M}$ , $\\mathrm{AP^L}$ for segment anything task. For efficiency metrics, we compare the number of model parameters or inference throughput.", + "bbox": [ + 496, + 409, + 895, + 681 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Main Results", + "text_level": 1, + "bbox": [ + 500, + 695, + 638, + 709 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Image Classification. To evaluate the effectiveness of our proposed techniques on the image classification task, we apply the proposed SAMI idea to ViT models and compare their performance over baselines on ImageNet-1K. As shown in Tab. 1, our SAMI is compared with pretraining methods like MAE, iBOT, CAE, and BEiT, and distillation methods including DeiT and SSTA. SAMI-B achieves $84.8\\%$ top-1 accuracy, which outperforms the pretraining baselines, MAE, DMAE, iBOT, CAE, and BEiT by $1.2\\%$ , $0.8\\%$ , $1.1\\%$ , $0.9\\%$ , and $0.4\\%$ respectively. Compared with distillation methods such as DeiT and SSTA, SAMI also shows large improvements. For lightweight models such as ViT-Tiny", + "bbox": [ + 496, + 719, + 893, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "16115", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/b278e3fd436ef17057d30228412536d513176c71d13b24b407a21f57c59e5a35.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodBackboneTraining DataAcc.(%)
DeiT-Ti[53]ViT-TinyIN1K74.5
SSTA-Ti[60]ViT-TinyIN1K75.2
DMAE-Ti[2]ViT-TinyIN1K70.0
MAE-Ti[26]ViT-TinyIN1K75.2
SAMI-T (ours)ViT-TinySA1B (11M) + IN1K76.8
DeiT-S[53]ViT-SmallIN1K81.2
SSTA-S[60]ViT-SmallIN1K81.4
DMAE-S[2]ViT-SmallIN1K79.3
MAE-S[26]ViT-SmallIN1K81.5
BEiT-S[3]ViT-SmallD250M+IN22K+IN1K81.7
CAE-S[12]ViT-SmallD250M+IN1K82.0
DINO-S[6]ViT-SmallIN1K82.0
iBOT-S[73]ViT-SmallIN22K+1N1K82.3
SAMI-S (ours)ViT-SmallSA1B (11M) + IN1K82.7
DeiT-B[53]ViT-BaseIN1K83.8
DMAE-B[2]ViT-BaseIN1K84.0
BootMAE[18]ViT-BaseIN1K84.2
MAE-B[26]ViT-BaseIN1K83.6
BEiT-B[3]ViT-BaseD250M+IN22K+IN1K83.7
CAE-B[12]ViT-BaseD250M+IN1K83.9
DINO-B[6]ViT-BaseIN1K82.8
iBOT-B[73]ViT-BaseIN22K+1N1K84.4
SAMI-B (ours)ViT-BaseSA1B (11M) + IN1K84.8
", + "bbox": [ + 89, + 87, + 460, + 380 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/2ac814eeab0049681b41d721b3d5c32dbdfec20c89b0e04077d5ae533a3c0549.jpg", + "table_caption": [ + "Table 1. Image classification results on ImageNet-1K. IN is short for ImageNet. We use IN1K for pretraining/binetuning and add SA1B to the training data to indicate the need for original SAM." + ], + "table_footnote": [], + "table_body": "
MethodBackboneAPbboxAPmask
MAE-Ti[26]ViT-Tiny37.934.9
SAMI-Ti(ours)ViT-Tiny44.740.0
MAE-S[26]ViT-Small45.340.8
DeiT-S[53]ViT-Small47.241.9
DINO-S[6]ViT-Small49.143.3
iBOT-S[73]ViT-Small49.744.0
SAMI-S (ours)ViT-Small49.844.2
MAE-B[26]ViT-Base51.645.9
SAMI-B (ours)ViT-Base52.546.5
", + "bbox": [ + 130, + 445, + 421, + 578 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "and ViT-Small, SAMI reports a substantial gain compared to DeiT, SSTA, DMAE, and MAE.", + "bbox": [ + 75, + 643, + 468, + 672 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Object Detection and Instance Segmentation. We also extend the SAMI-pretrained ViT backbones to the downstream object detection and instance segmentation task and compare it with previous pretraining baseline on COCO dataset to evaluate its efficacy. Specifically, we take the pretrained ViT backbones and adapt them to a simple feature pyramid in the Mask R-CNN framework[25] for constructing the detector, ViTDet[34]. Tab. 2 shows the overall comparison between our SAMI and other baselines. We can see that our SAMI consistently achieves better performance over other baselines. SAMI-B obtains $0.9\\mathrm{AP^{bbox}}$ and $0.6^{\\mathrm{mask}}$ gains compared with MAE-B. For light-weight backbones, SAMIS and SAMI-Ti report substantial gains compared to MAE-Ti and MAE-S. Moreover, SAMI-S significantly outperforms DeiT-S by $2.6\\mathrm{AP^{bbox}}$ and $2.3\\mathrm{AP^{mask}}$ . For other pretraining", + "bbox": [ + 75, + 674, + 472, + 902 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/bbb1c79e0dbd56b1db88a61f37a3b6b387950d222b24916ab630be64bdb9ab86.jpg", + "table_caption": [ + "Table 2. Object detection and instance segmentation results on the MS COCO using ViTDet." + ], + "table_footnote": [], + "table_body": "
MethodBackbonemIOU
MAE-Ti[26]ViT-Tiny39.0
SAMI-Ti(ours)ViT-Tiny42.7
MAE-S[26]ViT-Small44.1
SAMI-S (ours)ViT-Small48.8
MAE-B[26]ViT-Base49.3
SAMI-B (ours)ViT-Base51.8
", + "bbox": [ + 571, + 87, + 826, + 189 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/e9f56425f970434b3d5c3c9c71d70cd00c81a7fd2659a67b246b6942bc81a33a.jpg", + "table_caption": [ + "Table 3. Semantic segmentation results on the ADE20K dataset using Mask2former. The input resolution is $512 \\times 512$ ." + ], + "table_footnote": [], + "table_body": "
MethodCOCOLVIS
box1 click3 clickbox1 click3 click
SAM[31]78.455.674.178.959.875.2
MobileSAM[68]74.243.759.773.851.054.4
SAM-MAE-Ti[31]74.743.365.873.850.665.3
EfficientSAM-Ti (ours)75.745.567.274.352.766.8
EfficientSAM-S (ours)76.950.069.875.456.268.7
", + "bbox": [ + 513, + 241, + 883, + 321 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 4. Zero-shot single point valid mask evaluation results on COCO and LVIS. Following SAM[31], we uniformly sample random points within ground truth mask for click, and compute the tightest bounding box corresponding ground truth mask for box. SAM-MAE-Ti denotes SAM with pretrained MAE-Ti image encoder.", + "bbox": [ + 498, + 330, + 895, + 412 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "belines, our SAMI still compares favorably to DINO and iBOT. This set of experiments validate the effectiveness of the proposed SAMI for providing pretrained detector backbones in the object detection and instance segmentation task. Semantic Segmentation. We further extend the pretrained backbones to the semantic segmentation task to evaluate its effectiveness. Specifically, we use ViT models as the backbone in Mask2former [13] framework to benchmark on the ADE20K dataset. As shown in Tab. 3, Mask2former with SAMI-pretrained backbones achieve better mIoU, i.e., $\\uparrow 2.5$ , $\\uparrow 4.7$ , and $\\uparrow 3.7$ improvement over backbones with MAE pretraining [26] on ImageNet-1K. This set of experiments validate that our proposed techniques could be well generalized to various downstream tasks.", + "bbox": [ + 496, + 441, + 895, + 654 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. EfficientSAMs for Segment Anything Task", + "text_level": 1, + "bbox": [ + 498, + 664, + 862, + 681 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Segment Anything task is a process of promptable segmentation to produce segmentation masks based on any form of the prompt, including point set, rough boxes or mask, freeform text. We follow SAM [31] and focus on point-based and box-based prompt segmentation on COCO/LVIS. We now test the generalization abilities of our model on segment anything task including zero-shot single point valid mask evaluation and zero-shot instance segmentation. We take the SAMI-pretrained lightweight backbones as the image encoder of SAM for building efficient SAMs, EfficientSAMs. Then we finetune EfficientSAMs on SA-1B dataset and report the performance on zero-shot single point valid mask evaluation and zero-shot instance segmentation.", + "bbox": [ + 496, + 688, + 895, + 883 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Zero-Shot Single Point Valid Mask Evaluation. Similar", + "bbox": [ + 517, + 885, + 893, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "16116", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/4cdaf4d332667f6c77f6242065a20433bfbcec5e807a9bdaa0f254e2ce7c1d4b.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodCOCOLVIS
APAPSAPMAPLAPAPSAPMAPL
ViTDet-H[34]51.032.054.368.946.635.058.066.3
SAM[31]46.530.851.061.744.732.557.665.5
MobileSAM[68]38.723.742.254.334.423.844.953.7
FastSAM[71]37.923.943.450.034.524.646.250.8
EfficientSAM-Ti (ours)42.326.746.257.439.928.951.059.9
EfficientSAM-S (ours)44.428.448.360.142.330.854.062.3
", + "bbox": [ + 89, + 88, + 459, + 170 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 5. Zero-shot instance segmentation results on COCO/LVIS. ViTDet boxes are prompted to perform zero-shot segmentation.", + "bbox": [ + 76, + 179, + 470, + 207 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "to SAM[31], we evaluate segmenting an object from a single foreground point. For general interactive segmentation, we also consider object segmentation from a single box, and multiple points as introduced in [31]. To achieve this, we uniformly sample random points within ground truth mask for click, and compute the tightest bounding box corresponding to ground truth mask for box. Since our models are able to predict multiple masks, we only evaluate the most confident mask as SAM [31].", + "bbox": [ + 75, + 234, + 470, + 369 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Results. In Tab. 4, EfficientSAMs are compared with SAM, MobileSAM and SAM-MAE-Ti. On COCO, our EfficientSAM-Ti outperforms MobileSAM by $1.9\\mathrm{mIoU}$ on 1 click and $1.5\\mathrm{mIoU}$ on 1 box with comparable complexity. Our EfficientSAM-Ti with SAMI-pretrained weights also performs better than MAE-pretrained wights on COCO/LVIS interactive segmentation. We notice that our EfficientSAM-S only underperforms SAM by $1.5\\mathrm{mIoU}$ on COCO box and $3.5\\mathrm{mIoU}$ on LVIS box with $20x$ fewer parameters. We find that our EfficientSAMs also show promising performance on multiple click compared with MobileSAM and SAM-MAE-Ti.", + "bbox": [ + 75, + 369, + 470, + 551 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Zero-Shot Instance Segmentation. Following SAM [31], instance segmentation task is performed by taking the bounding box (bbox) generated by ViTDet[34] as the prompt. The mask with the highest Intersection over Union (IoU) with the_bbox as the predicted mask.", + "bbox": [ + 76, + 551, + 468, + 627 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Results. In Tab. 5, we report AP, $\\mathrm{APS}^{\\mathrm{s}}$ , $\\mathrm{APM}^{\\mathrm{s}}$ , $\\mathrm{APL}^{\\mathrm{s}}$ for zero-shot instance segmentation. We compare our EfficientSAM with MobileSAM and FastSAM. We can see that EfficientSAM-S obtains more than 6.5 AP on COCO and 7.8 AP on LVIS over FastSAM. For EfficientSAM-Ti, it still outperforms FastSAM by a large margin, 4.1 AP on COCO and 5.3 AP on LVIS, and MobileSAM by 3.6 AP on COCO and 5.5 AP on LVIS. Note that our EfficientSAMs are much light-weight than FastSAM, e.g., 9.8M parameters for efficientSAM-Ti vs 68M parameters for FastSAM. EfficientSAM-S also significantly reduces the gap between SAM with 0.6G parameters, only $\\sim2$ AP reduction. These results demonstrate the extraordinary benefits of EfficientSAMs for zero-shot instance segmentation and validate the advantages of our SAMI pretraining method.", + "bbox": [ + 75, + 628, + 470, + 854 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Qualitative Evaluation. We now provide the qualitative results for a complementary understanding of instance segmentation capabilities of EfficientSAMs. Some examples", + "bbox": [ + 76, + 854, + 470, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Input Image SAM[31] FastSAM[71] MobileSAM[68] EfficientSAM", + "bbox": [ + 508, + 75, + 885, + 87 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e462ee682d3ff6f7ff0ab4a8426c0e77b09e3a2f99f642f7ac0e58c012e2236d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 503, + 88, + 888, + 130 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/007be39bf465e5f70c2302f78b01e49a4bcecd1cce624b4055344345d330aaa1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 503, + 136, + 888, + 172 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/31d5a64b3dcf348335f9e0024e68dc5406d627be89a74dd20720a0f9b2db67d5.jpg", + "image_caption": [ + "Figure 3. Visualization result on point-prompt input with SAM, FastSAM, MobileSAM, and our EfficientSAM model." + ], + "image_footnote": [], + "bbox": [ + 503, + 178, + 888, + 220 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "can be seen in Fig. 3, Fig. 4, and Fig. 5. Specifically, we report the predicted masks with two types of prompts, point and box as in MobileSAM [68] and also segment everything results. More qualitative results can be found in the supplement. These results demonstrate that our EfficientSAMs have competing capabilities when comparing to SAM. Note that our EfficientSAMs are much lightweight than SAM, and our models can effectively give decent segmentation results. This indicates that our models can be served as a complementary version of SAM for many practical tasks.", + "bbox": [ + 496, + 286, + 892, + 438 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Input Image SAM[31] FastSAM[71] MobileSAM[68] EfficientSAM", + "bbox": [ + 509, + 443, + 885, + 454 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/4ec73f2b34416afb3937c1dd76997f041fcde170cb22b0a893af2eabadbf3f4f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 501, + 455, + 888, + 491 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/5b6559c846b5dfe10091dc90170fb0fca8c62424553ab3145d1b088eda60cb61.jpg", + "image_caption": [ + "Figure 4. Visualization result on box-prompt input with SAM, FastSAM, MobileSAM, and our EfficientSAM model." + ], + "image_footnote": [], + "bbox": [ + 503, + 497, + 888, + 540 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Salient Instance Segmentation. Salient object segmentation [4] aims to segment the most visually attractive objects from an image. We extend interactive instance segmentation to salient instance segmentation without manually creating points/boxes. Specifically, we take a state-of-the-art saliency object detection model, $\\mathrm{U}^2$ -net[45], to predict saliency map and uniformly sample 3 random points (3 click) within", + "bbox": [ + 496, + 595, + 892, + 702 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Input Image SAM[31] FastSAM[71] MobileSAM[68] EfficientSAM", + "bbox": [ + 509, + 710, + 885, + 720 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/594a71001852ba31b6dae70a744f92aae899c99728479baab48ae2c57bf791dc.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 503, + 723, + 885, + 765 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/8484e807ea89d4c8c5f90b459d6b1c59f7697f45f0eddd7e035c2c414587e8d8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 501, + 772, + 885, + 810 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/21978634147731669b65ad893441e496ca822e6c89f29b0f881d99bb22babbb2.jpg", + "image_caption": [ + "Figure 5. Visualization result on segment everything with SAM, FastSAM, MobileSAM, and our EfficientSAM model." + ], + "image_footnote": [], + "bbox": [ + 501, + 816, + 885, + 854 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "16117", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/c5f97f67a5834e353278e6a719ed8faa64ae1851e79a6a43dbe4295fb9ef923c.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodLossTop-1 Acc.(%)
SAMI-Ti1 - Cosine76.1
SAMI-TiMSE76.8
SAMI-S1 - Cosine82.3
SAMI-SMSE82.7
", + "bbox": [ + 158, + 88, + 392, + 154 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 6. Ablation study on training loss of SAMI. MSE loss gives better classification results on ImageNet-1K.", + "bbox": [ + 76, + 164, + 468, + 191 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "saliency map to perform instance segmentation with our EfficientSAM. In Fig. 6, we can see that our EfficientSAM can perform salient instance segmentation well. This preliminary exploration opens the potential to help people with hand impairments segment objects of interest in an image.", + "bbox": [ + 76, + 219, + 470, + 297 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/654530326519a5f0b225e53da68b6200ff6c243bc3f83a071828b48691c27888.jpg", + "image_caption": [ + "Figure 6. Saliency-based automatic instance segmentation results. With the assistance of saliency map generated from $\\mathbf{U}^2$ -net[45], our EfficientSAM is able to generate mask and perform automatic instance segmentation without manually creating points or boxes." + ], + "image_footnote": [], + "bbox": [ + 112, + 303, + 428, + 465 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4. Ablation Studies", + "text_level": 1, + "bbox": [ + 76, + 559, + 240, + 571 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We now analyze SAMI and EfficientSAMs through a series of ablation studies with ViT backbones.", + "bbox": [ + 75, + 582, + 468, + 612 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Reconstruction Loss. We study the effect of reconstruction loss on the performance of SAMI on ImageNet-1K. We compare our mean square error (MSE) reconstruction loss with cosine similarity loss. We find that MSE reconstruction loss performs better, shown in Tab. 6. This recommends a direct reconstruction of SAM features instead of the target with a high angular similarity.", + "bbox": [ + 75, + 613, + 468, + 719 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Cross-Attention Decoder. To reconstruct SAM features, we directly use the output tokens from encoder and only take decoder to transform the masked tokens with cross-attention. We study how the performance varies with cross-attention decoder. When changing the decoder module in MAE to our cross-attention decoder in SAMI, we find that SAMI-B improves the performance from $84.4\\%$ to $84.8\\%$ on ImageNet-1K. The cross-attention decoder of SAMI also shows consistent improvement for ViT-Tiny and ViT-Small. Analogy to anchor points in AnchorDETR[58], the output tokens from encoder are already learned well by directly aligning the SAM features, which can serve as anchor tokens", + "bbox": [ + 75, + 719, + 470, + 901 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "for help masked tokens align via cross-attention decoder.", + "bbox": [ + 500, + 90, + 875, + 106 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Mask Ratio. We explore how the performance varies with different mask ratio in SAMI. The observations are consistent with MAE [26] that a high mask ratio, $75\\%$ , tends to produce good results.", + "bbox": [ + 498, + 106, + 890, + 166 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Reconstruction Target. We study the impact of reconstruction target. We take a different encoder from CLIP [46] to generate features as the reconstruction target in SAMI. Aligning features from CLIP encoder can also outperform MAE by $0.8\\%$ for a ViT-Tiny model on ImageNet-1K. This demonstrates that masked image pretraining benefits from powerful guided reconstruction.", + "bbox": [ + 498, + 167, + 893, + 272 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/4120f34b8c2d8137670efcaea0b014cf992c37e92c1b7fd8d61207ac381ecb46.jpg", + "image_caption": [ + "Figure 7. Ablation study on training steps for EfficientSAMs on MS COCO dataset. Zero-shot single point valid mask evaluation with a single box prompt is performed for the ablation." + ], + "image_footnote": [], + "bbox": [ + 552, + 284, + 841, + 402 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effects of Finetuning Steps for EfficientSAMs. We explore the effect of fintuning steps for EfficientSAMs. As illustrated in Fig. 7, EfficientSAM-Ti and EfficientSAM-S achieve decent performance even at 0.1 epoch. For 1 epoch, the performance gain is larger than $2.5\\mathrm{mIoU}$ . The final performance of EfficientSAM-S reaches $76.9\\mathrm{mIoU}$ , which is only $1.5\\mathrm{mIoU}$ lower than SAM. These results demonstrate the advantages of SAMI-pretrained image encoders and our EfficientSAMs.", + "bbox": [ + 498, + 470, + 890, + 604 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Efficiency. We compare the throughput (images per second) of our EfficientSAM with SAM and other models on A100 in Fig. 1. Our EfficientSAM reduces the runtime/parameters of SAM by $\\sim 20\\mathrm{x}$ . We also observed similar speedup ( $\\sim 20\\mathrm{x}$ ) comparing to SAM w.r.t. FLOPs.", + "bbox": [ + 498, + 606, + 893, + 681 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 694, + 617, + 709 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We proposed a masked image pretraining approach, SAMI, to explore the potential of ViTs under the guidance of SAM foundation model. SAMI improves masked image pretraining by reconstructing the latent features from SAM image encoder to transfer knowledge from vision foundation model to ViTs. Extensive experiments on image classification, object detection and instance segmentation, semantic segmentation, and the segment anything task consistently validate SAMI's advantages. We also demonstrate that SAMI helps build efficient SAMs with pretrained light-weight encoders. Our preliminary work suggests that SAMI has potential applications beyond efficient segment anything task.", + "bbox": [ + 496, + 719, + 893, + 901 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "16118", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Pablo Arbelaez, Michael Maire, Charless Fowlkes, and Jitendra Malik. Contour detection and hierarchical image segmentation. IEEE transactions on pattern analysis and machine intelligence, 33(5):898-916, 2010. 1", + "[2] Yutong Bai, Zeyu Wang, Junfei Xiao, Chen Wei, Huiyu Wang, Alan L Yuille, Yuyin Zhou, and Cihang Xie. Masked autoencoders enable efficient knowledge distillers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24256-24265, 2023. 3, 5, 6", + "[3] Hangbo Bao, Li Dong, Songhao Piao, and Furu Wei. Beit: Bert pre-training of image transformers. arXiv preprint arXiv:2106.08254, 2021. 3, 5, 6", + "[4] Ali Borji, Ming-Ming Cheng, Qibin Hou, Huaizu Jiang, and Jia Li. Salient object detection: A survey. Computational visual media, 5:117-150, 2019. 7", + "[5] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In European conference on computer vision, pages 213-229. Springer, 2020. 2", + "[6] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9650-9660, 2021. 3, 5, 6", + "[7] Jun Cen, Yizheng Wu, Kewei Wang, Xingyi Li, Jingkang Yang, Yixuan Pei, Lingdong Kong, Ziwei Liu, and Qifeng Chen. Sad: Segment any rgbd. arXiv preprint arXiv:2305.14207, 2023. 2", + "[8] Jiaqi Chen, Zeyu Yang, and Li Zhang. Semantic segment anything. https://github.com/fudan-zvg/Semantic-Segment-Anything, 2023.2", + "[9] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International conference on machine learning, pages 1597-1607. PMLR, 2020. 3", + "[10] Tianrun Chen, Lanyun Zhu, Chaotao Deng, Runlong Cao, Yan Wang, Shangzhan Zhang, Zejian Li, Lingyun Sun, Ying Zang, and Papa Mao. Sam-adapter: Adapting segment anything in underperformed scenes. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3367-3375, 2023. 2", + "[11] Xinlei Chen and Kaiming He. Exploring simple siamese representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 15750-15758, 2021. 3", + "[12] Xiaokang Chen, Mingyu Ding, Xiaodi Wang, Ying Xin, Shentong Mo, Yunhao Wang, Shumin Han, Ping Luo, Gang Zeng, and Jingdong Wang. Context autoencoder for self-supervised representation learning. International Journal of Computer Vision, pages 1-16, 2023. 5, 6", + "[13] Bowen Cheng, Ishan Misra, Alexander G Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1290–1299, 2022. 5, 6" + ], + "bbox": [ + 78, + 114, + 470, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[14] Yangming Cheng, Lulei Li, Yuanyou Xu, Xiaodi Li, Zongxin Yang, Wenguan Wang, and Yi Yang. Segment and track anything. arXiv preprint arXiv:2305.06558, 2023. 2", + "[15] Ekin D Cubuk, Barret Zoph, Jonathon Shlens, and Quoc V Le. Randaugment: Practical automated data augmentation with a reduced search space. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops, pages 702-703, 2020. 5", + "[16] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 5", + "[17] Ruining Deng, Can Cui, Quan Liu, Tianyuan Yao, Lucas W Remedios, Shunxing Bao, Bennett A Landman, Lee E Whelless, Lori A Coburn, Keith T Wilson, et al. Segment anything model (sam) for digital pathology: Assess zero-shot segmentation on whole slide imaging. arXiv preprint arXiv:2304.04155, 2023. 2", + "[18] Xiaoyi Dong, Jianmin Bao, Ting Zhang, Dongdong Chen, Weiming Zhang, Lu Yuan, Dong Chen, Fang Wen, and Nenghai Yu. Bootstrapped masked autoencoders for vision bert pretraining. In European Conference on Computer Vision, pages 247-264. Springer, 2022. 6", + "[19] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 1, 2", + "[20] Haoqi Fan, Bo Xiong, Karttikeya Mangalam, Yanghao Li, Zhicheng Yan, Jitendra Malik, and Christoph Feichtenhofer. Multiscale vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6824-6835, 2021. 2", + "[21] Shanghua Gao, Zhijie Lin, Xingyu Xie, Pan Zhou, Ming-Ming Cheng, and Shuicheng Yan. Editanything: Empowering unparalleled flexibility in image editing and generation. In Proceedings of the 31st ACM International Conference on Multimedia, pages 9414–9416, 2023. 2", + "[22] Benjamin Graham, Alaaeldin El-Nouby, Hugo Touvron, Pierre Stock, Armand Joulin, Hervé Jégou, and Matthijs Douze. Levit: a vision transformer in convnet's clothing for faster inference. In Proceedings of the IEEE/CVF international conference on computer vision, pages 12259-12269, 2021. 3", + "[23] Agrim Gupta, Piotr Dólár, and Ross Girshick. Lvis: A dataset for large vocabulary instance segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2019. 5", + "[24] Dongsheng Han, Chaoning Zhang, Yu Qiao, Maryam Qamar, Yuna Jung, SeungKyu Lee, Sung-Ho Bae, and Choong Seon Hong. Segment anything model (sam) meets glass: Mirror and transparent objects cannot be easily detected. arXiv preprint arXiv:2305.00278, 2023. 1, 2", + "[25] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In Proceedings of the IEEE international conference on computer vision, pages 2961-2969, 2017. 6" + ], + "bbox": [ + 501, + 92, + 893, + 901 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "16119", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[26] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollar, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16000-16009, 2022. 2, 3, 4, 5, 6, 8", + "[27] Geoffrey Hinton, Oriol Vinyals, and Jeff Dean. Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531, 2015. 3", + "[28] Zejiang Hou, Fei Sun, Yen-Kuang Chen, Yuan Xie, and Sun-Yuan Kung. Milan: Masked image pretraining on language assisted representation. arXiv preprint arXiv:2208.06049, 2022.3", + "[29] Jiaxi Jiang and Christian Holz. Restore anything pipeline: Segment anything meets image restoration. arXiv preprint arXiv:2305.13093, 2023. 2", + "[30] Glenn Jocher, Ayush Chaurasia, and Jing Qiu. Yolo by ultralytics. https://github.com/ultralytics/ultralytics, 2023. 2", + "[31] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. arXiv preprint arXiv:2304.02643, 2023. 1, 2, 3, 4, 5, 6, 7", + "[32] Brett Koonce and Brett Koonce. *Mobilenetv3: Convolutional Neural Networks with Swift for Tensorflow: Image Recognition and Dataset Categorization*, pages 125–144, 2021. 2", + "[33] Jiashi Li, Xin Xia, Wei Li, Huixia Li, Xing Wang, Xuefeng Xiao, Rui Wang, Min Zheng, and Xin Pan. Next-vit: Next generation vision transformer for efficient deployment in realistic industrial scenarios. arXiv preprint arXiv:2207.05501, 2022. 3", + "[34] Yanghao Li, Hanzi Mao, Ross Girshick, and Kaiming He. Exploring plain vision transformer backbones for object detection. In European Conference on Computer Vision, pages 280-296. Springer, 2022. 2, 5, 6, 7", + "[35] Yanyu Li, Geng Yuan, Yang Wen, Ju Hu, Georgios Evangelidis, Sergey Tulyakov, Yanzhi Wang, and Jian Ren. Efficientformer: Vision transformers at mobilenet speed. Advances in Neural Information Processing Systems, 35:12934-12949, 2022. 3", + "[36] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In European conference on computer vision, pages 740-755. Springer, 2014. 5", + "[37] Ruiping Liu, Jiaming Zhang, Kunyu Peng, Junwei Zheng, Ke Cao, Yufan Chen, Kailun Yang, and Rainer Stiefelhagen. Open scene understanding: Grounded situation recognition meets segment anything for helping people with visual impairments. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1857-1867, 2023. 1, 2", + "[38] Xinyu Liu, Houwen Peng, Ningxin Zheng, Yuqing Yang, Han Hu, and Yixuan Yuan. Efficientvit: Memory efficient vision transformer with cascaded group attention. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14420-14430, 2023. 3" + ], + "bbox": [ + 78, + 90, + 470, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[39] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10012-10022, 2021. 2", + "[40] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In International Conference on Learning Representations, 2018. 5", + "[41] Jun Ma and Bo Wang. Segment anything in medical images. arXiv preprint arXiv:2304.12306, 2023. 1, 2", + "[42] Sachin Mehta and Mohammad Rastegari. Mobilevit: Lightweight, general-purpose, and mobile-friendly vision transformer. In International Conference on Learning Representations, 2021. 2", + "[43] Deepak Pathak, Philipp Krahenbuhl, Jeff Donahue, Trevor Darrell, and Alexei A Efros. Context encoders: Feature learning by inpainting. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2536-2544, 2016. 3", + "[44] Zhiliang Peng, Li Dong, Hangbo Bao, Qixiang Ye, and Furu Wei. Beit v2: Masked image modeling with vector-quantized visual tokenizers. arXiv preprint arXiv:2208.06366, 2022. 2, 3", + "[45] Xuebin Qin, Zichen Zhang, Chenyang Huang, Masood Dehghan, Osmar Zaiane, and Martin Jagersand. U2-net: Going deeper with nested u-structure for salient object detection. page 107404, 2020. 7, 8", + "[46] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In International Conference on Machine Learning, pages 8821-8831. PMLR, 2021. 8", + "[47] Adriana Romero, Nicolas Ballas, Samira Ebrahimi Kahou, Antoine Chassang, Carlo Gatta, and Yoshua Bengio. Fitnets: Hints for thin deep nets. arXiv preprint arXiv:1412.6550, 2014.3", + "[48] Mark Sandler, Andrew Howard, Menglong Zhu, Andrey Zh-moginov, and Liang-Chieh Chen. Mobilenetv2: Inverted residuals and linear bottlenecks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4510-4520, 2018. 2", + "[49] Qiuhong Shen, Xingyi Yang, and Xinchao Wang. Anything-3d: Towards single-view anything reconstruction in the wild. arXiv preprint arXiv:2304.10261, 2023. 2", + "[50] Ao Sun, Pingchuan Ma, Yuanyuan Yuan, and Shuai Wang. Explain any concept: Segment anything meets concept-based explanation. arXiv preprint arXiv:2305.10289, 2023. 1, 2", + "[51] Lv Tang, Haoke Xiao, and Bo Li. Can sam segment anything? when sam meets camouflaged object detection. arXiv preprint arXiv:2304.04709, 2023. 2", + "[52] Shehbaz Tariq, Brian Estadimas Arfeto, Chaoning Zhang, and Hyundong Shin. Segment anything meets semantic communication. arXiv preprint arXiv:2306.02094, 2023. 1, 2", + "[53] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Hervé Jégou. Training data-efficient image transformers & distillation through attention. In International conference on machine learning, pages 10347-10357. PMLR, 2021. 2, 5, 6" + ], + "bbox": [ + 501, + 92, + 893, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "16120", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[54] Koen EA Van de Sande, Jasper RR Uijlings, Theo Gevers, and Arnold WM Smeulders. Segmentation as selective search for object recognition. In 2011 international conference on computer vision, pages 1879-1886. IEEE, 2011. 1", + "[55] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 3", + "[56] Pascal Vincent, Hugo Larochelle, Isabelle Lajoie, Yoshua Bengio, Pierre-Antoine Manzagol, and León Bottou. Stacked denoising autoencoders: Learning useful representations in a deep network with a local denoising criterion. Journal of machine learning research, 11(12), 2010. 3", + "[57] Xinlong Wang, Rufeng Zhang, Chunhua Shen, Tao Kong, and Lei Li. Dense contrastive learning for self-supervised visual pre-training. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3024-3033, 2021. 3", + "[58] Yingming Wang, Xiangyu Zhang, Tong Yang, and Jian Sun. Anchor detr: Query design for transformer-based detector. In Proceedings of the AAAI conference on artificial intelligence, pages 2567-2575, 2022. 8", + "[59] Chen Wei, Haoqi Fan, Saining Xie, Chao-Yuan Wu, Alan Yuille, and Christoph Feichtenhofer. Masked feature prediction for self-supervised visual pre-training. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14668-14678, 2022. 3", + "[60] Haiyan Wu, Yuting Gao, Yinqi Zhang, Shaohui Lin, Yuan Xie, Xing Sun, and Ke Li. Self-supervised models are good teaching assistants for vision transformers. In Proceedings of the 39th International Conference on Machine Learning, pages 24031-24042. PMLR, 2022. 3, 5, 6", + "[61] Kan Wu, Jinnian Zhang, Houwen Peng, Mengchen Liu, Bin Xiao, Jianlong Fu, and Lu Yuan. Tinyvit: Fast pretraining distillation for small vision transformers. In European Conference on Computer Vision, pages 68-85. Springer, 2022. 3", + "[62] Zhenda Xie, Yutong Lin, Zheng Zhang, Yue Cao, Stephen Lin, and Han Hu. Propagate yourself: Exploring pixel-level consistency for unsupervised visual representation learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16684-16693, 2021. 3", + "[63] Zhenda Xie, Zheng Zhang, Yue Cao, Yutong Lin, Jianmin Bao, Zhuliang Yao, Qi Dai, and Han Hu. Simmim: A simple framework for masked image modeling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9653-9663, 2022. 3", + "[64] Jing Yang, Brais Martinez, Adrian Bulat, and Georgios Tzimiropoulos. Knowledge distillation via softmax regression representation learning. In International Conference on Learning Representations, 2021. 3", + "[65] Jinyu Yang, Mingqi Gao, Zhe Li, Shang Gao, Fangjing Wang, and Feng Zheng. Track anything: Segment anything meets videos. arXiv preprint arXiv:2304.11968, 2023. 2", + "[66] Haoran You, Yunyang Xiong, Xiaoliang Dai, Bichen Wu, Peizhao Zhang, Haoqi Fan, Peter Vajda, and Yingyan Celine Lin. Castling-vit: Compressing self-attention via switching towards linear-angular attention at vision transformer inference." + ], + "bbox": [ + 78, + 90, + 470, + 898 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14431-14442, 2023. 3", + "[67] Tao Yu, Runseng Feng, Ruoyu Feng, Jinming Liu, Xin Jin, Wenjun Zeng, and Zhibo Chen. Inpaint anything: Segment anything meets image inpainting. arXiv preprint arXiv:2304.06790, 2023. 2", + "[68] Chaoning Zhang, Dongshen Han, Yu Qiao, Jung Uk Kim, Sung-Ho Bae, Seungkyu Lee, and Choong Seon Hong. Faster segment anything: Towards lightweight sam for mobile applications. arXiv preprint arXiv:2306.14289, 2023. 2, 5, 6, 7", + "[69] Xiao Feng Zhang, Tian Yi Song, and Jia Wei Yao. Deshadow—anything: When segment anything model meets zero-shot shadow removal. arXiv preprint arXiv:2309.11715, 2023. 2", + "[70] Borui Zhao, Quan Cui, Renjie Song, Yiyu Qiu, and Jiajun Liang. Decoupled knowledge distillation. In Proceedings of the IEEE/CVF Conference on computer vision and pattern recognition, pages 11953-11962, 2022. 3", + "[71] Xu Zhao, Wenchao Ding, Yongqi An, Yinglong Du, Tao Yu, Min Li, Ming Tang, and Jinqiao Wang. Fast segment anything. arXiv preprint arXiv:2306.12156, 2023. 2, 5, 7", + "[72] Bolei Zhou, Hang Zhao, Xavier Puig, Sanja Fidler, Adela Barriuso, and Antonio Torralba. Scene parsing through ade20k dataset. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 633-641, 2017. 5", + "[73] Jinghao Zhou, Chen Wei, Huiyu Wang, Wei Shen, Cihang Xie, Alan Yuille, and Tao Kong. ibot: Image bert pre-training with online tokenizer. arXiv preprint arXiv:2111.07832, 2021. 5, 6" + ], + "bbox": [ + 503, + 92, + 893, + 501 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "16121", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2024/EfficientSAM_ Leveraged Masked Image Pretraining for Efficient Segment Anything/f1e649e9-e5d4-48b3-8018-af1f58065ff2_model.json b/2024/EfficientSAM_ Leveraged Masked Image Pretraining for Efficient Segment Anything/f1e649e9-e5d4-48b3-8018-af1f58065ff2_model.json new file mode 100644 index 0000000000000000000000000000000000000000..2526a71ad083a07994c068b329491abec259cbfe --- /dev/null +++ b/2024/EfficientSAM_ Leveraged Masked Image Pretraining for Efficient Segment Anything/f1e649e9-e5d4-48b3-8018-af1f58065ff2_model.json @@ -0,0 +1,2356 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.0, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.109, + 0.13, + 0.862, + 0.178 + ], + "angle": 0, + "content": "EfficientSAM: Leveraged Masked Image Pretraining for Efficient Segment Anything" + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.204, + 0.874, + 0.256 + ], + "angle": 0, + "content": "Yunyang Xiong, Bala Varadarajan*, Lemeng Wu*, Xiaoyu Xiang, Fanyi Xiao, Chenchen Zhu, Xiaoliang Dai, Dilin Wang, Fei Sun, Forrest Iandola, Raghuraman Krishnamoorthi, Vikas Chandra Meta AI Research" + }, + { + "type": "text", + "bbox": [ + 0.298, + 0.26, + 0.658, + 0.275 + ], + "angle": 0, + "content": "https://yformer.github.io/efficient-sam/" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.309, + 0.314, + 0.327 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.342, + 0.474, + 0.721 + ], + "angle": 0, + "content": "Segment Anything Model (SAM) has emerged as a powerful tool for numerous vision applications. A key component that drives the impressive performance for zero-shot transfer and high versatility is a super large Transformer model trained on the extensive high-quality SA-1B dataset. While beneficial, the huge computation cost of SAM model has limited its applications to wider real-world applications. To address this limitation, we propose EfficientSAMs, lightweight SAM models that exhibits decent performance with largely reduced complexity. Our idea is based on leveraging masked image pretraining, SAMI, which learns to reconstruct features from SAM image encoder for effective visual representation learning. Further, we take SAMI-pretrained light-weight image encoders and mask decoder to build EfficientSAMs, and finetune the models on SA-1B for segment anything task. We perform evaluations on multiple vision tasks including image classification, object detection, instance segmentation, and semantic segmentation, and find that our proposed pretraining method, SAMI, consistently outperforms other masked image pretraining methods. On segment anything task such as zero-shot instance segmentation, our EfficientSAMs with SAMI-pretrained lightweight image encoders perform favorably with a significant gain (e.g., \\(\\sim 4\\) AP on COCO/LVIS) over other fast SAM models. Our EfficientSAM code and models are available at here." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.747, + 0.21, + 0.763 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.772, + 0.472, + 0.879 + ], + "angle": 0, + "content": "Segment Anything Model (SAM) [31] has been very successful in the vision field, achieving state-of-the-art performance in a variety of image segmentation tasks such as zero-shot edge detection[1, 31], zero-shot object proposal generation[31, 54], and zero-shot instance segmentation[31], and many other real-world applications[24, 37, 41, 50-52]. The key feature of SAM is a prompt-based Vision Trans" + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.313, + 0.889, + 0.379 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.511, + 0.388, + 0.882, + 0.543 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.555, + 0.895, + 0.723 + ], + "angle": 0, + "content": "Figure 1. The comparative analysis result. (Top) The overview of EfficientSAM model by taking a well-pretrained light-weight image encoder for instance segmentation with largely reduced complexity. (Bottom) Throughput/Parameter/Performance comparison of EfficientSAM, MobileSAM, FastSAM, and SAM for zero-shot instance segmentation on COCO. We benchmark throughput (image per second) of all models on a single NVIDIA A100 with one box prompt. The input image resolution is \\(1024 \\times 1024\\). Our EfficientSAMs outperform MobileSAM and FastSAM by a large margin, \\(\\sim 4\\) AP, with comparable complexity. Our EfficientSAM-S reduces the inference time of SAM by \\(\\sim 20x\\) and the parameter size by \\(\\sim 20x\\) with a small performance drop, 44.4 AP vs 46.5 AP." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.749, + 0.895, + 0.825 + ], + "angle": 0, + "content": "former (ViT)[19]model trained on a large-scale visual dataset with more than 1B masks from 11M images, SA-1B[31], which allows segmenting any object on a given image. This ability of Segment Anything makes SAM a foundation model in vision and enables its applications even beyond vision." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.826, + 0.897, + 0.903 + ], + "angle": 0, + "content": "Despite the forgoing advantages, the model of SAM turns out to be a major efficiency bottleneck for practical deployment since the architecture of SAM, especially, the image encoder (e.g., ViT-H) is very expensive. Note that ViT-H image encoder in SAM has 632M parameters while the" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.888, + 0.21, + 0.9 + ], + "angle": 0, + "content": "*Joint second author." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "16111" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.471, + 0.152 + ], + "angle": 0, + "content": "prompt-based decoder only take 3.87M parameters. As a result, it leads to high computation and memory costs when using SAM to perform segment anything tasks in practice, which makes it challenging for real-time applications." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.153, + 0.472, + 0.259 + ], + "angle": 0, + "content": "To address this challenge, several recent works have proposed strategies that avoid incurring the huge cost when applying SAM for the prompt-based instance segmentation. For example, [68] suggests distilling the knowledge from the default ViT-H image encoder to a tiny ViT image encoder. In [71], the computation cost can be reduced with a real-time CNN-based architecture for Segment Anything task." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.26, + 0.473, + 0.518 + ], + "angle": 0, + "content": "In this paper, we propose using a well-pretrained lightweight ViT image encoder (e.g., ViT-Tiny/-Small[53]) to reduce the complexity of SAM while maintain decent performance. Our method, SAM-leveraged masked image pertraining (SAMI), produces desired pretrained lightweight ViT backbones for segment anything task. This is achieved by leveraging the celebrated MAE[26] pretraining method with SAM model to obtain high-quality pretrained ViT encoders. Specifically, our SAMI makes use of SAM encoder, ViT-H, to generate feature embedding and train a masked image model with lightweight encoders to reconstruct features from ViT-H of SAM instead of image patches. This leads to generalized ViT backbones, which can be used for downstream tasks such as image classification, object detection, and segment anything. Then we finetune the pretrained lightweight encoders with SAM decoders for segment anything task[31]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.518, + 0.472, + 0.819 + ], + "angle": 0, + "content": "To evaluate our method, we consider a transfer learning setting for masked image pretraining, where models are first pretrained with a reconstructive loss on ImageNet with image resolution \\(224 \\times 224\\), and then finetuned on target tasks using supervised data. Our SAMI learns lightweight encoders that generalize well. With SAMI pretraining, we can train models like ViT-Tiny/-Small/-Base on ImageNet-1K with improved generalization performance. For a ViT-Small model, we achieve \\(82.7\\%\\) top-1 accuracy when finetuned on ImageNet-1K with 100 epochs, which outperforms other state-of-the-art image pretraining baselines. We also finetune our pretrained models on object detection, instance segmentation, and semantic segmentation. Across all these tasks, our pretraining method achieves better results than other pretraining baselines, and more importantly, we observe significant gains for small models. Further, we evaluate our models on Segment Anything task. On zero-shot instance segmentation, our model performs well compared to recent lightweight SAM methods, including FastSAM, by a margin of 4.1 /5.2 AP on COCO/LVIS." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.821, + 0.452, + 0.835 + ], + "angle": 0, + "content": "Our main contribution can be summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.098, + 0.84, + 0.472, + 0.901 + ], + "angle": 0, + "content": "- We propose a SAM-leveraged masked image pretraining framework called SAMI, which trains the model to reconstruct features from SAM ViT-H image encoder. We show that this can substantially improve the" + }, + { + "type": "text", + "bbox": [ + 0.533, + 0.092, + 0.871, + 0.107 + ], + "angle": 0, + "content": "performance of image masked pretraining method." + }, + { + "type": "text", + "bbox": [ + 0.521, + 0.107, + 0.894, + 0.151 + ], + "angle": 0, + "content": "- We demonstrate that SAMI-pretrained backbones can generalize well to many tasks including image classification, object detection, and semantic segmentation." + }, + { + "type": "text", + "bbox": [ + 0.521, + 0.152, + 0.895, + 0.213 + ], + "angle": 0, + "content": "- We deliver EfficientSAMs, light-weight SAM models with state-of-the-art quality-efficiency trade-offs (Fig. 1), which is complementary to SAM for practical deployment." + }, + { + "type": "list", + "bbox": [ + 0.521, + 0.107, + 0.895, + 0.213 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.224, + 0.642, + 0.24 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.25, + 0.895, + 0.296 + ], + "angle": 0, + "content": "We briefly review relevant works on segment anything model, vision transformers, knowledge distillation, and masked image pretraining." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.303, + 0.733, + 0.319 + ], + "angle": 0, + "content": "2.1. Segment Anything Model" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.326, + 0.895, + 0.659 + ], + "angle": 0, + "content": "SAM[31] has been considered as a milestone vision foundation model, which can segment any object in the image based on interaction prompts. SAM has shown remarkable zero-shot transfer performance and high versatility for many vision tasks including a variety of segmentation application[7, 8, 10, 17], in-painting[67], image restoration[29], image editing[21], image shadow removal[69], object tracking[14, 65], and 3D object reconstruction[49]. There are many other works attempting to generalize SAM to real-world scenarios, including medical image segmentation[41], camouflaged object detection[51], transparent object detection[24], concept-based explanation[50], semantic communication[52], and helping people with visual impairments[37]. Due to its wide real-world applications, practical deployment of SAM has also gained increasing attention. Several recent works including [68, 71] have proposed strategies to reduce the computation costs of SAM. FastSAM[68] develops a CNN-based architecture, YOLOv8-seg[30], to segment all objects in an image for efficiency improvement. MobileSAM[71] presents a decoupled distillation for obtaining a lightweight image encoder of SAM. Our work focuses on dealing with this efficiency issue for practical deployment of SAM." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.666, + 0.694, + 0.681 + ], + "angle": 0, + "content": "2.2. Vision Transformers" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.689, + 0.895, + 0.901 + ], + "angle": 0, + "content": "ViTs [19] have achieved impressive performance in vision applications[5, 20, 26, 34, 39, 44]. ViTs demonstrate the advantages of and generalization over their CNN counterparts[26]. There are also a number of works on efficient ViTs for deployment. Smaller ViTs such as ViT-Small/DeiT-Small and ViT-Tiny/DeiT-Tiny are introduced in [53] for complementing ViT-Huge, ViT-Large, and ViT-Base in [19]. Motivated by the ability of convolution to capture local information with reduced parameters and computation costs, MobileViT[42] explore combining ViT with convolutions, which outperforms light-weight CNN models such as MobileNet-v2/v3[32, 48] with better task-level generalization properties and reduced memory size and computation cost. This trick has been used in many follow-up works" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "16112" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.091, + 0.472, + 0.152 + ], + "angle": 0, + "content": "including LeViT[22], EfficientFormer[35], Next-ViT[33], Tiny-ViT[61], Castling-ViT[66], EfficientViT[38]. This line of progress for designing efficient ViTs is orthogonal to our EfficientSAM work towards building efficient SAM." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.163, + 0.29, + 0.179 + ], + "angle": 0, + "content": "2.3. Knowledge Distillation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.187, + 0.473, + 0.534 + ], + "angle": 0, + "content": "Knowledge distillation (KD) is a technique to improve the performance of deep learning models without changing their architectures. [27] is a pioneering work to distill the dark knowledge from a larger teacher model to a smaller student model. The learning of a student model is supervised by the hard labels and the soft labels from a teacher model. This practice is followed by multiple works which aim to make better use of soft labels to transfer more knowledge. In [64], the distillation method decouples representation learning and classification. Decoupled knowledge distillation[70] separates the classical KD loss into two parts, target class knowledge distillation and non-target class knowledge distillation, which improves the effectiveness and flexibility of knowledge transfer. Another line of work is to transfer knowledge from intermediate features. FitNet [47] is a pioneering work by distilling the semantic information from the teacher model's intermediate feature directly. In [60], a self-supervised teaching assistant (SSTA) is introduced to guide the learning of a ViT-based student model with a supervised teacher together. [2] studies the potential of knowledge distillation from pre-training MAE model by aligning the intermediate features between the larger MAE teacher model and smaller MAE student model." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.545, + 0.321, + 0.562 + ], + "angle": 0, + "content": "2.4. Masked Image Pretraining" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.569, + 0.473, + 0.903 + ], + "angle": 0, + "content": "Self-supervised pretraining approaches [6] have attracted significant attention in computer vision. One line of work is contrastive learning methods[9, 11, 57, 62], which learn augmentation in-variance by imposing high similarity between different augmented views of a given image. While the learned representation shows good properties such as high linear separability, contrastive learning methods relies on strong augmentation and negative sampling. Another interesting line of work is masked image modeling (MIM), which helps models learn meaningful representations by reconstructing masked image patches. The MIM pioneering works focus on using denoising autoencoders[56] and context encoders[43] to train vision Transformer with masked prediction objectives. There are various promising works on using MIM for self-supervised image pretraining. BEiT[3] is the first one that adopts MIM for ViT pretraining to predict visual tokens. In BEiTv2[44], a semantic-rich image tokenizer is utilized for a better reconstruction target. In MaskFeat[59], reconstructing the local gradient features generated from HOG descriptor leads to effective visual pretraining. In SimMIM[63] and MAE[26], directly reconstructing the pixel values of the masked image patches achieves effec" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.168 + ], + "angle": 0, + "content": "tive visual representation learning. There are MAE-based follow-up works that use large teacher models to guide MAE pretraining[2, 28, 60]. Our work is built on MAE and finds that leveraging MAE to reconstruct the features from SAM image encoder enables the pretraining to be highly effective." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.182, + 0.61, + 0.2 + ], + "angle": 0, + "content": "3. Approach" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.208, + 0.63, + 0.224 + ], + "angle": 0, + "content": "3.1. Preliminary" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.231, + 0.895, + 0.429 + ], + "angle": 0, + "content": "Masked Autoencoders. Masked Autoencoders (MAE) model has two components, an encoder and a decoder. Both encoder and decoder are built on Transformer layers[55]. MAE takes image tokens, i.e., non-overlapping patches from the input images, as input. These input tokens are grouped to unmasked tokens and masked tokens with a given masking ratio. The unmasked tokens will be kept for encoder to extract features and the masked tokens will be set as the learning targets of the MAE decoder that need to be reconstructed during self-supervised learning (MIM). MAE[26] adopts a high mask ratio (e.g., \\(75\\%\\)), which prevents information leakage (e.g., simply extrapolating masked pixels based on the neighbors) in the pretraining stage." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.439, + 0.874, + 0.455 + ], + "angle": 0, + "content": "3.2. SAM-Leveraged Masked Image Pretraining" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.462, + 0.895, + 0.644 + ], + "angle": 0, + "content": "We now adapt MAE framework to obtain efficient image encoders for segment anything model. Motivated by the high versatility of SAM [31], we explore latent features from SAM image encoder as the reconstruction target to leverage MAE. Our method emphasizes transferring the knowledge embedded in SAM. Fig. 2 (top) illustrates an overview of the proposed SAM-leveraged masked image pretraining, SAMI. The encoder transforms the unmasked tokens into latent feature representation and the decoder reconstructs the representation of the masked tokens aided by the output feature embedding from the encoder. The representation learning of reconstruction is guided by latent features from SAM." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.644, + 0.895, + 0.825 + ], + "angle": 0, + "content": "Cross-Attention Decoder. With the supervision of SAM features, we observe that only masked tokens need to be reconstructed via decoder while the encoder's output can serve as anchors during the reconstruction. In the cross-attention decoder, queries come from masked tokens, and keys and values derive from both unmasked features from encoder and masked features. We merge the output features of masked tokens from cross-attention decoder and the output features of unmasked tokens from encoder for the MAE output embedding. Then, this combined features will be reordered to the original positions of input image tokens for the final MAE outputs." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.826, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Linear Projection Head. We obtain the image output through our encoder and cross-attention decoder. Then we feed such features into a small project head for aligning the features from SAM image encoder. For simplicity, we just use a linear projection head to address the feature dimension" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "16113" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.111, + 0.096, + 0.862, + 0.356 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.375, + 0.895, + 0.432 + ], + "angle": 0, + "content": "Figure 2. The overview of EfficientSAM framework. Our proposed EfficientSAM contains two stages: SAMI pretraining (top) on ImageNet and SAM finetuning (bottom) on SA-1B. For SAMI pretraining, the masked autoencoder takes the feature embeddings from SAM image encoder as the reconstruction target. After SAMI pretraining, the decoder is discarded and the light-weight encoder is served as the image encoder of EfficientSAM for finetuning on SA-1B." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.447, + 0.471, + 0.476 + ], + "angle": 0, + "content": "mismatch between the output of SAM image encoder and MAE." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.478, + 0.471, + 0.567 + ], + "angle": 0, + "content": "Reconstruction Loss. At each training iteration, SAMI consists of a feedforward feature extraction from SAM image encoder, and a feedforward and a backpropagation procedure of MAE. The outputs from SAM image encoder and MAE linear projection head are compared to compute the reconstruction loss." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.568, + 0.472, + 0.702 + ], + "angle": 0, + "content": "Let us denote the SAM image encoder as \\( f^{\\mathrm{sam}} \\), and the encoder and decoder of MAE as \\( g^{e} \\) with weights \\( W_{e} \\) and \\( g^{d} \\) with weights \\( W_{d} \\), linear projection head as \\( h^{\\theta} \\) with weights \\( W_{\\theta} \\) respectively. Assume the input tokens are denoted as \\( \\{\\mathbf{x}_i\\}_{i=1}^N \\), where \\( N \\) is the number of tokens. The input tokens are randomly grouped into the unmasked tokens, \\( \\{\\mathbf{x}_i\\}_{i \\in \\mathcal{U}} \\), the masked tokens \\( \\{\\mathbf{x}_i\\}_{i \\in \\mathcal{M}} \\) with a given masked ratio. Let the feature reordering operator be \\( \\phi \\), and the merging operator be \\( \\oplus \\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.704, + 0.472, + 0.793 + ], + "angle": 0, + "content": "Our target features from SAM image encoder can be written as \\( f^{\\mathrm{sam}}(\\mathbf{x}) = f^{\\mathrm{sam}}(\\{\\mathbf{x}_i\\}_{i=1}^N) \\), the output from MAE encoder is \\( g^{e}(\\{\\mathbf{x}_{i}\\}_{i\\in \\mathcal{U}}) \\), the decoder output is \\( g^{d}(\\{\\mathbf{x}_{i}\\}_{i\\in \\mathcal{M}}) \\). The output from linear projection head is \\( f^{h}(\\mathbf{x}) = h^{\\theta}(\\phi(g^{e}\\{\\mathbf{x}_{i}\\}_{i\\in \\mathcal{U}} \\oplus g^{d}\\{\\mathbf{x}_{i}\\}_{i\\in \\mathcal{M}})) \\). Therefore, our target reconstruction loss can be formulated as," + }, + { + "type": "equation", + "bbox": [ + 0.122, + 0.804, + 0.471, + 0.845 + ], + "angle": 0, + "content": "\\[\nL _ {W _ {e}, W _ {d}, W _ {\\theta}} = \\frac {1}{N} \\cdot \\sum_ {j = 1} ^ {N} \\left\\| f ^ {\\operatorname {s a m}} (\\mathbf {x}) - f ^ {h} (\\mathbf {x}) \\right\\| ^ {2}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.473, + 0.903 + ], + "angle": 0, + "content": "where \\(N\\) is the number of input tokens, \\(\\| \\cdot \\|\\) denotes a norm. We use \\(\\ell_2\\) norm for reconstruction loss in our experiments. By minimizing the reconstruction loss, \\(L_{W_e,W_d,W_\\theta}\\)," + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.447, + 0.895, + 0.537 + ], + "angle": 0, + "content": "our encoder \\( g^{e} \\) is optimized to serve as an image backbone to extract features as SAM image encoder. Our encoder, decoder, and linear projection head are optimized to learn context modeling ability from SAM image encoder. Optimizing the reconstruction loss on all tokens transfer the knowledge embedded in SAM." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.538, + 0.896, + 0.69 + ], + "angle": 0, + "content": "SAMI for EfficientSAM. After pretraining, our encoder extract feature representations for various vision tasks and the decoder is discarded. In particular, to build efficient SAM models for the segmentation anything task, we take the SAMI-pretrained lightweight encoder such as ViT-Tiny and ViT-Small as the image encoder and the default mask decoder of SAM for our EfficientSAM, as illustrated in Fig. 2 (bottom). We finetune our EfficientSAM models on SA-1B dataset for the segment anything task. The overview of our EfficientSAM framework is illustrated in Fig. 2." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.702, + 0.634, + 0.719 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.727, + 0.707, + 0.743 + ], + "angle": 0, + "content": "4.1. Experimental Settings" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.895, + 0.854 + ], + "angle": 0, + "content": "Pretraining Datasets. Our masked image pretraining method, SAMI, is conducted on ImageNet-1K training set with 1.2M images. Following masked image pretraining [26], we do not use the label information. We use the SAM ViT-H image encoders from [31] to generate reconstruction features when pretraining our ViT models, ViT-Tiny, ViT-Small, and ViT-Base." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.895, + 0.901 + ], + "angle": 0, + "content": "Pretraining Implementation Details. Our ViT models are pretrained with a mean squared error (MSE) loss for reconstruction. We use a batch size of 4096, AdamW optimizer" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "16114" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.091, + 0.472, + 0.243 + ], + "angle": 0, + "content": "[40] with learning rate \\(2.4\\mathrm{e} - 3\\), \\(\\beta_{1} = 0.9\\), \\(\\beta_{2} = 0.95\\), weight decay 0.05, linear learning rate warm-up over the first 40 epochs, cosine learning rate decay to update our models. We only adopt random resize crop to \\(224\\times 224\\) resolution, random horizontal flip, and normalization for data augmentation. The mask ratio is set to \\(75\\%\\) and the decoder contains 8 Transformer blocks with 512 dimensions as in [26]. We pretrain SAMI for 400 epochs using PyTorch framework on V100 machines. For reference, 1600-epoch pretraining is required for MAE[26]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.247, + 0.473, + 0.596 + ], + "angle": 0, + "content": "Downstream Tasks/Datasets/Models. Tasks and Datasets. We first consider three benchmarking datasets and several representative vision tasks to demonstrate the superiority of the proposed SAMI, including image classification on ImageNet dataset [16] with 1.2 million training and 50K validation images; Object detection and instance segmentation on COCO dataset [36] with 118K training and 5K validation images; Semantic segmentation on ADE20K dataset [72] with 20K/2K/3K images for training, validation, and testing, respectively. Then, we consider segment anything task to further show the advantages of our proposed SAMI. We finetune our pretrained lightweight image encoders for SAM on SA-1B dataset [31] with more than 1B masks from 11M high-resolution images, and test interactive instance segmentation and zero-shot instance segmentation ability of our EfficientSAMs on COCO and LVIS [23]. Models. We discard the decoder of SAMI while keeping the encoder as backbone to extract features for different tasks as in MAE[26]. We apply our well-pretrained ViT backbones for different tasks including ViTs for the classification, ViTDet [34] for the detection and instance segmentation, Mask2former [13] for the semantic segmentation task, and SAM for segment anything." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.599, + 0.473, + 0.902 + ], + "angle": 0, + "content": "Finetuning Settings. For the classification task, we use the AdamW optimizer with \\(\\beta_{1} = 0.9\\), \\(\\beta_{2} = 0.999\\), weight decay 0.05 to finetune ViTs for 100 epochs using 32 V100 GPUs, with each GPU having a batch size of 32. The initial learning rate is 1e-3 with first 5 epochs for linear warm-up and decays to zero by a cosine learning rate scheduler. We set the layer-wise learning rate decay factor to 0.75 for ViT-Small and ViT-Base. We do not apply layer-wise learning rate decay for ViT-Tiny. For data augmentation, we adopt RandAugment [15] and set label smoothing to 0.1, mixup to 0.8. For the detection and instance segmentation task, We follow the ViTDet [34] framework by adapting ViT backbones to a simple feature pyramid, for object detection and instance segmentation. We adopt AdamW optimizer with momentum \\(\\beta_{1} = 0.9\\), \\(\\beta_{2} = 0.999\\) and weight decay 0.1 to train models on COCO. All models are trained on 64 V100 GPUs for 100 epochs with each GPU having 1 batch size with image resolution \\(1024 \\times 1024\\). The initial learning rate is \\(2e - 4\\), linearly warmed up for the first 10 epochs, and decayed to 0 by a cosine learning rate schedule. Models are trained" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.409 + ], + "angle": 0, + "content": "for 100 epochs. For the segmentation task, Our pretrained ViT models serve as the backbone of Mask2former [13], which is finetuned together with the segmentation layers on ADE20K. We adopt the AdamW optimizer with \\(\\beta_{1} = 0.9\\), \\(\\beta_{2} = 0.999\\), a mini-batch size of 16, a weight decay of 0.05, and an initial learning rate of 2e-4. The learning rate is decayed to 0 by a poly learning rate schedule. A learning rate multiplier is set to 0.1 for the backbone. The input image resolution is \\(512 \\times 512\\). Models are trained for 160K iterations using 8 V100 GPUs. For the segmentation anything task, Following [31], we take our pretrained lightweight ViT models, ViT-Tiny and ViT-Small, as the image encoder of SAM framework and finetune the encoder and decoder together of our EfficientSAM on SA-1B dataset for 5 epochs. We use the AdamW optimizer with a momentum, \\((\\beta_{1} = 0.9\\), \\(\\beta_{2} = 0.999)\\), a mini-batch size of 128, and a initial lreanning rate of \\(4e - 4\\). The learning rate is decayed to 0 by a linear learning rate schedule. We set weight decay to 0.1. We do not apply data augmentation. The input image resolution is \\(1024 \\times 1024\\). Our EfficientSAMs are trained on 64 A100 GPUs with 40GB GPU memory." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.41, + 0.897, + 0.683 + ], + "angle": 0, + "content": "Baselines and Evaluation Metrics. Baselines. For the classification task, we compare the performance of ViT backbones from different pretraining/distillation methods including MAE[26], SSTA[60], DMAE[2], BEiT[3], CAE[12], DINO[6], iBOT[73], DeiT[53], etc. For the detection and instance semantic task, and semantic segmentation task, we also compare with several pretrained ViT backbones for ViTDet[34] and Mask2former[13]. For the segment every thing task, we compare with SAM[31], FastSAM[71], and MobileSAM[68]. Evaluation Metrics. We evaluate our method and all baselines in terms of accuracy. Specifically, the accuracy metrics refer to top-1 accuracy for the classification task; \\(\\mathrm{AP^{box}}\\), \\(\\mathrm{AP^{mask}}\\), for the detection and instance segmentation task (AP: average precision); mIoU, for the semantic segmentation task (mIoU: mean intersection over union); mIoU, AP, \\(\\mathrm{APS}\\), \\(\\mathrm{AP^M}\\), \\(\\mathrm{AP^L}\\) for segment anything task. For efficiency metrics, we compare the number of model parameters or inference throughput." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.696, + 0.64, + 0.71 + ], + "angle": 0, + "content": "4.2. Main Results" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Image Classification. To evaluate the effectiveness of our proposed techniques on the image classification task, we apply the proposed SAMI idea to ViT models and compare their performance over baselines on ImageNet-1K. As shown in Tab. 1, our SAMI is compared with pretraining methods like MAE, iBOT, CAE, and BEiT, and distillation methods including DeiT and SSTA. SAMI-B achieves \\(84.8\\%\\) top-1 accuracy, which outperforms the pretraining baselines, MAE, DMAE, iBOT, CAE, and BEiT by \\(1.2\\%\\), \\(0.8\\%\\), \\(1.1\\%\\), \\(0.9\\%\\), and \\(0.4\\%\\) respectively. Compared with distillation methods such as DeiT and SSTA, SAMI also shows large improvements. For lightweight models such as ViT-Tiny" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "16115" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.091, + 0.088, + 0.462, + 0.381 + ], + "angle": 0, + "content": "
MethodBackboneTraining DataAcc.(%)
DeiT-Ti[53]ViT-TinyIN1K74.5
SSTA-Ti[60]ViT-TinyIN1K75.2
DMAE-Ti[2]ViT-TinyIN1K70.0
MAE-Ti[26]ViT-TinyIN1K75.2
SAMI-T (ours)ViT-TinySA1B (11M) + IN1K76.8
DeiT-S[53]ViT-SmallIN1K81.2
SSTA-S[60]ViT-SmallIN1K81.4
DMAE-S[2]ViT-SmallIN1K79.3
MAE-S[26]ViT-SmallIN1K81.5
BEiT-S[3]ViT-SmallD250M+IN22K+IN1K81.7
CAE-S[12]ViT-SmallD250M+IN1K82.0
DINO-S[6]ViT-SmallIN1K82.0
iBOT-S[73]ViT-SmallIN22K+1N1K82.3
SAMI-S (ours)ViT-SmallSA1B (11M) + IN1K82.7
DeiT-B[53]ViT-BaseIN1K83.8
DMAE-B[2]ViT-BaseIN1K84.0
BootMAE[18]ViT-BaseIN1K84.2
MAE-B[26]ViT-BaseIN1K83.6
BEiT-B[3]ViT-BaseD250M+IN22K+IN1K83.7
CAE-B[12]ViT-BaseD250M+IN1K83.9
DINO-B[6]ViT-BaseIN1K82.8
iBOT-B[73]ViT-BaseIN22K+1N1K84.4
SAMI-B (ours)ViT-BaseSA1B (11M) + IN1K84.8
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.39, + 0.472, + 0.434 + ], + "angle": 0, + "content": "Table 1. Image classification results on ImageNet-1K. IN is short for ImageNet. We use IN1K for pretraining/binetuning and add SA1B to the training data to indicate the need for original SAM." + }, + { + "type": "table", + "bbox": [ + 0.131, + 0.446, + 0.422, + 0.579 + ], + "angle": 0, + "content": "
MethodBackboneAPbboxAPmask
MAE-Ti[26]ViT-Tiny37.934.9
SAMI-Ti(ours)ViT-Tiny44.740.0
MAE-S[26]ViT-Small45.340.8
DeiT-S[53]ViT-Small47.241.9
DINO-S[6]ViT-Small49.143.3
iBOT-S[73]ViT-Small49.744.0
SAMI-S (ours)ViT-Small49.844.2
MAE-B[26]ViT-Base51.645.9
SAMI-B (ours)ViT-Base52.546.5
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.589, + 0.47, + 0.618 + ], + "angle": 0, + "content": "Table 2. Object detection and instance segmentation results on the MS COCO using ViTDet." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.644, + 0.47, + 0.673 + ], + "angle": 0, + "content": "and ViT-Small, SAMI reports a substantial gain compared to DeiT, SSTA, DMAE, and MAE." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.675, + 0.473, + 0.903 + ], + "angle": 0, + "content": "Object Detection and Instance Segmentation. We also extend the SAMI-pretrained ViT backbones to the downstream object detection and instance segmentation task and compare it with previous pretraining baseline on COCO dataset to evaluate its efficacy. Specifically, we take the pretrained ViT backbones and adapt them to a simple feature pyramid in the Mask R-CNN framework[25] for constructing the detector, ViTDet[34]. Tab. 2 shows the overall comparison between our SAMI and other baselines. We can see that our SAMI consistently achieves better performance over other baselines. SAMI-B obtains \\(0.9\\mathrm{AP^{bbox}}\\) and \\(0.6^{\\mathrm{mask}}\\) gains compared with MAE-B. For light-weight backbones, SAMIS and SAMI-Ti report substantial gains compared to MAE-Ti and MAE-S. Moreover, SAMI-S significantly outperforms DeiT-S by \\(2.6\\mathrm{AP^{bbox}}\\) and \\(2.3\\mathrm{AP^{mask}}\\). For other pretraining" + }, + { + "type": "table", + "bbox": [ + 0.573, + 0.088, + 0.828, + 0.19 + ], + "angle": 0, + "content": "
MethodBackbonemIOU
MAE-Ti[26]ViT-Tiny39.0
SAMI-Ti(ours)ViT-Tiny42.7
MAE-S[26]ViT-Small44.1
SAMI-S (ours)ViT-Small48.8
MAE-B[26]ViT-Base49.3
SAMI-B (ours)ViT-Base51.8
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.2, + 0.894, + 0.228 + ], + "angle": 0, + "content": "Table 3. Semantic segmentation results on the ADE20K dataset using Mask2former. The input resolution is \\(512 \\times 512\\)." + }, + { + "type": "table", + "bbox": [ + 0.514, + 0.242, + 0.885, + 0.322 + ], + "angle": 0, + "content": "
MethodCOCOLVIS
box1 click3 clickbox1 click3 click
SAM[31]78.455.674.178.959.875.2
MobileSAM[68]74.243.759.773.851.054.4
SAM-MAE-Ti[31]74.743.365.873.850.665.3
EfficientSAM-Ti (ours)75.745.567.274.352.766.8
EfficientSAM-S (ours)76.950.069.875.456.268.7
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.331, + 0.896, + 0.414 + ], + "angle": 0, + "content": "Table 4. Zero-shot single point valid mask evaluation results on COCO and LVIS. Following SAM[31], we uniformly sample random points within ground truth mask for click, and compute the tightest bounding box corresponding ground truth mask for box. SAM-MAE-Ti denotes SAM with pretrained MAE-Ti image encoder." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.442, + 0.897, + 0.655 + ], + "angle": 0, + "content": "belines, our SAMI still compares favorably to DINO and iBOT. This set of experiments validate the effectiveness of the proposed SAMI for providing pretrained detector backbones in the object detection and instance segmentation task. Semantic Segmentation. We further extend the pretrained backbones to the semantic segmentation task to evaluate its effectiveness. Specifically, we use ViT models as the backbone in Mask2former [13] framework to benchmark on the ADE20K dataset. As shown in Tab. 3, Mask2former with SAMI-pretrained backbones achieve better mIoU, i.e., \\(\\uparrow 2.5\\), \\(\\uparrow 4.7\\), and \\(\\uparrow 3.7\\) improvement over backbones with MAE pretraining [26] on ImageNet-1K. This set of experiments validate that our proposed techniques could be well generalized to various downstream tasks." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.665, + 0.864, + 0.683 + ], + "angle": 0, + "content": "4.3. EfficientSAMs for Segment Anything Task" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.689, + 0.896, + 0.885 + ], + "angle": 0, + "content": "Segment Anything task is a process of promptable segmentation to produce segmentation masks based on any form of the prompt, including point set, rough boxes or mask, freeform text. We follow SAM [31] and focus on point-based and box-based prompt segmentation on COCO/LVIS. We now test the generalization abilities of our model on segment anything task including zero-shot single point valid mask evaluation and zero-shot instance segmentation. We take the SAMI-pretrained lightweight backbones as the image encoder of SAM for building efficient SAMs, EfficientSAMs. Then we finetune EfficientSAMs on SA-1B dataset and report the performance on zero-shot single point valid mask evaluation and zero-shot instance segmentation." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.886, + 0.894, + 0.902 + ], + "angle": 0, + "content": "Zero-Shot Single Point Valid Mask Evaluation. Similar" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "16116" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.091, + 0.089, + 0.46, + 0.171 + ], + "angle": 0, + "content": "
MethodCOCOLVIS
APAPSAPMAPLAPAPSAPMAPL
ViTDet-H[34]51.032.054.368.946.635.058.066.3
SAM[31]46.530.851.061.744.732.557.665.5
MobileSAM[68]38.723.742.254.334.423.844.953.7
FastSAM[71]37.923.943.450.034.524.646.250.8
EfficientSAM-Ti (ours)42.326.746.257.439.928.951.059.9
EfficientSAM-S (ours)44.428.448.360.142.330.854.062.3
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.18, + 0.471, + 0.208 + ], + "angle": 0, + "content": "Table 5. Zero-shot instance segmentation results on COCO/LVIS. ViTDet boxes are prompted to perform zero-shot segmentation." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.235, + 0.471, + 0.37 + ], + "angle": 0, + "content": "to SAM[31], we evaluate segmenting an object from a single foreground point. For general interactive segmentation, we also consider object segmentation from a single box, and multiple points as introduced in [31]. To achieve this, we uniformly sample random points within ground truth mask for click, and compute the tightest bounding box corresponding to ground truth mask for box. Since our models are able to predict multiple masks, we only evaluate the most confident mask as SAM [31]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.371, + 0.471, + 0.552 + ], + "angle": 0, + "content": "Results. In Tab. 4, EfficientSAMs are compared with SAM, MobileSAM and SAM-MAE-Ti. On COCO, our EfficientSAM-Ti outperforms MobileSAM by \\(1.9\\mathrm{mIoU}\\) on 1 click and \\(1.5\\mathrm{mIoU}\\) on 1 box with comparable complexity. Our EfficientSAM-Ti with SAMI-pretrained weights also performs better than MAE-pretrained wights on COCO/LVIS interactive segmentation. We notice that our EfficientSAM-S only underperforms SAM by \\(1.5\\mathrm{mIoU}\\) on COCO box and \\(3.5\\mathrm{mIoU}\\) on LVIS box with \\(20x\\) fewer parameters. We find that our EfficientSAMs also show promising performance on multiple click compared with MobileSAM and SAM-MAE-Ti." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.553, + 0.47, + 0.628 + ], + "angle": 0, + "content": "Zero-Shot Instance Segmentation. Following SAM [31], instance segmentation task is performed by taking the bounding box (bbox) generated by ViTDet[34] as the prompt. The mask with the highest Intersection over Union (IoU) with the_bbox as the predicted mask." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.629, + 0.471, + 0.856 + ], + "angle": 0, + "content": "Results. In Tab. 5, we report AP, \\(\\mathrm{APS}^{\\mathrm{s}}\\), \\(\\mathrm{APM}^{\\mathrm{s}}\\), \\(\\mathrm{APL}^{\\mathrm{s}}\\) for zero-shot instance segmentation. We compare our EfficientSAM with MobileSAM and FastSAM. We can see that EfficientSAM-S obtains more than 6.5 AP on COCO and 7.8 AP on LVIS over FastSAM. For EfficientSAM-Ti, it still outperforms FastSAM by a large margin, 4.1 AP on COCO and 5.3 AP on LVIS, and MobileSAM by 3.6 AP on COCO and 5.5 AP on LVIS. Note that our EfficientSAMs are much light-weight than FastSAM, e.g., 9.8M parameters for efficientSAM-Ti vs 68M parameters for FastSAM. EfficientSAM-S also significantly reduces the gap between SAM with 0.6G parameters, only \\(\\sim2\\) AP reduction. These results demonstrate the extraordinary benefits of EfficientSAMs for zero-shot instance segmentation and validate the advantages of our SAMI pretraining method." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.856, + 0.471, + 0.901 + ], + "angle": 0, + "content": "Qualitative Evaluation. We now provide the qualitative results for a complementary understanding of instance segmentation capabilities of EfficientSAMs. Some examples" + }, + { + "type": "text", + "bbox": [ + 0.509, + 0.076, + 0.887, + 0.088 + ], + "angle": 0, + "content": "Input Image SAM[31] FastSAM[71] MobileSAM[68] EfficientSAM" + }, + { + "type": "image", + "bbox": [ + 0.504, + 0.089, + 0.89, + 0.131 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.504, + 0.137, + 0.89, + 0.173 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.504, + 0.179, + 0.89, + 0.221 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.5, + 0.233, + 0.894, + 0.261 + ], + "angle": 0, + "content": "Figure 3. Visualization result on point-prompt input with SAM, FastSAM, MobileSAM, and our EfficientSAM model." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.287, + 0.893, + 0.439 + ], + "angle": 0, + "content": "can be seen in Fig. 3, Fig. 4, and Fig. 5. Specifically, we report the predicted masks with two types of prompts, point and box as in MobileSAM [68] and also segment everything results. More qualitative results can be found in the supplement. These results demonstrate that our EfficientSAMs have competing capabilities when comparing to SAM. Note that our EfficientSAMs are much lightweight than SAM, and our models can effectively give decent segmentation results. This indicates that our models can be served as a complementary version of SAM for many practical tasks." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.444, + 0.887, + 0.455 + ], + "angle": 0, + "content": "Input Image SAM[31] FastSAM[71] MobileSAM[68] EfficientSAM" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.456, + 0.89, + 0.492 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.504, + 0.498, + 0.89, + 0.541 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.553, + 0.893, + 0.581 + ], + "angle": 0, + "content": "Figure 4. Visualization result on box-prompt input with SAM, FastSAM, MobileSAM, and our EfficientSAM model." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.597, + 0.893, + 0.703 + ], + "angle": 0, + "content": "Salient Instance Segmentation. Salient object segmentation [4] aims to segment the most visually attractive objects from an image. We extend interactive instance segmentation to salient instance segmentation without manually creating points/boxes. Specifically, we take a state-of-the-art saliency object detection model, \\(\\mathrm{U}^2\\)-net[45], to predict saliency map and uniformly sample 3 random points (3 click) within" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.712, + 0.887, + 0.722 + ], + "angle": 0, + "content": "Input Image SAM[31] FastSAM[71] MobileSAM[68] EfficientSAM" + }, + { + "type": "image", + "bbox": [ + 0.504, + 0.724, + 0.887, + 0.766 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.773, + 0.887, + 0.811 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.818, + 0.887, + 0.856 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.869, + 0.893, + 0.896 + ], + "angle": 0, + "content": "Figure 5. Visualization result on segment everything with SAM, FastSAM, MobileSAM, and our EfficientSAM model." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16117" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.159, + 0.089, + 0.393, + 0.155 + ], + "angle": 0, + "content": "
MethodLossTop-1 Acc.(%)
SAMI-Ti1 - Cosine76.1
SAMI-TiMSE76.8
SAMI-S1 - Cosine82.3
SAMI-SMSE82.7
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.165, + 0.47, + 0.193 + ], + "angle": 0, + "content": "Table 6. Ablation study on training loss of SAMI. MSE loss gives better classification results on ImageNet-1K." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.22, + 0.471, + 0.298 + ], + "angle": 0, + "content": "saliency map to perform instance segmentation with our EfficientSAM. In Fig. 6, we can see that our EfficientSAM can perform salient instance segmentation well. This preliminary exploration opens the potential to help people with hand impairments segment objects of interest in an image." + }, + { + "type": "image", + "bbox": [ + 0.114, + 0.304, + 0.429, + 0.466 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.476, + 0.471, + 0.534 + ], + "angle": 0, + "content": "Figure 6. Saliency-based automatic instance segmentation results. With the assistance of saliency map generated from \\(\\mathbf{U}^2\\)-net[45], our EfficientSAM is able to generate mask and perform automatic instance segmentation without manually creating points or boxes." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.56, + 0.241, + 0.573 + ], + "angle": 0, + "content": "4.4. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.583, + 0.469, + 0.613 + ], + "angle": 0, + "content": "We now analyze SAMI and EfficientSAMs through a series of ablation studies with ViT backbones." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.614, + 0.47, + 0.72 + ], + "angle": 0, + "content": "Reconstruction Loss. We study the effect of reconstruction loss on the performance of SAMI on ImageNet-1K. We compare our mean square error (MSE) reconstruction loss with cosine similarity loss. We find that MSE reconstruction loss performs better, shown in Tab. 6. This recommends a direct reconstruction of SAM features instead of the target with a high angular similarity." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.72, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Cross-Attention Decoder. To reconstruct SAM features, we directly use the output tokens from encoder and only take decoder to transform the masked tokens with cross-attention. We study how the performance varies with cross-attention decoder. When changing the decoder module in MAE to our cross-attention decoder in SAMI, we find that SAMI-B improves the performance from \\(84.4\\%\\) to \\(84.8\\%\\) on ImageNet-1K. The cross-attention decoder of SAMI also shows consistent improvement for ViT-Tiny and ViT-Small. Analogy to anchor points in AnchorDETR[58], the output tokens from encoder are already learned well by directly aligning the SAM features, which can serve as anchor tokens" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.092, + 0.877, + 0.107 + ], + "angle": 0, + "content": "for help masked tokens align via cross-attention decoder." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.107, + 0.892, + 0.167 + ], + "angle": 0, + "content": "Mask Ratio. We explore how the performance varies with different mask ratio in SAMI. The observations are consistent with MAE [26] that a high mask ratio, \\(75\\%\\), tends to produce good results." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.168, + 0.895, + 0.273 + ], + "angle": 0, + "content": "Reconstruction Target. We study the impact of reconstruction target. We take a different encoder from CLIP [46] to generate features as the reconstruction target in SAMI. Aligning features from CLIP encoder can also outperform MAE by \\(0.8\\%\\) for a ViT-Tiny model on ImageNet-1K. This demonstrates that masked image pretraining benefits from powerful guided reconstruction." + }, + { + "type": "image", + "bbox": [ + 0.553, + 0.285, + 0.843, + 0.403 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.5, + 0.415, + 0.892, + 0.457 + ], + "angle": 0, + "content": "Figure 7. Ablation study on training steps for EfficientSAMs on MS COCO dataset. Zero-shot single point valid mask evaluation with a single box prompt is performed for the ablation." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.471, + 0.892, + 0.606 + ], + "angle": 0, + "content": "Effects of Finetuning Steps for EfficientSAMs. We explore the effect of fintuning steps for EfficientSAMs. As illustrated in Fig. 7, EfficientSAM-Ti and EfficientSAM-S achieve decent performance even at 0.1 epoch. For 1 epoch, the performance gain is larger than \\(2.5\\mathrm{mIoU}\\). The final performance of EfficientSAM-S reaches \\(76.9\\mathrm{mIoU}\\), which is only \\(1.5\\mathrm{mIoU}\\) lower than SAM. These results demonstrate the advantages of SAMI-pretrained image encoders and our EfficientSAMs." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.607, + 0.894, + 0.683 + ], + "angle": 0, + "content": "Efficiency. We compare the throughput (images per second) of our EfficientSAM with SAM and other models on A100 in Fig. 1. Our EfficientSAM reduces the runtime/parameters of SAM by \\(\\sim 20\\mathrm{x}\\). We also observed similar speedup (\\(\\sim 20\\mathrm{x}\\)) comparing to SAM w.r.t. FLOPs." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.695, + 0.619, + 0.71 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.895, + 0.902 + ], + "angle": 0, + "content": "We proposed a masked image pretraining approach, SAMI, to explore the potential of ViTs under the guidance of SAM foundation model. SAMI improves masked image pretraining by reconstructing the latent features from SAM image encoder to transfer knowledge from vision foundation model to ViTs. Extensive experiments on image classification, object detection and instance segmentation, semantic segmentation, and the segment anything task consistently validate SAMI's advantages. We also demonstrate that SAMI helps build efficient SAMs with pretrained light-weight encoders. Our preliminary work suggests that SAMI has potential applications beyond efficient segment anything task." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16118" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.472, + 0.17 + ], + "angle": 0, + "content": "[1] Pablo Arbelaez, Michael Maire, Charless Fowlkes, and Jitendra Malik. Contour detection and hierarchical image segmentation. IEEE transactions on pattern analysis and machine intelligence, 33(5):898-916, 2010. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.171, + 0.472, + 0.242 + ], + "angle": 0, + "content": "[2] Yutong Bai, Zeyu Wang, Junfei Xiao, Chen Wei, Huiyu Wang, Alan L Yuille, Yuyin Zhou, and Cihang Xie. Masked autoencoders enable efficient knowledge distillers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24256-24265, 2023. 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.242, + 0.472, + 0.283 + ], + "angle": 0, + "content": "[3] Hangbo Bao, Li Dong, Songhao Piao, and Furu Wei. Beit: Bert pre-training of image transformers. arXiv preprint arXiv:2106.08254, 2021. 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.284, + 0.472, + 0.325 + ], + "angle": 0, + "content": "[4] Ali Borji, Ming-Ming Cheng, Qibin Hou, Huaizu Jiang, and Jia Li. Salient object detection: A survey. Computational visual media, 5:117-150, 2019. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.326, + 0.472, + 0.395 + ], + "angle": 0, + "content": "[5] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In European conference on computer vision, pages 213-229. Springer, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.396, + 0.472, + 0.466 + ], + "angle": 0, + "content": "[6] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9650-9660, 2021. 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.467, + 0.472, + 0.522 + ], + "angle": 0, + "content": "[7] Jun Cen, Yizheng Wu, Kewei Wang, Xingyi Li, Jingkang Yang, Yixuan Pei, Lingdong Kong, Ziwei Liu, and Qifeng Chen. Sad: Segment any rgbd. arXiv preprint arXiv:2305.14207, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.523, + 0.472, + 0.564 + ], + "angle": 0, + "content": "[8] Jiaqi Chen, Zeyu Yang, and Li Zhang. Semantic segment anything. https://github.com/fudan-zvg/Semantic-Segment-Anything, 2023.2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.565, + 0.472, + 0.62 + ], + "angle": 0, + "content": "[9] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International conference on machine learning, pages 1597-1607. PMLR, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.621, + 0.472, + 0.704 + ], + "angle": 0, + "content": "[10] Tianrun Chen, Lanyun Zhu, Chaotao Deng, Runlong Cao, Yan Wang, Shangzhan Zhang, Zejian Li, Lingyun Sun, Ying Zang, and Papa Mao. Sam-adapter: Adapting segment anything in underperformed scenes. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3367-3375, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.705, + 0.472, + 0.761 + ], + "angle": 0, + "content": "[11] Xinlei Chen and Kaiming He. Exploring simple siamese representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 15750-15758, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.761, + 0.472, + 0.831 + ], + "angle": 0, + "content": "[12] Xiaokang Chen, Mingyu Ding, Xiaodi Wang, Ying Xin, Shentong Mo, Yunhao Wang, Shumin Han, Ping Luo, Gang Zeng, and Jingdong Wang. Context autoencoder for self-supervised representation learning. International Journal of Computer Vision, pages 1-16, 2023. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.832, + 0.472, + 0.902 + ], + "angle": 0, + "content": "[13] Bowen Cheng, Ishan Misra, Alexander G Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1290–1299, 2022. 5, 6" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.472, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.135 + ], + "angle": 0, + "content": "[14] Yangming Cheng, Lulei Li, Yuanyou Xu, Xiaodi Li, Zongxin Yang, Wenguan Wang, and Yi Yang. Segment and track anything. arXiv preprint arXiv:2305.06558, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.136, + 0.894, + 0.207 + ], + "angle": 0, + "content": "[15] Ekin D Cubuk, Barret Zoph, Jonathon Shlens, and Quoc V Le. Randaugment: Practical automated data augmentation with a reduced search space. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops, pages 702-703, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.207, + 0.894, + 0.263 + ], + "angle": 0, + "content": "[16] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.264, + 0.894, + 0.347 + ], + "angle": 0, + "content": "[17] Ruining Deng, Can Cui, Quan Liu, Tianyuan Yao, Lucas W Remedios, Shunxing Bao, Bennett A Landman, Lee E Whelless, Lori A Coburn, Keith T Wilson, et al. Segment anything model (sam) for digital pathology: Assess zero-shot segmentation on whole slide imaging. arXiv preprint arXiv:2304.04155, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.349, + 0.894, + 0.419 + ], + "angle": 0, + "content": "[18] Xiaoyi Dong, Jianmin Bao, Ting Zhang, Dongdong Chen, Weiming Zhang, Lu Yuan, Dong Chen, Fang Wen, and Nenghai Yu. Bootstrapped masked autoencoders for vision bert pretraining. In European Conference on Computer Vision, pages 247-264. Springer, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.42, + 0.894, + 0.503 + ], + "angle": 0, + "content": "[19] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.504, + 0.894, + 0.575 + ], + "angle": 0, + "content": "[20] Haoqi Fan, Bo Xiong, Karttikeya Mangalam, Yanghao Li, Zhicheng Yan, Jitendra Malik, and Christoph Feichtenhofer. Multiscale vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6824-6835, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.576, + 0.894, + 0.646 + ], + "angle": 0, + "content": "[21] Shanghua Gao, Zhijie Lin, Xingyu Xie, Pan Zhou, Ming-Ming Cheng, and Shuicheng Yan. Editanything: Empowering unparalleled flexibility in image editing and generation. In Proceedings of the 31st ACM International Conference on Multimedia, pages 9414–9416, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.647, + 0.894, + 0.729 + ], + "angle": 0, + "content": "[22] Benjamin Graham, Alaaeldin El-Nouby, Hugo Touvron, Pierre Stock, Armand Joulin, Hervé Jégou, and Matthijs Douze. Levit: a vision transformer in convnet's clothing for faster inference. In Proceedings of the IEEE/CVF international conference on computer vision, pages 12259-12269, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.73, + 0.894, + 0.788 + ], + "angle": 0, + "content": "[23] Agrim Gupta, Piotr Dólár, and Ross Girshick. Lvis: A dataset for large vocabulary instance segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.789, + 0.894, + 0.859 + ], + "angle": 0, + "content": "[24] Dongsheng Han, Chaoning Zhang, Yu Qiao, Maryam Qamar, Yuna Jung, SeungKyu Lee, Sung-Ho Bae, and Choong Seon Hong. Segment anything model (sam) meets glass: Mirror and transparent objects cannot be easily detected. arXiv preprint arXiv:2305.00278, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.86, + 0.894, + 0.902 + ], + "angle": 0, + "content": "[25] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In Proceedings of the IEEE international conference on computer vision, pages 2961-2969, 2017. 6" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "16119" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.161 + ], + "angle": 0, + "content": "[26] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollar, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16000-16009, 2022. 2, 3, 4, 5, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.163, + 0.472, + 0.204 + ], + "angle": 0, + "content": "[27] Geoffrey Hinton, Oriol Vinyals, and Jeff Dean. Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.207, + 0.472, + 0.261 + ], + "angle": 0, + "content": "[28] Zejiang Hou, Fei Sun, Yen-Kuang Chen, Yuan Xie, and Sun-Yuan Kung. Milan: Masked image pretraining on language assisted representation. arXiv preprint arXiv:2208.06049, 2022.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.264, + 0.472, + 0.304 + ], + "angle": 0, + "content": "[29] Jiaxi Jiang and Christian Holz. Restore anything pipeline: Segment anything meets image restoration. arXiv preprint arXiv:2305.13093, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.307, + 0.472, + 0.348 + ], + "angle": 0, + "content": "[30] Glenn Jocher, Ayush Chaurasia, and Jing Qiu. Yolo by ultralytics. https://github.com/ultralytics/ultralytics, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.35, + 0.472, + 0.417 + ], + "angle": 0, + "content": "[31] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. arXiv preprint arXiv:2304.02643, 2023. 1, 2, 3, 4, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.421, + 0.472, + 0.462 + ], + "angle": 0, + "content": "[32] Brett Koonce and Brett Koonce. *Mobilenetv3: Convolutional Neural Networks with Swift for Tensorflow: Image Recognition and Dataset Categorization*, pages 125–144, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.464, + 0.472, + 0.532 + ], + "angle": 0, + "content": "[33] Jiashi Li, Xin Xia, Wei Li, Huixia Li, Xing Wang, Xuefeng Xiao, Rui Wang, Min Zheng, and Xin Pan. Next-vit: Next generation vision transformer for efficient deployment in realistic industrial scenarios. arXiv preprint arXiv:2207.05501, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.535, + 0.472, + 0.589 + ], + "angle": 0, + "content": "[34] Yanghao Li, Hanzi Mao, Ross Girshick, and Kaiming He. Exploring plain vision transformer backbones for object detection. In European Conference on Computer Vision, pages 280-296. Springer, 2022. 2, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.592, + 0.472, + 0.66 + ], + "angle": 0, + "content": "[35] Yanyu Li, Geng Yuan, Yang Wen, Ju Hu, Georgios Evangelidis, Sergey Tulyakov, Yanzhi Wang, and Jian Ren. Efficientformer: Vision transformers at mobilenet speed. Advances in Neural Information Processing Systems, 35:12934-12949, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.663, + 0.472, + 0.732 + ], + "angle": 0, + "content": "[36] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In European conference on computer vision, pages 740-755. Springer, 2014. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.734, + 0.472, + 0.829 + ], + "angle": 0, + "content": "[37] Ruiping Liu, Jiaming Zhang, Kunyu Peng, Junwei Zheng, Ke Cao, Yufan Chen, Kailun Yang, and Rainer Stiefelhagen. Open scene understanding: Grounded situation recognition meets segment anything for helping people with visual impairments. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1857-1867, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.832, + 0.472, + 0.901 + ], + "angle": 0, + "content": "[38] Xinyu Liu, Houwen Peng, Ningxin Zheng, Yuqing Yang, Han Hu, and Yixuan Yuan. Efficientvit: Memory efficient vision transformer with cascaded group attention. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14420-14430, 2023. 3" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.893, + 0.162 + ], + "angle": 0, + "content": "[39] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10012-10022, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.163, + 0.894, + 0.204 + ], + "angle": 0, + "content": "[40] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In International Conference on Learning Representations, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.205, + 0.894, + 0.232 + ], + "angle": 0, + "content": "[41] Jun Ma and Bo Wang. Segment anything in medical images. arXiv preprint arXiv:2304.12306, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.233, + 0.894, + 0.287 + ], + "angle": 0, + "content": "[42] Sachin Mehta and Mohammad Rastegari. Mobilevit: Lightweight, general-purpose, and mobile-friendly vision transformer. In International Conference on Learning Representations, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.288, + 0.894, + 0.355 + ], + "angle": 0, + "content": "[43] Deepak Pathak, Philipp Krahenbuhl, Jeff Donahue, Trevor Darrell, and Alexei A Efros. Context encoders: Feature learning by inpainting. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2536-2544, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.357, + 0.894, + 0.412 + ], + "angle": 0, + "content": "[44] Zhiliang Peng, Li Dong, Hangbo Bao, Qixiang Ye, and Furu Wei. Beit v2: Masked image modeling with vector-quantized visual tokenizers. arXiv preprint arXiv:2208.06366, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.414, + 0.894, + 0.468 + ], + "angle": 0, + "content": "[45] Xuebin Qin, Zichen Zhang, Chenyang Huang, Masood Dehghan, Osmar Zaiane, and Martin Jagersand. U2-net: Going deeper with nested u-structure for salient object detection. page 107404, 2020. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.47, + 0.894, + 0.536 + ], + "angle": 0, + "content": "[46] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In International Conference on Machine Learning, pages 8821-8831. PMLR, 2021. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.539, + 0.894, + 0.593 + ], + "angle": 0, + "content": "[47] Adriana Romero, Nicolas Ballas, Samira Ebrahimi Kahou, Antoine Chassang, Carlo Gatta, and Yoshua Bengio. Fitnets: Hints for thin deep nets. arXiv preprint arXiv:1412.6550, 2014.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.595, + 0.894, + 0.662 + ], + "angle": 0, + "content": "[48] Mark Sandler, Andrew Howard, Menglong Zhu, Andrey Zh-moginov, and Liang-Chieh Chen. Mobilenetv2: Inverted residuals and linear bottlenecks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4510-4520, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.664, + 0.894, + 0.705 + ], + "angle": 0, + "content": "[49] Qiuhong Shen, Xingyi Yang, and Xinchao Wang. Anything-3d: Towards single-view anything reconstruction in the wild. arXiv preprint arXiv:2304.10261, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.707, + 0.894, + 0.748 + ], + "angle": 0, + "content": "[50] Ao Sun, Pingchuan Ma, Yuanyuan Yuan, and Shuai Wang. Explain any concept: Segment anything meets concept-based explanation. arXiv preprint arXiv:2305.10289, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.749, + 0.894, + 0.788 + ], + "angle": 0, + "content": "[51] Lv Tang, Haoke Xiao, and Bo Li. Can sam segment anything? when sam meets camouflaged object detection. arXiv preprint arXiv:2304.04709, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.79, + 0.894, + 0.831 + ], + "angle": 0, + "content": "[52] Shehbaz Tariq, Brian Estadimas Arfeto, Chaoning Zhang, and Hyundong Shin. Segment anything meets semantic communication. arXiv preprint arXiv:2306.02094, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.832, + 0.894, + 0.9 + ], + "angle": 0, + "content": "[53] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Hervé Jégou. Training data-efficient image transformers & distillation through attention. In International conference on machine learning, pages 10347-10357. PMLR, 2021. 2, 5, 6" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "16120" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.147 + ], + "angle": 0, + "content": "[54] Koen EA Van de Sande, Jasper RR Uijlings, Theo Gevers, and Arnold WM Smeulders. Segmentation as selective search for object recognition. In 2011 international conference on computer vision, pages 1879-1886. IEEE, 2011. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.149, + 0.472, + 0.204 + ], + "angle": 0, + "content": "[55] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.205, + 0.472, + 0.273 + ], + "angle": 0, + "content": "[56] Pascal Vincent, Hugo Larochelle, Isabelle Lajoie, Yoshua Bengio, Pierre-Antoine Manzagol, and León Bottou. Stacked denoising autoencoders: Learning useful representations in a deep network with a local denoising criterion. Journal of machine learning research, 11(12), 2010. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.274, + 0.471, + 0.342 + ], + "angle": 0, + "content": "[57] Xinlong Wang, Rufeng Zhang, Chunhua Shen, Tao Kong, and Lei Li. Dense contrastive learning for self-supervised visual pre-training. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3024-3033, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.344, + 0.471, + 0.399 + ], + "angle": 0, + "content": "[58] Yingming Wang, Xiangyu Zhang, Tong Yang, and Jian Sun. Anchor detr: Query design for transformer-based detector. In Proceedings of the AAAI conference on artificial intelligence, pages 2567-2575, 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.401, + 0.471, + 0.468 + ], + "angle": 0, + "content": "[59] Chen Wei, Haoqi Fan, Saining Xie, Chao-Yuan Wu, Alan Yuille, and Christoph Feichtenhofer. Masked feature prediction for self-supervised visual pre-training. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14668-14678, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.47, + 0.47, + 0.538 + ], + "angle": 0, + "content": "[60] Haiyan Wu, Yuting Gao, Yinqi Zhang, Shaohui Lin, Yuan Xie, Xing Sun, and Ke Li. Self-supervised models are good teaching assistants for vision transformers. In Proceedings of the 39th International Conference on Machine Learning, pages 24031-24042. PMLR, 2022. 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.54, + 0.47, + 0.607 + ], + "angle": 0, + "content": "[61] Kan Wu, Jinnian Zhang, Houwen Peng, Mengchen Liu, Bin Xiao, Jianlong Fu, and Lu Yuan. Tinyvit: Fast pretraining distillation for small vision transformers. In European Conference on Computer Vision, pages 68-85. Springer, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.609, + 0.47, + 0.678 + ], + "angle": 0, + "content": "[62] Zhenda Xie, Yutong Lin, Zheng Zhang, Yue Cao, Stephen Lin, and Han Hu. Propagate yourself: Exploring pixel-level consistency for unsupervised visual representation learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16684-16693, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.679, + 0.47, + 0.747 + ], + "angle": 0, + "content": "[63] Zhenda Xie, Zheng Zhang, Yue Cao, Yutong Lin, Jianmin Bao, Zhuliang Yao, Qi Dai, and Han Hu. Simmim: A simple framework for masked image modeling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9653-9663, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.749, + 0.47, + 0.803 + ], + "angle": 0, + "content": "[64] Jing Yang, Brais Martinez, Adrian Bulat, and Georgios Tzimiropoulos. Knowledge distillation via softmax regression representation learning. In International Conference on Learning Representations, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.805, + 0.47, + 0.845 + ], + "angle": 0, + "content": "[65] Jinyu Yang, Mingqi Gao, Zhe Li, Shang Gao, Fangjing Wang, and Feng Zheng. Track anything: Segment anything meets videos. arXiv preprint arXiv:2304.11968, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.847, + 0.47, + 0.9 + ], + "angle": 0, + "content": "[66] Haoran You, Yunyang Xiong, Xiaoliang Dai, Bichen Wu, Peizhao Zhang, Haoqi Fan, Peter Vajda, and Yingyan Celine Lin. Castling-vit: Compressing self-attention via switching towards linear-angular attention at vision transformer inference." + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.892, + 0.121 + ], + "angle": 0, + "content": "In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14431-14442, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.123, + 0.894, + 0.177 + ], + "angle": 0, + "content": "[67] Tao Yu, Runseng Feng, Ruoyu Feng, Jinming Liu, Xin Jin, Wenjun Zeng, and Zhibo Chen. Inpaint anything: Segment anything meets image inpainting. arXiv preprint arXiv:2304.06790, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.179, + 0.894, + 0.245 + ], + "angle": 0, + "content": "[68] Chaoning Zhang, Dongshen Han, Yu Qiao, Jung Uk Kim, Sung-Ho Bae, Seungkyu Lee, and Choong Seon Hong. Faster segment anything: Towards lightweight sam for mobile applications. arXiv preprint arXiv:2306.14289, 2023. 2, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.249, + 0.894, + 0.29 + ], + "angle": 0, + "content": "[69] Xiao Feng Zhang, Tian Yi Song, and Jia Wei Yao. Deshadow—anything: When segment anything model meets zero-shot shadow removal. arXiv preprint arXiv:2309.11715, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.292, + 0.894, + 0.347 + ], + "angle": 0, + "content": "[70] Borui Zhao, Quan Cui, Renjie Song, Yiyu Qiu, and Jiajun Liang. Decoupled knowledge distillation. In Proceedings of the IEEE/CVF Conference on computer vision and pattern recognition, pages 11953-11962, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.349, + 0.894, + 0.389 + ], + "angle": 0, + "content": "[71] Xu Zhao, Wenchao Ding, Yongqi An, Yinglong Du, Tao Yu, Min Li, Ming Tang, and Jinqiao Wang. Fast segment anything. arXiv preprint arXiv:2306.12156, 2023. 2, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.391, + 0.894, + 0.446 + ], + "angle": 0, + "content": "[72] Bolei Zhou, Hang Zhao, Xavier Puig, Sanja Fidler, Adela Barriuso, and Antonio Torralba. Scene parsing through ade20k dataset. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 633-641, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.448, + 0.894, + 0.502 + ], + "angle": 0, + "content": "[73] Jinghao Zhou, Chen Wei, Huiyu Wang, Wei Shen, Cihang Xie, Alan Yuille, and Tao Kong. ibot: Image bert pre-training with online tokenizer. arXiv preprint arXiv:2111.07832, 2021. 5, 6" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.894, + 0.502 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.518, + 0.957 + ], + "angle": 0, + "content": "16121" + } + ] +] \ No newline at end of file diff --git a/2024/EfficientSAM_ Leveraged Masked Image Pretraining for Efficient Segment Anything/f1e649e9-e5d4-48b3-8018-af1f58065ff2_origin.pdf b/2024/EfficientSAM_ Leveraged Masked Image Pretraining for Efficient Segment Anything/f1e649e9-e5d4-48b3-8018-af1f58065ff2_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..4cb7c1e7850f7d6bfcbe6554120ccd4fc966267e --- /dev/null +++ b/2024/EfficientSAM_ Leveraged Masked Image Pretraining for Efficient Segment Anything/f1e649e9-e5d4-48b3-8018-af1f58065ff2_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:593edf002923578ef5e3baa0e6c077f21102bfb0ad26c453d510e626ba326f61 +size 9586452 diff --git a/2024/EfficientSAM_ Leveraged Masked Image Pretraining for Efficient Segment Anything/full.md b/2024/EfficientSAM_ Leveraged Masked Image Pretraining for Efficient Segment Anything/full.md new file mode 100644 index 0000000000000000000000000000000000000000..224d14a8c190e7b624f0a924388cd1759f0d5b72 --- /dev/null +++ b/2024/EfficientSAM_ Leveraged Masked Image Pretraining for Efficient Segment Anything/full.md @@ -0,0 +1,307 @@ +# EfficientSAM: Leveraged Masked Image Pretraining for Efficient Segment Anything + +Yunyang Xiong, Bala Varadarajan*, Lemeng Wu*, Xiaoyu Xiang, Fanyi Xiao, Chenchen Zhu, Xiaoliang Dai, Dilin Wang, Fei Sun, Forrest Iandola, Raghuraman Krishnamoorthi, Vikas Chandra Meta AI Research + +https://yformer.github.io/efficient-sam/ + +# Abstract + +Segment Anything Model (SAM) has emerged as a powerful tool for numerous vision applications. A key component that drives the impressive performance for zero-shot transfer and high versatility is a super large Transformer model trained on the extensive high-quality SA-1B dataset. While beneficial, the huge computation cost of SAM model has limited its applications to wider real-world applications. To address this limitation, we propose EfficientSAMs, lightweight SAM models that exhibits decent performance with largely reduced complexity. Our idea is based on leveraging masked image pretraining, SAMI, which learns to reconstruct features from SAM image encoder for effective visual representation learning. Further, we take SAMI-pretrained light-weight image encoders and mask decoder to build EfficientSAMs, and finetune the models on SA-1B for segment anything task. We perform evaluations on multiple vision tasks including image classification, object detection, instance segmentation, and semantic segmentation, and find that our proposed pretraining method, SAMI, consistently outperforms other masked image pretraining methods. On segment anything task such as zero-shot instance segmentation, our EfficientSAMs with SAMI-pretrained lightweight image encoders perform favorably with a significant gain (e.g., $\sim 4$ AP on COCO/LVIS) over other fast SAM models. Our EfficientSAM code and models are available at here. + +# 1. Introduction + +Segment Anything Model (SAM) [31] has been very successful in the vision field, achieving state-of-the-art performance in a variety of image segmentation tasks such as zero-shot edge detection[1, 31], zero-shot object proposal generation[31, 54], and zero-shot instance segmentation[31], and many other real-world applications[24, 37, 41, 50-52]. The key feature of SAM is a prompt-based Vision Trans + +![](images/92e92a677f7a1fde0e71de11ef6055bfa06f71d67f0d8eadb98663cb69831ba3.jpg) + +![](images/7a852e0935c384ff68ee367b16efca629b1a790931f382b6c3f3626e5cbe12bb.jpg) +Figure 1. The comparative analysis result. (Top) The overview of EfficientSAM model by taking a well-pretrained light-weight image encoder for instance segmentation with largely reduced complexity. (Bottom) Throughput/Parameter/Performance comparison of EfficientSAM, MobileSAM, FastSAM, and SAM for zero-shot instance segmentation on COCO. We benchmark throughput (image per second) of all models on a single NVIDIA A100 with one box prompt. The input image resolution is $1024 \times 1024$ . Our EfficientSAMs outperform MobileSAM and FastSAM by a large margin, $\sim 4$ AP, with comparable complexity. Our EfficientSAM-S reduces the inference time of SAM by $\sim 20x$ and the parameter size by $\sim 20x$ with a small performance drop, 44.4 AP vs 46.5 AP. + +former (ViT)[19]model trained on a large-scale visual dataset with more than 1B masks from 11M images, SA-1B[31], which allows segmenting any object on a given image. This ability of Segment Anything makes SAM a foundation model in vision and enables its applications even beyond vision. + +Despite the forgoing advantages, the model of SAM turns out to be a major efficiency bottleneck for practical deployment since the architecture of SAM, especially, the image encoder (e.g., ViT-H) is very expensive. Note that ViT-H image encoder in SAM has 632M parameters while the + +prompt-based decoder only take 3.87M parameters. As a result, it leads to high computation and memory costs when using SAM to perform segment anything tasks in practice, which makes it challenging for real-time applications. + +To address this challenge, several recent works have proposed strategies that avoid incurring the huge cost when applying SAM for the prompt-based instance segmentation. For example, [68] suggests distilling the knowledge from the default ViT-H image encoder to a tiny ViT image encoder. In [71], the computation cost can be reduced with a real-time CNN-based architecture for Segment Anything task. + +In this paper, we propose using a well-pretrained lightweight ViT image encoder (e.g., ViT-Tiny/-Small[53]) to reduce the complexity of SAM while maintain decent performance. Our method, SAM-leveraged masked image pertraining (SAMI), produces desired pretrained lightweight ViT backbones for segment anything task. This is achieved by leveraging the celebrated MAE[26] pretraining method with SAM model to obtain high-quality pretrained ViT encoders. Specifically, our SAMI makes use of SAM encoder, ViT-H, to generate feature embedding and train a masked image model with lightweight encoders to reconstruct features from ViT-H of SAM instead of image patches. This leads to generalized ViT backbones, which can be used for downstream tasks such as image classification, object detection, and segment anything. Then we finetune the pretrained lightweight encoders with SAM decoders for segment anything task[31]. + +To evaluate our method, we consider a transfer learning setting for masked image pretraining, where models are first pretrained with a reconstructive loss on ImageNet with image resolution $224 \times 224$ , and then finetuned on target tasks using supervised data. Our SAMI learns lightweight encoders that generalize well. With SAMI pretraining, we can train models like ViT-Tiny/-Small/-Base on ImageNet-1K with improved generalization performance. For a ViT-Small model, we achieve $82.7\%$ top-1 accuracy when finetuned on ImageNet-1K with 100 epochs, which outperforms other state-of-the-art image pretraining baselines. We also finetune our pretrained models on object detection, instance segmentation, and semantic segmentation. Across all these tasks, our pretraining method achieves better results than other pretraining baselines, and more importantly, we observe significant gains for small models. Further, we evaluate our models on Segment Anything task. On zero-shot instance segmentation, our model performs well compared to recent lightweight SAM methods, including FastSAM, by a margin of 4.1 /5.2 AP on COCO/LVIS. + +Our main contribution can be summarized as follows: + +- We propose a SAM-leveraged masked image pretraining framework called SAMI, which trains the model to reconstruct features from SAM ViT-H image encoder. We show that this can substantially improve the + +performance of image masked pretraining method. + +- We demonstrate that SAMI-pretrained backbones can generalize well to many tasks including image classification, object detection, and semantic segmentation. +- We deliver EfficientSAMs, light-weight SAM models with state-of-the-art quality-efficiency trade-offs (Fig. 1), which is complementary to SAM for practical deployment. + +# 2. Related Work + +We briefly review relevant works on segment anything model, vision transformers, knowledge distillation, and masked image pretraining. + +# 2.1. Segment Anything Model + +SAM[31] has been considered as a milestone vision foundation model, which can segment any object in the image based on interaction prompts. SAM has shown remarkable zero-shot transfer performance and high versatility for many vision tasks including a variety of segmentation application[7, 8, 10, 17], in-painting[67], image restoration[29], image editing[21], image shadow removal[69], object tracking[14, 65], and 3D object reconstruction[49]. There are many other works attempting to generalize SAM to real-world scenarios, including medical image segmentation[41], camouflaged object detection[51], transparent object detection[24], concept-based explanation[50], semantic communication[52], and helping people with visual impairments[37]. Due to its wide real-world applications, practical deployment of SAM has also gained increasing attention. Several recent works including [68, 71] have proposed strategies to reduce the computation costs of SAM. FastSAM[68] develops a CNN-based architecture, YOLOv8-seg[30], to segment all objects in an image for efficiency improvement. MobileSAM[71] presents a decoupled distillation for obtaining a lightweight image encoder of SAM. Our work focuses on dealing with this efficiency issue for practical deployment of SAM. + +# 2.2. Vision Transformers + +ViTs [19] have achieved impressive performance in vision applications[5, 20, 26, 34, 39, 44]. ViTs demonstrate the advantages of and generalization over their CNN counterparts[26]. There are also a number of works on efficient ViTs for deployment. Smaller ViTs such as ViT-Small/DeiT-Small and ViT-Tiny/DeiT-Tiny are introduced in [53] for complementing ViT-Huge, ViT-Large, and ViT-Base in [19]. Motivated by the ability of convolution to capture local information with reduced parameters and computation costs, MobileViT[42] explore combining ViT with convolutions, which outperforms light-weight CNN models such as MobileNet-v2/v3[32, 48] with better task-level generalization properties and reduced memory size and computation cost. This trick has been used in many follow-up works + +including LeViT[22], EfficientFormer[35], Next-ViT[33], Tiny-ViT[61], Castling-ViT[66], EfficientViT[38]. This line of progress for designing efficient ViTs is orthogonal to our EfficientSAM work towards building efficient SAM. + +# 2.3. Knowledge Distillation + +Knowledge distillation (KD) is a technique to improve the performance of deep learning models without changing their architectures. [27] is a pioneering work to distill the dark knowledge from a larger teacher model to a smaller student model. The learning of a student model is supervised by the hard labels and the soft labels from a teacher model. This practice is followed by multiple works which aim to make better use of soft labels to transfer more knowledge. In [64], the distillation method decouples representation learning and classification. Decoupled knowledge distillation[70] separates the classical KD loss into two parts, target class knowledge distillation and non-target class knowledge distillation, which improves the effectiveness and flexibility of knowledge transfer. Another line of work is to transfer knowledge from intermediate features. FitNet [47] is a pioneering work by distilling the semantic information from the teacher model's intermediate feature directly. In [60], a self-supervised teaching assistant (SSTA) is introduced to guide the learning of a ViT-based student model with a supervised teacher together. [2] studies the potential of knowledge distillation from pre-training MAE model by aligning the intermediate features between the larger MAE teacher model and smaller MAE student model. + +# 2.4. Masked Image Pretraining + +Self-supervised pretraining approaches [6] have attracted significant attention in computer vision. One line of work is contrastive learning methods[9, 11, 57, 62], which learn augmentation in-variance by imposing high similarity between different augmented views of a given image. While the learned representation shows good properties such as high linear separability, contrastive learning methods relies on strong augmentation and negative sampling. Another interesting line of work is masked image modeling (MIM), which helps models learn meaningful representations by reconstructing masked image patches. The MIM pioneering works focus on using denoising autoencoders[56] and context encoders[43] to train vision Transformer with masked prediction objectives. There are various promising works on using MIM for self-supervised image pretraining. BEiT[3] is the first one that adopts MIM for ViT pretraining to predict visual tokens. In BEiTv2[44], a semantic-rich image tokenizer is utilized for a better reconstruction target. In MaskFeat[59], reconstructing the local gradient features generated from HOG descriptor leads to effective visual pretraining. In SimMIM[63] and MAE[26], directly reconstructing the pixel values of the masked image patches achieves effec + +tive visual representation learning. There are MAE-based follow-up works that use large teacher models to guide MAE pretraining[2, 28, 60]. Our work is built on MAE and finds that leveraging MAE to reconstruct the features from SAM image encoder enables the pretraining to be highly effective. + +# 3. Approach + +# 3.1. Preliminary + +Masked Autoencoders. Masked Autoencoders (MAE) model has two components, an encoder and a decoder. Both encoder and decoder are built on Transformer layers[55]. MAE takes image tokens, i.e., non-overlapping patches from the input images, as input. These input tokens are grouped to unmasked tokens and masked tokens with a given masking ratio. The unmasked tokens will be kept for encoder to extract features and the masked tokens will be set as the learning targets of the MAE decoder that need to be reconstructed during self-supervised learning (MIM). MAE[26] adopts a high mask ratio (e.g., $75\%$ ), which prevents information leakage (e.g., simply extrapolating masked pixels based on the neighbors) in the pretraining stage. + +# 3.2. SAM-Leveraged Masked Image Pretraining + +We now adapt MAE framework to obtain efficient image encoders for segment anything model. Motivated by the high versatility of SAM [31], we explore latent features from SAM image encoder as the reconstruction target to leverage MAE. Our method emphasizes transferring the knowledge embedded in SAM. Fig. 2 (top) illustrates an overview of the proposed SAM-leveraged masked image pretraining, SAMI. The encoder transforms the unmasked tokens into latent feature representation and the decoder reconstructs the representation of the masked tokens aided by the output feature embedding from the encoder. The representation learning of reconstruction is guided by latent features from SAM. + +Cross-Attention Decoder. With the supervision of SAM features, we observe that only masked tokens need to be reconstructed via decoder while the encoder's output can serve as anchors during the reconstruction. In the cross-attention decoder, queries come from masked tokens, and keys and values derive from both unmasked features from encoder and masked features. We merge the output features of masked tokens from cross-attention decoder and the output features of unmasked tokens from encoder for the MAE output embedding. Then, this combined features will be reordered to the original positions of input image tokens for the final MAE outputs. + +Linear Projection Head. We obtain the image output through our encoder and cross-attention decoder. Then we feed such features into a small project head for aligning the features from SAM image encoder. For simplicity, we just use a linear projection head to address the feature dimension + +![](images/4bda53427016eea619c4cdbf7e4079c640260ea75673c08a5f7cb216b9e50126.jpg) +Figure 2. The overview of EfficientSAM framework. Our proposed EfficientSAM contains two stages: SAMI pretraining (top) on ImageNet and SAM finetuning (bottom) on SA-1B. For SAMI pretraining, the masked autoencoder takes the feature embeddings from SAM image encoder as the reconstruction target. After SAMI pretraining, the decoder is discarded and the light-weight encoder is served as the image encoder of EfficientSAM for finetuning on SA-1B. + +mismatch between the output of SAM image encoder and MAE. + +Reconstruction Loss. At each training iteration, SAMI consists of a feedforward feature extraction from SAM image encoder, and a feedforward and a backpropagation procedure of MAE. The outputs from SAM image encoder and MAE linear projection head are compared to compute the reconstruction loss. + +Let us denote the SAM image encoder as $f^{\mathrm{sam}}$ , and the encoder and decoder of MAE as $g^{e}$ with weights $W_{e}$ and $g^{d}$ with weights $W_{d}$ , linear projection head as $h^{\theta}$ with weights $W_{\theta}$ respectively. Assume the input tokens are denoted as $\{\mathbf{x}_i\}_{i=1}^N$ , where $N$ is the number of tokens. The input tokens are randomly grouped into the unmasked tokens, $\{\mathbf{x}_i\}_{i \in \mathcal{U}}$ , the masked tokens $\{\mathbf{x}_i\}_{i \in \mathcal{M}}$ with a given masked ratio. Let the feature reordering operator be $\phi$ , and the merging operator be $\oplus$ . + +Our target features from SAM image encoder can be written as $f^{\mathrm{sam}}(\mathbf{x}) = f^{\mathrm{sam}}(\{\mathbf{x}_i\}_{i=1}^N)$ , the output from MAE encoder is $g^{e}(\{\mathbf{x}_{i}\}_{i\in \mathcal{U}})$ , the decoder output is $g^{d}(\{\mathbf{x}_{i}\}_{i\in \mathcal{M}})$ . The output from linear projection head is $f^{h}(\mathbf{x}) = h^{\theta}(\phi(g^{e}\{\mathbf{x}_{i}\}_{i\in \mathcal{U}} \oplus g^{d}\{\mathbf{x}_{i}\}_{i\in \mathcal{M}}))$ . Therefore, our target reconstruction loss can be formulated as, + +$$ +L _ {W _ {e}, W _ {d}, W _ {\theta}} = \frac {1}{N} \cdot \sum_ {j = 1} ^ {N} \left\| f ^ {\operatorname {s a m}} (\mathbf {x}) - f ^ {h} (\mathbf {x}) \right\| ^ {2}, \tag {1} +$$ + +where $N$ is the number of input tokens, $\| \cdot \|$ denotes a norm. We use $\ell_2$ norm for reconstruction loss in our experiments. By minimizing the reconstruction loss, $L_{W_e,W_d,W_\theta}$ , + +our encoder $g^{e}$ is optimized to serve as an image backbone to extract features as SAM image encoder. Our encoder, decoder, and linear projection head are optimized to learn context modeling ability from SAM image encoder. Optimizing the reconstruction loss on all tokens transfer the knowledge embedded in SAM. + +SAMI for EfficientSAM. After pretraining, our encoder extract feature representations for various vision tasks and the decoder is discarded. In particular, to build efficient SAM models for the segmentation anything task, we take the SAMI-pretrained lightweight encoder such as ViT-Tiny and ViT-Small as the image encoder and the default mask decoder of SAM for our EfficientSAM, as illustrated in Fig. 2 (bottom). We finetune our EfficientSAM models on SA-1B dataset for the segment anything task. The overview of our EfficientSAM framework is illustrated in Fig. 2. + +# 4. Experiments + +# 4.1. Experimental Settings + +Pretraining Datasets. Our masked image pretraining method, SAMI, is conducted on ImageNet-1K training set with 1.2M images. Following masked image pretraining [26], we do not use the label information. We use the SAM ViT-H image encoders from [31] to generate reconstruction features when pretraining our ViT models, ViT-Tiny, ViT-Small, and ViT-Base. + +Pretraining Implementation Details. Our ViT models are pretrained with a mean squared error (MSE) loss for reconstruction. We use a batch size of 4096, AdamW optimizer + +[40] with learning rate $2.4\mathrm{e} - 3$ , $\beta_{1} = 0.9$ , $\beta_{2} = 0.95$ , weight decay 0.05, linear learning rate warm-up over the first 40 epochs, cosine learning rate decay to update our models. We only adopt random resize crop to $224\times 224$ resolution, random horizontal flip, and normalization for data augmentation. The mask ratio is set to $75\%$ and the decoder contains 8 Transformer blocks with 512 dimensions as in [26]. We pretrain SAMI for 400 epochs using PyTorch framework on V100 machines. For reference, 1600-epoch pretraining is required for MAE[26]. + +Downstream Tasks/Datasets/Models. Tasks and Datasets. We first consider three benchmarking datasets and several representative vision tasks to demonstrate the superiority of the proposed SAMI, including image classification on ImageNet dataset [16] with 1.2 million training and 50K validation images; Object detection and instance segmentation on COCO dataset [36] with 118K training and 5K validation images; Semantic segmentation on ADE20K dataset [72] with 20K/2K/3K images for training, validation, and testing, respectively. Then, we consider segment anything task to further show the advantages of our proposed SAMI. We finetune our pretrained lightweight image encoders for SAM on SA-1B dataset [31] with more than 1B masks from 11M high-resolution images, and test interactive instance segmentation and zero-shot instance segmentation ability of our EfficientSAMs on COCO and LVIS [23]. Models. We discard the decoder of SAMI while keeping the encoder as backbone to extract features for different tasks as in MAE[26]. We apply our well-pretrained ViT backbones for different tasks including ViTs for the classification, ViTDet [34] for the detection and instance segmentation, Mask2former [13] for the semantic segmentation task, and SAM for segment anything. + +Finetuning Settings. For the classification task, we use the AdamW optimizer with $\beta_{1} = 0.9$ , $\beta_{2} = 0.999$ , weight decay 0.05 to finetune ViTs for 100 epochs using 32 V100 GPUs, with each GPU having a batch size of 32. The initial learning rate is 1e-3 with first 5 epochs for linear warm-up and decays to zero by a cosine learning rate scheduler. We set the layer-wise learning rate decay factor to 0.75 for ViT-Small and ViT-Base. We do not apply layer-wise learning rate decay for ViT-Tiny. For data augmentation, we adopt RandAugment [15] and set label smoothing to 0.1, mixup to 0.8. For the detection and instance segmentation task, We follow the ViTDet [34] framework by adapting ViT backbones to a simple feature pyramid, for object detection and instance segmentation. We adopt AdamW optimizer with momentum $\beta_{1} = 0.9$ , $\beta_{2} = 0.999$ and weight decay 0.1 to train models on COCO. All models are trained on 64 V100 GPUs for 100 epochs with each GPU having 1 batch size with image resolution $1024 \times 1024$ . The initial learning rate is $2e - 4$ , linearly warmed up for the first 10 epochs, and decayed to 0 by a cosine learning rate schedule. Models are trained + +for 100 epochs. For the segmentation task, Our pretrained ViT models serve as the backbone of Mask2former [13], which is finetuned together with the segmentation layers on ADE20K. We adopt the AdamW optimizer with $\beta_{1} = 0.9$ , $\beta_{2} = 0.999$ , a mini-batch size of 16, a weight decay of 0.05, and an initial learning rate of 2e-4. The learning rate is decayed to 0 by a poly learning rate schedule. A learning rate multiplier is set to 0.1 for the backbone. The input image resolution is $512 \times 512$ . Models are trained for 160K iterations using 8 V100 GPUs. For the segmentation anything task, Following [31], we take our pretrained lightweight ViT models, ViT-Tiny and ViT-Small, as the image encoder of SAM framework and finetune the encoder and decoder together of our EfficientSAM on SA-1B dataset for 5 epochs. We use the AdamW optimizer with a momentum, $(\beta_{1} = 0.9$ , $\beta_{2} = 0.999)$ , a mini-batch size of 128, and a initial lreanning rate of $4e - 4$ . The learning rate is decayed to 0 by a linear learning rate schedule. We set weight decay to 0.1. We do not apply data augmentation. The input image resolution is $1024 \times 1024$ . Our EfficientSAMs are trained on 64 A100 GPUs with 40GB GPU memory. + +Baselines and Evaluation Metrics. Baselines. For the classification task, we compare the performance of ViT backbones from different pretraining/distillation methods including MAE[26], SSTA[60], DMAE[2], BEiT[3], CAE[12], DINO[6], iBOT[73], DeiT[53], etc. For the detection and instance semantic task, and semantic segmentation task, we also compare with several pretrained ViT backbones for ViTDet[34] and Mask2former[13]. For the segment every thing task, we compare with SAM[31], FastSAM[71], and MobileSAM[68]. Evaluation Metrics. We evaluate our method and all baselines in terms of accuracy. Specifically, the accuracy metrics refer to top-1 accuracy for the classification task; $\mathrm{AP^{box}}$ , $\mathrm{AP^{mask}}$ , for the detection and instance segmentation task (AP: average precision); mIoU, for the semantic segmentation task (mIoU: mean intersection over union); mIoU, AP, $\mathrm{APS}$ , $\mathrm{AP^M}$ , $\mathrm{AP^L}$ for segment anything task. For efficiency metrics, we compare the number of model parameters or inference throughput. + +# 4.2. Main Results + +Image Classification. To evaluate the effectiveness of our proposed techniques on the image classification task, we apply the proposed SAMI idea to ViT models and compare their performance over baselines on ImageNet-1K. As shown in Tab. 1, our SAMI is compared with pretraining methods like MAE, iBOT, CAE, and BEiT, and distillation methods including DeiT and SSTA. SAMI-B achieves $84.8\%$ top-1 accuracy, which outperforms the pretraining baselines, MAE, DMAE, iBOT, CAE, and BEiT by $1.2\%$ , $0.8\%$ , $1.1\%$ , $0.9\%$ , and $0.4\%$ respectively. Compared with distillation methods such as DeiT and SSTA, SAMI also shows large improvements. For lightweight models such as ViT-Tiny + +
MethodBackboneTraining DataAcc.(%)
DeiT-Ti[53]ViT-TinyIN1K74.5
SSTA-Ti[60]ViT-TinyIN1K75.2
DMAE-Ti[2]ViT-TinyIN1K70.0
MAE-Ti[26]ViT-TinyIN1K75.2
SAMI-T (ours)ViT-TinySA1B (11M) + IN1K76.8
DeiT-S[53]ViT-SmallIN1K81.2
SSTA-S[60]ViT-SmallIN1K81.4
DMAE-S[2]ViT-SmallIN1K79.3
MAE-S[26]ViT-SmallIN1K81.5
BEiT-S[3]ViT-SmallD250M+IN22K+IN1K81.7
CAE-S[12]ViT-SmallD250M+IN1K82.0
DINO-S[6]ViT-SmallIN1K82.0
iBOT-S[73]ViT-SmallIN22K+1N1K82.3
SAMI-S (ours)ViT-SmallSA1B (11M) + IN1K82.7
DeiT-B[53]ViT-BaseIN1K83.8
DMAE-B[2]ViT-BaseIN1K84.0
BootMAE[18]ViT-BaseIN1K84.2
MAE-B[26]ViT-BaseIN1K83.6
BEiT-B[3]ViT-BaseD250M+IN22K+IN1K83.7
CAE-B[12]ViT-BaseD250M+IN1K83.9
DINO-B[6]ViT-BaseIN1K82.8
iBOT-B[73]ViT-BaseIN22K+1N1K84.4
SAMI-B (ours)ViT-BaseSA1B (11M) + IN1K84.8
+ +Table 1. Image classification results on ImageNet-1K. IN is short for ImageNet. We use IN1K for pretraining/binetuning and add SA1B to the training data to indicate the need for original SAM. + +
MethodBackboneAPbboxAPmask
MAE-Ti[26]ViT-Tiny37.934.9
SAMI-Ti(ours)ViT-Tiny44.740.0
MAE-S[26]ViT-Small45.340.8
DeiT-S[53]ViT-Small47.241.9
DINO-S[6]ViT-Small49.143.3
iBOT-S[73]ViT-Small49.744.0
SAMI-S (ours)ViT-Small49.844.2
MAE-B[26]ViT-Base51.645.9
SAMI-B (ours)ViT-Base52.546.5
+ +and ViT-Small, SAMI reports a substantial gain compared to DeiT, SSTA, DMAE, and MAE. + +Object Detection and Instance Segmentation. We also extend the SAMI-pretrained ViT backbones to the downstream object detection and instance segmentation task and compare it with previous pretraining baseline on COCO dataset to evaluate its efficacy. Specifically, we take the pretrained ViT backbones and adapt them to a simple feature pyramid in the Mask R-CNN framework[25] for constructing the detector, ViTDet[34]. Tab. 2 shows the overall comparison between our SAMI and other baselines. We can see that our SAMI consistently achieves better performance over other baselines. SAMI-B obtains $0.9\mathrm{AP^{bbox}}$ and $0.6^{\mathrm{mask}}$ gains compared with MAE-B. For light-weight backbones, SAMIS and SAMI-Ti report substantial gains compared to MAE-Ti and MAE-S. Moreover, SAMI-S significantly outperforms DeiT-S by $2.6\mathrm{AP^{bbox}}$ and $2.3\mathrm{AP^{mask}}$ . For other pretraining + +Table 2. Object detection and instance segmentation results on the MS COCO using ViTDet. + +
MethodBackbonemIOU
MAE-Ti[26]ViT-Tiny39.0
SAMI-Ti(ours)ViT-Tiny42.7
MAE-S[26]ViT-Small44.1
SAMI-S (ours)ViT-Small48.8
MAE-B[26]ViT-Base49.3
SAMI-B (ours)ViT-Base51.8
+ +Table 3. Semantic segmentation results on the ADE20K dataset using Mask2former. The input resolution is $512 \times 512$ . + +
MethodCOCOLVIS
box1 click3 clickbox1 click3 click
SAM[31]78.455.674.178.959.875.2
MobileSAM[68]74.243.759.773.851.054.4
SAM-MAE-Ti[31]74.743.365.873.850.665.3
EfficientSAM-Ti (ours)75.745.567.274.352.766.8
EfficientSAM-S (ours)76.950.069.875.456.268.7
+ +Table 4. Zero-shot single point valid mask evaluation results on COCO and LVIS. Following SAM[31], we uniformly sample random points within ground truth mask for click, and compute the tightest bounding box corresponding ground truth mask for box. SAM-MAE-Ti denotes SAM with pretrained MAE-Ti image encoder. + +belines, our SAMI still compares favorably to DINO and iBOT. This set of experiments validate the effectiveness of the proposed SAMI for providing pretrained detector backbones in the object detection and instance segmentation task. Semantic Segmentation. We further extend the pretrained backbones to the semantic segmentation task to evaluate its effectiveness. Specifically, we use ViT models as the backbone in Mask2former [13] framework to benchmark on the ADE20K dataset. As shown in Tab. 3, Mask2former with SAMI-pretrained backbones achieve better mIoU, i.e., $\uparrow 2.5$ , $\uparrow 4.7$ , and $\uparrow 3.7$ improvement over backbones with MAE pretraining [26] on ImageNet-1K. This set of experiments validate that our proposed techniques could be well generalized to various downstream tasks. + +# 4.3. EfficientSAMs for Segment Anything Task + +Segment Anything task is a process of promptable segmentation to produce segmentation masks based on any form of the prompt, including point set, rough boxes or mask, freeform text. We follow SAM [31] and focus on point-based and box-based prompt segmentation on COCO/LVIS. We now test the generalization abilities of our model on segment anything task including zero-shot single point valid mask evaluation and zero-shot instance segmentation. We take the SAMI-pretrained lightweight backbones as the image encoder of SAM for building efficient SAMs, EfficientSAMs. Then we finetune EfficientSAMs on SA-1B dataset and report the performance on zero-shot single point valid mask evaluation and zero-shot instance segmentation. + +Zero-Shot Single Point Valid Mask Evaluation. Similar + +
MethodCOCOLVIS
APAPSAPMAPLAPAPSAPMAPL
ViTDet-H[34]51.032.054.368.946.635.058.066.3
SAM[31]46.530.851.061.744.732.557.665.5
MobileSAM[68]38.723.742.254.334.423.844.953.7
FastSAM[71]37.923.943.450.034.524.646.250.8
EfficientSAM-Ti (ours)42.326.746.257.439.928.951.059.9
EfficientSAM-S (ours)44.428.448.360.142.330.854.062.3
+ +Table 5. Zero-shot instance segmentation results on COCO/LVIS. ViTDet boxes are prompted to perform zero-shot segmentation. + +to SAM[31], we evaluate segmenting an object from a single foreground point. For general interactive segmentation, we also consider object segmentation from a single box, and multiple points as introduced in [31]. To achieve this, we uniformly sample random points within ground truth mask for click, and compute the tightest bounding box corresponding to ground truth mask for box. Since our models are able to predict multiple masks, we only evaluate the most confident mask as SAM [31]. + +Results. In Tab. 4, EfficientSAMs are compared with SAM, MobileSAM and SAM-MAE-Ti. On COCO, our EfficientSAM-Ti outperforms MobileSAM by $1.9\mathrm{mIoU}$ on 1 click and $1.5\mathrm{mIoU}$ on 1 box with comparable complexity. Our EfficientSAM-Ti with SAMI-pretrained weights also performs better than MAE-pretrained wights on COCO/LVIS interactive segmentation. We notice that our EfficientSAM-S only underperforms SAM by $1.5\mathrm{mIoU}$ on COCO box and $3.5\mathrm{mIoU}$ on LVIS box with $20x$ fewer parameters. We find that our EfficientSAMs also show promising performance on multiple click compared with MobileSAM and SAM-MAE-Ti. + +Zero-Shot Instance Segmentation. Following SAM [31], instance segmentation task is performed by taking the bounding box (bbox) generated by ViTDet[34] as the prompt. The mask with the highest Intersection over Union (IoU) with the_bbox as the predicted mask. + +Results. In Tab. 5, we report AP, $\mathrm{APS}^{\mathrm{s}}$ , $\mathrm{APM}^{\mathrm{s}}$ , $\mathrm{APL}^{\mathrm{s}}$ for zero-shot instance segmentation. We compare our EfficientSAM with MobileSAM and FastSAM. We can see that EfficientSAM-S obtains more than 6.5 AP on COCO and 7.8 AP on LVIS over FastSAM. For EfficientSAM-Ti, it still outperforms FastSAM by a large margin, 4.1 AP on COCO and 5.3 AP on LVIS, and MobileSAM by 3.6 AP on COCO and 5.5 AP on LVIS. Note that our EfficientSAMs are much light-weight than FastSAM, e.g., 9.8M parameters for efficientSAM-Ti vs 68M parameters for FastSAM. EfficientSAM-S also significantly reduces the gap between SAM with 0.6G parameters, only $\sim2$ AP reduction. These results demonstrate the extraordinary benefits of EfficientSAMs for zero-shot instance segmentation and validate the advantages of our SAMI pretraining method. + +Qualitative Evaluation. We now provide the qualitative results for a complementary understanding of instance segmentation capabilities of EfficientSAMs. Some examples + +Input Image SAM[31] FastSAM[71] MobileSAM[68] EfficientSAM + +![](images/e462ee682d3ff6f7ff0ab4a8426c0e77b09e3a2f99f642f7ac0e58c012e2236d.jpg) + +![](images/007be39bf465e5f70c2302f78b01e49a4bcecd1cce624b4055344345d330aaa1.jpg) + +![](images/31d5a64b3dcf348335f9e0024e68dc5406d627be89a74dd20720a0f9b2db67d5.jpg) +Figure 3. Visualization result on point-prompt input with SAM, FastSAM, MobileSAM, and our EfficientSAM model. + +can be seen in Fig. 3, Fig. 4, and Fig. 5. Specifically, we report the predicted masks with two types of prompts, point and box as in MobileSAM [68] and also segment everything results. More qualitative results can be found in the supplement. These results demonstrate that our EfficientSAMs have competing capabilities when comparing to SAM. Note that our EfficientSAMs are much lightweight than SAM, and our models can effectively give decent segmentation results. This indicates that our models can be served as a complementary version of SAM for many practical tasks. + +Input Image SAM[31] FastSAM[71] MobileSAM[68] EfficientSAM + +![](images/4ec73f2b34416afb3937c1dd76997f041fcde170cb22b0a893af2eabadbf3f4f.jpg) + +![](images/5b6559c846b5dfe10091dc90170fb0fca8c62424553ab3145d1b088eda60cb61.jpg) +Figure 4. Visualization result on box-prompt input with SAM, FastSAM, MobileSAM, and our EfficientSAM model. + +Salient Instance Segmentation. Salient object segmentation [4] aims to segment the most visually attractive objects from an image. We extend interactive instance segmentation to salient instance segmentation without manually creating points/boxes. Specifically, we take a state-of-the-art saliency object detection model, $\mathrm{U}^2$ -net[45], to predict saliency map and uniformly sample 3 random points (3 click) within + +Input Image SAM[31] FastSAM[71] MobileSAM[68] EfficientSAM + +![](images/594a71001852ba31b6dae70a744f92aae899c99728479baab48ae2c57bf791dc.jpg) + +![](images/8484e807ea89d4c8c5f90b459d6b1c59f7697f45f0eddd7e035c2c414587e8d8.jpg) + +![](images/21978634147731669b65ad893441e496ca822e6c89f29b0f881d99bb22babbb2.jpg) +Figure 5. Visualization result on segment everything with SAM, FastSAM, MobileSAM, and our EfficientSAM model. + +
MethodLossTop-1 Acc.(%)
SAMI-Ti1 - Cosine76.1
SAMI-TiMSE76.8
SAMI-S1 - Cosine82.3
SAMI-SMSE82.7
+ +Table 6. Ablation study on training loss of SAMI. MSE loss gives better classification results on ImageNet-1K. + +saliency map to perform instance segmentation with our EfficientSAM. In Fig. 6, we can see that our EfficientSAM can perform salient instance segmentation well. This preliminary exploration opens the potential to help people with hand impairments segment objects of interest in an image. + +![](images/654530326519a5f0b225e53da68b6200ff6c243bc3f83a071828b48691c27888.jpg) +Figure 6. Saliency-based automatic instance segmentation results. With the assistance of saliency map generated from $\mathbf{U}^2$ -net[45], our EfficientSAM is able to generate mask and perform automatic instance segmentation without manually creating points or boxes. + +# 4.4. Ablation Studies + +We now analyze SAMI and EfficientSAMs through a series of ablation studies with ViT backbones. + +Reconstruction Loss. We study the effect of reconstruction loss on the performance of SAMI on ImageNet-1K. We compare our mean square error (MSE) reconstruction loss with cosine similarity loss. We find that MSE reconstruction loss performs better, shown in Tab. 6. This recommends a direct reconstruction of SAM features instead of the target with a high angular similarity. + +Cross-Attention Decoder. To reconstruct SAM features, we directly use the output tokens from encoder and only take decoder to transform the masked tokens with cross-attention. We study how the performance varies with cross-attention decoder. When changing the decoder module in MAE to our cross-attention decoder in SAMI, we find that SAMI-B improves the performance from $84.4\%$ to $84.8\%$ on ImageNet-1K. The cross-attention decoder of SAMI also shows consistent improvement for ViT-Tiny and ViT-Small. Analogy to anchor points in AnchorDETR[58], the output tokens from encoder are already learned well by directly aligning the SAM features, which can serve as anchor tokens + +for help masked tokens align via cross-attention decoder. + +Mask Ratio. We explore how the performance varies with different mask ratio in SAMI. The observations are consistent with MAE [26] that a high mask ratio, $75\%$ , tends to produce good results. + +Reconstruction Target. We study the impact of reconstruction target. We take a different encoder from CLIP [46] to generate features as the reconstruction target in SAMI. Aligning features from CLIP encoder can also outperform MAE by $0.8\%$ for a ViT-Tiny model on ImageNet-1K. This demonstrates that masked image pretraining benefits from powerful guided reconstruction. + +![](images/4120f34b8c2d8137670efcaea0b014cf992c37e92c1b7fd8d61207ac381ecb46.jpg) +Figure 7. Ablation study on training steps for EfficientSAMs on MS COCO dataset. Zero-shot single point valid mask evaluation with a single box prompt is performed for the ablation. + +Effects of Finetuning Steps for EfficientSAMs. We explore the effect of fintuning steps for EfficientSAMs. As illustrated in Fig. 7, EfficientSAM-Ti and EfficientSAM-S achieve decent performance even at 0.1 epoch. For 1 epoch, the performance gain is larger than $2.5\mathrm{mIoU}$ . The final performance of EfficientSAM-S reaches $76.9\mathrm{mIoU}$ , which is only $1.5\mathrm{mIoU}$ lower than SAM. These results demonstrate the advantages of SAMI-pretrained image encoders and our EfficientSAMs. + +Efficiency. We compare the throughput (images per second) of our EfficientSAM with SAM and other models on A100 in Fig. 1. Our EfficientSAM reduces the runtime/parameters of SAM by $\sim 20\mathrm{x}$ . We also observed similar speedup ( $\sim 20\mathrm{x}$ ) comparing to SAM w.r.t. FLOPs. + +# 5. Conclusion + +We proposed a masked image pretraining approach, SAMI, to explore the potential of ViTs under the guidance of SAM foundation model. SAMI improves masked image pretraining by reconstructing the latent features from SAM image encoder to transfer knowledge from vision foundation model to ViTs. Extensive experiments on image classification, object detection and instance segmentation, semantic segmentation, and the segment anything task consistently validate SAMI's advantages. We also demonstrate that SAMI helps build efficient SAMs with pretrained light-weight encoders. Our preliminary work suggests that SAMI has potential applications beyond efficient segment anything task. + +# References + +[1] Pablo Arbelaez, Michael Maire, Charless Fowlkes, and Jitendra Malik. Contour detection and hierarchical image segmentation. IEEE transactions on pattern analysis and machine intelligence, 33(5):898-916, 2010. 1 +[2] Yutong Bai, Zeyu Wang, Junfei Xiao, Chen Wei, Huiyu Wang, Alan L Yuille, Yuyin Zhou, and Cihang Xie. Masked autoencoders enable efficient knowledge distillers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24256-24265, 2023. 3, 5, 6 +[3] Hangbo Bao, Li Dong, Songhao Piao, and Furu Wei. Beit: Bert pre-training of image transformers. arXiv preprint arXiv:2106.08254, 2021. 3, 5, 6 +[4] Ali Borji, Ming-Ming Cheng, Qibin Hou, Huaizu Jiang, and Jia Li. Salient object detection: A survey. Computational visual media, 5:117-150, 2019. 7 +[5] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In European conference on computer vision, pages 213-229. Springer, 2020. 2 +[6] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9650-9660, 2021. 3, 5, 6 +[7] Jun Cen, Yizheng Wu, Kewei Wang, Xingyi Li, Jingkang Yang, Yixuan Pei, Lingdong Kong, Ziwei Liu, and Qifeng Chen. Sad: Segment any rgbd. arXiv preprint arXiv:2305.14207, 2023. 2 +[8] Jiaqi Chen, Zeyu Yang, and Li Zhang. Semantic segment anything. https://github.com/fudan-zvg/Semantic-Segment-Anything, 2023.2 +[9] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International conference on machine learning, pages 1597-1607. PMLR, 2020. 3 +[10] Tianrun Chen, Lanyun Zhu, Chaotao Deng, Runlong Cao, Yan Wang, Shangzhan Zhang, Zejian Li, Lingyun Sun, Ying Zang, and Papa Mao. Sam-adapter: Adapting segment anything in underperformed scenes. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3367-3375, 2023. 2 +[11] Xinlei Chen and Kaiming He. Exploring simple siamese representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 15750-15758, 2021. 3 +[12] Xiaokang Chen, Mingyu Ding, Xiaodi Wang, Ying Xin, Shentong Mo, Yunhao Wang, Shumin Han, Ping Luo, Gang Zeng, and Jingdong Wang. Context autoencoder for self-supervised representation learning. International Journal of Computer Vision, pages 1-16, 2023. 5, 6 +[13] Bowen Cheng, Ishan Misra, Alexander G Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1290–1299, 2022. 5, 6 + +[14] Yangming Cheng, Lulei Li, Yuanyou Xu, Xiaodi Li, Zongxin Yang, Wenguan Wang, and Yi Yang. Segment and track anything. arXiv preprint arXiv:2305.06558, 2023. 2 +[15] Ekin D Cubuk, Barret Zoph, Jonathon Shlens, and Quoc V Le. Randaugment: Practical automated data augmentation with a reduced search space. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops, pages 702-703, 2020. 5 +[16] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 5 +[17] Ruining Deng, Can Cui, Quan Liu, Tianyuan Yao, Lucas W Remedios, Shunxing Bao, Bennett A Landman, Lee E Whelless, Lori A Coburn, Keith T Wilson, et al. Segment anything model (sam) for digital pathology: Assess zero-shot segmentation on whole slide imaging. arXiv preprint arXiv:2304.04155, 2023. 2 +[18] Xiaoyi Dong, Jianmin Bao, Ting Zhang, Dongdong Chen, Weiming Zhang, Lu Yuan, Dong Chen, Fang Wen, and Nenghai Yu. Bootstrapped masked autoencoders for vision bert pretraining. In European Conference on Computer Vision, pages 247-264. Springer, 2022. 6 +[19] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 1, 2 +[20] Haoqi Fan, Bo Xiong, Karttikeya Mangalam, Yanghao Li, Zhicheng Yan, Jitendra Malik, and Christoph Feichtenhofer. Multiscale vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6824-6835, 2021. 2 +[21] Shanghua Gao, Zhijie Lin, Xingyu Xie, Pan Zhou, Ming-Ming Cheng, and Shuicheng Yan. Editanything: Empowering unparalleled flexibility in image editing and generation. In Proceedings of the 31st ACM International Conference on Multimedia, pages 9414–9416, 2023. 2 +[22] Benjamin Graham, Alaaeldin El-Nouby, Hugo Touvron, Pierre Stock, Armand Joulin, Hervé Jégou, and Matthijs Douze. Levit: a vision transformer in convnet's clothing for faster inference. In Proceedings of the IEEE/CVF international conference on computer vision, pages 12259-12269, 2021. 3 +[23] Agrim Gupta, Piotr Dólár, and Ross Girshick. Lvis: A dataset for large vocabulary instance segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2019. 5 +[24] Dongsheng Han, Chaoning Zhang, Yu Qiao, Maryam Qamar, Yuna Jung, SeungKyu Lee, Sung-Ho Bae, and Choong Seon Hong. Segment anything model (sam) meets glass: Mirror and transparent objects cannot be easily detected. arXiv preprint arXiv:2305.00278, 2023. 1, 2 +[25] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In Proceedings of the IEEE international conference on computer vision, pages 2961-2969, 2017. 6 + +[26] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollar, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16000-16009, 2022. 2, 3, 4, 5, 6, 8 +[27] Geoffrey Hinton, Oriol Vinyals, and Jeff Dean. Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531, 2015. 3 +[28] Zejiang Hou, Fei Sun, Yen-Kuang Chen, Yuan Xie, and Sun-Yuan Kung. Milan: Masked image pretraining on language assisted representation. arXiv preprint arXiv:2208.06049, 2022.3 +[29] Jiaxi Jiang and Christian Holz. Restore anything pipeline: Segment anything meets image restoration. arXiv preprint arXiv:2305.13093, 2023. 2 +[30] Glenn Jocher, Ayush Chaurasia, and Jing Qiu. Yolo by ultralytics. https://github.com/ultralytics/ultralytics, 2023. 2 +[31] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. arXiv preprint arXiv:2304.02643, 2023. 1, 2, 3, 4, 5, 6, 7 +[32] Brett Koonce and Brett Koonce. *Mobilenetv3: Convolutional Neural Networks with Swift for Tensorflow: Image Recognition and Dataset Categorization*, pages 125–144, 2021. 2 +[33] Jiashi Li, Xin Xia, Wei Li, Huixia Li, Xing Wang, Xuefeng Xiao, Rui Wang, Min Zheng, and Xin Pan. Next-vit: Next generation vision transformer for efficient deployment in realistic industrial scenarios. arXiv preprint arXiv:2207.05501, 2022. 3 +[34] Yanghao Li, Hanzi Mao, Ross Girshick, and Kaiming He. Exploring plain vision transformer backbones for object detection. In European Conference on Computer Vision, pages 280-296. Springer, 2022. 2, 5, 6, 7 +[35] Yanyu Li, Geng Yuan, Yang Wen, Ju Hu, Georgios Evangelidis, Sergey Tulyakov, Yanzhi Wang, and Jian Ren. Efficientformer: Vision transformers at mobilenet speed. Advances in Neural Information Processing Systems, 35:12934-12949, 2022. 3 +[36] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In European conference on computer vision, pages 740-755. Springer, 2014. 5 +[37] Ruiping Liu, Jiaming Zhang, Kunyu Peng, Junwei Zheng, Ke Cao, Yufan Chen, Kailun Yang, and Rainer Stiefelhagen. Open scene understanding: Grounded situation recognition meets segment anything for helping people with visual impairments. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1857-1867, 2023. 1, 2 +[38] Xinyu Liu, Houwen Peng, Ningxin Zheng, Yuqing Yang, Han Hu, and Yixuan Yuan. Efficientvit: Memory efficient vision transformer with cascaded group attention. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14420-14430, 2023. 3 + +[39] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10012-10022, 2021. 2 +[40] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In International Conference on Learning Representations, 2018. 5 +[41] Jun Ma and Bo Wang. Segment anything in medical images. arXiv preprint arXiv:2304.12306, 2023. 1, 2 +[42] Sachin Mehta and Mohammad Rastegari. Mobilevit: Lightweight, general-purpose, and mobile-friendly vision transformer. In International Conference on Learning Representations, 2021. 2 +[43] Deepak Pathak, Philipp Krahenbuhl, Jeff Donahue, Trevor Darrell, and Alexei A Efros. Context encoders: Feature learning by inpainting. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2536-2544, 2016. 3 +[44] Zhiliang Peng, Li Dong, Hangbo Bao, Qixiang Ye, and Furu Wei. Beit v2: Masked image modeling with vector-quantized visual tokenizers. arXiv preprint arXiv:2208.06366, 2022. 2, 3 +[45] Xuebin Qin, Zichen Zhang, Chenyang Huang, Masood Dehghan, Osmar Zaiane, and Martin Jagersand. U2-net: Going deeper with nested u-structure for salient object detection. page 107404, 2020. 7, 8 +[46] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In International Conference on Machine Learning, pages 8821-8831. PMLR, 2021. 8 +[47] Adriana Romero, Nicolas Ballas, Samira Ebrahimi Kahou, Antoine Chassang, Carlo Gatta, and Yoshua Bengio. Fitnets: Hints for thin deep nets. arXiv preprint arXiv:1412.6550, 2014.3 +[48] Mark Sandler, Andrew Howard, Menglong Zhu, Andrey Zh-moginov, and Liang-Chieh Chen. Mobilenetv2: Inverted residuals and linear bottlenecks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4510-4520, 2018. 2 +[49] Qiuhong Shen, Xingyi Yang, and Xinchao Wang. Anything-3d: Towards single-view anything reconstruction in the wild. arXiv preprint arXiv:2304.10261, 2023. 2 +[50] Ao Sun, Pingchuan Ma, Yuanyuan Yuan, and Shuai Wang. Explain any concept: Segment anything meets concept-based explanation. arXiv preprint arXiv:2305.10289, 2023. 1, 2 +[51] Lv Tang, Haoke Xiao, and Bo Li. Can sam segment anything? when sam meets camouflaged object detection. arXiv preprint arXiv:2304.04709, 2023. 2 +[52] Shehbaz Tariq, Brian Estadimas Arfeto, Chaoning Zhang, and Hyundong Shin. Segment anything meets semantic communication. arXiv preprint arXiv:2306.02094, 2023. 1, 2 +[53] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Hervé Jégou. Training data-efficient image transformers & distillation through attention. In International conference on machine learning, pages 10347-10357. PMLR, 2021. 2, 5, 6 + +[54] Koen EA Van de Sande, Jasper RR Uijlings, Theo Gevers, and Arnold WM Smeulders. Segmentation as selective search for object recognition. In 2011 international conference on computer vision, pages 1879-1886. IEEE, 2011. 1 +[55] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 3 +[56] Pascal Vincent, Hugo Larochelle, Isabelle Lajoie, Yoshua Bengio, Pierre-Antoine Manzagol, and León Bottou. Stacked denoising autoencoders: Learning useful representations in a deep network with a local denoising criterion. Journal of machine learning research, 11(12), 2010. 3 +[57] Xinlong Wang, Rufeng Zhang, Chunhua Shen, Tao Kong, and Lei Li. Dense contrastive learning for self-supervised visual pre-training. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3024-3033, 2021. 3 +[58] Yingming Wang, Xiangyu Zhang, Tong Yang, and Jian Sun. Anchor detr: Query design for transformer-based detector. In Proceedings of the AAAI conference on artificial intelligence, pages 2567-2575, 2022. 8 +[59] Chen Wei, Haoqi Fan, Saining Xie, Chao-Yuan Wu, Alan Yuille, and Christoph Feichtenhofer. Masked feature prediction for self-supervised visual pre-training. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14668-14678, 2022. 3 +[60] Haiyan Wu, Yuting Gao, Yinqi Zhang, Shaohui Lin, Yuan Xie, Xing Sun, and Ke Li. Self-supervised models are good teaching assistants for vision transformers. In Proceedings of the 39th International Conference on Machine Learning, pages 24031-24042. PMLR, 2022. 3, 5, 6 +[61] Kan Wu, Jinnian Zhang, Houwen Peng, Mengchen Liu, Bin Xiao, Jianlong Fu, and Lu Yuan. Tinyvit: Fast pretraining distillation for small vision transformers. In European Conference on Computer Vision, pages 68-85. Springer, 2022. 3 +[62] Zhenda Xie, Yutong Lin, Zheng Zhang, Yue Cao, Stephen Lin, and Han Hu. Propagate yourself: Exploring pixel-level consistency for unsupervised visual representation learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16684-16693, 2021. 3 +[63] Zhenda Xie, Zheng Zhang, Yue Cao, Yutong Lin, Jianmin Bao, Zhuliang Yao, Qi Dai, and Han Hu. Simmim: A simple framework for masked image modeling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9653-9663, 2022. 3 +[64] Jing Yang, Brais Martinez, Adrian Bulat, and Georgios Tzimiropoulos. Knowledge distillation via softmax regression representation learning. In International Conference on Learning Representations, 2021. 3 +[65] Jinyu Yang, Mingqi Gao, Zhe Li, Shang Gao, Fangjing Wang, and Feng Zheng. Track anything: Segment anything meets videos. arXiv preprint arXiv:2304.11968, 2023. 2 +[66] Haoran You, Yunyang Xiong, Xiaoliang Dai, Bichen Wu, Peizhao Zhang, Haoqi Fan, Peter Vajda, and Yingyan Celine Lin. Castling-vit: Compressing self-attention via switching towards linear-angular attention at vision transformer inference. + +In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14431-14442, 2023. 3 +[67] Tao Yu, Runseng Feng, Ruoyu Feng, Jinming Liu, Xin Jin, Wenjun Zeng, and Zhibo Chen. Inpaint anything: Segment anything meets image inpainting. arXiv preprint arXiv:2304.06790, 2023. 2 +[68] Chaoning Zhang, Dongshen Han, Yu Qiao, Jung Uk Kim, Sung-Ho Bae, Seungkyu Lee, and Choong Seon Hong. Faster segment anything: Towards lightweight sam for mobile applications. arXiv preprint arXiv:2306.14289, 2023. 2, 5, 6, 7 +[69] Xiao Feng Zhang, Tian Yi Song, and Jia Wei Yao. Deshadow—anything: When segment anything model meets zero-shot shadow removal. arXiv preprint arXiv:2309.11715, 2023. 2 +[70] Borui Zhao, Quan Cui, Renjie Song, Yiyu Qiu, and Jiajun Liang. Decoupled knowledge distillation. In Proceedings of the IEEE/CVF Conference on computer vision and pattern recognition, pages 11953-11962, 2022. 3 +[71] Xu Zhao, Wenchao Ding, Yongqi An, Yinglong Du, Tao Yu, Min Li, Ming Tang, and Jinqiao Wang. Fast segment anything. arXiv preprint arXiv:2306.12156, 2023. 2, 5, 7 +[72] Bolei Zhou, Hang Zhao, Xavier Puig, Sanja Fidler, Adela Barriuso, and Antonio Torralba. Scene parsing through ade20k dataset. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 633-641, 2017. 5 +[73] Jinghao Zhou, Chen Wei, Huiyu Wang, Wei Shen, Cihang Xie, Alan Yuille, and Tao Kong. ibot: Image bert pre-training with online tokenizer. arXiv preprint arXiv:2111.07832, 2021. 5, 6 \ No newline at end of file diff --git a/2024/EfficientSAM_ Leveraged Masked Image Pretraining for Efficient Segment Anything/images.zip b/2024/EfficientSAM_ Leveraged Masked Image Pretraining for Efficient Segment Anything/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..c5dd304abe4747ea8f7eae45dfb88fed44d4afe8 --- /dev/null +++ b/2024/EfficientSAM_ Leveraged Masked Image Pretraining for Efficient Segment Anything/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9a3365ee0c31381b515d68d8a3b841554d75abdd531936f90df5604ed6e0f7e0 +size 453354 diff --git a/2024/EfficientSAM_ Leveraged Masked Image Pretraining for Efficient Segment Anything/layout.json b/2024/EfficientSAM_ Leveraged Masked Image Pretraining for Efficient Segment Anything/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..e02416a5f9b01af1ec5e8201c4fff84b951ce6a3 --- /dev/null +++ b/2024/EfficientSAM_ Leveraged Masked Image Pretraining for Efficient Segment Anything/layout.json @@ -0,0 +1,8965 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 66, + 102, + 527, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 102, + 527, + 140 + ], + "spans": [ + { + "bbox": [ + 66, + 102, + 527, + 140 + ], + "type": "text", + "content": "EfficientSAM: Leveraged Masked Image Pretraining for Efficient Segment Anything" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 58, + 161, + 534, + 202 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 161, + 534, + 202 + ], + "spans": [ + { + "bbox": [ + 58, + 161, + 534, + 202 + ], + "type": "text", + "content": "Yunyang Xiong, Bala Varadarajan*, Lemeng Wu*, Xiaoyu Xiang, Fanyi Xiao, Chenchen Zhu, Xiaoliang Dai, Dilin Wang, Fei Sun, Forrest Iandola, Raghuraman Krishnamoorthi, Vikas Chandra Meta AI Research" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 182, + 205, + 402, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 182, + 205, + 402, + 217 + ], + "spans": [ + { + "bbox": [ + 182, + 205, + 402, + 217 + ], + "type": "text", + "content": "https://yformer.github.io/efficient-sam/" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 143, + 244, + 192, + 258 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 244, + 192, + 258 + ], + "spans": [ + { + "bbox": [ + 143, + 244, + 192, + 258 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 270, + 290, + 571 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 270, + 290, + 571 + ], + "spans": [ + { + "bbox": [ + 46, + 270, + 290, + 571 + ], + "type": "text", + "content": "Segment Anything Model (SAM) has emerged as a powerful tool for numerous vision applications. A key component that drives the impressive performance for zero-shot transfer and high versatility is a super large Transformer model trained on the extensive high-quality SA-1B dataset. While beneficial, the huge computation cost of SAM model has limited its applications to wider real-world applications. To address this limitation, we propose EfficientSAMs, lightweight SAM models that exhibits decent performance with largely reduced complexity. Our idea is based on leveraging masked image pretraining, SAMI, which learns to reconstruct features from SAM image encoder for effective visual representation learning. Further, we take SAMI-pretrained light-weight image encoders and mask decoder to build EfficientSAMs, and finetune the models on SA-1B for segment anything task. We perform evaluations on multiple vision tasks including image classification, object detection, instance segmentation, and semantic segmentation, and find that our proposed pretraining method, SAMI, consistently outperforms other masked image pretraining methods. On segment anything task such as zero-shot instance segmentation, our EfficientSAMs with SAMI-pretrained lightweight image encoders perform favorably with a significant gain (e.g., " + }, + { + "bbox": [ + 46, + 270, + 290, + 571 + ], + "type": "inline_equation", + "content": "\\sim 4" + }, + { + "bbox": [ + 46, + 270, + 290, + 571 + ], + "type": "text", + "content": " AP on COCO/LVIS) over other fast SAM models. Our EfficientSAM code and models are available at here." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 591, + 128, + 604 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 591, + 128, + 604 + ], + "spans": [ + { + "bbox": [ + 47, + 591, + 128, + 604 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 611, + 288, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 611, + 288, + 696 + ], + "spans": [ + { + "bbox": [ + 46, + 611, + 288, + 696 + ], + "type": "text", + "content": "Segment Anything Model (SAM) [31] has been very successful in the vision field, achieving state-of-the-art performance in a variety of image segmentation tasks such as zero-shot edge detection[1, 31], zero-shot object proposal generation[31, 54], and zero-shot instance segmentation[31], and many other real-world applications[24, 37, 41, 50-52]. The key feature of SAM is a prompt-based Vision Trans" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 310, + 247, + 544, + 300 + ], + "blocks": [ + { + "bbox": [ + 310, + 247, + 544, + 300 + ], + "lines": [ + { + "bbox": [ + 310, + 247, + 544, + 300 + ], + "spans": [ + { + "bbox": [ + 310, + 247, + 544, + 300 + ], + "type": "image", + "image_path": "92e92a677f7a1fde0e71de11ef6055bfa06f71d67f0d8eadb98663cb69831ba3.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 312, + 307, + 539, + 430 + ], + "blocks": [ + { + "bbox": [ + 312, + 307, + 539, + 430 + ], + "lines": [ + { + "bbox": [ + 312, + 307, + 539, + 430 + ], + "spans": [ + { + "bbox": [ + 312, + 307, + 539, + 430 + ], + "type": "image", + "image_path": "7a852e0935c384ff68ee367b16efca629b1a790931f382b6c3f3626e5cbe12bb.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 439, + 547, + 572 + ], + "lines": [ + { + "bbox": [ + 304, + 439, + 547, + 572 + ], + "spans": [ + { + "bbox": [ + 304, + 439, + 547, + 572 + ], + "type": "text", + "content": "Figure 1. The comparative analysis result. (Top) The overview of EfficientSAM model by taking a well-pretrained light-weight image encoder for instance segmentation with largely reduced complexity. (Bottom) Throughput/Parameter/Performance comparison of EfficientSAM, MobileSAM, FastSAM, and SAM for zero-shot instance segmentation on COCO. We benchmark throughput (image per second) of all models on a single NVIDIA A100 with one box prompt. The input image resolution is " + }, + { + "bbox": [ + 304, + 439, + 547, + 572 + ], + "type": "inline_equation", + "content": "1024 \\times 1024" + }, + { + "bbox": [ + 304, + 439, + 547, + 572 + ], + "type": "text", + "content": ". Our EfficientSAMs outperform MobileSAM and FastSAM by a large margin, " + }, + { + "bbox": [ + 304, + 439, + 547, + 572 + ], + "type": "inline_equation", + "content": "\\sim 4" + }, + { + "bbox": [ + 304, + 439, + 547, + 572 + ], + "type": "text", + "content": " AP, with comparable complexity. Our EfficientSAM-S reduces the inference time of SAM by " + }, + { + "bbox": [ + 304, + 439, + 547, + 572 + ], + "type": "inline_equation", + "content": "\\sim 20x" + }, + { + "bbox": [ + 304, + 439, + 547, + 572 + ], + "type": "text", + "content": " and the parameter size by " + }, + { + "bbox": [ + 304, + 439, + 547, + 572 + ], + "type": "inline_equation", + "content": "\\sim 20x" + }, + { + "bbox": [ + 304, + 439, + 547, + 572 + ], + "type": "text", + "content": " with a small performance drop, 44.4 AP vs 46.5 AP." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 593, + 547, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 593, + 547, + 653 + ], + "spans": [ + { + "bbox": [ + 304, + 593, + 547, + 653 + ], + "type": "text", + "content": "former (ViT)[19]model trained on a large-scale visual dataset with more than 1B masks from 11M images, SA-1B[31], which allows segmenting any object on a given image. This ability of Segment Anything makes SAM a foundation model in vision and enables its applications even beyond vision." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 654, + 548, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 548, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 548, + 715 + ], + "type": "text", + "content": "Despite the forgoing advantages, the model of SAM turns out to be a major efficiency bottleneck for practical deployment since the architecture of SAM, especially, the image encoder (e.g., ViT-H) is very expensive. Note that ViT-H image encoder in SAM has 632M parameters while the" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 703, + 128, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 703, + 128, + 712 + ], + "spans": [ + { + "bbox": [ + 58, + 703, + 128, + 712 + ], + "type": "text", + "content": "*Joint second author." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16111" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 288, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 288, + 120 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 288, + 120 + ], + "type": "text", + "content": "prompt-based decoder only take 3.87M parameters. As a result, it leads to high computation and memory costs when using SAM to perform segment anything tasks in practice, which makes it challenging for real-time applications." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 121, + 288, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 121, + 288, + 205 + ], + "spans": [ + { + "bbox": [ + 46, + 121, + 288, + 205 + ], + "type": "text", + "content": "To address this challenge, several recent works have proposed strategies that avoid incurring the huge cost when applying SAM for the prompt-based instance segmentation. For example, [68] suggests distilling the knowledge from the default ViT-H image encoder to a tiny ViT image encoder. In [71], the computation cost can be reduced with a real-time CNN-based architecture for Segment Anything task." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 205, + 289, + 410 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 205, + 289, + 410 + ], + "spans": [ + { + "bbox": [ + 46, + 205, + 289, + 410 + ], + "type": "text", + "content": "In this paper, we propose using a well-pretrained lightweight ViT image encoder (e.g., ViT-Tiny/-Small[53]) to reduce the complexity of SAM while maintain decent performance. Our method, SAM-leveraged masked image pertraining (SAMI), produces desired pretrained lightweight ViT backbones for segment anything task. This is achieved by leveraging the celebrated MAE[26] pretraining method with SAM model to obtain high-quality pretrained ViT encoders. Specifically, our SAMI makes use of SAM encoder, ViT-H, to generate feature embedding and train a masked image model with lightweight encoders to reconstruct features from ViT-H of SAM instead of image patches. This leads to generalized ViT backbones, which can be used for downstream tasks such as image classification, object detection, and segment anything. Then we finetune the pretrained lightweight encoders with SAM decoders for segment anything task[31]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 410, + 288, + 648 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 410, + 288, + 648 + ], + "spans": [ + { + "bbox": [ + 46, + 410, + 288, + 648 + ], + "type": "text", + "content": "To evaluate our method, we consider a transfer learning setting for masked image pretraining, where models are first pretrained with a reconstructive loss on ImageNet with image resolution " + }, + { + "bbox": [ + 46, + 410, + 288, + 648 + ], + "type": "inline_equation", + "content": "224 \\times 224" + }, + { + "bbox": [ + 46, + 410, + 288, + 648 + ], + "type": "text", + "content": ", and then finetuned on target tasks using supervised data. Our SAMI learns lightweight encoders that generalize well. With SAMI pretraining, we can train models like ViT-Tiny/-Small/-Base on ImageNet-1K with improved generalization performance. For a ViT-Small model, we achieve " + }, + { + "bbox": [ + 46, + 410, + 288, + 648 + ], + "type": "inline_equation", + "content": "82.7\\%" + }, + { + "bbox": [ + 46, + 410, + 288, + 648 + ], + "type": "text", + "content": " top-1 accuracy when finetuned on ImageNet-1K with 100 epochs, which outperforms other state-of-the-art image pretraining baselines. We also finetune our pretrained models on object detection, instance segmentation, and semantic segmentation. Across all these tasks, our pretraining method achieves better results than other pretraining baselines, and more importantly, we observe significant gains for small models. Further, we evaluate our models on Segment Anything task. On zero-shot instance segmentation, our model performs well compared to recent lightweight SAM methods, including FastSAM, by a margin of 4.1 /5.2 AP on COCO/LVIS." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 59, + 650, + 276, + 661 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 650, + 276, + 661 + ], + "spans": [ + { + "bbox": [ + 59, + 650, + 276, + 661 + ], + "type": "text", + "content": "Our main contribution can be summarized as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 59, + 665, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 665, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 59, + 665, + 288, + 713 + ], + "type": "text", + "content": "- We propose a SAM-leveraged masked image pretraining framework called SAMI, which trains the model to reconstruct features from SAM ViT-H image encoder. We show that this can substantially improve the" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 326, + 72, + 533, + 84 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 72, + 533, + 84 + ], + "spans": [ + { + "bbox": [ + 326, + 72, + 533, + 84 + ], + "type": "text", + "content": "performance of image masked pretraining method." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 318, + 84, + 547, + 168 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 318, + 84, + 547, + 119 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 84, + 547, + 119 + ], + "spans": [ + { + "bbox": [ + 318, + 84, + 547, + 119 + ], + "type": "text", + "content": "- We demonstrate that SAMI-pretrained backbones can generalize well to many tasks including image classification, object detection, and semantic segmentation." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 318, + 120, + 547, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 120, + 547, + 168 + ], + "spans": [ + { + "bbox": [ + 318, + 120, + 547, + 168 + ], + "type": "text", + "content": "- We deliver EfficientSAMs, light-weight SAM models with state-of-the-art quality-efficiency trade-offs (Fig. 1), which is complementary to SAM for practical deployment." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 177, + 392, + 190 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 177, + 392, + 190 + ], + "spans": [ + { + "bbox": [ + 306, + 177, + 392, + 190 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 198, + 547, + 234 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 198, + 547, + 234 + ], + "spans": [ + { + "bbox": [ + 305, + 198, + 547, + 234 + ], + "type": "text", + "content": "We briefly review relevant works on segment anything model, vision transformers, knowledge distillation, and masked image pretraining." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 239, + 448, + 252 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 239, + 448, + 252 + ], + "spans": [ + { + "bbox": [ + 306, + 239, + 448, + 252 + ], + "type": "text", + "content": "2.1. Segment Anything Model" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 258, + 547, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 258, + 547, + 521 + ], + "spans": [ + { + "bbox": [ + 304, + 258, + 547, + 521 + ], + "type": "text", + "content": "SAM[31] has been considered as a milestone vision foundation model, which can segment any object in the image based on interaction prompts. SAM has shown remarkable zero-shot transfer performance and high versatility for many vision tasks including a variety of segmentation application[7, 8, 10, 17], in-painting[67], image restoration[29], image editing[21], image shadow removal[69], object tracking[14, 65], and 3D object reconstruction[49]. There are many other works attempting to generalize SAM to real-world scenarios, including medical image segmentation[41], camouflaged object detection[51], transparent object detection[24], concept-based explanation[50], semantic communication[52], and helping people with visual impairments[37]. Due to its wide real-world applications, practical deployment of SAM has also gained increasing attention. Several recent works including [68, 71] have proposed strategies to reduce the computation costs of SAM. FastSAM[68] develops a CNN-based architecture, YOLOv8-seg[30], to segment all objects in an image for efficiency improvement. MobileSAM[71] presents a decoupled distillation for obtaining a lightweight image encoder of SAM. Our work focuses on dealing with this efficiency issue for practical deployment of SAM." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 527, + 424, + 539 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 527, + 424, + 539 + ], + "spans": [ + { + "bbox": [ + 306, + 527, + 424, + 539 + ], + "type": "text", + "content": "2.2. Vision Transformers" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 545, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 545, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 545, + 547, + 713 + ], + "type": "text", + "content": "ViTs [19] have achieved impressive performance in vision applications[5, 20, 26, 34, 39, 44]. ViTs demonstrate the advantages of and generalization over their CNN counterparts[26]. There are also a number of works on efficient ViTs for deployment. Smaller ViTs such as ViT-Small/DeiT-Small and ViT-Tiny/DeiT-Tiny are introduced in [53] for complementing ViT-Huge, ViT-Large, and ViT-Base in [19]. Motivated by the ability of convolution to capture local information with reduced parameters and computation costs, MobileViT[42] explore combining ViT with convolutions, which outperforms light-weight CNN models such as MobileNet-v2/v3[32, 48] with better task-level generalization properties and reduced memory size and computation cost. This trick has been used in many follow-up works" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "16112" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 288, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 288, + 120 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 288, + 120 + ], + "type": "text", + "content": "including LeViT[22], EfficientFormer[35], Next-ViT[33], Tiny-ViT[61], Castling-ViT[66], EfficientViT[38]. This line of progress for designing efficient ViTs is orthogonal to our EfficientSAM work towards building efficient SAM." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 129, + 177, + 141 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 129, + 177, + 141 + ], + "spans": [ + { + "bbox": [ + 47, + 129, + 177, + 141 + ], + "type": "text", + "content": "2.3. Knowledge Distillation" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 148, + 289, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 148, + 289, + 422 + ], + "spans": [ + { + "bbox": [ + 46, + 148, + 289, + 422 + ], + "type": "text", + "content": "Knowledge distillation (KD) is a technique to improve the performance of deep learning models without changing their architectures. [27] is a pioneering work to distill the dark knowledge from a larger teacher model to a smaller student model. The learning of a student model is supervised by the hard labels and the soft labels from a teacher model. This practice is followed by multiple works which aim to make better use of soft labels to transfer more knowledge. In [64], the distillation method decouples representation learning and classification. Decoupled knowledge distillation[70] separates the classical KD loss into two parts, target class knowledge distillation and non-target class knowledge distillation, which improves the effectiveness and flexibility of knowledge transfer. Another line of work is to transfer knowledge from intermediate features. FitNet [47] is a pioneering work by distilling the semantic information from the teacher model's intermediate feature directly. In [60], a self-supervised teaching assistant (SSTA) is introduced to guide the learning of a ViT-based student model with a supervised teacher together. [2] studies the potential of knowledge distillation from pre-training MAE model by aligning the intermediate features between the larger MAE teacher model and smaller MAE student model." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 431, + 196, + 445 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 431, + 196, + 445 + ], + "spans": [ + { + "bbox": [ + 47, + 431, + 196, + 445 + ], + "type": "text", + "content": "2.4. Masked Image Pretraining" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 450, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 450, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 450, + 289, + 715 + ], + "type": "text", + "content": "Self-supervised pretraining approaches [6] have attracted significant attention in computer vision. One line of work is contrastive learning methods[9, 11, 57, 62], which learn augmentation in-variance by imposing high similarity between different augmented views of a given image. While the learned representation shows good properties such as high linear separability, contrastive learning methods relies on strong augmentation and negative sampling. Another interesting line of work is masked image modeling (MIM), which helps models learn meaningful representations by reconstructing masked image patches. The MIM pioneering works focus on using denoising autoencoders[56] and context encoders[43] to train vision Transformer with masked prediction objectives. There are various promising works on using MIM for self-supervised image pretraining. BEiT[3] is the first one that adopts MIM for ViT pretraining to predict visual tokens. In BEiTv2[44], a semantic-rich image tokenizer is utilized for a better reconstruction target. In MaskFeat[59], reconstructing the local gradient features generated from HOG descriptor leads to effective visual pretraining. In SimMIM[63] and MAE[26], directly reconstructing the pixel values of the masked image patches achieves effec" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 72, + 547, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 133 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 133 + ], + "type": "text", + "content": "tive visual representation learning. There are MAE-based follow-up works that use large teacher models to guide MAE pretraining[2, 28, 60]. Our work is built on MAE and finds that leveraging MAE to reconstruct the features from SAM image encoder enables the pretraining to be highly effective." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 144, + 373, + 158 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 144, + 373, + 158 + ], + "spans": [ + { + "bbox": [ + 306, + 144, + 373, + 158 + ], + "type": "text", + "content": "3. Approach" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 164, + 385, + 177 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 164, + 385, + 177 + ], + "spans": [ + { + "bbox": [ + 306, + 164, + 385, + 177 + ], + "type": "text", + "content": "3.1. Preliminary" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 182, + 547, + 339 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 182, + 547, + 339 + ], + "spans": [ + { + "bbox": [ + 304, + 182, + 547, + 339 + ], + "type": "text", + "content": "Masked Autoencoders. Masked Autoencoders (MAE) model has two components, an encoder and a decoder. Both encoder and decoder are built on Transformer layers[55]. MAE takes image tokens, i.e., non-overlapping patches from the input images, as input. These input tokens are grouped to unmasked tokens and masked tokens with a given masking ratio. The unmasked tokens will be kept for encoder to extract features and the masked tokens will be set as the learning targets of the MAE decoder that need to be reconstructed during self-supervised learning (MIM). MAE[26] adopts a high mask ratio (e.g., " + }, + { + "bbox": [ + 304, + 182, + 547, + 339 + ], + "type": "inline_equation", + "content": "75\\%" + }, + { + "bbox": [ + 304, + 182, + 547, + 339 + ], + "type": "text", + "content": "), which prevents information leakage (e.g., simply extrapolating masked pixels based on the neighbors) in the pretraining stage." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 347, + 534, + 360 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 347, + 534, + 360 + ], + "spans": [ + { + "bbox": [ + 305, + 347, + 534, + 360 + ], + "type": "text", + "content": "3.2. SAM-Leveraged Masked Image Pretraining" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 365, + 547, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 365, + 547, + 510 + ], + "spans": [ + { + "bbox": [ + 304, + 365, + 547, + 510 + ], + "type": "text", + "content": "We now adapt MAE framework to obtain efficient image encoders for segment anything model. Motivated by the high versatility of SAM [31], we explore latent features from SAM image encoder as the reconstruction target to leverage MAE. Our method emphasizes transferring the knowledge embedded in SAM. Fig. 2 (top) illustrates an overview of the proposed SAM-leveraged masked image pretraining, SAMI. The encoder transforms the unmasked tokens into latent feature representation and the decoder reconstructs the representation of the masked tokens aided by the output feature embedding from the encoder. The representation learning of reconstruction is guided by latent features from SAM." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 510, + 547, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 510, + 547, + 653 + ], + "spans": [ + { + "bbox": [ + 304, + 510, + 547, + 653 + ], + "type": "text", + "content": "Cross-Attention Decoder. With the supervision of SAM features, we observe that only masked tokens need to be reconstructed via decoder while the encoder's output can serve as anchors during the reconstruction. In the cross-attention decoder, queries come from masked tokens, and keys and values derive from both unmasked features from encoder and masked features. We merge the output features of masked tokens from cross-attention decoder and the output features of unmasked tokens from encoder for the MAE output embedding. Then, this combined features will be reordered to the original positions of input image tokens for the final MAE outputs." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 654, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 547, + 715 + ], + "type": "text", + "content": "Linear Projection Head. We obtain the image output through our encoder and cross-attention decoder. Then we feed such features into a small project head for aligning the features from SAM image encoder. For simplicity, we just use a linear projection head to address the feature dimension" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "16113" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 67, + 76, + 527, + 281 + ], + "blocks": [ + { + "bbox": [ + 67, + 76, + 527, + 281 + ], + "lines": [ + { + "bbox": [ + 67, + 76, + 527, + 281 + ], + "spans": [ + { + "bbox": [ + 67, + 76, + 527, + 281 + ], + "type": "image", + "image_path": "4bda53427016eea619c4cdbf7e4079c640260ea75673c08a5f7cb216b9e50126.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 297, + 547, + 342 + ], + "lines": [ + { + "bbox": [ + 46, + 297, + 547, + 342 + ], + "spans": [ + { + "bbox": [ + 46, + 297, + 547, + 342 + ], + "type": "text", + "content": "Figure 2. The overview of EfficientSAM framework. Our proposed EfficientSAM contains two stages: SAMI pretraining (top) on ImageNet and SAM finetuning (bottom) on SA-1B. For SAMI pretraining, the masked autoencoder takes the feature embeddings from SAM image encoder as the reconstruction target. After SAMI pretraining, the decoder is discarded and the light-weight encoder is served as the image encoder of EfficientSAM for finetuning on SA-1B." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 354, + 288, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 354, + 288, + 376 + ], + "spans": [ + { + "bbox": [ + 46, + 354, + 288, + 376 + ], + "type": "text", + "content": "mismatch between the output of SAM image encoder and MAE." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 378, + 288, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 378, + 288, + 449 + ], + "spans": [ + { + "bbox": [ + 46, + 378, + 288, + 449 + ], + "type": "text", + "content": "Reconstruction Loss. At each training iteration, SAMI consists of a feedforward feature extraction from SAM image encoder, and a feedforward and a backpropagation procedure of MAE. The outputs from SAM image encoder and MAE linear projection head are compared to compute the reconstruction loss." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 449, + 288, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 449, + 288, + 555 + ], + "spans": [ + { + "bbox": [ + 46, + 449, + 288, + 555 + ], + "type": "text", + "content": "Let us denote the SAM image encoder as " + }, + { + "bbox": [ + 46, + 449, + 288, + 555 + ], + "type": "inline_equation", + "content": "f^{\\mathrm{sam}}" + }, + { + "bbox": [ + 46, + 449, + 288, + 555 + ], + "type": "text", + "content": ", and the encoder and decoder of MAE as " + }, + { + "bbox": [ + 46, + 449, + 288, + 555 + ], + "type": "inline_equation", + "content": "g^{e}" + }, + { + "bbox": [ + 46, + 449, + 288, + 555 + ], + "type": "text", + "content": " with weights " + }, + { + "bbox": [ + 46, + 449, + 288, + 555 + ], + "type": "inline_equation", + "content": "W_{e}" + }, + { + "bbox": [ + 46, + 449, + 288, + 555 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 449, + 288, + 555 + ], + "type": "inline_equation", + "content": "g^{d}" + }, + { + "bbox": [ + 46, + 449, + 288, + 555 + ], + "type": "text", + "content": " with weights " + }, + { + "bbox": [ + 46, + 449, + 288, + 555 + ], + "type": "inline_equation", + "content": "W_{d}" + }, + { + "bbox": [ + 46, + 449, + 288, + 555 + ], + "type": "text", + "content": ", linear projection head as " + }, + { + "bbox": [ + 46, + 449, + 288, + 555 + ], + "type": "inline_equation", + "content": "h^{\\theta}" + }, + { + "bbox": [ + 46, + 449, + 288, + 555 + ], + "type": "text", + "content": " with weights " + }, + { + "bbox": [ + 46, + 449, + 288, + 555 + ], + "type": "inline_equation", + "content": "W_{\\theta}" + }, + { + "bbox": [ + 46, + 449, + 288, + 555 + ], + "type": "text", + "content": " respectively. Assume the input tokens are denoted as " + }, + { + "bbox": [ + 46, + 449, + 288, + 555 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{x}_i\\}_{i=1}^N" + }, + { + "bbox": [ + 46, + 449, + 288, + 555 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 449, + 288, + 555 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 449, + 288, + 555 + ], + "type": "text", + "content": " is the number of tokens. The input tokens are randomly grouped into the unmasked tokens, " + }, + { + "bbox": [ + 46, + 449, + 288, + 555 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{x}_i\\}_{i \\in \\mathcal{U}}" + }, + { + "bbox": [ + 46, + 449, + 288, + 555 + ], + "type": "text", + "content": ", the masked tokens " + }, + { + "bbox": [ + 46, + 449, + 288, + 555 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{x}_i\\}_{i \\in \\mathcal{M}}" + }, + { + "bbox": [ + 46, + 449, + 288, + 555 + ], + "type": "text", + "content": " with a given masked ratio. Let the feature reordering operator be " + }, + { + "bbox": [ + 46, + 449, + 288, + 555 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 46, + 449, + 288, + 555 + ], + "type": "text", + "content": ", and the merging operator be " + }, + { + "bbox": [ + 46, + 449, + 288, + 555 + ], + "type": "inline_equation", + "content": "\\oplus" + }, + { + "bbox": [ + 46, + 449, + 288, + 555 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 557, + 288, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 557, + 288, + 628 + ], + "spans": [ + { + "bbox": [ + 46, + 557, + 288, + 628 + ], + "type": "text", + "content": "Our target features from SAM image encoder can be written as " + }, + { + "bbox": [ + 46, + 557, + 288, + 628 + ], + "type": "inline_equation", + "content": "f^{\\mathrm{sam}}(\\mathbf{x}) = f^{\\mathrm{sam}}(\\{\\mathbf{x}_i\\}_{i=1}^N)" + }, + { + "bbox": [ + 46, + 557, + 288, + 628 + ], + "type": "text", + "content": ", the output from MAE encoder is " + }, + { + "bbox": [ + 46, + 557, + 288, + 628 + ], + "type": "inline_equation", + "content": "g^{e}(\\{\\mathbf{x}_{i}\\}_{i\\in \\mathcal{U}})" + }, + { + "bbox": [ + 46, + 557, + 288, + 628 + ], + "type": "text", + "content": ", the decoder output is " + }, + { + "bbox": [ + 46, + 557, + 288, + 628 + ], + "type": "inline_equation", + "content": "g^{d}(\\{\\mathbf{x}_{i}\\}_{i\\in \\mathcal{M}})" + }, + { + "bbox": [ + 46, + 557, + 288, + 628 + ], + "type": "text", + "content": ". The output from linear projection head is " + }, + { + "bbox": [ + 46, + 557, + 288, + 628 + ], + "type": "inline_equation", + "content": "f^{h}(\\mathbf{x}) = h^{\\theta}(\\phi(g^{e}\\{\\mathbf{x}_{i}\\}_{i\\in \\mathcal{U}} \\oplus g^{d}\\{\\mathbf{x}_{i}\\}_{i\\in \\mathcal{M}}))" + }, + { + "bbox": [ + 46, + 557, + 288, + 628 + ], + "type": "text", + "content": ". Therefore, our target reconstruction loss can be formulated as," + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 74, + 636, + 288, + 669 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 74, + 636, + 288, + 669 + ], + "spans": [ + { + "bbox": [ + 74, + 636, + 288, + 669 + ], + "type": "interline_equation", + "content": "L _ {W _ {e}, W _ {d}, W _ {\\theta}} = \\frac {1}{N} \\cdot \\sum_ {j = 1} ^ {N} \\left\\| f ^ {\\operatorname {s a m}} (\\mathbf {x}) - f ^ {h} (\\mathbf {x}) \\right\\| ^ {2}, \\tag {1}", + "image_path": "58cce236e1c7665426fb361517943cc7b8e468034b9b56f1207f16246e1d60c7.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 677, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 289, + 715 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 677, + 289, + 715 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 677, + 289, + 715 + ], + "type": "text", + "content": " is the number of input tokens, " + }, + { + "bbox": [ + 46, + 677, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\| \\cdot \\|" + }, + { + "bbox": [ + 46, + 677, + 289, + 715 + ], + "type": "text", + "content": " denotes a norm. We use " + }, + { + "bbox": [ + 46, + 677, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\ell_2" + }, + { + "bbox": [ + 46, + 677, + 289, + 715 + ], + "type": "text", + "content": " norm for reconstruction loss in our experiments. By minimizing the reconstruction loss, " + }, + { + "bbox": [ + 46, + 677, + 289, + 715 + ], + "type": "inline_equation", + "content": "L_{W_e,W_d,W_\\theta}" + }, + { + "bbox": [ + 46, + 677, + 289, + 715 + ], + "type": "text", + "content": "," + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 354, + 547, + 425 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 354, + 547, + 425 + ], + "spans": [ + { + "bbox": [ + 304, + 354, + 547, + 425 + ], + "type": "text", + "content": "our encoder " + }, + { + "bbox": [ + 304, + 354, + 547, + 425 + ], + "type": "inline_equation", + "content": "g^{e}" + }, + { + "bbox": [ + 304, + 354, + 547, + 425 + ], + "type": "text", + "content": " is optimized to serve as an image backbone to extract features as SAM image encoder. Our encoder, decoder, and linear projection head are optimized to learn context modeling ability from SAM image encoder. Optimizing the reconstruction loss on all tokens transfer the knowledge embedded in SAM." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 426, + 548, + 546 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 426, + 548, + 546 + ], + "spans": [ + { + "bbox": [ + 304, + 426, + 548, + 546 + ], + "type": "text", + "content": "SAMI for EfficientSAM. After pretraining, our encoder extract feature representations for various vision tasks and the decoder is discarded. In particular, to build efficient SAM models for the segmentation anything task, we take the SAMI-pretrained lightweight encoder such as ViT-Tiny and ViT-Small as the image encoder and the default mask decoder of SAM for our EfficientSAM, as illustrated in Fig. 2 (bottom). We finetune our EfficientSAM models on SA-1B dataset for the segment anything task. The overview of our EfficientSAM framework is illustrated in Fig. 2." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 555, + 388, + 569 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 555, + 388, + 569 + ], + "spans": [ + { + "bbox": [ + 306, + 555, + 388, + 569 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 575, + 432, + 588 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 575, + 432, + 588 + ], + "spans": [ + { + "bbox": [ + 306, + 575, + 432, + 588 + ], + "type": "text", + "content": "4.1. Experimental Settings" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 594, + 547, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 547, + 676 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 547, + 676 + ], + "type": "text", + "content": "Pretraining Datasets. Our masked image pretraining method, SAMI, is conducted on ImageNet-1K training set with 1.2M images. Following masked image pretraining [26], we do not use the label information. We use the SAM ViT-H image encoders from [31] to generate reconstruction features when pretraining our ViT models, ViT-Tiny, ViT-Small, and ViT-Base." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 677, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 547, + 713 + ], + "type": "text", + "content": "Pretraining Implementation Details. Our ViT models are pretrained with a mean squared error (MSE) loss for reconstruction. We use a batch size of 4096, AdamW optimizer" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "16114" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 288, + 192 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 288, + 192 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 288, + 192 + ], + "type": "text", + "content": "[40] with learning rate " + }, + { + "bbox": [ + 46, + 72, + 288, + 192 + ], + "type": "inline_equation", + "content": "2.4\\mathrm{e} - 3" + }, + { + "bbox": [ + 46, + 72, + 288, + 192 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 72, + 288, + 192 + ], + "type": "inline_equation", + "content": "\\beta_{1} = 0.9" + }, + { + "bbox": [ + 46, + 72, + 288, + 192 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 72, + 288, + 192 + ], + "type": "inline_equation", + "content": "\\beta_{2} = 0.95" + }, + { + "bbox": [ + 46, + 72, + 288, + 192 + ], + "type": "text", + "content": ", weight decay 0.05, linear learning rate warm-up over the first 40 epochs, cosine learning rate decay to update our models. We only adopt random resize crop to " + }, + { + "bbox": [ + 46, + 72, + 288, + 192 + ], + "type": "inline_equation", + "content": "224\\times 224" + }, + { + "bbox": [ + 46, + 72, + 288, + 192 + ], + "type": "text", + "content": " resolution, random horizontal flip, and normalization for data augmentation. The mask ratio is set to " + }, + { + "bbox": [ + 46, + 72, + 288, + 192 + ], + "type": "inline_equation", + "content": "75\\%" + }, + { + "bbox": [ + 46, + 72, + 288, + 192 + ], + "type": "text", + "content": " and the decoder contains 8 Transformer blocks with 512 dimensions as in [26]. We pretrain SAMI for 400 epochs using PyTorch framework on V100 machines. For reference, 1600-epoch pretraining is required for MAE[26]." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 195, + 289, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 195, + 289, + 472 + ], + "spans": [ + { + "bbox": [ + 46, + 195, + 289, + 472 + ], + "type": "text", + "content": "Downstream Tasks/Datasets/Models. Tasks and Datasets. We first consider three benchmarking datasets and several representative vision tasks to demonstrate the superiority of the proposed SAMI, including image classification on ImageNet dataset [16] with 1.2 million training and 50K validation images; Object detection and instance segmentation on COCO dataset [36] with 118K training and 5K validation images; Semantic segmentation on ADE20K dataset [72] with 20K/2K/3K images for training, validation, and testing, respectively. Then, we consider segment anything task to further show the advantages of our proposed SAMI. We finetune our pretrained lightweight image encoders for SAM on SA-1B dataset [31] with more than 1B masks from 11M high-resolution images, and test interactive instance segmentation and zero-shot instance segmentation ability of our EfficientSAMs on COCO and LVIS [23]. Models. We discard the decoder of SAMI while keeping the encoder as backbone to extract features for different tasks as in MAE[26]. We apply our well-pretrained ViT backbones for different tasks including ViTs for the classification, ViTDet [34] for the detection and instance segmentation, Mask2former [13] for the semantic segmentation task, and SAM for segment anything." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 474, + 289, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 474, + 289, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 474, + 289, + 714 + ], + "type": "text", + "content": "Finetuning Settings. For the classification task, we use the AdamW optimizer with " + }, + { + "bbox": [ + 46, + 474, + 289, + 714 + ], + "type": "inline_equation", + "content": "\\beta_{1} = 0.9" + }, + { + "bbox": [ + 46, + 474, + 289, + 714 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 474, + 289, + 714 + ], + "type": "inline_equation", + "content": "\\beta_{2} = 0.999" + }, + { + "bbox": [ + 46, + 474, + 289, + 714 + ], + "type": "text", + "content": ", weight decay 0.05 to finetune ViTs for 100 epochs using 32 V100 GPUs, with each GPU having a batch size of 32. The initial learning rate is 1e-3 with first 5 epochs for linear warm-up and decays to zero by a cosine learning rate scheduler. We set the layer-wise learning rate decay factor to 0.75 for ViT-Small and ViT-Base. We do not apply layer-wise learning rate decay for ViT-Tiny. For data augmentation, we adopt RandAugment [15] and set label smoothing to 0.1, mixup to 0.8. For the detection and instance segmentation task, We follow the ViTDet [34] framework by adapting ViT backbones to a simple feature pyramid, for object detection and instance segmentation. We adopt AdamW optimizer with momentum " + }, + { + "bbox": [ + 46, + 474, + 289, + 714 + ], + "type": "inline_equation", + "content": "\\beta_{1} = 0.9" + }, + { + "bbox": [ + 46, + 474, + 289, + 714 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 474, + 289, + 714 + ], + "type": "inline_equation", + "content": "\\beta_{2} = 0.999" + }, + { + "bbox": [ + 46, + 474, + 289, + 714 + ], + "type": "text", + "content": " and weight decay 0.1 to train models on COCO. All models are trained on 64 V100 GPUs for 100 epochs with each GPU having 1 batch size with image resolution " + }, + { + "bbox": [ + 46, + 474, + 289, + 714 + ], + "type": "inline_equation", + "content": "1024 \\times 1024" + }, + { + "bbox": [ + 46, + 474, + 289, + 714 + ], + "type": "text", + "content": ". The initial learning rate is " + }, + { + "bbox": [ + 46, + 474, + 289, + 714 + ], + "type": "inline_equation", + "content": "2e - 4" + }, + { + "bbox": [ + 46, + 474, + 289, + 714 + ], + "type": "text", + "content": ", linearly warmed up for the first 10 epochs, and decayed to 0 by a cosine learning rate schedule. Models are trained" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 304, + 72, + 547, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 323 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 323 + ], + "type": "text", + "content": "for 100 epochs. For the segmentation task, Our pretrained ViT models serve as the backbone of Mask2former [13], which is finetuned together with the segmentation layers on ADE20K. We adopt the AdamW optimizer with " + }, + { + "bbox": [ + 304, + 72, + 547, + 323 + ], + "type": "inline_equation", + "content": "\\beta_{1} = 0.9" + }, + { + "bbox": [ + 304, + 72, + 547, + 323 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 72, + 547, + 323 + ], + "type": "inline_equation", + "content": "\\beta_{2} = 0.999" + }, + { + "bbox": [ + 304, + 72, + 547, + 323 + ], + "type": "text", + "content": ", a mini-batch size of 16, a weight decay of 0.05, and an initial learning rate of 2e-4. The learning rate is decayed to 0 by a poly learning rate schedule. A learning rate multiplier is set to 0.1 for the backbone. The input image resolution is " + }, + { + "bbox": [ + 304, + 72, + 547, + 323 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 304, + 72, + 547, + 323 + ], + "type": "text", + "content": ". Models are trained for 160K iterations using 8 V100 GPUs. For the segmentation anything task, Following [31], we take our pretrained lightweight ViT models, ViT-Tiny and ViT-Small, as the image encoder of SAM framework and finetune the encoder and decoder together of our EfficientSAM on SA-1B dataset for 5 epochs. We use the AdamW optimizer with a momentum, " + }, + { + "bbox": [ + 304, + 72, + 547, + 323 + ], + "type": "inline_equation", + "content": "(\\beta_{1} = 0.9" + }, + { + "bbox": [ + 304, + 72, + 547, + 323 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 72, + 547, + 323 + ], + "type": "inline_equation", + "content": "\\beta_{2} = 0.999)" + }, + { + "bbox": [ + 304, + 72, + 547, + 323 + ], + "type": "text", + "content": ", a mini-batch size of 128, and a initial lreanning rate of " + }, + { + "bbox": [ + 304, + 72, + 547, + 323 + ], + "type": "inline_equation", + "content": "4e - 4" + }, + { + "bbox": [ + 304, + 72, + 547, + 323 + ], + "type": "text", + "content": ". The learning rate is decayed to 0 by a linear learning rate schedule. We set weight decay to 0.1. We do not apply data augmentation. The input image resolution is " + }, + { + "bbox": [ + 304, + 72, + 547, + 323 + ], + "type": "inline_equation", + "content": "1024 \\times 1024" + }, + { + "bbox": [ + 304, + 72, + 547, + 323 + ], + "type": "text", + "content": ". Our EfficientSAMs are trained on 64 A100 GPUs with 40GB GPU memory." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 304, + 324, + 548, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 324, + 548, + 540 + ], + "spans": [ + { + "bbox": [ + 304, + 324, + 548, + 540 + ], + "type": "text", + "content": "Baselines and Evaluation Metrics. Baselines. For the classification task, we compare the performance of ViT backbones from different pretraining/distillation methods including MAE[26], SSTA[60], DMAE[2], BEiT[3], CAE[12], DINO[6], iBOT[73], DeiT[53], etc. For the detection and instance semantic task, and semantic segmentation task, we also compare with several pretrained ViT backbones for ViTDet[34] and Mask2former[13]. For the segment every thing task, we compare with SAM[31], FastSAM[71], and MobileSAM[68]. Evaluation Metrics. We evaluate our method and all baselines in terms of accuracy. Specifically, the accuracy metrics refer to top-1 accuracy for the classification task; " + }, + { + "bbox": [ + 304, + 324, + 548, + 540 + ], + "type": "inline_equation", + "content": "\\mathrm{AP^{box}}" + }, + { + "bbox": [ + 304, + 324, + 548, + 540 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 324, + 548, + 540 + ], + "type": "inline_equation", + "content": "\\mathrm{AP^{mask}}" + }, + { + "bbox": [ + 304, + 324, + 548, + 540 + ], + "type": "text", + "content": ", for the detection and instance segmentation task (AP: average precision); mIoU, for the semantic segmentation task (mIoU: mean intersection over union); mIoU, AP, " + }, + { + "bbox": [ + 304, + 324, + 548, + 540 + ], + "type": "inline_equation", + "content": "\\mathrm{APS}" + }, + { + "bbox": [ + 304, + 324, + 548, + 540 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 324, + 548, + 540 + ], + "type": "inline_equation", + "content": "\\mathrm{AP^M}" + }, + { + "bbox": [ + 304, + 324, + 548, + 540 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 324, + 548, + 540 + ], + "type": "inline_equation", + "content": "\\mathrm{AP^L}" + }, + { + "bbox": [ + 304, + 324, + 548, + 540 + ], + "type": "text", + "content": " for segment anything task. For efficiency metrics, we compare the number of model parameters or inference throughput." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 306, + 551, + 391, + 562 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 551, + 391, + 562 + ], + "spans": [ + { + "bbox": [ + 306, + 551, + 391, + 562 + ], + "type": "text", + "content": "4.2. Main Results" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 570, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 547, + 715 + ], + "type": "text", + "content": "Image Classification. To evaluate the effectiveness of our proposed techniques on the image classification task, we apply the proposed SAMI idea to ViT models and compare their performance over baselines on ImageNet-1K. As shown in Tab. 1, our SAMI is compared with pretraining methods like MAE, iBOT, CAE, and BEiT, and distillation methods including DeiT and SSTA. SAMI-B achieves " + }, + { + "bbox": [ + 304, + 570, + 547, + 715 + ], + "type": "inline_equation", + "content": "84.8\\%" + }, + { + "bbox": [ + 304, + 570, + 547, + 715 + ], + "type": "text", + "content": " top-1 accuracy, which outperforms the pretraining baselines, MAE, DMAE, iBOT, CAE, and BEiT by " + }, + { + "bbox": [ + 304, + 570, + 547, + 715 + ], + "type": "inline_equation", + "content": "1.2\\%" + }, + { + "bbox": [ + 304, + 570, + 547, + 715 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 570, + 547, + 715 + ], + "type": "inline_equation", + "content": "0.8\\%" + }, + { + "bbox": [ + 304, + 570, + 547, + 715 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 570, + 547, + 715 + ], + "type": "inline_equation", + "content": "1.1\\%" + }, + { + "bbox": [ + 304, + 570, + 547, + 715 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 570, + 547, + 715 + ], + "type": "inline_equation", + "content": "0.9\\%" + }, + { + "bbox": [ + 304, + 570, + 547, + 715 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 570, + 547, + 715 + ], + "type": "inline_equation", + "content": "0.4\\%" + }, + { + "bbox": [ + 304, + 570, + 547, + 715 + ], + "type": "text", + "content": " respectively. Compared with distillation methods such as DeiT and SSTA, SAMI also shows large improvements. For lightweight models such as ViT-Tiny" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "16115" + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 55, + 69, + 282, + 301 + ], + "blocks": [ + { + "bbox": [ + 55, + 69, + 282, + 301 + ], + "lines": [ + { + "bbox": [ + 55, + 69, + 282, + 301 + ], + "spans": [ + { + "bbox": [ + 55, + 69, + 282, + 301 + ], + "type": "table", + "html": "
MethodBackboneTraining DataAcc.(%)
DeiT-Ti[53]ViT-TinyIN1K74.5
SSTA-Ti[60]ViT-TinyIN1K75.2
DMAE-Ti[2]ViT-TinyIN1K70.0
MAE-Ti[26]ViT-TinyIN1K75.2
SAMI-T (ours)ViT-TinySA1B (11M) + IN1K76.8
DeiT-S[53]ViT-SmallIN1K81.2
SSTA-S[60]ViT-SmallIN1K81.4
DMAE-S[2]ViT-SmallIN1K79.3
MAE-S[26]ViT-SmallIN1K81.5
BEiT-S[3]ViT-SmallD250M+IN22K+IN1K81.7
CAE-S[12]ViT-SmallD250M+IN1K82.0
DINO-S[6]ViT-SmallIN1K82.0
iBOT-S[73]ViT-SmallIN22K+1N1K82.3
SAMI-S (ours)ViT-SmallSA1B (11M) + IN1K82.7
DeiT-B[53]ViT-BaseIN1K83.8
DMAE-B[2]ViT-BaseIN1K84.0
BootMAE[18]ViT-BaseIN1K84.2
MAE-B[26]ViT-BaseIN1K83.6
BEiT-B[3]ViT-BaseD250M+IN22K+IN1K83.7
CAE-B[12]ViT-BaseD250M+IN1K83.9
DINO-B[6]ViT-BaseIN1K82.8
iBOT-B[73]ViT-BaseIN22K+1N1K84.4
SAMI-B (ours)ViT-BaseSA1B (11M) + IN1K84.8
", + "image_path": "b278e3fd436ef17057d30228412536d513176c71d13b24b407a21f57c59e5a35.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 80, + 353, + 258, + 458 + ], + "blocks": [ + { + "bbox": [ + 46, + 308, + 288, + 343 + ], + "lines": [ + { + "bbox": [ + 46, + 308, + 288, + 343 + ], + "spans": [ + { + "bbox": [ + 46, + 308, + 288, + 343 + ], + "type": "text", + "content": "Table 1. Image classification results on ImageNet-1K. IN is short for ImageNet. We use IN1K for pretraining/binetuning and add SA1B to the training data to indicate the need for original SAM." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 80, + 353, + 258, + 458 + ], + "lines": [ + { + "bbox": [ + 80, + 353, + 258, + 458 + ], + "spans": [ + { + "bbox": [ + 80, + 353, + 258, + 458 + ], + "type": "table", + "html": "
MethodBackboneAPbboxAPmask
MAE-Ti[26]ViT-Tiny37.934.9
SAMI-Ti(ours)ViT-Tiny44.740.0
MAE-S[26]ViT-Small45.340.8
DeiT-S[53]ViT-Small47.241.9
DINO-S[6]ViT-Small49.143.3
iBOT-S[73]ViT-Small49.744.0
SAMI-S (ours)ViT-Small49.844.2
MAE-B[26]ViT-Base51.645.9
SAMI-B (ours)ViT-Base52.546.5
", + "image_path": "2ac814eeab0049681b41d721b3d5c32dbdfec20c89b0e04077d5ae533a3c0549.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 510, + 287, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 510, + 287, + 533 + ], + "spans": [ + { + "bbox": [ + 46, + 510, + 287, + 533 + ], + "type": "text", + "content": "and ViT-Small, SAMI reports a substantial gain compared to DeiT, SSTA, DMAE, and MAE." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 534, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 534, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 534, + 289, + 715 + ], + "type": "text", + "content": "Object Detection and Instance Segmentation. We also extend the SAMI-pretrained ViT backbones to the downstream object detection and instance segmentation task and compare it with previous pretraining baseline on COCO dataset to evaluate its efficacy. Specifically, we take the pretrained ViT backbones and adapt them to a simple feature pyramid in the Mask R-CNN framework[25] for constructing the detector, ViTDet[34]. Tab. 2 shows the overall comparison between our SAMI and other baselines. We can see that our SAMI consistently achieves better performance over other baselines. SAMI-B obtains " + }, + { + "bbox": [ + 46, + 534, + 289, + 715 + ], + "type": "inline_equation", + "content": "0.9\\mathrm{AP^{bbox}}" + }, + { + "bbox": [ + 46, + 534, + 289, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 534, + 289, + 715 + ], + "type": "inline_equation", + "content": "0.6^{\\mathrm{mask}}" + }, + { + "bbox": [ + 46, + 534, + 289, + 715 + ], + "type": "text", + "content": " gains compared with MAE-B. For light-weight backbones, SAMIS and SAMI-Ti report substantial gains compared to MAE-Ti and MAE-S. Moreover, SAMI-S significantly outperforms DeiT-S by " + }, + { + "bbox": [ + 46, + 534, + 289, + 715 + ], + "type": "inline_equation", + "content": "2.6\\mathrm{AP^{bbox}}" + }, + { + "bbox": [ + 46, + 534, + 289, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 534, + 289, + 715 + ], + "type": "inline_equation", + "content": "2.3\\mathrm{AP^{mask}}" + }, + { + "bbox": [ + 46, + 534, + 289, + 715 + ], + "type": "text", + "content": ". For other pretraining" + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 350, + 69, + 506, + 150 + ], + "blocks": [ + { + "bbox": [ + 46, + 466, + 287, + 489 + ], + "lines": [ + { + "bbox": [ + 46, + 466, + 287, + 489 + ], + "spans": [ + { + "bbox": [ + 46, + 466, + 287, + 489 + ], + "type": "text", + "content": "Table 2. Object detection and instance segmentation results on the MS COCO using ViTDet." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 350, + 69, + 506, + 150 + ], + "lines": [ + { + "bbox": [ + 350, + 69, + 506, + 150 + ], + "spans": [ + { + "bbox": [ + 350, + 69, + 506, + 150 + ], + "type": "table", + "html": "
MethodBackbonemIOU
MAE-Ti[26]ViT-Tiny39.0
SAMI-Ti(ours)ViT-Tiny42.7
MAE-S[26]ViT-Small44.1
SAMI-S (ours)ViT-Small48.8
MAE-B[26]ViT-Base49.3
SAMI-B (ours)ViT-Base51.8
", + "image_path": "bbb1c79e0dbd56b1db88a61f37a3b6b387950d222b24916ab630be64bdb9ab86.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 314, + 191, + 541, + 255 + ], + "blocks": [ + { + "bbox": [ + 305, + 158, + 547, + 180 + ], + "lines": [ + { + "bbox": [ + 305, + 158, + 547, + 180 + ], + "spans": [ + { + "bbox": [ + 305, + 158, + 547, + 180 + ], + "type": "text", + "content": "Table 3. Semantic segmentation results on the ADE20K dataset using Mask2former. The input resolution is " + }, + { + "bbox": [ + 305, + 158, + 547, + 180 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 305, + 158, + 547, + 180 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 314, + 191, + 541, + 255 + ], + "lines": [ + { + "bbox": [ + 314, + 191, + 541, + 255 + ], + "spans": [ + { + "bbox": [ + 314, + 191, + 541, + 255 + ], + "type": "table", + "html": "
MethodCOCOLVIS
box1 click3 clickbox1 click3 click
SAM[31]78.455.674.178.959.875.2
MobileSAM[68]74.243.759.773.851.054.4
SAM-MAE-Ti[31]74.743.365.873.850.665.3
EfficientSAM-Ti (ours)75.745.567.274.352.766.8
EfficientSAM-S (ours)76.950.069.875.456.268.7
", + "image_path": "e9f56425f970434b3d5c3c9c71d70cd00c81a7fd2659a67b246b6942bc81a33a.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 262, + 548, + 327 + ], + "lines": [ + { + "bbox": [ + 305, + 262, + 548, + 327 + ], + "spans": [ + { + "bbox": [ + 305, + 262, + 548, + 327 + ], + "type": "text", + "content": "Table 4. Zero-shot single point valid mask evaluation results on COCO and LVIS. Following SAM[31], we uniformly sample random points within ground truth mask for click, and compute the tightest bounding box corresponding ground truth mask for box. SAM-MAE-Ti denotes SAM with pretrained MAE-Ti image encoder." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 350, + 548, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 350, + 548, + 518 + ], + "spans": [ + { + "bbox": [ + 304, + 350, + 548, + 518 + ], + "type": "text", + "content": "belines, our SAMI still compares favorably to DINO and iBOT. This set of experiments validate the effectiveness of the proposed SAMI for providing pretrained detector backbones in the object detection and instance segmentation task. Semantic Segmentation. We further extend the pretrained backbones to the semantic segmentation task to evaluate its effectiveness. Specifically, we use ViT models as the backbone in Mask2former [13] framework to benchmark on the ADE20K dataset. As shown in Tab. 3, Mask2former with SAMI-pretrained backbones achieve better mIoU, i.e., " + }, + { + "bbox": [ + 304, + 350, + 548, + 518 + ], + "type": "inline_equation", + "content": "\\uparrow 2.5" + }, + { + "bbox": [ + 304, + 350, + 548, + 518 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 350, + 548, + 518 + ], + "type": "inline_equation", + "content": "\\uparrow 4.7" + }, + { + "bbox": [ + 304, + 350, + 548, + 518 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 350, + 548, + 518 + ], + "type": "inline_equation", + "content": "\\uparrow 3.7" + }, + { + "bbox": [ + 304, + 350, + 548, + 518 + ], + "type": "text", + "content": " improvement over backbones with MAE pretraining [26] on ImageNet-1K. This set of experiments validate that our proposed techniques could be well generalized to various downstream tasks." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 526, + 528, + 540 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 526, + 528, + 540 + ], + "spans": [ + { + "bbox": [ + 305, + 526, + 528, + 540 + ], + "type": "text", + "content": "4.3. EfficientSAMs for Segment Anything Task" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 545, + 548, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 545, + 548, + 700 + ], + "spans": [ + { + "bbox": [ + 304, + 545, + 548, + 700 + ], + "type": "text", + "content": "Segment Anything task is a process of promptable segmentation to produce segmentation masks based on any form of the prompt, including point set, rough boxes or mask, freeform text. We follow SAM [31] and focus on point-based and box-based prompt segmentation on COCO/LVIS. We now test the generalization abilities of our model on segment anything task including zero-shot single point valid mask evaluation and zero-shot instance segmentation. We take the SAMI-pretrained lightweight backbones as the image encoder of SAM for building efficient SAMs, EfficientSAMs. Then we finetune EfficientSAMs on SA-1B dataset and report the performance on zero-shot single point valid mask evaluation and zero-shot instance segmentation." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 317, + 701, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 701, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 317, + 701, + 547, + 714 + ], + "type": "text", + "content": "Zero-Shot Single Point Valid Mask Evaluation. Similar" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "16116" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 55, + 70, + 281, + 135 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 281, + 135 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 281, + 135 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 281, + 135 + ], + "type": "table", + "html": "
MethodCOCOLVIS
APAPSAPMAPLAPAPSAPMAPL
ViTDet-H[34]51.032.054.368.946.635.058.066.3
SAM[31]46.530.851.061.744.732.557.665.5
MobileSAM[68]38.723.742.254.334.423.844.953.7
FastSAM[71]37.923.943.450.034.524.646.250.8
EfficientSAM-Ti (ours)42.326.746.257.439.928.951.059.9
EfficientSAM-S (ours)44.428.448.360.142.330.854.062.3
", + "image_path": "4cdaf4d332667f6c77f6242065a20433bfbcec5e807a9bdaa0f254e2ce7c1d4b.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 142, + 288, + 164 + ], + "lines": [ + { + "bbox": [ + 47, + 142, + 288, + 164 + ], + "spans": [ + { + "bbox": [ + 47, + 142, + 288, + 164 + ], + "type": "text", + "content": "Table 5. Zero-shot instance segmentation results on COCO/LVIS. ViTDet boxes are prompted to perform zero-shot segmentation." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 186, + 288, + 293 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 186, + 288, + 293 + ], + "spans": [ + { + "bbox": [ + 46, + 186, + 288, + 293 + ], + "type": "text", + "content": "to SAM[31], we evaluate segmenting an object from a single foreground point. For general interactive segmentation, we also consider object segmentation from a single box, and multiple points as introduced in [31]. To achieve this, we uniformly sample random points within ground truth mask for click, and compute the tightest bounding box corresponding to ground truth mask for box. Since our models are able to predict multiple masks, we only evaluate the most confident mask as SAM [31]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 293, + 288, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 293, + 288, + 437 + ], + "spans": [ + { + "bbox": [ + 46, + 293, + 288, + 437 + ], + "type": "text", + "content": "Results. In Tab. 4, EfficientSAMs are compared with SAM, MobileSAM and SAM-MAE-Ti. On COCO, our EfficientSAM-Ti outperforms MobileSAM by " + }, + { + "bbox": [ + 46, + 293, + 288, + 437 + ], + "type": "inline_equation", + "content": "1.9\\mathrm{mIoU}" + }, + { + "bbox": [ + 46, + 293, + 288, + 437 + ], + "type": "text", + "content": " on 1 click and " + }, + { + "bbox": [ + 46, + 293, + 288, + 437 + ], + "type": "inline_equation", + "content": "1.5\\mathrm{mIoU}" + }, + { + "bbox": [ + 46, + 293, + 288, + 437 + ], + "type": "text", + "content": " on 1 box with comparable complexity. Our EfficientSAM-Ti with SAMI-pretrained weights also performs better than MAE-pretrained wights on COCO/LVIS interactive segmentation. We notice that our EfficientSAM-S only underperforms SAM by " + }, + { + "bbox": [ + 46, + 293, + 288, + 437 + ], + "type": "inline_equation", + "content": "1.5\\mathrm{mIoU}" + }, + { + "bbox": [ + 46, + 293, + 288, + 437 + ], + "type": "text", + "content": " on COCO box and " + }, + { + "bbox": [ + 46, + 293, + 288, + 437 + ], + "type": "inline_equation", + "content": "3.5\\mathrm{mIoU}" + }, + { + "bbox": [ + 46, + 293, + 288, + 437 + ], + "type": "text", + "content": " on LVIS box with " + }, + { + "bbox": [ + 46, + 293, + 288, + 437 + ], + "type": "inline_equation", + "content": "20x" + }, + { + "bbox": [ + 46, + 293, + 288, + 437 + ], + "type": "text", + "content": " fewer parameters. We find that our EfficientSAMs also show promising performance on multiple click compared with MobileSAM and SAM-MAE-Ti." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 437, + 287, + 497 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 437, + 287, + 497 + ], + "spans": [ + { + "bbox": [ + 47, + 437, + 287, + 497 + ], + "type": "text", + "content": "Zero-Shot Instance Segmentation. Following SAM [31], instance segmentation task is performed by taking the bounding box (bbox) generated by ViTDet[34] as the prompt. The mask with the highest Intersection over Union (IoU) with the_bbox as the predicted mask." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 498, + 288, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 498, + 288, + 677 + ], + "spans": [ + { + "bbox": [ + 46, + 498, + 288, + 677 + ], + "type": "text", + "content": "Results. In Tab. 5, we report AP, " + }, + { + "bbox": [ + 46, + 498, + 288, + 677 + ], + "type": "inline_equation", + "content": "\\mathrm{APS}^{\\mathrm{s}}" + }, + { + "bbox": [ + 46, + 498, + 288, + 677 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 498, + 288, + 677 + ], + "type": "inline_equation", + "content": "\\mathrm{APM}^{\\mathrm{s}}" + }, + { + "bbox": [ + 46, + 498, + 288, + 677 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 498, + 288, + 677 + ], + "type": "inline_equation", + "content": "\\mathrm{APL}^{\\mathrm{s}}" + }, + { + "bbox": [ + 46, + 498, + 288, + 677 + ], + "type": "text", + "content": " for zero-shot instance segmentation. We compare our EfficientSAM with MobileSAM and FastSAM. We can see that EfficientSAM-S obtains more than 6.5 AP on COCO and 7.8 AP on LVIS over FastSAM. For EfficientSAM-Ti, it still outperforms FastSAM by a large margin, 4.1 AP on COCO and 5.3 AP on LVIS, and MobileSAM by 3.6 AP on COCO and 5.5 AP on LVIS. Note that our EfficientSAMs are much light-weight than FastSAM, e.g., 9.8M parameters for efficientSAM-Ti vs 68M parameters for FastSAM. EfficientSAM-S also significantly reduces the gap between SAM with 0.6G parameters, only " + }, + { + "bbox": [ + 46, + 498, + 288, + 677 + ], + "type": "inline_equation", + "content": "\\sim2" + }, + { + "bbox": [ + 46, + 498, + 288, + 677 + ], + "type": "text", + "content": " AP reduction. These results demonstrate the extraordinary benefits of EfficientSAMs for zero-shot instance segmentation and validate the advantages of our SAMI pretraining method." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 677, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 677, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 677, + 288, + 713 + ], + "type": "text", + "content": "Qualitative Evaluation. We now provide the qualitative results for a complementary understanding of instance segmentation capabilities of EfficientSAMs. Some examples" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 311, + 60, + 542, + 69 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 60, + 542, + 69 + ], + "spans": [ + { + "bbox": [ + 311, + 60, + 542, + 69 + ], + "type": "text", + "content": "Input Image SAM[31] FastSAM[71] MobileSAM[68] EfficientSAM" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 308, + 70, + 544, + 103 + ], + "blocks": [ + { + "bbox": [ + 308, + 70, + 544, + 103 + ], + "lines": [ + { + "bbox": [ + 308, + 70, + 544, + 103 + ], + "spans": [ + { + "bbox": [ + 308, + 70, + 544, + 103 + ], + "type": "image", + "image_path": "e462ee682d3ff6f7ff0ab4a8426c0e77b09e3a2f99f642f7ac0e58c012e2236d.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 308, + 108, + 544, + 137 + ], + "blocks": [ + { + "bbox": [ + 308, + 108, + 544, + 137 + ], + "lines": [ + { + "bbox": [ + 308, + 108, + 544, + 137 + ], + "spans": [ + { + "bbox": [ + 308, + 108, + 544, + 137 + ], + "type": "image", + "image_path": "007be39bf465e5f70c2302f78b01e49a4bcecd1cce624b4055344345d330aaa1.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 308, + 141, + 544, + 175 + ], + "blocks": [ + { + "bbox": [ + 308, + 141, + 544, + 175 + ], + "lines": [ + { + "bbox": [ + 308, + 141, + 544, + 175 + ], + "spans": [ + { + "bbox": [ + 308, + 141, + 544, + 175 + ], + "type": "image", + "image_path": "31d5a64b3dcf348335f9e0024e68dc5406d627be89a74dd20720a0f9b2db67d5.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 306, + 184, + 547, + 206 + ], + "lines": [ + { + "bbox": [ + 306, + 184, + 547, + 206 + ], + "spans": [ + { + "bbox": [ + 306, + 184, + 547, + 206 + ], + "type": "text", + "content": "Figure 3. Visualization result on point-prompt input with SAM, FastSAM, MobileSAM, and our EfficientSAM model." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 227, + 546, + 347 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 227, + 546, + 347 + ], + "spans": [ + { + "bbox": [ + 304, + 227, + 546, + 347 + ], + "type": "text", + "content": "can be seen in Fig. 3, Fig. 4, and Fig. 5. Specifically, we report the predicted masks with two types of prompts, point and box as in MobileSAM [68] and also segment everything results. More qualitative results can be found in the supplement. These results demonstrate that our EfficientSAMs have competing capabilities when comparing to SAM. Note that our EfficientSAMs are much lightweight than SAM, and our models can effectively give decent segmentation results. This indicates that our models can be served as a complementary version of SAM for many practical tasks." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 312, + 351, + 542, + 360 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 351, + 542, + 360 + ], + "spans": [ + { + "bbox": [ + 312, + 351, + 542, + 360 + ], + "type": "text", + "content": "Input Image SAM[31] FastSAM[71] MobileSAM[68] EfficientSAM" + } + ] + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 307, + 361, + 544, + 389 + ], + "blocks": [ + { + "bbox": [ + 307, + 361, + 544, + 389 + ], + "lines": [ + { + "bbox": [ + 307, + 361, + 544, + 389 + ], + "spans": [ + { + "bbox": [ + 307, + 361, + 544, + 389 + ], + "type": "image", + "image_path": "4ec73f2b34416afb3937c1dd76997f041fcde170cb22b0a893af2eabadbf3f4f.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 308, + 394, + 544, + 428 + ], + "blocks": [ + { + "bbox": [ + 308, + 394, + 544, + 428 + ], + "lines": [ + { + "bbox": [ + 308, + 394, + 544, + 428 + ], + "spans": [ + { + "bbox": [ + 308, + 394, + 544, + 428 + ], + "type": "image", + "image_path": "5b6559c846b5dfe10091dc90170fb0fca8c62424553ab3145d1b088eda60cb61.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 437, + 546, + 460 + ], + "lines": [ + { + "bbox": [ + 305, + 437, + 546, + 460 + ], + "spans": [ + { + "bbox": [ + 305, + 437, + 546, + 460 + ], + "type": "text", + "content": "Figure 4. Visualization result on box-prompt input with SAM, FastSAM, MobileSAM, and our EfficientSAM model." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 472, + 546, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 472, + 546, + 556 + ], + "spans": [ + { + "bbox": [ + 304, + 472, + 546, + 556 + ], + "type": "text", + "content": "Salient Instance Segmentation. Salient object segmentation [4] aims to segment the most visually attractive objects from an image. We extend interactive instance segmentation to salient instance segmentation without manually creating points/boxes. Specifically, we take a state-of-the-art saliency object detection model, " + }, + { + "bbox": [ + 304, + 472, + 546, + 556 + ], + "type": "inline_equation", + "content": "\\mathrm{U}^2" + }, + { + "bbox": [ + 304, + 472, + 546, + 556 + ], + "type": "text", + "content": "-net[45], to predict saliency map and uniformly sample 3 random points (3 click) within" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 312, + 563, + 542, + 571 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 563, + 542, + 571 + ], + "spans": [ + { + "bbox": [ + 312, + 563, + 542, + 571 + ], + "type": "text", + "content": "Input Image SAM[31] FastSAM[71] MobileSAM[68] EfficientSAM" + } + ] + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 308, + 573, + 542, + 606 + ], + "blocks": [ + { + "bbox": [ + 308, + 573, + 542, + 606 + ], + "lines": [ + { + "bbox": [ + 308, + 573, + 542, + 606 + ], + "spans": [ + { + "bbox": [ + 308, + 573, + 542, + 606 + ], + "type": "image", + "image_path": "594a71001852ba31b6dae70a744f92aae899c99728479baab48ae2c57bf791dc.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 307, + 612, + 542, + 642 + ], + "blocks": [ + { + "bbox": [ + 307, + 612, + 542, + 642 + ], + "lines": [ + { + "bbox": [ + 307, + 612, + 542, + 642 + ], + "spans": [ + { + "bbox": [ + 307, + 612, + 542, + 642 + ], + "type": "image", + "image_path": "8484e807ea89d4c8c5f90b459d6b1c59f7697f45f0eddd7e035c2c414587e8d8.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 307, + 647, + 542, + 677 + ], + "blocks": [ + { + "bbox": [ + 307, + 647, + 542, + 677 + ], + "lines": [ + { + "bbox": [ + 307, + 647, + 542, + 677 + ], + "spans": [ + { + "bbox": [ + 307, + 647, + 542, + 677 + ], + "type": "image", + "image_path": "21978634147731669b65ad893441e496ca822e6c89f29b0f881d99bb22babbb2.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 688, + 546, + 709 + ], + "lines": [ + { + "bbox": [ + 305, + 688, + 546, + 709 + ], + "spans": [ + { + "bbox": [ + 305, + 688, + 546, + 709 + ], + "type": "text", + "content": "Figure 5. Visualization result on segment everything with SAM, FastSAM, MobileSAM, and our EfficientSAM model." + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16117" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 97, + 70, + 240, + 122 + ], + "blocks": [ + { + "bbox": [ + 97, + 70, + 240, + 122 + ], + "lines": [ + { + "bbox": [ + 97, + 70, + 240, + 122 + ], + "spans": [ + { + "bbox": [ + 97, + 70, + 240, + 122 + ], + "type": "table", + "html": "
MethodLossTop-1 Acc.(%)
SAMI-Ti1 - Cosine76.1
SAMI-TiMSE76.8
SAMI-S1 - Cosine82.3
SAMI-SMSE82.7
", + "image_path": "c5f97f67a5834e353278e6a719ed8faa64ae1851e79a6a43dbe4295fb9ef923c.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 130, + 287, + 152 + ], + "lines": [ + { + "bbox": [ + 47, + 130, + 287, + 152 + ], + "spans": [ + { + "bbox": [ + 47, + 130, + 287, + 152 + ], + "type": "text", + "content": "Table 6. Ablation study on training loss of SAMI. MSE loss gives better classification results on ImageNet-1K." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 47, + 174, + 288, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 174, + 288, + 236 + ], + "spans": [ + { + "bbox": [ + 47, + 174, + 288, + 236 + ], + "type": "text", + "content": "saliency map to perform instance segmentation with our EfficientSAM. In Fig. 6, we can see that our EfficientSAM can perform salient instance segmentation well. This preliminary exploration opens the potential to help people with hand impairments segment objects of interest in an image." + } + ] + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 69, + 240, + 262, + 369 + ], + "blocks": [ + { + "bbox": [ + 69, + 240, + 262, + 369 + ], + "lines": [ + { + "bbox": [ + 69, + 240, + 262, + 369 + ], + "spans": [ + { + "bbox": [ + 69, + 240, + 262, + 369 + ], + "type": "image", + "image_path": "654530326519a5f0b225e53da68b6200ff6c243bc3f83a071828b48691c27888.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 376, + 288, + 422 + ], + "lines": [ + { + "bbox": [ + 46, + 376, + 288, + 422 + ], + "spans": [ + { + "bbox": [ + 46, + 376, + 288, + 422 + ], + "type": "text", + "content": "Figure 6. Saliency-based automatic instance segmentation results. With the assistance of saliency map generated from " + }, + { + "bbox": [ + 46, + 376, + 288, + 422 + ], + "type": "inline_equation", + "content": "\\mathbf{U}^2" + }, + { + "bbox": [ + 46, + 376, + 288, + 422 + ], + "type": "text", + "content": "-net[45], our EfficientSAM is able to generate mask and perform automatic instance segmentation without manually creating points or boxes." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 443, + 147, + 453 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 443, + 147, + 453 + ], + "spans": [ + { + "bbox": [ + 47, + 443, + 147, + 453 + ], + "type": "text", + "content": "4.4. Ablation Studies" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 461, + 287, + 485 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 461, + 287, + 485 + ], + "spans": [ + { + "bbox": [ + 46, + 461, + 287, + 485 + ], + "type": "text", + "content": "We now analyze SAMI and EfficientSAMs through a series of ablation studies with ViT backbones." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 486, + 287, + 570 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 486, + 287, + 570 + ], + "spans": [ + { + "bbox": [ + 46, + 486, + 287, + 570 + ], + "type": "text", + "content": "Reconstruction Loss. We study the effect of reconstruction loss on the performance of SAMI on ImageNet-1K. We compare our mean square error (MSE) reconstruction loss with cosine similarity loss. We find that MSE reconstruction loss performs better, shown in Tab. 6. This recommends a direct reconstruction of SAM features instead of the target with a high angular similarity." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "type": "text", + "content": "Cross-Attention Decoder. To reconstruct SAM features, we directly use the output tokens from encoder and only take decoder to transform the masked tokens with cross-attention. We study how the performance varies with cross-attention decoder. When changing the decoder module in MAE to our cross-attention decoder in SAMI, we find that SAMI-B improves the performance from " + }, + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "type": "inline_equation", + "content": "84.4\\%" + }, + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "type": "inline_equation", + "content": "84.8\\%" + }, + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "type": "text", + "content": " on ImageNet-1K. The cross-attention decoder of SAMI also shows consistent improvement for ViT-Tiny and ViT-Small. Analogy to anchor points in AnchorDETR[58], the output tokens from encoder are already learned well by directly aligning the SAM features, which can serve as anchor tokens" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 72, + 536, + 84 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 536, + 84 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 536, + 84 + ], + "type": "text", + "content": "for help masked tokens align via cross-attention decoder." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 84, + 545, + 132 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 84, + 545, + 132 + ], + "spans": [ + { + "bbox": [ + 305, + 84, + 545, + 132 + ], + "type": "text", + "content": "Mask Ratio. We explore how the performance varies with different mask ratio in SAMI. The observations are consistent with MAE [26] that a high mask ratio, " + }, + { + "bbox": [ + 305, + 84, + 545, + 132 + ], + "type": "inline_equation", + "content": "75\\%" + }, + { + "bbox": [ + 305, + 84, + 545, + 132 + ], + "type": "text", + "content": ", tends to produce good results." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 133, + 547, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 133, + 547, + 216 + ], + "spans": [ + { + "bbox": [ + 305, + 133, + 547, + 216 + ], + "type": "text", + "content": "Reconstruction Target. We study the impact of reconstruction target. We take a different encoder from CLIP [46] to generate features as the reconstruction target in SAMI. Aligning features from CLIP encoder can also outperform MAE by " + }, + { + "bbox": [ + 305, + 133, + 547, + 216 + ], + "type": "inline_equation", + "content": "0.8\\%" + }, + { + "bbox": [ + 305, + 133, + 547, + 216 + ], + "type": "text", + "content": " for a ViT-Tiny model on ImageNet-1K. This demonstrates that masked image pretraining benefits from powerful guided reconstruction." + } + ] + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 338, + 225, + 515, + 319 + ], + "blocks": [ + { + "bbox": [ + 338, + 225, + 515, + 319 + ], + "lines": [ + { + "bbox": [ + 338, + 225, + 515, + 319 + ], + "spans": [ + { + "bbox": [ + 338, + 225, + 515, + 319 + ], + "type": "image", + "image_path": "4120f34b8c2d8137670efcaea0b014cf992c37e92c1b7fd8d61207ac381ecb46.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 306, + 328, + 545, + 361 + ], + "lines": [ + { + "bbox": [ + 306, + 328, + 545, + 361 + ], + "spans": [ + { + "bbox": [ + 306, + 328, + 545, + 361 + ], + "type": "text", + "content": "Figure 7. Ablation study on training steps for EfficientSAMs on MS COCO dataset. Zero-shot single point valid mask evaluation with a single box prompt is performed for the ablation." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 373, + 545, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 373, + 545, + 479 + ], + "spans": [ + { + "bbox": [ + 305, + 373, + 545, + 479 + ], + "type": "text", + "content": "Effects of Finetuning Steps for EfficientSAMs. We explore the effect of fintuning steps for EfficientSAMs. As illustrated in Fig. 7, EfficientSAM-Ti and EfficientSAM-S achieve decent performance even at 0.1 epoch. For 1 epoch, the performance gain is larger than " + }, + { + "bbox": [ + 305, + 373, + 545, + 479 + ], + "type": "inline_equation", + "content": "2.5\\mathrm{mIoU}" + }, + { + "bbox": [ + 305, + 373, + 545, + 479 + ], + "type": "text", + "content": ". The final performance of EfficientSAM-S reaches " + }, + { + "bbox": [ + 305, + 373, + 545, + 479 + ], + "type": "inline_equation", + "content": "76.9\\mathrm{mIoU}" + }, + { + "bbox": [ + 305, + 373, + 545, + 479 + ], + "type": "text", + "content": ", which is only " + }, + { + "bbox": [ + 305, + 373, + 545, + 479 + ], + "type": "inline_equation", + "content": "1.5\\mathrm{mIoU}" + }, + { + "bbox": [ + 305, + 373, + 545, + 479 + ], + "type": "text", + "content": " lower than SAM. These results demonstrate the advantages of SAMI-pretrained image encoders and our EfficientSAMs." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 480, + 547, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 480, + 547, + 540 + ], + "spans": [ + { + "bbox": [ + 305, + 480, + 547, + 540 + ], + "type": "text", + "content": "Efficiency. We compare the throughput (images per second) of our EfficientSAM with SAM and other models on A100 in Fig. 1. Our EfficientSAM reduces the runtime/parameters of SAM by " + }, + { + "bbox": [ + 305, + 480, + 547, + 540 + ], + "type": "inline_equation", + "content": "\\sim 20\\mathrm{x}" + }, + { + "bbox": [ + 305, + 480, + 547, + 540 + ], + "type": "text", + "content": ". We also observed similar speedup (" + }, + { + "bbox": [ + 305, + 480, + 547, + 540 + ], + "type": "inline_equation", + "content": "\\sim 20\\mathrm{x}" + }, + { + "bbox": [ + 305, + 480, + 547, + 540 + ], + "type": "text", + "content": ") comparing to SAM w.r.t. FLOPs." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 550, + 378, + 562 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 550, + 378, + 562 + ], + "spans": [ + { + "bbox": [ + 306, + 550, + 378, + 562 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 570, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 547, + 714 + ], + "type": "text", + "content": "We proposed a masked image pretraining approach, SAMI, to explore the potential of ViTs under the guidance of SAM foundation model. SAMI improves masked image pretraining by reconstructing the latent features from SAM image encoder to transfer knowledge from vision foundation model to ViTs. Extensive experiments on image classification, object detection and instance segmentation, semantic segmentation, and the segment anything task consistently validate SAMI's advantages. We also demonstrate that SAMI helps build efficient SAMs with pretrained light-weight encoders. Our preliminary work suggests that SAMI has potential applications beyond efficient segment anything task." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16118" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 714 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 53, + 91, + 288, + 134 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 288, + 134 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 288, + 134 + ], + "type": "text", + "content": "[1] Pablo Arbelaez, Michael Maire, Charless Fowlkes, and Jitendra Malik. Contour detection and hierarchical image segmentation. IEEE transactions on pattern analysis and machine intelligence, 33(5):898-916, 2010. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 135, + 288, + 191 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 135, + 288, + 191 + ], + "spans": [ + { + "bbox": [ + 53, + 135, + 288, + 191 + ], + "type": "text", + "content": "[2] Yutong Bai, Zeyu Wang, Junfei Xiao, Chen Wei, Huiyu Wang, Alan L Yuille, Yuyin Zhou, and Cihang Xie. Masked autoencoders enable efficient knowledge distillers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24256-24265, 2023. 3, 5, 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 191, + 288, + 224 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 191, + 288, + 224 + ], + "spans": [ + { + "bbox": [ + 53, + 191, + 288, + 224 + ], + "type": "text", + "content": "[3] Hangbo Bao, Li Dong, Songhao Piao, and Furu Wei. Beit: Bert pre-training of image transformers. arXiv preprint arXiv:2106.08254, 2021. 3, 5, 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 224, + 288, + 257 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 224, + 288, + 257 + ], + "spans": [ + { + "bbox": [ + 53, + 224, + 288, + 257 + ], + "type": "text", + "content": "[4] Ali Borji, Ming-Ming Cheng, Qibin Hou, Huaizu Jiang, and Jia Li. Salient object detection: A survey. Computational visual media, 5:117-150, 2019. 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 258, + 288, + 312 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 258, + 288, + 312 + ], + "spans": [ + { + "bbox": [ + 53, + 258, + 288, + 312 + ], + "type": "text", + "content": "[5] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In European conference on computer vision, pages 213-229. Springer, 2020. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 313, + 288, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 313, + 288, + 369 + ], + "spans": [ + { + "bbox": [ + 53, + 313, + 288, + 369 + ], + "type": "text", + "content": "[6] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9650-9660, 2021. 3, 5, 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 369, + 288, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 369, + 288, + 413 + ], + "spans": [ + { + "bbox": [ + 53, + 369, + 288, + 413 + ], + "type": "text", + "content": "[7] Jun Cen, Yizheng Wu, Kewei Wang, Xingyi Li, Jingkang Yang, Yixuan Pei, Lingdong Kong, Ziwei Liu, and Qifeng Chen. Sad: Segment any rgbd. arXiv preprint arXiv:2305.14207, 2023. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 414, + 288, + 446 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 414, + 288, + 446 + ], + "spans": [ + { + "bbox": [ + 53, + 414, + 288, + 446 + ], + "type": "text", + "content": "[8] Jiaqi Chen, Zeyu Yang, and Li Zhang. Semantic segment anything. https://github.com/fudan-zvg/Semantic-Segment-Anything, 2023.2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 447, + 288, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 447, + 288, + 491 + ], + "spans": [ + { + "bbox": [ + 53, + 447, + 288, + 491 + ], + "type": "text", + "content": "[9] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International conference on machine learning, pages 1597-1607. PMLR, 2020. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 491, + 288, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 491, + 288, + 557 + ], + "spans": [ + { + "bbox": [ + 48, + 491, + 288, + 557 + ], + "type": "text", + "content": "[10] Tianrun Chen, Lanyun Zhu, Chaotao Deng, Runlong Cao, Yan Wang, Shangzhan Zhang, Zejian Li, Lingyun Sun, Ying Zang, and Papa Mao. Sam-adapter: Adapting segment anything in underperformed scenes. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3367-3375, 2023. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 558, + 288, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 558, + 288, + 602 + ], + "spans": [ + { + "bbox": [ + 48, + 558, + 288, + 602 + ], + "type": "text", + "content": "[11] Xinlei Chen and Kaiming He. Exploring simple siamese representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 15750-15758, 2021. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 602, + 288, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 602, + 288, + 658 + ], + "spans": [ + { + "bbox": [ + 48, + 602, + 288, + 658 + ], + "type": "text", + "content": "[12] Xiaokang Chen, Mingyu Ding, Xiaodi Wang, Ying Xin, Shentong Mo, Yunhao Wang, Shumin Han, Ping Luo, Gang Zeng, and Jingdong Wang. Context autoencoder for self-supervised representation learning. International Journal of Computer Vision, pages 1-16, 2023. 5, 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 658, + 288, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 288, + 714 + ], + "type": "text", + "content": "[13] Bowen Cheng, Ishan Misra, Alexander G Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1290–1299, 2022. 5, 6" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 714 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "text", + "content": "[14] Yangming Cheng, Lulei Li, Yuanyou Xu, Xiaodi Li, Zongxin Yang, Wenguan Wang, and Yi Yang. Segment and track anything. arXiv preprint arXiv:2305.06558, 2023. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 107, + 547, + 163 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 107, + 547, + 163 + ], + "spans": [ + { + "bbox": [ + 307, + 107, + 547, + 163 + ], + "type": "text", + "content": "[15] Ekin D Cubuk, Barret Zoph, Jonathon Shlens, and Quoc V Le. Randaugment: Practical automated data augmentation with a reduced search space. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops, pages 702-703, 2020. 5" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 163, + 547, + 208 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 163, + 547, + 208 + ], + "spans": [ + { + "bbox": [ + 308, + 163, + 547, + 208 + ], + "type": "text", + "content": "[16] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 5" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 209, + 547, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 209, + 547, + 274 + ], + "spans": [ + { + "bbox": [ + 308, + 209, + 547, + 274 + ], + "type": "text", + "content": "[17] Ruining Deng, Can Cui, Quan Liu, Tianyuan Yao, Lucas W Remedios, Shunxing Bao, Bennett A Landman, Lee E Whelless, Lori A Coburn, Keith T Wilson, et al. Segment anything model (sam) for digital pathology: Assess zero-shot segmentation on whole slide imaging. arXiv preprint arXiv:2304.04155, 2023. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 276, + 547, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 276, + 547, + 331 + ], + "spans": [ + { + "bbox": [ + 308, + 276, + 547, + 331 + ], + "type": "text", + "content": "[18] Xiaoyi Dong, Jianmin Bao, Ting Zhang, Dongdong Chen, Weiming Zhang, Lu Yuan, Dong Chen, Fang Wen, and Nenghai Yu. Bootstrapped masked autoencoders for vision bert pretraining. In European Conference on Computer Vision, pages 247-264. Springer, 2022. 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 332, + 547, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 332, + 547, + 398 + ], + "spans": [ + { + "bbox": [ + 308, + 332, + 547, + 398 + ], + "type": "text", + "content": "[19] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 1, 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 399, + 547, + 455 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 399, + 547, + 455 + ], + "spans": [ + { + "bbox": [ + 308, + 399, + 547, + 455 + ], + "type": "text", + "content": "[20] Haoqi Fan, Bo Xiong, Karttikeya Mangalam, Yanghao Li, Zhicheng Yan, Jitendra Malik, and Christoph Feichtenhofer. Multiscale vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6824-6835, 2021. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 456, + 547, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 456, + 547, + 511 + ], + "spans": [ + { + "bbox": [ + 308, + 456, + 547, + 511 + ], + "type": "text", + "content": "[21] Shanghua Gao, Zhijie Lin, Xingyu Xie, Pan Zhou, Ming-Ming Cheng, and Shuicheng Yan. Editanything: Empowering unparalleled flexibility in image editing and generation. In Proceedings of the 31st ACM International Conference on Multimedia, pages 9414–9416, 2023. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 512, + 547, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 512, + 547, + 577 + ], + "spans": [ + { + "bbox": [ + 308, + 512, + 547, + 577 + ], + "type": "text", + "content": "[22] Benjamin Graham, Alaaeldin El-Nouby, Hugo Touvron, Pierre Stock, Armand Joulin, Hervé Jégou, and Matthijs Douze. Levit: a vision transformer in convnet's clothing for faster inference. In Proceedings of the IEEE/CVF international conference on computer vision, pages 12259-12269, 2021. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 578, + 547, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 578, + 547, + 624 + ], + "spans": [ + { + "bbox": [ + 308, + 578, + 547, + 624 + ], + "type": "text", + "content": "[23] Agrim Gupta, Piotr Dólár, and Ross Girshick. Lvis: A dataset for large vocabulary instance segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2019. 5" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 624, + 547, + 680 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 624, + 547, + 680 + ], + "spans": [ + { + "bbox": [ + 308, + 624, + 547, + 680 + ], + "type": "text", + "content": "[24] Dongsheng Han, Chaoning Zhang, Yu Qiao, Maryam Qamar, Yuna Jung, SeungKyu Lee, Sung-Ho Bae, and Choong Seon Hong. Segment anything model (sam) meets glass: Mirror and transparent objects cannot be easily detected. arXiv preprint arXiv:2305.00278, 2023. 1, 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 681, + 547, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 681, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 308, + 681, + 547, + 714 + ], + "type": "text", + "content": "[25] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In Proceedings of the IEEE international conference on computer vision, pages 2961-2969, 2017. 6" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "16119" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 288, + 127 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 288, + 127 + ], + "type": "text", + "content": "[26] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollar, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16000-16009, 2022. 2, 3, 4, 5, 6, 8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 129, + 288, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 129, + 288, + 161 + ], + "spans": [ + { + "bbox": [ + 48, + 129, + 288, + 161 + ], + "type": "text", + "content": "[27] Geoffrey Hinton, Oriol Vinyals, and Jeff Dean. Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531, 2015. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 163, + 288, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 163, + 288, + 206 + ], + "spans": [ + { + "bbox": [ + 48, + 163, + 288, + 206 + ], + "type": "text", + "content": "[28] Zejiang Hou, Fei Sun, Yen-Kuang Chen, Yuan Xie, and Sun-Yuan Kung. Milan: Masked image pretraining on language assisted representation. arXiv preprint arXiv:2208.06049, 2022.3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 209, + 288, + 240 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 209, + 288, + 240 + ], + "spans": [ + { + "bbox": [ + 48, + 209, + 288, + 240 + ], + "type": "text", + "content": "[29] Jiaxi Jiang and Christian Holz. Restore anything pipeline: Segment anything meets image restoration. arXiv preprint arXiv:2305.13093, 2023. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 243, + 288, + 275 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 243, + 288, + 275 + ], + "spans": [ + { + "bbox": [ + 48, + 243, + 288, + 275 + ], + "type": "text", + "content": "[30] Glenn Jocher, Ayush Chaurasia, and Jing Qiu. Yolo by ultralytics. https://github.com/ultralytics/ultralytics, 2023. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 277, + 288, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 277, + 288, + 330 + ], + "spans": [ + { + "bbox": [ + 48, + 277, + 288, + 330 + ], + "type": "text", + "content": "[31] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. arXiv preprint arXiv:2304.02643, 2023. 1, 2, 3, 4, 5, 6, 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 333, + 288, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 333, + 288, + 365 + ], + "spans": [ + { + "bbox": [ + 48, + 333, + 288, + 365 + ], + "type": "text", + "content": "[32] Brett Koonce and Brett Koonce. *Mobilenetv3: Convolutional Neural Networks with Swift for Tensorflow: Image Recognition and Dataset Categorization*, pages 125–144, 2021. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 367, + 288, + 421 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 367, + 288, + 421 + ], + "spans": [ + { + "bbox": [ + 48, + 367, + 288, + 421 + ], + "type": "text", + "content": "[33] Jiashi Li, Xin Xia, Wei Li, Huixia Li, Xing Wang, Xuefeng Xiao, Rui Wang, Min Zheng, and Xin Pan. Next-vit: Next generation vision transformer for efficient deployment in realistic industrial scenarios. arXiv preprint arXiv:2207.05501, 2022. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 423, + 288, + 466 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 423, + 288, + 466 + ], + "spans": [ + { + "bbox": [ + 48, + 423, + 288, + 466 + ], + "type": "text", + "content": "[34] Yanghao Li, Hanzi Mao, Ross Girshick, and Kaiming He. Exploring plain vision transformer backbones for object detection. In European Conference on Computer Vision, pages 280-296. Springer, 2022. 2, 5, 6, 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 468, + 288, + 522 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 468, + 288, + 522 + ], + "spans": [ + { + "bbox": [ + 48, + 468, + 288, + 522 + ], + "type": "text", + "content": "[35] Yanyu Li, Geng Yuan, Yang Wen, Ju Hu, Georgios Evangelidis, Sergey Tulyakov, Yanzhi Wang, and Jian Ren. Efficientformer: Vision transformers at mobilenet speed. Advances in Neural Information Processing Systems, 35:12934-12949, 2022. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 525, + 288, + 579 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 525, + 288, + 579 + ], + "spans": [ + { + "bbox": [ + 48, + 525, + 288, + 579 + ], + "type": "text", + "content": "[36] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In European conference on computer vision, pages 740-755. Springer, 2014. 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 581, + 288, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 581, + 288, + 656 + ], + "spans": [ + { + "bbox": [ + 48, + 581, + 288, + 656 + ], + "type": "text", + "content": "[37] Ruiping Liu, Jiaming Zhang, Kunyu Peng, Junwei Zheng, Ke Cao, Yufan Chen, Kailun Yang, and Rainer Stiefelhagen. Open scene understanding: Grounded situation recognition meets segment anything for helping people with visual impairments. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1857-1867, 2023. 1, 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 658, + 288, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 288, + 713 + ], + "type": "text", + "content": "[38] Xinyu Liu, Houwen Peng, Ningxin Zheng, Yuqing Yang, Han Hu, and Yixuan Yuan. Efficientvit: Memory efficient vision transformer with cascaded group attention. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14420-14430, 2023. 3" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 712 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 307, + 73, + 546, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 546, + 128 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 546, + 128 + ], + "type": "text", + "content": "[39] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10012-10022, 2021. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 129, + 547, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 129, + 547, + 161 + ], + "spans": [ + { + "bbox": [ + 307, + 129, + 547, + 161 + ], + "type": "text", + "content": "[40] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In International Conference on Learning Representations, 2018. 5" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 162, + 547, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 162, + 547, + 183 + ], + "spans": [ + { + "bbox": [ + 307, + 162, + 547, + 183 + ], + "type": "text", + "content": "[41] Jun Ma and Bo Wang. Segment anything in medical images. arXiv preprint arXiv:2304.12306, 2023. 1, 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 184, + 547, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 184, + 547, + 227 + ], + "spans": [ + { + "bbox": [ + 308, + 184, + 547, + 227 + ], + "type": "text", + "content": "[42] Sachin Mehta and Mohammad Rastegari. Mobilevit: Lightweight, general-purpose, and mobile-friendly vision transformer. In International Conference on Learning Representations, 2021. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 228, + 547, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 228, + 547, + 281 + ], + "spans": [ + { + "bbox": [ + 308, + 228, + 547, + 281 + ], + "type": "text", + "content": "[43] Deepak Pathak, Philipp Krahenbuhl, Jeff Donahue, Trevor Darrell, and Alexei A Efros. Context encoders: Feature learning by inpainting. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2536-2544, 2016. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 282, + 547, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 282, + 547, + 326 + ], + "spans": [ + { + "bbox": [ + 308, + 282, + 547, + 326 + ], + "type": "text", + "content": "[44] Zhiliang Peng, Li Dong, Hangbo Bao, Qixiang Ye, and Furu Wei. Beit v2: Masked image modeling with vector-quantized visual tokenizers. arXiv preprint arXiv:2208.06366, 2022. 2, 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 327, + 547, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 327, + 547, + 370 + ], + "spans": [ + { + "bbox": [ + 308, + 327, + 547, + 370 + ], + "type": "text", + "content": "[45] Xuebin Qin, Zichen Zhang, Chenyang Huang, Masood Dehghan, Osmar Zaiane, and Martin Jagersand. U2-net: Going deeper with nested u-structure for salient object detection. page 107404, 2020. 7, 8" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 372, + 547, + 424 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 372, + 547, + 424 + ], + "spans": [ + { + "bbox": [ + 308, + 372, + 547, + 424 + ], + "type": "text", + "content": "[46] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In International Conference on Machine Learning, pages 8821-8831. PMLR, 2021. 8" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 426, + 547, + 469 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 426, + 547, + 469 + ], + "spans": [ + { + "bbox": [ + 308, + 426, + 547, + 469 + ], + "type": "text", + "content": "[47] Adriana Romero, Nicolas Ballas, Samira Ebrahimi Kahou, Antoine Chassang, Carlo Gatta, and Yoshua Bengio. Fitnets: Hints for thin deep nets. arXiv preprint arXiv:1412.6550, 2014.3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 471, + 547, + 524 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 471, + 547, + 524 + ], + "spans": [ + { + "bbox": [ + 308, + 471, + 547, + 524 + ], + "type": "text", + "content": "[48] Mark Sandler, Andrew Howard, Menglong Zhu, Andrey Zh-moginov, and Liang-Chieh Chen. Mobilenetv2: Inverted residuals and linear bottlenecks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4510-4520, 2018. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 525, + 547, + 558 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 525, + 547, + 558 + ], + "spans": [ + { + "bbox": [ + 308, + 525, + 547, + 558 + ], + "type": "text", + "content": "[49] Qiuhong Shen, Xingyi Yang, and Xinchao Wang. Anything-3d: Towards single-view anything reconstruction in the wild. arXiv preprint arXiv:2304.10261, 2023. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 559, + 547, + 592 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 559, + 547, + 592 + ], + "spans": [ + { + "bbox": [ + 308, + 559, + 547, + 592 + ], + "type": "text", + "content": "[50] Ao Sun, Pingchuan Ma, Yuanyuan Yuan, and Shuai Wang. Explain any concept: Segment anything meets concept-based explanation. arXiv preprint arXiv:2305.10289, 2023. 1, 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 593, + 547, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 593, + 547, + 624 + ], + "spans": [ + { + "bbox": [ + 308, + 593, + 547, + 624 + ], + "type": "text", + "content": "[51] Lv Tang, Haoke Xiao, and Bo Li. Can sam segment anything? when sam meets camouflaged object detection. arXiv preprint arXiv:2304.04709, 2023. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 625, + 547, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 625, + 547, + 658 + ], + "spans": [ + { + "bbox": [ + 308, + 625, + 547, + 658 + ], + "type": "text", + "content": "[52] Shehbaz Tariq, Brian Estadimas Arfeto, Chaoning Zhang, and Hyundong Shin. Segment anything meets semantic communication. arXiv preprint arXiv:2306.02094, 2023. 1, 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 658, + 547, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 658, + 547, + 712 + ], + "spans": [ + { + "bbox": [ + 308, + 658, + 547, + 712 + ], + "type": "text", + "content": "[53] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Hervé Jégou. Training data-efficient image transformers & distillation through attention. In International conference on machine learning, pages 10347-10357. PMLR, 2021. 2, 5, 6" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "16120" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "text", + "content": "[54] Koen EA Van de Sande, Jasper RR Uijlings, Theo Gevers, and Arnold WM Smeulders. Segmentation as selective search for object recognition. In 2011 international conference on computer vision, pages 1879-1886. IEEE, 2011. 1" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 118, + 288, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 118, + 288, + 161 + ], + "spans": [ + { + "bbox": [ + 49, + 118, + 288, + 161 + ], + "type": "text", + "content": "[55] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 162, + 288, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 162, + 288, + 216 + ], + "spans": [ + { + "bbox": [ + 49, + 162, + 288, + 216 + ], + "type": "text", + "content": "[56] Pascal Vincent, Hugo Larochelle, Isabelle Lajoie, Yoshua Bengio, Pierre-Antoine Manzagol, and León Bottou. Stacked denoising autoencoders: Learning useful representations in a deep network with a local denoising criterion. Journal of machine learning research, 11(12), 2010. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 217, + 288, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 217, + 288, + 270 + ], + "spans": [ + { + "bbox": [ + 49, + 217, + 288, + 270 + ], + "type": "text", + "content": "[57] Xinlong Wang, Rufeng Zhang, Chunhua Shen, Tao Kong, and Lei Li. Dense contrastive learning for self-supervised visual pre-training. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3024-3033, 2021. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 49, + 272, + 288, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 272, + 288, + 316 + ], + "spans": [ + { + "bbox": [ + 49, + 272, + 288, + 316 + ], + "type": "text", + "content": "[58] Yingming Wang, Xiangyu Zhang, Tong Yang, and Jian Sun. Anchor detr: Query design for transformer-based detector. In Proceedings of the AAAI conference on artificial intelligence, pages 2567-2575, 2022. 8" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 49, + 317, + 288, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 317, + 288, + 370 + ], + "spans": [ + { + "bbox": [ + 49, + 317, + 288, + 370 + ], + "type": "text", + "content": "[59] Chen Wei, Haoqi Fan, Saining Xie, Chao-Yuan Wu, Alan Yuille, and Christoph Feichtenhofer. Masked feature prediction for self-supervised visual pre-training. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14668-14678, 2022. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 49, + 372, + 287, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 372, + 287, + 426 + ], + "spans": [ + { + "bbox": [ + 49, + 372, + 287, + 426 + ], + "type": "text", + "content": "[60] Haiyan Wu, Yuting Gao, Yinqi Zhang, Shaohui Lin, Yuan Xie, Xing Sun, and Ke Li. Self-supervised models are good teaching assistants for vision transformers. In Proceedings of the 39th International Conference on Machine Learning, pages 24031-24042. PMLR, 2022. 3, 5, 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 49, + 427, + 287, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 427, + 287, + 480 + ], + "spans": [ + { + "bbox": [ + 49, + 427, + 287, + 480 + ], + "type": "text", + "content": "[61] Kan Wu, Jinnian Zhang, Houwen Peng, Mengchen Liu, Bin Xiao, Jianlong Fu, and Lu Yuan. Tinyvit: Fast pretraining distillation for small vision transformers. In European Conference on Computer Vision, pages 68-85. Springer, 2022. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 49, + 482, + 287, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 482, + 287, + 536 + ], + "spans": [ + { + "bbox": [ + 49, + 482, + 287, + 536 + ], + "type": "text", + "content": "[62] Zhenda Xie, Yutong Lin, Zheng Zhang, Yue Cao, Stephen Lin, and Han Hu. Propagate yourself: Exploring pixel-level consistency for unsupervised visual representation learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16684-16693, 2021. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 49, + 537, + 287, + 591 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 537, + 287, + 591 + ], + "spans": [ + { + "bbox": [ + 49, + 537, + 287, + 591 + ], + "type": "text", + "content": "[63] Zhenda Xie, Zheng Zhang, Yue Cao, Yutong Lin, Jianmin Bao, Zhuliang Yao, Qi Dai, and Han Hu. Simmim: A simple framework for masked image modeling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9653-9663, 2022. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 49, + 593, + 287, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 593, + 287, + 635 + ], + "spans": [ + { + "bbox": [ + 49, + 593, + 287, + 635 + ], + "type": "text", + "content": "[64] Jing Yang, Brais Martinez, Adrian Bulat, and Georgios Tzimiropoulos. Knowledge distillation via softmax regression representation learning. In International Conference on Learning Representations, 2021. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 49, + 637, + 287, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 637, + 287, + 669 + ], + "spans": [ + { + "bbox": [ + 49, + 637, + 287, + 669 + ], + "type": "text", + "content": "[65] Jinyu Yang, Mingqi Gao, Zhe Li, Shang Gao, Fangjing Wang, and Feng Zheng. Track anything: Segment anything meets videos. arXiv preprint arXiv:2304.11968, 2023. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 49, + 670, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 670, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 49, + 670, + 287, + 712 + ], + "type": "text", + "content": "[66] Haoran You, Yunyang Xiong, Xiaoliang Dai, Bichen Wu, Peizhao Zhang, Haoqi Fan, Peter Vajda, and Yingyan Celine Lin. Castling-vit: Compressing self-attention via switching towards linear-angular attention at vision transformer inference." + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 547, + 397 + ], + "type": "list", + "angle": 0, + "index": 22, + "blocks": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "type": "text", + "content": "In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14431-14442, 2023. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 308, + 97, + 547, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 97, + 547, + 140 + ], + "spans": [ + { + "bbox": [ + 308, + 97, + 547, + 140 + ], + "type": "text", + "content": "[67] Tao Yu, Runseng Feng, Ruoyu Feng, Jinming Liu, Xin Jin, Wenjun Zeng, and Zhibo Chen. Inpaint anything: Segment anything meets image inpainting. arXiv preprint arXiv:2304.06790, 2023. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 141, + 547, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 141, + 547, + 194 + ], + "spans": [ + { + "bbox": [ + 308, + 141, + 547, + 194 + ], + "type": "text", + "content": "[68] Chaoning Zhang, Dongshen Han, Yu Qiao, Jung Uk Kim, Sung-Ho Bae, Seungkyu Lee, and Choong Seon Hong. Faster segment anything: Towards lightweight sam for mobile applications. arXiv preprint arXiv:2306.14289, 2023. 2, 5, 6, 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 197, + 547, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 197, + 547, + 229 + ], + "spans": [ + { + "bbox": [ + 308, + 197, + 547, + 229 + ], + "type": "text", + "content": "[69] Xiao Feng Zhang, Tian Yi Song, and Jia Wei Yao. Deshadow—anything: When segment anything model meets zero-shot shadow removal. arXiv preprint arXiv:2309.11715, 2023. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 231, + 547, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 231, + 547, + 274 + ], + "spans": [ + { + "bbox": [ + 308, + 231, + 547, + 274 + ], + "type": "text", + "content": "[70] Borui Zhao, Quan Cui, Renjie Song, Yiyu Qiu, and Jiajun Liang. Decoupled knowledge distillation. In Proceedings of the IEEE/CVF Conference on computer vision and pattern recognition, pages 11953-11962, 2022. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 276, + 547, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 276, + 547, + 308 + ], + "spans": [ + { + "bbox": [ + 308, + 276, + 547, + 308 + ], + "type": "text", + "content": "[71] Xu Zhao, Wenchao Ding, Yongqi An, Yinglong Du, Tao Yu, Min Li, Ming Tang, and Jinqiao Wang. Fast segment anything. arXiv preprint arXiv:2306.12156, 2023. 2, 5, 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 309, + 547, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 309, + 547, + 353 + ], + "spans": [ + { + "bbox": [ + 308, + 309, + 547, + 353 + ], + "type": "text", + "content": "[72] Bolei Zhou, Hang Zhao, Xavier Puig, Sanja Fidler, Adela Barriuso, and Antonio Torralba. Scene parsing through ade20k dataset. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 633-641, 2017. 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 354, + 547, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 354, + 547, + 397 + ], + "spans": [ + { + "bbox": [ + 308, + 354, + 547, + 397 + ], + "type": "text", + "content": "[73] Jinghao Zhou, Chen Wei, Huiyu Wang, Wei Shen, Cihang Xie, Alan Yuille, and Tao Kong. ibot: Image bert pre-training with online tokenizer. arXiv preprint arXiv:2111.07832, 2021. 5, 6" + } + ] + } + ], + "index": 21 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "16121" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Efficiently Assemble Normalization Layers and Regularization for Federated Domain Generalization/16ce0297-575e-48b3-acb7-4a9aa0af4fe6_content_list.json b/2024/Efficiently Assemble Normalization Layers and Regularization for Federated Domain Generalization/16ce0297-575e-48b3-acb7-4a9aa0af4fe6_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..a1195f5167f59ec34d4344e441684cfeac8a7431 --- /dev/null +++ b/2024/Efficiently Assemble Normalization Layers and Regularization for Federated Domain Generalization/16ce0297-575e-48b3-acb7-4a9aa0af4fe6_content_list.json @@ -0,0 +1,1278 @@ +[ + { + "type": "text", + "text": "Efficiently Assemble Normalization Layers and Regularization for Federated Domain Generalization", + "text_level": 1, + "bbox": [ + 151, + 130, + 818, + 174 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Khiem Le $^{1}$ , Long Ho $^{2}$ , Cuong Do $^{2}$ , Danh Le-Phuoc $^{3}$ , Kok-Seng Wong $^{2,*}$ $^{1}$ Department of Computer Science and Engineering, University of Notre Dame, IN, USA \n $^{2}$ College of Engineering and Computer Science, VinUniversity, Hanoi, Vietnam \n $^{3}$ Open Distributed Systems, Technical University Berlin, Berlin, Germany", + "bbox": [ + 130, + 202, + 841, + 275 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "kle3@nd.edu, dans.lephuoc@tu-berlin.de, {long.ht, cuong.dd, wong.ks}@vinuni.edu.vn", + "bbox": [ + 119, + 277, + 843, + 292 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 325, + 313, + 343 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Domain shift is a formidable issue in Machine Learning that causes a model to suffer from performance degradation when tested on unseen domains. Federated Domain Generalization (FedDG) attempts to train a global model using collaborative clients in a privacy-preserving manner that can generalize well to unseen clients possibly with domain shift. However, most existing FedDG methods either cause additional privacy risks of data leakage or induce significant costs in client communication and computation, which are major concerns in the Federated Learning paradigm. To circumvent these challenges, here we introduce a novel architectural method for FedDG, namely gPerXAN1, which relies on a normalization scheme working with a guiding regularizer. In particular, we carefully design Personalized eXplicitly Assessed Normalization to enforce client models selectively filtering domain-specific features that are biased towards local data while retaining discrimination of those features. Then, we incorporate a simple yet effective regularizer to guide these models in directly capturing domain-invariant representations that the global model's classifier can leverage. Extensive experimental results on two benchmark datasets, i.e., PACS and Office-Home, and a real-world medical dataset, Camelyon17, indicate that our proposed method outperforms other existing methods in addressing this particular problem.", + "bbox": [ + 76, + 358, + 473, + 737 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 763, + 209, + 780 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Over the past few decades, Machine Learning (ML) has demonstrated remarkable achievements across diverse areas such as Computer Vision, Natural Language and Speech Processing, or Robotics [5]. In general, most ML models rely on an over-simplified assumption, i.e., the train-", + "bbox": [ + 75, + 789, + 468, + 866 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ing and testing data are independent and identically distributed, which does not always reflect real-world practices. In practical scenarios where the distribution of testing data diverges from that of training data, the performance of ML models often drops catastrophically due to the domain shift issue [28]. Additionally, obtaining or identifying the testing data before model deployment can be challenging in numerous applications. For instance, in biomedical applications where data characteristics vary across different equipment and institutions, gathering data from all potential domains in advance is impractical. Therefore, it is essential to have a solution that can improve the generalization capability of such ML models to adapt effectively to unseen domains.", + "bbox": [ + 500, + 328, + 893, + 525 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Domain Generalization (DG) has been proposed to address the challenge of training ML models using data from single or multiple source domains with the expectation that these models will perform well on unseen domains [41]. The majority of existing DG methods fall under the category of domain-invariant representation learning approach [13, 20, 25, 29, 33]. This approach relies on a broadly acknowledged assumption that each domain contains its own domain-specific features, which are biased towards spurious relations in the data, and that all domains share domain-invariant features, which are general and robust to any unseen domains. From this assumption, previous works propose methods that remove domain-specific features and distill domain-invariant features to achieve the generalization ability. Alternative approaches for DG encompass data augmentation [23, 43, 44, 47], which involves exposing models to artificially generated domains, and meta-learning [1, 6, 12], an approach that emulates the domain shift during the meta-training phase. However, most methods require a centralized setting where all source domains are collected together. Consequently, these methods cannot be readily expanded to decentralized settings.", + "bbox": [ + 498, + 531, + 893, + 864 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Federated Learning (FL) [21] is an emerging decentralized learning paradigm widely adopted in various applica", + "bbox": [ + 500, + 869, + 893, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author: wong.rs@vinuni.edu.vn", + "bbox": [ + 94, + 875, + 392, + 887 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "$^{1}$ https://github.com/lhkhiem28/gPerXAN", + "bbox": [ + 96, + 887, + 315, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "6027", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/2d3d7aeaf622fa60cb0633c1b35aee5ab8449b09c55d3cafeff7f42fa3f3a00c.jpg", + "image_caption": [ + "Figure 1. An overview of our proposed gPerXAN method for solving the FedDG problem." + ], + "image_footnote": [], + "bbox": [ + 83, + 87, + 897, + 364 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "tions to cope with the increasing privacy concerns of data centralization [40]. Specifically, the paradigm works in a way that each client learns from their data and only aggregates local models' parameters at a certain frequency at the central server to generate a global model. Notably, all data samples are kept within each client during the FL training process. Due to the nature of data decentralization, where each client owns a single source domain, as illustrated in Figure 1, the FL paradigm poses further significant challenges for DG and limits the applicability of available DG methods. There have been some early attempts to address the DG problem in the FL scenario. For instance, Liu et al. [18] introduces a method that allows clients to share their image data in the frequency space with each other, thus relatively recovering the centralization process at each client. Similarly, Chen et al. [3] introduces another method that extracts and exchanges the style of local images among all clients. It is evident that these initial efforts employ a strategy that necessitates the sharing of partial client data, thereby compromising the data privacy constraints of FL to a certain extent. Although they show promising results, these methods can be overly complicated to implement in practice and lead to additional privacy risks during the FL training process.", + "bbox": [ + 75, + 412, + 472, + 775 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address the aforementioned challenges, this paper introduces a novel architectural method for domain-invariant representation learning within the FL framework. The proposed method enhances the generalization ability while upholding the fundamental privacy principles of FL. Based on the effectiveness of discarding domain-specific information from learned features [25, 30], we properly assemble Instance Normalization layers (IN) into Batch Normaliza", + "bbox": [ + 75, + 779, + 470, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "tion layers (BN) in well-known Convolutional Neural Networks (CNNs) using an explicit differential mixture as in Eqn (2). Moreover, thanks to the explicit property, the benefit of personalization in FL [27, 35] can be incorporated into the normalization scheme using local BN sides. Specifically, during the FL training process, while IN sides are globally aggregated along with other model parameters, BN sides are updated locally without broadcasting. In addition, we argue that only relying on the ability to filter domain-specific features of IN while lacking guidance to distill domain-invariant representations directly might lead to suboptimal performance. Based on this observation, we introduce a simple yet highly effective regularization term to guide client models to directly capture domain-invariant representations that can be used by the global model's classifier, which is aggregated from client models' classifiers.", + "bbox": [ + 496, + 412, + 893, + 655 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To summarize, our main contributions in this paper are highlighted as follows:", + "bbox": [ + 498, + 662, + 890, + 694 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Different from existing methods for DG in the FL scenario, we propose a novel method that concentrates on a personalized normalization scheme, global IN while local BN, for filtering domain-specific features and fully respecting the privacy-preserving principles of FL.", + "- Furthermore, we propose a simple yet effective regularization term to introduce clear guidance to client models for directly capturing domain-invariant representations, further improving performance on unseen domains.", + "- Finally, we conduct extensive experiments on two benchmark datasets, i.e., PACS and Office-Home, and a real-world medical dataset, Camelyon17, where our proposed method outperforms existing relevant ones." + ], + "bbox": [ + 504, + 703, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "6028", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 76, + 89, + 218, + 104 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.1. Domain Generalization", + "text_level": 1, + "bbox": [ + 76, + 114, + 292, + 130 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Domain Generalization is a challenging task requiring models to perform well on unseen domains. One dominant approach is domain-invariant representation learning, which aims to minimize the discrepancy among source domains, assuming that the resulting representation will be domain-invariant and generalize well on testing domains. Along this track, Muandet et al. [22] proposes to reduce the domain dissimilarity by using Maximum Mean Discrepancy. Pan et al. [25] combines IN and BN in popular CNNs to reap the benefits of removing domain-specific features while maintaining the ability to capture discriminative features. Meanwhile, Seo et al. [30] discloses that combining these normalization layers in a switchable mechanism can yield better performance. Another approach to DG is data augmentation, which augments source domains to a broader span of the training data space, enlarging the possibility of covering the span of the data in the testing domain. Zhou et al. [47] mixes styles of training instances, resulting in novel domains being synthesized implicitly. Recently, the meta-learning approach has drawn increasing attention from the DG community. Balaji et al. [1] proposes MetaReg that learns a regularization function, particularly for the network classification layer, while excluding the feature extractor. However, the majority of these methods rely on having access to a diverse set of source domains, making them not applicable to decentralized settings.", + "bbox": [ + 76, + 138, + 472, + 532 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. Federated Domain Generalization", + "text_level": 1, + "bbox": [ + 76, + 544, + 375, + 559 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "While there is a growing interest in addressing the DG problem within the FL framework, particularly in scenarios where each client possesses a single source domain, the existing literature on this subject remains relatively sparse. This FedDG problem was initially introduced by Liu et al. [18], who then proposed ELCFS that involves exchanging amplitude information in the frequency space among clients and leveraging episodic learning to further enhance performance. Chen et al. [3] proposes a similar mechanism CCST that extracts and exchanges the overall domain style of local images among all clients. Unfortunately, these early works require sharing partial information about the local data, which can be viewed as a form of data leakage and is undesirable in the FL setting. Moreover, these methods include broadcasting partial information and interpolating this information into new training data, which adds significant extra cost to communication and computation during the FL training process. Besides, Wu and Gong [42] proposes an architectural method COPA in which the global model consists of a domain-invariant representation extractor and an ensemble of domain-specific classifiers. Sun et al. [36] introduces FedKA that employs a server-side voting mech", + "bbox": [ + 76, + 568, + 472, + 902 + ], + "page_idx": 2 + }, + { + "type": "table", + "img_path": "images/acf83ca6a95d175a448784fcc11b2cd37e13bb4b21d88a8763975ce8fc09acc0.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodPrivacy RiskAdditional Cost
CommunicationComputation
ELCFS✓✓✓
CCST✓✓✓
COPA×✓✓
FedDG-GA××✓
gPerXAN (Ours)×××
", + "bbox": [ + 498, + 88, + 893, + 213 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Table 1. An advantage comparison of different methods.", + "bbox": [ + 527, + 223, + 862, + 237 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "anism that generates target domain pseudo-labels based on the consensus from clients to facilitate global model finetuning. Recently, Zhang et al. [46] proposes a novel global objective incorporating a variance reduction regularizer to encourage fairness, and then FedDG-GA is proposed to optimize this objective by dynamically calibrating the aggregation weights. Although relaxing from the data leakage issue, these methods also cause a large additional consumption of resources when the number of source domains and output classes increases. In this paper, we introduce an alternative architectural method that shares only the model update information during training. This ensures maximal data privacy and circumvents the communication and computation overhead issue mentioned earlier while achieving competitive results. In Table 1, we highlight the advantages of our method in comparison to previous ones.", + "bbox": [ + 496, + 263, + 893, + 503 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "One related research field to FedDG is Federated Domain Adaptation (FedDA). While both fields aim to maximize the model performance on unseen domains using existing source domains, FedDA can access the target domains while FedDG cannot see those data during training. Leveraging this assumption, FedDA methods such as FADA [26] or FMTDA [45], are typically able to align target and source features. However, this specific assumption also makes FedDA methods are inapplicable to FedDG.", + "bbox": [ + 496, + 505, + 893, + 642 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.3. Personalized Federated Learning", + "text_level": 1, + "bbox": [ + 500, + 650, + 790, + 666 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Another related and orthogonal line of work is Personalized Federated Learning (pFL). pFL aims to learn personalized models for different clients to tackle distribution shifts across client data, known as data heterogeneity. Li et al. [14] utilizes a regularization term in the local loss function so that the clients' trained models will not significantly differ from the global model. On the other side, Sun et al. [35] and Pillutla et al. [27] mitigate the heterogeneous distributions by only loading a subset of the global model's parameters rather than loading the entire model at each training round. Although related in terms of overcoming distribution shifts across clients, pFL focuses on improving the performance of participating clients, whereas our considered field FedDG focuses on improving the performance of unseen clients with unseen domains.", + "bbox": [ + 496, + 674, + 893, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "6029", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Methodology", + "text_level": 1, + "bbox": [ + 76, + 89, + 210, + 107 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Problem Formulation. First, we denote $X$ and $Y$ as the input space and the label space, respectively, of a specific task $\\mathcal{T}$ . In the standard FL setting, $N$ clients $\\{c_i\\}_{i=1}^N$ are involved in collaboratively constructing a global model $f_g$ for solving the task, where each client $c_i$ owns a dataset $D_i = \\{(x_k, y_k)\\}_{k=1,|D_i|}$ which is associated to a specific domain defined by a joint distribution $P_{X,Y}^{(i)}$ . Importantly, scattered datasets $\\{D_i\\}_{i=1}^N$ across clients satisfy:", + "bbox": [ + 76, + 114, + 468, + 243 + ], + "page_idx": 3 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- $P_{X,Y}^{(i)} \\neq P_{X,Y}^{(j)}$ with $1 \\leq i, j \\leq N$ and $i \\neq j$", + "- $P_{Y|X}^{(i)} = P_{Y|X}^{(j)}$ with $1 \\leq i, j \\leq N$ and $i \\neq j$ ." + ], + "bbox": [ + 78, + 247, + 383, + 297 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The objective of FedDG is leveraging scattered datasets to construct a model $f_{g}$ that can directly generalize to the unseen dataset $D_{U}$ with an unseen domain, which means:", + "bbox": [ + 76, + 301, + 468, + 347 + ], + "page_idx": 3 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- $P_{X,Y}^{(U)} \\neq P_{X,Y}^{(i)}$ with $1 \\leq i \\leq N$", + "- $P_{Y|X}^{(U)} = P_{Y|X}^{(i)}$ with $1 \\leq i \\leq N$" + ], + "bbox": [ + 78, + 349, + 302, + 400 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To this end, $N$ clients communicate with a central server for $T$ rounds. At each round, every client $c_{i}$ receives the same global model $f_{g}$ from the server and updates $f_{g}$ with their local dataset $D_{i}$ for $E$ epochs to establish its local model $f_{i}$ . The server then collects all trained models and aggregates them to update the global model. This process repeats until the global model converges. In this work, we consider the most popular FL framework, FedAvg [21], which aggregates client models as:", + "bbox": [ + 76, + 404, + 468, + 539 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nf _ {g} = \\sum_ {i = 1} ^ {N} \\frac {\\left| D _ {i} \\right|}{\\sum_ {1 \\leq j \\leq N} \\left| D _ {j} \\right|} f _ {i} \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 181, + 550, + 468, + 590 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Challenges. In the general spirit of DG, a model is expected to extensively explore multiple source domains to achieve domain-invariance in its learned latent representation. However, under the FL setting, each client is restricted to accessing only its own local data, which constrains to make full use of source domains and, consequently, limits learning of generalizable representation. Moreover, sharing data or even partial information about data among clients poses additional privacy risks and introduces communication and computation costs during the FL training process.", + "bbox": [ + 76, + 599, + 468, + 752 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1. Normalization Scheme", + "text_level": 1, + "bbox": [ + 76, + 760, + 287, + 776 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "eXplicitly Assessed Normalization. The first step towards solving the highlighted challenges, we introduce a novel normalization scheme called eXplicitly Assessed Normalization (XAN), which combines IN and BN as:", + "bbox": [ + 76, + 784, + 468, + 844 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {h} = w _ {i n} \\left(\\gamma_ {i n} \\frac {h - \\mu_ {i n}}{\\sqrt {\\sigma_ {i n} ^ {2} + \\epsilon}} + \\beta_ {i n}\\right) + w _ {b n} \\left(\\gamma_ {b n} \\frac {h - \\mu_ {b n}}{\\sqrt {\\sigma_ {b n} ^ {2} + \\epsilon}} + \\beta_ {b n}\\right), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 76, + 853, + 468, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $h, \\hat{h} \\in \\mathbb{R}^{B \\times C \\times W \\times H}$ are layer input and output activations, which are 4D tensors with dimensions of batch size $B$ , number of channels $C$ , width $W$ and height $H$ . Meanwhile, $\\mu$ and $\\sigma^2$ are means and variances captured by the normalization layers, respectively, $\\gamma$ and $\\beta$ are affine parameters, and $\\epsilon$ is for numerical stability.", + "bbox": [ + 498, + 89, + 890, + 181 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In essence, XAN is a sophisticated mixture mechanism of IN and BN to replace BN in the feature extractors of CNNs. In particular, $w_{in}$ and $w_{bn}$ are ratios to weight the mixture that allow the model to switch between IN and BN. These parameters are randomly initialized and optimized along with other model parameters during training in an end-to-end manner. Unlike previous works [19, 30], which uses an implicit mechanism to combine computed statistics of IN and BN, i.e., means and variances, XAN uses an explicit mechanism that combines the output activations of two normalization layers, this provides the model with a unique ability to separate IN from BN completely. Note that IN has shown great success in style transfer tasks [9] as it allows discarding the variability of visual styles such as object colors or textures from the content of images. This property makes IN beneficial for solving the DG problem. However, directly using IN to replace the conventional BN leads to the loss of discrimination in the learned features, resulting in performance degradation in classification tasks.", + "bbox": [ + 496, + 181, + 892, + 469 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Personalized eXplicitly Assessed Normalization. To accomplish our proposed normalization scheme, Personalized eXplicitly Assessed Normalization (PerXAN), we make a subtle modification to the conventional FedAvg [21] framework based on the explicit property of the above XAN. Specifically, during the FL training process, while IN sides of XAN layers are globally aggregated along with other model parameters, BN sides are updated locally, which means that parameters of BN sides are excluded from the broadcasting steps from the server. Notably, in inference time, the global model is generated by averaging all model parameters of clients. Our modification is motivated by a common observation that FedAvg normally results in poor convergence and performance in the presence of domain heterogeneity across clients [15, 48] with the major reason being that client models forget the acquired knowledge from previous rounds after aggregated [31]. The leading solution is to personalize a subset of the model, which benefits clients in learning better from their local data [27, 35]. Moreover, based on another finding that BN plays a crucial role in dealing with the domain shift issue in the centralized paradigm [16, 17], BN sides in XAN layers are finally made to be personalized. PerXAN is illustrated in Figure 1.", + "bbox": [ + 496, + 470, + 892, + 818 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Regularization as Guidance", + "text_level": 1, + "bbox": [ + 500, + 830, + 748, + 845 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "With the proposed normalization scheme PerXAN in place, now, we are ready to introduce our \"Regularization as Guidance\". It targets to induce clear guidance to client", + "bbox": [ + 500, + 854, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "6030", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "models through a regularizer, which brings an effect of domain alignment [33, 34], and guides these models to capture domain-invariant representations directly. We hypothesize that this will address a drawback of current DG methods that are later verified by our experiments in Section 4. The drawback is from DG methods that take advantage of the IN's function [25, 30] but do not actually equip the model with the capability of capturing domain-invariant features even if they demonstrated promising performance. Instead, it is only hoped that domain-invariant features would be distilled through achieving the goal of removing domain-specific features. This indirect learning purpose might affect the model's learning efficacy, especially in the FL setting, where each client only owns a separate single source domain. Specifically, we assume a classification model $f$ , either the global model $f_{g}$ or a client model $f_{i}$ , comprises a feature extractor $g$ and a classifier head $h$ . At each client $c_{i}$ , during local training, the client model $f_{i}$ is optimized on the local dataset $D_{i}$ using the following loss function:", + "bbox": [ + 76, + 90, + 472, + 380 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} L _ {i} = L _ {c l s} ^ {(i)} \\left(f _ {i}; D _ {i}\\right) + \\lambda L _ {r e g} ^ {(i)} \\left(g _ {i}, h _ {g}; D _ {i}\\right) (3) \\\\ = \\sum_ {k = 1} ^ {| D _ {i} |} \\ell \\left(f _ {i} \\left(x _ {k} ^ {(i)}\\right), y _ {k} ^ {(i)}\\right) + \\lambda \\ell \\left(h _ {g} \\left(g _ {i} \\left(x _ {k} ^ {(i)}\\right)\\right), y _ {k} ^ {(i)}\\right), (4) \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 94, + 390, + 468, + 457 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\ell$ is the base loss function, which is usually cross-entropy in classification tasks, and $\\lambda$ is a hyper-parameter to control the significance of the regularizer. In a deeper look, by freezing $h_g$ during local training, the auxiliary loss term $L_{reg}$ forces client models to arrive at representations that can be made use for classification by the same global classifier, hence, producing an alignment effect on these representations [33, 34]. Moreover, our proposed regularizer can also be interpreted as an implicit form of matching global knowledge to clients' knowledge, which has been demonstrated to bring performance gain [3, 18] for FedDG.", + "bbox": [ + 76, + 469, + 468, + 636 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Due to its orthogonality, our proposed regularizer can be easily integrated into various methods, as shown in Algorithm 1. Specifically, our experiments in Section 4 verify its compatibility with the normalization scheme that justifies our motivation. Moreover, by using only the global model's classifier to regularize feature extractors, gPerXAN saves major communication and computation resources, as well as memory usage at clients compared to others, which use an ensemble of classifiers [42] or entire global model [46].", + "bbox": [ + 75, + 637, + 470, + 773 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments and Results", + "text_level": 1, + "bbox": [ + 76, + 789, + 310, + 806 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Datasets", + "text_level": 1, + "bbox": [ + 76, + 815, + 179, + 830 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To evaluate the proposed method, we perform experiments on two standard DG benchmark datasets, i.e., PACS [11] and Office-Home [39], and a real-world medical image dataset, Camelyon17 [2], consisting of various sub-datasets", + "bbox": [ + 75, + 839, + 470, + 901 + ], + "page_idx": 4 + }, + { + "type": "code", + "sub_type": "algorithm", + "code_caption": [ + "Algorithm 1 gPerXAN" + ], + "code_body": "1: Input: A model $f$ uses PerXAN to replace BN layers in the feature extractor. $N$ clients with their local datasets $\\{D_i\\}_{i = 1}^N$ . Notably, $f^{(t)}$ and $f^{(l;t)}$ is the model $f$ and its $l^{th}$ layer at the communication round $t$ , respectively. \n2: Initialization: $f_{g}^{(0)}\\gets f$ \n3: for each round $t = 1,2,3,\\ldots ,T$ do \n4: for each client $i = 1,2,3,\\ldots ,N$ in parallel do \n5: for each layer in $f_{i}^{(t)}$ do \n6: if $f_{i}^{(l;t)}$ is not BN then \n7: $f_{i}^{(l;t)}\\gets f_{g}^{(l;t)}$ \n8: end if \n9: end for \n10: $f_{i}^{(t)}\\gets \\mathrm{LocalTraining}(f_{i}^{(t)},h_{g};D_{i})$ \n11: end for \n12: $f_{g}^{(t)} = \\sum_{i = 1}^{N}\\frac{|D_{i}|}{\\sum_{1\\leq j\\leq N}|D_{j}|} f_{i}^{(t)}$ // Eqn (2) \n13: end for \n14: return: $f_{g}^{(T)}$ \n15: LocalTraining $(f_i,h_g;D_i)$ \n16: for each epoch $e = 1,2,3,\\dots ,E$ do \n17: $L_{i} = L_{cls}^{(i)}(f_{i};D_{i}) + \\lambda L_{reg}^{(i)}(g_{i},h_{g};D_{i})$ // Eqn (3) \n18: $f_{i}\\gets f_{i} - \\eta \\nabla L_{i}$ \n19: end for \n20: return: $f_{i}$", + "bbox": [ + 506, + 109, + 893, + 520 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "that are considered as domains. Specifically, PACS is composed of 4 domains with large discrepancies from diverse image colors and textures, *Photo* (P), *Art painting* (A), *Cartoon* (C), and *Sketch* (S). Each domain contains 7 categories, with 9,991 images in total. Office-Home is also composed of 4 domains but with smaller discrepancies from various backgrounds and camera viewpoints, *Product* (P), *Art* (A), *Clipart* (C), and *Real-world* (R). Each domain contains a more extensive label set of 65 categories with 15,588 images. Camelyon17 is a binary tumor classification dataset containing 455,964 histology images with stains from 5 different hospitals worldwide. These datasets are at different difficulty levels and are commonly used in the literature.", + "bbox": [ + 496, + 547, + 893, + 746 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Experimental Settings", + "text_level": 1, + "bbox": [ + 498, + 755, + 707, + 772 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Evaluation. For a fair comparison purpose, we follow the common leave-one-domain-out evaluation protocol as considered in [3, 18, 46]. In particular, we sequentially choose one domain as the unseen domain, train the model on all remaining domains where a single domain is treated as a client, and evaluate the trained model on the chosen domain. For PACS and Office-Home datasets, we split $90\\%$ of the data of each source client as the training set and $10\\%$", + "bbox": [ + 496, + 779, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "6031", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/2b3d2ddbec3e495913dc3a46b19ec0add0a8306b26dde5910d05d9d70166c284.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodPACSOffice-Home
PACSAvgPACRAvgAvg
FedAvg (Baseline)95.2182.2378.2073.5682.3076.5365.9755.4078.0168.9875.64
FedAvg w/ MixStyle95.9385.9980.0375.4684.3575.8762.0957.9277.4868.3476.35
FedAvg w/ RSC95.2183.1578.2474.6282.8175.2662.3450.7977.4666.4674.63
ELCFS96.2383.9479.2773.3083.1976.8366.3255.6378.1269.2376.21
CCST96.6588.3378.2082.9086.5276.6166.3552.3980.0168.8477.68
COPA95.6284.8080.2882.8685.8975.8262.2756.0478.7268.2177.05
FedDG-GA96.8086.9181.2382.7486.9277.2365.1058.2978.8069.8678.42
gPerXAN (Ours)97.2786.5284.6883.2887.9478.9167.2457.7580.1571.0179.48
", + "bbox": [ + 76, + 88, + 895, + 308 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "of that as the validation set, while for unseen clients, the entire data is used for testing. For the large-scale Camelyon17 dataset, the ratios of training set and validation set at each source client are $80\\%$ and $20\\%$ , respectively, while the entire data is used for testing at unseen clients. In all experiments, we report the test accuracy on each unseen client by using the best validation model selected based on the average of accuracies on validation sets of source clients. The reported numerical values are averaged over 3 runs. Finally, we straightforwardly compare the proposed method with the vanilla FedAvg [21]. Two centralized DG methods, which are free from the requirement of data centralization, MixStyle [47] and RSC [10], are also evaluated under the integration into the FedAvg framework. Furthermore, we directly compare our method with state-of-the-art relevant ones that address the problem of DG in the FL setting as discussed in Section 2, including ELCFS [18], CCST [3], COPA [42], and FedDG-GA [46].", + "bbox": [ + 75, + 359, + 472, + 632 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation Details. We present architectural details and hyper-parameter values used for experiments in the paper. Following [3, 46], for PACS and Office-Home datasets, we use the image size of $224 \\times 224$ pixels and ResNet-50 [7] pre-trained on ImageNet as the backbone for the feature extractor and a linear layer as the classifier. BN layers in the first four and first two blocks in feature extractors are replaced with our PerXAN, respectively. For the Camelyon17 dataset, input images are resized to $96 \\times 96$ pixels, and DenseNet-121 [8] is used instead of ResNet-50, BN layers in the first dense block and the following transition block are replaced by PerXAN. Notably, parameters related to BN layers are initialized with ImageNet pre-trained weights. Client models in all experiments are optimized by an SGD optimizer with a learning rate of 2e-3 for 100 communication rounds (i.e. $T = 100$ ) with one local update epoch at clients (i.e. $E = 1$ ). During the training, simple", + "bbox": [ + 75, + 643, + 472, + 902 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/4abe91d7a862f8464364f1f58953e4df906ae73ff2976571738eec08275df989.jpg", + "table_caption": [ + "Table 2. Accuracy comparison on the PACS and Office-Home datasets in the leave-one-domain-out setting." + ], + "table_footnote": [], + "table_body": "
MethodCamelyon17
H1H2H3H4H5Avg
FedAvg (Baseline)97.091.889.994.281.090.8
FedAvg w/ MixStyle91.185.586.293.387.988.8
FedAvg w/ RSC90.690.688.394.593.391.5
ELCFS92.990.689.993.289.991.3
CCST91.590.287.394.691.691.0
COPA93.290.992.293.690.292.0
FedDG-GA97.290.791.092.390.592.3
gPerXAN (Ours)96.592.295.194.791.994.1
", + "bbox": [ + 498, + 356, + 895, + 575 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 3. Accuracy comparison on the Camelyon17 dataset.", + "bbox": [ + 519, + 587, + 870, + 602 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "data augmentation techniques are applied including random horizontal flipping and color jittering. The hyper-parameter of our regularization term $\\lambda$ is searched in the range [0, 1] with a step of 0.25. The MixStyle [47] and RSC [10] can be directly integrated into the FedAvg without further modifications. All settings of other compared methods, i.e., EL-CFS [18], CCST [3], COPA [42], and FedDG-GA [46] are chosen based on corresponding papers.", + "bbox": [ + 496, + 627, + 893, + 750 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Main Results", + "text_level": 1, + "bbox": [ + 498, + 756, + 638, + 771 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 2 presents the quantitative results for different testing domains on the PACS and Office-Home datasets. Particularly, each result column shows the test accuracy of the global model on the domain of the column name. First of all, we can empirically verify that centralized DG methods such as MixStyle [47] and RSC [10] are inconsistent under the FedAvg framework and even harmful in certain cases compared to the baseline. Since these methods are", + "bbox": [ + 496, + 779, + 893, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6032", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "designed for the centralized setting, they might need access to inter-domain knowledge to perform well, which is unavailable in the FL setting. In comparing state-of-the-art FedDG methods designed for the FL scenario, methods that share partial client data information, i.e., ELCFS [18] and CCST [3], show impressive results on the PACS dataset. However, they do not significantly affect the Office-Home dataset. A similar pattern is found with the architectural method COPA [42] and FedDG-GA [46]. Meanwhile, on the PACS and Office-Home datasets, our proposed method, gPerXAN achieves average accuracies across unseen clients of $87.94\\%$ and $71.01\\%$ , which are $1.02\\%$ and $1.15\\%$ better than the second-best ones, respectively. Although most of the considered methods can perform better than the baseline FedAvg in most cases, our method demonstrates a significant boost over others on both two standard benchmarks.", + "bbox": [ + 75, + 90, + 472, + 330 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In addition to standard PACS and Office-Home datasets, we further evaluate our proposed method on a medical imaging dataset Camelyon17 as presented in Table 3. On this real-world benchmark, we can observe that ELCFS [18] and CCST [3] show considerably inferior performance compared to architectural ones, i.e., COPA [42] and gPerXAN. This might be due to sophisticated and sensitive features in medical images that are more challenging to extract and interpolate than conventional datasets [38], leading to inefficiency in these information-sharing-based methods. Notably, although it yields impressive performance, COPA [42] involves several other advanced techniques, such as RandAugment [4], making the performance gain hard to justify. Meanwhile, our method achieves an average accuracy across unseen clients of $94.1\\%$ , outperforming FedDG-GA [46] by approximately $2\\%$ . In general, the above experimental results demonstrate the effectiveness of the proposed method across various applications.", + "bbox": [ + 75, + 333, + 472, + 604 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Ablation Studies", + "text_level": 1, + "bbox": [ + 76, + 617, + 243, + 633 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We conduct ablation studies on the PACS dataset to investigate the impact of each building component. Specifically, we first compare PerXAN with other centralized variants such as conventional BN, I-BN [25], and DSON [30] under the same implementation details, and then our guiding regularizer is applied to ELCFS [18] and CCST [3] to verify its compatibility with the rest of the method.", + "bbox": [ + 75, + 641, + 468, + 748 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4.1 Impact of the Normalization Scheme", + "text_level": 1, + "bbox": [ + 76, + 768, + 387, + 786 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 4 provides a quantitative comparison among considered normalization schemes where gFedAvg represents the FedAvg framework implemented with the guiding regularizer. This means that we sequentially utilize the conventional BN, I-BN [25], and DSON [30] to replace the PerXAN scheme in our proposed method. From this table, we can observe that I-BN [25] can yield better perfor", + "bbox": [ + 75, + 795, + 468, + 902 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/c248505276b24f47d6544d62d2057bf1e1137e6ea902898906b32c03ad4bb8a4.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodPACS
PACSAvg
gFedAvg w/ BN95.8584.3978.0476.4283.68
gFedAvg w/ I-BN94.6782.0880.4680.0784.32
gFedAvg w/ DSON93.7781.6479.9180.6383.99
gPerXAN (Ours)97.2786.5284.6883.2887.94
", + "bbox": [ + 498, + 88, + 895, + 224 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/94f9c628159021288fa16c579adab956b87c84502f575bb30bf9319039b2937d.jpg", + "image_caption": [ + "Table 4. Evaluation of different normalization schemes.", + "Figure 2. Contribution of the regularizer on different methods." + ], + "image_footnote": [], + "bbox": [ + 500, + 262, + 895, + 415 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "mance than DSON [30] and the conventional BN, which are slightly comparable to each other. Meanwhile, PerXAN shows optimal performance with a significant margin compared to others and largely contributes to the whole method.", + "bbox": [ + 498, + 464, + 893, + 527 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4.2 Impact of the Regularizer", + "text_level": 1, + "bbox": [ + 500, + 542, + 735, + 559 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To better understand the proposed regularization term's contribution, we sequentially employ it in FedAvg, ELCFS [18], and CCST [3] to observe performance changes. Figure 2 displays averaged accuracies across unseen clients of these methods without and with the involvement of our regularizer. It is straightforward to see that while regularizer improves the performance of FedAvg and PerXAN considerably, it does not enhance ELCFS [18] and CCST [3]. In information-sharing-based methods, i.e., ELCFS [18] and CCST [3], each client is exposed to other clients' data information, which means that clients can access global knowledge relatively, then making the effect of matching global knowledge to clients' knowledge redundant and harmful. Also, this ablation study empirically verifies the connection between the regularizer and normalization scheme.", + "bbox": [ + 496, + 566, + 893, + 794 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5. Analysis and Discussion", + "text_level": 1, + "bbox": [ + 500, + 806, + 725, + 824 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.1. Visualization", + "text_level": 1, + "bbox": [ + 500, + 832, + 638, + 847 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We now provide an in-depth ability analysis of the proposed method via representation visualization using t-SNE [37], a common dimensionality reduction technique. As shown in", + "bbox": [ + 498, + 854, + 893, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "6033", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ab506ec2940b2bee47a0333b246301000e144a6f2b9ba3cf857220fed241c919.jpg", + "image_caption": [ + "Figure 3. Visualization of representation extracted from the global model on the PACS dataset." + ], + "image_footnote": [], + "bbox": [ + 101, + 90, + 410, + 305 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/58675cc41bdd3feae2797f6bfb79fc42ce236980c17ffcee28fb3c14d4fee8af.jpg", + "image_caption": [ + "Figure 3, we compare the representations extracted from the global model obtained from our gPerXAN and the baseline FedAvg on the PACS dataset with the testing domain, a.k.a. target domain Photo (P). From this figure, we can observe that features derived from our method are semantically separated according to 7 categories in the PACS dataset for both source domains and the target domain. Moreover, features of each category on all domains tend to be close and grouped, demonstrating the ability of our method to distill more discriminative and domain-invariant representation, leading to a significant improvement of more than $2\\%$ in classification accuracy on the testing domain Photo (P)." + ], + "image_footnote": [], + "bbox": [ + 566, + 90, + 874, + 305 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.2. Privacy and Efficiency", + "text_level": 1, + "bbox": [ + 76, + 556, + 287, + 574 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Data privacy is a major concern in the field of FL, which is mitigated by only sharing client models' parameters instead of raw data. However, to resolve the FedDG problem, recently introduced methods sacrifice this principle by revealing partial information about client data. Specifically, leaking more information not only opens higher chances for attackers to perform an inversion attack [24], which aims to reconstruct the original data of clients, but also amplifies the risk of membership inference attack [32], which determines if a sample was in the model's training data. Furthermore, efficiency in communication and computation is critical in FL, especially in scenarios where resource-constrained clients are involved. Despite that, the common sharing mechanism in available methods introduces unacceptable extra costs to the FL training process. For COPA [42], using an ensemble of domain-specific classifiers in the model architecture results in an $O(N^2)$ increase in communication and computation complexities compared to $O(N)$ as conventional, according to the number of participating clients $N$ . Meanwhile, FedDG-GA [46] consumes a double of memory at clients. Compared with existing ones, in ad", + "bbox": [ + 75, + 583, + 472, + 902 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "dition to bypassing the limitations described above, our proposed method is more practical in terms of implementation yet provides competitive results in extensive evaluations.", + "bbox": [ + 496, + 358, + 890, + 405 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 498, + 415, + 619, + 431 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we introduce a novel architectural method, namely gPerXAN, to address the problem of FedDG. By explicitly assembling Instance Normalization layers into Batch Normalization layers in a personalized scheme and employing a simple yet effective guiding regularizer, our method allows the model to filter domain-specific features and actively distill domain-invariant representation for classification tasks. We conduct extensive experiments and in-depth analysis to quantitatively and qualitatively verify the effectiveness of our proposed method in solving this particular problem. Although our evaluation is more on cross-silo FL, our method can be easily extended to cross-device scenarios while Algorithm 1 remains unchanged. Unlike existing methods, due to the independence from imaging techniques, gPerXAN can be straightforwardly extended to diverse applications. A potential limitation of our method is its certain suitability with normalization-based models. In the future, investigating other forms of regularization terms is a promising research direction due to the available room for improvements. Moreover, the vulnerability of available methods under various attacks remains underexplored.", + "bbox": [ + 496, + 440, + 890, + 758 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgement", + "text_level": 1, + "bbox": [ + 500, + 768, + 660, + 787 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This work is supported by the Deutsche Forschungsgemeinschaft, German Research Foundation under grant number 453130567 (COSMO), by the Horizon Europe Research and Innovation Actions under grant number 101092908 (SmartEdge), and by the Federal Ministry for Education and Research, Germany under grant number 01IS18037A (BIFOLD).", + "bbox": [ + 496, + 794, + 890, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "6034", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Yogesh Balaji, Swami Sankaranarayanan, and Rama Chellappa. MetaReg: Towards Domain Generalization Using Meta-Regularization. Advances in Neural Information Processing Systems, 31, 2018. 1, 3", + "[2] Peter Bandi, Oscar Geessink, Quirine Manson, Marcory Van Dijk, Maschenka Balkenhol, Meyke Hermsen, Babak Ehteshami Bejnordi, Byungjae Lee, Kyunghyun Paeng, Aoxiao Zhong, et al. From Detection of Individual Metastases to Classification of Lymph Node Status at the Patient Level: The Camelyon17 Challenge. IEEE Transactions on Medical Imaging, 38(2):550-560, 2018. 5", + "[3] Junming Chen, Meirui Jiang, Qi Dou, and Qifeng Chen. Federated Domain Generalization for Image Recognition via Cross-Client Style Transfer. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 361–370, 2023. 2, 3, 5, 6, 7", + "[4] Ekin D Cubuk, Barret Zoph, Jonathon Shlens, and Quoc V Le. RandAugment: Practical Automated Data Augmentation With a Reduced Search Space. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, pages 702-703, 2020. 7", + "[5] Shi Dong, Ping Wang, and Khushnood Abbas. A Survey on Deep Learning and Its Applications. Computer Science Review, 40:100379, 2021. 1", + "[6] Qi Dou, Daniel Coelho de Castro, Konstantinos Kamnitsas, and Ben Glocker. Domain Generalization via Model-Agnostic Learning of Semantic Features. Advances in Neural Information Processing Systems, 32, 2019. 1", + "[7] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep Residual Learning for Image Recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 770-778, 2016. 6", + "[8] Gao Huang, Zhuang Liu, Laurens Van Der Maaten, and Kilian Q Weinberger. Densely Connected Convolutional Networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4700-4708, 2017. 6", + "[9] Xun Huang and Serge Belongie. Arbitrary Style Transfer in Real-Time With Adaptive Instance Normalization. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1501-1510, 2017. 4", + "[10] Zeyi Huang, Haohan Wang, Eric P Xing, and Dong Huang. Self-Challenging Improves Cross-Domain Generalization. In Proceedings of the European Conference on Computer Vision (ECCV), pages 124-140. Springer, 2020. 6", + "[11] Da Li, Yongxin Yang, Yi-Zhe Song, and Timothy M Hospedales. Deeper, Broader and Artier Domain Generalization. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5542-5550, 2017. 5", + "[12] Da Li, Jianshu Zhang, Yongxin Yang, Cong Liu, Yi-Zhe Song, and Timothy M Hospedales. Episodic Training for Domain Generalization. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1446-1455, 2019. 1", + "[13] Haoliang Li, Sinno Jialin Pan, Shiqi Wang, and Alex C Kot. Domain Generalization With Adversarial Feature Learning." + ], + "bbox": [ + 78, + 114, + 470, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5400-5409, 2018. 1", + "[14] Tian Li, Anit Kumar Sahu, Manzil Zaheer, Maziar Sanjabi, Ameet Talwalkar, and Virginia Smith. Federated Optimization in Heterogeneous Networks. Proceedings of Machine Learning and Systems, 2:429-450, 2020. 3", + "[15] Xiang Li, Kaixuan Huang, Wenhao Yang, Shusen Wang, and Zhihua Zhang. On the Convergence of Fedavg on Non-IID Data. In International Conference on Learning Representations, 2019. 4", + "[16] Yanghao Li, Naiyan Wang, Jianping Shi, Jiaying Liu, and Xiaodi Hou. Revisiting Batch Normalization for Practical Domain Adaptation. arXiv preprint arXiv:1603.04779, 2016.4", + "[17] Quande Liu, Qi Dou, Lequan Yu, and Pheng Ann Heng. MS-Net: Multi-Site Network for Improving Prostate Segmentation With Heterogeneous MRI Data. IEEE Transactions on Medical Imaging, 39(9):2713-2724, 2020. 4", + "[18] Quande Liu, Cheng Chen, Jing Qin, Qi Dou, and Pheng-Ann Heng. FedDG: Federated Domain Generalization on Medical Image Segmentation via Episodic Learning in Continuous Frequency Space. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1013–1023, 2021. 2, 3, 5, 6, 7", + "[19] Ping Luo, Jiamin Ren, Zhanglin Peng, Ruimao Zhang, and Jingyu Li. Differentiable Learning-To-Normalize via Switchable Normalization. In International Conference on Learning Representations, 2018. 4", + "[20] Divyat Mahajan, Shruti Tople, and Amit Sharma. Domain Generalization Using Causal Matching. In International Conference on Machine Learning, pages 7313-7324. PMLR, 2021. 1", + "[21] Brendan McMahan, Eider Moore, Daniel Ramage, Seth Hampson, and Blaise Aguera y Arcas. Communication-Efficient Learning of Deep Networks From Decentralized Data. In Artificial Intelligence and Statistics, pages 1273-1282. PMLR, 2017. 1, 4, 6", + "[22] Krikamol Muandet, David Balduzzi, and Bernhard Scholkopf. Domain Generalization via Invariant Feature Representation. In International Conference on Machine Learning, pages 10-18. PMLR, 2013. 3", + "[23] Hyeonseob Nam, HyunJae Lee, Jongchan Park, Wonjun Yoon, and Donggeun Yoo. Reducing Domain Gap by Reducing Style Bias. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8690-8699, 2021. 1", + "[24] Ngoc-Bao Nguyen, Keshigeyan Chandrasegaran, Milad Abdollahzadeh, and Ngai-Man Cheung. Re-thinking Model Inversion Attacks Against Deep Neural Networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, pages 16384-16393, 2023. 8", + "[25] Xingang Pan, Ping Luo, Jianping Shi, and Xiaou Tang. Two at Once: Enhancing Learning and Generalization Capacities via IBN-Net. In Proceedings of the European Conference on Computer Vision (ECCV), pages 464–479, 2018. 1, 2, 3, 5, 7" + ], + "bbox": [ + 503, + 92, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "6035", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[26] Xingchao Peng, Zijun Huang, Yizhe Zhu, and Kate Saenko. Federated Adversarial Domain Adaptation. In International Conference on Learning Representations, 2020. 3", + "[27] Krishna Pillutla, Kshitiz Malik, Abdel-Rahman Mohamed, Mike Rabbat, Maziar Sanjabi, and Lin Xiao. Federated Learning With Partial Model Personalization. In International Conference on Machine Learning, pages 17716-17758. PMLR, 2022. 2, 3, 4", + "[28] Joaquin Quinonero-Candela, Masashi Sugiyama, Anton Schwaighofer, and Neil D Lawrence. Dataset Shift in Machine Learning. Mit Press, 2008. 1", + "[29] Alexandre Rame, Coretin Dancette, and Matthieu Cord. Fishr: Invariant Gradient Variances for Out-Of-Distribution Generalization. In International Conference on Machine Learning, pages 18347–18377. PMLR, 2022. 1", + "[30] Seonguk Seo, Yumin Suh, Dongwan Kim, Geeho Kim, Jongwoo Han, and Bohyung Han. Learning to Optimize Domain Specific Normalization for Domain Generalization. In Proceedings of the European Conference on Computer Vision (ECCV), pages 68-83, 2020. 2, 3, 4, 5, 7", + "[31] Neta Shoham, Tomer Avidor, Aviv Keren, Nadav Israel, Daniel Benditkis, Liron Mor-Yosef, and Itai Zeitak. Overcoming Forgetting in Federated Learning on Non-IID Data. arXiv preprint arXiv:1910.07796, 2019. 4", + "[32] Reza Shokri, Marco Stronati, Congzheng Song, and Vitaly Shmatikov. membership inference attacks against machine learning models. In 2017 IEEE Symposium on Security and Privacy (SP), pages 3-18. IEEE, 2017. 8", + "[33] Baochen Sun and Kate Saenko. Deep Coral: Correlation Alignment for Deep Domain Adaptation. In Proceedings of the European Conference on Computer Vision (ECCV), pages 443-450, 2016. 1, 5", + "[34] Baochen Sun, Jiashi Feng, and Kate Saenko. Correlation Alignment for Unsupervised Domain Adaptation. Domain Adaptation in Computer Vision Applications, pages 153-171, 2017. 5", + "[35] Benyuan Sun, Hongxing Huo, Yi Yang, and Bo Bai. *PartialFed: Cross-Domain Personalized Federated Learning via Partial Initialization*. Advances in Neural Information Processing Systems, 34:23309-23320, 2021. 2, 3, 4", + "[36] Yuwei Sun, Ng Chong, and Hideya Ochiai. Feature Distribution Matching for Federated Domain Generalization. In Asian Conference on Machine Learning, pages 942-957. PMLR, 2023. 3", + "[37] Laurens Van der Maaten and Geoffrey Hinton. Visualizing Data Using t-SNE. Journal of Machine Learning Research, 9(11), 2008. 7", + "[38] Gael Varoquaux and Veronika Cheplygina. Machine Learning for Medical Imaging: Methodological Failures and Recommendations for the Future. Npj Digital Medicine, 5(1):48, 2022. 7", + "[39] Hemanth Venkateswara, Jose Eusebio, Shayok Chakraborty, and Sethuraman Panchanathan. Deep Hashing Network for Unsupervised Domain Adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5018-5027, 2017. 5", + "[40] Paul Voigt and Axel Von dem Bussche. The EU General Data Protection Regulation (GDPR). A Practical Guide, 1st Ed.," + ], + "bbox": [ + 78, + 90, + 470, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Cham: Springer International Publishing, 10(3152676):10-5555, 2017. 2", + "[41] Jindong Wang, Cuiling Lan, Chang Liu, Yidong Ouyang, Tao Qin, Wang Lu, Yiqiang Chen, Wenjun Zeng, and Philip Yu. Generalizing to Unseen Domains: A Survey on Domain Generalization. IEEE Transactions on Knowledge and Data Engineering, 2022. 1", + "[42] Guile Wu and Shaogang Gong. Collaborative Optimization and Aggregation for Decentralized Domain Generalization and Adaptation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6484-6493, 2021. 3, 5, 6, 7, 8", + "[43] Zhenlin Xu, Deyi Liu, Junlin Yang, Colin Raffel, and Marc Niethammer. Robust and Generalizable Visual Representation Learning via Random Convolutions. In International Conference on Learning Representations, 2020. 1", + "[44] Fu-En Yang, Yuan-Chia Cheng, Zu-Yun Shiau, and Yu-Chiang Frank Wang. Adversarial Teacher-Student Representation Learning for Domain Generalization. Advances in Neural Information Processing Systems, 34:19448-19460, 2021. 1", + "[45] Chun-Han Yao, Boqing Gong, Hang Qi, Yin Cui, Yukun Zhu, and Ming-Hsuan Yang. Federated Multi-Target Domain Adaptation. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1424–1433, 2022. 3", + "[46] Ruipeng Zhang, Qinwei Xu, Jiangchao Yao, Ya Zhang, Qi Tian, and Yanfeng Wang. Federated Domain Generalization with Generalization Adjustment. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3954-3963, 2023. 3, 5, 6, 7, 8", + "[47] Kaiyang Zhou, Yongxin Yang, Yu Qiao, and Tao Xiang. Domain Generalization With Mixstyle. In International Conference on Learning Representations, 2020. 1, 3, 6", + "[48] Weiming Zhuang, AI Sony, and Lingjuan Lyu. FedWon: Triumphing Multi-domain Federated Learning Without Normalization. In International Conference on Learning Representations, 2024. 4" + ], + "bbox": [ + 501, + 90, + 890, + 627 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "6036", + "bbox": [ + 482, + 945, + 516, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2024/Efficiently Assemble Normalization Layers and Regularization for Federated Domain Generalization/16ce0297-575e-48b3-acb7-4a9aa0af4fe6_model.json b/2024/Efficiently Assemble Normalization Layers and Regularization for Federated Domain Generalization/16ce0297-575e-48b3-acb7-4a9aa0af4fe6_model.json new file mode 100644 index 0000000000000000000000000000000000000000..7e1c8d2e2717750691efa5b359bbdb66dd32c15e --- /dev/null +++ b/2024/Efficiently Assemble Normalization Layers and Regularization for Federated Domain Generalization/16ce0297-575e-48b3-acb7-4a9aa0af4fe6_model.json @@ -0,0 +1,1881 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.153, + 0.131, + 0.82, + 0.175 + ], + "angle": 0, + "content": "Efficiently Assemble Normalization Layers and Regularization for Federated Domain Generalization" + }, + { + "type": "text", + "bbox": [ + 0.132, + 0.203, + 0.842, + 0.276 + ], + "angle": 0, + "content": "Khiem Le\\(^{1}\\), Long Ho\\(^{2}\\), Cuong Do\\(^{2}\\), Danh Le-Phuoc\\(^{3}\\), Kok-Seng Wong\\(^{2,*}\\) \n\\(^{1}\\) Department of Computer Science and Engineering, University of Notre Dame, IN, USA \n\\(^{2}\\) College of Engineering and Computer Science, VinUniversity, Hanoi, Vietnam \n\\(^{3}\\) Open Distributed Systems, Technical University Berlin, Berlin, Germany" + }, + { + "type": "text", + "bbox": [ + 0.12, + 0.278, + 0.844, + 0.293 + ], + "angle": 0, + "content": "kle3@nd.edu, dans.lephuoc@tu-berlin.de, {long.ht, cuong.dd, wong.ks}@vinuni.edu.vn" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.327, + 0.314, + 0.344 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.359, + 0.474, + 0.738 + ], + "angle": 0, + "content": "Domain shift is a formidable issue in Machine Learning that causes a model to suffer from performance degradation when tested on unseen domains. Federated Domain Generalization (FedDG) attempts to train a global model using collaborative clients in a privacy-preserving manner that can generalize well to unseen clients possibly with domain shift. However, most existing FedDG methods either cause additional privacy risks of data leakage or induce significant costs in client communication and computation, which are major concerns in the Federated Learning paradigm. To circumvent these challenges, here we introduce a novel architectural method for FedDG, namely gPerXAN1, which relies on a normalization scheme working with a guiding regularizer. In particular, we carefully design Personalized eXplicitly Assessed Normalization to enforce client models selectively filtering domain-specific features that are biased towards local data while retaining discrimination of those features. Then, we incorporate a simple yet effective regularizer to guide these models in directly capturing domain-invariant representations that the global model's classifier can leverage. Extensive experimental results on two benchmark datasets, i.e., PACS and Office-Home, and a real-world medical dataset, Camelyon17, indicate that our proposed method outperforms other existing methods in addressing this particular problem." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.765, + 0.21, + 0.781 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.79, + 0.47, + 0.867 + ], + "angle": 0, + "content": "Over the past few decades, Machine Learning (ML) has demonstrated remarkable achievements across diverse areas such as Computer Vision, Natural Language and Speech Processing, or Robotics [5]. In general, most ML models rely on an over-simplified assumption, i.e., the train-" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.329, + 0.895, + 0.526 + ], + "angle": 0, + "content": "ing and testing data are independent and identically distributed, which does not always reflect real-world practices. In practical scenarios where the distribution of testing data diverges from that of training data, the performance of ML models often drops catastrophically due to the domain shift issue [28]. Additionally, obtaining or identifying the testing data before model deployment can be challenging in numerous applications. For instance, in biomedical applications where data characteristics vary across different equipment and institutions, gathering data from all potential domains in advance is impractical. Therefore, it is essential to have a solution that can improve the generalization capability of such ML models to adapt effectively to unseen domains." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.532, + 0.895, + 0.865 + ], + "angle": 0, + "content": "Domain Generalization (DG) has been proposed to address the challenge of training ML models using data from single or multiple source domains with the expectation that these models will perform well on unseen domains [41]. The majority of existing DG methods fall under the category of domain-invariant representation learning approach [13, 20, 25, 29, 33]. This approach relies on a broadly acknowledged assumption that each domain contains its own domain-specific features, which are biased towards spurious relations in the data, and that all domains share domain-invariant features, which are general and robust to any unseen domains. From this assumption, previous works propose methods that remove domain-specific features and distill domain-invariant features to achieve the generalization ability. Alternative approaches for DG encompass data augmentation [23, 43, 44, 47], which involves exposing models to artificially generated domains, and meta-learning [1, 6, 12], an approach that emulates the domain shift during the meta-training phase. However, most methods require a centralized setting where all source domains are collected together. Consequently, these methods cannot be readily expanded to decentralized settings." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Federated Learning (FL) [21] is an emerging decentralized learning paradigm widely adopted in various applica" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.875, + 0.393, + 0.888 + ], + "angle": 0, + "content": "*Corresponding author: wong.rs@vinuni.edu.vn" + }, + { + "type": "page_footnote", + "bbox": [ + 0.097, + 0.888, + 0.316, + 0.9 + ], + "angle": 0, + "content": "\\(^{1}\\)https://github.com/lhkhiem28/gPerXAN" + }, + { + "type": "list", + "bbox": [ + 0.096, + 0.875, + 0.393, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6027" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.084, + 0.088, + 0.898, + 0.366 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.372, + 0.754, + 0.387 + ], + "angle": 0, + "content": "Figure 1. An overview of our proposed gPerXAN method for solving the FedDG problem." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.413, + 0.473, + 0.776 + ], + "angle": 0, + "content": "tions to cope with the increasing privacy concerns of data centralization [40]. Specifically, the paradigm works in a way that each client learns from their data and only aggregates local models' parameters at a certain frequency at the central server to generate a global model. Notably, all data samples are kept within each client during the FL training process. Due to the nature of data decentralization, where each client owns a single source domain, as illustrated in Figure 1, the FL paradigm poses further significant challenges for DG and limits the applicability of available DG methods. There have been some early attempts to address the DG problem in the FL scenario. For instance, Liu et al. [18] introduces a method that allows clients to share their image data in the frequency space with each other, thus relatively recovering the centralization process at each client. Similarly, Chen et al. [3] introduces another method that extracts and exchanges the style of local images among all clients. It is evident that these initial efforts employ a strategy that necessitates the sharing of partial client data, thereby compromising the data privacy constraints of FL to a certain extent. Although they show promising results, these methods can be overly complicated to implement in practice and lead to additional privacy risks during the FL training process." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.78, + 0.471, + 0.902 + ], + "angle": 0, + "content": "To address the aforementioned challenges, this paper introduces a novel architectural method for domain-invariant representation learning within the FL framework. The proposed method enhances the generalization ability while upholding the fundamental privacy principles of FL. Based on the effectiveness of discarding domain-specific information from learned features [25, 30], we properly assemble Instance Normalization layers (IN) into Batch Normaliza" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.413, + 0.895, + 0.656 + ], + "angle": 0, + "content": "tion layers (BN) in well-known Convolutional Neural Networks (CNNs) using an explicit differential mixture as in Eqn (2). Moreover, thanks to the explicit property, the benefit of personalization in FL [27, 35] can be incorporated into the normalization scheme using local BN sides. Specifically, during the FL training process, while IN sides are globally aggregated along with other model parameters, BN sides are updated locally without broadcasting. In addition, we argue that only relying on the ability to filter domain-specific features of IN while lacking guidance to distill domain-invariant representations directly might lead to suboptimal performance. Based on this observation, we introduce a simple yet highly effective regularization term to guide client models to directly capture domain-invariant representations that can be used by the global model's classifier, which is aggregated from client models' classifiers." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.664, + 0.892, + 0.695 + ], + "angle": 0, + "content": "To summarize, our main contributions in this paper are highlighted as follows:" + }, + { + "type": "text", + "bbox": [ + 0.506, + 0.704, + 0.892, + 0.78 + ], + "angle": 0, + "content": "- Different from existing methods for DG in the FL scenario, we propose a novel method that concentrates on a personalized normalization scheme, global IN while local BN, for filtering domain-specific features and fully respecting the privacy-preserving principles of FL." + }, + { + "type": "text", + "bbox": [ + 0.506, + 0.781, + 0.892, + 0.841 + ], + "angle": 0, + "content": "- Furthermore, we propose a simple yet effective regularization term to introduce clear guidance to client models for directly capturing domain-invariant representations, further improving performance on unseen domains." + }, + { + "type": "text", + "bbox": [ + 0.506, + 0.841, + 0.892, + 0.901 + ], + "angle": 0, + "content": "- Finally, we conduct extensive experiments on two benchmark datasets, i.e., PACS and Office-Home, and a real-world medical dataset, Camelyon17, where our proposed method outperforms existing relevant ones." + }, + { + "type": "list", + "bbox": [ + 0.506, + 0.704, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6028" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.078, + 0.09, + 0.22, + 0.106 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.115, + 0.294, + 0.131 + ], + "angle": 0, + "content": "2.1. Domain Generalization" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.14, + 0.473, + 0.534 + ], + "angle": 0, + "content": "Domain Generalization is a challenging task requiring models to perform well on unseen domains. One dominant approach is domain-invariant representation learning, which aims to minimize the discrepancy among source domains, assuming that the resulting representation will be domain-invariant and generalize well on testing domains. Along this track, Muandet et al. [22] proposes to reduce the domain dissimilarity by using Maximum Mean Discrepancy. Pan et al. [25] combines IN and BN in popular CNNs to reap the benefits of removing domain-specific features while maintaining the ability to capture discriminative features. Meanwhile, Seo et al. [30] discloses that combining these normalization layers in a switchable mechanism can yield better performance. Another approach to DG is data augmentation, which augments source domains to a broader span of the training data space, enlarging the possibility of covering the span of the data in the testing domain. Zhou et al. [47] mixes styles of training instances, resulting in novel domains being synthesized implicitly. Recently, the meta-learning approach has drawn increasing attention from the DG community. Balaji et al. [1] proposes MetaReg that learns a regularization function, particularly for the network classification layer, while excluding the feature extractor. However, the majority of these methods rely on having access to a diverse set of source domains, making them not applicable to decentralized settings." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.545, + 0.377, + 0.56 + ], + "angle": 0, + "content": "2.2. Federated Domain Generalization" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.569, + 0.473, + 0.903 + ], + "angle": 0, + "content": "While there is a growing interest in addressing the DG problem within the FL framework, particularly in scenarios where each client possesses a single source domain, the existing literature on this subject remains relatively sparse. This FedDG problem was initially introduced by Liu et al. [18], who then proposed ELCFS that involves exchanging amplitude information in the frequency space among clients and leveraging episodic learning to further enhance performance. Chen et al. [3] proposes a similar mechanism CCST that extracts and exchanges the overall domain style of local images among all clients. Unfortunately, these early works require sharing partial information about the local data, which can be viewed as a form of data leakage and is undesirable in the FL setting. Moreover, these methods include broadcasting partial information and interpolating this information into new training data, which adds significant extra cost to communication and computation during the FL training process. Besides, Wu and Gong [42] proposes an architectural method COPA in which the global model consists of a domain-invariant representation extractor and an ensemble of domain-specific classifiers. Sun et al. [36] introduces FedKA that employs a server-side voting mech" + }, + { + "type": "table", + "bbox": [ + 0.499, + 0.089, + 0.895, + 0.214 + ], + "angle": 0, + "content": "
MethodPrivacy RiskAdditional Cost
CommunicationComputation
ELCFS✓✓✓
CCST✓✓✓
COPA×✓✓
FedDG-GA××✓
gPerXAN (Ours)×××
" + }, + { + "type": "table_caption", + "bbox": [ + 0.528, + 0.224, + 0.864, + 0.238 + ], + "angle": 0, + "content": "Table 1. An advantage comparison of different methods." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.265, + 0.895, + 0.505 + ], + "angle": 0, + "content": "anism that generates target domain pseudo-labels based on the consensus from clients to facilitate global model finetuning. Recently, Zhang et al. [46] proposes a novel global objective incorporating a variance reduction regularizer to encourage fairness, and then FedDG-GA is proposed to optimize this objective by dynamically calibrating the aggregation weights. Although relaxing from the data leakage issue, these methods also cause a large additional consumption of resources when the number of source domains and output classes increases. In this paper, we introduce an alternative architectural method that shares only the model update information during training. This ensures maximal data privacy and circumvents the communication and computation overhead issue mentioned earlier while achieving competitive results. In Table 1, we highlight the advantages of our method in comparison to previous ones." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.506, + 0.895, + 0.643 + ], + "angle": 0, + "content": "One related research field to FedDG is Federated Domain Adaptation (FedDA). While both fields aim to maximize the model performance on unseen domains using existing source domains, FedDA can access the target domains while FedDG cannot see those data during training. Leveraging this assumption, FedDA methods such as FADA [26] or FMTDA [45], are typically able to align target and source features. However, this specific assumption also makes FedDA methods are inapplicable to FedDG." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.651, + 0.792, + 0.667 + ], + "angle": 0, + "content": "2.3. Personalized Federated Learning" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.675, + 0.895, + 0.901 + ], + "angle": 0, + "content": "Another related and orthogonal line of work is Personalized Federated Learning (pFL). pFL aims to learn personalized models for different clients to tackle distribution shifts across client data, known as data heterogeneity. Li et al. [14] utilizes a regularization term in the local loss function so that the clients' trained models will not significantly differ from the global model. On the other side, Sun et al. [35] and Pillutla et al. [27] mitigate the heterogeneous distributions by only loading a subset of the global model's parameters rather than loading the entire model at each training round. Although related in terms of overcoming distribution shifts across clients, pFL focuses on improving the performance of participating clients, whereas our considered field FedDG focuses on improving the performance of unseen clients with unseen domains." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6029" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.078, + 0.09, + 0.212, + 0.108 + ], + "angle": 0, + "content": "3. Methodology" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.115, + 0.47, + 0.244 + ], + "angle": 0, + "content": "Problem Formulation. First, we denote \\( X \\) and \\( Y \\) as the input space and the label space, respectively, of a specific task \\( \\mathcal{T} \\). In the standard FL setting, \\( N \\) clients \\( \\{c_i\\}_{i=1}^N \\) are involved in collaboratively constructing a global model \\( f_g \\) for solving the task, where each client \\( c_i \\) owns a dataset \\( D_i = \\{(x_k, y_k)\\}_{k=1,|D_i|} \\) which is associated to a specific domain defined by a joint distribution \\( P_{X,Y}^{(i)} \\). Importantly, scattered datasets \\( \\{D_i\\}_{i=1}^N \\) across clients satisfy:" + }, + { + "type": "text", + "bbox": [ + 0.08, + 0.248, + 0.384, + 0.268 + ], + "angle": 0, + "content": "- \\( P_{X,Y}^{(i)} \\neq P_{X,Y}^{(j)} \\) with \\( 1 \\leq i, j \\leq N \\) and \\( i \\neq j \\)" + }, + { + "type": "text", + "bbox": [ + 0.08, + 0.277, + 0.385, + 0.299 + ], + "angle": 0, + "content": "- \\( P_{Y|X}^{(i)} = P_{Y|X}^{(j)} \\) with \\( 1 \\leq i, j \\leq N \\) and \\( i \\neq j \\)." + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.248, + 0.385, + 0.299 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.302, + 0.469, + 0.348 + ], + "angle": 0, + "content": "The objective of FedDG is leveraging scattered datasets to construct a model \\( f_{g} \\) that can directly generalize to the unseen dataset \\( D_{U} \\) with an unseen domain, which means:" + }, + { + "type": "text", + "bbox": [ + 0.08, + 0.35, + 0.303, + 0.371 + ], + "angle": 0, + "content": "- \\( P_{X,Y}^{(U)} \\neq P_{X,Y}^{(i)} \\) with \\( 1 \\leq i \\leq N \\)" + }, + { + "type": "text", + "bbox": [ + 0.08, + 0.379, + 0.303, + 0.401 + ], + "angle": 0, + "content": "- \\( P_{Y|X}^{(U)} = P_{Y|X}^{(i)} \\) with \\( 1 \\leq i \\leq N \\)" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.35, + 0.303, + 0.401 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.405, + 0.469, + 0.54 + ], + "angle": 0, + "content": "To this end, \\(N\\) clients communicate with a central server for \\(T\\) rounds. At each round, every client \\(c_{i}\\) receives the same global model \\(f_{g}\\) from the server and updates \\(f_{g}\\) with their local dataset \\(D_{i}\\) for \\(E\\) epochs to establish its local model \\(f_{i}\\). The server then collects all trained models and aggregates them to update the global model. This process repeats until the global model converges. In this work, we consider the most popular FL framework, FedAvg [21], which aggregates client models as:" + }, + { + "type": "equation", + "bbox": [ + 0.182, + 0.551, + 0.469, + 0.592 + ], + "angle": 0, + "content": "\\[\nf _ {g} = \\sum_ {i = 1} ^ {N} \\frac {\\left| D _ {i} \\right|}{\\sum_ {1 \\leq j \\leq N} \\left| D _ {j} \\right|} f _ {i} \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.601, + 0.469, + 0.753 + ], + "angle": 0, + "content": "Challenges. In the general spirit of DG, a model is expected to extensively explore multiple source domains to achieve domain-invariance in its learned latent representation. However, under the FL setting, each client is restricted to accessing only its own local data, which constrains to make full use of source domains and, consequently, limits learning of generalizable representation. Moreover, sharing data or even partial information about data among clients poses additional privacy risks and introduces communication and computation costs during the FL training process." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.761, + 0.288, + 0.777 + ], + "angle": 0, + "content": "3.1. Normalization Scheme" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.785, + 0.469, + 0.845 + ], + "angle": 0, + "content": "eXplicitly Assessed Normalization. The first step towards solving the highlighted challenges, we introduce a novel normalization scheme called eXplicitly Assessed Normalization (XAN), which combines IN and BN as:" + }, + { + "type": "equation", + "bbox": [ + 0.077, + 0.854, + 0.469, + 0.901 + ], + "angle": 0, + "content": "\\[\n\\hat {h} = w _ {i n} \\left(\\gamma_ {i n} \\frac {h - \\mu_ {i n}}{\\sqrt {\\sigma_ {i n} ^ {2} + \\epsilon}} + \\beta_ {i n}\\right) + w _ {b n} \\left(\\gamma_ {b n} \\frac {h - \\mu_ {b n}}{\\sqrt {\\sigma_ {b n} ^ {2} + \\epsilon}} + \\beta_ {b n}\\right), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.09, + 0.892, + 0.182 + ], + "angle": 0, + "content": "where \\(h, \\hat{h} \\in \\mathbb{R}^{B \\times C \\times W \\times H}\\) are layer input and output activations, which are 4D tensors with dimensions of batch size \\(B\\), number of channels \\(C\\), width \\(W\\) and height \\(H\\). Meanwhile, \\(\\mu\\) and \\(\\sigma^2\\) are means and variances captured by the normalization layers, respectively, \\(\\gamma\\) and \\(\\beta\\) are affine parameters, and \\(\\epsilon\\) is for numerical stability." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.183, + 0.893, + 0.47 + ], + "angle": 0, + "content": "In essence, XAN is a sophisticated mixture mechanism of IN and BN to replace BN in the feature extractors of CNNs. In particular, \\( w_{in} \\) and \\( w_{bn} \\) are ratios to weight the mixture that allow the model to switch between IN and BN. These parameters are randomly initialized and optimized along with other model parameters during training in an end-to-end manner. Unlike previous works [19, 30], which uses an implicit mechanism to combine computed statistics of IN and BN, i.e., means and variances, XAN uses an explicit mechanism that combines the output activations of two normalization layers, this provides the model with a unique ability to separate IN from BN completely. Note that IN has shown great success in style transfer tasks [9] as it allows discarding the variability of visual styles such as object colors or textures from the content of images. This property makes IN beneficial for solving the DG problem. However, directly using IN to replace the conventional BN leads to the loss of discrimination in the learned features, resulting in performance degradation in classification tasks." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.472, + 0.893, + 0.819 + ], + "angle": 0, + "content": "Personalized eXplicitly Assessed Normalization. To accomplish our proposed normalization scheme, Personalized eXplicitly Assessed Normalization (PerXAN), we make a subtle modification to the conventional FedAvg [21] framework based on the explicit property of the above XAN. Specifically, during the FL training process, while IN sides of XAN layers are globally aggregated along with other model parameters, BN sides are updated locally, which means that parameters of BN sides are excluded from the broadcasting steps from the server. Notably, in inference time, the global model is generated by averaging all model parameters of clients. Our modification is motivated by a common observation that FedAvg normally results in poor convergence and performance in the presence of domain heterogeneity across clients [15, 48] with the major reason being that client models forget the acquired knowledge from previous rounds after aggregated [31]. The leading solution is to personalize a subset of the model, which benefits clients in learning better from their local data [27, 35]. Moreover, based on another finding that BN plays a crucial role in dealing with the domain shift issue in the centralized paradigm [16, 17], BN sides in XAN layers are finally made to be personalized. PerXAN is illustrated in Figure 1." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.831, + 0.749, + 0.847 + ], + "angle": 0, + "content": "3.2. Regularization as Guidance" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.856, + 0.892, + 0.901 + ], + "angle": 0, + "content": "With the proposed normalization scheme PerXAN in place, now, we are ready to introduce our \"Regularization as Guidance\". It targets to induce clear guidance to client" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6030" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.091, + 0.473, + 0.381 + ], + "angle": 0, + "content": "models through a regularizer, which brings an effect of domain alignment [33, 34], and guides these models to capture domain-invariant representations directly. We hypothesize that this will address a drawback of current DG methods that are later verified by our experiments in Section 4. The drawback is from DG methods that take advantage of the IN's function [25, 30] but do not actually equip the model with the capability of capturing domain-invariant features even if they demonstrated promising performance. Instead, it is only hoped that domain-invariant features would be distilled through achieving the goal of removing domain-specific features. This indirect learning purpose might affect the model's learning efficacy, especially in the FL setting, where each client only owns a separate single source domain. Specifically, we assume a classification model \\( f \\), either the global model \\( f_{g} \\) or a client model \\( f_{i} \\), comprises a feature extractor \\( g \\) and a classifier head \\( h \\). At each client \\( c_{i} \\), during local training, the client model \\( f_{i} \\) is optimized on the local dataset \\( D_{i} \\) using the following loss function:" + }, + { + "type": "equation", + "bbox": [ + 0.095, + 0.391, + 0.47, + 0.458 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} L _ {i} = L _ {c l s} ^ {(i)} \\left(f _ {i}; D _ {i}\\right) + \\lambda L _ {r e g} ^ {(i)} \\left(g _ {i}, h _ {g}; D _ {i}\\right) (3) \\\\ = \\sum_ {k = 1} ^ {| D _ {i} |} \\ell \\left(f _ {i} \\left(x _ {k} ^ {(i)}\\right), y _ {k} ^ {(i)}\\right) + \\lambda \\ell \\left(h _ {g} \\left(g _ {i} \\left(x _ {k} ^ {(i)}\\right)\\right), y _ {k} ^ {(i)}\\right), (4) \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.47, + 0.47, + 0.637 + ], + "angle": 0, + "content": "where \\(\\ell\\) is the base loss function, which is usually cross-entropy in classification tasks, and \\(\\lambda\\) is a hyper-parameter to control the significance of the regularizer. In a deeper look, by freezing \\(h_g\\) during local training, the auxiliary loss term \\(L_{reg}\\) forces client models to arrive at representations that can be made use for classification by the same global classifier, hence, producing an alignment effect on these representations [33, 34]. Moreover, our proposed regularizer can also be interpreted as an implicit form of matching global knowledge to clients' knowledge, which has been demonstrated to bring performance gain [3, 18] for FedDG." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.638, + 0.471, + 0.775 + ], + "angle": 0, + "content": "Due to its orthogonality, our proposed regularizer can be easily integrated into various methods, as shown in Algorithm 1. Specifically, our experiments in Section 4 verify its compatibility with the normalization scheme that justifies our motivation. Moreover, by using only the global model's classifier to regularize feature extractors, gPerXAN saves major communication and computation resources, as well as memory usage at clients compared to others, which use an ensemble of classifiers [42] or entire global model [46]." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.79, + 0.312, + 0.807 + ], + "angle": 0, + "content": "4. Experiments and Results" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.816, + 0.181, + 0.831 + ], + "angle": 0, + "content": "4.1. Datasets" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.471, + 0.902 + ], + "angle": 0, + "content": "To evaluate the proposed method, we perform experiments on two standard DG benchmark datasets, i.e., PACS [11] and Office-Home [39], and a real-world medical image dataset, Camelyon17 [2], consisting of various sub-datasets" + }, + { + "type": "code_caption", + "bbox": [ + 0.502, + 0.091, + 0.66, + 0.106 + ], + "angle": 0, + "content": "Algorithm 1 gPerXAN" + }, + { + "type": "algorithm", + "bbox": [ + 0.508, + 0.11, + 0.895, + 0.521 + ], + "angle": 0, + "content": "1: Input: A model \\(f\\) uses PerXAN to replace BN layers in the feature extractor. \\(N\\) clients with their local datasets \\(\\{D_i\\}_{i = 1}^N\\) . Notably, \\(f^{(t)}\\) and \\(f^{(l;t)}\\) is the model \\(f\\) and its \\(l^{th}\\) layer at the communication round \\(t\\) , respectively. \n2: Initialization: \\(f_{g}^{(0)}\\gets f\\) \n3: for each round \\(t = 1,2,3,\\ldots ,T\\) do \n4: for each client \\(i = 1,2,3,\\ldots ,N\\) in parallel do \n5: for each layer in \\(f_{i}^{(t)}\\) do \n6: if \\(f_{i}^{(l;t)}\\) is not BN then \n7: \\(f_{i}^{(l;t)}\\gets f_{g}^{(l;t)}\\) \n8: end if \n9: end for \n10: \\(f_{i}^{(t)}\\gets \\mathrm{LocalTraining}(f_{i}^{(t)},h_{g};D_{i})\\) \n11: end for \n12: \\(f_{g}^{(t)} = \\sum_{i = 1}^{N}\\frac{|D_{i}|}{\\sum_{1\\leq j\\leq N}|D_{j}|} f_{i}^{(t)}\\) // Eqn (2) \n13: end for \n14: return: \\(f_{g}^{(T)}\\) \n15: LocalTraining \\((f_i,h_g;D_i)\\) \n16: for each epoch \\(e = 1,2,3,\\dots ,E\\) do \n17: \\(L_{i} = L_{cls}^{(i)}(f_{i};D_{i}) + \\lambda L_{reg}^{(i)}(g_{i},h_{g};D_{i})\\) // Eqn (3) \n18: \\(f_{i}\\gets f_{i} - \\eta \\nabla L_{i}\\) \n19: end for \n20: return: \\(f_{i}\\)" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.549, + 0.895, + 0.747 + ], + "angle": 0, + "content": "that are considered as domains. Specifically, PACS is composed of 4 domains with large discrepancies from diverse image colors and textures, *Photo* (P), *Art painting* (A), *Cartoon* (C), and *Sketch* (S). Each domain contains 7 categories, with 9,991 images in total. Office-Home is also composed of 4 domains but with smaller discrepancies from various backgrounds and camera viewpoints, *Product* (P), *Art* (A), *Clipart* (C), and *Real-world* (R). Each domain contains a more extensive label set of 65 categories with 15,588 images. Camelyon17 is a binary tumor classification dataset containing 455,964 histology images with stains from 5 different hospitals worldwide. These datasets are at different difficulty levels and are commonly used in the literature." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.756, + 0.708, + 0.773 + ], + "angle": 0, + "content": "4.2. Experimental Settings" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.78, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Evaluation. For a fair comparison purpose, we follow the common leave-one-domain-out evaluation protocol as considered in [3, 18, 46]. In particular, we sequentially choose one domain as the unseen domain, train the model on all remaining domains where a single domain is treated as a client, and evaluate the trained model on the chosen domain. For PACS and Office-Home datasets, we split \\(90\\%\\) of the data of each source client as the training set and \\(10\\%\\)" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "6031" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.078, + 0.089, + 0.896, + 0.309 + ], + "angle": 0, + "content": "
MethodPACSOffice-Home
PACSAvgPACRAvgAvg
FedAvg (Baseline)95.2182.2378.2073.5682.3076.5365.9755.4078.0168.9875.64
FedAvg w/ MixStyle95.9385.9980.0375.4684.3575.8762.0957.9277.4868.3476.35
FedAvg w/ RSC95.2183.1578.2474.6282.8175.2662.3450.7977.4666.4674.63
ELCFS96.2383.9479.2773.3083.1976.8366.3255.6378.1269.2376.21
CCST96.6588.3378.2082.9086.5276.6166.3552.3980.0168.8477.68
COPA95.6284.8080.2882.8685.8975.8262.2756.0478.7268.2177.05
FedDG-GA96.8086.9181.2382.7486.9277.2365.1058.2978.8069.8678.42
gPerXAN (Ours)97.2786.5284.6883.2887.9478.9167.2457.7580.1571.0179.48
" + }, + { + "type": "table_caption", + "bbox": [ + 0.167, + 0.319, + 0.8, + 0.334 + ], + "angle": 0, + "content": "Table 2. Accuracy comparison on the PACS and Office-Home datasets in the leave-one-domain-out setting." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.36, + 0.473, + 0.633 + ], + "angle": 0, + "content": "of that as the validation set, while for unseen clients, the entire data is used for testing. For the large-scale Camelyon17 dataset, the ratios of training set and validation set at each source client are \\(80\\%\\) and \\(20\\%\\), respectively, while the entire data is used for testing at unseen clients. In all experiments, we report the test accuracy on each unseen client by using the best validation model selected based on the average of accuracies on validation sets of source clients. The reported numerical values are averaged over 3 runs. Finally, we straightforwardly compare the proposed method with the vanilla FedAvg [21]. Two centralized DG methods, which are free from the requirement of data centralization, MixStyle [47] and RSC [10], are also evaluated under the integration into the FedAvg framework. Furthermore, we directly compare our method with state-of-the-art relevant ones that address the problem of DG in the FL setting as discussed in Section 2, including ELCFS [18], CCST [3], COPA [42], and FedDG-GA [46]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.644, + 0.473, + 0.903 + ], + "angle": 0, + "content": "Implementation Details. We present architectural details and hyper-parameter values used for experiments in the paper. Following [3, 46], for PACS and Office-Home datasets, we use the image size of \\(224 \\times 224\\) pixels and ResNet-50 [7] pre-trained on ImageNet as the backbone for the feature extractor and a linear layer as the classifier. BN layers in the first four and first two blocks in feature extractors are replaced with our PerXAN, respectively. For the Camelyon17 dataset, input images are resized to \\(96 \\times 96\\) pixels, and DenseNet-121 [8] is used instead of ResNet-50, BN layers in the first dense block and the following transition block are replaced by PerXAN. Notably, parameters related to BN layers are initialized with ImageNet pre-trained weights. Client models in all experiments are optimized by an SGD optimizer with a learning rate of 2e-3 for 100 communication rounds (i.e. \\(T = 100\\)) with one local update epoch at clients (i.e. \\(E = 1\\)). During the training, simple" + }, + { + "type": "table", + "bbox": [ + 0.499, + 0.357, + 0.896, + 0.577 + ], + "angle": 0, + "content": "
MethodCamelyon17
H1H2H3H4H5Avg
FedAvg (Baseline)97.091.889.994.281.090.8
FedAvg w/ MixStyle91.185.586.293.387.988.8
FedAvg w/ RSC90.690.688.394.593.391.5
ELCFS92.990.689.993.289.991.3
CCST91.590.287.394.691.691.0
COPA93.290.992.293.690.292.0
FedDG-GA97.290.791.092.390.592.3
gPerXAN (Ours)96.592.295.194.791.994.1
" + }, + { + "type": "table_caption", + "bbox": [ + 0.52, + 0.588, + 0.871, + 0.603 + ], + "angle": 0, + "content": "Table 3. Accuracy comparison on the Camelyon17 dataset." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.628, + 0.894, + 0.75 + ], + "angle": 0, + "content": "data augmentation techniques are applied including random horizontal flipping and color jittering. The hyper-parameter of our regularization term \\(\\lambda\\) is searched in the range [0, 1] with a step of 0.25. The MixStyle [47] and RSC [10] can be directly integrated into the FedAvg without further modifications. All settings of other compared methods, i.e., EL-CFS [18], CCST [3], COPA [42], and FedDG-GA [46] are chosen based on corresponding papers." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.757, + 0.64, + 0.772 + ], + "angle": 0, + "content": "4.3. Main Results" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.78, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Table 2 presents the quantitative results for different testing domains on the PACS and Office-Home datasets. Particularly, each result column shows the test accuracy of the global model on the domain of the column name. First of all, we can empirically verify that centralized DG methods such as MixStyle [47] and RSC [10] are inconsistent under the FedAvg framework and even harmful in certain cases compared to the baseline. Since these methods are" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6032" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.332 + ], + "angle": 0, + "content": "designed for the centralized setting, they might need access to inter-domain knowledge to perform well, which is unavailable in the FL setting. In comparing state-of-the-art FedDG methods designed for the FL scenario, methods that share partial client data information, i.e., ELCFS [18] and CCST [3], show impressive results on the PACS dataset. However, they do not significantly affect the Office-Home dataset. A similar pattern is found with the architectural method COPA [42] and FedDG-GA [46]. Meanwhile, on the PACS and Office-Home datasets, our proposed method, gPerXAN achieves average accuracies across unseen clients of \\(87.94\\%\\) and \\(71.01\\%\\), which are \\(1.02\\%\\) and \\(1.15\\%\\) better than the second-best ones, respectively. Although most of the considered methods can perform better than the baseline FedAvg in most cases, our method demonstrates a significant boost over others on both two standard benchmarks." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.334, + 0.473, + 0.606 + ], + "angle": 0, + "content": "In addition to standard PACS and Office-Home datasets, we further evaluate our proposed method on a medical imaging dataset Camelyon17 as presented in Table 3. On this real-world benchmark, we can observe that ELCFS [18] and CCST [3] show considerably inferior performance compared to architectural ones, i.e., COPA [42] and gPerXAN. This might be due to sophisticated and sensitive features in medical images that are more challenging to extract and interpolate than conventional datasets [38], leading to inefficiency in these information-sharing-based methods. Notably, although it yields impressive performance, COPA [42] involves several other advanced techniques, such as RandAugment [4], making the performance gain hard to justify. Meanwhile, our method achieves an average accuracy across unseen clients of \\(94.1\\%\\), outperforming FedDG-GA [46] by approximately \\(2\\%\\). In general, the above experimental results demonstrate the effectiveness of the proposed method across various applications." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.618, + 0.245, + 0.634 + ], + "angle": 0, + "content": "4.4. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.642, + 0.47, + 0.749 + ], + "angle": 0, + "content": "We conduct ablation studies on the PACS dataset to investigate the impact of each building component. Specifically, we first compare PerXAN with other centralized variants such as conventional BN, I-BN [25], and DSON [30] under the same implementation details, and then our guiding regularizer is applied to ELCFS [18] and CCST [3] to verify its compatibility with the rest of the method." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.77, + 0.388, + 0.787 + ], + "angle": 0, + "content": "4.4.1 Impact of the Normalization Scheme" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.796, + 0.47, + 0.903 + ], + "angle": 0, + "content": "Table 4 provides a quantitative comparison among considered normalization schemes where gFedAvg represents the FedAvg framework implemented with the guiding regularizer. This means that we sequentially utilize the conventional BN, I-BN [25], and DSON [30] to replace the PerXAN scheme in our proposed method. From this table, we can observe that I-BN [25] can yield better perfor" + }, + { + "type": "table", + "bbox": [ + 0.499, + 0.089, + 0.897, + 0.226 + ], + "angle": 0, + "content": "
MethodPACS
PACSAvg
gFedAvg w/ BN95.8584.3978.0476.4283.68
gFedAvg w/ I-BN94.6782.0880.4680.0784.32
gFedAvg w/ DSON93.7781.6479.9180.6383.99
gPerXAN (Ours)97.2786.5284.6883.2887.94
" + }, + { + "type": "image_caption", + "bbox": [ + 0.53, + 0.235, + 0.862, + 0.25 + ], + "angle": 0, + "content": "Table 4. Evaluation of different normalization schemes." + }, + { + "type": "image", + "bbox": [ + 0.501, + 0.263, + 0.896, + 0.416 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.51, + 0.425, + 0.882, + 0.44 + ], + "angle": 0, + "content": "Figure 2. Contribution of the regularizer on different methods." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.465, + 0.895, + 0.529 + ], + "angle": 0, + "content": "mance than DSON [30] and the conventional BN, which are slightly comparable to each other. Meanwhile, PerXAN shows optimal performance with a significant margin compared to others and largely contributes to the whole method." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.544, + 0.736, + 0.56 + ], + "angle": 0, + "content": "4.4.2 Impact of the Regularizer" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.568, + 0.895, + 0.795 + ], + "angle": 0, + "content": "To better understand the proposed regularization term's contribution, we sequentially employ it in FedAvg, ELCFS [18], and CCST [3] to observe performance changes. Figure 2 displays averaged accuracies across unseen clients of these methods without and with the involvement of our regularizer. It is straightforward to see that while regularizer improves the performance of FedAvg and PerXAN considerably, it does not enhance ELCFS [18] and CCST [3]. In information-sharing-based methods, i.e., ELCFS [18] and CCST [3], each client is exposed to other clients' data information, which means that clients can access global knowledge relatively, then making the effect of matching global knowledge to clients' knowledge redundant and harmful. Also, this ablation study empirically verifies the connection between the regularizer and normalization scheme." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.808, + 0.727, + 0.825 + ], + "angle": 0, + "content": "5. Analysis and Discussion" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.833, + 0.64, + 0.848 + ], + "angle": 0, + "content": "5.1. Visualization" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.856, + 0.895, + 0.903 + ], + "angle": 0, + "content": "We now provide an in-depth ability analysis of the proposed method via representation visualization using t-SNE [37], a common dimensionality reduction technique. As shown in" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6033" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.102, + 0.091, + 0.411, + 0.306 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.568, + 0.091, + 0.875, + 0.306 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.204, + 0.318, + 0.766, + 0.334 + ], + "angle": 0, + "content": "Figure 3. Visualization of representation extracted from the global model on the PACS dataset." + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.359, + 0.473, + 0.542 + ], + "angle": 0, + "content": "Figure 3, we compare the representations extracted from the global model obtained from our gPerXAN and the baseline FedAvg on the PACS dataset with the testing domain, a.k.a. target domain Photo (P). From this figure, we can observe that features derived from our method are semantically separated according to 7 categories in the PACS dataset for both source domains and the target domain. Moreover, features of each category on all domains tend to be close and grouped, demonstrating the ability of our method to distill more discriminative and domain-invariant representation, leading to a significant improvement of more than \\(2\\%\\) in classification accuracy on the testing domain Photo (P)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.558, + 0.288, + 0.575 + ], + "angle": 0, + "content": "5.2. Privacy and Efficiency" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.584, + 0.473, + 0.903 + ], + "angle": 0, + "content": "Data privacy is a major concern in the field of FL, which is mitigated by only sharing client models' parameters instead of raw data. However, to resolve the FedDG problem, recently introduced methods sacrifice this principle by revealing partial information about client data. Specifically, leaking more information not only opens higher chances for attackers to perform an inversion attack [24], which aims to reconstruct the original data of clients, but also amplifies the risk of membership inference attack [32], which determines if a sample was in the model's training data. Furthermore, efficiency in communication and computation is critical in FL, especially in scenarios where resource-constrained clients are involved. Despite that, the common sharing mechanism in available methods introduces unacceptable extra costs to the FL training process. For COPA [42], using an ensemble of domain-specific classifiers in the model architecture results in an \\( O(N^2) \\) increase in communication and computation complexities compared to \\( O(N) \\) as conventional, according to the number of participating clients \\( N \\). Meanwhile, FedDG-GA [46] consumes a double of memory at clients. Compared with existing ones, in ad" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.359, + 0.892, + 0.406 + ], + "angle": 0, + "content": "dition to bypassing the limitations described above, our proposed method is more practical in terms of implementation yet provides competitive results in extensive evaluations." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.416, + 0.62, + 0.433 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.441, + 0.892, + 0.76 + ], + "angle": 0, + "content": "In this paper, we introduce a novel architectural method, namely gPerXAN, to address the problem of FedDG. By explicitly assembling Instance Normalization layers into Batch Normalization layers in a personalized scheme and employing a simple yet effective guiding regularizer, our method allows the model to filter domain-specific features and actively distill domain-invariant representation for classification tasks. We conduct extensive experiments and in-depth analysis to quantitatively and qualitatively verify the effectiveness of our proposed method in solving this particular problem. Although our evaluation is more on cross-silo FL, our method can be easily extended to cross-device scenarios while Algorithm 1 remains unchanged. Unlike existing methods, due to the independence from imaging techniques, gPerXAN can be straightforwardly extended to diverse applications. A potential limitation of our method is its certain suitability with normalization-based models. In the future, investigating other forms of regularization terms is a promising research direction due to the available room for improvements. Moreover, the vulnerability of available methods under various attacks remains underexplored." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.77, + 0.661, + 0.788 + ], + "angle": 0, + "content": "Acknowledgement" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.795, + 0.892, + 0.903 + ], + "angle": 0, + "content": "This work is supported by the Deutsche Forschungsgemeinschaft, German Research Foundation under grant number 453130567 (COSMO), by the Horizon Europe Research and Innovation Actions under grant number 101092908 (SmartEdge), and by the Federal Ministry for Education and Research, Germany under grant number 01IS18037A (BIFOLD)." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6034" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.105 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.17 + ], + "angle": 0, + "content": "[1] Yogesh Balaji, Swami Sankaranarayanan, and Rama Chellappa. MetaReg: Towards Domain Generalization Using Meta-Regularization. Advances in Neural Information Processing Systems, 31, 2018. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.171, + 0.472, + 0.268 + ], + "angle": 0, + "content": "[2] Peter Bandi, Oscar Geessink, Quirine Manson, Marcory Van Dijk, Maschenka Balkenhol, Meyke Hermsen, Babak Ehteshami Bejnordi, Byungjae Lee, Kyunghyun Paeng, Aoxiao Zhong, et al. From Detection of Individual Metastases to Classification of Lymph Node Status at the Patient Level: The Camelyon17 Challenge. IEEE Transactions on Medical Imaging, 38(2):550-560, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.269, + 0.471, + 0.338 + ], + "angle": 0, + "content": "[3] Junming Chen, Meirui Jiang, Qi Dou, and Qifeng Chen. Federated Domain Generalization for Image Recognition via Cross-Client Style Transfer. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 361–370, 2023. 2, 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.339, + 0.472, + 0.408 + ], + "angle": 0, + "content": "[4] Ekin D Cubuk, Barret Zoph, Jonathon Shlens, and Quoc V Le. RandAugment: Practical Automated Data Augmentation With a Reduced Search Space. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, pages 702-703, 2020. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.41, + 0.47, + 0.45 + ], + "angle": 0, + "content": "[5] Shi Dong, Ping Wang, and Khushnood Abbas. A Survey on Deep Learning and Its Applications. Computer Science Review, 40:100379, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.451, + 0.47, + 0.507 + ], + "angle": 0, + "content": "[6] Qi Dou, Daniel Coelho de Castro, Konstantinos Kamnitsas, and Ben Glocker. Domain Generalization via Model-Agnostic Learning of Semantic Features. Advances in Neural Information Processing Systems, 32, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.508, + 0.47, + 0.564 + ], + "angle": 0, + "content": "[7] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep Residual Learning for Image Recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 770-778, 2016. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.565, + 0.47, + 0.632 + ], + "angle": 0, + "content": "[8] Gao Huang, Zhuang Liu, Laurens Van Der Maaten, and Kilian Q Weinberger. Densely Connected Convolutional Networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4700-4708, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.634, + 0.47, + 0.689 + ], + "angle": 0, + "content": "[9] Xun Huang and Serge Belongie. Arbitrary Style Transfer in Real-Time With Adaptive Instance Normalization. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1501-1510, 2017. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.69, + 0.469, + 0.746 + ], + "angle": 0, + "content": "[10] Zeyi Huang, Haohan Wang, Eric P Xing, and Dong Huang. Self-Challenging Improves Cross-Domain Generalization. In Proceedings of the European Conference on Computer Vision (ECCV), pages 124-140. Springer, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.747, + 0.469, + 0.803 + ], + "angle": 0, + "content": "[11] Da Li, Yongxin Yang, Yi-Zhe Song, and Timothy M Hospedales. Deeper, Broader and Artier Domain Generalization. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5542-5550, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.804, + 0.469, + 0.871 + ], + "angle": 0, + "content": "[12] Da Li, Jianshu Zhang, Yongxin Yang, Cong Liu, Yi-Zhe Song, and Timothy M Hospedales. Episodic Training for Domain Generalization. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1446-1455, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.872, + 0.469, + 0.902 + ], + "angle": 0, + "content": "[13] Haoliang Li, Sinno Jialin Pan, Shiqi Wang, and Alex C Kot. Domain Generalization With Adversarial Feature Learning." + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.472, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.892, + 0.121 + ], + "angle": 0, + "content": "In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5400-5409, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.124, + 0.892, + 0.178 + ], + "angle": 0, + "content": "[14] Tian Li, Anit Kumar Sahu, Manzil Zaheer, Maziar Sanjabi, Ameet Talwalkar, and Virginia Smith. Federated Optimization in Heterogeneous Networks. Proceedings of Machine Learning and Systems, 2:429-450, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.181, + 0.892, + 0.235 + ], + "angle": 0, + "content": "[15] Xiang Li, Kaixuan Huang, Wenhao Yang, Shusen Wang, and Zhihua Zhang. On the Convergence of Fedavg on Non-IID Data. In International Conference on Learning Representations, 2019. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.239, + 0.892, + 0.293 + ], + "angle": 0, + "content": "[16] Yanghao Li, Naiyan Wang, Jianping Shi, Jiaying Liu, and Xiaodi Hou. Revisiting Batch Normalization for Practical Domain Adaptation. arXiv preprint arXiv:1603.04779, 2016.4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.298, + 0.892, + 0.353 + ], + "angle": 0, + "content": "[17] Quande Liu, Qi Dou, Lequan Yu, and Pheng Ann Heng. MS-Net: Multi-Site Network for Improving Prostate Segmentation With Heterogeneous MRI Data. IEEE Transactions on Medical Imaging, 39(9):2713-2724, 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.355, + 0.892, + 0.437 + ], + "angle": 0, + "content": "[18] Quande Liu, Cheng Chen, Jing Qin, Qi Dou, and Pheng-Ann Heng. FedDG: Federated Domain Generalization on Medical Image Segmentation via Episodic Learning in Continuous Frequency Space. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1013–1023, 2021. 2, 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.442, + 0.892, + 0.496 + ], + "angle": 0, + "content": "[19] Ping Luo, Jiamin Ren, Zhanglin Peng, Ruimao Zhang, and Jingyu Li. Differentiable Learning-To-Normalize via Switchable Normalization. In International Conference on Learning Representations, 2018. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.5, + 0.892, + 0.553 + ], + "angle": 0, + "content": "[20] Divyat Mahajan, Shruti Tople, and Amit Sharma. Domain Generalization Using Causal Matching. In International Conference on Machine Learning, pages 7313-7324. PMLR, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.557, + 0.892, + 0.626 + ], + "angle": 0, + "content": "[21] Brendan McMahan, Eider Moore, Daniel Ramage, Seth Hampson, and Blaise Aguera y Arcas. Communication-Efficient Learning of Deep Networks From Decentralized Data. In Artificial Intelligence and Statistics, pages 1273-1282. PMLR, 2017. 1, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.63, + 0.892, + 0.685 + ], + "angle": 0, + "content": "[22] Krikamol Muandet, David Balduzzi, and Bernhard Scholkopf. Domain Generalization via Invariant Feature Representation. In International Conference on Machine Learning, pages 10-18. PMLR, 2013. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.688, + 0.892, + 0.756 + ], + "angle": 0, + "content": "[23] Hyeonseob Nam, HyunJae Lee, Jongchan Park, Wonjun Yoon, and Donggeun Yoo. Reducing Domain Gap by Reducing Style Bias. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8690-8699, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.759, + 0.892, + 0.841 + ], + "angle": 0, + "content": "[24] Ngoc-Bao Nguyen, Keshigeyan Chandrasegaran, Milad Abdollahzadeh, and Ngai-Man Cheung. Re-thinking Model Inversion Attacks Against Deep Neural Networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, pages 16384-16393, 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.845, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[25] Xingang Pan, Ping Luo, Jianping Shi, and Xiaou Tang. Two at Once: Enhancing Learning and Generalization Capacities via IBN-Net. In Proceedings of the European Conference on Computer Vision (ECCV), pages 464–479, 2018. 1, 2, 3, 5, 7" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6035" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.134 + ], + "angle": 0, + "content": "[26] Xingchao Peng, Zijun Huang, Yizhe Zhu, and Kate Saenko. Federated Adversarial Domain Adaptation. In International Conference on Learning Representations, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.134, + 0.471, + 0.203 + ], + "angle": 0, + "content": "[27] Krishna Pillutla, Kshitiz Malik, Abdel-Rahman Mohamed, Mike Rabbat, Maziar Sanjabi, and Lin Xiao. Federated Learning With Partial Model Personalization. In International Conference on Machine Learning, pages 17716-17758. PMLR, 2022. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.204, + 0.471, + 0.245 + ], + "angle": 0, + "content": "[28] Joaquin Quinonero-Candela, Masashi Sugiyama, Anton Schwaighofer, and Neil D Lawrence. Dataset Shift in Machine Learning. Mit Press, 2008. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.246, + 0.471, + 0.302 + ], + "angle": 0, + "content": "[29] Alexandre Rame, Coretin Dancette, and Matthieu Cord. Fishr: Invariant Gradient Variances for Out-Of-Distribution Generalization. In International Conference on Machine Learning, pages 18347–18377. PMLR, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.302, + 0.471, + 0.371 + ], + "angle": 0, + "content": "[30] Seonguk Seo, Yumin Suh, Dongwan Kim, Geeho Kim, Jongwoo Han, and Bohyung Han. Learning to Optimize Domain Specific Normalization for Domain Generalization. In Proceedings of the European Conference on Computer Vision (ECCV), pages 68-83, 2020. 2, 3, 4, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.371, + 0.471, + 0.426 + ], + "angle": 0, + "content": "[31] Neta Shoham, Tomer Avidor, Aviv Keren, Nadav Israel, Daniel Benditkis, Liron Mor-Yosef, and Itai Zeitak. Overcoming Forgetting in Federated Learning on Non-IID Data. arXiv preprint arXiv:1910.07796, 2019. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.426, + 0.471, + 0.482 + ], + "angle": 0, + "content": "[32] Reza Shokri, Marco Stronati, Congzheng Song, and Vitaly Shmatikov. membership inference attacks against machine learning models. In 2017 IEEE Symposium on Security and Privacy (SP), pages 3-18. IEEE, 2017. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.482, + 0.471, + 0.538 + ], + "angle": 0, + "content": "[33] Baochen Sun and Kate Saenko. Deep Coral: Correlation Alignment for Deep Domain Adaptation. In Proceedings of the European Conference on Computer Vision (ECCV), pages 443-450, 2016. 1, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.539, + 0.471, + 0.593 + ], + "angle": 0, + "content": "[34] Baochen Sun, Jiashi Feng, and Kate Saenko. Correlation Alignment for Unsupervised Domain Adaptation. Domain Adaptation in Computer Vision Applications, pages 153-171, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.594, + 0.471, + 0.65 + ], + "angle": 0, + "content": "[35] Benyuan Sun, Hongxing Huo, Yi Yang, and Bo Bai. *PartialFed: Cross-Domain Personalized Federated Learning via Partial Initialization*. Advances in Neural Information Processing Systems, 34:23309-23320, 2021. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.65, + 0.471, + 0.704 + ], + "angle": 0, + "content": "[36] Yuwei Sun, Ng Chong, and Hideya Ochiai. Feature Distribution Matching for Federated Domain Generalization. In Asian Conference on Machine Learning, pages 942-957. PMLR, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.705, + 0.471, + 0.747 + ], + "angle": 0, + "content": "[37] Laurens Van der Maaten and Geoffrey Hinton. Visualizing Data Using t-SNE. Journal of Machine Learning Research, 9(11), 2008. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.747, + 0.471, + 0.802 + ], + "angle": 0, + "content": "[38] Gael Varoquaux and Veronika Cheplygina. Machine Learning for Medical Imaging: Methodological Failures and Recommendations for the Future. Npj Digital Medicine, 5(1):48, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.802, + 0.471, + 0.872 + ], + "angle": 0, + "content": "[39] Hemanth Venkateswara, Jose Eusebio, Shayok Chakraborty, and Sethuraman Panchanathan. Deep Hashing Network for Unsupervised Domain Adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5018-5027, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.872, + 0.471, + 0.901 + ], + "angle": 0, + "content": "[40] Paul Voigt and Axel Von dem Bussche. The EU General Data Protection Regulation (GDPR). A Practical Guide, 1st Ed.," + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.471, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.092, + 0.892, + 0.12 + ], + "angle": 0, + "content": "Cham: Springer International Publishing, 10(3152676):10-5555, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.121, + 0.892, + 0.19 + ], + "angle": 0, + "content": "[41] Jindong Wang, Cuiling Lan, Chang Liu, Yidong Ouyang, Tao Qin, Wang Lu, Yiqiang Chen, Wenjun Zeng, and Philip Yu. Generalizing to Unseen Domains: A Survey on Domain Generalization. IEEE Transactions on Knowledge and Data Engineering, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.191, + 0.892, + 0.261 + ], + "angle": 0, + "content": "[42] Guile Wu and Shaogang Gong. Collaborative Optimization and Aggregation for Decentralized Domain Generalization and Adaptation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6484-6493, 2021. 3, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.261, + 0.892, + 0.318 + ], + "angle": 0, + "content": "[43] Zhenlin Xu, Deyi Liu, Junlin Yang, Colin Raffel, and Marc Niethammer. Robust and Generalizable Visual Representation Learning via Random Convolutions. In International Conference on Learning Representations, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.319, + 0.892, + 0.387 + ], + "angle": 0, + "content": "[44] Fu-En Yang, Yuan-Chia Cheng, Zu-Yun Shiau, and Yu-Chiang Frank Wang. Adversarial Teacher-Student Representation Learning for Domain Generalization. Advances in Neural Information Processing Systems, 34:19448-19460, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.389, + 0.892, + 0.457 + ], + "angle": 0, + "content": "[45] Chun-Han Yao, Boqing Gong, Hang Qi, Yin Cui, Yukun Zhu, and Ming-Hsuan Yang. Federated Multi-Target Domain Adaptation. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1424–1433, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.459, + 0.892, + 0.53 + ], + "angle": 0, + "content": "[46] Ruipeng Zhang, Qinwei Xu, Jiangchao Yao, Ya Zhang, Qi Tian, and Yanfeng Wang. Federated Domain Generalization with Generalization Adjustment. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3954-3963, 2023. 3, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.53, + 0.892, + 0.573 + ], + "angle": 0, + "content": "[47] Kaiyang Zhou, Yongxin Yang, Yu Qiao, and Tao Xiang. Domain Generalization With Mixstyle. In International Conference on Learning Representations, 2020. 1, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.573, + 0.892, + 0.628 + ], + "angle": 0, + "content": "[48] Weiming Zhuang, AI Sony, and Lingjuan Lyu. FedWon: Triumphing Multi-domain Federated Learning Without Normalization. In International Conference on Learning Representations, 2024. 4" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.092, + 0.892, + 0.628 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.517, + 0.956 + ], + "angle": 0, + "content": "6036" + } + ] +] \ No newline at end of file diff --git a/2024/Efficiently Assemble Normalization Layers and Regularization for Federated Domain Generalization/16ce0297-575e-48b3-acb7-4a9aa0af4fe6_origin.pdf b/2024/Efficiently Assemble Normalization Layers and Regularization for Federated Domain Generalization/16ce0297-575e-48b3-acb7-4a9aa0af4fe6_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..562571eec7f6bdebe531fd28897b5bf83aba914a --- /dev/null +++ b/2024/Efficiently Assemble Normalization Layers and Regularization for Federated Domain Generalization/16ce0297-575e-48b3-acb7-4a9aa0af4fe6_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4d5c573a5b26bcae21023f4929f441c06eaed6db1c834570f7f3fe6194a41999 +size 1122726 diff --git a/2024/Efficiently Assemble Normalization Layers and Regularization for Federated Domain Generalization/full.md b/2024/Efficiently Assemble Normalization Layers and Regularization for Federated Domain Generalization/full.md new file mode 100644 index 0000000000000000000000000000000000000000..7db8d4883302edfe1b5095e588efcb66d86bc84c --- /dev/null +++ b/2024/Efficiently Assemble Normalization Layers and Regularization for Federated Domain Generalization/full.md @@ -0,0 +1,264 @@ +# Efficiently Assemble Normalization Layers and Regularization for Federated Domain Generalization + +Khiem Le $^{1}$ , Long Ho $^{2}$ , Cuong Do $^{2}$ , Danh Le-Phuoc $^{3}$ , Kok-Seng Wong $^{2,*}$ $^{1}$ Department of Computer Science and Engineering, University of Notre Dame, IN, USA + $^{2}$ College of Engineering and Computer Science, VinUniversity, Hanoi, Vietnam + $^{3}$ Open Distributed Systems, Technical University Berlin, Berlin, Germany + +kle3@nd.edu, dans.lephuoc@tu-berlin.de, {long.ht, cuong.dd, wong.ks}@vinuni.edu.vn + +# Abstract + +Domain shift is a formidable issue in Machine Learning that causes a model to suffer from performance degradation when tested on unseen domains. Federated Domain Generalization (FedDG) attempts to train a global model using collaborative clients in a privacy-preserving manner that can generalize well to unseen clients possibly with domain shift. However, most existing FedDG methods either cause additional privacy risks of data leakage or induce significant costs in client communication and computation, which are major concerns in the Federated Learning paradigm. To circumvent these challenges, here we introduce a novel architectural method for FedDG, namely gPerXAN1, which relies on a normalization scheme working with a guiding regularizer. In particular, we carefully design Personalized eXplicitly Assessed Normalization to enforce client models selectively filtering domain-specific features that are biased towards local data while retaining discrimination of those features. Then, we incorporate a simple yet effective regularizer to guide these models in directly capturing domain-invariant representations that the global model's classifier can leverage. Extensive experimental results on two benchmark datasets, i.e., PACS and Office-Home, and a real-world medical dataset, Camelyon17, indicate that our proposed method outperforms other existing methods in addressing this particular problem. + +# 1. Introduction + +Over the past few decades, Machine Learning (ML) has demonstrated remarkable achievements across diverse areas such as Computer Vision, Natural Language and Speech Processing, or Robotics [5]. In general, most ML models rely on an over-simplified assumption, i.e., the train- + +ing and testing data are independent and identically distributed, which does not always reflect real-world practices. In practical scenarios where the distribution of testing data diverges from that of training data, the performance of ML models often drops catastrophically due to the domain shift issue [28]. Additionally, obtaining or identifying the testing data before model deployment can be challenging in numerous applications. For instance, in biomedical applications where data characteristics vary across different equipment and institutions, gathering data from all potential domains in advance is impractical. Therefore, it is essential to have a solution that can improve the generalization capability of such ML models to adapt effectively to unseen domains. + +Domain Generalization (DG) has been proposed to address the challenge of training ML models using data from single or multiple source domains with the expectation that these models will perform well on unseen domains [41]. The majority of existing DG methods fall under the category of domain-invariant representation learning approach [13, 20, 25, 29, 33]. This approach relies on a broadly acknowledged assumption that each domain contains its own domain-specific features, which are biased towards spurious relations in the data, and that all domains share domain-invariant features, which are general and robust to any unseen domains. From this assumption, previous works propose methods that remove domain-specific features and distill domain-invariant features to achieve the generalization ability. Alternative approaches for DG encompass data augmentation [23, 43, 44, 47], which involves exposing models to artificially generated domains, and meta-learning [1, 6, 12], an approach that emulates the domain shift during the meta-training phase. However, most methods require a centralized setting where all source domains are collected together. Consequently, these methods cannot be readily expanded to decentralized settings. + +Federated Learning (FL) [21] is an emerging decentralized learning paradigm widely adopted in various applica + +![](images/2d3d7aeaf622fa60cb0633c1b35aee5ab8449b09c55d3cafeff7f42fa3f3a00c.jpg) +Figure 1. An overview of our proposed gPerXAN method for solving the FedDG problem. + +tions to cope with the increasing privacy concerns of data centralization [40]. Specifically, the paradigm works in a way that each client learns from their data and only aggregates local models' parameters at a certain frequency at the central server to generate a global model. Notably, all data samples are kept within each client during the FL training process. Due to the nature of data decentralization, where each client owns a single source domain, as illustrated in Figure 1, the FL paradigm poses further significant challenges for DG and limits the applicability of available DG methods. There have been some early attempts to address the DG problem in the FL scenario. For instance, Liu et al. [18] introduces a method that allows clients to share their image data in the frequency space with each other, thus relatively recovering the centralization process at each client. Similarly, Chen et al. [3] introduces another method that extracts and exchanges the style of local images among all clients. It is evident that these initial efforts employ a strategy that necessitates the sharing of partial client data, thereby compromising the data privacy constraints of FL to a certain extent. Although they show promising results, these methods can be overly complicated to implement in practice and lead to additional privacy risks during the FL training process. + +To address the aforementioned challenges, this paper introduces a novel architectural method for domain-invariant representation learning within the FL framework. The proposed method enhances the generalization ability while upholding the fundamental privacy principles of FL. Based on the effectiveness of discarding domain-specific information from learned features [25, 30], we properly assemble Instance Normalization layers (IN) into Batch Normaliza + +tion layers (BN) in well-known Convolutional Neural Networks (CNNs) using an explicit differential mixture as in Eqn (2). Moreover, thanks to the explicit property, the benefit of personalization in FL [27, 35] can be incorporated into the normalization scheme using local BN sides. Specifically, during the FL training process, while IN sides are globally aggregated along with other model parameters, BN sides are updated locally without broadcasting. In addition, we argue that only relying on the ability to filter domain-specific features of IN while lacking guidance to distill domain-invariant representations directly might lead to suboptimal performance. Based on this observation, we introduce a simple yet highly effective regularization term to guide client models to directly capture domain-invariant representations that can be used by the global model's classifier, which is aggregated from client models' classifiers. + +To summarize, our main contributions in this paper are highlighted as follows: + +- Different from existing methods for DG in the FL scenario, we propose a novel method that concentrates on a personalized normalization scheme, global IN while local BN, for filtering domain-specific features and fully respecting the privacy-preserving principles of FL. +- Furthermore, we propose a simple yet effective regularization term to introduce clear guidance to client models for directly capturing domain-invariant representations, further improving performance on unseen domains. +- Finally, we conduct extensive experiments on two benchmark datasets, i.e., PACS and Office-Home, and a real-world medical dataset, Camelyon17, where our proposed method outperforms existing relevant ones. + +# 2. Related Work + +# 2.1. Domain Generalization + +Domain Generalization is a challenging task requiring models to perform well on unseen domains. One dominant approach is domain-invariant representation learning, which aims to minimize the discrepancy among source domains, assuming that the resulting representation will be domain-invariant and generalize well on testing domains. Along this track, Muandet et al. [22] proposes to reduce the domain dissimilarity by using Maximum Mean Discrepancy. Pan et al. [25] combines IN and BN in popular CNNs to reap the benefits of removing domain-specific features while maintaining the ability to capture discriminative features. Meanwhile, Seo et al. [30] discloses that combining these normalization layers in a switchable mechanism can yield better performance. Another approach to DG is data augmentation, which augments source domains to a broader span of the training data space, enlarging the possibility of covering the span of the data in the testing domain. Zhou et al. [47] mixes styles of training instances, resulting in novel domains being synthesized implicitly. Recently, the meta-learning approach has drawn increasing attention from the DG community. Balaji et al. [1] proposes MetaReg that learns a regularization function, particularly for the network classification layer, while excluding the feature extractor. However, the majority of these methods rely on having access to a diverse set of source domains, making them not applicable to decentralized settings. + +# 2.2. Federated Domain Generalization + +While there is a growing interest in addressing the DG problem within the FL framework, particularly in scenarios where each client possesses a single source domain, the existing literature on this subject remains relatively sparse. This FedDG problem was initially introduced by Liu et al. [18], who then proposed ELCFS that involves exchanging amplitude information in the frequency space among clients and leveraging episodic learning to further enhance performance. Chen et al. [3] proposes a similar mechanism CCST that extracts and exchanges the overall domain style of local images among all clients. Unfortunately, these early works require sharing partial information about the local data, which can be viewed as a form of data leakage and is undesirable in the FL setting. Moreover, these methods include broadcasting partial information and interpolating this information into new training data, which adds significant extra cost to communication and computation during the FL training process. Besides, Wu and Gong [42] proposes an architectural method COPA in which the global model consists of a domain-invariant representation extractor and an ensemble of domain-specific classifiers. Sun et al. [36] introduces FedKA that employs a server-side voting mech + +
MethodPrivacy RiskAdditional Cost
CommunicationComputation
ELCFS✓✓✓
CCST✓✓✓
COPA×✓✓
FedDG-GA××✓
gPerXAN (Ours)×××
+ +Table 1. An advantage comparison of different methods. + +anism that generates target domain pseudo-labels based on the consensus from clients to facilitate global model finetuning. Recently, Zhang et al. [46] proposes a novel global objective incorporating a variance reduction regularizer to encourage fairness, and then FedDG-GA is proposed to optimize this objective by dynamically calibrating the aggregation weights. Although relaxing from the data leakage issue, these methods also cause a large additional consumption of resources when the number of source domains and output classes increases. In this paper, we introduce an alternative architectural method that shares only the model update information during training. This ensures maximal data privacy and circumvents the communication and computation overhead issue mentioned earlier while achieving competitive results. In Table 1, we highlight the advantages of our method in comparison to previous ones. + +One related research field to FedDG is Federated Domain Adaptation (FedDA). While both fields aim to maximize the model performance on unseen domains using existing source domains, FedDA can access the target domains while FedDG cannot see those data during training. Leveraging this assumption, FedDA methods such as FADA [26] or FMTDA [45], are typically able to align target and source features. However, this specific assumption also makes FedDA methods are inapplicable to FedDG. + +# 2.3. Personalized Federated Learning + +Another related and orthogonal line of work is Personalized Federated Learning (pFL). pFL aims to learn personalized models for different clients to tackle distribution shifts across client data, known as data heterogeneity. Li et al. [14] utilizes a regularization term in the local loss function so that the clients' trained models will not significantly differ from the global model. On the other side, Sun et al. [35] and Pillutla et al. [27] mitigate the heterogeneous distributions by only loading a subset of the global model's parameters rather than loading the entire model at each training round. Although related in terms of overcoming distribution shifts across clients, pFL focuses on improving the performance of participating clients, whereas our considered field FedDG focuses on improving the performance of unseen clients with unseen domains. + +# 3. Methodology + +Problem Formulation. First, we denote $X$ and $Y$ as the input space and the label space, respectively, of a specific task $\mathcal{T}$ . In the standard FL setting, $N$ clients $\{c_i\}_{i=1}^N$ are involved in collaboratively constructing a global model $f_g$ for solving the task, where each client $c_i$ owns a dataset $D_i = \{(x_k, y_k)\}_{k=1,|D_i|}$ which is associated to a specific domain defined by a joint distribution $P_{X,Y}^{(i)}$ . Importantly, scattered datasets $\{D_i\}_{i=1}^N$ across clients satisfy: + +- $P_{X,Y}^{(i)} \neq P_{X,Y}^{(j)}$ with $1 \leq i, j \leq N$ and $i \neq j$ +- $P_{Y|X}^{(i)} = P_{Y|X}^{(j)}$ with $1 \leq i, j \leq N$ and $i \neq j$ . + +The objective of FedDG is leveraging scattered datasets to construct a model $f_{g}$ that can directly generalize to the unseen dataset $D_{U}$ with an unseen domain, which means: + +- $P_{X,Y}^{(U)} \neq P_{X,Y}^{(i)}$ with $1 \leq i \leq N$ +- $P_{Y|X}^{(U)} = P_{Y|X}^{(i)}$ with $1 \leq i \leq N$ + +To this end, $N$ clients communicate with a central server for $T$ rounds. At each round, every client $c_{i}$ receives the same global model $f_{g}$ from the server and updates $f_{g}$ with their local dataset $D_{i}$ for $E$ epochs to establish its local model $f_{i}$ . The server then collects all trained models and aggregates them to update the global model. This process repeats until the global model converges. In this work, we consider the most popular FL framework, FedAvg [21], which aggregates client models as: + +$$ +f _ {g} = \sum_ {i = 1} ^ {N} \frac {\left| D _ {i} \right|}{\sum_ {1 \leq j \leq N} \left| D _ {j} \right|} f _ {i} \tag {1} +$$ + +Challenges. In the general spirit of DG, a model is expected to extensively explore multiple source domains to achieve domain-invariance in its learned latent representation. However, under the FL setting, each client is restricted to accessing only its own local data, which constrains to make full use of source domains and, consequently, limits learning of generalizable representation. Moreover, sharing data or even partial information about data among clients poses additional privacy risks and introduces communication and computation costs during the FL training process. + +# 3.1. Normalization Scheme + +eXplicitly Assessed Normalization. The first step towards solving the highlighted challenges, we introduce a novel normalization scheme called eXplicitly Assessed Normalization (XAN), which combines IN and BN as: + +$$ +\hat {h} = w _ {i n} \left(\gamma_ {i n} \frac {h - \mu_ {i n}}{\sqrt {\sigma_ {i n} ^ {2} + \epsilon}} + \beta_ {i n}\right) + w _ {b n} \left(\gamma_ {b n} \frac {h - \mu_ {b n}}{\sqrt {\sigma_ {b n} ^ {2} + \epsilon}} + \beta_ {b n}\right), \tag {2} +$$ + +where $h, \hat{h} \in \mathbb{R}^{B \times C \times W \times H}$ are layer input and output activations, which are 4D tensors with dimensions of batch size $B$ , number of channels $C$ , width $W$ and height $H$ . Meanwhile, $\mu$ and $\sigma^2$ are means and variances captured by the normalization layers, respectively, $\gamma$ and $\beta$ are affine parameters, and $\epsilon$ is for numerical stability. + +In essence, XAN is a sophisticated mixture mechanism of IN and BN to replace BN in the feature extractors of CNNs. In particular, $w_{in}$ and $w_{bn}$ are ratios to weight the mixture that allow the model to switch between IN and BN. These parameters are randomly initialized and optimized along with other model parameters during training in an end-to-end manner. Unlike previous works [19, 30], which uses an implicit mechanism to combine computed statistics of IN and BN, i.e., means and variances, XAN uses an explicit mechanism that combines the output activations of two normalization layers, this provides the model with a unique ability to separate IN from BN completely. Note that IN has shown great success in style transfer tasks [9] as it allows discarding the variability of visual styles such as object colors or textures from the content of images. This property makes IN beneficial for solving the DG problem. However, directly using IN to replace the conventional BN leads to the loss of discrimination in the learned features, resulting in performance degradation in classification tasks. + +Personalized eXplicitly Assessed Normalization. To accomplish our proposed normalization scheme, Personalized eXplicitly Assessed Normalization (PerXAN), we make a subtle modification to the conventional FedAvg [21] framework based on the explicit property of the above XAN. Specifically, during the FL training process, while IN sides of XAN layers are globally aggregated along with other model parameters, BN sides are updated locally, which means that parameters of BN sides are excluded from the broadcasting steps from the server. Notably, in inference time, the global model is generated by averaging all model parameters of clients. Our modification is motivated by a common observation that FedAvg normally results in poor convergence and performance in the presence of domain heterogeneity across clients [15, 48] with the major reason being that client models forget the acquired knowledge from previous rounds after aggregated [31]. The leading solution is to personalize a subset of the model, which benefits clients in learning better from their local data [27, 35]. Moreover, based on another finding that BN plays a crucial role in dealing with the domain shift issue in the centralized paradigm [16, 17], BN sides in XAN layers are finally made to be personalized. PerXAN is illustrated in Figure 1. + +# 3.2. Regularization as Guidance + +With the proposed normalization scheme PerXAN in place, now, we are ready to introduce our "Regularization as Guidance". It targets to induce clear guidance to client + +models through a regularizer, which brings an effect of domain alignment [33, 34], and guides these models to capture domain-invariant representations directly. We hypothesize that this will address a drawback of current DG methods that are later verified by our experiments in Section 4. The drawback is from DG methods that take advantage of the IN's function [25, 30] but do not actually equip the model with the capability of capturing domain-invariant features even if they demonstrated promising performance. Instead, it is only hoped that domain-invariant features would be distilled through achieving the goal of removing domain-specific features. This indirect learning purpose might affect the model's learning efficacy, especially in the FL setting, where each client only owns a separate single source domain. Specifically, we assume a classification model $f$ , either the global model $f_{g}$ or a client model $f_{i}$ , comprises a feature extractor $g$ and a classifier head $h$ . At each client $c_{i}$ , during local training, the client model $f_{i}$ is optimized on the local dataset $D_{i}$ using the following loss function: + +$$ +\begin{array}{l} L _ {i} = L _ {c l s} ^ {(i)} \left(f _ {i}; D _ {i}\right) + \lambda L _ {r e g} ^ {(i)} \left(g _ {i}, h _ {g}; D _ {i}\right) (3) \\ = \sum_ {k = 1} ^ {| D _ {i} |} \ell \left(f _ {i} \left(x _ {k} ^ {(i)}\right), y _ {k} ^ {(i)}\right) + \lambda \ell \left(h _ {g} \left(g _ {i} \left(x _ {k} ^ {(i)}\right)\right), y _ {k} ^ {(i)}\right), (4) \\ \end{array} +$$ + +where $\ell$ is the base loss function, which is usually cross-entropy in classification tasks, and $\lambda$ is a hyper-parameter to control the significance of the regularizer. In a deeper look, by freezing $h_g$ during local training, the auxiliary loss term $L_{reg}$ forces client models to arrive at representations that can be made use for classification by the same global classifier, hence, producing an alignment effect on these representations [33, 34]. Moreover, our proposed regularizer can also be interpreted as an implicit form of matching global knowledge to clients' knowledge, which has been demonstrated to bring performance gain [3, 18] for FedDG. + +Due to its orthogonality, our proposed regularizer can be easily integrated into various methods, as shown in Algorithm 1. Specifically, our experiments in Section 4 verify its compatibility with the normalization scheme that justifies our motivation. Moreover, by using only the global model's classifier to regularize feature extractors, gPerXAN saves major communication and computation resources, as well as memory usage at clients compared to others, which use an ensemble of classifiers [42] or entire global model [46]. + +# 4. Experiments and Results + +# 4.1. Datasets + +To evaluate the proposed method, we perform experiments on two standard DG benchmark datasets, i.e., PACS [11] and Office-Home [39], and a real-world medical image dataset, Camelyon17 [2], consisting of various sub-datasets + +Algorithm 1 gPerXAN +1: Input: A model $f$ uses PerXAN to replace BN layers in the feature extractor. $N$ clients with their local datasets $\{D_i\}_{i = 1}^N$ . Notably, $f^{(t)}$ and $f^{(l;t)}$ is the model $f$ and its $l^{th}$ layer at the communication round $t$ , respectively. +2: Initialization: $f_{g}^{(0)}\gets f$ +3: for each round $t = 1,2,3,\ldots ,T$ do +4: for each client $i = 1,2,3,\ldots ,N$ in parallel do +5: for each layer in $f_{i}^{(t)}$ do +6: if $f_{i}^{(l;t)}$ is not BN then +7: $f_{i}^{(l;t)}\gets f_{g}^{(l;t)}$ +8: end if +9: end for +10: $f_{i}^{(t)}\gets \mathrm{LocalTraining}(f_{i}^{(t)},h_{g};D_{i})$ +11: end for +12: $f_{g}^{(t)} = \sum_{i = 1}^{N}\frac{|D_{i}|}{\sum_{1\leq j\leq N}|D_{j}|} f_{i}^{(t)}$ // Eqn (2) +13: end for +14: return: $f_{g}^{(T)}$ +15: LocalTraining $(f_i,h_g;D_i)$ +16: for each epoch $e = 1,2,3,\dots ,E$ do +17: $L_{i} = L_{cls}^{(i)}(f_{i};D_{i}) + \lambda L_{reg}^{(i)}(g_{i},h_{g};D_{i})$ // Eqn (3) +18: $f_{i}\gets f_{i} - \eta \nabla L_{i}$ +19: end for +20: return: $f_{i}$ + +that are considered as domains. Specifically, PACS is composed of 4 domains with large discrepancies from diverse image colors and textures, *Photo* (P), *Art painting* (A), *Cartoon* (C), and *Sketch* (S). Each domain contains 7 categories, with 9,991 images in total. Office-Home is also composed of 4 domains but with smaller discrepancies from various backgrounds and camera viewpoints, *Product* (P), *Art* (A), *Clipart* (C), and *Real-world* (R). Each domain contains a more extensive label set of 65 categories with 15,588 images. Camelyon17 is a binary tumor classification dataset containing 455,964 histology images with stains from 5 different hospitals worldwide. These datasets are at different difficulty levels and are commonly used in the literature. + +# 4.2. Experimental Settings + +Evaluation. For a fair comparison purpose, we follow the common leave-one-domain-out evaluation protocol as considered in [3, 18, 46]. In particular, we sequentially choose one domain as the unseen domain, train the model on all remaining domains where a single domain is treated as a client, and evaluate the trained model on the chosen domain. For PACS and Office-Home datasets, we split $90\%$ of the data of each source client as the training set and $10\%$ + +
MethodPACSOffice-Home
PACSAvgPACRAvgAvg
FedAvg (Baseline)95.2182.2378.2073.5682.3076.5365.9755.4078.0168.9875.64
FedAvg w/ MixStyle95.9385.9980.0375.4684.3575.8762.0957.9277.4868.3476.35
FedAvg w/ RSC95.2183.1578.2474.6282.8175.2662.3450.7977.4666.4674.63
ELCFS96.2383.9479.2773.3083.1976.8366.3255.6378.1269.2376.21
CCST96.6588.3378.2082.9086.5276.6166.3552.3980.0168.8477.68
COPA95.6284.8080.2882.8685.8975.8262.2756.0478.7268.2177.05
FedDG-GA96.8086.9181.2382.7486.9277.2365.1058.2978.8069.8678.42
gPerXAN (Ours)97.2786.5284.6883.2887.9478.9167.2457.7580.1571.0179.48
+ +of that as the validation set, while for unseen clients, the entire data is used for testing. For the large-scale Camelyon17 dataset, the ratios of training set and validation set at each source client are $80\%$ and $20\%$ , respectively, while the entire data is used for testing at unseen clients. In all experiments, we report the test accuracy on each unseen client by using the best validation model selected based on the average of accuracies on validation sets of source clients. The reported numerical values are averaged over 3 runs. Finally, we straightforwardly compare the proposed method with the vanilla FedAvg [21]. Two centralized DG methods, which are free from the requirement of data centralization, MixStyle [47] and RSC [10], are also evaluated under the integration into the FedAvg framework. Furthermore, we directly compare our method with state-of-the-art relevant ones that address the problem of DG in the FL setting as discussed in Section 2, including ELCFS [18], CCST [3], COPA [42], and FedDG-GA [46]. + +Implementation Details. We present architectural details and hyper-parameter values used for experiments in the paper. Following [3, 46], for PACS and Office-Home datasets, we use the image size of $224 \times 224$ pixels and ResNet-50 [7] pre-trained on ImageNet as the backbone for the feature extractor and a linear layer as the classifier. BN layers in the first four and first two blocks in feature extractors are replaced with our PerXAN, respectively. For the Camelyon17 dataset, input images are resized to $96 \times 96$ pixels, and DenseNet-121 [8] is used instead of ResNet-50, BN layers in the first dense block and the following transition block are replaced by PerXAN. Notably, parameters related to BN layers are initialized with ImageNet pre-trained weights. Client models in all experiments are optimized by an SGD optimizer with a learning rate of 2e-3 for 100 communication rounds (i.e. $T = 100$ ) with one local update epoch at clients (i.e. $E = 1$ ). During the training, simple + +Table 2. Accuracy comparison on the PACS and Office-Home datasets in the leave-one-domain-out setting. + +
MethodCamelyon17
H1H2H3H4H5Avg
FedAvg (Baseline)97.091.889.994.281.090.8
FedAvg w/ MixStyle91.185.586.293.387.988.8
FedAvg w/ RSC90.690.688.394.593.391.5
ELCFS92.990.689.993.289.991.3
CCST91.590.287.394.691.691.0
COPA93.290.992.293.690.292.0
FedDG-GA97.290.791.092.390.592.3
gPerXAN (Ours)96.592.295.194.791.994.1
+ +Table 3. Accuracy comparison on the Camelyon17 dataset. + +data augmentation techniques are applied including random horizontal flipping and color jittering. The hyper-parameter of our regularization term $\lambda$ is searched in the range [0, 1] with a step of 0.25. The MixStyle [47] and RSC [10] can be directly integrated into the FedAvg without further modifications. All settings of other compared methods, i.e., EL-CFS [18], CCST [3], COPA [42], and FedDG-GA [46] are chosen based on corresponding papers. + +# 4.3. Main Results + +Table 2 presents the quantitative results for different testing domains on the PACS and Office-Home datasets. Particularly, each result column shows the test accuracy of the global model on the domain of the column name. First of all, we can empirically verify that centralized DG methods such as MixStyle [47] and RSC [10] are inconsistent under the FedAvg framework and even harmful in certain cases compared to the baseline. Since these methods are + +designed for the centralized setting, they might need access to inter-domain knowledge to perform well, which is unavailable in the FL setting. In comparing state-of-the-art FedDG methods designed for the FL scenario, methods that share partial client data information, i.e., ELCFS [18] and CCST [3], show impressive results on the PACS dataset. However, they do not significantly affect the Office-Home dataset. A similar pattern is found with the architectural method COPA [42] and FedDG-GA [46]. Meanwhile, on the PACS and Office-Home datasets, our proposed method, gPerXAN achieves average accuracies across unseen clients of $87.94\%$ and $71.01\%$ , which are $1.02\%$ and $1.15\%$ better than the second-best ones, respectively. Although most of the considered methods can perform better than the baseline FedAvg in most cases, our method demonstrates a significant boost over others on both two standard benchmarks. + +In addition to standard PACS and Office-Home datasets, we further evaluate our proposed method on a medical imaging dataset Camelyon17 as presented in Table 3. On this real-world benchmark, we can observe that ELCFS [18] and CCST [3] show considerably inferior performance compared to architectural ones, i.e., COPA [42] and gPerXAN. This might be due to sophisticated and sensitive features in medical images that are more challenging to extract and interpolate than conventional datasets [38], leading to inefficiency in these information-sharing-based methods. Notably, although it yields impressive performance, COPA [42] involves several other advanced techniques, such as RandAugment [4], making the performance gain hard to justify. Meanwhile, our method achieves an average accuracy across unseen clients of $94.1\%$ , outperforming FedDG-GA [46] by approximately $2\%$ . In general, the above experimental results demonstrate the effectiveness of the proposed method across various applications. + +# 4.4. Ablation Studies + +We conduct ablation studies on the PACS dataset to investigate the impact of each building component. Specifically, we first compare PerXAN with other centralized variants such as conventional BN, I-BN [25], and DSON [30] under the same implementation details, and then our guiding regularizer is applied to ELCFS [18] and CCST [3] to verify its compatibility with the rest of the method. + +# 4.4.1 Impact of the Normalization Scheme + +Table 4 provides a quantitative comparison among considered normalization schemes where gFedAvg represents the FedAvg framework implemented with the guiding regularizer. This means that we sequentially utilize the conventional BN, I-BN [25], and DSON [30] to replace the PerXAN scheme in our proposed method. From this table, we can observe that I-BN [25] can yield better perfor + +
MethodPACS
PACSAvg
gFedAvg w/ BN95.8584.3978.0476.4283.68
gFedAvg w/ I-BN94.6782.0880.4680.0784.32
gFedAvg w/ DSON93.7781.6479.9180.6383.99
gPerXAN (Ours)97.2786.5284.6883.2887.94
+ +![](images/94f9c628159021288fa16c579adab956b87c84502f575bb30bf9319039b2937d.jpg) +Table 4. Evaluation of different normalization schemes. +Figure 2. Contribution of the regularizer on different methods. + +mance than DSON [30] and the conventional BN, which are slightly comparable to each other. Meanwhile, PerXAN shows optimal performance with a significant margin compared to others and largely contributes to the whole method. + +# 4.4.2 Impact of the Regularizer + +To better understand the proposed regularization term's contribution, we sequentially employ it in FedAvg, ELCFS [18], and CCST [3] to observe performance changes. Figure 2 displays averaged accuracies across unseen clients of these methods without and with the involvement of our regularizer. It is straightforward to see that while regularizer improves the performance of FedAvg and PerXAN considerably, it does not enhance ELCFS [18] and CCST [3]. In information-sharing-based methods, i.e., ELCFS [18] and CCST [3], each client is exposed to other clients' data information, which means that clients can access global knowledge relatively, then making the effect of matching global knowledge to clients' knowledge redundant and harmful. Also, this ablation study empirically verifies the connection between the regularizer and normalization scheme. + +# 5. Analysis and Discussion + +# 5.1. Visualization + +We now provide an in-depth ability analysis of the proposed method via representation visualization using t-SNE [37], a common dimensionality reduction technique. As shown in + +![](images/ab506ec2940b2bee47a0333b246301000e144a6f2b9ba3cf857220fed241c919.jpg) +Figure 3. Visualization of representation extracted from the global model on the PACS dataset. + +![](images/58675cc41bdd3feae2797f6bfb79fc42ce236980c17ffcee28fb3c14d4fee8af.jpg) +Figure 3, we compare the representations extracted from the global model obtained from our gPerXAN and the baseline FedAvg on the PACS dataset with the testing domain, a.k.a. target domain Photo (P). From this figure, we can observe that features derived from our method are semantically separated according to 7 categories in the PACS dataset for both source domains and the target domain. Moreover, features of each category on all domains tend to be close and grouped, demonstrating the ability of our method to distill more discriminative and domain-invariant representation, leading to a significant improvement of more than $2\%$ in classification accuracy on the testing domain Photo (P). + +# 5.2. Privacy and Efficiency + +Data privacy is a major concern in the field of FL, which is mitigated by only sharing client models' parameters instead of raw data. However, to resolve the FedDG problem, recently introduced methods sacrifice this principle by revealing partial information about client data. Specifically, leaking more information not only opens higher chances for attackers to perform an inversion attack [24], which aims to reconstruct the original data of clients, but also amplifies the risk of membership inference attack [32], which determines if a sample was in the model's training data. Furthermore, efficiency in communication and computation is critical in FL, especially in scenarios where resource-constrained clients are involved. Despite that, the common sharing mechanism in available methods introduces unacceptable extra costs to the FL training process. For COPA [42], using an ensemble of domain-specific classifiers in the model architecture results in an $O(N^2)$ increase in communication and computation complexities compared to $O(N)$ as conventional, according to the number of participating clients $N$ . Meanwhile, FedDG-GA [46] consumes a double of memory at clients. Compared with existing ones, in ad + +dition to bypassing the limitations described above, our proposed method is more practical in terms of implementation yet provides competitive results in extensive evaluations. + +# 6. Conclusion + +In this paper, we introduce a novel architectural method, namely gPerXAN, to address the problem of FedDG. By explicitly assembling Instance Normalization layers into Batch Normalization layers in a personalized scheme and employing a simple yet effective guiding regularizer, our method allows the model to filter domain-specific features and actively distill domain-invariant representation for classification tasks. We conduct extensive experiments and in-depth analysis to quantitatively and qualitatively verify the effectiveness of our proposed method in solving this particular problem. Although our evaluation is more on cross-silo FL, our method can be easily extended to cross-device scenarios while Algorithm 1 remains unchanged. Unlike existing methods, due to the independence from imaging techniques, gPerXAN can be straightforwardly extended to diverse applications. A potential limitation of our method is its certain suitability with normalization-based models. In the future, investigating other forms of regularization terms is a promising research direction due to the available room for improvements. Moreover, the vulnerability of available methods under various attacks remains underexplored. + +# Acknowledgement + +This work is supported by the Deutsche Forschungsgemeinschaft, German Research Foundation under grant number 453130567 (COSMO), by the Horizon Europe Research and Innovation Actions under grant number 101092908 (SmartEdge), and by the Federal Ministry for Education and Research, Germany under grant number 01IS18037A (BIFOLD). + +# References + +[1] Yogesh Balaji, Swami Sankaranarayanan, and Rama Chellappa. MetaReg: Towards Domain Generalization Using Meta-Regularization. Advances in Neural Information Processing Systems, 31, 2018. 1, 3 +[2] Peter Bandi, Oscar Geessink, Quirine Manson, Marcory Van Dijk, Maschenka Balkenhol, Meyke Hermsen, Babak Ehteshami Bejnordi, Byungjae Lee, Kyunghyun Paeng, Aoxiao Zhong, et al. From Detection of Individual Metastases to Classification of Lymph Node Status at the Patient Level: The Camelyon17 Challenge. IEEE Transactions on Medical Imaging, 38(2):550-560, 2018. 5 +[3] Junming Chen, Meirui Jiang, Qi Dou, and Qifeng Chen. Federated Domain Generalization for Image Recognition via Cross-Client Style Transfer. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 361–370, 2023. 2, 3, 5, 6, 7 +[4] Ekin D Cubuk, Barret Zoph, Jonathon Shlens, and Quoc V Le. RandAugment: Practical Automated Data Augmentation With a Reduced Search Space. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, pages 702-703, 2020. 7 +[5] Shi Dong, Ping Wang, and Khushnood Abbas. A Survey on Deep Learning and Its Applications. Computer Science Review, 40:100379, 2021. 1 +[6] Qi Dou, Daniel Coelho de Castro, Konstantinos Kamnitsas, and Ben Glocker. Domain Generalization via Model-Agnostic Learning of Semantic Features. Advances in Neural Information Processing Systems, 32, 2019. 1 +[7] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep Residual Learning for Image Recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 770-778, 2016. 6 +[8] Gao Huang, Zhuang Liu, Laurens Van Der Maaten, and Kilian Q Weinberger. Densely Connected Convolutional Networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4700-4708, 2017. 6 +[9] Xun Huang and Serge Belongie. Arbitrary Style Transfer in Real-Time With Adaptive Instance Normalization. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1501-1510, 2017. 4 +[10] Zeyi Huang, Haohan Wang, Eric P Xing, and Dong Huang. Self-Challenging Improves Cross-Domain Generalization. In Proceedings of the European Conference on Computer Vision (ECCV), pages 124-140. Springer, 2020. 6 +[11] Da Li, Yongxin Yang, Yi-Zhe Song, and Timothy M Hospedales. Deeper, Broader and Artier Domain Generalization. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5542-5550, 2017. 5 +[12] Da Li, Jianshu Zhang, Yongxin Yang, Cong Liu, Yi-Zhe Song, and Timothy M Hospedales. Episodic Training for Domain Generalization. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1446-1455, 2019. 1 +[13] Haoliang Li, Sinno Jialin Pan, Shiqi Wang, and Alex C Kot. Domain Generalization With Adversarial Feature Learning. + +In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5400-5409, 2018. 1 +[14] Tian Li, Anit Kumar Sahu, Manzil Zaheer, Maziar Sanjabi, Ameet Talwalkar, and Virginia Smith. Federated Optimization in Heterogeneous Networks. Proceedings of Machine Learning and Systems, 2:429-450, 2020. 3 +[15] Xiang Li, Kaixuan Huang, Wenhao Yang, Shusen Wang, and Zhihua Zhang. On the Convergence of Fedavg on Non-IID Data. In International Conference on Learning Representations, 2019. 4 +[16] Yanghao Li, Naiyan Wang, Jianping Shi, Jiaying Liu, and Xiaodi Hou. Revisiting Batch Normalization for Practical Domain Adaptation. arXiv preprint arXiv:1603.04779, 2016.4 +[17] Quande Liu, Qi Dou, Lequan Yu, and Pheng Ann Heng. MS-Net: Multi-Site Network for Improving Prostate Segmentation With Heterogeneous MRI Data. IEEE Transactions on Medical Imaging, 39(9):2713-2724, 2020. 4 +[18] Quande Liu, Cheng Chen, Jing Qin, Qi Dou, and Pheng-Ann Heng. FedDG: Federated Domain Generalization on Medical Image Segmentation via Episodic Learning in Continuous Frequency Space. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1013–1023, 2021. 2, 3, 5, 6, 7 +[19] Ping Luo, Jiamin Ren, Zhanglin Peng, Ruimao Zhang, and Jingyu Li. Differentiable Learning-To-Normalize via Switchable Normalization. In International Conference on Learning Representations, 2018. 4 +[20] Divyat Mahajan, Shruti Tople, and Amit Sharma. Domain Generalization Using Causal Matching. In International Conference on Machine Learning, pages 7313-7324. PMLR, 2021. 1 +[21] Brendan McMahan, Eider Moore, Daniel Ramage, Seth Hampson, and Blaise Aguera y Arcas. Communication-Efficient Learning of Deep Networks From Decentralized Data. In Artificial Intelligence and Statistics, pages 1273-1282. PMLR, 2017. 1, 4, 6 +[22] Krikamol Muandet, David Balduzzi, and Bernhard Scholkopf. Domain Generalization via Invariant Feature Representation. In International Conference on Machine Learning, pages 10-18. PMLR, 2013. 3 +[23] Hyeonseob Nam, HyunJae Lee, Jongchan Park, Wonjun Yoon, and Donggeun Yoo. Reducing Domain Gap by Reducing Style Bias. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8690-8699, 2021. 1 +[24] Ngoc-Bao Nguyen, Keshigeyan Chandrasegaran, Milad Abdollahzadeh, and Ngai-Man Cheung. Re-thinking Model Inversion Attacks Against Deep Neural Networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, pages 16384-16393, 2023. 8 +[25] Xingang Pan, Ping Luo, Jianping Shi, and Xiaou Tang. Two at Once: Enhancing Learning and Generalization Capacities via IBN-Net. In Proceedings of the European Conference on Computer Vision (ECCV), pages 464–479, 2018. 1, 2, 3, 5, 7 + +[26] Xingchao Peng, Zijun Huang, Yizhe Zhu, and Kate Saenko. Federated Adversarial Domain Adaptation. In International Conference on Learning Representations, 2020. 3 +[27] Krishna Pillutla, Kshitiz Malik, Abdel-Rahman Mohamed, Mike Rabbat, Maziar Sanjabi, and Lin Xiao. Federated Learning With Partial Model Personalization. In International Conference on Machine Learning, pages 17716-17758. PMLR, 2022. 2, 3, 4 +[28] Joaquin Quinonero-Candela, Masashi Sugiyama, Anton Schwaighofer, and Neil D Lawrence. Dataset Shift in Machine Learning. Mit Press, 2008. 1 +[29] Alexandre Rame, Coretin Dancette, and Matthieu Cord. Fishr: Invariant Gradient Variances for Out-Of-Distribution Generalization. In International Conference on Machine Learning, pages 18347–18377. PMLR, 2022. 1 +[30] Seonguk Seo, Yumin Suh, Dongwan Kim, Geeho Kim, Jongwoo Han, and Bohyung Han. Learning to Optimize Domain Specific Normalization for Domain Generalization. In Proceedings of the European Conference on Computer Vision (ECCV), pages 68-83, 2020. 2, 3, 4, 5, 7 +[31] Neta Shoham, Tomer Avidor, Aviv Keren, Nadav Israel, Daniel Benditkis, Liron Mor-Yosef, and Itai Zeitak. Overcoming Forgetting in Federated Learning on Non-IID Data. arXiv preprint arXiv:1910.07796, 2019. 4 +[32] Reza Shokri, Marco Stronati, Congzheng Song, and Vitaly Shmatikov. membership inference attacks against machine learning models. In 2017 IEEE Symposium on Security and Privacy (SP), pages 3-18. IEEE, 2017. 8 +[33] Baochen Sun and Kate Saenko. Deep Coral: Correlation Alignment for Deep Domain Adaptation. In Proceedings of the European Conference on Computer Vision (ECCV), pages 443-450, 2016. 1, 5 +[34] Baochen Sun, Jiashi Feng, and Kate Saenko. Correlation Alignment for Unsupervised Domain Adaptation. Domain Adaptation in Computer Vision Applications, pages 153-171, 2017. 5 +[35] Benyuan Sun, Hongxing Huo, Yi Yang, and Bo Bai. *PartialFed: Cross-Domain Personalized Federated Learning via Partial Initialization*. Advances in Neural Information Processing Systems, 34:23309-23320, 2021. 2, 3, 4 +[36] Yuwei Sun, Ng Chong, and Hideya Ochiai. Feature Distribution Matching for Federated Domain Generalization. In Asian Conference on Machine Learning, pages 942-957. PMLR, 2023. 3 +[37] Laurens Van der Maaten and Geoffrey Hinton. Visualizing Data Using t-SNE. Journal of Machine Learning Research, 9(11), 2008. 7 +[38] Gael Varoquaux and Veronika Cheplygina. Machine Learning for Medical Imaging: Methodological Failures and Recommendations for the Future. Npj Digital Medicine, 5(1):48, 2022. 7 +[39] Hemanth Venkateswara, Jose Eusebio, Shayok Chakraborty, and Sethuraman Panchanathan. Deep Hashing Network for Unsupervised Domain Adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5018-5027, 2017. 5 +[40] Paul Voigt and Axel Von dem Bussche. The EU General Data Protection Regulation (GDPR). A Practical Guide, 1st Ed., + +Cham: Springer International Publishing, 10(3152676):10-5555, 2017. 2 +[41] Jindong Wang, Cuiling Lan, Chang Liu, Yidong Ouyang, Tao Qin, Wang Lu, Yiqiang Chen, Wenjun Zeng, and Philip Yu. Generalizing to Unseen Domains: A Survey on Domain Generalization. IEEE Transactions on Knowledge and Data Engineering, 2022. 1 +[42] Guile Wu and Shaogang Gong. Collaborative Optimization and Aggregation for Decentralized Domain Generalization and Adaptation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6484-6493, 2021. 3, 5, 6, 7, 8 +[43] Zhenlin Xu, Deyi Liu, Junlin Yang, Colin Raffel, and Marc Niethammer. Robust and Generalizable Visual Representation Learning via Random Convolutions. In International Conference on Learning Representations, 2020. 1 +[44] Fu-En Yang, Yuan-Chia Cheng, Zu-Yun Shiau, and Yu-Chiang Frank Wang. Adversarial Teacher-Student Representation Learning for Domain Generalization. Advances in Neural Information Processing Systems, 34:19448-19460, 2021. 1 +[45] Chun-Han Yao, Boqing Gong, Hang Qi, Yin Cui, Yukun Zhu, and Ming-Hsuan Yang. Federated Multi-Target Domain Adaptation. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1424–1433, 2022. 3 +[46] Ruipeng Zhang, Qinwei Xu, Jiangchao Yao, Ya Zhang, Qi Tian, and Yanfeng Wang. Federated Domain Generalization with Generalization Adjustment. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3954-3963, 2023. 3, 5, 6, 7, 8 +[47] Kaiyang Zhou, Yongxin Yang, Yu Qiao, and Tao Xiang. Domain Generalization With Mixstyle. In International Conference on Learning Representations, 2020. 1, 3, 6 +[48] Weiming Zhuang, AI Sony, and Lingjuan Lyu. FedWon: Triumphing Multi-domain Federated Learning Without Normalization. In International Conference on Learning Representations, 2024. 4 \ No newline at end of file diff --git a/2024/Efficiently Assemble Normalization Layers and Regularization for Federated Domain Generalization/images.zip b/2024/Efficiently Assemble Normalization Layers and Regularization for Federated Domain Generalization/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..36afeca30af3fa423ee53fb64aec2e80644cb6b6 --- /dev/null +++ b/2024/Efficiently Assemble Normalization Layers and Regularization for Federated Domain Generalization/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e32b4256af79b34de3645f41fc6ff90cbb810da1263da3b2ac6752ee23062399 +size 369399 diff --git a/2024/Efficiently Assemble Normalization Layers and Regularization for Federated Domain Generalization/layout.json b/2024/Efficiently Assemble Normalization Layers and Regularization for Federated Domain Generalization/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..a2b40d55e710f254424b5a738b168576cfb55426 --- /dev/null +++ b/2024/Efficiently Assemble Normalization Layers and Regularization for Federated Domain Generalization/layout.json @@ -0,0 +1,7653 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 93, + 103, + 501, + 138 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 93, + 103, + 501, + 138 + ], + "spans": [ + { + "bbox": [ + 93, + 103, + 501, + 138 + ], + "type": "text", + "content": "Efficiently Assemble Normalization Layers and Regularization for Federated Domain Generalization" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 80, + 160, + 515, + 218 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 80, + 160, + 515, + 218 + ], + "spans": [ + { + "bbox": [ + 80, + 160, + 515, + 218 + ], + "type": "text", + "content": "Khiem Le" + }, + { + "bbox": [ + 80, + 160, + 515, + 218 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 80, + 160, + 515, + 218 + ], + "type": "text", + "content": ", Long Ho" + }, + { + "bbox": [ + 80, + 160, + 515, + 218 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 80, + 160, + 515, + 218 + ], + "type": "text", + "content": ", Cuong Do" + }, + { + "bbox": [ + 80, + 160, + 515, + 218 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 80, + 160, + 515, + 218 + ], + "type": "text", + "content": ", Danh Le-Phuoc" + }, + { + "bbox": [ + 80, + 160, + 515, + 218 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 80, + 160, + 515, + 218 + ], + "type": "text", + "content": ", Kok-Seng Wong" + }, + { + "bbox": [ + 80, + 160, + 515, + 218 + ], + "type": "inline_equation", + "content": "^{2,*}" + }, + { + "bbox": [ + 80, + 160, + 515, + 218 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 80, + 160, + 515, + 218 + ], + "type": "text", + "content": " Department of Computer Science and Engineering, University of Notre Dame, IN, USA \n" + }, + { + "bbox": [ + 80, + 160, + 515, + 218 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 80, + 160, + 515, + 218 + ], + "type": "text", + "content": " College of Engineering and Computer Science, VinUniversity, Hanoi, Vietnam \n" + }, + { + "bbox": [ + 80, + 160, + 515, + 218 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 80, + 160, + 515, + 218 + ], + "type": "text", + "content": " Open Distributed Systems, Technical University Berlin, Berlin, Germany" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 73, + 220, + 516, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 73, + 220, + 516, + 232 + ], + "spans": [ + { + "bbox": [ + 73, + 220, + 516, + 232 + ], + "type": "text", + "content": "kle3@nd.edu, dans.lephuoc@tu-berlin.de, {long.ht, cuong.dd, wong.ks}@vinuni.edu.vn" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 143, + 258, + 192, + 272 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 258, + 192, + 272 + ], + "spans": [ + { + "bbox": [ + 143, + 258, + 192, + 272 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 284, + 290, + 584 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 284, + 290, + 584 + ], + "spans": [ + { + "bbox": [ + 47, + 284, + 290, + 584 + ], + "type": "text", + "content": "Domain shift is a formidable issue in Machine Learning that causes a model to suffer from performance degradation when tested on unseen domains. Federated Domain Generalization (FedDG) attempts to train a global model using collaborative clients in a privacy-preserving manner that can generalize well to unseen clients possibly with domain shift. However, most existing FedDG methods either cause additional privacy risks of data leakage or induce significant costs in client communication and computation, which are major concerns in the Federated Learning paradigm. To circumvent these challenges, here we introduce a novel architectural method for FedDG, namely gPerXAN1, which relies on a normalization scheme working with a guiding regularizer. In particular, we carefully design Personalized eXplicitly Assessed Normalization to enforce client models selectively filtering domain-specific features that are biased towards local data while retaining discrimination of those features. Then, we incorporate a simple yet effective regularizer to guide these models in directly capturing domain-invariant representations that the global model's classifier can leverage. Extensive experimental results on two benchmark datasets, i.e., PACS and Office-Home, and a real-world medical dataset, Camelyon17, indicate that our proposed method outperforms other existing methods in addressing this particular problem." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 605, + 128, + 618 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 605, + 128, + 618 + ], + "spans": [ + { + "bbox": [ + 47, + 605, + 128, + 618 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 625, + 287, + 686 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 625, + 287, + 686 + ], + "spans": [ + { + "bbox": [ + 46, + 625, + 287, + 686 + ], + "type": "text", + "content": "Over the past few decades, Machine Learning (ML) has demonstrated remarkable achievements across diverse areas such as Computer Vision, Natural Language and Speech Processing, or Robotics [5]. In general, most ML models rely on an over-simplified assumption, i.e., the train-" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 260, + 547, + 416 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 260, + 547, + 416 + ], + "spans": [ + { + "bbox": [ + 306, + 260, + 547, + 416 + ], + "type": "text", + "content": "ing and testing data are independent and identically distributed, which does not always reflect real-world practices. In practical scenarios where the distribution of testing data diverges from that of training data, the performance of ML models often drops catastrophically due to the domain shift issue [28]. Additionally, obtaining or identifying the testing data before model deployment can be challenging in numerous applications. For instance, in biomedical applications where data characteristics vary across different equipment and institutions, gathering data from all potential domains in advance is impractical. Therefore, it is essential to have a solution that can improve the generalization capability of such ML models to adapt effectively to unseen domains." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 421, + 547, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 421, + 547, + 685 + ], + "spans": [ + { + "bbox": [ + 305, + 421, + 547, + 685 + ], + "type": "text", + "content": "Domain Generalization (DG) has been proposed to address the challenge of training ML models using data from single or multiple source domains with the expectation that these models will perform well on unseen domains [41]. The majority of existing DG methods fall under the category of domain-invariant representation learning approach [13, 20, 25, 29, 33]. This approach relies on a broadly acknowledged assumption that each domain contains its own domain-specific features, which are biased towards spurious relations in the data, and that all domains share domain-invariant features, which are general and robust to any unseen domains. From this assumption, previous works propose methods that remove domain-specific features and distill domain-invariant features to achieve the generalization ability. Alternative approaches for DG encompass data augmentation [23, 43, 44, 47], which involves exposing models to artificially generated domains, and meta-learning [1, 6, 12], an approach that emulates the domain shift during the meta-training phase. However, most methods require a centralized setting where all source domains are collected together. Consequently, these methods cannot be readily expanded to decentralized settings." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 689, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 547, + 715 + ], + "type": "text", + "content": "Federated Learning (FL) [21] is an emerging decentralized learning paradigm widely adopted in various applica" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 693, + 240, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 693, + 240, + 703 + ], + "spans": [ + { + "bbox": [ + 58, + 693, + 240, + 703 + ], + "type": "text", + "content": "*Corresponding author: wong.rs@vinuni.edu.vn" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 59, + 703, + 193, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 703, + 193, + 712 + ], + "spans": [ + { + "bbox": [ + 59, + 703, + 193, + 712 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 59, + 703, + 193, + 712 + ], + "type": "text", + "content": "https://github.com/lhkhiem28/gPerXAN" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6027" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 51, + 69, + 549, + 289 + ], + "blocks": [ + { + "bbox": [ + 51, + 69, + 549, + 289 + ], + "lines": [ + { + "bbox": [ + 51, + 69, + 549, + 289 + ], + "spans": [ + { + "bbox": [ + 51, + 69, + 549, + 289 + ], + "type": "image", + "image_path": "2d3d7aeaf622fa60cb0633c1b35aee5ab8449b09c55d3cafeff7f42fa3f3a00c.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 294, + 461, + 306 + ], + "lines": [ + { + "bbox": [ + 131, + 294, + 461, + 306 + ], + "spans": [ + { + "bbox": [ + 131, + 294, + 461, + 306 + ], + "type": "text", + "content": "Figure 1. An overview of our proposed gPerXAN method for solving the FedDG problem." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 327, + 289, + 614 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 327, + 289, + 614 + ], + "spans": [ + { + "bbox": [ + 46, + 327, + 289, + 614 + ], + "type": "text", + "content": "tions to cope with the increasing privacy concerns of data centralization [40]. Specifically, the paradigm works in a way that each client learns from their data and only aggregates local models' parameters at a certain frequency at the central server to generate a global model. Notably, all data samples are kept within each client during the FL training process. Due to the nature of data decentralization, where each client owns a single source domain, as illustrated in Figure 1, the FL paradigm poses further significant challenges for DG and limits the applicability of available DG methods. There have been some early attempts to address the DG problem in the FL scenario. For instance, Liu et al. [18] introduces a method that allows clients to share their image data in the frequency space with each other, thus relatively recovering the centralization process at each client. Similarly, Chen et al. [3] introduces another method that extracts and exchanges the style of local images among all clients. It is evident that these initial efforts employ a strategy that necessitates the sharing of partial client data, thereby compromising the data privacy constraints of FL to a certain extent. Although they show promising results, these methods can be overly complicated to implement in practice and lead to additional privacy risks during the FL training process." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 617, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 617, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 617, + 288, + 714 + ], + "type": "text", + "content": "To address the aforementioned challenges, this paper introduces a novel architectural method for domain-invariant representation learning within the FL framework. The proposed method enhances the generalization ability while upholding the fundamental privacy principles of FL. Based on the effectiveness of discarding domain-specific information from learned features [25, 30], we properly assemble Instance Normalization layers (IN) into Batch Normaliza" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 304, + 327, + 547, + 519 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 327, + 547, + 519 + ], + "spans": [ + { + "bbox": [ + 304, + 327, + 547, + 519 + ], + "type": "text", + "content": "tion layers (BN) in well-known Convolutional Neural Networks (CNNs) using an explicit differential mixture as in Eqn (2). Moreover, thanks to the explicit property, the benefit of personalization in FL [27, 35] can be incorporated into the normalization scheme using local BN sides. Specifically, during the FL training process, while IN sides are globally aggregated along with other model parameters, BN sides are updated locally without broadcasting. In addition, we argue that only relying on the ability to filter domain-specific features of IN while lacking guidance to distill domain-invariant representations directly might lead to suboptimal performance. Based on this observation, we introduce a simple yet highly effective regularization term to guide client models to directly capture domain-invariant representations that can be used by the global model's classifier, which is aggregated from client models' classifiers." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 305, + 525, + 545, + 550 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 525, + 545, + 550 + ], + "spans": [ + { + "bbox": [ + 305, + 525, + 545, + 550 + ], + "type": "text", + "content": "To summarize, our main contributions in this paper are highlighted as follows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 309, + 557, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 309, + 557, + 545, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 309, + 557, + 545, + 617 + ], + "spans": [ + { + "bbox": [ + 309, + 557, + 545, + 617 + ], + "type": "text", + "content": "- Different from existing methods for DG in the FL scenario, we propose a novel method that concentrates on a personalized normalization scheme, global IN while local BN, for filtering domain-specific features and fully respecting the privacy-preserving principles of FL." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 309, + 618, + 545, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 309, + 618, + 545, + 666 + ], + "spans": [ + { + "bbox": [ + 309, + 618, + 545, + 666 + ], + "type": "text", + "content": "- Furthermore, we propose a simple yet effective regularization term to introduce clear guidance to client models for directly capturing domain-invariant representations, further improving performance on unseen domains." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 309, + 666, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 309, + 666, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 309, + 666, + 545, + 713 + ], + "type": "text", + "content": "- Finally, we conduct extensive experiments on two benchmark datasets, i.e., PACS and Office-Home, and a real-world medical dataset, Camelyon17, where our proposed method outperforms existing relevant ones." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6028" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 71, + 134, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 71, + 134, + 83 + ], + "spans": [ + { + "bbox": [ + 47, + 71, + 134, + 83 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 91, + 179, + 103 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 91, + 179, + 103 + ], + "spans": [ + { + "bbox": [ + 47, + 91, + 179, + 103 + ], + "type": "text", + "content": "2.1. Domain Generalization" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 110, + 289, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 110, + 289, + 422 + ], + "spans": [ + { + "bbox": [ + 47, + 110, + 289, + 422 + ], + "type": "text", + "content": "Domain Generalization is a challenging task requiring models to perform well on unseen domains. One dominant approach is domain-invariant representation learning, which aims to minimize the discrepancy among source domains, assuming that the resulting representation will be domain-invariant and generalize well on testing domains. Along this track, Muandet et al. [22] proposes to reduce the domain dissimilarity by using Maximum Mean Discrepancy. Pan et al. [25] combines IN and BN in popular CNNs to reap the benefits of removing domain-specific features while maintaining the ability to capture discriminative features. Meanwhile, Seo et al. [30] discloses that combining these normalization layers in a switchable mechanism can yield better performance. Another approach to DG is data augmentation, which augments source domains to a broader span of the training data space, enlarging the possibility of covering the span of the data in the testing domain. Zhou et al. [47] mixes styles of training instances, resulting in novel domains being synthesized implicitly. Recently, the meta-learning approach has drawn increasing attention from the DG community. Balaji et al. [1] proposes MetaReg that learns a regularization function, particularly for the network classification layer, while excluding the feature extractor. However, the majority of these methods rely on having access to a diverse set of source domains, making them not applicable to decentralized settings." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 431, + 230, + 443 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 431, + 230, + 443 + ], + "spans": [ + { + "bbox": [ + 47, + 431, + 230, + 443 + ], + "type": "text", + "content": "2.2. Federated Domain Generalization" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 450, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 450, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 47, + 450, + 289, + 715 + ], + "type": "text", + "content": "While there is a growing interest in addressing the DG problem within the FL framework, particularly in scenarios where each client possesses a single source domain, the existing literature on this subject remains relatively sparse. This FedDG problem was initially introduced by Liu et al. [18], who then proposed ELCFS that involves exchanging amplitude information in the frequency space among clients and leveraging episodic learning to further enhance performance. Chen et al. [3] proposes a similar mechanism CCST that extracts and exchanges the overall domain style of local images among all clients. Unfortunately, these early works require sharing partial information about the local data, which can be viewed as a form of data leakage and is undesirable in the FL setting. Moreover, these methods include broadcasting partial information and interpolating this information into new training data, which adds significant extra cost to communication and computation during the FL training process. Besides, Wu and Gong [42] proposes an architectural method COPA in which the global model consists of a domain-invariant representation extractor and an ensemble of domain-specific classifiers. Sun et al. [36] introduces FedKA that employs a server-side voting mech" + } + ] + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 305, + 70, + 547, + 169 + ], + "blocks": [ + { + "bbox": [ + 305, + 70, + 547, + 169 + ], + "lines": [ + { + "bbox": [ + 305, + 70, + 547, + 169 + ], + "spans": [ + { + "bbox": [ + 305, + 70, + 547, + 169 + ], + "type": "table", + "html": "
MethodPrivacy RiskAdditional Cost
CommunicationComputation
ELCFS✓✓✓
CCST✓✓✓
COPA×✓✓
FedDG-GA××✓
gPerXAN (Ours)×××
", + "image_path": "acf83ca6a95d175a448784fcc11b2cd37e13bb4b21d88a8763975ce8fc09acc0.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 323, + 177, + 528, + 188 + ], + "lines": [ + { + "bbox": [ + 323, + 177, + 528, + 188 + ], + "spans": [ + { + "bbox": [ + 323, + 177, + 528, + 188 + ], + "type": "text", + "content": "Table 1. An advantage comparison of different methods." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 209, + 547, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 209, + 547, + 399 + ], + "spans": [ + { + "bbox": [ + 304, + 209, + 547, + 399 + ], + "type": "text", + "content": "anism that generates target domain pseudo-labels based on the consensus from clients to facilitate global model finetuning. Recently, Zhang et al. [46] proposes a novel global objective incorporating a variance reduction regularizer to encourage fairness, and then FedDG-GA is proposed to optimize this objective by dynamically calibrating the aggregation weights. Although relaxing from the data leakage issue, these methods also cause a large additional consumption of resources when the number of source domains and output classes increases. In this paper, we introduce an alternative architectural method that shares only the model update information during training. This ensures maximal data privacy and circumvents the communication and computation overhead issue mentioned earlier while achieving competitive results. In Table 1, we highlight the advantages of our method in comparison to previous ones." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 400, + 547, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 400, + 547, + 509 + ], + "spans": [ + { + "bbox": [ + 304, + 400, + 547, + 509 + ], + "type": "text", + "content": "One related research field to FedDG is Federated Domain Adaptation (FedDA). While both fields aim to maximize the model performance on unseen domains using existing source domains, FedDA can access the target domains while FedDG cannot see those data during training. Leveraging this assumption, FedDA methods such as FADA [26] or FMTDA [45], are typically able to align target and source features. However, this specific assumption also makes FedDA methods are inapplicable to FedDG." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 515, + 484, + 528 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 515, + 484, + 528 + ], + "spans": [ + { + "bbox": [ + 306, + 515, + 484, + 528 + ], + "type": "text", + "content": "2.3. Personalized Federated Learning" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 534, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 534, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 534, + 547, + 713 + ], + "type": "text", + "content": "Another related and orthogonal line of work is Personalized Federated Learning (pFL). pFL aims to learn personalized models for different clients to tackle distribution shifts across client data, known as data heterogeneity. Li et al. [14] utilizes a regularization term in the local loss function so that the clients' trained models will not significantly differ from the global model. On the other side, Sun et al. [35] and Pillutla et al. [27] mitigate the heterogeneous distributions by only loading a subset of the global model's parameters rather than loading the entire model at each training round. Although related in terms of overcoming distribution shifts across clients, pFL focuses on improving the performance of participating clients, whereas our considered field FedDG focuses on improving the performance of unseen clients with unseen domains." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6029" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 71, + 129, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 71, + 129, + 85 + ], + "spans": [ + { + "bbox": [ + 47, + 71, + 129, + 85 + ], + "type": "text", + "content": "3. Methodology" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 91, + 287, + 193 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 91, + 287, + 193 + ], + "spans": [ + { + "bbox": [ + 47, + 91, + 287, + 193 + ], + "type": "text", + "content": "Problem Formulation. First, we denote " + }, + { + "bbox": [ + 47, + 91, + 287, + 193 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 47, + 91, + 287, + 193 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 91, + 287, + 193 + ], + "type": "inline_equation", + "content": "Y" + }, + { + "bbox": [ + 47, + 91, + 287, + 193 + ], + "type": "text", + "content": " as the input space and the label space, respectively, of a specific task " + }, + { + "bbox": [ + 47, + 91, + 287, + 193 + ], + "type": "inline_equation", + "content": "\\mathcal{T}" + }, + { + "bbox": [ + 47, + 91, + 287, + 193 + ], + "type": "text", + "content": ". In the standard FL setting, " + }, + { + "bbox": [ + 47, + 91, + 287, + 193 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 47, + 91, + 287, + 193 + ], + "type": "text", + "content": " clients " + }, + { + "bbox": [ + 47, + 91, + 287, + 193 + ], + "type": "inline_equation", + "content": "\\{c_i\\}_{i=1}^N" + }, + { + "bbox": [ + 47, + 91, + 287, + 193 + ], + "type": "text", + "content": " are involved in collaboratively constructing a global model " + }, + { + "bbox": [ + 47, + 91, + 287, + 193 + ], + "type": "inline_equation", + "content": "f_g" + }, + { + "bbox": [ + 47, + 91, + 287, + 193 + ], + "type": "text", + "content": " for solving the task, where each client " + }, + { + "bbox": [ + 47, + 91, + 287, + 193 + ], + "type": "inline_equation", + "content": "c_i" + }, + { + "bbox": [ + 47, + 91, + 287, + 193 + ], + "type": "text", + "content": " owns a dataset " + }, + { + "bbox": [ + 47, + 91, + 287, + 193 + ], + "type": "inline_equation", + "content": "D_i = \\{(x_k, y_k)\\}_{k=1,|D_i|}" + }, + { + "bbox": [ + 47, + 91, + 287, + 193 + ], + "type": "text", + "content": " which is associated to a specific domain defined by a joint distribution " + }, + { + "bbox": [ + 47, + 91, + 287, + 193 + ], + "type": "inline_equation", + "content": "P_{X,Y}^{(i)}" + }, + { + "bbox": [ + 47, + 91, + 287, + 193 + ], + "type": "text", + "content": ". Importantly, scattered datasets " + }, + { + "bbox": [ + 47, + 91, + 287, + 193 + ], + "type": "inline_equation", + "content": "\\{D_i\\}_{i=1}^N" + }, + { + "bbox": [ + 47, + 91, + 287, + 193 + ], + "type": "text", + "content": " across clients satisfy:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 196, + 235, + 236 + ], + "type": "list", + "angle": 0, + "index": 4, + "blocks": [ + { + "bbox": [ + 48, + 196, + 235, + 212 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 196, + 235, + 212 + ], + "spans": [ + { + "bbox": [ + 48, + 196, + 235, + 212 + ], + "type": "text", + "content": "- " + }, + { + "bbox": [ + 48, + 196, + 235, + 212 + ], + "type": "inline_equation", + "content": "P_{X,Y}^{(i)} \\neq P_{X,Y}^{(j)}" + }, + { + "bbox": [ + 48, + 196, + 235, + 212 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 48, + 196, + 235, + 212 + ], + "type": "inline_equation", + "content": "1 \\leq i, j \\leq N" + }, + { + "bbox": [ + 48, + 196, + 235, + 212 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 48, + 196, + 235, + 212 + ], + "type": "inline_equation", + "content": "i \\neq j" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 219, + 235, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 219, + 235, + 236 + ], + "spans": [ + { + "bbox": [ + 48, + 219, + 235, + 236 + ], + "type": "text", + "content": "- " + }, + { + "bbox": [ + 48, + 219, + 235, + 236 + ], + "type": "inline_equation", + "content": "P_{Y|X}^{(i)} = P_{Y|X}^{(j)}" + }, + { + "bbox": [ + 48, + 219, + 235, + 236 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 48, + 219, + 235, + 236 + ], + "type": "inline_equation", + "content": "1 \\leq i, j \\leq N" + }, + { + "bbox": [ + 48, + 219, + 235, + 236 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 48, + 219, + 235, + 236 + ], + "type": "inline_equation", + "content": "i \\neq j" + }, + { + "bbox": [ + 48, + 219, + 235, + 236 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 47, + 239, + 287, + 275 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 239, + 287, + 275 + ], + "spans": [ + { + "bbox": [ + 47, + 239, + 287, + 275 + ], + "type": "text", + "content": "The objective of FedDG is leveraging scattered datasets to construct a model " + }, + { + "bbox": [ + 47, + 239, + 287, + 275 + ], + "type": "inline_equation", + "content": "f_{g}" + }, + { + "bbox": [ + 47, + 239, + 287, + 275 + ], + "type": "text", + "content": " that can directly generalize to the unseen dataset " + }, + { + "bbox": [ + 47, + 239, + 287, + 275 + ], + "type": "inline_equation", + "content": "D_{U}" + }, + { + "bbox": [ + 47, + 239, + 287, + 275 + ], + "type": "text", + "content": " with an unseen domain, which means:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 277, + 185, + 317 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 48, + 277, + 185, + 293 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 277, + 185, + 293 + ], + "spans": [ + { + "bbox": [ + 48, + 277, + 185, + 293 + ], + "type": "text", + "content": "- " + }, + { + "bbox": [ + 48, + 277, + 185, + 293 + ], + "type": "inline_equation", + "content": "P_{X,Y}^{(U)} \\neq P_{X,Y}^{(i)}" + }, + { + "bbox": [ + 48, + 277, + 185, + 293 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 48, + 277, + 185, + 293 + ], + "type": "inline_equation", + "content": "1 \\leq i \\leq N" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 300, + 185, + 317 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 300, + 185, + 317 + ], + "spans": [ + { + "bbox": [ + 48, + 300, + 185, + 317 + ], + "type": "text", + "content": "- " + }, + { + "bbox": [ + 48, + 300, + 185, + 317 + ], + "type": "inline_equation", + "content": "P_{Y|X}^{(U)} = P_{Y|X}^{(i)}" + }, + { + "bbox": [ + 48, + 300, + 185, + 317 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 48, + 300, + 185, + 317 + ], + "type": "inline_equation", + "content": "1 \\leq i \\leq N" + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 47, + 320, + 287, + 427 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 320, + 287, + 427 + ], + "spans": [ + { + "bbox": [ + 47, + 320, + 287, + 427 + ], + "type": "text", + "content": "To this end, " + }, + { + "bbox": [ + 47, + 320, + 287, + 427 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 47, + 320, + 287, + 427 + ], + "type": "text", + "content": " clients communicate with a central server for " + }, + { + "bbox": [ + 47, + 320, + 287, + 427 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 47, + 320, + 287, + 427 + ], + "type": "text", + "content": " rounds. At each round, every client " + }, + { + "bbox": [ + 47, + 320, + 287, + 427 + ], + "type": "inline_equation", + "content": "c_{i}" + }, + { + "bbox": [ + 47, + 320, + 287, + 427 + ], + "type": "text", + "content": " receives the same global model " + }, + { + "bbox": [ + 47, + 320, + 287, + 427 + ], + "type": "inline_equation", + "content": "f_{g}" + }, + { + "bbox": [ + 47, + 320, + 287, + 427 + ], + "type": "text", + "content": " from the server and updates " + }, + { + "bbox": [ + 47, + 320, + 287, + 427 + ], + "type": "inline_equation", + "content": "f_{g}" + }, + { + "bbox": [ + 47, + 320, + 287, + 427 + ], + "type": "text", + "content": " with their local dataset " + }, + { + "bbox": [ + 47, + 320, + 287, + 427 + ], + "type": "inline_equation", + "content": "D_{i}" + }, + { + "bbox": [ + 47, + 320, + 287, + 427 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 47, + 320, + 287, + 427 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 47, + 320, + 287, + 427 + ], + "type": "text", + "content": " epochs to establish its local model " + }, + { + "bbox": [ + 47, + 320, + 287, + 427 + ], + "type": "inline_equation", + "content": "f_{i}" + }, + { + "bbox": [ + 47, + 320, + 287, + 427 + ], + "type": "text", + "content": ". The server then collects all trained models and aggregates them to update the global model. This process repeats until the global model converges. In this work, we consider the most popular FL framework, FedAvg [21], which aggregates client models as:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 111, + 436, + 287, + 468 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 111, + 436, + 287, + 468 + ], + "spans": [ + { + "bbox": [ + 111, + 436, + 287, + 468 + ], + "type": "interline_equation", + "content": "f _ {g} = \\sum_ {i = 1} ^ {N} \\frac {\\left| D _ {i} \\right|}{\\sum_ {1 \\leq j \\leq N} \\left| D _ {j} \\right|} f _ {i} \\tag {1}", + "image_path": "aded45a40633618f37eff9e762ee4e24255c311bd47a36b3aa4d7690be41cf29.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 475, + 287, + 596 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 475, + 287, + 596 + ], + "spans": [ + { + "bbox": [ + 47, + 475, + 287, + 596 + ], + "type": "text", + "content": "Challenges. In the general spirit of DG, a model is expected to extensively explore multiple source domains to achieve domain-invariance in its learned latent representation. However, under the FL setting, each client is restricted to accessing only its own local data, which constrains to make full use of source domains and, consequently, limits learning of generalizable representation. Moreover, sharing data or even partial information about data among clients poses additional privacy risks and introduces communication and computation costs during the FL training process." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 47, + 602, + 176, + 615 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 602, + 176, + 615 + ], + "spans": [ + { + "bbox": [ + 47, + 602, + 176, + 615 + ], + "type": "text", + "content": "3.1. Normalization Scheme" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 47, + 621, + 287, + 669 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 621, + 287, + 669 + ], + "spans": [ + { + "bbox": [ + 47, + 621, + 287, + 669 + ], + "type": "text", + "content": "eXplicitly Assessed Normalization. The first step towards solving the highlighted challenges, we introduce a novel normalization scheme called eXplicitly Assessed Normalization (XAN), which combines IN and BN as:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 47, + 676, + 287, + 713 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 676, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 676, + 287, + 713 + ], + "type": "interline_equation", + "content": "\\hat {h} = w _ {i n} \\left(\\gamma_ {i n} \\frac {h - \\mu_ {i n}}{\\sqrt {\\sigma_ {i n} ^ {2} + \\epsilon}} + \\beta_ {i n}\\right) + w _ {b n} \\left(\\gamma_ {b n} \\frac {h - \\mu_ {b n}}{\\sqrt {\\sigma_ {b n} ^ {2} + \\epsilon}} + \\beta_ {b n}\\right), \\tag {2}", + "image_path": "ee33ac02daba31b95f600e71ccbe65221d0db23f585907c3fb0ecbf9a3148951.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 71, + 545, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 71, + 545, + 144 + ], + "spans": [ + { + "bbox": [ + 305, + 71, + 545, + 144 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 71, + 545, + 144 + ], + "type": "inline_equation", + "content": "h, \\hat{h} \\in \\mathbb{R}^{B \\times C \\times W \\times H}" + }, + { + "bbox": [ + 305, + 71, + 545, + 144 + ], + "type": "text", + "content": " are layer input and output activations, which are 4D tensors with dimensions of batch size " + }, + { + "bbox": [ + 305, + 71, + 545, + 144 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 305, + 71, + 545, + 144 + ], + "type": "text", + "content": ", number of channels " + }, + { + "bbox": [ + 305, + 71, + 545, + 144 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 305, + 71, + 545, + 144 + ], + "type": "text", + "content": ", width " + }, + { + "bbox": [ + 305, + 71, + 545, + 144 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 305, + 71, + 545, + 144 + ], + "type": "text", + "content": " and height " + }, + { + "bbox": [ + 305, + 71, + 545, + 144 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 305, + 71, + 545, + 144 + ], + "type": "text", + "content": ". Meanwhile, " + }, + { + "bbox": [ + 305, + 71, + 545, + 144 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 305, + 71, + 545, + 144 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 71, + 545, + 144 + ], + "type": "inline_equation", + "content": "\\sigma^2" + }, + { + "bbox": [ + 305, + 71, + 545, + 144 + ], + "type": "text", + "content": " are means and variances captured by the normalization layers, respectively, " + }, + { + "bbox": [ + 305, + 71, + 545, + 144 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 305, + 71, + 545, + 144 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 71, + 545, + 144 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 305, + 71, + 545, + 144 + ], + "type": "text", + "content": " are affine parameters, and " + }, + { + "bbox": [ + 305, + 71, + 545, + 144 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 305, + 71, + 545, + 144 + ], + "type": "text", + "content": " is for numerical stability." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 144, + 546, + 372 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 144, + 546, + 372 + ], + "spans": [ + { + "bbox": [ + 304, + 144, + 546, + 372 + ], + "type": "text", + "content": "In essence, XAN is a sophisticated mixture mechanism of IN and BN to replace BN in the feature extractors of CNNs. In particular, " + }, + { + "bbox": [ + 304, + 144, + 546, + 372 + ], + "type": "inline_equation", + "content": "w_{in}" + }, + { + "bbox": [ + 304, + 144, + 546, + 372 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 144, + 546, + 372 + ], + "type": "inline_equation", + "content": "w_{bn}" + }, + { + "bbox": [ + 304, + 144, + 546, + 372 + ], + "type": "text", + "content": " are ratios to weight the mixture that allow the model to switch between IN and BN. These parameters are randomly initialized and optimized along with other model parameters during training in an end-to-end manner. Unlike previous works [19, 30], which uses an implicit mechanism to combine computed statistics of IN and BN, i.e., means and variances, XAN uses an explicit mechanism that combines the output activations of two normalization layers, this provides the model with a unique ability to separate IN from BN completely. Note that IN has shown great success in style transfer tasks [9] as it allows discarding the variability of visual styles such as object colors or textures from the content of images. This property makes IN beneficial for solving the DG problem. However, directly using IN to replace the conventional BN leads to the loss of discrimination in the learned features, resulting in performance degradation in classification tasks." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 373, + 546, + 648 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 373, + 546, + 648 + ], + "spans": [ + { + "bbox": [ + 304, + 373, + 546, + 648 + ], + "type": "text", + "content": "Personalized eXplicitly Assessed Normalization. To accomplish our proposed normalization scheme, Personalized eXplicitly Assessed Normalization (PerXAN), we make a subtle modification to the conventional FedAvg [21] framework based on the explicit property of the above XAN. Specifically, during the FL training process, while IN sides of XAN layers are globally aggregated along with other model parameters, BN sides are updated locally, which means that parameters of BN sides are excluded from the broadcasting steps from the server. Notably, in inference time, the global model is generated by averaging all model parameters of clients. Our modification is motivated by a common observation that FedAvg normally results in poor convergence and performance in the presence of domain heterogeneity across clients [15, 48] with the major reason being that client models forget the acquired knowledge from previous rounds after aggregated [31]. The leading solution is to personalize a subset of the model, which benefits clients in learning better from their local data [27, 35]. Moreover, based on another finding that BN plays a crucial role in dealing with the domain shift issue in the centralized paradigm [16, 17], BN sides in XAN layers are finally made to be personalized. PerXAN is illustrated in Figure 1." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 306, + 658, + 458, + 670 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 658, + 458, + 670 + ], + "spans": [ + { + "bbox": [ + 306, + 658, + 458, + 670 + ], + "type": "text", + "content": "3.2. Regularization as Guidance" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 306, + 677, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 677, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 677, + 545, + 713 + ], + "type": "text", + "content": "With the proposed normalization scheme PerXAN in place, now, we are ready to introduce our \"Regularization as Guidance\". It targets to induce clear guidance to client" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6030" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 289, + 301 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 289, + 301 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 289, + 301 + ], + "type": "text", + "content": "models through a regularizer, which brings an effect of domain alignment [33, 34], and guides these models to capture domain-invariant representations directly. We hypothesize that this will address a drawback of current DG methods that are later verified by our experiments in Section 4. The drawback is from DG methods that take advantage of the IN's function [25, 30] but do not actually equip the model with the capability of capturing domain-invariant features even if they demonstrated promising performance. Instead, it is only hoped that domain-invariant features would be distilled through achieving the goal of removing domain-specific features. This indirect learning purpose might affect the model's learning efficacy, especially in the FL setting, where each client only owns a separate single source domain. Specifically, we assume a classification model " + }, + { + "bbox": [ + 47, + 72, + 289, + 301 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 47, + 72, + 289, + 301 + ], + "type": "text", + "content": ", either the global model " + }, + { + "bbox": [ + 47, + 72, + 289, + 301 + ], + "type": "inline_equation", + "content": "f_{g}" + }, + { + "bbox": [ + 47, + 72, + 289, + 301 + ], + "type": "text", + "content": " or a client model " + }, + { + "bbox": [ + 47, + 72, + 289, + 301 + ], + "type": "inline_equation", + "content": "f_{i}" + }, + { + "bbox": [ + 47, + 72, + 289, + 301 + ], + "type": "text", + "content": ", comprises a feature extractor " + }, + { + "bbox": [ + 47, + 72, + 289, + 301 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 47, + 72, + 289, + 301 + ], + "type": "text", + "content": " and a classifier head " + }, + { + "bbox": [ + 47, + 72, + 289, + 301 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 47, + 72, + 289, + 301 + ], + "type": "text", + "content": ". At each client " + }, + { + "bbox": [ + 47, + 72, + 289, + 301 + ], + "type": "inline_equation", + "content": "c_{i}" + }, + { + "bbox": [ + 47, + 72, + 289, + 301 + ], + "type": "text", + "content": ", during local training, the client model " + }, + { + "bbox": [ + 47, + 72, + 289, + 301 + ], + "type": "inline_equation", + "content": "f_{i}" + }, + { + "bbox": [ + 47, + 72, + 289, + 301 + ], + "type": "text", + "content": " is optimized on the local dataset " + }, + { + "bbox": [ + 47, + 72, + 289, + 301 + ], + "type": "inline_equation", + "content": "D_{i}" + }, + { + "bbox": [ + 47, + 72, + 289, + 301 + ], + "type": "text", + "content": " using the following loss function:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 58, + 309, + 287, + 362 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 309, + 287, + 362 + ], + "spans": [ + { + "bbox": [ + 58, + 309, + 287, + 362 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} L _ {i} = L _ {c l s} ^ {(i)} \\left(f _ {i}; D _ {i}\\right) + \\lambda L _ {r e g} ^ {(i)} \\left(g _ {i}, h _ {g}; D _ {i}\\right) (3) \\\\ = \\sum_ {k = 1} ^ {| D _ {i} |} \\ell \\left(f _ {i} \\left(x _ {k} ^ {(i)}\\right), y _ {k} ^ {(i)}\\right) + \\lambda \\ell \\left(h _ {g} \\left(g _ {i} \\left(x _ {k} ^ {(i)}\\right)\\right), y _ {k} ^ {(i)}\\right), (4) \\\\ \\end{array}", + "image_path": "d223eb660c17b3e3d8d519e14f58be7e8b2f98fb4837f2fa5687dd1b986a3092.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 372, + 287, + 504 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 372, + 287, + 504 + ], + "spans": [ + { + "bbox": [ + 47, + 372, + 287, + 504 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 372, + 287, + 504 + ], + "type": "inline_equation", + "content": "\\ell" + }, + { + "bbox": [ + 47, + 372, + 287, + 504 + ], + "type": "text", + "content": " is the base loss function, which is usually cross-entropy in classification tasks, and " + }, + { + "bbox": [ + 47, + 372, + 287, + 504 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 47, + 372, + 287, + 504 + ], + "type": "text", + "content": " is a hyper-parameter to control the significance of the regularizer. In a deeper look, by freezing " + }, + { + "bbox": [ + 47, + 372, + 287, + 504 + ], + "type": "inline_equation", + "content": "h_g" + }, + { + "bbox": [ + 47, + 372, + 287, + 504 + ], + "type": "text", + "content": " during local training, the auxiliary loss term " + }, + { + "bbox": [ + 47, + 372, + 287, + 504 + ], + "type": "inline_equation", + "content": "L_{reg}" + }, + { + "bbox": [ + 47, + 372, + 287, + 504 + ], + "type": "text", + "content": " forces client models to arrive at representations that can be made use for classification by the same global classifier, hence, producing an alignment effect on these representations [33, 34]. Moreover, our proposed regularizer can also be interpreted as an implicit form of matching global knowledge to clients' knowledge, which has been demonstrated to bring performance gain [3, 18] for FedDG." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 505, + 288, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 505, + 288, + 613 + ], + "spans": [ + { + "bbox": [ + 46, + 505, + 288, + 613 + ], + "type": "text", + "content": "Due to its orthogonality, our proposed regularizer can be easily integrated into various methods, as shown in Algorithm 1. Specifically, our experiments in Section 4 verify its compatibility with the normalization scheme that justifies our motivation. Moreover, by using only the global model's classifier to regularize feature extractors, gPerXAN saves major communication and computation resources, as well as memory usage at clients compared to others, which use an ensemble of classifiers [42] or entire global model [46]." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 625, + 190, + 639 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 625, + 190, + 639 + ], + "spans": [ + { + "bbox": [ + 47, + 625, + 190, + 639 + ], + "type": "text", + "content": "4. Experiments and Results" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 646, + 110, + 658 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 646, + 110, + 658 + ], + "spans": [ + { + "bbox": [ + 47, + 646, + 110, + 658 + ], + "type": "text", + "content": "4.1. Datasets" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 665, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 288, + 714 + ], + "type": "text", + "content": "To evaluate the proposed method, we perform experiments on two standard DG benchmark datasets, i.e., PACS [11] and Office-Home [39], and a real-world medical image dataset, Camelyon17 [2], consisting of various sub-datasets" + } + ] + } + ], + "index": 6 + }, + { + "type": "code", + "bbox": [ + 310, + 87, + 547, + 412 + ], + "blocks": [ + { + "bbox": [ + 307, + 72, + 403, + 83 + ], + "lines": [ + { + "bbox": [ + 307, + 72, + 403, + 83 + ], + "spans": [ + { + "bbox": [ + 307, + 72, + 403, + 83 + ], + "type": "text", + "content": "Algorithm 1 gPerXAN" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "code_caption" + }, + { + "bbox": [ + 310, + 87, + 547, + 412 + ], + "lines": [ + { + "bbox": [ + 310, + 87, + 547, + 412 + ], + "spans": [ + { + "bbox": [ + 310, + 87, + 547, + 412 + ], + "type": "text", + "content": "1: Input: A model " + }, + { + "bbox": [ + 310, + 87, + 547, + 412 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 310, + 87, + 547, + 412 + ], + "type": "text", + "content": " uses PerXAN to replace BN layers in the feature extractor. " + }, + { + "bbox": [ + 310, + 87, + 547, + 412 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 310, + 87, + 547, + 412 + ], + "type": "text", + "content": " clients with their local datasets " + }, + { + "bbox": [ + 310, + 87, + 547, + 412 + ], + "type": "inline_equation", + "content": "\\{D_i\\}_{i = 1}^N" + }, + { + "bbox": [ + 310, + 87, + 547, + 412 + ], + "type": "text", + "content": " . Notably, " + }, + { + "bbox": [ + 310, + 87, + 547, + 412 + ], + "type": "inline_equation", + "content": "f^{(t)}" + }, + { + "bbox": [ + 310, + 87, + 547, + 412 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 310, + 87, + 547, + 412 + ], + "type": "inline_equation", + "content": "f^{(l;t)}" + }, + { + "bbox": [ + 310, + 87, + 547, + 412 + ], + "type": "text", + "content": " is the model " + }, + { + "bbox": [ + 310, + 87, + 547, + 412 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 310, + 87, + 547, + 412 + ], + "type": "text", + "content": " and its " + }, + { + "bbox": [ + 310, + 87, + 547, + 412 + ], + "type": "inline_equation", + "content": "l^{th}" + }, + { + "bbox": [ + 310, + 87, + 547, + 412 + ], + "type": "text", + "content": " layer at the communication round " + }, + { + "bbox": [ + 310, + 87, + 547, + 412 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 310, + 87, + 547, + 412 + ], + "type": "text", + "content": " , respectively. \n2: Initialization: " + }, + { + "bbox": [ + 310, + 87, + 547, + 412 + ], + "type": "inline_equation", + "content": "f_{g}^{(0)}\\gets f" + }, + { + "bbox": [ + 310, + 87, + 547, + 412 + ], + "type": "text", + "content": " \n3: for each round " + }, + { + "bbox": [ + 310, + 87, + 547, + 412 + ], + "type": "inline_equation", + "content": "t = 1,2,3,\\ldots ,T" + }, + { + "bbox": [ + 310, + 87, + 547, + 412 + ], + "type": "text", + "content": " do \n4: for each client " + }, + { + "bbox": [ + 310, + 87, + 547, + 412 + ], + "type": "inline_equation", + "content": "i = 1,2,3,\\ldots ,N" + }, + { + "bbox": [ + 310, + 87, + 547, + 412 + ], + "type": "text", + "content": " in parallel do \n5: for each layer in " + }, + { + "bbox": [ + 310, + 87, + 547, + 412 + ], + "type": "inline_equation", + "content": "f_{i}^{(t)}" + }, + { + "bbox": [ + 310, + 87, + 547, + 412 + ], + "type": "text", + "content": " do \n6: if " + }, + { + "bbox": [ + 310, + 87, + 547, + 412 + ], + "type": "inline_equation", + "content": "f_{i}^{(l;t)}" + }, + { + "bbox": [ + 310, + 87, + 547, + 412 + ], + "type": "text", + "content": " is not BN then \n7: " + }, + { + "bbox": [ + 310, + 87, + 547, + 412 + ], + "type": "inline_equation", + "content": "f_{i}^{(l;t)}\\gets f_{g}^{(l;t)}" + }, + { + "bbox": [ + 310, + 87, + 547, + 412 + ], + "type": "text", + "content": " \n8: end if \n9: end for \n10: " + }, + { + "bbox": [ + 310, + 87, + 547, + 412 + ], + "type": "inline_equation", + "content": "f_{i}^{(t)}\\gets \\mathrm{LocalTraining}(f_{i}^{(t)},h_{g};D_{i})" + }, + { + "bbox": [ + 310, + 87, + 547, + 412 + ], + "type": "text", + "content": " \n11: end for \n12: " + }, + { + "bbox": [ + 310, + 87, + 547, + 412 + ], + "type": "inline_equation", + "content": "f_{g}^{(t)} = \\sum_{i = 1}^{N}\\frac{|D_{i}|}{\\sum_{1\\leq j\\leq N}|D_{j}|} f_{i}^{(t)}" + }, + { + "bbox": [ + 310, + 87, + 547, + 412 + ], + "type": "text", + "content": " // Eqn (2) \n13: end for \n14: return: " + }, + { + "bbox": [ + 310, + 87, + 547, + 412 + ], + "type": "inline_equation", + "content": "f_{g}^{(T)}" + }, + { + "bbox": [ + 310, + 87, + 547, + 412 + ], + "type": "text", + "content": " \n15: LocalTraining " + }, + { + "bbox": [ + 310, + 87, + 547, + 412 + ], + "type": "inline_equation", + "content": "(f_i,h_g;D_i)" + }, + { + "bbox": [ + 310, + 87, + 547, + 412 + ], + "type": "text", + "content": " \n16: for each epoch " + }, + { + "bbox": [ + 310, + 87, + 547, + 412 + ], + "type": "inline_equation", + "content": "e = 1,2,3,\\dots ,E" + }, + { + "bbox": [ + 310, + 87, + 547, + 412 + ], + "type": "text", + "content": " do \n17: " + }, + { + "bbox": [ + 310, + 87, + 547, + 412 + ], + "type": "inline_equation", + "content": "L_{i} = L_{cls}^{(i)}(f_{i};D_{i}) + \\lambda L_{reg}^{(i)}(g_{i},h_{g};D_{i})" + }, + { + "bbox": [ + 310, + 87, + 547, + 412 + ], + "type": "text", + "content": " // Eqn (3) \n18: " + }, + { + "bbox": [ + 310, + 87, + 547, + 412 + ], + "type": "inline_equation", + "content": "f_{i}\\gets f_{i} - \\eta \\nabla L_{i}" + }, + { + "bbox": [ + 310, + 87, + 547, + 412 + ], + "type": "text", + "content": " \n19: end for \n20: return: " + }, + { + "bbox": [ + 310, + 87, + 547, + 412 + ], + "type": "inline_equation", + "content": "f_{i}" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "code_body" + } + ], + "index": 8, + "sub_type": "algorithm" + }, + { + "bbox": [ + 304, + 434, + 547, + 591 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 434, + 547, + 591 + ], + "spans": [ + { + "bbox": [ + 304, + 434, + 547, + 591 + ], + "type": "text", + "content": "that are considered as domains. Specifically, PACS is composed of 4 domains with large discrepancies from diverse image colors and textures, *Photo* (P), *Art painting* (A), *Cartoon* (C), and *Sketch* (S). Each domain contains 7 categories, with 9,991 images in total. Office-Home is also composed of 4 domains but with smaller discrepancies from various backgrounds and camera viewpoints, *Product* (P), *Art* (A), *Clipart* (C), and *Real-world* (R). Each domain contains a more extensive label set of 65 categories with 15,588 images. Camelyon17 is a binary tumor classification dataset containing 455,964 histology images with stains from 5 different hospitals worldwide. These datasets are at different difficulty levels and are commonly used in the literature." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 598, + 433, + 612 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 598, + 433, + 612 + ], + "spans": [ + { + "bbox": [ + 305, + 598, + 433, + 612 + ], + "type": "text", + "content": "4.2. Experimental Settings" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "content": "Evaluation. For a fair comparison purpose, we follow the common leave-one-domain-out evaluation protocol as considered in [3, 18, 46]. In particular, we sequentially choose one domain as the unseen domain, train the model on all remaining domains where a single domain is treated as a client, and evaluate the trained model on the chosen domain. For PACS and Office-Home datasets, we split " + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "inline_equation", + "content": "90\\%" + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "content": " of the data of each source client as the training set and " + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "inline_equation", + "content": "10\\%" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "6031" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 47, + 70, + 548, + 244 + ], + "blocks": [ + { + "bbox": [ + 47, + 70, + 548, + 244 + ], + "lines": [ + { + "bbox": [ + 47, + 70, + 548, + 244 + ], + "spans": [ + { + "bbox": [ + 47, + 70, + 548, + 244 + ], + "type": "table", + "html": "
MethodPACSOffice-Home
PACSAvgPACRAvgAvg
FedAvg (Baseline)95.2182.2378.2073.5682.3076.5365.9755.4078.0168.9875.64
FedAvg w/ MixStyle95.9385.9980.0375.4684.3575.8762.0957.9277.4868.3476.35
FedAvg w/ RSC95.2183.1578.2474.6282.8175.2662.3450.7977.4666.4674.63
ELCFS96.2383.9479.2773.3083.1976.8366.3255.6378.1269.2376.21
CCST96.6588.3378.2082.9086.5276.6166.3552.3980.0168.8477.68
COPA95.6284.8080.2882.8685.8975.8262.2756.0478.7268.2177.05
FedDG-GA96.8086.9181.2382.7486.9277.2365.1058.2978.8069.8678.42
gPerXAN (Ours)97.2786.5284.6883.2887.9478.9167.2457.7580.1571.0179.48
", + "image_path": "2b3d2ddbec3e495913dc3a46b19ec0add0a8306b26dde5910d05d9d70166c284.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 285, + 289, + 501 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 285, + 289, + 501 + ], + "spans": [ + { + "bbox": [ + 46, + 285, + 289, + 501 + ], + "type": "text", + "content": "of that as the validation set, while for unseen clients, the entire data is used for testing. For the large-scale Camelyon17 dataset, the ratios of training set and validation set at each source client are " + }, + { + "bbox": [ + 46, + 285, + 289, + 501 + ], + "type": "inline_equation", + "content": "80\\%" + }, + { + "bbox": [ + 46, + 285, + 289, + 501 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 285, + 289, + 501 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 46, + 285, + 289, + 501 + ], + "type": "text", + "content": ", respectively, while the entire data is used for testing at unseen clients. In all experiments, we report the test accuracy on each unseen client by using the best validation model selected based on the average of accuracies on validation sets of source clients. The reported numerical values are averaged over 3 runs. Finally, we straightforwardly compare the proposed method with the vanilla FedAvg [21]. Two centralized DG methods, which are free from the requirement of data centralization, MixStyle [47] and RSC [10], are also evaluated under the integration into the FedAvg framework. Furthermore, we directly compare our method with state-of-the-art relevant ones that address the problem of DG in the FL setting as discussed in Section 2, including ELCFS [18], CCST [3], COPA [42], and FedDG-GA [46]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 510, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 510, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 510, + 289, + 715 + ], + "type": "text", + "content": "Implementation Details. We present architectural details and hyper-parameter values used for experiments in the paper. Following [3, 46], for PACS and Office-Home datasets, we use the image size of " + }, + { + "bbox": [ + 46, + 510, + 289, + 715 + ], + "type": "inline_equation", + "content": "224 \\times 224" + }, + { + "bbox": [ + 46, + 510, + 289, + 715 + ], + "type": "text", + "content": " pixels and ResNet-50 [7] pre-trained on ImageNet as the backbone for the feature extractor and a linear layer as the classifier. BN layers in the first four and first two blocks in feature extractors are replaced with our PerXAN, respectively. For the Camelyon17 dataset, input images are resized to " + }, + { + "bbox": [ + 46, + 510, + 289, + 715 + ], + "type": "inline_equation", + "content": "96 \\times 96" + }, + { + "bbox": [ + 46, + 510, + 289, + 715 + ], + "type": "text", + "content": " pixels, and DenseNet-121 [8] is used instead of ResNet-50, BN layers in the first dense block and the following transition block are replaced by PerXAN. Notably, parameters related to BN layers are initialized with ImageNet pre-trained weights. Client models in all experiments are optimized by an SGD optimizer with a learning rate of 2e-3 for 100 communication rounds (i.e. " + }, + { + "bbox": [ + 46, + 510, + 289, + 715 + ], + "type": "inline_equation", + "content": "T = 100" + }, + { + "bbox": [ + 46, + 510, + 289, + 715 + ], + "type": "text", + "content": ") with one local update epoch at clients (i.e. " + }, + { + "bbox": [ + 46, + 510, + 289, + 715 + ], + "type": "inline_equation", + "content": "E = 1" + }, + { + "bbox": [ + 46, + 510, + 289, + 715 + ], + "type": "text", + "content": "). During the training, simple" + } + ] + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 305, + 282, + 548, + 456 + ], + "blocks": [ + { + "bbox": [ + 102, + 252, + 489, + 264 + ], + "lines": [ + { + "bbox": [ + 102, + 252, + 489, + 264 + ], + "spans": [ + { + "bbox": [ + 102, + 252, + 489, + 264 + ], + "type": "text", + "content": "Table 2. Accuracy comparison on the PACS and Office-Home datasets in the leave-one-domain-out setting." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 305, + 282, + 548, + 456 + ], + "lines": [ + { + "bbox": [ + 305, + 282, + 548, + 456 + ], + "spans": [ + { + "bbox": [ + 305, + 282, + 548, + 456 + ], + "type": "table", + "html": "
MethodCamelyon17
H1H2H3H4H5Avg
FedAvg (Baseline)97.091.889.994.281.090.8
FedAvg w/ MixStyle91.185.586.293.387.988.8
FedAvg w/ RSC90.690.688.394.593.391.5
ELCFS92.990.689.993.289.991.3
CCST91.590.287.394.691.691.0
COPA93.290.992.293.690.292.0
FedDG-GA97.290.791.092.390.592.3
gPerXAN (Ours)96.592.295.194.791.994.1
", + "image_path": "4abe91d7a862f8464364f1f58953e4df906ae73ff2976571738eec08275df989.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 318, + 465, + 533, + 477 + ], + "lines": [ + { + "bbox": [ + 318, + 465, + 533, + 477 + ], + "spans": [ + { + "bbox": [ + 318, + 465, + 533, + 477 + ], + "type": "text", + "content": "Table 3. Accuracy comparison on the Camelyon17 dataset." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 497, + 547, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 497, + 547, + 594 + ], + "spans": [ + { + "bbox": [ + 304, + 497, + 547, + 594 + ], + "type": "text", + "content": "data augmentation techniques are applied including random horizontal flipping and color jittering. The hyper-parameter of our regularization term " + }, + { + "bbox": [ + 304, + 497, + 547, + 594 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 304, + 497, + 547, + 594 + ], + "type": "text", + "content": " is searched in the range [0, 1] with a step of 0.25. The MixStyle [47] and RSC [10] can be directly integrated into the FedAvg without further modifications. All settings of other compared methods, i.e., EL-CFS [18], CCST [3], COPA [42], and FedDG-GA [46] are chosen based on corresponding papers." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 599, + 391, + 611 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 599, + 391, + 611 + ], + "spans": [ + { + "bbox": [ + 305, + 599, + 391, + 611 + ], + "type": "text", + "content": "4.3. Main Results" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 617, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 617, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 617, + 547, + 715 + ], + "type": "text", + "content": "Table 2 presents the quantitative results for different testing domains on the PACS and Office-Home datasets. Particularly, each result column shows the test accuracy of the global model on the domain of the column name. First of all, we can empirically verify that centralized DG methods such as MixStyle [47] and RSC [10] are inconsistent under the FedAvg framework and even harmful in certain cases compared to the baseline. Since these methods are" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6032" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 262 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 262 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 262 + ], + "type": "text", + "content": "designed for the centralized setting, they might need access to inter-domain knowledge to perform well, which is unavailable in the FL setting. In comparing state-of-the-art FedDG methods designed for the FL scenario, methods that share partial client data information, i.e., ELCFS [18] and CCST [3], show impressive results on the PACS dataset. However, they do not significantly affect the Office-Home dataset. A similar pattern is found with the architectural method COPA [42] and FedDG-GA [46]. Meanwhile, on the PACS and Office-Home datasets, our proposed method, gPerXAN achieves average accuracies across unseen clients of " + }, + { + "bbox": [ + 46, + 72, + 289, + 262 + ], + "type": "inline_equation", + "content": "87.94\\%" + }, + { + "bbox": [ + 46, + 72, + 289, + 262 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 72, + 289, + 262 + ], + "type": "inline_equation", + "content": "71.01\\%" + }, + { + "bbox": [ + 46, + 72, + 289, + 262 + ], + "type": "text", + "content": ", which are " + }, + { + "bbox": [ + 46, + 72, + 289, + 262 + ], + "type": "inline_equation", + "content": "1.02\\%" + }, + { + "bbox": [ + 46, + 72, + 289, + 262 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 72, + 289, + 262 + ], + "type": "inline_equation", + "content": "1.15\\%" + }, + { + "bbox": [ + 46, + 72, + 289, + 262 + ], + "type": "text", + "content": " better than the second-best ones, respectively. Although most of the considered methods can perform better than the baseline FedAvg in most cases, our method demonstrates a significant boost over others on both two standard benchmarks." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 264, + 289, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 264, + 289, + 479 + ], + "spans": [ + { + "bbox": [ + 46, + 264, + 289, + 479 + ], + "type": "text", + "content": "In addition to standard PACS and Office-Home datasets, we further evaluate our proposed method on a medical imaging dataset Camelyon17 as presented in Table 3. On this real-world benchmark, we can observe that ELCFS [18] and CCST [3] show considerably inferior performance compared to architectural ones, i.e., COPA [42] and gPerXAN. This might be due to sophisticated and sensitive features in medical images that are more challenging to extract and interpolate than conventional datasets [38], leading to inefficiency in these information-sharing-based methods. Notably, although it yields impressive performance, COPA [42] involves several other advanced techniques, such as RandAugment [4], making the performance gain hard to justify. Meanwhile, our method achieves an average accuracy across unseen clients of " + }, + { + "bbox": [ + 46, + 264, + 289, + 479 + ], + "type": "inline_equation", + "content": "94.1\\%" + }, + { + "bbox": [ + 46, + 264, + 289, + 479 + ], + "type": "text", + "content": ", outperforming FedDG-GA [46] by approximately " + }, + { + "bbox": [ + 46, + 264, + 289, + 479 + ], + "type": "inline_equation", + "content": "2\\%" + }, + { + "bbox": [ + 46, + 264, + 289, + 479 + ], + "type": "text", + "content": ". In general, the above experimental results demonstrate the effectiveness of the proposed method across various applications." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 489, + 149, + 502 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 489, + 149, + 502 + ], + "spans": [ + { + "bbox": [ + 47, + 489, + 149, + 502 + ], + "type": "text", + "content": "4.4. Ablation Studies" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 508, + 287, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 508, + 287, + 593 + ], + "spans": [ + { + "bbox": [ + 46, + 508, + 287, + 593 + ], + "type": "text", + "content": "We conduct ablation studies on the PACS dataset to investigate the impact of each building component. Specifically, we first compare PerXAN with other centralized variants such as conventional BN, I-BN [25], and DSON [30] under the same implementation details, and then our guiding regularizer is applied to ELCFS [18] and CCST [3] to verify its compatibility with the rest of the method." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 609, + 237, + 623 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 609, + 237, + 623 + ], + "spans": [ + { + "bbox": [ + 47, + 609, + 237, + 623 + ], + "type": "text", + "content": "4.4.1 Impact of the Normalization Scheme" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 630, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 630, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 630, + 287, + 715 + ], + "type": "text", + "content": "Table 4 provides a quantitative comparison among considered normalization schemes where gFedAvg represents the FedAvg framework implemented with the guiding regularizer. This means that we sequentially utilize the conventional BN, I-BN [25], and DSON [30] to replace the PerXAN scheme in our proposed method. From this table, we can observe that I-BN [25] can yield better perfor" + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 305, + 70, + 548, + 178 + ], + "blocks": [ + { + "bbox": [ + 305, + 70, + 548, + 178 + ], + "lines": [ + { + "bbox": [ + 305, + 70, + 548, + 178 + ], + "spans": [ + { + "bbox": [ + 305, + 70, + 548, + 178 + ], + "type": "table", + "html": "
MethodPACS
PACSAvg
gFedAvg w/ BN95.8584.3978.0476.4283.68
gFedAvg w/ I-BN94.6782.0880.4680.0784.32
gFedAvg w/ DSON93.7781.6479.9180.6383.99
gPerXAN (Ours)97.2786.5284.6883.2887.94
", + "image_path": "c248505276b24f47d6544d62d2057bf1e1137e6ea902898906b32c03ad4bb8a4.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 306, + 208, + 548, + 329 + ], + "blocks": [ + { + "bbox": [ + 324, + 186, + 527, + 198 + ], + "lines": [ + { + "bbox": [ + 324, + 186, + 527, + 198 + ], + "spans": [ + { + "bbox": [ + 324, + 186, + 527, + 198 + ], + "type": "text", + "content": "Table 4. Evaluation of different normalization schemes." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 306, + 208, + 548, + 329 + ], + "lines": [ + { + "bbox": [ + 306, + 208, + 548, + 329 + ], + "spans": [ + { + "bbox": [ + 306, + 208, + 548, + 329 + ], + "type": "image", + "image_path": "94f9c628159021288fa16c579adab956b87c84502f575bb30bf9319039b2937d.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 312, + 336, + 539, + 348 + ], + "lines": [ + { + "bbox": [ + 312, + 336, + 539, + 348 + ], + "spans": [ + { + "bbox": [ + 312, + 336, + 539, + 348 + ], + "type": "text", + "content": "Figure 2. Contribution of the regularizer on different methods." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 368, + 547, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 368, + 547, + 418 + ], + "spans": [ + { + "bbox": [ + 305, + 368, + 547, + 418 + ], + "type": "text", + "content": "mance than DSON [30] and the conventional BN, which are slightly comparable to each other. Meanwhile, PerXAN shows optimal performance with a significant margin compared to others and largely contributes to the whole method." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 430, + 450, + 443 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 430, + 450, + 443 + ], + "spans": [ + { + "bbox": [ + 306, + 430, + 450, + 443 + ], + "type": "text", + "content": "4.4.2 Impact of the Regularizer" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 449, + 547, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 449, + 547, + 629 + ], + "spans": [ + { + "bbox": [ + 304, + 449, + 547, + 629 + ], + "type": "text", + "content": "To better understand the proposed regularization term's contribution, we sequentially employ it in FedAvg, ELCFS [18], and CCST [3] to observe performance changes. Figure 2 displays averaged accuracies across unseen clients of these methods without and with the involvement of our regularizer. It is straightforward to see that while regularizer improves the performance of FedAvg and PerXAN considerably, it does not enhance ELCFS [18] and CCST [3]. In information-sharing-based methods, i.e., ELCFS [18] and CCST [3], each client is exposed to other clients' data information, which means that clients can access global knowledge relatively, then making the effect of matching global knowledge to clients' knowledge redundant and harmful. Also, this ablation study empirically verifies the connection between the regularizer and normalization scheme." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 639, + 444, + 653 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 639, + 444, + 653 + ], + "spans": [ + { + "bbox": [ + 306, + 639, + 444, + 653 + ], + "type": "text", + "content": "5. Analysis and Discussion" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 659, + 391, + 671 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 659, + 391, + 671 + ], + "spans": [ + { + "bbox": [ + 306, + 659, + 391, + 671 + ], + "type": "text", + "content": "5.1. Visualization" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 677, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 677, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 305, + 677, + 547, + 715 + ], + "type": "text", + "content": "We now provide an in-depth ability analysis of the proposed method via representation visualization using t-SNE [37], a common dimensionality reduction technique. As shown in" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6033" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 62, + 72, + 251, + 242 + ], + "blocks": [ + { + "bbox": [ + 62, + 72, + 251, + 242 + ], + "lines": [ + { + "bbox": [ + 62, + 72, + 251, + 242 + ], + "spans": [ + { + "bbox": [ + 62, + 72, + 251, + 242 + ], + "type": "image", + "image_path": "ab506ec2940b2bee47a0333b246301000e144a6f2b9ba3cf857220fed241c919.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 124, + 251, + 468, + 264 + ], + "lines": [ + { + "bbox": [ + 124, + 251, + 468, + 264 + ], + "spans": [ + { + "bbox": [ + 124, + 251, + 468, + 264 + ], + "type": "text", + "content": "Figure 3. Visualization of representation extracted from the global model on the PACS dataset." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 347, + 72, + 535, + 242 + ], + "blocks": [ + { + "bbox": [ + 347, + 72, + 535, + 242 + ], + "lines": [ + { + "bbox": [ + 347, + 72, + 535, + 242 + ], + "spans": [ + { + "bbox": [ + 347, + 72, + 535, + 242 + ], + "type": "image", + "image_path": "58675cc41bdd3feae2797f6bfb79fc42ce236980c17ffcee28fb3c14d4fee8af.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 284, + 289, + 429 + ], + "lines": [ + { + "bbox": [ + 46, + 284, + 289, + 429 + ], + "spans": [ + { + "bbox": [ + 46, + 284, + 289, + 429 + ], + "type": "text", + "content": "Figure 3, we compare the representations extracted from the global model obtained from our gPerXAN and the baseline FedAvg on the PACS dataset with the testing domain, a.k.a. target domain Photo (P). From this figure, we can observe that features derived from our method are semantically separated according to 7 categories in the PACS dataset for both source domains and the target domain. Moreover, features of each category on all domains tend to be close and grouped, demonstrating the ability of our method to distill more discriminative and domain-invariant representation, leading to a significant improvement of more than " + }, + { + "bbox": [ + 46, + 284, + 289, + 429 + ], + "type": "inline_equation", + "content": "2\\%" + }, + { + "bbox": [ + 46, + 284, + 289, + 429 + ], + "type": "text", + "content": " in classification accuracy on the testing domain Photo (P)." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 441, + 176, + 455 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 441, + 176, + 455 + ], + "spans": [ + { + "bbox": [ + 47, + 441, + 176, + 455 + ], + "type": "text", + "content": "5.2. Privacy and Efficiency" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 462, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 462, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 462, + 289, + 715 + ], + "type": "text", + "content": "Data privacy is a major concern in the field of FL, which is mitigated by only sharing client models' parameters instead of raw data. However, to resolve the FedDG problem, recently introduced methods sacrifice this principle by revealing partial information about client data. Specifically, leaking more information not only opens higher chances for attackers to perform an inversion attack [24], which aims to reconstruct the original data of clients, but also amplifies the risk of membership inference attack [32], which determines if a sample was in the model's training data. Furthermore, efficiency in communication and computation is critical in FL, especially in scenarios where resource-constrained clients are involved. Despite that, the common sharing mechanism in available methods introduces unacceptable extra costs to the FL training process. For COPA [42], using an ensemble of domain-specific classifiers in the model architecture results in an " + }, + { + "bbox": [ + 46, + 462, + 289, + 715 + ], + "type": "inline_equation", + "content": "O(N^2)" + }, + { + "bbox": [ + 46, + 462, + 289, + 715 + ], + "type": "text", + "content": " increase in communication and computation complexities compared to " + }, + { + "bbox": [ + 46, + 462, + 289, + 715 + ], + "type": "inline_equation", + "content": "O(N)" + }, + { + "bbox": [ + 46, + 462, + 289, + 715 + ], + "type": "text", + "content": " as conventional, according to the number of participating clients " + }, + { + "bbox": [ + 46, + 462, + 289, + 715 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 462, + 289, + 715 + ], + "type": "text", + "content": ". Meanwhile, FedDG-GA [46] consumes a double of memory at clients. Compared with existing ones, in ad" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 284, + 545, + 321 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 284, + 545, + 321 + ], + "spans": [ + { + "bbox": [ + 304, + 284, + 545, + 321 + ], + "type": "text", + "content": "dition to bypassing the limitations described above, our proposed method is more practical in terms of implementation yet provides competitive results in extensive evaluations." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 329, + 379, + 342 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 329, + 379, + 342 + ], + "spans": [ + { + "bbox": [ + 305, + 329, + 379, + 342 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 349, + 545, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 349, + 545, + 601 + ], + "spans": [ + { + "bbox": [ + 304, + 349, + 545, + 601 + ], + "type": "text", + "content": "In this paper, we introduce a novel architectural method, namely gPerXAN, to address the problem of FedDG. By explicitly assembling Instance Normalization layers into Batch Normalization layers in a personalized scheme and employing a simple yet effective guiding regularizer, our method allows the model to filter domain-specific features and actively distill domain-invariant representation for classification tasks. We conduct extensive experiments and in-depth analysis to quantitatively and qualitatively verify the effectiveness of our proposed method in solving this particular problem. Although our evaluation is more on cross-silo FL, our method can be easily extended to cross-device scenarios while Algorithm 1 remains unchanged. Unlike existing methods, due to the independence from imaging techniques, gPerXAN can be straightforwardly extended to diverse applications. A potential limitation of our method is its certain suitability with normalization-based models. In the future, investigating other forms of regularization terms is a promising research direction due to the available room for improvements. Moreover, the vulnerability of available methods under various attacks remains underexplored." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 609, + 404, + 624 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 609, + 404, + 624 + ], + "spans": [ + { + "bbox": [ + 306, + 609, + 404, + 624 + ], + "type": "text", + "content": "Acknowledgement" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 629, + 545, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 629, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 629, + 545, + 715 + ], + "type": "text", + "content": "This work is supported by the Deutsche Forschungsgemeinschaft, German Research Foundation under grant number 453130567 (COSMO), by the Horizon Europe Research and Innovation Actions under grant number 101092908 (SmartEdge), and by the Federal Ministry for Education and Research, Germany under grant number 01IS18037A (BIFOLD)." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6034" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 714 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 134 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 134 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 134 + ], + "type": "text", + "content": "[1] Yogesh Balaji, Swami Sankaranarayanan, and Rama Chellappa. MetaReg: Towards Domain Generalization Using Meta-Regularization. Advances in Neural Information Processing Systems, 31, 2018. 1, 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 135, + 288, + 212 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 135, + 288, + 212 + ], + "spans": [ + { + "bbox": [ + 53, + 135, + 288, + 212 + ], + "type": "text", + "content": "[2] Peter Bandi, Oscar Geessink, Quirine Manson, Marcory Van Dijk, Maschenka Balkenhol, Meyke Hermsen, Babak Ehteshami Bejnordi, Byungjae Lee, Kyunghyun Paeng, Aoxiao Zhong, et al. From Detection of Individual Metastases to Classification of Lymph Node Status at the Patient Level: The Camelyon17 Challenge. IEEE Transactions on Medical Imaging, 38(2):550-560, 2018. 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 213, + 288, + 267 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 213, + 288, + 267 + ], + "spans": [ + { + "bbox": [ + 53, + 213, + 288, + 267 + ], + "type": "text", + "content": "[3] Junming Chen, Meirui Jiang, Qi Dou, and Qifeng Chen. Federated Domain Generalization for Image Recognition via Cross-Client Style Transfer. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 361–370, 2023. 2, 3, 5, 6, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 268, + 288, + 323 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 268, + 288, + 323 + ], + "spans": [ + { + "bbox": [ + 53, + 268, + 288, + 323 + ], + "type": "text", + "content": "[4] Ekin D Cubuk, Barret Zoph, Jonathon Shlens, and Quoc V Le. RandAugment: Practical Automated Data Augmentation With a Reduced Search Space. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, pages 702-703, 2020. 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 324, + 287, + 356 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 324, + 287, + 356 + ], + "spans": [ + { + "bbox": [ + 53, + 324, + 287, + 356 + ], + "type": "text", + "content": "[5] Shi Dong, Ping Wang, and Khushnood Abbas. A Survey on Deep Learning and Its Applications. Computer Science Review, 40:100379, 2021. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 357, + 287, + 401 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 357, + 287, + 401 + ], + "spans": [ + { + "bbox": [ + 53, + 357, + 287, + 401 + ], + "type": "text", + "content": "[6] Qi Dou, Daniel Coelho de Castro, Konstantinos Kamnitsas, and Ben Glocker. Domain Generalization via Model-Agnostic Learning of Semantic Features. Advances in Neural Information Processing Systems, 32, 2019. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 402, + 287, + 446 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 402, + 287, + 446 + ], + "spans": [ + { + "bbox": [ + 53, + 402, + 287, + 446 + ], + "type": "text", + "content": "[7] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep Residual Learning for Image Recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 770-778, 2016. 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 447, + 287, + 500 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 447, + 287, + 500 + ], + "spans": [ + { + "bbox": [ + 53, + 447, + 287, + 500 + ], + "type": "text", + "content": "[8] Gao Huang, Zhuang Liu, Laurens Van Der Maaten, and Kilian Q Weinberger. Densely Connected Convolutional Networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4700-4708, 2017. 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 502, + 287, + 545 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 502, + 287, + 545 + ], + "spans": [ + { + "bbox": [ + 53, + 502, + 287, + 545 + ], + "type": "text", + "content": "[9] Xun Huang and Serge Belongie. Arbitrary Style Transfer in Real-Time With Adaptive Instance Normalization. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1501-1510, 2017. 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 546, + 287, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 546, + 287, + 590 + ], + "spans": [ + { + "bbox": [ + 48, + 546, + 287, + 590 + ], + "type": "text", + "content": "[10] Zeyi Huang, Haohan Wang, Eric P Xing, and Dong Huang. Self-Challenging Improves Cross-Domain Generalization. In Proceedings of the European Conference on Computer Vision (ECCV), pages 124-140. Springer, 2020. 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 591, + 287, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 591, + 287, + 635 + ], + "spans": [ + { + "bbox": [ + 48, + 591, + 287, + 635 + ], + "type": "text", + "content": "[11] Da Li, Yongxin Yang, Yi-Zhe Song, and Timothy M Hospedales. Deeper, Broader and Artier Domain Generalization. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5542-5550, 2017. 5" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 636, + 287, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 636, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 636, + 287, + 689 + ], + "type": "text", + "content": "[12] Da Li, Jianshu Zhang, Yongxin Yang, Cong Liu, Yi-Zhe Song, and Timothy M Hospedales. Episodic Training for Domain Generalization. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1446-1455, 2019. 1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 690, + 287, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 690, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 48, + 690, + 287, + 714 + ], + "type": "text", + "content": "[13] Haoliang Li, Sinno Jialin Pan, Shiqi Wang, and Alex C Kot. Domain Generalization With Adversarial Feature Learning." + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "type": "text", + "content": "In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5400-5409, 2018. 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 98, + 545, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 98, + 545, + 140 + ], + "spans": [ + { + "bbox": [ + 308, + 98, + 545, + 140 + ], + "type": "text", + "content": "[14] Tian Li, Anit Kumar Sahu, Manzil Zaheer, Maziar Sanjabi, Ameet Talwalkar, and Virginia Smith. Federated Optimization in Heterogeneous Networks. Proceedings of Machine Learning and Systems, 2:429-450, 2020. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 143, + 545, + 186 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 143, + 545, + 186 + ], + "spans": [ + { + "bbox": [ + 308, + 143, + 545, + 186 + ], + "type": "text", + "content": "[15] Xiang Li, Kaixuan Huang, Wenhao Yang, Shusen Wang, and Zhihua Zhang. On the Convergence of Fedavg on Non-IID Data. In International Conference on Learning Representations, 2019. 4" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 189, + 545, + 232 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 189, + 545, + 232 + ], + "spans": [ + { + "bbox": [ + 308, + 189, + 545, + 232 + ], + "type": "text", + "content": "[16] Yanghao Li, Naiyan Wang, Jianping Shi, Jiaying Liu, and Xiaodi Hou. Revisiting Batch Normalization for Practical Domain Adaptation. arXiv preprint arXiv:1603.04779, 2016.4" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 236, + 545, + 279 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 236, + 545, + 279 + ], + "spans": [ + { + "bbox": [ + 308, + 236, + 545, + 279 + ], + "type": "text", + "content": "[17] Quande Liu, Qi Dou, Lequan Yu, and Pheng Ann Heng. MS-Net: Multi-Site Network for Improving Prostate Segmentation With Heterogeneous MRI Data. IEEE Transactions on Medical Imaging, 39(9):2713-2724, 2020. 4" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 281, + 545, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 281, + 545, + 346 + ], + "spans": [ + { + "bbox": [ + 308, + 281, + 545, + 346 + ], + "type": "text", + "content": "[18] Quande Liu, Cheng Chen, Jing Qin, Qi Dou, and Pheng-Ann Heng. FedDG: Federated Domain Generalization on Medical Image Segmentation via Episodic Learning in Continuous Frequency Space. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1013–1023, 2021. 2, 3, 5, 6, 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 350, + 545, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 350, + 545, + 392 + ], + "spans": [ + { + "bbox": [ + 308, + 350, + 545, + 392 + ], + "type": "text", + "content": "[19] Ping Luo, Jiamin Ren, Zhanglin Peng, Ruimao Zhang, and Jingyu Li. Differentiable Learning-To-Normalize via Switchable Normalization. In International Conference on Learning Representations, 2018. 4" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 396, + 545, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 396, + 545, + 437 + ], + "spans": [ + { + "bbox": [ + 308, + 396, + 545, + 437 + ], + "type": "text", + "content": "[20] Divyat Mahajan, Shruti Tople, and Amit Sharma. Domain Generalization Using Causal Matching. In International Conference on Machine Learning, pages 7313-7324. PMLR, 2021. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 441, + 545, + 495 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 441, + 545, + 495 + ], + "spans": [ + { + "bbox": [ + 308, + 441, + 545, + 495 + ], + "type": "text", + "content": "[21] Brendan McMahan, Eider Moore, Daniel Ramage, Seth Hampson, and Blaise Aguera y Arcas. Communication-Efficient Learning of Deep Networks From Decentralized Data. In Artificial Intelligence and Statistics, pages 1273-1282. PMLR, 2017. 1, 4, 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 498, + 545, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 498, + 545, + 542 + ], + "spans": [ + { + "bbox": [ + 308, + 498, + 545, + 542 + ], + "type": "text", + "content": "[22] Krikamol Muandet, David Balduzzi, and Bernhard Scholkopf. Domain Generalization via Invariant Feature Representation. In International Conference on Machine Learning, pages 10-18. PMLR, 2013. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 544, + 545, + 598 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 544, + 545, + 598 + ], + "spans": [ + { + "bbox": [ + 308, + 544, + 545, + 598 + ], + "type": "text", + "content": "[23] Hyeonseob Nam, HyunJae Lee, Jongchan Park, Wonjun Yoon, and Donggeun Yoo. Reducing Domain Gap by Reducing Style Bias. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8690-8699, 2021. 1" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 601, + 545, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 601, + 545, + 666 + ], + "spans": [ + { + "bbox": [ + 308, + 601, + 545, + 666 + ], + "type": "text", + "content": "[24] Ngoc-Bao Nguyen, Keshigeyan Chandrasegaran, Milad Abdollahzadeh, and Ngai-Man Cheung. Re-thinking Model Inversion Attacks Against Deep Neural Networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, pages 16384-16393, 2023. 8" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 669, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 669, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 669, + 545, + 713 + ], + "type": "text", + "content": "[25] Xingang Pan, Ping Luo, Jianping Shi, and Xiaou Tang. Two at Once: Enhancing Learning and Generalization Capacities via IBN-Net. In Proceedings of the European Conference on Computer Vision (ECCV), pages 464–479, 2018. 1, 2, 3, 5, 7" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6035" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "type": "text", + "content": "[26] Xingchao Peng, Zijun Huang, Yizhe Zhu, and Kate Saenko. Federated Adversarial Domain Adaptation. In International Conference on Learning Representations, 2020. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 106, + 288, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 106, + 288, + 160 + ], + "spans": [ + { + "bbox": [ + 48, + 106, + 288, + 160 + ], + "type": "text", + "content": "[27] Krishna Pillutla, Kshitiz Malik, Abdel-Rahman Mohamed, Mike Rabbat, Maziar Sanjabi, and Lin Xiao. Federated Learning With Partial Model Personalization. In International Conference on Machine Learning, pages 17716-17758. PMLR, 2022. 2, 3, 4" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 161, + 288, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 161, + 288, + 194 + ], + "spans": [ + { + "bbox": [ + 48, + 161, + 288, + 194 + ], + "type": "text", + "content": "[28] Joaquin Quinonero-Candela, Masashi Sugiyama, Anton Schwaighofer, and Neil D Lawrence. Dataset Shift in Machine Learning. Mit Press, 2008. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 194, + 288, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 194, + 288, + 239 + ], + "spans": [ + { + "bbox": [ + 48, + 194, + 288, + 239 + ], + "type": "text", + "content": "[29] Alexandre Rame, Coretin Dancette, and Matthieu Cord. Fishr: Invariant Gradient Variances for Out-Of-Distribution Generalization. In International Conference on Machine Learning, pages 18347–18377. PMLR, 2022. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 239, + 288, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 239, + 288, + 293 + ], + "spans": [ + { + "bbox": [ + 48, + 239, + 288, + 293 + ], + "type": "text", + "content": "[30] Seonguk Seo, Yumin Suh, Dongwan Kim, Geeho Kim, Jongwoo Han, and Bohyung Han. Learning to Optimize Domain Specific Normalization for Domain Generalization. In Proceedings of the European Conference on Computer Vision (ECCV), pages 68-83, 2020. 2, 3, 4, 5, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 293, + 288, + 337 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 293, + 288, + 337 + ], + "spans": [ + { + "bbox": [ + 48, + 293, + 288, + 337 + ], + "type": "text", + "content": "[31] Neta Shoham, Tomer Avidor, Aviv Keren, Nadav Israel, Daniel Benditkis, Liron Mor-Yosef, and Itai Zeitak. Overcoming Forgetting in Federated Learning on Non-IID Data. arXiv preprint arXiv:1910.07796, 2019. 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 337, + 288, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 337, + 288, + 381 + ], + "spans": [ + { + "bbox": [ + 48, + 337, + 288, + 381 + ], + "type": "text", + "content": "[32] Reza Shokri, Marco Stronati, Congzheng Song, and Vitaly Shmatikov. membership inference attacks against machine learning models. In 2017 IEEE Symposium on Security and Privacy (SP), pages 3-18. IEEE, 2017. 8" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 381, + 288, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 381, + 288, + 426 + ], + "spans": [ + { + "bbox": [ + 48, + 381, + 288, + 426 + ], + "type": "text", + "content": "[33] Baochen Sun and Kate Saenko. Deep Coral: Correlation Alignment for Deep Domain Adaptation. In Proceedings of the European Conference on Computer Vision (ECCV), pages 443-450, 2016. 1, 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 426, + 288, + 469 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 426, + 288, + 469 + ], + "spans": [ + { + "bbox": [ + 48, + 426, + 288, + 469 + ], + "type": "text", + "content": "[34] Baochen Sun, Jiashi Feng, and Kate Saenko. Correlation Alignment for Unsupervised Domain Adaptation. Domain Adaptation in Computer Vision Applications, pages 153-171, 2017. 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 470, + 288, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 470, + 288, + 514 + ], + "spans": [ + { + "bbox": [ + 48, + 470, + 288, + 514 + ], + "type": "text", + "content": "[35] Benyuan Sun, Hongxing Huo, Yi Yang, and Bo Bai. *PartialFed: Cross-Domain Personalized Federated Learning via Partial Initialization*. Advances in Neural Information Processing Systems, 34:23309-23320, 2021. 2, 3, 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 514, + 288, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 514, + 288, + 557 + ], + "spans": [ + { + "bbox": [ + 48, + 514, + 288, + 557 + ], + "type": "text", + "content": "[36] Yuwei Sun, Ng Chong, and Hideya Ochiai. Feature Distribution Matching for Federated Domain Generalization. In Asian Conference on Machine Learning, pages 942-957. PMLR, 2023. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 558, + 288, + 591 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 558, + 288, + 591 + ], + "spans": [ + { + "bbox": [ + 48, + 558, + 288, + 591 + ], + "type": "text", + "content": "[37] Laurens Van der Maaten and Geoffrey Hinton. Visualizing Data Using t-SNE. Journal of Machine Learning Research, 9(11), 2008. 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 591, + 288, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 591, + 288, + 635 + ], + "spans": [ + { + "bbox": [ + 48, + 591, + 288, + 635 + ], + "type": "text", + "content": "[38] Gael Varoquaux and Veronika Cheplygina. Machine Learning for Medical Imaging: Methodological Failures and Recommendations for the Future. Npj Digital Medicine, 5(1):48, 2022. 7" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 635, + 288, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 635, + 288, + 690 + ], + "spans": [ + { + "bbox": [ + 48, + 635, + 288, + 690 + ], + "type": "text", + "content": "[39] Hemanth Venkateswara, Jose Eusebio, Shayok Chakraborty, and Sethuraman Panchanathan. Deep Hashing Network for Unsupervised Domain Adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5018-5027, 2017. 5" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 690, + 288, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 690, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 690, + 288, + 713 + ], + "type": "text", + "content": "[40] Paul Voigt and Axel Von dem Bussche. The EU General Data Protection Regulation (GDPR). A Practical Guide, 1st Ed.," + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 72, + 545, + 497 + ], + "type": "list", + "angle": 0, + "index": 25, + "blocks": [ + { + "bbox": [ + 327, + 72, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 72, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 327, + 72, + 545, + 95 + ], + "type": "text", + "content": "Cham: Springer International Publishing, 10(3152676):10-5555, 2017. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 95, + 545, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 95, + 545, + 150 + ], + "spans": [ + { + "bbox": [ + 307, + 95, + 545, + 150 + ], + "type": "text", + "content": "[41] Jindong Wang, Cuiling Lan, Chang Liu, Yidong Ouyang, Tao Qin, Wang Lu, Yiqiang Chen, Wenjun Zeng, and Philip Yu. Generalizing to Unseen Domains: A Survey on Domain Generalization. IEEE Transactions on Knowledge and Data Engineering, 2022. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 151, + 545, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 151, + 545, + 206 + ], + "spans": [ + { + "bbox": [ + 307, + 151, + 545, + 206 + ], + "type": "text", + "content": "[42] Guile Wu and Shaogang Gong. Collaborative Optimization and Aggregation for Decentralized Domain Generalization and Adaptation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6484-6493, 2021. 3, 5, 6, 7, 8" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 206, + 545, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 206, + 545, + 251 + ], + "spans": [ + { + "bbox": [ + 307, + 206, + 545, + 251 + ], + "type": "text", + "content": "[43] Zhenlin Xu, Deyi Liu, Junlin Yang, Colin Raffel, and Marc Niethammer. Robust and Generalizable Visual Representation Learning via Random Convolutions. In International Conference on Learning Representations, 2020. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 252, + 545, + 306 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 252, + 545, + 306 + ], + "spans": [ + { + "bbox": [ + 307, + 252, + 545, + 306 + ], + "type": "text", + "content": "[44] Fu-En Yang, Yuan-Chia Cheng, Zu-Yun Shiau, and Yu-Chiang Frank Wang. Adversarial Teacher-Student Representation Learning for Domain Generalization. Advances in Neural Information Processing Systems, 34:19448-19460, 2021. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 308, + 545, + 361 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 308, + 545, + 361 + ], + "spans": [ + { + "bbox": [ + 307, + 308, + 545, + 361 + ], + "type": "text", + "content": "[45] Chun-Han Yao, Boqing Gong, Hang Qi, Yin Cui, Yukun Zhu, and Ming-Hsuan Yang. Federated Multi-Target Domain Adaptation. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1424–1433, 2022. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 363, + 545, + 419 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 363, + 545, + 419 + ], + "spans": [ + { + "bbox": [ + 307, + 363, + 545, + 419 + ], + "type": "text", + "content": "[46] Ruipeng Zhang, Qinwei Xu, Jiangchao Yao, Ya Zhang, Qi Tian, and Yanfeng Wang. Federated Domain Generalization with Generalization Adjustment. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3954-3963, 2023. 3, 5, 6, 7, 8" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 419, + 545, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 419, + 545, + 453 + ], + "spans": [ + { + "bbox": [ + 307, + 419, + 545, + 453 + ], + "type": "text", + "content": "[47] Kaiyang Zhou, Yongxin Yang, Yu Qiao, and Tao Xiang. Domain Generalization With Mixstyle. In International Conference on Learning Representations, 2020. 1, 3, 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 453, + 545, + 497 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 453, + 545, + 497 + ], + "spans": [ + { + "bbox": [ + 307, + 453, + 545, + 497 + ], + "type": "text", + "content": "[48] Weiming Zhuang, AI Sony, and Lingjuan Lyu. FedWon: Triumphing Multi-domain Federated Learning Without Normalization. In International Conference on Learning Representations, 2024. 4" + } + ] + } + ], + "index": 24 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 316, + 757 + ], + "type": "text", + "content": "6036" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Ego-Exo4D_ Understanding Skilled Human Activity from First- and Third-Person Perspectives/742959dc-4498-434e-af9a-b1cc7f6e5bd2_content_list.json b/2024/Ego-Exo4D_ Understanding Skilled Human Activity from First- and Third-Person Perspectives/742959dc-4498-434e-af9a-b1cc7f6e5bd2_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..ddc47500b51e4b7b3571db3a7fac311cafb8b619 --- /dev/null +++ b/2024/Ego-Exo4D_ Understanding Skilled Human Activity from First- and Third-Person Perspectives/742959dc-4498-434e-af9a-b1cc7f6e5bd2_content_list.json @@ -0,0 +1,2281 @@ +[ + { + "type": "text", + "text": "Ego-Exo4D: Understanding Skilled Human Activity from First- and Third-Person Perspectives", + "text_level": 1, + "bbox": [ + 222, + 130, + 746, + 174 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Kristen Grauman $^{1,2}$ , Andrew Westbury $^{1}$ , Lorenzo Torresani $^{1}$ , Kris Kitani $^{1,3}$ , Jitendra Malik $^{1,4}$ , Triantafyllos Afouras $^{*1}$ , Kumar Ashutosh $^{*1,2}$ , Vijay Baiyya $^{5}$ , Siddhant Bansal $^{6,7}$ , Bikram Boote $^{8}$ , Eugene Byrne $^{*1,9}$ , Zach Chavis $^{*10}$ , Joya Chen $^{*11}$ , Feng Cheng $^{1}$ , Fu-Jen Chu $^{1}$ , Sean Crane $^{9}$ , Avijit Dasgupta $^{7}$ , Jing Dong $^{5}$ , Maria Escobar $^{*12}$ , Cristian Forigua $^{*12}$ , Abrham Gebreselasie $^{9}$ , Sanjay Haresh $^{13}$ , Jing Huang $^{1}$ , Md Mohaiminul Islam $^{14}$ , Suyog Jain $^{1}$ , Rawal Khirodkar $^{9}$ , Devansh Kukreja $^{1}$ , Kevin J Liang $^{1}$ , Jia-Wei Liu $^{11}$ , Sagnik Majumder $^{1,2}$ , Yongsen Mao $^{13}$ , Miguel Martin $^{1}$ , Effrosyni Mavroudi $^{1}$ , Tushar Nagarajan $^{1}$ , Francesco Ragusa $^{15}$ , Santhosh Kumar Ramakrishnan $^{2}$ , Luigi Seminara $^{15}$ , Arjun Somayazulu $^{2}$ , Yale Song $^{1}$ , Shan Su $^{16}$ , Zihui Xue $^{1,2}$ , Edward Zhang $^{16}$ , Jinxu Zhang $^{16}$ , Angela Castillo $^{12}$ , Changan Chen $^{2}$ , Xinzhu Fu $^{11}$ , Ryosuke Furuta $^{17}$ , Cristina González $^{12}$ , Prince Gupta $^{5}$ , Jiabo Hu $^{18}$ , Yifei Huang $^{17}$ , Yiming Huang $^{16}$ , Weslie Khoo $^{19}$ , Anush Kumar $^{10}$ , Robert Kuo $^{18}$ , Sach Lakhavani $^{5}$ , Miao Liu $^{18}$ , Mi Luo $^{2}$ , Zhengyi Luo $^{3}$ , Brighid Meredith $^{18}$ , Austin Miller $^{18}$ , Oluwatuminiu Oguntola $^{14}$ , Xiaqing Pan $^{5}$ , Penny Peng $^{18}$ , Shraman Pramanick $^{20}$ , Merey Ramazanova $^{21}$ , Fiona Ryan $^{22}$ , Wei Shan $^{14}$ , Kiran Somasundaram $^{5}$ , Chenan Song $^{11}$ , Audrey Southerland $^{22}$ , Masatoshi Tateno $^{17}$ , Huiyu Wang $^{1}$ , Yuchen Wang $^{19}$ , Takuma Yagi $^{17}$ , Mingfei Yan $^{5}$ , Xitong Yang $^{1}$ , Zecheng Yu $^{17}$ , Shengxin Cindy Zha $^{18}$ , Chen Zhao $^{21}$ , Ziwei Zhao $^{19}$ , Zhifan Zhu $^{6}$ , Jeff Zhuo $^{14}$ , Pablo Arbeláez† $^{12}$ , Gedas Bertasius† $^{14}$ , Dima Damen† $^{6}$ , Jakob Engel† $^{5}$ , Giovanni Maria Farinella† $^{15}$ , Antonino Furnari† $^{15}$ , Bernard Ghanem† $^{21}$ , Judy Hoffman† $^{22}$ , C. V. Jawahar† $^{7}$ , Richard Newcombe† $^{5}$ , Hyun Soo Park† $^{10}$ , James M. Rehg† $^{8}$ , Yoichi Sato† $^{17}$ , Manolis Savva† $^{13}$ , Jianbo Shi† $^{16}$ , Mike Zheng Shou† $^{11}$ and Michael Wray† $^{6}$", + "bbox": [ + 84, + 214, + 883, + 551 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ FAIR, Meta, $^{2}$ University of Texas at Austin, $^{3}$ Carnegie Mellon University, $^{4}$ University of California, Berkeley, $^{5}$ Project Aria, Meta, $^{6}$ University of Bristol, $^{7}$ International Institute of Information Technology, Hyderabad, $^{8}$ University of Illinois, Urbana Champaign, $^{9}$ Carnegie Mellon University, $^{10}$ University of Minnesota, $^{11}$ National University of Singapore, $^{12}$ Universidad de los Andes, $^{13}$ Simon Fraser University, $^{14}$ University of North Carolina, Chapel Hill, $^{15}$ University of Catania, $^{16}$ University of Pennsylvania, $^{17}$ University of Tokyo, $^{18}$ Meta, $^{19}$ Indiana University, $^{20}$ Johns Hopkins University, $^{21}$ King Abdullah University of Science and Technology, $^{22}$ Georgia Tech", + "bbox": [ + 76, + 560, + 893, + 633 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 655, + 313, + 671 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We present Ego-Exo4D, a diverse, large-scale multimodal multiview video dataset and benchmark challenge. Ego-Exo4D centers around simultaneously-captured ego-centric and exocentric video of skilled human activities (e.g., sports, music, dance, bike repair). 740 participants from 13 cities worldwide performed these activities in 123 different natural scene contexts, yielding long-form captures from 1 to 42 minutes each and 1,286 hours of video combined. The multimodal nature of the dataset is unprecedented: the video is accompanied by multichannel audio, eye gaze, 3D point clouds, camera poses, IMU, and multiple paired language descriptions—including a novel \"expert commentary\" done by coaches and teachers and tailored to the skilled-activity domain. To push the frontier of first-person video understanding of skilled hu", + "bbox": [ + 73, + 674, + 472, + 902 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "man activity, we also present a suite of benchmark tasks and their annotations, including fine-grained activity understanding, proficiency estimation, cross-view translation, and 3D hand/body pose. All resources are open sourced to fuel new research in the community.", + "bbox": [ + 496, + 678, + 893, + 755 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 500, + 768, + 632, + 784 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "A dancer leaps across a stage; Lionel Messi delivers a precise pass; your grandmother prepares her famous dumplings. We observe and seek human skills in a myriad of settings, from the practical (fixing a bike) to the aspirational (dancing beautifully). What would it mean for AI to understand human skills? And what would it take to get there?", + "bbox": [ + 496, + 794, + 893, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 44 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "19383", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/fac5d8236d3d229f991c89ab4fb2140794f3f9860ab60fe5efe2e1344f38e59f.jpg", + "image_caption": [ + "Figure 1. Ego-Exo4D offers egocentric video alongside multiple time-synchronized exocentric video streams for an array of skilled human activities—1,286 hours of ego and exo video in total. The data is both multiview and multimodal, and it is extensively annotated with language, 3D body and hand pose, keysteps, procedural dependencies, and proficiency ratings in support of our proposed benchmark tasks." + ], + "image_footnote": [], + "bbox": [ + 125, + 87, + 844, + 422 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Advances in AI understanding of human skill could facilitate many applications. In augmented reality (AR), a person wearing smart glasses could quickly pick up new skills with a virtual AI coach that provides real-time guidance. In robot learning, a robot watching people in its environment could acquire new dexterous manipulation skills with less physical experience. In social networks, new communities could form based on how people share their expertise and complementary skills in video.", + "bbox": [ + 75, + 500, + 470, + 636 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We contend that both the egocentric and exocentric viewpoints are critical for capturing human skill. Firstly, the two viewpoints are synergistic. The first-person (ego) perspective captures the details of close-by hand-object interactions and the camera wearer's attention, whereas the third-person (exo) perspective captures the full body pose and surrounding environment context. See Figure 1. Not coincidentally, instructional or \"how-to\" videos often alternate between a third-person view of the demonstrator and a close-up view of their near-field demonstration. For example, a chef may describe their approach and the equipment from an exo view, then cut to clips showing their hands manipulating the ingredients and tools from an ego-like view.", + "bbox": [ + 75, + 640, + 468, + 835 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Secondly, not only are the ego and exo viewpoints synergistic, but there is a need to translate fluently from one to the other when acquiring skill. For example, imagine watching an expert repair a bike tire, juggle a soccer ball,", + "bbox": [ + 75, + 840, + 470, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "or fold an origami swan—then mapping their steps to your own body. Cognitive science tells us that even from a very young age we can observe others' behavior (exo) and map it onto our own (ego) [42, 108], and this actor-observer translation remains the foundation of visual learning.", + "bbox": [ + 496, + 500, + 890, + 575 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Realizing this potential, however, is not possible using today's datasets and learning paradigms. Existing datasets comprised of both ego and exo views (i.e., ego-exo) are few [76, 77, 127, 139, 145], small in scale, lack synchronization across cameras, and/or are too staged or curated to be resilient to the diversity of the real world. Thus the current literature for activity understanding primarily attends to either the ego [28, 47] or exo [48, 67, 105, 149] view, leaving the ability to move fluidly between the first- and third-person perspectives out of reach. Instructional video datasets [103, 159, 204, 207] offer a compelling window into skilled human activity, but (like the above) are limited to single-viewpoint video, whether purely exocentric or mixed with \"ego-like\" views at certain time points.", + "bbox": [ + 496, + 579, + 892, + 790 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We introduce Ego-Exo4D, a foundational dataset to support research on ego-exo video learning and multimodal perception. The result of a two-year effort by a consortium of 15 research institutions, Ego-Exo4D is a first-of-its-kind large-scale multimodal multiview dataset and benchmark suite. It constitutes the largest public dataset of timesynchronized first- and third- person video, captured by 740", + "bbox": [ + 496, + 795, + 893, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "19384", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "diverse camera wearers in 123 distinct scenes and 13 cities worldwide. For every sequence, Ego-Exo4D provides both the camera wearer's egocentric video, as well as multiple (4-5) exocentric videos from tripod places around the camera wearer. All views are time-synchronized and precisely localized in a metric, gravity-aligned frame of reference. The total collection has 1,286 hours of video and 5,035 instances, each spanning 1 to $42\\mathrm{min}$ of continuous capture.", + "bbox": [ + 75, + 90, + 468, + 212 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Ego-Exo4D focuses on skilled single-person activities. The 740 participants perform skilled physical and/or procedural activities—dance, soccer, basketball, bouldering, music, cooking, bike repair, health care—in an unscripted manner and in natural settings (e.g., gym, soccer field, kitchens, bike shops, etc.), exhibiting a variety of skill levels from novice to expert. All video is recorded with rigorous privacy and ethics policies and formal consent of participants.", + "bbox": [ + 75, + 217, + 468, + 339 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Ego-Exo4D is not only multiview, it is also multimodal. Captured with the unique open-source Aria glasses [38], all ego video is accompanied by 7-channel audio, IMU, eye gaze, both RGB and two grayscale SLAM cameras, and 3D environment point clouds. Additionally, Ego-Exo4D provides multiple new video-language resources, all time indexed: first-person narrations by the camera wearers describing their own actions; third-person play-by-play descriptions of every camera wearer action; and third-person spoken expert commentary critiquing their performance. The latter is particularly novel: performed by domain-specific experienced coaches and teachers, it focuses on how an activity is executed rather than merely what is being done, surfacing subtleties in skilled execution not perceivable by the untrained eye. To our knowledge, there is no prior video resource with such extensive and high quality multimodal data.", + "bbox": [ + 75, + 344, + 470, + 601 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Alongside this data, we introduce benchmarks for foundational tasks for ego-exo video. We propose four families of tasks: 1) ego-exo relation, for relating the actions of a teacher (exo) to a learner (ego) by estimating semantic correspondences and translating viewpoints; 2) ego(-exo) recognition, for recognizing fine-grained keysteps and task structure; 3) ego(-exo) proficiency estimation, for inferring how well a person is executing a skill; and 4) ego pose, for recovering skilled 3D body and hand movements from ego-video. We provide annotations for each task—the result of more than 200,000 hours of annotator effort. To kick-start work in these new challenges, we also develop baseline models and report their results (Appendices). We are hosting the first public benchmark challenges in 2024.", + "bbox": [ + 75, + 607, + 468, + 819 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In summary, Ego-Exo4D is the community's first diverse, large-scale multimodal multiview video resource. We have open sourced all the data, annotations, camera rig protocol, and benchmarks. With this release, we aim to fuel new research in ego-exo, multimodal activity, and beyond.", + "bbox": [ + 75, + 825, + 468, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2. Related work", + "text_level": 1, + "bbox": [ + 500, + 89, + 635, + 104 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Next we review prior work in datasets, human skill, and cross-view analysis. Section 5 will discuss related work for each benchmark task. Table 2 in Appendix 9 summarizes Ego-Exo4D's properties vs. existing datasets.", + "bbox": [ + 498, + 114, + 890, + 176 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Egocentric datasets There has been a surge of interest in egocentric video understanding, facilitated by recent egovideo datasets showing unscripted daily-life activity as in Ego4D [47], EPIC-Kitchens [27, 28, 163], UT Ego [78], ADL [119], and KrishnaCam [147], or procedural activities as in EGTea [81], AssistQ [172], Meccano [126], CMUMMAC [77], and EgoProcel [10]. Unlike any of the above, Ego-Exo4D focuses on multimodal ego and exo capture, and it is focused on the domain of skilled activities.", + "bbox": [ + 498, + 185, + 892, + 321 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Multiview and ego-exo datasets Most existing multiview datasets focus on static scenes [20, 128, 151, 175, 176] and objects [133, 173], with limited (exo only) multiview human activity [26, 169]. CMU-MMAC [77] and CharadesEgo [145] are early efforts to capture both ego and exo video. CMU-MMAC [77] features 43 participants in mocap suits who cook 5 recipes in a lab kitchen. In CharadesEgo [145], 71 Mechanical Turkers record 34 hours of scripted scenarios (e.g., \"type on laptop, then pick up a pillow\") from the ego and exo perspectives sequentially, yielding unsynchronized videos with non-exact activity matches. More recent ego-exo efforts focus on specific activities in one or two environments. Assembly101 [139] and H2O [76] provide time-synced ego and exo video at a lab tabletop where people assemble toy cars or manipulate handheld objects, with 53 and 4 participants, and 513 and 5 hours of footage, respectively. Homage [127] provides 30 hours of ego-exo video from 27 participants in 2 homes doing household activities like laundry.", + "bbox": [ + 498, + 330, + 892, + 618 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Compared to any of the prior efforts, Ego-Exo4D offers an order of magnitude more participants, diverse locations, and hours of footage (740 participants, 123 unique scenes, 13 cities, 1,286 hours). Importantly, our focus on skilled tasks takes the participants out of the lab or home and into settings like soccer fields, dance studios, rock climbing walls, and bike repair shops. Such activities also yield a wide variety of full body poses and movements within the scene, beyond using objects at a tabletop. This variety means Ego-Exo4D augments existing 3D human body pose datasets [49, 66, 68, 80, 193]. Finally, compared to any prior ego-exo resource, Ego-Exo4D's suite of modalities and benchmark tasks are novel and will expand the research directions the community can take for egocentric and/or exocentric video understanding.", + "bbox": [ + 498, + 619, + 890, + 844 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Human skill and video learning Analyzing skill and action quality has received limited attention [12, 34, 35, 113, 120, 194]. Research in instructional or \"how-to\" videos is", + "bbox": [ + 500, + 854, + 892, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "19385", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/8e9f0c174b11764e0f15c304b5c22610240f5992f86298476a8de14579f1d267.jpg", + "image_caption": [ + "Cooking" + ], + "image_footnote": [], + "bbox": [ + 89, + 97, + 130, + 130 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/eebc1e0bf3ece0ea41c10a5452a90876a509030e552ddf750acc4e667d57939d.jpg", + "image_caption": [ + "x678", + "x173", + "8", + "Mx65", + "$\\mathbb{O}\\times 564\\mathrm{h}$" + ], + "image_footnote": [], + "bbox": [ + 202, + 89, + 352, + 145 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/a19d07b019f38fa7c8c76ad4a0301f9e336f516ac051e63806116e8155f497a2.jpg", + "image_caption": [ + "Music", + "x59", + "#", + "#" + ], + "image_footnote": [], + "bbox": [ + 359, + 97, + 419, + 132 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/984ff27e32171872ea2ada18f5aedee16cc6fa00b95998cfa31845245dfb8c80.jpg", + "image_caption": [ + "x276", + "L x 18Oh" + ], + "image_footnote": [], + "bbox": [ + 482, + 89, + 630, + 145 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/5a380045c4d5459df5d074c6037f9d8118652d1a20cf8e4e0f490735a0b03331.jpg", + "image_caption": [ + "Soccer", + "x282", + "x78", + "#", + "X14", + "$\\bullet \\times 66h$" + ], + "image_footnote": [], + "bbox": [ + 643, + 99, + 681, + 128 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/9a5e64d3c71b41b6225c4cef72f3bda9fcdcf5c474614aae5d0828f21ad7bd0c.jpg", + "image_caption": [ + "$\\therefore m = \\frac{3}{11}$ ;", + "Soccer in Pittsburgh" + ], + "image_footnote": [], + "bbox": [ + 816, + 90, + 890, + 145 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/c7aa36c3ac9b6150a2bd3f89c2eff2a7cd3f143b3eb72f3be85b17932dac5950.jpg", + "image_caption": [ + "Health", + "9x24" + ], + "image_footnote": [], + "bbox": [ + 89, + 154, + 132, + 190 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/693afb2f8315a87b6db52b87cf9b6d450f5b8030e72417b8027d01539b8e2897.jpg", + "image_caption": [ + "Bike Repair", + "4x363", + "x32", + "#", + "图" + ], + "image_footnote": [], + "bbox": [ + 84, + 220, + 140, + 253 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/bb9b9912947665d957dd000e5903a3a3fca5ab9d4d9ad8bfc1c1c303d813f372.jpg", + "image_caption": [ + "Making Chai in Hyderabad", + "x397", + "x122", + "中", + "X141" + ], + "image_footnote": [], + "bbox": [ + 202, + 152, + 349, + 209 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/0d8c3e0cc3b5f0e78255b64b060f0ba6fceaaf039156df74327dc10012f83a30.jpg", + "image_caption": [ + "X 82n", + "Removing a Wheel in Menlo Park", + "Figure 2. Ego-Exo4D captures skilled activity from 43 tasks and 689 keysteps in 8 domains, in a wide variety of 123 scenes in 13 cities in Japan, Colombia, Canada, India, Singapore, and 7 US states. Each domain is captured at multiple sites—from 2 to 64 unique locations. In total the dataset offers 1,286 hours of ego+exo video comprised of 5,035 takes from 740 camera wearers. An average take is 2.6 minutes." + ], + "image_footnote": [], + "bbox": [ + 199, + 212, + 349, + 271 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/2d79df213b3cc867df1653fb63954d69806914bb972f99f01dd47e5d2f31cf51.jpg", + "image_caption": [ + "Basketball", + "x910", + "x113", + "9x5", + "#" + ], + "image_footnote": [], + "bbox": [ + 374, + 154, + 415, + 191 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/8f5812255672053bbc201eee5ab4ee80e671699df439d8f665a728e94586e047.jpg", + "image_caption": [ + "Rock Climbing" + ], + "image_footnote": [], + "bbox": [ + 374, + 209, + 415, + 257 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/b3262b0b991a281b3481f1053e74925e8ce35455330367f75db85b66f831caf6.jpg", + "image_caption": [ + "5x78n" + ], + "image_footnote": [], + "bbox": [ + 429, + 209, + 478, + 271 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/cd72a671a831712395fef38e7bb491b1d1d9d8048b0ef6a9eecb25ba32e7671d.jpg", + "image_caption": [ + "Piano Playing in Bloomington" + ], + "image_footnote": [], + "bbox": [ + 488, + 152, + 629, + 209 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/28d54ba7c8516fa7d9299162bcbc4485a70940f8fbeaa3f9ac89ebd39d80f416.jpg", + "image_caption": [ + "Bouldering in Minneapolis", + "Total" + ], + "image_footnote": [], + "bbox": [ + 480, + 210, + 629, + 271 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/35e7474e239096906675b0aa12b24e7ad316f45ac7b0d3da3354e281c1610354.jpg", + "image_caption": [ + "Dance", + "9x7" + ], + "image_footnote": [], + "bbox": [ + 651, + 154, + 683, + 191 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/ef05d05169d1820987b79dcef743525e471d31daa226c8ce72abaeb09ed032e4.jpg", + "image_caption": [ + "4x728", + "x93", + "#", + "x106n", + "$\\therefore m = \\frac{3}{11}$ ;", + "Salsa in Bogota", + "4x5035 (#takes)", + "x740 (#participants)", + "9. x 123 (# sites)", + "x 1286h (#ego+exo hours)" + ], + "image_footnote": [], + "bbox": [ + 743, + 162, + 888, + 209 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "facilitated by (largely exo) datasets like HowTo100M [103] and others [11, 159, 204, 207]. Challenges include grounding keysteps [10, 36, 37, 89, 103, 104, 178, 207], procedural planning [15, 17, 22, 71, 143, 167, 196, 201], learning task structure [4, 9, 37, 107, 202, 205], and leveraging noisy narrations [89, 103, 104]. A portion of Ego-Exo4D is procedural activities, but unlike the above, it offers simultaneous ego-exo capture. The scale and diversity of our data—including its three forms of language descriptions—widen the avenues for skilled activity understanding research.", + "bbox": [ + 75, + 354, + 468, + 506 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Ego-exo cross-view modeling There is limited prior work on ego-exo cross-view modeling, arguably due to a lack of high-quality synchronized real-world data. Prior work explores matching people between videos [5, 6, 40, 170, 179] and learning view-invariant [7, 141, 144, 182, 184, 185] or ego features [82]. Beyond the specific case of ego-exo, cross-view methods are explored for translation [130, 131, 134, 157], novel view synthesis [19, 90, 135, 137, 164, 168, 171], and aerial to ground matching [86, 132]. Ego-Exo4D provides a testbed of unprecedented size and variety for cross-view modeling. In addition, our ego-exo relation tasks (cf. Section 5) surface new challenges in novel-view synthesis with widely varying viewpoints.", + "bbox": [ + 75, + 520, + 468, + 717 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3. Ego-Exo4D dataset", + "text_level": 1, + "bbox": [ + 76, + 732, + 261, + 750 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Next we introduce the dataset and its scope. Notably, the video capture was a distributed but coordinated effort performed by 12 research labs. We present the common framework, and reserve site-specific details for Appendix 10.", + "bbox": [ + 75, + 758, + 468, + 819 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1. Ego-exo camera rig", + "text_level": 1, + "bbox": [ + 76, + 830, + 264, + 848 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our goal is to capture simultaneous ego and exo video, together with multiple egocentric sensing modalities. One of our contributions is to create and share a low-cost (less than", + "bbox": [ + 75, + 854, + 468, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "\\(3,000), lightweight ego-exo rig with a user-friendly calibration and time sync procedure.", + "bbox": [ + 498, + 354, + 890, + 385 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our camera configuration features Aria glasses [38] for ego capture, leveraging their rich array of sensors, including an 8 MP RGB camera, two SLAM cameras, IMU, 7 microphones, and eye tracking (see Appendix 7). The ego camera is calibrated and time-synchronized with four to five (stationary) GoPros placed on tripods as the exo capture devices, allowing 3D reconstruction of the environment point clouds and the participant's body pose. The number and placement of the exocentric cameras is determined per scenario in order to allow maximal coverage of useful viewpoints without obstructing the participants' activity.", + "bbox": [ + 496, + 386, + 890, + 551 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our time sync and calibration design relies on a QR-code procedure to auto-sync the cameras and auto-separate the individual \"takes\", meaning instances of an activity. We can do continuous recordings of up to $\\sim 60$ minutes, based on the Aria battery life. See Appendix 8 for more details.", + "bbox": [ + 496, + 553, + 890, + 628 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Domains and environments", + "text_level": 1, + "bbox": [ + 500, + 641, + 745, + 655 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Ego-Exo4D focuses on skilled human activity. This is in contrast to existing ego-only efforts like Ego4D [47], which has a broad span of daily-life activities. We intentionally select the domains based on a few criteria: Will it illustrate skill and a variety of expertise? Is there visual variety to be expected across different instances? Will the ego and exo views offer complementary information? Will it present new challenges unaddressed by current datasets?", + "bbox": [ + 496, + 664, + 890, + 785 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Intersecting these criteria, we arrived at two broad categories1 of skilled activity: physical and procedural, together comprising eight total domains. The physical domains are soccer, basketball, dance, bouldering, and music. They emphasize body pose and movements as well as", + "bbox": [ + 496, + 787, + 890, + 862 + ], + "page_idx": 3 + }, + { + "type": "page_footnote", + "text": "1Note that in general physical and procedural are not mutually exclusive labels. An activity can both require physical skill and procedural steps.", + "bbox": [ + 500, + 875, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "19386", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "interaction with objects (e.g., a ball, musical instrument). The procedural domains are cooking, bike repair, and health care. They require performing a sequence of steps to reach a goal state (e.g., a completed recipe, a repaired bike) and generally entail intricate hand-object manipulations with a variety of objects (e.g., bike repair tools; cooking utensils, appliances, and ingredients).", + "bbox": [ + 75, + 90, + 467, + 196 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In total, we have 43 activities derived from the eight domains (see Appendix 9). For example, cooking is comprised of 14 recipes; soccer is comprised of 3 drills. The length of a take ranges from 8 sec to $42\\mathrm{min}$ , with procedural activities like cooking having the longest sustained captures.", + "bbox": [ + 75, + 196, + 467, + 272 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To achieve visual diversity in the data, multiple labs across our team (typically 3-5) captured each Ego-Exo4D domain. The data is collected in authentic settings—such as real-world bike shops, soccer pitches, or bouldering gyms—as opposed to lab environments. For example, we have videos of chefs in New York City, Vancouver, Philadelphia, Bogota, and others; soccer players in Tokyo, Chapel Hill, Hyderabad, Singapore, and Pittsburgh. See Figure 2.", + "bbox": [ + 75, + 273, + 467, + 393 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Participants: expertise and diversity", + "text_level": 1, + "bbox": [ + 76, + 404, + 390, + 420 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We recruited 740 total participants from the local communities of 12 labs. All scenarios feature real-world experts, where the camera-wearer participant has specific credentials, training, or expertise in the skill being demonstrated. For example, among the Ego-Exo4D camera wearers are professional and college athletes; jazz, salsa, and Chinese folk dancers and instructors; competitive boulders; professional chefs who work in industrial-scale kitchens; bike technicians who service dozens of bikes per day. Many of them have (individually) over 10 years of experience.", + "bbox": [ + 75, + 428, + 467, + 578 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Experts are prioritized given they are likely to conduct activities without mistakes or distractions, providing a strong ground truth for how to approach a given task. However, we also include capture from people with varying skill levels, as well—essential for our proposed skill proficiency estimation task (Section 5). Notably, Ego-Exo4D represents human intelligence in a new way by capturing domain-specific expertise—both in the video as well as the accompanying expert commentary (see Section 4)—portraying the evolution of a skill from beginners to experts.", + "bbox": [ + 75, + 579, + 467, + 729 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "According to the participant surveys (Appendix 11), the camera wearers range in age from 18 to 74 years old, with $37\\%$ self-identifying as female $60\\%$ male and $3\\%$ as nonbinary or preferring not to say. In total, the participants self report more than 24 different ethnicities.", + "bbox": [ + 75, + 729, + 467, + 805 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Privacy and ethics", + "text_level": 1, + "bbox": [ + 76, + 816, + 253, + 832 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Ego-Exo4D was collected following rigorous privacy and ethics standards. This included undergoing formal independent review processes at each institution to establish the standards for collection, management, and informed con", + "bbox": [ + 75, + 839, + 467, + 898 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/23c2068815e1aebf6b65f96acf44b75df4aa036595e21ce72aa799e712f3f696.jpg", + "image_caption": [ + "Expert commentary", + "Atomic action descriptions", + "Narrate and act", + "Figure 3. Ego-Exo4D offers 3 paired language corpora. Word cloud is from expert commentary which critiques the performance." + ], + "image_footnote": [], + "bbox": [ + 503, + 89, + 602, + 198 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "\"The dancer's hand is rotated inwardly a bit. Her palm should be facing to the ground, ...\"", + "bbox": [ + 606, + 102, + 745, + 125 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "\"C steps forward while shimmering her shoulders.\"", + "bbox": [ + 607, + 145, + 728, + 160 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "\"Here we go, left forward, right back for one phrase 1234...\"", + "bbox": [ + 607, + 179, + 743, + 195 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/090ac63719d41de68d3ac319ba1aaf42222509b8103f6d3c584ce5f72c4b69d7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 750, + 89, + 890, + 198 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "sent. Similarly, all Ego-Exo4D data collection adhered to the Project Aria Research Community Guidelines for responsible research. Since the scenarios allow for closed environments (e.g., no passerbys) nearly all video is available without de-identification. For information about each individual partners' protocols and restrictions, please see Appendix 10. Ego-Exo4D data is gated behind a license system, which defines permitted uses, restrictions, and consequences for non-compliance.", + "bbox": [ + 496, + 256, + 890, + 392 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Natural language descriptions", + "text_level": 1, + "bbox": [ + 498, + 406, + 774, + 422 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Ego-Exo4D also offers three kinds of paired natural language datasets, each time-indexed alongside the video. See Figure 3. These language annotations are not steered towards any single benchmark, but rather are a general resource that will support browsing and mining the dataset—as well as challenges in video-language learning like grounding actions and objects, self-supervised representation learning, video-conditioned language models, and skill assessment. See Appendix 12.", + "bbox": [ + 496, + 431, + 890, + 568 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The first language dataset is spoken expert commentary. The goal is to reveal nuances of the skill that are not always visible to non-experts. We recruited 52 experts (distinct from the participants) to critique the recorded videos, call out strengths and weaknesses, explain how the specific behavior of the participant (e.g., hand/body pose, use of objects) affects the performance, and provide spatial markings to support their commentary. The experts are not only well-credentialed in their areas of expertise, but also have coaching or teaching experience, which facilitates clear communication. They watch the video and pause every time they have a comment, typically 7 times per minute of video. Each piece of commentary is unbounded in length, and averages 4 sentences. We provide both the transcribed speech and the raw audio (interesting for its inflection and nonword utterances), as well as the experts' spatial drawings and numeric ratings of each participant's skill. Videos have expert commentary by 2-5 distinct experts, offering a variety of perspectives for the same content. In total, we have 117,812 pieces of time-stamped, video-aligned commentary. These commentaries are quite novel: they focus on how the activity is executed rather than what it entails,", + "bbox": [ + 496, + 569, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "19387", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/89e99b3d4d421fb985f90b50f51421249d01d29f4ac88415876f8ae480461290.jpg", + "image_caption": [ + "Figure 4. The ego-exo relation family consists of the tasks of correspondence (left) and translation (right)." + ], + "image_footnote": [], + "bbox": [ + 80, + 89, + 521, + 250 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/c32b41e20a7b16c60de14b52977b7500e1565171dc15dbf4cd0aa43a8c26dc8e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 529, + 90, + 888, + 250 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "capturing subtle differences in skilled execution. We believe this can unlock new fundamental problems (e.g., proficiency estimation below) and disruptive future applications (e.g., AI coaching).", + "bbox": [ + 75, + 295, + 468, + 354 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The second language dataset consists of narrate-and-act descriptions provided by the participants themselves. They are in the style of a tutorial or how-to video, where the participant explains what they are doing and why. Unlike the third-party expert commentary above, these are first-person reflections on the activity given by the people doing them. These narrations are available for about $10\\%$ of all takes in the dataset, since we wanted participants to execute the tasks without pausing for the bulk of the recordings.", + "bbox": [ + 75, + 357, + 468, + 492 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The third language dataset consists of atomic action descriptions. Whereas the commentary and narrate-and-act language reveals spoken opinions and reasons for the actions (the \"why and how\"), this stream of text is specifically about the \"what\". Inspired by Ego4D's narrations [47], these are short statements written by third-party (nondomain expert) annotators, timestamped for every atomic action performed by the participant for all videos in the dataset, for a total of 432K sentences. This data is valuable for mining for taxonomies of objects and actions in the data, indexing the videos with keywords for exploring the dataset, and for future research in video-language learning, as has been quite successful for the Ego4D narrations [8, 85, 123].", + "bbox": [ + 75, + 494, + 468, + 690 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Ego-Exo4D benchmark tasks", + "text_level": 1, + "bbox": [ + 76, + 707, + 344, + 724 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Our second major contribution is to define the core research challenges in the domain of egocentric perception of skilled activity, particularly when ego-exo data is available for training (if not testing). To that end, we devise a suite of foundational benchmark tasks organized into four task families: relation (Sec. 5.1), recognition (Sec. 5.2), proficiency (Sec. 5.3), and ego-pose (Sec. 5.4). For each task, we provide high quality annotations and baselines that provide a starting point from which the research community can build. We will run the first formal Ego-Exo4D challenges in 2024. Due to space limits, we briefly overview", + "bbox": [ + 75, + 734, + 468, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "each task; see the referenced Appendices for all details including baseline models and results. There are two publicly released versions of Ego-Exo4D annotations: v1 is used to train/test baselines in this paper; the larger v2 will be used for future challenge leaderboards (see Table 7 in Appendix).", + "bbox": [ + 496, + 295, + 890, + 371 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1. Ego-exo relation", + "text_level": 1, + "bbox": [ + 498, + 382, + 661, + 398 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Our ego-exo relation tasks deal with relating the video content across the extreme ego-exo viewpoint changes. They take the form of object-level matching (correspondence) and synthesis of one view from the other (translation).", + "bbox": [ + 496, + 407, + 890, + 467 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1.1 Ego-exo correspondence", + "text_level": 1, + "bbox": [ + 498, + 479, + 723, + 494 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Motivation. Establishing object-level correspondences between ego and exo viewpoints would allow AI assistants to provide visual instructions by matching third-person observations of objects from instructional videos to those in the user's first-person view. Compared to the general correspondence problem, our setting requires tackling a number of challenges: extreme viewpoint differences, high degrees of object occlusion, and many small objects (e.g., cooking utensils and bike repair tools).", + "bbox": [ + 496, + 506, + 890, + 642 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Task definition. Given a pair of synchronized ego-exo videos and a sequence of query masks of an object of interest in one of the videos, the task is to predict the corresponding mask for the same object in each synchronized frame of the other view, if it is visible. See Figure 4, left. The task can be posed with query objects in either the ego or exo video, with both directions presenting interesting challenges (e.g., high degree of occlusion in ego views, and small object size in exo views). See Appendix 13.A.1.", + "bbox": [ + 496, + 643, + 890, + 779 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Related work. Related tasks are image-level sparse correspondence given query points (instead of object masks) [65] and image-level object co-segmentation [166] for jointly segmenting semantically similar objects. Our task goes beyond static object correspondence, since the interplay between human pose and object state changes during manipulation necessitate using temporal context and tracking as the query object can be highly occluded or blurry [158].", + "bbox": [ + 496, + 780, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "19388", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/ff098b4a1c0e580f2546738c52b291d66c8636b675c4d68f04c6ea7151e35c9d.jpg", + "image_caption": [ + "Figure 5. Ego-exo keystep recognition. This family of tasks consists of fine-grained recognition (left, Section 5.2.1), procedure understanding (center, Section 5.2.3), and energy-efficient multimodal recognition (right, Section 5.2.2)." + ], + "image_footnote": [], + "bbox": [ + 78, + 88, + 344, + 265 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/fd1e6617f4d3344436a9a691e16b97009d3ab5c5043378dbf4e1a9fdf69f3ca2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 351, + 88, + 617, + 266 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/bf51d78274add40ba749ff7f90622abf5e662a32f03e54a7f68398be5de6235b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 622, + 88, + 890, + 265 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.1.2 Ego-exo translation", + "text_level": 1, + "bbox": [ + 76, + 332, + 267, + 347 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Motivation. Our translation task entails synthesizing a target ego clip from a given exo clip. We believe this problem will drive novel research for combining recognition and object synthesis. For example, in Figure 4 (right), the approach must make effective use of the hand's object-specific shape and appearance priors in order to synthesize the ego view of the fingertips—which are not visible in the exo clip. Furthermore, this task will stimulate advances in visual odometry, as the method must be able to infer the ego camera pose from the third-person clip. Ego-exo translation also holds strong application potential, as it may unlock the ability to generate first-person renderings of videos that were originally captured from a third-person perspective, e.g., benefitting robot perception or AR coaching.", + "bbox": [ + 75, + 356, + 468, + 568 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Task definition. We decompose ego-exo translation into two separate tasks: ego track prediction and ego clip generation (Figure 4, right). Ego track prediction estimates the segmentation mask of an object in the unobserved ego frames given the object masks in the observed exo clip. Ego clip generation must generate the image values (i.e., RGB) within the given ground-truth ego mask by making use of the exo clip and the object masks in those frames. This decomposition effectively splits the problem into two tasks: 1) predicting the location and shape of the object in the ego clip, and 2) synthesizing its appearance given the ground-truth position. For each, we consider a variant where the pose of the ego camera with respect to the exo camera is available to use at inference time. This simplifies the problem but reduces the applicability of the method, since this information is typically not available for arbitrary third-person videos. See Appendix 13.A.2.", + "bbox": [ + 75, + 568, + 470, + 824 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Related work. Ego-exo translation relates to cross-view image synthesis [96, 130, 157]. Within this genre, the problem of exo-to-ego generation was recently introduced for both images [93] and video [94, 97], and approached using GANs or diffusion conditioned on the input view. Our work", + "bbox": [ + 75, + 825, + 468, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "not only formalizes this task with ample data, but its formulation also draws attention to the need for a semantic basis to new view synthesis across extreme view changes.", + "bbox": [ + 498, + 332, + 890, + 378 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.2. Ego-exo keystep recognition", + "text_level": 1, + "bbox": [ + 500, + 388, + 751, + 404 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "This family of tasks centers around recognizing the keysteps of a procedural activity and modeling their dependencies.", + "bbox": [ + 498, + 411, + 890, + 441 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.2.1 Fine-grained keystep recognition", + "text_level": 1, + "bbox": [ + 500, + 452, + 782, + 468 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Motivation. Recognizing the step a camera wearer is performing is non-trivial: keysteps in the same activity may look similar (folding vs. smoothing the bedsheet) and may involve hand-object interactions with heavy occlusions and head motion. Models with access to multiple views during training can leverage their complementarity to account for the deficiencies of each one, by learning viewpoint invariant representations or distilling multi-view signals into a single model (e.g., human hands from ego; body pose from exo).", + "bbox": [ + 496, + 477, + 892, + 612 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Task definition. We study ego-exo for video recognition. During training, models have access to paired ego-exo data—time-synchronized captures of the same activity from multiple known viewpoints. Each training instance has one ego view, $N$ exo views, and a corresponding keystep label (e.g., \"flip the omelette\"). At test time, given only a trimmed egocentric video clip, the model must identify the keystep performed from a taxonomy of 689 keysteps across 17 procedural activities. See Figure 5, left. Importantly, all extra supervision (time-alignment, camera poses etc.) is only available at training time; inference is standard keystep recognition, but with models that benefit from cross-viewpoint training. See Appendix 13.B.1.", + "bbox": [ + 496, + 613, + 892, + 809 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Related work. Keystep recognition has been studied in first-person [10, 126, 145, 148] or third-person [9, 100, 159, 205, 207] videos; however, limited work considers both views together. Prior work considers cross-view learning with unpaired videos [7, 82, 182] and view-invariant feature learning on paired videos [144]. In contrast, we ex", + "bbox": [ + 496, + 810, + 893, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "19389", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "plore keystep recognition in large-scale, procedural activities with fully synchronized training videos.", + "bbox": [ + 76, + 90, + 468, + 122 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.2.2 Energy-efficient multimodal keystep recognition", + "text_level": 1, + "bbox": [ + 76, + 131, + 465, + 148 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Motivation. Current activity detection models assume access to densely sampled clips from the full video and ample computational resources to process them. These assumptions are incompatible with real-world devices (e.g., mobile phones, AR glasses) where the camera is not always on and the compute budget is limited by battery life. This task focuses on building energy-efficient video models to pave the way for feasibility on real-world hardware.", + "bbox": [ + 75, + 155, + 468, + 275 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Task definition. We formulate the problem as an online action detection task, with a given energy budget. See Figure 5, right. Given a stream of audio, IMU, and RGB video data, a model must identify the keystep being performed at each frame, as well as decide which sensor(s) to use for subsequent time-steps. This task will inspire models that are strategic about which modality to deploy when. Energy consumption is the sum of sensor energy (operating the camera/audio/IMU sensors), model inference costs, and memory transfer costs, and must be within $20\\mathrm{mW}$ to reflect real-world device power constraints. See Appendix 13.B.2.", + "bbox": [ + 75, + 276, + 468, + 443 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Related work. Prior work on efficient models considers light-weight architectures [41, 56, 101, 155, 165, 195], efficient input processing [43, 44, 73, 102, 156], or inference optimizations [39, 58, 121, 174, 206]. In all cases, they optimize computation (FLOPs), parameter count, or prediction throughput (FPS), which in isolation are insufficient to characterize running on real-world devices. To address this, we propose the first benchmark for energy-efficient video recognition that is tied to real-world, on-device constraints, and measures total power consumed.", + "bbox": [ + 75, + 443, + 468, + 594 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.2.3 Procedure understanding", + "text_level": 1, + "bbox": [ + 76, + 604, + 308, + 619 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Motivation. Automatically understanding the structure of a procedure from video (inferring keystep ordering, preconditions, etc.) would allow assisting AR users in a task or informing robots that learn from human demonstrations.", + "bbox": [ + 75, + 628, + 468, + 688 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Task definition. In our procedure understanding task, given a video segment $s_t$ and its previous video segment history, models have to 1) determine previous keysteps (to be performed before $s_t$ ); infer if $s_t$ is 2) optional or 3) a procedural mistake; 4) predict missing keysteps (should have been performed before $s_t$ but were not); and 5) next keysteps (for which dependencies are satisfied). The task offers two versions of weak supervision: instance-level: segments and their keystep labels are available for train/test; and procedure-level: only unlabeled segments and procedure-specific keystep names are given for train/test. See Figure 5 (center) and Appendix 13.B.3.", + "bbox": [ + 75, + 688, + 468, + 869 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Related work. Prior work focusing on procedural understanding learns an explicit graph [62, 150, 177] as", + "bbox": [ + 76, + 869, + 468, + 901 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/8179128309f83265ff9d2b17386b2dd0f72a15b935efdf359b6762bfc3f7347a.jpg", + "image_caption": [ + "Figure 6. Demonstrator and demonstration proficiency estimation." + ], + "image_footnote": [], + "bbox": [ + 504, + 89, + 890, + 234 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "ground truth or uses a task graph for representation learning [9, 107, 202] and short-term step understanding [9, 36, 202]. Other work [32, 139] studies mistake detection in a supervised setting. We are the first to propose procedural understanding to evaluate the long-term structure of the task in a weakly-supervised setting.", + "bbox": [ + 498, + 279, + 890, + 369 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.3. Ego-exo proficiency estimation", + "text_level": 1, + "bbox": [ + 500, + 380, + 769, + 396 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Motivation. Going beyond recognizing what a person is doing, this task aims to infer the user's skill level. Such an ability could lead to novel coaching tools that let people learn new skills more effectively, or new ways to evaluate human performance in domains like sports or music.", + "bbox": [ + 498, + 402, + 890, + 479 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Task definition. We consider two variants: (1) demonstrator and (2) demonstration proficiency estimation. Both tasks consider one egocentric and (optionally) $M$ exocentric videos synchronized in time as their inputs. Demonstrator proficiency is formulated as a video classification task, where the model has to output one of four labels (novice, early, intermediate, or late expert). Demonstration proficiency is formulated as a temporal action localization task where given an untrimmed video, the model must output a list of tuples, each containing a timestamp, a proficiency category (i.e., good execution or needs improvement), and its probability. Note that parts of the video that do not reveal the participant's skill are left unlabeled. See Figure 6 and Appendix 13.C.", + "bbox": [ + 498, + 479, + 892, + 690 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Related work. Prior work uses egocentric [12, 35] or exocentric [60, 114, 115] views for proficiency estimation in sports [12, 115, 120], health [60, 92, 191, 208], and others [35, 186]. We propose the first multi-view egocentric and exocentric proficiency estimation benchmark. Unlike prior work, our benchmark spans diverse, day-to-day physical and procedural scenarios and includes temporally localized annotations of (in)correct executions.", + "bbox": [ + 498, + 691, + 890, + 811 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.4. Ego pose", + "text_level": 1, + "bbox": [ + 500, + 821, + 604, + 838 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This family of tasks is motivated by recovering the skilled body movements of participants, even in the extreme setting of monocular ego-video input in dynamic environments.", + "bbox": [ + 498, + 845, + 890, + 890 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "19390", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/1153a25866145462bb4bdcae10f45145db80a004ab374bbddd7703c6dcaca210.jpg", + "image_caption": [ + "Figure 7. Hand and body keypoints for ego-pose estimation" + ], + "image_footnote": [], + "bbox": [ + 93, + 90, + 459, + 220 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Motivation. Estimating the physical state of a person's body—the 3D positions of the arms, legs, hands—from the ego view is essential for wearable AI systems that can support human activity. Challenges include subtle and flexible movements, frequent occlusion, and body parts out of view. Task definition. For both the body and hand pose (\"ego pose\") estimation tasks, the input is an ego video. The output is a set of 3D joint positions of the camera wearer's body and hands for each time step, parameterized as 17 3D body joint positions and 21 3D joint positions per hand, following the MS COCO convention [87]. To our knowledge, EgoExo4D offers the largest manual ground truth (GT) egocentric body and hand pose annotations to date. And, in total, it offers $\\sim 14\\mathrm{M}$ frames of 3D GT and pseudo-GT combined. See Figure 7 and Appendix 13.D.", + "bbox": [ + 75, + 253, + 468, + 479 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Related work. Limited prior work explores 3D body pose from a wearable camera. Some methods assume no body visibility [63, 80, 98, 187, 188], while others assume partial observability by modifying cameras to capture the body [3, 57, 136, 161, 181]. Our dataset can be used for both paradigms. Existing hand pose datasets use constrained environments [106, 146] with simple hand motion [50, 76, 109], whereas we include diverse real-world scenarios, e.g., with expert musicians and bike mechanics.", + "bbox": [ + 75, + 482, + 467, + 617 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "6. Conclusions", + "text_level": 1, + "bbox": [ + 76, + 632, + 204, + 648 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Ego-Exo4D provides a dataset of unprecedented scale and realism for ego-exo video learning. It offers a unique window into skilled human activity from 8 compelling domains by hundreds of real-world experts around the globe. Together with the proposed benchmarks, we hope that this new open source resource will set the stage for substantial new research for the years to come.", + "bbox": [ + 75, + 657, + 467, + 763 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Though we are motivated by skill learning, Ego-Exo4D is poised for even broader influence, beyond the proposed benchmarks. Whereas existing datasets lack activity modeling in real-world 3D contexts (e.g., restricted to mocap suits and/or lab settings), Ego-Exo4D is a resource for general 3D vision—such as environment reconstruction, camera relocalization, audio-visual mapping, and many others. Similarly, our novel video-language resources will offer many opportunities for grounding of actions and objects, multi", + "bbox": [ + 75, + 763, + 467, + 900 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "modal representation learning, and language generation. Finally, though our tasks prioritize perception from the \"ego-only\" perspective, the exo component of our data ensures its utility for the more traditional exo viewpoint too, e.g., for activity recognition and body pose estimation.", + "bbox": [ + 498, + 90, + 890, + 167 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Contribution statement", + "text_level": 1, + "bbox": [ + 500, + 179, + 700, + 193 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This project is the result of a large collaboration between many institutions over the last two years. Initial authors represent the leadership team of the project. Kristen Grauman initiated the project, served as the technical lead, initiated the recognition and proficiency benchmarks and expert commentary, and coordinated their working groups. Andrew Westbury served as the program manager and operations lead for all aspects of the project. Lorenzo Torresani led development of the capture domains, initiated the relation and ego-pose benchmarks, and coordinated their working groups. Kris Kitani led development of the multi-camera rig and supported the Ego-Exo4D engineering team on all aspects of the data annotation and organization. Jitendra Malik served as a scientific advisor. Authors with stars $(^{*})$ were key drivers of implementation, collection, and/or annotation development throughout the project. Authors with daggers $(^{\\dagger})$ are faculty and senior researcher PIs for the project. The Appendices detail the contributions of individual authors for the various benchmarks, data collection, and annotation pipelines.", + "bbox": [ + 498, + 204, + 890, + 474 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 500, + 488, + 666, + 503 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We gratefully acknowledge the following colleagues for valuable discussions and support of our project: Vittorio Caggiano, Ilé Danza, Ahmad Darkhalil, Zona de Bloque, Alex Dinh, Rene Martinez Doehner, Ivan Cruz, Matt Feiszli, Vance Feutz, Kelly Forbes, Rohit Girdhar, Pierre Gleize, Andrés Hernández, Shun Iwase, Bolin Lai, Vivian Lee, Brighid Meredith, Ashley Massie, Natalia Neverova, Manohar Paluri, Joelle Pineau, Artsiom Sanakoyeu, Paresh Shenoy, Jiaray Shi, Jiasheng Shi, Gaurav Shrivastava, Mitesh Singh, Manasi Swaminathan, Arjang Talattof, Ali Thabet, Laurens van der Maaten, Andrea Vedaldi, and Toby Zhu. We also sincerely thank the 52 experts who contributed to the expert commentary for their expertise and support; they are listed individually in Appendix 12. Thank you to the Common Visual Data Foundation (CVDF) for hosting the Ego-Exo4D dataset. Finally, thank you to the 740 participants who contributed to this dataset and shared their skills in video.", + "bbox": [ + 498, + 512, + 890, + 733 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Univ. of Bristol is supported in part by EPSRC UMPIRE (EP/T004991/1) and EPSRC PG Visual AI (EP/T028572/1). Z. Zhu is supported by UoB-CSC Scholarship. Univ. of Catalonia is supported in part by the project Future Artificial Intelligence Research (FAIR) - PNRR MUR Cod. PE0000013 - CUP: E63C22001940006. Simon Fraser Univ. is supported in part by the Canada Research Chairs Program (CRC-2019-00298) and NSERC Discovery (2019-06489). Georgia Tech is supported in part by NSF award CNS-2308994. IU is supported in part by NSF DRL-2112635 (AI Institute for Engaged Learning). UT Austin is supported in part by the IFML NSF AI Institute.", + "bbox": [ + 498, + 734, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "19391", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 137, + 171, + 152 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Ahmed Tawfik Aboukhadra, Jameel Malik, Ahmed Elhayek, Nadia Robertini, and Didier Stricker. Thor-net: End-to-end graformer-based realistic two hands and object reconstruction with self-supervision. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1001-1010, 2023. 58, 59", + "[2] Michael Abrash. Creating the future: Augmented reality, the next human-machine interface. In 2021 IEEE International Electron Devices Meeting (IEDM), 2021. 46", + "[3] Karan Ahuja, Chris Harrison, Mayank Goel, and Robert Xiao. Mecap: Whole-body digitization for low-cost vr/ar headsets. In Proceedings of the 32nd Annual ACM Symposium on User Interface Software and Technology, pages 453-462, 2019. 9", + "[4] Jean-Baptiste Alayrac, Piotr Bojanowski, Nishant Agrawal, Josef Sivic, Ivan Laptev, and Simon Lacoste-Julien. Unsupervised learning from narrated instruction videos. In CVPR, 2016. 4", + "[5] Shervin Ardeshir and Ali Borji. Ego2top: Matching viewers in egocentric and top-view videos. In ECCV, 2016. 4", + "[6] Shervin Ardeshir and Ali Borji. Egocentric meets top-view. IEEE transactions on pattern analysis and machine intelligence, 41(6), 2018. 4", + "[7] Shervin Ardeshir and Ali Borji. An exocentric look at egocentric actions and vice versa. Computer Vision and Image Understanding, 171, 2018. 4, 7", + "[8] Kumar Ashutosh, Rohit Girdhar, Lorenzo Torresani, and Kristen Grauman. Hiervl: Learning hierarchical videolanguage embeddings. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023. 6", + "[9] Kumar Ashutosh, Santhosh Kumar Ramakrishnan, Triantafyllos Afouras, and Kristen Grauman. Video-mined task graphs for keystep recognition in instructional videos. In NeurIPS, 2023. 4, 7, 8, 29, 49", + "[10] Siddhant Bansal, Chetan Arora, and C.V. Jawahar. My view is the best view: Procedure learning from egocentric videos. In European Conference on Computer Vision (ECCV), 2022. 3, 4, 7, 49", + "[11] Yizhak Ben-Shabat, Xin Yu, Fatemehsadat Saleh, Dylan Campbell, Cristian Rodriguez-Opazo, Hongdong Li, and Stephen Gould. The ikea asm dataset: Understanding people assembling furniture through actions, objects and pose. 2020. 4", + "[12] Gedas Bertasius, Hyun Soo Park, Stella Yu, and Jianbo Shi. Am i a baller? basketball performance assessment from first-person videos. In ICCV, 2017. 3, 8", + "[13] Gedas Bertasius, Heng Wang, and Lorenzo Torresani. Is space-time attention all you need for video understanding? In ICML, 2021. 39, 43, 44, 47, 54, 55, 56", + "[14] Jing Bi, Jiebo Luo, and Chenliang Xu. Procedure planning in instructional videos via contextual modeling and model" + ], + "bbox": [ + 86, + 162, + 468, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "based policy learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15611-15620, 2021. 49", + "[15] Jing Bi, Jiebo Luo, and Chenliang Xu. Procedure planning in instructional videos via contextual modeling and model-based policy learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15611-15620, 2021. 4", + "[16] Kay Henning Brodersen, Cheng Soon Ong, Klaas Enno Stephan, and Joachim M Buhmann. The balanced accuracy and its posterior distribution. In 2010 20th international conference on pattern recognition, pages 3121-3124. IEEE, 2010. 35", + "[17] Meng Cao, Tianyu Yang, Junwu Weng, Can Zhang, Jue Wang, and Yuexian Zou. Locvtp: Video-text pre-training for temporal localization. In European Conference on Computer Vision, 2022. 4", + "[18] Angela Castillo, Maria Escobar, Guillaume Jeanneret, Albert Pumarola, Pablo Arbeláez, Ali Thabet, and Artsiom Sanakoyeu. Bodiffusion: Diffusing sparse observations for full-body human motion synthesis. CV4Metaverse workshop, International Conference on Computer Vision, 2023. 57, 58", + "[19] Eric R Chan, Koki Nagano, Matthew A Chan, Alexander W Bergman, Jeong Joon Park, Axel Levy, Miika Aittala, Shalini De Mello, Tero Karras, and Gordon Wetzstein. Generative novel view synthesis with 3d-aware diffusion models. arXiv preprint arXiv:2304.02602, 2023. 4", + "[20] Angel Chang, Angela Dai, Tom Funkhouser, Matthias Nießner, Manolis Savva, Shuran Song, Andy Zeng, and Yinda Zhang. Matterport3d: Learning from rgb-d data in indoor environments. In Proceedings of the International Conference on 3D Vision (3DV), 2017. MatterPort3D dataset license available at: http://kaldir.vc.in.tum.de/matterport/MP_TOS.pdf.3", + "[21] Chien-Yi Chang, De-An Huang, Danfei Xu, Ehsan Adeli, Li Fei-Fei, and Juan Carlos Niebles. Procedure planning in instructional videos. In European Conference on Computer Vision, pages 334–350. Springer, 2020. 49", + "[22] Chien-Yi Chang, De-An Huang, Danfei Xu, Ehsan Adeli, Li Fei-Fei, and Juan Carlos Niebles. Procedure planning in instructional videos. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part XI, pages 334–350. Springer, 2020. 4", + "[23] Yu-Hsin Chen, Tien-Ju Yang, Joel Emer, and Vivienne Sze. Eyeriss v2: A flexible accelerator for emerging deep neural networks on mobile devices. IEEE Journal on Emerging and Selected Topics in Circuits and Systems, 2019. 46", + "[24] Ho Kei Cheng and Alexander G Schwing. Xmem: Long-term video object segmentation with an atkinson-shiffrin memory model. In European Conference on Computer Vision, pages 640-658. Springer, 2022. 36", + "[25] MMPose Contributors. Openmmlab pose estimation toolbox and benchmark. https://github.com/openmmlab/mmpose, 2020.56", + "[26] Kellie Corona, Katie Osterdahl, Roderic Collins, and Anthony Hoogs. Meva: A large-scale multiview, multimodal video dataset for activity detection. In WACV, 2021. 3, 7" + ], + "bbox": [ + 509, + 93, + 890, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "19392", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[27] Dima Damen, Hazel Doughty, Giovanni Maria Farinella, Sanja Fidler, Antonino Furnari, Evangelos Kazakos, Davide Moltisanti, Jonathan Munro, Toby Perrett, Will Price, and Michael Wray. Scaling egocentric vision: The epic-kitchens dataset. In European Conference on Computer Vision (ECCV), 2018. 3, 27", + "[28] Dima Damen, Hazel Doughty, Giovanni Maria Farinella, Antonino Furnari, Jian Ma, Evangelos Kazakos, Davide Moltisanti, Jonathan Munro, Toby Perrett, Will Price, and Michael Wray. Rescaling egocentric vision. *IJCV*, 2021. 2, 3", + "[29] Dima Damen, Hazel Doughty, Giovanni Maria Farinella, Antonino Furnari, Evangelos Kazakos, Jian Ma, Davide Moltisanti, Jonathan Munro, Toby Perrett, Will Price, et al. Rescaling egocentric vision: collection, pipeline and challenges for epic-kitchens-100. International Journal of Computer Vision, 130(1):33-55, 2022. 7", + "[30] Roeland De Geest, Efstratios Gavves, Amir Ghodrati, Zhenyang Li, Cees Snoek, and Tinne Tuytelaars. Online action detection. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part V 14, pages 269-284. Springer, 2016. 46, 51", + "[31] Radosvet Desislavov, Fernando Martínez-Plumed, and José Hernández-Orallo. Trends in ai inference energy consumption: Beyond the performance-vs-parameter laws of deep learning. Sustainable Computing: Informatics and Systems, 38:100857, 2023. 46", + "[32] Guodong Ding, Fadime Sener, Shugao Ma, and Angela Yao. Every mistake counts in assembly. arXiv preprint arXiv:2307.16453, 2023. 8", + "[33] Keyan Ding, Kede Ma, Shiqi Wang, and Eero P Simoncelli. Image quality assessment: Unifying structure and texture similarity. IEEE transactions on pattern analysis and machine intelligence, 44(5):2567-2581, 2020. 39", + "[34] H Doughty, D Damen, and W Mayol-Cuevas. Who's better? who's best? pairwise deep ranking for skill determination. In CVPR, 2018. 3", + "[35] Hazel Doughty, Walterio Mayol-Cuevas, and Dima Damen. The Pros and Cons: Rank-aware Temporal Attention for Skill Determination in Long Videos. 2019. 3, 8", + "[36] Nikita Dvornik, Isma Hadji, Hai Pham, Dhaivat Bhatt, Brais Martinez, Afsaneh Fazly, and Allan D Jepson. Flow graph to video grounding for weakly-supervised multi-step localization. In ECCV, pages 319-335. Springer, 2022. 4, 8", + "[37] Ehsan Elhamifar and Dat Huynh. Self-supervised multitask procedure learning from instructional videos. In European Conference on Computer Vision, pages 557-573. Springer, 2020. 4", + "[38] Jakob Engel, Kiran Somasundaram, Michael Goesele, Albert Sun, Alexander Gamino, Andrew Turner, Arjang Talattof, Arnie Yuan, Bilal Souti, Brighid Meredith, Cheng Peng, Chris Sweeney, Cole Wilson, Dan Barnes, Daniel DeTone, David Caruso, Derek Valleroy, Dinesh Ginjupalli, Duncan Frost, Edward Miller, Elias Mueggler, Evgeniy Oleinik, Fan Zhang, Guruprasad Somasundaram, Gustavo" + ], + "bbox": [ + 86, + 90, + 472, + 902 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Solaira, Harry Lanaras, Henry Howard-Jenkins, Huixuan Tang, Hyo Jin Kim, Jaime Rivera, Ji Luo, Jing Dong, Julian Straub, Kevin Bailey, Kevin Eckenhoff, Lingni Ma, Luis Pesqueira, Mark Schwesinger, Maurizio Monge, Nan Yang, Nick Charron, Nikhil Raina, Omkar Parkhi, Peter Borschowa, Pierre Moulon, Prince Gupta, Raul Mur-Artal, Robbie Pennington, Sachin Kulkarni, Sagar Miglani, Santosh Gondi, Saransh Solanki, Sean Diener, Shangyi Cheng, Simon Green, Steve Saarinen, Suvam Patra, Tassos Mourikis, Thomas Whelan, Tripti Singh, Vasileios Balntas, Vijay Baiyya, Wilson Dreewes, Xiaqing Pan, Yang Lou, Yipu Zhao, Yusuf Mansour, Yuyang Zou, Zhaoyang Lv, Zijian Wang, Mingfei Yan, Carl Ren, Renzo De Nardi, and Richard Newcombe. Project Aria: A new tool for egocentric multi-modal AI research, 2023. 3, 4, 2", + "[39] Steven K Esser, Jeffrey L McKinstry, Deepika Bablani, Rathinakumar Appuswamy, and Dharmendra S Modha. Learned step size quantization. arXiv preprint arXiv:1902.08153, 2019. 8", + "[40] Chenyou Fan, Jangwon Lee, Mingze Xu, Krishna Kumar Singh, Yong Jae Lee, David J Crandall, and Michael S Ryoo. Identifying first-person camera wearers in third-person videos. In CVPR, 2017. 4", + "[41] Christoph Feichtenhofer. X3d: Expanding architectures for efficient video recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 203-213, 2020. 8, 45, 46, 47, 48", + "[42] John H. Flavell, Eleanor R. Flavell, Frances L. Green, and Sharon A. Wilcox. The development of three spatial perspective-taking rules. Child Development, 1981. 2", + "[43] Ruohan Gao, Tae-Hyun Oh, Kristen Grauman, and Lorenzo Torresani. Listen to look: Action recognition by previewing audio. In CVPR, 2020. 8", + "[44] Amir Ghodrati, Babak Ehteshami Bejnordi, and Amirhossein Habibian. Frameexit: Conditional early exiting for efficient video recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2021. 8", + "[45] Rohit Girdhar, Mannat Singh, Nikhila Ravi, Laurens van der Maaten, Armand Joulin, and Ishan Misra. Omnivore: A single model for many visual modalities. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16102-16112, 2022. 54", + "[46] Raghav Goyal, Samira Ebrahimi Kahou, Vincent Michalski, Joanna Materzynska, Susanne Westphal, Heuna Kim, Valentin Haenel, Ingo Fruend, Peter Yianilos, Moritz Mueller-Freitag, et al. The \"something something\" video database for learning and evaluating visual common sense. In Proceedings of the IEEE international conference on computer vision, pages 5842-5850, 2017. 42", + "[47] Kristen Grauman, Andrew Westbury, Eugene Byrne, Zachary Chavis, Antonino Furnari, Rohit Girdhar, Jackson Hamburger, Hao Jiang, Miao Liu, Xingyu Liu, Miguel Martin, Tushar Nagarajan, Ilija Radosavovic, Santhosh Kumar Ramakrishnan, Fiona Ryan, Jayant Sharma, Michael Wray, Mengmeng Xu, Eric Zhongcong Xu, Chen Zhao, Siddhant Bansal, Dhruv Batra, Vincent Cartillier, Sean Crane, Tien Do, Morrie Doulaty, Akshay Erapalli, Christoph" + ], + "bbox": [ + 509, + 90, + 893, + 902 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "19393", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Feichtenhofer, Adriano Fragomeni, Qichen Fu, Abraham Gebreselasie, Cristina Gonzalez, James Hillis, Xuhua Huang, Yifei Huang, Wenqi Jia, Weslie Khoo, Jachym Kolář, Satwik Kottur, Anurag Kumar, Federico Landini, Chao Li, Yanghao Li, Zhenqiang Li, Karttikeya Mangalam, Raghava Modhugu, Jonathan Munro, Tullie Murrell, Takumi Nishiyasu, Will Price, Paola Ruiz, Merey Ramazanova, Leda Sari, Kiran Somasundaram, Audrey Southerland, Yusuke Sugano, Ruijie Tao, Minh Vo, Yuchen Wang, Xindi Wu, Takuma Yagi, Ziwei Zhao, Yunyi Zhu, Pablo Arbelaez, David Crandall, Dima Damen, Giovanni Maria Farinella, Christian Fuegen, Bernard Ghanem, Vamsi Krishna Ithapu, C. V. Jawahar, Hanbyul Joo, Kris Kitani, Haizhou Li, Richard Newcombe, Aude Oliva, Hyun Soo Park, James M. Rehg, Yoichi Sato, Jianbo Shi, Mike Zheng Shou, Antonio Torralba, Lorenzo Torresani, Mingfei Yan, and Jitendra Malik. Ego4D: Around the world in 3,000 hours of egocentric video. In CVPR, 2022. 2, 3, 4, 6, 7, 29, 34, 43", + "[48] Chunhui Gu, Chen Sun, David A. Ross, Carl Vondrick, Caroline Pantofaru, Yeqing Li, Sudheendra Vijayanarasimhan, George Toderici, Susanna Ricco, Rahul Sukthankar, Cordelia Schmid, and Jitendra Malik. Ava: A video dataset of spatio-temporally localized atomic visual actions. In CVPR, 2018. 2", + "[49] Vladimir Guzov, Aymen Mir, Torsten Sattler, and Gerard Pons-Moll. Human poseitioning system (hps): 3d human pose estimation and self-localization in large scenes from body-mounted sensors. In CVPR, 2021. 3", + "[50] Shreyas Hampali, Mahdi Rad, Markus Oberweger, and Vincent Lepetit. Honnotate: A method for 3d annotation of hand and object poses. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3196-3206, 2020. 9", + "[51] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, 2016. 36, 58", + "[52] Geoffrey Hinton, Oriol Vinyals, and Jeff Dean. Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531, 2015. 43, 44", + "[53] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in Neural Information Processing Systems, 33:6840-6851, 2020. 57", + "[54] Alain Hore and Djemel Ziou. Image quality metrics: Psnr vs. ssim. In 2010 20th international conference on pattern recognition, pages 2366-2369. IEEE, 2010. 39", + "[55] Mark Horowitz. 1.1 computing's energy problem (and what we can do about it). In 2014 IEEE international solid-state circuits conference digest of technical papers (ISSCC), 2014. 46", + "[56] Andrew G Howard, Menglong Zhu, Bo Chen, Dmitry Kalenichenko, Weijun Wang, Tobias Weyand, Marco Andretto, and Hartwig Adam. Mobilenets: Efficient convolutional neural networks for mobile vision applications. arXiv preprint arXiv:1704.04861, 2017. 8", + "[57] Dong-Hyun Hwang, Kohei Aso, Ye Yuan, Kris Kitani, and Hideki Koike. Monoeye: Multimodal human motion capture system using a single ultra-wide fisheye camera. In" + ], + "bbox": [ + 86, + 90, + 468, + 900 + ], + "page_idx": 11 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Proceedings of the 33rd Annual ACM Symposium on User Interface Software and Technology, pages 98-111, 2020. 9", + "[58] Forrest N Iandola, Song Han, Matthew W Moskewicz, Khalid Ashraf, William J Dally, and Kurt Keutzer. SqueezeNet: Alexnet-level accuracy with 50x fewer parameters and; 0.5 mb model size. arXiv preprint arXiv:1602.07360, 2016. 8", + "[59] Vladimir Iashin and Esa Rahtu. Multi-modal dense video captioning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops, pages 958-959, 2020. 26", + "[60] Hassan Ismail Fawaz, Germain Forestier, Jonathan Weber, Lhassane Idoumghar, and Pierre-Alain Muller. Evaluating surgical skills from kinematic data using convolutional neural networks. In Medical Image Computing and Computer Assisted Intervention - MICCAI 2018, pages 214-221, 2018. 8", + "[61] Phillip Isola, Jun-Yan Zhu, Tinghui Zhou, and Alexei A Efros. Image-to-image translation with conditional adversarial networks. CVPR, 2017. 39", + "[62] Yunseok Jang, Sungryull Sohn, Lajanugen Logeswaran, Tiange Luo, Moontae Lee, and Honglak Lee. Multimodal subtask graph generation from instructional videos. arXiv preprint arXiv:2302.08672, 2023. 8", + "[63] Hao Jiang and Kristen Grauman. Seeing invisible poses: Estimating 3d body pose from egocentric video. In CVPR, 2017. 9", + "[64] Jiaxi Jiang, Paul Streli, Huajian Qiu, Andreas Fender, Larissa Laich, Patrick Snape, and Christian Holz. Avatarposer: Articulated full-body pose tracking from sparse motion sensing. In European Conference on Computer Vision, pages 443-460. Springer, 2022. 57, 58", + "[65] Wei Jiang, Eduard Trulls, Jan Hosang, Andrea Tagliasacchi, and Kwang Moo Yi. Cotr: Correspondence transformer for matching across images. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6207-6217, 2021. 6, 35", + "[66] Hanbyul Joo, Tomas Simon, Xulong Li, Hao Liu, Lei Tan, Lin Gui, Sean Banerjee, Timothy Scott Godisart, Bart Nabbe, Iain Matthews, Takeo Kanade, Shohei Nobuhara, and Yaser Sheikh. Panoptic studio: A massively multiview system for social interaction capture. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2017. 3", + "[67] Will Kay, Joao Carreira, Karen Simonyan, Brian Zhang, Chloe Hillier, Sudheendra Vijayanarasimhan, Fabio Viola, Tim Green, Trevor Back, Paul Natsev, et al. The kinetics human action video dataset. arXiv preprint arXiv:1705.06950, 2017. 2, 42, 43", + "[68] Rawal Khirodkar, Aayush Bansal, Lingni Ma, Richard Newcombe, Minh Vo, and Kris Kitani. Egohumans: An egocentric 3d multi-human benchmark. In ICCV, 2023. 3", + "[69] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 36", + "[70] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C. Berg, Wan-Yen Lo, Piotr Dollar," + ], + "bbox": [ + 509, + 92, + 890, + 900 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "19394", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 11 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "and Ross Girshick. Segment anything. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 4015-4026, 2023. 33", + "[71] Dohwan Ko, Joonmyung Choi, Juyeon Ko, Shinyeong Noh, Kyoung-Woon On, Eun-Sol Kim, and Hyunwoo J Kim. Video-text representation learning via differentiable weak temporal alignment. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5016-5025, 2022. 4", + "[72] Nikos Kolotouros, Georgios Pavlakos, Michael J Black, and Kostas Daniilidis. Learning to reconstruct 3d human pose and shape via model-fitting in the loop. In Proceedings of the IEEE/CVF international conference on computer vision, pages 2252-2261, 2019. 59", + "[73] Bruno Korbar, Du Tran, and Lorenzo Torresani. Scsampler: Sampling salient clips from video for efficient action recognition. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2019. 8", + "[74] Zuzana Kukelova, Jan Heller, and Andrew Fitzgibbon. Efficient intersection of three quadrics and applications in computer vision. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 1799-1808, 2016. 3", + "[75] Iljung Kwak, Jian-Zhong Guo, Adam Hantman, David Kriegman, and Kristin Branson. Detecting the starting frame of actions in video. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 489-497, 2020. 54", + "[76] Taein Kwon, Bugra Tekin, Jan Stühmer, Federica Bogo, and Marc Pollefeys. H2o: Two hands manipulating objects for first person interaction recognition. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 10138-10148, 2021. 2, 3, 9, 7", + "[77] F. De la Torre, J. Hodgins, J. Montano, S. Valcarcel, R. Forcada, and J. Macey. Guide to the carnegie mellon university multimodal activity (cmu-mmac) database. In Tech. report CMU-RI-TR-08-22, Robotics Institute, Carnegie Mellon University, 2009. 2, 3, 7", + "[78] Y. J. Lee, J. Ghosh, and K. Grauman. Discovering important people and objects for egocentric video summarization. In CVPR, 2012. 3", + "[79] Jiefeng Li, Chao Xu, Zhicun Chen, Siyuan Bian, Lixin Yang, and Cewu Lu. Hybrik: A hybrid analytical-neural inverse kinematics solution for 3d human pose and shape estimation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3383-3393, 2021. 59", + "[80] Jiaman Li, Karen Liu, and Jiajun Wu. Ego-body pose estimation via ego-head pose estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17142-17151, 2023. 3, 9, 56, 57, 58", + "[81] Yin Li, Miao Liu, and James M. Rehg. In the eye of beholder: Joint learning of gaze and actions in first person video. In ECCV, 2018. 3, 7", + "[82] Yanghao Li, Tushar Nagarajan, Bo Xiong, and Kristen Grauman. Ego-exo: Transferring visual representations from third-person to first-person videos. In Proceedings of" + ], + "bbox": [ + 86, + 90, + 470, + 900 + ], + "page_idx": 12 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6943-6953, 2021. 4, 7, 21, 43, 44", + "[83] Junhua Liao, Haihan Duan, Kanghui Feng, Wanbing Zhao, Yanbing Yang, and Liangyin Chen. A light weight model for active speaker detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22932-22941, 2023. 47, 48", + "[84] Kevin Lin, Lijuan Wang, and Zicheng Liu. End-to-end human pose and mesh reconstruction with transformers. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1954-1963, 2021. 59", + "[85] Kevin Qinghong Lin, Alex Jinpeng Wang, Mattia Soldan, Michael Wray, Rui Yan, Eric Zhongcong Xu, Difei Gao, Rongcheng Tu, Wenzhe Zhao, Weijie Kong, et al. Egocentric video-language pretraining. NeurIPS, 2022. 6, 26, 50", + "[86] TY Lin, Y Cui, S Belongie, and J Hays. Learning deep representations for ground-to-aerial geolocation. In CVPR, 2015. 4", + "[87] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dolkar, and C Lawrence Zitnick. Microsoft COCO: Common objects in context. In ECCV, 2014. 9", + "[88] Tsung-Yi Lin, Piotr Dólar, Ross Girshick, Kaiming He, Bharath Hariharan, and Serge Belongie. Feature pyramid networks for object detection. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2117-2125, 2017. 58", + "[89] Xudong Lin, Fabio Petroni, Gedas Bertasius, Marcus Rohrbach, Shih-Fu Chang, and Lorenzo Torresani. Learning to recognize procedural activities with distant supervision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13853-13863, 2022. 4, 29", + "[90] Andrew Liu, Richard Tucker, Varun Jampani, Ameesh Makadia, Noah Snavely, and Angjoo Kanazawa. Infinite nature: Perpetual view generation of natural scenes from a single image. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14458-14467, 2021. 4", + "[91] Chiao Liu, Lyle Bainbridge, Andrew Berkovich, Song Chen, Wei Gao, Tsung-Hsun Tsai, Kazuya Mori, Rimon Ikeno, Masayuki Uno, Toshiyuki Isozaki, et al. A $4.6\\mu \\mathrm{m}$ , $512\\times 512$ , ultra-low power stacked digital pixel sensor with triple quantization and 127db dynamic range. In 2020 IEEE International Electron Devices Meeting (IEDM), 2020. 46", + "[92] Daochang Liu, Qiyue Li, Tingting Jiang, Yizhou Wang, Rulin Miao, Fei Shan, and Ziyu Li. Towards unified surgical skill assessment. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9522-9531, 2021. 8", + "[93] Gaowen Liu, Hao Tang, Hugo Latapie, and Yan Yan. Exocentric to egocentric image generation via parallel generative adversarial network. In ICASSP 2020-2020 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 1843-1847. IEEE, 2020. 7", + "[94] Gaowen Liu, Hao Tang, Hugo M Latapie, Jason J Corso, and Yan Yan. Cross-view exocentric to egocentric video" + ], + "bbox": [ + 509, + 92, + 890, + 900 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "19395", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 12 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "synthesis. In Proceedings of the 29th ACM International Conference on Multimedia, pages 974-982, 2021. 7", + "[95] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In International Conference on Learning Representations, 2019. 47", + "[96] Xiaohu Lu, Zuoyue Li, Zhaopeng Cui, Martin R. Oswald, Marc Pollefeys, and Rongjun Qin. Geometry-aware satellite-to-ground image synthesis for urban areas. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 7", + "[97] Mi Luo, Zihui Xue, Alex Dimakis, and Kristen Grauman. Put myself in your shoes: Lifting the egocentric perspective from exocentric videos. arXiv:2403.06351, 2024. 7", + "[98] Zhengyi Luo, Ryo Hachiuma, Ye Yuan, and Kris Kitani. Dynamics-regulated kinematic policy for egocentric pose estimation. In Advances in Neural Information Processing Systems, 2021. 9, 56, 57, 58", + "[99] Naureen Mahmood, Nima Ghorbani, Nikolaus F. Troje, Gerard Pons-Moll, and Michael J. Black. Amass: Archive of motion capture as surface shapes. In The IEEE International Conference on Computer Vision (ICCV), 2019. 57", + "[100] Effrosyni Mavroudi, Triantafyllos Afouras, and Lorenzo Torresani. Learning to ground instructional articles in videos through narrations. 2022. 7", + "[101] Sachin Mehta and Mohammad Rastegari. Mobilevit: lightweight, general-purpose, and mobile-friendly vision transformer. arXiv preprint arXiv:2110.02178, 2021. 8", + "[102] Yue Meng, Chung-Ching Lin, Rameswar Panda, Prasanna Sattigeri, Leonid Karlinsky, Aude Oliva, Kate Saenko, and Rogerio Feris. Ar-net: Adaptive frame resolution for efficient action recognition. In ECCV 2020, 2020. 8", + "[103] Antoine Miech, Dimitri Zhukov, Jean-Baptiste Alayrac, Makarand Tapaswi, Ivan Laptev, and Josef Sivic. HowTo100M: Learning a Text-Video Embedding by Watching Hundred Million Narrated Video Clips. In ICCV, 2019. 2, 4, 29", + "[104] Antoine Miech, Jean-Baptiste Alayrac, Lucas Smaira, Ivan Laptev, Josef Sivic, and Andrew Zisserman. End-to-end learning of visual representations from uncurated instructional videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9879-9889, 2020. 4, 29", + "[105] Mathew Monfort, Alex Andonian, Bolei Zhou, Kandan Ramakrishnan, Sarah Adel Bargal, Tom Yan, Lisa Brown, Quanfu Fan, Dan Gutfreund, Carl Vondrick, and Aude Oliva. Moments in time dataset: one million videos for event understanding. PAMI, 2019. 2", + "[106] Gyeongsik Moon, Shoou-I Yu, He Wen, Takaaki Shiratori, and Kyoung Mu Lee. Interhand2. 6m: A dataset and baseline for 3d interacting hand pose estimation from a single rgb image. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XX 16, pages 548-564. Springer, 2020. 9", + "[107] Medhini Narasimhan, Licheng Yu, Sean Bell, Ning Zhang, and Trevor Darrell. Learning and verification of task structure in instructional videos. arXiv preprint arXiv:2303.13519, 2023. 4, 8" + ], + "bbox": [ + 78, + 90, + 468, + 898 + ], + "page_idx": 13 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[108] Nora Newcombe. The development of spatial perspective taking. Advances in child development and behavior, 1989. 2", + "[109] Takehiko Ohkawa, Kun He, Fadime Sener, Tomas Hodan, Luan Tran, and Cem Keskin. Assemblyhands: Towards egocentric activity understanding via 3d hand pose estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12999-13008, 2023. 9, 7", + "[110] Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748, 2018. 43, 44", + "[111] Boxiao Pan, Haoye Cai, De-An Huang, Kuan-Hui Lee, Adrien Gaidon, Ehsan Adeli, and Juan Carlos Niebles. Spatio-temporal graph for video captioning with knowledge distillation. In CVPR, 2020. 26", + "[112] JoonKyu Park, Yeonguk Oh, Gyeongsik Moon, Hongsuk Choi, and Kyoung Mu Lee. Handoccnet: Occlusion-robust 3d hand mesh estimation network. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1496–1505, 2022. 58, 59", + "[113] Paritosh Parmar and Brendan Morris. Action quality assessment across multiple actions. In WACV, 2019. 3", + "[114] Paritosh Parmar and Brendan Tran Morris. Learning to score olympic events, 2017. 8", + "[115] Paritosh Parmar and Brendan Tran Morris. What and how well you performed? a multitask learning approach to action quality assessment. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 304-313, 2019. 8", + "[116] William Peebles and Saining Xie. Scalable diffusion models with transformers. arXiv preprint arXiv:2212.09748, 2022.39", + "[117] Federico Perazzi, Jordi Pont-Tuset, Brian McWilliams, Luc Van Gool, Markus Gross, and Alexander Sorkine-Hornung. A benchmark dataset and evaluation methodology for video object segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 724-732, 2016. 35, 39", + "[118] Ethan Perez, Florian Strub, Harm De Vries, Vincent Dumoulin, and Aaron Courville. Film: Visual reasoning with a general conditioning layer. In Proceedings of the AAAI conference on artificial intelligence, 2018. 39", + "[119] H. Pirsiavash and D. Ramanan. Detecting activities of daily living in first-person camera views. In CVPR, 2012. 3", + "[120] H. Piriavash, C. Vondrick, and A. Torralba. Assessing the quality of actions. In ECCV, 2014. 3, 8", + "[121] Antonio Polino, Razvan Pascanu, and Dan Alistarh. Model compression via distillation and quantization. arXiv preprint arXiv:1802.05668, 2018. 8", + "[122] Daniel Povey, Arnab Ghoshal, Gilles Boulianne, Lukávs. Burget, Ondrej Glembek, Nagendra Kumar Goel, Mirko Hannemann, Petr Motlícek, Yanmin Qian, Petr Schwarz, Jan Silovsky, Georg Stemmer, and Karel Vesely. The kaldi speech recognition toolkit. 2011. 47", + "[123] Shraman Pramanick, Yale Song, Sayan Nag, Kevin Qinghong Lin, Hardik Shah, Mike Zheng Shou," + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "19396", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 13 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Rama Chellappa, and Pengchuan Zhang. Egovlpv2: Egocentric video-language pre-training with fusion in the backbone. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5285-5297, 2023. 6, 26, 43, 44", + "[124] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 39", + "[125] Alec Radford, Jong Wook Kim, Tao Xu, Greg Brockman, Christine McLeavey, and Ilya Sutskever. Robust speech recognition via large-scale weak supervision. In International Conference on Machine Learning, pages 28492-28518. PMLR, 2023. 27", + "[126] Francesco Ragusa, Antonino Furnari, Salvatore Livatino, and Giovanni Maria Farinella. The meccano dataset: Understanding human-object interactions from egocentric videos in an industrial-like domain. In WACV, 2021. 3, 7", + "[127] N. Rai, H. Chen, J. Ji, R. Desai, K. Kozuka, S. Ishizaka, E. Adeli, and J.C. Niebles. Home action genome: Contrastive compositional action understanding. In CVPR, 2021. 2, 3", + "[128] Santhosh Kumar Ramakrishnan, Aaron Gokaslan, Erik Wijmans, Oleksandr Maksymets, Alexander Clegg, John M Turner, Eric Undersander, Wojciech Galuba, Andrew Westbury, Angel X Chang, Manolis Savva, Yili Zhao, and Dhruv Batra. Habitat-matterport 3d dataset (HM3d): 1000 large-scale 3d environments for embodied AI. In Thirty-fifth Conference on Neural Information Processing Systems Datasets and Benchmarks Track, 2021. 3", + "[129] Scott Reed, Honglak Lee, Dragomir Anguelov, Christian Szegedy, Dumitru Erhan, and Andrew Rabinovich. Training deep neural networks on noisy labels with bootstrapping. arXiv preprint arXiv:1412.6596, 2014. 39", + "[130] Krishna Regmi and Ali Borji. Cross-view image synthesis using conditional gans. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 4, 7", + "[131] Krishna Regmi and Ali Borji. Cross-view image synthesis using geometry-guided conditional gans. Computer Vision and Image Understanding, 2019. 4", + "[132] Krishna Regmi and Mubarak Shah. Bridging the domain gap for ground-to-aerial image matching. In ICCV, 2019. 4", + "[133] Jeremy Reizenstein, Roman Shapovalov, Philipp Henzler, Luca Sbordone, Patrick Labatut, and David Novotny. Common objects in 3d: Large-scale learning and evaluation of real-life 3d category reconstruction. In ICCV, 2021. 3", + "[134] Bin Ren, Hao Tang, and Nicu Sebe. Cascaded cross mlp-mixer gans for cross-view image translation. arXiv preprint arXiv:2110.10183, 2021. 4", + "[135] Xuanchi Ren and Xiaolong Wang. Look outside the room: Synthesizing a consistent long-term 3d scene video from a single image. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3563-3573, 2022. 4", + "[136] Helge Rhodin, Christian Richardt, Dan Casas, Eldar In-safutdinov, Mohammad Shafiei, Hans-Peter Seidel, Bernt" + ], + "bbox": [ + 78, + 90, + 468, + 898 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Schiele, and Christian Theobalt. Egocap: egocentric marker-less motion capture with two fisheye cameras. ACM Transactions on Graphics (TOG), 35(6):1-11, 2016. 9", + "[137] Robin Rombach, Patrick Esser, and Björn Ommer. Geometry-free view synthesis: Transformers and no 3d priors. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14356-14366, 2021. 4", + "[138] Javier Romero, Dimitrios Tzionas, and Michael J. Black. Embodied hands: Modeling and capturing hands and bodies together. ACM Transactions on Graphics, (Proc. SIGGRAPH Asia), 36(6), 2017. 59", + "[139] Fadime Sener, Dibyadip Chatterjee, Daniel Shelepov, Kun He, Dipika Singhania, Robert Wang, and Angela Yao. Assembly101: A large-scale multi-view video dataset for understanding procedural activities. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21096-21106, 2022. 2, 3, 8, 7", + "[140] Pierre Sermanet, Corey Lynch, Jasmine Hsu, and Sergey Levine. Time-contrastive networks: Self-supervised learning from multi-view observation. In 2017 IEEE Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), pages 486–487. IEEE, 2017. 43", + "[141] Pierre Sermanet, Corey Lynch, Yevgen Chebotar, Jasmine Hsu, Eric Jang, Stefan Schaal, and Sergey Levine. Time-contrastive networks: Self-supervised learning from video. Proceedings of International Conference in Robotics and Automation (ICRA), 2018. 4", + "[142] Xi Shen, Alexei A Efros, Armand Joulin, and Mathieu Aubry. Learning co-segmentation by segment swapping for retrieval and discovery. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5082-5092, 2022. 35, 36", + "[143] Nina Shvetsova, Brian Chen, Andrew Rouditchenko, Samuel Thomas, Brian Kingsbury, Rogerio S Feris, David Harwath, James Glass, and Hilde Kuehne. Everything at once-multi-modal fusion transformer for video retrieval. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20020-20029, 2022. 4", + "[144] Gunnar A Sigurdsson, Abhinav Gupta, Cordelia Schmid, Ali Farhadi, and Karteek Alahari. Actor and observer: Joint modeling of first and third-person videos. In CVPR, 2018, 4, 7, 43", + "[145] Gunnar A Sigurdsson, Abhinav Gupta, Cordelia Schmid, Ali Farhadi, and Karteek Alahari. Charades-ego: A large-scale dataset of paired third and first person videos. arXiv preprint arXiv:1804.09626, 2018. 2, 3, 7", + "[146] Tomas Simon, Hanbyul Joo, Iain Matthews, and Yaser Sheikh. Hand keypoint detection in single images using multiview bootstrapping. In Proceedings of the IEEE conference on Computer Vision and Pattern Recognition, pages 1145-1153, 2017. 9", + "[147] Krishna Kumar Singh, Kayvon Fatahalian, and Alexei A. Efros. Krishnamac: Using a longitudinal, single-person, egocentric dataset for scene understanding tasks. In WACV, 2016. 3", + "[148] Yale Song, Eugene Byrne, Tushar Nagarajan, Huiyu Wang, Miguel Martin, and Lorenzo Torresani. Ego4d goal-step:" + ], + "bbox": [ + 503, + 92, + 890, + 900 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "19397", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Toward hierarchical understanding of procedural activities. In NeurIPS, 2023. 7", + "[149] Khurram Soomro, Amir Roshan Zamir, and Mubarak Shah. Ucf101: A dataset of 101 human action classes from videos in the wild. In CRCV-TR-12-01, 2012. 2, 42", + "[150] Bilge Soran, Ali Farhadi, and Linda Shapiro. Generating notifications for missing actions: Don't forget to turn the lights off! In ICCV, pages 4669-4677, 2015. 8", + "[151] Julian Straub, Thomas Whelan, Lingni Ma, Yufan Chen, Erik Wijmans, Simon Green, Jakob J. Engel, Raul Mur-Artal, Carl Ren, Shobhit Verma, Anton Clarkson, Mingfei Yan, Brian Budge, Yajie Yan, Xiaqing Pan, June Yon, Yuyang Zou, Kimberly Leon, Nigel Carter, Jesus Briales, Tyler Gillingham, Elias Mueggler, Luis Pesqueira, Manolis Savva, Dhruv Batra, Hauke M. Strasdat, Renzo De Nardi, Michael Goesele, Steven Lovegrove, and Richard Newcombe. The Replica dataset: A digital replica of indoor spaces. arXiv preprint arXiv:1906.05797, 2019. 3", + "[152] Carole H Sudre, Wenqi Li, Tom Vercauteren, Sebastien Ourselin, and M Jorge Cardoso. Generalised dice overlap as a deep learning loss function for highly unbalanced segmentations. In Deep Learning in Medical Image Analysis and Multimodal Learning for Clinical Decision Support: Third International Workshop, DLMIA 2017, and 7th International Workshop, ML-CDS 2017, Held in conjunction with MICCAI 2017, Quebec City, QC, Canada, September 14, Proceedings 3, pages 240-248. Springer, 2017. 39", + "[153] Vivienne Sze, Yu-Hsin Chen, Tien-Ju Yang, and Joel S Emer. How to evaluate deep neural network processors: Tops/w (alone) considered harmful. IEEE Solid-State Circuits Magazine, 2020. 46", + "[154] Mukund Varma T, Peihao Wang, Xuxi Chen, Tianlong Chen, Subhashini Venugopalan, and Zhangyang Wang. Is attention all that neRF needs? In The Eleventh International Conference on Learning Representations, 2023. 39", + "[155] Mingxing Tan, Bo Chen, Ruoming Pang, Vijay Vasudevan, Mark Sandler, Andrew Howard, and Quoc V Le. Mnasnet: Platform-aware neural architecture search for mobile. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2820-2828, 2019. 8", + "[156] Shuhan Tan, Tushar Nagarajan, and Kristen Grauman. Egodistill: Egocentric head motion distillation for efficient video understanding. NeurIPS, 2023. 8", + "[157] Hao Tang, Dan Xu, Nicu Sebe, Yanzhi Wang, Jason J Corso, and Yan Yan. Multi-channel attention selection gan with cascaded semantic guidance for cross-view image translation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2417-2426, 2019. 4, 7", + "[158] Hao Tang, Kevin Liang, Kristen Grauman, Matt Feiszli, and Weiyao Wang. Egotracks: A long-term egocentric visual object tracking dataset. Advances in Neural Information Processing Systems, 2023. 6", + "[159] Yansong Tang, Jiwen Lu, and Jie Zhou. Comprehensive instructional video analysis: The coin dataset and performance evaluation. IEEE transactions on pattern analysis and machine intelligence, 2020. 2, 4, 7" + ], + "bbox": [ + 78, + 90, + 468, + 898 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[160] Zachary Teed and Jia Deng. DROID-SLAM: Deep Visual SLAM for Monocular, Stereo, and RGB-D Cameras. Advances in neural information processing systems, 2021. 57", + "[161] Denis Tome, Patrick Peluse, Lourdes Agapito, and Hernan Badino. xr-egopose: Egocentric 3d human pose from an hmd camera. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2019. 9", + "[162] Zhan Tong, Yibing Song, Jue Wang, and Limin Wang. Videomae: Masked autoencoders are data-efficient learners for self-supervised video pre-training. 2022. 43", + "[163] Vadim Tschernezki, Ahmad Darkhalil, Zhifan Zhu, David Fouhey, Iro Larina, Diane Larlus, Dima Damen, and Andrea Vedaldi. EPIC Fields: Marrying 3D Geometry and Video Understanding. In Proceedings of the Neural Information Processing Systems (NeurIPS), 2023. 3", + "[164] Hung-Yu Tseng, Qinbo Li, Changil Kim, Suhib Alsisan, Jia-Bin Huang, and Johannes Kopf. Consistent view synthesis with pose-guided diffusion models. arXiv preprint arXiv:2303.17598, 2023. 4", + "[165] Pavan Kumar Anasosalu Vasu, James Gabriel, Jeff Zhu, Oncel Tuzel, and Anurag Ranjan. Mobileone: An improved one millisecond mobile backbone. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7907-7917, 2023. 8", + "[166] Sara Vicente, Carsten Rother, and Vladimir Kolmogorov. Object cosegmentation. In CVPR 2011, pages 2217-2224. IEEE, 2011. 6", + "[167] Hanlin Wang, Yilu Wu, Sheng Guo, and Limin Wang. Pdpp: Projected diffusion for procedure planning in instructional videos. arXiv preprint arXiv:2303.14676, 2023. 4", + "[168] Daniel Watson, William Chan, Ricardo Martin-Brualla, Jonathan Ho, Andrea Tagliasacchi, and Mohammad Norouzi. Novel view synthesis with diffusion models. arXiv preprint arXiv:2210.04628, 2022. 4", + "[169] Daniel Weinland, Remi Ronfard, and Edmond Boyer. Free viewpoint action recognition using motion history volumes. Computer Vision and Image Understanding (CVIU), 2006. 3, 7", + "[170] Yangming Wen, Krishna Kumar Singh, Markham Anderson, Wei-Pang Jan, and Yong Jae Lee. Seeing the unseen: Predicting the first-person camera wearer's location and pose in third-person scenes. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) Workshops, pages 3446-3455, 2021. 4", + "[171] Olivia Wiles, Georgia Gkioxari, Richard Szeliski, and Justin Johnson. Synsin: End-to-end view synthesis from a single image. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7467-7477, 2020. 4", + "[172] Benita Wong, Joya Chen, You Wu, Stan Weixian Lei, Dongxing Mao, Difei Gao, and Mike Zheng Shou. Assistq: Affordance-centric question-driven task completion for egocentric assistant. In European Conference on Computer Vision, 2022. 3", + "[173] Zhirong Wu, Shuran Song, Aditya Khosla, Fisher Yu, Linguang Zhang, Xiaou Tang, and Jianxiong Xiao. 3d shapenets: A deep representation for volumetric shapes. In" + ], + "bbox": [ + 503, + 92, + 890, + 900 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "19398", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Computer Vision and Pattern Recognition, IEEE Conference on, 2015. 3", + "[174] Zuxuan Wu, Tushar Nagarajan, Abhishek Kumar, Steven Rennie, Larry S Davis, Kristen Grauman, and Rogerio Feris. Blockdrop: Dynamic inference paths in residual networks. In CVPR, 2018. 8", + "[175] Fei Xia, Amir R. Zamir, Zhi-Yang He, Alexander Sax, Jitendra Malik, and Silvio Savarese. Gibson Env: real-world perception for embodied agents. In CVPR. IEEE, 2018. Gibson license is available at http://svl.stanford.edu/gibson2/assets/GDS_agreement.pdf.3", + "[176] J. Xiao, A. Owens, and A. Torralba. Sun3d: A database of big spaces reconstructed using sfm and object labels. In ICCV, 2013. 3", + "[177] Frank F Xu, Lei Ji, Botian Shi, Junyi Du, Graham Neubig, Yonatan Bisk, and Nan Duan. A benchmark for structured procedural knowledge extraction from cooking videos. arXiv preprint arXiv:2005.00706, 2020. 8", + "[178] Hu Xu, Gargi Ghosh, Po-Yao Huang, Dmytro Okhonko, Armen Aghajanyan, Florian Metze, Luke Zettlemoyer, and Christoph Feichtenhofer. Videoclip: Contrastive pretraining for zero-shot video-text understanding. arXiv preprint arXiv:2109.14084, 2021. 4", + "[179] Mingze Xu, Chenyou Fan, Yuchen Wang, Michael S Ryoo, and David J Crandall. Joint person segmentation and identification in synchronized first-and third-person videos. In ECCV, 2018. 4", + "[180] Mingze Xu, Yuanjun Xiong, Hao Chen, Xinyu Li, Wei Xia, Zhuowen Tu, and Stefano Soatto. Long short-term transformer for online action detection. Advances in Neural Information Processing Systems, 34:1086-1099, 2021. 49", + "[181] Weipeng Xu, Avishek Chatterjee, Michael Zollhoefer, Helge Rhodin, Pascal Fua, Hans-Peter Seidel, and Christian Theobalt. Mo 2 cap 2: Real-time mobile 3d motion capture with a cap-mounted fisheye camera. IEEE transactions on visualization and computer graphics, 25(5):2093-2101, 2019. 9", + "[182] Zihui Xue and Kristen Grauman. Learning fine-grained view-invariant representations from unpaired ego-exo videos via temporal alignment. In NeurIPS, 2023. 4, 7", + "[183] Lita Yang, Robert M Radway, Yu-Hsin Chen, Tony F Wu, Huichu Liu, Elnaz Ansari, Vikas Chandra, Subhasish Mitra, and Edith Beigné. Three-dimensional stacked neural network accelerator architectures for ar/vr applications. IEEE Micro, 2022. 46", + "[184] Huangyue Yu, Minjie Cai, Yunfei Liu, and Feng Lu. What i see is what you see: Joint attention learning for first and third person video co-analysis. In ACM MM, 2019. 4", + "[185] Huangyue Yu, Minjie Cai, Yunfei Liu, and Feng Lu. First- and third-person video co-analysis by learning spatial-temporal joint attention. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2020. 4", + "[186] Xumin Yu, Yongming Rao, Wenliang Zhao, Jiwen Lu, and Jie Zhou. Group-aware contrastive regression for action quality assessment. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 7919-7928, 2021. 8" + ], + "bbox": [ + 78, + 90, + 468, + 898 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[187] Ye Yuan and Kris Kitani. 3d ego-pose estimation via imitation learning. In Proceedings of the European Conference on Computer Vision (ECCV), 2018. 9", + "[188] Ye Yuan and Kris Kitani. Ego-pose estimation and forecasting as real-time pd control. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2019. 9", + "[189] Chenlin Zhang, Jianxin Wu, and Yin Li. Actionformer: Localizing moments of actions with transformers. arXiv preprint arXiv:2202.07925, 2022. 21", + "[190] Chen-Lin Zhang, Jianxin Wu, and Yin Li. Actionformer: Localizing moments of actions with transformers. In European Conference on Computer Vision, pages 492-510, 2022. 54, 55, 57", + "[191] Qiang Zhang and Baoxin Li. Relative hidden markov models for evaluating motion skill. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2013. 8", + "[192] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, 2018. 39", + "[193] Siwei Zhang, Qianli Ma, Yan Zhang, Zhiyin Qian, Taein Kwon, Marc Pollefeys, Federica Bogo, and Siyu Tang. Egobody: Human body shape and motion of interacting people from head-mounted devices. In ECCV, 2022. 3", + "[194] Shiyi Zhang, Wenxun Dai, Sujia Wang, Xiangwei Shen, Jiwen Lu, Jie Zhou, and Yansong Tang. Logo: A long-form video dataset for group action quality assessment. In CVPR, 2023. 3", + "[195] Xiangyu Zhang, Xinyu Zhou, Mengxiao Lin, and Jian Sun. Shufflenet: An extremely efficient convolutional neural network for mobile devices. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6848-6856, 2018. 8", + "[196] He Zhao, Isma Hadji, Nikita Dvornik, Konstantinos G Derpanis, Richard P Wildes, and Allan D Jepson. P3iv: Probabilistic procedure planning from instructional videos with weak supervision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2938-2948, 2022. 4", + "[197] Weixi Zhao, Weiqiang Wang, and Yunjie Tian. Graformer: Graph-oriented transformer for 3d pose estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20438-20447, 2022. 58", + "[198] Yue Zhao and Philipp Krahenbuhl. Real-time online video detection with temporal smoothing transformers. In European Conference on Computer Vision, 2022. 49", + "[199] Yue Zhao, Ishan Misra, Philipp Krahenbuhl, and Rohit Girdhar. Learning video representations from large language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6586-6597, 2023. 26, 46, 47, 48", + "[200] Ce Zheng, Xianpeng Liu, Guo-Jun Qi, and Chen Chen. Potter: Pooling attention transformer for efficient human mesh recovery. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1611-1620, 2023. 59" + ], + "bbox": [ + 501, + 92, + 890, + 898 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "19399", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[201] Yiwu Zhong, Licheng Yu, Yang Bai, Shangwen Li, Xueling Yan, and Yin Li. Learning procedure-aware video representation from instructional videos and their narrations. arXiv preprint arXiv:2303.17839, 2023. 4", + "[202] Honglu Zhou, Roberto Martin-Martin, Mubbasir Kapadia, Silvio Savarese, and Juan Carlos Niebles. Procedure-aware pretraining for instructional video understanding. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023. 4, 8", + "[203] Honglu Zhou, Roberto Martin-Martin, Mubbasir Kapadia, Silvio Savarese, and Juan Carlos Niebles. Procedure-aware pretraining for instructional video understanding. In CVPR, pages 10727-10738, 2023. 49", + "[204] L. Zhou, N. Louis, and J. Corso. Weakly-supervised video object grounding from text by loss weighting and object interaction. In BMVC, 2018. 2, 4", + "[205] Luowei Zhou, Chenliang Xu, and Jason J Corso. Towards automatic learning of procedures from web instructional videos. In AAAI, 2018. 4, 7", + "[206] Michael Zhu and Suyog Gupta. To prune, or not to prune: exploring the efficacy of pruning for model compression. arXiv preprint arXiv:1710.01878, 2017. 8", + "[207] Dimitri Zhukov, Jean-Baptiste Alayrac, Ramazan Gokberk Cinbis, David Fouhey, Ivan Laptev, and Josef Sivic. Cross-task weakly supervised learning from instructional videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2019. 2, 4, 7", + "[208] Aneeq Zia, Yachna Sharma, Vinay Bettadapura, Eric L. Sarin, and Irfan A. Essa. Video and accelerometer-based motion analysis for automated surgical skills assessment. CoRR, abs/1702.07772, 2017. 8" + ], + "bbox": [ + 78, + 90, + 468, + 527 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "19400", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 17 + } +] \ No newline at end of file diff --git a/2024/Ego-Exo4D_ Understanding Skilled Human Activity from First- and Third-Person Perspectives/742959dc-4498-434e-af9a-b1cc7f6e5bd2_model.json b/2024/Ego-Exo4D_ Understanding Skilled Human Activity from First- and Third-Person Perspectives/742959dc-4498-434e-af9a-b1cc7f6e5bd2_model.json new file mode 100644 index 0000000000000000000000000000000000000000..d8279bc45dbf62d0a20b309836daa2fd0b15a0a0 --- /dev/null +++ b/2024/Ego-Exo4D_ Understanding Skilled Human Activity from First- and Third-Person Perspectives/742959dc-4498-434e-af9a-b1cc7f6e5bd2_model.json @@ -0,0 +1,5065 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.0, + 0.812, + 0.045 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.223, + 0.131, + 0.748, + 0.175 + ], + "angle": 0, + "content": "Ego-Exo4D: Understanding Skilled Human Activity from First- and Third-Person Perspectives" + }, + { + "type": "text", + "bbox": [ + 0.086, + 0.215, + 0.885, + 0.552 + ], + "angle": 0, + "content": "Kristen Grauman\\(^{1,2}\\), Andrew Westbury\\(^{1}\\), Lorenzo Torresani\\(^{1}\\), Kris Kitani\\(^{1,3}\\), Jitendra Malik\\(^{1,4}\\), Triantafyllos Afouras\\(^{*1}\\), Kumar Ashutosh\\(^{*1,2}\\), Vijay Baiyya\\(^{5}\\), Siddhant Bansal\\(^{6,7}\\), Bikram Boote\\(^{8}\\), Eugene Byrne\\(^{*1,9}\\), Zach Chavis\\(^{*10}\\), Joya Chen\\(^{*11}\\), Feng Cheng\\(^{1}\\), Fu-Jen Chu\\(^{1}\\), Sean Crane\\(^{9}\\), Avijit Dasgupta\\(^{7}\\), Jing Dong\\(^{5}\\), Maria Escobar\\(^{*12}\\), Cristian Forigua\\(^{*12}\\), Abrham Gebreselasie\\(^{9}\\), Sanjay Haresh\\(^{13}\\), Jing Huang\\(^{1}\\), Md Mohaiminul Islam\\(^{14}\\), Suyog Jain\\(^{1}\\), Rawal Khirodkar\\(^{9}\\), Devansh Kukreja\\(^{1}\\), Kevin J Liang\\(^{1}\\), Jia-Wei Liu\\(^{11}\\), Sagnik Majumder\\(^{1,2}\\), Yongsen Mao\\(^{13}\\), Miguel Martin\\(^{1}\\), Effrosyni Mavroudi\\(^{1}\\), Tushar Nagarajan\\(^{1}\\), Francesco Ragusa\\(^{15}\\), Santhosh Kumar Ramakrishnan\\(^{2}\\), Luigi Seminara\\(^{15}\\), Arjun Somayazulu\\(^{2}\\), Yale Song\\(^{1}\\), Shan Su\\(^{16}\\), Zihui Xue\\(^{1,2}\\), Edward Zhang\\(^{16}\\), Jinxu Zhang\\(^{16}\\), Angela Castillo\\(^{12}\\), Changan Chen\\(^{2}\\), Xinzhu Fu\\(^{11}\\), Ryosuke Furuta\\(^{17}\\), Cristina González\\(^{12}\\), Prince Gupta\\(^{5}\\), Jiabo Hu\\(^{18}\\), Yifei Huang\\(^{17}\\), Yiming Huang\\(^{16}\\), Weslie Khoo\\(^{19}\\), Anush Kumar\\(^{10}\\), Robert Kuo\\(^{18}\\), Sach Lakhavani\\(^{5}\\), Miao Liu\\(^{18}\\), Mi Luo\\(^{2}\\), Zhengyi Luo\\(^{3}\\), Brighid Meredith\\(^{18}\\), Austin Miller\\(^{18}\\), Oluwatuminiu Oguntola\\(^{14}\\), Xiaqing Pan\\(^{5}\\), Penny Peng\\(^{18}\\), Shraman Pramanick\\(^{20}\\), Merey Ramazanova\\(^{21}\\), Fiona Ryan\\(^{22}\\), Wei Shan\\(^{14}\\), Kiran Somasundaram\\(^{5}\\), Chenan Song\\(^{11}\\), Audrey Southerland\\(^{22}\\), Masatoshi Tateno\\(^{17}\\), Huiyu Wang\\(^{1}\\), Yuchen Wang\\(^{19}\\), Takuma Yagi\\(^{17}\\), Mingfei Yan\\(^{5}\\), Xitong Yang\\(^{1}\\), Zecheng Yu\\(^{17}\\), Shengxin Cindy Zha\\(^{18}\\), Chen Zhao\\(^{21}\\), Ziwei Zhao\\(^{19}\\), Zhifan Zhu\\(^{6}\\), Jeff Zhuo\\(^{14}\\), Pablo Arbeláez†\\(^{12}\\), Gedas Bertasius†\\(^{14}\\), Dima Damen†\\(^{6}\\), Jakob Engel†\\(^{5}\\), Giovanni Maria Farinella†\\(^{15}\\), Antonino Furnari†\\(^{15}\\), Bernard Ghanem†\\(^{21}\\), Judy Hoffman†\\(^{22}\\), C. V. Jawahar†\\(^{7}\\), Richard Newcombe†\\(^{5}\\), Hyun Soo Park†\\(^{10}\\), James M. Rehg†\\(^{8}\\), Yoichi Sato†\\(^{17}\\), Manolis Savva†\\(^{13}\\), Jianbo Shi†\\(^{16}\\), Mike Zheng Shou†\\(^{11}\\) and Michael Wray†\\(^{6}\\)" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.561, + 0.894, + 0.634 + ], + "angle": 0, + "content": "\\(^{1}\\)FAIR, Meta, \\(^{2}\\)University of Texas at Austin, \\(^{3}\\)Carnegie Mellon University, \\(^{4}\\)University of California, Berkeley, \\(^{5}\\)Project Aria, Meta, \\(^{6}\\)University of Bristol, \\(^{7}\\)International Institute of Information Technology, Hyderabad, \\(^{8}\\)University of Illinois, Urbana Champaign, \\(^{9}\\)Carnegie Mellon University, \\(^{10}\\)University of Minnesota, \\(^{11}\\)National University of Singapore, \\(^{12}\\)Universidad de los Andes, \\(^{13}\\)Simon Fraser University, \\(^{14}\\)University of North Carolina, Chapel Hill, \\(^{15}\\)University of Catania, \\(^{16}\\)University of Pennsylvania, \\(^{17}\\)University of Tokyo, \\(^{18}\\)Meta, \\(^{19}\\)Indiana University, \\(^{20}\\)Johns Hopkins University, \\(^{21}\\)King Abdullah University of Science and Technology, \\(^{22}\\)Georgia Tech" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.656, + 0.314, + 0.672 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.675, + 0.473, + 0.903 + ], + "angle": 0, + "content": "We present Ego-Exo4D, a diverse, large-scale multimodal multiview video dataset and benchmark challenge. Ego-Exo4D centers around simultaneously-captured ego-centric and exocentric video of skilled human activities (e.g., sports, music, dance, bike repair). 740 participants from 13 cities worldwide performed these activities in 123 different natural scene contexts, yielding long-form captures from 1 to 42 minutes each and 1,286 hours of video combined. The multimodal nature of the dataset is unprecedented: the video is accompanied by multichannel audio, eye gaze, 3D point clouds, camera poses, IMU, and multiple paired language descriptions—including a novel \"expert commentary\" done by coaches and teachers and tailored to the skilled-activity domain. To push the frontier of first-person video understanding of skilled hu" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.679, + 0.895, + 0.756 + ], + "angle": 0, + "content": "man activity, we also present a suite of benchmark tasks and their annotations, including fine-grained activity understanding, proficiency estimation, cross-view translation, and 3D hand/body pose. All resources are open sourced to fuel new research in the community." + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.769, + 0.633, + 0.785 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.795, + 0.895, + 0.901 + ], + "angle": 0, + "content": "A dancer leaps across a stage; Lionel Messi delivers a precise pass; your grandmother prepares her famous dumplings. We observe and seek human skills in a myriad of settings, from the practical (fixing a bike) to the aspirational (dancing beautifully). What would it mean for AI to understand human skills? And what would it take to get there?" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "19383" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.127, + 0.088, + 0.845, + 0.423 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.432, + 0.895, + 0.476 + ], + "angle": 0, + "content": "Figure 1. Ego-Exo4D offers egocentric video alongside multiple time-synchronized exocentric video streams for an array of skilled human activities—1,286 hours of ego and exo video in total. The data is both multiview and multimodal, and it is extensively annotated with language, 3D body and hand pose, keysteps, procedural dependencies, and proficiency ratings in support of our proposed benchmark tasks." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.501, + 0.471, + 0.637 + ], + "angle": 0, + "content": "Advances in AI understanding of human skill could facilitate many applications. In augmented reality (AR), a person wearing smart glasses could quickly pick up new skills with a virtual AI coach that provides real-time guidance. In robot learning, a robot watching people in its environment could acquire new dexterous manipulation skills with less physical experience. In social networks, new communities could form based on how people share their expertise and complementary skills in video." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.641, + 0.47, + 0.837 + ], + "angle": 0, + "content": "We contend that both the egocentric and exocentric viewpoints are critical for capturing human skill. Firstly, the two viewpoints are synergistic. The first-person (ego) perspective captures the details of close-by hand-object interactions and the camera wearer's attention, whereas the third-person (exo) perspective captures the full body pose and surrounding environment context. See Figure 1. Not coincidentally, instructional or \"how-to\" videos often alternate between a third-person view of the demonstrator and a close-up view of their near-field demonstration. For example, a chef may describe their approach and the equipment from an exo view, then cut to clips showing their hands manipulating the ingredients and tools from an ego-like view." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.841, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Secondly, not only are the ego and exo viewpoints synergistic, but there is a need to translate fluently from one to the other when acquiring skill. For example, imagine watching an expert repair a bike tire, juggle a soccer ball," + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.501, + 0.892, + 0.576 + ], + "angle": 0, + "content": "or fold an origami swan—then mapping their steps to your own body. Cognitive science tells us that even from a very young age we can observe others' behavior (exo) and map it onto our own (ego) [42, 108], and this actor-observer translation remains the foundation of visual learning." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.58, + 0.893, + 0.791 + ], + "angle": 0, + "content": "Realizing this potential, however, is not possible using today's datasets and learning paradigms. Existing datasets comprised of both ego and exo views (i.e., ego-exo) are few [76, 77, 127, 139, 145], small in scale, lack synchronization across cameras, and/or are too staged or curated to be resilient to the diversity of the real world. Thus the current literature for activity understanding primarily attends to either the ego [28, 47] or exo [48, 67, 105, 149] view, leaving the ability to move fluidly between the first- and third-person perspectives out of reach. Instructional video datasets [103, 159, 204, 207] offer a compelling window into skilled human activity, but (like the above) are limited to single-viewpoint video, whether purely exocentric or mixed with \"ego-like\" views at certain time points." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.796, + 0.895, + 0.903 + ], + "angle": 0, + "content": "We introduce Ego-Exo4D, a foundational dataset to support research on ego-exo video learning and multimodal perception. The result of a two-year effort by a consortium of 15 research institutions, Ego-Exo4D is a first-of-its-kind large-scale multimodal multiview dataset and benchmark suite. It constitutes the largest public dataset of timesynchronized first- and third- person video, captured by 740" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "19384" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.213 + ], + "angle": 0, + "content": "diverse camera wearers in 123 distinct scenes and 13 cities worldwide. For every sequence, Ego-Exo4D provides both the camera wearer's egocentric video, as well as multiple (4-5) exocentric videos from tripod places around the camera wearer. All views are time-synchronized and precisely localized in a metric, gravity-aligned frame of reference. The total collection has 1,286 hours of video and 5,035 instances, each spanning 1 to \\(42\\mathrm{min}\\) of continuous capture." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.218, + 0.47, + 0.34 + ], + "angle": 0, + "content": "Ego-Exo4D focuses on skilled single-person activities. The 740 participants perform skilled physical and/or procedural activities—dance, soccer, basketball, bouldering, music, cooking, bike repair, health care—in an unscripted manner and in natural settings (e.g., gym, soccer field, kitchens, bike shops, etc.), exhibiting a variety of skill levels from novice to expert. All video is recorded with rigorous privacy and ethics policies and formal consent of participants." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.345, + 0.471, + 0.602 + ], + "angle": 0, + "content": "Ego-Exo4D is not only multiview, it is also multimodal. Captured with the unique open-source Aria glasses [38], all ego video is accompanied by 7-channel audio, IMU, eye gaze, both RGB and two grayscale SLAM cameras, and 3D environment point clouds. Additionally, Ego-Exo4D provides multiple new video-language resources, all time indexed: first-person narrations by the camera wearers describing their own actions; third-person play-by-play descriptions of every camera wearer action; and third-person spoken expert commentary critiquing their performance. The latter is particularly novel: performed by domain-specific experienced coaches and teachers, it focuses on how an activity is executed rather than merely what is being done, surfacing subtleties in skilled execution not perceivable by the untrained eye. To our knowledge, there is no prior video resource with such extensive and high quality multimodal data." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.608, + 0.47, + 0.82 + ], + "angle": 0, + "content": "Alongside this data, we introduce benchmarks for foundational tasks for ego-exo video. We propose four families of tasks: 1) ego-exo relation, for relating the actions of a teacher (exo) to a learner (ego) by estimating semantic correspondences and translating viewpoints; 2) ego(-exo) recognition, for recognizing fine-grained keysteps and task structure; 3) ego(-exo) proficiency estimation, for inferring how well a person is executing a skill; and 4) ego pose, for recovering skilled 3D body and hand movements from ego-video. We provide annotations for each task—the result of more than 200,000 hours of annotator effort. To kick-start work in these new challenges, we also develop baseline models and report their results (Appendices). We are hosting the first public benchmark challenges in 2024." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.826, + 0.47, + 0.903 + ], + "angle": 0, + "content": "In summary, Ego-Exo4D is the community's first diverse, large-scale multimodal multiview video resource. We have open sourced all the data, annotations, camera rig protocol, and benchmarks. With this release, we aim to fuel new research in ego-exo, multimodal activity, and beyond." + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.09, + 0.637, + 0.105 + ], + "angle": 0, + "content": "2. Related work" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.116, + 0.892, + 0.177 + ], + "angle": 0, + "content": "Next we review prior work in datasets, human skill, and cross-view analysis. Section 5 will discuss related work for each benchmark task. Table 2 in Appendix 9 summarizes Ego-Exo4D's properties vs. existing datasets." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.186, + 0.893, + 0.322 + ], + "angle": 0, + "content": "Egocentric datasets There has been a surge of interest in egocentric video understanding, facilitated by recent egovideo datasets showing unscripted daily-life activity as in Ego4D [47], EPIC-Kitchens [27, 28, 163], UT Ego [78], ADL [119], and KrishnaCam [147], or procedural activities as in EGTea [81], AssistQ [172], Meccano [126], CMUMMAC [77], and EgoProcel [10]. Unlike any of the above, Ego-Exo4D focuses on multimodal ego and exo capture, and it is focused on the domain of skilled activities." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.332, + 0.893, + 0.619 + ], + "angle": 0, + "content": "Multiview and ego-exo datasets Most existing multiview datasets focus on static scenes [20, 128, 151, 175, 176] and objects [133, 173], with limited (exo only) multiview human activity [26, 169]. CMU-MMAC [77] and CharadesEgo [145] are early efforts to capture both ego and exo video. CMU-MMAC [77] features 43 participants in mocap suits who cook 5 recipes in a lab kitchen. In CharadesEgo [145], 71 Mechanical Turkers record 34 hours of scripted scenarios (e.g., \"type on laptop, then pick up a pillow\") from the ego and exo perspectives sequentially, yielding unsynchronized videos with non-exact activity matches. More recent ego-exo efforts focus on specific activities in one or two environments. Assembly101 [139] and H2O [76] provide time-synced ego and exo video at a lab tabletop where people assemble toy cars or manipulate handheld objects, with 53 and 4 participants, and 513 and 5 hours of footage, respectively. Homage [127] provides 30 hours of ego-exo video from 27 participants in 2 homes doing household activities like laundry." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.62, + 0.892, + 0.845 + ], + "angle": 0, + "content": "Compared to any of the prior efforts, Ego-Exo4D offers an order of magnitude more participants, diverse locations, and hours of footage (740 participants, 123 unique scenes, 13 cities, 1,286 hours). Importantly, our focus on skilled tasks takes the participants out of the lab or home and into settings like soccer fields, dance studios, rock climbing walls, and bike repair shops. Such activities also yield a wide variety of full body poses and movements within the scene, beyond using objects at a tabletop. This variety means Ego-Exo4D augments existing 3D human body pose datasets [49, 66, 68, 80, 193]. Finally, compared to any prior ego-exo resource, Ego-Exo4D's suite of modalities and benchmark tasks are novel and will expand the research directions the community can take for egocentric and/or exocentric video understanding." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.856, + 0.893, + 0.901 + ], + "angle": 0, + "content": "Human skill and video learning Analyzing skill and action quality has received limited attention [12, 34, 35, 113, 120, 194]. Research in instructional or \"how-to\" videos is" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19385" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.09, + 0.098, + 0.132, + 0.131 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.091, + 0.134, + 0.132, + 0.145 + ], + "angle": 0, + "content": "Cooking" + }, + { + "type": "image_caption", + "bbox": [ + 0.147, + 0.093, + 0.192, + 0.102 + ], + "angle": 0, + "content": "x678" + }, + { + "type": "image_caption", + "bbox": [ + 0.148, + 0.104, + 0.19, + 0.113 + ], + "angle": 0, + "content": "x173" + }, + { + "type": "image_caption", + "bbox": [ + 0.148, + 0.114, + 0.19, + 0.123 + ], + "angle": 0, + "content": "8" + }, + { + "type": "image_caption", + "bbox": [ + 0.148, + 0.124, + 0.19, + 0.133 + ], + "angle": 0, + "content": "Mx65" + }, + { + "type": "image_caption", + "bbox": [ + 0.148, + 0.134, + 0.197, + 0.144 + ], + "angle": 0, + "content": "\\(\\mathbb{O}\\times 564\\mathrm{h}\\)" + }, + { + "type": "image", + "bbox": [ + 0.203, + 0.09, + 0.353, + 0.146 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.24, + 0.147, + 0.306, + 0.153 + ], + "angle": 0, + "content": "Making Chai in Hyderabad" + }, + { + "type": "image", + "bbox": [ + 0.36, + 0.098, + 0.421, + 0.133 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.377, + 0.135, + 0.407, + 0.144 + ], + "angle": 0, + "content": "Music" + }, + { + "type": "image_caption", + "bbox": [ + 0.43, + 0.093, + 0.474, + 0.102 + ], + "angle": 0, + "content": "x276" + }, + { + "type": "image_caption", + "bbox": [ + 0.431, + 0.104, + 0.469, + 0.113 + ], + "angle": 0, + "content": "x59" + }, + { + "type": "image_caption", + "bbox": [ + 0.431, + 0.114, + 0.468, + 0.123 + ], + "angle": 0, + "content": "#" + }, + { + "type": "image_caption", + "bbox": [ + 0.431, + 0.124, + 0.468, + 0.133 + ], + "angle": 0, + "content": "#" + }, + { + "type": "image_caption", + "bbox": [ + 0.431, + 0.134, + 0.48, + 0.143 + ], + "angle": 0, + "content": "L x 18Oh" + }, + { + "type": "image", + "bbox": [ + 0.483, + 0.09, + 0.631, + 0.146 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.515, + 0.147, + 0.587, + 0.153 + ], + "angle": 0, + "content": "Piano Playing in Bloomington" + }, + { + "type": "image", + "bbox": [ + 0.644, + 0.101, + 0.683, + 0.13 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.649, + 0.134, + 0.681, + 0.143 + ], + "angle": 0, + "content": "Soccer" + }, + { + "type": "image_caption", + "bbox": [ + 0.695, + 0.094, + 0.736, + 0.103 + ], + "angle": 0, + "content": "x282" + }, + { + "type": "image_caption", + "bbox": [ + 0.695, + 0.104, + 0.731, + 0.113 + ], + "angle": 0, + "content": "x78" + }, + { + "type": "image_caption", + "bbox": [ + 0.695, + 0.114, + 0.729, + 0.123 + ], + "angle": 0, + "content": "#" + }, + { + "type": "image_caption", + "bbox": [ + 0.695, + 0.124, + 0.729, + 0.132 + ], + "angle": 0, + "content": "X14" + }, + { + "type": "image_caption", + "bbox": [ + 0.695, + 0.134, + 0.736, + 0.143 + ], + "angle": 0, + "content": "\\(\\bullet \\times 66h\\)" + }, + { + "type": "image_caption", + "bbox": [ + 0.787, + 0.091, + 0.816, + 0.1 + ], + "angle": 0, + "content": "\\( \\therefore m = \\frac{3}{11} \\) ;" + }, + { + "type": "image", + "bbox": [ + 0.817, + 0.091, + 0.891, + 0.146 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.787, + 0.148, + 0.837, + 0.154 + ], + "angle": 0, + "content": "Soccer in Pittsburgh" + }, + { + "type": "image", + "bbox": [ + 0.091, + 0.155, + 0.133, + 0.191 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.097, + 0.196, + 0.129, + 0.203 + ], + "angle": 0, + "content": "Health" + }, + { + "type": "image_caption", + "bbox": [ + 0.148, + 0.155, + 0.192, + 0.163 + ], + "angle": 0, + "content": "x397" + }, + { + "type": "image_caption", + "bbox": [ + 0.149, + 0.165, + 0.189, + 0.174 + ], + "angle": 0, + "content": "x122" + }, + { + "type": "image_caption", + "bbox": [ + 0.149, + 0.176, + 0.185, + 0.186 + ], + "angle": 0, + "content": "9x24" + }, + { + "type": "image_caption", + "bbox": [ + 0.149, + 0.188, + 0.194, + 0.198 + ], + "angle": 0, + "content": "中" + }, + { + "type": "image_caption", + "bbox": [ + 0.149, + 0.2, + 0.194, + 0.207 + ], + "angle": 0, + "content": "X141" + }, + { + "type": "list", + "bbox": [ + 0.148, + 0.155, + 0.194, + 0.207 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.085, + 0.221, + 0.142, + 0.255 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.086, + 0.259, + 0.141, + 0.268 + ], + "angle": 0, + "content": "Bike Repair" + }, + { + "type": "image_caption", + "bbox": [ + 0.149, + 0.218, + 0.191, + 0.226 + ], + "angle": 0, + "content": "4x363" + }, + { + "type": "image_caption", + "bbox": [ + 0.149, + 0.228, + 0.185, + 0.237 + ], + "angle": 0, + "content": "x32" + }, + { + "type": "image_caption", + "bbox": [ + 0.149, + 0.239, + 0.18, + 0.249 + ], + "angle": 0, + "content": "#" + }, + { + "type": "image_caption", + "bbox": [ + 0.149, + 0.251, + 0.18, + 0.259 + ], + "angle": 0, + "content": "图" + }, + { + "type": "list", + "bbox": [ + 0.149, + 0.228, + 0.185, + 0.259 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.149, + 0.261, + 0.191, + 0.268 + ], + "angle": 0, + "content": "X 82n" + }, + { + "type": "image", + "bbox": [ + 0.203, + 0.154, + 0.351, + 0.21 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.2, + 0.213, + 0.351, + 0.272 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.232, + 0.272, + 0.313, + 0.278 + ], + "angle": 0, + "content": "Removing a Wheel in Menlo Park" + }, + { + "type": "image", + "bbox": [ + 0.375, + 0.155, + 0.416, + 0.192 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.37, + 0.195, + 0.42, + 0.203 + ], + "angle": 0, + "content": "Basketball" + }, + { + "type": "image", + "bbox": [ + 0.375, + 0.21, + 0.416, + 0.258 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.36, + 0.26, + 0.429, + 0.268 + ], + "angle": 0, + "content": "Rock Climbing" + }, + { + "type": "image_caption", + "bbox": [ + 0.431, + 0.155, + 0.472, + 0.163 + ], + "angle": 0, + "content": "x910" + }, + { + "type": "image_caption", + "bbox": [ + 0.431, + 0.165, + 0.472, + 0.174 + ], + "angle": 0, + "content": "x113" + }, + { + "type": "image_caption", + "bbox": [ + 0.431, + 0.176, + 0.469, + 0.186 + ], + "angle": 0, + "content": "9x5" + }, + { + "type": "image_caption", + "bbox": [ + 0.431, + 0.188, + 0.474, + 0.197 + ], + "angle": 0, + "content": "#" + }, + { + "type": "image_caption", + "bbox": [ + 0.431, + 0.198, + 0.474, + 0.205 + ], + "angle": 0, + "content": "5x78n" + }, + { + "type": "list", + "bbox": [ + 0.431, + 0.155, + 0.474, + 0.205 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.431, + 0.21, + 0.479, + 0.272 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.49, + 0.154, + 0.63, + 0.21 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.482, + 0.211, + 0.63, + 0.272 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.519, + 0.272, + 0.584, + 0.278 + ], + "angle": 0, + "content": "Bouldering in Minneapolis" + }, + { + "type": "image", + "bbox": [ + 0.652, + 0.155, + 0.684, + 0.192 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.652, + 0.195, + 0.683, + 0.204 + ], + "angle": 0, + "content": "Dance" + }, + { + "type": "image_caption", + "bbox": [ + 0.706, + 0.155, + 0.736, + 0.163 + ], + "angle": 0, + "content": "4x728" + }, + { + "type": "image_caption", + "bbox": [ + 0.706, + 0.165, + 0.735, + 0.174 + ], + "angle": 0, + "content": "x93" + }, + { + "type": "image_caption", + "bbox": [ + 0.696, + 0.176, + 0.726, + 0.186 + ], + "angle": 0, + "content": "9x7" + }, + { + "type": "image_caption", + "bbox": [ + 0.696, + 0.188, + 0.741, + 0.197 + ], + "angle": 0, + "content": "#" + }, + { + "type": "image_caption", + "bbox": [ + 0.696, + 0.198, + 0.741, + 0.205 + ], + "angle": 0, + "content": "x106n" + }, + { + "type": "image_caption", + "bbox": [ + 0.787, + 0.154, + 0.837, + 0.162 + ], + "angle": 0, + "content": "\\( \\therefore m = \\frac{3}{11} \\) ;" + }, + { + "type": "image", + "bbox": [ + 0.745, + 0.163, + 0.89, + 0.21 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.793, + 0.211, + 0.831, + 0.216 + ], + "angle": 0, + "content": "Salsa in Bogota" + }, + { + "type": "image_caption", + "bbox": [ + 0.662, + 0.219, + 0.688, + 0.228 + ], + "angle": 0, + "content": "Total" + }, + { + "type": "image_caption", + "bbox": [ + 0.705, + 0.219, + 0.793, + 0.228 + ], + "angle": 0, + "content": "4x5035 (#takes)" + }, + { + "type": "image_caption", + "bbox": [ + 0.705, + 0.231, + 0.823, + 0.24 + ], + "angle": 0, + "content": "x740 (#participants)" + }, + { + "type": "image_caption", + "bbox": [ + 0.705, + 0.243, + 0.781, + 0.254 + ], + "angle": 0, + "content": "9. x 123 (# sites)" + }, + { + "type": "image_caption", + "bbox": [ + 0.705, + 0.257, + 0.84, + 0.268 + ], + "angle": 0, + "content": "x 1286h (#ego+exo hours)" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.286, + 0.894, + 0.329 + ], + "angle": 0, + "content": "Figure 2. Ego-Exo4D captures skilled activity from 43 tasks and 689 keysteps in 8 domains, in a wide variety of 123 scenes in 13 cities in Japan, Colombia, Canada, India, Singapore, and 7 US states. Each domain is captured at multiple sites—from 2 to 64 unique locations. In total the dataset offers 1,286 hours of ego+exo video comprised of 5,035 takes from 740 camera wearers. An average take is 2.6 minutes." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.355, + 0.47, + 0.507 + ], + "angle": 0, + "content": "facilitated by (largely exo) datasets like HowTo100M [103] and others [11, 159, 204, 207]. Challenges include grounding keysteps [10, 36, 37, 89, 103, 104, 178, 207], procedural planning [15, 17, 22, 71, 143, 167, 196, 201], learning task structure [4, 9, 37, 107, 202, 205], and leveraging noisy narrations [89, 103, 104]. A portion of Ego-Exo4D is procedural activities, but unlike the above, it offers simultaneous ego-exo capture. The scale and diversity of our data—including its three forms of language descriptions—widen the avenues for skilled activity understanding research." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.521, + 0.47, + 0.718 + ], + "angle": 0, + "content": "Ego-exo cross-view modeling There is limited prior work on ego-exo cross-view modeling, arguably due to a lack of high-quality synchronized real-world data. Prior work explores matching people between videos [5, 6, 40, 170, 179] and learning view-invariant [7, 141, 144, 182, 184, 185] or ego features [82]. Beyond the specific case of ego-exo, cross-view methods are explored for translation [130, 131, 134, 157], novel view synthesis [19, 90, 135, 137, 164, 168, 171], and aerial to ground matching [86, 132]. Ego-Exo4D provides a testbed of unprecedented size and variety for cross-view modeling. In addition, our ego-exo relation tasks (cf. Section 5) surface new challenges in novel-view synthesis with widely varying viewpoints." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.733, + 0.263, + 0.75 + ], + "angle": 0, + "content": "3. Ego-Exo4D dataset" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.759, + 0.469, + 0.82 + ], + "angle": 0, + "content": "Next we introduce the dataset and its scope. Notably, the video capture was a distributed but coordinated effort performed by 12 research labs. We present the common framework, and reserve site-specific details for Appendix 10." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.832, + 0.265, + 0.849 + ], + "angle": 0, + "content": "3.1. Ego-exo camera rig" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.47, + 0.901 + ], + "angle": 0, + "content": "Our goal is to capture simultaneous ego and exo video, together with multiple egocentric sensing modalities. One of our contributions is to create and share a low-cost (less than" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.355, + 0.891, + 0.386 + ], + "angle": 0, + "content": "\\(3,000), lightweight ego-exo rig with a user-friendly calibration and time sync procedure." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.387, + 0.892, + 0.553 + ], + "angle": 0, + "content": "Our camera configuration features Aria glasses [38] for ego capture, leveraging their rich array of sensors, including an 8 MP RGB camera, two SLAM cameras, IMU, 7 microphones, and eye tracking (see Appendix 7). The ego camera is calibrated and time-synchronized with four to five (stationary) GoPros placed on tripods as the exo capture devices, allowing 3D reconstruction of the environment point clouds and the participant's body pose. The number and placement of the exocentric cameras is determined per scenario in order to allow maximal coverage of useful viewpoints without obstructing the participants' activity." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.554, + 0.892, + 0.629 + ], + "angle": 0, + "content": "Our time sync and calibration design relies on a QR-code procedure to auto-sync the cameras and auto-separate the individual \"takes\", meaning instances of an activity. We can do continuous recordings of up to \\(\\sim 60\\) minutes, based on the Aria battery life. See Appendix 8 for more details." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.642, + 0.746, + 0.656 + ], + "angle": 0, + "content": "3.2. Domains and environments" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.665, + 0.892, + 0.786 + ], + "angle": 0, + "content": "Ego-Exo4D focuses on skilled human activity. This is in contrast to existing ego-only efforts like Ego4D [47], which has a broad span of daily-life activities. We intentionally select the domains based on a few criteria: Will it illustrate skill and a variety of expertise? Is there visual variety to be expected across different instances? Will the ego and exo views offer complementary information? Will it present new challenges unaddressed by current datasets?" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.788, + 0.892, + 0.863 + ], + "angle": 0, + "content": "Intersecting these criteria, we arrived at two broad categories1 of skilled activity: physical and procedural, together comprising eight total domains. The physical domains are soccer, basketball, dance, bouldering, and music. They emphasize body pose and movements as well as" + }, + { + "type": "page_footnote", + "bbox": [ + 0.5, + 0.876, + 0.892, + 0.901 + ], + "angle": 0, + "content": "1Note that in general physical and procedural are not mutually exclusive labels. An activity can both require physical skill and procedural steps." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19386" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.468, + 0.197 + ], + "angle": 0, + "content": "interaction with objects (e.g., a ball, musical instrument). The procedural domains are cooking, bike repair, and health care. They require performing a sequence of steps to reach a goal state (e.g., a completed recipe, a repaired bike) and generally entail intricate hand-object manipulations with a variety of objects (e.g., bike repair tools; cooking utensils, appliances, and ingredients)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.198, + 0.468, + 0.273 + ], + "angle": 0, + "content": "In total, we have 43 activities derived from the eight domains (see Appendix 9). For example, cooking is comprised of 14 recipes; soccer is comprised of 3 drills. The length of a take ranges from 8 sec to \\(42\\mathrm{min}\\), with procedural activities like cooking having the longest sustained captures." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.274, + 0.468, + 0.394 + ], + "angle": 0, + "content": "To achieve visual diversity in the data, multiple labs across our team (typically 3-5) captured each Ego-Exo4D domain. The data is collected in authentic settings—such as real-world bike shops, soccer pitches, or bouldering gyms—as opposed to lab environments. For example, we have videos of chefs in New York City, Vancouver, Philadelphia, Bogota, and others; soccer players in Tokyo, Chapel Hill, Hyderabad, Singapore, and Pittsburgh. See Figure 2." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.405, + 0.392, + 0.421 + ], + "angle": 0, + "content": "3.3. Participants: expertise and diversity" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.429, + 0.468, + 0.579 + ], + "angle": 0, + "content": "We recruited 740 total participants from the local communities of 12 labs. All scenarios feature real-world experts, where the camera-wearer participant has specific credentials, training, or expertise in the skill being demonstrated. For example, among the Ego-Exo4D camera wearers are professional and college athletes; jazz, salsa, and Chinese folk dancers and instructors; competitive boulders; professional chefs who work in industrial-scale kitchens; bike technicians who service dozens of bikes per day. Many of them have (individually) over 10 years of experience." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.58, + 0.468, + 0.731 + ], + "angle": 0, + "content": "Experts are prioritized given they are likely to conduct activities without mistakes or distractions, providing a strong ground truth for how to approach a given task. However, we also include capture from people with varying skill levels, as well—essential for our proposed skill proficiency estimation task (Section 5). Notably, Ego-Exo4D represents human intelligence in a new way by capturing domain-specific expertise—both in the video as well as the accompanying expert commentary (see Section 4)—portraying the evolution of a skill from beginners to experts." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.731, + 0.468, + 0.806 + ], + "angle": 0, + "content": "According to the participant surveys (Appendix 11), the camera wearers range in age from 18 to 74 years old, with \\(37\\%\\) self-identifying as female \\(60\\%\\) male and \\(3\\%\\) as nonbinary or preferring not to say. In total, the participants self report more than 24 different ethnicities." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.817, + 0.254, + 0.833 + ], + "angle": 0, + "content": "3.4. Privacy and ethics" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.468, + 0.9 + ], + "angle": 0, + "content": "Ego-Exo4D was collected following rigorous privacy and ethics standards. This included undergoing formal independent review processes at each institution to establish the standards for collection, management, and informed con" + }, + { + "type": "image", + "bbox": [ + 0.504, + 0.09, + 0.604, + 0.199 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.607, + 0.094, + 0.7, + 0.102 + ], + "angle": 0, + "content": "Expert commentary" + }, + { + "type": "text", + "bbox": [ + 0.607, + 0.103, + 0.746, + 0.126 + ], + "angle": 0, + "content": "\"The dancer's hand is rotated inwardly a bit. Her palm should be facing to the ground, ...\"" + }, + { + "type": "image_caption", + "bbox": [ + 0.608, + 0.137, + 0.734, + 0.145 + ], + "angle": 0, + "content": "Atomic action descriptions" + }, + { + "type": "text", + "bbox": [ + 0.608, + 0.146, + 0.729, + 0.161 + ], + "angle": 0, + "content": "\"C steps forward while shimmering her shoulders.\"" + }, + { + "type": "image_caption", + "bbox": [ + 0.608, + 0.172, + 0.682, + 0.179 + ], + "angle": 0, + "content": "Narrate and act" + }, + { + "type": "text", + "bbox": [ + 0.608, + 0.18, + 0.744, + 0.196 + ], + "angle": 0, + "content": "\"Here we go, left forward, right back for one phrase 1234...\"" + }, + { + "type": "image", + "bbox": [ + 0.75, + 0.09, + 0.891, + 0.199 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.211, + 0.892, + 0.239 + ], + "angle": 0, + "content": "Figure 3. Ego-Exo4D offers 3 paired language corpora. Word cloud is from expert commentary which critiques the performance." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.257, + 0.892, + 0.393 + ], + "angle": 0, + "content": "sent. Similarly, all Ego-Exo4D data collection adhered to the Project Aria Research Community Guidelines for responsible research. Since the scenarios allow for closed environments (e.g., no passerbys) nearly all video is available without de-identification. For information about each individual partners' protocols and restrictions, please see Appendix 10. Ego-Exo4D data is gated behind a license system, which defines permitted uses, restrictions, and consequences for non-compliance." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.407, + 0.776, + 0.424 + ], + "angle": 0, + "content": "4. Natural language descriptions" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.432, + 0.892, + 0.569 + ], + "angle": 0, + "content": "Ego-Exo4D also offers three kinds of paired natural language datasets, each time-indexed alongside the video. See Figure 3. These language annotations are not steered towards any single benchmark, but rather are a general resource that will support browsing and mining the dataset—as well as challenges in video-language learning like grounding actions and objects, self-supervised representation learning, video-conditioned language models, and skill assessment. See Appendix 12." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.57, + 0.892, + 0.901 + ], + "angle": 0, + "content": "The first language dataset is spoken expert commentary. The goal is to reveal nuances of the skill that are not always visible to non-experts. We recruited 52 experts (distinct from the participants) to critique the recorded videos, call out strengths and weaknesses, explain how the specific behavior of the participant (e.g., hand/body pose, use of objects) affects the performance, and provide spatial markings to support their commentary. The experts are not only well-credentialed in their areas of expertise, but also have coaching or teaching experience, which facilitates clear communication. They watch the video and pause every time they have a comment, typically 7 times per minute of video. Each piece of commentary is unbounded in length, and averages 4 sentences. We provide both the transcribed speech and the raw audio (interesting for its inflection and nonword utterances), as well as the experts' spatial drawings and numeric ratings of each participant's skill. Videos have expert commentary by 2-5 distinct experts, offering a variety of perspectives for the same content. In total, we have 117,812 pieces of time-stamped, video-aligned commentary. These commentaries are quite novel: they focus on how the activity is executed rather than what it entails," + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.956 + ], + "angle": 0, + "content": "19387" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.081, + 0.09, + 0.522, + 0.251 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.53, + 0.091, + 0.89, + 0.251 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.17, + 0.259, + 0.798, + 0.274 + ], + "angle": 0, + "content": "Figure 4. The ego-exo relation family consists of the tasks of correspondence (left) and translation (right)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.296, + 0.469, + 0.356 + ], + "angle": 0, + "content": "capturing subtle differences in skilled execution. We believe this can unlock new fundamental problems (e.g., proficiency estimation below) and disruptive future applications (e.g., AI coaching)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.358, + 0.469, + 0.493 + ], + "angle": 0, + "content": "The second language dataset consists of narrate-and-act descriptions provided by the participants themselves. They are in the style of a tutorial or how-to video, where the participant explains what they are doing and why. Unlike the third-party expert commentary above, these are first-person reflections on the activity given by the people doing them. These narrations are available for about \\(10\\%\\) of all takes in the dataset, since we wanted participants to execute the tasks without pausing for the bulk of the recordings." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.495, + 0.47, + 0.691 + ], + "angle": 0, + "content": "The third language dataset consists of atomic action descriptions. Whereas the commentary and narrate-and-act language reveals spoken opinions and reasons for the actions (the \"why and how\"), this stream of text is specifically about the \"what\". Inspired by Ego4D's narrations [47], these are short statements written by third-party (nondomain expert) annotators, timestamped for every atomic action performed by the participant for all videos in the dataset, for a total of 432K sentences. This data is valuable for mining for taxonomies of objects and actions in the data, indexing the videos with keywords for exploring the dataset, and for future research in video-language learning, as has been quite successful for the Ego4D narrations [8, 85, 123]." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.708, + 0.345, + 0.725 + ], + "angle": 0, + "content": "5. Ego-Exo4D benchmark tasks" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.735, + 0.47, + 0.901 + ], + "angle": 0, + "content": "Our second major contribution is to define the core research challenges in the domain of egocentric perception of skilled activity, particularly when ego-exo data is available for training (if not testing). To that end, we devise a suite of foundational benchmark tasks organized into four task families: relation (Sec. 5.1), recognition (Sec. 5.2), proficiency (Sec. 5.3), and ego-pose (Sec. 5.4). For each task, we provide high quality annotations and baselines that provide a starting point from which the research community can build. We will run the first formal Ego-Exo4D challenges in 2024. Due to space limits, we briefly overview" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.296, + 0.892, + 0.372 + ], + "angle": 0, + "content": "each task; see the referenced Appendices for all details including baseline models and results. There are two publicly released versions of Ego-Exo4D annotations: v1 is used to train/test baselines in this paper; the larger v2 will be used for future challenge leaderboards (see Table 7 in Appendix)." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.383, + 0.663, + 0.399 + ], + "angle": 0, + "content": "5.1. Ego-exo relation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.408, + 0.892, + 0.468 + ], + "angle": 0, + "content": "Our ego-exo relation tasks deal with relating the video content across the extreme ego-exo viewpoint changes. They take the form of object-level matching (correspondence) and synthesis of one view from the other (translation)." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.481, + 0.724, + 0.496 + ], + "angle": 0, + "content": "5.1.1 Ego-exo correspondence" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.507, + 0.892, + 0.643 + ], + "angle": 0, + "content": "Motivation. Establishing object-level correspondences between ego and exo viewpoints would allow AI assistants to provide visual instructions by matching third-person observations of objects from instructional videos to those in the user's first-person view. Compared to the general correspondence problem, our setting requires tackling a number of challenges: extreme viewpoint differences, high degrees of object occlusion, and many small objects (e.g., cooking utensils and bike repair tools)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.644, + 0.892, + 0.78 + ], + "angle": 0, + "content": "Task definition. Given a pair of synchronized ego-exo videos and a sequence of query masks of an object of interest in one of the videos, the task is to predict the corresponding mask for the same object in each synchronized frame of the other view, if it is visible. See Figure 4, left. The task can be posed with query objects in either the ego or exo video, with both directions presenting interesting challenges (e.g., high degree of occlusion in ego views, and small object size in exo views). See Appendix 13.A.1." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.781, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Related work. Related tasks are image-level sparse correspondence given query points (instead of object masks) [65] and image-level object co-segmentation [166] for jointly segmenting semantically similar objects. Our task goes beyond static object correspondence, since the interplay between human pose and object state changes during manipulation necessitate using temporal context and tracking as the query object can be highly occluded or blurry [158]." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19388" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.08, + 0.089, + 0.346, + 0.266 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.352, + 0.089, + 0.619, + 0.267 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.624, + 0.089, + 0.892, + 0.266 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.278, + 0.893, + 0.307 + ], + "angle": 0, + "content": "Figure 5. Ego-exo keystep recognition. This family of tasks consists of fine-grained recognition (left, Section 5.2.1), procedure understanding (center, Section 5.2.3), and energy-efficient multimodal recognition (right, Section 5.2.2)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.333, + 0.269, + 0.348 + ], + "angle": 0, + "content": "5.1.2 Ego-exo translation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.357, + 0.47, + 0.569 + ], + "angle": 0, + "content": "Motivation. Our translation task entails synthesizing a target ego clip from a given exo clip. We believe this problem will drive novel research for combining recognition and object synthesis. For example, in Figure 4 (right), the approach must make effective use of the hand's object-specific shape and appearance priors in order to synthesize the ego view of the fingertips—which are not visible in the exo clip. Furthermore, this task will stimulate advances in visual odometry, as the method must be able to infer the ego camera pose from the third-person clip. Ego-exo translation also holds strong application potential, as it may unlock the ability to generate first-person renderings of videos that were originally captured from a third-person perspective, e.g., benefitting robot perception or AR coaching." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.569, + 0.471, + 0.825 + ], + "angle": 0, + "content": "Task definition. We decompose ego-exo translation into two separate tasks: ego track prediction and ego clip generation (Figure 4, right). Ego track prediction estimates the segmentation mask of an object in the unobserved ego frames given the object masks in the observed exo clip. Ego clip generation must generate the image values (i.e., RGB) within the given ground-truth ego mask by making use of the exo clip and the object masks in those frames. This decomposition effectively splits the problem into two tasks: 1) predicting the location and shape of the object in the ego clip, and 2) synthesizing its appearance given the ground-truth position. For each, we consider a variant where the pose of the ego camera with respect to the exo camera is available to use at inference time. This simplifies the problem but reduces the applicability of the method, since this information is typically not available for arbitrary third-person videos. See Appendix 13.A.2." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.826, + 0.47, + 0.901 + ], + "angle": 0, + "content": "Related work. Ego-exo translation relates to cross-view image synthesis [96, 130, 157]. Within this genre, the problem of exo-to-ego generation was recently introduced for both images [93] and video [94, 97], and approached using GANs or diffusion conditioned on the input view. Our work" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.333, + 0.892, + 0.379 + ], + "angle": 0, + "content": "not only formalizes this task with ample data, but its formulation also draws attention to the need for a semantic basis to new view synthesis across extreme view changes." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.389, + 0.753, + 0.405 + ], + "angle": 0, + "content": "5.2. Ego-exo keystep recognition" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.412, + 0.892, + 0.443 + ], + "angle": 0, + "content": "This family of tasks centers around recognizing the keysteps of a procedural activity and modeling their dependencies." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.453, + 0.783, + 0.469 + ], + "angle": 0, + "content": "5.2.1 Fine-grained keystep recognition" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.478, + 0.893, + 0.613 + ], + "angle": 0, + "content": "Motivation. Recognizing the step a camera wearer is performing is non-trivial: keysteps in the same activity may look similar (folding vs. smoothing the bedsheet) and may involve hand-object interactions with heavy occlusions and head motion. Models with access to multiple views during training can leverage their complementarity to account for the deficiencies of each one, by learning viewpoint invariant representations or distilling multi-view signals into a single model (e.g., human hands from ego; body pose from exo)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.614, + 0.893, + 0.81 + ], + "angle": 0, + "content": "Task definition. We study ego-exo for video recognition. During training, models have access to paired ego-exo data—time-synchronized captures of the same activity from multiple known viewpoints. Each training instance has one ego view, \\(N\\) exo views, and a corresponding keystep label (e.g., \"flip the omelette\"). At test time, given only a trimmed egocentric video clip, the model must identify the keystep performed from a taxonomy of 689 keysteps across 17 procedural activities. See Figure 5, left. Importantly, all extra supervision (time-alignment, camera poses etc.) is only available at training time; inference is standard keystep recognition, but with models that benefit from cross-viewpoint training. See Appendix 13.B.1." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.811, + 0.894, + 0.901 + ], + "angle": 0, + "content": "Related work. Keystep recognition has been studied in first-person [10, 126, 145, 148] or third-person [9, 100, 159, 205, 207] videos; however, limited work considers both views together. Prior work considers cross-view learning with unpaired videos [7, 82, 182] and view-invariant feature learning on paired videos [144]. In contrast, we ex" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19389" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.47, + 0.123 + ], + "angle": 0, + "content": "plore keystep recognition in large-scale, procedural activities with fully synchronized training videos." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.132, + 0.467, + 0.149 + ], + "angle": 0, + "content": "5.2.2 Energy-efficient multimodal keystep recognition" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.156, + 0.469, + 0.276 + ], + "angle": 0, + "content": "Motivation. Current activity detection models assume access to densely sampled clips from the full video and ample computational resources to process them. These assumptions are incompatible with real-world devices (e.g., mobile phones, AR glasses) where the camera is not always on and the compute budget is limited by battery life. This task focuses on building energy-efficient video models to pave the way for feasibility on real-world hardware." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.277, + 0.469, + 0.444 + ], + "angle": 0, + "content": "Task definition. We formulate the problem as an online action detection task, with a given energy budget. See Figure 5, right. Given a stream of audio, IMU, and RGB video data, a model must identify the keystep being performed at each frame, as well as decide which sensor(s) to use for subsequent time-steps. This task will inspire models that are strategic about which modality to deploy when. Energy consumption is the sum of sensor energy (operating the camera/audio/IMU sensors), model inference costs, and memory transfer costs, and must be within \\(20\\mathrm{mW}\\) to reflect real-world device power constraints. See Appendix 13.B.2." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.444, + 0.469, + 0.595 + ], + "angle": 0, + "content": "Related work. Prior work on efficient models considers light-weight architectures [41, 56, 101, 155, 165, 195], efficient input processing [43, 44, 73, 102, 156], or inference optimizations [39, 58, 121, 174, 206]. In all cases, they optimize computation (FLOPs), parameter count, or prediction throughput (FPS), which in isolation are insufficient to characterize running on real-world devices. To address this, we propose the first benchmark for energy-efficient video recognition that is tied to real-world, on-device constraints, and measures total power consumed." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.606, + 0.31, + 0.621 + ], + "angle": 0, + "content": "5.2.3 Procedure understanding" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.629, + 0.47, + 0.689 + ], + "angle": 0, + "content": "Motivation. Automatically understanding the structure of a procedure from video (inferring keystep ordering, preconditions, etc.) would allow assisting AR users in a task or informing robots that learn from human demonstrations." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.689, + 0.469, + 0.87 + ], + "angle": 0, + "content": "Task definition. In our procedure understanding task, given a video segment \\( s_t \\) and its previous video segment history, models have to 1) determine previous keysteps (to be performed before \\( s_t \\)); infer if \\( s_t \\) is 2) optional or 3) a procedural mistake; 4) predict missing keysteps (should have been performed before \\( s_t \\) but were not); and 5) next keysteps (for which dependencies are satisfied). The task offers two versions of weak supervision: instance-level: segments and their keystep labels are available for train/test; and procedure-level: only unlabeled segments and procedure-specific keystep names are given for train/test. See Figure 5 (center) and Appendix 13.B.3." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.469, + 0.902 + ], + "angle": 0, + "content": "Related work. Prior work focusing on procedural understanding learns an explicit graph [62, 150, 177] as" + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.09, + 0.891, + 0.236 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.5, + 0.244, + 0.892, + 0.26 + ], + "angle": 0, + "content": "Figure 6. Demonstrator and demonstration proficiency estimation." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.28, + 0.892, + 0.371 + ], + "angle": 0, + "content": "ground truth or uses a task graph for representation learning [9, 107, 202] and short-term step understanding [9, 36, 202]. Other work [32, 139] studies mistake detection in a supervised setting. We are the first to propose procedural understanding to evaluate the long-term structure of the task in a weakly-supervised setting." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.381, + 0.771, + 0.397 + ], + "angle": 0, + "content": "5.3. Ego-exo proficiency estimation" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.404, + 0.892, + 0.48 + ], + "angle": 0, + "content": "Motivation. Going beyond recognizing what a person is doing, this task aims to infer the user's skill level. Such an ability could lead to novel coaching tools that let people learn new skills more effectively, or new ways to evaluate human performance in domains like sports or music." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.48, + 0.893, + 0.691 + ], + "angle": 0, + "content": "Task definition. We consider two variants: (1) demonstrator and (2) demonstration proficiency estimation. Both tasks consider one egocentric and (optionally) \\(M\\) exocentric videos synchronized in time as their inputs. Demonstrator proficiency is formulated as a video classification task, where the model has to output one of four labels (novice, early, intermediate, or late expert). Demonstration proficiency is formulated as a temporal action localization task where given an untrimmed video, the model must output a list of tuples, each containing a timestamp, a proficiency category (i.e., good execution or needs improvement), and its probability. Note that parts of the video that do not reveal the participant's skill are left unlabeled. See Figure 6 and Appendix 13.C." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.692, + 0.892, + 0.812 + ], + "angle": 0, + "content": "Related work. Prior work uses egocentric [12, 35] or exocentric [60, 114, 115] views for proficiency estimation in sports [12, 115, 120], health [60, 92, 191, 208], and others [35, 186]. We propose the first multi-view egocentric and exocentric proficiency estimation benchmark. Unlike prior work, our benchmark spans diverse, day-to-day physical and procedural scenarios and includes temporally localized annotations of (in)correct executions." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.822, + 0.606, + 0.839 + ], + "angle": 0, + "content": "5.4. Ego pose" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.846, + 0.892, + 0.891 + ], + "angle": 0, + "content": "This family of tasks is motivated by recovering the skilled body movements of participants, even in the extreme setting of monocular ego-video input in dynamic environments." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19390" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.094, + 0.091, + 0.46, + 0.222 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.094, + 0.228, + 0.449, + 0.242 + ], + "angle": 0, + "content": "Figure 7. Hand and body keypoints for ego-pose estimation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.255, + 0.47, + 0.481 + ], + "angle": 0, + "content": "Motivation. Estimating the physical state of a person's body—the 3D positions of the arms, legs, hands—from the ego view is essential for wearable AI systems that can support human activity. Challenges include subtle and flexible movements, frequent occlusion, and body parts out of view. Task definition. For both the body and hand pose (\"ego pose\") estimation tasks, the input is an ego video. The output is a set of 3D joint positions of the camera wearer's body and hands for each time step, parameterized as 17 3D body joint positions and 21 3D joint positions per hand, following the MS COCO convention [87]. To our knowledge, EgoExo4D offers the largest manual ground truth (GT) egocentric body and hand pose annotations to date. And, in total, it offers \\(\\sim 14\\mathrm{M}\\) frames of 3D GT and pseudo-GT combined. See Figure 7 and Appendix 13.D." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.483, + 0.468, + 0.618 + ], + "angle": 0, + "content": "Related work. Limited prior work explores 3D body pose from a wearable camera. Some methods assume no body visibility [63, 80, 98, 187, 188], while others assume partial observability by modifying cameras to capture the body [3, 57, 136, 161, 181]. Our dataset can be used for both paradigms. Existing hand pose datasets use constrained environments [106, 146] with simple hand motion [50, 76, 109], whereas we include diverse real-world scenarios, e.g., with expert musicians and bike mechanics." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.633, + 0.205, + 0.649 + ], + "angle": 0, + "content": "6. Conclusions" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.659, + 0.468, + 0.764 + ], + "angle": 0, + "content": "Ego-Exo4D provides a dataset of unprecedented scale and realism for ego-exo video learning. It offers a unique window into skilled human activity from 8 compelling domains by hundreds of real-world experts around the globe. Together with the proposed benchmarks, we hope that this new open source resource will set the stage for substantial new research for the years to come." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.468, + 0.901 + ], + "angle": 0, + "content": "Though we are motivated by skill learning, Ego-Exo4D is poised for even broader influence, beyond the proposed benchmarks. Whereas existing datasets lack activity modeling in real-world 3D contexts (e.g., restricted to mocap suits and/or lab settings), Ego-Exo4D is a resource for general 3D vision—such as environment reconstruction, camera relocalization, audio-visual mapping, and many others. Similarly, our novel video-language resources will offer many opportunities for grounding of actions and objects, multi" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.892, + 0.168 + ], + "angle": 0, + "content": "modal representation learning, and language generation. Finally, though our tasks prioritize perception from the \"ego-only\" perspective, the exo component of our data ensures its utility for the more traditional exo viewpoint too, e.g., for activity recognition and body pose estimation." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.18, + 0.702, + 0.194 + ], + "angle": 0, + "content": "Contribution statement" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.205, + 0.892, + 0.476 + ], + "angle": 0, + "content": "This project is the result of a large collaboration between many institutions over the last two years. Initial authors represent the leadership team of the project. Kristen Grauman initiated the project, served as the technical lead, initiated the recognition and proficiency benchmarks and expert commentary, and coordinated their working groups. Andrew Westbury served as the program manager and operations lead for all aspects of the project. Lorenzo Torresani led development of the capture domains, initiated the relation and ego-pose benchmarks, and coordinated their working groups. Kris Kitani led development of the multi-camera rig and supported the Ego-Exo4D engineering team on all aspects of the data annotation and organization. Jitendra Malik served as a scientific advisor. Authors with stars \\((^{*})\\) were key drivers of implementation, collection, and/or annotation development throughout the project. Authors with daggers \\((^{\\dagger})\\) are faculty and senior researcher PIs for the project. The Appendices detail the contributions of individual authors for the various benchmarks, data collection, and annotation pipelines." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.489, + 0.667, + 0.505 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.513, + 0.892, + 0.734 + ], + "angle": 0, + "content": "We gratefully acknowledge the following colleagues for valuable discussions and support of our project: Vittorio Caggiano, Ilé Danza, Ahmad Darkhalil, Zona de Bloque, Alex Dinh, Rene Martinez Doehner, Ivan Cruz, Matt Feiszli, Vance Feutz, Kelly Forbes, Rohit Girdhar, Pierre Gleize, Andrés Hernández, Shun Iwase, Bolin Lai, Vivian Lee, Brighid Meredith, Ashley Massie, Natalia Neverova, Manohar Paluri, Joelle Pineau, Artsiom Sanakoyeu, Paresh Shenoy, Jiaray Shi, Jiasheng Shi, Gaurav Shrivastava, Mitesh Singh, Manasi Swaminathan, Arjang Talattof, Ali Thabet, Laurens van der Maaten, Andrea Vedaldi, and Toby Zhu. We also sincerely thank the 52 experts who contributed to the expert commentary for their expertise and support; they are listed individually in Appendix 12. Thank you to the Common Visual Data Foundation (CVDF) for hosting the Ego-Exo4D dataset. Finally, thank you to the 740 participants who contributed to this dataset and shared their skills in video." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.736, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Univ. of Bristol is supported in part by EPSRC UMPIRE (EP/T004991/1) and EPSRC PG Visual AI (EP/T028572/1). Z. Zhu is supported by UoB-CSC Scholarship. Univ. of Catalonia is supported in part by the project Future Artificial Intelligence Research (FAIR) - PNRR MUR Cod. PE0000013 - CUP: E63C22001940006. Simon Fraser Univ. is supported in part by the Canada Research Chairs Program (CRC-2019-00298) and NSERC Discovery (2019-06489). Georgia Tech is supported in part by NSF award CNS-2308994. IU is supported in part by NSF DRL-2112635 (AI Institute for Engaged Learning). UT Austin is supported in part by the IFML NSF AI Institute." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "19391" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.079, + 0.138, + 0.173, + 0.153 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.096, + 0.163, + 0.47, + 0.246 + ], + "angle": 0, + "content": "[1] Ahmed Tawfik Aboukhadra, Jameel Malik, Ahmed Elhayek, Nadia Robertini, and Didier Stricker. Thor-net: End-to-end graformer-based realistic two hands and object reconstruction with self-supervision. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1001-1010, 2023. 58, 59" + }, + { + "type": "ref_text", + "bbox": [ + 0.096, + 0.248, + 0.47, + 0.289 + ], + "angle": 0, + "content": "[2] Michael Abrash. Creating the future: Augmented reality, the next human-machine interface. In 2021 IEEE International Electron Devices Meeting (IEDM), 2021. 46" + }, + { + "type": "ref_text", + "bbox": [ + 0.095, + 0.291, + 0.47, + 0.358 + ], + "angle": 0, + "content": "[3] Karan Ahuja, Chris Harrison, Mayank Goel, and Robert Xiao. Mecap: Whole-body digitization for low-cost vr/ar headsets. In Proceedings of the 32nd Annual ACM Symposium on User Interface Software and Technology, pages 453-462, 2019. 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.096, + 0.362, + 0.469, + 0.415 + ], + "angle": 0, + "content": "[4] Jean-Baptiste Alayrac, Piotr Bojanowski, Nishant Agrawal, Josef Sivic, Ivan Laptev, and Simon Lacoste-Julien. Unsupervised learning from narrated instruction videos. In CVPR, 2016. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.096, + 0.418, + 0.469, + 0.445 + ], + "angle": 0, + "content": "[5] Shervin Ardeshir and Ali Borji. Ego2top: Matching viewers in egocentric and top-view videos. In ECCV, 2016. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.096, + 0.447, + 0.469, + 0.488 + ], + "angle": 0, + "content": "[6] Shervin Ardeshir and Ali Borji. Egocentric meets top-view. IEEE transactions on pattern analysis and machine intelligence, 41(6), 2018. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.096, + 0.49, + 0.469, + 0.531 + ], + "angle": 0, + "content": "[7] Shervin Ardeshir and Ali Borji. An exocentric look at egocentric actions and vice versa. Computer Vision and Image Understanding, 171, 2018. 4, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.096, + 0.533, + 0.469, + 0.6 + ], + "angle": 0, + "content": "[8] Kumar Ashutosh, Rohit Girdhar, Lorenzo Torresani, and Kristen Grauman. Hiervl: Learning hierarchical videolanguage embeddings. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.096, + 0.604, + 0.469, + 0.657 + ], + "angle": 0, + "content": "[9] Kumar Ashutosh, Santhosh Kumar Ramakrishnan, Triantafyllos Afouras, and Kristen Grauman. Video-mined task graphs for keystep recognition in instructional videos. In NeurIPS, 2023. 4, 7, 8, 29, 49" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.661, + 0.469, + 0.715 + ], + "angle": 0, + "content": "[10] Siddhant Bansal, Chetan Arora, and C.V. Jawahar. My view is the best view: Procedure learning from egocentric videos. In European Conference on Computer Vision (ECCV), 2022. 3, 4, 7, 49" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.717, + 0.469, + 0.784 + ], + "angle": 0, + "content": "[11] Yizhak Ben-Shabat, Xin Yu, Fatemehsadat Saleh, Dylan Campbell, Cristian Rodriguez-Opazo, Hongdong Li, and Stephen Gould. The ikea asm dataset: Understanding people assembling furniture through actions, objects and pose. 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.788, + 0.469, + 0.829 + ], + "angle": 0, + "content": "[12] Gedas Bertasius, Hyun Soo Park, Stella Yu, and Jianbo Shi. Am i a baller? basketball performance assessment from first-person videos. In ICCV, 2017. 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.831, + 0.469, + 0.871 + ], + "angle": 0, + "content": "[13] Gedas Bertasius, Heng Wang, and Lorenzo Torresani. Is space-time attention all you need for video understanding? In ICML, 2021. 39, 43, 44, 47, 54, 55, 56" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.874, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[14] Jing Bi, Jiebo Luo, and Chenliang Xu. Procedure planning in instructional videos via contextual modeling and model" + }, + { + "type": "list", + "bbox": [ + 0.088, + 0.163, + 0.47, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.515, + 0.094, + 0.892, + 0.133 + ], + "angle": 0, + "content": "based policy learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15611-15620, 2021. 49" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.136, + 0.892, + 0.203 + ], + "angle": 0, + "content": "[15] Jing Bi, Jiebo Luo, and Chenliang Xu. Procedure planning in instructional videos via contextual modeling and model-based policy learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15611-15620, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.205, + 0.892, + 0.272 + ], + "angle": 0, + "content": "[16] Kay Henning Brodersen, Cheng Soon Ong, Klaas Enno Stephan, and Joachim M Buhmann. The balanced accuracy and its posterior distribution. In 2010 20th international conference on pattern recognition, pages 3121-3124. IEEE, 2010. 35" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.274, + 0.892, + 0.328 + ], + "angle": 0, + "content": "[17] Meng Cao, Tianyu Yang, Junwu Weng, Can Zhang, Jue Wang, and Yuexian Zou. Locvtp: Video-text pre-training for temporal localization. In European Conference on Computer Vision, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.331, + 0.892, + 0.411 + ], + "angle": 0, + "content": "[18] Angela Castillo, Maria Escobar, Guillaume Jeanneret, Albert Pumarola, Pablo Arbeláez, Ali Thabet, and Artsiom Sanakoyeu. Bodiffusion: Diffusing sparse observations for full-body human motion synthesis. CV4Metaverse workshop, International Conference on Computer Vision, 2023. 57, 58" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.414, + 0.892, + 0.482 + ], + "angle": 0, + "content": "[19] Eric R Chan, Koki Nagano, Matthew A Chan, Alexander W Bergman, Jeong Joon Park, Axel Levy, Miika Aittala, Shalini De Mello, Tero Karras, and Gordon Wetzstein. Generative novel view synthesis with 3d-aware diffusion models. arXiv preprint arXiv:2304.02602, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.484, + 0.892, + 0.58 + ], + "angle": 0, + "content": "[20] Angel Chang, Angela Dai, Tom Funkhouser, Matthias Nießner, Manolis Savva, Shuran Song, Andy Zeng, and Yinda Zhang. Matterport3d: Learning from rgb-d data in indoor environments. In Proceedings of the International Conference on 3D Vision (3DV), 2017. MatterPort3D dataset license available at: http://kaldir.vc.in.tum.de/matterport/MP_TOS.pdf.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.581, + 0.892, + 0.635 + ], + "angle": 0, + "content": "[21] Chien-Yi Chang, De-An Huang, Danfei Xu, Ehsan Adeli, Li Fei-Fei, and Juan Carlos Niebles. Procedure planning in instructional videos. In European Conference on Computer Vision, pages 334–350. Springer, 2020. 49" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.637, + 0.892, + 0.705 + ], + "angle": 0, + "content": "[22] Chien-Yi Chang, De-An Huang, Danfei Xu, Ehsan Adeli, Li Fei-Fei, and Juan Carlos Niebles. Procedure planning in instructional videos. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part XI, pages 334–350. Springer, 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.707, + 0.892, + 0.761 + ], + "angle": 0, + "content": "[23] Yu-Hsin Chen, Tien-Ju Yang, Joel Emer, and Vivienne Sze. Eyeriss v2: A flexible accelerator for emerging deep neural networks on mobile devices. IEEE Journal on Emerging and Selected Topics in Circuits and Systems, 2019. 46" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.763, + 0.892, + 0.816 + ], + "angle": 0, + "content": "[24] Ho Kei Cheng and Alexander G Schwing. Xmem: Long-term video object segmentation with an atkinson-shiffrin memory model. In European Conference on Computer Vision, pages 640-658. Springer, 2022. 36" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.818, + 0.892, + 0.858 + ], + "angle": 0, + "content": "[25] MMPose Contributors. Openmmlab pose estimation toolbox and benchmark. https://github.com/openmmlab/mmpose, 2020.56" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.86, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[26] Kellie Corona, Katie Osterdahl, Roderic Collins, and Anthony Hoogs. Meva: A large-scale multiview, multimodal video dataset for activity detection. In WACV, 2021. 3, 7" + }, + { + "type": "list", + "bbox": [ + 0.511, + 0.094, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19392" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.092, + 0.47, + 0.176 + ], + "angle": 0, + "content": "[27] Dima Damen, Hazel Doughty, Giovanni Maria Farinella, Sanja Fidler, Antonino Furnari, Evangelos Kazakos, Davide Moltisanti, Jonathan Munro, Toby Perrett, Will Price, and Michael Wray. Scaling egocentric vision: The epic-kitchens dataset. In European Conference on Computer Vision (ECCV), 2018. 3, 27" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.178, + 0.472, + 0.247 + ], + "angle": 0, + "content": "[28] Dima Damen, Hazel Doughty, Giovanni Maria Farinella, Antonino Furnari, Jian Ma, Evangelos Kazakos, Davide Moltisanti, Jonathan Munro, Toby Perrett, Will Price, and Michael Wray. Rescaling egocentric vision. *IJCV*, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.249, + 0.473, + 0.332 + ], + "angle": 0, + "content": "[29] Dima Damen, Hazel Doughty, Giovanni Maria Farinella, Antonino Furnari, Evangelos Kazakos, Jian Ma, Davide Moltisanti, Jonathan Munro, Toby Perrett, Will Price, et al. Rescaling egocentric vision: collection, pipeline and challenges for epic-kitchens-100. International Journal of Computer Vision, 130(1):33-55, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.334, + 0.472, + 0.417 + ], + "angle": 0, + "content": "[30] Roeland De Geest, Efstratios Gavves, Amir Ghodrati, Zhenyang Li, Cees Snoek, and Tinne Tuytelaars. Online action detection. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part V 14, pages 269-284. Springer, 2016. 46, 51" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.418, + 0.472, + 0.487 + ], + "angle": 0, + "content": "[31] Radosvet Desislavov, Fernando Martínez-Plumed, and José Hernández-Orallo. Trends in ai inference energy consumption: Beyond the performance-vs-parameter laws of deep learning. Sustainable Computing: Informatics and Systems, 38:100857, 2023. 46" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.489, + 0.472, + 0.53 + ], + "angle": 0, + "content": "[32] Guodong Ding, Fadime Sener, Shugao Ma, and Angela Yao. Every mistake counts in assembly. arXiv preprint arXiv:2307.16453, 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.532, + 0.472, + 0.588 + ], + "angle": 0, + "content": "[33] Keyan Ding, Kede Ma, Shiqi Wang, and Eero P Simoncelli. Image quality assessment: Unifying structure and texture similarity. IEEE transactions on pattern analysis and machine intelligence, 44(5):2567-2581, 2020. 39" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.59, + 0.472, + 0.63 + ], + "angle": 0, + "content": "[34] H Doughty, D Damen, and W Mayol-Cuevas. Who's better? who's best? pairwise deep ranking for skill determination. In CVPR, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.632, + 0.472, + 0.675 + ], + "angle": 0, + "content": "[35] Hazel Doughty, Walterio Mayol-Cuevas, and Dima Damen. The Pros and Cons: Rank-aware Temporal Attention for Skill Determination in Long Videos. 2019. 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.676, + 0.472, + 0.745 + ], + "angle": 0, + "content": "[36] Nikita Dvornik, Isma Hadji, Hai Pham, Dhaivat Bhatt, Brais Martinez, Afsaneh Fazly, and Allan D Jepson. Flow graph to video grounding for weakly-supervised multi-step localization. In ECCV, pages 319-335. Springer, 2022. 4, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.747, + 0.472, + 0.803 + ], + "angle": 0, + "content": "[37] Ehsan Elhamifar and Dat Huynh. Self-supervised multitask procedure learning from instructional videos. In European Conference on Computer Vision, pages 557-573. Springer, 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.805, + 0.472, + 0.903 + ], + "angle": 0, + "content": "[38] Jakob Engel, Kiran Somasundaram, Michael Goesele, Albert Sun, Alexander Gamino, Andrew Turner, Arjang Talattof, Arnie Yuan, Bilal Souti, Brighid Meredith, Cheng Peng, Chris Sweeney, Cole Wilson, Dan Barnes, Daniel DeTone, David Caruso, Derek Valleroy, Dinesh Ginjupalli, Duncan Frost, Edward Miller, Elias Mueggler, Evgeniy Oleinik, Fan Zhang, Guruprasad Somasundaram, Gustavo" + }, + { + "type": "list", + "bbox": [ + 0.087, + 0.092, + 0.473, + 0.903 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.541, + 0.092, + 0.895, + 0.3 + ], + "angle": 0, + "content": "Solaira, Harry Lanaras, Henry Howard-Jenkins, Huixuan Tang, Hyo Jin Kim, Jaime Rivera, Ji Luo, Jing Dong, Julian Straub, Kevin Bailey, Kevin Eckenhoff, Lingni Ma, Luis Pesqueira, Mark Schwesinger, Maurizio Monge, Nan Yang, Nick Charron, Nikhil Raina, Omkar Parkhi, Peter Borschowa, Pierre Moulon, Prince Gupta, Raul Mur-Artal, Robbie Pennington, Sachin Kulkarni, Sagar Miglani, Santosh Gondi, Saransh Solanki, Sean Diener, Shangyi Cheng, Simon Green, Steve Saarinen, Suvam Patra, Tassos Mourikis, Thomas Whelan, Tripti Singh, Vasileios Balntas, Vijay Baiyya, Wilson Dreewes, Xiaqing Pan, Yang Lou, Yipu Zhao, Yusuf Mansour, Yuyang Zou, Zhaoyang Lv, Zijian Wang, Mingfei Yan, Carl Ren, Renzo De Nardi, and Richard Newcombe. Project Aria: A new tool for egocentric multi-modal AI research, 2023. 3, 4, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.301, + 0.895, + 0.357 + ], + "angle": 0, + "content": "[39] Steven K Esser, Jeffrey L McKinstry, Deepika Bablani, Rathinakumar Appuswamy, and Dharmendra S Modha. Learned step size quantization. arXiv preprint arXiv:1902.08153, 2019. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.358, + 0.895, + 0.412 + ], + "angle": 0, + "content": "[40] Chenyou Fan, Jangwon Lee, Mingze Xu, Krishna Kumar Singh, Yong Jae Lee, David J Crandall, and Michael S Ryoo. Identifying first-person camera wearers in third-person videos. In CVPR, 2017. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.413, + 0.895, + 0.469 + ], + "angle": 0, + "content": "[41] Christoph Feichtenhofer. X3d: Expanding architectures for efficient video recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 203-213, 2020. 8, 45, 46, 47, 48" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.469, + 0.895, + 0.511 + ], + "angle": 0, + "content": "[42] John H. Flavell, Eleanor R. Flavell, Frances L. Green, and Sharon A. Wilcox. The development of three spatial perspective-taking rules. Child Development, 1981. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.511, + 0.895, + 0.552 + ], + "angle": 0, + "content": "[43] Ruohan Gao, Tae-Hyun Oh, Kristen Grauman, and Lorenzo Torresani. Listen to look: Action recognition by previewing audio. In CVPR, 2020. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.553, + 0.895, + 0.621 + ], + "angle": 0, + "content": "[44] Amir Ghodrati, Babak Ehteshami Bejnordi, and Amirhossein Habibian. Frameexit: Conditional early exiting for efficient video recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2021. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.622, + 0.895, + 0.692 + ], + "angle": 0, + "content": "[45] Rohit Girdhar, Mannat Singh, Nikhila Ravi, Laurens van der Maaten, Armand Joulin, and Ishan Misra. Omnivore: A single model for many visual modalities. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16102-16112, 2022. 54" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.692, + 0.895, + 0.789 + ], + "angle": 0, + "content": "[46] Raghav Goyal, Samira Ebrahimi Kahou, Vincent Michalski, Joanna Materzynska, Susanne Westphal, Heuna Kim, Valentin Haenel, Ingo Fruend, Peter Yianilos, Moritz Mueller-Freitag, et al. The \"something something\" video database for learning and evaluating visual common sense. In Proceedings of the IEEE international conference on computer vision, pages 5842-5850, 2017. 42" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.79, + 0.895, + 0.903 + ], + "angle": 0, + "content": "[47] Kristen Grauman, Andrew Westbury, Eugene Byrne, Zachary Chavis, Antonino Furnari, Rohit Girdhar, Jackson Hamburger, Hao Jiang, Miao Liu, Xingyu Liu, Miguel Martin, Tushar Nagarajan, Ilija Radosavovic, Santhosh Kumar Ramakrishnan, Fiona Ryan, Jayant Sharma, Michael Wray, Mengmeng Xu, Eric Zhongcong Xu, Chen Zhao, Siddhant Bansal, Dhruv Batra, Vincent Cartillier, Sean Crane, Tien Do, Morrie Doulaty, Akshay Erapalli, Christoph" + }, + { + "type": "list", + "bbox": [ + 0.51, + 0.092, + 0.895, + 0.903 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "19393" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.118, + 0.092, + 0.47, + 0.355 + ], + "angle": 0, + "content": "Feichtenhofer, Adriano Fragomeni, Qichen Fu, Abraham Gebreselasie, Cristina Gonzalez, James Hillis, Xuhua Huang, Yifei Huang, Wenqi Jia, Weslie Khoo, Jachym Kolář, Satwik Kottur, Anurag Kumar, Federico Landini, Chao Li, Yanghao Li, Zhenqiang Li, Karttikeya Mangalam, Raghava Modhugu, Jonathan Munro, Tullie Murrell, Takumi Nishiyasu, Will Price, Paola Ruiz, Merey Ramazanova, Leda Sari, Kiran Somasundaram, Audrey Southerland, Yusuke Sugano, Ruijie Tao, Minh Vo, Yuchen Wang, Xindi Wu, Takuma Yagi, Ziwei Zhao, Yunyi Zhu, Pablo Arbelaez, David Crandall, Dima Damen, Giovanni Maria Farinella, Christian Fuegen, Bernard Ghanem, Vamsi Krishna Ithapu, C. V. Jawahar, Hanbyul Joo, Kris Kitani, Haizhou Li, Richard Newcombe, Aude Oliva, Hyun Soo Park, James M. Rehg, Yoichi Sato, Jianbo Shi, Mike Zheng Shou, Antonio Torralba, Lorenzo Torresani, Mingfei Yan, and Jitendra Malik. Ego4D: Around the world in 3,000 hours of egocentric video. In CVPR, 2022. 2, 3, 4, 6, 7, 29, 34, 43" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.357, + 0.47, + 0.438 + ], + "angle": 0, + "content": "[48] Chunhui Gu, Chen Sun, David A. Ross, Carl Vondrick, Caroline Pantofaru, Yeqing Li, Sudheendra Vijayanarasimhan, George Toderici, Susanna Ricco, Rahul Sukthankar, Cordelia Schmid, and Jitendra Malik. Ava: A video dataset of spatio-temporally localized atomic visual actions. In CVPR, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.44, + 0.468, + 0.495 + ], + "angle": 0, + "content": "[49] Vladimir Guzov, Aymen Mir, Torsten Sattler, and Gerard Pons-Moll. Human poseitioning system (hps): 3d human pose estimation and self-localization in large scenes from body-mounted sensors. In CVPR, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.497, + 0.469, + 0.564 + ], + "angle": 0, + "content": "[50] Shreyas Hampali, Mahdi Rad, Markus Oberweger, and Vincent Lepetit. Honnotate: A method for 3d annotation of hand and object poses. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3196-3206, 2020. 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.566, + 0.468, + 0.606 + ], + "angle": 0, + "content": "[51] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, 2016. 36, 58" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.608, + 0.469, + 0.648 + ], + "angle": 0, + "content": "[52] Geoffrey Hinton, Oriol Vinyals, and Jeff Dean. Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531, 2015. 43, 44" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.65, + 0.469, + 0.691 + ], + "angle": 0, + "content": "[53] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in Neural Information Processing Systems, 33:6840-6851, 2020. 57" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.692, + 0.469, + 0.734 + ], + "angle": 0, + "content": "[54] Alain Hore and Djemel Ziou. Image quality metrics: Psnr vs. ssim. In 2010 20th international conference on pattern recognition, pages 2366-2369. IEEE, 2010. 39" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.735, + 0.469, + 0.789 + ], + "angle": 0, + "content": "[55] Mark Horowitz. 1.1 computing's energy problem (and what we can do about it). In 2014 IEEE international solid-state circuits conference digest of technical papers (ISSCC), 2014. 46" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.79, + 0.469, + 0.858 + ], + "angle": 0, + "content": "[56] Andrew G Howard, Menglong Zhu, Bo Chen, Dmitry Kalenichenko, Weijun Wang, Tobias Weyand, Marco Andretto, and Hartwig Adam. Mobilenets: Efficient convolutional neural networks for mobile vision applications. arXiv preprint arXiv:1704.04861, 2017. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.86, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[57] Dong-Hyun Hwang, Kohei Aso, Ye Yuan, Kris Kitani, and Hideki Koike. Monoeye: Multimodal human motion capture system using a single ultra-wide fisheye camera. In" + }, + { + "type": "list", + "bbox": [ + 0.088, + 0.092, + 0.47, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.541, + 0.093, + 0.892, + 0.121 + ], + "angle": 0, + "content": "Proceedings of the 33rd Annual ACM Symposium on User Interface Software and Technology, pages 98-111, 2020. 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.123, + 0.892, + 0.19 + ], + "angle": 0, + "content": "[58] Forrest N Iandola, Song Han, Matthew W Moskewicz, Khalid Ashraf, William J Dally, and Kurt Keutzer. SqueezeNet: Alexnet-level accuracy with 50x fewer parameters and; 0.5 mb model size. arXiv preprint arXiv:1602.07360, 2016. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.193, + 0.892, + 0.247 + ], + "angle": 0, + "content": "[59] Vladimir Iashin and Esa Rahtu. Multi-modal dense video captioning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops, pages 958-959, 2020. 26" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.25, + 0.892, + 0.331 + ], + "angle": 0, + "content": "[60] Hassan Ismail Fawaz, Germain Forestier, Jonathan Weber, Lhassane Idoumghar, and Pierre-Alain Muller. Evaluating surgical skills from kinematic data using convolutional neural networks. In Medical Image Computing and Computer Assisted Intervention - MICCAI 2018, pages 214-221, 2018. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.334, + 0.892, + 0.375 + ], + "angle": 0, + "content": "[61] Phillip Isola, Jun-Yan Zhu, Tinghui Zhou, and Alexei A Efros. Image-to-image translation with conditional adversarial networks. CVPR, 2017. 39" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.378, + 0.892, + 0.432 + ], + "angle": 0, + "content": "[62] Yunseok Jang, Sungryull Sohn, Lajanugen Logeswaran, Tiange Luo, Moontae Lee, and Honglak Lee. Multimodal subtask graph generation from instructional videos. arXiv preprint arXiv:2302.08672, 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.434, + 0.892, + 0.474 + ], + "angle": 0, + "content": "[63] Hao Jiang and Kristen Grauman. Seeing invisible poses: Estimating 3d body pose from egocentric video. In CVPR, 2017. 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.477, + 0.892, + 0.546 + ], + "angle": 0, + "content": "[64] Jiaxi Jiang, Paul Streli, Huajian Qiu, Andreas Fender, Larissa Laich, Patrick Snape, and Christian Holz. Avatarposer: Articulated full-body pose tracking from sparse motion sensing. In European Conference on Computer Vision, pages 443-460. Springer, 2022. 57, 58" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.548, + 0.892, + 0.616 + ], + "angle": 0, + "content": "[65] Wei Jiang, Eduard Trulls, Jan Hosang, Andrea Tagliasacchi, and Kwang Moo Yi. Cotr: Correspondence transformer for matching across images. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6207-6217, 2021. 6, 35" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.619, + 0.892, + 0.701 + ], + "angle": 0, + "content": "[66] Hanbyul Joo, Tomas Simon, Xulong Li, Hao Liu, Lei Tan, Lin Gui, Sean Banerjee, Timothy Scott Godisart, Bart Nabbe, Iain Matthews, Takeo Kanade, Shohei Nobuhara, and Yaser Sheikh. Panoptic studio: A massively multiview system for social interaction capture. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.703, + 0.892, + 0.771 + ], + "angle": 0, + "content": "[67] Will Kay, Joao Carreira, Karen Simonyan, Brian Zhang, Chloe Hillier, Sudheendra Vijayanarasimhan, Fabio Viola, Tim Green, Trevor Back, Paul Natsev, et al. The kinetics human action video dataset. arXiv preprint arXiv:1705.06950, 2017. 2, 42, 43" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.774, + 0.892, + 0.815 + ], + "angle": 0, + "content": "[68] Rawal Khirodkar, Aayush Bansal, Lingni Ma, Richard Newcombe, Minh Vo, and Kris Kitani. Egohumans: An egocentric 3d multi-human benchmark. In ICCV, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.817, + 0.892, + 0.857 + ], + "angle": 0, + "content": "[69] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 36" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.86, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[70] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C. Berg, Wan-Yen Lo, Piotr Dollar," + }, + { + "type": "list", + "bbox": [ + 0.51, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "19394" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.118, + 0.092, + 0.471, + 0.134 + ], + "angle": 0, + "content": "and Ross Girshick. Segment anything. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 4015-4026, 2023. 33" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.136, + 0.471, + 0.219 + ], + "angle": 0, + "content": "[71] Dohwan Ko, Joonmyung Choi, Juyeon Ko, Shinyeong Noh, Kyoung-Woon On, Eun-Sol Kim, and Hyunwoo J Kim. Video-text representation learning via differentiable weak temporal alignment. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5016-5025, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.221, + 0.471, + 0.29 + ], + "angle": 0, + "content": "[72] Nikos Kolotouros, Georgios Pavlakos, Michael J Black, and Kostas Daniilidis. Learning to reconstruct 3d human pose and shape via model-fitting in the loop. In Proceedings of the IEEE/CVF international conference on computer vision, pages 2252-2261, 2019. 59" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.292, + 0.471, + 0.347 + ], + "angle": 0, + "content": "[73] Bruno Korbar, Du Tran, and Lorenzo Torresani. Scsampler: Sampling salient clips from video for efficient action recognition. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2019. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.349, + 0.471, + 0.417 + ], + "angle": 0, + "content": "[74] Zuzana Kukelova, Jan Heller, and Andrew Fitzgibbon. Efficient intersection of three quadrics and applications in computer vision. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 1799-1808, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.42, + 0.471, + 0.488 + ], + "angle": 0, + "content": "[75] Iljung Kwak, Jian-Zhong Guo, Adam Hantman, David Kriegman, and Kristin Branson. Detecting the starting frame of actions in video. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 489-497, 2020. 54" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.49, + 0.471, + 0.559 + ], + "angle": 0, + "content": "[76] Taein Kwon, Bugra Tekin, Jan Stühmer, Federica Bogo, and Marc Pollefeys. H2o: Two hands manipulating objects for first person interaction recognition. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 10138-10148, 2021. 2, 3, 9, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.562, + 0.471, + 0.63 + ], + "angle": 0, + "content": "[77] F. De la Torre, J. Hodgins, J. Montano, S. Valcarcel, R. Forcada, and J. Macey. Guide to the carnegie mellon university multimodal activity (cmu-mmac) database. In Tech. report CMU-RI-TR-08-22, Robotics Institute, Carnegie Mellon University, 2009. 2, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.632, + 0.471, + 0.672 + ], + "angle": 0, + "content": "[78] Y. J. Lee, J. Ghosh, and K. Grauman. Discovering important people and objects for egocentric video summarization. In CVPR, 2012. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.675, + 0.471, + 0.757 + ], + "angle": 0, + "content": "[79] Jiefeng Li, Chao Xu, Zhicun Chen, Siyuan Bian, Lixin Yang, and Cewu Lu. Hybrik: A hybrid analytical-neural inverse kinematics solution for 3d human pose and shape estimation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3383-3393, 2021. 59" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.76, + 0.471, + 0.814 + ], + "angle": 0, + "content": "[80] Jiaman Li, Karen Liu, and Jiajun Wu. Ego-body pose estimation via ego-head pose estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17142-17151, 2023. 3, 9, 56, 57, 58" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.817, + 0.471, + 0.857 + ], + "angle": 0, + "content": "[81] Yin Li, Miao Liu, and James M. Rehg. In the eye of beholder: Joint learning of gaze and actions in first person video. In ECCV, 2018. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.86, + 0.471, + 0.901 + ], + "angle": 0, + "content": "[82] Yanghao Li, Tushar Nagarajan, Bo Xiong, and Kristen Grauman. Ego-exo: Transferring visual representations from third-person to first-person videos. In Proceedings of" + }, + { + "type": "list", + "bbox": [ + 0.088, + 0.092, + 0.471, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.541, + 0.093, + 0.892, + 0.121 + ], + "angle": 0, + "content": "the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6943-6953, 2021. 4, 7, 21, 43, 44" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.123, + 0.892, + 0.191 + ], + "angle": 0, + "content": "[83] Junhua Liao, Haihan Duan, Kanghui Feng, Wanbing Zhao, Yanbing Yang, and Liangyin Chen. A light weight model for active speaker detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22932-22941, 2023. 47, 48" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.193, + 0.892, + 0.248 + ], + "angle": 0, + "content": "[84] Kevin Lin, Lijuan Wang, and Zicheng Liu. End-to-end human pose and mesh reconstruction with transformers. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1954-1963, 2021. 59" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.251, + 0.892, + 0.306 + ], + "angle": 0, + "content": "[85] Kevin Qinghong Lin, Alex Jinpeng Wang, Mattia Soldan, Michael Wray, Rui Yan, Eric Zhongcong Xu, Difei Gao, Rongcheng Tu, Wenzhe Zhao, Weijie Kong, et al. Egocentric video-language pretraining. NeurIPS, 2022. 6, 26, 50" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.308, + 0.892, + 0.347 + ], + "angle": 0, + "content": "[86] TY Lin, Y Cui, S Belongie, and J Hays. Learning deep representations for ground-to-aerial geolocation. In CVPR, 2015. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.35, + 0.892, + 0.405 + ], + "angle": 0, + "content": "[87] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dolkar, and C Lawrence Zitnick. Microsoft COCO: Common objects in context. In ECCV, 2014. 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.408, + 0.892, + 0.475 + ], + "angle": 0, + "content": "[88] Tsung-Yi Lin, Piotr Dólar, Ross Girshick, Kaiming He, Bharath Hariharan, and Serge Belongie. Feature pyramid networks for object detection. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2117-2125, 2017. 58" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.478, + 0.892, + 0.559 + ], + "angle": 0, + "content": "[89] Xudong Lin, Fabio Petroni, Gedas Bertasius, Marcus Rohrbach, Shih-Fu Chang, and Lorenzo Torresani. Learning to recognize procedural activities with distant supervision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13853-13863, 2022. 4, 29" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.563, + 0.892, + 0.644 + ], + "angle": 0, + "content": "[90] Andrew Liu, Richard Tucker, Varun Jampani, Ameesh Makadia, Noah Snavely, and Angjoo Kanazawa. Infinite nature: Perpetual view generation of natural scenes from a single image. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14458-14467, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.647, + 0.892, + 0.73 + ], + "angle": 0, + "content": "[91] Chiao Liu, Lyle Bainbridge, Andrew Berkovich, Song Chen, Wei Gao, Tsung-Hsun Tsai, Kazuya Mori, Rimon Ikeno, Masayuki Uno, Toshiyuki Isozaki, et al. A \\(4.6\\mu \\mathrm{m}\\), \\(512\\times 512\\), ultra-low power stacked digital pixel sensor with triple quantization and 127db dynamic range. In 2020 IEEE International Electron Devices Meeting (IEDM), 2020. 46" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.733, + 0.892, + 0.801 + ], + "angle": 0, + "content": "[92] Daochang Liu, Qiyue Li, Tingting Jiang, Yizhou Wang, Rulin Miao, Fei Shan, and Ziyu Li. Towards unified surgical skill assessment. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9522-9531, 2021. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.803, + 0.892, + 0.871 + ], + "angle": 0, + "content": "[93] Gaowen Liu, Hao Tang, Hugo Latapie, and Yan Yan. Exocentric to egocentric image generation via parallel generative adversarial network. In ICASSP 2020-2020 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 1843-1847. IEEE, 2020. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.874, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[94] Gaowen Liu, Hao Tang, Hugo M Latapie, Jason J Corso, and Yan Yan. Cross-view exocentric to egocentric video" + }, + { + "type": "list", + "bbox": [ + 0.511, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19395" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.118, + 0.092, + 0.47, + 0.12 + ], + "angle": 0, + "content": "synthesis. In Proceedings of the 29th ACM International Conference on Multimedia, pages 974-982, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.122, + 0.47, + 0.163 + ], + "angle": 0, + "content": "[95] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In International Conference on Learning Representations, 2019. 47" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.165, + 0.469, + 0.234 + ], + "angle": 0, + "content": "[96] Xiaohu Lu, Zuoyue Li, Zhaopeng Cui, Martin R. Oswald, Marc Pollefeys, and Rongjun Qin. Geometry-aware satellite-to-ground image synthesis for urban areas. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.236, + 0.469, + 0.277 + ], + "angle": 0, + "content": "[97] Mi Luo, Zihui Xue, Alex Dimakis, and Kristen Grauman. Put myself in your shoes: Lifting the egocentric perspective from exocentric videos. arXiv:2403.06351, 2024. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.279, + 0.469, + 0.333 + ], + "angle": 0, + "content": "[98] Zhengyi Luo, Ryo Hachiuma, Ye Yuan, and Kris Kitani. Dynamics-regulated kinematic policy for egocentric pose estimation. In Advances in Neural Information Processing Systems, 2021. 9, 56, 57, 58" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.336, + 0.469, + 0.39 + ], + "angle": 0, + "content": "[99] Naureen Mahmood, Nima Ghorbani, Nikolaus F. Troje, Gerard Pons-Moll, and Michael J. Black. Amass: Archive of motion capture as surface shapes. In The IEEE International Conference on Computer Vision (ICCV), 2019. 57" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.392, + 0.469, + 0.434 + ], + "angle": 0, + "content": "[100] Effrosyni Mavroudi, Triantafyllos Afouras, and Lorenzo Torresani. Learning to ground instructional articles in videos through narrations. 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.435, + 0.469, + 0.478 + ], + "angle": 0, + "content": "[101] Sachin Mehta and Mohammad Rastegari. Mobilevit: lightweight, general-purpose, and mobile-friendly vision transformer. arXiv preprint arXiv:2110.02178, 2021. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.479, + 0.469, + 0.533 + ], + "angle": 0, + "content": "[102] Yue Meng, Chung-Ching Lin, Rameswar Panda, Prasanna Sattigeri, Leonid Karlinsky, Aude Oliva, Kate Saenko, and Rogerio Feris. Ar-net: Adaptive frame resolution for efficient action recognition. In ECCV 2020, 2020. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.535, + 0.469, + 0.604 + ], + "angle": 0, + "content": "[103] Antoine Miech, Dimitri Zhukov, Jean-Baptiste Alayrac, Makarand Tapaswi, Ivan Laptev, and Josef Sivic. HowTo100M: Learning a Text-Video Embedding by Watching Hundred Million Narrated Video Clips. In ICCV, 2019. 2, 4, 29" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.606, + 0.469, + 0.688 + ], + "angle": 0, + "content": "[104] Antoine Miech, Jean-Baptiste Alayrac, Lucas Smaira, Ivan Laptev, Josef Sivic, and Andrew Zisserman. End-to-end learning of visual representations from uncurated instructional videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9879-9889, 2020. 4, 29" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.69, + 0.469, + 0.759 + ], + "angle": 0, + "content": "[105] Mathew Monfort, Alex Andonian, Bolei Zhou, Kandan Ramakrishnan, Sarah Adel Bargal, Tom Yan, Lisa Brown, Quanfu Fan, Dan Gutfreund, Carl Vondrick, and Aude Oliva. Moments in time dataset: one million videos for event understanding. PAMI, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.761, + 0.469, + 0.844 + ], + "angle": 0, + "content": "[106] Gyeongsik Moon, Shoou-I Yu, He Wen, Takaaki Shiratori, and Kyoung Mu Lee. Interhand2. 6m: A dataset and baseline for 3d interacting hand pose estimation from a single rgb image. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XX 16, pages 548-564. Springer, 2020. 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.846, + 0.469, + 0.9 + ], + "angle": 0, + "content": "[107] Medhini Narasimhan, Licheng Yu, Sean Bell, Ning Zhang, and Trevor Darrell. Learning and verification of task structure in instructional videos. arXiv preprint arXiv:2303.13519, 2023. 4, 8" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.133 + ], + "angle": 0, + "content": "[108] Nora Newcombe. The development of spatial perspective taking. Advances in child development and behavior, 1989. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.136, + 0.892, + 0.218 + ], + "angle": 0, + "content": "[109] Takehiko Ohkawa, Kun He, Fadime Sener, Tomas Hodan, Luan Tran, and Cem Keskin. Assemblyhands: Towards egocentric activity understanding via 3d hand pose estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12999-13008, 2023. 9, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.22, + 0.892, + 0.261 + ], + "angle": 0, + "content": "[110] Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748, 2018. 43, 44" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.263, + 0.892, + 0.317 + ], + "angle": 0, + "content": "[111] Boxiao Pan, Haoye Cai, De-An Huang, Kuan-Hui Lee, Adrien Gaidon, Ehsan Adeli, and Juan Carlos Niebles. Spatio-temporal graph for video captioning with knowledge distillation. In CVPR, 2020. 26" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.32, + 0.892, + 0.388 + ], + "angle": 0, + "content": "[112] JoonKyu Park, Yeonguk Oh, Gyeongsik Moon, Hongsuk Choi, and Kyoung Mu Lee. Handoccnet: Occlusion-robust 3d hand mesh estimation network. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1496–1505, 2022. 58, 59" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.39, + 0.892, + 0.418 + ], + "angle": 0, + "content": "[113] Paritosh Parmar and Brendan Morris. Action quality assessment across multiple actions. In WACV, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.42, + 0.892, + 0.446 + ], + "angle": 0, + "content": "[114] Paritosh Parmar and Brendan Tran Morris. Learning to score olympic events, 2017. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.448, + 0.892, + 0.516 + ], + "angle": 0, + "content": "[115] Paritosh Parmar and Brendan Tran Morris. What and how well you performed? a multitask learning approach to action quality assessment. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 304-313, 2019. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.518, + 0.892, + 0.559 + ], + "angle": 0, + "content": "[116] William Peebles and Saining Xie. Scalable diffusion models with transformers. arXiv preprint arXiv:2212.09748, 2022.39" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.562, + 0.892, + 0.643 + ], + "angle": 0, + "content": "[117] Federico Perazzi, Jordi Pont-Tuset, Brian McWilliams, Luc Van Gool, Markus Gross, and Alexander Sorkine-Hornung. A benchmark dataset and evaluation methodology for video object segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 724-732, 2016. 35, 39" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.645, + 0.892, + 0.701 + ], + "angle": 0, + "content": "[118] Ethan Perez, Florian Strub, Harm De Vries, Vincent Dumoulin, and Aaron Courville. Film: Visual reasoning with a general conditioning layer. In Proceedings of the AAAI conference on artificial intelligence, 2018. 39" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.702, + 0.892, + 0.73 + ], + "angle": 0, + "content": "[119] H. Pirsiavash and D. Ramanan. Detecting activities of daily living in first-person camera views. In CVPR, 2012. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.732, + 0.892, + 0.759 + ], + "angle": 0, + "content": "[120] H. Piriavash, C. Vondrick, and A. Torralba. Assessing the quality of actions. In ECCV, 2014. 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.761, + 0.892, + 0.801 + ], + "angle": 0, + "content": "[121] Antonio Polino, Razvan Pascanu, and Dan Alistarh. Model compression via distillation and quantization. arXiv preprint arXiv:1802.05668, 2018. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.803, + 0.892, + 0.872 + ], + "angle": 0, + "content": "[122] Daniel Povey, Arnab Ghoshal, Gilles Boulianne, Lukávs. Burget, Ondrej Glembek, Nagendra Kumar Goel, Mirko Hannemann, Petr Motlícek, Yanmin Qian, Petr Schwarz, Jan Silovsky, Georg Stemmer, and Karel Vesely. The kaldi speech recognition toolkit. 2011. 47" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.874, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[123] Shraman Pramanick, Yale Song, Sayan Nag, Kevin Qinghong Lin, Hardik Shah, Mike Zheng Shou," + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "19396" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.118, + 0.092, + 0.47, + 0.161 + ], + "angle": 0, + "content": "Rama Chellappa, and Pengchuan Zhang. Egovlpv2: Egocentric video-language pre-training with fusion in the backbone. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5285-5297, 2023. 6, 26, 43, 44" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.164, + 0.47, + 0.246 + ], + "angle": 0, + "content": "[124] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 39" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.248, + 0.469, + 0.316 + ], + "angle": 0, + "content": "[125] Alec Radford, Jong Wook Kim, Tao Xu, Greg Brockman, Christine McLeavey, and Ilya Sutskever. Robust speech recognition via large-scale weak supervision. In International Conference on Machine Learning, pages 28492-28518. PMLR, 2023. 27" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.319, + 0.469, + 0.374 + ], + "angle": 0, + "content": "[126] Francesco Ragusa, Antonino Furnari, Salvatore Livatino, and Giovanni Maria Farinella. The meccano dataset: Understanding human-object interactions from egocentric videos in an industrial-like domain. In WACV, 2021. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.376, + 0.469, + 0.416 + ], + "angle": 0, + "content": "[127] N. Rai, H. Chen, J. Ji, R. Desai, K. Kozuka, S. Ishizaka, E. Adeli, and J.C. Niebles. Home action genome: Contrastive compositional action understanding. In CVPR, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.419, + 0.469, + 0.528 + ], + "angle": 0, + "content": "[128] Santhosh Kumar Ramakrishnan, Aaron Gokaslan, Erik Wijmans, Oleksandr Maksymets, Alexander Clegg, John M Turner, Eric Undersander, Wojciech Galuba, Andrew Westbury, Angel X Chang, Manolis Savva, Yili Zhao, and Dhruv Batra. Habitat-matterport 3d dataset (HM3d): 1000 large-scale 3d environments for embodied AI. In Thirty-fifth Conference on Neural Information Processing Systems Datasets and Benchmarks Track, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.531, + 0.469, + 0.586 + ], + "angle": 0, + "content": "[129] Scott Reed, Honglak Lee, Dragomir Anguelov, Christian Szegedy, Dumitru Erhan, and Andrew Rabinovich. Training deep neural networks on noisy labels with bootstrapping. arXiv preprint arXiv:1412.6596, 2014. 39" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.588, + 0.469, + 0.629 + ], + "angle": 0, + "content": "[130] Krishna Regmi and Ali Borji. Cross-view image synthesis using conditional gans. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 4, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.631, + 0.469, + 0.672 + ], + "angle": 0, + "content": "[131] Krishna Regmi and Ali Borji. Cross-view image synthesis using geometry-guided conditional gans. Computer Vision and Image Understanding, 2019. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.674, + 0.469, + 0.702 + ], + "angle": 0, + "content": "[132] Krishna Regmi and Mubarak Shah. Bridging the domain gap for ground-to-aerial image matching. In ICCV, 2019. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.703, + 0.469, + 0.758 + ], + "angle": 0, + "content": "[133] Jeremy Reizenstein, Roman Shapovalov, Philipp Henzler, Luca Sbordone, Patrick Labatut, and David Novotny. Common objects in 3d: Large-scale learning and evaluation of real-life 3d category reconstruction. In ICCV, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.76, + 0.469, + 0.8 + ], + "angle": 0, + "content": "[134] Bin Ren, Hao Tang, and Nicu Sebe. Cascaded cross mlp-mixer gans for cross-view image translation. arXiv preprint arXiv:2110.10183, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.803, + 0.469, + 0.87 + ], + "angle": 0, + "content": "[135] Xuanchi Ren and Xiaolong Wang. Look outside the room: Synthesizing a consistent long-term 3d scene video from a single image. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3563-3573, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.873, + 0.469, + 0.9 + ], + "angle": 0, + "content": "[136] Helge Rhodin, Christian Richardt, Dan Casas, Eldar In-safutdinov, Mohammad Shafiei, Hans-Peter Seidel, Bernt" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.541, + 0.093, + 0.892, + 0.134 + ], + "angle": 0, + "content": "Schiele, and Christian Theobalt. Egocap: egocentric marker-less motion capture with two fisheye cameras. ACM Transactions on Graphics (TOG), 35(6):1-11, 2016. 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.136, + 0.892, + 0.19 + ], + "angle": 0, + "content": "[137] Robin Rombach, Patrick Esser, and Björn Ommer. Geometry-free view synthesis: Transformers and no 3d priors. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14356-14366, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.192, + 0.892, + 0.245 + ], + "angle": 0, + "content": "[138] Javier Romero, Dimitrios Tzionas, and Michael J. Black. Embodied hands: Modeling and capturing hands and bodies together. ACM Transactions on Graphics, (Proc. SIGGRAPH Asia), 36(6), 2017. 59" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.247, + 0.892, + 0.329 + ], + "angle": 0, + "content": "[139] Fadime Sener, Dibyadip Chatterjee, Daniel Shelepov, Kun He, Dipika Singhania, Robert Wang, and Angela Yao. Assembly101: A large-scale multi-view video dataset for understanding procedural activities. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21096-21106, 2022. 2, 3, 8, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.331, + 0.892, + 0.399 + ], + "angle": 0, + "content": "[140] Pierre Sermanet, Corey Lynch, Jasmine Hsu, and Sergey Levine. Time-contrastive networks: Self-supervised learning from multi-view observation. In 2017 IEEE Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), pages 486–487. IEEE, 2017. 43" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.401, + 0.892, + 0.468 + ], + "angle": 0, + "content": "[141] Pierre Sermanet, Corey Lynch, Yevgen Chebotar, Jasmine Hsu, Eric Jang, Stefan Schaal, and Sergey Levine. Time-contrastive networks: Self-supervised learning from video. Proceedings of International Conference in Robotics and Automation (ICRA), 2018. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.47, + 0.892, + 0.537 + ], + "angle": 0, + "content": "[142] Xi Shen, Alexei A Efros, Armand Joulin, and Mathieu Aubry. Learning co-segmentation by segment swapping for retrieval and discovery. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5082-5092, 2022. 35, 36" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.539, + 0.892, + 0.633 + ], + "angle": 0, + "content": "[143] Nina Shvetsova, Brian Chen, Andrew Rouditchenko, Samuel Thomas, Brian Kingsbury, Rogerio S Feris, David Harwath, James Glass, and Hilde Kuehne. Everything at once-multi-modal fusion transformer for video retrieval. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20020-20029, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.636, + 0.892, + 0.69 + ], + "angle": 0, + "content": "[144] Gunnar A Sigurdsson, Abhinav Gupta, Cordelia Schmid, Ali Farhadi, and Karteek Alahari. Actor and observer: Joint modeling of first and third-person videos. In CVPR, 2018, 4, 7, 43" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.692, + 0.892, + 0.747 + ], + "angle": 0, + "content": "[145] Gunnar A Sigurdsson, Abhinav Gupta, Cordelia Schmid, Ali Farhadi, and Karteek Alahari. Charades-ego: A large-scale dataset of paired third and first person videos. arXiv preprint arXiv:1804.09626, 2018. 2, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.749, + 0.892, + 0.815 + ], + "angle": 0, + "content": "[146] Tomas Simon, Hanbyul Joo, Iain Matthews, and Yaser Sheikh. Hand keypoint detection in single images using multiview bootstrapping. In Proceedings of the IEEE conference on Computer Vision and Pattern Recognition, pages 1145-1153, 2017. 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.818, + 0.892, + 0.871 + ], + "angle": 0, + "content": "[147] Krishna Kumar Singh, Kayvon Fatahalian, and Alexei A. Efros. Krishnamac: Using a longitudinal, single-person, egocentric dataset for scene understanding tasks. In WACV, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.873, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[148] Yale Song, Eugene Byrne, Tushar Nagarajan, Huiyu Wang, Miguel Martin, and Lorenzo Torresani. Ego4d goal-step:" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19397" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.118, + 0.092, + 0.468, + 0.12 + ], + "angle": 0, + "content": "Toward hierarchical understanding of procedural activities. In NeurIPS, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.122, + 0.469, + 0.163 + ], + "angle": 0, + "content": "[149] Khurram Soomro, Amir Roshan Zamir, and Mubarak Shah. Ucf101: A dataset of 101 human action classes from videos in the wild. In CRCV-TR-12-01, 2012. 2, 42" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.166, + 0.469, + 0.207 + ], + "angle": 0, + "content": "[150] Bilge Soran, Ali Farhadi, and Linda Shapiro. Generating notifications for missing actions: Don't forget to turn the lights off! In ICCV, pages 4669-4677, 2015. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.209, + 0.469, + 0.347 + ], + "angle": 0, + "content": "[151] Julian Straub, Thomas Whelan, Lingni Ma, Yufan Chen, Erik Wijmans, Simon Green, Jakob J. Engel, Raul Mur-Artal, Carl Ren, Shobhit Verma, Anton Clarkson, Mingfei Yan, Brian Budge, Yajie Yan, Xiaqing Pan, June Yon, Yuyang Zou, Kimberly Leon, Nigel Carter, Jesus Briales, Tyler Gillingham, Elias Mueggler, Luis Pesqueira, Manolis Savva, Dhruv Batra, Hauke M. Strasdat, Renzo De Nardi, Michael Goesele, Steven Lovegrove, and Richard Newcombe. The Replica dataset: A digital replica of indoor spaces. arXiv preprint arXiv:1906.05797, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.349, + 0.469, + 0.473 + ], + "angle": 0, + "content": "[152] Carole H Sudre, Wenqi Li, Tom Vercauteren, Sebastien Ourselin, and M Jorge Cardoso. Generalised dice overlap as a deep learning loss function for highly unbalanced segmentations. In Deep Learning in Medical Image Analysis and Multimodal Learning for Clinical Decision Support: Third International Workshop, DLMIA 2017, and 7th International Workshop, ML-CDS 2017, Held in conjunction with MICCAI 2017, Quebec City, QC, Canada, September 14, Proceedings 3, pages 240-248. Springer, 2017. 39" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.476, + 0.469, + 0.53 + ], + "angle": 0, + "content": "[153] Vivienne Sze, Yu-Hsin Chen, Tien-Ju Yang, and Joel S Emer. How to evaluate deep neural network processors: Tops/w (alone) considered harmful. IEEE Solid-State Circuits Magazine, 2020. 46" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.533, + 0.469, + 0.587 + ], + "angle": 0, + "content": "[154] Mukund Varma T, Peihao Wang, Xuxi Chen, Tianlong Chen, Subhashini Venugopalan, and Zhangyang Wang. Is attention all that neRF needs? In The Eleventh International Conference on Learning Representations, 2023. 39" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.59, + 0.469, + 0.659 + ], + "angle": 0, + "content": "[155] Mingxing Tan, Bo Chen, Ruoming Pang, Vijay Vasudevan, Mark Sandler, Andrew Howard, and Quoc V Le. Mnasnet: Platform-aware neural architecture search for mobile. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2820-2828, 2019. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.661, + 0.469, + 0.702 + ], + "angle": 0, + "content": "[156] Shuhan Tan, Tushar Nagarajan, and Kristen Grauman. Egodistill: Egocentric head motion distillation for efficient video understanding. NeurIPS, 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.704, + 0.469, + 0.785 + ], + "angle": 0, + "content": "[157] Hao Tang, Dan Xu, Nicu Sebe, Yanzhi Wang, Jason J Corso, and Yan Yan. Multi-channel attention selection gan with cascaded semantic guidance for cross-view image translation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2417-2426, 2019. 4, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.789, + 0.469, + 0.843 + ], + "angle": 0, + "content": "[158] Hao Tang, Kevin Liang, Kristen Grauman, Matt Feiszli, and Weiyao Wang. Egotracks: A long-term egocentric visual object tracking dataset. Advances in Neural Information Processing Systems, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.846, + 0.469, + 0.9 + ], + "angle": 0, + "content": "[159] Yansong Tang, Jiwen Lu, and Jie Zhou. Comprehensive instructional video analysis: The coin dataset and performance evaluation. IEEE transactions on pattern analysis and machine intelligence, 2020. 2, 4, 7" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.469, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.093, + 0.892, + 0.134 + ], + "angle": 0, + "content": "[160] Zachary Teed and Jia Deng. DROID-SLAM: Deep Visual SLAM for Monocular, Stereo, and RGB-D Cameras. Advances in neural information processing systems, 2021. 57" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.137, + 0.892, + 0.191 + ], + "angle": 0, + "content": "[161] Denis Tome, Patrick Peluse, Lourdes Agapito, and Hernan Badino. xr-egopose: Egocentric 3d human pose from an hmd camera. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2019. 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.193, + 0.892, + 0.234 + ], + "angle": 0, + "content": "[162] Zhan Tong, Yibing Song, Jue Wang, and Limin Wang. Videomae: Masked autoencoders are data-efficient learners for self-supervised video pre-training. 2022. 43" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.236, + 0.892, + 0.305 + ], + "angle": 0, + "content": "[163] Vadim Tschernezki, Ahmad Darkhalil, Zhifan Zhu, David Fouhey, Iro Larina, Diane Larlus, Dima Damen, and Andrea Vedaldi. EPIC Fields: Marrying 3D Geometry and Video Understanding. In Proceedings of the Neural Information Processing Systems (NeurIPS), 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.307, + 0.892, + 0.36 + ], + "angle": 0, + "content": "[164] Hung-Yu Tseng, Qinbo Li, Changil Kim, Suhib Alsisan, Jia-Bin Huang, and Johannes Kopf. Consistent view synthesis with pose-guided diffusion models. arXiv preprint arXiv:2303.17598, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.363, + 0.892, + 0.432 + ], + "angle": 0, + "content": "[165] Pavan Kumar Anasosalu Vasu, James Gabriel, Jeff Zhu, Oncel Tuzel, and Anurag Ranjan. Mobileone: An improved one millisecond mobile backbone. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7907-7917, 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.435, + 0.892, + 0.474 + ], + "angle": 0, + "content": "[166] Sara Vicente, Carsten Rother, and Vladimir Kolmogorov. Object cosegmentation. In CVPR 2011, pages 2217-2224. IEEE, 2011. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.477, + 0.892, + 0.518 + ], + "angle": 0, + "content": "[167] Hanlin Wang, Yilu Wu, Sheng Guo, and Limin Wang. Pdpp: Projected diffusion for procedure planning in instructional videos. arXiv preprint arXiv:2303.14676, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.521, + 0.892, + 0.575 + ], + "angle": 0, + "content": "[168] Daniel Watson, William Chan, Ricardo Martin-Brualla, Jonathan Ho, Andrea Tagliasacchi, and Mohammad Norouzi. Novel view synthesis with diffusion models. arXiv preprint arXiv:2210.04628, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.577, + 0.892, + 0.631 + ], + "angle": 0, + "content": "[169] Daniel Weinland, Remi Ronfard, and Edmond Boyer. Free viewpoint action recognition using motion history volumes. Computer Vision and Image Understanding (CVIU), 2006. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.634, + 0.892, + 0.716 + ], + "angle": 0, + "content": "[170] Yangming Wen, Krishna Kumar Singh, Markham Anderson, Wei-Pang Jan, and Yong Jae Lee. Seeing the unseen: Predicting the first-person camera wearer's location and pose in third-person scenes. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) Workshops, pages 3446-3455, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.718, + 0.892, + 0.786 + ], + "angle": 0, + "content": "[171] Olivia Wiles, Georgia Gkioxari, Richard Szeliski, and Justin Johnson. Synsin: End-to-end view synthesis from a single image. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7467-7477, 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.789, + 0.892, + 0.857 + ], + "angle": 0, + "content": "[172] Benita Wong, Joya Chen, You Wu, Stan Weixian Lei, Dongxing Mao, Difei Gao, and Mike Zheng Shou. Assistq: Affordance-centric question-driven task completion for egocentric assistant. In European Conference on Computer Vision, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.86, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[173] Zhirong Wu, Shuran Song, Aditya Khosla, Fisher Yu, Linguang Zhang, Xiaou Tang, and Jianxiong Xiao. 3d shapenets: A deep representation for volumetric shapes. In" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19398" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.118, + 0.092, + 0.469, + 0.119 + ], + "angle": 0, + "content": "Computer Vision and Pattern Recognition, IEEE Conference on, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.121, + 0.469, + 0.175 + ], + "angle": 0, + "content": "[174] Zuxuan Wu, Tushar Nagarajan, Abhishek Kumar, Steven Rennie, Larry S Davis, Kristen Grauman, and Rogerio Feris. Blockdrop: Dynamic inference paths in residual networks. In CVPR, 2018. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.177, + 0.469, + 0.245 + ], + "angle": 0, + "content": "[175] Fei Xia, Amir R. Zamir, Zhi-Yang He, Alexander Sax, Jitendra Malik, and Silvio Savarese. Gibson Env: real-world perception for embodied agents. In CVPR. IEEE, 2018. Gibson license is available at http://svl.stanford.edu/gibson2/assets/GDS_agreement.pdf.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.247, + 0.469, + 0.286 + ], + "angle": 0, + "content": "[176] J. Xiao, A. Owens, and A. Torralba. Sun3d: A database of big spaces reconstructed using sfm and object labels. In ICCV, 2013. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.288, + 0.469, + 0.343 + ], + "angle": 0, + "content": "[177] Frank F Xu, Lei Ji, Botian Shi, Junyi Du, Graham Neubig, Yonatan Bisk, and Nan Duan. A benchmark for structured procedural knowledge extraction from cooking videos. arXiv preprint arXiv:2005.00706, 2020. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.344, + 0.469, + 0.412 + ], + "angle": 0, + "content": "[178] Hu Xu, Gargi Ghosh, Po-Yao Huang, Dmytro Okhonko, Armen Aghajanyan, Florian Metze, Luke Zettlemoyer, and Christoph Feichtenhofer. Videoclip: Contrastive pretraining for zero-shot video-text understanding. arXiv preprint arXiv:2109.14084, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.414, + 0.469, + 0.467 + ], + "angle": 0, + "content": "[179] Mingze Xu, Chenyou Fan, Yuchen Wang, Michael S Ryoo, and David J Crandall. Joint person segmentation and identification in synchronized first-and third-person videos. In ECCV, 2018. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.469, + 0.469, + 0.524 + ], + "angle": 0, + "content": "[180] Mingze Xu, Yuanjun Xiong, Hao Chen, Xinyu Li, Wei Xia, Zhuowen Tu, and Stefano Soatto. Long short-term transformer for online action detection. Advances in Neural Information Processing Systems, 34:1086-1099, 2021. 49" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.525, + 0.469, + 0.607 + ], + "angle": 0, + "content": "[181] Weipeng Xu, Avishek Chatterjee, Michael Zollhoefer, Helge Rhodin, Pascal Fua, Hans-Peter Seidel, and Christian Theobalt. Mo 2 cap 2: Real-time mobile 3d motion capture with a cap-mounted fisheye camera. IEEE transactions on visualization and computer graphics, 25(5):2093-2101, 2019. 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.608, + 0.469, + 0.661 + ], + "angle": 0, + "content": "[182] Zihui Xue and Kristen Grauman. Learning fine-grained view-invariant representations from unpaired ego-exo videos via temporal alignment. In NeurIPS, 2023. 4, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.664, + 0.469, + 0.732 + ], + "angle": 0, + "content": "[183] Lita Yang, Robert M Radway, Yu-Hsin Chen, Tony F Wu, Huichu Liu, Elnaz Ansari, Vikas Chandra, Subhasish Mitra, and Edith Beigné. Three-dimensional stacked neural network accelerator architectures for ar/vr applications. IEEE Micro, 2022. 46" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.734, + 0.469, + 0.774 + ], + "angle": 0, + "content": "[184] Huangyue Yu, Minjie Cai, Yunfei Liu, and Feng Lu. What i see is what you see: Joint attention learning for first and third person video co-analysis. In ACM MM, 2019. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.776, + 0.469, + 0.831 + ], + "angle": 0, + "content": "[185] Huangyue Yu, Minjie Cai, Yunfei Liu, and Feng Lu. First- and third-person video co-analysis by learning spatial-temporal joint attention. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.832, + 0.469, + 0.899 + ], + "angle": 0, + "content": "[186] Xumin Yu, Yongming Rao, Wenliang Zhao, Jiwen Lu, and Jie Zhou. Group-aware contrastive regression for action quality assessment. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 7919-7928, 2021. 8" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.469, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.134 + ], + "angle": 0, + "content": "[187] Ye Yuan and Kris Kitani. 3d ego-pose estimation via imitation learning. In Proceedings of the European Conference on Computer Vision (ECCV), 2018. 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.136, + 0.892, + 0.19 + ], + "angle": 0, + "content": "[188] Ye Yuan and Kris Kitani. Ego-pose estimation and forecasting as real-time pd control. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2019. 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.193, + 0.892, + 0.234 + ], + "angle": 0, + "content": "[189] Chenlin Zhang, Jianxin Wu, and Yin Li. Actionformer: Localizing moments of actions with transformers. arXiv preprint arXiv:2202.07925, 2022. 21" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.236, + 0.892, + 0.289 + ], + "angle": 0, + "content": "[190] Chen-Lin Zhang, Jianxin Wu, and Yin Li. Actionformer: Localizing moments of actions with transformers. In European Conference on Computer Vision, pages 492-510, 2022. 54, 55, 57" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.293, + 0.892, + 0.347 + ], + "angle": 0, + "content": "[191] Qiang Zhang and Baoxin Li. Relative hidden markov models for evaluating motion skill. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2013. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.349, + 0.892, + 0.39 + ], + "angle": 0, + "content": "[192] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, 2018. 39" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.392, + 0.892, + 0.447 + ], + "angle": 0, + "content": "[193] Siwei Zhang, Qianli Ma, Yan Zhang, Zhiyin Qian, Taein Kwon, Marc Pollefeys, Federica Bogo, and Siyu Tang. Egobody: Human body shape and motion of interacting people from head-mounted devices. In ECCV, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.449, + 0.892, + 0.503 + ], + "angle": 0, + "content": "[194] Shiyi Zhang, Wenxun Dai, Sujia Wang, Xiangwei Shen, Jiwen Lu, Jie Zhou, and Yansong Tang. Logo: A long-form video dataset for group action quality assessment. In CVPR, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.506, + 0.892, + 0.573 + ], + "angle": 0, + "content": "[195] Xiangyu Zhang, Xinyu Zhou, Mengxiao Lin, and Jian Sun. Shufflenet: An extremely efficient convolutional neural network for mobile devices. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6848-6856, 2018. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.576, + 0.892, + 0.658 + ], + "angle": 0, + "content": "[196] He Zhao, Isma Hadji, Nikita Dvornik, Konstantinos G Derpanis, Richard P Wildes, and Allan D Jepson. P3iv: Probabilistic procedure planning from instructional videos with weak supervision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2938-2948, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.661, + 0.892, + 0.716 + ], + "angle": 0, + "content": "[197] Weixi Zhao, Weiqiang Wang, and Yunjie Tian. Graformer: Graph-oriented transformer for 3d pose estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20438-20447, 2022. 58" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.718, + 0.892, + 0.759 + ], + "angle": 0, + "content": "[198] Yue Zhao and Philipp Krahenbuhl. Real-time online video detection with temporal smoothing transformers. In European Conference on Computer Vision, 2022. 49" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.761, + 0.892, + 0.829 + ], + "angle": 0, + "content": "[199] Yue Zhao, Ishan Misra, Philipp Krahenbuhl, and Rohit Girdhar. Learning video representations from large language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6586-6597, 2023. 26, 46, 47, 48" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.832, + 0.892, + 0.899 + ], + "angle": 0, + "content": "[200] Ce Zheng, Xianpeng Liu, Guo-Jun Qi, and Chen Chen. Potter: Pooling attention transformer for efficient human mesh recovery. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1611-1620, 2023. 59" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "19399" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.147 + ], + "angle": 0, + "content": "[201] Yiwu Zhong, Licheng Yu, Yang Bai, Shangwen Li, Xueling Yan, and Yin Li. Learning procedure-aware video representation from instructional videos and their narrations. arXiv preprint arXiv:2303.17839, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.149, + 0.47, + 0.219 + ], + "angle": 0, + "content": "[202] Honglu Zhou, Roberto Martin-Martin, Mubbasir Kapadia, Silvio Savarese, and Juan Carlos Niebles. Procedure-aware pretraining for instructional video understanding. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023. 4, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.22, + 0.469, + 0.275 + ], + "angle": 0, + "content": "[203] Honglu Zhou, Roberto Martin-Martin, Mubbasir Kapadia, Silvio Savarese, and Juan Carlos Niebles. Procedure-aware pretraining for instructional video understanding. In CVPR, pages 10727-10738, 2023. 49" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.277, + 0.469, + 0.317 + ], + "angle": 0, + "content": "[204] L. Zhou, N. Louis, and J. Corso. Weakly-supervised video object grounding from text by loss weighting and object interaction. In BMVC, 2018. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.319, + 0.469, + 0.359 + ], + "angle": 0, + "content": "[205] Luowei Zhou, Chenliang Xu, and Jason J Corso. Towards automatic learning of procedures from web instructional videos. In AAAI, 2018. 4, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.362, + 0.468, + 0.403 + ], + "angle": 0, + "content": "[206] Michael Zhu and Suyog Gupta. To prune, or not to prune: exploring the efficacy of pruning for model compression. arXiv preprint arXiv:1710.01878, 2017. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.405, + 0.469, + 0.473 + ], + "angle": 0, + "content": "[207] Dimitri Zhukov, Jean-Baptiste Alayrac, Ramazan Gokberk Cinbis, David Fouhey, Ivan Laptev, and Josef Sivic. Cross-task weakly supervised learning from instructional videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2019. 2, 4, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.476, + 0.469, + 0.529 + ], + "angle": 0, + "content": "[208] Aneeq Zia, Yachna Sharma, Vinay Bettadapura, Eric L. Sarin, and Irfan A. Essa. Video and accelerometer-based motion analysis for automated surgical skills assessment. CoRR, abs/1702.07772, 2017. 8" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.529 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.956 + ], + "angle": 0, + "content": "19400" + } + ] +] \ No newline at end of file diff --git a/2024/Ego-Exo4D_ Understanding Skilled Human Activity from First- and Third-Person Perspectives/742959dc-4498-434e-af9a-b1cc7f6e5bd2_origin.pdf b/2024/Ego-Exo4D_ Understanding Skilled Human Activity from First- and Third-Person Perspectives/742959dc-4498-434e-af9a-b1cc7f6e5bd2_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..63f9822e554934e38a541ddfd1aac47e623f05b9 --- /dev/null +++ b/2024/Ego-Exo4D_ Understanding Skilled Human Activity from First- and Third-Person Perspectives/742959dc-4498-434e-af9a-b1cc7f6e5bd2_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:036aa74f937612b01f1e0fcc411fccea58d5c82b3aff9d54f33b669329b6eb74 +size 7554745 diff --git a/2024/Ego-Exo4D_ Understanding Skilled Human Activity from First- and Third-Person Perspectives/full.md b/2024/Ego-Exo4D_ Understanding Skilled Human Activity from First- and Third-Person Perspectives/full.md new file mode 100644 index 0000000000000000000000000000000000000000..e1436b4d765a0a9f38cb0cc3a200a15947a087cb --- /dev/null +++ b/2024/Ego-Exo4D_ Understanding Skilled Human Activity from First- and Third-Person Perspectives/full.md @@ -0,0 +1,563 @@ +# Ego-Exo4D: Understanding Skilled Human Activity from First- and Third-Person Perspectives + +Kristen Grauman $^{1,2}$ , Andrew Westbury $^{1}$ , Lorenzo Torresani $^{1}$ , Kris Kitani $^{1,3}$ , Jitendra Malik $^{1,4}$ , Triantafyllos Afouras $^{*1}$ , Kumar Ashutosh $^{*1,2}$ , Vijay Baiyya $^{5}$ , Siddhant Bansal $^{6,7}$ , Bikram Boote $^{8}$ , Eugene Byrne $^{*1,9}$ , Zach Chavis $^{*10}$ , Joya Chen $^{*11}$ , Feng Cheng $^{1}$ , Fu-Jen Chu $^{1}$ , Sean Crane $^{9}$ , Avijit Dasgupta $^{7}$ , Jing Dong $^{5}$ , Maria Escobar $^{*12}$ , Cristian Forigua $^{*12}$ , Abrham Gebreselasie $^{9}$ , Sanjay Haresh $^{13}$ , Jing Huang $^{1}$ , Md Mohaiminul Islam $^{14}$ , Suyog Jain $^{1}$ , Rawal Khirodkar $^{9}$ , Devansh Kukreja $^{1}$ , Kevin J Liang $^{1}$ , Jia-Wei Liu $^{11}$ , Sagnik Majumder $^{1,2}$ , Yongsen Mao $^{13}$ , Miguel Martin $^{1}$ , Effrosyni Mavroudi $^{1}$ , Tushar Nagarajan $^{1}$ , Francesco Ragusa $^{15}$ , Santhosh Kumar Ramakrishnan $^{2}$ , Luigi Seminara $^{15}$ , Arjun Somayazulu $^{2}$ , Yale Song $^{1}$ , Shan Su $^{16}$ , Zihui Xue $^{1,2}$ , Edward Zhang $^{16}$ , Jinxu Zhang $^{16}$ , Angela Castillo $^{12}$ , Changan Chen $^{2}$ , Xinzhu Fu $^{11}$ , Ryosuke Furuta $^{17}$ , Cristina González $^{12}$ , Prince Gupta $^{5}$ , Jiabo Hu $^{18}$ , Yifei Huang $^{17}$ , Yiming Huang $^{16}$ , Weslie Khoo $^{19}$ , Anush Kumar $^{10}$ , Robert Kuo $^{18}$ , Sach Lakhavani $^{5}$ , Miao Liu $^{18}$ , Mi Luo $^{2}$ , Zhengyi Luo $^{3}$ , Brighid Meredith $^{18}$ , Austin Miller $^{18}$ , Oluwatuminiu Oguntola $^{14}$ , Xiaqing Pan $^{5}$ , Penny Peng $^{18}$ , Shraman Pramanick $^{20}$ , Merey Ramazanova $^{21}$ , Fiona Ryan $^{22}$ , Wei Shan $^{14}$ , Kiran Somasundaram $^{5}$ , Chenan Song $^{11}$ , Audrey Southerland $^{22}$ , Masatoshi Tateno $^{17}$ , Huiyu Wang $^{1}$ , Yuchen Wang $^{19}$ , Takuma Yagi $^{17}$ , Mingfei Yan $^{5}$ , Xitong Yang $^{1}$ , Zecheng Yu $^{17}$ , Shengxin Cindy Zha $^{18}$ , Chen Zhao $^{21}$ , Ziwei Zhao $^{19}$ , Zhifan Zhu $^{6}$ , Jeff Zhuo $^{14}$ , Pablo Arbeláez† $^{12}$ , Gedas Bertasius† $^{14}$ , Dima Damen† $^{6}$ , Jakob Engel† $^{5}$ , Giovanni Maria Farinella† $^{15}$ , Antonino Furnari† $^{15}$ , Bernard Ghanem† $^{21}$ , Judy Hoffman† $^{22}$ , C. V. Jawahar† $^{7}$ , Richard Newcombe† $^{5}$ , Hyun Soo Park† $^{10}$ , James M. Rehg† $^{8}$ , Yoichi Sato† $^{17}$ , Manolis Savva† $^{13}$ , Jianbo Shi† $^{16}$ , Mike Zheng Shou† $^{11}$ and Michael Wray† $^{6}$ + +$^{1}$ FAIR, Meta, $^{2}$ University of Texas at Austin, $^{3}$ Carnegie Mellon University, $^{4}$ University of California, Berkeley, $^{5}$ Project Aria, Meta, $^{6}$ University of Bristol, $^{7}$ International Institute of Information Technology, Hyderabad, $^{8}$ University of Illinois, Urbana Champaign, $^{9}$ Carnegie Mellon University, $^{10}$ University of Minnesota, $^{11}$ National University of Singapore, $^{12}$ Universidad de los Andes, $^{13}$ Simon Fraser University, $^{14}$ University of North Carolina, Chapel Hill, $^{15}$ University of Catania, $^{16}$ University of Pennsylvania, $^{17}$ University of Tokyo, $^{18}$ Meta, $^{19}$ Indiana University, $^{20}$ Johns Hopkins University, $^{21}$ King Abdullah University of Science and Technology, $^{22}$ Georgia Tech + +# Abstract + +We present Ego-Exo4D, a diverse, large-scale multimodal multiview video dataset and benchmark challenge. Ego-Exo4D centers around simultaneously-captured ego-centric and exocentric video of skilled human activities (e.g., sports, music, dance, bike repair). 740 participants from 13 cities worldwide performed these activities in 123 different natural scene contexts, yielding long-form captures from 1 to 42 minutes each and 1,286 hours of video combined. The multimodal nature of the dataset is unprecedented: the video is accompanied by multichannel audio, eye gaze, 3D point clouds, camera poses, IMU, and multiple paired language descriptions—including a novel "expert commentary" done by coaches and teachers and tailored to the skilled-activity domain. To push the frontier of first-person video understanding of skilled hu + +man activity, we also present a suite of benchmark tasks and their annotations, including fine-grained activity understanding, proficiency estimation, cross-view translation, and 3D hand/body pose. All resources are open sourced to fuel new research in the community. + +# 1. Introduction + +A dancer leaps across a stage; Lionel Messi delivers a precise pass; your grandmother prepares her famous dumplings. We observe and seek human skills in a myriad of settings, from the practical (fixing a bike) to the aspirational (dancing beautifully). What would it mean for AI to understand human skills? And what would it take to get there? + +![](images/fac5d8236d3d229f991c89ab4fb2140794f3f9860ab60fe5efe2e1344f38e59f.jpg) +Figure 1. Ego-Exo4D offers egocentric video alongside multiple time-synchronized exocentric video streams for an array of skilled human activities—1,286 hours of ego and exo video in total. The data is both multiview and multimodal, and it is extensively annotated with language, 3D body and hand pose, keysteps, procedural dependencies, and proficiency ratings in support of our proposed benchmark tasks. + +Advances in AI understanding of human skill could facilitate many applications. In augmented reality (AR), a person wearing smart glasses could quickly pick up new skills with a virtual AI coach that provides real-time guidance. In robot learning, a robot watching people in its environment could acquire new dexterous manipulation skills with less physical experience. In social networks, new communities could form based on how people share their expertise and complementary skills in video. + +We contend that both the egocentric and exocentric viewpoints are critical for capturing human skill. Firstly, the two viewpoints are synergistic. The first-person (ego) perspective captures the details of close-by hand-object interactions and the camera wearer's attention, whereas the third-person (exo) perspective captures the full body pose and surrounding environment context. See Figure 1. Not coincidentally, instructional or "how-to" videos often alternate between a third-person view of the demonstrator and a close-up view of their near-field demonstration. For example, a chef may describe their approach and the equipment from an exo view, then cut to clips showing their hands manipulating the ingredients and tools from an ego-like view. + +Secondly, not only are the ego and exo viewpoints synergistic, but there is a need to translate fluently from one to the other when acquiring skill. For example, imagine watching an expert repair a bike tire, juggle a soccer ball, + +or fold an origami swan—then mapping their steps to your own body. Cognitive science tells us that even from a very young age we can observe others' behavior (exo) and map it onto our own (ego) [42, 108], and this actor-observer translation remains the foundation of visual learning. + +Realizing this potential, however, is not possible using today's datasets and learning paradigms. Existing datasets comprised of both ego and exo views (i.e., ego-exo) are few [76, 77, 127, 139, 145], small in scale, lack synchronization across cameras, and/or are too staged or curated to be resilient to the diversity of the real world. Thus the current literature for activity understanding primarily attends to either the ego [28, 47] or exo [48, 67, 105, 149] view, leaving the ability to move fluidly between the first- and third-person perspectives out of reach. Instructional video datasets [103, 159, 204, 207] offer a compelling window into skilled human activity, but (like the above) are limited to single-viewpoint video, whether purely exocentric or mixed with "ego-like" views at certain time points. + +We introduce Ego-Exo4D, a foundational dataset to support research on ego-exo video learning and multimodal perception. The result of a two-year effort by a consortium of 15 research institutions, Ego-Exo4D is a first-of-its-kind large-scale multimodal multiview dataset and benchmark suite. It constitutes the largest public dataset of timesynchronized first- and third- person video, captured by 740 + +diverse camera wearers in 123 distinct scenes and 13 cities worldwide. For every sequence, Ego-Exo4D provides both the camera wearer's egocentric video, as well as multiple (4-5) exocentric videos from tripod places around the camera wearer. All views are time-synchronized and precisely localized in a metric, gravity-aligned frame of reference. The total collection has 1,286 hours of video and 5,035 instances, each spanning 1 to $42\mathrm{min}$ of continuous capture. + +Ego-Exo4D focuses on skilled single-person activities. The 740 participants perform skilled physical and/or procedural activities—dance, soccer, basketball, bouldering, music, cooking, bike repair, health care—in an unscripted manner and in natural settings (e.g., gym, soccer field, kitchens, bike shops, etc.), exhibiting a variety of skill levels from novice to expert. All video is recorded with rigorous privacy and ethics policies and formal consent of participants. + +Ego-Exo4D is not only multiview, it is also multimodal. Captured with the unique open-source Aria glasses [38], all ego video is accompanied by 7-channel audio, IMU, eye gaze, both RGB and two grayscale SLAM cameras, and 3D environment point clouds. Additionally, Ego-Exo4D provides multiple new video-language resources, all time indexed: first-person narrations by the camera wearers describing their own actions; third-person play-by-play descriptions of every camera wearer action; and third-person spoken expert commentary critiquing their performance. The latter is particularly novel: performed by domain-specific experienced coaches and teachers, it focuses on how an activity is executed rather than merely what is being done, surfacing subtleties in skilled execution not perceivable by the untrained eye. To our knowledge, there is no prior video resource with such extensive and high quality multimodal data. + +Alongside this data, we introduce benchmarks for foundational tasks for ego-exo video. We propose four families of tasks: 1) ego-exo relation, for relating the actions of a teacher (exo) to a learner (ego) by estimating semantic correspondences and translating viewpoints; 2) ego(-exo) recognition, for recognizing fine-grained keysteps and task structure; 3) ego(-exo) proficiency estimation, for inferring how well a person is executing a skill; and 4) ego pose, for recovering skilled 3D body and hand movements from ego-video. We provide annotations for each task—the result of more than 200,000 hours of annotator effort. To kick-start work in these new challenges, we also develop baseline models and report their results (Appendices). We are hosting the first public benchmark challenges in 2024. + +In summary, Ego-Exo4D is the community's first diverse, large-scale multimodal multiview video resource. We have open sourced all the data, annotations, camera rig protocol, and benchmarks. With this release, we aim to fuel new research in ego-exo, multimodal activity, and beyond. + +# 2. Related work + +Next we review prior work in datasets, human skill, and cross-view analysis. Section 5 will discuss related work for each benchmark task. Table 2 in Appendix 9 summarizes Ego-Exo4D's properties vs. existing datasets. + +Egocentric datasets There has been a surge of interest in egocentric video understanding, facilitated by recent egovideo datasets showing unscripted daily-life activity as in Ego4D [47], EPIC-Kitchens [27, 28, 163], UT Ego [78], ADL [119], and KrishnaCam [147], or procedural activities as in EGTea [81], AssistQ [172], Meccano [126], CMUMMAC [77], and EgoProcel [10]. Unlike any of the above, Ego-Exo4D focuses on multimodal ego and exo capture, and it is focused on the domain of skilled activities. + +Multiview and ego-exo datasets Most existing multiview datasets focus on static scenes [20, 128, 151, 175, 176] and objects [133, 173], with limited (exo only) multiview human activity [26, 169]. CMU-MMAC [77] and CharadesEgo [145] are early efforts to capture both ego and exo video. CMU-MMAC [77] features 43 participants in mocap suits who cook 5 recipes in a lab kitchen. In CharadesEgo [145], 71 Mechanical Turkers record 34 hours of scripted scenarios (e.g., "type on laptop, then pick up a pillow") from the ego and exo perspectives sequentially, yielding unsynchronized videos with non-exact activity matches. More recent ego-exo efforts focus on specific activities in one or two environments. Assembly101 [139] and H2O [76] provide time-synced ego and exo video at a lab tabletop where people assemble toy cars or manipulate handheld objects, with 53 and 4 participants, and 513 and 5 hours of footage, respectively. Homage [127] provides 30 hours of ego-exo video from 27 participants in 2 homes doing household activities like laundry. + +Compared to any of the prior efforts, Ego-Exo4D offers an order of magnitude more participants, diverse locations, and hours of footage (740 participants, 123 unique scenes, 13 cities, 1,286 hours). Importantly, our focus on skilled tasks takes the participants out of the lab or home and into settings like soccer fields, dance studios, rock climbing walls, and bike repair shops. Such activities also yield a wide variety of full body poses and movements within the scene, beyond using objects at a tabletop. This variety means Ego-Exo4D augments existing 3D human body pose datasets [49, 66, 68, 80, 193]. Finally, compared to any prior ego-exo resource, Ego-Exo4D's suite of modalities and benchmark tasks are novel and will expand the research directions the community can take for egocentric and/or exocentric video understanding. + +Human skill and video learning Analyzing skill and action quality has received limited attention [12, 34, 35, 113, 120, 194]. Research in instructional or "how-to" videos is + +![](images/8e9f0c174b11764e0f15c304b5c22610240f5992f86298476a8de14579f1d267.jpg) +Cooking + +![](images/eebc1e0bf3ece0ea41c10a5452a90876a509030e552ddf750acc4e667d57939d.jpg) +x678 +x173 +8 +Mx65 +$\mathbb{O}\times 564\mathrm{h}$ + +![](images/a19d07b019f38fa7c8c76ad4a0301f9e336f516ac051e63806116e8155f497a2.jpg) +Music +x59 +# +# + +![](images/984ff27e32171872ea2ada18f5aedee16cc6fa00b95998cfa31845245dfb8c80.jpg) +x276 +L x 18Oh + +![](images/5a380045c4d5459df5d074c6037f9d8118652d1a20cf8e4e0f490735a0b03331.jpg) +Soccer +x282 +x78 +# +X14 +$\bullet \times 66h$ + +![](images/9a5e64d3c71b41b6225c4cef72f3bda9fcdcf5c474614aae5d0828f21ad7bd0c.jpg) +$\therefore m = \frac{3}{11}$ ; +Soccer in Pittsburgh + +![](images/c7aa36c3ac9b6150a2bd3f89c2eff2a7cd3f143b3eb72f3be85b17932dac5950.jpg) +Health +9x24 + +![](images/693afb2f8315a87b6db52b87cf9b6d450f5b8030e72417b8027d01539b8e2897.jpg) +Bike Repair +4x363 +x32 +# +图 + +![](images/bb9b9912947665d957dd000e5903a3a3fca5ab9d4d9ad8bfc1c1c303d813f372.jpg) +Making Chai in Hyderabad +x397 +x122 +中 +X141 + +![](images/0d8c3e0cc3b5f0e78255b64b060f0ba6fceaaf039156df74327dc10012f83a30.jpg) +X 82n +Removing a Wheel in Menlo Park +Figure 2. Ego-Exo4D captures skilled activity from 43 tasks and 689 keysteps in 8 domains, in a wide variety of 123 scenes in 13 cities in Japan, Colombia, Canada, India, Singapore, and 7 US states. Each domain is captured at multiple sites—from 2 to 64 unique locations. In total the dataset offers 1,286 hours of ego+exo video comprised of 5,035 takes from 740 camera wearers. An average take is 2.6 minutes. + +![](images/2d79df213b3cc867df1653fb63954d69806914bb972f99f01dd47e5d2f31cf51.jpg) +Basketball +x910 +x113 +9x5 +# + +![](images/8f5812255672053bbc201eee5ab4ee80e671699df439d8f665a728e94586e047.jpg) +Rock Climbing + +![](images/b3262b0b991a281b3481f1053e74925e8ce35455330367f75db85b66f831caf6.jpg) +5x78n + +![](images/cd72a671a831712395fef38e7bb491b1d1d9d8048b0ef6a9eecb25ba32e7671d.jpg) +Piano Playing in Bloomington + +![](images/28d54ba7c8516fa7d9299162bcbc4485a70940f8fbeaa3f9ac89ebd39d80f416.jpg) +Bouldering in Minneapolis +Total + +![](images/35e7474e239096906675b0aa12b24e7ad316f45ac7b0d3da3354e281c1610354.jpg) +Dance +9x7 + +![](images/ef05d05169d1820987b79dcef743525e471d31daa226c8ce72abaeb09ed032e4.jpg) +4x728 +x93 +# +x106n +$\therefore m = \frac{3}{11}$ ; +Salsa in Bogota +4x5035 (#takes) +x740 (#participants) +9. x 123 (# sites) +x 1286h (#ego+exo hours) + +facilitated by (largely exo) datasets like HowTo100M [103] and others [11, 159, 204, 207]. Challenges include grounding keysteps [10, 36, 37, 89, 103, 104, 178, 207], procedural planning [15, 17, 22, 71, 143, 167, 196, 201], learning task structure [4, 9, 37, 107, 202, 205], and leveraging noisy narrations [89, 103, 104]. A portion of Ego-Exo4D is procedural activities, but unlike the above, it offers simultaneous ego-exo capture. The scale and diversity of our data—including its three forms of language descriptions—widen the avenues for skilled activity understanding research. + +Ego-exo cross-view modeling There is limited prior work on ego-exo cross-view modeling, arguably due to a lack of high-quality synchronized real-world data. Prior work explores matching people between videos [5, 6, 40, 170, 179] and learning view-invariant [7, 141, 144, 182, 184, 185] or ego features [82]. Beyond the specific case of ego-exo, cross-view methods are explored for translation [130, 131, 134, 157], novel view synthesis [19, 90, 135, 137, 164, 168, 171], and aerial to ground matching [86, 132]. Ego-Exo4D provides a testbed of unprecedented size and variety for cross-view modeling. In addition, our ego-exo relation tasks (cf. Section 5) surface new challenges in novel-view synthesis with widely varying viewpoints. + +# 3. Ego-Exo4D dataset + +Next we introduce the dataset and its scope. Notably, the video capture was a distributed but coordinated effort performed by 12 research labs. We present the common framework, and reserve site-specific details for Appendix 10. + +# 3.1. Ego-exo camera rig + +Our goal is to capture simultaneous ego and exo video, together with multiple egocentric sensing modalities. One of our contributions is to create and share a low-cost (less than + +\(3,000), lightweight ego-exo rig with a user-friendly calibration and time sync procedure. + +Our camera configuration features Aria glasses [38] for ego capture, leveraging their rich array of sensors, including an 8 MP RGB camera, two SLAM cameras, IMU, 7 microphones, and eye tracking (see Appendix 7). The ego camera is calibrated and time-synchronized with four to five (stationary) GoPros placed on tripods as the exo capture devices, allowing 3D reconstruction of the environment point clouds and the participant's body pose. The number and placement of the exocentric cameras is determined per scenario in order to allow maximal coverage of useful viewpoints without obstructing the participants' activity. + +Our time sync and calibration design relies on a QR-code procedure to auto-sync the cameras and auto-separate the individual "takes", meaning instances of an activity. We can do continuous recordings of up to $\sim 60$ minutes, based on the Aria battery life. See Appendix 8 for more details. + +# 3.2. Domains and environments + +Ego-Exo4D focuses on skilled human activity. This is in contrast to existing ego-only efforts like Ego4D [47], which has a broad span of daily-life activities. We intentionally select the domains based on a few criteria: Will it illustrate skill and a variety of expertise? Is there visual variety to be expected across different instances? Will the ego and exo views offer complementary information? Will it present new challenges unaddressed by current datasets? + +Intersecting these criteria, we arrived at two broad categories1 of skilled activity: physical and procedural, together comprising eight total domains. The physical domains are soccer, basketball, dance, bouldering, and music. They emphasize body pose and movements as well as + +interaction with objects (e.g., a ball, musical instrument). The procedural domains are cooking, bike repair, and health care. They require performing a sequence of steps to reach a goal state (e.g., a completed recipe, a repaired bike) and generally entail intricate hand-object manipulations with a variety of objects (e.g., bike repair tools; cooking utensils, appliances, and ingredients). + +In total, we have 43 activities derived from the eight domains (see Appendix 9). For example, cooking is comprised of 14 recipes; soccer is comprised of 3 drills. The length of a take ranges from 8 sec to $42\mathrm{min}$ , with procedural activities like cooking having the longest sustained captures. + +To achieve visual diversity in the data, multiple labs across our team (typically 3-5) captured each Ego-Exo4D domain. The data is collected in authentic settings—such as real-world bike shops, soccer pitches, or bouldering gyms—as opposed to lab environments. For example, we have videos of chefs in New York City, Vancouver, Philadelphia, Bogota, and others; soccer players in Tokyo, Chapel Hill, Hyderabad, Singapore, and Pittsburgh. See Figure 2. + +# 3.3. Participants: expertise and diversity + +We recruited 740 total participants from the local communities of 12 labs. All scenarios feature real-world experts, where the camera-wearer participant has specific credentials, training, or expertise in the skill being demonstrated. For example, among the Ego-Exo4D camera wearers are professional and college athletes; jazz, salsa, and Chinese folk dancers and instructors; competitive boulders; professional chefs who work in industrial-scale kitchens; bike technicians who service dozens of bikes per day. Many of them have (individually) over 10 years of experience. + +Experts are prioritized given they are likely to conduct activities without mistakes or distractions, providing a strong ground truth for how to approach a given task. However, we also include capture from people with varying skill levels, as well—essential for our proposed skill proficiency estimation task (Section 5). Notably, Ego-Exo4D represents human intelligence in a new way by capturing domain-specific expertise—both in the video as well as the accompanying expert commentary (see Section 4)—portraying the evolution of a skill from beginners to experts. + +According to the participant surveys (Appendix 11), the camera wearers range in age from 18 to 74 years old, with $37\%$ self-identifying as female $60\%$ male and $3\%$ as nonbinary or preferring not to say. In total, the participants self report more than 24 different ethnicities. + +# 3.4. Privacy and ethics + +Ego-Exo4D was collected following rigorous privacy and ethics standards. This included undergoing formal independent review processes at each institution to establish the standards for collection, management, and informed con + +![](images/23c2068815e1aebf6b65f96acf44b75df4aa036595e21ce72aa799e712f3f696.jpg) +Expert commentary +Atomic action descriptions +Narrate and act +Figure 3. Ego-Exo4D offers 3 paired language corpora. Word cloud is from expert commentary which critiques the performance. + +"The dancer's hand is rotated inwardly a bit. Her palm should be facing to the ground, ..." + +"C steps forward while shimmering her shoulders." + +"Here we go, left forward, right back for one phrase 1234..." + +![](images/090ac63719d41de68d3ac319ba1aaf42222509b8103f6d3c584ce5f72c4b69d7.jpg) + +sent. Similarly, all Ego-Exo4D data collection adhered to the Project Aria Research Community Guidelines for responsible research. Since the scenarios allow for closed environments (e.g., no passerbys) nearly all video is available without de-identification. For information about each individual partners' protocols and restrictions, please see Appendix 10. Ego-Exo4D data is gated behind a license system, which defines permitted uses, restrictions, and consequences for non-compliance. + +# 4. Natural language descriptions + +Ego-Exo4D also offers three kinds of paired natural language datasets, each time-indexed alongside the video. See Figure 3. These language annotations are not steered towards any single benchmark, but rather are a general resource that will support browsing and mining the dataset—as well as challenges in video-language learning like grounding actions and objects, self-supervised representation learning, video-conditioned language models, and skill assessment. See Appendix 12. + +The first language dataset is spoken expert commentary. The goal is to reveal nuances of the skill that are not always visible to non-experts. We recruited 52 experts (distinct from the participants) to critique the recorded videos, call out strengths and weaknesses, explain how the specific behavior of the participant (e.g., hand/body pose, use of objects) affects the performance, and provide spatial markings to support their commentary. The experts are not only well-credentialed in their areas of expertise, but also have coaching or teaching experience, which facilitates clear communication. They watch the video and pause every time they have a comment, typically 7 times per minute of video. Each piece of commentary is unbounded in length, and averages 4 sentences. We provide both the transcribed speech and the raw audio (interesting for its inflection and nonword utterances), as well as the experts' spatial drawings and numeric ratings of each participant's skill. Videos have expert commentary by 2-5 distinct experts, offering a variety of perspectives for the same content. In total, we have 117,812 pieces of time-stamped, video-aligned commentary. These commentaries are quite novel: they focus on how the activity is executed rather than what it entails, + +![](images/89e99b3d4d421fb985f90b50f51421249d01d29f4ac88415876f8ae480461290.jpg) +Figure 4. The ego-exo relation family consists of the tasks of correspondence (left) and translation (right). + +![](images/c32b41e20a7b16c60de14b52977b7500e1565171dc15dbf4cd0aa43a8c26dc8e.jpg) + +capturing subtle differences in skilled execution. We believe this can unlock new fundamental problems (e.g., proficiency estimation below) and disruptive future applications (e.g., AI coaching). + +The second language dataset consists of narrate-and-act descriptions provided by the participants themselves. They are in the style of a tutorial or how-to video, where the participant explains what they are doing and why. Unlike the third-party expert commentary above, these are first-person reflections on the activity given by the people doing them. These narrations are available for about $10\%$ of all takes in the dataset, since we wanted participants to execute the tasks without pausing for the bulk of the recordings. + +The third language dataset consists of atomic action descriptions. Whereas the commentary and narrate-and-act language reveals spoken opinions and reasons for the actions (the "why and how"), this stream of text is specifically about the "what". Inspired by Ego4D's narrations [47], these are short statements written by third-party (nondomain expert) annotators, timestamped for every atomic action performed by the participant for all videos in the dataset, for a total of 432K sentences. This data is valuable for mining for taxonomies of objects and actions in the data, indexing the videos with keywords for exploring the dataset, and for future research in video-language learning, as has been quite successful for the Ego4D narrations [8, 85, 123]. + +# 5. Ego-Exo4D benchmark tasks + +Our second major contribution is to define the core research challenges in the domain of egocentric perception of skilled activity, particularly when ego-exo data is available for training (if not testing). To that end, we devise a suite of foundational benchmark tasks organized into four task families: relation (Sec. 5.1), recognition (Sec. 5.2), proficiency (Sec. 5.3), and ego-pose (Sec. 5.4). For each task, we provide high quality annotations and baselines that provide a starting point from which the research community can build. We will run the first formal Ego-Exo4D challenges in 2024. Due to space limits, we briefly overview + +each task; see the referenced Appendices for all details including baseline models and results. There are two publicly released versions of Ego-Exo4D annotations: v1 is used to train/test baselines in this paper; the larger v2 will be used for future challenge leaderboards (see Table 7 in Appendix). + +# 5.1. Ego-exo relation + +Our ego-exo relation tasks deal with relating the video content across the extreme ego-exo viewpoint changes. They take the form of object-level matching (correspondence) and synthesis of one view from the other (translation). + +# 5.1.1 Ego-exo correspondence + +Motivation. Establishing object-level correspondences between ego and exo viewpoints would allow AI assistants to provide visual instructions by matching third-person observations of objects from instructional videos to those in the user's first-person view. Compared to the general correspondence problem, our setting requires tackling a number of challenges: extreme viewpoint differences, high degrees of object occlusion, and many small objects (e.g., cooking utensils and bike repair tools). + +Task definition. Given a pair of synchronized ego-exo videos and a sequence of query masks of an object of interest in one of the videos, the task is to predict the corresponding mask for the same object in each synchronized frame of the other view, if it is visible. See Figure 4, left. The task can be posed with query objects in either the ego or exo video, with both directions presenting interesting challenges (e.g., high degree of occlusion in ego views, and small object size in exo views). See Appendix 13.A.1. + +Related work. Related tasks are image-level sparse correspondence given query points (instead of object masks) [65] and image-level object co-segmentation [166] for jointly segmenting semantically similar objects. Our task goes beyond static object correspondence, since the interplay between human pose and object state changes during manipulation necessitate using temporal context and tracking as the query object can be highly occluded or blurry [158]. + +![](images/ff098b4a1c0e580f2546738c52b291d66c8636b675c4d68f04c6ea7151e35c9d.jpg) +Figure 5. Ego-exo keystep recognition. This family of tasks consists of fine-grained recognition (left, Section 5.2.1), procedure understanding (center, Section 5.2.3), and energy-efficient multimodal recognition (right, Section 5.2.2). + +![](images/fd1e6617f4d3344436a9a691e16b97009d3ab5c5043378dbf4e1a9fdf69f3ca2.jpg) + +![](images/bf51d78274add40ba749ff7f90622abf5e662a32f03e54a7f68398be5de6235b.jpg) + +# 5.1.2 Ego-exo translation + +Motivation. Our translation task entails synthesizing a target ego clip from a given exo clip. We believe this problem will drive novel research for combining recognition and object synthesis. For example, in Figure 4 (right), the approach must make effective use of the hand's object-specific shape and appearance priors in order to synthesize the ego view of the fingertips—which are not visible in the exo clip. Furthermore, this task will stimulate advances in visual odometry, as the method must be able to infer the ego camera pose from the third-person clip. Ego-exo translation also holds strong application potential, as it may unlock the ability to generate first-person renderings of videos that were originally captured from a third-person perspective, e.g., benefitting robot perception or AR coaching. + +Task definition. We decompose ego-exo translation into two separate tasks: ego track prediction and ego clip generation (Figure 4, right). Ego track prediction estimates the segmentation mask of an object in the unobserved ego frames given the object masks in the observed exo clip. Ego clip generation must generate the image values (i.e., RGB) within the given ground-truth ego mask by making use of the exo clip and the object masks in those frames. This decomposition effectively splits the problem into two tasks: 1) predicting the location and shape of the object in the ego clip, and 2) synthesizing its appearance given the ground-truth position. For each, we consider a variant where the pose of the ego camera with respect to the exo camera is available to use at inference time. This simplifies the problem but reduces the applicability of the method, since this information is typically not available for arbitrary third-person videos. See Appendix 13.A.2. + +Related work. Ego-exo translation relates to cross-view image synthesis [96, 130, 157]. Within this genre, the problem of exo-to-ego generation was recently introduced for both images [93] and video [94, 97], and approached using GANs or diffusion conditioned on the input view. Our work + +not only formalizes this task with ample data, but its formulation also draws attention to the need for a semantic basis to new view synthesis across extreme view changes. + +# 5.2. Ego-exo keystep recognition + +This family of tasks centers around recognizing the keysteps of a procedural activity and modeling their dependencies. + +# 5.2.1 Fine-grained keystep recognition + +Motivation. Recognizing the step a camera wearer is performing is non-trivial: keysteps in the same activity may look similar (folding vs. smoothing the bedsheet) and may involve hand-object interactions with heavy occlusions and head motion. Models with access to multiple views during training can leverage their complementarity to account for the deficiencies of each one, by learning viewpoint invariant representations or distilling multi-view signals into a single model (e.g., human hands from ego; body pose from exo). + +Task definition. We study ego-exo for video recognition. During training, models have access to paired ego-exo data—time-synchronized captures of the same activity from multiple known viewpoints. Each training instance has one ego view, $N$ exo views, and a corresponding keystep label (e.g., "flip the omelette"). At test time, given only a trimmed egocentric video clip, the model must identify the keystep performed from a taxonomy of 689 keysteps across 17 procedural activities. See Figure 5, left. Importantly, all extra supervision (time-alignment, camera poses etc.) is only available at training time; inference is standard keystep recognition, but with models that benefit from cross-viewpoint training. See Appendix 13.B.1. + +Related work. Keystep recognition has been studied in first-person [10, 126, 145, 148] or third-person [9, 100, 159, 205, 207] videos; however, limited work considers both views together. Prior work considers cross-view learning with unpaired videos [7, 82, 182] and view-invariant feature learning on paired videos [144]. In contrast, we ex + +plore keystep recognition in large-scale, procedural activities with fully synchronized training videos. + +# 5.2.2 Energy-efficient multimodal keystep recognition + +Motivation. Current activity detection models assume access to densely sampled clips from the full video and ample computational resources to process them. These assumptions are incompatible with real-world devices (e.g., mobile phones, AR glasses) where the camera is not always on and the compute budget is limited by battery life. This task focuses on building energy-efficient video models to pave the way for feasibility on real-world hardware. + +Task definition. We formulate the problem as an online action detection task, with a given energy budget. See Figure 5, right. Given a stream of audio, IMU, and RGB video data, a model must identify the keystep being performed at each frame, as well as decide which sensor(s) to use for subsequent time-steps. This task will inspire models that are strategic about which modality to deploy when. Energy consumption is the sum of sensor energy (operating the camera/audio/IMU sensors), model inference costs, and memory transfer costs, and must be within $20\mathrm{mW}$ to reflect real-world device power constraints. See Appendix 13.B.2. + +Related work. Prior work on efficient models considers light-weight architectures [41, 56, 101, 155, 165, 195], efficient input processing [43, 44, 73, 102, 156], or inference optimizations [39, 58, 121, 174, 206]. In all cases, they optimize computation (FLOPs), parameter count, or prediction throughput (FPS), which in isolation are insufficient to characterize running on real-world devices. To address this, we propose the first benchmark for energy-efficient video recognition that is tied to real-world, on-device constraints, and measures total power consumed. + +# 5.2.3 Procedure understanding + +Motivation. Automatically understanding the structure of a procedure from video (inferring keystep ordering, preconditions, etc.) would allow assisting AR users in a task or informing robots that learn from human demonstrations. + +Task definition. In our procedure understanding task, given a video segment $s_t$ and its previous video segment history, models have to 1) determine previous keysteps (to be performed before $s_t$ ); infer if $s_t$ is 2) optional or 3) a procedural mistake; 4) predict missing keysteps (should have been performed before $s_t$ but were not); and 5) next keysteps (for which dependencies are satisfied). The task offers two versions of weak supervision: instance-level: segments and their keystep labels are available for train/test; and procedure-level: only unlabeled segments and procedure-specific keystep names are given for train/test. See Figure 5 (center) and Appendix 13.B.3. + +Related work. Prior work focusing on procedural understanding learns an explicit graph [62, 150, 177] as + +![](images/8179128309f83265ff9d2b17386b2dd0f72a15b935efdf359b6762bfc3f7347a.jpg) +Figure 6. Demonstrator and demonstration proficiency estimation. + +ground truth or uses a task graph for representation learning [9, 107, 202] and short-term step understanding [9, 36, 202]. Other work [32, 139] studies mistake detection in a supervised setting. We are the first to propose procedural understanding to evaluate the long-term structure of the task in a weakly-supervised setting. + +# 5.3. Ego-exo proficiency estimation + +Motivation. Going beyond recognizing what a person is doing, this task aims to infer the user's skill level. Such an ability could lead to novel coaching tools that let people learn new skills more effectively, or new ways to evaluate human performance in domains like sports or music. + +Task definition. We consider two variants: (1) demonstrator and (2) demonstration proficiency estimation. Both tasks consider one egocentric and (optionally) $M$ exocentric videos synchronized in time as their inputs. Demonstrator proficiency is formulated as a video classification task, where the model has to output one of four labels (novice, early, intermediate, or late expert). Demonstration proficiency is formulated as a temporal action localization task where given an untrimmed video, the model must output a list of tuples, each containing a timestamp, a proficiency category (i.e., good execution or needs improvement), and its probability. Note that parts of the video that do not reveal the participant's skill are left unlabeled. See Figure 6 and Appendix 13.C. + +Related work. Prior work uses egocentric [12, 35] or exocentric [60, 114, 115] views for proficiency estimation in sports [12, 115, 120], health [60, 92, 191, 208], and others [35, 186]. We propose the first multi-view egocentric and exocentric proficiency estimation benchmark. Unlike prior work, our benchmark spans diverse, day-to-day physical and procedural scenarios and includes temporally localized annotations of (in)correct executions. + +# 5.4. Ego pose + +This family of tasks is motivated by recovering the skilled body movements of participants, even in the extreme setting of monocular ego-video input in dynamic environments. + +![](images/1153a25866145462bb4bdcae10f45145db80a004ab374bbddd7703c6dcaca210.jpg) +Figure 7. Hand and body keypoints for ego-pose estimation + +Motivation. Estimating the physical state of a person's body—the 3D positions of the arms, legs, hands—from the ego view is essential for wearable AI systems that can support human activity. Challenges include subtle and flexible movements, frequent occlusion, and body parts out of view. Task definition. For both the body and hand pose ("ego pose") estimation tasks, the input is an ego video. The output is a set of 3D joint positions of the camera wearer's body and hands for each time step, parameterized as 17 3D body joint positions and 21 3D joint positions per hand, following the MS COCO convention [87]. To our knowledge, EgoExo4D offers the largest manual ground truth (GT) egocentric body and hand pose annotations to date. And, in total, it offers $\sim 14\mathrm{M}$ frames of 3D GT and pseudo-GT combined. See Figure 7 and Appendix 13.D. + +Related work. Limited prior work explores 3D body pose from a wearable camera. Some methods assume no body visibility [63, 80, 98, 187, 188], while others assume partial observability by modifying cameras to capture the body [3, 57, 136, 161, 181]. Our dataset can be used for both paradigms. Existing hand pose datasets use constrained environments [106, 146] with simple hand motion [50, 76, 109], whereas we include diverse real-world scenarios, e.g., with expert musicians and bike mechanics. + +# 6. Conclusions + +Ego-Exo4D provides a dataset of unprecedented scale and realism for ego-exo video learning. It offers a unique window into skilled human activity from 8 compelling domains by hundreds of real-world experts around the globe. Together with the proposed benchmarks, we hope that this new open source resource will set the stage for substantial new research for the years to come. + +Though we are motivated by skill learning, Ego-Exo4D is poised for even broader influence, beyond the proposed benchmarks. Whereas existing datasets lack activity modeling in real-world 3D contexts (e.g., restricted to mocap suits and/or lab settings), Ego-Exo4D is a resource for general 3D vision—such as environment reconstruction, camera relocalization, audio-visual mapping, and many others. Similarly, our novel video-language resources will offer many opportunities for grounding of actions and objects, multi + +modal representation learning, and language generation. Finally, though our tasks prioritize perception from the "ego-only" perspective, the exo component of our data ensures its utility for the more traditional exo viewpoint too, e.g., for activity recognition and body pose estimation. + +# Contribution statement + +This project is the result of a large collaboration between many institutions over the last two years. Initial authors represent the leadership team of the project. Kristen Grauman initiated the project, served as the technical lead, initiated the recognition and proficiency benchmarks and expert commentary, and coordinated their working groups. Andrew Westbury served as the program manager and operations lead for all aspects of the project. Lorenzo Torresani led development of the capture domains, initiated the relation and ego-pose benchmarks, and coordinated their working groups. Kris Kitani led development of the multi-camera rig and supported the Ego-Exo4D engineering team on all aspects of the data annotation and organization. Jitendra Malik served as a scientific advisor. Authors with stars $(^{*})$ were key drivers of implementation, collection, and/or annotation development throughout the project. Authors with daggers $(^{\dagger})$ are faculty and senior researcher PIs for the project. The Appendices detail the contributions of individual authors for the various benchmarks, data collection, and annotation pipelines. + +# Acknowledgements + +We gratefully acknowledge the following colleagues for valuable discussions and support of our project: Vittorio Caggiano, Ilé Danza, Ahmad Darkhalil, Zona de Bloque, Alex Dinh, Rene Martinez Doehner, Ivan Cruz, Matt Feiszli, Vance Feutz, Kelly Forbes, Rohit Girdhar, Pierre Gleize, Andrés Hernández, Shun Iwase, Bolin Lai, Vivian Lee, Brighid Meredith, Ashley Massie, Natalia Neverova, Manohar Paluri, Joelle Pineau, Artsiom Sanakoyeu, Paresh Shenoy, Jiaray Shi, Jiasheng Shi, Gaurav Shrivastava, Mitesh Singh, Manasi Swaminathan, Arjang Talattof, Ali Thabet, Laurens van der Maaten, Andrea Vedaldi, and Toby Zhu. We also sincerely thank the 52 experts who contributed to the expert commentary for their expertise and support; they are listed individually in Appendix 12. Thank you to the Common Visual Data Foundation (CVDF) for hosting the Ego-Exo4D dataset. Finally, thank you to the 740 participants who contributed to this dataset and shared their skills in video. + +Univ. of Bristol is supported in part by EPSRC UMPIRE (EP/T004991/1) and EPSRC PG Visual AI (EP/T028572/1). Z. Zhu is supported by UoB-CSC Scholarship. Univ. of Catalonia is supported in part by the project Future Artificial Intelligence Research (FAIR) - PNRR MUR Cod. PE0000013 - CUP: E63C22001940006. Simon Fraser Univ. is supported in part by the Canada Research Chairs Program (CRC-2019-00298) and NSERC Discovery (2019-06489). Georgia Tech is supported in part by NSF award CNS-2308994. IU is supported in part by NSF DRL-2112635 (AI Institute for Engaged Learning). UT Austin is supported in part by the IFML NSF AI Institute. + +# References + +[1] Ahmed Tawfik Aboukhadra, Jameel Malik, Ahmed Elhayek, Nadia Robertini, and Didier Stricker. Thor-net: End-to-end graformer-based realistic two hands and object reconstruction with self-supervision. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1001-1010, 2023. 58, 59 +[2] Michael Abrash. Creating the future: Augmented reality, the next human-machine interface. In 2021 IEEE International Electron Devices Meeting (IEDM), 2021. 46 +[3] Karan Ahuja, Chris Harrison, Mayank Goel, and Robert Xiao. Mecap: Whole-body digitization for low-cost vr/ar headsets. In Proceedings of the 32nd Annual ACM Symposium on User Interface Software and Technology, pages 453-462, 2019. 9 +[4] Jean-Baptiste Alayrac, Piotr Bojanowski, Nishant Agrawal, Josef Sivic, Ivan Laptev, and Simon Lacoste-Julien. Unsupervised learning from narrated instruction videos. In CVPR, 2016. 4 +[5] Shervin Ardeshir and Ali Borji. Ego2top: Matching viewers in egocentric and top-view videos. In ECCV, 2016. 4 +[6] Shervin Ardeshir and Ali Borji. Egocentric meets top-view. IEEE transactions on pattern analysis and machine intelligence, 41(6), 2018. 4 +[7] Shervin Ardeshir and Ali Borji. An exocentric look at egocentric actions and vice versa. Computer Vision and Image Understanding, 171, 2018. 4, 7 +[8] Kumar Ashutosh, Rohit Girdhar, Lorenzo Torresani, and Kristen Grauman. Hiervl: Learning hierarchical videolanguage embeddings. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023. 6 +[9] Kumar Ashutosh, Santhosh Kumar Ramakrishnan, Triantafyllos Afouras, and Kristen Grauman. Video-mined task graphs for keystep recognition in instructional videos. In NeurIPS, 2023. 4, 7, 8, 29, 49 +[10] Siddhant Bansal, Chetan Arora, and C.V. Jawahar. My view is the best view: Procedure learning from egocentric videos. In European Conference on Computer Vision (ECCV), 2022. 3, 4, 7, 49 +[11] Yizhak Ben-Shabat, Xin Yu, Fatemehsadat Saleh, Dylan Campbell, Cristian Rodriguez-Opazo, Hongdong Li, and Stephen Gould. The ikea asm dataset: Understanding people assembling furniture through actions, objects and pose. 2020. 4 +[12] Gedas Bertasius, Hyun Soo Park, Stella Yu, and Jianbo Shi. Am i a baller? basketball performance assessment from first-person videos. In ICCV, 2017. 3, 8 +[13] Gedas Bertasius, Heng Wang, and Lorenzo Torresani. Is space-time attention all you need for video understanding? In ICML, 2021. 39, 43, 44, 47, 54, 55, 56 +[14] Jing Bi, Jiebo Luo, and Chenliang Xu. Procedure planning in instructional videos via contextual modeling and model + +based policy learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15611-15620, 2021. 49 +[15] Jing Bi, Jiebo Luo, and Chenliang Xu. Procedure planning in instructional videos via contextual modeling and model-based policy learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15611-15620, 2021. 4 +[16] Kay Henning Brodersen, Cheng Soon Ong, Klaas Enno Stephan, and Joachim M Buhmann. The balanced accuracy and its posterior distribution. In 2010 20th international conference on pattern recognition, pages 3121-3124. IEEE, 2010. 35 +[17] Meng Cao, Tianyu Yang, Junwu Weng, Can Zhang, Jue Wang, and Yuexian Zou. Locvtp: Video-text pre-training for temporal localization. In European Conference on Computer Vision, 2022. 4 +[18] Angela Castillo, Maria Escobar, Guillaume Jeanneret, Albert Pumarola, Pablo Arbeláez, Ali Thabet, and Artsiom Sanakoyeu. Bodiffusion: Diffusing sparse observations for full-body human motion synthesis. CV4Metaverse workshop, International Conference on Computer Vision, 2023. 57, 58 +[19] Eric R Chan, Koki Nagano, Matthew A Chan, Alexander W Bergman, Jeong Joon Park, Axel Levy, Miika Aittala, Shalini De Mello, Tero Karras, and Gordon Wetzstein. Generative novel view synthesis with 3d-aware diffusion models. arXiv preprint arXiv:2304.02602, 2023. 4 +[20] Angel Chang, Angela Dai, Tom Funkhouser, Matthias Nießner, Manolis Savva, Shuran Song, Andy Zeng, and Yinda Zhang. Matterport3d: Learning from rgb-d data in indoor environments. In Proceedings of the International Conference on 3D Vision (3DV), 2017. MatterPort3D dataset license available at: http://kaldir.vc.in.tum.de/matterport/MP_TOS.pdf.3 +[21] Chien-Yi Chang, De-An Huang, Danfei Xu, Ehsan Adeli, Li Fei-Fei, and Juan Carlos Niebles. Procedure planning in instructional videos. In European Conference on Computer Vision, pages 334–350. Springer, 2020. 49 +[22] Chien-Yi Chang, De-An Huang, Danfei Xu, Ehsan Adeli, Li Fei-Fei, and Juan Carlos Niebles. Procedure planning in instructional videos. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part XI, pages 334–350. Springer, 2020. 4 +[23] Yu-Hsin Chen, Tien-Ju Yang, Joel Emer, and Vivienne Sze. Eyeriss v2: A flexible accelerator for emerging deep neural networks on mobile devices. IEEE Journal on Emerging and Selected Topics in Circuits and Systems, 2019. 46 +[24] Ho Kei Cheng and Alexander G Schwing. Xmem: Long-term video object segmentation with an atkinson-shiffrin memory model. In European Conference on Computer Vision, pages 640-658. Springer, 2022. 36 +[25] MMPose Contributors. Openmmlab pose estimation toolbox and benchmark. https://github.com/openmmlab/mmpose, 2020.56 +[26] Kellie Corona, Katie Osterdahl, Roderic Collins, and Anthony Hoogs. Meva: A large-scale multiview, multimodal video dataset for activity detection. In WACV, 2021. 3, 7 + +[27] Dima Damen, Hazel Doughty, Giovanni Maria Farinella, Sanja Fidler, Antonino Furnari, Evangelos Kazakos, Davide Moltisanti, Jonathan Munro, Toby Perrett, Will Price, and Michael Wray. Scaling egocentric vision: The epic-kitchens dataset. In European Conference on Computer Vision (ECCV), 2018. 3, 27 +[28] Dima Damen, Hazel Doughty, Giovanni Maria Farinella, Antonino Furnari, Jian Ma, Evangelos Kazakos, Davide Moltisanti, Jonathan Munro, Toby Perrett, Will Price, and Michael Wray. Rescaling egocentric vision. *IJCV*, 2021. 2, 3 +[29] Dima Damen, Hazel Doughty, Giovanni Maria Farinella, Antonino Furnari, Evangelos Kazakos, Jian Ma, Davide Moltisanti, Jonathan Munro, Toby Perrett, Will Price, et al. Rescaling egocentric vision: collection, pipeline and challenges for epic-kitchens-100. International Journal of Computer Vision, 130(1):33-55, 2022. 7 +[30] Roeland De Geest, Efstratios Gavves, Amir Ghodrati, Zhenyang Li, Cees Snoek, and Tinne Tuytelaars. Online action detection. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part V 14, pages 269-284. Springer, 2016. 46, 51 +[31] Radosvet Desislavov, Fernando Martínez-Plumed, and José Hernández-Orallo. Trends in ai inference energy consumption: Beyond the performance-vs-parameter laws of deep learning. Sustainable Computing: Informatics and Systems, 38:100857, 2023. 46 +[32] Guodong Ding, Fadime Sener, Shugao Ma, and Angela Yao. Every mistake counts in assembly. arXiv preprint arXiv:2307.16453, 2023. 8 +[33] Keyan Ding, Kede Ma, Shiqi Wang, and Eero P Simoncelli. Image quality assessment: Unifying structure and texture similarity. IEEE transactions on pattern analysis and machine intelligence, 44(5):2567-2581, 2020. 39 +[34] H Doughty, D Damen, and W Mayol-Cuevas. Who's better? who's best? pairwise deep ranking for skill determination. In CVPR, 2018. 3 +[35] Hazel Doughty, Walterio Mayol-Cuevas, and Dima Damen. The Pros and Cons: Rank-aware Temporal Attention for Skill Determination in Long Videos. 2019. 3, 8 +[36] Nikita Dvornik, Isma Hadji, Hai Pham, Dhaivat Bhatt, Brais Martinez, Afsaneh Fazly, and Allan D Jepson. Flow graph to video grounding for weakly-supervised multi-step localization. In ECCV, pages 319-335. Springer, 2022. 4, 8 +[37] Ehsan Elhamifar and Dat Huynh. Self-supervised multitask procedure learning from instructional videos. In European Conference on Computer Vision, pages 557-573. Springer, 2020. 4 +[38] Jakob Engel, Kiran Somasundaram, Michael Goesele, Albert Sun, Alexander Gamino, Andrew Turner, Arjang Talattof, Arnie Yuan, Bilal Souti, Brighid Meredith, Cheng Peng, Chris Sweeney, Cole Wilson, Dan Barnes, Daniel DeTone, David Caruso, Derek Valleroy, Dinesh Ginjupalli, Duncan Frost, Edward Miller, Elias Mueggler, Evgeniy Oleinik, Fan Zhang, Guruprasad Somasundaram, Gustavo + +Solaira, Harry Lanaras, Henry Howard-Jenkins, Huixuan Tang, Hyo Jin Kim, Jaime Rivera, Ji Luo, Jing Dong, Julian Straub, Kevin Bailey, Kevin Eckenhoff, Lingni Ma, Luis Pesqueira, Mark Schwesinger, Maurizio Monge, Nan Yang, Nick Charron, Nikhil Raina, Omkar Parkhi, Peter Borschowa, Pierre Moulon, Prince Gupta, Raul Mur-Artal, Robbie Pennington, Sachin Kulkarni, Sagar Miglani, Santosh Gondi, Saransh Solanki, Sean Diener, Shangyi Cheng, Simon Green, Steve Saarinen, Suvam Patra, Tassos Mourikis, Thomas Whelan, Tripti Singh, Vasileios Balntas, Vijay Baiyya, Wilson Dreewes, Xiaqing Pan, Yang Lou, Yipu Zhao, Yusuf Mansour, Yuyang Zou, Zhaoyang Lv, Zijian Wang, Mingfei Yan, Carl Ren, Renzo De Nardi, and Richard Newcombe. Project Aria: A new tool for egocentric multi-modal AI research, 2023. 3, 4, 2 +[39] Steven K Esser, Jeffrey L McKinstry, Deepika Bablani, Rathinakumar Appuswamy, and Dharmendra S Modha. Learned step size quantization. arXiv preprint arXiv:1902.08153, 2019. 8 +[40] Chenyou Fan, Jangwon Lee, Mingze Xu, Krishna Kumar Singh, Yong Jae Lee, David J Crandall, and Michael S Ryoo. Identifying first-person camera wearers in third-person videos. In CVPR, 2017. 4 +[41] Christoph Feichtenhofer. X3d: Expanding architectures for efficient video recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 203-213, 2020. 8, 45, 46, 47, 48 +[42] John H. Flavell, Eleanor R. Flavell, Frances L. Green, and Sharon A. Wilcox. The development of three spatial perspective-taking rules. Child Development, 1981. 2 +[43] Ruohan Gao, Tae-Hyun Oh, Kristen Grauman, and Lorenzo Torresani. Listen to look: Action recognition by previewing audio. In CVPR, 2020. 8 +[44] Amir Ghodrati, Babak Ehteshami Bejnordi, and Amirhossein Habibian. Frameexit: Conditional early exiting for efficient video recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2021. 8 +[45] Rohit Girdhar, Mannat Singh, Nikhila Ravi, Laurens van der Maaten, Armand Joulin, and Ishan Misra. Omnivore: A single model for many visual modalities. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16102-16112, 2022. 54 +[46] Raghav Goyal, Samira Ebrahimi Kahou, Vincent Michalski, Joanna Materzynska, Susanne Westphal, Heuna Kim, Valentin Haenel, Ingo Fruend, Peter Yianilos, Moritz Mueller-Freitag, et al. The "something something" video database for learning and evaluating visual common sense. In Proceedings of the IEEE international conference on computer vision, pages 5842-5850, 2017. 42 +[47] Kristen Grauman, Andrew Westbury, Eugene Byrne, Zachary Chavis, Antonino Furnari, Rohit Girdhar, Jackson Hamburger, Hao Jiang, Miao Liu, Xingyu Liu, Miguel Martin, Tushar Nagarajan, Ilija Radosavovic, Santhosh Kumar Ramakrishnan, Fiona Ryan, Jayant Sharma, Michael Wray, Mengmeng Xu, Eric Zhongcong Xu, Chen Zhao, Siddhant Bansal, Dhruv Batra, Vincent Cartillier, Sean Crane, Tien Do, Morrie Doulaty, Akshay Erapalli, Christoph + +Feichtenhofer, Adriano Fragomeni, Qichen Fu, Abraham Gebreselasie, Cristina Gonzalez, James Hillis, Xuhua Huang, Yifei Huang, Wenqi Jia, Weslie Khoo, Jachym Kolář, Satwik Kottur, Anurag Kumar, Federico Landini, Chao Li, Yanghao Li, Zhenqiang Li, Karttikeya Mangalam, Raghava Modhugu, Jonathan Munro, Tullie Murrell, Takumi Nishiyasu, Will Price, Paola Ruiz, Merey Ramazanova, Leda Sari, Kiran Somasundaram, Audrey Southerland, Yusuke Sugano, Ruijie Tao, Minh Vo, Yuchen Wang, Xindi Wu, Takuma Yagi, Ziwei Zhao, Yunyi Zhu, Pablo Arbelaez, David Crandall, Dima Damen, Giovanni Maria Farinella, Christian Fuegen, Bernard Ghanem, Vamsi Krishna Ithapu, C. V. Jawahar, Hanbyul Joo, Kris Kitani, Haizhou Li, Richard Newcombe, Aude Oliva, Hyun Soo Park, James M. Rehg, Yoichi Sato, Jianbo Shi, Mike Zheng Shou, Antonio Torralba, Lorenzo Torresani, Mingfei Yan, and Jitendra Malik. Ego4D: Around the world in 3,000 hours of egocentric video. In CVPR, 2022. 2, 3, 4, 6, 7, 29, 34, 43 +[48] Chunhui Gu, Chen Sun, David A. Ross, Carl Vondrick, Caroline Pantofaru, Yeqing Li, Sudheendra Vijayanarasimhan, George Toderici, Susanna Ricco, Rahul Sukthankar, Cordelia Schmid, and Jitendra Malik. Ava: A video dataset of spatio-temporally localized atomic visual actions. In CVPR, 2018. 2 +[49] Vladimir Guzov, Aymen Mir, Torsten Sattler, and Gerard Pons-Moll. Human poseitioning system (hps): 3d human pose estimation and self-localization in large scenes from body-mounted sensors. In CVPR, 2021. 3 +[50] Shreyas Hampali, Mahdi Rad, Markus Oberweger, and Vincent Lepetit. Honnotate: A method for 3d annotation of hand and object poses. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3196-3206, 2020. 9 +[51] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, 2016. 36, 58 +[52] Geoffrey Hinton, Oriol Vinyals, and Jeff Dean. Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531, 2015. 43, 44 +[53] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in Neural Information Processing Systems, 33:6840-6851, 2020. 57 +[54] Alain Hore and Djemel Ziou. Image quality metrics: Psnr vs. ssim. In 2010 20th international conference on pattern recognition, pages 2366-2369. IEEE, 2010. 39 +[55] Mark Horowitz. 1.1 computing's energy problem (and what we can do about it). In 2014 IEEE international solid-state circuits conference digest of technical papers (ISSCC), 2014. 46 +[56] Andrew G Howard, Menglong Zhu, Bo Chen, Dmitry Kalenichenko, Weijun Wang, Tobias Weyand, Marco Andretto, and Hartwig Adam. Mobilenets: Efficient convolutional neural networks for mobile vision applications. arXiv preprint arXiv:1704.04861, 2017. 8 +[57] Dong-Hyun Hwang, Kohei Aso, Ye Yuan, Kris Kitani, and Hideki Koike. Monoeye: Multimodal human motion capture system using a single ultra-wide fisheye camera. In + +Proceedings of the 33rd Annual ACM Symposium on User Interface Software and Technology, pages 98-111, 2020. 9 +[58] Forrest N Iandola, Song Han, Matthew W Moskewicz, Khalid Ashraf, William J Dally, and Kurt Keutzer. SqueezeNet: Alexnet-level accuracy with 50x fewer parameters and; 0.5 mb model size. arXiv preprint arXiv:1602.07360, 2016. 8 +[59] Vladimir Iashin and Esa Rahtu. Multi-modal dense video captioning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops, pages 958-959, 2020. 26 +[60] Hassan Ismail Fawaz, Germain Forestier, Jonathan Weber, Lhassane Idoumghar, and Pierre-Alain Muller. Evaluating surgical skills from kinematic data using convolutional neural networks. In Medical Image Computing and Computer Assisted Intervention - MICCAI 2018, pages 214-221, 2018. 8 +[61] Phillip Isola, Jun-Yan Zhu, Tinghui Zhou, and Alexei A Efros. Image-to-image translation with conditional adversarial networks. CVPR, 2017. 39 +[62] Yunseok Jang, Sungryull Sohn, Lajanugen Logeswaran, Tiange Luo, Moontae Lee, and Honglak Lee. Multimodal subtask graph generation from instructional videos. arXiv preprint arXiv:2302.08672, 2023. 8 +[63] Hao Jiang and Kristen Grauman. Seeing invisible poses: Estimating 3d body pose from egocentric video. In CVPR, 2017. 9 +[64] Jiaxi Jiang, Paul Streli, Huajian Qiu, Andreas Fender, Larissa Laich, Patrick Snape, and Christian Holz. Avatarposer: Articulated full-body pose tracking from sparse motion sensing. In European Conference on Computer Vision, pages 443-460. Springer, 2022. 57, 58 +[65] Wei Jiang, Eduard Trulls, Jan Hosang, Andrea Tagliasacchi, and Kwang Moo Yi. Cotr: Correspondence transformer for matching across images. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6207-6217, 2021. 6, 35 +[66] Hanbyul Joo, Tomas Simon, Xulong Li, Hao Liu, Lei Tan, Lin Gui, Sean Banerjee, Timothy Scott Godisart, Bart Nabbe, Iain Matthews, Takeo Kanade, Shohei Nobuhara, and Yaser Sheikh. Panoptic studio: A massively multiview system for social interaction capture. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2017. 3 +[67] Will Kay, Joao Carreira, Karen Simonyan, Brian Zhang, Chloe Hillier, Sudheendra Vijayanarasimhan, Fabio Viola, Tim Green, Trevor Back, Paul Natsev, et al. The kinetics human action video dataset. arXiv preprint arXiv:1705.06950, 2017. 2, 42, 43 +[68] Rawal Khirodkar, Aayush Bansal, Lingni Ma, Richard Newcombe, Minh Vo, and Kris Kitani. Egohumans: An egocentric 3d multi-human benchmark. In ICCV, 2023. 3 +[69] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 36 +[70] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C. Berg, Wan-Yen Lo, Piotr Dollar, + +and Ross Girshick. Segment anything. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 4015-4026, 2023. 33 +[71] Dohwan Ko, Joonmyung Choi, Juyeon Ko, Shinyeong Noh, Kyoung-Woon On, Eun-Sol Kim, and Hyunwoo J Kim. Video-text representation learning via differentiable weak temporal alignment. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5016-5025, 2022. 4 +[72] Nikos Kolotouros, Georgios Pavlakos, Michael J Black, and Kostas Daniilidis. Learning to reconstruct 3d human pose and shape via model-fitting in the loop. In Proceedings of the IEEE/CVF international conference on computer vision, pages 2252-2261, 2019. 59 +[73] Bruno Korbar, Du Tran, and Lorenzo Torresani. Scsampler: Sampling salient clips from video for efficient action recognition. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2019. 8 +[74] Zuzana Kukelova, Jan Heller, and Andrew Fitzgibbon. Efficient intersection of three quadrics and applications in computer vision. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 1799-1808, 2016. 3 +[75] Iljung Kwak, Jian-Zhong Guo, Adam Hantman, David Kriegman, and Kristin Branson. Detecting the starting frame of actions in video. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 489-497, 2020. 54 +[76] Taein Kwon, Bugra Tekin, Jan Stühmer, Federica Bogo, and Marc Pollefeys. H2o: Two hands manipulating objects for first person interaction recognition. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 10138-10148, 2021. 2, 3, 9, 7 +[77] F. De la Torre, J. Hodgins, J. Montano, S. Valcarcel, R. Forcada, and J. Macey. Guide to the carnegie mellon university multimodal activity (cmu-mmac) database. In Tech. report CMU-RI-TR-08-22, Robotics Institute, Carnegie Mellon University, 2009. 2, 3, 7 +[78] Y. J. Lee, J. Ghosh, and K. Grauman. Discovering important people and objects for egocentric video summarization. In CVPR, 2012. 3 +[79] Jiefeng Li, Chao Xu, Zhicun Chen, Siyuan Bian, Lixin Yang, and Cewu Lu. Hybrik: A hybrid analytical-neural inverse kinematics solution for 3d human pose and shape estimation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3383-3393, 2021. 59 +[80] Jiaman Li, Karen Liu, and Jiajun Wu. Ego-body pose estimation via ego-head pose estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17142-17151, 2023. 3, 9, 56, 57, 58 +[81] Yin Li, Miao Liu, and James M. Rehg. In the eye of beholder: Joint learning of gaze and actions in first person video. In ECCV, 2018. 3, 7 +[82] Yanghao Li, Tushar Nagarajan, Bo Xiong, and Kristen Grauman. Ego-exo: Transferring visual representations from third-person to first-person videos. In Proceedings of + +the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6943-6953, 2021. 4, 7, 21, 43, 44 +[83] Junhua Liao, Haihan Duan, Kanghui Feng, Wanbing Zhao, Yanbing Yang, and Liangyin Chen. A light weight model for active speaker detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22932-22941, 2023. 47, 48 +[84] Kevin Lin, Lijuan Wang, and Zicheng Liu. End-to-end human pose and mesh reconstruction with transformers. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1954-1963, 2021. 59 +[85] Kevin Qinghong Lin, Alex Jinpeng Wang, Mattia Soldan, Michael Wray, Rui Yan, Eric Zhongcong Xu, Difei Gao, Rongcheng Tu, Wenzhe Zhao, Weijie Kong, et al. Egocentric video-language pretraining. NeurIPS, 2022. 6, 26, 50 +[86] TY Lin, Y Cui, S Belongie, and J Hays. Learning deep representations for ground-to-aerial geolocation. In CVPR, 2015. 4 +[87] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dolkar, and C Lawrence Zitnick. Microsoft COCO: Common objects in context. In ECCV, 2014. 9 +[88] Tsung-Yi Lin, Piotr Dólar, Ross Girshick, Kaiming He, Bharath Hariharan, and Serge Belongie. Feature pyramid networks for object detection. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2117-2125, 2017. 58 +[89] Xudong Lin, Fabio Petroni, Gedas Bertasius, Marcus Rohrbach, Shih-Fu Chang, and Lorenzo Torresani. Learning to recognize procedural activities with distant supervision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13853-13863, 2022. 4, 29 +[90] Andrew Liu, Richard Tucker, Varun Jampani, Ameesh Makadia, Noah Snavely, and Angjoo Kanazawa. Infinite nature: Perpetual view generation of natural scenes from a single image. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14458-14467, 2021. 4 +[91] Chiao Liu, Lyle Bainbridge, Andrew Berkovich, Song Chen, Wei Gao, Tsung-Hsun Tsai, Kazuya Mori, Rimon Ikeno, Masayuki Uno, Toshiyuki Isozaki, et al. A $4.6\mu \mathrm{m}$ , $512\times 512$ , ultra-low power stacked digital pixel sensor with triple quantization and 127db dynamic range. In 2020 IEEE International Electron Devices Meeting (IEDM), 2020. 46 +[92] Daochang Liu, Qiyue Li, Tingting Jiang, Yizhou Wang, Rulin Miao, Fei Shan, and Ziyu Li. Towards unified surgical skill assessment. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9522-9531, 2021. 8 +[93] Gaowen Liu, Hao Tang, Hugo Latapie, and Yan Yan. Exocentric to egocentric image generation via parallel generative adversarial network. In ICASSP 2020-2020 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 1843-1847. IEEE, 2020. 7 +[94] Gaowen Liu, Hao Tang, Hugo M Latapie, Jason J Corso, and Yan Yan. Cross-view exocentric to egocentric video + +synthesis. In Proceedings of the 29th ACM International Conference on Multimedia, pages 974-982, 2021. 7 +[95] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In International Conference on Learning Representations, 2019. 47 +[96] Xiaohu Lu, Zuoyue Li, Zhaopeng Cui, Martin R. Oswald, Marc Pollefeys, and Rongjun Qin. Geometry-aware satellite-to-ground image synthesis for urban areas. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 7 +[97] Mi Luo, Zihui Xue, Alex Dimakis, and Kristen Grauman. Put myself in your shoes: Lifting the egocentric perspective from exocentric videos. arXiv:2403.06351, 2024. 7 +[98] Zhengyi Luo, Ryo Hachiuma, Ye Yuan, and Kris Kitani. Dynamics-regulated kinematic policy for egocentric pose estimation. In Advances in Neural Information Processing Systems, 2021. 9, 56, 57, 58 +[99] Naureen Mahmood, Nima Ghorbani, Nikolaus F. Troje, Gerard Pons-Moll, and Michael J. Black. Amass: Archive of motion capture as surface shapes. In The IEEE International Conference on Computer Vision (ICCV), 2019. 57 +[100] Effrosyni Mavroudi, Triantafyllos Afouras, and Lorenzo Torresani. Learning to ground instructional articles in videos through narrations. 2022. 7 +[101] Sachin Mehta and Mohammad Rastegari. Mobilevit: lightweight, general-purpose, and mobile-friendly vision transformer. arXiv preprint arXiv:2110.02178, 2021. 8 +[102] Yue Meng, Chung-Ching Lin, Rameswar Panda, Prasanna Sattigeri, Leonid Karlinsky, Aude Oliva, Kate Saenko, and Rogerio Feris. Ar-net: Adaptive frame resolution for efficient action recognition. In ECCV 2020, 2020. 8 +[103] Antoine Miech, Dimitri Zhukov, Jean-Baptiste Alayrac, Makarand Tapaswi, Ivan Laptev, and Josef Sivic. HowTo100M: Learning a Text-Video Embedding by Watching Hundred Million Narrated Video Clips. In ICCV, 2019. 2, 4, 29 +[104] Antoine Miech, Jean-Baptiste Alayrac, Lucas Smaira, Ivan Laptev, Josef Sivic, and Andrew Zisserman. End-to-end learning of visual representations from uncurated instructional videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9879-9889, 2020. 4, 29 +[105] Mathew Monfort, Alex Andonian, Bolei Zhou, Kandan Ramakrishnan, Sarah Adel Bargal, Tom Yan, Lisa Brown, Quanfu Fan, Dan Gutfreund, Carl Vondrick, and Aude Oliva. Moments in time dataset: one million videos for event understanding. PAMI, 2019. 2 +[106] Gyeongsik Moon, Shoou-I Yu, He Wen, Takaaki Shiratori, and Kyoung Mu Lee. Interhand2. 6m: A dataset and baseline for 3d interacting hand pose estimation from a single rgb image. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XX 16, pages 548-564. Springer, 2020. 9 +[107] Medhini Narasimhan, Licheng Yu, Sean Bell, Ning Zhang, and Trevor Darrell. Learning and verification of task structure in instructional videos. arXiv preprint arXiv:2303.13519, 2023. 4, 8 + +[108] Nora Newcombe. The development of spatial perspective taking. Advances in child development and behavior, 1989. 2 +[109] Takehiko Ohkawa, Kun He, Fadime Sener, Tomas Hodan, Luan Tran, and Cem Keskin. Assemblyhands: Towards egocentric activity understanding via 3d hand pose estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12999-13008, 2023. 9, 7 +[110] Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748, 2018. 43, 44 +[111] Boxiao Pan, Haoye Cai, De-An Huang, Kuan-Hui Lee, Adrien Gaidon, Ehsan Adeli, and Juan Carlos Niebles. Spatio-temporal graph for video captioning with knowledge distillation. In CVPR, 2020. 26 +[112] JoonKyu Park, Yeonguk Oh, Gyeongsik Moon, Hongsuk Choi, and Kyoung Mu Lee. Handoccnet: Occlusion-robust 3d hand mesh estimation network. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1496–1505, 2022. 58, 59 +[113] Paritosh Parmar and Brendan Morris. Action quality assessment across multiple actions. In WACV, 2019. 3 +[114] Paritosh Parmar and Brendan Tran Morris. Learning to score olympic events, 2017. 8 +[115] Paritosh Parmar and Brendan Tran Morris. What and how well you performed? a multitask learning approach to action quality assessment. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 304-313, 2019. 8 +[116] William Peebles and Saining Xie. Scalable diffusion models with transformers. arXiv preprint arXiv:2212.09748, 2022.39 +[117] Federico Perazzi, Jordi Pont-Tuset, Brian McWilliams, Luc Van Gool, Markus Gross, and Alexander Sorkine-Hornung. A benchmark dataset and evaluation methodology for video object segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 724-732, 2016. 35, 39 +[118] Ethan Perez, Florian Strub, Harm De Vries, Vincent Dumoulin, and Aaron Courville. Film: Visual reasoning with a general conditioning layer. In Proceedings of the AAAI conference on artificial intelligence, 2018. 39 +[119] H. Pirsiavash and D. Ramanan. Detecting activities of daily living in first-person camera views. In CVPR, 2012. 3 +[120] H. Piriavash, C. Vondrick, and A. Torralba. Assessing the quality of actions. In ECCV, 2014. 3, 8 +[121] Antonio Polino, Razvan Pascanu, and Dan Alistarh. Model compression via distillation and quantization. arXiv preprint arXiv:1802.05668, 2018. 8 +[122] Daniel Povey, Arnab Ghoshal, Gilles Boulianne, Lukávs. Burget, Ondrej Glembek, Nagendra Kumar Goel, Mirko Hannemann, Petr Motlícek, Yanmin Qian, Petr Schwarz, Jan Silovsky, Georg Stemmer, and Karel Vesely. The kaldi speech recognition toolkit. 2011. 47 +[123] Shraman Pramanick, Yale Song, Sayan Nag, Kevin Qinghong Lin, Hardik Shah, Mike Zheng Shou, + +Rama Chellappa, and Pengchuan Zhang. Egovlpv2: Egocentric video-language pre-training with fusion in the backbone. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5285-5297, 2023. 6, 26, 43, 44 +[124] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 39 +[125] Alec Radford, Jong Wook Kim, Tao Xu, Greg Brockman, Christine McLeavey, and Ilya Sutskever. Robust speech recognition via large-scale weak supervision. In International Conference on Machine Learning, pages 28492-28518. PMLR, 2023. 27 +[126] Francesco Ragusa, Antonino Furnari, Salvatore Livatino, and Giovanni Maria Farinella. The meccano dataset: Understanding human-object interactions from egocentric videos in an industrial-like domain. In WACV, 2021. 3, 7 +[127] N. Rai, H. Chen, J. Ji, R. Desai, K. Kozuka, S. Ishizaka, E. Adeli, and J.C. Niebles. Home action genome: Contrastive compositional action understanding. In CVPR, 2021. 2, 3 +[128] Santhosh Kumar Ramakrishnan, Aaron Gokaslan, Erik Wijmans, Oleksandr Maksymets, Alexander Clegg, John M Turner, Eric Undersander, Wojciech Galuba, Andrew Westbury, Angel X Chang, Manolis Savva, Yili Zhao, and Dhruv Batra. Habitat-matterport 3d dataset (HM3d): 1000 large-scale 3d environments for embodied AI. In Thirty-fifth Conference on Neural Information Processing Systems Datasets and Benchmarks Track, 2021. 3 +[129] Scott Reed, Honglak Lee, Dragomir Anguelov, Christian Szegedy, Dumitru Erhan, and Andrew Rabinovich. Training deep neural networks on noisy labels with bootstrapping. arXiv preprint arXiv:1412.6596, 2014. 39 +[130] Krishna Regmi and Ali Borji. Cross-view image synthesis using conditional gans. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 4, 7 +[131] Krishna Regmi and Ali Borji. Cross-view image synthesis using geometry-guided conditional gans. Computer Vision and Image Understanding, 2019. 4 +[132] Krishna Regmi and Mubarak Shah. Bridging the domain gap for ground-to-aerial image matching. In ICCV, 2019. 4 +[133] Jeremy Reizenstein, Roman Shapovalov, Philipp Henzler, Luca Sbordone, Patrick Labatut, and David Novotny. Common objects in 3d: Large-scale learning and evaluation of real-life 3d category reconstruction. In ICCV, 2021. 3 +[134] Bin Ren, Hao Tang, and Nicu Sebe. Cascaded cross mlp-mixer gans for cross-view image translation. arXiv preprint arXiv:2110.10183, 2021. 4 +[135] Xuanchi Ren and Xiaolong Wang. Look outside the room: Synthesizing a consistent long-term 3d scene video from a single image. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3563-3573, 2022. 4 +[136] Helge Rhodin, Christian Richardt, Dan Casas, Eldar In-safutdinov, Mohammad Shafiei, Hans-Peter Seidel, Bernt + +Schiele, and Christian Theobalt. Egocap: egocentric marker-less motion capture with two fisheye cameras. ACM Transactions on Graphics (TOG), 35(6):1-11, 2016. 9 +[137] Robin Rombach, Patrick Esser, and Björn Ommer. Geometry-free view synthesis: Transformers and no 3d priors. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14356-14366, 2021. 4 +[138] Javier Romero, Dimitrios Tzionas, and Michael J. Black. Embodied hands: Modeling and capturing hands and bodies together. ACM Transactions on Graphics, (Proc. SIGGRAPH Asia), 36(6), 2017. 59 +[139] Fadime Sener, Dibyadip Chatterjee, Daniel Shelepov, Kun He, Dipika Singhania, Robert Wang, and Angela Yao. Assembly101: A large-scale multi-view video dataset for understanding procedural activities. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21096-21106, 2022. 2, 3, 8, 7 +[140] Pierre Sermanet, Corey Lynch, Jasmine Hsu, and Sergey Levine. Time-contrastive networks: Self-supervised learning from multi-view observation. In 2017 IEEE Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), pages 486–487. IEEE, 2017. 43 +[141] Pierre Sermanet, Corey Lynch, Yevgen Chebotar, Jasmine Hsu, Eric Jang, Stefan Schaal, and Sergey Levine. Time-contrastive networks: Self-supervised learning from video. Proceedings of International Conference in Robotics and Automation (ICRA), 2018. 4 +[142] Xi Shen, Alexei A Efros, Armand Joulin, and Mathieu Aubry. Learning co-segmentation by segment swapping for retrieval and discovery. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5082-5092, 2022. 35, 36 +[143] Nina Shvetsova, Brian Chen, Andrew Rouditchenko, Samuel Thomas, Brian Kingsbury, Rogerio S Feris, David Harwath, James Glass, and Hilde Kuehne. Everything at once-multi-modal fusion transformer for video retrieval. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20020-20029, 2022. 4 +[144] Gunnar A Sigurdsson, Abhinav Gupta, Cordelia Schmid, Ali Farhadi, and Karteek Alahari. Actor and observer: Joint modeling of first and third-person videos. In CVPR, 2018, 4, 7, 43 +[145] Gunnar A Sigurdsson, Abhinav Gupta, Cordelia Schmid, Ali Farhadi, and Karteek Alahari. Charades-ego: A large-scale dataset of paired third and first person videos. arXiv preprint arXiv:1804.09626, 2018. 2, 3, 7 +[146] Tomas Simon, Hanbyul Joo, Iain Matthews, and Yaser Sheikh. Hand keypoint detection in single images using multiview bootstrapping. In Proceedings of the IEEE conference on Computer Vision and Pattern Recognition, pages 1145-1153, 2017. 9 +[147] Krishna Kumar Singh, Kayvon Fatahalian, and Alexei A. Efros. Krishnamac: Using a longitudinal, single-person, egocentric dataset for scene understanding tasks. In WACV, 2016. 3 +[148] Yale Song, Eugene Byrne, Tushar Nagarajan, Huiyu Wang, Miguel Martin, and Lorenzo Torresani. Ego4d goal-step: + +Toward hierarchical understanding of procedural activities. In NeurIPS, 2023. 7 +[149] Khurram Soomro, Amir Roshan Zamir, and Mubarak Shah. Ucf101: A dataset of 101 human action classes from videos in the wild. In CRCV-TR-12-01, 2012. 2, 42 +[150] Bilge Soran, Ali Farhadi, and Linda Shapiro. Generating notifications for missing actions: Don't forget to turn the lights off! In ICCV, pages 4669-4677, 2015. 8 +[151] Julian Straub, Thomas Whelan, Lingni Ma, Yufan Chen, Erik Wijmans, Simon Green, Jakob J. Engel, Raul Mur-Artal, Carl Ren, Shobhit Verma, Anton Clarkson, Mingfei Yan, Brian Budge, Yajie Yan, Xiaqing Pan, June Yon, Yuyang Zou, Kimberly Leon, Nigel Carter, Jesus Briales, Tyler Gillingham, Elias Mueggler, Luis Pesqueira, Manolis Savva, Dhruv Batra, Hauke M. Strasdat, Renzo De Nardi, Michael Goesele, Steven Lovegrove, and Richard Newcombe. The Replica dataset: A digital replica of indoor spaces. arXiv preprint arXiv:1906.05797, 2019. 3 +[152] Carole H Sudre, Wenqi Li, Tom Vercauteren, Sebastien Ourselin, and M Jorge Cardoso. Generalised dice overlap as a deep learning loss function for highly unbalanced segmentations. In Deep Learning in Medical Image Analysis and Multimodal Learning for Clinical Decision Support: Third International Workshop, DLMIA 2017, and 7th International Workshop, ML-CDS 2017, Held in conjunction with MICCAI 2017, Quebec City, QC, Canada, September 14, Proceedings 3, pages 240-248. Springer, 2017. 39 +[153] Vivienne Sze, Yu-Hsin Chen, Tien-Ju Yang, and Joel S Emer. How to evaluate deep neural network processors: Tops/w (alone) considered harmful. IEEE Solid-State Circuits Magazine, 2020. 46 +[154] Mukund Varma T, Peihao Wang, Xuxi Chen, Tianlong Chen, Subhashini Venugopalan, and Zhangyang Wang. Is attention all that neRF needs? In The Eleventh International Conference on Learning Representations, 2023. 39 +[155] Mingxing Tan, Bo Chen, Ruoming Pang, Vijay Vasudevan, Mark Sandler, Andrew Howard, and Quoc V Le. Mnasnet: Platform-aware neural architecture search for mobile. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2820-2828, 2019. 8 +[156] Shuhan Tan, Tushar Nagarajan, and Kristen Grauman. Egodistill: Egocentric head motion distillation for efficient video understanding. NeurIPS, 2023. 8 +[157] Hao Tang, Dan Xu, Nicu Sebe, Yanzhi Wang, Jason J Corso, and Yan Yan. Multi-channel attention selection gan with cascaded semantic guidance for cross-view image translation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2417-2426, 2019. 4, 7 +[158] Hao Tang, Kevin Liang, Kristen Grauman, Matt Feiszli, and Weiyao Wang. Egotracks: A long-term egocentric visual object tracking dataset. Advances in Neural Information Processing Systems, 2023. 6 +[159] Yansong Tang, Jiwen Lu, and Jie Zhou. Comprehensive instructional video analysis: The coin dataset and performance evaluation. IEEE transactions on pattern analysis and machine intelligence, 2020. 2, 4, 7 + +[160] Zachary Teed and Jia Deng. DROID-SLAM: Deep Visual SLAM for Monocular, Stereo, and RGB-D Cameras. Advances in neural information processing systems, 2021. 57 +[161] Denis Tome, Patrick Peluse, Lourdes Agapito, and Hernan Badino. xr-egopose: Egocentric 3d human pose from an hmd camera. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2019. 9 +[162] Zhan Tong, Yibing Song, Jue Wang, and Limin Wang. Videomae: Masked autoencoders are data-efficient learners for self-supervised video pre-training. 2022. 43 +[163] Vadim Tschernezki, Ahmad Darkhalil, Zhifan Zhu, David Fouhey, Iro Larina, Diane Larlus, Dima Damen, and Andrea Vedaldi. EPIC Fields: Marrying 3D Geometry and Video Understanding. In Proceedings of the Neural Information Processing Systems (NeurIPS), 2023. 3 +[164] Hung-Yu Tseng, Qinbo Li, Changil Kim, Suhib Alsisan, Jia-Bin Huang, and Johannes Kopf. Consistent view synthesis with pose-guided diffusion models. arXiv preprint arXiv:2303.17598, 2023. 4 +[165] Pavan Kumar Anasosalu Vasu, James Gabriel, Jeff Zhu, Oncel Tuzel, and Anurag Ranjan. Mobileone: An improved one millisecond mobile backbone. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7907-7917, 2023. 8 +[166] Sara Vicente, Carsten Rother, and Vladimir Kolmogorov. Object cosegmentation. In CVPR 2011, pages 2217-2224. IEEE, 2011. 6 +[167] Hanlin Wang, Yilu Wu, Sheng Guo, and Limin Wang. Pdpp: Projected diffusion for procedure planning in instructional videos. arXiv preprint arXiv:2303.14676, 2023. 4 +[168] Daniel Watson, William Chan, Ricardo Martin-Brualla, Jonathan Ho, Andrea Tagliasacchi, and Mohammad Norouzi. Novel view synthesis with diffusion models. arXiv preprint arXiv:2210.04628, 2022. 4 +[169] Daniel Weinland, Remi Ronfard, and Edmond Boyer. Free viewpoint action recognition using motion history volumes. Computer Vision and Image Understanding (CVIU), 2006. 3, 7 +[170] Yangming Wen, Krishna Kumar Singh, Markham Anderson, Wei-Pang Jan, and Yong Jae Lee. Seeing the unseen: Predicting the first-person camera wearer's location and pose in third-person scenes. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) Workshops, pages 3446-3455, 2021. 4 +[171] Olivia Wiles, Georgia Gkioxari, Richard Szeliski, and Justin Johnson. Synsin: End-to-end view synthesis from a single image. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7467-7477, 2020. 4 +[172] Benita Wong, Joya Chen, You Wu, Stan Weixian Lei, Dongxing Mao, Difei Gao, and Mike Zheng Shou. Assistq: Affordance-centric question-driven task completion for egocentric assistant. In European Conference on Computer Vision, 2022. 3 +[173] Zhirong Wu, Shuran Song, Aditya Khosla, Fisher Yu, Linguang Zhang, Xiaou Tang, and Jianxiong Xiao. 3d shapenets: A deep representation for volumetric shapes. In + +Computer Vision and Pattern Recognition, IEEE Conference on, 2015. 3 +[174] Zuxuan Wu, Tushar Nagarajan, Abhishek Kumar, Steven Rennie, Larry S Davis, Kristen Grauman, and Rogerio Feris. Blockdrop: Dynamic inference paths in residual networks. In CVPR, 2018. 8 +[175] Fei Xia, Amir R. Zamir, Zhi-Yang He, Alexander Sax, Jitendra Malik, and Silvio Savarese. Gibson Env: real-world perception for embodied agents. In CVPR. IEEE, 2018. Gibson license is available at http://svl.stanford.edu/gibson2/assets/GDS_agreement.pdf.3 +[176] J. Xiao, A. Owens, and A. Torralba. Sun3d: A database of big spaces reconstructed using sfm and object labels. In ICCV, 2013. 3 +[177] Frank F Xu, Lei Ji, Botian Shi, Junyi Du, Graham Neubig, Yonatan Bisk, and Nan Duan. A benchmark for structured procedural knowledge extraction from cooking videos. arXiv preprint arXiv:2005.00706, 2020. 8 +[178] Hu Xu, Gargi Ghosh, Po-Yao Huang, Dmytro Okhonko, Armen Aghajanyan, Florian Metze, Luke Zettlemoyer, and Christoph Feichtenhofer. Videoclip: Contrastive pretraining for zero-shot video-text understanding. arXiv preprint arXiv:2109.14084, 2021. 4 +[179] Mingze Xu, Chenyou Fan, Yuchen Wang, Michael S Ryoo, and David J Crandall. Joint person segmentation and identification in synchronized first-and third-person videos. In ECCV, 2018. 4 +[180] Mingze Xu, Yuanjun Xiong, Hao Chen, Xinyu Li, Wei Xia, Zhuowen Tu, and Stefano Soatto. Long short-term transformer for online action detection. Advances in Neural Information Processing Systems, 34:1086-1099, 2021. 49 +[181] Weipeng Xu, Avishek Chatterjee, Michael Zollhoefer, Helge Rhodin, Pascal Fua, Hans-Peter Seidel, and Christian Theobalt. Mo 2 cap 2: Real-time mobile 3d motion capture with a cap-mounted fisheye camera. IEEE transactions on visualization and computer graphics, 25(5):2093-2101, 2019. 9 +[182] Zihui Xue and Kristen Grauman. Learning fine-grained view-invariant representations from unpaired ego-exo videos via temporal alignment. In NeurIPS, 2023. 4, 7 +[183] Lita Yang, Robert M Radway, Yu-Hsin Chen, Tony F Wu, Huichu Liu, Elnaz Ansari, Vikas Chandra, Subhasish Mitra, and Edith Beigné. Three-dimensional stacked neural network accelerator architectures for ar/vr applications. IEEE Micro, 2022. 46 +[184] Huangyue Yu, Minjie Cai, Yunfei Liu, and Feng Lu. What i see is what you see: Joint attention learning for first and third person video co-analysis. In ACM MM, 2019. 4 +[185] Huangyue Yu, Minjie Cai, Yunfei Liu, and Feng Lu. First- and third-person video co-analysis by learning spatial-temporal joint attention. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2020. 4 +[186] Xumin Yu, Yongming Rao, Wenliang Zhao, Jiwen Lu, and Jie Zhou. Group-aware contrastive regression for action quality assessment. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 7919-7928, 2021. 8 + +[187] Ye Yuan and Kris Kitani. 3d ego-pose estimation via imitation learning. In Proceedings of the European Conference on Computer Vision (ECCV), 2018. 9 +[188] Ye Yuan and Kris Kitani. Ego-pose estimation and forecasting as real-time pd control. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2019. 9 +[189] Chenlin Zhang, Jianxin Wu, and Yin Li. Actionformer: Localizing moments of actions with transformers. arXiv preprint arXiv:2202.07925, 2022. 21 +[190] Chen-Lin Zhang, Jianxin Wu, and Yin Li. Actionformer: Localizing moments of actions with transformers. In European Conference on Computer Vision, pages 492-510, 2022. 54, 55, 57 +[191] Qiang Zhang and Baoxin Li. Relative hidden markov models for evaluating motion skill. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2013. 8 +[192] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, 2018. 39 +[193] Siwei Zhang, Qianli Ma, Yan Zhang, Zhiyin Qian, Taein Kwon, Marc Pollefeys, Federica Bogo, and Siyu Tang. Egobody: Human body shape and motion of interacting people from head-mounted devices. In ECCV, 2022. 3 +[194] Shiyi Zhang, Wenxun Dai, Sujia Wang, Xiangwei Shen, Jiwen Lu, Jie Zhou, and Yansong Tang. Logo: A long-form video dataset for group action quality assessment. In CVPR, 2023. 3 +[195] Xiangyu Zhang, Xinyu Zhou, Mengxiao Lin, and Jian Sun. Shufflenet: An extremely efficient convolutional neural network for mobile devices. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6848-6856, 2018. 8 +[196] He Zhao, Isma Hadji, Nikita Dvornik, Konstantinos G Derpanis, Richard P Wildes, and Allan D Jepson. P3iv: Probabilistic procedure planning from instructional videos with weak supervision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2938-2948, 2022. 4 +[197] Weixi Zhao, Weiqiang Wang, and Yunjie Tian. Graformer: Graph-oriented transformer for 3d pose estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20438-20447, 2022. 58 +[198] Yue Zhao and Philipp Krahenbuhl. Real-time online video detection with temporal smoothing transformers. In European Conference on Computer Vision, 2022. 49 +[199] Yue Zhao, Ishan Misra, Philipp Krahenbuhl, and Rohit Girdhar. Learning video representations from large language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6586-6597, 2023. 26, 46, 47, 48 +[200] Ce Zheng, Xianpeng Liu, Guo-Jun Qi, and Chen Chen. Potter: Pooling attention transformer for efficient human mesh recovery. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1611-1620, 2023. 59 + +[201] Yiwu Zhong, Licheng Yu, Yang Bai, Shangwen Li, Xueling Yan, and Yin Li. Learning procedure-aware video representation from instructional videos and their narrations. arXiv preprint arXiv:2303.17839, 2023. 4 +[202] Honglu Zhou, Roberto Martin-Martin, Mubbasir Kapadia, Silvio Savarese, and Juan Carlos Niebles. Procedure-aware pretraining for instructional video understanding. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023. 4, 8 +[203] Honglu Zhou, Roberto Martin-Martin, Mubbasir Kapadia, Silvio Savarese, and Juan Carlos Niebles. Procedure-aware pretraining for instructional video understanding. In CVPR, pages 10727-10738, 2023. 49 +[204] L. Zhou, N. Louis, and J. Corso. Weakly-supervised video object grounding from text by loss weighting and object interaction. In BMVC, 2018. 2, 4 +[205] Luowei Zhou, Chenliang Xu, and Jason J Corso. Towards automatic learning of procedures from web instructional videos. In AAAI, 2018. 4, 7 +[206] Michael Zhu and Suyog Gupta. To prune, or not to prune: exploring the efficacy of pruning for model compression. arXiv preprint arXiv:1710.01878, 2017. 8 +[207] Dimitri Zhukov, Jean-Baptiste Alayrac, Ramazan Gokberk Cinbis, David Fouhey, Ivan Laptev, and Josef Sivic. Cross-task weakly supervised learning from instructional videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2019. 2, 4, 7 +[208] Aneeq Zia, Yachna Sharma, Vinay Bettadapura, Eric L. Sarin, and Irfan A. Essa. Video and accelerometer-based motion analysis for automated surgical skills assessment. CoRR, abs/1702.07772, 2017. 8 \ No newline at end of file diff --git a/2024/Ego-Exo4D_ Understanding Skilled Human Activity from First- and Third-Person Perspectives/images.zip b/2024/Ego-Exo4D_ Understanding Skilled Human Activity from First- and Third-Person Perspectives/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..d3c26b5ee9ac4bdf9ea83e59b6735ff3909768dd --- /dev/null +++ b/2024/Ego-Exo4D_ Understanding Skilled Human Activity from First- and Third-Person Perspectives/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a8de9d02de70be0b59c1f219192de36f41043849d52c432ba674f7c0da5d0c8f +size 496880 diff --git a/2024/Ego-Exo4D_ Understanding Skilled Human Activity from First- and Third-Person Perspectives/layout.json b/2024/Ego-Exo4D_ Understanding Skilled Human Activity from First- and Third-Person Perspectives/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..8122cc3078dc42f76aca56c66a81425d6b4bfa39 --- /dev/null +++ b/2024/Ego-Exo4D_ Understanding Skilled Human Activity from First- and Third-Person Perspectives/layout.json @@ -0,0 +1,18007 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 136, + 103, + 457, + 138 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 103, + 457, + 138 + ], + "spans": [ + { + "bbox": [ + 136, + 103, + 457, + 138 + ], + "type": "text", + "content": "Ego-Exo4D: Understanding Skilled Human Activity from First- and Third-Person Perspectives" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 52, + 170, + 541, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 52, + 170, + 541, + 437 + ], + "spans": [ + { + "bbox": [ + 52, + 170, + 541, + 437 + ], + "type": "text", + "content": "Kristen Grauman" + }, + { + "bbox": [ + 52, + 170, + 541, + 437 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 52, + 170, + 541, + 437 + ], + "type": "text", + "content": ", Andrew Westbury" + }, + { + "bbox": [ + 52, + 170, + 541, + 437 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 52, + 170, + 541, + 437 + ], + "type": "text", + "content": ", Lorenzo Torresani" + }, + { + "bbox": [ + 52, + 170, + 541, + 437 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 52, + 170, + 541, + 437 + ], + "type": "text", + "content": ", Kris Kitani" + }, + { + "bbox": [ + 52, + 170, + 541, + 437 + ], + "type": "inline_equation", + "content": "^{1,3}" + }, + { + "bbox": [ + 52, + 170, + 541, + 437 + ], + "type": "text", + "content": ", Jitendra Malik" + }, + { + "bbox": [ + 52, + 170, + 541, + 437 + ], + "type": "inline_equation", + "content": "^{1,4}" + }, + { + "bbox": [ + 52, + 170, + 541, + 437 + ], + "type": "text", + "content": ", Triantafyllos Afouras" + }, + { + "bbox": [ + 52, + 170, + 541, + 437 + ], + "type": "inline_equation", + "content": "^{*1}" + }, + { + "bbox": [ + 52, + 170, + 541, + 437 + ], + "type": "text", + "content": ", Kumar Ashutosh" + }, + { + "bbox": [ + 52, + 170, + 541, + 437 + ], + "type": "inline_equation", + "content": "^{*1,2}" + }, + { + "bbox": [ + 52, + 170, + 541, + 437 + ], + "type": "text", + "content": ", Vijay Baiyya" + }, + { + "bbox": [ + 52, + 170, + 541, + 437 + ], + "type": "inline_equation", + "content": "^{5}" + }, + { + "bbox": [ + 52, + 170, + 541, + 437 + ], + "type": "text", + "content": ", Siddhant Bansal" + }, + { + "bbox": [ + 52, + 170, + 541, + 437 + ], + "type": "inline_equation", + "content": "^{6,7}" + }, + { + "bbox": [ + 52, + 170, + 541, + 437 + ], + "type": "text", + "content": ", Bikram Boote" + }, + { + "bbox": [ + 52, + 170, + 541, + 437 + ], + "type": "inline_equation", + "content": "^{8}" + }, + { + "bbox": [ + 52, + 170, + 541, + 437 + ], + "type": "text", + "content": ", Eugene Byrne" + }, + { + "bbox": [ + 52, + 170, + 541, + 437 + ], + "type": "inline_equation", + "content": "^{*1,9}" + }, + { + "bbox": [ + 52, + 170, + 541, + 437 + ], + "type": "text", + "content": ", Zach Chavis" + }, + { + "bbox": [ + 52, + 170, + 541, + 437 + ], + "type": "inline_equation", + "content": "^{*10}" + }, + { + "bbox": [ + 52, + 170, + 541, + 437 + ], + "type": "text", + "content": ", Joya Chen" + }, + { + "bbox": [ + 52, + 170, + 541, + 437 + ], + "type": "inline_equation", + "content": "^{*11}" + }, + { + "bbox": [ + 52, + 170, + 541, + 437 + ], + "type": "text", + "content": ", Feng Cheng" + }, + { + "bbox": [ + 52, + 170, + 541, + 437 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 52, + 170, + 541, + 437 + ], + "type": "text", + "content": ", Fu-Jen Chu" + }, + { + "bbox": [ + 52, + 170, + 541, + 437 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 52, + 170, + 541, + 437 + ], + "type": "text", + "content": ", Sean Crane" + }, + { + "bbox": [ + 52, + 170, + 541, + 437 + ], + "type": "inline_equation", + "content": "^{9}" + }, + { + "bbox": [ + 52, + 170, + 541, + 437 + ], + "type": "text", + "content": ", Avijit Dasgupta" + }, + { + "bbox": [ + 52, + 170, + 541, + 437 + ], + "type": "inline_equation", + "content": "^{7}" + }, + { + "bbox": [ + 52, + 170, + 541, + 437 + ], + "type": "text", + "content": ", Jing Dong" + }, + { + "bbox": [ + 52, + 170, + 541, + 437 + ], + "type": "inline_equation", + "content": "^{5}" + }, + { + "bbox": [ + 52, + 170, + 541, + 437 + ], + "type": "text", + "content": ", Maria Escobar" + }, + { + "bbox": [ + 52, + 170, + 541, + 437 + ], + "type": "inline_equation", + "content": "^{*12}" + }, + { + "bbox": [ + 52, + 170, + 541, + 437 + ], + "type": "text", + "content": ", Cristian Forigua" + }, + { + "bbox": [ + 52, + 170, + 541, + 437 + ], + "type": "inline_equation", + "content": "^{*12}" + }, + { + "bbox": [ + 52, + 170, + 541, + 437 + ], + "type": "text", + "content": ", Abrham Gebreselasie" + }, + { + "bbox": [ + 52, + 170, + 541, + 437 + ], + "type": "inline_equation", + "content": "^{9}" + }, + { + "bbox": [ + 52, + 170, + 541, + 437 + ], + "type": "text", + "content": ", Sanjay Haresh" + }, + { + "bbox": [ + 52, + 170, + 541, + 437 + ], + "type": "inline_equation", + "content": "^{13}" + }, + { + "bbox": [ + 52, + 170, + 541, + 437 + ], + "type": "text", + "content": ", Jing Huang" + }, + { + "bbox": [ + 52, + 170, + 541, + 437 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 52, + 170, + 541, + 437 + ], + "type": "text", + "content": ", Md Mohaiminul Islam" + }, + { + "bbox": [ + 52, + 170, + 541, + 437 + ], + "type": "inline_equation", + "content": "^{14}" + }, + { + "bbox": [ + 52, + 170, + 541, + 437 + ], + "type": "text", + "content": ", Suyog Jain" + }, + { + "bbox": [ + 52, + 170, + 541, + 437 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 52, + 170, + 541, + 437 + ], + "type": "text", + "content": ", Rawal Khirodkar" + }, + { + "bbox": [ + 52, + 170, + 541, + 437 + ], + "type": "inline_equation", + "content": "^{9}" + }, + { + "bbox": [ + 52, + 170, + 541, + 437 + ], + "type": "text", + "content": ", Devansh Kukreja" + }, + { + "bbox": [ + 52, + 170, + 541, + 437 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 52, + 170, + 541, + 437 + ], + "type": "text", + "content": ", Kevin J Liang" + }, + { + "bbox": [ + 52, + 170, + 541, + 437 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 52, + 170, + 541, + 437 + ], + "type": "text", + "content": ", Jia-Wei Liu" + }, + { + "bbox": [ + 52, + 170, + 541, + 437 + ], + "type": "inline_equation", + "content": "^{11}" + }, + { + "bbox": [ + 52, + 170, + 541, + 437 + ], + "type": "text", + "content": ", Sagnik Majumder" + }, + { + "bbox": [ + 52, + 170, + 541, + 437 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 52, + 170, + 541, + 437 + ], + "type": "text", + "content": ", Yongsen Mao" + }, + { + "bbox": [ + 52, + 170, + 541, + 437 + ], + "type": "inline_equation", + "content": "^{13}" + }, + { + "bbox": [ + 52, + 170, + 541, + 437 + ], + "type": "text", + "content": ", Miguel Martin" + }, + { + "bbox": [ + 52, + 170, + 541, + 437 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 52, + 170, + 541, + 437 + ], + "type": "text", + "content": ", Effrosyni Mavroudi" + }, + { + "bbox": [ + 52, + 170, + 541, + 437 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 52, + 170, + 541, + 437 + ], + "type": "text", + "content": ", Tushar Nagarajan" + }, + { + "bbox": [ + 52, + 170, + 541, + 437 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 52, + 170, + 541, + 437 + ], + "type": "text", + "content": ", Francesco Ragusa" + }, + { + "bbox": [ + 52, + 170, + 541, + 437 + ], + "type": "inline_equation", + "content": "^{15}" + }, + { + "bbox": [ + 52, + 170, + 541, + 437 + ], + "type": "text", + "content": ", Santhosh Kumar Ramakrishnan" + }, + { + "bbox": [ + 52, + 170, + 541, + 437 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 52, + 170, + 541, + 437 + ], + "type": "text", + "content": ", Luigi Seminara" + }, + { + "bbox": [ + 52, + 170, + 541, + 437 + ], + "type": "inline_equation", + "content": "^{15}" + }, + { + "bbox": [ + 52, + 170, + 541, + 437 + ], + "type": "text", + "content": ", Arjun Somayazulu" + }, + { + "bbox": [ + 52, + 170, + 541, + 437 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 52, + 170, + 541, + 437 + ], + "type": "text", + "content": ", Yale Song" + }, + { + "bbox": [ + 52, + 170, + 541, + 437 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 52, + 170, + 541, + 437 + ], + "type": "text", + "content": ", Shan Su" + }, + { + "bbox": [ + 52, + 170, + 541, + 437 + ], + "type": "inline_equation", + "content": "^{16}" + }, + { + "bbox": [ + 52, + 170, + 541, + 437 + ], + "type": "text", + "content": ", Zihui Xue" + }, + { + "bbox": [ + 52, + 170, + 541, + 437 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 52, + 170, + 541, + 437 + ], + "type": "text", + "content": ", Edward Zhang" + }, + { + "bbox": [ + 52, + 170, + 541, + 437 + ], + "type": "inline_equation", + "content": "^{16}" + }, + { + "bbox": [ + 52, + 170, + 541, + 437 + ], + "type": "text", + "content": ", Jinxu Zhang" + }, + { + "bbox": [ + 52, + 170, + 541, + 437 + ], + "type": "inline_equation", + "content": "^{16}" + }, + { + "bbox": [ + 52, + 170, + 541, + 437 + ], + "type": "text", + "content": ", Angela Castillo" + }, + { + "bbox": [ + 52, + 170, + 541, + 437 + ], + "type": "inline_equation", + "content": "^{12}" + }, + { + "bbox": [ + 52, + 170, + 541, + 437 + ], + "type": "text", + "content": ", Changan Chen" + }, + { + "bbox": [ + 52, + 170, + 541, + 437 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 52, + 170, + 541, + 437 + ], + "type": "text", + "content": ", Xinzhu Fu" + }, + { + "bbox": [ + 52, + 170, + 541, + 437 + ], + "type": "inline_equation", + "content": "^{11}" + }, + { + "bbox": [ + 52, + 170, + 541, + 437 + ], + "type": "text", + "content": ", Ryosuke Furuta" + }, + { + "bbox": [ + 52, + 170, + 541, + 437 + ], + "type": "inline_equation", + "content": "^{17}" + }, + { + "bbox": [ + 52, + 170, + 541, + 437 + ], + "type": "text", + "content": ", Cristina González" + }, + { + "bbox": [ + 52, + 170, + 541, + 437 + ], + "type": "inline_equation", + "content": "^{12}" + }, + { + "bbox": [ + 52, + 170, + 541, + 437 + ], + "type": "text", + "content": ", Prince Gupta" + }, + { + "bbox": [ + 52, + 170, + 541, + 437 + ], + "type": "inline_equation", + "content": "^{5}" + }, + { + "bbox": [ + 52, + 170, + 541, + 437 + ], + "type": "text", + "content": ", Jiabo Hu" + }, + { + "bbox": [ + 52, + 170, + 541, + 437 + ], + "type": "inline_equation", + "content": "^{18}" + }, + { + "bbox": [ + 52, + 170, + 541, + 437 + ], + "type": "text", + "content": ", Yifei Huang" + }, + { + "bbox": [ + 52, + 170, + 541, + 437 + ], + "type": "inline_equation", + "content": "^{17}" + }, + { + "bbox": [ + 52, + 170, + 541, + 437 + ], + "type": "text", + "content": ", Yiming Huang" + }, + { + "bbox": [ + 52, + 170, + 541, + 437 + ], + "type": "inline_equation", + "content": "^{16}" + }, + { + "bbox": [ + 52, + 170, + 541, + 437 + ], + "type": "text", + "content": ", Weslie Khoo" + }, + { + "bbox": [ + 52, + 170, + 541, + 437 + ], + "type": "inline_equation", + "content": "^{19}" + }, + { + "bbox": [ + 52, + 170, + 541, + 437 + ], + "type": "text", + "content": ", Anush Kumar" + }, + { + "bbox": [ + 52, + 170, + 541, + 437 + ], + "type": "inline_equation", + "content": "^{10}" + }, + { + "bbox": [ + 52, + 170, + 541, + 437 + ], + "type": "text", + "content": ", Robert Kuo" + }, + { + "bbox": [ + 52, + 170, + 541, + 437 + ], + "type": "inline_equation", + "content": "^{18}" + }, + { + "bbox": [ + 52, + 170, + 541, + 437 + ], + "type": "text", + "content": ", Sach Lakhavani" + }, + { + "bbox": [ + 52, + 170, + 541, + 437 + ], + "type": "inline_equation", + "content": "^{5}" + }, + { + "bbox": [ + 52, + 170, + 541, + 437 + ], + "type": "text", + "content": ", Miao Liu" + }, + { + "bbox": [ + 52, + 170, + 541, + 437 + ], + "type": "inline_equation", + "content": "^{18}" + }, + { + "bbox": [ + 52, + 170, + 541, + 437 + ], + "type": "text", + "content": ", Mi Luo" + }, + { + "bbox": [ + 52, + 170, + 541, + 437 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 52, + 170, + 541, + 437 + ], + "type": "text", + "content": ", Zhengyi Luo" + }, + { + "bbox": [ + 52, + 170, + 541, + 437 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 52, + 170, + 541, + 437 + ], + "type": "text", + "content": ", Brighid Meredith" + }, + { + "bbox": [ + 52, + 170, + 541, + 437 + ], + "type": "inline_equation", + "content": "^{18}" + }, + { + "bbox": [ + 52, + 170, + 541, + 437 + ], + "type": "text", + "content": ", Austin Miller" + }, + { + "bbox": [ + 52, + 170, + 541, + 437 + ], + "type": "inline_equation", + "content": "^{18}" + }, + { + "bbox": [ + 52, + 170, + 541, + 437 + ], + "type": "text", + "content": ", Oluwatuminiu Oguntola" + }, + { + "bbox": [ + 52, + 170, + 541, + 437 + ], + "type": "inline_equation", + "content": "^{14}" + }, + { + "bbox": [ + 52, + 170, + 541, + 437 + ], + "type": "text", + "content": ", Xiaqing Pan" + }, + { + "bbox": [ + 52, + 170, + 541, + 437 + ], + "type": "inline_equation", + "content": "^{5}" + }, + { + "bbox": [ + 52, + 170, + 541, + 437 + ], + "type": "text", + "content": ", Penny Peng" + }, + { + "bbox": [ + 52, + 170, + 541, + 437 + ], + "type": "inline_equation", + "content": "^{18}" + }, + { + "bbox": [ + 52, + 170, + 541, + 437 + ], + "type": "text", + "content": ", Shraman Pramanick" + }, + { + "bbox": [ + 52, + 170, + 541, + 437 + ], + "type": "inline_equation", + "content": "^{20}" + }, + { + "bbox": [ + 52, + 170, + 541, + 437 + ], + "type": "text", + "content": ", Merey Ramazanova" + }, + { + "bbox": [ + 52, + 170, + 541, + 437 + ], + "type": "inline_equation", + "content": "^{21}" + }, + { + "bbox": [ + 52, + 170, + 541, + 437 + ], + "type": "text", + "content": ", Fiona Ryan" + }, + { + "bbox": [ + 52, + 170, + 541, + 437 + ], + "type": "inline_equation", + "content": "^{22}" + }, + { + "bbox": [ + 52, + 170, + 541, + 437 + ], + "type": "text", + "content": ", Wei Shan" + }, + { + "bbox": [ + 52, + 170, + 541, + 437 + ], + "type": "inline_equation", + "content": "^{14}" + }, + { + "bbox": [ + 52, + 170, + 541, + 437 + ], + "type": "text", + "content": ", Kiran Somasundaram" + }, + { + "bbox": [ + 52, + 170, + 541, + 437 + ], + "type": "inline_equation", + "content": "^{5}" + }, + { + "bbox": [ + 52, + 170, + 541, + 437 + ], + "type": "text", + "content": ", Chenan Song" + }, + { + "bbox": [ + 52, + 170, + 541, + 437 + ], + "type": "inline_equation", + "content": "^{11}" + }, + { + "bbox": [ + 52, + 170, + 541, + 437 + ], + "type": "text", + "content": ", Audrey Southerland" + }, + { + "bbox": [ + 52, + 170, + 541, + 437 + ], + "type": "inline_equation", + "content": "^{22}" + }, + { + "bbox": [ + 52, + 170, + 541, + 437 + ], + "type": "text", + "content": ", Masatoshi Tateno" + }, + { + "bbox": [ + 52, + 170, + 541, + 437 + ], + "type": "inline_equation", + "content": "^{17}" + }, + { + "bbox": [ + 52, + 170, + 541, + 437 + ], + "type": "text", + "content": ", Huiyu Wang" + }, + { + "bbox": [ + 52, + 170, + 541, + 437 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 52, + 170, + 541, + 437 + ], + "type": "text", + "content": ", Yuchen Wang" + }, + { + "bbox": [ + 52, + 170, + 541, + 437 + ], + "type": "inline_equation", + "content": "^{19}" + }, + { + "bbox": [ + 52, + 170, + 541, + 437 + ], + "type": "text", + "content": ", Takuma Yagi" + }, + { + "bbox": [ + 52, + 170, + 541, + 437 + ], + "type": "inline_equation", + "content": "^{17}" + }, + { + "bbox": [ + 52, + 170, + 541, + 437 + ], + "type": "text", + "content": ", Mingfei Yan" + }, + { + "bbox": [ + 52, + 170, + 541, + 437 + ], + "type": "inline_equation", + "content": "^{5}" + }, + { + "bbox": [ + 52, + 170, + 541, + 437 + ], + "type": "text", + "content": ", Xitong Yang" + }, + { + "bbox": [ + 52, + 170, + 541, + 437 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 52, + 170, + 541, + 437 + ], + "type": "text", + "content": ", Zecheng Yu" + }, + { + "bbox": [ + 52, + 170, + 541, + 437 + ], + "type": "inline_equation", + "content": "^{17}" + }, + { + "bbox": [ + 52, + 170, + 541, + 437 + ], + "type": "text", + "content": ", Shengxin Cindy Zha" + }, + { + "bbox": [ + 52, + 170, + 541, + 437 + ], + "type": "inline_equation", + "content": "^{18}" + }, + { + "bbox": [ + 52, + 170, + 541, + 437 + ], + "type": "text", + "content": ", Chen Zhao" + }, + { + "bbox": [ + 52, + 170, + 541, + 437 + ], + "type": "inline_equation", + "content": "^{21}" + }, + { + "bbox": [ + 52, + 170, + 541, + 437 + ], + "type": "text", + "content": ", Ziwei Zhao" + }, + { + "bbox": [ + 52, + 170, + 541, + 437 + ], + "type": "inline_equation", + "content": "^{19}" + }, + { + "bbox": [ + 52, + 170, + 541, + 437 + ], + "type": "text", + "content": ", Zhifan Zhu" + }, + { + "bbox": [ + 52, + 170, + 541, + 437 + ], + "type": "inline_equation", + "content": "^{6}" + }, + { + "bbox": [ + 52, + 170, + 541, + 437 + ], + "type": "text", + "content": ", Jeff Zhuo" + }, + { + "bbox": [ + 52, + 170, + 541, + 437 + ], + "type": "inline_equation", + "content": "^{14}" + }, + { + "bbox": [ + 52, + 170, + 541, + 437 + ], + "type": "text", + "content": ", Pablo Arbeláez†" + }, + { + "bbox": [ + 52, + 170, + 541, + 437 + ], + "type": "inline_equation", + "content": "^{12}" + }, + { + "bbox": [ + 52, + 170, + 541, + 437 + ], + "type": "text", + "content": ", Gedas Bertasius†" + }, + { + "bbox": [ + 52, + 170, + 541, + 437 + ], + "type": "inline_equation", + "content": "^{14}" + }, + { + "bbox": [ + 52, + 170, + 541, + 437 + ], + "type": "text", + "content": ", Dima Damen†" + }, + { + "bbox": [ + 52, + 170, + 541, + 437 + ], + "type": "inline_equation", + "content": "^{6}" + }, + { + "bbox": [ + 52, + 170, + 541, + 437 + ], + "type": "text", + "content": ", Jakob Engel†" + }, + { + "bbox": [ + 52, + 170, + 541, + 437 + ], + "type": "inline_equation", + "content": "^{5}" + }, + { + "bbox": [ + 52, + 170, + 541, + 437 + ], + "type": "text", + "content": ", Giovanni Maria Farinella†" + }, + { + "bbox": [ + 52, + 170, + 541, + 437 + ], + "type": "inline_equation", + "content": "^{15}" + }, + { + "bbox": [ + 52, + 170, + 541, + 437 + ], + "type": "text", + "content": ", Antonino Furnari†" + }, + { + "bbox": [ + 52, + 170, + 541, + 437 + ], + "type": "inline_equation", + "content": "^{15}" + }, + { + "bbox": [ + 52, + 170, + 541, + 437 + ], + "type": "text", + "content": ", Bernard Ghanem†" + }, + { + "bbox": [ + 52, + 170, + 541, + 437 + ], + "type": "inline_equation", + "content": "^{21}" + }, + { + "bbox": [ + 52, + 170, + 541, + 437 + ], + "type": "text", + "content": ", Judy Hoffman†" + }, + { + "bbox": [ + 52, + 170, + 541, + 437 + ], + "type": "inline_equation", + "content": "^{22}" + }, + { + "bbox": [ + 52, + 170, + 541, + 437 + ], + "type": "text", + "content": ", C. V. Jawahar†" + }, + { + "bbox": [ + 52, + 170, + 541, + 437 + ], + "type": "inline_equation", + "content": "^{7}" + }, + { + "bbox": [ + 52, + 170, + 541, + 437 + ], + "type": "text", + "content": ", Richard Newcombe†" + }, + { + "bbox": [ + 52, + 170, + 541, + 437 + ], + "type": "inline_equation", + "content": "^{5}" + }, + { + "bbox": [ + 52, + 170, + 541, + 437 + ], + "type": "text", + "content": ", Hyun Soo Park†" + }, + { + "bbox": [ + 52, + 170, + 541, + 437 + ], + "type": "inline_equation", + "content": "^{10}" + }, + { + "bbox": [ + 52, + 170, + 541, + 437 + ], + "type": "text", + "content": ", James M. Rehg†" + }, + { + "bbox": [ + 52, + 170, + 541, + 437 + ], + "type": "inline_equation", + "content": "^{8}" + }, + { + "bbox": [ + 52, + 170, + 541, + 437 + ], + "type": "text", + "content": ", Yoichi Sato†" + }, + { + "bbox": [ + 52, + 170, + 541, + 437 + ], + "type": "inline_equation", + "content": "^{17}" + }, + { + "bbox": [ + 52, + 170, + 541, + 437 + ], + "type": "text", + "content": ", Manolis Savva†" + }, + { + "bbox": [ + 52, + 170, + 541, + 437 + ], + "type": "inline_equation", + "content": "^{13}" + }, + { + "bbox": [ + 52, + 170, + 541, + 437 + ], + "type": "text", + "content": ", Jianbo Shi†" + }, + { + "bbox": [ + 52, + 170, + 541, + 437 + ], + "type": "inline_equation", + "content": "^{16}" + }, + { + "bbox": [ + 52, + 170, + 541, + 437 + ], + "type": "text", + "content": ", Mike Zheng Shou†" + }, + { + "bbox": [ + 52, + 170, + 541, + 437 + ], + "type": "inline_equation", + "content": "^{11}" + }, + { + "bbox": [ + 52, + 170, + 541, + 437 + ], + "type": "text", + "content": " and Michael Wray†" + }, + { + "bbox": [ + 52, + 170, + 541, + 437 + ], + "type": "inline_equation", + "content": "^{6}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 444, + 547, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 444, + 547, + 502 + ], + "spans": [ + { + "bbox": [ + 47, + 444, + 547, + 502 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 47, + 444, + 547, + 502 + ], + "type": "text", + "content": "FAIR, Meta, " + }, + { + "bbox": [ + 47, + 444, + 547, + 502 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 47, + 444, + 547, + 502 + ], + "type": "text", + "content": "University of Texas at Austin, " + }, + { + "bbox": [ + 47, + 444, + 547, + 502 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 47, + 444, + 547, + 502 + ], + "type": "text", + "content": "Carnegie Mellon University, " + }, + { + "bbox": [ + 47, + 444, + 547, + 502 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 47, + 444, + 547, + 502 + ], + "type": "text", + "content": "University of California, Berkeley, " + }, + { + "bbox": [ + 47, + 444, + 547, + 502 + ], + "type": "inline_equation", + "content": "^{5}" + }, + { + "bbox": [ + 47, + 444, + 547, + 502 + ], + "type": "text", + "content": "Project Aria, Meta, " + }, + { + "bbox": [ + 47, + 444, + 547, + 502 + ], + "type": "inline_equation", + "content": "^{6}" + }, + { + "bbox": [ + 47, + 444, + 547, + 502 + ], + "type": "text", + "content": "University of Bristol, " + }, + { + "bbox": [ + 47, + 444, + 547, + 502 + ], + "type": "inline_equation", + "content": "^{7}" + }, + { + "bbox": [ + 47, + 444, + 547, + 502 + ], + "type": "text", + "content": "International Institute of Information Technology, Hyderabad, " + }, + { + "bbox": [ + 47, + 444, + 547, + 502 + ], + "type": "inline_equation", + "content": "^{8}" + }, + { + "bbox": [ + 47, + 444, + 547, + 502 + ], + "type": "text", + "content": "University of Illinois, Urbana Champaign, " + }, + { + "bbox": [ + 47, + 444, + 547, + 502 + ], + "type": "inline_equation", + "content": "^{9}" + }, + { + "bbox": [ + 47, + 444, + 547, + 502 + ], + "type": "text", + "content": "Carnegie Mellon University, " + }, + { + "bbox": [ + 47, + 444, + 547, + 502 + ], + "type": "inline_equation", + "content": "^{10}" + }, + { + "bbox": [ + 47, + 444, + 547, + 502 + ], + "type": "text", + "content": "University of Minnesota, " + }, + { + "bbox": [ + 47, + 444, + 547, + 502 + ], + "type": "inline_equation", + "content": "^{11}" + }, + { + "bbox": [ + 47, + 444, + 547, + 502 + ], + "type": "text", + "content": "National University of Singapore, " + }, + { + "bbox": [ + 47, + 444, + 547, + 502 + ], + "type": "inline_equation", + "content": "^{12}" + }, + { + "bbox": [ + 47, + 444, + 547, + 502 + ], + "type": "text", + "content": "Universidad de los Andes, " + }, + { + "bbox": [ + 47, + 444, + 547, + 502 + ], + "type": "inline_equation", + "content": "^{13}" + }, + { + "bbox": [ + 47, + 444, + 547, + 502 + ], + "type": "text", + "content": "Simon Fraser University, " + }, + { + "bbox": [ + 47, + 444, + 547, + 502 + ], + "type": "inline_equation", + "content": "^{14}" + }, + { + "bbox": [ + 47, + 444, + 547, + 502 + ], + "type": "text", + "content": "University of North Carolina, Chapel Hill, " + }, + { + "bbox": [ + 47, + 444, + 547, + 502 + ], + "type": "inline_equation", + "content": "^{15}" + }, + { + "bbox": [ + 47, + 444, + 547, + 502 + ], + "type": "text", + "content": "University of Catania, " + }, + { + "bbox": [ + 47, + 444, + 547, + 502 + ], + "type": "inline_equation", + "content": "^{16}" + }, + { + "bbox": [ + 47, + 444, + 547, + 502 + ], + "type": "text", + "content": "University of Pennsylvania, " + }, + { + "bbox": [ + 47, + 444, + 547, + 502 + ], + "type": "inline_equation", + "content": "^{17}" + }, + { + "bbox": [ + 47, + 444, + 547, + 502 + ], + "type": "text", + "content": "University of Tokyo, " + }, + { + "bbox": [ + 47, + 444, + 547, + 502 + ], + "type": "inline_equation", + "content": "^{18}" + }, + { + "bbox": [ + 47, + 444, + 547, + 502 + ], + "type": "text", + "content": "Meta, " + }, + { + "bbox": [ + 47, + 444, + 547, + 502 + ], + "type": "inline_equation", + "content": "^{19}" + }, + { + "bbox": [ + 47, + 444, + 547, + 502 + ], + "type": "text", + "content": "Indiana University, " + }, + { + "bbox": [ + 47, + 444, + 547, + 502 + ], + "type": "inline_equation", + "content": "^{20}" + }, + { + "bbox": [ + 47, + 444, + 547, + 502 + ], + "type": "text", + "content": "Johns Hopkins University, " + }, + { + "bbox": [ + 47, + 444, + 547, + 502 + ], + "type": "inline_equation", + "content": "^{21}" + }, + { + "bbox": [ + 47, + 444, + 547, + 502 + ], + "type": "text", + "content": "King Abdullah University of Science and Technology, " + }, + { + "bbox": [ + 47, + 444, + 547, + 502 + ], + "type": "inline_equation", + "content": "^{22}" + }, + { + "bbox": [ + 47, + 444, + 547, + 502 + ], + "type": "text", + "content": "Georgia Tech" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 143, + 519, + 192, + 532 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 519, + 192, + 532 + ], + "spans": [ + { + "bbox": [ + 143, + 519, + 192, + 532 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 45, + 534, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 534, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 45, + 534, + 289, + 715 + ], + "type": "text", + "content": "We present Ego-Exo4D, a diverse, large-scale multimodal multiview video dataset and benchmark challenge. Ego-Exo4D centers around simultaneously-captured ego-centric and exocentric video of skilled human activities (e.g., sports, music, dance, bike repair). 740 participants from 13 cities worldwide performed these activities in 123 different natural scene contexts, yielding long-form captures from 1 to 42 minutes each and 1,286 hours of video combined. The multimodal nature of the dataset is unprecedented: the video is accompanied by multichannel audio, eye gaze, 3D point clouds, camera poses, IMU, and multiple paired language descriptions—including a novel \"expert commentary\" done by coaches and teachers and tailored to the skilled-activity domain. To push the frontier of first-person video understanding of skilled hu" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 537, + 547, + 598 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 537, + 547, + 598 + ], + "spans": [ + { + "bbox": [ + 304, + 537, + 547, + 598 + ], + "type": "text", + "content": "man activity, we also present a suite of benchmark tasks and their annotations, including fine-grained activity understanding, proficiency estimation, cross-view translation, and 3D hand/body pose. All resources are open sourced to fuel new research in the community." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 609, + 387, + 621 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 609, + 387, + 621 + ], + "spans": [ + { + "bbox": [ + 306, + 609, + 387, + 621 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 629, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 629, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 629, + 547, + 713 + ], + "type": "text", + "content": "A dancer leaps across a stage; Lionel Messi delivers a precise pass; your grandmother prepares her famous dumplings. We observe and seek human skills in a myriad of settings, from the practical (fixing a bike) to the aspirational (dancing beautifully). What would it mean for AI to understand human skills? And what would it take to get there?" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "19383" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 77, + 69, + 517, + 335 + ], + "blocks": [ + { + "bbox": [ + 77, + 69, + 517, + 335 + ], + "lines": [ + { + "bbox": [ + 77, + 69, + 517, + 335 + ], + "spans": [ + { + "bbox": [ + 77, + 69, + 517, + 335 + ], + "type": "image", + "image_path": "fac5d8236d3d229f991c89ab4fb2140794f3f9860ab60fe5efe2e1344f38e59f.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 342, + 547, + 376 + ], + "lines": [ + { + "bbox": [ + 46, + 342, + 547, + 376 + ], + "spans": [ + { + "bbox": [ + 46, + 342, + 547, + 376 + ], + "type": "text", + "content": "Figure 1. Ego-Exo4D offers egocentric video alongside multiple time-synchronized exocentric video streams for an array of skilled human activities—1,286 hours of ego and exo video in total. The data is both multiview and multimodal, and it is extensively annotated with language, 3D body and hand pose, keysteps, procedural dependencies, and proficiency ratings in support of our proposed benchmark tasks." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 396, + 288, + 504 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 396, + 288, + 504 + ], + "spans": [ + { + "bbox": [ + 46, + 396, + 288, + 504 + ], + "type": "text", + "content": "Advances in AI understanding of human skill could facilitate many applications. In augmented reality (AR), a person wearing smart glasses could quickly pick up new skills with a virtual AI coach that provides real-time guidance. In robot learning, a robot watching people in its environment could acquire new dexterous manipulation skills with less physical experience. In social networks, new communities could form based on how people share their expertise and complementary skills in video." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 507, + 287, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 507, + 287, + 662 + ], + "spans": [ + { + "bbox": [ + 46, + 507, + 287, + 662 + ], + "type": "text", + "content": "We contend that both the egocentric and exocentric viewpoints are critical for capturing human skill. Firstly, the two viewpoints are synergistic. The first-person (ego) perspective captures the details of close-by hand-object interactions and the camera wearer's attention, whereas the third-person (exo) perspective captures the full body pose and surrounding environment context. See Figure 1. Not coincidentally, instructional or \"how-to\" videos often alternate between a third-person view of the demonstrator and a close-up view of their near-field demonstration. For example, a chef may describe their approach and the equipment from an exo view, then cut to clips showing their hands manipulating the ingredients and tools from an ego-like view." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 666, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 666, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 666, + 288, + 715 + ], + "type": "text", + "content": "Secondly, not only are the ego and exo viewpoints synergistic, but there is a need to translate fluently from one to the other when acquiring skill. For example, imagine watching an expert repair a bike tire, juggle a soccer ball," + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 396, + 545, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 396, + 545, + 456 + ], + "spans": [ + { + "bbox": [ + 304, + 396, + 545, + 456 + ], + "type": "text", + "content": "or fold an origami swan—then mapping their steps to your own body. Cognitive science tells us that even from a very young age we can observe others' behavior (exo) and map it onto our own (ego) [42, 108], and this actor-observer translation remains the foundation of visual learning." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 459, + 546, + 626 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 459, + 546, + 626 + ], + "spans": [ + { + "bbox": [ + 304, + 459, + 546, + 626 + ], + "type": "text", + "content": "Realizing this potential, however, is not possible using today's datasets and learning paradigms. Existing datasets comprised of both ego and exo views (i.e., ego-exo) are few [76, 77, 127, 139, 145], small in scale, lack synchronization across cameras, and/or are too staged or curated to be resilient to the diversity of the real world. Thus the current literature for activity understanding primarily attends to either the ego [28, 47] or exo [48, 67, 105, 149] view, leaving the ability to move fluidly between the first- and third-person perspectives out of reach. Instructional video datasets [103, 159, 204, 207] offer a compelling window into skilled human activity, but (like the above) are limited to single-viewpoint video, whether purely exocentric or mixed with \"ego-like\" views at certain time points." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 630, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 630, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 630, + 547, + 715 + ], + "type": "text", + "content": "We introduce Ego-Exo4D, a foundational dataset to support research on ego-exo video learning and multimodal perception. The result of a two-year effort by a consortium of 15 research institutions, Ego-Exo4D is a first-of-its-kind large-scale multimodal multiview dataset and benchmark suite. It constitutes the largest public dataset of timesynchronized first- and third- person video, captured by 740" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "19384" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 168 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 168 + ], + "type": "text", + "content": "diverse camera wearers in 123 distinct scenes and 13 cities worldwide. For every sequence, Ego-Exo4D provides both the camera wearer's egocentric video, as well as multiple (4-5) exocentric videos from tripod places around the camera wearer. All views are time-synchronized and precisely localized in a metric, gravity-aligned frame of reference. The total collection has 1,286 hours of video and 5,035 instances, each spanning 1 to " + }, + { + "bbox": [ + 46, + 72, + 287, + 168 + ], + "type": "inline_equation", + "content": "42\\mathrm{min}" + }, + { + "bbox": [ + 46, + 72, + 287, + 168 + ], + "type": "text", + "content": " of continuous capture." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 172, + 287, + 269 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 172, + 287, + 269 + ], + "spans": [ + { + "bbox": [ + 46, + 172, + 287, + 269 + ], + "type": "text", + "content": "Ego-Exo4D focuses on skilled single-person activities. The 740 participants perform skilled physical and/or procedural activities—dance, soccer, basketball, bouldering, music, cooking, bike repair, health care—in an unscripted manner and in natural settings (e.g., gym, soccer field, kitchens, bike shops, etc.), exhibiting a variety of skill levels from novice to expert. All video is recorded with rigorous privacy and ethics policies and formal consent of participants." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 273, + 288, + 476 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 273, + 288, + 476 + ], + "spans": [ + { + "bbox": [ + 46, + 273, + 288, + 476 + ], + "type": "text", + "content": "Ego-Exo4D is not only multiview, it is also multimodal. Captured with the unique open-source Aria glasses [38], all ego video is accompanied by 7-channel audio, IMU, eye gaze, both RGB and two grayscale SLAM cameras, and 3D environment point clouds. Additionally, Ego-Exo4D provides multiple new video-language resources, all time indexed: first-person narrations by the camera wearers describing their own actions; third-person play-by-play descriptions of every camera wearer action; and third-person spoken expert commentary critiquing their performance. The latter is particularly novel: performed by domain-specific experienced coaches and teachers, it focuses on how an activity is executed rather than merely what is being done, surfacing subtleties in skilled execution not perceivable by the untrained eye. To our knowledge, there is no prior video resource with such extensive and high quality multimodal data." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 481, + 287, + 649 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 481, + 287, + 649 + ], + "spans": [ + { + "bbox": [ + 46, + 481, + 287, + 649 + ], + "type": "text", + "content": "Alongside this data, we introduce benchmarks for foundational tasks for ego-exo video. We propose four families of tasks: 1) ego-exo relation, for relating the actions of a teacher (exo) to a learner (ego) by estimating semantic correspondences and translating viewpoints; 2) ego(-exo) recognition, for recognizing fine-grained keysteps and task structure; 3) ego(-exo) proficiency estimation, for inferring how well a person is executing a skill; and 4) ego pose, for recovering skilled 3D body and hand movements from ego-video. We provide annotations for each task—the result of more than 200,000 hours of annotator effort. To kick-start work in these new challenges, we also develop baseline models and report their results (Appendices). We are hosting the first public benchmark challenges in 2024." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 654, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 654, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 654, + 287, + 715 + ], + "type": "text", + "content": "In summary, Ego-Exo4D is the community's first diverse, large-scale multimodal multiview video resource. We have open sourced all the data, annotations, camera rig protocol, and benchmarks. With this release, we aim to fuel new research in ego-exo, multimodal activity, and beyond." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 306, + 71, + 389, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 71, + 389, + 83 + ], + "spans": [ + { + "bbox": [ + 306, + 71, + 389, + 83 + ], + "type": "text", + "content": "2. Related work" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 305, + 91, + 545, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 91, + 545, + 140 + ], + "spans": [ + { + "bbox": [ + 305, + 91, + 545, + 140 + ], + "type": "text", + "content": "Next we review prior work in datasets, human skill, and cross-view analysis. Section 5 will discuss related work for each benchmark task. Table 2 in Appendix 9 summarizes Ego-Exo4D's properties vs. existing datasets." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 147, + 546, + 255 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 147, + 546, + 255 + ], + "spans": [ + { + "bbox": [ + 305, + 147, + 546, + 255 + ], + "type": "text", + "content": "Egocentric datasets There has been a surge of interest in egocentric video understanding, facilitated by recent egovideo datasets showing unscripted daily-life activity as in Ego4D [47], EPIC-Kitchens [27, 28, 163], UT Ego [78], ADL [119], and KrishnaCam [147], or procedural activities as in EGTea [81], AssistQ [172], Meccano [126], CMUMMAC [77], and EgoProcel [10]. Unlike any of the above, Ego-Exo4D focuses on multimodal ego and exo capture, and it is focused on the domain of skilled activities." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 262, + 546, + 490 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 262, + 546, + 490 + ], + "spans": [ + { + "bbox": [ + 305, + 262, + 546, + 490 + ], + "type": "text", + "content": "Multiview and ego-exo datasets Most existing multiview datasets focus on static scenes [20, 128, 151, 175, 176] and objects [133, 173], with limited (exo only) multiview human activity [26, 169]. CMU-MMAC [77] and CharadesEgo [145] are early efforts to capture both ego and exo video. CMU-MMAC [77] features 43 participants in mocap suits who cook 5 recipes in a lab kitchen. In CharadesEgo [145], 71 Mechanical Turkers record 34 hours of scripted scenarios (e.g., \"type on laptop, then pick up a pillow\") from the ego and exo perspectives sequentially, yielding unsynchronized videos with non-exact activity matches. More recent ego-exo efforts focus on specific activities in one or two environments. Assembly101 [139] and H2O [76] provide time-synced ego and exo video at a lab tabletop where people assemble toy cars or manipulate handheld objects, with 53 and 4 participants, and 513 and 5 hours of footage, respectively. Homage [127] provides 30 hours of ego-exo video from 27 participants in 2 homes doing household activities like laundry." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 491, + 545, + 669 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 491, + 545, + 669 + ], + "spans": [ + { + "bbox": [ + 305, + 491, + 545, + 669 + ], + "type": "text", + "content": "Compared to any of the prior efforts, Ego-Exo4D offers an order of magnitude more participants, diverse locations, and hours of footage (740 participants, 123 unique scenes, 13 cities, 1,286 hours). Importantly, our focus on skilled tasks takes the participants out of the lab or home and into settings like soccer fields, dance studios, rock climbing walls, and bike repair shops. Such activities also yield a wide variety of full body poses and movements within the scene, beyond using objects at a tabletop. This variety means Ego-Exo4D augments existing 3D human body pose datasets [49, 66, 68, 80, 193]. Finally, compared to any prior ego-exo resource, Ego-Exo4D's suite of modalities and benchmark tasks are novel and will expand the research directions the community can take for egocentric and/or exocentric video understanding." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 677, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 677, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 677, + 546, + 713 + ], + "type": "text", + "content": "Human skill and video learning Analyzing skill and action quality has received limited attention [12, 34, 35, 113, 120, 194]. Research in instructional or \"how-to\" videos is" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19385" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 55, + 77, + 80, + 103 + ], + "blocks": [ + { + "bbox": [ + 55, + 77, + 80, + 103 + ], + "lines": [ + { + "bbox": [ + 55, + 77, + 80, + 103 + ], + "spans": [ + { + "bbox": [ + 55, + 77, + 80, + 103 + ], + "type": "image", + "image_path": "8e9f0c174b11764e0f15c304b5c22610240f5992f86298476a8de14579f1d267.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 106, + 80, + 114 + ], + "lines": [ + { + "bbox": [ + 55, + 106, + 80, + 114 + ], + "spans": [ + { + "bbox": [ + 55, + 106, + 80, + 114 + ], + "type": "text", + "content": "Cooking" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 124, + 71, + 216, + 115 + ], + "blocks": [ + { + "bbox": [ + 89, + 73, + 117, + 80 + ], + "lines": [ + { + "bbox": [ + 89, + 73, + 117, + 80 + ], + "spans": [ + { + "bbox": [ + 89, + 73, + 117, + 80 + ], + "type": "text", + "content": "x678" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 90, + 82, + 116, + 89 + ], + "lines": [ + { + "bbox": [ + 90, + 82, + 116, + 89 + ], + "spans": [ + { + "bbox": [ + 90, + 82, + 116, + 89 + ], + "type": "text", + "content": "x173" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 90, + 90, + 116, + 97 + ], + "lines": [ + { + "bbox": [ + 90, + 90, + 116, + 97 + ], + "spans": [ + { + "bbox": [ + 90, + 90, + 116, + 97 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 90, + 98, + 116, + 105 + ], + "lines": [ + { + "bbox": [ + 90, + 98, + 116, + 105 + ], + "spans": [ + { + "bbox": [ + 90, + 98, + 116, + 105 + ], + "type": "text", + "content": "Mx65" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 90, + 106, + 120, + 114 + ], + "lines": [ + { + "bbox": [ + 90, + 106, + 120, + 114 + ], + "spans": [ + { + "bbox": [ + 90, + 106, + 120, + 114 + ], + "type": "inline_equation", + "content": "\\mathbb{O}\\times 564\\mathrm{h}" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 124, + 71, + 216, + 115 + ], + "lines": [ + { + "bbox": [ + 124, + 71, + 216, + 115 + ], + "spans": [ + { + "bbox": [ + 124, + 71, + 216, + 115 + ], + "type": "image", + "image_path": "eebc1e0bf3ece0ea41c10a5452a90876a509030e552ddf750acc4e667d57939d.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 220, + 77, + 257, + 105 + ], + "blocks": [ + { + "bbox": [ + 220, + 77, + 257, + 105 + ], + "lines": [ + { + "bbox": [ + 220, + 77, + 257, + 105 + ], + "spans": [ + { + "bbox": [ + 220, + 77, + 257, + 105 + ], + "type": "image", + "image_path": "a19d07b019f38fa7c8c76ad4a0301f9e336f516ac051e63806116e8155f497a2.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 230, + 106, + 249, + 114 + ], + "lines": [ + { + "bbox": [ + 230, + 106, + 249, + 114 + ], + "spans": [ + { + "bbox": [ + 230, + 106, + 249, + 114 + ], + "type": "text", + "content": "Music" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 263, + 82, + 287, + 89 + ], + "lines": [ + { + "bbox": [ + 263, + 82, + 287, + 89 + ], + "spans": [ + { + "bbox": [ + 263, + 82, + 287, + 89 + ], + "type": "text", + "content": "x59" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 263, + 90, + 286, + 97 + ], + "lines": [ + { + "bbox": [ + 263, + 90, + 286, + 97 + ], + "spans": [ + { + "bbox": [ + 263, + 90, + 286, + 97 + ], + "type": "text", + "content": "#" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 263, + 98, + 286, + 105 + ], + "lines": [ + { + "bbox": [ + 263, + 98, + 286, + 105 + ], + "spans": [ + { + "bbox": [ + 263, + 98, + 286, + 105 + ], + "type": "text", + "content": "#" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 295, + 71, + 386, + 115 + ], + "blocks": [ + { + "bbox": [ + 263, + 73, + 290, + 80 + ], + "lines": [ + { + "bbox": [ + 263, + 73, + 290, + 80 + ], + "spans": [ + { + "bbox": [ + 263, + 73, + 290, + 80 + ], + "type": "text", + "content": "x276" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 263, + 106, + 293, + 113 + ], + "lines": [ + { + "bbox": [ + 263, + 106, + 293, + 113 + ], + "spans": [ + { + "bbox": [ + 263, + 106, + 293, + 113 + ], + "type": "text", + "content": "L x 18Oh" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 295, + 71, + 386, + 115 + ], + "lines": [ + { + "bbox": [ + 295, + 71, + 386, + 115 + ], + "spans": [ + { + "bbox": [ + 295, + 71, + 386, + 115 + ], + "type": "image", + "image_path": "984ff27e32171872ea2ada18f5aedee16cc6fa00b95998cfa31845245dfb8c80.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 394, + 79, + 417, + 102 + ], + "blocks": [ + { + "bbox": [ + 394, + 79, + 417, + 102 + ], + "lines": [ + { + "bbox": [ + 394, + 79, + 417, + 102 + ], + "spans": [ + { + "bbox": [ + 394, + 79, + 417, + 102 + ], + "type": "image", + "image_path": "5a380045c4d5459df5d074c6037f9d8118652d1a20cf8e4e0f490735a0b03331.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 397, + 106, + 416, + 113 + ], + "lines": [ + { + "bbox": [ + 397, + 106, + 416, + 113 + ], + "spans": [ + { + "bbox": [ + 397, + 106, + 416, + 113 + ], + "type": "text", + "content": "Soccer" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 425, + 74, + 450, + 81 + ], + "lines": [ + { + "bbox": [ + 425, + 74, + 450, + 81 + ], + "spans": [ + { + "bbox": [ + 425, + 74, + 450, + 81 + ], + "type": "text", + "content": "x282" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 425, + 82, + 447, + 89 + ], + "lines": [ + { + "bbox": [ + 425, + 82, + 447, + 89 + ], + "spans": [ + { + "bbox": [ + 425, + 82, + 447, + 89 + ], + "type": "text", + "content": "x78" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 425, + 90, + 446, + 97 + ], + "lines": [ + { + "bbox": [ + 425, + 90, + 446, + 97 + ], + "spans": [ + { + "bbox": [ + 425, + 90, + 446, + 97 + ], + "type": "text", + "content": "#" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 425, + 98, + 446, + 104 + ], + "lines": [ + { + "bbox": [ + 425, + 98, + 446, + 104 + ], + "spans": [ + { + "bbox": [ + 425, + 98, + 446, + 104 + ], + "type": "text", + "content": "X14" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 425, + 106, + 450, + 113 + ], + "lines": [ + { + "bbox": [ + 425, + 106, + 450, + 113 + ], + "spans": [ + { + "bbox": [ + 425, + 106, + 450, + 113 + ], + "type": "inline_equation", + "content": "\\bullet \\times 66h" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 500, + 72, + 545, + 115 + ], + "blocks": [ + { + "bbox": [ + 481, + 72, + 499, + 79 + ], + "lines": [ + { + "bbox": [ + 481, + 72, + 499, + 79 + ], + "spans": [ + { + "bbox": [ + 481, + 72, + 499, + 79 + ], + "type": "inline_equation", + "content": "\\therefore m = \\frac{3}{11}" + }, + { + "bbox": [ + 481, + 72, + 499, + 79 + ], + "type": "text", + "content": " ;" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 500, + 72, + 545, + 115 + ], + "lines": [ + { + "bbox": [ + 500, + 72, + 545, + 115 + ], + "spans": [ + { + "bbox": [ + 500, + 72, + 545, + 115 + ], + "type": "image", + "image_path": "9a5e64d3c71b41b6225c4cef72f3bda9fcdcf5c474614aae5d0828f21ad7bd0c.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 481, + 117, + 512, + 121 + ], + "lines": [ + { + "bbox": [ + 481, + 117, + 512, + 121 + ], + "spans": [ + { + "bbox": [ + 481, + 117, + 512, + 121 + ], + "type": "text", + "content": "Soccer in Pittsburgh" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_caption" + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 55, + 122, + 81, + 151 + ], + "blocks": [ + { + "bbox": [ + 55, + 122, + 81, + 151 + ], + "lines": [ + { + "bbox": [ + 55, + 122, + 81, + 151 + ], + "spans": [ + { + "bbox": [ + 55, + 122, + 81, + 151 + ], + "type": "image", + "image_path": "c7aa36c3ac9b6150a2bd3f89c2eff2a7cd3f143b3eb72f3be85b17932dac5950.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 59, + 155, + 78, + 160 + ], + "lines": [ + { + "bbox": [ + 59, + 155, + 78, + 160 + ], + "spans": [ + { + "bbox": [ + 59, + 155, + 78, + 160 + ], + "type": "text", + "content": "Health" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 91, + 139, + 113, + 147 + ], + "lines": [ + { + "bbox": [ + 91, + 139, + 113, + 147 + ], + "spans": [ + { + "bbox": [ + 91, + 139, + 113, + 147 + ], + "type": "text", + "content": "9x24" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_caption" + } + ], + "index": 28 + }, + { + "type": "image", + "bbox": [ + 52, + 175, + 86, + 201 + ], + "blocks": [ + { + "bbox": [ + 52, + 175, + 86, + 201 + ], + "lines": [ + { + "bbox": [ + 52, + 175, + 86, + 201 + ], + "spans": [ + { + "bbox": [ + 52, + 175, + 86, + 201 + ], + "type": "image", + "image_path": "693afb2f8315a87b6db52b87cf9b6d450f5b8030e72417b8027d01539b8e2897.jpg" + } + ] + } + ], + "index": 36, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 52, + 205, + 86, + 212 + ], + "lines": [ + { + "bbox": [ + 52, + 205, + 86, + 212 + ], + "spans": [ + { + "bbox": [ + 52, + 205, + 86, + 212 + ], + "type": "text", + "content": "Bike Repair" + } + ] + } + ], + "index": 37, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 91, + 172, + 116, + 178 + ], + "lines": [ + { + "bbox": [ + 91, + 172, + 116, + 178 + ], + "spans": [ + { + "bbox": [ + 91, + 172, + 116, + 178 + ], + "type": "text", + "content": "4x363" + } + ] + } + ], + "index": 38, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 91, + 180, + 113, + 187 + ], + "lines": [ + { + "bbox": [ + 91, + 180, + 113, + 187 + ], + "spans": [ + { + "bbox": [ + 91, + 180, + 113, + 187 + ], + "type": "text", + "content": "x32" + } + ] + } + ], + "index": 39, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 91, + 189, + 110, + 197 + ], + "lines": [ + { + "bbox": [ + 91, + 189, + 110, + 197 + ], + "spans": [ + { + "bbox": [ + 91, + 189, + 110, + 197 + ], + "type": "text", + "content": "#" + } + ] + } + ], + "index": 40, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 91, + 198, + 110, + 205 + ], + "lines": [ + { + "bbox": [ + 91, + 198, + 110, + 205 + ], + "spans": [ + { + "bbox": [ + 91, + 198, + 110, + 205 + ], + "type": "text", + "content": "图" + } + ] + } + ], + "index": 41, + "angle": 0, + "type": "image_caption" + } + ], + "index": 36 + }, + { + "type": "image", + "bbox": [ + 124, + 121, + 214, + 166 + ], + "blocks": [ + { + "bbox": [ + 146, + 116, + 187, + 121 + ], + "lines": [ + { + "bbox": [ + 146, + 116, + 187, + 121 + ], + "spans": [ + { + "bbox": [ + 146, + 116, + 187, + 121 + ], + "type": "text", + "content": "Making Chai in Hyderabad" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 90, + 122, + 117, + 129 + ], + "lines": [ + { + "bbox": [ + 90, + 122, + 117, + 129 + ], + "spans": [ + { + "bbox": [ + 90, + 122, + 117, + 129 + ], + "type": "text", + "content": "x397" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 91, + 130, + 115, + 137 + ], + "lines": [ + { + "bbox": [ + 91, + 130, + 115, + 137 + ], + "spans": [ + { + "bbox": [ + 91, + 130, + 115, + 137 + ], + "type": "text", + "content": "x122" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 91, + 148, + 118, + 156 + ], + "lines": [ + { + "bbox": [ + 91, + 148, + 118, + 156 + ], + "spans": [ + { + "bbox": [ + 91, + 148, + 118, + 156 + ], + "type": "text", + "content": "中" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 91, + 158, + 118, + 163 + ], + "lines": [ + { + "bbox": [ + 91, + 158, + 118, + 163 + ], + "spans": [ + { + "bbox": [ + 91, + 158, + 118, + 163 + ], + "type": "text", + "content": "X141" + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 124, + 121, + 214, + 166 + ], + "lines": [ + { + "bbox": [ + 124, + 121, + 214, + 166 + ], + "spans": [ + { + "bbox": [ + 124, + 121, + 214, + 166 + ], + "type": "image", + "image_path": "bb9b9912947665d957dd000e5903a3a3fca5ab9d4d9ad8bfc1c1c303d813f372.jpg" + } + ] + } + ], + "index": 44, + "angle": 0, + "type": "image_body" + } + ], + "index": 44 + }, + { + "type": "image", + "bbox": [ + 122, + 168, + 214, + 215 + ], + "blocks": [ + { + "bbox": [ + 91, + 206, + 116, + 212 + ], + "lines": [ + { + "bbox": [ + 91, + 206, + 116, + 212 + ], + "spans": [ + { + "bbox": [ + 91, + 206, + 116, + 212 + ], + "type": "text", + "content": "X 82n" + } + ] + } + ], + "index": 43, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 122, + 168, + 214, + 215 + ], + "lines": [ + { + "bbox": [ + 122, + 168, + 214, + 215 + ], + "spans": [ + { + "bbox": [ + 122, + 168, + 214, + 215 + ], + "type": "image", + "image_path": "0d8c3e0cc3b5f0e78255b64b060f0ba6fceaaf039156df74327dc10012f83a30.jpg" + } + ] + } + ], + "index": 45, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 141, + 215, + 191, + 220 + ], + "lines": [ + { + "bbox": [ + 141, + 215, + 191, + 220 + ], + "spans": [ + { + "bbox": [ + 141, + 215, + 191, + 220 + ], + "type": "text", + "content": "Removing a Wheel in Menlo Park" + } + ] + } + ], + "index": 46, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 226, + 547, + 260 + ], + "lines": [ + { + "bbox": [ + 46, + 226, + 547, + 260 + ], + "spans": [ + { + "bbox": [ + 46, + 226, + 547, + 260 + ], + "type": "text", + "content": "Figure 2. Ego-Exo4D captures skilled activity from 43 tasks and 689 keysteps in 8 domains, in a wide variety of 123 scenes in 13 cities in Japan, Colombia, Canada, India, Singapore, and 7 US states. Each domain is captured at multiple sites—from 2 to 64 unique locations. In total the dataset offers 1,286 hours of ego+exo video comprised of 5,035 takes from 740 camera wearers. An average take is 2.6 minutes." + } + ] + } + ], + "index": 76, + "angle": 0, + "type": "image_caption" + } + ], + "index": 45 + }, + { + "type": "image", + "bbox": [ + 229, + 122, + 254, + 152 + ], + "blocks": [ + { + "bbox": [ + 229, + 122, + 254, + 152 + ], + "lines": [ + { + "bbox": [ + 229, + 122, + 254, + 152 + ], + "spans": [ + { + "bbox": [ + 229, + 122, + 254, + 152 + ], + "type": "image", + "image_path": "2d79df213b3cc867df1653fb63954d69806914bb972f99f01dd47e5d2f31cf51.jpg" + } + ] + } + ], + "index": 47, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 226, + 154, + 257, + 160 + ], + "lines": [ + { + "bbox": [ + 226, + 154, + 257, + 160 + ], + "spans": [ + { + "bbox": [ + 226, + 154, + 257, + 160 + ], + "type": "text", + "content": "Basketball" + } + ] + } + ], + "index": 48, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 263, + 122, + 288, + 129 + ], + "lines": [ + { + "bbox": [ + 263, + 122, + 288, + 129 + ], + "spans": [ + { + "bbox": [ + 263, + 122, + 288, + 129 + ], + "type": "text", + "content": "x910" + } + ] + } + ], + "index": 51, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 263, + 130, + 288, + 137 + ], + "lines": [ + { + "bbox": [ + 263, + 130, + 288, + 137 + ], + "spans": [ + { + "bbox": [ + 263, + 130, + 288, + 137 + ], + "type": "text", + "content": "x113" + } + ] + } + ], + "index": 52, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 263, + 139, + 287, + 147 + ], + "lines": [ + { + "bbox": [ + 263, + 139, + 287, + 147 + ], + "spans": [ + { + "bbox": [ + 263, + 139, + 287, + 147 + ], + "type": "text", + "content": "9x5" + } + ] + } + ], + "index": 53, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 263, + 148, + 290, + 156 + ], + "lines": [ + { + "bbox": [ + 263, + 148, + 290, + 156 + ], + "spans": [ + { + "bbox": [ + 263, + 148, + 290, + 156 + ], + "type": "text", + "content": "#" + } + ] + } + ], + "index": 54, + "angle": 0, + "type": "image_caption" + } + ], + "index": 47 + }, + { + "type": "image", + "bbox": [ + 229, + 166, + 254, + 204 + ], + "blocks": [ + { + "bbox": [ + 229, + 166, + 254, + 204 + ], + "lines": [ + { + "bbox": [ + 229, + 166, + 254, + 204 + ], + "spans": [ + { + "bbox": [ + 229, + 166, + 254, + 204 + ], + "type": "image", + "image_path": "8f5812255672053bbc201eee5ab4ee80e671699df439d8f665a728e94586e047.jpg" + } + ] + } + ], + "index": 49, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 220, + 205, + 262, + 212 + ], + "lines": [ + { + "bbox": [ + 220, + 205, + 262, + 212 + ], + "spans": [ + { + "bbox": [ + 220, + 205, + 262, + 212 + ], + "type": "text", + "content": "Rock Climbing" + } + ] + } + ], + "index": 50, + "angle": 0, + "type": "image_caption" + } + ], + "index": 49 + }, + { + "type": "image", + "bbox": [ + 263, + 166, + 293, + 215 + ], + "blocks": [ + { + "bbox": [ + 263, + 156, + 290, + 162 + ], + "lines": [ + { + "bbox": [ + 263, + 156, + 290, + 162 + ], + "spans": [ + { + "bbox": [ + 263, + 156, + 290, + 162 + ], + "type": "text", + "content": "5x78n" + } + ] + } + ], + "index": 55, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 263, + 166, + 293, + 215 + ], + "lines": [ + { + "bbox": [ + 263, + 166, + 293, + 215 + ], + "spans": [ + { + "bbox": [ + 263, + 166, + 293, + 215 + ], + "type": "image", + "image_path": "b3262b0b991a281b3481f1053e74925e8ce35455330367f75db85b66f831caf6.jpg" + } + ] + } + ], + "index": 57, + "angle": 0, + "type": "image_body" + } + ], + "index": 57 + }, + { + "type": "image", + "bbox": [ + 299, + 121, + 385, + 166 + ], + "blocks": [ + { + "bbox": [ + 315, + 116, + 359, + 121 + ], + "lines": [ + { + "bbox": [ + 315, + 116, + 359, + 121 + ], + "spans": [ + { + "bbox": [ + 315, + 116, + 359, + 121 + ], + "type": "text", + "content": "Piano Playing in Bloomington" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 299, + 121, + 385, + 166 + ], + "lines": [ + { + "bbox": [ + 299, + 121, + 385, + 166 + ], + "spans": [ + { + "bbox": [ + 299, + 121, + 385, + 166 + ], + "type": "image", + "image_path": "cd72a671a831712395fef38e7bb491b1d1d9d8048b0ef6a9eecb25ba32e7671d.jpg" + } + ] + } + ], + "index": 58, + "angle": 0, + "type": "image_body" + } + ], + "index": 58 + }, + { + "type": "image", + "bbox": [ + 294, + 167, + 385, + 215 + ], + "blocks": [ + { + "bbox": [ + 294, + 167, + 385, + 215 + ], + "lines": [ + { + "bbox": [ + 294, + 167, + 385, + 215 + ], + "spans": [ + { + "bbox": [ + 294, + 167, + 385, + 215 + ], + "type": "image", + "image_path": "28d54ba7c8516fa7d9299162bcbc4485a70940f8fbeaa3f9ac89ebd39d80f416.jpg" + } + ] + } + ], + "index": 59, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 317, + 215, + 357, + 220 + ], + "lines": [ + { + "bbox": [ + 317, + 215, + 357, + 220 + ], + "spans": [ + { + "bbox": [ + 317, + 215, + 357, + 220 + ], + "type": "text", + "content": "Bouldering in Minneapolis" + } + ] + } + ], + "index": 60, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 405, + 173, + 421, + 180 + ], + "lines": [ + { + "bbox": [ + 405, + 173, + 421, + 180 + ], + "spans": [ + { + "bbox": [ + 405, + 173, + 421, + 180 + ], + "type": "text", + "content": "Total" + } + ] + } + ], + "index": 71, + "angle": 0, + "type": "image_caption" + } + ], + "index": 59 + }, + { + "type": "image", + "bbox": [ + 399, + 122, + 418, + 152 + ], + "blocks": [ + { + "bbox": [ + 399, + 122, + 418, + 152 + ], + "lines": [ + { + "bbox": [ + 399, + 122, + 418, + 152 + ], + "spans": [ + { + "bbox": [ + 399, + 122, + 418, + 152 + ], + "type": "image", + "image_path": "35e7474e239096906675b0aa12b24e7ad316f45ac7b0d3da3354e281c1610354.jpg" + } + ] + } + ], + "index": 61, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 399, + 154, + 417, + 161 + ], + "lines": [ + { + "bbox": [ + 399, + 154, + 417, + 161 + ], + "spans": [ + { + "bbox": [ + 399, + 154, + 417, + 161 + ], + "type": "text", + "content": "Dance" + } + ] + } + ], + "index": 62, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 425, + 139, + 444, + 147 + ], + "lines": [ + { + "bbox": [ + 425, + 139, + 444, + 147 + ], + "spans": [ + { + "bbox": [ + 425, + 139, + 444, + 147 + ], + "type": "text", + "content": "9x7" + } + ] + } + ], + "index": 65, + "angle": 0, + "type": "image_caption" + } + ], + "index": 61 + }, + { + "type": "image", + "bbox": [ + 455, + 129, + 544, + 166 + ], + "blocks": [ + { + "bbox": [ + 432, + 122, + 450, + 129 + ], + "lines": [ + { + "bbox": [ + 432, + 122, + 450, + 129 + ], + "spans": [ + { + "bbox": [ + 432, + 122, + 450, + 129 + ], + "type": "text", + "content": "4x728" + } + ] + } + ], + "index": 63, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 432, + 130, + 449, + 137 + ], + "lines": [ + { + "bbox": [ + 432, + 130, + 449, + 137 + ], + "spans": [ + { + "bbox": [ + 432, + 130, + 449, + 137 + ], + "type": "text", + "content": "x93" + } + ] + } + ], + "index": 64, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 425, + 148, + 453, + 156 + ], + "lines": [ + { + "bbox": [ + 425, + 148, + 453, + 156 + ], + "spans": [ + { + "bbox": [ + 425, + 148, + 453, + 156 + ], + "type": "text", + "content": "#" + } + ] + } + ], + "index": 66, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 425, + 156, + 453, + 162 + ], + "lines": [ + { + "bbox": [ + 425, + 156, + 453, + 162 + ], + "spans": [ + { + "bbox": [ + 425, + 156, + 453, + 162 + ], + "type": "text", + "content": "x106n" + } + ] + } + ], + "index": 67, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 481, + 121, + 512, + 128 + ], + "lines": [ + { + "bbox": [ + 481, + 121, + 512, + 128 + ], + "spans": [ + { + "bbox": [ + 481, + 121, + 512, + 128 + ], + "type": "inline_equation", + "content": "\\therefore m = \\frac{3}{11}" + }, + { + "bbox": [ + 481, + 121, + 512, + 128 + ], + "type": "text", + "content": " ;" + } + ] + } + ], + "index": 68, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 455, + 129, + 544, + 166 + ], + "lines": [ + { + "bbox": [ + 455, + 129, + 544, + 166 + ], + "spans": [ + { + "bbox": [ + 455, + 129, + 544, + 166 + ], + "type": "image", + "image_path": "ef05d05169d1820987b79dcef743525e471d31daa226c8ce72abaeb09ed032e4.jpg" + } + ] + } + ], + "index": 69, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 485, + 167, + 508, + 171 + ], + "lines": [ + { + "bbox": [ + 485, + 167, + 508, + 171 + ], + "spans": [ + { + "bbox": [ + 485, + 167, + 508, + 171 + ], + "type": "text", + "content": "Salsa in Bogota" + } + ] + } + ], + "index": 70, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 431, + 173, + 485, + 180 + ], + "lines": [ + { + "bbox": [ + 431, + 173, + 485, + 180 + ], + "spans": [ + { + "bbox": [ + 431, + 173, + 485, + 180 + ], + "type": "text", + "content": "4x5035 (#takes)" + } + ] + } + ], + "index": 72, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 431, + 182, + 503, + 190 + ], + "lines": [ + { + "bbox": [ + 431, + 182, + 503, + 190 + ], + "spans": [ + { + "bbox": [ + 431, + 182, + 503, + 190 + ], + "type": "text", + "content": "x740 (#participants)" + } + ] + } + ], + "index": 73, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 431, + 192, + 477, + 201 + ], + "lines": [ + { + "bbox": [ + 431, + 192, + 477, + 201 + ], + "spans": [ + { + "bbox": [ + 431, + 192, + 477, + 201 + ], + "type": "text", + "content": "9. x 123 (# sites)" + } + ] + } + ], + "index": 74, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 431, + 203, + 514, + 212 + ], + "lines": [ + { + "bbox": [ + 431, + 203, + 514, + 212 + ], + "spans": [ + { + "bbox": [ + 431, + 203, + 514, + 212 + ], + "type": "text", + "content": "x 1286h (#ego+exo hours)" + } + ] + } + ], + "index": 75, + "angle": 0, + "type": "image_caption" + } + ], + "index": 69 + }, + { + "bbox": [ + 46, + 281, + 287, + 401 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 281, + 287, + 401 + ], + "spans": [ + { + "bbox": [ + 46, + 281, + 287, + 401 + ], + "type": "text", + "content": "facilitated by (largely exo) datasets like HowTo100M [103] and others [11, 159, 204, 207]. Challenges include grounding keysteps [10, 36, 37, 89, 103, 104, 178, 207], procedural planning [15, 17, 22, 71, 143, 167, 196, 201], learning task structure [4, 9, 37, 107, 202, 205], and leveraging noisy narrations [89, 103, 104]. A portion of Ego-Exo4D is procedural activities, but unlike the above, it offers simultaneous ego-exo capture. The scale and diversity of our data—including its three forms of language descriptions—widen the avenues for skilled activity understanding research." + } + ] + } + ], + "index": 77 + }, + { + "bbox": [ + 46, + 412, + 287, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 412, + 287, + 568 + ], + "spans": [ + { + "bbox": [ + 46, + 412, + 287, + 568 + ], + "type": "text", + "content": "Ego-exo cross-view modeling There is limited prior work on ego-exo cross-view modeling, arguably due to a lack of high-quality synchronized real-world data. Prior work explores matching people between videos [5, 6, 40, 170, 179] and learning view-invariant [7, 141, 144, 182, 184, 185] or ego features [82]. Beyond the specific case of ego-exo, cross-view methods are explored for translation [130, 131, 134, 157], novel view synthesis [19, 90, 135, 137, 164, 168, 171], and aerial to ground matching [86, 132]. Ego-Exo4D provides a testbed of unprecedented size and variety for cross-view modeling. In addition, our ego-exo relation tasks (cf. Section 5) surface new challenges in novel-view synthesis with widely varying viewpoints." + } + ] + } + ], + "index": 78 + }, + { + "bbox": [ + 47, + 580, + 160, + 594 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 580, + 160, + 594 + ], + "spans": [ + { + "bbox": [ + 47, + 580, + 160, + 594 + ], + "type": "text", + "content": "3. Ego-Exo4D dataset" + } + ] + } + ], + "index": 79 + }, + { + "bbox": [ + 46, + 601, + 287, + 649 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 601, + 287, + 649 + ], + "spans": [ + { + "bbox": [ + 46, + 601, + 287, + 649 + ], + "type": "text", + "content": "Next we introduce the dataset and its scope. Notably, the video capture was a distributed but coordinated effort performed by 12 research labs. We present the common framework, and reserve site-specific details for Appendix 10." + } + ] + } + ], + "index": 80 + }, + { + "bbox": [ + 47, + 658, + 162, + 672 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 658, + 162, + 672 + ], + "spans": [ + { + "bbox": [ + 47, + 658, + 162, + 672 + ], + "type": "text", + "content": "3.1. Ego-exo camera rig" + } + ] + } + ], + "index": 81 + }, + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "text", + "content": "Our goal is to capture simultaneous ego and exo video, together with multiple egocentric sensing modalities. One of our contributions is to create and share a low-cost (less than" + } + ] + } + ], + "index": 82 + }, + { + "bbox": [ + 305, + 281, + 545, + 305 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 281, + 545, + 305 + ], + "spans": [ + { + "bbox": [ + 305, + 281, + 545, + 305 + ], + "type": "text", + "content": "\\(3,000), lightweight ego-exo rig with a user-friendly calibration and time sync procedure." + } + ] + } + ], + "index": 83 + }, + { + "bbox": [ + 304, + 306, + 545, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 306, + 545, + 437 + ], + "spans": [ + { + "bbox": [ + 304, + 306, + 545, + 437 + ], + "type": "text", + "content": "Our camera configuration features Aria glasses [38] for ego capture, leveraging their rich array of sensors, including an 8 MP RGB camera, two SLAM cameras, IMU, 7 microphones, and eye tracking (see Appendix 7). The ego camera is calibrated and time-synchronized with four to five (stationary) GoPros placed on tripods as the exo capture devices, allowing 3D reconstruction of the environment point clouds and the participant's body pose. The number and placement of the exocentric cameras is determined per scenario in order to allow maximal coverage of useful viewpoints without obstructing the participants' activity." + } + ] + } + ], + "index": 84 + }, + { + "bbox": [ + 304, + 438, + 545, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 438, + 545, + 498 + ], + "spans": [ + { + "bbox": [ + 304, + 438, + 545, + 498 + ], + "type": "text", + "content": "Our time sync and calibration design relies on a QR-code procedure to auto-sync the cameras and auto-separate the individual \"takes\", meaning instances of an activity. We can do continuous recordings of up to " + }, + { + "bbox": [ + 304, + 438, + 545, + 498 + ], + "type": "inline_equation", + "content": "\\sim 60" + }, + { + "bbox": [ + 304, + 438, + 545, + 498 + ], + "type": "text", + "content": " minutes, based on the Aria battery life. See Appendix 8 for more details." + } + ] + } + ], + "index": 85 + }, + { + "bbox": [ + 306, + 508, + 456, + 519 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 508, + 456, + 519 + ], + "spans": [ + { + "bbox": [ + 306, + 508, + 456, + 519 + ], + "type": "text", + "content": "3.2. Domains and environments" + } + ] + } + ], + "index": 86 + }, + { + "bbox": [ + 304, + 526, + 545, + 622 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 526, + 545, + 622 + ], + "spans": [ + { + "bbox": [ + 304, + 526, + 545, + 622 + ], + "type": "text", + "content": "Ego-Exo4D focuses on skilled human activity. This is in contrast to existing ego-only efforts like Ego4D [47], which has a broad span of daily-life activities. We intentionally select the domains based on a few criteria: Will it illustrate skill and a variety of expertise? Is there visual variety to be expected across different instances? Will the ego and exo views offer complementary information? Will it present new challenges unaddressed by current datasets?" + } + ] + } + ], + "index": 87 + }, + { + "bbox": [ + 304, + 624, + 545, + 683 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 624, + 545, + 683 + ], + "spans": [ + { + "bbox": [ + 304, + 624, + 545, + 683 + ], + "type": "text", + "content": "Intersecting these criteria, we arrived at two broad categories1 of skilled activity: physical and procedural, together comprising eight total domains. The physical domains are soccer, basketball, dance, bouldering, and music. They emphasize body pose and movements as well as" + } + ] + } + ], + "index": 88 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 306, + 693, + 545, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 693, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 693, + 545, + 713 + ], + "type": "text", + "content": "1Note that in general physical and procedural are not mutually exclusive labels. An activity can both require physical skill and procedural steps." + } + ] + } + ], + "index": 89 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19386" + } + ] + } + ], + "index": 90 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 286, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 286, + 156 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 286, + 156 + ], + "type": "text", + "content": "interaction with objects (e.g., a ball, musical instrument). The procedural domains are cooking, bike repair, and health care. They require performing a sequence of steps to reach a goal state (e.g., a completed recipe, a repaired bike) and generally entail intricate hand-object manipulations with a variety of objects (e.g., bike repair tools; cooking utensils, appliances, and ingredients)." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 156, + 286, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 156, + 286, + 216 + ], + "spans": [ + { + "bbox": [ + 46, + 156, + 286, + 216 + ], + "type": "text", + "content": "In total, we have 43 activities derived from the eight domains (see Appendix 9). For example, cooking is comprised of 14 recipes; soccer is comprised of 3 drills. The length of a take ranges from 8 sec to " + }, + { + "bbox": [ + 46, + 156, + 286, + 216 + ], + "type": "inline_equation", + "content": "42\\mathrm{min}" + }, + { + "bbox": [ + 46, + 156, + 286, + 216 + ], + "type": "text", + "content": ", with procedural activities like cooking having the longest sustained captures." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 217, + 286, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 217, + 286, + 312 + ], + "spans": [ + { + "bbox": [ + 46, + 217, + 286, + 312 + ], + "type": "text", + "content": "To achieve visual diversity in the data, multiple labs across our team (typically 3-5) captured each Ego-Exo4D domain. The data is collected in authentic settings—such as real-world bike shops, soccer pitches, or bouldering gyms—as opposed to lab environments. For example, we have videos of chefs in New York City, Vancouver, Philadelphia, Bogota, and others; soccer players in Tokyo, Chapel Hill, Hyderabad, Singapore, and Pittsburgh. See Figure 2." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 320, + 239, + 333 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 320, + 239, + 333 + ], + "spans": [ + { + "bbox": [ + 47, + 320, + 239, + 333 + ], + "type": "text", + "content": "3.3. Participants: expertise and diversity" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 339, + 286, + 458 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 339, + 286, + 458 + ], + "spans": [ + { + "bbox": [ + 46, + 339, + 286, + 458 + ], + "type": "text", + "content": "We recruited 740 total participants from the local communities of 12 labs. All scenarios feature real-world experts, where the camera-wearer participant has specific credentials, training, or expertise in the skill being demonstrated. For example, among the Ego-Exo4D camera wearers are professional and college athletes; jazz, salsa, and Chinese folk dancers and instructors; competitive boulders; professional chefs who work in industrial-scale kitchens; bike technicians who service dozens of bikes per day. Many of them have (individually) over 10 years of experience." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 459, + 286, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 459, + 286, + 578 + ], + "spans": [ + { + "bbox": [ + 46, + 459, + 286, + 578 + ], + "type": "text", + "content": "Experts are prioritized given they are likely to conduct activities without mistakes or distractions, providing a strong ground truth for how to approach a given task. However, we also include capture from people with varying skill levels, as well—essential for our proposed skill proficiency estimation task (Section 5). Notably, Ego-Exo4D represents human intelligence in a new way by capturing domain-specific expertise—both in the video as well as the accompanying expert commentary (see Section 4)—portraying the evolution of a skill from beginners to experts." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 578, + 286, + 638 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 578, + 286, + 638 + ], + "spans": [ + { + "bbox": [ + 46, + 578, + 286, + 638 + ], + "type": "text", + "content": "According to the participant surveys (Appendix 11), the camera wearers range in age from 18 to 74 years old, with " + }, + { + "bbox": [ + 46, + 578, + 286, + 638 + ], + "type": "inline_equation", + "content": "37\\%" + }, + { + "bbox": [ + 46, + 578, + 286, + 638 + ], + "type": "text", + "content": " self-identifying as female " + }, + { + "bbox": [ + 46, + 578, + 286, + 638 + ], + "type": "inline_equation", + "content": "60\\%" + }, + { + "bbox": [ + 46, + 578, + 286, + 638 + ], + "type": "text", + "content": " male and " + }, + { + "bbox": [ + 46, + 578, + 286, + 638 + ], + "type": "inline_equation", + "content": "3\\%" + }, + { + "bbox": [ + 46, + 578, + 286, + 638 + ], + "type": "text", + "content": " as nonbinary or preferring not to say. In total, the participants self report more than 24 different ethnicities." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 647, + 155, + 659 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 647, + 155, + 659 + ], + "spans": [ + { + "bbox": [ + 47, + 647, + 155, + 659 + ], + "type": "text", + "content": "3.4. Privacy and ethics" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 665, + 286, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 286, + 712 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 286, + 712 + ], + "type": "text", + "content": "Ego-Exo4D was collected following rigorous privacy and ethics standards. This included undergoing formal independent review processes at each institution to establish the standards for collection, management, and informed con" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 308, + 71, + 369, + 157 + ], + "blocks": [ + { + "bbox": [ + 308, + 71, + 369, + 157 + ], + "lines": [ + { + "bbox": [ + 308, + 71, + 369, + 157 + ], + "spans": [ + { + "bbox": [ + 308, + 71, + 369, + 157 + ], + "type": "image", + "image_path": "23c2068815e1aebf6b65f96acf44b75df4aa036595e21ce72aa799e712f3f696.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 371, + 74, + 428, + 80 + ], + "lines": [ + { + "bbox": [ + 371, + 74, + 428, + 80 + ], + "spans": [ + { + "bbox": [ + 371, + 74, + 428, + 80 + ], + "type": "text", + "content": "Expert commentary" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 372, + 108, + 449, + 114 + ], + "lines": [ + { + "bbox": [ + 372, + 108, + 449, + 114 + ], + "spans": [ + { + "bbox": [ + 372, + 108, + 449, + 114 + ], + "type": "text", + "content": "Atomic action descriptions" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 372, + 136, + 417, + 141 + ], + "lines": [ + { + "bbox": [ + 372, + 136, + 417, + 141 + ], + "spans": [ + { + "bbox": [ + 372, + 136, + 417, + 141 + ], + "type": "text", + "content": "Narrate and act" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 305, + 167, + 545, + 189 + ], + "lines": [ + { + "bbox": [ + 305, + 167, + 545, + 189 + ], + "spans": [ + { + "bbox": [ + 305, + 167, + 545, + 189 + ], + "type": "text", + "content": "Figure 3. Ego-Exo4D offers 3 paired language corpora. Word cloud is from expert commentary which critiques the performance." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 371, + 81, + 456, + 99 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 371, + 81, + 456, + 99 + ], + "spans": [ + { + "bbox": [ + 371, + 81, + 456, + 99 + ], + "type": "text", + "content": "\"The dancer's hand is rotated inwardly a bit. Her palm should be facing to the ground, ...\"" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 372, + 115, + 446, + 127 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 372, + 115, + 446, + 127 + ], + "spans": [ + { + "bbox": [ + 372, + 115, + 446, + 127 + ], + "type": "text", + "content": "\"C steps forward while shimmering her shoulders.\"" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 372, + 142, + 455, + 155 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 372, + 142, + 455, + 155 + ], + "spans": [ + { + "bbox": [ + 372, + 142, + 455, + 155 + ], + "type": "text", + "content": "\"Here we go, left forward, right back for one phrase 1234...\"" + } + ] + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 459, + 71, + 545, + 157 + ], + "blocks": [ + { + "bbox": [ + 459, + 71, + 545, + 157 + ], + "lines": [ + { + "bbox": [ + 459, + 71, + 545, + 157 + ], + "spans": [ + { + "bbox": [ + 459, + 71, + 545, + 157 + ], + "type": "image", + "image_path": "090ac63719d41de68d3ac319ba1aaf42222509b8103f6d3c584ce5f72c4b69d7.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 203, + 545, + 311 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 203, + 545, + 311 + ], + "spans": [ + { + "bbox": [ + 304, + 203, + 545, + 311 + ], + "type": "text", + "content": "sent. Similarly, all Ego-Exo4D data collection adhered to the Project Aria Research Community Guidelines for responsible research. Since the scenarios allow for closed environments (e.g., no passerbys) nearly all video is available without de-identification. For information about each individual partners' protocols and restrictions, please see Appendix 10. Ego-Exo4D data is gated behind a license system, which defines permitted uses, restrictions, and consequences for non-compliance." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 305, + 322, + 474, + 335 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 322, + 474, + 335 + ], + "spans": [ + { + "bbox": [ + 305, + 322, + 474, + 335 + ], + "type": "text", + "content": "4. Natural language descriptions" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 342, + 545, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 342, + 545, + 450 + ], + "spans": [ + { + "bbox": [ + 304, + 342, + 545, + 450 + ], + "type": "text", + "content": "Ego-Exo4D also offers three kinds of paired natural language datasets, each time-indexed alongside the video. See Figure 3. These language annotations are not steered towards any single benchmark, but rather are a general resource that will support browsing and mining the dataset—as well as challenges in video-language learning like grounding actions and objects, self-supervised representation learning, video-conditioned language models, and skill assessment. See Appendix 12." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 451, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 451, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 451, + 545, + 713 + ], + "type": "text", + "content": "The first language dataset is spoken expert commentary. The goal is to reveal nuances of the skill that are not always visible to non-experts. We recruited 52 experts (distinct from the participants) to critique the recorded videos, call out strengths and weaknesses, explain how the specific behavior of the participant (e.g., hand/body pose, use of objects) affects the performance, and provide spatial markings to support their commentary. The experts are not only well-credentialed in their areas of expertise, but also have coaching or teaching experience, which facilitates clear communication. They watch the video and pause every time they have a comment, typically 7 times per minute of video. Each piece of commentary is unbounded in length, and averages 4 sentences. We provide both the transcribed speech and the raw audio (interesting for its inflection and nonword utterances), as well as the experts' spatial drawings and numeric ratings of each participant's skill. Videos have expert commentary by 2-5 distinct experts, offering a variety of perspectives for the same content. In total, we have 117,812 pieces of time-stamped, video-aligned commentary. These commentaries are quite novel: they focus on how the activity is executed rather than what it entails," + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19387" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 49, + 71, + 319, + 198 + ], + "blocks": [ + { + "bbox": [ + 49, + 71, + 319, + 198 + ], + "lines": [ + { + "bbox": [ + 49, + 71, + 319, + 198 + ], + "spans": [ + { + "bbox": [ + 49, + 71, + 319, + 198 + ], + "type": "image", + "image_path": "89e99b3d4d421fb985f90b50f51421249d01d29f4ac88415876f8ae480461290.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 104, + 205, + 488, + 217 + ], + "lines": [ + { + "bbox": [ + 104, + 205, + 488, + 217 + ], + "spans": [ + { + "bbox": [ + 104, + 205, + 488, + 217 + ], + "type": "text", + "content": "Figure 4. The ego-exo relation family consists of the tasks of correspondence (left) and translation (right)." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 324, + 72, + 544, + 198 + ], + "blocks": [ + { + "bbox": [ + 324, + 72, + 544, + 198 + ], + "lines": [ + { + "bbox": [ + 324, + 72, + 544, + 198 + ], + "spans": [ + { + "bbox": [ + 324, + 72, + 544, + 198 + ], + "type": "image", + "image_path": "c32b41e20a7b16c60de14b52977b7500e1565171dc15dbf4cd0aa43a8c26dc8e.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 234, + 287, + 281 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 234, + 287, + 281 + ], + "spans": [ + { + "bbox": [ + 46, + 234, + 287, + 281 + ], + "type": "text", + "content": "capturing subtle differences in skilled execution. We believe this can unlock new fundamental problems (e.g., proficiency estimation below) and disruptive future applications (e.g., AI coaching)." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 283, + 287, + 390 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 283, + 287, + 390 + ], + "spans": [ + { + "bbox": [ + 46, + 283, + 287, + 390 + ], + "type": "text", + "content": "The second language dataset consists of narrate-and-act descriptions provided by the participants themselves. They are in the style of a tutorial or how-to video, where the participant explains what they are doing and why. Unlike the third-party expert commentary above, these are first-person reflections on the activity given by the people doing them. These narrations are available for about " + }, + { + "bbox": [ + 46, + 283, + 287, + 390 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 46, + 283, + 287, + 390 + ], + "type": "text", + "content": " of all takes in the dataset, since we wanted participants to execute the tasks without pausing for the bulk of the recordings." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 392, + 287, + 547 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 392, + 287, + 547 + ], + "spans": [ + { + "bbox": [ + 46, + 392, + 287, + 547 + ], + "type": "text", + "content": "The third language dataset consists of atomic action descriptions. Whereas the commentary and narrate-and-act language reveals spoken opinions and reasons for the actions (the \"why and how\"), this stream of text is specifically about the \"what\". Inspired by Ego4D's narrations [47], these are short statements written by third-party (nondomain expert) annotators, timestamped for every atomic action performed by the participant for all videos in the dataset, for a total of 432K sentences. This data is valuable for mining for taxonomies of objects and actions in the data, indexing the videos with keywords for exploring the dataset, and for future research in video-language learning, as has been quite successful for the Ego4D narrations [8, 85, 123]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 560, + 211, + 574 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 560, + 211, + 574 + ], + "spans": [ + { + "bbox": [ + 47, + 560, + 211, + 574 + ], + "type": "text", + "content": "5. Ego-Exo4D benchmark tasks" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 582, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 582, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 582, + 287, + 713 + ], + "type": "text", + "content": "Our second major contribution is to define the core research challenges in the domain of egocentric perception of skilled activity, particularly when ego-exo data is available for training (if not testing). To that end, we devise a suite of foundational benchmark tasks organized into four task families: relation (Sec. 5.1), recognition (Sec. 5.2), proficiency (Sec. 5.3), and ego-pose (Sec. 5.4). For each task, we provide high quality annotations and baselines that provide a starting point from which the research community can build. We will run the first formal Ego-Exo4D challenges in 2024. Due to space limits, we briefly overview" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 234, + 545, + 294 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 234, + 545, + 294 + ], + "spans": [ + { + "bbox": [ + 304, + 234, + 545, + 294 + ], + "type": "text", + "content": "each task; see the referenced Appendices for all details including baseline models and results. There are two publicly released versions of Ego-Exo4D annotations: v1 is used to train/test baselines in this paper; the larger v2 will be used for future challenge leaderboards (see Table 7 in Appendix)." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 303, + 405, + 316 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 303, + 405, + 316 + ], + "spans": [ + { + "bbox": [ + 305, + 303, + 405, + 316 + ], + "type": "text", + "content": "5.1. Ego-exo relation" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 323, + 545, + 370 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 323, + 545, + 370 + ], + "spans": [ + { + "bbox": [ + 304, + 323, + 545, + 370 + ], + "type": "text", + "content": "Our ego-exo relation tasks deal with relating the video content across the extreme ego-exo viewpoint changes. They take the form of object-level matching (correspondence) and synthesis of one view from the other (translation)." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 380, + 443, + 392 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 380, + 443, + 392 + ], + "spans": [ + { + "bbox": [ + 305, + 380, + 443, + 392 + ], + "type": "text", + "content": "5.1.1 Ego-exo correspondence" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 401, + 545, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 401, + 545, + 509 + ], + "spans": [ + { + "bbox": [ + 304, + 401, + 545, + 509 + ], + "type": "text", + "content": "Motivation. Establishing object-level correspondences between ego and exo viewpoints would allow AI assistants to provide visual instructions by matching third-person observations of objects from instructional videos to those in the user's first-person view. Compared to the general correspondence problem, our setting requires tackling a number of challenges: extreme viewpoint differences, high degrees of object occlusion, and many small objects (e.g., cooking utensils and bike repair tools)." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 510, + 545, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 510, + 545, + 617 + ], + "spans": [ + { + "bbox": [ + 304, + 510, + 545, + 617 + ], + "type": "text", + "content": "Task definition. Given a pair of synchronized ego-exo videos and a sequence of query masks of an object of interest in one of the videos, the task is to predict the corresponding mask for the same object in each synchronized frame of the other view, if it is visible. See Figure 4, left. The task can be posed with query objects in either the ego or exo video, with both directions presenting interesting challenges (e.g., high degree of occlusion in ego views, and small object size in exo views). See Appendix 13.A.1." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 618, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 618, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 618, + 545, + 713 + ], + "type": "text", + "content": "Related work. Related tasks are image-level sparse correspondence given query points (instead of object masks) [65] and image-level object co-segmentation [166] for jointly segmenting semantically similar objects. Our task goes beyond static object correspondence, since the interplay between human pose and object state changes during manipulation necessitate using temporal context and tracking as the query object can be highly occluded or blurry [158]." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19388" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 48, + 70, + 211, + 210 + ], + "blocks": [ + { + "bbox": [ + 48, + 70, + 211, + 210 + ], + "lines": [ + { + "bbox": [ + 48, + 70, + 211, + 210 + ], + "spans": [ + { + "bbox": [ + 48, + 70, + 211, + 210 + ], + "type": "image", + "image_path": "ff098b4a1c0e580f2546738c52b291d66c8636b675c4d68f04c6ea7151e35c9d.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 220, + 546, + 243 + ], + "lines": [ + { + "bbox": [ + 46, + 220, + 546, + 243 + ], + "spans": [ + { + "bbox": [ + 46, + 220, + 546, + 243 + ], + "type": "text", + "content": "Figure 5. Ego-exo keystep recognition. This family of tasks consists of fine-grained recognition (left, Section 5.2.1), procedure understanding (center, Section 5.2.3), and energy-efficient multimodal recognition (right, Section 5.2.2)." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 215, + 70, + 378, + 211 + ], + "blocks": [ + { + "bbox": [ + 215, + 70, + 378, + 211 + ], + "lines": [ + { + "bbox": [ + 215, + 70, + 378, + 211 + ], + "spans": [ + { + "bbox": [ + 215, + 70, + 378, + 211 + ], + "type": "image", + "image_path": "fd1e6617f4d3344436a9a691e16b97009d3ab5c5043378dbf4e1a9fdf69f3ca2.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 381, + 70, + 545, + 210 + ], + "blocks": [ + { + "bbox": [ + 381, + 70, + 545, + 210 + ], + "lines": [ + { + "bbox": [ + 381, + 70, + 545, + 210 + ], + "spans": [ + { + "bbox": [ + 381, + 70, + 545, + 210 + ], + "type": "image", + "image_path": "bf51d78274add40ba749ff7f90622abf5e662a32f03e54a7f68398be5de6235b.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 263, + 164, + 275 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 263, + 164, + 275 + ], + "spans": [ + { + "bbox": [ + 47, + 263, + 164, + 275 + ], + "type": "text", + "content": "5.1.2 Ego-exo translation" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 282, + 287, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 282, + 287, + 450 + ], + "spans": [ + { + "bbox": [ + 46, + 282, + 287, + 450 + ], + "type": "text", + "content": "Motivation. Our translation task entails synthesizing a target ego clip from a given exo clip. We believe this problem will drive novel research for combining recognition and object synthesis. For example, in Figure 4 (right), the approach must make effective use of the hand's object-specific shape and appearance priors in order to synthesize the ego view of the fingertips—which are not visible in the exo clip. Furthermore, this task will stimulate advances in visual odometry, as the method must be able to infer the ego camera pose from the third-person clip. Ego-exo translation also holds strong application potential, as it may unlock the ability to generate first-person renderings of videos that were originally captured from a third-person perspective, e.g., benefitting robot perception or AR coaching." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 450, + 288, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 450, + 288, + 653 + ], + "spans": [ + { + "bbox": [ + 46, + 450, + 288, + 653 + ], + "type": "text", + "content": "Task definition. We decompose ego-exo translation into two separate tasks: ego track prediction and ego clip generation (Figure 4, right). Ego track prediction estimates the segmentation mask of an object in the unobserved ego frames given the object masks in the observed exo clip. Ego clip generation must generate the image values (i.e., RGB) within the given ground-truth ego mask by making use of the exo clip and the object masks in those frames. This decomposition effectively splits the problem into two tasks: 1) predicting the location and shape of the object in the ego clip, and 2) synthesizing its appearance given the ground-truth position. For each, we consider a variant where the pose of the ego camera with respect to the exo camera is available to use at inference time. This simplifies the problem but reduces the applicability of the method, since this information is typically not available for arbitrary third-person videos. See Appendix 13.A.2." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 654, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 654, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 654, + 287, + 713 + ], + "type": "text", + "content": "Related work. Ego-exo translation relates to cross-view image synthesis [96, 130, 157]. Within this genre, the problem of exo-to-ego generation was recently introduced for both images [93] and video [94, 97], and approached using GANs or diffusion conditioned on the input view. Our work" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 263, + 545, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 263, + 545, + 300 + ], + "spans": [ + { + "bbox": [ + 305, + 263, + 545, + 300 + ], + "type": "text", + "content": "not only formalizes this task with ample data, but its formulation also draws attention to the need for a semantic basis to new view synthesis across extreme view changes." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 308, + 460, + 320 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 308, + 460, + 320 + ], + "spans": [ + { + "bbox": [ + 306, + 308, + 460, + 320 + ], + "type": "text", + "content": "5.2. Ego-exo keystep recognition" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 326, + 545, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 326, + 545, + 350 + ], + "spans": [ + { + "bbox": [ + 305, + 326, + 545, + 350 + ], + "type": "text", + "content": "This family of tasks centers around recognizing the keysteps of a procedural activity and modeling their dependencies." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 358, + 479, + 371 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 358, + 479, + 371 + ], + "spans": [ + { + "bbox": [ + 306, + 358, + 479, + 371 + ], + "type": "text", + "content": "5.2.1 Fine-grained keystep recognition" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 378, + 546, + 485 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 378, + 546, + 485 + ], + "spans": [ + { + "bbox": [ + 304, + 378, + 546, + 485 + ], + "type": "text", + "content": "Motivation. Recognizing the step a camera wearer is performing is non-trivial: keysteps in the same activity may look similar (folding vs. smoothing the bedsheet) and may involve hand-object interactions with heavy occlusions and head motion. Models with access to multiple views during training can leverage their complementarity to account for the deficiencies of each one, by learning viewpoint invariant representations or distilling multi-view signals into a single model (e.g., human hands from ego; body pose from exo)." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 486, + 546, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 486, + 546, + 641 + ], + "spans": [ + { + "bbox": [ + 304, + 486, + 546, + 641 + ], + "type": "text", + "content": "Task definition. We study ego-exo for video recognition. During training, models have access to paired ego-exo data—time-synchronized captures of the same activity from multiple known viewpoints. Each training instance has one ego view, " + }, + { + "bbox": [ + 304, + 486, + 546, + 641 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 486, + 546, + 641 + ], + "type": "text", + "content": " exo views, and a corresponding keystep label (e.g., \"flip the omelette\"). At test time, given only a trimmed egocentric video clip, the model must identify the keystep performed from a taxonomy of 689 keysteps across 17 procedural activities. See Figure 5, left. Importantly, all extra supervision (time-alignment, camera poses etc.) is only available at training time; inference is standard keystep recognition, but with models that benefit from cross-viewpoint training. See Appendix 13.B.1." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 642, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 642, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 642, + 547, + 713 + ], + "type": "text", + "content": "Related work. Keystep recognition has been studied in first-person [10, 126, 145, 148] or third-person [9, 100, 159, 205, 207] videos; however, limited work considers both views together. Prior work considers cross-view learning with unpaired videos [7, 82, 182] and view-invariant feature learning on paired videos [144]. In contrast, we ex" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19389" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 287, + 97 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 287, + 97 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 287, + 97 + ], + "type": "text", + "content": "plore keystep recognition in large-scale, procedural activities with fully synchronized training videos." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 104, + 285, + 118 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 104, + 285, + 118 + ], + "spans": [ + { + "bbox": [ + 47, + 104, + 285, + 118 + ], + "type": "text", + "content": "5.2.2 Energy-efficient multimodal keystep recognition" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 123, + 287, + 218 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 123, + 287, + 218 + ], + "spans": [ + { + "bbox": [ + 46, + 123, + 287, + 218 + ], + "type": "text", + "content": "Motivation. Current activity detection models assume access to densely sampled clips from the full video and ample computational resources to process them. These assumptions are incompatible with real-world devices (e.g., mobile phones, AR glasses) where the camera is not always on and the compute budget is limited by battery life. This task focuses on building energy-efficient video models to pave the way for feasibility on real-world hardware." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 219, + 287, + 351 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 219, + 287, + 351 + ], + "spans": [ + { + "bbox": [ + 46, + 219, + 287, + 351 + ], + "type": "text", + "content": "Task definition. We formulate the problem as an online action detection task, with a given energy budget. See Figure 5, right. Given a stream of audio, IMU, and RGB video data, a model must identify the keystep being performed at each frame, as well as decide which sensor(s) to use for subsequent time-steps. This task will inspire models that are strategic about which modality to deploy when. Energy consumption is the sum of sensor energy (operating the camera/audio/IMU sensors), model inference costs, and memory transfer costs, and must be within " + }, + { + "bbox": [ + 46, + 219, + 287, + 351 + ], + "type": "inline_equation", + "content": "20\\mathrm{mW}" + }, + { + "bbox": [ + 46, + 219, + 287, + 351 + ], + "type": "text", + "content": " to reflect real-world device power constraints. See Appendix 13.B.2." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 351, + 287, + 471 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 351, + 287, + 471 + ], + "spans": [ + { + "bbox": [ + 46, + 351, + 287, + 471 + ], + "type": "text", + "content": "Related work. Prior work on efficient models considers light-weight architectures [41, 56, 101, 155, 165, 195], efficient input processing [43, 44, 73, 102, 156], or inference optimizations [39, 58, 121, 174, 206]. In all cases, they optimize computation (FLOPs), parameter count, or prediction throughput (FPS), which in isolation are insufficient to characterize running on real-world devices. To address this, we propose the first benchmark for energy-efficient video recognition that is tied to real-world, on-device constraints, and measures total power consumed." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 479, + 189, + 491 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 479, + 189, + 491 + ], + "spans": [ + { + "bbox": [ + 47, + 479, + 189, + 491 + ], + "type": "text", + "content": "5.2.3 Procedure understanding" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 498, + 287, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 498, + 287, + 545 + ], + "spans": [ + { + "bbox": [ + 46, + 498, + 287, + 545 + ], + "type": "text", + "content": "Motivation. Automatically understanding the structure of a procedure from video (inferring keystep ordering, preconditions, etc.) would allow assisting AR users in a task or informing robots that learn from human demonstrations." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 545, + 287, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 545, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 46, + 545, + 287, + 689 + ], + "type": "text", + "content": "Task definition. In our procedure understanding task, given a video segment " + }, + { + "bbox": [ + 46, + 545, + 287, + 689 + ], + "type": "inline_equation", + "content": "s_t" + }, + { + "bbox": [ + 46, + 545, + 287, + 689 + ], + "type": "text", + "content": " and its previous video segment history, models have to 1) determine previous keysteps (to be performed before " + }, + { + "bbox": [ + 46, + 545, + 287, + 689 + ], + "type": "inline_equation", + "content": "s_t" + }, + { + "bbox": [ + 46, + 545, + 287, + 689 + ], + "type": "text", + "content": "); infer if " + }, + { + "bbox": [ + 46, + 545, + 287, + 689 + ], + "type": "inline_equation", + "content": "s_t" + }, + { + "bbox": [ + 46, + 545, + 287, + 689 + ], + "type": "text", + "content": " is 2) optional or 3) a procedural mistake; 4) predict missing keysteps (should have been performed before " + }, + { + "bbox": [ + 46, + 545, + 287, + 689 + ], + "type": "inline_equation", + "content": "s_t" + }, + { + "bbox": [ + 46, + 545, + 287, + 689 + ], + "type": "text", + "content": " but were not); and 5) next keysteps (for which dependencies are satisfied). The task offers two versions of weak supervision: instance-level: segments and their keystep labels are available for train/test; and procedure-level: only unlabeled segments and procedure-specific keystep names are given for train/test. See Figure 5 (center) and Appendix 13.B.3." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "type": "text", + "content": "Related work. Prior work focusing on procedural understanding learns an explicit graph [62, 150, 177] as" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 309, + 71, + 545, + 186 + ], + "blocks": [ + { + "bbox": [ + 309, + 71, + 545, + 186 + ], + "lines": [ + { + "bbox": [ + 309, + 71, + 545, + 186 + ], + "spans": [ + { + "bbox": [ + 309, + 71, + 545, + 186 + ], + "type": "image", + "image_path": "8179128309f83265ff9d2b17386b2dd0f72a15b935efdf359b6762bfc3f7347a.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 306, + 193, + 545, + 205 + ], + "lines": [ + { + "bbox": [ + 306, + 193, + 545, + 205 + ], + "spans": [ + { + "bbox": [ + 306, + 193, + 545, + 205 + ], + "type": "text", + "content": "Figure 6. Demonstrator and demonstration proficiency estimation." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 221, + 545, + 293 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 221, + 545, + 293 + ], + "spans": [ + { + "bbox": [ + 305, + 221, + 545, + 293 + ], + "type": "text", + "content": "ground truth or uses a task graph for representation learning [9, 107, 202] and short-term step understanding [9, 36, 202]. Other work [32, 139] studies mistake detection in a supervised setting. We are the first to propose procedural understanding to evaluate the long-term structure of the task in a weakly-supervised setting." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 301, + 471, + 314 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 301, + 471, + 314 + ], + "spans": [ + { + "bbox": [ + 306, + 301, + 471, + 314 + ], + "type": "text", + "content": "5.3. Ego-exo proficiency estimation" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 319, + 545, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 319, + 545, + 380 + ], + "spans": [ + { + "bbox": [ + 305, + 319, + 545, + 380 + ], + "type": "text", + "content": "Motivation. Going beyond recognizing what a person is doing, this task aims to infer the user's skill level. Such an ability could lead to novel coaching tools that let people learn new skills more effectively, or new ways to evaluate human performance in domains like sports or music." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 380, + 546, + 547 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 380, + 546, + 547 + ], + "spans": [ + { + "bbox": [ + 305, + 380, + 546, + 547 + ], + "type": "text", + "content": "Task definition. We consider two variants: (1) demonstrator and (2) demonstration proficiency estimation. Both tasks consider one egocentric and (optionally) " + }, + { + "bbox": [ + 305, + 380, + 546, + 547 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 305, + 380, + 546, + 547 + ], + "type": "text", + "content": " exocentric videos synchronized in time as their inputs. Demonstrator proficiency is formulated as a video classification task, where the model has to output one of four labels (novice, early, intermediate, or late expert). Demonstration proficiency is formulated as a temporal action localization task where given an untrimmed video, the model must output a list of tuples, each containing a timestamp, a proficiency category (i.e., good execution or needs improvement), and its probability. Note that parts of the video that do not reveal the participant's skill are left unlabeled. See Figure 6 and Appendix 13.C." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 548, + 545, + 643 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 548, + 545, + 643 + ], + "spans": [ + { + "bbox": [ + 305, + 548, + 545, + 643 + ], + "type": "text", + "content": "Related work. Prior work uses egocentric [12, 35] or exocentric [60, 114, 115] views for proficiency estimation in sports [12, 115, 120], health [60, 92, 191, 208], and others [35, 186]. We propose the first multi-view egocentric and exocentric proficiency estimation benchmark. Unlike prior work, our benchmark spans diverse, day-to-day physical and procedural scenarios and includes temporally localized annotations of (in)correct executions." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 651, + 370, + 664 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 651, + 370, + 664 + ], + "spans": [ + { + "bbox": [ + 306, + 651, + 370, + 664 + ], + "type": "text", + "content": "5.4. Ego pose" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 305, + 670, + 545, + 705 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 670, + 545, + 705 + ], + "spans": [ + { + "bbox": [ + 305, + 670, + 545, + 705 + ], + "type": "text", + "content": "This family of tasks is motivated by recovering the skilled body movements of participants, even in the extreme setting of monocular ego-video input in dynamic environments." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19390" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 72, + 281, + 175 + ], + "blocks": [ + { + "bbox": [ + 57, + 72, + 281, + 175 + ], + "lines": [ + { + "bbox": [ + 57, + 72, + 281, + 175 + ], + "spans": [ + { + "bbox": [ + 57, + 72, + 281, + 175 + ], + "type": "image", + "image_path": "1153a25866145462bb4bdcae10f45145db80a004ab374bbddd7703c6dcaca210.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 57, + 180, + 274, + 191 + ], + "lines": [ + { + "bbox": [ + 57, + 180, + 274, + 191 + ], + "spans": [ + { + "bbox": [ + 57, + 180, + 274, + 191 + ], + "type": "text", + "content": "Figure 7. Hand and body keypoints for ego-pose estimation" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 201, + 287, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 201, + 287, + 380 + ], + "spans": [ + { + "bbox": [ + 46, + 201, + 287, + 380 + ], + "type": "text", + "content": "Motivation. Estimating the physical state of a person's body—the 3D positions of the arms, legs, hands—from the ego view is essential for wearable AI systems that can support human activity. Challenges include subtle and flexible movements, frequent occlusion, and body parts out of view. Task definition. For both the body and hand pose (\"ego pose\") estimation tasks, the input is an ego video. The output is a set of 3D joint positions of the camera wearer's body and hands for each time step, parameterized as 17 3D body joint positions and 21 3D joint positions per hand, following the MS COCO convention [87]. To our knowledge, EgoExo4D offers the largest manual ground truth (GT) egocentric body and hand pose annotations to date. And, in total, it offers " + }, + { + "bbox": [ + 46, + 201, + 287, + 380 + ], + "type": "inline_equation", + "content": "\\sim 14\\mathrm{M}" + }, + { + "bbox": [ + 46, + 201, + 287, + 380 + ], + "type": "text", + "content": " frames of 3D GT and pseudo-GT combined. See Figure 7 and Appendix 13.D." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 382, + 286, + 489 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 382, + 286, + 489 + ], + "spans": [ + { + "bbox": [ + 46, + 382, + 286, + 489 + ], + "type": "text", + "content": "Related work. Limited prior work explores 3D body pose from a wearable camera. Some methods assume no body visibility [63, 80, 98, 187, 188], while others assume partial observability by modifying cameras to capture the body [3, 57, 136, 161, 181]. Our dataset can be used for both paradigms. Existing hand pose datasets use constrained environments [106, 146] with simple hand motion [50, 76, 109], whereas we include diverse real-world scenarios, e.g., with expert musicians and bike mechanics." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 501, + 125, + 514 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 501, + 125, + 514 + ], + "spans": [ + { + "bbox": [ + 47, + 501, + 125, + 514 + ], + "type": "text", + "content": "6. Conclusions" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 521, + 286, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 521, + 286, + 605 + ], + "spans": [ + { + "bbox": [ + 46, + 521, + 286, + 605 + ], + "type": "text", + "content": "Ego-Exo4D provides a dataset of unprecedented scale and realism for ego-exo video learning. It offers a unique window into skilled human activity from 8 compelling domains by hundreds of real-world experts around the globe. Together with the proposed benchmarks, we hope that this new open source resource will set the stage for substantial new research for the years to come." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 605, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 286, + 713 + ], + "type": "text", + "content": "Though we are motivated by skill learning, Ego-Exo4D is poised for even broader influence, beyond the proposed benchmarks. Whereas existing datasets lack activity modeling in real-world 3D contexts (e.g., restricted to mocap suits and/or lab settings), Ego-Exo4D is a resource for general 3D vision—such as environment reconstruction, camera relocalization, audio-visual mapping, and many others. Similarly, our novel video-language resources will offer many opportunities for grounding of actions and objects, multi" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 72, + 545, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 133 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 133 + ], + "type": "text", + "content": "modal representation learning, and language generation. Finally, though our tasks prioritize perception from the \"ego-only\" perspective, the exo component of our data ensures its utility for the more traditional exo viewpoint too, e.g., for activity recognition and body pose estimation." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 142, + 429, + 153 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 142, + 429, + 153 + ], + "spans": [ + { + "bbox": [ + 306, + 142, + 429, + 153 + ], + "type": "text", + "content": "Contribution statement" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 162, + 545, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 162, + 545, + 376 + ], + "spans": [ + { + "bbox": [ + 305, + 162, + 545, + 376 + ], + "type": "text", + "content": "This project is the result of a large collaboration between many institutions over the last two years. Initial authors represent the leadership team of the project. Kristen Grauman initiated the project, served as the technical lead, initiated the recognition and proficiency benchmarks and expert commentary, and coordinated their working groups. Andrew Westbury served as the program manager and operations lead for all aspects of the project. Lorenzo Torresani led development of the capture domains, initiated the relation and ego-pose benchmarks, and coordinated their working groups. Kris Kitani led development of the multi-camera rig and supported the Ego-Exo4D engineering team on all aspects of the data annotation and organization. Jitendra Malik served as a scientific advisor. Authors with stars " + }, + { + "bbox": [ + 305, + 162, + 545, + 376 + ], + "type": "inline_equation", + "content": "(^{*})" + }, + { + "bbox": [ + 305, + 162, + 545, + 376 + ], + "type": "text", + "content": " were key drivers of implementation, collection, and/or annotation development throughout the project. Authors with daggers " + }, + { + "bbox": [ + 305, + 162, + 545, + 376 + ], + "type": "inline_equation", + "content": "(^{\\dagger})" + }, + { + "bbox": [ + 305, + 162, + 545, + 376 + ], + "type": "text", + "content": " are faculty and senior researcher PIs for the project. The Appendices detail the contributions of individual authors for the various benchmarks, data collection, and annotation pipelines." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 387, + 408, + 399 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 387, + 408, + 399 + ], + "spans": [ + { + "bbox": [ + 306, + 387, + 408, + 399 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 406, + 545, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 406, + 545, + 581 + ], + "spans": [ + { + "bbox": [ + 305, + 406, + 545, + 581 + ], + "type": "text", + "content": "We gratefully acknowledge the following colleagues for valuable discussions and support of our project: Vittorio Caggiano, Ilé Danza, Ahmad Darkhalil, Zona de Bloque, Alex Dinh, Rene Martinez Doehner, Ivan Cruz, Matt Feiszli, Vance Feutz, Kelly Forbes, Rohit Girdhar, Pierre Gleize, Andrés Hernández, Shun Iwase, Bolin Lai, Vivian Lee, Brighid Meredith, Ashley Massie, Natalia Neverova, Manohar Paluri, Joelle Pineau, Artsiom Sanakoyeu, Paresh Shenoy, Jiaray Shi, Jiasheng Shi, Gaurav Shrivastava, Mitesh Singh, Manasi Swaminathan, Arjang Talattof, Ali Thabet, Laurens van der Maaten, Andrea Vedaldi, and Toby Zhu. We also sincerely thank the 52 experts who contributed to the expert commentary for their expertise and support; they are listed individually in Appendix 12. Thank you to the Common Visual Data Foundation (CVDF) for hosting the Ego-Exo4D dataset. Finally, thank you to the 740 participants who contributed to this dataset and shared their skills in video." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 582, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 582, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 582, + 545, + 713 + ], + "type": "text", + "content": "Univ. of Bristol is supported in part by EPSRC UMPIRE (EP/T004991/1) and EPSRC PG Visual AI (EP/T028572/1). Z. Zhu is supported by UoB-CSC Scholarship. Univ. of Catalonia is supported in part by the project Future Artificial Intelligence Research (FAIR) - PNRR MUR Cod. PE0000013 - CUP: E63C22001940006. Simon Fraser Univ. is supported in part by the Canada Research Chairs Program (CRC-2019-00298) and NSERC Discovery (2019-06489). Georgia Tech is supported in part by NSF award CNS-2308994. IU is supported in part by NSF DRL-2112635 (AI Institute for Engaged Learning). UT Austin is supported in part by the IFML NSF AI Institute." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19391" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 109, + 105, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 109, + 105, + 121 + ], + "spans": [ + { + "bbox": [ + 48, + 109, + 105, + 121 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 53, + 129, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 58, + 129, + 287, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 129, + 287, + 194 + ], + "spans": [ + { + "bbox": [ + 58, + 129, + 287, + 194 + ], + "type": "text", + "content": "[1] Ahmed Tawfik Aboukhadra, Jameel Malik, Ahmed Elhayek, Nadia Robertini, and Didier Stricker. Thor-net: End-to-end graformer-based realistic two hands and object reconstruction with self-supervision. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1001-1010, 2023. 58, 59" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 196, + 287, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 196, + 287, + 228 + ], + "spans": [ + { + "bbox": [ + 58, + 196, + 287, + 228 + ], + "type": "text", + "content": "[2] Michael Abrash. Creating the future: Augmented reality, the next human-machine interface. In 2021 IEEE International Electron Devices Meeting (IEDM), 2021. 46" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 58, + 230, + 287, + 283 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 230, + 287, + 283 + ], + "spans": [ + { + "bbox": [ + 58, + 230, + 287, + 283 + ], + "type": "text", + "content": "[3] Karan Ahuja, Chris Harrison, Mayank Goel, and Robert Xiao. Mecap: Whole-body digitization for low-cost vr/ar headsets. In Proceedings of the 32nd Annual ACM Symposium on User Interface Software and Technology, pages 453-462, 2019. 9" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 58, + 286, + 287, + 328 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 286, + 287, + 328 + ], + "spans": [ + { + "bbox": [ + 58, + 286, + 287, + 328 + ], + "type": "text", + "content": "[4] Jean-Baptiste Alayrac, Piotr Bojanowski, Nishant Agrawal, Josef Sivic, Ivan Laptev, and Simon Lacoste-Julien. Unsupervised learning from narrated instruction videos. In CVPR, 2016. 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 58, + 331, + 287, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 331, + 287, + 352 + ], + "spans": [ + { + "bbox": [ + 58, + 331, + 287, + 352 + ], + "type": "text", + "content": "[5] Shervin Ardeshir and Ali Borji. Ego2top: Matching viewers in egocentric and top-view videos. In ECCV, 2016. 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 58, + 354, + 287, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 354, + 287, + 386 + ], + "spans": [ + { + "bbox": [ + 58, + 354, + 287, + 386 + ], + "type": "text", + "content": "[6] Shervin Ardeshir and Ali Borji. Egocentric meets top-view. IEEE transactions on pattern analysis and machine intelligence, 41(6), 2018. 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 58, + 388, + 287, + 420 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 388, + 287, + 420 + ], + "spans": [ + { + "bbox": [ + 58, + 388, + 287, + 420 + ], + "type": "text", + "content": "[7] Shervin Ardeshir and Ali Borji. An exocentric look at egocentric actions and vice versa. Computer Vision and Image Understanding, 171, 2018. 4, 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 58, + 422, + 287, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 422, + 287, + 475 + ], + "spans": [ + { + "bbox": [ + 58, + 422, + 287, + 475 + ], + "type": "text", + "content": "[8] Kumar Ashutosh, Rohit Girdhar, Lorenzo Torresani, and Kristen Grauman. Hiervl: Learning hierarchical videolanguage embeddings. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023. 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 58, + 478, + 287, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 478, + 287, + 520 + ], + "spans": [ + { + "bbox": [ + 58, + 478, + 287, + 520 + ], + "type": "text", + "content": "[9] Kumar Ashutosh, Santhosh Kumar Ramakrishnan, Triantafyllos Afouras, and Kristen Grauman. Video-mined task graphs for keystep recognition in instructional videos. In NeurIPS, 2023. 4, 7, 8, 29, 49" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 53, + 523, + 287, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 523, + 287, + 566 + ], + "spans": [ + { + "bbox": [ + 53, + 523, + 287, + 566 + ], + "type": "text", + "content": "[10] Siddhant Bansal, Chetan Arora, and C.V. Jawahar. My view is the best view: Procedure learning from egocentric videos. In European Conference on Computer Vision (ECCV), 2022. 3, 4, 7, 49" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 53, + 567, + 287, + 620 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 567, + 287, + 620 + ], + "spans": [ + { + "bbox": [ + 53, + 567, + 287, + 620 + ], + "type": "text", + "content": "[11] Yizhak Ben-Shabat, Xin Yu, Fatemehsadat Saleh, Dylan Campbell, Cristian Rodriguez-Opazo, Hongdong Li, and Stephen Gould. The ikea asm dataset: Understanding people assembling furniture through actions, objects and pose. 2020. 4" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 53, + 624, + 287, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 624, + 287, + 656 + ], + "spans": [ + { + "bbox": [ + 53, + 624, + 287, + 656 + ], + "type": "text", + "content": "[12] Gedas Bertasius, Hyun Soo Park, Stella Yu, and Jianbo Shi. Am i a baller? basketball performance assessment from first-person videos. In ICCV, 2017. 3, 8" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 53, + 658, + 287, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 658, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 53, + 658, + 287, + 689 + ], + "type": "text", + "content": "[13] Gedas Bertasius, Heng Wang, and Lorenzo Torresani. Is space-time attention all you need for video understanding? In ICML, 2021. 39, 43, 44, 47, 54, 55, 56" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 53, + 692, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 692, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 53, + 692, + 287, + 713 + ], + "type": "text", + "content": "[14] Jing Bi, Jiebo Luo, and Chenliang Xu. Procedure planning in instructional videos via contextual modeling and model" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 312, + 74, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 315, + 74, + 545, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 74, + 545, + 105 + ], + "spans": [ + { + "bbox": [ + 315, + 74, + 545, + 105 + ], + "type": "text", + "content": "based policy learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15611-15620, 2021. 49" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 312, + 107, + 545, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 107, + 545, + 160 + ], + "spans": [ + { + "bbox": [ + 312, + 107, + 545, + 160 + ], + "type": "text", + "content": "[15] Jing Bi, Jiebo Luo, and Chenliang Xu. Procedure planning in instructional videos via contextual modeling and model-based policy learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15611-15620, 2021. 4" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 312, + 162, + 545, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 162, + 545, + 215 + ], + "spans": [ + { + "bbox": [ + 312, + 162, + 545, + 215 + ], + "type": "text", + "content": "[16] Kay Henning Brodersen, Cheng Soon Ong, Klaas Enno Stephan, and Joachim M Buhmann. The balanced accuracy and its posterior distribution. In 2010 20th international conference on pattern recognition, pages 3121-3124. IEEE, 2010. 35" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 312, + 217, + 545, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 217, + 545, + 259 + ], + "spans": [ + { + "bbox": [ + 312, + 217, + 545, + 259 + ], + "type": "text", + "content": "[17] Meng Cao, Tianyu Yang, Junwu Weng, Can Zhang, Jue Wang, and Yuexian Zou. Locvtp: Video-text pre-training for temporal localization. In European Conference on Computer Vision, 2022. 4" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 312, + 262, + 545, + 325 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 262, + 545, + 325 + ], + "spans": [ + { + "bbox": [ + 312, + 262, + 545, + 325 + ], + "type": "text", + "content": "[18] Angela Castillo, Maria Escobar, Guillaume Jeanneret, Albert Pumarola, Pablo Arbeláez, Ali Thabet, and Artsiom Sanakoyeu. Bodiffusion: Diffusing sparse observations for full-body human motion synthesis. CV4Metaverse workshop, International Conference on Computer Vision, 2023. 57, 58" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 312, + 327, + 545, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 327, + 545, + 381 + ], + "spans": [ + { + "bbox": [ + 312, + 327, + 545, + 381 + ], + "type": "text", + "content": "[19] Eric R Chan, Koki Nagano, Matthew A Chan, Alexander W Bergman, Jeong Joon Park, Axel Levy, Miika Aittala, Shalini De Mello, Tero Karras, and Gordon Wetzstein. Generative novel view synthesis with 3d-aware diffusion models. arXiv preprint arXiv:2304.02602, 2023. 4" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 312, + 383, + 545, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 383, + 545, + 459 + ], + "spans": [ + { + "bbox": [ + 312, + 383, + 545, + 459 + ], + "type": "text", + "content": "[20] Angel Chang, Angela Dai, Tom Funkhouser, Matthias Nießner, Manolis Savva, Shuran Song, Andy Zeng, and Yinda Zhang. Matterport3d: Learning from rgb-d data in indoor environments. In Proceedings of the International Conference on 3D Vision (3DV), 2017. MatterPort3D dataset license available at: http://kaldir.vc.in.tum.de/matterport/MP_TOS.pdf.3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 312, + 460, + 545, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 460, + 545, + 502 + ], + "spans": [ + { + "bbox": [ + 312, + 460, + 545, + 502 + ], + "type": "text", + "content": "[21] Chien-Yi Chang, De-An Huang, Danfei Xu, Ehsan Adeli, Li Fei-Fei, and Juan Carlos Niebles. Procedure planning in instructional videos. In European Conference on Computer Vision, pages 334–350. Springer, 2020. 49" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 312, + 504, + 545, + 558 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 504, + 545, + 558 + ], + "spans": [ + { + "bbox": [ + 312, + 504, + 545, + 558 + ], + "type": "text", + "content": "[22] Chien-Yi Chang, De-An Huang, Danfei Xu, Ehsan Adeli, Li Fei-Fei, and Juan Carlos Niebles. Procedure planning in instructional videos. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part XI, pages 334–350. Springer, 2020. 4" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 312, + 559, + 545, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 559, + 545, + 602 + ], + "spans": [ + { + "bbox": [ + 312, + 559, + 545, + 602 + ], + "type": "text", + "content": "[23] Yu-Hsin Chen, Tien-Ju Yang, Joel Emer, and Vivienne Sze. Eyeriss v2: A flexible accelerator for emerging deep neural networks on mobile devices. IEEE Journal on Emerging and Selected Topics in Circuits and Systems, 2019. 46" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 312, + 604, + 545, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 604, + 545, + 646 + ], + "spans": [ + { + "bbox": [ + 312, + 604, + 545, + 646 + ], + "type": "text", + "content": "[24] Ho Kei Cheng and Alexander G Schwing. Xmem: Long-term video object segmentation with an atkinson-shiffrin memory model. In European Conference on Computer Vision, pages 640-658. Springer, 2022. 36" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 312, + 647, + 545, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 647, + 545, + 679 + ], + "spans": [ + { + "bbox": [ + 312, + 647, + 545, + 679 + ], + "type": "text", + "content": "[25] MMPose Contributors. Openmmlab pose estimation toolbox and benchmark. https://github.com/openmmlab/mmpose, 2020.56" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 312, + 681, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 681, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 312, + 681, + 545, + 712 + ], + "type": "text", + "content": "[26] Kellie Corona, Katie Osterdahl, Roderic Collins, and Anthony Hoogs. Meva: A large-scale multiview, multimodal video dataset for activity detection. In WACV, 2021. 3, 7" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19392" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 53, + 72, + 289, + 715 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 53, + 72, + 287, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 72, + 287, + 139 + ], + "spans": [ + { + "bbox": [ + 53, + 72, + 287, + 139 + ], + "type": "text", + "content": "[27] Dima Damen, Hazel Doughty, Giovanni Maria Farinella, Sanja Fidler, Antonino Furnari, Evangelos Kazakos, Davide Moltisanti, Jonathan Munro, Toby Perrett, Will Price, and Michael Wray. Scaling egocentric vision: The epic-kitchens dataset. In European Conference on Computer Vision (ECCV), 2018. 3, 27" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 53, + 140, + 288, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 140, + 288, + 195 + ], + "spans": [ + { + "bbox": [ + 53, + 140, + 288, + 195 + ], + "type": "text", + "content": "[28] Dima Damen, Hazel Doughty, Giovanni Maria Farinella, Antonino Furnari, Jian Ma, Evangelos Kazakos, Davide Moltisanti, Jonathan Munro, Toby Perrett, Will Price, and Michael Wray. Rescaling egocentric vision. *IJCV*, 2021. 2, 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 197, + 289, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 197, + 289, + 262 + ], + "spans": [ + { + "bbox": [ + 53, + 197, + 289, + 262 + ], + "type": "text", + "content": "[29] Dima Damen, Hazel Doughty, Giovanni Maria Farinella, Antonino Furnari, Evangelos Kazakos, Jian Ma, Davide Moltisanti, Jonathan Munro, Toby Perrett, Will Price, et al. Rescaling egocentric vision: collection, pipeline and challenges for epic-kitchens-100. International Journal of Computer Vision, 130(1):33-55, 2022. 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 264, + 288, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 264, + 288, + 330 + ], + "spans": [ + { + "bbox": [ + 53, + 264, + 288, + 330 + ], + "type": "text", + "content": "[30] Roeland De Geest, Efstratios Gavves, Amir Ghodrati, Zhenyang Li, Cees Snoek, and Tinne Tuytelaars. Online action detection. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part V 14, pages 269-284. Springer, 2016. 46, 51" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 331, + 288, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 331, + 288, + 385 + ], + "spans": [ + { + "bbox": [ + 53, + 331, + 288, + 385 + ], + "type": "text", + "content": "[31] Radosvet Desislavov, Fernando Martínez-Plumed, and José Hernández-Orallo. Trends in ai inference energy consumption: Beyond the performance-vs-parameter laws of deep learning. Sustainable Computing: Informatics and Systems, 38:100857, 2023. 46" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 387, + 288, + 419 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 387, + 288, + 419 + ], + "spans": [ + { + "bbox": [ + 53, + 387, + 288, + 419 + ], + "type": "text", + "content": "[32] Guodong Ding, Fadime Sener, Shugao Ma, and Angela Yao. Every mistake counts in assembly. arXiv preprint arXiv:2307.16453, 2023. 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 421, + 288, + 465 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 421, + 288, + 465 + ], + "spans": [ + { + "bbox": [ + 53, + 421, + 288, + 465 + ], + "type": "text", + "content": "[33] Keyan Ding, Kede Ma, Shiqi Wang, and Eero P Simoncelli. Image quality assessment: Unifying structure and texture similarity. IEEE transactions on pattern analysis and machine intelligence, 44(5):2567-2581, 2020. 39" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 467, + 288, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 467, + 288, + 498 + ], + "spans": [ + { + "bbox": [ + 53, + 467, + 288, + 498 + ], + "type": "text", + "content": "[34] H Doughty, D Damen, and W Mayol-Cuevas. Who's better? who's best? pairwise deep ranking for skill determination. In CVPR, 2018. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 500, + 288, + 534 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 500, + 288, + 534 + ], + "spans": [ + { + "bbox": [ + 53, + 500, + 288, + 534 + ], + "type": "text", + "content": "[35] Hazel Doughty, Walterio Mayol-Cuevas, and Dima Damen. The Pros and Cons: Rank-aware Temporal Attention for Skill Determination in Long Videos. 2019. 3, 8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 535, + 288, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 535, + 288, + 590 + ], + "spans": [ + { + "bbox": [ + 53, + 535, + 288, + 590 + ], + "type": "text", + "content": "[36] Nikita Dvornik, Isma Hadji, Hai Pham, Dhaivat Bhatt, Brais Martinez, Afsaneh Fazly, and Allan D Jepson. Flow graph to video grounding for weakly-supervised multi-step localization. In ECCV, pages 319-335. Springer, 2022. 4, 8" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 53, + 591, + 288, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 591, + 288, + 635 + ], + "spans": [ + { + "bbox": [ + 53, + 591, + 288, + 635 + ], + "type": "text", + "content": "[37] Ehsan Elhamifar and Dat Huynh. Self-supervised multitask procedure learning from instructional videos. In European Conference on Computer Vision, pages 557-573. Springer, 2020. 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 53, + 637, + 288, + 715 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 637, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 53, + 637, + 288, + 715 + ], + "type": "text", + "content": "[38] Jakob Engel, Kiran Somasundaram, Michael Goesele, Albert Sun, Alexander Gamino, Andrew Turner, Arjang Talattof, Arnie Yuan, Bilal Souti, Brighid Meredith, Cheng Peng, Chris Sweeney, Cole Wilson, Dan Barnes, Daniel DeTone, David Caruso, Derek Valleroy, Dinesh Ginjupalli, Duncan Frost, Edward Miller, Elias Mueggler, Evgeniy Oleinik, Fan Zhang, Guruprasad Somasundaram, Gustavo" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 312, + 72, + 547, + 715 + ], + "type": "list", + "angle": 0, + "index": 23, + "blocks": [ + { + "bbox": [ + 331, + 72, + 547, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 331, + 72, + 547, + 237 + ], + "spans": [ + { + "bbox": [ + 331, + 72, + 547, + 237 + ], + "type": "text", + "content": "Solaira, Harry Lanaras, Henry Howard-Jenkins, Huixuan Tang, Hyo Jin Kim, Jaime Rivera, Ji Luo, Jing Dong, Julian Straub, Kevin Bailey, Kevin Eckenhoff, Lingni Ma, Luis Pesqueira, Mark Schwesinger, Maurizio Monge, Nan Yang, Nick Charron, Nikhil Raina, Omkar Parkhi, Peter Borschowa, Pierre Moulon, Prince Gupta, Raul Mur-Artal, Robbie Pennington, Sachin Kulkarni, Sagar Miglani, Santosh Gondi, Saransh Solanki, Sean Diener, Shangyi Cheng, Simon Green, Steve Saarinen, Suvam Patra, Tassos Mourikis, Thomas Whelan, Tripti Singh, Vasileios Balntas, Vijay Baiyya, Wilson Dreewes, Xiaqing Pan, Yang Lou, Yipu Zhao, Yusuf Mansour, Yuyang Zou, Zhaoyang Lv, Zijian Wang, Mingfei Yan, Carl Ren, Renzo De Nardi, and Richard Newcombe. Project Aria: A new tool for egocentric multi-modal AI research, 2023. 3, 4, 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 312, + 238, + 547, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 238, + 547, + 282 + ], + "spans": [ + { + "bbox": [ + 312, + 238, + 547, + 282 + ], + "type": "text", + "content": "[39] Steven K Esser, Jeffrey L McKinstry, Deepika Bablani, Rathinakumar Appuswamy, and Dharmendra S Modha. Learned step size quantization. arXiv preprint arXiv:1902.08153, 2019. 8" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 312, + 283, + 547, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 283, + 547, + 326 + ], + "spans": [ + { + "bbox": [ + 312, + 283, + 547, + 326 + ], + "type": "text", + "content": "[40] Chenyou Fan, Jangwon Lee, Mingze Xu, Krishna Kumar Singh, Yong Jae Lee, David J Crandall, and Michael S Ryoo. Identifying first-person camera wearers in third-person videos. In CVPR, 2017. 4" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 312, + 327, + 547, + 371 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 327, + 547, + 371 + ], + "spans": [ + { + "bbox": [ + 312, + 327, + 547, + 371 + ], + "type": "text", + "content": "[41] Christoph Feichtenhofer. X3d: Expanding architectures for efficient video recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 203-213, 2020. 8, 45, 46, 47, 48" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 312, + 371, + 547, + 404 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 371, + 547, + 404 + ], + "spans": [ + { + "bbox": [ + 312, + 371, + 547, + 404 + ], + "type": "text", + "content": "[42] John H. Flavell, Eleanor R. Flavell, Frances L. Green, and Sharon A. Wilcox. The development of three spatial perspective-taking rules. Child Development, 1981. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 312, + 404, + 547, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 404, + 547, + 437 + ], + "spans": [ + { + "bbox": [ + 312, + 404, + 547, + 437 + ], + "type": "text", + "content": "[43] Ruohan Gao, Tae-Hyun Oh, Kristen Grauman, and Lorenzo Torresani. Listen to look: Action recognition by previewing audio. In CVPR, 2020. 8" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 312, + 437, + 547, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 437, + 547, + 491 + ], + "spans": [ + { + "bbox": [ + 312, + 437, + 547, + 491 + ], + "type": "text", + "content": "[44] Amir Ghodrati, Babak Ehteshami Bejnordi, and Amirhossein Habibian. Frameexit: Conditional early exiting for efficient video recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2021. 8" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 312, + 492, + 547, + 548 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 492, + 547, + 548 + ], + "spans": [ + { + "bbox": [ + 312, + 492, + 547, + 548 + ], + "type": "text", + "content": "[45] Rohit Girdhar, Mannat Singh, Nikhila Ravi, Laurens van der Maaten, Armand Joulin, and Ishan Misra. Omnivore: A single model for many visual modalities. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16102-16112, 2022. 54" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 312, + 548, + 547, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 548, + 547, + 624 + ], + "spans": [ + { + "bbox": [ + 312, + 548, + 547, + 624 + ], + "type": "text", + "content": "[46] Raghav Goyal, Samira Ebrahimi Kahou, Vincent Michalski, Joanna Materzynska, Susanne Westphal, Heuna Kim, Valentin Haenel, Ingo Fruend, Peter Yianilos, Moritz Mueller-Freitag, et al. The \"something something\" video database for learning and evaluating visual common sense. In Proceedings of the IEEE international conference on computer vision, pages 5842-5850, 2017. 42" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 312, + 625, + 547, + 715 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 625, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 312, + 625, + 547, + 715 + ], + "type": "text", + "content": "[47] Kristen Grauman, Andrew Westbury, Eugene Byrne, Zachary Chavis, Antonino Furnari, Rohit Girdhar, Jackson Hamburger, Hao Jiang, Miao Liu, Xingyu Liu, Miguel Martin, Tushar Nagarajan, Ilija Radosavovic, Santhosh Kumar Ramakrishnan, Fiona Ryan, Jayant Sharma, Michael Wray, Mengmeng Xu, Eric Zhongcong Xu, Chen Zhao, Siddhant Bansal, Dhruv Batra, Vincent Cartillier, Sean Crane, Tien Do, Morrie Doulaty, Akshay Erapalli, Christoph" + } + ] + } + ], + "index": 22 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "19393" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 53, + 72, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 72, + 72, + 287, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 72, + 72, + 287, + 281 + ], + "spans": [ + { + "bbox": [ + 72, + 72, + 287, + 281 + ], + "type": "text", + "content": "Feichtenhofer, Adriano Fragomeni, Qichen Fu, Abraham Gebreselasie, Cristina Gonzalez, James Hillis, Xuhua Huang, Yifei Huang, Wenqi Jia, Weslie Khoo, Jachym Kolář, Satwik Kottur, Anurag Kumar, Federico Landini, Chao Li, Yanghao Li, Zhenqiang Li, Karttikeya Mangalam, Raghava Modhugu, Jonathan Munro, Tullie Murrell, Takumi Nishiyasu, Will Price, Paola Ruiz, Merey Ramazanova, Leda Sari, Kiran Somasundaram, Audrey Southerland, Yusuke Sugano, Ruijie Tao, Minh Vo, Yuchen Wang, Xindi Wu, Takuma Yagi, Ziwei Zhao, Yunyi Zhu, Pablo Arbelaez, David Crandall, Dima Damen, Giovanni Maria Farinella, Christian Fuegen, Bernard Ghanem, Vamsi Krishna Ithapu, C. V. Jawahar, Hanbyul Joo, Kris Kitani, Haizhou Li, Richard Newcombe, Aude Oliva, Hyun Soo Park, James M. Rehg, Yoichi Sato, Jianbo Shi, Mike Zheng Shou, Antonio Torralba, Lorenzo Torresani, Mingfei Yan, and Jitendra Malik. Ego4D: Around the world in 3,000 hours of egocentric video. In CVPR, 2022. 2, 3, 4, 6, 7, 29, 34, 43" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 53, + 282, + 287, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 282, + 287, + 346 + ], + "spans": [ + { + "bbox": [ + 53, + 282, + 287, + 346 + ], + "type": "text", + "content": "[48] Chunhui Gu, Chen Sun, David A. Ross, Carl Vondrick, Caroline Pantofaru, Yeqing Li, Sudheendra Vijayanarasimhan, George Toderici, Susanna Ricco, Rahul Sukthankar, Cordelia Schmid, and Jitendra Malik. Ava: A video dataset of spatio-temporally localized atomic visual actions. In CVPR, 2018. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 348, + 286, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 348, + 286, + 392 + ], + "spans": [ + { + "bbox": [ + 53, + 348, + 286, + 392 + ], + "type": "text", + "content": "[49] Vladimir Guzov, Aymen Mir, Torsten Sattler, and Gerard Pons-Moll. Human poseitioning system (hps): 3d human pose estimation and self-localization in large scenes from body-mounted sensors. In CVPR, 2021. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 393, + 287, + 446 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 393, + 287, + 446 + ], + "spans": [ + { + "bbox": [ + 53, + 393, + 287, + 446 + ], + "type": "text", + "content": "[50] Shreyas Hampali, Mahdi Rad, Markus Oberweger, and Vincent Lepetit. Honnotate: A method for 3d annotation of hand and object poses. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3196-3206, 2020. 9" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 448, + 286, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 448, + 286, + 479 + ], + "spans": [ + { + "bbox": [ + 53, + 448, + 286, + 479 + ], + "type": "text", + "content": "[51] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, 2016. 36, 58" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 481, + 287, + 513 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 481, + 287, + 513 + ], + "spans": [ + { + "bbox": [ + 53, + 481, + 287, + 513 + ], + "type": "text", + "content": "[52] Geoffrey Hinton, Oriol Vinyals, and Jeff Dean. Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531, 2015. 43, 44" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 514, + 287, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 514, + 287, + 547 + ], + "spans": [ + { + "bbox": [ + 53, + 514, + 287, + 547 + ], + "type": "text", + "content": "[53] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in Neural Information Processing Systems, 33:6840-6851, 2020. 57" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 548, + 287, + 581 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 548, + 287, + 581 + ], + "spans": [ + { + "bbox": [ + 53, + 548, + 287, + 581 + ], + "type": "text", + "content": "[54] Alain Hore and Djemel Ziou. Image quality metrics: Psnr vs. ssim. In 2010 20th international conference on pattern recognition, pages 2366-2369. IEEE, 2010. 39" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 582, + 287, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 582, + 287, + 624 + ], + "spans": [ + { + "bbox": [ + 53, + 582, + 287, + 624 + ], + "type": "text", + "content": "[55] Mark Horowitz. 1.1 computing's energy problem (and what we can do about it). In 2014 IEEE international solid-state circuits conference digest of technical papers (ISSCC), 2014. 46" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 625, + 287, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 625, + 287, + 679 + ], + "spans": [ + { + "bbox": [ + 53, + 625, + 287, + 679 + ], + "type": "text", + "content": "[56] Andrew G Howard, Menglong Zhu, Bo Chen, Dmitry Kalenichenko, Weijun Wang, Tobias Weyand, Marco Andretto, and Hartwig Adam. Mobilenets: Efficient convolutional neural networks for mobile vision applications. arXiv preprint arXiv:1704.04861, 2017. 8" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 53, + 681, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 681, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 53, + 681, + 287, + 713 + ], + "type": "text", + "content": "[57] Dong-Hyun Hwang, Kohei Aso, Ye Yuan, Kris Kitani, and Hideki Koike. Monoeye: Multimodal human motion capture system using a single ultra-wide fisheye camera. In" + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 312, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 331, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 331, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 331, + 73, + 545, + 95 + ], + "type": "text", + "content": "Proceedings of the 33rd Annual ACM Symposium on User Interface Software and Technology, pages 98-111, 2020. 9" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 312, + 97, + 545, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 97, + 545, + 150 + ], + "spans": [ + { + "bbox": [ + 312, + 97, + 545, + 150 + ], + "type": "text", + "content": "[58] Forrest N Iandola, Song Han, Matthew W Moskewicz, Khalid Ashraf, William J Dally, and Kurt Keutzer. SqueezeNet: Alexnet-level accuracy with 50x fewer parameters and; 0.5 mb model size. arXiv preprint arXiv:1602.07360, 2016. 8" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 312, + 152, + 545, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 152, + 545, + 195 + ], + "spans": [ + { + "bbox": [ + 312, + 152, + 545, + 195 + ], + "type": "text", + "content": "[59] Vladimir Iashin and Esa Rahtu. Multi-modal dense video captioning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops, pages 958-959, 2020. 26" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 312, + 198, + 545, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 198, + 545, + 262 + ], + "spans": [ + { + "bbox": [ + 312, + 198, + 545, + 262 + ], + "type": "text", + "content": "[60] Hassan Ismail Fawaz, Germain Forestier, Jonathan Weber, Lhassane Idoumghar, and Pierre-Alain Muller. Evaluating surgical skills from kinematic data using convolutional neural networks. In Medical Image Computing and Computer Assisted Intervention - MICCAI 2018, pages 214-221, 2018. 8" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 312, + 264, + 545, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 264, + 545, + 297 + ], + "spans": [ + { + "bbox": [ + 312, + 264, + 545, + 297 + ], + "type": "text", + "content": "[61] Phillip Isola, Jun-Yan Zhu, Tinghui Zhou, and Alexei A Efros. Image-to-image translation with conditional adversarial networks. CVPR, 2017. 39" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 312, + 299, + 545, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 299, + 545, + 342 + ], + "spans": [ + { + "bbox": [ + 312, + 299, + 545, + 342 + ], + "type": "text", + "content": "[62] Yunseok Jang, Sungryull Sohn, Lajanugen Logeswaran, Tiange Luo, Moontae Lee, and Honglak Lee. Multimodal subtask graph generation from instructional videos. arXiv preprint arXiv:2302.08672, 2023. 8" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 312, + 343, + 545, + 375 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 343, + 545, + 375 + ], + "spans": [ + { + "bbox": [ + 312, + 343, + 545, + 375 + ], + "type": "text", + "content": "[63] Hao Jiang and Kristen Grauman. Seeing invisible poses: Estimating 3d body pose from egocentric video. In CVPR, 2017. 9" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 312, + 377, + 545, + 432 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 377, + 545, + 432 + ], + "spans": [ + { + "bbox": [ + 312, + 377, + 545, + 432 + ], + "type": "text", + "content": "[64] Jiaxi Jiang, Paul Streli, Huajian Qiu, Andreas Fender, Larissa Laich, Patrick Snape, and Christian Holz. Avatarposer: Articulated full-body pose tracking from sparse motion sensing. In European Conference on Computer Vision, pages 443-460. Springer, 2022. 57, 58" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 312, + 434, + 545, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 434, + 545, + 487 + ], + "spans": [ + { + "bbox": [ + 312, + 434, + 545, + 487 + ], + "type": "text", + "content": "[65] Wei Jiang, Eduard Trulls, Jan Hosang, Andrea Tagliasacchi, and Kwang Moo Yi. Cotr: Correspondence transformer for matching across images. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6207-6217, 2021. 6, 35" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 312, + 490, + 545, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 490, + 545, + 555 + ], + "spans": [ + { + "bbox": [ + 312, + 490, + 545, + 555 + ], + "type": "text", + "content": "[66] Hanbyul Joo, Tomas Simon, Xulong Li, Hao Liu, Lei Tan, Lin Gui, Sean Banerjee, Timothy Scott Godisart, Bart Nabbe, Iain Matthews, Takeo Kanade, Shohei Nobuhara, and Yaser Sheikh. Panoptic studio: A massively multiview system for social interaction capture. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2017. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 312, + 556, + 545, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 556, + 545, + 610 + ], + "spans": [ + { + "bbox": [ + 312, + 556, + 545, + 610 + ], + "type": "text", + "content": "[67] Will Kay, Joao Carreira, Karen Simonyan, Brian Zhang, Chloe Hillier, Sudheendra Vijayanarasimhan, Fabio Viola, Tim Green, Trevor Back, Paul Natsev, et al. The kinetics human action video dataset. arXiv preprint arXiv:1705.06950, 2017. 2, 42, 43" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 312, + 613, + 545, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 613, + 545, + 645 + ], + "spans": [ + { + "bbox": [ + 312, + 613, + 545, + 645 + ], + "type": "text", + "content": "[68] Rawal Khirodkar, Aayush Bansal, Lingni Ma, Richard Newcombe, Minh Vo, and Kris Kitani. Egohumans: An egocentric 3d multi-human benchmark. In ICCV, 2023. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 312, + 647, + 545, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 647, + 545, + 678 + ], + "spans": [ + { + "bbox": [ + 312, + 647, + 545, + 678 + ], + "type": "text", + "content": "[69] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 36" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 312, + 681, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 681, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 312, + 681, + 545, + 713 + ], + "type": "text", + "content": "[70] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C. Berg, Wan-Yen Lo, Piotr Dollar," + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "19394" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "bbox": [ + 53, + 72, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 72, + 72, + 288, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 72, + 72, + 288, + 106 + ], + "spans": [ + { + "bbox": [ + 72, + 72, + 288, + 106 + ], + "type": "text", + "content": "and Ross Girshick. Segment anything. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 4015-4026, 2023. 33" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 53, + 107, + 288, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 107, + 288, + 173 + ], + "spans": [ + { + "bbox": [ + 53, + 107, + 288, + 173 + ], + "type": "text", + "content": "[71] Dohwan Ko, Joonmyung Choi, Juyeon Ko, Shinyeong Noh, Kyoung-Woon On, Eun-Sol Kim, and Hyunwoo J Kim. Video-text representation learning via differentiable weak temporal alignment. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5016-5025, 2022. 4" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 175, + 288, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 175, + 288, + 229 + ], + "spans": [ + { + "bbox": [ + 53, + 175, + 288, + 229 + ], + "type": "text", + "content": "[72] Nikos Kolotouros, Georgios Pavlakos, Michael J Black, and Kostas Daniilidis. Learning to reconstruct 3d human pose and shape via model-fitting in the loop. In Proceedings of the IEEE/CVF international conference on computer vision, pages 2252-2261, 2019. 59" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 231, + 288, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 231, + 288, + 274 + ], + "spans": [ + { + "bbox": [ + 53, + 231, + 288, + 274 + ], + "type": "text", + "content": "[73] Bruno Korbar, Du Tran, and Lorenzo Torresani. Scsampler: Sampling salient clips from video for efficient action recognition. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2019. 8" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 276, + 288, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 276, + 288, + 330 + ], + "spans": [ + { + "bbox": [ + 53, + 276, + 288, + 330 + ], + "type": "text", + "content": "[74] Zuzana Kukelova, Jan Heller, and Andrew Fitzgibbon. Efficient intersection of three quadrics and applications in computer vision. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 1799-1808, 2016. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 332, + 288, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 332, + 288, + 386 + ], + "spans": [ + { + "bbox": [ + 53, + 332, + 288, + 386 + ], + "type": "text", + "content": "[75] Iljung Kwak, Jian-Zhong Guo, Adam Hantman, David Kriegman, and Kristin Branson. Detecting the starting frame of actions in video. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 489-497, 2020. 54" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 388, + 288, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 388, + 288, + 442 + ], + "spans": [ + { + "bbox": [ + 53, + 388, + 288, + 442 + ], + "type": "text", + "content": "[76] Taein Kwon, Bugra Tekin, Jan Stühmer, Federica Bogo, and Marc Pollefeys. H2o: Two hands manipulating objects for first person interaction recognition. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 10138-10148, 2021. 2, 3, 9, 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 445, + 288, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 445, + 288, + 498 + ], + "spans": [ + { + "bbox": [ + 53, + 445, + 288, + 498 + ], + "type": "text", + "content": "[77] F. De la Torre, J. Hodgins, J. Montano, S. Valcarcel, R. Forcada, and J. Macey. Guide to the carnegie mellon university multimodal activity (cmu-mmac) database. In Tech. report CMU-RI-TR-08-22, Robotics Institute, Carnegie Mellon University, 2009. 2, 3, 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 500, + 288, + 532 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 500, + 288, + 532 + ], + "spans": [ + { + "bbox": [ + 53, + 500, + 288, + 532 + ], + "type": "text", + "content": "[78] Y. J. Lee, J. Ghosh, and K. Grauman. Discovering important people and objects for egocentric video summarization. In CVPR, 2012. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 534, + 288, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 534, + 288, + 599 + ], + "spans": [ + { + "bbox": [ + 53, + 534, + 288, + 599 + ], + "type": "text", + "content": "[79] Jiefeng Li, Chao Xu, Zhicun Chen, Siyuan Bian, Lixin Yang, and Cewu Lu. Hybrik: A hybrid analytical-neural inverse kinematics solution for 3d human pose and shape estimation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3383-3393, 2021. 59" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 53, + 601, + 288, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 601, + 288, + 644 + ], + "spans": [ + { + "bbox": [ + 53, + 601, + 288, + 644 + ], + "type": "text", + "content": "[80] Jiaman Li, Karen Liu, and Jiajun Wu. Ego-body pose estimation via ego-head pose estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17142-17151, 2023. 3, 9, 56, 57, 58" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 53, + 647, + 288, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 647, + 288, + 678 + ], + "spans": [ + { + "bbox": [ + 53, + 647, + 288, + 678 + ], + "type": "text", + "content": "[81] Yin Li, Miao Liu, and James M. Rehg. In the eye of beholder: Joint learning of gaze and actions in first person video. In ECCV, 2018. 3, 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 53, + 681, + 288, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 681, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 53, + 681, + 288, + 713 + ], + "type": "text", + "content": "[82] Yanghao Li, Tushar Nagarajan, Bo Xiong, and Kristen Grauman. Ego-exo: Transferring visual representations from third-person to first-person videos. In Proceedings of" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 312, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 331, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 331, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 331, + 73, + 545, + 95 + ], + "type": "text", + "content": "the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6943-6953, 2021. 4, 7, 21, 43, 44" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 312, + 97, + 545, + 151 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 97, + 545, + 151 + ], + "spans": [ + { + "bbox": [ + 312, + 97, + 545, + 151 + ], + "type": "text", + "content": "[83] Junhua Liao, Haihan Duan, Kanghui Feng, Wanbing Zhao, Yanbing Yang, and Liangyin Chen. A light weight model for active speaker detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22932-22941, 2023. 47, 48" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 312, + 152, + 545, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 152, + 545, + 196 + ], + "spans": [ + { + "bbox": [ + 312, + 152, + 545, + 196 + ], + "type": "text", + "content": "[84] Kevin Lin, Lijuan Wang, and Zicheng Liu. End-to-end human pose and mesh reconstruction with transformers. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1954-1963, 2021. 59" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 312, + 198, + 545, + 242 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 198, + 545, + 242 + ], + "spans": [ + { + "bbox": [ + 312, + 198, + 545, + 242 + ], + "type": "text", + "content": "[85] Kevin Qinghong Lin, Alex Jinpeng Wang, Mattia Soldan, Michael Wray, Rui Yan, Eric Zhongcong Xu, Difei Gao, Rongcheng Tu, Wenzhe Zhao, Weijie Kong, et al. Egocentric video-language pretraining. NeurIPS, 2022. 6, 26, 50" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 312, + 243, + 545, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 243, + 545, + 274 + ], + "spans": [ + { + "bbox": [ + 312, + 243, + 545, + 274 + ], + "type": "text", + "content": "[86] TY Lin, Y Cui, S Belongie, and J Hays. Learning deep representations for ground-to-aerial geolocation. In CVPR, 2015. 4" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 312, + 277, + 545, + 320 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 277, + 545, + 320 + ], + "spans": [ + { + "bbox": [ + 312, + 277, + 545, + 320 + ], + "type": "text", + "content": "[87] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dolkar, and C Lawrence Zitnick. Microsoft COCO: Common objects in context. In ECCV, 2014. 9" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 312, + 323, + 545, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 323, + 545, + 376 + ], + "spans": [ + { + "bbox": [ + 312, + 323, + 545, + 376 + ], + "type": "text", + "content": "[88] Tsung-Yi Lin, Piotr Dólar, Ross Girshick, Kaiming He, Bharath Hariharan, and Serge Belongie. Feature pyramid networks for object detection. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2117-2125, 2017. 58" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 312, + 378, + 545, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 378, + 545, + 442 + ], + "spans": [ + { + "bbox": [ + 312, + 378, + 545, + 442 + ], + "type": "text", + "content": "[89] Xudong Lin, Fabio Petroni, Gedas Bertasius, Marcus Rohrbach, Shih-Fu Chang, and Lorenzo Torresani. Learning to recognize procedural activities with distant supervision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13853-13863, 2022. 4, 29" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 312, + 445, + 545, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 445, + 545, + 510 + ], + "spans": [ + { + "bbox": [ + 312, + 445, + 545, + 510 + ], + "type": "text", + "content": "[90] Andrew Liu, Richard Tucker, Varun Jampani, Ameesh Makadia, Noah Snavely, and Angjoo Kanazawa. Infinite nature: Perpetual view generation of natural scenes from a single image. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14458-14467, 2021. 4" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 312, + 512, + 545, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 512, + 545, + 578 + ], + "spans": [ + { + "bbox": [ + 312, + 512, + 545, + 578 + ], + "type": "text", + "content": "[91] Chiao Liu, Lyle Bainbridge, Andrew Berkovich, Song Chen, Wei Gao, Tsung-Hsun Tsai, Kazuya Mori, Rimon Ikeno, Masayuki Uno, Toshiyuki Isozaki, et al. A " + }, + { + "bbox": [ + 312, + 512, + 545, + 578 + ], + "type": "inline_equation", + "content": "4.6\\mu \\mathrm{m}" + }, + { + "bbox": [ + 312, + 512, + 545, + 578 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 312, + 512, + 545, + 578 + ], + "type": "inline_equation", + "content": "512\\times 512" + }, + { + "bbox": [ + 312, + 512, + 545, + 578 + ], + "type": "text", + "content": ", ultra-low power stacked digital pixel sensor with triple quantization and 127db dynamic range. In 2020 IEEE International Electron Devices Meeting (IEDM), 2020. 46" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 312, + 580, + 545, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 580, + 545, + 634 + ], + "spans": [ + { + "bbox": [ + 312, + 580, + 545, + 634 + ], + "type": "text", + "content": "[92] Daochang Liu, Qiyue Li, Tingting Jiang, Yizhou Wang, Rulin Miao, Fei Shan, and Ziyu Li. Towards unified surgical skill assessment. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9522-9531, 2021. 8" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 312, + 635, + 545, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 635, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 312, + 635, + 545, + 689 + ], + "type": "text", + "content": "[93] Gaowen Liu, Hao Tang, Hugo Latapie, and Yan Yan. Exocentric to egocentric image generation via parallel generative adversarial network. In ICASSP 2020-2020 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 1843-1847. IEEE, 2020. 7" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 312, + 692, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 692, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 312, + 692, + 545, + 713 + ], + "type": "text", + "content": "[94] Gaowen Liu, Hao Tang, Hugo M Latapie, Jason J Corso, and Yan Yan. Cross-view exocentric to egocentric video" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19395" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 72, + 72, + 287, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 72, + 72, + 287, + 95 + ], + "spans": [ + { + "bbox": [ + 72, + 72, + 287, + 95 + ], + "type": "text", + "content": "synthesis. In Proceedings of the 29th ACM International Conference on Multimedia, pages 974-982, 2021. 7" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 53, + 96, + 287, + 129 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 96, + 287, + 129 + ], + "spans": [ + { + "bbox": [ + 53, + 96, + 287, + 129 + ], + "type": "text", + "content": "[95] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In International Conference on Learning Representations, 2019. 47" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 130, + 287, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 130, + 287, + 185 + ], + "spans": [ + { + "bbox": [ + 53, + 130, + 287, + 185 + ], + "type": "text", + "content": "[96] Xiaohu Lu, Zuoyue Li, Zhaopeng Cui, Martin R. Oswald, Marc Pollefeys, and Rongjun Qin. Geometry-aware satellite-to-ground image synthesis for urban areas. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 186, + 287, + 219 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 186, + 287, + 219 + ], + "spans": [ + { + "bbox": [ + 53, + 186, + 287, + 219 + ], + "type": "text", + "content": "[97] Mi Luo, Zihui Xue, Alex Dimakis, and Kristen Grauman. Put myself in your shoes: Lifting the egocentric perspective from exocentric videos. arXiv:2403.06351, 2024. 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 220, + 287, + 263 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 220, + 287, + 263 + ], + "spans": [ + { + "bbox": [ + 53, + 220, + 287, + 263 + ], + "type": "text", + "content": "[98] Zhengyi Luo, Ryo Hachiuma, Ye Yuan, and Kris Kitani. Dynamics-regulated kinematic policy for egocentric pose estimation. In Advances in Neural Information Processing Systems, 2021. 9, 56, 57, 58" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 266, + 287, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 266, + 287, + 308 + ], + "spans": [ + { + "bbox": [ + 53, + 266, + 287, + 308 + ], + "type": "text", + "content": "[99] Naureen Mahmood, Nima Ghorbani, Nikolaus F. Troje, Gerard Pons-Moll, and Michael J. Black. Amass: Archive of motion capture as surface shapes. In The IEEE International Conference on Computer Vision (ICCV), 2019. 57" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 310, + 287, + 343 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 310, + 287, + 343 + ], + "spans": [ + { + "bbox": [ + 48, + 310, + 287, + 343 + ], + "type": "text", + "content": "[100] Effrosyni Mavroudi, Triantafyllos Afouras, and Lorenzo Torresani. Learning to ground instructional articles in videos through narrations. 2022. 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 344, + 287, + 378 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 344, + 287, + 378 + ], + "spans": [ + { + "bbox": [ + 48, + 344, + 287, + 378 + ], + "type": "text", + "content": "[101] Sachin Mehta and Mohammad Rastegari. Mobilevit: lightweight, general-purpose, and mobile-friendly vision transformer. arXiv preprint arXiv:2110.02178, 2021. 8" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 379, + 287, + 422 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 379, + 287, + 422 + ], + "spans": [ + { + "bbox": [ + 48, + 379, + 287, + 422 + ], + "type": "text", + "content": "[102] Yue Meng, Chung-Ching Lin, Rameswar Panda, Prasanna Sattigeri, Leonid Karlinsky, Aude Oliva, Kate Saenko, and Rogerio Feris. Ar-net: Adaptive frame resolution for efficient action recognition. In ECCV 2020, 2020. 8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 423, + 287, + 478 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 423, + 287, + 478 + ], + "spans": [ + { + "bbox": [ + 48, + 423, + 287, + 478 + ], + "type": "text", + "content": "[103] Antoine Miech, Dimitri Zhukov, Jean-Baptiste Alayrac, Makarand Tapaswi, Ivan Laptev, and Josef Sivic. HowTo100M: Learning a Text-Video Embedding by Watching Hundred Million Narrated Video Clips. In ICCV, 2019. 2, 4, 29" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 479, + 287, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 479, + 287, + 544 + ], + "spans": [ + { + "bbox": [ + 48, + 479, + 287, + 544 + ], + "type": "text", + "content": "[104] Antoine Miech, Jean-Baptiste Alayrac, Lucas Smaira, Ivan Laptev, Josef Sivic, and Andrew Zisserman. End-to-end learning of visual representations from uncurated instructional videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9879-9889, 2020. 4, 29" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 546, + 287, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 546, + 287, + 601 + ], + "spans": [ + { + "bbox": [ + 48, + 546, + 287, + 601 + ], + "type": "text", + "content": "[105] Mathew Monfort, Alex Andonian, Bolei Zhou, Kandan Ramakrishnan, Sarah Adel Bargal, Tom Yan, Lisa Brown, Quanfu Fan, Dan Gutfreund, Carl Vondrick, and Aude Oliva. Moments in time dataset: one million videos for event understanding. PAMI, 2019. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 602, + 287, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 602, + 287, + 668 + ], + "spans": [ + { + "bbox": [ + 48, + 602, + 287, + 668 + ], + "type": "text", + "content": "[106] Gyeongsik Moon, Shoou-I Yu, He Wen, Takaaki Shiratori, and Kyoung Mu Lee. Interhand2. 6m: A dataset and baseline for 3d interacting hand pose estimation from a single rgb image. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XX 16, pages 548-564. Springer, 2020. 9" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "type": "text", + "content": "[107] Medhini Narasimhan, Licheng Yu, Sean Bell, Ning Zhang, and Trevor Darrell. Learning and verification of task structure in instructional videos. arXiv preprint arXiv:2303.13519, 2023. 4, 8" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 31, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 105 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 105 + ], + "type": "text", + "content": "[108] Nora Newcombe. The development of spatial perspective taking. Advances in child development and behavior, 1989. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 107, + 545, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 107, + 545, + 172 + ], + "spans": [ + { + "bbox": [ + 307, + 107, + 545, + 172 + ], + "type": "text", + "content": "[109] Takehiko Ohkawa, Kun He, Fadime Sener, Tomas Hodan, Luan Tran, and Cem Keskin. Assemblyhands: Towards egocentric activity understanding via 3d hand pose estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12999-13008, 2023. 9, 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 174, + 545, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 174, + 545, + 206 + ], + "spans": [ + { + "bbox": [ + 307, + 174, + 545, + 206 + ], + "type": "text", + "content": "[110] Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748, 2018. 43, 44" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 208, + 545, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 208, + 545, + 251 + ], + "spans": [ + { + "bbox": [ + 307, + 208, + 545, + 251 + ], + "type": "text", + "content": "[111] Boxiao Pan, Haoye Cai, De-An Huang, Kuan-Hui Lee, Adrien Gaidon, Ehsan Adeli, and Juan Carlos Niebles. Spatio-temporal graph for video captioning with knowledge distillation. In CVPR, 2020. 26" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 253, + 545, + 307 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 253, + 545, + 307 + ], + "spans": [ + { + "bbox": [ + 307, + 253, + 545, + 307 + ], + "type": "text", + "content": "[112] JoonKyu Park, Yeonguk Oh, Gyeongsik Moon, Hongsuk Choi, and Kyoung Mu Lee. Handoccnet: Occlusion-robust 3d hand mesh estimation network. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1496–1505, 2022. 58, 59" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 308, + 545, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 308, + 545, + 331 + ], + "spans": [ + { + "bbox": [ + 307, + 308, + 545, + 331 + ], + "type": "text", + "content": "[113] Paritosh Parmar and Brendan Morris. Action quality assessment across multiple actions. In WACV, 2019. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 332, + 545, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 332, + 545, + 353 + ], + "spans": [ + { + "bbox": [ + 307, + 332, + 545, + 353 + ], + "type": "text", + "content": "[114] Paritosh Parmar and Brendan Tran Morris. Learning to score olympic events, 2017. 8" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 354, + 545, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 354, + 545, + 408 + ], + "spans": [ + { + "bbox": [ + 307, + 354, + 545, + 408 + ], + "type": "text", + "content": "[115] Paritosh Parmar and Brendan Tran Morris. What and how well you performed? a multitask learning approach to action quality assessment. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 304-313, 2019. 8" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 410, + 545, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 410, + 545, + 442 + ], + "spans": [ + { + "bbox": [ + 307, + 410, + 545, + 442 + ], + "type": "text", + "content": "[116] William Peebles and Saining Xie. Scalable diffusion models with transformers. arXiv preprint arXiv:2212.09748, 2022.39" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 445, + 545, + 509 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 445, + 545, + 509 + ], + "spans": [ + { + "bbox": [ + 307, + 445, + 545, + 509 + ], + "type": "text", + "content": "[117] Federico Perazzi, Jordi Pont-Tuset, Brian McWilliams, Luc Van Gool, Markus Gross, and Alexander Sorkine-Hornung. A benchmark dataset and evaluation methodology for video object segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 724-732, 2016. 35, 39" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 510, + 545, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 510, + 545, + 555 + ], + "spans": [ + { + "bbox": [ + 307, + 510, + 545, + 555 + ], + "type": "text", + "content": "[118] Ethan Perez, Florian Strub, Harm De Vries, Vincent Dumoulin, and Aaron Courville. Film: Visual reasoning with a general conditioning layer. In Proceedings of the AAAI conference on artificial intelligence, 2018. 39" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 555, + 545, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 555, + 545, + 578 + ], + "spans": [ + { + "bbox": [ + 307, + 555, + 545, + 578 + ], + "type": "text", + "content": "[119] H. Pirsiavash and D. Ramanan. Detecting activities of daily living in first-person camera views. In CVPR, 2012. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 579, + 545, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 579, + 545, + 601 + ], + "spans": [ + { + "bbox": [ + 307, + 579, + 545, + 601 + ], + "type": "text", + "content": "[120] H. Piriavash, C. Vondrick, and A. Torralba. Assessing the quality of actions. In ECCV, 2014. 3, 8" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 602, + 545, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 602, + 545, + 634 + ], + "spans": [ + { + "bbox": [ + 307, + 602, + 545, + 634 + ], + "type": "text", + "content": "[121] Antonio Polino, Razvan Pascanu, and Dan Alistarh. Model compression via distillation and quantization. arXiv preprint arXiv:1802.05668, 2018. 8" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 635, + 545, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 635, + 545, + 690 + ], + "spans": [ + { + "bbox": [ + 307, + 635, + 545, + 690 + ], + "type": "text", + "content": "[122] Daniel Povey, Arnab Ghoshal, Gilles Boulianne, Lukávs. Burget, Ondrej Glembek, Nagendra Kumar Goel, Mirko Hannemann, Petr Motlícek, Yanmin Qian, Petr Schwarz, Jan Silovsky, Georg Stemmer, and Karel Vesely. The kaldi speech recognition toolkit. 2011. 47" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 692, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 692, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 692, + 545, + 713 + ], + "type": "text", + "content": "[123] Shraman Pramanick, Yale Song, Sayan Nag, Kevin Qinghong Lin, Hardik Shah, Mike Zheng Shou," + } + ] + } + ], + "index": 30 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "19396" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 72, + 72, + 287, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 72, + 72, + 287, + 127 + ], + "spans": [ + { + "bbox": [ + 72, + 72, + 287, + 127 + ], + "type": "text", + "content": "Rama Chellappa, and Pengchuan Zhang. Egovlpv2: Egocentric video-language pre-training with fusion in the backbone. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5285-5297, 2023. 6, 26, 43, 44" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 129, + 287, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 129, + 287, + 194 + ], + "spans": [ + { + "bbox": [ + 48, + 129, + 287, + 194 + ], + "type": "text", + "content": "[124] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 39" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 196, + 287, + 250 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 196, + 287, + 250 + ], + "spans": [ + { + "bbox": [ + 49, + 196, + 287, + 250 + ], + "type": "text", + "content": "[125] Alec Radford, Jong Wook Kim, Tao Xu, Greg Brockman, Christine McLeavey, and Ilya Sutskever. Robust speech recognition via large-scale weak supervision. In International Conference on Machine Learning, pages 28492-28518. PMLR, 2023. 27" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 252, + 287, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 252, + 287, + 296 + ], + "spans": [ + { + "bbox": [ + 49, + 252, + 287, + 296 + ], + "type": "text", + "content": "[126] Francesco Ragusa, Antonino Furnari, Salvatore Livatino, and Giovanni Maria Farinella. The meccano dataset: Understanding human-object interactions from egocentric videos in an industrial-like domain. In WACV, 2021. 3, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 49, + 297, + 287, + 329 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 297, + 287, + 329 + ], + "spans": [ + { + "bbox": [ + 49, + 297, + 287, + 329 + ], + "type": "text", + "content": "[127] N. Rai, H. Chen, J. Ji, R. Desai, K. Kozuka, S. Ishizaka, E. Adeli, and J.C. Niebles. Home action genome: Contrastive compositional action understanding. In CVPR, 2021. 2, 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 49, + 331, + 287, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 331, + 287, + 418 + ], + "spans": [ + { + "bbox": [ + 49, + 331, + 287, + 418 + ], + "type": "text", + "content": "[128] Santhosh Kumar Ramakrishnan, Aaron Gokaslan, Erik Wijmans, Oleksandr Maksymets, Alexander Clegg, John M Turner, Eric Undersander, Wojciech Galuba, Andrew Westbury, Angel X Chang, Manolis Savva, Yili Zhao, and Dhruv Batra. Habitat-matterport 3d dataset (HM3d): 1000 large-scale 3d environments for embodied AI. In Thirty-fifth Conference on Neural Information Processing Systems Datasets and Benchmarks Track, 2021. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 49, + 420, + 287, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 420, + 287, + 464 + ], + "spans": [ + { + "bbox": [ + 49, + 420, + 287, + 464 + ], + "type": "text", + "content": "[129] Scott Reed, Honglak Lee, Dragomir Anguelov, Christian Szegedy, Dumitru Erhan, and Andrew Rabinovich. Training deep neural networks on noisy labels with bootstrapping. arXiv preprint arXiv:1412.6596, 2014. 39" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 49, + 465, + 287, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 465, + 287, + 498 + ], + "spans": [ + { + "bbox": [ + 49, + 465, + 287, + 498 + ], + "type": "text", + "content": "[130] Krishna Regmi and Ali Borji. Cross-view image synthesis using conditional gans. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 4, 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 49, + 499, + 287, + 532 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 499, + 287, + 532 + ], + "spans": [ + { + "bbox": [ + 49, + 499, + 287, + 532 + ], + "type": "text", + "content": "[131] Krishna Regmi and Ali Borji. Cross-view image synthesis using geometry-guided conditional gans. Computer Vision and Image Understanding, 2019. 4" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 49, + 533, + 287, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 533, + 287, + 555 + ], + "spans": [ + { + "bbox": [ + 49, + 533, + 287, + 555 + ], + "type": "text", + "content": "[132] Krishna Regmi and Mubarak Shah. Bridging the domain gap for ground-to-aerial image matching. In ICCV, 2019. 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 49, + 556, + 287, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 556, + 287, + 600 + ], + "spans": [ + { + "bbox": [ + 49, + 556, + 287, + 600 + ], + "type": "text", + "content": "[133] Jeremy Reizenstein, Roman Shapovalov, Philipp Henzler, Luca Sbordone, Patrick Labatut, and David Novotny. Common objects in 3d: Large-scale learning and evaluation of real-life 3d category reconstruction. In ICCV, 2021. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 49, + 601, + 287, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 601, + 287, + 633 + ], + "spans": [ + { + "bbox": [ + 49, + 601, + 287, + 633 + ], + "type": "text", + "content": "[134] Bin Ren, Hao Tang, and Nicu Sebe. Cascaded cross mlp-mixer gans for cross-view image translation. arXiv preprint arXiv:2110.10183, 2021. 4" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 49, + 635, + 287, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 635, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 49, + 635, + 287, + 689 + ], + "type": "text", + "content": "[135] Xuanchi Ren and Xiaolong Wang. Look outside the room: Synthesizing a consistent long-term 3d scene video from a single image. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3563-3573, 2022. 4" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 49, + 691, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 691, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 49, + 691, + 287, + 712 + ], + "type": "text", + "content": "[136] Helge Rhodin, Christian Richardt, Dan Casas, Eldar In-safutdinov, Mohammad Shafiei, Hans-Peter Seidel, Bernt" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 331, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 331, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 331, + 73, + 545, + 106 + ], + "type": "text", + "content": "Schiele, and Christian Theobalt. Egocap: egocentric marker-less motion capture with two fisheye cameras. ACM Transactions on Graphics (TOG), 35(6):1-11, 2016. 9" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 107, + 545, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 107, + 545, + 150 + ], + "spans": [ + { + "bbox": [ + 308, + 107, + 545, + 150 + ], + "type": "text", + "content": "[137] Robin Rombach, Patrick Esser, and Björn Ommer. Geometry-free view synthesis: Transformers and no 3d priors. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14356-14366, 2021. 4" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 152, + 545, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 152, + 545, + 194 + ], + "spans": [ + { + "bbox": [ + 308, + 152, + 545, + 194 + ], + "type": "text", + "content": "[138] Javier Romero, Dimitrios Tzionas, and Michael J. Black. Embodied hands: Modeling and capturing hands and bodies together. ACM Transactions on Graphics, (Proc. SIGGRAPH Asia), 36(6), 2017. 59" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 195, + 545, + 260 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 195, + 545, + 260 + ], + "spans": [ + { + "bbox": [ + 308, + 195, + 545, + 260 + ], + "type": "text", + "content": "[139] Fadime Sener, Dibyadip Chatterjee, Daniel Shelepov, Kun He, Dipika Singhania, Robert Wang, and Angela Yao. Assembly101: A large-scale multi-view video dataset for understanding procedural activities. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21096-21106, 2022. 2, 3, 8, 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 262, + 545, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 262, + 545, + 316 + ], + "spans": [ + { + "bbox": [ + 308, + 262, + 545, + 316 + ], + "type": "text", + "content": "[140] Pierre Sermanet, Corey Lynch, Jasmine Hsu, and Sergey Levine. Time-contrastive networks: Self-supervised learning from multi-view observation. In 2017 IEEE Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), pages 486–487. IEEE, 2017. 43" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 317, + 545, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 317, + 545, + 370 + ], + "spans": [ + { + "bbox": [ + 308, + 317, + 545, + 370 + ], + "type": "text", + "content": "[141] Pierre Sermanet, Corey Lynch, Yevgen Chebotar, Jasmine Hsu, Eric Jang, Stefan Schaal, and Sergey Levine. Time-contrastive networks: Self-supervised learning from video. Proceedings of International Conference in Robotics and Automation (ICRA), 2018. 4" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 372, + 545, + 425 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 372, + 545, + 425 + ], + "spans": [ + { + "bbox": [ + 308, + 372, + 545, + 425 + ], + "type": "text", + "content": "[142] Xi Shen, Alexei A Efros, Armand Joulin, and Mathieu Aubry. Learning co-segmentation by segment swapping for retrieval and discovery. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5082-5092, 2022. 35, 36" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 426, + 545, + 501 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 426, + 545, + 501 + ], + "spans": [ + { + "bbox": [ + 308, + 426, + 545, + 501 + ], + "type": "text", + "content": "[143] Nina Shvetsova, Brian Chen, Andrew Rouditchenko, Samuel Thomas, Brian Kingsbury, Rogerio S Feris, David Harwath, James Glass, and Hilde Kuehne. Everything at once-multi-modal fusion transformer for video retrieval. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20020-20029, 2022. 4" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 503, + 545, + 546 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 503, + 545, + 546 + ], + "spans": [ + { + "bbox": [ + 308, + 503, + 545, + 546 + ], + "type": "text", + "content": "[144] Gunnar A Sigurdsson, Abhinav Gupta, Cordelia Schmid, Ali Farhadi, and Karteek Alahari. Actor and observer: Joint modeling of first and third-person videos. In CVPR, 2018, 4, 7, 43" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 548, + 545, + 591 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 548, + 545, + 591 + ], + "spans": [ + { + "bbox": [ + 308, + 548, + 545, + 591 + ], + "type": "text", + "content": "[145] Gunnar A Sigurdsson, Abhinav Gupta, Cordelia Schmid, Ali Farhadi, and Karteek Alahari. Charades-ego: A large-scale dataset of paired third and first person videos. arXiv preprint arXiv:1804.09626, 2018. 2, 3, 7" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 593, + 545, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 593, + 545, + 645 + ], + "spans": [ + { + "bbox": [ + 308, + 593, + 545, + 645 + ], + "type": "text", + "content": "[146] Tomas Simon, Hanbyul Joo, Iain Matthews, and Yaser Sheikh. Hand keypoint detection in single images using multiview bootstrapping. In Proceedings of the IEEE conference on Computer Vision and Pattern Recognition, pages 1145-1153, 2017. 9" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 647, + 545, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 647, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 308, + 647, + 545, + 689 + ], + "type": "text", + "content": "[147] Krishna Kumar Singh, Kayvon Fatahalian, and Alexei A. Efros. Krishnamac: Using a longitudinal, single-person, egocentric dataset for scene understanding tasks. In WACV, 2016. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 691, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 691, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 691, + 545, + 713 + ], + "type": "text", + "content": "[148] Yale Song, Eugene Byrne, Tushar Nagarajan, Huiyu Wang, Miguel Martin, and Lorenzo Torresani. Ego4d goal-step:" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19397" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 72, + 72, + 286, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 72, + 72, + 286, + 95 + ], + "spans": [ + { + "bbox": [ + 72, + 72, + 286, + 95 + ], + "type": "text", + "content": "Toward hierarchical understanding of procedural activities. In NeurIPS, 2023. 7" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 96, + 287, + 129 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 96, + 287, + 129 + ], + "spans": [ + { + "bbox": [ + 48, + 96, + 287, + 129 + ], + "type": "text", + "content": "[149] Khurram Soomro, Amir Roshan Zamir, and Mubarak Shah. Ucf101: A dataset of 101 human action classes from videos in the wild. In CRCV-TR-12-01, 2012. 2, 42" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 131, + 287, + 163 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 131, + 287, + 163 + ], + "spans": [ + { + "bbox": [ + 49, + 131, + 287, + 163 + ], + "type": "text", + "content": "[150] Bilge Soran, Ali Farhadi, and Linda Shapiro. Generating notifications for missing actions: Don't forget to turn the lights off! In ICCV, pages 4669-4677, 2015. 8" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 165, + 287, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 165, + 287, + 274 + ], + "spans": [ + { + "bbox": [ + 49, + 165, + 287, + 274 + ], + "type": "text", + "content": "[151] Julian Straub, Thomas Whelan, Lingni Ma, Yufan Chen, Erik Wijmans, Simon Green, Jakob J. Engel, Raul Mur-Artal, Carl Ren, Shobhit Verma, Anton Clarkson, Mingfei Yan, Brian Budge, Yajie Yan, Xiaqing Pan, June Yon, Yuyang Zou, Kimberly Leon, Nigel Carter, Jesus Briales, Tyler Gillingham, Elias Mueggler, Luis Pesqueira, Manolis Savva, Dhruv Batra, Hauke M. Strasdat, Renzo De Nardi, Michael Goesele, Steven Lovegrove, and Richard Newcombe. The Replica dataset: A digital replica of indoor spaces. arXiv preprint arXiv:1906.05797, 2019. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 49, + 276, + 287, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 276, + 287, + 374 + ], + "spans": [ + { + "bbox": [ + 49, + 276, + 287, + 374 + ], + "type": "text", + "content": "[152] Carole H Sudre, Wenqi Li, Tom Vercauteren, Sebastien Ourselin, and M Jorge Cardoso. Generalised dice overlap as a deep learning loss function for highly unbalanced segmentations. In Deep Learning in Medical Image Analysis and Multimodal Learning for Clinical Decision Support: Third International Workshop, DLMIA 2017, and 7th International Workshop, ML-CDS 2017, Held in conjunction with MICCAI 2017, Quebec City, QC, Canada, September 14, Proceedings 3, pages 240-248. Springer, 2017. 39" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 49, + 376, + 287, + 419 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 376, + 287, + 419 + ], + "spans": [ + { + "bbox": [ + 49, + 376, + 287, + 419 + ], + "type": "text", + "content": "[153] Vivienne Sze, Yu-Hsin Chen, Tien-Ju Yang, and Joel S Emer. How to evaluate deep neural network processors: Tops/w (alone) considered harmful. IEEE Solid-State Circuits Magazine, 2020. 46" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 49, + 422, + 287, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 422, + 287, + 464 + ], + "spans": [ + { + "bbox": [ + 49, + 422, + 287, + 464 + ], + "type": "text", + "content": "[154] Mukund Varma T, Peihao Wang, Xuxi Chen, Tianlong Chen, Subhashini Venugopalan, and Zhangyang Wang. Is attention all that neRF needs? In The Eleventh International Conference on Learning Representations, 2023. 39" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 49, + 467, + 287, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 467, + 287, + 521 + ], + "spans": [ + { + "bbox": [ + 49, + 467, + 287, + 521 + ], + "type": "text", + "content": "[155] Mingxing Tan, Bo Chen, Ruoming Pang, Vijay Vasudevan, Mark Sandler, Andrew Howard, and Quoc V Le. Mnasnet: Platform-aware neural architecture search for mobile. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2820-2828, 2019. 8" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 49, + 523, + 287, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 523, + 287, + 555 + ], + "spans": [ + { + "bbox": [ + 49, + 523, + 287, + 555 + ], + "type": "text", + "content": "[156] Shuhan Tan, Tushar Nagarajan, and Kristen Grauman. Egodistill: Egocentric head motion distillation for efficient video understanding. NeurIPS, 2023. 8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 49, + 557, + 287, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 557, + 287, + 621 + ], + "spans": [ + { + "bbox": [ + 49, + 557, + 287, + 621 + ], + "type": "text", + "content": "[157] Hao Tang, Dan Xu, Nicu Sebe, Yanzhi Wang, Jason J Corso, and Yan Yan. Multi-channel attention selection gan with cascaded semantic guidance for cross-view image translation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2417-2426, 2019. 4, 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 49, + 624, + 287, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 624, + 287, + 667 + ], + "spans": [ + { + "bbox": [ + 49, + 624, + 287, + 667 + ], + "type": "text", + "content": "[158] Hao Tang, Kevin Liang, Kristen Grauman, Matt Feiszli, and Weiyao Wang. Egotracks: A long-term egocentric visual object tracking dataset. Advances in Neural Information Processing Systems, 2023. 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 49, + 670, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 670, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 49, + 670, + 287, + 712 + ], + "type": "text", + "content": "[159] Yansong Tang, Jiwen Lu, and Jie Zhou. Comprehensive instructional video analysis: The coin dataset and performance evaluation. IEEE transactions on pattern analysis and machine intelligence, 2020. 2, 4, 7" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 308, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 308, + 73, + 545, + 106 + ], + "type": "text", + "content": "[160] Zachary Teed and Jia Deng. DROID-SLAM: Deep Visual SLAM for Monocular, Stereo, and RGB-D Cameras. Advances in neural information processing systems, 2021. 57" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 308, + 108, + 545, + 151 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 108, + 545, + 151 + ], + "spans": [ + { + "bbox": [ + 308, + 108, + 545, + 151 + ], + "type": "text", + "content": "[161] Denis Tome, Patrick Peluse, Lourdes Agapito, and Hernan Badino. xr-egopose: Egocentric 3d human pose from an hmd camera. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2019. 9" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 308, + 152, + 545, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 152, + 545, + 185 + ], + "spans": [ + { + "bbox": [ + 308, + 152, + 545, + 185 + ], + "type": "text", + "content": "[162] Zhan Tong, Yibing Song, Jue Wang, and Limin Wang. Videomae: Masked autoencoders are data-efficient learners for self-supervised video pre-training. 2022. 43" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 186, + 545, + 241 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 186, + 545, + 241 + ], + "spans": [ + { + "bbox": [ + 308, + 186, + 545, + 241 + ], + "type": "text", + "content": "[163] Vadim Tschernezki, Ahmad Darkhalil, Zhifan Zhu, David Fouhey, Iro Larina, Diane Larlus, Dima Damen, and Andrea Vedaldi. EPIC Fields: Marrying 3D Geometry and Video Understanding. In Proceedings of the Neural Information Processing Systems (NeurIPS), 2023. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 243, + 545, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 243, + 545, + 285 + ], + "spans": [ + { + "bbox": [ + 308, + 243, + 545, + 285 + ], + "type": "text", + "content": "[164] Hung-Yu Tseng, Qinbo Li, Changil Kim, Suhib Alsisan, Jia-Bin Huang, and Johannes Kopf. Consistent view synthesis with pose-guided diffusion models. arXiv preprint arXiv:2303.17598, 2023. 4" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 287, + 545, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 287, + 545, + 342 + ], + "spans": [ + { + "bbox": [ + 308, + 287, + 545, + 342 + ], + "type": "text", + "content": "[165] Pavan Kumar Anasosalu Vasu, James Gabriel, Jeff Zhu, Oncel Tuzel, and Anurag Ranjan. Mobileone: An improved one millisecond mobile backbone. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7907-7917, 2023. 8" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 344, + 545, + 375 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 344, + 545, + 375 + ], + "spans": [ + { + "bbox": [ + 308, + 344, + 545, + 375 + ], + "type": "text", + "content": "[166] Sara Vicente, Carsten Rother, and Vladimir Kolmogorov. Object cosegmentation. In CVPR 2011, pages 2217-2224. IEEE, 2011. 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 377, + 545, + 410 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 377, + 545, + 410 + ], + "spans": [ + { + "bbox": [ + 308, + 377, + 545, + 410 + ], + "type": "text", + "content": "[167] Hanlin Wang, Yilu Wu, Sheng Guo, and Limin Wang. Pdpp: Projected diffusion for procedure planning in instructional videos. arXiv preprint arXiv:2303.14676, 2023. 4" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 412, + 545, + 455 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 412, + 545, + 455 + ], + "spans": [ + { + "bbox": [ + 308, + 412, + 545, + 455 + ], + "type": "text", + "content": "[168] Daniel Watson, William Chan, Ricardo Martin-Brualla, Jonathan Ho, Andrea Tagliasacchi, and Mohammad Norouzi. Novel view synthesis with diffusion models. arXiv preprint arXiv:2210.04628, 2022. 4" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 456, + 545, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 456, + 545, + 499 + ], + "spans": [ + { + "bbox": [ + 308, + 456, + 545, + 499 + ], + "type": "text", + "content": "[169] Daniel Weinland, Remi Ronfard, and Edmond Boyer. Free viewpoint action recognition using motion history volumes. Computer Vision and Image Understanding (CVIU), 2006. 3, 7" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 502, + 545, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 502, + 545, + 567 + ], + "spans": [ + { + "bbox": [ + 308, + 502, + 545, + 567 + ], + "type": "text", + "content": "[170] Yangming Wen, Krishna Kumar Singh, Markham Anderson, Wei-Pang Jan, and Yong Jae Lee. Seeing the unseen: Predicting the first-person camera wearer's location and pose in third-person scenes. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) Workshops, pages 3446-3455, 2021. 4" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 568, + 545, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 568, + 545, + 622 + ], + "spans": [ + { + "bbox": [ + 308, + 568, + 545, + 622 + ], + "type": "text", + "content": "[171] Olivia Wiles, Georgia Gkioxari, Richard Szeliski, and Justin Johnson. Synsin: End-to-end view synthesis from a single image. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7467-7477, 2020. 4" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 624, + 545, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 624, + 545, + 678 + ], + "spans": [ + { + "bbox": [ + 308, + 624, + 545, + 678 + ], + "type": "text", + "content": "[172] Benita Wong, Joya Chen, You Wu, Stan Weixian Lei, Dongxing Mao, Difei Gao, and Mike Zheng Shou. Assistq: Affordance-centric question-driven task completion for egocentric assistant. In European Conference on Computer Vision, 2022. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 681, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 681, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 681, + 545, + 713 + ], + "type": "text", + "content": "[173] Zhirong Wu, Shuran Song, Aditya Khosla, Fisher Yu, Linguang Zhang, Xiaou Tang, and Jianxiong Xiao. 3d shapenets: A deep representation for volumetric shapes. In" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19398" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 72, + 72, + 287, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 72, + 72, + 287, + 94 + ], + "spans": [ + { + "bbox": [ + 72, + 72, + 287, + 94 + ], + "type": "text", + "content": "Computer Vision and Pattern Recognition, IEEE Conference on, 2015. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 95, + 287, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 95, + 287, + 138 + ], + "spans": [ + { + "bbox": [ + 48, + 95, + 287, + 138 + ], + "type": "text", + "content": "[174] Zuxuan Wu, Tushar Nagarajan, Abhishek Kumar, Steven Rennie, Larry S Davis, Kristen Grauman, and Rogerio Feris. Blockdrop: Dynamic inference paths in residual networks. In CVPR, 2018. 8" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 140, + 287, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 140, + 287, + 194 + ], + "spans": [ + { + "bbox": [ + 48, + 140, + 287, + 194 + ], + "type": "text", + "content": "[175] Fei Xia, Amir R. Zamir, Zhi-Yang He, Alexander Sax, Jitendra Malik, and Silvio Savarese. Gibson Env: real-world perception for embodied agents. In CVPR. IEEE, 2018. Gibson license is available at http://svl.stanford.edu/gibson2/assets/GDS_agreement.pdf.3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 195, + 287, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 195, + 287, + 226 + ], + "spans": [ + { + "bbox": [ + 48, + 195, + 287, + 226 + ], + "type": "text", + "content": "[176] J. Xiao, A. Owens, and A. Torralba. Sun3d: A database of big spaces reconstructed using sfm and object labels. In ICCV, 2013. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 228, + 287, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 228, + 287, + 271 + ], + "spans": [ + { + "bbox": [ + 48, + 228, + 287, + 271 + ], + "type": "text", + "content": "[177] Frank F Xu, Lei Ji, Botian Shi, Junyi Du, Graham Neubig, Yonatan Bisk, and Nan Duan. A benchmark for structured procedural knowledge extraction from cooking videos. arXiv preprint arXiv:2005.00706, 2020. 8" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 272, + 287, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 272, + 287, + 326 + ], + "spans": [ + { + "bbox": [ + 48, + 272, + 287, + 326 + ], + "type": "text", + "content": "[178] Hu Xu, Gargi Ghosh, Po-Yao Huang, Dmytro Okhonko, Armen Aghajanyan, Florian Metze, Luke Zettlemoyer, and Christoph Feichtenhofer. Videoclip: Contrastive pretraining for zero-shot video-text understanding. arXiv preprint arXiv:2109.14084, 2021. 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 327, + 287, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 327, + 287, + 369 + ], + "spans": [ + { + "bbox": [ + 48, + 327, + 287, + 369 + ], + "type": "text", + "content": "[179] Mingze Xu, Chenyou Fan, Yuchen Wang, Michael S Ryoo, and David J Crandall. Joint person segmentation and identification in synchronized first-and third-person videos. In ECCV, 2018. 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 371, + 287, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 371, + 287, + 415 + ], + "spans": [ + { + "bbox": [ + 48, + 371, + 287, + 415 + ], + "type": "text", + "content": "[180] Mingze Xu, Yuanjun Xiong, Hao Chen, Xinyu Li, Wei Xia, Zhuowen Tu, and Stefano Soatto. Long short-term transformer for online action detection. Advances in Neural Information Processing Systems, 34:1086-1099, 2021. 49" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 415, + 287, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 415, + 287, + 480 + ], + "spans": [ + { + "bbox": [ + 48, + 415, + 287, + 480 + ], + "type": "text", + "content": "[181] Weipeng Xu, Avishek Chatterjee, Michael Zollhoefer, Helge Rhodin, Pascal Fua, Hans-Peter Seidel, and Christian Theobalt. Mo 2 cap 2: Real-time mobile 3d motion capture with a cap-mounted fisheye camera. IEEE transactions on visualization and computer graphics, 25(5):2093-2101, 2019. 9" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 481, + 287, + 523 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 481, + 287, + 523 + ], + "spans": [ + { + "bbox": [ + 48, + 481, + 287, + 523 + ], + "type": "text", + "content": "[182] Zihui Xue and Kristen Grauman. Learning fine-grained view-invariant representations from unpaired ego-exo videos via temporal alignment. In NeurIPS, 2023. 4, 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 525, + 287, + 579 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 525, + 287, + 579 + ], + "spans": [ + { + "bbox": [ + 48, + 525, + 287, + 579 + ], + "type": "text", + "content": "[183] Lita Yang, Robert M Radway, Yu-Hsin Chen, Tony F Wu, Huichu Liu, Elnaz Ansari, Vikas Chandra, Subhasish Mitra, and Edith Beigné. Three-dimensional stacked neural network accelerator architectures for ar/vr applications. IEEE Micro, 2022. 46" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 581, + 287, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 581, + 287, + 613 + ], + "spans": [ + { + "bbox": [ + 48, + 581, + 287, + 613 + ], + "type": "text", + "content": "[184] Huangyue Yu, Minjie Cai, Yunfei Liu, and Feng Lu. What i see is what you see: Joint attention learning for first and third person video co-analysis. In ACM MM, 2019. 4" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 614, + 287, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 614, + 287, + 658 + ], + "spans": [ + { + "bbox": [ + 48, + 614, + 287, + 658 + ], + "type": "text", + "content": "[185] Huangyue Yu, Minjie Cai, Yunfei Liu, and Feng Lu. First- and third-person video co-analysis by learning spatial-temporal joint attention. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2020. 4" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 658, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 287, + 712 + ], + "type": "text", + "content": "[186] Xumin Yu, Yongming Rao, Wenliang Zhao, Jiwen Lu, and Jie Zhou. Group-aware contrastive regression for action quality assessment. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 7919-7928, 2021. 8" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "text", + "content": "[187] Ye Yuan and Kris Kitani. 3d ego-pose estimation via imitation learning. In Proceedings of the European Conference on Computer Vision (ECCV), 2018. 9" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 107, + 545, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 107, + 545, + 150 + ], + "spans": [ + { + "bbox": [ + 307, + 107, + 545, + 150 + ], + "type": "text", + "content": "[188] Ye Yuan and Kris Kitani. Ego-pose estimation and forecasting as real-time pd control. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2019. 9" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 152, + 545, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 152, + 545, + 185 + ], + "spans": [ + { + "bbox": [ + 307, + 152, + 545, + 185 + ], + "type": "text", + "content": "[189] Chenlin Zhang, Jianxin Wu, and Yin Li. Actionformer: Localizing moments of actions with transformers. arXiv preprint arXiv:2202.07925, 2022. 21" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 186, + 545, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 186, + 545, + 228 + ], + "spans": [ + { + "bbox": [ + 307, + 186, + 545, + 228 + ], + "type": "text", + "content": "[190] Chen-Lin Zhang, Jianxin Wu, and Yin Li. Actionformer: Localizing moments of actions with transformers. In European Conference on Computer Vision, pages 492-510, 2022. 54, 55, 57" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 232, + 545, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 232, + 545, + 274 + ], + "spans": [ + { + "bbox": [ + 307, + 232, + 545, + 274 + ], + "type": "text", + "content": "[191] Qiang Zhang and Baoxin Li. Relative hidden markov models for evaluating motion skill. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2013. 8" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 276, + 545, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 276, + 545, + 308 + ], + "spans": [ + { + "bbox": [ + 307, + 276, + 545, + 308 + ], + "type": "text", + "content": "[192] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, 2018. 39" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 310, + 545, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 310, + 545, + 354 + ], + "spans": [ + { + "bbox": [ + 307, + 310, + 545, + 354 + ], + "type": "text", + "content": "[193] Siwei Zhang, Qianli Ma, Yan Zhang, Zhiyin Qian, Taein Kwon, Marc Pollefeys, Federica Bogo, and Siyu Tang. Egobody: Human body shape and motion of interacting people from head-mounted devices. In ECCV, 2022. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 355, + 545, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 355, + 545, + 398 + ], + "spans": [ + { + "bbox": [ + 307, + 355, + 545, + 398 + ], + "type": "text", + "content": "[194] Shiyi Zhang, Wenxun Dai, Sujia Wang, Xiangwei Shen, Jiwen Lu, Jie Zhou, and Yansong Tang. Logo: A long-form video dataset for group action quality assessment. In CVPR, 2023. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 400, + 545, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 400, + 545, + 453 + ], + "spans": [ + { + "bbox": [ + 307, + 400, + 545, + 453 + ], + "type": "text", + "content": "[195] Xiangyu Zhang, Xinyu Zhou, Mengxiao Lin, and Jian Sun. Shufflenet: An extremely efficient convolutional neural network for mobile devices. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6848-6856, 2018. 8" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 456, + 545, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 456, + 545, + 521 + ], + "spans": [ + { + "bbox": [ + 307, + 456, + 545, + 521 + ], + "type": "text", + "content": "[196] He Zhao, Isma Hadji, Nikita Dvornik, Konstantinos G Derpanis, Richard P Wildes, and Allan D Jepson. P3iv: Probabilistic procedure planning from instructional videos with weak supervision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2938-2948, 2022. 4" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 523, + 545, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 523, + 545, + 567 + ], + "spans": [ + { + "bbox": [ + 307, + 523, + 545, + 567 + ], + "type": "text", + "content": "[197] Weixi Zhao, Weiqiang Wang, and Yunjie Tian. Graformer: Graph-oriented transformer for 3d pose estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20438-20447, 2022. 58" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 568, + 545, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 568, + 545, + 601 + ], + "spans": [ + { + "bbox": [ + 307, + 568, + 545, + 601 + ], + "type": "text", + "content": "[198] Yue Zhao and Philipp Krahenbuhl. Real-time online video detection with temporal smoothing transformers. In European Conference on Computer Vision, 2022. 49" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 602, + 545, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 602, + 545, + 656 + ], + "spans": [ + { + "bbox": [ + 307, + 602, + 545, + 656 + ], + "type": "text", + "content": "[199] Yue Zhao, Ishan Misra, Philipp Krahenbuhl, and Rohit Girdhar. Learning video representations from large language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6586-6597, 2023. 26, 46, 47, 48" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 658, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 658, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 307, + 658, + 545, + 712 + ], + "type": "text", + "content": "[200] Ce Zheng, Xianpeng Liu, Guo-Jun Qi, and Chen Chen. Potter: Pooling attention transformer for efficient human mesh recovery. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1611-1620, 2023. 59" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "19399" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 418 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "text", + "content": "[201] Yiwu Zhong, Licheng Yu, Yang Bai, Shangwen Li, Xueling Yan, and Yin Li. Learning procedure-aware video representation from instructional videos and their narrations. arXiv preprint arXiv:2303.17839, 2023. 4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 118, + 287, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 118, + 287, + 173 + ], + "spans": [ + { + "bbox": [ + 49, + 118, + 287, + 173 + ], + "type": "text", + "content": "[202] Honglu Zhou, Roberto Martin-Martin, Mubbasir Kapadia, Silvio Savarese, and Juan Carlos Niebles. Procedure-aware pretraining for instructional video understanding. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023. 4, 8" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 174, + 287, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 174, + 287, + 217 + ], + "spans": [ + { + "bbox": [ + 49, + 174, + 287, + 217 + ], + "type": "text", + "content": "[203] Honglu Zhou, Roberto Martin-Martin, Mubbasir Kapadia, Silvio Savarese, and Juan Carlos Niebles. Procedure-aware pretraining for instructional video understanding. In CVPR, pages 10727-10738, 2023. 49" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 219, + 287, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 219, + 287, + 251 + ], + "spans": [ + { + "bbox": [ + 49, + 219, + 287, + 251 + ], + "type": "text", + "content": "[204] L. Zhou, N. Louis, and J. Corso. Weakly-supervised video object grounding from text by loss weighting and object interaction. In BMVC, 2018. 2, 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 49, + 252, + 287, + 284 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 252, + 287, + 284 + ], + "spans": [ + { + "bbox": [ + 49, + 252, + 287, + 284 + ], + "type": "text", + "content": "[205] Luowei Zhou, Chenliang Xu, and Jason J Corso. Towards automatic learning of procedures from web instructional videos. In AAAI, 2018. 4, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 49, + 286, + 286, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 286, + 286, + 319 + ], + "spans": [ + { + "bbox": [ + 49, + 286, + 286, + 319 + ], + "type": "text", + "content": "[206] Michael Zhu and Suyog Gupta. To prune, or not to prune: exploring the efficacy of pruning for model compression. arXiv preprint arXiv:1710.01878, 2017. 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 49, + 320, + 287, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 320, + 287, + 374 + ], + "spans": [ + { + "bbox": [ + 49, + 320, + 287, + 374 + ], + "type": "text", + "content": "[207] Dimitri Zhukov, Jean-Baptiste Alayrac, Ramazan Gokberk Cinbis, David Fouhey, Ivan Laptev, and Josef Sivic. Cross-task weakly supervised learning from instructional videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2019. 2, 4, 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 49, + 376, + 287, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 376, + 287, + 418 + ], + "spans": [ + { + "bbox": [ + 49, + 376, + 287, + 418 + ], + "type": "text", + "content": "[208] Aneeq Zia, Yachna Sharma, Vinay Bettadapura, Eric L. Sarin, and Irfan A. Essa. Video and accelerometer-based motion analysis for automated surgical skills assessment. CoRR, abs/1702.07772, 2017. 8" + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19400" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/EgoExoLearn_ A Dataset for Bridging Asynchronous Ego- and Exo-centric View of Procedural Activities in Real World/134e2089-370f-4ad6-9686-84a1d7af0ed4_content_list.json b/2024/EgoExoLearn_ A Dataset for Bridging Asynchronous Ego- and Exo-centric View of Procedural Activities in Real World/134e2089-370f-4ad6-9686-84a1d7af0ed4_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..755ff603238adbbfdb2f87eea4b1354b6a0990ea --- /dev/null +++ b/2024/EgoExoLearn_ A Dataset for Bridging Asynchronous Ego- and Exo-centric View of Procedural Activities in Real World/134e2089-370f-4ad6-9686-84a1d7af0ed4_content_list.json @@ -0,0 +1,1651 @@ +[ + { + "type": "text", + "text": "EgoExoLearn: A Dataset for Bridging Asynchronous Ego- and Exo-centric View of Procedural Activities in Real World", + "text_level": 1, + "bbox": [ + 153, + 130, + 816, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yifei Huang†‡, Guo Chen‡, Jilan Xu‡, Mingfang Zhang‡, Lijin Yang, Baoqi Pei Hongjie Zhang, Lu Dong, Yali Wang*, Limin Wang*, Yu Qiao* OpenGVLab, Shanghai AI Laboratory Shenzhen Institutes of Advanced Technology, CAS Nanjing University", + "bbox": [ + 174, + 202, + 795, + 276 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 309, + 313, + 325 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Being able to map the activities of others into one's own point of view is a fundamental human skill even from a very early age. Taking a step toward understanding this human ability, we introduce EgoExoLearn, a large-scale dataset that emulates the human demonstration following process, in which individuals record egocentric videos as they execute tasks guided by exocentric-view demonstration videos. Focusing on the potential applications in daily assistance and professional support, EgoExoLearn contains egocentric and demonstration video data spanning 120 hours captured in daily life scenarios and specialized laboratories. Along with the videos we record high-quality gaze data and provide detailed multimodal annotations, formulating a playground for modeling the human ability to bridge asynchronous procedural actions from different viewpoints. To this end, we present benchmarks such as cross-view association, cross-view action planning, and cross-view referenced skill assessment, along with detailed analysis. We expect EgoExoLearn can serve as an important resource for bridging the actions across views, thus paving the way for creating AI agents capable of seamlessly learning by observing humans in the real world. The dataset and benchmark codes are available at https://github.com/OpenGVLab/EgoExoLearn.", + "bbox": [ + 75, + 329, + 473, + 691 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 710, + 209, + 726 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Even as a child, humans can observe the actions of others and then map them to their own view [4, 34, 100, 105]. With this ability to asynchronously bridge activities from egocentric and exocentric views [95, 103], humans can watch others' demonstrations and replicate the procedures in a new environment. This ability is especially beneficial when actual physical trials carry the potential of high costs [24], e.g., conducting dangerous chemical experiments.", + "bbox": [ + 75, + 736, + 470, + 857 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In the wake of recent advancements in AI systems, one", + "bbox": [ + 94, + 859, + 468, + 875 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "goal for the next generation of AI agents is to perform tasks in a more embodied setting [94]. However, different from humans, training these AI agents usually requires demonstration videos taken in a similar environment [75, 119] and from a congruent perspective with the AI agents, (e.g., the egocentric point of view [43, 59, 108, 134]). While great effort has been made into the collection of egocentric data in different scenarios [18, 30, 106], it remains crucial for the AI agents to directly learn from demonstration videos taken in a different place and from a different viewpoint [35, 148]. Realizing this capability can unleash the full potential of public instructional video data [82] and is also useful in the human-robot cooperation scenario, especially in novel environments [57, 77, 128].", + "bbox": [ + 496, + 311, + 893, + 521 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Current works towards this goal can be roughly divided into two directions. One way is to learn models in simulated environments [11, 64, 79, 89, 92], but it remains difficult for models in this setting to generalize in the real world [126]. The other direction is to learn from human activity in real-world scenarios. However, attempts to directly combine existing multiview datasets often yield datasets of inferior quality or scale [124, 143]. Meanwhile, the few existing datasets in this direction [99, 106, 109] only record ego- and exo-view videos in the same environment and in a time-synchronized manner. In reality, when following demonstrations it is often needed to bridge a series of procedural actions performed in a different place and at a different time. However, currently, no dataset is available for the exploration of how to bridge asynchronous procedural activities in realistic egocentric and exocentric viewpoints.", + "bbox": [ + 496, + 522, + 895, + 763 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To address this lack of dataset issue, we introduce EgoExoLearn, a large-scale dataset containing demonstration videos and corresponding egocentric videos where the camera wearers follow the demonstrations and perform the same task in a different environment, as shown in Figure 1. Targeting two potential applications, i.e., daily assistance and professional support, EgoExoLearn consists of 747 video sequences spanning a total of 120 hours of footage, ranging from daily food-making to specialized laboratory experi", + "bbox": [ + 496, + 763, + 895, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding authors. †Project lead. ‡Equal key contributions.", + "bbox": [ + 76, + 886, + 426, + 901 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "22072", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/31594f139b11ce421ea3883f07660feaa1a57473de5d0f9f5fd264319f7ffe46.jpg", + "image_caption": [ + "Figure 1. EgoExoLearn emulates the human asynchronous demonstration following process. It contains demonstration videos of multiple tasks, together with egocentric videos recorded by participants replicating the procedure after watching the demonstrations. The dataset comprises gaze signals and fine-grained multi-level multi-modal annotations, enabling the exploration of key features in this context such as cross-view association and cross-view action planning." + ], + "image_footnote": [], + "bbox": [ + 83, + 73, + 890, + 277 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "ments. Notably, the egocentric videos in EgoExoLearn contain eye gaze signals showing humans' visual attention while performing the task. This provides a valuable cue for better bridging the actions in ego- and exo-viewpoints.", + "bbox": [ + 75, + 359, + 468, + 419 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We take one more step forward by analyzing human ability in bridging asynchronous ego- and exo-view actions and, accordingly, introduce new tasks and benchmarks that we believe can form building blocks for the development of next-stage embodied AI agents with similar abilities. When humans perform an action, he/she can associate and describe the undergoing action in the egocentric view with the corresponding action in the demonstration. With the knowledge from demonstration videos, humans can know the needed action steps and predict what the next steps should be. Besides, through the comparison with the demonstration, humans can also assess their level of skills.", + "bbox": [ + 75, + 420, + 473, + 599 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Based on the above analysis, we design benchmarks of 1) cross-view association, 2) cross-view action understanding, 3) cross-view referenced skill assessment, and 4) cross-view referenced video captioning. Each benchmark is meticulously defined, annotated, and supported by baseline implementations. In addition, we pioneeringly explore the role of gaze in these tasks. We hope our dataset can provide resources for future work for bridging asynchronous procedural actions in ego- and exo-centric perspectives, thereby inspiring the design of AI agents adept at learning from real-world human demonstrations and mapping the procedural actions into robot-centric views.", + "bbox": [ + 75, + 601, + 470, + 781 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 76, + 796, + 218, + 811 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Ego-exo datasets. While there exist works that associate existing datasets to explore how activities can be bridged between them, these associated datasets are often limited in scale [14, 144, 147] or quality [124], meanwhile focusing only on single actions captured from the same view [81, 88,", + "bbox": [ + 75, + 825, + 472, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "131]. As for actions from different views, apart from multiview fixed camera datasets [6, 15, 54, 61], there also exist datasets with both ego- exo-centric view videos [30, 31, 99, 106, 109]. These datasets are either recorded in the same environment [45, 99, 109] or record time-synced multi-view videos in the same environment with primary focuses on pose/activity understanding grounded in the 3D world [55, 106, 154]. Our dataset offers a more challenging and realistic scenario, where egocentric camera wearers learn to complete the tasks demonstrated by exocentric demonstration videos. This setting complements these datasets by focusing on high-level procedural actions.", + "bbox": [ + 496, + 359, + 895, + 541 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The only dataset conceptually similar to ours is the recently proposed AE2 dataset [143], where the goal is to learn view-invariant representation from unpaired ego and exo videos. This dataset combines ego and exo videos from five public datasets [18, 20, 53, 55, 155] and a newly collected ego tennis forehand dataset. However, due to the difficulty in associating existing ego-exo datasets, the AE2 dataset is relatively small where the largest subset contains only 322 clips. Also, this dataset only focuses on clip-level actions, and thus cannot feature the real-world demonstration following setting, which usually requires multimodal, task-centric procedural knowledge. Instead, our EgoExoLearn is much larger in scale (100x more clips), while offering gaze and fine-grained multimodal annotations facilitating multi-faceted analysis of ego-exo action understanding.", + "bbox": [ + 496, + 545, + 895, + 772 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Egocentric video datasets. In line with the recent development in wearable cameras [110], multiple egocentric video datasets [5, 17, 19, 30, 46, 78, 98, 112, 150] have been proposed. Different from previous egocentric datasets, the egocentric videos in EgoExoLearn feature a demonstration-following setting. We believe EgoExoLearn provides a playground for developing tools to bridge asynchronous procedural activities from ego- and exocentric viewpoints.", + "bbox": [ + 496, + 780, + 895, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "22073", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The setting of our EgoExoLearn complements existing datasets like Ego4D and can benefit from their rich knowledge and representations.", + "bbox": [ + 75, + 90, + 470, + 136 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Egocentric gaze. Gaze can indicate visual attention and contains valuable information about human intent [36, 156], thus is used in a diverse range of areas such as human-computer interaction [42, 51, 151], and augmented reality [90, 102]. In computer vision, efforts have been made to leverage gaze in various tasks [37-39, 56, 65-67, 83, 101, 139]. However, with the previous absence of large-scale egocentric datasets that include gaze, this avenue of research is currently under-explored [44, 68, 94, 154, 160]. Our EgoExoLearn offers calibrated gaze positions for all egocentric videos. Thanks to our unique setting, our dataset enables the integration of gaze in egocentric video understanding and the exploration of the role of gaze in the cross-view context.", + "bbox": [ + 75, + 141, + 470, + 335 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Egocentric and ego-exo video understanding. The unique recording perspective of egocentric videos presents a series of challenges including but not limited to action understanding [13, 26, 27, 40, 50, 93, 96, 113, 120, 121, 125], hand detection [12, 29, 107, 153], and video-language understanding [41, 48]. These form fundamental building block techniques of embodied AI [85], VR/AR [9, 47, 74, 123], and human-robot interaction [63, 80, 86, 137]. Since most egocentric datasets are smaller in scale compared with general datasets which contain mostly exocentric view videos [3, 142, 152], it is possible to leverage exocentric video data to improve model performance on egocentric videos [132]. There are typically three main directions: joint view-invariant learning [124, 140, 142], domain adaptation [136], and knowledge distillation [69]. In this work, we evaluate all these directions in our benchmarks.", + "bbox": [ + 75, + 342, + 472, + 583 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Dataset", + "text_level": 1, + "bbox": [ + 76, + 599, + 163, + 614 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Data Collection", + "text_level": 1, + "bbox": [ + 76, + 626, + 232, + 640 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Scenarios and tasks. We consider procedural goal-oriented tasks ranging from daily food-making to specialized laboratory-based experiments. This selection is grounded in their exemplification of two prospective areas where future embodied AI agents would need the ability to bridge ego-exo activities: daily-life assistance and professional support. Specifically, EgoExoLearn incorporates 5 types of daily tasks (e.g., cooking) and 3 types of specialized laboratory tasks (e.g., solid-phase peptide synthesis). We record egocentric videos in 4 different kitchens and 3 different labs. Other details are provided in the supplementary.", + "bbox": [ + 75, + 654, + 470, + 820 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Data collection procedure. Before the start of each collection session, participants are required to complete a questionnaire gathering basic demographic information and their self-evaluated expertise in executing the designated task. This questionnaire also highlights the ethical, privacy, and secu", + "bbox": [ + 75, + 825, + 470, + 901 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/41bf1cd76182fcc4e543357f9fff5b20ff68656802bdde7eb624f84544efa2ad.jpg", + "image_caption": [ + "Figure 2. The number of videos per task (left) and the average duration of each video per task (right). Task1 to Task5 represent the 5 daily tasks and the remaining are three tasks in specialized laboratories. In each recording session of the egocentric video, one participant may learn from multiple demonstration videos and one demonstration video may be watched by several participants." + ], + "image_footnote": [], + "bbox": [ + 500, + 90, + 700, + 268 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/6e81b914ecc921aa6bfb1c83be55963cf5f84e4eaf459922900c1d7c7a7f5cfe.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 707, + 90, + 890, + 265 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "rity considerations. Then in each session, participants will be asked to choose one or several exocentric view demonstration videos from a provided list and carefully learn the detailed procedures. Once they feel ready, they will wear Pupil Invisible Glasses [49], complete the gaze calibration, and begin to replicate the task performed in the demonstration videos. While not encouraged, participants are permitted to revisit the demonstration video during the recording.", + "bbox": [ + 496, + 376, + 893, + 497 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "After each recording session, the participants are asked to re-do the gaze calibration to ensure gaze data fidelity. For the 5 daily tasks, the exocentric demonstration videos are manually curated from online video platforms such as YouTube. For the lab experiments, the exocentric demonstration videos are tutorials recorded by senior lab members.", + "bbox": [ + 496, + 498, + 893, + 587 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Figure 2 shows the distribution of the 120 hours of data. Since most demonstration videos are meticulously edited to remove repeated steps, the average length of demonstration videos is lower than the egocentric videos which record the full procedure. As a result, the EgoExoLearn contains 432 egocentric videos totaling 96.5 hours and 315 demonstration videos spanning 23.5 hours. This difference in video length poses a unique challenge when bridging ego- and exo-view activities for future research endeavors.", + "bbox": [ + 496, + 589, + 893, + 724 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Annotation", + "text_level": 1, + "bbox": [ + 500, + 736, + 624, + 750 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To facilitate our dataset in the development of algorithms that can effectively bridge the gap between ego and exo viewpoints, we provide detailed multi-modal human annotations. Our pipeline of annotation contains four stages detailed in the following paragraphs. Each of these stages is subject to a rigorous manual quality check involving no fewer than two individuals for verification and validation.", + "bbox": [ + 496, + 760, + 893, + 864 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Coarse-level language annotation. In this step, we ask annotators to annotate the coarse actions in the videos. Like", + "bbox": [ + 500, + 869, + 890, + 898 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "22074", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/6a91f604f0dc85ff41960cbd533e77603029baf2ce792b1ff2d4376c7596dd7b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 76, + 87, + 483, + 213 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/1d69c8321a354769981259b26af409b5416cbfbe34be289b772a54d9988d2775.jpg", + "image_caption": [ + "Number of nouns (left axis) and noun durations (right axis) of right hand", + "Figure 3. Occurrence and duration distribution of the annotated fine-level verbs and nouns associated with the right hand." + ], + "image_footnote": [], + "bbox": [ + 488, + 97, + 892, + 215 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "the previous works [106, 126], the coarse actions are defined as a middle-level step for accomplishing a task and can be divided into multiple fine actions. For instance, \"Prepare the pork\" in the task of twice-cooked pork, and \"Suction filtration\" in solid-phase peptide synthesis. Three types of annotation are given in this step: 1) the temporal interval, consisting of start and end timestamps; 2) the action label; and 3) a language description of the video within the annotated interval. We specifically request annotators to focus on elucidating \"what is done,\" \"how it is done,\" and \"the purpose of this step\" in their language descriptions. We define a total of 39 categories of coarse-level actions in this stage and acquire 41.2 coarse-level annotations per video with an average length of 21.5 seconds.", + "bbox": [ + 75, + 246, + 473, + 458 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Fine-level language annotation. Based on the coarse-level annotations, in this step we request the annotators to provide annotations for the fine-level actions. The fine-level actions are the atomic actions like \"take knife\" or \"pull syringe plunger\". Unlike the first step, annotators are instructed to furnish language descriptions that specifically emphasize \"which hand is used\", \"what object is used\" and \"why it is used\". For the first two steps, we employ a two-round manual annotation checking to ensure the annotation quality.", + "bbox": [ + 75, + 467, + 470, + 604 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Translation & parsing. To ensure linguistic precision, all the annotators give the language description using their native language [18]. For non-English annotations, we employ ChatGPT and Google Translation API to translate them into English. Subsequently, for the fine-level annotations, we employ specific rules and utilize tools such as NLTK [10] and Spacy [116] to extract the verbs and nouns associated with each segment. During this stage, we confine the selection of verbs and nouns to the predefined taxonomy offered by Ego4D [30], while also manually introducing supplementary verbs and nouns that are absent in the taxonomy. Since our manual annotation specifies the engagement of the left/right hand, we can extract multiple verbs and nouns for each segment, meanwhile attributing them to the respective involvement of the left or right hand. After manual checking, we obtain a total of 95 verb and 254 noun categories in the fine-level annotation. In Figure 3 we show the occurrence of the top 30 categories of verbs and nouns attributed to the right hand. More statistics can be found in the supplementary.", + "bbox": [ + 75, + 613, + 473, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Skill level annotation. Since self-assessed skill level is not perfectly suitable for skill assessment, we identify several representative skills and assign human annotators to assess their skills. The annotation follows a pairwise ranking scheme, where annotators are presented with pairs of videos of the same action, and instructed to determine which video demonstrates a higher skill level. We prepare 40,191 video pairs and ensure that 4 different annotators annotate each pair. After filtering out pairs with less than 3 consistent opinions, we get a collection of 34,239 valid video pairs.", + "bbox": [ + 496, + 265, + 895, + 417 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Statistics & Comparisons", + "text_level": 1, + "bbox": [ + 500, + 426, + 732, + 443 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To the best of our knowledge, there is no dataset that follows the same setting as ours for a direct comparison. Therefore, we enumerate various aspects of our dataset and conduct a comparative analysis with relevant datasets in Tables 1 and 2. EgoExoLearn distinctively enriches the domain with its \"visual demonstration following\" setting. Beyond this unique setting, it stands as the first egocentric dataset that includes temporal bounded language captions, annotated cross-view associations, and multi-label video segments.", + "bbox": [ + 496, + 450, + 893, + 588 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Dataset Properties & Benchmarks", + "text_level": 1, + "bbox": [ + 498, + 602, + 812, + 619 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1. Dataset Properties", + "text_level": 1, + "bbox": [ + 500, + 627, + 679, + 643 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "EgoExoLearn stands out from current egocentric and ego-exo datasets due to several unique properties.", + "bbox": [ + 496, + 651, + 893, + 681 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Ego-Exo demonstration following setting. The most distinguished property of EgoExoLearn is the ego-exo demonstration following context. Egocentric video recorders are instructed to follow the steps in exocentric demonstration videos to perform the same task but in a different environment. This setting closely emulates the human observational learning process [4, 34] and can be instrumental in designing embodied AI agents that learn from alternative perspectives while executing tasks from their own viewpoint.", + "bbox": [ + 496, + 685, + 895, + 821 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Fine-grained vision-language annotations with gaze. To facilitate a deeper analysis, we equip our dataset with rich, multimodal, fine-grained annotations. EgoExoLearn is the first egocentric dataset featuring high-quality captions, evidenced by the number of words per segment in Tab. 2.", + "bbox": [ + 496, + 825, + 895, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "22075", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/bfdcfc0a8d661c4c31503579290085220379affceab77774f9ee9a5fe924216a.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
DatasetSettingsUnique HoursEgo +Exo?Instruction following?Visual instruction?GazeCoarse ActionFine ActionDense CaptionAssociationSkill
Meccano [98]Industry7X✓XXX✓XXX
EGTEA [67]Cooking28X✓X✓X✓XXX
EK-100 [18]Cooking100XXXXX✓XXX
HoloAssist [126]Assistive166X✓X✓✓✓XX✓
Ego4D [30]Multiple3670XXX◇◇◇◇XX
H2O [55]Desk1✓XXXX✓XXX
LEMMA [45]Daily10✓XXXX✓XXX
HOMEAGE [99]Daily25✓XXX✓✓XXX
CharadesEgo [109]Daily34✓XXXX✓XXX
Assembly101 [55]Desk42✓✓XX✓✓XXX
EgoExoLearn (ours)Daily & Lab120✓✓✓✓✓✓✓✓✓
", + "bbox": [ + 81, + 88, + 890, + 250 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/783d0cbc9cd7bb9df1f1d23d9ecca658354c1d6c13f38b66722d622af3bec05b.jpg", + "table_caption": [ + "Table 1. Comparison to related datasets on settings (left) and annotations (right). \"Unique hours\" refers to the cumulative duration of distinct video recordings, counting only one camera's footage for simultaneous recordings of the same activity. $\\diamond$ : Partially included." + ], + "table_footnote": [], + "table_body": "
Dataset# videosAvg. min# segsAvg. secverb classesnoun classes# verb /seg#noun /seg#words /seg
CharadesEgo [109]78600.569k1.83237110
HOMEAGE [99]57000.926k-2986--0
EK-100 [18]7008.590k3.19730011.23.0
Assembly101 [106]43217.183k1.72490110
Ego4D [30]99126.477k8.0115478117.4
Ours-ego43213.464k4.6952541.82.516.9
Ours-exo3154.514k4.7822512.33.019.4
", + "bbox": [ + 81, + 301, + 467, + 412 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Table 2. Contemporary egocentric datasets. We show only the fine-level actions for a fair comparison. For Ego4D [30], we select the closest subtask of \"forecasting\" following [106].", + "bbox": [ + 75, + 422, + 470, + 465 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Different from Ego4D where captions are associated with only single timestamps, our captions come with manually annotated start and end timestamps. For better visual perception across ego-exo views, we give verb and noun labels associated with the specific hand. One aspect from which we can analyze the human's ability to bridge ego-exo activities is through gaze. EgoExoLearn is also augmented with calibrated eye-gaze signals. These annotations enable the understanding of human ability to bridge ego-exo activities from diverse perspectives, which we posit will benefit the next-generation embodied AI agents [21, 62].", + "bbox": [ + 75, + 486, + 470, + 652 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Benchmarks", + "text_level": 1, + "bbox": [ + 76, + 665, + 210, + 679 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To evaluate the ability of bridging asynchronous ego-exo procedural activities, we introduce 4 new benchmarks: 1) cross-view association, 2) cross-view action understanding, 3) cross-view referenced skill assessment, and 4) cross-view referenced captioning. The cross-view action understanding benchmark is further subdivided into three subtasks: cross-view action anticipation, cross-view action planning, and cross-view action segmentation. Additionally, we explore the role of gaze in assisting these tasks. We also benchmark models on zero-shot and supervised fine-grained action recognition tasks for reference, following [18, 126]. Note that we carefully split our dataset to eliminate annotation leak across benchmarks. Due to the space limit, we only provide partial content of definition, annotation, results, and analysis,", + "bbox": [ + 75, + 688, + 470, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "leaving more complete details in the supplementary.", + "bbox": [ + 500, + 305, + 843, + 321 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2.1 Cross-view association", + "text_level": 1, + "bbox": [ + 500, + 344, + 710, + 358 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Motivation. One straightforward indicator of the ego-exo activity bridging is the ability to associate the same semantics across ego- and exo-views. This benchmark focuses on equipping models with this cross-view association ability. An application of this ability is assistants in AR that can show expert demonstration videos when the human is confused [157]. Another potential application is the embodied AI agent that can explain its decision [118, 127, 146].", + "bbox": [ + 496, + 373, + 893, + 494 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Problem settings. We formulate this association benchmark as a cross-view multiple-choice association problem. Specifically, we consider two different cross-view association settings: ego2exo and exo2ego. In the case of ego2exo, given an egocentric video, the model is asked to predict the corresponding exocentric video performing the same action from a candidate choice set of exocentric samples, and vice versa for the exo2ego setting. For both ego2exo and exo2ego settings, we use 20 candidate samples for each query. The evaluation metric is the averaged Top-1 accuracy.", + "bbox": [ + 496, + 500, + 893, + 650 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Annotations. We meticulously construct the ground-truth ego-exo pairs via a semantic-aware matching process. It is composed of five stages with details in the supplementary material: (1) Scenario Matching. (2) Noun and Verb Matching. (3) Sentence Matching with LLM. (4) Negative Sampling. (5) Two-round Manual Verification. Notably, we do not provide such pairs for the training set and leave the modeling of cross-view association on unpaired samples to be further explored for the community.", + "bbox": [ + 496, + 654, + 893, + 790 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Baseline model. We adopt three types of baseline models, i.e., ego-only, exo-only, and ego-exo, which refer to the data we used during training. Under all the settings, we leverage the paired video and caption and jointly train a video encoder and a text encoder using the contrastive loss [97]. We use TimeSformer-B [8] as the video encoder and clip-text [97] as the text encoder. We initialize both encoders", + "bbox": [ + 496, + 794, + 895, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "22076", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/91014facf6f8bcfa9554cb9dad5ad362d552cd3af1a70c237964fa8454581a74.jpg", + "image_caption": [ + "Figure 4. Concept of the 3 benchmarks of cross-view association (Sec. 4.2.1), cross-view action anticipation & planning (Sec. 4.2.2) and cross-view reference skill assessment (Sec.4.2.3) in this section. Other benchmarks can be found in the supplementary material." + ], + "image_footnote": [], + "bbox": [ + 83, + 90, + 361, + 215 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/5929e0bdfdd454c39af711fa6adb1ae67b6f37f7d0c96e68327bb792422ed6a8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 369, + 90, + 629, + 215 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/e3d6d14183859995b34ad8502e15e3d09f2b00be8121242e9fa5eab2f01469eb.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 638, + 90, + 888, + 215 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "using the EgoVLP [72] pre-trained weights. During testing, we obtain the ego/exocentric video representations using the video encoder. The prediction is defined as the one with the highest normalized cross-view video feature similarity among all candidates. On top of these models, we evaluate the effectiveness of gaze in associating egocentric video and exocentric video. This is achieved by replacing the original egocentric video with spatially cropped videos using the gaze positions as the cropping center.", + "bbox": [ + 75, + 265, + 468, + 400 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Related work. Prior works on pre-trained vision-language models for multiple choice association/questioning [60, 72, 122, 130, 138, 158] are generally pre-trained on either exocentric [3, 82] or egocentric [30] datasets, only revealing weak cross-view bridging ability. Another line of works explores cross-view learning [2, 33, 135, 141] by either transferring the knowledge from one view to the other [69] or training view-invariant video understanding models [124, 143]. Different from previous works, our cross-view association benchmark is in a more realistic but challenging setting, by evaluating the model's ability to associate asynchronous activities across ego- and exo-views.", + "bbox": [ + 75, + 407, + 470, + 589 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Experiment results. We first evaluate several vision-language models via zero-shot transfer. These models are pre-trained either on egocentric videos, i.e. EgoVLP [72] and LaViLa [158], or exocentric videos, i.e., InternVideo [129]. As shown in Tab. 3, without using gaze, EgoVLP generally outperforms the others on both validation and test sets. By introducing gaze information, InternVideo receives a decent improvement, especially on Exo2Ego.", + "bbox": [ + 75, + 595, + 468, + 715 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "As for fine-tuned models, Tab. 3, reveals that models trained solely on single-view data struggle with cross-view association. Training ego-only models using gaze-cropped egocentric videos results in substantial improvements, significantly outperforming those trained on center-cropped videos. This again highlights the importance of gaze in enhancing cross-view association. Based on our observation, the regions around gaze help the cluttered ego videos become visually similar to exo videos where the primary object is salient. Last, we show the baseline result when co-trained on both egocentric and exocentric videos. The model in this setting shows the strongest cross-view association ability over", + "bbox": [ + 75, + 719, + 470, + 901 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/3db0d1c8bb14337be8ff259c1b7bf9418393f6123267379f5f7711b0b2bf2a65.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodGazeValTest
Ego2ExoExo2EgoEgo2ExoExo2Ego
Zero-shot
Random✘12.715.014.113.4
EgoVLP [72]✘28.827.232.128.9
LaViLa [158]✘22.624.928.725.7
InternVideo [129]✘27.021.230.621.7
EgoVLP [72]✓28.829.731.528.9
LaViLa [158]✓21.921.430.325.9
InternVideo [129]✓30.932.333.332.2
Fine-tuned
Exo-only✘42.941.745.446.9
Ego-only✘33.637.140.335.8
Ego-only✓34.638.745.641.8
Ego-onlyCenter25.422.824.724.2
EgoExo✘42.945.449.045.3
EgoExo✓47.948.855.351.1
", + "bbox": [ + 504, + 261, + 888, + 494 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 3. Association accuracy in the cross-view association benchmark. In the fine-tuned setting, we adopt three kinds of data sources for training, i.e., ego-only, exo-only, and hybrid ego-exo data. By leveraging gaze information during training, the model outperforms the baseline (w/o gaze) and the center-crop counterpart.", + "bbox": [ + 498, + 505, + 892, + 575 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "single-view models. Findings from this benchmark underline the limitation of current models in associating activities across ego and exo views, and point towards the potential benefits of integrating gaze into the association.", + "bbox": [ + 498, + 590, + 893, + 650 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2.2 Cross-view action anticipation & planning", + "text_level": 1, + "bbox": [ + 500, + 671, + 846, + 686 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Motivation. The procedural actions are not guaranteed to be identical between the two views due to practical constraints. Thus, we design benchmarks for cross-view action anticipation and planning to enable a thorough understanding and transfer of necessary steps (or actions) for task completion, bridging the gap between views and considering real-world conditions. A practical application of this benchmark can be seen in human-robot collaboration scenarios. For instance, an embodied AI agent, after observing a human perform the first part of a task, could effectively take over and complete the remaining half of the task based on the particular environmental situation.", + "bbox": [ + 496, + 699, + 893, + 878 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Problem settings. For both the cross-view anticipation and", + "bbox": [ + 500, + 885, + 892, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "22077", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/79d9f1fe7c9237ded359ee3193ca1a430d4bad06545cc88009b740d7d7cb89c4.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodGazeAnticipation↑Planning↓
Ego-VEgo-NExo-VExo-NEgoExo
Exo-onlyX29.923.640.940.584.776.1
Ego-onlyX33.437.828.917.683.484.5
Ego-only✓40.552.837.637.680.082.6
Ego-onlyCenter33.238.634.132.782.684.7
Unsupervised Domain Adaption
Ego2ExoX33.638.135.428.783.084.1
Exo2EgoX30.423.639.239.883.979.0
Ego2Exo✓40.854.238.737.182.884.3
Exo2Ego✓33.531.339.140.182.478.8
Knowledge Distillation
Ego2ExoX29.624.941.645.284.375.5
Exo2EgoX34.038.428.618.683.184.3
Ego2Exo✓29.925.041.245.184.875.1
Exo2Ego✓41.056.137.739.179.582.6
Co-training
Ego & ExoX33.537.439.644.383.276.0
Ego & Exo✓43.853.340.344.479.075.6
", + "bbox": [ + 76, + 88, + 470, + 345 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 4. Results of cross-view action anticipation and planning benchmarks. For anticipation, the class-mean Top-5 recall is used as the evaluation metric (higher is better). For planning, the Edit distance is used as the evaluation metric (lower is better).", + "bbox": [ + 75, + 356, + 470, + 412 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "cross-view planning, the goal is to anticipate the future activities in one view, given labeled training data only in another view. For the cross-view anticipation task, we focus on predicting the verb and noun categories of the next fine-level action $\\tau = 1$ second into the future. Given the multilabel nature of our verb and noun annotations in each fine-level segment, we perform multiclass anticipation. Performance is evaluated by class-mean Top-5 recall as per [18]. For the cross-view planning task, we aim to generate the next $K = 8$ steps of coarse-level actions. We adopt ED@ $K$ as the evaluation metric following the setting of Ego4D LTA [30].", + "bbox": [ + 75, + 425, + 470, + 592 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Annotations. For action anticipation, we use the fine-level verb and noun annotations, and then take their intersection between ego and exo videos to constrain them in the same closed set. To evaluate the model more effectively, we further control the long-tail degree of the data by and filter out the tail categories that occur less than $\\frac{1}{100}$ of the highest occurrence category. For the action planning task, we directly adopt the coarse-level action annotations and regard the start timestamp of each segment as one action step. More details can be found in the supplementary material.", + "bbox": [ + 75, + 595, + 470, + 747 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Baseline model. We explore three distinct directions to implement cross-view baseline models. The first direction is based on unsupervised domain adaptation (UDA), treating one view as the source domain and the other as the target domain. This method operates within an unsupervised training framework, using labels from the source domain and video data from both domains [16, 52, 84, 104, 111, 131, 145]. We adopt CLIP [97] + TA3N [14] as the baseline model. The second direction entails knowledge distillation (KD) [28, 32, 87], allowing the model trained on one view to learn knowl", + "bbox": [ + 75, + 750, + 473, + 901 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "edge of the other view, under the assumption that a teacher model of the other view is available. We equip CLIP with a distillation approach based on [69] to transfer knowledge from the teacher model to a newly created student model. The third and most straightforward direction is co-training (CT) using the data from both views to encourage the model to discover correlations between them directly. For using gaze, we also crop the video based on the gaze positions.", + "bbox": [ + 496, + 90, + 893, + 212 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We comprehensively consider four evaluation settings. The \"Ego-only\" and \"Exo-only\" settings do not involve cross-view understanding, thus we use zero-shot evaluation serving as the references. The Ego2Exo and Exo2Ego settings are the cross-view settings. For UDA, \"Ego2Exo\" is defined as utilizing the egocentric view as the source domain and the exocentric view as the target domain. In the context of KD, \"Ego2Exo\" indicates we initially train a teacher model on egocentric data, followed by the training and distillation of the student model on exocentric data. For CT, we merge both egocentric and exocentric datasets through direct concatenation. We report results on the test set and put the validation set results in the supplementary.", + "bbox": [ + 496, + 212, + 895, + 409 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Related work. Prior works [25, 54, 106, 117] on multi-view action understanding mainly focus on synchronized multiview videos. Some work studied transferring knowledge [69] from one view to the other or training the view-invariant [124, 143, 159] video models. Our cross-view benchmarks seek to evaluate the ability of models to bridge asynchronous actions across views, which is more challenging yet realistic.", + "bbox": [ + 496, + 411, + 893, + 518 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Experiment results. Table 4 presents the results of action anticipation and planning on the test set. The first block of results shows a significant performance gap when models trained exclusively on one view are tested on the other view. This underscores the inherent differences in activities captured in the two views. Remarkably, even without relying on any specific cross-view method, the inclusion of gaze information markedly diminishes the disparity between egocentric and exocentric data. Leveraging techniques such as UDA or KD we can see improved performance compared with direct zero-shot inference in the first block. Since CT can utilize both egocentric and exocentric data, it achieves the best performance in all setups. Moreover, from the comparison in all settings using or not using gaze, it is clear that gaze serves as a surprisingly effective signal to mitigate the gap between activities in the two views, although naively designed. These results take the first step in the potential directions for better bridging the cross-view activities, setting a foundation for future advancements in the field of cross-view action understanding.", + "bbox": [ + 496, + 521, + 895, + 824 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2.3 Cross-view referenced skill assessment", + "text_level": 1, + "bbox": [ + 498, + 842, + 821, + 857 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Motivation and problem setting. We propose a novel task of cross-view referenced skill assessment leveraging the", + "bbox": [ + 496, + 869, + 893, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "22078", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "unique setting and annotations of our dataset. This task goes beyond the traditional pairwise ranking often used in skill assessment [22, 58] by incorporating an expert demonstration video as a reference point. This demonstration provides a model of the ideal execution of an action, offering a standard against which to compare skill levels. In this task, the input is a pair of egocentric video clips $C_{ego1}$ , $C_{ego2}$ of the same action and an exo-view demonstration video clip $C_{exo}$ as the reference, and the output is a choice $c \\in \\{ego1, ego2\\}$ of the egocentric clip that demonstrates a higher skill level. Moreover, we can also assess the skill level by the relation between action and gaze. This benchmark evaluates a model's ability to bridge asynchronous ego-exo dynamics. A practical application is an AR system that helps humans in skill acquisition by providing targeted feedback based on skill level assessment and expert demonstration.", + "bbox": [ + 75, + 90, + 472, + 333 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Annotations. For the cross-view referenced skill assessment benchmark, we concentrate on four types of representative actions, as detailed in Tab. 5. We use the skill level annotations in Sec. 3.2 in this benchmark. For each pair of videos, the annotation is provided by four distinct annotators to minimize subjective bias. We ensure credibility by checking the transitivity of annotations and removing the pairs with less than 3 agreements among 4 annotators. We then append an exo-view demonstration video clip of the same action, forming video triplets as the model input.", + "bbox": [ + 75, + 338, + 472, + 491 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Baseline model. Our baseline model is built upon a pairwise ranking skill assessment model RAAN [23]. Without loss of generality, assuming egol is the video showing a higher skill level, we apply the following approaches to leverage the reference exo-view demonstration video: 1) Triplet loss (TL). The feature distance between $C_{exo}$ and $C_{ego1}$ should be closer to the distance between $C_{exo}$ and $C_{ego2}$ . 2) Relation network (RN). Inspired by [114], we employ a relation network that concatenates the features of the ego and exo clips. This network is designed to discern which of the two egocentric video clips bears a closer relation to the demonstration video in terms of skill level. The way to gaze is consistent with the other benchmarks.", + "bbox": [ + 75, + 494, + 472, + 691 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Related work. Several previous works on skill assessment aim to directly regress a score based on professional ratings [1, 71, 73, 91, 115, 149]. We adopt a more general approach of pairwise ranking since no absolute score is available in most real-world skills [7, 70]. Previous works in this direction only use a pair of videos that are in either egocentric view [22] or in exo view [23, 70]. Differently from their works, we explore the cross-view activity bridging ability by examining how demonstration videos from exo view can benefit skill assessment in ego views.", + "bbox": [ + 75, + 698, + 472, + 849 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Experiment results. We first evaluate the performance of previous works under the conventional pairwise setting. In the upper block of Table 5, both methods [22, 23] receive", + "bbox": [ + 75, + 854, + 472, + 902 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/5279d34b49c02b1db117aed65e7923bffd208d3b1d335273556b809779f5bf5e.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodGazeEgg CrackingPeelingStir-fryCutting
Ego pairs only
Who's better [22]✘74.7975.9877.5476.85
RAAN [23]✘78.4578.5282.5379.09
Who's better [22]✓77.9176.7079.1377.02
RAAN [23]✓82.0879.3783.9279.36
Ego pairs + Exo
RAAN [23] + RN✘77.9277.0981.5478.26
RAAN [23] + TL✘78.8179.4682.5078.73
RAAN [23] + RN✓82.1479.3283.5179.44
RAAN [23] + TL✓82.1679.5984.0579.29
", + "bbox": [ + 504, + 88, + 890, + 241 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 5. Ranking accuracy of cross-view referenced skill assessment. In the upper part of the table, only ego video pairs are used, while in the lower part, exo demonstrations are incorporated by \"RN\": relation network and \"TL\": triplet loss.", + "bbox": [ + 496, + 250, + 893, + 306 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "a clear performance boost when gaze is used. This aligns with behavioral science findings that experts and novices have different gaze patterns in the same task [76, 133]. With the exocentric demonstration as reference, both the relation network method (RN) and the triplet loss method (TL) can leverage the reference video and bring performance improvement. Further adding gaze we can get the best performance. However, the marginal improvement observed with the inclusion of the exocentric reference suggests that current models may still struggle to fully bridge asynchronous activities across egocentric and exocentric views. There remains ample room for new improvement and innovation.", + "bbox": [ + 496, + 321, + 895, + 503 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 518, + 619, + 535 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The ability to bridge asynchronous procedural activities in ego- and exo-views is imperative for next-generation embodied AI in executing sophisticated tasks in the real world. As a fundamental step, our EgoExoLearn encompasses a rich collection of egocentric videos, each captured when replicating procedures of exocentric demonstration videos, but performed in different environments and at different times. This realistic setup, combined with our multimodal annotations, allows us to construct 4 novel benchmarks, serving as a versatile platform for investigating how cross-view asynchronous activities can be bridged. EgoExoLearn also enables new research directions e.g., how to better leverage gaze and hand-associated annotations. Results from the benchmarks show weaknesses of current models in bridging ego- and exo-view asynchronous activities, leaving significant room for future work to improve upon.", + "bbox": [ + 496, + 545, + 893, + 787 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgement. This work is supported by the National Key R&D Program of China (No.2022ZD0160102) and the Industry Collaboration Projects Grant, Shanghai Committee of Science and Technology, China (No.22YF1461500).", + "bbox": [ + 496, + 787, + 890, + 844 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Key contribution statement: Guo made key contributions to annotation processing and 3 action benchmarks. Jilan made key contributions to the association and caption benchmarks. Mingfang contributed primarily to the skill benchmark.", + "bbox": [ + 496, + 844, + 893, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "22079", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Narges Ahmidi, Lingling Tao, Shahin Sefati, Yixin Gao, Colin Lea, Benjamin Bejar Haro, Luca Zappella, Sanjeev Khudanpur, René Vidal, and Gregory D Hager. A dataset and benchmarks for segmentation and recognition of gestures in robotic surgery. IEEE Transactions on Biomedical Engineering, 64(9):2025-2041, 2017. 8", + "[2] Shervin Ardeshir and Ali Borji. An exocentric look at egocentric actions and vice versa. Computer Vision and Image Understanding, 171:61-68, 2018. 6", + "[3] Max Bain, Arsha Nagrani, Gül Varol, and Andrew Zisserman. Frozen in time: A joint video and image encoder for end-to-end retrieval. In Proceedings of the International Conference on Computer Vision (ICCV), 2021. 3, 6", + "[4] Albert Bandura. Observational learning. The international encyclopedia of communication, 2008. 1, 4", + "[5] Siddhant Bansal, Chetan Arora, and C.V. Jawahar. My view is the best view: Procedure learning from egocentric videos. In Proceedings of the European Conference on Computer Vision (ECCV), 2022. 2", + "[6] Yizhak Ben-Shabat, Xin Yu, Fatemeh Saleh, Dylan Campbell, Cristian Rodriguez-Opazo, Hongdong Li, and Stephen Gould. The ikaa asm dataset: Understanding people assembling furniture through actions, objects and pose. In Proceedings of the Winter Conference on Applications of Computer Vision (WACV), 2021. 2", + "[7] Gedas Bertasius, Hyun Soo Park, Stella X Yu, and Jianbo Shi. Am i a baller? basketball performance assessment from first-person videos. In Proceedings of the International Conference on Computer Vision (ICCV), 2017. 8", + "[8] Gedas Bertasius, Heng Wang, and Lorenzo Torresani. Is space-time attention all you need for video understanding? In Proceedings of the International Conference on Machine Learning (ICML), 2021. 5", + "[9] Vinay Bettadapura, Irfan Essa, and Caroline Pantofaru. Ego-centric field-of-view localization using first-person point-of-view devices. In Proceedings of the Winter Conference on Applications of Computer Vision (WACV), 2015. 3", + "[10] Steven Bird. Nltk: the natural language toolkit. In Proceedings of the COLING/ACL Interactive Presentation Sessions, 2006. 4", + "[11] Greg Brockman, Vicki Cheung, Ludwig Pettersson, Jonas Schneider, John Schulman, Jie Tang, and Wojciech Zaremba. Openai gym. arXiv preprint arXiv:1606.01540, 2016. 1", + "[12] Minjie Cai, Kris M Kitani, and Yoichi Sato. Understanding hand-object manipulation with grasp types and object attributes. In Robotics: Science and Systems, 2016. 3", + "[13] Guo Chen, Yifei Huang, Jilan Xu, Baoqi Pei, Zhe Chen, Zhiqi Li, Jiahao Wang, Kunchang Li, Tong Lu, and Limin Wang. Video mamba suite: State space model as a versatile alternative for video understanding. arXiv preprint arXiv:2403.09626, 2024. 3", + "[14] Min-Hung Chen, Zsolt Kira, Ghassan AlRegib, Jaekwon Yoo, Ruxin Chen, and Jian Zheng. Temporal attentive alignment for large-scale video domain adaptation. In Proceedings of the International Conference on Computer Vision (ICCV), 2019. 2, 7" + ], + "bbox": [ + 86, + 114, + 470, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[15] Kellie Corona, Katie Osterdahl, Roderic Collins, and Anthony Hoogs. Meva: A large-scale multiview, multimodal video dataset for activity detection. In Proceedings of the Winter Conference on Applications of Computer Vision (WACV), 2021. 2", + "[16] Victor G Turrisi da Costa, Giacomo Zara, Paolo Rota, Thiago Oliveira-Santos, Nicu Sebe, Vittorio Murino, and Elisa Ricci. Dual-head contrastive domain adaptation for video action recognition. In Proceedings of the Winter Conference on Applications of Computer Vision (WACV), 2022. 7", + "[17] Dima Damen, Hazel Doughty, Giovanni Maria Farinella, Sanja Fidler, Antonino Furnari, Evangelos Kazakos, Davide Moltisanti, Jonathan Munro, Toby Perrett, Will Price, et al. Scaling egocentric vision: The epic-kitchens dataset. In Proceedings of the European Conference on Computer Vision (ECCV), 2018. 2", + "[18] Dima Damen, Hazel Doughty, Giovanni Maria Farinella, Antonino Furnari, Jian Ma, Evangelos Kazakos, Davide Moltisanti, Jonathan Munro, Toby Perrett, Will Price, and Michael Wray. Rescaling egocentric vision: Collection, pipeline and challenges for epic-kitchens-100. International Journal of Computer Vision (IJCV), 130:33-55, 2022. 1, 2, 4, 5, 7", + "[19] Ahmad Darkhalil, Dandan Shan, Bin Zhu, Jian Ma, Amlan Kar, Richard Higgins, Sanja Fidler, David Fouhey, and Dima Damen. Epic-kitchens visor benchmark: Video segmentations and object relations. In Proceedings of the Neural Information Processing Systems (NeurIPS) Track on Datasets and Benchmarks, 2022. 2", + "[20] Fernando De la Torre, Jessica Hodgins, Adam Bargteil, Xavier Martin, Justin Macey, Alex Collado, and Pep Beltran. Guide to the carnegie mellon university multimodal activity (cmu-mmac) database. 2009. 2", + "[21] Matt Deitke, Winson Han, Alvaro Herrasti, Aniruddha Kembhavi, Eric Kolve, Roozbeh Mottaghi, Jordi Salvador, Dustin Schwenk, Eli VanderBilt, Matthew Wallingford, et al. Robothor: An open simulation-to-real embodied ai platform. In Proceedings of the Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 5", + "[22] Hazel Doughty, Dima Damen, and Walterio Mayol-Cuevas. Who's better? who's best? pairwise deep ranking for skill determination. In Proceedings of the Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 8", + "[23] Hazel Doughty, Walterio Mayol-Cuevas, and Dima Damen. The pros and cons: Rank-aware temporal attention for skill determination in long videos. In Proceedings of the Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 8", + "[24] Mitch J Fryling, Cristin Johnston, and Linda J Hayes. Understanding observational learning: An interbehavioral approach. The Analysis of verbal behavior, 27:191, 2011. 1", + "[25] Yanwei Fu, Yanwen Guo, Yanshu Zhu, Feng Liu, Chuanming Song, and Zhi-Hua Zhou. Multi-view video summarization. IEEE Transactions on Multimedia, 12(7):717-729, 2010. 7", + "[26] Antonino Furnari and Giovanni Maria Farinella. What would you expect? anticipating egocentric actions with rolling-unrolling lstms and modality attention. In Proceedings of" + ], + "bbox": [ + 509, + 92, + 893, + 901 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "22080", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "the International Conference on Computer Vision (ICCV), 2019.3", + "[27] Rohit Girdhar and Kristen Grauman. Anticipative video transformer. In Proceedings of the International Conference on Computer Vision (ICCV), 2021. 3", + "[28] Rohit Girdhar, Du Tran, Lorenzo Torresani, and Deva Ramanan. Distinct: Learning video representations without a single labeled video. In Proceedings of the International Conference on Computer Vision (ICCV), 2019. 7", + "[29] Mohit Goyal, Sahil Modi, Rishabh Goyal, and Saurabh Gupta. Human hands as probes for interactive object understanding. In Proceedings of the Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 3", + "[30] Kristen Grauman, Andrew Westbury, and Eugene Byrne et al. Ego4d: Around the World in 3,000 Hours of Egocentric Video. In Proceedings of the Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 1, 2, 4, 5, 6, 7", + "[31] Kristen Grauman, Andrew Westbury, et al. Ego-exo4d: Understanding skilled human activity from first-and third-person perspectives. arXiv preprint arXiv:2311.18259, 2023. 2", + "[32] Geoffrey Hinton, Oriol Vinyals, and Jeff Dean. Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531, 2015. 7", + "[33] Hsuan-I Ho, Wei-Chen Chiu, and Yu-Chiang Frank Wang. Summarizing first-person videos from third persons' points of view. In Proceedings of the European Conference on Computer Vision (ECCV), 2018. 6", + "[34] Nicola J Hodges, A Mark Williams, Spencer J Hayes, and Gavin Breslin. What is modelled during observational learning? Journal of sports sciences, 25(5):531-545, 2007. 1, 4", + "[35] Jiang Hua, Liangcai Zeng, Gongfa Li, and Zhaojie Ju. Learning for a robot: Deep reinforcement learning, imitation learning, transfer learning. Sensors, 21(4):1278, 2021. 1", + "[36] Chien-Ming Huang, Sean Andrist, Allison Sauppé, and Bilge Mutlu. Using gaze patterns to predict task intent in collaboration. Frontiers in psychology, 6:1049, 2015. 3", + "[37] Yifei Huang, Minjie Cai, Zhenqiang Li, and Yoichi Sato. Predicting gaze in egocentric video by learning task-dependent attention transition. In Proceedings of the European Conference on Computer Vision (ECCV), 2018. 3", + "[38] Yifei Huang, Minjie Cai, Zhenqiang Li, Feng Lu, and Yoichi Sato. Mutual context network for jointly estimating egocentric gaze and action. IEEE Transactions on Image Processing, 29:7795-7806, 2020.", + "[39] Yifei Huang, Minjie Cai, and Yoichi Sato. An ego-vision system for discovering human joint attention. IEEE Transactions on Human-Machine Systems, 50(4):306-316, 2020. 3", + "[40] Yifei Huang, Lijin Yang, and Yoichi Sato. Compound prototype matching for few-shot action recognition. In Proceedings of the European Conference on Computer Vision (ECCV), 2022. 3", + "[41] Yifei Huang, Lijin Yang, and Yoichi Sato. Weakly supervised temporal sentence grounding with uncertainty-guided self-training. In Proceedings of the Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 3" + ], + "bbox": [ + 86, + 90, + 470, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[42] Thomas E Hutchinson, K Preston White, Worthy N Martin, Kelly C Reichert, and Lisa A Frey. Human-computer interaction using eye-gaze input. IEEE Transactions on systems, man, and cybernetics, 19(6):1527-1534, 1989. 3", + "[43] Eric Jang, Alex Irpan, Mohi Khansari, Daniel Kappler, Frederik Ebert, Corey Lynch, Sergey Levine, and Chelsea Finn. Bc-z: Zero-shot task generalization with robotic imitation learning. In Proceedings of the Conference on Robot Learning, 2022. 1", + "[44] Youngkyoon Jang, Brian Sullivan, Casimir Ludwig, Iain Gilchrist, Dima Damen, and Walterio Mayol-Cuevas. Epicent: An egocentric video dataset for camping tent assembly. In Proceedings of the International Conference on Computer Vision Workshops (ICCVW), 2019. 3", + "[45] Baoxiong Jia, Yixin Chen, Siyuan Huang, Yixin Zhu, and Song-Chun Zhu. Lemma: A multi-view dataset for the arning multi-agent multi-task activities. In Proceedings of the European Conference on Computer Vision (ECCV), 2020. 2, 5", + "[46] Baoxiong Jia, Ting Lei, Song-Chun Zhu, and Siyuan Huang. Egotaskqa: Understanding human tasks in egocentric videos. In Proceedings of the Advances in Neural Information Processing Systems (NeurIPS), 2022. 2", + "[47] Hao Jiang and Vamsi Krishna Ithapu. Egocentric pose estimation from human vision span. In Proceedings of the International Conference on Computer Vision (ICCV), 2021. 3", + "[48] Soo-Han Kang and Ji-Hyeong Han. Video captioning based on both egocentric and exocentric views of robot vision for human-robot interaction. International Journal of Social Robotics, pages 1–11, 2021. 3", + "[49] Moritz Kassner, William Patera, and Andreas Bulling. Pupil: An open source platform for pervasive eye tracking and mobile gaze-based interaction. In Adjunct Proceedings of the 2014 ACM International Joint Conference on Pervasive and Ubiquitous Computing, 2014. 3", + "[50] Evangelos Kazakos, Arsha Nagrani, Andrew Zisserman, and Dima Damen. Epic-fusion: Audio-visual temporal binding for egocentric action recognition. In Proceedings of the International Conference on Computer Vision (ICCV), 2019. 3", + "[51] Daekyum Kim, Brian Byunghyun Kang, Kyu Bum Kim, Hyungmin Choi, Jeesoo Ha, Kyu-Jin Cho, and Sungho Jo. Eyes are faster than hands: A soft wearable robot learns user intention from the egocentric view. Science Robotics, 4(26), 2019. 3", + "[52] Donghyun Kim, Yi-Hsuan Tsai, Bingbing Zhuang, Xiang Yu, Stan Sclaroff, Kate Saenko, and Manmohan Chandraker. Learning cross-modal contrastive features for video domain adaptation. In Proceedings of the International Conference on Computer Vision (ICCV), 2021. 7", + "[53] Hildegard Kuehne, Hueihan Jhuang, Estíbaliz Garrote, Tomaso Poggio, and Thomas Serre. Hmdb: a large video database for human motion recognition. In Proceedings of the International Conference on Computer Vision (ICCV), 2011. 2", + "[54] Hilde Kuehne, Ali Arslan, and Thomas Serre. The language of actions: Recovering the syntax and semantics of goal" + ], + "bbox": [ + 509, + 90, + 893, + 901 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "22081", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "directed human activities. In Proceedings of the Conference on Computer Vision and Pattern Recognition (CVPR), 2014, 2, 7", + "[55] Taein Kwon, Bugra Tekin, Jan Stühmer, Federica Bogo, and Marc Pollefeys. H2o: Two hands manipulating objects for first person interaction recognition. In Proceedings of the International Conference on Computer Vision (ICCV), 2021. 2, 5", + "[56] Bolin Lai, Miao Liu, Fiona Ryan, and James M Rehg. In the eye of transformer: Global-local correlation for egocentric gaze estimation and beyond. International Journal of Computer Vision (IJCV), pages 1-18, 2023. 3", + "[57] Stephane Lallée, Eiichi Yoshida, Anthony Mallet, Francesco Nori, Lorenzo Natale, Giorgio Metta, Felix Warneken, and Peter Ford Dominey. Human-robot cooperation based on interaction learning. From motor learning to interaction learning in robots, pages 491–536, 2010. 1", + "[58] Kyle Lam, Junhong Chen, Zeyu Wang, Fahad M Iqbal, Ara Darzi, Benny Lo, Sanjay Purkayastha, and James M Kinross. Machine learning for technical skill assessment in surgery: a systematic review. NPJ digital medicine, 5(1):24, 2022. 8", + "[59] Jangwon Lee and Michael S Ryoo. Learning robot activities from first-person human videos using convolutional future regression. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), 2017. 1", + "[60] Jie Lei, Linjie Li, Luowei Zhou, Zhe Gan, Tamara L Berg, Mohit Bansal, and Jingjing Liu. Less is more: Clipbert for video-and-language learning via sparse sampling. In Proceedings of the Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 6", + "[61] Chengshu Li, Fei Xia, Roberto Martin-Martin, Michael Lingelbach, Sanjana Srivastava, Bokui Shen, Kent Vainio, Cem Gokmen, Gokul Dharan, Tanish Jain, et al. igibson 2.0: Object-centric simulation for robot learning of everyday household tasks. arXiv preprint arXiv:2108.03272, 2021. 2", + "[62] Chengshu Li, Ruohan Zhang, Josiah Wong, Cem Gokmen, Sanjana Srivastava, Roberto Martin-Martin, Chen Wang, Gabrael Levine, Michael Lingelbach, Jiankai Sun, et al. Behavior-1k: A benchmark for embodied ai with 1,000 everyday activities and realistic simulation. In Proceedings of the Conference on Robot Learning, 2023. 5", + "[63] Haoxin Li, Yijun Cai, and Wei-Shi Zheng. Deep dual relation modeling for egocentric interaction recognition. In Proceedings of the Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 3", + "[64] Jiayi Li, Tao Lu, Xiaoge Cao, Yinghao Cai, and Shuo Wang. Meta-imitation learning by watching video demonstrations. In Proceedings of the International Conference on Learning Representations, 2021. 1", + "[65] Yin Li, Alireza Fathi, and James M Rehg. Learning to predict gaze in egocentric video. In Proceedings of the International Conference on Computer Vision (ICCV), 2013. 3", + "[66] Yin Li, Zhefan Ye, and James M Rehg. Delving into egocentric actions. In Proceedings of the Conference on Computer Vision and Pattern Recognition (CVPR), 2015." + ], + "bbox": [ + 86, + 90, + 470, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[67] Yin Li, Miao Liu, and James M Rehg. In the eye of beholder: Joint learning of gaze and actions in first person video. In Proceedings of the European Conference on Computer Vision (ECCV), 2018. 3, 5", + "[68] Yin Li, Miao Liu, and Jame Rehg. In the eye of the beholder: Gaze and actions in first person video. IEEE Transaction on Pattern Analysis and Machine Intelligence (TPAMI), 2021. 3", + "[69] Yanghao Li, Tushar Nagarajan, Bo Xiong, and Kristen Grauman. Ego-exo: Transferring visual representations from third-person to first-person videos. In Proceedings of the Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 3, 6, 7", + "[70] Zhenqiang Li, Yifei Huang, Minjie Cai, and Yoichi Sato. Manipulation-skill assessment from videos with spatial attention network. In Proceedings of the International Conference on Computer Vision Workshops (ICCVW), 2019. 8", + "[71] Zhenqiang Li, Lin Gu, Weimin Wang, Ryosuke Nakamura, and Yoichi Sato. Surgical skill assessment via video semantic aggregation. In Proceedings of the International Conference on Medical Image Computing and Computer-Assisted Intervention (MICCAI), 2022. 8", + "[72] Kevin Qinghong Lin, Jinpeng Wang, Mattia Soldan, Michael Wray, Rui Yan, Eric Z XU, Difei Gao, Rong-Cheng Tu, Wenzhe Zhao, Weijie Kong, et al. Egocentric video-language pretraining. In Proceedings of the Advances in Neural Information Processing Systems (NeurIPS), 2022. 6", + "[73] Daochang Liu, Qiyue Li, Tingting Jiang, Yizhou Wang, Rulin Miao, Fei Shan, and Ziyu Li. Towards unified surgical skill assessment. In Proceedings of the Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 8", + "[74] Shaowei Liu, Subarna Tripathi, Somdeb Majumdar, and Xiaolong Wang. Joint hand motion and interaction hotspots prediction from egocentric videos. In Proceedings of the Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 3", + "[75] Xiaotian Liu, Hector Palacios, and Christian Muise. Egocentric planning for scalable embodied task achievement. arXiv preprint arXiv:2306.01295, 2023. 1", + "[76] Yan Liu, Pei Yun Hsueh, Jennifer Lai, Mirweis Sangin, Marc Antoine Nussli, and Pierre Dillenbourg. Who is the expert? analyzing gaze data to predict expertise level in collaborative applications. In Proceedings of the IEEE International Conference on Multimedia & Expo (ICME), 2009. 8", + "[77] Yueyue Liu, Zhijun Li, Huaping Liu, and Zhen Kan. Skill transfer learning for autonomous robots and human-robot cooperation: A survey. Robotics and Autonomous Systems, 128:103515, 2020. 1", + "[78] Yunze Liu, Yun Liu, Che Jiang, Kangbo Lyu, Weikang Wan, Hao Shen, Boqiang Liang, Zhoujie Fu, He Wang, and Li Yi. Hoi4d: A 4d egocentric dataset for category-level human-object interaction. In Proceedings of the Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2", + "[79] Manolis Savva*, Abhishek Kadian*, Oleksandr Maksymets*, Yili Zhao, Erik Wijmans, Bhavana Jain, Julian Straub, Jia Liu, Vladlen Koltun, Jitendra Malik, Devi Parikh, and Dhruv Batra. Habitat: A Platform for Embodied AI" + ], + "bbox": [ + 509, + 90, + 893, + 900 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "22082", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Research. In Proceedings of the International Conference on Computer Vision (ICCV), 2019. 1", + "[80] Javier Marina-Miranda and V Javier Traver. Head and eye egocentric gesture recognition for human-robot interaction using eyewear cameras. IEEE Robotics and Automation Letters, 7(3):7067-7074, 2022. 3", + "[81] Joanna Materzynska, Guillaume Berger, Ingo Bax, and Roland Memisevic. The jester dataset: A large-scale video dataset of human gestures. In Proceedings of the International Conference on Computer Vision Workshops (ICCVW), 2019. 2", + "[82] Antoine Miech, Dimitri Zhukov, Jean-Baptiste Alayrac, Makarand Tapaswi, Ivan Laptev, and Josef Sivic. HowTo100M: Learning a Text-Video Embedding by Watching Hundred Million Narrated Video Clips. In Proceedings of the International Conference on Computer Vision (ICCV), 2019. 1, 6", + "[83] Kyle Min and Jason J Corso. Integrating human gaze into attention for egocentric activity recognition. In Proceedings of the Winter Conference on Applications of Computer Vision (WACV), pages 1069-1078, 2021. 3", + "[84] Jonathan Munro and Dima Damen. Multi-modal domain adaptation for fine-grained action recognition. In Proceedings of the Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 7", + "[85] Tushar Nagarajan and Kristen Grauman. Shaping embodied agent behavior with activity-context priors from egocentric video. In Proceedings of the Advances in Neural Information Processing Systems (NeurIPS), 2021. 3", + "[86] Tushar Nagarajan, Yanghao Li, Christoph Feichtenhofer, and Kristen Grauman. Ego-topo: Environment affordances from egocentric video. In Proceedings of the Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 3", + "[87] Boxiao Pan, Haoye Cai, De-An Huang, Kuan-Hui Lee, Adrien Gaidon, Ehsan Adeli, and Juan Carlos Niebles. Spatio-temporal graph for video captioning with knowledge distillation. In Proceedings of the Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 7", + "[88] Boxiao Pan, Zhangjie Cao, Ehsan Adeli, and Juan Carlos Niebles. Adversarial cross-domain action recognition with co-attention. In Proceedings of the AAAI Conference on Artificial Intelligence (AAAI), 2020. 2", + "[89] Boxiao Pan, Bokui Shen, Davis Rempe, Despoina Paschalidou, Kaichun Mo, Yanchao Yang, and Leonidas J Guibas. Copilot: Human-environment collision prediction and localization from egocentric videos. In Proceedings of the International Conference on Computer Vision (ICCV), 2023. 1", + "[90] Hyung Min Park, Seok Han Lee, and Jong Soo Choi. Wearable augmented reality system using gaze interaction. In Proceedings of the IEEE/ACM International Symposium on Mixed and Augmented Reality, 2008. 3", + "[91] Paritosh Parmar and Brendan Tran Morris. What and how well you performed? a multitask learning approach to action quality assessment. In Proceedings of the Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 8" + ], + "bbox": [ + 86, + 90, + 470, + 900 + ], + "page_idx": 11 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[92] Leo Pauly, Wisdom C Agboh, David C Hogg, and Raul Fuentes. O2a: one-shot observational learning with action vectors. Frontiers in Robotics and AI, 2021. 1", + "[93] Chiara Plizzari, Mirco Planamente, Gabriele Goletto, Marco Cannici, Emanuele Gusso, Matteo Matteucci, and Barbara Caputo. E2 (go) motion: Motion augmented event stream for egocentric action recognition. In Proceedings of the Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 3", + "[94] Chiara Plizzari, Gabriele Goletto, Antonino Furnari, Siddhant Bansal, Francesco Ragusa, Giovanni Maria Farinella, Dima Damen, and Tatiana Tommasi. An outlook into the future of egocentric vision. arXiv preprint arXiv:2308.07123, 2023. 1, 3", + "[95] David Premack and Guy Woodruff. Does the chimpanzee have a theory of mind? Behavioral and brain sciences, 1(4): 515-526, 1978. 1", + "[96] Gorjan Radevski, Dusan Grujicic, Matthew Blaschko, Marie-Francine Moens, and Tinne Tuytelaars. Multimodal distillation for egocentric action recognition. In Proceedings of the International Conference on Computer Vision (ICCV), 2023. 3", + "[97] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In Proceedings of the International Conference on Machine Learning (ICML), 2021. 5, 7", + "[98] Francesco Ragusa, Antonino Furnari, Salvatore Livatino, and Giovanni Maria Farinella. The meccano dataset: Understanding human-object interactions from egocentric videos in an industrial-like domain. In Proceedings of the Winter Conference on Applications of Computer Vision (WACV), 2021. 2, 5", + "[99] Nishant Rai, Haofeng Chen, Jingwei Ji, Rishi Desai, Kazuki Kozuka, Shun Ishizaka, Ehsan Adeli, and Juan Carlos Niebles. Home action genome: Cooperative compositional action understanding. In Proceedings of the Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 1, 2, 5", + "[100] Richard Ramsey, David M Kaplan, and Emily S Cross. Watch and learn: the cognitive neuroscience of learning from others' actions. Trends in Neurosciences, 44(6):478-491, 2021. 1", + "[101] Yosef Razin and Karen Feigh. Learning to predict intent from gaze during robotic hand-eye coordination. In Proceedings of the AAAI Conference on Artificial Intelligence (AAAI), 2017. 3", + "[102] Radiah Rivu, Yasmeen Abdrabou, Ken Pfeuffer, Augusto Esteves, Stefanie Meitner, and Florian Alt. Stare: gaze-assisted face-to-face communication in augmented reality. In Proceedings of the ACM Symposium on Eye Tracking Research and Applications (ETRA), 2020. 3", + "[103] Giacomo Rizzolatti and Laila Craighero. The mirror-neuron system. Annual Review of Neuroscience, 27:169-192, 2004. 1", + "[104] Aadarsh Sahoo, Rutav Shah, Rameswar Panda, Kate Saenko, and Abir Das. Contrast and mix: Temporal contrastive video" + ], + "bbox": [ + 503, + 92, + 893, + 900 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "22083", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 11 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "domain adaptation with background mixing. In Proceedings of the Advances in Neural Information Processing Systems (NeurIPS), 2021. 7", + "[105] Stefan Schaal. Learning from demonstration. In Proceedings of the Advances in Neural Information Processing Systems (NeurIPS), 1996. 1", + "[106] F. Sener, D. Chatterjee, D. Shelepov, K. He, D. Singhania, R. Wang, and A. Yao. Assembly101: A large-scale multiview video dataset for understanding procedural activities. In Proceedings of the Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 1, 2, 4, 5, 7", + "[107] Dandan Shan, Jiaqi Geng, Michelle Shu, and David F Fouhey. Understanding human hands in contact at internet scale. In Proceedings of the Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 3", + "[108] Mohit Shridhar, Jesse Thomason, Daniel Gordon, Yonatan Bisk, Winson Han, Roozbeh Mottaghi, Luke Zettlemoyer, and Dieter Fox. ALFRED: A Benchmark for Interpreting Grounded Instructions for Everyday Tasks. In Proceedings of the Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 1", + "[109] Gunnar A Sigurdsson, Abhinav Gupta, Cordelia Schmid, Ali Farhadi, and Karteek Alahari. Charades-ego: A large-scale dataset of paired third and first person videos. arXiv preprint arXiv:1804.09626, 2018. 1, 2, 5", + "[110] Kiran Somasundaram, Jing Dong, Huixuan Tang, Julian Straub, Mingfei Yan, Michael Goesele, Jakob Julian Engel, Renzo De Nardi, and Richard Newcombe. Project aria: A new tool for egocentric multi-modal ai research. arXiv preprint arXiv:2308.13561, 2023. 2", + "[111] Xiaolin Song, Sicheng Zhao, Jingyu Yang, Huanjing Yue, Pengfei Xu, Runbo Hu, and Hua Chai. Spatio-temporal contrastive domain adaptation for action recognition. In Proceedings of the Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 7", + "[112] Yale Song, Eugene Byrne, Tushar Nagarajan, Huiyu Wang, Miguel Martin, and Lorenzo Torresani. Ego4d goal-step: Toward hierarchical understanding of procedural activities. Proceedings of the Advances in Neural Information Processing Systems (NeurIPS), 2024. 2", + "[113] Swathikiran Sudhakaran, Sergio Escalera, and Oswald Lanz. Lsta: Long short-term attention for egocentric action recognition. In Proceedings of the Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 3", + "[114] Flood Sung, Yongxin Yang, Li Zhang, Tao Xiang, Philip HS Torr, and Timothy M Hospedales. Learning to compare: Relation network for few-shot learning. In Proceedings of the Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 8", + "[115] Yansong Tang, Zanlin Ni, Jiahuan Zhou, Danyang Zhang, Jiwen Lu, Ying Wu, and Jie Zhou. Uncertainty-aware score distribution learning for action quality assessment. In Proceedings of the Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 8", + "[116] Yuli Vasiliev. Natural language processing with Python and spaCy: A practical introduction. 2020. 4", + "[117] Shruti Vyas, Yogesh S Rawat, and Mubarak Shah. Multiview action recognition using cross-view video prediction." + ], + "bbox": [ + 78, + 92, + 470, + 900 + ], + "page_idx": 12 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "In Proceedings of the European Conference on Computer Vision (ECCV), 2020. 7", + "[118] Sebastian Wallkötter, Silvia Tulli, Ginevra Castellano, Ana Paiva, and Mohamed Chetouani. Explainable embodied agents through social cues: a review. ACM Transactions on Human-Robot Interaction (THRI), 10(3):1-24, 2021. 5", + "[119] Chen Wang, Linxi Fan, Jiankai Sun, Ruohan Zhang, Li Fei-Fei, Danfei Xu, Yuke Zhu, and Anima Anandkumar. Mimicplay: Long-horizon imitation learning by watching human play. arXiv preprint arXiv:2302.12422, 2023. 1", + "[120] Huiyu Wang, Mitesh Kumar Singh, and Lorenzo Torresani. Ego-only: Egocentric action detection without exocentric transferring. In Proceedings of the International Conference on Computer Vision (ICCV), 2023. 3", + "[121] Jiahao Wang, Guo Chen, Yifei Huang, Limin Wang, and Tong Lu. Memory-and-anticipation transformer for online action understanding. In Proceedings of the International Conference on Computer Vision (ICCV), 2023. 3", + "[122] Jinpeng Wang, Yixiao Ge, Rui Yan, Yuying Ge, Kevin Qinghong Lin, Satoshi Tsutsui, Xudong Lin, Guanyu Cai, Jianping Wu, Ying Shan, et al. All in one: Exploring unified video-language pre-training. In Proceedings of the Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 6", + "[123] Jian Wang, Diogo Luvizon, Weipeng Xu, Lingjie Liu, Kripasindhu Sarkar, and Christian Theobalt. Scene-aware egocentric 3d human pose estimation. In Proceedings of the Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 3", + "[124] Qitong Wang, Long Zhao, Liangzhe Yuan, Ting Liu, and Xi Peng. Learning from semantic alignment between unpaired multiviews for egocentric video recognition. In Proceedings of the International Conference on Computer Vision (ICCV), 2023. 1, 2, 3, 6, 7", + "[125] Xiaohan Wang, Linchao Zhu, Heng Wang, and Yi Yang. Interactive prototype learning for egocentric action recognition. In Proceedings of the International Conference on Computer Vision (ICCV), 2021. 3", + "[126] Xin Wang, Taein Kwon, Mahdi Rad, Bowen Pan, Ishani Chakraborty, Sean Andrist, Dan Bohus, Ashley Feniello, Bugra Tekin, Felipe Vieira Frujeri, Neel Joshi, and Marc Pollefeys. Holoassist: an egocentric human interaction dataset for interactive ai assistants in the real world. In Proceedings of the International Conference on Computer Vision (ICCV), 2023. 1, 4, 5", + "[127] Xiaohan Wang, Yuehu Liu, Xinhang Song, Beibei Wang, and Shuqiang Jiang. Generating explanations for embodied action decision from visual observation. In Proceedings of the 31st ACM International Conference on Multimedia, 2023. 5", + "[128] Yeping Wang, Gopika Ajaykumar, and Chien-Ming Huang. See what i see: Enabling user-centric robotic assistance using first-person demonstrations. In Proceedings of the 2020 ACM/IEEE International Conference on Human-Robot Interaction, pages 639-648, 2020. 1", + "[129] Yi Wang, Kunchang Li, Yizhuo Li, Yinan He, Bingkun Huang, Zhiyu Zhao, Hongjie Zhang, Jilan Xu, Yi Liu, Zun" + ], + "bbox": [ + 503, + 92, + 890, + 900 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "22084", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 12 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Wang, et al. Intervideo: General video foundation models via generative and discriminative learning. arXiv preprint arXiv:2212.03191, 2022. 6", + "[130] Zirui Wang, Jiahui Yu, Adams Wei Yu, Zihang Dai, Yulia Tsvetkov, and Yuan Cao. Simvlm: Simple visual language model pretraining with weak supervision. arXiv preprint arXiv:2108.10904, 2021. 6", + "[131] Pengfei Wei, Lingdong Kong, Xinghua Qu, Yi Ren, Jing Jiang, Xiang Yin, et al. Unsupervised video domain adaptation for action recognition: A disentanglement perspective. In Proceedings of the Advances in Neural Information Processing Systems (NeurIPS), 2023. 2, 7", + "[132] Daniel Weinland, Mustafa Özuysal, and Pascal Fua. Making action recognition robust to occlusions and viewpoint changes. In Proceedings of the European Conference on Computer Vision (ECCV), 2010. 3", + "[133] Mark Wilson, John Mcgrath, Samuel Vine, James Brewer, David Defriend, and Richard Masters. Psychomotor control in a virtual laparoscopic surgery training environment: gaze control parameters differentiate novices from experts. Surgical Endoscopy, 24(10):2458-2464, 2010. 8", + "[134] Benita Wong, Joya Chen, You Wu, Stan Weixian Lei, Dongxing Mao, Difei Gao, and Mike Zheng Shou. Assistq: Affordance-centric question-driven task completion for egocentric assistant. In Proceedings of the European Conference on Computer Vision (ECCV), 2022. 1", + "[135] Xinxiao Wu, Han Wang, Cuiwei Liu, and Yunde Jia. Cross-view action recognition over heterogeneous feature spaces. In Proceedings of the International Conference on Computer Vision (ICCV), 2013. 6", + "[136] Haifeng Xia, Pu Wang, and Zhengming Ding. Incomplete multi-view domain adaptation via channel enhancement and knowledge transfer. In Proceedings of the European Conference on Computer Vision (ECCV), 2022. 3", + "[137] Jianjia Xin, Lichun Wang, Kai Xu, Chao Yang, and Baocai Yin. Learning interaction regions and motion trajectories simultaneously from egocentric demonstration videos. IEEE Robotics and Automation Letters, 2023. 3", + "[138] Hu Xu, Gargi Ghosh, Po-Yao Huang, Dmytro Okhonko, Armen Aghajanyan, Florian Metze, Luke Zettlemoyer, and Christoph Feichtenhofer. Videoclip: Contrastive pre-training for zero-shot video-text understanding. arXiv preprint arXiv:2109.14084, 2021. 6", + "[139] Jia Xu, Lopamudra Mukherjee, Yin Li, Jamieson Warner, James M Rehg, and Vikas Singh. Gaze-enabled egocentric video summarization via constrained submodular maximization. In Proceedings of the Conference on Computer Vision and Pattern Recognition (CVPR), 2015. 3", + "[140] Jilan Xu, Yifei Huang, Junlin Hou, Guo Chen, Yuejie Zhang, Rui Feng, and Weidi Xie. Retrieval-augmented egocentric video captioning. arXiv preprint arXiv:2401.00789, 2024.3", + "[141] Mingze Xu, Chenyou Fan, Yuchen Wang, Michael S Ryoo, and David J Crandall. Joint person segmentation and identification in synchronized first-and third-person videos. In Proceedings of the European Conference on Computer Vision (ECCV), 2018. 6", + "[142] Hongwei Xue, Tiankai Hang, Yanhong Zeng, Yuchong Sun, Bei Liu, Huan Yang, Jianlong Fu, and Baining Guo." + ], + "bbox": [ + 78, + 90, + 470, + 901 + ], + "page_idx": 13 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Advancing high-resolution video-language representation with large-scale video transcriptions. In Proceedings of the Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 3", + "[143] Zihui Xue and Kristen Grauman. Learning fine-grained view-invariant representations from unpaired ego-exo videos via temporal alignment. In Proceedings of the Advances in Neural Information Processing Systems (NeurIPS), 2023. 1, 2, 6, 7", + "[144] Takuma Yagi, Misaki Ohashi, Yifei Huang, Ryosuke Furuta, Shungo Adachi, Toutai Mitsuyama, and Yoichi Sato. Finebio: A fine-grained video dataset of biological experiments with hierarchical annotations. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), 2023. 2", + "[145] Lijin Yang, Yifei Huang, Yusuke Sugano, and Yoichi Sato. Interact before align: Leveraging cross-modal knowledge for domain adaptive action recognition. In Proceedings of the Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 7", + "[146] Sherry Yang, Jacob Walker, Jack Parker-Holder, Yilun Du, Jake Bruce, Andre Barreto, Pieter Abbeel, and Dale Schuurmans. Video as the new language for real-world decision making. arXiv preprint arXiv:2402.17139, 2024. 5", + "[147] Li Yingzhen and Stephan Mandt. Disentangled sequential autoencoder. In Proceedings of the International Conference on Machine Learning (ICML), 2018. 2", + "[148] Tianhe Yu, Chelsea Finn, Annie Xie, Sudeep Dasari, Tianhao Zhang, Pieter Abbeel, and Sergey Levine. One-shot imitation from observing humans via domain-adaptive meta-learning. arXiv preprint arXiv:1802.01557, 2018. 1", + "[149] Xumin Yu, Yongming Rao, Wenliang Zhao, Jiwen Lu, and Jie Zhou. Group-aware contrastive regression for action quality assessment. In Proceedings of the International Conference on Computer Vision (ICCV), 2021. 8", + "[150] Zecheng Yu, Yifei Huang, Ryosuke Furuta, Takuma Yagi, Yusuke Goutsu, and Yoichi Sato. Fine-grained affordance annotation for egocentric hand-object interaction videos. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, 2023. 2", + "[151] Thorsten O Zander, Matti Gaertner, Christian Kothe, and Roman Vilimek. Combining eye gaze input with a brain-computer interface for touchless human-computer interaction. International Journal of Human-Computer Interaction, 27(1):38-51, 2010. 3", + "[152] Rowan Zellers, Ximing Lu, Jack Hessel, Youngjae Yu, Jae Sung Park, Jize Cao, Ali Farhadi, and Yejin Choi. Merlot: Multimodal neural script knowledge models. In Proceedings of the Advances in Neural Information Processing Systems (NeurIPS), 2021. 3", + "[153] Lingzhi Zhang, Shenghao Zhou, Simon Stent, and Jianbo Shi. Fine-grained egocentric hand-object segmentation: Dataset, model, and applications. In Proceedings of the European Conference on Computer Vision (ECCV), 2022. 3", + "[154] Siwei Zhang, Qianli Ma, Yan Zhang, Zhiyin Qian, Taein Kwon, Marc Pollefeys, Federica Bogo, and Siyu Tang. Ego-body: Human body shape and motion of interacting people" + ], + "bbox": [ + 501, + 92, + 893, + 901 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "22085", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 13 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "from head-mounted devices. In Proceedings of the European Conference on Computer Vision (ECCV), 2022. 2, 3", + "[155] Weiyu Zhang, Menglong Zhu, and KG Derpanis. From actemes to action: A strongly supervised representation for detailed action understanding. In Proceedings of the International Conference on Computer Vision (ICCV), 2013.", + "[156] Zehua Zhang, David Crandall, Michael Proulx, Sachin Talathi, and Abhishek Sharma. Can gaze inform egocentric action recognition? In Proceedings of the Symposium on Eye Tracking Research and Applications (ETRA), 2022. 3", + "[157] Zhenning Zhang, Zhigeng Pan, Weiqing Li, and Zhiyong Su. X-board: an egocentric adaptive ar assistant for perception in indoor environments. Virtual Reality, 27(2):1327-1343, 2023. 5", + "[158] Yue Zhao, Ishan Misra, Philipp Krahenbuhl, and Rohit Girdhar. Learning video representations from large language models. In Proceedings of the Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 6", + "[159] Jingjing Zheng and Zhuolin Jiang. Learning view-invariant sparse representations for cross-view action recognition. In Proceedings of the International Conference on Computer Vision (ICCV), 2013. 7", + "[160] Yang Zheng, Yanchao Yang, Kaichun Mo, Jiaman Li, Tao Yu, Yebin Liu, Karen Liu, and Leonidas J Guibas. Gimo: Gaze-informed human motion prediction in context. In Proceedings of the European Conference on Computer Vision (ECCV), 2022. 3" + ], + "bbox": [ + 78, + 90, + 470, + 486 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "22086", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 14 + } +] \ No newline at end of file diff --git a/2024/EgoExoLearn_ A Dataset for Bridging Asynchronous Ego- and Exo-centric View of Procedural Activities in Real World/134e2089-370f-4ad6-9686-84a1d7af0ed4_model.json b/2024/EgoExoLearn_ A Dataset for Bridging Asynchronous Ego- and Exo-centric View of Procedural Activities in Real World/134e2089-370f-4ad6-9686-84a1d7af0ed4_model.json new file mode 100644 index 0000000000000000000000000000000000000000..7c5ff58d6a48d8f29d416b5935987a663f99194f --- /dev/null +++ b/2024/EgoExoLearn_ A Dataset for Bridging Asynchronous Ego- and Exo-centric View of Procedural Activities in Real World/134e2089-370f-4ad6-9686-84a1d7af0ed4_model.json @@ -0,0 +1,3343 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.0, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.155, + 0.131, + 0.817, + 0.177 + ], + "angle": 0, + "content": "EgoExoLearn: A Dataset for Bridging Asynchronous Ego- and Exo-centric View of Procedural Activities in Real World" + }, + { + "type": "text", + "bbox": [ + 0.175, + 0.203, + 0.796, + 0.277 + ], + "angle": 0, + "content": "Yifei Huang†‡, Guo Chen‡, Jilan Xu‡, Mingfang Zhang‡, Lijin Yang, Baoqi Pei Hongjie Zhang, Lu Dong, Yali Wang*, Limin Wang*, Yu Qiao* OpenGVLab, Shanghai AI Laboratory Shenzhen Institutes of Advanced Technology, CAS Nanjing University" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.31, + 0.314, + 0.326 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.33, + 0.474, + 0.692 + ], + "angle": 0, + "content": "Being able to map the activities of others into one's own point of view is a fundamental human skill even from a very early age. Taking a step toward understanding this human ability, we introduce EgoExoLearn, a large-scale dataset that emulates the human demonstration following process, in which individuals record egocentric videos as they execute tasks guided by exocentric-view demonstration videos. Focusing on the potential applications in daily assistance and professional support, EgoExoLearn contains egocentric and demonstration video data spanning 120 hours captured in daily life scenarios and specialized laboratories. Along with the videos we record high-quality gaze data and provide detailed multimodal annotations, formulating a playground for modeling the human ability to bridge asynchronous procedural actions from different viewpoints. To this end, we present benchmarks such as cross-view association, cross-view action planning, and cross-view referenced skill assessment, along with detailed analysis. We expect EgoExoLearn can serve as an important resource for bridging the actions across views, thus paving the way for creating AI agents capable of seamlessly learning by observing humans in the real world. The dataset and benchmark codes are available at https://github.com/OpenGVLab/EgoExoLearn." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.711, + 0.21, + 0.727 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.737, + 0.471, + 0.858 + ], + "angle": 0, + "content": "Even as a child, humans can observe the actions of others and then map them to their own view [4, 34, 100, 105]. With this ability to asynchronously bridge activities from egocentric and exocentric views [95, 103], humans can watch others' demonstrations and replicate the procedures in a new environment. This ability is especially beneficial when actual physical trials carry the potential of high costs [24], e.g., conducting dangerous chemical experiments." + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.86, + 0.47, + 0.875 + ], + "angle": 0, + "content": "In the wake of recent advancements in AI systems, one" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.312, + 0.895, + 0.522 + ], + "angle": 0, + "content": "goal for the next generation of AI agents is to perform tasks in a more embodied setting [94]. However, different from humans, training these AI agents usually requires demonstration videos taken in a similar environment [75, 119] and from a congruent perspective with the AI agents, (e.g., the egocentric point of view [43, 59, 108, 134]). While great effort has been made into the collection of egocentric data in different scenarios [18, 30, 106], it remains crucial for the AI agents to directly learn from demonstration videos taken in a different place and from a different viewpoint [35, 148]. Realizing this capability can unleash the full potential of public instructional video data [82] and is also useful in the human-robot cooperation scenario, especially in novel environments [57, 77, 128]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.523, + 0.896, + 0.765 + ], + "angle": 0, + "content": "Current works towards this goal can be roughly divided into two directions. One way is to learn models in simulated environments [11, 64, 79, 89, 92], but it remains difficult for models in this setting to generalize in the real world [126]. The other direction is to learn from human activity in real-world scenarios. However, attempts to directly combine existing multiview datasets often yield datasets of inferior quality or scale [124, 143]. Meanwhile, the few existing datasets in this direction [99, 106, 109] only record ego- and exo-view videos in the same environment and in a time-synchronized manner. In reality, when following demonstrations it is often needed to bridge a series of procedural actions performed in a different place and at a different time. However, currently, no dataset is available for the exploration of how to bridge asynchronous procedural activities in realistic egocentric and exocentric viewpoints." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.897, + 0.903 + ], + "angle": 0, + "content": "To address this lack of dataset issue, we introduce EgoExoLearn, a large-scale dataset containing demonstration videos and corresponding egocentric videos where the camera wearers follow the demonstrations and perform the same task in a different environment, as shown in Figure 1. Targeting two potential applications, i.e., daily assistance and professional support, EgoExoLearn consists of 747 video sequences spanning a total of 120 hours of footage, ranging from daily food-making to specialized laboratory experi" + }, + { + "type": "page_footnote", + "bbox": [ + 0.078, + 0.887, + 0.428, + 0.902 + ], + "angle": 0, + "content": "*Corresponding authors. †Project lead. ‡Equal key contributions." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "22072" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.084, + 0.074, + 0.891, + 0.278 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.29, + 0.897, + 0.348 + ], + "angle": 0, + "content": "Figure 1. EgoExoLearn emulates the human asynchronous demonstration following process. It contains demonstration videos of multiple tasks, together with egocentric videos recorded by participants replicating the procedure after watching the demonstrations. The dataset comprises gaze signals and fine-grained multi-level multi-modal annotations, enabling the exploration of key features in this context such as cross-view association and cross-view action planning." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.36, + 0.47, + 0.42 + ], + "angle": 0, + "content": "ments. Notably, the egocentric videos in EgoExoLearn contain eye gaze signals showing humans' visual attention while performing the task. This provides a valuable cue for better bridging the actions in ego- and exo-viewpoints." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.421, + 0.474, + 0.6 + ], + "angle": 0, + "content": "We take one more step forward by analyzing human ability in bridging asynchronous ego- and exo-view actions and, accordingly, introduce new tasks and benchmarks that we believe can form building blocks for the development of next-stage embodied AI agents with similar abilities. When humans perform an action, he/she can associate and describe the undergoing action in the egocentric view with the corresponding action in the demonstration. With the knowledge from demonstration videos, humans can know the needed action steps and predict what the next steps should be. Besides, through the comparison with the demonstration, humans can also assess their level of skills." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.602, + 0.472, + 0.782 + ], + "angle": 0, + "content": "Based on the above analysis, we design benchmarks of 1) cross-view association, 2) cross-view action understanding, 3) cross-view referenced skill assessment, and 4) cross-view referenced video captioning. Each benchmark is meticulously defined, annotated, and supported by baseline implementations. In addition, we pioneeringly explore the role of gaze in these tasks. We hope our dataset can provide resources for future work for bridging asynchronous procedural actions in ego- and exo-centric perspectives, thereby inspiring the design of AI agents adept at learning from real-world human demonstrations and mapping the procedural actions into robot-centric views." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.797, + 0.22, + 0.813 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.826, + 0.473, + 0.903 + ], + "angle": 0, + "content": "Ego-exo datasets. While there exist works that associate existing datasets to explore how activities can be bridged between them, these associated datasets are often limited in scale [14, 144, 147] or quality [124], meanwhile focusing only on single actions captured from the same view [81, 88," + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.36, + 0.896, + 0.542 + ], + "angle": 0, + "content": "131]. As for actions from different views, apart from multiview fixed camera datasets [6, 15, 54, 61], there also exist datasets with both ego- exo-centric view videos [30, 31, 99, 106, 109]. These datasets are either recorded in the same environment [45, 99, 109] or record time-synced multi-view videos in the same environment with primary focuses on pose/activity understanding grounded in the 3D world [55, 106, 154]. Our dataset offers a more challenging and realistic scenario, where egocentric camera wearers learn to complete the tasks demonstrated by exocentric demonstration videos. This setting complements these datasets by focusing on high-level procedural actions." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.546, + 0.897, + 0.773 + ], + "angle": 0, + "content": "The only dataset conceptually similar to ours is the recently proposed AE2 dataset [143], where the goal is to learn view-invariant representation from unpaired ego and exo videos. This dataset combines ego and exo videos from five public datasets [18, 20, 53, 55, 155] and a newly collected ego tennis forehand dataset. However, due to the difficulty in associating existing ego-exo datasets, the AE2 dataset is relatively small where the largest subset contains only 322 clips. Also, this dataset only focuses on clip-level actions, and thus cannot feature the real-world demonstration following setting, which usually requires multimodal, task-centric procedural knowledge. Instead, our EgoExoLearn is much larger in scale (100x more clips), while offering gaze and fine-grained multimodal annotations facilitating multi-faceted analysis of ego-exo action understanding." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.781, + 0.897, + 0.903 + ], + "angle": 0, + "content": "Egocentric video datasets. In line with the recent development in wearable cameras [110], multiple egocentric video datasets [5, 17, 19, 30, 46, 78, 98, 112, 150] have been proposed. Different from previous egocentric datasets, the egocentric videos in EgoExoLearn feature a demonstration-following setting. We believe EgoExoLearn provides a playground for developing tools to bridge asynchronous procedural activities from ego- and exocentric viewpoints." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "22073" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.471, + 0.137 + ], + "angle": 0, + "content": "The setting of our EgoExoLearn complements existing datasets like Ego4D and can benefit from their rich knowledge and representations." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.142, + 0.472, + 0.337 + ], + "angle": 0, + "content": "Egocentric gaze. Gaze can indicate visual attention and contains valuable information about human intent [36, 156], thus is used in a diverse range of areas such as human-computer interaction [42, 51, 151], and augmented reality [90, 102]. In computer vision, efforts have been made to leverage gaze in various tasks [37-39, 56, 65-67, 83, 101, 139]. However, with the previous absence of large-scale egocentric datasets that include gaze, this avenue of research is currently under-explored [44, 68, 94, 154, 160]. Our EgoExoLearn offers calibrated gaze positions for all egocentric videos. Thanks to our unique setting, our dataset enables the integration of gaze in egocentric video understanding and the exploration of the role of gaze in the cross-view context." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.343, + 0.473, + 0.584 + ], + "angle": 0, + "content": "Egocentric and ego-exo video understanding. The unique recording perspective of egocentric videos presents a series of challenges including but not limited to action understanding [13, 26, 27, 40, 50, 93, 96, 113, 120, 121, 125], hand detection [12, 29, 107, 153], and video-language understanding [41, 48]. These form fundamental building block techniques of embodied AI [85], VR/AR [9, 47, 74, 123], and human-robot interaction [63, 80, 86, 137]. Since most egocentric datasets are smaller in scale compared with general datasets which contain mostly exocentric view videos [3, 142, 152], it is possible to leverage exocentric video data to improve model performance on egocentric videos [132]. There are typically three main directions: joint view-invariant learning [124, 140, 142], domain adaptation [136], and knowledge distillation [69]. In this work, we evaluate all these directions in our benchmarks." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.601, + 0.165, + 0.616 + ], + "angle": 0, + "content": "3. Dataset" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.627, + 0.233, + 0.641 + ], + "angle": 0, + "content": "3.1. Data Collection" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.655, + 0.472, + 0.821 + ], + "angle": 0, + "content": "Scenarios and tasks. We consider procedural goal-oriented tasks ranging from daily food-making to specialized laboratory-based experiments. This selection is grounded in their exemplification of two prospective areas where future embodied AI agents would need the ability to bridge ego-exo activities: daily-life assistance and professional support. Specifically, EgoExoLearn incorporates 5 types of daily tasks (e.g., cooking) and 3 types of specialized laboratory tasks (e.g., solid-phase peptide synthesis). We record egocentric videos in 4 different kitchens and 3 different labs. Other details are provided in the supplementary." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.826, + 0.472, + 0.902 + ], + "angle": 0, + "content": "Data collection procedure. Before the start of each collection session, participants are required to complete a questionnaire gathering basic demographic information and their self-evaluated expertise in executing the designated task. This questionnaire also highlights the ethical, privacy, and secu" + }, + { + "type": "image", + "bbox": [ + 0.501, + 0.092, + 0.701, + 0.269 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.709, + 0.092, + 0.892, + 0.266 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.278, + 0.892, + 0.362 + ], + "angle": 0, + "content": "Figure 2. The number of videos per task (left) and the average duration of each video per task (right). Task1 to Task5 represent the 5 daily tasks and the remaining are three tasks in specialized laboratories. In each recording session of the egocentric video, one participant may learn from multiple demonstration videos and one demonstration video may be watched by several participants." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.377, + 0.895, + 0.498 + ], + "angle": 0, + "content": "rity considerations. Then in each session, participants will be asked to choose one or several exocentric view demonstration videos from a provided list and carefully learn the detailed procedures. Once they feel ready, they will wear Pupil Invisible Glasses [49], complete the gaze calibration, and begin to replicate the task performed in the demonstration videos. While not encouraged, participants are permitted to revisit the demonstration video during the recording." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.499, + 0.895, + 0.588 + ], + "angle": 0, + "content": "After each recording session, the participants are asked to re-do the gaze calibration to ensure gaze data fidelity. For the 5 daily tasks, the exocentric demonstration videos are manually curated from online video platforms such as YouTube. For the lab experiments, the exocentric demonstration videos are tutorials recorded by senior lab members." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.59, + 0.894, + 0.725 + ], + "angle": 0, + "content": "Figure 2 shows the distribution of the 120 hours of data. Since most demonstration videos are meticulously edited to remove repeated steps, the average length of demonstration videos is lower than the egocentric videos which record the full procedure. As a result, the EgoExoLearn contains 432 egocentric videos totaling 96.5 hours and 315 demonstration videos spanning 23.5 hours. This difference in video length poses a unique challenge when bridging ego- and exo-view activities for future research endeavors." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.737, + 0.625, + 0.751 + ], + "angle": 0, + "content": "3.2. Annotation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.761, + 0.894, + 0.866 + ], + "angle": 0, + "content": "To facilitate our dataset in the development of algorithms that can effectively bridge the gap between ego and exo viewpoints, we provide detailed multi-modal human annotations. Our pipeline of annotation contains four stages detailed in the following paragraphs. Each of these stages is subject to a rigorous manual quality check involving no fewer than two individuals for verification and validation." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.892, + 0.9 + ], + "angle": 0, + "content": "Coarse-level language annotation. In this step, we ask annotators to annotate the coarse actions in the videos. Like" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "22074" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.078, + 0.088, + 0.484, + 0.214 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.489, + 0.089, + 0.866, + 0.101 + ], + "angle": 0, + "content": "Number of nouns (left axis) and noun durations (right axis) of right hand" + }, + { + "type": "image", + "bbox": [ + 0.489, + 0.098, + 0.893, + 0.216 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.126, + 0.219, + 0.842, + 0.233 + ], + "angle": 0, + "content": "Figure 3. Occurrence and duration distribution of the annotated fine-level verbs and nouns associated with the right hand." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.247, + 0.474, + 0.459 + ], + "angle": 0, + "content": "the previous works [106, 126], the coarse actions are defined as a middle-level step for accomplishing a task and can be divided into multiple fine actions. For instance, \"Prepare the pork\" in the task of twice-cooked pork, and \"Suction filtration\" in solid-phase peptide synthesis. Three types of annotation are given in this step: 1) the temporal interval, consisting of start and end timestamps; 2) the action label; and 3) a language description of the video within the annotated interval. We specifically request annotators to focus on elucidating \"what is done,\" \"how it is done,\" and \"the purpose of this step\" in their language descriptions. We define a total of 39 categories of coarse-level actions in this stage and acquire 41.2 coarse-level annotations per video with an average length of 21.5 seconds." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.468, + 0.471, + 0.606 + ], + "angle": 0, + "content": "Fine-level language annotation. Based on the coarse-level annotations, in this step we request the annotators to provide annotations for the fine-level actions. The fine-level actions are the atomic actions like \"take knife\" or \"pull syringe plunger\". Unlike the first step, annotators are instructed to furnish language descriptions that specifically emphasize \"which hand is used\", \"what object is used\" and \"why it is used\". For the first two steps, we employ a two-round manual annotation checking to ensure the annotation quality." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.614, + 0.474, + 0.903 + ], + "angle": 0, + "content": "Translation & parsing. To ensure linguistic precision, all the annotators give the language description using their native language [18]. For non-English annotations, we employ ChatGPT and Google Translation API to translate them into English. Subsequently, for the fine-level annotations, we employ specific rules and utilize tools such as NLTK [10] and Spacy [116] to extract the verbs and nouns associated with each segment. During this stage, we confine the selection of verbs and nouns to the predefined taxonomy offered by Ego4D [30], while also manually introducing supplementary verbs and nouns that are absent in the taxonomy. Since our manual annotation specifies the engagement of the left/right hand, we can extract multiple verbs and nouns for each segment, meanwhile attributing them to the respective involvement of the left or right hand. After manual checking, we obtain a total of 95 verb and 254 noun categories in the fine-level annotation. In Figure 3 we show the occurrence of the top 30 categories of verbs and nouns attributed to the right hand. More statistics can be found in the supplementary." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.266, + 0.896, + 0.419 + ], + "angle": 0, + "content": "Skill level annotation. Since self-assessed skill level is not perfectly suitable for skill assessment, we identify several representative skills and assign human annotators to assess their skills. The annotation follows a pairwise ranking scheme, where annotators are presented with pairs of videos of the same action, and instructed to determine which video demonstrates a higher skill level. We prepare 40,191 video pairs and ensure that 4 different annotators annotate each pair. After filtering out pairs with less than 3 consistent opinions, we get a collection of 34,239 valid video pairs." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.428, + 0.733, + 0.444 + ], + "angle": 0, + "content": "3.3. Statistics & Comparisons" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.452, + 0.895, + 0.589 + ], + "angle": 0, + "content": "To the best of our knowledge, there is no dataset that follows the same setting as ours for a direct comparison. Therefore, we enumerate various aspects of our dataset and conduct a comparative analysis with relevant datasets in Tables 1 and 2. EgoExoLearn distinctively enriches the domain with its \"visual demonstration following\" setting. Beyond this unique setting, it stands as the first egocentric dataset that includes temporal bounded language captions, annotated cross-view associations, and multi-label video segments." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.603, + 0.813, + 0.62 + ], + "angle": 0, + "content": "4. Dataset Properties & Benchmarks" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.628, + 0.681, + 0.644 + ], + "angle": 0, + "content": "4.1. Dataset Properties" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.652, + 0.895, + 0.682 + ], + "angle": 0, + "content": "EgoExoLearn stands out from current egocentric and ego-exo datasets due to several unique properties." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.686, + 0.896, + 0.822 + ], + "angle": 0, + "content": "Ego-Exo demonstration following setting. The most distinguished property of EgoExoLearn is the ego-exo demonstration following context. Egocentric video recorders are instructed to follow the steps in exocentric demonstration videos to perform the same task but in a different environment. This setting closely emulates the human observational learning process [4, 34] and can be instrumental in designing embodied AI agents that learn from alternative perspectives while executing tasks from their own viewpoint." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.826, + 0.897, + 0.903 + ], + "angle": 0, + "content": "Fine-grained vision-language annotations with gaze. To facilitate a deeper analysis, we equip our dataset with rich, multimodal, fine-grained annotations. EgoExoLearn is the first egocentric dataset featuring high-quality captions, evidenced by the number of words per segment in Tab. 2." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "22075" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.082, + 0.089, + 0.891, + 0.251 + ], + "angle": 0, + "content": "
DatasetSettingsUnique HoursEgo +Exo?Instruction following?Visual instruction?GazeCoarse ActionFine ActionDense CaptionAssociationSkill
Meccano [98]Industry7X✓XXX✓XXX
EGTEA [67]Cooking28X✓X✓X✓XXX
EK-100 [18]Cooking100XXXXX✓XXX
HoloAssist [126]Assistive166X✓X✓✓✓XX✓
Ego4D [30]Multiple3670XXX◇◇◇◇XX
H2O [55]Desk1✓XXXX✓XXX
LEMMA [45]Daily10✓XXXX✓XXX
HOMEAGE [99]Daily25✓XXX✓✓XXX
CharadesEgo [109]Daily34✓XXXX✓XXX
Assembly101 [55]Desk42✓✓XX✓✓XXX
EgoExoLearn (ours)Daily & Lab120✓✓✓✓✓✓✓✓✓
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.259, + 0.893, + 0.289 + ], + "angle": 0, + "content": "Table 1. Comparison to related datasets on settings (left) and annotations (right). \"Unique hours\" refers to the cumulative duration of distinct video recordings, counting only one camera's footage for simultaneous recordings of the same activity. \\(\\diamond\\): Partially included." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.302, + 0.468, + 0.414 + ], + "angle": 0, + "content": "
Dataset# videosAvg. min# segsAvg. secverb classesnoun classes# verb /seg#noun /seg#words /seg
CharadesEgo [109]78600.569k1.83237110
HOMEAGE [99]57000.926k-2986--0
EK-100 [18]7008.590k3.19730011.23.0
Assembly101 [106]43217.183k1.72490110
Ego4D [30]99126.477k8.0115478117.4
Ours-ego43213.464k4.6952541.82.516.9
Ours-exo3154.514k4.7822512.33.019.4
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.424, + 0.471, + 0.467 + ], + "angle": 0, + "content": "Table 2. Contemporary egocentric datasets. We show only the fine-level actions for a fair comparison. For Ego4D [30], we select the closest subtask of \"forecasting\" following [106]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.487, + 0.472, + 0.653 + ], + "angle": 0, + "content": "Different from Ego4D where captions are associated with only single timestamps, our captions come with manually annotated start and end timestamps. For better visual perception across ego-exo views, we give verb and noun labels associated with the specific hand. One aspect from which we can analyze the human's ability to bridge ego-exo activities is through gaze. EgoExoLearn is also augmented with calibrated eye-gaze signals. These annotations enable the understanding of human ability to bridge ego-exo activities from diverse perspectives, which we posit will benefit the next-generation embodied AI agents [21, 62]." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.666, + 0.212, + 0.68 + ], + "angle": 0, + "content": "4.2. Benchmarks" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.689, + 0.472, + 0.901 + ], + "angle": 0, + "content": "To evaluate the ability of bridging asynchronous ego-exo procedural activities, we introduce 4 new benchmarks: 1) cross-view association, 2) cross-view action understanding, 3) cross-view referenced skill assessment, and 4) cross-view referenced captioning. The cross-view action understanding benchmark is further subdivided into three subtasks: cross-view action anticipation, cross-view action planning, and cross-view action segmentation. Additionally, we explore the role of gaze in assisting these tasks. We also benchmark models on zero-shot and supervised fine-grained action recognition tasks for reference, following [18, 126]. Note that we carefully split our dataset to eliminate annotation leak across benchmarks. Due to the space limit, we only provide partial content of definition, annotation, results, and analysis," + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.306, + 0.844, + 0.322 + ], + "angle": 0, + "content": "leaving more complete details in the supplementary." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.345, + 0.712, + 0.359 + ], + "angle": 0, + "content": "4.2.1 Cross-view association" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.374, + 0.895, + 0.496 + ], + "angle": 0, + "content": "Motivation. One straightforward indicator of the ego-exo activity bridging is the ability to associate the same semantics across ego- and exo-views. This benchmark focuses on equipping models with this cross-view association ability. An application of this ability is assistants in AR that can show expert demonstration videos when the human is confused [157]. Another potential application is the embodied AI agent that can explain its decision [118, 127, 146]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.5, + 0.895, + 0.651 + ], + "angle": 0, + "content": "Problem settings. We formulate this association benchmark as a cross-view multiple-choice association problem. Specifically, we consider two different cross-view association settings: ego2exo and exo2ego. In the case of ego2exo, given an egocentric video, the model is asked to predict the corresponding exocentric video performing the same action from a candidate choice set of exocentric samples, and vice versa for the exo2ego setting. For both ego2exo and exo2ego settings, we use 20 candidate samples for each query. The evaluation metric is the averaged Top-1 accuracy." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.655, + 0.895, + 0.791 + ], + "angle": 0, + "content": "Annotations. We meticulously construct the ground-truth ego-exo pairs via a semantic-aware matching process. It is composed of five stages with details in the supplementary material: (1) Scenario Matching. (2) Noun and Verb Matching. (3) Sentence Matching with LLM. (4) Negative Sampling. (5) Two-round Manual Verification. Notably, we do not provide such pairs for the training set and leave the modeling of cross-view association on unpaired samples to be further explored for the community." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.795, + 0.897, + 0.901 + ], + "angle": 0, + "content": "Baseline model. We adopt three types of baseline models, i.e., ego-only, exo-only, and ego-exo, which refer to the data we used during training. Under all the settings, we leverage the paired video and caption and jointly train a video encoder and a text encoder using the contrastive loss [97]. We use TimeSformer-B [8] as the video encoder and clip-text [97] as the text encoder. We initialize both encoders" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "22076" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.084, + 0.092, + 0.362, + 0.216 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.37, + 0.092, + 0.63, + 0.216 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.64, + 0.092, + 0.889, + 0.216 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.223, + 0.893, + 0.252 + ], + "angle": 0, + "content": "Figure 4. Concept of the 3 benchmarks of cross-view association (Sec. 4.2.1), cross-view action anticipation & planning (Sec. 4.2.2) and cross-view reference skill assessment (Sec.4.2.3) in this section. Other benchmarks can be found in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.266, + 0.47, + 0.401 + ], + "angle": 0, + "content": "using the EgoVLP [72] pre-trained weights. During testing, we obtain the ego/exocentric video representations using the video encoder. The prediction is defined as the one with the highest normalized cross-view video feature similarity among all candidates. On top of these models, we evaluate the effectiveness of gaze in associating egocentric video and exocentric video. This is achieved by replacing the original egocentric video with spatially cropped videos using the gaze positions as the cropping center." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.408, + 0.471, + 0.59 + ], + "angle": 0, + "content": "Related work. Prior works on pre-trained vision-language models for multiple choice association/questioning [60, 72, 122, 130, 138, 158] are generally pre-trained on either exocentric [3, 82] or egocentric [30] datasets, only revealing weak cross-view bridging ability. Another line of works explores cross-view learning [2, 33, 135, 141] by either transferring the knowledge from one view to the other [69] or training view-invariant video understanding models [124, 143]. Different from previous works, our cross-view association benchmark is in a more realistic but challenging setting, by evaluating the model's ability to associate asynchronous activities across ego- and exo-views." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.596, + 0.47, + 0.717 + ], + "angle": 0, + "content": "Experiment results. We first evaluate several vision-language models via zero-shot transfer. These models are pre-trained either on egocentric videos, i.e. EgoVLP [72] and LaViLa [158], or exocentric videos, i.e., InternVideo [129]. As shown in Tab. 3, without using gaze, EgoVLP generally outperforms the others on both validation and test sets. By introducing gaze information, InternVideo receives a decent improvement, especially on Exo2Ego." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.72, + 0.471, + 0.902 + ], + "angle": 0, + "content": "As for fine-tuned models, Tab. 3, reveals that models trained solely on single-view data struggle with cross-view association. Training ego-only models using gaze-cropped egocentric videos results in substantial improvements, significantly outperforming those trained on center-cropped videos. This again highlights the importance of gaze in enhancing cross-view association. Based on our observation, the regions around gaze help the cluttered ego videos become visually similar to exo videos where the primary object is salient. Last, we show the baseline result when co-trained on both egocentric and exocentric videos. The model in this setting shows the strongest cross-view association ability over" + }, + { + "type": "table", + "bbox": [ + 0.505, + 0.262, + 0.89, + 0.495 + ], + "angle": 0, + "content": "
MethodGazeValTest
Ego2ExoExo2EgoEgo2ExoExo2Ego
Zero-shot
Random✘12.715.014.113.4
EgoVLP [72]✘28.827.232.128.9
LaViLa [158]✘22.624.928.725.7
InternVideo [129]✘27.021.230.621.7
EgoVLP [72]✓28.829.731.528.9
LaViLa [158]✓21.921.430.325.9
InternVideo [129]✓30.932.333.332.2
Fine-tuned
Exo-only✘42.941.745.446.9
Ego-only✘33.637.140.335.8
Ego-only✓34.638.745.641.8
Ego-onlyCenter25.422.824.724.2
EgoExo✘42.945.449.045.3
EgoExo✓47.948.855.351.1
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.506, + 0.893, + 0.576 + ], + "angle": 0, + "content": "Table 3. Association accuracy in the cross-view association benchmark. In the fine-tuned setting, we adopt three kinds of data sources for training, i.e., ego-only, exo-only, and hybrid ego-exo data. By leveraging gaze information during training, the model outperforms the baseline (w/o gaze) and the center-crop counterpart." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.591, + 0.894, + 0.651 + ], + "angle": 0, + "content": "single-view models. Findings from this benchmark underline the limitation of current models in associating activities across ego and exo views, and point towards the potential benefits of integrating gaze into the association." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.672, + 0.848, + 0.688 + ], + "angle": 0, + "content": "4.2.2 Cross-view action anticipation & planning" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.7, + 0.894, + 0.88 + ], + "angle": 0, + "content": "Motivation. The procedural actions are not guaranteed to be identical between the two views due to practical constraints. Thus, we design benchmarks for cross-view action anticipation and planning to enable a thorough understanding and transfer of necessary steps (or actions) for task completion, bridging the gap between views and considering real-world conditions. A practical application of this benchmark can be seen in human-robot collaboration scenarios. For instance, an embodied AI agent, after observing a human perform the first part of a task, could effectively take over and complete the remaining half of the task based on the particular environmental situation." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.886, + 0.893, + 0.901 + ], + "angle": 0, + "content": "Problem settings. For both the cross-view anticipation and" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "22077" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.078, + 0.089, + 0.472, + 0.346 + ], + "angle": 0, + "content": "
MethodGazeAnticipation↑Planning↓
Ego-VEgo-NExo-VExo-NEgoExo
Exo-onlyX29.923.640.940.584.776.1
Ego-onlyX33.437.828.917.683.484.5
Ego-only✓40.552.837.637.680.082.6
Ego-onlyCenter33.238.634.132.782.684.7
Unsupervised Domain Adaption
Ego2ExoX33.638.135.428.783.084.1
Exo2EgoX30.423.639.239.883.979.0
Ego2Exo✓40.854.238.737.182.884.3
Exo2Ego✓33.531.339.140.182.478.8
Knowledge Distillation
Ego2ExoX29.624.941.645.284.375.5
Exo2EgoX34.038.428.618.683.184.3
Ego2Exo✓29.925.041.245.184.875.1
Exo2Ego✓41.056.137.739.179.582.6
Co-training
Ego & ExoX33.537.439.644.383.276.0
Ego & Exo✓43.853.340.344.479.075.6
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.357, + 0.472, + 0.414 + ], + "angle": 0, + "content": "Table 4. Results of cross-view action anticipation and planning benchmarks. For anticipation, the class-mean Top-5 recall is used as the evaluation metric (higher is better). For planning, the Edit distance is used as the evaluation metric (lower is better)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.426, + 0.472, + 0.593 + ], + "angle": 0, + "content": "cross-view planning, the goal is to anticipate the future activities in one view, given labeled training data only in another view. For the cross-view anticipation task, we focus on predicting the verb and noun categories of the next fine-level action \\(\\tau = 1\\) second into the future. Given the multilabel nature of our verb and noun annotations in each fine-level segment, we perform multiclass anticipation. Performance is evaluated by class-mean Top-5 recall as per [18]. For the cross-view planning task, we aim to generate the next \\(K = 8\\) steps of coarse-level actions. We adopt ED@ \\(K\\) as the evaluation metric following the setting of Ego4D LTA [30]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.596, + 0.472, + 0.748 + ], + "angle": 0, + "content": "Annotations. For action anticipation, we use the fine-level verb and noun annotations, and then take their intersection between ego and exo videos to constrain them in the same closed set. To evaluate the model more effectively, we further control the long-tail degree of the data by and filter out the tail categories that occur less than \\( \\frac{1}{100} \\) of the highest occurrence category. For the action planning task, we directly adopt the coarse-level action annotations and regard the start timestamp of each segment as one action step. More details can be found in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.474, + 0.902 + ], + "angle": 0, + "content": "Baseline model. We explore three distinct directions to implement cross-view baseline models. The first direction is based on unsupervised domain adaptation (UDA), treating one view as the source domain and the other as the target domain. This method operates within an unsupervised training framework, using labels from the source domain and video data from both domains [16, 52, 84, 104, 111, 131, 145]. We adopt CLIP [97] + TA3N [14] as the baseline model. The second direction entails knowledge distillation (KD) [28, 32, 87], allowing the model trained on one view to learn knowl" + }, + { + "type": "text", + "bbox": [ + 0.497, + 0.092, + 0.895, + 0.213 + ], + "angle": 0, + "content": "edge of the other view, under the assumption that a teacher model of the other view is available. We equip CLIP with a distillation approach based on [69] to transfer knowledge from the teacher model to a newly created student model. The third and most straightforward direction is co-training (CT) using the data from both views to encourage the model to discover correlations between them directly. For using gaze, we also crop the video based on the gaze positions." + }, + { + "type": "text", + "bbox": [ + 0.497, + 0.213, + 0.897, + 0.41 + ], + "angle": 0, + "content": "We comprehensively consider four evaluation settings. The \"Ego-only\" and \"Exo-only\" settings do not involve cross-view understanding, thus we use zero-shot evaluation serving as the references. The Ego2Exo and Exo2Ego settings are the cross-view settings. For UDA, \"Ego2Exo\" is defined as utilizing the egocentric view as the source domain and the exocentric view as the target domain. In the context of KD, \"Ego2Exo\" indicates we initially train a teacher model on egocentric data, followed by the training and distillation of the student model on exocentric data. For CT, we merge both egocentric and exocentric datasets through direct concatenation. We report results on the test set and put the validation set results in the supplementary." + }, + { + "type": "text", + "bbox": [ + 0.497, + 0.412, + 0.895, + 0.52 + ], + "angle": 0, + "content": "Related work. Prior works [25, 54, 106, 117] on multi-view action understanding mainly focus on synchronized multiview videos. Some work studied transferring knowledge [69] from one view to the other or training the view-invariant [124, 143, 159] video models. Our cross-view benchmarks seek to evaluate the ability of models to bridge asynchronous actions across views, which is more challenging yet realistic." + }, + { + "type": "text", + "bbox": [ + 0.497, + 0.522, + 0.897, + 0.825 + ], + "angle": 0, + "content": "Experiment results. Table 4 presents the results of action anticipation and planning on the test set. The first block of results shows a significant performance gap when models trained exclusively on one view are tested on the other view. This underscores the inherent differences in activities captured in the two views. Remarkably, even without relying on any specific cross-view method, the inclusion of gaze information markedly diminishes the disparity between egocentric and exocentric data. Leveraging techniques such as UDA or KD we can see improved performance compared with direct zero-shot inference in the first block. Since CT can utilize both egocentric and exocentric data, it achieves the best performance in all setups. Moreover, from the comparison in all settings using or not using gaze, it is clear that gaze serves as a surprisingly effective signal to mitigate the gap between activities in the two views, although naively designed. These results take the first step in the potential directions for better bridging the cross-view activities, setting a foundation for future advancements in the field of cross-view action understanding." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.843, + 0.822, + 0.858 + ], + "angle": 0, + "content": "4.2.3 Cross-view referenced skill assessment" + }, + { + "type": "text", + "bbox": [ + 0.497, + 0.871, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Motivation and problem setting. We propose a novel task of cross-view referenced skill assessment leveraging the" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "22078" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.334 + ], + "angle": 0, + "content": "unique setting and annotations of our dataset. This task goes beyond the traditional pairwise ranking often used in skill assessment [22, 58] by incorporating an expert demonstration video as a reference point. This demonstration provides a model of the ideal execution of an action, offering a standard against which to compare skill levels. In this task, the input is a pair of egocentric video clips \\( C_{ego1} \\), \\( C_{ego2} \\) of the same action and an exo-view demonstration video clip \\( C_{exo} \\) as the reference, and the output is a choice \\( c \\in \\{ego1, ego2\\} \\) of the egocentric clip that demonstrates a higher skill level. Moreover, we can also assess the skill level by the relation between action and gaze. This benchmark evaluates a model's ability to bridge asynchronous ego-exo dynamics. A practical application is an AR system that helps humans in skill acquisition by providing targeted feedback based on skill level assessment and expert demonstration." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.339, + 0.473, + 0.492 + ], + "angle": 0, + "content": "Annotations. For the cross-view referenced skill assessment benchmark, we concentrate on four types of representative actions, as detailed in Tab. 5. We use the skill level annotations in Sec. 3.2 in this benchmark. For each pair of videos, the annotation is provided by four distinct annotators to minimize subjective bias. We ensure credibility by checking the transitivity of annotations and removing the pairs with less than 3 agreements among 4 annotators. We then append an exo-view demonstration video clip of the same action, forming video triplets as the model input." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.496, + 0.473, + 0.693 + ], + "angle": 0, + "content": "Baseline model. Our baseline model is built upon a pairwise ranking skill assessment model RAAN [23]. Without loss of generality, assuming egol is the video showing a higher skill level, we apply the following approaches to leverage the reference exo-view demonstration video: 1) Triplet loss (TL). The feature distance between \\( C_{exo} \\) and \\( C_{ego1} \\) should be closer to the distance between \\( C_{exo} \\) and \\( C_{ego2} \\). 2) Relation network (RN). Inspired by [114], we employ a relation network that concatenates the features of the ego and exo clips. This network is designed to discern which of the two egocentric video clips bears a closer relation to the demonstration video in terms of skill level. The way to gaze is consistent with the other benchmarks." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.699, + 0.473, + 0.851 + ], + "angle": 0, + "content": "Related work. Several previous works on skill assessment aim to directly regress a score based on professional ratings [1, 71, 73, 91, 115, 149]. We adopt a more general approach of pairwise ranking since no absolute score is available in most real-world skills [7, 70]. Previous works in this direction only use a pair of videos that are in either egocentric view [22] or in exo view [23, 70]. Differently from their works, we explore the cross-view activity bridging ability by examining how demonstration videos from exo view can benefit skill assessment in ego views." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.473, + 0.903 + ], + "angle": 0, + "content": "Experiment results. We first evaluate the performance of previous works under the conventional pairwise setting. In the upper block of Table 5, both methods [22, 23] receive" + }, + { + "type": "table", + "bbox": [ + 0.505, + 0.089, + 0.892, + 0.242 + ], + "angle": 0, + "content": "
MethodGazeEgg CrackingPeelingStir-fryCutting
Ego pairs only
Who's better [22]✘74.7975.9877.5476.85
RAAN [23]✘78.4578.5282.5379.09
Who's better [22]✓77.9176.7079.1377.02
RAAN [23]✓82.0879.3783.9279.36
Ego pairs + Exo
RAAN [23] + RN✘77.9277.0981.5478.26
RAAN [23] + TL✘78.8179.4682.5078.73
RAAN [23] + RN✓82.1479.3283.5179.44
RAAN [23] + TL✓82.1679.5984.0579.29
" + }, + { + "type": "table_caption", + "bbox": [ + 0.498, + 0.251, + 0.895, + 0.308 + ], + "angle": 0, + "content": "Table 5. Ranking accuracy of cross-view referenced skill assessment. In the upper part of the table, only ego video pairs are used, while in the lower part, exo demonstrations are incorporated by \"RN\": relation network and \"TL\": triplet loss." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.322, + 0.897, + 0.504 + ], + "angle": 0, + "content": "a clear performance boost when gaze is used. This aligns with behavioral science findings that experts and novices have different gaze patterns in the same task [76, 133]. With the exocentric demonstration as reference, both the relation network method (RN) and the triplet loss method (TL) can leverage the reference video and bring performance improvement. Further adding gaze we can get the best performance. However, the marginal improvement observed with the inclusion of the exocentric reference suggests that current models may still struggle to fully bridge asynchronous activities across egocentric and exocentric views. There remains ample room for new improvement and innovation." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.52, + 0.62, + 0.536 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.546, + 0.895, + 0.789 + ], + "angle": 0, + "content": "The ability to bridge asynchronous procedural activities in ego- and exo-views is imperative for next-generation embodied AI in executing sophisticated tasks in the real world. As a fundamental step, our EgoExoLearn encompasses a rich collection of egocentric videos, each captured when replicating procedures of exocentric demonstration videos, but performed in different environments and at different times. This realistic setup, combined with our multimodal annotations, allows us to construct 4 novel benchmarks, serving as a versatile platform for investigating how cross-view asynchronous activities can be bridged. EgoExoLearn also enables new research directions e.g., how to better leverage gaze and hand-associated annotations. Results from the benchmarks show weaknesses of current models in bridging ego- and exo-view asynchronous activities, leaving significant room for future work to improve upon." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.789, + 0.892, + 0.845 + ], + "angle": 0, + "content": "Acknowledgement. This work is supported by the National Key R&D Program of China (No.2022ZD0160102) and the Industry Collaboration Projects Grant, Shanghai Committee of Science and Technology, China (No.22YF1461500)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.845, + 0.895, + 0.901 + ], + "angle": 0, + "content": "Key contribution statement: Guo made key contributions to annotation processing and 3 action benchmarks. Jilan made key contributions to the association and caption benchmarks. Mingfang contributed primarily to the skill benchmark." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "22079" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.096, + 0.116, + 0.47, + 0.198 + ], + "angle": 0, + "content": "[1] Narges Ahmidi, Lingling Tao, Shahin Sefati, Yixin Gao, Colin Lea, Benjamin Bejar Haro, Luca Zappella, Sanjeev Khudanpur, René Vidal, and Gregory D Hager. A dataset and benchmarks for segmentation and recognition of gestures in robotic surgery. IEEE Transactions on Biomedical Engineering, 64(9):2025-2041, 2017. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.096, + 0.2, + 0.471, + 0.24 + ], + "angle": 0, + "content": "[2] Shervin Ardeshir and Ali Borji. An exocentric look at egocentric actions and vice versa. Computer Vision and Image Understanding, 171:61-68, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.096, + 0.242, + 0.471, + 0.297 + ], + "angle": 0, + "content": "[3] Max Bain, Arsha Nagrani, Gül Varol, and Andrew Zisserman. Frozen in time: A joint video and image encoder for end-to-end retrieval. In Proceedings of the International Conference on Computer Vision (ICCV), 2021. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.096, + 0.298, + 0.469, + 0.325 + ], + "angle": 0, + "content": "[4] Albert Bandura. Observational learning. The international encyclopedia of communication, 2008. 1, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.096, + 0.327, + 0.47, + 0.381 + ], + "angle": 0, + "content": "[5] Siddhant Bansal, Chetan Arora, and C.V. Jawahar. My view is the best view: Procedure learning from egocentric videos. In Proceedings of the European Conference on Computer Vision (ECCV), 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.096, + 0.383, + 0.471, + 0.465 + ], + "angle": 0, + "content": "[6] Yizhak Ben-Shabat, Xin Yu, Fatemeh Saleh, Dylan Campbell, Cristian Rodriguez-Opazo, Hongdong Li, and Stephen Gould. The ikaa asm dataset: Understanding people assembling furniture through actions, objects and pose. In Proceedings of the Winter Conference on Applications of Computer Vision (WACV), 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.096, + 0.467, + 0.469, + 0.522 + ], + "angle": 0, + "content": "[7] Gedas Bertasius, Hyun Soo Park, Stella X Yu, and Jianbo Shi. Am i a baller? basketball performance assessment from first-person videos. In Proceedings of the International Conference on Computer Vision (ICCV), 2017. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.096, + 0.524, + 0.469, + 0.577 + ], + "angle": 0, + "content": "[8] Gedas Bertasius, Heng Wang, and Lorenzo Torresani. Is space-time attention all you need for video understanding? In Proceedings of the International Conference on Machine Learning (ICML), 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.096, + 0.579, + 0.469, + 0.634 + ], + "angle": 0, + "content": "[9] Vinay Bettadapura, Irfan Essa, and Caroline Pantofaru. Ego-centric field-of-view localization using first-person point-of-view devices. In Proceedings of the Winter Conference on Applications of Computer Vision (WACV), 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.636, + 0.471, + 0.675 + ], + "angle": 0, + "content": "[10] Steven Bird. Nltk: the natural language toolkit. In Proceedings of the COLING/ACL Interactive Presentation Sessions, 2006. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.677, + 0.471, + 0.718 + ], + "angle": 0, + "content": "[11] Greg Brockman, Vicki Cheung, Ludwig Pettersson, Jonas Schneider, John Schulman, Jie Tang, and Wojciech Zaremba. Openai gym. arXiv preprint arXiv:1606.01540, 2016. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.72, + 0.471, + 0.761 + ], + "angle": 0, + "content": "[12] Minjie Cai, Kris M Kitani, and Yoichi Sato. Understanding hand-object manipulation with grasp types and object attributes. In Robotics: Science and Systems, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.762, + 0.471, + 0.83 + ], + "angle": 0, + "content": "[13] Guo Chen, Yifei Huang, Jilan Xu, Baoqi Pei, Zhe Chen, Zhiqi Li, Jiahao Wang, Kunchang Li, Tong Lu, and Limin Wang. Video mamba suite: State space model as a versatile alternative for video understanding. arXiv preprint arXiv:2403.09626, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.832, + 0.471, + 0.9 + ], + "angle": 0, + "content": "[14] Min-Hung Chen, Zsolt Kira, Ghassan AlRegib, Jaekwon Yoo, Ruxin Chen, and Jian Zheng. Temporal attentive alignment for large-scale video domain adaptation. In Proceedings of the International Conference on Computer Vision (ICCV), 2019. 2, 7" + }, + { + "type": "list", + "bbox": [ + 0.088, + 0.116, + 0.471, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.093, + 0.894, + 0.161 + ], + "angle": 0, + "content": "[15] Kellie Corona, Katie Osterdahl, Roderic Collins, and Anthony Hoogs. Meva: A large-scale multiview, multimodal video dataset for activity detection. In Proceedings of the Winter Conference on Applications of Computer Vision (WACV), 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.163, + 0.894, + 0.232 + ], + "angle": 0, + "content": "[16] Victor G Turrisi da Costa, Giacomo Zara, Paolo Rota, Thiago Oliveira-Santos, Nicu Sebe, Vittorio Murino, and Elisa Ricci. Dual-head contrastive domain adaptation for video action recognition. In Proceedings of the Winter Conference on Applications of Computer Vision (WACV), 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.233, + 0.894, + 0.315 + ], + "angle": 0, + "content": "[17] Dima Damen, Hazel Doughty, Giovanni Maria Farinella, Sanja Fidler, Antonino Furnari, Evangelos Kazakos, Davide Moltisanti, Jonathan Munro, Toby Perrett, Will Price, et al. Scaling egocentric vision: The epic-kitchens dataset. In Proceedings of the European Conference on Computer Vision (ECCV), 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.317, + 0.894, + 0.412 + ], + "angle": 0, + "content": "[18] Dima Damen, Hazel Doughty, Giovanni Maria Farinella, Antonino Furnari, Jian Ma, Evangelos Kazakos, Davide Moltisanti, Jonathan Munro, Toby Perrett, Will Price, and Michael Wray. Rescaling egocentric vision: Collection, pipeline and challenges for epic-kitchens-100. International Journal of Computer Vision (IJCV), 130:33-55, 2022. 1, 2, 4, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.414, + 0.893, + 0.495 + ], + "angle": 0, + "content": "[19] Ahmad Darkhalil, Dandan Shan, Bin Zhu, Jian Ma, Amlan Kar, Richard Higgins, Sanja Fidler, David Fouhey, and Dima Damen. Epic-kitchens visor benchmark: Video segmentations and object relations. In Proceedings of the Neural Information Processing Systems (NeurIPS) Track on Datasets and Benchmarks, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.497, + 0.893, + 0.551 + ], + "angle": 0, + "content": "[20] Fernando De la Torre, Jessica Hodgins, Adam Bargteil, Xavier Martin, Justin Macey, Alex Collado, and Pep Beltran. Guide to the carnegie mellon university multimodal activity (cmu-mmac) database. 2009. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.553, + 0.894, + 0.635 + ], + "angle": 0, + "content": "[21] Matt Deitke, Winson Han, Alvaro Herrasti, Aniruddha Kembhavi, Eric Kolve, Roozbeh Mottaghi, Jordi Salvador, Dustin Schwenk, Eli VanderBilt, Matthew Wallingford, et al. Robothor: An open simulation-to-real embodied ai platform. In Proceedings of the Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.637, + 0.894, + 0.691 + ], + "angle": 0, + "content": "[22] Hazel Doughty, Dima Damen, and Walterio Mayol-Cuevas. Who's better? who's best? pairwise deep ranking for skill determination. In Proceedings of the Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.692, + 0.894, + 0.76 + ], + "angle": 0, + "content": "[23] Hazel Doughty, Walterio Mayol-Cuevas, and Dima Damen. The pros and cons: Rank-aware temporal attention for skill determination in long videos. In Proceedings of the Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.762, + 0.893, + 0.803 + ], + "angle": 0, + "content": "[24] Mitch J Fryling, Cristin Johnston, and Linda J Hayes. Understanding observational learning: An interbehavioral approach. The Analysis of verbal behavior, 27:191, 2011. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.804, + 0.893, + 0.858 + ], + "angle": 0, + "content": "[25] Yanwei Fu, Yanwen Guo, Yanshu Zhu, Feng Liu, Chuanming Song, and Zhi-Hua Zhou. Multi-view video summarization. IEEE Transactions on Multimedia, 12(7):717-729, 2010. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.86, + 0.894, + 0.902 + ], + "angle": 0, + "content": "[26] Antonino Furnari and Giovanni Maria Farinella. What would you expect? anticipating egocentric actions with rolling-unrolling lstms and modality attention. In Proceedings of" + }, + { + "type": "list", + "bbox": [ + 0.511, + 0.093, + 0.894, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "22080" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.118, + 0.092, + 0.47, + 0.119 + ], + "angle": 0, + "content": "the International Conference on Computer Vision (ICCV), 2019.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.121, + 0.472, + 0.162 + ], + "angle": 0, + "content": "[27] Rohit Girdhar and Kristen Grauman. Anticipative video transformer. In Proceedings of the International Conference on Computer Vision (ICCV), 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.163, + 0.472, + 0.218 + ], + "angle": 0, + "content": "[28] Rohit Girdhar, Du Tran, Lorenzo Torresani, and Deva Ramanan. Distinct: Learning video representations without a single labeled video. In Proceedings of the International Conference on Computer Vision (ICCV), 2019. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.219, + 0.472, + 0.273 + ], + "angle": 0, + "content": "[29] Mohit Goyal, Sahil Modi, Rishabh Goyal, and Saurabh Gupta. Human hands as probes for interactive object understanding. In Proceedings of the Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.274, + 0.472, + 0.329 + ], + "angle": 0, + "content": "[30] Kristen Grauman, Andrew Westbury, and Eugene Byrne et al. Ego4d: Around the World in 3,000 Hours of Egocentric Video. In Proceedings of the Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 1, 2, 4, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.33, + 0.472, + 0.383 + ], + "angle": 0, + "content": "[31] Kristen Grauman, Andrew Westbury, et al. Ego-exo4d: Understanding skilled human activity from first-and third-person perspectives. arXiv preprint arXiv:2311.18259, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.385, + 0.472, + 0.426 + ], + "angle": 0, + "content": "[32] Geoffrey Hinton, Oriol Vinyals, and Jeff Dean. Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531, 2015. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.428, + 0.472, + 0.482 + ], + "angle": 0, + "content": "[33] Hsuan-I Ho, Wei-Chen Chiu, and Yu-Chiang Frank Wang. Summarizing first-person videos from third persons' points of view. In Proceedings of the European Conference on Computer Vision (ECCV), 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.484, + 0.472, + 0.536 + ], + "angle": 0, + "content": "[34] Nicola J Hodges, A Mark Williams, Spencer J Hayes, and Gavin Breslin. What is modelled during observational learning? Journal of sports sciences, 25(5):531-545, 2007. 1, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.538, + 0.472, + 0.58 + ], + "angle": 0, + "content": "[35] Jiang Hua, Liangcai Zeng, Gongfa Li, and Zhaojie Ju. Learning for a robot: Deep reinforcement learning, imitation learning, transfer learning. Sensors, 21(4):1278, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.581, + 0.472, + 0.622 + ], + "angle": 0, + "content": "[36] Chien-Ming Huang, Sean Andrist, Allison Sauppé, and Bilge Mutlu. Using gaze patterns to predict task intent in collaboration. Frontiers in psychology, 6:1049, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.623, + 0.472, + 0.678 + ], + "angle": 0, + "content": "[37] Yifei Huang, Minjie Cai, Zhenqiang Li, and Yoichi Sato. Predicting gaze in egocentric video by learning task-dependent attention transition. In Proceedings of the European Conference on Computer Vision (ECCV), 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.679, + 0.472, + 0.734 + ], + "angle": 0, + "content": "[38] Yifei Huang, Minjie Cai, Zhenqiang Li, Feng Lu, and Yoichi Sato. Mutual context network for jointly estimating egocentric gaze and action. IEEE Transactions on Image Processing, 29:7795-7806, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.735, + 0.472, + 0.788 + ], + "angle": 0, + "content": "[39] Yifei Huang, Minjie Cai, and Yoichi Sato. An ego-vision system for discovering human joint attention. IEEE Transactions on Human-Machine Systems, 50(4):306-316, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.79, + 0.472, + 0.844 + ], + "angle": 0, + "content": "[40] Yifei Huang, Lijin Yang, and Yoichi Sato. Compound prototype matching for few-shot action recognition. In Proceedings of the European Conference on Computer Vision (ECCV), 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.845, + 0.472, + 0.9 + ], + "angle": 0, + "content": "[41] Yifei Huang, Lijin Yang, and Yoichi Sato. Weakly supervised temporal sentence grounding with uncertainty-guided self-training. In Proceedings of the Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 3" + }, + { + "type": "list", + "bbox": [ + 0.088, + 0.092, + 0.472, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.092, + 0.894, + 0.147 + ], + "angle": 0, + "content": "[42] Thomas E Hutchinson, K Preston White, Worthy N Martin, Kelly C Reichert, and Lisa A Frey. Human-computer interaction using eye-gaze input. IEEE Transactions on systems, man, and cybernetics, 19(6):1527-1534, 1989. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.149, + 0.895, + 0.218 + ], + "angle": 0, + "content": "[43] Eric Jang, Alex Irpan, Mohi Khansari, Daniel Kappler, Frederik Ebert, Corey Lynch, Sergey Levine, and Chelsea Finn. Bc-z: Zero-shot task generalization with robotic imitation learning. In Proceedings of the Conference on Robot Learning, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.219, + 0.895, + 0.287 + ], + "angle": 0, + "content": "[44] Youngkyoon Jang, Brian Sullivan, Casimir Ludwig, Iain Gilchrist, Dima Damen, and Walterio Mayol-Cuevas. Epicent: An egocentric video dataset for camping tent assembly. In Proceedings of the International Conference on Computer Vision Workshops (ICCVW), 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.288, + 0.894, + 0.355 + ], + "angle": 0, + "content": "[45] Baoxiong Jia, Yixin Chen, Siyuan Huang, Yixin Zhu, and Song-Chun Zhu. Lemma: A multi-view dataset for the arning multi-agent multi-task activities. In Proceedings of the European Conference on Computer Vision (ECCV), 2020. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.357, + 0.894, + 0.413 + ], + "angle": 0, + "content": "[46] Baoxiong Jia, Ting Lei, Song-Chun Zhu, and Siyuan Huang. Egotaskqa: Understanding human tasks in egocentric videos. In Proceedings of the Advances in Neural Information Processing Systems (NeurIPS), 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.414, + 0.894, + 0.467 + ], + "angle": 0, + "content": "[47] Hao Jiang and Vamsi Krishna Ithapu. Egocentric pose estimation from human vision span. In Proceedings of the International Conference on Computer Vision (ICCV), 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.469, + 0.894, + 0.525 + ], + "angle": 0, + "content": "[48] Soo-Han Kang and Ji-Hyeong Han. Video captioning based on both egocentric and exocentric views of robot vision for human-robot interaction. International Journal of Social Robotics, pages 1–11, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.526, + 0.894, + 0.594 + ], + "angle": 0, + "content": "[49] Moritz Kassner, William Patera, and Andreas Bulling. Pupil: An open source platform for pervasive eye tracking and mobile gaze-based interaction. In Adjunct Proceedings of the 2014 ACM International Joint Conference on Pervasive and Ubiquitous Computing, 2014. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.595, + 0.894, + 0.661 + ], + "angle": 0, + "content": "[50] Evangelos Kazakos, Arsha Nagrani, Andrew Zisserman, and Dima Damen. Epic-fusion: Audio-visual temporal binding for egocentric action recognition. In Proceedings of the International Conference on Computer Vision (ICCV), 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.664, + 0.894, + 0.733 + ], + "angle": 0, + "content": "[51] Daekyum Kim, Brian Byunghyun Kang, Kyu Bum Kim, Hyungmin Choi, Jeesoo Ha, Kyu-Jin Cho, and Sungho Jo. Eyes are faster than hands: A soft wearable robot learns user intention from the egocentric view. Science Robotics, 4(26), 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.734, + 0.894, + 0.803 + ], + "angle": 0, + "content": "[52] Donghyun Kim, Yi-Hsuan Tsai, Bingbing Zhuang, Xiang Yu, Stan Sclaroff, Kate Saenko, and Manmohan Chandraker. Learning cross-modal contrastive features for video domain adaptation. In Proceedings of the International Conference on Computer Vision (ICCV), 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.804, + 0.894, + 0.871 + ], + "angle": 0, + "content": "[53] Hildegard Kuehne, Hueihan Jhuang, Estíbaliz Garrote, Tomaso Poggio, and Thomas Serre. Hmdb: a large video database for human motion recognition. In Proceedings of the International Conference on Computer Vision (ICCV), 2011. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.873, + 0.894, + 0.902 + ], + "angle": 0, + "content": "[54] Hilde Kuehne, Ali Arslan, and Thomas Serre. The language of actions: Recovering the syntax and semantics of goal" + }, + { + "type": "list", + "bbox": [ + 0.51, + 0.092, + 0.895, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.518, + 0.957 + ], + "angle": 0, + "content": "22081" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.118, + 0.092, + 0.472, + 0.134 + ], + "angle": 0, + "content": "directed human activities. In Proceedings of the Conference on Computer Vision and Pattern Recognition (CVPR), 2014, 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.136, + 0.472, + 0.205 + ], + "angle": 0, + "content": "[55] Taein Kwon, Bugra Tekin, Jan Stühmer, Federica Bogo, and Marc Pollefeys. H2o: Two hands manipulating objects for first person interaction recognition. In Proceedings of the International Conference on Computer Vision (ICCV), 2021. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.207, + 0.472, + 0.263 + ], + "angle": 0, + "content": "[56] Bolin Lai, Miao Liu, Fiona Ryan, and James M Rehg. In the eye of transformer: Global-local correlation for egocentric gaze estimation and beyond. International Journal of Computer Vision (IJCV), pages 1-18, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.264, + 0.472, + 0.332 + ], + "angle": 0, + "content": "[57] Stephane Lallée, Eiichi Yoshida, Anthony Mallet, Francesco Nori, Lorenzo Natale, Giorgio Metta, Felix Warneken, and Peter Ford Dominey. Human-robot cooperation based on interaction learning. From motor learning to interaction learning in robots, pages 491–536, 2010. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.334, + 0.472, + 0.39 + ], + "angle": 0, + "content": "[58] Kyle Lam, Junhong Chen, Zeyu Wang, Fahad M Iqbal, Ara Darzi, Benny Lo, Sanjay Purkayastha, and James M Kinross. Machine learning for technical skill assessment in surgery: a systematic review. NPJ digital medicine, 5(1):24, 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.392, + 0.472, + 0.459 + ], + "angle": 0, + "content": "[59] Jangwon Lee and Michael S Ryoo. Learning robot activities from first-person human videos using convolutional future regression. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.462, + 0.472, + 0.531 + ], + "angle": 0, + "content": "[60] Jie Lei, Linjie Li, Luowei Zhou, Zhe Gan, Tamara L Berg, Mohit Bansal, and Jingjing Liu. Less is more: Clipbert for video-and-language learning via sparse sampling. In Proceedings of the Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.533, + 0.472, + 0.603 + ], + "angle": 0, + "content": "[61] Chengshu Li, Fei Xia, Roberto Martin-Martin, Michael Lingelbach, Sanjana Srivastava, Bokui Shen, Kent Vainio, Cem Gokmen, Gokul Dharan, Tanish Jain, et al. igibson 2.0: Object-centric simulation for robot learning of everyday household tasks. arXiv preprint arXiv:2108.03272, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.604, + 0.472, + 0.687 + ], + "angle": 0, + "content": "[62] Chengshu Li, Ruohan Zhang, Josiah Wong, Cem Gokmen, Sanjana Srivastava, Roberto Martin-Martin, Chen Wang, Gabrael Levine, Michael Lingelbach, Jiankai Sun, et al. Behavior-1k: A benchmark for embodied ai with 1,000 everyday activities and realistic simulation. In Proceedings of the Conference on Robot Learning, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.689, + 0.472, + 0.744 + ], + "angle": 0, + "content": "[63] Haoxin Li, Yijun Cai, and Wei-Shi Zheng. Deep dual relation modeling for egocentric interaction recognition. In Proceedings of the Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.746, + 0.472, + 0.801 + ], + "angle": 0, + "content": "[64] Jiayi Li, Tao Lu, Xiaoge Cao, Yinghao Cai, and Shuo Wang. Meta-imitation learning by watching video demonstrations. In Proceedings of the International Conference on Learning Representations, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.803, + 0.472, + 0.856 + ], + "angle": 0, + "content": "[65] Yin Li, Alireza Fathi, and James M Rehg. Learning to predict gaze in egocentric video. In Proceedings of the International Conference on Computer Vision (ICCV), 2013. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.859, + 0.472, + 0.901 + ], + "angle": 0, + "content": "[66] Yin Li, Zhefan Ye, and James M Rehg. Delving into egocentric actions. In Proceedings of the Conference on Computer Vision and Pattern Recognition (CVPR), 2015." + }, + { + "type": "list", + "bbox": [ + 0.088, + 0.092, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.092, + 0.894, + 0.147 + ], + "angle": 0, + "content": "[67] Yin Li, Miao Liu, and James M Rehg. In the eye of beholder: Joint learning of gaze and actions in first person video. In Proceedings of the European Conference on Computer Vision (ECCV), 2018. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.149, + 0.894, + 0.204 + ], + "angle": 0, + "content": "[68] Yin Li, Miao Liu, and Jame Rehg. In the eye of the beholder: Gaze and actions in first person video. IEEE Transaction on Pattern Analysis and Machine Intelligence (TPAMI), 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.206, + 0.894, + 0.275 + ], + "angle": 0, + "content": "[69] Yanghao Li, Tushar Nagarajan, Bo Xiong, and Kristen Grauman. Ego-exo: Transferring visual representations from third-person to first-person videos. In Proceedings of the Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.278, + 0.894, + 0.332 + ], + "angle": 0, + "content": "[70] Zhenqiang Li, Yifei Huang, Minjie Cai, and Yoichi Sato. Manipulation-skill assessment from videos with spatial attention network. In Proceedings of the International Conference on Computer Vision Workshops (ICCVW), 2019. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.334, + 0.894, + 0.403 + ], + "angle": 0, + "content": "[71] Zhenqiang Li, Lin Gu, Weimin Wang, Ryosuke Nakamura, and Yoichi Sato. Surgical skill assessment via video semantic aggregation. In Proceedings of the International Conference on Medical Image Computing and Computer-Assisted Intervention (MICCAI), 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.405, + 0.894, + 0.473 + ], + "angle": 0, + "content": "[72] Kevin Qinghong Lin, Jinpeng Wang, Mattia Soldan, Michael Wray, Rui Yan, Eric Z XU, Difei Gao, Rong-Cheng Tu, Wenzhe Zhao, Weijie Kong, et al. Egocentric video-language pretraining. In Proceedings of the Advances in Neural Information Processing Systems (NeurIPS), 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.476, + 0.894, + 0.532 + ], + "angle": 0, + "content": "[73] Daochang Liu, Qiyue Li, Tingting Jiang, Yizhou Wang, Rulin Miao, Fei Shan, and Ziyu Li. Towards unified surgical skill assessment. In Proceedings of the Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.533, + 0.894, + 0.602 + ], + "angle": 0, + "content": "[74] Shaowei Liu, Subarna Tripathi, Somdeb Majumdar, and Xiaolong Wang. Joint hand motion and interaction hotspots prediction from egocentric videos. In Proceedings of the Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.604, + 0.894, + 0.646 + ], + "angle": 0, + "content": "[75] Xiaotian Liu, Hector Palacios, and Christian Muise. Egocentric planning for scalable embodied task achievement. arXiv preprint arXiv:2306.01295, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.647, + 0.894, + 0.716 + ], + "angle": 0, + "content": "[76] Yan Liu, Pei Yun Hsueh, Jennifer Lai, Mirweis Sangin, Marc Antoine Nussli, and Pierre Dillenbourg. Who is the expert? analyzing gaze data to predict expertise level in collaborative applications. In Proceedings of the IEEE International Conference on Multimedia & Expo (ICME), 2009. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.718, + 0.894, + 0.772 + ], + "angle": 0, + "content": "[77] Yueyue Liu, Zhijun Li, Huaping Liu, and Zhen Kan. Skill transfer learning for autonomous robots and human-robot cooperation: A survey. Robotics and Autonomous Systems, 128:103515, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.774, + 0.894, + 0.845 + ], + "angle": 0, + "content": "[78] Yunze Liu, Yun Liu, Che Jiang, Kangbo Lyu, Weikang Wan, Hao Shen, Boqiang Liang, Zhoujie Fu, He Wang, and Li Yi. Hoi4d: A 4d egocentric dataset for category-level human-object interaction. In Proceedings of the Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.846, + 0.894, + 0.901 + ], + "angle": 0, + "content": "[79] Manolis Savva*, Abhishek Kadian*, Oleksandr Maksymets*, Yili Zhao, Erik Wijmans, Bhavana Jain, Julian Straub, Jia Liu, Vladlen Koltun, Jitendra Malik, Devi Parikh, and Dhruv Batra. Habitat: A Platform for Embodied AI" + }, + { + "type": "list", + "bbox": [ + 0.51, + 0.092, + 0.894, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "22082" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.118, + 0.092, + 0.469, + 0.12 + ], + "angle": 0, + "content": "Research. In Proceedings of the International Conference on Computer Vision (ICCV), 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.123, + 0.47, + 0.178 + ], + "angle": 0, + "content": "[80] Javier Marina-Miranda and V Javier Traver. Head and eye egocentric gesture recognition for human-robot interaction using eyewear cameras. IEEE Robotics and Automation Letters, 7(3):7067-7074, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.181, + 0.471, + 0.248 + ], + "angle": 0, + "content": "[81] Joanna Materzynska, Guillaume Berger, Ingo Bax, and Roland Memisevic. The jester dataset: A large-scale video dataset of human gestures. In Proceedings of the International Conference on Computer Vision Workshops (ICCVW), 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.253, + 0.471, + 0.335 + ], + "angle": 0, + "content": "[82] Antoine Miech, Dimitri Zhukov, Jean-Baptiste Alayrac, Makarand Tapaswi, Ivan Laptev, and Josef Sivic. HowTo100M: Learning a Text-Video Embedding by Watching Hundred Million Narrated Video Clips. In Proceedings of the International Conference on Computer Vision (ICCV), 2019. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.339, + 0.47, + 0.394 + ], + "angle": 0, + "content": "[83] Kyle Min and Jason J Corso. Integrating human gaze into attention for egocentric activity recognition. In Proceedings of the Winter Conference on Applications of Computer Vision (WACV), pages 1069-1078, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.397, + 0.471, + 0.452 + ], + "angle": 0, + "content": "[84] Jonathan Munro and Dima Damen. Multi-modal domain adaptation for fine-grained action recognition. In Proceedings of the Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.455, + 0.47, + 0.51 + ], + "angle": 0, + "content": "[85] Tushar Nagarajan and Kristen Grauman. Shaping embodied agent behavior with activity-context priors from egocentric video. In Proceedings of the Advances in Neural Information Processing Systems (NeurIPS), 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.513, + 0.471, + 0.568 + ], + "angle": 0, + "content": "[86] Tushar Nagarajan, Yanghao Li, Christoph Feichtenhofer, and Kristen Grauman. Ego-topo: Environment affordances from egocentric video. In Proceedings of the Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.572, + 0.47, + 0.64 + ], + "angle": 0, + "content": "[87] Boxiao Pan, Haoye Cai, De-An Huang, Kuan-Hui Lee, Adrien Gaidon, Ehsan Adeli, and Juan Carlos Niebles. Spatio-temporal graph for video captioning with knowledge distillation. In Proceedings of the Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.644, + 0.47, + 0.699 + ], + "angle": 0, + "content": "[88] Boxiao Pan, Zhangjie Cao, Ehsan Adeli, and Juan Carlos Niebles. Adversarial cross-domain action recognition with co-attention. In Proceedings of the AAAI Conference on Artificial Intelligence (AAAI), 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.702, + 0.471, + 0.783 + ], + "angle": 0, + "content": "[89] Boxiao Pan, Bokui Shen, Davis Rempe, Despoina Paschalidou, Kaichun Mo, Yanchao Yang, and Leonidas J Guibas. Copilot: Human-environment collision prediction and localization from egocentric videos. In Proceedings of the International Conference on Computer Vision (ICCV), 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.787, + 0.471, + 0.842 + ], + "angle": 0, + "content": "[90] Hyung Min Park, Seok Han Lee, and Jong Soo Choi. Wearable augmented reality system using gaze interaction. In Proceedings of the IEEE/ACM International Symposium on Mixed and Augmented Reality, 2008. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.846, + 0.47, + 0.901 + ], + "angle": 0, + "content": "[91] Paritosh Parmar and Brendan Tran Morris. What and how well you performed? a multitask learning approach to action quality assessment. In Proceedings of the Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 8" + }, + { + "type": "list", + "bbox": [ + 0.088, + 0.092, + 0.471, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.093, + 0.892, + 0.133 + ], + "angle": 0, + "content": "[92] Leo Pauly, Wisdom C Agboh, David C Hogg, and Raul Fuentes. O2a: one-shot observational learning with action vectors. Frontiers in Robotics and AI, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.136, + 0.892, + 0.217 + ], + "angle": 0, + "content": "[93] Chiara Plizzari, Mirco Planamente, Gabriele Goletto, Marco Cannici, Emanuele Gusso, Matteo Matteucci, and Barbara Caputo. E2 (go) motion: Motion augmented event stream for egocentric action recognition. In Proceedings of the Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.219, + 0.895, + 0.285 + ], + "angle": 0, + "content": "[94] Chiara Plizzari, Gabriele Goletto, Antonino Furnari, Siddhant Bansal, Francesco Ragusa, Giovanni Maria Farinella, Dima Damen, and Tatiana Tommasi. An outlook into the future of egocentric vision. arXiv preprint arXiv:2308.07123, 2023. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.288, + 0.895, + 0.327 + ], + "angle": 0, + "content": "[95] David Premack and Guy Woodruff. Does the chimpanzee have a theory of mind? Behavioral and brain sciences, 1(4): 515-526, 1978. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.33, + 0.895, + 0.396 + ], + "angle": 0, + "content": "[96] Gorjan Radevski, Dusan Grujicic, Matthew Blaschko, Marie-Francine Moens, and Tinne Tuytelaars. Multimodal distillation for egocentric action recognition. In Proceedings of the International Conference on Computer Vision (ICCV), 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.4, + 0.895, + 0.482 + ], + "angle": 0, + "content": "[97] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In Proceedings of the International Conference on Machine Learning (ICML), 2021. 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.484, + 0.895, + 0.564 + ], + "angle": 0, + "content": "[98] Francesco Ragusa, Antonino Furnari, Salvatore Livatino, and Giovanni Maria Farinella. The meccano dataset: Understanding human-object interactions from egocentric videos in an industrial-like domain. In Proceedings of the Winter Conference on Applications of Computer Vision (WACV), 2021. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.567, + 0.895, + 0.648 + ], + "angle": 0, + "content": "[99] Nishant Rai, Haofeng Chen, Jingwei Ji, Rishi Desai, Kazuki Kozuka, Shun Ishizaka, Ehsan Adeli, and Juan Carlos Niebles. Home action genome: Cooperative compositional action understanding. In Proceedings of the Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 1, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.65, + 0.895, + 0.704 + ], + "angle": 0, + "content": "[100] Richard Ramsey, David M Kaplan, and Emily S Cross. Watch and learn: the cognitive neuroscience of learning from others' actions. Trends in Neurosciences, 44(6):478-491, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.707, + 0.895, + 0.761 + ], + "angle": 0, + "content": "[101] Yosef Razin and Karen Feigh. Learning to predict intent from gaze during robotic hand-eye coordination. In Proceedings of the AAAI Conference on Artificial Intelligence (AAAI), 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.763, + 0.895, + 0.831 + ], + "angle": 0, + "content": "[102] Radiah Rivu, Yasmeen Abdrabou, Ken Pfeuffer, Augusto Esteves, Stefanie Meitner, and Florian Alt. Stare: gaze-assisted face-to-face communication in augmented reality. In Proceedings of the ACM Symposium on Eye Tracking Research and Applications (ETRA), 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.833, + 0.895, + 0.871 + ], + "angle": 0, + "content": "[103] Giacomo Rizzolatti and Laila Craighero. The mirror-neuron system. Annual Review of Neuroscience, 27:169-192, 2004. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.874, + 0.895, + 0.901 + ], + "angle": 0, + "content": "[104] Aadarsh Sahoo, Rutav Shah, Rameswar Panda, Kate Saenko, and Abir Das. Contrast and mix: Temporal contrastive video" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.895, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.518, + 0.957 + ], + "angle": 0, + "content": "22083" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.118, + 0.093, + 0.469, + 0.134 + ], + "angle": 0, + "content": "domain adaptation with background mixing. In Proceedings of the Advances in Neural Information Processing Systems (NeurIPS), 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.136, + 0.469, + 0.175 + ], + "angle": 0, + "content": "[105] Stefan Schaal. Learning from demonstration. In Proceedings of the Advances in Neural Information Processing Systems (NeurIPS), 1996. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.178, + 0.47, + 0.245 + ], + "angle": 0, + "content": "[106] F. Sener, D. Chatterjee, D. Shelepov, K. He, D. Singhania, R. Wang, and A. Yao. Assembly101: A large-scale multiview video dataset for understanding procedural activities. In Proceedings of the Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 1, 2, 4, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.247, + 0.471, + 0.302 + ], + "angle": 0, + "content": "[107] Dandan Shan, Jiaqi Geng, Michelle Shu, and David F Fouhey. Understanding human hands in contact at internet scale. In Proceedings of the Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.303, + 0.471, + 0.384 + ], + "angle": 0, + "content": "[108] Mohit Shridhar, Jesse Thomason, Daniel Gordon, Yonatan Bisk, Winson Han, Roozbeh Mottaghi, Luke Zettlemoyer, and Dieter Fox. ALFRED: A Benchmark for Interpreting Grounded Instructions for Everyday Tasks. In Proceedings of the Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.386, + 0.469, + 0.44 + ], + "angle": 0, + "content": "[109] Gunnar A Sigurdsson, Abhinav Gupta, Cordelia Schmid, Ali Farhadi, and Karteek Alahari. Charades-ego: A large-scale dataset of paired third and first person videos. arXiv preprint arXiv:1804.09626, 2018. 1, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.442, + 0.47, + 0.51 + ], + "angle": 0, + "content": "[110] Kiran Somasundaram, Jing Dong, Huixuan Tang, Julian Straub, Mingfei Yan, Michael Goesele, Jakob Julian Engel, Renzo De Nardi, and Richard Newcombe. Project aria: A new tool for egocentric multi-modal ai research. arXiv preprint arXiv:2308.13561, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.511, + 0.47, + 0.58 + ], + "angle": 0, + "content": "[111] Xiaolin Song, Sicheng Zhao, Jingyu Yang, Huanjing Yue, Pengfei Xu, Runbo Hu, and Hua Chai. Spatio-temporal contrastive domain adaptation for action recognition. In Proceedings of the Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.581, + 0.471, + 0.649 + ], + "angle": 0, + "content": "[112] Yale Song, Eugene Byrne, Tushar Nagarajan, Huiyu Wang, Miguel Martin, and Lorenzo Torresani. Ego4d goal-step: Toward hierarchical understanding of procedural activities. Proceedings of the Advances in Neural Information Processing Systems (NeurIPS), 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.65, + 0.47, + 0.705 + ], + "angle": 0, + "content": "[113] Swathikiran Sudhakaran, Sergio Escalera, and Oswald Lanz. Lsta: Long short-term attention for egocentric action recognition. In Proceedings of the Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.707, + 0.471, + 0.774 + ], + "angle": 0, + "content": "[114] Flood Sung, Yongxin Yang, Li Zhang, Tao Xiang, Philip HS Torr, and Timothy M Hospedales. Learning to compare: Relation network for few-shot learning. In Proceedings of the Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.776, + 0.471, + 0.844 + ], + "angle": 0, + "content": "[115] Yansong Tang, Zanlin Ni, Jiahuan Zhou, Danyang Zhang, Jiwen Lu, Ying Wu, and Jie Zhou. Uncertainty-aware score distribution learning for action quality assessment. In Proceedings of the Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.846, + 0.47, + 0.873 + ], + "angle": 0, + "content": "[116] Yuli Vasiliev. Natural language processing with Python and spaCy: A practical introduction. 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.874, + 0.471, + 0.901 + ], + "angle": 0, + "content": "[117] Shruti Vyas, Yogesh S Rawat, and Mubarak Shah. Multiview action recognition using cross-view video prediction." + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.093, + 0.471, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.541, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "In Proceedings of the European Conference on Computer Vision (ECCV), 2020. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.122, + 0.892, + 0.177 + ], + "angle": 0, + "content": "[118] Sebastian Wallkötter, Silvia Tulli, Ginevra Castellano, Ana Paiva, and Mohamed Chetouani. Explainable embodied agents through social cues: a review. ACM Transactions on Human-Robot Interaction (THRI), 10(3):1-24, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.179, + 0.892, + 0.235 + ], + "angle": 0, + "content": "[119] Chen Wang, Linxi Fan, Jiankai Sun, Ruohan Zhang, Li Fei-Fei, Danfei Xu, Yuke Zhu, and Anima Anandkumar. Mimicplay: Long-horizon imitation learning by watching human play. arXiv preprint arXiv:2302.12422, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.236, + 0.892, + 0.292 + ], + "angle": 0, + "content": "[120] Huiyu Wang, Mitesh Kumar Singh, and Lorenzo Torresani. Ego-only: Egocentric action detection without exocentric transferring. In Proceedings of the International Conference on Computer Vision (ICCV), 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.293, + 0.892, + 0.348 + ], + "angle": 0, + "content": "[121] Jiahao Wang, Guo Chen, Yifei Huang, Limin Wang, and Tong Lu. Memory-and-anticipation transformer for online action understanding. In Proceedings of the International Conference on Computer Vision (ICCV), 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.35, + 0.892, + 0.433 + ], + "angle": 0, + "content": "[122] Jinpeng Wang, Yixiao Ge, Rui Yan, Yuying Ge, Kevin Qinghong Lin, Satoshi Tsutsui, Xudong Lin, Guanyu Cai, Jianping Wu, Ying Shan, et al. All in one: Exploring unified video-language pre-training. In Proceedings of the Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.435, + 0.892, + 0.503 + ], + "angle": 0, + "content": "[123] Jian Wang, Diogo Luvizon, Weipeng Xu, Lingjie Liu, Kripasindhu Sarkar, and Christian Theobalt. Scene-aware egocentric 3d human pose estimation. In Proceedings of the Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.505, + 0.892, + 0.574 + ], + "angle": 0, + "content": "[124] Qitong Wang, Long Zhao, Liangzhe Yuan, Ting Liu, and Xi Peng. Learning from semantic alignment between unpaired multiviews for egocentric video recognition. In Proceedings of the International Conference on Computer Vision (ICCV), 2023. 1, 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.576, + 0.892, + 0.631 + ], + "angle": 0, + "content": "[125] Xiaohan Wang, Linchao Zhu, Heng Wang, and Yi Yang. Interactive prototype learning for egocentric action recognition. In Proceedings of the International Conference on Computer Vision (ICCV), 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.633, + 0.892, + 0.73 + ], + "angle": 0, + "content": "[126] Xin Wang, Taein Kwon, Mahdi Rad, Bowen Pan, Ishani Chakraborty, Sean Andrist, Dan Bohus, Ashley Feniello, Bugra Tekin, Felipe Vieira Frujeri, Neel Joshi, and Marc Pollefeys. Holoassist: an egocentric human interaction dataset for interactive ai assistants in the real world. In Proceedings of the International Conference on Computer Vision (ICCV), 2023. 1, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.732, + 0.892, + 0.8 + ], + "angle": 0, + "content": "[127] Xiaohan Wang, Yuehu Liu, Xinhang Song, Beibei Wang, and Shuqiang Jiang. Generating explanations for embodied action decision from visual observation. In Proceedings of the 31st ACM International Conference on Multimedia, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.802, + 0.892, + 0.871 + ], + "angle": 0, + "content": "[128] Yeping Wang, Gopika Ajaykumar, and Chien-Ming Huang. See what i see: Enabling user-centric robotic assistance using first-person demonstrations. In Proceedings of the 2020 ACM/IEEE International Conference on Human-Robot Interaction, pages 639-648, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.873, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[129] Yi Wang, Kunchang Li, Yizhuo Li, Yinan He, Bingkun Huang, Zhiyu Zhao, Hongjie Zhang, Jilan Xu, Yi Liu, Zun" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "22084" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.118, + 0.092, + 0.47, + 0.134 + ], + "angle": 0, + "content": "Wang, et al. Intervideo: General video foundation models via generative and discriminative learning. arXiv preprint arXiv:2212.03191, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.135, + 0.471, + 0.189 + ], + "angle": 0, + "content": "[130] Zirui Wang, Jiahui Yu, Adams Wei Yu, Zihang Dai, Yulia Tsvetkov, and Yuan Cao. Simvlm: Simple visual language model pretraining with weak supervision. arXiv preprint arXiv:2108.10904, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.19, + 0.472, + 0.259 + ], + "angle": 0, + "content": "[131] Pengfei Wei, Lingdong Kong, Xinghua Qu, Yi Ren, Jing Jiang, Xiang Yin, et al. Unsupervised video domain adaptation for action recognition: A disentanglement perspective. In Proceedings of the Advances in Neural Information Processing Systems (NeurIPS), 2023. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.26, + 0.472, + 0.315 + ], + "angle": 0, + "content": "[132] Daniel Weinland, Mustafa Özuysal, and Pascal Fua. Making action recognition robust to occlusions and viewpoint changes. In Proceedings of the European Conference on Computer Vision (ECCV), 2010. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.316, + 0.472, + 0.385 + ], + "angle": 0, + "content": "[133] Mark Wilson, John Mcgrath, Samuel Vine, James Brewer, David Defriend, and Richard Masters. Psychomotor control in a virtual laparoscopic surgery training environment: gaze control parameters differentiate novices from experts. Surgical Endoscopy, 24(10):2458-2464, 2010. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.386, + 0.472, + 0.455 + ], + "angle": 0, + "content": "[134] Benita Wong, Joya Chen, You Wu, Stan Weixian Lei, Dongxing Mao, Difei Gao, and Mike Zheng Shou. Assistq: Affordance-centric question-driven task completion for egocentric assistant. In Proceedings of the European Conference on Computer Vision (ECCV), 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.455, + 0.472, + 0.51 + ], + "angle": 0, + "content": "[135] Xinxiao Wu, Han Wang, Cuiwei Liu, and Yunde Jia. Cross-view action recognition over heterogeneous feature spaces. In Proceedings of the International Conference on Computer Vision (ICCV), 2013. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.511, + 0.472, + 0.566 + ], + "angle": 0, + "content": "[136] Haifeng Xia, Pu Wang, and Zhengming Ding. Incomplete multi-view domain adaptation via channel enhancement and knowledge transfer. In Proceedings of the European Conference on Computer Vision (ECCV), 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.567, + 0.472, + 0.621 + ], + "angle": 0, + "content": "[137] Jianjia Xin, Lichun Wang, Kai Xu, Chao Yang, and Baocai Yin. Learning interaction regions and motion trajectories simultaneously from egocentric demonstration videos. IEEE Robotics and Automation Letters, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.622, + 0.472, + 0.691 + ], + "angle": 0, + "content": "[138] Hu Xu, Gargi Ghosh, Po-Yao Huang, Dmytro Okhonko, Armen Aghajanyan, Florian Metze, Luke Zettlemoyer, and Christoph Feichtenhofer. Videoclip: Contrastive pre-training for zero-shot video-text understanding. arXiv preprint arXiv:2109.14084, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.692, + 0.472, + 0.761 + ], + "angle": 0, + "content": "[139] Jia Xu, Lopamudra Mukherjee, Yin Li, Jamieson Warner, James M Rehg, and Vikas Singh. Gaze-enabled egocentric video summarization via constrained submodular maximization. In Proceedings of the Conference on Computer Vision and Pattern Recognition (CVPR), 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.761, + 0.472, + 0.803 + ], + "angle": 0, + "content": "[140] Jilan Xu, Yifei Huang, Junlin Hou, Guo Chen, Yuejie Zhang, Rui Feng, and Weidi Xie. Retrieval-augmented egocentric video captioning. arXiv preprint arXiv:2401.00789, 2024.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.804, + 0.472, + 0.872 + ], + "angle": 0, + "content": "[141] Mingze Xu, Chenyou Fan, Yuchen Wang, Michael S Ryoo, and David J Crandall. Joint person segmentation and identification in synchronized first-and third-person videos. In Proceedings of the European Conference on Computer Vision (ECCV), 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.472, + 0.902 + ], + "angle": 0, + "content": "[142] Hongwei Xue, Tiankai Hang, Yanhong Zeng, Yuchong Sun, Bei Liu, Huan Yang, Jianlong Fu, and Baining Guo." + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.541, + 0.093, + 0.892, + 0.148 + ], + "angle": 0, + "content": "Advancing high-resolution video-language representation with large-scale video transcriptions. In Proceedings of the Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.15, + 0.894, + 0.218 + ], + "angle": 0, + "content": "[143] Zihui Xue and Kristen Grauman. Learning fine-grained view-invariant representations from unpaired ego-exo videos via temporal alignment. In Proceedings of the Advances in Neural Information Processing Systems (NeurIPS), 2023. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.221, + 0.894, + 0.303 + ], + "angle": 0, + "content": "[144] Takuma Yagi, Misaki Ohashi, Yifei Huang, Ryosuke Furuta, Shungo Adachi, Toutai Mitsuyama, and Yoichi Sato. Finebio: A fine-grained video dataset of biological experiments with hierarchical annotations. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.305, + 0.894, + 0.374 + ], + "angle": 0, + "content": "[145] Lijin Yang, Yifei Huang, Yusuke Sugano, and Yoichi Sato. Interact before align: Leveraging cross-modal knowledge for domain adaptive action recognition. In Proceedings of the Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.376, + 0.894, + 0.432 + ], + "angle": 0, + "content": "[146] Sherry Yang, Jacob Walker, Jack Parker-Holder, Yilun Du, Jake Bruce, Andre Barreto, Pieter Abbeel, and Dale Schuurmans. Video as the new language for real-world decision making. arXiv preprint arXiv:2402.17139, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.433, + 0.892, + 0.474 + ], + "angle": 0, + "content": "[147] Li Yingzhen and Stephan Mandt. Disentangled sequential autoencoder. In Proceedings of the International Conference on Machine Learning (ICML), 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.476, + 0.894, + 0.532 + ], + "angle": 0, + "content": "[148] Tianhe Yu, Chelsea Finn, Annie Xie, Sudeep Dasari, Tianhao Zhang, Pieter Abbeel, and Sergey Levine. One-shot imitation from observing humans via domain-adaptive meta-learning. arXiv preprint arXiv:1802.01557, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.533, + 0.892, + 0.588 + ], + "angle": 0, + "content": "[149] Xumin Yu, Yongming Rao, Wenliang Zhao, Jiwen Lu, and Jie Zhou. Group-aware contrastive regression for action quality assessment. In Proceedings of the International Conference on Computer Vision (ICCV), 2021. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.59, + 0.894, + 0.659 + ], + "angle": 0, + "content": "[150] Zecheng Yu, Yifei Huang, Ryosuke Furuta, Takuma Yagi, Yusuke Goutsu, and Yoichi Sato. Fine-grained affordance annotation for egocentric hand-object interaction videos. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.66, + 0.894, + 0.73 + ], + "angle": 0, + "content": "[151] Thorsten O Zander, Matti Gaertner, Christian Kothe, and Roman Vilimek. Combining eye gaze input with a brain-computer interface for touchless human-computer interaction. International Journal of Human-Computer Interaction, 27(1):38-51, 2010. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.732, + 0.894, + 0.8 + ], + "angle": 0, + "content": "[152] Rowan Zellers, Ximing Lu, Jack Hessel, Youngjae Yu, Jae Sung Park, Jize Cao, Ali Farhadi, and Yejin Choi. Merlot: Multimodal neural script knowledge models. In Proceedings of the Advances in Neural Information Processing Systems (NeurIPS), 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.802, + 0.894, + 0.858 + ], + "angle": 0, + "content": "[153] Lingzhi Zhang, Shenghao Zhou, Simon Stent, and Jianbo Shi. Fine-grained egocentric hand-object segmentation: Dataset, model, and applications. In Proceedings of the European Conference on Computer Vision (ECCV), 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.86, + 0.894, + 0.902 + ], + "angle": 0, + "content": "[154] Siwei Zhang, Qianli Ma, Yan Zhang, Zhiyin Qian, Taein Kwon, Marc Pollefeys, Federica Bogo, and Siyu Tang. Ego-body: Human body shape and motion of interacting people" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "22085" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.118, + 0.092, + 0.469, + 0.12 + ], + "angle": 0, + "content": "from head-mounted devices. In Proceedings of the European Conference on Computer Vision (ECCV), 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.122, + 0.47, + 0.188 + ], + "angle": 0, + "content": "[155] Weiyu Zhang, Menglong Zhu, and KG Derpanis. From actemes to action: A strongly supervised representation for detailed action understanding. In Proceedings of the International Conference on Computer Vision (ICCV), 2013." + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.192, + 0.471, + 0.247 + ], + "angle": 0, + "content": "[156] Zehua Zhang, David Crandall, Michael Proulx, Sachin Talathi, and Abhishek Sharma. Can gaze inform egocentric action recognition? In Proceedings of the Symposium on Eye Tracking Research and Applications (ETRA), 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.249, + 0.471, + 0.303 + ], + "angle": 0, + "content": "[157] Zhenning Zhang, Zhigeng Pan, Weiqing Li, and Zhiyong Su. X-board: an egocentric adaptive ar assistant for perception in indoor environments. Virtual Reality, 27(2):1327-1343, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.305, + 0.471, + 0.36 + ], + "angle": 0, + "content": "[158] Yue Zhao, Ishan Misra, Philipp Krahenbuhl, and Rohit Girdhar. Learning video representations from large language models. In Proceedings of the Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.362, + 0.47, + 0.416 + ], + "angle": 0, + "content": "[159] Jingjing Zheng and Zhuolin Jiang. Learning view-invariant sparse representations for cross-view action recognition. In Proceedings of the International Conference on Computer Vision (ICCV), 2013. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.419, + 0.471, + 0.487 + ], + "angle": 0, + "content": "[160] Yang Zheng, Yanchao Yang, Kaichun Mo, Jiaman Li, Tao Yu, Yebin Liu, Karen Liu, and Leonidas J Guibas. Gimo: Gaze-informed human motion prediction in context. In Proceedings of the European Conference on Computer Vision (ECCV), 2022. 3" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.471, + 0.487 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "22086" + } + ] +] \ No newline at end of file diff --git a/2024/EgoExoLearn_ A Dataset for Bridging Asynchronous Ego- and Exo-centric View of Procedural Activities in Real World/134e2089-370f-4ad6-9686-84a1d7af0ed4_origin.pdf b/2024/EgoExoLearn_ A Dataset for Bridging Asynchronous Ego- and Exo-centric View of Procedural Activities in Real World/134e2089-370f-4ad6-9686-84a1d7af0ed4_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..6e093ad736e234fc49eafd2d9376e8689e98a304 --- /dev/null +++ b/2024/EgoExoLearn_ A Dataset for Bridging Asynchronous Ego- and Exo-centric View of Procedural Activities in Real World/134e2089-370f-4ad6-9686-84a1d7af0ed4_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8eee97cb088939e37a1b8983b520639187af2f7dba46b39a57cc3d6ad4ca4b68 +size 3004093 diff --git a/2024/EgoExoLearn_ A Dataset for Bridging Asynchronous Ego- and Exo-centric View of Procedural Activities in Real World/full.md b/2024/EgoExoLearn_ A Dataset for Bridging Asynchronous Ego- and Exo-centric View of Procedural Activities in Real World/full.md new file mode 100644 index 0000000000000000000000000000000000000000..b3fee38d7dcf2263ad7983a554649f202e852741 --- /dev/null +++ b/2024/EgoExoLearn_ A Dataset for Bridging Asynchronous Ego- and Exo-centric View of Procedural Activities in Real World/full.md @@ -0,0 +1,379 @@ +# EgoExoLearn: A Dataset for Bridging Asynchronous Ego- and Exo-centric View of Procedural Activities in Real World + +Yifei Huang†‡, Guo Chen‡, Jilan Xu‡, Mingfang Zhang‡, Lijin Yang, Baoqi Pei Hongjie Zhang, Lu Dong, Yali Wang*, Limin Wang*, Yu Qiao* OpenGVLab, Shanghai AI Laboratory Shenzhen Institutes of Advanced Technology, CAS Nanjing University + +# Abstract + +Being able to map the activities of others into one's own point of view is a fundamental human skill even from a very early age. Taking a step toward understanding this human ability, we introduce EgoExoLearn, a large-scale dataset that emulates the human demonstration following process, in which individuals record egocentric videos as they execute tasks guided by exocentric-view demonstration videos. Focusing on the potential applications in daily assistance and professional support, EgoExoLearn contains egocentric and demonstration video data spanning 120 hours captured in daily life scenarios and specialized laboratories. Along with the videos we record high-quality gaze data and provide detailed multimodal annotations, formulating a playground for modeling the human ability to bridge asynchronous procedural actions from different viewpoints. To this end, we present benchmarks such as cross-view association, cross-view action planning, and cross-view referenced skill assessment, along with detailed analysis. We expect EgoExoLearn can serve as an important resource for bridging the actions across views, thus paving the way for creating AI agents capable of seamlessly learning by observing humans in the real world. The dataset and benchmark codes are available at https://github.com/OpenGVLab/EgoExoLearn. + +# 1. Introduction + +Even as a child, humans can observe the actions of others and then map them to their own view [4, 34, 100, 105]. With this ability to asynchronously bridge activities from egocentric and exocentric views [95, 103], humans can watch others' demonstrations and replicate the procedures in a new environment. This ability is especially beneficial when actual physical trials carry the potential of high costs [24], e.g., conducting dangerous chemical experiments. + +In the wake of recent advancements in AI systems, one + +goal for the next generation of AI agents is to perform tasks in a more embodied setting [94]. However, different from humans, training these AI agents usually requires demonstration videos taken in a similar environment [75, 119] and from a congruent perspective with the AI agents, (e.g., the egocentric point of view [43, 59, 108, 134]). While great effort has been made into the collection of egocentric data in different scenarios [18, 30, 106], it remains crucial for the AI agents to directly learn from demonstration videos taken in a different place and from a different viewpoint [35, 148]. Realizing this capability can unleash the full potential of public instructional video data [82] and is also useful in the human-robot cooperation scenario, especially in novel environments [57, 77, 128]. + +Current works towards this goal can be roughly divided into two directions. One way is to learn models in simulated environments [11, 64, 79, 89, 92], but it remains difficult for models in this setting to generalize in the real world [126]. The other direction is to learn from human activity in real-world scenarios. However, attempts to directly combine existing multiview datasets often yield datasets of inferior quality or scale [124, 143]. Meanwhile, the few existing datasets in this direction [99, 106, 109] only record ego- and exo-view videos in the same environment and in a time-synchronized manner. In reality, when following demonstrations it is often needed to bridge a series of procedural actions performed in a different place and at a different time. However, currently, no dataset is available for the exploration of how to bridge asynchronous procedural activities in realistic egocentric and exocentric viewpoints. + +To address this lack of dataset issue, we introduce EgoExoLearn, a large-scale dataset containing demonstration videos and corresponding egocentric videos where the camera wearers follow the demonstrations and perform the same task in a different environment, as shown in Figure 1. Targeting two potential applications, i.e., daily assistance and professional support, EgoExoLearn consists of 747 video sequences spanning a total of 120 hours of footage, ranging from daily food-making to specialized laboratory experi + +![](images/31594f139b11ce421ea3883f07660feaa1a57473de5d0f9f5fd264319f7ffe46.jpg) +Figure 1. EgoExoLearn emulates the human asynchronous demonstration following process. It contains demonstration videos of multiple tasks, together with egocentric videos recorded by participants replicating the procedure after watching the demonstrations. The dataset comprises gaze signals and fine-grained multi-level multi-modal annotations, enabling the exploration of key features in this context such as cross-view association and cross-view action planning. + +ments. Notably, the egocentric videos in EgoExoLearn contain eye gaze signals showing humans' visual attention while performing the task. This provides a valuable cue for better bridging the actions in ego- and exo-viewpoints. + +We take one more step forward by analyzing human ability in bridging asynchronous ego- and exo-view actions and, accordingly, introduce new tasks and benchmarks that we believe can form building blocks for the development of next-stage embodied AI agents with similar abilities. When humans perform an action, he/she can associate and describe the undergoing action in the egocentric view with the corresponding action in the demonstration. With the knowledge from demonstration videos, humans can know the needed action steps and predict what the next steps should be. Besides, through the comparison with the demonstration, humans can also assess their level of skills. + +Based on the above analysis, we design benchmarks of 1) cross-view association, 2) cross-view action understanding, 3) cross-view referenced skill assessment, and 4) cross-view referenced video captioning. Each benchmark is meticulously defined, annotated, and supported by baseline implementations. In addition, we pioneeringly explore the role of gaze in these tasks. We hope our dataset can provide resources for future work for bridging asynchronous procedural actions in ego- and exo-centric perspectives, thereby inspiring the design of AI agents adept at learning from real-world human demonstrations and mapping the procedural actions into robot-centric views. + +# 2. Related Work + +Ego-exo datasets. While there exist works that associate existing datasets to explore how activities can be bridged between them, these associated datasets are often limited in scale [14, 144, 147] or quality [124], meanwhile focusing only on single actions captured from the same view [81, 88, + +131]. As for actions from different views, apart from multiview fixed camera datasets [6, 15, 54, 61], there also exist datasets with both ego- exo-centric view videos [30, 31, 99, 106, 109]. These datasets are either recorded in the same environment [45, 99, 109] or record time-synced multi-view videos in the same environment with primary focuses on pose/activity understanding grounded in the 3D world [55, 106, 154]. Our dataset offers a more challenging and realistic scenario, where egocentric camera wearers learn to complete the tasks demonstrated by exocentric demonstration videos. This setting complements these datasets by focusing on high-level procedural actions. + +The only dataset conceptually similar to ours is the recently proposed AE2 dataset [143], where the goal is to learn view-invariant representation from unpaired ego and exo videos. This dataset combines ego and exo videos from five public datasets [18, 20, 53, 55, 155] and a newly collected ego tennis forehand dataset. However, due to the difficulty in associating existing ego-exo datasets, the AE2 dataset is relatively small where the largest subset contains only 322 clips. Also, this dataset only focuses on clip-level actions, and thus cannot feature the real-world demonstration following setting, which usually requires multimodal, task-centric procedural knowledge. Instead, our EgoExoLearn is much larger in scale (100x more clips), while offering gaze and fine-grained multimodal annotations facilitating multi-faceted analysis of ego-exo action understanding. + +Egocentric video datasets. In line with the recent development in wearable cameras [110], multiple egocentric video datasets [5, 17, 19, 30, 46, 78, 98, 112, 150] have been proposed. Different from previous egocentric datasets, the egocentric videos in EgoExoLearn feature a demonstration-following setting. We believe EgoExoLearn provides a playground for developing tools to bridge asynchronous procedural activities from ego- and exocentric viewpoints. + +The setting of our EgoExoLearn complements existing datasets like Ego4D and can benefit from their rich knowledge and representations. + +Egocentric gaze. Gaze can indicate visual attention and contains valuable information about human intent [36, 156], thus is used in a diverse range of areas such as human-computer interaction [42, 51, 151], and augmented reality [90, 102]. In computer vision, efforts have been made to leverage gaze in various tasks [37-39, 56, 65-67, 83, 101, 139]. However, with the previous absence of large-scale egocentric datasets that include gaze, this avenue of research is currently under-explored [44, 68, 94, 154, 160]. Our EgoExoLearn offers calibrated gaze positions for all egocentric videos. Thanks to our unique setting, our dataset enables the integration of gaze in egocentric video understanding and the exploration of the role of gaze in the cross-view context. + +Egocentric and ego-exo video understanding. The unique recording perspective of egocentric videos presents a series of challenges including but not limited to action understanding [13, 26, 27, 40, 50, 93, 96, 113, 120, 121, 125], hand detection [12, 29, 107, 153], and video-language understanding [41, 48]. These form fundamental building block techniques of embodied AI [85], VR/AR [9, 47, 74, 123], and human-robot interaction [63, 80, 86, 137]. Since most egocentric datasets are smaller in scale compared with general datasets which contain mostly exocentric view videos [3, 142, 152], it is possible to leverage exocentric video data to improve model performance on egocentric videos [132]. There are typically three main directions: joint view-invariant learning [124, 140, 142], domain adaptation [136], and knowledge distillation [69]. In this work, we evaluate all these directions in our benchmarks. + +# 3. Dataset + +# 3.1. Data Collection + +Scenarios and tasks. We consider procedural goal-oriented tasks ranging from daily food-making to specialized laboratory-based experiments. This selection is grounded in their exemplification of two prospective areas where future embodied AI agents would need the ability to bridge ego-exo activities: daily-life assistance and professional support. Specifically, EgoExoLearn incorporates 5 types of daily tasks (e.g., cooking) and 3 types of specialized laboratory tasks (e.g., solid-phase peptide synthesis). We record egocentric videos in 4 different kitchens and 3 different labs. Other details are provided in the supplementary. + +Data collection procedure. Before the start of each collection session, participants are required to complete a questionnaire gathering basic demographic information and their self-evaluated expertise in executing the designated task. This questionnaire also highlights the ethical, privacy, and secu + +![](images/41bf1cd76182fcc4e543357f9fff5b20ff68656802bdde7eb624f84544efa2ad.jpg) +Figure 2. The number of videos per task (left) and the average duration of each video per task (right). Task1 to Task5 represent the 5 daily tasks and the remaining are three tasks in specialized laboratories. In each recording session of the egocentric video, one participant may learn from multiple demonstration videos and one demonstration video may be watched by several participants. + +![](images/6e81b914ecc921aa6bfb1c83be55963cf5f84e4eaf459922900c1d7c7a7f5cfe.jpg) + +rity considerations. Then in each session, participants will be asked to choose one or several exocentric view demonstration videos from a provided list and carefully learn the detailed procedures. Once they feel ready, they will wear Pupil Invisible Glasses [49], complete the gaze calibration, and begin to replicate the task performed in the demonstration videos. While not encouraged, participants are permitted to revisit the demonstration video during the recording. + +After each recording session, the participants are asked to re-do the gaze calibration to ensure gaze data fidelity. For the 5 daily tasks, the exocentric demonstration videos are manually curated from online video platforms such as YouTube. For the lab experiments, the exocentric demonstration videos are tutorials recorded by senior lab members. + +Figure 2 shows the distribution of the 120 hours of data. Since most demonstration videos are meticulously edited to remove repeated steps, the average length of demonstration videos is lower than the egocentric videos which record the full procedure. As a result, the EgoExoLearn contains 432 egocentric videos totaling 96.5 hours and 315 demonstration videos spanning 23.5 hours. This difference in video length poses a unique challenge when bridging ego- and exo-view activities for future research endeavors. + +# 3.2. Annotation + +To facilitate our dataset in the development of algorithms that can effectively bridge the gap between ego and exo viewpoints, we provide detailed multi-modal human annotations. Our pipeline of annotation contains four stages detailed in the following paragraphs. Each of these stages is subject to a rigorous manual quality check involving no fewer than two individuals for verification and validation. + +Coarse-level language annotation. In this step, we ask annotators to annotate the coarse actions in the videos. Like + +![](images/6a91f604f0dc85ff41960cbd533e77603029baf2ce792b1ff2d4376c7596dd7b.jpg) + +![](images/1d69c8321a354769981259b26af409b5416cbfbe34be289b772a54d9988d2775.jpg) +Number of nouns (left axis) and noun durations (right axis) of right hand +Figure 3. Occurrence and duration distribution of the annotated fine-level verbs and nouns associated with the right hand. + +the previous works [106, 126], the coarse actions are defined as a middle-level step for accomplishing a task and can be divided into multiple fine actions. For instance, "Prepare the pork" in the task of twice-cooked pork, and "Suction filtration" in solid-phase peptide synthesis. Three types of annotation are given in this step: 1) the temporal interval, consisting of start and end timestamps; 2) the action label; and 3) a language description of the video within the annotated interval. We specifically request annotators to focus on elucidating "what is done," "how it is done," and "the purpose of this step" in their language descriptions. We define a total of 39 categories of coarse-level actions in this stage and acquire 41.2 coarse-level annotations per video with an average length of 21.5 seconds. + +Fine-level language annotation. Based on the coarse-level annotations, in this step we request the annotators to provide annotations for the fine-level actions. The fine-level actions are the atomic actions like "take knife" or "pull syringe plunger". Unlike the first step, annotators are instructed to furnish language descriptions that specifically emphasize "which hand is used", "what object is used" and "why it is used". For the first two steps, we employ a two-round manual annotation checking to ensure the annotation quality. + +Translation & parsing. To ensure linguistic precision, all the annotators give the language description using their native language [18]. For non-English annotations, we employ ChatGPT and Google Translation API to translate them into English. Subsequently, for the fine-level annotations, we employ specific rules and utilize tools such as NLTK [10] and Spacy [116] to extract the verbs and nouns associated with each segment. During this stage, we confine the selection of verbs and nouns to the predefined taxonomy offered by Ego4D [30], while also manually introducing supplementary verbs and nouns that are absent in the taxonomy. Since our manual annotation specifies the engagement of the left/right hand, we can extract multiple verbs and nouns for each segment, meanwhile attributing them to the respective involvement of the left or right hand. After manual checking, we obtain a total of 95 verb and 254 noun categories in the fine-level annotation. In Figure 3 we show the occurrence of the top 30 categories of verbs and nouns attributed to the right hand. More statistics can be found in the supplementary. + +Skill level annotation. Since self-assessed skill level is not perfectly suitable for skill assessment, we identify several representative skills and assign human annotators to assess their skills. The annotation follows a pairwise ranking scheme, where annotators are presented with pairs of videos of the same action, and instructed to determine which video demonstrates a higher skill level. We prepare 40,191 video pairs and ensure that 4 different annotators annotate each pair. After filtering out pairs with less than 3 consistent opinions, we get a collection of 34,239 valid video pairs. + +# 3.3. Statistics & Comparisons + +To the best of our knowledge, there is no dataset that follows the same setting as ours for a direct comparison. Therefore, we enumerate various aspects of our dataset and conduct a comparative analysis with relevant datasets in Tables 1 and 2. EgoExoLearn distinctively enriches the domain with its "visual demonstration following" setting. Beyond this unique setting, it stands as the first egocentric dataset that includes temporal bounded language captions, annotated cross-view associations, and multi-label video segments. + +# 4. Dataset Properties & Benchmarks + +# 4.1. Dataset Properties + +EgoExoLearn stands out from current egocentric and ego-exo datasets due to several unique properties. + +Ego-Exo demonstration following setting. The most distinguished property of EgoExoLearn is the ego-exo demonstration following context. Egocentric video recorders are instructed to follow the steps in exocentric demonstration videos to perform the same task but in a different environment. This setting closely emulates the human observational learning process [4, 34] and can be instrumental in designing embodied AI agents that learn from alternative perspectives while executing tasks from their own viewpoint. + +Fine-grained vision-language annotations with gaze. To facilitate a deeper analysis, we equip our dataset with rich, multimodal, fine-grained annotations. EgoExoLearn is the first egocentric dataset featuring high-quality captions, evidenced by the number of words per segment in Tab. 2. + +
DatasetSettingsUnique HoursEgo +Exo?Instruction following?Visual instruction?GazeCoarse ActionFine ActionDense CaptionAssociationSkill
Meccano [98]Industry7X✓XXX✓XXX
EGTEA [67]Cooking28X✓X✓X✓XXX
EK-100 [18]Cooking100XXXXX✓XXX
HoloAssist [126]Assistive166X✓X✓✓✓XX✓
Ego4D [30]Multiple3670XXX◇◇◇◇XX
H2O [55]Desk1✓XXXX✓XXX
LEMMA [45]Daily10✓XXXX✓XXX
HOMEAGE [99]Daily25✓XXX✓✓XXX
CharadesEgo [109]Daily34✓XXXX✓XXX
Assembly101 [55]Desk42✓✓XX✓✓XXX
EgoExoLearn (ours)Daily & Lab120✓✓✓✓✓✓✓✓✓
+ +Table 1. Comparison to related datasets on settings (left) and annotations (right). "Unique hours" refers to the cumulative duration of distinct video recordings, counting only one camera's footage for simultaneous recordings of the same activity. $\diamond$ : Partially included. + +
Dataset# videosAvg. min# segsAvg. secverb classesnoun classes# verb /seg#noun /seg#words /seg
CharadesEgo [109]78600.569k1.83237110
HOMEAGE [99]57000.926k-2986--0
EK-100 [18]7008.590k3.19730011.23.0
Assembly101 [106]43217.183k1.72490110
Ego4D [30]99126.477k8.0115478117.4
Ours-ego43213.464k4.6952541.82.516.9
Ours-exo3154.514k4.7822512.33.019.4
+ +Table 2. Contemporary egocentric datasets. We show only the fine-level actions for a fair comparison. For Ego4D [30], we select the closest subtask of "forecasting" following [106]. + +Different from Ego4D where captions are associated with only single timestamps, our captions come with manually annotated start and end timestamps. For better visual perception across ego-exo views, we give verb and noun labels associated with the specific hand. One aspect from which we can analyze the human's ability to bridge ego-exo activities is through gaze. EgoExoLearn is also augmented with calibrated eye-gaze signals. These annotations enable the understanding of human ability to bridge ego-exo activities from diverse perspectives, which we posit will benefit the next-generation embodied AI agents [21, 62]. + +# 4.2. Benchmarks + +To evaluate the ability of bridging asynchronous ego-exo procedural activities, we introduce 4 new benchmarks: 1) cross-view association, 2) cross-view action understanding, 3) cross-view referenced skill assessment, and 4) cross-view referenced captioning. The cross-view action understanding benchmark is further subdivided into three subtasks: cross-view action anticipation, cross-view action planning, and cross-view action segmentation. Additionally, we explore the role of gaze in assisting these tasks. We also benchmark models on zero-shot and supervised fine-grained action recognition tasks for reference, following [18, 126]. Note that we carefully split our dataset to eliminate annotation leak across benchmarks. Due to the space limit, we only provide partial content of definition, annotation, results, and analysis, + +leaving more complete details in the supplementary. + +# 4.2.1 Cross-view association + +Motivation. One straightforward indicator of the ego-exo activity bridging is the ability to associate the same semantics across ego- and exo-views. This benchmark focuses on equipping models with this cross-view association ability. An application of this ability is assistants in AR that can show expert demonstration videos when the human is confused [157]. Another potential application is the embodied AI agent that can explain its decision [118, 127, 146]. + +Problem settings. We formulate this association benchmark as a cross-view multiple-choice association problem. Specifically, we consider two different cross-view association settings: ego2exo and exo2ego. In the case of ego2exo, given an egocentric video, the model is asked to predict the corresponding exocentric video performing the same action from a candidate choice set of exocentric samples, and vice versa for the exo2ego setting. For both ego2exo and exo2ego settings, we use 20 candidate samples for each query. The evaluation metric is the averaged Top-1 accuracy. + +Annotations. We meticulously construct the ground-truth ego-exo pairs via a semantic-aware matching process. It is composed of five stages with details in the supplementary material: (1) Scenario Matching. (2) Noun and Verb Matching. (3) Sentence Matching with LLM. (4) Negative Sampling. (5) Two-round Manual Verification. Notably, we do not provide such pairs for the training set and leave the modeling of cross-view association on unpaired samples to be further explored for the community. + +Baseline model. We adopt three types of baseline models, i.e., ego-only, exo-only, and ego-exo, which refer to the data we used during training. Under all the settings, we leverage the paired video and caption and jointly train a video encoder and a text encoder using the contrastive loss [97]. We use TimeSformer-B [8] as the video encoder and clip-text [97] as the text encoder. We initialize both encoders + +![](images/91014facf6f8bcfa9554cb9dad5ad362d552cd3af1a70c237964fa8454581a74.jpg) +Figure 4. Concept of the 3 benchmarks of cross-view association (Sec. 4.2.1), cross-view action anticipation & planning (Sec. 4.2.2) and cross-view reference skill assessment (Sec.4.2.3) in this section. Other benchmarks can be found in the supplementary material. + +![](images/5929e0bdfdd454c39af711fa6adb1ae67b6f37f7d0c96e68327bb792422ed6a8.jpg) + +![](images/e3d6d14183859995b34ad8502e15e3d09f2b00be8121242e9fa5eab2f01469eb.jpg) + +using the EgoVLP [72] pre-trained weights. During testing, we obtain the ego/exocentric video representations using the video encoder. The prediction is defined as the one with the highest normalized cross-view video feature similarity among all candidates. On top of these models, we evaluate the effectiveness of gaze in associating egocentric video and exocentric video. This is achieved by replacing the original egocentric video with spatially cropped videos using the gaze positions as the cropping center. + +Related work. Prior works on pre-trained vision-language models for multiple choice association/questioning [60, 72, 122, 130, 138, 158] are generally pre-trained on either exocentric [3, 82] or egocentric [30] datasets, only revealing weak cross-view bridging ability. Another line of works explores cross-view learning [2, 33, 135, 141] by either transferring the knowledge from one view to the other [69] or training view-invariant video understanding models [124, 143]. Different from previous works, our cross-view association benchmark is in a more realistic but challenging setting, by evaluating the model's ability to associate asynchronous activities across ego- and exo-views. + +Experiment results. We first evaluate several vision-language models via zero-shot transfer. These models are pre-trained either on egocentric videos, i.e. EgoVLP [72] and LaViLa [158], or exocentric videos, i.e., InternVideo [129]. As shown in Tab. 3, without using gaze, EgoVLP generally outperforms the others on both validation and test sets. By introducing gaze information, InternVideo receives a decent improvement, especially on Exo2Ego. + +As for fine-tuned models, Tab. 3, reveals that models trained solely on single-view data struggle with cross-view association. Training ego-only models using gaze-cropped egocentric videos results in substantial improvements, significantly outperforming those trained on center-cropped videos. This again highlights the importance of gaze in enhancing cross-view association. Based on our observation, the regions around gaze help the cluttered ego videos become visually similar to exo videos where the primary object is salient. Last, we show the baseline result when co-trained on both egocentric and exocentric videos. The model in this setting shows the strongest cross-view association ability over + +
MethodGazeValTest
Ego2ExoExo2EgoEgo2ExoExo2Ego
Zero-shot
Random✘12.715.014.113.4
EgoVLP [72]✘28.827.232.128.9
LaViLa [158]✘22.624.928.725.7
InternVideo [129]✘27.021.230.621.7
EgoVLP [72]✓28.829.731.528.9
LaViLa [158]✓21.921.430.325.9
InternVideo [129]✓30.932.333.332.2
Fine-tuned
Exo-only✘42.941.745.446.9
Ego-only✘33.637.140.335.8
Ego-only✓34.638.745.641.8
Ego-onlyCenter25.422.824.724.2
EgoExo✘42.945.449.045.3
EgoExo✓47.948.855.351.1
+ +Table 3. Association accuracy in the cross-view association benchmark. In the fine-tuned setting, we adopt three kinds of data sources for training, i.e., ego-only, exo-only, and hybrid ego-exo data. By leveraging gaze information during training, the model outperforms the baseline (w/o gaze) and the center-crop counterpart. + +single-view models. Findings from this benchmark underline the limitation of current models in associating activities across ego and exo views, and point towards the potential benefits of integrating gaze into the association. + +# 4.2.2 Cross-view action anticipation & planning + +Motivation. The procedural actions are not guaranteed to be identical between the two views due to practical constraints. Thus, we design benchmarks for cross-view action anticipation and planning to enable a thorough understanding and transfer of necessary steps (or actions) for task completion, bridging the gap between views and considering real-world conditions. A practical application of this benchmark can be seen in human-robot collaboration scenarios. For instance, an embodied AI agent, after observing a human perform the first part of a task, could effectively take over and complete the remaining half of the task based on the particular environmental situation. + +Problem settings. For both the cross-view anticipation and + +
MethodGazeAnticipation↑Planning↓
Ego-VEgo-NExo-VExo-NEgoExo
Exo-onlyX29.923.640.940.584.776.1
Ego-onlyX33.437.828.917.683.484.5
Ego-only✓40.552.837.637.680.082.6
Ego-onlyCenter33.238.634.132.782.684.7
Unsupervised Domain Adaption
Ego2ExoX33.638.135.428.783.084.1
Exo2EgoX30.423.639.239.883.979.0
Ego2Exo✓40.854.238.737.182.884.3
Exo2Ego✓33.531.339.140.182.478.8
Knowledge Distillation
Ego2ExoX29.624.941.645.284.375.5
Exo2EgoX34.038.428.618.683.184.3
Ego2Exo✓29.925.041.245.184.875.1
Exo2Ego✓41.056.137.739.179.582.6
Co-training
Ego & ExoX33.537.439.644.383.276.0
Ego & Exo✓43.853.340.344.479.075.6
+ +Table 4. Results of cross-view action anticipation and planning benchmarks. For anticipation, the class-mean Top-5 recall is used as the evaluation metric (higher is better). For planning, the Edit distance is used as the evaluation metric (lower is better). + +cross-view planning, the goal is to anticipate the future activities in one view, given labeled training data only in another view. For the cross-view anticipation task, we focus on predicting the verb and noun categories of the next fine-level action $\tau = 1$ second into the future. Given the multilabel nature of our verb and noun annotations in each fine-level segment, we perform multiclass anticipation. Performance is evaluated by class-mean Top-5 recall as per [18]. For the cross-view planning task, we aim to generate the next $K = 8$ steps of coarse-level actions. We adopt ED@ $K$ as the evaluation metric following the setting of Ego4D LTA [30]. + +Annotations. For action anticipation, we use the fine-level verb and noun annotations, and then take their intersection between ego and exo videos to constrain them in the same closed set. To evaluate the model more effectively, we further control the long-tail degree of the data by and filter out the tail categories that occur less than $\frac{1}{100}$ of the highest occurrence category. For the action planning task, we directly adopt the coarse-level action annotations and regard the start timestamp of each segment as one action step. More details can be found in the supplementary material. + +Baseline model. We explore three distinct directions to implement cross-view baseline models. The first direction is based on unsupervised domain adaptation (UDA), treating one view as the source domain and the other as the target domain. This method operates within an unsupervised training framework, using labels from the source domain and video data from both domains [16, 52, 84, 104, 111, 131, 145]. We adopt CLIP [97] + TA3N [14] as the baseline model. The second direction entails knowledge distillation (KD) [28, 32, 87], allowing the model trained on one view to learn knowl + +edge of the other view, under the assumption that a teacher model of the other view is available. We equip CLIP with a distillation approach based on [69] to transfer knowledge from the teacher model to a newly created student model. The third and most straightforward direction is co-training (CT) using the data from both views to encourage the model to discover correlations between them directly. For using gaze, we also crop the video based on the gaze positions. + +We comprehensively consider four evaluation settings. The "Ego-only" and "Exo-only" settings do not involve cross-view understanding, thus we use zero-shot evaluation serving as the references. The Ego2Exo and Exo2Ego settings are the cross-view settings. For UDA, "Ego2Exo" is defined as utilizing the egocentric view as the source domain and the exocentric view as the target domain. In the context of KD, "Ego2Exo" indicates we initially train a teacher model on egocentric data, followed by the training and distillation of the student model on exocentric data. For CT, we merge both egocentric and exocentric datasets through direct concatenation. We report results on the test set and put the validation set results in the supplementary. + +Related work. Prior works [25, 54, 106, 117] on multi-view action understanding mainly focus on synchronized multiview videos. Some work studied transferring knowledge [69] from one view to the other or training the view-invariant [124, 143, 159] video models. Our cross-view benchmarks seek to evaluate the ability of models to bridge asynchronous actions across views, which is more challenging yet realistic. + +Experiment results. Table 4 presents the results of action anticipation and planning on the test set. The first block of results shows a significant performance gap when models trained exclusively on one view are tested on the other view. This underscores the inherent differences in activities captured in the two views. Remarkably, even without relying on any specific cross-view method, the inclusion of gaze information markedly diminishes the disparity between egocentric and exocentric data. Leveraging techniques such as UDA or KD we can see improved performance compared with direct zero-shot inference in the first block. Since CT can utilize both egocentric and exocentric data, it achieves the best performance in all setups. Moreover, from the comparison in all settings using or not using gaze, it is clear that gaze serves as a surprisingly effective signal to mitigate the gap between activities in the two views, although naively designed. These results take the first step in the potential directions for better bridging the cross-view activities, setting a foundation for future advancements in the field of cross-view action understanding. + +# 4.2.3 Cross-view referenced skill assessment + +Motivation and problem setting. We propose a novel task of cross-view referenced skill assessment leveraging the + +unique setting and annotations of our dataset. This task goes beyond the traditional pairwise ranking often used in skill assessment [22, 58] by incorporating an expert demonstration video as a reference point. This demonstration provides a model of the ideal execution of an action, offering a standard against which to compare skill levels. In this task, the input is a pair of egocentric video clips $C_{ego1}$ , $C_{ego2}$ of the same action and an exo-view demonstration video clip $C_{exo}$ as the reference, and the output is a choice $c \in \{ego1, ego2\}$ of the egocentric clip that demonstrates a higher skill level. Moreover, we can also assess the skill level by the relation between action and gaze. This benchmark evaluates a model's ability to bridge asynchronous ego-exo dynamics. A practical application is an AR system that helps humans in skill acquisition by providing targeted feedback based on skill level assessment and expert demonstration. + +Annotations. For the cross-view referenced skill assessment benchmark, we concentrate on four types of representative actions, as detailed in Tab. 5. We use the skill level annotations in Sec. 3.2 in this benchmark. For each pair of videos, the annotation is provided by four distinct annotators to minimize subjective bias. We ensure credibility by checking the transitivity of annotations and removing the pairs with less than 3 agreements among 4 annotators. We then append an exo-view demonstration video clip of the same action, forming video triplets as the model input. + +Baseline model. Our baseline model is built upon a pairwise ranking skill assessment model RAAN [23]. Without loss of generality, assuming egol is the video showing a higher skill level, we apply the following approaches to leverage the reference exo-view demonstration video: 1) Triplet loss (TL). The feature distance between $C_{exo}$ and $C_{ego1}$ should be closer to the distance between $C_{exo}$ and $C_{ego2}$ . 2) Relation network (RN). Inspired by [114], we employ a relation network that concatenates the features of the ego and exo clips. This network is designed to discern which of the two egocentric video clips bears a closer relation to the demonstration video in terms of skill level. The way to gaze is consistent with the other benchmarks. + +Related work. Several previous works on skill assessment aim to directly regress a score based on professional ratings [1, 71, 73, 91, 115, 149]. We adopt a more general approach of pairwise ranking since no absolute score is available in most real-world skills [7, 70]. Previous works in this direction only use a pair of videos that are in either egocentric view [22] or in exo view [23, 70]. Differently from their works, we explore the cross-view activity bridging ability by examining how demonstration videos from exo view can benefit skill assessment in ego views. + +Experiment results. We first evaluate the performance of previous works under the conventional pairwise setting. In the upper block of Table 5, both methods [22, 23] receive + +
MethodGazeEgg CrackingPeelingStir-fryCutting
Ego pairs only
Who's better [22]✘74.7975.9877.5476.85
RAAN [23]✘78.4578.5282.5379.09
Who's better [22]✓77.9176.7079.1377.02
RAAN [23]✓82.0879.3783.9279.36
Ego pairs + Exo
RAAN [23] + RN✘77.9277.0981.5478.26
RAAN [23] + TL✘78.8179.4682.5078.73
RAAN [23] + RN✓82.1479.3283.5179.44
RAAN [23] + TL✓82.1679.5984.0579.29
+ +Table 5. Ranking accuracy of cross-view referenced skill assessment. In the upper part of the table, only ego video pairs are used, while in the lower part, exo demonstrations are incorporated by "RN": relation network and "TL": triplet loss. + +a clear performance boost when gaze is used. This aligns with behavioral science findings that experts and novices have different gaze patterns in the same task [76, 133]. With the exocentric demonstration as reference, both the relation network method (RN) and the triplet loss method (TL) can leverage the reference video and bring performance improvement. Further adding gaze we can get the best performance. However, the marginal improvement observed with the inclusion of the exocentric reference suggests that current models may still struggle to fully bridge asynchronous activities across egocentric and exocentric views. There remains ample room for new improvement and innovation. + +# 5. Conclusion + +The ability to bridge asynchronous procedural activities in ego- and exo-views is imperative for next-generation embodied AI in executing sophisticated tasks in the real world. As a fundamental step, our EgoExoLearn encompasses a rich collection of egocentric videos, each captured when replicating procedures of exocentric demonstration videos, but performed in different environments and at different times. This realistic setup, combined with our multimodal annotations, allows us to construct 4 novel benchmarks, serving as a versatile platform for investigating how cross-view asynchronous activities can be bridged. EgoExoLearn also enables new research directions e.g., how to better leverage gaze and hand-associated annotations. Results from the benchmarks show weaknesses of current models in bridging ego- and exo-view asynchronous activities, leaving significant room for future work to improve upon. + +Acknowledgement. This work is supported by the National Key R&D Program of China (No.2022ZD0160102) and the Industry Collaboration Projects Grant, Shanghai Committee of Science and Technology, China (No.22YF1461500). + +Key contribution statement: Guo made key contributions to annotation processing and 3 action benchmarks. Jilan made key contributions to the association and caption benchmarks. Mingfang contributed primarily to the skill benchmark. + +# References + +[1] Narges Ahmidi, Lingling Tao, Shahin Sefati, Yixin Gao, Colin Lea, Benjamin Bejar Haro, Luca Zappella, Sanjeev Khudanpur, René Vidal, and Gregory D Hager. A dataset and benchmarks for segmentation and recognition of gestures in robotic surgery. IEEE Transactions on Biomedical Engineering, 64(9):2025-2041, 2017. 8 +[2] Shervin Ardeshir and Ali Borji. An exocentric look at egocentric actions and vice versa. Computer Vision and Image Understanding, 171:61-68, 2018. 6 +[3] Max Bain, Arsha Nagrani, Gül Varol, and Andrew Zisserman. Frozen in time: A joint video and image encoder for end-to-end retrieval. In Proceedings of the International Conference on Computer Vision (ICCV), 2021. 3, 6 +[4] Albert Bandura. Observational learning. The international encyclopedia of communication, 2008. 1, 4 +[5] Siddhant Bansal, Chetan Arora, and C.V. Jawahar. My view is the best view: Procedure learning from egocentric videos. In Proceedings of the European Conference on Computer Vision (ECCV), 2022. 2 +[6] Yizhak Ben-Shabat, Xin Yu, Fatemeh Saleh, Dylan Campbell, Cristian Rodriguez-Opazo, Hongdong Li, and Stephen Gould. The ikaa asm dataset: Understanding people assembling furniture through actions, objects and pose. In Proceedings of the Winter Conference on Applications of Computer Vision (WACV), 2021. 2 +[7] Gedas Bertasius, Hyun Soo Park, Stella X Yu, and Jianbo Shi. Am i a baller? basketball performance assessment from first-person videos. In Proceedings of the International Conference on Computer Vision (ICCV), 2017. 8 +[8] Gedas Bertasius, Heng Wang, and Lorenzo Torresani. Is space-time attention all you need for video understanding? In Proceedings of the International Conference on Machine Learning (ICML), 2021. 5 +[9] Vinay Bettadapura, Irfan Essa, and Caroline Pantofaru. Ego-centric field-of-view localization using first-person point-of-view devices. In Proceedings of the Winter Conference on Applications of Computer Vision (WACV), 2015. 3 +[10] Steven Bird. Nltk: the natural language toolkit. In Proceedings of the COLING/ACL Interactive Presentation Sessions, 2006. 4 +[11] Greg Brockman, Vicki Cheung, Ludwig Pettersson, Jonas Schneider, John Schulman, Jie Tang, and Wojciech Zaremba. Openai gym. arXiv preprint arXiv:1606.01540, 2016. 1 +[12] Minjie Cai, Kris M Kitani, and Yoichi Sato. Understanding hand-object manipulation with grasp types and object attributes. In Robotics: Science and Systems, 2016. 3 +[13] Guo Chen, Yifei Huang, Jilan Xu, Baoqi Pei, Zhe Chen, Zhiqi Li, Jiahao Wang, Kunchang Li, Tong Lu, and Limin Wang. Video mamba suite: State space model as a versatile alternative for video understanding. arXiv preprint arXiv:2403.09626, 2024. 3 +[14] Min-Hung Chen, Zsolt Kira, Ghassan AlRegib, Jaekwon Yoo, Ruxin Chen, and Jian Zheng. Temporal attentive alignment for large-scale video domain adaptation. In Proceedings of the International Conference on Computer Vision (ICCV), 2019. 2, 7 + +[15] Kellie Corona, Katie Osterdahl, Roderic Collins, and Anthony Hoogs. Meva: A large-scale multiview, multimodal video dataset for activity detection. In Proceedings of the Winter Conference on Applications of Computer Vision (WACV), 2021. 2 +[16] Victor G Turrisi da Costa, Giacomo Zara, Paolo Rota, Thiago Oliveira-Santos, Nicu Sebe, Vittorio Murino, and Elisa Ricci. Dual-head contrastive domain adaptation for video action recognition. In Proceedings of the Winter Conference on Applications of Computer Vision (WACV), 2022. 7 +[17] Dima Damen, Hazel Doughty, Giovanni Maria Farinella, Sanja Fidler, Antonino Furnari, Evangelos Kazakos, Davide Moltisanti, Jonathan Munro, Toby Perrett, Will Price, et al. Scaling egocentric vision: The epic-kitchens dataset. In Proceedings of the European Conference on Computer Vision (ECCV), 2018. 2 +[18] Dima Damen, Hazel Doughty, Giovanni Maria Farinella, Antonino Furnari, Jian Ma, Evangelos Kazakos, Davide Moltisanti, Jonathan Munro, Toby Perrett, Will Price, and Michael Wray. Rescaling egocentric vision: Collection, pipeline and challenges for epic-kitchens-100. International Journal of Computer Vision (IJCV), 130:33-55, 2022. 1, 2, 4, 5, 7 +[19] Ahmad Darkhalil, Dandan Shan, Bin Zhu, Jian Ma, Amlan Kar, Richard Higgins, Sanja Fidler, David Fouhey, and Dima Damen. Epic-kitchens visor benchmark: Video segmentations and object relations. In Proceedings of the Neural Information Processing Systems (NeurIPS) Track on Datasets and Benchmarks, 2022. 2 +[20] Fernando De la Torre, Jessica Hodgins, Adam Bargteil, Xavier Martin, Justin Macey, Alex Collado, and Pep Beltran. Guide to the carnegie mellon university multimodal activity (cmu-mmac) database. 2009. 2 +[21] Matt Deitke, Winson Han, Alvaro Herrasti, Aniruddha Kembhavi, Eric Kolve, Roozbeh Mottaghi, Jordi Salvador, Dustin Schwenk, Eli VanderBilt, Matthew Wallingford, et al. Robothor: An open simulation-to-real embodied ai platform. In Proceedings of the Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 5 +[22] Hazel Doughty, Dima Damen, and Walterio Mayol-Cuevas. Who's better? who's best? pairwise deep ranking for skill determination. In Proceedings of the Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 8 +[23] Hazel Doughty, Walterio Mayol-Cuevas, and Dima Damen. The pros and cons: Rank-aware temporal attention for skill determination in long videos. In Proceedings of the Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 8 +[24] Mitch J Fryling, Cristin Johnston, and Linda J Hayes. Understanding observational learning: An interbehavioral approach. The Analysis of verbal behavior, 27:191, 2011. 1 +[25] Yanwei Fu, Yanwen Guo, Yanshu Zhu, Feng Liu, Chuanming Song, and Zhi-Hua Zhou. Multi-view video summarization. IEEE Transactions on Multimedia, 12(7):717-729, 2010. 7 +[26] Antonino Furnari and Giovanni Maria Farinella. What would you expect? anticipating egocentric actions with rolling-unrolling lstms and modality attention. In Proceedings of + +the International Conference on Computer Vision (ICCV), 2019.3 +[27] Rohit Girdhar and Kristen Grauman. Anticipative video transformer. In Proceedings of the International Conference on Computer Vision (ICCV), 2021. 3 +[28] Rohit Girdhar, Du Tran, Lorenzo Torresani, and Deva Ramanan. Distinct: Learning video representations without a single labeled video. In Proceedings of the International Conference on Computer Vision (ICCV), 2019. 7 +[29] Mohit Goyal, Sahil Modi, Rishabh Goyal, and Saurabh Gupta. Human hands as probes for interactive object understanding. In Proceedings of the Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 3 +[30] Kristen Grauman, Andrew Westbury, and Eugene Byrne et al. Ego4d: Around the World in 3,000 Hours of Egocentric Video. In Proceedings of the Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 1, 2, 4, 5, 6, 7 +[31] Kristen Grauman, Andrew Westbury, et al. Ego-exo4d: Understanding skilled human activity from first-and third-person perspectives. arXiv preprint arXiv:2311.18259, 2023. 2 +[32] Geoffrey Hinton, Oriol Vinyals, and Jeff Dean. Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531, 2015. 7 +[33] Hsuan-I Ho, Wei-Chen Chiu, and Yu-Chiang Frank Wang. Summarizing first-person videos from third persons' points of view. In Proceedings of the European Conference on Computer Vision (ECCV), 2018. 6 +[34] Nicola J Hodges, A Mark Williams, Spencer J Hayes, and Gavin Breslin. What is modelled during observational learning? Journal of sports sciences, 25(5):531-545, 2007. 1, 4 +[35] Jiang Hua, Liangcai Zeng, Gongfa Li, and Zhaojie Ju. Learning for a robot: Deep reinforcement learning, imitation learning, transfer learning. Sensors, 21(4):1278, 2021. 1 +[36] Chien-Ming Huang, Sean Andrist, Allison Sauppé, and Bilge Mutlu. Using gaze patterns to predict task intent in collaboration. Frontiers in psychology, 6:1049, 2015. 3 +[37] Yifei Huang, Minjie Cai, Zhenqiang Li, and Yoichi Sato. Predicting gaze in egocentric video by learning task-dependent attention transition. In Proceedings of the European Conference on Computer Vision (ECCV), 2018. 3 +[38] Yifei Huang, Minjie Cai, Zhenqiang Li, Feng Lu, and Yoichi Sato. Mutual context network for jointly estimating egocentric gaze and action. IEEE Transactions on Image Processing, 29:7795-7806, 2020. +[39] Yifei Huang, Minjie Cai, and Yoichi Sato. An ego-vision system for discovering human joint attention. IEEE Transactions on Human-Machine Systems, 50(4):306-316, 2020. 3 +[40] Yifei Huang, Lijin Yang, and Yoichi Sato. Compound prototype matching for few-shot action recognition. In Proceedings of the European Conference on Computer Vision (ECCV), 2022. 3 +[41] Yifei Huang, Lijin Yang, and Yoichi Sato. Weakly supervised temporal sentence grounding with uncertainty-guided self-training. In Proceedings of the Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 3 + +[42] Thomas E Hutchinson, K Preston White, Worthy N Martin, Kelly C Reichert, and Lisa A Frey. Human-computer interaction using eye-gaze input. IEEE Transactions on systems, man, and cybernetics, 19(6):1527-1534, 1989. 3 +[43] Eric Jang, Alex Irpan, Mohi Khansari, Daniel Kappler, Frederik Ebert, Corey Lynch, Sergey Levine, and Chelsea Finn. Bc-z: Zero-shot task generalization with robotic imitation learning. In Proceedings of the Conference on Robot Learning, 2022. 1 +[44] Youngkyoon Jang, Brian Sullivan, Casimir Ludwig, Iain Gilchrist, Dima Damen, and Walterio Mayol-Cuevas. Epicent: An egocentric video dataset for camping tent assembly. In Proceedings of the International Conference on Computer Vision Workshops (ICCVW), 2019. 3 +[45] Baoxiong Jia, Yixin Chen, Siyuan Huang, Yixin Zhu, and Song-Chun Zhu. Lemma: A multi-view dataset for the arning multi-agent multi-task activities. In Proceedings of the European Conference on Computer Vision (ECCV), 2020. 2, 5 +[46] Baoxiong Jia, Ting Lei, Song-Chun Zhu, and Siyuan Huang. Egotaskqa: Understanding human tasks in egocentric videos. In Proceedings of the Advances in Neural Information Processing Systems (NeurIPS), 2022. 2 +[47] Hao Jiang and Vamsi Krishna Ithapu. Egocentric pose estimation from human vision span. In Proceedings of the International Conference on Computer Vision (ICCV), 2021. 3 +[48] Soo-Han Kang and Ji-Hyeong Han. Video captioning based on both egocentric and exocentric views of robot vision for human-robot interaction. International Journal of Social Robotics, pages 1–11, 2021. 3 +[49] Moritz Kassner, William Patera, and Andreas Bulling. Pupil: An open source platform for pervasive eye tracking and mobile gaze-based interaction. In Adjunct Proceedings of the 2014 ACM International Joint Conference on Pervasive and Ubiquitous Computing, 2014. 3 +[50] Evangelos Kazakos, Arsha Nagrani, Andrew Zisserman, and Dima Damen. Epic-fusion: Audio-visual temporal binding for egocentric action recognition. In Proceedings of the International Conference on Computer Vision (ICCV), 2019. 3 +[51] Daekyum Kim, Brian Byunghyun Kang, Kyu Bum Kim, Hyungmin Choi, Jeesoo Ha, Kyu-Jin Cho, and Sungho Jo. Eyes are faster than hands: A soft wearable robot learns user intention from the egocentric view. Science Robotics, 4(26), 2019. 3 +[52] Donghyun Kim, Yi-Hsuan Tsai, Bingbing Zhuang, Xiang Yu, Stan Sclaroff, Kate Saenko, and Manmohan Chandraker. Learning cross-modal contrastive features for video domain adaptation. In Proceedings of the International Conference on Computer Vision (ICCV), 2021. 7 +[53] Hildegard Kuehne, Hueihan Jhuang, Estíbaliz Garrote, Tomaso Poggio, and Thomas Serre. Hmdb: a large video database for human motion recognition. In Proceedings of the International Conference on Computer Vision (ICCV), 2011. 2 +[54] Hilde Kuehne, Ali Arslan, and Thomas Serre. The language of actions: Recovering the syntax and semantics of goal + +directed human activities. In Proceedings of the Conference on Computer Vision and Pattern Recognition (CVPR), 2014, 2, 7 +[55] Taein Kwon, Bugra Tekin, Jan Stühmer, Federica Bogo, and Marc Pollefeys. H2o: Two hands manipulating objects for first person interaction recognition. In Proceedings of the International Conference on Computer Vision (ICCV), 2021. 2, 5 +[56] Bolin Lai, Miao Liu, Fiona Ryan, and James M Rehg. In the eye of transformer: Global-local correlation for egocentric gaze estimation and beyond. International Journal of Computer Vision (IJCV), pages 1-18, 2023. 3 +[57] Stephane Lallée, Eiichi Yoshida, Anthony Mallet, Francesco Nori, Lorenzo Natale, Giorgio Metta, Felix Warneken, and Peter Ford Dominey. Human-robot cooperation based on interaction learning. From motor learning to interaction learning in robots, pages 491–536, 2010. 1 +[58] Kyle Lam, Junhong Chen, Zeyu Wang, Fahad M Iqbal, Ara Darzi, Benny Lo, Sanjay Purkayastha, and James M Kinross. Machine learning for technical skill assessment in surgery: a systematic review. NPJ digital medicine, 5(1):24, 2022. 8 +[59] Jangwon Lee and Michael S Ryoo. Learning robot activities from first-person human videos using convolutional future regression. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), 2017. 1 +[60] Jie Lei, Linjie Li, Luowei Zhou, Zhe Gan, Tamara L Berg, Mohit Bansal, and Jingjing Liu. Less is more: Clipbert for video-and-language learning via sparse sampling. In Proceedings of the Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 6 +[61] Chengshu Li, Fei Xia, Roberto Martin-Martin, Michael Lingelbach, Sanjana Srivastava, Bokui Shen, Kent Vainio, Cem Gokmen, Gokul Dharan, Tanish Jain, et al. igibson 2.0: Object-centric simulation for robot learning of everyday household tasks. arXiv preprint arXiv:2108.03272, 2021. 2 +[62] Chengshu Li, Ruohan Zhang, Josiah Wong, Cem Gokmen, Sanjana Srivastava, Roberto Martin-Martin, Chen Wang, Gabrael Levine, Michael Lingelbach, Jiankai Sun, et al. Behavior-1k: A benchmark for embodied ai with 1,000 everyday activities and realistic simulation. In Proceedings of the Conference on Robot Learning, 2023. 5 +[63] Haoxin Li, Yijun Cai, and Wei-Shi Zheng. Deep dual relation modeling for egocentric interaction recognition. In Proceedings of the Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 3 +[64] Jiayi Li, Tao Lu, Xiaoge Cao, Yinghao Cai, and Shuo Wang. Meta-imitation learning by watching video demonstrations. In Proceedings of the International Conference on Learning Representations, 2021. 1 +[65] Yin Li, Alireza Fathi, and James M Rehg. Learning to predict gaze in egocentric video. In Proceedings of the International Conference on Computer Vision (ICCV), 2013. 3 +[66] Yin Li, Zhefan Ye, and James M Rehg. Delving into egocentric actions. In Proceedings of the Conference on Computer Vision and Pattern Recognition (CVPR), 2015. + +[67] Yin Li, Miao Liu, and James M Rehg. In the eye of beholder: Joint learning of gaze and actions in first person video. In Proceedings of the European Conference on Computer Vision (ECCV), 2018. 3, 5 +[68] Yin Li, Miao Liu, and Jame Rehg. In the eye of the beholder: Gaze and actions in first person video. IEEE Transaction on Pattern Analysis and Machine Intelligence (TPAMI), 2021. 3 +[69] Yanghao Li, Tushar Nagarajan, Bo Xiong, and Kristen Grauman. Ego-exo: Transferring visual representations from third-person to first-person videos. In Proceedings of the Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 3, 6, 7 +[70] Zhenqiang Li, Yifei Huang, Minjie Cai, and Yoichi Sato. Manipulation-skill assessment from videos with spatial attention network. In Proceedings of the International Conference on Computer Vision Workshops (ICCVW), 2019. 8 +[71] Zhenqiang Li, Lin Gu, Weimin Wang, Ryosuke Nakamura, and Yoichi Sato. Surgical skill assessment via video semantic aggregation. In Proceedings of the International Conference on Medical Image Computing and Computer-Assisted Intervention (MICCAI), 2022. 8 +[72] Kevin Qinghong Lin, Jinpeng Wang, Mattia Soldan, Michael Wray, Rui Yan, Eric Z XU, Difei Gao, Rong-Cheng Tu, Wenzhe Zhao, Weijie Kong, et al. Egocentric video-language pretraining. In Proceedings of the Advances in Neural Information Processing Systems (NeurIPS), 2022. 6 +[73] Daochang Liu, Qiyue Li, Tingting Jiang, Yizhou Wang, Rulin Miao, Fei Shan, and Ziyu Li. Towards unified surgical skill assessment. In Proceedings of the Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 8 +[74] Shaowei Liu, Subarna Tripathi, Somdeb Majumdar, and Xiaolong Wang. Joint hand motion and interaction hotspots prediction from egocentric videos. In Proceedings of the Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 3 +[75] Xiaotian Liu, Hector Palacios, and Christian Muise. Egocentric planning for scalable embodied task achievement. arXiv preprint arXiv:2306.01295, 2023. 1 +[76] Yan Liu, Pei Yun Hsueh, Jennifer Lai, Mirweis Sangin, Marc Antoine Nussli, and Pierre Dillenbourg. Who is the expert? analyzing gaze data to predict expertise level in collaborative applications. In Proceedings of the IEEE International Conference on Multimedia & Expo (ICME), 2009. 8 +[77] Yueyue Liu, Zhijun Li, Huaping Liu, and Zhen Kan. Skill transfer learning for autonomous robots and human-robot cooperation: A survey. Robotics and Autonomous Systems, 128:103515, 2020. 1 +[78] Yunze Liu, Yun Liu, Che Jiang, Kangbo Lyu, Weikang Wan, Hao Shen, Boqiang Liang, Zhoujie Fu, He Wang, and Li Yi. Hoi4d: A 4d egocentric dataset for category-level human-object interaction. In Proceedings of the Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2 +[79] Manolis Savva*, Abhishek Kadian*, Oleksandr Maksymets*, Yili Zhao, Erik Wijmans, Bhavana Jain, Julian Straub, Jia Liu, Vladlen Koltun, Jitendra Malik, Devi Parikh, and Dhruv Batra. Habitat: A Platform for Embodied AI + +Research. In Proceedings of the International Conference on Computer Vision (ICCV), 2019. 1 +[80] Javier Marina-Miranda and V Javier Traver. Head and eye egocentric gesture recognition for human-robot interaction using eyewear cameras. IEEE Robotics and Automation Letters, 7(3):7067-7074, 2022. 3 +[81] Joanna Materzynska, Guillaume Berger, Ingo Bax, and Roland Memisevic. The jester dataset: A large-scale video dataset of human gestures. In Proceedings of the International Conference on Computer Vision Workshops (ICCVW), 2019. 2 +[82] Antoine Miech, Dimitri Zhukov, Jean-Baptiste Alayrac, Makarand Tapaswi, Ivan Laptev, and Josef Sivic. HowTo100M: Learning a Text-Video Embedding by Watching Hundred Million Narrated Video Clips. In Proceedings of the International Conference on Computer Vision (ICCV), 2019. 1, 6 +[83] Kyle Min and Jason J Corso. Integrating human gaze into attention for egocentric activity recognition. In Proceedings of the Winter Conference on Applications of Computer Vision (WACV), pages 1069-1078, 2021. 3 +[84] Jonathan Munro and Dima Damen. Multi-modal domain adaptation for fine-grained action recognition. In Proceedings of the Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 7 +[85] Tushar Nagarajan and Kristen Grauman. Shaping embodied agent behavior with activity-context priors from egocentric video. In Proceedings of the Advances in Neural Information Processing Systems (NeurIPS), 2021. 3 +[86] Tushar Nagarajan, Yanghao Li, Christoph Feichtenhofer, and Kristen Grauman. Ego-topo: Environment affordances from egocentric video. In Proceedings of the Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 3 +[87] Boxiao Pan, Haoye Cai, De-An Huang, Kuan-Hui Lee, Adrien Gaidon, Ehsan Adeli, and Juan Carlos Niebles. Spatio-temporal graph for video captioning with knowledge distillation. In Proceedings of the Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 7 +[88] Boxiao Pan, Zhangjie Cao, Ehsan Adeli, and Juan Carlos Niebles. Adversarial cross-domain action recognition with co-attention. In Proceedings of the AAAI Conference on Artificial Intelligence (AAAI), 2020. 2 +[89] Boxiao Pan, Bokui Shen, Davis Rempe, Despoina Paschalidou, Kaichun Mo, Yanchao Yang, and Leonidas J Guibas. Copilot: Human-environment collision prediction and localization from egocentric videos. In Proceedings of the International Conference on Computer Vision (ICCV), 2023. 1 +[90] Hyung Min Park, Seok Han Lee, and Jong Soo Choi. Wearable augmented reality system using gaze interaction. In Proceedings of the IEEE/ACM International Symposium on Mixed and Augmented Reality, 2008. 3 +[91] Paritosh Parmar and Brendan Tran Morris. What and how well you performed? a multitask learning approach to action quality assessment. In Proceedings of the Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 8 + +[92] Leo Pauly, Wisdom C Agboh, David C Hogg, and Raul Fuentes. O2a: one-shot observational learning with action vectors. Frontiers in Robotics and AI, 2021. 1 +[93] Chiara Plizzari, Mirco Planamente, Gabriele Goletto, Marco Cannici, Emanuele Gusso, Matteo Matteucci, and Barbara Caputo. E2 (go) motion: Motion augmented event stream for egocentric action recognition. In Proceedings of the Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 3 +[94] Chiara Plizzari, Gabriele Goletto, Antonino Furnari, Siddhant Bansal, Francesco Ragusa, Giovanni Maria Farinella, Dima Damen, and Tatiana Tommasi. An outlook into the future of egocentric vision. arXiv preprint arXiv:2308.07123, 2023. 1, 3 +[95] David Premack and Guy Woodruff. Does the chimpanzee have a theory of mind? Behavioral and brain sciences, 1(4): 515-526, 1978. 1 +[96] Gorjan Radevski, Dusan Grujicic, Matthew Blaschko, Marie-Francine Moens, and Tinne Tuytelaars. Multimodal distillation for egocentric action recognition. In Proceedings of the International Conference on Computer Vision (ICCV), 2023. 3 +[97] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In Proceedings of the International Conference on Machine Learning (ICML), 2021. 5, 7 +[98] Francesco Ragusa, Antonino Furnari, Salvatore Livatino, and Giovanni Maria Farinella. The meccano dataset: Understanding human-object interactions from egocentric videos in an industrial-like domain. In Proceedings of the Winter Conference on Applications of Computer Vision (WACV), 2021. 2, 5 +[99] Nishant Rai, Haofeng Chen, Jingwei Ji, Rishi Desai, Kazuki Kozuka, Shun Ishizaka, Ehsan Adeli, and Juan Carlos Niebles. Home action genome: Cooperative compositional action understanding. In Proceedings of the Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 1, 2, 5 +[100] Richard Ramsey, David M Kaplan, and Emily S Cross. Watch and learn: the cognitive neuroscience of learning from others' actions. Trends in Neurosciences, 44(6):478-491, 2021. 1 +[101] Yosef Razin and Karen Feigh. Learning to predict intent from gaze during robotic hand-eye coordination. In Proceedings of the AAAI Conference on Artificial Intelligence (AAAI), 2017. 3 +[102] Radiah Rivu, Yasmeen Abdrabou, Ken Pfeuffer, Augusto Esteves, Stefanie Meitner, and Florian Alt. Stare: gaze-assisted face-to-face communication in augmented reality. In Proceedings of the ACM Symposium on Eye Tracking Research and Applications (ETRA), 2020. 3 +[103] Giacomo Rizzolatti and Laila Craighero. The mirror-neuron system. Annual Review of Neuroscience, 27:169-192, 2004. 1 +[104] Aadarsh Sahoo, Rutav Shah, Rameswar Panda, Kate Saenko, and Abir Das. Contrast and mix: Temporal contrastive video + +domain adaptation with background mixing. In Proceedings of the Advances in Neural Information Processing Systems (NeurIPS), 2021. 7 +[105] Stefan Schaal. Learning from demonstration. In Proceedings of the Advances in Neural Information Processing Systems (NeurIPS), 1996. 1 +[106] F. Sener, D. Chatterjee, D. Shelepov, K. He, D. Singhania, R. Wang, and A. Yao. Assembly101: A large-scale multiview video dataset for understanding procedural activities. In Proceedings of the Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 1, 2, 4, 5, 7 +[107] Dandan Shan, Jiaqi Geng, Michelle Shu, and David F Fouhey. Understanding human hands in contact at internet scale. In Proceedings of the Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 3 +[108] Mohit Shridhar, Jesse Thomason, Daniel Gordon, Yonatan Bisk, Winson Han, Roozbeh Mottaghi, Luke Zettlemoyer, and Dieter Fox. ALFRED: A Benchmark for Interpreting Grounded Instructions for Everyday Tasks. In Proceedings of the Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 1 +[109] Gunnar A Sigurdsson, Abhinav Gupta, Cordelia Schmid, Ali Farhadi, and Karteek Alahari. Charades-ego: A large-scale dataset of paired third and first person videos. arXiv preprint arXiv:1804.09626, 2018. 1, 2, 5 +[110] Kiran Somasundaram, Jing Dong, Huixuan Tang, Julian Straub, Mingfei Yan, Michael Goesele, Jakob Julian Engel, Renzo De Nardi, and Richard Newcombe. Project aria: A new tool for egocentric multi-modal ai research. arXiv preprint arXiv:2308.13561, 2023. 2 +[111] Xiaolin Song, Sicheng Zhao, Jingyu Yang, Huanjing Yue, Pengfei Xu, Runbo Hu, and Hua Chai. Spatio-temporal contrastive domain adaptation for action recognition. In Proceedings of the Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 7 +[112] Yale Song, Eugene Byrne, Tushar Nagarajan, Huiyu Wang, Miguel Martin, and Lorenzo Torresani. Ego4d goal-step: Toward hierarchical understanding of procedural activities. Proceedings of the Advances in Neural Information Processing Systems (NeurIPS), 2024. 2 +[113] Swathikiran Sudhakaran, Sergio Escalera, and Oswald Lanz. Lsta: Long short-term attention for egocentric action recognition. In Proceedings of the Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 3 +[114] Flood Sung, Yongxin Yang, Li Zhang, Tao Xiang, Philip HS Torr, and Timothy M Hospedales. Learning to compare: Relation network for few-shot learning. In Proceedings of the Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 8 +[115] Yansong Tang, Zanlin Ni, Jiahuan Zhou, Danyang Zhang, Jiwen Lu, Ying Wu, and Jie Zhou. Uncertainty-aware score distribution learning for action quality assessment. In Proceedings of the Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 8 +[116] Yuli Vasiliev. Natural language processing with Python and spaCy: A practical introduction. 2020. 4 +[117] Shruti Vyas, Yogesh S Rawat, and Mubarak Shah. Multiview action recognition using cross-view video prediction. + +In Proceedings of the European Conference on Computer Vision (ECCV), 2020. 7 +[118] Sebastian Wallkötter, Silvia Tulli, Ginevra Castellano, Ana Paiva, and Mohamed Chetouani. Explainable embodied agents through social cues: a review. ACM Transactions on Human-Robot Interaction (THRI), 10(3):1-24, 2021. 5 +[119] Chen Wang, Linxi Fan, Jiankai Sun, Ruohan Zhang, Li Fei-Fei, Danfei Xu, Yuke Zhu, and Anima Anandkumar. Mimicplay: Long-horizon imitation learning by watching human play. arXiv preprint arXiv:2302.12422, 2023. 1 +[120] Huiyu Wang, Mitesh Kumar Singh, and Lorenzo Torresani. Ego-only: Egocentric action detection without exocentric transferring. In Proceedings of the International Conference on Computer Vision (ICCV), 2023. 3 +[121] Jiahao Wang, Guo Chen, Yifei Huang, Limin Wang, and Tong Lu. Memory-and-anticipation transformer for online action understanding. In Proceedings of the International Conference on Computer Vision (ICCV), 2023. 3 +[122] Jinpeng Wang, Yixiao Ge, Rui Yan, Yuying Ge, Kevin Qinghong Lin, Satoshi Tsutsui, Xudong Lin, Guanyu Cai, Jianping Wu, Ying Shan, et al. All in one: Exploring unified video-language pre-training. In Proceedings of the Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 6 +[123] Jian Wang, Diogo Luvizon, Weipeng Xu, Lingjie Liu, Kripasindhu Sarkar, and Christian Theobalt. Scene-aware egocentric 3d human pose estimation. In Proceedings of the Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 3 +[124] Qitong Wang, Long Zhao, Liangzhe Yuan, Ting Liu, and Xi Peng. Learning from semantic alignment between unpaired multiviews for egocentric video recognition. In Proceedings of the International Conference on Computer Vision (ICCV), 2023. 1, 2, 3, 6, 7 +[125] Xiaohan Wang, Linchao Zhu, Heng Wang, and Yi Yang. Interactive prototype learning for egocentric action recognition. In Proceedings of the International Conference on Computer Vision (ICCV), 2021. 3 +[126] Xin Wang, Taein Kwon, Mahdi Rad, Bowen Pan, Ishani Chakraborty, Sean Andrist, Dan Bohus, Ashley Feniello, Bugra Tekin, Felipe Vieira Frujeri, Neel Joshi, and Marc Pollefeys. Holoassist: an egocentric human interaction dataset for interactive ai assistants in the real world. In Proceedings of the International Conference on Computer Vision (ICCV), 2023. 1, 4, 5 +[127] Xiaohan Wang, Yuehu Liu, Xinhang Song, Beibei Wang, and Shuqiang Jiang. Generating explanations for embodied action decision from visual observation. In Proceedings of the 31st ACM International Conference on Multimedia, 2023. 5 +[128] Yeping Wang, Gopika Ajaykumar, and Chien-Ming Huang. See what i see: Enabling user-centric robotic assistance using first-person demonstrations. In Proceedings of the 2020 ACM/IEEE International Conference on Human-Robot Interaction, pages 639-648, 2020. 1 +[129] Yi Wang, Kunchang Li, Yizhuo Li, Yinan He, Bingkun Huang, Zhiyu Zhao, Hongjie Zhang, Jilan Xu, Yi Liu, Zun + +Wang, et al. Intervideo: General video foundation models via generative and discriminative learning. arXiv preprint arXiv:2212.03191, 2022. 6 +[130] Zirui Wang, Jiahui Yu, Adams Wei Yu, Zihang Dai, Yulia Tsvetkov, and Yuan Cao. Simvlm: Simple visual language model pretraining with weak supervision. arXiv preprint arXiv:2108.10904, 2021. 6 +[131] Pengfei Wei, Lingdong Kong, Xinghua Qu, Yi Ren, Jing Jiang, Xiang Yin, et al. Unsupervised video domain adaptation for action recognition: A disentanglement perspective. In Proceedings of the Advances in Neural Information Processing Systems (NeurIPS), 2023. 2, 7 +[132] Daniel Weinland, Mustafa Özuysal, and Pascal Fua. Making action recognition robust to occlusions and viewpoint changes. In Proceedings of the European Conference on Computer Vision (ECCV), 2010. 3 +[133] Mark Wilson, John Mcgrath, Samuel Vine, James Brewer, David Defriend, and Richard Masters. Psychomotor control in a virtual laparoscopic surgery training environment: gaze control parameters differentiate novices from experts. Surgical Endoscopy, 24(10):2458-2464, 2010. 8 +[134] Benita Wong, Joya Chen, You Wu, Stan Weixian Lei, Dongxing Mao, Difei Gao, and Mike Zheng Shou. Assistq: Affordance-centric question-driven task completion for egocentric assistant. In Proceedings of the European Conference on Computer Vision (ECCV), 2022. 1 +[135] Xinxiao Wu, Han Wang, Cuiwei Liu, and Yunde Jia. Cross-view action recognition over heterogeneous feature spaces. In Proceedings of the International Conference on Computer Vision (ICCV), 2013. 6 +[136] Haifeng Xia, Pu Wang, and Zhengming Ding. Incomplete multi-view domain adaptation via channel enhancement and knowledge transfer. In Proceedings of the European Conference on Computer Vision (ECCV), 2022. 3 +[137] Jianjia Xin, Lichun Wang, Kai Xu, Chao Yang, and Baocai Yin. Learning interaction regions and motion trajectories simultaneously from egocentric demonstration videos. IEEE Robotics and Automation Letters, 2023. 3 +[138] Hu Xu, Gargi Ghosh, Po-Yao Huang, Dmytro Okhonko, Armen Aghajanyan, Florian Metze, Luke Zettlemoyer, and Christoph Feichtenhofer. Videoclip: Contrastive pre-training for zero-shot video-text understanding. arXiv preprint arXiv:2109.14084, 2021. 6 +[139] Jia Xu, Lopamudra Mukherjee, Yin Li, Jamieson Warner, James M Rehg, and Vikas Singh. Gaze-enabled egocentric video summarization via constrained submodular maximization. In Proceedings of the Conference on Computer Vision and Pattern Recognition (CVPR), 2015. 3 +[140] Jilan Xu, Yifei Huang, Junlin Hou, Guo Chen, Yuejie Zhang, Rui Feng, and Weidi Xie. Retrieval-augmented egocentric video captioning. arXiv preprint arXiv:2401.00789, 2024.3 +[141] Mingze Xu, Chenyou Fan, Yuchen Wang, Michael S Ryoo, and David J Crandall. Joint person segmentation and identification in synchronized first-and third-person videos. In Proceedings of the European Conference on Computer Vision (ECCV), 2018. 6 +[142] Hongwei Xue, Tiankai Hang, Yanhong Zeng, Yuchong Sun, Bei Liu, Huan Yang, Jianlong Fu, and Baining Guo. + +Advancing high-resolution video-language representation with large-scale video transcriptions. In Proceedings of the Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 3 +[143] Zihui Xue and Kristen Grauman. Learning fine-grained view-invariant representations from unpaired ego-exo videos via temporal alignment. In Proceedings of the Advances in Neural Information Processing Systems (NeurIPS), 2023. 1, 2, 6, 7 +[144] Takuma Yagi, Misaki Ohashi, Yifei Huang, Ryosuke Furuta, Shungo Adachi, Toutai Mitsuyama, and Yoichi Sato. Finebio: A fine-grained video dataset of biological experiments with hierarchical annotations. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), 2023. 2 +[145] Lijin Yang, Yifei Huang, Yusuke Sugano, and Yoichi Sato. Interact before align: Leveraging cross-modal knowledge for domain adaptive action recognition. In Proceedings of the Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 7 +[146] Sherry Yang, Jacob Walker, Jack Parker-Holder, Yilun Du, Jake Bruce, Andre Barreto, Pieter Abbeel, and Dale Schuurmans. Video as the new language for real-world decision making. arXiv preprint arXiv:2402.17139, 2024. 5 +[147] Li Yingzhen and Stephan Mandt. Disentangled sequential autoencoder. In Proceedings of the International Conference on Machine Learning (ICML), 2018. 2 +[148] Tianhe Yu, Chelsea Finn, Annie Xie, Sudeep Dasari, Tianhao Zhang, Pieter Abbeel, and Sergey Levine. One-shot imitation from observing humans via domain-adaptive meta-learning. arXiv preprint arXiv:1802.01557, 2018. 1 +[149] Xumin Yu, Yongming Rao, Wenliang Zhao, Jiwen Lu, and Jie Zhou. Group-aware contrastive regression for action quality assessment. In Proceedings of the International Conference on Computer Vision (ICCV), 2021. 8 +[150] Zecheng Yu, Yifei Huang, Ryosuke Furuta, Takuma Yagi, Yusuke Goutsu, and Yoichi Sato. Fine-grained affordance annotation for egocentric hand-object interaction videos. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, 2023. 2 +[151] Thorsten O Zander, Matti Gaertner, Christian Kothe, and Roman Vilimek. Combining eye gaze input with a brain-computer interface for touchless human-computer interaction. International Journal of Human-Computer Interaction, 27(1):38-51, 2010. 3 +[152] Rowan Zellers, Ximing Lu, Jack Hessel, Youngjae Yu, Jae Sung Park, Jize Cao, Ali Farhadi, and Yejin Choi. Merlot: Multimodal neural script knowledge models. In Proceedings of the Advances in Neural Information Processing Systems (NeurIPS), 2021. 3 +[153] Lingzhi Zhang, Shenghao Zhou, Simon Stent, and Jianbo Shi. Fine-grained egocentric hand-object segmentation: Dataset, model, and applications. In Proceedings of the European Conference on Computer Vision (ECCV), 2022. 3 +[154] Siwei Zhang, Qianli Ma, Yan Zhang, Zhiyin Qian, Taein Kwon, Marc Pollefeys, Federica Bogo, and Siyu Tang. Ego-body: Human body shape and motion of interacting people + +from head-mounted devices. In Proceedings of the European Conference on Computer Vision (ECCV), 2022. 2, 3 +[155] Weiyu Zhang, Menglong Zhu, and KG Derpanis. From actemes to action: A strongly supervised representation for detailed action understanding. In Proceedings of the International Conference on Computer Vision (ICCV), 2013. +[156] Zehua Zhang, David Crandall, Michael Proulx, Sachin Talathi, and Abhishek Sharma. Can gaze inform egocentric action recognition? In Proceedings of the Symposium on Eye Tracking Research and Applications (ETRA), 2022. 3 +[157] Zhenning Zhang, Zhigeng Pan, Weiqing Li, and Zhiyong Su. X-board: an egocentric adaptive ar assistant for perception in indoor environments. Virtual Reality, 27(2):1327-1343, 2023. 5 +[158] Yue Zhao, Ishan Misra, Philipp Krahenbuhl, and Rohit Girdhar. Learning video representations from large language models. In Proceedings of the Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 6 +[159] Jingjing Zheng and Zhuolin Jiang. Learning view-invariant sparse representations for cross-view action recognition. In Proceedings of the International Conference on Computer Vision (ICCV), 2013. 7 +[160] Yang Zheng, Yanchao Yang, Kaichun Mo, Jiaman Li, Tao Yu, Yebin Liu, Karen Liu, and Leonidas J Guibas. Gimo: Gaze-informed human motion prediction in context. In Proceedings of the European Conference on Computer Vision (ECCV), 2022. 3 \ No newline at end of file diff --git a/2024/EgoExoLearn_ A Dataset for Bridging Asynchronous Ego- and Exo-centric View of Procedural Activities in Real World/images.zip b/2024/EgoExoLearn_ A Dataset for Bridging Asynchronous Ego- and Exo-centric View of Procedural Activities in Real World/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..a65bd77c0c4f9624377be754d6b195d1c7487936 --- /dev/null +++ b/2024/EgoExoLearn_ A Dataset for Bridging Asynchronous Ego- and Exo-centric View of Procedural Activities in Real World/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a83c006b48cb828f818bfc5566e0bf2a5b7418114fbf753cfc5431b4b2b72ce3 +size 493803 diff --git a/2024/EgoExoLearn_ A Dataset for Bridging Asynchronous Ego- and Exo-centric View of Procedural Activities in Real World/layout.json b/2024/EgoExoLearn_ A Dataset for Bridging Asynchronous Ego- and Exo-centric View of Procedural Activities in Real World/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..56f9630731171ec5ab7e98e03b6e0ef92e99cd6e --- /dev/null +++ b/2024/EgoExoLearn_ A Dataset for Bridging Asynchronous Ego- and Exo-centric View of Procedural Activities in Real World/layout.json @@ -0,0 +1,10278 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 94, + 103, + 500, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 94, + 103, + 500, + 140 + ], + "spans": [ + { + "bbox": [ + 94, + 103, + 500, + 140 + ], + "type": "text", + "content": "EgoExoLearn: A Dataset for Bridging Asynchronous Ego- and Exo-centric View of Procedural Activities in Real World" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 107, + 160, + 487, + 219 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 107, + 160, + 487, + 219 + ], + "spans": [ + { + "bbox": [ + 107, + 160, + 487, + 219 + ], + "type": "text", + "content": "Yifei Huang†‡, Guo Chen‡, Jilan Xu‡, Mingfang Zhang‡, Lijin Yang, Baoqi Pei Hongjie Zhang, Lu Dong, Yali Wang*, Limin Wang*, Yu Qiao* OpenGVLab, Shanghai AI Laboratory Shenzhen Institutes of Advanced Technology, CAS Nanjing University" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 143, + 245, + 192, + 258 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 245, + 192, + 258 + ], + "spans": [ + { + "bbox": [ + 143, + 245, + 192, + 258 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 261, + 290, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 261, + 290, + 548 + ], + "spans": [ + { + "bbox": [ + 46, + 261, + 290, + 548 + ], + "type": "text", + "content": "Being able to map the activities of others into one's own point of view is a fundamental human skill even from a very early age. Taking a step toward understanding this human ability, we introduce EgoExoLearn, a large-scale dataset that emulates the human demonstration following process, in which individuals record egocentric videos as they execute tasks guided by exocentric-view demonstration videos. Focusing on the potential applications in daily assistance and professional support, EgoExoLearn contains egocentric and demonstration video data spanning 120 hours captured in daily life scenarios and specialized laboratories. Along with the videos we record high-quality gaze data and provide detailed multimodal annotations, formulating a playground for modeling the human ability to bridge asynchronous procedural actions from different viewpoints. To this end, we present benchmarks such as cross-view association, cross-view action planning, and cross-view referenced skill assessment, along with detailed analysis. We expect EgoExoLearn can serve as an important resource for bridging the actions across views, thus paving the way for creating AI agents capable of seamlessly learning by observing humans in the real world. The dataset and benchmark codes are available at https://github.com/OpenGVLab/EgoExoLearn." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 563, + 128, + 575 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 563, + 128, + 575 + ], + "spans": [ + { + "bbox": [ + 47, + 563, + 128, + 575 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 583, + 288, + 679 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 583, + 288, + 679 + ], + "spans": [ + { + "bbox": [ + 46, + 583, + 288, + 679 + ], + "type": "text", + "content": "Even as a child, humans can observe the actions of others and then map them to their own view [4, 34, 100, 105]. With this ability to asynchronously bridge activities from egocentric and exocentric views [95, 103], humans can watch others' demonstrations and replicate the procedures in a new environment. This ability is especially beneficial when actual physical trials carry the potential of high costs [24], e.g., conducting dangerous chemical experiments." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 58, + 681, + 287, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 681, + 287, + 693 + ], + "spans": [ + { + "bbox": [ + 58, + 681, + 287, + 693 + ], + "type": "text", + "content": "In the wake of recent advancements in AI systems, one" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 247, + 547, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 247, + 547, + 413 + ], + "spans": [ + { + "bbox": [ + 304, + 247, + 547, + 413 + ], + "type": "text", + "content": "goal for the next generation of AI agents is to perform tasks in a more embodied setting [94]. However, different from humans, training these AI agents usually requires demonstration videos taken in a similar environment [75, 119] and from a congruent perspective with the AI agents, (e.g., the egocentric point of view [43, 59, 108, 134]). While great effort has been made into the collection of egocentric data in different scenarios [18, 30, 106], it remains crucial for the AI agents to directly learn from demonstration videos taken in a different place and from a different viewpoint [35, 148]. Realizing this capability can unleash the full potential of public instructional video data [82] and is also useful in the human-robot cooperation scenario, especially in novel environments [57, 77, 128]." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 414, + 548, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 414, + 548, + 605 + ], + "spans": [ + { + "bbox": [ + 304, + 414, + 548, + 605 + ], + "type": "text", + "content": "Current works towards this goal can be roughly divided into two directions. One way is to learn models in simulated environments [11, 64, 79, 89, 92], but it remains difficult for models in this setting to generalize in the real world [126]. The other direction is to learn from human activity in real-world scenarios. However, attempts to directly combine existing multiview datasets often yield datasets of inferior quality or scale [124, 143]. Meanwhile, the few existing datasets in this direction [99, 106, 109] only record ego- and exo-view videos in the same environment and in a time-synchronized manner. In reality, when following demonstrations it is often needed to bridge a series of procedural actions performed in a different place and at a different time. However, currently, no dataset is available for the exploration of how to bridge asynchronous procedural activities in realistic egocentric and exocentric viewpoints." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 605, + 548, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 548, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 548, + 715 + ], + "type": "text", + "content": "To address this lack of dataset issue, we introduce EgoExoLearn, a large-scale dataset containing demonstration videos and corresponding egocentric videos where the camera wearers follow the demonstrations and perform the same task in a different environment, as shown in Figure 1. Targeting two potential applications, i.e., daily assistance and professional support, EgoExoLearn consists of 747 video sequences spanning a total of 120 hours of footage, ranging from daily food-making to specialized laboratory experi" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 702, + 261, + 714 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 702, + 261, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 702, + 261, + 714 + ], + "type": "text", + "content": "*Corresponding authors. †Project lead. ‡Equal key contributions." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "22072" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 51, + 58, + 545, + 220 + ], + "blocks": [ + { + "bbox": [ + 51, + 58, + 545, + 220 + ], + "lines": [ + { + "bbox": [ + 51, + 58, + 545, + 220 + ], + "spans": [ + { + "bbox": [ + 51, + 58, + 545, + 220 + ], + "type": "image", + "image_path": "31594f139b11ce421ea3883f07660feaa1a57473de5d0f9f5fd264319f7ffe46.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 229, + 548, + 275 + ], + "lines": [ + { + "bbox": [ + 46, + 229, + 548, + 275 + ], + "spans": [ + { + "bbox": [ + 46, + 229, + 548, + 275 + ], + "type": "text", + "content": "Figure 1. EgoExoLearn emulates the human asynchronous demonstration following process. It contains demonstration videos of multiple tasks, together with egocentric videos recorded by participants replicating the procedure after watching the demonstrations. The dataset comprises gaze signals and fine-grained multi-level multi-modal annotations, enabling the exploration of key features in this context such as cross-view association and cross-view action planning." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 285, + 287, + 332 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 285, + 287, + 332 + ], + "spans": [ + { + "bbox": [ + 46, + 285, + 287, + 332 + ], + "type": "text", + "content": "ments. Notably, the egocentric videos in EgoExoLearn contain eye gaze signals showing humans' visual attention while performing the task. This provides a valuable cue for better bridging the actions in ego- and exo-viewpoints." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 333, + 290, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 333, + 290, + 475 + ], + "spans": [ + { + "bbox": [ + 46, + 333, + 290, + 475 + ], + "type": "text", + "content": "We take one more step forward by analyzing human ability in bridging asynchronous ego- and exo-view actions and, accordingly, introduce new tasks and benchmarks that we believe can form building blocks for the development of next-stage embodied AI agents with similar abilities. When humans perform an action, he/she can associate and describe the undergoing action in the egocentric view with the corresponding action in the demonstration. With the knowledge from demonstration videos, humans can know the needed action steps and predict what the next steps should be. Besides, through the comparison with the demonstration, humans can also assess their level of skills." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 476, + 288, + 619 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 476, + 288, + 619 + ], + "spans": [ + { + "bbox": [ + 46, + 476, + 288, + 619 + ], + "type": "text", + "content": "Based on the above analysis, we design benchmarks of 1) cross-view association, 2) cross-view action understanding, 3) cross-view referenced skill assessment, and 4) cross-view referenced video captioning. Each benchmark is meticulously defined, annotated, and supported by baseline implementations. In addition, we pioneeringly explore the role of gaze in these tasks. We hope our dataset can provide resources for future work for bridging asynchronous procedural actions in ego- and exo-centric perspectives, thereby inspiring the design of AI agents adept at learning from real-world human demonstrations and mapping the procedural actions into robot-centric views." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 631, + 134, + 643 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 631, + 134, + 643 + ], + "spans": [ + { + "bbox": [ + 47, + 631, + 134, + 643 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 654, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 654, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 654, + 289, + 715 + ], + "type": "text", + "content": "Ego-exo datasets. While there exist works that associate existing datasets to explore how activities can be bridged between them, these associated datasets are often limited in scale [14, 144, 147] or quality [124], meanwhile focusing only on single actions captured from the same view [81, 88," + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 285, + 548, + 429 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 285, + 548, + 429 + ], + "spans": [ + { + "bbox": [ + 304, + 285, + 548, + 429 + ], + "type": "text", + "content": "131]. As for actions from different views, apart from multiview fixed camera datasets [6, 15, 54, 61], there also exist datasets with both ego- exo-centric view videos [30, 31, 99, 106, 109]. These datasets are either recorded in the same environment [45, 99, 109] or record time-synced multi-view videos in the same environment with primary focuses on pose/activity understanding grounded in the 3D world [55, 106, 154]. Our dataset offers a more challenging and realistic scenario, where egocentric camera wearers learn to complete the tasks demonstrated by exocentric demonstration videos. This setting complements these datasets by focusing on high-level procedural actions." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 432, + 548, + 612 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 432, + 548, + 612 + ], + "spans": [ + { + "bbox": [ + 304, + 432, + 548, + 612 + ], + "type": "text", + "content": "The only dataset conceptually similar to ours is the recently proposed AE2 dataset [143], where the goal is to learn view-invariant representation from unpaired ego and exo videos. This dataset combines ego and exo videos from five public datasets [18, 20, 53, 55, 155] and a newly collected ego tennis forehand dataset. However, due to the difficulty in associating existing ego-exo datasets, the AE2 dataset is relatively small where the largest subset contains only 322 clips. Also, this dataset only focuses on clip-level actions, and thus cannot feature the real-world demonstration following setting, which usually requires multimodal, task-centric procedural knowledge. Instead, our EgoExoLearn is much larger in scale (100x more clips), while offering gaze and fine-grained multimodal annotations facilitating multi-faceted analysis of ego-exo action understanding." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 618, + 548, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 618, + 548, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 618, + 548, + 715 + ], + "type": "text", + "content": "Egocentric video datasets. In line with the recent development in wearable cameras [110], multiple egocentric video datasets [5, 17, 19, 30, 46, 78, 98, 112, 150] have been proposed. Different from previous egocentric datasets, the egocentric videos in EgoExoLearn feature a demonstration-following setting. We believe EgoExoLearn provides a playground for developing tools to bridge asynchronous procedural activities from ego- and exocentric viewpoints." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "22073" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 288, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 288, + 108 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 288, + 108 + ], + "type": "text", + "content": "The setting of our EgoExoLearn complements existing datasets like Ego4D and can benefit from their rich knowledge and representations." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 112, + 288, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 112, + 288, + 266 + ], + "spans": [ + { + "bbox": [ + 46, + 112, + 288, + 266 + ], + "type": "text", + "content": "Egocentric gaze. Gaze can indicate visual attention and contains valuable information about human intent [36, 156], thus is used in a diverse range of areas such as human-computer interaction [42, 51, 151], and augmented reality [90, 102]. In computer vision, efforts have been made to leverage gaze in various tasks [37-39, 56, 65-67, 83, 101, 139]. However, with the previous absence of large-scale egocentric datasets that include gaze, this avenue of research is currently under-explored [44, 68, 94, 154, 160]. Our EgoExoLearn offers calibrated gaze positions for all egocentric videos. Thanks to our unique setting, our dataset enables the integration of gaze in egocentric video understanding and the exploration of the role of gaze in the cross-view context." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 271, + 289, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 271, + 289, + 462 + ], + "spans": [ + { + "bbox": [ + 46, + 271, + 289, + 462 + ], + "type": "text", + "content": "Egocentric and ego-exo video understanding. The unique recording perspective of egocentric videos presents a series of challenges including but not limited to action understanding [13, 26, 27, 40, 50, 93, 96, 113, 120, 121, 125], hand detection [12, 29, 107, 153], and video-language understanding [41, 48]. These form fundamental building block techniques of embodied AI [85], VR/AR [9, 47, 74, 123], and human-robot interaction [63, 80, 86, 137]. Since most egocentric datasets are smaller in scale compared with general datasets which contain mostly exocentric view videos [3, 142, 152], it is possible to leverage exocentric video data to improve model performance on egocentric videos [132]. There are typically three main directions: joint view-invariant learning [124, 140, 142], domain adaptation [136], and knowledge distillation [69]. In this work, we evaluate all these directions in our benchmarks." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 475, + 100, + 487 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 475, + 100, + 487 + ], + "spans": [ + { + "bbox": [ + 47, + 475, + 100, + 487 + ], + "type": "text", + "content": "3. Dataset" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 496, + 142, + 507 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 496, + 142, + 507 + ], + "spans": [ + { + "bbox": [ + 47, + 496, + 142, + 507 + ], + "type": "text", + "content": "3.1. Data Collection" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 518, + 288, + 650 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 518, + 288, + 650 + ], + "spans": [ + { + "bbox": [ + 46, + 518, + 288, + 650 + ], + "type": "text", + "content": "Scenarios and tasks. We consider procedural goal-oriented tasks ranging from daily food-making to specialized laboratory-based experiments. This selection is grounded in their exemplification of two prospective areas where future embodied AI agents would need the ability to bridge ego-exo activities: daily-life assistance and professional support. Specifically, EgoExoLearn incorporates 5 types of daily tasks (e.g., cooking) and 3 types of specialized laboratory tasks (e.g., solid-phase peptide synthesis). We record egocentric videos in 4 different kitchens and 3 different labs. Other details are provided in the supplementary." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 654, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 654, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 654, + 288, + 714 + ], + "type": "text", + "content": "Data collection procedure. Before the start of each collection session, participants are required to complete a questionnaire gathering basic demographic information and their self-evaluated expertise in executing the designated task. This questionnaire also highlights the ethical, privacy, and secu" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 306, + 72, + 429, + 213 + ], + "blocks": [ + { + "bbox": [ + 306, + 72, + 429, + 213 + ], + "lines": [ + { + "bbox": [ + 306, + 72, + 429, + 213 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 429, + 213 + ], + "type": "image", + "image_path": "41bf1cd76182fcc4e543357f9fff5b20ff68656802bdde7eb624f84544efa2ad.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 220, + 545, + 286 + ], + "lines": [ + { + "bbox": [ + 304, + 220, + 545, + 286 + ], + "spans": [ + { + "bbox": [ + 304, + 220, + 545, + 286 + ], + "type": "text", + "content": "Figure 2. The number of videos per task (left) and the average duration of each video per task (right). Task1 to Task5 represent the 5 daily tasks and the remaining are three tasks in specialized laboratories. In each recording session of the egocentric video, one participant may learn from multiple demonstration videos and one demonstration video may be watched by several participants." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 433, + 72, + 545, + 210 + ], + "blocks": [ + { + "bbox": [ + 433, + 72, + 545, + 210 + ], + "lines": [ + { + "bbox": [ + 433, + 72, + 545, + 210 + ], + "spans": [ + { + "bbox": [ + 433, + 72, + 545, + 210 + ], + "type": "image", + "image_path": "6e81b914ecc921aa6bfb1c83be55963cf5f84e4eaf459922900c1d7c7a7f5cfe.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 298, + 547, + 394 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 298, + 547, + 394 + ], + "spans": [ + { + "bbox": [ + 304, + 298, + 547, + 394 + ], + "type": "text", + "content": "rity considerations. Then in each session, participants will be asked to choose one or several exocentric view demonstration videos from a provided list and carefully learn the detailed procedures. Once they feel ready, they will wear Pupil Invisible Glasses [49], complete the gaze calibration, and begin to replicate the task performed in the demonstration videos. While not encouraged, participants are permitted to revisit the demonstration video during the recording." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 395, + 547, + 465 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 395, + 547, + 465 + ], + "spans": [ + { + "bbox": [ + 304, + 395, + 547, + 465 + ], + "type": "text", + "content": "After each recording session, the participants are asked to re-do the gaze calibration to ensure gaze data fidelity. For the 5 daily tasks, the exocentric demonstration videos are manually curated from online video platforms such as YouTube. For the lab experiments, the exocentric demonstration videos are tutorials recorded by senior lab members." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 467, + 547, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 467, + 547, + 574 + ], + "spans": [ + { + "bbox": [ + 304, + 467, + 547, + 574 + ], + "type": "text", + "content": "Figure 2 shows the distribution of the 120 hours of data. Since most demonstration videos are meticulously edited to remove repeated steps, the average length of demonstration videos is lower than the egocentric videos which record the full procedure. As a result, the EgoExoLearn contains 432 egocentric videos totaling 96.5 hours and 315 demonstration videos spanning 23.5 hours. This difference in video length poses a unique challenge when bridging ego- and exo-view activities for future research endeavors." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 583, + 382, + 594 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 583, + 382, + 594 + ], + "spans": [ + { + "bbox": [ + 306, + 583, + 382, + 594 + ], + "type": "text", + "content": "3.2. Annotation" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 602, + 547, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 602, + 547, + 685 + ], + "spans": [ + { + "bbox": [ + 304, + 602, + 547, + 685 + ], + "type": "text", + "content": "To facilitate our dataset in the development of algorithms that can effectively bridge the gap between ego and exo viewpoints, we provide detailed multi-modal human annotations. Our pipeline of annotation contains four stages detailed in the following paragraphs. Each of these stages is subject to a rigorous manual quality check involving no fewer than two individuals for verification and validation." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 689, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 545, + 712 + ], + "type": "text", + "content": "Coarse-level language annotation. In this step, we ask annotators to annotate the coarse actions in the videos. Like" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "22074" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 47, + 69, + 296, + 169 + ], + "blocks": [ + { + "bbox": [ + 47, + 69, + 296, + 169 + ], + "lines": [ + { + "bbox": [ + 47, + 69, + 296, + 169 + ], + "spans": [ + { + "bbox": [ + 47, + 69, + 296, + 169 + ], + "type": "image", + "image_path": "6a91f604f0dc85ff41960cbd533e77603029baf2ce792b1ff2d4376c7596dd7b.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 299, + 77, + 546, + 171 + ], + "blocks": [ + { + "bbox": [ + 299, + 70, + 529, + 79 + ], + "lines": [ + { + "bbox": [ + 299, + 70, + 529, + 79 + ], + "spans": [ + { + "bbox": [ + 299, + 70, + 529, + 79 + ], + "type": "text", + "content": "Number of nouns (left axis) and noun durations (right axis) of right hand" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 299, + 77, + 546, + 171 + ], + "lines": [ + { + "bbox": [ + 299, + 77, + 546, + 171 + ], + "spans": [ + { + "bbox": [ + 299, + 77, + 546, + 171 + ], + "type": "image", + "image_path": "1d69c8321a354769981259b26af409b5416cbfbe34be289b772a54d9988d2775.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 77, + 173, + 515, + 184 + ], + "lines": [ + { + "bbox": [ + 77, + 173, + 515, + 184 + ], + "spans": [ + { + "bbox": [ + 77, + 173, + 515, + 184 + ], + "type": "text", + "content": "Figure 3. Occurrence and duration distribution of the annotated fine-level verbs and nouns associated with the right hand." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 195, + 290, + 363 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 195, + 290, + 363 + ], + "spans": [ + { + "bbox": [ + 46, + 195, + 290, + 363 + ], + "type": "text", + "content": "the previous works [106, 126], the coarse actions are defined as a middle-level step for accomplishing a task and can be divided into multiple fine actions. For instance, \"Prepare the pork\" in the task of twice-cooked pork, and \"Suction filtration\" in solid-phase peptide synthesis. Three types of annotation are given in this step: 1) the temporal interval, consisting of start and end timestamps; 2) the action label; and 3) a language description of the video within the annotated interval. We specifically request annotators to focus on elucidating \"what is done,\" \"how it is done,\" and \"the purpose of this step\" in their language descriptions. We define a total of 39 categories of coarse-level actions in this stage and acquire 41.2 coarse-level annotations per video with an average length of 21.5 seconds." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 370, + 288, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 370, + 288, + 479 + ], + "spans": [ + { + "bbox": [ + 46, + 370, + 288, + 479 + ], + "type": "text", + "content": "Fine-level language annotation. Based on the coarse-level annotations, in this step we request the annotators to provide annotations for the fine-level actions. The fine-level actions are the atomic actions like \"take knife\" or \"pull syringe plunger\". Unlike the first step, annotators are instructed to furnish language descriptions that specifically emphasize \"which hand is used\", \"what object is used\" and \"why it is used\". For the first two steps, we employ a two-round manual annotation checking to ensure the annotation quality." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 486, + 290, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 486, + 290, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 486, + 290, + 715 + ], + "type": "text", + "content": "Translation & parsing. To ensure linguistic precision, all the annotators give the language description using their native language [18]. For non-English annotations, we employ ChatGPT and Google Translation API to translate them into English. Subsequently, for the fine-level annotations, we employ specific rules and utilize tools such as NLTK [10] and Spacy [116] to extract the verbs and nouns associated with each segment. During this stage, we confine the selection of verbs and nouns to the predefined taxonomy offered by Ego4D [30], while also manually introducing supplementary verbs and nouns that are absent in the taxonomy. Since our manual annotation specifies the engagement of the left/right hand, we can extract multiple verbs and nouns for each segment, meanwhile attributing them to the respective involvement of the left or right hand. After manual checking, we obtain a total of 95 verb and 254 noun categories in the fine-level annotation. In Figure 3 we show the occurrence of the top 30 categories of verbs and nouns attributed to the right hand. More statistics can be found in the supplementary." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 210, + 548, + 331 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 210, + 548, + 331 + ], + "spans": [ + { + "bbox": [ + 304, + 210, + 548, + 331 + ], + "type": "text", + "content": "Skill level annotation. Since self-assessed skill level is not perfectly suitable for skill assessment, we identify several representative skills and assign human annotators to assess their skills. The annotation follows a pairwise ranking scheme, where annotators are presented with pairs of videos of the same action, and instructed to determine which video demonstrates a higher skill level. We prepare 40,191 video pairs and ensure that 4 different annotators annotate each pair. After filtering out pairs with less than 3 consistent opinions, we get a collection of 34,239 valid video pairs." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 338, + 448, + 351 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 338, + 448, + 351 + ], + "spans": [ + { + "bbox": [ + 306, + 338, + 448, + 351 + ], + "type": "text", + "content": "3.3. Statistics & Comparisons" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 357, + 547, + 466 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 357, + 547, + 466 + ], + "spans": [ + { + "bbox": [ + 304, + 357, + 547, + 466 + ], + "type": "text", + "content": "To the best of our knowledge, there is no dataset that follows the same setting as ours for a direct comparison. Therefore, we enumerate various aspects of our dataset and conduct a comparative analysis with relevant datasets in Tables 1 and 2. EgoExoLearn distinctively enriches the domain with its \"visual demonstration following\" setting. Beyond this unique setting, it stands as the first egocentric dataset that includes temporal bounded language captions, annotated cross-view associations, and multi-label video segments." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 477, + 497, + 491 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 477, + 497, + 491 + ], + "spans": [ + { + "bbox": [ + 305, + 477, + 497, + 491 + ], + "type": "text", + "content": "4. Dataset Properties & Benchmarks" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 497, + 416, + 510 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 497, + 416, + 510 + ], + "spans": [ + { + "bbox": [ + 306, + 497, + 416, + 510 + ], + "type": "text", + "content": "4.1. Dataset Properties" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 516, + 547, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 516, + 547, + 540 + ], + "spans": [ + { + "bbox": [ + 304, + 516, + 547, + 540 + ], + "type": "text", + "content": "EgoExoLearn stands out from current egocentric and ego-exo datasets due to several unique properties." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 543, + 548, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 543, + 548, + 651 + ], + "spans": [ + { + "bbox": [ + 304, + 543, + 548, + 651 + ], + "type": "text", + "content": "Ego-Exo demonstration following setting. The most distinguished property of EgoExoLearn is the ego-exo demonstration following context. Egocentric video recorders are instructed to follow the steps in exocentric demonstration videos to perform the same task but in a different environment. This setting closely emulates the human observational learning process [4, 34] and can be instrumental in designing embodied AI agents that learn from alternative perspectives while executing tasks from their own viewpoint." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 654, + 548, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 548, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 548, + 715 + ], + "type": "text", + "content": "Fine-grained vision-language annotations with gaze. To facilitate a deeper analysis, we equip our dataset with rich, multimodal, fine-grained annotations. EgoExoLearn is the first egocentric dataset featuring high-quality captions, evidenced by the number of words per segment in Tab. 2." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "22075" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 50, + 70, + 545, + 198 + ], + "blocks": [ + { + "bbox": [ + 50, + 70, + 545, + 198 + ], + "lines": [ + { + "bbox": [ + 50, + 70, + 545, + 198 + ], + "spans": [ + { + "bbox": [ + 50, + 70, + 545, + 198 + ], + "type": "table", + "html": "
DatasetSettingsUnique HoursEgo +Exo?Instruction following?Visual instruction?GazeCoarse ActionFine ActionDense CaptionAssociationSkill
Meccano [98]Industry7X✓XXX✓XXX
EGTEA [67]Cooking28X✓X✓X✓XXX
EK-100 [18]Cooking100XXXXX✓XXX
HoloAssist [126]Assistive166X✓X✓✓✓XX✓
Ego4D [30]Multiple3670XXX◇◇◇◇XX
H2O [55]Desk1✓XXXX✓XXX
LEMMA [45]Daily10✓XXXX✓XXX
HOMEAGE [99]Daily25✓XXX✓✓XXX
CharadesEgo [109]Daily34✓XXXX✓XXX
Assembly101 [55]Desk42✓✓XX✓✓XXX
EgoExoLearn (ours)Daily & Lab120✓✓✓✓✓✓✓✓✓
", + "image_path": "bfdcfc0a8d661c4c31503579290085220379affceab77774f9ee9a5fe924216a.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 50, + 239, + 286, + 327 + ], + "blocks": [ + { + "bbox": [ + 46, + 205, + 546, + 228 + ], + "lines": [ + { + "bbox": [ + 46, + 205, + 546, + 228 + ], + "spans": [ + { + "bbox": [ + 46, + 205, + 546, + 228 + ], + "type": "text", + "content": "Table 1. Comparison to related datasets on settings (left) and annotations (right). \"Unique hours\" refers to the cumulative duration of distinct video recordings, counting only one camera's footage for simultaneous recordings of the same activity. " + }, + { + "bbox": [ + 46, + 205, + 546, + 228 + ], + "type": "inline_equation", + "content": "\\diamond" + }, + { + "bbox": [ + 46, + 205, + 546, + 228 + ], + "type": "text", + "content": ": Partially included." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 50, + 239, + 286, + 327 + ], + "lines": [ + { + "bbox": [ + 50, + 239, + 286, + 327 + ], + "spans": [ + { + "bbox": [ + 50, + 239, + 286, + 327 + ], + "type": "table", + "html": "
Dataset# videosAvg. min# segsAvg. secverb classesnoun classes# verb /seg#noun /seg#words /seg
CharadesEgo [109]78600.569k1.83237110
HOMEAGE [99]57000.926k-2986--0
EK-100 [18]7008.590k3.19730011.23.0
Assembly101 [106]43217.183k1.72490110
Ego4D [30]99126.477k8.0115478117.4
Ours-ego43213.464k4.6952541.82.516.9
Ours-exo3154.514k4.7822512.33.019.4
", + "image_path": "783d0cbc9cd7bb9df1f1d23d9ecca658354c1d6c13f38b66722d622af3bec05b.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 335, + 288, + 369 + ], + "lines": [ + { + "bbox": [ + 46, + 335, + 288, + 369 + ], + "spans": [ + { + "bbox": [ + 46, + 335, + 288, + 369 + ], + "type": "text", + "content": "Table 2. Contemporary egocentric datasets. We show only the fine-level actions for a fair comparison. For Ego4D [30], we select the closest subtask of \"forecasting\" following [106]." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 385, + 288, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 385, + 288, + 517 + ], + "spans": [ + { + "bbox": [ + 46, + 385, + 288, + 517 + ], + "type": "text", + "content": "Different from Ego4D where captions are associated with only single timestamps, our captions come with manually annotated start and end timestamps. For better visual perception across ego-exo views, we give verb and noun labels associated with the specific hand. One aspect from which we can analyze the human's ability to bridge ego-exo activities is through gaze. EgoExoLearn is also augmented with calibrated eye-gaze signals. These annotations enable the understanding of human ability to bridge ego-exo activities from diverse perspectives, which we posit will benefit the next-generation embodied AI agents [21, 62]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 527, + 129, + 538 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 527, + 129, + 538 + ], + "spans": [ + { + "bbox": [ + 47, + 527, + 129, + 538 + ], + "type": "text", + "content": "4.2. Benchmarks" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 545, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 545, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 545, + 288, + 713 + ], + "type": "text", + "content": "To evaluate the ability of bridging asynchronous ego-exo procedural activities, we introduce 4 new benchmarks: 1) cross-view association, 2) cross-view action understanding, 3) cross-view referenced skill assessment, and 4) cross-view referenced captioning. The cross-view action understanding benchmark is further subdivided into three subtasks: cross-view action anticipation, cross-view action planning, and cross-view action segmentation. Additionally, we explore the role of gaze in assisting these tasks. We also benchmark models on zero-shot and supervised fine-grained action recognition tasks for reference, following [18, 126]. Note that we carefully split our dataset to eliminate annotation leak across benchmarks. Due to the space limit, we only provide partial content of definition, annotation, results, and analysis," + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 242, + 516, + 255 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 242, + 516, + 255 + ], + "spans": [ + { + "bbox": [ + 306, + 242, + 516, + 255 + ], + "type": "text", + "content": "leaving more complete details in the supplementary." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 273, + 435, + 284 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 273, + 435, + 284 + ], + "spans": [ + { + "bbox": [ + 306, + 273, + 435, + 284 + ], + "type": "text", + "content": "4.2.1 Cross-view association" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 296, + 547, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 296, + 547, + 392 + ], + "spans": [ + { + "bbox": [ + 304, + 296, + 547, + 392 + ], + "type": "text", + "content": "Motivation. One straightforward indicator of the ego-exo activity bridging is the ability to associate the same semantics across ego- and exo-views. This benchmark focuses on equipping models with this cross-view association ability. An application of this ability is assistants in AR that can show expert demonstration videos when the human is confused [157]. Another potential application is the embodied AI agent that can explain its decision [118, 127, 146]." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 396, + 547, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 396, + 547, + 515 + ], + "spans": [ + { + "bbox": [ + 304, + 396, + 547, + 515 + ], + "type": "text", + "content": "Problem settings. We formulate this association benchmark as a cross-view multiple-choice association problem. Specifically, we consider two different cross-view association settings: ego2exo and exo2ego. In the case of ego2exo, given an egocentric video, the model is asked to predict the corresponding exocentric video performing the same action from a candidate choice set of exocentric samples, and vice versa for the exo2ego setting. For both ego2exo and exo2ego settings, we use 20 candidate samples for each query. The evaluation metric is the averaged Top-1 accuracy." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 518, + 547, + 626 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 518, + 547, + 626 + ], + "spans": [ + { + "bbox": [ + 304, + 518, + 547, + 626 + ], + "type": "text", + "content": "Annotations. We meticulously construct the ground-truth ego-exo pairs via a semantic-aware matching process. It is composed of five stages with details in the supplementary material: (1) Scenario Matching. (2) Noun and Verb Matching. (3) Sentence Matching with LLM. (4) Negative Sampling. (5) Two-round Manual Verification. Notably, we do not provide such pairs for the training set and leave the modeling of cross-view association on unpaired samples to be further explored for the community." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 629, + 548, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 629, + 548, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 629, + 548, + 713 + ], + "type": "text", + "content": "Baseline model. We adopt three types of baseline models, i.e., ego-only, exo-only, and ego-exo, which refer to the data we used during training. Under all the settings, we leverage the paired video and caption and jointly train a video encoder and a text encoder using the contrastive loss [97]. We use TimeSformer-B [8] as the video encoder and clip-text [97] as the text encoder. We initialize both encoders" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "22076" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 51, + 72, + 221, + 171 + ], + "blocks": [ + { + "bbox": [ + 51, + 72, + 221, + 171 + ], + "lines": [ + { + "bbox": [ + 51, + 72, + 221, + 171 + ], + "spans": [ + { + "bbox": [ + 51, + 72, + 221, + 171 + ], + "type": "image", + "image_path": "91014facf6f8bcfa9554cb9dad5ad362d552cd3af1a70c237964fa8454581a74.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 176, + 546, + 199 + ], + "lines": [ + { + "bbox": [ + 46, + 176, + 546, + 199 + ], + "spans": [ + { + "bbox": [ + 46, + 176, + 546, + 199 + ], + "type": "text", + "content": "Figure 4. Concept of the 3 benchmarks of cross-view association (Sec. 4.2.1), cross-view action anticipation & planning (Sec. 4.2.2) and cross-view reference skill assessment (Sec.4.2.3) in this section. Other benchmarks can be found in the supplementary material." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 226, + 72, + 385, + 171 + ], + "blocks": [ + { + "bbox": [ + 226, + 72, + 385, + 171 + ], + "lines": [ + { + "bbox": [ + 226, + 72, + 385, + 171 + ], + "spans": [ + { + "bbox": [ + 226, + 72, + 385, + 171 + ], + "type": "image", + "image_path": "5929e0bdfdd454c39af711fa6adb1ae67b6f37f7d0c96e68327bb792422ed6a8.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 391, + 72, + 544, + 171 + ], + "blocks": [ + { + "bbox": [ + 391, + 72, + 544, + 171 + ], + "lines": [ + { + "bbox": [ + 391, + 72, + 544, + 171 + ], + "spans": [ + { + "bbox": [ + 391, + 72, + 544, + 171 + ], + "type": "image", + "image_path": "e3d6d14183859995b34ad8502e15e3d09f2b00be8121242e9fa5eab2f01469eb.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 210, + 287, + 317 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 210, + 287, + 317 + ], + "spans": [ + { + "bbox": [ + 46, + 210, + 287, + 317 + ], + "type": "text", + "content": "using the EgoVLP [72] pre-trained weights. During testing, we obtain the ego/exocentric video representations using the video encoder. The prediction is defined as the one with the highest normalized cross-view video feature similarity among all candidates. On top of these models, we evaluate the effectiveness of gaze in associating egocentric video and exocentric video. This is achieved by replacing the original egocentric video with spatially cropped videos using the gaze positions as the cropping center." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 323, + 288, + 467 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 323, + 288, + 467 + ], + "spans": [ + { + "bbox": [ + 46, + 323, + 288, + 467 + ], + "type": "text", + "content": "Related work. Prior works on pre-trained vision-language models for multiple choice association/questioning [60, 72, 122, 130, 138, 158] are generally pre-trained on either exocentric [3, 82] or egocentric [30] datasets, only revealing weak cross-view bridging ability. Another line of works explores cross-view learning [2, 33, 135, 141] by either transferring the knowledge from one view to the other [69] or training view-invariant video understanding models [124, 143]. Different from previous works, our cross-view association benchmark is in a more realistic but challenging setting, by evaluating the model's ability to associate asynchronous activities across ego- and exo-views." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 472, + 287, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 472, + 287, + 567 + ], + "spans": [ + { + "bbox": [ + 46, + 472, + 287, + 567 + ], + "type": "text", + "content": "Experiment results. We first evaluate several vision-language models via zero-shot transfer. These models are pre-trained either on egocentric videos, i.e. EgoVLP [72] and LaViLa [158], or exocentric videos, i.e., InternVideo [129]. As shown in Tab. 3, without using gaze, EgoVLP generally outperforms the others on both validation and test sets. By introducing gaze information, InternVideo receives a decent improvement, especially on Exo2Ego." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "type": "text", + "content": "As for fine-tuned models, Tab. 3, reveals that models trained solely on single-view data struggle with cross-view association. Training ego-only models using gaze-cropped egocentric videos results in substantial improvements, significantly outperforming those trained on center-cropped videos. This again highlights the importance of gaze in enhancing cross-view association. Based on our observation, the regions around gaze help the cluttered ego videos become visually similar to exo videos where the primary object is salient. Last, we show the baseline result when co-trained on both egocentric and exocentric videos. The model in this setting shows the strongest cross-view association ability over" + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 309, + 207, + 544, + 392 + ], + "blocks": [ + { + "bbox": [ + 309, + 207, + 544, + 392 + ], + "lines": [ + { + "bbox": [ + 309, + 207, + 544, + 392 + ], + "spans": [ + { + "bbox": [ + 309, + 207, + 544, + 392 + ], + "type": "table", + "html": "
MethodGazeValTest
Ego2ExoExo2EgoEgo2ExoExo2Ego
Zero-shot
Random✘12.715.014.113.4
EgoVLP [72]✘28.827.232.128.9
LaViLa [158]✘22.624.928.725.7
InternVideo [129]✘27.021.230.621.7
EgoVLP [72]✓28.829.731.528.9
LaViLa [158]✓21.921.430.325.9
InternVideo [129]✓30.932.333.332.2
Fine-tuned
Exo-only✘42.941.745.446.9
Ego-only✘33.637.140.335.8
Ego-only✓34.638.745.641.8
Ego-onlyCenter25.422.824.724.2
EgoExo✘42.945.449.045.3
EgoExo✓47.948.855.351.1
", + "image_path": "3db0d1c8bb14337be8ff259c1b7bf9418393f6123267379f5f7711b0b2bf2a65.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 400, + 546, + 456 + ], + "lines": [ + { + "bbox": [ + 305, + 400, + 546, + 456 + ], + "spans": [ + { + "bbox": [ + 305, + 400, + 546, + 456 + ], + "type": "text", + "content": "Table 3. Association accuracy in the cross-view association benchmark. In the fine-tuned setting, we adopt three kinds of data sources for training, i.e., ego-only, exo-only, and hybrid ego-exo data. By leveraging gaze information during training, the model outperforms the baseline (w/o gaze) and the center-crop counterpart." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 305, + 468, + 547, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 468, + 547, + 515 + ], + "spans": [ + { + "bbox": [ + 305, + 468, + 547, + 515 + ], + "type": "text", + "content": "single-view models. Findings from this benchmark underline the limitation of current models in associating activities across ego and exo views, and point towards the potential benefits of integrating gaze into the association." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 532, + 518, + 544 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 532, + 518, + 544 + ], + "spans": [ + { + "bbox": [ + 306, + 532, + 518, + 544 + ], + "type": "text", + "content": "4.2.2 Cross-view action anticipation & planning" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 554, + 547, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 554, + 547, + 696 + ], + "spans": [ + { + "bbox": [ + 304, + 554, + 547, + 696 + ], + "type": "text", + "content": "Motivation. The procedural actions are not guaranteed to be identical between the two views due to practical constraints. Thus, we design benchmarks for cross-view action anticipation and planning to enable a thorough understanding and transfer of necessary steps (or actions) for task completion, bridging the gap between views and considering real-world conditions. A practical application of this benchmark can be seen in human-robot collaboration scenarios. For instance, an embodied AI agent, after observing a human perform the first part of a task, could effectively take over and complete the remaining half of the task based on the particular environmental situation." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 701, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 701, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 701, + 546, + 713 + ], + "type": "text", + "content": "Problem settings. For both the cross-view anticipation and" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "22077" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 47, + 70, + 288, + 274 + ], + "blocks": [ + { + "bbox": [ + 47, + 70, + 288, + 274 + ], + "lines": [ + { + "bbox": [ + 47, + 70, + 288, + 274 + ], + "spans": [ + { + "bbox": [ + 47, + 70, + 288, + 274 + ], + "type": "table", + "html": "
MethodGazeAnticipation↑Planning↓
Ego-VEgo-NExo-VExo-NEgoExo
Exo-onlyX29.923.640.940.584.776.1
Ego-onlyX33.437.828.917.683.484.5
Ego-only✓40.552.837.637.680.082.6
Ego-onlyCenter33.238.634.132.782.684.7
Unsupervised Domain Adaption
Ego2ExoX33.638.135.428.783.084.1
Exo2EgoX30.423.639.239.883.979.0
Ego2Exo✓40.854.238.737.182.884.3
Exo2Ego✓33.531.339.140.182.478.8
Knowledge Distillation
Ego2ExoX29.624.941.645.284.375.5
Exo2EgoX34.038.428.618.683.184.3
Ego2Exo✓29.925.041.245.184.875.1
Exo2Ego✓41.056.137.739.179.582.6
Co-training
Ego & ExoX33.537.439.644.383.276.0
Ego & Exo✓43.853.340.344.479.075.6
", + "image_path": "79d9f1fe7c9237ded359ee3193ca1a430d4bad06545cc88009b740d7d7cb89c4.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 282, + 288, + 327 + ], + "lines": [ + { + "bbox": [ + 46, + 282, + 288, + 327 + ], + "spans": [ + { + "bbox": [ + 46, + 282, + 288, + 327 + ], + "type": "text", + "content": "Table 4. Results of cross-view action anticipation and planning benchmarks. For anticipation, the class-mean Top-5 recall is used as the evaluation metric (higher is better). For planning, the Edit distance is used as the evaluation metric (lower is better)." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 337, + 288, + 469 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 337, + 288, + 469 + ], + "spans": [ + { + "bbox": [ + 46, + 337, + 288, + 469 + ], + "type": "text", + "content": "cross-view planning, the goal is to anticipate the future activities in one view, given labeled training data only in another view. For the cross-view anticipation task, we focus on predicting the verb and noun categories of the next fine-level action " + }, + { + "bbox": [ + 46, + 337, + 288, + 469 + ], + "type": "inline_equation", + "content": "\\tau = 1" + }, + { + "bbox": [ + 46, + 337, + 288, + 469 + ], + "type": "text", + "content": " second into the future. Given the multilabel nature of our verb and noun annotations in each fine-level segment, we perform multiclass anticipation. Performance is evaluated by class-mean Top-5 recall as per [18]. For the cross-view planning task, we aim to generate the next " + }, + { + "bbox": [ + 46, + 337, + 288, + 469 + ], + "type": "inline_equation", + "content": "K = 8" + }, + { + "bbox": [ + 46, + 337, + 288, + 469 + ], + "type": "text", + "content": " steps of coarse-level actions. We adopt ED@ " + }, + { + "bbox": [ + 46, + 337, + 288, + 469 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 46, + 337, + 288, + 469 + ], + "type": "text", + "content": " as the evaluation metric following the setting of Ego4D LTA [30]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 472, + 288, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 472, + 288, + 592 + ], + "spans": [ + { + "bbox": [ + 46, + 472, + 288, + 592 + ], + "type": "text", + "content": "Annotations. For action anticipation, we use the fine-level verb and noun annotations, and then take their intersection between ego and exo videos to constrain them in the same closed set. To evaluate the model more effectively, we further control the long-tail degree of the data by and filter out the tail categories that occur less than " + }, + { + "bbox": [ + 46, + 472, + 288, + 592 + ], + "type": "inline_equation", + "content": "\\frac{1}{100}" + }, + { + "bbox": [ + 46, + 472, + 288, + 592 + ], + "type": "text", + "content": " of the highest occurrence category. For the action planning task, we directly adopt the coarse-level action annotations and regard the start timestamp of each segment as one action step. More details can be found in the supplementary material." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 594, + 290, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 290, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 290, + 714 + ], + "type": "text", + "content": "Baseline model. We explore three distinct directions to implement cross-view baseline models. The first direction is based on unsupervised domain adaptation (UDA), treating one view as the source domain and the other as the target domain. This method operates within an unsupervised training framework, using labels from the source domain and video data from both domains [16, 52, 84, 104, 111, 131, 145]. We adopt CLIP [97] + TA3N [14] as the baseline model. The second direction entails knowledge distillation (KD) [28, 32, 87], allowing the model trained on one view to learn knowl" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 72, + 547, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 168 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 168 + ], + "type": "text", + "content": "edge of the other view, under the assumption that a teacher model of the other view is available. We equip CLIP with a distillation approach based on [69] to transfer knowledge from the teacher model to a newly created student model. The third and most straightforward direction is co-training (CT) using the data from both views to encourage the model to discover correlations between them directly. For using gaze, we also crop the video based on the gaze positions." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 168, + 548, + 324 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 168, + 548, + 324 + ], + "spans": [ + { + "bbox": [ + 304, + 168, + 548, + 324 + ], + "type": "text", + "content": "We comprehensively consider four evaluation settings. The \"Ego-only\" and \"Exo-only\" settings do not involve cross-view understanding, thus we use zero-shot evaluation serving as the references. The Ego2Exo and Exo2Ego settings are the cross-view settings. For UDA, \"Ego2Exo\" is defined as utilizing the egocentric view as the source domain and the exocentric view as the target domain. In the context of KD, \"Ego2Exo\" indicates we initially train a teacher model on egocentric data, followed by the training and distillation of the student model on exocentric data. For CT, we merge both egocentric and exocentric datasets through direct concatenation. We report results on the test set and put the validation set results in the supplementary." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 326, + 547, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 326, + 547, + 411 + ], + "spans": [ + { + "bbox": [ + 304, + 326, + 547, + 411 + ], + "type": "text", + "content": "Related work. Prior works [25, 54, 106, 117] on multi-view action understanding mainly focus on synchronized multiview videos. Some work studied transferring knowledge [69] from one view to the other or training the view-invariant [124, 143, 159] video models. Our cross-view benchmarks seek to evaluate the ability of models to bridge asynchronous actions across views, which is more challenging yet realistic." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 413, + 548, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 413, + 548, + 653 + ], + "spans": [ + { + "bbox": [ + 304, + 413, + 548, + 653 + ], + "type": "text", + "content": "Experiment results. Table 4 presents the results of action anticipation and planning on the test set. The first block of results shows a significant performance gap when models trained exclusively on one view are tested on the other view. This underscores the inherent differences in activities captured in the two views. Remarkably, even without relying on any specific cross-view method, the inclusion of gaze information markedly diminishes the disparity between egocentric and exocentric data. Leveraging techniques such as UDA or KD we can see improved performance compared with direct zero-shot inference in the first block. Since CT can utilize both egocentric and exocentric data, it achieves the best performance in all setups. Moreover, from the comparison in all settings using or not using gaze, it is clear that gaze serves as a surprisingly effective signal to mitigate the gap between activities in the two views, although naively designed. These results take the first step in the potential directions for better bridging the cross-view activities, setting a foundation for future advancements in the field of cross-view action understanding." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 667, + 503, + 679 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 667, + 503, + 679 + ], + "spans": [ + { + "bbox": [ + 305, + 667, + 503, + 679 + ], + "type": "text", + "content": "4.2.3 Cross-view referenced skill assessment" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 689, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 689, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 689, + 547, + 715 + ], + "type": "text", + "content": "Motivation and problem setting. We propose a novel task of cross-view referenced skill assessment leveraging the" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "22078" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 264 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 264 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 264 + ], + "type": "text", + "content": "unique setting and annotations of our dataset. This task goes beyond the traditional pairwise ranking often used in skill assessment [22, 58] by incorporating an expert demonstration video as a reference point. This demonstration provides a model of the ideal execution of an action, offering a standard against which to compare skill levels. In this task, the input is a pair of egocentric video clips " + }, + { + "bbox": [ + 46, + 72, + 289, + 264 + ], + "type": "inline_equation", + "content": "C_{ego1}" + }, + { + "bbox": [ + 46, + 72, + 289, + 264 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 72, + 289, + 264 + ], + "type": "inline_equation", + "content": "C_{ego2}" + }, + { + "bbox": [ + 46, + 72, + 289, + 264 + ], + "type": "text", + "content": " of the same action and an exo-view demonstration video clip " + }, + { + "bbox": [ + 46, + 72, + 289, + 264 + ], + "type": "inline_equation", + "content": "C_{exo}" + }, + { + "bbox": [ + 46, + 72, + 289, + 264 + ], + "type": "text", + "content": " as the reference, and the output is a choice " + }, + { + "bbox": [ + 46, + 72, + 289, + 264 + ], + "type": "inline_equation", + "content": "c \\in \\{ego1, ego2\\}" + }, + { + "bbox": [ + 46, + 72, + 289, + 264 + ], + "type": "text", + "content": " of the egocentric clip that demonstrates a higher skill level. Moreover, we can also assess the skill level by the relation between action and gaze. This benchmark evaluates a model's ability to bridge asynchronous ego-exo dynamics. A practical application is an AR system that helps humans in skill acquisition by providing targeted feedback based on skill level assessment and expert demonstration." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 268, + 289, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 268, + 289, + 389 + ], + "spans": [ + { + "bbox": [ + 46, + 268, + 289, + 389 + ], + "type": "text", + "content": "Annotations. For the cross-view referenced skill assessment benchmark, we concentrate on four types of representative actions, as detailed in Tab. 5. We use the skill level annotations in Sec. 3.2 in this benchmark. For each pair of videos, the annotation is provided by four distinct annotators to minimize subjective bias. We ensure credibility by checking the transitivity of annotations and removing the pairs with less than 3 agreements among 4 annotators. We then append an exo-view demonstration video clip of the same action, forming video triplets as the model input." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 392, + 289, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 392, + 289, + 548 + ], + "spans": [ + { + "bbox": [ + 46, + 392, + 289, + 548 + ], + "type": "text", + "content": "Baseline model. Our baseline model is built upon a pairwise ranking skill assessment model RAAN [23]. Without loss of generality, assuming egol is the video showing a higher skill level, we apply the following approaches to leverage the reference exo-view demonstration video: 1) Triplet loss (TL). The feature distance between " + }, + { + "bbox": [ + 46, + 392, + 289, + 548 + ], + "type": "inline_equation", + "content": "C_{exo}" + }, + { + "bbox": [ + 46, + 392, + 289, + 548 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 392, + 289, + 548 + ], + "type": "inline_equation", + "content": "C_{ego1}" + }, + { + "bbox": [ + 46, + 392, + 289, + 548 + ], + "type": "text", + "content": " should be closer to the distance between " + }, + { + "bbox": [ + 46, + 392, + 289, + 548 + ], + "type": "inline_equation", + "content": "C_{exo}" + }, + { + "bbox": [ + 46, + 392, + 289, + 548 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 392, + 289, + 548 + ], + "type": "inline_equation", + "content": "C_{ego2}" + }, + { + "bbox": [ + 46, + 392, + 289, + 548 + ], + "type": "text", + "content": ". 2) Relation network (RN). Inspired by [114], we employ a relation network that concatenates the features of the ego and exo clips. This network is designed to discern which of the two egocentric video clips bears a closer relation to the demonstration video in terms of skill level. The way to gaze is consistent with the other benchmarks." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 553, + 289, + 673 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 553, + 289, + 673 + ], + "spans": [ + { + "bbox": [ + 46, + 553, + 289, + 673 + ], + "type": "text", + "content": "Related work. Several previous works on skill assessment aim to directly regress a score based on professional ratings [1, 71, 73, 91, 115, 149]. We adopt a more general approach of pairwise ranking since no absolute score is available in most real-world skills [7, 70]. Previous works in this direction only use a pair of videos that are in either egocentric view [22] or in exo view [23, 70]. Differently from their works, we explore the cross-view activity bridging ability by examining how demonstration videos from exo view can benefit skill assessment in ego views." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 677, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 289, + 715 + ], + "type": "text", + "content": "Experiment results. We first evaluate the performance of previous works under the conventional pairwise setting. In the upper block of Table 5, both methods [22, 23] receive" + } + ] + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 309, + 70, + 545, + 191 + ], + "blocks": [ + { + "bbox": [ + 309, + 70, + 545, + 191 + ], + "lines": [ + { + "bbox": [ + 309, + 70, + 545, + 191 + ], + "spans": [ + { + "bbox": [ + 309, + 70, + 545, + 191 + ], + "type": "table", + "html": "
MethodGazeEgg CrackingPeelingStir-fryCutting
Ego pairs only
Who's better [22]✘74.7975.9877.5476.85
RAAN [23]✘78.4578.5282.5379.09
Who's better [22]✓77.9176.7079.1377.02
RAAN [23]✓82.0879.3783.9279.36
Ego pairs + Exo
RAAN [23] + RN✘77.9277.0981.5478.26
RAAN [23] + TL✘78.8179.4682.5078.73
RAAN [23] + RN✓82.1479.3283.5179.44
RAAN [23] + TL✓82.1679.5984.0579.29
", + "image_path": "5279d34b49c02b1db117aed65e7923bffd208d3b1d335273556b809779f5bf5e.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 198, + 547, + 243 + ], + "lines": [ + { + "bbox": [ + 304, + 198, + 547, + 243 + ], + "spans": [ + { + "bbox": [ + 304, + 198, + 547, + 243 + ], + "type": "text", + "content": "Table 5. Ranking accuracy of cross-view referenced skill assessment. In the upper part of the table, only ego video pairs are used, while in the lower part, exo demonstrations are incorporated by \"RN\": relation network and \"TL\": triplet loss." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 255, + 548, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 255, + 548, + 399 + ], + "spans": [ + { + "bbox": [ + 304, + 255, + 548, + 399 + ], + "type": "text", + "content": "a clear performance boost when gaze is used. This aligns with behavioral science findings that experts and novices have different gaze patterns in the same task [76, 133]. With the exocentric demonstration as reference, both the relation network method (RN) and the triplet loss method (TL) can leverage the reference video and bring performance improvement. Further adding gaze we can get the best performance. However, the marginal improvement observed with the inclusion of the exocentric reference suggests that current models may still struggle to fully bridge asynchronous activities across egocentric and exocentric views. There remains ample room for new improvement and innovation." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 411, + 379, + 424 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 411, + 379, + 424 + ], + "spans": [ + { + "bbox": [ + 306, + 411, + 379, + 424 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 432, + 547, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 432, + 547, + 624 + ], + "spans": [ + { + "bbox": [ + 304, + 432, + 547, + 624 + ], + "type": "text", + "content": "The ability to bridge asynchronous procedural activities in ego- and exo-views is imperative for next-generation embodied AI in executing sophisticated tasks in the real world. As a fundamental step, our EgoExoLearn encompasses a rich collection of egocentric videos, each captured when replicating procedures of exocentric demonstration videos, but performed in different environments and at different times. This realistic setup, combined with our multimodal annotations, allows us to construct 4 novel benchmarks, serving as a versatile platform for investigating how cross-view asynchronous activities can be bridged. EgoExoLearn also enables new research directions e.g., how to better leverage gaze and hand-associated annotations. Results from the benchmarks show weaknesses of current models in bridging ego- and exo-view asynchronous activities, leaving significant room for future work to improve upon." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 624, + 545, + 669 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 624, + 545, + 669 + ], + "spans": [ + { + "bbox": [ + 304, + 624, + 545, + 669 + ], + "type": "text", + "content": "Acknowledgement. This work is supported by the National Key R&D Program of China (No.2022ZD0160102) and the Industry Collaboration Projects Grant, Shanghai Committee of Science and Technology, China (No.22YF1461500)." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 669, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 669, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 669, + 547, + 713 + ], + "type": "text", + "content": "Key contribution statement: Guo made key contributions to annotation processing and 3 action benchmarks. Jilan made key contributions to the association and caption benchmarks. Mingfang contributed primarily to the skill benchmark." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "22079" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 53, + 91, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 58, + 91, + 287, + 156 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 91, + 287, + 156 + ], + "spans": [ + { + "bbox": [ + 58, + 91, + 287, + 156 + ], + "type": "text", + "content": "[1] Narges Ahmidi, Lingling Tao, Shahin Sefati, Yixin Gao, Colin Lea, Benjamin Bejar Haro, Luca Zappella, Sanjeev Khudanpur, René Vidal, and Gregory D Hager. A dataset and benchmarks for segmentation and recognition of gestures in robotic surgery. IEEE Transactions on Biomedical Engineering, 64(9):2025-2041, 2017. 8" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 158, + 288, + 190 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 158, + 288, + 190 + ], + "spans": [ + { + "bbox": [ + 58, + 158, + 288, + 190 + ], + "type": "text", + "content": "[2] Shervin Ardeshir and Ali Borji. An exocentric look at egocentric actions and vice versa. Computer Vision and Image Understanding, 171:61-68, 2018. 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 58, + 191, + 288, + 235 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 191, + 288, + 235 + ], + "spans": [ + { + "bbox": [ + 58, + 191, + 288, + 235 + ], + "type": "text", + "content": "[3] Max Bain, Arsha Nagrani, Gül Varol, and Andrew Zisserman. Frozen in time: A joint video and image encoder for end-to-end retrieval. In Proceedings of the International Conference on Computer Vision (ICCV), 2021. 3, 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 58, + 236, + 287, + 257 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 236, + 287, + 257 + ], + "spans": [ + { + "bbox": [ + 58, + 236, + 287, + 257 + ], + "type": "text", + "content": "[4] Albert Bandura. Observational learning. The international encyclopedia of communication, 2008. 1, 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 58, + 258, + 287, + 301 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 258, + 287, + 301 + ], + "spans": [ + { + "bbox": [ + 58, + 258, + 287, + 301 + ], + "type": "text", + "content": "[5] Siddhant Bansal, Chetan Arora, and C.V. Jawahar. My view is the best view: Procedure learning from egocentric videos. In Proceedings of the European Conference on Computer Vision (ECCV), 2022. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 58, + 303, + 288, + 368 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 303, + 288, + 368 + ], + "spans": [ + { + "bbox": [ + 58, + 303, + 288, + 368 + ], + "type": "text", + "content": "[6] Yizhak Ben-Shabat, Xin Yu, Fatemeh Saleh, Dylan Campbell, Cristian Rodriguez-Opazo, Hongdong Li, and Stephen Gould. The ikaa asm dataset: Understanding people assembling furniture through actions, objects and pose. In Proceedings of the Winter Conference on Applications of Computer Vision (WACV), 2021. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 58, + 369, + 287, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 369, + 287, + 413 + ], + "spans": [ + { + "bbox": [ + 58, + 369, + 287, + 413 + ], + "type": "text", + "content": "[7] Gedas Bertasius, Hyun Soo Park, Stella X Yu, and Jianbo Shi. Am i a baller? basketball performance assessment from first-person videos. In Proceedings of the International Conference on Computer Vision (ICCV), 2017. 8" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 58, + 415, + 287, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 415, + 287, + 456 + ], + "spans": [ + { + "bbox": [ + 58, + 415, + 287, + 456 + ], + "type": "text", + "content": "[8] Gedas Bertasius, Heng Wang, and Lorenzo Torresani. Is space-time attention all you need for video understanding? In Proceedings of the International Conference on Machine Learning (ICML), 2021. 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 58, + 458, + 287, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 458, + 287, + 502 + ], + "spans": [ + { + "bbox": [ + 58, + 458, + 287, + 502 + ], + "type": "text", + "content": "[9] Vinay Bettadapura, Irfan Essa, and Caroline Pantofaru. Ego-centric field-of-view localization using first-person point-of-view devices. In Proceedings of the Winter Conference on Applications of Computer Vision (WACV), 2015. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 53, + 503, + 288, + 534 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 503, + 288, + 534 + ], + "spans": [ + { + "bbox": [ + 53, + 503, + 288, + 534 + ], + "type": "text", + "content": "[10] Steven Bird. Nltk: the natural language toolkit. In Proceedings of the COLING/ACL Interactive Presentation Sessions, 2006. 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 53, + 536, + 288, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 536, + 288, + 568 + ], + "spans": [ + { + "bbox": [ + 53, + 536, + 288, + 568 + ], + "type": "text", + "content": "[11] Greg Brockman, Vicki Cheung, Ludwig Pettersson, Jonas Schneider, John Schulman, Jie Tang, and Wojciech Zaremba. Openai gym. arXiv preprint arXiv:1606.01540, 2016. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 53, + 570, + 288, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 570, + 288, + 602 + ], + "spans": [ + { + "bbox": [ + 53, + 570, + 288, + 602 + ], + "type": "text", + "content": "[12] Minjie Cai, Kris M Kitani, and Yoichi Sato. Understanding hand-object manipulation with grasp types and object attributes. In Robotics: Science and Systems, 2016. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 53, + 603, + 288, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 603, + 288, + 657 + ], + "spans": [ + { + "bbox": [ + 53, + 603, + 288, + 657 + ], + "type": "text", + "content": "[13] Guo Chen, Yifei Huang, Jilan Xu, Baoqi Pei, Zhe Chen, Zhiqi Li, Jiahao Wang, Kunchang Li, Tong Lu, and Limin Wang. Video mamba suite: State space model as a versatile alternative for video understanding. arXiv preprint arXiv:2403.09626, 2024. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 53, + 658, + 288, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 658, + 288, + 712 + ], + "spans": [ + { + "bbox": [ + 53, + 658, + 288, + 712 + ], + "type": "text", + "content": "[14] Min-Hung Chen, Zsolt Kira, Ghassan AlRegib, Jaekwon Yoo, Ruxin Chen, and Jian Zheng. Temporal attentive alignment for large-scale video domain adaptation. In Proceedings of the International Conference on Computer Vision (ICCV), 2019. 2, 7" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 312, + 73, + 547, + 714 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 312, + 73, + 547, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 73, + 547, + 127 + ], + "spans": [ + { + "bbox": [ + 312, + 73, + 547, + 127 + ], + "type": "text", + "content": "[15] Kellie Corona, Katie Osterdahl, Roderic Collins, and Anthony Hoogs. Meva: A large-scale multiview, multimodal video dataset for activity detection. In Proceedings of the Winter Conference on Applications of Computer Vision (WACV), 2021. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 312, + 129, + 547, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 129, + 547, + 183 + ], + "spans": [ + { + "bbox": [ + 312, + 129, + 547, + 183 + ], + "type": "text", + "content": "[16] Victor G Turrisi da Costa, Giacomo Zara, Paolo Rota, Thiago Oliveira-Santos, Nicu Sebe, Vittorio Murino, and Elisa Ricci. Dual-head contrastive domain adaptation for video action recognition. In Proceedings of the Winter Conference on Applications of Computer Vision (WACV), 2022. 7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 312, + 184, + 547, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 184, + 547, + 249 + ], + "spans": [ + { + "bbox": [ + 312, + 184, + 547, + 249 + ], + "type": "text", + "content": "[17] Dima Damen, Hazel Doughty, Giovanni Maria Farinella, Sanja Fidler, Antonino Furnari, Evangelos Kazakos, Davide Moltisanti, Jonathan Munro, Toby Perrett, Will Price, et al. Scaling egocentric vision: The epic-kitchens dataset. In Proceedings of the European Conference on Computer Vision (ECCV), 2018. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 312, + 251, + 547, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 251, + 547, + 326 + ], + "spans": [ + { + "bbox": [ + 312, + 251, + 547, + 326 + ], + "type": "text", + "content": "[18] Dima Damen, Hazel Doughty, Giovanni Maria Farinella, Antonino Furnari, Jian Ma, Evangelos Kazakos, Davide Moltisanti, Jonathan Munro, Toby Perrett, Will Price, and Michael Wray. Rescaling egocentric vision: Collection, pipeline and challenges for epic-kitchens-100. International Journal of Computer Vision (IJCV), 130:33-55, 2022. 1, 2, 4, 5, 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 312, + 327, + 546, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 327, + 546, + 392 + ], + "spans": [ + { + "bbox": [ + 312, + 327, + 546, + 392 + ], + "type": "text", + "content": "[19] Ahmad Darkhalil, Dandan Shan, Bin Zhu, Jian Ma, Amlan Kar, Richard Higgins, Sanja Fidler, David Fouhey, and Dima Damen. Epic-kitchens visor benchmark: Video segmentations and object relations. In Proceedings of the Neural Information Processing Systems (NeurIPS) Track on Datasets and Benchmarks, 2022. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 312, + 393, + 546, + 436 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 393, + 546, + 436 + ], + "spans": [ + { + "bbox": [ + 312, + 393, + 546, + 436 + ], + "type": "text", + "content": "[20] Fernando De la Torre, Jessica Hodgins, Adam Bargteil, Xavier Martin, Justin Macey, Alex Collado, and Pep Beltran. Guide to the carnegie mellon university multimodal activity (cmu-mmac) database. 2009. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 312, + 437, + 547, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 437, + 547, + 502 + ], + "spans": [ + { + "bbox": [ + 312, + 437, + 547, + 502 + ], + "type": "text", + "content": "[21] Matt Deitke, Winson Han, Alvaro Herrasti, Aniruddha Kembhavi, Eric Kolve, Roozbeh Mottaghi, Jordi Salvador, Dustin Schwenk, Eli VanderBilt, Matthew Wallingford, et al. Robothor: An open simulation-to-real embodied ai platform. In Proceedings of the Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 5" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 312, + 504, + 547, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 504, + 547, + 547 + ], + "spans": [ + { + "bbox": [ + 312, + 504, + 547, + 547 + ], + "type": "text", + "content": "[22] Hazel Doughty, Dima Damen, and Walterio Mayol-Cuevas. Who's better? who's best? pairwise deep ranking for skill determination. In Proceedings of the Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 8" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 312, + 548, + 547, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 548, + 547, + 601 + ], + "spans": [ + { + "bbox": [ + 312, + 548, + 547, + 601 + ], + "type": "text", + "content": "[23] Hazel Doughty, Walterio Mayol-Cuevas, and Dima Damen. The pros and cons: Rank-aware temporal attention for skill determination in long videos. In Proceedings of the Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 8" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 312, + 603, + 546, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 603, + 546, + 635 + ], + "spans": [ + { + "bbox": [ + 312, + 603, + 546, + 635 + ], + "type": "text", + "content": "[24] Mitch J Fryling, Cristin Johnston, and Linda J Hayes. Understanding observational learning: An interbehavioral approach. The Analysis of verbal behavior, 27:191, 2011. 1" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 312, + 636, + 546, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 636, + 546, + 679 + ], + "spans": [ + { + "bbox": [ + 312, + 636, + 546, + 679 + ], + "type": "text", + "content": "[25] Yanwei Fu, Yanwen Guo, Yanshu Zhu, Feng Liu, Chuanming Song, and Zhi-Hua Zhou. Multi-view video summarization. IEEE Transactions on Multimedia, 12(7):717-729, 2010. 7" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 312, + 681, + 547, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 681, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 312, + 681, + 547, + 714 + ], + "type": "text", + "content": "[26] Antonino Furnari and Giovanni Maria Farinella. What would you expect? anticipating egocentric actions with rolling-unrolling lstms and modality attention. In Proceedings of" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "22080" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 53, + 72, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 72, + 72, + 287, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 72, + 72, + 287, + 94 + ], + "spans": [ + { + "bbox": [ + 72, + 72, + 287, + 94 + ], + "type": "text", + "content": "the International Conference on Computer Vision (ICCV), 2019.3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 53, + 95, + 288, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 95, + 288, + 128 + ], + "spans": [ + { + "bbox": [ + 53, + 95, + 288, + 128 + ], + "type": "text", + "content": "[27] Rohit Girdhar and Kristen Grauman. Anticipative video transformer. In Proceedings of the International Conference on Computer Vision (ICCV), 2021. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 129, + 288, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 129, + 288, + 172 + ], + "spans": [ + { + "bbox": [ + 53, + 129, + 288, + 172 + ], + "type": "text", + "content": "[28] Rohit Girdhar, Du Tran, Lorenzo Torresani, and Deva Ramanan. Distinct: Learning video representations without a single labeled video. In Proceedings of the International Conference on Computer Vision (ICCV), 2019. 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 173, + 288, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 173, + 288, + 216 + ], + "spans": [ + { + "bbox": [ + 53, + 173, + 288, + 216 + ], + "type": "text", + "content": "[29] Mohit Goyal, Sahil Modi, Rishabh Goyal, and Saurabh Gupta. Human hands as probes for interactive object understanding. In Proceedings of the Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 217, + 288, + 260 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 217, + 288, + 260 + ], + "spans": [ + { + "bbox": [ + 53, + 217, + 288, + 260 + ], + "type": "text", + "content": "[30] Kristen Grauman, Andrew Westbury, and Eugene Byrne et al. Ego4d: Around the World in 3,000 Hours of Egocentric Video. In Proceedings of the Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 1, 2, 4, 5, 6, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 261, + 288, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 261, + 288, + 303 + ], + "spans": [ + { + "bbox": [ + 53, + 261, + 288, + 303 + ], + "type": "text", + "content": "[31] Kristen Grauman, Andrew Westbury, et al. Ego-exo4d: Understanding skilled human activity from first-and third-person perspectives. arXiv preprint arXiv:2311.18259, 2023. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 304, + 288, + 337 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 304, + 288, + 337 + ], + "spans": [ + { + "bbox": [ + 53, + 304, + 288, + 337 + ], + "type": "text", + "content": "[32] Geoffrey Hinton, Oriol Vinyals, and Jeff Dean. Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531, 2015. 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 338, + 288, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 338, + 288, + 381 + ], + "spans": [ + { + "bbox": [ + 53, + 338, + 288, + 381 + ], + "type": "text", + "content": "[33] Hsuan-I Ho, Wei-Chen Chiu, and Yu-Chiang Frank Wang. Summarizing first-person videos from third persons' points of view. In Proceedings of the European Conference on Computer Vision (ECCV), 2018. 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 383, + 288, + 424 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 383, + 288, + 424 + ], + "spans": [ + { + "bbox": [ + 53, + 383, + 288, + 424 + ], + "type": "text", + "content": "[34] Nicola J Hodges, A Mark Williams, Spencer J Hayes, and Gavin Breslin. What is modelled during observational learning? Journal of sports sciences, 25(5):531-545, 2007. 1, 4" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 426, + 288, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 426, + 288, + 459 + ], + "spans": [ + { + "bbox": [ + 53, + 426, + 288, + 459 + ], + "type": "text", + "content": "[35] Jiang Hua, Liangcai Zeng, Gongfa Li, and Zhaojie Ju. Learning for a robot: Deep reinforcement learning, imitation learning, transfer learning. Sensors, 21(4):1278, 2021. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 53, + 460, + 288, + 492 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 460, + 288, + 492 + ], + "spans": [ + { + "bbox": [ + 53, + 460, + 288, + 492 + ], + "type": "text", + "content": "[36] Chien-Ming Huang, Sean Andrist, Allison Sauppé, and Bilge Mutlu. Using gaze patterns to predict task intent in collaboration. Frontiers in psychology, 6:1049, 2015. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 53, + 493, + 288, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 493, + 288, + 536 + ], + "spans": [ + { + "bbox": [ + 53, + 493, + 288, + 536 + ], + "type": "text", + "content": "[37] Yifei Huang, Minjie Cai, Zhenqiang Li, and Yoichi Sato. Predicting gaze in egocentric video by learning task-dependent attention transition. In Proceedings of the European Conference on Computer Vision (ECCV), 2018. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 53, + 537, + 288, + 581 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 537, + 288, + 581 + ], + "spans": [ + { + "bbox": [ + 53, + 537, + 288, + 581 + ], + "type": "text", + "content": "[38] Yifei Huang, Minjie Cai, Zhenqiang Li, Feng Lu, and Yoichi Sato. Mutual context network for jointly estimating egocentric gaze and action. IEEE Transactions on Image Processing, 29:7795-7806, 2020." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 53, + 582, + 288, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 582, + 288, + 624 + ], + "spans": [ + { + "bbox": [ + 53, + 582, + 288, + 624 + ], + "type": "text", + "content": "[39] Yifei Huang, Minjie Cai, and Yoichi Sato. An ego-vision system for discovering human joint attention. IEEE Transactions on Human-Machine Systems, 50(4):306-316, 2020. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 53, + 625, + 288, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 625, + 288, + 668 + ], + "spans": [ + { + "bbox": [ + 53, + 625, + 288, + 668 + ], + "type": "text", + "content": "[40] Yifei Huang, Lijin Yang, and Yoichi Sato. Compound prototype matching for few-shot action recognition. In Proceedings of the European Conference on Computer Vision (ECCV), 2022. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 53, + 669, + 288, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 669, + 288, + 712 + ], + "spans": [ + { + "bbox": [ + 53, + 669, + 288, + 712 + ], + "type": "text", + "content": "[41] Yifei Huang, Lijin Yang, and Yoichi Sato. Weakly supervised temporal sentence grounding with uncertainty-guided self-training. In Proceedings of the Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 3" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 312, + 72, + 547, + 714 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 312, + 72, + 547, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 72, + 547, + 116 + ], + "spans": [ + { + "bbox": [ + 312, + 72, + 547, + 116 + ], + "type": "text", + "content": "[42] Thomas E Hutchinson, K Preston White, Worthy N Martin, Kelly C Reichert, and Lisa A Frey. Human-computer interaction using eye-gaze input. IEEE Transactions on systems, man, and cybernetics, 19(6):1527-1534, 1989. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 312, + 118, + 547, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 118, + 547, + 172 + ], + "spans": [ + { + "bbox": [ + 312, + 118, + 547, + 172 + ], + "type": "text", + "content": "[43] Eric Jang, Alex Irpan, Mohi Khansari, Daniel Kappler, Frederik Ebert, Corey Lynch, Sergey Levine, and Chelsea Finn. Bc-z: Zero-shot task generalization with robotic imitation learning. In Proceedings of the Conference on Robot Learning, 2022. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 312, + 173, + 547, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 173, + 547, + 227 + ], + "spans": [ + { + "bbox": [ + 312, + 173, + 547, + 227 + ], + "type": "text", + "content": "[44] Youngkyoon Jang, Brian Sullivan, Casimir Ludwig, Iain Gilchrist, Dima Damen, and Walterio Mayol-Cuevas. Epicent: An egocentric video dataset for camping tent assembly. In Proceedings of the International Conference on Computer Vision Workshops (ICCVW), 2019. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 312, + 228, + 547, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 228, + 547, + 281 + ], + "spans": [ + { + "bbox": [ + 312, + 228, + 547, + 281 + ], + "type": "text", + "content": "[45] Baoxiong Jia, Yixin Chen, Siyuan Huang, Yixin Zhu, and Song-Chun Zhu. Lemma: A multi-view dataset for the arning multi-agent multi-task activities. In Proceedings of the European Conference on Computer Vision (ECCV), 2020. 2, 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 312, + 282, + 547, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 282, + 547, + 327 + ], + "spans": [ + { + "bbox": [ + 312, + 282, + 547, + 327 + ], + "type": "text", + "content": "[46] Baoxiong Jia, Ting Lei, Song-Chun Zhu, and Siyuan Huang. Egotaskqa: Understanding human tasks in egocentric videos. In Proceedings of the Advances in Neural Information Processing Systems (NeurIPS), 2022. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 312, + 327, + 547, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 327, + 547, + 369 + ], + "spans": [ + { + "bbox": [ + 312, + 327, + 547, + 369 + ], + "type": "text", + "content": "[47] Hao Jiang and Vamsi Krishna Ithapu. Egocentric pose estimation from human vision span. In Proceedings of the International Conference on Computer Vision (ICCV), 2021. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 312, + 371, + 547, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 371, + 547, + 415 + ], + "spans": [ + { + "bbox": [ + 312, + 371, + 547, + 415 + ], + "type": "text", + "content": "[48] Soo-Han Kang and Ji-Hyeong Han. Video captioning based on both egocentric and exocentric views of robot vision for human-robot interaction. International Journal of Social Robotics, pages 1–11, 2021. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 312, + 416, + 547, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 416, + 547, + 470 + ], + "spans": [ + { + "bbox": [ + 312, + 416, + 547, + 470 + ], + "type": "text", + "content": "[49] Moritz Kassner, William Patera, and Andreas Bulling. Pupil: An open source platform for pervasive eye tracking and mobile gaze-based interaction. In Adjunct Proceedings of the 2014 ACM International Joint Conference on Pervasive and Ubiquitous Computing, 2014. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 312, + 471, + 547, + 523 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 471, + 547, + 523 + ], + "spans": [ + { + "bbox": [ + 312, + 471, + 547, + 523 + ], + "type": "text", + "content": "[50] Evangelos Kazakos, Arsha Nagrani, Andrew Zisserman, and Dima Damen. Epic-fusion: Audio-visual temporal binding for egocentric action recognition. In Proceedings of the International Conference on Computer Vision (ICCV), 2019. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 312, + 525, + 547, + 580 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 525, + 547, + 580 + ], + "spans": [ + { + "bbox": [ + 312, + 525, + 547, + 580 + ], + "type": "text", + "content": "[51] Daekyum Kim, Brian Byunghyun Kang, Kyu Bum Kim, Hyungmin Choi, Jeesoo Ha, Kyu-Jin Cho, and Sungho Jo. Eyes are faster than hands: A soft wearable robot learns user intention from the egocentric view. Science Robotics, 4(26), 2019. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 312, + 581, + 547, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 581, + 547, + 635 + ], + "spans": [ + { + "bbox": [ + 312, + 581, + 547, + 635 + ], + "type": "text", + "content": "[52] Donghyun Kim, Yi-Hsuan Tsai, Bingbing Zhuang, Xiang Yu, Stan Sclaroff, Kate Saenko, and Manmohan Chandraker. Learning cross-modal contrastive features for video domain adaptation. In Proceedings of the International Conference on Computer Vision (ICCV), 2021. 7" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 312, + 636, + 547, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 636, + 547, + 689 + ], + "spans": [ + { + "bbox": [ + 312, + 636, + 547, + 689 + ], + "type": "text", + "content": "[53] Hildegard Kuehne, Hueihan Jhuang, Estíbaliz Garrote, Tomaso Poggio, and Thomas Serre. Hmdb: a large video database for human motion recognition. In Proceedings of the International Conference on Computer Vision (ICCV), 2011. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 312, + 691, + 547, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 691, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 312, + 691, + 547, + 714 + ], + "type": "text", + "content": "[54] Hilde Kuehne, Ali Arslan, and Thomas Serre. The language of actions: Recovering the syntax and semantics of goal" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "22081" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 53, + 72, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 72, + 72, + 288, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 72, + 72, + 288, + 106 + ], + "spans": [ + { + "bbox": [ + 72, + 72, + 288, + 106 + ], + "type": "text", + "content": "directed human activities. In Proceedings of the Conference on Computer Vision and Pattern Recognition (CVPR), 2014, 2, 7" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 53, + 107, + 288, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 107, + 288, + 162 + ], + "spans": [ + { + "bbox": [ + 53, + 107, + 288, + 162 + ], + "type": "text", + "content": "[55] Taein Kwon, Bugra Tekin, Jan Stühmer, Federica Bogo, and Marc Pollefeys. H2o: Two hands manipulating objects for first person interaction recognition. In Proceedings of the International Conference on Computer Vision (ICCV), 2021. 2, 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 163, + 288, + 208 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 163, + 288, + 208 + ], + "spans": [ + { + "bbox": [ + 53, + 163, + 288, + 208 + ], + "type": "text", + "content": "[56] Bolin Lai, Miao Liu, Fiona Ryan, and James M Rehg. In the eye of transformer: Global-local correlation for egocentric gaze estimation and beyond. International Journal of Computer Vision (IJCV), pages 1-18, 2023. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 209, + 288, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 209, + 288, + 262 + ], + "spans": [ + { + "bbox": [ + 53, + 209, + 288, + 262 + ], + "type": "text", + "content": "[57] Stephane Lallée, Eiichi Yoshida, Anthony Mallet, Francesco Nori, Lorenzo Natale, Giorgio Metta, Felix Warneken, and Peter Ford Dominey. Human-robot cooperation based on interaction learning. From motor learning to interaction learning in robots, pages 491–536, 2010. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 264, + 288, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 264, + 288, + 308 + ], + "spans": [ + { + "bbox": [ + 53, + 264, + 288, + 308 + ], + "type": "text", + "content": "[58] Kyle Lam, Junhong Chen, Zeyu Wang, Fahad M Iqbal, Ara Darzi, Benny Lo, Sanjay Purkayastha, and James M Kinross. Machine learning for technical skill assessment in surgery: a systematic review. NPJ digital medicine, 5(1):24, 2022. 8" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 310, + 288, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 310, + 288, + 363 + ], + "spans": [ + { + "bbox": [ + 53, + 310, + 288, + 363 + ], + "type": "text", + "content": "[59] Jangwon Lee and Michael S Ryoo. Learning robot activities from first-person human videos using convolutional future regression. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), 2017. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 365, + 288, + 420 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 365, + 288, + 420 + ], + "spans": [ + { + "bbox": [ + 53, + 365, + 288, + 420 + ], + "type": "text", + "content": "[60] Jie Lei, Linjie Li, Luowei Zhou, Zhe Gan, Tamara L Berg, Mohit Bansal, and Jingjing Liu. Less is more: Clipbert for video-and-language learning via sparse sampling. In Proceedings of the Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 422, + 288, + 477 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 422, + 288, + 477 + ], + "spans": [ + { + "bbox": [ + 53, + 422, + 288, + 477 + ], + "type": "text", + "content": "[61] Chengshu Li, Fei Xia, Roberto Martin-Martin, Michael Lingelbach, Sanjana Srivastava, Bokui Shen, Kent Vainio, Cem Gokmen, Gokul Dharan, Tanish Jain, et al. igibson 2.0: Object-centric simulation for robot learning of everyday household tasks. arXiv preprint arXiv:2108.03272, 2021. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 478, + 288, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 478, + 288, + 544 + ], + "spans": [ + { + "bbox": [ + 53, + 478, + 288, + 544 + ], + "type": "text", + "content": "[62] Chengshu Li, Ruohan Zhang, Josiah Wong, Cem Gokmen, Sanjana Srivastava, Roberto Martin-Martin, Chen Wang, Gabrael Levine, Michael Lingelbach, Jiankai Sun, et al. Behavior-1k: A benchmark for embodied ai with 1,000 everyday activities and realistic simulation. In Proceedings of the Conference on Robot Learning, 2023. 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 545, + 288, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 545, + 288, + 589 + ], + "spans": [ + { + "bbox": [ + 53, + 545, + 288, + 589 + ], + "type": "text", + "content": "[63] Haoxin Li, Yijun Cai, and Wei-Shi Zheng. Deep dual relation modeling for egocentric interaction recognition. In Proceedings of the Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 53, + 590, + 288, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 590, + 288, + 634 + ], + "spans": [ + { + "bbox": [ + 53, + 590, + 288, + 634 + ], + "type": "text", + "content": "[64] Jiayi Li, Tao Lu, Xiaoge Cao, Yinghao Cai, and Shuo Wang. Meta-imitation learning by watching video demonstrations. In Proceedings of the International Conference on Learning Representations, 2021. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 53, + 635, + 288, + 677 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 635, + 288, + 677 + ], + "spans": [ + { + "bbox": [ + 53, + 635, + 288, + 677 + ], + "type": "text", + "content": "[65] Yin Li, Alireza Fathi, and James M Rehg. Learning to predict gaze in egocentric video. In Proceedings of the International Conference on Computer Vision (ICCV), 2013. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 53, + 680, + 288, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 680, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 53, + 680, + 288, + 713 + ], + "type": "text", + "content": "[66] Yin Li, Zhefan Ye, and James M Rehg. Delving into egocentric actions. In Proceedings of the Conference on Computer Vision and Pattern Recognition (CVPR), 2015." + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 312, + 72, + 547, + 713 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 312, + 72, + 547, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 72, + 547, + 116 + ], + "spans": [ + { + "bbox": [ + 312, + 72, + 547, + 116 + ], + "type": "text", + "content": "[67] Yin Li, Miao Liu, and James M Rehg. In the eye of beholder: Joint learning of gaze and actions in first person video. In Proceedings of the European Conference on Computer Vision (ECCV), 2018. 3, 5" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 312, + 118, + 547, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 118, + 547, + 161 + ], + "spans": [ + { + "bbox": [ + 312, + 118, + 547, + 161 + ], + "type": "text", + "content": "[68] Yin Li, Miao Liu, and Jame Rehg. In the eye of the beholder: Gaze and actions in first person video. IEEE Transaction on Pattern Analysis and Machine Intelligence (TPAMI), 2021. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 312, + 163, + 547, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 163, + 547, + 217 + ], + "spans": [ + { + "bbox": [ + 312, + 163, + 547, + 217 + ], + "type": "text", + "content": "[69] Yanghao Li, Tushar Nagarajan, Bo Xiong, and Kristen Grauman. Ego-exo: Transferring visual representations from third-person to first-person videos. In Proceedings of the Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 3, 6, 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 312, + 220, + 547, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 220, + 547, + 262 + ], + "spans": [ + { + "bbox": [ + 312, + 220, + 547, + 262 + ], + "type": "text", + "content": "[70] Zhenqiang Li, Yifei Huang, Minjie Cai, and Yoichi Sato. Manipulation-skill assessment from videos with spatial attention network. In Proceedings of the International Conference on Computer Vision Workshops (ICCVW), 2019. 8" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 312, + 264, + 547, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 264, + 547, + 319 + ], + "spans": [ + { + "bbox": [ + 312, + 264, + 547, + 319 + ], + "type": "text", + "content": "[71] Zhenqiang Li, Lin Gu, Weimin Wang, Ryosuke Nakamura, and Yoichi Sato. Surgical skill assessment via video semantic aggregation. In Proceedings of the International Conference on Medical Image Computing and Computer-Assisted Intervention (MICCAI), 2022. 8" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 312, + 320, + 547, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 320, + 547, + 374 + ], + "spans": [ + { + "bbox": [ + 312, + 320, + 547, + 374 + ], + "type": "text", + "content": "[72] Kevin Qinghong Lin, Jinpeng Wang, Mattia Soldan, Michael Wray, Rui Yan, Eric Z XU, Difei Gao, Rong-Cheng Tu, Wenzhe Zhao, Weijie Kong, et al. Egocentric video-language pretraining. In Proceedings of the Advances in Neural Information Processing Systems (NeurIPS), 2022. 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 312, + 376, + 547, + 421 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 376, + 547, + 421 + ], + "spans": [ + { + "bbox": [ + 312, + 376, + 547, + 421 + ], + "type": "text", + "content": "[73] Daochang Liu, Qiyue Li, Tingting Jiang, Yizhou Wang, Rulin Miao, Fei Shan, and Ziyu Li. Towards unified surgical skill assessment. In Proceedings of the Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 8" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 312, + 422, + 547, + 476 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 422, + 547, + 476 + ], + "spans": [ + { + "bbox": [ + 312, + 422, + 547, + 476 + ], + "type": "text", + "content": "[74] Shaowei Liu, Subarna Tripathi, Somdeb Majumdar, and Xiaolong Wang. Joint hand motion and interaction hotspots prediction from egocentric videos. In Proceedings of the Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 312, + 478, + 547, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 478, + 547, + 511 + ], + "spans": [ + { + "bbox": [ + 312, + 478, + 547, + 511 + ], + "type": "text", + "content": "[75] Xiaotian Liu, Hector Palacios, and Christian Muise. Egocentric planning for scalable embodied task achievement. arXiv preprint arXiv:2306.01295, 2023. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 312, + 512, + 547, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 512, + 547, + 567 + ], + "spans": [ + { + "bbox": [ + 312, + 512, + 547, + 567 + ], + "type": "text", + "content": "[76] Yan Liu, Pei Yun Hsueh, Jennifer Lai, Mirweis Sangin, Marc Antoine Nussli, and Pierre Dillenbourg. Who is the expert? analyzing gaze data to predict expertise level in collaborative applications. In Proceedings of the IEEE International Conference on Multimedia & Expo (ICME), 2009. 8" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 312, + 568, + 547, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 568, + 547, + 611 + ], + "spans": [ + { + "bbox": [ + 312, + 568, + 547, + 611 + ], + "type": "text", + "content": "[77] Yueyue Liu, Zhijun Li, Huaping Liu, and Zhen Kan. Skill transfer learning for autonomous robots and human-robot cooperation: A survey. Robotics and Autonomous Systems, 128:103515, 2020. 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 312, + 613, + 547, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 613, + 547, + 669 + ], + "spans": [ + { + "bbox": [ + 312, + 613, + 547, + 669 + ], + "type": "text", + "content": "[78] Yunze Liu, Yun Liu, Che Jiang, Kangbo Lyu, Weikang Wan, Hao Shen, Boqiang Liang, Zhoujie Fu, He Wang, and Li Yi. Hoi4d: A 4d egocentric dataset for category-level human-object interaction. In Proceedings of the Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 312, + 670, + 547, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 670, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 312, + 670, + 547, + 713 + ], + "type": "text", + "content": "[79] Manolis Savva*, Abhishek Kadian*, Oleksandr Maksymets*, Yili Zhao, Erik Wijmans, Bhavana Jain, Julian Straub, Jia Liu, Vladlen Koltun, Jitendra Malik, Devi Parikh, and Dhruv Batra. Habitat: A Platform for Embodied AI" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "22082" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 53, + 72, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 72, + 72, + 287, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 72, + 72, + 287, + 95 + ], + "spans": [ + { + "bbox": [ + 72, + 72, + 287, + 95 + ], + "type": "text", + "content": "Research. In Proceedings of the International Conference on Computer Vision (ICCV), 2019. 1" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 53, + 97, + 287, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 97, + 287, + 140 + ], + "spans": [ + { + "bbox": [ + 53, + 97, + 287, + 140 + ], + "type": "text", + "content": "[80] Javier Marina-Miranda and V Javier Traver. Head and eye egocentric gesture recognition for human-robot interaction using eyewear cameras. IEEE Robotics and Automation Letters, 7(3):7067-7074, 2022. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 143, + 288, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 143, + 288, + 196 + ], + "spans": [ + { + "bbox": [ + 53, + 143, + 288, + 196 + ], + "type": "text", + "content": "[81] Joanna Materzynska, Guillaume Berger, Ingo Bax, and Roland Memisevic. The jester dataset: A large-scale video dataset of human gestures. In Proceedings of the International Conference on Computer Vision Workshops (ICCVW), 2019. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 200, + 288, + 265 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 200, + 288, + 265 + ], + "spans": [ + { + "bbox": [ + 53, + 200, + 288, + 265 + ], + "type": "text", + "content": "[82] Antoine Miech, Dimitri Zhukov, Jean-Baptiste Alayrac, Makarand Tapaswi, Ivan Laptev, and Josef Sivic. HowTo100M: Learning a Text-Video Embedding by Watching Hundred Million Narrated Video Clips. In Proceedings of the International Conference on Computer Vision (ICCV), 2019. 1, 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 268, + 287, + 312 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 268, + 287, + 312 + ], + "spans": [ + { + "bbox": [ + 53, + 268, + 287, + 312 + ], + "type": "text", + "content": "[83] Kyle Min and Jason J Corso. Integrating human gaze into attention for egocentric activity recognition. In Proceedings of the Winter Conference on Applications of Computer Vision (WACV), pages 1069-1078, 2021. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 314, + 288, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 314, + 288, + 357 + ], + "spans": [ + { + "bbox": [ + 53, + 314, + 288, + 357 + ], + "type": "text", + "content": "[84] Jonathan Munro and Dima Damen. Multi-modal domain adaptation for fine-grained action recognition. In Proceedings of the Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 360, + 287, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 360, + 287, + 403 + ], + "spans": [ + { + "bbox": [ + 53, + 360, + 287, + 403 + ], + "type": "text", + "content": "[85] Tushar Nagarajan and Kristen Grauman. Shaping embodied agent behavior with activity-context priors from egocentric video. In Proceedings of the Advances in Neural Information Processing Systems (NeurIPS), 2021. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 406, + 288, + 449 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 406, + 288, + 449 + ], + "spans": [ + { + "bbox": [ + 53, + 406, + 288, + 449 + ], + "type": "text", + "content": "[86] Tushar Nagarajan, Yanghao Li, Christoph Feichtenhofer, and Kristen Grauman. Ego-topo: Environment affordances from egocentric video. In Proceedings of the Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 453, + 287, + 506 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 453, + 287, + 506 + ], + "spans": [ + { + "bbox": [ + 53, + 453, + 287, + 506 + ], + "type": "text", + "content": "[87] Boxiao Pan, Haoye Cai, De-An Huang, Kuan-Hui Lee, Adrien Gaidon, Ehsan Adeli, and Juan Carlos Niebles. Spatio-temporal graph for video captioning with knowledge distillation. In Proceedings of the Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 510, + 287, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 510, + 287, + 553 + ], + "spans": [ + { + "bbox": [ + 53, + 510, + 287, + 553 + ], + "type": "text", + "content": "[88] Boxiao Pan, Zhangjie Cao, Ehsan Adeli, and Juan Carlos Niebles. Adversarial cross-domain action recognition with co-attention. In Proceedings of the AAAI Conference on Artificial Intelligence (AAAI), 2020. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 53, + 555, + 288, + 620 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 555, + 288, + 620 + ], + "spans": [ + { + "bbox": [ + 53, + 555, + 288, + 620 + ], + "type": "text", + "content": "[89] Boxiao Pan, Bokui Shen, Davis Rempe, Despoina Paschalidou, Kaichun Mo, Yanchao Yang, and Leonidas J Guibas. Copilot: Human-environment collision prediction and localization from egocentric videos. In Proceedings of the International Conference on Computer Vision (ICCV), 2023. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 53, + 623, + 288, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 623, + 288, + 666 + ], + "spans": [ + { + "bbox": [ + 53, + 623, + 288, + 666 + ], + "type": "text", + "content": "[90] Hyung Min Park, Seok Han Lee, and Jong Soo Choi. Wearable augmented reality system using gaze interaction. In Proceedings of the IEEE/ACM International Symposium on Mixed and Augmented Reality, 2008. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 53, + 670, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 670, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 53, + 670, + 287, + 713 + ], + "type": "text", + "content": "[91] Paritosh Parmar and Brendan Tran Morris. What and how well you performed? a multitask learning approach to action quality assessment. In Proceedings of the Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 8" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 547, + 713 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 312, + 73, + 545, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 73, + 545, + 105 + ], + "spans": [ + { + "bbox": [ + 312, + 73, + 545, + 105 + ], + "type": "text", + "content": "[92] Leo Pauly, Wisdom C Agboh, David C Hogg, and Raul Fuentes. O2a: one-shot observational learning with action vectors. Frontiers in Robotics and AI, 2021. 1" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 312, + 107, + 545, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 107, + 545, + 171 + ], + "spans": [ + { + "bbox": [ + 312, + 107, + 545, + 171 + ], + "type": "text", + "content": "[93] Chiara Plizzari, Mirco Planamente, Gabriele Goletto, Marco Cannici, Emanuele Gusso, Matteo Matteucci, and Barbara Caputo. E2 (go) motion: Motion augmented event stream for egocentric action recognition. In Proceedings of the Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 312, + 173, + 547, + 225 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 173, + 547, + 225 + ], + "spans": [ + { + "bbox": [ + 312, + 173, + 547, + 225 + ], + "type": "text", + "content": "[94] Chiara Plizzari, Gabriele Goletto, Antonino Furnari, Siddhant Bansal, Francesco Ragusa, Giovanni Maria Farinella, Dima Damen, and Tatiana Tommasi. An outlook into the future of egocentric vision. arXiv preprint arXiv:2308.07123, 2023. 1, 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 312, + 228, + 547, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 228, + 547, + 258 + ], + "spans": [ + { + "bbox": [ + 312, + 228, + 547, + 258 + ], + "type": "text", + "content": "[95] David Premack and Guy Woodruff. Does the chimpanzee have a theory of mind? Behavioral and brain sciences, 1(4): 515-526, 1978. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 312, + 261, + 547, + 313 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 261, + 547, + 313 + ], + "spans": [ + { + "bbox": [ + 312, + 261, + 547, + 313 + ], + "type": "text", + "content": "[96] Gorjan Radevski, Dusan Grujicic, Matthew Blaschko, Marie-Francine Moens, and Tinne Tuytelaars. Multimodal distillation for egocentric action recognition. In Proceedings of the International Conference on Computer Vision (ICCV), 2023. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 312, + 316, + 547, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 316, + 547, + 381 + ], + "spans": [ + { + "bbox": [ + 312, + 316, + 547, + 381 + ], + "type": "text", + "content": "[97] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In Proceedings of the International Conference on Machine Learning (ICML), 2021. 5, 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 312, + 383, + 547, + 446 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 383, + 547, + 446 + ], + "spans": [ + { + "bbox": [ + 312, + 383, + 547, + 446 + ], + "type": "text", + "content": "[98] Francesco Ragusa, Antonino Furnari, Salvatore Livatino, and Giovanni Maria Farinella. The meccano dataset: Understanding human-object interactions from egocentric videos in an industrial-like domain. In Proceedings of the Winter Conference on Applications of Computer Vision (WACV), 2021. 2, 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 312, + 449, + 547, + 513 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 449, + 547, + 513 + ], + "spans": [ + { + "bbox": [ + 312, + 449, + 547, + 513 + ], + "type": "text", + "content": "[99] Nishant Rai, Haofeng Chen, Jingwei Ji, Rishi Desai, Kazuki Kozuka, Shun Ishizaka, Ehsan Adeli, and Juan Carlos Niebles. Home action genome: Cooperative compositional action understanding. In Proceedings of the Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 1, 2, 5" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 514, + 547, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 514, + 547, + 557 + ], + "spans": [ + { + "bbox": [ + 308, + 514, + 547, + 557 + ], + "type": "text", + "content": "[100] Richard Ramsey, David M Kaplan, and Emily S Cross. Watch and learn: the cognitive neuroscience of learning from others' actions. Trends in Neurosciences, 44(6):478-491, 2021. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 559, + 547, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 559, + 547, + 602 + ], + "spans": [ + { + "bbox": [ + 308, + 559, + 547, + 602 + ], + "type": "text", + "content": "[101] Yosef Razin and Karen Feigh. Learning to predict intent from gaze during robotic hand-eye coordination. In Proceedings of the AAAI Conference on Artificial Intelligence (AAAI), 2017. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 604, + 547, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 604, + 547, + 658 + ], + "spans": [ + { + "bbox": [ + 308, + 604, + 547, + 658 + ], + "type": "text", + "content": "[102] Radiah Rivu, Yasmeen Abdrabou, Ken Pfeuffer, Augusto Esteves, Stefanie Meitner, and Florian Alt. Stare: gaze-assisted face-to-face communication in augmented reality. In Proceedings of the ACM Symposium on Eye Tracking Research and Applications (ETRA), 2020. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 659, + 547, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 659, + 547, + 689 + ], + "spans": [ + { + "bbox": [ + 308, + 659, + 547, + 689 + ], + "type": "text", + "content": "[103] Giacomo Rizzolatti and Laila Craighero. The mirror-neuron system. Annual Review of Neuroscience, 27:169-192, 2004. 1" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 692, + 547, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 692, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 692, + 547, + 713 + ], + "type": "text", + "content": "[104] Aadarsh Sahoo, Rutav Shah, Rameswar Panda, Kate Saenko, and Abir Das. Contrast and mix: Temporal contrastive video" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "22083" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 73, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 72, + 73, + 287, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 72, + 73, + 287, + 106 + ], + "spans": [ + { + "bbox": [ + 72, + 73, + 287, + 106 + ], + "type": "text", + "content": "domain adaptation with background mixing. In Proceedings of the Advances in Neural Information Processing Systems (NeurIPS), 2021. 7" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 107, + 287, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 107, + 287, + 138 + ], + "spans": [ + { + "bbox": [ + 48, + 107, + 287, + 138 + ], + "type": "text", + "content": "[105] Stefan Schaal. Learning from demonstration. In Proceedings of the Advances in Neural Information Processing Systems (NeurIPS), 1996. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 140, + 287, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 140, + 287, + 194 + ], + "spans": [ + { + "bbox": [ + 49, + 140, + 287, + 194 + ], + "type": "text", + "content": "[106] F. Sener, D. Chatterjee, D. Shelepov, K. He, D. Singhania, R. Wang, and A. Yao. Assembly101: A large-scale multiview video dataset for understanding procedural activities. In Proceedings of the Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 1, 2, 4, 5, 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 195, + 288, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 195, + 288, + 239 + ], + "spans": [ + { + "bbox": [ + 49, + 195, + 288, + 239 + ], + "type": "text", + "content": "[107] Dandan Shan, Jiaqi Geng, Michelle Shu, and David F Fouhey. Understanding human hands in contact at internet scale. In Proceedings of the Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 49, + 239, + 288, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 239, + 288, + 304 + ], + "spans": [ + { + "bbox": [ + 49, + 239, + 288, + 304 + ], + "type": "text", + "content": "[108] Mohit Shridhar, Jesse Thomason, Daniel Gordon, Yonatan Bisk, Winson Han, Roozbeh Mottaghi, Luke Zettlemoyer, and Dieter Fox. ALFRED: A Benchmark for Interpreting Grounded Instructions for Everyday Tasks. In Proceedings of the Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 49, + 305, + 287, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 305, + 287, + 348 + ], + "spans": [ + { + "bbox": [ + 49, + 305, + 287, + 348 + ], + "type": "text", + "content": "[109] Gunnar A Sigurdsson, Abhinav Gupta, Cordelia Schmid, Ali Farhadi, and Karteek Alahari. Charades-ego: A large-scale dataset of paired third and first person videos. arXiv preprint arXiv:1804.09626, 2018. 1, 2, 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 49, + 350, + 287, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 350, + 287, + 403 + ], + "spans": [ + { + "bbox": [ + 49, + 350, + 287, + 403 + ], + "type": "text", + "content": "[110] Kiran Somasundaram, Jing Dong, Huixuan Tang, Julian Straub, Mingfei Yan, Michael Goesele, Jakob Julian Engel, Renzo De Nardi, and Richard Newcombe. Project aria: A new tool for egocentric multi-modal ai research. arXiv preprint arXiv:2308.13561, 2023. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 49, + 404, + 287, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 404, + 287, + 459 + ], + "spans": [ + { + "bbox": [ + 49, + 404, + 287, + 459 + ], + "type": "text", + "content": "[111] Xiaolin Song, Sicheng Zhao, Jingyu Yang, Huanjing Yue, Pengfei Xu, Runbo Hu, and Hua Chai. Spatio-temporal contrastive domain adaptation for action recognition. In Proceedings of the Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 49, + 460, + 288, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 460, + 288, + 514 + ], + "spans": [ + { + "bbox": [ + 49, + 460, + 288, + 514 + ], + "type": "text", + "content": "[112] Yale Song, Eugene Byrne, Tushar Nagarajan, Huiyu Wang, Miguel Martin, and Lorenzo Torresani. Ego4d goal-step: Toward hierarchical understanding of procedural activities. Proceedings of the Advances in Neural Information Processing Systems (NeurIPS), 2024. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 49, + 514, + 287, + 558 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 514, + 287, + 558 + ], + "spans": [ + { + "bbox": [ + 49, + 514, + 287, + 558 + ], + "type": "text", + "content": "[113] Swathikiran Sudhakaran, Sergio Escalera, and Oswald Lanz. Lsta: Long short-term attention for egocentric action recognition. In Proceedings of the Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 49, + 559, + 288, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 559, + 288, + 613 + ], + "spans": [ + { + "bbox": [ + 49, + 559, + 288, + 613 + ], + "type": "text", + "content": "[114] Flood Sung, Yongxin Yang, Li Zhang, Tao Xiang, Philip HS Torr, and Timothy M Hospedales. Learning to compare: Relation network for few-shot learning. In Proceedings of the Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 8" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 49, + 614, + 288, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 614, + 288, + 668 + ], + "spans": [ + { + "bbox": [ + 49, + 614, + 288, + 668 + ], + "type": "text", + "content": "[115] Yansong Tang, Zanlin Ni, Jiahuan Zhou, Danyang Zhang, Jiwen Lu, Ying Wu, and Jie Zhou. Uncertainty-aware score distribution learning for action quality assessment. In Proceedings of the Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 8" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 49, + 670, + 287, + 691 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 670, + 287, + 691 + ], + "spans": [ + { + "bbox": [ + 49, + 670, + 287, + 691 + ], + "type": "text", + "content": "[116] Yuli Vasiliev. Natural language processing with Python and spaCy: A practical introduction. 2020. 4" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 49, + 692, + 288, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 692, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 49, + 692, + 288, + 713 + ], + "type": "text", + "content": "[117] Shruti Vyas, Yogesh S Rawat, and Mubarak Shah. Multiview action recognition using cross-view video prediction." + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 331, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 331, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 331, + 73, + 545, + 95 + ], + "type": "text", + "content": "In Proceedings of the European Conference on Computer Vision (ECCV), 2020. 7" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 96, + 545, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 96, + 545, + 140 + ], + "spans": [ + { + "bbox": [ + 308, + 96, + 545, + 140 + ], + "type": "text", + "content": "[118] Sebastian Wallkötter, Silvia Tulli, Ginevra Castellano, Ana Paiva, and Mohamed Chetouani. Explainable embodied agents through social cues: a review. ACM Transactions on Human-Robot Interaction (THRI), 10(3):1-24, 2021. 5" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 141, + 545, + 186 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 141, + 545, + 186 + ], + "spans": [ + { + "bbox": [ + 308, + 141, + 545, + 186 + ], + "type": "text", + "content": "[119] Chen Wang, Linxi Fan, Jiankai Sun, Ruohan Zhang, Li Fei-Fei, Danfei Xu, Yuke Zhu, and Anima Anandkumar. Mimicplay: Long-horizon imitation learning by watching human play. arXiv preprint arXiv:2302.12422, 2023. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 186, + 545, + 231 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 186, + 545, + 231 + ], + "spans": [ + { + "bbox": [ + 308, + 186, + 545, + 231 + ], + "type": "text", + "content": "[120] Huiyu Wang, Mitesh Kumar Singh, and Lorenzo Torresani. Ego-only: Egocentric action detection without exocentric transferring. In Proceedings of the International Conference on Computer Vision (ICCV), 2023. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 232, + 545, + 275 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 232, + 545, + 275 + ], + "spans": [ + { + "bbox": [ + 308, + 232, + 545, + 275 + ], + "type": "text", + "content": "[121] Jiahao Wang, Guo Chen, Yifei Huang, Limin Wang, and Tong Lu. Memory-and-anticipation transformer for online action understanding. In Proceedings of the International Conference on Computer Vision (ICCV), 2023. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 277, + 545, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 277, + 545, + 342 + ], + "spans": [ + { + "bbox": [ + 308, + 277, + 545, + 342 + ], + "type": "text", + "content": "[122] Jinpeng Wang, Yixiao Ge, Rui Yan, Yuying Ge, Kevin Qinghong Lin, Satoshi Tsutsui, Xudong Lin, Guanyu Cai, Jianping Wu, Ying Shan, et al. All in one: Exploring unified video-language pre-training. In Proceedings of the Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 344, + 545, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 344, + 545, + 398 + ], + "spans": [ + { + "bbox": [ + 308, + 344, + 545, + 398 + ], + "type": "text", + "content": "[123] Jian Wang, Diogo Luvizon, Weipeng Xu, Lingjie Liu, Kripasindhu Sarkar, and Christian Theobalt. Scene-aware egocentric 3d human pose estimation. In Proceedings of the Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 399, + 545, + 454 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 399, + 545, + 454 + ], + "spans": [ + { + "bbox": [ + 308, + 399, + 545, + 454 + ], + "type": "text", + "content": "[124] Qitong Wang, Long Zhao, Liangzhe Yuan, Ting Liu, and Xi Peng. Learning from semantic alignment between unpaired multiviews for egocentric video recognition. In Proceedings of the International Conference on Computer Vision (ICCV), 2023. 1, 2, 3, 6, 7" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 456, + 545, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 456, + 545, + 499 + ], + "spans": [ + { + "bbox": [ + 308, + 456, + 545, + 499 + ], + "type": "text", + "content": "[125] Xiaohan Wang, Linchao Zhu, Heng Wang, and Yi Yang. Interactive prototype learning for egocentric action recognition. In Proceedings of the International Conference on Computer Vision (ICCV), 2021. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 501, + 545, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 501, + 545, + 578 + ], + "spans": [ + { + "bbox": [ + 308, + 501, + 545, + 578 + ], + "type": "text", + "content": "[126] Xin Wang, Taein Kwon, Mahdi Rad, Bowen Pan, Ishani Chakraborty, Sean Andrist, Dan Bohus, Ashley Feniello, Bugra Tekin, Felipe Vieira Frujeri, Neel Joshi, and Marc Pollefeys. Holoassist: an egocentric human interaction dataset for interactive ai assistants in the real world. In Proceedings of the International Conference on Computer Vision (ICCV), 2023. 1, 4, 5" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 579, + 545, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 579, + 545, + 633 + ], + "spans": [ + { + "bbox": [ + 308, + 579, + 545, + 633 + ], + "type": "text", + "content": "[127] Xiaohan Wang, Yuehu Liu, Xinhang Song, Beibei Wang, and Shuqiang Jiang. Generating explanations for embodied action decision from visual observation. In Proceedings of the 31st ACM International Conference on Multimedia, 2023. 5" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 635, + 545, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 635, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 308, + 635, + 545, + 689 + ], + "type": "text", + "content": "[128] Yeping Wang, Gopika Ajaykumar, and Chien-Ming Huang. See what i see: Enabling user-centric robotic assistance using first-person demonstrations. In Proceedings of the 2020 ACM/IEEE International Conference on Human-Robot Interaction, pages 639-648, 2020. 1" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 691, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 691, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 691, + 545, + 713 + ], + "type": "text", + "content": "[129] Yi Wang, Kunchang Li, Yizhuo Li, Yinan He, Bingkun Huang, Zhiyu Zhao, Hongjie Zhang, Jilan Xu, Yi Liu, Zun" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "22084" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 714 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 72, + 72, + 287, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 72, + 72, + 287, + 106 + ], + "spans": [ + { + "bbox": [ + 72, + 72, + 287, + 106 + ], + "type": "text", + "content": "Wang, et al. Intervideo: General video foundation models via generative and discriminative learning. arXiv preprint arXiv:2212.03191, 2022. 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 106, + 288, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 106, + 288, + 149 + ], + "spans": [ + { + "bbox": [ + 48, + 106, + 288, + 149 + ], + "type": "text", + "content": "[130] Zirui Wang, Jiahui Yu, Adams Wei Yu, Zihang Dai, Yulia Tsvetkov, and Yuan Cao. Simvlm: Simple visual language model pretraining with weak supervision. arXiv preprint arXiv:2108.10904, 2021. 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 150, + 288, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 150, + 288, + 205 + ], + "spans": [ + { + "bbox": [ + 48, + 150, + 288, + 205 + ], + "type": "text", + "content": "[131] Pengfei Wei, Lingdong Kong, Xinghua Qu, Yi Ren, Jing Jiang, Xiang Yin, et al. Unsupervised video domain adaptation for action recognition: A disentanglement perspective. In Proceedings of the Advances in Neural Information Processing Systems (NeurIPS), 2023. 2, 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 205, + 288, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 205, + 288, + 249 + ], + "spans": [ + { + "bbox": [ + 48, + 205, + 288, + 249 + ], + "type": "text", + "content": "[132] Daniel Weinland, Mustafa Özuysal, and Pascal Fua. Making action recognition robust to occlusions and viewpoint changes. In Proceedings of the European Conference on Computer Vision (ECCV), 2010. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 250, + 288, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 250, + 288, + 304 + ], + "spans": [ + { + "bbox": [ + 48, + 250, + 288, + 304 + ], + "type": "text", + "content": "[133] Mark Wilson, John Mcgrath, Samuel Vine, James Brewer, David Defriend, and Richard Masters. Psychomotor control in a virtual laparoscopic surgery training environment: gaze control parameters differentiate novices from experts. Surgical Endoscopy, 24(10):2458-2464, 2010. 8" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 305, + 288, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 305, + 288, + 360 + ], + "spans": [ + { + "bbox": [ + 48, + 305, + 288, + 360 + ], + "type": "text", + "content": "[134] Benita Wong, Joya Chen, You Wu, Stan Weixian Lei, Dongxing Mao, Difei Gao, and Mike Zheng Shou. Assistq: Affordance-centric question-driven task completion for egocentric assistant. In Proceedings of the European Conference on Computer Vision (ECCV), 2022. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 360, + 288, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 360, + 288, + 403 + ], + "spans": [ + { + "bbox": [ + 48, + 360, + 288, + 403 + ], + "type": "text", + "content": "[135] Xinxiao Wu, Han Wang, Cuiwei Liu, and Yunde Jia. Cross-view action recognition over heterogeneous feature spaces. In Proceedings of the International Conference on Computer Vision (ICCV), 2013. 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 404, + 288, + 448 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 404, + 288, + 448 + ], + "spans": [ + { + "bbox": [ + 48, + 404, + 288, + 448 + ], + "type": "text", + "content": "[136] Haifeng Xia, Pu Wang, and Zhengming Ding. Incomplete multi-view domain adaptation via channel enhancement and knowledge transfer. In Proceedings of the European Conference on Computer Vision (ECCV), 2022. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 449, + 288, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 449, + 288, + 491 + ], + "spans": [ + { + "bbox": [ + 48, + 449, + 288, + 491 + ], + "type": "text", + "content": "[137] Jianjia Xin, Lichun Wang, Kai Xu, Chao Yang, and Baocai Yin. Learning interaction regions and motion trajectories simultaneously from egocentric demonstration videos. IEEE Robotics and Automation Letters, 2023. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 492, + 288, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 492, + 288, + 547 + ], + "spans": [ + { + "bbox": [ + 48, + 492, + 288, + 547 + ], + "type": "text", + "content": "[138] Hu Xu, Gargi Ghosh, Po-Yao Huang, Dmytro Okhonko, Armen Aghajanyan, Florian Metze, Luke Zettlemoyer, and Christoph Feichtenhofer. Videoclip: Contrastive pre-training for zero-shot video-text understanding. arXiv preprint arXiv:2109.14084, 2021. 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 548, + 288, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 548, + 288, + 602 + ], + "spans": [ + { + "bbox": [ + 48, + 548, + 288, + 602 + ], + "type": "text", + "content": "[139] Jia Xu, Lopamudra Mukherjee, Yin Li, Jamieson Warner, James M Rehg, and Vikas Singh. Gaze-enabled egocentric video summarization via constrained submodular maximization. In Proceedings of the Conference on Computer Vision and Pattern Recognition (CVPR), 2015. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 602, + 288, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 602, + 288, + 635 + ], + "spans": [ + { + "bbox": [ + 48, + 602, + 288, + 635 + ], + "type": "text", + "content": "[140] Jilan Xu, Yifei Huang, Junlin Hou, Guo Chen, Yuejie Zhang, Rui Feng, and Weidi Xie. Retrieval-augmented egocentric video captioning. arXiv preprint arXiv:2401.00789, 2024.3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 636, + 288, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 636, + 288, + 690 + ], + "spans": [ + { + "bbox": [ + 48, + 636, + 288, + 690 + ], + "type": "text", + "content": "[141] Mingze Xu, Chenyou Fan, Yuchen Wang, Michael S Ryoo, and David J Crandall. Joint person segmentation and identification in synchronized first-and third-person videos. In Proceedings of the European Conference on Computer Vision (ECCV), 2018. 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 691, + 288, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 288, + 714 + ], + "type": "text", + "content": "[142] Hongwei Xue, Tiankai Hang, Yanhong Zeng, Yuchong Sun, Bei Liu, Huan Yang, Jianlong Fu, and Baining Guo." + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 714 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 331, + 73, + 545, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 331, + 73, + 545, + 117 + ], + "spans": [ + { + "bbox": [ + 331, + 73, + 545, + 117 + ], + "type": "text", + "content": "Advancing high-resolution video-language representation with large-scale video transcriptions. In Proceedings of the Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 118, + 547, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 547, + 172 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 547, + 172 + ], + "type": "text", + "content": "[143] Zihui Xue and Kristen Grauman. Learning fine-grained view-invariant representations from unpaired ego-exo videos via temporal alignment. In Proceedings of the Advances in Neural Information Processing Systems (NeurIPS), 2023. 1, 2, 6, 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 175, + 547, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 175, + 547, + 239 + ], + "spans": [ + { + "bbox": [ + 307, + 175, + 547, + 239 + ], + "type": "text", + "content": "[144] Takuma Yagi, Misaki Ohashi, Yifei Huang, Ryosuke Furuta, Shungo Adachi, Toutai Mitsuyama, and Yoichi Sato. Finebio: A fine-grained video dataset of biological experiments with hierarchical annotations. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), 2023. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 241, + 547, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 241, + 547, + 296 + ], + "spans": [ + { + "bbox": [ + 307, + 241, + 547, + 296 + ], + "type": "text", + "content": "[145] Lijin Yang, Yifei Huang, Yusuke Sugano, and Yoichi Sato. Interact before align: Leveraging cross-modal knowledge for domain adaptive action recognition. In Proceedings of the Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 297, + 547, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 297, + 547, + 342 + ], + "spans": [ + { + "bbox": [ + 307, + 297, + 547, + 342 + ], + "type": "text", + "content": "[146] Sherry Yang, Jacob Walker, Jack Parker-Holder, Yilun Du, Jake Bruce, Andre Barreto, Pieter Abbeel, and Dale Schuurmans. Video as the new language for real-world decision making. arXiv preprint arXiv:2402.17139, 2024. 5" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 342, + 545, + 375 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 342, + 545, + 375 + ], + "spans": [ + { + "bbox": [ + 307, + 342, + 545, + 375 + ], + "type": "text", + "content": "[147] Li Yingzhen and Stephan Mandt. Disentangled sequential autoencoder. In Proceedings of the International Conference on Machine Learning (ICML), 2018. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 376, + 547, + 421 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 376, + 547, + 421 + ], + "spans": [ + { + "bbox": [ + 307, + 376, + 547, + 421 + ], + "type": "text", + "content": "[148] Tianhe Yu, Chelsea Finn, Annie Xie, Sudeep Dasari, Tianhao Zhang, Pieter Abbeel, and Sergey Levine. One-shot imitation from observing humans via domain-adaptive meta-learning. arXiv preprint arXiv:1802.01557, 2018. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 422, + 545, + 465 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 422, + 545, + 465 + ], + "spans": [ + { + "bbox": [ + 307, + 422, + 545, + 465 + ], + "type": "text", + "content": "[149] Xumin Yu, Yongming Rao, Wenliang Zhao, Jiwen Lu, and Jie Zhou. Group-aware contrastive regression for action quality assessment. In Proceedings of the International Conference on Computer Vision (ICCV), 2021. 8" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 467, + 547, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 467, + 547, + 521 + ], + "spans": [ + { + "bbox": [ + 307, + 467, + 547, + 521 + ], + "type": "text", + "content": "[150] Zecheng Yu, Yifei Huang, Ryosuke Furuta, Takuma Yagi, Yusuke Goutsu, and Yoichi Sato. Fine-grained affordance annotation for egocentric hand-object interaction videos. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, 2023. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 522, + 547, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 522, + 547, + 578 + ], + "spans": [ + { + "bbox": [ + 307, + 522, + 547, + 578 + ], + "type": "text", + "content": "[151] Thorsten O Zander, Matti Gaertner, Christian Kothe, and Roman Vilimek. Combining eye gaze input with a brain-computer interface for touchless human-computer interaction. International Journal of Human-Computer Interaction, 27(1):38-51, 2010. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 579, + 547, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 579, + 547, + 633 + ], + "spans": [ + { + "bbox": [ + 307, + 579, + 547, + 633 + ], + "type": "text", + "content": "[152] Rowan Zellers, Ximing Lu, Jack Hessel, Youngjae Yu, Jae Sung Park, Jize Cao, Ali Farhadi, and Yejin Choi. Merlot: Multimodal neural script knowledge models. In Proceedings of the Advances in Neural Information Processing Systems (NeurIPS), 2021. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 635, + 547, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 635, + 547, + 679 + ], + "spans": [ + { + "bbox": [ + 307, + 635, + 547, + 679 + ], + "type": "text", + "content": "[153] Lingzhi Zhang, Shenghao Zhou, Simon Stent, and Jianbo Shi. Fine-grained egocentric hand-object segmentation: Dataset, model, and applications. In Proceedings of the European Conference on Computer Vision (ECCV), 2022. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 681, + 547, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 681, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 307, + 681, + 547, + 714 + ], + "type": "text", + "content": "[154] Siwei Zhang, Qianli Ma, Yan Zhang, Zhiyin Qian, Taein Kwon, Marc Pollefeys, Federica Bogo, and Siyu Tang. Ego-body: Human body shape and motion of interacting people" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "22085" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 385 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 72, + 72, + 287, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 72, + 72, + 287, + 95 + ], + "spans": [ + { + "bbox": [ + 72, + 72, + 287, + 95 + ], + "type": "text", + "content": "from head-mounted devices. In Proceedings of the European Conference on Computer Vision (ECCV), 2022. 2, 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 96, + 287, + 148 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 96, + 287, + 148 + ], + "spans": [ + { + "bbox": [ + 48, + 96, + 287, + 148 + ], + "type": "text", + "content": "[155] Weiyu Zhang, Menglong Zhu, and KG Derpanis. From actemes to action: A strongly supervised representation for detailed action understanding. In Proceedings of the International Conference on Computer Vision (ICCV), 2013." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 152, + 288, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 152, + 288, + 195 + ], + "spans": [ + { + "bbox": [ + 49, + 152, + 288, + 195 + ], + "type": "text", + "content": "[156] Zehua Zhang, David Crandall, Michael Proulx, Sachin Talathi, and Abhishek Sharma. Can gaze inform egocentric action recognition? In Proceedings of the Symposium on Eye Tracking Research and Applications (ETRA), 2022. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 197, + 288, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 197, + 288, + 239 + ], + "spans": [ + { + "bbox": [ + 49, + 197, + 288, + 239 + ], + "type": "text", + "content": "[157] Zhenning Zhang, Zhigeng Pan, Weiqing Li, and Zhiyong Su. X-board: an egocentric adaptive ar assistant for perception in indoor environments. Virtual Reality, 27(2):1327-1343, 2023. 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 49, + 241, + 288, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 241, + 288, + 285 + ], + "spans": [ + { + "bbox": [ + 49, + 241, + 288, + 285 + ], + "type": "text", + "content": "[158] Yue Zhao, Ishan Misra, Philipp Krahenbuhl, and Rohit Girdhar. Learning video representations from large language models. In Proceedings of the Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 49, + 286, + 287, + 329 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 286, + 287, + 329 + ], + "spans": [ + { + "bbox": [ + 49, + 286, + 287, + 329 + ], + "type": "text", + "content": "[159] Jingjing Zheng and Zhuolin Jiang. Learning view-invariant sparse representations for cross-view action recognition. In Proceedings of the International Conference on Computer Vision (ICCV), 2013. 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 331, + 288, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 331, + 288, + 385 + ], + "spans": [ + { + "bbox": [ + 48, + 331, + 288, + 385 + ], + "type": "text", + "content": "[160] Yang Zheng, Yanchao Yang, Kaichun Mo, Jiaman Li, Tao Yu, Yebin Liu, Karen Liu, and Leonidas J Guibas. Gimo: Gaze-informed human motion prediction in context. In Proceedings of the European Conference on Computer Vision (ECCV), 2022. 3" + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "22086" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/EgoGen_ An Egocentric Synthetic Data Generator/717ad579-fa9d-4ce5-996d-77c046c2cdbf_content_list.json b/2024/EgoGen_ An Egocentric Synthetic Data Generator/717ad579-fa9d-4ce5-996d-77c046c2cdbf_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..4e0b479776dbb91836519f44367f43519a52d1c2 --- /dev/null +++ b/2024/EgoGen_ An Egocentric Synthetic Data Generator/717ad579-fa9d-4ce5-996d-77c046c2cdbf_content_list.json @@ -0,0 +1,1717 @@ +[ + { + "type": "text", + "text": "EgoGen: An Egocentric Synthetic Data Generator", + "text_level": 1, + "bbox": [ + 230, + 130, + 740, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Gen Li $^{1}$ Kaifeng Zhao $^{1}$ Siwei Zhang $^{1}$ Xiaozhong Lyu $^{1}$ \nMihai Dusmanu $^{2}$ Yan Zhang $^{1}$ Marc Pollefeys $^{1,2}$ Siyu Tang $^{1}$ $^{1}$ ETH Zürich \n $^{2}$ Microsoft \nhttps://ego-gen.github.io/", + "bbox": [ + 225, + 179, + 754, + 256 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/20f96637edb2d3d562a65def7cff75f240c545540ad7a1c0a8763394d5c27659.jpg", + "image_caption": [ + "RGB", + "RGB with Motion Blur", + "Depth" + ], + "image_footnote": [], + "bbox": [ + 78, + 273, + 483, + 483 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/73eaa166df6ddb80755c410a54a4b1ab7074c0acff1935e1445efd19d432d69c.jpg", + "image_caption": [ + "Surface Normal", + "Figure 1. EgoGen: a scalable synthetic data generation system for egocentric perception tasks, with rich multi-modal data and accurate annotations. We simulate camera rigs for head-mounted devices (HMDs) and render from the perspective of the camera wearer with various sensors. Top to bottom: middle and right camera sensors in the rig. Left to right: photo-realistic RGB image, RGB with simulated motion blur, depth map, surface normal, segmentation mask, and world position for fisheye cameras widely used in HMDs." + ], + "image_footnote": [], + "bbox": [ + 483, + 273, + 617, + 483 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/5d1f2404d006586ba9eb281ec93f4ef6684b3f107eb264300d5fb6b0cd250652.jpg", + "image_caption": [ + "Segmentation Mask" + ], + "image_footnote": [], + "bbox": [ + 617, + 273, + 751, + 483 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/a97948112dd8e0dedfe94b08c70c361c3b1c45a52a4fc02f4560d6042536f524.jpg", + "image_caption": [ + "World Position" + ], + "image_footnote": [], + "bbox": [ + 753, + 273, + 890, + 483 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 584, + 313, + 599 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Understanding the world in first-person view is fundamental in Augmented Reality (AR). This immersive perspective brings dramatic visual changes and unique challenges compared to third-person views. Synthetic data has empowered third-person-view vision models, but its application to embodied egocentric perception tasks remains largely unexplored. A critical challenge lies in simulating natural human movements and behaviors that effectively steer the embodied cameras to capture a faithful egocentric representation of the 3D world. To address this challenge, we introduce EgoGen, a new synthetic data generator that can produce accurate and rich ground-truth training data for egocentric perception tasks. At the heart of EgoGen is a novel human motion synthesis model that directly leverages egocentric visual inputs of a virtual human to sense the 3D environment. Combined with collision-avoiding motion primitives and a two-stage reinforcement learning approach, our motion synthesis model offers a closed-loop so", + "bbox": [ + 75, + 628, + 472, + 902 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "lution where the embodied perception and movement of the virtual human are seamlessly coupled. Compared to previous works, our model eliminates the need for a pre-defined global path, and is directly applicable to dynamic environments. Combined with our easy-to-use and scalable data generation pipeline, we demonstrate EgoGen's efficacy in three tasks: mapping and localization for head-mounted cameras, egocentric camera tracking, and human mesh recovery from egocentric views. EgoGen will be fully open-sourced, offering a practical solution for creating realistic egocentric training data and aiming to serve as a useful tool for egocentric computer vision research.", + "bbox": [ + 496, + 585, + 893, + 767 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 500, + 784, + 632, + 800 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The analysis of visual input from front-facing egocentric cameras is crucial for applications that benefit from a first-person perspective, mirroring the natural human experience [23, 43, 115]. AR devices, for instance, can utilize this viewpoint to amplify user immersion. Such cameras can also cater to individual preferences, providing custom", + "bbox": [ + 496, + 809, + 893, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "14497", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "visual assistance for those with impaired vision [19, 120].", + "bbox": [ + 76, + 90, + 460, + 107 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Despite its potential, egocentric perception faces challenges, primarily due to the scarcity of labeled data. Although datasets like Ego4D [23], ADT [61], EpicKitchen [13] and HoloAssist [102] exist, creating such datasets with rich and accurate annotations is costly and raises privacy concerns [115]. Alternatively, using graphics techniques to render synthetic multi-modal visual data has proven to be cost-effective and successful in training deep learning models, such as 3D human body estimation [8] and facial landmark detection [106].", + "bbox": [ + 75, + 108, + 467, + 258 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Creating egocentric synthetic data is challenging because egocentric cameras capture the complex interplay of body movements and the environment from the camera wearer's viewpoint. Modeling the intricate details and variations in human behavior presents a significant challenge.", + "bbox": [ + 75, + 261, + 467, + 335 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To tackle this problem, we introduce EgoGen, an egocentric synthetic data generation approach that simulates data from embodied sensors, i.e., front-facing cameras in head-mounted devices (HMD). While the ultimate goal is to simulate human behaviors that are indistinguishable from reality, in this work, we focus on creating virtual humans (i.e., camera wearers) that can explore and avoid obstacles in the 3D world that is not only complex and dynamic but could potentially include other moving virtual humans.", + "bbox": [ + 75, + 338, + 467, + 473 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Specifically, we propose a novel generative human motion model. Our key insight is that body movement and embodied perception should be seamlessly coupled. As William Gibson aptly stated, \"We see in order to move; we move in order to see,\" our egocentric perception is crucial for identifying obstacles, navigating in an environment, and planning actions. Our body movements are not solely a response to visual stimuli; they also change our egocentric perception. Therefore, the key idea of our motion model is to enable virtual humans to see their environment with egocentric visual inputs and respond accordingly by learning a policy to control a set of collision-avoiding motion primitives (CAMPs) that are composable for synthesizing long-term, diverse human motions. Due to the unbounded and high-dimensional latent action space of our generative motion primitive model, direct policy training through rendered egocentric images is often unstable [122]. Therefore, we propose a two-stage reinforcement learning scheme using an efficient egocentric visual proxy to couple egocentric visual cues and body movements seamlessly. In addition, we use an \"attention\" reward to incentivize egocentric perception behaviors, i.e., looking in the desired direction.", + "bbox": [ + 75, + 476, + 467, + 808 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Empirical results showcase the benefits of our egocentric perception-driven motion framework, which does not require a pre-calculated walking path in 3D scenes as in [29, 53, 119]. Instead, it empowers virtual humans to perceive the environment from their own viewpoint, enabling them to navigate, circumvent obstacles, and plan move", + "bbox": [ + 75, + 810, + 467, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "ments to reach the destination. Moreover, our model generalizes well to dynamic environments, even with training limited to static settings. By training virtual humans independently using CAMPs, our method synthesizes emergent multi-human behaviors without relying on multi-agent reinforcement learning algorithms. Egocentric visual cues are essential to build exploratory and generalizable motion models that unify navigational planning and movement control in complex and dynamic environments.", + "bbox": [ + 496, + 90, + 890, + 226 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Building upon CAMPs, we further create a scalable data generation pipeline for EgoGen that outfits virtual humans with clothing, automates cloth animation, and integrates 3D assets from various sources. We validate EgoGen's efficacy across three egocentric perception tasks. The high-quality synthetic data with precise ground truth annotations consistently improve the performance of state-of-the-art methods. In summary, the contributions of this work are:", + "bbox": [ + 496, + 227, + 890, + 347 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. We introduce EgoGen, a generative and scalable synthetic data generation approach specifically tailored for egocentric perception tasks.", + "2. We introduce novel motion primitives based on egocentric visual cues, enabling diverse and realistic human motion synthesis in 3D scenes. These primitives empower virtual humans to handle complex scenarios, such as dynamic environments and crowd motion without relying on multi-agent reinforcement learning.", + "3. EgoGen enables us to augment existing real-world egocentric datasets with synthetic images. Quantitative results demonstrate enhanced performance in state-of-the-art algorithms on mapping and localization for head-mounted cameras, egocentric camera tracking, and human mesh recovery from egocentric views." + ], + "bbox": [ + 500, + 348, + 890, + 573 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 588, + 640, + 603 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Human-Related Simulators and Synthetic Data. Previous works primarily focus on simulating robots [54, 69, 78, 92, 95] and autonomous cars [7, 16, 20, 74]. While some incorporated animated digital humans, like in [5, 20, 67], these efforts relied on pre-recorded motion sequences. Rendering images of people to train perception models has been widely studied such as [5, 17, 18, 43, 68, 82, 86]. In particular, [97] offers large-scale synthetic data for egocentric camera wearer pose estimation but relies on mocap data, lacking realistic and spontaneous interactions with the digital world. In contrast, EgoGen closes the gap in egocentric synthetic data generation for head-mounted devices. Please refer to Supp. Mat. for detailed comparisons.", + "bbox": [ + 496, + 613, + 890, + 809 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Human Motion Synthesis. Generating high-fidelity human motions and interactions with 3D scenes is widely studied in graphics [10, 31, 32, 42, 89, 90]. While they can generate high-quality motion, it's usually deterministic. Synthesizing physically plausible human motions has been extensively studied [12, 30, 64, 73, 96, 105]. However,", + "bbox": [ + 496, + 810, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "14498", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "they struggle with generalization to different body shapes. For example, [73] explicitly created 2048 humanoids to improve body shape generalizability. Time series models [65, 94, 114] synthesize the stochastic motions of diverse people well. However, in [94, 114], their generated motion sequences have limited lengths and human-scene interactions are not explicitly considered. Autoregressive methods [45, 72, 117, 119] can produce perpetual motions. In particular, [117] can generalize to diverse body shapes and synthesize long-term human motions.", + "bbox": [ + 75, + 90, + 472, + 241 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our egocentric perception-driven motion synthesis model is closely related to [117, 119], but distinguishes itself w.r.t.: (1) Enabling virtual humans to explore using egocentric visual cues, without predefined paths. (2) Synthesizing egocentric perception behaviors beyond locomotion, e.g., looking in certain directions. (3) Handling dynamic environments and multi-agent behavior without re-training.", + "bbox": [ + 75, + 242, + 472, + 348 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Mapping and Localization for AR. Localization and mapping from images is a long-standing problem known as: Photogrammetry [2, 26]; Structure-from-Motion (SfM) [22, 66, 79, 87, 104]; Simultaneous Localization and Mapping (SLAM) [14, 37, 56, 60]. Researchers have worked to make SLAM amenable for edge hand-held or head-mounted devices [6, 21, 37]. Cloud-based services like Google's Visual Positioning System [71], Niantic's Lightship [59], and Microsoft's Azure Spatial Anchors [34] have made visual localization and mapping more accessible. Benchmarking efforts have arisen for small-scale AR scenarios [36, 85], touristic landmarks [33, 80], and large-scale AR-device based localization [76, 77] to evaluate these systems.", + "bbox": [ + 75, + 348, + 472, + 542 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Egocentric Human Pose Estimation. Estimating 3D bodies from RGB images is widely studied from third-person views [9, 35, 38-41, 44, 63], and egocentric views [25, 46, 47, 57, 83, 97, 98, 100, 112, 113, 116], mostly requiring expensive real-world data paired with ground truth annotations. Besides RGB images, depth images offer explicit 3D information, mitigating scale and shape ambiguity, with the potential to enable broader AR/VR applications. However, depth-based methods, especially for the egocentric view, are underexplored. Most existing works [28, 49, 55, 70, 84, 101, 107, 111] predict 3D body skeletons without expressive body meshes, struggling with challenges like severe body truncations and scene occlusions typical in egocentric views. Such limited attention mainly stems from the scarcity of data, as obtaining high-quality human mesh annotations for real-world depth images is labor-intensive.", + "bbox": [ + 75, + 544, + 472, + 801 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Ego-Sensing Driven Motion Synthesis", + "text_level": 1, + "bbox": [ + 76, + 814, + 416, + 832 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To close the loop for the interdependence between egocentric synthetic image data and human motion synthesis, we use deep reinforcement learning (RL), integrating egocentric vision cues to synthesize human motions as described in", + "bbox": [ + 75, + 839, + 470, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Sec. 3.1 and 3.2. Subsequently, we extend learned policies to generate emergent multi-agent behaviors, as in Sec. 3.3.", + "bbox": [ + 498, + 90, + 890, + 122 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Ego-Sensing Driven Motion Primitives", + "text_level": 1, + "bbox": [ + 498, + 133, + 831, + 148 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Generating realistic egocentric data requires diverse and lifelike human motion synthesis. In this work, we consider arguably the most common everyday behaviors: navigating towards goals with egocentric perception while avoiding collisions with obstacles and people in dynamic 3D scenes.", + "bbox": [ + 496, + 157, + 890, + 232 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Overview. Following recent literature [45, 110, 117, 119], we employ deep RL to train control policies on learned latent spaces that characterize natural human motions. However, unlike these previous works that only consider simple static scenes, we leverage egocentric perception and propose collision-avoiding motion primitives (CAMPs) to enable virtual humans to self-exlore and navigate in a dynamic environment. Specifically, CAMPs are trained jointly to produce collision-free motion sequences. At each timestep $t$ , the agent observes the state $\\mathbf{s}_t$ , performs an action $\\mathbf{a}_t$ , and receives a reward $r_t = r(\\mathbf{s}_t, \\mathbf{a}_t, \\mathbf{s}_{t+1})$ , where $\\mathbf{s}_{t+1}$ represents the next state of the environment due to $\\mathbf{a}_t$ .", + "bbox": [ + 496, + 233, + 890, + 415 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Egocentric Sensing As Depth Proxy. We aim to sample actions given by a policy to synthesize realistic human motions. Egocentric perception-driven motion synthesis should arguably use egocentric vision as input. However, depth rendering is costly and RL requires billions of samples to converge [3, 103]. Besides, directly training RL with visual data can be unstable [122]. We thereby use a cheap-to-compute egocentric sensing $\\mathcal{E}_t$ as a proxy for depth images as illustrated in Fig. 2. $N$ rays are cast evenly from the midpoint of two eyeballs, i.e., the location of the egocentric camera. The field of view $[\\theta_{min},\\theta_{max}]$ , centered on the 2D projection of the viewing direction $\\vec{\\nu}$ , limits the agent's perception to the front area. Rays stop at collisions, with collision detection in 2D. See more details in Supp. Mat.", + "bbox": [ + 496, + 416, + 890, + 627 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Agent Representation. The agent is a virtual human represented by an SMPL-X mesh [63]. We further compact the body representation by selecting $M = 67$ body surface markers $\\mathbf{x} \\in \\mathbb{R}^{M \\times 3}$ on the mesh following [118].", + "bbox": [ + 496, + 628, + 890, + 688 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Environment. Inspired by [108], we aim to learn a library of composable CAMPs. We implement a finite-horizon environment based on the generative motion primitive model from [117]. Specifically, a motion primitive is defined as a 0.5-second motion clip containing $T = 20$ frames in the canonical coordinate, and each frame contains a single agent representation. The primitive model $\\mathcal{P}$ is based on the C-VAE framework [88], which takes the first $T_{s} = 2$ frames as the condition, and models a conditional probability of the next $T - T_{s}$ frames. Compared to [117] trained on the AMASS dataset [48] with many sport motion sequences, we train $\\mathcal{P}$ using the SAMP dataset [29], which focuses on daily activities, better suited for HMD use cases. Our action space $\\mathcal{A}$ is the pretrained 128D latent space of $\\mathcal{P}$ , and the", + "bbox": [ + 496, + 688, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "14499", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/13ff6411c240f09adaa375b54aefc78e76a61979c69f8352c6ed6db654396e22.jpg", + "image_caption": [ + "Figure 2. Policy network architecture. We learn a generalizable mapping from motion seed body markers $\\mathbf{X}_t^S$ , marker directions $\\mathbf{X}_t^{S^D}$ , egocentric sensing $\\mathcal{E}_t$ , and distance to the target $d_t$ to CAMPs. The policy learns a stochastic collision avoiding action space to predict future body markers $\\mathbf{X}_t^F$ . For illustration purposes, we visualize only one frame of $\\mathbf{X}_t^S$ and $\\mathcal{E}_t$ . See Sec. 3.1 and 3.2 for details." + ], + "image_footnote": [], + "bbox": [ + 76, + 85, + 893, + 310 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "action $\\mathbf{a}_t$ can be randomly sampled from $\\mathcal{A}$ .", + "bbox": [ + 75, + 382, + 372, + 397 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "With the input of a random action $\\mathbf{a}_t\\in \\mathbb{R}^{128}$ and a motion seed $\\mathbf{X}_t^S = [\\mathbf{x}_t^0,\\mathbf{x}_t^{T_s - 1}]$ (history frames), $\\mathcal{P}$ predicts future frames $\\mathbf{X}_t^F = [\\mathbf{x}_t^{T_s},\\dots ,\\mathbf{x}_t^{T - 1}]$ of the current motion primitive $\\mathbf{X}_t = [\\mathbf{X}_t^S,\\mathbf{X}_t^F ]\\in \\mathbb{R}^{T\\times M\\times 3}$ , which represents a short sequence of human motion spanning $0.5\\mathrm{s}$ :", + "bbox": [ + 75, + 397, + 470, + 474 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {X} _ {t} ^ {F} = \\mathcal {P} (\\mathbf {X} _ {t} ^ {S}, \\mathbf {a} _ {t})\n$$\n", + "text_format": "latex", + "bbox": [ + 210, + 478, + 336, + 498 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "State. To preserve Markov property [62], the state is defined as $\\mathbf{s}_t = \\{\\mathbf{X}_t^S, \\mathbf{X}_t^{S^D}, \\mathcal{E}_t, d_t, \\tau_t\\}$ , in which $\\mathbf{X}_t^{S^D} \\in \\mathbb{R}^{T_s \\times M \\times 3}$ denotes the normalized direction of each marker seed to the target, $\\mathcal{E}_t \\in \\mathbb{R}^{T_s \\times N}$ denotes the egocentric sensing depth proxy, $d_t$ denotes the distance from the pelvis to the target, and $\\tau_t$ denotes the remaining time. See Fig. 2.", + "bbox": [ + 75, + 503, + 468, + 595 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Reward. To synthesize egocentric perception behaviors, we use an \"attention reward\" to incentivize the virtual human to look in specific directions: $r_{\\text{attention}} = \\cos \\langle \\vec{\\mathbf{v}}, \\vec{\\mathbf{a}} \\rangle$ , where $\\vec{\\mathbf{a}}$ is the attention direction from the head joint to the viewing target. The reward function is defined as:", + "bbox": [ + 75, + 595, + 468, + 672 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nr _ {t} = r _ {c o n t.} + r _ {d i s t} + r _ {o r i} + r _ {a t t e n t i o n} + r _ {p e n e} + r _ {p o s e} + r _ {s u c c},\n$$\n", + "text_format": "latex", + "bbox": [ + 75, + 680, + 468, + 696 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $r_{cont}$ enforces valid foot contact and minimizes foot skating; $r_{dist}$ encourages reaching the target; $r_{ori}$ aligns the body forward direction with the target; $r_{pene}$ guides collision avoidance; $r_{pose}$ reduces unrealistic human poses; and $r_{succ}$ is a sparse reward when reaching the target.", + "bbox": [ + 75, + 700, + 468, + 777 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Episode Termination. To handle collisions beyond [119], we employ multiple termination signals to conclude an episode if the generated motion primitive $\\mathbf{X}_t$ satisfies any of the following conditions:", + "bbox": [ + 75, + 777, + 468, + 837 + ], + "page_idx": 3 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Success: The virtual human reached the target.", + "- Penetration: The virtual human collides with the obstacle.", + "- Timeout: The virtual human did not reach the target within the maximum timesteps." + ], + "bbox": [ + 76, + 837, + 468, + 898 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Training Collision-Avoiding Stochastic Policies", + "text_level": 1, + "bbox": [ + 498, + 381, + 890, + 398 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Algorithm. We use Proximal Policy Optimization [81] (PPO) to learn a generalizable mapping from various egocentric sensing and body configurations to CAMPs. Instead of extensive manual data collection for all possible input combinations, we leverage the stochastic nature of the PPO policy. Through exploration and sampling actions, the agent traverses the scene and generates varied egocentric sensing and body configurations, diversifying the training data.", + "bbox": [ + 496, + 405, + 890, + 526 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Instead of training each CAMP independently for every single step, we use PPO to train a sequence of CAMPs jointly in multi-step collision avoidance tasks. This approach can benefit choosing a more favorable CAMP which makes the subsequent action easier.", + "bbox": [ + 496, + 526, + 890, + 601 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Network. The network architecture is shown in Fig 2. The actor and critic network share a feature extraction trunk to encode the state $\\mathbf{s}_t$ : the motion seed $(\\mathbf{X}_t^S$ and $\\mathbf{X}_t^{S^D}$ ) and the egocentric sensing $\\mathcal{E}_t$ are encoded using RNNs; the rest of scalar states are encoded using positional encoding [99]. The actor predicts a stochastic policy $\\mathbf{a}_t \\sim \\pi(\\mathbf{z}_t|\\mathbf{s}_t)$ conditioned on the current state $\\mathbf{s}_t$ , where $\\mathbf{z}_t \\sim \\mathcal{N}(\\mu, \\Sigma)$ . $\\mu$ and $\\Sigma$ are the mean and variance of the learned action space.", + "bbox": [ + 496, + 602, + 890, + 723 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Objective Function. The objective function includes the policy surrogate $L^{CLIP}$ , the value function error term $L^{VF}$ to evaluate the value prediction $V_{\\theta}$ , and an entropy bonus $L^{S}$ to encourage exploration:", + "bbox": [ + 496, + 723, + 890, + 784 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nL = L ^ {C L I P} + c _ {1} L ^ {V F} + c _ {2} L ^ {S}\n$$\n", + "text_format": "latex", + "bbox": [ + 591, + 794, + 799, + 811 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $c_{1}, c_{2}$ are coefficients. See more details in Supp. Mat. RL Pre-training and Finetuning. Training in crowded scenes, e.g. Replica [91], requires additional steps. Because the action space $\\mathcal{A}$ is an unbounded Gaussian, RL exploration while predicting reasonable human poses can", + "bbox": [ + 496, + 825, + 890, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "14500", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 3 + }, + { + "type": "code", + "sub_type": "algorithm", + "code_caption": [ + "Algorithm 1 Crowd motion synthesis with learned CAMPs" + ], + "code_body": "Result: Multi-human locomotion w/ collision avoidance; \nInit: crowd size $C$ , marker seed for each human $\\mathbf{X}_c^S$ ; \nfor step $\\leftarrow 1$ to max_steps do env. finite horizon \nfor $c \\gets 1$ to $C$ do for each human update all locations with $\\{bbox(\\mathbf{X}_c^S)\\}_{c=1:C}$ compute egocentric sensing $\\mathcal{E}_c$ ; execute the action that maximizes the expected return, and produce one CAMP; end for \nend for", + "bbox": [ + 93, + 109, + 468, + 260 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "be challenging. We first pretrain the policy with a higher $r_{pene}$ weight without penetration termination. After convergence, we finetune it with strict termination constraints using a signed distance field (SDF) for penetration detection. Please refer to Supp. Mat. for the formulation and weighting of each reward and training detail.", + "bbox": [ + 75, + 294, + 468, + 385 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Compositing Learned Motion Primitives", + "text_level": 1, + "bbox": [ + 76, + 402, + 423, + 417 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Although CAMPs are trained solely with static scenes, their direct application to dynamic settings is achieved by decomposing jointly trained CAMPs into individual motion primitives and re-compositing them. Our model demonstrates effective generalization by selecting the next best motion primitive from the learned CAMP library to maximize the expected return, provided that the egocentric sensing is updated with the most recent obstacle location at each timestep. Furthermore, our model is directly applicable to tasks involving complex interactions with other virtual humans. To synthesize crowd motion (Alg. 1), each virtual human employs the same policy to navigate and avoid others. To a specific virtual human, others are seen as dynamic obstacles, represented by body bounding boxes for avoidance. Acknowledging the inherent delay in human reactions when avoiding dynamic obstacles [4], agents take a single CAMP sequentially, instead of in parallel, i.e. the first agent generates its first CAMP and waits for others to complete their first CAMP before all agents move on to prepare their second CAMP. To ensure successful collision avoidance, the agent's egocentric sensing is updated before taking a new action. This composition of CAMPs synthesizes emergent multi-human behaviors without multi-agent RL algorithms (see Sec. 5.1), enhancing the generalization and scalability.", + "bbox": [ + 75, + 428, + 468, + 790 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Egocentric Synthetic Data Generation", + "text_level": 1, + "bbox": [ + 76, + 811, + 419, + 829 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Synthesizing realistic egocentric perception-driven human motions (as detailed in Sec. 3) forms the foundation of simulating egocentric synthetic data. An overview of our egocentric data generation pipeline EgoGen, is shown in Fig. 3.", + "bbox": [ + 75, + 839, + 468, + 901 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/06018b31962397ed8eb422593026e3d987b0a8165b470871706a4442b584c09e.jpg", + "image_caption": [ + "Figure 3. Overview of EgoGen. Through generative motion synthesis (Sec. 3), we further enhance egocentric data diversity by randomly sampling diverse body textures (ethnicity, gender) and 3D textured clothing through an automated clothing simulation pipeline (Sec. 4.2). With high-quality scenes and different egocentric cameras, we can render photorealistic egocentric synthetic data with rich and accurate ground truth annotations (Sec. 4.3)." + ], + "image_footnote": [], + "bbox": [ + 501, + 88, + 890, + 325 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Embodied Camera Placement", + "text_level": 1, + "bbox": [ + 500, + 454, + 766, + 468 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Similar to existing AR devices, we use the head pose to define the egocentric viewing direction $\\vec{\\mathbf{v}}$ . Our development is based on Blender [11]. We use the SMPL-X [63] mesh to position the egocentric camera between the two eyeballs. The camera's viewing direction $(\\vec{\\mathbf{v}})$ is perpendicular to the plane determined by the two eye bones in the armature. We also support simulating multi-camera rigs as shown in Fig. 1. When the body moves (Sec. 3.3), we can synthesize egocentric videos with continuously updated camera poses.", + "bbox": [ + 496, + 479, + 890, + 614 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Body Texture and Clothing", + "text_level": 1, + "bbox": [ + 500, + 631, + 745, + 647 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To enhance EgoGen's synthetic data realism, we dress virtual humans using human textures and 3D clothing assets from BEDLAM [1, 8], including 50 male and 50 female skin albedo textures from seven ethnic groups.", + "bbox": [ + 496, + 656, + 890, + 717 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Unlike prior works [8, 109] relying on unscalable commercial software for clothing dynamics simulation, we automate it for diverse synthesized motions and body shapes, minimizing manual effort. Each garment mesh is in a consistent rest pose, i.e., A-Pose (See Fig. 3 middle-left). For each motion sequence, we first repose it to match the body pose in the first frame using linear blend skinning. This involves initializing the clothing geometry by sampling pose and shape blend shapes, along with skinning weights from the nearest multiple SMPL-X vertices in A-Pose. Then we simulate upper and lower garments separately using a state-of-the-art clothing simulation network [24].", + "bbox": [ + 496, + 719, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "14501", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/cd9e91a4dadcea46de16ec657329bf2a75b5be609615d8ad07ae3aa8897f4b31.jpg", + "table_caption": [ + "Table 1. Evaluation of motion synthesis in scenes with moving obstacles, multiple humans, and path diversity. $\\downarrow$ : lower is better; $\\uparrow$ : higher is better. The best results in each scenario are in boldface. * denotes an improved version for fair comparison. (Sec. 5.1)" + ], + "table_footnote": [], + "table_body": "
EvaluationMetricsGAMMA* [117]DIMOS* [119]Ours
Mov. obs.SR (%) ↑9683100
Dist. (m) ↓0.290.550.06
Cont. ↑0.950.960.97
Pene-S. (%) ↓9.28.43.4
2 humansSR (%) ↑9588100
Dist. (m) ↓0.320.410.07
Cont. ↑0.960.980.97
Pene-H. ↓27.610.70
4 humansSR (%) ↑9270100
Dist. (m) ↓0.410.790.07
Cont. ↑0.940.950.96
Pene-H. ↓60.441.70
DiversitySR (%) ↑968497
Std Dev ↑0.9871.051.21
", + "bbox": [ + 89, + 155, + 457, + 325 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Rendering and Annotations", + "text_level": 1, + "bbox": [ + 76, + 338, + 326, + 353 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "EgoGen supports simulating diverse head-mounted devices with different camera models, such as fisheye and pinhole cameras. Given the camera's intrinsic parameters and relative poses within the camera rig, we can simulate AR devices like Project Aria glasses [50] and HoloLens [51], facilitating synthetic data generation for real-world applications. Camera extrinsic is determined by our generative human motion model. We use Blender [11] to render photorealistic egocentric image sequences with motion blur. We also render out a rich set of ground truth annotations, such as depth maps, surface normals, segmentation masks, world positions, optical flow, etc for egocentric perception tasks.", + "bbox": [ + 75, + 361, + 468, + 542 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 76, + 556, + 209, + 571 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We assess the motion quality, generalizability, and diversity of our motion model, highlighting its ability to generalize to unseen complex tasks and comparing it with recent baselines (Sec. 5.1). We evaluate our proposed egocentric sensing as a depth proxy for enhancing agent exploration (Sec. 5.2) and conduct ablation studies (Sec. 5.3).", + "bbox": [ + 75, + 580, + 468, + 671 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We further demonstrate the effectiveness of EgoGen on three egocentric computer vision tasks in Sec. 5.4, and 5.5. By incorporating synthesized egocentric images, we can enhance the performance of the state-of-the-art algorithms.", + "bbox": [ + 75, + 671, + 468, + 733 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1. Evaluation of Learned CAMPs", + "text_level": 1, + "bbox": [ + 76, + 741, + 351, + 756 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We assess CAMPs' generalizability in dynamic scenes, including scenes with moving obstacles and scenes with multiple individuals. In tests with moving obstacles, the obstacle blocks the person's path by moving between the person and the goal. In multiple human test scenes, lines from their starting and goal locations intersect in the middle, requiring solving human-human penetrations. See detail in Sup. Mat.", + "bbox": [ + 75, + 763, + 468, + 869 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In Tab. 1, we compare goal-reaching behaviors with two recent baselines: GAMMA [117] and DIMOS [119]. Base", + "bbox": [ + 76, + 869, + 468, + 900 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/3f089a401ed28703c01383da3789c575ae76917eafb30193c6b53e2c8b2839f1.jpg", + "table_caption": [ + "Table 2. Evaluation of egocentric sensing. (Sec. 5.2)" + ], + "table_footnote": [], + "table_body": "
Method (sensing range)SR (%) ↑Dist. (m) ↓
Local map [119] (0.8 m)780.35
Local map* (7 m)43.04
Egocentric sensing (ours) (7 m)950.12
", + "bbox": [ + 506, + 113, + 875, + 178 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "line methods use navigation meshes and path planning for static scenes, while CAMPs can autonomously avoid dynamic obstacles (Sec. 3.3). For fair comparison in dynamic scenes, we extend the baselines by updating navigation meshes and performing on-the-fly path planning at each time step. The tree-based search as in [117] is disabled for all the methods. Metrics: (1) $SR$ : Success rate for reaching the goal location within a $0.3\\mathrm{m}$ threshold. (2) Dist.: Average distance of the final pelvis location to the goal. (3) Cont.: The contact metric [117] that measures foot-floor contact and foot skating. (4) Pene-S.: Percentage of frames with detected human-scene penetration in moving obstacle scenes. (5) Pene-H.: Accurate human-human penetration evaluation metric using COAP [52] in multiple human scenes. Please refer to Supp. Mat. for metric details.", + "bbox": [ + 496, + 198, + 890, + 425 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "CAMPs outperform the two baselines in dynamic scenarios with moving obstacles and multiple humans, exhibiting lower human-scene and human-human penetrations and a higher goal-reaching success rate. In multiple human scenarios, we observe that in the baselines, dynamically redoing path planning for each human independently can not effectively solve human-human penetration. In contrast, com-posable CAMPs can generalize well in dynamic settings without using multi-agent RL to synthesize crowd motions.", + "bbox": [ + 496, + 428, + 890, + 565 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We assess walking path diversity using the standard deviation of pelvis locations for the same start-target pairs in scenes with a single static box obstacle. As shown in Tab. 1 (Diversity), our approach does not require a pre-computed global path and allows agents to self-exlore without being constrained by predefined paths, achieving higher walking path diversity and success rate. This fosters diverse synthetic data generation via more diverse synthesized motion.", + "bbox": [ + 496, + 568, + 890, + 689 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.2. Evaluation of Egocentric Sensing", + "text_level": 1, + "bbox": [ + 500, + 708, + 790, + 724 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We assess the exploration ability of our egocentric sensing $\\mathcal{E}_t$ in Replica [91] scenes. In Tab. 2, we replace $\\mathcal{E}_t$ with a local map [119] in our state $\\mathbf{s}_t$ , following their encoding method. Relying on local information can trap agents in local optima, e.g., walls beyond their sensing range, resulting in lower SR. Our egocentric sensing acts as a depth proxy, allowing the agent to avoid local optima, explore more effectively than local maps [119] or scandots [3], and achieve higher SR. In addition, our compact representation is more scalable as the sensing range increases, while quadratic local map growth can hinder the policy network's learning.", + "bbox": [ + 496, + 734, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "14502", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/f9ccde6c7f475f2c74da82ae063fbb22ab61c88e988f644dcc5ff7c39f5c6ffe.jpg", + "table_caption": [ + "Table 3. Ablation studies. Note: in our observation, $\\| VP\\| _2 > 15$ indicates abnormal human poses. (Sec. 5.3)" + ], + "table_footnote": [], + "table_body": "
SR (%)↑||VP||2↓cos(▼,▲)↑
Egocentric depth813.640.049
No pretraining9028.770.918
No attention reward9012.260.891
Our policy9210.570.940
", + "bbox": [ + 84, + 127, + 452, + 204 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.3. Ablation Studies", + "text_level": 1, + "bbox": [ + 76, + 219, + 240, + 234 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We compare our policy with several ablations in Tab. 3:", + "bbox": [ + 76, + 243, + 444, + 257 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Egocentric depth: an ablation training an egocentric depth image-based policy without the depth sensing proxy. Egocentric depth images are encoded with a CNN;", + "bbox": [ + 75, + 258, + 468, + 303 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "No pretraining: an ablation training collision avoidance in crowded scenes with strict penetration termination directly; No attention reward: an ablation for the viewing direction.", + "bbox": [ + 76, + 304, + 468, + 349 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We assess pose naturalness with the maximum pose embedding norm encoded with VPoser [63] and evaluate the attention reward with the cosine similarity between the viewing direction $\\vec{\\mathbf{v}}$ and the attention direction $\\vec{\\mathbf{a}}$ (Sec. 3.1).", + "bbox": [ + 75, + 349, + 468, + 410 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Directly training RL with egocentric depth images is ineffective due to our high-dimensional action space, emphasizing the value of the compact egocentric sensing representation. Training agents with strict penetration constraints in crowded scenes directly can result in exploring unreasonable action subspaces, given its unbounded Gaussian nature, leading to unrealistic human poses, highlighting the effectiveness of our two-stage RL training scheme. Without the attention reward, the virtual human's capability to attend to a specific direction decreases. All ablation studies are evaluated in Replica. See visuals in Supp. Vid. and Supp. Mat.", + "bbox": [ + 75, + 411, + 468, + 578 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.4. Mapping, Localization, and Tracking for HMD", + "text_level": 1, + "bbox": [ + 76, + 588, + 468, + 604 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Mapping and localization. LaMAR [76] is the first mapping and localization benchmark dataset for AR in large-scale scenes. Despite over a year of extensive data collection, the dataset still lacks exhaustive scene coverage, especially in large open spaces. EgoGen can let virtual humans explore large-scale scenes, render dense egocentric views, and build a more complete SfM map by extracting image feature points with SuperPoint [15] and matching images with SuperGlue [75]. Despite synthetic images being noisier due to scene quality, SuperGlue [75] matching can filter out noisy feature points and yield reliable matches.", + "bbox": [ + 75, + 612, + 468, + 777 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In Tab. 4, we evaluate EgoGen by assessing the localization recall at $(1^{\\circ}, 10cm)$ on the validation set in a lobby of $\\sim 120$ sqm of the LaMAR CAB location. In addition, we report the number of triangulated 3D points (#P3D) and track length. EgoGen improves the 3D reconstruction by yielding more points for a slightly improved track length and also a significantly better localization performance compared to using the real data only. Ng et al. [58] augments mapping", + "bbox": [ + 75, + 780, + 468, + 901 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/2156192db9bd92b1b053820fcaf8d5a26c0c2109a9791b25adbeec780be69440.jpg", + "table_caption": [ + "Table 4. Mapping and localization evaluation. We augment LaMAR with the same amount of images (248 frames) and report the localization recall at $(1^{\\circ},10cm)$ on the validation set. EgoGen achieves the highest track length and recall. (Sec. 5.4)" + ], + "table_footnote": [], + "table_body": "
#P3D ↑Track length ↑Recall (%) ↑
LaMAR19297395.194666.9
Ng et al. [58]19377585.194074.9
EgoGen19361695.210576.7
", + "bbox": [ + 508, + 155, + 875, + 219 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/f3e7593e91b2cd328a8b31d399b1f798261a010da8696f8b0a762311fd59aa07.jpg", + "table_caption": [ + "Table 5. Egocentric camera tracking evaluation of models trained with and without synthetic data from EgoGen. (Sec. 5.4)" + ], + "table_footnote": [], + "table_body": "
Pose ↓Rotation ↓Transl (mm) ↓
Scratch1.830.741303
+ EgoGen pretrain1.670.621305
", + "bbox": [ + 508, + 265, + 875, + 314 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "images by perturbing real-world camera poses with noise, which may generate unrealistic camera poses (e.g., stuck in a wall or facing the ceiling), limiting egocentric localization effectiveness. Their method also assumes the availability of initial camera poses, which may not always be feasible. In contrast, EgoGen augments by virtual humans randomly exploring scenes. Our approach holds promise for creating AR mapping and localization datasets for digital twin scenes without manual data collection, providing enhanced privacy preservation, e.g. no need for anonymization. Refer to Supp. Mat. for visualization and implementation details.", + "bbox": [ + 496, + 325, + 890, + 492 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Egocentric camera tracking. Egocentric camera tracking for HMD aims to yield device pose trajectories in 3D scenes given egocentric video observations. Recovering camera poses from monocular RGB videos using SLAM [93] is a challenging and ill-posed problem due to scale ambiguity. EgoEgo [43] leverages the knowledge of human motion to address the egocentric HMD tracking problem. Specifically, EgoEgo trains a neural network to infer the translation scaling and rotations from egocentric videos, which improves the HMD tracking performance. However, training this model requires jointly captured data of ground truth HMD trajectories and egocentric videos, which are costly to collect. We address this limitation by using EgoGen to synthesize quantities of egocentric videos with accurate camera trajectories to pretrain the model, which proves to improve the tracking performance on real data. We conduct experiments on the GIMO [121] dataset that contains $\\sim 200$ short sequences of paired motion-video data in 19 scenes. Using EgoGen, we synthesize $\\sim 4k$ sequences of human movements in their scenes and render corresponding egocentric videos using the same camera intrinsic as GIMO and the embodied camera placement described in Sec. 4.1. We also slightly perturb the camera placement location and orientation to simulate the diversity of how people wear HMDs in real data and avoid overfitting to one specific camera placement. We first pretrain the model with synthetic data generated by EgoGen, then finetune it on the real GIMO data.", + "bbox": [ + 496, + 493, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "14503", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Tab. 5 shows the egocentric camera tracking performances for models trained with and without synthetic data. Definitions of evaluation metrics can be found in Supp. Mat. The finetuned model benefits from EgoGen synthetic data and predicts more accurate camera poses compared to the model trained using real data only.", + "bbox": [ + 75, + 90, + 470, + 181 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.5. Human Mesh Recovery from Egocentric views", + "text_level": 1, + "bbox": [ + 76, + 193, + 468, + 209 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Human mesh recovery (HMR) is the key to human behavior understanding from the egocentric view, thus crucial for applications in robotics and AR/VR. Given an egocentric RGB or depth image of a target subject, HMR aims to reconstruct the subject's 3D body pose and shape. However, acquiring and annotating real-world data is expensive, demanding, and time-consuming, with egocentric data being particularly scarce. EgoBody [115] is a recent egocentric dataset capturing two-people interactions, with egocentric depth/RGB frames annotated with SMPL-X body meshes. EgoBody provides $\\sim 180\\mathrm{k}$ egocentric RGB frames, and merely $\\sim 23\\mathrm{k}$ depth frames due to the low frame rate of the depth sensor, with $\\sim 90\\mathrm{k} / \\sim 10\\mathrm{k}$ in the RGB/depth training set. Such limited data is insufficient to train a learning-based model from scratch. In contrast, with EgoGen, largescale synthetic egocentric data can be generated in a time-efficient way. We leverage EgoGen to generate quantities of training frames (300k RGB, 105k depth) of humans moving in EgoBody 3D scenes, rendered from the egocentric view, and annotated by SMPL-X parameters of the target subject. Specifically, RGB images are rendered with lifelike human body textures and 3D clothing, with random lighting.", + "bbox": [ + 75, + 217, + 468, + 549 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "With the recent HMR regressor, ProHMR [41], we show that pre-training with our synthetic data from EgoGen enhances the existing method's capability to generalize on real-world scenarios. Evaluated on the real-world EgoBody test set, we compare two training schemes: (1) trained from scratch on the real-world EgoBody training set (\"-scratch\"), and (2) pre-trained on synthetic data from EgoGen and finetuned on the real-world EgoBody training set (\"-ft\").", + "bbox": [ + 75, + 551, + 468, + 672 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "HMR from depth. As no existing methods were proposed for depth-based HMR task, we adapt ProHMR [41] to the depth input by changing the channel number of the first convolution layer. To mimic real-world sensor noise, synthetic noise [27] is added to the rendered depth. G-MPJPE is additionally reported for depth-based HMR as depth images provide global information. As shown in Tab. 6, compared to the model trained only with a limited amount of real-world data (Depth-scratch), errors are significantly reduced for the model pre-trained with our large-scale synthetic data (Depth-ft), in terms of global translation (22.9% lower G-MPJPE), local pose (20.7% lower MJPE), and body shape (19.5% lower V2V).", + "bbox": [ + 75, + 672, + 468, + 867 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "HMR from RGB. For training with RGB images, we apply various data augmentation techniques similar to [8].", + "bbox": [ + 76, + 869, + 468, + 900 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/5a0cccada08b1fdf35aeb231140523d5c0db7c8b38e2f38ef203b0c0c29eefe4.jpg", + "table_caption": [ + "Table 6. Evaluation of HMR on EgoBody test set. “*-scratch” denotes the model trained from scratch with the Egobody training set, and “*-ft” denotes the model pre-trained with EgoGen synthetic data. The units for all metrics are in mm. (Sec. 5.5)" + ], + "table_footnote": [], + "table_body": "
G-MPJPE ↓MPJPE ↓PA-MPJPE ↓V2V ↓
Depth-scratch117.782.254.1100.6
Depth-ft90.765.247.381.0
RGB-scratch-90.759.9102.1
RGB-ft-85.356.297.2
", + "bbox": [ + 506, + 148, + 875, + 217 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Tab. 6 indicates that the RGB-based model pre-trained with large-scale synthetic data (\"RGB-ft\") also outperforms the model trained only on real-world data (\"RGB-scratch\"), for both body pose and shape accuracy.", + "bbox": [ + 496, + 227, + 890, + 287 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The enhanced performance highlights that EgoGen's synthetic data effectively compensates for the lack of real-world training data, boosting the performance of current methods when test on real-world data. We will release both of our synthetic EgoBody datasets. See dataset statistics, qualitative visualizations, and training details in Supp. Mat.", + "bbox": [ + 496, + 287, + 890, + 378 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion and Future Work", + "text_level": 1, + "bbox": [ + 500, + 392, + 767, + 407 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We propose a novel egocentric synthetic data generation approach, EgoGen, that uses embodied sensors, a parametric body model, and a generative egocentric perception-driven human motion synthesis method to create egocentric training data with accurate and rich ground truth annotations. By integrating deep reinforcement learning and collision-avoiding motion primitives with egocentric depth proxy, EgoGen synthesizes robust human motion and emergent multi-agent behaviors. This paves the way to an efficient and scalable data generation solution that may have a profound impact on egocentric perception tasks.", + "bbox": [ + 496, + 417, + 890, + 583 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Human-scene interaction in EgoGen is currently coarse. We aim to extend the current method to simulate more detailed human motion driven by egocentric perception, such as hand manipulation, sitting, lying, etc, to facilitate more realistic egocentric synthetic data. We use fixed attention goals to model human attention. Predicting human intention through historical egocentric perception and synthesizing viewing directions based on predicted intention holds significant potential. Synthesizing gaze direction for predicting human intent is valuable but presently hampered by data requirements; we will revisit this when resources allow.", + "bbox": [ + 496, + 584, + 890, + 748 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We will explore many other egocentric vision tasks with EgoGen as this area grows rapidly such as social understanding and forecasting. EgoGen could benefit humanrobot interaction, e.g., our generative human motion model and lifelike human appearances can be integrated into [67] to close the sim2real gap for robotic agents further.", + "bbox": [ + 496, + 750, + 890, + 839 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements. This work is supported by the SNSF project grant 200021 204840 and the SDSC PhD fellowship. We sincerely thank Linfei Pan, Yunke Ao, Qianli Ma, and Maxime Raafat for the valuable discussions.", + "bbox": [ + 496, + 839, + 890, + 898 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "14504", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Meshcapade GmbH, Tübingen, Germany. https://meshcapade.com, 2022.5", + "[2] Mohammed Abdel-Wahab, Konrad Wenzel, and Dieter Fritsch. Efficient reconstruction of large unordered image datasets for high accuracy photogrammetric applications. In ISPRS Annals of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Melbourne, Australia. XXII ISPRS Congress, 2012. 3", + "[3] Ananye Agarwal, Ashish Kumar, Jitendra Malik, and Deepak Pathak. Legged locomotion in challenging terrains using egocentric vision. In Conference on Robot Learning, CoRL 2022, 14-18 December 2022, Auckland, New Zealand, pages 403-415. PMLR, 2022. 3, 6", + "[4] M. Pilar Aivar, Eli Brenner, and Jeroen B. J. Smeets. Avoiding moving obstacles. Experimental Brain Research, 190 (3):251-264, 2008. 5", + "[5] Hiroyasu Akada, Jian Wang, Soshi Shimada, Masaki Takahashi, Christian Theobalt, and Vladislav Golyanik. Unrealego: A new dataset for robust egocentric 3d human motion capture. In European Conference on Computer Vision (ECCV), 2022. 2", + "[6] Apple. ARKit. https://developer.apple.com/arkit/, 2017.3", + "[7] Sonia Baltodano, Srinath Sibi, Nikolas Martelaro, Nikhil Gowda, and Wendy Ju. The rroads platform: a real road autonomous driving simulator. In Proceedings of the 7th International Conference on Automotive User Interfaces and Interactive Vehicular Applications, pages 281-288, 2015. 2", + "[8] Michael J. Black, Priyanka Patel, Joachim Tesch, and Jinlong Yang. BEDLAM: A synthetic dataset of bodies exhibiting detailed lifelike animated motion. In Proceedings IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR), 2023. 2, 5, 8", + "[9] Federica Bogo, Angjoo Kanazawa, Christoph Lassner, Peter Gehler, Javier Romero, and Michael J. Black. Keep it SMPL: Automatic estimation of 3D human pose and shape from a single image. In European Conference on Computer Vision, pages 561-578, 2016. 3", + "[10] Simon Clavet. Motion matching and the road to next-gen animation. In Proc. of GDC, 2016. 2", + "[11] Blender Online Community. Blender - a 3D modelling and rendering package. Blender Foundation, Stichting Blender Foundation, Amsterdam, 2018. 5, 6", + "[12] Erwin Coumans and Yunfei Bai. Pybullet, a python module for physics simulation for games, robotics and machine learning. http://pybullet.org, 2016-2021. 2", + "[13] Dima Damen, Hazel Doughty, Giovanni Maria Farinella, Sanja Fidler, Antonino Furnari, Evangelos Kazakos, Davide Moltisanti, Jonathan Munro, Toby Perrett, Will Price, and Michael Wray. The epic-kitchens dataset: Collection, challenges and baselines. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 43(11):4125-4141, 2021. 2", + "[14] Andrew J. Davison. Real-time simultaneous localisation and mapping with a single camera. In ICCV, 2003. 3" + ], + "bbox": [ + 86, + 114, + 467, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[15] Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. Superpoint: Self-supervised interest point detection and description, 2018. 7", + "[16] Alexey Dosovitskiy, German Ros, Felipe Codevilla, Antonio Lopez, and Vladlen Koltun. Carla: An open urban driving simulator, 2017. 2", + "[17] Salehe Erfanian Ebadi, Saurav Dhakad, Sanjay Vishwakarma, Chunpu Wang, You-Cyuan Jhang, Maciek Chociej, Adam Crespi, Alex Thaman, and Sujoy Ganguly. Psp-hdri+: A synthetic dataset generator for pre-training of human-centric computer vision models. In First Workshop on Pre-training: Perspectives, Pitfalls, and Paths Forward at ICML 2022, 2022. 2", + "[18] Matteo Fabbri, Fabio Lanzi, Simone Calderara, Andrea Palazzi, Roberto Vezzani, and Rita Cucchiara. Learning to detect and track visible and occluded body joints in a virtual world. In European Conference on Computer Vision (ECCV), 2018. 2", + "[19] Alexander Fiannaca, Ilias Apostolopoulous, and Eelke Folmer. Headlock: A wearable navigation aid that helps blind cane users traverse large open spaces. In Proceedings of the 16th international ACM SIGACCESS conference on Computers & accessibility, pages 19-26, 2014. 2", + "[20] Zahra Ghodsi, Siva Kumar Sastry Hari, Iuri Frosio, Timothy Tsai, Alejandro Troccoli, Stephen W. Keckler, Siddharth Garg, and Anima Anandkumar. Generating and characterizing scenarios for safety testing of autonomous vehicles, 2021. 2", + "[21] Google. ARCore. https://developers.google.com/ar/, 2018.3", + "[22] Venu Madhav Govindu. Combining two-view constraints for motion estimation. In CVPR, 2001. 3", + "[23] Kristen Grauman, Andrew Westbury, Eugene Byrne, Zachary Chavis, Antonino Furnari, Rohit Girdhar, Jackson Hamburger, Hao Jiang, Miao Liu, Xingyu Liu, Miguel Martin, Tushar Nagarajan, Iija Radosavovic, Santhosh Kumar Ramakrishnan, Fiona Ryan, Jayant Sharma, Michael Wray, Mengmeng Xu, Eric Zhongcong Xu, Chen Zhao, Siddhant Bansal, Dhruv Batra, Vincent Cartillier, Sean Crane, Tien Do, Morrie Doulaty, Akshay Erapalli, Christoph Feichtenhofer, Adriano Fragomeni, Qichen Fu, Christian Fuegen, Abrham Gebreselasie, Cristina Gonzalez, James Hillis, Xuhua Huang, Yifei Huang, Wenqi Jia, Weslie Khoo, Jachym Kolar, Satwik Kottur, Anurag Kumar, Federico Landini, Chao Li, Yanghao Li, Zhenqiang Li, Karttikeya Mangalam, Raghava Modhugu, Jonathan Munro, Tullie Murrell, Takumi Nishiyasu, Will Price, Paola Ruiz Puentes, Merey Ramazanova, Leda Sari, Kiran Somasundaram, Audrey Southerland, Yusuke Sugano, Ruijie Tao, Minh Vo, Yuchen Wang, Xindi Wu, Takuma Yagi, Yunyi Zhu, Pablo Arbelaez, David Crandall, Dima Damen, Giovanni Maria Farinella, Bernard Ghanem, Vamsi Krishna Ithapu, C. V. Jawahar, Hanbyul Joo, Kris Kitani, Haizhou Li, Richard Newcombe, Aude Oliva, Hyun Soo Park, James M. Rehg, Yoichi Sato, Jianbo Shi, Mike Zheng Shou, Antonio Torralba, Lorenzo Torresani, Mingfei Yan, and Jitendra Malik. Ego4d: Around the World in 3,000 Hours of Egocentric" + ], + "bbox": [ + 509, + 92, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "14505", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Video. In IEEE/CVF Computer Vision and Pattern Recognition (CVPR), 2022. 1, 2", + "[24] Artur Grigorev, Michael J. Black, and Otmar Hilliges. HOOD: hierarchical graphs for generalized modelling of clothing dynamics. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2023, Vancouver, BC, Canada, June 17-24, 2023, pages 16965-16974. IEEE, 2023. 5", + "[25] Vladimir Guzov, Aymen Mir, Torsten Sattler, and Gerard Pons-Moll. Human poseitioning system (hps): 3d human pose estimation and self-localization in large scenes from body-mounted sensors. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4318-4329, 2021. 3", + "[26] Norbert Haala, Michael Cramer, Florian Weimer, and Martin Trittler. Performance test on uav-based photogrammetric data collection. The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, 2012. 3", + "[27] Ankur Handa, Thomas Whelan, John McDonald, and Andrew J Davison. A benchmark for rgb-d visual odometry, 3d reconstruction and slam. ICRA, 2014. 8", + "[28] Albert Haque, Boya Peng, Zelun Luo, Alexandre Alahi, Serena Yeung, and Li Fei-Fei. Towards viewpoint invariant 3d human pose estimation. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11–14, 2016, Proceedings, Part I 14, pages 160–177. Springer, 2016. 3", + "[29] Mohamed Hassan, Duygu Ceylan, Ruben Villegas, Jun Saito, Jimei Yang, Yi Zhou, and Michael Black. Stochastic scene-aware motion prediction. In Proceedings of the International Conference on Computer Vision 2021, 2, 3", + "[30] Mohamed Hassan, Yunrong Guo, Tingwu Wang, Michael Black, Sanja Fidler, and Xue Bin Peng. Synthesizing physical character-scene interactions. arXiv preprint arXiv:2302.00883, 2023. 2", + "[31] Daniel Holden, Taku Komura, and Jun Saito. Phase-functional neural networks for character control. ACM Trans. Graph., 36(4):1-13, 2017. 2", + "[32] Daniel Holden, Oussama Kanoun, Maksym Perepichka, and Tiberiu Popa. Learned motion matching. ACM Transactions on Graphics (TOG), 39(4):53-1, 2020. 2", + "[33] Yuhe Jin, Dmytro Mishkin, Anastasiia Mishchuk, Jiri Matas, Pascal Fua, Kwang Moo Yi, and Eduard Trulls. Image Matching across Wide Baselines: From Paper to Practice. IJCV, 2021. 3", + "[34] Neena Kamath. Announcing Azure Spatial Anchors for collaborative, cross-platform mixed reality apps. https://azure.microsoft.com/en-us/blog/announcing-azure-spatial-anchors-for-collaborative-cross-platform-mixed-reality-apps/, 2019.3", + "[35] Angjoo Kanazawa, Michael J Black, David W Jacobs, and Jitendra Malik. End-to-end recovery of human shape and pose. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 7122-7131, 2018. 3" + ], + "bbox": [ + 86, + 90, + 470, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[36] Alex Kendall, Matthew Grimes, and Roberto Cipolla. PoseNet: A Convolutional Network for Real-Time 6-DoF Camera Relocalization. In ICCV, 2015. 3", + "[37] Georg Klein and David Murray. Parallel tracking and mapping for small ar workspaces. In IEEE and ACM International Symposium on Mixed and Augmented Reality, 2007. 3", + "[38] Muhammed Kocabas, Chun-Hao P. Huang, Otmar Hilliges, and Michael J. Black. PARE: Part attention regressor for 3D human body estimation. In Proceedings International Conference on Computer Vision (ICCV), pages 11127-11137. IEEE, 2021. 3", + "[39] Nikos Kolotouros, Georgios Pavlakos, Michael J Black, and Kostas Daniilidis. Learning to reconstruct 3D human pose and shape via model-fitting in the loop. In Proceedings of the IEEE International Conference on Computer Vision, pages 2252-2261, 2019.", + "[40] Nikos Kolotouros, Georgios Pavlakos, and Kostas Dani-ilidis. Convolutional mesh regression for single-image human shape reconstruction. In CVPR, 2019.", + "[41] Nikos Kolotouros, Georgios Pavlakos, Dinesh Jayaraman, and Kostas Daniilidis. Probabilistic modeling for human mesh recovery. In Proceedings of the IEEE/CVF international conference on computer vision, pages 11605-11614, 2021. 3, 8", + "[42] Lucas Kovar, Michael Gleicher, and Frédéric H. Pighin. Motion graphs. In International Conference on Computer Graphics and Interactive Techniques, SIGGRAPH 2008, Los Angeles, California, USA, August 11-15, 2008, Classes, pages 51:1-51:10. ACM, 2008. 2", + "[43] Jiaman Li, C. Karen Liu, and Jiajun Wu. Ego-body pose estimation via ego-head pose estimation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2023, Vancouver, BC, Canada, June 17-24, 2023, pages 17142-17151. IEEE, 2023. 1, 2, 7", + "[44] Zhihao Li, Jianzhuang Liu, Zhensong Zhang, Songcen Xu, and Youliang Yan. CLIFF: Carrying location information in full frames into human pose and shape estimation. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part V, pages 590-606. Springer, 2022. 3", + "[45] Hung Yu Ling, Fabio Zinno, George Cheng, and Michiel van de Panne. Character controllers using motion vaes. ACM Trans. Graph., 39(4), 2020. 3", + "[46] Miao Liu, Dexin Yang, Yan Zhang, Zhaopeng Cui, James M Rehg, and Siyu Tang. 4d human body capture from egocentric video via 3d scene grounding. In 2021 International Conference on 3D Vision (3DV), pages 930-939. IEEE, 2021. 3", + "[47] Zhengyi Luo, Ryo Hachiuma, Ye Yuan, Shun Iwase, and Kris M Kitani. Kinematics-guided reinforcement learning for object-aware 3d ego-pose estimation. arXiv preprint arXiv:2011.04837, 2020. 3", + "[48] Naureen Mahmood, Nima Ghorbani, Nikolaus F Troje, Gerard Pons-Moll, and Michael J Black. Amass: Archive of motion capture as surface shapes. In Proceedings of the IEEE/CVF international conference on computer vision, pages 5442-5451, 2019. 3" + ], + "bbox": [ + 509, + 92, + 890, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "14506", + "bbox": [ + 480, + 945, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[49] Angel Martínez-González, Michael Villamizar, Olivier Canévet, and Jean-Marc Odobez. Residual pose: A decoupled approach for depth-based 3d human pose estimation. In 2020 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 10313-10318. IEEE, 2020. 3", + "[50] Meta. Project Aria Glasses. https://www.projectaria.com/, 2023.6", + "[51] Microsoft. HoloLens 2. https://www.microsoft.com/en-us/hololens, 2019.6", + "[52] Marko Mihajlovic, Shunsuke Saito, Aayush Bansal, Michael Zollhoefer, and Siyu Tang. COAP: Compositional articulated occupancy of people. In Proceedings IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2022. 6", + "[53] Aymen Mir, Xavier Puig, Angjoo Kanazawa, and Gerard Pons-Moll. Generating continual human motion in diverse 3d scenes. CoRR, abs/2304.02061, 2023. 2", + "[54] Michael Montemerlo, Nicholas Roy, and Sebastian Thrun. Perspectives on standardization in mobile robot programming: the Carnegie mellon navigation (CARMEN) toolkit. In 2003 IEEE/RSJ International Conference on Intelligent Robots and Systems, Las Vegas, Nevada, USA, October 27 - November 1, 2003, pages 2436-2441. IEEE, 2003. 2", + "[55] Gyeongsik Moon, Ju Yong Chang, and Kyoung Mu Lee. V2v-posenet: Voxel-to-voxel prediction network for accurate 3d hand and human pose estimation from a single depth map. In Proceedings of the IEEE conference on computer vision and pattern Recognition, pages 5079-5088, 2018. 3", + "[56] Etienne Mouragnon, Maxime Lhuillier, Michel Dhome, Fabien Dekeyser, and Patrick Sayd. Real time localization and 3d reconstruction. In CVPR, 2006. 3", + "[57] Evonne Ng, Donglai Xiang, Hanbyul Joo, and Kristen Grauman. You2me: Inferring body pose in egocentric video via first and second person interactions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9890-9900, 2020. 3", + "[58] Tony Ng, Adrian Lopez-Rodriguez, Vassileios Balntas, and Krystian Mikolajczyk. Reassessing the limitations of cnn methods for camera pose regression, 2021. 7", + "[59] Niantic. Niantic Expands Developer Platform and AR Tools with Niantic Lightship. https://nianticlabs.com/news/lightship/, 2021.3", + "[60] David Nister, Oleg Naroditsky, and James Bergen. Visual odometry. In CVPR, 2004. 3", + "[61] Xiaqing Pan, Nicholas Charron, Yongqian Yang, Scott Peters, Thomas Whelan, Chen Kong, Omkar M. Parkhi, Richard A. Newcombe, and Carl Yuheng Ren. Aria digital twin: A new benchmark dataset for egocentric 3d machine perception. CoRR, abs/2306.06362, 2023. 2", + "[62] Fabio Pardo, Arash Tavakoli, Vitaly Levdik, and Petar Kormushev. Time limits in reinforcement learning. In Proceedings of the 35th International Conference on Machine Learning, pages 4045-4054, Stockholm, Sweden, 2018. PMLR. 4", + "[63] Georgios Pavlakos, Vasileios Choutas, Nima Ghorbani, Timo Bolkart, Ahmed A. A. Osman, Dimitrios Tzionas," + ], + "bbox": [ + 86, + 90, + 468, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "and Michael J. Black. Expressive body capture: 3D hands, face, and body from a single image. In Proceedings IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), pages 10975-10985, 2019. 3, 5, 7", + "[64] Xue Bin Peng, Ze Ma, Pieter Abbeel, Sergey Levine, and Angjoo Kanazawa. Amp: Adversarial motion priors for stylized physics-based character control. ACM Transactions on Graphics (TOG), 40(4):1-20, 2021. 2", + "[65] Mathis Petrovich, Michael J. Black, and Gül Varol. Action-conditioned 3d human motion synthesis with transformer vae, 2021. 3", + "[66] Marc Pollefeys, Luc Van Gool, Maarten Vergauwen, Frank Verbiest, Kurt Cornelis, Jan Tops, and Reinhard Koch. Visual modeling with a hand-held camera. IJCV, 2004. 3", + "[67] Xavier Puig, Eric Undersander, Andrew Szot, Mikael Dallaire Cote, Tsung-Yen Yang, Ruslan Partsey, Ruta Desai, Alexander William Clegg, Michal Hlavac, So Yeon Min, Vladimir Vondrus, Théophile Gervet, Vincent-Pierre Berges, John M. Turner, Oleksandr Maksymets, Zsolt Kira, Mrinal Kalakrishnan, Jitendra Malik, Devendra Singh Chaplot, Unnat Jain, Dhruv Batra, Akshara Rai, and Roozbeh Mottaghi. Habitat 3.0: A co-habitat for humans, avatars and robots. CoRR, abs/2310.13724, 2023. 2, 8", + "[68] Albert Pumarola, Jordi Sanchez, Gary Choi, Alberto Sanfeliu, and Francesc Moreno-Noguer. 3DPeople: Modeling the Geometry of Dressed Humans. In International Conference in Computer Vision (ICCV), 2019. 2", + "[69] Morgan Quigley, Ken Conley, Brian Gerkey, Josh Faust, Tully Foote, Jeremy Leibs, Rob Wheeler, Andrew Y Ng, et al. Ros: an open-source robot operating system. In ICRA workshop on open source software, page 5. Kobe, Japan, 2009. 2", + "[70] Umer Rafi, Juergen Gall, and Bastian Leibe. A semantic occlusion model for human pose estimation from a single depth image. In Proceedings of the IEEE conference on computer vision and pattern recognition workshops, pages 67-74, 2015. 3", + "[71] Tilman Reinhardt. Google Visual Positioning Service. https://ai.googleblog.com/2019/02/using-global-localization-to-improve.html, 2019.3", + "[72] Davis Rempe, Tolga Birdal, Aaron Hertzmann, Jimei Yang, Srinath Sridhar, and Leonidas J. Guibas. Humor: 3d human motion model for robust pose estimation. In International Conference on Computer Vision (ICCV), 2021. 3", + "[73] Davis Rempe, Zhengyi Luo, Xue Bin Peng, Ye Yuan, Kris Kitani, Karsten Kreis, Sanja Fidler, and Or Litany. Trace and pace: Controllable pedestrian animation via guided trajectory diffusion, 2023. 2, 3", + "[74] Guodong Rong, Byung Hyun Shin, Hadi Tabatabaee, Qiang Lu, Steve Lemke, Martiš Možeiko, Eric Boise, Geehoon Uhm, Mark Gerow, Shalin Mehta, et al. Lgsvl simulator: A high fidelity simulator for autonomous driving. In 2020 IEEE 23rd International conference on intelligent transportation systems (ITSC), pages 1-6. IEEE, 2020. 2", + "[75] Paul-Edouard Sarlin, Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. Superglue: Learning feature matching with graph neural networks, 2020. 7" + ], + "bbox": [ + 509, + 90, + 890, + 900 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "14507", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[76] Paul-Edouard Sarlin, Mihai Dusmanu, Johannes L. Schonberger, Pablo Speciale, Lukas Gruber, Viktor Larsson, Ondrej Miksik, and Marc Pollefeys. LaMAR: Benchmarking Localization and Mapping for Augmented Reality. In ECCV, 2022. 3, 7", + "[77] Torsten Sattler, Will Maddern, Carl Toft, Akihiko Torii, Lars Hammarstrand, Erik Stenberg, Daniel Safari, Masatoshi Okutomi, Marc Pollefeys, Josef Sivic, Fredrik Kahl, and Tomas Pajdla. Benchmarking 6DoF outdoor visual localization in changing conditions. In CVPR, 2018. 3", + "[78] Manolis Savva, Abhishek Kadian, Oleksandr Maksymets, Yili Zhao, Erik Wijmans, Bhavana Jain, Julian Straub, Jia Liu, Vladlen Koltun, Jitendra Malik, Devi Parikh, and Dhruv Batra. Habitat: A platform for embodied ai research, 2019. 2", + "[79] Johannes Lutz Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In CVPR, 2016. 3", + "[80] Johannes L. Schonberger, Hans Hardmeier, Torsten Sattler, and Marc Pollefeys. Comparative evaluation of handcrafted and learned local features. In CVPR, 2017. 3", + "[81] John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Radford, and Oleg Klimov. Proximal policy optimization algorithms, 2017. 4", + "[82] Akash Sengupta, Ignas Budvytis, and Roberto Cipolla. Synthetic training for accurate 3d human pose and shape estimation in the wild. In British Machine Vision Conference (BMVC), 2020. 2", + "[83] Takaaki Shiratori, Hyun Soo Park, Leonid Sigal, Yaser Sheikh, and Jessica K. Hodgins. Motion capture from body-mounted cameras. ACM Trans. Graph., 30(4):31, 2011. 3", + "[84] Jamie Shotton, Ross Girshick, Andrew Fitzgibbon, Toby Sharp, Mat Cook, Mark Finocchio, Richard Moore, Pushmeet Kohli, Antonio Criminisi, Alex Kipman, et al. Efficient human pose estimation from single depth images. IEEE transactions on pattern analysis and machine intelligence, 35(12):2821-2840, 2012. 3", + "[85] Jamie Shotton, Ben Glocker, Christopher Zach, Shahram Izadi, Antonio Criminisi, and Andrew Fitzgibbon. Scene Coordinate Regression Forests for Camera Relocalization in RGB-D Images. In CVPR, 2013. 3", + "[86] Jamie Shotton, Toby Sharp, Alex Kipman, Andrew Fitzgibbon, Mark Finocchio, Andrew Blake, Mat Cook, and Richard Moore. Real-time human pose recognition in parts from single depth images. Commun. ACM, 56(1):116-124, 2013. 2", + "[87] Noah Snavely, Steven M. Seitz, and Richard Szeliski. Photo tourism: exploring photo collections in 3d. ACM Trans. Graph., 25(3):835-846, 2006. 3", + "[88] Kihyuk Sohn, Honglak Lee, and Xinchen Yan. Learning structured output representation using deep conditional generative models. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2015. 3", + "[89] Sebastian Starke, He Zhang, Taku Komura, and Jun Saito. Neural state machine for character-scene interactions. ACM Trans. Graph., 38(6):209-1, 2019. 2", + "[90] Sebastian Starke, Ian Mason, and Taku Komura. Deepphase: Periodic autoencoders for learning motion phase" + ], + "bbox": [ + 86, + 92, + 470, + 902 + ], + "page_idx": 11 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "manifolds. ACM Transactions on Graphics (TOG), 41(4): 1-13, 2022. 2", + "[91] Julian Straub, Thomas Whelan, Lingni Ma, Yufan Chen, Erik Wijmans, Simon Green, Jakob J. Engel, Raul Mur-Artal, Carl Ren, Shobhit Verma, Anton Clarkson, Mingfei Yan, Brian Budge, Yajie Yan, Xiaqing Pan, June Yon, Yuyang Zou, Kimberly Leon, Nigel Carter, Jesus Briales, Tyler Gillingham, Elias Mueggler, Luis Pesqueira, Manolis Savva, Dhruv Batra, Hauke M. Strasdat, Renzo De Nardi, Michael Goesele, Steven Lovegrove, and Richard Newcombe. The Replica dataset: A digital replica of indoor spaces. arXiv preprint arXiv:1906.05797, 2019. 4, 6", + "[92] Andrew Szot, Alex Clegg, Eric Undersander, Erik Wijmans, Yili Zhao, John Turner, Noah Maestre, Mustafa Mukadam, Devendra Chaplot, Oleksandr Maksymets, Aaron Gokaslan, Vladimir Vondrus, Sameer Dharur, Franziska Meier, Wojciech Galuba, Angel Chang, Zsolt Kira, Vladlen Koltun, Jitendra Malik, Manolis Savva, and Dhruv Batra. Habitat 2.0: Training home assistants to rearrange their habitat. In Advances in Neural Information Processing Systems (NeurIPS), 2021. 2", + "[93] Zachary Teed and Jia Deng. DROID-SLAM: Deep Visual SLAM for Monocular, Stereo, and RGB-D Cameras. Advances in neural information processing systems, 2021. 7", + "[94] Guy Tevet, Sigal Raab, Brian Gordon, Yonatan Shafir, Daniel Cohen-Or, and Amit H Bermano. Human motion diffusion model. arXiv preprint arXiv:2209.14916, 2022.3", + "[95] Emanuel Todorov, Tom Erez, and Yuval Tassa. Mujoco: A physics engine for model-based control. In 2012 IEEE/RSJ International Conference on Intelligent Robots and Systems, pages 5026-5033, 2012. 2", + "[96] Emanuel Todorov, Tom Erez, and Yuval Tassa. Mujoco: A physics engine for model-based control. In 2012 IEEE/RSJ International Conference on Intelligent Robots and Systems, pages 5026-5033. IEEE, 2012. 2", + "[97] Denis Tome, Patrick Peluse, Lourdes Agapito, and Hernan Badino. xr-egopose: Egocentric 3d human pose from an hmd camera. In Proceedings of the IEEE International Conference on Computer Vision, pages 7728-7738, 2019. 2, 3", + "[98] Denis Tome, Thiemo Alldieck, Patrick Peluse, Gerard Pons-Moll, Lourdes Agapito, Hernan Badino, and Fernando De la Torre. Selfpose: 3d egocentric pose estimation from a headset mounted camera. arXiv preprint arXiv:2011.01519, 2020. 3", + "[99] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Advances in neural information processing systems, pages 5998-6008, 2017. 4", + "[100] Jian Wang, Diogo Luvizon, Weipeng Xu, Lingjie Liu, Kripasindhu Sarkar, and Christian Theobalt. Scene-aware egocentric 3d human pose estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13031-13040, 2023. 3", + "[101] Keze Wang, Shengfu Zhai, Hui Cheng, Xiaodan Liang, and Liang Lin. Human pose estimation from depth images via" + ], + "bbox": [ + 501, + 92, + 893, + 902 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "14508", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 11 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "inference embedded multi-task learning. In Proceedings of the 24th ACM international conference on Multimedia, pages 1227-1236, 2016. 3", + "[102] Xin Wang, Taein Kwon, Mahdi Rad, Bowen Pan, Ishani Chakraborty, Sean Andrist, Dan Bohus, Ashley Feniello, Bugra Tekin, Felipe Vieira Frujeri, Neel Joshi, and Marc Pollefeys. Holoassist: an egocentric human interaction dataset for interactive ai assistants in the real world. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 20270-20281, 2023. 2", + "[103] Erik Wijmans, Abhishek Kadian, Ari Morcos, Stefan Lee, Irfan Essa, Devi Parikh, Manolis Savva, and Dhruv Batra. DD-PPO: learning near-perfect pointgoal navigators from 2.5 billion frames. In 8th International Conference on Learning Representations, ICLR 2020, Addis Ababa, Ethiopia, April 26-30, 2020. OpenReview.net, 2020. 3", + "[104] Kyle Wilson and Noah Snavely. Robust global translations with 1dsfm. In ECCV, 2014. 3", + "[105] Jungdam Won, Deepak Gopinath, and Jessica Hodgins. Physics-based character controllers using conditional vaes. ACM Trans. Graph., 41(4), 2022. 2", + "[106] Erroll Wood, Tadas Baltrusaitis, Charlie Hewitt, Sebastian Dziadzio, Thomas J Cashman, and Jamie Shotton. Fake it till you make it: face analysis in the wild using synthetic data alone. In Proceedings of the IEEE/CVF international conference on computer vision, pages 3681-3691, 2021. 2", + "[107] Fu Xiong, Boshen Zhang, Yang Xiao, Zhiguo Cao, Taidong Yu, Joey Tianyi Zhou, and Junsong Yuan. A2j: Anchor-to-joint regression network for 3d articulated pose estimation from a single depth image. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 793-802, 2019. 3", + "[108] Mengyuan Yan, Gen Li, Yilin Zhu, and Jeannette Bohg. Learning topological motion primitives for knot planning. In IEEE/RSJ International Conference on Intelligent Robots and Systems, IROS 2020, Las Vegas, NV, USA, October 24, 2020 - January 24, 2021, pages 9457-9464. IEEE, 2020. 3", + "[109] Zhitao Yang, Zhongang Cai, Haiyi Mei, Shuai Liu, Zhaoxi Chen, Weiye Xiao, Yukun Wei, Zhongfei Qing, Chen Wei, Bo Dai, Wayne Wu, Chen Qian, Dahua Lin, Ziwei Liu, and Lei Yang. Synbody: Synthetic dataset with layered human models for 3d human perception and modeling, 2023. 5", + "[110] Heyuan Yao, Zhenhua Song, Baoquan Chen, and Libin Liu. Controlvae: Model-based learning of generative controllers for physics-based characters. ACM Trans. Graph., 41(6), 2022. 3", + "[111] Mao Ye, Xianwang Wang, Ruigang Yang, Liu Ren, and Marc Pollefeys. Accurate 3d pose estimation from a single depth image. In 2011 International Conference on Computer Vision, pages 731-738. IEEE, 2011. 3", + "[112] Vickie Ye, Georgios Pavlakos, Jitendra Malik, and Angjoo Kanazawa. Decoupling human and camera motion from videos in the wild. arXiv preprint arXiv:2302.12827, 2023. 3", + "[113] Ye Yuan and Kris Kitani. Ego-posed estimation and forecasting as real-time pd control. In Proceedings of the" + ], + "bbox": [ + 78, + 92, + 467, + 900 + ], + "page_idx": 12 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "IEEE/CVF International Conference on Computer Vision, pages 10082-10092, 2019. 3", + "[114] Mingyuan Zhang, Zhongang Cai, Liang Pan, Fangzhou Hong, Xinying Guo, Lei Yang, and Ziwei Liu. Motiondiffuse: Text-driven human motion generation with diffusion model. arXiv preprint arXiv:2208.15001, 2022. 3", + "[115] Siwei Zhang, Qianli Ma, Yan Zhang, Zhiyin Qian, Taein Kwon, Marc Pollefeys, Federica Bogo, and Siyu Tang. Egobody: Human body shape and motion of interacting people from head-mounted devices. In European conference on computer vision (ECCV), 2022. 1, 2, 8", + "[116] Siwei Zhang, Qianli Ma, Yan Zhang, Sadegh Aliakbarian, Darren Cosker, and Siyu Tang. Probabilistic human mesh recovery in 3d scenes from egocentric views. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2023. 3", + "[117] Yan Zhang and Siyu Tang. The wanderings of odysseus in 3d scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20481-20491, 2022. 3, 6", + "[118] Yan Zhang, Michael J Black, and Siyu Tang. We are more than our joints: Predicting how 3d bodies move. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3372-3382, 2021. 3", + "[119] Kaifeng Zhao, Yan Zhang, Shaofei Wang, Thabo Beeler, and Siyu Tang. Synthesizing diverse human motions in 3d indoor scenes. In International conference on computer vision (ICCV), 2023. 2, 3, 4, 6", + "[120] Yuhang Zhao, Elizabeth Kupferstein, Brenda Veronica Castro, Steven Feiner, and Shiri Azenkot. Designing ar visualizations to facilitate stair navigation for people with low vision. In Proceedings of the 32nd annual ACM symposium on user interface software and technology, pages 387-402, 2019. 2", + "[121] Yang Zheng, Yanchao Yang, Kaichun Mo, Jiaman Li, Tao Yu, Yebin Liu, C Karen Liu, and Leonidas J Guibas. Gimo: Gaze-informed human motion prediction in context. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XIII, pages 676-694. Springer, 2022. 7", + "[122] Ziwen Zhuang, Zipeng Fu, Jianren Wang, Christopher Atkeson, Soren Schwertfeger, Chelsea Finn, and Hang Zhao. Robot parkour learning. In Conference on Robot Learning (CoRL), 2023. 2, 3" + ], + "bbox": [ + 503, + 92, + 890, + 712 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "14509", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 12 + } +] \ No newline at end of file diff --git a/2024/EgoGen_ An Egocentric Synthetic Data Generator/717ad579-fa9d-4ce5-996d-77c046c2cdbf_model.json b/2024/EgoGen_ An Egocentric Synthetic Data Generator/717ad579-fa9d-4ce5-996d-77c046c2cdbf_model.json new file mode 100644 index 0000000000000000000000000000000000000000..bed07d1e6cc3e9e01ae0d6aca6139ffe0143953f --- /dev/null +++ b/2024/EgoGen_ An Egocentric Synthetic Data Generator/717ad579-fa9d-4ce5-996d-77c046c2cdbf_model.json @@ -0,0 +1,3141 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.231, + 0.131, + 0.741, + 0.154 + ], + "angle": 0, + "content": "EgoGen: An Egocentric Synthetic Data Generator" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.18, + 0.756, + 0.257 + ], + "angle": 0, + "content": "Gen Li\\(^{1}\\) Kaifeng Zhao\\(^{1}\\) Siwei Zhang\\(^{1}\\) Xiaozhong Lyu\\(^{1}\\) \nMihai Dusmanu\\(^{2}\\) Yan Zhang\\(^{1}\\) Marc Pollefeys\\(^{1,2}\\) Siyu Tang\\(^{1}\\) \n\\(^{1}\\)ETH Zürich \n\\(^{2}\\)Microsoft \nhttps://ego-gen.github.io/" + }, + { + "type": "image", + "bbox": [ + 0.08, + 0.274, + 0.485, + 0.484 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.135, + 0.489, + 0.16, + 0.498 + ], + "angle": 0, + "content": "RGB" + }, + { + "type": "image_caption", + "bbox": [ + 0.232, + 0.489, + 0.333, + 0.5 + ], + "angle": 0, + "content": "RGB with Motion Blur" + }, + { + "type": "image_caption", + "bbox": [ + 0.402, + 0.489, + 0.432, + 0.5 + ], + "angle": 0, + "content": "Depth" + }, + { + "type": "image", + "bbox": [ + 0.484, + 0.274, + 0.618, + 0.484 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.517, + 0.489, + 0.587, + 0.499 + ], + "angle": 0, + "content": "Surface Normal" + }, + { + "type": "image", + "bbox": [ + 0.619, + 0.274, + 0.753, + 0.484 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.643, + 0.489, + 0.73, + 0.5 + ], + "angle": 0, + "content": "Segmentation Mask" + }, + { + "type": "image", + "bbox": [ + 0.754, + 0.274, + 0.892, + 0.484 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.789, + 0.489, + 0.854, + 0.499 + ], + "angle": 0, + "content": "World Position" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.516, + 0.895, + 0.572 + ], + "angle": 0, + "content": "Figure 1. EgoGen: a scalable synthetic data generation system for egocentric perception tasks, with rich multi-modal data and accurate annotations. We simulate camera rigs for head-mounted devices (HMDs) and render from the perspective of the camera wearer with various sensors. Top to bottom: middle and right camera sensors in the rig. Left to right: photo-realistic RGB image, RGB with simulated motion blur, depth map, surface normal, segmentation mask, and world position for fisheye cameras widely used in HMDs." + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.585, + 0.314, + 0.6 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.629, + 0.473, + 0.903 + ], + "angle": 0, + "content": "Understanding the world in first-person view is fundamental in Augmented Reality (AR). This immersive perspective brings dramatic visual changes and unique challenges compared to third-person views. Synthetic data has empowered third-person-view vision models, but its application to embodied egocentric perception tasks remains largely unexplored. A critical challenge lies in simulating natural human movements and behaviors that effectively steer the embodied cameras to capture a faithful egocentric representation of the 3D world. To address this challenge, we introduce EgoGen, a new synthetic data generator that can produce accurate and rich ground-truth training data for egocentric perception tasks. At the heart of EgoGen is a novel human motion synthesis model that directly leverages egocentric visual inputs of a virtual human to sense the 3D environment. Combined with collision-avoiding motion primitives and a two-stage reinforcement learning approach, our motion synthesis model offers a closed-loop so" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.586, + 0.895, + 0.768 + ], + "angle": 0, + "content": "lution where the embodied perception and movement of the virtual human are seamlessly coupled. Compared to previous works, our model eliminates the need for a pre-defined global path, and is directly applicable to dynamic environments. Combined with our easy-to-use and scalable data generation pipeline, we demonstrate EgoGen's efficacy in three tasks: mapping and localization for head-mounted cameras, egocentric camera tracking, and human mesh recovery from egocentric views. EgoGen will be fully open-sourced, offering a practical solution for creating realistic egocentric training data and aiming to serve as a useful tool for egocentric computer vision research." + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.785, + 0.633, + 0.801 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.81, + 0.895, + 0.903 + ], + "angle": 0, + "content": "The analysis of visual input from front-facing egocentric cameras is crucial for applications that benefit from a first-person perspective, mirroring the natural human experience [23, 43, 115]. AR devices, for instance, can utilize this viewpoint to amplify user immersion. Such cameras can also cater to individual preferences, providing custom" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "14497" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.461, + 0.108 + ], + "angle": 0, + "content": "visual assistance for those with impaired vision [19, 120]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.109, + 0.468, + 0.26 + ], + "angle": 0, + "content": "Despite its potential, egocentric perception faces challenges, primarily due to the scarcity of labeled data. Although datasets like Ego4D [23], ADT [61], EpicKitchen [13] and HoloAssist [102] exist, creating such datasets with rich and accurate annotations is costly and raises privacy concerns [115]. Alternatively, using graphics techniques to render synthetic multi-modal visual data has proven to be cost-effective and successful in training deep learning models, such as 3D human body estimation [8] and facial landmark detection [106]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.262, + 0.468, + 0.337 + ], + "angle": 0, + "content": "Creating egocentric synthetic data is challenging because egocentric cameras capture the complex interplay of body movements and the environment from the camera wearer's viewpoint. Modeling the intricate details and variations in human behavior presents a significant challenge." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.339, + 0.468, + 0.474 + ], + "angle": 0, + "content": "To tackle this problem, we introduce EgoGen, an egocentric synthetic data generation approach that simulates data from embodied sensors, i.e., front-facing cameras in head-mounted devices (HMD). While the ultimate goal is to simulate human behaviors that are indistinguishable from reality, in this work, we focus on creating virtual humans (i.e., camera wearers) that can explore and avoid obstacles in the 3D world that is not only complex and dynamic but could potentially include other moving virtual humans." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.477, + 0.468, + 0.809 + ], + "angle": 0, + "content": "Specifically, we propose a novel generative human motion model. Our key insight is that body movement and embodied perception should be seamlessly coupled. As William Gibson aptly stated, \"We see in order to move; we move in order to see,\" our egocentric perception is crucial for identifying obstacles, navigating in an environment, and planning actions. Our body movements are not solely a response to visual stimuli; they also change our egocentric perception. Therefore, the key idea of our motion model is to enable virtual humans to see their environment with egocentric visual inputs and respond accordingly by learning a policy to control a set of collision-avoiding motion primitives (CAMPs) that are composable for synthesizing long-term, diverse human motions. Due to the unbounded and high-dimensional latent action space of our generative motion primitive model, direct policy training through rendered egocentric images is often unstable [122]. Therefore, we propose a two-stage reinforcement learning scheme using an efficient egocentric visual proxy to couple egocentric visual cues and body movements seamlessly. In addition, we use an \"attention\" reward to incentivize egocentric perception behaviors, i.e., looking in the desired direction." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.811, + 0.468, + 0.901 + ], + "angle": 0, + "content": "Empirical results showcase the benefits of our egocentric perception-driven motion framework, which does not require a pre-calculated walking path in 3D scenes as in [29, 53, 119]. Instead, it empowers virtual humans to perceive the environment from their own viewpoint, enabling them to navigate, circumvent obstacles, and plan move" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.227 + ], + "angle": 0, + "content": "ments to reach the destination. Moreover, our model generalizes well to dynamic environments, even with training limited to static settings. By training virtual humans independently using CAMPs, our method synthesizes emergent multi-human behaviors without relying on multi-agent reinforcement learning algorithms. Egocentric visual cues are essential to build exploratory and generalizable motion models that unify navigational planning and movement control in complex and dynamic environments." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.228, + 0.892, + 0.348 + ], + "angle": 0, + "content": "Building upon CAMPs, we further create a scalable data generation pipeline for EgoGen that outfits virtual humans with clothing, automates cloth animation, and integrates 3D assets from various sources. We validate EgoGen's efficacy across three egocentric perception tasks. The high-quality synthetic data with precise ground truth annotations consistently improve the performance of state-of-the-art methods. In summary, the contributions of this work are:" + }, + { + "type": "text", + "bbox": [ + 0.501, + 0.349, + 0.892, + 0.393 + ], + "angle": 0, + "content": "1. We introduce EgoGen, a generative and scalable synthetic data generation approach specifically tailored for egocentric perception tasks." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.394, + 0.892, + 0.483 + ], + "angle": 0, + "content": "2. We introduce novel motion primitives based on egocentric visual cues, enabling diverse and realistic human motion synthesis in 3D scenes. These primitives empower virtual humans to handle complex scenarios, such as dynamic environments and crowd motion without relying on multi-agent reinforcement learning." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.484, + 0.892, + 0.574 + ], + "angle": 0, + "content": "3. EgoGen enables us to augment existing real-world egocentric datasets with synthetic images. Quantitative results demonstrate enhanced performance in state-of-the-art algorithms on mapping and localization for head-mounted cameras, egocentric camera tracking, and human mesh recovery from egocentric views." + }, + { + "type": "list", + "bbox": [ + 0.5, + 0.349, + 0.892, + 0.574 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.589, + 0.642, + 0.604 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.614, + 0.892, + 0.81 + ], + "angle": 0, + "content": "Human-Related Simulators and Synthetic Data. Previous works primarily focus on simulating robots [54, 69, 78, 92, 95] and autonomous cars [7, 16, 20, 74]. While some incorporated animated digital humans, like in [5, 20, 67], these efforts relied on pre-recorded motion sequences. Rendering images of people to train perception models has been widely studied such as [5, 17, 18, 43, 68, 82, 86]. In particular, [97] offers large-scale synthetic data for egocentric camera wearer pose estimation but relies on mocap data, lacking realistic and spontaneous interactions with the digital world. In contrast, EgoGen closes the gap in egocentric synthetic data generation for head-mounted devices. Please refer to Supp. Mat. for detailed comparisons." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.811, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Human Motion Synthesis. Generating high-fidelity human motions and interactions with 3D scenes is widely studied in graphics [10, 31, 32, 42, 89, 90]. While they can generate high-quality motion, it's usually deterministic. Synthesizing physically plausible human motions has been extensively studied [12, 30, 64, 73, 96, 105]. However," + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "14498" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.242 + ], + "angle": 0, + "content": "they struggle with generalization to different body shapes. For example, [73] explicitly created 2048 humanoids to improve body shape generalizability. Time series models [65, 94, 114] synthesize the stochastic motions of diverse people well. However, in [94, 114], their generated motion sequences have limited lengths and human-scene interactions are not explicitly considered. Autoregressive methods [45, 72, 117, 119] can produce perpetual motions. In particular, [117] can generalize to diverse body shapes and synthesize long-term human motions." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.243, + 0.473, + 0.349 + ], + "angle": 0, + "content": "Our egocentric perception-driven motion synthesis model is closely related to [117, 119], but distinguishes itself w.r.t.: (1) Enabling virtual humans to explore using egocentric visual cues, without predefined paths. (2) Synthesizing egocentric perception behaviors beyond locomotion, e.g., looking in certain directions. (3) Handling dynamic environments and multi-agent behavior without re-training." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.349, + 0.473, + 0.544 + ], + "angle": 0, + "content": "Mapping and Localization for AR. Localization and mapping from images is a long-standing problem known as: Photogrammetry [2, 26]; Structure-from-Motion (SfM) [22, 66, 79, 87, 104]; Simultaneous Localization and Mapping (SLAM) [14, 37, 56, 60]. Researchers have worked to make SLAM amenable for edge hand-held or head-mounted devices [6, 21, 37]. Cloud-based services like Google's Visual Positioning System [71], Niantic's Lightship [59], and Microsoft's Azure Spatial Anchors [34] have made visual localization and mapping more accessible. Benchmarking efforts have arisen for small-scale AR scenarios [36, 85], touristic landmarks [33, 80], and large-scale AR-device based localization [76, 77] to evaluate these systems." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.545, + 0.473, + 0.803 + ], + "angle": 0, + "content": "Egocentric Human Pose Estimation. Estimating 3D bodies from RGB images is widely studied from third-person views [9, 35, 38-41, 44, 63], and egocentric views [25, 46, 47, 57, 83, 97, 98, 100, 112, 113, 116], mostly requiring expensive real-world data paired with ground truth annotations. Besides RGB images, depth images offer explicit 3D information, mitigating scale and shape ambiguity, with the potential to enable broader AR/VR applications. However, depth-based methods, especially for the egocentric view, are underexplored. Most existing works [28, 49, 55, 70, 84, 101, 107, 111] predict 3D body skeletons without expressive body meshes, struggling with challenges like severe body truncations and scene occlusions typical in egocentric views. Such limited attention mainly stems from the scarcity of data, as obtaining high-quality human mesh annotations for real-world depth images is labor-intensive." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.815, + 0.418, + 0.833 + ], + "angle": 0, + "content": "3. Ego-Sensing Driven Motion Synthesis" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.472, + 0.903 + ], + "angle": 0, + "content": "To close the loop for the interdependence between egocentric synthetic image data and human motion synthesis, we use deep reinforcement learning (RL), integrating egocentric vision cues to synthesize human motions as described in" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.892, + 0.123 + ], + "angle": 0, + "content": "Sec. 3.1 and 3.2. Subsequently, we extend learned policies to generate emergent multi-agent behaviors, as in Sec. 3.3." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.134, + 0.833, + 0.15 + ], + "angle": 0, + "content": "3.1. Ego-Sensing Driven Motion Primitives" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.158, + 0.892, + 0.233 + ], + "angle": 0, + "content": "Generating realistic egocentric data requires diverse and lifelike human motion synthesis. In this work, we consider arguably the most common everyday behaviors: navigating towards goals with egocentric perception while avoiding collisions with obstacles and people in dynamic 3D scenes." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.234, + 0.892, + 0.416 + ], + "angle": 0, + "content": "Overview. Following recent literature [45, 110, 117, 119], we employ deep RL to train control policies on learned latent spaces that characterize natural human motions. However, unlike these previous works that only consider simple static scenes, we leverage egocentric perception and propose collision-avoiding motion primitives (CAMPs) to enable virtual humans to self-exlore and navigate in a dynamic environment. Specifically, CAMPs are trained jointly to produce collision-free motion sequences. At each timestep \\( t \\), the agent observes the state \\( \\mathbf{s}_t \\), performs an action \\( \\mathbf{a}_t \\), and receives a reward \\( r_t = r(\\mathbf{s}_t, \\mathbf{a}_t, \\mathbf{s}_{t+1}) \\), where \\( \\mathbf{s}_{t+1} \\) represents the next state of the environment due to \\( \\mathbf{a}_t \\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.417, + 0.892, + 0.628 + ], + "angle": 0, + "content": "Egocentric Sensing As Depth Proxy. We aim to sample actions given by a policy to synthesize realistic human motions. Egocentric perception-driven motion synthesis should arguably use egocentric vision as input. However, depth rendering is costly and RL requires billions of samples to converge [3, 103]. Besides, directly training RL with visual data can be unstable [122]. We thereby use a cheap-to-compute egocentric sensing \\(\\mathcal{E}_t\\) as a proxy for depth images as illustrated in Fig. 2. \\(N\\) rays are cast evenly from the midpoint of two eyeballs, i.e., the location of the egocentric camera. The field of view \\([\\theta_{min},\\theta_{max}]\\), centered on the 2D projection of the viewing direction \\(\\vec{\\nu}\\), limits the agent's perception to the front area. Rays stop at collisions, with collision detection in 2D. See more details in Supp. Mat." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.629, + 0.892, + 0.689 + ], + "angle": 0, + "content": "Agent Representation. The agent is a virtual human represented by an SMPL-X mesh [63]. We further compact the body representation by selecting \\( M = 67 \\) body surface markers \\( \\mathbf{x} \\in \\mathbb{R}^{M \\times 3} \\) on the mesh following [118]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.689, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Environment. Inspired by [108], we aim to learn a library of composable CAMPs. We implement a finite-horizon environment based on the generative motion primitive model from [117]. Specifically, a motion primitive is defined as a 0.5-second motion clip containing \\( T = 20 \\) frames in the canonical coordinate, and each frame contains a single agent representation. The primitive model \\( \\mathcal{P} \\) is based on the C-VAE framework [88], which takes the first \\( T_{s} = 2 \\) frames as the condition, and models a conditional probability of the next \\( T - T_{s} \\) frames. Compared to [117] trained on the AMASS dataset [48] with many sport motion sequences, we train \\( \\mathcal{P} \\) using the SAMP dataset [29], which focuses on daily activities, better suited for HMD use cases. Our action space \\( \\mathcal{A} \\) is the pretrained 128D latent space of \\( \\mathcal{P} \\), and the" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "14499" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.078, + 0.087, + 0.895, + 0.311 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.319, + 0.895, + 0.366 + ], + "angle": 0, + "content": "Figure 2. Policy network architecture. We learn a generalizable mapping from motion seed body markers \\(\\mathbf{X}_t^S\\), marker directions \\(\\mathbf{X}_t^{S^D}\\), egocentric sensing \\(\\mathcal{E}_t\\), and distance to the target \\(d_t\\) to CAMPs. The policy learns a stochastic collision avoiding action space to predict future body markers \\(\\mathbf{X}_t^F\\). For illustration purposes, we visualize only one frame of \\(\\mathbf{X}_t^S\\) and \\(\\mathcal{E}_t\\). See Sec. 3.1 and 3.2 for details." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.383, + 0.373, + 0.398 + ], + "angle": 0, + "content": "action \\(\\mathbf{a}_t\\) can be randomly sampled from \\(\\mathcal{A}\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.398, + 0.471, + 0.475 + ], + "angle": 0, + "content": "With the input of a random action \\(\\mathbf{a}_t\\in \\mathbb{R}^{128}\\) and a motion seed \\(\\mathbf{X}_t^S = [\\mathbf{x}_t^0,\\mathbf{x}_t^{T_s - 1}]\\) (history frames), \\(\\mathcal{P}\\) predicts future frames \\(\\mathbf{X}_t^F = [\\mathbf{x}_t^{T_s},\\dots ,\\mathbf{x}_t^{T - 1}]\\) of the current motion primitive \\(\\mathbf{X}_t = [\\mathbf{X}_t^S,\\mathbf{X}_t^F ]\\in \\mathbb{R}^{T\\times M\\times 3}\\), which represents a short sequence of human motion spanning \\(0.5\\mathrm{s}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.211, + 0.479, + 0.337, + 0.499 + ], + "angle": 0, + "content": "\\[\n\\mathbf {X} _ {t} ^ {F} = \\mathcal {P} (\\mathbf {X} _ {t} ^ {S}, \\mathbf {a} _ {t})\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.504, + 0.47, + 0.596 + ], + "angle": 0, + "content": "State. To preserve Markov property [62], the state is defined as \\(\\mathbf{s}_t = \\{\\mathbf{X}_t^S, \\mathbf{X}_t^{S^D}, \\mathcal{E}_t, d_t, \\tau_t\\}\\), in which \\(\\mathbf{X}_t^{S^D} \\in \\mathbb{R}^{T_s \\times M \\times 3}\\) denotes the normalized direction of each marker seed to the target, \\(\\mathcal{E}_t \\in \\mathbb{R}^{T_s \\times N}\\) denotes the egocentric sensing depth proxy, \\(d_t\\) denotes the distance from the pelvis to the target, and \\(\\tau_t\\) denotes the remaining time. See Fig. 2." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.597, + 0.47, + 0.673 + ], + "angle": 0, + "content": "Reward. To synthesize egocentric perception behaviors, we use an \"attention reward\" to incentivize the virtual human to look in specific directions: \\( r_{\\text{attention}} = \\cos \\langle \\vec{\\mathbf{v}}, \\vec{\\mathbf{a}} \\rangle \\), where \\( \\vec{\\mathbf{a}} \\) is the attention direction from the head joint to the viewing target. The reward function is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.076, + 0.681, + 0.47, + 0.697 + ], + "angle": 0, + "content": "\\[\nr _ {t} = r _ {c o n t.} + r _ {d i s t} + r _ {o r i} + r _ {a t t e n t i o n} + r _ {p e n e} + r _ {p o s e} + r _ {s u c c},\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.702, + 0.47, + 0.778 + ], + "angle": 0, + "content": "where \\( r_{cont} \\) enforces valid foot contact and minimizes foot skating; \\( r_{dist} \\) encourages reaching the target; \\( r_{ori} \\) aligns the body forward direction with the target; \\( r_{pene} \\) guides collision avoidance; \\( r_{pose} \\) reduces unrealistic human poses; and \\( r_{succ} \\) is a sparse reward when reaching the target." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.778, + 0.47, + 0.838 + ], + "angle": 0, + "content": "Episode Termination. To handle collisions beyond [119], we employ multiple termination signals to conclude an episode if the generated motion primitive \\(\\mathbf{X}_t\\) satisfies any of the following conditions:" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.838, + 0.402, + 0.853 + ], + "angle": 0, + "content": "- Success: The virtual human reached the target." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.853, + 0.469, + 0.868 + ], + "angle": 0, + "content": "- Penetration: The virtual human collides with the obstacle." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.868, + 0.469, + 0.899 + ], + "angle": 0, + "content": "- Timeout: The virtual human did not reach the target within the maximum timesteps." + }, + { + "type": "list", + "bbox": [ + 0.078, + 0.838, + 0.469, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.382, + 0.892, + 0.399 + ], + "angle": 0, + "content": "3.2. Training Collision-Avoiding Stochastic Policies" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.406, + 0.892, + 0.527 + ], + "angle": 0, + "content": "Algorithm. We use Proximal Policy Optimization [81] (PPO) to learn a generalizable mapping from various egocentric sensing and body configurations to CAMPs. Instead of extensive manual data collection for all possible input combinations, we leverage the stochastic nature of the PPO policy. Through exploration and sampling actions, the agent traverses the scene and generates varied egocentric sensing and body configurations, diversifying the training data." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.527, + 0.892, + 0.602 + ], + "angle": 0, + "content": "Instead of training each CAMP independently for every single step, we use PPO to train a sequence of CAMPs jointly in multi-step collision avoidance tasks. This approach can benefit choosing a more favorable CAMP which makes the subsequent action easier." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.603, + 0.892, + 0.724 + ], + "angle": 0, + "content": "Network. The network architecture is shown in Fig 2. The actor and critic network share a feature extraction trunk to encode the state \\(\\mathbf{s}_t\\): the motion seed \\((\\mathbf{X}_t^S\\) and \\(\\mathbf{X}_t^{S^D}\\)) and the egocentric sensing \\(\\mathcal{E}_t\\) are encoded using RNNs; the rest of scalar states are encoded using positional encoding [99]. The actor predicts a stochastic policy \\(\\mathbf{a}_t \\sim \\pi(\\mathbf{z}_t|\\mathbf{s}_t)\\) conditioned on the current state \\(\\mathbf{s}_t\\), where \\(\\mathbf{z}_t \\sim \\mathcal{N}(\\mu, \\Sigma)\\). \\(\\mu\\) and \\(\\Sigma\\) are the mean and variance of the learned action space." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.724, + 0.892, + 0.785 + ], + "angle": 0, + "content": "Objective Function. The objective function includes the policy surrogate \\( L^{CLIP} \\), the value function error term \\( L^{VF} \\) to evaluate the value prediction \\( V_{\\theta} \\), and an entropy bonus \\( L^{S} \\) to encourage exploration:" + }, + { + "type": "equation", + "bbox": [ + 0.592, + 0.795, + 0.8, + 0.813 + ], + "angle": 0, + "content": "\\[\nL = L ^ {C L I P} + c _ {1} L ^ {V F} + c _ {2} L ^ {S}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.826, + 0.892, + 0.902 + ], + "angle": 0, + "content": "where \\(c_{1}, c_{2}\\) are coefficients. See more details in Supp. Mat. RL Pre-training and Finetuning. Training in crowded scenes, e.g. Replica [91], requires additional steps. Because the action space \\(\\mathcal{A}\\) is an unbounded Gaussian, RL exploration while predicting reasonable human poses can" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "14500" + } + ], + [ + { + "type": "code_caption", + "bbox": [ + 0.08, + 0.091, + 0.468, + 0.107 + ], + "angle": 0, + "content": "Algorithm 1 Crowd motion synthesis with learned CAMPs" + }, + { + "type": "algorithm", + "bbox": [ + 0.094, + 0.111, + 0.47, + 0.261 + ], + "angle": 0, + "content": "Result: Multi-human locomotion w/ collision avoidance; \nInit: crowd size \\(C\\), marker seed for each human \\(\\mathbf{X}_c^S\\); \nfor step \\(\\leftarrow 1\\) to max_steps do env. finite horizon \nfor \\(c \\gets 1\\) to \\(C\\) do for each human update all locations with \\(\\{bbox(\\mathbf{X}_c^S)\\}_{c=1:C}\\) compute egocentric sensing \\(\\mathcal{E}_c\\); execute the action that maximizes the expected return, and produce one CAMP; end for \nend for" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.295, + 0.47, + 0.386 + ], + "angle": 0, + "content": "be challenging. We first pretrain the policy with a higher \\( r_{pene} \\) weight without penetration termination. After convergence, we finetune it with strict termination constraints using a signed distance field (SDF) for penetration detection. Please refer to Supp. Mat. for the formulation and weighting of each reward and training detail." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.403, + 0.424, + 0.419 + ], + "angle": 0, + "content": "3.3. Compositing Learned Motion Primitives" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.429, + 0.47, + 0.791 + ], + "angle": 0, + "content": "Although CAMPs are trained solely with static scenes, their direct application to dynamic settings is achieved by decomposing jointly trained CAMPs into individual motion primitives and re-compositing them. Our model demonstrates effective generalization by selecting the next best motion primitive from the learned CAMP library to maximize the expected return, provided that the egocentric sensing is updated with the most recent obstacle location at each timestep. Furthermore, our model is directly applicable to tasks involving complex interactions with other virtual humans. To synthesize crowd motion (Alg. 1), each virtual human employs the same policy to navigate and avoid others. To a specific virtual human, others are seen as dynamic obstacles, represented by body bounding boxes for avoidance. Acknowledging the inherent delay in human reactions when avoiding dynamic obstacles [4], agents take a single CAMP sequentially, instead of in parallel, i.e. the first agent generates its first CAMP and waits for others to complete their first CAMP before all agents move on to prepare their second CAMP. To ensure successful collision avoidance, the agent's egocentric sensing is updated before taking a new action. This composition of CAMPs synthesizes emergent multi-human behaviors without multi-agent RL algorithms (see Sec. 5.1), enhancing the generalization and scalability." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.813, + 0.421, + 0.83 + ], + "angle": 0, + "content": "4. Egocentric Synthetic Data Generation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.469, + 0.902 + ], + "angle": 0, + "content": "Synthesizing realistic egocentric perception-driven human motions (as detailed in Sec. 3) forms the foundation of simulating egocentric synthetic data. An overview of our egocentric data generation pipeline EgoGen, is shown in Fig. 3." + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.089, + 0.892, + 0.326 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.337, + 0.892, + 0.435 + ], + "angle": 0, + "content": "Figure 3. Overview of EgoGen. Through generative motion synthesis (Sec. 3), we further enhance egocentric data diversity by randomly sampling diverse body textures (ethnicity, gender) and 3D textured clothing through an automated clothing simulation pipeline (Sec. 4.2). With high-quality scenes and different egocentric cameras, we can render photorealistic egocentric synthetic data with rich and accurate ground truth annotations (Sec. 4.3)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.455, + 0.767, + 0.469 + ], + "angle": 0, + "content": "4.1. Embodied Camera Placement" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.48, + 0.892, + 0.616 + ], + "angle": 0, + "content": "Similar to existing AR devices, we use the head pose to define the egocentric viewing direction \\(\\vec{\\mathbf{v}}\\). Our development is based on Blender [11]. We use the SMPL-X [63] mesh to position the egocentric camera between the two eyeballs. The camera's viewing direction \\((\\vec{\\mathbf{v}})\\) is perpendicular to the plane determined by the two eye bones in the armature. We also support simulating multi-camera rigs as shown in Fig. 1. When the body moves (Sec. 3.3), we can synthesize egocentric videos with continuously updated camera poses." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.632, + 0.746, + 0.648 + ], + "angle": 0, + "content": "4.2. Body Texture and Clothing" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.657, + 0.891, + 0.718 + ], + "angle": 0, + "content": "To enhance EgoGen's synthetic data realism, we dress virtual humans using human textures and 3D clothing assets from BEDLAM [1, 8], including 50 male and 50 female skin albedo textures from seven ethnic groups." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Unlike prior works [8, 109] relying on unscalable commercial software for clothing dynamics simulation, we automate it for diverse synthesized motions and body shapes, minimizing manual effort. Each garment mesh is in a consistent rest pose, i.e., A-Pose (See Fig. 3 middle-left). For each motion sequence, we first repose it to match the body pose in the first frame using linear blend skinning. This involves initializing the clothing geometry by sampling pose and shape blend shapes, along with skinning weights from the nearest multiple SMPL-X vertices in A-Pose. Then we simulate upper and lower garments separately using a state-of-the-art clothing simulation network [24]." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "14501" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.089, + 0.47, + 0.146 + ], + "angle": 0, + "content": "Table 1. Evaluation of motion synthesis in scenes with moving obstacles, multiple humans, and path diversity. \\( \\downarrow \\) : lower is better; \\( \\uparrow \\) : higher is better. The best results in each scenario are in boldface. * denotes an improved version for fair comparison. (Sec. 5.1)" + }, + { + "type": "table", + "bbox": [ + 0.091, + 0.156, + 0.458, + 0.326 + ], + "angle": 0, + "content": "
EvaluationMetricsGAMMA* [117]DIMOS* [119]Ours
Mov. obs.SR (%) ↑9683100
Dist. (m) ↓0.290.550.06
Cont. ↑0.950.960.97
Pene-S. (%) ↓9.28.43.4
2 humansSR (%) ↑9588100
Dist. (m) ↓0.320.410.07
Cont. ↑0.960.980.97
Pene-H. ↓27.610.70
4 humansSR (%) ↑9270100
Dist. (m) ↓0.410.790.07
Cont. ↑0.940.950.96
Pene-H. ↓60.441.70
DiversitySR (%) ↑968497
Std Dev ↑0.9871.051.21
" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.339, + 0.327, + 0.354 + ], + "angle": 0, + "content": "4.3. Rendering and Annotations" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.362, + 0.47, + 0.543 + ], + "angle": 0, + "content": "EgoGen supports simulating diverse head-mounted devices with different camera models, such as fisheye and pinhole cameras. Given the camera's intrinsic parameters and relative poses within the camera rig, we can simulate AR devices like Project Aria glasses [50] and HoloLens [51], facilitating synthetic data generation for real-world applications. Camera extrinsic is determined by our generative human motion model. We use Blender [11] to render photorealistic egocentric image sequences with motion blur. We also render out a rich set of ground truth annotations, such as depth maps, surface normals, segmentation masks, world positions, optical flow, etc for egocentric perception tasks." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.557, + 0.21, + 0.573 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.582, + 0.469, + 0.672 + ], + "angle": 0, + "content": "We assess the motion quality, generalizability, and diversity of our motion model, highlighting its ability to generalize to unseen complex tasks and comparing it with recent baselines (Sec. 5.1). We evaluate our proposed egocentric sensing as a depth proxy for enhancing agent exploration (Sec. 5.2) and conduct ablation studies (Sec. 5.3)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.672, + 0.469, + 0.734 + ], + "angle": 0, + "content": "We further demonstrate the effectiveness of EgoGen on three egocentric computer vision tasks in Sec. 5.4, and 5.5. By incorporating synthesized egocentric images, we can enhance the performance of the state-of-the-art algorithms." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.742, + 0.352, + 0.757 + ], + "angle": 0, + "content": "5.1. Evaluation of Learned CAMPs" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.469, + 0.871 + ], + "angle": 0, + "content": "We assess CAMPs' generalizability in dynamic scenes, including scenes with moving obstacles and scenes with multiple individuals. In tests with moving obstacles, the obstacle blocks the person's path by moving between the person and the goal. In multiple human test scenes, lines from their starting and goal locations intersect in the middle, requiring solving human-human penetrations. See detail in Sup. Mat." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.469, + 0.901 + ], + "angle": 0, + "content": "In Tab. 1, we compare goal-reaching behaviors with two recent baselines: GAMMA [117] and DIMOS [119]. Base" + }, + { + "type": "table_caption", + "bbox": [ + 0.54, + 0.09, + 0.853, + 0.104 + ], + "angle": 0, + "content": "Table 2. Evaluation of egocentric sensing. (Sec. 5.2)" + }, + { + "type": "table", + "bbox": [ + 0.508, + 0.114, + 0.877, + 0.179 + ], + "angle": 0, + "content": "
Method (sensing range)SR (%) ↑Dist. (m) ↓
Local map [119] (0.8 m)780.35
Local map* (7 m)43.04
Egocentric sensing (ours) (7 m)950.12
" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.199, + 0.892, + 0.426 + ], + "angle": 0, + "content": "line methods use navigation meshes and path planning for static scenes, while CAMPs can autonomously avoid dynamic obstacles (Sec. 3.3). For fair comparison in dynamic scenes, we extend the baselines by updating navigation meshes and performing on-the-fly path planning at each time step. The tree-based search as in [117] is disabled for all the methods. Metrics: (1) \\( SR \\): Success rate for reaching the goal location within a \\( 0.3\\mathrm{m} \\) threshold. (2) Dist.: Average distance of the final pelvis location to the goal. (3) Cont.: The contact metric [117] that measures foot-floor contact and foot skating. (4) Pene-S.: Percentage of frames with detected human-scene penetration in moving obstacle scenes. (5) Pene-H.: Accurate human-human penetration evaluation metric using COAP [52] in multiple human scenes. Please refer to Supp. Mat. for metric details." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.429, + 0.892, + 0.566 + ], + "angle": 0, + "content": "CAMPs outperform the two baselines in dynamic scenarios with moving obstacles and multiple humans, exhibiting lower human-scene and human-human penetrations and a higher goal-reaching success rate. In multiple human scenarios, we observe that in the baselines, dynamically redoing path planning for each human independently can not effectively solve human-human penetration. In contrast, com-posable CAMPs can generalize well in dynamic settings without using multi-agent RL to synthesize crowd motions." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.569, + 0.892, + 0.69 + ], + "angle": 0, + "content": "We assess walking path diversity using the standard deviation of pelvis locations for the same start-target pairs in scenes with a single static box obstacle. As shown in Tab. 1 (Diversity), our approach does not require a pre-computed global path and allows agents to self-exlore without being constrained by predefined paths, achieving higher walking path diversity and success rate. This fosters diverse synthetic data generation via more diverse synthesized motion." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.709, + 0.792, + 0.725 + ], + "angle": 0, + "content": "5.2. Evaluation of Egocentric Sensing" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.735, + 0.892, + 0.901 + ], + "angle": 0, + "content": "We assess the exploration ability of our egocentric sensing \\(\\mathcal{E}_t\\) in Replica [91] scenes. In Tab. 2, we replace \\(\\mathcal{E}_t\\) with a local map [119] in our state \\(\\mathbf{s}_t\\), following their encoding method. Relying on local information can trap agents in local optima, e.g., walls beyond their sensing range, resulting in lower SR. Our egocentric sensing acts as a depth proxy, allowing the agent to avoid local optima, explore more effectively than local maps [119] or scandots [3], and achieve higher SR. In addition, our compact representation is more scalable as the sensing range increases, while quadratic local map growth can hinder the policy network's learning." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "14502" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.089, + 0.47, + 0.118 + ], + "angle": 0, + "content": "Table 3. Ablation studies. Note: in our observation, \\( \\| VP\\| _2 > 15 \\) indicates abnormal human poses. (Sec. 5.3)" + }, + { + "type": "table", + "bbox": [ + 0.086, + 0.128, + 0.454, + 0.205 + ], + "angle": 0, + "content": "
SR (%)↑||VP||2↓cos(▼,▲)↑
Egocentric depth813.640.049
No pretraining9028.770.918
No attention reward9012.260.891
Our policy9210.570.940
" + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.22, + 0.241, + 0.235 + ], + "angle": 0, + "content": "5.3. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.244, + 0.445, + 0.258 + ], + "angle": 0, + "content": "We compare our policy with several ablations in Tab. 3:" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.259, + 0.469, + 0.304 + ], + "angle": 0, + "content": "Egocentric depth: an ablation training an egocentric depth image-based policy without the depth sensing proxy. Egocentric depth images are encoded with a CNN;" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.305, + 0.469, + 0.35 + ], + "angle": 0, + "content": "No pretraining: an ablation training collision avoidance in crowded scenes with strict penetration termination directly; No attention reward: an ablation for the viewing direction." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.351, + 0.469, + 0.411 + ], + "angle": 0, + "content": "We assess pose naturalness with the maximum pose embedding norm encoded with VPoser [63] and evaluate the attention reward with the cosine similarity between the viewing direction \\(\\vec{\\mathbf{v}}\\) and the attention direction \\(\\vec{\\mathbf{a}}\\) (Sec. 3.1)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.412, + 0.469, + 0.579 + ], + "angle": 0, + "content": "Directly training RL with egocentric depth images is ineffective due to our high-dimensional action space, emphasizing the value of the compact egocentric sensing representation. Training agents with strict penetration constraints in crowded scenes directly can result in exploring unreasonable action subspaces, given its unbounded Gaussian nature, leading to unrealistic human poses, highlighting the effectiveness of our two-stage RL training scheme. Without the attention reward, the virtual human's capability to attend to a specific direction decreases. All ablation studies are evaluated in Replica. See visuals in Supp. Vid. and Supp. Mat." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.589, + 0.469, + 0.606 + ], + "angle": 0, + "content": "5.4. Mapping, Localization, and Tracking for HMD" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.613, + 0.469, + 0.779 + ], + "angle": 0, + "content": "Mapping and localization. LaMAR [76] is the first mapping and localization benchmark dataset for AR in large-scale scenes. Despite over a year of extensive data collection, the dataset still lacks exhaustive scene coverage, especially in large open spaces. EgoGen can let virtual humans explore large-scale scenes, render dense egocentric views, and build a more complete SfM map by extracting image feature points with SuperPoint [15] and matching images with SuperGlue [75]. Despite synthetic images being noisier due to scene quality, SuperGlue [75] matching can filter out noisy feature points and yield reliable matches." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.781, + 0.469, + 0.902 + ], + "angle": 0, + "content": "In Tab. 4, we evaluate EgoGen by assessing the localization recall at \\((1^{\\circ}, 10cm)\\) on the validation set in a lobby of \\(\\sim 120\\) sqm of the LaMAR CAB location. In addition, we report the number of triangulated 3D points (#P3D) and track length. EgoGen improves the 3D reconstruction by yielding more points for a slightly improved track length and also a significantly better localization performance compared to using the real data only. Ng et al. [58] augments mapping" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.089, + 0.892, + 0.146 + ], + "angle": 0, + "content": "Table 4. Mapping and localization evaluation. We augment LaMAR with the same amount of images (248 frames) and report the localization recall at \\((1^{\\circ},10cm)\\) on the validation set. EgoGen achieves the highest track length and recall. (Sec. 5.4)" + }, + { + "type": "table", + "bbox": [ + 0.509, + 0.156, + 0.876, + 0.22 + ], + "angle": 0, + "content": "
#P3D ↑Track length ↑Recall (%) ↑
LaMAR19297395.194666.9
Ng et al. [58]19377585.194074.9
EgoGen19361695.210576.7
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.227, + 0.892, + 0.255 + ], + "angle": 0, + "content": "Table 5. Egocentric camera tracking evaluation of models trained with and without synthetic data from EgoGen. (Sec. 5.4)" + }, + { + "type": "table", + "bbox": [ + 0.509, + 0.266, + 0.876, + 0.315 + ], + "angle": 0, + "content": "
Pose ↓Rotation ↓Transl (mm) ↓
Scratch1.830.741303
+ EgoGen pretrain1.670.621305
" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.327, + 0.892, + 0.493 + ], + "angle": 0, + "content": "images by perturbing real-world camera poses with noise, which may generate unrealistic camera poses (e.g., stuck in a wall or facing the ceiling), limiting egocentric localization effectiveness. Their method also assumes the availability of initial camera poses, which may not always be feasible. In contrast, EgoGen augments by virtual humans randomly exploring scenes. Our approach holds promise for creating AR mapping and localization datasets for digital twin scenes without manual data collection, providing enhanced privacy preservation, e.g. no need for anonymization. Refer to Supp. Mat. for visualization and implementation details." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.494, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Egocentric camera tracking. Egocentric camera tracking for HMD aims to yield device pose trajectories in 3D scenes given egocentric video observations. Recovering camera poses from monocular RGB videos using SLAM [93] is a challenging and ill-posed problem due to scale ambiguity. EgoEgo [43] leverages the knowledge of human motion to address the egocentric HMD tracking problem. Specifically, EgoEgo trains a neural network to infer the translation scaling and rotations from egocentric videos, which improves the HMD tracking performance. However, training this model requires jointly captured data of ground truth HMD trajectories and egocentric videos, which are costly to collect. We address this limitation by using EgoGen to synthesize quantities of egocentric videos with accurate camera trajectories to pretrain the model, which proves to improve the tracking performance on real data. We conduct experiments on the GIMO [121] dataset that contains \\(\\sim 200\\) short sequences of paired motion-video data in 19 scenes. Using EgoGen, we synthesize \\(\\sim 4k\\) sequences of human movements in their scenes and render corresponding egocentric videos using the same camera intrinsic as GIMO and the embodied camera placement described in Sec. 4.1. We also slightly perturb the camera placement location and orientation to simulate the diversity of how people wear HMDs in real data and avoid overfitting to one specific camera placement. We first pretrain the model with synthetic data generated by EgoGen, then finetune it on the real GIMO data." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "14503" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.471, + 0.182 + ], + "angle": 0, + "content": "Tab. 5 shows the egocentric camera tracking performances for models trained with and without synthetic data. Definitions of evaluation metrics can be found in Supp. Mat. The finetuned model benefits from EgoGen synthetic data and predicts more accurate camera poses compared to the model trained using real data only." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.194, + 0.469, + 0.21 + ], + "angle": 0, + "content": "5.5. Human Mesh Recovery from Egocentric views" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.218, + 0.47, + 0.55 + ], + "angle": 0, + "content": "Human mesh recovery (HMR) is the key to human behavior understanding from the egocentric view, thus crucial for applications in robotics and AR/VR. Given an egocentric RGB or depth image of a target subject, HMR aims to reconstruct the subject's 3D body pose and shape. However, acquiring and annotating real-world data is expensive, demanding, and time-consuming, with egocentric data being particularly scarce. EgoBody [115] is a recent egocentric dataset capturing two-people interactions, with egocentric depth/RGB frames annotated with SMPL-X body meshes. EgoBody provides \\(\\sim 180\\mathrm{k}\\) egocentric RGB frames, and merely \\(\\sim 23\\mathrm{k}\\) depth frames due to the low frame rate of the depth sensor, with \\(\\sim 90\\mathrm{k} / \\sim 10\\mathrm{k}\\) in the RGB/depth training set. Such limited data is insufficient to train a learning-based model from scratch. In contrast, with EgoGen, largescale synthetic egocentric data can be generated in a time-efficient way. We leverage EgoGen to generate quantities of training frames (300k RGB, 105k depth) of humans moving in EgoBody 3D scenes, rendered from the egocentric view, and annotated by SMPL-X parameters of the target subject. Specifically, RGB images are rendered with lifelike human body textures and 3D clothing, with random lighting." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.552, + 0.469, + 0.673 + ], + "angle": 0, + "content": "With the recent HMR regressor, ProHMR [41], we show that pre-training with our synthetic data from EgoGen enhances the existing method's capability to generalize on real-world scenarios. Evaluated on the real-world EgoBody test set, we compare two training schemes: (1) trained from scratch on the real-world EgoBody training set (\"-scratch\"), and (2) pre-trained on synthetic data from EgoGen and finetuned on the real-world EgoBody training set (\"-ft\")." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.674, + 0.469, + 0.868 + ], + "angle": 0, + "content": "HMR from depth. As no existing methods were proposed for depth-based HMR task, we adapt ProHMR [41] to the depth input by changing the channel number of the first convolution layer. To mimic real-world sensor noise, synthetic noise [27] is added to the rendered depth. G-MPJPE is additionally reported for depth-based HMR as depth images provide global information. As shown in Tab. 6, compared to the model trained only with a limited amount of real-world data (Depth-scratch), errors are significantly reduced for the model pre-trained with our large-scale synthetic data (Depth-ft), in terms of global translation (22.9% lower G-MPJPE), local pose (20.7% lower MJPE), and body shape (19.5% lower V2V)." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.469, + 0.901 + ], + "angle": 0, + "content": "HMR from RGB. For training with RGB images, we apply various data augmentation techniques similar to [8]." + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.09, + 0.892, + 0.145 + ], + "angle": 0, + "content": "Table 6. Evaluation of HMR on EgoBody test set. “*-scratch” denotes the model trained from scratch with the Egobody training set, and “*-ft” denotes the model pre-trained with EgoGen synthetic data. The units for all metrics are in mm. (Sec. 5.5)" + }, + { + "type": "table", + "bbox": [ + 0.508, + 0.149, + 0.876, + 0.218 + ], + "angle": 0, + "content": "
G-MPJPE ↓MPJPE ↓PA-MPJPE ↓V2V ↓
Depth-scratch117.782.254.1100.6
Depth-ft90.765.247.381.0
RGB-scratch-90.759.9102.1
RGB-ft-85.356.297.2
" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.228, + 0.892, + 0.288 + ], + "angle": 0, + "content": "Tab. 6 indicates that the RGB-based model pre-trained with large-scale synthetic data (\"RGB-ft\") also outperforms the model trained only on real-world data (\"RGB-scratch\"), for both body pose and shape accuracy." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.289, + 0.892, + 0.38 + ], + "angle": 0, + "content": "The enhanced performance highlights that EgoGen's synthetic data effectively compensates for the lack of real-world training data, boosting the performance of current methods when test on real-world data. We will release both of our synthetic EgoBody datasets. See dataset statistics, qualitative visualizations, and training details in Supp. Mat." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.393, + 0.769, + 0.408 + ], + "angle": 0, + "content": "6. Conclusion and Future Work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.418, + 0.892, + 0.584 + ], + "angle": 0, + "content": "We propose a novel egocentric synthetic data generation approach, EgoGen, that uses embodied sensors, a parametric body model, and a generative egocentric perception-driven human motion synthesis method to create egocentric training data with accurate and rich ground truth annotations. By integrating deep reinforcement learning and collision-avoiding motion primitives with egocentric depth proxy, EgoGen synthesizes robust human motion and emergent multi-agent behaviors. This paves the way to an efficient and scalable data generation solution that may have a profound impact on egocentric perception tasks." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.585, + 0.892, + 0.749 + ], + "angle": 0, + "content": "Human-scene interaction in EgoGen is currently coarse. We aim to extend the current method to simulate more detailed human motion driven by egocentric perception, such as hand manipulation, sitting, lying, etc, to facilitate more realistic egocentric synthetic data. We use fixed attention goals to model human attention. Predicting human intention through historical egocentric perception and synthesizing viewing directions based on predicted intention holds significant potential. Synthesizing gaze direction for predicting human intent is valuable but presently hampered by data requirements; we will revisit this when resources allow." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.892, + 0.84 + ], + "angle": 0, + "content": "We will explore many other egocentric vision tasks with EgoGen as this area grows rapidly such as social understanding and forecasting. EgoGen could benefit humanrobot interaction, e.g., our generative human motion model and lifelike human appearances can be integrated into [67] to close the sim2real gap for robotic agents further." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.892, + 0.9 + ], + "angle": 0, + "content": "Acknowledgements. This work is supported by the SNSF project grant 200021 204840 and the SDSC PhD fellowship. We sincerely thank Linfei Pan, Yunke Ao, Qianli Ma, and Maxime Raafat for the valuable discussions." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "14504" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.096, + 0.116, + 0.468, + 0.142 + ], + "angle": 0, + "content": "[1] Meshcapade GmbH, Tübingen, Germany. https://meshcapade.com, 2022.5" + }, + { + "type": "ref_text", + "bbox": [ + 0.096, + 0.146, + 0.468, + 0.227 + ], + "angle": 0, + "content": "[2] Mohammed Abdel-Wahab, Konrad Wenzel, and Dieter Fritsch. Efficient reconstruction of large unordered image datasets for high accuracy photogrammetric applications. In ISPRS Annals of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Melbourne, Australia. XXII ISPRS Congress, 2012. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.096, + 0.231, + 0.468, + 0.299 + ], + "angle": 0, + "content": "[3] Ananye Agarwal, Ashish Kumar, Jitendra Malik, and Deepak Pathak. Legged locomotion in challenging terrains using egocentric vision. In Conference on Robot Learning, CoRL 2022, 14-18 December 2022, Auckland, New Zealand, pages 403-415. PMLR, 2022. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.096, + 0.302, + 0.468, + 0.341 + ], + "angle": 0, + "content": "[4] M. Pilar Aivar, Eli Brenner, and Jeroen B. J. Smeets. Avoiding moving obstacles. Experimental Brain Research, 190 (3):251-264, 2008. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.096, + 0.345, + 0.468, + 0.413 + ], + "angle": 0, + "content": "[5] Hiroyasu Akada, Jian Wang, Soshi Shimada, Masaki Takahashi, Christian Theobalt, and Vladislav Golyanik. Unrealego: A new dataset for robust egocentric 3d human motion capture. In European Conference on Computer Vision (ECCV), 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.096, + 0.416, + 0.468, + 0.442 + ], + "angle": 0, + "content": "[6] Apple. ARKit. https://developer.apple.com/arkit/, 2017.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.096, + 0.446, + 0.468, + 0.514 + ], + "angle": 0, + "content": "[7] Sonia Baltodano, Srinath Sibi, Nikolas Martelaro, Nikhil Gowda, and Wendy Ju. The rroads platform: a real road autonomous driving simulator. In Proceedings of the 7th International Conference on Automotive User Interfaces and Interactive Vehicular Applications, pages 281-288, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.096, + 0.517, + 0.468, + 0.584 + ], + "angle": 0, + "content": "[8] Michael J. Black, Priyanka Patel, Joachim Tesch, and Jinlong Yang. BEDLAM: A synthetic dataset of bodies exhibiting detailed lifelike animated motion. In Proceedings IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR), 2023. 2, 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.096, + 0.588, + 0.468, + 0.656 + ], + "angle": 0, + "content": "[9] Federica Bogo, Angjoo Kanazawa, Christoph Lassner, Peter Gehler, Javier Romero, and Michael J. Black. Keep it SMPL: Automatic estimation of 3D human pose and shape from a single image. In European Conference on Computer Vision, pages 561-578, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.659, + 0.468, + 0.685 + ], + "angle": 0, + "content": "[10] Simon Clavet. Motion matching and the road to next-gen animation. In Proc. of GDC, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.688, + 0.468, + 0.728 + ], + "angle": 0, + "content": "[11] Blender Online Community. Blender - a 3D modelling and rendering package. Blender Foundation, Stichting Blender Foundation, Amsterdam, 2018. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.732, + 0.468, + 0.772 + ], + "angle": 0, + "content": "[12] Erwin Coumans and Yunfei Bai. Pybullet, a python module for physics simulation for games, robotics and machine learning. http://pybullet.org, 2016-2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.776, + 0.468, + 0.87 + ], + "angle": 0, + "content": "[13] Dima Damen, Hazel Doughty, Giovanni Maria Farinella, Sanja Fidler, Antonino Furnari, Evangelos Kazakos, Davide Moltisanti, Jonathan Munro, Toby Perrett, Will Price, and Michael Wray. The epic-kitchens dataset: Collection, challenges and baselines. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 43(11):4125-4141, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.874, + 0.468, + 0.901 + ], + "angle": 0, + "content": "[14] Andrew J. Davison. Real-time simultaneous localisation and mapping with a single camera. In ICCV, 2003. 3" + }, + { + "type": "list", + "bbox": [ + 0.088, + 0.116, + 0.468, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.093, + 0.891, + 0.134 + ], + "angle": 0, + "content": "[15] Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. Superpoint: Self-supervised interest point detection and description, 2018. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.137, + 0.891, + 0.178 + ], + "angle": 0, + "content": "[16] Alexey Dosovitskiy, German Ros, Felipe Codevilla, Antonio Lopez, and Vladlen Koltun. Carla: An open urban driving simulator, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.181, + 0.892, + 0.277 + ], + "angle": 0, + "content": "[17] Salehe Erfanian Ebadi, Saurav Dhakad, Sanjay Vishwakarma, Chunpu Wang, You-Cyuan Jhang, Maciek Chociej, Adam Crespi, Alex Thaman, and Sujoy Ganguly. Psp-hdri+: A synthetic dataset generator for pre-training of human-centric computer vision models. In First Workshop on Pre-training: Perspectives, Pitfalls, and Paths Forward at ICML 2022, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.281, + 0.891, + 0.348 + ], + "angle": 0, + "content": "[18] Matteo Fabbri, Fabio Lanzi, Simone Calderara, Andrea Palazzi, Roberto Vezzani, and Rita Cucchiara. Learning to detect and track visible and occluded body joints in a virtual world. In European Conference on Computer Vision (ECCV), 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.352, + 0.891, + 0.42 + ], + "angle": 0, + "content": "[19] Alexander Fiannaca, Ilias Apostolopoulous, and Eelke Folmer. Headlock: A wearable navigation aid that helps blind cane users traverse large open spaces. In Proceedings of the 16th international ACM SIGACCESS conference on Computers & accessibility, pages 19-26, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.423, + 0.891, + 0.491 + ], + "angle": 0, + "content": "[20] Zahra Ghodsi, Siva Kumar Sastry Hari, Iuri Frosio, Timothy Tsai, Alejandro Troccoli, Stephen W. Keckler, Siddharth Garg, and Anima Anandkumar. Generating and characterizing scenarios for safety testing of autonomous vehicles, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.496, + 0.888, + 0.521 + ], + "angle": 0, + "content": "[21] Google. ARCore. https://developers.google.com/ar/, 2018.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.525, + 0.891, + 0.551 + ], + "angle": 0, + "content": "[22] Venu Madhav Govindu. Combining two-view constraints for motion estimation. In CVPR, 2001. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.555, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[23] Kristen Grauman, Andrew Westbury, Eugene Byrne, Zachary Chavis, Antonino Furnari, Rohit Girdhar, Jackson Hamburger, Hao Jiang, Miao Liu, Xingyu Liu, Miguel Martin, Tushar Nagarajan, Iija Radosavovic, Santhosh Kumar Ramakrishnan, Fiona Ryan, Jayant Sharma, Michael Wray, Mengmeng Xu, Eric Zhongcong Xu, Chen Zhao, Siddhant Bansal, Dhruv Batra, Vincent Cartillier, Sean Crane, Tien Do, Morrie Doulaty, Akshay Erapalli, Christoph Feichtenhofer, Adriano Fragomeni, Qichen Fu, Christian Fuegen, Abrham Gebreselasie, Cristina Gonzalez, James Hillis, Xuhua Huang, Yifei Huang, Wenqi Jia, Weslie Khoo, Jachym Kolar, Satwik Kottur, Anurag Kumar, Federico Landini, Chao Li, Yanghao Li, Zhenqiang Li, Karttikeya Mangalam, Raghava Modhugu, Jonathan Munro, Tullie Murrell, Takumi Nishiyasu, Will Price, Paola Ruiz Puentes, Merey Ramazanova, Leda Sari, Kiran Somasundaram, Audrey Southerland, Yusuke Sugano, Ruijie Tao, Minh Vo, Yuchen Wang, Xindi Wu, Takuma Yagi, Yunyi Zhu, Pablo Arbelaez, David Crandall, Dima Damen, Giovanni Maria Farinella, Bernard Ghanem, Vamsi Krishna Ithapu, C. V. Jawahar, Hanbyul Joo, Kris Kitani, Haizhou Li, Richard Newcombe, Aude Oliva, Hyun Soo Park, James M. Rehg, Yoichi Sato, Jianbo Shi, Mike Zheng Shou, Antonio Torralba, Lorenzo Torresani, Mingfei Yan, and Jitendra Malik. Ego4d: Around the World in 3,000 Hours of Egocentric" + }, + { + "type": "list", + "bbox": [ + 0.511, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "14505" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.118, + 0.092, + 0.468, + 0.12 + ], + "angle": 0, + "content": "Video. In IEEE/CVF Computer Vision and Pattern Recognition (CVPR), 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.122, + 0.47, + 0.205 + ], + "angle": 0, + "content": "[24] Artur Grigorev, Michael J. Black, and Otmar Hilliges. HOOD: hierarchical graphs for generalized modelling of clothing dynamics. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2023, Vancouver, BC, Canada, June 17-24, 2023, pages 16965-16974. IEEE, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.207, + 0.471, + 0.29 + ], + "angle": 0, + "content": "[25] Vladimir Guzov, Aymen Mir, Torsten Sattler, and Gerard Pons-Moll. Human poseitioning system (hps): 3d human pose estimation and self-localization in large scenes from body-mounted sensors. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4318-4329, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.089, + 0.292, + 0.47, + 0.36 + ], + "angle": 0, + "content": "[26] Norbert Haala, Michael Cramer, Florian Weimer, and Martin Trittler. Performance test on uav-based photogrammetric data collection. The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, 2012. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.089, + 0.362, + 0.468, + 0.403 + ], + "angle": 0, + "content": "[27] Ankur Handa, Thomas Whelan, John McDonald, and Andrew J Davison. A benchmark for rgb-d visual odometry, 3d reconstruction and slam. ICRA, 2014. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.405, + 0.469, + 0.488 + ], + "angle": 0, + "content": "[28] Albert Haque, Boya Peng, Zelun Luo, Alexandre Alahi, Serena Yeung, and Li Fei-Fei. Towards viewpoint invariant 3d human pose estimation. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11–14, 2016, Proceedings, Part I 14, pages 160–177. Springer, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.49, + 0.469, + 0.558 + ], + "angle": 0, + "content": "[29] Mohamed Hassan, Duygu Ceylan, Ruben Villegas, Jun Saito, Jimei Yang, Yi Zhou, and Michael Black. Stochastic scene-aware motion prediction. In Proceedings of the International Conference on Computer Vision 2021, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.561, + 0.469, + 0.615 + ], + "angle": 0, + "content": "[30] Mohamed Hassan, Yunrong Guo, Tingwu Wang, Michael Black, Sanja Fidler, and Xue Bin Peng. Synthesizing physical character-scene interactions. arXiv preprint arXiv:2302.00883, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.618, + 0.469, + 0.659 + ], + "angle": 0, + "content": "[31] Daniel Holden, Taku Komura, and Jun Saito. Phase-functional neural networks for character control. ACM Trans. Graph., 36(4):1-13, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.661, + 0.469, + 0.702 + ], + "angle": 0, + "content": "[32] Daniel Holden, Oussama Kanoun, Maksym Perepichka, and Tiberiu Popa. Learned motion matching. ACM Transactions on Graphics (TOG), 39(4):53-1, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.704, + 0.469, + 0.758 + ], + "angle": 0, + "content": "[33] Yuhe Jin, Dmytro Mishkin, Anastasiia Mishchuk, Jiri Matas, Pascal Fua, Kwang Moo Yi, and Eduard Trulls. Image Matching across Wide Baselines: From Paper to Practice. IJCV, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.761, + 0.469, + 0.844 + ], + "angle": 0, + "content": "[34] Neena Kamath. Announcing Azure Spatial Anchors for collaborative, cross-platform mixed reality apps. https://azure.microsoft.com/en-us/blog/announcing-azure-spatial-anchors-for-collaborative-cross-platform-mixed-reality-apps/, 2019.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.846, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[35] Angjoo Kanazawa, Michael J Black, David W Jacobs, and Jitendra Malik. End-to-end recovery of human shape and pose. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 7122-7131, 2018. 3" + }, + { + "type": "list", + "bbox": [ + 0.088, + 0.092, + 0.471, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.093, + 0.892, + 0.134 + ], + "angle": 0, + "content": "[36] Alex Kendall, Matthew Grimes, and Roberto Cipolla. PoseNet: A Convolutional Network for Real-Time 6-DoF Camera Relocalization. In ICCV, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.136, + 0.892, + 0.188 + ], + "angle": 0, + "content": "[37] Georg Klein and David Murray. Parallel tracking and mapping for small ar workspaces. In IEEE and ACM International Symposium on Mixed and Augmented Reality, 2007. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.19, + 0.892, + 0.259 + ], + "angle": 0, + "content": "[38] Muhammed Kocabas, Chun-Hao P. Huang, Otmar Hilliges, and Michael J. Black. PARE: Part attention regressor for 3D human body estimation. In Proceedings International Conference on Computer Vision (ICCV), pages 11127-11137. IEEE, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.261, + 0.892, + 0.328 + ], + "angle": 0, + "content": "[39] Nikos Kolotouros, Georgios Pavlakos, Michael J Black, and Kostas Daniilidis. Learning to reconstruct 3D human pose and shape via model-fitting in the loop. In Proceedings of the IEEE International Conference on Computer Vision, pages 2252-2261, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.33, + 0.892, + 0.371 + ], + "angle": 0, + "content": "[40] Nikos Kolotouros, Georgios Pavlakos, and Kostas Dani-ilidis. Convolutional mesh regression for single-image human shape reconstruction. In CVPR, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.373, + 0.892, + 0.44 + ], + "angle": 0, + "content": "[41] Nikos Kolotouros, Georgios Pavlakos, Dinesh Jayaraman, and Kostas Daniilidis. Probabilistic modeling for human mesh recovery. In Proceedings of the IEEE/CVF international conference on computer vision, pages 11605-11614, 2021. 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.442, + 0.892, + 0.511 + ], + "angle": 0, + "content": "[42] Lucas Kovar, Michael Gleicher, and Frédéric H. Pighin. Motion graphs. In International Conference on Computer Graphics and Interactive Techniques, SIGGRAPH 2008, Los Angeles, California, USA, August 11-15, 2008, Classes, pages 51:1-51:10. ACM, 2008. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.512, + 0.892, + 0.58 + ], + "angle": 0, + "content": "[43] Jiaman Li, C. Karen Liu, and Jiajun Wu. Ego-body pose estimation via ego-head pose estimation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2023, Vancouver, BC, Canada, June 17-24, 2023, pages 17142-17151. IEEE, 2023. 1, 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.581, + 0.892, + 0.664 + ], + "angle": 0, + "content": "[44] Zhihao Li, Jianzhuang Liu, Zhensong Zhang, Songcen Xu, and Youliang Yan. CLIFF: Carrying location information in full frames into human pose and shape estimation. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part V, pages 590-606. Springer, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.665, + 0.892, + 0.705 + ], + "angle": 0, + "content": "[45] Hung Yu Ling, Fabio Zinno, George Cheng, and Michiel van de Panne. Character controllers using motion vaes. ACM Trans. Graph., 39(4), 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.707, + 0.892, + 0.774 + ], + "angle": 0, + "content": "[46] Miao Liu, Dexin Yang, Yan Zhang, Zhaopeng Cui, James M Rehg, and Siyu Tang. 4d human body capture from egocentric video via 3d scene grounding. In 2021 International Conference on 3D Vision (3DV), pages 930-939. IEEE, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.776, + 0.892, + 0.831 + ], + "angle": 0, + "content": "[47] Zhengyi Luo, Ryo Hachiuma, Ye Yuan, Shun Iwase, and Kris M Kitani. Kinematics-guided reinforcement learning for object-aware 3d ego-pose estimation. arXiv preprint arXiv:2011.04837, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.832, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[48] Naureen Mahmood, Nima Ghorbani, Nikolaus F Troje, Gerard Pons-Moll, and Michael J Black. Amass: Archive of motion capture as surface shapes. In Proceedings of the IEEE/CVF international conference on computer vision, pages 5442-5451, 2019. 3" + }, + { + "type": "list", + "bbox": [ + 0.511, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "14506" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.092, + 0.47, + 0.175 + ], + "angle": 0, + "content": "[49] Angel Martínez-González, Michael Villamizar, Olivier Canévet, and Jean-Marc Odobez. Residual pose: A decoupled approach for depth-based 3d human pose estimation. In 2020 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 10313-10318. IEEE, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.177, + 0.468, + 0.204 + ], + "angle": 0, + "content": "[50] Meta. Project Aria Glasses. https://www.projectaria.com/, 2023.6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.206, + 0.468, + 0.234 + ], + "angle": 0, + "content": "[51] Microsoft. HoloLens 2. https://www.microsoft.com/en-us/hololens, 2019.6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.235, + 0.47, + 0.304 + ], + "angle": 0, + "content": "[52] Marko Mihajlovic, Shunsuke Saito, Aayush Bansal, Michael Zollhoefer, and Siyu Tang. COAP: Compositional articulated occupancy of people. In Proceedings IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.305, + 0.47, + 0.347 + ], + "angle": 0, + "content": "[53] Aymen Mir, Xavier Puig, Angjoo Kanazawa, and Gerard Pons-Moll. Generating continual human motion in diverse 3d scenes. CoRR, abs/2304.02061, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.349, + 0.47, + 0.431 + ], + "angle": 0, + "content": "[54] Michael Montemerlo, Nicholas Roy, and Sebastian Thrun. Perspectives on standardization in mobile robot programming: the Carnegie mellon navigation (CARMEN) toolkit. In 2003 IEEE/RSJ International Conference on Intelligent Robots and Systems, Las Vegas, Nevada, USA, October 27 - November 1, 2003, pages 2436-2441. IEEE, 2003. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.433, + 0.47, + 0.502 + ], + "angle": 0, + "content": "[55] Gyeongsik Moon, Ju Yong Chang, and Kyoung Mu Lee. V2v-posenet: Voxel-to-voxel prediction network for accurate 3d hand and human pose estimation from a single depth map. In Proceedings of the IEEE conference on computer vision and pattern Recognition, pages 5079-5088, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.504, + 0.47, + 0.544 + ], + "angle": 0, + "content": "[56] Etienne Mouragnon, Maxime Lhuillier, Michel Dhome, Fabien Dekeyser, and Patrick Sayd. Real time localization and 3d reconstruction. In CVPR, 2006. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.546, + 0.47, + 0.616 + ], + "angle": 0, + "content": "[57] Evonne Ng, Donglai Xiang, Hanbyul Joo, and Kristen Grauman. You2me: Inferring body pose in egocentric video via first and second person interactions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9890-9900, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.617, + 0.47, + 0.659 + ], + "angle": 0, + "content": "[58] Tony Ng, Adrian Lopez-Rodriguez, Vassileios Balntas, and Krystian Mikolajczyk. Reassessing the limitations of cnn methods for camera pose regression, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.66, + 0.468, + 0.702 + ], + "angle": 0, + "content": "[59] Niantic. Niantic Expands Developer Platform and AR Tools with Niantic Lightship. https://nianticlabs.com/news/lightship/, 2021.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.703, + 0.468, + 0.731 + ], + "angle": 0, + "content": "[60] David Nister, Oleg Naroditsky, and James Bergen. Visual odometry. In CVPR, 2004. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.732, + 0.468, + 0.802 + ], + "angle": 0, + "content": "[61] Xiaqing Pan, Nicholas Charron, Yongqian Yang, Scott Peters, Thomas Whelan, Chen Kong, Omkar M. Parkhi, Richard A. Newcombe, and Carl Yuheng Ren. Aria digital twin: A new benchmark dataset for egocentric 3d machine perception. CoRR, abs/2306.06362, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.803, + 0.468, + 0.871 + ], + "angle": 0, + "content": "[62] Fabio Pardo, Arash Tavakoli, Vitaly Levdik, and Petar Kormushev. Time limits in reinforcement learning. In Proceedings of the 35th International Conference on Machine Learning, pages 4045-4054, Stockholm, Sweden, 2018. PMLR. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.872, + 0.468, + 0.901 + ], + "angle": 0, + "content": "[63] Georgios Pavlakos, Vasileios Choutas, Nima Ghorbani, Timo Bolkart, Ahmed A. A. Osman, Dimitrios Tzionas," + }, + { + "type": "list", + "bbox": [ + 0.087, + 0.092, + 0.47, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.541, + 0.092, + 0.892, + 0.148 + ], + "angle": 0, + "content": "and Michael J. Black. Expressive body capture: 3D hands, face, and body from a single image. In Proceedings IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), pages 10975-10985, 2019. 3, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.149, + 0.892, + 0.204 + ], + "angle": 0, + "content": "[64] Xue Bin Peng, Ze Ma, Pieter Abbeel, Sergey Levine, and Angjoo Kanazawa. Amp: Adversarial motion priors for stylized physics-based character control. ACM Transactions on Graphics (TOG), 40(4):1-20, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.205, + 0.892, + 0.245 + ], + "angle": 0, + "content": "[65] Mathis Petrovich, Michael J. Black, and Gül Varol. Action-conditioned 3d human motion synthesis with transformer vae, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.247, + 0.892, + 0.287 + ], + "angle": 0, + "content": "[66] Marc Pollefeys, Luc Van Gool, Maarten Vergauwen, Frank Verbiest, Kurt Cornelis, Jan Tops, and Reinhard Koch. Visual modeling with a hand-held camera. IJCV, 2004. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.288, + 0.892, + 0.413 + ], + "angle": 0, + "content": "[67] Xavier Puig, Eric Undersander, Andrew Szot, Mikael Dallaire Cote, Tsung-Yen Yang, Ruslan Partsey, Ruta Desai, Alexander William Clegg, Michal Hlavac, So Yeon Min, Vladimir Vondrus, Théophile Gervet, Vincent-Pierre Berges, John M. Turner, Oleksandr Maksymets, Zsolt Kira, Mrinal Kalakrishnan, Jitendra Malik, Devendra Singh Chaplot, Unnat Jain, Dhruv Batra, Akshara Rai, and Roozbeh Mottaghi. Habitat 3.0: A co-habitat for humans, avatars and robots. CoRR, abs/2310.13724, 2023. 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.414, + 0.892, + 0.469 + ], + "angle": 0, + "content": "[68] Albert Pumarola, Jordi Sanchez, Gary Choi, Alberto Sanfeliu, and Francesc Moreno-Noguer. 3DPeople: Modeling the Geometry of Dressed Humans. In International Conference in Computer Vision (ICCV), 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.47, + 0.892, + 0.537 + ], + "angle": 0, + "content": "[69] Morgan Quigley, Ken Conley, Brian Gerkey, Josh Faust, Tully Foote, Jeremy Leibs, Rob Wheeler, Andrew Y Ng, et al. Ros: an open-source robot operating system. In ICRA workshop on open source software, page 5. Kobe, Japan, 2009. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.539, + 0.892, + 0.607 + ], + "angle": 0, + "content": "[70] Umer Rafi, Juergen Gall, and Bastian Leibe. A semantic occlusion model for human pose estimation from a single depth image. In Proceedings of the IEEE conference on computer vision and pattern recognition workshops, pages 67-74, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.608, + 0.892, + 0.663 + ], + "angle": 0, + "content": "[71] Tilman Reinhardt. Google Visual Positioning Service. https://ai.googleblog.com/2019/02/using-global-localization-to-improve.html, 2019.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.664, + 0.892, + 0.719 + ], + "angle": 0, + "content": "[72] Davis Rempe, Tolga Birdal, Aaron Hertzmann, Jimei Yang, Srinath Sridhar, and Leonidas J. Guibas. Humor: 3d human motion model for robust pose estimation. In International Conference on Computer Vision (ICCV), 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.72, + 0.892, + 0.775 + ], + "angle": 0, + "content": "[73] Davis Rempe, Zhengyi Luo, Xue Bin Peng, Ye Yuan, Kris Kitani, Karsten Kreis, Sanja Fidler, and Or Litany. Trace and pace: Controllable pedestrian animation via guided trajectory diffusion, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.776, + 0.892, + 0.859 + ], + "angle": 0, + "content": "[74] Guodong Rong, Byung Hyun Shin, Hadi Tabatabaee, Qiang Lu, Steve Lemke, Martiš Možeiko, Eric Boise, Geehoon Uhm, Mark Gerow, Shalin Mehta, et al. Lgsvl simulator: A high fidelity simulator for autonomous driving. In 2020 IEEE 23rd International conference on intelligent transportation systems (ITSC), pages 1-6. IEEE, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.86, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[75] Paul-Edouard Sarlin, Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. Superglue: Learning feature matching with graph neural networks, 2020. 7" + }, + { + "type": "list", + "bbox": [ + 0.511, + 0.092, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "14507" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.093, + 0.47, + 0.161 + ], + "angle": 0, + "content": "[76] Paul-Edouard Sarlin, Mihai Dusmanu, Johannes L. Schonberger, Pablo Speciale, Lukas Gruber, Viktor Larsson, Ondrej Miksik, and Marc Pollefeys. LaMAR: Benchmarking Localization and Mapping for Augmented Reality. In ECCV, 2022. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.162, + 0.472, + 0.245 + ], + "angle": 0, + "content": "[77] Torsten Sattler, Will Maddern, Carl Toft, Akihiko Torii, Lars Hammarstrand, Erik Stenberg, Daniel Safari, Masatoshi Okutomi, Marc Pollefeys, Josef Sivic, Fredrik Kahl, and Tomas Pajdla. Benchmarking 6DoF outdoor visual localization in changing conditions. In CVPR, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.246, + 0.472, + 0.314 + ], + "angle": 0, + "content": "[78] Manolis Savva, Abhishek Kadian, Oleksandr Maksymets, Yili Zhao, Erik Wijmans, Bhavana Jain, Julian Straub, Jia Liu, Vladlen Koltun, Jitendra Malik, Devi Parikh, and Dhruv Batra. Habitat: A platform for embodied ai research, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.315, + 0.472, + 0.343 + ], + "angle": 0, + "content": "[79] Johannes Lutz Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In CVPR, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.343, + 0.472, + 0.384 + ], + "angle": 0, + "content": "[80] Johannes L. Schonberger, Hans Hardmeier, Torsten Sattler, and Marc Pollefeys. Comparative evaluation of handcrafted and learned local features. In CVPR, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.385, + 0.472, + 0.426 + ], + "angle": 0, + "content": "[81] John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Radford, and Oleg Klimov. Proximal policy optimization algorithms, 2017. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.426, + 0.472, + 0.482 + ], + "angle": 0, + "content": "[82] Akash Sengupta, Ignas Budvytis, and Roberto Cipolla. Synthetic training for accurate 3d human pose and shape estimation in the wild. In British Machine Vision Conference (BMVC), 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.483, + 0.472, + 0.525 + ], + "angle": 0, + "content": "[83] Takaaki Shiratori, Hyun Soo Park, Leonid Sigal, Yaser Sheikh, and Jessica K. Hodgins. Motion capture from body-mounted cameras. ACM Trans. Graph., 30(4):31, 2011. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.525, + 0.472, + 0.607 + ], + "angle": 0, + "content": "[84] Jamie Shotton, Ross Girshick, Andrew Fitzgibbon, Toby Sharp, Mat Cook, Mark Finocchio, Richard Moore, Pushmeet Kohli, Antonio Criminisi, Alex Kipman, et al. Efficient human pose estimation from single depth images. IEEE transactions on pattern analysis and machine intelligence, 35(12):2821-2840, 2012. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.608, + 0.472, + 0.663 + ], + "angle": 0, + "content": "[85] Jamie Shotton, Ben Glocker, Christopher Zach, Shahram Izadi, Antonio Criminisi, and Andrew Fitzgibbon. Scene Coordinate Regression Forests for Camera Relocalization in RGB-D Images. In CVPR, 2013. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.664, + 0.472, + 0.732 + ], + "angle": 0, + "content": "[86] Jamie Shotton, Toby Sharp, Alex Kipman, Andrew Fitzgibbon, Mark Finocchio, Andrew Blake, Mat Cook, and Richard Moore. Real-time human pose recognition in parts from single depth images. Commun. ACM, 56(1):116-124, 2013. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.733, + 0.472, + 0.774 + ], + "angle": 0, + "content": "[87] Noah Snavely, Steven M. Seitz, and Richard Szeliski. Photo tourism: exploring photo collections in 3d. ACM Trans. Graph., 25(3):835-846, 2006. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.775, + 0.472, + 0.831 + ], + "angle": 0, + "content": "[88] Kihyuk Sohn, Honglak Lee, and Xinchen Yan. Learning structured output representation using deep conditional generative models. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.831, + 0.472, + 0.872 + ], + "angle": 0, + "content": "[89] Sebastian Starke, He Zhang, Taku Komura, and Jun Saito. Neural state machine for character-scene interactions. ACM Trans. Graph., 38(6):209-1, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.873, + 0.472, + 0.903 + ], + "angle": 0, + "content": "[90] Sebastian Starke, Ian Mason, and Taku Komura. Deepphase: Periodic autoencoders for learning motion phase" + }, + { + "type": "list", + "bbox": [ + 0.087, + 0.093, + 0.472, + 0.903 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.541, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "manifolds. ACM Transactions on Graphics (TOG), 41(4): 1-13, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.122, + 0.894, + 0.261 + ], + "angle": 0, + "content": "[91] Julian Straub, Thomas Whelan, Lingni Ma, Yufan Chen, Erik Wijmans, Simon Green, Jakob J. Engel, Raul Mur-Artal, Carl Ren, Shobhit Verma, Anton Clarkson, Mingfei Yan, Brian Budge, Yajie Yan, Xiaqing Pan, June Yon, Yuyang Zou, Kimberly Leon, Nigel Carter, Jesus Briales, Tyler Gillingham, Elias Mueggler, Luis Pesqueira, Manolis Savva, Dhruv Batra, Hauke M. Strasdat, Renzo De Nardi, Michael Goesele, Steven Lovegrove, and Richard Newcombe. The Replica dataset: A digital replica of indoor spaces. arXiv preprint arXiv:1906.05797, 2019. 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.262, + 0.894, + 0.387 + ], + "angle": 0, + "content": "[92] Andrew Szot, Alex Clegg, Eric Undersander, Erik Wijmans, Yili Zhao, John Turner, Noah Maestre, Mustafa Mukadam, Devendra Chaplot, Oleksandr Maksymets, Aaron Gokaslan, Vladimir Vondrus, Sameer Dharur, Franziska Meier, Wojciech Galuba, Angel Chang, Zsolt Kira, Vladlen Koltun, Jitendra Malik, Manolis Savva, and Dhruv Batra. Habitat 2.0: Training home assistants to rearrange their habitat. In Advances in Neural Information Processing Systems (NeurIPS), 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.388, + 0.892, + 0.43 + ], + "angle": 0, + "content": "[93] Zachary Teed and Jia Deng. DROID-SLAM: Deep Visual SLAM for Monocular, Stereo, and RGB-D Cameras. Advances in neural information processing systems, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.431, + 0.892, + 0.473 + ], + "angle": 0, + "content": "[94] Guy Tevet, Sigal Raab, Brian Gordon, Yonatan Shafir, Daniel Cohen-Or, and Amit H Bermano. Human motion diffusion model. arXiv preprint arXiv:2209.14916, 2022.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.474, + 0.892, + 0.531 + ], + "angle": 0, + "content": "[95] Emanuel Todorov, Tom Erez, and Yuval Tassa. Mujoco: A physics engine for model-based control. In 2012 IEEE/RSJ International Conference on Intelligent Robots and Systems, pages 5026-5033, 2012. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.532, + 0.892, + 0.588 + ], + "angle": 0, + "content": "[96] Emanuel Todorov, Tom Erez, and Yuval Tassa. Mujoco: A physics engine for model-based control. In 2012 IEEE/RSJ International Conference on Intelligent Robots and Systems, pages 5026-5033. IEEE, 2012. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.589, + 0.892, + 0.657 + ], + "angle": 0, + "content": "[97] Denis Tome, Patrick Peluse, Lourdes Agapito, and Hernan Badino. xr-egopose: Egocentric 3d human pose from an hmd camera. In Proceedings of the IEEE International Conference on Computer Vision, pages 7728-7738, 2019. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.66, + 0.892, + 0.729 + ], + "angle": 0, + "content": "[98] Denis Tome, Thiemo Alldieck, Patrick Peluse, Gerard Pons-Moll, Lourdes Agapito, Hernan Badino, and Fernando De la Torre. Selfpose: 3d egocentric pose estimation from a headset mounted camera. arXiv preprint arXiv:2011.01519, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.73, + 0.892, + 0.8 + ], + "angle": 0, + "content": "[99] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Advances in neural information processing systems, pages 5998-6008, 2017. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.801, + 0.892, + 0.871 + ], + "angle": 0, + "content": "[100] Jian Wang, Diogo Luvizon, Weipeng Xu, Lingjie Liu, Kripasindhu Sarkar, and Christian Theobalt. Scene-aware egocentric 3d human pose estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13031-13040, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.872, + 0.892, + 0.903 + ], + "angle": 0, + "content": "[101] Keze Wang, Shengfu Zhai, Hui Cheng, Xiaodan Liang, and Liang Lin. Human pose estimation from depth images via" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.903 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "14508" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.118, + 0.093, + 0.468, + 0.134 + ], + "angle": 0, + "content": "inference embedded multi-task learning. In Proceedings of the 24th ACM international conference on Multimedia, pages 1227-1236, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.136, + 0.468, + 0.233 + ], + "angle": 0, + "content": "[102] Xin Wang, Taein Kwon, Mahdi Rad, Bowen Pan, Ishani Chakraborty, Sean Andrist, Dan Bohus, Ashley Feniello, Bugra Tekin, Felipe Vieira Frujeri, Neel Joshi, and Marc Pollefeys. Holoassist: an egocentric human interaction dataset for interactive ai assistants in the real world. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 20270-20281, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.234, + 0.468, + 0.317 + ], + "angle": 0, + "content": "[103] Erik Wijmans, Abhishek Kadian, Ari Morcos, Stefan Lee, Irfan Essa, Devi Parikh, Manolis Savva, and Dhruv Batra. DD-PPO: learning near-perfect pointgoal navigators from 2.5 billion frames. In 8th International Conference on Learning Representations, ICLR 2020, Addis Ababa, Ethiopia, April 26-30, 2020. OpenReview.net, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.319, + 0.468, + 0.346 + ], + "angle": 0, + "content": "[104] Kyle Wilson and Noah Snavely. Robust global translations with 1dsfm. In ECCV, 2014. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.349, + 0.468, + 0.389 + ], + "angle": 0, + "content": "[105] Jungdam Won, Deepak Gopinath, and Jessica Hodgins. Physics-based character controllers using conditional vaes. ACM Trans. Graph., 41(4), 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.392, + 0.468, + 0.46 + ], + "angle": 0, + "content": "[106] Erroll Wood, Tadas Baltrusaitis, Charlie Hewitt, Sebastian Dziadzio, Thomas J Cashman, and Jamie Shotton. Fake it till you make it: face analysis in the wild using synthetic data alone. In Proceedings of the IEEE/CVF international conference on computer vision, pages 3681-3691, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.462, + 0.468, + 0.544 + ], + "angle": 0, + "content": "[107] Fu Xiong, Boshen Zhang, Yang Xiao, Zhiguo Cao, Taidong Yu, Joey Tianyi Zhou, and Junsong Yuan. A2j: Anchor-to-joint regression network for 3d articulated pose estimation from a single depth image. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 793-802, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.547, + 0.468, + 0.629 + ], + "angle": 0, + "content": "[108] Mengyuan Yan, Gen Li, Yilin Zhu, and Jeannette Bohg. Learning topological motion primitives for knot planning. In IEEE/RSJ International Conference on Intelligent Robots and Systems, IROS 2020, Las Vegas, NV, USA, October 24, 2020 - January 24, 2021, pages 9457-9464. IEEE, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.632, + 0.468, + 0.701 + ], + "angle": 0, + "content": "[109] Zhitao Yang, Zhongang Cai, Haiyi Mei, Shuai Liu, Zhaoxi Chen, Weiye Xiao, Yukun Wei, Zhongfei Qing, Chen Wei, Bo Dai, Wayne Wu, Chen Qian, Dahua Lin, Ziwei Liu, and Lei Yang. Synbody: Synthetic dataset with layered human models for 3d human perception and modeling, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.703, + 0.468, + 0.757 + ], + "angle": 0, + "content": "[110] Heyuan Yao, Zhenhua Song, Baoquan Chen, and Libin Liu. Controlvae: Model-based learning of generative controllers for physics-based characters. ACM Trans. Graph., 41(6), 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.76, + 0.468, + 0.814 + ], + "angle": 0, + "content": "[111] Mao Ye, Xianwang Wang, Ruigang Yang, Liu Ren, and Marc Pollefeys. Accurate 3d pose estimation from a single depth image. In 2011 International Conference on Computer Vision, pages 731-738. IEEE, 2011. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.817, + 0.468, + 0.87 + ], + "angle": 0, + "content": "[112] Vickie Ye, Georgios Pavlakos, Jitendra Malik, and Angjoo Kanazawa. Decoupling human and camera motion from videos in the wild. arXiv preprint arXiv:2302.12827, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.873, + 0.468, + 0.901 + ], + "angle": 0, + "content": "[113] Ye Yuan and Kris Kitani. Ego-posed estimation and forecasting as real-time pd control. In Proceedings of the" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.093, + 0.468, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.541, + 0.093, + 0.891, + 0.12 + ], + "angle": 0, + "content": "IEEE/CVF International Conference on Computer Vision, pages 10082-10092, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.122, + 0.892, + 0.177 + ], + "angle": 0, + "content": "[114] Mingyuan Zhang, Zhongang Cai, Liang Pan, Fangzhou Hong, Xinying Guo, Lei Yang, and Ziwei Liu. Motiondiffuse: Text-driven human motion generation with diffusion model. arXiv preprint arXiv:2208.15001, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.179, + 0.892, + 0.247 + ], + "angle": 0, + "content": "[115] Siwei Zhang, Qianli Ma, Yan Zhang, Zhiyin Qian, Taein Kwon, Marc Pollefeys, Federica Bogo, and Siyu Tang. Egobody: Human body shape and motion of interacting people from head-mounted devices. In European conference on computer vision (ECCV), 2022. 1, 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.249, + 0.892, + 0.317 + ], + "angle": 0, + "content": "[116] Siwei Zhang, Qianli Ma, Yan Zhang, Sadegh Aliakbarian, Darren Cosker, and Siyu Tang. Probabilistic human mesh recovery in 3d scenes from egocentric views. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.32, + 0.892, + 0.373 + ], + "angle": 0, + "content": "[117] Yan Zhang and Siyu Tang. The wanderings of odysseus in 3d scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20481-20491, 2022. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.376, + 0.892, + 0.431 + ], + "angle": 0, + "content": "[118] Yan Zhang, Michael J Black, and Siyu Tang. We are more than our joints: Predicting how 3d bodies move. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3372-3382, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.433, + 0.892, + 0.487 + ], + "angle": 0, + "content": "[119] Kaifeng Zhao, Yan Zhang, Shaofei Wang, Thabo Beeler, and Siyu Tang. Synthesizing diverse human motions in 3d indoor scenes. In International conference on computer vision (ICCV), 2023. 2, 3, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.489, + 0.892, + 0.571 + ], + "angle": 0, + "content": "[120] Yuhang Zhao, Elizabeth Kupferstein, Brenda Veronica Castro, Steven Feiner, and Shiri Azenkot. Designing ar visualizations to facilitate stair navigation for people with low vision. In Proceedings of the 32nd annual ACM symposium on user interface software and technology, pages 387-402, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.574, + 0.892, + 0.656 + ], + "angle": 0, + "content": "[121] Yang Zheng, Yanchao Yang, Kaichun Mo, Jiaman Li, Tao Yu, Yebin Liu, C Karen Liu, and Leonidas J Guibas. Gimo: Gaze-informed human motion prediction in context. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XIII, pages 676-694. Springer, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.658, + 0.892, + 0.713 + ], + "angle": 0, + "content": "[122] Ziwen Zhuang, Zipeng Fu, Jianren Wang, Christopher Atkeson, Soren Schwertfeger, Chelsea Finn, and Hang Zhao. Robot parkour learning. In Conference on Robot Learning (CoRL), 2023. 2, 3" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.892, + 0.713 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "14509" + } + ] +] \ No newline at end of file diff --git a/2024/EgoGen_ An Egocentric Synthetic Data Generator/717ad579-fa9d-4ce5-996d-77c046c2cdbf_origin.pdf b/2024/EgoGen_ An Egocentric Synthetic Data Generator/717ad579-fa9d-4ce5-996d-77c046c2cdbf_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..4824eb44fafce9ed13852e47f23f1ff350f3a289 --- /dev/null +++ b/2024/EgoGen_ An Egocentric Synthetic Data Generator/717ad579-fa9d-4ce5-996d-77c046c2cdbf_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8db2f75662081e350e6f9b3d715d726bf1ce8a5f24367f72fda15591aeee2aca +size 3832470 diff --git a/2024/EgoGen_ An Egocentric Synthetic Data Generator/full.md b/2024/EgoGen_ An Egocentric Synthetic Data Generator/full.md new file mode 100644 index 0000000000000000000000000000000000000000..83b3c92bbe3d1365ba1ff122871ad0de4dc3ba14 --- /dev/null +++ b/2024/EgoGen_ An Egocentric Synthetic Data Generator/full.md @@ -0,0 +1,393 @@ +# EgoGen: An Egocentric Synthetic Data Generator + +Gen Li $^{1}$ Kaifeng Zhao $^{1}$ Siwei Zhang $^{1}$ Xiaozhong Lyu $^{1}$ +Mihai Dusmanu $^{2}$ Yan Zhang $^{1}$ Marc Pollefeys $^{1,2}$ Siyu Tang $^{1}$ $^{1}$ ETH Zürich + $^{2}$ Microsoft +https://ego-gen.github.io/ + +![](images/20f96637edb2d3d562a65def7cff75f240c545540ad7a1c0a8763394d5c27659.jpg) +RGB +RGB with Motion Blur +Depth + +![](images/73eaa166df6ddb80755c410a54a4b1ab7074c0acff1935e1445efd19d432d69c.jpg) +Surface Normal +Figure 1. EgoGen: a scalable synthetic data generation system for egocentric perception tasks, with rich multi-modal data and accurate annotations. We simulate camera rigs for head-mounted devices (HMDs) and render from the perspective of the camera wearer with various sensors. Top to bottom: middle and right camera sensors in the rig. Left to right: photo-realistic RGB image, RGB with simulated motion blur, depth map, surface normal, segmentation mask, and world position for fisheye cameras widely used in HMDs. + +![](images/5d1f2404d006586ba9eb281ec93f4ef6684b3f107eb264300d5fb6b0cd250652.jpg) +Segmentation Mask + +![](images/a97948112dd8e0dedfe94b08c70c361c3b1c45a52a4fc02f4560d6042536f524.jpg) +World Position + +# Abstract + +Understanding the world in first-person view is fundamental in Augmented Reality (AR). This immersive perspective brings dramatic visual changes and unique challenges compared to third-person views. Synthetic data has empowered third-person-view vision models, but its application to embodied egocentric perception tasks remains largely unexplored. A critical challenge lies in simulating natural human movements and behaviors that effectively steer the embodied cameras to capture a faithful egocentric representation of the 3D world. To address this challenge, we introduce EgoGen, a new synthetic data generator that can produce accurate and rich ground-truth training data for egocentric perception tasks. At the heart of EgoGen is a novel human motion synthesis model that directly leverages egocentric visual inputs of a virtual human to sense the 3D environment. Combined with collision-avoiding motion primitives and a two-stage reinforcement learning approach, our motion synthesis model offers a closed-loop so + +lution where the embodied perception and movement of the virtual human are seamlessly coupled. Compared to previous works, our model eliminates the need for a pre-defined global path, and is directly applicable to dynamic environments. Combined with our easy-to-use and scalable data generation pipeline, we demonstrate EgoGen's efficacy in three tasks: mapping and localization for head-mounted cameras, egocentric camera tracking, and human mesh recovery from egocentric views. EgoGen will be fully open-sourced, offering a practical solution for creating realistic egocentric training data and aiming to serve as a useful tool for egocentric computer vision research. + +# 1. Introduction + +The analysis of visual input from front-facing egocentric cameras is crucial for applications that benefit from a first-person perspective, mirroring the natural human experience [23, 43, 115]. AR devices, for instance, can utilize this viewpoint to amplify user immersion. Such cameras can also cater to individual preferences, providing custom + +visual assistance for those with impaired vision [19, 120]. + +Despite its potential, egocentric perception faces challenges, primarily due to the scarcity of labeled data. Although datasets like Ego4D [23], ADT [61], EpicKitchen [13] and HoloAssist [102] exist, creating such datasets with rich and accurate annotations is costly and raises privacy concerns [115]. Alternatively, using graphics techniques to render synthetic multi-modal visual data has proven to be cost-effective and successful in training deep learning models, such as 3D human body estimation [8] and facial landmark detection [106]. + +Creating egocentric synthetic data is challenging because egocentric cameras capture the complex interplay of body movements and the environment from the camera wearer's viewpoint. Modeling the intricate details and variations in human behavior presents a significant challenge. + +To tackle this problem, we introduce EgoGen, an egocentric synthetic data generation approach that simulates data from embodied sensors, i.e., front-facing cameras in head-mounted devices (HMD). While the ultimate goal is to simulate human behaviors that are indistinguishable from reality, in this work, we focus on creating virtual humans (i.e., camera wearers) that can explore and avoid obstacles in the 3D world that is not only complex and dynamic but could potentially include other moving virtual humans. + +Specifically, we propose a novel generative human motion model. Our key insight is that body movement and embodied perception should be seamlessly coupled. As William Gibson aptly stated, "We see in order to move; we move in order to see," our egocentric perception is crucial for identifying obstacles, navigating in an environment, and planning actions. Our body movements are not solely a response to visual stimuli; they also change our egocentric perception. Therefore, the key idea of our motion model is to enable virtual humans to see their environment with egocentric visual inputs and respond accordingly by learning a policy to control a set of collision-avoiding motion primitives (CAMPs) that are composable for synthesizing long-term, diverse human motions. Due to the unbounded and high-dimensional latent action space of our generative motion primitive model, direct policy training through rendered egocentric images is often unstable [122]. Therefore, we propose a two-stage reinforcement learning scheme using an efficient egocentric visual proxy to couple egocentric visual cues and body movements seamlessly. In addition, we use an "attention" reward to incentivize egocentric perception behaviors, i.e., looking in the desired direction. + +Empirical results showcase the benefits of our egocentric perception-driven motion framework, which does not require a pre-calculated walking path in 3D scenes as in [29, 53, 119]. Instead, it empowers virtual humans to perceive the environment from their own viewpoint, enabling them to navigate, circumvent obstacles, and plan move + +ments to reach the destination. Moreover, our model generalizes well to dynamic environments, even with training limited to static settings. By training virtual humans independently using CAMPs, our method synthesizes emergent multi-human behaviors without relying on multi-agent reinforcement learning algorithms. Egocentric visual cues are essential to build exploratory and generalizable motion models that unify navigational planning and movement control in complex and dynamic environments. + +Building upon CAMPs, we further create a scalable data generation pipeline for EgoGen that outfits virtual humans with clothing, automates cloth animation, and integrates 3D assets from various sources. We validate EgoGen's efficacy across three egocentric perception tasks. The high-quality synthetic data with precise ground truth annotations consistently improve the performance of state-of-the-art methods. In summary, the contributions of this work are: + +1. We introduce EgoGen, a generative and scalable synthetic data generation approach specifically tailored for egocentric perception tasks. +2. We introduce novel motion primitives based on egocentric visual cues, enabling diverse and realistic human motion synthesis in 3D scenes. These primitives empower virtual humans to handle complex scenarios, such as dynamic environments and crowd motion without relying on multi-agent reinforcement learning. +3. EgoGen enables us to augment existing real-world egocentric datasets with synthetic images. Quantitative results demonstrate enhanced performance in state-of-the-art algorithms on mapping and localization for head-mounted cameras, egocentric camera tracking, and human mesh recovery from egocentric views. + +# 2. Related Work + +Human-Related Simulators and Synthetic Data. Previous works primarily focus on simulating robots [54, 69, 78, 92, 95] and autonomous cars [7, 16, 20, 74]. While some incorporated animated digital humans, like in [5, 20, 67], these efforts relied on pre-recorded motion sequences. Rendering images of people to train perception models has been widely studied such as [5, 17, 18, 43, 68, 82, 86]. In particular, [97] offers large-scale synthetic data for egocentric camera wearer pose estimation but relies on mocap data, lacking realistic and spontaneous interactions with the digital world. In contrast, EgoGen closes the gap in egocentric synthetic data generation for head-mounted devices. Please refer to Supp. Mat. for detailed comparisons. + +Human Motion Synthesis. Generating high-fidelity human motions and interactions with 3D scenes is widely studied in graphics [10, 31, 32, 42, 89, 90]. While they can generate high-quality motion, it's usually deterministic. Synthesizing physically plausible human motions has been extensively studied [12, 30, 64, 73, 96, 105]. However, + +they struggle with generalization to different body shapes. For example, [73] explicitly created 2048 humanoids to improve body shape generalizability. Time series models [65, 94, 114] synthesize the stochastic motions of diverse people well. However, in [94, 114], their generated motion sequences have limited lengths and human-scene interactions are not explicitly considered. Autoregressive methods [45, 72, 117, 119] can produce perpetual motions. In particular, [117] can generalize to diverse body shapes and synthesize long-term human motions. + +Our egocentric perception-driven motion synthesis model is closely related to [117, 119], but distinguishes itself w.r.t.: (1) Enabling virtual humans to explore using egocentric visual cues, without predefined paths. (2) Synthesizing egocentric perception behaviors beyond locomotion, e.g., looking in certain directions. (3) Handling dynamic environments and multi-agent behavior without re-training. + +Mapping and Localization for AR. Localization and mapping from images is a long-standing problem known as: Photogrammetry [2, 26]; Structure-from-Motion (SfM) [22, 66, 79, 87, 104]; Simultaneous Localization and Mapping (SLAM) [14, 37, 56, 60]. Researchers have worked to make SLAM amenable for edge hand-held or head-mounted devices [6, 21, 37]. Cloud-based services like Google's Visual Positioning System [71], Niantic's Lightship [59], and Microsoft's Azure Spatial Anchors [34] have made visual localization and mapping more accessible. Benchmarking efforts have arisen for small-scale AR scenarios [36, 85], touristic landmarks [33, 80], and large-scale AR-device based localization [76, 77] to evaluate these systems. + +Egocentric Human Pose Estimation. Estimating 3D bodies from RGB images is widely studied from third-person views [9, 35, 38-41, 44, 63], and egocentric views [25, 46, 47, 57, 83, 97, 98, 100, 112, 113, 116], mostly requiring expensive real-world data paired with ground truth annotations. Besides RGB images, depth images offer explicit 3D information, mitigating scale and shape ambiguity, with the potential to enable broader AR/VR applications. However, depth-based methods, especially for the egocentric view, are underexplored. Most existing works [28, 49, 55, 70, 84, 101, 107, 111] predict 3D body skeletons without expressive body meshes, struggling with challenges like severe body truncations and scene occlusions typical in egocentric views. Such limited attention mainly stems from the scarcity of data, as obtaining high-quality human mesh annotations for real-world depth images is labor-intensive. + +# 3. Ego-Sensing Driven Motion Synthesis + +To close the loop for the interdependence between egocentric synthetic image data and human motion synthesis, we use deep reinforcement learning (RL), integrating egocentric vision cues to synthesize human motions as described in + +Sec. 3.1 and 3.2. Subsequently, we extend learned policies to generate emergent multi-agent behaviors, as in Sec. 3.3. + +# 3.1. Ego-Sensing Driven Motion Primitives + +Generating realistic egocentric data requires diverse and lifelike human motion synthesis. In this work, we consider arguably the most common everyday behaviors: navigating towards goals with egocentric perception while avoiding collisions with obstacles and people in dynamic 3D scenes. + +Overview. Following recent literature [45, 110, 117, 119], we employ deep RL to train control policies on learned latent spaces that characterize natural human motions. However, unlike these previous works that only consider simple static scenes, we leverage egocentric perception and propose collision-avoiding motion primitives (CAMPs) to enable virtual humans to self-exlore and navigate in a dynamic environment. Specifically, CAMPs are trained jointly to produce collision-free motion sequences. At each timestep $t$ , the agent observes the state $\mathbf{s}_t$ , performs an action $\mathbf{a}_t$ , and receives a reward $r_t = r(\mathbf{s}_t, \mathbf{a}_t, \mathbf{s}_{t+1})$ , where $\mathbf{s}_{t+1}$ represents the next state of the environment due to $\mathbf{a}_t$ . + +Egocentric Sensing As Depth Proxy. We aim to sample actions given by a policy to synthesize realistic human motions. Egocentric perception-driven motion synthesis should arguably use egocentric vision as input. However, depth rendering is costly and RL requires billions of samples to converge [3, 103]. Besides, directly training RL with visual data can be unstable [122]. We thereby use a cheap-to-compute egocentric sensing $\mathcal{E}_t$ as a proxy for depth images as illustrated in Fig. 2. $N$ rays are cast evenly from the midpoint of two eyeballs, i.e., the location of the egocentric camera. The field of view $[\theta_{min},\theta_{max}]$ , centered on the 2D projection of the viewing direction $\vec{\nu}$ , limits the agent's perception to the front area. Rays stop at collisions, with collision detection in 2D. See more details in Supp. Mat. + +Agent Representation. The agent is a virtual human represented by an SMPL-X mesh [63]. We further compact the body representation by selecting $M = 67$ body surface markers $\mathbf{x} \in \mathbb{R}^{M \times 3}$ on the mesh following [118]. + +Environment. Inspired by [108], we aim to learn a library of composable CAMPs. We implement a finite-horizon environment based on the generative motion primitive model from [117]. Specifically, a motion primitive is defined as a 0.5-second motion clip containing $T = 20$ frames in the canonical coordinate, and each frame contains a single agent representation. The primitive model $\mathcal{P}$ is based on the C-VAE framework [88], which takes the first $T_{s} = 2$ frames as the condition, and models a conditional probability of the next $T - T_{s}$ frames. Compared to [117] trained on the AMASS dataset [48] with many sport motion sequences, we train $\mathcal{P}$ using the SAMP dataset [29], which focuses on daily activities, better suited for HMD use cases. Our action space $\mathcal{A}$ is the pretrained 128D latent space of $\mathcal{P}$ , and the + +![](images/13ff6411c240f09adaa375b54aefc78e76a61979c69f8352c6ed6db654396e22.jpg) +Figure 2. Policy network architecture. We learn a generalizable mapping from motion seed body markers $\mathbf{X}_t^S$ , marker directions $\mathbf{X}_t^{S^D}$ , egocentric sensing $\mathcal{E}_t$ , and distance to the target $d_t$ to CAMPs. The policy learns a stochastic collision avoiding action space to predict future body markers $\mathbf{X}_t^F$ . For illustration purposes, we visualize only one frame of $\mathbf{X}_t^S$ and $\mathcal{E}_t$ . See Sec. 3.1 and 3.2 for details. + +action $\mathbf{a}_t$ can be randomly sampled from $\mathcal{A}$ . + +With the input of a random action $\mathbf{a}_t\in \mathbb{R}^{128}$ and a motion seed $\mathbf{X}_t^S = [\mathbf{x}_t^0,\mathbf{x}_t^{T_s - 1}]$ (history frames), $\mathcal{P}$ predicts future frames $\mathbf{X}_t^F = [\mathbf{x}_t^{T_s},\dots ,\mathbf{x}_t^{T - 1}]$ of the current motion primitive $\mathbf{X}_t = [\mathbf{X}_t^S,\mathbf{X}_t^F ]\in \mathbb{R}^{T\times M\times 3}$ , which represents a short sequence of human motion spanning $0.5\mathrm{s}$ : + +$$ +\mathbf {X} _ {t} ^ {F} = \mathcal {P} (\mathbf {X} _ {t} ^ {S}, \mathbf {a} _ {t}) +$$ + +State. To preserve Markov property [62], the state is defined as $\mathbf{s}_t = \{\mathbf{X}_t^S, \mathbf{X}_t^{S^D}, \mathcal{E}_t, d_t, \tau_t\}$ , in which $\mathbf{X}_t^{S^D} \in \mathbb{R}^{T_s \times M \times 3}$ denotes the normalized direction of each marker seed to the target, $\mathcal{E}_t \in \mathbb{R}^{T_s \times N}$ denotes the egocentric sensing depth proxy, $d_t$ denotes the distance from the pelvis to the target, and $\tau_t$ denotes the remaining time. See Fig. 2. + +Reward. To synthesize egocentric perception behaviors, we use an "attention reward" to incentivize the virtual human to look in specific directions: $r_{\text{attention}} = \cos \langle \vec{\mathbf{v}}, \vec{\mathbf{a}} \rangle$ , where $\vec{\mathbf{a}}$ is the attention direction from the head joint to the viewing target. The reward function is defined as: + +$$ +r _ {t} = r _ {c o n t.} + r _ {d i s t} + r _ {o r i} + r _ {a t t e n t i o n} + r _ {p e n e} + r _ {p o s e} + r _ {s u c c}, +$$ + +where $r_{cont}$ enforces valid foot contact and minimizes foot skating; $r_{dist}$ encourages reaching the target; $r_{ori}$ aligns the body forward direction with the target; $r_{pene}$ guides collision avoidance; $r_{pose}$ reduces unrealistic human poses; and $r_{succ}$ is a sparse reward when reaching the target. + +Episode Termination. To handle collisions beyond [119], we employ multiple termination signals to conclude an episode if the generated motion primitive $\mathbf{X}_t$ satisfies any of the following conditions: + +- Success: The virtual human reached the target. +- Penetration: The virtual human collides with the obstacle. +- Timeout: The virtual human did not reach the target within the maximum timesteps. + +# 3.2. Training Collision-Avoiding Stochastic Policies + +Algorithm. We use Proximal Policy Optimization [81] (PPO) to learn a generalizable mapping from various egocentric sensing and body configurations to CAMPs. Instead of extensive manual data collection for all possible input combinations, we leverage the stochastic nature of the PPO policy. Through exploration and sampling actions, the agent traverses the scene and generates varied egocentric sensing and body configurations, diversifying the training data. + +Instead of training each CAMP independently for every single step, we use PPO to train a sequence of CAMPs jointly in multi-step collision avoidance tasks. This approach can benefit choosing a more favorable CAMP which makes the subsequent action easier. + +Network. The network architecture is shown in Fig 2. The actor and critic network share a feature extraction trunk to encode the state $\mathbf{s}_t$ : the motion seed $(\mathbf{X}_t^S$ and $\mathbf{X}_t^{S^D}$ ) and the egocentric sensing $\mathcal{E}_t$ are encoded using RNNs; the rest of scalar states are encoded using positional encoding [99]. The actor predicts a stochastic policy $\mathbf{a}_t \sim \pi(\mathbf{z}_t|\mathbf{s}_t)$ conditioned on the current state $\mathbf{s}_t$ , where $\mathbf{z}_t \sim \mathcal{N}(\mu, \Sigma)$ . $\mu$ and $\Sigma$ are the mean and variance of the learned action space. + +Objective Function. The objective function includes the policy surrogate $L^{CLIP}$ , the value function error term $L^{VF}$ to evaluate the value prediction $V_{\theta}$ , and an entropy bonus $L^{S}$ to encourage exploration: + +$$ +L = L ^ {C L I P} + c _ {1} L ^ {V F} + c _ {2} L ^ {S} +$$ + +where $c_{1}, c_{2}$ are coefficients. See more details in Supp. Mat. RL Pre-training and Finetuning. Training in crowded scenes, e.g. Replica [91], requires additional steps. Because the action space $\mathcal{A}$ is an unbounded Gaussian, RL exploration while predicting reasonable human poses can + +Algorithm 1 Crowd motion synthesis with learned CAMPs +Result: Multi-human locomotion w/ collision avoidance; +Init: crowd size $C$ , marker seed for each human $\mathbf{X}_c^S$ ; +for step $\leftarrow 1$ to max_steps do env. finite horizon +for $c \gets 1$ to $C$ do for each human update all locations with $\{bbox(\mathbf{X}_c^S)\}_{c=1:C}$ compute egocentric sensing $\mathcal{E}_c$ ; execute the action that maximizes the expected return, and produce one CAMP; end for +end for + +be challenging. We first pretrain the policy with a higher $r_{pene}$ weight without penetration termination. After convergence, we finetune it with strict termination constraints using a signed distance field (SDF) for penetration detection. Please refer to Supp. Mat. for the formulation and weighting of each reward and training detail. + +# 3.3. Compositing Learned Motion Primitives + +Although CAMPs are trained solely with static scenes, their direct application to dynamic settings is achieved by decomposing jointly trained CAMPs into individual motion primitives and re-compositing them. Our model demonstrates effective generalization by selecting the next best motion primitive from the learned CAMP library to maximize the expected return, provided that the egocentric sensing is updated with the most recent obstacle location at each timestep. Furthermore, our model is directly applicable to tasks involving complex interactions with other virtual humans. To synthesize crowd motion (Alg. 1), each virtual human employs the same policy to navigate and avoid others. To a specific virtual human, others are seen as dynamic obstacles, represented by body bounding boxes for avoidance. Acknowledging the inherent delay in human reactions when avoiding dynamic obstacles [4], agents take a single CAMP sequentially, instead of in parallel, i.e. the first agent generates its first CAMP and waits for others to complete their first CAMP before all agents move on to prepare their second CAMP. To ensure successful collision avoidance, the agent's egocentric sensing is updated before taking a new action. This composition of CAMPs synthesizes emergent multi-human behaviors without multi-agent RL algorithms (see Sec. 5.1), enhancing the generalization and scalability. + +# 4. Egocentric Synthetic Data Generation + +Synthesizing realistic egocentric perception-driven human motions (as detailed in Sec. 3) forms the foundation of simulating egocentric synthetic data. An overview of our egocentric data generation pipeline EgoGen, is shown in Fig. 3. + +![](images/06018b31962397ed8eb422593026e3d987b0a8165b470871706a4442b584c09e.jpg) +Figure 3. Overview of EgoGen. Through generative motion synthesis (Sec. 3), we further enhance egocentric data diversity by randomly sampling diverse body textures (ethnicity, gender) and 3D textured clothing through an automated clothing simulation pipeline (Sec. 4.2). With high-quality scenes and different egocentric cameras, we can render photorealistic egocentric synthetic data with rich and accurate ground truth annotations (Sec. 4.3). + +# 4.1. Embodied Camera Placement + +Similar to existing AR devices, we use the head pose to define the egocentric viewing direction $\vec{\mathbf{v}}$ . Our development is based on Blender [11]. We use the SMPL-X [63] mesh to position the egocentric camera between the two eyeballs. The camera's viewing direction $(\vec{\mathbf{v}})$ is perpendicular to the plane determined by the two eye bones in the armature. We also support simulating multi-camera rigs as shown in Fig. 1. When the body moves (Sec. 3.3), we can synthesize egocentric videos with continuously updated camera poses. + +# 4.2. Body Texture and Clothing + +To enhance EgoGen's synthetic data realism, we dress virtual humans using human textures and 3D clothing assets from BEDLAM [1, 8], including 50 male and 50 female skin albedo textures from seven ethnic groups. + +Unlike prior works [8, 109] relying on unscalable commercial software for clothing dynamics simulation, we automate it for diverse synthesized motions and body shapes, minimizing manual effort. Each garment mesh is in a consistent rest pose, i.e., A-Pose (See Fig. 3 middle-left). For each motion sequence, we first repose it to match the body pose in the first frame using linear blend skinning. This involves initializing the clothing geometry by sampling pose and shape blend shapes, along with skinning weights from the nearest multiple SMPL-X vertices in A-Pose. Then we simulate upper and lower garments separately using a state-of-the-art clothing simulation network [24]. + +Table 1. Evaluation of motion synthesis in scenes with moving obstacles, multiple humans, and path diversity. $\downarrow$ : lower is better; $\uparrow$ : higher is better. The best results in each scenario are in boldface. * denotes an improved version for fair comparison. (Sec. 5.1) + +
EvaluationMetricsGAMMA* [117]DIMOS* [119]Ours
Mov. obs.SR (%) ↑9683100
Dist. (m) ↓0.290.550.06
Cont. ↑0.950.960.97
Pene-S. (%) ↓9.28.43.4
2 humansSR (%) ↑9588100
Dist. (m) ↓0.320.410.07
Cont. ↑0.960.980.97
Pene-H. ↓27.610.70
4 humansSR (%) ↑9270100
Dist. (m) ↓0.410.790.07
Cont. ↑0.940.950.96
Pene-H. ↓60.441.70
DiversitySR (%) ↑968497
Std Dev ↑0.9871.051.21
+ +# 4.3. Rendering and Annotations + +EgoGen supports simulating diverse head-mounted devices with different camera models, such as fisheye and pinhole cameras. Given the camera's intrinsic parameters and relative poses within the camera rig, we can simulate AR devices like Project Aria glasses [50] and HoloLens [51], facilitating synthetic data generation for real-world applications. Camera extrinsic is determined by our generative human motion model. We use Blender [11] to render photorealistic egocentric image sequences with motion blur. We also render out a rich set of ground truth annotations, such as depth maps, surface normals, segmentation masks, world positions, optical flow, etc for egocentric perception tasks. + +# 5. Experiments + +We assess the motion quality, generalizability, and diversity of our motion model, highlighting its ability to generalize to unseen complex tasks and comparing it with recent baselines (Sec. 5.1). We evaluate our proposed egocentric sensing as a depth proxy for enhancing agent exploration (Sec. 5.2) and conduct ablation studies (Sec. 5.3). + +We further demonstrate the effectiveness of EgoGen on three egocentric computer vision tasks in Sec. 5.4, and 5.5. By incorporating synthesized egocentric images, we can enhance the performance of the state-of-the-art algorithms. + +# 5.1. Evaluation of Learned CAMPs + +We assess CAMPs' generalizability in dynamic scenes, including scenes with moving obstacles and scenes with multiple individuals. In tests with moving obstacles, the obstacle blocks the person's path by moving between the person and the goal. In multiple human test scenes, lines from their starting and goal locations intersect in the middle, requiring solving human-human penetrations. See detail in Sup. Mat. + +In Tab. 1, we compare goal-reaching behaviors with two recent baselines: GAMMA [117] and DIMOS [119]. Base + +Table 2. Evaluation of egocentric sensing. (Sec. 5.2) + +
Method (sensing range)SR (%) ↑Dist. (m) ↓
Local map [119] (0.8 m)780.35
Local map* (7 m)43.04
Egocentric sensing (ours) (7 m)950.12
+ +line methods use navigation meshes and path planning for static scenes, while CAMPs can autonomously avoid dynamic obstacles (Sec. 3.3). For fair comparison in dynamic scenes, we extend the baselines by updating navigation meshes and performing on-the-fly path planning at each time step. The tree-based search as in [117] is disabled for all the methods. Metrics: (1) $SR$ : Success rate for reaching the goal location within a $0.3\mathrm{m}$ threshold. (2) Dist.: Average distance of the final pelvis location to the goal. (3) Cont.: The contact metric [117] that measures foot-floor contact and foot skating. (4) Pene-S.: Percentage of frames with detected human-scene penetration in moving obstacle scenes. (5) Pene-H.: Accurate human-human penetration evaluation metric using COAP [52] in multiple human scenes. Please refer to Supp. Mat. for metric details. + +CAMPs outperform the two baselines in dynamic scenarios with moving obstacles and multiple humans, exhibiting lower human-scene and human-human penetrations and a higher goal-reaching success rate. In multiple human scenarios, we observe that in the baselines, dynamically redoing path planning for each human independently can not effectively solve human-human penetration. In contrast, com-posable CAMPs can generalize well in dynamic settings without using multi-agent RL to synthesize crowd motions. + +We assess walking path diversity using the standard deviation of pelvis locations for the same start-target pairs in scenes with a single static box obstacle. As shown in Tab. 1 (Diversity), our approach does not require a pre-computed global path and allows agents to self-exlore without being constrained by predefined paths, achieving higher walking path diversity and success rate. This fosters diverse synthetic data generation via more diverse synthesized motion. + +# 5.2. Evaluation of Egocentric Sensing + +We assess the exploration ability of our egocentric sensing $\mathcal{E}_t$ in Replica [91] scenes. In Tab. 2, we replace $\mathcal{E}_t$ with a local map [119] in our state $\mathbf{s}_t$ , following their encoding method. Relying on local information can trap agents in local optima, e.g., walls beyond their sensing range, resulting in lower SR. Our egocentric sensing acts as a depth proxy, allowing the agent to avoid local optima, explore more effectively than local maps [119] or scandots [3], and achieve higher SR. In addition, our compact representation is more scalable as the sensing range increases, while quadratic local map growth can hinder the policy network's learning. + +Table 3. Ablation studies. Note: in our observation, $\| VP\| _2 > 15$ indicates abnormal human poses. (Sec. 5.3) + +
SR (%)↑||VP||2↓cos(▼,▲)↑
Egocentric depth813.640.049
No pretraining9028.770.918
No attention reward9012.260.891
Our policy9210.570.940
+ +# 5.3. Ablation Studies + +We compare our policy with several ablations in Tab. 3: + +Egocentric depth: an ablation training an egocentric depth image-based policy without the depth sensing proxy. Egocentric depth images are encoded with a CNN; + +No pretraining: an ablation training collision avoidance in crowded scenes with strict penetration termination directly; No attention reward: an ablation for the viewing direction. + +We assess pose naturalness with the maximum pose embedding norm encoded with VPoser [63] and evaluate the attention reward with the cosine similarity between the viewing direction $\vec{\mathbf{v}}$ and the attention direction $\vec{\mathbf{a}}$ (Sec. 3.1). + +Directly training RL with egocentric depth images is ineffective due to our high-dimensional action space, emphasizing the value of the compact egocentric sensing representation. Training agents with strict penetration constraints in crowded scenes directly can result in exploring unreasonable action subspaces, given its unbounded Gaussian nature, leading to unrealistic human poses, highlighting the effectiveness of our two-stage RL training scheme. Without the attention reward, the virtual human's capability to attend to a specific direction decreases. All ablation studies are evaluated in Replica. See visuals in Supp. Vid. and Supp. Mat. + +# 5.4. Mapping, Localization, and Tracking for HMD + +Mapping and localization. LaMAR [76] is the first mapping and localization benchmark dataset for AR in large-scale scenes. Despite over a year of extensive data collection, the dataset still lacks exhaustive scene coverage, especially in large open spaces. EgoGen can let virtual humans explore large-scale scenes, render dense egocentric views, and build a more complete SfM map by extracting image feature points with SuperPoint [15] and matching images with SuperGlue [75]. Despite synthetic images being noisier due to scene quality, SuperGlue [75] matching can filter out noisy feature points and yield reliable matches. + +In Tab. 4, we evaluate EgoGen by assessing the localization recall at $(1^{\circ}, 10cm)$ on the validation set in a lobby of $\sim 120$ sqm of the LaMAR CAB location. In addition, we report the number of triangulated 3D points (#P3D) and track length. EgoGen improves the 3D reconstruction by yielding more points for a slightly improved track length and also a significantly better localization performance compared to using the real data only. Ng et al. [58] augments mapping + +Table 4. Mapping and localization evaluation. We augment LaMAR with the same amount of images (248 frames) and report the localization recall at $(1^{\circ},10cm)$ on the validation set. EgoGen achieves the highest track length and recall. (Sec. 5.4) + +
#P3D ↑Track length ↑Recall (%) ↑
LaMAR19297395.194666.9
Ng et al. [58]19377585.194074.9
EgoGen19361695.210576.7
+ +Table 5. Egocentric camera tracking evaluation of models trained with and without synthetic data from EgoGen. (Sec. 5.4) + +
Pose ↓Rotation ↓Transl (mm) ↓
Scratch1.830.741303
+ EgoGen pretrain1.670.621305
+ +images by perturbing real-world camera poses with noise, which may generate unrealistic camera poses (e.g., stuck in a wall or facing the ceiling), limiting egocentric localization effectiveness. Their method also assumes the availability of initial camera poses, which may not always be feasible. In contrast, EgoGen augments by virtual humans randomly exploring scenes. Our approach holds promise for creating AR mapping and localization datasets for digital twin scenes without manual data collection, providing enhanced privacy preservation, e.g. no need for anonymization. Refer to Supp. Mat. for visualization and implementation details. + +Egocentric camera tracking. Egocentric camera tracking for HMD aims to yield device pose trajectories in 3D scenes given egocentric video observations. Recovering camera poses from monocular RGB videos using SLAM [93] is a challenging and ill-posed problem due to scale ambiguity. EgoEgo [43] leverages the knowledge of human motion to address the egocentric HMD tracking problem. Specifically, EgoEgo trains a neural network to infer the translation scaling and rotations from egocentric videos, which improves the HMD tracking performance. However, training this model requires jointly captured data of ground truth HMD trajectories and egocentric videos, which are costly to collect. We address this limitation by using EgoGen to synthesize quantities of egocentric videos with accurate camera trajectories to pretrain the model, which proves to improve the tracking performance on real data. We conduct experiments on the GIMO [121] dataset that contains $\sim 200$ short sequences of paired motion-video data in 19 scenes. Using EgoGen, we synthesize $\sim 4k$ sequences of human movements in their scenes and render corresponding egocentric videos using the same camera intrinsic as GIMO and the embodied camera placement described in Sec. 4.1. We also slightly perturb the camera placement location and orientation to simulate the diversity of how people wear HMDs in real data and avoid overfitting to one specific camera placement. We first pretrain the model with synthetic data generated by EgoGen, then finetune it on the real GIMO data. + +Tab. 5 shows the egocentric camera tracking performances for models trained with and without synthetic data. Definitions of evaluation metrics can be found in Supp. Mat. The finetuned model benefits from EgoGen synthetic data and predicts more accurate camera poses compared to the model trained using real data only. + +# 5.5. Human Mesh Recovery from Egocentric views + +Human mesh recovery (HMR) is the key to human behavior understanding from the egocentric view, thus crucial for applications in robotics and AR/VR. Given an egocentric RGB or depth image of a target subject, HMR aims to reconstruct the subject's 3D body pose and shape. However, acquiring and annotating real-world data is expensive, demanding, and time-consuming, with egocentric data being particularly scarce. EgoBody [115] is a recent egocentric dataset capturing two-people interactions, with egocentric depth/RGB frames annotated with SMPL-X body meshes. EgoBody provides $\sim 180\mathrm{k}$ egocentric RGB frames, and merely $\sim 23\mathrm{k}$ depth frames due to the low frame rate of the depth sensor, with $\sim 90\mathrm{k} / \sim 10\mathrm{k}$ in the RGB/depth training set. Such limited data is insufficient to train a learning-based model from scratch. In contrast, with EgoGen, largescale synthetic egocentric data can be generated in a time-efficient way. We leverage EgoGen to generate quantities of training frames (300k RGB, 105k depth) of humans moving in EgoBody 3D scenes, rendered from the egocentric view, and annotated by SMPL-X parameters of the target subject. Specifically, RGB images are rendered with lifelike human body textures and 3D clothing, with random lighting. + +With the recent HMR regressor, ProHMR [41], we show that pre-training with our synthetic data from EgoGen enhances the existing method's capability to generalize on real-world scenarios. Evaluated on the real-world EgoBody test set, we compare two training schemes: (1) trained from scratch on the real-world EgoBody training set ("-scratch"), and (2) pre-trained on synthetic data from EgoGen and finetuned on the real-world EgoBody training set ("-ft"). + +HMR from depth. As no existing methods were proposed for depth-based HMR task, we adapt ProHMR [41] to the depth input by changing the channel number of the first convolution layer. To mimic real-world sensor noise, synthetic noise [27] is added to the rendered depth. G-MPJPE is additionally reported for depth-based HMR as depth images provide global information. As shown in Tab. 6, compared to the model trained only with a limited amount of real-world data (Depth-scratch), errors are significantly reduced for the model pre-trained with our large-scale synthetic data (Depth-ft), in terms of global translation (22.9% lower G-MPJPE), local pose (20.7% lower MJPE), and body shape (19.5% lower V2V). + +HMR from RGB. For training with RGB images, we apply various data augmentation techniques similar to [8]. + +Table 6. Evaluation of HMR on EgoBody test set. “*-scratch” denotes the model trained from scratch with the Egobody training set, and “*-ft” denotes the model pre-trained with EgoGen synthetic data. The units for all metrics are in mm. (Sec. 5.5) + +
G-MPJPE ↓MPJPE ↓PA-MPJPE ↓V2V ↓
Depth-scratch117.782.254.1100.6
Depth-ft90.765.247.381.0
RGB-scratch-90.759.9102.1
RGB-ft-85.356.297.2
+ +Tab. 6 indicates that the RGB-based model pre-trained with large-scale synthetic data ("RGB-ft") also outperforms the model trained only on real-world data ("RGB-scratch"), for both body pose and shape accuracy. + +The enhanced performance highlights that EgoGen's synthetic data effectively compensates for the lack of real-world training data, boosting the performance of current methods when test on real-world data. We will release both of our synthetic EgoBody datasets. See dataset statistics, qualitative visualizations, and training details in Supp. Mat. + +# 6. Conclusion and Future Work + +We propose a novel egocentric synthetic data generation approach, EgoGen, that uses embodied sensors, a parametric body model, and a generative egocentric perception-driven human motion synthesis method to create egocentric training data with accurate and rich ground truth annotations. By integrating deep reinforcement learning and collision-avoiding motion primitives with egocentric depth proxy, EgoGen synthesizes robust human motion and emergent multi-agent behaviors. This paves the way to an efficient and scalable data generation solution that may have a profound impact on egocentric perception tasks. + +Human-scene interaction in EgoGen is currently coarse. We aim to extend the current method to simulate more detailed human motion driven by egocentric perception, such as hand manipulation, sitting, lying, etc, to facilitate more realistic egocentric synthetic data. We use fixed attention goals to model human attention. Predicting human intention through historical egocentric perception and synthesizing viewing directions based on predicted intention holds significant potential. Synthesizing gaze direction for predicting human intent is valuable but presently hampered by data requirements; we will revisit this when resources allow. + +We will explore many other egocentric vision tasks with EgoGen as this area grows rapidly such as social understanding and forecasting. EgoGen could benefit humanrobot interaction, e.g., our generative human motion model and lifelike human appearances can be integrated into [67] to close the sim2real gap for robotic agents further. + +Acknowledgements. This work is supported by the SNSF project grant 200021 204840 and the SDSC PhD fellowship. We sincerely thank Linfei Pan, Yunke Ao, Qianli Ma, and Maxime Raafat for the valuable discussions. + +# References + +[1] Meshcapade GmbH, Tübingen, Germany. https://meshcapade.com, 2022.5 +[2] Mohammed Abdel-Wahab, Konrad Wenzel, and Dieter Fritsch. Efficient reconstruction of large unordered image datasets for high accuracy photogrammetric applications. In ISPRS Annals of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Melbourne, Australia. XXII ISPRS Congress, 2012. 3 +[3] Ananye Agarwal, Ashish Kumar, Jitendra Malik, and Deepak Pathak. Legged locomotion in challenging terrains using egocentric vision. In Conference on Robot Learning, CoRL 2022, 14-18 December 2022, Auckland, New Zealand, pages 403-415. PMLR, 2022. 3, 6 +[4] M. Pilar Aivar, Eli Brenner, and Jeroen B. J. Smeets. Avoiding moving obstacles. Experimental Brain Research, 190 (3):251-264, 2008. 5 +[5] Hiroyasu Akada, Jian Wang, Soshi Shimada, Masaki Takahashi, Christian Theobalt, and Vladislav Golyanik. Unrealego: A new dataset for robust egocentric 3d human motion capture. In European Conference on Computer Vision (ECCV), 2022. 2 +[6] Apple. ARKit. https://developer.apple.com/arkit/, 2017.3 +[7] Sonia Baltodano, Srinath Sibi, Nikolas Martelaro, Nikhil Gowda, and Wendy Ju. The rroads platform: a real road autonomous driving simulator. In Proceedings of the 7th International Conference on Automotive User Interfaces and Interactive Vehicular Applications, pages 281-288, 2015. 2 +[8] Michael J. Black, Priyanka Patel, Joachim Tesch, and Jinlong Yang. BEDLAM: A synthetic dataset of bodies exhibiting detailed lifelike animated motion. In Proceedings IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR), 2023. 2, 5, 8 +[9] Federica Bogo, Angjoo Kanazawa, Christoph Lassner, Peter Gehler, Javier Romero, and Michael J. Black. Keep it SMPL: Automatic estimation of 3D human pose and shape from a single image. In European Conference on Computer Vision, pages 561-578, 2016. 3 +[10] Simon Clavet. Motion matching and the road to next-gen animation. In Proc. of GDC, 2016. 2 +[11] Blender Online Community. Blender - a 3D modelling and rendering package. Blender Foundation, Stichting Blender Foundation, Amsterdam, 2018. 5, 6 +[12] Erwin Coumans and Yunfei Bai. Pybullet, a python module for physics simulation for games, robotics and machine learning. http://pybullet.org, 2016-2021. 2 +[13] Dima Damen, Hazel Doughty, Giovanni Maria Farinella, Sanja Fidler, Antonino Furnari, Evangelos Kazakos, Davide Moltisanti, Jonathan Munro, Toby Perrett, Will Price, and Michael Wray. The epic-kitchens dataset: Collection, challenges and baselines. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 43(11):4125-4141, 2021. 2 +[14] Andrew J. Davison. Real-time simultaneous localisation and mapping with a single camera. In ICCV, 2003. 3 + +[15] Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. Superpoint: Self-supervised interest point detection and description, 2018. 7 +[16] Alexey Dosovitskiy, German Ros, Felipe Codevilla, Antonio Lopez, and Vladlen Koltun. Carla: An open urban driving simulator, 2017. 2 +[17] Salehe Erfanian Ebadi, Saurav Dhakad, Sanjay Vishwakarma, Chunpu Wang, You-Cyuan Jhang, Maciek Chociej, Adam Crespi, Alex Thaman, and Sujoy Ganguly. Psp-hdri+: A synthetic dataset generator for pre-training of human-centric computer vision models. In First Workshop on Pre-training: Perspectives, Pitfalls, and Paths Forward at ICML 2022, 2022. 2 +[18] Matteo Fabbri, Fabio Lanzi, Simone Calderara, Andrea Palazzi, Roberto Vezzani, and Rita Cucchiara. Learning to detect and track visible and occluded body joints in a virtual world. In European Conference on Computer Vision (ECCV), 2018. 2 +[19] Alexander Fiannaca, Ilias Apostolopoulous, and Eelke Folmer. Headlock: A wearable navigation aid that helps blind cane users traverse large open spaces. In Proceedings of the 16th international ACM SIGACCESS conference on Computers & accessibility, pages 19-26, 2014. 2 +[20] Zahra Ghodsi, Siva Kumar Sastry Hari, Iuri Frosio, Timothy Tsai, Alejandro Troccoli, Stephen W. Keckler, Siddharth Garg, and Anima Anandkumar. Generating and characterizing scenarios for safety testing of autonomous vehicles, 2021. 2 +[21] Google. ARCore. https://developers.google.com/ar/, 2018.3 +[22] Venu Madhav Govindu. Combining two-view constraints for motion estimation. In CVPR, 2001. 3 +[23] Kristen Grauman, Andrew Westbury, Eugene Byrne, Zachary Chavis, Antonino Furnari, Rohit Girdhar, Jackson Hamburger, Hao Jiang, Miao Liu, Xingyu Liu, Miguel Martin, Tushar Nagarajan, Iija Radosavovic, Santhosh Kumar Ramakrishnan, Fiona Ryan, Jayant Sharma, Michael Wray, Mengmeng Xu, Eric Zhongcong Xu, Chen Zhao, Siddhant Bansal, Dhruv Batra, Vincent Cartillier, Sean Crane, Tien Do, Morrie Doulaty, Akshay Erapalli, Christoph Feichtenhofer, Adriano Fragomeni, Qichen Fu, Christian Fuegen, Abrham Gebreselasie, Cristina Gonzalez, James Hillis, Xuhua Huang, Yifei Huang, Wenqi Jia, Weslie Khoo, Jachym Kolar, Satwik Kottur, Anurag Kumar, Federico Landini, Chao Li, Yanghao Li, Zhenqiang Li, Karttikeya Mangalam, Raghava Modhugu, Jonathan Munro, Tullie Murrell, Takumi Nishiyasu, Will Price, Paola Ruiz Puentes, Merey Ramazanova, Leda Sari, Kiran Somasundaram, Audrey Southerland, Yusuke Sugano, Ruijie Tao, Minh Vo, Yuchen Wang, Xindi Wu, Takuma Yagi, Yunyi Zhu, Pablo Arbelaez, David Crandall, Dima Damen, Giovanni Maria Farinella, Bernard Ghanem, Vamsi Krishna Ithapu, C. V. Jawahar, Hanbyul Joo, Kris Kitani, Haizhou Li, Richard Newcombe, Aude Oliva, Hyun Soo Park, James M. Rehg, Yoichi Sato, Jianbo Shi, Mike Zheng Shou, Antonio Torralba, Lorenzo Torresani, Mingfei Yan, and Jitendra Malik. Ego4d: Around the World in 3,000 Hours of Egocentric + +Video. In IEEE/CVF Computer Vision and Pattern Recognition (CVPR), 2022. 1, 2 +[24] Artur Grigorev, Michael J. Black, and Otmar Hilliges. HOOD: hierarchical graphs for generalized modelling of clothing dynamics. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2023, Vancouver, BC, Canada, June 17-24, 2023, pages 16965-16974. IEEE, 2023. 5 +[25] Vladimir Guzov, Aymen Mir, Torsten Sattler, and Gerard Pons-Moll. Human poseitioning system (hps): 3d human pose estimation and self-localization in large scenes from body-mounted sensors. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4318-4329, 2021. 3 +[26] Norbert Haala, Michael Cramer, Florian Weimer, and Martin Trittler. Performance test on uav-based photogrammetric data collection. The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, 2012. 3 +[27] Ankur Handa, Thomas Whelan, John McDonald, and Andrew J Davison. A benchmark for rgb-d visual odometry, 3d reconstruction and slam. ICRA, 2014. 8 +[28] Albert Haque, Boya Peng, Zelun Luo, Alexandre Alahi, Serena Yeung, and Li Fei-Fei. Towards viewpoint invariant 3d human pose estimation. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11–14, 2016, Proceedings, Part I 14, pages 160–177. Springer, 2016. 3 +[29] Mohamed Hassan, Duygu Ceylan, Ruben Villegas, Jun Saito, Jimei Yang, Yi Zhou, and Michael Black. Stochastic scene-aware motion prediction. In Proceedings of the International Conference on Computer Vision 2021, 2, 3 +[30] Mohamed Hassan, Yunrong Guo, Tingwu Wang, Michael Black, Sanja Fidler, and Xue Bin Peng. Synthesizing physical character-scene interactions. arXiv preprint arXiv:2302.00883, 2023. 2 +[31] Daniel Holden, Taku Komura, and Jun Saito. Phase-functional neural networks for character control. ACM Trans. Graph., 36(4):1-13, 2017. 2 +[32] Daniel Holden, Oussama Kanoun, Maksym Perepichka, and Tiberiu Popa. Learned motion matching. ACM Transactions on Graphics (TOG), 39(4):53-1, 2020. 2 +[33] Yuhe Jin, Dmytro Mishkin, Anastasiia Mishchuk, Jiri Matas, Pascal Fua, Kwang Moo Yi, and Eduard Trulls. Image Matching across Wide Baselines: From Paper to Practice. IJCV, 2021. 3 +[34] Neena Kamath. Announcing Azure Spatial Anchors for collaborative, cross-platform mixed reality apps. https://azure.microsoft.com/en-us/blog/announcing-azure-spatial-anchors-for-collaborative-cross-platform-mixed-reality-apps/, 2019.3 +[35] Angjoo Kanazawa, Michael J Black, David W Jacobs, and Jitendra Malik. End-to-end recovery of human shape and pose. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 7122-7131, 2018. 3 + +[36] Alex Kendall, Matthew Grimes, and Roberto Cipolla. PoseNet: A Convolutional Network for Real-Time 6-DoF Camera Relocalization. In ICCV, 2015. 3 +[37] Georg Klein and David Murray. Parallel tracking and mapping for small ar workspaces. In IEEE and ACM International Symposium on Mixed and Augmented Reality, 2007. 3 +[38] Muhammed Kocabas, Chun-Hao P. Huang, Otmar Hilliges, and Michael J. Black. PARE: Part attention regressor for 3D human body estimation. In Proceedings International Conference on Computer Vision (ICCV), pages 11127-11137. IEEE, 2021. 3 +[39] Nikos Kolotouros, Georgios Pavlakos, Michael J Black, and Kostas Daniilidis. Learning to reconstruct 3D human pose and shape via model-fitting in the loop. In Proceedings of the IEEE International Conference on Computer Vision, pages 2252-2261, 2019. +[40] Nikos Kolotouros, Georgios Pavlakos, and Kostas Dani-ilidis. Convolutional mesh regression for single-image human shape reconstruction. In CVPR, 2019. +[41] Nikos Kolotouros, Georgios Pavlakos, Dinesh Jayaraman, and Kostas Daniilidis. Probabilistic modeling for human mesh recovery. In Proceedings of the IEEE/CVF international conference on computer vision, pages 11605-11614, 2021. 3, 8 +[42] Lucas Kovar, Michael Gleicher, and Frédéric H. Pighin. Motion graphs. In International Conference on Computer Graphics and Interactive Techniques, SIGGRAPH 2008, Los Angeles, California, USA, August 11-15, 2008, Classes, pages 51:1-51:10. ACM, 2008. 2 +[43] Jiaman Li, C. Karen Liu, and Jiajun Wu. Ego-body pose estimation via ego-head pose estimation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2023, Vancouver, BC, Canada, June 17-24, 2023, pages 17142-17151. IEEE, 2023. 1, 2, 7 +[44] Zhihao Li, Jianzhuang Liu, Zhensong Zhang, Songcen Xu, and Youliang Yan. CLIFF: Carrying location information in full frames into human pose and shape estimation. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part V, pages 590-606. Springer, 2022. 3 +[45] Hung Yu Ling, Fabio Zinno, George Cheng, and Michiel van de Panne. Character controllers using motion vaes. ACM Trans. Graph., 39(4), 2020. 3 +[46] Miao Liu, Dexin Yang, Yan Zhang, Zhaopeng Cui, James M Rehg, and Siyu Tang. 4d human body capture from egocentric video via 3d scene grounding. In 2021 International Conference on 3D Vision (3DV), pages 930-939. IEEE, 2021. 3 +[47] Zhengyi Luo, Ryo Hachiuma, Ye Yuan, Shun Iwase, and Kris M Kitani. Kinematics-guided reinforcement learning for object-aware 3d ego-pose estimation. arXiv preprint arXiv:2011.04837, 2020. 3 +[48] Naureen Mahmood, Nima Ghorbani, Nikolaus F Troje, Gerard Pons-Moll, and Michael J Black. Amass: Archive of motion capture as surface shapes. In Proceedings of the IEEE/CVF international conference on computer vision, pages 5442-5451, 2019. 3 + +[49] Angel Martínez-González, Michael Villamizar, Olivier Canévet, and Jean-Marc Odobez. Residual pose: A decoupled approach for depth-based 3d human pose estimation. In 2020 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 10313-10318. IEEE, 2020. 3 +[50] Meta. Project Aria Glasses. https://www.projectaria.com/, 2023.6 +[51] Microsoft. HoloLens 2. https://www.microsoft.com/en-us/hololens, 2019.6 +[52] Marko Mihajlovic, Shunsuke Saito, Aayush Bansal, Michael Zollhoefer, and Siyu Tang. COAP: Compositional articulated occupancy of people. In Proceedings IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2022. 6 +[53] Aymen Mir, Xavier Puig, Angjoo Kanazawa, and Gerard Pons-Moll. Generating continual human motion in diverse 3d scenes. CoRR, abs/2304.02061, 2023. 2 +[54] Michael Montemerlo, Nicholas Roy, and Sebastian Thrun. Perspectives on standardization in mobile robot programming: the Carnegie mellon navigation (CARMEN) toolkit. In 2003 IEEE/RSJ International Conference on Intelligent Robots and Systems, Las Vegas, Nevada, USA, October 27 - November 1, 2003, pages 2436-2441. IEEE, 2003. 2 +[55] Gyeongsik Moon, Ju Yong Chang, and Kyoung Mu Lee. V2v-posenet: Voxel-to-voxel prediction network for accurate 3d hand and human pose estimation from a single depth map. In Proceedings of the IEEE conference on computer vision and pattern Recognition, pages 5079-5088, 2018. 3 +[56] Etienne Mouragnon, Maxime Lhuillier, Michel Dhome, Fabien Dekeyser, and Patrick Sayd. Real time localization and 3d reconstruction. In CVPR, 2006. 3 +[57] Evonne Ng, Donglai Xiang, Hanbyul Joo, and Kristen Grauman. You2me: Inferring body pose in egocentric video via first and second person interactions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9890-9900, 2020. 3 +[58] Tony Ng, Adrian Lopez-Rodriguez, Vassileios Balntas, and Krystian Mikolajczyk. Reassessing the limitations of cnn methods for camera pose regression, 2021. 7 +[59] Niantic. Niantic Expands Developer Platform and AR Tools with Niantic Lightship. https://nianticlabs.com/news/lightship/, 2021.3 +[60] David Nister, Oleg Naroditsky, and James Bergen. Visual odometry. In CVPR, 2004. 3 +[61] Xiaqing Pan, Nicholas Charron, Yongqian Yang, Scott Peters, Thomas Whelan, Chen Kong, Omkar M. Parkhi, Richard A. Newcombe, and Carl Yuheng Ren. Aria digital twin: A new benchmark dataset for egocentric 3d machine perception. CoRR, abs/2306.06362, 2023. 2 +[62] Fabio Pardo, Arash Tavakoli, Vitaly Levdik, and Petar Kormushev. Time limits in reinforcement learning. In Proceedings of the 35th International Conference on Machine Learning, pages 4045-4054, Stockholm, Sweden, 2018. PMLR. 4 +[63] Georgios Pavlakos, Vasileios Choutas, Nima Ghorbani, Timo Bolkart, Ahmed A. A. Osman, Dimitrios Tzionas, + +and Michael J. Black. Expressive body capture: 3D hands, face, and body from a single image. In Proceedings IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), pages 10975-10985, 2019. 3, 5, 7 +[64] Xue Bin Peng, Ze Ma, Pieter Abbeel, Sergey Levine, and Angjoo Kanazawa. Amp: Adversarial motion priors for stylized physics-based character control. ACM Transactions on Graphics (TOG), 40(4):1-20, 2021. 2 +[65] Mathis Petrovich, Michael J. Black, and Gül Varol. Action-conditioned 3d human motion synthesis with transformer vae, 2021. 3 +[66] Marc Pollefeys, Luc Van Gool, Maarten Vergauwen, Frank Verbiest, Kurt Cornelis, Jan Tops, and Reinhard Koch. Visual modeling with a hand-held camera. IJCV, 2004. 3 +[67] Xavier Puig, Eric Undersander, Andrew Szot, Mikael Dallaire Cote, Tsung-Yen Yang, Ruslan Partsey, Ruta Desai, Alexander William Clegg, Michal Hlavac, So Yeon Min, Vladimir Vondrus, Théophile Gervet, Vincent-Pierre Berges, John M. Turner, Oleksandr Maksymets, Zsolt Kira, Mrinal Kalakrishnan, Jitendra Malik, Devendra Singh Chaplot, Unnat Jain, Dhruv Batra, Akshara Rai, and Roozbeh Mottaghi. Habitat 3.0: A co-habitat for humans, avatars and robots. CoRR, abs/2310.13724, 2023. 2, 8 +[68] Albert Pumarola, Jordi Sanchez, Gary Choi, Alberto Sanfeliu, and Francesc Moreno-Noguer. 3DPeople: Modeling the Geometry of Dressed Humans. In International Conference in Computer Vision (ICCV), 2019. 2 +[69] Morgan Quigley, Ken Conley, Brian Gerkey, Josh Faust, Tully Foote, Jeremy Leibs, Rob Wheeler, Andrew Y Ng, et al. Ros: an open-source robot operating system. In ICRA workshop on open source software, page 5. Kobe, Japan, 2009. 2 +[70] Umer Rafi, Juergen Gall, and Bastian Leibe. A semantic occlusion model for human pose estimation from a single depth image. In Proceedings of the IEEE conference on computer vision and pattern recognition workshops, pages 67-74, 2015. 3 +[71] Tilman Reinhardt. Google Visual Positioning Service. https://ai.googleblog.com/2019/02/using-global-localization-to-improve.html, 2019.3 +[72] Davis Rempe, Tolga Birdal, Aaron Hertzmann, Jimei Yang, Srinath Sridhar, and Leonidas J. Guibas. Humor: 3d human motion model for robust pose estimation. In International Conference on Computer Vision (ICCV), 2021. 3 +[73] Davis Rempe, Zhengyi Luo, Xue Bin Peng, Ye Yuan, Kris Kitani, Karsten Kreis, Sanja Fidler, and Or Litany. Trace and pace: Controllable pedestrian animation via guided trajectory diffusion, 2023. 2, 3 +[74] Guodong Rong, Byung Hyun Shin, Hadi Tabatabaee, Qiang Lu, Steve Lemke, Martiš Možeiko, Eric Boise, Geehoon Uhm, Mark Gerow, Shalin Mehta, et al. Lgsvl simulator: A high fidelity simulator for autonomous driving. In 2020 IEEE 23rd International conference on intelligent transportation systems (ITSC), pages 1-6. IEEE, 2020. 2 +[75] Paul-Edouard Sarlin, Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. Superglue: Learning feature matching with graph neural networks, 2020. 7 + +[76] Paul-Edouard Sarlin, Mihai Dusmanu, Johannes L. Schonberger, Pablo Speciale, Lukas Gruber, Viktor Larsson, Ondrej Miksik, and Marc Pollefeys. LaMAR: Benchmarking Localization and Mapping for Augmented Reality. In ECCV, 2022. 3, 7 +[77] Torsten Sattler, Will Maddern, Carl Toft, Akihiko Torii, Lars Hammarstrand, Erik Stenberg, Daniel Safari, Masatoshi Okutomi, Marc Pollefeys, Josef Sivic, Fredrik Kahl, and Tomas Pajdla. Benchmarking 6DoF outdoor visual localization in changing conditions. In CVPR, 2018. 3 +[78] Manolis Savva, Abhishek Kadian, Oleksandr Maksymets, Yili Zhao, Erik Wijmans, Bhavana Jain, Julian Straub, Jia Liu, Vladlen Koltun, Jitendra Malik, Devi Parikh, and Dhruv Batra. Habitat: A platform for embodied ai research, 2019. 2 +[79] Johannes Lutz Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In CVPR, 2016. 3 +[80] Johannes L. Schonberger, Hans Hardmeier, Torsten Sattler, and Marc Pollefeys. Comparative evaluation of handcrafted and learned local features. In CVPR, 2017. 3 +[81] John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Radford, and Oleg Klimov. Proximal policy optimization algorithms, 2017. 4 +[82] Akash Sengupta, Ignas Budvytis, and Roberto Cipolla. Synthetic training for accurate 3d human pose and shape estimation in the wild. In British Machine Vision Conference (BMVC), 2020. 2 +[83] Takaaki Shiratori, Hyun Soo Park, Leonid Sigal, Yaser Sheikh, and Jessica K. Hodgins. Motion capture from body-mounted cameras. ACM Trans. Graph., 30(4):31, 2011. 3 +[84] Jamie Shotton, Ross Girshick, Andrew Fitzgibbon, Toby Sharp, Mat Cook, Mark Finocchio, Richard Moore, Pushmeet Kohli, Antonio Criminisi, Alex Kipman, et al. Efficient human pose estimation from single depth images. IEEE transactions on pattern analysis and machine intelligence, 35(12):2821-2840, 2012. 3 +[85] Jamie Shotton, Ben Glocker, Christopher Zach, Shahram Izadi, Antonio Criminisi, and Andrew Fitzgibbon. Scene Coordinate Regression Forests for Camera Relocalization in RGB-D Images. In CVPR, 2013. 3 +[86] Jamie Shotton, Toby Sharp, Alex Kipman, Andrew Fitzgibbon, Mark Finocchio, Andrew Blake, Mat Cook, and Richard Moore. Real-time human pose recognition in parts from single depth images. Commun. ACM, 56(1):116-124, 2013. 2 +[87] Noah Snavely, Steven M. Seitz, and Richard Szeliski. Photo tourism: exploring photo collections in 3d. ACM Trans. Graph., 25(3):835-846, 2006. 3 +[88] Kihyuk Sohn, Honglak Lee, and Xinchen Yan. Learning structured output representation using deep conditional generative models. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2015. 3 +[89] Sebastian Starke, He Zhang, Taku Komura, and Jun Saito. Neural state machine for character-scene interactions. ACM Trans. Graph., 38(6):209-1, 2019. 2 +[90] Sebastian Starke, Ian Mason, and Taku Komura. Deepphase: Periodic autoencoders for learning motion phase + +manifolds. ACM Transactions on Graphics (TOG), 41(4): 1-13, 2022. 2 +[91] Julian Straub, Thomas Whelan, Lingni Ma, Yufan Chen, Erik Wijmans, Simon Green, Jakob J. Engel, Raul Mur-Artal, Carl Ren, Shobhit Verma, Anton Clarkson, Mingfei Yan, Brian Budge, Yajie Yan, Xiaqing Pan, June Yon, Yuyang Zou, Kimberly Leon, Nigel Carter, Jesus Briales, Tyler Gillingham, Elias Mueggler, Luis Pesqueira, Manolis Savva, Dhruv Batra, Hauke M. Strasdat, Renzo De Nardi, Michael Goesele, Steven Lovegrove, and Richard Newcombe. The Replica dataset: A digital replica of indoor spaces. arXiv preprint arXiv:1906.05797, 2019. 4, 6 +[92] Andrew Szot, Alex Clegg, Eric Undersander, Erik Wijmans, Yili Zhao, John Turner, Noah Maestre, Mustafa Mukadam, Devendra Chaplot, Oleksandr Maksymets, Aaron Gokaslan, Vladimir Vondrus, Sameer Dharur, Franziska Meier, Wojciech Galuba, Angel Chang, Zsolt Kira, Vladlen Koltun, Jitendra Malik, Manolis Savva, and Dhruv Batra. Habitat 2.0: Training home assistants to rearrange their habitat. In Advances in Neural Information Processing Systems (NeurIPS), 2021. 2 +[93] Zachary Teed and Jia Deng. DROID-SLAM: Deep Visual SLAM for Monocular, Stereo, and RGB-D Cameras. Advances in neural information processing systems, 2021. 7 +[94] Guy Tevet, Sigal Raab, Brian Gordon, Yonatan Shafir, Daniel Cohen-Or, and Amit H Bermano. Human motion diffusion model. arXiv preprint arXiv:2209.14916, 2022.3 +[95] Emanuel Todorov, Tom Erez, and Yuval Tassa. Mujoco: A physics engine for model-based control. In 2012 IEEE/RSJ International Conference on Intelligent Robots and Systems, pages 5026-5033, 2012. 2 +[96] Emanuel Todorov, Tom Erez, and Yuval Tassa. Mujoco: A physics engine for model-based control. In 2012 IEEE/RSJ International Conference on Intelligent Robots and Systems, pages 5026-5033. IEEE, 2012. 2 +[97] Denis Tome, Patrick Peluse, Lourdes Agapito, and Hernan Badino. xr-egopose: Egocentric 3d human pose from an hmd camera. In Proceedings of the IEEE International Conference on Computer Vision, pages 7728-7738, 2019. 2, 3 +[98] Denis Tome, Thiemo Alldieck, Patrick Peluse, Gerard Pons-Moll, Lourdes Agapito, Hernan Badino, and Fernando De la Torre. Selfpose: 3d egocentric pose estimation from a headset mounted camera. arXiv preprint arXiv:2011.01519, 2020. 3 +[99] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Advances in neural information processing systems, pages 5998-6008, 2017. 4 +[100] Jian Wang, Diogo Luvizon, Weipeng Xu, Lingjie Liu, Kripasindhu Sarkar, and Christian Theobalt. Scene-aware egocentric 3d human pose estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13031-13040, 2023. 3 +[101] Keze Wang, Shengfu Zhai, Hui Cheng, Xiaodan Liang, and Liang Lin. Human pose estimation from depth images via + +inference embedded multi-task learning. In Proceedings of the 24th ACM international conference on Multimedia, pages 1227-1236, 2016. 3 +[102] Xin Wang, Taein Kwon, Mahdi Rad, Bowen Pan, Ishani Chakraborty, Sean Andrist, Dan Bohus, Ashley Feniello, Bugra Tekin, Felipe Vieira Frujeri, Neel Joshi, and Marc Pollefeys. Holoassist: an egocentric human interaction dataset for interactive ai assistants in the real world. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 20270-20281, 2023. 2 +[103] Erik Wijmans, Abhishek Kadian, Ari Morcos, Stefan Lee, Irfan Essa, Devi Parikh, Manolis Savva, and Dhruv Batra. DD-PPO: learning near-perfect pointgoal navigators from 2.5 billion frames. In 8th International Conference on Learning Representations, ICLR 2020, Addis Ababa, Ethiopia, April 26-30, 2020. OpenReview.net, 2020. 3 +[104] Kyle Wilson and Noah Snavely. Robust global translations with 1dsfm. In ECCV, 2014. 3 +[105] Jungdam Won, Deepak Gopinath, and Jessica Hodgins. Physics-based character controllers using conditional vaes. ACM Trans. Graph., 41(4), 2022. 2 +[106] Erroll Wood, Tadas Baltrusaitis, Charlie Hewitt, Sebastian Dziadzio, Thomas J Cashman, and Jamie Shotton. Fake it till you make it: face analysis in the wild using synthetic data alone. In Proceedings of the IEEE/CVF international conference on computer vision, pages 3681-3691, 2021. 2 +[107] Fu Xiong, Boshen Zhang, Yang Xiao, Zhiguo Cao, Taidong Yu, Joey Tianyi Zhou, and Junsong Yuan. A2j: Anchor-to-joint regression network for 3d articulated pose estimation from a single depth image. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 793-802, 2019. 3 +[108] Mengyuan Yan, Gen Li, Yilin Zhu, and Jeannette Bohg. Learning topological motion primitives for knot planning. In IEEE/RSJ International Conference on Intelligent Robots and Systems, IROS 2020, Las Vegas, NV, USA, October 24, 2020 - January 24, 2021, pages 9457-9464. IEEE, 2020. 3 +[109] Zhitao Yang, Zhongang Cai, Haiyi Mei, Shuai Liu, Zhaoxi Chen, Weiye Xiao, Yukun Wei, Zhongfei Qing, Chen Wei, Bo Dai, Wayne Wu, Chen Qian, Dahua Lin, Ziwei Liu, and Lei Yang. Synbody: Synthetic dataset with layered human models for 3d human perception and modeling, 2023. 5 +[110] Heyuan Yao, Zhenhua Song, Baoquan Chen, and Libin Liu. Controlvae: Model-based learning of generative controllers for physics-based characters. ACM Trans. Graph., 41(6), 2022. 3 +[111] Mao Ye, Xianwang Wang, Ruigang Yang, Liu Ren, and Marc Pollefeys. Accurate 3d pose estimation from a single depth image. In 2011 International Conference on Computer Vision, pages 731-738. IEEE, 2011. 3 +[112] Vickie Ye, Georgios Pavlakos, Jitendra Malik, and Angjoo Kanazawa. Decoupling human and camera motion from videos in the wild. arXiv preprint arXiv:2302.12827, 2023. 3 +[113] Ye Yuan and Kris Kitani. Ego-posed estimation and forecasting as real-time pd control. In Proceedings of the + +IEEE/CVF International Conference on Computer Vision, pages 10082-10092, 2019. 3 +[114] Mingyuan Zhang, Zhongang Cai, Liang Pan, Fangzhou Hong, Xinying Guo, Lei Yang, and Ziwei Liu. Motiondiffuse: Text-driven human motion generation with diffusion model. arXiv preprint arXiv:2208.15001, 2022. 3 +[115] Siwei Zhang, Qianli Ma, Yan Zhang, Zhiyin Qian, Taein Kwon, Marc Pollefeys, Federica Bogo, and Siyu Tang. Egobody: Human body shape and motion of interacting people from head-mounted devices. In European conference on computer vision (ECCV), 2022. 1, 2, 8 +[116] Siwei Zhang, Qianli Ma, Yan Zhang, Sadegh Aliakbarian, Darren Cosker, and Siyu Tang. Probabilistic human mesh recovery in 3d scenes from egocentric views. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2023. 3 +[117] Yan Zhang and Siyu Tang. The wanderings of odysseus in 3d scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20481-20491, 2022. 3, 6 +[118] Yan Zhang, Michael J Black, and Siyu Tang. We are more than our joints: Predicting how 3d bodies move. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3372-3382, 2021. 3 +[119] Kaifeng Zhao, Yan Zhang, Shaofei Wang, Thabo Beeler, and Siyu Tang. Synthesizing diverse human motions in 3d indoor scenes. In International conference on computer vision (ICCV), 2023. 2, 3, 4, 6 +[120] Yuhang Zhao, Elizabeth Kupferstein, Brenda Veronica Castro, Steven Feiner, and Shiri Azenkot. Designing ar visualizations to facilitate stair navigation for people with low vision. In Proceedings of the 32nd annual ACM symposium on user interface software and technology, pages 387-402, 2019. 2 +[121] Yang Zheng, Yanchao Yang, Kaichun Mo, Jiaman Li, Tao Yu, Yebin Liu, C Karen Liu, and Leonidas J Guibas. Gimo: Gaze-informed human motion prediction in context. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XIII, pages 676-694. Springer, 2022. 7 +[122] Ziwen Zhuang, Zipeng Fu, Jianren Wang, Christopher Atkeson, Soren Schwertfeger, Chelsea Finn, and Hang Zhao. Robot parkour learning. In Conference on Robot Learning (CoRL), 2023. 2, 3 \ No newline at end of file diff --git a/2024/EgoGen_ An Egocentric Synthetic Data Generator/images.zip b/2024/EgoGen_ An Egocentric Synthetic Data Generator/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..c2d0167a8ed8c39c4b7e9f67b5e18e64bd27dd40 --- /dev/null +++ b/2024/EgoGen_ An Egocentric Synthetic Data Generator/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:23a8f170ab3a81ae2a74691b64b94fd74d33f5d977afb868e7bf6874a178de00 +size 312506 diff --git a/2024/EgoGen_ An Egocentric Synthetic Data Generator/layout.json b/2024/EgoGen_ An Egocentric Synthetic Data Generator/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..0affd4004be4d75c2bc32fe11e61b4cd19a5e64c --- /dev/null +++ b/2024/EgoGen_ An Egocentric Synthetic Data Generator/layout.json @@ -0,0 +1,11436 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 141, + 103, + 453, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 103, + 453, + 121 + ], + "spans": [ + { + "bbox": [ + 141, + 103, + 453, + 121 + ], + "type": "text", + "content": "EgoGen: An Egocentric Synthetic Data Generator" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 142, + 462, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 142, + 462, + 203 + ], + "spans": [ + { + "bbox": [ + 138, + 142, + 462, + 203 + ], + "type": "text", + "content": "Gen Li" + }, + { + "bbox": [ + 138, + 142, + 462, + 203 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 138, + 142, + 462, + 203 + ], + "type": "text", + "content": " Kaifeng Zhao" + }, + { + "bbox": [ + 138, + 142, + 462, + 203 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 138, + 142, + 462, + 203 + ], + "type": "text", + "content": " Siwei Zhang" + }, + { + "bbox": [ + 138, + 142, + 462, + 203 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 138, + 142, + 462, + 203 + ], + "type": "text", + "content": " Xiaozhong Lyu" + }, + { + "bbox": [ + 138, + 142, + 462, + 203 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 138, + 142, + 462, + 203 + ], + "type": "text", + "content": " \nMihai Dusmanu" + }, + { + "bbox": [ + 138, + 142, + 462, + 203 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 138, + 142, + 462, + 203 + ], + "type": "text", + "content": " Yan Zhang" + }, + { + "bbox": [ + 138, + 142, + 462, + 203 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 138, + 142, + 462, + 203 + ], + "type": "text", + "content": " Marc Pollefeys" + }, + { + "bbox": [ + 138, + 142, + 462, + 203 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 138, + 142, + 462, + 203 + ], + "type": "text", + "content": " Siyu Tang" + }, + { + "bbox": [ + 138, + 142, + 462, + 203 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 138, + 142, + 462, + 203 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 138, + 142, + 462, + 203 + ], + "type": "text", + "content": "ETH Zürich \n" + }, + { + "bbox": [ + 138, + 142, + 462, + 203 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 138, + 142, + 462, + 203 + ], + "type": "text", + "content": "Microsoft \nhttps://ego-gen.github.io/" + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 48, + 217, + 296, + 383 + ], + "blocks": [ + { + "bbox": [ + 48, + 217, + 296, + 383 + ], + "lines": [ + { + "bbox": [ + 48, + 217, + 296, + 383 + ], + "spans": [ + { + "bbox": [ + 48, + 217, + 296, + 383 + ], + "type": "image", + "image_path": "20f96637edb2d3d562a65def7cff75f240c545540ad7a1c0a8763394d5c27659.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 82, + 387, + 97, + 394 + ], + "lines": [ + { + "bbox": [ + 82, + 387, + 97, + 394 + ], + "spans": [ + { + "bbox": [ + 82, + 387, + 97, + 394 + ], + "type": "text", + "content": "RGB" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 141, + 387, + 203, + 396 + ], + "lines": [ + { + "bbox": [ + 141, + 387, + 203, + 396 + ], + "spans": [ + { + "bbox": [ + 141, + 387, + 203, + 396 + ], + "type": "text", + "content": "RGB with Motion Blur" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 246, + 387, + 264, + 396 + ], + "lines": [ + { + "bbox": [ + 246, + 387, + 264, + 396 + ], + "spans": [ + { + "bbox": [ + 246, + 387, + 264, + 396 + ], + "type": "text", + "content": "Depth" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 296, + 217, + 378, + 383 + ], + "blocks": [ + { + "bbox": [ + 296, + 217, + 378, + 383 + ], + "lines": [ + { + "bbox": [ + 296, + 217, + 378, + 383 + ], + "spans": [ + { + "bbox": [ + 296, + 217, + 378, + 383 + ], + "type": "image", + "image_path": "73eaa166df6ddb80755c410a54a4b1ab7074c0acff1935e1445efd19d432d69c.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 316, + 387, + 359, + 395 + ], + "lines": [ + { + "bbox": [ + 316, + 387, + 359, + 395 + ], + "spans": [ + { + "bbox": [ + 316, + 387, + 359, + 395 + ], + "type": "text", + "content": "Surface Normal" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 408, + 547, + 453 + ], + "lines": [ + { + "bbox": [ + 46, + 408, + 547, + 453 + ], + "spans": [ + { + "bbox": [ + 46, + 408, + 547, + 453 + ], + "type": "text", + "content": "Figure 1. EgoGen: a scalable synthetic data generation system for egocentric perception tasks, with rich multi-modal data and accurate annotations. We simulate camera rigs for head-mounted devices (HMDs) and render from the perspective of the camera wearer with various sensors. Top to bottom: middle and right camera sensors in the rig. Left to right: photo-realistic RGB image, RGB with simulated motion blur, depth map, surface normal, segmentation mask, and world position for fisheye cameras widely used in HMDs." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 378, + 217, + 460, + 383 + ], + "blocks": [ + { + "bbox": [ + 378, + 217, + 460, + 383 + ], + "lines": [ + { + "bbox": [ + 378, + 217, + 460, + 383 + ], + "spans": [ + { + "bbox": [ + 378, + 217, + 460, + 383 + ], + "type": "image", + "image_path": "5d1f2404d006586ba9eb281ec93f4ef6684b3f107eb264300d5fb6b0cd250652.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 393, + 387, + 446, + 396 + ], + "lines": [ + { + "bbox": [ + 393, + 387, + 446, + 396 + ], + "spans": [ + { + "bbox": [ + 393, + 387, + 446, + 396 + ], + "type": "text", + "content": "Segmentation Mask" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 461, + 217, + 545, + 383 + ], + "blocks": [ + { + "bbox": [ + 461, + 217, + 545, + 383 + ], + "lines": [ + { + "bbox": [ + 461, + 217, + 545, + 383 + ], + "spans": [ + { + "bbox": [ + 461, + 217, + 545, + 383 + ], + "type": "image", + "image_path": "a97948112dd8e0dedfe94b08c70c361c3b1c45a52a4fc02f4560d6042536f524.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 482, + 387, + 522, + 395 + ], + "lines": [ + { + "bbox": [ + 482, + 387, + 522, + 395 + ], + "spans": [ + { + "bbox": [ + 482, + 387, + 522, + 395 + ], + "type": "text", + "content": "World Position" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "bbox": [ + 143, + 463, + 192, + 475 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 463, + 192, + 475 + ], + "spans": [ + { + "bbox": [ + 143, + 463, + 192, + 475 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 46, + 498, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 498, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 498, + 289, + 715 + ], + "type": "text", + "content": "Understanding the world in first-person view is fundamental in Augmented Reality (AR). This immersive perspective brings dramatic visual changes and unique challenges compared to third-person views. Synthetic data has empowered third-person-view vision models, but its application to embodied egocentric perception tasks remains largely unexplored. A critical challenge lies in simulating natural human movements and behaviors that effectively steer the embodied cameras to capture a faithful egocentric representation of the 3D world. To address this challenge, we introduce EgoGen, a new synthetic data generator that can produce accurate and rich ground-truth training data for egocentric perception tasks. At the heart of EgoGen is a novel human motion synthesis model that directly leverages egocentric visual inputs of a virtual human to sense the 3D environment. Combined with collision-avoiding motion primitives and a two-stage reinforcement learning approach, our motion synthesis model offers a closed-loop so" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 464, + 547, + 608 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 464, + 547, + 608 + ], + "spans": [ + { + "bbox": [ + 304, + 464, + 547, + 608 + ], + "type": "text", + "content": "lution where the embodied perception and movement of the virtual human are seamlessly coupled. Compared to previous works, our model eliminates the need for a pre-defined global path, and is directly applicable to dynamic environments. Combined with our easy-to-use and scalable data generation pipeline, we demonstrate EgoGen's efficacy in three tasks: mapping and localization for head-mounted cameras, egocentric camera tracking, and human mesh recovery from egocentric views. EgoGen will be fully open-sourced, offering a practical solution for creating realistic egocentric training data and aiming to serve as a useful tool for egocentric computer vision research." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 306, + 621, + 387, + 634 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 621, + 387, + 634 + ], + "spans": [ + { + "bbox": [ + 306, + 621, + 387, + 634 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 641, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 641, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 641, + 547, + 715 + ], + "type": "text", + "content": "The analysis of visual input from front-facing egocentric cameras is crucial for applications that benefit from a first-person perspective, mirroring the natural human experience [23, 43, 115]. AR devices, for instance, can utilize this viewpoint to amplify user immersion. Such cameras can also cater to individual preferences, providing custom" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14497" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 282, + 85 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 282, + 85 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 282, + 85 + ], + "type": "text", + "content": "visual assistance for those with impaired vision [19, 120]." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 86, + 286, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 86, + 286, + 205 + ], + "spans": [ + { + "bbox": [ + 46, + 86, + 286, + 205 + ], + "type": "text", + "content": "Despite its potential, egocentric perception faces challenges, primarily due to the scarcity of labeled data. Although datasets like Ego4D [23], ADT [61], EpicKitchen [13] and HoloAssist [102] exist, creating such datasets with rich and accurate annotations is costly and raises privacy concerns [115]. Alternatively, using graphics techniques to render synthetic multi-modal visual data has proven to be cost-effective and successful in training deep learning models, such as 3D human body estimation [8] and facial landmark detection [106]." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 207, + 286, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 207, + 286, + 266 + ], + "spans": [ + { + "bbox": [ + 46, + 207, + 286, + 266 + ], + "type": "text", + "content": "Creating egocentric synthetic data is challenging because egocentric cameras capture the complex interplay of body movements and the environment from the camera wearer's viewpoint. Modeling the intricate details and variations in human behavior presents a significant challenge." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 268, + 286, + 375 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 268, + 286, + 375 + ], + "spans": [ + { + "bbox": [ + 46, + 268, + 286, + 375 + ], + "type": "text", + "content": "To tackle this problem, we introduce EgoGen, an egocentric synthetic data generation approach that simulates data from embodied sensors, i.e., front-facing cameras in head-mounted devices (HMD). While the ultimate goal is to simulate human behaviors that are indistinguishable from reality, in this work, we focus on creating virtual humans (i.e., camera wearers) that can explore and avoid obstacles in the 3D world that is not only complex and dynamic but could potentially include other moving virtual humans." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 377, + 286, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 377, + 286, + 640 + ], + "spans": [ + { + "bbox": [ + 46, + 377, + 286, + 640 + ], + "type": "text", + "content": "Specifically, we propose a novel generative human motion model. Our key insight is that body movement and embodied perception should be seamlessly coupled. As William Gibson aptly stated, \"We see in order to move; we move in order to see,\" our egocentric perception is crucial for identifying obstacles, navigating in an environment, and planning actions. Our body movements are not solely a response to visual stimuli; they also change our egocentric perception. Therefore, the key idea of our motion model is to enable virtual humans to see their environment with egocentric visual inputs and respond accordingly by learning a policy to control a set of collision-avoiding motion primitives (CAMPs) that are composable for synthesizing long-term, diverse human motions. Due to the unbounded and high-dimensional latent action space of our generative motion primitive model, direct policy training through rendered egocentric images is often unstable [122]. Therefore, we propose a two-stage reinforcement learning scheme using an efficient egocentric visual proxy to couple egocentric visual cues and body movements seamlessly. In addition, we use an \"attention\" reward to incentivize egocentric perception behaviors, i.e., looking in the desired direction." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 642, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 642, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 642, + 286, + 713 + ], + "type": "text", + "content": "Empirical results showcase the benefits of our egocentric perception-driven motion framework, which does not require a pre-calculated walking path in 3D scenes as in [29, 53, 119]. Instead, it empowers virtual humans to perceive the environment from their own viewpoint, enabling them to navigate, circumvent obstacles, and plan move" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 72, + 545, + 179 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 179 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 179 + ], + "type": "text", + "content": "ments to reach the destination. Moreover, our model generalizes well to dynamic environments, even with training limited to static settings. By training virtual humans independently using CAMPs, our method synthesizes emergent multi-human behaviors without relying on multi-agent reinforcement learning algorithms. Egocentric visual cues are essential to build exploratory and generalizable motion models that unify navigational planning and movement control in complex and dynamic environments." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 180, + 545, + 275 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 180, + 545, + 275 + ], + "spans": [ + { + "bbox": [ + 304, + 180, + 545, + 275 + ], + "type": "text", + "content": "Building upon CAMPs, we further create a scalable data generation pipeline for EgoGen that outfits virtual humans with clothing, automates cloth animation, and integrates 3D assets from various sources. We validate EgoGen's efficacy across three egocentric perception tasks. The high-quality synthetic data with precise ground truth annotations consistently improve the performance of state-of-the-art methods. In summary, the contributions of this work are:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 276, + 545, + 454 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 306, + 276, + 545, + 311 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 276, + 545, + 311 + ], + "spans": [ + { + "bbox": [ + 306, + 276, + 545, + 311 + ], + "type": "text", + "content": "1. We introduce EgoGen, a generative and scalable synthetic data generation approach specifically tailored for egocentric perception tasks." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 312, + 545, + 382 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 312, + 545, + 382 + ], + "spans": [ + { + "bbox": [ + 306, + 312, + 545, + 382 + ], + "type": "text", + "content": "2. We introduce novel motion primitives based on egocentric visual cues, enabling diverse and realistic human motion synthesis in 3D scenes. These primitives empower virtual humans to handle complex scenarios, such as dynamic environments and crowd motion without relying on multi-agent reinforcement learning." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 383, + 545, + 454 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 383, + 545, + 454 + ], + "spans": [ + { + "bbox": [ + 306, + 383, + 545, + 454 + ], + "type": "text", + "content": "3. EgoGen enables us to augment existing real-world egocentric datasets with synthetic images. Quantitative results demonstrate enhanced performance in state-of-the-art algorithms on mapping and localization for head-mounted cameras, egocentric camera tracking, and human mesh recovery from egocentric views." + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 466, + 392, + 478 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 466, + 392, + 478 + ], + "spans": [ + { + "bbox": [ + 306, + 466, + 392, + 478 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 486, + 545, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 486, + 545, + 641 + ], + "spans": [ + { + "bbox": [ + 304, + 486, + 545, + 641 + ], + "type": "text", + "content": "Human-Related Simulators and Synthetic Data. Previous works primarily focus on simulating robots [54, 69, 78, 92, 95] and autonomous cars [7, 16, 20, 74]. While some incorporated animated digital humans, like in [5, 20, 67], these efforts relied on pre-recorded motion sequences. Rendering images of people to train perception models has been widely studied such as [5, 17, 18, 43, 68, 82, 86]. In particular, [97] offers large-scale synthetic data for egocentric camera wearer pose estimation but relies on mocap data, lacking realistic and spontaneous interactions with the digital world. In contrast, EgoGen closes the gap in egocentric synthetic data generation for head-mounted devices. Please refer to Supp. Mat. for detailed comparisons." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "type": "text", + "content": "Human Motion Synthesis. Generating high-fidelity human motions and interactions with 3D scenes is widely studied in graphics [10, 31, 32, 42, 89, 90]. While they can generate high-quality motion, it's usually deterministic. Synthesizing physically plausible human motions has been extensively studied [12, 30, 64, 73, 96, 105]. However," + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14498" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 191 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 191 + ], + "type": "text", + "content": "they struggle with generalization to different body shapes. For example, [73] explicitly created 2048 humanoids to improve body shape generalizability. Time series models [65, 94, 114] synthesize the stochastic motions of diverse people well. However, in [94, 114], their generated motion sequences have limited lengths and human-scene interactions are not explicitly considered. Autoregressive methods [45, 72, 117, 119] can produce perpetual motions. In particular, [117] can generalize to diverse body shapes and synthesize long-term human motions." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 192, + 289, + 276 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 192, + 289, + 276 + ], + "spans": [ + { + "bbox": [ + 46, + 192, + 289, + 276 + ], + "type": "text", + "content": "Our egocentric perception-driven motion synthesis model is closely related to [117, 119], but distinguishes itself w.r.t.: (1) Enabling virtual humans to explore using egocentric visual cues, without predefined paths. (2) Synthesizing egocentric perception behaviors beyond locomotion, e.g., looking in certain directions. (3) Handling dynamic environments and multi-agent behavior without re-training." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 276, + 289, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 276, + 289, + 430 + ], + "spans": [ + { + "bbox": [ + 46, + 276, + 289, + 430 + ], + "type": "text", + "content": "Mapping and Localization for AR. Localization and mapping from images is a long-standing problem known as: Photogrammetry [2, 26]; Structure-from-Motion (SfM) [22, 66, 79, 87, 104]; Simultaneous Localization and Mapping (SLAM) [14, 37, 56, 60]. Researchers have worked to make SLAM amenable for edge hand-held or head-mounted devices [6, 21, 37]. Cloud-based services like Google's Visual Positioning System [71], Niantic's Lightship [59], and Microsoft's Azure Spatial Anchors [34] have made visual localization and mapping more accessible. Benchmarking efforts have arisen for small-scale AR scenarios [36, 85], touristic landmarks [33, 80], and large-scale AR-device based localization [76, 77] to evaluate these systems." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 431, + 289, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 431, + 289, + 635 + ], + "spans": [ + { + "bbox": [ + 46, + 431, + 289, + 635 + ], + "type": "text", + "content": "Egocentric Human Pose Estimation. Estimating 3D bodies from RGB images is widely studied from third-person views [9, 35, 38-41, 44, 63], and egocentric views [25, 46, 47, 57, 83, 97, 98, 100, 112, 113, 116], mostly requiring expensive real-world data paired with ground truth annotations. Besides RGB images, depth images offer explicit 3D information, mitigating scale and shape ambiguity, with the potential to enable broader AR/VR applications. However, depth-based methods, especially for the egocentric view, are underexplored. Most existing works [28, 49, 55, 70, 84, 101, 107, 111] predict 3D body skeletons without expressive body meshes, struggling with challenges like severe body truncations and scene occlusions typical in egocentric views. Such limited attention mainly stems from the scarcity of data, as obtaining high-quality human mesh annotations for real-world depth images is labor-intensive." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 645, + 255, + 659 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 645, + 255, + 659 + ], + "spans": [ + { + "bbox": [ + 47, + 645, + 255, + 659 + ], + "type": "text", + "content": "3. Ego-Sensing Driven Motion Synthesis" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "type": "text", + "content": "To close the loop for the interdependence between egocentric synthetic image data and human motion synthesis, we use deep reinforcement learning (RL), integrating egocentric vision cues to synthesize human motions as described in" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 305, + 72, + 545, + 97 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 97 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 97 + ], + "type": "text", + "content": "Sec. 3.1 and 3.2. Subsequently, we extend learned policies to generate emergent multi-agent behaviors, as in Sec. 3.3." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 106, + 509, + 118 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 106, + 509, + 118 + ], + "spans": [ + { + "bbox": [ + 305, + 106, + 509, + 118 + ], + "type": "text", + "content": "3.1. Ego-Sensing Driven Motion Primitives" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 125, + 545, + 184 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 125, + 545, + 184 + ], + "spans": [ + { + "bbox": [ + 304, + 125, + 545, + 184 + ], + "type": "text", + "content": "Generating realistic egocentric data requires diverse and lifelike human motion synthesis. In this work, we consider arguably the most common everyday behaviors: navigating towards goals with egocentric perception while avoiding collisions with obstacles and people in dynamic 3D scenes." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 185, + 545, + 329 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 185, + 545, + 329 + ], + "spans": [ + { + "bbox": [ + 304, + 185, + 545, + 329 + ], + "type": "text", + "content": "Overview. Following recent literature [45, 110, 117, 119], we employ deep RL to train control policies on learned latent spaces that characterize natural human motions. However, unlike these previous works that only consider simple static scenes, we leverage egocentric perception and propose collision-avoiding motion primitives (CAMPs) to enable virtual humans to self-exlore and navigate in a dynamic environment. Specifically, CAMPs are trained jointly to produce collision-free motion sequences. At each timestep " + }, + { + "bbox": [ + 304, + 185, + 545, + 329 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 185, + 545, + 329 + ], + "type": "text", + "content": ", the agent observes the state " + }, + { + "bbox": [ + 304, + 185, + 545, + 329 + ], + "type": "inline_equation", + "content": "\\mathbf{s}_t" + }, + { + "bbox": [ + 304, + 185, + 545, + 329 + ], + "type": "text", + "content": ", performs an action " + }, + { + "bbox": [ + 304, + 185, + 545, + 329 + ], + "type": "inline_equation", + "content": "\\mathbf{a}_t" + }, + { + "bbox": [ + 304, + 185, + 545, + 329 + ], + "type": "text", + "content": ", and receives a reward " + }, + { + "bbox": [ + 304, + 185, + 545, + 329 + ], + "type": "inline_equation", + "content": "r_t = r(\\mathbf{s}_t, \\mathbf{a}_t, \\mathbf{s}_{t+1})" + }, + { + "bbox": [ + 304, + 185, + 545, + 329 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 185, + 545, + 329 + ], + "type": "inline_equation", + "content": "\\mathbf{s}_{t+1}" + }, + { + "bbox": [ + 304, + 185, + 545, + 329 + ], + "type": "text", + "content": " represents the next state of the environment due to " + }, + { + "bbox": [ + 304, + 185, + 545, + 329 + ], + "type": "inline_equation", + "content": "\\mathbf{a}_t" + }, + { + "bbox": [ + 304, + 185, + 545, + 329 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 330, + 545, + 497 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 330, + 545, + 497 + ], + "spans": [ + { + "bbox": [ + 304, + 330, + 545, + 497 + ], + "type": "text", + "content": "Egocentric Sensing As Depth Proxy. We aim to sample actions given by a policy to synthesize realistic human motions. Egocentric perception-driven motion synthesis should arguably use egocentric vision as input. However, depth rendering is costly and RL requires billions of samples to converge [3, 103]. Besides, directly training RL with visual data can be unstable [122]. We thereby use a cheap-to-compute egocentric sensing " + }, + { + "bbox": [ + 304, + 330, + 545, + 497 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_t" + }, + { + "bbox": [ + 304, + 330, + 545, + 497 + ], + "type": "text", + "content": " as a proxy for depth images as illustrated in Fig. 2. " + }, + { + "bbox": [ + 304, + 330, + 545, + 497 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 330, + 545, + 497 + ], + "type": "text", + "content": " rays are cast evenly from the midpoint of two eyeballs, i.e., the location of the egocentric camera. The field of view " + }, + { + "bbox": [ + 304, + 330, + 545, + 497 + ], + "type": "inline_equation", + "content": "[\\theta_{min},\\theta_{max}]" + }, + { + "bbox": [ + 304, + 330, + 545, + 497 + ], + "type": "text", + "content": ", centered on the 2D projection of the viewing direction " + }, + { + "bbox": [ + 304, + 330, + 545, + 497 + ], + "type": "inline_equation", + "content": "\\vec{\\nu}" + }, + { + "bbox": [ + 304, + 330, + 545, + 497 + ], + "type": "text", + "content": ", limits the agent's perception to the front area. Rays stop at collisions, with collision detection in 2D. See more details in Supp. Mat." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 498, + 545, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 498, + 545, + 545 + ], + "spans": [ + { + "bbox": [ + 304, + 498, + 545, + 545 + ], + "type": "text", + "content": "Agent Representation. The agent is a virtual human represented by an SMPL-X mesh [63]. We further compact the body representation by selecting " + }, + { + "bbox": [ + 304, + 498, + 545, + 545 + ], + "type": "inline_equation", + "content": "M = 67" + }, + { + "bbox": [ + 304, + 498, + 545, + 545 + ], + "type": "text", + "content": " body surface markers " + }, + { + "bbox": [ + 304, + 498, + 545, + 545 + ], + "type": "inline_equation", + "content": "\\mathbf{x} \\in \\mathbb{R}^{M \\times 3}" + }, + { + "bbox": [ + 304, + 498, + 545, + 545 + ], + "type": "text", + "content": " on the mesh following [118]." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 545, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 545, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 545, + 545, + 713 + ], + "type": "text", + "content": "Environment. Inspired by [108], we aim to learn a library of composable CAMPs. We implement a finite-horizon environment based on the generative motion primitive model from [117]. Specifically, a motion primitive is defined as a 0.5-second motion clip containing " + }, + { + "bbox": [ + 304, + 545, + 545, + 713 + ], + "type": "inline_equation", + "content": "T = 20" + }, + { + "bbox": [ + 304, + 545, + 545, + 713 + ], + "type": "text", + "content": " frames in the canonical coordinate, and each frame contains a single agent representation. The primitive model " + }, + { + "bbox": [ + 304, + 545, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 304, + 545, + 545, + 713 + ], + "type": "text", + "content": " is based on the C-VAE framework [88], which takes the first " + }, + { + "bbox": [ + 304, + 545, + 545, + 713 + ], + "type": "inline_equation", + "content": "T_{s} = 2" + }, + { + "bbox": [ + 304, + 545, + 545, + 713 + ], + "type": "text", + "content": " frames as the condition, and models a conditional probability of the next " + }, + { + "bbox": [ + 304, + 545, + 545, + 713 + ], + "type": "inline_equation", + "content": "T - T_{s}" + }, + { + "bbox": [ + 304, + 545, + 545, + 713 + ], + "type": "text", + "content": " frames. Compared to [117] trained on the AMASS dataset [48] with many sport motion sequences, we train " + }, + { + "bbox": [ + 304, + 545, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 304, + 545, + 545, + 713 + ], + "type": "text", + "content": " using the SAMP dataset [29], which focuses on daily activities, better suited for HMD use cases. Our action space " + }, + { + "bbox": [ + 304, + 545, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{A}" + }, + { + "bbox": [ + 304, + 545, + 545, + 713 + ], + "type": "text", + "content": " is the pretrained 128D latent space of " + }, + { + "bbox": [ + 304, + 545, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 304, + 545, + 545, + 713 + ], + "type": "text", + "content": ", and the" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "14499" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 47, + 68, + 547, + 246 + ], + "blocks": [ + { + "bbox": [ + 47, + 68, + 547, + 246 + ], + "lines": [ + { + "bbox": [ + 47, + 68, + 547, + 246 + ], + "spans": [ + { + "bbox": [ + 47, + 68, + 547, + 246 + ], + "type": "image", + "image_path": "13ff6411c240f09adaa375b54aefc78e76a61979c69f8352c6ed6db654396e22.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 252, + 547, + 289 + ], + "lines": [ + { + "bbox": [ + 46, + 252, + 547, + 289 + ], + "spans": [ + { + "bbox": [ + 46, + 252, + 547, + 289 + ], + "type": "text", + "content": "Figure 2. Policy network architecture. We learn a generalizable mapping from motion seed body markers " + }, + { + "bbox": [ + 46, + 252, + 547, + 289 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_t^S" + }, + { + "bbox": [ + 46, + 252, + 547, + 289 + ], + "type": "text", + "content": ", marker directions " + }, + { + "bbox": [ + 46, + 252, + 547, + 289 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_t^{S^D}" + }, + { + "bbox": [ + 46, + 252, + 547, + 289 + ], + "type": "text", + "content": ", egocentric sensing " + }, + { + "bbox": [ + 46, + 252, + 547, + 289 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_t" + }, + { + "bbox": [ + 46, + 252, + 547, + 289 + ], + "type": "text", + "content": ", and distance to the target " + }, + { + "bbox": [ + 46, + 252, + 547, + 289 + ], + "type": "inline_equation", + "content": "d_t" + }, + { + "bbox": [ + 46, + 252, + 547, + 289 + ], + "type": "text", + "content": " to CAMPs. The policy learns a stochastic collision avoiding action space to predict future body markers " + }, + { + "bbox": [ + 46, + 252, + 547, + 289 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_t^F" + }, + { + "bbox": [ + 46, + 252, + 547, + 289 + ], + "type": "text", + "content": ". For illustration purposes, we visualize only one frame of " + }, + { + "bbox": [ + 46, + 252, + 547, + 289 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_t^S" + }, + { + "bbox": [ + 46, + 252, + 547, + 289 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 252, + 547, + 289 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_t" + }, + { + "bbox": [ + 46, + 252, + 547, + 289 + ], + "type": "text", + "content": ". See Sec. 3.1 and 3.2 for details." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 303, + 228, + 315 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 303, + 228, + 315 + ], + "spans": [ + { + "bbox": [ + 46, + 303, + 228, + 315 + ], + "type": "text", + "content": "action " + }, + { + "bbox": [ + 46, + 303, + 228, + 315 + ], + "type": "inline_equation", + "content": "\\mathbf{a}_t" + }, + { + "bbox": [ + 46, + 303, + 228, + 315 + ], + "type": "text", + "content": " can be randomly sampled from " + }, + { + "bbox": [ + 46, + 303, + 228, + 315 + ], + "type": "inline_equation", + "content": "\\mathcal{A}" + }, + { + "bbox": [ + 46, + 303, + 228, + 315 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 315, + 288, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 315, + 288, + 376 + ], + "spans": [ + { + "bbox": [ + 46, + 315, + 288, + 376 + ], + "type": "text", + "content": "With the input of a random action " + }, + { + "bbox": [ + 46, + 315, + 288, + 376 + ], + "type": "inline_equation", + "content": "\\mathbf{a}_t\\in \\mathbb{R}^{128}" + }, + { + "bbox": [ + 46, + 315, + 288, + 376 + ], + "type": "text", + "content": " and a motion seed " + }, + { + "bbox": [ + 46, + 315, + 288, + 376 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_t^S = [\\mathbf{x}_t^0,\\mathbf{x}_t^{T_s - 1}]" + }, + { + "bbox": [ + 46, + 315, + 288, + 376 + ], + "type": "text", + "content": " (history frames), " + }, + { + "bbox": [ + 46, + 315, + 288, + 376 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 46, + 315, + 288, + 376 + ], + "type": "text", + "content": " predicts future frames " + }, + { + "bbox": [ + 46, + 315, + 288, + 376 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_t^F = [\\mathbf{x}_t^{T_s},\\dots ,\\mathbf{x}_t^{T - 1}]" + }, + { + "bbox": [ + 46, + 315, + 288, + 376 + ], + "type": "text", + "content": " of the current motion primitive " + }, + { + "bbox": [ + 46, + 315, + 288, + 376 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_t = [\\mathbf{X}_t^S,\\mathbf{X}_t^F ]\\in \\mathbb{R}^{T\\times M\\times 3}" + }, + { + "bbox": [ + 46, + 315, + 288, + 376 + ], + "type": "text", + "content": ", which represents a short sequence of human motion spanning " + }, + { + "bbox": [ + 46, + 315, + 288, + 376 + ], + "type": "inline_equation", + "content": "0.5\\mathrm{s}" + }, + { + "bbox": [ + 46, + 315, + 288, + 376 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 129, + 379, + 206, + 395 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 129, + 379, + 206, + 395 + ], + "spans": [ + { + "bbox": [ + 129, + 379, + 206, + 395 + ], + "type": "interline_equation", + "content": "\\mathbf {X} _ {t} ^ {F} = \\mathcal {P} (\\mathbf {X} _ {t} ^ {S}, \\mathbf {a} _ {t})", + "image_path": "109bd7f4555cff3ba10b2cc3217d4dae3352e7320b305368ed6a2f971f884a46.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 399, + 287, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 399, + 287, + 472 + ], + "spans": [ + { + "bbox": [ + 46, + 399, + 287, + 472 + ], + "type": "text", + "content": "State. To preserve Markov property [62], the state is defined as " + }, + { + "bbox": [ + 46, + 399, + 287, + 472 + ], + "type": "inline_equation", + "content": "\\mathbf{s}_t = \\{\\mathbf{X}_t^S, \\mathbf{X}_t^{S^D}, \\mathcal{E}_t, d_t, \\tau_t\\}" + }, + { + "bbox": [ + 46, + 399, + 287, + 472 + ], + "type": "text", + "content": ", in which " + }, + { + "bbox": [ + 46, + 399, + 287, + 472 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_t^{S^D} \\in \\mathbb{R}^{T_s \\times M \\times 3}" + }, + { + "bbox": [ + 46, + 399, + 287, + 472 + ], + "type": "text", + "content": " denotes the normalized direction of each marker seed to the target, " + }, + { + "bbox": [ + 46, + 399, + 287, + 472 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_t \\in \\mathbb{R}^{T_s \\times N}" + }, + { + "bbox": [ + 46, + 399, + 287, + 472 + ], + "type": "text", + "content": " denotes the egocentric sensing depth proxy, " + }, + { + "bbox": [ + 46, + 399, + 287, + 472 + ], + "type": "inline_equation", + "content": "d_t" + }, + { + "bbox": [ + 46, + 399, + 287, + 472 + ], + "type": "text", + "content": " denotes the distance from the pelvis to the target, and " + }, + { + "bbox": [ + 46, + 399, + 287, + 472 + ], + "type": "inline_equation", + "content": "\\tau_t" + }, + { + "bbox": [ + 46, + 399, + 287, + 472 + ], + "type": "text", + "content": " denotes the remaining time. See Fig. 2." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 472, + 287, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 472, + 287, + 533 + ], + "spans": [ + { + "bbox": [ + 46, + 472, + 287, + 533 + ], + "type": "text", + "content": "Reward. To synthesize egocentric perception behaviors, we use an \"attention reward\" to incentivize the virtual human to look in specific directions: " + }, + { + "bbox": [ + 46, + 472, + 287, + 533 + ], + "type": "inline_equation", + "content": "r_{\\text{attention}} = \\cos \\langle \\vec{\\mathbf{v}}, \\vec{\\mathbf{a}} \\rangle" + }, + { + "bbox": [ + 46, + 472, + 287, + 533 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 472, + 287, + 533 + ], + "type": "inline_equation", + "content": "\\vec{\\mathbf{a}}" + }, + { + "bbox": [ + 46, + 472, + 287, + 533 + ], + "type": "text", + "content": " is the attention direction from the head joint to the viewing target. The reward function is defined as:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 539, + 287, + 552 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 539, + 287, + 552 + ], + "spans": [ + { + "bbox": [ + 46, + 539, + 287, + 552 + ], + "type": "interline_equation", + "content": "r _ {t} = r _ {c o n t.} + r _ {d i s t} + r _ {o r i} + r _ {a t t e n t i o n} + r _ {p e n e} + r _ {p o s e} + r _ {s u c c},", + "image_path": "90ba247e710712433ffc454ebc4f746058587295d62f041f6937418cd43b3820.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 555, + 287, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 555, + 287, + 616 + ], + "spans": [ + { + "bbox": [ + 46, + 555, + 287, + 616 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 555, + 287, + 616 + ], + "type": "inline_equation", + "content": "r_{cont}" + }, + { + "bbox": [ + 46, + 555, + 287, + 616 + ], + "type": "text", + "content": " enforces valid foot contact and minimizes foot skating; " + }, + { + "bbox": [ + 46, + 555, + 287, + 616 + ], + "type": "inline_equation", + "content": "r_{dist}" + }, + { + "bbox": [ + 46, + 555, + 287, + 616 + ], + "type": "text", + "content": " encourages reaching the target; " + }, + { + "bbox": [ + 46, + 555, + 287, + 616 + ], + "type": "inline_equation", + "content": "r_{ori}" + }, + { + "bbox": [ + 46, + 555, + 287, + 616 + ], + "type": "text", + "content": " aligns the body forward direction with the target; " + }, + { + "bbox": [ + 46, + 555, + 287, + 616 + ], + "type": "inline_equation", + "content": "r_{pene}" + }, + { + "bbox": [ + 46, + 555, + 287, + 616 + ], + "type": "text", + "content": " guides collision avoidance; " + }, + { + "bbox": [ + 46, + 555, + 287, + 616 + ], + "type": "inline_equation", + "content": "r_{pose}" + }, + { + "bbox": [ + 46, + 555, + 287, + 616 + ], + "type": "text", + "content": " reduces unrealistic human poses; and " + }, + { + "bbox": [ + 46, + 555, + 287, + 616 + ], + "type": "inline_equation", + "content": "r_{succ}" + }, + { + "bbox": [ + 46, + 555, + 287, + 616 + ], + "type": "text", + "content": " is a sparse reward when reaching the target." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 616, + 287, + 663 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 616, + 287, + 663 + ], + "spans": [ + { + "bbox": [ + 46, + 616, + 287, + 663 + ], + "type": "text", + "content": "Episode Termination. To handle collisions beyond [119], we employ multiple termination signals to conclude an episode if the generated motion primitive " + }, + { + "bbox": [ + 46, + 616, + 287, + 663 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_t" + }, + { + "bbox": [ + 46, + 616, + 287, + 663 + ], + "type": "text", + "content": " satisfies any of the following conditions:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 663, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 47, + 663, + 246, + 675 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 663, + 246, + 675 + ], + "spans": [ + { + "bbox": [ + 47, + 663, + 246, + 675 + ], + "type": "text", + "content": "- Success: The virtual human reached the target." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 675, + 287, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 675, + 287, + 687 + ], + "spans": [ + { + "bbox": [ + 47, + 675, + 287, + 687 + ], + "type": "text", + "content": "- Penetration: The virtual human collides with the obstacle." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 47, + 687, + 287, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 687, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 47, + 687, + 287, + 712 + ], + "type": "text", + "content": "- Timeout: The virtual human did not reach the target within the maximum timesteps." + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 305, + 302, + 545, + 316 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 302, + 545, + 316 + ], + "spans": [ + { + "bbox": [ + 305, + 302, + 545, + 316 + ], + "type": "text", + "content": "3.2. Training Collision-Avoiding Stochastic Policies" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 321, + 545, + 417 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 321, + 545, + 417 + ], + "spans": [ + { + "bbox": [ + 304, + 321, + 545, + 417 + ], + "type": "text", + "content": "Algorithm. We use Proximal Policy Optimization [81] (PPO) to learn a generalizable mapping from various egocentric sensing and body configurations to CAMPs. Instead of extensive manual data collection for all possible input combinations, we leverage the stochastic nature of the PPO policy. Through exploration and sampling actions, the agent traverses the scene and generates varied egocentric sensing and body configurations, diversifying the training data." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 417, + 545, + 476 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 417, + 545, + 476 + ], + "spans": [ + { + "bbox": [ + 304, + 417, + 545, + 476 + ], + "type": "text", + "content": "Instead of training each CAMP independently for every single step, we use PPO to train a sequence of CAMPs jointly in multi-step collision avoidance tasks. This approach can benefit choosing a more favorable CAMP which makes the subsequent action easier." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 477, + 545, + 573 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 477, + 545, + 573 + ], + "spans": [ + { + "bbox": [ + 304, + 477, + 545, + 573 + ], + "type": "text", + "content": "Network. The network architecture is shown in Fig 2. The actor and critic network share a feature extraction trunk to encode the state " + }, + { + "bbox": [ + 304, + 477, + 545, + 573 + ], + "type": "inline_equation", + "content": "\\mathbf{s}_t" + }, + { + "bbox": [ + 304, + 477, + 545, + 573 + ], + "type": "text", + "content": ": the motion seed " + }, + { + "bbox": [ + 304, + 477, + 545, + 573 + ], + "type": "inline_equation", + "content": "(\\mathbf{X}_t^S" + }, + { + "bbox": [ + 304, + 477, + 545, + 573 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 477, + 545, + 573 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_t^{S^D}" + }, + { + "bbox": [ + 304, + 477, + 545, + 573 + ], + "type": "text", + "content": ") and the egocentric sensing " + }, + { + "bbox": [ + 304, + 477, + 545, + 573 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_t" + }, + { + "bbox": [ + 304, + 477, + 545, + 573 + ], + "type": "text", + "content": " are encoded using RNNs; the rest of scalar states are encoded using positional encoding [99]. The actor predicts a stochastic policy " + }, + { + "bbox": [ + 304, + 477, + 545, + 573 + ], + "type": "inline_equation", + "content": "\\mathbf{a}_t \\sim \\pi(\\mathbf{z}_t|\\mathbf{s}_t)" + }, + { + "bbox": [ + 304, + 477, + 545, + 573 + ], + "type": "text", + "content": " conditioned on the current state " + }, + { + "bbox": [ + 304, + 477, + 545, + 573 + ], + "type": "inline_equation", + "content": "\\mathbf{s}_t" + }, + { + "bbox": [ + 304, + 477, + 545, + 573 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 477, + 545, + 573 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_t \\sim \\mathcal{N}(\\mu, \\Sigma)" + }, + { + "bbox": [ + 304, + 477, + 545, + 573 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 304, + 477, + 545, + 573 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 304, + 477, + 545, + 573 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 477, + 545, + 573 + ], + "type": "inline_equation", + "content": "\\Sigma" + }, + { + "bbox": [ + 304, + 477, + 545, + 573 + ], + "type": "text", + "content": " are the mean and variance of the learned action space." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 573, + 545, + 621 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 573, + 545, + 621 + ], + "spans": [ + { + "bbox": [ + 304, + 573, + 545, + 621 + ], + "type": "text", + "content": "Objective Function. The objective function includes the policy surrogate " + }, + { + "bbox": [ + 304, + 573, + 545, + 621 + ], + "type": "inline_equation", + "content": "L^{CLIP}" + }, + { + "bbox": [ + 304, + 573, + 545, + 621 + ], + "type": "text", + "content": ", the value function error term " + }, + { + "bbox": [ + 304, + 573, + 545, + 621 + ], + "type": "inline_equation", + "content": "L^{VF}" + }, + { + "bbox": [ + 304, + 573, + 545, + 621 + ], + "type": "text", + "content": " to evaluate the value prediction " + }, + { + "bbox": [ + 304, + 573, + 545, + 621 + ], + "type": "inline_equation", + "content": "V_{\\theta}" + }, + { + "bbox": [ + 304, + 573, + 545, + 621 + ], + "type": "text", + "content": ", and an entropy bonus " + }, + { + "bbox": [ + 304, + 573, + 545, + 621 + ], + "type": "inline_equation", + "content": "L^{S}" + }, + { + "bbox": [ + 304, + 573, + 545, + 621 + ], + "type": "text", + "content": " to encourage exploration:" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 362, + 629, + 489, + 643 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 362, + 629, + 489, + 643 + ], + "spans": [ + { + "bbox": [ + 362, + 629, + 489, + 643 + ], + "type": "interline_equation", + "content": "L = L ^ {C L I P} + c _ {1} L ^ {V F} + c _ {2} L ^ {S}", + "image_path": "93d1d66636522b136014e5ce3dcde954d8b41b35d3aa2a1582b91886c489af04.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 654, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 545, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 654, + 545, + 714 + ], + "type": "inline_equation", + "content": "c_{1}, c_{2}" + }, + { + "bbox": [ + 304, + 654, + 545, + 714 + ], + "type": "text", + "content": " are coefficients. See more details in Supp. Mat. RL Pre-training and Finetuning. Training in crowded scenes, e.g. Replica [91], requires additional steps. Because the action space " + }, + { + "bbox": [ + 304, + 654, + 545, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{A}" + }, + { + "bbox": [ + 304, + 654, + 545, + 714 + ], + "type": "text", + "content": " is an unbounded Gaussian, RL exploration while predicting reasonable human poses can" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "14500" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "code", + "bbox": [ + 57, + 87, + 287, + 206 + ], + "blocks": [ + { + "bbox": [ + 48, + 72, + 286, + 84 + ], + "lines": [ + { + "bbox": [ + 48, + 72, + 286, + 84 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 286, + 84 + ], + "type": "text", + "content": "Algorithm 1 Crowd motion synthesis with learned CAMPs" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "code_caption" + }, + { + "bbox": [ + 57, + 87, + 287, + 206 + ], + "lines": [ + { + "bbox": [ + 57, + 87, + 287, + 206 + ], + "spans": [ + { + "bbox": [ + 57, + 87, + 287, + 206 + ], + "type": "text", + "content": "Result: Multi-human locomotion w/ collision avoidance; \nInit: crowd size " + }, + { + "bbox": [ + 57, + 87, + 287, + 206 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 57, + 87, + 287, + 206 + ], + "type": "text", + "content": ", marker seed for each human " + }, + { + "bbox": [ + 57, + 87, + 287, + 206 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_c^S" + }, + { + "bbox": [ + 57, + 87, + 287, + 206 + ], + "type": "text", + "content": "; \nfor step " + }, + { + "bbox": [ + 57, + 87, + 287, + 206 + ], + "type": "inline_equation", + "content": "\\leftarrow 1" + }, + { + "bbox": [ + 57, + 87, + 287, + 206 + ], + "type": "text", + "content": " to max_steps do env. finite horizon \nfor " + }, + { + "bbox": [ + 57, + 87, + 287, + 206 + ], + "type": "inline_equation", + "content": "c \\gets 1" + }, + { + "bbox": [ + 57, + 87, + 287, + 206 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 57, + 87, + 287, + 206 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 57, + 87, + 287, + 206 + ], + "type": "text", + "content": " do for each human update all locations with " + }, + { + "bbox": [ + 57, + 87, + 287, + 206 + ], + "type": "inline_equation", + "content": "\\{bbox(\\mathbf{X}_c^S)\\}_{c=1:C}" + }, + { + "bbox": [ + 57, + 87, + 287, + 206 + ], + "type": "text", + "content": " compute egocentric sensing " + }, + { + "bbox": [ + 57, + 87, + 287, + 206 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_c" + }, + { + "bbox": [ + 57, + 87, + 287, + 206 + ], + "type": "text", + "content": "; execute the action that maximizes the expected return, and produce one CAMP; end for \nend for" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "code_body" + } + ], + "index": 1, + "sub_type": "algorithm" + }, + { + "bbox": [ + 46, + 233, + 287, + 305 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 233, + 287, + 305 + ], + "spans": [ + { + "bbox": [ + 46, + 233, + 287, + 305 + ], + "type": "text", + "content": "be challenging. We first pretrain the policy with a higher " + }, + { + "bbox": [ + 46, + 233, + 287, + 305 + ], + "type": "inline_equation", + "content": "r_{pene}" + }, + { + "bbox": [ + 46, + 233, + 287, + 305 + ], + "type": "text", + "content": " weight without penetration termination. After convergence, we finetune it with strict termination constraints using a signed distance field (SDF) for penetration detection. Please refer to Supp. Mat. for the formulation and weighting of each reward and training detail." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 319, + 259, + 331 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 319, + 259, + 331 + ], + "spans": [ + { + "bbox": [ + 47, + 319, + 259, + 331 + ], + "type": "text", + "content": "3.3. Compositing Learned Motion Primitives" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 339, + 287, + 626 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 339, + 287, + 626 + ], + "spans": [ + { + "bbox": [ + 46, + 339, + 287, + 626 + ], + "type": "text", + "content": "Although CAMPs are trained solely with static scenes, their direct application to dynamic settings is achieved by decomposing jointly trained CAMPs into individual motion primitives and re-compositing them. Our model demonstrates effective generalization by selecting the next best motion primitive from the learned CAMP library to maximize the expected return, provided that the egocentric sensing is updated with the most recent obstacle location at each timestep. Furthermore, our model is directly applicable to tasks involving complex interactions with other virtual humans. To synthesize crowd motion (Alg. 1), each virtual human employs the same policy to navigate and avoid others. To a specific virtual human, others are seen as dynamic obstacles, represented by body bounding boxes for avoidance. Acknowledging the inherent delay in human reactions when avoiding dynamic obstacles [4], agents take a single CAMP sequentially, instead of in parallel, i.e. the first agent generates its first CAMP and waits for others to complete their first CAMP before all agents move on to prepare their second CAMP. To ensure successful collision avoidance, the agent's egocentric sensing is updated before taking a new action. This composition of CAMPs synthesizes emergent multi-human behaviors without multi-agent RL algorithms (see Sec. 5.1), enhancing the generalization and scalability." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 643, + 257, + 657 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 643, + 257, + 657 + ], + "spans": [ + { + "bbox": [ + 47, + 643, + 257, + 657 + ], + "type": "text", + "content": "4. Egocentric Synthetic Data Generation" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "text", + "content": "Synthesizing realistic egocentric perception-driven human motions (as detailed in Sec. 3) forms the foundation of simulating egocentric synthetic data. An overview of our egocentric data generation pipeline EgoGen, is shown in Fig. 3." + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 307, + 70, + 545, + 258 + ], + "blocks": [ + { + "bbox": [ + 307, + 70, + 545, + 258 + ], + "lines": [ + { + "bbox": [ + 307, + 70, + 545, + 258 + ], + "spans": [ + { + "bbox": [ + 307, + 70, + 545, + 258 + ], + "type": "image", + "image_path": "06018b31962397ed8eb422593026e3d987b0a8165b470871706a4442b584c09e.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 266, + 545, + 344 + ], + "lines": [ + { + "bbox": [ + 304, + 266, + 545, + 344 + ], + "spans": [ + { + "bbox": [ + 304, + 266, + 545, + 344 + ], + "type": "text", + "content": "Figure 3. Overview of EgoGen. Through generative motion synthesis (Sec. 3), we further enhance egocentric data diversity by randomly sampling diverse body textures (ethnicity, gender) and 3D textured clothing through an automated clothing simulation pipeline (Sec. 4.2). With high-quality scenes and different egocentric cameras, we can render photorealistic egocentric synthetic data with rich and accurate ground truth annotations (Sec. 4.3)." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 360, + 469, + 371 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 360, + 469, + 371 + ], + "spans": [ + { + "bbox": [ + 306, + 360, + 469, + 371 + ], + "type": "text", + "content": "4.1. Embodied Camera Placement" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 380, + 545, + 487 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 380, + 545, + 487 + ], + "spans": [ + { + "bbox": [ + 304, + 380, + 545, + 487 + ], + "type": "text", + "content": "Similar to existing AR devices, we use the head pose to define the egocentric viewing direction " + }, + { + "bbox": [ + 304, + 380, + 545, + 487 + ], + "type": "inline_equation", + "content": "\\vec{\\mathbf{v}}" + }, + { + "bbox": [ + 304, + 380, + 545, + 487 + ], + "type": "text", + "content": ". Our development is based on Blender [11]. We use the SMPL-X [63] mesh to position the egocentric camera between the two eyeballs. The camera's viewing direction " + }, + { + "bbox": [ + 304, + 380, + 545, + 487 + ], + "type": "inline_equation", + "content": "(\\vec{\\mathbf{v}})" + }, + { + "bbox": [ + 304, + 380, + 545, + 487 + ], + "type": "text", + "content": " is perpendicular to the plane determined by the two eye bones in the armature. We also support simulating multi-camera rigs as shown in Fig. 1. When the body moves (Sec. 3.3), we can synthesize egocentric videos with continuously updated camera poses." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 500, + 456, + 513 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 500, + 456, + 513 + ], + "spans": [ + { + "bbox": [ + 306, + 500, + 456, + 513 + ], + "type": "text", + "content": "4.2. Body Texture and Clothing" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 520, + 545, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 520, + 545, + 568 + ], + "spans": [ + { + "bbox": [ + 304, + 520, + 545, + 568 + ], + "type": "text", + "content": "To enhance EgoGen's synthetic data realism, we dress virtual humans using human textures and 3D clothing assets from BEDLAM [1, 8], including 50 male and 50 female skin albedo textures from seven ethnic groups." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "text", + "content": "Unlike prior works [8, 109] relying on unscalable commercial software for clothing dynamics simulation, we automate it for diverse synthesized motions and body shapes, minimizing manual effort. Each garment mesh is in a consistent rest pose, i.e., A-Pose (See Fig. 3 middle-left). For each motion sequence, we first repose it to match the body pose in the first frame using linear blend skinning. This involves initializing the clothing geometry by sampling pose and shape blend shapes, along with skinning weights from the nearest multiple SMPL-X vertices in A-Pose. Then we simulate upper and lower garments separately using a state-of-the-art clothing simulation network [24]." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14501" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 55, + 123, + 280, + 258 + ], + "blocks": [ + { + "bbox": [ + 47, + 70, + 287, + 115 + ], + "lines": [ + { + "bbox": [ + 47, + 70, + 287, + 115 + ], + "spans": [ + { + "bbox": [ + 47, + 70, + 287, + 115 + ], + "type": "text", + "content": "Table 1. Evaluation of motion synthesis in scenes with moving obstacles, multiple humans, and path diversity. " + }, + { + "bbox": [ + 47, + 70, + 287, + 115 + ], + "type": "inline_equation", + "content": "\\downarrow" + }, + { + "bbox": [ + 47, + 70, + 287, + 115 + ], + "type": "text", + "content": " : lower is better; " + }, + { + "bbox": [ + 47, + 70, + 287, + 115 + ], + "type": "inline_equation", + "content": "\\uparrow" + }, + { + "bbox": [ + 47, + 70, + 287, + 115 + ], + "type": "text", + "content": " : higher is better. The best results in each scenario are in boldface. * denotes an improved version for fair comparison. (Sec. 5.1)" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 55, + 123, + 280, + 258 + ], + "lines": [ + { + "bbox": [ + 55, + 123, + 280, + 258 + ], + "spans": [ + { + "bbox": [ + 55, + 123, + 280, + 258 + ], + "type": "table", + "html": "
EvaluationMetricsGAMMA* [117]DIMOS* [119]Ours
Mov. obs.SR (%) ↑9683100
Dist. (m) ↓0.290.550.06
Cont. ↑0.950.960.97
Pene-S. (%) ↓9.28.43.4
2 humansSR (%) ↑9588100
Dist. (m) ↓0.320.410.07
Cont. ↑0.960.980.97
Pene-H. ↓27.610.70
4 humansSR (%) ↑9270100
Dist. (m) ↓0.410.790.07
Cont. ↑0.940.950.96
Pene-H. ↓60.441.70
DiversitySR (%) ↑968497
Std Dev ↑0.9871.051.21
", + "image_path": "cd9e91a4dadcea46de16ec657329bf2a75b5be609615d8ad07ae3aa8897f4b31.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 268, + 200, + 280 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 268, + 200, + 280 + ], + "spans": [ + { + "bbox": [ + 47, + 268, + 200, + 280 + ], + "type": "text", + "content": "4.3. Rendering and Annotations" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 286, + 287, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 286, + 287, + 430 + ], + "spans": [ + { + "bbox": [ + 46, + 286, + 287, + 430 + ], + "type": "text", + "content": "EgoGen supports simulating diverse head-mounted devices with different camera models, such as fisheye and pinhole cameras. Given the camera's intrinsic parameters and relative poses within the camera rig, we can simulate AR devices like Project Aria glasses [50] and HoloLens [51], facilitating synthetic data generation for real-world applications. Camera extrinsic is determined by our generative human motion model. We use Blender [11] to render photorealistic egocentric image sequences with motion blur. We also render out a rich set of ground truth annotations, such as depth maps, surface normals, segmentation masks, world positions, optical flow, etc for egocentric perception tasks." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 441, + 128, + 453 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 441, + 128, + 453 + ], + "spans": [ + { + "bbox": [ + 47, + 441, + 128, + 453 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 460, + 287, + 532 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 460, + 287, + 532 + ], + "spans": [ + { + "bbox": [ + 46, + 460, + 287, + 532 + ], + "type": "text", + "content": "We assess the motion quality, generalizability, and diversity of our motion model, highlighting its ability to generalize to unseen complex tasks and comparing it with recent baselines (Sec. 5.1). We evaluate our proposed egocentric sensing as a depth proxy for enhancing agent exploration (Sec. 5.2) and conduct ablation studies (Sec. 5.3)." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 532, + 287, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 532, + 287, + 581 + ], + "spans": [ + { + "bbox": [ + 46, + 532, + 287, + 581 + ], + "type": "text", + "content": "We further demonstrate the effectiveness of EgoGen on three egocentric computer vision tasks in Sec. 5.4, and 5.5. By incorporating synthesized egocentric images, we can enhance the performance of the state-of-the-art algorithms." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 587, + 215, + 599 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 587, + 215, + 599 + ], + "spans": [ + { + "bbox": [ + 47, + 587, + 215, + 599 + ], + "type": "text", + "content": "5.1. Evaluation of Learned CAMPs" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 605, + 287, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 287, + 689 + ], + "type": "text", + "content": "We assess CAMPs' generalizability in dynamic scenes, including scenes with moving obstacles and scenes with multiple individuals. In tests with moving obstacles, the obstacle blocks the person's path by moving between the person and the goal. In multiple human test scenes, lines from their starting and goal locations intersect in the middle, requiring solving human-human penetrations. See detail in Sup. Mat." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "content": "In Tab. 1, we compare goal-reaching behaviors with two recent baselines: GAMMA [117] and DIMOS [119]. Base" + } + ] + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 310, + 90, + 536, + 141 + ], + "blocks": [ + { + "bbox": [ + 330, + 71, + 522, + 82 + ], + "lines": [ + { + "bbox": [ + 330, + 71, + 522, + 82 + ], + "spans": [ + { + "bbox": [ + 330, + 71, + 522, + 82 + ], + "type": "text", + "content": "Table 2. Evaluation of egocentric sensing. (Sec. 5.2)" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 310, + 90, + 536, + 141 + ], + "lines": [ + { + "bbox": [ + 310, + 90, + 536, + 141 + ], + "spans": [ + { + "bbox": [ + 310, + 90, + 536, + 141 + ], + "type": "table", + "html": "
Method (sensing range)SR (%) ↑Dist. (m) ↓
Local map [119] (0.8 m)780.35
Local map* (7 m)43.04
Egocentric sensing (ours) (7 m)950.12
", + "image_path": "3f089a401ed28703c01383da3789c575ae76917eafb30193c6b53e2c8b2839f1.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 157, + 545, + 337 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 157, + 545, + 337 + ], + "spans": [ + { + "bbox": [ + 304, + 157, + 545, + 337 + ], + "type": "text", + "content": "line methods use navigation meshes and path planning for static scenes, while CAMPs can autonomously avoid dynamic obstacles (Sec. 3.3). For fair comparison in dynamic scenes, we extend the baselines by updating navigation meshes and performing on-the-fly path planning at each time step. The tree-based search as in [117] is disabled for all the methods. Metrics: (1) " + }, + { + "bbox": [ + 304, + 157, + 545, + 337 + ], + "type": "inline_equation", + "content": "SR" + }, + { + "bbox": [ + 304, + 157, + 545, + 337 + ], + "type": "text", + "content": ": Success rate for reaching the goal location within a " + }, + { + "bbox": [ + 304, + 157, + 545, + 337 + ], + "type": "inline_equation", + "content": "0.3\\mathrm{m}" + }, + { + "bbox": [ + 304, + 157, + 545, + 337 + ], + "type": "text", + "content": " threshold. (2) Dist.: Average distance of the final pelvis location to the goal. (3) Cont.: The contact metric [117] that measures foot-floor contact and foot skating. (4) Pene-S.: Percentage of frames with detected human-scene penetration in moving obstacle scenes. (5) Pene-H.: Accurate human-human penetration evaluation metric using COAP [52] in multiple human scenes. Please refer to Supp. Mat. for metric details." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 339, + 545, + 448 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 339, + 545, + 448 + ], + "spans": [ + { + "bbox": [ + 304, + 339, + 545, + 448 + ], + "type": "text", + "content": "CAMPs outperform the two baselines in dynamic scenarios with moving obstacles and multiple humans, exhibiting lower human-scene and human-human penetrations and a higher goal-reaching success rate. In multiple human scenarios, we observe that in the baselines, dynamically redoing path planning for each human independently can not effectively solve human-human penetration. In contrast, com-posable CAMPs can generalize well in dynamic settings without using multi-agent RL to synthesize crowd motions." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 450, + 545, + 546 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 450, + 545, + 546 + ], + "spans": [ + { + "bbox": [ + 304, + 450, + 545, + 546 + ], + "type": "text", + "content": "We assess walking path diversity using the standard deviation of pelvis locations for the same start-target pairs in scenes with a single static box obstacle. As shown in Tab. 1 (Diversity), our approach does not require a pre-computed global path and allows agents to self-exlore without being constrained by predefined paths, achieving higher walking path diversity and success rate. This fosters diverse synthetic data generation via more diverse synthesized motion." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 561, + 484, + 574 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 561, + 484, + 574 + ], + "spans": [ + { + "bbox": [ + 306, + 561, + 484, + 574 + ], + "type": "text", + "content": "5.2. Evaluation of Egocentric Sensing" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "content": "We assess the exploration ability of our egocentric sensing " + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_t" + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "content": " in Replica [91] scenes. In Tab. 2, we replace " + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_t" + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "content": " with a local map [119] in our state " + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{s}_t" + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "content": ", following their encoding method. Relying on local information can trap agents in local optima, e.g., walls beyond their sensing range, resulting in lower SR. Our egocentric sensing acts as a depth proxy, allowing the agent to avoid local optima, explore more effectively than local maps [119] or scandots [3], and achieve higher SR. In addition, our compact representation is more scalable as the sensing range increases, while quadratic local map growth can hinder the policy network's learning." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14502" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 52, + 101, + 277, + 162 + ], + "blocks": [ + { + "bbox": [ + 47, + 70, + 287, + 93 + ], + "lines": [ + { + "bbox": [ + 47, + 70, + 287, + 93 + ], + "spans": [ + { + "bbox": [ + 47, + 70, + 287, + 93 + ], + "type": "text", + "content": "Table 3. Ablation studies. Note: in our observation, " + }, + { + "bbox": [ + 47, + 70, + 287, + 93 + ], + "type": "inline_equation", + "content": "\\| VP\\| _2 > 15" + }, + { + "bbox": [ + 47, + 70, + 287, + 93 + ], + "type": "text", + "content": " indicates abnormal human poses. (Sec. 5.3)" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 52, + 101, + 277, + 162 + ], + "lines": [ + { + "bbox": [ + 52, + 101, + 277, + 162 + ], + "spans": [ + { + "bbox": [ + 52, + 101, + 277, + 162 + ], + "type": "table", + "html": "
SR (%)↑||VP||2↓cos(▼,▲)↑
Egocentric depth813.640.049
No pretraining9028.770.918
No attention reward9012.260.891
Our policy9210.570.940
", + "image_path": "f9ccde6c7f475f2c74da82ae063fbb22ab61c88e988f644dcc5ff7c39f5c6ffe.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 174, + 147, + 186 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 174, + 147, + 186 + ], + "spans": [ + { + "bbox": [ + 47, + 174, + 147, + 186 + ], + "type": "text", + "content": "5.3. Ablation Studies" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 193, + 272, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 193, + 272, + 204 + ], + "spans": [ + { + "bbox": [ + 47, + 193, + 272, + 204 + ], + "type": "text", + "content": "We compare our policy with several ablations in Tab. 3:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 205, + 287, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 205, + 287, + 240 + ], + "spans": [ + { + "bbox": [ + 46, + 205, + 287, + 240 + ], + "type": "text", + "content": "Egocentric depth: an ablation training an egocentric depth image-based policy without the depth sensing proxy. Egocentric depth images are encoded with a CNN;" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 241, + 287, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 241, + 287, + 277 + ], + "spans": [ + { + "bbox": [ + 47, + 241, + 287, + 277 + ], + "type": "text", + "content": "No pretraining: an ablation training collision avoidance in crowded scenes with strict penetration termination directly; No attention reward: an ablation for the viewing direction." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 277, + 287, + 325 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 277, + 287, + 325 + ], + "spans": [ + { + "bbox": [ + 46, + 277, + 287, + 325 + ], + "type": "text", + "content": "We assess pose naturalness with the maximum pose embedding norm encoded with VPoser [63] and evaluate the attention reward with the cosine similarity between the viewing direction " + }, + { + "bbox": [ + 46, + 277, + 287, + 325 + ], + "type": "inline_equation", + "content": "\\vec{\\mathbf{v}}" + }, + { + "bbox": [ + 46, + 277, + 287, + 325 + ], + "type": "text", + "content": " and the attention direction " + }, + { + "bbox": [ + 46, + 277, + 287, + 325 + ], + "type": "inline_equation", + "content": "\\vec{\\mathbf{a}}" + }, + { + "bbox": [ + 46, + 277, + 287, + 325 + ], + "type": "text", + "content": " (Sec. 3.1)." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 326, + 287, + 458 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 326, + 287, + 458 + ], + "spans": [ + { + "bbox": [ + 46, + 326, + 287, + 458 + ], + "type": "text", + "content": "Directly training RL with egocentric depth images is ineffective due to our high-dimensional action space, emphasizing the value of the compact egocentric sensing representation. Training agents with strict penetration constraints in crowded scenes directly can result in exploring unreasonable action subspaces, given its unbounded Gaussian nature, leading to unrealistic human poses, highlighting the effectiveness of our two-stage RL training scheme. Without the attention reward, the virtual human's capability to attend to a specific direction decreases. All ablation studies are evaluated in Replica. See visuals in Supp. Vid. and Supp. Mat." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 466, + 287, + 479 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 466, + 287, + 479 + ], + "spans": [ + { + "bbox": [ + 47, + 466, + 287, + 479 + ], + "type": "text", + "content": "5.4. Mapping, Localization, and Tracking for HMD" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 485, + 287, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 485, + 287, + 616 + ], + "spans": [ + { + "bbox": [ + 46, + 485, + 287, + 616 + ], + "type": "text", + "content": "Mapping and localization. LaMAR [76] is the first mapping and localization benchmark dataset for AR in large-scale scenes. Despite over a year of extensive data collection, the dataset still lacks exhaustive scene coverage, especially in large open spaces. EgoGen can let virtual humans explore large-scale scenes, render dense egocentric views, and build a more complete SfM map by extracting image feature points with SuperPoint [15] and matching images with SuperGlue [75]. Despite synthetic images being noisier due to scene quality, SuperGlue [75] matching can filter out noisy feature points and yield reliable matches." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 618, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 618, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 618, + 287, + 714 + ], + "type": "text", + "content": "In Tab. 4, we evaluate EgoGen by assessing the localization recall at " + }, + { + "bbox": [ + 46, + 618, + 287, + 714 + ], + "type": "inline_equation", + "content": "(1^{\\circ}, 10cm)" + }, + { + "bbox": [ + 46, + 618, + 287, + 714 + ], + "type": "text", + "content": " on the validation set in a lobby of " + }, + { + "bbox": [ + 46, + 618, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\sim 120" + }, + { + "bbox": [ + 46, + 618, + 287, + 714 + ], + "type": "text", + "content": " sqm of the LaMAR CAB location. In addition, we report the number of triangulated 3D points (#P3D) and track length. EgoGen improves the 3D reconstruction by yielding more points for a slightly improved track length and also a significantly better localization performance compared to using the real data only. Ng et al. [58] augments mapping" + } + ] + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 311, + 123, + 536, + 174 + ], + "blocks": [ + { + "bbox": [ + 305, + 70, + 545, + 115 + ], + "lines": [ + { + "bbox": [ + 305, + 70, + 545, + 115 + ], + "spans": [ + { + "bbox": [ + 305, + 70, + 545, + 115 + ], + "type": "text", + "content": "Table 4. Mapping and localization evaluation. We augment LaMAR with the same amount of images (248 frames) and report the localization recall at " + }, + { + "bbox": [ + 305, + 70, + 545, + 115 + ], + "type": "inline_equation", + "content": "(1^{\\circ},10cm)" + }, + { + "bbox": [ + 305, + 70, + 545, + 115 + ], + "type": "text", + "content": " on the validation set. EgoGen achieves the highest track length and recall. (Sec. 5.4)" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 311, + 123, + 536, + 174 + ], + "lines": [ + { + "bbox": [ + 311, + 123, + 536, + 174 + ], + "spans": [ + { + "bbox": [ + 311, + 123, + 536, + 174 + ], + "type": "table", + "html": "
#P3D ↑Track length ↑Recall (%) ↑
LaMAR19297395.194666.9
Ng et al. [58]19377585.194074.9
EgoGen19361695.210576.7
", + "image_path": "2156192db9bd92b1b053820fcaf8d5a26c0c2109a9791b25adbeec780be69440.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_body" + } + ], + "index": 12 + }, + { + "type": "table", + "bbox": [ + 311, + 210, + 536, + 249 + ], + "blocks": [ + { + "bbox": [ + 305, + 179, + 545, + 201 + ], + "lines": [ + { + "bbox": [ + 305, + 179, + 545, + 201 + ], + "spans": [ + { + "bbox": [ + 305, + 179, + 545, + 201 + ], + "type": "text", + "content": "Table 5. Egocentric camera tracking evaluation of models trained with and without synthetic data from EgoGen. (Sec. 5.4)" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 311, + 210, + 536, + 249 + ], + "lines": [ + { + "bbox": [ + 311, + 210, + 536, + 249 + ], + "spans": [ + { + "bbox": [ + 311, + 210, + 536, + 249 + ], + "type": "table", + "html": "
Pose ↓Rotation ↓Transl (mm) ↓
Scratch1.830.741303
+ EgoGen pretrain1.670.621305
", + "image_path": "f3e7593e91b2cd328a8b31d399b1f798261a010da8696f8b0a762311fd59aa07.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "table_body" + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 258, + 545, + 390 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 258, + 545, + 390 + ], + "spans": [ + { + "bbox": [ + 304, + 258, + 545, + 390 + ], + "type": "text", + "content": "images by perturbing real-world camera poses with noise, which may generate unrealistic camera poses (e.g., stuck in a wall or facing the ceiling), limiting egocentric localization effectiveness. Their method also assumes the availability of initial camera poses, which may not always be feasible. In contrast, EgoGen augments by virtual humans randomly exploring scenes. Our approach holds promise for creating AR mapping and localization datasets for digital twin scenes without manual data collection, providing enhanced privacy preservation, e.g. no need for anonymization. Refer to Supp. Mat. for visualization and implementation details." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 391, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 391, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 391, + 545, + 713 + ], + "type": "text", + "content": "Egocentric camera tracking. Egocentric camera tracking for HMD aims to yield device pose trajectories in 3D scenes given egocentric video observations. Recovering camera poses from monocular RGB videos using SLAM [93] is a challenging and ill-posed problem due to scale ambiguity. EgoEgo [43] leverages the knowledge of human motion to address the egocentric HMD tracking problem. Specifically, EgoEgo trains a neural network to infer the translation scaling and rotations from egocentric videos, which improves the HMD tracking performance. However, training this model requires jointly captured data of ground truth HMD trajectories and egocentric videos, which are costly to collect. We address this limitation by using EgoGen to synthesize quantities of egocentric videos with accurate camera trajectories to pretrain the model, which proves to improve the tracking performance on real data. We conduct experiments on the GIMO [121] dataset that contains " + }, + { + "bbox": [ + 304, + 391, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\sim 200" + }, + { + "bbox": [ + 304, + 391, + 545, + 713 + ], + "type": "text", + "content": " short sequences of paired motion-video data in 19 scenes. Using EgoGen, we synthesize " + }, + { + "bbox": [ + 304, + 391, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\sim 4k" + }, + { + "bbox": [ + 304, + 391, + 545, + 713 + ], + "type": "text", + "content": " sequences of human movements in their scenes and render corresponding egocentric videos using the same camera intrinsic as GIMO and the embodied camera placement described in Sec. 4.1. We also slightly perturb the camera placement location and orientation to simulate the diversity of how people wear HMDs in real data and avoid overfitting to one specific camera placement. We first pretrain the model with synthetic data generated by EgoGen, then finetune it on the real GIMO data." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14503" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 288, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 288, + 144 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 288, + 144 + ], + "type": "text", + "content": "Tab. 5 shows the egocentric camera tracking performances for models trained with and without synthetic data. Definitions of evaluation metrics can be found in Supp. Mat. The finetuned model benefits from EgoGen synthetic data and predicts more accurate camera poses compared to the model trained using real data only." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 153, + 287, + 166 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 153, + 287, + 166 + ], + "spans": [ + { + "bbox": [ + 47, + 153, + 287, + 166 + ], + "type": "text", + "content": "5.5. Human Mesh Recovery from Egocentric views" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 172, + 287, + 435 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 172, + 287, + 435 + ], + "spans": [ + { + "bbox": [ + 46, + 172, + 287, + 435 + ], + "type": "text", + "content": "Human mesh recovery (HMR) is the key to human behavior understanding from the egocentric view, thus crucial for applications in robotics and AR/VR. Given an egocentric RGB or depth image of a target subject, HMR aims to reconstruct the subject's 3D body pose and shape. However, acquiring and annotating real-world data is expensive, demanding, and time-consuming, with egocentric data being particularly scarce. EgoBody [115] is a recent egocentric dataset capturing two-people interactions, with egocentric depth/RGB frames annotated with SMPL-X body meshes. EgoBody provides " + }, + { + "bbox": [ + 46, + 172, + 287, + 435 + ], + "type": "inline_equation", + "content": "\\sim 180\\mathrm{k}" + }, + { + "bbox": [ + 46, + 172, + 287, + 435 + ], + "type": "text", + "content": " egocentric RGB frames, and merely " + }, + { + "bbox": [ + 46, + 172, + 287, + 435 + ], + "type": "inline_equation", + "content": "\\sim 23\\mathrm{k}" + }, + { + "bbox": [ + 46, + 172, + 287, + 435 + ], + "type": "text", + "content": " depth frames due to the low frame rate of the depth sensor, with " + }, + { + "bbox": [ + 46, + 172, + 287, + 435 + ], + "type": "inline_equation", + "content": "\\sim 90\\mathrm{k} / \\sim 10\\mathrm{k}" + }, + { + "bbox": [ + 46, + 172, + 287, + 435 + ], + "type": "text", + "content": " in the RGB/depth training set. Such limited data is insufficient to train a learning-based model from scratch. In contrast, with EgoGen, largescale synthetic egocentric data can be generated in a time-efficient way. We leverage EgoGen to generate quantities of training frames (300k RGB, 105k depth) of humans moving in EgoBody 3D scenes, rendered from the egocentric view, and annotated by SMPL-X parameters of the target subject. Specifically, RGB images are rendered with lifelike human body textures and 3D clothing, with random lighting." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 437, + 287, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 437, + 287, + 533 + ], + "spans": [ + { + "bbox": [ + 46, + 437, + 287, + 533 + ], + "type": "text", + "content": "With the recent HMR regressor, ProHMR [41], we show that pre-training with our synthetic data from EgoGen enhances the existing method's capability to generalize on real-world scenarios. Evaluated on the real-world EgoBody test set, we compare two training schemes: (1) trained from scratch on the real-world EgoBody training set (\"-scratch\"), and (2) pre-trained on synthetic data from EgoGen and finetuned on the real-world EgoBody training set (\"-ft\")." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 533, + 287, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 533, + 287, + 687 + ], + "spans": [ + { + "bbox": [ + 46, + 533, + 287, + 687 + ], + "type": "text", + "content": "HMR from depth. As no existing methods were proposed for depth-based HMR task, we adapt ProHMR [41] to the depth input by changing the channel number of the first convolution layer. To mimic real-world sensor noise, synthetic noise [27] is added to the rendered depth. G-MPJPE is additionally reported for depth-based HMR as depth images provide global information. As shown in Tab. 6, compared to the model trained only with a limited amount of real-world data (Depth-scratch), errors are significantly reduced for the model pre-trained with our large-scale synthetic data (Depth-ft), in terms of global translation (22.9% lower G-MPJPE), local pose (20.7% lower MJPE), and body shape (19.5% lower V2V)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "content": "HMR from RGB. For training with RGB images, we apply various data augmentation techniques similar to [8]." + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 310, + 118, + 536, + 172 + ], + "blocks": [ + { + "bbox": [ + 305, + 71, + 545, + 114 + ], + "lines": [ + { + "bbox": [ + 305, + 71, + 545, + 114 + ], + "spans": [ + { + "bbox": [ + 305, + 71, + 545, + 114 + ], + "type": "text", + "content": "Table 6. Evaluation of HMR on EgoBody test set. “*-scratch” denotes the model trained from scratch with the Egobody training set, and “*-ft” denotes the model pre-trained with EgoGen synthetic data. The units for all metrics are in mm. (Sec. 5.5)" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 310, + 118, + 536, + 172 + ], + "lines": [ + { + "bbox": [ + 310, + 118, + 536, + 172 + ], + "spans": [ + { + "bbox": [ + 310, + 118, + 536, + 172 + ], + "type": "table", + "html": "
G-MPJPE ↓MPJPE ↓PA-MPJPE ↓V2V ↓
Depth-scratch117.782.254.1100.6
Depth-ft90.765.247.381.0
RGB-scratch-90.759.9102.1
RGB-ft-85.356.297.2
", + "image_path": "5a0cccada08b1fdf35aeb231140523d5c0db7c8b38e2f38ef203b0c0c29eefe4.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 180, + 545, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 180, + 545, + 228 + ], + "spans": [ + { + "bbox": [ + 304, + 180, + 545, + 228 + ], + "type": "text", + "content": "Tab. 6 indicates that the RGB-based model pre-trained with large-scale synthetic data (\"RGB-ft\") also outperforms the model trained only on real-world data (\"RGB-scratch\"), for both body pose and shape accuracy." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 228, + 545, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 228, + 545, + 300 + ], + "spans": [ + { + "bbox": [ + 304, + 228, + 545, + 300 + ], + "type": "text", + "content": "The enhanced performance highlights that EgoGen's synthetic data effectively compensates for the lack of real-world training data, boosting the performance of current methods when test on real-world data. We will release both of our synthetic EgoBody datasets. See dataset statistics, qualitative visualizations, and training details in Supp. Mat." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 311, + 470, + 323 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 311, + 470, + 323 + ], + "spans": [ + { + "bbox": [ + 306, + 311, + 470, + 323 + ], + "type": "text", + "content": "6. Conclusion and Future Work" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 331, + 545, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 331, + 545, + 462 + ], + "spans": [ + { + "bbox": [ + 304, + 331, + 545, + 462 + ], + "type": "text", + "content": "We propose a novel egocentric synthetic data generation approach, EgoGen, that uses embodied sensors, a parametric body model, and a generative egocentric perception-driven human motion synthesis method to create egocentric training data with accurate and rich ground truth annotations. By integrating deep reinforcement learning and collision-avoiding motion primitives with egocentric depth proxy, EgoGen synthesizes robust human motion and emergent multi-agent behaviors. This paves the way to an efficient and scalable data generation solution that may have a profound impact on egocentric perception tasks." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 463, + 545, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 463, + 545, + 593 + ], + "spans": [ + { + "bbox": [ + 304, + 463, + 545, + 593 + ], + "type": "text", + "content": "Human-scene interaction in EgoGen is currently coarse. We aim to extend the current method to simulate more detailed human motion driven by egocentric perception, such as hand manipulation, sitting, lying, etc, to facilitate more realistic egocentric synthetic data. We use fixed attention goals to model human attention. Predicting human intention through historical egocentric perception and synthesizing viewing directions based on predicted intention holds significant potential. Synthesizing gaze direction for predicting human intent is valuable but presently hampered by data requirements; we will revisit this when resources allow." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 594, + 545, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 545, + 665 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 545, + 665 + ], + "type": "text", + "content": "We will explore many other egocentric vision tasks with EgoGen as this area grows rapidly such as social understanding and forecasting. EgoGen could benefit humanrobot interaction, e.g., our generative human motion model and lifelike human appearances can be integrated into [67] to close the sim2real gap for robotic agents further." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 665, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 545, + 712 + ], + "type": "text", + "content": "Acknowledgements. This work is supported by the SNSF project grant 200021 204840 and the SDSC PhD fellowship. We sincerely thank Linfei Pan, Yunke Ao, Qianli Ma, and Maxime Raafat for the valuable discussions." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14504" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 53, + 91, + 286, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 58, + 91, + 286, + 112 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 91, + 286, + 112 + ], + "spans": [ + { + "bbox": [ + 58, + 91, + 286, + 112 + ], + "type": "text", + "content": "[1] Meshcapade GmbH, Tübingen, Germany. https://meshcapade.com, 2022.5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 115, + 286, + 179 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 115, + 286, + 179 + ], + "spans": [ + { + "bbox": [ + 58, + 115, + 286, + 179 + ], + "type": "text", + "content": "[2] Mohammed Abdel-Wahab, Konrad Wenzel, and Dieter Fritsch. Efficient reconstruction of large unordered image datasets for high accuracy photogrammetric applications. In ISPRS Annals of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Melbourne, Australia. XXII ISPRS Congress, 2012. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 58, + 182, + 286, + 236 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 182, + 286, + 236 + ], + "spans": [ + { + "bbox": [ + 58, + 182, + 286, + 236 + ], + "type": "text", + "content": "[3] Ananye Agarwal, Ashish Kumar, Jitendra Malik, and Deepak Pathak. Legged locomotion in challenging terrains using egocentric vision. In Conference on Robot Learning, CoRL 2022, 14-18 December 2022, Auckland, New Zealand, pages 403-415. PMLR, 2022. 3, 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 58, + 239, + 286, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 239, + 286, + 270 + ], + "spans": [ + { + "bbox": [ + 58, + 239, + 286, + 270 + ], + "type": "text", + "content": "[4] M. Pilar Aivar, Eli Brenner, and Jeroen B. J. Smeets. Avoiding moving obstacles. Experimental Brain Research, 190 (3):251-264, 2008. 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 58, + 273, + 286, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 273, + 286, + 327 + ], + "spans": [ + { + "bbox": [ + 58, + 273, + 286, + 327 + ], + "type": "text", + "content": "[5] Hiroyasu Akada, Jian Wang, Soshi Shimada, Masaki Takahashi, Christian Theobalt, and Vladislav Golyanik. Unrealego: A new dataset for robust egocentric 3d human motion capture. In European Conference on Computer Vision (ECCV), 2022. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 58, + 329, + 286, + 350 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 329, + 286, + 350 + ], + "spans": [ + { + "bbox": [ + 58, + 329, + 286, + 350 + ], + "type": "text", + "content": "[6] Apple. ARKit. https://developer.apple.com/arkit/, 2017.3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 58, + 353, + 286, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 353, + 286, + 407 + ], + "spans": [ + { + "bbox": [ + 58, + 353, + 286, + 407 + ], + "type": "text", + "content": "[7] Sonia Baltodano, Srinath Sibi, Nikolas Martelaro, Nikhil Gowda, and Wendy Ju. The rroads platform: a real road autonomous driving simulator. In Proceedings of the 7th International Conference on Automotive User Interfaces and Interactive Vehicular Applications, pages 281-288, 2015. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 58, + 409, + 286, + 462 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 409, + 286, + 462 + ], + "spans": [ + { + "bbox": [ + 58, + 409, + 286, + 462 + ], + "type": "text", + "content": "[8] Michael J. Black, Priyanka Patel, Joachim Tesch, and Jinlong Yang. BEDLAM: A synthetic dataset of bodies exhibiting detailed lifelike animated motion. In Proceedings IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR), 2023. 2, 5, 8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 58, + 465, + 286, + 519 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 465, + 286, + 519 + ], + "spans": [ + { + "bbox": [ + 58, + 465, + 286, + 519 + ], + "type": "text", + "content": "[9] Federica Bogo, Angjoo Kanazawa, Christoph Lassner, Peter Gehler, Javier Romero, and Michael J. Black. Keep it SMPL: Automatic estimation of 3D human pose and shape from a single image. In European Conference on Computer Vision, pages 561-578, 2016. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 53, + 521, + 286, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 521, + 286, + 542 + ], + "spans": [ + { + "bbox": [ + 53, + 521, + 286, + 542 + ], + "type": "text", + "content": "[10] Simon Clavet. Motion matching and the road to next-gen animation. In Proc. of GDC, 2016. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 53, + 544, + 286, + 576 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 544, + 286, + 576 + ], + "spans": [ + { + "bbox": [ + 53, + 544, + 286, + 576 + ], + "type": "text", + "content": "[11] Blender Online Community. Blender - a 3D modelling and rendering package. Blender Foundation, Stichting Blender Foundation, Amsterdam, 2018. 5, 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 53, + 579, + 286, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 579, + 286, + 611 + ], + "spans": [ + { + "bbox": [ + 53, + 579, + 286, + 611 + ], + "type": "text", + "content": "[12] Erwin Coumans and Yunfei Bai. Pybullet, a python module for physics simulation for games, robotics and machine learning. http://pybullet.org, 2016-2021. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 53, + 614, + 286, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 614, + 286, + 689 + ], + "spans": [ + { + "bbox": [ + 53, + 614, + 286, + 689 + ], + "type": "text", + "content": "[13] Dima Damen, Hazel Doughty, Giovanni Maria Farinella, Sanja Fidler, Antonino Furnari, Evangelos Kazakos, Davide Moltisanti, Jonathan Munro, Toby Perrett, Will Price, and Michael Wray. The epic-kitchens dataset: Collection, challenges and baselines. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 43(11):4125-4141, 2021. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 53, + 692, + 286, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 692, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 53, + 692, + 286, + 713 + ], + "type": "text", + "content": "[14] Andrew J. Davison. Real-time simultaneous localisation and mapping with a single camera. In ICCV, 2003. 3" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 312, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 25, + "blocks": [ + { + "bbox": [ + 312, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 312, + 73, + 545, + 106 + ], + "type": "text", + "content": "[15] Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. Superpoint: Self-supervised interest point detection and description, 2018. 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 312, + 108, + 545, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 108, + 545, + 140 + ], + "spans": [ + { + "bbox": [ + 312, + 108, + 545, + 140 + ], + "type": "text", + "content": "[16] Alexey Dosovitskiy, German Ros, Felipe Codevilla, Antonio Lopez, and Vladlen Koltun. Carla: An open urban driving simulator, 2017. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 312, + 143, + 545, + 219 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 143, + 545, + 219 + ], + "spans": [ + { + "bbox": [ + 312, + 143, + 545, + 219 + ], + "type": "text", + "content": "[17] Salehe Erfanian Ebadi, Saurav Dhakad, Sanjay Vishwakarma, Chunpu Wang, You-Cyuan Jhang, Maciek Chociej, Adam Crespi, Alex Thaman, and Sujoy Ganguly. Psp-hdri+: A synthetic dataset generator for pre-training of human-centric computer vision models. In First Workshop on Pre-training: Perspectives, Pitfalls, and Paths Forward at ICML 2022, 2022. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 312, + 222, + 545, + 275 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 222, + 545, + 275 + ], + "spans": [ + { + "bbox": [ + 312, + 222, + 545, + 275 + ], + "type": "text", + "content": "[18] Matteo Fabbri, Fabio Lanzi, Simone Calderara, Andrea Palazzi, Roberto Vezzani, and Rita Cucchiara. Learning to detect and track visible and occluded body joints in a virtual world. In European Conference on Computer Vision (ECCV), 2018. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 312, + 278, + 545, + 332 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 278, + 545, + 332 + ], + "spans": [ + { + "bbox": [ + 312, + 278, + 545, + 332 + ], + "type": "text", + "content": "[19] Alexander Fiannaca, Ilias Apostolopoulous, and Eelke Folmer. Headlock: A wearable navigation aid that helps blind cane users traverse large open spaces. In Proceedings of the 16th international ACM SIGACCESS conference on Computers & accessibility, pages 19-26, 2014. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 312, + 335, + 545, + 388 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 335, + 545, + 388 + ], + "spans": [ + { + "bbox": [ + 312, + 335, + 545, + 388 + ], + "type": "text", + "content": "[20] Zahra Ghodsi, Siva Kumar Sastry Hari, Iuri Frosio, Timothy Tsai, Alejandro Troccoli, Stephen W. Keckler, Siddharth Garg, and Anima Anandkumar. Generating and characterizing scenarios for safety testing of autonomous vehicles, 2021. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 312, + 392, + 543, + 412 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 392, + 543, + 412 + ], + "spans": [ + { + "bbox": [ + 312, + 392, + 543, + 412 + ], + "type": "text", + "content": "[21] Google. ARCore. https://developers.google.com/ar/, 2018.3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 312, + 415, + 545, + 436 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 415, + 545, + 436 + ], + "spans": [ + { + "bbox": [ + 312, + 415, + 545, + 436 + ], + "type": "text", + "content": "[22] Venu Madhav Govindu. Combining two-view constraints for motion estimation. In CVPR, 2001. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 312, + 439, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 439, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 312, + 439, + 545, + 713 + ], + "type": "text", + "content": "[23] Kristen Grauman, Andrew Westbury, Eugene Byrne, Zachary Chavis, Antonino Furnari, Rohit Girdhar, Jackson Hamburger, Hao Jiang, Miao Liu, Xingyu Liu, Miguel Martin, Tushar Nagarajan, Iija Radosavovic, Santhosh Kumar Ramakrishnan, Fiona Ryan, Jayant Sharma, Michael Wray, Mengmeng Xu, Eric Zhongcong Xu, Chen Zhao, Siddhant Bansal, Dhruv Batra, Vincent Cartillier, Sean Crane, Tien Do, Morrie Doulaty, Akshay Erapalli, Christoph Feichtenhofer, Adriano Fragomeni, Qichen Fu, Christian Fuegen, Abrham Gebreselasie, Cristina Gonzalez, James Hillis, Xuhua Huang, Yifei Huang, Wenqi Jia, Weslie Khoo, Jachym Kolar, Satwik Kottur, Anurag Kumar, Federico Landini, Chao Li, Yanghao Li, Zhenqiang Li, Karttikeya Mangalam, Raghava Modhugu, Jonathan Munro, Tullie Murrell, Takumi Nishiyasu, Will Price, Paola Ruiz Puentes, Merey Ramazanova, Leda Sari, Kiran Somasundaram, Audrey Southerland, Yusuke Sugano, Ruijie Tao, Minh Vo, Yuchen Wang, Xindi Wu, Takuma Yagi, Yunyi Zhu, Pablo Arbelaez, David Crandall, Dima Damen, Giovanni Maria Farinella, Bernard Ghanem, Vamsi Krishna Ithapu, C. V. Jawahar, Hanbyul Joo, Kris Kitani, Haizhou Li, Richard Newcombe, Aude Oliva, Hyun Soo Park, James M. Rehg, Yoichi Sato, Jianbo Shi, Mike Zheng Shou, Antonio Torralba, Lorenzo Torresani, Mingfei Yan, and Jitendra Malik. Ego4d: Around the World in 3,000 Hours of Egocentric" + } + ] + } + ], + "index": 24 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14505" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 53, + 72, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 72, + 72, + 286, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 72, + 72, + 286, + 95 + ], + "spans": [ + { + "bbox": [ + 72, + 72, + 286, + 95 + ], + "type": "text", + "content": "Video. In IEEE/CVF Computer Vision and Pattern Recognition (CVPR), 2022. 1, 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 53, + 96, + 287, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 96, + 287, + 162 + ], + "spans": [ + { + "bbox": [ + 53, + 96, + 287, + 162 + ], + "type": "text", + "content": "[24] Artur Grigorev, Michael J. Black, and Otmar Hilliges. HOOD: hierarchical graphs for generalized modelling of clothing dynamics. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2023, Vancouver, BC, Canada, June 17-24, 2023, pages 16965-16974. IEEE, 2023. 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 163, + 288, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 163, + 288, + 229 + ], + "spans": [ + { + "bbox": [ + 53, + 163, + 288, + 229 + ], + "type": "text", + "content": "[25] Vladimir Guzov, Aymen Mir, Torsten Sattler, and Gerard Pons-Moll. Human poseitioning system (hps): 3d human pose estimation and self-localization in large scenes from body-mounted sensors. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4318-4329, 2021. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 231, + 287, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 231, + 287, + 285 + ], + "spans": [ + { + "bbox": [ + 54, + 231, + 287, + 285 + ], + "type": "text", + "content": "[26] Norbert Haala, Michael Cramer, Florian Weimer, and Martin Trittler. Performance test on uav-based photogrammetric data collection. The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, 2012. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 286, + 286, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 286, + 286, + 319 + ], + "spans": [ + { + "bbox": [ + 54, + 286, + 286, + 319 + ], + "type": "text", + "content": "[27] Ankur Handa, Thomas Whelan, John McDonald, and Andrew J Davison. A benchmark for rgb-d visual odometry, 3d reconstruction and slam. ICRA, 2014. 8" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 320, + 287, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 320, + 287, + 386 + ], + "spans": [ + { + "bbox": [ + 53, + 320, + 287, + 386 + ], + "type": "text", + "content": "[28] Albert Haque, Boya Peng, Zelun Luo, Alexandre Alahi, Serena Yeung, and Li Fei-Fei. Towards viewpoint invariant 3d human pose estimation. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11–14, 2016, Proceedings, Part I 14, pages 160–177. Springer, 2016. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 388, + 287, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 388, + 287, + 441 + ], + "spans": [ + { + "bbox": [ + 53, + 388, + 287, + 441 + ], + "type": "text", + "content": "[29] Mohamed Hassan, Duygu Ceylan, Ruben Villegas, Jun Saito, Jimei Yang, Yi Zhou, and Michael Black. Stochastic scene-aware motion prediction. In Proceedings of the International Conference on Computer Vision 2021, 2, 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 444, + 287, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 444, + 287, + 487 + ], + "spans": [ + { + "bbox": [ + 53, + 444, + 287, + 487 + ], + "type": "text", + "content": "[30] Mohamed Hassan, Yunrong Guo, Tingwu Wang, Michael Black, Sanja Fidler, and Xue Bin Peng. Synthesizing physical character-scene interactions. arXiv preprint arXiv:2302.00883, 2023. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 489, + 287, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 489, + 287, + 521 + ], + "spans": [ + { + "bbox": [ + 53, + 489, + 287, + 521 + ], + "type": "text", + "content": "[31] Daniel Holden, Taku Komura, and Jun Saito. Phase-functional neural networks for character control. ACM Trans. Graph., 36(4):1-13, 2017. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 523, + 287, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 523, + 287, + 555 + ], + "spans": [ + { + "bbox": [ + 53, + 523, + 287, + 555 + ], + "type": "text", + "content": "[32] Daniel Holden, Oussama Kanoun, Maksym Perepichka, and Tiberiu Popa. Learned motion matching. ACM Transactions on Graphics (TOG), 39(4):53-1, 2020. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 53, + 557, + 287, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 557, + 287, + 600 + ], + "spans": [ + { + "bbox": [ + 53, + 557, + 287, + 600 + ], + "type": "text", + "content": "[33] Yuhe Jin, Dmytro Mishkin, Anastasiia Mishchuk, Jiri Matas, Pascal Fua, Kwang Moo Yi, and Eduard Trulls. Image Matching across Wide Baselines: From Paper to Practice. IJCV, 2021. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 53, + 602, + 287, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 602, + 287, + 668 + ], + "spans": [ + { + "bbox": [ + 53, + 602, + 287, + 668 + ], + "type": "text", + "content": "[34] Neena Kamath. Announcing Azure Spatial Anchors for collaborative, cross-platform mixed reality apps. https://azure.microsoft.com/en-us/blog/announcing-azure-spatial-anchors-for-collaborative-cross-platform-mixed-reality-apps/, 2019.3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 53, + 670, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 670, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 53, + 670, + 287, + 713 + ], + "type": "text", + "content": "[35] Angjoo Kanazawa, Michael J Black, David W Jacobs, and Jitendra Malik. End-to-end recovery of human shape and pose. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 7122-7131, 2018. 3" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 312, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 312, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 312, + 73, + 545, + 106 + ], + "type": "text", + "content": "[36] Alex Kendall, Matthew Grimes, and Roberto Cipolla. PoseNet: A Convolutional Network for Real-Time 6-DoF Camera Relocalization. In ICCV, 2015. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 312, + 107, + 545, + 148 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 107, + 545, + 148 + ], + "spans": [ + { + "bbox": [ + 312, + 107, + 545, + 148 + ], + "type": "text", + "content": "[37] Georg Klein and David Murray. Parallel tracking and mapping for small ar workspaces. In IEEE and ACM International Symposium on Mixed and Augmented Reality, 2007. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 312, + 150, + 545, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 150, + 545, + 205 + ], + "spans": [ + { + "bbox": [ + 312, + 150, + 545, + 205 + ], + "type": "text", + "content": "[38] Muhammed Kocabas, Chun-Hao P. Huang, Otmar Hilliges, and Michael J. Black. PARE: Part attention regressor for 3D human body estimation. In Proceedings International Conference on Computer Vision (ICCV), pages 11127-11137. IEEE, 2021. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 312, + 206, + 545, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 206, + 545, + 259 + ], + "spans": [ + { + "bbox": [ + 312, + 206, + 545, + 259 + ], + "type": "text", + "content": "[39] Nikos Kolotouros, Georgios Pavlakos, Michael J Black, and Kostas Daniilidis. Learning to reconstruct 3D human pose and shape via model-fitting in the loop. In Proceedings of the IEEE International Conference on Computer Vision, pages 2252-2261, 2019." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 312, + 261, + 545, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 261, + 545, + 293 + ], + "spans": [ + { + "bbox": [ + 312, + 261, + 545, + 293 + ], + "type": "text", + "content": "[40] Nikos Kolotouros, Georgios Pavlakos, and Kostas Dani-ilidis. Convolutional mesh regression for single-image human shape reconstruction. In CVPR, 2019." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 312, + 295, + 545, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 295, + 545, + 348 + ], + "spans": [ + { + "bbox": [ + 312, + 295, + 545, + 348 + ], + "type": "text", + "content": "[41] Nikos Kolotouros, Georgios Pavlakos, Dinesh Jayaraman, and Kostas Daniilidis. Probabilistic modeling for human mesh recovery. In Proceedings of the IEEE/CVF international conference on computer vision, pages 11605-11614, 2021. 3, 8" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 312, + 350, + 545, + 404 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 350, + 545, + 404 + ], + "spans": [ + { + "bbox": [ + 312, + 350, + 545, + 404 + ], + "type": "text", + "content": "[42] Lucas Kovar, Michael Gleicher, and Frédéric H. Pighin. Motion graphs. In International Conference on Computer Graphics and Interactive Techniques, SIGGRAPH 2008, Los Angeles, California, USA, August 11-15, 2008, Classes, pages 51:1-51:10. ACM, 2008. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 312, + 405, + 545, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 405, + 545, + 459 + ], + "spans": [ + { + "bbox": [ + 312, + 405, + 545, + 459 + ], + "type": "text", + "content": "[43] Jiaman Li, C. Karen Liu, and Jiajun Wu. Ego-body pose estimation via ego-head pose estimation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2023, Vancouver, BC, Canada, June 17-24, 2023, pages 17142-17151. IEEE, 2023. 1, 2, 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 312, + 460, + 545, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 460, + 545, + 525 + ], + "spans": [ + { + "bbox": [ + 312, + 460, + 545, + 525 + ], + "type": "text", + "content": "[44] Zhihao Li, Jianzhuang Liu, Zhensong Zhang, Songcen Xu, and Youliang Yan. CLIFF: Carrying location information in full frames into human pose and shape estimation. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part V, pages 590-606. Springer, 2022. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 312, + 526, + 545, + 558 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 526, + 545, + 558 + ], + "spans": [ + { + "bbox": [ + 312, + 526, + 545, + 558 + ], + "type": "text", + "content": "[45] Hung Yu Ling, Fabio Zinno, George Cheng, and Michiel van de Panne. Character controllers using motion vaes. ACM Trans. Graph., 39(4), 2020. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 312, + 559, + 545, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 559, + 545, + 613 + ], + "spans": [ + { + "bbox": [ + 312, + 559, + 545, + 613 + ], + "type": "text", + "content": "[46] Miao Liu, Dexin Yang, Yan Zhang, Zhaopeng Cui, James M Rehg, and Siyu Tang. 4d human body capture from egocentric video via 3d scene grounding. In 2021 International Conference on 3D Vision (3DV), pages 930-939. IEEE, 2021. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 312, + 614, + 545, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 614, + 545, + 658 + ], + "spans": [ + { + "bbox": [ + 312, + 614, + 545, + 658 + ], + "type": "text", + "content": "[47] Zhengyi Luo, Ryo Hachiuma, Ye Yuan, Shun Iwase, and Kris M Kitani. Kinematics-guided reinforcement learning for object-aware 3d ego-pose estimation. arXiv preprint arXiv:2011.04837, 2020. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 312, + 658, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 658, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 312, + 658, + 545, + 713 + ], + "type": "text", + "content": "[48] Naureen Mahmood, Nima Ghorbani, Nikolaus F Troje, Gerard Pons-Moll, and Michael J Black. Amass: Archive of motion capture as surface shapes. In Proceedings of the IEEE/CVF international conference on computer vision, pages 5442-5451, 2019. 3" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "text", + "content": "14506" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 53, + 72, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 53, + 72, + 287, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 72, + 287, + 138 + ], + "spans": [ + { + "bbox": [ + 53, + 72, + 287, + 138 + ], + "type": "text", + "content": "[49] Angel Martínez-González, Michael Villamizar, Olivier Canévet, and Jean-Marc Odobez. Residual pose: A decoupled approach for depth-based 3d human pose estimation. In 2020 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 10313-10318. IEEE, 2020. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 53, + 140, + 286, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 140, + 286, + 161 + ], + "spans": [ + { + "bbox": [ + 53, + 140, + 286, + 161 + ], + "type": "text", + "content": "[50] Meta. Project Aria Glasses. https://www.projectaria.com/, 2023.6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 163, + 286, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 163, + 286, + 185 + ], + "spans": [ + { + "bbox": [ + 53, + 163, + 286, + 185 + ], + "type": "text", + "content": "[51] Microsoft. HoloLens 2. https://www.microsoft.com/en-us/hololens, 2019.6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 186, + 287, + 240 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 186, + 287, + 240 + ], + "spans": [ + { + "bbox": [ + 53, + 186, + 287, + 240 + ], + "type": "text", + "content": "[52] Marko Mihajlovic, Shunsuke Saito, Aayush Bansal, Michael Zollhoefer, and Siyu Tang. COAP: Compositional articulated occupancy of people. In Proceedings IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2022. 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 241, + 287, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 241, + 287, + 274 + ], + "spans": [ + { + "bbox": [ + 53, + 241, + 287, + 274 + ], + "type": "text", + "content": "[53] Aymen Mir, Xavier Puig, Angjoo Kanazawa, and Gerard Pons-Moll. Generating continual human motion in diverse 3d scenes. CoRR, abs/2304.02061, 2023. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 276, + 287, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 276, + 287, + 341 + ], + "spans": [ + { + "bbox": [ + 53, + 276, + 287, + 341 + ], + "type": "text", + "content": "[54] Michael Montemerlo, Nicholas Roy, and Sebastian Thrun. Perspectives on standardization in mobile robot programming: the Carnegie mellon navigation (CARMEN) toolkit. In 2003 IEEE/RSJ International Conference on Intelligent Robots and Systems, Las Vegas, Nevada, USA, October 27 - November 1, 2003, pages 2436-2441. IEEE, 2003. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 342, + 287, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 342, + 287, + 397 + ], + "spans": [ + { + "bbox": [ + 53, + 342, + 287, + 397 + ], + "type": "text", + "content": "[55] Gyeongsik Moon, Ju Yong Chang, and Kyoung Mu Lee. V2v-posenet: Voxel-to-voxel prediction network for accurate 3d hand and human pose estimation from a single depth map. In Proceedings of the IEEE conference on computer vision and pattern Recognition, pages 5079-5088, 2018. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 399, + 287, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 399, + 287, + 430 + ], + "spans": [ + { + "bbox": [ + 53, + 399, + 287, + 430 + ], + "type": "text", + "content": "[56] Etienne Mouragnon, Maxime Lhuillier, Michel Dhome, Fabien Dekeyser, and Patrick Sayd. Real time localization and 3d reconstruction. In CVPR, 2006. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 432, + 287, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 432, + 287, + 487 + ], + "spans": [ + { + "bbox": [ + 53, + 432, + 287, + 487 + ], + "type": "text", + "content": "[57] Evonne Ng, Donglai Xiang, Hanbyul Joo, and Kristen Grauman. You2me: Inferring body pose in egocentric video via first and second person interactions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9890-9900, 2020. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 488, + 287, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 488, + 287, + 521 + ], + "spans": [ + { + "bbox": [ + 53, + 488, + 287, + 521 + ], + "type": "text", + "content": "[58] Tony Ng, Adrian Lopez-Rodriguez, Vassileios Balntas, and Krystian Mikolajczyk. Reassessing the limitations of cnn methods for camera pose regression, 2021. 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 53, + 522, + 286, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 522, + 286, + 555 + ], + "spans": [ + { + "bbox": [ + 53, + 522, + 286, + 555 + ], + "type": "text", + "content": "[59] Niantic. Niantic Expands Developer Platform and AR Tools with Niantic Lightship. https://nianticlabs.com/news/lightship/, 2021.3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 53, + 556, + 286, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 556, + 286, + 578 + ], + "spans": [ + { + "bbox": [ + 53, + 556, + 286, + 578 + ], + "type": "text", + "content": "[60] David Nister, Oleg Naroditsky, and James Bergen. Visual odometry. In CVPR, 2004. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 53, + 579, + 286, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 579, + 286, + 635 + ], + "spans": [ + { + "bbox": [ + 53, + 579, + 286, + 635 + ], + "type": "text", + "content": "[61] Xiaqing Pan, Nicholas Charron, Yongqian Yang, Scott Peters, Thomas Whelan, Chen Kong, Omkar M. Parkhi, Richard A. Newcombe, and Carl Yuheng Ren. Aria digital twin: A new benchmark dataset for egocentric 3d machine perception. CoRR, abs/2306.06362, 2023. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 53, + 635, + 286, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 635, + 286, + 689 + ], + "spans": [ + { + "bbox": [ + 53, + 635, + 286, + 689 + ], + "type": "text", + "content": "[62] Fabio Pardo, Arash Tavakoli, Vitaly Levdik, and Petar Kormushev. Time limits in reinforcement learning. In Proceedings of the 35th International Conference on Machine Learning, pages 4045-4054, Stockholm, Sweden, 2018. PMLR. 4" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 53, + 690, + 286, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 690, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 53, + 690, + 286, + 713 + ], + "type": "text", + "content": "[63] Georgios Pavlakos, Vasileios Choutas, Nima Ghorbani, Timo Bolkart, Ahmed A. A. Osman, Dimitrios Tzionas," + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 312, + 72, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 331, + 72, + 545, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 331, + 72, + 545, + 117 + ], + "spans": [ + { + "bbox": [ + 331, + 72, + 545, + 117 + ], + "type": "text", + "content": "and Michael J. Black. Expressive body capture: 3D hands, face, and body from a single image. In Proceedings IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), pages 10975-10985, 2019. 3, 5, 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 312, + 118, + 545, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 118, + 545, + 161 + ], + "spans": [ + { + "bbox": [ + 312, + 118, + 545, + 161 + ], + "type": "text", + "content": "[64] Xue Bin Peng, Ze Ma, Pieter Abbeel, Sergey Levine, and Angjoo Kanazawa. Amp: Adversarial motion priors for stylized physics-based character control. ACM Transactions on Graphics (TOG), 40(4):1-20, 2021. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 312, + 162, + 545, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 162, + 545, + 194 + ], + "spans": [ + { + "bbox": [ + 312, + 162, + 545, + 194 + ], + "type": "text", + "content": "[65] Mathis Petrovich, Michael J. Black, and Gül Varol. Action-conditioned 3d human motion synthesis with transformer vae, 2021. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 312, + 195, + 545, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 195, + 545, + 227 + ], + "spans": [ + { + "bbox": [ + 312, + 195, + 545, + 227 + ], + "type": "text", + "content": "[66] Marc Pollefeys, Luc Van Gool, Maarten Vergauwen, Frank Verbiest, Kurt Cornelis, Jan Tops, and Reinhard Koch. Visual modeling with a hand-held camera. IJCV, 2004. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 312, + 228, + 545, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 228, + 545, + 327 + ], + "spans": [ + { + "bbox": [ + 312, + 228, + 545, + 327 + ], + "type": "text", + "content": "[67] Xavier Puig, Eric Undersander, Andrew Szot, Mikael Dallaire Cote, Tsung-Yen Yang, Ruslan Partsey, Ruta Desai, Alexander William Clegg, Michal Hlavac, So Yeon Min, Vladimir Vondrus, Théophile Gervet, Vincent-Pierre Berges, John M. Turner, Oleksandr Maksymets, Zsolt Kira, Mrinal Kalakrishnan, Jitendra Malik, Devendra Singh Chaplot, Unnat Jain, Dhruv Batra, Akshara Rai, and Roozbeh Mottaghi. Habitat 3.0: A co-habitat for humans, avatars and robots. CoRR, abs/2310.13724, 2023. 2, 8" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 312, + 327, + 545, + 371 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 327, + 545, + 371 + ], + "spans": [ + { + "bbox": [ + 312, + 327, + 545, + 371 + ], + "type": "text", + "content": "[68] Albert Pumarola, Jordi Sanchez, Gary Choi, Alberto Sanfeliu, and Francesc Moreno-Noguer. 3DPeople: Modeling the Geometry of Dressed Humans. In International Conference in Computer Vision (ICCV), 2019. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 312, + 372, + 545, + 425 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 372, + 545, + 425 + ], + "spans": [ + { + "bbox": [ + 312, + 372, + 545, + 425 + ], + "type": "text", + "content": "[69] Morgan Quigley, Ken Conley, Brian Gerkey, Josh Faust, Tully Foote, Jeremy Leibs, Rob Wheeler, Andrew Y Ng, et al. Ros: an open-source robot operating system. In ICRA workshop on open source software, page 5. Kobe, Japan, 2009. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 312, + 426, + 545, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 426, + 545, + 480 + ], + "spans": [ + { + "bbox": [ + 312, + 426, + 545, + 480 + ], + "type": "text", + "content": "[70] Umer Rafi, Juergen Gall, and Bastian Leibe. A semantic occlusion model for human pose estimation from a single depth image. In Proceedings of the IEEE conference on computer vision and pattern recognition workshops, pages 67-74, 2015. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 312, + 481, + 545, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 481, + 545, + 525 + ], + "spans": [ + { + "bbox": [ + 312, + 481, + 545, + 525 + ], + "type": "text", + "content": "[71] Tilman Reinhardt. Google Visual Positioning Service. https://ai.googleblog.com/2019/02/using-global-localization-to-improve.html, 2019.3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 312, + 525, + 545, + 569 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 525, + 545, + 569 + ], + "spans": [ + { + "bbox": [ + 312, + 525, + 545, + 569 + ], + "type": "text", + "content": "[72] Davis Rempe, Tolga Birdal, Aaron Hertzmann, Jimei Yang, Srinath Sridhar, and Leonidas J. Guibas. Humor: 3d human motion model for robust pose estimation. In International Conference on Computer Vision (ICCV), 2021. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 312, + 570, + 545, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 570, + 545, + 613 + ], + "spans": [ + { + "bbox": [ + 312, + 570, + 545, + 613 + ], + "type": "text", + "content": "[73] Davis Rempe, Zhengyi Luo, Xue Bin Peng, Ye Yuan, Kris Kitani, Karsten Kreis, Sanja Fidler, and Or Litany. Trace and pace: Controllable pedestrian animation via guided trajectory diffusion, 2023. 2, 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 312, + 614, + 545, + 680 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 614, + 545, + 680 + ], + "spans": [ + { + "bbox": [ + 312, + 614, + 545, + 680 + ], + "type": "text", + "content": "[74] Guodong Rong, Byung Hyun Shin, Hadi Tabatabaee, Qiang Lu, Steve Lemke, Martiš Možeiko, Eric Boise, Geehoon Uhm, Mark Gerow, Shalin Mehta, et al. Lgsvl simulator: A high fidelity simulator for autonomous driving. In 2020 IEEE 23rd International conference on intelligent transportation systems (ITSC), pages 1-6. IEEE, 2020. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 312, + 681, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 681, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 312, + 681, + 545, + 713 + ], + "type": "text", + "content": "[75] Paul-Edouard Sarlin, Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. Superglue: Learning feature matching with graph neural networks, 2020. 7" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "14507" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 53, + 73, + 288, + 715 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 53, + 73, + 287, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 73, + 287, + 127 + ], + "spans": [ + { + "bbox": [ + 53, + 73, + 287, + 127 + ], + "type": "text", + "content": "[76] Paul-Edouard Sarlin, Mihai Dusmanu, Johannes L. Schonberger, Pablo Speciale, Lukas Gruber, Viktor Larsson, Ondrej Miksik, and Marc Pollefeys. LaMAR: Benchmarking Localization and Mapping for Augmented Reality. In ECCV, 2022. 3, 7" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 53, + 128, + 288, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 128, + 288, + 194 + ], + "spans": [ + { + "bbox": [ + 53, + 128, + 288, + 194 + ], + "type": "text", + "content": "[77] Torsten Sattler, Will Maddern, Carl Toft, Akihiko Torii, Lars Hammarstrand, Erik Stenberg, Daniel Safari, Masatoshi Okutomi, Marc Pollefeys, Josef Sivic, Fredrik Kahl, and Tomas Pajdla. Benchmarking 6DoF outdoor visual localization in changing conditions. In CVPR, 2018. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 194, + 288, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 194, + 288, + 248 + ], + "spans": [ + { + "bbox": [ + 53, + 194, + 288, + 248 + ], + "type": "text", + "content": "[78] Manolis Savva, Abhishek Kadian, Oleksandr Maksymets, Yili Zhao, Erik Wijmans, Bhavana Jain, Julian Straub, Jia Liu, Vladlen Koltun, Jitendra Malik, Devi Parikh, and Dhruv Batra. Habitat: A platform for embodied ai research, 2019. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 249, + 288, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 249, + 288, + 271 + ], + "spans": [ + { + "bbox": [ + 53, + 249, + 288, + 271 + ], + "type": "text", + "content": "[79] Johannes Lutz Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In CVPR, 2016. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 271, + 288, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 271, + 288, + 304 + ], + "spans": [ + { + "bbox": [ + 53, + 271, + 288, + 304 + ], + "type": "text", + "content": "[80] Johannes L. Schonberger, Hans Hardmeier, Torsten Sattler, and Marc Pollefeys. Comparative evaluation of handcrafted and learned local features. In CVPR, 2017. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 304, + 288, + 337 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 304, + 288, + 337 + ], + "spans": [ + { + "bbox": [ + 53, + 304, + 288, + 337 + ], + "type": "text", + "content": "[81] John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Radford, and Oleg Klimov. Proximal policy optimization algorithms, 2017. 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 337, + 288, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 337, + 288, + 381 + ], + "spans": [ + { + "bbox": [ + 53, + 337, + 288, + 381 + ], + "type": "text", + "content": "[82] Akash Sengupta, Ignas Budvytis, and Roberto Cipolla. Synthetic training for accurate 3d human pose and shape estimation in the wild. In British Machine Vision Conference (BMVC), 2020. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 382, + 288, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 382, + 288, + 415 + ], + "spans": [ + { + "bbox": [ + 53, + 382, + 288, + 415 + ], + "type": "text", + "content": "[83] Takaaki Shiratori, Hyun Soo Park, Leonid Sigal, Yaser Sheikh, and Jessica K. Hodgins. Motion capture from body-mounted cameras. ACM Trans. Graph., 30(4):31, 2011. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 415, + 288, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 415, + 288, + 480 + ], + "spans": [ + { + "bbox": [ + 53, + 415, + 288, + 480 + ], + "type": "text", + "content": "[84] Jamie Shotton, Ross Girshick, Andrew Fitzgibbon, Toby Sharp, Mat Cook, Mark Finocchio, Richard Moore, Pushmeet Kohli, Antonio Criminisi, Alex Kipman, et al. Efficient human pose estimation from single depth images. IEEE transactions on pattern analysis and machine intelligence, 35(12):2821-2840, 2012. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 481, + 288, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 481, + 288, + 525 + ], + "spans": [ + { + "bbox": [ + 53, + 481, + 288, + 525 + ], + "type": "text", + "content": "[85] Jamie Shotton, Ben Glocker, Christopher Zach, Shahram Izadi, Antonio Criminisi, and Andrew Fitzgibbon. Scene Coordinate Regression Forests for Camera Relocalization in RGB-D Images. In CVPR, 2013. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 53, + 525, + 288, + 579 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 525, + 288, + 579 + ], + "spans": [ + { + "bbox": [ + 53, + 525, + 288, + 579 + ], + "type": "text", + "content": "[86] Jamie Shotton, Toby Sharp, Alex Kipman, Andrew Fitzgibbon, Mark Finocchio, Andrew Blake, Mat Cook, and Richard Moore. Real-time human pose recognition in parts from single depth images. Commun. ACM, 56(1):116-124, 2013. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 53, + 580, + 288, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 580, + 288, + 613 + ], + "spans": [ + { + "bbox": [ + 53, + 580, + 288, + 613 + ], + "type": "text", + "content": "[87] Noah Snavely, Steven M. Seitz, and Richard Szeliski. Photo tourism: exploring photo collections in 3d. ACM Trans. Graph., 25(3):835-846, 2006. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 53, + 613, + 288, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 613, + 288, + 658 + ], + "spans": [ + { + "bbox": [ + 53, + 613, + 288, + 658 + ], + "type": "text", + "content": "[88] Kihyuk Sohn, Honglak Lee, and Xinchen Yan. Learning structured output representation using deep conditional generative models. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2015. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 53, + 658, + 288, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 658, + 288, + 690 + ], + "spans": [ + { + "bbox": [ + 53, + 658, + 288, + 690 + ], + "type": "text", + "content": "[89] Sebastian Starke, He Zhang, Taku Komura, and Jun Saito. Neural state machine for character-scene interactions. ACM Trans. Graph., 38(6):209-1, 2019. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 53, + 691, + 288, + 715 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 691, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 53, + 691, + 288, + 715 + ], + "type": "text", + "content": "[90] Sebastian Starke, Ian Mason, and Taku Komura. Deepphase: Periodic autoencoders for learning motion phase" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 715 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 331, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 331, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 331, + 73, + 545, + 95 + ], + "type": "text", + "content": "manifolds. ACM Transactions on Graphics (TOG), 41(4): 1-13, 2022. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 312, + 96, + 547, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 96, + 547, + 206 + ], + "spans": [ + { + "bbox": [ + 312, + 96, + 547, + 206 + ], + "type": "text", + "content": "[91] Julian Straub, Thomas Whelan, Lingni Ma, Yufan Chen, Erik Wijmans, Simon Green, Jakob J. Engel, Raul Mur-Artal, Carl Ren, Shobhit Verma, Anton Clarkson, Mingfei Yan, Brian Budge, Yajie Yan, Xiaqing Pan, June Yon, Yuyang Zou, Kimberly Leon, Nigel Carter, Jesus Briales, Tyler Gillingham, Elias Mueggler, Luis Pesqueira, Manolis Savva, Dhruv Batra, Hauke M. Strasdat, Renzo De Nardi, Michael Goesele, Steven Lovegrove, and Richard Newcombe. The Replica dataset: A digital replica of indoor spaces. arXiv preprint arXiv:1906.05797, 2019. 4, 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 312, + 207, + 547, + 306 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 207, + 547, + 306 + ], + "spans": [ + { + "bbox": [ + 312, + 207, + 547, + 306 + ], + "type": "text", + "content": "[92] Andrew Szot, Alex Clegg, Eric Undersander, Erik Wijmans, Yili Zhao, John Turner, Noah Maestre, Mustafa Mukadam, Devendra Chaplot, Oleksandr Maksymets, Aaron Gokaslan, Vladimir Vondrus, Sameer Dharur, Franziska Meier, Wojciech Galuba, Angel Chang, Zsolt Kira, Vladlen Koltun, Jitendra Malik, Manolis Savva, and Dhruv Batra. Habitat 2.0: Training home assistants to rearrange their habitat. In Advances in Neural Information Processing Systems (NeurIPS), 2021. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 312, + 307, + 545, + 340 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 307, + 545, + 340 + ], + "spans": [ + { + "bbox": [ + 312, + 307, + 545, + 340 + ], + "type": "text", + "content": "[93] Zachary Teed and Jia Deng. DROID-SLAM: Deep Visual SLAM for Monocular, Stereo, and RGB-D Cameras. Advances in neural information processing systems, 2021. 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 312, + 341, + 545, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 341, + 545, + 374 + ], + "spans": [ + { + "bbox": [ + 312, + 341, + 545, + 374 + ], + "type": "text", + "content": "[94] Guy Tevet, Sigal Raab, Brian Gordon, Yonatan Shafir, Daniel Cohen-Or, and Amit H Bermano. Human motion diffusion model. arXiv preprint arXiv:2209.14916, 2022.3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 312, + 375, + 545, + 420 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 375, + 545, + 420 + ], + "spans": [ + { + "bbox": [ + 312, + 375, + 545, + 420 + ], + "type": "text", + "content": "[95] Emanuel Todorov, Tom Erez, and Yuval Tassa. Mujoco: A physics engine for model-based control. In 2012 IEEE/RSJ International Conference on Intelligent Robots and Systems, pages 5026-5033, 2012. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 312, + 421, + 545, + 465 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 421, + 545, + 465 + ], + "spans": [ + { + "bbox": [ + 312, + 421, + 545, + 465 + ], + "type": "text", + "content": "[96] Emanuel Todorov, Tom Erez, and Yuval Tassa. Mujoco: A physics engine for model-based control. In 2012 IEEE/RSJ International Conference on Intelligent Robots and Systems, pages 5026-5033. IEEE, 2012. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 312, + 466, + 545, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 466, + 545, + 520 + ], + "spans": [ + { + "bbox": [ + 312, + 466, + 545, + 520 + ], + "type": "text", + "content": "[97] Denis Tome, Patrick Peluse, Lourdes Agapito, and Hernan Badino. xr-egopose: Egocentric 3d human pose from an hmd camera. In Proceedings of the IEEE International Conference on Computer Vision, pages 7728-7738, 2019. 2, 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 312, + 522, + 545, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 522, + 545, + 577 + ], + "spans": [ + { + "bbox": [ + 312, + 522, + 545, + 577 + ], + "type": "text", + "content": "[98] Denis Tome, Thiemo Alldieck, Patrick Peluse, Gerard Pons-Moll, Lourdes Agapito, Hernan Badino, and Fernando De la Torre. Selfpose: 3d egocentric pose estimation from a headset mounted camera. arXiv preprint arXiv:2011.01519, 2020. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 312, + 578, + 545, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 578, + 545, + 633 + ], + "spans": [ + { + "bbox": [ + 312, + 578, + 545, + 633 + ], + "type": "text", + "content": "[99] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Advances in neural information processing systems, pages 5998-6008, 2017. 4" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 634, + 545, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 634, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 307, + 634, + 545, + 689 + ], + "type": "text", + "content": "[100] Jian Wang, Diogo Luvizon, Weipeng Xu, Lingjie Liu, Kripasindhu Sarkar, and Christian Theobalt. Scene-aware egocentric 3d human pose estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13031-13040, 2023. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 690, + 545, + 715 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 690, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 307, + 690, + 545, + 715 + ], + "type": "text", + "content": "[101] Keze Wang, Shengfu Zhai, Hui Cheng, Xiaodan Liang, and Liang Lin. Human pose estimation from depth images via" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "14508" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 73, + 286, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 72, + 73, + 286, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 72, + 73, + 286, + 106 + ], + "spans": [ + { + "bbox": [ + 72, + 73, + 286, + 106 + ], + "type": "text", + "content": "inference embedded multi-task learning. In Proceedings of the 24th ACM international conference on Multimedia, pages 1227-1236, 2016. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 107, + 286, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 107, + 286, + 184 + ], + "spans": [ + { + "bbox": [ + 48, + 107, + 286, + 184 + ], + "type": "text", + "content": "[102] Xin Wang, Taein Kwon, Mahdi Rad, Bowen Pan, Ishani Chakraborty, Sean Andrist, Dan Bohus, Ashley Feniello, Bugra Tekin, Felipe Vieira Frujeri, Neel Joshi, and Marc Pollefeys. Holoassist: an egocentric human interaction dataset for interactive ai assistants in the real world. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 20270-20281, 2023. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 185, + 286, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 185, + 286, + 251 + ], + "spans": [ + { + "bbox": [ + 49, + 185, + 286, + 251 + ], + "type": "text", + "content": "[103] Erik Wijmans, Abhishek Kadian, Ari Morcos, Stefan Lee, Irfan Essa, Devi Parikh, Manolis Savva, and Dhruv Batra. DD-PPO: learning near-perfect pointgoal navigators from 2.5 billion frames. In 8th International Conference on Learning Representations, ICLR 2020, Addis Ababa, Ethiopia, April 26-30, 2020. OpenReview.net, 2020. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 252, + 286, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 252, + 286, + 274 + ], + "spans": [ + { + "bbox": [ + 49, + 252, + 286, + 274 + ], + "type": "text", + "content": "[104] Kyle Wilson and Noah Snavely. Robust global translations with 1dsfm. In ECCV, 2014. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 49, + 276, + 286, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 276, + 286, + 308 + ], + "spans": [ + { + "bbox": [ + 49, + 276, + 286, + 308 + ], + "type": "text", + "content": "[105] Jungdam Won, Deepak Gopinath, and Jessica Hodgins. Physics-based character controllers using conditional vaes. ACM Trans. Graph., 41(4), 2022. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 49, + 310, + 286, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 310, + 286, + 364 + ], + "spans": [ + { + "bbox": [ + 49, + 310, + 286, + 364 + ], + "type": "text", + "content": "[106] Erroll Wood, Tadas Baltrusaitis, Charlie Hewitt, Sebastian Dziadzio, Thomas J Cashman, and Jamie Shotton. Fake it till you make it: face analysis in the wild using synthetic data alone. In Proceedings of the IEEE/CVF international conference on computer vision, pages 3681-3691, 2021. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 49, + 365, + 286, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 365, + 286, + 430 + ], + "spans": [ + { + "bbox": [ + 49, + 365, + 286, + 430 + ], + "type": "text", + "content": "[107] Fu Xiong, Boshen Zhang, Yang Xiao, Zhiguo Cao, Taidong Yu, Joey Tianyi Zhou, and Junsong Yuan. A2j: Anchor-to-joint regression network for 3d articulated pose estimation from a single depth image. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 793-802, 2019. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 49, + 433, + 286, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 433, + 286, + 498 + ], + "spans": [ + { + "bbox": [ + 49, + 433, + 286, + 498 + ], + "type": "text", + "content": "[108] Mengyuan Yan, Gen Li, Yilin Zhu, and Jeannette Bohg. Learning topological motion primitives for knot planning. In IEEE/RSJ International Conference on Intelligent Robots and Systems, IROS 2020, Las Vegas, NV, USA, October 24, 2020 - January 24, 2021, pages 9457-9464. IEEE, 2020. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 49, + 500, + 286, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 500, + 286, + 555 + ], + "spans": [ + { + "bbox": [ + 49, + 500, + 286, + 555 + ], + "type": "text", + "content": "[109] Zhitao Yang, Zhongang Cai, Haiyi Mei, Shuai Liu, Zhaoxi Chen, Weiye Xiao, Yukun Wei, Zhongfei Qing, Chen Wei, Bo Dai, Wayne Wu, Chen Qian, Dahua Lin, Ziwei Liu, and Lei Yang. Synbody: Synthetic dataset with layered human models for 3d human perception and modeling, 2023. 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 49, + 556, + 286, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 556, + 286, + 599 + ], + "spans": [ + { + "bbox": [ + 49, + 556, + 286, + 599 + ], + "type": "text", + "content": "[110] Heyuan Yao, Zhenhua Song, Baoquan Chen, and Libin Liu. Controlvae: Model-based learning of generative controllers for physics-based characters. ACM Trans. Graph., 41(6), 2022. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 49, + 601, + 286, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 601, + 286, + 644 + ], + "spans": [ + { + "bbox": [ + 49, + 601, + 286, + 644 + ], + "type": "text", + "content": "[111] Mao Ye, Xianwang Wang, Ruigang Yang, Liu Ren, and Marc Pollefeys. Accurate 3d pose estimation from a single depth image. In 2011 International Conference on Computer Vision, pages 731-738. IEEE, 2011. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 49, + 647, + 286, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 647, + 286, + 689 + ], + "spans": [ + { + "bbox": [ + 49, + 647, + 286, + 689 + ], + "type": "text", + "content": "[112] Vickie Ye, Georgios Pavlakos, Jitendra Malik, and Angjoo Kanazawa. Decoupling human and camera motion from videos in the wild. arXiv preprint arXiv:2302.12827, 2023. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 49, + 691, + 286, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 691, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 49, + 691, + 286, + 713 + ], + "type": "text", + "content": "[113] Ye Yuan and Kris Kitani. Ego-posed estimation and forecasting as real-time pd control. In Proceedings of the" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 545, + 564 + ], + "type": "list", + "angle": 0, + "index": 24, + "blocks": [ + { + "bbox": [ + 331, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 331, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 331, + 73, + 545, + 95 + ], + "type": "text", + "content": "IEEE/CVF International Conference on Computer Vision, pages 10082-10092, 2019. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 308, + 96, + 545, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 96, + 545, + 140 + ], + "spans": [ + { + "bbox": [ + 308, + 96, + 545, + 140 + ], + "type": "text", + "content": "[114] Mingyuan Zhang, Zhongang Cai, Liang Pan, Fangzhou Hong, Xinying Guo, Lei Yang, and Ziwei Liu. Motiondiffuse: Text-driven human motion generation with diffusion model. arXiv preprint arXiv:2208.15001, 2022. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 141, + 545, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 141, + 545, + 195 + ], + "spans": [ + { + "bbox": [ + 308, + 141, + 545, + 195 + ], + "type": "text", + "content": "[115] Siwei Zhang, Qianli Ma, Yan Zhang, Zhiyin Qian, Taein Kwon, Marc Pollefeys, Federica Bogo, and Siyu Tang. Egobody: Human body shape and motion of interacting people from head-mounted devices. In European conference on computer vision (ECCV), 2022. 1, 2, 8" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 197, + 545, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 197, + 545, + 251 + ], + "spans": [ + { + "bbox": [ + 308, + 197, + 545, + 251 + ], + "type": "text", + "content": "[116] Siwei Zhang, Qianli Ma, Yan Zhang, Sadegh Aliakbarian, Darren Cosker, and Siyu Tang. Probabilistic human mesh recovery in 3d scenes from egocentric views. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2023. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 253, + 545, + 295 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 253, + 545, + 295 + ], + "spans": [ + { + "bbox": [ + 308, + 253, + 545, + 295 + ], + "type": "text", + "content": "[117] Yan Zhang and Siyu Tang. The wanderings of odysseus in 3d scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20481-20491, 2022. 3, 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 297, + 545, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 297, + 545, + 341 + ], + "spans": [ + { + "bbox": [ + 308, + 297, + 545, + 341 + ], + "type": "text", + "content": "[118] Yan Zhang, Michael J Black, and Siyu Tang. We are more than our joints: Predicting how 3d bodies move. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3372-3382, 2021. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 342, + 545, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 342, + 545, + 385 + ], + "spans": [ + { + "bbox": [ + 308, + 342, + 545, + 385 + ], + "type": "text", + "content": "[119] Kaifeng Zhao, Yan Zhang, Shaofei Wang, Thabo Beeler, and Siyu Tang. Synthesizing diverse human motions in 3d indoor scenes. In International conference on computer vision (ICCV), 2023. 2, 3, 4, 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 387, + 545, + 452 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 387, + 545, + 452 + ], + "spans": [ + { + "bbox": [ + 308, + 387, + 545, + 452 + ], + "type": "text", + "content": "[120] Yuhang Zhao, Elizabeth Kupferstein, Brenda Veronica Castro, Steven Feiner, and Shiri Azenkot. Designing ar visualizations to facilitate stair navigation for people with low vision. In Proceedings of the 32nd annual ACM symposium on user interface software and technology, pages 387-402, 2019. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 454, + 545, + 519 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 454, + 545, + 519 + ], + "spans": [ + { + "bbox": [ + 308, + 454, + 545, + 519 + ], + "type": "text", + "content": "[121] Yang Zheng, Yanchao Yang, Kaichun Mo, Jiaman Li, Tao Yu, Yebin Liu, C Karen Liu, and Leonidas J Guibas. Gimo: Gaze-informed human motion prediction in context. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XIII, pages 676-694. Springer, 2022. 7" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 521, + 545, + 564 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 521, + 545, + 564 + ], + "spans": [ + { + "bbox": [ + 308, + 521, + 545, + 564 + ], + "type": "text", + "content": "[122] Ziwen Zhuang, Zipeng Fu, Jianren Wang, Christopher Atkeson, Soren Schwertfeger, Chelsea Finn, and Hang Zhao. Robot parkour learning. In Conference on Robot Learning (CoRL), 2023. 2, 3" + } + ] + } + ], + "index": 23 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "14509" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/EgoThink_ Evaluating First-Person Perspective Thinking Capability of Vision-Language Models/2ff0d147-b907-475e-8aa3-5584b71f3014_content_list.json b/2024/EgoThink_ Evaluating First-Person Perspective Thinking Capability of Vision-Language Models/2ff0d147-b907-475e-8aa3-5584b71f3014_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..f92bfe3ea8105a315d4f5e684294f45d33b77861 --- /dev/null +++ b/2024/EgoThink_ Evaluating First-Person Perspective Thinking Capability of Vision-Language Models/2ff0d147-b907-475e-8aa3-5584b71f3014_content_list.json @@ -0,0 +1,1390 @@ +[ + { + "type": "text", + "text": "EgoThink: Evaluating First-Person Perspective Thinking Capability of Vision-Language Models", + "text_level": 1, + "bbox": [ + 127, + 128, + 843, + 175 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Sijie Cheng $^{1,2,5 \\ddagger}$ , Zhicheng Guo $^{1,2*}$ , Jingwen Wu $^{3*}$ , Kechen Fang $^{4}$ , Peng Li $^{2 \\boxtimes}$ , Huaping Liu $^{1}$ , Yang Liu $^{1,2 \\boxtimes}$", + "bbox": [ + 215, + 200, + 751, + 238 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Department of Computer Science and Technology, Tsinghua University $^{2}$ Institute for AI Industry Research (AIR), Tsinghua University $^{3}$ Department of Electrical and Computer Engineering, University of Toronto $^{4}$ Zhili College, Tsinghua University $^{5}$ 01.AI csj23@mails.tsinghua.edu.cn", + "bbox": [ + 184, + 239, + 785, + 325 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 361, + 313, + 378 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Vision-language models (VLMs) have recently shown promising results in traditional downstream tasks. Evaluation studies have emerged to assess their abilities, with the majority focusing on the third-person perspective, and only a few addressing specific tasks from the first-person perspective. However, the capability of VLMs to \"think\" from a first-person perspective, a crucial attribute for advancing autonomous agents and robotics, remains largely unexplored. To bridge this research gap, we introduce EgoThink, a novel visual question-answering benchmark that encompasses six core capabilities with twelve detailed dimensions. The benchmark is constructed using selected clips from egocentric videos, with manually annotated question-answer pairs containing first-person information. To comprehensively assess VLMs, we evaluate twenty-one popular VLMs on EgoThink. Moreover, given the open-ended format of the answers, we use GPT-4 as the automatic judge to compute single-answer grading. Experimental results indicate that although GPT-4V leads in numerous dimensions, all evaluated VLMs still possess considerable potential for improvement in first-person perspective tasks. Meanwhile, enlarging the number of trainable parameters has the most significant impact on model performance on EgoThink. In conclusion, EgoThink serves as a valuable addition to existing evaluation benchmarks for VLMs, providing an indispensable resource for future research in the realm of embodied artificial intelligence and robotics.", + "bbox": [ + 76, + 398, + 473, + 806 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/590199f49b69a6a83ae1ad2c7d310819f3dae15d279bc8054b176f41a5f5c11d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 506, + 364, + 890, + 523 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/083095910b6751a0b160ced125d24313e8a4b0ad4320ce4081a9a23578260616.jpg", + "image_caption": [ + "Figure 1. The main categories of our EgoThink benchmark to comprehensively assess the capability of thinking from a first-person perspective." + ], + "image_footnote": [], + "bbox": [ + 658, + 531, + 686, + 553 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "EgoThink", + "text_level": 1, + "bbox": [ + 687, + 537, + 759, + 554 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 500, + 628, + 632, + 645 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Benefiting from the rapid development of large language models (LLMs) [8, 60, 73], vision-language models (VLMs) [2, 15, 43, 80] have shown remarkable progress in both conventional vision-language downstream tasks [2, 15, 43, 80] and following diverse human instructions [13, 42, 48, 81, 89]. Their application has expanded into broader domains such as robotics [21, 31, 40] and embodied artificial intelligence (EAI) [71, 78]. As a result, the thorough evaluation of VLMs has become increasingly important and challenging. Observing and understanding the world from a first-person perspective is a natural approach for both humans and artificial intelligence agents. We propose that the ability to \"think\" from a first-person perspective, especially when interpreting egocentric images, is crucial for VLMs.", + "bbox": [ + 496, + 656, + 890, + 868 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "However, as shown in Table 1, the ability to think from a first-person perspective is not adequately addressed by cur", + "bbox": [ + 500, + 869, + 890, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 44 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Equal contribution, ‡ Project lead, Corresponding author", + "bbox": [ + 94, + 849, + 413, + 864 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "Project page: https://adacheng.github.io/EgoThink/", + "bbox": [ + 102, + 864, + 372, + 875 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "GitHub page: https://github.com/AdaCheng/EgoThink/", + "bbox": [ + 102, + 876, + 393, + 887 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "Dataset page: https://huggingface.co/datasets/EgoThink/EgoThink/", + "bbox": [ + 102, + 888, + 455, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "14291", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "table", + "img_path": "images/2149d4d569ab524c47420b74a71e41afd9e015987c6cc9f20342e807df195494.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
BenchmarkCapabilityPerspectiveData SourceAnswer TypeEvaluatorSize
VL-CheckList [84]Object / Attribute / RelationThirdDatasetsPSAccuracy410k
LVLM-eHub [77]General Multi-ModalityThirdDatasetsMC / OEMetrics / LLMs / User332k
MME [19]General Multi-ModalityThirdHandcraftMCAccuracy2,194
Tiny LVLM-eHub [68]General Multi-ModalityThirdDatasetsOELLMs2,100
MMBench [54]General Multi-ModalityThirdDatasets / Handcraft / LLMsMCLLMs2,974
PCA-EVAL [11]Decision-MakingThirdHandcraftOEAccuracy / User300
EgoTaskQA [34]Spatial / Temporal / CausalFirstCrowdsourcingOECrowdsourcing40k
EgoVQA [16]Object / Action / PersonThird / FirstHandcraftMCAccuracy520
EgoThink (Ours)First-Person ThinkingFirstHandcraftOELLMs700
", + "bbox": [ + 81, + 88, + 893, + 224 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Table 1. Comparison of recent comprehensive evaluation benchmarks of VLMs and our proposed benchmark EgoThink. Third and first indicate third-person and first-person perspectives. Datasets/Handcraft/LLMs denote existing datasets, manual annotation, and automatic generation by LLMs. PS/MC/OE indicate pairwise scoring, multi-choice, and open-ended question-answering, respectively.", + "bbox": [ + 75, + 233, + 892, + 277 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "rent evaluation benchmarks for VLMs. On one hand, most of these benchmarks (six out of nine, as listed in Table 1) focus solely on the third-person perspective. On the other hand, those benchmarks that do consider the first-person perspective only encompass a limited range of capabilities. For instance, EgoTaskQA [34] examines spatial, temporal, and causal aspects, whereas EgoVQA [16] is limited to object, action, and person aspects. Therefore, there is a clear need to develop a comprehensive benchmark to evaluate the first-person capabilities of VLMs more effectively.", + "bbox": [ + 73, + 291, + 468, + 443 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this work, we introduce a new benchmark for VLMs from a first-person perspective, named EgoThink. The initial step in developing this benchmark involves determining the necessary capabilities to assess. Humans, when interacting with the real world, consider a series of questions centered on themselves, ranging from \"What is around me?\", \"What am I doing?\", \"Where am I?\", \"What about the situation around me?\", \"What will happen to me?\" to \"How will I do?\". Drawing inspiration from this, we evaluate six core capabilities of VLMs, namely object, activity, localization, reasoning, forecasting, and planning. Each capability corresponds to one of the aforementioned questions, as illustrated in Figure 1. The next step is constructing the benchmark. We first categorize the six core capabilities into twelve detailed dimensions. We then select a minimum of 50 distinct and clear clips from egocentric videos for each dimension and manually annotate them with relevant first-person question-answer pairs. This approach ensures the quality and variety of the benchmark. The final step is evaluating VLM performance on this benchmark. Building on recent studies [7, 12, 68], we use GPT-4 [60] as an automatic evaluator. The Pearson correlation coefficient, when compared with human evaluation, shows a value of 0.68, indicating that the evaluation results are dependable.", + "bbox": [ + 73, + 445, + 470, + 808 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Based on our proposed EgoThink benchmark, we conduct comprehensive experiments to evaluate the first-person capabilities of twenty-one popular VLMs with varying model and data compositions. The findings indicate that GPT-4V stands out as the most effective model in various aspects. However, it shows less impressive results in spe", + "bbox": [ + 75, + 810, + 470, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "cific capabilities such as activity and counting. Additionally, we observed that no single VLM consistently surpasses others in every aspect. For instance, GPT-4V is less effective than BLIP-2-11B for localization. Increasing the language model portion of the VLMs generally leads to better performance, but this improvement is not uniform across all models. Finally, our results highlight a significant potential for further enhancing the first-person capabilities of VLMs.", + "bbox": [ + 496, + 291, + 893, + 412 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 431, + 640, + 448 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Vision-Language Models. Inspired by the impressive success of LLMs [8, 61, 75], the recent popular VLMs tend to regard the powerful LLMs as the core backbone. At the beginning, VLMs usually use large-scale image-text pairwise datasets [9, 35, 46] or arbitrarily interleaved visual and textual data [2, 90] to pre-train. Furthermore, thanks to the availability of enormous image-text instruction datasets [41, 49], recent studies [13, 42, 48, 81, 89] further apply instruction tuning to help VLMs generate satisfactory answers. Benefiting from the two-stage training process, recent VLMs can achieve stunning performance on downstream vision-language tasks [3, 33, 46, 64].", + "bbox": [ + 496, + 460, + 890, + 642 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Evaluations of VLMs. To evaluate the abilities of VLMs, there are diverse types of vision language downstream tasks. Conventional benchmarks, such as image caption tasks [29, 82] and visual question reasoning tasks [23, 67], mainly probe specific abilities of VLMs from the third-person perspective. Meanwhile, specialized analytical studies comprehensively evaluate the performance of VLMs from the third-person perspective, where Vblue [87] consists of five fundamental tasks and Lvmehub [77] evaluates six categories of capabilities on 47 standard vision-language benchmarks. As for the first-person perspective, there are some egocentric evaluation benchmarks in the computer vision field to assess some visual capabilities [32, 53, 83, 88]. In terms of multi-modality, there are a few benchmarks, such as EgoVQA[16] and EgoTaskQA [34], where mainly specific tasks without an overall understanding. In this paper, we mainly focus on exploring the comprehensive capa", + "bbox": [ + 496, + 643, + 892, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "14292", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/000713ea875894cd593e908d6fbae158eb919b691d82eb9ec6920158683cf253.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 78, + 88, + 480, + 268 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/f4e2c9ff8b5c15d045074ce18a376f0c84be14de1697d851ce51296e1a125c81.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 80, + 268, + 480, + 335 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/1ec81f93bd954bd3db813f746843b2f2cc2edd9bdcc9993fd93f54f145d82ec1.jpg", + "image_caption": [ + "Figure 2. Categories with fine-grained dimensions and their corresponding examples of EgoThink benchmark." + ], + "image_footnote": [], + "bbox": [ + 78, + 335, + 480, + 463 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/a0d8999f7132f208ed4af268ab802c7dd76884612dba1181fe4d5b8681d14cc7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 486, + 88, + 887, + 268 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/de1d8420fbed3468228b674dc89328b43fdbe4a715ccc6dc4269debb8affe0c6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 486, + 268, + 887, + 335 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/6966aa81ea571d068b91902ba9375c3d662181e2df522d9107d712d207bfffb1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 486, + 335, + 887, + 463 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "bilities of VLMs to think from a first-person perspective, as a supplement to previous evaluation benchmarks.", + "bbox": [ + 76, + 505, + 468, + 535 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. EgoThink Benchmark", + "text_level": 1, + "bbox": [ + 76, + 550, + 287, + 566 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we first elaborate on the core capabilities of thinking from a first-person perspective. Then, we introduce the process to manually construct our proposed benchmark EgoThink, which asks VLMs to generate open-ended answers according to first-person images and questions.", + "bbox": [ + 76, + 575, + 468, + 652 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Core Capabilities", + "text_level": 1, + "bbox": [ + 76, + 662, + 246, + 678 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "As shown in Figure 2, we specifically design six categories with twelve fine-grained dimensions from the first-person perspective for quantitative evaluation.", + "bbox": [ + 76, + 686, + 468, + 731 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "- Object: What is around me? Recognizing objects in the real world is a preliminary ability of the human visual system [50, 85, 91]. Images from a first-person or egocentric perspective [53, 65, 88] pay more attention to the objects surrounding the subject or in hands. Moreover, we further divide the object category into three fine-grained dimensions: (1) Existence, predicting whether there is an object as described in the images; (2) Attribute [17, 37], detecting properties or characteristics (e.g., color) of an object; (3) Affordance [28, 56], predicting potential actions that a human can apply to an object.", + "bbox": [ + 76, + 734, + 468, + 901 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Activity: What am I doing? Activity recognition is to automatically recognize specific human activities in video frames or still images [36, 38, 74]. From the egocentric perspective, we mainly focus on actions or activities based on object-hand interaction [6, 18, 59].", + "- Localization: Where am I? In reality, localization is a critical capability for navigation and scene understanding in the real world [55, 66]. Here we investigate the localization capability from two aspects, Location and Spatial Relationship. Location indicates detecting the scene surrounding the subject [14, 26]. Spatial reasoning contains allocentric and egocentric perspectives [24, 39, 57, 58]. We focus on the egocentric perspective, i.e., the position of the object with respect to the subject.", + "- Reasoning: What about the situation around me? During the complex decision-making process, reasoning lies everywhere in our lives. Here we mainly focus on Counting, Comparison, and Situated Reasoning. Due to the first-person perspective, we generally count or compare objects in our hands or surrounding ourselves. As for situated reasoning, we employ cases that cannot be answered directly from the information in the images and require further reasoning processes.", + "- Forecasting: What will happen to me? Forecasting [20, 25, 51, 52] is a critical skill in the real world. From an egocentric view, forecasting always predicts the future of" + ], + "bbox": [ + 500, + 505, + 890, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "14293", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "table", + "img_path": "images/e93175a76128cfd0a94f429aa0b7c9b37c165980170fd486502352f68d04b021.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
VLMsImage EncoderLLMAlignment ModuleTTPToPDataset SizeEgoDataVideo
Image-TextInstruction
API-based Model
GPT-4V [60]Unknown
~7B Models
OpenFlamingo-7B [2, 5]CLIP-ViT-LMPT7BAttention1.4B8.1B2B-X✓
BLIP-2-6.7B [43]EVA-CLIP-ViT-gOPT6.7BQ-Former108M7.8B129M-XX
VideoChat-7B [44]BLIP2-VEVicuna7BQ-Former205M8B25M18KX✓
LLaVA-1.5-7B [47]CLIP-ViT-L-336pxLlama27BMLP6.8B7.1B558k665kXX
MiniGPT-4-7B [89]BLIP2-VELlama27BLinear23M7.7B5M3.5kXX
InstructBLIP-7B [13]EVA-CLIP-ViT-gVicuna7BQ-Former189M7.9B-16MXX
LLaMA-Adapter-7B [22]CLIP-ViT-LLLaMA7BEarly Fusion14M7.2B567k52kXX
Otter-I-7B [42]CLIP-ViT-LMPT7BAttention1.4B8.1B-2.8BX✓
PandaGPT-7B [70]ImageBindVicuna7BLinear + LLM LoRA38M7.9B-160k✓✓
mPLUG-owl-7B [81]CLIP-ViT-LLLaMA7BAttention4M7.1B204M158kXX
Video-LLaVA-7B [45]LanguageBindVicuna7BLinear6.8B7.5B1260k765kX✓
LLaVA-7B [48]CLIP-ViT-LLlama27BLinear6.7B7.1B595k158kXX
ShareGPT4V-7B [10]CLIP-ViT-L-336pxVicuna7BMLP6.7B6.7B1.2M665kXX
~13B Models
InstructBLIP-13B [13]EVA-CLIP-ViT-gVicuna13BQ-Former189M14.2B-16MXX
PandaGPT-13B [70]ImageBindVicuna13BLinear+LLM LoRA52M13.1B-160k✓✓
LLaVA-13B-Vicuna [48]CLIP-ViT-L-336pxVicuna13BLinear13.0B13.3B595k158kXX
BLIP-2-11B [43]EVA-CLIP-ViT-gFlanT5XXLQ-Former108M12.2B129M-XX
InstructBLIP-11B [13]EVA-CLIP-ViT-gFlanT5XXLQ-Former189M12.3B-16MXX
LLaVA-13B-Llama2 [48]CLIP-ViT-LLlama213BLinear13.0B13.3B595k158kXX
LLaVA-1.5-13B [47]CLIP-ViT-L-336pxLlama213BMLP13.0B13.4B558k665kXX
", + "bbox": [ + 80, + 88, + 890, + 465 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Table 2. Statistics of compared API-based and open-source VLMs, where TTP and ToP indicate Total Trainable Parameters and Total Parameters, respectively. Moreover, EgoData and Video indicate that there are egocentric visual data and video data for training, respectively.", + "bbox": [ + 75, + 474, + 892, + 505 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "object-state transformation or hand-object interactions.", + "bbox": [ + 89, + 518, + 452, + 534 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "- Planning: How will I do? In reality, planning [1, 30, 69] is an important capability to deal with complex problems, typically applied in Navigation [62, 63, 72] and Assistance [27, 76]. Navigation is going to a goal location from a start position, while assistance is offering instructions to solve daily problems.", + "bbox": [ + 76, + 534, + 468, + 625 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Data Collection", + "text_level": 1, + "bbox": [ + 76, + 635, + 233, + 648 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this section, we mainly introduce the detailed processing to construct our EgoThink benchmark.", + "bbox": [ + 75, + 659, + 468, + 688 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Collecting first-person visual data. Firstly, we leverage a popular and large egocentric video dataset Ego4D [25], which is designed to advance the field of first-person perception in computer vision. To obtain a diverse representation in different scenarios, Ego4D encompasses 3,670 hours of video from 931 unique camera wearers spanning 74 global locations across 9 countries. To collect first-person visual data, we begin by extracting every frame from a subset of the Ego4D video dataset, yielding a diverse raw image dataset. Please note that our current focus is solely on images, as most VLMs today do not support video input. We intend to expand our scope to include videos in our future work. Considering the heavy human labor and the diversity of scenarios, we sample images every few dozen frames.", + "bbox": [ + 75, + 688, + 468, + 900 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/c09b13c3d87b0f3d46b8e8ff00183e2229ba9b88e7a294737a86fdf647f89c75.jpg", + "image_caption": [ + "Figure 3. This chart illustrates the distribution of various scene categories within the EgoThink dataset. The 'others' category encompasses 13 different scene types, each representing less than one percent of total scenes." + ], + "image_footnote": [], + "bbox": [ + 522, + 518, + 867, + 675 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To ensure high quality, we apply strict criteria for selecting the extracted frames. We first exclude images that lack clarity or fail to exhibit egocentric characteristics. Then, to obtain the high diversity within the dataset, we conduct a further screening to ensure that at most two images per video are included in the filtered image set. Finally, we obtain enormous high-quality images with exhibit egocentric characteristics as first-person image candidates.", + "bbox": [ + 496, + 748, + 890, + 869 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Annotating question-answer pairs. Upon receiving a substantial collection of first-person image candidates, we", + "bbox": [ + 498, + 869, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "14294", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/55df987711440910bcfbef23efc488d20cba05712ba741dcba12781ffcb51414.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodsObjectActivityLocalizationReasoningForecastingPlanningAverage
ExistAttrAffordLocSpatialCountComparSituatedNavAssist
API-based model
GPT-4V62.082.058.059.586.062.042.048.083.055.064.084.065.5
~7B Models
OpenFlamingo-7B16.055.037.015.034.034.021.040.021.031.011.011.027.2
BLIP-2-6.7B49.029.039.033.560.031.03.021.033.025.08.06.028.1
VideoChat-7B46.044.036.045.061.042.036.039.032.026.513.021.036.8
LLaVA-1.5-7B33.047.054.035.535.049.020.047.037.027.029.054.039.0
MiniGPT-4-7B50.056.046.039.055.049.014.048.031.041.514.044.040.6
InstructBLIP-7B50.033.045.047.577.038.018.043.067.040.519.031.042.4
LLaMA-Adapter-7B37.060.046.034.548.051.029.039.025.041.542.057.042.5
Otter-I-7B48.056.039.044.060.044.039.048.042.038.031.055.045.3
PandaGPT-7B40.056.041.037.061.052.019.052.053.043.039.061.046.2
mPLUG-owl-7B56.058.047.053.060.053.025.049.044.049.533.058.048.8
Video-LLaVA-7B56.060.053.045.086.060.039.038.060.046.511.038.049.4
LLaVA-7B63.058.050.047.081.045.024.036.047.049.535.060.049.6
ShareGPT4V-7B67.075.053.055.577.062.030.038.066.047.041.063.051.9
~13B Models
InstructBLIP-13B52.055.049.054.063.049.011.033.059.044.019.025.042.8
PandaGPT-13B35.052.041.040.568.031.032.040.047.045.516.069.043.1
LLaVA-13B-Vicuna54.062.052.046.053.046.026.044.029.044.035.066.046.4
BLIP-2-11B52.062.041.049.590.066.025.050.070.048.018.024.049.6
InstructBLIP-11B74.068.048.049.586.052.032.049.073.053.016.017.051.5
LLaVA-13B-Llama265.061.045.056.077.053.034.034.066.050.549.071.055.1
LLaVA-1.5-13B66.055.051.055.082.057.032.056.067.048.539.055.055.3
", + "bbox": [ + 80, + 88, + 890, + 441 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Table 3. Combined single-answer grading scores on zero-shot setups for various dimensions. The **bold** indicates the best performance while the **underline** indicates the second-best performance. Exist, Attr, Afford, Loc, Spatial, Count, Compar, Situated, Nav and Assist represent existence, attribute, affordance, location, spatial relationship, counting, comparison, situated reasoning, navigation, and assistance.", + "bbox": [ + 75, + 452, + 892, + 494 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "engage six annotators to manually label question-answer pairs. Given that the EgoThink benchmark is composed of twelve dimensions, annotators were responsible for two specific dimensions. The annotators can access all the image candidates and are asked to select appropriate images to annotate their corresponding question-answer pairs to relevant categories. Once the image is selected, it will be removed from the candidates to ensure no repetition. Moreover, to ensure the correctness of our annotations, we have three additional annotators to review the question-answer pairs after the first annotation process. The annotation will not be reserved until the three annotators all agree that the first-person visual data and the assigned question-answer pairs meet the definition of a specific dimension.", + "bbox": [ + 75, + 521, + 468, + 733 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Statistics. The EgoThink benchmark comprises a collection of 700 images across six categories with twelve fine-grained dimensions. These images are extracted from 595 videos, ensuring a broad representation of scenarios. To guarantee diversity, a wide range of scenes and concepts has been deliberately selected. As depicted in Figure 3, the dataset encompasses a diverse range of scenes, covering key scenarios relevant to EAI. Furthermore, we have meticulously crafted question and answer for each image in the EgoThink benchmark, aiming to closely replicate real-life conversations. This involves employing different", + "bbox": [ + 75, + 734, + 470, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "question types, varying questions in length and complexity, paired with well-reasoned and accurate answers. In addition, the size of the dataset is a well-balanced trade-off between benchmark diversity and high cost of open-ended QA evaluation. We ensure our 700 examples were diverse in concepts, scenes, and videos to provide a robust performance estimation within practical limits. Detailed statistics of the EgoThink benchmark are presented in Appendix A.", + "bbox": [ + 496, + 521, + 890, + 643 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 500, + 656, + 633, + 672 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Experimental Setups", + "text_level": 1, + "bbox": [ + 500, + 680, + 697, + 698 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Vision-Language Models. We collect the most popular eighteen types of representative VLMs to assess as shown in Table 2. Due to the possible effects of model parameters, we divide models into $\\sim 7\\mathrm{B}$ and $\\sim 13\\mathrm{B}$ for a fair comparison. Detailed information about VLMs can be found in Appendix B. We conduct zero-shot setups for all VLMs across our EgoThink benchmark. The prompts used for each VLM are shown in Appendix C.", + "bbox": [ + 496, + 703, + 890, + 824 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Single-answer grading. Considering evaluating open-ended model generations is not a trivial problem [54, 68], we propose to use GPT-4 [60] as the automatic evaluator [86] to better measure generated answers. In this protocol, we want to measure how close one model output", + "bbox": [ + 496, + 825, + 893, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "14295", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/d44fe81b279c39d982f12209a2ca04758d000397b08cb8f750d16f0e4d381ccc.jpg", + "image_caption": [ + "Figure 4. Case study (wrong) of GPT-4V in the existence dimension of the object ability. In both the question formats of yes/no and what, GPT-4V can not correctly detect the chopsticks in my right hand, while InstructBLIP-11B and LLaVA-7B can." + ], + "image_footnote": [], + "bbox": [ + 76, + 87, + 468, + 388 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "is to the reference. Different from traditional similarity-based methods, GPT-4 pays more attention to semantics. In the detailed implementation, we format the question, the model output, and the reference in a prompt as shown in Appendix D and feed it into the GPT-4 evaluator. The GPT-4 evaluator is asked to assign a score of 0 (wrong), 0.5 (partially correct), or 1 (correct) to the model output. Additionally, we further discuss to use of GPT-3.5-Turbo, Claude-2, and humans as evaluators in Section 5.2.", + "bbox": [ + 75, + 468, + 468, + 604 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Results", + "text_level": 1, + "bbox": [ + 76, + 617, + 171, + 633 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We first present the overall results of the evaluated models on our EgoThink benchmark as shown in Table 3. Despite having improved over the years, VLMs are still difficult to think from a first-person perspective, even GPT-4V. Among the six categories, only the scores on planning and localization are relatively high, the performance in other capabilities can only reach around 60 points at best. Among the better models, GPT-4V generally performs much better than other models, only falling short in localization dimension compared to BLIP-2-11B. We will further introduce the detailed scores across different dimensions. More case studies can be found in Appendix E.", + "bbox": [ + 75, + 642, + 470, + 823 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Results on object. In detail, we evaluate through three dimensions, including existence, attribute, and affordance. For existence, InstructBLIP-11B and ShareGPT4V-7B achieve the top-2 performance, indicating that they can predict the object precisely from the first-person perspec", + "bbox": [ + 75, + 824, + 468, + 902 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/4da87f053f0a04d7b2caff67fb5573d56f164be4d01d086d3f266a4994ba5c23.jpg", + "image_caption": [ + "Figure 5. Case studies (wrong) in the attribute, affordance, and activity dimensions. The top case demonstrates some VLMs locate the wrong place in the attribute and affordance dimension. The bottom case illustrates that in the activity dimension, some VLMs fail to detect the specific action." + ], + "image_footnote": [], + "bbox": [ + 501, + 87, + 890, + 455 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "tive. As for GPT-4V, as illustrated in Figure 4, we observe that its performance in handed object detection leaves room for improvement. As for both attribute and affordance, the GPT-4V model has demonstrated superior performance, especially in the attribute dimension. In both dimensions, some open-source models as shown at the top of Figure 5 locate the wrong place or only answer the type of the object rather than its attribute or affordance.", + "bbox": [ + 496, + 550, + 892, + 670 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Results on activity. The performance of GPT-4V outperforms all open-source VLMs in the activity dimension. Among the $\\sim 7\\mathrm{B}$ models, ShareGPT4V-7B and mPLUGowl-7B significantly outperform other VLMs and even achieve superior or comparable performance to $\\sim 13\\mathrm{B}$ models. Overall, $\\sim 13\\mathrm{B}$ models tend to perform better than $\\sim 7\\mathrm{B}$ model in the activity dimension, but their scores are just below the passing line. The most possible reason is that detecting the specific action is difficult for VLMs as shown at the bottom of Figure 5.", + "bbox": [ + 496, + 672, + 892, + 823 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Results on localization. In general, BLIP-2-11B has shown obvious advantages among all VLMs, even surpassing GPT-4V in both location and spatial relationship dimensions. In the location dimension, BLIP-2-11B, GPT-4V, and InstructBLIP-11B demonstrate superior ability to achieve", + "bbox": [ + 496, + 824, + 893, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "14296", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/f1b05be3e4f0cca5aaea07d275409330047ef68ca19bbeb0b8590619fcd3cdc3.jpg", + "image_caption": [ + "Figure 6. Case studies (wrong) in counting (top) and situated reasoning (bottom) dimensions. The top case shows VLMs can count but can not understand their relative position. The bottom case displays that the question requires commonsense knowledge and image understanding where only GPT-4V can answer it correctly." + ], + "image_footnote": [], + "bbox": [ + 76, + 88, + 468, + 479 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "around 90 points. However, perceiving the spatial relationship of an object relative to oneself is much more difficult. This phenomenon can be also reflected in the top of Figure 5 where VLMs hard to distinguish left or right hand.", + "bbox": [ + 75, + 577, + 468, + 638 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Results on reasoning. Counting is the most difficult ability [80] among all evaluated dimensions. The best-performing model, GPT-4V, only scores 42.0, far away from satisfaction. Under the first-person perspective setup, VLMs need to not only count but also understand the relative position to oneself, as shown in the top case of Figure 6. Meanwhile, the comparison dimension also reflects the high difficulty, where the best score of 56.0 is obtained by LLaVA-1.5-13B. As for situated reasoning, GPT-4V demonstrates its strong commonsense reasoning ability to answer complex questions at the bottom of Figure 6.", + "bbox": [ + 75, + 641, + 468, + 806 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Results on forecasting. Achieving high performance seems to be challenging as the best score achieved by GPT-4V is only 55.0. InsturctBLIP-11B achieves a relatively high score of 53.0 which is close to that of GPT-4V. We observe that the VLMs mainly suffer from two problems: recognizing objects incorrectly or forecasting too far as shown", + "bbox": [ + 75, + 810, + 470, + 902 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c5c3cecf1be63fbe682bd6df7053c614d6f0d9a7688f6686d59b58e73bfc3be0.jpg", + "image_caption": [ + "Figure 7. Case studies (wrong) in the forecast (top) and navigation dimension (bottom). The top case shows VLMs might detect the glove as a cloth, while the bottom case indicates VLMs lack navigation details and overlook image information." + ], + "image_footnote": [], + "bbox": [ + 500, + 88, + 890, + 537 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "in the top of Figure 7.", + "bbox": [ + 500, + 613, + 647, + 627 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Results on planning. In both navigation and assistance dimensions, the highest scores are achieved by GPT-4V with 60.0 and 84.0, respectively. LLaVA-13B-Llama2 behaves well in both dimensions with the second-best performance but its score is still 10 points lower than that of GPT-4V. The most possible reason is that answers provided by most open-source VLMs lack crucial details or overlook important information given in the images, as illustrated at the bottom of Figure 7.", + "bbox": [ + 496, + 628, + 890, + 763 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5. Analysis", + "text_level": 1, + "bbox": [ + 500, + 777, + 596, + 792 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.1. Effects of Components", + "text_level": 1, + "bbox": [ + 500, + 801, + 710, + 816 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "As shown in Table 2, VLMs consist of multiple key components. In this section, we probe the influence of different components on our EgoThink benchmark.", + "bbox": [ + 496, + 825, + 890, + 869 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The total parameters of LLMs. Here we compare the performance of $\\sim 7\\mathrm{B}$ and $\\sim 13\\mathrm{B}$ variants of four VLMs. Note", + "bbox": [ + 500, + 869, + 890, + 898 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "14297", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/31240c3a0da56d36cf88d2c198cda83630df797e727752ce45d4e909d61048d3.jpg", + "image_caption": [ + "Figure 8. Impact of LLMs sizes (above the dash-line) and instruction-tuning (below the dash-line) on model performance. Average scores across all capabilities are reported." + ], + "image_footnote": [], + "bbox": [ + 78, + 87, + 472, + 270 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "that the increase in the number of parameters mainly falls in the LLMs. Firstly, as shown in the top part of Figure 8, scaling does not lead to significant improvement for PandaGPT and InstructBLIP, while LLaVA (LLaVA-7B and LLava13B-Llama2) and LLaVA-1.5 benefit a lot from scaling. We hypothesize that this is because LLaVA series models do not freeze their language models during instruction tuning, indicating that enlarging the number of trainable parameters can help improve both performance and generalization. In other words, one can see that simply scaling up language models without better alignment may not help.", + "bbox": [ + 75, + 335, + 468, + 501 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Instruction tuning. We directly compare the performance of BLIP-2-11B and InstructBLIP-11B, because these two models differ only in instruction tuning and additional instruction-aware tokens. As presented in the bottom part of Figure 8, InstructBLIP-11B outperforms BLIP-2-11B after instruction tuning, despite an unexpectedly small margin. This may be because much of the instruction tuning data employed by InstructBLIP is collected from specific downstream tasks, whose data distributions are very different from our first-person perspective data.", + "bbox": [ + 75, + 502, + 468, + 652 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The information of image encoder. Considering that there is no ablation version of VLMs for image encoder, following Set-of-Mark [79], we probe the effect of visual grounding information (i.e., a set of marks) in our setups. As presented in Figure 9, GPT-4V with additional segmentation information can correctly detect the mentioned location and objects, indicating that supplemented image information can be helpful in some situations. More discussion about quantitative experiments can be found in Appendix F.", + "bbox": [ + 75, + 652, + 468, + 789 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.2. Agreements between Human and Evaluators", + "text_level": 1, + "bbox": [ + 76, + 801, + 457, + 816 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this section, we further assess the model performance on object and planning dimensions using GPT-3.5-Turbo, Claude-2, and human annotators. Due to the heavy human labor, we ask three annotators to evaluate the performance of GPT-4V, which is the overall best model. Human anno", + "bbox": [ + 75, + 825, + 468, + 898 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/009c8433541138dc8a35fec52d71473cc708944105bc5c953b6ab0430514af07.jpg", + "image_caption": [ + "Figure 9. Case study (wrong) in the adding visual grounding information with images. The segmentation can help VLMs better locate the objects in question." + ], + "image_footnote": [], + "bbox": [ + 500, + 87, + 890, + 247 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "tators consider the following aspects to evaluate: accuracy, completeness, logical soundness, and grammatical correctness. Our annotation system and detailed guidelines can be found in Appendix G. We further conduct GPT-3.5-Turbo and Claude-2 with the same evaluation prompt as GPT-4. The Pearson correlation coefficients between automatic evaluators (i.e., GPT-4, GPT-3.5-Turbo, Claude-2) and humans are 0.68, 0.43, and 0.68, respectively. The Cohen's Kappa coefficient among the three annotators is 0.81. This shows that evaluations made by GPT-4 and Claude-2 have a high correlation with humans. We hypothesize that recent well-performant LLMs can evaluate highly aligned with humans, given that most answers in our benchmark are relatively short and precise. Detailed scores of all evaluators and their correlations are discussed in Appendix H.", + "bbox": [ + 496, + 319, + 892, + 546 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 558, + 619, + 573 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To pave the way for the development of VLMs in the field of EAI and robotics, we introduce a comprehensive benchmark, EgoThink. Designed to evaluate the capacity of VLMs to \"think\" from a first-person perspective, EgoThink encompasses six core capabilities across twelve detailed dimensions. We assess eighteen popular VLMs and find that even the top-performing VLMs in most dimensions achieve only around a score of 60. GPT-4V achieves the best overall performance, but can not consistently surpass other open-source VLMs across all dimensions. In the analysis, we further probe the impact of various components on model performance and find that the total number of trainable parameters in LLMs has the most significant effect. Despite the human agreement with automatic evaluators being high, the evaluation of planning is difficult due to the detailed information in the answers. In future research, we aim to improve the evaluation method and further explore the essential capabilities of VLMs in the EAI and robotics fields. Acknowledgment: The work is supported by the National Key R&D Program of China (2022ZD0160502) and the National Natural Science Foundation of China (No.61925601).", + "bbox": [ + 496, + 583, + 892, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "14298", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 174, + 106 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Michael Ahn, Anthony Brohan, Noah Brown, Yevgen Chebotar, Omar Cortes, Byron David, Chelsea Finn, Keerthana Gopalakrishnan, Karol Hausman, Alex Herzog, et al. Do as i can, not as i say: Grounding language in robotic affordances. arXiv preprint arXiv:2204.01691, 2022. 4", + "[2] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. Advances in Neural Information Processing Systems, 35:23716-23736, 2022. 1, 2, 4", + "[3] Stanislaw Antol, Aishwarya Agrawal, Jiasen Lu, Margaret Mitchell, Dhruv Batra, C Lawrence Zitnick, and Devi Parikh. Vqa: Visual question answering. In Proceedings of the IEEE international conference on computer vision, pages 2425-2433, 2015. 2", + "[4] Anas Awadalla, Irena Gao, Joshua Gardner, Jack Hessel, Yusuf Hanafy, Wanrong Zhu, Kalyani Marathe, Yonatan Bitton, Samir Gadre, Jenia Jitsev, Simon Kornblith, Pang Wei Koh, Gabriel Ilharco, Mitchell Wortsman, and Ludwig Schmidt. Openflamingo, 2023. 1", + "[5] Anas Awadalla, Irena Gao, Josh Gardner, Jack Hessel, Yusuf Hanafy, Wanrong Zhu, Kalyani Marathe, Yonatan Bitton, Samir Gadre, Shiori Sagawa, et al. Openflamingo: An opensource framework for training large autoregressive vision-language models. arXiv preprint arXiv:2308.01390, 2023. 4", + "[6] Sven Bambach, Stefan Lee, David J Crandall, and Chen Yu. Lending a hand: Detecting hands and recognizing activities in complex egocentric interactions. In Proceedings of the IEEE international conference on computer vision, pages 1949-1957, 2015. 3", + "[7] William Berrios, Gautam Mittal, Tristan Thrush, Douwe Kiela, and Amanpreet Singh. Towards language models that can see: Computer vision through the lens of natural language. arXiv preprint arXiv:2306.16410, 2023. 2", + "[8] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. Advances in neural information processing systems, 33:1877-1901, 2020. 1, 2", + "[9] Soravit Changpinyo, Piyush Sharma, Nan Ding, and Radu Soricut. Conceptual 12m: Pushing web-scale image-text pretraining to recognize long-tail visual concepts. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3558-3568, 2021. 2", + "[10] Lin Chen, Jisong Li, Xiaoyi Dong, Pan Zhang, Conghui He, Jiaqi Wang, Feng Zhao, and Dahua Lin. Sharegpt4v: Improving large multi-modal models with better captions. arXiv preprint arXiv:2311.12793, 2023. 4, 1", + "[11] Liang Chen, Yichi Zhang, Shuhuai Ren, Haozhe Zhao, Zefan Cai, Yuchi Wang, Peiyi Wang, Tianyu Liu, and Baobao Chang. Towards end-to-end embodied decision making via multi-modal large language model: Explorations with gpt4- vision and beyond, 2023. 2" + ], + "bbox": [ + 78, + 114, + 470, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[12] Wei-Lin Chiang, Zhuohan Li, Zi Lin, Ying Sheng, Zhang-hao Wu, Hao Zhang, Lianmin Zheng, Siyuan Zhuang, Yong-hao Zhuang, Joseph E. Gonzalez, Ion Stoica, and Eric P. Xing. Vicuna: An open-source chatbot impressing gpt-4 with $90\\%$ * chatgpt quality, 2023. 2", + "[13] Wenliang Dai, Junnan Li, Dongxu Li, Anthony Meng Huat Tiong, Junqi Zhao, Weisheng Wang, Boyang Li, Pascale Fung, and Steven Hoi. Instructlip: Towards general-purpose vision-language models with instruction tuning, 2023. 1, 2, 4", + "[14] Tien Do, Ondrej Miksik, Joseph DeGol, Hyun Soo Park, and Sudipta N Sinha. Learning to detect scene landmarks for camera localization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11132-11142, 2022. 3", + "[15] Danny Driess, Fei Xia, Mehdi SM Sajjadi, Corey Lynch, Aakanksha Chowdhery, Brian Ichter, Ayzaan Wahid, Jonathan Thompson, Quan Vuong, Tianhe Yu, et al. Palme: An embodied multimodal language model. arXiv preprint arXiv:2303.03378, 2023. 1", + "[16] Chenyou Fan. Egovqa-an egocentric video question answering benchmark dataset. In Proceedings of the IEEE/CVF International Conference on Computer Vision Workshops, pages 0-0, 2019. 2", + "[17] Ali Farhadi, Ian Endres, Derek Hoiem, and David Forsyth. Describing objects by their attributes. In 2009 IEEE conference on computer vision and pattern recognition, pages 1778-1785. IEEE, 2009. 3", + "[18] Alireza Fathi, Ali Farhadi, and James M Rehg. Understanding egocentric activities. In 2011 international conference on computer vision, pages 407-414. IEEE, 2011. 3", + "[19] Chaoyou Fu, Peixian Chen, Yunhang Shen, Yulei Qin, Mengdan Zhang, Xu Lin, Zhenyu Qiu, Wei Lin, Jinrui Yang, Xiawu Zheng, et al. Mme: A comprehensive evaluation benchmark for multimodal large language models. arXiv preprint arXiv:2306.13394, 2023. 2", + "[20] Antonino Furnari, Sebastiano Battiato, Kristen Grauman, and Giovanni Maria Farinella. Next-active-object prediction from egocentric videos. Journal of Visual Communication and Image Representation, 49:401-411, 2017. 3", + "[21] Jensen Gao, Bidipta Sarkar, Fei Xia, Ted Xiao, Jiajun Wu, Brian Ichter, Anirudha Majumdar, and Dorsa Sadigh. Physically grounded vision-language models for robotic manipulation. arXiv preprint arXiv:2309.02561, 2023. 1", + "[22] Peng Gao, Jiaming Han, Renrui Zhang, Ziyi Lin, Shijie Geng, Aojun Zhou, Wei Zhang, Pan Lu, Conghui He, Xi-angyu Yue, et al. Llama-adapter v2: Parameter-efficient visual instruction model. arXiv preprint arXiv:2304.15010, 2023. 4, 1", + "[23] Yash Goyal, Tejas Khot, Douglas Summers-Stay, Dhruv Bartra, and Devi Parikh. Making the v in vqa matter: Elevating the role of image understanding in visual question answering. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6904-6913, 2017. 2", + "[24] Stéphane Grade, Mauro Pesenti, and Martin G Edwards. Evidence for the embodiment of space perception: concurrent hand but not arm action moderates reachability and egocen-" + ], + "bbox": [ + 501, + 92, + 893, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "14299", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "tric distance perception. Frontiers in Psychology, 6:862, 2015. 3", + "[25] Kristen Grauman, Andrew Westbury, Eugene Byrne, Zachary Chavis, Antonino Furnari, Rohit Girdhar, Jackson Hamburger, Hao Jiang, Miao Liu, Xingyu Liu, et al. Ego4d: Around the world in 3,000 hours of egocentric video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18995-19012, 2022. 3, 4", + "[26] Jean-Bernard Hayet, Frédéric Lerasle, and Michel Devy. Visual landmarks detection and recognition for mobile robot navigation. In 2003 IEEE Computer Society Conference on Computer Vision and Pattern Recognition, 2003. Proceedings., pages II-II. IEEE, 2003. 3", + "[27] Evert Helms, Rolf Dieter Schraft, and M Hagele. rob@ work: Robot assistant in industrial environments. In Proceedings. 11th IEEE International Workshop on Robot and Human Interactive Communication, pages 399-404. IEEE, 2002. 4", + "[28] Tucker Hermans, James M Rehg, and Aaron Bobick. Affordance prediction via learned object attributes. In IEEE international conference on robotics and automation (ICRA): Workshop on semantic perception, mapping, and exploration, pages 181-184. CiteSeer, 2011. 3", + "[29] MD Zakir Hossain, Ferdous Sohel, Mohd Fairuz Shiratuddin, and Hamid Laga. A comprehensive survey of deep learning for image captioning. ACM Computing Surveys (CsUR), 51(6):1-36, 2019. 2", + "[30] Wenlong Huang, Fei Xia, Ted Xiao, Harris Chan, Jacky Liang, Pete Florence, Andy Zeng, Jonathan Tompson, Igor Mordatch, Yevgen Chebotar, et al. Inner monologue: Embodied reasoning through planning with language models. arXiv preprint arXiv:2207.05608, 2022. 4", + "[31] Wenlong Huang, Chen Wang, Ruohan Zhang, Yunzhu Li, Jiajun Wu, and Li Fei-Fei. Voxposer: Composable 3d value maps for robotic manipulation with language models. arXiv preprint arXiv:2307.05973, 2023. 1", + "[32] Yichao Huang, Xiaorui Liu, Xin Zhang, and Lianwen Jin. A pointing gesture based egocentric interaction system: Dataset, approach and application. In Proceedings of the IEEE conference on computer vision and pattern recognition workshops, pages 16-23, 2016. 2", + "[33] Drew A Hudson and Christopher D Manning. Gqa: A new dataset for real-world visual reasoning and compositional question answering. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6700-6709, 2019. 2", + "[34] Baoxiong Jia, Ting Lei, Song-Chun Zhu, and Siyuan Huang. Egotaskqa: Understanding human tasks in egocentric videos. Advances in Neural Information Processing Systems, 35: 3343-3360, 2022. 2", + "[35] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In International conference on machine learning, pages 4904-4916. PMLR, 2021. 2" + ], + "bbox": [ + 78, + 90, + 470, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[36] Charmi Jobanputra, Jatna Bavishi, and Nishant Doshi. Human activity recognition: A survey. Procedia Computer Science, 155:698-703, 2019. 3", + "[37] Nancy Kanwisher and Jon Driver. Objects, attributes, and visual attention: Which, what, and where. Current Directions in Psychological Science, 1(1):26-31, 1992. 3", + "[38] Shian-Ru Ke, Hoang Le Uyen Thuc, Yong-Jin Lee, Jenq-Neng Hwang, Jang-Hee Yoo, and Kyoung-Ho Choi. A review on video-based human activity recognition. Computers, 2(2):88-131, 2013. 3", + "[39] Roberta L Klatzky. Allocentric and egocentric spatial representations: Definitions, distinctions, and interconnections. In Spatial cognition: An interdisciplinary approach to representing and processing spatial knowledge, pages 1-17. Springer, 1998. 3", + "[40] Weicheng Kuo, Yin Cui, Xiuye Gu, AJ Piergiovanni, and Anelia Angelova. F-vlm: Open-vocabulary object detection upon frozen vision and language models. arXiv preprint arXiv:2209.15639, 2022. 1", + "[41] Bo Li, Yuanhan Zhang, Liangyu Chen, Jinghao Wang, Fanyi Pu, Jingkang Yang, Chunyuan Li, and Ziwei Liu. Mimic-it: Multi-modal in-context instruction tuning. arXiv preprint arXiv:2306.05425, 2023. 2", + "[42] Bo Li, Yuanhan Zhang, Liangyu Chen, Jinghao Wang, Jingkang Yang, and Ziwei Liu. Otter: A multi-modal model with in-context instruction tuning. arXiv preprint arXiv:2305.03726, 2023. 1, 2, 4", + "[43] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. arXiv preprint arXiv:2301.12597, 2023. 1, 4", + "[44] KunChang Li, Yinan He, Yi Wang, Yizhuo Li, Wenhai Wang, Ping Luo, Yali Wang, Limin Wang, and Yu Qiao. Videochat: Chat-centric video understanding. arXiv preprint arXiv:2305.06355, 2023. 4, 1", + "[45] Bin Lin, Bin Zhu, Yang Ye, Munan Ning, Peng Jin, and Li Yuan. Video-llava: Learning united visual representation by alignment before projection. arXiv preprint arXiv:2311.10122, 2023. 4, 1", + "[46] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 2", + "[47] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. arXiv preprint arXiv:2310.03744, 2023. 4, 1", + "[48] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning, 2023. 1, 2, 4", + "[49] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. arXiv preprint arXiv:2304.08485, 2023. 2", + "[50] Li Liu, Wanli Ouyang, Xiaogang Wang, Paul Fieguth, Jie Chen, Xinwang Liu, and Matti Pietikainen. Deep learning for generic object detection: A survey. International journal of computer vision, 128:261-318, 2020. 3" + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "14300", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[51] Miao Liu, Siyu Tang, Yin Li, and James M Rehg. Forecasting human-object interaction: Joint prediction of motor attention and egocentric activity. Computer Vision-ECCV 2020, 12346:704-721, 2020. 3", + "[52] Shaowei Liu, Subarna Tripathi, Somdeb Majumdar, and Xiaolong Wang. Joint hand motion and interaction hotspots prediction from egocentric videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3282-3292, 2022. 3", + "[53] Yunze Liu, Yun Liu, Che Jiang, Kangbo Lyu, Weikang Wan, Hao Shen, Boqiang Liang, Zhoujie Fu, He Wang, and Li Yi. Hoi4d: A 4d egocentric dataset for category-level human-object interaction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21013-21022, 2022. 2, 3", + "[54] Yuan Liu, Haodong Duan, Yuanhan Zhang, Bo Li, Songyang Zhang, Wangbo Zhao, Yike Yuan, Jiaqi Wang, Conghui He, Ziwei Liu, et al. Mmbench: Is your multi-modal model an all-around player? arXiv preprint arXiv:2307.06281, 2023. 2, 5", + "[55] Michael Montemerlo and Sebastian Thrun. FastSLAM: A scalable method for the simultaneous localization and mapping problem in robotics. Springer, 2007. 3", + "[56] Luis Montesano, Manuel Lopes, Alexandre Bernardino, and José Santos-Victor. Learning object affordances: from sensory-motor coordination to imitation. IEEE Transactions on Robotics, 24(1):15-26, 2008. 3", + "[57] Francesca Morganti, Stefano Stefanini, and Giuseppe Riva. From allo-to egocentric spatial ability in early alzheimer's disease: a study with virtual reality spatial tasks. Cognitive neuroscience, 4(3-4):171-180, 2013. 3", + "[58] Weimin Mou, Timothy P McNamara, Christine M Vali-quette, and Björn Rump. Allocentric and egocentric updating of spatial memories. Journal of experimental psychology: Learning, Memory, and Cognition, 30(1):142, 2004. 3", + "[59] Thi-Hoa-Cuc Nguyen, Jean-Christophe Nebel, and Francisco Florez-Revuelta. Recognition of activities of daily living with egocentric vision: A review. Sensors, 16(1):72, 2016. 3", + "[60] OpenAI. Gpt-4 technical report, 2023. 1, 2, 4, 5", + "[61] Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, et al. Training language models to follow instructions with human feedback. Advances in Neural Information Processing Systems, 35: 27730-27744, 2022. 2", + "[62] Anish Pandey, Rakesh Kumar Sonkar, Krishna Kant Pandey, and DR Parhi. Path planning navigation of mobile robot with obstacles avoidance using fuzzy logic controller. In 2014 IEEE 8th international conference on intelligent systems and control (ISCO), pages 39-41. IEEE, 2014. 4", + "[63] BK Patle, Anish Pandey, DRK Parhi, AJDT Jagadeesh, et al. A review: On path planning strategies for navigation of mobile robot. Defence Technology, 15(4):582-606, 2019. 4", + "[64] Bryan A Plummer, Liwei Wang, Chris M Cervantes, Juan C Caicedo, Julia Hockenmaier, and Svetlana Lazebnik. Flickr30k entities: Collecting region-to-phrase corre" + ], + "bbox": [ + 78, + 90, + 468, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "spondences for richer image-to-sentence models. In Proceedings of the IEEE international conference on computer vision, pages 2641-2649, 2015. 2", + "[65] Vignesh Ramanathan, Anmol Kalia, Vladan Petrovic, Yi Wen, Baixue Zheng, Baishan Guo, Rui Wang, Aaron Marquez, Rama Kovvuri, Abhishek Kadian, et al. Paco: Parts and attributes of common objects. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7141-7151, 2023. 3", + "[66] Sajad Saeedi, Michael Trentini, Mae Seto, and Howard Li. Multiple-robot simultaneous localization and mapping: A review. Journal of Field Robotics, 33(1):3-46, 2016. 3", + "[67] Dustin Schwenk, Apoorv Khandelwal, Christopher Clark, Kenneth Marino, and Roozbeh Mottaghi. A-okvqa: A benchmark for visual question answering using world knowledge. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23–27, 2022, Proceedings, Part VIII, pages 146–162. Springer, 2022. 2", + "[68] Wenqi Shao, Yutao Hu, Peng Gao, Meng Lei, Kaipeng Zhang, Fanqing Meng, Peng Xu, Siyuan Huang, Hongsheng Li, Yu Qiao, et al. Tiny lvlm-ehub: Early multimodal experiments with bard. arXiv preprint arXiv:2308.03729, 2023. 2, 5", + "[69] Chan Hee Song, Jiaman Wu, Clayton Washington, Brian M Sadler, Wei-Lun Chao, and Yu Su. Llm-planner: Few-shot grounded planning for embodied agents with large language models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2998-3009, 2023. 4", + "[70] Yixuan Su, Tian Lan, Huayang Li, Jialu Xu, Yan Wang, and Deng Cai. Pandagpt: One model to instruction-follow them all. arXiv preprint arXiv:2305.16355, 2023. 4, 1", + "[71] Theodore Sumers, Kenneth Marino, Arun Ahuja, Rob Fergus, and Ishita Dasgupta. Distilling internet-scale vision-language models into embodied agents. arXiv preprint arXiv:2301.12507, 2023. 1", + "[72] Loreto Susperregi, Izaskun Fernandez, Ane Fernandez, Santiago Fernandez, Iñaki Maurtua, and Irene Lopez de Vallejo. Interacting with a robot: a guide robot understanding natural language instructions. In International Conference on Ubiquitous Computing and Ambient Intelligence, pages 185-192. Springer, 2012. 4", + "[73] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 1", + "[74] Michalis Vrigkas, Christophoros Nikou, and Ioannis A Kakadiaris. A review of human activity recognition methods. Frontiers in Robotics and AI, 2:28, 2015. 3", + "[75] Guan Wang, Sijie Cheng, Xianyuan Zhan, Xiangang Li, Sen Song, and Yang Liu. Openchat: Advancing open-source language models with mixed-quality data. arXiv preprint arXiv:2309.11235, 2023. 2", + "[76] Weronika Wojtak, Flora Ferreira, Paulo Vicente, Luís Louro, Estela Bicho, and Wolfram Erlhagen. A neural integrator model for planning and value-based decision making of a robotics assistant. Neural Computing and Applications, 33: 3737-3756, 2021. 4" + ], + "bbox": [ + 503, + 92, + 890, + 898 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "14301", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[77] Peng Xu, Wenqi Shao, Kaipeng Zhang, Peng Gao, Shuo Liu, Meng Lei, Fanqing Meng, Siyuan Huang, Yu Qiao, and Ping Luo. Lvlm-ehub: A comprehensive evaluation benchmark for large vision-language models. arXiv preprint arXiv:2306.09265, 2023. 2", + "[78] Jingkang Yang, Yuhao Dong, Shuai Liu, Bo Li, Ziyue Wang, Chencheng Jiang, Haoran Tan, Jiamu Kang, Yuanhan Zhang, Kaiyang Zhou, et al. Octopus: Embodied vision-language programmer from environmental feedback. arXiv preprint arXiv:2310.08588, 2023. 1", + "[79] Jianwei Yang, Hao Zhang, Feng Li, Xueyan Zou, Chunyuan Li, and Jianfeng Gao. Set-of-mark prompting unleashes extraordinary visual grounding in gpt-4v. arXiv preprint arXiv:2310.11441, 2023. 8", + "[80] Zhengyuan Yang, Linjie Li, Kevin Lin, Jianfeng Wang, Chung-Ching Lin, Zicheng Liu, and Lijuan Wang. The dawn of Imms: Preliminary explorations with gpt-4v (isdiction). arXiv preprint arXiv:2309.17421, 2023. 1, 7", + "[81] Qinghao Ye, Haiyang Xu, Guohai Xu, Jiabo Ye, Ming Yan, Yiyang Zhou, Junyang Wang, Anwen Hu, Pengcheng Shi, Yaya Shi, Chaoya Jiang, Chenliang Li, Yuanhong Xu, Hehong Chen, Junfeng Tian, Qian Qi, Ji Zhang, and Fei Huang. mplug-owl: Modularization empowers large language models with multimodality, 2023. 1, 2, 4", + "[82] Quanzeng You, Hailin Jin, Zhaowen Wang, Chen Fang, and Jiebo Luo. Image captioning with semantic attention. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4651-4659, 2016. 2", + "[83] Yifan Zhang, Congqi Cao, Jian Cheng, and Hanqing Lu. Egogesture: A new dataset and benchmark for egocentric hand gesture recognition. IEEE Transactions on Multimedia, 20(5):1038-1050, 2018. 2", + "[84] Tiancheng Zhao, Tianqi Zhang, Mingwei Zhu, Haozhan Shen, Kyusong Lee, Xiaopeng Lu, and Jianwei Yin. Vl-checklist: Evaluating pre-trained vision-language models with objects, attributes and relations. arXiv preprint arXiv:2207.00221, 2022. 2", + "[85] Zhong-Qiu Zhao, Peng Zheng, Shou-tao Xu, and Xindong Wu. Object detection with deep learning: A review. IEEE transactions on neural networks and learning systems, 30 (11):3212-3232, 2019. 3", + "[86] Lianmin Zheng, Wei-Lin Chiang, Ying Sheng, Siyuan Zhuang, Zhanghao Wu, Yonghao Zhuang, Zi Lin, Zhuohan Li, Dacheng Li, Eric Xing, et al. Judging llm-as-a-judge with mt-bench and chatbot arena. arXiv preprint arXiv:2306.05685, 2023. 5, 2", + "[87] Wangchunshu Zhou, Yan Zeng, Shizhe Diao, and Xinsong Zhang. Vlue: A multi-task benchmark for evaluating vision-language models. arXiv preprint arXiv:2205.15237, 2022. 2", + "[88] Chenchen Zhu, Fanyi Xiao, Andres Alvarado, Yasmine Babaei, Jiabo Hu, Hichem El-Mohri, Sean Culatana, Roshan Sumbaly, and Zhicheng Yan. Egoobjects: A large-scale egocentric dataset for fine-grained object understanding. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 20110-20120, 2023. 2, 3", + "[89] Deyao Zhu, Jun Chen, Xiaogian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language" + ], + "bbox": [ + 78, + 90, + 468, + 902 + ], + "page_idx": 11 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "understanding with advanced large language models. arXiv preprint arXiv:2304.10592, 2023. 1, 2, 4", + "[90] Wanrong Zhu, Jack Hessel, Anas Awadalla, Samir Yitzhak Gadre, Jesse Dodge, Alex Fang, Youngjae Yu, Ludwig Schmidt, William Yang Wang, and Yejin Choi. Multimodal c4: An open, billion-scale corpus of images interleaved with text. arXiv preprint arXiv:2304.06939, 2023. 2", + "[91] Zhengxia Zou, Keyan Chen, Zhenwei Shi, Yuhong Guo, and Jieping Ye. Object detection in 20 years: A survey. Proceedings of the IEEE, 2023. 3" + ], + "bbox": [ + 501, + 90, + 890, + 233 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "14302", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 11 + } +] \ No newline at end of file diff --git a/2024/EgoThink_ Evaluating First-Person Perspective Thinking Capability of Vision-Language Models/2ff0d147-b907-475e-8aa3-5584b71f3014_model.json b/2024/EgoThink_ Evaluating First-Person Perspective Thinking Capability of Vision-Language Models/2ff0d147-b907-475e-8aa3-5584b71f3014_model.json new file mode 100644 index 0000000000000000000000000000000000000000..fa2871aeeb05f85606cde580b91aeb0d1685d444 --- /dev/null +++ b/2024/EgoThink_ Evaluating First-Person Perspective Thinking Capability of Vision-Language Models/2ff0d147-b907-475e-8aa3-5584b71f3014_model.json @@ -0,0 +1,2413 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.0, + 0.812, + 0.045 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.128, + 0.13, + 0.844, + 0.176 + ], + "angle": 0, + "content": "EgoThink: Evaluating First-Person Perspective Thinking Capability of Vision-Language Models" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.202, + 0.753, + 0.239 + ], + "angle": 0, + "content": "Sijie Cheng\\(^{1,2,5 \\ddagger}\\), Zhicheng Guo\\(^{1,2*}\\), Jingwen Wu\\(^{3*}\\), Kechen Fang\\(^{4}\\), Peng Li\\(^{2 \\boxtimes}\\), Huaping Liu\\(^{1}\\), Yang Liu\\(^{1,2 \\boxtimes}\\)" + }, + { + "type": "text", + "bbox": [ + 0.185, + 0.24, + 0.786, + 0.327 + ], + "angle": 0, + "content": "\\(^{1}\\)Department of Computer Science and Technology, Tsinghua University \\(^{2}\\)Institute for AI Industry Research (AIR), Tsinghua University \\(^{3}\\)Department of Electrical and Computer Engineering, University of Toronto \\(^{4}\\)Zhili College, Tsinghua University \\(^{5}\\)01.AI csj23@mails.tsinghua.edu.cn" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.362, + 0.314, + 0.379 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.399, + 0.474, + 0.807 + ], + "angle": 0, + "content": "Vision-language models (VLMs) have recently shown promising results in traditional downstream tasks. Evaluation studies have emerged to assess their abilities, with the majority focusing on the third-person perspective, and only a few addressing specific tasks from the first-person perspective. However, the capability of VLMs to \"think\" from a first-person perspective, a crucial attribute for advancing autonomous agents and robotics, remains largely unexplored. To bridge this research gap, we introduce EgoThink, a novel visual question-answering benchmark that encompasses six core capabilities with twelve detailed dimensions. The benchmark is constructed using selected clips from egocentric videos, with manually annotated question-answer pairs containing first-person information. To comprehensively assess VLMs, we evaluate twenty-one popular VLMs on EgoThink. Moreover, given the open-ended format of the answers, we use GPT-4 as the automatic judge to compute single-answer grading. Experimental results indicate that although GPT-4V leads in numerous dimensions, all evaluated VLMs still possess considerable potential for improvement in first-person perspective tasks. Meanwhile, enlarging the number of trainable parameters has the most significant impact on model performance on EgoThink. In conclusion, EgoThink serves as a valuable addition to existing evaluation benchmarks for VLMs, providing an indispensable resource for future research in the realm of embodied artificial intelligence and robotics." + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.365, + 0.891, + 0.525 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.659, + 0.532, + 0.687, + 0.554 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.688, + 0.539, + 0.76, + 0.555 + ], + "angle": 0, + "content": "EgoThink" + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.559, + 0.895, + 0.603 + ], + "angle": 0, + "content": "Figure 1. The main categories of our EgoThink benchmark to comprehensively assess the capability of thinking from a first-person perspective." + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.63, + 0.633, + 0.646 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.657, + 0.892, + 0.869 + ], + "angle": 0, + "content": "Benefiting from the rapid development of large language models (LLMs) [8, 60, 73], vision-language models (VLMs) [2, 15, 43, 80] have shown remarkable progress in both conventional vision-language downstream tasks [2, 15, 43, 80] and following diverse human instructions [13, 42, 48, 81, 89]. Their application has expanded into broader domains such as robotics [21, 31, 40] and embodied artificial intelligence (EAI) [71, 78]. As a result, the thorough evaluation of VLMs has become increasingly important and challenging. Observing and understanding the world from a first-person perspective is a natural approach for both humans and artificial intelligence agents. We propose that the ability to \"think\" from a first-person perspective, especially when interpreting egocentric images, is crucial for VLMs." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.892, + 0.902 + ], + "angle": 0, + "content": "However, as shown in Table 1, the ability to think from a first-person perspective is not adequately addressed by cur" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.85, + 0.415, + 0.865 + ], + "angle": 0, + "content": "*Equal contribution, ‡ Project lead, Corresponding author" + }, + { + "type": "page_footnote", + "bbox": [ + 0.103, + 0.865, + 0.373, + 0.876 + ], + "angle": 0, + "content": "Project page: https://adacheng.github.io/EgoThink/" + }, + { + "type": "page_footnote", + "bbox": [ + 0.104, + 0.877, + 0.394, + 0.888 + ], + "angle": 0, + "content": "GitHub page: https://github.com/AdaCheng/EgoThink/" + }, + { + "type": "page_footnote", + "bbox": [ + 0.104, + 0.889, + 0.457, + 0.9 + ], + "angle": 0, + "content": "Dataset page: https://huggingface.co/datasets/EgoThink/EgoThink/" + }, + { + "type": "list", + "bbox": [ + 0.096, + 0.85, + 0.457, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "14291" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.082, + 0.089, + 0.894, + 0.226 + ], + "angle": 0, + "content": "
BenchmarkCapabilityPerspectiveData SourceAnswer TypeEvaluatorSize
VL-CheckList [84]Object / Attribute / RelationThirdDatasetsPSAccuracy410k
LVLM-eHub [77]General Multi-ModalityThirdDatasetsMC / OEMetrics / LLMs / User332k
MME [19]General Multi-ModalityThirdHandcraftMCAccuracy2,194
Tiny LVLM-eHub [68]General Multi-ModalityThirdDatasetsOELLMs2,100
MMBench [54]General Multi-ModalityThirdDatasets / Handcraft / LLMsMCLLMs2,974
PCA-EVAL [11]Decision-MakingThirdHandcraftOEAccuracy / User300
EgoTaskQA [34]Spatial / Temporal / CausalFirstCrowdsourcingOECrowdsourcing40k
EgoVQA [16]Object / Action / PersonThird / FirstHandcraftMCAccuracy520
EgoThink (Ours)First-Person ThinkingFirstHandcraftOELLMs700
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.234, + 0.893, + 0.278 + ], + "angle": 0, + "content": "Table 1. Comparison of recent comprehensive evaluation benchmarks of VLMs and our proposed benchmark EgoThink. Third and first indicate third-person and first-person perspectives. Datasets/Handcraft/LLMs denote existing datasets, manual annotation, and automatic generation by LLMs. PS/MC/OE indicate pairwise scoring, multi-choice, and open-ended question-answering, respectively." + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.292, + 0.47, + 0.444 + ], + "angle": 0, + "content": "rent evaluation benchmarks for VLMs. On one hand, most of these benchmarks (six out of nine, as listed in Table 1) focus solely on the third-person perspective. On the other hand, those benchmarks that do consider the first-person perspective only encompass a limited range of capabilities. For instance, EgoTaskQA [34] examines spatial, temporal, and causal aspects, whereas EgoVQA [16] is limited to object, action, and person aspects. Therefore, there is a clear need to develop a comprehensive benchmark to evaluate the first-person capabilities of VLMs more effectively." + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.446, + 0.471, + 0.809 + ], + "angle": 0, + "content": "In this work, we introduce a new benchmark for VLMs from a first-person perspective, named EgoThink. The initial step in developing this benchmark involves determining the necessary capabilities to assess. Humans, when interacting with the real world, consider a series of questions centered on themselves, ranging from \"What is around me?\", \"What am I doing?\", \"Where am I?\", \"What about the situation around me?\", \"What will happen to me?\" to \"How will I do?\". Drawing inspiration from this, we evaluate six core capabilities of VLMs, namely object, activity, localization, reasoning, forecasting, and planning. Each capability corresponds to one of the aforementioned questions, as illustrated in Figure 1. The next step is constructing the benchmark. We first categorize the six core capabilities into twelve detailed dimensions. We then select a minimum of 50 distinct and clear clips from egocentric videos for each dimension and manually annotate them with relevant first-person question-answer pairs. This approach ensures the quality and variety of the benchmark. The final step is evaluating VLM performance on this benchmark. Building on recent studies [7, 12, 68], we use GPT-4 [60] as an automatic evaluator. The Pearson correlation coefficient, when compared with human evaluation, shows a value of 0.68, indicating that the evaluation results are dependable." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.811, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Based on our proposed EgoThink benchmark, we conduct comprehensive experiments to evaluate the first-person capabilities of twenty-one popular VLMs with varying model and data compositions. The findings indicate that GPT-4V stands out as the most effective model in various aspects. However, it shows less impressive results in spe" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.292, + 0.895, + 0.414 + ], + "angle": 0, + "content": "cific capabilities such as activity and counting. Additionally, we observed that no single VLM consistently surpasses others in every aspect. For instance, GPT-4V is less effective than BLIP-2-11B for localization. Increasing the language model portion of the VLMs generally leads to better performance, but this improvement is not uniform across all models. Finally, our results highlight a significant potential for further enhancing the first-person capabilities of VLMs." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.433, + 0.642, + 0.449 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.461, + 0.892, + 0.643 + ], + "angle": 0, + "content": "Vision-Language Models. Inspired by the impressive success of LLMs [8, 61, 75], the recent popular VLMs tend to regard the powerful LLMs as the core backbone. At the beginning, VLMs usually use large-scale image-text pairwise datasets [9, 35, 46] or arbitrarily interleaved visual and textual data [2, 90] to pre-train. Furthermore, thanks to the availability of enormous image-text instruction datasets [41, 49], recent studies [13, 42, 48, 81, 89] further apply instruction tuning to help VLMs generate satisfactory answers. Benefiting from the two-stage training process, recent VLMs can achieve stunning performance on downstream vision-language tasks [3, 33, 46, 64]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.644, + 0.893, + 0.903 + ], + "angle": 0, + "content": "Evaluations of VLMs. To evaluate the abilities of VLMs, there are diverse types of vision language downstream tasks. Conventional benchmarks, such as image caption tasks [29, 82] and visual question reasoning tasks [23, 67], mainly probe specific abilities of VLMs from the third-person perspective. Meanwhile, specialized analytical studies comprehensively evaluate the performance of VLMs from the third-person perspective, where Vblue [87] consists of five fundamental tasks and Lvmehub [77] evaluates six categories of capabilities on 47 standard vision-language benchmarks. As for the first-person perspective, there are some egocentric evaluation benchmarks in the computer vision field to assess some visual capabilities [32, 53, 83, 88]. In terms of multi-modality, there are a few benchmarks, such as EgoVQA[16] and EgoTaskQA [34], where mainly specific tasks without an overall understanding. In this paper, we mainly focus on exploring the comprehensive capa" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "14292" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.08, + 0.089, + 0.482, + 0.269 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.081, + 0.27, + 0.482, + 0.337 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.08, + 0.337, + 0.482, + 0.464 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.487, + 0.089, + 0.888, + 0.269 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.487, + 0.27, + 0.888, + 0.337 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.487, + 0.337, + 0.888, + 0.464 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.158, + 0.475, + 0.809, + 0.49 + ], + "angle": 0, + "content": "Figure 2. Categories with fine-grained dimensions and their corresponding examples of EgoThink benchmark." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.506, + 0.47, + 0.536 + ], + "angle": 0, + "content": "bilities of VLMs to think from a first-person perspective, as a supplement to previous evaluation benchmarks." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.551, + 0.289, + 0.567 + ], + "angle": 0, + "content": "3. EgoThink Benchmark" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.577, + 0.469, + 0.653 + ], + "angle": 0, + "content": "In this section, we first elaborate on the core capabilities of thinking from a first-person perspective. Then, we introduce the process to manually construct our proposed benchmark EgoThink, which asks VLMs to generate open-ended answers according to first-person images and questions." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.663, + 0.248, + 0.679 + ], + "angle": 0, + "content": "3.1. Core Capabilities" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.687, + 0.469, + 0.732 + ], + "angle": 0, + "content": "As shown in Figure 2, we specifically design six categories with twelve fine-grained dimensions from the first-person perspective for quantitative evaluation." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.735, + 0.47, + 0.902 + ], + "angle": 0, + "content": "- Object: What is around me? Recognizing objects in the real world is a preliminary ability of the human visual system [50, 85, 91]. Images from a first-person or egocentric perspective [53, 65, 88] pay more attention to the objects surrounding the subject or in hands. Moreover, we further divide the object category into three fine-grained dimensions: (1) Existence, predicting whether there is an object as described in the images; (2) Attribute [17, 37], detecting properties or characteristics (e.g., color) of an object; (3) Affordance [28, 56], predicting potential actions that a human can apply to an object." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.506, + 0.891, + 0.581 + ], + "angle": 0, + "content": "- Activity: What am I doing? Activity recognition is to automatically recognize specific human activities in video frames or still images [36, 38, 74]. From the egocentric perspective, we mainly focus on actions or activities based on object-hand interaction [6, 18, 59]." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.581, + 0.892, + 0.717 + ], + "angle": 0, + "content": "- Localization: Where am I? In reality, localization is a critical capability for navigation and scene understanding in the real world [55, 66]. Here we investigate the localization capability from two aspects, Location and Spatial Relationship. Location indicates detecting the scene surrounding the subject [14, 26]. Spatial reasoning contains allocentric and egocentric perspectives [24, 39, 57, 58]. We focus on the egocentric perspective, i.e., the position of the object with respect to the subject." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.72, + 0.892, + 0.856 + ], + "angle": 0, + "content": "- Reasoning: What about the situation around me? During the complex decision-making process, reasoning lies everywhere in our lives. Here we mainly focus on Counting, Comparison, and Situated Reasoning. Due to the first-person perspective, we generally count or compare objects in our hands or surrounding ourselves. As for situated reasoning, we employ cases that cannot be answered directly from the information in the images and require further reasoning processes." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.857, + 0.892, + 0.902 + ], + "angle": 0, + "content": "- Forecasting: What will happen to me? Forecasting [20, 25, 51, 52] is a critical skill in the real world. From an egocentric view, forecasting always predicts the future of" + }, + { + "type": "list", + "bbox": [ + 0.5, + 0.506, + 0.892, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "14293" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.081, + 0.089, + 0.892, + 0.466 + ], + "angle": 0, + "content": "
VLMsImage EncoderLLMAlignment ModuleTTPToPDataset SizeEgoDataVideo
Image-TextInstruction
API-based Model
GPT-4V [60]Unknown
~7B Models
OpenFlamingo-7B [2, 5]CLIP-ViT-LMPT7BAttention1.4B8.1B2B-X✓
BLIP-2-6.7B [43]EVA-CLIP-ViT-gOPT6.7BQ-Former108M7.8B129M-XX
VideoChat-7B [44]BLIP2-VEVicuna7BQ-Former205M8B25M18KX✓
LLaVA-1.5-7B [47]CLIP-ViT-L-336pxLlama27BMLP6.8B7.1B558k665kXX
MiniGPT-4-7B [89]BLIP2-VELlama27BLinear23M7.7B5M3.5kXX
InstructBLIP-7B [13]EVA-CLIP-ViT-gVicuna7BQ-Former189M7.9B-16MXX
LLaMA-Adapter-7B [22]CLIP-ViT-LLLaMA7BEarly Fusion14M7.2B567k52kXX
Otter-I-7B [42]CLIP-ViT-LMPT7BAttention1.4B8.1B-2.8BX✓
PandaGPT-7B [70]ImageBindVicuna7BLinear + LLM LoRA38M7.9B-160k✓✓
mPLUG-owl-7B [81]CLIP-ViT-LLLaMA7BAttention4M7.1B204M158kXX
Video-LLaVA-7B [45]LanguageBindVicuna7BLinear6.8B7.5B1260k765kX✓
LLaVA-7B [48]CLIP-ViT-LLlama27BLinear6.7B7.1B595k158kXX
ShareGPT4V-7B [10]CLIP-ViT-L-336pxVicuna7BMLP6.7B6.7B1.2M665kXX
~13B Models
InstructBLIP-13B [13]EVA-CLIP-ViT-gVicuna13BQ-Former189M14.2B-16MXX
PandaGPT-13B [70]ImageBindVicuna13BLinear+LLM LoRA52M13.1B-160k✓✓
LLaVA-13B-Vicuna [48]CLIP-ViT-L-336pxVicuna13BLinear13.0B13.3B595k158kXX
BLIP-2-11B [43]EVA-CLIP-ViT-gFlanT5XXLQ-Former108M12.2B129M-XX
InstructBLIP-11B [13]EVA-CLIP-ViT-gFlanT5XXLQ-Former189M12.3B-16MXX
LLaVA-13B-Llama2 [48]CLIP-ViT-LLlama213BLinear13.0B13.3B595k158kXX
LLaVA-1.5-13B [47]CLIP-ViT-L-336pxLlama213BMLP13.0B13.4B558k665kXX
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.476, + 0.893, + 0.506 + ], + "angle": 0, + "content": "Table 2. Statistics of compared API-based and open-source VLMs, where TTP and ToP indicate Total Trainable Parameters and Total Parameters, respectively. Moreover, EgoData and Video indicate that there are egocentric visual data and video data for training, respectively." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.52, + 0.454, + 0.535 + ], + "angle": 0, + "content": "object-state transformation or hand-object interactions." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.535, + 0.47, + 0.626 + ], + "angle": 0, + "content": "- Planning: How will I do? In reality, planning [1, 30, 69] is an important capability to deal with complex problems, typically applied in Navigation [62, 63, 72] and Assistance [27, 76]. Navigation is going to a goal location from a start position, while assistance is offering instructions to solve daily problems." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.636, + 0.235, + 0.65 + ], + "angle": 0, + "content": "3.2. Data Collection" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.66, + 0.469, + 0.689 + ], + "angle": 0, + "content": "In this section, we mainly introduce the detailed processing to construct our EgoThink benchmark." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.689, + 0.47, + 0.901 + ], + "angle": 0, + "content": "Collecting first-person visual data. Firstly, we leverage a popular and large egocentric video dataset Ego4D [25], which is designed to advance the field of first-person perception in computer vision. To obtain a diverse representation in different scenarios, Ego4D encompasses 3,670 hours of video from 931 unique camera wearers spanning 74 global locations across 9 countries. To collect first-person visual data, we begin by extracting every frame from a subset of the Ego4D video dataset, yielding a diverse raw image dataset. Please note that our current focus is solely on images, as most VLMs today do not support video input. We intend to expand our scope to include videos in our future work. Considering the heavy human labor and the diversity of scenarios, we sample images every few dozen frames." + }, + { + "type": "image", + "bbox": [ + 0.523, + 0.519, + 0.868, + 0.676 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.679, + 0.892, + 0.735 + ], + "angle": 0, + "content": "Figure 3. This chart illustrates the distribution of various scene categories within the EgoThink dataset. The 'others' category encompasses 13 different scene types, each representing less than one percent of total scenes." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.749, + 0.892, + 0.87 + ], + "angle": 0, + "content": "To ensure high quality, we apply strict criteria for selecting the extracted frames. We first exclude images that lack clarity or fail to exhibit egocentric characteristics. Then, to obtain the high diversity within the dataset, we conduct a further screening to ensure that at most two images per video are included in the filtered image set. Finally, we obtain enormous high-quality images with exhibit egocentric characteristics as first-person image candidates." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.871, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Annotating question-answer pairs. Upon receiving a substantial collection of first-person image candidates, we" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "14294" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.081, + 0.089, + 0.892, + 0.442 + ], + "angle": 0, + "content": "
MethodsObjectActivityLocalizationReasoningForecastingPlanningAverage
ExistAttrAffordLocSpatialCountComparSituatedNavAssist
API-based model
GPT-4V62.082.058.059.586.062.042.048.083.055.064.084.065.5
~7B Models
OpenFlamingo-7B16.055.037.015.034.034.021.040.021.031.011.011.027.2
BLIP-2-6.7B49.029.039.033.560.031.03.021.033.025.08.06.028.1
VideoChat-7B46.044.036.045.061.042.036.039.032.026.513.021.036.8
LLaVA-1.5-7B33.047.054.035.535.049.020.047.037.027.029.054.039.0
MiniGPT-4-7B50.056.046.039.055.049.014.048.031.041.514.044.040.6
InstructBLIP-7B50.033.045.047.577.038.018.043.067.040.519.031.042.4
LLaMA-Adapter-7B37.060.046.034.548.051.029.039.025.041.542.057.042.5
Otter-I-7B48.056.039.044.060.044.039.048.042.038.031.055.045.3
PandaGPT-7B40.056.041.037.061.052.019.052.053.043.039.061.046.2
mPLUG-owl-7B56.058.047.053.060.053.025.049.044.049.533.058.048.8
Video-LLaVA-7B56.060.053.045.086.060.039.038.060.046.511.038.049.4
LLaVA-7B63.058.050.047.081.045.024.036.047.049.535.060.049.6
ShareGPT4V-7B67.075.053.055.577.062.030.038.066.047.041.063.051.9
~13B Models
InstructBLIP-13B52.055.049.054.063.049.011.033.059.044.019.025.042.8
PandaGPT-13B35.052.041.040.568.031.032.040.047.045.516.069.043.1
LLaVA-13B-Vicuna54.062.052.046.053.046.026.044.029.044.035.066.046.4
BLIP-2-11B52.062.041.049.590.066.025.050.070.048.018.024.049.6
InstructBLIP-11B74.068.048.049.586.052.032.049.073.053.016.017.051.5
LLaVA-13B-Llama265.061.045.056.077.053.034.034.066.050.549.071.055.1
LLaVA-1.5-13B66.055.051.055.082.057.032.056.067.048.539.055.055.3
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.453, + 0.893, + 0.496 + ], + "angle": 0, + "content": "Table 3. Combined single-answer grading scores on zero-shot setups for various dimensions. The **bold** indicates the best performance while the **underline** indicates the second-best performance. Exist, Attr, Afford, Loc, Spatial, Count, Compar, Situated, Nav and Assist represent existence, attribute, affordance, location, spatial relationship, counting, comparison, situated reasoning, navigation, and assistance." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.522, + 0.47, + 0.734 + ], + "angle": 0, + "content": "engage six annotators to manually label question-answer pairs. Given that the EgoThink benchmark is composed of twelve dimensions, annotators were responsible for two specific dimensions. The annotators can access all the image candidates and are asked to select appropriate images to annotate their corresponding question-answer pairs to relevant categories. Once the image is selected, it will be removed from the candidates to ensure no repetition. Moreover, to ensure the correctness of our annotations, we have three additional annotators to review the question-answer pairs after the first annotation process. The annotation will not be reserved until the three annotators all agree that the first-person visual data and the assigned question-answer pairs meet the definition of a specific dimension." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.735, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Statistics. The EgoThink benchmark comprises a collection of 700 images across six categories with twelve fine-grained dimensions. These images are extracted from 595 videos, ensuring a broad representation of scenarios. To guarantee diversity, a wide range of scenes and concepts has been deliberately selected. As depicted in Figure 3, the dataset encompasses a diverse range of scenes, covering key scenarios relevant to EAI. Furthermore, we have meticulously crafted question and answer for each image in the EgoThink benchmark, aiming to closely replicate real-life conversations. This involves employing different" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.522, + 0.892, + 0.645 + ], + "angle": 0, + "content": "question types, varying questions in length and complexity, paired with well-reasoned and accurate answers. In addition, the size of the dataset is a well-balanced trade-off between benchmark diversity and high cost of open-ended QA evaluation. We ensure our 700 examples were diverse in concepts, scenes, and videos to provide a robust performance estimation within practical limits. Detailed statistics of the EgoThink benchmark are presented in Appendix A." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.657, + 0.634, + 0.673 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.681, + 0.698, + 0.699 + ], + "angle": 0, + "content": "4.1. Experimental Setups" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.704, + 0.892, + 0.825 + ], + "angle": 0, + "content": "Vision-Language Models. We collect the most popular eighteen types of representative VLMs to assess as shown in Table 2. Due to the possible effects of model parameters, we divide models into \\(\\sim 7\\mathrm{B}\\) and \\(\\sim 13\\mathrm{B}\\) for a fair comparison. Detailed information about VLMs can be found in Appendix B. We conduct zero-shot setups for all VLMs across our EgoThink benchmark. The prompts used for each VLM are shown in Appendix C." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.826, + 0.894, + 0.903 + ], + "angle": 0, + "content": "Single-answer grading. Considering evaluating open-ended model generations is not a trivial problem [54, 68], we propose to use GPT-4 [60] as the automatic evaluator [86] to better measure generated answers. In this protocol, we want to measure how close one model output" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "14295" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.078, + 0.088, + 0.47, + 0.389 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.395, + 0.471, + 0.452 + ], + "angle": 0, + "content": "Figure 4. Case study (wrong) of GPT-4V in the existence dimension of the object ability. In both the question formats of yes/no and what, GPT-4V can not correctly detect the chopsticks in my right hand, while InstructBLIP-11B and LLaVA-7B can." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.469, + 0.47, + 0.606 + ], + "angle": 0, + "content": "is to the reference. Different from traditional similarity-based methods, GPT-4 pays more attention to semantics. In the detailed implementation, we format the question, the model output, and the reference in a prompt as shown in Appendix D and feed it into the GPT-4 evaluator. The GPT-4 evaluator is asked to assign a score of 0 (wrong), 0.5 (partially correct), or 1 (correct) to the model output. Additionally, we further discuss to use of GPT-3.5-Turbo, Claude-2, and humans as evaluators in Section 5.2." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.618, + 0.172, + 0.634 + ], + "angle": 0, + "content": "4.2. Results" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.643, + 0.471, + 0.824 + ], + "angle": 0, + "content": "We first present the overall results of the evaluated models on our EgoThink benchmark as shown in Table 3. Despite having improved over the years, VLMs are still difficult to think from a first-person perspective, even GPT-4V. Among the six categories, only the scores on planning and localization are relatively high, the performance in other capabilities can only reach around 60 points at best. Among the better models, GPT-4V generally performs much better than other models, only falling short in localization dimension compared to BLIP-2-11B. We will further introduce the detailed scores across different dimensions. More case studies can be found in Appendix E." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.47, + 0.903 + ], + "angle": 0, + "content": "Results on object. In detail, we evaluate through three dimensions, including existence, attribute, and affordance. For existence, InstructBLIP-11B and ShareGPT4V-7B achieve the top-2 performance, indicating that they can predict the object precisely from the first-person perspec" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.088, + 0.891, + 0.456 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.462, + 0.893, + 0.533 + ], + "angle": 0, + "content": "Figure 5. Case studies (wrong) in the attribute, affordance, and activity dimensions. The top case demonstrates some VLMs locate the wrong place in the attribute and affordance dimension. The bottom case illustrates that in the activity dimension, some VLMs fail to detect the specific action." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.551, + 0.893, + 0.671 + ], + "angle": 0, + "content": "tive. As for GPT-4V, as illustrated in Figure 4, we observe that its performance in handed object detection leaves room for improvement. As for both attribute and affordance, the GPT-4V model has demonstrated superior performance, especially in the attribute dimension. In both dimensions, some open-source models as shown at the top of Figure 5 locate the wrong place or only answer the type of the object rather than its attribute or affordance." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.673, + 0.893, + 0.824 + ], + "angle": 0, + "content": "Results on activity. The performance of GPT-4V outperforms all open-source VLMs in the activity dimension. Among the \\(\\sim 7\\mathrm{B}\\) models, ShareGPT4V-7B and mPLUGowl-7B significantly outperform other VLMs and even achieve superior or comparable performance to \\(\\sim 13\\mathrm{B}\\) models. Overall, \\(\\sim 13\\mathrm{B}\\) models tend to perform better than \\(\\sim 7\\mathrm{B}\\) model in the activity dimension, but their scores are just below the passing line. The most possible reason is that detecting the specific action is difficult for VLMs as shown at the bottom of Figure 5." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.825, + 0.894, + 0.903 + ], + "angle": 0, + "content": "Results on localization. In general, BLIP-2-11B has shown obvious advantages among all VLMs, even surpassing GPT-4V in both location and spatial relationship dimensions. In the location dimension, BLIP-2-11B, GPT-4V, and InstructBLIP-11B demonstrate superior ability to achieve" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "14296" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.078, + 0.089, + 0.47, + 0.48 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.486, + 0.471, + 0.557 + ], + "angle": 0, + "content": "Figure 6. Case studies (wrong) in counting (top) and situated reasoning (bottom) dimensions. The top case shows VLMs can count but can not understand their relative position. The bottom case displays that the question requires commonsense knowledge and image understanding where only GPT-4V can answer it correctly." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.578, + 0.47, + 0.639 + ], + "angle": 0, + "content": "around 90 points. However, perceiving the spatial relationship of an object relative to oneself is much more difficult. This phenomenon can be also reflected in the top of Figure 5 where VLMs hard to distinguish left or right hand." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.642, + 0.47, + 0.808 + ], + "angle": 0, + "content": "Results on reasoning. Counting is the most difficult ability [80] among all evaluated dimensions. The best-performing model, GPT-4V, only scores 42.0, far away from satisfaction. Under the first-person perspective setup, VLMs need to not only count but also understand the relative position to oneself, as shown in the top case of Figure 6. Meanwhile, the comparison dimension also reflects the high difficulty, where the best score of 56.0 is obtained by LLaVA-1.5-13B. As for situated reasoning, GPT-4V demonstrates its strong commonsense reasoning ability to answer complex questions at the bottom of Figure 6." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.811, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Results on forecasting. Achieving high performance seems to be challenging as the best score achieved by GPT-4V is only 55.0. InsturctBLIP-11B achieves a relatively high score of 53.0 which is close to that of GPT-4V. We observe that the VLMs mainly suffer from two problems: recognizing objects incorrectly or forecasting too far as shown" + }, + { + "type": "image", + "bbox": [ + 0.501, + 0.089, + 0.891, + 0.539 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.543, + 0.894, + 0.6 + ], + "angle": 0, + "content": "Figure 7. Case studies (wrong) in the forecast (top) and navigation dimension (bottom). The top case shows VLMs might detect the glove as a cloth, while the bottom case indicates VLMs lack navigation details and overlook image information." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.614, + 0.648, + 0.628 + ], + "angle": 0, + "content": "in the top of Figure 7." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.629, + 0.892, + 0.765 + ], + "angle": 0, + "content": "Results on planning. In both navigation and assistance dimensions, the highest scores are achieved by GPT-4V with 60.0 and 84.0, respectively. LLaVA-13B-Llama2 behaves well in both dimensions with the second-best performance but its score is still 10 points lower than that of GPT-4V. The most possible reason is that answers provided by most open-source VLMs lack crucial details or overlook important information given in the images, as illustrated at the bottom of Figure 7." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.778, + 0.597, + 0.794 + ], + "angle": 0, + "content": "5. Analysis" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.802, + 0.712, + 0.818 + ], + "angle": 0, + "content": "5.1. Effects of Components" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.826, + 0.892, + 0.871 + ], + "angle": 0, + "content": "As shown in Table 2, VLMs consist of multiple key components. In this section, we probe the influence of different components on our EgoThink benchmark." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.892, + 0.9 + ], + "angle": 0, + "content": "The total parameters of LLMs. Here we compare the performance of \\(\\sim 7\\mathrm{B}\\) and \\(\\sim 13\\mathrm{B}\\) variants of four VLMs. Note" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "14297" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.08, + 0.088, + 0.473, + 0.271 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.277, + 0.471, + 0.319 + ], + "angle": 0, + "content": "Figure 8. Impact of LLMs sizes (above the dash-line) and instruction-tuning (below the dash-line) on model performance. Average scores across all capabilities are reported." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.336, + 0.47, + 0.502 + ], + "angle": 0, + "content": "that the increase in the number of parameters mainly falls in the LLMs. Firstly, as shown in the top part of Figure 8, scaling does not lead to significant improvement for PandaGPT and InstructBLIP, while LLaVA (LLaVA-7B and LLava13B-Llama2) and LLaVA-1.5 benefit a lot from scaling. We hypothesize that this is because LLaVA series models do not freeze their language models during instruction tuning, indicating that enlarging the number of trainable parameters can help improve both performance and generalization. In other words, one can see that simply scaling up language models without better alignment may not help." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.503, + 0.47, + 0.653 + ], + "angle": 0, + "content": "Instruction tuning. We directly compare the performance of BLIP-2-11B and InstructBLIP-11B, because these two models differ only in instruction tuning and additional instruction-aware tokens. As presented in the bottom part of Figure 8, InstructBLIP-11B outperforms BLIP-2-11B after instruction tuning, despite an unexpectedly small margin. This may be because much of the instruction tuning data employed by InstructBLIP is collected from specific downstream tasks, whose data distributions are very different from our first-person perspective data." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.654, + 0.47, + 0.79 + ], + "angle": 0, + "content": "The information of image encoder. Considering that there is no ablation version of VLMs for image encoder, following Set-of-Mark [79], we probe the effect of visual grounding information (i.e., a set of marks) in our setups. As presented in Figure 9, GPT-4V with additional segmentation information can correctly detect the mentioned location and objects, indicating that supplemented image information can be helpful in some situations. More discussion about quantitative experiments can be found in Appendix F." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.802, + 0.458, + 0.817 + ], + "angle": 0, + "content": "5.2. Agreements between Human and Evaluators" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.826, + 0.47, + 0.9 + ], + "angle": 0, + "content": "In this section, we further assess the model performance on object and planning dimensions using GPT-3.5-Turbo, Claude-2, and human annotators. Due to the heavy human labor, we ask three annotators to evaluate the performance of GPT-4V, which is the overall best model. Human anno" + }, + { + "type": "image", + "bbox": [ + 0.5, + 0.088, + 0.892, + 0.248 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.263, + 0.892, + 0.306 + ], + "angle": 0, + "content": "Figure 9. Case study (wrong) in the adding visual grounding information with images. The segmentation can help VLMs better locate the objects in question." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.32, + 0.893, + 0.547 + ], + "angle": 0, + "content": "tators consider the following aspects to evaluate: accuracy, completeness, logical soundness, and grammatical correctness. Our annotation system and detailed guidelines can be found in Appendix G. We further conduct GPT-3.5-Turbo and Claude-2 with the same evaluation prompt as GPT-4. The Pearson correlation coefficients between automatic evaluators (i.e., GPT-4, GPT-3.5-Turbo, Claude-2) and humans are 0.68, 0.43, and 0.68, respectively. The Cohen's Kappa coefficient among the three annotators is 0.81. This shows that evaluations made by GPT-4 and Claude-2 have a high correlation with humans. We hypothesize that recent well-performant LLMs can evaluate highly aligned with humans, given that most answers in our benchmark are relatively short and precise. Detailed scores of all evaluators and their correlations are discussed in Appendix H." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.559, + 0.62, + 0.574 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.584, + 0.893, + 0.901 + ], + "angle": 0, + "content": "To pave the way for the development of VLMs in the field of EAI and robotics, we introduce a comprehensive benchmark, EgoThink. Designed to evaluate the capacity of VLMs to \"think\" from a first-person perspective, EgoThink encompasses six core capabilities across twelve detailed dimensions. We assess eighteen popular VLMs and find that even the top-performing VLMs in most dimensions achieve only around a score of 60. GPT-4V achieves the best overall performance, but can not consistently surpass other open-source VLMs across all dimensions. In the analysis, we further probe the impact of various components on model performance and find that the total number of trainable parameters in LLMs has the most significant effect. Despite the human agreement with automatic evaluators being high, the evaluation of planning is difficult due to the detailed information in the answers. In future research, we aim to improve the evaluation method and further explore the essential capabilities of VLMs in the EAI and robotics fields. Acknowledgment: The work is supported by the National Key R&D Program of China (2022ZD0160502) and the National Natural Science Foundation of China (No.61925601)." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "14298" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.079, + 0.09, + 0.176, + 0.107 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.185 + ], + "angle": 0, + "content": "[1] Michael Ahn, Anthony Brohan, Noah Brown, Yevgen Chebotar, Omar Cortes, Byron David, Chelsea Finn, Keerthana Gopalakrishnan, Karol Hausman, Alex Herzog, et al. Do as i can, not as i say: Grounding language in robotic affordances. arXiv preprint arXiv:2204.01691, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.187, + 0.472, + 0.269 + ], + "angle": 0, + "content": "[2] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. Advances in Neural Information Processing Systems, 35:23716-23736, 2022. 1, 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.272, + 0.472, + 0.342 + ], + "angle": 0, + "content": "[3] Stanislaw Antol, Aishwarya Agrawal, Jiasen Lu, Margaret Mitchell, Dhruv Batra, C Lawrence Zitnick, and Devi Parikh. Vqa: Visual question answering. In Proceedings of the IEEE international conference on computer vision, pages 2425-2433, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.344, + 0.472, + 0.413 + ], + "angle": 0, + "content": "[4] Anas Awadalla, Irena Gao, Joshua Gardner, Jack Hessel, Yusuf Hanafy, Wanrong Zhu, Kalyani Marathe, Yonatan Bitton, Samir Gadre, Jenia Jitsev, Simon Kornblith, Pang Wei Koh, Gabriel Ilharco, Mitchell Wortsman, and Ludwig Schmidt. Openflamingo, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.415, + 0.472, + 0.497 + ], + "angle": 0, + "content": "[5] Anas Awadalla, Irena Gao, Josh Gardner, Jack Hessel, Yusuf Hanafy, Wanrong Zhu, Kalyani Marathe, Yonatan Bitton, Samir Gadre, Shiori Sagawa, et al. Openflamingo: An opensource framework for training large autoregressive vision-language models. arXiv preprint arXiv:2308.01390, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.5, + 0.472, + 0.57 + ], + "angle": 0, + "content": "[6] Sven Bambach, Stefan Lee, David J Crandall, and Chen Yu. Lending a hand: Detecting hands and recognizing activities in complex egocentric interactions. In Proceedings of the IEEE international conference on computer vision, pages 1949-1957, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.572, + 0.472, + 0.628 + ], + "angle": 0, + "content": "[7] William Berrios, Gautam Mittal, Tristan Thrush, Douwe Kiela, and Amanpreet Singh. Towards language models that can see: Computer vision through the lens of natural language. arXiv preprint arXiv:2306.16410, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.63, + 0.472, + 0.7 + ], + "angle": 0, + "content": "[8] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. Advances in neural information processing systems, 33:1877-1901, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.702, + 0.472, + 0.771 + ], + "angle": 0, + "content": "[9] Soravit Changpinyo, Piyush Sharma, Nan Ding, and Radu Soricut. Conceptual 12m: Pushing web-scale image-text pretraining to recognize long-tail visual concepts. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3558-3568, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.773, + 0.472, + 0.829 + ], + "angle": 0, + "content": "[10] Lin Chen, Jisong Li, Xiaoyi Dong, Pan Zhang, Conghui He, Jiaqi Wang, Feng Zhao, and Dahua Lin. Sharegpt4v: Improving large multi-modal models with better captions. arXiv preprint arXiv:2311.12793, 2023. 4, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.831, + 0.472, + 0.901 + ], + "angle": 0, + "content": "[11] Liang Chen, Yichi Zhang, Shuhuai Ren, Haozhe Zhao, Zefan Cai, Yuchi Wang, Peiyi Wang, Tianyu Liu, and Baobao Chang. Towards end-to-end embodied decision making via multi-modal large language model: Explorations with gpt4- vision and beyond, 2023. 2" + }, + { + "type": "list", + "bbox": [ + 0.079, + 0.115, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.893, + 0.162 + ], + "angle": 0, + "content": "[12] Wei-Lin Chiang, Zhuohan Li, Zi Lin, Ying Sheng, Zhang-hao Wu, Hao Zhang, Lianmin Zheng, Siyuan Zhuang, Yong-hao Zhuang, Joseph E. Gonzalez, Ion Stoica, and Eric P. Xing. Vicuna: An open-source chatbot impressing gpt-4 with \\(90\\%\\) * chatgpt quality, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.164, + 0.894, + 0.232 + ], + "angle": 0, + "content": "[13] Wenliang Dai, Junnan Li, Dongxu Li, Anthony Meng Huat Tiong, Junqi Zhao, Weisheng Wang, Boyang Li, Pascale Fung, and Steven Hoi. Instructlip: Towards general-purpose vision-language models with instruction tuning, 2023. 1, 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.234, + 0.893, + 0.303 + ], + "angle": 0, + "content": "[14] Tien Do, Ondrej Miksik, Joseph DeGol, Hyun Soo Park, and Sudipta N Sinha. Learning to detect scene landmarks for camera localization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11132-11142, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.305, + 0.893, + 0.374 + ], + "angle": 0, + "content": "[15] Danny Driess, Fei Xia, Mehdi SM Sajjadi, Corey Lynch, Aakanksha Chowdhery, Brian Ichter, Ayzaan Wahid, Jonathan Thompson, Quan Vuong, Tianhe Yu, et al. Palme: An embodied multimodal language model. arXiv preprint arXiv:2303.03378, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.375, + 0.893, + 0.432 + ], + "angle": 0, + "content": "[16] Chenyou Fan. Egovqa-an egocentric video question answering benchmark dataset. In Proceedings of the IEEE/CVF International Conference on Computer Vision Workshops, pages 0-0, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.434, + 0.893, + 0.488 + ], + "angle": 0, + "content": "[17] Ali Farhadi, Ian Endres, Derek Hoiem, and David Forsyth. Describing objects by their attributes. In 2009 IEEE conference on computer vision and pattern recognition, pages 1778-1785. IEEE, 2009. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.49, + 0.893, + 0.532 + ], + "angle": 0, + "content": "[18] Alireza Fathi, Ali Farhadi, and James M Rehg. Understanding egocentric activities. In 2011 international conference on computer vision, pages 407-414. IEEE, 2011. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.533, + 0.893, + 0.602 + ], + "angle": 0, + "content": "[19] Chaoyou Fu, Peixian Chen, Yunhang Shen, Yulei Qin, Mengdan Zhang, Xu Lin, Zhenyu Qiu, Wei Lin, Jinrui Yang, Xiawu Zheng, et al. Mme: A comprehensive evaluation benchmark for multimodal large language models. arXiv preprint arXiv:2306.13394, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.604, + 0.893, + 0.659 + ], + "angle": 0, + "content": "[20] Antonino Furnari, Sebastiano Battiato, Kristen Grauman, and Giovanni Maria Farinella. Next-active-object prediction from egocentric videos. Journal of Visual Communication and Image Representation, 49:401-411, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.661, + 0.893, + 0.716 + ], + "angle": 0, + "content": "[21] Jensen Gao, Bidipta Sarkar, Fei Xia, Ted Xiao, Jiajun Wu, Brian Ichter, Anirudha Majumdar, and Dorsa Sadigh. Physically grounded vision-language models for robotic manipulation. arXiv preprint arXiv:2309.02561, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.718, + 0.893, + 0.786 + ], + "angle": 0, + "content": "[22] Peng Gao, Jiaming Han, Renrui Zhang, Ziyi Lin, Shijie Geng, Aojun Zhou, Wei Zhang, Pan Lu, Conghui He, Xi-angyu Yue, et al. Llama-adapter v2: Parameter-efficient visual instruction model. arXiv preprint arXiv:2304.15010, 2023. 4, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.789, + 0.893, + 0.858 + ], + "angle": 0, + "content": "[23] Yash Goyal, Tejas Khot, Douglas Summers-Stay, Dhruv Bartra, and Devi Parikh. Making the v in vqa matter: Elevating the role of image understanding in visual question answering. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6904-6913, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.86, + 0.893, + 0.901 + ], + "angle": 0, + "content": "[24] Stéphane Grade, Mauro Pesenti, and Martin G Edwards. Evidence for the embodiment of space perception: concurrent hand but not arm action moderates reachability and egocen-" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "14299" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.092, + 0.469, + 0.12 + ], + "angle": 0, + "content": "tric distance perception. Frontiers in Psychology, 6:862, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.123, + 0.47, + 0.207 + ], + "angle": 0, + "content": "[25] Kristen Grauman, Andrew Westbury, Eugene Byrne, Zachary Chavis, Antonino Furnari, Rohit Girdhar, Jackson Hamburger, Hao Jiang, Miao Liu, Xingyu Liu, et al. Ego4d: Around the world in 3,000 hours of egocentric video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18995-19012, 2022. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.209, + 0.471, + 0.279 + ], + "angle": 0, + "content": "[26] Jean-Bernard Hayet, Frédéric Lerasle, and Michel Devy. Visual landmarks detection and recognition for mobile robot navigation. In 2003 IEEE Computer Society Conference on Computer Vision and Pattern Recognition, 2003. Proceedings., pages II-II. IEEE, 2003. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.282, + 0.471, + 0.35 + ], + "angle": 0, + "content": "[27] Evert Helms, Rolf Dieter Schraft, and M Hagele. rob@ work: Robot assistant in industrial environments. In Proceedings. 11th IEEE International Workshop on Robot and Human Interactive Communication, pages 399-404. IEEE, 2002. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.354, + 0.47, + 0.424 + ], + "angle": 0, + "content": "[28] Tucker Hermans, James M Rehg, and Aaron Bobick. Affordance prediction via learned object attributes. In IEEE international conference on robotics and automation (ICRA): Workshop on semantic perception, mapping, and exploration, pages 181-184. CiteSeer, 2011. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.426, + 0.469, + 0.481 + ], + "angle": 0, + "content": "[29] MD Zakir Hossain, Ferdous Sohel, Mohd Fairuz Shiratuddin, and Hamid Laga. A comprehensive survey of deep learning for image captioning. ACM Computing Surveys (CsUR), 51(6):1-36, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.484, + 0.469, + 0.554 + ], + "angle": 0, + "content": "[30] Wenlong Huang, Fei Xia, Ted Xiao, Harris Chan, Jacky Liang, Pete Florence, Andy Zeng, Jonathan Tompson, Igor Mordatch, Yevgen Chebotar, et al. Inner monologue: Embodied reasoning through planning with language models. arXiv preprint arXiv:2207.05608, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.557, + 0.469, + 0.613 + ], + "angle": 0, + "content": "[31] Wenlong Huang, Chen Wang, Ruohan Zhang, Yunzhu Li, Jiajun Wu, and Li Fei-Fei. Voxposer: Composable 3d value maps for robotic manipulation with language models. arXiv preprint arXiv:2307.05973, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.616, + 0.469, + 0.684 + ], + "angle": 0, + "content": "[32] Yichao Huang, Xiaorui Liu, Xin Zhang, and Lianwen Jin. A pointing gesture based egocentric interaction system: Dataset, approach and application. In Proceedings of the IEEE conference on computer vision and pattern recognition workshops, pages 16-23, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.687, + 0.469, + 0.757 + ], + "angle": 0, + "content": "[33] Drew A Hudson and Christopher D Manning. Gqa: A new dataset for real-world visual reasoning and compositional question answering. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6700-6709, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.759, + 0.469, + 0.814 + ], + "angle": 0, + "content": "[34] Baoxiong Jia, Ting Lei, Song-Chun Zhu, and Siyuan Huang. Egotaskqa: Understanding human tasks in egocentric videos. Advances in Neural Information Processing Systems, 35: 3343-3360, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.818, + 0.469, + 0.899 + ], + "angle": 0, + "content": "[35] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In International conference on machine learning, pages 4904-4916. PMLR, 2021. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.471, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.134 + ], + "angle": 0, + "content": "[36] Charmi Jobanputra, Jatna Bavishi, and Nishant Doshi. Human activity recognition: A survey. Procedia Computer Science, 155:698-703, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.136, + 0.892, + 0.177 + ], + "angle": 0, + "content": "[37] Nancy Kanwisher and Jon Driver. Objects, attributes, and visual attention: Which, what, and where. Current Directions in Psychological Science, 1(1):26-31, 1992. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.179, + 0.892, + 0.233 + ], + "angle": 0, + "content": "[38] Shian-Ru Ke, Hoang Le Uyen Thuc, Yong-Jin Lee, Jenq-Neng Hwang, Jang-Hee Yoo, and Kyoung-Ho Choi. A review on video-based human activity recognition. Computers, 2(2):88-131, 2013. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.235, + 0.892, + 0.304 + ], + "angle": 0, + "content": "[39] Roberta L Klatzky. Allocentric and egocentric spatial representations: Definitions, distinctions, and interconnections. In Spatial cognition: An interdisciplinary approach to representing and processing spatial knowledge, pages 1-17. Springer, 1998. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.305, + 0.892, + 0.36 + ], + "angle": 0, + "content": "[40] Weicheng Kuo, Yin Cui, Xiuye Gu, AJ Piergiovanni, and Anelia Angelova. F-vlm: Open-vocabulary object detection upon frozen vision and language models. arXiv preprint arXiv:2209.15639, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.362, + 0.892, + 0.417 + ], + "angle": 0, + "content": "[41] Bo Li, Yuanhan Zhang, Liangyu Chen, Jinghao Wang, Fanyi Pu, Jingkang Yang, Chunyuan Li, and Ziwei Liu. Mimic-it: Multi-modal in-context instruction tuning. arXiv preprint arXiv:2306.05425, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.419, + 0.892, + 0.473 + ], + "angle": 0, + "content": "[42] Bo Li, Yuanhan Zhang, Liangyu Chen, Jinghao Wang, Jingkang Yang, and Ziwei Liu. Otter: A multi-modal model with in-context instruction tuning. arXiv preprint arXiv:2305.03726, 2023. 1, 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.476, + 0.892, + 0.532 + ], + "angle": 0, + "content": "[43] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. arXiv preprint arXiv:2301.12597, 2023. 1, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.533, + 0.892, + 0.587 + ], + "angle": 0, + "content": "[44] KunChang Li, Yinan He, Yi Wang, Yizhuo Li, Wenhai Wang, Ping Luo, Yali Wang, Limin Wang, and Yu Qiao. Videochat: Chat-centric video understanding. arXiv preprint arXiv:2305.06355, 2023. 4, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.589, + 0.892, + 0.644 + ], + "angle": 0, + "content": "[45] Bin Lin, Bin Zhu, Yang Ye, Munan Ning, Peng Jin, and Li Yuan. Video-llava: Learning united visual representation by alignment before projection. arXiv preprint arXiv:2311.10122, 2023. 4, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.646, + 0.892, + 0.73 + ], + "angle": 0, + "content": "[46] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.732, + 0.892, + 0.773 + ], + "angle": 0, + "content": "[47] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. arXiv preprint arXiv:2310.03744, 2023. 4, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.774, + 0.892, + 0.802 + ], + "angle": 0, + "content": "[48] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning, 2023. 1, 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.803, + 0.892, + 0.843 + ], + "angle": 0, + "content": "[49] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. arXiv preprint arXiv:2304.08485, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.845, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[50] Li Liu, Wanli Ouyang, Xiaogang Wang, Paul Fieguth, Jie Chen, Xinwang Liu, and Matti Pietikainen. Deep learning for generic object detection: A survey. International journal of computer vision, 128:261-318, 2020. 3" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "14300" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.147 + ], + "angle": 0, + "content": "[51] Miao Liu, Siyu Tang, Yin Li, and James M Rehg. Forecasting human-object interaction: Joint prediction of motor attention and egocentric activity. Computer Vision-ECCV 2020, 12346:704-721, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.149, + 0.47, + 0.219 + ], + "angle": 0, + "content": "[52] Shaowei Liu, Subarna Tripathi, Somdeb Majumdar, and Xiaolong Wang. Joint hand motion and interaction hotspots prediction from egocentric videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3282-3292, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.22, + 0.47, + 0.302 + ], + "angle": 0, + "content": "[53] Yunze Liu, Yun Liu, Che Jiang, Kangbo Lyu, Weikang Wan, Hao Shen, Boqiang Liang, Zhoujie Fu, He Wang, and Li Yi. Hoi4d: A 4d egocentric dataset for category-level human-object interaction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21013-21022, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.304, + 0.47, + 0.373 + ], + "angle": 0, + "content": "[54] Yuan Liu, Haodong Duan, Yuanhan Zhang, Bo Li, Songyang Zhang, Wangbo Zhao, Yike Yuan, Jiaqi Wang, Conghui He, Ziwei Liu, et al. Mmbench: Is your multi-modal model an all-around player? arXiv preprint arXiv:2307.06281, 2023. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.375, + 0.469, + 0.417 + ], + "angle": 0, + "content": "[55] Michael Montemerlo and Sebastian Thrun. FastSLAM: A scalable method for the simultaneous localization and mapping problem in robotics. Springer, 2007. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.418, + 0.469, + 0.473 + ], + "angle": 0, + "content": "[56] Luis Montesano, Manuel Lopes, Alexandre Bernardino, and José Santos-Victor. Learning object affordances: from sensory-motor coordination to imitation. IEEE Transactions on Robotics, 24(1):15-26, 2008. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.475, + 0.469, + 0.53 + ], + "angle": 0, + "content": "[57] Francesca Morganti, Stefano Stefanini, and Giuseppe Riva. From allo-to egocentric spatial ability in early alzheimer's disease: a study with virtual reality spatial tasks. Cognitive neuroscience, 4(3-4):171-180, 2013. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.531, + 0.469, + 0.587 + ], + "angle": 0, + "content": "[58] Weimin Mou, Timothy P McNamara, Christine M Vali-quette, and Björn Rump. Allocentric and egocentric updating of spatial memories. Journal of experimental psychology: Learning, Memory, and Cognition, 30(1):142, 2004. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.589, + 0.469, + 0.643 + ], + "angle": 0, + "content": "[59] Thi-Hoa-Cuc Nguyen, Jean-Christophe Nebel, and Francisco Florez-Revuelta. Recognition of activities of daily living with egocentric vision: A review. Sensors, 16(1):72, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.645, + 0.398, + 0.659 + ], + "angle": 0, + "content": "[60] OpenAI. Gpt-4 technical report, 2023. 1, 2, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.661, + 0.469, + 0.743 + ], + "angle": 0, + "content": "[61] Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, et al. Training language models to follow instructions with human feedback. Advances in Neural Information Processing Systems, 35: 27730-27744, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.745, + 0.469, + 0.814 + ], + "angle": 0, + "content": "[62] Anish Pandey, Rakesh Kumar Sonkar, Krishna Kant Pandey, and DR Parhi. Path planning navigation of mobile robot with obstacles avoidance using fuzzy logic controller. In 2014 IEEE 8th international conference on intelligent systems and control (ISCO), pages 39-41. IEEE, 2014. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.816, + 0.469, + 0.857 + ], + "angle": 0, + "content": "[63] BK Patle, Anish Pandey, DRK Parhi, AJDT Jagadeesh, et al. A review: On path planning strategies for navigation of mobile robot. Defence Technology, 15(4):582-606, 2019. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.859, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[64] Bryan A Plummer, Liwei Wang, Chris M Cervantes, Juan C Caicedo, Julia Hockenmaier, and Svetlana Lazebnik. Flickr30k entities: Collecting region-to-phrase corre" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.892, + 0.134 + ], + "angle": 0, + "content": "spondences for richer image-to-sentence models. In Proceedings of the IEEE international conference on computer vision, pages 2641-2649, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.135, + 0.892, + 0.217 + ], + "angle": 0, + "content": "[65] Vignesh Ramanathan, Anmol Kalia, Vladan Petrovic, Yi Wen, Baixue Zheng, Baishan Guo, Rui Wang, Aaron Marquez, Rama Kovvuri, Abhishek Kadian, et al. Paco: Parts and attributes of common objects. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7141-7151, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.218, + 0.892, + 0.259 + ], + "angle": 0, + "content": "[66] Sajad Saeedi, Michael Trentini, Mae Seto, and Howard Li. Multiple-robot simultaneous localization and mapping: A review. Journal of Field Robotics, 33(1):3-46, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.26, + 0.892, + 0.343 + ], + "angle": 0, + "content": "[67] Dustin Schwenk, Apoorv Khandelwal, Christopher Clark, Kenneth Marino, and Roozbeh Mottaghi. A-okvqa: A benchmark for visual question answering using world knowledge. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23–27, 2022, Proceedings, Part VIII, pages 146–162. Springer, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.344, + 0.892, + 0.41 + ], + "angle": 0, + "content": "[68] Wenqi Shao, Yutao Hu, Peng Gao, Meng Lei, Kaipeng Zhang, Fanqing Meng, Peng Xu, Siyuan Huang, Hongsheng Li, Yu Qiao, et al. Tiny lvlm-ehub: Early multimodal experiments with bard. arXiv preprint arXiv:2308.03729, 2023. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.413, + 0.892, + 0.482 + ], + "angle": 0, + "content": "[69] Chan Hee Song, Jiaman Wu, Clayton Washington, Brian M Sadler, Wei-Lun Chao, and Yu Su. Llm-planner: Few-shot grounded planning for embodied agents with large language models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2998-3009, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.483, + 0.892, + 0.524 + ], + "angle": 0, + "content": "[70] Yixuan Su, Tian Lan, Huayang Li, Jialu Xu, Yan Wang, and Deng Cai. Pandagpt: One model to instruction-follow them all. arXiv preprint arXiv:2305.16355, 2023. 4, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.525, + 0.892, + 0.579 + ], + "angle": 0, + "content": "[71] Theodore Sumers, Kenneth Marino, Arun Ahuja, Rob Fergus, and Ishita Dasgupta. Distilling internet-scale vision-language models into embodied agents. arXiv preprint arXiv:2301.12507, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.581, + 0.892, + 0.663 + ], + "angle": 0, + "content": "[72] Loreto Susperregi, Izaskun Fernandez, Ane Fernandez, Santiago Fernandez, Iñaki Maurtua, and Irene Lopez de Vallejo. Interacting with a robot: a guide robot understanding natural language instructions. In International Conference on Ubiquitous Computing and Ambient Intelligence, pages 185-192. Springer, 2012. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.664, + 0.892, + 0.733 + ], + "angle": 0, + "content": "[73] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.734, + 0.892, + 0.774 + ], + "angle": 0, + "content": "[74] Michalis Vrigkas, Christophoros Nikou, and Ioannis A Kakadiaris. A review of human activity recognition methods. Frontiers in Robotics and AI, 2:28, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.775, + 0.892, + 0.83 + ], + "angle": 0, + "content": "[75] Guan Wang, Sijie Cheng, Xianyuan Zhan, Xiangang Li, Sen Song, and Yang Liu. Openchat: Advancing open-source language models with mixed-quality data. arXiv preprint arXiv:2309.11235, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.832, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[76] Weronika Wojtak, Flora Ferreira, Paulo Vicente, Luís Louro, Estela Bicho, and Wolfram Erlhagen. A neural integrator model for planning and value-based decision making of a robotics assistant. Neural Computing and Applications, 33: 3737-3756, 2021. 4" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "14301" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.161 + ], + "angle": 0, + "content": "[77] Peng Xu, Wenqi Shao, Kaipeng Zhang, Peng Gao, Shuo Liu, Meng Lei, Fanqing Meng, Siyuan Huang, Yu Qiao, and Ping Luo. Lvlm-ehub: A comprehensive evaluation benchmark for large vision-language models. arXiv preprint arXiv:2306.09265, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.162, + 0.47, + 0.232 + ], + "angle": 0, + "content": "[78] Jingkang Yang, Yuhao Dong, Shuai Liu, Bo Li, Ziyue Wang, Chencheng Jiang, Haoran Tan, Jiamu Kang, Yuanhan Zhang, Kaiyang Zhou, et al. Octopus: Embodied vision-language programmer from environmental feedback. arXiv preprint arXiv:2310.08588, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.232, + 0.47, + 0.287 + ], + "angle": 0, + "content": "[79] Jianwei Yang, Hao Zhang, Feng Li, Xueyan Zou, Chunyuan Li, and Jianfeng Gao. Set-of-mark prompting unleashes extraordinary visual grounding in gpt-4v. arXiv preprint arXiv:2310.11441, 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.288, + 0.47, + 0.343 + ], + "angle": 0, + "content": "[80] Zhengyuan Yang, Linjie Li, Kevin Lin, Jianfeng Wang, Chung-Ching Lin, Zicheng Liu, and Lijuan Wang. The dawn of Imms: Preliminary explorations with gpt-4v (isdiction). arXiv preprint arXiv:2309.17421, 2023. 1, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.344, + 0.47, + 0.427 + ], + "angle": 0, + "content": "[81] Qinghao Ye, Haiyang Xu, Guohai Xu, Jiabo Ye, Ming Yan, Yiyang Zhou, Junyang Wang, Anwen Hu, Pengcheng Shi, Yaya Shi, Chaoya Jiang, Chenliang Li, Yuanhong Xu, Hehong Chen, Junfeng Tian, Qian Qi, Ji Zhang, and Fei Huang. mplug-owl: Modularization empowers large language models with multimodality, 2023. 1, 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.427, + 0.47, + 0.483 + ], + "angle": 0, + "content": "[82] Quanzeng You, Hailin Jin, Zhaowen Wang, Chen Fang, and Jiebo Luo. Image captioning with semantic attention. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4651-4659, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.484, + 0.47, + 0.538 + ], + "angle": 0, + "content": "[83] Yifan Zhang, Congqi Cao, Jian Cheng, and Hanqing Lu. Egogesture: A new dataset and benchmark for egocentric hand gesture recognition. IEEE Transactions on Multimedia, 20(5):1038-1050, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.539, + 0.47, + 0.607 + ], + "angle": 0, + "content": "[84] Tiancheng Zhao, Tianqi Zhang, Mingwei Zhu, Haozhan Shen, Kyusong Lee, Xiaopeng Lu, and Jianwei Yin. Vl-checklist: Evaluating pre-trained vision-language models with objects, attributes and relations. arXiv preprint arXiv:2207.00221, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.608, + 0.47, + 0.663 + ], + "angle": 0, + "content": "[85] Zhong-Qiu Zhao, Peng Zheng, Shou-tao Xu, and Xindong Wu. Object detection with deep learning: A review. IEEE transactions on neural networks and learning systems, 30 (11):3212-3232, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.664, + 0.47, + 0.734 + ], + "angle": 0, + "content": "[86] Lianmin Zheng, Wei-Lin Chiang, Ying Sheng, Siyuan Zhuang, Zhanghao Wu, Yonghao Zhuang, Zi Lin, Zhuohan Li, Dacheng Li, Eric Xing, et al. Judging llm-as-a-judge with mt-bench and chatbot arena. arXiv preprint arXiv:2306.05685, 2023. 5, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.734, + 0.47, + 0.788 + ], + "angle": 0, + "content": "[87] Wangchunshu Zhou, Yan Zeng, Shizhe Diao, and Xinsong Zhang. Vlue: A multi-task benchmark for evaluating vision-language models. arXiv preprint arXiv:2205.15237, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.789, + 0.47, + 0.873 + ], + "angle": 0, + "content": "[88] Chenchen Zhu, Fanyi Xiao, Andres Alvarado, Yasmine Babaei, Jiabo Hu, Hichem El-Mohri, Sean Culatana, Roshan Sumbaly, and Zhicheng Yan. Egoobjects: A large-scale egocentric dataset for fine-grained object understanding. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 20110-20120, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.47, + 0.903 + ], + "angle": 0, + "content": "[89] Deyao Zhu, Jun Chen, Xiaogian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.903 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.092, + 0.892, + 0.121 + ], + "angle": 0, + "content": "understanding with advanced large language models. arXiv preprint arXiv:2304.10592, 2023. 1, 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.122, + 0.892, + 0.19 + ], + "angle": 0, + "content": "[90] Wanrong Zhu, Jack Hessel, Anas Awadalla, Samir Yitzhak Gadre, Jesse Dodge, Alex Fang, Youngjae Yu, Ludwig Schmidt, William Yang Wang, and Yejin Choi. Multimodal c4: An open, billion-scale corpus of images interleaved with text. arXiv preprint arXiv:2304.06939, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.192, + 0.892, + 0.234 + ], + "angle": 0, + "content": "[91] Zhengxia Zou, Keyan Chen, Zhenwei Shi, Yuhong Guo, and Jieping Ye. Object detection in 20 years: A survey. Proceedings of the IEEE, 2023. 3" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.092, + 0.892, + 0.234 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "14302" + } + ] +] \ No newline at end of file diff --git a/2024/EgoThink_ Evaluating First-Person Perspective Thinking Capability of Vision-Language Models/2ff0d147-b907-475e-8aa3-5584b71f3014_origin.pdf b/2024/EgoThink_ Evaluating First-Person Perspective Thinking Capability of Vision-Language Models/2ff0d147-b907-475e-8aa3-5584b71f3014_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..c4aeb2c012bfa6c4291ac4181f5d341fa17496e5 --- /dev/null +++ b/2024/EgoThink_ Evaluating First-Person Perspective Thinking Capability of Vision-Language Models/2ff0d147-b907-475e-8aa3-5584b71f3014_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a7ee847e9102e754f433edf09a4a8e879c830ced407e4b098cbd0c84610b362d +size 1674687 diff --git a/2024/EgoThink_ Evaluating First-Person Perspective Thinking Capability of Vision-Language Models/full.md b/2024/EgoThink_ Evaluating First-Person Perspective Thinking Capability of Vision-Language Models/full.md new file mode 100644 index 0000000000000000000000000000000000000000..977b704373b7f6ba3e881e412db4061f86b0518a --- /dev/null +++ b/2024/EgoThink_ Evaluating First-Person Perspective Thinking Capability of Vision-Language Models/full.md @@ -0,0 +1,279 @@ +# EgoThink: Evaluating First-Person Perspective Thinking Capability of Vision-Language Models + +Sijie Cheng $^{1,2,5 \ddagger}$ , Zhicheng Guo $^{1,2*}$ , Jingwen Wu $^{3*}$ , Kechen Fang $^{4}$ , Peng Li $^{2 \boxtimes}$ , Huaping Liu $^{1}$ , Yang Liu $^{1,2 \boxtimes}$ + +$^{1}$ Department of Computer Science and Technology, Tsinghua University $^{2}$ Institute for AI Industry Research (AIR), Tsinghua University $^{3}$ Department of Electrical and Computer Engineering, University of Toronto $^{4}$ Zhili College, Tsinghua University $^{5}$ 01.AI csj23@mails.tsinghua.edu.cn + +# Abstract + +Vision-language models (VLMs) have recently shown promising results in traditional downstream tasks. Evaluation studies have emerged to assess their abilities, with the majority focusing on the third-person perspective, and only a few addressing specific tasks from the first-person perspective. However, the capability of VLMs to "think" from a first-person perspective, a crucial attribute for advancing autonomous agents and robotics, remains largely unexplored. To bridge this research gap, we introduce EgoThink, a novel visual question-answering benchmark that encompasses six core capabilities with twelve detailed dimensions. The benchmark is constructed using selected clips from egocentric videos, with manually annotated question-answer pairs containing first-person information. To comprehensively assess VLMs, we evaluate twenty-one popular VLMs on EgoThink. Moreover, given the open-ended format of the answers, we use GPT-4 as the automatic judge to compute single-answer grading. Experimental results indicate that although GPT-4V leads in numerous dimensions, all evaluated VLMs still possess considerable potential for improvement in first-person perspective tasks. Meanwhile, enlarging the number of trainable parameters has the most significant impact on model performance on EgoThink. In conclusion, EgoThink serves as a valuable addition to existing evaluation benchmarks for VLMs, providing an indispensable resource for future research in the realm of embodied artificial intelligence and robotics. + +![](images/590199f49b69a6a83ae1ad2c7d310819f3dae15d279bc8054b176f41a5f5c11d.jpg) + +![](images/083095910b6751a0b160ced125d24313e8a4b0ad4320ce4081a9a23578260616.jpg) +Figure 1. The main categories of our EgoThink benchmark to comprehensively assess the capability of thinking from a first-person perspective. + +# EgoThink + +# 1. Introduction + +Benefiting from the rapid development of large language models (LLMs) [8, 60, 73], vision-language models (VLMs) [2, 15, 43, 80] have shown remarkable progress in both conventional vision-language downstream tasks [2, 15, 43, 80] and following diverse human instructions [13, 42, 48, 81, 89]. Their application has expanded into broader domains such as robotics [21, 31, 40] and embodied artificial intelligence (EAI) [71, 78]. As a result, the thorough evaluation of VLMs has become increasingly important and challenging. Observing and understanding the world from a first-person perspective is a natural approach for both humans and artificial intelligence agents. We propose that the ability to "think" from a first-person perspective, especially when interpreting egocentric images, is crucial for VLMs. + +However, as shown in Table 1, the ability to think from a first-person perspective is not adequately addressed by cur + +
BenchmarkCapabilityPerspectiveData SourceAnswer TypeEvaluatorSize
VL-CheckList [84]Object / Attribute / RelationThirdDatasetsPSAccuracy410k
LVLM-eHub [77]General Multi-ModalityThirdDatasetsMC / OEMetrics / LLMs / User332k
MME [19]General Multi-ModalityThirdHandcraftMCAccuracy2,194
Tiny LVLM-eHub [68]General Multi-ModalityThirdDatasetsOELLMs2,100
MMBench [54]General Multi-ModalityThirdDatasets / Handcraft / LLMsMCLLMs2,974
PCA-EVAL [11]Decision-MakingThirdHandcraftOEAccuracy / User300
EgoTaskQA [34]Spatial / Temporal / CausalFirstCrowdsourcingOECrowdsourcing40k
EgoVQA [16]Object / Action / PersonThird / FirstHandcraftMCAccuracy520
EgoThink (Ours)First-Person ThinkingFirstHandcraftOELLMs700
+ +Table 1. Comparison of recent comprehensive evaluation benchmarks of VLMs and our proposed benchmark EgoThink. Third and first indicate third-person and first-person perspectives. Datasets/Handcraft/LLMs denote existing datasets, manual annotation, and automatic generation by LLMs. PS/MC/OE indicate pairwise scoring, multi-choice, and open-ended question-answering, respectively. + +rent evaluation benchmarks for VLMs. On one hand, most of these benchmarks (six out of nine, as listed in Table 1) focus solely on the third-person perspective. On the other hand, those benchmarks that do consider the first-person perspective only encompass a limited range of capabilities. For instance, EgoTaskQA [34] examines spatial, temporal, and causal aspects, whereas EgoVQA [16] is limited to object, action, and person aspects. Therefore, there is a clear need to develop a comprehensive benchmark to evaluate the first-person capabilities of VLMs more effectively. + +In this work, we introduce a new benchmark for VLMs from a first-person perspective, named EgoThink. The initial step in developing this benchmark involves determining the necessary capabilities to assess. Humans, when interacting with the real world, consider a series of questions centered on themselves, ranging from "What is around me?", "What am I doing?", "Where am I?", "What about the situation around me?", "What will happen to me?" to "How will I do?". Drawing inspiration from this, we evaluate six core capabilities of VLMs, namely object, activity, localization, reasoning, forecasting, and planning. Each capability corresponds to one of the aforementioned questions, as illustrated in Figure 1. The next step is constructing the benchmark. We first categorize the six core capabilities into twelve detailed dimensions. We then select a minimum of 50 distinct and clear clips from egocentric videos for each dimension and manually annotate them with relevant first-person question-answer pairs. This approach ensures the quality and variety of the benchmark. The final step is evaluating VLM performance on this benchmark. Building on recent studies [7, 12, 68], we use GPT-4 [60] as an automatic evaluator. The Pearson correlation coefficient, when compared with human evaluation, shows a value of 0.68, indicating that the evaluation results are dependable. + +Based on our proposed EgoThink benchmark, we conduct comprehensive experiments to evaluate the first-person capabilities of twenty-one popular VLMs with varying model and data compositions. The findings indicate that GPT-4V stands out as the most effective model in various aspects. However, it shows less impressive results in spe + +cific capabilities such as activity and counting. Additionally, we observed that no single VLM consistently surpasses others in every aspect. For instance, GPT-4V is less effective than BLIP-2-11B for localization. Increasing the language model portion of the VLMs generally leads to better performance, but this improvement is not uniform across all models. Finally, our results highlight a significant potential for further enhancing the first-person capabilities of VLMs. + +# 2. Related Work + +Vision-Language Models. Inspired by the impressive success of LLMs [8, 61, 75], the recent popular VLMs tend to regard the powerful LLMs as the core backbone. At the beginning, VLMs usually use large-scale image-text pairwise datasets [9, 35, 46] or arbitrarily interleaved visual and textual data [2, 90] to pre-train. Furthermore, thanks to the availability of enormous image-text instruction datasets [41, 49], recent studies [13, 42, 48, 81, 89] further apply instruction tuning to help VLMs generate satisfactory answers. Benefiting from the two-stage training process, recent VLMs can achieve stunning performance on downstream vision-language tasks [3, 33, 46, 64]. + +Evaluations of VLMs. To evaluate the abilities of VLMs, there are diverse types of vision language downstream tasks. Conventional benchmarks, such as image caption tasks [29, 82] and visual question reasoning tasks [23, 67], mainly probe specific abilities of VLMs from the third-person perspective. Meanwhile, specialized analytical studies comprehensively evaluate the performance of VLMs from the third-person perspective, where Vblue [87] consists of five fundamental tasks and Lvmehub [77] evaluates six categories of capabilities on 47 standard vision-language benchmarks. As for the first-person perspective, there are some egocentric evaluation benchmarks in the computer vision field to assess some visual capabilities [32, 53, 83, 88]. In terms of multi-modality, there are a few benchmarks, such as EgoVQA[16] and EgoTaskQA [34], where mainly specific tasks without an overall understanding. In this paper, we mainly focus on exploring the comprehensive capa + +![](images/000713ea875894cd593e908d6fbae158eb919b691d82eb9ec6920158683cf253.jpg) + +![](images/f4e2c9ff8b5c15d045074ce18a376f0c84be14de1697d851ce51296e1a125c81.jpg) + +![](images/1ec81f93bd954bd3db813f746843b2f2cc2edd9bdcc9993fd93f54f145d82ec1.jpg) +Figure 2. Categories with fine-grained dimensions and their corresponding examples of EgoThink benchmark. + +![](images/a0d8999f7132f208ed4af268ab802c7dd76884612dba1181fe4d5b8681d14cc7.jpg) + +![](images/de1d8420fbed3468228b674dc89328b43fdbe4a715ccc6dc4269debb8affe0c6.jpg) + +![](images/6966aa81ea571d068b91902ba9375c3d662181e2df522d9107d712d207bfffb1.jpg) + +bilities of VLMs to think from a first-person perspective, as a supplement to previous evaluation benchmarks. + +# 3. EgoThink Benchmark + +In this section, we first elaborate on the core capabilities of thinking from a first-person perspective. Then, we introduce the process to manually construct our proposed benchmark EgoThink, which asks VLMs to generate open-ended answers according to first-person images and questions. + +# 3.1. Core Capabilities + +As shown in Figure 2, we specifically design six categories with twelve fine-grained dimensions from the first-person perspective for quantitative evaluation. + +- Object: What is around me? Recognizing objects in the real world is a preliminary ability of the human visual system [50, 85, 91]. Images from a first-person or egocentric perspective [53, 65, 88] pay more attention to the objects surrounding the subject or in hands. Moreover, we further divide the object category into three fine-grained dimensions: (1) Existence, predicting whether there is an object as described in the images; (2) Attribute [17, 37], detecting properties or characteristics (e.g., color) of an object; (3) Affordance [28, 56], predicting potential actions that a human can apply to an object. + +- Activity: What am I doing? Activity recognition is to automatically recognize specific human activities in video frames or still images [36, 38, 74]. From the egocentric perspective, we mainly focus on actions or activities based on object-hand interaction [6, 18, 59]. +- Localization: Where am I? In reality, localization is a critical capability for navigation and scene understanding in the real world [55, 66]. Here we investigate the localization capability from two aspects, Location and Spatial Relationship. Location indicates detecting the scene surrounding the subject [14, 26]. Spatial reasoning contains allocentric and egocentric perspectives [24, 39, 57, 58]. We focus on the egocentric perspective, i.e., the position of the object with respect to the subject. +- Reasoning: What about the situation around me? During the complex decision-making process, reasoning lies everywhere in our lives. Here we mainly focus on Counting, Comparison, and Situated Reasoning. Due to the first-person perspective, we generally count or compare objects in our hands or surrounding ourselves. As for situated reasoning, we employ cases that cannot be answered directly from the information in the images and require further reasoning processes. +- Forecasting: What will happen to me? Forecasting [20, 25, 51, 52] is a critical skill in the real world. From an egocentric view, forecasting always predicts the future of + +
VLMsImage EncoderLLMAlignment ModuleTTPToPDataset SizeEgoDataVideo
Image-TextInstruction
API-based Model
GPT-4V [60]Unknown
~7B Models
OpenFlamingo-7B [2, 5]CLIP-ViT-LMPT7BAttention1.4B8.1B2B-X✓
BLIP-2-6.7B [43]EVA-CLIP-ViT-gOPT6.7BQ-Former108M7.8B129M-XX
VideoChat-7B [44]BLIP2-VEVicuna7BQ-Former205M8B25M18KX✓
LLaVA-1.5-7B [47]CLIP-ViT-L-336pxLlama27BMLP6.8B7.1B558k665kXX
MiniGPT-4-7B [89]BLIP2-VELlama27BLinear23M7.7B5M3.5kXX
InstructBLIP-7B [13]EVA-CLIP-ViT-gVicuna7BQ-Former189M7.9B-16MXX
LLaMA-Adapter-7B [22]CLIP-ViT-LLLaMA7BEarly Fusion14M7.2B567k52kXX
Otter-I-7B [42]CLIP-ViT-LMPT7BAttention1.4B8.1B-2.8BX✓
PandaGPT-7B [70]ImageBindVicuna7BLinear + LLM LoRA38M7.9B-160k✓✓
mPLUG-owl-7B [81]CLIP-ViT-LLLaMA7BAttention4M7.1B204M158kXX
Video-LLaVA-7B [45]LanguageBindVicuna7BLinear6.8B7.5B1260k765kX✓
LLaVA-7B [48]CLIP-ViT-LLlama27BLinear6.7B7.1B595k158kXX
ShareGPT4V-7B [10]CLIP-ViT-L-336pxVicuna7BMLP6.7B6.7B1.2M665kXX
~13B Models
InstructBLIP-13B [13]EVA-CLIP-ViT-gVicuna13BQ-Former189M14.2B-16MXX
PandaGPT-13B [70]ImageBindVicuna13BLinear+LLM LoRA52M13.1B-160k✓✓
LLaVA-13B-Vicuna [48]CLIP-ViT-L-336pxVicuna13BLinear13.0B13.3B595k158kXX
BLIP-2-11B [43]EVA-CLIP-ViT-gFlanT5XXLQ-Former108M12.2B129M-XX
InstructBLIP-11B [13]EVA-CLIP-ViT-gFlanT5XXLQ-Former189M12.3B-16MXX
LLaVA-13B-Llama2 [48]CLIP-ViT-LLlama213BLinear13.0B13.3B595k158kXX
LLaVA-1.5-13B [47]CLIP-ViT-L-336pxLlama213BMLP13.0B13.4B558k665kXX
+ +Table 2. Statistics of compared API-based and open-source VLMs, where TTP and ToP indicate Total Trainable Parameters and Total Parameters, respectively. Moreover, EgoData and Video indicate that there are egocentric visual data and video data for training, respectively. + +object-state transformation or hand-object interactions. + +- Planning: How will I do? In reality, planning [1, 30, 69] is an important capability to deal with complex problems, typically applied in Navigation [62, 63, 72] and Assistance [27, 76]. Navigation is going to a goal location from a start position, while assistance is offering instructions to solve daily problems. + +# 3.2. Data Collection + +In this section, we mainly introduce the detailed processing to construct our EgoThink benchmark. + +Collecting first-person visual data. Firstly, we leverage a popular and large egocentric video dataset Ego4D [25], which is designed to advance the field of first-person perception in computer vision. To obtain a diverse representation in different scenarios, Ego4D encompasses 3,670 hours of video from 931 unique camera wearers spanning 74 global locations across 9 countries. To collect first-person visual data, we begin by extracting every frame from a subset of the Ego4D video dataset, yielding a diverse raw image dataset. Please note that our current focus is solely on images, as most VLMs today do not support video input. We intend to expand our scope to include videos in our future work. Considering the heavy human labor and the diversity of scenarios, we sample images every few dozen frames. + +![](images/c09b13c3d87b0f3d46b8e8ff00183e2229ba9b88e7a294737a86fdf647f89c75.jpg) +Figure 3. This chart illustrates the distribution of various scene categories within the EgoThink dataset. The 'others' category encompasses 13 different scene types, each representing less than one percent of total scenes. + +To ensure high quality, we apply strict criteria for selecting the extracted frames. We first exclude images that lack clarity or fail to exhibit egocentric characteristics. Then, to obtain the high diversity within the dataset, we conduct a further screening to ensure that at most two images per video are included in the filtered image set. Finally, we obtain enormous high-quality images with exhibit egocentric characteristics as first-person image candidates. + +Annotating question-answer pairs. Upon receiving a substantial collection of first-person image candidates, we + +
MethodsObjectActivityLocalizationReasoningForecastingPlanningAverage
ExistAttrAffordLocSpatialCountComparSituatedNavAssist
API-based model
GPT-4V62.082.058.059.586.062.042.048.083.055.064.084.065.5
~7B Models
OpenFlamingo-7B16.055.037.015.034.034.021.040.021.031.011.011.027.2
BLIP-2-6.7B49.029.039.033.560.031.03.021.033.025.08.06.028.1
VideoChat-7B46.044.036.045.061.042.036.039.032.026.513.021.036.8
LLaVA-1.5-7B33.047.054.035.535.049.020.047.037.027.029.054.039.0
MiniGPT-4-7B50.056.046.039.055.049.014.048.031.041.514.044.040.6
InstructBLIP-7B50.033.045.047.577.038.018.043.067.040.519.031.042.4
LLaMA-Adapter-7B37.060.046.034.548.051.029.039.025.041.542.057.042.5
Otter-I-7B48.056.039.044.060.044.039.048.042.038.031.055.045.3
PandaGPT-7B40.056.041.037.061.052.019.052.053.043.039.061.046.2
mPLUG-owl-7B56.058.047.053.060.053.025.049.044.049.533.058.048.8
Video-LLaVA-7B56.060.053.045.086.060.039.038.060.046.511.038.049.4
LLaVA-7B63.058.050.047.081.045.024.036.047.049.535.060.049.6
ShareGPT4V-7B67.075.053.055.577.062.030.038.066.047.041.063.051.9
~13B Models
InstructBLIP-13B52.055.049.054.063.049.011.033.059.044.019.025.042.8
PandaGPT-13B35.052.041.040.568.031.032.040.047.045.516.069.043.1
LLaVA-13B-Vicuna54.062.052.046.053.046.026.044.029.044.035.066.046.4
BLIP-2-11B52.062.041.049.590.066.025.050.070.048.018.024.049.6
InstructBLIP-11B74.068.048.049.586.052.032.049.073.053.016.017.051.5
LLaVA-13B-Llama265.061.045.056.077.053.034.034.066.050.549.071.055.1
LLaVA-1.5-13B66.055.051.055.082.057.032.056.067.048.539.055.055.3
+ +Table 3. Combined single-answer grading scores on zero-shot setups for various dimensions. The **bold** indicates the best performance while the **underline** indicates the second-best performance. Exist, Attr, Afford, Loc, Spatial, Count, Compar, Situated, Nav and Assist represent existence, attribute, affordance, location, spatial relationship, counting, comparison, situated reasoning, navigation, and assistance. + +engage six annotators to manually label question-answer pairs. Given that the EgoThink benchmark is composed of twelve dimensions, annotators were responsible for two specific dimensions. The annotators can access all the image candidates and are asked to select appropriate images to annotate their corresponding question-answer pairs to relevant categories. Once the image is selected, it will be removed from the candidates to ensure no repetition. Moreover, to ensure the correctness of our annotations, we have three additional annotators to review the question-answer pairs after the first annotation process. The annotation will not be reserved until the three annotators all agree that the first-person visual data and the assigned question-answer pairs meet the definition of a specific dimension. + +Statistics. The EgoThink benchmark comprises a collection of 700 images across six categories with twelve fine-grained dimensions. These images are extracted from 595 videos, ensuring a broad representation of scenarios. To guarantee diversity, a wide range of scenes and concepts has been deliberately selected. As depicted in Figure 3, the dataset encompasses a diverse range of scenes, covering key scenarios relevant to EAI. Furthermore, we have meticulously crafted question and answer for each image in the EgoThink benchmark, aiming to closely replicate real-life conversations. This involves employing different + +question types, varying questions in length and complexity, paired with well-reasoned and accurate answers. In addition, the size of the dataset is a well-balanced trade-off between benchmark diversity and high cost of open-ended QA evaluation. We ensure our 700 examples were diverse in concepts, scenes, and videos to provide a robust performance estimation within practical limits. Detailed statistics of the EgoThink benchmark are presented in Appendix A. + +# 4. Experiments + +# 4.1. Experimental Setups + +Vision-Language Models. We collect the most popular eighteen types of representative VLMs to assess as shown in Table 2. Due to the possible effects of model parameters, we divide models into $\sim 7\mathrm{B}$ and $\sim 13\mathrm{B}$ for a fair comparison. Detailed information about VLMs can be found in Appendix B. We conduct zero-shot setups for all VLMs across our EgoThink benchmark. The prompts used for each VLM are shown in Appendix C. + +Single-answer grading. Considering evaluating open-ended model generations is not a trivial problem [54, 68], we propose to use GPT-4 [60] as the automatic evaluator [86] to better measure generated answers. In this protocol, we want to measure how close one model output + +![](images/d44fe81b279c39d982f12209a2ca04758d000397b08cb8f750d16f0e4d381ccc.jpg) +Figure 4. Case study (wrong) of GPT-4V in the existence dimension of the object ability. In both the question formats of yes/no and what, GPT-4V can not correctly detect the chopsticks in my right hand, while InstructBLIP-11B and LLaVA-7B can. + +is to the reference. Different from traditional similarity-based methods, GPT-4 pays more attention to semantics. In the detailed implementation, we format the question, the model output, and the reference in a prompt as shown in Appendix D and feed it into the GPT-4 evaluator. The GPT-4 evaluator is asked to assign a score of 0 (wrong), 0.5 (partially correct), or 1 (correct) to the model output. Additionally, we further discuss to use of GPT-3.5-Turbo, Claude-2, and humans as evaluators in Section 5.2. + +# 4.2. Results + +We first present the overall results of the evaluated models on our EgoThink benchmark as shown in Table 3. Despite having improved over the years, VLMs are still difficult to think from a first-person perspective, even GPT-4V. Among the six categories, only the scores on planning and localization are relatively high, the performance in other capabilities can only reach around 60 points at best. Among the better models, GPT-4V generally performs much better than other models, only falling short in localization dimension compared to BLIP-2-11B. We will further introduce the detailed scores across different dimensions. More case studies can be found in Appendix E. + +Results on object. In detail, we evaluate through three dimensions, including existence, attribute, and affordance. For existence, InstructBLIP-11B and ShareGPT4V-7B achieve the top-2 performance, indicating that they can predict the object precisely from the first-person perspec + +![](images/4da87f053f0a04d7b2caff67fb5573d56f164be4d01d086d3f266a4994ba5c23.jpg) +Figure 5. Case studies (wrong) in the attribute, affordance, and activity dimensions. The top case demonstrates some VLMs locate the wrong place in the attribute and affordance dimension. The bottom case illustrates that in the activity dimension, some VLMs fail to detect the specific action. + +tive. As for GPT-4V, as illustrated in Figure 4, we observe that its performance in handed object detection leaves room for improvement. As for both attribute and affordance, the GPT-4V model has demonstrated superior performance, especially in the attribute dimension. In both dimensions, some open-source models as shown at the top of Figure 5 locate the wrong place or only answer the type of the object rather than its attribute or affordance. + +Results on activity. The performance of GPT-4V outperforms all open-source VLMs in the activity dimension. Among the $\sim 7\mathrm{B}$ models, ShareGPT4V-7B and mPLUGowl-7B significantly outperform other VLMs and even achieve superior or comparable performance to $\sim 13\mathrm{B}$ models. Overall, $\sim 13\mathrm{B}$ models tend to perform better than $\sim 7\mathrm{B}$ model in the activity dimension, but their scores are just below the passing line. The most possible reason is that detecting the specific action is difficult for VLMs as shown at the bottom of Figure 5. + +Results on localization. In general, BLIP-2-11B has shown obvious advantages among all VLMs, even surpassing GPT-4V in both location and spatial relationship dimensions. In the location dimension, BLIP-2-11B, GPT-4V, and InstructBLIP-11B demonstrate superior ability to achieve + +![](images/f1b05be3e4f0cca5aaea07d275409330047ef68ca19bbeb0b8590619fcd3cdc3.jpg) +Figure 6. Case studies (wrong) in counting (top) and situated reasoning (bottom) dimensions. The top case shows VLMs can count but can not understand their relative position. The bottom case displays that the question requires commonsense knowledge and image understanding where only GPT-4V can answer it correctly. + +around 90 points. However, perceiving the spatial relationship of an object relative to oneself is much more difficult. This phenomenon can be also reflected in the top of Figure 5 where VLMs hard to distinguish left or right hand. + +Results on reasoning. Counting is the most difficult ability [80] among all evaluated dimensions. The best-performing model, GPT-4V, only scores 42.0, far away from satisfaction. Under the first-person perspective setup, VLMs need to not only count but also understand the relative position to oneself, as shown in the top case of Figure 6. Meanwhile, the comparison dimension also reflects the high difficulty, where the best score of 56.0 is obtained by LLaVA-1.5-13B. As for situated reasoning, GPT-4V demonstrates its strong commonsense reasoning ability to answer complex questions at the bottom of Figure 6. + +Results on forecasting. Achieving high performance seems to be challenging as the best score achieved by GPT-4V is only 55.0. InsturctBLIP-11B achieves a relatively high score of 53.0 which is close to that of GPT-4V. We observe that the VLMs mainly suffer from two problems: recognizing objects incorrectly or forecasting too far as shown + +![](images/c5c3cecf1be63fbe682bd6df7053c614d6f0d9a7688f6686d59b58e73bfc3be0.jpg) +Figure 7. Case studies (wrong) in the forecast (top) and navigation dimension (bottom). The top case shows VLMs might detect the glove as a cloth, while the bottom case indicates VLMs lack navigation details and overlook image information. + +in the top of Figure 7. + +Results on planning. In both navigation and assistance dimensions, the highest scores are achieved by GPT-4V with 60.0 and 84.0, respectively. LLaVA-13B-Llama2 behaves well in both dimensions with the second-best performance but its score is still 10 points lower than that of GPT-4V. The most possible reason is that answers provided by most open-source VLMs lack crucial details or overlook important information given in the images, as illustrated at the bottom of Figure 7. + +# 5. Analysis + +# 5.1. Effects of Components + +As shown in Table 2, VLMs consist of multiple key components. In this section, we probe the influence of different components on our EgoThink benchmark. + +The total parameters of LLMs. Here we compare the performance of $\sim 7\mathrm{B}$ and $\sim 13\mathrm{B}$ variants of four VLMs. Note + +![](images/31240c3a0da56d36cf88d2c198cda83630df797e727752ce45d4e909d61048d3.jpg) +Figure 8. Impact of LLMs sizes (above the dash-line) and instruction-tuning (below the dash-line) on model performance. Average scores across all capabilities are reported. + +that the increase in the number of parameters mainly falls in the LLMs. Firstly, as shown in the top part of Figure 8, scaling does not lead to significant improvement for PandaGPT and InstructBLIP, while LLaVA (LLaVA-7B and LLava13B-Llama2) and LLaVA-1.5 benefit a lot from scaling. We hypothesize that this is because LLaVA series models do not freeze their language models during instruction tuning, indicating that enlarging the number of trainable parameters can help improve both performance and generalization. In other words, one can see that simply scaling up language models without better alignment may not help. + +Instruction tuning. We directly compare the performance of BLIP-2-11B and InstructBLIP-11B, because these two models differ only in instruction tuning and additional instruction-aware tokens. As presented in the bottom part of Figure 8, InstructBLIP-11B outperforms BLIP-2-11B after instruction tuning, despite an unexpectedly small margin. This may be because much of the instruction tuning data employed by InstructBLIP is collected from specific downstream tasks, whose data distributions are very different from our first-person perspective data. + +The information of image encoder. Considering that there is no ablation version of VLMs for image encoder, following Set-of-Mark [79], we probe the effect of visual grounding information (i.e., a set of marks) in our setups. As presented in Figure 9, GPT-4V with additional segmentation information can correctly detect the mentioned location and objects, indicating that supplemented image information can be helpful in some situations. More discussion about quantitative experiments can be found in Appendix F. + +# 5.2. Agreements between Human and Evaluators + +In this section, we further assess the model performance on object and planning dimensions using GPT-3.5-Turbo, Claude-2, and human annotators. Due to the heavy human labor, we ask three annotators to evaluate the performance of GPT-4V, which is the overall best model. Human anno + +![](images/009c8433541138dc8a35fec52d71473cc708944105bc5c953b6ab0430514af07.jpg) +Figure 9. Case study (wrong) in the adding visual grounding information with images. The segmentation can help VLMs better locate the objects in question. + +tators consider the following aspects to evaluate: accuracy, completeness, logical soundness, and grammatical correctness. Our annotation system and detailed guidelines can be found in Appendix G. We further conduct GPT-3.5-Turbo and Claude-2 with the same evaluation prompt as GPT-4. The Pearson correlation coefficients between automatic evaluators (i.e., GPT-4, GPT-3.5-Turbo, Claude-2) and humans are 0.68, 0.43, and 0.68, respectively. The Cohen's Kappa coefficient among the three annotators is 0.81. This shows that evaluations made by GPT-4 and Claude-2 have a high correlation with humans. We hypothesize that recent well-performant LLMs can evaluate highly aligned with humans, given that most answers in our benchmark are relatively short and precise. Detailed scores of all evaluators and their correlations are discussed in Appendix H. + +# 6. Conclusion + +To pave the way for the development of VLMs in the field of EAI and robotics, we introduce a comprehensive benchmark, EgoThink. Designed to evaluate the capacity of VLMs to "think" from a first-person perspective, EgoThink encompasses six core capabilities across twelve detailed dimensions. We assess eighteen popular VLMs and find that even the top-performing VLMs in most dimensions achieve only around a score of 60. GPT-4V achieves the best overall performance, but can not consistently surpass other open-source VLMs across all dimensions. In the analysis, we further probe the impact of various components on model performance and find that the total number of trainable parameters in LLMs has the most significant effect. Despite the human agreement with automatic evaluators being high, the evaluation of planning is difficult due to the detailed information in the answers. In future research, we aim to improve the evaluation method and further explore the essential capabilities of VLMs in the EAI and robotics fields. Acknowledgment: The work is supported by the National Key R&D Program of China (2022ZD0160502) and the National Natural Science Foundation of China (No.61925601). + +# References + +[1] Michael Ahn, Anthony Brohan, Noah Brown, Yevgen Chebotar, Omar Cortes, Byron David, Chelsea Finn, Keerthana Gopalakrishnan, Karol Hausman, Alex Herzog, et al. Do as i can, not as i say: Grounding language in robotic affordances. arXiv preprint arXiv:2204.01691, 2022. 4 +[2] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. Advances in Neural Information Processing Systems, 35:23716-23736, 2022. 1, 2, 4 +[3] Stanislaw Antol, Aishwarya Agrawal, Jiasen Lu, Margaret Mitchell, Dhruv Batra, C Lawrence Zitnick, and Devi Parikh. Vqa: Visual question answering. In Proceedings of the IEEE international conference on computer vision, pages 2425-2433, 2015. 2 +[4] Anas Awadalla, Irena Gao, Joshua Gardner, Jack Hessel, Yusuf Hanafy, Wanrong Zhu, Kalyani Marathe, Yonatan Bitton, Samir Gadre, Jenia Jitsev, Simon Kornblith, Pang Wei Koh, Gabriel Ilharco, Mitchell Wortsman, and Ludwig Schmidt. Openflamingo, 2023. 1 +[5] Anas Awadalla, Irena Gao, Josh Gardner, Jack Hessel, Yusuf Hanafy, Wanrong Zhu, Kalyani Marathe, Yonatan Bitton, Samir Gadre, Shiori Sagawa, et al. Openflamingo: An opensource framework for training large autoregressive vision-language models. arXiv preprint arXiv:2308.01390, 2023. 4 +[6] Sven Bambach, Stefan Lee, David J Crandall, and Chen Yu. Lending a hand: Detecting hands and recognizing activities in complex egocentric interactions. In Proceedings of the IEEE international conference on computer vision, pages 1949-1957, 2015. 3 +[7] William Berrios, Gautam Mittal, Tristan Thrush, Douwe Kiela, and Amanpreet Singh. Towards language models that can see: Computer vision through the lens of natural language. arXiv preprint arXiv:2306.16410, 2023. 2 +[8] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. Advances in neural information processing systems, 33:1877-1901, 2020. 1, 2 +[9] Soravit Changpinyo, Piyush Sharma, Nan Ding, and Radu Soricut. Conceptual 12m: Pushing web-scale image-text pretraining to recognize long-tail visual concepts. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3558-3568, 2021. 2 +[10] Lin Chen, Jisong Li, Xiaoyi Dong, Pan Zhang, Conghui He, Jiaqi Wang, Feng Zhao, and Dahua Lin. Sharegpt4v: Improving large multi-modal models with better captions. arXiv preprint arXiv:2311.12793, 2023. 4, 1 +[11] Liang Chen, Yichi Zhang, Shuhuai Ren, Haozhe Zhao, Zefan Cai, Yuchi Wang, Peiyi Wang, Tianyu Liu, and Baobao Chang. Towards end-to-end embodied decision making via multi-modal large language model: Explorations with gpt4- vision and beyond, 2023. 2 + +[12] Wei-Lin Chiang, Zhuohan Li, Zi Lin, Ying Sheng, Zhang-hao Wu, Hao Zhang, Lianmin Zheng, Siyuan Zhuang, Yong-hao Zhuang, Joseph E. Gonzalez, Ion Stoica, and Eric P. Xing. Vicuna: An open-source chatbot impressing gpt-4 with $90\%$ * chatgpt quality, 2023. 2 +[13] Wenliang Dai, Junnan Li, Dongxu Li, Anthony Meng Huat Tiong, Junqi Zhao, Weisheng Wang, Boyang Li, Pascale Fung, and Steven Hoi. Instructlip: Towards general-purpose vision-language models with instruction tuning, 2023. 1, 2, 4 +[14] Tien Do, Ondrej Miksik, Joseph DeGol, Hyun Soo Park, and Sudipta N Sinha. Learning to detect scene landmarks for camera localization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11132-11142, 2022. 3 +[15] Danny Driess, Fei Xia, Mehdi SM Sajjadi, Corey Lynch, Aakanksha Chowdhery, Brian Ichter, Ayzaan Wahid, Jonathan Thompson, Quan Vuong, Tianhe Yu, et al. Palme: An embodied multimodal language model. arXiv preprint arXiv:2303.03378, 2023. 1 +[16] Chenyou Fan. Egovqa-an egocentric video question answering benchmark dataset. In Proceedings of the IEEE/CVF International Conference on Computer Vision Workshops, pages 0-0, 2019. 2 +[17] Ali Farhadi, Ian Endres, Derek Hoiem, and David Forsyth. Describing objects by their attributes. In 2009 IEEE conference on computer vision and pattern recognition, pages 1778-1785. IEEE, 2009. 3 +[18] Alireza Fathi, Ali Farhadi, and James M Rehg. Understanding egocentric activities. In 2011 international conference on computer vision, pages 407-414. IEEE, 2011. 3 +[19] Chaoyou Fu, Peixian Chen, Yunhang Shen, Yulei Qin, Mengdan Zhang, Xu Lin, Zhenyu Qiu, Wei Lin, Jinrui Yang, Xiawu Zheng, et al. Mme: A comprehensive evaluation benchmark for multimodal large language models. arXiv preprint arXiv:2306.13394, 2023. 2 +[20] Antonino Furnari, Sebastiano Battiato, Kristen Grauman, and Giovanni Maria Farinella. Next-active-object prediction from egocentric videos. Journal of Visual Communication and Image Representation, 49:401-411, 2017. 3 +[21] Jensen Gao, Bidipta Sarkar, Fei Xia, Ted Xiao, Jiajun Wu, Brian Ichter, Anirudha Majumdar, and Dorsa Sadigh. Physically grounded vision-language models for robotic manipulation. arXiv preprint arXiv:2309.02561, 2023. 1 +[22] Peng Gao, Jiaming Han, Renrui Zhang, Ziyi Lin, Shijie Geng, Aojun Zhou, Wei Zhang, Pan Lu, Conghui He, Xi-angyu Yue, et al. Llama-adapter v2: Parameter-efficient visual instruction model. arXiv preprint arXiv:2304.15010, 2023. 4, 1 +[23] Yash Goyal, Tejas Khot, Douglas Summers-Stay, Dhruv Bartra, and Devi Parikh. Making the v in vqa matter: Elevating the role of image understanding in visual question answering. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6904-6913, 2017. 2 +[24] Stéphane Grade, Mauro Pesenti, and Martin G Edwards. Evidence for the embodiment of space perception: concurrent hand but not arm action moderates reachability and egocen- + +tric distance perception. Frontiers in Psychology, 6:862, 2015. 3 +[25] Kristen Grauman, Andrew Westbury, Eugene Byrne, Zachary Chavis, Antonino Furnari, Rohit Girdhar, Jackson Hamburger, Hao Jiang, Miao Liu, Xingyu Liu, et al. Ego4d: Around the world in 3,000 hours of egocentric video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18995-19012, 2022. 3, 4 +[26] Jean-Bernard Hayet, Frédéric Lerasle, and Michel Devy. Visual landmarks detection and recognition for mobile robot navigation. In 2003 IEEE Computer Society Conference on Computer Vision and Pattern Recognition, 2003. Proceedings., pages II-II. IEEE, 2003. 3 +[27] Evert Helms, Rolf Dieter Schraft, and M Hagele. rob@ work: Robot assistant in industrial environments. In Proceedings. 11th IEEE International Workshop on Robot and Human Interactive Communication, pages 399-404. IEEE, 2002. 4 +[28] Tucker Hermans, James M Rehg, and Aaron Bobick. Affordance prediction via learned object attributes. In IEEE international conference on robotics and automation (ICRA): Workshop on semantic perception, mapping, and exploration, pages 181-184. CiteSeer, 2011. 3 +[29] MD Zakir Hossain, Ferdous Sohel, Mohd Fairuz Shiratuddin, and Hamid Laga. A comprehensive survey of deep learning for image captioning. ACM Computing Surveys (CsUR), 51(6):1-36, 2019. 2 +[30] Wenlong Huang, Fei Xia, Ted Xiao, Harris Chan, Jacky Liang, Pete Florence, Andy Zeng, Jonathan Tompson, Igor Mordatch, Yevgen Chebotar, et al. Inner monologue: Embodied reasoning through planning with language models. arXiv preprint arXiv:2207.05608, 2022. 4 +[31] Wenlong Huang, Chen Wang, Ruohan Zhang, Yunzhu Li, Jiajun Wu, and Li Fei-Fei. Voxposer: Composable 3d value maps for robotic manipulation with language models. arXiv preprint arXiv:2307.05973, 2023. 1 +[32] Yichao Huang, Xiaorui Liu, Xin Zhang, and Lianwen Jin. A pointing gesture based egocentric interaction system: Dataset, approach and application. In Proceedings of the IEEE conference on computer vision and pattern recognition workshops, pages 16-23, 2016. 2 +[33] Drew A Hudson and Christopher D Manning. Gqa: A new dataset for real-world visual reasoning and compositional question answering. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6700-6709, 2019. 2 +[34] Baoxiong Jia, Ting Lei, Song-Chun Zhu, and Siyuan Huang. Egotaskqa: Understanding human tasks in egocentric videos. Advances in Neural Information Processing Systems, 35: 3343-3360, 2022. 2 +[35] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In International conference on machine learning, pages 4904-4916. PMLR, 2021. 2 + +[36] Charmi Jobanputra, Jatna Bavishi, and Nishant Doshi. Human activity recognition: A survey. Procedia Computer Science, 155:698-703, 2019. 3 +[37] Nancy Kanwisher and Jon Driver. Objects, attributes, and visual attention: Which, what, and where. Current Directions in Psychological Science, 1(1):26-31, 1992. 3 +[38] Shian-Ru Ke, Hoang Le Uyen Thuc, Yong-Jin Lee, Jenq-Neng Hwang, Jang-Hee Yoo, and Kyoung-Ho Choi. A review on video-based human activity recognition. Computers, 2(2):88-131, 2013. 3 +[39] Roberta L Klatzky. Allocentric and egocentric spatial representations: Definitions, distinctions, and interconnections. In Spatial cognition: An interdisciplinary approach to representing and processing spatial knowledge, pages 1-17. Springer, 1998. 3 +[40] Weicheng Kuo, Yin Cui, Xiuye Gu, AJ Piergiovanni, and Anelia Angelova. F-vlm: Open-vocabulary object detection upon frozen vision and language models. arXiv preprint arXiv:2209.15639, 2022. 1 +[41] Bo Li, Yuanhan Zhang, Liangyu Chen, Jinghao Wang, Fanyi Pu, Jingkang Yang, Chunyuan Li, and Ziwei Liu. Mimic-it: Multi-modal in-context instruction tuning. arXiv preprint arXiv:2306.05425, 2023. 2 +[42] Bo Li, Yuanhan Zhang, Liangyu Chen, Jinghao Wang, Jingkang Yang, and Ziwei Liu. Otter: A multi-modal model with in-context instruction tuning. arXiv preprint arXiv:2305.03726, 2023. 1, 2, 4 +[43] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. arXiv preprint arXiv:2301.12597, 2023. 1, 4 +[44] KunChang Li, Yinan He, Yi Wang, Yizhuo Li, Wenhai Wang, Ping Luo, Yali Wang, Limin Wang, and Yu Qiao. Videochat: Chat-centric video understanding. arXiv preprint arXiv:2305.06355, 2023. 4, 1 +[45] Bin Lin, Bin Zhu, Yang Ye, Munan Ning, Peng Jin, and Li Yuan. Video-llava: Learning united visual representation by alignment before projection. arXiv preprint arXiv:2311.10122, 2023. 4, 1 +[46] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 2 +[47] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. arXiv preprint arXiv:2310.03744, 2023. 4, 1 +[48] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning, 2023. 1, 2, 4 +[49] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. arXiv preprint arXiv:2304.08485, 2023. 2 +[50] Li Liu, Wanli Ouyang, Xiaogang Wang, Paul Fieguth, Jie Chen, Xinwang Liu, and Matti Pietikainen. Deep learning for generic object detection: A survey. International journal of computer vision, 128:261-318, 2020. 3 + +[51] Miao Liu, Siyu Tang, Yin Li, and James M Rehg. Forecasting human-object interaction: Joint prediction of motor attention and egocentric activity. Computer Vision-ECCV 2020, 12346:704-721, 2020. 3 +[52] Shaowei Liu, Subarna Tripathi, Somdeb Majumdar, and Xiaolong Wang. Joint hand motion and interaction hotspots prediction from egocentric videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3282-3292, 2022. 3 +[53] Yunze Liu, Yun Liu, Che Jiang, Kangbo Lyu, Weikang Wan, Hao Shen, Boqiang Liang, Zhoujie Fu, He Wang, and Li Yi. Hoi4d: A 4d egocentric dataset for category-level human-object interaction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21013-21022, 2022. 2, 3 +[54] Yuan Liu, Haodong Duan, Yuanhan Zhang, Bo Li, Songyang Zhang, Wangbo Zhao, Yike Yuan, Jiaqi Wang, Conghui He, Ziwei Liu, et al. Mmbench: Is your multi-modal model an all-around player? arXiv preprint arXiv:2307.06281, 2023. 2, 5 +[55] Michael Montemerlo and Sebastian Thrun. FastSLAM: A scalable method for the simultaneous localization and mapping problem in robotics. Springer, 2007. 3 +[56] Luis Montesano, Manuel Lopes, Alexandre Bernardino, and José Santos-Victor. Learning object affordances: from sensory-motor coordination to imitation. IEEE Transactions on Robotics, 24(1):15-26, 2008. 3 +[57] Francesca Morganti, Stefano Stefanini, and Giuseppe Riva. From allo-to egocentric spatial ability in early alzheimer's disease: a study with virtual reality spatial tasks. Cognitive neuroscience, 4(3-4):171-180, 2013. 3 +[58] Weimin Mou, Timothy P McNamara, Christine M Vali-quette, and Björn Rump. Allocentric and egocentric updating of spatial memories. Journal of experimental psychology: Learning, Memory, and Cognition, 30(1):142, 2004. 3 +[59] Thi-Hoa-Cuc Nguyen, Jean-Christophe Nebel, and Francisco Florez-Revuelta. Recognition of activities of daily living with egocentric vision: A review. Sensors, 16(1):72, 2016. 3 +[60] OpenAI. Gpt-4 technical report, 2023. 1, 2, 4, 5 +[61] Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, et al. Training language models to follow instructions with human feedback. Advances in Neural Information Processing Systems, 35: 27730-27744, 2022. 2 +[62] Anish Pandey, Rakesh Kumar Sonkar, Krishna Kant Pandey, and DR Parhi. Path planning navigation of mobile robot with obstacles avoidance using fuzzy logic controller. In 2014 IEEE 8th international conference on intelligent systems and control (ISCO), pages 39-41. IEEE, 2014. 4 +[63] BK Patle, Anish Pandey, DRK Parhi, AJDT Jagadeesh, et al. A review: On path planning strategies for navigation of mobile robot. Defence Technology, 15(4):582-606, 2019. 4 +[64] Bryan A Plummer, Liwei Wang, Chris M Cervantes, Juan C Caicedo, Julia Hockenmaier, and Svetlana Lazebnik. Flickr30k entities: Collecting region-to-phrase corre + +spondences for richer image-to-sentence models. In Proceedings of the IEEE international conference on computer vision, pages 2641-2649, 2015. 2 +[65] Vignesh Ramanathan, Anmol Kalia, Vladan Petrovic, Yi Wen, Baixue Zheng, Baishan Guo, Rui Wang, Aaron Marquez, Rama Kovvuri, Abhishek Kadian, et al. Paco: Parts and attributes of common objects. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7141-7151, 2023. 3 +[66] Sajad Saeedi, Michael Trentini, Mae Seto, and Howard Li. Multiple-robot simultaneous localization and mapping: A review. Journal of Field Robotics, 33(1):3-46, 2016. 3 +[67] Dustin Schwenk, Apoorv Khandelwal, Christopher Clark, Kenneth Marino, and Roozbeh Mottaghi. A-okvqa: A benchmark for visual question answering using world knowledge. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23–27, 2022, Proceedings, Part VIII, pages 146–162. Springer, 2022. 2 +[68] Wenqi Shao, Yutao Hu, Peng Gao, Meng Lei, Kaipeng Zhang, Fanqing Meng, Peng Xu, Siyuan Huang, Hongsheng Li, Yu Qiao, et al. Tiny lvlm-ehub: Early multimodal experiments with bard. arXiv preprint arXiv:2308.03729, 2023. 2, 5 +[69] Chan Hee Song, Jiaman Wu, Clayton Washington, Brian M Sadler, Wei-Lun Chao, and Yu Su. Llm-planner: Few-shot grounded planning for embodied agents with large language models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2998-3009, 2023. 4 +[70] Yixuan Su, Tian Lan, Huayang Li, Jialu Xu, Yan Wang, and Deng Cai. Pandagpt: One model to instruction-follow them all. arXiv preprint arXiv:2305.16355, 2023. 4, 1 +[71] Theodore Sumers, Kenneth Marino, Arun Ahuja, Rob Fergus, and Ishita Dasgupta. Distilling internet-scale vision-language models into embodied agents. arXiv preprint arXiv:2301.12507, 2023. 1 +[72] Loreto Susperregi, Izaskun Fernandez, Ane Fernandez, Santiago Fernandez, Iñaki Maurtua, and Irene Lopez de Vallejo. Interacting with a robot: a guide robot understanding natural language instructions. In International Conference on Ubiquitous Computing and Ambient Intelligence, pages 185-192. Springer, 2012. 4 +[73] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 1 +[74] Michalis Vrigkas, Christophoros Nikou, and Ioannis A Kakadiaris. A review of human activity recognition methods. Frontiers in Robotics and AI, 2:28, 2015. 3 +[75] Guan Wang, Sijie Cheng, Xianyuan Zhan, Xiangang Li, Sen Song, and Yang Liu. Openchat: Advancing open-source language models with mixed-quality data. arXiv preprint arXiv:2309.11235, 2023. 2 +[76] Weronika Wojtak, Flora Ferreira, Paulo Vicente, Luís Louro, Estela Bicho, and Wolfram Erlhagen. A neural integrator model for planning and value-based decision making of a robotics assistant. Neural Computing and Applications, 33: 3737-3756, 2021. 4 + +[77] Peng Xu, Wenqi Shao, Kaipeng Zhang, Peng Gao, Shuo Liu, Meng Lei, Fanqing Meng, Siyuan Huang, Yu Qiao, and Ping Luo. Lvlm-ehub: A comprehensive evaluation benchmark for large vision-language models. arXiv preprint arXiv:2306.09265, 2023. 2 +[78] Jingkang Yang, Yuhao Dong, Shuai Liu, Bo Li, Ziyue Wang, Chencheng Jiang, Haoran Tan, Jiamu Kang, Yuanhan Zhang, Kaiyang Zhou, et al. Octopus: Embodied vision-language programmer from environmental feedback. arXiv preprint arXiv:2310.08588, 2023. 1 +[79] Jianwei Yang, Hao Zhang, Feng Li, Xueyan Zou, Chunyuan Li, and Jianfeng Gao. Set-of-mark prompting unleashes extraordinary visual grounding in gpt-4v. arXiv preprint arXiv:2310.11441, 2023. 8 +[80] Zhengyuan Yang, Linjie Li, Kevin Lin, Jianfeng Wang, Chung-Ching Lin, Zicheng Liu, and Lijuan Wang. The dawn of Imms: Preliminary explorations with gpt-4v (isdiction). arXiv preprint arXiv:2309.17421, 2023. 1, 7 +[81] Qinghao Ye, Haiyang Xu, Guohai Xu, Jiabo Ye, Ming Yan, Yiyang Zhou, Junyang Wang, Anwen Hu, Pengcheng Shi, Yaya Shi, Chaoya Jiang, Chenliang Li, Yuanhong Xu, Hehong Chen, Junfeng Tian, Qian Qi, Ji Zhang, and Fei Huang. mplug-owl: Modularization empowers large language models with multimodality, 2023. 1, 2, 4 +[82] Quanzeng You, Hailin Jin, Zhaowen Wang, Chen Fang, and Jiebo Luo. Image captioning with semantic attention. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4651-4659, 2016. 2 +[83] Yifan Zhang, Congqi Cao, Jian Cheng, and Hanqing Lu. Egogesture: A new dataset and benchmark for egocentric hand gesture recognition. IEEE Transactions on Multimedia, 20(5):1038-1050, 2018. 2 +[84] Tiancheng Zhao, Tianqi Zhang, Mingwei Zhu, Haozhan Shen, Kyusong Lee, Xiaopeng Lu, and Jianwei Yin. Vl-checklist: Evaluating pre-trained vision-language models with objects, attributes and relations. arXiv preprint arXiv:2207.00221, 2022. 2 +[85] Zhong-Qiu Zhao, Peng Zheng, Shou-tao Xu, and Xindong Wu. Object detection with deep learning: A review. IEEE transactions on neural networks and learning systems, 30 (11):3212-3232, 2019. 3 +[86] Lianmin Zheng, Wei-Lin Chiang, Ying Sheng, Siyuan Zhuang, Zhanghao Wu, Yonghao Zhuang, Zi Lin, Zhuohan Li, Dacheng Li, Eric Xing, et al. Judging llm-as-a-judge with mt-bench and chatbot arena. arXiv preprint arXiv:2306.05685, 2023. 5, 2 +[87] Wangchunshu Zhou, Yan Zeng, Shizhe Diao, and Xinsong Zhang. Vlue: A multi-task benchmark for evaluating vision-language models. arXiv preprint arXiv:2205.15237, 2022. 2 +[88] Chenchen Zhu, Fanyi Xiao, Andres Alvarado, Yasmine Babaei, Jiabo Hu, Hichem El-Mohri, Sean Culatana, Roshan Sumbaly, and Zhicheng Yan. Egoobjects: A large-scale egocentric dataset for fine-grained object understanding. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 20110-20120, 2023. 2, 3 +[89] Deyao Zhu, Jun Chen, Xiaogian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language + +understanding with advanced large language models. arXiv preprint arXiv:2304.10592, 2023. 1, 2, 4 +[90] Wanrong Zhu, Jack Hessel, Anas Awadalla, Samir Yitzhak Gadre, Jesse Dodge, Alex Fang, Youngjae Yu, Ludwig Schmidt, William Yang Wang, and Yejin Choi. Multimodal c4: An open, billion-scale corpus of images interleaved with text. arXiv preprint arXiv:2304.06939, 2023. 2 +[91] Zhengxia Zou, Keyan Chen, Zhenwei Shi, Yuhong Guo, and Jieping Ye. Object detection in 20 years: A survey. Proceedings of the IEEE, 2023. 3 \ No newline at end of file diff --git a/2024/EgoThink_ Evaluating First-Person Perspective Thinking Capability of Vision-Language Models/images.zip b/2024/EgoThink_ Evaluating First-Person Perspective Thinking Capability of Vision-Language Models/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..1fdcc1dc9833a1cc3a431dd7cfbe72954583e191 --- /dev/null +++ b/2024/EgoThink_ Evaluating First-Person Perspective Thinking Capability of Vision-Language Models/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:42d18f48a4b58819116da90091af8fb6a3a8181fd08033bb07f6ca8882d3b52f +size 1016363 diff --git a/2024/EgoThink_ Evaluating First-Person Perspective Thinking Capability of Vision-Language Models/layout.json b/2024/EgoThink_ Evaluating First-Person Perspective Thinking Capability of Vision-Language Models/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..25c33e080054e6764e8fc037d5dfa9159838110a --- /dev/null +++ b/2024/EgoThink_ Evaluating First-Person Perspective Thinking Capability of Vision-Language Models/layout.json @@ -0,0 +1,7714 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 78, + 102, + 516, + 139 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 78, + 102, + 516, + 139 + ], + "spans": [ + { + "bbox": [ + 78, + 102, + 516, + 139 + ], + "type": "text", + "content": "EgoThink: Evaluating First-Person Perspective Thinking Capability of Vision-Language Models" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 159, + 460, + 189 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 159, + 460, + 189 + ], + "spans": [ + { + "bbox": [ + 132, + 159, + 460, + 189 + ], + "type": "text", + "content": "Sijie Cheng" + }, + { + "bbox": [ + 132, + 159, + 460, + 189 + ], + "type": "inline_equation", + "content": "^{1,2,5 \\ddagger}" + }, + { + "bbox": [ + 132, + 159, + 460, + 189 + ], + "type": "text", + "content": ", Zhicheng Guo" + }, + { + "bbox": [ + 132, + 159, + 460, + 189 + ], + "type": "inline_equation", + "content": "^{1,2*}" + }, + { + "bbox": [ + 132, + 159, + 460, + 189 + ], + "type": "text", + "content": ", Jingwen Wu" + }, + { + "bbox": [ + 132, + 159, + 460, + 189 + ], + "type": "inline_equation", + "content": "^{3*}" + }, + { + "bbox": [ + 132, + 159, + 460, + 189 + ], + "type": "text", + "content": ", Kechen Fang" + }, + { + "bbox": [ + 132, + 159, + 460, + 189 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 132, + 159, + 460, + 189 + ], + "type": "text", + "content": ", Peng Li" + }, + { + "bbox": [ + 132, + 159, + 460, + 189 + ], + "type": "inline_equation", + "content": "^{2 \\boxtimes}" + }, + { + "bbox": [ + 132, + 159, + 460, + 189 + ], + "type": "text", + "content": ", Huaping Liu" + }, + { + "bbox": [ + 132, + 159, + 460, + 189 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 132, + 159, + 460, + 189 + ], + "type": "text", + "content": ", Yang Liu" + }, + { + "bbox": [ + 132, + 159, + 460, + 189 + ], + "type": "inline_equation", + "content": "^{1,2 \\boxtimes}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 113, + 190, + 481, + 258 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 113, + 190, + 481, + 258 + ], + "spans": [ + { + "bbox": [ + 113, + 190, + 481, + 258 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 113, + 190, + 481, + 258 + ], + "type": "text", + "content": "Department of Computer Science and Technology, Tsinghua University " + }, + { + "bbox": [ + 113, + 190, + 481, + 258 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 113, + 190, + 481, + 258 + ], + "type": "text", + "content": "Institute for AI Industry Research (AIR), Tsinghua University " + }, + { + "bbox": [ + 113, + 190, + 481, + 258 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 113, + 190, + 481, + 258 + ], + "type": "text", + "content": "Department of Electrical and Computer Engineering, University of Toronto " + }, + { + "bbox": [ + 113, + 190, + 481, + 258 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 113, + 190, + 481, + 258 + ], + "type": "text", + "content": "Zhili College, Tsinghua University " + }, + { + "bbox": [ + 113, + 190, + 481, + 258 + ], + "type": "inline_equation", + "content": "^{5}" + }, + { + "bbox": [ + 113, + 190, + 481, + 258 + ], + "type": "text", + "content": "01.AI csj23@mails.tsinghua.edu.cn" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 143, + 286, + 192, + 300 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 286, + 192, + 300 + ], + "spans": [ + { + "bbox": [ + 143, + 286, + 192, + 300 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 316, + 290, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 316, + 290, + 639 + ], + "spans": [ + { + "bbox": [ + 47, + 316, + 290, + 639 + ], + "type": "text", + "content": "Vision-language models (VLMs) have recently shown promising results in traditional downstream tasks. Evaluation studies have emerged to assess their abilities, with the majority focusing on the third-person perspective, and only a few addressing specific tasks from the first-person perspective. However, the capability of VLMs to \"think\" from a first-person perspective, a crucial attribute for advancing autonomous agents and robotics, remains largely unexplored. To bridge this research gap, we introduce EgoThink, a novel visual question-answering benchmark that encompasses six core capabilities with twelve detailed dimensions. The benchmark is constructed using selected clips from egocentric videos, with manually annotated question-answer pairs containing first-person information. To comprehensively assess VLMs, we evaluate twenty-one popular VLMs on EgoThink. Moreover, given the open-ended format of the answers, we use GPT-4 as the automatic judge to compute single-answer grading. Experimental results indicate that although GPT-4V leads in numerous dimensions, all evaluated VLMs still possess considerable potential for improvement in first-person perspective tasks. Meanwhile, enlarging the number of trainable parameters has the most significant impact on model performance on EgoThink. In conclusion, EgoThink serves as a valuable addition to existing evaluation benchmarks for VLMs, providing an indispensable resource for future research in the realm of embodied artificial intelligence and robotics." + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 310, + 289, + 545, + 415 + ], + "blocks": [ + { + "bbox": [ + 310, + 289, + 545, + 415 + ], + "lines": [ + { + "bbox": [ + 310, + 289, + 545, + 415 + ], + "spans": [ + { + "bbox": [ + 310, + 289, + 545, + 415 + ], + "type": "image", + "image_path": "590199f49b69a6a83ae1ad2c7d310819f3dae15d279bc8054b176f41a5f5c11d.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 403, + 421, + 420, + 438 + ], + "blocks": [ + { + "bbox": [ + 403, + 421, + 420, + 438 + ], + "lines": [ + { + "bbox": [ + 403, + 421, + 420, + 438 + ], + "spans": [ + { + "bbox": [ + 403, + 421, + 420, + 438 + ], + "type": "image", + "image_path": "083095910b6751a0b160ced125d24313e8a4b0ad4320ce4081a9a23578260616.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 442, + 547, + 477 + ], + "lines": [ + { + "bbox": [ + 305, + 442, + 547, + 477 + ], + "spans": [ + { + "bbox": [ + 305, + 442, + 547, + 477 + ], + "type": "text", + "content": "Figure 1. The main categories of our EgoThink benchmark to comprehensively assess the capability of thinking from a first-person perspective." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 421, + 426, + 465, + 439 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 421, + 426, + 465, + 439 + ], + "spans": [ + { + "bbox": [ + 421, + 426, + 465, + 439 + ], + "type": "text", + "content": "EgoThink" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 498, + 387, + 511 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 498, + 387, + 511 + ], + "spans": [ + { + "bbox": [ + 306, + 498, + 387, + 511 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 520, + 545, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 520, + 545, + 688 + ], + "spans": [ + { + "bbox": [ + 304, + 520, + 545, + 688 + ], + "type": "text", + "content": "Benefiting from the rapid development of large language models (LLMs) [8, 60, 73], vision-language models (VLMs) [2, 15, 43, 80] have shown remarkable progress in both conventional vision-language downstream tasks [2, 15, 43, 80] and following diverse human instructions [13, 42, 48, 81, 89]. Their application has expanded into broader domains such as robotics [21, 31, 40] and embodied artificial intelligence (EAI) [71, 78]. As a result, the thorough evaluation of VLMs has become increasingly important and challenging. Observing and understanding the world from a first-person perspective is a natural approach for both humans and artificial intelligence agents. We propose that the ability to \"think\" from a first-person perspective, especially when interpreting egocentric images, is crucial for VLMs." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 689, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 545, + 714 + ], + "type": "text", + "content": "However, as shown in Table 1, the ability to think from a first-person perspective is not adequately addressed by cur" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 673, + 253, + 685 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 673, + 253, + 685 + ], + "spans": [ + { + "bbox": [ + 58, + 673, + 253, + 685 + ], + "type": "text", + "content": "*Equal contribution, ‡ Project lead, Corresponding author" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 63, + 685, + 228, + 693 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 685, + 228, + 693 + ], + "spans": [ + { + "bbox": [ + 63, + 685, + 228, + 693 + ], + "type": "text", + "content": "Project page: https://adacheng.github.io/EgoThink/" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 63, + 694, + 241, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 694, + 241, + 703 + ], + "spans": [ + { + "bbox": [ + 63, + 694, + 241, + 703 + ], + "type": "text", + "content": "GitHub page: https://github.com/AdaCheng/EgoThink/" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 63, + 704, + 279, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 704, + 279, + 712 + ], + "spans": [ + { + "bbox": [ + 63, + 704, + 279, + 712 + ], + "type": "text", + "content": "Dataset page: https://huggingface.co/datasets/EgoThink/EgoThink/" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14291" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 50, + 70, + 547, + 178 + ], + "blocks": [ + { + "bbox": [ + 50, + 70, + 547, + 178 + ], + "lines": [ + { + "bbox": [ + 50, + 70, + 547, + 178 + ], + "spans": [ + { + "bbox": [ + 50, + 70, + 547, + 178 + ], + "type": "table", + "html": "
BenchmarkCapabilityPerspectiveData SourceAnswer TypeEvaluatorSize
VL-CheckList [84]Object / Attribute / RelationThirdDatasetsPSAccuracy410k
LVLM-eHub [77]General Multi-ModalityThirdDatasetsMC / OEMetrics / LLMs / User332k
MME [19]General Multi-ModalityThirdHandcraftMCAccuracy2,194
Tiny LVLM-eHub [68]General Multi-ModalityThirdDatasetsOELLMs2,100
MMBench [54]General Multi-ModalityThirdDatasets / Handcraft / LLMsMCLLMs2,974
PCA-EVAL [11]Decision-MakingThirdHandcraftOEAccuracy / User300
EgoTaskQA [34]Spatial / Temporal / CausalFirstCrowdsourcingOECrowdsourcing40k
EgoVQA [16]Object / Action / PersonThird / FirstHandcraftMCAccuracy520
EgoThink (Ours)First-Person ThinkingFirstHandcraftOELLMs700
", + "image_path": "2149d4d569ab524c47420b74a71e41afd9e015987c6cc9f20342e807df195494.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 185, + 546, + 220 + ], + "lines": [ + { + "bbox": [ + 46, + 185, + 546, + 220 + ], + "spans": [ + { + "bbox": [ + 46, + 185, + 546, + 220 + ], + "type": "text", + "content": "Table 1. Comparison of recent comprehensive evaluation benchmarks of VLMs and our proposed benchmark EgoThink. Third and first indicate third-person and first-person perspectives. Datasets/Handcraft/LLMs denote existing datasets, manual annotation, and automatic generation by LLMs. PS/MC/OE indicate pairwise scoring, multi-choice, and open-ended question-answering, respectively." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 45, + 231, + 287, + 351 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 231, + 287, + 351 + ], + "spans": [ + { + "bbox": [ + 45, + 231, + 287, + 351 + ], + "type": "text", + "content": "rent evaluation benchmarks for VLMs. On one hand, most of these benchmarks (six out of nine, as listed in Table 1) focus solely on the third-person perspective. On the other hand, those benchmarks that do consider the first-person perspective only encompass a limited range of capabilities. For instance, EgoTaskQA [34] examines spatial, temporal, and causal aspects, whereas EgoVQA [16] is limited to object, action, and person aspects. Therefore, there is a clear need to develop a comprehensive benchmark to evaluate the first-person capabilities of VLMs more effectively." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 45, + 353, + 288, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 353, + 288, + 640 + ], + "spans": [ + { + "bbox": [ + 45, + 353, + 288, + 640 + ], + "type": "text", + "content": "In this work, we introduce a new benchmark for VLMs from a first-person perspective, named EgoThink. The initial step in developing this benchmark involves determining the necessary capabilities to assess. Humans, when interacting with the real world, consider a series of questions centered on themselves, ranging from \"What is around me?\", \"What am I doing?\", \"Where am I?\", \"What about the situation around me?\", \"What will happen to me?\" to \"How will I do?\". Drawing inspiration from this, we evaluate six core capabilities of VLMs, namely object, activity, localization, reasoning, forecasting, and planning. Each capability corresponds to one of the aforementioned questions, as illustrated in Figure 1. The next step is constructing the benchmark. We first categorize the six core capabilities into twelve detailed dimensions. We then select a minimum of 50 distinct and clear clips from egocentric videos for each dimension and manually annotate them with relevant first-person question-answer pairs. This approach ensures the quality and variety of the benchmark. The final step is evaluating VLM performance on this benchmark. Building on recent studies [7, 12, 68], we use GPT-4 [60] as an automatic evaluator. The Pearson correlation coefficient, when compared with human evaluation, shows a value of 0.68, indicating that the evaluation results are dependable." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 642, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 642, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 642, + 288, + 715 + ], + "type": "text", + "content": "Based on our proposed EgoThink benchmark, we conduct comprehensive experiments to evaluate the first-person capabilities of twenty-one popular VLMs with varying model and data compositions. The findings indicate that GPT-4V stands out as the most effective model in various aspects. However, it shows less impressive results in spe" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 231, + 547, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 231, + 547, + 327 + ], + "spans": [ + { + "bbox": [ + 304, + 231, + 547, + 327 + ], + "type": "text", + "content": "cific capabilities such as activity and counting. Additionally, we observed that no single VLM consistently surpasses others in every aspect. For instance, GPT-4V is less effective than BLIP-2-11B for localization. Increasing the language model portion of the VLMs generally leads to better performance, but this improvement is not uniform across all models. Finally, our results highlight a significant potential for further enhancing the first-person capabilities of VLMs." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 342, + 392, + 355 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 342, + 392, + 355 + ], + "spans": [ + { + "bbox": [ + 306, + 342, + 392, + 355 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 365, + 545, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 365, + 545, + 509 + ], + "spans": [ + { + "bbox": [ + 304, + 365, + 545, + 509 + ], + "type": "text", + "content": "Vision-Language Models. Inspired by the impressive success of LLMs [8, 61, 75], the recent popular VLMs tend to regard the powerful LLMs as the core backbone. At the beginning, VLMs usually use large-scale image-text pairwise datasets [9, 35, 46] or arbitrarily interleaved visual and textual data [2, 90] to pre-train. Furthermore, thanks to the availability of enormous image-text instruction datasets [41, 49], recent studies [13, 42, 48, 81, 89] further apply instruction tuning to help VLMs generate satisfactory answers. Benefiting from the two-stage training process, recent VLMs can achieve stunning performance on downstream vision-language tasks [3, 33, 46, 64]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 510, + 546, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 510, + 546, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 510, + 546, + 715 + ], + "type": "text", + "content": "Evaluations of VLMs. To evaluate the abilities of VLMs, there are diverse types of vision language downstream tasks. Conventional benchmarks, such as image caption tasks [29, 82] and visual question reasoning tasks [23, 67], mainly probe specific abilities of VLMs from the third-person perspective. Meanwhile, specialized analytical studies comprehensively evaluate the performance of VLMs from the third-person perspective, where Vblue [87] consists of five fundamental tasks and Lvmehub [77] evaluates six categories of capabilities on 47 standard vision-language benchmarks. As for the first-person perspective, there are some egocentric evaluation benchmarks in the computer vision field to assess some visual capabilities [32, 53, 83, 88]. In terms of multi-modality, there are a few benchmarks, such as EgoVQA[16] and EgoTaskQA [34], where mainly specific tasks without an overall understanding. In this paper, we mainly focus on exploring the comprehensive capa" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "14292" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 48, + 70, + 294, + 213 + ], + "blocks": [ + { + "bbox": [ + 48, + 70, + 294, + 213 + ], + "lines": [ + { + "bbox": [ + 48, + 70, + 294, + 213 + ], + "spans": [ + { + "bbox": [ + 48, + 70, + 294, + 213 + ], + "type": "image", + "image_path": "000713ea875894cd593e908d6fbae158eb919b691d82eb9ec6920158683cf253.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 49, + 213, + 294, + 266 + ], + "blocks": [ + { + "bbox": [ + 49, + 213, + 294, + 266 + ], + "lines": [ + { + "bbox": [ + 49, + 213, + 294, + 266 + ], + "spans": [ + { + "bbox": [ + 49, + 213, + 294, + 266 + ], + "type": "image", + "image_path": "f4e2c9ff8b5c15d045074ce18a376f0c84be14de1697d851ce51296e1a125c81.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 48, + 266, + 294, + 367 + ], + "blocks": [ + { + "bbox": [ + 48, + 266, + 294, + 367 + ], + "lines": [ + { + "bbox": [ + 48, + 266, + 294, + 367 + ], + "spans": [ + { + "bbox": [ + 48, + 266, + 294, + 367 + ], + "type": "image", + "image_path": "1ec81f93bd954bd3db813f746843b2f2cc2edd9bdcc9993fd93f54f145d82ec1.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 96, + 376, + 495, + 388 + ], + "lines": [ + { + "bbox": [ + 96, + 376, + 495, + 388 + ], + "spans": [ + { + "bbox": [ + 96, + 376, + 495, + 388 + ], + "type": "text", + "content": "Figure 2. Categories with fine-grained dimensions and their corresponding examples of EgoThink benchmark." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 298, + 70, + 543, + 213 + ], + "blocks": [ + { + "bbox": [ + 298, + 70, + 543, + 213 + ], + "lines": [ + { + "bbox": [ + 298, + 70, + 543, + 213 + ], + "spans": [ + { + "bbox": [ + 298, + 70, + 543, + 213 + ], + "type": "image", + "image_path": "a0d8999f7132f208ed4af268ab802c7dd76884612dba1181fe4d5b8681d14cc7.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 298, + 213, + 543, + 266 + ], + "blocks": [ + { + "bbox": [ + 298, + 213, + 543, + 266 + ], + "lines": [ + { + "bbox": [ + 298, + 213, + 543, + 266 + ], + "spans": [ + { + "bbox": [ + 298, + 213, + 543, + 266 + ], + "type": "image", + "image_path": "de1d8420fbed3468228b674dc89328b43fdbe4a715ccc6dc4269debb8affe0c6.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 298, + 266, + 543, + 367 + ], + "blocks": [ + { + "bbox": [ + 298, + 266, + 543, + 367 + ], + "lines": [ + { + "bbox": [ + 298, + 266, + 543, + 367 + ], + "spans": [ + { + "bbox": [ + 298, + 266, + 543, + 367 + ], + "type": "image", + "image_path": "6966aa81ea571d068b91902ba9375c3d662181e2df522d9107d712d207bfffb1.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 400, + 287, + 424 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 400, + 287, + 424 + ], + "spans": [ + { + "bbox": [ + 47, + 400, + 287, + 424 + ], + "type": "text", + "content": "bilities of VLMs to think from a first-person perspective, as a supplement to previous evaluation benchmarks." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 436, + 176, + 449 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 436, + 176, + 449 + ], + "spans": [ + { + "bbox": [ + 47, + 436, + 176, + 449 + ], + "type": "text", + "content": "3. EgoThink Benchmark" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 456, + 287, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 456, + 287, + 517 + ], + "spans": [ + { + "bbox": [ + 47, + 456, + 287, + 517 + ], + "type": "text", + "content": "In this section, we first elaborate on the core capabilities of thinking from a first-person perspective. Then, we introduce the process to manually construct our proposed benchmark EgoThink, which asks VLMs to generate open-ended answers according to first-person images and questions." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 525, + 151, + 537 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 525, + 151, + 537 + ], + "spans": [ + { + "bbox": [ + 47, + 525, + 151, + 537 + ], + "type": "text", + "content": "3.1. Core Capabilities" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 544, + 287, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 544, + 287, + 579 + ], + "spans": [ + { + "bbox": [ + 47, + 544, + 287, + 579 + ], + "type": "text", + "content": "As shown in Figure 2, we specifically design six categories with twelve fine-grained dimensions from the first-person perspective for quantitative evaluation." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 47, + 582, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 582, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 582, + 287, + 714 + ], + "type": "text", + "content": "- Object: What is around me? Recognizing objects in the real world is a preliminary ability of the human visual system [50, 85, 91]. Images from a first-person or egocentric perspective [53, 65, 88] pay more attention to the objects surrounding the subject or in hands. Moreover, we further divide the object category into three fine-grained dimensions: (1) Existence, predicting whether there is an object as described in the images; (2) Attribute [17, 37], detecting properties or characteristics (e.g., color) of an object; (3) Affordance [28, 56], predicting potential actions that a human can apply to an object." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 400, + 545, + 714 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 306, + 400, + 545, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 400, + 545, + 460 + ], + "spans": [ + { + "bbox": [ + 306, + 400, + 545, + 460 + ], + "type": "text", + "content": "- Activity: What am I doing? Activity recognition is to automatically recognize specific human activities in video frames or still images [36, 38, 74]. From the egocentric perspective, we mainly focus on actions or activities based on object-hand interaction [6, 18, 59]." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 460, + 545, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 460, + 545, + 567 + ], + "spans": [ + { + "bbox": [ + 306, + 460, + 545, + 567 + ], + "type": "text", + "content": "- Localization: Where am I? In reality, localization is a critical capability for navigation and scene understanding in the real world [55, 66]. Here we investigate the localization capability from two aspects, Location and Spatial Relationship. Location indicates detecting the scene surrounding the subject [14, 26]. Spatial reasoning contains allocentric and egocentric perspectives [24, 39, 57, 58]. We focus on the egocentric perspective, i.e., the position of the object with respect to the subject." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 570, + 545, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 570, + 545, + 677 + ], + "spans": [ + { + "bbox": [ + 306, + 570, + 545, + 677 + ], + "type": "text", + "content": "- Reasoning: What about the situation around me? During the complex decision-making process, reasoning lies everywhere in our lives. Here we mainly focus on Counting, Comparison, and Situated Reasoning. Due to the first-person perspective, we generally count or compare objects in our hands or surrounding ourselves. As for situated reasoning, we employ cases that cannot be answered directly from the information in the images and require further reasoning processes." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 678, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 678, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 306, + 678, + 545, + 714 + ], + "type": "text", + "content": "- Forecasting: What will happen to me? Forecasting [20, 25, 51, 52] is a critical skill in the real world. From an egocentric view, forecasting always predicts the future of" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14293" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 49, + 70, + 545, + 369 + ], + "blocks": [ + { + "bbox": [ + 49, + 70, + 545, + 369 + ], + "lines": [ + { + "bbox": [ + 49, + 70, + 545, + 369 + ], + "spans": [ + { + "bbox": [ + 49, + 70, + 545, + 369 + ], + "type": "table", + "html": "
VLMsImage EncoderLLMAlignment ModuleTTPToPDataset SizeEgoDataVideo
Image-TextInstruction
API-based Model
GPT-4V [60]Unknown
~7B Models
OpenFlamingo-7B [2, 5]CLIP-ViT-LMPT7BAttention1.4B8.1B2B-X✓
BLIP-2-6.7B [43]EVA-CLIP-ViT-gOPT6.7BQ-Former108M7.8B129M-XX
VideoChat-7B [44]BLIP2-VEVicuna7BQ-Former205M8B25M18KX✓
LLaVA-1.5-7B [47]CLIP-ViT-L-336pxLlama27BMLP6.8B7.1B558k665kXX
MiniGPT-4-7B [89]BLIP2-VELlama27BLinear23M7.7B5M3.5kXX
InstructBLIP-7B [13]EVA-CLIP-ViT-gVicuna7BQ-Former189M7.9B-16MXX
LLaMA-Adapter-7B [22]CLIP-ViT-LLLaMA7BEarly Fusion14M7.2B567k52kXX
Otter-I-7B [42]CLIP-ViT-LMPT7BAttention1.4B8.1B-2.8BX✓
PandaGPT-7B [70]ImageBindVicuna7BLinear + LLM LoRA38M7.9B-160k✓✓
mPLUG-owl-7B [81]CLIP-ViT-LLLaMA7BAttention4M7.1B204M158kXX
Video-LLaVA-7B [45]LanguageBindVicuna7BLinear6.8B7.5B1260k765kX✓
LLaVA-7B [48]CLIP-ViT-LLlama27BLinear6.7B7.1B595k158kXX
ShareGPT4V-7B [10]CLIP-ViT-L-336pxVicuna7BMLP6.7B6.7B1.2M665kXX
~13B Models
InstructBLIP-13B [13]EVA-CLIP-ViT-gVicuna13BQ-Former189M14.2B-16MXX
PandaGPT-13B [70]ImageBindVicuna13BLinear+LLM LoRA52M13.1B-160k✓✓
LLaVA-13B-Vicuna [48]CLIP-ViT-L-336pxVicuna13BLinear13.0B13.3B595k158kXX
BLIP-2-11B [43]EVA-CLIP-ViT-gFlanT5XXLQ-Former108M12.2B129M-XX
InstructBLIP-11B [13]EVA-CLIP-ViT-gFlanT5XXLQ-Former189M12.3B-16MXX
LLaVA-13B-Llama2 [48]CLIP-ViT-LLlama213BLinear13.0B13.3B595k158kXX
LLaVA-1.5-13B [47]CLIP-ViT-L-336pxLlama213BMLP13.0B13.4B558k665kXX
", + "image_path": "e93175a76128cfd0a94f429aa0b7c9b37c165980170fd486502352f68d04b021.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 376, + 546, + 400 + ], + "lines": [ + { + "bbox": [ + 46, + 376, + 546, + 400 + ], + "spans": [ + { + "bbox": [ + 46, + 376, + 546, + 400 + ], + "type": "text", + "content": "Table 2. Statistics of compared API-based and open-source VLMs, where TTP and ToP indicate Total Trainable Parameters and Total Parameters, respectively. Moreover, EgoData and Video indicate that there are egocentric visual data and video data for training, respectively." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 411, + 277, + 423 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 411, + 277, + 423 + ], + "spans": [ + { + "bbox": [ + 55, + 411, + 277, + 423 + ], + "type": "text", + "content": "object-state transformation or hand-object interactions." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 423, + 287, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 423, + 287, + 495 + ], + "spans": [ + { + "bbox": [ + 47, + 423, + 287, + 495 + ], + "type": "text", + "content": "- Planning: How will I do? In reality, planning [1, 30, 69] is an important capability to deal with complex problems, typically applied in Navigation [62, 63, 72] and Assistance [27, 76]. Navigation is going to a goal location from a start position, while assistance is offering instructions to solve daily problems." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 503, + 143, + 514 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 503, + 143, + 514 + ], + "spans": [ + { + "bbox": [ + 47, + 503, + 143, + 514 + ], + "type": "text", + "content": "3.2. Data Collection" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 522, + 287, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 522, + 287, + 545 + ], + "spans": [ + { + "bbox": [ + 46, + 522, + 287, + 545 + ], + "type": "text", + "content": "In this section, we mainly introduce the detailed processing to construct our EgoThink benchmark." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 545, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 545, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 545, + 287, + 713 + ], + "type": "text", + "content": "Collecting first-person visual data. Firstly, we leverage a popular and large egocentric video dataset Ego4D [25], which is designed to advance the field of first-person perception in computer vision. To obtain a diverse representation in different scenarios, Ego4D encompasses 3,670 hours of video from 931 unique camera wearers spanning 74 global locations across 9 countries. To collect first-person visual data, we begin by extracting every frame from a subset of the Ego4D video dataset, yielding a diverse raw image dataset. Please note that our current focus is solely on images, as most VLMs today do not support video input. We intend to expand our scope to include videos in our future work. Considering the heavy human labor and the diversity of scenarios, we sample images every few dozen frames." + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 320, + 411, + 531, + 535 + ], + "blocks": [ + { + "bbox": [ + 320, + 411, + 531, + 535 + ], + "lines": [ + { + "bbox": [ + 320, + 411, + 531, + 535 + ], + "spans": [ + { + "bbox": [ + 320, + 411, + 531, + 535 + ], + "type": "image", + "image_path": "c09b13c3d87b0f3d46b8e8ff00183e2229ba9b88e7a294737a86fdf647f89c75.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 537, + 545, + 582 + ], + "lines": [ + { + "bbox": [ + 304, + 537, + 545, + 582 + ], + "spans": [ + { + "bbox": [ + 304, + 537, + 545, + 582 + ], + "type": "text", + "content": "Figure 3. This chart illustrates the distribution of various scene categories within the EgoThink dataset. The 'others' category encompasses 13 different scene types, each representing less than one percent of total scenes." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 593, + 545, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 593, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 304, + 593, + 545, + 689 + ], + "type": "text", + "content": "To ensure high quality, we apply strict criteria for selecting the extracted frames. We first exclude images that lack clarity or fail to exhibit egocentric characteristics. Then, to obtain the high diversity within the dataset, we conduct a further screening to ensure that at most two images per video are included in the filtered image set. Finally, we obtain enormous high-quality images with exhibit egocentric characteristics as first-person image candidates." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "text", + "content": "Annotating question-answer pairs. Upon receiving a substantial collection of first-person image candidates, we" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14294" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 49, + 70, + 545, + 350 + ], + "blocks": [ + { + "bbox": [ + 49, + 70, + 545, + 350 + ], + "lines": [ + { + "bbox": [ + 49, + 70, + 545, + 350 + ], + "spans": [ + { + "bbox": [ + 49, + 70, + 545, + 350 + ], + "type": "table", + "html": "
MethodsObjectActivityLocalizationReasoningForecastingPlanningAverage
ExistAttrAffordLocSpatialCountComparSituatedNavAssist
API-based model
GPT-4V62.082.058.059.586.062.042.048.083.055.064.084.065.5
~7B Models
OpenFlamingo-7B16.055.037.015.034.034.021.040.021.031.011.011.027.2
BLIP-2-6.7B49.029.039.033.560.031.03.021.033.025.08.06.028.1
VideoChat-7B46.044.036.045.061.042.036.039.032.026.513.021.036.8
LLaVA-1.5-7B33.047.054.035.535.049.020.047.037.027.029.054.039.0
MiniGPT-4-7B50.056.046.039.055.049.014.048.031.041.514.044.040.6
InstructBLIP-7B50.033.045.047.577.038.018.043.067.040.519.031.042.4
LLaMA-Adapter-7B37.060.046.034.548.051.029.039.025.041.542.057.042.5
Otter-I-7B48.056.039.044.060.044.039.048.042.038.031.055.045.3
PandaGPT-7B40.056.041.037.061.052.019.052.053.043.039.061.046.2
mPLUG-owl-7B56.058.047.053.060.053.025.049.044.049.533.058.048.8
Video-LLaVA-7B56.060.053.045.086.060.039.038.060.046.511.038.049.4
LLaVA-7B63.058.050.047.081.045.024.036.047.049.535.060.049.6
ShareGPT4V-7B67.075.053.055.577.062.030.038.066.047.041.063.051.9
~13B Models
InstructBLIP-13B52.055.049.054.063.049.011.033.059.044.019.025.042.8
PandaGPT-13B35.052.041.040.568.031.032.040.047.045.516.069.043.1
LLaVA-13B-Vicuna54.062.052.046.053.046.026.044.029.044.035.066.046.4
BLIP-2-11B52.062.041.049.590.066.025.050.070.048.018.024.049.6
InstructBLIP-11B74.068.048.049.586.052.032.049.073.053.016.017.051.5
LLaVA-13B-Llama265.061.045.056.077.053.034.034.066.050.549.071.055.1
LLaVA-1.5-13B66.055.051.055.082.057.032.056.067.048.539.055.055.3
", + "image_path": "55df987711440910bcfbef23efc488d20cba05712ba741dcba12781ffcb51414.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 358, + 546, + 392 + ], + "lines": [ + { + "bbox": [ + 46, + 358, + 546, + 392 + ], + "spans": [ + { + "bbox": [ + 46, + 358, + 546, + 392 + ], + "type": "text", + "content": "Table 3. Combined single-answer grading scores on zero-shot setups for various dimensions. The **bold** indicates the best performance while the **underline** indicates the second-best performance. Exist, Attr, Afford, Loc, Spatial, Count, Compar, Situated, Nav and Assist represent existence, attribute, affordance, location, spatial relationship, counting, comparison, situated reasoning, navigation, and assistance." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 413, + 287, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 413, + 287, + 581 + ], + "spans": [ + { + "bbox": [ + 46, + 413, + 287, + 581 + ], + "type": "text", + "content": "engage six annotators to manually label question-answer pairs. Given that the EgoThink benchmark is composed of twelve dimensions, annotators were responsible for two specific dimensions. The annotators can access all the image candidates and are asked to select appropriate images to annotate their corresponding question-answer pairs to relevant categories. Once the image is selected, it will be removed from the candidates to ensure no repetition. Moreover, to ensure the correctness of our annotations, we have three additional annotators to review the question-answer pairs after the first annotation process. The annotation will not be reserved until the three annotators all agree that the first-person visual data and the assigned question-answer pairs meet the definition of a specific dimension." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 582, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 582, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 582, + 288, + 714 + ], + "type": "text", + "content": "Statistics. The EgoThink benchmark comprises a collection of 700 images across six categories with twelve fine-grained dimensions. These images are extracted from 595 videos, ensuring a broad representation of scenarios. To guarantee diversity, a wide range of scenes and concepts has been deliberately selected. As depicted in Figure 3, the dataset encompasses a diverse range of scenes, covering key scenarios relevant to EAI. Furthermore, we have meticulously crafted question and answer for each image in the EgoThink benchmark, aiming to closely replicate real-life conversations. This involves employing different" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 304, + 413, + 545, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 413, + 545, + 510 + ], + "spans": [ + { + "bbox": [ + 304, + 413, + 545, + 510 + ], + "type": "text", + "content": "question types, varying questions in length and complexity, paired with well-reasoned and accurate answers. In addition, the size of the dataset is a well-balanced trade-off between benchmark diversity and high cost of open-ended QA evaluation. We ensure our 700 examples were diverse in concepts, scenes, and videos to provide a robust performance estimation within practical limits. Detailed statistics of the EgoThink benchmark are presented in Appendix A." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 306, + 520, + 388, + 533 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 520, + 388, + 533 + ], + "spans": [ + { + "bbox": [ + 306, + 520, + 388, + 533 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 539, + 427, + 553 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 539, + 427, + 553 + ], + "spans": [ + { + "bbox": [ + 306, + 539, + 427, + 553 + ], + "type": "text", + "content": "4.1. Experimental Setups" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 557, + 545, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 557, + 545, + 653 + ], + "spans": [ + { + "bbox": [ + 304, + 557, + 545, + 653 + ], + "type": "text", + "content": "Vision-Language Models. We collect the most popular eighteen types of representative VLMs to assess as shown in Table 2. Due to the possible effects of model parameters, we divide models into " + }, + { + "bbox": [ + 304, + 557, + 545, + 653 + ], + "type": "inline_equation", + "content": "\\sim 7\\mathrm{B}" + }, + { + "bbox": [ + 304, + 557, + 545, + 653 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 557, + 545, + 653 + ], + "type": "inline_equation", + "content": "\\sim 13\\mathrm{B}" + }, + { + "bbox": [ + 304, + 557, + 545, + 653 + ], + "type": "text", + "content": " for a fair comparison. Detailed information about VLMs can be found in Appendix B. We conduct zero-shot setups for all VLMs across our EgoThink benchmark. The prompts used for each VLM are shown in Appendix C." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 654, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 547, + 715 + ], + "type": "text", + "content": "Single-answer grading. Considering evaluating open-ended model generations is not a trivial problem [54, 68], we propose to use GPT-4 [60] as the automatic evaluator [86] to better measure generated answers. In this protocol, we want to measure how close one model output" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14295" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 47, + 69, + 287, + 308 + ], + "blocks": [ + { + "bbox": [ + 47, + 69, + 287, + 308 + ], + "lines": [ + { + "bbox": [ + 47, + 69, + 287, + 308 + ], + "spans": [ + { + "bbox": [ + 47, + 69, + 287, + 308 + ], + "type": "image", + "image_path": "d44fe81b279c39d982f12209a2ca04758d000397b08cb8f750d16f0e4d381ccc.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 312, + 288, + 357 + ], + "lines": [ + { + "bbox": [ + 46, + 312, + 288, + 357 + ], + "spans": [ + { + "bbox": [ + 46, + 312, + 288, + 357 + ], + "type": "text", + "content": "Figure 4. Case study (wrong) of GPT-4V in the existence dimension of the object ability. In both the question formats of yes/no and what, GPT-4V can not correctly detect the chopsticks in my right hand, while InstructBLIP-11B and LLaVA-7B can." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 371, + 287, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 371, + 287, + 479 + ], + "spans": [ + { + "bbox": [ + 46, + 371, + 287, + 479 + ], + "type": "text", + "content": "is to the reference. Different from traditional similarity-based methods, GPT-4 pays more attention to semantics. In the detailed implementation, we format the question, the model output, and the reference in a prompt as shown in Appendix D and feed it into the GPT-4 evaluator. The GPT-4 evaluator is asked to assign a score of 0 (wrong), 0.5 (partially correct), or 1 (correct) to the model output. Additionally, we further discuss to use of GPT-3.5-Turbo, Claude-2, and humans as evaluators in Section 5.2." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 489, + 105, + 502 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 489, + 105, + 502 + ], + "spans": [ + { + "bbox": [ + 47, + 489, + 105, + 502 + ], + "type": "text", + "content": "4.2. Results" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 509, + 288, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 509, + 288, + 652 + ], + "spans": [ + { + "bbox": [ + 46, + 509, + 288, + 652 + ], + "type": "text", + "content": "We first present the overall results of the evaluated models on our EgoThink benchmark as shown in Table 3. Despite having improved over the years, VLMs are still difficult to think from a first-person perspective, even GPT-4V. Among the six categories, only the scores on planning and localization are relatively high, the performance in other capabilities can only reach around 60 points at best. Among the better models, GPT-4V generally performs much better than other models, only falling short in localization dimension compared to BLIP-2-11B. We will further introduce the detailed scores across different dimensions. More case studies can be found in Appendix E." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 653, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 287, + 715 + ], + "type": "text", + "content": "Results on object. In detail, we evaluate through three dimensions, including existence, attribute, and affordance. For existence, InstructBLIP-11B and ShareGPT4V-7B achieve the top-2 performance, indicating that they can predict the object precisely from the first-person perspec" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 307, + 69, + 545, + 361 + ], + "blocks": [ + { + "bbox": [ + 307, + 69, + 545, + 361 + ], + "lines": [ + { + "bbox": [ + 307, + 69, + 545, + 361 + ], + "spans": [ + { + "bbox": [ + 307, + 69, + 545, + 361 + ], + "type": "image", + "image_path": "4da87f053f0a04d7b2caff67fb5573d56f164be4d01d086d3f266a4994ba5c23.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 365, + 546, + 422 + ], + "lines": [ + { + "bbox": [ + 304, + 365, + 546, + 422 + ], + "spans": [ + { + "bbox": [ + 304, + 365, + 546, + 422 + ], + "type": "text", + "content": "Figure 5. Case studies (wrong) in the attribute, affordance, and activity dimensions. The top case demonstrates some VLMs locate the wrong place in the attribute and affordance dimension. The bottom case illustrates that in the activity dimension, some VLMs fail to detect the specific action." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 436, + 546, + 531 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 436, + 546, + 531 + ], + "spans": [ + { + "bbox": [ + 304, + 436, + 546, + 531 + ], + "type": "text", + "content": "tive. As for GPT-4V, as illustrated in Figure 4, we observe that its performance in handed object detection leaves room for improvement. As for both attribute and affordance, the GPT-4V model has demonstrated superior performance, especially in the attribute dimension. In both dimensions, some open-source models as shown at the top of Figure 5 locate the wrong place or only answer the type of the object rather than its attribute or affordance." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 533, + 546, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 533, + 546, + 652 + ], + "spans": [ + { + "bbox": [ + 304, + 533, + 546, + 652 + ], + "type": "text", + "content": "Results on activity. The performance of GPT-4V outperforms all open-source VLMs in the activity dimension. Among the " + }, + { + "bbox": [ + 304, + 533, + 546, + 652 + ], + "type": "inline_equation", + "content": "\\sim 7\\mathrm{B}" + }, + { + "bbox": [ + 304, + 533, + 546, + 652 + ], + "type": "text", + "content": " models, ShareGPT4V-7B and mPLUGowl-7B significantly outperform other VLMs and even achieve superior or comparable performance to " + }, + { + "bbox": [ + 304, + 533, + 546, + 652 + ], + "type": "inline_equation", + "content": "\\sim 13\\mathrm{B}" + }, + { + "bbox": [ + 304, + 533, + 546, + 652 + ], + "type": "text", + "content": " models. Overall, " + }, + { + "bbox": [ + 304, + 533, + 546, + 652 + ], + "type": "inline_equation", + "content": "\\sim 13\\mathrm{B}" + }, + { + "bbox": [ + 304, + 533, + 546, + 652 + ], + "type": "text", + "content": " models tend to perform better than " + }, + { + "bbox": [ + 304, + 533, + 546, + 652 + ], + "type": "inline_equation", + "content": "\\sim 7\\mathrm{B}" + }, + { + "bbox": [ + 304, + 533, + 546, + 652 + ], + "type": "text", + "content": " model in the activity dimension, but their scores are just below the passing line. The most possible reason is that detecting the specific action is difficult for VLMs as shown at the bottom of Figure 5." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 653, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 653, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 653, + 547, + 715 + ], + "type": "text", + "content": "Results on localization. In general, BLIP-2-11B has shown obvious advantages among all VLMs, even surpassing GPT-4V in both location and spatial relationship dimensions. In the location dimension, BLIP-2-11B, GPT-4V, and InstructBLIP-11B demonstrate superior ability to achieve" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "14296" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 47, + 70, + 287, + 380 + ], + "blocks": [ + { + "bbox": [ + 47, + 70, + 287, + 380 + ], + "lines": [ + { + "bbox": [ + 47, + 70, + 287, + 380 + ], + "spans": [ + { + "bbox": [ + 47, + 70, + 287, + 380 + ], + "type": "image", + "image_path": "f1b05be3e4f0cca5aaea07d275409330047ef68ca19bbeb0b8590619fcd3cdc3.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 384, + 288, + 441 + ], + "lines": [ + { + "bbox": [ + 46, + 384, + 288, + 441 + ], + "spans": [ + { + "bbox": [ + 46, + 384, + 288, + 441 + ], + "type": "text", + "content": "Figure 6. Case studies (wrong) in counting (top) and situated reasoning (bottom) dimensions. The top case shows VLMs can count but can not understand their relative position. The bottom case displays that the question requires commonsense knowledge and image understanding where only GPT-4V can answer it correctly." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 457, + 287, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 457, + 287, + 506 + ], + "spans": [ + { + "bbox": [ + 46, + 457, + 287, + 506 + ], + "type": "text", + "content": "around 90 points. However, perceiving the spatial relationship of an object relative to oneself is much more difficult. This phenomenon can be also reflected in the top of Figure 5 where VLMs hard to distinguish left or right hand." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 508, + 287, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 508, + 287, + 639 + ], + "spans": [ + { + "bbox": [ + 46, + 508, + 287, + 639 + ], + "type": "text", + "content": "Results on reasoning. Counting is the most difficult ability [80] among all evaluated dimensions. The best-performing model, GPT-4V, only scores 42.0, far away from satisfaction. Under the first-person perspective setup, VLMs need to not only count but also understand the relative position to oneself, as shown in the top case of Figure 6. Meanwhile, the comparison dimension also reflects the high difficulty, where the best score of 56.0 is obtained by LLaVA-1.5-13B. As for situated reasoning, GPT-4V demonstrates its strong commonsense reasoning ability to answer complex questions at the bottom of Figure 6." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 642, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 642, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 642, + 288, + 715 + ], + "type": "text", + "content": "Results on forecasting. Achieving high performance seems to be challenging as the best score achieved by GPT-4V is only 55.0. InsturctBLIP-11B achieves a relatively high score of 53.0 which is close to that of GPT-4V. We observe that the VLMs mainly suffer from two problems: recognizing objects incorrectly or forecasting too far as shown" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 306, + 70, + 545, + 426 + ], + "blocks": [ + { + "bbox": [ + 306, + 70, + 545, + 426 + ], + "lines": [ + { + "bbox": [ + 306, + 70, + 545, + 426 + ], + "spans": [ + { + "bbox": [ + 306, + 70, + 545, + 426 + ], + "type": "image", + "image_path": "c5c3cecf1be63fbe682bd6df7053c614d6f0d9a7688f6686d59b58e73bfc3be0.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 430, + 547, + 475 + ], + "lines": [ + { + "bbox": [ + 305, + 430, + 547, + 475 + ], + "spans": [ + { + "bbox": [ + 305, + 430, + 547, + 475 + ], + "type": "text", + "content": "Figure 7. Case studies (wrong) in the forecast (top) and navigation dimension (bottom). The top case shows VLMs might detect the glove as a cloth, while the bottom case indicates VLMs lack navigation details and overlook image information." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 486, + 396, + 497 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 486, + 396, + 497 + ], + "spans": [ + { + "bbox": [ + 306, + 486, + 396, + 497 + ], + "type": "text", + "content": "in the top of Figure 7." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 498, + 545, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 498, + 545, + 605 + ], + "spans": [ + { + "bbox": [ + 304, + 498, + 545, + 605 + ], + "type": "text", + "content": "Results on planning. In both navigation and assistance dimensions, the highest scores are achieved by GPT-4V with 60.0 and 84.0, respectively. LLaVA-13B-Llama2 behaves well in both dimensions with the second-best performance but its score is still 10 points lower than that of GPT-4V. The most possible reason is that answers provided by most open-source VLMs lack crucial details or overlook important information given in the images, as illustrated at the bottom of Figure 7." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 616, + 365, + 628 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 616, + 365, + 628 + ], + "spans": [ + { + "bbox": [ + 306, + 616, + 365, + 628 + ], + "type": "text", + "content": "5. Analysis" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 635, + 435, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 635, + 435, + 647 + ], + "spans": [ + { + "bbox": [ + 306, + 635, + 435, + 647 + ], + "type": "text", + "content": "5.1. Effects of Components" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 654, + 545, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 545, + 689 + ], + "type": "text", + "content": "As shown in Table 2, VLMs consist of multiple key components. In this section, we probe the influence of different components on our EgoThink benchmark." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 689, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 545, + 712 + ], + "type": "text", + "content": "The total parameters of LLMs. Here we compare the performance of " + }, + { + "bbox": [ + 306, + 689, + 545, + 712 + ], + "type": "inline_equation", + "content": "\\sim 7\\mathrm{B}" + }, + { + "bbox": [ + 306, + 689, + 545, + 712 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 306, + 689, + 545, + 712 + ], + "type": "inline_equation", + "content": "\\sim 13\\mathrm{B}" + }, + { + "bbox": [ + 306, + 689, + 545, + 712 + ], + "type": "text", + "content": " variants of four VLMs. Note" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14297" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 48, + 69, + 289, + 214 + ], + "blocks": [ + { + "bbox": [ + 48, + 69, + 289, + 214 + ], + "lines": [ + { + "bbox": [ + 48, + 69, + 289, + 214 + ], + "spans": [ + { + "bbox": [ + 48, + 69, + 289, + 214 + ], + "type": "image", + "image_path": "31240c3a0da56d36cf88d2c198cda83630df797e727752ce45d4e909d61048d3.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 219, + 288, + 252 + ], + "lines": [ + { + "bbox": [ + 46, + 219, + 288, + 252 + ], + "spans": [ + { + "bbox": [ + 46, + 219, + 288, + 252 + ], + "type": "text", + "content": "Figure 8. Impact of LLMs sizes (above the dash-line) and instruction-tuning (below the dash-line) on model performance. Average scores across all capabilities are reported." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 266, + 287, + 397 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 266, + 287, + 397 + ], + "spans": [ + { + "bbox": [ + 46, + 266, + 287, + 397 + ], + "type": "text", + "content": "that the increase in the number of parameters mainly falls in the LLMs. Firstly, as shown in the top part of Figure 8, scaling does not lead to significant improvement for PandaGPT and InstructBLIP, while LLaVA (LLaVA-7B and LLava13B-Llama2) and LLaVA-1.5 benefit a lot from scaling. We hypothesize that this is because LLaVA series models do not freeze their language models during instruction tuning, indicating that enlarging the number of trainable parameters can help improve both performance and generalization. In other words, one can see that simply scaling up language models without better alignment may not help." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 398, + 287, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 398, + 287, + 517 + ], + "spans": [ + { + "bbox": [ + 46, + 398, + 287, + 517 + ], + "type": "text", + "content": "Instruction tuning. We directly compare the performance of BLIP-2-11B and InstructBLIP-11B, because these two models differ only in instruction tuning and additional instruction-aware tokens. As presented in the bottom part of Figure 8, InstructBLIP-11B outperforms BLIP-2-11B after instruction tuning, despite an unexpectedly small margin. This may be because much of the instruction tuning data employed by InstructBLIP is collected from specific downstream tasks, whose data distributions are very different from our first-person perspective data." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 517, + 287, + 625 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 517, + 287, + 625 + ], + "spans": [ + { + "bbox": [ + 46, + 517, + 287, + 625 + ], + "type": "text", + "content": "The information of image encoder. Considering that there is no ablation version of VLMs for image encoder, following Set-of-Mark [79], we probe the effect of visual grounding information (i.e., a set of marks) in our setups. As presented in Figure 9, GPT-4V with additional segmentation information can correctly detect the mentioned location and objects, indicating that supplemented image information can be helpful in some situations. More discussion about quantitative experiments can be found in Appendix F." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 635, + 280, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 635, + 280, + 647 + ], + "spans": [ + { + "bbox": [ + 47, + 635, + 280, + 647 + ], + "type": "text", + "content": "5.2. Agreements between Human and Evaluators" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 654, + 287, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 654, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 46, + 654, + 287, + 712 + ], + "type": "text", + "content": "In this section, we further assess the model performance on object and planning dimensions using GPT-3.5-Turbo, Claude-2, and human annotators. Due to the heavy human labor, we ask three annotators to evaluate the performance of GPT-4V, which is the overall best model. Human anno" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 306, + 69, + 545, + 196 + ], + "blocks": [ + { + "bbox": [ + 306, + 69, + 545, + 196 + ], + "lines": [ + { + "bbox": [ + 306, + 69, + 545, + 196 + ], + "spans": [ + { + "bbox": [ + 306, + 69, + 545, + 196 + ], + "type": "image", + "image_path": "009c8433541138dc8a35fec52d71473cc708944105bc5c953b6ab0430514af07.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 208, + 545, + 242 + ], + "lines": [ + { + "bbox": [ + 305, + 208, + 545, + 242 + ], + "spans": [ + { + "bbox": [ + 305, + 208, + 545, + 242 + ], + "type": "text", + "content": "Figure 9. Case study (wrong) in the adding visual grounding information with images. The segmentation can help VLMs better locate the objects in question." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 253, + 546, + 433 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 253, + 546, + 433 + ], + "spans": [ + { + "bbox": [ + 304, + 253, + 546, + 433 + ], + "type": "text", + "content": "tators consider the following aspects to evaluate: accuracy, completeness, logical soundness, and grammatical correctness. Our annotation system and detailed guidelines can be found in Appendix G. We further conduct GPT-3.5-Turbo and Claude-2 with the same evaluation prompt as GPT-4. The Pearson correlation coefficients between automatic evaluators (i.e., GPT-4, GPT-3.5-Turbo, Claude-2) and humans are 0.68, 0.43, and 0.68, respectively. The Cohen's Kappa coefficient among the three annotators is 0.81. This shows that evaluations made by GPT-4 and Claude-2 have a high correlation with humans. We hypothesize that recent well-performant LLMs can evaluate highly aligned with humans, given that most answers in our benchmark are relatively short and precise. Detailed scores of all evaluators and their correlations are discussed in Appendix H." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 442, + 379, + 454 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 442, + 379, + 454 + ], + "spans": [ + { + "bbox": [ + 306, + 442, + 379, + 454 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 462, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 462, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 462, + 546, + 713 + ], + "type": "text", + "content": "To pave the way for the development of VLMs in the field of EAI and robotics, we introduce a comprehensive benchmark, EgoThink. Designed to evaluate the capacity of VLMs to \"think\" from a first-person perspective, EgoThink encompasses six core capabilities across twelve detailed dimensions. We assess eighteen popular VLMs and find that even the top-performing VLMs in most dimensions achieve only around a score of 60. GPT-4V achieves the best overall performance, but can not consistently surpass other open-source VLMs across all dimensions. In the analysis, we further probe the impact of various components on model performance and find that the total number of trainable parameters in LLMs has the most significant effect. Despite the human agreement with automatic evaluators being high, the evaluation of planning is difficult due to the detailed information in the answers. In future research, we aim to improve the evaluation method and further explore the essential capabilities of VLMs in the EAI and robotics fields. Acknowledgment: The work is supported by the National Key R&D Program of China (2022ZD0160502) and the National Natural Science Foundation of China (No.61925601)." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14298" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 146 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 146 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 146 + ], + "type": "text", + "content": "[1] Michael Ahn, Anthony Brohan, Noah Brown, Yevgen Chebotar, Omar Cortes, Byron David, Chelsea Finn, Keerthana Gopalakrishnan, Karol Hausman, Alex Herzog, et al. Do as i can, not as i say: Grounding language in robotic affordances. arXiv preprint arXiv:2204.01691, 2022. 4" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 148, + 288, + 213 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 148, + 288, + 213 + ], + "spans": [ + { + "bbox": [ + 53, + 148, + 288, + 213 + ], + "type": "text", + "content": "[2] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. Advances in Neural Information Processing Systems, 35:23716-23736, 2022. 1, 2, 4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 215, + 288, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 215, + 288, + 270 + ], + "spans": [ + { + "bbox": [ + 53, + 215, + 288, + 270 + ], + "type": "text", + "content": "[3] Stanislaw Antol, Aishwarya Agrawal, Jiasen Lu, Margaret Mitchell, Dhruv Batra, C Lawrence Zitnick, and Devi Parikh. Vqa: Visual question answering. In Proceedings of the IEEE international conference on computer vision, pages 2425-2433, 2015. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 272, + 288, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 272, + 288, + 327 + ], + "spans": [ + { + "bbox": [ + 53, + 272, + 288, + 327 + ], + "type": "text", + "content": "[4] Anas Awadalla, Irena Gao, Joshua Gardner, Jack Hessel, Yusuf Hanafy, Wanrong Zhu, Kalyani Marathe, Yonatan Bitton, Samir Gadre, Jenia Jitsev, Simon Kornblith, Pang Wei Koh, Gabriel Ilharco, Mitchell Wortsman, and Ludwig Schmidt. Openflamingo, 2023. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 328, + 288, + 393 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 328, + 288, + 393 + ], + "spans": [ + { + "bbox": [ + 53, + 328, + 288, + 393 + ], + "type": "text", + "content": "[5] Anas Awadalla, Irena Gao, Josh Gardner, Jack Hessel, Yusuf Hanafy, Wanrong Zhu, Kalyani Marathe, Yonatan Bitton, Samir Gadre, Shiori Sagawa, et al. Openflamingo: An opensource framework for training large autoregressive vision-language models. arXiv preprint arXiv:2308.01390, 2023. 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 396, + 288, + 451 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 396, + 288, + 451 + ], + "spans": [ + { + "bbox": [ + 53, + 396, + 288, + 451 + ], + "type": "text", + "content": "[6] Sven Bambach, Stefan Lee, David J Crandall, and Chen Yu. Lending a hand: Detecting hands and recognizing activities in complex egocentric interactions. In Proceedings of the IEEE international conference on computer vision, pages 1949-1957, 2015. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 453, + 288, + 497 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 453, + 288, + 497 + ], + "spans": [ + { + "bbox": [ + 53, + 453, + 288, + 497 + ], + "type": "text", + "content": "[7] William Berrios, Gautam Mittal, Tristan Thrush, Douwe Kiela, and Amanpreet Singh. Towards language models that can see: Computer vision through the lens of natural language. arXiv preprint arXiv:2306.16410, 2023. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 498, + 288, + 554 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 498, + 288, + 554 + ], + "spans": [ + { + "bbox": [ + 53, + 498, + 288, + 554 + ], + "type": "text", + "content": "[8] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. Advances in neural information processing systems, 33:1877-1901, 2020. 1, 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 555, + 288, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 555, + 288, + 610 + ], + "spans": [ + { + "bbox": [ + 53, + 555, + 288, + 610 + ], + "type": "text", + "content": "[9] Soravit Changpinyo, Piyush Sharma, Nan Ding, and Radu Soricut. Conceptual 12m: Pushing web-scale image-text pretraining to recognize long-tail visual concepts. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3558-3568, 2021. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 612, + 288, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 612, + 288, + 656 + ], + "spans": [ + { + "bbox": [ + 48, + 612, + 288, + 656 + ], + "type": "text", + "content": "[10] Lin Chen, Jisong Li, Xiaoyi Dong, Pan Zhang, Conghui He, Jiaqi Wang, Feng Zhao, and Dahua Lin. Sharegpt4v: Improving large multi-modal models with better captions. arXiv preprint arXiv:2311.12793, 2023. 4, 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 658, + 288, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 288, + 713 + ], + "type": "text", + "content": "[11] Liang Chen, Yichi Zhang, Shuhuai Ren, Haozhe Zhao, Zefan Cai, Yuchi Wang, Peiyi Wang, Tianyu Liu, and Baobao Chang. Towards end-to-end embodied decision making via multi-modal large language model: Explorations with gpt4- vision and beyond, 2023. 2" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 713 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 307, + 73, + 546, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 546, + 128 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 546, + 128 + ], + "type": "text", + "content": "[12] Wei-Lin Chiang, Zhuohan Li, Zi Lin, Ying Sheng, Zhang-hao Wu, Hao Zhang, Lianmin Zheng, Siyuan Zhuang, Yong-hao Zhuang, Joseph E. Gonzalez, Ion Stoica, and Eric P. Xing. Vicuna: An open-source chatbot impressing gpt-4 with " + }, + { + "bbox": [ + 307, + 73, + 546, + 128 + ], + "type": "inline_equation", + "content": "90\\%" + }, + { + "bbox": [ + 307, + 73, + 546, + 128 + ], + "type": "text", + "content": " * chatgpt quality, 2023. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 307, + 129, + 547, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 129, + 547, + 183 + ], + "spans": [ + { + "bbox": [ + 307, + 129, + 547, + 183 + ], + "type": "text", + "content": "[13] Wenliang Dai, Junnan Li, Dongxu Li, Anthony Meng Huat Tiong, Junqi Zhao, Weisheng Wang, Boyang Li, Pascale Fung, and Steven Hoi. Instructlip: Towards general-purpose vision-language models with instruction tuning, 2023. 1, 2, 4" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 185, + 546, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 185, + 546, + 239 + ], + "spans": [ + { + "bbox": [ + 307, + 185, + 546, + 239 + ], + "type": "text", + "content": "[14] Tien Do, Ondrej Miksik, Joseph DeGol, Hyun Soo Park, and Sudipta N Sinha. Learning to detect scene landmarks for camera localization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11132-11142, 2022. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 241, + 546, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 241, + 546, + 296 + ], + "spans": [ + { + "bbox": [ + 307, + 241, + 546, + 296 + ], + "type": "text", + "content": "[15] Danny Driess, Fei Xia, Mehdi SM Sajjadi, Corey Lynch, Aakanksha Chowdhery, Brian Ichter, Ayzaan Wahid, Jonathan Thompson, Quan Vuong, Tianhe Yu, et al. Palme: An embodied multimodal language model. arXiv preprint arXiv:2303.03378, 2023. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 297, + 546, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 297, + 546, + 342 + ], + "spans": [ + { + "bbox": [ + 307, + 297, + 546, + 342 + ], + "type": "text", + "content": "[16] Chenyou Fan. Egovqa-an egocentric video question answering benchmark dataset. In Proceedings of the IEEE/CVF International Conference on Computer Vision Workshops, pages 0-0, 2019. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 343, + 546, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 343, + 546, + 386 + ], + "spans": [ + { + "bbox": [ + 307, + 343, + 546, + 386 + ], + "type": "text", + "content": "[17] Ali Farhadi, Ian Endres, Derek Hoiem, and David Forsyth. Describing objects by their attributes. In 2009 IEEE conference on computer vision and pattern recognition, pages 1778-1785. IEEE, 2009. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 388, + 546, + 421 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 388, + 546, + 421 + ], + "spans": [ + { + "bbox": [ + 307, + 388, + 546, + 421 + ], + "type": "text", + "content": "[18] Alireza Fathi, Ali Farhadi, and James M Rehg. Understanding egocentric activities. In 2011 international conference on computer vision, pages 407-414. IEEE, 2011. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 422, + 546, + 476 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 422, + 546, + 476 + ], + "spans": [ + { + "bbox": [ + 307, + 422, + 546, + 476 + ], + "type": "text", + "content": "[19] Chaoyou Fu, Peixian Chen, Yunhang Shen, Yulei Qin, Mengdan Zhang, Xu Lin, Zhenyu Qiu, Wei Lin, Jinrui Yang, Xiawu Zheng, et al. Mme: A comprehensive evaluation benchmark for multimodal large language models. arXiv preprint arXiv:2306.13394, 2023. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 478, + 546, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 478, + 546, + 521 + ], + "spans": [ + { + "bbox": [ + 307, + 478, + 546, + 521 + ], + "type": "text", + "content": "[20] Antonino Furnari, Sebastiano Battiato, Kristen Grauman, and Giovanni Maria Farinella. Next-active-object prediction from egocentric videos. Journal of Visual Communication and Image Representation, 49:401-411, 2017. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 523, + 546, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 523, + 546, + 567 + ], + "spans": [ + { + "bbox": [ + 307, + 523, + 546, + 567 + ], + "type": "text", + "content": "[21] Jensen Gao, Bidipta Sarkar, Fei Xia, Ted Xiao, Jiajun Wu, Brian Ichter, Anirudha Majumdar, and Dorsa Sadigh. Physically grounded vision-language models for robotic manipulation. arXiv preprint arXiv:2309.02561, 2023. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 568, + 546, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 568, + 546, + 622 + ], + "spans": [ + { + "bbox": [ + 307, + 568, + 546, + 622 + ], + "type": "text", + "content": "[22] Peng Gao, Jiaming Han, Renrui Zhang, Ziyi Lin, Shijie Geng, Aojun Zhou, Wei Zhang, Pan Lu, Conghui He, Xi-angyu Yue, et al. Llama-adapter v2: Parameter-efficient visual instruction model. arXiv preprint arXiv:2304.15010, 2023. 4, 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 624, + 546, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 624, + 546, + 679 + ], + "spans": [ + { + "bbox": [ + 307, + 624, + 546, + 679 + ], + "type": "text", + "content": "[23] Yash Goyal, Tejas Khot, Douglas Summers-Stay, Dhruv Bartra, and Devi Parikh. Making the v in vqa matter: Elevating the role of image understanding in visual question answering. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6904-6913, 2017. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 681, + 546, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 681, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 681, + 546, + 713 + ], + "type": "text", + "content": "[24] Stéphane Grade, Mauro Pesenti, and Martin G Edwards. Evidence for the embodiment of space perception: concurrent hand but not arm action moderates reachability and egocen-" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "14299" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 66, + 72, + 287, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 72, + 287, + 95 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 287, + 95 + ], + "type": "text", + "content": "tric distance perception. Frontiers in Psychology, 6:862, 2015. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 97, + 287, + 163 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 97, + 287, + 163 + ], + "spans": [ + { + "bbox": [ + 48, + 97, + 287, + 163 + ], + "type": "text", + "content": "[25] Kristen Grauman, Andrew Westbury, Eugene Byrne, Zachary Chavis, Antonino Furnari, Rohit Girdhar, Jackson Hamburger, Hao Jiang, Miao Liu, Xingyu Liu, et al. Ego4d: Around the world in 3,000 hours of egocentric video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18995-19012, 2022. 3, 4" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 165, + 288, + 220 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 165, + 288, + 220 + ], + "spans": [ + { + "bbox": [ + 48, + 165, + 288, + 220 + ], + "type": "text", + "content": "[26] Jean-Bernard Hayet, Frédéric Lerasle, and Michel Devy. Visual landmarks detection and recognition for mobile robot navigation. In 2003 IEEE Computer Society Conference on Computer Vision and Pattern Recognition, 2003. Proceedings., pages II-II. IEEE, 2003. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 223, + 288, + 277 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 223, + 288, + 277 + ], + "spans": [ + { + "bbox": [ + 48, + 223, + 288, + 277 + ], + "type": "text", + "content": "[27] Evert Helms, Rolf Dieter Schraft, and M Hagele. rob@ work: Robot assistant in industrial environments. In Proceedings. 11th IEEE International Workshop on Robot and Human Interactive Communication, pages 399-404. IEEE, 2002. 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 280, + 287, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 280, + 287, + 335 + ], + "spans": [ + { + "bbox": [ + 48, + 280, + 287, + 335 + ], + "type": "text", + "content": "[28] Tucker Hermans, James M Rehg, and Aaron Bobick. Affordance prediction via learned object attributes. In IEEE international conference on robotics and automation (ICRA): Workshop on semantic perception, mapping, and exploration, pages 181-184. CiteSeer, 2011. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 337, + 287, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 337, + 287, + 380 + ], + "spans": [ + { + "bbox": [ + 48, + 337, + 287, + 380 + ], + "type": "text", + "content": "[29] MD Zakir Hossain, Ferdous Sohel, Mohd Fairuz Shiratuddin, and Hamid Laga. A comprehensive survey of deep learning for image captioning. ACM Computing Surveys (CsUR), 51(6):1-36, 2019. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 383, + 287, + 438 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 383, + 287, + 438 + ], + "spans": [ + { + "bbox": [ + 48, + 383, + 287, + 438 + ], + "type": "text", + "content": "[30] Wenlong Huang, Fei Xia, Ted Xiao, Harris Chan, Jacky Liang, Pete Florence, Andy Zeng, Jonathan Tompson, Igor Mordatch, Yevgen Chebotar, et al. Inner monologue: Embodied reasoning through planning with language models. arXiv preprint arXiv:2207.05608, 2022. 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 441, + 287, + 485 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 441, + 287, + 485 + ], + "spans": [ + { + "bbox": [ + 48, + 441, + 287, + 485 + ], + "type": "text", + "content": "[31] Wenlong Huang, Chen Wang, Ruohan Zhang, Yunzhu Li, Jiajun Wu, and Li Fei-Fei. Voxposer: Composable 3d value maps for robotic manipulation with language models. arXiv preprint arXiv:2307.05973, 2023. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 487, + 287, + 541 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 487, + 287, + 541 + ], + "spans": [ + { + "bbox": [ + 48, + 487, + 287, + 541 + ], + "type": "text", + "content": "[32] Yichao Huang, Xiaorui Liu, Xin Zhang, and Lianwen Jin. A pointing gesture based egocentric interaction system: Dataset, approach and application. In Proceedings of the IEEE conference on computer vision and pattern recognition workshops, pages 16-23, 2016. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 544, + 287, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 544, + 287, + 599 + ], + "spans": [ + { + "bbox": [ + 48, + 544, + 287, + 599 + ], + "type": "text", + "content": "[33] Drew A Hudson and Christopher D Manning. Gqa: A new dataset for real-world visual reasoning and compositional question answering. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6700-6709, 2019. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 601, + 287, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 601, + 287, + 644 + ], + "spans": [ + { + "bbox": [ + 48, + 601, + 287, + 644 + ], + "type": "text", + "content": "[34] Baoxiong Jia, Ting Lei, Song-Chun Zhu, and Siyuan Huang. Egotaskqa: Understanding human tasks in egocentric videos. Advances in Neural Information Processing Systems, 35: 3343-3360, 2022. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 647, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 647, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 647, + 287, + 712 + ], + "type": "text", + "content": "[35] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In International conference on machine learning, pages 4904-4916. PMLR, 2021. 2" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "text", + "content": "[36] Charmi Jobanputra, Jatna Bavishi, and Nishant Doshi. Human activity recognition: A survey. Procedia Computer Science, 155:698-703, 2019. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 307, + 107, + 545, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 107, + 545, + 140 + ], + "spans": [ + { + "bbox": [ + 307, + 107, + 545, + 140 + ], + "type": "text", + "content": "[37] Nancy Kanwisher and Jon Driver. Objects, attributes, and visual attention: Which, what, and where. Current Directions in Psychological Science, 1(1):26-31, 1992. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 141, + 545, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 141, + 545, + 184 + ], + "spans": [ + { + "bbox": [ + 307, + 141, + 545, + 184 + ], + "type": "text", + "content": "[38] Shian-Ru Ke, Hoang Le Uyen Thuc, Yong-Jin Lee, Jenq-Neng Hwang, Jang-Hee Yoo, and Kyoung-Ho Choi. A review on video-based human activity recognition. Computers, 2(2):88-131, 2013. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 186, + 545, + 240 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 186, + 545, + 240 + ], + "spans": [ + { + "bbox": [ + 307, + 186, + 545, + 240 + ], + "type": "text", + "content": "[39] Roberta L Klatzky. Allocentric and egocentric spatial representations: Definitions, distinctions, and interconnections. In Spatial cognition: An interdisciplinary approach to representing and processing spatial knowledge, pages 1-17. Springer, 1998. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 241, + 545, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 241, + 545, + 285 + ], + "spans": [ + { + "bbox": [ + 307, + 241, + 545, + 285 + ], + "type": "text", + "content": "[40] Weicheng Kuo, Yin Cui, Xiuye Gu, AJ Piergiovanni, and Anelia Angelova. F-vlm: Open-vocabulary object detection upon frozen vision and language models. arXiv preprint arXiv:2209.15639, 2022. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 286, + 545, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 286, + 545, + 330 + ], + "spans": [ + { + "bbox": [ + 307, + 286, + 545, + 330 + ], + "type": "text", + "content": "[41] Bo Li, Yuanhan Zhang, Liangyu Chen, Jinghao Wang, Fanyi Pu, Jingkang Yang, Chunyuan Li, and Ziwei Liu. Mimic-it: Multi-modal in-context instruction tuning. arXiv preprint arXiv:2306.05425, 2023. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 331, + 545, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 331, + 545, + 374 + ], + "spans": [ + { + "bbox": [ + 307, + 331, + 545, + 374 + ], + "type": "text", + "content": "[42] Bo Li, Yuanhan Zhang, Liangyu Chen, Jinghao Wang, Jingkang Yang, and Ziwei Liu. Otter: A multi-modal model with in-context instruction tuning. arXiv preprint arXiv:2305.03726, 2023. 1, 2, 4" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 376, + 545, + 421 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 376, + 545, + 421 + ], + "spans": [ + { + "bbox": [ + 307, + 376, + 545, + 421 + ], + "type": "text", + "content": "[43] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. arXiv preprint arXiv:2301.12597, 2023. 1, 4" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 422, + 545, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 422, + 545, + 464 + ], + "spans": [ + { + "bbox": [ + 307, + 422, + 545, + 464 + ], + "type": "text", + "content": "[44] KunChang Li, Yinan He, Yi Wang, Yizhuo Li, Wenhai Wang, Ping Luo, Yali Wang, Limin Wang, and Yu Qiao. Videochat: Chat-centric video understanding. arXiv preprint arXiv:2305.06355, 2023. 4, 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 466, + 545, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 466, + 545, + 510 + ], + "spans": [ + { + "bbox": [ + 307, + 466, + 545, + 510 + ], + "type": "text", + "content": "[45] Bin Lin, Bin Zhu, Yang Ye, Munan Ning, Peng Jin, and Li Yuan. Video-llava: Learning united visual representation by alignment before projection. arXiv preprint arXiv:2311.10122, 2023. 4, 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 511, + 545, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 511, + 545, + 578 + ], + "spans": [ + { + "bbox": [ + 307, + 511, + 545, + 578 + ], + "type": "text", + "content": "[46] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 579, + 545, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 579, + 545, + 612 + ], + "spans": [ + { + "bbox": [ + 307, + 579, + 545, + 612 + ], + "type": "text", + "content": "[47] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. arXiv preprint arXiv:2310.03744, 2023. 4, 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 613, + 545, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 613, + 545, + 635 + ], + "spans": [ + { + "bbox": [ + 307, + 613, + 545, + 635 + ], + "type": "text", + "content": "[48] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning, 2023. 1, 2, 4" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 635, + 545, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 635, + 545, + 667 + ], + "spans": [ + { + "bbox": [ + 307, + 635, + 545, + 667 + ], + "type": "text", + "content": "[49] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. arXiv preprint arXiv:2304.08485, 2023. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 669, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 669, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 669, + 545, + 713 + ], + "type": "text", + "content": "[50] Li Liu, Wanli Ouyang, Xiaogang Wang, Paul Fieguth, Jie Chen, Xinwang Liu, and Matti Pietikainen. Deep learning for generic object detection: A survey. International journal of computer vision, 128:261-318, 2020. 3" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "14300" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "text", + "content": "[51] Miao Liu, Siyu Tang, Yin Li, and James M Rehg. Forecasting human-object interaction: Joint prediction of motor attention and egocentric activity. Computer Vision-ECCV 2020, 12346:704-721, 2020. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 118, + 287, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 118, + 287, + 173 + ], + "spans": [ + { + "bbox": [ + 48, + 118, + 287, + 173 + ], + "type": "text", + "content": "[52] Shaowei Liu, Subarna Tripathi, Somdeb Majumdar, and Xiaolong Wang. Joint hand motion and interaction hotspots prediction from egocentric videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3282-3292, 2022. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 174, + 287, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 174, + 287, + 239 + ], + "spans": [ + { + "bbox": [ + 48, + 174, + 287, + 239 + ], + "type": "text", + "content": "[53] Yunze Liu, Yun Liu, Che Jiang, Kangbo Lyu, Weikang Wan, Hao Shen, Boqiang Liang, Zhoujie Fu, He Wang, and Li Yi. Hoi4d: A 4d egocentric dataset for category-level human-object interaction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21013-21022, 2022. 2, 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 240, + 287, + 295 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 240, + 287, + 295 + ], + "spans": [ + { + "bbox": [ + 48, + 240, + 287, + 295 + ], + "type": "text", + "content": "[54] Yuan Liu, Haodong Duan, Yuanhan Zhang, Bo Li, Songyang Zhang, Wangbo Zhao, Yike Yuan, Jiaqi Wang, Conghui He, Ziwei Liu, et al. Mmbench: Is your multi-modal model an all-around player? arXiv preprint arXiv:2307.06281, 2023. 2, 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 297, + 287, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 297, + 287, + 330 + ], + "spans": [ + { + "bbox": [ + 48, + 297, + 287, + 330 + ], + "type": "text", + "content": "[55] Michael Montemerlo and Sebastian Thrun. FastSLAM: A scalable method for the simultaneous localization and mapping problem in robotics. Springer, 2007. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 331, + 287, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 331, + 287, + 374 + ], + "spans": [ + { + "bbox": [ + 48, + 331, + 287, + 374 + ], + "type": "text", + "content": "[56] Luis Montesano, Manuel Lopes, Alexandre Bernardino, and José Santos-Victor. Learning object affordances: from sensory-motor coordination to imitation. IEEE Transactions on Robotics, 24(1):15-26, 2008. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 376, + 287, + 419 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 376, + 287, + 419 + ], + "spans": [ + { + "bbox": [ + 48, + 376, + 287, + 419 + ], + "type": "text", + "content": "[57] Francesca Morganti, Stefano Stefanini, and Giuseppe Riva. From allo-to egocentric spatial ability in early alzheimer's disease: a study with virtual reality spatial tasks. Cognitive neuroscience, 4(3-4):171-180, 2013. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 420, + 287, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 420, + 287, + 464 + ], + "spans": [ + { + "bbox": [ + 48, + 420, + 287, + 464 + ], + "type": "text", + "content": "[58] Weimin Mou, Timothy P McNamara, Christine M Vali-quette, and Björn Rump. Allocentric and egocentric updating of spatial memories. Journal of experimental psychology: Learning, Memory, and Cognition, 30(1):142, 2004. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 466, + 287, + 509 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 466, + 287, + 509 + ], + "spans": [ + { + "bbox": [ + 48, + 466, + 287, + 509 + ], + "type": "text", + "content": "[59] Thi-Hoa-Cuc Nguyen, Jean-Christophe Nebel, and Francisco Florez-Revuelta. Recognition of activities of daily living with egocentric vision: A review. Sensors, 16(1):72, 2016. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 510, + 243, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 510, + 243, + 521 + ], + "spans": [ + { + "bbox": [ + 48, + 510, + 243, + 521 + ], + "type": "text", + "content": "[60] OpenAI. Gpt-4 technical report, 2023. 1, 2, 4, 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 523, + 287, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 523, + 287, + 588 + ], + "spans": [ + { + "bbox": [ + 48, + 523, + 287, + 588 + ], + "type": "text", + "content": "[61] Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, et al. Training language models to follow instructions with human feedback. Advances in Neural Information Processing Systems, 35: 27730-27744, 2022. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 590, + 287, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 590, + 287, + 644 + ], + "spans": [ + { + "bbox": [ + 48, + 590, + 287, + 644 + ], + "type": "text", + "content": "[62] Anish Pandey, Rakesh Kumar Sonkar, Krishna Kant Pandey, and DR Parhi. Path planning navigation of mobile robot with obstacles avoidance using fuzzy logic controller. In 2014 IEEE 8th international conference on intelligent systems and control (ISCO), pages 39-41. IEEE, 2014. 4" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 646, + 287, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 646, + 287, + 678 + ], + "spans": [ + { + "bbox": [ + 48, + 646, + 287, + 678 + ], + "type": "text", + "content": "[63] BK Patle, Anish Pandey, DRK Parhi, AJDT Jagadeesh, et al. A review: On path planning strategies for navigation of mobile robot. Defence Technology, 15(4):582-606, 2019. 4" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "type": "text", + "content": "[64] Bryan A Plummer, Liwei Wang, Chris M Cervantes, Juan C Caicedo, Julia Hockenmaier, and Svetlana Lazebnik. Flickr30k entities: Collecting region-to-phrase corre" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 327, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 545, + 106 + ], + "type": "text", + "content": "spondences for richer image-to-sentence models. In Proceedings of the IEEE international conference on computer vision, pages 2641-2649, 2015. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 106, + 545, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 106, + 545, + 171 + ], + "spans": [ + { + "bbox": [ + 308, + 106, + 545, + 171 + ], + "type": "text", + "content": "[65] Vignesh Ramanathan, Anmol Kalia, Vladan Petrovic, Yi Wen, Baixue Zheng, Baishan Guo, Rui Wang, Aaron Marquez, Rama Kovvuri, Abhishek Kadian, et al. Paco: Parts and attributes of common objects. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7141-7151, 2023. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 172, + 545, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 172, + 545, + 205 + ], + "spans": [ + { + "bbox": [ + 308, + 172, + 545, + 205 + ], + "type": "text", + "content": "[66] Sajad Saeedi, Michael Trentini, Mae Seto, and Howard Li. Multiple-robot simultaneous localization and mapping: A review. Journal of Field Robotics, 33(1):3-46, 2016. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 205, + 545, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 205, + 545, + 271 + ], + "spans": [ + { + "bbox": [ + 308, + 205, + 545, + 271 + ], + "type": "text", + "content": "[67] Dustin Schwenk, Apoorv Khandelwal, Christopher Clark, Kenneth Marino, and Roozbeh Mottaghi. A-okvqa: A benchmark for visual question answering using world knowledge. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23–27, 2022, Proceedings, Part VIII, pages 146–162. Springer, 2022. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 272, + 545, + 324 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 272, + 545, + 324 + ], + "spans": [ + { + "bbox": [ + 308, + 272, + 545, + 324 + ], + "type": "text", + "content": "[68] Wenqi Shao, Yutao Hu, Peng Gao, Meng Lei, Kaipeng Zhang, Fanqing Meng, Peng Xu, Siyuan Huang, Hongsheng Li, Yu Qiao, et al. Tiny lvlm-ehub: Early multimodal experiments with bard. arXiv preprint arXiv:2308.03729, 2023. 2, 5" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 327, + 545, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 327, + 545, + 381 + ], + "spans": [ + { + "bbox": [ + 308, + 327, + 545, + 381 + ], + "type": "text", + "content": "[69] Chan Hee Song, Jiaman Wu, Clayton Washington, Brian M Sadler, Wei-Lun Chao, and Yu Su. Llm-planner: Few-shot grounded planning for embodied agents with large language models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2998-3009, 2023. 4" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 382, + 545, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 382, + 545, + 415 + ], + "spans": [ + { + "bbox": [ + 308, + 382, + 545, + 415 + ], + "type": "text", + "content": "[70] Yixuan Su, Tian Lan, Huayang Li, Jialu Xu, Yan Wang, and Deng Cai. Pandagpt: One model to instruction-follow them all. arXiv preprint arXiv:2305.16355, 2023. 4, 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 415, + 545, + 458 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 415, + 545, + 458 + ], + "spans": [ + { + "bbox": [ + 308, + 415, + 545, + 458 + ], + "type": "text", + "content": "[71] Theodore Sumers, Kenneth Marino, Arun Ahuja, Rob Fergus, and Ishita Dasgupta. Distilling internet-scale vision-language models into embodied agents. arXiv preprint arXiv:2301.12507, 2023. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 460, + 545, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 460, + 545, + 525 + ], + "spans": [ + { + "bbox": [ + 308, + 460, + 545, + 525 + ], + "type": "text", + "content": "[72] Loreto Susperregi, Izaskun Fernandez, Ane Fernandez, Santiago Fernandez, Iñaki Maurtua, and Irene Lopez de Vallejo. Interacting with a robot: a guide robot understanding natural language instructions. In International Conference on Ubiquitous Computing and Ambient Intelligence, pages 185-192. Springer, 2012. 4" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 525, + 545, + 580 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 525, + 545, + 580 + ], + "spans": [ + { + "bbox": [ + 308, + 525, + 545, + 580 + ], + "type": "text", + "content": "[73] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 581, + 545, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 581, + 545, + 613 + ], + "spans": [ + { + "bbox": [ + 308, + 581, + 545, + 613 + ], + "type": "text", + "content": "[74] Michalis Vrigkas, Christophoros Nikou, and Ioannis A Kakadiaris. A review of human activity recognition methods. Frontiers in Robotics and AI, 2:28, 2015. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 613, + 545, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 613, + 545, + 657 + ], + "spans": [ + { + "bbox": [ + 308, + 613, + 545, + 657 + ], + "type": "text", + "content": "[75] Guan Wang, Sijie Cheng, Xianyuan Zhan, Xiangang Li, Sen Song, and Yang Liu. Openchat: Advancing open-source language models with mixed-quality data. arXiv preprint arXiv:2309.11235, 2023. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 658, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 658, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 308, + 658, + 545, + 712 + ], + "type": "text", + "content": "[76] Weronika Wojtak, Flora Ferreira, Paulo Vicente, Luís Louro, Estela Bicho, and Wolfram Erlhagen. A neural integrator model for planning and value-based decision making of a robotics assistant. Neural Computing and Applications, 33: 3737-3756, 2021. 4" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14301" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 715 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "text", + "content": "[77] Peng Xu, Wenqi Shao, Kaipeng Zhang, Peng Gao, Shuo Liu, Meng Lei, Fanqing Meng, Siyuan Huang, Yu Qiao, and Ping Luo. Lvlm-ehub: A comprehensive evaluation benchmark for large vision-language models. arXiv preprint arXiv:2306.09265, 2023. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 128, + 287, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 128, + 287, + 183 + ], + "spans": [ + { + "bbox": [ + 48, + 128, + 287, + 183 + ], + "type": "text", + "content": "[78] Jingkang Yang, Yuhao Dong, Shuai Liu, Bo Li, Ziyue Wang, Chencheng Jiang, Haoran Tan, Jiamu Kang, Yuanhan Zhang, Kaiyang Zhou, et al. Octopus: Embodied vision-language programmer from environmental feedback. arXiv preprint arXiv:2310.08588, 2023. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 183, + 287, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 183, + 287, + 227 + ], + "spans": [ + { + "bbox": [ + 48, + 183, + 287, + 227 + ], + "type": "text", + "content": "[79] Jianwei Yang, Hao Zhang, Feng Li, Xueyan Zou, Chunyuan Li, and Jianfeng Gao. Set-of-mark prompting unleashes extraordinary visual grounding in gpt-4v. arXiv preprint arXiv:2310.11441, 2023. 8" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 228, + 287, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 228, + 287, + 271 + ], + "spans": [ + { + "bbox": [ + 48, + 228, + 287, + 271 + ], + "type": "text", + "content": "[80] Zhengyuan Yang, Linjie Li, Kevin Lin, Jianfeng Wang, Chung-Ching Lin, Zicheng Liu, and Lijuan Wang. The dawn of Imms: Preliminary explorations with gpt-4v (isdiction). arXiv preprint arXiv:2309.17421, 2023. 1, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 272, + 287, + 338 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 272, + 287, + 338 + ], + "spans": [ + { + "bbox": [ + 48, + 272, + 287, + 338 + ], + "type": "text", + "content": "[81] Qinghao Ye, Haiyang Xu, Guohai Xu, Jiabo Ye, Ming Yan, Yiyang Zhou, Junyang Wang, Anwen Hu, Pengcheng Shi, Yaya Shi, Chaoya Jiang, Chenliang Li, Yuanhong Xu, Hehong Chen, Junfeng Tian, Qian Qi, Ji Zhang, and Fei Huang. mplug-owl: Modularization empowers large language models with multimodality, 2023. 1, 2, 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 338, + 287, + 382 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 338, + 287, + 382 + ], + "spans": [ + { + "bbox": [ + 48, + 338, + 287, + 382 + ], + "type": "text", + "content": "[82] Quanzeng You, Hailin Jin, Zhaowen Wang, Chen Fang, and Jiebo Luo. Image captioning with semantic attention. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4651-4659, 2016. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 383, + 287, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 383, + 287, + 426 + ], + "spans": [ + { + "bbox": [ + 48, + 383, + 287, + 426 + ], + "type": "text", + "content": "[83] Yifan Zhang, Congqi Cao, Jian Cheng, and Hanqing Lu. Egogesture: A new dataset and benchmark for egocentric hand gesture recognition. IEEE Transactions on Multimedia, 20(5):1038-1050, 2018. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 426, + 287, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 426, + 287, + 480 + ], + "spans": [ + { + "bbox": [ + 48, + 426, + 287, + 480 + ], + "type": "text", + "content": "[84] Tiancheng Zhao, Tianqi Zhang, Mingwei Zhu, Haozhan Shen, Kyusong Lee, Xiaopeng Lu, and Jianwei Yin. Vl-checklist: Evaluating pre-trained vision-language models with objects, attributes and relations. arXiv preprint arXiv:2207.00221, 2022. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 481, + 287, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 481, + 287, + 525 + ], + "spans": [ + { + "bbox": [ + 48, + 481, + 287, + 525 + ], + "type": "text", + "content": "[85] Zhong-Qiu Zhao, Peng Zheng, Shou-tao Xu, and Xindong Wu. Object detection with deep learning: A review. IEEE transactions on neural networks and learning systems, 30 (11):3212-3232, 2019. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 525, + 287, + 581 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 525, + 287, + 581 + ], + "spans": [ + { + "bbox": [ + 48, + 525, + 287, + 581 + ], + "type": "text", + "content": "[86] Lianmin Zheng, Wei-Lin Chiang, Ying Sheng, Siyuan Zhuang, Zhanghao Wu, Yonghao Zhuang, Zi Lin, Zhuohan Li, Dacheng Li, Eric Xing, et al. Judging llm-as-a-judge with mt-bench and chatbot arena. arXiv preprint arXiv:2306.05685, 2023. 5, 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 581, + 287, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 581, + 287, + 624 + ], + "spans": [ + { + "bbox": [ + 48, + 581, + 287, + 624 + ], + "type": "text", + "content": "[87] Wangchunshu Zhou, Yan Zeng, Shizhe Diao, and Xinsong Zhang. Vlue: A multi-task benchmark for evaluating vision-language models. arXiv preprint arXiv:2205.15237, 2022. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 624, + 287, + 691 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 624, + 287, + 691 + ], + "spans": [ + { + "bbox": [ + 48, + 624, + 287, + 691 + ], + "type": "text", + "content": "[88] Chenchen Zhu, Fanyi Xiao, Andres Alvarado, Yasmine Babaei, Jiabo Hu, Hichem El-Mohri, Sean Culatana, Roshan Sumbaly, and Zhicheng Yan. Egoobjects: A large-scale egocentric dataset for fine-grained object understanding. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 20110-20120, 2023. 2, 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 691, + 287, + 715 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 287, + 715 + ], + "type": "text", + "content": "[89] Deyao Zhu, Jun Chen, Xiaogian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 72, + 545, + 185 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 326, + 72, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 72, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 326, + 72, + 545, + 95 + ], + "type": "text", + "content": "understanding with advanced large language models. arXiv preprint arXiv:2304.10592, 2023. 1, 2, 4" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 96, + 545, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 96, + 545, + 150 + ], + "spans": [ + { + "bbox": [ + 307, + 96, + 545, + 150 + ], + "type": "text", + "content": "[90] Wanrong Zhu, Jack Hessel, Anas Awadalla, Samir Yitzhak Gadre, Jesse Dodge, Alex Fang, Youngjae Yu, Ludwig Schmidt, William Yang Wang, and Yejin Choi. Multimodal c4: An open, billion-scale corpus of images interleaved with text. arXiv preprint arXiv:2304.06939, 2023. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 152, + 545, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 152, + 545, + 185 + ], + "spans": [ + { + "bbox": [ + 307, + 152, + 545, + 185 + ], + "type": "text", + "content": "[91] Zhengxia Zou, Keyan Chen, Zhenwei Shi, Yuhong Guo, and Jieping Ye. Object detection in 20 years: A survey. Proceedings of the IEEE, 2023. 3" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "14302" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Egocentric Whole-Body Motion Capture with FisheyeViT and Diffusion-Based Motion Refinement/90ece1d3-5ea4-4b56-9fe8-5a7d9a524eef_content_list.json b/2024/Egocentric Whole-Body Motion Capture with FisheyeViT and Diffusion-Based Motion Refinement/90ece1d3-5ea4-4b56-9fe8-5a7d9a524eef_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..447bf44331f612628a71598264abec995c649bb3 --- /dev/null +++ b/2024/Egocentric Whole-Body Motion Capture with FisheyeViT and Diffusion-Based Motion Refinement/90ece1d3-5ea4-4b56-9fe8-5a7d9a524eef_content_list.json @@ -0,0 +1,1981 @@ +[ + { + "type": "text", + "text": "Egocentric Whole-Body Motion Capture with FisheyeViT and Diffusion-Based Motion Refinement", + "text_level": 1, + "bbox": [ + 89, + 130, + 880, + 174 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Jian Wang $^{1,4}$", + "bbox": [ + 233, + 202, + 336, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Kripasindhu Sarkar2", + "bbox": [ + 161, + 222, + 323, + 238 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Zhe Cao2", + "bbox": [ + 367, + 203, + 444, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Danhang Tang", + "bbox": [ + 349, + 220, + 472, + 238 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Diogo Luvizon1,4", + "bbox": [ + 470, + 203, + 611, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Thabo Beeler2", + "bbox": [ + 501, + 220, + 616, + 237 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Lingjie Liu3", + "bbox": [ + 635, + 203, + 735, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Christian Theobalt1,4", + "bbox": [ + 643, + 220, + 808, + 237 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1MPI Informatics & Saarland Informatics Campus 2Google 3University of Pennsylvania", + "bbox": [ + 120, + 238, + 849, + 256 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{4}$ Saarbrücken Research Center for Visual Computing, Interaction and Artificial Intelligence", + "bbox": [ + 122, + 256, + 841, + 273 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Project Page: https://jianwang-mpi.github.io/egowholemocap", + "bbox": [ + 281, + 282, + 686, + 297 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/2c7264137600ce24089392af138465890eff1926ff7756b800fdf6649855f9de.jpg", + "image_caption": [ + "Figure 1. From an image sequence captured by a single head-mounted fisheye camera, our method can predict accurate and temporally coherent whole-body motion, including human body and hand poses. The SMPL-X parameters are obtained using inverse kinematics." + ], + "image_footnote": [], + "bbox": [ + 78, + 309, + 205, + 424 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/36b3ea0ad948dd017e6c69274be3bab28c3522b0410a21094784063b16913666.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 205, + 309, + 403, + 424 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/3680f9941ff67bc2aed8921004f6f30027b0736c4f72dd633f1953249cb8c4a6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 406, + 309, + 480, + 424 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/d1184db6933fba68a3c957ead65e29b2421106ab5e8f41b12ea82315382b41a3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 490, + 309, + 614, + 424 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/e229efc3cee3db2f232971165b7e0bf10a4d372ff2372289390bd48b23a8ebe0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 616, + 309, + 813, + 424 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/bda988de8371450c63d11fa9d74da986daaabeb32f388b3626bbe0e66cfca2a7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 815, + 309, + 888, + 424 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 476, + 313, + 492 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this work, we explore egocentric whole-body motion capture using a single fisheye camera, which simultaneously estimates human body and hand motion. This task presents significant challenges due to three factors: the lack of high-quality datasets, fisheye camera distortion, and human body self-occlusion. To address these challenges, we propose a novel approach that leverages FisheyeViT to extract fisheye image features, which are subsequently converted into pixel-aligned 3D heatmap representations for 3D human body pose prediction. For hand tracking, we incorporate dedicated hand detection and hand pose estimation networks for regressing 3D hand poses. Finally, we develop a diffusion-based whole-body motion prior model to refine the estimated whole-body motion while accounting for joint uncertainties. To train these networks, we collect a large synthetic dataset, EgoWholeBody, comprising 840,000 high-quality egocentric images captured across a diverse range of whole-body motion sequences. Quantitative and qualitative evaluations demonstrate the effectiveness of our method in producing high-quality whole-body motion estimates from a single egocentric camera.", + "bbox": [ + 75, + 508, + 472, + 828 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 844, + 209, + 859 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Egocentric 3D human motion estimation using head-mounted devices [47, 54] has garnered significant traction", + "bbox": [ + 75, + 869, + 470, + 902 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "in recent years, driven by its diverse applications in VR/AR. Immersed in a virtual world, we can traverse virtual environments, interact with virtual objects, and even simulate real-world interactions. To fully capture the intricacies of human motion during such interaction, understanding both body and hand movements is essential. While existing egocentric motion capture methods [30, 47, 50-52, 54] focus solely on body motion, neglecting the hands, this work proposes the task of egocentric whole-body motion capture, i.e. simultaneous estimation of the body motion and hand motion from a single head-mounted fisheye camera (shown in Fig. 1). This task is extremely challenging due to three factors: First, the fisheye image introduces significant distortion, making it difficult for existing networks, which are designed for non-distorted images, to extract features. Second, the egocentric camera perspective frequently leads to the occlusion of body parts, such as the feet and hands, further complicating the task of whole-body motion capture. Lastly, large-scale training data with ground truth annotations for both body and hand poses is absent in existing datasets [4, 29, 47, 51, 54].", + "bbox": [ + 496, + 478, + 893, + 796 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this work, we propose a novel egocentric whole-body motion capture method to address the aforementioned challenges. To effectively address fisheye distortion, we propose FisheyeViT for extracting image features, along with a joint regressor employing pixel-aligned 3D heatmap for predicting 3D body poses. Instead of attempting to undis-", + "bbox": [ + 496, + 809, + 893, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "777", + "bbox": [ + 485, + 945, + 509, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "tort the entire fisheye image, which is impractical due to the fisheye lens's large field of view (FOV), we opt to partition the image into smaller patches aligned with a specific FOV range. This approach enables individual patch-level undistortion and seamlessly aligns with the vision transformer architecture that is employed for extracting the complete image feature map. We further propose an egocentric 3D pose regressor utilizing 3D heatmap representations. Unlike the existing approach [52] that projects image features into 3D space through fisheye reprojection functions and regresses 3D heatmaps with V2V networks [33] leading to intricate network learning and high computational complexity—our proposed egocentric pose regressor adopts a simpler approach. It employs deconvolutional layers to obtain pixel-aligned 3D heatmaps. Notably, the voxels in the 3D heatmap directly correspond to pixels in 2D features, subsequently linking to image patches in FisheyeViT. This streamlined approach significantly simplifies network training. Joint locations from the pixel-aligned 3D heatmap are finally transformed with the fisheye camera model to obtain the 3D human body poses. Due to the large size difference between body and hands, we train a hand detection network and a hand pose estimation network to accurately regress 3D hand poses.", + "bbox": [ + 76, + 90, + 472, + 454 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To overcome the challenges posed by self-occlusion and improve the accuracy of pose estimation, we propose a novel method for refining the whole-body motion predictions by incorporating temporal context and a motion prior. Our method learns a whole-body motion prior with the diffusion model [18] from a collection of diverse human motion sequences, capturing intrinsic correlations between hand and body movements. Following this, we extract the joint uncertainties from the pixel-aligned 3D heatmap and utilize them to guide the refinement of the whole-body motion. The joint uncertainties act as indicators of the trustworthiness of the pose regressor's predictions. By conditioning on joints with low uncertainty, our whole-body motion diffusion model selectively refines joints with high uncertainty. This strategy substantially improves the quality of whole-body pose estimations and effectively mitigates the effects of self-occlusion.", + "bbox": [ + 75, + 455, + 472, + 714 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In response to the absence of the egocentric whole-body motion capture datasets, we present EgoWholeBody, a new large-scale high-quality synthetic dataset. This dataset encompasses a wide range of whole-body motions, comprising over 870k frames, which significantly surpasses the size of previous egocentric training datasets. EgoWholeBody could serve as a valuable resource for advancing research in egocentric whole-body motion capture.", + "bbox": [ + 75, + 715, + 468, + 838 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "A thorough evaluation across a range of datasets, including SceneEgo [52], GlobalEgoMocap [50] and $\\mathrm{Mo}^2\\mathrm{Cap}^2$ [54], has demonstrated the remarkable improvements of our method in estimating egocentric whole-body", + "bbox": [ + 76, + 840, + 470, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "motion compared to previous approaches. This substantiates the effectiveness of our approach in addressing the special challenges encountered in egocentric views, including the fisheye distortion and self-occlusion.", + "bbox": [ + 496, + 90, + 890, + 151 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In summary, our key contributions are the following:", + "bbox": [ + 498, + 152, + 849, + 166 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- The first egocentric whole-body motion capture method that predicts accurate and temporarily coherent egocentric body and hand motion;", + "- FisheyeViT for alleviating fisheye camera distortion and pose regressor using pixel-aligned 3D heatmaps for accurate egocentric body pose estimation from a single image;", + "- Uncertainty-aware refinement method based on motion diffusion models for correcting initial pose estimations and predicting plausible motions even under occlusion;", + "- EgoWholeBody, a new high-quality synthetic dataset for egocentric whole-body motion capture." + ], + "bbox": [ + 500, + 167, + 890, + 335 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 345, + 640, + 363 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Egocentric 3D Human Body Pose Estimation. Recently, there has been growing interest in estimating egocentric 3D poses from body-worn cameras. Some methods [21, 25, 31, 35, 59, 60] use front-facing cameras and infer the human body motion from the camera view. However, since the user's body is often unobserved by the camera, these methods fail when the human body is not roaming around. Millerdurai et al. [32] leverage event cameras for estimating egocentric body pose. Other methods [4, 5, 7, 23, 39, 65] use head-mounted down-facing stereo cameras to estimate body poses. However, stereo camera setups introduce extra burdens of weight and energy consumption.", + "bbox": [ + 496, + 372, + 890, + 551 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Xu et al. [54] and Tome et al. [47] introduce the single head-mounted down-facing fisheye camera setup for the egocentric 3D human pose estimation task. Zhang et al. [64] regressed fisheye camera parameters and 3D human pose simultaneously. To address the self-occlusion issue, Park et al. [36] leveraged the temporal information with the spatio-temporal self-attention network, and Liu et al. [30] introduced diffusion model to generate 3D human pose conditioned on egocentric image features. Wang et al. [50] and Liu et al. [28] combined the SLAM and egocentric pose estimation methods to estimate human body poses in the world coordinate. Wang et al. [51] and Liu et al. [29] leverage the synchronized egocentric camera and external cameras to collect large-scale egocentric pose estimation datasets with pseudo-ground truth. Considering the human-scene interaction, Wang et al. [52] estimated the scene geometry from the egocentric camera and constrained the 3D human pose with it.", + "bbox": [ + 496, + 553, + 892, + 825 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "These methods only focus on estimating human body poses while omitting the hand motion, and they still suffer from fisheye camera distortion since they directly put the highly distorted fisheye images into the neural network. Our proposed method can capture whole-body motion and", + "bbox": [ + 496, + 825, + 893, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "778", + "bbox": [ + 486, + 945, + 509, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/0db61e7538dc84f37f270d5fe48c34a96602a8642274b91ef635766c67b4affc.jpg", + "image_caption": [ + "Figure 2. Overview of our whole-body motion capture pipeline. We first use FisheyeViT to undistort the input image and generate image feature tokens (3.1.1). Next, we use a 1D convolutional network to convert the image features to a pixel-aligned 3D heatmap and use soft-argmax and fisheye camera undistortion function to obtain the 3D body joins positions and uncertainty (3.1.2). We further detect the hand location and regress the 3D hand poses from the input image (3.1.3). Finally, the estimated hand motion and human body motion are combined and the uncertainty-aware diffusion model is applied to refine the estimated whole-body motion (3.2)." + ], + "image_footnote": [], + "bbox": [ + 80, + 88, + 890, + 303 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "resolve the fisheye camera distortion issue with the FisheyeViT and pixel-aligned 3D heatmap.", + "bbox": [ + 75, + 398, + 468, + 428 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Whole-Body 3D Pose Estimation. Whole-body 3D pose estimation aims to estimate the 3D human body, face, and hands parameters from input images. This task is crucial for many applications, e.g., modeling human activities and human-scene interactions. Some methods [37, 53] fit the 2D body joints estimated from images with optimization algorithms, while these methods suffer from high computation overhead and can fall into local optima. Some other learning-based methods [6, 9, 15, 27, 40, 45, 66] use the neural network to regress the SMPL-X [37] parameters from input images. For example, ExPose [9] introduced body-driven attention to extract face and hand crops and used a refinement module to regress whole-body pose. OSX [27] proposed a one-stage pipeline for whole-body mesh recovery without separate networks for each part. SMPLer-X [6] propose a foundation model for whole-body pose estimation trained with the large model and big data.", + "bbox": [ + 75, + 429, + 468, + 686 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Though much progress has been made on whole-body pose estimation from an external view, the task from an ego-centric view is still unexplored. In this paper, we introduce the first whole-body 3D pose estimation method from a single egocentric image and also incorporate temporal information with diffusion-based motion refinement.", + "bbox": [ + 75, + 686, + 468, + 777 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Diffusion Models for Pose Estimation. Recently, some methods [8, 10, 16, 17, 19, 42] have effectively applied Denoising Diffusion Probabilistic Models (DDPM) [18] to human pose estimation tasks. Building on the success of motion diffusion models in human pose estimation, many methods have extended this approach to egocentric pose estimation, where the human body is only partially visible from RGB cameras or VR sensors. Zhang et al.'s work [63]", + "bbox": [ + 75, + 779, + 468, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "uses a diffusion model to generate realistic human poses considering scene geometry. AGROL [14] generates body motion based on head and hand 6D pose estimates from a VR headset. EgoEgo [25] estimates head poses from a head-mounted front-facing camera and uses them to generate body poses. EgoHMR [30] extracts image features and uses them as a condition for the diffusion denoising process.", + "bbox": [ + 496, + 398, + 890, + 503 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "However, the aforementioned pose estimation methods train the conditioned diffusion model with image features or IMU signals. This cannot be generalized since the trained network only accepts one specific condition format and is inclined to learn domain-specific distributions of condition features. ZeDO [22] tackles this issue with a zero-shot diffusion-based optimization approach that doesn't require training with 2D-3D or image-3D pairs. Our method leverages the uncertainty value given by the single-frame pose estimation network and refines the initial motion estimation with the uncertainty of each joint. Moreover, different from previous methods that only focus on human body motion, we train a whole-body motion diffusion model to construct the relationship between hand and body motion.", + "bbox": [ + 496, + 503, + 892, + 715 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 500, + 728, + 589, + 743 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we propose a new method for predicting accurate egocentric whole-body poses from egocentric image sequences. An overview of our approach is shown in Fig. 2.", + "bbox": [ + 496, + 753, + 890, + 799 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Single Image Based Egocentric Pose Estimation", + "text_level": 1, + "bbox": [ + 500, + 808, + 890, + 824 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1.1 FisheyeViT", + "text_level": 1, + "bbox": [ + 500, + 830, + 633, + 845 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we introduce FisheyeViT, which is specially designed to alleviate the fisheye distortion issue. Instead of undistorting the entire fisheye image, we extract undistorted", + "bbox": [ + 496, + 854, + 892, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "779", + "bbox": [ + 485, + 945, + 509, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/9012d87f36fd44a5ad44837bd96db815a063baf3c73872f6d559e4d84768d640.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 86, + 88, + 264, + 202 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/a150273fbd21bd7d66a798fb15085301439938063a88b8ed59b7b973e509a2aa.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 267, + 88, + 460, + 204 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/785fbab66b2e9d0cfba8efc7a302c7cbc151a05f21cd9c45d1c6f3ea85e1794c.jpg", + "image_caption": [ + "Figure 3. The detailed illustration of FisheyeViT (Sec. 3.1.1)." + ], + "image_footnote": [], + "bbox": [ + 86, + 205, + 263, + 321 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/3897e648c0b37f699b97766125124aa2e6c52ad507a6c3128e166b975a50c9ed.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 267, + 205, + 460, + 321 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "image patches from the fisheye image and then fit these patches as tokens into the transformer network [13]. To get the undistorted patches, we first warp the fisheye image to a unit semi-sphere, then get the patches with the gnomonic projection (see Fig. 2). The FisheyeViT can be split into five steps, the first four of which are illustrated in Fig. 3.", + "bbox": [ + 75, + 359, + 467, + 450 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Step 1. Given an input image $\\mathbf{I}$ with size $H\\times W$ , we first evenly sample $N\\times N$ patch center points: $\\{\\mathbf{C}_{ij} = (u_i,v_j) = \\left(\\frac{H}{N} (i + \\frac{1}{2}),\\frac{W}{N} (j + \\frac{1}{2})\\right)|i,j\\in 0,\\dots,N - 1\\}$ . Then, the patch center points $\\mathbf{C}_{ij}$ are projected onto a unit sphere with the fisheye reprojection function: $\\mathbf{P}_{ij}^{c} = (x_{ij}^{c},y_{ij}^{c},z_{ij}^{c}) = \\mathcal{P}^{-1}(u_i,v_j,1)$ . The fisheye camera model is described in Sec. 8 of the supplementary material. Given a point $\\mathbf{P}_{ij}^{c}$ on the unit sphere, the tangent plane $\\mathbf{T}_{ij}$ that passes through the point is defined by the normal vector $\\mathbf{v}_{ij}^{c} = (x_{ij}^{c},y_{ij}^{c},z_{ij}^{c})$ . In the following steps, we implement the gnomonic projection by sampling grid points in the plane and projecting them back onto the fisheye image.", + "bbox": [ + 76, + 450, + 468, + 633 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Step 2. In this step, we determine the orientation of the grid points in the tangent plane, ensuring that the grid points from different tangent planes $\\mathbf{T}_{ij}$ have the same orientation when projected back onto the fisheye image. To achieve this, we select a 2D point $\\mathbf{U}_{ij} = (u_i + d,v_j)$ in the fisheye image space that is $d$ pixels to the right of the patch center point and project it to the unit sphere using the fisheye reprojection function: $\\mathbf{P}_{ij}^{u} = (x_{ij}^{u},y_{ij}^{u},z_{ij}^{u}) = \\mathcal{P}^{-1}(u_i + d,v_j,1)$ . We then calculate the intersection point $\\mathbf{P}_{ij}^{x}$ between the vector $\\mathbf{v}_{ij}^{u} = (x_{ij}^{u},y_{ij}^{u},z_{ij}^{u})$ that is passing the origin and the tangent plane $\\mathbf{T}_{ij}$ :", + "bbox": [ + 75, + 633, + 468, + 801 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {P} _ {i j} ^ {x} = \\frac {\\left\\langle \\mathbf {P} _ {i j} ^ {c} , \\mathbf {v} _ {i j} ^ {c} \\right\\rangle}{\\left\\langle \\mathbf {v} _ {i j} ^ {u} , \\mathbf {v} _ {i j} ^ {c} \\right\\rangle} \\mathbf {v} _ {i j} ^ {u} = \\frac {1}{\\left\\langle \\mathbf {v} _ {i j} ^ {u} , \\mathbf {v} _ {i j} ^ {c} \\right\\rangle} \\mathbf {v} _ {i j} ^ {u}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 140, + 809, + 468, + 849 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\langle \\cdot ,\\cdot \\rangle$ denotes the inner product.", + "bbox": [ + 76, + 854, + 328, + 869 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Step 3. Based on the center point $\\mathbf{P}_{ij}^{c}$ and intersection point $\\mathbf{P}_{ij}^{x}$ on the tangent plane $\\mathbf{T}_{ij}$ , we build a coordinate", + "bbox": [ + 76, + 869, + 468, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "system with the $x$ axis: $\\mathbf{v}_{ij}^{x} = \\mathrm{Norm}(\\mathbf{P}_{ij}^{x} - \\mathbf{P}_{ij}^{c})$ , the $z$ axis: $\\mathbf{v}_{ij}^{z} = \\mathrm{Norm}(\\mathbf{v}_{ij}^{c})$ and the $y$ axis: $\\mathbf{v}_{ij}^{y} = \\mathbf{v}_{ij}^{x} \\times \\mathbf{v}_{ij}^{x}$ where Norm denotes the normalize operation. We grid-sample $M \\times M$ points in a $l \\times l$ square on the $x-y$ plane:", + "bbox": [ + 498, + 90, + 890, + 152 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\left\\{\\mathbf {P} _ {i j} ^ {m n} = \\mathbf {P} _ {i j} ^ {c} + \\left(l \\frac {m}{M} \\mathbf {v} _ {i j} ^ {x}, l \\frac {n}{M} \\mathbf {v} _ {i j} ^ {y}\\right) \\right\\} \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 576, + 162, + 890, + 189 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $m,n\\in -\\frac{1}{2} (M - 1),\\ldots , - \\frac{3}{2}, - \\frac{1}{2},\\frac{1}{2},\\frac{3}{2},\\ldots \\frac{1}{2} (M - 1)$", + "bbox": [ + 498, + 200, + 890, + 218 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Step 4. The points $\\mathbf{P}_{ij}^{mn}$ are projected back to the fisheye image with the fisheye projection function: $\\mathbf{C}_{ij}^{mn} = \\mathcal{P}(\\mathbf{P}_{ij}^{mn})$ . We then apply bilinear sampling to obtain the colors at points $\\mathbf{C}_{ij}^{mn}$ of the input image $\\mathbf{I}$ , yielding the undistorted image patch $\\mathbf{I}_{ij}^{\\mathrm{undis}}$ . Please also see the supplementary video for a visual demonstration of undistorted image patches and their movement on the fisheye image.", + "bbox": [ + 498, + 218, + 890, + 325 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Step 5. The image patches $\\{\\mathbf{I}_{ij}^{\\mathrm{undis}}\\}$ are sent to a ViT transformer network [13] to obtain the feature tokens $\\{\\mathbf{F}_{ij}\\}$ . The feature token is further reshaped in $i\\times j$ matrix and obtain the image feature $\\mathbf{F}$ . In the FisheyeViT, we empirically chose $N = 16$ ; $M = 16$ ; $d = 8$ ; $l = 0.2m$ given the image size $H = W = 256$ .", + "bbox": [ + 496, + 325, + 890, + 415 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Note that $\\mathbf{C}_{ij}^{mn}$ is independent of the image $\\mathbf{I}$ . This means that, given a fixed fisheye camera model, we can precompute $\\mathbf{C}_{ij}^{mn}$ for all combinations of $m,n$ and $i,j$ in advance. This significantly speeds up both the training and evaluation processes. Furthermore, the number and dimensions of image patches $\\{\\mathbf{I}_{ij}^{\\mathrm{undis}}\\}$ match exactly with those in the traditional ViT network. This compatibility allows us to finetune existing ViT networks on our egocentric datasets. Our sampling strategy ensures that each image patch $\\mathbf{I}_{ij}^{\\mathrm{undis}}$ corresponds to the same FOV range in the fisheye camera. In our ablation study in Sec. 5.3, we show that FisheyeViT enhances the performance of the pose estimation network when applied to egocentric fisheye images.", + "bbox": [ + 496, + 416, + 892, + 613 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1.2 Pose Regressor with Pixel-Aligned 3D Heatmap", + "text_level": 1, + "bbox": [ + 498, + 633, + 883, + 648 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "After collecting image features with FisheyeViT, we utilize a 3D heatmap-based network to estimate the body poses. The existing 3D heatmap-based pose regressors [34, 44] are designed for the weak-perspective cameras and predict the 3D heatmap in $xyz$ space. Directly applying these regressors will result in misalignment between 3D heatmap features in $xyz$ space and 2D image features in the fisheye image space. Therefore, we introduce a novel egocentric pose regressor that relies on the pixel-aligned 3D heatmap, tailored to address the needs of fisheye cameras. The idea is to regress the 3D heatmap in $uvd$ space rather than traditional $xyz$ space, where $uv$ corresponds to the fisheye image $uv$ space. Specifically, given a feature map $\\mathbf{F} \\in \\mathbb{R}^{C \\times N \\times N}$ , where $C$ is the channel number, $N$ is feature map height and width, we firstly use two deconvolutional layers to convert the feature map $\\mathbf{F}$ into shape", + "bbox": [ + 496, + 657, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "780", + "bbox": [ + 486, + 945, + 509, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "$(D_h\\times J,H_h,W_h)$ , and further reshape it to pixel-aligned 3D heatmap $\\mathbf{H}\\in \\mathbb{R}^{J\\times D_h\\times H_h\\times W_h}$ , where $J$ is the joint number and $D_{h},H_{h},W_{h}$ is the 3D heatmap depth, height and width. The illustration of pixel-aligned 3D heatmap is shown in Fig. 2. Next, we obtain the max-value positions $\\tilde{\\mathbf{J}}_b = \\{(u_i,v_i,d_i)\\mid i\\in 0,1,2,\\dots,J\\}$ from $\\mathbf{H}$ by the differentiable soft-argmax operation [44]. Here, we note that $u_{i}$ and $v_{i}$ correspond to the uv-coordinate of the 3D body joint projected in the fisheye image space, and $d_{i}$ denotes the distance of the joint to the fisheye camera. Finally, the 3D body joints $\\hat{\\mathbf{J}}_b = \\{(x_i,y_i,z_i)\\mid i\\in 0,1,2,\\dots,J\\}$ are recovered with the fisheye reprojection function: $(x_{i},y_{i},z_{i}) =$ $\\mathcal{P}^{-1}(u_i,v_i,d_i)$ . The predicted body pose $\\hat{\\mathbf{J}}_b$ is finally compared with the ground truth body pose $\\mathbf{J}_b$ with the MSE loss. By first regressing 3D body poses in uvd space and then reproducing it, we ensure that the 3D heatmap is pixelaligned with the end-to-end training.", + "bbox": [ + 76, + 90, + 472, + 347 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "With the pixel-aligned heatmap, our proposed 3D pose regressor solves problems in all three types of previous egocentric joint regressors. First, $\\mathrm{Mo}^2\\mathrm{Cap}^2$ [54] employs separate networks to predict 2D joint positions and joint distances. However, this method can yield unrealistic joint estimations because small errors in 2D joints can result in large errors in 3D joints due to the projection effect. Second, $xR$ -egopose [47] and EgoHMR [30] directly regress the 3D joint positions. However, this method is agnostic to the fisheye camera parameters, making it suitable only for a specific camera configuration (e.g., camera parameters, head-mounted position, etc.). Third, SceneEgo [52] projects 2D features into 3D voxel space and uses a V2V network to regress 3D poses. Because of these, the SceneEgo method suffers from low accuracy and large computation overhead. Different from previous methods, our pose regressor with pixel-aligned 3D heatmap is versatile and efficient since it directly estimates 3D joints while also incorporating an explicitly parametrized fisheye camera model. Moreover, it can preserve the uncertainty of the estimated joints, which will be used in our uncertainty-aware motion refinement method (Sec. 3.2.2). Detailed comparison with other pose prediction heads is shown in Table 3.", + "bbox": [ + 76, + 349, + 472, + 696 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.1.3 Egocentric Hand Pose Estimation", + "text_level": 1, + "bbox": [ + 76, + 722, + 366, + 738 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this section, we first train a network to detect hand pose locations and then train a 3D hand pose estimation network to regress 3D hand poses. Then, we describe how to integrate the estimated hand and body poses.", + "bbox": [ + 76, + 748, + 468, + 808 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Hand Detection. Given an input image $\\mathbf{I}$ , we finetune the HRNet [49] network to regress the 2D hand poses of left hand $\\mathbf{J}_{lh}^{2d}$ and right hand $\\mathbf{J}_{rh}^{2d}$ . From the hand poses, we obtain the center point of left hand $\\mathbf{C}_{lh}$ and right hand $\\mathbf{C}_{rh}$ along with the bounding box sizes, $d_{lh}$ and $d_{rh}$ . We use our approach described in Sec. 3.1.1 to compute undistorted", + "bbox": [ + 76, + 810, + 470, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "image patches of left $\\mathbf{I}_{lh}$ and right hands $\\mathbf{I}_{rh}$ .", + "bbox": [ + 498, + 90, + 799, + 106 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Hand Pose Estimation. Given the cropped image $\\mathbf{I}_{lh}$ or $\\mathbf{I}_{rh}$ , we regress the 3D hand poses $\\hat{\\mathbf{J}}_{lh}^{loc}$ and $\\hat{\\mathbf{J}}_{rh}^{loc}$ with the Hand4Whole [34] network, which is fine-tuned on our Ego-FullBody dataset.", + "bbox": [ + 498, + 106, + 890, + 165 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Integration of Body and Hand Positions. It is not straightforward to integrate the hand poses with the body pose in the egocentric camera view primarily due to the fisheye camera's perspective effects. Take the left hand as an example. Following Step 3 in Sec. 3.1.1, we establish a local coordinate system on the tangent plane of the left-hand image with XYZ axes as follows: $x:\\mathbf{v}_{lh}^{x};y:\\mathbf{v}_{lh}^{y};z:\\mathbf{v}_{lh}^{z}$ . We define a rotation matrix, denoted as $\\mathbf{R}$ , that represents the transformation between the root coordinate system and the local coordinate system on the tangent plane. The estimated hand pose is first rotated with the rotation matrix $\\hat{\\mathbf{J}}_{lh} = \\mathbf{R}\\hat{\\mathbf{J}}_{lh}^{loc}$ and then translated to align the wrist location of the human body. This same process is also applied to the right hand to get the right hand pose $\\hat{\\mathbf{J}}_{rh}$ . The whole-body joints $\\hat{\\mathbf{J}}$ are obtained by combining $\\hat{\\mathbf{J}}_b$ , $\\hat{\\mathbf{J}}_{lh}$ , and $\\hat{\\mathbf{J}}_{rh}$ . The uncertainty of whole-body joints $\\hat{\\mathbf{U}}$ is also obtained from the maximal value of the 3D heatmap in pose estimation modules.", + "bbox": [ + 496, + 167, + 892, + 422 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2. Diffusion-Based Motion Refinement", + "text_level": 1, + "bbox": [ + 498, + 431, + 812, + 446 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We notice that the single-frame estimations in Sec. 3.1 suffer from inaccuracies and temporal instabilities. In this section, we propose a diffusion-based motion refinement method to tackle this problem. We first learn the whole-body motion prior with the motion diffusion model in Sec. 3.2.1 and then introduce an uncertainty-aware zero-shot motion refinement method in Sec. 3.2.2.", + "bbox": [ + 496, + 454, + 890, + 559 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2.1 Whole-Body Motion Diffusion Model", + "text_level": 1, + "bbox": [ + 498, + 578, + 812, + 593 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We follow DDPM [18] as our diffusion approach to capture the whole-body motion prior $q(\\mathbf{x})$ . DDPM learns a distribution of whole-body motion $\\mathbf{x}$ through a forward diffusion process and an inverse denoising process. The forward diffusion process is a Markov process of adding Gaussian noise over $t \\in \\{0, 1, \\dots, T - 1\\}$ steps:", + "bbox": [ + 496, + 601, + 890, + 693 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nq \\left(\\mathbf {x} _ {t} \\mid \\mathbf {x} _ {t - 1}\\right) = \\mathcal {N} \\left(\\sqrt {\\alpha_ {t}} \\mathbf {x} _ {t - 1}, (1 - \\alpha_ {t}) I\\right) \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 563, + 700, + 890, + 718 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathbf{x}_t$ denotes the whole-body motion sequence at step $t$ , the variance $(1 - \\alpha_{t}) \\in (0,1]$ denotes a constant hyperparameter increases with $t$ .", + "bbox": [ + 496, + 724, + 890, + 770 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The inverse process uses a denoising network $D(\\cdot)$ to remove the added Gaussian noise at each time step $t$ . Here we use the transformer-based framework in EDGE [48] as the motion-denoising network $D(\\cdot)$ . We follow Ramesh et al.'s work [38] to make the network predict the original signal itself, i.e. $\\hat{\\mathbf{x}}_0 = D(\\mathbf{x}_t,t)$ and train it with the simple objective [18]:", + "bbox": [ + 496, + 771, + 890, + 876 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {s i m p l e}} = E _ {\\mathbf {x} _ {0} \\sim q (\\mathbf {x} _ {0}), t \\sim [ 1, T ]} \\left[ \\| \\mathbf {x} _ {0} - D (\\mathbf {x} _ {t}, t) \\| _ {2} ^ {2} \\right] \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 526, + 883, + 890, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "781", + "bbox": [ + 485, + 945, + 509, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2.2 Uncertainty-Aware Motion Refinement", + "text_level": 1, + "bbox": [ + 76, + 90, + 403, + 104 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Given the learned whole-body motion prior, we leverage the uncertainty value for each pose to guide the diffusion denoising process with the classifier-guided diffusion sampling [12]. Given an initial sequence of whole-body pose estimation $\\mathbf{x}_e = \\{\\hat{\\mathbf{J}}_i\\}$ and the uncertainty value for each pose $\\mathbf{u} = \\{\\hat{\\mathbf{U}}_i\\}$ , where $i$ denotes the $i$ th pose in the sequence, we keep the joints with low uncertainty but use the diffusion model to generate joints with high uncertainty conditioned on the low-uncertainty joints. Specifically, in the $t$ th sampling step of the diffusion process, the denoising network predicts $\\hat{\\mathbf{x}}_0 = D(\\mathbf{x}_t,t)$ , which is noised back to $\\mathbf{x}_{t - 1}$ by sampling from the Gaussian distribution:", + "bbox": [ + 75, + 114, + 472, + 297 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {x} _ {t - 1} \\sim \\mathcal {N} \\left(\\hat {\\mathbf {x}} _ {0} + \\mathbf {w} \\left(\\mathbf {x} _ {e} - \\hat {\\mathbf {x}} _ {0}\\right), \\Sigma_ {t}\\right) \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 156, + 306, + 468, + 324 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\Sigma_{t}$ is a scheduled Gaussian distribution in DDPM [18] and $\\mathbf{w}$ controls the weight of a specific joint between the predicted motion $\\hat{\\mathbf{x}}_0$ and the estimated motion $\\mathbf{x}_e$ . Generally, we expect $\\mathbf{w} \\rightarrow \\overrightarrow{0}$ when $t \\rightarrow 0$ such that the temporal stability is guaranteed through the generation of the denoising process, and $\\mathbf{w} \\rightarrow \\overrightarrow{1}$ when $t \\rightarrow T$ such that the denoising process is initialized by the estimated motion $\\mathbf{x}_e$ . We also expect that $w_{ij} = \\mathbf{w}[i][j]$ , which is the weight of $j$ th joints in the $i$ th pose, is smaller when the uncertainty value $u_{ij} = \\mathbf{u}[i][j]$ of the $j$ th joints in the $i$ th pose is large. Based on this requirement, we design $\\mathbf{w}$ as:", + "bbox": [ + 75, + 332, + 472, + 503 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {w} = 1 / \\left(1 + e ^ {- k (t - T \\mathbf {u})}\\right) \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 184, + 511, + 468, + 537 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $T$ is the overall diffusion steps, $k$ is a hyperparameter which is empirically set to 0.1. From the experimental results in Sec. 5, we demonstrate the effectiveness of uncertain-aware motion refinement and our uncertainty-guided diffusion sampling strategy.", + "bbox": [ + 75, + 545, + 468, + 623 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. EgoWholeBody Dataset", + "text_level": 1, + "bbox": [ + 76, + 633, + 302, + 650 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this section, we introduce EgoWholeBody, a large-scale high-quality synthetic dataset built for the task of egocentric whole-body motion capture. The EgoWholeBody dataset is organized into two sections. The first part, containing over 700k frames, is rendered with 14 different rigged Renderpeople [3] models driven by 2367 Mixamo [2] motion sequences. The second part focuses on hand motions and contains 170k frames with the SMPL-X model. This data is constructed from 24 different shapes and textures, driven by 262 motion sequences selected from the GRAB [46] and TCDHandMocap dataset [20]. We also created synthetic test sequences, which include 133k images rendered with 3 Renderpeople models and Mixamo motions.", + "bbox": [ + 75, + 659, + 468, + 854 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "During the rendering process, we first attach a virtual fisheye camera to the forehead of human body models and render the images, semantic labels, and depth map with", + "bbox": [ + 75, + 854, + 470, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Blender [1]. Our dataset is larger and more diverse than previous egocentric training datasets—see Sec. 10 in the supplementary material for a detailed comparison.", + "bbox": [ + 496, + 90, + 892, + 137 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 498, + 151, + 633, + 167 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1. Datasets and Evaluation Metrics", + "text_level": 1, + "bbox": [ + 498, + 176, + 785, + 191 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Training Datasets. To train our body pose estimation module (Sec. 3.1.1 and Sec. 3.1.2), we use our EgoWholeBody dataset and the EgoPW dataset [50]. Additionally, the EgoWholeBody dataset is used to train the hand pose estimation module in Sec. 3.1.3. For training the whole-body diffusion model (Sec. 3.2), we utilize a combined motion capture dataset that includes EgoBody [62], Mixamo [2], TCDHandMocap dataset [20] and GRAB dataset [46].", + "bbox": [ + 496, + 199, + 892, + 321 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation Datasets. In our experiment, we evaluate our methods on four datasets: the GlobalEgoMocap test datasets [50], the $\\mathrm{Mo}^2\\mathrm{Cap}^2$ test dataset [54], the SceneEgo test dataset [52] and out EgoWholeBody test dataset. The details of the datasets are shown in Sec. 12 of supplementary materials. Note that evaluating whole-body poses requires accurate annotations for human hands, which is absent in real-world datasets. To resolve the issue, we request the multi-view videos of the SceneEgo test dataset [52] from the authors and use a multi-view motion capture system to obtain the hand motion. The hand pose annotations will be made publicly available.", + "bbox": [ + 496, + 321, + 893, + 503 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation Metrics. To evaluate the precision of human body poses on the SceneEgo test dataset [52], we use MPJPE and PA-MPJPE. For the GlobalEgoMocap test dataset [50] and $\\mathrm{Mo}^2\\mathrm{Cap}^2$ test dataset [54], where egocentric camera poses are unavailable, we evaluate PA-MPJPE and BA-MPJPE. For hand pose accuracy, we align the predicted and ground truth hand poses at the root position, followed by computing MPJPE and PA-MPJPE. Detailed explanations of these metrics are in Sec. 11 of the supplementary materials. All reported metrics are in millimeters.", + "bbox": [ + 496, + 503, + 893, + 655 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.2. Comparisons on Whole-Body Pose Estimation", + "text_level": 1, + "bbox": [ + 498, + 665, + 888, + 681 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For a fair comparison with existing methods focusing solely on body or hand pose, we split our evaluation into two parts, reporting results of body poses in Table 1 and hand pose in Table 2. We first compare the accuracy of the human body poses with state-of-the-art methods, including EgoPW [51] and SceneEgo [52], on EgoWholeBody and SceneEgo [52] test datasets. The comparison with more previous methods [47, 50, 54] and on more evaluation datasets [50, 54] are shown in Sec. 7 of the supplementary materials. Since our motion refinement method incorporates random Gaussian noise, we generate five samples and calculate the average MPJPE values. The standard deviation is low ( $< 0.01\\mathrm{mm}$ ) and is discussed in Sec. 13 of supplementary materials. Results are presented in Table 1, where our single-frame re", + "bbox": [ + 496, + 688, + 893, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "782", + "bbox": [ + 486, + 945, + 509, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/2b36abfc7293d272285e3cd4a5923fe4d3975b24615caa51d348f4182cec6dee.jpg", + "image_caption": [ + "Input" + ], + "image_footnote": [], + "bbox": [ + 83, + 88, + 166, + 212 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/7296f8f261090bbba4cc565a89c0a1263b78b4e35d582f0d76b7be038e6db2f8.jpg", + "image_caption": [ + "EgoPW" + ], + "image_footnote": [], + "bbox": [ + 173, + 93, + 227, + 210 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/fd3208ba132e94b485215577d61f0e4ae8c98e8d844925559f66b5d3f0706c32.jpg", + "image_caption": [ + "SceneEgo" + ], + "image_footnote": [], + "bbox": [ + 254, + 90, + 305, + 210 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/0416137e264b082666a001f0d877bdecb1233463c15eece3a3b594c8ffb40dc2.jpg", + "image_caption": [ + "Ours-Single" + ], + "image_footnote": [], + "bbox": [ + 333, + 90, + 387, + 210 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d116b8c08e886af64ea3404515370357c512628eee4cf8c1bec4e8f6b123554b.jpg", + "image_caption": [ + "Ours-Refined" + ], + "image_footnote": [], + "bbox": [ + 410, + 90, + 462, + 210 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/eec0ee177d1a6d86bdbe3f5c7fd62a8b1c16990dd0ea4d339fd8768b2a0881fc.jpg", + "image_caption": [ + "Input" + ], + "image_footnote": [], + "bbox": [ + 485, + 88, + 568, + 212 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/78bc9457063ba21f824a921e8d7d897ae814735a0288d48f7fd304888c07b4d0.jpg", + "image_caption": [ + "EgoPW" + ], + "image_footnote": [], + "bbox": [ + 576, + 90, + 632, + 210 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/2db5bb07a47b9f653267dd865fcf90623f3e95809148e7cc681c9f05a7829dea.jpg", + "image_caption": [ + "SceneEgo" + ], + "image_footnote": [], + "bbox": [ + 658, + 90, + 715, + 210 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/1f6fc40d8fa32c0cf3eb18c47bf63e527bb44f52c316f76a32a6410d14e1f311.jpg", + "image_caption": [ + "Ours-Single" + ], + "image_footnote": [], + "bbox": [ + 736, + 90, + 800, + 210 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/0aef160c92c409446aaee3cae33e11765fb8998fec395ea53b2961ef82f61071.jpg", + "image_caption": [ + "Ours-Refined" + ], + "image_footnote": [], + "bbox": [ + 816, + 90, + 880, + 210 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a705b4d478e7ae29a419d8aef00da1809a863112089c2bd25dd790aeac296a91.jpg", + "image_caption": [ + "Figure 4. Qualitative comparison on human body pose estimations between our methods and the state-of-the-art egocentric pose estimation methods on in-the-studio (left column) and in-the-wild scenes (right column). The red skeleton is the ground truth while the green skeleton is the predicted pose. Our methods predict more accurate body poses compared with EgoPW [51] and SceneEgo [52].", + "Input", + "Figure 5. Qualitative comparison on human hand pose estimations between our methods and the state-of-the-art third-view pose estimation methods. Our single-view and refined hand poses are more accurate than the poses from Hand4Whole [34] method. The red skeleton is the ground truth while the green skeleton is the predicted pose." + ], + "image_footnote": [], + "bbox": [ + 83, + 277, + 169, + 407 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/2511ddddfbb45f663922c3d6275ba93aaedca7fedc5dd82520a337637611a853.jpg", + "image_caption": [ + "Hand Image" + ], + "image_footnote": [], + "bbox": [ + 178, + 277, + 279, + 407 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a3e0d8f6c237c8e18267565ef03ef52150f8a31fa73290579d211711de55b529.jpg", + "image_caption": [ + "Hand4Whole" + ], + "image_footnote": [], + "bbox": [ + 287, + 281, + 352, + 406 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d1567b2df1aaabf7c88b08f6808222173c19e81a7842c75318e92e5a5153e0da.jpg", + "image_caption": [ + "Ours-Single" + ], + "image_footnote": [], + "bbox": [ + 372, + 281, + 429, + 406 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/9d14d446c812f3ba79dff90426f0c487f1622cead460c6f80a98f5546f25a5a9.jpg", + "image_caption": [ + "Ours-Refined" + ], + "image_footnote": [], + "bbox": [ + 454, + 280, + 509, + 406 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/868694d3382e8384e9bc0acb6559028c3eeebd996d190f073c96360431b5b60f.jpg", + "image_caption": [ + "Rrnnnne", + "Hand Image", + "Hand4Whole" + ], + "image_footnote": [], + "bbox": [ + 555, + 279, + 640, + 407 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/0aab8a411ce9f29fe58d8547125f2a1b5b383cff52ba3a7078c372552d76111f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 648, + 281, + 715, + 406 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/973eb23b413d2231b829f02d5a8402bd11582fa3bae0b981c051d6b8702eff50.jpg", + "image_caption": [ + "Ours-Single" + ], + "image_footnote": [], + "bbox": [ + 723, + 280, + 797, + 406 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ebd9da3b584bfa078c7c9111266a0314999a58041405dabbcfef33dd46c35036.jpg", + "image_caption": [ + "Ours-Refined" + ], + "image_footnote": [], + "bbox": [ + 808, + 280, + 880, + 407 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "sults are labeled as \"Ours-Single\" and our refinement results are labeled as \"Ours-Refined\". Our single-frame body pose estimation method outperforms all previous methods by a large margin. Our diffusion-based motion refinement method can further improve the accuracy of body poses estimated by the single-frame methods.", + "bbox": [ + 75, + 489, + 468, + 580 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Note that previous methods [47, 50-52, 54] use training datasets different from each other. For a fair comparison, we re-train previous methods with our training datasets in Sec. 5.1 and show the results with “*” in Table 1. This retraining led to significant improvements across all previous methods, demonstrating our dataset's broad applicability. However, these methods still underperformed compared to ours, highlighting our approach's superiority.", + "bbox": [ + 75, + 580, + 468, + 702 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To evaluate the accuracy of our hand pose estimation method, we first crop the hand images with the hand detection method in Sec. 3.1.3. Then we show the results of our single-frame hand pose estimation (labeled as \"Ours-Single\") and whole-body motion refinement methods (labeled as \"Ours-Refined\") in Table 2. Our single-frame hand pose estimation method outperforms the state-of-the-art method Hands4Whole [34], demonstrating the effectiveness of training the network on our EgoWholeBody dataset. Our whole-body motion refinement method can also enhance the accuracy of hand motion.", + "bbox": [ + 75, + 703, + 468, + 868 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "For a qualitative comparison, we compare the body and hand poses of our method with existing methods on the", + "bbox": [ + 76, + 869, + 468, + 900 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/1064052aec9da9a3ec86e7da2a1eb808e52d1671886366acc035472711ae59a1.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodMPJPEPA-MPJPE
SceneEgo test dataset [52]
EgoPW [51]189.6105.3
SceneEgo [52]118.592.75
EgoPW* [51]90.9664.33
SceneEgo* [52]89.0670.10
Ours-Single64.1950.06
Ours-Refined57.5946.55
EgoWholeBody test dataset
EgoPW* [51]84.2163.02
SceneEgo* [52]87.5769.46
Ours-Single66.2843.14
Ours-Refined60.3240.35
", + "bbox": [ + 504, + 487, + 888, + 676 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 1. Egocentric human body pose accuracy of our method on SceneEgo test datasets and EgoWholeBody test dataset. Our method outperforms all previous state-of-the-art methods. * denotes the method trained with the datasets in Sec. 5.1.", + "bbox": [ + 498, + 686, + 890, + 742 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "SceneEgo dataset and the in-the-wild EgoPW [51] evaluation sequences. The results are shown in Fig. 4 and Fig. 5, showing that our method can predict high-quality whole-body poses from an egocentric camera. Please refer to our supplementary video for more qualitative evaluation results.", + "bbox": [ + 498, + 758, + 890, + 835 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.3. Ablation Study", + "text_level": 1, + "bbox": [ + 500, + 845, + 651, + 862 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "EgoWholeBody Dataset. Compared to existing egocentric datasets, our EgoWholeBody dataset contains diverse body", + "bbox": [ + 500, + 869, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "783", + "bbox": [ + 485, + 945, + 509, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/fa1a92a804b3af566e94da576c291e5a2311a813afb72d94b28a4958b5b182ca.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodMPJPEPA-MPJPE
SceneEgo test dataset [52]
Hand4Whole [34]49.6613.85
Ours-Single23.639.59
Ours-Refined19.379.05
EgoWholeBody test dataset
Hand4Whole [34]52.8535.04
Ours-Single33.1019.68
Ours-Refined28.2914.51
", + "bbox": [ + 81, + 88, + 465, + 220 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/6ee6f25a3c413d7716b537137e4bf74163cfd84641f9784c1945624c873bc71f.jpg", + "table_caption": [ + "Table 2. Egocentric hand pose accuracy of our method. Our method outperforms the Hand4Whole [34] on both datasets." + ], + "table_footnote": [], + "table_body": "
MethodMPJPEPA-MPJPE
Body Pose Results
w/o EgoWholeBody75.1058.62
w/o FisheyeViT67.3653.44
w/ Mo2Cap2 [54] head87.4765.10
w/ xR-egopose [47] head116.595.78
w/ SceneEgo [52] head77.7362.69
Ours-Single64.1950.06
w/ GlobalEgoMocap†69.8356.73
w/o uncert. guidance†62.1648.40
Only body diffusion58.9547.03
Ours-Refined†57.5946.55
Hand Pose Results
Only hand diffusion21.699.24
Ours-Refined19.379.05
", + "bbox": [ + 81, + 266, + 465, + 482 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 3. Ablation Study on SceneEgo test dataset [52]. ${}^{ \\dagger }$ denotes the temporal-based method.", + "bbox": [ + 76, + 493, + 467, + 522 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "and hand motions, larger quantity of images, and higher image quality. We show this by training our body pose estimation network without our dataset, using the $\\mathrm{Mo}^2\\mathrm{Cap}^2$ [54] and EgoPW [51] training dataset. The results, labeled as \"w/o EgoWholeBody\" in Table 3, show that performance without the EgoWholeBody dataset is inferior to our proposed method. This highlights that training with our EgoW-holeBody dataset enhances the performance of the pose estimation method. We also compare this result with existing methods on the SceneEgo test set (Table 1). Trained without EgoWholeBody, our approach still outperforms previous methods, showing the effectiveness of our method.", + "bbox": [ + 75, + 537, + 467, + 717 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "FisheyeViT and Pose Regressor with Pixel-Aligned 3D Heatmap. To assess the individual contributions of FisheyeViT and the pixel-aligned 3D heatmap in our single-frame pose estimation pipeline, we perform experiments to measure their impact on the overall performance. First, we substitute the FisheyeViT module in our single-frame pose estimation method to ViT [13]. The result is shown in \"w/o FisheyeViT\" in Table 3 and it is worse than our full method. This demonstrates the effectiveness of FisheyeViT in addressing fisheye distortion and feature extraction.", + "bbox": [ + 75, + 719, + 467, + 868 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Next, we analyze the performance of the single-frame pose estimation network when substituting our pose regress-", + "bbox": [ + 76, + 869, + 467, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "sor based on pixel-aligned 3D heatmap with the pose estimation heads of previous works [47, 52, 54]. The results of the three experiments, labeled as \"w/ Mo²Cap² head\", \"w/ xR-egopose head\" and \"w/ SceneEgo head\", show a performance drop compared to our full method. This emphasizes the crucial role of the pixel-aligned 3D heatmap in accurately estimating egocentric 3D body joint positions.", + "bbox": [ + 496, + 90, + 890, + 196 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Diffusion-based Motion Refinement. We assess the effectiveness of our diffusion-based motion refinement with the following experiments: First, we compare the performance of our diffusion-based motion refinement with GlobalEgoMocap [50] by applying the GlobalEgoMocap optimizer on the single-frame body pose estimation results. The result, labeled as \"w GlobalEgoMocap\" in Table 3, indicates that our refinement method outperforms GlobalEgoMocap.", + "bbox": [ + 496, + 196, + 890, + 316 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Second, we remove the uncertainty-aware guidance in the motion refinement. Instead, we use fixed Gaussian denoising steps to refine the motion. The result \"w/o uncert. guidance\" in Table 3, shows that our uncertainty-aware refinement method performs better. Our approach relies on the uncertainty values for each joint, using low-uncertainty joints to guide the generation of high-uncertainty joints. This helps reduce errors in joint predictions caused by egocentric self-occlusion, leading to improved results.", + "bbox": [ + 496, + 316, + 890, + 452 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Third, we replace our whole-body motion diffusion model with the separate human body and left/right-hand diffusion models and show the accuracy of refined body and hand motion in \"Only body diffusion\" and \"Only hand diffusion\" in Table 3. From the results, we observe improvements in the accuracy of motion refined by our whole-body diffusion method, proving that learning the whole-body motion prior can help both the refinement of the body and hand motion by learning the correlation between them.", + "bbox": [ + 496, + 454, + 890, + 589 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 603, + 617, + 619 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this work, we have introduced an innovative approach to capture egocentric whole-body human motion. Our method comprises a single-frame-based whole-body pose estimation process, which includes FisheyeViT and pixel-aligned 3D heatmap representations. To enhance the initial whole-body pose estimates, we have integrated an uncertainty-aware diffusion-based motion refinement technique. Our experimental results demonstrate that both our single-frame method and the temporal-based method surpass all existing state-of-the-art techniques in terms of both quality and accuracy. Looking ahead, we see the potential for extending the applications of FisheyeViT to other vision tasks involving fisheye cameras. Future work could also involve incorporating facial expressions in whole-body motion capture.", + "bbox": [ + 496, + 628, + 890, + 840 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments This project was supported by the Saarbrücken Research Center for Visual Computing, Interaction and AI. Christian Theobalt was supported by ERC Consolidator Grant 4DReply (770784).", + "bbox": [ + 496, + 840, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "784", + "bbox": [ + 485, + 944, + 509, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Blender. https://www.blender.org. 6", + "[2] Mixamo. https://www MIXamo.com. 6", + "[3] Renderpeople. https://renderpeople.com. 6, 3", + "[4] Hiroyasu Akada, Jian Wang, Soshi Shimada, Masaki Takahashi, Christian Theobalt, and Vladislav Golyanik. Unrealego: A new dataset for robust egocentric 3d human motion capture. In European Conference on Computer Vision, pages 1-17. Springer, 2022. 1, 2", + "[5] Hiroyasu Akada, Jian Wang, Vladislav Golyanik, and Christian Theobalt. 3d human pose perception from egocentric stereo videos. In 37th IEEE/CVF Conference on Computer Vision and Pattern Recognition. IEEE, 2024. 2", + "[6] Zhongang Cai, Wanqi Yin, Ailing Zeng, Chen Wei, Qingping Sun, Yanjun Wang, Hui En Pang, Haiyi Mei, Mingyuan Zhang, Lei Zhang, et al. Smpler-x: Scaling up expressive human pose and shape estimation. arXiv preprint arXiv:2309.17448, 2023. 3", + "[7] Young-Woon Cha, True Price, Zhen Wei, Xinran Lu, Nicholas Rewkowski, Rohan Chabra, Zihe Qin, Hyounghun Kim, Zhaoqi Su, Yebin Liu, et al. Towards fully mobile 3d face, body, and environment capture using only head-worn cameras. IEEE transactions on visualization and computer graphics, 24(11):2993-3004, 2018. 2", + "[8] Jeongjun Choi, Dongseok Shim, and H Jin Kim. Diffupose: Monocular 3d human pose estimation via denoising diffusion probabilistic model. arXiv preprint arXiv:2212.02796, 2022. 3", + "[9] Vasileios Choutas, Georgios Pavlakos, Timo Bolkart, Dimitrios Tzionas, and Michael J Black. Monocular expressive body regression through body-driven attention. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part X 16, pages 20-40. Springer, 2020. 3", + "[10] Hai Ci, Mingdong Wu, Wentao Zhu, Xiaoxuan Ma, Hao Dong, Fangwei Zhong, and Yizhou Wang. Gfpose: Learning 3d human pose prior with gradient fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4800-4810, 2023. 3", + "[11] Benjamin Coors, Alexandru Paul Condurache, and Andreas Geiger. Spherenet: Learning spherical representations for detection and classification in omnidirectional images. In Proceedings of the European conference on computer vision (ECCV), pages 518-533, 2018. 5", + "[12] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in neural information processing systems, 34:8780-8794, 2021. 6", + "[13] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020.4, 8", + "[14] Yuming Du, Robin Kips, Albert Pumarola, Sebastian Starke, Ali Thabet, and Artsiom Sanakoyeu. Avatars grow legs: Generating smooth human motion from sparse tracking inputs with diffusion model. In Proceedings of the IEEE/CVF" + ], + "bbox": [ + 78, + 114, + 468, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Conference on Computer Vision and Pattern Recognition, pages 481-490, 2023. 3", + "[15] Yao Feng, Vasileios Choutas, Timo Bolkart, Dimitrios Tzionas, and Michael J Black. Collaborative regression of expressive bodies using moderation. In 2021 International Conference on 3D Vision (3DV), pages 792-804. IEEE, 2021. 3", + "[16] Lin Geng Foo, Jia Gong, Hossein Rahmani, and Jun Liu. Distribution-aligned diffusion for human mesh recovery. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9221-9232, 2023. 3", + "[17] Jia Gong, Lin Geng Foo, Zhipeng Fan, Qiuhong Ke, Hossein Rahmani, and Jun Liu. Diffpose: Toward more reliable 3d pose estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13041-13051, 2023. 3", + "[18] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 2, 3, 5, 6", + "[19] Karl Holmquist and Bastian Wandt. Diffpose: Multi-hypothesis human pose estimation using diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15977-15987, 2023. 3", + "[20] Ludovic Hoyet, Kenneth Ryall, Rachel McDonnell, and Carol O'Sullivan. Sleight of hand: perception of finger motion from reduced marker sets. In Proceedings of the ACM SIGGRAPH symposium on interactive 3D graphics and games, pages 79-86, 2012. 6", + "[21] Hao Jiang and Kristen Grauman. Seeing invisible poses: Estimating 3d body pose from egocentric video. In 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 3501-3509. IEEE, 2017. 2", + "[22] Zhongyu Jiang, Zhuoran Zhou, Lei Li, Wenhao Chai, ChengYen Yang, and Jenq-Neng Hwang. Back to optimization: Diffusion-based zero-shot 3d human pose estimation. arXiv preprint arXiv:2307.03833, 2023. 3", + "[23] Taeho Kang, Kyungjin Lee, Jinrui Zhang, and Youngki Lee. Ego3dpose: Capturing 3d cues from binocular egocentric views. arXiv preprint arXiv:2309.11962, 2023. 2", + "[24] David G Kendall. A survey of the statistical theory of shape. Statistical Science, 4(2):87-99, 1989. 3", + "[25] Jiaman Li, Karen Liu, and Jiajun Wu. Ego-body pose estimation via ego-head pose estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17142-17151, 2023. 2, 3", + "[26] Yuyan Li, Yuliang Guo, Zhixin Yan, Xinyu Huang, Ye Duan, and Liu Ren. Omnifusion: 360 monocular depth estimation via geometry-aware fusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2801-2810, 2022. 5", + "[27] Jing Lin, Ailing Zeng, Haoqian Wang, Lei Zhang, and Yu Li. One-stage 3d whole-body mesh recovery with component aware transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21159-21168, 2023. 3", + "[28] Yuxuan Liu, Jianxin Yang, Xiao Gu, Yao Guo, and Guang-Zhong Yang. Ego+ x: An egocentric vision system for global" + ], + "bbox": [ + 501, + 93, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "785", + "bbox": [ + 486, + 945, + 509, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "3d human pose estimation and social interaction characterization. In 2022 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 5271-5277. IEEE, 2022. 2", + "[29] Yuxuan Liu, Jianxin Yang, Xiao Gu, Yijun Chen, Yao Guo, and Guang-Zhong Yang. Egofish3d: Egocentric 3d pose estimation from a fisheye camera via self-supervised learning. IEEE Transactions on Multimedia, 2023. 1, 2, 4", + "[30] Yuxuan Liu, Jianxin Yang, Xiao Gu, Yao Guo, and GuangZhong Yang. Egohmr: Egocentric human mesh recovery via hierarchical latent diffusion model. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 9807-9813. IEEE, 2023. 1, 2, 3, 5", + "[31] Zhengyi Luo, Ryo Hachiuma, Ye Yuan, and Kris Kitani. Dynamics-regulated kinematic policy for egocentric pose estimation. Advances in Neural Information Processing Systems, 34:25019-25032, 2021. 2", + "[32] Christen Millerdurai, Hiroyasu Akada, Jian Wang, Diogo Luvizon, Christian Theobalt, and Vladislav Golyanik. Evento3d: 3d human motion capture from egocentric event streams. In 37th IEEE/CVF Conference on Computer Vision and Pattern Recognition. IEEE, 2024. 2", + "[33] Gyeongsik Moon, Ju Yong Chang, and Kyoung Mu Lee. V2v-posenet: Voxel-to-voxel prediction network for accurate 3d hand and human pose estimation from a single depth map. In Proceedings of the IEEE conference on computer vision and pattern Recognition, pages 5079–5088, 2018. 2", + "[34] Gyeongsik Moon, Hongsuk Choi, and Kyoung Mu Lee. Accurate 3d hand pose estimation for whole-body 3d human mesh estimation. In Computer Vision and Pattern Recognition Workshop (CVPRW), 2022. 4, 5, 7, 8, 2", + "[35] Evonne Ng, Donglai Xiang, Hanbyul Joo, and Kristen Grauman. You2me: Inferring body pose in egocentric video via first and second person interactions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9890-9900, 2020. 2", + "[36] Jinman Park, Kimathi Kaai, Saad Hossain, Norikatsu Sumi, Sirisha Rambhatla, and Paul Fieguth. Domain-guided spatiotemporal self-attention for egocentric 3d pose estimation. In Proceedings of the 29th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, pages 1837–1849, 2023. 2, 1", + "[37] Georgios Pavlakos, Vasileios Choutas, Nima Ghorbani, Timo Bolkart, Ahmed AA Osman, Dimitrios Tzionas, and Michael J Black. Expressive body capture: 3d hands, face, and body from a single image. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10975-10985, 2019. 3", + "[38] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 1 (2):3, 2022. 5", + "[39] Helge Rhodin, Christian Richardt, Dan Casas, Eldar Insafutdinov, Mohammad Shafiei, Hans-Peter Seidel, Bernt Schiele, and Christian Theobalt. Egocap: egocentric marker-less motion capture with two fisheye cameras. ACM Transactions on Graphics (TOG), 35(6):1-11, 2016. 2" + ], + "bbox": [ + 78, + 90, + 468, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[40] Yu Rong, Takaaki Shiratori, and Hanbyul Joo. Frankmocap: A monocular 3d whole-body pose estimation system via regression and integration. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1749-1759, 2021. 3", + "[41] Davide Scaramuzza, Agostino Martinelli, and Roland Siegwart. A toolbox for easily calibrating omnidirectional cameras. In 2006 IEEE/RSJ International Conference on Intelligent Robots and Systems, pages 5695-5701. IEEE, 2006. 1", + "[42] Wenkang Shan, Zhenhua Liu, Xinfeng Zhang, Zhao Wang, Kai Han, Shanshe Wang, Siwei Ma, and Wen Gao. Diffusion-based 3d human pose estimation with multi-hypothesis aggregation. arXiv preprint arXiv:2303.11579, 2023.3", + "[43] Soshi Shimada, Vladislav Golyanik, Weipeng Xu, and Christian Theobalt. Physcap: Physically plausible monocular 3d motion capture in real time. ACM Transactions on Graphics (ToG), 39(6):1-16, 2020. 5", + "[44] Xiao Sun, Bin Xiao, Fangyin Wei, Shuang Liang, and Yichen Wei. Integral human pose regression. In Proceedings of the European conference on computer vision (ECCV), pages 529-545, 2018. 4, 5", + "[45] Yu Sun, Tianyu Huang, Qian Bao, Wu Liu, Wenpeng Gao, and Yili Fu. Learning monocular mesh recovery of multiple body parts via synthesis. In ICASSP 2022-2022 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 2669-2673. IEEE, 2022. 3", + "[46] Omid Taheri, Nima Ghorbani, Michael J Black, and Dimitrios Tzionas. Grab: A dataset of whole-body human grasping of objects. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part IV 16, pages 581-600. Springer, 2020. 6", + "[47] Denis Tomè, Patrick Peluse, Lourdes Agapito, and Hernán Badino. xr-egopose: Egocentric 3d human pose from an HMD camera. In ICCV, pages 7727-7737, 2019. 1, 2, 5, 6, 7, 8, 4", + "[48] Jonathan Tseng, Rodrigo Castellon, and Karen Liu. Edge: Editable dance generation from music. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 448-458, 2023. 5, 2", + "[49] Jingdong Wang, Ke Sun, Tianheng Cheng, Borui Jiang, Chaorui Deng, Yang Zhao, Dong Liu, Yadong Mu, Mingkui Tan, Xinggang Wang, et al. Deep high-resolution representation learning for visual recognition. IEEE transactions on pattern analysis and machine intelligence, 43(10):3349-3364, 2020. 5", + "[50] Jian Wang, Lingjie Liu, Weipeng Xu, Kripasindhu Sarkar, and Christian Theobalt. Estimating egocentric 3d human pose in global space. ICCV, 2021. 1, 2, 6, 7, 8, 3, 4, 5", + "[51] Jian Wang, Lingjie Liu, Weipeng Xu, Kripasindhu Sarkar, Diogo Luvizon, and Christian Theobalt. Estimating egocentric 3d human pose in the wild with external weak supervision. In CVPR, pages 13157-13166, 2022. 1, 2, 6, 7, 8, 4", + "[52] Jian Wang, Diogo Luvizon, Weipeng Xu, Lingjie Liu, Kripasindhu Sarkar, and Christian Theobalt. Scene-aware egocentric 3d human pose estimation. In Proceedings of the" + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "786", + "bbox": [ + 486, + 945, + 511, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13031-13040, 2023. 1, 2, 5, 6, 7, 8, 3, 4", + "[53] Donglai Xiang, Hanbyul Joo, and Yaser Sheikh. Monocular total capture: Posing face, body, and hands in the wild. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10965-10974, 2019. 3", + "[54] Weipeng Xu, Avishek Chatterjee, Michael Zollhöfer, Helge Rhodin, Pascal Fua, Hans-Peter Seidel, and Christian Theobalt. $\\mathrm{Mo}^2\\mathrm{cap}^2$ : Real-time mobile 3d motion capture with a cap-mounted fisheye camera. IEEE Trans. Vis. Comput. Graph., 25(5):2093-2101, 2019. 1, 2, 5, 6, 7, 8, 3, 4", + "[55] Yufei Xu, Jing Zhang, Qiming Zhang, and Dacheng Tao. Vitpose: Simple vision transformer baselines for human pose estimation. Advances in Neural Information Processing Systems, 35:38571-38584, 2022. 2", + "[56] Dianyi Yang, Jiadong Tang, Yu Gao, Yi Yang, and Mengyin Fu. Sector patch embedding: An embedding module conforming to the distortion pattern of fisheye image. arXiv preprint arXiv:2303.14645, 2023. 5", + "[57] Shangrong Yang, Chunyu Lin, Kang Liao, and Yao Zhao. Dual diffusion architecture for fisheye image rectification: Synthetic-to-real generalization. arXiv preprint arXiv:2301.11785, 2023.", + "[58] Fanghua Yu, Xintao Wang, Mingdeng Cao, Gen Li, Ying Shan, and Chao Dong. Osrt: Omnidirectional image superresolution with distortion-aware transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13283-13292, 2023. 5", + "[59] Ye Yuan and Kris Kitani. 3d ego-pose estimation via imitation learning. In Proceedings of the European Conference on Computer Vision (ECCV), pages 735–750, 2018. 2", + "[60] Ye Yuan and Kris Kitani. Ego-pose estimation and forecasting as real-time pd control. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10082-10092, 2019. 2", + "[61] Ye Yuan, Jiaming Song, Umar Iqbal, Arash Vahdat, and Jan Kautz. Physdiff: Physics-guided human motion diffusion model. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 16010-16021, 2023. 5", + "[62] Siwei Zhang, Qianli Ma, Yan Zhang, Zhiyin Qian, Taein Kwon, Marc Pollefeys, Federica Bogo, and Siyu Tang. Ego-body: Human body shape and motion of interacting people from head-mounted devices. In European Conference on Computer Vision, pages 180–200. Springer, 2022. 6", + "[63] Siwei Zhang, Qianli Ma, Yan Zhang, Sadegh Aliakbarian, Darren Cosker, and Siyu Tang. Probabilistic human mesh recovery in 3d scenes from egocentric views. arXiv preprint arXiv:2304.06024, 2023. 3", + "[64] Yahui Zhang, Shaodi You, and Theo Gevers. Automatic calibration of the fisheye camera for egocentric 3d human pose estimation from a single image. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1772-1781, 2021. 2", + "[65] Dongxu Zhao, Zhen Wei, Jisan Mahmud, and Jan-Michael Frahm. Egoglass: Egocentric-view human pose estimation from an eyeglass frame. In 2021 International Conference on 3D Vision (3DV), pages 32-41. IEEE, 2021. 2" + ], + "bbox": [ + 78, + 92, + 467, + 898 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[66] Yuxiao Zhou, Marc Habermann, Ikhsanul Habibie, Ayush Tewari, Christian Theobalt, and Feng Xu. Monocular real-time full body capture with inter-part correlations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4811-4822, 2021. 3", + "[67] Christian Zimmermann, Duygu Ceylan, Jimei Yang, Bryan Russell, Max Argus, and Thomas Brox. Freihand: A dataset for markerless capture of hand pose and shape from single rgb images. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 813-822, 2019. 2" + ], + "bbox": [ + 501, + 92, + 890, + 231 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "787", + "bbox": [ + 486, + 945, + 509, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2024/Egocentric Whole-Body Motion Capture with FisheyeViT and Diffusion-Based Motion Refinement/90ece1d3-5ea4-4b56-9fe8-5a7d9a524eef_model.json b/2024/Egocentric Whole-Body Motion Capture with FisheyeViT and Diffusion-Based Motion Refinement/90ece1d3-5ea4-4b56-9fe8-5a7d9a524eef_model.json new file mode 100644 index 0000000000000000000000000000000000000000..ceda96b694f7a977dd8e9bba2d05321f30a0d96b --- /dev/null +++ b/2024/Egocentric Whole-Body Motion Capture with FisheyeViT and Diffusion-Based Motion Refinement/90ece1d3-5ea4-4b56-9fe8-5a7d9a524eef_model.json @@ -0,0 +1,2873 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.131, + 0.882, + 0.175 + ], + "angle": 0, + "content": "Egocentric Whole-Body Motion Capture with FisheyeViT and Diffusion-Based Motion Refinement" + }, + { + "type": "text", + "bbox": [ + 0.234, + 0.203, + 0.338, + 0.222 + ], + "angle": 0, + "content": "Jian Wang\\(^{1,4}\\)" + }, + { + "type": "text", + "bbox": [ + 0.162, + 0.223, + 0.325, + 0.239 + ], + "angle": 0, + "content": "Kripasindhu Sarkar2" + }, + { + "type": "text", + "bbox": [ + 0.368, + 0.204, + 0.445, + 0.221 + ], + "angle": 0, + "content": "Zhe Cao2" + }, + { + "type": "text", + "bbox": [ + 0.351, + 0.222, + 0.473, + 0.239 + ], + "angle": 0, + "content": "Danhang Tang" + }, + { + "type": "text", + "bbox": [ + 0.472, + 0.204, + 0.612, + 0.222 + ], + "angle": 0, + "content": "Diogo Luvizon1,4" + }, + { + "type": "text", + "bbox": [ + 0.502, + 0.222, + 0.617, + 0.238 + ], + "angle": 0, + "content": "Thabo Beeler2" + }, + { + "type": "text", + "bbox": [ + 0.636, + 0.204, + 0.736, + 0.222 + ], + "angle": 0, + "content": "Lingjie Liu3" + }, + { + "type": "text", + "bbox": [ + 0.644, + 0.222, + 0.809, + 0.238 + ], + "angle": 0, + "content": "Christian Theobalt1,4" + }, + { + "type": "text", + "bbox": [ + 0.122, + 0.239, + 0.851, + 0.257 + ], + "angle": 0, + "content": "1MPI Informatics & Saarland Informatics Campus 2Google 3University of Pennsylvania" + }, + { + "type": "text", + "bbox": [ + 0.124, + 0.257, + 0.842, + 0.275 + ], + "angle": 0, + "content": "\\(^{4}\\)Saarbrücken Research Center for Visual Computing, Interaction and Artificial Intelligence" + }, + { + "type": "text", + "bbox": [ + 0.282, + 0.283, + 0.687, + 0.299 + ], + "angle": 0, + "content": "Project Page: https://jianwang-mpi.github.io/egowholemocap" + }, + { + "type": "image", + "bbox": [ + 0.08, + 0.31, + 0.207, + 0.425 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.207, + 0.31, + 0.405, + 0.425 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.408, + 0.31, + 0.481, + 0.425 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.491, + 0.31, + 0.616, + 0.425 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.617, + 0.31, + 0.814, + 0.425 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.816, + 0.31, + 0.889, + 0.425 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.439, + 0.893, + 0.468 + ], + "angle": 0, + "content": "Figure 1. From an image sequence captured by a single head-mounted fisheye camera, our method can predict accurate and temporally coherent whole-body motion, including human body and hand poses. The SMPL-X parameters are obtained using inverse kinematics." + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.477, + 0.314, + 0.493 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.51, + 0.473, + 0.829 + ], + "angle": 0, + "content": "In this work, we explore egocentric whole-body motion capture using a single fisheye camera, which simultaneously estimates human body and hand motion. This task presents significant challenges due to three factors: the lack of high-quality datasets, fisheye camera distortion, and human body self-occlusion. To address these challenges, we propose a novel approach that leverages FisheyeViT to extract fisheye image features, which are subsequently converted into pixel-aligned 3D heatmap representations for 3D human body pose prediction. For hand tracking, we incorporate dedicated hand detection and hand pose estimation networks for regressing 3D hand poses. Finally, we develop a diffusion-based whole-body motion prior model to refine the estimated whole-body motion while accounting for joint uncertainties. To train these networks, we collect a large synthetic dataset, EgoWholeBody, comprising 840,000 high-quality egocentric images captured across a diverse range of whole-body motion sequences. Quantitative and qualitative evaluations demonstrate the effectiveness of our method in producing high-quality whole-body motion estimates from a single egocentric camera." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.845, + 0.21, + 0.86 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.871, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Egocentric 3D human motion estimation using head-mounted devices [47, 54] has garnered significant traction" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.479, + 0.895, + 0.797 + ], + "angle": 0, + "content": "in recent years, driven by its diverse applications in VR/AR. Immersed in a virtual world, we can traverse virtual environments, interact with virtual objects, and even simulate real-world interactions. To fully capture the intricacies of human motion during such interaction, understanding both body and hand movements is essential. While existing egocentric motion capture methods [30, 47, 50-52, 54] focus solely on body motion, neglecting the hands, this work proposes the task of egocentric whole-body motion capture, i.e. simultaneous estimation of the body motion and hand motion from a single head-mounted fisheye camera (shown in Fig. 1). This task is extremely challenging due to three factors: First, the fisheye image introduces significant distortion, making it difficult for existing networks, which are designed for non-distorted images, to extract features. Second, the egocentric camera perspective frequently leads to the occlusion of body parts, such as the feet and hands, further complicating the task of whole-body motion capture. Lastly, large-scale training data with ground truth annotations for both body and hand poses is absent in existing datasets [4, 29, 47, 51, 54]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.81, + 0.895, + 0.903 + ], + "angle": 0, + "content": "In this work, we propose a novel egocentric whole-body motion capture method to address the aforementioned challenges. To effectively address fisheye distortion, we propose FisheyeViT for extracting image features, along with a joint regressor employing pixel-aligned 3D heatmap for predicting 3D body poses. Instead of attempting to undis-" + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.946, + 0.511, + 0.957 + ], + "angle": 0, + "content": "777" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.473, + 0.455 + ], + "angle": 0, + "content": "tort the entire fisheye image, which is impractical due to the fisheye lens's large field of view (FOV), we opt to partition the image into smaller patches aligned with a specific FOV range. This approach enables individual patch-level undistortion and seamlessly aligns with the vision transformer architecture that is employed for extracting the complete image feature map. We further propose an egocentric 3D pose regressor utilizing 3D heatmap representations. Unlike the existing approach [52] that projects image features into 3D space through fisheye reprojection functions and regresses 3D heatmaps with V2V networks [33] leading to intricate network learning and high computational complexity—our proposed egocentric pose regressor adopts a simpler approach. It employs deconvolutional layers to obtain pixel-aligned 3D heatmaps. Notably, the voxels in the 3D heatmap directly correspond to pixels in 2D features, subsequently linking to image patches in FisheyeViT. This streamlined approach significantly simplifies network training. Joint locations from the pixel-aligned 3D heatmap are finally transformed with the fisheye camera model to obtain the 3D human body poses. Due to the large size difference between body and hands, we train a hand detection network and a hand pose estimation network to accurately regress 3D hand poses." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.457, + 0.473, + 0.715 + ], + "angle": 0, + "content": "To overcome the challenges posed by self-occlusion and improve the accuracy of pose estimation, we propose a novel method for refining the whole-body motion predictions by incorporating temporal context and a motion prior. Our method learns a whole-body motion prior with the diffusion model [18] from a collection of diverse human motion sequences, capturing intrinsic correlations between hand and body movements. Following this, we extract the joint uncertainties from the pixel-aligned 3D heatmap and utilize them to guide the refinement of the whole-body motion. The joint uncertainties act as indicators of the trustworthiness of the pose regressor's predictions. By conditioning on joints with low uncertainty, our whole-body motion diffusion model selectively refines joints with high uncertainty. This strategy substantially improves the quality of whole-body pose estimations and effectively mitigates the effects of self-occlusion." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.717, + 0.47, + 0.839 + ], + "angle": 0, + "content": "In response to the absence of the egocentric whole-body motion capture datasets, we present EgoWholeBody, a new large-scale high-quality synthetic dataset. This dataset encompasses a wide range of whole-body motions, comprising over 870k frames, which significantly surpasses the size of previous egocentric training datasets. EgoWholeBody could serve as a valuable resource for advancing research in egocentric whole-body motion capture." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.841, + 0.471, + 0.903 + ], + "angle": 0, + "content": "A thorough evaluation across a range of datasets, including SceneEgo [52], GlobalEgoMocap [50] and \\(\\mathrm{Mo}^2\\mathrm{Cap}^2\\) [54], has demonstrated the remarkable improvements of our method in estimating egocentric whole-body" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.152 + ], + "angle": 0, + "content": "motion compared to previous approaches. This substantiates the effectiveness of our approach in addressing the special challenges encountered in egocentric views, including the fisheye distortion and self-occlusion." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.153, + 0.85, + 0.167 + ], + "angle": 0, + "content": "In summary, our key contributions are the following:" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.169, + 0.892, + 0.214 + ], + "angle": 0, + "content": "- The first egocentric whole-body motion capture method that predicts accurate and temporarily coherent egocentric body and hand motion;" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.215, + 0.892, + 0.26 + ], + "angle": 0, + "content": "- FisheyeViT for alleviating fisheye camera distortion and pose regressor using pixel-aligned 3D heatmaps for accurate egocentric body pose estimation from a single image;" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.26, + 0.892, + 0.305 + ], + "angle": 0, + "content": "- Uncertainty-aware refinement method based on motion diffusion models for correcting initial pose estimations and predicting plausible motions even under occlusion;" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.305, + 0.892, + 0.336 + ], + "angle": 0, + "content": "- EgoWholeBody, a new high-quality synthetic dataset for egocentric whole-body motion capture." + }, + { + "type": "list", + "bbox": [ + 0.5, + 0.169, + 0.892, + 0.336 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.347, + 0.642, + 0.364 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.373, + 0.892, + 0.553 + ], + "angle": 0, + "content": "Egocentric 3D Human Body Pose Estimation. Recently, there has been growing interest in estimating egocentric 3D poses from body-worn cameras. Some methods [21, 25, 31, 35, 59, 60] use front-facing cameras and infer the human body motion from the camera view. However, since the user's body is often unobserved by the camera, these methods fail when the human body is not roaming around. Millerdurai et al. [32] leverage event cameras for estimating egocentric body pose. Other methods [4, 5, 7, 23, 39, 65] use head-mounted down-facing stereo cameras to estimate body poses. However, stereo camera setups introduce extra burdens of weight and energy consumption." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.554, + 0.893, + 0.827 + ], + "angle": 0, + "content": "Xu et al. [54] and Tome et al. [47] introduce the single head-mounted down-facing fisheye camera setup for the egocentric 3D human pose estimation task. Zhang et al. [64] regressed fisheye camera parameters and 3D human pose simultaneously. To address the self-occlusion issue, Park et al. [36] leveraged the temporal information with the spatio-temporal self-attention network, and Liu et al. [30] introduced diffusion model to generate 3D human pose conditioned on egocentric image features. Wang et al. [50] and Liu et al. [28] combined the SLAM and egocentric pose estimation methods to estimate human body poses in the world coordinate. Wang et al. [51] and Liu et al. [29] leverage the synchronized egocentric camera and external cameras to collect large-scale egocentric pose estimation datasets with pseudo-ground truth. Considering the human-scene interaction, Wang et al. [52] estimated the scene geometry from the egocentric camera and constrained the 3D human pose with it." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.826, + 0.894, + 0.903 + ], + "angle": 0, + "content": "These methods only focus on estimating human body poses while omitting the hand motion, and they still suffer from fisheye camera distortion since they directly put the highly distorted fisheye images into the neural network. Our proposed method can capture whole-body motion and" + }, + { + "type": "page_number", + "bbox": [ + 0.487, + 0.946, + 0.511, + 0.957 + ], + "angle": 0, + "content": "778" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.081, + 0.089, + 0.891, + 0.304 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.313, + 0.893, + 0.385 + ], + "angle": 0, + "content": "Figure 2. Overview of our whole-body motion capture pipeline. We first use FisheyeViT to undistort the input image and generate image feature tokens (3.1.1). Next, we use a 1D convolutional network to convert the image features to a pixel-aligned 3D heatmap and use soft-argmax and fisheye camera undistortion function to obtain the 3D body joins positions and uncertainty (3.1.2). We further detect the hand location and regress the 3D hand poses from the input image (3.1.3). Finally, the estimated hand motion and human body motion are combined and the uncertainty-aware diffusion model is applied to refine the estimated whole-body motion (3.2)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.399, + 0.469, + 0.429 + ], + "angle": 0, + "content": "resolve the fisheye camera distortion issue with the FisheyeViT and pixel-aligned 3D heatmap." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.43, + 0.469, + 0.687 + ], + "angle": 0, + "content": "Whole-Body 3D Pose Estimation. Whole-body 3D pose estimation aims to estimate the 3D human body, face, and hands parameters from input images. This task is crucial for many applications, e.g., modeling human activities and human-scene interactions. Some methods [37, 53] fit the 2D body joints estimated from images with optimization algorithms, while these methods suffer from high computation overhead and can fall into local optima. Some other learning-based methods [6, 9, 15, 27, 40, 45, 66] use the neural network to regress the SMPL-X [37] parameters from input images. For example, ExPose [9] introduced body-driven attention to extract face and hand crops and used a refinement module to regress whole-body pose. OSX [27] proposed a one-stage pipeline for whole-body mesh recovery without separate networks for each part. SMPLer-X [6] propose a foundation model for whole-body pose estimation trained with the large model and big data." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.688, + 0.469, + 0.779 + ], + "angle": 0, + "content": "Though much progress has been made on whole-body pose estimation from an external view, the task from an ego-centric view is still unexplored. In this paper, we introduce the first whole-body 3D pose estimation method from a single egocentric image and also incorporate temporal information with diffusion-based motion refinement." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.78, + 0.47, + 0.902 + ], + "angle": 0, + "content": "Diffusion Models for Pose Estimation. Recently, some methods [8, 10, 16, 17, 19, 42] have effectively applied Denoising Diffusion Probabilistic Models (DDPM) [18] to human pose estimation tasks. Building on the success of motion diffusion models in human pose estimation, many methods have extended this approach to egocentric pose estimation, where the human body is only partially visible from RGB cameras or VR sensors. Zhang et al.'s work [63]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.399, + 0.892, + 0.504 + ], + "angle": 0, + "content": "uses a diffusion model to generate realistic human poses considering scene geometry. AGROL [14] generates body motion based on head and hand 6D pose estimates from a VR headset. EgoEgo [25] estimates head poses from a head-mounted front-facing camera and uses them to generate body poses. EgoHMR [30] extracts image features and uses them as a condition for the diffusion denoising process." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.505, + 0.893, + 0.716 + ], + "angle": 0, + "content": "However, the aforementioned pose estimation methods train the conditioned diffusion model with image features or IMU signals. This cannot be generalized since the trained network only accepts one specific condition format and is inclined to learn domain-specific distributions of condition features. ZeDO [22] tackles this issue with a zero-shot diffusion-based optimization approach that doesn't require training with 2D-3D or image-3D pairs. Our method leverages the uncertainty value given by the single-frame pose estimation network and refines the initial motion estimation with the uncertainty of each joint. Moreover, different from previous methods that only focus on human body motion, we train a whole-body motion diffusion model to construct the relationship between hand and body motion." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.729, + 0.591, + 0.744 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.755, + 0.892, + 0.8 + ], + "angle": 0, + "content": "In this section, we propose a new method for predicting accurate egocentric whole-body poses from egocentric image sequences. An overview of our approach is shown in Fig. 2." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.809, + 0.891, + 0.825 + ], + "angle": 0, + "content": "3.1. Single Image Based Egocentric Pose Estimation" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.832, + 0.634, + 0.847 + ], + "angle": 0, + "content": "3.1.1 FisheyeViT" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.893, + 0.901 + ], + "angle": 0, + "content": "In this section, we introduce FisheyeViT, which is specially designed to alleviate the fisheye distortion issue. Instead of undistorting the entire fisheye image, we extract undistorted" + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.946, + 0.511, + 0.957 + ], + "angle": 0, + "content": "779" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.088, + 0.089, + 0.265, + 0.203 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.269, + 0.089, + 0.462, + 0.205 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.087, + 0.206, + 0.264, + 0.323 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.269, + 0.206, + 0.462, + 0.323 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.088, + 0.334, + 0.453, + 0.348 + ], + "angle": 0, + "content": "Figure 3. The detailed illustration of FisheyeViT (Sec. 3.1.1)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.361, + 0.468, + 0.451 + ], + "angle": 0, + "content": "image patches from the fisheye image and then fit these patches as tokens into the transformer network [13]. To get the undistorted patches, we first warp the fisheye image to a unit semi-sphere, then get the patches with the gnomonic projection (see Fig. 2). The FisheyeViT can be split into five steps, the first four of which are illustrated in Fig. 3." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.452, + 0.469, + 0.635 + ], + "angle": 0, + "content": "Step 1. Given an input image \\(\\mathbf{I}\\) with size \\(H\\times W\\), we first evenly sample \\(N\\times N\\) patch center points: \\(\\{\\mathbf{C}_{ij} = (u_i,v_j) = \\left(\\frac{H}{N} (i + \\frac{1}{2}),\\frac{W}{N} (j + \\frac{1}{2})\\right)|i,j\\in 0,\\dots,N - 1\\}\\). Then, the patch center points \\(\\mathbf{C}_{ij}\\) are projected onto a unit sphere with the fisheye reprojection function: \\(\\mathbf{P}_{ij}^{c} = (x_{ij}^{c},y_{ij}^{c},z_{ij}^{c}) = \\mathcal{P}^{-1}(u_i,v_j,1)\\). The fisheye camera model is described in Sec. 8 of the supplementary material. Given a point \\(\\mathbf{P}_{ij}^{c}\\) on the unit sphere, the tangent plane \\(\\mathbf{T}_{ij}\\) that passes through the point is defined by the normal vector \\(\\mathbf{v}_{ij}^{c} = (x_{ij}^{c},y_{ij}^{c},z_{ij}^{c})\\). In the following steps, we implement the gnomonic projection by sampling grid points in the plane and projecting them back onto the fisheye image." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.635, + 0.469, + 0.803 + ], + "angle": 0, + "content": "Step 2. In this step, we determine the orientation of the grid points in the tangent plane, ensuring that the grid points from different tangent planes \\(\\mathbf{T}_{ij}\\) have the same orientation when projected back onto the fisheye image. To achieve this, we select a 2D point \\(\\mathbf{U}_{ij} = (u_i + d,v_j)\\) in the fisheye image space that is \\(d\\) pixels to the right of the patch center point and project it to the unit sphere using the fisheye reprojection function: \\(\\mathbf{P}_{ij}^{u} = (x_{ij}^{u},y_{ij}^{u},z_{ij}^{u}) = \\mathcal{P}^{-1}(u_i + d,v_j,1)\\). We then calculate the intersection point \\(\\mathbf{P}_{ij}^{x}\\) between the vector \\(\\mathbf{v}_{ij}^{u} = (x_{ij}^{u},y_{ij}^{u},z_{ij}^{u})\\) that is passing the origin and the tangent plane \\(\\mathbf{T}_{ij}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.141, + 0.81, + 0.469, + 0.85 + ], + "angle": 0, + "content": "\\[\n\\mathbf {P} _ {i j} ^ {x} = \\frac {\\left\\langle \\mathbf {P} _ {i j} ^ {c} , \\mathbf {v} _ {i j} ^ {c} \\right\\rangle}{\\left\\langle \\mathbf {v} _ {i j} ^ {u} , \\mathbf {v} _ {i j} ^ {c} \\right\\rangle} \\mathbf {v} _ {i j} ^ {u} = \\frac {1}{\\left\\langle \\mathbf {v} _ {i j} ^ {u} , \\mathbf {v} _ {i j} ^ {c} \\right\\rangle} \\mathbf {v} _ {i j} ^ {u}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.856, + 0.33, + 0.871 + ], + "angle": 0, + "content": "where \\(\\langle \\cdot ,\\cdot \\rangle\\) denotes the inner product." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.469, + 0.903 + ], + "angle": 0, + "content": "Step 3. Based on the center point \\(\\mathbf{P}_{ij}^{c}\\) and intersection point \\(\\mathbf{P}_{ij}^{x}\\) on the tangent plane \\(\\mathbf{T}_{ij}\\), we build a coordinate" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.892, + 0.153 + ], + "angle": 0, + "content": "system with the \\(x\\) axis: \\(\\mathbf{v}_{ij}^{x} = \\mathrm{Norm}(\\mathbf{P}_{ij}^{x} - \\mathbf{P}_{ij}^{c})\\), the \\(z\\) axis: \\(\\mathbf{v}_{ij}^{z} = \\mathrm{Norm}(\\mathbf{v}_{ij}^{c})\\) and the \\(y\\) axis: \\(\\mathbf{v}_{ij}^{y} = \\mathbf{v}_{ij}^{x} \\times \\mathbf{v}_{ij}^{x}\\) where Norm denotes the normalize operation. We grid-sample \\(M \\times M\\) points in a \\(l \\times l\\) square on the \\(x-y\\) plane:" + }, + { + "type": "equation", + "bbox": [ + 0.578, + 0.163, + 0.892, + 0.19 + ], + "angle": 0, + "content": "\\[\n\\left\\{\\mathbf {P} _ {i j} ^ {m n} = \\mathbf {P} _ {i j} ^ {c} + \\left(l \\frac {m}{M} \\mathbf {v} _ {i j} ^ {x}, l \\frac {n}{M} \\mathbf {v} _ {i j} ^ {y}\\right) \\right\\} \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.201, + 0.891, + 0.219 + ], + "angle": 0, + "content": "where \\(m,n\\in -\\frac{1}{2} (M - 1),\\ldots , - \\frac{3}{2}, - \\frac{1}{2},\\frac{1}{2},\\frac{3}{2},\\ldots \\frac{1}{2} (M - 1)\\)" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.219, + 0.892, + 0.326 + ], + "angle": 0, + "content": "Step 4. The points \\(\\mathbf{P}_{ij}^{mn}\\) are projected back to the fisheye image with the fisheye projection function: \\(\\mathbf{C}_{ij}^{mn} = \\mathcal{P}(\\mathbf{P}_{ij}^{mn})\\). We then apply bilinear sampling to obtain the colors at points \\(\\mathbf{C}_{ij}^{mn}\\) of the input image \\(\\mathbf{I}\\), yielding the undistorted image patch \\(\\mathbf{I}_{ij}^{\\mathrm{undis}}\\). Please also see the supplementary video for a visual demonstration of undistorted image patches and their movement on the fisheye image." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.326, + 0.892, + 0.416 + ], + "angle": 0, + "content": "Step 5. The image patches \\(\\{\\mathbf{I}_{ij}^{\\mathrm{undis}}\\}\\) are sent to a ViT transformer network [13] to obtain the feature tokens \\(\\{\\mathbf{F}_{ij}\\}\\). The feature token is further reshaped in \\(i\\times j\\) matrix and obtain the image feature \\(\\mathbf{F}\\). In the FisheyeViT, we empirically chose \\(N = 16\\); \\(M = 16\\); \\(d = 8\\); \\(l = 0.2m\\) given the image size \\(H = W = 256\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.417, + 0.893, + 0.614 + ], + "angle": 0, + "content": "Note that \\(\\mathbf{C}_{ij}^{mn}\\) is independent of the image \\(\\mathbf{I}\\). This means that, given a fixed fisheye camera model, we can precompute \\(\\mathbf{C}_{ij}^{mn}\\) for all combinations of \\(m,n\\) and \\(i,j\\) in advance. This significantly speeds up both the training and evaluation processes. Furthermore, the number and dimensions of image patches \\(\\{\\mathbf{I}_{ij}^{\\mathrm{undis}}\\}\\) match exactly with those in the traditional ViT network. This compatibility allows us to finetune existing ViT networks on our egocentric datasets. Our sampling strategy ensures that each image patch \\(\\mathbf{I}_{ij}^{\\mathrm{undis}}\\) corresponds to the same FOV range in the fisheye camera. In our ablation study in Sec. 5.3, we show that FisheyeViT enhances the performance of the pose estimation network when applied to egocentric fisheye images." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.635, + 0.885, + 0.65 + ], + "angle": 0, + "content": "3.1.2 Pose Regressor with Pixel-Aligned 3D Heatmap" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.659, + 0.892, + 0.901 + ], + "angle": 0, + "content": "After collecting image features with FisheyeViT, we utilize a 3D heatmap-based network to estimate the body poses. The existing 3D heatmap-based pose regressors [34, 44] are designed for the weak-perspective cameras and predict the 3D heatmap in \\( xyz \\) space. Directly applying these regressors will result in misalignment between 3D heatmap features in \\( xyz \\) space and 2D image features in the fisheye image space. Therefore, we introduce a novel egocentric pose regressor that relies on the pixel-aligned 3D heatmap, tailored to address the needs of fisheye cameras. The idea is to regress the 3D heatmap in \\( uvd \\) space rather than traditional \\( xyz \\) space, where \\( uv \\) corresponds to the fisheye image \\( uv \\) space. Specifically, given a feature map \\( \\mathbf{F} \\in \\mathbb{R}^{C \\times N \\times N} \\), where \\( C \\) is the channel number, \\( N \\) is feature map height and width, we firstly use two deconvolutional layers to convert the feature map \\( \\mathbf{F} \\) into shape" + }, + { + "type": "page_number", + "bbox": [ + 0.487, + 0.946, + 0.511, + 0.957 + ], + "angle": 0, + "content": "780" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.091, + 0.473, + 0.348 + ], + "angle": 0, + "content": "\\((D_h\\times J,H_h,W_h)\\) , and further reshape it to pixel-aligned 3D heatmap \\(\\mathbf{H}\\in \\mathbb{R}^{J\\times D_h\\times H_h\\times W_h}\\) , where \\(J\\) is the joint number and \\(D_{h},H_{h},W_{h}\\) is the 3D heatmap depth, height and width. The illustration of pixel-aligned 3D heatmap is shown in Fig. 2. Next, we obtain the max-value positions \\(\\tilde{\\mathbf{J}}_b = \\{(u_i,v_i,d_i)\\mid i\\in 0,1,2,\\dots,J\\}\\) from \\(\\mathbf{H}\\) by the differentiable soft-argmax operation [44]. Here, we note that \\(u_{i}\\) and \\(v_{i}\\) correspond to the uv-coordinate of the 3D body joint projected in the fisheye image space, and \\(d_{i}\\) denotes the distance of the joint to the fisheye camera. Finally, the 3D body joints \\(\\hat{\\mathbf{J}}_b = \\{(x_i,y_i,z_i)\\mid i\\in 0,1,2,\\dots,J\\}\\) are recovered with the fisheye reprojection function: \\((x_{i},y_{i},z_{i}) =\\) \\(\\mathcal{P}^{-1}(u_i,v_i,d_i)\\) . The predicted body pose \\(\\hat{\\mathbf{J}}_b\\) is finally compared with the ground truth body pose \\(\\mathbf{J}_b\\) with the MSE loss. By first regressing 3D body poses in uvd space and then reproducing it, we ensure that the 3D heatmap is pixelaligned with the end-to-end training." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.35, + 0.473, + 0.698 + ], + "angle": 0, + "content": "With the pixel-aligned heatmap, our proposed 3D pose regressor solves problems in all three types of previous egocentric joint regressors. First, \\(\\mathrm{Mo}^2\\mathrm{Cap}^2\\) [54] employs separate networks to predict 2D joint positions and joint distances. However, this method can yield unrealistic joint estimations because small errors in 2D joints can result in large errors in 3D joints due to the projection effect. Second, \\(xR\\)-egopose [47] and EgoHMR [30] directly regress the 3D joint positions. However, this method is agnostic to the fisheye camera parameters, making it suitable only for a specific camera configuration (e.g., camera parameters, head-mounted position, etc.). Third, SceneEgo [52] projects 2D features into 3D voxel space and uses a V2V network to regress 3D poses. Because of these, the SceneEgo method suffers from low accuracy and large computation overhead. Different from previous methods, our pose regressor with pixel-aligned 3D heatmap is versatile and efficient since it directly estimates 3D joints while also incorporating an explicitly parametrized fisheye camera model. Moreover, it can preserve the uncertainty of the estimated joints, which will be used in our uncertainty-aware motion refinement method (Sec. 3.2.2). Detailed comparison with other pose prediction heads is shown in Table 3." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.723, + 0.367, + 0.739 + ], + "angle": 0, + "content": "3.1.3 Egocentric Hand Pose Estimation" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.749, + 0.47, + 0.809 + ], + "angle": 0, + "content": "In this section, we first train a network to detect hand pose locations and then train a 3D hand pose estimation network to regress 3D hand poses. Then, we describe how to integrate the estimated hand and body poses." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.811, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Hand Detection. Given an input image \\(\\mathbf{I}\\), we finetune the HRNet [49] network to regress the 2D hand poses of left hand \\(\\mathbf{J}_{lh}^{2d}\\) and right hand \\(\\mathbf{J}_{rh}^{2d}\\). From the hand poses, we obtain the center point of left hand \\(\\mathbf{C}_{lh}\\) and right hand \\(\\mathbf{C}_{rh}\\) along with the bounding box sizes, \\(d_{lh}\\) and \\(d_{rh}\\). We use our approach described in Sec. 3.1.1 to compute undistorted" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.8, + 0.107 + ], + "angle": 0, + "content": "image patches of left \\(\\mathbf{I}_{lh}\\) and right hands \\(\\mathbf{I}_{rh}\\)." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.107, + 0.892, + 0.166 + ], + "angle": 0, + "content": "Hand Pose Estimation. Given the cropped image \\(\\mathbf{I}_{lh}\\) or \\(\\mathbf{I}_{rh}\\), we regress the 3D hand poses \\(\\hat{\\mathbf{J}}_{lh}^{loc}\\) and \\(\\hat{\\mathbf{J}}_{rh}^{loc}\\) with the Hand4Whole [34] network, which is fine-tuned on our Ego-FullBody dataset." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.168, + 0.893, + 0.424 + ], + "angle": 0, + "content": "Integration of Body and Hand Positions. It is not straightforward to integrate the hand poses with the body pose in the egocentric camera view primarily due to the fisheye camera's perspective effects. Take the left hand as an example. Following Step 3 in Sec. 3.1.1, we establish a local coordinate system on the tangent plane of the left-hand image with XYZ axes as follows: \\( x:\\mathbf{v}_{lh}^{x};y:\\mathbf{v}_{lh}^{y};z:\\mathbf{v}_{lh}^{z} \\). We define a rotation matrix, denoted as \\( \\mathbf{R} \\), that represents the transformation between the root coordinate system and the local coordinate system on the tangent plane. The estimated hand pose is first rotated with the rotation matrix \\( \\hat{\\mathbf{J}}_{lh} = \\mathbf{R}\\hat{\\mathbf{J}}_{lh}^{loc} \\) and then translated to align the wrist location of the human body. This same process is also applied to the right hand to get the right hand pose \\( \\hat{\\mathbf{J}}_{rh} \\). The whole-body joints \\( \\hat{\\mathbf{J}} \\) are obtained by combining \\( \\hat{\\mathbf{J}}_b \\), \\( \\hat{\\mathbf{J}}_{lh} \\), and \\( \\hat{\\mathbf{J}}_{rh} \\). The uncertainty of whole-body joints \\( \\hat{\\mathbf{U}} \\) is also obtained from the maximal value of the 3D heatmap in pose estimation modules." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.432, + 0.813, + 0.447 + ], + "angle": 0, + "content": "3.2. Diffusion-Based Motion Refinement" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.455, + 0.892, + 0.56 + ], + "angle": 0, + "content": "We notice that the single-frame estimations in Sec. 3.1 suffer from inaccuracies and temporal instabilities. In this section, we propose a diffusion-based motion refinement method to tackle this problem. We first learn the whole-body motion prior with the motion diffusion model in Sec. 3.2.1 and then introduce an uncertainty-aware zero-shot motion refinement method in Sec. 3.2.2." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.579, + 0.813, + 0.594 + ], + "angle": 0, + "content": "3.2.1 Whole-Body Motion Diffusion Model" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.602, + 0.892, + 0.694 + ], + "angle": 0, + "content": "We follow DDPM [18] as our diffusion approach to capture the whole-body motion prior \\( q(\\mathbf{x}) \\). DDPM learns a distribution of whole-body motion \\( \\mathbf{x} \\) through a forward diffusion process and an inverse denoising process. The forward diffusion process is a Markov process of adding Gaussian noise over \\( t \\in \\{0, 1, \\dots, T - 1\\} \\) steps:" + }, + { + "type": "equation", + "bbox": [ + 0.564, + 0.701, + 0.891, + 0.719 + ], + "angle": 0, + "content": "\\[\nq \\left(\\mathbf {x} _ {t} \\mid \\mathbf {x} _ {t - 1}\\right) = \\mathcal {N} \\left(\\sqrt {\\alpha_ {t}} \\mathbf {x} _ {t - 1}, (1 - \\alpha_ {t}) I\\right) \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.726, + 0.892, + 0.771 + ], + "angle": 0, + "content": "where \\(\\mathbf{x}_t\\) denotes the whole-body motion sequence at step \\(t\\), the variance \\((1 - \\alpha_{t}) \\in (0,1]\\) denotes a constant hyperparameter increases with \\(t\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.772, + 0.892, + 0.877 + ], + "angle": 0, + "content": "The inverse process uses a denoising network \\( D(\\cdot) \\) to remove the added Gaussian noise at each time step \\( t \\). Here we use the transformer-based framework in EDGE [48] as the motion-denoising network \\( D(\\cdot) \\). We follow Ramesh et al.'s work [38] to make the network predict the original signal itself, i.e. \\( \\hat{\\mathbf{x}}_0 = D(\\mathbf{x}_t,t) \\) and train it with the simple objective [18]:" + }, + { + "type": "equation", + "bbox": [ + 0.527, + 0.884, + 0.892, + 0.904 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {s i m p l e}} = E _ {\\mathbf {x} _ {0} \\sim q (\\mathbf {x} _ {0}), t \\sim [ 1, T ]} \\left[ \\| \\mathbf {x} _ {0} - D (\\mathbf {x} _ {t}, t) \\| _ {2} ^ {2} \\right] \\tag {4}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.946, + 0.51, + 0.957 + ], + "angle": 0, + "content": "781" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.077, + 0.092, + 0.405, + 0.106 + ], + "angle": 0, + "content": "3.2.2 Uncertainty-Aware Motion Refinement" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.115, + 0.473, + 0.298 + ], + "angle": 0, + "content": "Given the learned whole-body motion prior, we leverage the uncertainty value for each pose to guide the diffusion denoising process with the classifier-guided diffusion sampling [12]. Given an initial sequence of whole-body pose estimation \\(\\mathbf{x}_e = \\{\\hat{\\mathbf{J}}_i\\}\\) and the uncertainty value for each pose \\(\\mathbf{u} = \\{\\hat{\\mathbf{U}}_i\\}\\), where \\(i\\) denotes the \\(i\\)th pose in the sequence, we keep the joints with low uncertainty but use the diffusion model to generate joints with high uncertainty conditioned on the low-uncertainty joints. Specifically, in the \\(t\\)th sampling step of the diffusion process, the denoising network predicts \\(\\hat{\\mathbf{x}}_0 = D(\\mathbf{x}_t,t)\\), which is noised back to \\(\\mathbf{x}_{t - 1}\\) by sampling from the Gaussian distribution:" + }, + { + "type": "equation", + "bbox": [ + 0.157, + 0.307, + 0.47, + 0.325 + ], + "angle": 0, + "content": "\\[\n\\mathbf {x} _ {t - 1} \\sim \\mathcal {N} \\left(\\hat {\\mathbf {x}} _ {0} + \\mathbf {w} \\left(\\mathbf {x} _ {e} - \\hat {\\mathbf {x}} _ {0}\\right), \\Sigma_ {t}\\right) \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.333, + 0.473, + 0.504 + ], + "angle": 0, + "content": "where \\(\\Sigma_{t}\\) is a scheduled Gaussian distribution in DDPM [18] and \\(\\mathbf{w}\\) controls the weight of a specific joint between the predicted motion \\(\\hat{\\mathbf{x}}_0\\) and the estimated motion \\(\\mathbf{x}_e\\). Generally, we expect \\(\\mathbf{w} \\rightarrow \\overrightarrow{0}\\) when \\(t \\rightarrow 0\\) such that the temporal stability is guaranteed through the generation of the denoising process, and \\(\\mathbf{w} \\rightarrow \\overrightarrow{1}\\) when \\(t \\rightarrow T\\) such that the denoising process is initialized by the estimated motion \\(\\mathbf{x}_e\\). We also expect that \\(w_{ij} = \\mathbf{w}[i][j]\\), which is the weight of \\(j\\)th joints in the \\(i\\)th pose, is smaller when the uncertainty value \\(u_{ij} = \\mathbf{u}[i][j]\\) of the \\(j\\)th joints in the \\(i\\)th pose is large. Based on this requirement, we design \\(\\mathbf{w}\\) as:" + }, + { + "type": "equation", + "bbox": [ + 0.185, + 0.512, + 0.47, + 0.539 + ], + "angle": 0, + "content": "\\[\n\\mathbf {w} = 1 / \\left(1 + e ^ {- k (t - T \\mathbf {u})}\\right) \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.546, + 0.47, + 0.624 + ], + "angle": 0, + "content": "where \\(T\\) is the overall diffusion steps, \\(k\\) is a hyperparameter which is empirically set to 0.1. From the experimental results in Sec. 5, we demonstrate the effectiveness of uncertain-aware motion refinement and our uncertainty-guided diffusion sampling strategy." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.634, + 0.303, + 0.651 + ], + "angle": 0, + "content": "4. EgoWholeBody Dataset" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.66, + 0.47, + 0.856 + ], + "angle": 0, + "content": "In this section, we introduce EgoWholeBody, a large-scale high-quality synthetic dataset built for the task of egocentric whole-body motion capture. The EgoWholeBody dataset is organized into two sections. The first part, containing over 700k frames, is rendered with 14 different rigged Renderpeople [3] models driven by 2367 Mixamo [2] motion sequences. The second part focuses on hand motions and contains 170k frames with the SMPL-X model. This data is constructed from 24 different shapes and textures, driven by 262 motion sequences selected from the GRAB [46] and TCDHandMocap dataset [20]. We also created synthetic test sequences, which include 133k images rendered with 3 Renderpeople models and Mixamo motions." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.471, + 0.902 + ], + "angle": 0, + "content": "During the rendering process, we first attach a virtual fisheye camera to the forehead of human body models and render the images, semantic labels, and depth map with" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.893, + 0.138 + ], + "angle": 0, + "content": "Blender [1]. Our dataset is larger and more diverse than previous egocentric training datasets—see Sec. 10 in the supplementary material for a detailed comparison." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.152, + 0.634, + 0.169 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.178, + 0.787, + 0.193 + ], + "angle": 0, + "content": "5.1. Datasets and Evaluation Metrics" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.2, + 0.893, + 0.322 + ], + "angle": 0, + "content": "Training Datasets. To train our body pose estimation module (Sec. 3.1.1 and Sec. 3.1.2), we use our EgoWholeBody dataset and the EgoPW dataset [50]. Additionally, the EgoWholeBody dataset is used to train the hand pose estimation module in Sec. 3.1.3. For training the whole-body diffusion model (Sec. 3.2), we utilize a combined motion capture dataset that includes EgoBody [62], Mixamo [2], TCDHandMocap dataset [20] and GRAB dataset [46]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.322, + 0.894, + 0.504 + ], + "angle": 0, + "content": "Evaluation Datasets. In our experiment, we evaluate our methods on four datasets: the GlobalEgoMocap test datasets [50], the \\(\\mathrm{Mo}^2\\mathrm{Cap}^2\\) test dataset [54], the SceneEgo test dataset [52] and out EgoWholeBody test dataset. The details of the datasets are shown in Sec. 12 of supplementary materials. Note that evaluating whole-body poses requires accurate annotations for human hands, which is absent in real-world datasets. To resolve the issue, we request the multi-view videos of the SceneEgo test dataset [52] from the authors and use a multi-view motion capture system to obtain the hand motion. The hand pose annotations will be made publicly available." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.504, + 0.894, + 0.656 + ], + "angle": 0, + "content": "Evaluation Metrics. To evaluate the precision of human body poses on the SceneEgo test dataset [52], we use MPJPE and PA-MPJPE. For the GlobalEgoMocap test dataset [50] and \\(\\mathrm{Mo}^2\\mathrm{Cap}^2\\) test dataset [54], where egocentric camera poses are unavailable, we evaluate PA-MPJPE and BA-MPJPE. For hand pose accuracy, we align the predicted and ground truth hand poses at the root position, followed by computing MPJPE and PA-MPJPE. Detailed explanations of these metrics are in Sec. 11 of the supplementary materials. All reported metrics are in millimeters." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.666, + 0.89, + 0.682 + ], + "angle": 0, + "content": "5.2. Comparisons on Whole-Body Pose Estimation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.689, + 0.895, + 0.902 + ], + "angle": 0, + "content": "For a fair comparison with existing methods focusing solely on body or hand pose, we split our evaluation into two parts, reporting results of body poses in Table 1 and hand pose in Table 2. We first compare the accuracy of the human body poses with state-of-the-art methods, including EgoPW [51] and SceneEgo [52], on EgoWholeBody and SceneEgo [52] test datasets. The comparison with more previous methods [47, 50, 54] and on more evaluation datasets [50, 54] are shown in Sec. 7 of the supplementary materials. Since our motion refinement method incorporates random Gaussian noise, we generate five samples and calculate the average MPJPE values. The standard deviation is low (\\(< 0.01\\mathrm{mm}\\)) and is discussed in Sec. 13 of supplementary materials. Results are presented in Table 1, where our single-frame re" + }, + { + "type": "page_number", + "bbox": [ + 0.487, + 0.946, + 0.511, + 0.957 + ], + "angle": 0, + "content": "782" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.084, + 0.089, + 0.167, + 0.213 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.111, + 0.214, + 0.138, + 0.223 + ], + "angle": 0, + "content": "Input" + }, + { + "type": "image", + "bbox": [ + 0.174, + 0.094, + 0.228, + 0.212 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.186, + 0.214, + 0.221, + 0.223 + ], + "angle": 0, + "content": "EgoPW" + }, + { + "type": "image", + "bbox": [ + 0.255, + 0.091, + 0.306, + 0.211 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.26, + 0.213, + 0.305, + 0.223 + ], + "angle": 0, + "content": "SceneEgo" + }, + { + "type": "image", + "bbox": [ + 0.334, + 0.091, + 0.388, + 0.212 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.334, + 0.213, + 0.388, + 0.223 + ], + "angle": 0, + "content": "Ours-Single" + }, + { + "type": "image", + "bbox": [ + 0.411, + 0.092, + 0.464, + 0.212 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.409, + 0.213, + 0.471, + 0.223 + ], + "angle": 0, + "content": "Ours-Refined" + }, + { + "type": "image", + "bbox": [ + 0.486, + 0.089, + 0.57, + 0.213 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.515, + 0.214, + 0.541, + 0.223 + ], + "angle": 0, + "content": "Input" + }, + { + "type": "image", + "bbox": [ + 0.578, + 0.091, + 0.633, + 0.212 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.59, + 0.213, + 0.623, + 0.223 + ], + "angle": 0, + "content": "EgoPW" + }, + { + "type": "image", + "bbox": [ + 0.659, + 0.091, + 0.716, + 0.211 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.663, + 0.213, + 0.708, + 0.223 + ], + "angle": 0, + "content": "SceneEgo" + }, + { + "type": "image", + "bbox": [ + 0.738, + 0.091, + 0.801, + 0.211 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.741, + 0.213, + 0.795, + 0.223 + ], + "angle": 0, + "content": "Ours-Single" + }, + { + "type": "image", + "bbox": [ + 0.818, + 0.091, + 0.882, + 0.211 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.816, + 0.213, + 0.877, + 0.223 + ], + "angle": 0, + "content": "Ours-Refined" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.234, + 0.894, + 0.276 + ], + "angle": 0, + "content": "Figure 4. Qualitative comparison on human body pose estimations between our methods and the state-of-the-art egocentric pose estimation methods on in-the-studio (left column) and in-the-wild scenes (right column). The red skeleton is the ground truth while the green skeleton is the predicted pose. Our methods predict more accurate body poses compared with EgoPW [51] and SceneEgo [52]." + }, + { + "type": "image", + "bbox": [ + 0.084, + 0.279, + 0.17, + 0.409 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.111, + 0.412, + 0.142, + 0.422 + ], + "angle": 0, + "content": "Input" + }, + { + "type": "image", + "bbox": [ + 0.179, + 0.279, + 0.281, + 0.409 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.203, + 0.412, + 0.27, + 0.422 + ], + "angle": 0, + "content": "Hand Image" + }, + { + "type": "image", + "bbox": [ + 0.289, + 0.282, + 0.354, + 0.407 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.283, + 0.412, + 0.355, + 0.422 + ], + "angle": 0, + "content": "Hand4Whole" + }, + { + "type": "image", + "bbox": [ + 0.374, + 0.282, + 0.431, + 0.407 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.368, + 0.412, + 0.433, + 0.422 + ], + "angle": 0, + "content": "Ours-Single" + }, + { + "type": "image", + "bbox": [ + 0.455, + 0.281, + 0.511, + 0.407 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.446, + 0.412, + 0.518, + 0.422 + ], + "angle": 0, + "content": "Ours-Refined" + }, + { + "type": "image_caption", + "bbox": [ + 0.535, + 0.323, + 0.548, + 0.37 + ], + "angle": 0, + "content": "Rrnnnne" + }, + { + "type": "image", + "bbox": [ + 0.556, + 0.28, + 0.641, + 0.409 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.565, + 0.412, + 0.631, + 0.422 + ], + "angle": 0, + "content": "Hand Image" + }, + { + "type": "image", + "bbox": [ + 0.649, + 0.282, + 0.717, + 0.407 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.644, + 0.412, + 0.717, + 0.422 + ], + "angle": 0, + "content": "Hand4Whole" + }, + { + "type": "image", + "bbox": [ + 0.725, + 0.281, + 0.799, + 0.407 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.73, + 0.412, + 0.794, + 0.422 + ], + "angle": 0, + "content": "Ours-Single" + }, + { + "type": "image", + "bbox": [ + 0.81, + 0.281, + 0.881, + 0.409 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.81, + 0.412, + 0.881, + 0.422 + ], + "angle": 0, + "content": "Ours-Refined" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.434, + 0.893, + 0.476 + ], + "angle": 0, + "content": "Figure 5. Qualitative comparison on human hand pose estimations between our methods and the state-of-the-art third-view pose estimation methods. Our single-view and refined hand poses are more accurate than the poses from Hand4Whole [34] method. The red skeleton is the ground truth while the green skeleton is the predicted pose." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.491, + 0.469, + 0.581 + ], + "angle": 0, + "content": "sults are labeled as \"Ours-Single\" and our refinement results are labeled as \"Ours-Refined\". Our single-frame body pose estimation method outperforms all previous methods by a large margin. Our diffusion-based motion refinement method can further improve the accuracy of body poses estimated by the single-frame methods." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.582, + 0.469, + 0.703 + ], + "angle": 0, + "content": "Note that previous methods [47, 50-52, 54] use training datasets different from each other. For a fair comparison, we re-train previous methods with our training datasets in Sec. 5.1 and show the results with “*” in Table 1. This retraining led to significant improvements across all previous methods, demonstrating our dataset's broad applicability. However, these methods still underperformed compared to ours, highlighting our approach's superiority." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.704, + 0.469, + 0.869 + ], + "angle": 0, + "content": "To evaluate the accuracy of our hand pose estimation method, we first crop the hand images with the hand detection method in Sec. 3.1.3. Then we show the results of our single-frame hand pose estimation (labeled as \"Ours-Single\") and whole-body motion refinement methods (labeled as \"Ours-Refined\") in Table 2. Our single-frame hand pose estimation method outperforms the state-of-the-art method Hands4Whole [34], demonstrating the effectiveness of training the network on our EgoWholeBody dataset. Our whole-body motion refinement method can also enhance the accuracy of hand motion." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.469, + 0.901 + ], + "angle": 0, + "content": "For a qualitative comparison, we compare the body and hand poses of our method with existing methods on the" + }, + { + "type": "table", + "bbox": [ + 0.505, + 0.488, + 0.89, + 0.678 + ], + "angle": 0, + "content": "
MethodMPJPEPA-MPJPE
SceneEgo test dataset [52]
EgoPW [51]189.6105.3
SceneEgo [52]118.592.75
EgoPW* [51]90.9664.33
SceneEgo* [52]89.0670.10
Ours-Single64.1950.06
Ours-Refined57.5946.55
EgoWholeBody test dataset
EgoPW* [51]84.2163.02
SceneEgo* [52]87.5769.46
Ours-Single66.2843.14
Ours-Refined60.3240.35
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.688, + 0.892, + 0.743 + ], + "angle": 0, + "content": "Table 1. Egocentric human body pose accuracy of our method on SceneEgo test datasets and EgoWholeBody test dataset. Our method outperforms all previous state-of-the-art methods. * denotes the method trained with the datasets in Sec. 5.1." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.759, + 0.892, + 0.836 + ], + "angle": 0, + "content": "SceneEgo dataset and the in-the-wild EgoPW [51] evaluation sequences. The results are shown in Fig. 4 and Fig. 5, showing that our method can predict high-quality whole-body poses from an egocentric camera. Please refer to our supplementary video for more qualitative evaluation results." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.847, + 0.653, + 0.863 + ], + "angle": 0, + "content": "5.3. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.892, + 0.901 + ], + "angle": 0, + "content": "EgoWholeBody Dataset. Compared to existing egocentric datasets, our EgoWholeBody dataset contains diverse body" + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.946, + 0.511, + 0.957 + ], + "angle": 0, + "content": "783" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.082, + 0.089, + 0.467, + 0.221 + ], + "angle": 0, + "content": "
MethodMPJPEPA-MPJPE
SceneEgo test dataset [52]
Hand4Whole [34]49.6613.85
Ours-Single23.639.59
Ours-Refined19.379.05
EgoWholeBody test dataset
Hand4Whole [34]52.8535.04
Ours-Single33.1019.68
Ours-Refined28.2914.51
" + }, + { + "type": "table_caption", + "bbox": [ + 0.078, + 0.231, + 0.468, + 0.259 + ], + "angle": 0, + "content": "Table 2. Egocentric hand pose accuracy of our method. Our method outperforms the Hand4Whole [34] on both datasets." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.267, + 0.466, + 0.483 + ], + "angle": 0, + "content": "
MethodMPJPEPA-MPJPE
Body Pose Results
w/o EgoWholeBody75.1058.62
w/o FisheyeViT67.3653.44
w/ Mo2Cap2 [54] head87.4765.10
w/ xR-egopose [47] head116.595.78
w/ SceneEgo [52] head77.7362.69
Ours-Single64.1950.06
w/ GlobalEgoMocap†69.8356.73
w/o uncert. guidance†62.1648.40
Only body diffusion58.9547.03
Ours-Refined†57.5946.55
Hand Pose Results
Only hand diffusion21.699.24
Ours-Refined19.379.05
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.494, + 0.468, + 0.523 + ], + "angle": 0, + "content": "Table 3. Ablation Study on SceneEgo test dataset [52]. \\( {}^{ \\dagger } \\) denotes the temporal-based method." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.539, + 0.468, + 0.718 + ], + "angle": 0, + "content": "and hand motions, larger quantity of images, and higher image quality. We show this by training our body pose estimation network without our dataset, using the \\(\\mathrm{Mo}^2\\mathrm{Cap}^2\\) [54] and EgoPW [51] training dataset. The results, labeled as \"w/o EgoWholeBody\" in Table 3, show that performance without the EgoWholeBody dataset is inferior to our proposed method. This highlights that training with our EgoW-holeBody dataset enhances the performance of the pose estimation method. We also compare this result with existing methods on the SceneEgo test set (Table 1). Trained without EgoWholeBody, our approach still outperforms previous methods, showing the effectiveness of our method." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.72, + 0.468, + 0.869 + ], + "angle": 0, + "content": "FisheyeViT and Pose Regressor with Pixel-Aligned 3D Heatmap. To assess the individual contributions of FisheyeViT and the pixel-aligned 3D heatmap in our single-frame pose estimation pipeline, we perform experiments to measure their impact on the overall performance. First, we substitute the FisheyeViT module in our single-frame pose estimation method to ViT [13]. The result is shown in \"w/o FisheyeViT\" in Table 3 and it is worse than our full method. This demonstrates the effectiveness of FisheyeViT in addressing fisheye distortion and feature extraction." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.468, + 0.901 + ], + "angle": 0, + "content": "Next, we analyze the performance of the single-frame pose estimation network when substituting our pose regress-" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.197 + ], + "angle": 0, + "content": "sor based on pixel-aligned 3D heatmap with the pose estimation heads of previous works [47, 52, 54]. The results of the three experiments, labeled as \"w/ Mo²Cap² head\", \"w/ xR-egopose head\" and \"w/ SceneEgo head\", show a performance drop compared to our full method. This emphasizes the crucial role of the pixel-aligned 3D heatmap in accurately estimating egocentric 3D body joint positions." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.198, + 0.892, + 0.318 + ], + "angle": 0, + "content": "Diffusion-based Motion Refinement. We assess the effectiveness of our diffusion-based motion refinement with the following experiments: First, we compare the performance of our diffusion-based motion refinement with GlobalEgoMocap [50] by applying the GlobalEgoMocap optimizer on the single-frame body pose estimation results. The result, labeled as \"w GlobalEgoMocap\" in Table 3, indicates that our refinement method outperforms GlobalEgoMocap." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.318, + 0.892, + 0.453 + ], + "angle": 0, + "content": "Second, we remove the uncertainty-aware guidance in the motion refinement. Instead, we use fixed Gaussian denoising steps to refine the motion. The result \"w/o uncert. guidance\" in Table 3, shows that our uncertainty-aware refinement method performs better. Our approach relies on the uncertainty values for each joint, using low-uncertainty joints to guide the generation of high-uncertainty joints. This helps reduce errors in joint predictions caused by egocentric self-occlusion, leading to improved results." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.455, + 0.892, + 0.59 + ], + "angle": 0, + "content": "Third, we replace our whole-body motion diffusion model with the separate human body and left/right-hand diffusion models and show the accuracy of refined body and hand motion in \"Only body diffusion\" and \"Only hand diffusion\" in Table 3. From the results, we observe improvements in the accuracy of motion refined by our whole-body diffusion method, proving that learning the whole-body motion prior can help both the refinement of the body and hand motion by learning the correlation between them." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.604, + 0.618, + 0.62 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.629, + 0.892, + 0.841 + ], + "angle": 0, + "content": "In this work, we have introduced an innovative approach to capture egocentric whole-body human motion. Our method comprises a single-frame-based whole-body pose estimation process, which includes FisheyeViT and pixel-aligned 3D heatmap representations. To enhance the initial whole-body pose estimates, we have integrated an uncertainty-aware diffusion-based motion refinement technique. Our experimental results demonstrate that both our single-frame method and the temporal-based method surpass all existing state-of-the-art techniques in terms of both quality and accuracy. Looking ahead, we see the potential for extending the applications of FisheyeViT to other vision tasks involving fisheye cameras. Future work could also involve incorporating facial expressions in whole-body motion capture." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.841, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Acknowledgments This project was supported by the Saarbrücken Research Center for Visual Computing, Interaction and AI. Christian Theobalt was supported by ERC Consolidator Grant 4DReply (770784)." + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.945, + 0.511, + 0.956 + ], + "angle": 0, + "content": "784" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.105 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.116, + 0.328, + 0.129 + ], + "angle": 0, + "content": "[1] Blender. https://www.blender.org. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.131, + 0.343, + 0.143 + ], + "angle": 0, + "content": "[2] Mixamo. https://www MIXamo.com. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.145, + 0.383, + 0.158 + ], + "angle": 0, + "content": "[3] Renderpeople. https://renderpeople.com. 6, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.16, + 0.469, + 0.228 + ], + "angle": 0, + "content": "[4] Hiroyasu Akada, Jian Wang, Soshi Shimada, Masaki Takahashi, Christian Theobalt, and Vladislav Golyanik. Unrealego: A new dataset for robust egocentric 3d human motion capture. In European Conference on Computer Vision, pages 1-17. Springer, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.23, + 0.469, + 0.284 + ], + "angle": 0, + "content": "[5] Hiroyasu Akada, Jian Wang, Vladislav Golyanik, and Christian Theobalt. 3d human pose perception from egocentric stereo videos. In 37th IEEE/CVF Conference on Computer Vision and Pattern Recognition. IEEE, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.286, + 0.469, + 0.354 + ], + "angle": 0, + "content": "[6] Zhongang Cai, Wanqi Yin, Ailing Zeng, Chen Wei, Qingping Sun, Yanjun Wang, Hui En Pang, Haiyi Mei, Mingyuan Zhang, Lei Zhang, et al. Smpler-x: Scaling up expressive human pose and shape estimation. arXiv preprint arXiv:2309.17448, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.355, + 0.469, + 0.438 + ], + "angle": 0, + "content": "[7] Young-Woon Cha, True Price, Zhen Wei, Xinran Lu, Nicholas Rewkowski, Rohan Chabra, Zihe Qin, Hyounghun Kim, Zhaoqi Su, Yebin Liu, et al. Towards fully mobile 3d face, body, and environment capture using only head-worn cameras. IEEE transactions on visualization and computer graphics, 24(11):2993-3004, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.44, + 0.469, + 0.493 + ], + "angle": 0, + "content": "[8] Jeongjun Choi, Dongseok Shim, and H Jin Kim. Diffupose: Monocular 3d human pose estimation via denoising diffusion probabilistic model. arXiv preprint arXiv:2212.02796, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.496, + 0.469, + 0.579 + ], + "angle": 0, + "content": "[9] Vasileios Choutas, Georgios Pavlakos, Timo Bolkart, Dimitrios Tzionas, and Michael J Black. Monocular expressive body regression through body-driven attention. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part X 16, pages 20-40. Springer, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.58, + 0.47, + 0.648 + ], + "angle": 0, + "content": "[10] Hai Ci, Mingdong Wu, Wentao Zhu, Xiaoxuan Ma, Hao Dong, Fangwei Zhong, and Yizhou Wang. Gfpose: Learning 3d human pose prior with gradient fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4800-4810, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.65, + 0.469, + 0.718 + ], + "angle": 0, + "content": "[11] Benjamin Coors, Alexandru Paul Condurache, and Andreas Geiger. Spherenet: Learning spherical representations for detection and classification in omnidirectional images. In Proceedings of the European conference on computer vision (ECCV), pages 518-533, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.72, + 0.469, + 0.761 + ], + "angle": 0, + "content": "[12] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in neural information processing systems, 34:8780-8794, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.762, + 0.469, + 0.844 + ], + "angle": 0, + "content": "[13] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020.4, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.846, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[14] Yuming Du, Robin Kips, Albert Pumarola, Sebastian Starke, Ali Thabet, and Artsiom Sanakoyeu. Avatars grow legs: Generating smooth human motion from sparse tracking inputs with diffusion model. In Proceedings of the IEEE/CVF" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.116, + 0.47, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.094, + 0.892, + 0.12 + ], + "angle": 0, + "content": "Conference on Computer Vision and Pattern Recognition, pages 481-490, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.122, + 0.892, + 0.189 + ], + "angle": 0, + "content": "[15] Yao Feng, Vasileios Choutas, Timo Bolkart, Dimitrios Tzionas, and Michael J Black. Collaborative regression of expressive bodies using moderation. In 2021 International Conference on 3D Vision (3DV), pages 792-804. IEEE, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.193, + 0.892, + 0.247 + ], + "angle": 0, + "content": "[16] Lin Geng Foo, Jia Gong, Hossein Rahmani, and Jun Liu. Distribution-aligned diffusion for human mesh recovery. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9221-9232, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.25, + 0.892, + 0.317 + ], + "angle": 0, + "content": "[17] Jia Gong, Lin Geng Foo, Zhipeng Fan, Qiuhong Ke, Hossein Rahmani, and Jun Liu. Diffpose: Toward more reliable 3d pose estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13041-13051, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.32, + 0.892, + 0.361 + ], + "angle": 0, + "content": "[18] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 2, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.363, + 0.892, + 0.418 + ], + "angle": 0, + "content": "[19] Karl Holmquist and Bastian Wandt. Diffpose: Multi-hypothesis human pose estimation using diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15977-15987, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.42, + 0.892, + 0.489 + ], + "angle": 0, + "content": "[20] Ludovic Hoyet, Kenneth Ryall, Rachel McDonnell, and Carol O'Sullivan. Sleight of hand: perception of finger motion from reduced marker sets. In Proceedings of the ACM SIGGRAPH symposium on interactive 3D graphics and games, pages 79-86, 2012. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.49, + 0.892, + 0.545 + ], + "angle": 0, + "content": "[21] Hao Jiang and Kristen Grauman. Seeing invisible poses: Estimating 3d body pose from egocentric video. In 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 3501-3509. IEEE, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.547, + 0.892, + 0.602 + ], + "angle": 0, + "content": "[22] Zhongyu Jiang, Zhuoran Zhou, Lei Li, Wenhao Chai, ChengYen Yang, and Jenq-Neng Hwang. Back to optimization: Diffusion-based zero-shot 3d human pose estimation. arXiv preprint arXiv:2307.03833, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.604, + 0.892, + 0.644 + ], + "angle": 0, + "content": "[23] Taeho Kang, Kyungjin Lee, Jinrui Zhang, and Youngki Lee. Ego3dpose: Capturing 3d cues from binocular egocentric views. arXiv preprint arXiv:2309.11962, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.647, + 0.892, + 0.673 + ], + "angle": 0, + "content": "[24] David G Kendall. A survey of the statistical theory of shape. Statistical Science, 4(2):87-99, 1989. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.676, + 0.892, + 0.731 + ], + "angle": 0, + "content": "[25] Jiaman Li, Karen Liu, and Jiajun Wu. Ego-body pose estimation via ego-head pose estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17142-17151, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.733, + 0.892, + 0.801 + ], + "angle": 0, + "content": "[26] Yuyan Li, Yuliang Guo, Zhixin Yan, Xinyu Huang, Ye Duan, and Liu Ren. Omnifusion: 360 monocular depth estimation via geometry-aware fusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2801-2810, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.803, + 0.892, + 0.87 + ], + "angle": 0, + "content": "[27] Jing Lin, Ailing Zeng, Haoqian Wang, Lei Zhang, and Yu Li. One-stage 3d whole-body mesh recovery with component aware transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21159-21168, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.873, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[28] Yuxuan Liu, Jianxin Yang, Xiao Gu, Yao Guo, and Guang-Zhong Yang. Ego+ x: An egocentric vision system for global" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.094, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.487, + 0.946, + 0.511, + 0.957 + ], + "angle": 0, + "content": "785" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.11, + 0.092, + 0.469, + 0.147 + ], + "angle": 0, + "content": "3d human pose estimation and social interaction characterization. In 2022 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 5271-5277. IEEE, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.15, + 0.469, + 0.205 + ], + "angle": 0, + "content": "[29] Yuxuan Liu, Jianxin Yang, Xiao Gu, Yijun Chen, Yao Guo, and Guang-Zhong Yang. Egofish3d: Egocentric 3d pose estimation from a fisheye camera via self-supervised learning. IEEE Transactions on Multimedia, 2023. 1, 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.208, + 0.469, + 0.276 + ], + "angle": 0, + "content": "[30] Yuxuan Liu, Jianxin Yang, Xiao Gu, Yao Guo, and GuangZhong Yang. Egohmr: Egocentric human mesh recovery via hierarchical latent diffusion model. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 9807-9813. IEEE, 2023. 1, 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.279, + 0.469, + 0.332 + ], + "angle": 0, + "content": "[31] Zhengyi Luo, Ryo Hachiuma, Ye Yuan, and Kris Kitani. Dynamics-regulated kinematic policy for egocentric pose estimation. Advances in Neural Information Processing Systems, 34:25019-25032, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.335, + 0.469, + 0.404 + ], + "angle": 0, + "content": "[32] Christen Millerdurai, Hiroyasu Akada, Jian Wang, Diogo Luvizon, Christian Theobalt, and Vladislav Golyanik. Evento3d: 3d human motion capture from egocentric event streams. In 37th IEEE/CVF Conference on Computer Vision and Pattern Recognition. IEEE, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.406, + 0.469, + 0.476 + ], + "angle": 0, + "content": "[33] Gyeongsik Moon, Ju Yong Chang, and Kyoung Mu Lee. V2v-posenet: Voxel-to-voxel prediction network for accurate 3d hand and human pose estimation from a single depth map. In Proceedings of the IEEE conference on computer vision and pattern Recognition, pages 5079–5088, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.478, + 0.469, + 0.532 + ], + "angle": 0, + "content": "[34] Gyeongsik Moon, Hongsuk Choi, and Kyoung Mu Lee. Accurate 3d hand pose estimation for whole-body 3d human mesh estimation. In Computer Vision and Pattern Recognition Workshop (CVPRW), 2022. 4, 5, 7, 8, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.535, + 0.47, + 0.603 + ], + "angle": 0, + "content": "[35] Evonne Ng, Donglai Xiang, Hanbyul Joo, and Kristen Grauman. You2me: Inferring body pose in egocentric video via first and second person interactions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9890-9900, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.606, + 0.469, + 0.687 + ], + "angle": 0, + "content": "[36] Jinman Park, Kimathi Kaai, Saad Hossain, Norikatsu Sumi, Sirisha Rambhatla, and Paul Fieguth. Domain-guided spatiotemporal self-attention for egocentric 3d pose estimation. In Proceedings of the 29th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, pages 1837–1849, 2023. 2, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.69, + 0.469, + 0.772 + ], + "angle": 0, + "content": "[37] Georgios Pavlakos, Vasileios Choutas, Nima Ghorbani, Timo Bolkart, Ahmed AA Osman, Dimitrios Tzionas, and Michael J Black. Expressive body capture: 3d hands, face, and body from a single image. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10975-10985, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.776, + 0.469, + 0.829 + ], + "angle": 0, + "content": "[38] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 1 (2):3, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.832, + 0.469, + 0.9 + ], + "angle": 0, + "content": "[39] Helge Rhodin, Christian Richardt, Dan Casas, Eldar Insafutdinov, Mohammad Shafiei, Hans-Peter Seidel, Bernt Schiele, and Christian Theobalt. Egocap: egocentric marker-less motion capture with two fisheye cameras. ACM Transactions on Graphics (TOG), 35(6):1-11, 2016. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.161 + ], + "angle": 0, + "content": "[40] Yu Rong, Takaaki Shiratori, and Hanbyul Joo. Frankmocap: A monocular 3d whole-body pose estimation system via regression and integration. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1749-1759, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.163, + 0.892, + 0.232 + ], + "angle": 0, + "content": "[41] Davide Scaramuzza, Agostino Martinelli, and Roland Siegwart. A toolbox for easily calibrating omnidirectional cameras. In 2006 IEEE/RSJ International Conference on Intelligent Robots and Systems, pages 5695-5701. IEEE, 2006. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.233, + 0.892, + 0.3 + ], + "angle": 0, + "content": "[42] Wenkang Shan, Zhenhua Liu, Xinfeng Zhang, Zhao Wang, Kai Han, Shanshe Wang, Siwei Ma, and Wen Gao. Diffusion-based 3d human pose estimation with multi-hypothesis aggregation. arXiv preprint arXiv:2303.11579, 2023.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.302, + 0.892, + 0.357 + ], + "angle": 0, + "content": "[43] Soshi Shimada, Vladislav Golyanik, Weipeng Xu, and Christian Theobalt. Physcap: Physically plausible monocular 3d motion capture in real time. ACM Transactions on Graphics (ToG), 39(6):1-16, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.358, + 0.892, + 0.412 + ], + "angle": 0, + "content": "[44] Xiao Sun, Bin Xiao, Fangyin Wei, Shuang Liang, and Yichen Wei. Integral human pose regression. In Proceedings of the European conference on computer vision (ECCV), pages 529-545, 2018. 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.414, + 0.892, + 0.483 + ], + "angle": 0, + "content": "[45] Yu Sun, Tianyu Huang, Qian Bao, Wu Liu, Wenpeng Gao, and Yili Fu. Learning monocular mesh recovery of multiple body parts via synthesis. In ICASSP 2022-2022 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 2669-2673. IEEE, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.484, + 0.892, + 0.552 + ], + "angle": 0, + "content": "[46] Omid Taheri, Nima Ghorbani, Michael J Black, and Dimitrios Tzionas. Grab: A dataset of whole-body human grasping of objects. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part IV 16, pages 581-600. Springer, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.554, + 0.892, + 0.607 + ], + "angle": 0, + "content": "[47] Denis Tomè, Patrick Peluse, Lourdes Agapito, and Hernán Badino. xr-egopose: Egocentric 3d human pose from an HMD camera. In ICCV, pages 7727-7737, 2019. 1, 2, 5, 6, 7, 8, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.609, + 0.892, + 0.663 + ], + "angle": 0, + "content": "[48] Jonathan Tseng, Rodrigo Castellon, and Karen Liu. Edge: Editable dance generation from music. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 448-458, 2023. 5, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.665, + 0.892, + 0.747 + ], + "angle": 0, + "content": "[49] Jingdong Wang, Ke Sun, Tianheng Cheng, Borui Jiang, Chaorui Deng, Yang Zhao, Dong Liu, Yadong Mu, Mingkui Tan, Xinggang Wang, et al. Deep high-resolution representation learning for visual recognition. IEEE transactions on pattern analysis and machine intelligence, 43(10):3349-3364, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.749, + 0.892, + 0.789 + ], + "angle": 0, + "content": "[50] Jian Wang, Lingjie Liu, Weipeng Xu, Kripasindhu Sarkar, and Christian Theobalt. Estimating egocentric 3d human pose in global space. ICCV, 2021. 1, 2, 6, 7, 8, 3, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.79, + 0.892, + 0.857 + ], + "angle": 0, + "content": "[51] Jian Wang, Lingjie Liu, Weipeng Xu, Kripasindhu Sarkar, Diogo Luvizon, and Christian Theobalt. Estimating egocentric 3d human pose in the wild with external weak supervision. In CVPR, pages 13157-13166, 2022. 1, 2, 6, 7, 8, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.86, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[52] Jian Wang, Diogo Luvizon, Weipeng Xu, Lingjie Liu, Kripasindhu Sarkar, and Christian Theobalt. Scene-aware egocentric 3d human pose estimation. In Proceedings of the" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.487, + 0.946, + 0.512, + 0.956 + ], + "angle": 0, + "content": "786" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.11, + 0.093, + 0.468, + 0.12 + ], + "angle": 0, + "content": "IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13031-13040, 2023. 1, 2, 5, 6, 7, 8, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.122, + 0.468, + 0.177 + ], + "angle": 0, + "content": "[53] Donglai Xiang, Hanbyul Joo, and Yaser Sheikh. Monocular total capture: Posing face, body, and hands in the wild. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10965-10974, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.179, + 0.468, + 0.248 + ], + "angle": 0, + "content": "[54] Weipeng Xu, Avishek Chatterjee, Michael Zollhöfer, Helge Rhodin, Pascal Fua, Hans-Peter Seidel, and Christian Theobalt. \\(\\mathrm{Mo}^2\\mathrm{cap}^2\\): Real-time mobile 3d motion capture with a cap-mounted fisheye camera. IEEE Trans. Vis. Comput. Graph., 25(5):2093-2101, 2019. 1, 2, 5, 6, 7, 8, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.251, + 0.468, + 0.305 + ], + "angle": 0, + "content": "[55] Yufei Xu, Jing Zhang, Qiming Zhang, and Dacheng Tao. Vitpose: Simple vision transformer baselines for human pose estimation. Advances in Neural Information Processing Systems, 35:38571-38584, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.308, + 0.468, + 0.362 + ], + "angle": 0, + "content": "[56] Dianyi Yang, Jiadong Tang, Yu Gao, Yi Yang, and Mengyin Fu. Sector patch embedding: An embedding module conforming to the distortion pattern of fisheye image. arXiv preprint arXiv:2303.14645, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.365, + 0.468, + 0.418 + ], + "angle": 0, + "content": "[57] Shangrong Yang, Chunyu Lin, Kang Liao, and Yao Zhao. Dual diffusion architecture for fisheye image rectification: Synthetic-to-real generalization. arXiv preprint arXiv:2301.11785, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.421, + 0.468, + 0.49 + ], + "angle": 0, + "content": "[58] Fanghua Yu, Xintao Wang, Mingdeng Cao, Gen Li, Ying Shan, and Chao Dong. Osrt: Omnidirectional image superresolution with distortion-aware transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13283-13292, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.492, + 0.468, + 0.533 + ], + "angle": 0, + "content": "[59] Ye Yuan and Kris Kitani. 3d ego-pose estimation via imitation learning. In Proceedings of the European Conference on Computer Vision (ECCV), pages 735–750, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.535, + 0.468, + 0.588 + ], + "angle": 0, + "content": "[60] Ye Yuan and Kris Kitani. Ego-pose estimation and forecasting as real-time pd control. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10082-10092, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.591, + 0.468, + 0.646 + ], + "angle": 0, + "content": "[61] Ye Yuan, Jiaming Song, Umar Iqbal, Arash Vahdat, and Jan Kautz. Physdiff: Physics-guided human motion diffusion model. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 16010-16021, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.648, + 0.468, + 0.716 + ], + "angle": 0, + "content": "[62] Siwei Zhang, Qianli Ma, Yan Zhang, Zhiyin Qian, Taein Kwon, Marc Pollefeys, Federica Bogo, and Siyu Tang. Ego-body: Human body shape and motion of interacting people from head-mounted devices. In European Conference on Computer Vision, pages 180–200. Springer, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.719, + 0.468, + 0.772 + ], + "angle": 0, + "content": "[63] Siwei Zhang, Qianli Ma, Yan Zhang, Sadegh Aliakbarian, Darren Cosker, and Siyu Tang. Probabilistic human mesh recovery in 3d scenes from egocentric views. arXiv preprint arXiv:2304.06024, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.775, + 0.468, + 0.843 + ], + "angle": 0, + "content": "[64] Yahui Zhang, Shaodi You, and Theo Gevers. Automatic calibration of the fisheye camera for egocentric 3d human pose estimation from a single image. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1772-1781, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.846, + 0.468, + 0.9 + ], + "angle": 0, + "content": "[65] Dongxu Zhao, Zhen Wei, Jisan Mahmud, and Jan-Michael Frahm. Egoglass: Egocentric-view human pose estimation from an eyeglass frame. In 2021 International Conference on 3D Vision (3DV), pages 32-41. IEEE, 2021. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.093, + 0.468, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.162 + ], + "angle": 0, + "content": "[66] Yuxiao Zhou, Marc Habermann, Ikhsanul Habibie, Ayush Tewari, Christian Theobalt, and Feng Xu. Monocular real-time full body capture with inter-part correlations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4811-4822, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.164, + 0.892, + 0.232 + ], + "angle": 0, + "content": "[67] Christian Zimmermann, Duygu Ceylan, Jimei Yang, Bryan Russell, Max Argus, and Thomas Brox. Freihand: A dataset for markerless capture of hand pose and shape from single rgb images. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 813-822, 2019. 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.232 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.487, + 0.946, + 0.511, + 0.956 + ], + "angle": 0, + "content": "787" + } + ] +] \ No newline at end of file diff --git a/2024/Egocentric Whole-Body Motion Capture with FisheyeViT and Diffusion-Based Motion Refinement/90ece1d3-5ea4-4b56-9fe8-5a7d9a524eef_origin.pdf b/2024/Egocentric Whole-Body Motion Capture with FisheyeViT and Diffusion-Based Motion Refinement/90ece1d3-5ea4-4b56-9fe8-5a7d9a524eef_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..5df581c830d9c6fcfabc4765fa93aa83803885a4 --- /dev/null +++ b/2024/Egocentric Whole-Body Motion Capture with FisheyeViT and Diffusion-Based Motion Refinement/90ece1d3-5ea4-4b56-9fe8-5a7d9a524eef_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0a406f79f0708b4b5b06137429e314eedd6292d1173caab62d1f1de9b779f34a +size 1630474 diff --git a/2024/Egocentric Whole-Body Motion Capture with FisheyeViT and Diffusion-Based Motion Refinement/full.md b/2024/Egocentric Whole-Body Motion Capture with FisheyeViT and Diffusion-Based Motion Refinement/full.md new file mode 100644 index 0000000000000000000000000000000000000000..85a46679233c96fae691ee9ccd22ce131c31d39e --- /dev/null +++ b/2024/Egocentric Whole-Body Motion Capture with FisheyeViT and Diffusion-Based Motion Refinement/full.md @@ -0,0 +1,397 @@ +# Egocentric Whole-Body Motion Capture with FisheyeViT and Diffusion-Based Motion Refinement + +Jian Wang $^{1,4}$ + +Kripasindhu Sarkar2 + +Zhe Cao2 + +Danhang Tang + +Diogo Luvizon1,4 + +Thabo Beeler2 + +Lingjie Liu3 + +Christian Theobalt1,4 + +1MPI Informatics & Saarland Informatics Campus 2Google 3University of Pennsylvania + +$^{4}$ Saarbrücken Research Center for Visual Computing, Interaction and Artificial Intelligence + +Project Page: https://jianwang-mpi.github.io/egowholemocap + +![](images/2c7264137600ce24089392af138465890eff1926ff7756b800fdf6649855f9de.jpg) +Figure 1. From an image sequence captured by a single head-mounted fisheye camera, our method can predict accurate and temporally coherent whole-body motion, including human body and hand poses. The SMPL-X parameters are obtained using inverse kinematics. + +![](images/36b3ea0ad948dd017e6c69274be3bab28c3522b0410a21094784063b16913666.jpg) + +![](images/3680f9941ff67bc2aed8921004f6f30027b0736c4f72dd633f1953249cb8c4a6.jpg) + +![](images/d1184db6933fba68a3c957ead65e29b2421106ab5e8f41b12ea82315382b41a3.jpg) + +![](images/e229efc3cee3db2f232971165b7e0bf10a4d372ff2372289390bd48b23a8ebe0.jpg) + +![](images/bda988de8371450c63d11fa9d74da986daaabeb32f388b3626bbe0e66cfca2a7.jpg) + +# Abstract + +In this work, we explore egocentric whole-body motion capture using a single fisheye camera, which simultaneously estimates human body and hand motion. This task presents significant challenges due to three factors: the lack of high-quality datasets, fisheye camera distortion, and human body self-occlusion. To address these challenges, we propose a novel approach that leverages FisheyeViT to extract fisheye image features, which are subsequently converted into pixel-aligned 3D heatmap representations for 3D human body pose prediction. For hand tracking, we incorporate dedicated hand detection and hand pose estimation networks for regressing 3D hand poses. Finally, we develop a diffusion-based whole-body motion prior model to refine the estimated whole-body motion while accounting for joint uncertainties. To train these networks, we collect a large synthetic dataset, EgoWholeBody, comprising 840,000 high-quality egocentric images captured across a diverse range of whole-body motion sequences. Quantitative and qualitative evaluations demonstrate the effectiveness of our method in producing high-quality whole-body motion estimates from a single egocentric camera. + +# 1. Introduction + +Egocentric 3D human motion estimation using head-mounted devices [47, 54] has garnered significant traction + +in recent years, driven by its diverse applications in VR/AR. Immersed in a virtual world, we can traverse virtual environments, interact with virtual objects, and even simulate real-world interactions. To fully capture the intricacies of human motion during such interaction, understanding both body and hand movements is essential. While existing egocentric motion capture methods [30, 47, 50-52, 54] focus solely on body motion, neglecting the hands, this work proposes the task of egocentric whole-body motion capture, i.e. simultaneous estimation of the body motion and hand motion from a single head-mounted fisheye camera (shown in Fig. 1). This task is extremely challenging due to three factors: First, the fisheye image introduces significant distortion, making it difficult for existing networks, which are designed for non-distorted images, to extract features. Second, the egocentric camera perspective frequently leads to the occlusion of body parts, such as the feet and hands, further complicating the task of whole-body motion capture. Lastly, large-scale training data with ground truth annotations for both body and hand poses is absent in existing datasets [4, 29, 47, 51, 54]. + +In this work, we propose a novel egocentric whole-body motion capture method to address the aforementioned challenges. To effectively address fisheye distortion, we propose FisheyeViT for extracting image features, along with a joint regressor employing pixel-aligned 3D heatmap for predicting 3D body poses. Instead of attempting to undis- + +tort the entire fisheye image, which is impractical due to the fisheye lens's large field of view (FOV), we opt to partition the image into smaller patches aligned with a specific FOV range. This approach enables individual patch-level undistortion and seamlessly aligns with the vision transformer architecture that is employed for extracting the complete image feature map. We further propose an egocentric 3D pose regressor utilizing 3D heatmap representations. Unlike the existing approach [52] that projects image features into 3D space through fisheye reprojection functions and regresses 3D heatmaps with V2V networks [33] leading to intricate network learning and high computational complexity—our proposed egocentric pose regressor adopts a simpler approach. It employs deconvolutional layers to obtain pixel-aligned 3D heatmaps. Notably, the voxels in the 3D heatmap directly correspond to pixels in 2D features, subsequently linking to image patches in FisheyeViT. This streamlined approach significantly simplifies network training. Joint locations from the pixel-aligned 3D heatmap are finally transformed with the fisheye camera model to obtain the 3D human body poses. Due to the large size difference between body and hands, we train a hand detection network and a hand pose estimation network to accurately regress 3D hand poses. + +To overcome the challenges posed by self-occlusion and improve the accuracy of pose estimation, we propose a novel method for refining the whole-body motion predictions by incorporating temporal context and a motion prior. Our method learns a whole-body motion prior with the diffusion model [18] from a collection of diverse human motion sequences, capturing intrinsic correlations between hand and body movements. Following this, we extract the joint uncertainties from the pixel-aligned 3D heatmap and utilize them to guide the refinement of the whole-body motion. The joint uncertainties act as indicators of the trustworthiness of the pose regressor's predictions. By conditioning on joints with low uncertainty, our whole-body motion diffusion model selectively refines joints with high uncertainty. This strategy substantially improves the quality of whole-body pose estimations and effectively mitigates the effects of self-occlusion. + +In response to the absence of the egocentric whole-body motion capture datasets, we present EgoWholeBody, a new large-scale high-quality synthetic dataset. This dataset encompasses a wide range of whole-body motions, comprising over 870k frames, which significantly surpasses the size of previous egocentric training datasets. EgoWholeBody could serve as a valuable resource for advancing research in egocentric whole-body motion capture. + +A thorough evaluation across a range of datasets, including SceneEgo [52], GlobalEgoMocap [50] and $\mathrm{Mo}^2\mathrm{Cap}^2$ [54], has demonstrated the remarkable improvements of our method in estimating egocentric whole-body + +motion compared to previous approaches. This substantiates the effectiveness of our approach in addressing the special challenges encountered in egocentric views, including the fisheye distortion and self-occlusion. + +In summary, our key contributions are the following: + +- The first egocentric whole-body motion capture method that predicts accurate and temporarily coherent egocentric body and hand motion; +- FisheyeViT for alleviating fisheye camera distortion and pose regressor using pixel-aligned 3D heatmaps for accurate egocentric body pose estimation from a single image; +- Uncertainty-aware refinement method based on motion diffusion models for correcting initial pose estimations and predicting plausible motions even under occlusion; +- EgoWholeBody, a new high-quality synthetic dataset for egocentric whole-body motion capture. + +# 2. Related Work + +Egocentric 3D Human Body Pose Estimation. Recently, there has been growing interest in estimating egocentric 3D poses from body-worn cameras. Some methods [21, 25, 31, 35, 59, 60] use front-facing cameras and infer the human body motion from the camera view. However, since the user's body is often unobserved by the camera, these methods fail when the human body is not roaming around. Millerdurai et al. [32] leverage event cameras for estimating egocentric body pose. Other methods [4, 5, 7, 23, 39, 65] use head-mounted down-facing stereo cameras to estimate body poses. However, stereo camera setups introduce extra burdens of weight and energy consumption. + +Xu et al. [54] and Tome et al. [47] introduce the single head-mounted down-facing fisheye camera setup for the egocentric 3D human pose estimation task. Zhang et al. [64] regressed fisheye camera parameters and 3D human pose simultaneously. To address the self-occlusion issue, Park et al. [36] leveraged the temporal information with the spatio-temporal self-attention network, and Liu et al. [30] introduced diffusion model to generate 3D human pose conditioned on egocentric image features. Wang et al. [50] and Liu et al. [28] combined the SLAM and egocentric pose estimation methods to estimate human body poses in the world coordinate. Wang et al. [51] and Liu et al. [29] leverage the synchronized egocentric camera and external cameras to collect large-scale egocentric pose estimation datasets with pseudo-ground truth. Considering the human-scene interaction, Wang et al. [52] estimated the scene geometry from the egocentric camera and constrained the 3D human pose with it. + +These methods only focus on estimating human body poses while omitting the hand motion, and they still suffer from fisheye camera distortion since they directly put the highly distorted fisheye images into the neural network. Our proposed method can capture whole-body motion and + +![](images/0db61e7538dc84f37f270d5fe48c34a96602a8642274b91ef635766c67b4affc.jpg) +Figure 2. Overview of our whole-body motion capture pipeline. We first use FisheyeViT to undistort the input image and generate image feature tokens (3.1.1). Next, we use a 1D convolutional network to convert the image features to a pixel-aligned 3D heatmap and use soft-argmax and fisheye camera undistortion function to obtain the 3D body joins positions and uncertainty (3.1.2). We further detect the hand location and regress the 3D hand poses from the input image (3.1.3). Finally, the estimated hand motion and human body motion are combined and the uncertainty-aware diffusion model is applied to refine the estimated whole-body motion (3.2). + +resolve the fisheye camera distortion issue with the FisheyeViT and pixel-aligned 3D heatmap. + +Whole-Body 3D Pose Estimation. Whole-body 3D pose estimation aims to estimate the 3D human body, face, and hands parameters from input images. This task is crucial for many applications, e.g., modeling human activities and human-scene interactions. Some methods [37, 53] fit the 2D body joints estimated from images with optimization algorithms, while these methods suffer from high computation overhead and can fall into local optima. Some other learning-based methods [6, 9, 15, 27, 40, 45, 66] use the neural network to regress the SMPL-X [37] parameters from input images. For example, ExPose [9] introduced body-driven attention to extract face and hand crops and used a refinement module to regress whole-body pose. OSX [27] proposed a one-stage pipeline for whole-body mesh recovery without separate networks for each part. SMPLer-X [6] propose a foundation model for whole-body pose estimation trained with the large model and big data. + +Though much progress has been made on whole-body pose estimation from an external view, the task from an ego-centric view is still unexplored. In this paper, we introduce the first whole-body 3D pose estimation method from a single egocentric image and also incorporate temporal information with diffusion-based motion refinement. + +Diffusion Models for Pose Estimation. Recently, some methods [8, 10, 16, 17, 19, 42] have effectively applied Denoising Diffusion Probabilistic Models (DDPM) [18] to human pose estimation tasks. Building on the success of motion diffusion models in human pose estimation, many methods have extended this approach to egocentric pose estimation, where the human body is only partially visible from RGB cameras or VR sensors. Zhang et al.'s work [63] + +uses a diffusion model to generate realistic human poses considering scene geometry. AGROL [14] generates body motion based on head and hand 6D pose estimates from a VR headset. EgoEgo [25] estimates head poses from a head-mounted front-facing camera and uses them to generate body poses. EgoHMR [30] extracts image features and uses them as a condition for the diffusion denoising process. + +However, the aforementioned pose estimation methods train the conditioned diffusion model with image features or IMU signals. This cannot be generalized since the trained network only accepts one specific condition format and is inclined to learn domain-specific distributions of condition features. ZeDO [22] tackles this issue with a zero-shot diffusion-based optimization approach that doesn't require training with 2D-3D or image-3D pairs. Our method leverages the uncertainty value given by the single-frame pose estimation network and refines the initial motion estimation with the uncertainty of each joint. Moreover, different from previous methods that only focus on human body motion, we train a whole-body motion diffusion model to construct the relationship between hand and body motion. + +# 3. Method + +In this section, we propose a new method for predicting accurate egocentric whole-body poses from egocentric image sequences. An overview of our approach is shown in Fig. 2. + +# 3.1. Single Image Based Egocentric Pose Estimation + +# 3.1.1 FisheyeViT + +In this section, we introduce FisheyeViT, which is specially designed to alleviate the fisheye distortion issue. Instead of undistorting the entire fisheye image, we extract undistorted + +![](images/9012d87f36fd44a5ad44837bd96db815a063baf3c73872f6d559e4d84768d640.jpg) + +![](images/a150273fbd21bd7d66a798fb15085301439938063a88b8ed59b7b973e509a2aa.jpg) + +![](images/785fbab66b2e9d0cfba8efc7a302c7cbc151a05f21cd9c45d1c6f3ea85e1794c.jpg) +Figure 3. The detailed illustration of FisheyeViT (Sec. 3.1.1). + +![](images/3897e648c0b37f699b97766125124aa2e6c52ad507a6c3128e166b975a50c9ed.jpg) + +image patches from the fisheye image and then fit these patches as tokens into the transformer network [13]. To get the undistorted patches, we first warp the fisheye image to a unit semi-sphere, then get the patches with the gnomonic projection (see Fig. 2). The FisheyeViT can be split into five steps, the first four of which are illustrated in Fig. 3. + +Step 1. Given an input image $\mathbf{I}$ with size $H\times W$ , we first evenly sample $N\times N$ patch center points: $\{\mathbf{C}_{ij} = (u_i,v_j) = \left(\frac{H}{N} (i + \frac{1}{2}),\frac{W}{N} (j + \frac{1}{2})\right)|i,j\in 0,\dots,N - 1\}$ . Then, the patch center points $\mathbf{C}_{ij}$ are projected onto a unit sphere with the fisheye reprojection function: $\mathbf{P}_{ij}^{c} = (x_{ij}^{c},y_{ij}^{c},z_{ij}^{c}) = \mathcal{P}^{-1}(u_i,v_j,1)$ . The fisheye camera model is described in Sec. 8 of the supplementary material. Given a point $\mathbf{P}_{ij}^{c}$ on the unit sphere, the tangent plane $\mathbf{T}_{ij}$ that passes through the point is defined by the normal vector $\mathbf{v}_{ij}^{c} = (x_{ij}^{c},y_{ij}^{c},z_{ij}^{c})$ . In the following steps, we implement the gnomonic projection by sampling grid points in the plane and projecting them back onto the fisheye image. + +Step 2. In this step, we determine the orientation of the grid points in the tangent plane, ensuring that the grid points from different tangent planes $\mathbf{T}_{ij}$ have the same orientation when projected back onto the fisheye image. To achieve this, we select a 2D point $\mathbf{U}_{ij} = (u_i + d,v_j)$ in the fisheye image space that is $d$ pixels to the right of the patch center point and project it to the unit sphere using the fisheye reprojection function: $\mathbf{P}_{ij}^{u} = (x_{ij}^{u},y_{ij}^{u},z_{ij}^{u}) = \mathcal{P}^{-1}(u_i + d,v_j,1)$ . We then calculate the intersection point $\mathbf{P}_{ij}^{x}$ between the vector $\mathbf{v}_{ij}^{u} = (x_{ij}^{u},y_{ij}^{u},z_{ij}^{u})$ that is passing the origin and the tangent plane $\mathbf{T}_{ij}$ : + +$$ +\mathbf {P} _ {i j} ^ {x} = \frac {\left\langle \mathbf {P} _ {i j} ^ {c} , \mathbf {v} _ {i j} ^ {c} \right\rangle}{\left\langle \mathbf {v} _ {i j} ^ {u} , \mathbf {v} _ {i j} ^ {c} \right\rangle} \mathbf {v} _ {i j} ^ {u} = \frac {1}{\left\langle \mathbf {v} _ {i j} ^ {u} , \mathbf {v} _ {i j} ^ {c} \right\rangle} \mathbf {v} _ {i j} ^ {u}, \tag {1} +$$ + +where $\langle \cdot ,\cdot \rangle$ denotes the inner product. + +Step 3. Based on the center point $\mathbf{P}_{ij}^{c}$ and intersection point $\mathbf{P}_{ij}^{x}$ on the tangent plane $\mathbf{T}_{ij}$ , we build a coordinate + +system with the $x$ axis: $\mathbf{v}_{ij}^{x} = \mathrm{Norm}(\mathbf{P}_{ij}^{x} - \mathbf{P}_{ij}^{c})$ , the $z$ axis: $\mathbf{v}_{ij}^{z} = \mathrm{Norm}(\mathbf{v}_{ij}^{c})$ and the $y$ axis: $\mathbf{v}_{ij}^{y} = \mathbf{v}_{ij}^{x} \times \mathbf{v}_{ij}^{x}$ where Norm denotes the normalize operation. We grid-sample $M \times M$ points in a $l \times l$ square on the $x-y$ plane: + +$$ +\left\{\mathbf {P} _ {i j} ^ {m n} = \mathbf {P} _ {i j} ^ {c} + \left(l \frac {m}{M} \mathbf {v} _ {i j} ^ {x}, l \frac {n}{M} \mathbf {v} _ {i j} ^ {y}\right) \right\} \tag {2} +$$ + +where $m,n\in -\frac{1}{2} (M - 1),\ldots , - \frac{3}{2}, - \frac{1}{2},\frac{1}{2},\frac{3}{2},\ldots \frac{1}{2} (M - 1)$ + +Step 4. The points $\mathbf{P}_{ij}^{mn}$ are projected back to the fisheye image with the fisheye projection function: $\mathbf{C}_{ij}^{mn} = \mathcal{P}(\mathbf{P}_{ij}^{mn})$ . We then apply bilinear sampling to obtain the colors at points $\mathbf{C}_{ij}^{mn}$ of the input image $\mathbf{I}$ , yielding the undistorted image patch $\mathbf{I}_{ij}^{\mathrm{undis}}$ . Please also see the supplementary video for a visual demonstration of undistorted image patches and their movement on the fisheye image. + +Step 5. The image patches $\{\mathbf{I}_{ij}^{\mathrm{undis}}\}$ are sent to a ViT transformer network [13] to obtain the feature tokens $\{\mathbf{F}_{ij}\}$ . The feature token is further reshaped in $i\times j$ matrix and obtain the image feature $\mathbf{F}$ . In the FisheyeViT, we empirically chose $N = 16$ ; $M = 16$ ; $d = 8$ ; $l = 0.2m$ given the image size $H = W = 256$ . + +Note that $\mathbf{C}_{ij}^{mn}$ is independent of the image $\mathbf{I}$ . This means that, given a fixed fisheye camera model, we can precompute $\mathbf{C}_{ij}^{mn}$ for all combinations of $m,n$ and $i,j$ in advance. This significantly speeds up both the training and evaluation processes. Furthermore, the number and dimensions of image patches $\{\mathbf{I}_{ij}^{\mathrm{undis}}\}$ match exactly with those in the traditional ViT network. This compatibility allows us to finetune existing ViT networks on our egocentric datasets. Our sampling strategy ensures that each image patch $\mathbf{I}_{ij}^{\mathrm{undis}}$ corresponds to the same FOV range in the fisheye camera. In our ablation study in Sec. 5.3, we show that FisheyeViT enhances the performance of the pose estimation network when applied to egocentric fisheye images. + +# 3.1.2 Pose Regressor with Pixel-Aligned 3D Heatmap + +After collecting image features with FisheyeViT, we utilize a 3D heatmap-based network to estimate the body poses. The existing 3D heatmap-based pose regressors [34, 44] are designed for the weak-perspective cameras and predict the 3D heatmap in $xyz$ space. Directly applying these regressors will result in misalignment between 3D heatmap features in $xyz$ space and 2D image features in the fisheye image space. Therefore, we introduce a novel egocentric pose regressor that relies on the pixel-aligned 3D heatmap, tailored to address the needs of fisheye cameras. The idea is to regress the 3D heatmap in $uvd$ space rather than traditional $xyz$ space, where $uv$ corresponds to the fisheye image $uv$ space. Specifically, given a feature map $\mathbf{F} \in \mathbb{R}^{C \times N \times N}$ , where $C$ is the channel number, $N$ is feature map height and width, we firstly use two deconvolutional layers to convert the feature map $\mathbf{F}$ into shape + +$(D_h\times J,H_h,W_h)$ , and further reshape it to pixel-aligned 3D heatmap $\mathbf{H}\in \mathbb{R}^{J\times D_h\times H_h\times W_h}$ , where $J$ is the joint number and $D_{h},H_{h},W_{h}$ is the 3D heatmap depth, height and width. The illustration of pixel-aligned 3D heatmap is shown in Fig. 2. Next, we obtain the max-value positions $\tilde{\mathbf{J}}_b = \{(u_i,v_i,d_i)\mid i\in 0,1,2,\dots,J\}$ from $\mathbf{H}$ by the differentiable soft-argmax operation [44]. Here, we note that $u_{i}$ and $v_{i}$ correspond to the uv-coordinate of the 3D body joint projected in the fisheye image space, and $d_{i}$ denotes the distance of the joint to the fisheye camera. Finally, the 3D body joints $\hat{\mathbf{J}}_b = \{(x_i,y_i,z_i)\mid i\in 0,1,2,\dots,J\}$ are recovered with the fisheye reprojection function: $(x_{i},y_{i},z_{i}) =$ $\mathcal{P}^{-1}(u_i,v_i,d_i)$ . The predicted body pose $\hat{\mathbf{J}}_b$ is finally compared with the ground truth body pose $\mathbf{J}_b$ with the MSE loss. By first regressing 3D body poses in uvd space and then reproducing it, we ensure that the 3D heatmap is pixelaligned with the end-to-end training. + +With the pixel-aligned heatmap, our proposed 3D pose regressor solves problems in all three types of previous egocentric joint regressors. First, $\mathrm{Mo}^2\mathrm{Cap}^2$ [54] employs separate networks to predict 2D joint positions and joint distances. However, this method can yield unrealistic joint estimations because small errors in 2D joints can result in large errors in 3D joints due to the projection effect. Second, $xR$ -egopose [47] and EgoHMR [30] directly regress the 3D joint positions. However, this method is agnostic to the fisheye camera parameters, making it suitable only for a specific camera configuration (e.g., camera parameters, head-mounted position, etc.). Third, SceneEgo [52] projects 2D features into 3D voxel space and uses a V2V network to regress 3D poses. Because of these, the SceneEgo method suffers from low accuracy and large computation overhead. Different from previous methods, our pose regressor with pixel-aligned 3D heatmap is versatile and efficient since it directly estimates 3D joints while also incorporating an explicitly parametrized fisheye camera model. Moreover, it can preserve the uncertainty of the estimated joints, which will be used in our uncertainty-aware motion refinement method (Sec. 3.2.2). Detailed comparison with other pose prediction heads is shown in Table 3. + +# 3.1.3 Egocentric Hand Pose Estimation + +In this section, we first train a network to detect hand pose locations and then train a 3D hand pose estimation network to regress 3D hand poses. Then, we describe how to integrate the estimated hand and body poses. + +Hand Detection. Given an input image $\mathbf{I}$ , we finetune the HRNet [49] network to regress the 2D hand poses of left hand $\mathbf{J}_{lh}^{2d}$ and right hand $\mathbf{J}_{rh}^{2d}$ . From the hand poses, we obtain the center point of left hand $\mathbf{C}_{lh}$ and right hand $\mathbf{C}_{rh}$ along with the bounding box sizes, $d_{lh}$ and $d_{rh}$ . We use our approach described in Sec. 3.1.1 to compute undistorted + +image patches of left $\mathbf{I}_{lh}$ and right hands $\mathbf{I}_{rh}$ . + +Hand Pose Estimation. Given the cropped image $\mathbf{I}_{lh}$ or $\mathbf{I}_{rh}$ , we regress the 3D hand poses $\hat{\mathbf{J}}_{lh}^{loc}$ and $\hat{\mathbf{J}}_{rh}^{loc}$ with the Hand4Whole [34] network, which is fine-tuned on our Ego-FullBody dataset. + +Integration of Body and Hand Positions. It is not straightforward to integrate the hand poses with the body pose in the egocentric camera view primarily due to the fisheye camera's perspective effects. Take the left hand as an example. Following Step 3 in Sec. 3.1.1, we establish a local coordinate system on the tangent plane of the left-hand image with XYZ axes as follows: $x:\mathbf{v}_{lh}^{x};y:\mathbf{v}_{lh}^{y};z:\mathbf{v}_{lh}^{z}$ . We define a rotation matrix, denoted as $\mathbf{R}$ , that represents the transformation between the root coordinate system and the local coordinate system on the tangent plane. The estimated hand pose is first rotated with the rotation matrix $\hat{\mathbf{J}}_{lh} = \mathbf{R}\hat{\mathbf{J}}_{lh}^{loc}$ and then translated to align the wrist location of the human body. This same process is also applied to the right hand to get the right hand pose $\hat{\mathbf{J}}_{rh}$ . The whole-body joints $\hat{\mathbf{J}}$ are obtained by combining $\hat{\mathbf{J}}_b$ , $\hat{\mathbf{J}}_{lh}$ , and $\hat{\mathbf{J}}_{rh}$ . The uncertainty of whole-body joints $\hat{\mathbf{U}}$ is also obtained from the maximal value of the 3D heatmap in pose estimation modules. + +# 3.2. Diffusion-Based Motion Refinement + +We notice that the single-frame estimations in Sec. 3.1 suffer from inaccuracies and temporal instabilities. In this section, we propose a diffusion-based motion refinement method to tackle this problem. We first learn the whole-body motion prior with the motion diffusion model in Sec. 3.2.1 and then introduce an uncertainty-aware zero-shot motion refinement method in Sec. 3.2.2. + +# 3.2.1 Whole-Body Motion Diffusion Model + +We follow DDPM [18] as our diffusion approach to capture the whole-body motion prior $q(\mathbf{x})$ . DDPM learns a distribution of whole-body motion $\mathbf{x}$ through a forward diffusion process and an inverse denoising process. The forward diffusion process is a Markov process of adding Gaussian noise over $t \in \{0, 1, \dots, T - 1\}$ steps: + +$$ +q \left(\mathbf {x} _ {t} \mid \mathbf {x} _ {t - 1}\right) = \mathcal {N} \left(\sqrt {\alpha_ {t}} \mathbf {x} _ {t - 1}, (1 - \alpha_ {t}) I\right) \tag {3} +$$ + +where $\mathbf{x}_t$ denotes the whole-body motion sequence at step $t$ , the variance $(1 - \alpha_{t}) \in (0,1]$ denotes a constant hyperparameter increases with $t$ . + +The inverse process uses a denoising network $D(\cdot)$ to remove the added Gaussian noise at each time step $t$ . Here we use the transformer-based framework in EDGE [48] as the motion-denoising network $D(\cdot)$ . We follow Ramesh et al.'s work [38] to make the network predict the original signal itself, i.e. $\hat{\mathbf{x}}_0 = D(\mathbf{x}_t,t)$ and train it with the simple objective [18]: + +$$ +\mathcal {L} _ {\text {s i m p l e}} = E _ {\mathbf {x} _ {0} \sim q (\mathbf {x} _ {0}), t \sim [ 1, T ]} \left[ \| \mathbf {x} _ {0} - D (\mathbf {x} _ {t}, t) \| _ {2} ^ {2} \right] \tag {4} +$$ + +# 3.2.2 Uncertainty-Aware Motion Refinement + +Given the learned whole-body motion prior, we leverage the uncertainty value for each pose to guide the diffusion denoising process with the classifier-guided diffusion sampling [12]. Given an initial sequence of whole-body pose estimation $\mathbf{x}_e = \{\hat{\mathbf{J}}_i\}$ and the uncertainty value for each pose $\mathbf{u} = \{\hat{\mathbf{U}}_i\}$ , where $i$ denotes the $i$ th pose in the sequence, we keep the joints with low uncertainty but use the diffusion model to generate joints with high uncertainty conditioned on the low-uncertainty joints. Specifically, in the $t$ th sampling step of the diffusion process, the denoising network predicts $\hat{\mathbf{x}}_0 = D(\mathbf{x}_t,t)$ , which is noised back to $\mathbf{x}_{t - 1}$ by sampling from the Gaussian distribution: + +$$ +\mathbf {x} _ {t - 1} \sim \mathcal {N} \left(\hat {\mathbf {x}} _ {0} + \mathbf {w} \left(\mathbf {x} _ {e} - \hat {\mathbf {x}} _ {0}\right), \Sigma_ {t}\right) \tag {5} +$$ + +where $\Sigma_{t}$ is a scheduled Gaussian distribution in DDPM [18] and $\mathbf{w}$ controls the weight of a specific joint between the predicted motion $\hat{\mathbf{x}}_0$ and the estimated motion $\mathbf{x}_e$ . Generally, we expect $\mathbf{w} \rightarrow \overrightarrow{0}$ when $t \rightarrow 0$ such that the temporal stability is guaranteed through the generation of the denoising process, and $\mathbf{w} \rightarrow \overrightarrow{1}$ when $t \rightarrow T$ such that the denoising process is initialized by the estimated motion $\mathbf{x}_e$ . We also expect that $w_{ij} = \mathbf{w}[i][j]$ , which is the weight of $j$ th joints in the $i$ th pose, is smaller when the uncertainty value $u_{ij} = \mathbf{u}[i][j]$ of the $j$ th joints in the $i$ th pose is large. Based on this requirement, we design $\mathbf{w}$ as: + +$$ +\mathbf {w} = 1 / \left(1 + e ^ {- k (t - T \mathbf {u})}\right) \tag {6} +$$ + +where $T$ is the overall diffusion steps, $k$ is a hyperparameter which is empirically set to 0.1. From the experimental results in Sec. 5, we demonstrate the effectiveness of uncertain-aware motion refinement and our uncertainty-guided diffusion sampling strategy. + +# 4. EgoWholeBody Dataset + +In this section, we introduce EgoWholeBody, a large-scale high-quality synthetic dataset built for the task of egocentric whole-body motion capture. The EgoWholeBody dataset is organized into two sections. The first part, containing over 700k frames, is rendered with 14 different rigged Renderpeople [3] models driven by 2367 Mixamo [2] motion sequences. The second part focuses on hand motions and contains 170k frames with the SMPL-X model. This data is constructed from 24 different shapes and textures, driven by 262 motion sequences selected from the GRAB [46] and TCDHandMocap dataset [20]. We also created synthetic test sequences, which include 133k images rendered with 3 Renderpeople models and Mixamo motions. + +During the rendering process, we first attach a virtual fisheye camera to the forehead of human body models and render the images, semantic labels, and depth map with + +Blender [1]. Our dataset is larger and more diverse than previous egocentric training datasets—see Sec. 10 in the supplementary material for a detailed comparison. + +# 5. Experiments + +# 5.1. Datasets and Evaluation Metrics + +Training Datasets. To train our body pose estimation module (Sec. 3.1.1 and Sec. 3.1.2), we use our EgoWholeBody dataset and the EgoPW dataset [50]. Additionally, the EgoWholeBody dataset is used to train the hand pose estimation module in Sec. 3.1.3. For training the whole-body diffusion model (Sec. 3.2), we utilize a combined motion capture dataset that includes EgoBody [62], Mixamo [2], TCDHandMocap dataset [20] and GRAB dataset [46]. + +Evaluation Datasets. In our experiment, we evaluate our methods on four datasets: the GlobalEgoMocap test datasets [50], the $\mathrm{Mo}^2\mathrm{Cap}^2$ test dataset [54], the SceneEgo test dataset [52] and out EgoWholeBody test dataset. The details of the datasets are shown in Sec. 12 of supplementary materials. Note that evaluating whole-body poses requires accurate annotations for human hands, which is absent in real-world datasets. To resolve the issue, we request the multi-view videos of the SceneEgo test dataset [52] from the authors and use a multi-view motion capture system to obtain the hand motion. The hand pose annotations will be made publicly available. + +Evaluation Metrics. To evaluate the precision of human body poses on the SceneEgo test dataset [52], we use MPJPE and PA-MPJPE. For the GlobalEgoMocap test dataset [50] and $\mathrm{Mo}^2\mathrm{Cap}^2$ test dataset [54], where egocentric camera poses are unavailable, we evaluate PA-MPJPE and BA-MPJPE. For hand pose accuracy, we align the predicted and ground truth hand poses at the root position, followed by computing MPJPE and PA-MPJPE. Detailed explanations of these metrics are in Sec. 11 of the supplementary materials. All reported metrics are in millimeters. + +# 5.2. Comparisons on Whole-Body Pose Estimation + +For a fair comparison with existing methods focusing solely on body or hand pose, we split our evaluation into two parts, reporting results of body poses in Table 1 and hand pose in Table 2. We first compare the accuracy of the human body poses with state-of-the-art methods, including EgoPW [51] and SceneEgo [52], on EgoWholeBody and SceneEgo [52] test datasets. The comparison with more previous methods [47, 50, 54] and on more evaluation datasets [50, 54] are shown in Sec. 7 of the supplementary materials. Since our motion refinement method incorporates random Gaussian noise, we generate five samples and calculate the average MPJPE values. The standard deviation is low ( $< 0.01\mathrm{mm}$ ) and is discussed in Sec. 13 of supplementary materials. Results are presented in Table 1, where our single-frame re + +![](images/2b36abfc7293d272285e3cd4a5923fe4d3975b24615caa51d348f4182cec6dee.jpg) +Input + +![](images/7296f8f261090bbba4cc565a89c0a1263b78b4e35d582f0d76b7be038e6db2f8.jpg) +EgoPW + +![](images/fd3208ba132e94b485215577d61f0e4ae8c98e8d844925559f66b5d3f0706c32.jpg) +SceneEgo + +![](images/0416137e264b082666a001f0d877bdecb1233463c15eece3a3b594c8ffb40dc2.jpg) +Ours-Single + +![](images/d116b8c08e886af64ea3404515370357c512628eee4cf8c1bec4e8f6b123554b.jpg) +Ours-Refined + +![](images/eec0ee177d1a6d86bdbe3f5c7fd62a8b1c16990dd0ea4d339fd8768b2a0881fc.jpg) +Input + +![](images/78bc9457063ba21f824a921e8d7d897ae814735a0288d48f7fd304888c07b4d0.jpg) +EgoPW + +![](images/2db5bb07a47b9f653267dd865fcf90623f3e95809148e7cc681c9f05a7829dea.jpg) +SceneEgo + +![](images/1f6fc40d8fa32c0cf3eb18c47bf63e527bb44f52c316f76a32a6410d14e1f311.jpg) +Ours-Single + +![](images/0aef160c92c409446aaee3cae33e11765fb8998fec395ea53b2961ef82f61071.jpg) +Ours-Refined + +![](images/a705b4d478e7ae29a419d8aef00da1809a863112089c2bd25dd790aeac296a91.jpg) +Figure 4. Qualitative comparison on human body pose estimations between our methods and the state-of-the-art egocentric pose estimation methods on in-the-studio (left column) and in-the-wild scenes (right column). The red skeleton is the ground truth while the green skeleton is the predicted pose. Our methods predict more accurate body poses compared with EgoPW [51] and SceneEgo [52]. +Input +Figure 5. Qualitative comparison on human hand pose estimations between our methods and the state-of-the-art third-view pose estimation methods. Our single-view and refined hand poses are more accurate than the poses from Hand4Whole [34] method. The red skeleton is the ground truth while the green skeleton is the predicted pose. + +![](images/2511ddddfbb45f663922c3d6275ba93aaedca7fedc5dd82520a337637611a853.jpg) +Hand Image + +![](images/a3e0d8f6c237c8e18267565ef03ef52150f8a31fa73290579d211711de55b529.jpg) +Hand4Whole + +![](images/d1567b2df1aaabf7c88b08f6808222173c19e81a7842c75318e92e5a5153e0da.jpg) +Ours-Single + +![](images/9d14d446c812f3ba79dff90426f0c487f1622cead460c6f80a98f5546f25a5a9.jpg) +Ours-Refined + +![](images/868694d3382e8384e9bc0acb6559028c3eeebd996d190f073c96360431b5b60f.jpg) +Rrnnnne +Hand Image +Hand4Whole + +![](images/0aab8a411ce9f29fe58d8547125f2a1b5b383cff52ba3a7078c372552d76111f.jpg) + +![](images/973eb23b413d2231b829f02d5a8402bd11582fa3bae0b981c051d6b8702eff50.jpg) +Ours-Single + +![](images/ebd9da3b584bfa078c7c9111266a0314999a58041405dabbcfef33dd46c35036.jpg) +Ours-Refined + +sults are labeled as "Ours-Single" and our refinement results are labeled as "Ours-Refined". Our single-frame body pose estimation method outperforms all previous methods by a large margin. Our diffusion-based motion refinement method can further improve the accuracy of body poses estimated by the single-frame methods. + +Note that previous methods [47, 50-52, 54] use training datasets different from each other. For a fair comparison, we re-train previous methods with our training datasets in Sec. 5.1 and show the results with “*” in Table 1. This retraining led to significant improvements across all previous methods, demonstrating our dataset's broad applicability. However, these methods still underperformed compared to ours, highlighting our approach's superiority. + +To evaluate the accuracy of our hand pose estimation method, we first crop the hand images with the hand detection method in Sec. 3.1.3. Then we show the results of our single-frame hand pose estimation (labeled as "Ours-Single") and whole-body motion refinement methods (labeled as "Ours-Refined") in Table 2. Our single-frame hand pose estimation method outperforms the state-of-the-art method Hands4Whole [34], demonstrating the effectiveness of training the network on our EgoWholeBody dataset. Our whole-body motion refinement method can also enhance the accuracy of hand motion. + +For a qualitative comparison, we compare the body and hand poses of our method with existing methods on the + +
MethodMPJPEPA-MPJPE
SceneEgo test dataset [52]
EgoPW [51]189.6105.3
SceneEgo [52]118.592.75
EgoPW* [51]90.9664.33
SceneEgo* [52]89.0670.10
Ours-Single64.1950.06
Ours-Refined57.5946.55
EgoWholeBody test dataset
EgoPW* [51]84.2163.02
SceneEgo* [52]87.5769.46
Ours-Single66.2843.14
Ours-Refined60.3240.35
+ +Table 1. Egocentric human body pose accuracy of our method on SceneEgo test datasets and EgoWholeBody test dataset. Our method outperforms all previous state-of-the-art methods. * denotes the method trained with the datasets in Sec. 5.1. + +SceneEgo dataset and the in-the-wild EgoPW [51] evaluation sequences. The results are shown in Fig. 4 and Fig. 5, showing that our method can predict high-quality whole-body poses from an egocentric camera. Please refer to our supplementary video for more qualitative evaluation results. + +# 5.3. Ablation Study + +EgoWholeBody Dataset. Compared to existing egocentric datasets, our EgoWholeBody dataset contains diverse body + +
MethodMPJPEPA-MPJPE
SceneEgo test dataset [52]
Hand4Whole [34]49.6613.85
Ours-Single23.639.59
Ours-Refined19.379.05
EgoWholeBody test dataset
Hand4Whole [34]52.8535.04
Ours-Single33.1019.68
Ours-Refined28.2914.51
+ +Table 2. Egocentric hand pose accuracy of our method. Our method outperforms the Hand4Whole [34] on both datasets. + +
MethodMPJPEPA-MPJPE
Body Pose Results
w/o EgoWholeBody75.1058.62
w/o FisheyeViT67.3653.44
w/ Mo2Cap2 [54] head87.4765.10
w/ xR-egopose [47] head116.595.78
w/ SceneEgo [52] head77.7362.69
Ours-Single64.1950.06
w/ GlobalEgoMocap†69.8356.73
w/o uncert. guidance†62.1648.40
Only body diffusion58.9547.03
Ours-Refined†57.5946.55
Hand Pose Results
Only hand diffusion21.699.24
Ours-Refined19.379.05
+ +Table 3. Ablation Study on SceneEgo test dataset [52]. ${}^{ \dagger }$ denotes the temporal-based method. + +and hand motions, larger quantity of images, and higher image quality. We show this by training our body pose estimation network without our dataset, using the $\mathrm{Mo}^2\mathrm{Cap}^2$ [54] and EgoPW [51] training dataset. The results, labeled as "w/o EgoWholeBody" in Table 3, show that performance without the EgoWholeBody dataset is inferior to our proposed method. This highlights that training with our EgoW-holeBody dataset enhances the performance of the pose estimation method. We also compare this result with existing methods on the SceneEgo test set (Table 1). Trained without EgoWholeBody, our approach still outperforms previous methods, showing the effectiveness of our method. + +FisheyeViT and Pose Regressor with Pixel-Aligned 3D Heatmap. To assess the individual contributions of FisheyeViT and the pixel-aligned 3D heatmap in our single-frame pose estimation pipeline, we perform experiments to measure their impact on the overall performance. First, we substitute the FisheyeViT module in our single-frame pose estimation method to ViT [13]. The result is shown in "w/o FisheyeViT" in Table 3 and it is worse than our full method. This demonstrates the effectiveness of FisheyeViT in addressing fisheye distortion and feature extraction. + +Next, we analyze the performance of the single-frame pose estimation network when substituting our pose regress- + +sor based on pixel-aligned 3D heatmap with the pose estimation heads of previous works [47, 52, 54]. The results of the three experiments, labeled as "w/ Mo²Cap² head", "w/ xR-egopose head" and "w/ SceneEgo head", show a performance drop compared to our full method. This emphasizes the crucial role of the pixel-aligned 3D heatmap in accurately estimating egocentric 3D body joint positions. + +Diffusion-based Motion Refinement. We assess the effectiveness of our diffusion-based motion refinement with the following experiments: First, we compare the performance of our diffusion-based motion refinement with GlobalEgoMocap [50] by applying the GlobalEgoMocap optimizer on the single-frame body pose estimation results. The result, labeled as "w GlobalEgoMocap" in Table 3, indicates that our refinement method outperforms GlobalEgoMocap. + +Second, we remove the uncertainty-aware guidance in the motion refinement. Instead, we use fixed Gaussian denoising steps to refine the motion. The result "w/o uncert. guidance" in Table 3, shows that our uncertainty-aware refinement method performs better. Our approach relies on the uncertainty values for each joint, using low-uncertainty joints to guide the generation of high-uncertainty joints. This helps reduce errors in joint predictions caused by egocentric self-occlusion, leading to improved results. + +Third, we replace our whole-body motion diffusion model with the separate human body and left/right-hand diffusion models and show the accuracy of refined body and hand motion in "Only body diffusion" and "Only hand diffusion" in Table 3. From the results, we observe improvements in the accuracy of motion refined by our whole-body diffusion method, proving that learning the whole-body motion prior can help both the refinement of the body and hand motion by learning the correlation between them. + +# 6. Conclusion + +In this work, we have introduced an innovative approach to capture egocentric whole-body human motion. Our method comprises a single-frame-based whole-body pose estimation process, which includes FisheyeViT and pixel-aligned 3D heatmap representations. To enhance the initial whole-body pose estimates, we have integrated an uncertainty-aware diffusion-based motion refinement technique. Our experimental results demonstrate that both our single-frame method and the temporal-based method surpass all existing state-of-the-art techniques in terms of both quality and accuracy. Looking ahead, we see the potential for extending the applications of FisheyeViT to other vision tasks involving fisheye cameras. Future work could also involve incorporating facial expressions in whole-body motion capture. + +Acknowledgments This project was supported by the Saarbrücken Research Center for Visual Computing, Interaction and AI. Christian Theobalt was supported by ERC Consolidator Grant 4DReply (770784). + +# References + +[1] Blender. https://www.blender.org. 6 +[2] Mixamo. https://www MIXamo.com. 6 +[3] Renderpeople. https://renderpeople.com. 6, 3 +[4] Hiroyasu Akada, Jian Wang, Soshi Shimada, Masaki Takahashi, Christian Theobalt, and Vladislav Golyanik. Unrealego: A new dataset for robust egocentric 3d human motion capture. In European Conference on Computer Vision, pages 1-17. Springer, 2022. 1, 2 +[5] Hiroyasu Akada, Jian Wang, Vladislav Golyanik, and Christian Theobalt. 3d human pose perception from egocentric stereo videos. In 37th IEEE/CVF Conference on Computer Vision and Pattern Recognition. IEEE, 2024. 2 +[6] Zhongang Cai, Wanqi Yin, Ailing Zeng, Chen Wei, Qingping Sun, Yanjun Wang, Hui En Pang, Haiyi Mei, Mingyuan Zhang, Lei Zhang, et al. Smpler-x: Scaling up expressive human pose and shape estimation. arXiv preprint arXiv:2309.17448, 2023. 3 +[7] Young-Woon Cha, True Price, Zhen Wei, Xinran Lu, Nicholas Rewkowski, Rohan Chabra, Zihe Qin, Hyounghun Kim, Zhaoqi Su, Yebin Liu, et al. Towards fully mobile 3d face, body, and environment capture using only head-worn cameras. IEEE transactions on visualization and computer graphics, 24(11):2993-3004, 2018. 2 +[8] Jeongjun Choi, Dongseok Shim, and H Jin Kim. Diffupose: Monocular 3d human pose estimation via denoising diffusion probabilistic model. arXiv preprint arXiv:2212.02796, 2022. 3 +[9] Vasileios Choutas, Georgios Pavlakos, Timo Bolkart, Dimitrios Tzionas, and Michael J Black. Monocular expressive body regression through body-driven attention. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part X 16, pages 20-40. Springer, 2020. 3 +[10] Hai Ci, Mingdong Wu, Wentao Zhu, Xiaoxuan Ma, Hao Dong, Fangwei Zhong, and Yizhou Wang. Gfpose: Learning 3d human pose prior with gradient fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4800-4810, 2023. 3 +[11] Benjamin Coors, Alexandru Paul Condurache, and Andreas Geiger. Spherenet: Learning spherical representations for detection and classification in omnidirectional images. In Proceedings of the European conference on computer vision (ECCV), pages 518-533, 2018. 5 +[12] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in neural information processing systems, 34:8780-8794, 2021. 6 +[13] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020.4, 8 +[14] Yuming Du, Robin Kips, Albert Pumarola, Sebastian Starke, Ali Thabet, and Artsiom Sanakoyeu. Avatars grow legs: Generating smooth human motion from sparse tracking inputs with diffusion model. In Proceedings of the IEEE/CVF + +Conference on Computer Vision and Pattern Recognition, pages 481-490, 2023. 3 +[15] Yao Feng, Vasileios Choutas, Timo Bolkart, Dimitrios Tzionas, and Michael J Black. Collaborative regression of expressive bodies using moderation. In 2021 International Conference on 3D Vision (3DV), pages 792-804. IEEE, 2021. 3 +[16] Lin Geng Foo, Jia Gong, Hossein Rahmani, and Jun Liu. Distribution-aligned diffusion for human mesh recovery. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9221-9232, 2023. 3 +[17] Jia Gong, Lin Geng Foo, Zhipeng Fan, Qiuhong Ke, Hossein Rahmani, and Jun Liu. Diffpose: Toward more reliable 3d pose estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13041-13051, 2023. 3 +[18] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 2, 3, 5, 6 +[19] Karl Holmquist and Bastian Wandt. Diffpose: Multi-hypothesis human pose estimation using diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15977-15987, 2023. 3 +[20] Ludovic Hoyet, Kenneth Ryall, Rachel McDonnell, and Carol O'Sullivan. Sleight of hand: perception of finger motion from reduced marker sets. In Proceedings of the ACM SIGGRAPH symposium on interactive 3D graphics and games, pages 79-86, 2012. 6 +[21] Hao Jiang and Kristen Grauman. Seeing invisible poses: Estimating 3d body pose from egocentric video. In 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 3501-3509. IEEE, 2017. 2 +[22] Zhongyu Jiang, Zhuoran Zhou, Lei Li, Wenhao Chai, ChengYen Yang, and Jenq-Neng Hwang. Back to optimization: Diffusion-based zero-shot 3d human pose estimation. arXiv preprint arXiv:2307.03833, 2023. 3 +[23] Taeho Kang, Kyungjin Lee, Jinrui Zhang, and Youngki Lee. Ego3dpose: Capturing 3d cues from binocular egocentric views. arXiv preprint arXiv:2309.11962, 2023. 2 +[24] David G Kendall. A survey of the statistical theory of shape. Statistical Science, 4(2):87-99, 1989. 3 +[25] Jiaman Li, Karen Liu, and Jiajun Wu. Ego-body pose estimation via ego-head pose estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17142-17151, 2023. 2, 3 +[26] Yuyan Li, Yuliang Guo, Zhixin Yan, Xinyu Huang, Ye Duan, and Liu Ren. Omnifusion: 360 monocular depth estimation via geometry-aware fusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2801-2810, 2022. 5 +[27] Jing Lin, Ailing Zeng, Haoqian Wang, Lei Zhang, and Yu Li. One-stage 3d whole-body mesh recovery with component aware transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21159-21168, 2023. 3 +[28] Yuxuan Liu, Jianxin Yang, Xiao Gu, Yao Guo, and Guang-Zhong Yang. Ego+ x: An egocentric vision system for global + +3d human pose estimation and social interaction characterization. In 2022 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 5271-5277. IEEE, 2022. 2 +[29] Yuxuan Liu, Jianxin Yang, Xiao Gu, Yijun Chen, Yao Guo, and Guang-Zhong Yang. Egofish3d: Egocentric 3d pose estimation from a fisheye camera via self-supervised learning. IEEE Transactions on Multimedia, 2023. 1, 2, 4 +[30] Yuxuan Liu, Jianxin Yang, Xiao Gu, Yao Guo, and GuangZhong Yang. Egohmr: Egocentric human mesh recovery via hierarchical latent diffusion model. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 9807-9813. IEEE, 2023. 1, 2, 3, 5 +[31] Zhengyi Luo, Ryo Hachiuma, Ye Yuan, and Kris Kitani. Dynamics-regulated kinematic policy for egocentric pose estimation. Advances in Neural Information Processing Systems, 34:25019-25032, 2021. 2 +[32] Christen Millerdurai, Hiroyasu Akada, Jian Wang, Diogo Luvizon, Christian Theobalt, and Vladislav Golyanik. Evento3d: 3d human motion capture from egocentric event streams. In 37th IEEE/CVF Conference on Computer Vision and Pattern Recognition. IEEE, 2024. 2 +[33] Gyeongsik Moon, Ju Yong Chang, and Kyoung Mu Lee. V2v-posenet: Voxel-to-voxel prediction network for accurate 3d hand and human pose estimation from a single depth map. In Proceedings of the IEEE conference on computer vision and pattern Recognition, pages 5079–5088, 2018. 2 +[34] Gyeongsik Moon, Hongsuk Choi, and Kyoung Mu Lee. Accurate 3d hand pose estimation for whole-body 3d human mesh estimation. In Computer Vision and Pattern Recognition Workshop (CVPRW), 2022. 4, 5, 7, 8, 2 +[35] Evonne Ng, Donglai Xiang, Hanbyul Joo, and Kristen Grauman. You2me: Inferring body pose in egocentric video via first and second person interactions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9890-9900, 2020. 2 +[36] Jinman Park, Kimathi Kaai, Saad Hossain, Norikatsu Sumi, Sirisha Rambhatla, and Paul Fieguth. Domain-guided spatiotemporal self-attention for egocentric 3d pose estimation. In Proceedings of the 29th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, pages 1837–1849, 2023. 2, 1 +[37] Georgios Pavlakos, Vasileios Choutas, Nima Ghorbani, Timo Bolkart, Ahmed AA Osman, Dimitrios Tzionas, and Michael J Black. Expressive body capture: 3d hands, face, and body from a single image. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10975-10985, 2019. 3 +[38] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 1 (2):3, 2022. 5 +[39] Helge Rhodin, Christian Richardt, Dan Casas, Eldar Insafutdinov, Mohammad Shafiei, Hans-Peter Seidel, Bernt Schiele, and Christian Theobalt. Egocap: egocentric marker-less motion capture with two fisheye cameras. ACM Transactions on Graphics (TOG), 35(6):1-11, 2016. 2 + +[40] Yu Rong, Takaaki Shiratori, and Hanbyul Joo. Frankmocap: A monocular 3d whole-body pose estimation system via regression and integration. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1749-1759, 2021. 3 +[41] Davide Scaramuzza, Agostino Martinelli, and Roland Siegwart. A toolbox for easily calibrating omnidirectional cameras. In 2006 IEEE/RSJ International Conference on Intelligent Robots and Systems, pages 5695-5701. IEEE, 2006. 1 +[42] Wenkang Shan, Zhenhua Liu, Xinfeng Zhang, Zhao Wang, Kai Han, Shanshe Wang, Siwei Ma, and Wen Gao. Diffusion-based 3d human pose estimation with multi-hypothesis aggregation. arXiv preprint arXiv:2303.11579, 2023.3 +[43] Soshi Shimada, Vladislav Golyanik, Weipeng Xu, and Christian Theobalt. Physcap: Physically plausible monocular 3d motion capture in real time. ACM Transactions on Graphics (ToG), 39(6):1-16, 2020. 5 +[44] Xiao Sun, Bin Xiao, Fangyin Wei, Shuang Liang, and Yichen Wei. Integral human pose regression. In Proceedings of the European conference on computer vision (ECCV), pages 529-545, 2018. 4, 5 +[45] Yu Sun, Tianyu Huang, Qian Bao, Wu Liu, Wenpeng Gao, and Yili Fu. Learning monocular mesh recovery of multiple body parts via synthesis. In ICASSP 2022-2022 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 2669-2673. IEEE, 2022. 3 +[46] Omid Taheri, Nima Ghorbani, Michael J Black, and Dimitrios Tzionas. Grab: A dataset of whole-body human grasping of objects. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part IV 16, pages 581-600. Springer, 2020. 6 +[47] Denis Tomè, Patrick Peluse, Lourdes Agapito, and Hernán Badino. xr-egopose: Egocentric 3d human pose from an HMD camera. In ICCV, pages 7727-7737, 2019. 1, 2, 5, 6, 7, 8, 4 +[48] Jonathan Tseng, Rodrigo Castellon, and Karen Liu. Edge: Editable dance generation from music. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 448-458, 2023. 5, 2 +[49] Jingdong Wang, Ke Sun, Tianheng Cheng, Borui Jiang, Chaorui Deng, Yang Zhao, Dong Liu, Yadong Mu, Mingkui Tan, Xinggang Wang, et al. Deep high-resolution representation learning for visual recognition. IEEE transactions on pattern analysis and machine intelligence, 43(10):3349-3364, 2020. 5 +[50] Jian Wang, Lingjie Liu, Weipeng Xu, Kripasindhu Sarkar, and Christian Theobalt. Estimating egocentric 3d human pose in global space. ICCV, 2021. 1, 2, 6, 7, 8, 3, 4, 5 +[51] Jian Wang, Lingjie Liu, Weipeng Xu, Kripasindhu Sarkar, Diogo Luvizon, and Christian Theobalt. Estimating egocentric 3d human pose in the wild with external weak supervision. In CVPR, pages 13157-13166, 2022. 1, 2, 6, 7, 8, 4 +[52] Jian Wang, Diogo Luvizon, Weipeng Xu, Lingjie Liu, Kripasindhu Sarkar, and Christian Theobalt. Scene-aware egocentric 3d human pose estimation. In Proceedings of the + +IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13031-13040, 2023. 1, 2, 5, 6, 7, 8, 3, 4 +[53] Donglai Xiang, Hanbyul Joo, and Yaser Sheikh. Monocular total capture: Posing face, body, and hands in the wild. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10965-10974, 2019. 3 +[54] Weipeng Xu, Avishek Chatterjee, Michael Zollhöfer, Helge Rhodin, Pascal Fua, Hans-Peter Seidel, and Christian Theobalt. $\mathrm{Mo}^2\mathrm{cap}^2$ : Real-time mobile 3d motion capture with a cap-mounted fisheye camera. IEEE Trans. Vis. Comput. Graph., 25(5):2093-2101, 2019. 1, 2, 5, 6, 7, 8, 3, 4 +[55] Yufei Xu, Jing Zhang, Qiming Zhang, and Dacheng Tao. Vitpose: Simple vision transformer baselines for human pose estimation. Advances in Neural Information Processing Systems, 35:38571-38584, 2022. 2 +[56] Dianyi Yang, Jiadong Tang, Yu Gao, Yi Yang, and Mengyin Fu. Sector patch embedding: An embedding module conforming to the distortion pattern of fisheye image. arXiv preprint arXiv:2303.14645, 2023. 5 +[57] Shangrong Yang, Chunyu Lin, Kang Liao, and Yao Zhao. Dual diffusion architecture for fisheye image rectification: Synthetic-to-real generalization. arXiv preprint arXiv:2301.11785, 2023. +[58] Fanghua Yu, Xintao Wang, Mingdeng Cao, Gen Li, Ying Shan, and Chao Dong. Osrt: Omnidirectional image superresolution with distortion-aware transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13283-13292, 2023. 5 +[59] Ye Yuan and Kris Kitani. 3d ego-pose estimation via imitation learning. In Proceedings of the European Conference on Computer Vision (ECCV), pages 735–750, 2018. 2 +[60] Ye Yuan and Kris Kitani. Ego-pose estimation and forecasting as real-time pd control. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10082-10092, 2019. 2 +[61] Ye Yuan, Jiaming Song, Umar Iqbal, Arash Vahdat, and Jan Kautz. Physdiff: Physics-guided human motion diffusion model. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 16010-16021, 2023. 5 +[62] Siwei Zhang, Qianli Ma, Yan Zhang, Zhiyin Qian, Taein Kwon, Marc Pollefeys, Federica Bogo, and Siyu Tang. Ego-body: Human body shape and motion of interacting people from head-mounted devices. In European Conference on Computer Vision, pages 180–200. Springer, 2022. 6 +[63] Siwei Zhang, Qianli Ma, Yan Zhang, Sadegh Aliakbarian, Darren Cosker, and Siyu Tang. Probabilistic human mesh recovery in 3d scenes from egocentric views. arXiv preprint arXiv:2304.06024, 2023. 3 +[64] Yahui Zhang, Shaodi You, and Theo Gevers. Automatic calibration of the fisheye camera for egocentric 3d human pose estimation from a single image. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1772-1781, 2021. 2 +[65] Dongxu Zhao, Zhen Wei, Jisan Mahmud, and Jan-Michael Frahm. Egoglass: Egocentric-view human pose estimation from an eyeglass frame. In 2021 International Conference on 3D Vision (3DV), pages 32-41. IEEE, 2021. 2 + +[66] Yuxiao Zhou, Marc Habermann, Ikhsanul Habibie, Ayush Tewari, Christian Theobalt, and Feng Xu. Monocular real-time full body capture with inter-part correlations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4811-4822, 2021. 3 +[67] Christian Zimmermann, Duygu Ceylan, Jimei Yang, Bryan Russell, Max Argus, and Thomas Brox. Freihand: A dataset for markerless capture of hand pose and shape from single rgb images. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 813-822, 2019. 2 \ No newline at end of file diff --git a/2024/Egocentric Whole-Body Motion Capture with FisheyeViT and Diffusion-Based Motion Refinement/images.zip b/2024/Egocentric Whole-Body Motion Capture with FisheyeViT and Diffusion-Based Motion Refinement/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..6fcc4817174d993236d5b5f3d2493c850cfd62d5 --- /dev/null +++ b/2024/Egocentric Whole-Body Motion Capture with FisheyeViT and Diffusion-Based Motion Refinement/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:62bc6996fd5469a4cecd7c964f2616f0ceb775199f8fe64f688c0d82aebffc74 +size 471618 diff --git a/2024/Egocentric Whole-Body Motion Capture with FisheyeViT and Diffusion-Based Motion Refinement/layout.json b/2024/Egocentric Whole-Body Motion Capture with FisheyeViT and Diffusion-Based Motion Refinement/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..9f4397707f7580dc23cfef18d8f9bc9fdcdadfea --- /dev/null +++ b/2024/Egocentric Whole-Body Motion Capture with FisheyeViT and Diffusion-Based Motion Refinement/layout.json @@ -0,0 +1,11776 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 55, + 103, + 539, + 138 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 103, + 539, + 138 + ], + "spans": [ + { + "bbox": [ + 55, + 103, + 539, + 138 + ], + "type": "text", + "content": "Egocentric Whole-Body Motion Capture with FisheyeViT and Diffusion-Based Motion Refinement" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 143, + 160, + 206, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 160, + 206, + 175 + ], + "spans": [ + { + "bbox": [ + 143, + 160, + 206, + 175 + ], + "type": "text", + "content": "Jian Wang" + }, + { + "bbox": [ + 143, + 160, + 206, + 175 + ], + "type": "inline_equation", + "content": "^{1,4}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 99, + 176, + 198, + 189 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 99, + 176, + 198, + 189 + ], + "spans": [ + { + "bbox": [ + 99, + 176, + 198, + 189 + ], + "type": "text", + "content": "Kripasindhu Sarkar2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 225, + 161, + 272, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 225, + 161, + 272, + 175 + ], + "spans": [ + { + "bbox": [ + 225, + 161, + 272, + 175 + ], + "type": "text", + "content": "Zhe Cao2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 214, + 175, + 289, + 189 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 214, + 175, + 289, + 189 + ], + "spans": [ + { + "bbox": [ + 214, + 175, + 289, + 189 + ], + "type": "text", + "content": "Danhang Tang" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 288, + 161, + 374, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 288, + 161, + 374, + 175 + ], + "spans": [ + { + "bbox": [ + 288, + 161, + 374, + 175 + ], + "type": "text", + "content": "Diogo Luvizon1,4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 307, + 175, + 377, + 188 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 175, + 377, + 188 + ], + "spans": [ + { + "bbox": [ + 307, + 175, + 377, + 188 + ], + "type": "text", + "content": "Thabo Beeler2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 389, + 161, + 450, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 389, + 161, + 450, + 175 + ], + "spans": [ + { + "bbox": [ + 389, + 161, + 450, + 175 + ], + "type": "text", + "content": "Lingjie Liu3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 394, + 175, + 495, + 188 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 394, + 175, + 495, + 188 + ], + "spans": [ + { + "bbox": [ + 394, + 175, + 495, + 188 + ], + "type": "text", + "content": "Christian Theobalt1,4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 74, + 189, + 520, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 74, + 189, + 520, + 203 + ], + "spans": [ + { + "bbox": [ + 74, + 189, + 520, + 203 + ], + "type": "text", + "content": "1MPI Informatics & Saarland Informatics Campus 2Google 3University of Pennsylvania" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 75, + 203, + 515, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 203, + 515, + 217 + ], + "spans": [ + { + "bbox": [ + 75, + 203, + 515, + 217 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 75, + 203, + 515, + 217 + ], + "type": "text", + "content": "Saarbrücken Research Center for Visual Computing, Interaction and Artificial Intelligence" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 172, + 224, + 420, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 172, + 224, + 420, + 236 + ], + "spans": [ + { + "bbox": [ + 172, + 224, + 420, + 236 + ], + "type": "text", + "content": "Project Page: https://jianwang-mpi.github.io/egowholemocap" + } + ] + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 48, + 245, + 126, + 336 + ], + "blocks": [ + { + "bbox": [ + 48, + 245, + 126, + 336 + ], + "lines": [ + { + "bbox": [ + 48, + 245, + 126, + 336 + ], + "spans": [ + { + "bbox": [ + 48, + 245, + 126, + 336 + ], + "type": "image", + "image_path": "2c7264137600ce24089392af138465890eff1926ff7756b800fdf6649855f9de.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 347, + 546, + 370 + ], + "lines": [ + { + "bbox": [ + 46, + 347, + 546, + 370 + ], + "spans": [ + { + "bbox": [ + 46, + 347, + 546, + 370 + ], + "type": "text", + "content": "Figure 1. From an image sequence captured by a single head-mounted fisheye camera, our method can predict accurate and temporally coherent whole-body motion, including human body and hand poses. The SMPL-X parameters are obtained using inverse kinematics." + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 126, + 245, + 247, + 336 + ], + "blocks": [ + { + "bbox": [ + 126, + 245, + 247, + 336 + ], + "lines": [ + { + "bbox": [ + 126, + 245, + 247, + 336 + ], + "spans": [ + { + "bbox": [ + 126, + 245, + 247, + 336 + ], + "type": "image", + "image_path": "36b3ea0ad948dd017e6c69274be3bab28c3522b0410a21094784063b16913666.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 249, + 245, + 294, + 336 + ], + "blocks": [ + { + "bbox": [ + 249, + 245, + 294, + 336 + ], + "lines": [ + { + "bbox": [ + 249, + 245, + 294, + 336 + ], + "spans": [ + { + "bbox": [ + 249, + 245, + 294, + 336 + ], + "type": "image", + "image_path": "3680f9941ff67bc2aed8921004f6f30027b0736c4f72dd633f1953249cb8c4a6.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 300, + 245, + 376, + 336 + ], + "blocks": [ + { + "bbox": [ + 300, + 245, + 376, + 336 + ], + "lines": [ + { + "bbox": [ + 300, + 245, + 376, + 336 + ], + "spans": [ + { + "bbox": [ + 300, + 245, + 376, + 336 + ], + "type": "image", + "image_path": "d1184db6933fba68a3c957ead65e29b2421106ab5e8f41b12ea82315382b41a3.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 377, + 245, + 498, + 336 + ], + "blocks": [ + { + "bbox": [ + 377, + 245, + 498, + 336 + ], + "lines": [ + { + "bbox": [ + 377, + 245, + 498, + 336 + ], + "spans": [ + { + "bbox": [ + 377, + 245, + 498, + 336 + ], + "type": "image", + "image_path": "e229efc3cee3db2f232971165b7e0bf10a4d372ff2372289390bd48b23a8ebe0.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 499, + 245, + 544, + 336 + ], + "blocks": [ + { + "bbox": [ + 499, + 245, + 544, + 336 + ], + "lines": [ + { + "bbox": [ + 499, + 245, + 544, + 336 + ], + "spans": [ + { + "bbox": [ + 499, + 245, + 544, + 336 + ], + "type": "image", + "image_path": "bda988de8371450c63d11fa9d74da986daaabeb32f388b3626bbe0e66cfca2a7.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "bbox": [ + 143, + 377, + 192, + 390 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 377, + 192, + 390 + ], + "spans": [ + { + "bbox": [ + 143, + 377, + 192, + 390 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 46, + 403, + 289, + 656 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 403, + 289, + 656 + ], + "spans": [ + { + "bbox": [ + 46, + 403, + 289, + 656 + ], + "type": "text", + "content": "In this work, we explore egocentric whole-body motion capture using a single fisheye camera, which simultaneously estimates human body and hand motion. This task presents significant challenges due to three factors: the lack of high-quality datasets, fisheye camera distortion, and human body self-occlusion. To address these challenges, we propose a novel approach that leverages FisheyeViT to extract fisheye image features, which are subsequently converted into pixel-aligned 3D heatmap representations for 3D human body pose prediction. For hand tracking, we incorporate dedicated hand detection and hand pose estimation networks for regressing 3D hand poses. Finally, we develop a diffusion-based whole-body motion prior model to refine the estimated whole-body motion while accounting for joint uncertainties. To train these networks, we collect a large synthetic dataset, EgoWholeBody, comprising 840,000 high-quality egocentric images captured across a diverse range of whole-body motion sequences. Quantitative and qualitative evaluations demonstrate the effectiveness of our method in producing high-quality whole-body motion estimates from a single egocentric camera." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 47, + 669, + 128, + 681 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 669, + 128, + 681 + ], + "spans": [ + { + "bbox": [ + 47, + 669, + 128, + 681 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 46, + 689, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 689, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 689, + 288, + 715 + ], + "type": "text", + "content": "Egocentric 3D human motion estimation using head-mounted devices [47, 54] has garnered significant traction" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 304, + 379, + 547, + 631 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 379, + 547, + 631 + ], + "spans": [ + { + "bbox": [ + 304, + 379, + 547, + 631 + ], + "type": "text", + "content": "in recent years, driven by its diverse applications in VR/AR. Immersed in a virtual world, we can traverse virtual environments, interact with virtual objects, and even simulate real-world interactions. To fully capture the intricacies of human motion during such interaction, understanding both body and hand movements is essential. While existing egocentric motion capture methods [30, 47, 50-52, 54] focus solely on body motion, neglecting the hands, this work proposes the task of egocentric whole-body motion capture, i.e. simultaneous estimation of the body motion and hand motion from a single head-mounted fisheye camera (shown in Fig. 1). This task is extremely challenging due to three factors: First, the fisheye image introduces significant distortion, making it difficult for existing networks, which are designed for non-distorted images, to extract features. Second, the egocentric camera perspective frequently leads to the occlusion of body parts, such as the feet and hands, further complicating the task of whole-body motion capture. Lastly, large-scale training data with ground truth annotations for both body and hand poses is absent in existing datasets [4, 29, 47, 51, 54]." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 304, + 641, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 641, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 641, + 547, + 715 + ], + "type": "text", + "content": "In this work, we propose a novel egocentric whole-body motion capture method to address the aforementioned challenges. To effectively address fisheye distortion, we propose FisheyeViT for extracting image features, along with a joint regressor employing pixel-aligned 3D heatmap for predicting 3D body poses. Instead of attempting to undis-" + } + ] + } + ], + "index": 26 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "text", + "content": "777" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 289, + 360 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 289, + 360 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 289, + 360 + ], + "type": "text", + "content": "tort the entire fisheye image, which is impractical due to the fisheye lens's large field of view (FOV), we opt to partition the image into smaller patches aligned with a specific FOV range. This approach enables individual patch-level undistortion and seamlessly aligns with the vision transformer architecture that is employed for extracting the complete image feature map. We further propose an egocentric 3D pose regressor utilizing 3D heatmap representations. Unlike the existing approach [52] that projects image features into 3D space through fisheye reprojection functions and regresses 3D heatmaps with V2V networks [33] leading to intricate network learning and high computational complexity—our proposed egocentric pose regressor adopts a simpler approach. It employs deconvolutional layers to obtain pixel-aligned 3D heatmaps. Notably, the voxels in the 3D heatmap directly correspond to pixels in 2D features, subsequently linking to image patches in FisheyeViT. This streamlined approach significantly simplifies network training. Joint locations from the pixel-aligned 3D heatmap are finally transformed with the fisheye camera model to obtain the 3D human body poses. Due to the large size difference between body and hands, we train a hand detection network and a hand pose estimation network to accurately regress 3D hand poses." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 361, + 289, + 566 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 361, + 289, + 566 + ], + "spans": [ + { + "bbox": [ + 46, + 361, + 289, + 566 + ], + "type": "text", + "content": "To overcome the challenges posed by self-occlusion and improve the accuracy of pose estimation, we propose a novel method for refining the whole-body motion predictions by incorporating temporal context and a motion prior. Our method learns a whole-body motion prior with the diffusion model [18] from a collection of diverse human motion sequences, capturing intrinsic correlations between hand and body movements. Following this, we extract the joint uncertainties from the pixel-aligned 3D heatmap and utilize them to guide the refinement of the whole-body motion. The joint uncertainties act as indicators of the trustworthiness of the pose regressor's predictions. By conditioning on joints with low uncertainty, our whole-body motion diffusion model selectively refines joints with high uncertainty. This strategy substantially improves the quality of whole-body pose estimations and effectively mitigates the effects of self-occlusion." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 567, + 287, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 567, + 287, + 664 + ], + "spans": [ + { + "bbox": [ + 46, + 567, + 287, + 664 + ], + "type": "text", + "content": "In response to the absence of the egocentric whole-body motion capture datasets, we present EgoWholeBody, a new large-scale high-quality synthetic dataset. This dataset encompasses a wide range of whole-body motions, comprising over 870k frames, which significantly surpasses the size of previous egocentric training datasets. EgoWholeBody could serve as a valuable resource for advancing research in egocentric whole-body motion capture." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 666, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 666, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 47, + 666, + 288, + 715 + ], + "type": "text", + "content": "A thorough evaluation across a range of datasets, including SceneEgo [52], GlobalEgoMocap [50] and " + }, + { + "bbox": [ + 47, + 666, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\mathrm{Mo}^2\\mathrm{Cap}^2" + }, + { + "bbox": [ + 47, + 666, + 288, + 715 + ], + "type": "text", + "content": " [54], has demonstrated the remarkable improvements of our method in estimating egocentric whole-body" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "text", + "content": "motion compared to previous approaches. This substantiates the effectiveness of our approach in addressing the special challenges encountered in egocentric views, including the fisheye distortion and self-occlusion." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 305, + 121, + 520, + 132 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 121, + 520, + 132 + ], + "spans": [ + { + "bbox": [ + 305, + 121, + 520, + 132 + ], + "type": "text", + "content": "In summary, our key contributions are the following:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 133, + 545, + 266 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 306, + 133, + 545, + 169 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 133, + 545, + 169 + ], + "spans": [ + { + "bbox": [ + 306, + 133, + 545, + 169 + ], + "type": "text", + "content": "- The first egocentric whole-body motion capture method that predicts accurate and temporarily coherent egocentric body and hand motion;" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 170, + 545, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 170, + 545, + 205 + ], + "spans": [ + { + "bbox": [ + 306, + 170, + 545, + 205 + ], + "type": "text", + "content": "- FisheyeViT for alleviating fisheye camera distortion and pose regressor using pixel-aligned 3D heatmaps for accurate egocentric body pose estimation from a single image;" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 205, + 545, + 241 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 205, + 545, + 241 + ], + "spans": [ + { + "bbox": [ + 306, + 205, + 545, + 241 + ], + "type": "text", + "content": "- Uncertainty-aware refinement method based on motion diffusion models for correcting initial pose estimations and predicting plausible motions even under occlusion;" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 241, + 545, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 241, + 545, + 266 + ], + "spans": [ + { + "bbox": [ + 306, + 241, + 545, + 266 + ], + "type": "text", + "content": "- EgoWholeBody, a new high-quality synthetic dataset for egocentric whole-body motion capture." + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 274, + 392, + 288 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 274, + 392, + 288 + ], + "spans": [ + { + "bbox": [ + 306, + 274, + 392, + 288 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 295, + 545, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 295, + 545, + 437 + ], + "spans": [ + { + "bbox": [ + 304, + 295, + 545, + 437 + ], + "type": "text", + "content": "Egocentric 3D Human Body Pose Estimation. Recently, there has been growing interest in estimating egocentric 3D poses from body-worn cameras. Some methods [21, 25, 31, 35, 59, 60] use front-facing cameras and infer the human body motion from the camera view. However, since the user's body is often unobserved by the camera, these methods fail when the human body is not roaming around. Millerdurai et al. [32] leverage event cameras for estimating egocentric body pose. Other methods [4, 5, 7, 23, 39, 65] use head-mounted down-facing stereo cameras to estimate body poses. However, stereo camera setups introduce extra burdens of weight and energy consumption." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 438, + 546, + 654 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 438, + 546, + 654 + ], + "spans": [ + { + "bbox": [ + 304, + 438, + 546, + 654 + ], + "type": "text", + "content": "Xu et al. [54] and Tome et al. [47] introduce the single head-mounted down-facing fisheye camera setup for the egocentric 3D human pose estimation task. Zhang et al. [64] regressed fisheye camera parameters and 3D human pose simultaneously. To address the self-occlusion issue, Park et al. [36] leveraged the temporal information with the spatio-temporal self-attention network, and Liu et al. [30] introduced diffusion model to generate 3D human pose conditioned on egocentric image features. Wang et al. [50] and Liu et al. [28] combined the SLAM and egocentric pose estimation methods to estimate human body poses in the world coordinate. Wang et al. [51] and Liu et al. [29] leverage the synchronized egocentric camera and external cameras to collect large-scale egocentric pose estimation datasets with pseudo-ground truth. Considering the human-scene interaction, Wang et al. [52] estimated the scene geometry from the egocentric camera and constrained the 3D human pose with it." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 654, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 547, + 715 + ], + "type": "text", + "content": "These methods only focus on estimating human body poses while omitting the hand motion, and they still suffer from fisheye camera distortion since they directly put the highly distorted fisheye images into the neural network. Our proposed method can capture whole-body motion and" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "text", + "content": "778" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 49, + 70, + 545, + 240 + ], + "blocks": [ + { + "bbox": [ + 49, + 70, + 545, + 240 + ], + "lines": [ + { + "bbox": [ + 49, + 70, + 545, + 240 + ], + "spans": [ + { + "bbox": [ + 49, + 70, + 545, + 240 + ], + "type": "image", + "image_path": "0db61e7538dc84f37f270d5fe48c34a96602a8642274b91ef635766c67b4affc.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 247, + 546, + 304 + ], + "lines": [ + { + "bbox": [ + 46, + 247, + 546, + 304 + ], + "spans": [ + { + "bbox": [ + 46, + 247, + 546, + 304 + ], + "type": "text", + "content": "Figure 2. Overview of our whole-body motion capture pipeline. We first use FisheyeViT to undistort the input image and generate image feature tokens (3.1.1). Next, we use a 1D convolutional network to convert the image features to a pixel-aligned 3D heatmap and use soft-argmax and fisheye camera undistortion function to obtain the 3D body joins positions and uncertainty (3.1.2). We further detect the hand location and regress the 3D hand poses from the input image (3.1.3). Finally, the estimated hand motion and human body motion are combined and the uncertainty-aware diffusion model is applied to refine the estimated whole-body motion (3.2)." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 316, + 287, + 339 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 316, + 287, + 339 + ], + "spans": [ + { + "bbox": [ + 46, + 316, + 287, + 339 + ], + "type": "text", + "content": "resolve the fisheye camera distortion issue with the FisheyeViT and pixel-aligned 3D heatmap." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 340, + 287, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 340, + 287, + 544 + ], + "spans": [ + { + "bbox": [ + 46, + 340, + 287, + 544 + ], + "type": "text", + "content": "Whole-Body 3D Pose Estimation. Whole-body 3D pose estimation aims to estimate the 3D human body, face, and hands parameters from input images. This task is crucial for many applications, e.g., modeling human activities and human-scene interactions. Some methods [37, 53] fit the 2D body joints estimated from images with optimization algorithms, while these methods suffer from high computation overhead and can fall into local optima. Some other learning-based methods [6, 9, 15, 27, 40, 45, 66] use the neural network to regress the SMPL-X [37] parameters from input images. For example, ExPose [9] introduced body-driven attention to extract face and hand crops and used a refinement module to regress whole-body pose. OSX [27] proposed a one-stage pipeline for whole-body mesh recovery without separate networks for each part. SMPLer-X [6] propose a foundation model for whole-body pose estimation trained with the large model and big data." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 544, + 287, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 544, + 287, + 616 + ], + "spans": [ + { + "bbox": [ + 46, + 544, + 287, + 616 + ], + "type": "text", + "content": "Though much progress has been made on whole-body pose estimation from an external view, the task from an ego-centric view is still unexplored. In this paper, we introduce the first whole-body 3D pose estimation method from a single egocentric image and also incorporate temporal information with diffusion-based motion refinement." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 617, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 617, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 617, + 287, + 714 + ], + "type": "text", + "content": "Diffusion Models for Pose Estimation. Recently, some methods [8, 10, 16, 17, 19, 42] have effectively applied Denoising Diffusion Probabilistic Models (DDPM) [18] to human pose estimation tasks. Building on the success of motion diffusion models in human pose estimation, many methods have extended this approach to egocentric pose estimation, where the human body is only partially visible from RGB cameras or VR sensors. Zhang et al.'s work [63]" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 316, + 545, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 316, + 545, + 399 + ], + "spans": [ + { + "bbox": [ + 304, + 316, + 545, + 399 + ], + "type": "text", + "content": "uses a diffusion model to generate realistic human poses considering scene geometry. AGROL [14] generates body motion based on head and hand 6D pose estimates from a VR headset. EgoEgo [25] estimates head poses from a head-mounted front-facing camera and uses them to generate body poses. EgoHMR [30] extracts image features and uses them as a condition for the diffusion denoising process." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 399, + 546, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 399, + 546, + 567 + ], + "spans": [ + { + "bbox": [ + 304, + 399, + 546, + 567 + ], + "type": "text", + "content": "However, the aforementioned pose estimation methods train the conditioned diffusion model with image features or IMU signals. This cannot be generalized since the trained network only accepts one specific condition format and is inclined to learn domain-specific distributions of condition features. ZeDO [22] tackles this issue with a zero-shot diffusion-based optimization approach that doesn't require training with 2D-3D or image-3D pairs. Our method leverages the uncertainty value given by the single-frame pose estimation network and refines the initial motion estimation with the uncertainty of each joint. Moreover, different from previous methods that only focus on human body motion, we train a whole-body motion diffusion model to construct the relationship between hand and body motion." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 577, + 361, + 589 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 577, + 361, + 589 + ], + "spans": [ + { + "bbox": [ + 306, + 577, + 361, + 589 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 597, + 545, + 633 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 597, + 545, + 633 + ], + "spans": [ + { + "bbox": [ + 304, + 597, + 545, + 633 + ], + "type": "text", + "content": "In this section, we propose a new method for predicting accurate egocentric whole-body poses from egocentric image sequences. An overview of our approach is shown in Fig. 2." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 640, + 545, + 653 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 640, + 545, + 653 + ], + "spans": [ + { + "bbox": [ + 306, + 640, + 545, + 653 + ], + "type": "text", + "content": "3.1. Single Image Based Egocentric Pose Estimation" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 658, + 388, + 670 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 658, + 388, + 670 + ], + "spans": [ + { + "bbox": [ + 306, + 658, + 388, + 670 + ], + "type": "text", + "content": "3.1.1 FisheyeViT" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 677, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 546, + 713 + ], + "type": "text", + "content": "In this section, we introduce FisheyeViT, which is specially designed to alleviate the fisheye distortion issue. Instead of undistorting the entire fisheye image, we extract undistorted" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "text", + "content": "779" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 53, + 70, + 162, + 160 + ], + "blocks": [ + { + "bbox": [ + 53, + 70, + 162, + 160 + ], + "lines": [ + { + "bbox": [ + 53, + 70, + 162, + 160 + ], + "spans": [ + { + "bbox": [ + 53, + 70, + 162, + 160 + ], + "type": "image", + "image_path": "9012d87f36fd44a5ad44837bd96db815a063baf3c73872f6d559e4d84768d640.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 164, + 70, + 282, + 162 + ], + "blocks": [ + { + "bbox": [ + 164, + 70, + 282, + 162 + ], + "lines": [ + { + "bbox": [ + 164, + 70, + 282, + 162 + ], + "spans": [ + { + "bbox": [ + 164, + 70, + 282, + 162 + ], + "type": "image", + "image_path": "a150273fbd21bd7d66a798fb15085301439938063a88b8ed59b7b973e509a2aa.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 53, + 163, + 161, + 255 + ], + "blocks": [ + { + "bbox": [ + 53, + 163, + 161, + 255 + ], + "lines": [ + { + "bbox": [ + 53, + 163, + 161, + 255 + ], + "spans": [ + { + "bbox": [ + 53, + 163, + 161, + 255 + ], + "type": "image", + "image_path": "785fbab66b2e9d0cfba8efc7a302c7cbc151a05f21cd9c45d1c6f3ea85e1794c.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 53, + 264, + 277, + 275 + ], + "lines": [ + { + "bbox": [ + 53, + 264, + 277, + 275 + ], + "spans": [ + { + "bbox": [ + 53, + 264, + 277, + 275 + ], + "type": "text", + "content": "Figure 3. The detailed illustration of FisheyeViT (Sec. 3.1.1)." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 164, + 163, + 282, + 255 + ], + "blocks": [ + { + "bbox": [ + 164, + 163, + 282, + 255 + ], + "lines": [ + { + "bbox": [ + 164, + 163, + 282, + 255 + ], + "spans": [ + { + "bbox": [ + 164, + 163, + 282, + 255 + ], + "type": "image", + "image_path": "3897e648c0b37f699b97766125124aa2e6c52ad507a6c3128e166b975a50c9ed.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 285, + 286, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 285, + 286, + 357 + ], + "spans": [ + { + "bbox": [ + 46, + 285, + 286, + 357 + ], + "type": "text", + "content": "image patches from the fisheye image and then fit these patches as tokens into the transformer network [13]. To get the undistorted patches, we first warp the fisheye image to a unit semi-sphere, then get the patches with the gnomonic projection (see Fig. 2). The FisheyeViT can be split into five steps, the first four of which are illustrated in Fig. 3." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 357, + 287, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 357, + 287, + 502 + ], + "spans": [ + { + "bbox": [ + 47, + 357, + 287, + 502 + ], + "type": "text", + "content": "Step 1. Given an input image " + }, + { + "bbox": [ + 47, + 357, + 287, + 502 + ], + "type": "inline_equation", + "content": "\\mathbf{I}" + }, + { + "bbox": [ + 47, + 357, + 287, + 502 + ], + "type": "text", + "content": " with size " + }, + { + "bbox": [ + 47, + 357, + 287, + 502 + ], + "type": "inline_equation", + "content": "H\\times W" + }, + { + "bbox": [ + 47, + 357, + 287, + 502 + ], + "type": "text", + "content": ", we first evenly sample " + }, + { + "bbox": [ + 47, + 357, + 287, + 502 + ], + "type": "inline_equation", + "content": "N\\times N" + }, + { + "bbox": [ + 47, + 357, + 287, + 502 + ], + "type": "text", + "content": " patch center points: " + }, + { + "bbox": [ + 47, + 357, + 287, + 502 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{C}_{ij} = (u_i,v_j) = \\left(\\frac{H}{N} (i + \\frac{1}{2}),\\frac{W}{N} (j + \\frac{1}{2})\\right)|i,j\\in 0,\\dots,N - 1\\}" + }, + { + "bbox": [ + 47, + 357, + 287, + 502 + ], + "type": "text", + "content": ". Then, the patch center points " + }, + { + "bbox": [ + 47, + 357, + 287, + 502 + ], + "type": "inline_equation", + "content": "\\mathbf{C}_{ij}" + }, + { + "bbox": [ + 47, + 357, + 287, + 502 + ], + "type": "text", + "content": " are projected onto a unit sphere with the fisheye reprojection function: " + }, + { + "bbox": [ + 47, + 357, + 287, + 502 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_{ij}^{c} = (x_{ij}^{c},y_{ij}^{c},z_{ij}^{c}) = \\mathcal{P}^{-1}(u_i,v_j,1)" + }, + { + "bbox": [ + 47, + 357, + 287, + 502 + ], + "type": "text", + "content": ". The fisheye camera model is described in Sec. 8 of the supplementary material. Given a point " + }, + { + "bbox": [ + 47, + 357, + 287, + 502 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_{ij}^{c}" + }, + { + "bbox": [ + 47, + 357, + 287, + 502 + ], + "type": "text", + "content": " on the unit sphere, the tangent plane " + }, + { + "bbox": [ + 47, + 357, + 287, + 502 + ], + "type": "inline_equation", + "content": "\\mathbf{T}_{ij}" + }, + { + "bbox": [ + 47, + 357, + 287, + 502 + ], + "type": "text", + "content": " that passes through the point is defined by the normal vector " + }, + { + "bbox": [ + 47, + 357, + 287, + 502 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_{ij}^{c} = (x_{ij}^{c},y_{ij}^{c},z_{ij}^{c})" + }, + { + "bbox": [ + 47, + 357, + 287, + 502 + ], + "type": "text", + "content": ". In the following steps, we implement the gnomonic projection by sampling grid points in the plane and projecting them back onto the fisheye image." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 502, + 287, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 502, + 287, + 635 + ], + "spans": [ + { + "bbox": [ + 46, + 502, + 287, + 635 + ], + "type": "text", + "content": "Step 2. In this step, we determine the orientation of the grid points in the tangent plane, ensuring that the grid points from different tangent planes " + }, + { + "bbox": [ + 46, + 502, + 287, + 635 + ], + "type": "inline_equation", + "content": "\\mathbf{T}_{ij}" + }, + { + "bbox": [ + 46, + 502, + 287, + 635 + ], + "type": "text", + "content": " have the same orientation when projected back onto the fisheye image. To achieve this, we select a 2D point " + }, + { + "bbox": [ + 46, + 502, + 287, + 635 + ], + "type": "inline_equation", + "content": "\\mathbf{U}_{ij} = (u_i + d,v_j)" + }, + { + "bbox": [ + 46, + 502, + 287, + 635 + ], + "type": "text", + "content": " in the fisheye image space that is " + }, + { + "bbox": [ + 46, + 502, + 287, + 635 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 46, + 502, + 287, + 635 + ], + "type": "text", + "content": " pixels to the right of the patch center point and project it to the unit sphere using the fisheye reprojection function: " + }, + { + "bbox": [ + 46, + 502, + 287, + 635 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_{ij}^{u} = (x_{ij}^{u},y_{ij}^{u},z_{ij}^{u}) = \\mathcal{P}^{-1}(u_i + d,v_j,1)" + }, + { + "bbox": [ + 46, + 502, + 287, + 635 + ], + "type": "text", + "content": ". We then calculate the intersection point " + }, + { + "bbox": [ + 46, + 502, + 287, + 635 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_{ij}^{x}" + }, + { + "bbox": [ + 46, + 502, + 287, + 635 + ], + "type": "text", + "content": " between the vector " + }, + { + "bbox": [ + 46, + 502, + 287, + 635 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_{ij}^{u} = (x_{ij}^{u},y_{ij}^{u},z_{ij}^{u})" + }, + { + "bbox": [ + 46, + 502, + 287, + 635 + ], + "type": "text", + "content": " that is passing the origin and the tangent plane " + }, + { + "bbox": [ + 46, + 502, + 287, + 635 + ], + "type": "inline_equation", + "content": "\\mathbf{T}_{ij}" + }, + { + "bbox": [ + 46, + 502, + 287, + 635 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 86, + 641, + 287, + 673 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 86, + 641, + 287, + 673 + ], + "spans": [ + { + "bbox": [ + 86, + 641, + 287, + 673 + ], + "type": "interline_equation", + "content": "\\mathbf {P} _ {i j} ^ {x} = \\frac {\\left\\langle \\mathbf {P} _ {i j} ^ {c} , \\mathbf {v} _ {i j} ^ {c} \\right\\rangle}{\\left\\langle \\mathbf {v} _ {i j} ^ {u} , \\mathbf {v} _ {i j} ^ {c} \\right\\rangle} \\mathbf {v} _ {i j} ^ {u} = \\frac {1}{\\left\\langle \\mathbf {v} _ {i j} ^ {u} , \\mathbf {v} _ {i j} ^ {c} \\right\\rangle} \\mathbf {v} _ {i j} ^ {u}, \\tag {1}", + "image_path": "34cc5c2f2f009ab97118734dbf3dacacad6e4f7a074bb44a8f69743d20ebd629.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 677, + 201, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 677, + 201, + 689 + ], + "spans": [ + { + "bbox": [ + 47, + 677, + 201, + 689 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 677, + 201, + 689 + ], + "type": "inline_equation", + "content": "\\langle \\cdot ,\\cdot \\rangle" + }, + { + "bbox": [ + 47, + 677, + 201, + 689 + ], + "type": "text", + "content": " denotes the inner product." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 689, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 287, + 715 + ], + "type": "text", + "content": "Step 3. Based on the center point " + }, + { + "bbox": [ + 47, + 689, + 287, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_{ij}^{c}" + }, + { + "bbox": [ + 47, + 689, + 287, + 715 + ], + "type": "text", + "content": " and intersection point " + }, + { + "bbox": [ + 47, + 689, + 287, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_{ij}^{x}" + }, + { + "bbox": [ + 47, + 689, + 287, + 715 + ], + "type": "text", + "content": " on the tangent plane " + }, + { + "bbox": [ + 47, + 689, + 287, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{T}_{ij}" + }, + { + "bbox": [ + 47, + 689, + 287, + 715 + ], + "type": "text", + "content": ", we build a coordinate" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 72, + 545, + 121 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 121 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 121 + ], + "type": "text", + "content": "system with the " + }, + { + "bbox": [ + 305, + 72, + 545, + 121 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 305, + 72, + 545, + 121 + ], + "type": "text", + "content": " axis: " + }, + { + "bbox": [ + 305, + 72, + 545, + 121 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_{ij}^{x} = \\mathrm{Norm}(\\mathbf{P}_{ij}^{x} - \\mathbf{P}_{ij}^{c})" + }, + { + "bbox": [ + 305, + 72, + 545, + 121 + ], + "type": "text", + "content": ", the " + }, + { + "bbox": [ + 305, + 72, + 545, + 121 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 305, + 72, + 545, + 121 + ], + "type": "text", + "content": " axis: " + }, + { + "bbox": [ + 305, + 72, + 545, + 121 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_{ij}^{z} = \\mathrm{Norm}(\\mathbf{v}_{ij}^{c})" + }, + { + "bbox": [ + 305, + 72, + 545, + 121 + ], + "type": "text", + "content": " and the " + }, + { + "bbox": [ + 305, + 72, + 545, + 121 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 305, + 72, + 545, + 121 + ], + "type": "text", + "content": " axis: " + }, + { + "bbox": [ + 305, + 72, + 545, + 121 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_{ij}^{y} = \\mathbf{v}_{ij}^{x} \\times \\mathbf{v}_{ij}^{x}" + }, + { + "bbox": [ + 305, + 72, + 545, + 121 + ], + "type": "text", + "content": " where Norm denotes the normalize operation. We grid-sample " + }, + { + "bbox": [ + 305, + 72, + 545, + 121 + ], + "type": "inline_equation", + "content": "M \\times M" + }, + { + "bbox": [ + 305, + 72, + 545, + 121 + ], + "type": "text", + "content": " points in a " + }, + { + "bbox": [ + 305, + 72, + 545, + 121 + ], + "type": "inline_equation", + "content": "l \\times l" + }, + { + "bbox": [ + 305, + 72, + 545, + 121 + ], + "type": "text", + "content": " square on the " + }, + { + "bbox": [ + 305, + 72, + 545, + 121 + ], + "type": "inline_equation", + "content": "x-y" + }, + { + "bbox": [ + 305, + 72, + 545, + 121 + ], + "type": "text", + "content": " plane:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 353, + 129, + 545, + 150 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 353, + 129, + 545, + 150 + ], + "spans": [ + { + "bbox": [ + 353, + 129, + 545, + 150 + ], + "type": "interline_equation", + "content": "\\left\\{\\mathbf {P} _ {i j} ^ {m n} = \\mathbf {P} _ {i j} ^ {c} + \\left(l \\frac {m}{M} \\mathbf {v} _ {i j} ^ {x}, l \\frac {n}{M} \\mathbf {v} _ {i j} ^ {y}\\right) \\right\\} \\tag {2}", + "image_path": "3fc013f8fd1d67b225fe04043fb2244594ee8994561631060b6dbde3c7f1e234.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 159, + 545, + 173 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 159, + 545, + 173 + ], + "spans": [ + { + "bbox": [ + 305, + 159, + 545, + 173 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 159, + 545, + 173 + ], + "type": "inline_equation", + "content": "m,n\\in -\\frac{1}{2} (M - 1),\\ldots , - \\frac{3}{2}, - \\frac{1}{2},\\frac{1}{2},\\frac{3}{2},\\ldots \\frac{1}{2} (M - 1)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 173, + 545, + 258 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 173, + 545, + 258 + ], + "spans": [ + { + "bbox": [ + 305, + 173, + 545, + 258 + ], + "type": "text", + "content": "Step 4. The points " + }, + { + "bbox": [ + 305, + 173, + 545, + 258 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_{ij}^{mn}" + }, + { + "bbox": [ + 305, + 173, + 545, + 258 + ], + "type": "text", + "content": " are projected back to the fisheye image with the fisheye projection function: " + }, + { + "bbox": [ + 305, + 173, + 545, + 258 + ], + "type": "inline_equation", + "content": "\\mathbf{C}_{ij}^{mn} = \\mathcal{P}(\\mathbf{P}_{ij}^{mn})" + }, + { + "bbox": [ + 305, + 173, + 545, + 258 + ], + "type": "text", + "content": ". We then apply bilinear sampling to obtain the colors at points " + }, + { + "bbox": [ + 305, + 173, + 545, + 258 + ], + "type": "inline_equation", + "content": "\\mathbf{C}_{ij}^{mn}" + }, + { + "bbox": [ + 305, + 173, + 545, + 258 + ], + "type": "text", + "content": " of the input image " + }, + { + "bbox": [ + 305, + 173, + 545, + 258 + ], + "type": "inline_equation", + "content": "\\mathbf{I}" + }, + { + "bbox": [ + 305, + 173, + 545, + 258 + ], + "type": "text", + "content": ", yielding the undistorted image patch " + }, + { + "bbox": [ + 305, + 173, + 545, + 258 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_{ij}^{\\mathrm{undis}}" + }, + { + "bbox": [ + 305, + 173, + 545, + 258 + ], + "type": "text", + "content": ". Please also see the supplementary video for a visual demonstration of undistorted image patches and their movement on the fisheye image." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 258, + 545, + 329 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 258, + 545, + 329 + ], + "spans": [ + { + "bbox": [ + 304, + 258, + 545, + 329 + ], + "type": "text", + "content": "Step 5. The image patches " + }, + { + "bbox": [ + 304, + 258, + 545, + 329 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{I}_{ij}^{\\mathrm{undis}}\\}" + }, + { + "bbox": [ + 304, + 258, + 545, + 329 + ], + "type": "text", + "content": " are sent to a ViT transformer network [13] to obtain the feature tokens " + }, + { + "bbox": [ + 304, + 258, + 545, + 329 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{F}_{ij}\\}" + }, + { + "bbox": [ + 304, + 258, + 545, + 329 + ], + "type": "text", + "content": ". The feature token is further reshaped in " + }, + { + "bbox": [ + 304, + 258, + 545, + 329 + ], + "type": "inline_equation", + "content": "i\\times j" + }, + { + "bbox": [ + 304, + 258, + 545, + 329 + ], + "type": "text", + "content": " matrix and obtain the image feature " + }, + { + "bbox": [ + 304, + 258, + 545, + 329 + ], + "type": "inline_equation", + "content": "\\mathbf{F}" + }, + { + "bbox": [ + 304, + 258, + 545, + 329 + ], + "type": "text", + "content": ". In the FisheyeViT, we empirically chose " + }, + { + "bbox": [ + 304, + 258, + 545, + 329 + ], + "type": "inline_equation", + "content": "N = 16" + }, + { + "bbox": [ + 304, + 258, + 545, + 329 + ], + "type": "text", + "content": "; " + }, + { + "bbox": [ + 304, + 258, + 545, + 329 + ], + "type": "inline_equation", + "content": "M = 16" + }, + { + "bbox": [ + 304, + 258, + 545, + 329 + ], + "type": "text", + "content": "; " + }, + { + "bbox": [ + 304, + 258, + 545, + 329 + ], + "type": "inline_equation", + "content": "d = 8" + }, + { + "bbox": [ + 304, + 258, + 545, + 329 + ], + "type": "text", + "content": "; " + }, + { + "bbox": [ + 304, + 258, + 545, + 329 + ], + "type": "inline_equation", + "content": "l = 0.2m" + }, + { + "bbox": [ + 304, + 258, + 545, + 329 + ], + "type": "text", + "content": " given the image size " + }, + { + "bbox": [ + 304, + 258, + 545, + 329 + ], + "type": "inline_equation", + "content": "H = W = 256" + }, + { + "bbox": [ + 304, + 258, + 545, + 329 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 330, + 546, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 330, + 546, + 486 + ], + "spans": [ + { + "bbox": [ + 304, + 330, + 546, + 486 + ], + "type": "text", + "content": "Note that " + }, + { + "bbox": [ + 304, + 330, + 546, + 486 + ], + "type": "inline_equation", + "content": "\\mathbf{C}_{ij}^{mn}" + }, + { + "bbox": [ + 304, + 330, + 546, + 486 + ], + "type": "text", + "content": " is independent of the image " + }, + { + "bbox": [ + 304, + 330, + 546, + 486 + ], + "type": "inline_equation", + "content": "\\mathbf{I}" + }, + { + "bbox": [ + 304, + 330, + 546, + 486 + ], + "type": "text", + "content": ". This means that, given a fixed fisheye camera model, we can precompute " + }, + { + "bbox": [ + 304, + 330, + 546, + 486 + ], + "type": "inline_equation", + "content": "\\mathbf{C}_{ij}^{mn}" + }, + { + "bbox": [ + 304, + 330, + 546, + 486 + ], + "type": "text", + "content": " for all combinations of " + }, + { + "bbox": [ + 304, + 330, + 546, + 486 + ], + "type": "inline_equation", + "content": "m,n" + }, + { + "bbox": [ + 304, + 330, + 546, + 486 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 330, + 546, + 486 + ], + "type": "inline_equation", + "content": "i,j" + }, + { + "bbox": [ + 304, + 330, + 546, + 486 + ], + "type": "text", + "content": " in advance. This significantly speeds up both the training and evaluation processes. Furthermore, the number and dimensions of image patches " + }, + { + "bbox": [ + 304, + 330, + 546, + 486 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{I}_{ij}^{\\mathrm{undis}}\\}" + }, + { + "bbox": [ + 304, + 330, + 546, + 486 + ], + "type": "text", + "content": " match exactly with those in the traditional ViT network. This compatibility allows us to finetune existing ViT networks on our egocentric datasets. Our sampling strategy ensures that each image patch " + }, + { + "bbox": [ + 304, + 330, + 546, + 486 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_{ij}^{\\mathrm{undis}}" + }, + { + "bbox": [ + 304, + 330, + 546, + 486 + ], + "type": "text", + "content": " corresponds to the same FOV range in the fisheye camera. In our ablation study in Sec. 5.3, we show that FisheyeViT enhances the performance of the pose estimation network when applied to egocentric fisheye images." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 305, + 502, + 541, + 514 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 502, + 541, + 514 + ], + "spans": [ + { + "bbox": [ + 305, + 502, + 541, + 514 + ], + "type": "text", + "content": "3.1.2 Pose Regressor with Pixel-Aligned 3D Heatmap" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 521, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 521, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 521, + 545, + 713 + ], + "type": "text", + "content": "After collecting image features with FisheyeViT, we utilize a 3D heatmap-based network to estimate the body poses. The existing 3D heatmap-based pose regressors [34, 44] are designed for the weak-perspective cameras and predict the 3D heatmap in " + }, + { + "bbox": [ + 304, + 521, + 545, + 713 + ], + "type": "inline_equation", + "content": "xyz" + }, + { + "bbox": [ + 304, + 521, + 545, + 713 + ], + "type": "text", + "content": " space. Directly applying these regressors will result in misalignment between 3D heatmap features in " + }, + { + "bbox": [ + 304, + 521, + 545, + 713 + ], + "type": "inline_equation", + "content": "xyz" + }, + { + "bbox": [ + 304, + 521, + 545, + 713 + ], + "type": "text", + "content": " space and 2D image features in the fisheye image space. Therefore, we introduce a novel egocentric pose regressor that relies on the pixel-aligned 3D heatmap, tailored to address the needs of fisheye cameras. The idea is to regress the 3D heatmap in " + }, + { + "bbox": [ + 304, + 521, + 545, + 713 + ], + "type": "inline_equation", + "content": "uvd" + }, + { + "bbox": [ + 304, + 521, + 545, + 713 + ], + "type": "text", + "content": " space rather than traditional " + }, + { + "bbox": [ + 304, + 521, + 545, + 713 + ], + "type": "inline_equation", + "content": "xyz" + }, + { + "bbox": [ + 304, + 521, + 545, + 713 + ], + "type": "text", + "content": " space, where " + }, + { + "bbox": [ + 304, + 521, + 545, + 713 + ], + "type": "inline_equation", + "content": "uv" + }, + { + "bbox": [ + 304, + 521, + 545, + 713 + ], + "type": "text", + "content": " corresponds to the fisheye image " + }, + { + "bbox": [ + 304, + 521, + 545, + 713 + ], + "type": "inline_equation", + "content": "uv" + }, + { + "bbox": [ + 304, + 521, + 545, + 713 + ], + "type": "text", + "content": " space. Specifically, given a feature map " + }, + { + "bbox": [ + 304, + 521, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{F} \\in \\mathbb{R}^{C \\times N \\times N}" + }, + { + "bbox": [ + 304, + 521, + 545, + 713 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 521, + 545, + 713 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 304, + 521, + 545, + 713 + ], + "type": "text", + "content": " is the channel number, " + }, + { + "bbox": [ + 304, + 521, + 545, + 713 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 521, + 545, + 713 + ], + "type": "text", + "content": " is feature map height and width, we firstly use two deconvolutional layers to convert the feature map " + }, + { + "bbox": [ + 304, + 521, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{F}" + }, + { + "bbox": [ + 304, + 521, + 545, + 713 + ], + "type": "text", + "content": " into shape" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "text", + "content": "780" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 289, + 275 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 289, + 275 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 289, + 275 + ], + "type": "inline_equation", + "content": "(D_h\\times J,H_h,W_h)" + }, + { + "bbox": [ + 47, + 72, + 289, + 275 + ], + "type": "text", + "content": " , and further reshape it to pixel-aligned 3D heatmap " + }, + { + "bbox": [ + 47, + 72, + 289, + 275 + ], + "type": "inline_equation", + "content": "\\mathbf{H}\\in \\mathbb{R}^{J\\times D_h\\times H_h\\times W_h}" + }, + { + "bbox": [ + 47, + 72, + 289, + 275 + ], + "type": "text", + "content": " , where " + }, + { + "bbox": [ + 47, + 72, + 289, + 275 + ], + "type": "inline_equation", + "content": "J" + }, + { + "bbox": [ + 47, + 72, + 289, + 275 + ], + "type": "text", + "content": " is the joint number and " + }, + { + "bbox": [ + 47, + 72, + 289, + 275 + ], + "type": "inline_equation", + "content": "D_{h},H_{h},W_{h}" + }, + { + "bbox": [ + 47, + 72, + 289, + 275 + ], + "type": "text", + "content": " is the 3D heatmap depth, height and width. The illustration of pixel-aligned 3D heatmap is shown in Fig. 2. Next, we obtain the max-value positions " + }, + { + "bbox": [ + 47, + 72, + 289, + 275 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{J}}_b = \\{(u_i,v_i,d_i)\\mid i\\in 0,1,2,\\dots,J\\}" + }, + { + "bbox": [ + 47, + 72, + 289, + 275 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 47, + 72, + 289, + 275 + ], + "type": "inline_equation", + "content": "\\mathbf{H}" + }, + { + "bbox": [ + 47, + 72, + 289, + 275 + ], + "type": "text", + "content": " by the differentiable soft-argmax operation [44]. Here, we note that " + }, + { + "bbox": [ + 47, + 72, + 289, + 275 + ], + "type": "inline_equation", + "content": "u_{i}" + }, + { + "bbox": [ + 47, + 72, + 289, + 275 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 72, + 289, + 275 + ], + "type": "inline_equation", + "content": "v_{i}" + }, + { + "bbox": [ + 47, + 72, + 289, + 275 + ], + "type": "text", + "content": " correspond to the uv-coordinate of the 3D body joint projected in the fisheye image space, and " + }, + { + "bbox": [ + 47, + 72, + 289, + 275 + ], + "type": "inline_equation", + "content": "d_{i}" + }, + { + "bbox": [ + 47, + 72, + 289, + 275 + ], + "type": "text", + "content": " denotes the distance of the joint to the fisheye camera. Finally, the 3D body joints " + }, + { + "bbox": [ + 47, + 72, + 289, + 275 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{J}}_b = \\{(x_i,y_i,z_i)\\mid i\\in 0,1,2,\\dots,J\\}" + }, + { + "bbox": [ + 47, + 72, + 289, + 275 + ], + "type": "text", + "content": " are recovered with the fisheye reprojection function: " + }, + { + "bbox": [ + 47, + 72, + 289, + 275 + ], + "type": "inline_equation", + "content": "(x_{i},y_{i},z_{i}) =" + }, + { + "bbox": [ + 47, + 72, + 289, + 275 + ], + "type": "inline_equation", + "content": "\\mathcal{P}^{-1}(u_i,v_i,d_i)" + }, + { + "bbox": [ + 47, + 72, + 289, + 275 + ], + "type": "text", + "content": " . The predicted body pose " + }, + { + "bbox": [ + 47, + 72, + 289, + 275 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{J}}_b" + }, + { + "bbox": [ + 47, + 72, + 289, + 275 + ], + "type": "text", + "content": " is finally compared with the ground truth body pose " + }, + { + "bbox": [ + 47, + 72, + 289, + 275 + ], + "type": "inline_equation", + "content": "\\mathbf{J}_b" + }, + { + "bbox": [ + 47, + 72, + 289, + 275 + ], + "type": "text", + "content": " with the MSE loss. By first regressing 3D body poses in uvd space and then reproducing it, we ensure that the 3D heatmap is pixelaligned with the end-to-end training." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 277, + 289, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 277, + 289, + 552 + ], + "spans": [ + { + "bbox": [ + 47, + 277, + 289, + 552 + ], + "type": "text", + "content": "With the pixel-aligned heatmap, our proposed 3D pose regressor solves problems in all three types of previous egocentric joint regressors. First, " + }, + { + "bbox": [ + 47, + 277, + 289, + 552 + ], + "type": "inline_equation", + "content": "\\mathrm{Mo}^2\\mathrm{Cap}^2" + }, + { + "bbox": [ + 47, + 277, + 289, + 552 + ], + "type": "text", + "content": " [54] employs separate networks to predict 2D joint positions and joint distances. However, this method can yield unrealistic joint estimations because small errors in 2D joints can result in large errors in 3D joints due to the projection effect. Second, " + }, + { + "bbox": [ + 47, + 277, + 289, + 552 + ], + "type": "inline_equation", + "content": "xR" + }, + { + "bbox": [ + 47, + 277, + 289, + 552 + ], + "type": "text", + "content": "-egopose [47] and EgoHMR [30] directly regress the 3D joint positions. However, this method is agnostic to the fisheye camera parameters, making it suitable only for a specific camera configuration (e.g., camera parameters, head-mounted position, etc.). Third, SceneEgo [52] projects 2D features into 3D voxel space and uses a V2V network to regress 3D poses. Because of these, the SceneEgo method suffers from low accuracy and large computation overhead. Different from previous methods, our pose regressor with pixel-aligned 3D heatmap is versatile and efficient since it directly estimates 3D joints while also incorporating an explicitly parametrized fisheye camera model. Moreover, it can preserve the uncertainty of the estimated joints, which will be used in our uncertainty-aware motion refinement method (Sec. 3.2.2). Detailed comparison with other pose prediction heads is shown in Table 3." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 572, + 224, + 585 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 572, + 224, + 585 + ], + "spans": [ + { + "bbox": [ + 47, + 572, + 224, + 585 + ], + "type": "text", + "content": "3.1.3 Egocentric Hand Pose Estimation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 593, + 287, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 593, + 287, + 640 + ], + "spans": [ + { + "bbox": [ + 47, + 593, + 287, + 640 + ], + "type": "text", + "content": "In this section, we first train a network to detect hand pose locations and then train a 3D hand pose estimation network to regress 3D hand poses. Then, we describe how to integrate the estimated hand and body poses." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 642, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 642, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 642, + 288, + 714 + ], + "type": "text", + "content": "Hand Detection. Given an input image " + }, + { + "bbox": [ + 47, + 642, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{I}" + }, + { + "bbox": [ + 47, + 642, + 288, + 714 + ], + "type": "text", + "content": ", we finetune the HRNet [49] network to regress the 2D hand poses of left hand " + }, + { + "bbox": [ + 47, + 642, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{J}_{lh}^{2d}" + }, + { + "bbox": [ + 47, + 642, + 288, + 714 + ], + "type": "text", + "content": " and right hand " + }, + { + "bbox": [ + 47, + 642, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{J}_{rh}^{2d}" + }, + { + "bbox": [ + 47, + 642, + 288, + 714 + ], + "type": "text", + "content": ". From the hand poses, we obtain the center point of left hand " + }, + { + "bbox": [ + 47, + 642, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{C}_{lh}" + }, + { + "bbox": [ + 47, + 642, + 288, + 714 + ], + "type": "text", + "content": " and right hand " + }, + { + "bbox": [ + 47, + 642, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{C}_{rh}" + }, + { + "bbox": [ + 47, + 642, + 288, + 714 + ], + "type": "text", + "content": " along with the bounding box sizes, " + }, + { + "bbox": [ + 47, + 642, + 288, + 714 + ], + "type": "inline_equation", + "content": "d_{lh}" + }, + { + "bbox": [ + 47, + 642, + 288, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 642, + 288, + 714 + ], + "type": "inline_equation", + "content": "d_{rh}" + }, + { + "bbox": [ + 47, + 642, + 288, + 714 + ], + "type": "text", + "content": ". We use our approach described in Sec. 3.1.1 to compute undistorted" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 305, + 72, + 489, + 84 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 489, + 84 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 489, + 84 + ], + "type": "text", + "content": "image patches of left " + }, + { + "bbox": [ + 305, + 72, + 489, + 84 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_{lh}" + }, + { + "bbox": [ + 305, + 72, + 489, + 84 + ], + "type": "text", + "content": " and right hands " + }, + { + "bbox": [ + 305, + 72, + 489, + 84 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_{rh}" + }, + { + "bbox": [ + 305, + 72, + 489, + 84 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 305, + 84, + 545, + 131 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 84, + 545, + 131 + ], + "spans": [ + { + "bbox": [ + 305, + 84, + 545, + 131 + ], + "type": "text", + "content": "Hand Pose Estimation. Given the cropped image " + }, + { + "bbox": [ + 305, + 84, + 545, + 131 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_{lh}" + }, + { + "bbox": [ + 305, + 84, + 545, + 131 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 305, + 84, + 545, + 131 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_{rh}" + }, + { + "bbox": [ + 305, + 84, + 545, + 131 + ], + "type": "text", + "content": ", we regress the 3D hand poses " + }, + { + "bbox": [ + 305, + 84, + 545, + 131 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{J}}_{lh}^{loc}" + }, + { + "bbox": [ + 305, + 84, + 545, + 131 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 84, + 545, + 131 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{J}}_{rh}^{loc}" + }, + { + "bbox": [ + 305, + 84, + 545, + 131 + ], + "type": "text", + "content": " with the Hand4Whole [34] network, which is fine-tuned on our Ego-FullBody dataset." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 133, + 546, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 133, + 546, + 335 + ], + "spans": [ + { + "bbox": [ + 304, + 133, + 546, + 335 + ], + "type": "text", + "content": "Integration of Body and Hand Positions. It is not straightforward to integrate the hand poses with the body pose in the egocentric camera view primarily due to the fisheye camera's perspective effects. Take the left hand as an example. Following Step 3 in Sec. 3.1.1, we establish a local coordinate system on the tangent plane of the left-hand image with XYZ axes as follows: " + }, + { + "bbox": [ + 304, + 133, + 546, + 335 + ], + "type": "inline_equation", + "content": "x:\\mathbf{v}_{lh}^{x};y:\\mathbf{v}_{lh}^{y};z:\\mathbf{v}_{lh}^{z}" + }, + { + "bbox": [ + 304, + 133, + 546, + 335 + ], + "type": "text", + "content": ". We define a rotation matrix, denoted as " + }, + { + "bbox": [ + 304, + 133, + 546, + 335 + ], + "type": "inline_equation", + "content": "\\mathbf{R}" + }, + { + "bbox": [ + 304, + 133, + 546, + 335 + ], + "type": "text", + "content": ", that represents the transformation between the root coordinate system and the local coordinate system on the tangent plane. The estimated hand pose is first rotated with the rotation matrix " + }, + { + "bbox": [ + 304, + 133, + 546, + 335 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{J}}_{lh} = \\mathbf{R}\\hat{\\mathbf{J}}_{lh}^{loc}" + }, + { + "bbox": [ + 304, + 133, + 546, + 335 + ], + "type": "text", + "content": " and then translated to align the wrist location of the human body. This same process is also applied to the right hand to get the right hand pose " + }, + { + "bbox": [ + 304, + 133, + 546, + 335 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{J}}_{rh}" + }, + { + "bbox": [ + 304, + 133, + 546, + 335 + ], + "type": "text", + "content": ". The whole-body joints " + }, + { + "bbox": [ + 304, + 133, + 546, + 335 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{J}}" + }, + { + "bbox": [ + 304, + 133, + 546, + 335 + ], + "type": "text", + "content": " are obtained by combining " + }, + { + "bbox": [ + 304, + 133, + 546, + 335 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{J}}_b" + }, + { + "bbox": [ + 304, + 133, + 546, + 335 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 133, + 546, + 335 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{J}}_{lh}" + }, + { + "bbox": [ + 304, + 133, + 546, + 335 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 133, + 546, + 335 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{J}}_{rh}" + }, + { + "bbox": [ + 304, + 133, + 546, + 335 + ], + "type": "text", + "content": ". The uncertainty of whole-body joints " + }, + { + "bbox": [ + 304, + 133, + 546, + 335 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{U}}" + }, + { + "bbox": [ + 304, + 133, + 546, + 335 + ], + "type": "text", + "content": " is also obtained from the maximal value of the 3D heatmap in pose estimation modules." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 342, + 497, + 354 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 342, + 497, + 354 + ], + "spans": [ + { + "bbox": [ + 305, + 342, + 497, + 354 + ], + "type": "text", + "content": "3.2. Diffusion-Based Motion Refinement" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 360, + 545, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 360, + 545, + 443 + ], + "spans": [ + { + "bbox": [ + 304, + 360, + 545, + 443 + ], + "type": "text", + "content": "We notice that the single-frame estimations in Sec. 3.1 suffer from inaccuracies and temporal instabilities. In this section, we propose a diffusion-based motion refinement method to tackle this problem. We first learn the whole-body motion prior with the motion diffusion model in Sec. 3.2.1 and then introduce an uncertainty-aware zero-shot motion refinement method in Sec. 3.2.2." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 458, + 497, + 470 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 458, + 497, + 470 + ], + "spans": [ + { + "bbox": [ + 305, + 458, + 497, + 470 + ], + "type": "text", + "content": "3.2.1 Whole-Body Motion Diffusion Model" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 476, + 545, + 549 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 476, + 545, + 549 + ], + "spans": [ + { + "bbox": [ + 304, + 476, + 545, + 549 + ], + "type": "text", + "content": "We follow DDPM [18] as our diffusion approach to capture the whole-body motion prior " + }, + { + "bbox": [ + 304, + 476, + 545, + 549 + ], + "type": "inline_equation", + "content": "q(\\mathbf{x})" + }, + { + "bbox": [ + 304, + 476, + 545, + 549 + ], + "type": "text", + "content": ". DDPM learns a distribution of whole-body motion " + }, + { + "bbox": [ + 304, + 476, + 545, + 549 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 304, + 476, + 545, + 549 + ], + "type": "text", + "content": " through a forward diffusion process and an inverse denoising process. The forward diffusion process is a Markov process of adding Gaussian noise over " + }, + { + "bbox": [ + 304, + 476, + 545, + 549 + ], + "type": "inline_equation", + "content": "t \\in \\{0, 1, \\dots, T - 1\\}" + }, + { + "bbox": [ + 304, + 476, + 545, + 549 + ], + "type": "text", + "content": " steps:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 345, + 555, + 545, + 569 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 345, + 555, + 545, + 569 + ], + "spans": [ + { + "bbox": [ + 345, + 555, + 545, + 569 + ], + "type": "interline_equation", + "content": "q \\left(\\mathbf {x} _ {t} \\mid \\mathbf {x} _ {t - 1}\\right) = \\mathcal {N} \\left(\\sqrt {\\alpha_ {t}} \\mathbf {x} _ {t - 1}, (1 - \\alpha_ {t}) I\\right) \\tag {3}", + "image_path": "6d6438e60cd647dcb083efbea71e426cb52d6ef522de7d56c740dd7165880b56.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 574, + 545, + 610 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 574, + 545, + 610 + ], + "spans": [ + { + "bbox": [ + 304, + 574, + 545, + 610 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 574, + 545, + 610 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_t" + }, + { + "bbox": [ + 304, + 574, + 545, + 610 + ], + "type": "text", + "content": " denotes the whole-body motion sequence at step " + }, + { + "bbox": [ + 304, + 574, + 545, + 610 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 574, + 545, + 610 + ], + "type": "text", + "content": ", the variance " + }, + { + "bbox": [ + 304, + 574, + 545, + 610 + ], + "type": "inline_equation", + "content": "(1 - \\alpha_{t}) \\in (0,1]" + }, + { + "bbox": [ + 304, + 574, + 545, + 610 + ], + "type": "text", + "content": " denotes a constant hyperparameter increases with " + }, + { + "bbox": [ + 304, + 574, + 545, + 610 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 574, + 545, + 610 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 611, + 545, + 694 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 611, + 545, + 694 + ], + "spans": [ + { + "bbox": [ + 304, + 611, + 545, + 694 + ], + "type": "text", + "content": "The inverse process uses a denoising network " + }, + { + "bbox": [ + 304, + 611, + 545, + 694 + ], + "type": "inline_equation", + "content": "D(\\cdot)" + }, + { + "bbox": [ + 304, + 611, + 545, + 694 + ], + "type": "text", + "content": " to remove the added Gaussian noise at each time step " + }, + { + "bbox": [ + 304, + 611, + 545, + 694 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 611, + 545, + 694 + ], + "type": "text", + "content": ". Here we use the transformer-based framework in EDGE [48] as the motion-denoising network " + }, + { + "bbox": [ + 304, + 611, + 545, + 694 + ], + "type": "inline_equation", + "content": "D(\\cdot)" + }, + { + "bbox": [ + 304, + 611, + 545, + 694 + ], + "type": "text", + "content": ". We follow Ramesh et al.'s work [38] to make the network predict the original signal itself, i.e. " + }, + { + "bbox": [ + 304, + 611, + 545, + 694 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{x}}_0 = D(\\mathbf{x}_t,t)" + }, + { + "bbox": [ + 304, + 611, + 545, + 694 + ], + "type": "text", + "content": " and train it with the simple objective [18]:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 322, + 700, + 545, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 322, + 700, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 322, + 700, + 545, + 715 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {s i m p l e}} = E _ {\\mathbf {x} _ {0} \\sim q (\\mathbf {x} _ {0}), t \\sim [ 1, T ]} \\left[ \\| \\mathbf {x} _ {0} - D (\\mathbf {x} _ {t}, t) \\| _ {2} ^ {2} \\right] \\tag {4}", + "image_path": "d7a1a90619ffb89c7340c53f1c8ec651d7ab89aaa50428c9e6ad1b5a8f248ad4.jpg" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "text", + "content": "781" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 247, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 247, + 83 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 247, + 83 + ], + "type": "text", + "content": "3.2.2 Uncertainty-Aware Motion Refinement" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 91, + 289, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 91, + 289, + 236 + ], + "spans": [ + { + "bbox": [ + 46, + 91, + 289, + 236 + ], + "type": "text", + "content": "Given the learned whole-body motion prior, we leverage the uncertainty value for each pose to guide the diffusion denoising process with the classifier-guided diffusion sampling [12]. Given an initial sequence of whole-body pose estimation " + }, + { + "bbox": [ + 46, + 91, + 289, + 236 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_e = \\{\\hat{\\mathbf{J}}_i\\}" + }, + { + "bbox": [ + 46, + 91, + 289, + 236 + ], + "type": "text", + "content": " and the uncertainty value for each pose " + }, + { + "bbox": [ + 46, + 91, + 289, + 236 + ], + "type": "inline_equation", + "content": "\\mathbf{u} = \\{\\hat{\\mathbf{U}}_i\\}" + }, + { + "bbox": [ + 46, + 91, + 289, + 236 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 91, + 289, + 236 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 91, + 289, + 236 + ], + "type": "text", + "content": " denotes the " + }, + { + "bbox": [ + 46, + 91, + 289, + 236 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 91, + 289, + 236 + ], + "type": "text", + "content": "th pose in the sequence, we keep the joints with low uncertainty but use the diffusion model to generate joints with high uncertainty conditioned on the low-uncertainty joints. Specifically, in the " + }, + { + "bbox": [ + 46, + 91, + 289, + 236 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 91, + 289, + 236 + ], + "type": "text", + "content": "th sampling step of the diffusion process, the denoising network predicts " + }, + { + "bbox": [ + 46, + 91, + 289, + 236 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{x}}_0 = D(\\mathbf{x}_t,t)" + }, + { + "bbox": [ + 46, + 91, + 289, + 236 + ], + "type": "text", + "content": ", which is noised back to " + }, + { + "bbox": [ + 46, + 91, + 289, + 236 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{t - 1}" + }, + { + "bbox": [ + 46, + 91, + 289, + 236 + ], + "type": "text", + "content": " by sampling from the Gaussian distribution:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 96, + 243, + 287, + 257 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 96, + 243, + 287, + 257 + ], + "spans": [ + { + "bbox": [ + 96, + 243, + 287, + 257 + ], + "type": "interline_equation", + "content": "\\mathbf {x} _ {t - 1} \\sim \\mathcal {N} \\left(\\hat {\\mathbf {x}} _ {0} + \\mathbf {w} \\left(\\mathbf {x} _ {e} - \\hat {\\mathbf {x}} _ {0}\\right), \\Sigma_ {t}\\right) \\tag {5}", + "image_path": "aa5377739aca1fd427957838c0adfcd1f68a558d70696105d21dfc8ff4db7921.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 263, + 289, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 263, + 289, + 399 + ], + "spans": [ + { + "bbox": [ + 46, + 263, + 289, + 399 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 263, + 289, + 399 + ], + "type": "inline_equation", + "content": "\\Sigma_{t}" + }, + { + "bbox": [ + 46, + 263, + 289, + 399 + ], + "type": "text", + "content": " is a scheduled Gaussian distribution in DDPM [18] and " + }, + { + "bbox": [ + 46, + 263, + 289, + 399 + ], + "type": "inline_equation", + "content": "\\mathbf{w}" + }, + { + "bbox": [ + 46, + 263, + 289, + 399 + ], + "type": "text", + "content": " controls the weight of a specific joint between the predicted motion " + }, + { + "bbox": [ + 46, + 263, + 289, + 399 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{x}}_0" + }, + { + "bbox": [ + 46, + 263, + 289, + 399 + ], + "type": "text", + "content": " and the estimated motion " + }, + { + "bbox": [ + 46, + 263, + 289, + 399 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_e" + }, + { + "bbox": [ + 46, + 263, + 289, + 399 + ], + "type": "text", + "content": ". Generally, we expect " + }, + { + "bbox": [ + 46, + 263, + 289, + 399 + ], + "type": "inline_equation", + "content": "\\mathbf{w} \\rightarrow \\overrightarrow{0}" + }, + { + "bbox": [ + 46, + 263, + 289, + 399 + ], + "type": "text", + "content": " when " + }, + { + "bbox": [ + 46, + 263, + 289, + 399 + ], + "type": "inline_equation", + "content": "t \\rightarrow 0" + }, + { + "bbox": [ + 46, + 263, + 289, + 399 + ], + "type": "text", + "content": " such that the temporal stability is guaranteed through the generation of the denoising process, and " + }, + { + "bbox": [ + 46, + 263, + 289, + 399 + ], + "type": "inline_equation", + "content": "\\mathbf{w} \\rightarrow \\overrightarrow{1}" + }, + { + "bbox": [ + 46, + 263, + 289, + 399 + ], + "type": "text", + "content": " when " + }, + { + "bbox": [ + 46, + 263, + 289, + 399 + ], + "type": "inline_equation", + "content": "t \\rightarrow T" + }, + { + "bbox": [ + 46, + 263, + 289, + 399 + ], + "type": "text", + "content": " such that the denoising process is initialized by the estimated motion " + }, + { + "bbox": [ + 46, + 263, + 289, + 399 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_e" + }, + { + "bbox": [ + 46, + 263, + 289, + 399 + ], + "type": "text", + "content": ". We also expect that " + }, + { + "bbox": [ + 46, + 263, + 289, + 399 + ], + "type": "inline_equation", + "content": "w_{ij} = \\mathbf{w}[i][j]" + }, + { + "bbox": [ + 46, + 263, + 289, + 399 + ], + "type": "text", + "content": ", which is the weight of " + }, + { + "bbox": [ + 46, + 263, + 289, + 399 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 46, + 263, + 289, + 399 + ], + "type": "text", + "content": "th joints in the " + }, + { + "bbox": [ + 46, + 263, + 289, + 399 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 263, + 289, + 399 + ], + "type": "text", + "content": "th pose, is smaller when the uncertainty value " + }, + { + "bbox": [ + 46, + 263, + 289, + 399 + ], + "type": "inline_equation", + "content": "u_{ij} = \\mathbf{u}[i][j]" + }, + { + "bbox": [ + 46, + 263, + 289, + 399 + ], + "type": "text", + "content": " of the " + }, + { + "bbox": [ + 46, + 263, + 289, + 399 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 46, + 263, + 289, + 399 + ], + "type": "text", + "content": "th joints in the " + }, + { + "bbox": [ + 46, + 263, + 289, + 399 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 263, + 289, + 399 + ], + "type": "text", + "content": "th pose is large. Based on this requirement, we design " + }, + { + "bbox": [ + 46, + 263, + 289, + 399 + ], + "type": "inline_equation", + "content": "\\mathbf{w}" + }, + { + "bbox": [ + 46, + 263, + 289, + 399 + ], + "type": "text", + "content": " as:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 113, + 405, + 287, + 426 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 113, + 405, + 287, + 426 + ], + "spans": [ + { + "bbox": [ + 113, + 405, + 287, + 426 + ], + "type": "interline_equation", + "content": "\\mathbf {w} = 1 / \\left(1 + e ^ {- k (t - T \\mathbf {u})}\\right) \\tag {6}", + "image_path": "d452c397bc9e957c32701b72eb535a3b27054d21a9025b7611b46683f4e61cda.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 432, + 287, + 494 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 432, + 287, + 494 + ], + "spans": [ + { + "bbox": [ + 46, + 432, + 287, + 494 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 432, + 287, + 494 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 46, + 432, + 287, + 494 + ], + "type": "text", + "content": " is the overall diffusion steps, " + }, + { + "bbox": [ + 46, + 432, + 287, + 494 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 432, + 287, + 494 + ], + "type": "text", + "content": " is a hyperparameter which is empirically set to 0.1. From the experimental results in Sec. 5, we demonstrate the effectiveness of uncertain-aware motion refinement and our uncertainty-guided diffusion sampling strategy." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 502, + 185, + 515 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 502, + 185, + 515 + ], + "spans": [ + { + "bbox": [ + 47, + 502, + 185, + 515 + ], + "type": "text", + "content": "4. EgoWholeBody Dataset" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 522, + 287, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 522, + 287, + 677 + ], + "spans": [ + { + "bbox": [ + 46, + 522, + 287, + 677 + ], + "type": "text", + "content": "In this section, we introduce EgoWholeBody, a large-scale high-quality synthetic dataset built for the task of egocentric whole-body motion capture. The EgoWholeBody dataset is organized into two sections. The first part, containing over 700k frames, is rendered with 14 different rigged Renderpeople [3] models driven by 2367 Mixamo [2] motion sequences. The second part focuses on hand motions and contains 170k frames with the SMPL-X model. This data is constructed from 24 different shapes and textures, driven by 262 motion sequences selected from the GRAB [46] and TCDHandMocap dataset [20]. We also created synthetic test sequences, which include 133k images rendered with 3 Renderpeople models and Mixamo motions." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 677, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 288, + 714 + ], + "type": "text", + "content": "During the rendering process, we first attach a virtual fisheye camera to the forehead of human body models and render the images, semantic labels, and depth map with" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 72, + 546, + 109 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 546, + 109 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 546, + 109 + ], + "type": "text", + "content": "Blender [1]. Our dataset is larger and more diverse than previous egocentric training datasets—see Sec. 10 in the supplementary material for a detailed comparison." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 120, + 388, + 133 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 120, + 388, + 133 + ], + "spans": [ + { + "bbox": [ + 305, + 120, + 388, + 133 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 140, + 481, + 152 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 140, + 481, + 152 + ], + "spans": [ + { + "bbox": [ + 305, + 140, + 481, + 152 + ], + "type": "text", + "content": "5.1. Datasets and Evaluation Metrics" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 158, + 546, + 255 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 158, + 546, + 255 + ], + "spans": [ + { + "bbox": [ + 304, + 158, + 546, + 255 + ], + "type": "text", + "content": "Training Datasets. To train our body pose estimation module (Sec. 3.1.1 and Sec. 3.1.2), we use our EgoWholeBody dataset and the EgoPW dataset [50]. Additionally, the EgoWholeBody dataset is used to train the hand pose estimation module in Sec. 3.1.3. For training the whole-body diffusion model (Sec. 3.2), we utilize a combined motion capture dataset that includes EgoBody [62], Mixamo [2], TCDHandMocap dataset [20] and GRAB dataset [46]." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 255, + 547, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 255, + 547, + 399 + ], + "spans": [ + { + "bbox": [ + 304, + 255, + 547, + 399 + ], + "type": "text", + "content": "Evaluation Datasets. In our experiment, we evaluate our methods on four datasets: the GlobalEgoMocap test datasets [50], the " + }, + { + "bbox": [ + 304, + 255, + 547, + 399 + ], + "type": "inline_equation", + "content": "\\mathrm{Mo}^2\\mathrm{Cap}^2" + }, + { + "bbox": [ + 304, + 255, + 547, + 399 + ], + "type": "text", + "content": " test dataset [54], the SceneEgo test dataset [52] and out EgoWholeBody test dataset. The details of the datasets are shown in Sec. 12 of supplementary materials. Note that evaluating whole-body poses requires accurate annotations for human hands, which is absent in real-world datasets. To resolve the issue, we request the multi-view videos of the SceneEgo test dataset [52] from the authors and use a multi-view motion capture system to obtain the hand motion. The hand pose annotations will be made publicly available." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 399, + 547, + 519 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 399, + 547, + 519 + ], + "spans": [ + { + "bbox": [ + 304, + 399, + 547, + 519 + ], + "type": "text", + "content": "Evaluation Metrics. To evaluate the precision of human body poses on the SceneEgo test dataset [52], we use MPJPE and PA-MPJPE. For the GlobalEgoMocap test dataset [50] and " + }, + { + "bbox": [ + 304, + 399, + 547, + 519 + ], + "type": "inline_equation", + "content": "\\mathrm{Mo}^2\\mathrm{Cap}^2" + }, + { + "bbox": [ + 304, + 399, + 547, + 519 + ], + "type": "text", + "content": " test dataset [54], where egocentric camera poses are unavailable, we evaluate PA-MPJPE and BA-MPJPE. For hand pose accuracy, we align the predicted and ground truth hand poses at the root position, followed by computing MPJPE and PA-MPJPE. Detailed explanations of these metrics are in Sec. 11 of the supplementary materials. All reported metrics are in millimeters." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 527, + 544, + 540 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 527, + 544, + 540 + ], + "spans": [ + { + "bbox": [ + 305, + 527, + 544, + 540 + ], + "type": "text", + "content": "5.2. Comparisons on Whole-Body Pose Estimation" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 545, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 545, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 545, + 547, + 714 + ], + "type": "text", + "content": "For a fair comparison with existing methods focusing solely on body or hand pose, we split our evaluation into two parts, reporting results of body poses in Table 1 and hand pose in Table 2. We first compare the accuracy of the human body poses with state-of-the-art methods, including EgoPW [51] and SceneEgo [52], on EgoWholeBody and SceneEgo [52] test datasets. The comparison with more previous methods [47, 50, 54] and on more evaluation datasets [50, 54] are shown in Sec. 7 of the supplementary materials. Since our motion refinement method incorporates random Gaussian noise, we generate five samples and calculate the average MPJPE values. The standard deviation is low (" + }, + { + "bbox": [ + 304, + 545, + 547, + 714 + ], + "type": "inline_equation", + "content": "< 0.01\\mathrm{mm}" + }, + { + "bbox": [ + 304, + 545, + 547, + 714 + ], + "type": "text", + "content": ") and is discussed in Sec. 13 of supplementary materials. Results are presented in Table 1, where our single-frame re" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "text", + "content": "782" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 51, + 70, + 102, + 168 + ], + "blocks": [ + { + "bbox": [ + 51, + 70, + 102, + 168 + ], + "lines": [ + { + "bbox": [ + 51, + 70, + 102, + 168 + ], + "spans": [ + { + "bbox": [ + 51, + 70, + 102, + 168 + ], + "type": "image", + "image_path": "2b36abfc7293d272285e3cd4a5923fe4d3975b24615caa51d348f4182cec6dee.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 67, + 169, + 84, + 176 + ], + "lines": [ + { + "bbox": [ + 67, + 169, + 84, + 176 + ], + "spans": [ + { + "bbox": [ + 67, + 169, + 84, + 176 + ], + "type": "text", + "content": "Input" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 106, + 74, + 139, + 167 + ], + "blocks": [ + { + "bbox": [ + 106, + 74, + 139, + 167 + ], + "lines": [ + { + "bbox": [ + 106, + 74, + 139, + 167 + ], + "spans": [ + { + "bbox": [ + 106, + 74, + 139, + 167 + ], + "type": "image", + "image_path": "7296f8f261090bbba4cc565a89c0a1263b78b4e35d582f0d76b7be038e6db2f8.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 113, + 169, + 135, + 176 + ], + "lines": [ + { + "bbox": [ + 113, + 169, + 135, + 176 + ], + "spans": [ + { + "bbox": [ + 113, + 169, + 135, + 176 + ], + "type": "text", + "content": "EgoPW" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 156, + 72, + 187, + 167 + ], + "blocks": [ + { + "bbox": [ + 156, + 72, + 187, + 167 + ], + "lines": [ + { + "bbox": [ + 156, + 72, + 187, + 167 + ], + "spans": [ + { + "bbox": [ + 156, + 72, + 187, + 167 + ], + "type": "image", + "image_path": "fd3208ba132e94b485215577d61f0e4ae8c98e8d844925559f66b5d3f0706c32.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 159, + 168, + 186, + 176 + ], + "lines": [ + { + "bbox": [ + 159, + 168, + 186, + 176 + ], + "spans": [ + { + "bbox": [ + 159, + 168, + 186, + 176 + ], + "type": "text", + "content": "SceneEgo" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 204, + 72, + 237, + 167 + ], + "blocks": [ + { + "bbox": [ + 204, + 72, + 237, + 167 + ], + "lines": [ + { + "bbox": [ + 204, + 72, + 237, + 167 + ], + "spans": [ + { + "bbox": [ + 204, + 72, + 237, + 167 + ], + "type": "image", + "image_path": "0416137e264b082666a001f0d877bdecb1233463c15eece3a3b594c8ffb40dc2.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 204, + 168, + 237, + 176 + ], + "lines": [ + { + "bbox": [ + 204, + 168, + 237, + 176 + ], + "spans": [ + { + "bbox": [ + 204, + 168, + 237, + 176 + ], + "type": "text", + "content": "Ours-Single" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 251, + 72, + 283, + 167 + ], + "blocks": [ + { + "bbox": [ + 251, + 72, + 283, + 167 + ], + "lines": [ + { + "bbox": [ + 251, + 72, + 283, + 167 + ], + "spans": [ + { + "bbox": [ + 251, + 72, + 283, + 167 + ], + "type": "image", + "image_path": "d116b8c08e886af64ea3404515370357c512628eee4cf8c1bec4e8f6b123554b.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 250, + 168, + 288, + 176 + ], + "lines": [ + { + "bbox": [ + 250, + 168, + 288, + 176 + ], + "spans": [ + { + "bbox": [ + 250, + 168, + 288, + 176 + ], + "type": "text", + "content": "Ours-Refined" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 297, + 70, + 348, + 168 + ], + "blocks": [ + { + "bbox": [ + 297, + 70, + 348, + 168 + ], + "lines": [ + { + "bbox": [ + 297, + 70, + 348, + 168 + ], + "spans": [ + { + "bbox": [ + 297, + 70, + 348, + 168 + ], + "type": "image", + "image_path": "eec0ee177d1a6d86bdbe3f5c7fd62a8b1c16990dd0ea4d339fd8768b2a0881fc.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 315, + 169, + 331, + 176 + ], + "lines": [ + { + "bbox": [ + 315, + 169, + 331, + 176 + ], + "spans": [ + { + "bbox": [ + 315, + 169, + 331, + 176 + ], + "type": "text", + "content": "Input" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 353, + 72, + 387, + 167 + ], + "blocks": [ + { + "bbox": [ + 353, + 72, + 387, + 167 + ], + "lines": [ + { + "bbox": [ + 353, + 72, + 387, + 167 + ], + "spans": [ + { + "bbox": [ + 353, + 72, + 387, + 167 + ], + "type": "image", + "image_path": "78bc9457063ba21f824a921e8d7d897ae814735a0288d48f7fd304888c07b4d0.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 361, + 168, + 381, + 176 + ], + "lines": [ + { + "bbox": [ + 361, + 168, + 381, + 176 + ], + "spans": [ + { + "bbox": [ + 361, + 168, + 381, + 176 + ], + "type": "text", + "content": "EgoPW" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 403, + 72, + 438, + 167 + ], + "blocks": [ + { + "bbox": [ + 403, + 72, + 438, + 167 + ], + "lines": [ + { + "bbox": [ + 403, + 72, + 438, + 167 + ], + "spans": [ + { + "bbox": [ + 403, + 72, + 438, + 167 + ], + "type": "image", + "image_path": "2db5bb07a47b9f653267dd865fcf90623f3e95809148e7cc681c9f05a7829dea.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 405, + 168, + 433, + 176 + ], + "lines": [ + { + "bbox": [ + 405, + 168, + 433, + 176 + ], + "spans": [ + { + "bbox": [ + 405, + 168, + 433, + 176 + ], + "type": "text", + "content": "SceneEgo" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 451, + 72, + 490, + 167 + ], + "blocks": [ + { + "bbox": [ + 451, + 72, + 490, + 167 + ], + "lines": [ + { + "bbox": [ + 451, + 72, + 490, + 167 + ], + "spans": [ + { + "bbox": [ + 451, + 72, + 490, + 167 + ], + "type": "image", + "image_path": "1f6fc40d8fa32c0cf3eb18c47bf63e527bb44f52c316f76a32a6410d14e1f311.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 453, + 168, + 486, + 176 + ], + "lines": [ + { + "bbox": [ + 453, + 168, + 486, + 176 + ], + "spans": [ + { + "bbox": [ + 453, + 168, + 486, + 176 + ], + "type": "text", + "content": "Ours-Single" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 500, + 72, + 539, + 167 + ], + "blocks": [ + { + "bbox": [ + 500, + 72, + 539, + 167 + ], + "lines": [ + { + "bbox": [ + 500, + 72, + 539, + 167 + ], + "spans": [ + { + "bbox": [ + 500, + 72, + 539, + 167 + ], + "type": "image", + "image_path": "0aef160c92c409446aaee3cae33e11765fb8998fec395ea53b2961ef82f61071.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 499, + 168, + 536, + 176 + ], + "lines": [ + { + "bbox": [ + 499, + 168, + 536, + 176 + ], + "spans": [ + { + "bbox": [ + 499, + 168, + 536, + 176 + ], + "type": "text", + "content": "Ours-Refined" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 51, + 220, + 104, + 323 + ], + "blocks": [ + { + "bbox": [ + 46, + 185, + 547, + 218 + ], + "lines": [ + { + "bbox": [ + 46, + 185, + 547, + 218 + ], + "spans": [ + { + "bbox": [ + 46, + 185, + 547, + 218 + ], + "type": "text", + "content": "Figure 4. Qualitative comparison on human body pose estimations between our methods and the state-of-the-art egocentric pose estimation methods on in-the-studio (left column) and in-the-wild scenes (right column). The red skeleton is the ground truth while the green skeleton is the predicted pose. Our methods predict more accurate body poses compared with EgoPW [51] and SceneEgo [52]." + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 51, + 220, + 104, + 323 + ], + "lines": [ + { + "bbox": [ + 51, + 220, + 104, + 323 + ], + "spans": [ + { + "bbox": [ + 51, + 220, + 104, + 323 + ], + "type": "image", + "image_path": "a705b4d478e7ae29a419d8aef00da1809a863112089c2bd25dd790aeac296a91.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 67, + 326, + 86, + 334 + ], + "lines": [ + { + "bbox": [ + 67, + 326, + 86, + 334 + ], + "spans": [ + { + "bbox": [ + 67, + 326, + 86, + 334 + ], + "type": "text", + "content": "Input" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 343, + 546, + 376 + ], + "lines": [ + { + "bbox": [ + 46, + 343, + 546, + 376 + ], + "spans": [ + { + "bbox": [ + 46, + 343, + 546, + 376 + ], + "type": "text", + "content": "Figure 5. Qualitative comparison on human hand pose estimations between our methods and the state-of-the-art third-view pose estimation methods. Our single-view and refined hand poses are more accurate than the poses from Hand4Whole [34] method. The red skeleton is the ground truth while the green skeleton is the predicted pose." + } + ] + } + ], + "index": 40, + "angle": 0, + "type": "image_caption" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 109, + 220, + 171, + 323 + ], + "blocks": [ + { + "bbox": [ + 109, + 220, + 171, + 323 + ], + "lines": [ + { + "bbox": [ + 109, + 220, + 171, + 323 + ], + "spans": [ + { + "bbox": [ + 109, + 220, + 171, + 323 + ], + "type": "image", + "image_path": "2511ddddfbb45f663922c3d6275ba93aaedca7fedc5dd82520a337637611a853.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 124, + 326, + 165, + 334 + ], + "lines": [ + { + "bbox": [ + 124, + 326, + 165, + 334 + ], + "spans": [ + { + "bbox": [ + 124, + 326, + 165, + 334 + ], + "type": "text", + "content": "Hand Image" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 176, + 223, + 216, + 322 + ], + "blocks": [ + { + "bbox": [ + 176, + 223, + 216, + 322 + ], + "lines": [ + { + "bbox": [ + 176, + 223, + 216, + 322 + ], + "spans": [ + { + "bbox": [ + 176, + 223, + 216, + 322 + ], + "type": "image", + "image_path": "a3e0d8f6c237c8e18267565ef03ef52150f8a31fa73290579d211711de55b529.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 173, + 326, + 217, + 334 + ], + "lines": [ + { + "bbox": [ + 173, + 326, + 217, + 334 + ], + "spans": [ + { + "bbox": [ + 173, + 326, + 217, + 334 + ], + "type": "text", + "content": "Hand4Whole" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_caption" + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 228, + 223, + 263, + 322 + ], + "blocks": [ + { + "bbox": [ + 228, + 223, + 263, + 322 + ], + "lines": [ + { + "bbox": [ + 228, + 223, + 263, + 322 + ], + "spans": [ + { + "bbox": [ + 228, + 223, + 263, + 322 + ], + "type": "image", + "image_path": "d1567b2df1aaabf7c88b08f6808222173c19e81a7842c75318e92e5a5153e0da.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 225, + 326, + 264, + 334 + ], + "lines": [ + { + "bbox": [ + 225, + 326, + 264, + 334 + ], + "spans": [ + { + "bbox": [ + 225, + 326, + 264, + 334 + ], + "type": "text", + "content": "Ours-Single" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_caption" + } + ], + "index": 27 + }, + { + "type": "image", + "bbox": [ + 278, + 222, + 312, + 322 + ], + "blocks": [ + { + "bbox": [ + 278, + 222, + 312, + 322 + ], + "lines": [ + { + "bbox": [ + 278, + 222, + 312, + 322 + ], + "spans": [ + { + "bbox": [ + 278, + 222, + 312, + 322 + ], + "type": "image", + "image_path": "9d14d446c812f3ba79dff90426f0c487f1622cead460c6f80a98f5546f25a5a9.jpg" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 272, + 326, + 317, + 334 + ], + "lines": [ + { + "bbox": [ + 272, + 326, + 317, + 334 + ], + "spans": [ + { + "bbox": [ + 272, + 326, + 317, + 334 + ], + "type": "text", + "content": "Ours-Refined" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_caption" + } + ], + "index": 29 + }, + { + "type": "image", + "bbox": [ + 340, + 221, + 392, + 323 + ], + "blocks": [ + { + "bbox": [ + 327, + 255, + 335, + 293 + ], + "lines": [ + { + "bbox": [ + 327, + 255, + 335, + 293 + ], + "spans": [ + { + "bbox": [ + 327, + 255, + 335, + 293 + ], + "type": "text", + "content": "Rrnnnne" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 340, + 221, + 392, + 323 + ], + "lines": [ + { + "bbox": [ + 340, + 221, + 392, + 323 + ], + "spans": [ + { + "bbox": [ + 340, + 221, + 392, + 323 + ], + "type": "image", + "image_path": "868694d3382e8384e9bc0acb6559028c3eeebd996d190f073c96360431b5b60f.jpg" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 345, + 326, + 386, + 334 + ], + "lines": [ + { + "bbox": [ + 345, + 326, + 386, + 334 + ], + "spans": [ + { + "bbox": [ + 345, + 326, + 386, + 334 + ], + "type": "text", + "content": "Hand Image" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 394, + 326, + 438, + 334 + ], + "lines": [ + { + "bbox": [ + 394, + 326, + 438, + 334 + ], + "spans": [ + { + "bbox": [ + 394, + 326, + 438, + 334 + ], + "type": "text", + "content": "Hand4Whole" + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "image_caption" + } + ], + "index": 32 + }, + { + "type": "image", + "bbox": [ + 397, + 223, + 438, + 322 + ], + "blocks": [ + { + "bbox": [ + 397, + 223, + 438, + 322 + ], + "lines": [ + { + "bbox": [ + 397, + 223, + 438, + 322 + ], + "spans": [ + { + "bbox": [ + 397, + 223, + 438, + 322 + ], + "type": "image", + "image_path": "0aab8a411ce9f29fe58d8547125f2a1b5b383cff52ba3a7078c372552d76111f.jpg" + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_body" + } + ], + "index": 34 + }, + { + "type": "image", + "bbox": [ + 443, + 222, + 488, + 322 + ], + "blocks": [ + { + "bbox": [ + 443, + 222, + 488, + 322 + ], + "lines": [ + { + "bbox": [ + 443, + 222, + 488, + 322 + ], + "spans": [ + { + "bbox": [ + 443, + 222, + 488, + 322 + ], + "type": "image", + "image_path": "973eb23b413d2231b829f02d5a8402bd11582fa3bae0b981c051d6b8702eff50.jpg" + } + ] + } + ], + "index": 36, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 446, + 326, + 485, + 334 + ], + "lines": [ + { + "bbox": [ + 446, + 326, + 485, + 334 + ], + "spans": [ + { + "bbox": [ + 446, + 326, + 485, + 334 + ], + "type": "text", + "content": "Ours-Single" + } + ] + } + ], + "index": 37, + "angle": 0, + "type": "image_caption" + } + ], + "index": 36 + }, + { + "type": "image", + "bbox": [ + 495, + 222, + 539, + 323 + ], + "blocks": [ + { + "bbox": [ + 495, + 222, + 539, + 323 + ], + "lines": [ + { + "bbox": [ + 495, + 222, + 539, + 323 + ], + "spans": [ + { + "bbox": [ + 495, + 222, + 539, + 323 + ], + "type": "image", + "image_path": "ebd9da3b584bfa078c7c9111266a0314999a58041405dabbcfef33dd46c35036.jpg" + } + ] + } + ], + "index": 38, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 495, + 326, + 539, + 334 + ], + "lines": [ + { + "bbox": [ + 495, + 326, + 539, + 334 + ], + "spans": [ + { + "bbox": [ + 495, + 326, + 539, + 334 + ], + "type": "text", + "content": "Ours-Refined" + } + ] + } + ], + "index": 39, + "angle": 0, + "type": "image_caption" + } + ], + "index": 38 + }, + { + "bbox": [ + 46, + 388, + 287, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 388, + 287, + 460 + ], + "spans": [ + { + "bbox": [ + 46, + 388, + 287, + 460 + ], + "type": "text", + "content": "sults are labeled as \"Ours-Single\" and our refinement results are labeled as \"Ours-Refined\". Our single-frame body pose estimation method outperforms all previous methods by a large margin. Our diffusion-based motion refinement method can further improve the accuracy of body poses estimated by the single-frame methods." + } + ] + } + ], + "index": 41 + }, + { + "bbox": [ + 46, + 460, + 287, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 460, + 287, + 556 + ], + "spans": [ + { + "bbox": [ + 46, + 460, + 287, + 556 + ], + "type": "text", + "content": "Note that previous methods [47, 50-52, 54] use training datasets different from each other. For a fair comparison, we re-train previous methods with our training datasets in Sec. 5.1 and show the results with “*” in Table 1. This retraining led to significant improvements across all previous methods, demonstrating our dataset's broad applicability. However, these methods still underperformed compared to ours, highlighting our approach's superiority." + } + ] + } + ], + "index": 42 + }, + { + "bbox": [ + 46, + 557, + 287, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 557, + 287, + 688 + ], + "spans": [ + { + "bbox": [ + 46, + 557, + 287, + 688 + ], + "type": "text", + "content": "To evaluate the accuracy of our hand pose estimation method, we first crop the hand images with the hand detection method in Sec. 3.1.3. Then we show the results of our single-frame hand pose estimation (labeled as \"Ours-Single\") and whole-body motion refinement methods (labeled as \"Ours-Refined\") in Table 2. Our single-frame hand pose estimation method outperforms the state-of-the-art method Hands4Whole [34], demonstrating the effectiveness of training the network on our EgoWholeBody dataset. Our whole-body motion refinement method can also enhance the accuracy of hand motion." + } + ] + } + ], + "index": 43 + }, + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "content": "For a qualitative comparison, we compare the body and hand poses of our method with existing methods on the" + } + ] + } + ], + "index": 44 + }, + { + "type": "table", + "bbox": [ + 309, + 386, + 544, + 536 + ], + "blocks": [ + { + "bbox": [ + 309, + 386, + 544, + 536 + ], + "lines": [ + { + "bbox": [ + 309, + 386, + 544, + 536 + ], + "spans": [ + { + "bbox": [ + 309, + 386, + 544, + 536 + ], + "type": "table", + "html": "
MethodMPJPEPA-MPJPE
SceneEgo test dataset [52]
EgoPW [51]189.6105.3
SceneEgo [52]118.592.75
EgoPW* [51]90.9664.33
SceneEgo* [52]89.0670.10
Ours-Single64.1950.06
Ours-Refined57.5946.55
EgoWholeBody test dataset
EgoPW* [51]84.2163.02
SceneEgo* [52]87.5769.46
Ours-Single66.2843.14
Ours-Refined60.3240.35
", + "image_path": "1064052aec9da9a3ec86e7da2a1eb808e52d1671886366acc035472711ae59a1.jpg" + } + ] + } + ], + "index": 45, + "angle": 0, + "type": "table_body" + } + ], + "index": 45 + }, + { + "bbox": [ + 305, + 544, + 545, + 588 + ], + "lines": [ + { + "bbox": [ + 305, + 544, + 545, + 588 + ], + "spans": [ + { + "bbox": [ + 305, + 544, + 545, + 588 + ], + "type": "text", + "content": "Table 1. Egocentric human body pose accuracy of our method on SceneEgo test datasets and EgoWholeBody test dataset. Our method outperforms all previous state-of-the-art methods. * denotes the method trained with the datasets in Sec. 5.1." + } + ] + } + ], + "index": 46, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 305, + 601, + 545, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 601, + 545, + 662 + ], + "spans": [ + { + "bbox": [ + 305, + 601, + 545, + 662 + ], + "type": "text", + "content": "SceneEgo dataset and the in-the-wild EgoPW [51] evaluation sequences. The results are shown in Fig. 4 and Fig. 5, showing that our method can predict high-quality whole-body poses from an egocentric camera. Please refer to our supplementary video for more qualitative evaluation results." + } + ] + } + ], + "index": 47 + }, + { + "bbox": [ + 306, + 670, + 399, + 683 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 670, + 399, + 683 + ], + "spans": [ + { + "bbox": [ + 306, + 670, + 399, + 683 + ], + "type": "text", + "content": "5.3. Ablation Study" + } + ] + } + ], + "index": 48 + }, + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "content": "EgoWholeBody Dataset. Compared to existing egocentric datasets, our EgoWholeBody dataset contains diverse body" + } + ] + } + ], + "index": 49 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "text", + "content": "783" + } + ] + } + ], + "index": 50 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 50, + 70, + 285, + 175 + ], + "blocks": [ + { + "bbox": [ + 50, + 70, + 285, + 175 + ], + "lines": [ + { + "bbox": [ + 50, + 70, + 285, + 175 + ], + "spans": [ + { + "bbox": [ + 50, + 70, + 285, + 175 + ], + "type": "table", + "html": "
MethodMPJPEPA-MPJPE
SceneEgo test dataset [52]
Hand4Whole [34]49.6613.85
Ours-Single23.639.59
Ours-Refined19.379.05
EgoWholeBody test dataset
Hand4Whole [34]52.8535.04
Ours-Single33.1019.68
Ours-Refined28.2914.51
", + "image_path": "fa1a92a804b3af566e94da576c291e5a2311a813afb72d94b28a4958b5b182ca.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 50, + 211, + 285, + 382 + ], + "blocks": [ + { + "bbox": [ + 47, + 182, + 286, + 205 + ], + "lines": [ + { + "bbox": [ + 47, + 182, + 286, + 205 + ], + "spans": [ + { + "bbox": [ + 47, + 182, + 286, + 205 + ], + "type": "text", + "content": "Table 2. Egocentric hand pose accuracy of our method. Our method outperforms the Hand4Whole [34] on both datasets." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 50, + 211, + 285, + 382 + ], + "lines": [ + { + "bbox": [ + 50, + 211, + 285, + 382 + ], + "spans": [ + { + "bbox": [ + 50, + 211, + 285, + 382 + ], + "type": "table", + "html": "
MethodMPJPEPA-MPJPE
Body Pose Results
w/o EgoWholeBody75.1058.62
w/o FisheyeViT67.3653.44
w/ Mo2Cap2 [54] head87.4765.10
w/ xR-egopose [47] head116.595.78
w/ SceneEgo [52] head77.7362.69
Ours-Single64.1950.06
w/ GlobalEgoMocap†69.8356.73
w/o uncert. guidance†62.1648.40
Only body diffusion58.9547.03
Ours-Refined†57.5946.55
Hand Pose Results
Only hand diffusion21.699.24
Ours-Refined19.379.05
", + "image_path": "6ee6f25a3c413d7716b537137e4bf74163cfd84641f9784c1945624c873bc71f.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 391, + 286, + 414 + ], + "lines": [ + { + "bbox": [ + 47, + 391, + 286, + 414 + ], + "spans": [ + { + "bbox": [ + 47, + 391, + 286, + 414 + ], + "type": "text", + "content": "Table 3. Ablation Study on SceneEgo test dataset [52]. " + }, + { + "bbox": [ + 47, + 391, + 286, + 414 + ], + "type": "inline_equation", + "content": "{}^{ \\dagger }" + }, + { + "bbox": [ + 47, + 391, + 286, + 414 + ], + "type": "text", + "content": " denotes the temporal-based method." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 426, + 286, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 426, + 286, + 568 + ], + "spans": [ + { + "bbox": [ + 46, + 426, + 286, + 568 + ], + "type": "text", + "content": "and hand motions, larger quantity of images, and higher image quality. We show this by training our body pose estimation network without our dataset, using the " + }, + { + "bbox": [ + 46, + 426, + 286, + 568 + ], + "type": "inline_equation", + "content": "\\mathrm{Mo}^2\\mathrm{Cap}^2" + }, + { + "bbox": [ + 46, + 426, + 286, + 568 + ], + "type": "text", + "content": " [54] and EgoPW [51] training dataset. The results, labeled as \"w/o EgoWholeBody\" in Table 3, show that performance without the EgoWholeBody dataset is inferior to our proposed method. This highlights that training with our EgoW-holeBody dataset enhances the performance of the pose estimation method. We also compare this result with existing methods on the SceneEgo test set (Table 1). Trained without EgoWholeBody, our approach still outperforms previous methods, showing the effectiveness of our method." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 570, + 286, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 570, + 286, + 688 + ], + "spans": [ + { + "bbox": [ + 46, + 570, + 286, + 688 + ], + "type": "text", + "content": "FisheyeViT and Pose Regressor with Pixel-Aligned 3D Heatmap. To assess the individual contributions of FisheyeViT and the pixel-aligned 3D heatmap in our single-frame pose estimation pipeline, we perform experiments to measure their impact on the overall performance. First, we substitute the FisheyeViT module in our single-frame pose estimation method to ViT [13]. The result is shown in \"w/o FisheyeViT\" in Table 3 and it is worse than our full method. This demonstrates the effectiveness of FisheyeViT in addressing fisheye distortion and feature extraction." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 689, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 286, + 713 + ], + "type": "text", + "content": "Next, we analyze the performance of the single-frame pose estimation network when substituting our pose regress-" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "text", + "content": "sor based on pixel-aligned 3D heatmap with the pose estimation heads of previous works [47, 52, 54]. The results of the three experiments, labeled as \"w/ Mo²Cap² head\", \"w/ xR-egopose head\" and \"w/ SceneEgo head\", show a performance drop compared to our full method. This emphasizes the crucial role of the pixel-aligned 3D heatmap in accurately estimating egocentric 3D body joint positions." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 156, + 545, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 156, + 545, + 251 + ], + "spans": [ + { + "bbox": [ + 304, + 156, + 545, + 251 + ], + "type": "text", + "content": "Diffusion-based Motion Refinement. We assess the effectiveness of our diffusion-based motion refinement with the following experiments: First, we compare the performance of our diffusion-based motion refinement with GlobalEgoMocap [50] by applying the GlobalEgoMocap optimizer on the single-frame body pose estimation results. The result, labeled as \"w GlobalEgoMocap\" in Table 3, indicates that our refinement method outperforms GlobalEgoMocap." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 251, + 545, + 358 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 251, + 545, + 358 + ], + "spans": [ + { + "bbox": [ + 304, + 251, + 545, + 358 + ], + "type": "text", + "content": "Second, we remove the uncertainty-aware guidance in the motion refinement. Instead, we use fixed Gaussian denoising steps to refine the motion. The result \"w/o uncert. guidance\" in Table 3, shows that our uncertainty-aware refinement method performs better. Our approach relies on the uncertainty values for each joint, using low-uncertainty joints to guide the generation of high-uncertainty joints. This helps reduce errors in joint predictions caused by egocentric self-occlusion, leading to improved results." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 360, + 545, + 467 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 360, + 545, + 467 + ], + "spans": [ + { + "bbox": [ + 304, + 360, + 545, + 467 + ], + "type": "text", + "content": "Third, we replace our whole-body motion diffusion model with the separate human body and left/right-hand diffusion models and show the accuracy of refined body and hand motion in \"Only body diffusion\" and \"Only hand diffusion\" in Table 3. From the results, we observe improvements in the accuracy of motion refined by our whole-body diffusion method, proving that learning the whole-body motion prior can help both the refinement of the body and hand motion by learning the correlation between them." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 478, + 378, + 491 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 478, + 378, + 491 + ], + "spans": [ + { + "bbox": [ + 306, + 478, + 378, + 491 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 498, + 545, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 498, + 545, + 666 + ], + "spans": [ + { + "bbox": [ + 304, + 498, + 545, + 666 + ], + "type": "text", + "content": "In this work, we have introduced an innovative approach to capture egocentric whole-body human motion. Our method comprises a single-frame-based whole-body pose estimation process, which includes FisheyeViT and pixel-aligned 3D heatmap representations. To enhance the initial whole-body pose estimates, we have integrated an uncertainty-aware diffusion-based motion refinement technique. Our experimental results demonstrate that both our single-frame method and the temporal-based method surpass all existing state-of-the-art techniques in terms of both quality and accuracy. Looking ahead, we see the potential for extending the applications of FisheyeViT to other vision tasks involving fisheye cameras. Future work could also involve incorporating facial expressions in whole-body motion capture." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 666, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 666, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 666, + 545, + 713 + ], + "type": "text", + "content": "Acknowledgments This project was supported by the Saarbrücken Research Center for Visual Computing, Interaction and AI. Christian Theobalt was supported by ERC Consolidator Grant 4DReply (770784)." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 748, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 748, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 748, + 312, + 757 + ], + "type": "text", + "content": "784" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 53, + 91, + 200, + 102 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 200, + 102 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 200, + 102 + ], + "type": "text", + "content": "[1] Blender. https://www.blender.org. 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 103, + 209, + 113 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 103, + 209, + 113 + ], + "spans": [ + { + "bbox": [ + 53, + 103, + 209, + 113 + ], + "type": "text", + "content": "[2] Mixamo. https://www MIXamo.com. 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 114, + 234, + 125 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 114, + 234, + 125 + ], + "spans": [ + { + "bbox": [ + 53, + 114, + 234, + 125 + ], + "type": "text", + "content": "[3] Renderpeople. https://renderpeople.com. 6, 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 126, + 287, + 180 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 126, + 287, + 180 + ], + "spans": [ + { + "bbox": [ + 53, + 126, + 287, + 180 + ], + "type": "text", + "content": "[4] Hiroyasu Akada, Jian Wang, Soshi Shimada, Masaki Takahashi, Christian Theobalt, and Vladislav Golyanik. Unrealego: A new dataset for robust egocentric 3d human motion capture. In European Conference on Computer Vision, pages 1-17. Springer, 2022. 1, 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 182, + 287, + 224 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 182, + 287, + 224 + ], + "spans": [ + { + "bbox": [ + 53, + 182, + 287, + 224 + ], + "type": "text", + "content": "[5] Hiroyasu Akada, Jian Wang, Vladislav Golyanik, and Christian Theobalt. 3d human pose perception from egocentric stereo videos. In 37th IEEE/CVF Conference on Computer Vision and Pattern Recognition. IEEE, 2024. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 226, + 287, + 280 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 226, + 287, + 280 + ], + "spans": [ + { + "bbox": [ + 53, + 226, + 287, + 280 + ], + "type": "text", + "content": "[6] Zhongang Cai, Wanqi Yin, Ailing Zeng, Chen Wei, Qingping Sun, Yanjun Wang, Hui En Pang, Haiyi Mei, Mingyuan Zhang, Lei Zhang, et al. Smpler-x: Scaling up expressive human pose and shape estimation. arXiv preprint arXiv:2309.17448, 2023. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 281, + 287, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 281, + 287, + 346 + ], + "spans": [ + { + "bbox": [ + 53, + 281, + 287, + 346 + ], + "type": "text", + "content": "[7] Young-Woon Cha, True Price, Zhen Wei, Xinran Lu, Nicholas Rewkowski, Rohan Chabra, Zihe Qin, Hyounghun Kim, Zhaoqi Su, Yebin Liu, et al. Towards fully mobile 3d face, body, and environment capture using only head-worn cameras. IEEE transactions on visualization and computer graphics, 24(11):2993-3004, 2018. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 348, + 287, + 390 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 348, + 287, + 390 + ], + "spans": [ + { + "bbox": [ + 53, + 348, + 287, + 390 + ], + "type": "text", + "content": "[8] Jeongjun Choi, Dongseok Shim, and H Jin Kim. Diffupose: Monocular 3d human pose estimation via denoising diffusion probabilistic model. arXiv preprint arXiv:2212.02796, 2022. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 392, + 287, + 458 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 392, + 287, + 458 + ], + "spans": [ + { + "bbox": [ + 53, + 392, + 287, + 458 + ], + "type": "text", + "content": "[9] Vasileios Choutas, Georgios Pavlakos, Timo Bolkart, Dimitrios Tzionas, and Michael J Black. Monocular expressive body regression through body-driven attention. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part X 16, pages 20-40. Springer, 2020. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 459, + 287, + 513 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 459, + 287, + 513 + ], + "spans": [ + { + "bbox": [ + 48, + 459, + 287, + 513 + ], + "type": "text", + "content": "[10] Hai Ci, Mingdong Wu, Wentao Zhu, Xiaoxuan Ma, Hao Dong, Fangwei Zhong, and Yizhou Wang. Gfpose: Learning 3d human pose prior with gradient fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4800-4810, 2023. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 514, + 287, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 514, + 287, + 568 + ], + "spans": [ + { + "bbox": [ + 48, + 514, + 287, + 568 + ], + "type": "text", + "content": "[11] Benjamin Coors, Alexandru Paul Condurache, and Andreas Geiger. Spherenet: Learning spherical representations for detection and classification in omnidirectional images. In Proceedings of the European conference on computer vision (ECCV), pages 518-533, 2018. 5" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 570, + 287, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 570, + 287, + 602 + ], + "spans": [ + { + "bbox": [ + 48, + 570, + 287, + 602 + ], + "type": "text", + "content": "[12] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in neural information processing systems, 34:8780-8794, 2021. 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 603, + 287, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 603, + 287, + 668 + ], + "spans": [ + { + "bbox": [ + 48, + 603, + 287, + 668 + ], + "type": "text", + "content": "[13] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020.4, 8" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 670, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 670, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 670, + 287, + 713 + ], + "type": "text", + "content": "[14] Yuming Du, Robin Kips, Albert Pumarola, Sebastian Starke, Ali Thabet, and Artsiom Sanakoyeu. Avatars grow legs: Generating smooth human motion from sparse tracking inputs with diffusion model. In Proceedings of the IEEE/CVF" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 74, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 31, + "blocks": [ + { + "bbox": [ + 326, + 74, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 74, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 326, + 74, + 545, + 95 + ], + "type": "text", + "content": "Conference on Computer Vision and Pattern Recognition, pages 481-490, 2023. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 96, + 545, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 96, + 545, + 149 + ], + "spans": [ + { + "bbox": [ + 308, + 96, + 545, + 149 + ], + "type": "text", + "content": "[15] Yao Feng, Vasileios Choutas, Timo Bolkart, Dimitrios Tzionas, and Michael J Black. Collaborative regression of expressive bodies using moderation. In 2021 International Conference on 3D Vision (3DV), pages 792-804. IEEE, 2021. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 152, + 545, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 152, + 545, + 195 + ], + "spans": [ + { + "bbox": [ + 307, + 152, + 545, + 195 + ], + "type": "text", + "content": "[16] Lin Geng Foo, Jia Gong, Hossein Rahmani, and Jun Liu. Distribution-aligned diffusion for human mesh recovery. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9221-9232, 2023. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 198, + 545, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 198, + 545, + 251 + ], + "spans": [ + { + "bbox": [ + 307, + 198, + 545, + 251 + ], + "type": "text", + "content": "[17] Jia Gong, Lin Geng Foo, Zhipeng Fan, Qiuhong Ke, Hossein Rahmani, and Jun Liu. Diffpose: Toward more reliable 3d pose estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13041-13051, 2023. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 253, + 545, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 253, + 545, + 285 + ], + "spans": [ + { + "bbox": [ + 308, + 253, + 545, + 285 + ], + "type": "text", + "content": "[18] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 2, 3, 5, 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 287, + 545, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 287, + 545, + 331 + ], + "spans": [ + { + "bbox": [ + 308, + 287, + 545, + 331 + ], + "type": "text", + "content": "[19] Karl Holmquist and Bastian Wandt. Diffpose: Multi-hypothesis human pose estimation using diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15977-15987, 2023. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 332, + 545, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 332, + 545, + 387 + ], + "spans": [ + { + "bbox": [ + 308, + 332, + 545, + 387 + ], + "type": "text", + "content": "[20] Ludovic Hoyet, Kenneth Ryall, Rachel McDonnell, and Carol O'Sullivan. Sleight of hand: perception of finger motion from reduced marker sets. In Proceedings of the ACM SIGGRAPH symposium on interactive 3D graphics and games, pages 79-86, 2012. 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 388, + 545, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 388, + 545, + 431 + ], + "spans": [ + { + "bbox": [ + 308, + 388, + 545, + 431 + ], + "type": "text", + "content": "[21] Hao Jiang and Kristen Grauman. Seeing invisible poses: Estimating 3d body pose from egocentric video. In 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 3501-3509. IEEE, 2017. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 433, + 545, + 476 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 433, + 545, + 476 + ], + "spans": [ + { + "bbox": [ + 308, + 433, + 545, + 476 + ], + "type": "text", + "content": "[22] Zhongyu Jiang, Zhuoran Zhou, Lei Li, Wenhao Chai, ChengYen Yang, and Jenq-Neng Hwang. Back to optimization: Diffusion-based zero-shot 3d human pose estimation. arXiv preprint arXiv:2307.03833, 2023. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 478, + 545, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 478, + 545, + 510 + ], + "spans": [ + { + "bbox": [ + 308, + 478, + 545, + 510 + ], + "type": "text", + "content": "[23] Taeho Kang, Kyungjin Lee, Jinrui Zhang, and Youngki Lee. Ego3dpose: Capturing 3d cues from binocular egocentric views. arXiv preprint arXiv:2309.11962, 2023. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 512, + 545, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 512, + 545, + 533 + ], + "spans": [ + { + "bbox": [ + 308, + 512, + 545, + 533 + ], + "type": "text", + "content": "[24] David G Kendall. A survey of the statistical theory of shape. Statistical Science, 4(2):87-99, 1989. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 535, + 545, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 535, + 545, + 578 + ], + "spans": [ + { + "bbox": [ + 308, + 535, + 545, + 578 + ], + "type": "text", + "content": "[25] Jiaman Li, Karen Liu, and Jiajun Wu. Ego-body pose estimation via ego-head pose estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17142-17151, 2023. 2, 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 580, + 545, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 580, + 545, + 634 + ], + "spans": [ + { + "bbox": [ + 308, + 580, + 545, + 634 + ], + "type": "text", + "content": "[26] Yuyan Li, Yuliang Guo, Zhixin Yan, Xinyu Huang, Ye Duan, and Liu Ren. Omnifusion: 360 monocular depth estimation via geometry-aware fusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2801-2810, 2022. 5" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 635, + 545, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 635, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 308, + 635, + 545, + 689 + ], + "type": "text", + "content": "[27] Jing Lin, Ailing Zeng, Haoqian Wang, Lei Zhang, and Yu Li. One-stage 3d whole-body mesh recovery with component aware transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21159-21168, 2023. 3" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 308, + 691, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 691, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 691, + 545, + 713 + ], + "type": "text", + "content": "[28] Yuxuan Liu, Jianxin Yang, Xiao Gu, Yao Guo, and Guang-Zhong Yang. Ego+ x: An egocentric vision system for global" + } + ] + } + ], + "index": 30 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "text", + "content": "785" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 67, + 72, + 287, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 72, + 287, + 116 + ], + "spans": [ + { + "bbox": [ + 67, + 72, + 287, + 116 + ], + "type": "text", + "content": "3d human pose estimation and social interaction characterization. In 2022 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 5271-5277. IEEE, 2022. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 118, + 287, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 118, + 287, + 162 + ], + "spans": [ + { + "bbox": [ + 48, + 118, + 287, + 162 + ], + "type": "text", + "content": "[29] Yuxuan Liu, Jianxin Yang, Xiao Gu, Yijun Chen, Yao Guo, and Guang-Zhong Yang. Egofish3d: Egocentric 3d pose estimation from a fisheye camera via self-supervised learning. IEEE Transactions on Multimedia, 2023. 1, 2, 4" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 164, + 287, + 218 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 164, + 287, + 218 + ], + "spans": [ + { + "bbox": [ + 49, + 164, + 287, + 218 + ], + "type": "text", + "content": "[30] Yuxuan Liu, Jianxin Yang, Xiao Gu, Yao Guo, and GuangZhong Yang. Egohmr: Egocentric human mesh recovery via hierarchical latent diffusion model. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 9807-9813. IEEE, 2023. 1, 2, 3, 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 220, + 287, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 220, + 287, + 262 + ], + "spans": [ + { + "bbox": [ + 49, + 220, + 287, + 262 + ], + "type": "text", + "content": "[31] Zhengyi Luo, Ryo Hachiuma, Ye Yuan, and Kris Kitani. Dynamics-regulated kinematic policy for egocentric pose estimation. Advances in Neural Information Processing Systems, 34:25019-25032, 2021. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 265, + 287, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 265, + 287, + 319 + ], + "spans": [ + { + "bbox": [ + 48, + 265, + 287, + 319 + ], + "type": "text", + "content": "[32] Christen Millerdurai, Hiroyasu Akada, Jian Wang, Diogo Luvizon, Christian Theobalt, and Vladislav Golyanik. Evento3d: 3d human motion capture from egocentric event streams. In 37th IEEE/CVF Conference on Computer Vision and Pattern Recognition. IEEE, 2024. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 321, + 287, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 321, + 287, + 376 + ], + "spans": [ + { + "bbox": [ + 48, + 321, + 287, + 376 + ], + "type": "text", + "content": "[33] Gyeongsik Moon, Ju Yong Chang, and Kyoung Mu Lee. V2v-posenet: Voxel-to-voxel prediction network for accurate 3d hand and human pose estimation from a single depth map. In Proceedings of the IEEE conference on computer vision and pattern Recognition, pages 5079–5088, 2018. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 378, + 287, + 421 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 378, + 287, + 421 + ], + "spans": [ + { + "bbox": [ + 48, + 378, + 287, + 421 + ], + "type": "text", + "content": "[34] Gyeongsik Moon, Hongsuk Choi, and Kyoung Mu Lee. Accurate 3d hand pose estimation for whole-body 3d human mesh estimation. In Computer Vision and Pattern Recognition Workshop (CVPRW), 2022. 4, 5, 7, 8, 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 423, + 287, + 477 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 423, + 287, + 477 + ], + "spans": [ + { + "bbox": [ + 48, + 423, + 287, + 477 + ], + "type": "text", + "content": "[35] Evonne Ng, Donglai Xiang, Hanbyul Joo, and Kristen Grauman. You2me: Inferring body pose in egocentric video via first and second person interactions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9890-9900, 2020. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 479, + 287, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 479, + 287, + 544 + ], + "spans": [ + { + "bbox": [ + 48, + 479, + 287, + 544 + ], + "type": "text", + "content": "[36] Jinman Park, Kimathi Kaai, Saad Hossain, Norikatsu Sumi, Sirisha Rambhatla, and Paul Fieguth. Domain-guided spatiotemporal self-attention for egocentric 3d pose estimation. In Proceedings of the 29th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, pages 1837–1849, 2023. 2, 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 546, + 287, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 546, + 287, + 611 + ], + "spans": [ + { + "bbox": [ + 48, + 546, + 287, + 611 + ], + "type": "text", + "content": "[37] Georgios Pavlakos, Vasileios Choutas, Nima Ghorbani, Timo Bolkart, Ahmed AA Osman, Dimitrios Tzionas, and Michael J Black. Expressive body capture: 3d hands, face, and body from a single image. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10975-10985, 2019. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 614, + 287, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 614, + 287, + 656 + ], + "spans": [ + { + "bbox": [ + 48, + 614, + 287, + 656 + ], + "type": "text", + "content": "[38] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 1 (2):3, 2022. 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 658, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 287, + 712 + ], + "type": "text", + "content": "[39] Helge Rhodin, Christian Richardt, Dan Casas, Eldar Insafutdinov, Mohammad Shafiei, Hans-Peter Seidel, Bernt Schiele, and Christian Theobalt. Egocap: egocentric marker-less motion capture with two fisheye cameras. ACM Transactions on Graphics (TOG), 35(6):1-11, 2016. 2" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 127 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 127 + ], + "type": "text", + "content": "[40] Yu Rong, Takaaki Shiratori, and Hanbyul Joo. Frankmocap: A monocular 3d whole-body pose estimation system via regression and integration. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1749-1759, 2021. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 307, + 129, + 545, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 129, + 545, + 183 + ], + "spans": [ + { + "bbox": [ + 307, + 129, + 545, + 183 + ], + "type": "text", + "content": "[41] Davide Scaramuzza, Agostino Martinelli, and Roland Siegwart. A toolbox for easily calibrating omnidirectional cameras. In 2006 IEEE/RSJ International Conference on Intelligent Robots and Systems, pages 5695-5701. IEEE, 2006. 1" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 308, + 184, + 545, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 184, + 545, + 237 + ], + "spans": [ + { + "bbox": [ + 308, + 184, + 545, + 237 + ], + "type": "text", + "content": "[42] Wenkang Shan, Zhenhua Liu, Xinfeng Zhang, Zhao Wang, Kai Han, Shanshe Wang, Siwei Ma, and Wen Gao. Diffusion-based 3d human pose estimation with multi-hypothesis aggregation. arXiv preprint arXiv:2303.11579, 2023.3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 239, + 545, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 239, + 545, + 282 + ], + "spans": [ + { + "bbox": [ + 308, + 239, + 545, + 282 + ], + "type": "text", + "content": "[43] Soshi Shimada, Vladislav Golyanik, Weipeng Xu, and Christian Theobalt. Physcap: Physically plausible monocular 3d motion capture in real time. ACM Transactions on Graphics (ToG), 39(6):1-16, 2020. 5" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 283, + 545, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 283, + 545, + 326 + ], + "spans": [ + { + "bbox": [ + 308, + 283, + 545, + 326 + ], + "type": "text", + "content": "[44] Xiao Sun, Bin Xiao, Fangyin Wei, Shuang Liang, and Yichen Wei. Integral human pose regression. In Proceedings of the European conference on computer vision (ECCV), pages 529-545, 2018. 4, 5" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 327, + 545, + 382 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 327, + 545, + 382 + ], + "spans": [ + { + "bbox": [ + 308, + 327, + 545, + 382 + ], + "type": "text", + "content": "[45] Yu Sun, Tianyu Huang, Qian Bao, Wu Liu, Wenpeng Gao, and Yili Fu. Learning monocular mesh recovery of multiple body parts via synthesis. In ICASSP 2022-2022 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 2669-2673. IEEE, 2022. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 383, + 545, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 383, + 545, + 437 + ], + "spans": [ + { + "bbox": [ + 308, + 383, + 545, + 437 + ], + "type": "text", + "content": "[46] Omid Taheri, Nima Ghorbani, Michael J Black, and Dimitrios Tzionas. Grab: A dataset of whole-body human grasping of objects. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part IV 16, pages 581-600. Springer, 2020. 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 438, + 545, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 438, + 545, + 480 + ], + "spans": [ + { + "bbox": [ + 308, + 438, + 545, + 480 + ], + "type": "text", + "content": "[47] Denis Tomè, Patrick Peluse, Lourdes Agapito, and Hernán Badino. xr-egopose: Egocentric 3d human pose from an HMD camera. In ICCV, pages 7727-7737, 2019. 1, 2, 5, 6, 7, 8, 4" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 482, + 545, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 482, + 545, + 525 + ], + "spans": [ + { + "bbox": [ + 308, + 482, + 545, + 525 + ], + "type": "text", + "content": "[48] Jonathan Tseng, Rodrigo Castellon, and Karen Liu. Edge: Editable dance generation from music. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 448-458, 2023. 5, 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 526, + 545, + 591 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 526, + 545, + 591 + ], + "spans": [ + { + "bbox": [ + 308, + 526, + 545, + 591 + ], + "type": "text", + "content": "[49] Jingdong Wang, Ke Sun, Tianheng Cheng, Borui Jiang, Chaorui Deng, Yang Zhao, Dong Liu, Yadong Mu, Mingkui Tan, Xinggang Wang, et al. Deep high-resolution representation learning for visual recognition. IEEE transactions on pattern analysis and machine intelligence, 43(10):3349-3364, 2020. 5" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 593, + 545, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 593, + 545, + 624 + ], + "spans": [ + { + "bbox": [ + 308, + 593, + 545, + 624 + ], + "type": "text", + "content": "[50] Jian Wang, Lingjie Liu, Weipeng Xu, Kripasindhu Sarkar, and Christian Theobalt. Estimating egocentric 3d human pose in global space. ICCV, 2021. 1, 2, 6, 7, 8, 3, 4, 5" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 625, + 545, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 625, + 545, + 678 + ], + "spans": [ + { + "bbox": [ + 308, + 625, + 545, + 678 + ], + "type": "text", + "content": "[51] Jian Wang, Lingjie Liu, Weipeng Xu, Kripasindhu Sarkar, Diogo Luvizon, and Christian Theobalt. Estimating egocentric 3d human pose in the wild with external weak supervision. In CVPR, pages 13157-13166, 2022. 1, 2, 6, 7, 8, 4" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 681, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 681, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 681, + 545, + 713 + ], + "type": "text", + "content": "[52] Jian Wang, Diogo Luvizon, Weipeng Xu, Lingjie Liu, Kripasindhu Sarkar, and Christian Theobalt. Scene-aware egocentric 3d human pose estimation. In Proceedings of the" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 298, + 749, + 313, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 749, + 313, + 757 + ], + "spans": [ + { + "bbox": [ + 298, + 749, + 313, + 757 + ], + "type": "text", + "content": "786" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 73, + 286, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 67, + 73, + 286, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 73, + 286, + 95 + ], + "spans": [ + { + "bbox": [ + 67, + 73, + 286, + 95 + ], + "type": "text", + "content": "IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13031-13040, 2023. 1, 2, 5, 6, 7, 8, 3, 4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 96, + 286, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 96, + 286, + 140 + ], + "spans": [ + { + "bbox": [ + 48, + 96, + 286, + 140 + ], + "type": "text", + "content": "[53] Donglai Xiang, Hanbyul Joo, and Yaser Sheikh. Monocular total capture: Posing face, body, and hands in the wild. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10965-10974, 2019. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 141, + 286, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 141, + 286, + 196 + ], + "spans": [ + { + "bbox": [ + 48, + 141, + 286, + 196 + ], + "type": "text", + "content": "[54] Weipeng Xu, Avishek Chatterjee, Michael Zollhöfer, Helge Rhodin, Pascal Fua, Hans-Peter Seidel, and Christian Theobalt. " + }, + { + "bbox": [ + 48, + 141, + 286, + 196 + ], + "type": "inline_equation", + "content": "\\mathrm{Mo}^2\\mathrm{cap}^2" + }, + { + "bbox": [ + 48, + 141, + 286, + 196 + ], + "type": "text", + "content": ": Real-time mobile 3d motion capture with a cap-mounted fisheye camera. IEEE Trans. Vis. Comput. Graph., 25(5):2093-2101, 2019. 1, 2, 5, 6, 7, 8, 3, 4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 198, + 286, + 241 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 198, + 286, + 241 + ], + "spans": [ + { + "bbox": [ + 48, + 198, + 286, + 241 + ], + "type": "text", + "content": "[55] Yufei Xu, Jing Zhang, Qiming Zhang, and Dacheng Tao. Vitpose: Simple vision transformer baselines for human pose estimation. Advances in Neural Information Processing Systems, 35:38571-38584, 2022. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 243, + 286, + 286 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 243, + 286, + 286 + ], + "spans": [ + { + "bbox": [ + 48, + 243, + 286, + 286 + ], + "type": "text", + "content": "[56] Dianyi Yang, Jiadong Tang, Yu Gao, Yi Yang, and Mengyin Fu. Sector patch embedding: An embedding module conforming to the distortion pattern of fisheye image. arXiv preprint arXiv:2303.14645, 2023. 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 289, + 286, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 289, + 286, + 331 + ], + "spans": [ + { + "bbox": [ + 48, + 289, + 286, + 331 + ], + "type": "text", + "content": "[57] Shangrong Yang, Chunyu Lin, Kang Liao, and Yao Zhao. Dual diffusion architecture for fisheye image rectification: Synthetic-to-real generalization. arXiv preprint arXiv:2301.11785, 2023." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 333, + 286, + 388 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 333, + 286, + 388 + ], + "spans": [ + { + "bbox": [ + 48, + 333, + 286, + 388 + ], + "type": "text", + "content": "[58] Fanghua Yu, Xintao Wang, Mingdeng Cao, Gen Li, Ying Shan, and Chao Dong. Osrt: Omnidirectional image superresolution with distortion-aware transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13283-13292, 2023. 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 389, + 286, + 422 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 389, + 286, + 422 + ], + "spans": [ + { + "bbox": [ + 48, + 389, + 286, + 422 + ], + "type": "text", + "content": "[59] Ye Yuan and Kris Kitani. 3d ego-pose estimation via imitation learning. In Proceedings of the European Conference on Computer Vision (ECCV), pages 735–750, 2018. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 423, + 286, + 465 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 423, + 286, + 465 + ], + "spans": [ + { + "bbox": [ + 48, + 423, + 286, + 465 + ], + "type": "text", + "content": "[60] Ye Yuan and Kris Kitani. Ego-pose estimation and forecasting as real-time pd control. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10082-10092, 2019. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 468, + 286, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 468, + 286, + 511 + ], + "spans": [ + { + "bbox": [ + 48, + 468, + 286, + 511 + ], + "type": "text", + "content": "[61] Ye Yuan, Jiaming Song, Umar Iqbal, Arash Vahdat, and Jan Kautz. Physdiff: Physics-guided human motion diffusion model. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 16010-16021, 2023. 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 513, + 286, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 513, + 286, + 567 + ], + "spans": [ + { + "bbox": [ + 48, + 513, + 286, + 567 + ], + "type": "text", + "content": "[62] Siwei Zhang, Qianli Ma, Yan Zhang, Zhiyin Qian, Taein Kwon, Marc Pollefeys, Federica Bogo, and Siyu Tang. Ego-body: Human body shape and motion of interacting people from head-mounted devices. In European Conference on Computer Vision, pages 180–200. Springer, 2022. 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 569, + 286, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 569, + 286, + 611 + ], + "spans": [ + { + "bbox": [ + 48, + 569, + 286, + 611 + ], + "type": "text", + "content": "[63] Siwei Zhang, Qianli Ma, Yan Zhang, Sadegh Aliakbarian, Darren Cosker, and Siyu Tang. Probabilistic human mesh recovery in 3d scenes from egocentric views. arXiv preprint arXiv:2304.06024, 2023. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 613, + 286, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 613, + 286, + 667 + ], + "spans": [ + { + "bbox": [ + 48, + 613, + 286, + 667 + ], + "type": "text", + "content": "[64] Yahui Zhang, Shaodi You, and Theo Gevers. Automatic calibration of the fisheye camera for egocentric 3d human pose estimation from a single image. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1772-1781, 2021. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 670, + 286, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 670, + 286, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 670, + 286, + 712 + ], + "type": "text", + "content": "[65] Dongxu Zhao, Zhen Wei, Jisan Mahmud, and Jan-Michael Frahm. Egoglass: Egocentric-view human pose estimation from an eyeglass frame. In 2021 International Conference on 3D Vision (3DV), pages 32-41. IEEE, 2021. 2" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 183 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 128 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 128 + ], + "type": "text", + "content": "[66] Yuxiao Zhou, Marc Habermann, Ikhsanul Habibie, Ayush Tewari, Christian Theobalt, and Feng Xu. Monocular real-time full body capture with inter-part correlations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4811-4822, 2021. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 129, + 545, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 129, + 545, + 183 + ], + "spans": [ + { + "bbox": [ + 307, + 129, + 545, + 183 + ], + "type": "text", + "content": "[67] Christian Zimmermann, Duygu Ceylan, Jimei Yang, Bryan Russell, Max Argus, and Thomas Brox. Freihand: A dataset for markerless capture of hand pose and shape from single rgb images. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 813-822, 2019. 2" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "text", + "content": "787" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/ElasticDiffusion_ Training-free Arbitrary Size Image Generation through Global-Local Content Separation/de40da54-b1b7-40fa-9aec-0f4f1bc4f406_content_list.json b/2024/ElasticDiffusion_ Training-free Arbitrary Size Image Generation through Global-Local Content Separation/de40da54-b1b7-40fa-9aec-0f4f1bc4f406_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..f998dc718724d101fd71a00138394b0d834a611c --- /dev/null +++ b/2024/ElasticDiffusion_ Training-free Arbitrary Size Image Generation through Global-Local Content Separation/de40da54-b1b7-40fa-9aec-0f4f1bc4f406_content_list.json @@ -0,0 +1,1415 @@ +[ + { + "type": "text", + "text": "ElasticDiffusion: Training-free Arbitrary Size Image Generation through Global-Local Content Separation", + "text_level": 1, + "bbox": [ + 158, + 130, + 810, + 175 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Moayed Haji-Ali", + "bbox": [ + 155, + 203, + 292, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Guha Balakrishnan Rice University", + "bbox": [ + 408, + 203, + 562, + 238 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{mh155, guha, vicenteor}@rice.edu", + "bbox": [ + 341, + 241, + 630, + 255 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Vicente Ordonez", + "bbox": [ + 678, + 203, + 813, + 219 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/6332836812d5c3aa5ab1d905bbb973969a8793b5267a835e2d10f1c30630b170.jpg", + "image_caption": [ + "Figure 1. ElasticDiffusion generates high quality images at arbitrary sizes using a pretrained diffusion model trained on a single image size, with equivalent memory footprint and no further training. These results are based on Stable Diffusion1,4, which was trained to generate $512 \\times 512$ images. The examples shown in this collage are presented without any image cropping, stretching, or post-processing." + ], + "image_footnote": [], + "bbox": [ + 78, + 271, + 890, + 616 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 679, + 313, + 695 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Diffusion models have revolutionized image generation in recent years, yet they are still limited to a few sizes and aspect ratios. We propose ElasticDiffusion, a novel training-free decoding method that enables pretrained text-to-image diffusion models to generate images with various sizes. ElasticDiffusion attempts to decouple the generation trajectory of a pretrained model into local and global signals. The local signal controls low-level pixel information and can be estimated on local patches, while the global signal is used to maintain overall structural consistency and is estimated with a reference image. We test our method on CelebA-HQ (faces) and LAION-COCO (objects/indoor/outdoor scenes). Our experiments and qualita", + "bbox": [ + 73, + 704, + 470, + 901 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "tive results show superior image coherence quality across aspect ratios compared to MultiDiffusion and the standard decoding strategy of Stable Diffusion. Project Webpage: https://elasticdiffusion.github.io/", + "bbox": [ + 498, + 681, + 890, + 742 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 500, + 751, + 630, + 766 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Diffusion models are a powerful family of algorithms that achieve remarkable quality and obtain the current state-of-the-art performance on various image synthesis tasks. As is the case with most deep neural networks, diffusion models are typically trained on one or a few image resolutions. For instance, Stable Diffusion (SD) [39], one of the most widely adopted diffusion models, is trained on a square images of size $512 \\times 512$ , yet fails to maintain its performance at dif-", + "bbox": [ + 496, + 779, + 892, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "6603", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ferent aspect ratios during inference time. In practice, many applications require a wide aspect ratio or portrait mode, such as digital billboards, wearable devices, automotive displays, and any application relying on a computer screen.", + "bbox": [ + 75, + 90, + 468, + 151 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recent studies address the issue of variable image size in different ways. SDXL [35] and Any-Size-Diffusion [62] explicitly finetune models on images with a range of aspect ratios, which requires extensive computation, a quadratic memory footprint, and larger training data. In addition, this strategy requires a set of resolutions to be specified upfront during training time, while the models still struggle to generalize to new resolutions during inference, often resulting in artifacts. Recent works also show remarkable results in generating panoramic images using pretrained diffusion models by overlapping generated patches into a larger image [3, 63]. These methods work well for landscape images with repetitive patterns. However, their lack of global guidance limits their abilities to generate images of single objects or faces where global structure is important. Recent work [24, 50, 51] aimed at extending pre-trained diffusion models capabilities to others domains, often in a training-free way [4, 7, 13, 28, 30, 34, 47, 54, 61]. Few concurrent work [9, 10, 12, 31, 60] focus on adapting them to higher resolutions, yet they are constrained to square images.", + "bbox": [ + 75, + 154, + 472, + 457 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this work, we propose ElasticDiffusion, a novel decoding strategy that takes a pretrained diffusion model and generate images at arbitrary sizes during inference using a constant memory footprint. To achieve this, we revisit the guidance mechanism of conditional diffusion models to decouple global and local content generation. Global content controls the high-level aspects of the image, whereas local content adds finer, more granular details. This separation facilitates generating the local content in patches for images of varying sizes, all while being guided with global content that we derive from a reference image at the diffusion model pretraining resolution. This enables the synthesis of images at diverse resolutions and aspect ratios while adhering the diffusion forward calls to the model's initial training size. To aid in this task, we introduce several techniques including an efficient patch fusion method for smooth boundaries, a novel guidance strategy to reduce image artifacts, and a global content resampling technique that amplifies the resolution of diffusion models up to 4X the training size.", + "bbox": [ + 75, + 459, + 472, + 747 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Figure 1 shows a diverse array of images generated with ElasticDiffusion. Several of theses images include a single object or were generated with extreme aspect ratios, showcasing our method's ability to produce coherent images under various sizes. Quantitatively, ElasticDiffusion outperforms baselines across most aspect ratios. More importantly, despite relying on Stable Diffusion1,4, ElasticDiffusion obtains comparable FID (228.87 vs 230.21) and CLIP scores (26.07 vs 28.06) as SDXL when generating images at $1024 \\times 1024$ , which is the native resolution of SDXL.", + "bbox": [ + 75, + 750, + 472, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 498, + 89, + 642, + 106 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Diffusion Models have been widely adopted for their high-quality outputs in generative tasks [8, 11, 15, 20, 22, 35, 37, 39, 43, 55, 58]. These models involve iterative decoding with many steps leading to high compute and memory requirements. Recent work addressed these issues by devising faster sampling strategies [26, 48], hierarchical models [19, 37, 43], progressive training at increasing resolutions [33, 35, 39], and two-stage models [35, 39, 42, 53]. Stable Diffusion (SD) [39] trains a variational auto-encoder to compress images into a low-dimensional $64 \\times 64$ latent space and trains a diffusion model on this latent space. To train for higher resolutions, models are initially trained at a $256 \\times 256$ resolution, before fine-tuning them at $512 \\times 512$ and $1024 \\times 1024$ in the case of SDXL [35]. SD is one of the few large-scale diffusion models that released trained parameters, making it the building block for many subsequent work [5, 16, 40, 44, 52, 53]. However, these models, including SD, are confined to specific resolutions and do not generalize well to aspect ratios unseen during training. Interestingly, despite being presented with a $256 \\times 256$ resolution during their early training stages, both SD and SDXL fail to generate realistic images at this resolution after being fine-tuned for larger outputs. ElasticDiffusion enables high quality generation at unseen resolutions including re-enabling consistent high quality outputs for SD at $256 \\times 256$ .", + "bbox": [ + 496, + 114, + 893, + 492 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Mixture of diffusers [63] and MultiDiffusion [3] generate panoramic images using a pre-trained diffusion model by generating overlapping crops and combining the generation signal of the overlapping regions. Blending multiple generation signals spatially has been of interest in many previous work [1, 63]. SDXL [35] and Any-Size-Diffusion [62] fine-tune a pre-trained SD on a fixed set of resolutions. A concurrent work [14] uses dilated convolution kernels to reduce SD artifacts when generating non-square images. ElasticDiffusion extends a pre-trained SD to generate images of various sizes at a constant memory requirement and without additional training, all while ensuring global coherence.", + "bbox": [ + 496, + 493, + 893, + 674 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Guided diffusion models devise strategies to condition image generation based on text and other modalities [2, 2, 6, 18, 23, 25, 32, 39, 57, 59]. Classifier guidance [32] uses a pre-trained classifier on noisy images to guide the generation process. Classifier-free guidance [18] eliminates the need for a pretrained classifier but requires training a conditional diffusion model, limiting its applications. Universal Guidance [2] applies a pre-trained classifier on the noise-free images produced by DDIM [48], bypassing training on noisy images. StableSR [53] uses LoRA [21] to condition the generation on low-resolution inputs to achieve superb image super-resolution. Inspired by this, we propose Reduce-Resolution Guidance (Sec. 4) to constrain the generation of high-res images using a lower resolution version, substantially reducing artifacts without extra training.", + "bbox": [ + 496, + 674, + 895, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "6604", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3. Background: Diffusion Models", + "text_level": 1, + "bbox": [ + 76, + 89, + 361, + 107 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "A conditional diffusion model $\\epsilon_{\\theta} \\colon \\mathcal{X} \\times \\mathcal{C} \\to \\mathcal{X}$ predicts a less noisy version of the input image $x \\in \\mathbb{R}^{H \\times W \\times 3}$ , conditioned on variable $c \\in \\mathbb{R}^D$ (e.g. a text embedding). Starting with $x_{T} \\sim \\mathcal{N}(0, \\mathbf{I})$ , the reverse diffusion process progressively denoise $x_{T}$ over $T$ steps to generate a realistic image $x_0$ that conforms to the input condition $c$ through:", + "bbox": [ + 76, + 114, + 468, + 205 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nx _ {t - 1} = \\epsilon_ {\\theta} (x _ {t}, c) \\quad \\text {f o r} t = T, T - 1, \\dots , 1,\n$$\n", + "text_format": "latex", + "bbox": [ + 119, + 217, + 424, + 233 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\epsilon_{\\theta}$ is the denoising network. Standard diffusion models operate in the pixel space, but others like Latent Diffusion Models [39] instead operate on a latent image encoding space to reduce memory footprints. A U-Net architecture is commonly used to implement the denoising network and is typically trained on images at a fixed resolution $H\\times W$ [8, 39, 43]. The convolutional architecture of a U-Net allows for inputs and outputs of arbitrary spatial dimensions. In order to generate an image of a different size $\\bar{H}\\times \\bar{W}$ at inference time, one can sample an initial noise variable $\\bar{x}_T\\in \\mathbb{R}^{\\bar{H}\\times \\bar{W}\\times 3}$ and follow the same diffusion process. However, we find that this works poorly in practice, resulting in a significant degradation in output quality.", + "bbox": [ + 76, + 243, + 468, + 439 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Denoising Diffusion Implicit Models (DDIMs) introduce a faster non-Markovian sampling strategy, bypassing denoising steps. The reverse diffusion step in DDIM is:", + "bbox": [ + 76, + 440, + 468, + 484 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nx _ {t - 1} = \\sqrt {\\bar {\\alpha} _ {t - 1}} \\underbrace {\\left(x _ {t} - \\sqrt {1 - \\bar {\\alpha} _ {t} \\epsilon_ {\\theta} ^ {(t)} \\left(x _ {t}\\right)}\\right)} _ {\\text {p r e d i c t e d} \\hat {x} _ {0} ^ {t}} + \\underbrace {\\sqrt {1 - \\bar {\\alpha} _ {t - 1}} ,} _ {\\text {d i r e c t i o n p o i n t i n g t o} x _ {t}}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 86, + 505, + 468, + 583 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\bar{\\alpha}_t = \\prod_{i=1}^t 1 - \\beta_i$ is a cumulative product of the noise levels using a predetermined variance schedule $\\beta$ . $\\hat{x}_0^t$ is a noise-free estimation of $x_t$ obtained by subtracting the predicted noise scaled for step $t$ . We assume a deterministic sampling process in the DDIM formula [48] for simplicity.", + "bbox": [ + 76, + 592, + 468, + 667 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Classifier-Free Guidance updates the reverse diffusion process to condition it on a given condition $c$ as:", + "bbox": [ + 76, + 667, + 468, + 699 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\epsilon} _ {\\theta} ^ {(t)} (x _ {t}) = \\underbrace {\\epsilon_ {\\theta} ^ {(t)} (x _ {t})} _ {\\text {u n c o n d i t i o n a l}} + (1 + w) \\cdot \\underbrace {\\left(\\epsilon_ {\\theta} ^ {(t)} (x _ {t} , c) - \\epsilon_ {\\theta} ^ {(t)} (x _ {t})\\right)} _ {\\Delta_ {\\mathcal {C}} (x, c): \\text {c l a s s d i r e c t i o n s c o r e}}. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 84, + 720, + 468, + 767 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "$\\epsilon_{\\theta}(x,c)$ is a pretrained conditional diffusion model, and $w$ is a scaling factor. The difference between the conditional $\\epsilon_{\\theta}(x,c)$ and unconditional $\\epsilon_{\\theta}(x)$ scores, denoted as class direction score, gives the guidance direction towards $c$ .", + "bbox": [ + 76, + 775, + 468, + 838 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4. ElasticDiffusion", + "text_level": 1, + "bbox": [ + 76, + 844, + 235, + 859 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The aim of this work is to develop a method capable of synthesizing images at arbitrary size $\\bar{H} \\times \\bar{W}$ , and conforming", + "bbox": [ + 76, + 869, + 468, + 901 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/2170657f1a26f15a373285c892f64f17278698b15e753998df65798b2b4c97dd.jpg", + "image_caption": [ + "Figure 2. PCA of diffusion scores: class-direction score (top) dictates global content by clustering on semantic parts, while the unconditional score (bottom) lacks pixel correlations." + ], + "image_footnote": [], + "bbox": [ + 504, + 90, + 890, + 212 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "to a global condition $c$ using a pre-trained diffusion model that is limited at inference to its training resolution $H \\times W$ . To achieve this, we observe a main insight with respect to the diffusion scores in Eq. (2) that we visualize in Fig. 2. The class direction score $(\\Delta_c)$ primarily dictates the image's overall composition by showing clustering along semantic regions (e.g. hair) and edges. Thus, upscaling this score maintains global content in higher resolutions. In contrast, the unconditional score lacks inter-pixel correlation, suggesting a localized influence in enhancing detail at the pixel level. Thus, computing it requires a pixel-level precision. We study the global and local behaviour of these scores in Supp. Sec. 2.1. Our intuition behind this observation is that latent pixels, specifically in early diffusion steps where noise is high, display weak and short-range correlations. This leads the unconditional score to focus on local patterns due to these limited pixel interactions. Meanwhile, as $\\Delta_c$ represents a latent direction towards a specified global condition, it exerts a global influence by guiding clusters of pixels with similar directions towards semantic regions, overriding their initial noisy state. Leveraging this insight, we propose a method to decouple the generation of local and global content during the diffusion process. Specifically, we compute the unconditional score on local patches of size $H \\times W$ while simultaneously resizing a class direction score, originally derived for a reference latent of the dimensions $H \\times W$ as well. This dual strategy facilitates the generation of images at varied sizes using a pretrained diffusion model at its native resolution, all while maintaining the same memory requirement and without further training. We first present our approach for computing the unconditional score (Sec. 4.1). We then detail our method to estimate (Sec. 4.2) and upscale the resolution (Sec. 4.3) of the class direction score. Finally, we combine the two estimated scores with a novel guidance strategy to generate images at arbitrary sizes (Sec. 4.4). The generation process of ElasticDiffusion is illustrated in Fig. 3.", + "bbox": [ + 496, + 265, + 890, + 824 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4.1. Estimating the Unconditional Score", + "text_level": 1, + "bbox": [ + 500, + 833, + 808, + 849 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Building upon previous work [3, 53, 63], we estimate the unconditional score for a large latent signal $\\bar{x}_t\\in \\mathbb{R}^{\\bar{H}\\times \\bar{W}\\times 3}$ by concatenating scores derived from local patches. Specif", + "bbox": [ + 498, + 854, + 890, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "6605", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/ccc04c596e6212ceff276749c13a60680845a2627e4d227da6bd28139b0c0b42.jpg", + "image_caption": [ + "Figure 3. Illustration of ElasticDiffusion: We generate images at various sizes by generating local and global content separately. For local content, we partition the latent $\\bar{x}_t$ into non-overlapping patches $p_k$ , each concatenated with context $c_k$ to estimate their unconditional score. For global content, we downsample $\\bar{x}_t$ to $\\mathbf{x}_t$ , pad to a square size $(\\hat{\\mathbf{x}}_t)$ , compute class-direction score $(\\Delta_c)$ , and upscale to match $\\bar{x}_t$ ." + ], + "image_footnote": [], + "bbox": [ + 117, + 88, + 851, + 417 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "ically, we divide image $\\bar{x}_t$ to $K$ patches $P_{k} \\in \\mathbb{R}^{H \\times W \\times 3}$ and compute the score as $\\epsilon_{\\theta}(\\bar{x}_t) = \\{\\epsilon_{\\theta}(P_k); \\forall k \\leq K\\}$ . A common challenge encountered with this implementation is discontinuities at boundaries, as illustrated in Fig. 4 (A). To address this, earlier research calculated the diffusion model score on explicitly overlapping patches and averaged their scores in the intersecting regions [3, 53, 63]. While this strategy mitigates discontinuities, it requires large overlap between patches, thereby substantially increasing inference time and blurring details. To speed up the process, we introduce a more effective method that enhances boundary transitions in local patches by incorporating contextual information from adjacent patches, thus negating the need for signal averaging in overlapping areas. Specifically, we select patches smaller than the full size $p_{k} \\in \\mathbb{R}^{h \\times w \\times 3}$ with $h < H$ and $w < W$ , and concatenate them with context pixels from adjacent patches, denoted as $c_{k} \\in \\mathbb{R}^{(H - h) \\times (W - w) \\times 3}$ , to compute the diffusion model unconditional score $\\mathbf{S}_{u}$ as:", + "bbox": [ + 75, + 494, + 472, + 750 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {\\epsilon} _ {\\theta} \\left(x _ {t}\\right) = \\left\\{\\epsilon_ {\\theta} \\left(p _ {k} \\mid c _ {k}\\right) \\mid \\bar {x} _ {t} = \\left\\{p _ {k}; \\forall_ {k \\leq K} \\right\\} \\right\\}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 133, + 763, + 468, + 780 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $p_i$ is a local patch and $c_i$ are the context pixels surrounding it. This substantially increases the efficiency of the process. For instance, to generate an image of size $1024 \\times 1024$ , previous methods [3, 53, 63] used $87.5\\%$ overlap between adjacent patches, necessitating 81 forward diffusion calls per decoding step. In comparison, ElasticDiffusion achieves comparable results with only 9 forward calls,", + "bbox": [ + 76, + 794, + 470, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "as demonstrated in Fig. 4 (D). Employing the same number of calls, previous techniques result in obvious boundary discontinuities, as depicted in Fig. 4 (B).", + "bbox": [ + 496, + 496, + 890, + 542 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.2. Estimating the Class Direction Score", + "text_level": 1, + "bbox": [ + 498, + 550, + 816, + 566 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "A simple approach to estimate a class direction score of an intermediate latent signal $\\bar{x}_t \\in \\mathbb{R}^{\\bar{H} \\times \\bar{W} \\times 3}$ is to upsample the score from a reference latent $x_t \\in \\mathbb{R}^{N \\times M \\times 3}$ to $\\bar{H} \\times \\bar{W}$ . This is possible due to our observation that the class direction score represents a latent direction that can be shared between nearby pixels. We validate this observation empirically in Supplementary. We choose $N < H$ and $M < W$ such that $N \\times M$ is as close as possible to $H \\times W$ and $\\frac{\\bar{H}}{W} = \\frac{N}{M}$ . This is important to maintain the aspect ratio and prevent stretching the global content. Formally, we compute the class direction score $\\mathbf{S}_d$ as:", + "bbox": [ + 496, + 574, + 890, + 741 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathbf {x} _ {t} \\leftarrow \\text {D o w n s a m p l e} (\\bar {x} _ {t}, N \\times M), \\\\ \\Delta_ {\\mathcal {C}} (\\bar {x} _ {t}, c) = \\operatorname {U p s a m p l e} \\left(\\Delta_ {\\mathcal {C}} \\left(\\mathbf {x} _ {t}, c\\right), \\bar {H} \\times \\bar {W}\\right), \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 545, + 750, + 844, + 784 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "(4)", + "bbox": [ + 867, + 765, + 888, + 777 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\Delta_{\\mathcal{C}}(.)$ is the class direction score from Eq. (2), Downsample and Upsample are downsampling and upsampling operations. We use a nearest-neighbors approach to prevent altering the statistics of the latent signal. In order to maintain the input to the diffusion models at the size $H\\times W$ we dynamically pad the downsampled latent $\\mathbf{x}_t$ using a random background with a constant color.", + "bbox": [ + 496, + 794, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "6606", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/33797549b612d2d3da94f381f50662943123a329d503c4b31cf3843b6a85e416.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 89, + 89, + 269, + 154 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/3469265d4222a88721addc3eb441fe98dd1f99cb17554f84889c04b4a31b6bba.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 274, + 90, + 457, + 154 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/5487e88e0556e1adb268bd3e0d86267314fc29918e6a34d0891cc14fe3629a50.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 89, + 156, + 267, + 219 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/268bb81e8397b84389674ad6e464b572f4b5659ae0bdb4274b8f5ad1f87e057a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 274, + 155, + 457, + 219 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/366814e38e08459d15ca539142fbb5864edc21686a6e85cd7ae51e9236dd3cc3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 89, + 227, + 267, + 290 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/04299a38c6bfe4e4aa1b6da90de60d081fdf9e6afb77e272fe224f020dc99136.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 274, + 226, + 457, + 290 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/c91b5e67b6f85819184ca6bad9e0ed4de0f089a461e40f59fafae3a36ddd5150.jpg", + "image_caption": [ + "Intermediate step", + "Figure 4. Comparing strategies for calculating diffusion model score on a local patch. No overlap between adjacent patches (A) leads to discontinuities at the boundaries. Strategies (B) and (C), explicitly overlap nearby patches, necessitating substantial overlap to be effective. Our implicit overlapping method (D) achieves superior results with computational demand similar to (B)." + ], + "image_footnote": [], + "bbox": [ + 89, + 297, + 267, + 359 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/974809efb764c7055cac428db041fcf6e0d341f2c45eab564f16e7ed4255538d.jpg", + "image_caption": [ + "Generated Image" + ], + "image_footnote": [], + "bbox": [ + 274, + 295, + 457, + 359 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "This encourages the model to concentrate content generation within the center area. We then crop the extended parts from the predicted noise to the target image resolution $N \\times M$ . Formally we modify the forward call for the reverse diffusion step as:", + "bbox": [ + 75, + 474, + 468, + 549 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\dot {\\mathbf {x}} _ {t} \\leftarrow \\operatorname {P a d} \\left(\\mathbf {x} _ {t}, \\mathcal {A} _ {t} \\sqrt {\\bar {\\alpha} _ {t}} + \\sqrt {1 - \\bar {\\alpha} _ {t}} \\cdot \\mathcal {Z} _ {t}\\right),\n$$\n", + "text_format": "latex", + "bbox": [ + 147, + 561, + 423, + 579 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\epsilon_ {\\theta} \\left(\\mathbf {x} _ {t}, c\\right) = \\operatorname {C r o p} \\left(\\epsilon_ {\\theta} \\left(\\hat {\\mathbf {x}} _ {t}, c\\right), N \\times M\\right), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 104, + 580, + 468, + 597 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathcal{Z}_t\\sim \\mathcal{N}(0,I)$ represents the injected Gaussian noise at each step, and $\\mathcal{A}_t$ represents a background image of size $(H - N)\\times (W - M)$ with a constant color value $\\mathcal{V}$ randomly sampled from a uniform distribution. This simple padding operation guarantees that the input to the diffusion model is kept at $H\\times W$ , while encouraging the diffusion model to keep the generated content within the cropped $N\\times M$ center that we are interested in.", + "bbox": [ + 75, + 609, + 468, + 729 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.3. Refined Class Direction Score", + "text_level": 1, + "bbox": [ + 76, + 741, + 341, + 755 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Sharing the class direction score among nearby pixels can result in over-smoothed images. To mitigate this, we propose an iterative resampling technique that increases the resolution of the estimated class direction score by extrapolating missing image components from their surrounding context, following [29]. Our technique involves a gradual enhancement of the class direction score's resolution by estimating and integrating it for newly sampled pixels. Specifically, in each iteration, we replace $n\\%$ of the pixels", + "bbox": [ + 75, + 763, + 468, + 900 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/f1eecfec9c801373a03fa89b89c30f1b9f0ed689a4b8c335368fe9440d1c942b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 501, + 90, + 890, + 157 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/4c3a5bc0c419ed0226ee95dc439211064e8c2d0018f3862702e4231afca6093b.jpg", + "image_caption": [ + "Figure 5. The Effect of Reduced-Resolution Guidance (RRG). Higher RRG weights effectively eliminates emerging artifacts albeit at the cost of slightly blurrier outputs. $\\delta = 200$ strikes a good balance. Improvements are more noticeable when zooming in.", + "Figure 6. Effect of Resampling. Applying more resampling steps noticeably enhances detail in the generated images." + ], + "image_footnote": [], + "bbox": [ + 501, + 229, + 890, + 275 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "in $\\mathbf{x}_t$ with newly sampled ones from $x_{t}$ to get an updated version $\\tilde{\\mathbf{x}}_t$ . Following each update, the direction score is recalculated and blended with the previously calculated score. Formally, we consider $\\mathbf{S}_d^0 = \\mathbf{S}_d$ and define the iterative resampling as:", + "bbox": [ + 498, + 330, + 890, + 406 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {S} _ {d} ^ {r + 1} = \\mathbf {S} _ {d} ^ {r} \\odot m + \\tilde {\\mathbf {S}} _ {d} \\odot (1 - m), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 578, + 419, + 890, + 439 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $m \\in \\{0,1\\}^{\\bar{H} \\times \\bar{W}}$ is a mask with a value of 1 at the positions of the newly sampled pixels and 0 elsewhere. $\\tilde{\\mathbf{S}}_d$ represents the recalculated class direction score on $\\tilde{\\mathbf{x}}_t$ as per Eq. (4). This method estimates the class direction score of $n\\%$ new pixels while retaining the information of the previously estimated score, thereby increasing the score's overall resolution. In our experiments, we set $n$ to $20\\%$ and repeat the process $R$ times, depending on the target generation resolution. Fig. 6 demonstrates the effectiveness of our method in enhancing the details of the generated images.", + "bbox": [ + 496, + 450, + 890, + 604 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.4. Reduced-Resolution Guidance (RRG)", + "text_level": 1, + "bbox": [ + 500, + 614, + 823, + 630 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We effectively estimate the unconditional score signal and concurrently steer the image generation using the class direction score. However, inaccuracies in unconditional score estimation or fluctuations in local content, especially from distant patches, can lead to artifacts. To enhance image coherence and diminish artifacts, we consider a downsampled version of the latent at each decoding step as a reference and aim to align the decoded latent with it through our reduced-resolution latent update strategy. Specifically, we utilize the noise-free sample $\\hat{x}_0^t$ of the latent $x_{t}$ from Eq. (1) and generate a corresponding noise-free sample $\\hat{\\mathbf{x}}_0^t$ from its downsampled counterpart $\\mathbf{x}^t$ in Eq. (4) as:", + "bbox": [ + 496, + 637, + 890, + 805 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathbf {x}} _ {0} ^ {t} = \\frac {1}{\\sqrt {\\bar {\\alpha} _ {t}}} (\\mathbf {x} _ {t} - \\sqrt {1 - \\bar {\\alpha} _ {t}} \\cdot (\\epsilon_ {\\theta} (\\mathbf {x} _ {t}) + (1 + w) \\cdot \\Delta_ {\\mathcal {C}} (\\mathbf {x} _ {t}, c))),\n$$\n", + "text_format": "latex", + "bbox": [ + 506, + 815, + 882, + 844 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Here, both $\\hat{x}_0^t$ and $\\hat{\\mathbf{x}}_0^t$ corresponds to the same latent update at different resolutions. Due to its smaller dimension, $\\hat{\\mathbf{x}}_0^t$ has a broader context when computing the unconditional", + "bbox": [ + 500, + 854, + 890, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "6607", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/59eb54a5213e8dbd5fe4904d66b7320fc934009382dd44d7d1833a563cfd971f.jpg", + "image_caption": [ + "Figure 7. Qualitative comparison at various resolutions on CelebAHQ faces. ElasticDiffusion consistently generates coherent images at all resolutions. StableDiffusion, and MultiDiffusion produce repeating body parts at higher resolutions, while StableDiffusion-XL fails to maintain its quality at lower resolutions, resulting in noisy outputs at $256 \\times 256$ . We exclude MultiDiffusion results at $256 \\times 256$ as it is not designed to produce images at lower resolutions." + ], + "image_footnote": [], + "bbox": [ + 96, + 92, + 872, + 400 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "signal. To guide $x_{t}$ with its downsampled reference, we refine the latent $x_{t}$ with the direction that minimizes $L2$ difference between $\\hat{\\mathbf{x}}_0^t$ and $\\hat{x}_0^t$ . Formally,", + "bbox": [ + 75, + 486, + 470, + 532 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\bar {x} _ {t - 1} \\leftarrow \\bar {x} _ {t - 1} - \\delta_ {t} \\nabla_ {x _ {t}} | | \\text {U p s a m p l e} \\left(\\hat {\\mathbf {x}} _ {0} ^ {t}, \\bar {H} \\times \\bar {W}\\right) - \\hat {x} _ {0} ^ {t} | |, \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 83, + 537, + 468, + 556 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\delta_t$ represent the weight of the guidance. Since the overall image structure is determined in the early diffusion steps, we start with $\\delta_t = 400$ and follow a cosine scheduler to decrease this weight for later diffusion steps. This scheduling strategy mitigates potential quality degradation from matching the generated image with a lower-resolution version while allowing the model to fill-in higher-resolution details in the later decoding stages. Fig. 5 illustrates how RRG eliminates emerging artifacts.", + "bbox": [ + 75, + 561, + 468, + 699 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 76, + 708, + 209, + 726 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "ElasticDiffusion supports generating images across different resolutions and aspect ratios. Our experiments focus on: (1) square images at multiple resolutions, and (2) images with varied aspect ratios and resolutions.", + "bbox": [ + 75, + 734, + 468, + 795 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Datasets. We evaluate the generation of square images on the Multi-Modal CelebAHQ dataset [56], which includes high-resolution square face images accompanied with text descriptions. We evaluate different aspect ratio generation using the LAION-COCO dataset [46], derived from the web-crawled LAION-5B dataset [45], which includes a variety of image aspect ratios and contains landscapes, people,", + "bbox": [ + 75, + 795, + 470, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "objects, and everyday scenes, each paired with a synthetic caption generated using BLIP [27]. We consider four common aspect ratios: 9:16, 16:9, 3:4 and 4:3.", + "bbox": [ + 496, + 486, + 890, + 532 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation Metrics. Following prior text-to-image synthesis works [3, 33, 35, 38, 43], we use automatic evaluation metrics Frechet Inception Distance (FID) and CLIP-score. FID [17] measures both the realism and diversity of the generated images by calculating the difference between features of the real and generated images computed using Inception-V3 [49] pretrained on ImageNet [41]. CLIP-score uses a pretrained text-image CLIP model [36] to measure alignment between the generated images and input prompts. We use 10,000 images to compute these scores.", + "bbox": [ + 496, + 534, + 892, + 686 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Baselines. We compare our approach against prior diffusion model generation methods, specifically focusing on Stable Diffusion (SD) and MultiDiffusion (MD). SD follows the standard reverse diffusion process on an image latent space. MD uses a pretrained $SD$ for panoramic image generation, by creating smaller, overlapping patches and averaging the Diffusion Model scores in intersecting areas. For baseline comparisons, we fix the pre-trained diffusion model to $SD_{1.4}$ , which is trained to generate images at a resolution of $512 \\times 512$ . Additionally, we compare our method with SDXL, an enhanced SD model that is three times larger than the standard one. SDXL is trained at a higher resolution of $1024p$ , and fine-tuned on a specific set of aspect ratios with pixel sums close to $1024^2$ . We exclude", + "bbox": [ + 496, + 688, + 893, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6608", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/6596084cb3897a01d054b5e051ecb42341e87ca58efe25e350836c561a05a7d3.jpg", + "image_caption": [ + "Figure 8. Qualitative comparison on various aspect ratios. ElasticDiffusion effectively handles a variety of aspect ratios. In comparison, $SD$ and MultiDiffusion generate unrealistic images, while SDXL outputs exhibit a decline in the perceptual quality." + ], + "image_footnote": [], + "bbox": [ + 81, + 88, + 467, + 500 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "the latent refinement module in $SDXL$ and focus our analysis on the base model. Throughout our experiments, we employ a DDIM sampling strategy with 50 steps and use 7.0 for the classifier-free guidance scaling factor. We also ensure consistency in seeds and captions for all baselines.", + "bbox": [ + 75, + 578, + 470, + 655 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.1. Qualitative Results", + "text_level": 1, + "bbox": [ + 76, + 664, + 261, + 681 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We show square image generation samples in Fig. 7 generated from the CelebAHQ dataset at various resolutions. Both MultiDiffusion and Stable Diffusion have a tendency to replicate textures and body parts at higher resolutions, resulting in images that lack coherence. At resolutions lower than its training resolution $512 \\times 512$ , StableDiffusion aligns poorly with the provided captions and produces unappealing images. SDXL, trained for $1024 \\times 1024$ resolution, has a similar trend of reduced perceptual quality at lower sizes, eventually producing complete noise at a resolution of $256 \\times 256$ . In contrast, ElasticDiffusion maintains image coherence across all tested resolutions, and yields results comparable to SDXL at $1024 \\times 1024$ , despite using a less powerful base model and lower memory. We excluded", + "bbox": [ + 75, + 688, + 472, + 902 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 1. Quantitative comparison of on CelebA-HQ and LAION-COCO at different resolution. We indicate the best performances in bold, and underline the second best ones. #Calls represent the number of diffusion model calls required at each decoding steps.", + "bbox": [ + 498, + 88, + 890, + 146 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/aa400ad434a010a46ffbea54e5211893b749a2aa5d0288d81af08a59046f7b98.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Res.MethodCelebA-HQLAION-COCO#CallsMem.
FID ↓CLIP ↑FID ↓CLIP ↑
256 × 256SD1.4258.4320.1454.0621.4327.2GB
SDXL368.0614.40175.8714.60218.5GB
Ours1.4235.2323.8823.7726.3028.6GB
512 × 512SD1.4233.4024.0020.5027.3328.6GB
SDXL240.2021.5742.5825.34221.6GB
768 × 768SD1.4238.8723.4529.8927.01211.1GB
MD1.4240.5622.8229.9827.31508.6GB
SDXL225.4824.2323.3127.88224.5GB
Ours1.4225.8626.6625.7825.93178.6GB
1024 × 1024SD1.4266.0121.9047.0125.70114.7GB
MD1.4264.5721.5537.7026.961628.6GB
SDXL230.2124.6225.5828.06127.5GB
Ours1.4228.8723.7427.7626.07338.6GB
", + "bbox": [ + 503, + 155, + 890, + 498 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "results at $512 \\times 512$ because both our method and Multi-Diffusion generate same outputs as the base Stable Diffusion model. We also omitted the results of MultiDiffusion at $256 \\times 256$ since it is not designed to generate images at sizes smaller than the base model. Fig. 8 presents generated samples at various aspect ratios. We observe a similar trend of pattern repetition and reduced perceptual quality for images generated by the baselines, in contrast to those generated by ElasticDiffusion. Finally, we apply our method using SDXL as the base model to enable Full-HD image generation ( $1920 \\times 1080$ ) and provide examples in Supp. Sec. 3.3.", + "bbox": [ + 496, + 505, + 893, + 674 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.2. Quantitative Results", + "text_level": 1, + "bbox": [ + 498, + 680, + 694, + 696 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 1 shows quantitative evaluations. StableDiffusion shows increasing image quality degradation, as indicated by the FID metric when processing images of sizes different from its training resolution $512 \\times 512$ . MultiDiffusion slightly improves FID at the expense of a substantially more forward calls. SDXL demonstrates similar declines in quality at resolutions far from its fine-tuning size of $1024 \\times 1024$ . ElasticDiffusion, however, improves FID while maintaining a comparable CLIP-score to the base model. ElasticDiffusion also significantly improves the performance for lower resolutions $256 \\times 256$ , while achieving comparable results to SDXL at its training resolution $1024 \\times 1024$ , with only $\\sim 31\\%$ of the memory requirement for SDXL.", + "bbox": [ + 496, + 703, + 893, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "6609", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/955f8e0bc6d5a70596f004f778f377e9c1f4af826b69af30ab1a822c159443f5.jpg", + "table_caption": [ + "Table 2. Quantitative comparison of on LAION-COCO datasets at various aspect ratios (A) and resolutions (R). best performances are in bold, and the second best are underlined. Vertical means the resolution is transposed from H:W to W:H." + ], + "table_footnote": [], + "table_body": "
ARMethodHorizontalVertical
FID ↓CLIP↑FID ↓CLIP↑
3:4384 × 512SD1.438.8624.6317.6626.54
MD1.4----
SDXL104.8421.3368.8422.40
Ours1.435.1024.9115.5027.33
512 × 680SD1.445.5424.9616.8127.33
MD1.443.4424.5619.1326.80
SDXL62.8024.2028.2326.17
Ours1.441.0624.4018.9026.83
768 × 1024SD1.471.0024.0928.8326.30
MD1.454.8925.0226.3526.95
SDXL47.0525.7919.5027.41
Ours1.447.0324.9122.5225.80
9:16288 × 512SD1.423.5024.6924.0124.89
MD1.4----
SDXL121.8317.65112.4118.54
Ours1.423.2325.2622.8626.30
512 × 904SD1.429.8625.3427.4526.01
MD1.426.3525.7026.7025.28
SDXL29.6025.4027.2726.08
Ours1.422.8525.0126.6826.12
", + "bbox": [ + 81, + 155, + 465, + 577 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 2 provides an evaluation on a variety of aspect ratios and resolutions for both horizontal and vertical image resolutions. StableDiffusion obtains worse FID score with increasing resolutions, while surprisingly maintaining or improving its CLIP-score. We posit that since CLIP-score quantifies the agreement between the input prompt and the generated image, StableDiffusion benefits from generating repetitive textures and artifacts that align closely with the prompt. For example, a photo of repeated lipsticks might align more with the caption \"lipstick\" despite its poor composition. MultiDiffusion generally enhances image quality but does not achieve satisfactory performance. MultiDiffusion struggles with objects that span the entire image (e.g. Fig. 8). Our method consistently improves FID over the baseline on horizontal and most vertical resolutions while preserving fidelity to the input prompts, thereby attaining comparable or superior CLIP-scores. Remarkably, even with a larger model size and explicit fine-tuning at a similar resolution of $768 \\times 1280$ , SDXL only marginally surpasses our method at the $768 \\times 1024$ resolution.", + "bbox": [ + 75, + 598, + 468, + 900 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/802cad3ac0b2ab9322ae3dc171970570a045452587c5674b9b97134ba8427517.jpg", + "table_caption": [ + "Table 3. Ablation analysis of ElasticDiffusion on 500 images." + ], + "table_footnote": [], + "table_body": "
Model DetailsFID ↓CLIP ↑
ElasticDiffusion133.6725.82
w/o Resampling150.0123.82
w/o RRG150.6424.34
w/o Imp. overlap, w/ exp. overlap141.4225.82
", + "bbox": [ + 529, + 114, + 854, + 195 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.3. Ablation study", + "text_level": 1, + "bbox": [ + 500, + 207, + 650, + 223 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Tab. 3 presents results of our method when excluding key components, demonstrating that each element improves performance. Fig. 4 illustrates the effectiveness of our proposed implicit overlap method in resolving boundary discontinuities at a reduced computational cost. Fig. 5 highlights the effectiveness of Reduced-Resolution Guidance in removing emerging artifacts. Fig. 6 shows the effectiveness of our iterative resampling technique in enhancing details.", + "bbox": [ + 496, + 231, + 890, + 353 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Discussion and Conclusion", + "text_level": 1, + "bbox": [ + 498, + 369, + 746, + 386 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Experimental results highlight the adaptability and effectiveness of ElasticDiffusion at steering diffusion models to produce an array of resolutions and aspect ratios. ElasticDiffusion requires no fine-tuning, consumes a constant memory footprint, enables both increased and reduced resolutions, and can generate a variety of aspect ratios.", + "bbox": [ + 496, + 396, + 890, + 487 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "ElasticDiffusion, however, does have several practical limitations. First, inaccuracies in estimating the global and local signals may result in artifacts. Although we attempt to mitigate artifacts with our Reduced-resolution guidance, it can still generate blurrier outputs. Second, since the global content guidance is initially estimated at the original training resolution of the underlying diffusion model, our method is less effective in generating images at significantly extended sizes beyond the training resolution. In particular, at extreme resolutions beyond 4X, our method produces artifacts and images of a lower perceptual quality. We provide examples of these failure cases in Supp. Sec. 10.4.", + "bbox": [ + 496, + 488, + 892, + 669 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The main insight underpinning our method is a novel reinterpretation of classifier-free guidance in somewhat disentangling both global and local content. Our comprehensive evaluations demonstrate the feasibility of disentangling these signals, yet the full extent of their separation offers an avenue for further exploration in this direction. We hope that our findings inspire future work in investigating the separation of global and local content guidance signals for image synthesis. This separation holds potential for various applications such as selectively manipulating local and global content or enhancing style transfer. Additionally, the rich semantic representation in the class-direction score has the potential for improving discriminative models.", + "bbox": [ + 496, + 670, + 890, + 867 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments. This work was partially funded by NSF Award #2201710.", + "bbox": [ + 500, + 869, + 890, + 898 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "6610", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Omri Avrahami, Dani Lischinski, and Ohad Fried. Blended diffusion for text-driven editing of natural images. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). IEEE, 2022. 2", + "[2] Arpit Bansal, Hong-Min Chu, Avi Schwarzschild, Soumyadip Sengupta, Micah Goldblum, Jonas Geiping, and Tom Goldstein. Universal guidance for diffusion models, 2023. 2", + "[3] Omer Bar-Tal, Lior Yariv, Yaron Lipman, and Tali Dekel. Multidiffusion: Fusing diffusion paths for controlled image generation. In ICML, 2023. 2, 3, 4, 6", + "[4] Angela Castillo, Jonas Kohler, Juan C. Pérez, Juan Pablo Pérez, Albert Pumarola, Bernard Ghanem, Pablo Arbeláez, and Ali Thabet. Adaptive guidance: Training-free acceleration of conditional diffusion models, 2023. 2", + "[5] Duygu Ceylan, Chun-Hao Paul Huang, and Niloy J. Mitra. Pix2video: Video editing using image diffusion. In ICCV, 2023. 2", + "[6] Hyungjin Chung, Jeongsol Kim, Michael T. McCann, Marc L. Klasky, and Jong Chul Ye. Diffusion posterior sampling for general noisy inverse problems, 2023. 2", + "[7] Jiwoo Chung, Sangeek Hyun, and Jae-Pil Heo. Style injection in diffusion: A training-free approach for adapting large-scale diffusion models for style transfer, 2024. 2", + "[8] Prafulla Dhariwal and Alex Nichol. Diffusion models beat gans on image synthesis. In NeurIPS, 2021. 2, 3", + "[9] Ruoyi Du, Dongliang Chang, Timothy Hospedales, Yi-Zhe Song, and Zhanyu Ma. Demofusion: Democratising high-resolution image generation with no $$$. In CVPR, 2024. 2", + "[10] Alexandros Graikos, Srikar Yellapragada, Minh-Quan Le, Saarthak Kapse, Prateek Prasanna, Joel Saltz, and Dimitris Samaras. Learned representation-guided diffusion models for large-image generation, 2024. 2", + "[11] Jiatao Gu, Qingzhe Gao, Shuangfei Zhai, Baoquan Chen, Lingjie Liu, and Josh Susskind. Control3diff: Learning controllable 3d diffusion models from single-view images. In 3DV24, 2023. 2", + "[12] Lanqing Guo, Yingqing He, Haoxin Chen, Menghan Xia, Xiaodong Cun, Yufei Wang, Siyu Huang, Yong Zhang, Xintao Wang, Qifeng Chen, Ying Shan, and Bihan Wen. Make a cheap scaling: A self-cascade diffusion model for higher-resolution adaptation, 2024. 2", + "[13] Feihong He, Gang Li, Lingyu Si, Leilei Yan, Shimeng Hou, Hongwei Dong, and Fanzhang Li. Cartooniff: Training-free cartoon image generation with diffusion transformer models, 2023. 2", + "[14] Yingqing He, Shaoshu Yang, Haoxin Chen, Xiaodong Cun, Menghan Xia, Yong Zhang, Xintao Wang, Ran He, Qifeng Chen, and Ying Shan. Scalecrafter: Tuning-free higher-resolution visual generation with diffusion models, 2023. 2", + "[15] Yingqing He, Tianyu Yang, Yong Zhang, Ying Shan, and Qifeng Chen. Latent video diffusion models for high-fidelity long video generation, 2023. 2" + ], + "bbox": [ + 78, + 114, + 468, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[16] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-prompt image editing with cross attention control. In ICLR, 2022. 2", + "[17] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. In NIPS, 2017. 6", + "[18] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance, 2022. 2", + "[19] Jonathan Ho, Chitwan Sahara, William Chan, David J. Fleet, Mohammad Norouzi, and Tim Salimans. Cascaded diffusion models for high fidelity image generation. Journal of Machine Learning Research, 23(47):1-33, 2022. 2", + "[20] Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J. Fleet. Video diffusion models. In NeurIPS, 2022. 2", + "[21] Edward J. Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. In ICLR, 2022. 2", + "[22] Rongjie Huang, Jiawei Huang, Dongchao Yang, Yi Ren, Luping Liu, Mingze Li, Zhenhui Ye, Jinglin Liu, Xiang Yin, and Zhou Zhao. Make-an-audio: Text-to-audio generation with prompt-enhanced diffusion models. In ICML, 2023. 2", + "[23] Yujin Jeong, Wonjeong Ryoo, Seunghyun Lee, Dabin Seo, Wonmin Byeon, Sanggil Kim, and Jinkyu Kim. The power of sound (tpos): Audio reactive video generation with stable diffusion. In ICCV, 2023. 2", + "[24] Zhiyu Jin, Xuli Shen, Bin Li, and Xiangyang Xue. Training-free diffusion model adaptation for variable-sized text-to-image synthesis, 2023. 2", + "[25] Gwanghyun Kim, Taesung Kwon, and Jong Chul Ye. Diffusionclip: Text-guided diffusion models for robust image manipulation. In CVPR, 2022. 2", + "[26] Zhifeng Kong and Wei Ping. On fast sampling of diffusion probabilistic models. In ICML, 2021. 2", + "[27] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation, 2022. 6", + "[28] Jin Liu, Huaibo Huang, Chao Jin, and Ran He. Portrait diffusion: Training-free face stylization with chain-of-painting, 2023. 2", + "[29] Andreas Lugmayr, Martin Danelljan, Andres Romero, Fisher Yu, Radu Timofte, and Luc Van Gool. Repaint: Inpainting using denoising diffusion probabilistic models. In CVPR, 2022. 5", + "[30] Sicheng Mo, Fangzhou Mu, Kuan Heng Lin, Yanli Liu, Bochen Guan, Yin Li, and Bolei Zhou. Freecontrol: Training-free spatial control of any text-to-image diffusion model with any condition, 2023. 2", + "[31] Brian B. Moser, Stanislav Frolov, Federico Raue, Sebastian Palacio, and Andreas Dengel. Dynamic attention-guided diffusion for image super-resolution, 2024. 2", + "[32] Alex Nichol and Prafulla Dhariwal. Improved denoising diffusion probabilistic models. In ICML, 2021. 2", + "[33] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and" + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "6611", + "bbox": [ + 482, + 945, + 513, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. In ICML, 2022. 2, 6", + "[34] Bo Peng, Xinyuan Chen, Yaohui Wang, Chaochao Lu, and Yu Qiao. Conditionvideo: Training-free condition-guided text-to-video generation, 2023. 2", + "[35] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis, 2023. 2, 6", + "[36] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In ICML, 2021. 6", + "[37] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents, 2022. 2", + "[38] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents, 2022. 6", + "[39] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 1, 2, 3", + "[40] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In CVPR, 2023. 2", + "[41] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, Alexander C. Berg, and Li Fei-Fei. Imagenet large scale visual recognition challenge, 2015. 6", + "[42] Chitwan Sahara, Jonathan Ho, William Chan, Tim Salimans, David J. Fleet, and Mohammad Norouzi. Image superresolution via iterative refinement, 2021. 2", + "[43] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, Seyed Kamyar Seyed Ghasemipour, Burcu Karagol Ayan, S. Sara Mahdavi, Rapha Gontijo Lopes, Tim Salimans, Jonathan Ho, David J Fleet, and Mohammad Norouzi. Photorealistic text-to-image diffusion models with deep language understanding. In NeurIPS, 2022. 2, 3, 6", + "[44] Axel Sauer, Frederic Boesel, Tim Dockhorn, Andreas Blattmann, Patrick Esser, and Robin Rombach. Fast high-resolution image synthesis with latent adversarial diffusion distillation, 2024. 2", + "[45] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, Patrick Schramowski, Srivatsa Kundurthy, Katherine Crowson, Ludwig Schmidt, Robert Kaczmarczyk, and Jenia Jitsev. Laion-5b: An open large-scale dataset for training next generation image-text models, 2022. 6", + "[46] Christoph Schuhmann, Andreas A. Köpf, Richard Vencu, Theo Coombes, and Ross Beaumont. Laion coco: 600m synthetic captions from laion2b-en, 2023. 6" + ], + "bbox": [ + 78, + 90, + 468, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[47] Fengyuan Shi, Jiaxi Gu, Hang Xu, Songcen Xu, Wei Zhang, and Limin Wang. Bivdiff: A training-free framework for general-purpose video synthesis via bridging image and video diffusion models, 2023. 2", + "[48] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In ICLR, 2021. 2, 3", + "[49] Christian Szegedy, Vincent Vanhoucke, Sergey Ioffe, Jonathon Shlens, and Zbigniew Wojna. Rethinking the inception architecture for computer vision. In CVPR, 2015. 6", + "[50] Andrey Voynov, Amir Hertz, Moab Arar, Shlomi Fruchter, and Daniel Cohen-Or. Anylens: A generative diffusion model with any rendering lens, 2023. 2", + "[51] Bingyuan Wang, Hengyu Meng, Zeyu Cai, Lanjiong Li, Yue Ma, Qifeng Chen, and Zeyu Wang. Magicscrol: Nontypical aspect-ratio image generation for visual storytelling via multi-layered semantic-aware denoising, 2023. 2", + "[52] Fu-Yun Wang, Zhaoyang Huang, Xiaoyu Shi, Weikang Bian, Guanglu Song, Yu Liu, and Hongsheng Li. *Animatelcm: Accelerating the animation of personalized diffusion models and adapters with decoupled consistency learning*, 2024. 2", + "[53] Jianyi Wang, Zongsheng Yue, Shangchen Zhou, Kelvin C. K. Chan, and Chen Change Loy. Exploiting diffusion prior for real-world image super-resolution, 2023. 2, 3, 4", + "[54] Lezhong Wang, Jeppe Revall Frisvad, Mark Bo Jensen, and Siavash Arjomand Bigdeli. Stereodiffusion: Training-free stereo image generation using latent diffusion models, 2024. 2", + "[55] Daniel Watson, William Chan, Ricardo Martin-Brualla, Jonathan Ho, Andrea Tagliasacchi, and Mohammad Norouzi. Novel view synthesis with diffusion models. In ICLR, 2023. 2", + "[56] Weihao Xia, Yujiu Yang, Jing-Hao Xue, and Baoyuan Wu. Tedigan: Text-guided diverse face image generation and manipulation, 2021. 6", + "[57] Guy Yariv, Itai Gat, Lior Wolf, Yossi Adi, and Idan Schwartz. Audiotoken: Adaptation of text-conditioned diffusion models for audio-to-image generation, 2023. 2", + "[58] Ahmet Burak Yildirim, Vedat Baday, Erkut Erdem, Aykut Erdem, and Aysegul Dundar. Inst-inpaint: Instructing to remove objects with diffusion models, 2023. 2", + "[59] Jiwen Yu, Yinhuai Wang, Chen Zhao, Bernard Ghanem, and Jian Zhang. Freedom: Training-free energy-guided conditional diffusion model, 2023. 2", + "[60] Shen Zhang, Zhaowei Chen, Zhenyu Zhao, Zhenyuan Chen, Yao Tang, Yuhao Chen, Wengang Cao, and Jiajun Liang. Hidiffusion: Unlocking high-resolution creativity and efficiency in low-resolution trained diffusion models, 2023. 2", + "[61] Peiang Zhao, Han Li, Ruiyang Jin, and S. Kevin Zhou. Loco: Locally constrained training-free layout-to-image synthesis, 2024. 2", + "[62] Qingping Zheng, Yuanfan Guo, Jiankang Deng, Jianhua Han, Ying Li, Songcen Xu, and Hang Xu. Any-sized diffusion: Toward efficient text-driven synthesis for any-size hd images, 2023. 2", + "[63] Álvaro Barbero Jiménez. Mixture of diffusers for scene composition and high resolution image generation, 2023. 2, 3, 4" + ], + "bbox": [ + 501, + 92, + 890, + 901 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "6612", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2024/ElasticDiffusion_ Training-free Arbitrary Size Image Generation through Global-Local Content Separation/de40da54-b1b7-40fa-9aec-0f4f1bc4f406_model.json b/2024/ElasticDiffusion_ Training-free Arbitrary Size Image Generation through Global-Local Content Separation/de40da54-b1b7-40fa-9aec-0f4f1bc4f406_model.json new file mode 100644 index 0000000000000000000000000000000000000000..48b705768ee5ed1efc3090b0dd695071dee2906e --- /dev/null +++ b/2024/ElasticDiffusion_ Training-free Arbitrary Size Image Generation through Global-Local Content Separation/de40da54-b1b7-40fa-9aec-0f4f1bc4f406_model.json @@ -0,0 +1,2112 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.16, + 0.131, + 0.812, + 0.176 + ], + "angle": 0, + "content": "ElasticDiffusion: Training-free Arbitrary Size Image Generation through Global-Local Content Separation" + }, + { + "type": "text", + "bbox": [ + 0.156, + 0.204, + 0.294, + 0.222 + ], + "angle": 0, + "content": "Moayed Haji-Ali" + }, + { + "type": "text", + "bbox": [ + 0.409, + 0.204, + 0.563, + 0.239 + ], + "angle": 0, + "content": "Guha Balakrishnan Rice University" + }, + { + "type": "text", + "bbox": [ + 0.342, + 0.242, + 0.631, + 0.256 + ], + "angle": 0, + "content": "{mh155, guha, vicenteor}@rice.edu" + }, + { + "type": "text", + "bbox": [ + 0.679, + 0.204, + 0.815, + 0.22 + ], + "angle": 0, + "content": "Vicente Ordonez" + }, + { + "type": "image", + "bbox": [ + 0.08, + 0.272, + 0.891, + 0.617 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.625, + 0.893, + 0.669 + ], + "angle": 0, + "content": "Figure 1. ElasticDiffusion generates high quality images at arbitrary sizes using a pretrained diffusion model trained on a single image size, with equivalent memory footprint and no further training. These results are based on Stable Diffusion1,4, which was trained to generate \\(512 \\times 512\\) images. The examples shown in this collage are presented without any image cropping, stretching, or post-processing." + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.68, + 0.314, + 0.696 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.705, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Diffusion models have revolutionized image generation in recent years, yet they are still limited to a few sizes and aspect ratios. We propose ElasticDiffusion, a novel training-free decoding method that enables pretrained text-to-image diffusion models to generate images with various sizes. ElasticDiffusion attempts to decouple the generation trajectory of a pretrained model into local and global signals. The local signal controls low-level pixel information and can be estimated on local patches, while the global signal is used to maintain overall structural consistency and is estimated with a reference image. We test our method on CelebA-HQ (faces) and LAION-COCO (objects/indoor/outdoor scenes). Our experiments and qualita" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.682, + 0.892, + 0.743 + ], + "angle": 0, + "content": "tive results show superior image coherence quality across aspect ratios compared to MultiDiffusion and the standard decoding strategy of Stable Diffusion. Project Webpage: https://elasticdiffusion.github.io/" + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.752, + 0.631, + 0.767 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.78, + 0.893, + 0.901 + ], + "angle": 0, + "content": "Diffusion models are a powerful family of algorithms that achieve remarkable quality and obtain the current state-of-the-art performance on various image synthesis tasks. As is the case with most deep neural networks, diffusion models are typically trained on one or a few image resolutions. For instance, Stable Diffusion (SD) [39], one of the most widely adopted diffusion models, is trained on a square images of size \\(512 \\times 512\\), yet fails to maintain its performance at dif-" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6603" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.152 + ], + "angle": 0, + "content": "ferent aspect ratios during inference time. In practice, many applications require a wide aspect ratio or portrait mode, such as digital billboards, wearable devices, automotive displays, and any application relying on a computer screen." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.155, + 0.473, + 0.458 + ], + "angle": 0, + "content": "Recent studies address the issue of variable image size in different ways. SDXL [35] and Any-Size-Diffusion [62] explicitly finetune models on images with a range of aspect ratios, which requires extensive computation, a quadratic memory footprint, and larger training data. In addition, this strategy requires a set of resolutions to be specified upfront during training time, while the models still struggle to generalize to new resolutions during inference, often resulting in artifacts. Recent works also show remarkable results in generating panoramic images using pretrained diffusion models by overlapping generated patches into a larger image [3, 63]. These methods work well for landscape images with repetitive patterns. However, their lack of global guidance limits their abilities to generate images of single objects or faces where global structure is important. Recent work [24, 50, 51] aimed at extending pre-trained diffusion models capabilities to others domains, often in a training-free way [4, 7, 13, 28, 30, 34, 47, 54, 61]. Few concurrent work [9, 10, 12, 31, 60] focus on adapting them to higher resolutions, yet they are constrained to square images." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.46, + 0.473, + 0.748 + ], + "angle": 0, + "content": "In this work, we propose ElasticDiffusion, a novel decoding strategy that takes a pretrained diffusion model and generate images at arbitrary sizes during inference using a constant memory footprint. To achieve this, we revisit the guidance mechanism of conditional diffusion models to decouple global and local content generation. Global content controls the high-level aspects of the image, whereas local content adds finer, more granular details. This separation facilitates generating the local content in patches for images of varying sizes, all while being guided with global content that we derive from a reference image at the diffusion model pretraining resolution. This enables the synthesis of images at diverse resolutions and aspect ratios while adhering the diffusion forward calls to the model's initial training size. To aid in this task, we introduce several techniques including an efficient patch fusion method for smooth boundaries, a novel guidance strategy to reduce image artifacts, and a global content resampling technique that amplifies the resolution of diffusion models up to 4X the training size." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.473, + 0.903 + ], + "angle": 0, + "content": "Figure 1 shows a diverse array of images generated with ElasticDiffusion. Several of theses images include a single object or were generated with extreme aspect ratios, showcasing our method's ability to produce coherent images under various sizes. Quantitatively, ElasticDiffusion outperforms baselines across most aspect ratios. More importantly, despite relying on Stable Diffusion1,4, ElasticDiffusion obtains comparable FID (228.87 vs 230.21) and CLIP scores (26.07 vs 28.06) as SDXL when generating images at \\(1024 \\times 1024\\), which is the native resolution of SDXL." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.09, + 0.643, + 0.107 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.116, + 0.895, + 0.493 + ], + "angle": 0, + "content": "Diffusion Models have been widely adopted for their high-quality outputs in generative tasks [8, 11, 15, 20, 22, 35, 37, 39, 43, 55, 58]. These models involve iterative decoding with many steps leading to high compute and memory requirements. Recent work addressed these issues by devising faster sampling strategies [26, 48], hierarchical models [19, 37, 43], progressive training at increasing resolutions [33, 35, 39], and two-stage models [35, 39, 42, 53]. Stable Diffusion (SD) [39] trains a variational auto-encoder to compress images into a low-dimensional \\(64 \\times 64\\) latent space and trains a diffusion model on this latent space. To train for higher resolutions, models are initially trained at a \\(256 \\times 256\\) resolution, before fine-tuning them at \\(512 \\times 512\\) and \\(1024 \\times 1024\\) in the case of SDXL [35]. SD is one of the few large-scale diffusion models that released trained parameters, making it the building block for many subsequent work [5, 16, 40, 44, 52, 53]. However, these models, including SD, are confined to specific resolutions and do not generalize well to aspect ratios unseen during training. Interestingly, despite being presented with a \\(256 \\times 256\\) resolution during their early training stages, both SD and SDXL fail to generate realistic images at this resolution after being fine-tuned for larger outputs. ElasticDiffusion enables high quality generation at unseen resolutions including re-enabling consistent high quality outputs for SD at \\(256 \\times 256\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.494, + 0.895, + 0.675 + ], + "angle": 0, + "content": "Mixture of diffusers [63] and MultiDiffusion [3] generate panoramic images using a pre-trained diffusion model by generating overlapping crops and combining the generation signal of the overlapping regions. Blending multiple generation signals spatially has been of interest in many previous work [1, 63]. SDXL [35] and Any-Size-Diffusion [62] fine-tune a pre-trained SD on a fixed set of resolutions. A concurrent work [14] uses dilated convolution kernels to reduce SD artifacts when generating non-square images. ElasticDiffusion extends a pre-trained SD to generate images of various sizes at a constant memory requirement and without additional training, all while ensuring global coherence." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.675, + 0.897, + 0.903 + ], + "angle": 0, + "content": "Guided diffusion models devise strategies to condition image generation based on text and other modalities [2, 2, 6, 18, 23, 25, 32, 39, 57, 59]. Classifier guidance [32] uses a pre-trained classifier on noisy images to guide the generation process. Classifier-free guidance [18] eliminates the need for a pretrained classifier but requires training a conditional diffusion model, limiting its applications. Universal Guidance [2] applies a pre-trained classifier on the noise-free images produced by DDIM [48], bypassing training on noisy images. StableSR [53] uses LoRA [21] to condition the generation on low-resolution inputs to achieve superb image super-resolution. Inspired by this, we propose Reduce-Resolution Guidance (Sec. 4) to constrain the generation of high-res images using a lower resolution version, substantially reducing artifacts without extra training." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.957 + ], + "angle": 0, + "content": "6604" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.078, + 0.09, + 0.362, + 0.108 + ], + "angle": 0, + "content": "3. Background: Diffusion Models" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.115, + 0.47, + 0.207 + ], + "angle": 0, + "content": "A conditional diffusion model \\(\\epsilon_{\\theta} \\colon \\mathcal{X} \\times \\mathcal{C} \\to \\mathcal{X}\\) predicts a less noisy version of the input image \\(x \\in \\mathbb{R}^{H \\times W \\times 3}\\), conditioned on variable \\(c \\in \\mathbb{R}^D\\) (e.g. a text embedding). Starting with \\(x_{T} \\sim \\mathcal{N}(0, \\mathbf{I})\\), the reverse diffusion process progressively denoise \\(x_{T}\\) over \\(T\\) steps to generate a realistic image \\(x_0\\) that conforms to the input condition \\(c\\) through:" + }, + { + "type": "equation", + "bbox": [ + 0.12, + 0.218, + 0.426, + 0.234 + ], + "angle": 0, + "content": "\\[\nx _ {t - 1} = \\epsilon_ {\\theta} (x _ {t}, c) \\quad \\text {f o r} t = T, T - 1, \\dots , 1,\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.244, + 0.469, + 0.44 + ], + "angle": 0, + "content": "where \\(\\epsilon_{\\theta}\\) is the denoising network. Standard diffusion models operate in the pixel space, but others like Latent Diffusion Models [39] instead operate on a latent image encoding space to reduce memory footprints. A U-Net architecture is commonly used to implement the denoising network and is typically trained on images at a fixed resolution \\(H\\times W\\) [8, 39, 43]. The convolutional architecture of a U-Net allows for inputs and outputs of arbitrary spatial dimensions. In order to generate an image of a different size \\(\\bar{H}\\times \\bar{W}\\) at inference time, one can sample an initial noise variable \\(\\bar{x}_T\\in \\mathbb{R}^{\\bar{H}\\times \\bar{W}\\times 3}\\) and follow the same diffusion process. However, we find that this works poorly in practice, resulting in a significant degradation in output quality." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.441, + 0.469, + 0.486 + ], + "angle": 0, + "content": "Denoising Diffusion Implicit Models (DDIMs) introduce a faster non-Markovian sampling strategy, bypassing denoising steps. The reverse diffusion step in DDIM is:" + }, + { + "type": "equation", + "bbox": [ + 0.087, + 0.506, + 0.47, + 0.584 + ], + "angle": 0, + "content": "\\[\nx _ {t - 1} = \\sqrt {\\bar {\\alpha} _ {t - 1}} \\underbrace {\\left(x _ {t} - \\sqrt {1 - \\bar {\\alpha} _ {t} \\epsilon_ {\\theta} ^ {(t)} \\left(x _ {t}\\right)}\\right)} _ {\\text {p r e d i c t e d} \\hat {x} _ {0} ^ {t}} + \\underbrace {\\sqrt {1 - \\bar {\\alpha} _ {t - 1}} ,} _ {\\text {d i r e c t i o n p o i n t i n g t o} x _ {t}}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.593, + 0.469, + 0.669 + ], + "angle": 0, + "content": "where \\(\\bar{\\alpha}_t = \\prod_{i=1}^t 1 - \\beta_i\\) is a cumulative product of the noise levels using a predetermined variance schedule \\(\\beta\\). \\(\\hat{x}_0^t\\) is a noise-free estimation of \\(x_t\\) obtained by subtracting the predicted noise scaled for step \\(t\\). We assume a deterministic sampling process in the DDIM formula [48] for simplicity." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.669, + 0.469, + 0.7 + ], + "angle": 0, + "content": "Classifier-Free Guidance updates the reverse diffusion process to condition it on a given condition \\( c \\) as:" + }, + { + "type": "equation", + "bbox": [ + 0.085, + 0.721, + 0.47, + 0.768 + ], + "angle": 0, + "content": "\\[\n\\hat {\\epsilon} _ {\\theta} ^ {(t)} (x _ {t}) = \\underbrace {\\epsilon_ {\\theta} ^ {(t)} (x _ {t})} _ {\\text {u n c o n d i t i o n a l}} + (1 + w) \\cdot \\underbrace {\\left(\\epsilon_ {\\theta} ^ {(t)} (x _ {t} , c) - \\epsilon_ {\\theta} ^ {(t)} (x _ {t})\\right)} _ {\\Delta_ {\\mathcal {C}} (x, c): \\text {c l a s s d i r e c t i o n s c o r e}}. \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.776, + 0.469, + 0.839 + ], + "angle": 0, + "content": "\\(\\epsilon_{\\theta}(x,c)\\) is a pretrained conditional diffusion model, and \\(w\\) is a scaling factor. The difference between the conditional \\(\\epsilon_{\\theta}(x,c)\\) and unconditional \\(\\epsilon_{\\theta}(x)\\) scores, denoted as class direction score, gives the guidance direction towards \\(c\\)." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.845, + 0.236, + 0.861 + ], + "angle": 0, + "content": "4. ElasticDiffusion" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.469, + 0.902 + ], + "angle": 0, + "content": "The aim of this work is to develop a method capable of synthesizing images at arbitrary size \\(\\bar{H} \\times \\bar{W}\\), and conforming" + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.091, + 0.892, + 0.213 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.216, + 0.892, + 0.258 + ], + "angle": 0, + "content": "Figure 2. PCA of diffusion scores: class-direction score (top) dictates global content by clustering on semantic parts, while the unconditional score (bottom) lacks pixel correlations." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.266, + 0.892, + 0.825 + ], + "angle": 0, + "content": "to a global condition \\( c \\) using a pre-trained diffusion model that is limited at inference to its training resolution \\( H \\times W \\). To achieve this, we observe a main insight with respect to the diffusion scores in Eq. (2) that we visualize in Fig. 2. The class direction score \\( (\\Delta_c) \\) primarily dictates the image's overall composition by showing clustering along semantic regions (e.g. hair) and edges. Thus, upscaling this score maintains global content in higher resolutions. In contrast, the unconditional score lacks inter-pixel correlation, suggesting a localized influence in enhancing detail at the pixel level. Thus, computing it requires a pixel-level precision. We study the global and local behaviour of these scores in Supp. Sec. 2.1. Our intuition behind this observation is that latent pixels, specifically in early diffusion steps where noise is high, display weak and short-range correlations. This leads the unconditional score to focus on local patterns due to these limited pixel interactions. Meanwhile, as \\( \\Delta_c \\) represents a latent direction towards a specified global condition, it exerts a global influence by guiding clusters of pixels with similar directions towards semantic regions, overriding their initial noisy state. Leveraging this insight, we propose a method to decouple the generation of local and global content during the diffusion process. Specifically, we compute the unconditional score on local patches of size \\( H \\times W \\) while simultaneously resizing a class direction score, originally derived for a reference latent of the dimensions \\( H \\times W \\) as well. This dual strategy facilitates the generation of images at varied sizes using a pretrained diffusion model at its native resolution, all while maintaining the same memory requirement and without further training. We first present our approach for computing the unconditional score (Sec. 4.1). We then detail our method to estimate (Sec. 4.2) and upscale the resolution (Sec. 4.3) of the class direction score. Finally, we combine the two estimated scores with a novel guidance strategy to generate images at arbitrary sizes (Sec. 4.4). The generation process of ElasticDiffusion is illustrated in Fig. 3." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.834, + 0.809, + 0.85 + ], + "angle": 0, + "content": "4.1. Estimating the Unconditional Score" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.856, + 0.892, + 0.902 + ], + "angle": 0, + "content": "Building upon previous work [3, 53, 63], we estimate the unconditional score for a large latent signal \\(\\bar{x}_t\\in \\mathbb{R}^{\\bar{H}\\times \\bar{W}\\times 3}\\) by concatenating scores derived from local patches. Specif" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6605" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.119, + 0.089, + 0.852, + 0.419 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.429, + 0.895, + 0.473 + ], + "angle": 0, + "content": "Figure 3. Illustration of ElasticDiffusion: We generate images at various sizes by generating local and global content separately. For local content, we partition the latent \\(\\bar{x}_t\\) into non-overlapping patches \\(p_k\\), each concatenated with context \\(c_k\\) to estimate their unconditional score. For global content, we downsample \\(\\bar{x}_t\\) to \\(\\mathbf{x}_t\\), pad to a square size \\((\\hat{\\mathbf{x}}_t)\\), compute class-direction score \\((\\Delta_c)\\), and upscale to match \\(\\bar{x}_t\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.496, + 0.473, + 0.75 + ], + "angle": 0, + "content": "ically, we divide image \\(\\bar{x}_t\\) to \\(K\\) patches \\(P_{k} \\in \\mathbb{R}^{H \\times W \\times 3}\\) and compute the score as \\(\\epsilon_{\\theta}(\\bar{x}_t) = \\{\\epsilon_{\\theta}(P_k); \\forall k \\leq K\\}\\). A common challenge encountered with this implementation is discontinuities at boundaries, as illustrated in Fig. 4 (A). To address this, earlier research calculated the diffusion model score on explicitly overlapping patches and averaged their scores in the intersecting regions [3, 53, 63]. While this strategy mitigates discontinuities, it requires large overlap between patches, thereby substantially increasing inference time and blurring details. To speed up the process, we introduce a more effective method that enhances boundary transitions in local patches by incorporating contextual information from adjacent patches, thus negating the need for signal averaging in overlapping areas. Specifically, we select patches smaller than the full size \\(p_{k} \\in \\mathbb{R}^{h \\times w \\times 3}\\) with \\(h < H\\) and \\(w < W\\), and concatenate them with context pixels from adjacent patches, denoted as \\(c_{k} \\in \\mathbb{R}^{(H - h) \\times (W - w) \\times 3}\\), to compute the diffusion model unconditional score \\(\\mathbf{S}_{u}\\) as:" + }, + { + "type": "equation", + "bbox": [ + 0.135, + 0.765, + 0.47, + 0.781 + ], + "angle": 0, + "content": "\\[\n\\tilde {\\epsilon} _ {\\theta} \\left(x _ {t}\\right) = \\left\\{\\epsilon_ {\\theta} \\left(p _ {k} \\mid c _ {k}\\right) \\mid \\bar {x} _ {t} = \\left\\{p _ {k}; \\forall_ {k \\leq K} \\right\\} \\right\\}, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.795, + 0.471, + 0.903 + ], + "angle": 0, + "content": "where \\( p_i \\) is a local patch and \\( c_i \\) are the context pixels surrounding it. This substantially increases the efficiency of the process. For instance, to generate an image of size \\( 1024 \\times 1024 \\), previous methods [3, 53, 63] used \\( 87.5\\% \\) overlap between adjacent patches, necessitating 81 forward diffusion calls per decoding step. In comparison, ElasticDiffusion achieves comparable results with only 9 forward calls," + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.497, + 0.892, + 0.543 + ], + "angle": 0, + "content": "as demonstrated in Fig. 4 (D). Employing the same number of calls, previous techniques result in obvious boundary discontinuities, as depicted in Fig. 4 (B)." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.551, + 0.818, + 0.567 + ], + "angle": 0, + "content": "4.2. Estimating the Class Direction Score" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.575, + 0.892, + 0.742 + ], + "angle": 0, + "content": "A simple approach to estimate a class direction score of an intermediate latent signal \\(\\bar{x}_t \\in \\mathbb{R}^{\\bar{H} \\times \\bar{W} \\times 3}\\) is to upsample the score from a reference latent \\(x_t \\in \\mathbb{R}^{N \\times M \\times 3}\\) to \\(\\bar{H} \\times \\bar{W}\\). This is possible due to our observation that the class direction score represents a latent direction that can be shared between nearby pixels. We validate this observation empirically in Supplementary. We choose \\(N < H\\) and \\(M < W\\) such that \\(N \\times M\\) is as close as possible to \\(H \\times W\\) and \\(\\frac{\\bar{H}}{W} = \\frac{N}{M}\\). This is important to maintain the aspect ratio and prevent stretching the global content. Formally, we compute the class direction score \\(\\mathbf{S}_d\\) as:" + }, + { + "type": "equation", + "bbox": [ + 0.547, + 0.751, + 0.845, + 0.785 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathbf {x} _ {t} \\leftarrow \\text {D o w n s a m p l e} (\\bar {x} _ {t}, N \\times M), \\\\ \\Delta_ {\\mathcal {C}} (\\bar {x} _ {t}, c) = \\operatorname {U p s a m p l e} \\left(\\Delta_ {\\mathcal {C}} \\left(\\mathbf {x} _ {t}, c\\right), \\bar {H} \\times \\bar {W}\\right), \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.868, + 0.766, + 0.89, + 0.779 + ], + "angle": 0, + "content": "(4)" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.795, + 0.892, + 0.901 + ], + "angle": 0, + "content": "where \\(\\Delta_{\\mathcal{C}}(.)\\) is the class direction score from Eq. (2), Downsample and Upsample are downsampling and upsampling operations. We use a nearest-neighbors approach to prevent altering the statistics of the latent signal. In order to maintain the input to the diffusion models at the size \\(H\\times W\\) we dynamically pad the downsampled latent \\(\\mathbf{x}_t\\) using a random background with a constant color." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6606" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.091, + 0.09, + 0.27, + 0.155 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.276, + 0.091, + 0.458, + 0.155 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.091, + 0.157, + 0.269, + 0.22 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.276, + 0.156, + 0.458, + 0.22 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.091, + 0.228, + 0.269, + 0.291 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.276, + 0.227, + 0.458, + 0.291 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.091, + 0.298, + 0.269, + 0.36 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.157, + 0.362, + 0.219, + 0.37 + ], + "angle": 0, + "content": "Intermediate step" + }, + { + "type": "image", + "bbox": [ + 0.276, + 0.296, + 0.458, + 0.36 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.329, + 0.362, + 0.391, + 0.371 + ], + "angle": 0, + "content": "Generated Image" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.382, + 0.47, + 0.466 + ], + "angle": 0, + "content": "Figure 4. Comparing strategies for calculating diffusion model score on a local patch. No overlap between adjacent patches (A) leads to discontinuities at the boundaries. Strategies (B) and (C), explicitly overlap nearby patches, necessitating substantial overlap to be effective. Our implicit overlapping method (D) achieves superior results with computational demand similar to (B)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.475, + 0.469, + 0.55 + ], + "angle": 0, + "content": "This encourages the model to concentrate content generation within the center area. We then crop the extended parts from the predicted noise to the target image resolution \\( N \\times M \\). Formally we modify the forward call for the reverse diffusion step as:" + }, + { + "type": "equation", + "bbox": [ + 0.148, + 0.562, + 0.424, + 0.58 + ], + "angle": 0, + "content": "\\[\n\\dot {\\mathbf {x}} _ {t} \\leftarrow \\operatorname {P a d} \\left(\\mathbf {x} _ {t}, \\mathcal {A} _ {t} \\sqrt {\\bar {\\alpha} _ {t}} + \\sqrt {1 - \\bar {\\alpha} _ {t}} \\cdot \\mathcal {Z} _ {t}\\right),\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.106, + 0.581, + 0.469, + 0.598 + ], + "angle": 0, + "content": "\\[\n\\epsilon_ {\\theta} \\left(\\mathbf {x} _ {t}, c\\right) = \\operatorname {C r o p} \\left(\\epsilon_ {\\theta} \\left(\\hat {\\mathbf {x}} _ {t}, c\\right), N \\times M\\right), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.61, + 0.469, + 0.731 + ], + "angle": 0, + "content": "where \\(\\mathcal{Z}_t\\sim \\mathcal{N}(0,I)\\) represents the injected Gaussian noise at each step, and \\(\\mathcal{A}_t\\) represents a background image of size \\((H - N)\\times (W - M)\\) with a constant color value \\(\\mathcal{V}\\) randomly sampled from a uniform distribution. This simple padding operation guarantees that the input to the diffusion model is kept at \\(H\\times W\\), while encouraging the diffusion model to keep the generated content within the cropped \\(N\\times M\\) center that we are interested in." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.742, + 0.342, + 0.756 + ], + "angle": 0, + "content": "4.3. Refined Class Direction Score" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Sharing the class direction score among nearby pixels can result in over-smoothed images. To mitigate this, we propose an iterative resampling technique that increases the resolution of the estimated class direction score by extrapolating missing image components from their surrounding context, following [29]. Our technique involves a gradual enhancement of the class direction score's resolution by estimating and integrating it for newly sampled pixels. Specifically, in each iteration, we replace \\( n\\% \\) of the pixels" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.091, + 0.891, + 0.159 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.167, + 0.892, + 0.223 + ], + "angle": 0, + "content": "Figure 5. The Effect of Reduced-Resolution Guidance (RRG). Higher RRG weights effectively eliminates emerging artifacts albeit at the cost of slightly blurrier outputs. \\(\\delta = 200\\) strikes a good balance. Improvements are more noticeable when zooming in." + }, + { + "type": "image", + "bbox": [ + 0.502, + 0.23, + 0.892, + 0.276 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.5, + 0.277, + 0.891, + 0.303 + ], + "angle": 0, + "content": "Figure 6. Effect of Resampling. Applying more resampling steps noticeably enhances detail in the generated images." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.331, + 0.892, + 0.407 + ], + "angle": 0, + "content": "in \\(\\mathbf{x}_t\\) with newly sampled ones from \\(x_{t}\\) to get an updated version \\(\\tilde{\\mathbf{x}}_t\\). Following each update, the direction score is recalculated and blended with the previously calculated score. Formally, we consider \\(\\mathbf{S}_d^0 = \\mathbf{S}_d\\) and define the iterative resampling as:" + }, + { + "type": "equation", + "bbox": [ + 0.58, + 0.42, + 0.891, + 0.44 + ], + "angle": 0, + "content": "\\[\n\\mathbf {S} _ {d} ^ {r + 1} = \\mathbf {S} _ {d} ^ {r} \\odot m + \\tilde {\\mathbf {S}} _ {d} \\odot (1 - m), \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.452, + 0.892, + 0.605 + ], + "angle": 0, + "content": "where \\(m \\in \\{0,1\\}^{\\bar{H} \\times \\bar{W}}\\) is a mask with a value of 1 at the positions of the newly sampled pixels and 0 elsewhere. \\(\\tilde{\\mathbf{S}}_d\\) represents the recalculated class direction score on \\(\\tilde{\\mathbf{x}}_t\\) as per Eq. (4). This method estimates the class direction score of \\(n\\%\\) new pixels while retaining the information of the previously estimated score, thereby increasing the score's overall resolution. In our experiments, we set \\(n\\) to \\(20\\%\\) and repeat the process \\(R\\) times, depending on the target generation resolution. Fig. 6 demonstrates the effectiveness of our method in enhancing the details of the generated images." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.615, + 0.825, + 0.631 + ], + "angle": 0, + "content": "4.4. Reduced-Resolution Guidance (RRG)" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.638, + 0.892, + 0.806 + ], + "angle": 0, + "content": "We effectively estimate the unconditional score signal and concurrently steer the image generation using the class direction score. However, inaccuracies in unconditional score estimation or fluctuations in local content, especially from distant patches, can lead to artifacts. To enhance image coherence and diminish artifacts, we consider a downsampled version of the latent at each decoding step as a reference and aim to align the decoded latent with it through our reduced-resolution latent update strategy. Specifically, we utilize the noise-free sample \\(\\hat{x}_0^t\\) of the latent \\(x_{t}\\) from Eq. (1) and generate a corresponding noise-free sample \\(\\hat{\\mathbf{x}}_0^t\\) from its downsampled counterpart \\(\\mathbf{x}^t\\) in Eq. (4) as:" + }, + { + "type": "equation", + "bbox": [ + 0.507, + 0.816, + 0.883, + 0.845 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathbf {x}} _ {0} ^ {t} = \\frac {1}{\\sqrt {\\bar {\\alpha} _ {t}}} (\\mathbf {x} _ {t} - \\sqrt {1 - \\bar {\\alpha} _ {t}} \\cdot (\\epsilon_ {\\theta} (\\mathbf {x} _ {t}) + (1 + w) \\cdot \\Delta_ {\\mathcal {C}} (\\mathbf {x} _ {t}, c))),\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.856, + 0.892, + 0.902 + ], + "angle": 0, + "content": "Here, both \\(\\hat{x}_0^t\\) and \\(\\hat{\\mathbf{x}}_0^t\\) corresponds to the same latent update at different resolutions. Due to its smaller dimension, \\(\\hat{\\mathbf{x}}_0^t\\) has a broader context when computing the unconditional" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6607" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.098, + 0.093, + 0.873, + 0.401 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.405, + 0.893, + 0.463 + ], + "angle": 0, + "content": "Figure 7. Qualitative comparison at various resolutions on CelebAHQ faces. ElasticDiffusion consistently generates coherent images at all resolutions. StableDiffusion, and MultiDiffusion produce repeating body parts at higher resolutions, while StableDiffusion-XL fails to maintain its quality at lower resolutions, resulting in noisy outputs at \\(256 \\times 256\\). We exclude MultiDiffusion results at \\(256 \\times 256\\) as it is not designed to produce images at lower resolutions." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.487, + 0.471, + 0.534 + ], + "angle": 0, + "content": "signal. To guide \\( x_{t} \\) with its downsampled reference, we refine the latent \\( x_{t} \\) with the direction that minimizes \\( L2 \\) difference between \\( \\hat{\\mathbf{x}}_0^t \\) and \\( \\hat{x}_0^t \\). Formally," + }, + { + "type": "equation", + "bbox": [ + 0.084, + 0.539, + 0.47, + 0.557 + ], + "angle": 0, + "content": "\\[\n\\bar {x} _ {t - 1} \\leftarrow \\bar {x} _ {t - 1} - \\delta_ {t} \\nabla_ {x _ {t}} | | \\text {U p s a m p l e} \\left(\\hat {\\mathbf {x}} _ {0} ^ {t}, \\bar {H} \\times \\bar {W}\\right) - \\hat {x} _ {0} ^ {t} | |, \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.562, + 0.47, + 0.7 + ], + "angle": 0, + "content": "where \\(\\delta_t\\) represent the weight of the guidance. Since the overall image structure is determined in the early diffusion steps, we start with \\(\\delta_t = 400\\) and follow a cosine scheduler to decrease this weight for later diffusion steps. This scheduling strategy mitigates potential quality degradation from matching the generated image with a lower-resolution version while allowing the model to fill-in higher-resolution details in the later decoding stages. Fig. 5 illustrates how RRG eliminates emerging artifacts." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.709, + 0.21, + 0.727 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.735, + 0.469, + 0.796 + ], + "angle": 0, + "content": "ElasticDiffusion supports generating images across different resolutions and aspect ratios. Our experiments focus on: (1) square images at multiple resolutions, and (2) images with varied aspect ratios and resolutions." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.796, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Datasets. We evaluate the generation of square images on the Multi-Modal CelebAHQ dataset [56], which includes high-resolution square face images accompanied with text descriptions. We evaluate different aspect ratio generation using the LAION-COCO dataset [46], derived from the web-crawled LAION-5B dataset [45], which includes a variety of image aspect ratios and contains landscapes, people," + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.487, + 0.892, + 0.533 + ], + "angle": 0, + "content": "objects, and everyday scenes, each paired with a synthetic caption generated using BLIP [27]. We consider four common aspect ratios: 9:16, 16:9, 3:4 and 4:3." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.535, + 0.893, + 0.687 + ], + "angle": 0, + "content": "Evaluation Metrics. Following prior text-to-image synthesis works [3, 33, 35, 38, 43], we use automatic evaluation metrics Frechet Inception Distance (FID) and CLIP-score. FID [17] measures both the realism and diversity of the generated images by calculating the difference between features of the real and generated images computed using Inception-V3 [49] pretrained on ImageNet [41]. CLIP-score uses a pretrained text-image CLIP model [36] to measure alignment between the generated images and input prompts. We use 10,000 images to compute these scores." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.689, + 0.895, + 0.901 + ], + "angle": 0, + "content": "Baselines. We compare our approach against prior diffusion model generation methods, specifically focusing on Stable Diffusion (SD) and MultiDiffusion (MD). SD follows the standard reverse diffusion process on an image latent space. MD uses a pretrained \\( SD \\) for panoramic image generation, by creating smaller, overlapping patches and averaging the Diffusion Model scores in intersecting areas. For baseline comparisons, we fix the pre-trained diffusion model to \\( SD_{1.4} \\), which is trained to generate images at a resolution of \\( 512 \\times 512 \\). Additionally, we compare our method with SDXL, an enhanced SD model that is three times larger than the standard one. SDXL is trained at a higher resolution of \\( 1024p \\), and fine-tuned on a specific set of aspect ratios with pixel sums close to \\( 1024^2 \\). We exclude" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "6608" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.089, + 0.468, + 0.5 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.513, + 0.47, + 0.571 + ], + "angle": 0, + "content": "Figure 8. Qualitative comparison on various aspect ratios. ElasticDiffusion effectively handles a variety of aspect ratios. In comparison, \\( SD \\) and MultiDiffusion generate unrealistic images, while SDXL outputs exhibit a decline in the perceptual quality." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.579, + 0.471, + 0.656 + ], + "angle": 0, + "content": "the latent refinement module in \\(SDXL\\) and focus our analysis on the base model. Throughout our experiments, we employ a DDIM sampling strategy with 50 steps and use 7.0 for the classifier-free guidance scaling factor. We also ensure consistency in seeds and captions for all baselines." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.665, + 0.262, + 0.683 + ], + "angle": 0, + "content": "5.1. Qualitative Results" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.689, + 0.473, + 0.903 + ], + "angle": 0, + "content": "We show square image generation samples in Fig. 7 generated from the CelebAHQ dataset at various resolutions. Both MultiDiffusion and Stable Diffusion have a tendency to replicate textures and body parts at higher resolutions, resulting in images that lack coherence. At resolutions lower than its training resolution \\(512 \\times 512\\), StableDiffusion aligns poorly with the provided captions and produces unappealing images. SDXL, trained for \\(1024 \\times 1024\\) resolution, has a similar trend of reduced perceptual quality at lower sizes, eventually producing complete noise at a resolution of \\(256 \\times 256\\). In contrast, ElasticDiffusion maintains image coherence across all tested resolutions, and yields results comparable to SDXL at \\(1024 \\times 1024\\), despite using a less powerful base model and lower memory. We excluded" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.089, + 0.892, + 0.147 + ], + "angle": 0, + "content": "Table 1. Quantitative comparison of on CelebA-HQ and LAION-COCO at different resolution. We indicate the best performances in bold, and underline the second best ones. #Calls represent the number of diffusion model calls required at each decoding steps." + }, + { + "type": "table", + "bbox": [ + 0.504, + 0.156, + 0.892, + 0.499 + ], + "angle": 0, + "content": "
Res.MethodCelebA-HQLAION-COCO#CallsMem.
FID ↓CLIP ↑FID ↓CLIP ↑
256 × 256SD1.4258.4320.1454.0621.4327.2GB
SDXL368.0614.40175.8714.60218.5GB
Ours1.4235.2323.8823.7726.3028.6GB
512 × 512SD1.4233.4024.0020.5027.3328.6GB
SDXL240.2021.5742.5825.34221.6GB
768 × 768SD1.4238.8723.4529.8927.01211.1GB
MD1.4240.5622.8229.9827.31508.6GB
SDXL225.4824.2323.3127.88224.5GB
Ours1.4225.8626.6625.7825.93178.6GB
1024 × 1024SD1.4266.0121.9047.0125.70114.7GB
MD1.4264.5721.5537.7026.961628.6GB
SDXL230.2124.6225.5828.06127.5GB
Ours1.4228.8723.7427.7626.07338.6GB
" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.506, + 0.895, + 0.675 + ], + "angle": 0, + "content": "results at \\(512 \\times 512\\) because both our method and Multi-Diffusion generate same outputs as the base Stable Diffusion model. We also omitted the results of MultiDiffusion at \\(256 \\times 256\\) since it is not designed to generate images at sizes smaller than the base model. Fig. 8 presents generated samples at various aspect ratios. We observe a similar trend of pattern repetition and reduced perceptual quality for images generated by the baselines, in contrast to those generated by ElasticDiffusion. Finally, we apply our method using SDXL as the base model to enable Full-HD image generation (\\(1920 \\times 1080\\)) and provide examples in Supp. Sec. 3.3." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.681, + 0.695, + 0.698 + ], + "angle": 0, + "content": "5.2. Quantitative Results" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.704, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Table 1 shows quantitative evaluations. StableDiffusion shows increasing image quality degradation, as indicated by the FID metric when processing images of sizes different from its training resolution \\(512 \\times 512\\). MultiDiffusion slightly improves FID at the expense of a substantially more forward calls. SDXL demonstrates similar declines in quality at resolutions far from its fine-tuning size of \\(1024 \\times 1024\\). ElasticDiffusion, however, improves FID while maintaining a comparable CLIP-score to the base model. ElasticDiffusion also significantly improves the performance for lower resolutions \\(256 \\times 256\\), while achieving comparable results to SDXL at its training resolution \\(1024 \\times 1024\\), with only \\(\\sim 31\\%\\) of the memory requirement for SDXL." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6609" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.089, + 0.472, + 0.146 + ], + "angle": 0, + "content": "Table 2. Quantitative comparison of on LAION-COCO datasets at various aspect ratios (A) and resolutions (R). best performances are in bold, and the second best are underlined. Vertical means the resolution is transposed from H:W to W:H." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.156, + 0.467, + 0.578 + ], + "angle": 0, + "content": "
ARMethodHorizontalVertical
FID ↓CLIP↑FID ↓CLIP↑
3:4384 × 512SD1.438.8624.6317.6626.54
MD1.4----
SDXL104.8421.3368.8422.40
Ours1.435.1024.9115.5027.33
512 × 680SD1.445.5424.9616.8127.33
MD1.443.4424.5619.1326.80
SDXL62.8024.2028.2326.17
Ours1.441.0624.4018.9026.83
768 × 1024SD1.471.0024.0928.8326.30
MD1.454.8925.0226.3526.95
SDXL47.0525.7919.5027.41
Ours1.447.0324.9122.5225.80
9:16288 × 512SD1.423.5024.6924.0124.89
MD1.4----
SDXL121.8317.65112.4118.54
Ours1.423.2325.2622.8626.30
512 × 904SD1.429.8625.3427.4526.01
MD1.426.3525.7026.7025.28
SDXL29.6025.4027.2726.08
Ours1.422.8525.0126.6826.12
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.599, + 0.47, + 0.901 + ], + "angle": 0, + "content": "Table 2 provides an evaluation on a variety of aspect ratios and resolutions for both horizontal and vertical image resolutions. StableDiffusion obtains worse FID score with increasing resolutions, while surprisingly maintaining or improving its CLIP-score. We posit that since CLIP-score quantifies the agreement between the input prompt and the generated image, StableDiffusion benefits from generating repetitive textures and artifacts that align closely with the prompt. For example, a photo of repeated lipsticks might align more with the caption \"lipstick\" despite its poor composition. MultiDiffusion generally enhances image quality but does not achieve satisfactory performance. MultiDiffusion struggles with objects that span the entire image (e.g. Fig. 8). Our method consistently improves FID over the baseline on horizontal and most vertical resolutions while preserving fidelity to the input prompts, thereby attaining comparable or superior CLIP-scores. Remarkably, even with a larger model size and explicit fine-tuning at a similar resolution of \\(768 \\times 1280\\), SDXL only marginally surpasses our method at the \\(768 \\times 1024\\) resolution." + }, + { + "type": "table_caption", + "bbox": [ + 0.508, + 0.089, + 0.884, + 0.105 + ], + "angle": 0, + "content": "Table 3. Ablation analysis of ElasticDiffusion on 500 images." + }, + { + "type": "table", + "bbox": [ + 0.53, + 0.115, + 0.855, + 0.196 + ], + "angle": 0, + "content": "
Model DetailsFID ↓CLIP ↑
ElasticDiffusion133.6725.82
w/o Resampling150.0123.82
w/o RRG150.6424.34
w/o Imp. overlap, w/ exp. overlap141.4225.82
" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.208, + 0.651, + 0.224 + ], + "angle": 0, + "content": "5.3. Ablation study" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.232, + 0.892, + 0.354 + ], + "angle": 0, + "content": "Tab. 3 presents results of our method when excluding key components, demonstrating that each element improves performance. Fig. 4 illustrates the effectiveness of our proposed implicit overlap method in resolving boundary discontinuities at a reduced computational cost. Fig. 5 highlights the effectiveness of Reduced-Resolution Guidance in removing emerging artifacts. Fig. 6 shows the effectiveness of our iterative resampling technique in enhancing details." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.371, + 0.748, + 0.387 + ], + "angle": 0, + "content": "6. Discussion and Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.397, + 0.892, + 0.488 + ], + "angle": 0, + "content": "Experimental results highlight the adaptability and effectiveness of ElasticDiffusion at steering diffusion models to produce an array of resolutions and aspect ratios. ElasticDiffusion requires no fine-tuning, consumes a constant memory footprint, enables both increased and reduced resolutions, and can generate a variety of aspect ratios." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.489, + 0.893, + 0.67 + ], + "angle": 0, + "content": "ElasticDiffusion, however, does have several practical limitations. First, inaccuracies in estimating the global and local signals may result in artifacts. Although we attempt to mitigate artifacts with our Reduced-resolution guidance, it can still generate blurrier outputs. Second, since the global content guidance is initially estimated at the original training resolution of the underlying diffusion model, our method is less effective in generating images at significantly extended sizes beyond the training resolution. In particular, at extreme resolutions beyond 4X, our method produces artifacts and images of a lower perceptual quality. We provide examples of these failure cases in Supp. Sec. 10.4." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.671, + 0.892, + 0.868 + ], + "angle": 0, + "content": "The main insight underpinning our method is a novel reinterpretation of classifier-free guidance in somewhat disentangling both global and local content. Our comprehensive evaluations demonstrate the feasibility of disentangling these signals, yet the full extent of their separation offers an avenue for further exploration in this direction. We hope that our findings inspire future work in investigating the separation of global and local content guidance signals for image synthesis. This separation holds potential for various applications such as selectively manipulating local and global content or enhancing style transfer. Additionally, the rich semantic representation in the class-direction score has the potential for improving discriminative models." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.892, + 0.9 + ], + "angle": 0, + "content": "Acknowledgments. This work was partially funded by NSF Award #2201710." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "6610" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.171 + ], + "angle": 0, + "content": "[1] Omri Avrahami, Dani Lischinski, and Ohad Fried. Blended diffusion for text-driven editing of natural images. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). IEEE, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.173, + 0.47, + 0.228 + ], + "angle": 0, + "content": "[2] Arpit Bansal, Hong-Min Chu, Avi Schwarzschild, Soumyadip Sengupta, Micah Goldblum, Jonas Geiping, and Tom Goldstein. Universal guidance for diffusion models, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.232, + 0.468, + 0.273 + ], + "angle": 0, + "content": "[3] Omer Bar-Tal, Lior Yariv, Yaron Lipman, and Tali Dekel. Multidiffusion: Fusing diffusion paths for controlled image generation. In ICML, 2023. 2, 3, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.276, + 0.468, + 0.33 + ], + "angle": 0, + "content": "[4] Angela Castillo, Jonas Kohler, Juan C. Pérez, Juan Pablo Pérez, Albert Pumarola, Bernard Ghanem, Pablo Arbeláez, and Ali Thabet. Adaptive guidance: Training-free acceleration of conditional diffusion models, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.334, + 0.468, + 0.374 + ], + "angle": 0, + "content": "[5] Duygu Ceylan, Chun-Hao Paul Huang, and Niloy J. Mitra. Pix2video: Video editing using image diffusion. In ICCV, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.378, + 0.468, + 0.42 + ], + "angle": 0, + "content": "[6] Hyungjin Chung, Jeongsol Kim, Michael T. McCann, Marc L. Klasky, and Jong Chul Ye. Diffusion posterior sampling for general noisy inverse problems, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.422, + 0.468, + 0.464 + ], + "angle": 0, + "content": "[7] Jiwoo Chung, Sangeek Hyun, and Jae-Pil Heo. Style injection in diffusion: A training-free approach for adapting large-scale diffusion models for style transfer, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.467, + 0.468, + 0.494 + ], + "angle": 0, + "content": "[8] Prafulla Dhariwal and Alex Nichol. Diffusion models beat gans on image synthesis. In NeurIPS, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.497, + 0.468, + 0.551 + ], + "angle": 0, + "content": "[9] Ruoyi Du, Dongliang Chang, Timothy Hospedales, Yi-Zhe Song, and Zhanyu Ma. Demofusion: Democratising high-resolution image generation with no $$$. In CVPR, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.555, + 0.468, + 0.61 + ], + "angle": 0, + "content": "[10] Alexandros Graikos, Srikar Yellapragada, Minh-Quan Le, Saarthak Kapse, Prateek Prasanna, Joel Saltz, and Dimitris Samaras. Learned representation-guided diffusion models for large-image generation, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.613, + 0.468, + 0.667 + ], + "angle": 0, + "content": "[11] Jiatao Gu, Qingzhe Gao, Shuangfei Zhai, Baoquan Chen, Lingjie Liu, and Josh Susskind. Control3diff: Learning controllable 3d diffusion models from single-view images. In 3DV24, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.671, + 0.468, + 0.74 + ], + "angle": 0, + "content": "[12] Lanqing Guo, Yingqing He, Haoxin Chen, Menghan Xia, Xiaodong Cun, Yufei Wang, Siyu Huang, Yong Zhang, Xintao Wang, Qifeng Chen, Ying Shan, and Bihan Wen. Make a cheap scaling: A self-cascade diffusion model for higher-resolution adaptation, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.743, + 0.468, + 0.797 + ], + "angle": 0, + "content": "[13] Feihong He, Gang Li, Lingyu Si, Leilei Yan, Shimeng Hou, Hongwei Dong, and Fanzhang Li. Cartooniff: Training-free cartoon image generation with diffusion transformer models, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.801, + 0.468, + 0.856 + ], + "angle": 0, + "content": "[14] Yingqing He, Shaoshu Yang, Haoxin Chen, Xiaodong Cun, Menghan Xia, Yong Zhang, Xintao Wang, Ran He, Qifeng Chen, and Ying Shan. Scalecrafter: Tuning-free higher-resolution visual generation with diffusion models, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.859, + 0.468, + 0.901 + ], + "angle": 0, + "content": "[15] Yingqing He, Tianyu Yang, Yong Zhang, Ying Shan, and Qifeng Chen. Latent video diffusion models for high-fidelity long video generation, 2023. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.47, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.135 + ], + "angle": 0, + "content": "[16] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-prompt image editing with cross attention control. In ICLR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.136, + 0.892, + 0.19 + ], + "angle": 0, + "content": "[17] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. In NIPS, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.193, + 0.892, + 0.219 + ], + "angle": 0, + "content": "[18] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.221, + 0.892, + 0.275 + ], + "angle": 0, + "content": "[19] Jonathan Ho, Chitwan Sahara, William Chan, David J. Fleet, Mohammad Norouzi, and Tim Salimans. Cascaded diffusion models for high fidelity image generation. Journal of Machine Learning Research, 23(47):1-33, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.278, + 0.892, + 0.317 + ], + "angle": 0, + "content": "[20] Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J. Fleet. Video diffusion models. In NeurIPS, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.32, + 0.892, + 0.374 + ], + "angle": 0, + "content": "[21] Edward J. Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. In ICLR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.377, + 0.892, + 0.432 + ], + "angle": 0, + "content": "[22] Rongjie Huang, Jiawei Huang, Dongchao Yang, Yi Ren, Luping Liu, Mingze Li, Zhenhui Ye, Jinglin Liu, Xiang Yin, and Zhou Zhao. Make-an-audio: Text-to-audio generation with prompt-enhanced diffusion models. In ICML, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.434, + 0.892, + 0.487 + ], + "angle": 0, + "content": "[23] Yujin Jeong, Wonjeong Ryoo, Seunghyun Lee, Dabin Seo, Wonmin Byeon, Sanggil Kim, and Jinkyu Kim. The power of sound (tpos): Audio reactive video generation with stable diffusion. In ICCV, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.49, + 0.892, + 0.531 + ], + "angle": 0, + "content": "[24] Zhiyu Jin, Xuli Shen, Bin Li, and Xiangyang Xue. Training-free diffusion model adaptation for variable-sized text-to-image synthesis, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.533, + 0.892, + 0.573 + ], + "angle": 0, + "content": "[25] Gwanghyun Kim, Taesung Kwon, and Jong Chul Ye. Diffusionclip: Text-guided diffusion models for robust image manipulation. In CVPR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.575, + 0.892, + 0.602 + ], + "angle": 0, + "content": "[26] Zhifeng Kong and Wei Ping. On fast sampling of diffusion probabilistic models. In ICML, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.604, + 0.892, + 0.645 + ], + "angle": 0, + "content": "[27] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.647, + 0.892, + 0.686 + ], + "angle": 0, + "content": "[28] Jin Liu, Huaibo Huang, Chao Jin, and Ran He. Portrait diffusion: Training-free face stylization with chain-of-painting, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.689, + 0.892, + 0.743 + ], + "angle": 0, + "content": "[29] Andreas Lugmayr, Martin Danelljan, Andres Romero, Fisher Yu, Radu Timofte, and Luc Van Gool. Repaint: Inpainting using denoising diffusion probabilistic models. In CVPR, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.746, + 0.892, + 0.8 + ], + "angle": 0, + "content": "[30] Sicheng Mo, Fangzhou Mu, Kuan Heng Lin, Yanli Liu, Bochen Guan, Yin Li, and Bolei Zhou. Freecontrol: Training-free spatial control of any text-to-image diffusion model with any condition, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.802, + 0.892, + 0.843 + ], + "angle": 0, + "content": "[31] Brian B. Moser, Stanislav Frolov, Federico Raue, Sebastian Palacio, and Andreas Dengel. Dynamic attention-guided diffusion for image super-resolution, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.845, + 0.892, + 0.872 + ], + "angle": 0, + "content": "[32] Alex Nichol and Prafulla Dhariwal. Improved denoising diffusion probabilistic models. In ICML, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.874, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[33] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.514, + 0.957 + ], + "angle": 0, + "content": "6611" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.092, + 0.469, + 0.133 + ], + "angle": 0, + "content": "Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. In ICML, 2022. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.134, + 0.469, + 0.175 + ], + "angle": 0, + "content": "[34] Bo Peng, Xinyuan Chen, Yaohui Wang, Chaochao Lu, and Yu Qiao. Conditionvideo: Training-free condition-guided text-to-video generation, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.177, + 0.469, + 0.232 + ], + "angle": 0, + "content": "[35] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis, 2023. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.233, + 0.469, + 0.314 + ], + "angle": 0, + "content": "[36] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In ICML, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.316, + 0.469, + 0.357 + ], + "angle": 0, + "content": "[37] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.359, + 0.469, + 0.399 + ], + "angle": 0, + "content": "[38] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.401, + 0.469, + 0.44 + ], + "angle": 0, + "content": "[39] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.442, + 0.469, + 0.496 + ], + "angle": 0, + "content": "[40] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In CVPR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.498, + 0.469, + 0.565 + ], + "angle": 0, + "content": "[41] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, Alexander C. Berg, and Li Fei-Fei. Imagenet large scale visual recognition challenge, 2015. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.567, + 0.469, + 0.607 + ], + "angle": 0, + "content": "[42] Chitwan Sahara, Jonathan Ho, William Chan, Tim Salimans, David J. Fleet, and Mohammad Norouzi. Image superresolution via iterative refinement, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.609, + 0.469, + 0.705 + ], + "angle": 0, + "content": "[43] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, Seyed Kamyar Seyed Ghasemipour, Burcu Karagol Ayan, S. Sara Mahdavi, Rapha Gontijo Lopes, Tim Salimans, Jonathan Ho, David J Fleet, and Mohammad Norouzi. Photorealistic text-to-image diffusion models with deep language understanding. In NeurIPS, 2022. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.707, + 0.469, + 0.761 + ], + "angle": 0, + "content": "[44] Axel Sauer, Frederic Boesel, Tim Dockhorn, Andreas Blattmann, Patrick Esser, and Robin Rombach. Fast high-resolution image synthesis with latent adversarial diffusion distillation, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.763, + 0.469, + 0.858 + ], + "angle": 0, + "content": "[45] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, Patrick Schramowski, Srivatsa Kundurthy, Katherine Crowson, Ludwig Schmidt, Robert Kaczmarczyk, and Jenia Jitsev. Laion-5b: An open large-scale dataset for training next generation image-text models, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.86, + 0.469, + 0.9 + ], + "angle": 0, + "content": "[46] Christoph Schuhmann, Andreas A. Köpf, Richard Vencu, Theo Coombes, and Ross Beaumont. Laion coco: 600m synthetic captions from laion2b-en, 2023. 6" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.469, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.147 + ], + "angle": 0, + "content": "[47] Fengyuan Shi, Jiaxi Gu, Hang Xu, Songcen Xu, Wei Zhang, and Limin Wang. Bivdiff: A training-free framework for general-purpose video synthesis via bridging image and video diffusion models, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.149, + 0.892, + 0.177 + ], + "angle": 0, + "content": "[48] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In ICLR, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.179, + 0.892, + 0.232 + ], + "angle": 0, + "content": "[49] Christian Szegedy, Vincent Vanhoucke, Sergey Ioffe, Jonathon Shlens, and Zbigniew Wojna. Rethinking the inception architecture for computer vision. In CVPR, 2015. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.235, + 0.892, + 0.276 + ], + "angle": 0, + "content": "[50] Andrey Voynov, Amir Hertz, Moab Arar, Shlomi Fruchter, and Daniel Cohen-Or. Anylens: A generative diffusion model with any rendering lens, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.278, + 0.892, + 0.332 + ], + "angle": 0, + "content": "[51] Bingyuan Wang, Hengyu Meng, Zeyu Cai, Lanjiong Li, Yue Ma, Qifeng Chen, and Zeyu Wang. Magicscrol: Nontypical aspect-ratio image generation for visual storytelling via multi-layered semantic-aware denoising, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.334, + 0.892, + 0.389 + ], + "angle": 0, + "content": "[52] Fu-Yun Wang, Zhaoyang Huang, Xiaoyu Shi, Weikang Bian, Guanglu Song, Yu Liu, and Hongsheng Li. *Animatelcm: Accelerating the animation of personalized diffusion models and adapters with decoupled consistency learning*, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.391, + 0.892, + 0.432 + ], + "angle": 0, + "content": "[53] Jianyi Wang, Zongsheng Yue, Shangchen Zhou, Kelvin C. K. Chan, and Chen Change Loy. Exploiting diffusion prior for real-world image super-resolution, 2023. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.434, + 0.892, + 0.487 + ], + "angle": 0, + "content": "[54] Lezhong Wang, Jeppe Revall Frisvad, Mark Bo Jensen, and Siavash Arjomand Bigdeli. Stereodiffusion: Training-free stereo image generation using latent diffusion models, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.49, + 0.892, + 0.544 + ], + "angle": 0, + "content": "[55] Daniel Watson, William Chan, Ricardo Martin-Brualla, Jonathan Ho, Andrea Tagliasacchi, and Mohammad Norouzi. Novel view synthesis with diffusion models. In ICLR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.547, + 0.892, + 0.587 + ], + "angle": 0, + "content": "[56] Weihao Xia, Yujiu Yang, Jing-Hao Xue, and Baoyuan Wu. Tedigan: Text-guided diverse face image generation and manipulation, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.59, + 0.892, + 0.631 + ], + "angle": 0, + "content": "[57] Guy Yariv, Itai Gat, Lior Wolf, Yossi Adi, and Idan Schwartz. Audiotoken: Adaptation of text-conditioned diffusion models for audio-to-image generation, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.633, + 0.892, + 0.673 + ], + "angle": 0, + "content": "[58] Ahmet Burak Yildirim, Vedat Baday, Erkut Erdem, Aykut Erdem, and Aysegul Dundar. Inst-inpaint: Instructing to remove objects with diffusion models, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.676, + 0.892, + 0.715 + ], + "angle": 0, + "content": "[59] Jiwen Yu, Yinhuai Wang, Chen Zhao, Bernard Ghanem, and Jian Zhang. Freedom: Training-free energy-guided conditional diffusion model, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.718, + 0.892, + 0.772 + ], + "angle": 0, + "content": "[60] Shen Zhang, Zhaowei Chen, Zhenyu Zhao, Zhenyuan Chen, Yao Tang, Yuhao Chen, Wengang Cao, and Jiajun Liang. Hidiffusion: Unlocking high-resolution creativity and efficiency in low-resolution trained diffusion models, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.774, + 0.892, + 0.814 + ], + "angle": 0, + "content": "[61] Peiang Zhao, Han Li, Ruiyang Jin, and S. Kevin Zhou. Loco: Locally constrained training-free layout-to-image synthesis, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.817, + 0.892, + 0.871 + ], + "angle": 0, + "content": "[62] Qingping Zheng, Yuanfan Guo, Jiankang Deng, Jianhua Han, Ying Li, Songcen Xu, and Hang Xu. Any-sized diffusion: Toward efficient text-driven synthesis for any-size hd images, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.873, + 0.892, + 0.902 + ], + "angle": 0, + "content": "[63] Álvaro Barbero Jiménez. Mixture of diffusers for scene composition and high resolution image generation, 2023. 2, 3, 4" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6612" + } + ] +] \ No newline at end of file diff --git a/2024/ElasticDiffusion_ Training-free Arbitrary Size Image Generation through Global-Local Content Separation/de40da54-b1b7-40fa-9aec-0f4f1bc4f406_origin.pdf b/2024/ElasticDiffusion_ Training-free Arbitrary Size Image Generation through Global-Local Content Separation/de40da54-b1b7-40fa-9aec-0f4f1bc4f406_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..56a7719e24520e230ff9c6af9ddf55222fa895f7 --- /dev/null +++ b/2024/ElasticDiffusion_ Training-free Arbitrary Size Image Generation through Global-Local Content Separation/de40da54-b1b7-40fa-9aec-0f4f1bc4f406_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:233729556bd5309e4ecec7a98f6c7ca509e3cc1668d44170561fbfc39d1b82b6 +size 9979398 diff --git a/2024/ElasticDiffusion_ Training-free Arbitrary Size Image Generation through Global-Local Content Separation/full.md b/2024/ElasticDiffusion_ Training-free Arbitrary Size Image Generation through Global-Local Content Separation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..d32a249cbf705e2551387a4ffd4595c674da1d43 --- /dev/null +++ b/2024/ElasticDiffusion_ Training-free Arbitrary Size Image Generation through Global-Local Content Separation/full.md @@ -0,0 +1,297 @@ +# ElasticDiffusion: Training-free Arbitrary Size Image Generation through Global-Local Content Separation + +Moayed Haji-Ali + +Guha Balakrishnan Rice University + +{mh155, guha, vicenteor}@rice.edu + +Vicente Ordonez + +![](images/6332836812d5c3aa5ab1d905bbb973969a8793b5267a835e2d10f1c30630b170.jpg) +Figure 1. ElasticDiffusion generates high quality images at arbitrary sizes using a pretrained diffusion model trained on a single image size, with equivalent memory footprint and no further training. These results are based on Stable Diffusion1,4, which was trained to generate $512 \times 512$ images. The examples shown in this collage are presented without any image cropping, stretching, or post-processing. + +# Abstract + +Diffusion models have revolutionized image generation in recent years, yet they are still limited to a few sizes and aspect ratios. We propose ElasticDiffusion, a novel training-free decoding method that enables pretrained text-to-image diffusion models to generate images with various sizes. ElasticDiffusion attempts to decouple the generation trajectory of a pretrained model into local and global signals. The local signal controls low-level pixel information and can be estimated on local patches, while the global signal is used to maintain overall structural consistency and is estimated with a reference image. We test our method on CelebA-HQ (faces) and LAION-COCO (objects/indoor/outdoor scenes). Our experiments and qualita + +tive results show superior image coherence quality across aspect ratios compared to MultiDiffusion and the standard decoding strategy of Stable Diffusion. Project Webpage: https://elasticdiffusion.github.io/ + +# 1. Introduction + +Diffusion models are a powerful family of algorithms that achieve remarkable quality and obtain the current state-of-the-art performance on various image synthesis tasks. As is the case with most deep neural networks, diffusion models are typically trained on one or a few image resolutions. For instance, Stable Diffusion (SD) [39], one of the most widely adopted diffusion models, is trained on a square images of size $512 \times 512$ , yet fails to maintain its performance at dif- + +ferent aspect ratios during inference time. In practice, many applications require a wide aspect ratio or portrait mode, such as digital billboards, wearable devices, automotive displays, and any application relying on a computer screen. + +Recent studies address the issue of variable image size in different ways. SDXL [35] and Any-Size-Diffusion [62] explicitly finetune models on images with a range of aspect ratios, which requires extensive computation, a quadratic memory footprint, and larger training data. In addition, this strategy requires a set of resolutions to be specified upfront during training time, while the models still struggle to generalize to new resolutions during inference, often resulting in artifacts. Recent works also show remarkable results in generating panoramic images using pretrained diffusion models by overlapping generated patches into a larger image [3, 63]. These methods work well for landscape images with repetitive patterns. However, their lack of global guidance limits their abilities to generate images of single objects or faces where global structure is important. Recent work [24, 50, 51] aimed at extending pre-trained diffusion models capabilities to others domains, often in a training-free way [4, 7, 13, 28, 30, 34, 47, 54, 61]. Few concurrent work [9, 10, 12, 31, 60] focus on adapting them to higher resolutions, yet they are constrained to square images. + +In this work, we propose ElasticDiffusion, a novel decoding strategy that takes a pretrained diffusion model and generate images at arbitrary sizes during inference using a constant memory footprint. To achieve this, we revisit the guidance mechanism of conditional diffusion models to decouple global and local content generation. Global content controls the high-level aspects of the image, whereas local content adds finer, more granular details. This separation facilitates generating the local content in patches for images of varying sizes, all while being guided with global content that we derive from a reference image at the diffusion model pretraining resolution. This enables the synthesis of images at diverse resolutions and aspect ratios while adhering the diffusion forward calls to the model's initial training size. To aid in this task, we introduce several techniques including an efficient patch fusion method for smooth boundaries, a novel guidance strategy to reduce image artifacts, and a global content resampling technique that amplifies the resolution of diffusion models up to 4X the training size. + +Figure 1 shows a diverse array of images generated with ElasticDiffusion. Several of theses images include a single object or were generated with extreme aspect ratios, showcasing our method's ability to produce coherent images under various sizes. Quantitatively, ElasticDiffusion outperforms baselines across most aspect ratios. More importantly, despite relying on Stable Diffusion1,4, ElasticDiffusion obtains comparable FID (228.87 vs 230.21) and CLIP scores (26.07 vs 28.06) as SDXL when generating images at $1024 \times 1024$ , which is the native resolution of SDXL. + +# 2. Related Work + +Diffusion Models have been widely adopted for their high-quality outputs in generative tasks [8, 11, 15, 20, 22, 35, 37, 39, 43, 55, 58]. These models involve iterative decoding with many steps leading to high compute and memory requirements. Recent work addressed these issues by devising faster sampling strategies [26, 48], hierarchical models [19, 37, 43], progressive training at increasing resolutions [33, 35, 39], and two-stage models [35, 39, 42, 53]. Stable Diffusion (SD) [39] trains a variational auto-encoder to compress images into a low-dimensional $64 \times 64$ latent space and trains a diffusion model on this latent space. To train for higher resolutions, models are initially trained at a $256 \times 256$ resolution, before fine-tuning them at $512 \times 512$ and $1024 \times 1024$ in the case of SDXL [35]. SD is one of the few large-scale diffusion models that released trained parameters, making it the building block for many subsequent work [5, 16, 40, 44, 52, 53]. However, these models, including SD, are confined to specific resolutions and do not generalize well to aspect ratios unseen during training. Interestingly, despite being presented with a $256 \times 256$ resolution during their early training stages, both SD and SDXL fail to generate realistic images at this resolution after being fine-tuned for larger outputs. ElasticDiffusion enables high quality generation at unseen resolutions including re-enabling consistent high quality outputs for SD at $256 \times 256$ . + +Mixture of diffusers [63] and MultiDiffusion [3] generate panoramic images using a pre-trained diffusion model by generating overlapping crops and combining the generation signal of the overlapping regions. Blending multiple generation signals spatially has been of interest in many previous work [1, 63]. SDXL [35] and Any-Size-Diffusion [62] fine-tune a pre-trained SD on a fixed set of resolutions. A concurrent work [14] uses dilated convolution kernels to reduce SD artifacts when generating non-square images. ElasticDiffusion extends a pre-trained SD to generate images of various sizes at a constant memory requirement and without additional training, all while ensuring global coherence. + +Guided diffusion models devise strategies to condition image generation based on text and other modalities [2, 2, 6, 18, 23, 25, 32, 39, 57, 59]. Classifier guidance [32] uses a pre-trained classifier on noisy images to guide the generation process. Classifier-free guidance [18] eliminates the need for a pretrained classifier but requires training a conditional diffusion model, limiting its applications. Universal Guidance [2] applies a pre-trained classifier on the noise-free images produced by DDIM [48], bypassing training on noisy images. StableSR [53] uses LoRA [21] to condition the generation on low-resolution inputs to achieve superb image super-resolution. Inspired by this, we propose Reduce-Resolution Guidance (Sec. 4) to constrain the generation of high-res images using a lower resolution version, substantially reducing artifacts without extra training. + +# 3. Background: Diffusion Models + +A conditional diffusion model $\epsilon_{\theta} \colon \mathcal{X} \times \mathcal{C} \to \mathcal{X}$ predicts a less noisy version of the input image $x \in \mathbb{R}^{H \times W \times 3}$ , conditioned on variable $c \in \mathbb{R}^D$ (e.g. a text embedding). Starting with $x_{T} \sim \mathcal{N}(0, \mathbf{I})$ , the reverse diffusion process progressively denoise $x_{T}$ over $T$ steps to generate a realistic image $x_0$ that conforms to the input condition $c$ through: + +$$ +x _ {t - 1} = \epsilon_ {\theta} (x _ {t}, c) \quad \text {f o r} t = T, T - 1, \dots , 1, +$$ + +where $\epsilon_{\theta}$ is the denoising network. Standard diffusion models operate in the pixel space, but others like Latent Diffusion Models [39] instead operate on a latent image encoding space to reduce memory footprints. A U-Net architecture is commonly used to implement the denoising network and is typically trained on images at a fixed resolution $H\times W$ [8, 39, 43]. The convolutional architecture of a U-Net allows for inputs and outputs of arbitrary spatial dimensions. In order to generate an image of a different size $\bar{H}\times \bar{W}$ at inference time, one can sample an initial noise variable $\bar{x}_T\in \mathbb{R}^{\bar{H}\times \bar{W}\times 3}$ and follow the same diffusion process. However, we find that this works poorly in practice, resulting in a significant degradation in output quality. + +Denoising Diffusion Implicit Models (DDIMs) introduce a faster non-Markovian sampling strategy, bypassing denoising steps. The reverse diffusion step in DDIM is: + +$$ +x _ {t - 1} = \sqrt {\bar {\alpha} _ {t - 1}} \underbrace {\left(x _ {t} - \sqrt {1 - \bar {\alpha} _ {t} \epsilon_ {\theta} ^ {(t)} \left(x _ {t}\right)}\right)} _ {\text {p r e d i c t e d} \hat {x} _ {0} ^ {t}} + \underbrace {\sqrt {1 - \bar {\alpha} _ {t - 1}} ,} _ {\text {d i r e c t i o n p o i n t i n g t o} x _ {t}}, \tag {1} +$$ + +where $\bar{\alpha}_t = \prod_{i=1}^t 1 - \beta_i$ is a cumulative product of the noise levels using a predetermined variance schedule $\beta$ . $\hat{x}_0^t$ is a noise-free estimation of $x_t$ obtained by subtracting the predicted noise scaled for step $t$ . We assume a deterministic sampling process in the DDIM formula [48] for simplicity. + +Classifier-Free Guidance updates the reverse diffusion process to condition it on a given condition $c$ as: + +$$ +\hat {\epsilon} _ {\theta} ^ {(t)} (x _ {t}) = \underbrace {\epsilon_ {\theta} ^ {(t)} (x _ {t})} _ {\text {u n c o n d i t i o n a l}} + (1 + w) \cdot \underbrace {\left(\epsilon_ {\theta} ^ {(t)} (x _ {t} , c) - \epsilon_ {\theta} ^ {(t)} (x _ {t})\right)} _ {\Delta_ {\mathcal {C}} (x, c): \text {c l a s s d i r e c t i o n s c o r e}}. \tag {2} +$$ + +$\epsilon_{\theta}(x,c)$ is a pretrained conditional diffusion model, and $w$ is a scaling factor. The difference between the conditional $\epsilon_{\theta}(x,c)$ and unconditional $\epsilon_{\theta}(x)$ scores, denoted as class direction score, gives the guidance direction towards $c$ . + +# 4. ElasticDiffusion + +The aim of this work is to develop a method capable of synthesizing images at arbitrary size $\bar{H} \times \bar{W}$ , and conforming + +![](images/2170657f1a26f15a373285c892f64f17278698b15e753998df65798b2b4c97dd.jpg) +Figure 2. PCA of diffusion scores: class-direction score (top) dictates global content by clustering on semantic parts, while the unconditional score (bottom) lacks pixel correlations. + +to a global condition $c$ using a pre-trained diffusion model that is limited at inference to its training resolution $H \times W$ . To achieve this, we observe a main insight with respect to the diffusion scores in Eq. (2) that we visualize in Fig. 2. The class direction score $(\Delta_c)$ primarily dictates the image's overall composition by showing clustering along semantic regions (e.g. hair) and edges. Thus, upscaling this score maintains global content in higher resolutions. In contrast, the unconditional score lacks inter-pixel correlation, suggesting a localized influence in enhancing detail at the pixel level. Thus, computing it requires a pixel-level precision. We study the global and local behaviour of these scores in Supp. Sec. 2.1. Our intuition behind this observation is that latent pixels, specifically in early diffusion steps where noise is high, display weak and short-range correlations. This leads the unconditional score to focus on local patterns due to these limited pixel interactions. Meanwhile, as $\Delta_c$ represents a latent direction towards a specified global condition, it exerts a global influence by guiding clusters of pixels with similar directions towards semantic regions, overriding their initial noisy state. Leveraging this insight, we propose a method to decouple the generation of local and global content during the diffusion process. Specifically, we compute the unconditional score on local patches of size $H \times W$ while simultaneously resizing a class direction score, originally derived for a reference latent of the dimensions $H \times W$ as well. This dual strategy facilitates the generation of images at varied sizes using a pretrained diffusion model at its native resolution, all while maintaining the same memory requirement and without further training. We first present our approach for computing the unconditional score (Sec. 4.1). We then detail our method to estimate (Sec. 4.2) and upscale the resolution (Sec. 4.3) of the class direction score. Finally, we combine the two estimated scores with a novel guidance strategy to generate images at arbitrary sizes (Sec. 4.4). The generation process of ElasticDiffusion is illustrated in Fig. 3. + +# 4.1. Estimating the Unconditional Score + +Building upon previous work [3, 53, 63], we estimate the unconditional score for a large latent signal $\bar{x}_t\in \mathbb{R}^{\bar{H}\times \bar{W}\times 3}$ by concatenating scores derived from local patches. Specif + +![](images/ccc04c596e6212ceff276749c13a60680845a2627e4d227da6bd28139b0c0b42.jpg) +Figure 3. Illustration of ElasticDiffusion: We generate images at various sizes by generating local and global content separately. For local content, we partition the latent $\bar{x}_t$ into non-overlapping patches $p_k$ , each concatenated with context $c_k$ to estimate their unconditional score. For global content, we downsample $\bar{x}_t$ to $\mathbf{x}_t$ , pad to a square size $(\hat{\mathbf{x}}_t)$ , compute class-direction score $(\Delta_c)$ , and upscale to match $\bar{x}_t$ . + +ically, we divide image $\bar{x}_t$ to $K$ patches $P_{k} \in \mathbb{R}^{H \times W \times 3}$ and compute the score as $\epsilon_{\theta}(\bar{x}_t) = \{\epsilon_{\theta}(P_k); \forall k \leq K\}$ . A common challenge encountered with this implementation is discontinuities at boundaries, as illustrated in Fig. 4 (A). To address this, earlier research calculated the diffusion model score on explicitly overlapping patches and averaged their scores in the intersecting regions [3, 53, 63]. While this strategy mitigates discontinuities, it requires large overlap between patches, thereby substantially increasing inference time and blurring details. To speed up the process, we introduce a more effective method that enhances boundary transitions in local patches by incorporating contextual information from adjacent patches, thus negating the need for signal averaging in overlapping areas. Specifically, we select patches smaller than the full size $p_{k} \in \mathbb{R}^{h \times w \times 3}$ with $h < H$ and $w < W$ , and concatenate them with context pixels from adjacent patches, denoted as $c_{k} \in \mathbb{R}^{(H - h) \times (W - w) \times 3}$ , to compute the diffusion model unconditional score $\mathbf{S}_{u}$ as: + +$$ +\tilde {\epsilon} _ {\theta} \left(x _ {t}\right) = \left\{\epsilon_ {\theta} \left(p _ {k} \mid c _ {k}\right) \mid \bar {x} _ {t} = \left\{p _ {k}; \forall_ {k \leq K} \right\} \right\}, \tag {3} +$$ + +where $p_i$ is a local patch and $c_i$ are the context pixels surrounding it. This substantially increases the efficiency of the process. For instance, to generate an image of size $1024 \times 1024$ , previous methods [3, 53, 63] used $87.5\%$ overlap between adjacent patches, necessitating 81 forward diffusion calls per decoding step. In comparison, ElasticDiffusion achieves comparable results with only 9 forward calls, + +as demonstrated in Fig. 4 (D). Employing the same number of calls, previous techniques result in obvious boundary discontinuities, as depicted in Fig. 4 (B). + +# 4.2. Estimating the Class Direction Score + +A simple approach to estimate a class direction score of an intermediate latent signal $\bar{x}_t \in \mathbb{R}^{\bar{H} \times \bar{W} \times 3}$ is to upsample the score from a reference latent $x_t \in \mathbb{R}^{N \times M \times 3}$ to $\bar{H} \times \bar{W}$ . This is possible due to our observation that the class direction score represents a latent direction that can be shared between nearby pixels. We validate this observation empirically in Supplementary. We choose $N < H$ and $M < W$ such that $N \times M$ is as close as possible to $H \times W$ and $\frac{\bar{H}}{W} = \frac{N}{M}$ . This is important to maintain the aspect ratio and prevent stretching the global content. Formally, we compute the class direction score $\mathbf{S}_d$ as: + +$$ +\begin{array}{l} \mathbf {x} _ {t} \leftarrow \text {D o w n s a m p l e} (\bar {x} _ {t}, N \times M), \\ \Delta_ {\mathcal {C}} (\bar {x} _ {t}, c) = \operatorname {U p s a m p l e} \left(\Delta_ {\mathcal {C}} \left(\mathbf {x} _ {t}, c\right), \bar {H} \times \bar {W}\right), \\ \end{array} +$$ + +(4) + +where $\Delta_{\mathcal{C}}(.)$ is the class direction score from Eq. (2), Downsample and Upsample are downsampling and upsampling operations. We use a nearest-neighbors approach to prevent altering the statistics of the latent signal. In order to maintain the input to the diffusion models at the size $H\times W$ we dynamically pad the downsampled latent $\mathbf{x}_t$ using a random background with a constant color. + +![](images/33797549b612d2d3da94f381f50662943123a329d503c4b31cf3843b6a85e416.jpg) + +![](images/3469265d4222a88721addc3eb441fe98dd1f99cb17554f84889c04b4a31b6bba.jpg) + +![](images/5487e88e0556e1adb268bd3e0d86267314fc29918e6a34d0891cc14fe3629a50.jpg) + +![](images/268bb81e8397b84389674ad6e464b572f4b5659ae0bdb4274b8f5ad1f87e057a.jpg) + +![](images/366814e38e08459d15ca539142fbb5864edc21686a6e85cd7ae51e9236dd3cc3.jpg) + +![](images/04299a38c6bfe4e4aa1b6da90de60d081fdf9e6afb77e272fe224f020dc99136.jpg) + +![](images/c91b5e67b6f85819184ca6bad9e0ed4de0f089a461e40f59fafae3a36ddd5150.jpg) +Intermediate step +Figure 4. Comparing strategies for calculating diffusion model score on a local patch. No overlap between adjacent patches (A) leads to discontinuities at the boundaries. Strategies (B) and (C), explicitly overlap nearby patches, necessitating substantial overlap to be effective. Our implicit overlapping method (D) achieves superior results with computational demand similar to (B). + +![](images/974809efb764c7055cac428db041fcf6e0d341f2c45eab564f16e7ed4255538d.jpg) +Generated Image + +This encourages the model to concentrate content generation within the center area. We then crop the extended parts from the predicted noise to the target image resolution $N \times M$ . Formally we modify the forward call for the reverse diffusion step as: + +$$ +\dot {\mathbf {x}} _ {t} \leftarrow \operatorname {P a d} \left(\mathbf {x} _ {t}, \mathcal {A} _ {t} \sqrt {\bar {\alpha} _ {t}} + \sqrt {1 - \bar {\alpha} _ {t}} \cdot \mathcal {Z} _ {t}\right), +$$ + +$$ +\epsilon_ {\theta} \left(\mathbf {x} _ {t}, c\right) = \operatorname {C r o p} \left(\epsilon_ {\theta} \left(\hat {\mathbf {x}} _ {t}, c\right), N \times M\right), \tag {5} +$$ + +where $\mathcal{Z}_t\sim \mathcal{N}(0,I)$ represents the injected Gaussian noise at each step, and $\mathcal{A}_t$ represents a background image of size $(H - N)\times (W - M)$ with a constant color value $\mathcal{V}$ randomly sampled from a uniform distribution. This simple padding operation guarantees that the input to the diffusion model is kept at $H\times W$ , while encouraging the diffusion model to keep the generated content within the cropped $N\times M$ center that we are interested in. + +# 4.3. Refined Class Direction Score + +Sharing the class direction score among nearby pixels can result in over-smoothed images. To mitigate this, we propose an iterative resampling technique that increases the resolution of the estimated class direction score by extrapolating missing image components from their surrounding context, following [29]. Our technique involves a gradual enhancement of the class direction score's resolution by estimating and integrating it for newly sampled pixels. Specifically, in each iteration, we replace $n\%$ of the pixels + +![](images/f1eecfec9c801373a03fa89b89c30f1b9f0ed689a4b8c335368fe9440d1c942b.jpg) + +![](images/4c3a5bc0c419ed0226ee95dc439211064e8c2d0018f3862702e4231afca6093b.jpg) +Figure 5. The Effect of Reduced-Resolution Guidance (RRG). Higher RRG weights effectively eliminates emerging artifacts albeit at the cost of slightly blurrier outputs. $\delta = 200$ strikes a good balance. Improvements are more noticeable when zooming in. +Figure 6. Effect of Resampling. Applying more resampling steps noticeably enhances detail in the generated images. + +in $\mathbf{x}_t$ with newly sampled ones from $x_{t}$ to get an updated version $\tilde{\mathbf{x}}_t$ . Following each update, the direction score is recalculated and blended with the previously calculated score. Formally, we consider $\mathbf{S}_d^0 = \mathbf{S}_d$ and define the iterative resampling as: + +$$ +\mathbf {S} _ {d} ^ {r + 1} = \mathbf {S} _ {d} ^ {r} \odot m + \tilde {\mathbf {S}} _ {d} \odot (1 - m), \tag {6} +$$ + +where $m \in \{0,1\}^{\bar{H} \times \bar{W}}$ is a mask with a value of 1 at the positions of the newly sampled pixels and 0 elsewhere. $\tilde{\mathbf{S}}_d$ represents the recalculated class direction score on $\tilde{\mathbf{x}}_t$ as per Eq. (4). This method estimates the class direction score of $n\%$ new pixels while retaining the information of the previously estimated score, thereby increasing the score's overall resolution. In our experiments, we set $n$ to $20\%$ and repeat the process $R$ times, depending on the target generation resolution. Fig. 6 demonstrates the effectiveness of our method in enhancing the details of the generated images. + +# 4.4. Reduced-Resolution Guidance (RRG) + +We effectively estimate the unconditional score signal and concurrently steer the image generation using the class direction score. However, inaccuracies in unconditional score estimation or fluctuations in local content, especially from distant patches, can lead to artifacts. To enhance image coherence and diminish artifacts, we consider a downsampled version of the latent at each decoding step as a reference and aim to align the decoded latent with it through our reduced-resolution latent update strategy. Specifically, we utilize the noise-free sample $\hat{x}_0^t$ of the latent $x_{t}$ from Eq. (1) and generate a corresponding noise-free sample $\hat{\mathbf{x}}_0^t$ from its downsampled counterpart $\mathbf{x}^t$ in Eq. (4) as: + +$$ +\hat {\mathbf {x}} _ {0} ^ {t} = \frac {1}{\sqrt {\bar {\alpha} _ {t}}} (\mathbf {x} _ {t} - \sqrt {1 - \bar {\alpha} _ {t}} \cdot (\epsilon_ {\theta} (\mathbf {x} _ {t}) + (1 + w) \cdot \Delta_ {\mathcal {C}} (\mathbf {x} _ {t}, c))), +$$ + +Here, both $\hat{x}_0^t$ and $\hat{\mathbf{x}}_0^t$ corresponds to the same latent update at different resolutions. Due to its smaller dimension, $\hat{\mathbf{x}}_0^t$ has a broader context when computing the unconditional + +![](images/59eb54a5213e8dbd5fe4904d66b7320fc934009382dd44d7d1833a563cfd971f.jpg) +Figure 7. Qualitative comparison at various resolutions on CelebAHQ faces. ElasticDiffusion consistently generates coherent images at all resolutions. StableDiffusion, and MultiDiffusion produce repeating body parts at higher resolutions, while StableDiffusion-XL fails to maintain its quality at lower resolutions, resulting in noisy outputs at $256 \times 256$ . We exclude MultiDiffusion results at $256 \times 256$ as it is not designed to produce images at lower resolutions. + +signal. To guide $x_{t}$ with its downsampled reference, we refine the latent $x_{t}$ with the direction that minimizes $L2$ difference between $\hat{\mathbf{x}}_0^t$ and $\hat{x}_0^t$ . Formally, + +$$ +\bar {x} _ {t - 1} \leftarrow \bar {x} _ {t - 1} - \delta_ {t} \nabla_ {x _ {t}} | | \text {U p s a m p l e} \left(\hat {\mathbf {x}} _ {0} ^ {t}, \bar {H} \times \bar {W}\right) - \hat {x} _ {0} ^ {t} | |, \tag {7} +$$ + +where $\delta_t$ represent the weight of the guidance. Since the overall image structure is determined in the early diffusion steps, we start with $\delta_t = 400$ and follow a cosine scheduler to decrease this weight for later diffusion steps. This scheduling strategy mitigates potential quality degradation from matching the generated image with a lower-resolution version while allowing the model to fill-in higher-resolution details in the later decoding stages. Fig. 5 illustrates how RRG eliminates emerging artifacts. + +# 5. Experiments + +ElasticDiffusion supports generating images across different resolutions and aspect ratios. Our experiments focus on: (1) square images at multiple resolutions, and (2) images with varied aspect ratios and resolutions. + +Datasets. We evaluate the generation of square images on the Multi-Modal CelebAHQ dataset [56], which includes high-resolution square face images accompanied with text descriptions. We evaluate different aspect ratio generation using the LAION-COCO dataset [46], derived from the web-crawled LAION-5B dataset [45], which includes a variety of image aspect ratios and contains landscapes, people, + +objects, and everyday scenes, each paired with a synthetic caption generated using BLIP [27]. We consider four common aspect ratios: 9:16, 16:9, 3:4 and 4:3. + +Evaluation Metrics. Following prior text-to-image synthesis works [3, 33, 35, 38, 43], we use automatic evaluation metrics Frechet Inception Distance (FID) and CLIP-score. FID [17] measures both the realism and diversity of the generated images by calculating the difference between features of the real and generated images computed using Inception-V3 [49] pretrained on ImageNet [41]. CLIP-score uses a pretrained text-image CLIP model [36] to measure alignment between the generated images and input prompts. We use 10,000 images to compute these scores. + +Baselines. We compare our approach against prior diffusion model generation methods, specifically focusing on Stable Diffusion (SD) and MultiDiffusion (MD). SD follows the standard reverse diffusion process on an image latent space. MD uses a pretrained $SD$ for panoramic image generation, by creating smaller, overlapping patches and averaging the Diffusion Model scores in intersecting areas. For baseline comparisons, we fix the pre-trained diffusion model to $SD_{1.4}$ , which is trained to generate images at a resolution of $512 \times 512$ . Additionally, we compare our method with SDXL, an enhanced SD model that is three times larger than the standard one. SDXL is trained at a higher resolution of $1024p$ , and fine-tuned on a specific set of aspect ratios with pixel sums close to $1024^2$ . We exclude + +![](images/6596084cb3897a01d054b5e051ecb42341e87ca58efe25e350836c561a05a7d3.jpg) +Figure 8. Qualitative comparison on various aspect ratios. ElasticDiffusion effectively handles a variety of aspect ratios. In comparison, $SD$ and MultiDiffusion generate unrealistic images, while SDXL outputs exhibit a decline in the perceptual quality. + +the latent refinement module in $SDXL$ and focus our analysis on the base model. Throughout our experiments, we employ a DDIM sampling strategy with 50 steps and use 7.0 for the classifier-free guidance scaling factor. We also ensure consistency in seeds and captions for all baselines. + +# 5.1. Qualitative Results + +We show square image generation samples in Fig. 7 generated from the CelebAHQ dataset at various resolutions. Both MultiDiffusion and Stable Diffusion have a tendency to replicate textures and body parts at higher resolutions, resulting in images that lack coherence. At resolutions lower than its training resolution $512 \times 512$ , StableDiffusion aligns poorly with the provided captions and produces unappealing images. SDXL, trained for $1024 \times 1024$ resolution, has a similar trend of reduced perceptual quality at lower sizes, eventually producing complete noise at a resolution of $256 \times 256$ . In contrast, ElasticDiffusion maintains image coherence across all tested resolutions, and yields results comparable to SDXL at $1024 \times 1024$ , despite using a less powerful base model and lower memory. We excluded + +Table 1. Quantitative comparison of on CelebA-HQ and LAION-COCO at different resolution. We indicate the best performances in bold, and underline the second best ones. #Calls represent the number of diffusion model calls required at each decoding steps. + +
Res.MethodCelebA-HQLAION-COCO#CallsMem.
FID ↓CLIP ↑FID ↓CLIP ↑
256 × 256SD1.4258.4320.1454.0621.4327.2GB
SDXL368.0614.40175.8714.60218.5GB
Ours1.4235.2323.8823.7726.3028.6GB
512 × 512SD1.4233.4024.0020.5027.3328.6GB
SDXL240.2021.5742.5825.34221.6GB
768 × 768SD1.4238.8723.4529.8927.01211.1GB
MD1.4240.5622.8229.9827.31508.6GB
SDXL225.4824.2323.3127.88224.5GB
Ours1.4225.8626.6625.7825.93178.6GB
1024 × 1024SD1.4266.0121.9047.0125.70114.7GB
MD1.4264.5721.5537.7026.961628.6GB
SDXL230.2124.6225.5828.06127.5GB
Ours1.4228.8723.7427.7626.07338.6GB
+ +results at $512 \times 512$ because both our method and Multi-Diffusion generate same outputs as the base Stable Diffusion model. We also omitted the results of MultiDiffusion at $256 \times 256$ since it is not designed to generate images at sizes smaller than the base model. Fig. 8 presents generated samples at various aspect ratios. We observe a similar trend of pattern repetition and reduced perceptual quality for images generated by the baselines, in contrast to those generated by ElasticDiffusion. Finally, we apply our method using SDXL as the base model to enable Full-HD image generation ( $1920 \times 1080$ ) and provide examples in Supp. Sec. 3.3. + +# 5.2. Quantitative Results + +Table 1 shows quantitative evaluations. StableDiffusion shows increasing image quality degradation, as indicated by the FID metric when processing images of sizes different from its training resolution $512 \times 512$ . MultiDiffusion slightly improves FID at the expense of a substantially more forward calls. SDXL demonstrates similar declines in quality at resolutions far from its fine-tuning size of $1024 \times 1024$ . ElasticDiffusion, however, improves FID while maintaining a comparable CLIP-score to the base model. ElasticDiffusion also significantly improves the performance for lower resolutions $256 \times 256$ , while achieving comparable results to SDXL at its training resolution $1024 \times 1024$ , with only $\sim 31\%$ of the memory requirement for SDXL. + +Table 2. Quantitative comparison of on LAION-COCO datasets at various aspect ratios (A) and resolutions (R). best performances are in bold, and the second best are underlined. Vertical means the resolution is transposed from H:W to W:H. + +
ARMethodHorizontalVertical
FID ↓CLIP↑FID ↓CLIP↑
3:4384 × 512SD1.438.8624.6317.6626.54
MD1.4----
SDXL104.8421.3368.8422.40
Ours1.435.1024.9115.5027.33
512 × 680SD1.445.5424.9616.8127.33
MD1.443.4424.5619.1326.80
SDXL62.8024.2028.2326.17
Ours1.441.0624.4018.9026.83
768 × 1024SD1.471.0024.0928.8326.30
MD1.454.8925.0226.3526.95
SDXL47.0525.7919.5027.41
Ours1.447.0324.9122.5225.80
9:16288 × 512SD1.423.5024.6924.0124.89
MD1.4----
SDXL121.8317.65112.4118.54
Ours1.423.2325.2622.8626.30
512 × 904SD1.429.8625.3427.4526.01
MD1.426.3525.7026.7025.28
SDXL29.6025.4027.2726.08
Ours1.422.8525.0126.6826.12
+ +Table 2 provides an evaluation on a variety of aspect ratios and resolutions for both horizontal and vertical image resolutions. StableDiffusion obtains worse FID score with increasing resolutions, while surprisingly maintaining or improving its CLIP-score. We posit that since CLIP-score quantifies the agreement between the input prompt and the generated image, StableDiffusion benefits from generating repetitive textures and artifacts that align closely with the prompt. For example, a photo of repeated lipsticks might align more with the caption "lipstick" despite its poor composition. MultiDiffusion generally enhances image quality but does not achieve satisfactory performance. MultiDiffusion struggles with objects that span the entire image (e.g. Fig. 8). Our method consistently improves FID over the baseline on horizontal and most vertical resolutions while preserving fidelity to the input prompts, thereby attaining comparable or superior CLIP-scores. Remarkably, even with a larger model size and explicit fine-tuning at a similar resolution of $768 \times 1280$ , SDXL only marginally surpasses our method at the $768 \times 1024$ resolution. + +Table 3. Ablation analysis of ElasticDiffusion on 500 images. + +
Model DetailsFID ↓CLIP ↑
ElasticDiffusion133.6725.82
w/o Resampling150.0123.82
w/o RRG150.6424.34
w/o Imp. overlap, w/ exp. overlap141.4225.82
+ +# 5.3. Ablation study + +Tab. 3 presents results of our method when excluding key components, demonstrating that each element improves performance. Fig. 4 illustrates the effectiveness of our proposed implicit overlap method in resolving boundary discontinuities at a reduced computational cost. Fig. 5 highlights the effectiveness of Reduced-Resolution Guidance in removing emerging artifacts. Fig. 6 shows the effectiveness of our iterative resampling technique in enhancing details. + +# 6. Discussion and Conclusion + +Experimental results highlight the adaptability and effectiveness of ElasticDiffusion at steering diffusion models to produce an array of resolutions and aspect ratios. ElasticDiffusion requires no fine-tuning, consumes a constant memory footprint, enables both increased and reduced resolutions, and can generate a variety of aspect ratios. + +ElasticDiffusion, however, does have several practical limitations. First, inaccuracies in estimating the global and local signals may result in artifacts. Although we attempt to mitigate artifacts with our Reduced-resolution guidance, it can still generate blurrier outputs. Second, since the global content guidance is initially estimated at the original training resolution of the underlying diffusion model, our method is less effective in generating images at significantly extended sizes beyond the training resolution. In particular, at extreme resolutions beyond 4X, our method produces artifacts and images of a lower perceptual quality. We provide examples of these failure cases in Supp. Sec. 10.4. + +The main insight underpinning our method is a novel reinterpretation of classifier-free guidance in somewhat disentangling both global and local content. Our comprehensive evaluations demonstrate the feasibility of disentangling these signals, yet the full extent of their separation offers an avenue for further exploration in this direction. We hope that our findings inspire future work in investigating the separation of global and local content guidance signals for image synthesis. This separation holds potential for various applications such as selectively manipulating local and global content or enhancing style transfer. Additionally, the rich semantic representation in the class-direction score has the potential for improving discriminative models. + +Acknowledgments. This work was partially funded by NSF Award #2201710. + +# References + +[1] Omri Avrahami, Dani Lischinski, and Ohad Fried. Blended diffusion for text-driven editing of natural images. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). IEEE, 2022. 2 +[2] Arpit Bansal, Hong-Min Chu, Avi Schwarzschild, Soumyadip Sengupta, Micah Goldblum, Jonas Geiping, and Tom Goldstein. Universal guidance for diffusion models, 2023. 2 +[3] Omer Bar-Tal, Lior Yariv, Yaron Lipman, and Tali Dekel. Multidiffusion: Fusing diffusion paths for controlled image generation. In ICML, 2023. 2, 3, 4, 6 +[4] Angela Castillo, Jonas Kohler, Juan C. Pérez, Juan Pablo Pérez, Albert Pumarola, Bernard Ghanem, Pablo Arbeláez, and Ali Thabet. Adaptive guidance: Training-free acceleration of conditional diffusion models, 2023. 2 +[5] Duygu Ceylan, Chun-Hao Paul Huang, and Niloy J. Mitra. Pix2video: Video editing using image diffusion. In ICCV, 2023. 2 +[6] Hyungjin Chung, Jeongsol Kim, Michael T. McCann, Marc L. Klasky, and Jong Chul Ye. Diffusion posterior sampling for general noisy inverse problems, 2023. 2 +[7] Jiwoo Chung, Sangeek Hyun, and Jae-Pil Heo. Style injection in diffusion: A training-free approach for adapting large-scale diffusion models for style transfer, 2024. 2 +[8] Prafulla Dhariwal and Alex Nichol. Diffusion models beat gans on image synthesis. In NeurIPS, 2021. 2, 3 +[9] Ruoyi Du, Dongliang Chang, Timothy Hospedales, Yi-Zhe Song, and Zhanyu Ma. Demofusion: Democratising high-resolution image generation with no $$$. In CVPR, 2024. 2 +[10] Alexandros Graikos, Srikar Yellapragada, Minh-Quan Le, Saarthak Kapse, Prateek Prasanna, Joel Saltz, and Dimitris Samaras. Learned representation-guided diffusion models for large-image generation, 2024. 2 +[11] Jiatao Gu, Qingzhe Gao, Shuangfei Zhai, Baoquan Chen, Lingjie Liu, and Josh Susskind. Control3diff: Learning controllable 3d diffusion models from single-view images. In 3DV24, 2023. 2 +[12] Lanqing Guo, Yingqing He, Haoxin Chen, Menghan Xia, Xiaodong Cun, Yufei Wang, Siyu Huang, Yong Zhang, Xintao Wang, Qifeng Chen, Ying Shan, and Bihan Wen. Make a cheap scaling: A self-cascade diffusion model for higher-resolution adaptation, 2024. 2 +[13] Feihong He, Gang Li, Lingyu Si, Leilei Yan, Shimeng Hou, Hongwei Dong, and Fanzhang Li. Cartooniff: Training-free cartoon image generation with diffusion transformer models, 2023. 2 +[14] Yingqing He, Shaoshu Yang, Haoxin Chen, Xiaodong Cun, Menghan Xia, Yong Zhang, Xintao Wang, Ran He, Qifeng Chen, and Ying Shan. Scalecrafter: Tuning-free higher-resolution visual generation with diffusion models, 2023. 2 +[15] Yingqing He, Tianyu Yang, Yong Zhang, Ying Shan, and Qifeng Chen. Latent video diffusion models for high-fidelity long video generation, 2023. 2 + +[16] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-prompt image editing with cross attention control. In ICLR, 2022. 2 +[17] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. In NIPS, 2017. 6 +[18] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance, 2022. 2 +[19] Jonathan Ho, Chitwan Sahara, William Chan, David J. Fleet, Mohammad Norouzi, and Tim Salimans. Cascaded diffusion models for high fidelity image generation. Journal of Machine Learning Research, 23(47):1-33, 2022. 2 +[20] Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J. Fleet. Video diffusion models. In NeurIPS, 2022. 2 +[21] Edward J. Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. In ICLR, 2022. 2 +[22] Rongjie Huang, Jiawei Huang, Dongchao Yang, Yi Ren, Luping Liu, Mingze Li, Zhenhui Ye, Jinglin Liu, Xiang Yin, and Zhou Zhao. Make-an-audio: Text-to-audio generation with prompt-enhanced diffusion models. In ICML, 2023. 2 +[23] Yujin Jeong, Wonjeong Ryoo, Seunghyun Lee, Dabin Seo, Wonmin Byeon, Sanggil Kim, and Jinkyu Kim. The power of sound (tpos): Audio reactive video generation with stable diffusion. In ICCV, 2023. 2 +[24] Zhiyu Jin, Xuli Shen, Bin Li, and Xiangyang Xue. Training-free diffusion model adaptation for variable-sized text-to-image synthesis, 2023. 2 +[25] Gwanghyun Kim, Taesung Kwon, and Jong Chul Ye. Diffusionclip: Text-guided diffusion models for robust image manipulation. In CVPR, 2022. 2 +[26] Zhifeng Kong and Wei Ping. On fast sampling of diffusion probabilistic models. In ICML, 2021. 2 +[27] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation, 2022. 6 +[28] Jin Liu, Huaibo Huang, Chao Jin, and Ran He. Portrait diffusion: Training-free face stylization with chain-of-painting, 2023. 2 +[29] Andreas Lugmayr, Martin Danelljan, Andres Romero, Fisher Yu, Radu Timofte, and Luc Van Gool. Repaint: Inpainting using denoising diffusion probabilistic models. In CVPR, 2022. 5 +[30] Sicheng Mo, Fangzhou Mu, Kuan Heng Lin, Yanli Liu, Bochen Guan, Yin Li, and Bolei Zhou. Freecontrol: Training-free spatial control of any text-to-image diffusion model with any condition, 2023. 2 +[31] Brian B. Moser, Stanislav Frolov, Federico Raue, Sebastian Palacio, and Andreas Dengel. Dynamic attention-guided diffusion for image super-resolution, 2024. 2 +[32] Alex Nichol and Prafulla Dhariwal. Improved denoising diffusion probabilistic models. In ICML, 2021. 2 +[33] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and + +Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. In ICML, 2022. 2, 6 +[34] Bo Peng, Xinyuan Chen, Yaohui Wang, Chaochao Lu, and Yu Qiao. Conditionvideo: Training-free condition-guided text-to-video generation, 2023. 2 +[35] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis, 2023. 2, 6 +[36] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In ICML, 2021. 6 +[37] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents, 2022. 2 +[38] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents, 2022. 6 +[39] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 1, 2, 3 +[40] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In CVPR, 2023. 2 +[41] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, Alexander C. Berg, and Li Fei-Fei. Imagenet large scale visual recognition challenge, 2015. 6 +[42] Chitwan Sahara, Jonathan Ho, William Chan, Tim Salimans, David J. Fleet, and Mohammad Norouzi. Image superresolution via iterative refinement, 2021. 2 +[43] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, Seyed Kamyar Seyed Ghasemipour, Burcu Karagol Ayan, S. Sara Mahdavi, Rapha Gontijo Lopes, Tim Salimans, Jonathan Ho, David J Fleet, and Mohammad Norouzi. Photorealistic text-to-image diffusion models with deep language understanding. In NeurIPS, 2022. 2, 3, 6 +[44] Axel Sauer, Frederic Boesel, Tim Dockhorn, Andreas Blattmann, Patrick Esser, and Robin Rombach. Fast high-resolution image synthesis with latent adversarial diffusion distillation, 2024. 2 +[45] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, Patrick Schramowski, Srivatsa Kundurthy, Katherine Crowson, Ludwig Schmidt, Robert Kaczmarczyk, and Jenia Jitsev. Laion-5b: An open large-scale dataset for training next generation image-text models, 2022. 6 +[46] Christoph Schuhmann, Andreas A. Köpf, Richard Vencu, Theo Coombes, and Ross Beaumont. Laion coco: 600m synthetic captions from laion2b-en, 2023. 6 + +[47] Fengyuan Shi, Jiaxi Gu, Hang Xu, Songcen Xu, Wei Zhang, and Limin Wang. Bivdiff: A training-free framework for general-purpose video synthesis via bridging image and video diffusion models, 2023. 2 +[48] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In ICLR, 2021. 2, 3 +[49] Christian Szegedy, Vincent Vanhoucke, Sergey Ioffe, Jonathon Shlens, and Zbigniew Wojna. Rethinking the inception architecture for computer vision. In CVPR, 2015. 6 +[50] Andrey Voynov, Amir Hertz, Moab Arar, Shlomi Fruchter, and Daniel Cohen-Or. Anylens: A generative diffusion model with any rendering lens, 2023. 2 +[51] Bingyuan Wang, Hengyu Meng, Zeyu Cai, Lanjiong Li, Yue Ma, Qifeng Chen, and Zeyu Wang. Magicscrol: Nontypical aspect-ratio image generation for visual storytelling via multi-layered semantic-aware denoising, 2023. 2 +[52] Fu-Yun Wang, Zhaoyang Huang, Xiaoyu Shi, Weikang Bian, Guanglu Song, Yu Liu, and Hongsheng Li. *Animatelcm: Accelerating the animation of personalized diffusion models and adapters with decoupled consistency learning*, 2024. 2 +[53] Jianyi Wang, Zongsheng Yue, Shangchen Zhou, Kelvin C. K. Chan, and Chen Change Loy. Exploiting diffusion prior for real-world image super-resolution, 2023. 2, 3, 4 +[54] Lezhong Wang, Jeppe Revall Frisvad, Mark Bo Jensen, and Siavash Arjomand Bigdeli. Stereodiffusion: Training-free stereo image generation using latent diffusion models, 2024. 2 +[55] Daniel Watson, William Chan, Ricardo Martin-Brualla, Jonathan Ho, Andrea Tagliasacchi, and Mohammad Norouzi. Novel view synthesis with diffusion models. In ICLR, 2023. 2 +[56] Weihao Xia, Yujiu Yang, Jing-Hao Xue, and Baoyuan Wu. Tedigan: Text-guided diverse face image generation and manipulation, 2021. 6 +[57] Guy Yariv, Itai Gat, Lior Wolf, Yossi Adi, and Idan Schwartz. Audiotoken: Adaptation of text-conditioned diffusion models for audio-to-image generation, 2023. 2 +[58] Ahmet Burak Yildirim, Vedat Baday, Erkut Erdem, Aykut Erdem, and Aysegul Dundar. Inst-inpaint: Instructing to remove objects with diffusion models, 2023. 2 +[59] Jiwen Yu, Yinhuai Wang, Chen Zhao, Bernard Ghanem, and Jian Zhang. Freedom: Training-free energy-guided conditional diffusion model, 2023. 2 +[60] Shen Zhang, Zhaowei Chen, Zhenyu Zhao, Zhenyuan Chen, Yao Tang, Yuhao Chen, Wengang Cao, and Jiajun Liang. Hidiffusion: Unlocking high-resolution creativity and efficiency in low-resolution trained diffusion models, 2023. 2 +[61] Peiang Zhao, Han Li, Ruiyang Jin, and S. Kevin Zhou. Loco: Locally constrained training-free layout-to-image synthesis, 2024. 2 +[62] Qingping Zheng, Yuanfan Guo, Jiankang Deng, Jianhua Han, Ying Li, Songcen Xu, and Hang Xu. Any-sized diffusion: Toward efficient text-driven synthesis for any-size hd images, 2023. 2 +[63] Álvaro Barbero Jiménez. Mixture of diffusers for scene composition and high resolution image generation, 2023. 2, 3, 4 \ No newline at end of file diff --git a/2024/ElasticDiffusion_ Training-free Arbitrary Size Image Generation through Global-Local Content Separation/images.zip b/2024/ElasticDiffusion_ Training-free Arbitrary Size Image Generation through Global-Local Content Separation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..c492d1ba12649f2fba11a722407719d0f5890789 --- /dev/null +++ b/2024/ElasticDiffusion_ Training-free Arbitrary Size Image Generation through Global-Local Content Separation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:cbbee42ebc7b834cd498990020b23d928a816c42513f6172c0541e32a5f9d51a +size 958274 diff --git a/2024/ElasticDiffusion_ Training-free Arbitrary Size Image Generation through Global-Local Content Separation/layout.json b/2024/ElasticDiffusion_ Training-free Arbitrary Size Image Generation through Global-Local Content Separation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..2d9ad1cfe8fc3e35cdaf965ca2d975af44c70377 --- /dev/null +++ b/2024/ElasticDiffusion_ Training-free Arbitrary Size Image Generation through Global-Local Content Separation/layout.json @@ -0,0 +1,9129 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 97, + 103, + 496, + 139 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 97, + 103, + 496, + 139 + ], + "spans": [ + { + "bbox": [ + 97, + 103, + 496, + 139 + ], + "type": "text", + "content": "ElasticDiffusion: Training-free Arbitrary Size Image Generation through Global-Local Content Separation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 95, + 161, + 179, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 95, + 161, + 179, + 175 + ], + "spans": [ + { + "bbox": [ + 95, + 161, + 179, + 175 + ], + "type": "text", + "content": "Moayed Haji-Ali" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 250, + 161, + 344, + 189 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 250, + 161, + 344, + 189 + ], + "spans": [ + { + "bbox": [ + 250, + 161, + 344, + 189 + ], + "type": "text", + "content": "Guha Balakrishnan Rice University" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 209, + 191, + 386, + 202 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 209, + 191, + 386, + 202 + ], + "spans": [ + { + "bbox": [ + 209, + 191, + 386, + 202 + ], + "type": "text", + "content": "{mh155, guha, vicenteor}@rice.edu" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 415, + 161, + 498, + 174 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 415, + 161, + 498, + 174 + ], + "spans": [ + { + "bbox": [ + 415, + 161, + 498, + 174 + ], + "type": "text", + "content": "Vicente Ordonez" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 48, + 215, + 545, + 488 + ], + "blocks": [ + { + "bbox": [ + 48, + 215, + 545, + 488 + ], + "lines": [ + { + "bbox": [ + 48, + 215, + 545, + 488 + ], + "spans": [ + { + "bbox": [ + 48, + 215, + 545, + 488 + ], + "type": "image", + "image_path": "6332836812d5c3aa5ab1d905bbb973969a8793b5267a835e2d10f1c30630b170.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 495, + 546, + 529 + ], + "lines": [ + { + "bbox": [ + 46, + 495, + 546, + 529 + ], + "spans": [ + { + "bbox": [ + 46, + 495, + 546, + 529 + ], + "type": "text", + "content": "Figure 1. ElasticDiffusion generates high quality images at arbitrary sizes using a pretrained diffusion model trained on a single image size, with equivalent memory footprint and no further training. These results are based on Stable Diffusion1,4, which was trained to generate " + }, + { + "bbox": [ + 46, + 495, + 546, + 529 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 46, + 495, + 546, + 529 + ], + "type": "text", + "content": " images. The examples shown in this collage are presented without any image cropping, stretching, or post-processing." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 143, + 538, + 192, + 551 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 538, + 192, + 551 + ], + "spans": [ + { + "bbox": [ + 143, + 538, + 192, + 551 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 45, + 558, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 558, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 45, + 558, + 288, + 714 + ], + "type": "text", + "content": "Diffusion models have revolutionized image generation in recent years, yet they are still limited to a few sizes and aspect ratios. We propose ElasticDiffusion, a novel training-free decoding method that enables pretrained text-to-image diffusion models to generate images with various sizes. ElasticDiffusion attempts to decouple the generation trajectory of a pretrained model into local and global signals. The local signal controls low-level pixel information and can be estimated on local patches, while the global signal is used to maintain overall structural consistency and is estimated with a reference image. We test our method on CelebA-HQ (faces) and LAION-COCO (objects/indoor/outdoor scenes). Our experiments and qualita" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 540, + 545, + 588 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 540, + 545, + 588 + ], + "spans": [ + { + "bbox": [ + 305, + 540, + 545, + 588 + ], + "type": "text", + "content": "tive results show superior image coherence quality across aspect ratios compared to MultiDiffusion and the standard decoding strategy of Stable Diffusion. Project Webpage: https://elasticdiffusion.github.io/" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 595, + 386, + 607 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 595, + 386, + 607 + ], + "spans": [ + { + "bbox": [ + 306, + 595, + 386, + 607 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 617, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 617, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 617, + 546, + 713 + ], + "type": "text", + "content": "Diffusion models are a powerful family of algorithms that achieve remarkable quality and obtain the current state-of-the-art performance on various image synthesis tasks. As is the case with most deep neural networks, diffusion models are typically trained on one or a few image resolutions. For instance, Stable Diffusion (SD) [39], one of the most widely adopted diffusion models, is trained on a square images of size " + }, + { + "bbox": [ + 304, + 617, + 546, + 713 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 304, + 617, + 546, + 713 + ], + "type": "text", + "content": ", yet fails to maintain its performance at dif-" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "6603" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "text", + "content": "ferent aspect ratios during inference time. In practice, many applications require a wide aspect ratio or portrait mode, such as digital billboards, wearable devices, automotive displays, and any application relying on a computer screen." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 122, + 289, + 362 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 122, + 289, + 362 + ], + "spans": [ + { + "bbox": [ + 46, + 122, + 289, + 362 + ], + "type": "text", + "content": "Recent studies address the issue of variable image size in different ways. SDXL [35] and Any-Size-Diffusion [62] explicitly finetune models on images with a range of aspect ratios, which requires extensive computation, a quadratic memory footprint, and larger training data. In addition, this strategy requires a set of resolutions to be specified upfront during training time, while the models still struggle to generalize to new resolutions during inference, often resulting in artifacts. Recent works also show remarkable results in generating panoramic images using pretrained diffusion models by overlapping generated patches into a larger image [3, 63]. These methods work well for landscape images with repetitive patterns. However, their lack of global guidance limits their abilities to generate images of single objects or faces where global structure is important. Recent work [24, 50, 51] aimed at extending pre-trained diffusion models capabilities to others domains, often in a training-free way [4, 7, 13, 28, 30, 34, 47, 54, 61]. Few concurrent work [9, 10, 12, 31, 60] focus on adapting them to higher resolutions, yet they are constrained to square images." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 364, + 289, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 364, + 289, + 592 + ], + "spans": [ + { + "bbox": [ + 46, + 364, + 289, + 592 + ], + "type": "text", + "content": "In this work, we propose ElasticDiffusion, a novel decoding strategy that takes a pretrained diffusion model and generate images at arbitrary sizes during inference using a constant memory footprint. To achieve this, we revisit the guidance mechanism of conditional diffusion models to decouple global and local content generation. Global content controls the high-level aspects of the image, whereas local content adds finer, more granular details. This separation facilitates generating the local content in patches for images of varying sizes, all while being guided with global content that we derive from a reference image at the diffusion model pretraining resolution. This enables the synthesis of images at diverse resolutions and aspect ratios while adhering the diffusion forward calls to the model's initial training size. To aid in this task, we introduce several techniques including an efficient patch fusion method for smooth boundaries, a novel guidance strategy to reduce image artifacts, and a global content resampling technique that amplifies the resolution of diffusion models up to 4X the training size." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 594, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 289, + 715 + ], + "type": "text", + "content": "Figure 1 shows a diverse array of images generated with ElasticDiffusion. Several of theses images include a single object or were generated with extreme aspect ratios, showcasing our method's ability to produce coherent images under various sizes. Quantitatively, ElasticDiffusion outperforms baselines across most aspect ratios. More importantly, despite relying on Stable Diffusion1,4, ElasticDiffusion obtains comparable FID (228.87 vs 230.21) and CLIP scores (26.07 vs 28.06) as SDXL when generating images at " + }, + { + "bbox": [ + 46, + 594, + 289, + 715 + ], + "type": "inline_equation", + "content": "1024 \\times 1024" + }, + { + "bbox": [ + 46, + 594, + 289, + 715 + ], + "type": "text", + "content": ", which is the native resolution of SDXL." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 305, + 71, + 393, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 71, + 393, + 84 + ], + "spans": [ + { + "bbox": [ + 305, + 71, + 393, + 84 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 91, + 547, + 390 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 91, + 547, + 390 + ], + "spans": [ + { + "bbox": [ + 304, + 91, + 547, + 390 + ], + "type": "text", + "content": "Diffusion Models have been widely adopted for their high-quality outputs in generative tasks [8, 11, 15, 20, 22, 35, 37, 39, 43, 55, 58]. These models involve iterative decoding with many steps leading to high compute and memory requirements. Recent work addressed these issues by devising faster sampling strategies [26, 48], hierarchical models [19, 37, 43], progressive training at increasing resolutions [33, 35, 39], and two-stage models [35, 39, 42, 53]. Stable Diffusion (SD) [39] trains a variational auto-encoder to compress images into a low-dimensional " + }, + { + "bbox": [ + 304, + 91, + 547, + 390 + ], + "type": "inline_equation", + "content": "64 \\times 64" + }, + { + "bbox": [ + 304, + 91, + 547, + 390 + ], + "type": "text", + "content": " latent space and trains a diffusion model on this latent space. To train for higher resolutions, models are initially trained at a " + }, + { + "bbox": [ + 304, + 91, + 547, + 390 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 304, + 91, + 547, + 390 + ], + "type": "text", + "content": " resolution, before fine-tuning them at " + }, + { + "bbox": [ + 304, + 91, + 547, + 390 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 304, + 91, + 547, + 390 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 91, + 547, + 390 + ], + "type": "inline_equation", + "content": "1024 \\times 1024" + }, + { + "bbox": [ + 304, + 91, + 547, + 390 + ], + "type": "text", + "content": " in the case of SDXL [35]. SD is one of the few large-scale diffusion models that released trained parameters, making it the building block for many subsequent work [5, 16, 40, 44, 52, 53]. However, these models, including SD, are confined to specific resolutions and do not generalize well to aspect ratios unseen during training. Interestingly, despite being presented with a " + }, + { + "bbox": [ + 304, + 91, + 547, + 390 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 304, + 91, + 547, + 390 + ], + "type": "text", + "content": " resolution during their early training stages, both SD and SDXL fail to generate realistic images at this resolution after being fine-tuned for larger outputs. ElasticDiffusion enables high quality generation at unseen resolutions including re-enabling consistent high quality outputs for SD at " + }, + { + "bbox": [ + 304, + 91, + 547, + 390 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 304, + 91, + 547, + 390 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 391, + 547, + 534 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 391, + 547, + 534 + ], + "spans": [ + { + "bbox": [ + 304, + 391, + 547, + 534 + ], + "type": "text", + "content": "Mixture of diffusers [63] and MultiDiffusion [3] generate panoramic images using a pre-trained diffusion model by generating overlapping crops and combining the generation signal of the overlapping regions. Blending multiple generation signals spatially has been of interest in many previous work [1, 63]. SDXL [35] and Any-Size-Diffusion [62] fine-tune a pre-trained SD on a fixed set of resolutions. A concurrent work [14] uses dilated convolution kernels to reduce SD artifacts when generating non-square images. ElasticDiffusion extends a pre-trained SD to generate images of various sizes at a constant memory requirement and without additional training, all while ensuring global coherence." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 534, + 548, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 534, + 548, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 534, + 548, + 715 + ], + "type": "text", + "content": "Guided diffusion models devise strategies to condition image generation based on text and other modalities [2, 2, 6, 18, 23, 25, 32, 39, 57, 59]. Classifier guidance [32] uses a pre-trained classifier on noisy images to guide the generation process. Classifier-free guidance [18] eliminates the need for a pretrained classifier but requires training a conditional diffusion model, limiting its applications. Universal Guidance [2] applies a pre-trained classifier on the noise-free images produced by DDIM [48], bypassing training on noisy images. StableSR [53] uses LoRA [21] to condition the generation on low-resolution inputs to achieve superb image super-resolution. Inspired by this, we propose Reduce-Resolution Guidance (Sec. 4) to constrain the generation of high-res images using a lower resolution version, substantially reducing artifacts without extra training." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "6604" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 71, + 221, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 71, + 221, + 85 + ], + "spans": [ + { + "bbox": [ + 47, + 71, + 221, + 85 + ], + "type": "text", + "content": "3. Background: Diffusion Models" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 91, + 287, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 91, + 287, + 163 + ], + "spans": [ + { + "bbox": [ + 47, + 91, + 287, + 163 + ], + "type": "text", + "content": "A conditional diffusion model " + }, + { + "bbox": [ + 47, + 91, + 287, + 163 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta} \\colon \\mathcal{X} \\times \\mathcal{C} \\to \\mathcal{X}" + }, + { + "bbox": [ + 47, + 91, + 287, + 163 + ], + "type": "text", + "content": " predicts a less noisy version of the input image " + }, + { + "bbox": [ + 47, + 91, + 287, + 163 + ], + "type": "inline_equation", + "content": "x \\in \\mathbb{R}^{H \\times W \\times 3}" + }, + { + "bbox": [ + 47, + 91, + 287, + 163 + ], + "type": "text", + "content": ", conditioned on variable " + }, + { + "bbox": [ + 47, + 91, + 287, + 163 + ], + "type": "inline_equation", + "content": "c \\in \\mathbb{R}^D" + }, + { + "bbox": [ + 47, + 91, + 287, + 163 + ], + "type": "text", + "content": " (e.g. a text embedding). Starting with " + }, + { + "bbox": [ + 47, + 91, + 287, + 163 + ], + "type": "inline_equation", + "content": "x_{T} \\sim \\mathcal{N}(0, \\mathbf{I})" + }, + { + "bbox": [ + 47, + 91, + 287, + 163 + ], + "type": "text", + "content": ", the reverse diffusion process progressively denoise " + }, + { + "bbox": [ + 47, + 91, + 287, + 163 + ], + "type": "inline_equation", + "content": "x_{T}" + }, + { + "bbox": [ + 47, + 91, + 287, + 163 + ], + "type": "text", + "content": " over " + }, + { + "bbox": [ + 47, + 91, + 287, + 163 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 47, + 91, + 287, + 163 + ], + "type": "text", + "content": " steps to generate a realistic image " + }, + { + "bbox": [ + 47, + 91, + 287, + 163 + ], + "type": "inline_equation", + "content": "x_0" + }, + { + "bbox": [ + 47, + 91, + 287, + 163 + ], + "type": "text", + "content": " that conforms to the input condition " + }, + { + "bbox": [ + 47, + 91, + 287, + 163 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 47, + 91, + 287, + 163 + ], + "type": "text", + "content": " through:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 73, + 172, + 260, + 185 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 73, + 172, + 260, + 185 + ], + "spans": [ + { + "bbox": [ + 73, + 172, + 260, + 185 + ], + "type": "interline_equation", + "content": "x _ {t - 1} = \\epsilon_ {\\theta} (x _ {t}, c) \\quad \\text {f o r} t = T, T - 1, \\dots , 1,", + "image_path": "2e8f41e073afd87660fdbfdb8e3c30f0435724688710da4e6d84776acacad2cd.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 193, + 287, + 348 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 193, + 287, + 348 + ], + "spans": [ + { + "bbox": [ + 47, + 193, + 287, + 348 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 193, + 287, + 348 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}" + }, + { + "bbox": [ + 47, + 193, + 287, + 348 + ], + "type": "text", + "content": " is the denoising network. Standard diffusion models operate in the pixel space, but others like Latent Diffusion Models [39] instead operate on a latent image encoding space to reduce memory footprints. A U-Net architecture is commonly used to implement the denoising network and is typically trained on images at a fixed resolution " + }, + { + "bbox": [ + 47, + 193, + 287, + 348 + ], + "type": "inline_equation", + "content": "H\\times W" + }, + { + "bbox": [ + 47, + 193, + 287, + 348 + ], + "type": "text", + "content": " [8, 39, 43]. The convolutional architecture of a U-Net allows for inputs and outputs of arbitrary spatial dimensions. In order to generate an image of a different size " + }, + { + "bbox": [ + 47, + 193, + 287, + 348 + ], + "type": "inline_equation", + "content": "\\bar{H}\\times \\bar{W}" + }, + { + "bbox": [ + 47, + 193, + 287, + 348 + ], + "type": "text", + "content": " at inference time, one can sample an initial noise variable " + }, + { + "bbox": [ + 47, + 193, + 287, + 348 + ], + "type": "inline_equation", + "content": "\\bar{x}_T\\in \\mathbb{R}^{\\bar{H}\\times \\bar{W}\\times 3}" + }, + { + "bbox": [ + 47, + 193, + 287, + 348 + ], + "type": "text", + "content": " and follow the same diffusion process. However, we find that this works poorly in practice, resulting in a significant degradation in output quality." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 349, + 287, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 349, + 287, + 384 + ], + "spans": [ + { + "bbox": [ + 47, + 349, + 287, + 384 + ], + "type": "text", + "content": "Denoising Diffusion Implicit Models (DDIMs) introduce a faster non-Markovian sampling strategy, bypassing denoising steps. The reverse diffusion step in DDIM is:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 400, + 287, + 462 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 400, + 287, + 462 + ], + "spans": [ + { + "bbox": [ + 53, + 400, + 287, + 462 + ], + "type": "interline_equation", + "content": "x _ {t - 1} = \\sqrt {\\bar {\\alpha} _ {t - 1}} \\underbrace {\\left(x _ {t} - \\sqrt {1 - \\bar {\\alpha} _ {t} \\epsilon_ {\\theta} ^ {(t)} \\left(x _ {t}\\right)}\\right)} _ {\\text {p r e d i c t e d} \\hat {x} _ {0} ^ {t}} + \\underbrace {\\sqrt {1 - \\bar {\\alpha} _ {t - 1}} ,} _ {\\text {d i r e c t i o n p o i n t i n g t o} x _ {t}}, \\tag {1}", + "image_path": "c6c5422cdf67e3c0a570d684f93cc3568b918096bd3f6857a38898dbbc8868e0.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 469, + 287, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 469, + 287, + 529 + ], + "spans": [ + { + "bbox": [ + 47, + 469, + 287, + 529 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 469, + 287, + 529 + ], + "type": "inline_equation", + "content": "\\bar{\\alpha}_t = \\prod_{i=1}^t 1 - \\beta_i" + }, + { + "bbox": [ + 47, + 469, + 287, + 529 + ], + "type": "text", + "content": " is a cumulative product of the noise levels using a predetermined variance schedule " + }, + { + "bbox": [ + 47, + 469, + 287, + 529 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 47, + 469, + 287, + 529 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 47, + 469, + 287, + 529 + ], + "type": "inline_equation", + "content": "\\hat{x}_0^t" + }, + { + "bbox": [ + 47, + 469, + 287, + 529 + ], + "type": "text", + "content": " is a noise-free estimation of " + }, + { + "bbox": [ + 47, + 469, + 287, + 529 + ], + "type": "inline_equation", + "content": "x_t" + }, + { + "bbox": [ + 47, + 469, + 287, + 529 + ], + "type": "text", + "content": " obtained by subtracting the predicted noise scaled for step " + }, + { + "bbox": [ + 47, + 469, + 287, + 529 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 47, + 469, + 287, + 529 + ], + "type": "text", + "content": ". We assume a deterministic sampling process in the DDIM formula [48] for simplicity." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 529, + 287, + 554 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 529, + 287, + 554 + ], + "spans": [ + { + "bbox": [ + 47, + 529, + 287, + 554 + ], + "type": "text", + "content": "Classifier-Free Guidance updates the reverse diffusion process to condition it on a given condition " + }, + { + "bbox": [ + 47, + 529, + 287, + 554 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 47, + 529, + 287, + 554 + ], + "type": "text", + "content": " as:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 52, + 571, + 287, + 608 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 52, + 571, + 287, + 608 + ], + "spans": [ + { + "bbox": [ + 52, + 571, + 287, + 608 + ], + "type": "interline_equation", + "content": "\\hat {\\epsilon} _ {\\theta} ^ {(t)} (x _ {t}) = \\underbrace {\\epsilon_ {\\theta} ^ {(t)} (x _ {t})} _ {\\text {u n c o n d i t i o n a l}} + (1 + w) \\cdot \\underbrace {\\left(\\epsilon_ {\\theta} ^ {(t)} (x _ {t} , c) - \\epsilon_ {\\theta} ^ {(t)} (x _ {t})\\right)} _ {\\Delta_ {\\mathcal {C}} (x, c): \\text {c l a s s d i r e c t i o n s c o r e}}. \\tag {2}", + "image_path": "d8556583f8e3072cc59d8d5ffba9daea8b21f8292dd5c92adcd9952aacf54e48.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 614, + 287, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 614, + 287, + 664 + ], + "spans": [ + { + "bbox": [ + 47, + 614, + 287, + 664 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}(x,c)" + }, + { + "bbox": [ + 47, + 614, + 287, + 664 + ], + "type": "text", + "content": " is a pretrained conditional diffusion model, and " + }, + { + "bbox": [ + 47, + 614, + 287, + 664 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 47, + 614, + 287, + 664 + ], + "type": "text", + "content": " is a scaling factor. The difference between the conditional " + }, + { + "bbox": [ + 47, + 614, + 287, + 664 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}(x,c)" + }, + { + "bbox": [ + 47, + 614, + 287, + 664 + ], + "type": "text", + "content": " and unconditional " + }, + { + "bbox": [ + 47, + 614, + 287, + 664 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}(x)" + }, + { + "bbox": [ + 47, + 614, + 287, + 664 + ], + "type": "text", + "content": " scores, denoted as class direction score, gives the guidance direction towards " + }, + { + "bbox": [ + 47, + 614, + 287, + 664 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 47, + 614, + 287, + 664 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 669, + 144, + 681 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 669, + 144, + 681 + ], + "spans": [ + { + "bbox": [ + 47, + 669, + 144, + 681 + ], + "type": "text", + "content": "4. ElasticDiffusion" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "type": "text", + "content": "The aim of this work is to develop a method capable of synthesizing images at arbitrary size " + }, + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\bar{H} \\times \\bar{W}" + }, + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "type": "text", + "content": ", and conforming" + } + ] + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 309, + 72, + 545, + 168 + ], + "blocks": [ + { + "bbox": [ + 309, + 72, + 545, + 168 + ], + "lines": [ + { + "bbox": [ + 309, + 72, + 545, + 168 + ], + "spans": [ + { + "bbox": [ + 309, + 72, + 545, + 168 + ], + "type": "image", + "image_path": "2170657f1a26f15a373285c892f64f17278698b15e753998df65798b2b4c97dd.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 171, + 545, + 204 + ], + "lines": [ + { + "bbox": [ + 305, + 171, + 545, + 204 + ], + "spans": [ + { + "bbox": [ + 305, + 171, + 545, + 204 + ], + "type": "text", + "content": "Figure 2. PCA of diffusion scores: class-direction score (top) dictates global content by clustering on semantic parts, while the unconditional score (bottom) lacks pixel correlations." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 210, + 545, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 210, + 545, + 653 + ], + "spans": [ + { + "bbox": [ + 304, + 210, + 545, + 653 + ], + "type": "text", + "content": "to a global condition " + }, + { + "bbox": [ + 304, + 210, + 545, + 653 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 304, + 210, + 545, + 653 + ], + "type": "text", + "content": " using a pre-trained diffusion model that is limited at inference to its training resolution " + }, + { + "bbox": [ + 304, + 210, + 545, + 653 + ], + "type": "inline_equation", + "content": "H \\times W" + }, + { + "bbox": [ + 304, + 210, + 545, + 653 + ], + "type": "text", + "content": ". To achieve this, we observe a main insight with respect to the diffusion scores in Eq. (2) that we visualize in Fig. 2. The class direction score " + }, + { + "bbox": [ + 304, + 210, + 545, + 653 + ], + "type": "inline_equation", + "content": "(\\Delta_c)" + }, + { + "bbox": [ + 304, + 210, + 545, + 653 + ], + "type": "text", + "content": " primarily dictates the image's overall composition by showing clustering along semantic regions (e.g. hair) and edges. Thus, upscaling this score maintains global content in higher resolutions. In contrast, the unconditional score lacks inter-pixel correlation, suggesting a localized influence in enhancing detail at the pixel level. Thus, computing it requires a pixel-level precision. We study the global and local behaviour of these scores in Supp. Sec. 2.1. Our intuition behind this observation is that latent pixels, specifically in early diffusion steps where noise is high, display weak and short-range correlations. This leads the unconditional score to focus on local patterns due to these limited pixel interactions. Meanwhile, as " + }, + { + "bbox": [ + 304, + 210, + 545, + 653 + ], + "type": "inline_equation", + "content": "\\Delta_c" + }, + { + "bbox": [ + 304, + 210, + 545, + 653 + ], + "type": "text", + "content": " represents a latent direction towards a specified global condition, it exerts a global influence by guiding clusters of pixels with similar directions towards semantic regions, overriding their initial noisy state. Leveraging this insight, we propose a method to decouple the generation of local and global content during the diffusion process. Specifically, we compute the unconditional score on local patches of size " + }, + { + "bbox": [ + 304, + 210, + 545, + 653 + ], + "type": "inline_equation", + "content": "H \\times W" + }, + { + "bbox": [ + 304, + 210, + 545, + 653 + ], + "type": "text", + "content": " while simultaneously resizing a class direction score, originally derived for a reference latent of the dimensions " + }, + { + "bbox": [ + 304, + 210, + 545, + 653 + ], + "type": "inline_equation", + "content": "H \\times W" + }, + { + "bbox": [ + 304, + 210, + 545, + 653 + ], + "type": "text", + "content": " as well. This dual strategy facilitates the generation of images at varied sizes using a pretrained diffusion model at its native resolution, all while maintaining the same memory requirement and without further training. We first present our approach for computing the unconditional score (Sec. 4.1). We then detail our method to estimate (Sec. 4.2) and upscale the resolution (Sec. 4.3) of the class direction score. Finally, we combine the two estimated scores with a novel guidance strategy to generate images at arbitrary sizes (Sec. 4.4). The generation process of ElasticDiffusion is illustrated in Fig. 3." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 660, + 495, + 673 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 660, + 495, + 673 + ], + "spans": [ + { + "bbox": [ + 306, + 660, + 495, + 673 + ], + "type": "text", + "content": "4.1. Estimating the Unconditional Score" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 305, + 677, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 677, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 305, + 677, + 545, + 714 + ], + "type": "text", + "content": "Building upon previous work [3, 53, 63], we estimate the unconditional score for a large latent signal " + }, + { + "bbox": [ + 305, + 677, + 545, + 714 + ], + "type": "inline_equation", + "content": "\\bar{x}_t\\in \\mathbb{R}^{\\bar{H}\\times \\bar{W}\\times 3}" + }, + { + "bbox": [ + 305, + 677, + 545, + 714 + ], + "type": "text", + "content": " by concatenating scores derived from local patches. Specif" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6605" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 72, + 70, + 521, + 331 + ], + "blocks": [ + { + "bbox": [ + 72, + 70, + 521, + 331 + ], + "lines": [ + { + "bbox": [ + 72, + 70, + 521, + 331 + ], + "spans": [ + { + "bbox": [ + 72, + 70, + 521, + 331 + ], + "type": "image", + "image_path": "ccc04c596e6212ceff276749c13a60680845a2627e4d227da6bd28139b0c0b42.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 339, + 547, + 374 + ], + "lines": [ + { + "bbox": [ + 46, + 339, + 547, + 374 + ], + "spans": [ + { + "bbox": [ + 46, + 339, + 547, + 374 + ], + "type": "text", + "content": "Figure 3. Illustration of ElasticDiffusion: We generate images at various sizes by generating local and global content separately. For local content, we partition the latent " + }, + { + "bbox": [ + 46, + 339, + 547, + 374 + ], + "type": "inline_equation", + "content": "\\bar{x}_t" + }, + { + "bbox": [ + 46, + 339, + 547, + 374 + ], + "type": "text", + "content": " into non-overlapping patches " + }, + { + "bbox": [ + 46, + 339, + 547, + 374 + ], + "type": "inline_equation", + "content": "p_k" + }, + { + "bbox": [ + 46, + 339, + 547, + 374 + ], + "type": "text", + "content": ", each concatenated with context " + }, + { + "bbox": [ + 46, + 339, + 547, + 374 + ], + "type": "inline_equation", + "content": "c_k" + }, + { + "bbox": [ + 46, + 339, + 547, + 374 + ], + "type": "text", + "content": " to estimate their unconditional score. For global content, we downsample " + }, + { + "bbox": [ + 46, + 339, + 547, + 374 + ], + "type": "inline_equation", + "content": "\\bar{x}_t" + }, + { + "bbox": [ + 46, + 339, + 547, + 374 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 46, + 339, + 547, + 374 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_t" + }, + { + "bbox": [ + 46, + 339, + 547, + 374 + ], + "type": "text", + "content": ", pad to a square size " + }, + { + "bbox": [ + 46, + 339, + 547, + 374 + ], + "type": "inline_equation", + "content": "(\\hat{\\mathbf{x}}_t)" + }, + { + "bbox": [ + 46, + 339, + 547, + 374 + ], + "type": "text", + "content": ", compute class-direction score " + }, + { + "bbox": [ + 46, + 339, + 547, + 374 + ], + "type": "inline_equation", + "content": "(\\Delta_c)" + }, + { + "bbox": [ + 46, + 339, + 547, + 374 + ], + "type": "text", + "content": ", and upscale to match " + }, + { + "bbox": [ + 46, + 339, + 547, + 374 + ], + "type": "inline_equation", + "content": "\\bar{x}_t" + }, + { + "bbox": [ + 46, + 339, + 547, + 374 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 392, + 289, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 392, + 289, + 594 + ], + "spans": [ + { + "bbox": [ + 46, + 392, + 289, + 594 + ], + "type": "text", + "content": "ically, we divide image " + }, + { + "bbox": [ + 46, + 392, + 289, + 594 + ], + "type": "inline_equation", + "content": "\\bar{x}_t" + }, + { + "bbox": [ + 46, + 392, + 289, + 594 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 46, + 392, + 289, + 594 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 46, + 392, + 289, + 594 + ], + "type": "text", + "content": " patches " + }, + { + "bbox": [ + 46, + 392, + 289, + 594 + ], + "type": "inline_equation", + "content": "P_{k} \\in \\mathbb{R}^{H \\times W \\times 3}" + }, + { + "bbox": [ + 46, + 392, + 289, + 594 + ], + "type": "text", + "content": " and compute the score as " + }, + { + "bbox": [ + 46, + 392, + 289, + 594 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}(\\bar{x}_t) = \\{\\epsilon_{\\theta}(P_k); \\forall k \\leq K\\}" + }, + { + "bbox": [ + 46, + 392, + 289, + 594 + ], + "type": "text", + "content": ". A common challenge encountered with this implementation is discontinuities at boundaries, as illustrated in Fig. 4 (A). To address this, earlier research calculated the diffusion model score on explicitly overlapping patches and averaged their scores in the intersecting regions [3, 53, 63]. While this strategy mitigates discontinuities, it requires large overlap between patches, thereby substantially increasing inference time and blurring details. To speed up the process, we introduce a more effective method that enhances boundary transitions in local patches by incorporating contextual information from adjacent patches, thus negating the need for signal averaging in overlapping areas. Specifically, we select patches smaller than the full size " + }, + { + "bbox": [ + 46, + 392, + 289, + 594 + ], + "type": "inline_equation", + "content": "p_{k} \\in \\mathbb{R}^{h \\times w \\times 3}" + }, + { + "bbox": [ + 46, + 392, + 289, + 594 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 46, + 392, + 289, + 594 + ], + "type": "inline_equation", + "content": "h < H" + }, + { + "bbox": [ + 46, + 392, + 289, + 594 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 392, + 289, + 594 + ], + "type": "inline_equation", + "content": "w < W" + }, + { + "bbox": [ + 46, + 392, + 289, + 594 + ], + "type": "text", + "content": ", and concatenate them with context pixels from adjacent patches, denoted as " + }, + { + "bbox": [ + 46, + 392, + 289, + 594 + ], + "type": "inline_equation", + "content": "c_{k} \\in \\mathbb{R}^{(H - h) \\times (W - w) \\times 3}" + }, + { + "bbox": [ + 46, + 392, + 289, + 594 + ], + "type": "text", + "content": ", to compute the diffusion model unconditional score " + }, + { + "bbox": [ + 46, + 392, + 289, + 594 + ], + "type": "inline_equation", + "content": "\\mathbf{S}_{u}" + }, + { + "bbox": [ + 46, + 392, + 289, + 594 + ], + "type": "text", + "content": " as:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 82, + 605, + 287, + 618 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 82, + 605, + 287, + 618 + ], + "spans": [ + { + "bbox": [ + 82, + 605, + 287, + 618 + ], + "type": "interline_equation", + "content": "\\tilde {\\epsilon} _ {\\theta} \\left(x _ {t}\\right) = \\left\\{\\epsilon_ {\\theta} \\left(p _ {k} \\mid c _ {k}\\right) \\mid \\bar {x} _ {t} = \\left\\{p _ {k}; \\forall_ {k \\leq K} \\right\\} \\right\\}, \\tag {3}", + "image_path": "ded14ec3e71dbd64ddfcd9bae72adb05d1953593c0a7e1b84b4be798a72eb47d.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 629, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 629, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 47, + 629, + 288, + 715 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 629, + 288, + 715 + ], + "type": "inline_equation", + "content": "p_i" + }, + { + "bbox": [ + 47, + 629, + 288, + 715 + ], + "type": "text", + "content": " is a local patch and " + }, + { + "bbox": [ + 47, + 629, + 288, + 715 + ], + "type": "inline_equation", + "content": "c_i" + }, + { + "bbox": [ + 47, + 629, + 288, + 715 + ], + "type": "text", + "content": " are the context pixels surrounding it. This substantially increases the efficiency of the process. For instance, to generate an image of size " + }, + { + "bbox": [ + 47, + 629, + 288, + 715 + ], + "type": "inline_equation", + "content": "1024 \\times 1024" + }, + { + "bbox": [ + 47, + 629, + 288, + 715 + ], + "type": "text", + "content": ", previous methods [3, 53, 63] used " + }, + { + "bbox": [ + 47, + 629, + 288, + 715 + ], + "type": "inline_equation", + "content": "87.5\\%" + }, + { + "bbox": [ + 47, + 629, + 288, + 715 + ], + "type": "text", + "content": " overlap between adjacent patches, necessitating 81 forward diffusion calls per decoding step. In comparison, ElasticDiffusion achieves comparable results with only 9 forward calls," + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 393, + 545, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 393, + 545, + 430 + ], + "spans": [ + { + "bbox": [ + 304, + 393, + 545, + 430 + ], + "type": "text", + "content": "as demonstrated in Fig. 4 (D). Employing the same number of calls, previous techniques result in obvious boundary discontinuities, as depicted in Fig. 4 (B)." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 305, + 436, + 500, + 449 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 436, + 500, + 449 + ], + "spans": [ + { + "bbox": [ + 305, + 436, + 500, + 449 + ], + "type": "text", + "content": "4.2. Estimating the Class Direction Score" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 455, + 545, + 587 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 455, + 545, + 587 + ], + "spans": [ + { + "bbox": [ + 304, + 455, + 545, + 587 + ], + "type": "text", + "content": "A simple approach to estimate a class direction score of an intermediate latent signal " + }, + { + "bbox": [ + 304, + 455, + 545, + 587 + ], + "type": "inline_equation", + "content": "\\bar{x}_t \\in \\mathbb{R}^{\\bar{H} \\times \\bar{W} \\times 3}" + }, + { + "bbox": [ + 304, + 455, + 545, + 587 + ], + "type": "text", + "content": " is to upsample the score from a reference latent " + }, + { + "bbox": [ + 304, + 455, + 545, + 587 + ], + "type": "inline_equation", + "content": "x_t \\in \\mathbb{R}^{N \\times M \\times 3}" + }, + { + "bbox": [ + 304, + 455, + 545, + 587 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 304, + 455, + 545, + 587 + ], + "type": "inline_equation", + "content": "\\bar{H} \\times \\bar{W}" + }, + { + "bbox": [ + 304, + 455, + 545, + 587 + ], + "type": "text", + "content": ". This is possible due to our observation that the class direction score represents a latent direction that can be shared between nearby pixels. We validate this observation empirically in Supplementary. We choose " + }, + { + "bbox": [ + 304, + 455, + 545, + 587 + ], + "type": "inline_equation", + "content": "N < H" + }, + { + "bbox": [ + 304, + 455, + 545, + 587 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 455, + 545, + 587 + ], + "type": "inline_equation", + "content": "M < W" + }, + { + "bbox": [ + 304, + 455, + 545, + 587 + ], + "type": "text", + "content": " such that " + }, + { + "bbox": [ + 304, + 455, + 545, + 587 + ], + "type": "inline_equation", + "content": "N \\times M" + }, + { + "bbox": [ + 304, + 455, + 545, + 587 + ], + "type": "text", + "content": " is as close as possible to " + }, + { + "bbox": [ + 304, + 455, + 545, + 587 + ], + "type": "inline_equation", + "content": "H \\times W" + }, + { + "bbox": [ + 304, + 455, + 545, + 587 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 455, + 545, + 587 + ], + "type": "inline_equation", + "content": "\\frac{\\bar{H}}{W} = \\frac{N}{M}" + }, + { + "bbox": [ + 304, + 455, + 545, + 587 + ], + "type": "text", + "content": ". This is important to maintain the aspect ratio and prevent stretching the global content. Formally, we compute the class direction score " + }, + { + "bbox": [ + 304, + 455, + 545, + 587 + ], + "type": "inline_equation", + "content": "\\mathbf{S}_d" + }, + { + "bbox": [ + 304, + 455, + 545, + 587 + ], + "type": "text", + "content": " as:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 334, + 594, + 517, + 621 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 334, + 594, + 517, + 621 + ], + "spans": [ + { + "bbox": [ + 334, + 594, + 517, + 621 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathbf {x} _ {t} \\leftarrow \\text {D o w n s a m p l e} (\\bar {x} _ {t}, N \\times M), \\\\ \\Delta_ {\\mathcal {C}} (\\bar {x} _ {t}, c) = \\operatorname {U p s a m p l e} \\left(\\Delta_ {\\mathcal {C}} \\left(\\mathbf {x} _ {t}, c\\right), \\bar {H} \\times \\bar {W}\\right), \\\\ \\end{array}", + "image_path": "2bd4ef904ef6e90268ba7b629d23979008d01592cf045d6bf4c16ab6b9effad3.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 531, + 606, + 544, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 531, + 606, + 544, + 616 + ], + "spans": [ + { + "bbox": [ + 531, + 606, + 544, + 616 + ], + "type": "text", + "content": "(4)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\Delta_{\\mathcal{C}}(.)" + }, + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "text", + "content": " is the class direction score from Eq. (2), Downsample and Upsample are downsampling and upsampling operations. We use a nearest-neighbors approach to prevent altering the statistics of the latent signal. In order to maintain the input to the diffusion models at the size " + }, + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "inline_equation", + "content": "H\\times W" + }, + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "text", + "content": " we dynamically pad the downsampled latent " + }, + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_t" + }, + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "text", + "content": " using a random background with a constant color." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "6606" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 55, + 71, + 165, + 122 + ], + "blocks": [ + { + "bbox": [ + 55, + 71, + 165, + 122 + ], + "lines": [ + { + "bbox": [ + 55, + 71, + 165, + 122 + ], + "spans": [ + { + "bbox": [ + 55, + 71, + 165, + 122 + ], + "type": "image", + "image_path": "33797549b612d2d3da94f381f50662943123a329d503c4b31cf3843b6a85e416.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 168, + 72, + 280, + 122 + ], + "blocks": [ + { + "bbox": [ + 168, + 72, + 280, + 122 + ], + "lines": [ + { + "bbox": [ + 168, + 72, + 280, + 122 + ], + "spans": [ + { + "bbox": [ + 168, + 72, + 280, + 122 + ], + "type": "image", + "image_path": "3469265d4222a88721addc3eb441fe98dd1f99cb17554f84889c04b4a31b6bba.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 55, + 124, + 164, + 174 + ], + "blocks": [ + { + "bbox": [ + 55, + 124, + 164, + 174 + ], + "lines": [ + { + "bbox": [ + 55, + 124, + 164, + 174 + ], + "spans": [ + { + "bbox": [ + 55, + 124, + 164, + 174 + ], + "type": "image", + "image_path": "5487e88e0556e1adb268bd3e0d86267314fc29918e6a34d0891cc14fe3629a50.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 168, + 123, + 280, + 174 + ], + "blocks": [ + { + "bbox": [ + 168, + 123, + 280, + 174 + ], + "lines": [ + { + "bbox": [ + 168, + 123, + 280, + 174 + ], + "spans": [ + { + "bbox": [ + 168, + 123, + 280, + 174 + ], + "type": "image", + "image_path": "268bb81e8397b84389674ad6e464b572f4b5659ae0bdb4274b8f5ad1f87e057a.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 55, + 180, + 164, + 230 + ], + "blocks": [ + { + "bbox": [ + 55, + 180, + 164, + 230 + ], + "lines": [ + { + "bbox": [ + 55, + 180, + 164, + 230 + ], + "spans": [ + { + "bbox": [ + 55, + 180, + 164, + 230 + ], + "type": "image", + "image_path": "366814e38e08459d15ca539142fbb5864edc21686a6e85cd7ae51e9236dd3cc3.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 168, + 179, + 280, + 230 + ], + "blocks": [ + { + "bbox": [ + 168, + 179, + 280, + 230 + ], + "lines": [ + { + "bbox": [ + 168, + 179, + 280, + 230 + ], + "spans": [ + { + "bbox": [ + 168, + 179, + 280, + 230 + ], + "type": "image", + "image_path": "04299a38c6bfe4e4aa1b6da90de60d081fdf9e6afb77e272fe224f020dc99136.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 55, + 236, + 164, + 285 + ], + "blocks": [ + { + "bbox": [ + 55, + 236, + 164, + 285 + ], + "lines": [ + { + "bbox": [ + 55, + 236, + 164, + 285 + ], + "spans": [ + { + "bbox": [ + 55, + 236, + 164, + 285 + ], + "type": "image", + "image_path": "c91b5e67b6f85819184ca6bad9e0ed4de0f089a461e40f59fafae3a36ddd5150.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 96, + 286, + 134, + 293 + ], + "lines": [ + { + "bbox": [ + 96, + 286, + 134, + 293 + ], + "spans": [ + { + "bbox": [ + 96, + 286, + 134, + 293 + ], + "type": "text", + "content": "Intermediate step" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 302, + 287, + 369 + ], + "lines": [ + { + "bbox": [ + 46, + 302, + 287, + 369 + ], + "spans": [ + { + "bbox": [ + 46, + 302, + 287, + 369 + ], + "type": "text", + "content": "Figure 4. Comparing strategies for calculating diffusion model score on a local patch. No overlap between adjacent patches (A) leads to discontinuities at the boundaries. Strategies (B) and (C), explicitly overlap nearby patches, necessitating substantial overlap to be effective. Our implicit overlapping method (D) achieves superior results with computational demand similar to (B)." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 168, + 234, + 280, + 285 + ], + "blocks": [ + { + "bbox": [ + 168, + 234, + 280, + 285 + ], + "lines": [ + { + "bbox": [ + 168, + 234, + 280, + 285 + ], + "spans": [ + { + "bbox": [ + 168, + 234, + 280, + 285 + ], + "type": "image", + "image_path": "974809efb764c7055cac428db041fcf6e0d341f2c45eab564f16e7ed4255538d.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 201, + 286, + 239, + 293 + ], + "lines": [ + { + "bbox": [ + 201, + 286, + 239, + 293 + ], + "spans": [ + { + "bbox": [ + 201, + 286, + 239, + 293 + ], + "type": "text", + "content": "Generated Image" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 376, + 287, + 435 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 376, + 287, + 435 + ], + "spans": [ + { + "bbox": [ + 46, + 376, + 287, + 435 + ], + "type": "text", + "content": "This encourages the model to concentrate content generation within the center area. We then crop the extended parts from the predicted noise to the target image resolution " + }, + { + "bbox": [ + 46, + 376, + 287, + 435 + ], + "type": "inline_equation", + "content": "N \\times M" + }, + { + "bbox": [ + 46, + 376, + 287, + 435 + ], + "type": "text", + "content": ". Formally we modify the forward call for the reverse diffusion step as:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 90, + 445, + 259, + 459 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 90, + 445, + 259, + 459 + ], + "spans": [ + { + "bbox": [ + 90, + 445, + 259, + 459 + ], + "type": "interline_equation", + "content": "\\dot {\\mathbf {x}} _ {t} \\leftarrow \\operatorname {P a d} \\left(\\mathbf {x} _ {t}, \\mathcal {A} _ {t} \\sqrt {\\bar {\\alpha} _ {t}} + \\sqrt {1 - \\bar {\\alpha} _ {t}} \\cdot \\mathcal {Z} _ {t}\\right),", + "image_path": "92314ee7cf802033fbbfbe06be32676afb8518d62edd901b9019421faf97a4d3.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 64, + 460, + 287, + 473 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 460, + 287, + 473 + ], + "spans": [ + { + "bbox": [ + 64, + 460, + 287, + 473 + ], + "type": "interline_equation", + "content": "\\epsilon_ {\\theta} \\left(\\mathbf {x} _ {t}, c\\right) = \\operatorname {C r o p} \\left(\\epsilon_ {\\theta} \\left(\\hat {\\mathbf {x}} _ {t}, c\\right), N \\times M\\right), \\tag {5}", + "image_path": "2fc9395cc3b6b4993319f0953cb399ad32bc47fb03dc1734076f7e1be9bc38e4.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 46, + 483, + 287, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 483, + 287, + 578 + ], + "spans": [ + { + "bbox": [ + 46, + 483, + 287, + 578 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 483, + 287, + 578 + ], + "type": "inline_equation", + "content": "\\mathcal{Z}_t\\sim \\mathcal{N}(0,I)" + }, + { + "bbox": [ + 46, + 483, + 287, + 578 + ], + "type": "text", + "content": " represents the injected Gaussian noise at each step, and " + }, + { + "bbox": [ + 46, + 483, + 287, + 578 + ], + "type": "inline_equation", + "content": "\\mathcal{A}_t" + }, + { + "bbox": [ + 46, + 483, + 287, + 578 + ], + "type": "text", + "content": " represents a background image of size " + }, + { + "bbox": [ + 46, + 483, + 287, + 578 + ], + "type": "inline_equation", + "content": "(H - N)\\times (W - M)" + }, + { + "bbox": [ + 46, + 483, + 287, + 578 + ], + "type": "text", + "content": " with a constant color value " + }, + { + "bbox": [ + 46, + 483, + 287, + 578 + ], + "type": "inline_equation", + "content": "\\mathcal{V}" + }, + { + "bbox": [ + 46, + 483, + 287, + 578 + ], + "type": "text", + "content": " randomly sampled from a uniform distribution. This simple padding operation guarantees that the input to the diffusion model is kept at " + }, + { + "bbox": [ + 46, + 483, + 287, + 578 + ], + "type": "inline_equation", + "content": "H\\times W" + }, + { + "bbox": [ + 46, + 483, + 287, + 578 + ], + "type": "text", + "content": ", while encouraging the diffusion model to keep the generated content within the cropped " + }, + { + "bbox": [ + 46, + 483, + 287, + 578 + ], + "type": "inline_equation", + "content": "N\\times M" + }, + { + "bbox": [ + 46, + 483, + 287, + 578 + ], + "type": "text", + "content": " center that we are interested in." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 47, + 587, + 209, + 598 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 587, + 209, + 598 + ], + "spans": [ + { + "bbox": [ + 47, + 587, + 209, + 598 + ], + "type": "text", + "content": "4.3. Refined Class Direction Score" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "type": "text", + "content": "Sharing the class direction score among nearby pixels can result in over-smoothed images. To mitigate this, we propose an iterative resampling technique that increases the resolution of the estimated class direction score by extrapolating missing image components from their surrounding context, following [29]. Our technique involves a gradual enhancement of the class direction score's resolution by estimating and integrating it for newly sampled pixels. Specifically, in each iteration, we replace " + }, + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "type": "inline_equation", + "content": "n\\%" + }, + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "type": "text", + "content": " of the pixels" + } + ] + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 307, + 72, + 545, + 125 + ], + "blocks": [ + { + "bbox": [ + 307, + 72, + 545, + 125 + ], + "lines": [ + { + "bbox": [ + 307, + 72, + 545, + 125 + ], + "spans": [ + { + "bbox": [ + 307, + 72, + 545, + 125 + ], + "type": "image", + "image_path": "f1eecfec9c801373a03fa89b89c30f1b9f0ed689a4b8c335368fe9440d1c942b.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 307, + 182, + 545, + 218 + ], + "blocks": [ + { + "bbox": [ + 305, + 132, + 545, + 176 + ], + "lines": [ + { + "bbox": [ + 305, + 132, + 545, + 176 + ], + "spans": [ + { + "bbox": [ + 305, + 132, + 545, + 176 + ], + "type": "text", + "content": "Figure 5. The Effect of Reduced-Resolution Guidance (RRG). Higher RRG weights effectively eliminates emerging artifacts albeit at the cost of slightly blurrier outputs. " + }, + { + "bbox": [ + 305, + 132, + 545, + 176 + ], + "type": "inline_equation", + "content": "\\delta = 200" + }, + { + "bbox": [ + 305, + 132, + 545, + 176 + ], + "type": "text", + "content": " strikes a good balance. Improvements are more noticeable when zooming in." + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 307, + 182, + 545, + 218 + ], + "lines": [ + { + "bbox": [ + 307, + 182, + 545, + 218 + ], + "spans": [ + { + "bbox": [ + 307, + 182, + 545, + 218 + ], + "type": "image", + "image_path": "4c3a5bc0c419ed0226ee95dc439211064e8c2d0018f3862702e4231afca6093b.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 306, + 219, + 545, + 239 + ], + "lines": [ + { + "bbox": [ + 306, + 219, + 545, + 239 + ], + "spans": [ + { + "bbox": [ + 306, + 219, + 545, + 239 + ], + "type": "text", + "content": "Figure 6. Effect of Resampling. Applying more resampling steps noticeably enhances detail in the generated images." + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "bbox": [ + 305, + 262, + 545, + 322 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 262, + 545, + 322 + ], + "spans": [ + { + "bbox": [ + 305, + 262, + 545, + 322 + ], + "type": "text", + "content": "in " + }, + { + "bbox": [ + 305, + 262, + 545, + 322 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_t" + }, + { + "bbox": [ + 305, + 262, + 545, + 322 + ], + "type": "text", + "content": " with newly sampled ones from " + }, + { + "bbox": [ + 305, + 262, + 545, + 322 + ], + "type": "inline_equation", + "content": "x_{t}" + }, + { + "bbox": [ + 305, + 262, + 545, + 322 + ], + "type": "text", + "content": " to get an updated version " + }, + { + "bbox": [ + 305, + 262, + 545, + 322 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{x}}_t" + }, + { + "bbox": [ + 305, + 262, + 545, + 322 + ], + "type": "text", + "content": ". Following each update, the direction score is recalculated and blended with the previously calculated score. Formally, we consider " + }, + { + "bbox": [ + 305, + 262, + 545, + 322 + ], + "type": "inline_equation", + "content": "\\mathbf{S}_d^0 = \\mathbf{S}_d" + }, + { + "bbox": [ + 305, + 262, + 545, + 322 + ], + "type": "text", + "content": " and define the iterative resampling as:" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 354, + 332, + 545, + 348 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 354, + 332, + 545, + 348 + ], + "spans": [ + { + "bbox": [ + 354, + 332, + 545, + 348 + ], + "type": "interline_equation", + "content": "\\mathbf {S} _ {d} ^ {r + 1} = \\mathbf {S} _ {d} ^ {r} \\odot m + \\tilde {\\mathbf {S}} _ {d} \\odot (1 - m), \\tag {6}", + "image_path": "0989bff3c86c3ac7dbba347313b4c53a1ee12a4f8a49d95a481db35bcb16f638.jpg" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 304, + 357, + 545, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 357, + 545, + 479 + ], + "spans": [ + { + "bbox": [ + 304, + 357, + 545, + 479 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 357, + 545, + 479 + ], + "type": "inline_equation", + "content": "m \\in \\{0,1\\}^{\\bar{H} \\times \\bar{W}}" + }, + { + "bbox": [ + 304, + 357, + 545, + 479 + ], + "type": "text", + "content": " is a mask with a value of 1 at the positions of the newly sampled pixels and 0 elsewhere. " + }, + { + "bbox": [ + 304, + 357, + 545, + 479 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{S}}_d" + }, + { + "bbox": [ + 304, + 357, + 545, + 479 + ], + "type": "text", + "content": " represents the recalculated class direction score on " + }, + { + "bbox": [ + 304, + 357, + 545, + 479 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{x}}_t" + }, + { + "bbox": [ + 304, + 357, + 545, + 479 + ], + "type": "text", + "content": " as per Eq. (4). This method estimates the class direction score of " + }, + { + "bbox": [ + 304, + 357, + 545, + 479 + ], + "type": "inline_equation", + "content": "n\\%" + }, + { + "bbox": [ + 304, + 357, + 545, + 479 + ], + "type": "text", + "content": " new pixels while retaining the information of the previously estimated score, thereby increasing the score's overall resolution. In our experiments, we set " + }, + { + "bbox": [ + 304, + 357, + 545, + 479 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 304, + 357, + 545, + 479 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 304, + 357, + 545, + 479 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 304, + 357, + 545, + 479 + ], + "type": "text", + "content": " and repeat the process " + }, + { + "bbox": [ + 304, + 357, + 545, + 479 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 304, + 357, + 545, + 479 + ], + "type": "text", + "content": " times, depending on the target generation resolution. Fig. 6 demonstrates the effectiveness of our method in enhancing the details of the generated images." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 306, + 487, + 504, + 499 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 487, + 504, + 499 + ], + "spans": [ + { + "bbox": [ + 306, + 487, + 504, + 499 + ], + "type": "text", + "content": "4.4. Reduced-Resolution Guidance (RRG)" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 304, + 505, + 545, + 638 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 505, + 545, + 638 + ], + "spans": [ + { + "bbox": [ + 304, + 505, + 545, + 638 + ], + "type": "text", + "content": "We effectively estimate the unconditional score signal and concurrently steer the image generation using the class direction score. However, inaccuracies in unconditional score estimation or fluctuations in local content, especially from distant patches, can lead to artifacts. To enhance image coherence and diminish artifacts, we consider a downsampled version of the latent at each decoding step as a reference and aim to align the decoded latent with it through our reduced-resolution latent update strategy. Specifically, we utilize the noise-free sample " + }, + { + "bbox": [ + 304, + 505, + 545, + 638 + ], + "type": "inline_equation", + "content": "\\hat{x}_0^t" + }, + { + "bbox": [ + 304, + 505, + 545, + 638 + ], + "type": "text", + "content": " of the latent " + }, + { + "bbox": [ + 304, + 505, + 545, + 638 + ], + "type": "inline_equation", + "content": "x_{t}" + }, + { + "bbox": [ + 304, + 505, + 545, + 638 + ], + "type": "text", + "content": " from Eq. (1) and generate a corresponding noise-free sample " + }, + { + "bbox": [ + 304, + 505, + 545, + 638 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{x}}_0^t" + }, + { + "bbox": [ + 304, + 505, + 545, + 638 + ], + "type": "text", + "content": " from its downsampled counterpart " + }, + { + "bbox": [ + 304, + 505, + 545, + 638 + ], + "type": "inline_equation", + "content": "\\mathbf{x}^t" + }, + { + "bbox": [ + 304, + 505, + 545, + 638 + ], + "type": "text", + "content": " in Eq. (4) as:" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 310, + 646, + 540, + 669 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 310, + 646, + 540, + 669 + ], + "spans": [ + { + "bbox": [ + 310, + 646, + 540, + 669 + ], + "type": "interline_equation", + "content": "\\hat {\\mathbf {x}} _ {0} ^ {t} = \\frac {1}{\\sqrt {\\bar {\\alpha} _ {t}}} (\\mathbf {x} _ {t} - \\sqrt {1 - \\bar {\\alpha} _ {t}} \\cdot (\\epsilon_ {\\theta} (\\mathbf {x} _ {t}) + (1 + w) \\cdot \\Delta_ {\\mathcal {C}} (\\mathbf {x} _ {t}, c))),", + "image_path": "8a08c7b3e6c359ef713c76c6510cba07ce5b8ce98ec23b60098df1be0d59f2bc.jpg" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 306, + 677, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 677, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 306, + 677, + 545, + 714 + ], + "type": "text", + "content": "Here, both " + }, + { + "bbox": [ + 306, + 677, + 545, + 714 + ], + "type": "inline_equation", + "content": "\\hat{x}_0^t" + }, + { + "bbox": [ + 306, + 677, + 545, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 306, + 677, + 545, + 714 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{x}}_0^t" + }, + { + "bbox": [ + 306, + 677, + 545, + 714 + ], + "type": "text", + "content": " corresponds to the same latent update at different resolutions. Due to its smaller dimension, " + }, + { + "bbox": [ + 306, + 677, + 545, + 714 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{x}}_0^t" + }, + { + "bbox": [ + 306, + 677, + 545, + 714 + ], + "type": "text", + "content": " has a broader context when computing the unconditional" + } + ] + } + ], + "index": 27 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6607" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 73, + 534, + 317 + ], + "blocks": [ + { + "bbox": [ + 59, + 73, + 534, + 317 + ], + "lines": [ + { + "bbox": [ + 59, + 73, + 534, + 317 + ], + "spans": [ + { + "bbox": [ + 59, + 73, + 534, + 317 + ], + "type": "image", + "image_path": "59eb54a5213e8dbd5fe4904d66b7320fc934009382dd44d7d1833a563cfd971f.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 320, + 546, + 366 + ], + "lines": [ + { + "bbox": [ + 46, + 320, + 546, + 366 + ], + "spans": [ + { + "bbox": [ + 46, + 320, + 546, + 366 + ], + "type": "text", + "content": "Figure 7. Qualitative comparison at various resolutions on CelebAHQ faces. ElasticDiffusion consistently generates coherent images at all resolutions. StableDiffusion, and MultiDiffusion produce repeating body parts at higher resolutions, while StableDiffusion-XL fails to maintain its quality at lower resolutions, resulting in noisy outputs at " + }, + { + "bbox": [ + 46, + 320, + 546, + 366 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 46, + 320, + 546, + 366 + ], + "type": "text", + "content": ". We exclude MultiDiffusion results at " + }, + { + "bbox": [ + 46, + 320, + 546, + 366 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 46, + 320, + 546, + 366 + ], + "type": "text", + "content": " as it is not designed to produce images at lower resolutions." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 385, + 288, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 385, + 288, + 422 + ], + "spans": [ + { + "bbox": [ + 46, + 385, + 288, + 422 + ], + "type": "text", + "content": "signal. To guide " + }, + { + "bbox": [ + 46, + 385, + 288, + 422 + ], + "type": "inline_equation", + "content": "x_{t}" + }, + { + "bbox": [ + 46, + 385, + 288, + 422 + ], + "type": "text", + "content": " with its downsampled reference, we refine the latent " + }, + { + "bbox": [ + 46, + 385, + 288, + 422 + ], + "type": "inline_equation", + "content": "x_{t}" + }, + { + "bbox": [ + 46, + 385, + 288, + 422 + ], + "type": "text", + "content": " with the direction that minimizes " + }, + { + "bbox": [ + 46, + 385, + 288, + 422 + ], + "type": "inline_equation", + "content": "L2" + }, + { + "bbox": [ + 46, + 385, + 288, + 422 + ], + "type": "text", + "content": " difference between " + }, + { + "bbox": [ + 46, + 385, + 288, + 422 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{x}}_0^t" + }, + { + "bbox": [ + 46, + 385, + 288, + 422 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 385, + 288, + 422 + ], + "type": "inline_equation", + "content": "\\hat{x}_0^t" + }, + { + "bbox": [ + 46, + 385, + 288, + 422 + ], + "type": "text", + "content": ". Formally," + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 51, + 426, + 287, + 441 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 51, + 426, + 287, + 441 + ], + "spans": [ + { + "bbox": [ + 51, + 426, + 287, + 441 + ], + "type": "interline_equation", + "content": "\\bar {x} _ {t - 1} \\leftarrow \\bar {x} _ {t - 1} - \\delta_ {t} \\nabla_ {x _ {t}} | | \\text {U p s a m p l e} \\left(\\hat {\\mathbf {x}} _ {0} ^ {t}, \\bar {H} \\times \\bar {W}\\right) - \\hat {x} _ {0} ^ {t} | |, \\tag {7}", + "image_path": "47e7efed30f65611666ce2712c42cb6d40383a665ee39d8ed1f7ab0d53ae648a.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 445, + 287, + 554 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 445, + 287, + 554 + ], + "spans": [ + { + "bbox": [ + 46, + 445, + 287, + 554 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 445, + 287, + 554 + ], + "type": "inline_equation", + "content": "\\delta_t" + }, + { + "bbox": [ + 46, + 445, + 287, + 554 + ], + "type": "text", + "content": " represent the weight of the guidance. Since the overall image structure is determined in the early diffusion steps, we start with " + }, + { + "bbox": [ + 46, + 445, + 287, + 554 + ], + "type": "inline_equation", + "content": "\\delta_t = 400" + }, + { + "bbox": [ + 46, + 445, + 287, + 554 + ], + "type": "text", + "content": " and follow a cosine scheduler to decrease this weight for later diffusion steps. This scheduling strategy mitigates potential quality degradation from matching the generated image with a lower-resolution version while allowing the model to fill-in higher-resolution details in the later decoding stages. Fig. 5 illustrates how RRG eliminates emerging artifacts." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 561, + 128, + 575 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 561, + 128, + 575 + ], + "spans": [ + { + "bbox": [ + 47, + 561, + 128, + 575 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 582, + 287, + 630 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 582, + 287, + 630 + ], + "spans": [ + { + "bbox": [ + 46, + 582, + 287, + 630 + ], + "type": "text", + "content": "ElasticDiffusion supports generating images across different resolutions and aspect ratios. Our experiments focus on: (1) square images at multiple resolutions, and (2) images with varied aspect ratios and resolutions." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 630, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 630, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 630, + 288, + 715 + ], + "type": "text", + "content": "Datasets. We evaluate the generation of square images on the Multi-Modal CelebAHQ dataset [56], which includes high-resolution square face images accompanied with text descriptions. We evaluate different aspect ratio generation using the LAION-COCO dataset [46], derived from the web-crawled LAION-5B dataset [45], which includes a variety of image aspect ratios and contains landscapes, people," + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 385, + 545, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 385, + 545, + 422 + ], + "spans": [ + { + "bbox": [ + 304, + 385, + 545, + 422 + ], + "type": "text", + "content": "objects, and everyday scenes, each paired with a synthetic caption generated using BLIP [27]. We consider four common aspect ratios: 9:16, 16:9, 3:4 and 4:3." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 423, + 546, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 423, + 546, + 544 + ], + "spans": [ + { + "bbox": [ + 304, + 423, + 546, + 544 + ], + "type": "text", + "content": "Evaluation Metrics. Following prior text-to-image synthesis works [3, 33, 35, 38, 43], we use automatic evaluation metrics Frechet Inception Distance (FID) and CLIP-score. FID [17] measures both the realism and diversity of the generated images by calculating the difference between features of the real and generated images computed using Inception-V3 [49] pretrained on ImageNet [41]. CLIP-score uses a pretrained text-image CLIP model [36] to measure alignment between the generated images and input prompts. We use 10,000 images to compute these scores." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 545, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 545, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 545, + 547, + 713 + ], + "type": "text", + "content": "Baselines. We compare our approach against prior diffusion model generation methods, specifically focusing on Stable Diffusion (SD) and MultiDiffusion (MD). SD follows the standard reverse diffusion process on an image latent space. MD uses a pretrained " + }, + { + "bbox": [ + 304, + 545, + 547, + 713 + ], + "type": "inline_equation", + "content": "SD" + }, + { + "bbox": [ + 304, + 545, + 547, + 713 + ], + "type": "text", + "content": " for panoramic image generation, by creating smaller, overlapping patches and averaging the Diffusion Model scores in intersecting areas. For baseline comparisons, we fix the pre-trained diffusion model to " + }, + { + "bbox": [ + 304, + 545, + 547, + 713 + ], + "type": "inline_equation", + "content": "SD_{1.4}" + }, + { + "bbox": [ + 304, + 545, + 547, + 713 + ], + "type": "text", + "content": ", which is trained to generate images at a resolution of " + }, + { + "bbox": [ + 304, + 545, + 547, + 713 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 304, + 545, + 547, + 713 + ], + "type": "text", + "content": ". Additionally, we compare our method with SDXL, an enhanced SD model that is three times larger than the standard one. SDXL is trained at a higher resolution of " + }, + { + "bbox": [ + 304, + 545, + 547, + 713 + ], + "type": "inline_equation", + "content": "1024p" + }, + { + "bbox": [ + 304, + 545, + 547, + 713 + ], + "type": "text", + "content": ", and fine-tuned on a specific set of aspect ratios with pixel sums close to " + }, + { + "bbox": [ + 304, + 545, + 547, + 713 + ], + "type": "inline_equation", + "content": "1024^2" + }, + { + "bbox": [ + 304, + 545, + 547, + 713 + ], + "type": "text", + "content": ". We exclude" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6608" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 70, + 286, + 396 + ], + "blocks": [ + { + "bbox": [ + 50, + 70, + 286, + 396 + ], + "lines": [ + { + "bbox": [ + 50, + 70, + 286, + 396 + ], + "spans": [ + { + "bbox": [ + 50, + 70, + 286, + 396 + ], + "type": "image", + "image_path": "6596084cb3897a01d054b5e051ecb42341e87ca58efe25e350836c561a05a7d3.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 406, + 287, + 452 + ], + "lines": [ + { + "bbox": [ + 46, + 406, + 287, + 452 + ], + "spans": [ + { + "bbox": [ + 46, + 406, + 287, + 452 + ], + "type": "text", + "content": "Figure 8. Qualitative comparison on various aspect ratios. ElasticDiffusion effectively handles a variety of aspect ratios. In comparison, " + }, + { + "bbox": [ + 46, + 406, + 287, + 452 + ], + "type": "inline_equation", + "content": "SD" + }, + { + "bbox": [ + 46, + 406, + 287, + 452 + ], + "type": "text", + "content": " and MultiDiffusion generate unrealistic images, while SDXL outputs exhibit a decline in the perceptual quality." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 458, + 288, + 519 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 458, + 288, + 519 + ], + "spans": [ + { + "bbox": [ + 46, + 458, + 288, + 519 + ], + "type": "text", + "content": "the latent refinement module in " + }, + { + "bbox": [ + 46, + 458, + 288, + 519 + ], + "type": "inline_equation", + "content": "SDXL" + }, + { + "bbox": [ + 46, + 458, + 288, + 519 + ], + "type": "text", + "content": " and focus our analysis on the base model. Throughout our experiments, we employ a DDIM sampling strategy with 50 steps and use 7.0 for the classifier-free guidance scaling factor. We also ensure consistency in seeds and captions for all baselines." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 526, + 160, + 540 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 526, + 160, + 540 + ], + "spans": [ + { + "bbox": [ + 47, + 526, + 160, + 540 + ], + "type": "text", + "content": "5.1. Qualitative Results" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 545, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 545, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 545, + 289, + 715 + ], + "type": "text", + "content": "We show square image generation samples in Fig. 7 generated from the CelebAHQ dataset at various resolutions. Both MultiDiffusion and Stable Diffusion have a tendency to replicate textures and body parts at higher resolutions, resulting in images that lack coherence. At resolutions lower than its training resolution " + }, + { + "bbox": [ + 46, + 545, + 289, + 715 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 46, + 545, + 289, + 715 + ], + "type": "text", + "content": ", StableDiffusion aligns poorly with the provided captions and produces unappealing images. SDXL, trained for " + }, + { + "bbox": [ + 46, + 545, + 289, + 715 + ], + "type": "inline_equation", + "content": "1024 \\times 1024" + }, + { + "bbox": [ + 46, + 545, + 289, + 715 + ], + "type": "text", + "content": " resolution, has a similar trend of reduced perceptual quality at lower sizes, eventually producing complete noise at a resolution of " + }, + { + "bbox": [ + 46, + 545, + 289, + 715 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 46, + 545, + 289, + 715 + ], + "type": "text", + "content": ". In contrast, ElasticDiffusion maintains image coherence across all tested resolutions, and yields results comparable to SDXL at " + }, + { + "bbox": [ + 46, + 545, + 289, + 715 + ], + "type": "inline_equation", + "content": "1024 \\times 1024" + }, + { + "bbox": [ + 46, + 545, + 289, + 715 + ], + "type": "text", + "content": ", despite using a less powerful base model and lower memory. We excluded" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 305, + 70, + 545, + 116 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 70, + 545, + 116 + ], + "spans": [ + { + "bbox": [ + 305, + 70, + 545, + 116 + ], + "type": "text", + "content": "Table 1. Quantitative comparison of on CelebA-HQ and LAION-COCO at different resolution. We indicate the best performances in bold, and underline the second best ones. #Calls represent the number of diffusion model calls required at each decoding steps." + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 308, + 123, + 545, + 395 + ], + "blocks": [ + { + "bbox": [ + 308, + 123, + 545, + 395 + ], + "lines": [ + { + "bbox": [ + 308, + 123, + 545, + 395 + ], + "spans": [ + { + "bbox": [ + 308, + 123, + 545, + 395 + ], + "type": "table", + "html": "
Res.MethodCelebA-HQLAION-COCO#CallsMem.
FID ↓CLIP ↑FID ↓CLIP ↑
256 × 256SD1.4258.4320.1454.0621.4327.2GB
SDXL368.0614.40175.8714.60218.5GB
Ours1.4235.2323.8823.7726.3028.6GB
512 × 512SD1.4233.4024.0020.5027.3328.6GB
SDXL240.2021.5742.5825.34221.6GB
768 × 768SD1.4238.8723.4529.8927.01211.1GB
MD1.4240.5622.8229.9827.31508.6GB
SDXL225.4824.2323.3127.88224.5GB
Ours1.4225.8626.6625.7825.93178.6GB
1024 × 1024SD1.4266.0121.9047.0125.70114.7GB
MD1.4264.5721.5537.7026.961628.6GB
SDXL230.2124.6225.5828.06127.5GB
Ours1.4228.8723.7427.7626.07338.6GB
", + "image_path": "aa400ad434a010a46ffbea54e5211893b749a2aa5d0288d81af08a59046f7b98.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 400, + 547, + 534 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 400, + 547, + 534 + ], + "spans": [ + { + "bbox": [ + 304, + 400, + 547, + 534 + ], + "type": "text", + "content": "results at " + }, + { + "bbox": [ + 304, + 400, + 547, + 534 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 304, + 400, + 547, + 534 + ], + "type": "text", + "content": " because both our method and Multi-Diffusion generate same outputs as the base Stable Diffusion model. We also omitted the results of MultiDiffusion at " + }, + { + "bbox": [ + 304, + 400, + 547, + 534 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 304, + 400, + 547, + 534 + ], + "type": "text", + "content": " since it is not designed to generate images at sizes smaller than the base model. Fig. 8 presents generated samples at various aspect ratios. We observe a similar trend of pattern repetition and reduced perceptual quality for images generated by the baselines, in contrast to those generated by ElasticDiffusion. Finally, we apply our method using SDXL as the base model to enable Full-HD image generation (" + }, + { + "bbox": [ + 304, + 400, + 547, + 534 + ], + "type": "inline_equation", + "content": "1920 \\times 1080" + }, + { + "bbox": [ + 304, + 400, + 547, + 534 + ], + "type": "text", + "content": ") and provide examples in Supp. Sec. 3.3." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 539, + 425, + 552 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 539, + 425, + 552 + ], + "spans": [ + { + "bbox": [ + 305, + 539, + 425, + 552 + ], + "type": "text", + "content": "5.2. Quantitative Results" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 557, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 557, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 557, + 547, + 715 + ], + "type": "text", + "content": "Table 1 shows quantitative evaluations. StableDiffusion shows increasing image quality degradation, as indicated by the FID metric when processing images of sizes different from its training resolution " + }, + { + "bbox": [ + 304, + 557, + 547, + 715 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 304, + 557, + 547, + 715 + ], + "type": "text", + "content": ". MultiDiffusion slightly improves FID at the expense of a substantially more forward calls. SDXL demonstrates similar declines in quality at resolutions far from its fine-tuning size of " + }, + { + "bbox": [ + 304, + 557, + 547, + 715 + ], + "type": "inline_equation", + "content": "1024 \\times 1024" + }, + { + "bbox": [ + 304, + 557, + 547, + 715 + ], + "type": "text", + "content": ". ElasticDiffusion, however, improves FID while maintaining a comparable CLIP-score to the base model. ElasticDiffusion also significantly improves the performance for lower resolutions " + }, + { + "bbox": [ + 304, + 557, + 547, + 715 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 304, + 557, + 547, + 715 + ], + "type": "text", + "content": ", while achieving comparable results to SDXL at its training resolution " + }, + { + "bbox": [ + 304, + 557, + 547, + 715 + ], + "type": "inline_equation", + "content": "1024 \\times 1024" + }, + { + "bbox": [ + 304, + 557, + 547, + 715 + ], + "type": "text", + "content": ", with only " + }, + { + "bbox": [ + 304, + 557, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\sim 31\\%" + }, + { + "bbox": [ + 304, + 557, + 547, + 715 + ], + "type": "text", + "content": " of the memory requirement for SDXL." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6609" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 50, + 123, + 285, + 457 + ], + "blocks": [ + { + "bbox": [ + 47, + 70, + 288, + 115 + ], + "lines": [ + { + "bbox": [ + 47, + 70, + 288, + 115 + ], + "spans": [ + { + "bbox": [ + 47, + 70, + 288, + 115 + ], + "type": "text", + "content": "Table 2. Quantitative comparison of on LAION-COCO datasets at various aspect ratios (A) and resolutions (R). best performances are in bold, and the second best are underlined. Vertical means the resolution is transposed from H:W to W:H." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 50, + 123, + 285, + 457 + ], + "lines": [ + { + "bbox": [ + 50, + 123, + 285, + 457 + ], + "spans": [ + { + "bbox": [ + 50, + 123, + 285, + 457 + ], + "type": "table", + "html": "
ARMethodHorizontalVertical
FID ↓CLIP↑FID ↓CLIP↑
3:4384 × 512SD1.438.8624.6317.6626.54
MD1.4----
SDXL104.8421.3368.8422.40
Ours1.435.1024.9115.5027.33
512 × 680SD1.445.5424.9616.8127.33
MD1.443.4424.5619.1326.80
SDXL62.8024.2028.2326.17
Ours1.441.0624.4018.9026.83
768 × 1024SD1.471.0024.0928.8326.30
MD1.454.8925.0226.3526.95
SDXL47.0525.7919.5027.41
Ours1.447.0324.9122.5225.80
9:16288 × 512SD1.423.5024.6924.0124.89
MD1.4----
SDXL121.8317.65112.4118.54
Ours1.423.2325.2622.8626.30
512 × 904SD1.429.8625.3427.4526.01
MD1.426.3525.7026.7025.28
SDXL29.6025.4027.2726.08
Ours1.422.8525.0126.6826.12
", + "image_path": "955f8e0bc6d5a70596f004f778f377e9c1f4af826b69af30ab1a822c159443f5.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 474, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 474, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 474, + 287, + 713 + ], + "type": "text", + "content": "Table 2 provides an evaluation on a variety of aspect ratios and resolutions for both horizontal and vertical image resolutions. StableDiffusion obtains worse FID score with increasing resolutions, while surprisingly maintaining or improving its CLIP-score. We posit that since CLIP-score quantifies the agreement between the input prompt and the generated image, StableDiffusion benefits from generating repetitive textures and artifacts that align closely with the prompt. For example, a photo of repeated lipsticks might align more with the caption \"lipstick\" despite its poor composition. MultiDiffusion generally enhances image quality but does not achieve satisfactory performance. MultiDiffusion struggles with objects that span the entire image (e.g. Fig. 8). Our method consistently improves FID over the baseline on horizontal and most vertical resolutions while preserving fidelity to the input prompts, thereby attaining comparable or superior CLIP-scores. Remarkably, even with a larger model size and explicit fine-tuning at a similar resolution of " + }, + { + "bbox": [ + 46, + 474, + 287, + 713 + ], + "type": "inline_equation", + "content": "768 \\times 1280" + }, + { + "bbox": [ + 46, + 474, + 287, + 713 + ], + "type": "text", + "content": ", SDXL only marginally surpasses our method at the " + }, + { + "bbox": [ + 46, + 474, + 287, + 713 + ], + "type": "inline_equation", + "content": "768 \\times 1024" + }, + { + "bbox": [ + 46, + 474, + 287, + 713 + ], + "type": "text", + "content": " resolution." + } + ] + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 324, + 91, + 523, + 155 + ], + "blocks": [ + { + "bbox": [ + 310, + 70, + 541, + 83 + ], + "lines": [ + { + "bbox": [ + 310, + 70, + 541, + 83 + ], + "spans": [ + { + "bbox": [ + 310, + 70, + 541, + 83 + ], + "type": "text", + "content": "Table 3. Ablation analysis of ElasticDiffusion on 500 images." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 324, + 91, + 523, + 155 + ], + "lines": [ + { + "bbox": [ + 324, + 91, + 523, + 155 + ], + "spans": [ + { + "bbox": [ + 324, + 91, + 523, + 155 + ], + "type": "table", + "html": "
Model DetailsFID ↓CLIP ↑
ElasticDiffusion133.6725.82
w/o Resampling150.0123.82
w/o RRG150.6424.34
w/o Imp. overlap, w/ exp. overlap141.4225.82
", + "image_path": "802cad3ac0b2ab9322ae3dc171970570a045452587c5674b9b97134ba8427517.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 306, + 164, + 398, + 177 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 164, + 398, + 177 + ], + "spans": [ + { + "bbox": [ + 306, + 164, + 398, + 177 + ], + "type": "text", + "content": "5.3. Ablation study" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 183, + 545, + 280 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 183, + 545, + 280 + ], + "spans": [ + { + "bbox": [ + 304, + 183, + 545, + 280 + ], + "type": "text", + "content": "Tab. 3 presents results of our method when excluding key components, demonstrating that each element improves performance. Fig. 4 illustrates the effectiveness of our proposed implicit overlap method in resolving boundary discontinuities at a reduced computational cost. Fig. 5 highlights the effectiveness of Reduced-Resolution Guidance in removing emerging artifacts. Fig. 6 shows the effectiveness of our iterative resampling technique in enhancing details." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 293, + 457, + 306 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 293, + 457, + 306 + ], + "spans": [ + { + "bbox": [ + 305, + 293, + 457, + 306 + ], + "type": "text", + "content": "6. Discussion and Conclusion" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 314, + 545, + 386 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 314, + 545, + 386 + ], + "spans": [ + { + "bbox": [ + 304, + 314, + 545, + 386 + ], + "type": "text", + "content": "Experimental results highlight the adaptability and effectiveness of ElasticDiffusion at steering diffusion models to produce an array of resolutions and aspect ratios. ElasticDiffusion requires no fine-tuning, consumes a constant memory footprint, enables both increased and reduced resolutions, and can generate a variety of aspect ratios." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 387, + 546, + 530 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 387, + 546, + 530 + ], + "spans": [ + { + "bbox": [ + 304, + 387, + 546, + 530 + ], + "type": "text", + "content": "ElasticDiffusion, however, does have several practical limitations. First, inaccuracies in estimating the global and local signals may result in artifacts. Although we attempt to mitigate artifacts with our Reduced-resolution guidance, it can still generate blurrier outputs. Second, since the global content guidance is initially estimated at the original training resolution of the underlying diffusion model, our method is less effective in generating images at significantly extended sizes beyond the training resolution. In particular, at extreme resolutions beyond 4X, our method produces artifacts and images of a lower perceptual quality. We provide examples of these failure cases in Supp. Sec. 10.4." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 531, + 545, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 531, + 545, + 687 + ], + "spans": [ + { + "bbox": [ + 304, + 531, + 545, + 687 + ], + "type": "text", + "content": "The main insight underpinning our method is a novel reinterpretation of classifier-free guidance in somewhat disentangling both global and local content. Our comprehensive evaluations demonstrate the feasibility of disentangling these signals, yet the full extent of their separation offers an avenue for further exploration in this direction. We hope that our findings inspire future work in investigating the separation of global and local content guidance signals for image synthesis. This separation holds potential for various applications such as selectively manipulating local and global content or enhancing style transfer. Additionally, the rich semantic representation in the class-direction score has the potential for improving discriminative models." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 689, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 545, + 712 + ], + "type": "text", + "content": "Acknowledgments. This work was partially funded by NSF Award #2201710." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6610" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 135 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 135 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 135 + ], + "type": "text", + "content": "[1] Omri Avrahami, Dani Lischinski, and Ohad Fried. Blended diffusion for text-driven editing of natural images. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). IEEE, 2022. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 137, + 287, + 180 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 137, + 287, + 180 + ], + "spans": [ + { + "bbox": [ + 53, + 137, + 287, + 180 + ], + "type": "text", + "content": "[2] Arpit Bansal, Hong-Min Chu, Avi Schwarzschild, Soumyadip Sengupta, Micah Goldblum, Jonas Geiping, and Tom Goldstein. Universal guidance for diffusion models, 2023. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 183, + 286, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 183, + 286, + 216 + ], + "spans": [ + { + "bbox": [ + 53, + 183, + 286, + 216 + ], + "type": "text", + "content": "[3] Omer Bar-Tal, Lior Yariv, Yaron Lipman, and Tali Dekel. Multidiffusion: Fusing diffusion paths for controlled image generation. In ICML, 2023. 2, 3, 4, 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 218, + 286, + 261 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 218, + 286, + 261 + ], + "spans": [ + { + "bbox": [ + 53, + 218, + 286, + 261 + ], + "type": "text", + "content": "[4] Angela Castillo, Jonas Kohler, Juan C. Pérez, Juan Pablo Pérez, Albert Pumarola, Bernard Ghanem, Pablo Arbeláez, and Ali Thabet. Adaptive guidance: Training-free acceleration of conditional diffusion models, 2023. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 264, + 286, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 264, + 286, + 296 + ], + "spans": [ + { + "bbox": [ + 53, + 264, + 286, + 296 + ], + "type": "text", + "content": "[5] Duygu Ceylan, Chun-Hao Paul Huang, and Niloy J. Mitra. Pix2video: Video editing using image diffusion. In ICCV, 2023. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 299, + 286, + 332 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 299, + 286, + 332 + ], + "spans": [ + { + "bbox": [ + 53, + 299, + 286, + 332 + ], + "type": "text", + "content": "[6] Hyungjin Chung, Jeongsol Kim, Michael T. McCann, Marc L. Klasky, and Jong Chul Ye. Diffusion posterior sampling for general noisy inverse problems, 2023. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 334, + 286, + 367 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 334, + 286, + 367 + ], + "spans": [ + { + "bbox": [ + 53, + 334, + 286, + 367 + ], + "type": "text", + "content": "[7] Jiwoo Chung, Sangeek Hyun, and Jae-Pil Heo. Style injection in diffusion: A training-free approach for adapting large-scale diffusion models for style transfer, 2024. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 369, + 286, + 391 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 369, + 286, + 391 + ], + "spans": [ + { + "bbox": [ + 53, + 369, + 286, + 391 + ], + "type": "text", + "content": "[8] Prafulla Dhariwal and Alex Nichol. Diffusion models beat gans on image synthesis. In NeurIPS, 2021. 2, 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 393, + 286, + 436 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 393, + 286, + 436 + ], + "spans": [ + { + "bbox": [ + 53, + 393, + 286, + 436 + ], + "type": "text", + "content": "[9] Ruoyi Du, Dongliang Chang, Timothy Hospedales, Yi-Zhe Song, and Zhanyu Ma. Demofusion: Democratising high-resolution image generation with no $$$. In CVPR, 2024. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 439, + 286, + 483 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 439, + 286, + 483 + ], + "spans": [ + { + "bbox": [ + 48, + 439, + 286, + 483 + ], + "type": "text", + "content": "[10] Alexandros Graikos, Srikar Yellapragada, Minh-Quan Le, Saarthak Kapse, Prateek Prasanna, Joel Saltz, and Dimitris Samaras. Learned representation-guided diffusion models for large-image generation, 2024. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 485, + 286, + 528 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 485, + 286, + 528 + ], + "spans": [ + { + "bbox": [ + 48, + 485, + 286, + 528 + ], + "type": "text", + "content": "[11] Jiatao Gu, Qingzhe Gao, Shuangfei Zhai, Baoquan Chen, Lingjie Liu, and Josh Susskind. Control3diff: Learning controllable 3d diffusion models from single-view images. In 3DV24, 2023. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 531, + 286, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 531, + 286, + 586 + ], + "spans": [ + { + "bbox": [ + 48, + 531, + 286, + 586 + ], + "type": "text", + "content": "[12] Lanqing Guo, Yingqing He, Haoxin Chen, Menghan Xia, Xiaodong Cun, Yufei Wang, Siyu Huang, Yong Zhang, Xintao Wang, Qifeng Chen, Ying Shan, and Bihan Wen. Make a cheap scaling: A self-cascade diffusion model for higher-resolution adaptation, 2024. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 588, + 286, + 631 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 588, + 286, + 631 + ], + "spans": [ + { + "bbox": [ + 48, + 588, + 286, + 631 + ], + "type": "text", + "content": "[13] Feihong He, Gang Li, Lingyu Si, Leilei Yan, Shimeng Hou, Hongwei Dong, and Fanzhang Li. Cartooniff: Training-free cartoon image generation with diffusion transformer models, 2023. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 634, + 286, + 677 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 634, + 286, + 677 + ], + "spans": [ + { + "bbox": [ + 48, + 634, + 286, + 677 + ], + "type": "text", + "content": "[14] Yingqing He, Shaoshu Yang, Haoxin Chen, Xiaodong Cun, Menghan Xia, Yong Zhang, Xintao Wang, Ran He, Qifeng Chen, and Ying Shan. Scalecrafter: Tuning-free higher-resolution visual generation with diffusion models, 2023. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 680, + 286, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 680, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 680, + 286, + 713 + ], + "type": "text", + "content": "[15] Yingqing He, Tianyu Yang, Yong Zhang, Ying Shan, and Qifeng Chen. Latent video diffusion models for high-fidelity long video generation, 2023. 2" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 35, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "text", + "content": "[16] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-prompt image editing with cross attention control. In ICLR, 2022. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 107, + 545, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 107, + 545, + 150 + ], + "spans": [ + { + "bbox": [ + 307, + 107, + 545, + 150 + ], + "type": "text", + "content": "[17] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. In NIPS, 2017. 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 152, + 545, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 152, + 545, + 173 + ], + "spans": [ + { + "bbox": [ + 307, + 152, + 545, + 173 + ], + "type": "text", + "content": "[18] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance, 2022. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 175, + 545, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 175, + 545, + 217 + ], + "spans": [ + { + "bbox": [ + 307, + 175, + 545, + 217 + ], + "type": "text", + "content": "[19] Jonathan Ho, Chitwan Sahara, William Chan, David J. Fleet, Mohammad Norouzi, and Tim Salimans. Cascaded diffusion models for high fidelity image generation. Journal of Machine Learning Research, 23(47):1-33, 2022. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 220, + 545, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 220, + 545, + 251 + ], + "spans": [ + { + "bbox": [ + 307, + 220, + 545, + 251 + ], + "type": "text", + "content": "[20] Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J. Fleet. Video diffusion models. In NeurIPS, 2022. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 253, + 545, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 253, + 545, + 296 + ], + "spans": [ + { + "bbox": [ + 307, + 253, + 545, + 296 + ], + "type": "text", + "content": "[21] Edward J. Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. In ICLR, 2022. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 298, + 545, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 298, + 545, + 342 + ], + "spans": [ + { + "bbox": [ + 307, + 298, + 545, + 342 + ], + "type": "text", + "content": "[22] Rongjie Huang, Jiawei Huang, Dongchao Yang, Yi Ren, Luping Liu, Mingze Li, Zhenhui Ye, Jinglin Liu, Xiang Yin, and Zhou Zhao. Make-an-audio: Text-to-audio generation with prompt-enhanced diffusion models. In ICML, 2023. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 343, + 545, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 343, + 545, + 385 + ], + "spans": [ + { + "bbox": [ + 307, + 343, + 545, + 385 + ], + "type": "text", + "content": "[23] Yujin Jeong, Wonjeong Ryoo, Seunghyun Lee, Dabin Seo, Wonmin Byeon, Sanggil Kim, and Jinkyu Kim. The power of sound (tpos): Audio reactive video generation with stable diffusion. In ICCV, 2023. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 388, + 545, + 420 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 388, + 545, + 420 + ], + "spans": [ + { + "bbox": [ + 307, + 388, + 545, + 420 + ], + "type": "text", + "content": "[24] Zhiyu Jin, Xuli Shen, Bin Li, and Xiangyang Xue. Training-free diffusion model adaptation for variable-sized text-to-image synthesis, 2023. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 422, + 545, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 422, + 545, + 453 + ], + "spans": [ + { + "bbox": [ + 307, + 422, + 545, + 453 + ], + "type": "text", + "content": "[25] Gwanghyun Kim, Taesung Kwon, and Jong Chul Ye. Diffusionclip: Text-guided diffusion models for robust image manipulation. In CVPR, 2022. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 455, + 545, + 476 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 455, + 545, + 476 + ], + "spans": [ + { + "bbox": [ + 307, + 455, + 545, + 476 + ], + "type": "text", + "content": "[26] Zhifeng Kong and Wei Ping. On fast sampling of diffusion probabilistic models. In ICML, 2021. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 478, + 545, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 478, + 545, + 510 + ], + "spans": [ + { + "bbox": [ + 307, + 478, + 545, + 510 + ], + "type": "text", + "content": "[27] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation, 2022. 6" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 512, + 545, + 543 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 512, + 545, + 543 + ], + "spans": [ + { + "bbox": [ + 307, + 512, + 545, + 543 + ], + "type": "text", + "content": "[28] Jin Liu, Huaibo Huang, Chao Jin, and Ran He. Portrait diffusion: Training-free face stylization with chain-of-painting, 2023. 2" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 545, + 545, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 545, + 545, + 588 + ], + "spans": [ + { + "bbox": [ + 307, + 545, + 545, + 588 + ], + "type": "text", + "content": "[29] Andreas Lugmayr, Martin Danelljan, Andres Romero, Fisher Yu, Radu Timofte, and Luc Van Gool. Repaint: Inpainting using denoising diffusion probabilistic models. In CVPR, 2022. 5" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 307, + 590, + 545, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 590, + 545, + 633 + ], + "spans": [ + { + "bbox": [ + 307, + 590, + 545, + 633 + ], + "type": "text", + "content": "[30] Sicheng Mo, Fangzhou Mu, Kuan Heng Lin, Yanli Liu, Bochen Guan, Yin Li, and Bolei Zhou. Freecontrol: Training-free spatial control of any text-to-image diffusion model with any condition, 2023. 2" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 307, + 635, + 545, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 635, + 545, + 667 + ], + "spans": [ + { + "bbox": [ + 307, + 635, + 545, + 667 + ], + "type": "text", + "content": "[31] Brian B. Moser, Stanislav Frolov, Federico Raue, Sebastian Palacio, and Andreas Dengel. Dynamic attention-guided diffusion for image super-resolution, 2024. 2" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 307, + 669, + 545, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 669, + 545, + 690 + ], + "spans": [ + { + "bbox": [ + 307, + 669, + 545, + 690 + ], + "type": "text", + "content": "[32] Alex Nichol and Prafulla Dhariwal. Improved denoising diffusion probabilistic models. In ICML, 2021. 2" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 307, + 692, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 692, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 692, + 545, + 713 + ], + "type": "text", + "content": "[33] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and" + } + ] + } + ], + "index": 34 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 314, + 757 + ], + "type": "text", + "content": "6611" + } + ] + } + ], + "index": 36 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 66, + 72, + 287, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 72, + 287, + 105 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 287, + 105 + ], + "type": "text", + "content": "Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. In ICML, 2022. 2, 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 106, + 287, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 106, + 287, + 138 + ], + "spans": [ + { + "bbox": [ + 48, + 106, + 287, + 138 + ], + "type": "text", + "content": "[34] Bo Peng, Xinyuan Chen, Yaohui Wang, Chaochao Lu, and Yu Qiao. Conditionvideo: Training-free condition-guided text-to-video generation, 2023. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 140, + 287, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 140, + 287, + 183 + ], + "spans": [ + { + "bbox": [ + 48, + 140, + 287, + 183 + ], + "type": "text", + "content": "[35] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis, 2023. 2, 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 184, + 287, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 184, + 287, + 248 + ], + "spans": [ + { + "bbox": [ + 48, + 184, + 287, + 248 + ], + "type": "text", + "content": "[36] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In ICML, 2021. 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 250, + 287, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 250, + 287, + 282 + ], + "spans": [ + { + "bbox": [ + 48, + 250, + 287, + 282 + ], + "type": "text", + "content": "[37] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents, 2022. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 284, + 287, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 284, + 287, + 316 + ], + "spans": [ + { + "bbox": [ + 48, + 284, + 287, + 316 + ], + "type": "text", + "content": "[38] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents, 2022. 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 317, + 287, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 317, + 287, + 348 + ], + "spans": [ + { + "bbox": [ + 48, + 317, + 287, + 348 + ], + "type": "text", + "content": "[39] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 1, 2, 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 350, + 287, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 350, + 287, + 392 + ], + "spans": [ + { + "bbox": [ + 48, + 350, + 287, + 392 + ], + "type": "text", + "content": "[40] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In CVPR, 2023. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 394, + 287, + 447 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 394, + 287, + 447 + ], + "spans": [ + { + "bbox": [ + 48, + 394, + 287, + 447 + ], + "type": "text", + "content": "[41] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, Alexander C. Berg, and Li Fei-Fei. Imagenet large scale visual recognition challenge, 2015. 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 449, + 287, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 449, + 287, + 480 + ], + "spans": [ + { + "bbox": [ + 48, + 449, + 287, + 480 + ], + "type": "text", + "content": "[42] Chitwan Sahara, Jonathan Ho, William Chan, Tim Salimans, David J. Fleet, and Mohammad Norouzi. Image superresolution via iterative refinement, 2021. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 482, + 287, + 558 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 482, + 287, + 558 + ], + "spans": [ + { + "bbox": [ + 48, + 482, + 287, + 558 + ], + "type": "text", + "content": "[43] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, Seyed Kamyar Seyed Ghasemipour, Burcu Karagol Ayan, S. Sara Mahdavi, Rapha Gontijo Lopes, Tim Salimans, Jonathan Ho, David J Fleet, and Mohammad Norouzi. Photorealistic text-to-image diffusion models with deep language understanding. In NeurIPS, 2022. 2, 3, 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 559, + 287, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 559, + 287, + 602 + ], + "spans": [ + { + "bbox": [ + 48, + 559, + 287, + 602 + ], + "type": "text", + "content": "[44] Axel Sauer, Frederic Boesel, Tim Dockhorn, Andreas Blattmann, Patrick Esser, and Robin Rombach. Fast high-resolution image synthesis with latent adversarial diffusion distillation, 2024. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 604, + 287, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 604, + 287, + 679 + ], + "spans": [ + { + "bbox": [ + 48, + 604, + 287, + 679 + ], + "type": "text", + "content": "[45] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, Patrick Schramowski, Srivatsa Kundurthy, Katherine Crowson, Ludwig Schmidt, Robert Kaczmarczyk, and Jenia Jitsev. Laion-5b: An open large-scale dataset for training next generation image-text models, 2022. 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 681, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 681, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 681, + 287, + 712 + ], + "type": "text", + "content": "[46] Christoph Schuhmann, Andreas A. Köpf, Richard Vencu, Theo Coombes, and Ross Beaumont. Laion coco: 600m synthetic captions from laion2b-en, 2023. 6" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 714 + ], + "type": "list", + "angle": 0, + "index": 32, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "text", + "content": "[47] Fengyuan Shi, Jiaxi Gu, Hang Xu, Songcen Xu, Wei Zhang, and Limin Wang. Bivdiff: A training-free framework for general-purpose video synthesis via bridging image and video diffusion models, 2023. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 118, + 545, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 545, + 140 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 545, + 140 + ], + "type": "text", + "content": "[48] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In ICLR, 2021. 2, 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 141, + 545, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 141, + 545, + 183 + ], + "spans": [ + { + "bbox": [ + 307, + 141, + 545, + 183 + ], + "type": "text", + "content": "[49] Christian Szegedy, Vincent Vanhoucke, Sergey Ioffe, Jonathon Shlens, and Zbigniew Wojna. Rethinking the inception architecture for computer vision. In CVPR, 2015. 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 186, + 545, + 218 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 186, + 545, + 218 + ], + "spans": [ + { + "bbox": [ + 307, + 186, + 545, + 218 + ], + "type": "text", + "content": "[50] Andrey Voynov, Amir Hertz, Moab Arar, Shlomi Fruchter, and Daniel Cohen-Or. Anylens: A generative diffusion model with any rendering lens, 2023. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 220, + 545, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 220, + 545, + 262 + ], + "spans": [ + { + "bbox": [ + 307, + 220, + 545, + 262 + ], + "type": "text", + "content": "[51] Bingyuan Wang, Hengyu Meng, Zeyu Cai, Lanjiong Li, Yue Ma, Qifeng Chen, and Zeyu Wang. Magicscrol: Nontypical aspect-ratio image generation for visual storytelling via multi-layered semantic-aware denoising, 2023. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 264, + 545, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 264, + 545, + 308 + ], + "spans": [ + { + "bbox": [ + 307, + 264, + 545, + 308 + ], + "type": "text", + "content": "[52] Fu-Yun Wang, Zhaoyang Huang, Xiaoyu Shi, Weikang Bian, Guanglu Song, Yu Liu, and Hongsheng Li. *Animatelcm: Accelerating the animation of personalized diffusion models and adapters with decoupled consistency learning*, 2024. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 309, + 545, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 309, + 545, + 342 + ], + "spans": [ + { + "bbox": [ + 307, + 309, + 545, + 342 + ], + "type": "text", + "content": "[53] Jianyi Wang, Zongsheng Yue, Shangchen Zhou, Kelvin C. K. Chan, and Chen Change Loy. Exploiting diffusion prior for real-world image super-resolution, 2023. 2, 3, 4" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 343, + 545, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 343, + 545, + 385 + ], + "spans": [ + { + "bbox": [ + 307, + 343, + 545, + 385 + ], + "type": "text", + "content": "[54] Lezhong Wang, Jeppe Revall Frisvad, Mark Bo Jensen, and Siavash Arjomand Bigdeli. Stereodiffusion: Training-free stereo image generation using latent diffusion models, 2024. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 388, + 545, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 388, + 545, + 430 + ], + "spans": [ + { + "bbox": [ + 307, + 388, + 545, + 430 + ], + "type": "text", + "content": "[55] Daniel Watson, William Chan, Ricardo Martin-Brualla, Jonathan Ho, Andrea Tagliasacchi, and Mohammad Norouzi. Novel view synthesis with diffusion models. In ICLR, 2023. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 433, + 545, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 433, + 545, + 464 + ], + "spans": [ + { + "bbox": [ + 307, + 433, + 545, + 464 + ], + "type": "text", + "content": "[56] Weihao Xia, Yujiu Yang, Jing-Hao Xue, and Baoyuan Wu. Tedigan: Text-guided diverse face image generation and manipulation, 2021. 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 467, + 545, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 467, + 545, + 499 + ], + "spans": [ + { + "bbox": [ + 307, + 467, + 545, + 499 + ], + "type": "text", + "content": "[57] Guy Yariv, Itai Gat, Lior Wolf, Yossi Adi, and Idan Schwartz. Audiotoken: Adaptation of text-conditioned diffusion models for audio-to-image generation, 2023. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 501, + 545, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 501, + 545, + 533 + ], + "spans": [ + { + "bbox": [ + 307, + 501, + 545, + 533 + ], + "type": "text", + "content": "[58] Ahmet Burak Yildirim, Vedat Baday, Erkut Erdem, Aykut Erdem, and Aysegul Dundar. Inst-inpaint: Instructing to remove objects with diffusion models, 2023. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 535, + 545, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 535, + 545, + 566 + ], + "spans": [ + { + "bbox": [ + 307, + 535, + 545, + 566 + ], + "type": "text", + "content": "[59] Jiwen Yu, Yinhuai Wang, Chen Zhao, Bernard Ghanem, and Jian Zhang. Freedom: Training-free energy-guided conditional diffusion model, 2023. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 568, + 545, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 568, + 545, + 611 + ], + "spans": [ + { + "bbox": [ + 307, + 568, + 545, + 611 + ], + "type": "text", + "content": "[60] Shen Zhang, Zhaowei Chen, Zhenyu Zhao, Zhenyuan Chen, Yao Tang, Yuhao Chen, Wengang Cao, and Jiajun Liang. Hidiffusion: Unlocking high-resolution creativity and efficiency in low-resolution trained diffusion models, 2023. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 613, + 545, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 613, + 545, + 644 + ], + "spans": [ + { + "bbox": [ + 307, + 613, + 545, + 644 + ], + "type": "text", + "content": "[61] Peiang Zhao, Han Li, Ruiyang Jin, and S. Kevin Zhou. Loco: Locally constrained training-free layout-to-image synthesis, 2024. 2" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 647, + 545, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 647, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 307, + 647, + 545, + 689 + ], + "type": "text", + "content": "[62] Qingping Zheng, Yuanfan Guo, Jiankang Deng, Jianhua Han, Ying Li, Songcen Xu, and Hang Xu. Any-sized diffusion: Toward efficient text-driven synthesis for any-size hd images, 2023. 2" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 307, + 691, + 545, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 691, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 307, + 691, + 545, + 714 + ], + "type": "text", + "content": "[63] Álvaro Barbero Jiménez. Mixture of diffusers for scene composition and high resolution image generation, 2023. 2, 3, 4" + } + ] + } + ], + "index": 31 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6612" + } + ] + } + ], + "index": 33 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Elite360D_ Towards Efficient 360 Depth Estimation via Semantic- and Distance-Aware Bi-Projection Fusion/b933bca2-9ff4-42fc-9bd7-be64c39e1936_content_list.json b/2024/Elite360D_ Towards Efficient 360 Depth Estimation via Semantic- and Distance-Aware Bi-Projection Fusion/b933bca2-9ff4-42fc-9bd7-be64c39e1936_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..584b1ec3fa62046b35541f7a98d773316d020c8c --- /dev/null +++ b/2024/Elite360D_ Towards Efficient 360 Depth Estimation via Semantic- and Distance-Aware Bi-Projection Fusion/b933bca2-9ff4-42fc-9bd7-be64c39e1936_content_list.json @@ -0,0 +1,1570 @@ +[ + { + "type": "text", + "text": "Elite360D: Towards Efficient 360 Depth Estimation via Semantic- and Distance-Aware Bi-Projection Fusion", + "text_level": 1, + "bbox": [ + 133, + 128, + 836, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Hao Ai $^{1}$ Lin Wang $^{1,2*}$", + "bbox": [ + 388, + 202, + 584, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ AI Thrust, HKUST(GZ) $^{2}$ Dept. of CSE, HKUST", + "bbox": [ + 276, + 220, + 692, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "hai033@connect.hkust-gz.edu.cn, linwang@ust.hk", + "bbox": [ + 279, + 241, + 689, + 255 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 291, + 313, + 306 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "360 depth estimation has recently received great attention for 3D reconstruction owing to its omnidirectional field of view (FoV). Recent approaches are predominantly focused on cross-projection fusion with geometry-based reprojection: they fuse 360 images with equirectangular projection (ERP) and another projection type, e.g., cubemap projection to estimate depth with the ERP format. However, these methods suffer from 1) limited local receptive fields, making it hardly possible to capture large FoV scenes, and 2) prohibitive computational cost, caused by the complex cross-projection fusion module design. In this paper, we propose Elite360D, a novel framework that inputs the ERP image and icosahedron projection (ICOSAP) point set, which is undistorted and spatially continuous. Elite360D is superior in its capacity in learning a representation from a local-with-global perspective. With a flexible ERP image encoder, it includes an ICOSAP point encoder, and a Bi-projection Bi-attention Fusion (B2F) module (totally $\\sim 1M$ parameters). Specifically, the ERP image encoder can take various perspective image-trained backbones (e.g., ResNet, Transformer) to extract local features. The point encoder extracts the global features from the ICOSAP. Then, the B2F module captures the semantic- and distance-aware dependencies between each pixel of the ERP feature and the entire ICOSAP feature set. Without specific backbone design and obvious computational cost increase, Elite360D outperforms the prior arts on several benchmark datasets.", + "bbox": [ + 75, + 323, + 473, + 731 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Multimedia Material", + "text_level": 1, + "bbox": [ + 76, + 739, + 258, + 755 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "For videos, code, demo and more information, you can visit https://VLIS2022.github.io/Elite360D/", + "bbox": [ + 75, + 758, + 468, + 790 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 794, + 209, + 809 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$360^{\\circ}$ images capture the complete surrounding environment in one shot with an ultra-wide field-of-view (FoV) of $180^{\\circ} \\times 360^{\\circ}$ , which is broadly applied to applications, e.g., autonomous driving [1, 19, 42, 43, 49] and virtual reality", + "bbox": [ + 76, + 816, + 470, + 878 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/4b8c1128a92aaa7822a410fa93cb01b5cfaaf43b3b1e06e4afc0f851443c59e5.jpg", + "image_caption": [ + "Model #FLOPs(G)" + ], + "image_footnote": [], + "bbox": [ + 501, + 290, + 696, + 430 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/6ccd0c6784e3ed90697fd7c4da8aba8d21897d4ef0eaceaa2b3eb9f282f127f6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 697, + 291, + 888, + 430 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/df0039b601e777013f7ebd4a22ecf72e05ba18dd39bdc61f392415cb7e9bdbdb.jpg", + "image_caption": [ + "ERP Image" + ], + "image_footnote": [], + "bbox": [ + 504, + 446, + 687, + 518 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/bb551f9bef0ee005a962d3966ac4eed6aac8138259cd520f9cf94b3ede5d6865.jpg", + "image_caption": [ + "(a)", + "GT Depth" + ], + "image_footnote": [], + "bbox": [ + 702, + 446, + 887, + 518 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/bfd43e37b788093d55dfe53c89684b1fd3d1f5de07820cf265f93fcbb2556269.jpg", + "image_caption": [ + "Equi-Baseline", + "Figure 1. (a) Performance (RMSE error) curves on M3D test dataset [5]. Larger circles mean more parameters (e.g., ResNet18, ResNet34, ResNet50) and lower errors mean better performance. (b) Comparison with the ResNet34 as the ERP encoder backbone. With only 1M more parameters, our depth result is more accurate." + ], + "image_footnote": [], + "bbox": [ + 504, + 531, + 687, + 603 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/698131984424d9e7b1c0bd995bd3c5a496ed09ed2cd3b34c8dee377974d61f94.jpg", + "image_caption": [ + "Our Elite360D", + "(b)" + ], + "image_footnote": [], + "bbox": [ + 702, + 531, + 887, + 603 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "[4, 23, 26]. The ability to infer the 3D structure of a surrounding scene has inspired active research for monocular 360 depth estimation. Equirectangular projection (ERP) is the commonly used projection type that can provide a complete view of a scene. An ERP image, a.k.a., panorama, samples pixels with a higher density at the poles compared to the equator, resulting in spherical distortions (Fig. 2(c)).", + "bbox": [ + 496, + 702, + 890, + 808 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In recent years, to address the spherical distortions, several ah-hoc designs [33, 36, 50, 51] focus on the distortion-aware convolutional kernels and sample grids from undistorted tangent planes on the sphere. These grids are backprojected to the corresponding locations of the ERP image. Also, some methods [28, 34, 44] partition the ERP", + "bbox": [ + 496, + 810, + 893, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author (e-mail: linwang@ust.hk)", + "bbox": [ + 94, + 886, + 357, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "9926", + "bbox": [ + 482, + 989, + 514, + 1000 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "input into multiple vertical slices and extract a compact representation from these slices to extend the receptive field of each pixel for the ERP representation. Unfortunately, these methods result in substantial computational costs. Meanwhile, the limited local receptive field of the convolutional filters is insufficient to provide the global perception of the panorama. Therefore, PanoFormer [32] and EGFormer [45] design specific transformer architectures to effectively model the long-range dependencies between the global context of large-FoV ERP images. Nevertheless, the receptive fields are confined by the sizes of local attention windows. Furthermore, the aforementioned data-driven methods often yield sub-optimal results when trained from scratch on limited $360^{\\circ}$ depth dataset (See Tab. 2).", + "bbox": [ + 75, + 90, + 472, + 303 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "With the power of existing large-scale perspective images [11], several methods [2, 16, 37, 38] employ the pretrained models - designed for perspective images - as encoder backbones to extract ERP feature maps, and propose the cross-projection fusion to rectify distortions in the ERP feature maps. Specifically, these methods introduce the less-distorted planar projection data, i.e., cubemap projection (CP) patches or tangent projection (TP) patches (See Fig. 2(d), (e)), and unify the spatial dimensions between different projections for the cross-projection fusion. BiFuse and UniFuse [16, 37, 38] propose the C2E module to reproject the content in patch-wise CP features into ERP grid following the spherical geometric relationships. Furthermore, they leverage the fusion between ERP feature maps and C2E feature maps to enhance depth estimation accuracy. However, their concatenation-based feature fusion across multiple scales brings a significant computational burden. Based on the feature similarity, HRDFuse [2] spatially aligns the ERP pixel features and TP patch features. This makes it possible to significantly marry the global context in the ERP feature map and regional structural details in the TP patch features. However, the decoder-level fusion remarkably increases the computational memory and cost.", + "bbox": [ + 75, + 306, + 472, + 654 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "These less-distorted gnomonic projections, i.e., CP and TP patches, are spatially discontinuous and require complex re-projection operations for the desired ERP format predictions. Inspired by these issues, we introduce a more powerful projection, the icosahedron projection (ICOSAP) [21, 46], see Fig. 2(f). Importantly, ICOSAP is a spatially continuous and globally perceptive non-Euclidean projection for $360^{\\circ}$ images. In light of this, we propose an efficient and effective fusion-based framework, named Elite360D that takes the best of ERP and ICOSAP by learning a representation from a local-with-global perspective. It comprises three components: an ERP encoder, an ICOSAP encoder, and the bi-projection bi-attention fusion (B2F) module. For the ERP image encoder, our Elite360D flexibly supports a wide range of off-the-shelf 2D models, e.g., ResNet [14], Swin transformer [24], as the backbone to", + "bbox": [ + 75, + 657, + 472, + 901 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/559fd2dd5357f51f53b367b8c3c250f69194382d4865fadbb1f82775ba3e7437.jpg", + "image_caption": [ + "(a) Panorama" + ], + "image_footnote": [], + "bbox": [ + 531, + 88, + 596, + 140 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/517fb55fd3973852e3fd942b48bd7c0440f89f20555a0d131e6bcd5e61239cdd.jpg", + "image_caption": [ + "(b) ERP" + ], + "image_footnote": [], + "bbox": [ + 647, + 89, + 702, + 138 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/1c8897cd52e20ba6a64eeecdbe62c477c8f42d61c95b12daacf070ab3f16ac4d.jpg", + "image_caption": [ + "(c) ERP (flatten)" + ], + "image_footnote": [], + "bbox": [ + 756, + 90, + 859, + 140 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/8683cec09fd50640b30f1a71c3147380a3f1f434af1c3e190bcc7d0e79d4b8ed.jpg", + "image_caption": [ + "(d) CP" + ], + "image_footnote": [], + "bbox": [ + 529, + 156, + 596, + 205 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/adb587f419b2ea73a279429ddeaab0947ee794a867d4317b62e7cd5fd837f3b6.jpg", + "image_caption": [ + "(e) TP" + ], + "image_footnote": [], + "bbox": [ + 643, + 156, + 710, + 205 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/ed904e61df0c4ff3dd2dfdcd9325d330125cf773612ec50c478beac5f4aa4ec8.jpg", + "image_caption": [ + "(f) ICOSAP (level=1)", + "Figure 2. Different projections of a spherical imaging panorama." + ], + "image_footnote": [], + "bbox": [ + 777, + 157, + 839, + 205 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "extract ERP feature maps. This potentially reduces the overfitting problems, particularly on small-scale 360 depth estimation datasets [3, 5]. For efficiency, we represent the ICOSAP spheres as the discrete point sets rather than icosahedron meshes [10, 15, 46] or unfolded representations [21, 43]. As such, we can avoid semantic information redundancy due to dense ERP pixels and maintain the spatial position information. Then, to enable each ERP pixel feature with local receptive fields to perceive the whole scene, B2F module captures the semantic- and distance-aware dependencies between each ERP pixel feature and entire ICOSAP feature set.", + "bbox": [ + 496, + 239, + 890, + 421 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We conduct extensive experiments on three different datasets with different encoder backbones to demonstrate the flexibility and effectiveness of our Elite360D. The experimental results indicate that our Elite360D significantly improves plain-backbones' performance with minimal computational memory (only about 1M parameters) (See Fig. 1(a)). Note that, based on the simple ERP depth estimation baseline, Elite360D only performs bi-projection feature fusion at the last feature layer and achieves results that are on par with leading methods, such as [2, 41].", + "bbox": [ + 496, + 422, + 892, + 573 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In summary, our main contributions are three-fold: (I) We introduce the ICOSAP with spatial continuity and global perception and represent them as the discrete point sets to reduce the computation cost and maintain the spatial information; (II) We propose the B2F module that jointly models the semantic and spatial dependencies between ICOSAP and ERP features and learn the representations with global-with-local receptive fields for large-FoV scenes; (III) Building upon the B2F module, we propose a novel framework that supports diverse off-the-shelf models as encoder backbones, showing better flexibility than the specially designed models, e.g., [32, 45].", + "bbox": [ + 496, + 573, + 893, + 755 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 500, + 768, + 650, + 784 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Monocular 360 depth estimation Existing methods can be categorized into two types: one with single projection input, and the other with bi-projection inputs.", + "bbox": [ + 496, + 794, + 890, + 839 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "1) Single Projection Input With the ERP images as the inputs, many methods focus on mitigating the inherent spherical distortion issues. Several methods [8, 36, 51] rectify distortions by modifying the receptive fields of tradi", + "bbox": [ + 496, + 839, + 893, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "9927", + "bbox": [ + 480, + 988, + 514, + 1000 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/4b8c68120290d19dcc009978e2f56cedab404557aec0de5506e1feb4ff66701a.jpg", + "image_caption": [ + "Figure 3. (a) An overview of our Elite360D framework, comprising image-based ERP feature extraction (Sec. 3.1), point-based ICOSAP feature extraction (Sec. 3.2), and Bi-projection Bi-attention fusion (B2F) (Sec. 3.3). For better visualization, we do not show the skip connections [30] at the decoding stage. (b) Illustration of the B2F module, consisting of three parts: semantic-aware affinity attention block (Fig. 5), distance-aware affinity attention block (Fig. 6) and gated fusion (Eq. 5)." + ], + "image_footnote": [], + "bbox": [ + 106, + 88, + 861, + 305 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "tional convolutional kernels. Besides, OmniFusion [41] and 360MonoDepth [29] follow [13] to predict depth maps from less distorted TP patches. Similarly, PanelNet [44] predicts depth maps on the vertical slices of ERP images and then aggregates them. Moreover, some methods [28, 34, 50] focus on tackling the small receptive fields of convolution filters for processing large FoV scenes. More recently, PanoFormer [32] and EGFormer [45] design transformer architectures [12] for ERP images to model the long-range dependencies. By contrast, $\\mathbf{S}^2\\mathrm{Net}$ [22] extracts robust features from ERP images with the Swin transformer [24] and projects the ERP feature maps onto the spherical surface to minimize distortions and maintain spatial consistency.", + "bbox": [ + 75, + 369, + 472, + 566 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2) Bi-Projection Inputs. BiFuse [37] utilizes a dense fusion strategy to bidirectionally fuse ERP and CP features at both encoding and decoding stages. In contrast, BiFuse++[38] and UniFuse [16] fuse ERP and CP features at the encoding stage. Recently, HRDFuse [2] achieved the SOTA performance based on the adaptive fusion of ERP and TP predictions. By contrast, we introduce a non-Euclidean projection, ICOSAP, which is less distorted and spatially continuous. We leverage the global perception capacity of ICOSAP to enable each ERP pixel-wise feature, with a limited local receptive field, to capture the entire scene.", + "bbox": [ + 75, + 566, + 470, + 733 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Representations for 360 Images ERP is the most commonly used projection, which projects a panorama onto a cylinder (Fig. 2(b)) and then unfolds it into a plane (Fig. 2(c)). However, ERP maps latitudes and longitudes onto the vertical and horizontal axes with equal spacing on the plane, leading to distortions. CP [27] projects the panoramas onto six faces of a cube where each face is less distorted (Fig. 2(d)). Building upon CP, some padding methods [7, 37] remove the boundary inconsistency between the adjacent CP patches. Recently, Eder et al. [13] proposed the tangent projection (TP) (Fig. 2(e)), to signif", + "bbox": [ + 75, + 734, + 470, + 902 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/1da8f6d0fed53242d5b8ed7279e74011204c8d01b3a5c7e818bfa7bc7c833b0c.jpg", + "image_caption": [ + "(a) $l = 0$" + ], + "image_footnote": [], + "bbox": [ + 501, + 367, + 584, + 435 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/f3406636289f7ea1271712b536fd94cec36ba985895a4f4c2edc4b318f99eab3.jpg", + "image_caption": [ + "(b) $l = 1$" + ], + "image_footnote": [], + "bbox": [ + 598, + 368, + 684, + 435 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/ec00a3b3e4a7e4e596147cdb250833a5ee0e0ab2b8e160f70d55ca63400e035b.jpg", + "image_caption": [ + "(c) $l = 2$" + ], + "image_footnote": [], + "bbox": [ + 697, + 368, + 787, + 435 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/e0b40d58d35bd335423683da34282925bc08dc2157f1abffe2b3a017281bf8db.jpg", + "image_caption": [ + "(d) $l = 3$" + ], + "image_footnote": [], + "bbox": [ + 802, + 368, + 890, + 435 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Figure 4. The subdivision of icosahedron at different resolution $l$ . icantly mitigate spherical distortion. Besides, [21, 31, 43] employ the polyhedrons, e.g., octahedron and icosahedron, to represent the panoramas and unfold these polyhedron representations on the plane for processing. In Elite360D, we select popular ICOSAP (See Fig. 4) as our input. Notably, we represent ICOSAP as discrete points, significantly reducing the computational cost while preserving spatial information and global awareness. Meanwhile, with the dense pixels in the ERP panoramas, discrete ICOSAP points can avoid semantic information redundancy.", + "bbox": [ + 498, + 460, + 890, + 628 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Cross-Attention Mechanism It is widely used for efficient multi-modal feature fusion. For instance, Chen et al. [6] built cross-attention blocks to align and fuse 2D image features and 3D point cloud features for 3D object recognition. Similarly, BEVGuide [25], using BEV embedding as a guided query, employs a cross-attention block to fuse information across different sensors. We propose the B2F module to model the relationships between each pixel of the ERP feature and the whole ICOSAP point feature set, based on the semantic-aware affinity attention and distance-aware affinity attention simultaneously. With the B2F module, each ERP pixel feature can perceive the spatial information and semantic information of the large FoV scenes.", + "bbox": [ + 496, + 628, + 892, + 825 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. The Proposed Framework", + "text_level": 1, + "bbox": [ + 500, + 834, + 743, + 849 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Overview. Our goal is to investigate an efficient and effective bi-projection fusion framework for 360 depth estimation. To this end, as illustrated in Fig. 3, the proposed", + "bbox": [ + 498, + 854, + 893, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "9928", + "bbox": [ + 480, + 989, + 514, + 1000 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Elite360D consists of three key components: image-based ERP feature extraction, point-based ICOSAP feature extraction and bi-projection bi-attention fusion (B2F). For the ERP feature $F^{E}$ extraction, Elite360D flexibly accommodates a wide range of 2D models as encoder backbones and takes advantage of pre-training to reduce the over-fitting problems. For the ICOSAP feature extraction, we represent ICOSAP sphere as the point set, which significantly reduces the computation cost and preserves the spatial information and global perception. Then we directly employ the simple point encoder, i.e. Point transformer [47], to extract the ICOSAP point feature set $\\{F_{n}^{I}\\}_{n = 1,\\dots ,N}$ , $N$ is the point number, denoted as $\\{F_n^I\\}_N$ . Next, to facilitate bi-projection feature fusion and learn a powerful representation $F^{GL}$ from a local-with-global perspective, B2F module comprises semantic-aware affinity attention block, distance-aware affinity attention block, and gated fusion block. These two attention blocks model the semantic- and distance-aware dependencies between each ERP pixel-wise feature and entire ICOSAP feature set, respectively and the gated fusion block adaptively fuses the bi-attention outputs. Lastly, we employ a simple decoder with up-sampling and skip-connections to predict the final depth map from the fused representation $F^{GL}$ . We now describe the details.", + "bbox": [ + 76, + 90, + 472, + 454 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1. ERP Feature Extraction", + "text_level": 1, + "bbox": [ + 76, + 460, + 302, + 474 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Taking an ERP image with the resolution size of $H \\times W$ as the input, the encoder extracts feature map $F^{E} \\in \\mathbb{R}^{h \\times w \\times C}$ , where $h = H / s$ , $w = W / s$ , $s$ is the down-sampling scale factor and $C$ is the channel number. In particular, as treating ERP images as 2D perspective images, our encoder backbone is compatible with a wide range of robust 2D models that have been pre-trained on large-scale perspective image datasets [11], including CNNs and vision transformers, e.g., ResNet [14], EfficientNet [35], Swin transformer [24]. Notably, as repeated local operations cannot substitute for a global operator [40, 45], the finite size of convolutional kernels or the limited size of local attention windows results in ERP pixel-wise features lacking sufficient global receptive fields. Additionally, these 2D model performance suffers from the spherical distortions (See Tab. 1).", + "bbox": [ + 75, + 479, + 472, + 705 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. ICOSAP Feature Extraction", + "text_level": 1, + "bbox": [ + 76, + 714, + 333, + 729 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Preliminary. Since Lee et al. [21] found that polyhedron representations with a greater number of initial faces exhibit lower distortion, this paper focuses on the ICOSAP with the most initial faces. Especially, we introduce the ICOSAP to offer a comprehensive and high-quality global perception. For the resolution, ICOSAP data consists of $20 \\times 4^l$ faces and $12 \\times 4^l$ vertices at the subdivision level $l$ (See Fig. 4). Given the spatial relationships between ICOSAP and ERP within the spherical space, RGB values of ICOSAP vertices can be derived from corresponding ERP pixels.", + "bbox": [ + 75, + 733, + 468, + 885 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Feature extraction. Existing works focus on processing", + "bbox": [ + 76, + 886, + 470, + 901 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/7aa4770334ef310092503896260a80e21b73aac7e5baf5998c60ad04f2aeee9d.jpg", + "image_caption": [ + "Figure 5. The architecture of semantic-aware affinity attention. Especially, $\\mathbf{Q}_{i,j}^{S}\\in \\mathbb{R}^{1\\times d}$ , $\\mathbf{K}^S\\in \\mathbb{R}^{N\\times d}$ , and $\\mathbf{V}^S\\in \\mathbb{R}^{N\\times d}$ , where $d$ is the dimension and $N$ is the ICOSAP point number." + ], + "image_footnote": [], + "bbox": [ + 519, + 89, + 870, + 227 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "ICOSAP data on the plane, i.e., unfolded mesh representation [46] and spherical polyhedron representations [21]. However, these methods require specially designed and computationally intensive operations, such as convolutions, pooling [21], and up-sampling [31], to process the planar representations. By contrast, we propose to represent an ICOSAP input as a point set, as illustrated at the bottom of Fig. 3(a). ERP images provide dense pixel values and discrete ICOSAP point set can prevent redundancy in semantic information while preserving spatial information and global perception. Therefore, we take full advantage of two projections for efficient and accurate 360 depth estimation. In detail, we employ the central points of each face to represent the ICOSAP sphere. Firstly, we obtain the $20 \\times 4^l$ faces of an ICOSAP sphere at the default subdivision level $l$ , where each face is composed of three vertices. Then, we calculate the spatial coordinates and RGB values of each face center by averaging them of each three vertices, as each face is an equilateral triangle. As a result, we obtain the ICOSCAP point set $\\{P^{ICO}\\} \\in \\mathbb{R}^{(20 \\times 4^l) \\times 6}$ , where $20 \\times 4^l$ is the point number and 6 represents the coordinates $[x, y, z]$ and RGB channels. With the input point set $\\{P^{ICO}\\}$ , we directly employ the encoder of Point Transformer [47] to extract ICOSAP point feature set $\\{F_n^I\\}_N \\in \\mathbb{R}^{N \\times C}$ , where $N$ represents the number of point features and $C$ corresponds to the same channel number as the ERP feature map $F^E$ . Especially, for simplicity and efficiency, we do not opt for complex networks for this purpose.", + "bbox": [ + 496, + 277, + 893, + 704 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Bi-Projection Bi-Attention Fusion", + "text_level": 1, + "bbox": [ + 500, + 710, + 795, + 727 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In earlier methods [37] and [16], bi-projection feature fusion primarily depends on the geometric relationships between CP and ERP. They first apply the C2E function to re-project CP feature patches into ERP format feature maps and subsequently perform pixel-wise feature concatenation. While this fusion is effective, it brings several problems: 1) geometry-based re-projection and concatenation-based feature fusion significantly increase the computational costs (See Tab. 2); 2) the geometry-based fusion process from CP patches to ERP panorama restricts ERP pixels from the global perception, where each ERP pixel perceives scene in-", + "bbox": [ + 496, + 734, + 893, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "9929", + "bbox": [ + 480, + 989, + 514, + 1000 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/87abc69907542577fed6c570359ef5063eddeebd335922cfd3351ef9949c2dd0.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
DatasetsBackboneMethod#Params (M)#FLOPs (G)Abs Rel ↓Sq Rel ↓RMSE ↓δ1(%) ↑δ2(%) ↑δ3(%) ↑
M3D [5]ResNet-18 [14]Equi-Baseline14.3939.840.14280.12030.560780.9794.6898.26
Ours15.4345.910.12720.10700.527085.2895.2898.49
Δ+1.04+6.07-11.06%-9.73%-6.01%+4.31+0.60+0.23
ResNet-34 [14]Equi-Baseline24.5059.270.12550.10480.517385.5296.1098.55
Ours25.5465.290.11150.09140.487588.1596.4698.74
Δ+1.04+6.02-11.16%-12.79%-5.76%+2.63+0.36+0.19
ResNet-50* [14]Equi-Baseline29.6375.130.13700.12090.543282.5194.8398.07
Ours42.99170.110.11120.09800.487086.7096.0198.61
Δ+13.36+94.98-18.75%-18.94%-10.35%+4.19+1.18+0.54
EfficientNet-B5 [35]Equi-Baseline33.1018.810.10340.08310.463889.9296.8098.80
Ours34.1124.880.10480.08050.452489.9297.0799.14
Δ+1.01+6.07+1.35%-3.13%-2.46%-0.00+0.27+0.34
SWin-B* [24]Equi-Baseline90.75187.370.13010.12500.555784.0694.7697.72
Ours94.30211.280.12490.12100.546285.3495.0797.77
Δ+3.55+23.91-4.00%-3.20%-1.71%+1.28+0.31+0.05
DilateFormer-T [17]Equi-Baseline19.9346.960.14290.12740.574881.1594.4697.79
Ours20.9453.020.14230.12510.551782.6895.0098.14
Δ+1.01+6.06-0.42%-1.81%-4.02%+1.53+0.54+0.35
S2D3D [3]ResNet-34 [14]Equi-Baseline24.5059.220.12030.07540.372487.4196.5198.77
Ours25.5165.280.11820.07280.375688.7296.8498.92
Δ+1.01+6.06-9.21%-3.45%+0.86%+1.31+0.33+0.15
EfficientNet-B5 [35]Equi-Baseline33.1018.810.10260.06380.358089.4397.0699.16
Ours34.1124.880.10180.06030.357589.4797.2399.22
Δ+1.01+6.07-0.78%-5.49%-0.14%+0.03+0.17+0.06
Struct3D [48]ResNet-34 [14]Equi-Baseline24.5059.220.22560.39100.764176.9089.9694.49
Ours25.5165.280.14800.22150.496187.4194.3496.66
Δ+1.01+6.06-34.40%-43.35%-35.07%+10.51+4.38+2.17
EfficientNet-B5 [35]Equi-Baseline33.1018.810.13120.19380.431288.5395.2197.40
Ours34.1124.880.12770.19300.415189.1695.3397.43
Δ+1.01+6.07-2.67%-0.41%-3.73%+0.63+0.12+0.03
", + "bbox": [ + 138, + 88, + 831, + 441 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Table 1. Quantitative comparison with ERP-based depth baseline. For most conditions, we set the channel number $C$ to 64. Especially, we set $C$ to 256 for ResNet-50* and to 128 for Swin-B*. Bold indicates performance improvement. Red indicates performance decline.", + "bbox": [ + 76, + 446, + 890, + 477 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "formation only from its corresponding small-FoV CP patch without considering other patches; 3) this one-to-one alignment pattern only emphasizes spatial consistency, with no consideration for the semantic similarity. Thus, we design the Bi-Projection Bi-Attention Fusion (B2F) module to solve the above problems. As shown in Fig. 3(b), B2F module first leverages semantic-aware affinity attention and distance-aware affinity attention to model the semantic and spatial dependencies between each ERP pixel-wise feature $F_{i,j}^{E}$ and ICOSAP point feature set $\\{F_n^I\\}_N$ . Here $(i,j)$ indicates the coordinate of a pixel in the ERP feature map, $i \\in (1,h), j \\in (1,w)$ , and $N$ is the point number of the ICOSAP feature set. Consequently, a gated fusion block is employed to adaptively marry semantic- and distance-related information. Semantic-aware affinity attention. The semantic-aware affinity attention, as shown in Fig. 5, follows the standard dot-product attention [12]. In detail, given the extracted ERP feature map $F^{E} \\in \\mathbb{R}^{h \\times w \\times C}$ and ICOSAP point feature set $\\{F_n^I\\}_N \\in \\mathbb{R}^{N \\times C}$ (for simplicity, we denote it as $[F^I]$ ), we generate the query $\\mathbf{Q}_{i,j}^{S}$ from each ERP pixel-wise feature $F_{i,j}^{E} \\in \\mathbb{R}^{1 \\times C}$ and produce the key $\\mathbf{K}^{S}$ and value $\\mathbf{V}^{S}$ from the whole ICOSAP point feature set $[F^{I}]$ . Then, we calculate the attention weight $\\mathbf{A}_{i,j}^{S}$ based on $\\mathbf{Q}_{i,j}^{S}$ and $\\mathbf{K}^{S}$ , and obtain the affinity feature $F_{i,j}^{SA}$ with $F_{i,j}^{SA} = \\mathbf{A}_{i,j}^{S} * \\mathbf{V}^{S}, \\mathbf{A}_{i,j}^{S} = \\text{softmax}\\left(\\frac{\\mathbf{Q}_{i,j}^{S} \\mathbf{K}^{ST}}{\\sqrt{d}}\\right)$ , where $d$ is the channel number and", + "bbox": [ + 76, + 488, + 472, + 898 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/14496925d91d219f5dbb38074c0ad79bdf92be77a9e4c00a00e5b7d077ea33a8.jpg", + "image_caption": [ + "Figure 6. The architecture of distance-aware affinity attention, especially, $Q_{i,j}^{D} \\in R^{1 \\times d}$ , $K^{D} \\in \\mathbb{R}^{N \\times d}$ , and $\\mathbf{V}^{D} \\in \\mathbb{R}^{N \\times d}$ ." + ], + "image_footnote": [], + "bbox": [ + 531, + 484, + 857, + 654 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "$d = C$ . After querying all the ERP pixel-wise features, we can obtain the ERP format feature map $F^{SA}$ with the dimension of $\\mathbb{R}^{h\\times w\\times d}$ . The attention weight $A_{i,j}^{S}$ captures affinities based on the semantic-aware feature similarities, and the output $F^{SA}$ effectively integrates the global and local receptive fields.", + "bbox": [ + 496, + 700, + 890, + 792 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Distance-aware affinity attention. Distance-aware affinity attention captures the differences between ERP and ICOSAP in both spatial and semantic information. This can better utilize the geometric prior knowledge of panoramicas and precisely measure the distances between two different projection features. As depicted in Fig. 6, the distance-aware affinity attention is built upon the subtraction-based", + "bbox": [ + 496, + 795, + 892, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "9930", + "bbox": [ + 482, + 989, + 514, + 1000 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/2722c8d233fe17e6a2b7161296b07e606b3c7b7b0358cbec27865f5f80df9bf3.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
DatasetsBackboneMethodPub'Year#Params (M)#FLOPs (G)Abs Rel ↓Sq Rel ↓RMSE ↓δ1(%) ↑δ2(%) ↑δ3(%) ↑
M3D [5]TransformerEGFormer [45]ICCV'2315.3966.210.14730.15170.602581.5893.9097.35
PanoFormer [32]ECCV'2220.3881.090.10510.09660.492989.0896.2398.31
ResNet-18 [14]BiFuse [37]CVPR'2035.80165.660.13600.12020.548883.2795.1298.10
UniFuse [16]RAL'2130.2662.600.11910.10300.515886.0495.8498.30
Omnifusion [41]CVPR'2232.3598.680.12090.10900.505586.5895.8198.36
HRDFuse† [2]CVPR'2326.0950.590.14140.12410.550781.4894.8998.20
Ours-15.4345.910.12720.10700.527085.2895.2898.49
ResNet-34 [14]BiFuse [37]CVPR'2056.01199.580.11260.09920.502788.0096.1398.47
BiFuse++ [38]TPAMI'2252.4987.480.11230.09150.485388.1296.5698.69
UniFuse [16]RAL'2150.4896.520.11440.09360.483587.8596.5998.73
Omnifusion [41]CVPR'2242.46142.290.11610.10070.493187.7296.1598.44
HRDFuse† [2]CVPR'2346.3180.870.11720.09710.502586.7496.1798.49
Ours-25.5465.290.11150.09140.487588.1596.4698.74
ResNet-50* [14]BiFuse [37]CVPR'20253.08775.240.11790.09810.497086.7496.2798.66
UniFuse [16]RAL'21131.30222.300.11850.09840.502486.6696.1898.50
Ours-42.99170.110.11120.09800.487086.7096.0198.61
S2D3D [3]TransformerEGFormer [45]ICCV'2315.3966.210.15280.14080.497481.8593.3897.36
PanoFormer [32]ECCV'2220.3881.090.11220.07860.394588.7495.8498.59
ResNet-34 [14]Omnifusion [41]CVPR'2242.46142.290.11540.07750.380986.7496.0398.71
UniFuse [16]RAL'2150.4896.520.11240.07090.355587.0697.0498.99
Ours-25.5165.280.11820.07280.375688.7296.8498.92
Struct3D [48]TransformerEGFormer [45]ICCV'2315.3966.210.22050.45090.684179.7990.7194.55
PanoFormer [32]ECCV'2220.3881.090.25490.49490.793774.7089.1593.97
ResNet-34 [14]BiFuse [37]CVPR'2056.01199.580.15730.24550.521385.9194.0096.72
UniFuse [16]RAL'2150.4896.520.15060.23190.501685.4293.9996.76
Ours-25.5165.280.14800.22150.496187.4194.3496.66
", + "bbox": [ + 138, + 88, + 831, + 371 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 2. Quantitative comparison with the SOTA methods. ${}^{ \\dagger }$ means that we modify the HRDFuse network structure for a fair comparison. Green represents the best performance under the given encoder backbone.", + "bbox": [ + 75, + 372, + 888, + 402 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "cross-attention mechanism [39]. Given the ERP pixel-wise feature $F_{i,j}^{E}$ and ICOSAP point feature set $[\\pmb{F}^{\\pmb{I}}]$ , we firstly calculate the spatial distance embedding $Dis^{SP}$ from the spatial coordinates of ERP pixel and ICOSAP point set, i.e. $[x_{i,j}^{e},y_{i,j}^{e},z_{i,j}^{e}]$ and $\\{[x_n^{ico},y_n^{ico},z_n^{ico}]\\}_{N}$ , as:", + "bbox": [ + 76, + 409, + 468, + 491 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nD i s _ {i, j} ^ {S P} = \\left[ e ^ {- \\Delta x _ {i, j}}, e ^ {- \\Delta y _ {i, j}}, e ^ {- \\Delta z _ {i, j}} \\right] \\mathbf {W} ^ {S P}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 127, + 494, + 468, + 516 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where linear projection $\\mathbf{W}^{SP}\\in \\mathbb{R}^{3\\times d}$ $Dis_{i,j}^{SP}\\in \\mathbb{R}^{1\\times N\\times d}$ and $[\\Delta x_{i,j},\\Delta y_{i,j},\\Delta z_{i,j}]\\in \\mathbb{R}^{1\\times N\\times 3}$ is the distances between $[x_{i,j}^e,y_{i,j}^e,z_{i,j}^e ]$ and $\\{[x_n^{ico},y_n^{ico},z_n^{ico}]\\}_{N}$ . In particular, the operation $e^{-(\\cdot)}$ allows $Dis^{SP}$ to pay more attention to the close parts between ERP pixels and ICOSAP point set. After that, we produce the query $\\mathbf{Q}_{i,j}^{D}$ and key $\\mathbf{K}^D$ from $F_{i,j}^{E}$ and $[\\pmb {F}^{\\pmb{I}}]$ , respectively and calculate the semantic distance embedding $Dis_{i,j}^{SE}$ :", + "bbox": [ + 76, + 525, + 468, + 654 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nD i s _ {i, j} ^ {S E} = e ^ {- \\left\\| \\mathbf {Q} _ {i, j} ^ {D} - \\mathbf {K} ^ {D} \\right\\|}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 187, + 656, + 468, + 679 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\mathbf{W}_Q^D$ , $\\mathbf{W}_K^D \\in \\mathbb{R}^{C \\times d}$ are linear projections, $\\mathbf{Q}_{i,j}^D \\in \\mathbb{R}^{1 \\times d}$ , $\\mathbf{K}^D \\in \\mathbb{R}^{N \\times d}$ , and $Dis_{i,j}^{SE} \\in \\mathbb{R}^{1 \\times N \\times d}$ . Lastly, the distance-aware attention weight $\\mathbf{A}_{i,j}^D$ is generated with spatial and semantic distance embeddings, and the distance-aware affinity feature vector $F_{i,j}^{DA}$ is obtained from the attention weight $\\mathbf{A}_{i,j}^D$ and the value $\\mathbf{V}^D$ :", + "bbox": [ + 76, + 680, + 468, + 779 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {A} _ {i, j} ^ {D} = \\operatorname {s o f t m a x} \\left(\\frac {\\sum \\left(D i s _ {i , j} ^ {S P} + D i s _ {i , j} ^ {S E}\\right)}{\\sqrt {d}}\\right), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 148, + 781, + 468, + 813 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {V} ^ {D} = F ^ {I} \\mathbf {W} _ {V} ^ {D}, F _ {i, j} ^ {D A} = \\mathbf {A} _ {i, j} ^ {D} * \\mathbf {V} ^ {D}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 163, + 820, + 468, + 837 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\sum$ means the sum for the channel dimension, i.e. $\\sum (Dis_{i,j}^{SP} + Dis_{i,j}^{SE})\\in \\mathbb{R}^{1\\times N}$ . After querying all ERP pixel-wise features, we obtain the distance-aware aggregated feature $F^{DA}\\in \\mathbb{R}^{h\\times w\\times d}$ , $d = C$ .", + "bbox": [ + 76, + 839, + 468, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Gated fusion. Since direct average or concatenation may compromise the original representation ability, inspired by [9], we propose the gated fusion block to adaptively fuse $F^{SA}$ and $F^{DA}$ and obtain the representations $F^{GL}$ from a local-with-global perspective, formulated as:", + "bbox": [ + 498, + 409, + 890, + 479 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nF ^ {G L} = g ^ {S A} * F ^ {S A} + g ^ {D A} * F ^ {D A}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 584, + 481, + 890, + 500 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\ng ^ {S A} = \\sigma_ {S A} \\left(W _ {g} ^ {S A} \\cdot \\left[ F ^ {S A}; F ^ {D A} \\right]\\right),\n$$\n", + "text_format": "latex", + "bbox": [ + 588, + 502, + 802, + 520 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\ng ^ {D A} = \\sigma_ {D A} \\left(W _ {g} ^ {D A} \\cdot \\left[ F ^ {S A}; F ^ {D A} \\right]\\right),\n$$\n", + "text_format": "latex", + "bbox": [ + 586, + 522, + 803, + 540 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $W_{g}^{SA}$ and $W_{g}^{DA}$ are linear projections, $\\sigma_{SA}(\\cdot)$ and $\\sigma_{DA}(\\cdot)$ are the sigmoid functions.", + "bbox": [ + 500, + 547, + 890, + 579 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.4. Optimization", + "text_level": 1, + "bbox": [ + 500, + 588, + 638, + 604 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "With the fused feature $F^{GL}$ and multi-scale ERP feature maps in the ERP encoder backbone, we feed them into a decoder [30] with several up-sampling blocks and skipconnections to output the final depth. For the depth supervision, we follow existing works [28, 32] and employ the combination of Berhu loss [20] and gradient loss [28]. (Details of training loss can be found in the suppl. material.).", + "bbox": [ + 496, + 609, + 890, + 717 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 500, + 724, + 633, + 739 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Datasets, Metrics, and Implementation Details", + "text_level": 1, + "bbox": [ + 496, + 744, + 890, + 761 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Datasets and Metrics. We evaluate Elite360D on three datasets: two real-world datasets, Matterport3D(M3D) [5], Stanford2D3D(S2D3D) [3], and a recently proposed large-scale synthetic dataset, Structure3D(Struct3D) [48]. For the evaluation metrics, we follow previous works [16, 32, 45] to employ some standard metrics, including absolute relative error (Abs Rel), squared relative error (Sq Rel), root mean squared error (RMSE), and three threshold percentage $\\delta < \\alpha^t$ ( $\\alpha = 1.25, t = 1, 2, 3$ ), denoted as $\\delta_t$ . Addi", + "bbox": [ + 496, + 763, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "9931", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/9853f330a79315c7421f1e6ce11598740b021fe41cd8c5ee4da8eefa63278e18.jpg", + "image_caption": [ + "Figure 7. Qualitative results (with ResNet-34 as the backbone) on Matterport3D (top), Stanford2D3D (middle) and Structure3D (bottom)." + ], + "image_footnote": [], + "bbox": [ + 107, + 87, + 857, + 434 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/48f90647e65920518b98c294cb20cecb049f087a37b6857ea43f042d92ba1be3.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
BackbonePre-trainedAbs Rel↓Sq Rel↓RMSE↓δ1↑
ResNet-34X0.15960.14520.585681.36
✓0.11150.09140.487588.15
EfficientNet-B5X0.12110.10870.513186.80
✓0.10480.08050.452489.92
DilateFormerX0.15150.14150.569480.95
✓0.14230.12510.551782.68
", + "bbox": [ + 112, + 458, + 431, + 551 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 3. The impact of pre-training on ERP encoder backbones.", + "bbox": [ + 81, + 553, + 460, + 566 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "tionally, we measure the number of parameters and FLOPS to evaluate the efficiency of our method.", + "bbox": [ + 75, + 575, + 467, + 604 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Training Details. We use diverse ERP encoder backbones, including CNNs (ResNet-18, 34, 50 [14], EfficientNet B5 [35]), and transformers (Swin-B [24], DilateFormer-T [17]). All backbones are pre-trained on ImageNet1K [11]. We set the default channel number $C$ to 64 and default subdivision level of ICOSAP as $l = 4$ . For the ICOSAP encoder, we employ the one of Point transformer [47] with three down-sample blocks. Following [16], we use Adam optimizer [18] and a constant learning rate of $1\\mathrm{e}^{-4}$ . Considering the unfair comparisons stemming from variations in hyper-parameters and validation procedures used across different methods, we re-train the existing methods from scratch and validate them, following the unified training and validation settings [16]. (Due to page limit, detailed training and validation settings can be found in suppl. mat.).", + "bbox": [ + 73, + 606, + 468, + 847 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. Performance Comparison", + "text_level": 1, + "bbox": [ + 75, + 849, + 310, + 866 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Comparisons with ERP-based depth baselines. As shown in Tab. 1, with an increase of only $\\sim 1\\mathrm{M}$ parameters", + "bbox": [ + 75, + 869, + 468, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "$(C = 64)$ , our Elite360D demonstrates substantial advancements over the ERP-based baselines across different ERP encoder backbones on all three datasets. Specifically, for the Matterport3D dataset, Elite360D achieves reductions exceeding $10\\%$ in Abs Rel error (ResNet-18, 34), along with reductions of $4.00\\%$ in Abs Rel error (Swin-B) and $4.02\\%$ in RMSE error (DilateFormer-T). Besides, with the larger channel number $C = 256$ (ResNet-50), Elite360D outperforms ERP baseline by $18.75\\%$ (Abl Rel), $18.94\\%$ (Sq Rel). For the small-scale S2D3D dataset, Elite360D outperforms ERP baseline by $9.21\\%$ in Abs Rel error and $1.31\\%$ in accuracy $\\delta_{1}$ (ResNet-34), as well as $5.49\\%$ in Sq Rel error (EfficientNet-B5). Remarkably, on the larger-scale Structure3D, Elite360 performs favorably against the baseline by a significant margin, especially with ResNet-34. Comparisons with prevalent methods. In Tab. 2, we conduct a comprehensive comparison with prevalent supervised methods. From the results, we can observe that our approach achieves similar or even superior performance compared to existing both bi-projection fusion methods and single input methods at a significantly lower cost, particularly on two large-scale datasets, Matterport3D and Structure3D. Specifically, for the Matterport3D dataset, our Elite360D with ResNet-34 outperforms UniFuse by $2.53\\%$ (Abs Rel) and with ResNet-50 outperforms BiFuse by $2.01\\%$ (RMSE). For the Structure3D dataset, our Elite360D with ResNet-34 outperforms UniFuse by $4.48\\%$ (Seq Rel), $1.99\\%$ ( $\\delta_{1}$ ). For performance on the Stanford2D3D dataset, we suspect it might be related to the ICOSAP point encoder.", + "bbox": [ + 496, + 460, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "9932", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/4d3823bbf76775462d399272a6df7a31810226282a3da49a9d7a64af39743eec.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Bi-projection feature fusionAbs Rel ↓Sq Rel ↓RMSE ↓δ1 ↑
SFA [2] + Add0.12760.10020.515084.27
SFA [2] + Concat0.11910.10190.514386.52
Only SA0.12040.10140.512186.26
Only DA0.11840.09720.494487.06
Our B2F (SA + DA)0.11150.09140.487588.15
", + "bbox": [ + 117, + 89, + 426, + 155 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/1f17d042b37e505a705e320f88f52adc1ae8601c54035a0b62faf115f6e1f004.jpg", + "table_caption": [ + "Table 4. The ablation results for B2F module." + ], + "table_footnote": [], + "table_body": "
Final fusionAbs Rel ↓Sq Rel ↓RMSE ↓δ1 ↑
Add0.16850.14810.580974.60
Average0.11980.09180.489386.65
Concatenation0.11450.09370.488087.66
Adaptive fusion [2]0.12440.09680.489186.08
Our gated fusion0.11150.09140.487588.15
", + "bbox": [ + 117, + 180, + 426, + 253 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The limited data of Stanford2D3D dataset restricts the ability of the transformer-based point encoder to provide accurate global perception. Moreover, in Fig. 7, we present the qualitative comparisons. Our Elite360D can predict more accurate depth values based on the local-with-global perception capabilities (e.g., flowers, shelves and doors). Additional qualitative results and inference time comparisons can be found in the suppl. material.", + "bbox": [ + 75, + 279, + 468, + 400 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.3. Ablation Study and Analyses", + "text_level": 1, + "bbox": [ + 76, + 409, + 334, + 425 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Most of ablation experiments are conducted on the Mateport3D test dataset with ResNet34 as the backbone.", + "bbox": [ + 75, + 429, + 468, + 459 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The Effect of pre-training. We verify the effectiveness of ImageNet [11] pre-training with different encoder backbones. As observed from Tab. 3, the pre-training results in a significant improvement for all encoder backbones, e.g., $6.79\\%$ improvement in accuracy $\\delta_{1}$ (ResNet-34). Notably, pre-training has a relatively small impact on DilateFormer. Combined with the results in Tab. 1, the explanation of this phenomenon is that the default input resolution in pretrained models is different from actual input, thereby impacting the resolution-related position embeddings. In general, pre-training based on large-scale perspective images can effectively enhance the performance of models based on $360^{\\circ}$ images and reduce the risk of overfitting.", + "bbox": [ + 75, + 460, + 468, + 656 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The effectiveness of B2F module. In Tab. 4, we compare four available bi-projection feature fusion modules. To align the spatial dimensions between ICOSAP point feature set and ERP feature map, we introduce SFA module from [2]. After that, we employ direct addition and concatenation to aggregate these two projections. We also achieve the bi-projection feature fusion with semantic-aware affinity attention (SA) alone and distance-aware affinity attention (DA) alone. Compared to the methods based solely on semantic-aware feature similarities (The first three rows), single distance-aware affinity attention can achieve better performance, which indicates that the spatial positional relationships boost the bi-projection feature fusion. Overall, our B2F module achieves the best performance.", + "bbox": [ + 75, + 657, + 468, + 868 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The superiority of ICOSAP. As only CP/TP's patch centers lie on the sphere's surface, we extract the feature em", + "bbox": [ + 76, + 869, + 468, + 900 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/ac2cdbf031fdf200100fe4e62f215ef275779ac7a42479d3100c118439eb33c3.jpg", + "table_caption": [ + "Table 5. The ablation results for the fusion of B2F module." + ], + "table_footnote": [], + "table_body": "
Method#Param(M)#FLOPs(G)Abs Rel ↓RMSE ↓δ1 ↑
ERP-CP25.6654.150.13690.540183.69
ERP-TP (N=18)25.6650.580.13280.538583.87
ERP-ICOSAP (Ours)15.4345.910.12720.527085.28
", + "bbox": [ + 503, + 89, + 890, + 148 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/564183ad451fc53238d9afe250a8e56c71311b2198b2ad657615a96c8f46fc70.jpg", + "table_caption": [ + "Table 6. The comparison of different projections on Matterport3D." + ], + "table_footnote": [], + "table_body": "
N of {F^I}#Params (M)#FLOPs (G)Abs Rel↓Sq Rel↓RMSE↓δ1↑
2027.4166.290.11570.09950.502487.12
8025.5465.290.11150.09140.487588.15
32024.9864.320.11530.09430.490587.85
", + "bbox": [ + 503, + 171, + 890, + 222 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 7. Impact of the ICOSAP point-wise feature number $N$ . Larger $N$ , fewer down-sampling blocks in the point encoder.", + "bbox": [ + 500, + 224, + 888, + 252 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "bedding from each CP/TP patch and employs the patch center coordinates and feature embedding as the input of B2F module. In Tab. 6, we show the results with ResNet18 backbone. Our Elite360D, utilizing the ICOSAP point set, marginally outperforms models with CP and TP patches, while exhibiting fewer parameters and FLOPs.", + "bbox": [ + 496, + 257, + 890, + 348 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The effectiveness of gated fusion. We conduct an ablation study for the gated fusion block, outlined in Tab. 5. With the feature maps $F^{SA}$ and $F^{DA}$ , We compare it with the direct addition, average fusion, concatenation, and the adaptive fusion in [2]. The gated fusion performs best.", + "bbox": [ + 496, + 348, + 890, + 422 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "ICOSAP point feature number $N$ . We study the effect of the ICOSAP feature point number (See Tab. 7). Too few points $(N = 20)$ lead to the over-concentrated global contextual information resulting from excessive down-sampling blocks, while too many points $(N = 320)$ lead to underconcentrated condition, resulting in insufficient perception of ERP pixel features. Best performance can be observed when $N = 80$ and we used $N = 80$ as default in this paper.", + "bbox": [ + 496, + 424, + 890, + 545 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion and Future Work", + "text_level": 1, + "bbox": [ + 500, + 551, + 767, + 566 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we proposed a novel bi-projection fusion solution for efficient 360 depth estimation. To address the limited local receptive field of ERP pixel-wise features and avoid expensive bi-projection fusion modules, we proposed a compact yet effective B2F module to learn the representations with local-with-global perceptions from ERP and ICOSAP. With an increase of 1M parameters, we significantly improved the performance of the ERP-based depth estimation baseline. Remarkably, our approach achieved performance on par with complex state-of-the-art methods. Future Work: From the experimental results, we observed that ERP-based depth baseline, with pre-trained Efficient-Net backbone, even outperforms existing specifically designed methods. Therefore, in the future, we will explore how to fully leverage different projections and successful perspective models for $360^{\\circ}$ community.", + "bbox": [ + 496, + 570, + 890, + 811 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgement", + "text_level": 1, + "bbox": [ + 500, + 820, + 660, + 837 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This paper is supported by the National Natural Science Foundation of China (NSF) under Grant No. NSFC22FYT45 and the Guangzhou City, University and Enterprise Joint Fund under Grant No.SL2022A03J01278.", + "bbox": [ + 496, + 839, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "9933", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Hao Ai, Zidong Cao, Jin Zhu, Haotian Bai, Yucheng Chen, and Ling Wang. Deep learning for omnidirectional vision: A survey and new perspectives. ArXiv, abs/2205.10468, 2022. 1", + "[2] Hao Ai, Zidong Cao, Yan-Pei Cao, Ying Shan, and Lin Wang. Hrdfuse: Monocular $360^{\\circ}$ depth estimation by collaboratively learning holistic-with-regional depth distributions. 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13273-13282, 2023. 2, 3, 6, 8", + "[3] Iro Armeni, Sasha Sax, Amir Roshan Zamir, and Silvio Savarese. Joint 2d-3d-semantic data for indoor scene understanding. CoRR, abs/1702.01105, 2017. 2, 5, 6", + "[4] Zidong Cao, Hao Ai, Yan Cao, Ying Shan, Xiaohu Qie, and Lin Wang. Omnizoomer: Learning to move and zoom in on sphere at high-resolution. 2023 IEEE/CVF International Conference on Computer Vision (ICCV), pages 12851-12861, 2023. 1", + "[5] Angel X. Chang, Angela Dai, Thomas A. Funkhouser, Maciej Halber, Matthias Nießner, Manolis Savva, Shuran Song, Andy Zeng, and Yinda Zhang. Matterport3d: Learning from RGB-D data in indoor environments. In 3DV, pages 667-676. IEEE Computer Society, 2017. 1, 2, 5, 6", + "[6] Zehui Chen, Zhenyu Li, Shiquan Zhang, Liangji Fang, Qinghong Jiang, Feng Zhao, Bolei Zhou, and Hang Zhao. Autoalign: Pixel-instance feature aggregation for multimodal 3d object detection. In International Joint Conference on Artificial Intelligence, 2022. 3", + "[7] Hsien-Tzu Cheng, Chun-Hung Chao, Jin-Dong Dong, Hao-Kai Wen, Tyng-Luh Liu, and Min Sun. Cube padding for weakly-supervised saliency prediction in $360^{\\circ}$ videos. 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1420-1429, 2018. 3", + "[8] Xinjing Cheng, Peng Wang, Yanqi Zhou, Chenye Guan, and Ruigang Yang. Omnidirectional depth extension networks. In 2020 IEEE International Conference on Robotics and Automation (ICRA), pages 589-595, 2020. 2", + "[9] Yanhua Cheng, Rui Cai, Zhiwei Li, Xin Zhao, and Kaiqi Huang. Locality-sensitive deconvolution networks with gated fusion for rgb-d indoor semantic segmentation. 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 1475-1483, 2017. 6", + "[10] Taco Cohen, Maurice Weiler, Berkay Kicanaoglu, and Max Welling. Gauge equivariant convolutional networks and the icosahedral cnn. In International Conference on Machine Learning, 2019. 2", + "[11] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. computer vision and pattern recognition, 2009. 2, 4, 7, 8", + "[12] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at" + ], + "bbox": [ + 78, + 114, + 470, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "scale. In International Conference on Learning Representations, 2021. 3, 5", + "[13] Marc Eder, Mykhailo Shvets, John Lim, and Jan-Michael Frahm. Tangent images for mitigating spherical distortion. 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12423-12431, 2019. 3", + "[14] Kaiming He, X. Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 770-778, 2015. 2, 4, 5, 6, 7", + "[15] ChiyuMax Jiang, Jingwei Huang, Karthik Kashinath, Prabhat Prabhat, Philip Marcus, and Matthias Niessner. Spherical cnns on unstructured grids. International Conference on Learning Representations, International Conference on Learning Representations, 2019. 2", + "[16] Hualie Jiang, Zhe Sheng, Siyu Zhu, Zilong Dong, and Rui Huang. Unifuse: Unidirectional fusion for $360^{\\circ}$ panorama depth estimation. IEEE Robotics and Automation Letters, 6: 1519-1526, 2021. 2, 3, 4, 6, 7", + "[17] Jiayu Jiao, Yu-Ming Tang, Kun-Yu Lin, Yipeng Gao, Jinhua Ma, Yaowei Wang, and Wei-Shi Zheng. Dilateformer: Multi-scale dilated transformer for visual recognition. IEEE Transactions on Multimedia, pages 1-14, 2023. 5, 7", + "[18] Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. CoRR, abs/1412.6980, 2014. 7", + "[19] Varun Ravi Kumar, Senthil Kumar Yogamani, Hazem Rashed, Ganesh Sitsu, Christian Witt, Isabelle Leang, Stefan Milz, and Patrick Mäder. Omnidet: Surround view cameras based multi-task visual perception network for autonomous driving. IEEE Robotics and Automation Letters, 6:2830-2837, 2021. 1", + "[20] Iro Laina, C. Rupprecht, Vasileios Belagiannis, Federico Tombari, and Nassir Navab. Deeper depth prediction with fully convolutional residual networks. 3DV 2016, pages 239-248, 2016. 6", + "[21] Yeonkun Lee, Jaeseok Jeong, Jong Seob Yun, Wonjune Cho, and Kuk jin Yoon. Spherephd: Applying cnns on a spherical polyhedron representation of $360^{\\circ}$ images. 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9173-9181, 2018. 2, 3, 4", + "[22] Meng Li, Senbo Wang, Weihao Yuan, Weichao Shen, Zhe Sheng, and Zilong Dong. $S^2$ net: Accurate panorama depth estimation on spherical surface. IEEE Robotics and Automation Letters, 8:1053-1060, 2023. 3", + "[23] Yunhao Li, Wei Shen, Zhongpai Gao, Yucheng Zhu, Guangtao Zhai, and Guodong Guo. Looking here or there? gaze following in 360-degree images. 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 3722-3731, 2021. 1", + "[24] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 9992-10002, 2021. 2, 3, 4, 5, 7", + "[25] Yunze Man, Liangyan Gui, and Yu-Xiong Wang. Bev-guided multi-modality fusion for driving perception. 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 21960-21969, 2023. 3" + ], + "bbox": [ + 503, + 92, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9934", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[26] Daniel Martin, Ana Serrano, Alexander W. Bergman, Gordon Wetzstein, and Belén Masiá. Scangan360: A generative model of realistic scanpaths for $360^{\\circ}$ images. IEEE Transactions on Visualization and Computer Graphics, 28:2003-2013, 2021. 1", + "[27] Rafael Monroy, Sebastian Lutz, Tejo Chalasani, and Aljoscha Smolic. Salnet360: Saliency maps for omnidirectional images with cnn. Signal Process. Image Commun., 69:26-34, 2017. 3", + "[28] Giovanni Pintore, Eva Almansa, and Jens Schneider. Slicenet: deep dense depth estimation from a single indoor panorama using a slice-based representation. 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11531-11540, 2021. 1, 3, 6", + "[29] Manuel Rey-Area, Mingze Yuan, and Christian Richardt. 360monodepth: High-resolution $360^{\\circ}$ monocular depth estimation. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3752-3762, 2022. 3", + "[30] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In MICCAI (3), pages 234-241. Springer, 2015. 3, 6", + "[31] Mehran Shakerinava and Siamak Ravanbakhsh. Equivariant networks for pixelized spheres. Proceedings of the 38th International Conference on Machine Learning, ICML, abs/2106.06662, 2021. 3, 4", + "[32] Zhijie Shen, Chunyu Lin, Kang Liao, Lang Nie, Zishuo Zheng, and Yao Zhao. Panofformer: Panorama transformer for indoor $360^{\\circ}$ depth estimation. In European Conference on Computer Vision, 2022. 2, 3, 6", + "[33] Yu-Chuan Su and Kristen Grauman. Kernel transformer networks for compact spherical convolution. 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9434-9443, 2018. 1", + "[34] Cheng Sun, Min Sun, and Hwann-Tzong Chen. Hohonet: 360 indoor holistic understanding with latent horizontal features. 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2573-2582, 2020. 1, 3", + "[35] Mingxing Tan and Quoc V. Le. Efficientnet: Rethinking model scaling for convolutional neural networks. In ICML, pages 6105-6114. PMLR, 2019. 4, 5, 7", + "[36] Keisuke Tateno, Nassir Navab, and Federico Tombari. Distortion-aware convolutional filters for dense prediction in panoramic images. In European Conference on Computer Vision, 2018. 1, 2", + "[37] Fu-En Wang, Yu-Hsuan Yeh, Min Sun, Wei-Chen Chiu, and Yi-Hsuan Tsai. Bifuse: Monocular 360 depth estimation via bi-projection fusion. 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 459–468, 2020. 2, 3, 4, 6", + "[38] Fu-En Wang, Yu-Hsuan Yeh, Yi-Hsuan Tsai, Wei-Chen Chiu, and Min Sun. Bifuse++: Self-supervised and efficient bi-projection fusion for $360^{\\circ}$ depth estimation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45: 5448-5460, 2022. 2, 3, 6", + "[39] Sirui Wang, Di Liang, Jian Song, Yuntao Li, and Wei Wu. DABERT: dual attention enhanced BERT for semantic matching. In COLING, pages 1645-1654. International Committee on Computational Linguistics, 2022. 6" + ], + "bbox": [ + 78, + 90, + 470, + 902 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[40] X. Wang, Ross B. Girshick, Abhinav Kumar Gupta, and Kaiming He. Non-local neural networks. 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7794-7803, 2017. 4", + "[41] Yu yang Li, Yuliang Guo, Zhixin Yan, Xinyu Huang, Ye Duan, and Liu Ren. Omnifusion: 360 monocular depth estimation via geometry-aware fusion. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2791-2800, 2022. 2, 3, 6", + "[42] Xiaoqing Ye, Mao Shu, Hanyu Li, Yifeng Shi, Yingying Li, Guan-Sheng Wang, Xiao Tan, and Errui Ding. Rope3d: The roadside perception dataset for autonomous driving and monocular 3d object detection task. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 21309-21318, 2022. 1", + "[43] Youngho Yoon, Inchul Chung, Lin Wang, and Kuk-Jin Yoon. Spheresr: $360^{\\circ}$ image super-resolution with arbitrary projection via continuous spherical image representation. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5667-5676, 2021. 1, 2, 3", + "[44] Haozheng Yu, Lu He, Bing Jian, Weiwei Feng, and Shanghua Liu. Panelnet: Understanding 360 indoor environment via panel representation. 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 878-887, 2023. 1, 3", + "[45] Ilwi Yun, Chan-Yong Shin, Hyunku Lee, Hyuk-Jae Lee, and Chae-Eun Rhee. Egformer: Equirectangular geometry-biased transformer for 360 depth estimation. ArXiv, abs/2304.07803, 2023. 2, 3, 4, 6", + "[46] Chao Zhang, Stephan Liwicki, William Smith, and Roberto Cipolla. Orientation-aware semantic segmentation on icosahedron spheres. 2019 IEEE/CVF International Conference on Computer Vision (ICCV), pages 3532-3540, 2019. 2, 4", + "[47] Hengshuang Zhao, Li Jiang, Jiaya Jia, Philip H. S. Torr, and Vladlen Koltun. Point transformer. 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 16239-16248, 2020. 4, 7", + "[48] Jia Zheng, Junfei Zhang, Jing Li, Rui Tang, Shenghua Gao, and Zihan Zhou. Structured3d: A large photo-realistic dataset for structured 3d modeling. In European Conference on Computer Vision, 2019. 5, 6", + "[49] Xu Zheng, Jinjing Zhu, Ye-Peng Liu, Zidong Cao, Chong Fu, and Lin Wang. Both style and distortion matter: Dualpath unsupervised domain adaptation for panoramic semantic segmentation. 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1285-1295, 2023. 1", + "[50] Chuanqing Zhuang, Zhengda Lu, Yiqun Wang, Jun Xiao, and Ying Wang. Acdnet: Adaptively combined dilated convolution for monocular panorama depth estimation. CoRR, abs/2112.14440, 2021. 1, 3", + "[51] Nikolaos Zioulis, Antonis Karakottas, Dimitrios Zarpalas, and Petros Daras. Omndepth: Dense depth estimation for indoors spherical panoramicas. In ECCV (6), pages 453-471. Springer, 2018. 1, 2" + ], + "bbox": [ + 501, + 90, + 892, + 881 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "9935", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2024/Elite360D_ Towards Efficient 360 Depth Estimation via Semantic- and Distance-Aware Bi-Projection Fusion/b933bca2-9ff4-42fc-9bd7-be64c39e1936_model.json b/2024/Elite360D_ Towards Efficient 360 Depth Estimation via Semantic- and Distance-Aware Bi-Projection Fusion/b933bca2-9ff4-42fc-9bd7-be64c39e1936_model.json new file mode 100644 index 0000000000000000000000000000000000000000..1457c2c681c2e389942e28ad60a5049cd7009794 --- /dev/null +++ b/2024/Elite360D_ Towards Efficient 360 Depth Estimation via Semantic- and Distance-Aware Bi-Projection Fusion/b933bca2-9ff4-42fc-9bd7-be64c39e1936_model.json @@ -0,0 +1,2255 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.135, + 0.13, + 0.838, + 0.177 + ], + "angle": 0, + "content": "Elite360D: Towards Efficient 360 Depth Estimation via Semantic- and Distance-Aware Bi-Projection Fusion" + }, + { + "type": "text", + "bbox": [ + 0.39, + 0.203, + 0.586, + 0.221 + ], + "angle": 0, + "content": "Hao Ai\\(^{1}\\) Lin Wang\\(^{1,2*}\\)" + }, + { + "type": "text", + "bbox": [ + 0.277, + 0.221, + 0.693, + 0.24 + ], + "angle": 0, + "content": "\\(^{1}\\)AI Thrust, HKUST(GZ) \\(^{2}\\)Dept. of CSE, HKUST" + }, + { + "type": "text", + "bbox": [ + 0.28, + 0.242, + 0.69, + 0.256 + ], + "angle": 0, + "content": "hai033@connect.hkust-gz.edu.cn, linwang@ust.hk" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.292, + 0.314, + 0.308 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.324, + 0.474, + 0.732 + ], + "angle": 0, + "content": "360 depth estimation has recently received great attention for 3D reconstruction owing to its omnidirectional field of view (FoV). Recent approaches are predominantly focused on cross-projection fusion with geometry-based reprojection: they fuse 360 images with equirectangular projection (ERP) and another projection type, e.g., cubemap projection to estimate depth with the ERP format. However, these methods suffer from 1) limited local receptive fields, making it hardly possible to capture large FoV scenes, and 2) prohibitive computational cost, caused by the complex cross-projection fusion module design. In this paper, we propose Elite360D, a novel framework that inputs the ERP image and icosahedron projection (ICOSAP) point set, which is undistorted and spatially continuous. Elite360D is superior in its capacity in learning a representation from a local-with-global perspective. With a flexible ERP image encoder, it includes an ICOSAP point encoder, and a Bi-projection Bi-attention Fusion (B2F) module (totally \\(\\sim 1M\\) parameters). Specifically, the ERP image encoder can take various perspective image-trained backbones (e.g., ResNet, Transformer) to extract local features. The point encoder extracts the global features from the ICOSAP. Then, the B2F module captures the semantic- and distance-aware dependencies between each pixel of the ERP feature and the entire ICOSAP feature set. Without specific backbone design and obvious computational cost increase, Elite360D outperforms the prior arts on several benchmark datasets." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.741, + 0.259, + 0.756 + ], + "angle": 0, + "content": "Multimedia Material" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.76, + 0.47, + 0.791 + ], + "angle": 0, + "content": "For videos, code, demo and more information, you can visit https://VLIS2022.github.io/Elite360D/" + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.795, + 0.21, + 0.81 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.817, + 0.471, + 0.879 + ], + "angle": 0, + "content": "\\(360^{\\circ}\\) images capture the complete surrounding environment in one shot with an ultra-wide field-of-view (FoV) of \\(180^{\\circ} \\times 360^{\\circ}\\), which is broadly applied to applications, e.g., autonomous driving [1, 19, 42, 43, 49] and virtual reality" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.291, + 0.697, + 0.431 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.56, + 0.433, + 0.648, + 0.443 + ], + "angle": 0, + "content": "Model #FLOPs(G)" + }, + { + "type": "image", + "bbox": [ + 0.698, + 0.292, + 0.89, + 0.431 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.691, + 0.434, + 0.708, + 0.447 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.448, + 0.689, + 0.52 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.571, + 0.52, + 0.627, + 0.53 + ], + "angle": 0, + "content": "ERP Image" + }, + { + "type": "image", + "bbox": [ + 0.704, + 0.448, + 0.888, + 0.52 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.773, + 0.52, + 0.821, + 0.53 + ], + "angle": 0, + "content": "GT Depth" + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.532, + 0.689, + 0.604 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.562, + 0.605, + 0.633, + 0.615 + ], + "angle": 0, + "content": "Equi-Baseline" + }, + { + "type": "image", + "bbox": [ + 0.704, + 0.532, + 0.888, + 0.604 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.763, + 0.605, + 0.832, + 0.614 + ], + "angle": 0, + "content": "Our Elite360D" + }, + { + "type": "image_caption", + "bbox": [ + 0.691, + 0.61, + 0.708, + 0.622 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image_caption", + "bbox": [ + 0.5, + 0.623, + 0.892, + 0.693 + ], + "angle": 0, + "content": "Figure 1. (a) Performance (RMSE error) curves on M3D test dataset [5]. Larger circles mean more parameters (e.g., ResNet18, ResNet34, ResNet50) and lower errors mean better performance. (b) Comparison with the ResNet34 as the ERP encoder backbone. With only 1M more parameters, our depth result is more accurate." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.703, + 0.892, + 0.809 + ], + "angle": 0, + "content": "[4, 23, 26]. The ability to infer the 3D structure of a surrounding scene has inspired active research for monocular 360 depth estimation. Equirectangular projection (ERP) is the commonly used projection type that can provide a complete view of a scene. An ERP image, a.k.a., panorama, samples pixels with a higher density at the poles compared to the equator, resulting in spherical distortions (Fig. 2(c))." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.811, + 0.895, + 0.902 + ], + "angle": 0, + "content": "In recent years, to address the spherical distortions, several ah-hoc designs [33, 36, 50, 51] focus on the distortion-aware convolutional kernels and sample grids from undistorted tangent planes on the sphere. These grids are backprojected to the corresponding locations of the ERP image. Also, some methods [28, 34, 44] partition the ERP" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.887, + 0.359, + 0.901 + ], + "angle": 0, + "content": "*Corresponding author (e-mail: linwang@ust.hk)" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.99, + 0.516, + 1.0 + ], + "angle": 0, + "content": "9926" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.304 + ], + "angle": 0, + "content": "input into multiple vertical slices and extract a compact representation from these slices to extend the receptive field of each pixel for the ERP representation. Unfortunately, these methods result in substantial computational costs. Meanwhile, the limited local receptive field of the convolutional filters is insufficient to provide the global perception of the panorama. Therefore, PanoFormer [32] and EGFormer [45] design specific transformer architectures to effectively model the long-range dependencies between the global context of large-FoV ERP images. Nevertheless, the receptive fields are confined by the sizes of local attention windows. Furthermore, the aforementioned data-driven methods often yield sub-optimal results when trained from scratch on limited \\(360^{\\circ}\\) depth dataset (See Tab. 2)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.307, + 0.473, + 0.655 + ], + "angle": 0, + "content": "With the power of existing large-scale perspective images [11], several methods [2, 16, 37, 38] employ the pretrained models - designed for perspective images - as encoder backbones to extract ERP feature maps, and propose the cross-projection fusion to rectify distortions in the ERP feature maps. Specifically, these methods introduce the less-distorted planar projection data, i.e., cubemap projection (CP) patches or tangent projection (TP) patches (See Fig. 2(d), (e)), and unify the spatial dimensions between different projections for the cross-projection fusion. BiFuse and UniFuse [16, 37, 38] propose the C2E module to reproject the content in patch-wise CP features into ERP grid following the spherical geometric relationships. Furthermore, they leverage the fusion between ERP feature maps and C2E feature maps to enhance depth estimation accuracy. However, their concatenation-based feature fusion across multiple scales brings a significant computational burden. Based on the feature similarity, HRDFuse [2] spatially aligns the ERP pixel features and TP patch features. This makes it possible to significantly marry the global context in the ERP feature map and regional structural details in the TP patch features. However, the decoder-level fusion remarkably increases the computational memory and cost." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.659, + 0.473, + 0.902 + ], + "angle": 0, + "content": "These less-distorted gnomonic projections, i.e., CP and TP patches, are spatially discontinuous and require complex re-projection operations for the desired ERP format predictions. Inspired by these issues, we introduce a more powerful projection, the icosahedron projection (ICOSAP) [21, 46], see Fig. 2(f). Importantly, ICOSAP is a spatially continuous and globally perceptive non-Euclidean projection for \\(360^{\\circ}\\) images. In light of this, we propose an efficient and effective fusion-based framework, named Elite360D that takes the best of ERP and ICOSAP by learning a representation from a local-with-global perspective. It comprises three components: an ERP encoder, an ICOSAP encoder, and the bi-projection bi-attention fusion (B2F) module. For the ERP image encoder, our Elite360D flexibly supports a wide range of off-the-shelf 2D models, e.g., ResNet [14], Swin transformer [24], as the backbone to" + }, + { + "type": "image", + "bbox": [ + 0.532, + 0.089, + 0.597, + 0.141 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.533, + 0.143, + 0.596, + 0.152 + ], + "angle": 0, + "content": "(a) Panorama" + }, + { + "type": "image", + "bbox": [ + 0.648, + 0.09, + 0.704, + 0.14 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.66, + 0.143, + 0.696, + 0.151 + ], + "angle": 0, + "content": "(b) ERP" + }, + { + "type": "image", + "bbox": [ + 0.757, + 0.091, + 0.86, + 0.141 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.773, + 0.143, + 0.845, + 0.151 + ], + "angle": 0, + "content": "(c) ERP (flatten)" + }, + { + "type": "image", + "bbox": [ + 0.53, + 0.157, + 0.598, + 0.207 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.55, + 0.209, + 0.579, + 0.217 + ], + "angle": 0, + "content": "(d) CP" + }, + { + "type": "image", + "bbox": [ + 0.645, + 0.157, + 0.711, + 0.207 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.665, + 0.208, + 0.693, + 0.216 + ], + "angle": 0, + "content": "(e) TP" + }, + { + "type": "image", + "bbox": [ + 0.778, + 0.158, + 0.841, + 0.207 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.765, + 0.209, + 0.854, + 0.217 + ], + "angle": 0, + "content": "(f) ICOSAP (level=1)" + }, + { + "type": "image_caption", + "bbox": [ + 0.503, + 0.22, + 0.888, + 0.234 + ], + "angle": 0, + "content": "Figure 2. Different projections of a spherical imaging panorama." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.241, + 0.892, + 0.422 + ], + "angle": 0, + "content": "extract ERP feature maps. This potentially reduces the overfitting problems, particularly on small-scale 360 depth estimation datasets [3, 5]. For efficiency, we represent the ICOSAP spheres as the discrete point sets rather than icosahedron meshes [10, 15, 46] or unfolded representations [21, 43]. As such, we can avoid semantic information redundancy due to dense ERP pixels and maintain the spatial position information. Then, to enable each ERP pixel feature with local receptive fields to perceive the whole scene, B2F module captures the semantic- and distance-aware dependencies between each ERP pixel feature and entire ICOSAP feature set." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.423, + 0.893, + 0.574 + ], + "angle": 0, + "content": "We conduct extensive experiments on three different datasets with different encoder backbones to demonstrate the flexibility and effectiveness of our Elite360D. The experimental results indicate that our Elite360D significantly improves plain-backbones' performance with minimal computational memory (only about 1M parameters) (See Fig. 1(a)). Note that, based on the simple ERP depth estimation baseline, Elite360D only performs bi-projection feature fusion at the last feature layer and achieves results that are on par with leading methods, such as [2, 41]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.574, + 0.894, + 0.756 + ], + "angle": 0, + "content": "In summary, our main contributions are three-fold: (I) We introduce the ICOSAP with spatial continuity and global perception and represent them as the discrete point sets to reduce the computation cost and maintain the spatial information; (II) We propose the B2F module that jointly models the semantic and spatial dependencies between ICOSAP and ERP features and learn the representations with global-with-local receptive fields for large-FoV scenes; (III) Building upon the B2F module, we propose a novel framework that supports diverse off-the-shelf models as encoder backbones, showing better flexibility than the specially designed models, e.g., [32, 45]." + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.77, + 0.651, + 0.785 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.795, + 0.892, + 0.84 + ], + "angle": 0, + "content": "Monocular 360 depth estimation Existing methods can be categorized into two types: one with single projection input, and the other with bi-projection inputs." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.894, + 0.901 + ], + "angle": 0, + "content": "1) Single Projection Input With the ERP images as the inputs, many methods focus on mitigating the inherent spherical distortion issues. Several methods [8, 36, 51] rectify distortions by modifying the receptive fields of tradi" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.989, + 0.516, + 1.0 + ], + "angle": 0, + "content": "9927" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.107, + 0.089, + 0.862, + 0.306 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.307, + 0.894, + 0.364 + ], + "angle": 0, + "content": "Figure 3. (a) An overview of our Elite360D framework, comprising image-based ERP feature extraction (Sec. 3.1), point-based ICOSAP feature extraction (Sec. 3.2), and Bi-projection Bi-attention fusion (B2F) (Sec. 3.3). For better visualization, we do not show the skip connections [30] at the decoding stage. (b) Illustration of the B2F module, consisting of three parts: semantic-aware affinity attention block (Fig. 5), distance-aware affinity attention block (Fig. 6) and gated fusion (Eq. 5)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.37, + 0.473, + 0.567 + ], + "angle": 0, + "content": "tional convolutional kernels. Besides, OmniFusion [41] and 360MonoDepth [29] follow [13] to predict depth maps from less distorted TP patches. Similarly, PanelNet [44] predicts depth maps on the vertical slices of ERP images and then aggregates them. Moreover, some methods [28, 34, 50] focus on tackling the small receptive fields of convolution filters for processing large FoV scenes. More recently, PanoFormer [32] and EGFormer [45] design transformer architectures [12] for ERP images to model the long-range dependencies. By contrast, \\(\\mathbf{S}^2\\mathrm{Net}\\) [22] extracts robust features from ERP images with the Swin transformer [24] and projects the ERP feature maps onto the spherical surface to minimize distortions and maintain spatial consistency." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.568, + 0.472, + 0.734 + ], + "angle": 0, + "content": "2) Bi-Projection Inputs. BiFuse [37] utilizes a dense fusion strategy to bidirectionally fuse ERP and CP features at both encoding and decoding stages. In contrast, BiFuse++[38] and UniFuse [16] fuse ERP and CP features at the encoding stage. Recently, HRDFuse [2] achieved the SOTA performance based on the adaptive fusion of ERP and TP predictions. By contrast, we introduce a non-Euclidean projection, ICOSAP, which is less distorted and spatially continuous. We leverage the global perception capacity of ICOSAP to enable each ERP pixel-wise feature, with a limited local receptive field, to capture the entire scene." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.735, + 0.472, + 0.903 + ], + "angle": 0, + "content": "Representations for 360 Images ERP is the most commonly used projection, which projects a panorama onto a cylinder (Fig. 2(b)) and then unfolds it into a plane (Fig. 2(c)). However, ERP maps latitudes and longitudes onto the vertical and horizontal axes with equal spacing on the plane, leading to distortions. CP [27] projects the panoramas onto six faces of a cube where each face is less distorted (Fig. 2(d)). Building upon CP, some padding methods [7, 37] remove the boundary inconsistency between the adjacent CP patches. Recently, Eder et al. [13] proposed the tangent projection (TP) (Fig. 2(e)), to signif" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.368, + 0.585, + 0.436 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.519, + 0.438, + 0.567, + 0.451 + ], + "angle": 0, + "content": "(a) \\(l = 0\\)" + }, + { + "type": "image", + "bbox": [ + 0.599, + 0.369, + 0.686, + 0.436 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.617, + 0.438, + 0.665, + 0.451 + ], + "angle": 0, + "content": "(b) \\(l = 1\\)" + }, + { + "type": "image", + "bbox": [ + 0.699, + 0.369, + 0.788, + 0.436 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.719, + 0.438, + 0.766, + 0.451 + ], + "angle": 0, + "content": "(c) \\(l = 2\\)" + }, + { + "type": "image", + "bbox": [ + 0.803, + 0.369, + 0.892, + 0.436 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.822, + 0.438, + 0.871, + 0.451 + ], + "angle": 0, + "content": "(d) \\(l = 3\\)" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.461, + 0.892, + 0.629 + ], + "angle": 0, + "content": "Figure 4. The subdivision of icosahedron at different resolution \\(l\\). icantly mitigate spherical distortion. Besides, [21, 31, 43] employ the polyhedrons, e.g., octahedron and icosahedron, to represent the panoramas and unfold these polyhedron representations on the plane for processing. In Elite360D, we select popular ICOSAP (See Fig. 4) as our input. Notably, we represent ICOSAP as discrete points, significantly reducing the computational cost while preserving spatial information and global awareness. Meanwhile, with the dense pixels in the ERP panoramas, discrete ICOSAP points can avoid semantic information redundancy." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.63, + 0.893, + 0.827 + ], + "angle": 0, + "content": "Cross-Attention Mechanism It is widely used for efficient multi-modal feature fusion. For instance, Chen et al. [6] built cross-attention blocks to align and fuse 2D image features and 3D point cloud features for 3D object recognition. Similarly, BEVGuide [25], using BEV embedding as a guided query, employs a cross-attention block to fuse information across different sensors. We propose the B2F module to model the relationships between each pixel of the ERP feature and the whole ICOSAP point feature set, based on the semantic-aware affinity attention and distance-aware affinity attention simultaneously. With the B2F module, each ERP pixel feature can perceive the spatial information and semantic information of the large FoV scenes." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.835, + 0.745, + 0.851 + ], + "angle": 0, + "content": "3. The Proposed Framework" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.856, + 0.894, + 0.901 + ], + "angle": 0, + "content": "Overview. Our goal is to investigate an efficient and effective bi-projection fusion framework for 360 depth estimation. To this end, as illustrated in Fig. 3, the proposed" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.99, + 0.516, + 1.0 + ], + "angle": 0, + "content": "9928" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.473, + 0.455 + ], + "angle": 0, + "content": "Elite360D consists of three key components: image-based ERP feature extraction, point-based ICOSAP feature extraction and bi-projection bi-attention fusion (B2F). For the ERP feature \\( F^{E} \\) extraction, Elite360D flexibly accommodates a wide range of 2D models as encoder backbones and takes advantage of pre-training to reduce the over-fitting problems. For the ICOSAP feature extraction, we represent ICOSAP sphere as the point set, which significantly reduces the computation cost and preserves the spatial information and global perception. Then we directly employ the simple point encoder, i.e. Point transformer [47], to extract the ICOSAP point feature set \\( \\{F_{n}^{I}\\}_{n = 1,\\dots ,N} \\), \\( N \\) is the point number, denoted as \\( \\{F_n^I\\}_N \\). Next, to facilitate bi-projection feature fusion and learn a powerful representation \\( F^{GL} \\) from a local-with-global perspective, B2F module comprises semantic-aware affinity attention block, distance-aware affinity attention block, and gated fusion block. These two attention blocks model the semantic- and distance-aware dependencies between each ERP pixel-wise feature and entire ICOSAP feature set, respectively and the gated fusion block adaptively fuses the bi-attention outputs. Lastly, we employ a simple decoder with up-sampling and skip-connections to predict the final depth map from the fused representation \\( F^{GL} \\). We now describe the details." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.461, + 0.303, + 0.476 + ], + "angle": 0, + "content": "3.1. ERP Feature Extraction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.481, + 0.473, + 0.707 + ], + "angle": 0, + "content": "Taking an ERP image with the resolution size of \\( H \\times W \\) as the input, the encoder extracts feature map \\( F^{E} \\in \\mathbb{R}^{h \\times w \\times C} \\), where \\( h = H / s \\), \\( w = W / s \\), \\( s \\) is the down-sampling scale factor and \\( C \\) is the channel number. In particular, as treating ERP images as 2D perspective images, our encoder backbone is compatible with a wide range of robust 2D models that have been pre-trained on large-scale perspective image datasets [11], including CNNs and vision transformers, e.g., ResNet [14], EfficientNet [35], Swin transformer [24]. Notably, as repeated local operations cannot substitute for a global operator [40, 45], the finite size of convolutional kernels or the limited size of local attention windows results in ERP pixel-wise features lacking sufficient global receptive fields. Additionally, these 2D model performance suffers from the spherical distortions (See Tab. 1)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.715, + 0.334, + 0.73 + ], + "angle": 0, + "content": "3.2. ICOSAP Feature Extraction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.734, + 0.47, + 0.886 + ], + "angle": 0, + "content": "Preliminary. Since Lee et al. [21] found that polyhedron representations with a greater number of initial faces exhibit lower distortion, this paper focuses on the ICOSAP with the most initial faces. Especially, we introduce the ICOSAP to offer a comprehensive and high-quality global perception. For the resolution, ICOSAP data consists of \\(20 \\times 4^l\\) faces and \\(12 \\times 4^l\\) vertices at the subdivision level \\(l\\) (See Fig. 4). Given the spatial relationships between ICOSAP and ERP within the spherical space, RGB values of ICOSAP vertices can be derived from corresponding ERP pixels." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.887, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Feature extraction. Existing works focus on processing" + }, + { + "type": "image", + "bbox": [ + 0.52, + 0.09, + 0.872, + 0.228 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.23, + 0.892, + 0.272 + ], + "angle": 0, + "content": "Figure 5. The architecture of semantic-aware affinity attention. Especially, \\(\\mathbf{Q}_{i,j}^{S}\\in \\mathbb{R}^{1\\times d}\\), \\(\\mathbf{K}^S\\in \\mathbb{R}^{N\\times d}\\), and \\(\\mathbf{V}^S\\in \\mathbb{R}^{N\\times d}\\), where \\(d\\) is the dimension and \\(N\\) is the ICOSAP point number." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.278, + 0.895, + 0.705 + ], + "angle": 0, + "content": "ICOSAP data on the plane, i.e., unfolded mesh representation [46] and spherical polyhedron representations [21]. However, these methods require specially designed and computationally intensive operations, such as convolutions, pooling [21], and up-sampling [31], to process the planar representations. By contrast, we propose to represent an ICOSAP input as a point set, as illustrated at the bottom of Fig. 3(a). ERP images provide dense pixel values and discrete ICOSAP point set can prevent redundancy in semantic information while preserving spatial information and global perception. Therefore, we take full advantage of two projections for efficient and accurate 360 depth estimation. In detail, we employ the central points of each face to represent the ICOSAP sphere. Firstly, we obtain the \\(20 \\times 4^l\\) faces of an ICOSAP sphere at the default subdivision level \\(l\\), where each face is composed of three vertices. Then, we calculate the spatial coordinates and RGB values of each face center by averaging them of each three vertices, as each face is an equilateral triangle. As a result, we obtain the ICOSCAP point set \\(\\{P^{ICO}\\} \\in \\mathbb{R}^{(20 \\times 4^l) \\times 6}\\), where \\(20 \\times 4^l\\) is the point number and 6 represents the coordinates \\([x, y, z]\\) and RGB channels. With the input point set \\(\\{P^{ICO}\\}\\), we directly employ the encoder of Point Transformer [47] to extract ICOSAP point feature set \\(\\{F_n^I\\}_N \\in \\mathbb{R}^{N \\times C}\\), where \\(N\\) represents the number of point features and \\(C\\) corresponds to the same channel number as the ERP feature map \\(F^E\\). Especially, for simplicity and efficiency, we do not opt for complex networks for this purpose." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.712, + 0.796, + 0.728 + ], + "angle": 0, + "content": "3.3. Bi-Projection Bi-Attention Fusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.735, + 0.895, + 0.903 + ], + "angle": 0, + "content": "In earlier methods [37] and [16], bi-projection feature fusion primarily depends on the geometric relationships between CP and ERP. They first apply the C2E function to re-project CP feature patches into ERP format feature maps and subsequently perform pixel-wise feature concatenation. While this fusion is effective, it brings several problems: 1) geometry-based re-projection and concatenation-based feature fusion significantly increase the computational costs (See Tab. 2); 2) the geometry-based fusion process from CP patches to ERP panorama restricts ERP pixels from the global perception, where each ERP pixel perceives scene in-" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.99, + 0.516, + 1.0 + ], + "angle": 0, + "content": "9929" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.14, + 0.089, + 0.832, + 0.442 + ], + "angle": 0, + "content": "
DatasetsBackboneMethod#Params (M)#FLOPs (G)Abs Rel ↓Sq Rel ↓RMSE ↓δ1(%) ↑δ2(%) ↑δ3(%) ↑
M3D [5]ResNet-18 [14]Equi-Baseline14.3939.840.14280.12030.560780.9794.6898.26
Ours15.4345.910.12720.10700.527085.2895.2898.49
Δ+1.04+6.07-11.06%-9.73%-6.01%+4.31+0.60+0.23
ResNet-34 [14]Equi-Baseline24.5059.270.12550.10480.517385.5296.1098.55
Ours25.5465.290.11150.09140.487588.1596.4698.74
Δ+1.04+6.02-11.16%-12.79%-5.76%+2.63+0.36+0.19
ResNet-50* [14]Equi-Baseline29.6375.130.13700.12090.543282.5194.8398.07
Ours42.99170.110.11120.09800.487086.7096.0198.61
Δ+13.36+94.98-18.75%-18.94%-10.35%+4.19+1.18+0.54
EfficientNet-B5 [35]Equi-Baseline33.1018.810.10340.08310.463889.9296.8098.80
Ours34.1124.880.10480.08050.452489.9297.0799.14
Δ+1.01+6.07+1.35%-3.13%-2.46%-0.00+0.27+0.34
SWin-B* [24]Equi-Baseline90.75187.370.13010.12500.555784.0694.7697.72
Ours94.30211.280.12490.12100.546285.3495.0797.77
Δ+3.55+23.91-4.00%-3.20%-1.71%+1.28+0.31+0.05
DilateFormer-T [17]Equi-Baseline19.9346.960.14290.12740.574881.1594.4697.79
Ours20.9453.020.14230.12510.551782.6895.0098.14
Δ+1.01+6.06-0.42%-1.81%-4.02%+1.53+0.54+0.35
S2D3D [3]ResNet-34 [14]Equi-Baseline24.5059.220.12030.07540.372487.4196.5198.77
Ours25.5165.280.11820.07280.375688.7296.8498.92
Δ+1.01+6.06-9.21%-3.45%+0.86%+1.31+0.33+0.15
EfficientNet-B5 [35]Equi-Baseline33.1018.810.10260.06380.358089.4397.0699.16
Ours34.1124.880.10180.06030.357589.4797.2399.22
Δ+1.01+6.07-0.78%-5.49%-0.14%+0.03+0.17+0.06
Struct3D [48]ResNet-34 [14]Equi-Baseline24.5059.220.22560.39100.764176.9089.9694.49
Ours25.5165.280.14800.22150.496187.4194.3496.66
Δ+1.01+6.06-34.40%-43.35%-35.07%+10.51+4.38+2.17
EfficientNet-B5 [35]Equi-Baseline33.1018.810.13120.19380.431288.5395.2197.40
Ours34.1124.880.12770.19300.415189.1695.3397.43
Δ+1.01+6.07-2.67%-0.41%-3.73%+0.63+0.12+0.03
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.447, + 0.892, + 0.478 + ], + "angle": 0, + "content": "Table 1. Quantitative comparison with ERP-based depth baseline. For most conditions, we set the channel number \\( C \\) to 64. Especially, we set \\( C \\) to 256 for ResNet-50* and to 128 for Swin-B*. Bold indicates performance improvement. Red indicates performance decline." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.489, + 0.473, + 0.899 + ], + "angle": 0, + "content": "formation only from its corresponding small-FoV CP patch without considering other patches; 3) this one-to-one alignment pattern only emphasizes spatial consistency, with no consideration for the semantic similarity. Thus, we design the Bi-Projection Bi-Attention Fusion (B2F) module to solve the above problems. As shown in Fig. 3(b), B2F module first leverages semantic-aware affinity attention and distance-aware affinity attention to model the semantic and spatial dependencies between each ERP pixel-wise feature \\( F_{i,j}^{E} \\) and ICOSAP point feature set \\( \\{F_n^I\\}_N \\). Here \\( (i,j) \\) indicates the coordinate of a pixel in the ERP feature map, \\( i \\in (1,h), j \\in (1,w) \\), and \\( N \\) is the point number of the ICOSAP feature set. Consequently, a gated fusion block is employed to adaptively marry semantic- and distance-related information. Semantic-aware affinity attention. The semantic-aware affinity attention, as shown in Fig. 5, follows the standard dot-product attention [12]. In detail, given the extracted ERP feature map \\( F^{E} \\in \\mathbb{R}^{h \\times w \\times C} \\) and ICOSAP point feature set \\( \\{F_n^I\\}_N \\in \\mathbb{R}^{N \\times C} \\) (for simplicity, we denote it as \\( [F^I] \\)), we generate the query \\( \\mathbf{Q}_{i,j}^{S} \\) from each ERP pixel-wise feature \\( F_{i,j}^{E} \\in \\mathbb{R}^{1 \\times C} \\) and produce the key \\( \\mathbf{K}^{S} \\) and value \\( \\mathbf{V}^{S} \\) from the whole ICOSAP point feature set \\( [F^{I}] \\). Then, we calculate the attention weight \\( \\mathbf{A}_{i,j}^{S} \\) based on \\( \\mathbf{Q}_{i,j}^{S} \\) and \\( \\mathbf{K}^{S} \\), and obtain the affinity feature \\( F_{i,j}^{SA} \\) with \\( F_{i,j}^{SA} = \\mathbf{A}_{i,j}^{S} * \\mathbf{V}^{S}, \\mathbf{A}_{i,j}^{S} = \\text{softmax}\\left(\\frac{\\mathbf{Q}_{i,j}^{S} \\mathbf{K}^{ST}}{\\sqrt{d}}\\right) \\), where \\( d \\) is the channel number and" + }, + { + "type": "image", + "bbox": [ + 0.532, + 0.486, + 0.859, + 0.655 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.66, + 0.892, + 0.691 + ], + "angle": 0, + "content": "Figure 6. The architecture of distance-aware affinity attention, especially, \\( Q_{i,j}^{D} \\in R^{1 \\times d} \\), \\( K^{D} \\in \\mathbb{R}^{N \\times d} \\), and \\( \\mathbf{V}^{D} \\in \\mathbb{R}^{N \\times d} \\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.702, + 0.892, + 0.793 + ], + "angle": 0, + "content": "\\(d = C\\). After querying all the ERP pixel-wise features, we can obtain the ERP format feature map \\(F^{SA}\\) with the dimension of \\(\\mathbb{R}^{h\\times w\\times d}\\). The attention weight \\(A_{i,j}^{S}\\) captures affinities based on the semantic-aware feature similarities, and the output \\(F^{SA}\\) effectively integrates the global and local receptive fields." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.796, + 0.893, + 0.903 + ], + "angle": 0, + "content": "Distance-aware affinity attention. Distance-aware affinity attention captures the differences between ERP and ICOSAP in both spatial and semantic information. This can better utilize the geometric prior knowledge of panoramicas and precisely measure the distances between two different projection features. As depicted in Fig. 6, the distance-aware affinity attention is built upon the subtraction-based" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.99, + 0.516, + 1.0 + ], + "angle": 0, + "content": "9930" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.14, + 0.089, + 0.833, + 0.372 + ], + "angle": 0, + "content": "
DatasetsBackboneMethodPub'Year#Params (M)#FLOPs (G)Abs Rel ↓Sq Rel ↓RMSE ↓δ1(%) ↑δ2(%) ↑δ3(%) ↑
M3D [5]TransformerEGFormer [45]ICCV'2315.3966.210.14730.15170.602581.5893.9097.35
PanoFormer [32]ECCV'2220.3881.090.10510.09660.492989.0896.2398.31
ResNet-18 [14]BiFuse [37]CVPR'2035.80165.660.13600.12020.548883.2795.1298.10
UniFuse [16]RAL'2130.2662.600.11910.10300.515886.0495.8498.30
Omnifusion [41]CVPR'2232.3598.680.12090.10900.505586.5895.8198.36
HRDFuse† [2]CVPR'2326.0950.590.14140.12410.550781.4894.8998.20
Ours-15.4345.910.12720.10700.527085.2895.2898.49
ResNet-34 [14]BiFuse [37]CVPR'2056.01199.580.11260.09920.502788.0096.1398.47
BiFuse++ [38]TPAMI'2252.4987.480.11230.09150.485388.1296.5698.69
UniFuse [16]RAL'2150.4896.520.11440.09360.483587.8596.5998.73
Omnifusion [41]CVPR'2242.46142.290.11610.10070.493187.7296.1598.44
HRDFuse† [2]CVPR'2346.3180.870.11720.09710.502586.7496.1798.49
Ours-25.5465.290.11150.09140.487588.1596.4698.74
ResNet-50* [14]BiFuse [37]CVPR'20253.08775.240.11790.09810.497086.7496.2798.66
UniFuse [16]RAL'21131.30222.300.11850.09840.502486.6696.1898.50
Ours-42.99170.110.11120.09800.487086.7096.0198.61
S2D3D [3]TransformerEGFormer [45]ICCV'2315.3966.210.15280.14080.497481.8593.3897.36
PanoFormer [32]ECCV'2220.3881.090.11220.07860.394588.7495.8498.59
ResNet-34 [14]Omnifusion [41]CVPR'2242.46142.290.11540.07750.380986.7496.0398.71
UniFuse [16]RAL'2150.4896.520.11240.07090.355587.0697.0498.99
Ours-25.5165.280.11820.07280.375688.7296.8498.92
Struct3D [48]TransformerEGFormer [45]ICCV'2315.3966.210.22050.45090.684179.7990.7194.55
PanoFormer [32]ECCV'2220.3881.090.25490.49490.793774.7089.1593.97
ResNet-34 [14]BiFuse [37]CVPR'2056.01199.580.15730.24550.521385.9194.0096.72
UniFuse [16]RAL'2150.4896.520.15060.23190.501685.4293.9996.76
Ours-25.5165.280.14800.22150.496187.4194.3496.66
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.373, + 0.89, + 0.403 + ], + "angle": 0, + "content": "Table 2. Quantitative comparison with the SOTA methods. \\( {}^{ \\dagger } \\) means that we modify the HRDFuse network structure for a fair comparison. Green represents the best performance under the given encoder backbone." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.41, + 0.47, + 0.492 + ], + "angle": 0, + "content": "cross-attention mechanism [39]. Given the ERP pixel-wise feature \\( F_{i,j}^{E} \\) and ICOSAP point feature set \\( [\\pmb{F}^{\\pmb{I}}] \\), we firstly calculate the spatial distance embedding \\( Dis^{SP} \\) from the spatial coordinates of ERP pixel and ICOSAP point set, i.e. \\( [x_{i,j}^{e},y_{i,j}^{e},z_{i,j}^{e}] \\) and \\( \\{[x_n^{ico},y_n^{ico},z_n^{ico}]\\}_{N} \\), as:" + }, + { + "type": "equation", + "bbox": [ + 0.128, + 0.496, + 0.469, + 0.517 + ], + "angle": 0, + "content": "\\[\nD i s _ {i, j} ^ {S P} = \\left[ e ^ {- \\Delta x _ {i, j}}, e ^ {- \\Delta y _ {i, j}}, e ^ {- \\Delta z _ {i, j}} \\right] \\mathbf {W} ^ {S P}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.526, + 0.47, + 0.655 + ], + "angle": 0, + "content": "where linear projection \\(\\mathbf{W}^{SP}\\in \\mathbb{R}^{3\\times d}\\) \\(Dis_{i,j}^{SP}\\in \\mathbb{R}^{1\\times N\\times d}\\) and \\([\\Delta x_{i,j},\\Delta y_{i,j},\\Delta z_{i,j}]\\in \\mathbb{R}^{1\\times N\\times 3}\\) is the distances between \\([x_{i,j}^e,y_{i,j}^e,z_{i,j}^e ]\\) and \\(\\{[x_n^{ico},y_n^{ico},z_n^{ico}]\\}_{N}\\). In particular, the operation \\(e^{-(\\cdot)}\\) allows \\(Dis^{SP}\\) to pay more attention to the close parts between ERP pixels and ICOSAP point set. After that, we produce the query \\(\\mathbf{Q}_{i,j}^{D}\\) and key \\(\\mathbf{K}^D\\) from \\(F_{i,j}^{E}\\) and \\([\\pmb {F}^{\\pmb{I}}]\\), respectively and calculate the semantic distance embedding \\(Dis_{i,j}^{SE}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.188, + 0.657, + 0.469, + 0.68 + ], + "angle": 0, + "content": "\\[\nD i s _ {i, j} ^ {S E} = e ^ {- \\left\\| \\mathbf {Q} _ {i, j} ^ {D} - \\mathbf {K} ^ {D} \\right\\|}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.681, + 0.469, + 0.78 + ], + "angle": 0, + "content": "where \\(\\mathbf{W}_Q^D\\), \\(\\mathbf{W}_K^D \\in \\mathbb{R}^{C \\times d}\\) are linear projections, \\(\\mathbf{Q}_{i,j}^D \\in \\mathbb{R}^{1 \\times d}\\), \\(\\mathbf{K}^D \\in \\mathbb{R}^{N \\times d}\\), and \\(Dis_{i,j}^{SE} \\in \\mathbb{R}^{1 \\times N \\times d}\\). Lastly, the distance-aware attention weight \\(\\mathbf{A}_{i,j}^D\\) is generated with spatial and semantic distance embeddings, and the distance-aware affinity feature vector \\(F_{i,j}^{DA}\\) is obtained from the attention weight \\(\\mathbf{A}_{i,j}^D\\) and the value \\(\\mathbf{V}^D\\):" + }, + { + "type": "equation", + "bbox": [ + 0.15, + 0.782, + 0.469, + 0.814 + ], + "angle": 0, + "content": "\\[\n\\mathbf {A} _ {i, j} ^ {D} = \\operatorname {s o f t m a x} \\left(\\frac {\\sum \\left(D i s _ {i , j} ^ {S P} + D i s _ {i , j} ^ {S E}\\right)}{\\sqrt {d}}\\right), \\tag {3}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.164, + 0.821, + 0.469, + 0.838 + ], + "angle": 0, + "content": "\\[\n\\mathbf {V} ^ {D} = F ^ {I} \\mathbf {W} _ {V} ^ {D}, F _ {i, j} ^ {D A} = \\mathbf {A} _ {i, j} ^ {D} * \\mathbf {V} ^ {D}, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.84, + 0.469, + 0.901 + ], + "angle": 0, + "content": "where \\(\\sum\\) means the sum for the channel dimension, i.e. \\(\\sum (Dis_{i,j}^{SP} + Dis_{i,j}^{SE})\\in \\mathbb{R}^{1\\times N}\\). After querying all ERP pixel-wise features, we obtain the distance-aware aggregated feature \\(F^{DA}\\in \\mathbb{R}^{h\\times w\\times d}\\), \\(d = C\\)." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.41, + 0.892, + 0.48 + ], + "angle": 0, + "content": "Gated fusion. Since direct average or concatenation may compromise the original representation ability, inspired by [9], we propose the gated fusion block to adaptively fuse \\( F^{SA} \\) and \\( F^{DA} \\) and obtain the representations \\( F^{GL} \\) from a local-with-global perspective, formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.586, + 0.482, + 0.891, + 0.5 + ], + "angle": 0, + "content": "\\[\nF ^ {G L} = g ^ {S A} * F ^ {S A} + g ^ {D A} * F ^ {D A}, \\tag {5}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.589, + 0.503, + 0.803, + 0.521 + ], + "angle": 0, + "content": "\\[\ng ^ {S A} = \\sigma_ {S A} \\left(W _ {g} ^ {S A} \\cdot \\left[ F ^ {S A}; F ^ {D A} \\right]\\right),\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.587, + 0.523, + 0.805, + 0.541 + ], + "angle": 0, + "content": "\\[\ng ^ {D A} = \\sigma_ {D A} \\left(W _ {g} ^ {D A} \\cdot \\left[ F ^ {S A}; F ^ {D A} \\right]\\right),\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.548, + 0.892, + 0.58 + ], + "angle": 0, + "content": "where \\(W_{g}^{SA}\\) and \\(W_{g}^{DA}\\) are linear projections, \\(\\sigma_{SA}(\\cdot)\\) and \\(\\sigma_{DA}(\\cdot)\\) are the sigmoid functions." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.589, + 0.64, + 0.605 + ], + "angle": 0, + "content": "3.4. Optimization" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.611, + 0.892, + 0.718 + ], + "angle": 0, + "content": "With the fused feature \\( F^{GL} \\) and multi-scale ERP feature maps in the ERP encoder backbone, we feed them into a decoder [30] with several up-sampling blocks and skipconnections to output the final depth. For the depth supervision, we follow existing works [28, 32] and employ the combination of Berhu loss [20] and gradient loss [28]. (Details of training loss can be found in the suppl. material.)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.725, + 0.634, + 0.741 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.498, + 0.745, + 0.892, + 0.762 + ], + "angle": 0, + "content": "4.1. Datasets, Metrics, and Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Datasets and Metrics. We evaluate Elite360D on three datasets: two real-world datasets, Matterport3D(M3D) [5], Stanford2D3D(S2D3D) [3], and a recently proposed large-scale synthetic dataset, Structure3D(Struct3D) [48]. For the evaluation metrics, we follow previous works [16, 32, 45] to employ some standard metrics, including absolute relative error (Abs Rel), squared relative error (Sq Rel), root mean squared error (RMSE), and three threshold percentage \\(\\delta < \\alpha^t\\) (\\(\\alpha = 1.25, t = 1, 2, 3\\)), denoted as \\(\\delta_t\\). Addi" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "9931" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.108, + 0.088, + 0.859, + 0.435 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.078, + 0.437, + 0.891, + 0.451 + ], + "angle": 0, + "content": "Figure 7. Qualitative results (with ResNet-34 as the backbone) on Matterport3D (top), Stanford2D3D (middle) and Structure3D (bottom)." + }, + { + "type": "table", + "bbox": [ + 0.114, + 0.459, + 0.432, + 0.552 + ], + "angle": 0, + "content": "
BackbonePre-trainedAbs Rel↓Sq Rel↓RMSE↓δ1↑
ResNet-34X0.15960.14520.585681.36
✓0.11150.09140.487588.15
EfficientNet-B5X0.12110.10870.513186.80
✓0.10480.08050.452489.92
DilateFormerX0.15150.14150.569480.95
✓0.14230.12510.551782.68
" + }, + { + "type": "table_caption", + "bbox": [ + 0.083, + 0.554, + 0.462, + 0.567 + ], + "angle": 0, + "content": "Table 3. The impact of pre-training on ERP encoder backbones." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.577, + 0.468, + 0.606 + ], + "angle": 0, + "content": "tionally, we measure the number of parameters and FLOPS to evaluate the efficiency of our method." + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.607, + 0.469, + 0.848 + ], + "angle": 0, + "content": "Training Details. We use diverse ERP encoder backbones, including CNNs (ResNet-18, 34, 50 [14], EfficientNet B5 [35]), and transformers (Swin-B [24], DilateFormer-T [17]). All backbones are pre-trained on ImageNet1K [11]. We set the default channel number \\( C \\) to 64 and default subdivision level of ICOSAP as \\( l = 4 \\). For the ICOSAP encoder, we employ the one of Point transformer [47] with three down-sample blocks. Following [16], we use Adam optimizer [18] and a constant learning rate of \\( 1\\mathrm{e}^{-4} \\). Considering the unfair comparisons stemming from variations in hyper-parameters and validation procedures used across different methods, we re-train the existing methods from scratch and validate them, following the unified training and validation settings [16]. (Due to page limit, detailed training and validation settings can be found in suppl. mat.)." + }, + { + "type": "title", + "bbox": [ + 0.076, + 0.851, + 0.312, + 0.867 + ], + "angle": 0, + "content": "4.2. Performance Comparison" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.871, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Comparisons with ERP-based depth baselines. As shown in Tab. 1, with an increase of only \\(\\sim 1\\mathrm{M}\\) parameters" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.462, + 0.892, + 0.901 + ], + "angle": 0, + "content": "\\((C = 64)\\), our Elite360D demonstrates substantial advancements over the ERP-based baselines across different ERP encoder backbones on all three datasets. Specifically, for the Matterport3D dataset, Elite360D achieves reductions exceeding \\(10\\%\\) in Abs Rel error (ResNet-18, 34), along with reductions of \\(4.00\\%\\) in Abs Rel error (Swin-B) and \\(4.02\\%\\) in RMSE error (DilateFormer-T). Besides, with the larger channel number \\(C = 256\\) (ResNet-50), Elite360D outperforms ERP baseline by \\(18.75\\%\\) (Abl Rel), \\(18.94\\%\\) (Sq Rel). For the small-scale S2D3D dataset, Elite360D outperforms ERP baseline by \\(9.21\\%\\) in Abs Rel error and \\(1.31\\%\\) in accuracy \\(\\delta_{1}\\) (ResNet-34), as well as \\(5.49\\%\\) in Sq Rel error (EfficientNet-B5). Remarkably, on the larger-scale Structure3D, Elite360 performs favorably against the baseline by a significant margin, especially with ResNet-34. Comparisons with prevalent methods. In Tab. 2, we conduct a comprehensive comparison with prevalent supervised methods. From the results, we can observe that our approach achieves similar or even superior performance compared to existing both bi-projection fusion methods and single input methods at a significantly lower cost, particularly on two large-scale datasets, Matterport3D and Structure3D. Specifically, for the Matterport3D dataset, our Elite360D with ResNet-34 outperforms UniFuse by \\(2.53\\%\\) (Abs Rel) and with ResNet-50 outperforms BiFuse by \\(2.01\\%\\) (RMSE). For the Structure3D dataset, our Elite360D with ResNet-34 outperforms UniFuse by \\(4.48\\%\\) (Seq Rel), \\(1.99\\%\\) (\\(\\delta_{1}\\)). For performance on the Stanford2D3D dataset, we suspect it might be related to the ICOSAP point encoder." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "9932" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.119, + 0.09, + 0.428, + 0.156 + ], + "angle": 0, + "content": "
Bi-projection feature fusionAbs Rel ↓Sq Rel ↓RMSE ↓δ1 ↑
SFA [2] + Add0.12760.10020.515084.27
SFA [2] + Concat0.11910.10190.514386.52
Only SA0.12040.10140.512186.26
Only DA0.11840.09720.494487.06
Our B2F (SA + DA)0.11150.09140.487588.15
" + }, + { + "type": "table_caption", + "bbox": [ + 0.138, + 0.158, + 0.409, + 0.17 + ], + "angle": 0, + "content": "Table 4. The ablation results for B2F module." + }, + { + "type": "table", + "bbox": [ + 0.119, + 0.181, + 0.428, + 0.255 + ], + "angle": 0, + "content": "
Final fusionAbs Rel ↓Sq Rel ↓RMSE ↓δ1 ↑
Add0.16850.14810.580974.60
Average0.11980.09180.489386.65
Concatenation0.11450.09370.488087.66
Adaptive fusion [2]0.12440.09680.489186.08
Our gated fusion0.11150.09140.487588.15
" + }, + { + "type": "table_caption", + "bbox": [ + 0.098, + 0.257, + 0.447, + 0.27 + ], + "angle": 0, + "content": "Table 5. The ablation results for the fusion of B2F module." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.28, + 0.47, + 0.401 + ], + "angle": 0, + "content": "The limited data of Stanford2D3D dataset restricts the ability of the transformer-based point encoder to provide accurate global perception. Moreover, in Fig. 7, we present the qualitative comparisons. Our Elite360D can predict more accurate depth values based on the local-with-global perception capabilities (e.g., flowers, shelves and doors). Additional qualitative results and inference time comparisons can be found in the suppl. material." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.41, + 0.336, + 0.426 + ], + "angle": 0, + "content": "4.3. Ablation Study and Analyses" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.43, + 0.469, + 0.46 + ], + "angle": 0, + "content": "Most of ablation experiments are conducted on the Mateport3D test dataset with ResNet34 as the backbone." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.462, + 0.469, + 0.657 + ], + "angle": 0, + "content": "The Effect of pre-training. We verify the effectiveness of ImageNet [11] pre-training with different encoder backbones. As observed from Tab. 3, the pre-training results in a significant improvement for all encoder backbones, e.g., \\(6.79\\%\\) improvement in accuracy \\(\\delta_{1}\\) (ResNet-34). Notably, pre-training has a relatively small impact on DilateFormer. Combined with the results in Tab. 1, the explanation of this phenomenon is that the default input resolution in pretrained models is different from actual input, thereby impacting the resolution-related position embeddings. In general, pre-training based on large-scale perspective images can effectively enhance the performance of models based on \\(360^{\\circ}\\) images and reduce the risk of overfitting." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.659, + 0.469, + 0.869 + ], + "angle": 0, + "content": "The effectiveness of B2F module. In Tab. 4, we compare four available bi-projection feature fusion modules. To align the spatial dimensions between ICOSAP point feature set and ERP feature map, we introduce SFA module from [2]. After that, we employ direct addition and concatenation to aggregate these two projections. We also achieve the bi-projection feature fusion with semantic-aware affinity attention (SA) alone and distance-aware affinity attention (DA) alone. Compared to the methods based solely on semantic-aware feature similarities (The first three rows), single distance-aware affinity attention can achieve better performance, which indicates that the spatial positional relationships boost the bi-projection feature fusion. Overall, our B2F module achieves the best performance." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.469, + 0.901 + ], + "angle": 0, + "content": "The superiority of ICOSAP. As only CP/TP's patch centers lie on the sphere's surface, we extract the feature em" + }, + { + "type": "table", + "bbox": [ + 0.504, + 0.09, + 0.892, + 0.149 + ], + "angle": 0, + "content": "
Method#Param(M)#FLOPs(G)Abs Rel ↓RMSE ↓δ1 ↑
ERP-CP25.6654.150.13690.540183.69
ERP-TP (N=18)25.6650.580.13280.538583.87
ERP-ICOSAP (Ours)15.4345.910.12720.527085.28
" + }, + { + "type": "table_caption", + "bbox": [ + 0.501, + 0.155, + 0.89, + 0.169 + ], + "angle": 0, + "content": "Table 6. The comparison of different projections on Matterport3D." + }, + { + "type": "table", + "bbox": [ + 0.504, + 0.172, + 0.892, + 0.223 + ], + "angle": 0, + "content": "
N of {F^I}#Params (M)#FLOPs (G)Abs Rel↓Sq Rel↓RMSE↓δ1↑
2027.4166.290.11570.09950.502487.12
8025.5465.290.11150.09140.487588.15
32024.9864.320.11530.09430.490587.85
" + }, + { + "type": "table_caption", + "bbox": [ + 0.5, + 0.225, + 0.89, + 0.253 + ], + "angle": 0, + "content": "Table 7. Impact of the ICOSAP point-wise feature number \\(N\\). Larger \\(N\\), fewer down-sampling blocks in the point encoder." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.258, + 0.892, + 0.349 + ], + "angle": 0, + "content": "bedding from each CP/TP patch and employs the patch center coordinates and feature embedding as the input of B2F module. In Tab. 6, we show the results with ResNet18 backbone. Our Elite360D, utilizing the ICOSAP point set, marginally outperforms models with CP and TP patches, while exhibiting fewer parameters and FLOPs." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.349, + 0.892, + 0.424 + ], + "angle": 0, + "content": "The effectiveness of gated fusion. We conduct an ablation study for the gated fusion block, outlined in Tab. 5. With the feature maps \\( F^{SA} \\) and \\( F^{DA} \\), We compare it with the direct addition, average fusion, concatenation, and the adaptive fusion in [2]. The gated fusion performs best." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.425, + 0.892, + 0.546 + ], + "angle": 0, + "content": "ICOSAP point feature number \\(N\\). We study the effect of the ICOSAP feature point number (See Tab. 7). Too few points \\((N = 20)\\) lead to the over-concentrated global contextual information resulting from excessive down-sampling blocks, while too many points \\((N = 320)\\) lead to underconcentrated condition, resulting in insufficient perception of ERP pixel features. Best performance can be observed when \\(N = 80\\) and we used \\(N = 80\\) as default in this paper." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.553, + 0.769, + 0.568 + ], + "angle": 0, + "content": "5. Conclusion and Future Work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.571, + 0.892, + 0.813 + ], + "angle": 0, + "content": "In this paper, we proposed a novel bi-projection fusion solution for efficient 360 depth estimation. To address the limited local receptive field of ERP pixel-wise features and avoid expensive bi-projection fusion modules, we proposed a compact yet effective B2F module to learn the representations with local-with-global perceptions from ERP and ICOSAP. With an increase of 1M parameters, we significantly improved the performance of the ERP-based depth estimation baseline. Remarkably, our approach achieved performance on par with complex state-of-the-art methods. Future Work: From the experimental results, we observed that ERP-based depth baseline, with pre-trained Efficient-Net backbone, even outperforms existing specifically designed methods. Therefore, in the future, we will explore how to fully leverage different projections and successful perspective models for \\(360^{\\circ}\\) community." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.821, + 0.661, + 0.838 + ], + "angle": 0, + "content": "Acknowledgement" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.892, + 0.901 + ], + "angle": 0, + "content": "This paper is supported by the National Natural Science Foundation of China (NSF) under Grant No. NSFC22FYT45 and the Guangzhou City, University and Enterprise Joint Fund under Grant No.SL2022A03J01278." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9933" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.105 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.169 + ], + "angle": 0, + "content": "[1] Hao Ai, Zidong Cao, Jin Zhu, Haotian Bai, Yucheng Chen, and Ling Wang. Deep learning for omnidirectional vision: A survey and new perspectives. ArXiv, abs/2205.10468, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.173, + 0.472, + 0.255 + ], + "angle": 0, + "content": "[2] Hao Ai, Zidong Cao, Yan-Pei Cao, Ying Shan, and Lin Wang. Hrdfuse: Monocular \\(360^{\\circ}\\) depth estimation by collaboratively learning holistic-with-regional depth distributions. 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13273-13282, 2023. 2, 3, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.258, + 0.471, + 0.3 + ], + "angle": 0, + "content": "[3] Iro Armeni, Sasha Sax, Amir Roshan Zamir, and Silvio Savarese. Joint 2d-3d-semantic data for indoor scene understanding. CoRR, abs/1702.01105, 2017. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.302, + 0.47, + 0.37 + ], + "angle": 0, + "content": "[4] Zidong Cao, Hao Ai, Yan Cao, Ying Shan, Xiaohu Qie, and Lin Wang. Omnizoomer: Learning to move and zoom in on sphere at high-resolution. 2023 IEEE/CVF International Conference on Computer Vision (ICCV), pages 12851-12861, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.373, + 0.47, + 0.442 + ], + "angle": 0, + "content": "[5] Angel X. Chang, Angela Dai, Thomas A. Funkhouser, Maciej Halber, Matthias Nießner, Manolis Savva, Shuran Song, Andy Zeng, and Yinda Zhang. Matterport3d: Learning from RGB-D data in indoor environments. In 3DV, pages 667-676. IEEE Computer Society, 2017. 1, 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.445, + 0.47, + 0.514 + ], + "angle": 0, + "content": "[6] Zehui Chen, Zhenyu Li, Shiquan Zhang, Liangji Fang, Qinghong Jiang, Feng Zhao, Bolei Zhou, and Hang Zhao. Autoalign: Pixel-instance feature aggregation for multimodal 3d object detection. In International Joint Conference on Artificial Intelligence, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.516, + 0.47, + 0.585 + ], + "angle": 0, + "content": "[7] Hsien-Tzu Cheng, Chun-Hung Chao, Jin-Dong Dong, Hao-Kai Wen, Tyng-Luh Liu, and Min Sun. Cube padding for weakly-supervised saliency prediction in \\(360^{\\circ}\\) videos. 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1420-1429, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.588, + 0.47, + 0.643 + ], + "angle": 0, + "content": "[8] Xinjing Cheng, Peng Wang, Yanqi Zhou, Chenye Guan, and Ruigang Yang. Omnidirectional depth extension networks. In 2020 IEEE International Conference on Robotics and Automation (ICRA), pages 589-595, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.645, + 0.47, + 0.714 + ], + "angle": 0, + "content": "[9] Yanhua Cheng, Rui Cai, Zhiwei Li, Xin Zhao, and Kaiqi Huang. Locality-sensitive deconvolution networks with gated fusion for rgb-d indoor semantic segmentation. 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 1475-1483, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.717, + 0.47, + 0.771 + ], + "angle": 0, + "content": "[10] Taco Cohen, Maurice Weiler, Berkay Kicanaoglu, and Max Welling. Gauge equivariant convolutional networks and the icosahedral cnn. In International Conference on Machine Learning, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.774, + 0.47, + 0.829 + ], + "angle": 0, + "content": "[11] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. computer vision and pattern recognition, 2009. 2, 4, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.832, + 0.47, + 0.901 + ], + "angle": 0, + "content": "[12] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "scale. In International Conference on Learning Representations, 2021. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.121, + 0.892, + 0.176 + ], + "angle": 0, + "content": "[13] Marc Eder, Mykhailo Shvets, John Lim, and Jan-Michael Frahm. Tangent images for mitigating spherical distortion. 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12423-12431, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.178, + 0.892, + 0.232 + ], + "angle": 0, + "content": "[14] Kaiming He, X. Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 770-778, 2015. 2, 4, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.233, + 0.892, + 0.302 + ], + "angle": 0, + "content": "[15] ChiyuMax Jiang, Jingwei Huang, Karthik Kashinath, Prabhat Prabhat, Philip Marcus, and Matthias Niessner. Spherical cnns on unstructured grids. International Conference on Learning Representations, International Conference on Learning Representations, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.303, + 0.892, + 0.357 + ], + "angle": 0, + "content": "[16] Hualie Jiang, Zhe Sheng, Siyu Zhu, Zilong Dong, and Rui Huang. Unifuse: Unidirectional fusion for \\(360^{\\circ}\\) panorama depth estimation. IEEE Robotics and Automation Letters, 6: 1519-1526, 2021. 2, 3, 4, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.359, + 0.892, + 0.413 + ], + "angle": 0, + "content": "[17] Jiayu Jiao, Yu-Ming Tang, Kun-Yu Lin, Yipeng Gao, Jinhua Ma, Yaowei Wang, and Wei-Shi Zheng. Dilateformer: Multi-scale dilated transformer for visual recognition. IEEE Transactions on Multimedia, pages 1-14, 2023. 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.414, + 0.892, + 0.441 + ], + "angle": 0, + "content": "[18] Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. CoRR, abs/1412.6980, 2014. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.442, + 0.892, + 0.523 + ], + "angle": 0, + "content": "[19] Varun Ravi Kumar, Senthil Kumar Yogamani, Hazem Rashed, Ganesh Sitsu, Christian Witt, Isabelle Leang, Stefan Milz, and Patrick Mäder. Omnidet: Surround view cameras based multi-task visual perception network for autonomous driving. IEEE Robotics and Automation Letters, 6:2830-2837, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.525, + 0.892, + 0.579 + ], + "angle": 0, + "content": "[20] Iro Laina, C. Rupprecht, Vasileios Belagiannis, Federico Tombari, and Nassir Navab. Deeper depth prediction with fully convolutional residual networks. 3DV 2016, pages 239-248, 2016. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.581, + 0.892, + 0.65 + ], + "angle": 0, + "content": "[21] Yeonkun Lee, Jaeseok Jeong, Jong Seob Yun, Wonjune Cho, and Kuk jin Yoon. Spherephd: Applying cnns on a spherical polyhedron representation of \\(360^{\\circ}\\) images. 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9173-9181, 2018. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.651, + 0.892, + 0.705 + ], + "angle": 0, + "content": "[22] Meng Li, Senbo Wang, Weihao Yuan, Weichao Shen, Zhe Sheng, and Zilong Dong. \\(S^2\\) net: Accurate panorama depth estimation on spherical surface. IEEE Robotics and Automation Letters, 8:1053-1060, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.706, + 0.892, + 0.774 + ], + "angle": 0, + "content": "[23] Yunhao Li, Wei Shen, Zhongpai Gao, Yucheng Zhu, Guangtao Zhai, and Guodong Guo. Looking here or there? gaze following in 360-degree images. 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 3722-3731, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.776, + 0.892, + 0.845 + ], + "angle": 0, + "content": "[24] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 9992-10002, 2021. 2, 3, 4, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.846, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[25] Yunze Man, Liangyan Gui, and Yu-Xiong Wang. Bev-guided multi-modality fusion for driving perception. 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 21960-21969, 2023. 3" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "9934" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.161 + ], + "angle": 0, + "content": "[26] Daniel Martin, Ana Serrano, Alexander W. Bergman, Gordon Wetzstein, and Belén Masiá. Scangan360: A generative model of realistic scanpaths for \\(360^{\\circ}\\) images. IEEE Transactions on Visualization and Computer Graphics, 28:2003-2013, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.162, + 0.471, + 0.218 + ], + "angle": 0, + "content": "[27] Rafael Monroy, Sebastian Lutz, Tejo Chalasani, and Aljoscha Smolic. Salnet360: Saliency maps for omnidirectional images with cnn. Signal Process. Image Commun., 69:26-34, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.218, + 0.471, + 0.288 + ], + "angle": 0, + "content": "[28] Giovanni Pintore, Eva Almansa, and Jens Schneider. Slicenet: deep dense depth estimation from a single indoor panorama using a slice-based representation. 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11531-11540, 2021. 1, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.288, + 0.471, + 0.344 + ], + "angle": 0, + "content": "[29] Manuel Rey-Area, Mingze Yuan, and Christian Richardt. 360monodepth: High-resolution \\(360^{\\circ}\\) monocular depth estimation. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3752-3762, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.344, + 0.471, + 0.386 + ], + "angle": 0, + "content": "[30] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In MICCAI (3), pages 234-241. Springer, 2015. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.386, + 0.471, + 0.442 + ], + "angle": 0, + "content": "[31] Mehran Shakerinava and Siamak Ravanbakhsh. Equivariant networks for pixelized spheres. Proceedings of the 38th International Conference on Machine Learning, ICML, abs/2106.06662, 2021. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.442, + 0.471, + 0.497 + ], + "angle": 0, + "content": "[32] Zhijie Shen, Chunyu Lin, Kang Liao, Lang Nie, Zishuo Zheng, and Yao Zhao. Panofformer: Panorama transformer for indoor \\(360^{\\circ}\\) depth estimation. In European Conference on Computer Vision, 2022. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.497, + 0.471, + 0.553 + ], + "angle": 0, + "content": "[33] Yu-Chuan Su and Kristen Grauman. Kernel transformer networks for compact spherical convolution. 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9434-9443, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.553, + 0.471, + 0.609 + ], + "angle": 0, + "content": "[34] Cheng Sun, Min Sun, and Hwann-Tzong Chen. Hohonet: 360 indoor holistic understanding with latent horizontal features. 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2573-2582, 2020. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.608, + 0.471, + 0.651 + ], + "angle": 0, + "content": "[35] Mingxing Tan and Quoc V. Le. Efficientnet: Rethinking model scaling for convolutional neural networks. In ICML, pages 6105-6114. PMLR, 2019. 4, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.651, + 0.471, + 0.706 + ], + "angle": 0, + "content": "[36] Keisuke Tateno, Nassir Navab, and Federico Tombari. Distortion-aware convolutional filters for dense prediction in panoramic images. In European Conference on Computer Vision, 2018. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.706, + 0.471, + 0.775 + ], + "angle": 0, + "content": "[37] Fu-En Wang, Yu-Hsuan Yeh, Min Sun, Wei-Chen Chiu, and Yi-Hsuan Tsai. Bifuse: Monocular 360 depth estimation via bi-projection fusion. 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 459–468, 2020. 2, 3, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.776, + 0.471, + 0.845 + ], + "angle": 0, + "content": "[38] Fu-En Wang, Yu-Hsuan Yeh, Yi-Hsuan Tsai, Wei-Chen Chiu, and Min Sun. Bifuse++: Self-supervised and efficient bi-projection fusion for \\(360^{\\circ}\\) depth estimation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45: 5448-5460, 2022. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.845, + 0.471, + 0.903 + ], + "angle": 0, + "content": "[39] Sirui Wang, Di Liang, Jian Song, Yuntao Li, and Wei Wu. DABERT: dual attention enhanced BERT for semantic matching. In COLING, pages 1645-1654. International Committee on Computational Linguistics, 2022. 6" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.471, + 0.903 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.092, + 0.893, + 0.149 + ], + "angle": 0, + "content": "[40] X. Wang, Ross B. Girshick, Abhinav Kumar Gupta, and Kaiming He. Non-local neural networks. 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7794-7803, 2017. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.149, + 0.893, + 0.219 + ], + "angle": 0, + "content": "[41] Yu yang Li, Yuliang Guo, Zhixin Yan, Xinyu Huang, Ye Duan, and Liu Ren. Omnifusion: 360 monocular depth estimation via geometry-aware fusion. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2791-2800, 2022. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.219, + 0.893, + 0.303 + ], + "angle": 0, + "content": "[42] Xiaoqing Ye, Mao Shu, Hanyu Li, Yifeng Shi, Yingying Li, Guan-Sheng Wang, Xiao Tan, and Errui Ding. Rope3d: The roadside perception dataset for autonomous driving and monocular 3d object detection task. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 21309-21318, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.303, + 0.893, + 0.387 + ], + "angle": 0, + "content": "[43] Youngho Yoon, Inchul Chung, Lin Wang, and Kuk-Jin Yoon. Spheresr: \\(360^{\\circ}\\) image super-resolution with arbitrary projection via continuous spherical image representation. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5667-5676, 2021. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.387, + 0.893, + 0.458 + ], + "angle": 0, + "content": "[44] Haozheng Yu, Lu He, Bing Jian, Weiwei Feng, and Shanghua Liu. Panelnet: Understanding 360 indoor environment via panel representation. 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 878-887, 2023. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.458, + 0.893, + 0.514 + ], + "angle": 0, + "content": "[45] Ilwi Yun, Chan-Yong Shin, Hyunku Lee, Hyuk-Jae Lee, and Chae-Eun Rhee. Egformer: Equirectangular geometry-biased transformer for 360 depth estimation. ArXiv, abs/2304.07803, 2023. 2, 3, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.514, + 0.893, + 0.572 + ], + "angle": 0, + "content": "[46] Chao Zhang, Stephan Liwicki, William Smith, and Roberto Cipolla. Orientation-aware semantic segmentation on icosahedron spheres. 2019 IEEE/CVF International Conference on Computer Vision (ICCV), pages 3532-3540, 2019. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.572, + 0.893, + 0.627 + ], + "angle": 0, + "content": "[47] Hengshuang Zhao, Li Jiang, Jiaya Jia, Philip H. S. Torr, and Vladlen Koltun. Point transformer. 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 16239-16248, 2020. 4, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.627, + 0.893, + 0.685 + ], + "angle": 0, + "content": "[48] Jia Zheng, Junfei Zhang, Jing Li, Rui Tang, Shenghua Gao, and Zihan Zhou. Structured3d: A large photo-realistic dataset for structured 3d modeling. In European Conference on Computer Vision, 2019. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.684, + 0.893, + 0.768 + ], + "angle": 0, + "content": "[49] Xu Zheng, Jinjing Zhu, Ye-Peng Liu, Zidong Cao, Chong Fu, and Lin Wang. Both style and distortion matter: Dualpath unsupervised domain adaptation for panoramic semantic segmentation. 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1285-1295, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.768, + 0.893, + 0.825 + ], + "angle": 0, + "content": "[50] Chuanqing Zhuang, Zhengda Lu, Yiqun Wang, Jun Xiao, and Ying Wang. Acdnet: Adaptively combined dilated convolution for monocular panorama depth estimation. CoRR, abs/2112.14440, 2021. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.825, + 0.893, + 0.882 + ], + "angle": 0, + "content": "[51] Nikolaos Zioulis, Antonis Karakottas, Dimitrios Zarpalas, and Petros Daras. Omndepth: Dense depth estimation for indoors spherical panoramicas. In ECCV (6), pages 453-471. Springer, 2018. 1, 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.092, + 0.893, + 0.882 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9935" + } + ] +] \ No newline at end of file diff --git a/2024/Elite360D_ Towards Efficient 360 Depth Estimation via Semantic- and Distance-Aware Bi-Projection Fusion/b933bca2-9ff4-42fc-9bd7-be64c39e1936_origin.pdf b/2024/Elite360D_ Towards Efficient 360 Depth Estimation via Semantic- and Distance-Aware Bi-Projection Fusion/b933bca2-9ff4-42fc-9bd7-be64c39e1936_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..63bdba44b4d9eea15eb1d735ff657cede2f5e2ea --- /dev/null +++ b/2024/Elite360D_ Towards Efficient 360 Depth Estimation via Semantic- and Distance-Aware Bi-Projection Fusion/b933bca2-9ff4-42fc-9bd7-be64c39e1936_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:808f4bc2d63f9237bd5b6b469e1c9944a0be40b78bed8b74b027e1508632dcc1 +size 9369982 diff --git a/2024/Elite360D_ Towards Efficient 360 Depth Estimation via Semantic- and Distance-Aware Bi-Projection Fusion/full.md b/2024/Elite360D_ Towards Efficient 360 Depth Estimation via Semantic- and Distance-Aware Bi-Projection Fusion/full.md new file mode 100644 index 0000000000000000000000000000000000000000..3be055a35968a4e99fc77f6baee4a7deb94d5bca --- /dev/null +++ b/2024/Elite360D_ Towards Efficient 360 Depth Estimation via Semantic- and Distance-Aware Bi-Projection Fusion/full.md @@ -0,0 +1,314 @@ +# Elite360D: Towards Efficient 360 Depth Estimation via Semantic- and Distance-Aware Bi-Projection Fusion + +Hao Ai $^{1}$ Lin Wang $^{1,2*}$ + +$^{1}$ AI Thrust, HKUST(GZ) $^{2}$ Dept. of CSE, HKUST + +hai033@connect.hkust-gz.edu.cn, linwang@ust.hk + +# Abstract + +360 depth estimation has recently received great attention for 3D reconstruction owing to its omnidirectional field of view (FoV). Recent approaches are predominantly focused on cross-projection fusion with geometry-based reprojection: they fuse 360 images with equirectangular projection (ERP) and another projection type, e.g., cubemap projection to estimate depth with the ERP format. However, these methods suffer from 1) limited local receptive fields, making it hardly possible to capture large FoV scenes, and 2) prohibitive computational cost, caused by the complex cross-projection fusion module design. In this paper, we propose Elite360D, a novel framework that inputs the ERP image and icosahedron projection (ICOSAP) point set, which is undistorted and spatially continuous. Elite360D is superior in its capacity in learning a representation from a local-with-global perspective. With a flexible ERP image encoder, it includes an ICOSAP point encoder, and a Bi-projection Bi-attention Fusion (B2F) module (totally $\sim 1M$ parameters). Specifically, the ERP image encoder can take various perspective image-trained backbones (e.g., ResNet, Transformer) to extract local features. The point encoder extracts the global features from the ICOSAP. Then, the B2F module captures the semantic- and distance-aware dependencies between each pixel of the ERP feature and the entire ICOSAP feature set. Without specific backbone design and obvious computational cost increase, Elite360D outperforms the prior arts on several benchmark datasets. + +# Multimedia Material + +For videos, code, demo and more information, you can visit https://VLIS2022.github.io/Elite360D/ + +# 1. Introduction + +$360^{\circ}$ images capture the complete surrounding environment in one shot with an ultra-wide field-of-view (FoV) of $180^{\circ} \times 360^{\circ}$ , which is broadly applied to applications, e.g., autonomous driving [1, 19, 42, 43, 49] and virtual reality + +![](images/4b8c1128a92aaa7822a410fa93cb01b5cfaaf43b3b1e06e4afc0f851443c59e5.jpg) +Model #FLOPs(G) + +![](images/6ccd0c6784e3ed90697fd7c4da8aba8d21897d4ef0eaceaa2b3eb9f282f127f6.jpg) + +![](images/df0039b601e777013f7ebd4a22ecf72e05ba18dd39bdc61f392415cb7e9bdbdb.jpg) +ERP Image + +![](images/bb551f9bef0ee005a962d3966ac4eed6aac8138259cd520f9cf94b3ede5d6865.jpg) +(a) +GT Depth + +![](images/bfd43e37b788093d55dfe53c89684b1fd3d1f5de07820cf265f93fcbb2556269.jpg) +Equi-Baseline +Figure 1. (a) Performance (RMSE error) curves on M3D test dataset [5]. Larger circles mean more parameters (e.g., ResNet18, ResNet34, ResNet50) and lower errors mean better performance. (b) Comparison with the ResNet34 as the ERP encoder backbone. With only 1M more parameters, our depth result is more accurate. + +![](images/698131984424d9e7b1c0bd995bd3c5a496ed09ed2cd3b34c8dee377974d61f94.jpg) +Our Elite360D +(b) + +[4, 23, 26]. The ability to infer the 3D structure of a surrounding scene has inspired active research for monocular 360 depth estimation. Equirectangular projection (ERP) is the commonly used projection type that can provide a complete view of a scene. An ERP image, a.k.a., panorama, samples pixels with a higher density at the poles compared to the equator, resulting in spherical distortions (Fig. 2(c)). + +In recent years, to address the spherical distortions, several ah-hoc designs [33, 36, 50, 51] focus on the distortion-aware convolutional kernels and sample grids from undistorted tangent planes on the sphere. These grids are backprojected to the corresponding locations of the ERP image. Also, some methods [28, 34, 44] partition the ERP + +input into multiple vertical slices and extract a compact representation from these slices to extend the receptive field of each pixel for the ERP representation. Unfortunately, these methods result in substantial computational costs. Meanwhile, the limited local receptive field of the convolutional filters is insufficient to provide the global perception of the panorama. Therefore, PanoFormer [32] and EGFormer [45] design specific transformer architectures to effectively model the long-range dependencies between the global context of large-FoV ERP images. Nevertheless, the receptive fields are confined by the sizes of local attention windows. Furthermore, the aforementioned data-driven methods often yield sub-optimal results when trained from scratch on limited $360^{\circ}$ depth dataset (See Tab. 2). + +With the power of existing large-scale perspective images [11], several methods [2, 16, 37, 38] employ the pretrained models - designed for perspective images - as encoder backbones to extract ERP feature maps, and propose the cross-projection fusion to rectify distortions in the ERP feature maps. Specifically, these methods introduce the less-distorted planar projection data, i.e., cubemap projection (CP) patches or tangent projection (TP) patches (See Fig. 2(d), (e)), and unify the spatial dimensions between different projections for the cross-projection fusion. BiFuse and UniFuse [16, 37, 38] propose the C2E module to reproject the content in patch-wise CP features into ERP grid following the spherical geometric relationships. Furthermore, they leverage the fusion between ERP feature maps and C2E feature maps to enhance depth estimation accuracy. However, their concatenation-based feature fusion across multiple scales brings a significant computational burden. Based on the feature similarity, HRDFuse [2] spatially aligns the ERP pixel features and TP patch features. This makes it possible to significantly marry the global context in the ERP feature map and regional structural details in the TP patch features. However, the decoder-level fusion remarkably increases the computational memory and cost. + +These less-distorted gnomonic projections, i.e., CP and TP patches, are spatially discontinuous and require complex re-projection operations for the desired ERP format predictions. Inspired by these issues, we introduce a more powerful projection, the icosahedron projection (ICOSAP) [21, 46], see Fig. 2(f). Importantly, ICOSAP is a spatially continuous and globally perceptive non-Euclidean projection for $360^{\circ}$ images. In light of this, we propose an efficient and effective fusion-based framework, named Elite360D that takes the best of ERP and ICOSAP by learning a representation from a local-with-global perspective. It comprises three components: an ERP encoder, an ICOSAP encoder, and the bi-projection bi-attention fusion (B2F) module. For the ERP image encoder, our Elite360D flexibly supports a wide range of off-the-shelf 2D models, e.g., ResNet [14], Swin transformer [24], as the backbone to + +![](images/559fd2dd5357f51f53b367b8c3c250f69194382d4865fadbb1f82775ba3e7437.jpg) +(a) Panorama + +![](images/517fb55fd3973852e3fd942b48bd7c0440f89f20555a0d131e6bcd5e61239cdd.jpg) +(b) ERP + +![](images/1c8897cd52e20ba6a64eeecdbe62c477c8f42d61c95b12daacf070ab3f16ac4d.jpg) +(c) ERP (flatten) + +![](images/8683cec09fd50640b30f1a71c3147380a3f1f434af1c3e190bcc7d0e79d4b8ed.jpg) +(d) CP + +![](images/adb587f419b2ea73a279429ddeaab0947ee794a867d4317b62e7cd5fd837f3b6.jpg) +(e) TP + +![](images/ed904e61df0c4ff3dd2dfdcd9325d330125cf773612ec50c478beac5f4aa4ec8.jpg) +(f) ICOSAP (level=1) +Figure 2. Different projections of a spherical imaging panorama. + +extract ERP feature maps. This potentially reduces the overfitting problems, particularly on small-scale 360 depth estimation datasets [3, 5]. For efficiency, we represent the ICOSAP spheres as the discrete point sets rather than icosahedron meshes [10, 15, 46] or unfolded representations [21, 43]. As such, we can avoid semantic information redundancy due to dense ERP pixels and maintain the spatial position information. Then, to enable each ERP pixel feature with local receptive fields to perceive the whole scene, B2F module captures the semantic- and distance-aware dependencies between each ERP pixel feature and entire ICOSAP feature set. + +We conduct extensive experiments on three different datasets with different encoder backbones to demonstrate the flexibility and effectiveness of our Elite360D. The experimental results indicate that our Elite360D significantly improves plain-backbones' performance with minimal computational memory (only about 1M parameters) (See Fig. 1(a)). Note that, based on the simple ERP depth estimation baseline, Elite360D only performs bi-projection feature fusion at the last feature layer and achieves results that are on par with leading methods, such as [2, 41]. + +In summary, our main contributions are three-fold: (I) We introduce the ICOSAP with spatial continuity and global perception and represent them as the discrete point sets to reduce the computation cost and maintain the spatial information; (II) We propose the B2F module that jointly models the semantic and spatial dependencies between ICOSAP and ERP features and learn the representations with global-with-local receptive fields for large-FoV scenes; (III) Building upon the B2F module, we propose a novel framework that supports diverse off-the-shelf models as encoder backbones, showing better flexibility than the specially designed models, e.g., [32, 45]. + +# 2. Related Works + +Monocular 360 depth estimation Existing methods can be categorized into two types: one with single projection input, and the other with bi-projection inputs. + +1) Single Projection Input With the ERP images as the inputs, many methods focus on mitigating the inherent spherical distortion issues. Several methods [8, 36, 51] rectify distortions by modifying the receptive fields of tradi + +![](images/4b8c68120290d19dcc009978e2f56cedab404557aec0de5506e1feb4ff66701a.jpg) +Figure 3. (a) An overview of our Elite360D framework, comprising image-based ERP feature extraction (Sec. 3.1), point-based ICOSAP feature extraction (Sec. 3.2), and Bi-projection Bi-attention fusion (B2F) (Sec. 3.3). For better visualization, we do not show the skip connections [30] at the decoding stage. (b) Illustration of the B2F module, consisting of three parts: semantic-aware affinity attention block (Fig. 5), distance-aware affinity attention block (Fig. 6) and gated fusion (Eq. 5). + +tional convolutional kernels. Besides, OmniFusion [41] and 360MonoDepth [29] follow [13] to predict depth maps from less distorted TP patches. Similarly, PanelNet [44] predicts depth maps on the vertical slices of ERP images and then aggregates them. Moreover, some methods [28, 34, 50] focus on tackling the small receptive fields of convolution filters for processing large FoV scenes. More recently, PanoFormer [32] and EGFormer [45] design transformer architectures [12] for ERP images to model the long-range dependencies. By contrast, $\mathbf{S}^2\mathrm{Net}$ [22] extracts robust features from ERP images with the Swin transformer [24] and projects the ERP feature maps onto the spherical surface to minimize distortions and maintain spatial consistency. + +2) Bi-Projection Inputs. BiFuse [37] utilizes a dense fusion strategy to bidirectionally fuse ERP and CP features at both encoding and decoding stages. In contrast, BiFuse++[38] and UniFuse [16] fuse ERP and CP features at the encoding stage. Recently, HRDFuse [2] achieved the SOTA performance based on the adaptive fusion of ERP and TP predictions. By contrast, we introduce a non-Euclidean projection, ICOSAP, which is less distorted and spatially continuous. We leverage the global perception capacity of ICOSAP to enable each ERP pixel-wise feature, with a limited local receptive field, to capture the entire scene. + +Representations for 360 Images ERP is the most commonly used projection, which projects a panorama onto a cylinder (Fig. 2(b)) and then unfolds it into a plane (Fig. 2(c)). However, ERP maps latitudes and longitudes onto the vertical and horizontal axes with equal spacing on the plane, leading to distortions. CP [27] projects the panoramas onto six faces of a cube where each face is less distorted (Fig. 2(d)). Building upon CP, some padding methods [7, 37] remove the boundary inconsistency between the adjacent CP patches. Recently, Eder et al. [13] proposed the tangent projection (TP) (Fig. 2(e)), to signif + +![](images/1da8f6d0fed53242d5b8ed7279e74011204c8d01b3a5c7e818bfa7bc7c833b0c.jpg) +(a) $l = 0$ + +![](images/f3406636289f7ea1271712b536fd94cec36ba985895a4f4c2edc4b318f99eab3.jpg) +(b) $l = 1$ + +![](images/ec00a3b3e4a7e4e596147cdb250833a5ee0e0ab2b8e160f70d55ca63400e035b.jpg) +(c) $l = 2$ + +![](images/e0b40d58d35bd335423683da34282925bc08dc2157f1abffe2b3a017281bf8db.jpg) +(d) $l = 3$ + +Figure 4. The subdivision of icosahedron at different resolution $l$ . icantly mitigate spherical distortion. Besides, [21, 31, 43] employ the polyhedrons, e.g., octahedron and icosahedron, to represent the panoramas and unfold these polyhedron representations on the plane for processing. In Elite360D, we select popular ICOSAP (See Fig. 4) as our input. Notably, we represent ICOSAP as discrete points, significantly reducing the computational cost while preserving spatial information and global awareness. Meanwhile, with the dense pixels in the ERP panoramas, discrete ICOSAP points can avoid semantic information redundancy. + +Cross-Attention Mechanism It is widely used for efficient multi-modal feature fusion. For instance, Chen et al. [6] built cross-attention blocks to align and fuse 2D image features and 3D point cloud features for 3D object recognition. Similarly, BEVGuide [25], using BEV embedding as a guided query, employs a cross-attention block to fuse information across different sensors. We propose the B2F module to model the relationships between each pixel of the ERP feature and the whole ICOSAP point feature set, based on the semantic-aware affinity attention and distance-aware affinity attention simultaneously. With the B2F module, each ERP pixel feature can perceive the spatial information and semantic information of the large FoV scenes. + +# 3. The Proposed Framework + +Overview. Our goal is to investigate an efficient and effective bi-projection fusion framework for 360 depth estimation. To this end, as illustrated in Fig. 3, the proposed + +Elite360D consists of three key components: image-based ERP feature extraction, point-based ICOSAP feature extraction and bi-projection bi-attention fusion (B2F). For the ERP feature $F^{E}$ extraction, Elite360D flexibly accommodates a wide range of 2D models as encoder backbones and takes advantage of pre-training to reduce the over-fitting problems. For the ICOSAP feature extraction, we represent ICOSAP sphere as the point set, which significantly reduces the computation cost and preserves the spatial information and global perception. Then we directly employ the simple point encoder, i.e. Point transformer [47], to extract the ICOSAP point feature set $\{F_{n}^{I}\}_{n = 1,\dots ,N}$ , $N$ is the point number, denoted as $\{F_n^I\}_N$ . Next, to facilitate bi-projection feature fusion and learn a powerful representation $F^{GL}$ from a local-with-global perspective, B2F module comprises semantic-aware affinity attention block, distance-aware affinity attention block, and gated fusion block. These two attention blocks model the semantic- and distance-aware dependencies between each ERP pixel-wise feature and entire ICOSAP feature set, respectively and the gated fusion block adaptively fuses the bi-attention outputs. Lastly, we employ a simple decoder with up-sampling and skip-connections to predict the final depth map from the fused representation $F^{GL}$ . We now describe the details. + +# 3.1. ERP Feature Extraction + +Taking an ERP image with the resolution size of $H \times W$ as the input, the encoder extracts feature map $F^{E} \in \mathbb{R}^{h \times w \times C}$ , where $h = H / s$ , $w = W / s$ , $s$ is the down-sampling scale factor and $C$ is the channel number. In particular, as treating ERP images as 2D perspective images, our encoder backbone is compatible with a wide range of robust 2D models that have been pre-trained on large-scale perspective image datasets [11], including CNNs and vision transformers, e.g., ResNet [14], EfficientNet [35], Swin transformer [24]. Notably, as repeated local operations cannot substitute for a global operator [40, 45], the finite size of convolutional kernels or the limited size of local attention windows results in ERP pixel-wise features lacking sufficient global receptive fields. Additionally, these 2D model performance suffers from the spherical distortions (See Tab. 1). + +# 3.2. ICOSAP Feature Extraction + +Preliminary. Since Lee et al. [21] found that polyhedron representations with a greater number of initial faces exhibit lower distortion, this paper focuses on the ICOSAP with the most initial faces. Especially, we introduce the ICOSAP to offer a comprehensive and high-quality global perception. For the resolution, ICOSAP data consists of $20 \times 4^l$ faces and $12 \times 4^l$ vertices at the subdivision level $l$ (See Fig. 4). Given the spatial relationships between ICOSAP and ERP within the spherical space, RGB values of ICOSAP vertices can be derived from corresponding ERP pixels. + +Feature extraction. Existing works focus on processing + +![](images/7aa4770334ef310092503896260a80e21b73aac7e5baf5998c60ad04f2aeee9d.jpg) +Figure 5. The architecture of semantic-aware affinity attention. Especially, $\mathbf{Q}_{i,j}^{S}\in \mathbb{R}^{1\times d}$ , $\mathbf{K}^S\in \mathbb{R}^{N\times d}$ , and $\mathbf{V}^S\in \mathbb{R}^{N\times d}$ , where $d$ is the dimension and $N$ is the ICOSAP point number. + +ICOSAP data on the plane, i.e., unfolded mesh representation [46] and spherical polyhedron representations [21]. However, these methods require specially designed and computationally intensive operations, such as convolutions, pooling [21], and up-sampling [31], to process the planar representations. By contrast, we propose to represent an ICOSAP input as a point set, as illustrated at the bottom of Fig. 3(a). ERP images provide dense pixel values and discrete ICOSAP point set can prevent redundancy in semantic information while preserving spatial information and global perception. Therefore, we take full advantage of two projections for efficient and accurate 360 depth estimation. In detail, we employ the central points of each face to represent the ICOSAP sphere. Firstly, we obtain the $20 \times 4^l$ faces of an ICOSAP sphere at the default subdivision level $l$ , where each face is composed of three vertices. Then, we calculate the spatial coordinates and RGB values of each face center by averaging them of each three vertices, as each face is an equilateral triangle. As a result, we obtain the ICOSCAP point set $\{P^{ICO}\} \in \mathbb{R}^{(20 \times 4^l) \times 6}$ , where $20 \times 4^l$ is the point number and 6 represents the coordinates $[x, y, z]$ and RGB channels. With the input point set $\{P^{ICO}\}$ , we directly employ the encoder of Point Transformer [47] to extract ICOSAP point feature set $\{F_n^I\}_N \in \mathbb{R}^{N \times C}$ , where $N$ represents the number of point features and $C$ corresponds to the same channel number as the ERP feature map $F^E$ . Especially, for simplicity and efficiency, we do not opt for complex networks for this purpose. + +# 3.3. Bi-Projection Bi-Attention Fusion + +In earlier methods [37] and [16], bi-projection feature fusion primarily depends on the geometric relationships between CP and ERP. They first apply the C2E function to re-project CP feature patches into ERP format feature maps and subsequently perform pixel-wise feature concatenation. While this fusion is effective, it brings several problems: 1) geometry-based re-projection and concatenation-based feature fusion significantly increase the computational costs (See Tab. 2); 2) the geometry-based fusion process from CP patches to ERP panorama restricts ERP pixels from the global perception, where each ERP pixel perceives scene in- + +
DatasetsBackboneMethod#Params (M)#FLOPs (G)Abs Rel ↓Sq Rel ↓RMSE ↓δ1(%) ↑δ2(%) ↑δ3(%) ↑
M3D [5]ResNet-18 [14]Equi-Baseline14.3939.840.14280.12030.560780.9794.6898.26
Ours15.4345.910.12720.10700.527085.2895.2898.49
Δ+1.04+6.07-11.06%-9.73%-6.01%+4.31+0.60+0.23
ResNet-34 [14]Equi-Baseline24.5059.270.12550.10480.517385.5296.1098.55
Ours25.5465.290.11150.09140.487588.1596.4698.74
Δ+1.04+6.02-11.16%-12.79%-5.76%+2.63+0.36+0.19
ResNet-50* [14]Equi-Baseline29.6375.130.13700.12090.543282.5194.8398.07
Ours42.99170.110.11120.09800.487086.7096.0198.61
Δ+13.36+94.98-18.75%-18.94%-10.35%+4.19+1.18+0.54
EfficientNet-B5 [35]Equi-Baseline33.1018.810.10340.08310.463889.9296.8098.80
Ours34.1124.880.10480.08050.452489.9297.0799.14
Δ+1.01+6.07+1.35%-3.13%-2.46%-0.00+0.27+0.34
SWin-B* [24]Equi-Baseline90.75187.370.13010.12500.555784.0694.7697.72
Ours94.30211.280.12490.12100.546285.3495.0797.77
Δ+3.55+23.91-4.00%-3.20%-1.71%+1.28+0.31+0.05
DilateFormer-T [17]Equi-Baseline19.9346.960.14290.12740.574881.1594.4697.79
Ours20.9453.020.14230.12510.551782.6895.0098.14
Δ+1.01+6.06-0.42%-1.81%-4.02%+1.53+0.54+0.35
S2D3D [3]ResNet-34 [14]Equi-Baseline24.5059.220.12030.07540.372487.4196.5198.77
Ours25.5165.280.11820.07280.375688.7296.8498.92
Δ+1.01+6.06-9.21%-3.45%+0.86%+1.31+0.33+0.15
EfficientNet-B5 [35]Equi-Baseline33.1018.810.10260.06380.358089.4397.0699.16
Ours34.1124.880.10180.06030.357589.4797.2399.22
Δ+1.01+6.07-0.78%-5.49%-0.14%+0.03+0.17+0.06
Struct3D [48]ResNet-34 [14]Equi-Baseline24.5059.220.22560.39100.764176.9089.9694.49
Ours25.5165.280.14800.22150.496187.4194.3496.66
Δ+1.01+6.06-34.40%-43.35%-35.07%+10.51+4.38+2.17
EfficientNet-B5 [35]Equi-Baseline33.1018.810.13120.19380.431288.5395.2197.40
Ours34.1124.880.12770.19300.415189.1695.3397.43
Δ+1.01+6.07-2.67%-0.41%-3.73%+0.63+0.12+0.03
+ +Table 1. Quantitative comparison with ERP-based depth baseline. For most conditions, we set the channel number $C$ to 64. Especially, we set $C$ to 256 for ResNet-50* and to 128 for Swin-B*. Bold indicates performance improvement. Red indicates performance decline. + +formation only from its corresponding small-FoV CP patch without considering other patches; 3) this one-to-one alignment pattern only emphasizes spatial consistency, with no consideration for the semantic similarity. Thus, we design the Bi-Projection Bi-Attention Fusion (B2F) module to solve the above problems. As shown in Fig. 3(b), B2F module first leverages semantic-aware affinity attention and distance-aware affinity attention to model the semantic and spatial dependencies between each ERP pixel-wise feature $F_{i,j}^{E}$ and ICOSAP point feature set $\{F_n^I\}_N$ . Here $(i,j)$ indicates the coordinate of a pixel in the ERP feature map, $i \in (1,h), j \in (1,w)$ , and $N$ is the point number of the ICOSAP feature set. Consequently, a gated fusion block is employed to adaptively marry semantic- and distance-related information. Semantic-aware affinity attention. The semantic-aware affinity attention, as shown in Fig. 5, follows the standard dot-product attention [12]. In detail, given the extracted ERP feature map $F^{E} \in \mathbb{R}^{h \times w \times C}$ and ICOSAP point feature set $\{F_n^I\}_N \in \mathbb{R}^{N \times C}$ (for simplicity, we denote it as $[F^I]$ ), we generate the query $\mathbf{Q}_{i,j}^{S}$ from each ERP pixel-wise feature $F_{i,j}^{E} \in \mathbb{R}^{1 \times C}$ and produce the key $\mathbf{K}^{S}$ and value $\mathbf{V}^{S}$ from the whole ICOSAP point feature set $[F^{I}]$ . Then, we calculate the attention weight $\mathbf{A}_{i,j}^{S}$ based on $\mathbf{Q}_{i,j}^{S}$ and $\mathbf{K}^{S}$ , and obtain the affinity feature $F_{i,j}^{SA}$ with $F_{i,j}^{SA} = \mathbf{A}_{i,j}^{S} * \mathbf{V}^{S}, \mathbf{A}_{i,j}^{S} = \text{softmax}\left(\frac{\mathbf{Q}_{i,j}^{S} \mathbf{K}^{ST}}{\sqrt{d}}\right)$ , where $d$ is the channel number and + +![](images/14496925d91d219f5dbb38074c0ad79bdf92be77a9e4c00a00e5b7d077ea33a8.jpg) +Figure 6. The architecture of distance-aware affinity attention, especially, $Q_{i,j}^{D} \in R^{1 \times d}$ , $K^{D} \in \mathbb{R}^{N \times d}$ , and $\mathbf{V}^{D} \in \mathbb{R}^{N \times d}$ . + +$d = C$ . After querying all the ERP pixel-wise features, we can obtain the ERP format feature map $F^{SA}$ with the dimension of $\mathbb{R}^{h\times w\times d}$ . The attention weight $A_{i,j}^{S}$ captures affinities based on the semantic-aware feature similarities, and the output $F^{SA}$ effectively integrates the global and local receptive fields. + +Distance-aware affinity attention. Distance-aware affinity attention captures the differences between ERP and ICOSAP in both spatial and semantic information. This can better utilize the geometric prior knowledge of panoramicas and precisely measure the distances between two different projection features. As depicted in Fig. 6, the distance-aware affinity attention is built upon the subtraction-based + +
DatasetsBackboneMethodPub'Year#Params (M)#FLOPs (G)Abs Rel ↓Sq Rel ↓RMSE ↓δ1(%) ↑δ2(%) ↑δ3(%) ↑
M3D [5]TransformerEGFormer [45]ICCV'2315.3966.210.14730.15170.602581.5893.9097.35
PanoFormer [32]ECCV'2220.3881.090.10510.09660.492989.0896.2398.31
ResNet-18 [14]BiFuse [37]CVPR'2035.80165.660.13600.12020.548883.2795.1298.10
UniFuse [16]RAL'2130.2662.600.11910.10300.515886.0495.8498.30
Omnifusion [41]CVPR'2232.3598.680.12090.10900.505586.5895.8198.36
HRDFuse† [2]CVPR'2326.0950.590.14140.12410.550781.4894.8998.20
Ours-15.4345.910.12720.10700.527085.2895.2898.49
ResNet-34 [14]BiFuse [37]CVPR'2056.01199.580.11260.09920.502788.0096.1398.47
BiFuse++ [38]TPAMI'2252.4987.480.11230.09150.485388.1296.5698.69
UniFuse [16]RAL'2150.4896.520.11440.09360.483587.8596.5998.73
Omnifusion [41]CVPR'2242.46142.290.11610.10070.493187.7296.1598.44
HRDFuse† [2]CVPR'2346.3180.870.11720.09710.502586.7496.1798.49
Ours-25.5465.290.11150.09140.487588.1596.4698.74
ResNet-50* [14]BiFuse [37]CVPR'20253.08775.240.11790.09810.497086.7496.2798.66
UniFuse [16]RAL'21131.30222.300.11850.09840.502486.6696.1898.50
Ours-42.99170.110.11120.09800.487086.7096.0198.61
S2D3D [3]TransformerEGFormer [45]ICCV'2315.3966.210.15280.14080.497481.8593.3897.36
PanoFormer [32]ECCV'2220.3881.090.11220.07860.394588.7495.8498.59
ResNet-34 [14]Omnifusion [41]CVPR'2242.46142.290.11540.07750.380986.7496.0398.71
UniFuse [16]RAL'2150.4896.520.11240.07090.355587.0697.0498.99
Ours-25.5165.280.11820.07280.375688.7296.8498.92
Struct3D [48]TransformerEGFormer [45]ICCV'2315.3966.210.22050.45090.684179.7990.7194.55
PanoFormer [32]ECCV'2220.3881.090.25490.49490.793774.7089.1593.97
ResNet-34 [14]BiFuse [37]CVPR'2056.01199.580.15730.24550.521385.9194.0096.72
UniFuse [16]RAL'2150.4896.520.15060.23190.501685.4293.9996.76
Ours-25.5165.280.14800.22150.496187.4194.3496.66
+ +Table 2. Quantitative comparison with the SOTA methods. ${}^{ \dagger }$ means that we modify the HRDFuse network structure for a fair comparison. Green represents the best performance under the given encoder backbone. + +cross-attention mechanism [39]. Given the ERP pixel-wise feature $F_{i,j}^{E}$ and ICOSAP point feature set $[\pmb{F}^{\pmb{I}}]$ , we firstly calculate the spatial distance embedding $Dis^{SP}$ from the spatial coordinates of ERP pixel and ICOSAP point set, i.e. $[x_{i,j}^{e},y_{i,j}^{e},z_{i,j}^{e}]$ and $\{[x_n^{ico},y_n^{ico},z_n^{ico}]\}_{N}$ , as: + +$$ +D i s _ {i, j} ^ {S P} = \left[ e ^ {- \Delta x _ {i, j}}, e ^ {- \Delta y _ {i, j}}, e ^ {- \Delta z _ {i, j}} \right] \mathbf {W} ^ {S P}, \tag {1} +$$ + +where linear projection $\mathbf{W}^{SP}\in \mathbb{R}^{3\times d}$ $Dis_{i,j}^{SP}\in \mathbb{R}^{1\times N\times d}$ and $[\Delta x_{i,j},\Delta y_{i,j},\Delta z_{i,j}]\in \mathbb{R}^{1\times N\times 3}$ is the distances between $[x_{i,j}^e,y_{i,j}^e,z_{i,j}^e ]$ and $\{[x_n^{ico},y_n^{ico},z_n^{ico}]\}_{N}$ . In particular, the operation $e^{-(\cdot)}$ allows $Dis^{SP}$ to pay more attention to the close parts between ERP pixels and ICOSAP point set. After that, we produce the query $\mathbf{Q}_{i,j}^{D}$ and key $\mathbf{K}^D$ from $F_{i,j}^{E}$ and $[\pmb {F}^{\pmb{I}}]$ , respectively and calculate the semantic distance embedding $Dis_{i,j}^{SE}$ : + +$$ +D i s _ {i, j} ^ {S E} = e ^ {- \left\| \mathbf {Q} _ {i, j} ^ {D} - \mathbf {K} ^ {D} \right\|}, \tag {2} +$$ + +where $\mathbf{W}_Q^D$ , $\mathbf{W}_K^D \in \mathbb{R}^{C \times d}$ are linear projections, $\mathbf{Q}_{i,j}^D \in \mathbb{R}^{1 \times d}$ , $\mathbf{K}^D \in \mathbb{R}^{N \times d}$ , and $Dis_{i,j}^{SE} \in \mathbb{R}^{1 \times N \times d}$ . Lastly, the distance-aware attention weight $\mathbf{A}_{i,j}^D$ is generated with spatial and semantic distance embeddings, and the distance-aware affinity feature vector $F_{i,j}^{DA}$ is obtained from the attention weight $\mathbf{A}_{i,j}^D$ and the value $\mathbf{V}^D$ : + +$$ +\mathbf {A} _ {i, j} ^ {D} = \operatorname {s o f t m a x} \left(\frac {\sum \left(D i s _ {i , j} ^ {S P} + D i s _ {i , j} ^ {S E}\right)}{\sqrt {d}}\right), \tag {3} +$$ + +$$ +\mathbf {V} ^ {D} = F ^ {I} \mathbf {W} _ {V} ^ {D}, F _ {i, j} ^ {D A} = \mathbf {A} _ {i, j} ^ {D} * \mathbf {V} ^ {D}, \tag {4} +$$ + +where $\sum$ means the sum for the channel dimension, i.e. $\sum (Dis_{i,j}^{SP} + Dis_{i,j}^{SE})\in \mathbb{R}^{1\times N}$ . After querying all ERP pixel-wise features, we obtain the distance-aware aggregated feature $F^{DA}\in \mathbb{R}^{h\times w\times d}$ , $d = C$ . + +Gated fusion. Since direct average or concatenation may compromise the original representation ability, inspired by [9], we propose the gated fusion block to adaptively fuse $F^{SA}$ and $F^{DA}$ and obtain the representations $F^{GL}$ from a local-with-global perspective, formulated as: + +$$ +F ^ {G L} = g ^ {S A} * F ^ {S A} + g ^ {D A} * F ^ {D A}, \tag {5} +$$ + +$$ +g ^ {S A} = \sigma_ {S A} \left(W _ {g} ^ {S A} \cdot \left[ F ^ {S A}; F ^ {D A} \right]\right), +$$ + +$$ +g ^ {D A} = \sigma_ {D A} \left(W _ {g} ^ {D A} \cdot \left[ F ^ {S A}; F ^ {D A} \right]\right), +$$ + +where $W_{g}^{SA}$ and $W_{g}^{DA}$ are linear projections, $\sigma_{SA}(\cdot)$ and $\sigma_{DA}(\cdot)$ are the sigmoid functions. + +# 3.4. Optimization + +With the fused feature $F^{GL}$ and multi-scale ERP feature maps in the ERP encoder backbone, we feed them into a decoder [30] with several up-sampling blocks and skipconnections to output the final depth. For the depth supervision, we follow existing works [28, 32] and employ the combination of Berhu loss [20] and gradient loss [28]. (Details of training loss can be found in the suppl. material.). + +# 4. Experiments + +# 4.1. Datasets, Metrics, and Implementation Details + +Datasets and Metrics. We evaluate Elite360D on three datasets: two real-world datasets, Matterport3D(M3D) [5], Stanford2D3D(S2D3D) [3], and a recently proposed large-scale synthetic dataset, Structure3D(Struct3D) [48]. For the evaluation metrics, we follow previous works [16, 32, 45] to employ some standard metrics, including absolute relative error (Abs Rel), squared relative error (Sq Rel), root mean squared error (RMSE), and three threshold percentage $\delta < \alpha^t$ ( $\alpha = 1.25, t = 1, 2, 3$ ), denoted as $\delta_t$ . Addi + +![](images/9853f330a79315c7421f1e6ce11598740b021fe41cd8c5ee4da8eefa63278e18.jpg) +Figure 7. Qualitative results (with ResNet-34 as the backbone) on Matterport3D (top), Stanford2D3D (middle) and Structure3D (bottom). + +
BackbonePre-trainedAbs Rel↓Sq Rel↓RMSE↓δ1↑
ResNet-34X0.15960.14520.585681.36
✓0.11150.09140.487588.15
EfficientNet-B5X0.12110.10870.513186.80
✓0.10480.08050.452489.92
DilateFormerX0.15150.14150.569480.95
✓0.14230.12510.551782.68
+ +Table 3. The impact of pre-training on ERP encoder backbones. + +tionally, we measure the number of parameters and FLOPS to evaluate the efficiency of our method. + +Training Details. We use diverse ERP encoder backbones, including CNNs (ResNet-18, 34, 50 [14], EfficientNet B5 [35]), and transformers (Swin-B [24], DilateFormer-T [17]). All backbones are pre-trained on ImageNet1K [11]. We set the default channel number $C$ to 64 and default subdivision level of ICOSAP as $l = 4$ . For the ICOSAP encoder, we employ the one of Point transformer [47] with three down-sample blocks. Following [16], we use Adam optimizer [18] and a constant learning rate of $1\mathrm{e}^{-4}$ . Considering the unfair comparisons stemming from variations in hyper-parameters and validation procedures used across different methods, we re-train the existing methods from scratch and validate them, following the unified training and validation settings [16]. (Due to page limit, detailed training and validation settings can be found in suppl. mat.). + +# 4.2. Performance Comparison + +Comparisons with ERP-based depth baselines. As shown in Tab. 1, with an increase of only $\sim 1\mathrm{M}$ parameters + +$(C = 64)$ , our Elite360D demonstrates substantial advancements over the ERP-based baselines across different ERP encoder backbones on all three datasets. Specifically, for the Matterport3D dataset, Elite360D achieves reductions exceeding $10\%$ in Abs Rel error (ResNet-18, 34), along with reductions of $4.00\%$ in Abs Rel error (Swin-B) and $4.02\%$ in RMSE error (DilateFormer-T). Besides, with the larger channel number $C = 256$ (ResNet-50), Elite360D outperforms ERP baseline by $18.75\%$ (Abl Rel), $18.94\%$ (Sq Rel). For the small-scale S2D3D dataset, Elite360D outperforms ERP baseline by $9.21\%$ in Abs Rel error and $1.31\%$ in accuracy $\delta_{1}$ (ResNet-34), as well as $5.49\%$ in Sq Rel error (EfficientNet-B5). Remarkably, on the larger-scale Structure3D, Elite360 performs favorably against the baseline by a significant margin, especially with ResNet-34. Comparisons with prevalent methods. In Tab. 2, we conduct a comprehensive comparison with prevalent supervised methods. From the results, we can observe that our approach achieves similar or even superior performance compared to existing both bi-projection fusion methods and single input methods at a significantly lower cost, particularly on two large-scale datasets, Matterport3D and Structure3D. Specifically, for the Matterport3D dataset, our Elite360D with ResNet-34 outperforms UniFuse by $2.53\%$ (Abs Rel) and with ResNet-50 outperforms BiFuse by $2.01\%$ (RMSE). For the Structure3D dataset, our Elite360D with ResNet-34 outperforms UniFuse by $4.48\%$ (Seq Rel), $1.99\%$ ( $\delta_{1}$ ). For performance on the Stanford2D3D dataset, we suspect it might be related to the ICOSAP point encoder. + +
Bi-projection feature fusionAbs Rel ↓Sq Rel ↓RMSE ↓δ1 ↑
SFA [2] + Add0.12760.10020.515084.27
SFA [2] + Concat0.11910.10190.514386.52
Only SA0.12040.10140.512186.26
Only DA0.11840.09720.494487.06
Our B2F (SA + DA)0.11150.09140.487588.15
+ +Table 4. The ablation results for B2F module. + +
Final fusionAbs Rel ↓Sq Rel ↓RMSE ↓δ1 ↑
Add0.16850.14810.580974.60
Average0.11980.09180.489386.65
Concatenation0.11450.09370.488087.66
Adaptive fusion [2]0.12440.09680.489186.08
Our gated fusion0.11150.09140.487588.15
+ +The limited data of Stanford2D3D dataset restricts the ability of the transformer-based point encoder to provide accurate global perception. Moreover, in Fig. 7, we present the qualitative comparisons. Our Elite360D can predict more accurate depth values based on the local-with-global perception capabilities (e.g., flowers, shelves and doors). Additional qualitative results and inference time comparisons can be found in the suppl. material. + +# 4.3. Ablation Study and Analyses + +Most of ablation experiments are conducted on the Mateport3D test dataset with ResNet34 as the backbone. + +The Effect of pre-training. We verify the effectiveness of ImageNet [11] pre-training with different encoder backbones. As observed from Tab. 3, the pre-training results in a significant improvement for all encoder backbones, e.g., $6.79\%$ improvement in accuracy $\delta_{1}$ (ResNet-34). Notably, pre-training has a relatively small impact on DilateFormer. Combined with the results in Tab. 1, the explanation of this phenomenon is that the default input resolution in pretrained models is different from actual input, thereby impacting the resolution-related position embeddings. In general, pre-training based on large-scale perspective images can effectively enhance the performance of models based on $360^{\circ}$ images and reduce the risk of overfitting. + +The effectiveness of B2F module. In Tab. 4, we compare four available bi-projection feature fusion modules. To align the spatial dimensions between ICOSAP point feature set and ERP feature map, we introduce SFA module from [2]. After that, we employ direct addition and concatenation to aggregate these two projections. We also achieve the bi-projection feature fusion with semantic-aware affinity attention (SA) alone and distance-aware affinity attention (DA) alone. Compared to the methods based solely on semantic-aware feature similarities (The first three rows), single distance-aware affinity attention can achieve better performance, which indicates that the spatial positional relationships boost the bi-projection feature fusion. Overall, our B2F module achieves the best performance. + +The superiority of ICOSAP. As only CP/TP's patch centers lie on the sphere's surface, we extract the feature em + +Table 5. The ablation results for the fusion of B2F module. + +
Method#Param(M)#FLOPs(G)Abs Rel ↓RMSE ↓δ1 ↑
ERP-CP25.6654.150.13690.540183.69
ERP-TP (N=18)25.6650.580.13280.538583.87
ERP-ICOSAP (Ours)15.4345.910.12720.527085.28
+ +Table 6. The comparison of different projections on Matterport3D. + +
N of {F^I}#Params (M)#FLOPs (G)Abs Rel↓Sq Rel↓RMSE↓δ1↑
2027.4166.290.11570.09950.502487.12
8025.5465.290.11150.09140.487588.15
32024.9864.320.11530.09430.490587.85
+ +Table 7. Impact of the ICOSAP point-wise feature number $N$ . Larger $N$ , fewer down-sampling blocks in the point encoder. + +bedding from each CP/TP patch and employs the patch center coordinates and feature embedding as the input of B2F module. In Tab. 6, we show the results with ResNet18 backbone. Our Elite360D, utilizing the ICOSAP point set, marginally outperforms models with CP and TP patches, while exhibiting fewer parameters and FLOPs. + +The effectiveness of gated fusion. We conduct an ablation study for the gated fusion block, outlined in Tab. 5. With the feature maps $F^{SA}$ and $F^{DA}$ , We compare it with the direct addition, average fusion, concatenation, and the adaptive fusion in [2]. The gated fusion performs best. + +ICOSAP point feature number $N$ . We study the effect of the ICOSAP feature point number (See Tab. 7). Too few points $(N = 20)$ lead to the over-concentrated global contextual information resulting from excessive down-sampling blocks, while too many points $(N = 320)$ lead to underconcentrated condition, resulting in insufficient perception of ERP pixel features. Best performance can be observed when $N = 80$ and we used $N = 80$ as default in this paper. + +# 5. Conclusion and Future Work + +In this paper, we proposed a novel bi-projection fusion solution for efficient 360 depth estimation. To address the limited local receptive field of ERP pixel-wise features and avoid expensive bi-projection fusion modules, we proposed a compact yet effective B2F module to learn the representations with local-with-global perceptions from ERP and ICOSAP. With an increase of 1M parameters, we significantly improved the performance of the ERP-based depth estimation baseline. Remarkably, our approach achieved performance on par with complex state-of-the-art methods. Future Work: From the experimental results, we observed that ERP-based depth baseline, with pre-trained Efficient-Net backbone, even outperforms existing specifically designed methods. Therefore, in the future, we will explore how to fully leverage different projections and successful perspective models for $360^{\circ}$ community. + +# Acknowledgement + +This paper is supported by the National Natural Science Foundation of China (NSF) under Grant No. NSFC22FYT45 and the Guangzhou City, University and Enterprise Joint Fund under Grant No.SL2022A03J01278. + +# References + +[1] Hao Ai, Zidong Cao, Jin Zhu, Haotian Bai, Yucheng Chen, and Ling Wang. Deep learning for omnidirectional vision: A survey and new perspectives. ArXiv, abs/2205.10468, 2022. 1 +[2] Hao Ai, Zidong Cao, Yan-Pei Cao, Ying Shan, and Lin Wang. Hrdfuse: Monocular $360^{\circ}$ depth estimation by collaboratively learning holistic-with-regional depth distributions. 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13273-13282, 2023. 2, 3, 6, 8 +[3] Iro Armeni, Sasha Sax, Amir Roshan Zamir, and Silvio Savarese. Joint 2d-3d-semantic data for indoor scene understanding. CoRR, abs/1702.01105, 2017. 2, 5, 6 +[4] Zidong Cao, Hao Ai, Yan Cao, Ying Shan, Xiaohu Qie, and Lin Wang. Omnizoomer: Learning to move and zoom in on sphere at high-resolution. 2023 IEEE/CVF International Conference on Computer Vision (ICCV), pages 12851-12861, 2023. 1 +[5] Angel X. Chang, Angela Dai, Thomas A. Funkhouser, Maciej Halber, Matthias Nießner, Manolis Savva, Shuran Song, Andy Zeng, and Yinda Zhang. Matterport3d: Learning from RGB-D data in indoor environments. In 3DV, pages 667-676. IEEE Computer Society, 2017. 1, 2, 5, 6 +[6] Zehui Chen, Zhenyu Li, Shiquan Zhang, Liangji Fang, Qinghong Jiang, Feng Zhao, Bolei Zhou, and Hang Zhao. Autoalign: Pixel-instance feature aggregation for multimodal 3d object detection. In International Joint Conference on Artificial Intelligence, 2022. 3 +[7] Hsien-Tzu Cheng, Chun-Hung Chao, Jin-Dong Dong, Hao-Kai Wen, Tyng-Luh Liu, and Min Sun. Cube padding for weakly-supervised saliency prediction in $360^{\circ}$ videos. 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1420-1429, 2018. 3 +[8] Xinjing Cheng, Peng Wang, Yanqi Zhou, Chenye Guan, and Ruigang Yang. Omnidirectional depth extension networks. In 2020 IEEE International Conference on Robotics and Automation (ICRA), pages 589-595, 2020. 2 +[9] Yanhua Cheng, Rui Cai, Zhiwei Li, Xin Zhao, and Kaiqi Huang. Locality-sensitive deconvolution networks with gated fusion for rgb-d indoor semantic segmentation. 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 1475-1483, 2017. 6 +[10] Taco Cohen, Maurice Weiler, Berkay Kicanaoglu, and Max Welling. Gauge equivariant convolutional networks and the icosahedral cnn. In International Conference on Machine Learning, 2019. 2 +[11] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. computer vision and pattern recognition, 2009. 2, 4, 7, 8 +[12] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at + +scale. In International Conference on Learning Representations, 2021. 3, 5 +[13] Marc Eder, Mykhailo Shvets, John Lim, and Jan-Michael Frahm. Tangent images for mitigating spherical distortion. 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12423-12431, 2019. 3 +[14] Kaiming He, X. Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 770-778, 2015. 2, 4, 5, 6, 7 +[15] ChiyuMax Jiang, Jingwei Huang, Karthik Kashinath, Prabhat Prabhat, Philip Marcus, and Matthias Niessner. Spherical cnns on unstructured grids. International Conference on Learning Representations, International Conference on Learning Representations, 2019. 2 +[16] Hualie Jiang, Zhe Sheng, Siyu Zhu, Zilong Dong, and Rui Huang. Unifuse: Unidirectional fusion for $360^{\circ}$ panorama depth estimation. IEEE Robotics and Automation Letters, 6: 1519-1526, 2021. 2, 3, 4, 6, 7 +[17] Jiayu Jiao, Yu-Ming Tang, Kun-Yu Lin, Yipeng Gao, Jinhua Ma, Yaowei Wang, and Wei-Shi Zheng. Dilateformer: Multi-scale dilated transformer for visual recognition. IEEE Transactions on Multimedia, pages 1-14, 2023. 5, 7 +[18] Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. CoRR, abs/1412.6980, 2014. 7 +[19] Varun Ravi Kumar, Senthil Kumar Yogamani, Hazem Rashed, Ganesh Sitsu, Christian Witt, Isabelle Leang, Stefan Milz, and Patrick Mäder. Omnidet: Surround view cameras based multi-task visual perception network for autonomous driving. IEEE Robotics and Automation Letters, 6:2830-2837, 2021. 1 +[20] Iro Laina, C. Rupprecht, Vasileios Belagiannis, Federico Tombari, and Nassir Navab. Deeper depth prediction with fully convolutional residual networks. 3DV 2016, pages 239-248, 2016. 6 +[21] Yeonkun Lee, Jaeseok Jeong, Jong Seob Yun, Wonjune Cho, and Kuk jin Yoon. Spherephd: Applying cnns on a spherical polyhedron representation of $360^{\circ}$ images. 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9173-9181, 2018. 2, 3, 4 +[22] Meng Li, Senbo Wang, Weihao Yuan, Weichao Shen, Zhe Sheng, and Zilong Dong. $S^2$ net: Accurate panorama depth estimation on spherical surface. IEEE Robotics and Automation Letters, 8:1053-1060, 2023. 3 +[23] Yunhao Li, Wei Shen, Zhongpai Gao, Yucheng Zhu, Guangtao Zhai, and Guodong Guo. Looking here or there? gaze following in 360-degree images. 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 3722-3731, 2021. 1 +[24] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 9992-10002, 2021. 2, 3, 4, 5, 7 +[25] Yunze Man, Liangyan Gui, and Yu-Xiong Wang. Bev-guided multi-modality fusion for driving perception. 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 21960-21969, 2023. 3 + +[26] Daniel Martin, Ana Serrano, Alexander W. Bergman, Gordon Wetzstein, and Belén Masiá. Scangan360: A generative model of realistic scanpaths for $360^{\circ}$ images. IEEE Transactions on Visualization and Computer Graphics, 28:2003-2013, 2021. 1 +[27] Rafael Monroy, Sebastian Lutz, Tejo Chalasani, and Aljoscha Smolic. Salnet360: Saliency maps for omnidirectional images with cnn. Signal Process. Image Commun., 69:26-34, 2017. 3 +[28] Giovanni Pintore, Eva Almansa, and Jens Schneider. Slicenet: deep dense depth estimation from a single indoor panorama using a slice-based representation. 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11531-11540, 2021. 1, 3, 6 +[29] Manuel Rey-Area, Mingze Yuan, and Christian Richardt. 360monodepth: High-resolution $360^{\circ}$ monocular depth estimation. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3752-3762, 2022. 3 +[30] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In MICCAI (3), pages 234-241. Springer, 2015. 3, 6 +[31] Mehran Shakerinava and Siamak Ravanbakhsh. Equivariant networks for pixelized spheres. Proceedings of the 38th International Conference on Machine Learning, ICML, abs/2106.06662, 2021. 3, 4 +[32] Zhijie Shen, Chunyu Lin, Kang Liao, Lang Nie, Zishuo Zheng, and Yao Zhao. Panofformer: Panorama transformer for indoor $360^{\circ}$ depth estimation. In European Conference on Computer Vision, 2022. 2, 3, 6 +[33] Yu-Chuan Su and Kristen Grauman. Kernel transformer networks for compact spherical convolution. 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9434-9443, 2018. 1 +[34] Cheng Sun, Min Sun, and Hwann-Tzong Chen. Hohonet: 360 indoor holistic understanding with latent horizontal features. 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2573-2582, 2020. 1, 3 +[35] Mingxing Tan and Quoc V. Le. Efficientnet: Rethinking model scaling for convolutional neural networks. In ICML, pages 6105-6114. PMLR, 2019. 4, 5, 7 +[36] Keisuke Tateno, Nassir Navab, and Federico Tombari. Distortion-aware convolutional filters for dense prediction in panoramic images. In European Conference on Computer Vision, 2018. 1, 2 +[37] Fu-En Wang, Yu-Hsuan Yeh, Min Sun, Wei-Chen Chiu, and Yi-Hsuan Tsai. Bifuse: Monocular 360 depth estimation via bi-projection fusion. 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 459–468, 2020. 2, 3, 4, 6 +[38] Fu-En Wang, Yu-Hsuan Yeh, Yi-Hsuan Tsai, Wei-Chen Chiu, and Min Sun. Bifuse++: Self-supervised and efficient bi-projection fusion for $360^{\circ}$ depth estimation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45: 5448-5460, 2022. 2, 3, 6 +[39] Sirui Wang, Di Liang, Jian Song, Yuntao Li, and Wei Wu. DABERT: dual attention enhanced BERT for semantic matching. In COLING, pages 1645-1654. International Committee on Computational Linguistics, 2022. 6 + +[40] X. Wang, Ross B. Girshick, Abhinav Kumar Gupta, and Kaiming He. Non-local neural networks. 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7794-7803, 2017. 4 +[41] Yu yang Li, Yuliang Guo, Zhixin Yan, Xinyu Huang, Ye Duan, and Liu Ren. Omnifusion: 360 monocular depth estimation via geometry-aware fusion. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2791-2800, 2022. 2, 3, 6 +[42] Xiaoqing Ye, Mao Shu, Hanyu Li, Yifeng Shi, Yingying Li, Guan-Sheng Wang, Xiao Tan, and Errui Ding. Rope3d: The roadside perception dataset for autonomous driving and monocular 3d object detection task. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 21309-21318, 2022. 1 +[43] Youngho Yoon, Inchul Chung, Lin Wang, and Kuk-Jin Yoon. Spheresr: $360^{\circ}$ image super-resolution with arbitrary projection via continuous spherical image representation. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5667-5676, 2021. 1, 2, 3 +[44] Haozheng Yu, Lu He, Bing Jian, Weiwei Feng, and Shanghua Liu. Panelnet: Understanding 360 indoor environment via panel representation. 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 878-887, 2023. 1, 3 +[45] Ilwi Yun, Chan-Yong Shin, Hyunku Lee, Hyuk-Jae Lee, and Chae-Eun Rhee. Egformer: Equirectangular geometry-biased transformer for 360 depth estimation. ArXiv, abs/2304.07803, 2023. 2, 3, 4, 6 +[46] Chao Zhang, Stephan Liwicki, William Smith, and Roberto Cipolla. Orientation-aware semantic segmentation on icosahedron spheres. 2019 IEEE/CVF International Conference on Computer Vision (ICCV), pages 3532-3540, 2019. 2, 4 +[47] Hengshuang Zhao, Li Jiang, Jiaya Jia, Philip H. S. Torr, and Vladlen Koltun. Point transformer. 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 16239-16248, 2020. 4, 7 +[48] Jia Zheng, Junfei Zhang, Jing Li, Rui Tang, Shenghua Gao, and Zihan Zhou. Structured3d: A large photo-realistic dataset for structured 3d modeling. In European Conference on Computer Vision, 2019. 5, 6 +[49] Xu Zheng, Jinjing Zhu, Ye-Peng Liu, Zidong Cao, Chong Fu, and Lin Wang. Both style and distortion matter: Dualpath unsupervised domain adaptation for panoramic semantic segmentation. 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1285-1295, 2023. 1 +[50] Chuanqing Zhuang, Zhengda Lu, Yiqun Wang, Jun Xiao, and Ying Wang. Acdnet: Adaptively combined dilated convolution for monocular panorama depth estimation. CoRR, abs/2112.14440, 2021. 1, 3 +[51] Nikolaos Zioulis, Antonis Karakottas, Dimitrios Zarpalas, and Petros Daras. Omndepth: Dense depth estimation for indoors spherical panoramicas. In ECCV (6), pages 453-471. Springer, 2018. 1, 2 \ No newline at end of file diff --git a/2024/Elite360D_ Towards Efficient 360 Depth Estimation via Semantic- and Distance-Aware Bi-Projection Fusion/images.zip b/2024/Elite360D_ Towards Efficient 360 Depth Estimation via Semantic- and Distance-Aware Bi-Projection Fusion/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..0e7eaf0bcfbc7925f043abfa8f7de3a4c1b3d890 --- /dev/null +++ b/2024/Elite360D_ Towards Efficient 360 Depth Estimation via Semantic- and Distance-Aware Bi-Projection Fusion/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:fde0456957b8a74a26518be20c0012603a9f2967912a44d9cf7a131ff0133115 +size 840974 diff --git a/2024/Elite360D_ Towards Efficient 360 Depth Estimation via Semantic- and Distance-Aware Bi-Projection Fusion/layout.json b/2024/Elite360D_ Towards Efficient 360 Depth Estimation via Semantic- and Distance-Aware Bi-Projection Fusion/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..ec780a3550595c6d04661a5e67a005efae443bd4 --- /dev/null +++ b/2024/Elite360D_ Towards Efficient 360 Depth Estimation via Semantic- and Distance-Aware Bi-Projection Fusion/layout.json @@ -0,0 +1,10157 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 82, + 102, + 512, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 82, + 102, + 512, + 140 + ], + "spans": [ + { + "bbox": [ + 82, + 102, + 512, + 140 + ], + "type": "text", + "content": "Elite360D: Towards Efficient 360 Depth Estimation via Semantic- and Distance-Aware Bi-Projection Fusion" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 238, + 160, + 358, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 238, + 160, + 358, + 175 + ], + "spans": [ + { + "bbox": [ + 238, + 160, + 358, + 175 + ], + "type": "text", + "content": "Hao Ai" + }, + { + "bbox": [ + 238, + 160, + 358, + 175 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 238, + 160, + 358, + 175 + ], + "type": "text", + "content": " Lin Wang" + }, + { + "bbox": [ + 238, + 160, + 358, + 175 + ], + "type": "inline_equation", + "content": "^{1,2*}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 169, + 175, + 424, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 169, + 175, + 424, + 190 + ], + "spans": [ + { + "bbox": [ + 169, + 175, + 424, + 190 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 169, + 175, + 424, + 190 + ], + "type": "text", + "content": "AI Thrust, HKUST(GZ) " + }, + { + "bbox": [ + 169, + 175, + 424, + 190 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 169, + 175, + 424, + 190 + ], + "type": "text", + "content": "Dept. of CSE, HKUST" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 171, + 191, + 422, + 202 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 171, + 191, + 422, + 202 + ], + "spans": [ + { + "bbox": [ + 171, + 191, + 422, + 202 + ], + "type": "text", + "content": "hai033@connect.hkust-gz.edu.cn, linwang@ust.hk" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 143, + 231, + 192, + 243 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 231, + 192, + 243 + ], + "spans": [ + { + "bbox": [ + 143, + 231, + 192, + 243 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 256, + 290, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 256, + 290, + 579 + ], + "spans": [ + { + "bbox": [ + 46, + 256, + 290, + 579 + ], + "type": "text", + "content": "360 depth estimation has recently received great attention for 3D reconstruction owing to its omnidirectional field of view (FoV). Recent approaches are predominantly focused on cross-projection fusion with geometry-based reprojection: they fuse 360 images with equirectangular projection (ERP) and another projection type, e.g., cubemap projection to estimate depth with the ERP format. However, these methods suffer from 1) limited local receptive fields, making it hardly possible to capture large FoV scenes, and 2) prohibitive computational cost, caused by the complex cross-projection fusion module design. In this paper, we propose Elite360D, a novel framework that inputs the ERP image and icosahedron projection (ICOSAP) point set, which is undistorted and spatially continuous. Elite360D is superior in its capacity in learning a representation from a local-with-global perspective. With a flexible ERP image encoder, it includes an ICOSAP point encoder, and a Bi-projection Bi-attention Fusion (B2F) module (totally " + }, + { + "bbox": [ + 46, + 256, + 290, + 579 + ], + "type": "inline_equation", + "content": "\\sim 1M" + }, + { + "bbox": [ + 46, + 256, + 290, + 579 + ], + "type": "text", + "content": " parameters). Specifically, the ERP image encoder can take various perspective image-trained backbones (e.g., ResNet, Transformer) to extract local features. The point encoder extracts the global features from the ICOSAP. Then, the B2F module captures the semantic- and distance-aware dependencies between each pixel of the ERP feature and the entire ICOSAP feature set. Without specific backbone design and obvious computational cost increase, Elite360D outperforms the prior arts on several benchmark datasets." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 586, + 158, + 598 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 586, + 158, + 598 + ], + "spans": [ + { + "bbox": [ + 47, + 586, + 158, + 598 + ], + "type": "text", + "content": "Multimedia Material" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 601, + 287, + 626 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 601, + 287, + 626 + ], + "spans": [ + { + "bbox": [ + 46, + 601, + 287, + 626 + ], + "type": "text", + "content": "For videos, code, demo and more information, you can visit https://VLIS2022.github.io/Elite360D/" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 629, + 128, + 641 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 629, + 128, + 641 + ], + "spans": [ + { + "bbox": [ + 47, + 629, + 128, + 641 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 647, + 288, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 647, + 288, + 696 + ], + "spans": [ + { + "bbox": [ + 47, + 647, + 288, + 696 + ], + "type": "inline_equation", + "content": "360^{\\circ}" + }, + { + "bbox": [ + 47, + 647, + 288, + 696 + ], + "type": "text", + "content": " images capture the complete surrounding environment in one shot with an ultra-wide field-of-view (FoV) of " + }, + { + "bbox": [ + 47, + 647, + 288, + 696 + ], + "type": "inline_equation", + "content": "180^{\\circ} \\times 360^{\\circ}" + }, + { + "bbox": [ + 47, + 647, + 288, + 696 + ], + "type": "text", + "content": ", which is broadly applied to applications, e.g., autonomous driving [1, 19, 42, 43, 49] and virtual reality" + } + ] + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 307, + 230, + 426, + 341 + ], + "blocks": [ + { + "bbox": [ + 307, + 230, + 426, + 341 + ], + "lines": [ + { + "bbox": [ + 307, + 230, + 426, + 341 + ], + "spans": [ + { + "bbox": [ + 307, + 230, + 426, + 341 + ], + "type": "image", + "image_path": "4b8c1128a92aaa7822a410fa93cb01b5cfaaf43b3b1e06e4afc0f851443c59e5.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 342, + 342, + 396, + 350 + ], + "lines": [ + { + "bbox": [ + 342, + 342, + 396, + 350 + ], + "spans": [ + { + "bbox": [ + 342, + 342, + 396, + 350 + ], + "type": "text", + "content": "Model #FLOPs(G)" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 427, + 231, + 544, + 341 + ], + "blocks": [ + { + "bbox": [ + 427, + 231, + 544, + 341 + ], + "lines": [ + { + "bbox": [ + 427, + 231, + 544, + 341 + ], + "spans": [ + { + "bbox": [ + 427, + 231, + 544, + 341 + ], + "type": "image", + "image_path": "6ccd0c6784e3ed90697fd7c4da8aba8d21897d4ef0eaceaa2b3eb9f282f127f6.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 309, + 354, + 421, + 411 + ], + "blocks": [ + { + "bbox": [ + 309, + 354, + 421, + 411 + ], + "lines": [ + { + "bbox": [ + 309, + 354, + 421, + 411 + ], + "spans": [ + { + "bbox": [ + 309, + 354, + 421, + 411 + ], + "type": "image", + "image_path": "df0039b601e777013f7ebd4a22ecf72e05ba18dd39bdc61f392415cb7e9bdbdb.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 349, + 411, + 383, + 419 + ], + "lines": [ + { + "bbox": [ + 349, + 411, + 383, + 419 + ], + "spans": [ + { + "bbox": [ + 349, + 411, + 383, + 419 + ], + "type": "text", + "content": "ERP Image" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 430, + 354, + 543, + 411 + ], + "blocks": [ + { + "bbox": [ + 422, + 343, + 433, + 354 + ], + "lines": [ + { + "bbox": [ + 422, + 343, + 433, + 354 + ], + "spans": [ + { + "bbox": [ + 422, + 343, + 433, + 354 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 430, + 354, + 543, + 411 + ], + "lines": [ + { + "bbox": [ + 430, + 354, + 543, + 411 + ], + "spans": [ + { + "bbox": [ + 430, + 354, + 543, + 411 + ], + "type": "image", + "image_path": "bb551f9bef0ee005a962d3966ac4eed6aac8138259cd520f9cf94b3ede5d6865.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 473, + 411, + 502, + 419 + ], + "lines": [ + { + "bbox": [ + 473, + 411, + 502, + 419 + ], + "spans": [ + { + "bbox": [ + 473, + 411, + 502, + 419 + ], + "type": "text", + "content": "GT Depth" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 309, + 421, + 421, + 478 + ], + "blocks": [ + { + "bbox": [ + 309, + 421, + 421, + 478 + ], + "lines": [ + { + "bbox": [ + 309, + 421, + 421, + 478 + ], + "spans": [ + { + "bbox": [ + 309, + 421, + 421, + 478 + ], + "type": "image", + "image_path": "bfd43e37b788093d55dfe53c89684b1fd3d1f5de07820cf265f93fcbb2556269.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 343, + 479, + 387, + 487 + ], + "lines": [ + { + "bbox": [ + 343, + 479, + 387, + 487 + ], + "spans": [ + { + "bbox": [ + 343, + 479, + 387, + 487 + ], + "type": "text", + "content": "Equi-Baseline" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 306, + 493, + 545, + 548 + ], + "lines": [ + { + "bbox": [ + 306, + 493, + 545, + 548 + ], + "spans": [ + { + "bbox": [ + 306, + 493, + 545, + 548 + ], + "type": "text", + "content": "Figure 1. (a) Performance (RMSE error) curves on M3D test dataset [5]. Larger circles mean more parameters (e.g., ResNet18, ResNet34, ResNet50) and lower errors mean better performance. (b) Comparison with the ResNet34 as the ERP encoder backbone. With only 1M more parameters, our depth result is more accurate." + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 430, + 421, + 543, + 478 + ], + "blocks": [ + { + "bbox": [ + 430, + 421, + 543, + 478 + ], + "lines": [ + { + "bbox": [ + 430, + 421, + 543, + 478 + ], + "spans": [ + { + "bbox": [ + 430, + 421, + 543, + 478 + ], + "type": "image", + "image_path": "698131984424d9e7b1c0bd995bd3c5a496ed09ed2cd3b34c8dee377974d61f94.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 466, + 479, + 509, + 486 + ], + "lines": [ + { + "bbox": [ + 466, + 479, + 509, + 486 + ], + "spans": [ + { + "bbox": [ + 466, + 479, + 509, + 486 + ], + "type": "text", + "content": "Our Elite360D" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 422, + 483, + 433, + 492 + ], + "lines": [ + { + "bbox": [ + 422, + 483, + 433, + 492 + ], + "spans": [ + { + "bbox": [ + 422, + 483, + 433, + 492 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + } + ], + "index": 22 + }, + { + "bbox": [ + 304, + 556, + 545, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 556, + 545, + 640 + ], + "spans": [ + { + "bbox": [ + 304, + 556, + 545, + 640 + ], + "type": "text", + "content": "[4, 23, 26]. The ability to infer the 3D structure of a surrounding scene has inspired active research for monocular 360 depth estimation. Equirectangular projection (ERP) is the commonly used projection type that can provide a complete view of a scene. An ERP image, a.k.a., panorama, samples pixels with a higher density at the poles compared to the equator, resulting in spherical distortions (Fig. 2(c))." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 304, + 642, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 642, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 642, + 547, + 714 + ], + "type": "text", + "content": "In recent years, to address the spherical distortions, several ah-hoc designs [33, 36, 50, 51] focus on the distortion-aware convolutional kernels and sample grids from undistorted tangent planes on the sphere. These grids are backprojected to the corresponding locations of the ERP image. Also, some methods [28, 34, 44] partition the ERP" + } + ] + } + ], + "index": 27 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 702, + 219, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 702, + 219, + 713 + ], + "spans": [ + { + "bbox": [ + 58, + 702, + 219, + 713 + ], + "type": "text", + "content": "*Corresponding author (e-mail: linwang@ust.hk)" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 295, + 784, + 315, + 792 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 784, + 315, + 792 + ], + "spans": [ + { + "bbox": [ + 295, + 784, + 315, + 792 + ], + "type": "text", + "content": "9926" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 240 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 240 + ], + "type": "text", + "content": "input into multiple vertical slices and extract a compact representation from these slices to extend the receptive field of each pixel for the ERP representation. Unfortunately, these methods result in substantial computational costs. Meanwhile, the limited local receptive field of the convolutional filters is insufficient to provide the global perception of the panorama. Therefore, PanoFormer [32] and EGFormer [45] design specific transformer architectures to effectively model the long-range dependencies between the global context of large-FoV ERP images. Nevertheless, the receptive fields are confined by the sizes of local attention windows. Furthermore, the aforementioned data-driven methods often yield sub-optimal results when trained from scratch on limited " + }, + { + "bbox": [ + 46, + 72, + 289, + 240 + ], + "type": "inline_equation", + "content": "360^{\\circ}" + }, + { + "bbox": [ + 46, + 72, + 289, + 240 + ], + "type": "text", + "content": " depth dataset (See Tab. 2)." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 243, + 289, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 243, + 289, + 518 + ], + "spans": [ + { + "bbox": [ + 46, + 243, + 289, + 518 + ], + "type": "text", + "content": "With the power of existing large-scale perspective images [11], several methods [2, 16, 37, 38] employ the pretrained models - designed for perspective images - as encoder backbones to extract ERP feature maps, and propose the cross-projection fusion to rectify distortions in the ERP feature maps. Specifically, these methods introduce the less-distorted planar projection data, i.e., cubemap projection (CP) patches or tangent projection (TP) patches (See Fig. 2(d), (e)), and unify the spatial dimensions between different projections for the cross-projection fusion. BiFuse and UniFuse [16, 37, 38] propose the C2E module to reproject the content in patch-wise CP features into ERP grid following the spherical geometric relationships. Furthermore, they leverage the fusion between ERP feature maps and C2E feature maps to enhance depth estimation accuracy. However, their concatenation-based feature fusion across multiple scales brings a significant computational burden. Based on the feature similarity, HRDFuse [2] spatially aligns the ERP pixel features and TP patch features. This makes it possible to significantly marry the global context in the ERP feature map and regional structural details in the TP patch features. However, the decoder-level fusion remarkably increases the computational memory and cost." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 521, + 289, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 521, + 289, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 521, + 289, + 714 + ], + "type": "text", + "content": "These less-distorted gnomonic projections, i.e., CP and TP patches, are spatially discontinuous and require complex re-projection operations for the desired ERP format predictions. Inspired by these issues, we introduce a more powerful projection, the icosahedron projection (ICOSAP) [21, 46], see Fig. 2(f). Importantly, ICOSAP is a spatially continuous and globally perceptive non-Euclidean projection for " + }, + { + "bbox": [ + 46, + 521, + 289, + 714 + ], + "type": "inline_equation", + "content": "360^{\\circ}" + }, + { + "bbox": [ + 46, + 521, + 289, + 714 + ], + "type": "text", + "content": " images. In light of this, we propose an efficient and effective fusion-based framework, named Elite360D that takes the best of ERP and ICOSAP by learning a representation from a local-with-global perspective. It comprises three components: an ERP encoder, an ICOSAP encoder, and the bi-projection bi-attention fusion (B2F) module. For the ERP image encoder, our Elite360D flexibly supports a wide range of off-the-shelf 2D models, e.g., ResNet [14], Swin transformer [24], as the backbone to" + } + ] + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 325, + 70, + 365, + 111 + ], + "blocks": [ + { + "bbox": [ + 325, + 70, + 365, + 111 + ], + "lines": [ + { + "bbox": [ + 325, + 70, + 365, + 111 + ], + "spans": [ + { + "bbox": [ + 325, + 70, + 365, + 111 + ], + "type": "image", + "image_path": "559fd2dd5357f51f53b367b8c3c250f69194382d4865fadbb1f82775ba3e7437.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 326, + 113, + 364, + 120 + ], + "lines": [ + { + "bbox": [ + 326, + 113, + 364, + 120 + ], + "spans": [ + { + "bbox": [ + 326, + 113, + 364, + 120 + ], + "type": "text", + "content": "(a) Panorama" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 396, + 71, + 430, + 110 + ], + "blocks": [ + { + "bbox": [ + 396, + 71, + 430, + 110 + ], + "lines": [ + { + "bbox": [ + 396, + 71, + 430, + 110 + ], + "spans": [ + { + "bbox": [ + 396, + 71, + 430, + 110 + ], + "type": "image", + "image_path": "517fb55fd3973852e3fd942b48bd7c0440f89f20555a0d131e6bcd5e61239cdd.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 403, + 113, + 425, + 119 + ], + "lines": [ + { + "bbox": [ + 403, + 113, + 425, + 119 + ], + "spans": [ + { + "bbox": [ + 403, + 113, + 425, + 119 + ], + "type": "text", + "content": "(b) ERP" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 463, + 72, + 526, + 111 + ], + "blocks": [ + { + "bbox": [ + 463, + 72, + 526, + 111 + ], + "lines": [ + { + "bbox": [ + 463, + 72, + 526, + 111 + ], + "spans": [ + { + "bbox": [ + 463, + 72, + 526, + 111 + ], + "type": "image", + "image_path": "1c8897cd52e20ba6a64eeecdbe62c477c8f42d61c95b12daacf070ab3f16ac4d.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 473, + 113, + 517, + 119 + ], + "lines": [ + { + "bbox": [ + 473, + 113, + 517, + 119 + ], + "spans": [ + { + "bbox": [ + 473, + 113, + 517, + 119 + ], + "type": "text", + "content": "(c) ERP (flatten)" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 324, + 124, + 365, + 163 + ], + "blocks": [ + { + "bbox": [ + 324, + 124, + 365, + 163 + ], + "lines": [ + { + "bbox": [ + 324, + 124, + 365, + 163 + ], + "spans": [ + { + "bbox": [ + 324, + 124, + 365, + 163 + ], + "type": "image", + "image_path": "8683cec09fd50640b30f1a71c3147380a3f1f434af1c3e190bcc7d0e79d4b8ed.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 336, + 165, + 354, + 171 + ], + "lines": [ + { + "bbox": [ + 336, + 165, + 354, + 171 + ], + "spans": [ + { + "bbox": [ + 336, + 165, + 354, + 171 + ], + "type": "text", + "content": "(d) CP" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 394, + 124, + 435, + 163 + ], + "blocks": [ + { + "bbox": [ + 394, + 124, + 435, + 163 + ], + "lines": [ + { + "bbox": [ + 394, + 124, + 435, + 163 + ], + "spans": [ + { + "bbox": [ + 394, + 124, + 435, + 163 + ], + "type": "image", + "image_path": "adb587f419b2ea73a279429ddeaab0947ee794a867d4317b62e7cd5fd837f3b6.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 406, + 164, + 424, + 171 + ], + "lines": [ + { + "bbox": [ + 406, + 164, + 424, + 171 + ], + "spans": [ + { + "bbox": [ + 406, + 164, + 424, + 171 + ], + "type": "text", + "content": "(e) TP" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 476, + 125, + 514, + 163 + ], + "blocks": [ + { + "bbox": [ + 476, + 125, + 514, + 163 + ], + "lines": [ + { + "bbox": [ + 476, + 125, + 514, + 163 + ], + "spans": [ + { + "bbox": [ + 476, + 125, + 514, + 163 + ], + "type": "image", + "image_path": "ed904e61df0c4ff3dd2dfdcd9325d330125cf773612ec50c478beac5f4aa4ec8.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 468, + 165, + 522, + 171 + ], + "lines": [ + { + "bbox": [ + 468, + 165, + 522, + 171 + ], + "spans": [ + { + "bbox": [ + 468, + 165, + 522, + 171 + ], + "type": "text", + "content": "(f) ICOSAP (level=1)" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 307, + 174, + 543, + 185 + ], + "lines": [ + { + "bbox": [ + 307, + 174, + 543, + 185 + ], + "spans": [ + { + "bbox": [ + 307, + 174, + 543, + 185 + ], + "type": "text", + "content": "Figure 2. Different projections of a spherical imaging panorama." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 190, + 545, + 334 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 190, + 545, + 334 + ], + "spans": [ + { + "bbox": [ + 304, + 190, + 545, + 334 + ], + "type": "text", + "content": "extract ERP feature maps. This potentially reduces the overfitting problems, particularly on small-scale 360 depth estimation datasets [3, 5]. For efficiency, we represent the ICOSAP spheres as the discrete point sets rather than icosahedron meshes [10, 15, 46] or unfolded representations [21, 43]. As such, we can avoid semantic information redundancy due to dense ERP pixels and maintain the spatial position information. Then, to enable each ERP pixel feature with local receptive fields to perceive the whole scene, B2F module captures the semantic- and distance-aware dependencies between each ERP pixel feature and entire ICOSAP feature set." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 335, + 546, + 454 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 335, + 546, + 454 + ], + "spans": [ + { + "bbox": [ + 304, + 335, + 546, + 454 + ], + "type": "text", + "content": "We conduct extensive experiments on three different datasets with different encoder backbones to demonstrate the flexibility and effectiveness of our Elite360D. The experimental results indicate that our Elite360D significantly improves plain-backbones' performance with minimal computational memory (only about 1M parameters) (See Fig. 1(a)). Note that, based on the simple ERP depth estimation baseline, Elite360D only performs bi-projection feature fusion at the last feature layer and achieves results that are on par with leading methods, such as [2, 41]." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 454, + 547, + 598 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 454, + 547, + 598 + ], + "spans": [ + { + "bbox": [ + 304, + 454, + 547, + 598 + ], + "type": "text", + "content": "In summary, our main contributions are three-fold: (I) We introduce the ICOSAP with spatial continuity and global perception and represent them as the discrete point sets to reduce the computation cost and maintain the spatial information; (II) We propose the B2F module that jointly models the semantic and spatial dependencies between ICOSAP and ERP features and learn the representations with global-with-local receptive fields for large-FoV scenes; (III) Building upon the B2F module, we propose a novel framework that supports diverse off-the-shelf models as encoder backbones, showing better flexibility than the specially designed models, e.g., [32, 45]." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 306, + 609, + 398, + 621 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 609, + 398, + 621 + ], + "spans": [ + { + "bbox": [ + 306, + 609, + 398, + 621 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 629, + 545, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 629, + 545, + 665 + ], + "spans": [ + { + "bbox": [ + 304, + 629, + 545, + 665 + ], + "type": "text", + "content": "Monocular 360 depth estimation Existing methods can be categorized into two types: one with single projection input, and the other with bi-projection inputs." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "type": "text", + "content": "1) Single Projection Input With the ERP images as the inputs, many methods focus on mitigating the inherent spherical distortion issues. Several methods [8, 36, 51] rectify distortions by modifying the receptive fields of tradi" + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 783, + 315, + 792 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 783, + 315, + 792 + ], + "spans": [ + { + "bbox": [ + 294, + 783, + 315, + 792 + ], + "type": "text", + "content": "9927" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 65, + 70, + 527, + 242 + ], + "blocks": [ + { + "bbox": [ + 65, + 70, + 527, + 242 + ], + "lines": [ + { + "bbox": [ + 65, + 70, + 527, + 242 + ], + "spans": [ + { + "bbox": [ + 65, + 70, + 527, + 242 + ], + "type": "image", + "image_path": "4b8c68120290d19dcc009978e2f56cedab404557aec0de5506e1feb4ff66701a.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 243, + 547, + 288 + ], + "lines": [ + { + "bbox": [ + 46, + 243, + 547, + 288 + ], + "spans": [ + { + "bbox": [ + 46, + 243, + 547, + 288 + ], + "type": "text", + "content": "Figure 3. (a) An overview of our Elite360D framework, comprising image-based ERP feature extraction (Sec. 3.1), point-based ICOSAP feature extraction (Sec. 3.2), and Bi-projection Bi-attention fusion (B2F) (Sec. 3.3). For better visualization, we do not show the skip connections [30] at the decoding stage. (b) Illustration of the B2F module, consisting of three parts: semantic-aware affinity attention block (Fig. 5), distance-aware affinity attention block (Fig. 6) and gated fusion (Eq. 5)." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 293, + 289, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 293, + 289, + 449 + ], + "spans": [ + { + "bbox": [ + 46, + 293, + 289, + 449 + ], + "type": "text", + "content": "tional convolutional kernels. Besides, OmniFusion [41] and 360MonoDepth [29] follow [13] to predict depth maps from less distorted TP patches. Similarly, PanelNet [44] predicts depth maps on the vertical slices of ERP images and then aggregates them. Moreover, some methods [28, 34, 50] focus on tackling the small receptive fields of convolution filters for processing large FoV scenes. More recently, PanoFormer [32] and EGFormer [45] design transformer architectures [12] for ERP images to model the long-range dependencies. By contrast, " + }, + { + "bbox": [ + 46, + 293, + 289, + 449 + ], + "type": "inline_equation", + "content": "\\mathbf{S}^2\\mathrm{Net}" + }, + { + "bbox": [ + 46, + 293, + 289, + 449 + ], + "type": "text", + "content": " [22] extracts robust features from ERP images with the Swin transformer [24] and projects the ERP feature maps onto the spherical surface to minimize distortions and maintain spatial consistency." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 449, + 288, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 449, + 288, + 581 + ], + "spans": [ + { + "bbox": [ + 46, + 449, + 288, + 581 + ], + "type": "text", + "content": "2) Bi-Projection Inputs. BiFuse [37] utilizes a dense fusion strategy to bidirectionally fuse ERP and CP features at both encoding and decoding stages. In contrast, BiFuse++[38] and UniFuse [16] fuse ERP and CP features at the encoding stage. Recently, HRDFuse [2] achieved the SOTA performance based on the adaptive fusion of ERP and TP predictions. By contrast, we introduce a non-Euclidean projection, ICOSAP, which is less distorted and spatially continuous. We leverage the global perception capacity of ICOSAP to enable each ERP pixel-wise feature, with a limited local receptive field, to capture the entire scene." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 582, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 582, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 582, + 288, + 715 + ], + "type": "text", + "content": "Representations for 360 Images ERP is the most commonly used projection, which projects a panorama onto a cylinder (Fig. 2(b)) and then unfolds it into a plane (Fig. 2(c)). However, ERP maps latitudes and longitudes onto the vertical and horizontal axes with equal spacing on the plane, leading to distortions. CP [27] projects the panoramas onto six faces of a cube where each face is less distorted (Fig. 2(d)). Building upon CP, some padding methods [7, 37] remove the boundary inconsistency between the adjacent CP patches. Recently, Eder et al. [13] proposed the tangent projection (TP) (Fig. 2(e)), to signif" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 307, + 291, + 358, + 345 + ], + "blocks": [ + { + "bbox": [ + 307, + 291, + 358, + 345 + ], + "lines": [ + { + "bbox": [ + 307, + 291, + 358, + 345 + ], + "spans": [ + { + "bbox": [ + 307, + 291, + 358, + 345 + ], + "type": "image", + "image_path": "1da8f6d0fed53242d5b8ed7279e74011204c8d01b3a5c7e818bfa7bc7c833b0c.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 317, + 346, + 347, + 357 + ], + "lines": [ + { + "bbox": [ + 317, + 346, + 347, + 357 + ], + "spans": [ + { + "bbox": [ + 317, + 346, + 347, + 357 + ], + "type": "text", + "content": "(a) " + }, + { + "bbox": [ + 317, + 346, + 347, + 357 + ], + "type": "inline_equation", + "content": "l = 0" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 366, + 292, + 419, + 345 + ], + "blocks": [ + { + "bbox": [ + 366, + 292, + 419, + 345 + ], + "lines": [ + { + "bbox": [ + 366, + 292, + 419, + 345 + ], + "spans": [ + { + "bbox": [ + 366, + 292, + 419, + 345 + ], + "type": "image", + "image_path": "f3406636289f7ea1271712b536fd94cec36ba985895a4f4c2edc4b318f99eab3.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 377, + 346, + 406, + 357 + ], + "lines": [ + { + "bbox": [ + 377, + 346, + 406, + 357 + ], + "spans": [ + { + "bbox": [ + 377, + 346, + 406, + 357 + ], + "type": "text", + "content": "(b) " + }, + { + "bbox": [ + 377, + 346, + 406, + 357 + ], + "type": "inline_equation", + "content": "l = 1" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 427, + 292, + 482, + 345 + ], + "blocks": [ + { + "bbox": [ + 427, + 292, + 482, + 345 + ], + "lines": [ + { + "bbox": [ + 427, + 292, + 482, + 345 + ], + "spans": [ + { + "bbox": [ + 427, + 292, + 482, + 345 + ], + "type": "image", + "image_path": "ec00a3b3e4a7e4e596147cdb250833a5ee0e0ab2b8e160f70d55ca63400e035b.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 440, + 346, + 468, + 357 + ], + "lines": [ + { + "bbox": [ + 440, + 346, + 468, + 357 + ], + "spans": [ + { + "bbox": [ + 440, + 346, + 468, + 357 + ], + "type": "text", + "content": "(c) " + }, + { + "bbox": [ + 440, + 346, + 468, + 357 + ], + "type": "inline_equation", + "content": "l = 2" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 491, + 292, + 545, + 345 + ], + "blocks": [ + { + "bbox": [ + 491, + 292, + 545, + 345 + ], + "lines": [ + { + "bbox": [ + 491, + 292, + 545, + 345 + ], + "spans": [ + { + "bbox": [ + 491, + 292, + 545, + 345 + ], + "type": "image", + "image_path": "e0b40d58d35bd335423683da34282925bc08dc2157f1abffe2b3a017281bf8db.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 503, + 346, + 533, + 357 + ], + "lines": [ + { + "bbox": [ + 503, + 346, + 533, + 357 + ], + "spans": [ + { + "bbox": [ + 503, + 346, + 533, + 357 + ], + "type": "text", + "content": "(d) " + }, + { + "bbox": [ + 503, + 346, + 533, + 357 + ], + "type": "inline_equation", + "content": "l = 3" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 365, + 545, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 365, + 545, + 498 + ], + "spans": [ + { + "bbox": [ + 305, + 365, + 545, + 498 + ], + "type": "text", + "content": "Figure 4. The subdivision of icosahedron at different resolution " + }, + { + "bbox": [ + 305, + 365, + 545, + 498 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 305, + 365, + 545, + 498 + ], + "type": "text", + "content": ". icantly mitigate spherical distortion. Besides, [21, 31, 43] employ the polyhedrons, e.g., octahedron and icosahedron, to represent the panoramas and unfold these polyhedron representations on the plane for processing. In Elite360D, we select popular ICOSAP (See Fig. 4) as our input. Notably, we represent ICOSAP as discrete points, significantly reducing the computational cost while preserving spatial information and global awareness. Meanwhile, with the dense pixels in the ERP panoramas, discrete ICOSAP points can avoid semantic information redundancy." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 498, + 546, + 654 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 498, + 546, + 654 + ], + "spans": [ + { + "bbox": [ + 304, + 498, + 546, + 654 + ], + "type": "text", + "content": "Cross-Attention Mechanism It is widely used for efficient multi-modal feature fusion. For instance, Chen et al. [6] built cross-attention blocks to align and fuse 2D image features and 3D point cloud features for 3D object recognition. Similarly, BEVGuide [25], using BEV embedding as a guided query, employs a cross-attention block to fuse information across different sensors. We propose the B2F module to model the relationships between each pixel of the ERP feature and the whole ICOSAP point feature set, based on the semantic-aware affinity attention and distance-aware affinity attention simultaneously. With the B2F module, each ERP pixel feature can perceive the spatial information and semantic information of the large FoV scenes." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 661, + 455, + 673 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 661, + 455, + 673 + ], + "spans": [ + { + "bbox": [ + 306, + 661, + 455, + 673 + ], + "type": "text", + "content": "3. The Proposed Framework" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 305, + 677, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 677, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 677, + 547, + 713 + ], + "type": "text", + "content": "Overview. Our goal is to investigate an efficient and effective bi-projection fusion framework for 360 depth estimation. To this end, as illustrated in Fig. 3, the proposed" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 784, + 315, + 792 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 784, + 315, + 792 + ], + "spans": [ + { + "bbox": [ + 294, + 784, + 315, + 792 + ], + "type": "text", + "content": "9928" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 289, + 360 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 289, + 360 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 289, + 360 + ], + "type": "text", + "content": "Elite360D consists of three key components: image-based ERP feature extraction, point-based ICOSAP feature extraction and bi-projection bi-attention fusion (B2F). For the ERP feature " + }, + { + "bbox": [ + 47, + 72, + 289, + 360 + ], + "type": "inline_equation", + "content": "F^{E}" + }, + { + "bbox": [ + 47, + 72, + 289, + 360 + ], + "type": "text", + "content": " extraction, Elite360D flexibly accommodates a wide range of 2D models as encoder backbones and takes advantage of pre-training to reduce the over-fitting problems. For the ICOSAP feature extraction, we represent ICOSAP sphere as the point set, which significantly reduces the computation cost and preserves the spatial information and global perception. Then we directly employ the simple point encoder, i.e. Point transformer [47], to extract the ICOSAP point feature set " + }, + { + "bbox": [ + 47, + 72, + 289, + 360 + ], + "type": "inline_equation", + "content": "\\{F_{n}^{I}\\}_{n = 1,\\dots ,N}" + }, + { + "bbox": [ + 47, + 72, + 289, + 360 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 47, + 72, + 289, + 360 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 47, + 72, + 289, + 360 + ], + "type": "text", + "content": " is the point number, denoted as " + }, + { + "bbox": [ + 47, + 72, + 289, + 360 + ], + "type": "inline_equation", + "content": "\\{F_n^I\\}_N" + }, + { + "bbox": [ + 47, + 72, + 289, + 360 + ], + "type": "text", + "content": ". Next, to facilitate bi-projection feature fusion and learn a powerful representation " + }, + { + "bbox": [ + 47, + 72, + 289, + 360 + ], + "type": "inline_equation", + "content": "F^{GL}" + }, + { + "bbox": [ + 47, + 72, + 289, + 360 + ], + "type": "text", + "content": " from a local-with-global perspective, B2F module comprises semantic-aware affinity attention block, distance-aware affinity attention block, and gated fusion block. These two attention blocks model the semantic- and distance-aware dependencies between each ERP pixel-wise feature and entire ICOSAP feature set, respectively and the gated fusion block adaptively fuses the bi-attention outputs. Lastly, we employ a simple decoder with up-sampling and skip-connections to predict the final depth map from the fused representation " + }, + { + "bbox": [ + 47, + 72, + 289, + 360 + ], + "type": "inline_equation", + "content": "F^{GL}" + }, + { + "bbox": [ + 47, + 72, + 289, + 360 + ], + "type": "text", + "content": ". We now describe the details." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 365, + 185, + 376 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 365, + 185, + 376 + ], + "spans": [ + { + "bbox": [ + 47, + 365, + 185, + 376 + ], + "type": "text", + "content": "3.1. ERP Feature Extraction" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 380, + 289, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 380, + 289, + 559 + ], + "spans": [ + { + "bbox": [ + 46, + 380, + 289, + 559 + ], + "type": "text", + "content": "Taking an ERP image with the resolution size of " + }, + { + "bbox": [ + 46, + 380, + 289, + 559 + ], + "type": "inline_equation", + "content": "H \\times W" + }, + { + "bbox": [ + 46, + 380, + 289, + 559 + ], + "type": "text", + "content": " as the input, the encoder extracts feature map " + }, + { + "bbox": [ + 46, + 380, + 289, + 559 + ], + "type": "inline_equation", + "content": "F^{E} \\in \\mathbb{R}^{h \\times w \\times C}" + }, + { + "bbox": [ + 46, + 380, + 289, + 559 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 380, + 289, + 559 + ], + "type": "inline_equation", + "content": "h = H / s" + }, + { + "bbox": [ + 46, + 380, + 289, + 559 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 380, + 289, + 559 + ], + "type": "inline_equation", + "content": "w = W / s" + }, + { + "bbox": [ + 46, + 380, + 289, + 559 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 380, + 289, + 559 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 46, + 380, + 289, + 559 + ], + "type": "text", + "content": " is the down-sampling scale factor and " + }, + { + "bbox": [ + 46, + 380, + 289, + 559 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 46, + 380, + 289, + 559 + ], + "type": "text", + "content": " is the channel number. In particular, as treating ERP images as 2D perspective images, our encoder backbone is compatible with a wide range of robust 2D models that have been pre-trained on large-scale perspective image datasets [11], including CNNs and vision transformers, e.g., ResNet [14], EfficientNet [35], Swin transformer [24]. Notably, as repeated local operations cannot substitute for a global operator [40, 45], the finite size of convolutional kernels or the limited size of local attention windows results in ERP pixel-wise features lacking sufficient global receptive fields. Additionally, these 2D model performance suffers from the spherical distortions (See Tab. 1)." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 566, + 204, + 578 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 566, + 204, + 578 + ], + "spans": [ + { + "bbox": [ + 47, + 566, + 204, + 578 + ], + "type": "text", + "content": "3.2. ICOSAP Feature Extraction" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 581, + 287, + 701 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 581, + 287, + 701 + ], + "spans": [ + { + "bbox": [ + 46, + 581, + 287, + 701 + ], + "type": "text", + "content": "Preliminary. Since Lee et al. [21] found that polyhedron representations with a greater number of initial faces exhibit lower distortion, this paper focuses on the ICOSAP with the most initial faces. Especially, we introduce the ICOSAP to offer a comprehensive and high-quality global perception. For the resolution, ICOSAP data consists of " + }, + { + "bbox": [ + 46, + 581, + 287, + 701 + ], + "type": "inline_equation", + "content": "20 \\times 4^l" + }, + { + "bbox": [ + 46, + 581, + 287, + 701 + ], + "type": "text", + "content": " faces and " + }, + { + "bbox": [ + 46, + 581, + 287, + 701 + ], + "type": "inline_equation", + "content": "12 \\times 4^l" + }, + { + "bbox": [ + 46, + 581, + 287, + 701 + ], + "type": "text", + "content": " vertices at the subdivision level " + }, + { + "bbox": [ + 46, + 581, + 287, + 701 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 46, + 581, + 287, + 701 + ], + "type": "text", + "content": " (See Fig. 4). Given the spatial relationships between ICOSAP and ERP within the spherical space, RGB values of ICOSAP vertices can be derived from corresponding ERP pixels." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 702, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 702, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 702, + 288, + 714 + ], + "type": "text", + "content": "Feature extraction. Existing works focus on processing" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 318, + 71, + 533, + 180 + ], + "blocks": [ + { + "bbox": [ + 318, + 71, + 533, + 180 + ], + "lines": [ + { + "bbox": [ + 318, + 71, + 533, + 180 + ], + "spans": [ + { + "bbox": [ + 318, + 71, + 533, + 180 + ], + "type": "image", + "image_path": "7aa4770334ef310092503896260a80e21b73aac7e5baf5998c60ad04f2aeee9d.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 182, + 545, + 215 + ], + "lines": [ + { + "bbox": [ + 305, + 182, + 545, + 215 + ], + "spans": [ + { + "bbox": [ + 305, + 182, + 545, + 215 + ], + "type": "text", + "content": "Figure 5. The architecture of semantic-aware affinity attention. Especially, " + }, + { + "bbox": [ + 305, + 182, + 545, + 215 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}_{i,j}^{S}\\in \\mathbb{R}^{1\\times d}" + }, + { + "bbox": [ + 305, + 182, + 545, + 215 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 305, + 182, + 545, + 215 + ], + "type": "inline_equation", + "content": "\\mathbf{K}^S\\in \\mathbb{R}^{N\\times d}" + }, + { + "bbox": [ + 305, + 182, + 545, + 215 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 305, + 182, + 545, + 215 + ], + "type": "inline_equation", + "content": "\\mathbf{V}^S\\in \\mathbb{R}^{N\\times d}" + }, + { + "bbox": [ + 305, + 182, + 545, + 215 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 305, + 182, + 545, + 215 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 305, + 182, + 545, + 215 + ], + "type": "text", + "content": " is the dimension and " + }, + { + "bbox": [ + 305, + 182, + 545, + 215 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 305, + 182, + 545, + 215 + ], + "type": "text", + "content": " is the ICOSAP point number." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 220, + 547, + 558 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 220, + 547, + 558 + ], + "spans": [ + { + "bbox": [ + 304, + 220, + 547, + 558 + ], + "type": "text", + "content": "ICOSAP data on the plane, i.e., unfolded mesh representation [46] and spherical polyhedron representations [21]. However, these methods require specially designed and computationally intensive operations, such as convolutions, pooling [21], and up-sampling [31], to process the planar representations. By contrast, we propose to represent an ICOSAP input as a point set, as illustrated at the bottom of Fig. 3(a). ERP images provide dense pixel values and discrete ICOSAP point set can prevent redundancy in semantic information while preserving spatial information and global perception. Therefore, we take full advantage of two projections for efficient and accurate 360 depth estimation. In detail, we employ the central points of each face to represent the ICOSAP sphere. Firstly, we obtain the " + }, + { + "bbox": [ + 304, + 220, + 547, + 558 + ], + "type": "inline_equation", + "content": "20 \\times 4^l" + }, + { + "bbox": [ + 304, + 220, + 547, + 558 + ], + "type": "text", + "content": " faces of an ICOSAP sphere at the default subdivision level " + }, + { + "bbox": [ + 304, + 220, + 547, + 558 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 304, + 220, + 547, + 558 + ], + "type": "text", + "content": ", where each face is composed of three vertices. Then, we calculate the spatial coordinates and RGB values of each face center by averaging them of each three vertices, as each face is an equilateral triangle. As a result, we obtain the ICOSCAP point set " + }, + { + "bbox": [ + 304, + 220, + 547, + 558 + ], + "type": "inline_equation", + "content": "\\{P^{ICO}\\} \\in \\mathbb{R}^{(20 \\times 4^l) \\times 6}" + }, + { + "bbox": [ + 304, + 220, + 547, + 558 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 220, + 547, + 558 + ], + "type": "inline_equation", + "content": "20 \\times 4^l" + }, + { + "bbox": [ + 304, + 220, + 547, + 558 + ], + "type": "text", + "content": " is the point number and 6 represents the coordinates " + }, + { + "bbox": [ + 304, + 220, + 547, + 558 + ], + "type": "inline_equation", + "content": "[x, y, z]" + }, + { + "bbox": [ + 304, + 220, + 547, + 558 + ], + "type": "text", + "content": " and RGB channels. With the input point set " + }, + { + "bbox": [ + 304, + 220, + 547, + 558 + ], + "type": "inline_equation", + "content": "\\{P^{ICO}\\}" + }, + { + "bbox": [ + 304, + 220, + 547, + 558 + ], + "type": "text", + "content": ", we directly employ the encoder of Point Transformer [47] to extract ICOSAP point feature set " + }, + { + "bbox": [ + 304, + 220, + 547, + 558 + ], + "type": "inline_equation", + "content": "\\{F_n^I\\}_N \\in \\mathbb{R}^{N \\times C}" + }, + { + "bbox": [ + 304, + 220, + 547, + 558 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 220, + 547, + 558 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 220, + 547, + 558 + ], + "type": "text", + "content": " represents the number of point features and " + }, + { + "bbox": [ + 304, + 220, + 547, + 558 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 304, + 220, + 547, + 558 + ], + "type": "text", + "content": " corresponds to the same channel number as the ERP feature map " + }, + { + "bbox": [ + 304, + 220, + 547, + 558 + ], + "type": "inline_equation", + "content": "F^E" + }, + { + "bbox": [ + 304, + 220, + 547, + 558 + ], + "type": "text", + "content": ". Especially, for simplicity and efficiency, we do not opt for complex networks for this purpose." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 563, + 487, + 576 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 563, + 487, + 576 + ], + "spans": [ + { + "bbox": [ + 306, + 563, + 487, + 576 + ], + "type": "text", + "content": "3.3. Bi-Projection Bi-Attention Fusion" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 582, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 582, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 582, + 547, + 715 + ], + "type": "text", + "content": "In earlier methods [37] and [16], bi-projection feature fusion primarily depends on the geometric relationships between CP and ERP. They first apply the C2E function to re-project CP feature patches into ERP format feature maps and subsequently perform pixel-wise feature concatenation. While this fusion is effective, it brings several problems: 1) geometry-based re-projection and concatenation-based feature fusion significantly increase the computational costs (See Tab. 2); 2) the geometry-based fusion process from CP patches to ERP panorama restricts ERP pixels from the global perception, where each ERP pixel perceives scene in-" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 784, + 315, + 792 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 784, + 315, + 792 + ], + "spans": [ + { + "bbox": [ + 294, + 784, + 315, + 792 + ], + "type": "text", + "content": "9929" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 85, + 70, + 509, + 350 + ], + "blocks": [ + { + "bbox": [ + 85, + 70, + 509, + 350 + ], + "lines": [ + { + "bbox": [ + 85, + 70, + 509, + 350 + ], + "spans": [ + { + "bbox": [ + 85, + 70, + 509, + 350 + ], + "type": "table", + "html": "
DatasetsBackboneMethod#Params (M)#FLOPs (G)Abs Rel ↓Sq Rel ↓RMSE ↓δ1(%) ↑δ2(%) ↑δ3(%) ↑
M3D [5]ResNet-18 [14]Equi-Baseline14.3939.840.14280.12030.560780.9794.6898.26
Ours15.4345.910.12720.10700.527085.2895.2898.49
Δ+1.04+6.07-11.06%-9.73%-6.01%+4.31+0.60+0.23
ResNet-34 [14]Equi-Baseline24.5059.270.12550.10480.517385.5296.1098.55
Ours25.5465.290.11150.09140.487588.1596.4698.74
Δ+1.04+6.02-11.16%-12.79%-5.76%+2.63+0.36+0.19
ResNet-50* [14]Equi-Baseline29.6375.130.13700.12090.543282.5194.8398.07
Ours42.99170.110.11120.09800.487086.7096.0198.61
Δ+13.36+94.98-18.75%-18.94%-10.35%+4.19+1.18+0.54
EfficientNet-B5 [35]Equi-Baseline33.1018.810.10340.08310.463889.9296.8098.80
Ours34.1124.880.10480.08050.452489.9297.0799.14
Δ+1.01+6.07+1.35%-3.13%-2.46%-0.00+0.27+0.34
SWin-B* [24]Equi-Baseline90.75187.370.13010.12500.555784.0694.7697.72
Ours94.30211.280.12490.12100.546285.3495.0797.77
Δ+3.55+23.91-4.00%-3.20%-1.71%+1.28+0.31+0.05
DilateFormer-T [17]Equi-Baseline19.9346.960.14290.12740.574881.1594.4697.79
Ours20.9453.020.14230.12510.551782.6895.0098.14
Δ+1.01+6.06-0.42%-1.81%-4.02%+1.53+0.54+0.35
S2D3D [3]ResNet-34 [14]Equi-Baseline24.5059.220.12030.07540.372487.4196.5198.77
Ours25.5165.280.11820.07280.375688.7296.8498.92
Δ+1.01+6.06-9.21%-3.45%+0.86%+1.31+0.33+0.15
EfficientNet-B5 [35]Equi-Baseline33.1018.810.10260.06380.358089.4397.0699.16
Ours34.1124.880.10180.06030.357589.4797.2399.22
Δ+1.01+6.07-0.78%-5.49%-0.14%+0.03+0.17+0.06
Struct3D [48]ResNet-34 [14]Equi-Baseline24.5059.220.22560.39100.764176.9089.9694.49
Ours25.5165.280.14800.22150.496187.4194.3496.66
Δ+1.01+6.06-34.40%-43.35%-35.07%+10.51+4.38+2.17
EfficientNet-B5 [35]Equi-Baseline33.1018.810.13120.19380.431288.5395.2197.40
Ours34.1124.880.12770.19300.415189.1695.3397.43
Δ+1.01+6.07-2.67%-0.41%-3.73%+0.63+0.12+0.03
", + "image_path": "87abc69907542577fed6c570359ef5063eddeebd335922cfd3351ef9949c2dd0.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 354, + 545, + 378 + ], + "lines": [ + { + "bbox": [ + 47, + 354, + 545, + 378 + ], + "spans": [ + { + "bbox": [ + 47, + 354, + 545, + 378 + ], + "type": "text", + "content": "Table 1. Quantitative comparison with ERP-based depth baseline. For most conditions, we set the channel number " + }, + { + "bbox": [ + 47, + 354, + 545, + 378 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 47, + 354, + 545, + 378 + ], + "type": "text", + "content": " to 64. Especially, we set " + }, + { + "bbox": [ + 47, + 354, + 545, + 378 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 47, + 354, + 545, + 378 + ], + "type": "text", + "content": " to 256 for ResNet-50* and to 128 for Swin-B*. Bold indicates performance improvement. Red indicates performance decline." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 47, + 387, + 289, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 387, + 289, + 712 + ], + "spans": [ + { + "bbox": [ + 47, + 387, + 289, + 712 + ], + "type": "text", + "content": "formation only from its corresponding small-FoV CP patch without considering other patches; 3) this one-to-one alignment pattern only emphasizes spatial consistency, with no consideration for the semantic similarity. Thus, we design the Bi-Projection Bi-Attention Fusion (B2F) module to solve the above problems. As shown in Fig. 3(b), B2F module first leverages semantic-aware affinity attention and distance-aware affinity attention to model the semantic and spatial dependencies between each ERP pixel-wise feature " + }, + { + "bbox": [ + 47, + 387, + 289, + 712 + ], + "type": "inline_equation", + "content": "F_{i,j}^{E}" + }, + { + "bbox": [ + 47, + 387, + 289, + 712 + ], + "type": "text", + "content": " and ICOSAP point feature set " + }, + { + "bbox": [ + 47, + 387, + 289, + 712 + ], + "type": "inline_equation", + "content": "\\{F_n^I\\}_N" + }, + { + "bbox": [ + 47, + 387, + 289, + 712 + ], + "type": "text", + "content": ". Here " + }, + { + "bbox": [ + 47, + 387, + 289, + 712 + ], + "type": "inline_equation", + "content": "(i,j)" + }, + { + "bbox": [ + 47, + 387, + 289, + 712 + ], + "type": "text", + "content": " indicates the coordinate of a pixel in the ERP feature map, " + }, + { + "bbox": [ + 47, + 387, + 289, + 712 + ], + "type": "inline_equation", + "content": "i \\in (1,h), j \\in (1,w)" + }, + { + "bbox": [ + 47, + 387, + 289, + 712 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 47, + 387, + 289, + 712 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 47, + 387, + 289, + 712 + ], + "type": "text", + "content": " is the point number of the ICOSAP feature set. Consequently, a gated fusion block is employed to adaptively marry semantic- and distance-related information. Semantic-aware affinity attention. The semantic-aware affinity attention, as shown in Fig. 5, follows the standard dot-product attention [12]. In detail, given the extracted ERP feature map " + }, + { + "bbox": [ + 47, + 387, + 289, + 712 + ], + "type": "inline_equation", + "content": "F^{E} \\in \\mathbb{R}^{h \\times w \\times C}" + }, + { + "bbox": [ + 47, + 387, + 289, + 712 + ], + "type": "text", + "content": " and ICOSAP point feature set " + }, + { + "bbox": [ + 47, + 387, + 289, + 712 + ], + "type": "inline_equation", + "content": "\\{F_n^I\\}_N \\in \\mathbb{R}^{N \\times C}" + }, + { + "bbox": [ + 47, + 387, + 289, + 712 + ], + "type": "text", + "content": " (for simplicity, we denote it as " + }, + { + "bbox": [ + 47, + 387, + 289, + 712 + ], + "type": "inline_equation", + "content": "[F^I]" + }, + { + "bbox": [ + 47, + 387, + 289, + 712 + ], + "type": "text", + "content": "), we generate the query " + }, + { + "bbox": [ + 47, + 387, + 289, + 712 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}_{i,j}^{S}" + }, + { + "bbox": [ + 47, + 387, + 289, + 712 + ], + "type": "text", + "content": " from each ERP pixel-wise feature " + }, + { + "bbox": [ + 47, + 387, + 289, + 712 + ], + "type": "inline_equation", + "content": "F_{i,j}^{E} \\in \\mathbb{R}^{1 \\times C}" + }, + { + "bbox": [ + 47, + 387, + 289, + 712 + ], + "type": "text", + "content": " and produce the key " + }, + { + "bbox": [ + 47, + 387, + 289, + 712 + ], + "type": "inline_equation", + "content": "\\mathbf{K}^{S}" + }, + { + "bbox": [ + 47, + 387, + 289, + 712 + ], + "type": "text", + "content": " and value " + }, + { + "bbox": [ + 47, + 387, + 289, + 712 + ], + "type": "inline_equation", + "content": "\\mathbf{V}^{S}" + }, + { + "bbox": [ + 47, + 387, + 289, + 712 + ], + "type": "text", + "content": " from the whole ICOSAP point feature set " + }, + { + "bbox": [ + 47, + 387, + 289, + 712 + ], + "type": "inline_equation", + "content": "[F^{I}]" + }, + { + "bbox": [ + 47, + 387, + 289, + 712 + ], + "type": "text", + "content": ". Then, we calculate the attention weight " + }, + { + "bbox": [ + 47, + 387, + 289, + 712 + ], + "type": "inline_equation", + "content": "\\mathbf{A}_{i,j}^{S}" + }, + { + "bbox": [ + 47, + 387, + 289, + 712 + ], + "type": "text", + "content": " based on " + }, + { + "bbox": [ + 47, + 387, + 289, + 712 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}_{i,j}^{S}" + }, + { + "bbox": [ + 47, + 387, + 289, + 712 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 387, + 289, + 712 + ], + "type": "inline_equation", + "content": "\\mathbf{K}^{S}" + }, + { + "bbox": [ + 47, + 387, + 289, + 712 + ], + "type": "text", + "content": ", and obtain the affinity feature " + }, + { + "bbox": [ + 47, + 387, + 289, + 712 + ], + "type": "inline_equation", + "content": "F_{i,j}^{SA}" + }, + { + "bbox": [ + 47, + 387, + 289, + 712 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 47, + 387, + 289, + 712 + ], + "type": "inline_equation", + "content": "F_{i,j}^{SA} = \\mathbf{A}_{i,j}^{S} * \\mathbf{V}^{S}, \\mathbf{A}_{i,j}^{S} = \\text{softmax}\\left(\\frac{\\mathbf{Q}_{i,j}^{S} \\mathbf{K}^{ST}}{\\sqrt{d}}\\right)" + }, + { + "bbox": [ + 47, + 387, + 289, + 712 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 47, + 387, + 289, + 712 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 47, + 387, + 289, + 712 + ], + "type": "text", + "content": " is the channel number and" + } + ] + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 325, + 384, + 525, + 518 + ], + "blocks": [ + { + "bbox": [ + 325, + 384, + 525, + 518 + ], + "lines": [ + { + "bbox": [ + 325, + 384, + 525, + 518 + ], + "spans": [ + { + "bbox": [ + 325, + 384, + 525, + 518 + ], + "type": "image", + "image_path": "14496925d91d219f5dbb38074c0ad79bdf92be77a9e4c00a00e5b7d077ea33a8.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 522, + 545, + 547 + ], + "lines": [ + { + "bbox": [ + 305, + 522, + 545, + 547 + ], + "spans": [ + { + "bbox": [ + 305, + 522, + 545, + 547 + ], + "type": "text", + "content": "Figure 6. The architecture of distance-aware affinity attention, especially, " + }, + { + "bbox": [ + 305, + 522, + 545, + 547 + ], + "type": "inline_equation", + "content": "Q_{i,j}^{D} \\in R^{1 \\times d}" + }, + { + "bbox": [ + 305, + 522, + 545, + 547 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 305, + 522, + 545, + 547 + ], + "type": "inline_equation", + "content": "K^{D} \\in \\mathbb{R}^{N \\times d}" + }, + { + "bbox": [ + 305, + 522, + 545, + 547 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 305, + 522, + 545, + 547 + ], + "type": "inline_equation", + "content": "\\mathbf{V}^{D} \\in \\mathbb{R}^{N \\times d}" + }, + { + "bbox": [ + 305, + 522, + 545, + 547 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "bbox": [ + 304, + 555, + 545, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 555, + 545, + 628 + ], + "spans": [ + { + "bbox": [ + 304, + 555, + 545, + 628 + ], + "type": "inline_equation", + "content": "d = C" + }, + { + "bbox": [ + 304, + 555, + 545, + 628 + ], + "type": "text", + "content": ". After querying all the ERP pixel-wise features, we can obtain the ERP format feature map " + }, + { + "bbox": [ + 304, + 555, + 545, + 628 + ], + "type": "inline_equation", + "content": "F^{SA}" + }, + { + "bbox": [ + 304, + 555, + 545, + 628 + ], + "type": "text", + "content": " with the dimension of " + }, + { + "bbox": [ + 304, + 555, + 545, + 628 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^{h\\times w\\times d}" + }, + { + "bbox": [ + 304, + 555, + 545, + 628 + ], + "type": "text", + "content": ". The attention weight " + }, + { + "bbox": [ + 304, + 555, + 545, + 628 + ], + "type": "inline_equation", + "content": "A_{i,j}^{S}" + }, + { + "bbox": [ + 304, + 555, + 545, + 628 + ], + "type": "text", + "content": " captures affinities based on the semantic-aware feature similarities, and the output " + }, + { + "bbox": [ + 304, + 555, + 545, + 628 + ], + "type": "inline_equation", + "content": "F^{SA}" + }, + { + "bbox": [ + 304, + 555, + 545, + 628 + ], + "type": "text", + "content": " effectively integrates the global and local receptive fields." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 630, + 546, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 630, + 546, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 630, + 546, + 715 + ], + "type": "text", + "content": "Distance-aware affinity attention. Distance-aware affinity attention captures the differences between ERP and ICOSAP in both spatial and semantic information. This can better utilize the geometric prior knowledge of panoramicas and precisely measure the distances between two different projection features. As depicted in Fig. 6, the distance-aware affinity attention is built upon the subtraction-based" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 784, + 315, + 792 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 784, + 315, + 792 + ], + "spans": [ + { + "bbox": [ + 295, + 784, + 315, + 792 + ], + "type": "text", + "content": "9930" + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 85, + 70, + 509, + 294 + ], + "blocks": [ + { + "bbox": [ + 85, + 70, + 509, + 294 + ], + "lines": [ + { + "bbox": [ + 85, + 70, + 509, + 294 + ], + "spans": [ + { + "bbox": [ + 85, + 70, + 509, + 294 + ], + "type": "table", + "html": "
DatasetsBackboneMethodPub'Year#Params (M)#FLOPs (G)Abs Rel ↓Sq Rel ↓RMSE ↓δ1(%) ↑δ2(%) ↑δ3(%) ↑
M3D [5]TransformerEGFormer [45]ICCV'2315.3966.210.14730.15170.602581.5893.9097.35
PanoFormer [32]ECCV'2220.3881.090.10510.09660.492989.0896.2398.31
ResNet-18 [14]BiFuse [37]CVPR'2035.80165.660.13600.12020.548883.2795.1298.10
UniFuse [16]RAL'2130.2662.600.11910.10300.515886.0495.8498.30
Omnifusion [41]CVPR'2232.3598.680.12090.10900.505586.5895.8198.36
HRDFuse† [2]CVPR'2326.0950.590.14140.12410.550781.4894.8998.20
Ours-15.4345.910.12720.10700.527085.2895.2898.49
ResNet-34 [14]BiFuse [37]CVPR'2056.01199.580.11260.09920.502788.0096.1398.47
BiFuse++ [38]TPAMI'2252.4987.480.11230.09150.485388.1296.5698.69
UniFuse [16]RAL'2150.4896.520.11440.09360.483587.8596.5998.73
Omnifusion [41]CVPR'2242.46142.290.11610.10070.493187.7296.1598.44
HRDFuse† [2]CVPR'2346.3180.870.11720.09710.502586.7496.1798.49
Ours-25.5465.290.11150.09140.487588.1596.4698.74
ResNet-50* [14]BiFuse [37]CVPR'20253.08775.240.11790.09810.497086.7496.2798.66
UniFuse [16]RAL'21131.30222.300.11850.09840.502486.6696.1898.50
Ours-42.99170.110.11120.09800.487086.7096.0198.61
S2D3D [3]TransformerEGFormer [45]ICCV'2315.3966.210.15280.14080.497481.8593.3897.36
PanoFormer [32]ECCV'2220.3881.090.11220.07860.394588.7495.8498.59
ResNet-34 [14]Omnifusion [41]CVPR'2242.46142.290.11540.07750.380986.7496.0398.71
UniFuse [16]RAL'2150.4896.520.11240.07090.355587.0697.0498.99
Ours-25.5165.280.11820.07280.375688.7296.8498.92
Struct3D [48]TransformerEGFormer [45]ICCV'2315.3966.210.22050.45090.684179.7990.7194.55
PanoFormer [32]ECCV'2220.3881.090.25490.49490.793774.7089.1593.97
ResNet-34 [14]BiFuse [37]CVPR'2056.01199.580.15730.24550.521385.9194.0096.72
UniFuse [16]RAL'2150.4896.520.15060.23190.501685.4293.9996.76
Ours-25.5165.280.14800.22150.496187.4194.3496.66
", + "image_path": "2722c8d233fe17e6a2b7161296b07e606b3c7b7b0358cbec27865f5f80df9bf3.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 295, + 544, + 319 + ], + "lines": [ + { + "bbox": [ + 46, + 295, + 544, + 319 + ], + "spans": [ + { + "bbox": [ + 46, + 295, + 544, + 319 + ], + "type": "text", + "content": "Table 2. Quantitative comparison with the SOTA methods. " + }, + { + "bbox": [ + 46, + 295, + 544, + 319 + ], + "type": "inline_equation", + "content": "{}^{ \\dagger }" + }, + { + "bbox": [ + 46, + 295, + 544, + 319 + ], + "type": "text", + "content": " means that we modify the HRDFuse network structure for a fair comparison. Green represents the best performance under the given encoder backbone." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 47, + 324, + 287, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 324, + 287, + 389 + ], + "spans": [ + { + "bbox": [ + 47, + 324, + 287, + 389 + ], + "type": "text", + "content": "cross-attention mechanism [39]. Given the ERP pixel-wise feature " + }, + { + "bbox": [ + 47, + 324, + 287, + 389 + ], + "type": "inline_equation", + "content": "F_{i,j}^{E}" + }, + { + "bbox": [ + 47, + 324, + 287, + 389 + ], + "type": "text", + "content": " and ICOSAP point feature set " + }, + { + "bbox": [ + 47, + 324, + 287, + 389 + ], + "type": "inline_equation", + "content": "[\\pmb{F}^{\\pmb{I}}]" + }, + { + "bbox": [ + 47, + 324, + 287, + 389 + ], + "type": "text", + "content": ", we firstly calculate the spatial distance embedding " + }, + { + "bbox": [ + 47, + 324, + 287, + 389 + ], + "type": "inline_equation", + "content": "Dis^{SP}" + }, + { + "bbox": [ + 47, + 324, + 287, + 389 + ], + "type": "text", + "content": " from the spatial coordinates of ERP pixel and ICOSAP point set, i.e. " + }, + { + "bbox": [ + 47, + 324, + 287, + 389 + ], + "type": "inline_equation", + "content": "[x_{i,j}^{e},y_{i,j}^{e},z_{i,j}^{e}]" + }, + { + "bbox": [ + 47, + 324, + 287, + 389 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 324, + 287, + 389 + ], + "type": "inline_equation", + "content": "\\{[x_n^{ico},y_n^{ico},z_n^{ico}]\\}_{N}" + }, + { + "bbox": [ + 47, + 324, + 287, + 389 + ], + "type": "text", + "content": ", as:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 78, + 392, + 287, + 409 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 78, + 392, + 287, + 409 + ], + "spans": [ + { + "bbox": [ + 78, + 392, + 287, + 409 + ], + "type": "interline_equation", + "content": "D i s _ {i, j} ^ {S P} = \\left[ e ^ {- \\Delta x _ {i, j}}, e ^ {- \\Delta y _ {i, j}}, e ^ {- \\Delta z _ {i, j}} \\right] \\mathbf {W} ^ {S P}, \\tag {1}", + "image_path": "e6337dc3c508ade6d2ec7f5e29994578a7e0a8e8ca8ae23d68d3efeb7c52cb29.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 416, + 287, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 416, + 287, + 518 + ], + "spans": [ + { + "bbox": [ + 47, + 416, + 287, + 518 + ], + "type": "text", + "content": "where linear projection " + }, + { + "bbox": [ + 47, + 416, + 287, + 518 + ], + "type": "inline_equation", + "content": "\\mathbf{W}^{SP}\\in \\mathbb{R}^{3\\times d}" + }, + { + "bbox": [ + 47, + 416, + 287, + 518 + ], + "type": "inline_equation", + "content": "Dis_{i,j}^{SP}\\in \\mathbb{R}^{1\\times N\\times d}" + }, + { + "bbox": [ + 47, + 416, + 287, + 518 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 416, + 287, + 518 + ], + "type": "inline_equation", + "content": "[\\Delta x_{i,j},\\Delta y_{i,j},\\Delta z_{i,j}]\\in \\mathbb{R}^{1\\times N\\times 3}" + }, + { + "bbox": [ + 47, + 416, + 287, + 518 + ], + "type": "text", + "content": " is the distances between " + }, + { + "bbox": [ + 47, + 416, + 287, + 518 + ], + "type": "inline_equation", + "content": "[x_{i,j}^e,y_{i,j}^e,z_{i,j}^e ]" + }, + { + "bbox": [ + 47, + 416, + 287, + 518 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 416, + 287, + 518 + ], + "type": "inline_equation", + "content": "\\{[x_n^{ico},y_n^{ico},z_n^{ico}]\\}_{N}" + }, + { + "bbox": [ + 47, + 416, + 287, + 518 + ], + "type": "text", + "content": ". In particular, the operation " + }, + { + "bbox": [ + 47, + 416, + 287, + 518 + ], + "type": "inline_equation", + "content": "e^{-(\\cdot)}" + }, + { + "bbox": [ + 47, + 416, + 287, + 518 + ], + "type": "text", + "content": " allows " + }, + { + "bbox": [ + 47, + 416, + 287, + 518 + ], + "type": "inline_equation", + "content": "Dis^{SP}" + }, + { + "bbox": [ + 47, + 416, + 287, + 518 + ], + "type": "text", + "content": " to pay more attention to the close parts between ERP pixels and ICOSAP point set. After that, we produce the query " + }, + { + "bbox": [ + 47, + 416, + 287, + 518 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}_{i,j}^{D}" + }, + { + "bbox": [ + 47, + 416, + 287, + 518 + ], + "type": "text", + "content": " and key " + }, + { + "bbox": [ + 47, + 416, + 287, + 518 + ], + "type": "inline_equation", + "content": "\\mathbf{K}^D" + }, + { + "bbox": [ + 47, + 416, + 287, + 518 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 47, + 416, + 287, + 518 + ], + "type": "inline_equation", + "content": "F_{i,j}^{E}" + }, + { + "bbox": [ + 47, + 416, + 287, + 518 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 416, + 287, + 518 + ], + "type": "inline_equation", + "content": "[\\pmb {F}^{\\pmb{I}}]" + }, + { + "bbox": [ + 47, + 416, + 287, + 518 + ], + "type": "text", + "content": ", respectively and calculate the semantic distance embedding " + }, + { + "bbox": [ + 47, + 416, + 287, + 518 + ], + "type": "inline_equation", + "content": "Dis_{i,j}^{SE}" + }, + { + "bbox": [ + 47, + 416, + 287, + 518 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 115, + 520, + 287, + 538 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 115, + 520, + 287, + 538 + ], + "spans": [ + { + "bbox": [ + 115, + 520, + 287, + 538 + ], + "type": "interline_equation", + "content": "D i s _ {i, j} ^ {S E} = e ^ {- \\left\\| \\mathbf {Q} _ {i, j} ^ {D} - \\mathbf {K} ^ {D} \\right\\|}, \\tag {2}", + "image_path": "ed2867998618550dfd75178e2cf71ac96809acfdaaa1e9a7251889346a89170e.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 539, + 287, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 539, + 287, + 617 + ], + "spans": [ + { + "bbox": [ + 47, + 539, + 287, + 617 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 539, + 287, + 617 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_Q^D" + }, + { + "bbox": [ + 47, + 539, + 287, + 617 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 47, + 539, + 287, + 617 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_K^D \\in \\mathbb{R}^{C \\times d}" + }, + { + "bbox": [ + 47, + 539, + 287, + 617 + ], + "type": "text", + "content": " are linear projections, " + }, + { + "bbox": [ + 47, + 539, + 287, + 617 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}_{i,j}^D \\in \\mathbb{R}^{1 \\times d}" + }, + { + "bbox": [ + 47, + 539, + 287, + 617 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 47, + 539, + 287, + 617 + ], + "type": "inline_equation", + "content": "\\mathbf{K}^D \\in \\mathbb{R}^{N \\times d}" + }, + { + "bbox": [ + 47, + 539, + 287, + 617 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 47, + 539, + 287, + 617 + ], + "type": "inline_equation", + "content": "Dis_{i,j}^{SE} \\in \\mathbb{R}^{1 \\times N \\times d}" + }, + { + "bbox": [ + 47, + 539, + 287, + 617 + ], + "type": "text", + "content": ". Lastly, the distance-aware attention weight " + }, + { + "bbox": [ + 47, + 539, + 287, + 617 + ], + "type": "inline_equation", + "content": "\\mathbf{A}_{i,j}^D" + }, + { + "bbox": [ + 47, + 539, + 287, + 617 + ], + "type": "text", + "content": " is generated with spatial and semantic distance embeddings, and the distance-aware affinity feature vector " + }, + { + "bbox": [ + 47, + 539, + 287, + 617 + ], + "type": "inline_equation", + "content": "F_{i,j}^{DA}" + }, + { + "bbox": [ + 47, + 539, + 287, + 617 + ], + "type": "text", + "content": " is obtained from the attention weight " + }, + { + "bbox": [ + 47, + 539, + 287, + 617 + ], + "type": "inline_equation", + "content": "\\mathbf{A}_{i,j}^D" + }, + { + "bbox": [ + 47, + 539, + 287, + 617 + ], + "type": "text", + "content": " and the value " + }, + { + "bbox": [ + 47, + 539, + 287, + 617 + ], + "type": "inline_equation", + "content": "\\mathbf{V}^D" + }, + { + "bbox": [ + 47, + 539, + 287, + 617 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 91, + 619, + 287, + 644 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 619, + 287, + 644 + ], + "spans": [ + { + "bbox": [ + 91, + 619, + 287, + 644 + ], + "type": "interline_equation", + "content": "\\mathbf {A} _ {i, j} ^ {D} = \\operatorname {s o f t m a x} \\left(\\frac {\\sum \\left(D i s _ {i , j} ^ {S P} + D i s _ {i , j} ^ {S E}\\right)}{\\sqrt {d}}\\right), \\tag {3}", + "image_path": "935f3216fd896f0124a53c222f59fc3a0797c32076388af0a5eebdb19351ee52.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 100, + 650, + 287, + 663 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 100, + 650, + 287, + 663 + ], + "spans": [ + { + "bbox": [ + 100, + 650, + 287, + 663 + ], + "type": "interline_equation", + "content": "\\mathbf {V} ^ {D} = F ^ {I} \\mathbf {W} _ {V} ^ {D}, F _ {i, j} ^ {D A} = \\mathbf {A} _ {i, j} ^ {D} * \\mathbf {V} ^ {D}, \\tag {4}", + "image_path": "df5e60ee4e76217ac7fe504896a09625e5372fac0c1dd2ab75f0d7c94524066a.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 665, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 665, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 665, + 287, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 665, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\sum" + }, + { + "bbox": [ + 47, + 665, + 287, + 713 + ], + "type": "text", + "content": " means the sum for the channel dimension, i.e. " + }, + { + "bbox": [ + 47, + 665, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\sum (Dis_{i,j}^{SP} + Dis_{i,j}^{SE})\\in \\mathbb{R}^{1\\times N}" + }, + { + "bbox": [ + 47, + 665, + 287, + 713 + ], + "type": "text", + "content": ". After querying all ERP pixel-wise features, we obtain the distance-aware aggregated feature " + }, + { + "bbox": [ + 47, + 665, + 287, + 713 + ], + "type": "inline_equation", + "content": "F^{DA}\\in \\mathbb{R}^{h\\times w\\times d}" + }, + { + "bbox": [ + 47, + 665, + 287, + 713 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 47, + 665, + 287, + 713 + ], + "type": "inline_equation", + "content": "d = C" + }, + { + "bbox": [ + 47, + 665, + 287, + 713 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 324, + 545, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 324, + 545, + 380 + ], + "spans": [ + { + "bbox": [ + 305, + 324, + 545, + 380 + ], + "type": "text", + "content": "Gated fusion. Since direct average or concatenation may compromise the original representation ability, inspired by [9], we propose the gated fusion block to adaptively fuse " + }, + { + "bbox": [ + 305, + 324, + 545, + 380 + ], + "type": "inline_equation", + "content": "F^{SA}" + }, + { + "bbox": [ + 305, + 324, + 545, + 380 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 324, + 545, + 380 + ], + "type": "inline_equation", + "content": "F^{DA}" + }, + { + "bbox": [ + 305, + 324, + 545, + 380 + ], + "type": "text", + "content": " and obtain the representations " + }, + { + "bbox": [ + 305, + 324, + 545, + 380 + ], + "type": "inline_equation", + "content": "F^{GL}" + }, + { + "bbox": [ + 305, + 324, + 545, + 380 + ], + "type": "text", + "content": " from a local-with-global perspective, formulated as:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 358, + 381, + 545, + 396 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 358, + 381, + 545, + 396 + ], + "spans": [ + { + "bbox": [ + 358, + 381, + 545, + 396 + ], + "type": "interline_equation", + "content": "F ^ {G L} = g ^ {S A} * F ^ {S A} + g ^ {D A} * F ^ {D A}, \\tag {5}", + "image_path": "0383dd2a9f789fb1c429db2ba11168b17ad09fd38e6e44cc01363d12d06a1121.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 360, + 398, + 491, + 412 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 360, + 398, + 491, + 412 + ], + "spans": [ + { + "bbox": [ + 360, + 398, + 491, + 412 + ], + "type": "interline_equation", + "content": "g ^ {S A} = \\sigma_ {S A} \\left(W _ {g} ^ {S A} \\cdot \\left[ F ^ {S A}; F ^ {D A} \\right]\\right),", + "image_path": "e9f46377684d8be63ad11b1fe375ec282c04530e58b7c0262876c86fd0aa4ffe.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 359, + 414, + 492, + 428 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 359, + 414, + 492, + 428 + ], + "spans": [ + { + "bbox": [ + 359, + 414, + 492, + 428 + ], + "type": "interline_equation", + "content": "g ^ {D A} = \\sigma_ {D A} \\left(W _ {g} ^ {D A} \\cdot \\left[ F ^ {S A}; F ^ {D A} \\right]\\right),", + "image_path": "7a5365bbc7a47194a56e39c9687fdffe0ede6c95f16f9a46b0656ec4270cdcde.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 434, + 545, + 459 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 434, + 545, + 459 + ], + "spans": [ + { + "bbox": [ + 306, + 434, + 545, + 459 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 306, + 434, + 545, + 459 + ], + "type": "inline_equation", + "content": "W_{g}^{SA}" + }, + { + "bbox": [ + 306, + 434, + 545, + 459 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 306, + 434, + 545, + 459 + ], + "type": "inline_equation", + "content": "W_{g}^{DA}" + }, + { + "bbox": [ + 306, + 434, + 545, + 459 + ], + "type": "text", + "content": " are linear projections, " + }, + { + "bbox": [ + 306, + 434, + 545, + 459 + ], + "type": "inline_equation", + "content": "\\sigma_{SA}(\\cdot)" + }, + { + "bbox": [ + 306, + 434, + 545, + 459 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 306, + 434, + 545, + 459 + ], + "type": "inline_equation", + "content": "\\sigma_{DA}(\\cdot)" + }, + { + "bbox": [ + 306, + 434, + 545, + 459 + ], + "type": "text", + "content": " are the sigmoid functions." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 466, + 391, + 479 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 466, + 391, + 479 + ], + "spans": [ + { + "bbox": [ + 306, + 466, + 391, + 479 + ], + "type": "text", + "content": "3.4. Optimization" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 483, + 545, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 483, + 545, + 568 + ], + "spans": [ + { + "bbox": [ + 304, + 483, + 545, + 568 + ], + "type": "text", + "content": "With the fused feature " + }, + { + "bbox": [ + 304, + 483, + 545, + 568 + ], + "type": "inline_equation", + "content": "F^{GL}" + }, + { + "bbox": [ + 304, + 483, + 545, + 568 + ], + "type": "text", + "content": " and multi-scale ERP feature maps in the ERP encoder backbone, we feed them into a decoder [30] with several up-sampling blocks and skipconnections to output the final depth. For the depth supervision, we follow existing works [28, 32] and employ the combination of Berhu loss [20] and gradient loss [28]. (Details of training loss can be found in the suppl. material.)." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 574, + 388, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 574, + 388, + 586 + ], + "spans": [ + { + "bbox": [ + 306, + 574, + 388, + 586 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 590, + 545, + 603 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 590, + 545, + 603 + ], + "spans": [ + { + "bbox": [ + 304, + 590, + 545, + 603 + ], + "type": "text", + "content": "4.1. Datasets, Metrics, and Implementation Details" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": "Datasets and Metrics. We evaluate Elite360D on three datasets: two real-world datasets, Matterport3D(M3D) [5], Stanford2D3D(S2D3D) [3], and a recently proposed large-scale synthetic dataset, Structure3D(Struct3D) [48]. For the evaluation metrics, we follow previous works [16, 32, 45] to employ some standard metrics, including absolute relative error (Abs Rel), squared relative error (Sq Rel), root mean squared error (RMSE), and three threshold percentage " + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\delta < \\alpha^t" + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\alpha = 1.25, t = 1, 2, 3" + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": "), denoted as " + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\delta_t" + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": ". Addi" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "9931" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 66, + 69, + 525, + 344 + ], + "blocks": [ + { + "bbox": [ + 66, + 69, + 525, + 344 + ], + "lines": [ + { + "bbox": [ + 66, + 69, + 525, + 344 + ], + "spans": [ + { + "bbox": [ + 66, + 69, + 525, + 344 + ], + "type": "image", + "image_path": "9853f330a79315c7421f1e6ce11598740b021fe41cd8c5ee4da8eefa63278e18.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 47, + 346, + 545, + 357 + ], + "lines": [ + { + "bbox": [ + 47, + 346, + 545, + 357 + ], + "spans": [ + { + "bbox": [ + 47, + 346, + 545, + 357 + ], + "type": "text", + "content": "Figure 7. Qualitative results (with ResNet-34 as the backbone) on Matterport3D (top), Stanford2D3D (middle) and Structure3D (bottom)." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 69, + 363, + 264, + 437 + ], + "blocks": [ + { + "bbox": [ + 69, + 363, + 264, + 437 + ], + "lines": [ + { + "bbox": [ + 69, + 363, + 264, + 437 + ], + "spans": [ + { + "bbox": [ + 69, + 363, + 264, + 437 + ], + "type": "table", + "html": "
BackbonePre-trainedAbs Rel↓Sq Rel↓RMSE↓δ1↑
ResNet-34X0.15960.14520.585681.36
✓0.11150.09140.487588.15
EfficientNet-B5X0.12110.10870.513186.80
✓0.10480.08050.452489.92
DilateFormerX0.15150.14150.569480.95
✓0.14230.12510.551782.68
", + "image_path": "48f90647e65920518b98c294cb20cecb049f087a37b6857ea43f042d92ba1be3.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 50, + 438, + 282, + 449 + ], + "lines": [ + { + "bbox": [ + 50, + 438, + 282, + 449 + ], + "spans": [ + { + "bbox": [ + 50, + 438, + 282, + 449 + ], + "type": "text", + "content": "Table 3. The impact of pre-training on ERP encoder backbones." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 456, + 286, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 456, + 286, + 479 + ], + "spans": [ + { + "bbox": [ + 46, + 456, + 286, + 479 + ], + "type": "text", + "content": "tionally, we measure the number of parameters and FLOPS to evaluate the efficiency of our method." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 45, + 480, + 287, + 671 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 480, + 287, + 671 + ], + "spans": [ + { + "bbox": [ + 45, + 480, + 287, + 671 + ], + "type": "text", + "content": "Training Details. We use diverse ERP encoder backbones, including CNNs (ResNet-18, 34, 50 [14], EfficientNet B5 [35]), and transformers (Swin-B [24], DilateFormer-T [17]). All backbones are pre-trained on ImageNet1K [11]. We set the default channel number " + }, + { + "bbox": [ + 45, + 480, + 287, + 671 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 45, + 480, + 287, + 671 + ], + "type": "text", + "content": " to 64 and default subdivision level of ICOSAP as " + }, + { + "bbox": [ + 45, + 480, + 287, + 671 + ], + "type": "inline_equation", + "content": "l = 4" + }, + { + "bbox": [ + 45, + 480, + 287, + 671 + ], + "type": "text", + "content": ". For the ICOSAP encoder, we employ the one of Point transformer [47] with three down-sample blocks. Following [16], we use Adam optimizer [18] and a constant learning rate of " + }, + { + "bbox": [ + 45, + 480, + 287, + 671 + ], + "type": "inline_equation", + "content": "1\\mathrm{e}^{-4}" + }, + { + "bbox": [ + 45, + 480, + 287, + 671 + ], + "type": "text", + "content": ". Considering the unfair comparisons stemming from variations in hyper-parameters and validation procedures used across different methods, we re-train the existing methods from scratch and validate them, following the unified training and validation settings [16]. (Due to page limit, detailed training and validation settings can be found in suppl. mat.)." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 673, + 190, + 686 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 673, + 190, + 686 + ], + "spans": [ + { + "bbox": [ + 46, + 673, + 190, + 686 + ], + "type": "text", + "content": "4.2. Performance Comparison" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 689, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 689, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 689, + 287, + 713 + ], + "type": "text", + "content": "Comparisons with ERP-based depth baselines. As shown in Tab. 1, with an increase of only " + }, + { + "bbox": [ + 46, + 689, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\sim 1\\mathrm{M}" + }, + { + "bbox": [ + 46, + 689, + 287, + 713 + ], + "type": "text", + "content": " parameters" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 365, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 365, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 365, + 545, + 713 + ], + "type": "inline_equation", + "content": "(C = 64)" + }, + { + "bbox": [ + 304, + 365, + 545, + 713 + ], + "type": "text", + "content": ", our Elite360D demonstrates substantial advancements over the ERP-based baselines across different ERP encoder backbones on all three datasets. Specifically, for the Matterport3D dataset, Elite360D achieves reductions exceeding " + }, + { + "bbox": [ + 304, + 365, + 545, + 713 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 304, + 365, + 545, + 713 + ], + "type": "text", + "content": " in Abs Rel error (ResNet-18, 34), along with reductions of " + }, + { + "bbox": [ + 304, + 365, + 545, + 713 + ], + "type": "inline_equation", + "content": "4.00\\%" + }, + { + "bbox": [ + 304, + 365, + 545, + 713 + ], + "type": "text", + "content": " in Abs Rel error (Swin-B) and " + }, + { + "bbox": [ + 304, + 365, + 545, + 713 + ], + "type": "inline_equation", + "content": "4.02\\%" + }, + { + "bbox": [ + 304, + 365, + 545, + 713 + ], + "type": "text", + "content": " in RMSE error (DilateFormer-T). Besides, with the larger channel number " + }, + { + "bbox": [ + 304, + 365, + 545, + 713 + ], + "type": "inline_equation", + "content": "C = 256" + }, + { + "bbox": [ + 304, + 365, + 545, + 713 + ], + "type": "text", + "content": " (ResNet-50), Elite360D outperforms ERP baseline by " + }, + { + "bbox": [ + 304, + 365, + 545, + 713 + ], + "type": "inline_equation", + "content": "18.75\\%" + }, + { + "bbox": [ + 304, + 365, + 545, + 713 + ], + "type": "text", + "content": " (Abl Rel), " + }, + { + "bbox": [ + 304, + 365, + 545, + 713 + ], + "type": "inline_equation", + "content": "18.94\\%" + }, + { + "bbox": [ + 304, + 365, + 545, + 713 + ], + "type": "text", + "content": " (Sq Rel). For the small-scale S2D3D dataset, Elite360D outperforms ERP baseline by " + }, + { + "bbox": [ + 304, + 365, + 545, + 713 + ], + "type": "inline_equation", + "content": "9.21\\%" + }, + { + "bbox": [ + 304, + 365, + 545, + 713 + ], + "type": "text", + "content": " in Abs Rel error and " + }, + { + "bbox": [ + 304, + 365, + 545, + 713 + ], + "type": "inline_equation", + "content": "1.31\\%" + }, + { + "bbox": [ + 304, + 365, + 545, + 713 + ], + "type": "text", + "content": " in accuracy " + }, + { + "bbox": [ + 304, + 365, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\delta_{1}" + }, + { + "bbox": [ + 304, + 365, + 545, + 713 + ], + "type": "text", + "content": " (ResNet-34), as well as " + }, + { + "bbox": [ + 304, + 365, + 545, + 713 + ], + "type": "inline_equation", + "content": "5.49\\%" + }, + { + "bbox": [ + 304, + 365, + 545, + 713 + ], + "type": "text", + "content": " in Sq Rel error (EfficientNet-B5). Remarkably, on the larger-scale Structure3D, Elite360 performs favorably against the baseline by a significant margin, especially with ResNet-34. Comparisons with prevalent methods. In Tab. 2, we conduct a comprehensive comparison with prevalent supervised methods. From the results, we can observe that our approach achieves similar or even superior performance compared to existing both bi-projection fusion methods and single input methods at a significantly lower cost, particularly on two large-scale datasets, Matterport3D and Structure3D. Specifically, for the Matterport3D dataset, our Elite360D with ResNet-34 outperforms UniFuse by " + }, + { + "bbox": [ + 304, + 365, + 545, + 713 + ], + "type": "inline_equation", + "content": "2.53\\%" + }, + { + "bbox": [ + 304, + 365, + 545, + 713 + ], + "type": "text", + "content": " (Abs Rel) and with ResNet-50 outperforms BiFuse by " + }, + { + "bbox": [ + 304, + 365, + 545, + 713 + ], + "type": "inline_equation", + "content": "2.01\\%" + }, + { + "bbox": [ + 304, + 365, + 545, + 713 + ], + "type": "text", + "content": " (RMSE). For the Structure3D dataset, our Elite360D with ResNet-34 outperforms UniFuse by " + }, + { + "bbox": [ + 304, + 365, + 545, + 713 + ], + "type": "inline_equation", + "content": "4.48\\%" + }, + { + "bbox": [ + 304, + 365, + 545, + 713 + ], + "type": "text", + "content": " (Seq Rel), " + }, + { + "bbox": [ + 304, + 365, + 545, + 713 + ], + "type": "inline_equation", + "content": "1.99\\%" + }, + { + "bbox": [ + 304, + 365, + 545, + 713 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 304, + 365, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\delta_{1}" + }, + { + "bbox": [ + 304, + 365, + 545, + 713 + ], + "type": "text", + "content": "). For performance on the Stanford2D3D dataset, we suspect it might be related to the ICOSAP point encoder." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "9932" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 72, + 71, + 261, + 123 + ], + "blocks": [ + { + "bbox": [ + 72, + 71, + 261, + 123 + ], + "lines": [ + { + "bbox": [ + 72, + 71, + 261, + 123 + ], + "spans": [ + { + "bbox": [ + 72, + 71, + 261, + 123 + ], + "type": "table", + "html": "
Bi-projection feature fusionAbs Rel ↓Sq Rel ↓RMSE ↓δ1 ↑
SFA [2] + Add0.12760.10020.515084.27
SFA [2] + Concat0.11910.10190.514386.52
Only SA0.12040.10140.512186.26
Only DA0.11840.09720.494487.06
Our B2F (SA + DA)0.11150.09140.487588.15
", + "image_path": "4d3823bbf76775462d399272a6df7a31810226282a3da49a9d7a64af39743eec.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 72, + 143, + 261, + 201 + ], + "blocks": [ + { + "bbox": [ + 84, + 125, + 250, + 134 + ], + "lines": [ + { + "bbox": [ + 84, + 125, + 250, + 134 + ], + "spans": [ + { + "bbox": [ + 84, + 125, + 250, + 134 + ], + "type": "text", + "content": "Table 4. The ablation results for B2F module." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 72, + 143, + 261, + 201 + ], + "lines": [ + { + "bbox": [ + 72, + 143, + 261, + 201 + ], + "spans": [ + { + "bbox": [ + 72, + 143, + 261, + 201 + ], + "type": "table", + "html": "
Final fusionAbs Rel ↓Sq Rel ↓RMSE ↓δ1 ↑
Add0.16850.14810.580974.60
Average0.11980.09180.489386.65
Concatenation0.11450.09370.488087.66
Adaptive fusion [2]0.12440.09680.489186.08
Our gated fusion0.11150.09140.487588.15
", + "image_path": "1f17d042b37e505a705e320f88f52adc1ae8601c54035a0b62faf115f6e1f004.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 221, + 287, + 317 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 221, + 287, + 317 + ], + "spans": [ + { + "bbox": [ + 46, + 221, + 287, + 317 + ], + "type": "text", + "content": "The limited data of Stanford2D3D dataset restricts the ability of the transformer-based point encoder to provide accurate global perception. Moreover, in Fig. 7, we present the qualitative comparisons. Our Elite360D can predict more accurate depth values based on the local-with-global perception capabilities (e.g., flowers, shelves and doors). Additional qualitative results and inference time comparisons can be found in the suppl. material." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 324, + 205, + 337 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 324, + 205, + 337 + ], + "spans": [ + { + "bbox": [ + 47, + 324, + 205, + 337 + ], + "type": "text", + "content": "4.3. Ablation Study and Analyses" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 340, + 287, + 364 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 340, + 287, + 364 + ], + "spans": [ + { + "bbox": [ + 46, + 340, + 287, + 364 + ], + "type": "text", + "content": "Most of ablation experiments are conducted on the Mateport3D test dataset with ResNet34 as the backbone." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 365, + 287, + 520 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 365, + 287, + 520 + ], + "spans": [ + { + "bbox": [ + 46, + 365, + 287, + 520 + ], + "type": "text", + "content": "The Effect of pre-training. We verify the effectiveness of ImageNet [11] pre-training with different encoder backbones. As observed from Tab. 3, the pre-training results in a significant improvement for all encoder backbones, e.g., " + }, + { + "bbox": [ + 46, + 365, + 287, + 520 + ], + "type": "inline_equation", + "content": "6.79\\%" + }, + { + "bbox": [ + 46, + 365, + 287, + 520 + ], + "type": "text", + "content": " improvement in accuracy " + }, + { + "bbox": [ + 46, + 365, + 287, + 520 + ], + "type": "inline_equation", + "content": "\\delta_{1}" + }, + { + "bbox": [ + 46, + 365, + 287, + 520 + ], + "type": "text", + "content": " (ResNet-34). Notably, pre-training has a relatively small impact on DilateFormer. Combined with the results in Tab. 1, the explanation of this phenomenon is that the default input resolution in pretrained models is different from actual input, thereby impacting the resolution-related position embeddings. In general, pre-training based on large-scale perspective images can effectively enhance the performance of models based on " + }, + { + "bbox": [ + 46, + 365, + 287, + 520 + ], + "type": "inline_equation", + "content": "360^{\\circ}" + }, + { + "bbox": [ + 46, + 365, + 287, + 520 + ], + "type": "text", + "content": " images and reduce the risk of overfitting." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 521, + 287, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 521, + 287, + 688 + ], + "spans": [ + { + "bbox": [ + 46, + 521, + 287, + 688 + ], + "type": "text", + "content": "The effectiveness of B2F module. In Tab. 4, we compare four available bi-projection feature fusion modules. To align the spatial dimensions between ICOSAP point feature set and ERP feature map, we introduce SFA module from [2]. After that, we employ direct addition and concatenation to aggregate these two projections. We also achieve the bi-projection feature fusion with semantic-aware affinity attention (SA) alone and distance-aware affinity attention (DA) alone. Compared to the methods based solely on semantic-aware feature similarities (The first three rows), single distance-aware affinity attention can achieve better performance, which indicates that the spatial positional relationships boost the bi-projection feature fusion. Overall, our B2F module achieves the best performance." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "content": "The superiority of ICOSAP. As only CP/TP's patch centers lie on the sphere's surface, we extract the feature em" + } + ] + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 308, + 71, + 545, + 118 + ], + "blocks": [ + { + "bbox": [ + 59, + 203, + 273, + 213 + ], + "lines": [ + { + "bbox": [ + 59, + 203, + 273, + 213 + ], + "spans": [ + { + "bbox": [ + 59, + 203, + 273, + 213 + ], + "type": "text", + "content": "Table 5. The ablation results for the fusion of B2F module." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 308, + 71, + 545, + 118 + ], + "lines": [ + { + "bbox": [ + 308, + 71, + 545, + 118 + ], + "spans": [ + { + "bbox": [ + 308, + 71, + 545, + 118 + ], + "type": "table", + "html": "
Method#Param(M)#FLOPs(G)Abs Rel ↓RMSE ↓δ1 ↑
ERP-CP25.6654.150.13690.540183.69
ERP-TP (N=18)25.6650.580.13280.538583.87
ERP-ICOSAP (Ours)15.4345.910.12720.527085.28
", + "image_path": "ac2cdbf031fdf200100fe4e62f215ef275779ac7a42479d3100c118439eb33c3.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 308, + 136, + 545, + 176 + ], + "blocks": [ + { + "bbox": [ + 306, + 122, + 544, + 133 + ], + "lines": [ + { + "bbox": [ + 306, + 122, + 544, + 133 + ], + "spans": [ + { + "bbox": [ + 306, + 122, + 544, + 133 + ], + "type": "text", + "content": "Table 6. The comparison of different projections on Matterport3D." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 308, + 136, + 545, + 176 + ], + "lines": [ + { + "bbox": [ + 308, + 136, + 545, + 176 + ], + "spans": [ + { + "bbox": [ + 308, + 136, + 545, + 176 + ], + "type": "table", + "html": "
N of {F^I}#Params (M)#FLOPs (G)Abs Rel↓Sq Rel↓RMSE↓δ1↑
2027.4166.290.11570.09950.502487.12
8025.5465.290.11150.09140.487588.15
32024.9864.320.11530.09430.490587.85
", + "image_path": "564183ad451fc53238d9afe250a8e56c71311b2198b2ad657615a96c8f46fc70.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_body" + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 178, + 544, + 200 + ], + "lines": [ + { + "bbox": [ + 306, + 178, + 544, + 200 + ], + "spans": [ + { + "bbox": [ + 306, + 178, + 544, + 200 + ], + "type": "text", + "content": "Table 7. Impact of the ICOSAP point-wise feature number " + }, + { + "bbox": [ + 306, + 178, + 544, + 200 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 306, + 178, + 544, + 200 + ], + "type": "text", + "content": ". Larger " + }, + { + "bbox": [ + 306, + 178, + 544, + 200 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 306, + 178, + 544, + 200 + ], + "type": "text", + "content": ", fewer down-sampling blocks in the point encoder." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 204, + 545, + 276 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 204, + 545, + 276 + ], + "spans": [ + { + "bbox": [ + 304, + 204, + 545, + 276 + ], + "type": "text", + "content": "bedding from each CP/TP patch and employs the patch center coordinates and feature embedding as the input of B2F module. In Tab. 6, we show the results with ResNet18 backbone. Our Elite360D, utilizing the ICOSAP point set, marginally outperforms models with CP and TP patches, while exhibiting fewer parameters and FLOPs." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 276, + 545, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 276, + 545, + 335 + ], + "spans": [ + { + "bbox": [ + 304, + 276, + 545, + 335 + ], + "type": "text", + "content": "The effectiveness of gated fusion. We conduct an ablation study for the gated fusion block, outlined in Tab. 5. With the feature maps " + }, + { + "bbox": [ + 304, + 276, + 545, + 335 + ], + "type": "inline_equation", + "content": "F^{SA}" + }, + { + "bbox": [ + 304, + 276, + 545, + 335 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 276, + 545, + 335 + ], + "type": "inline_equation", + "content": "F^{DA}" + }, + { + "bbox": [ + 304, + 276, + 545, + 335 + ], + "type": "text", + "content": ", We compare it with the direct addition, average fusion, concatenation, and the adaptive fusion in [2]. The gated fusion performs best." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 336, + 545, + 432 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 336, + 545, + 432 + ], + "spans": [ + { + "bbox": [ + 304, + 336, + 545, + 432 + ], + "type": "text", + "content": "ICOSAP point feature number " + }, + { + "bbox": [ + 304, + 336, + 545, + 432 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 336, + 545, + 432 + ], + "type": "text", + "content": ". We study the effect of the ICOSAP feature point number (See Tab. 7). Too few points " + }, + { + "bbox": [ + 304, + 336, + 545, + 432 + ], + "type": "inline_equation", + "content": "(N = 20)" + }, + { + "bbox": [ + 304, + 336, + 545, + 432 + ], + "type": "text", + "content": " lead to the over-concentrated global contextual information resulting from excessive down-sampling blocks, while too many points " + }, + { + "bbox": [ + 304, + 336, + 545, + 432 + ], + "type": "inline_equation", + "content": "(N = 320)" + }, + { + "bbox": [ + 304, + 336, + 545, + 432 + ], + "type": "text", + "content": " lead to underconcentrated condition, resulting in insufficient perception of ERP pixel features. Best performance can be observed when " + }, + { + "bbox": [ + 304, + 336, + 545, + 432 + ], + "type": "inline_equation", + "content": "N = 80" + }, + { + "bbox": [ + 304, + 336, + 545, + 432 + ], + "type": "text", + "content": " and we used " + }, + { + "bbox": [ + 304, + 336, + 545, + 432 + ], + "type": "inline_equation", + "content": "N = 80" + }, + { + "bbox": [ + 304, + 336, + 545, + 432 + ], + "type": "text", + "content": " as default in this paper." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 437, + 470, + 449 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 437, + 470, + 449 + ], + "spans": [ + { + "bbox": [ + 306, + 437, + 470, + 449 + ], + "type": "text", + "content": "5. Conclusion and Future Work" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 452, + 545, + 643 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 452, + 545, + 643 + ], + "spans": [ + { + "bbox": [ + 304, + 452, + 545, + 643 + ], + "type": "text", + "content": "In this paper, we proposed a novel bi-projection fusion solution for efficient 360 depth estimation. To address the limited local receptive field of ERP pixel-wise features and avoid expensive bi-projection fusion modules, we proposed a compact yet effective B2F module to learn the representations with local-with-global perceptions from ERP and ICOSAP. With an increase of 1M parameters, we significantly improved the performance of the ERP-based depth estimation baseline. Remarkably, our approach achieved performance on par with complex state-of-the-art methods. Future Work: From the experimental results, we observed that ERP-based depth baseline, with pre-trained Efficient-Net backbone, even outperforms existing specifically designed methods. Therefore, in the future, we will explore how to fully leverage different projections and successful perspective models for " + }, + { + "bbox": [ + 304, + 452, + 545, + 643 + ], + "type": "inline_equation", + "content": "360^{\\circ}" + }, + { + "bbox": [ + 304, + 452, + 545, + 643 + ], + "type": "text", + "content": " community." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 306, + 650, + 404, + 663 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 650, + 404, + 663 + ], + "spans": [ + { + "bbox": [ + 306, + 650, + 404, + 663 + ], + "type": "text", + "content": "Acknowledgement" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "content": "This paper is supported by the National Natural Science Foundation of China (NSF) under Grant No. NSFC22FYT45 and the Guangzhou City, University and Enterprise Joint Fund under Grant No.SL2022A03J01278." + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "9933" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 133 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 133 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 133 + ], + "type": "text", + "content": "[1] Hao Ai, Zidong Cao, Jin Zhu, Haotian Bai, Yucheng Chen, and Ling Wang. Deep learning for omnidirectional vision: A survey and new perspectives. ArXiv, abs/2205.10468, 2022. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 137, + 288, + 201 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 137, + 288, + 201 + ], + "spans": [ + { + "bbox": [ + 53, + 137, + 288, + 201 + ], + "type": "text", + "content": "[2] Hao Ai, Zidong Cao, Yan-Pei Cao, Ying Shan, and Lin Wang. Hrdfuse: Monocular " + }, + { + "bbox": [ + 53, + 137, + 288, + 201 + ], + "type": "inline_equation", + "content": "360^{\\circ}" + }, + { + "bbox": [ + 53, + 137, + 288, + 201 + ], + "type": "text", + "content": " depth estimation by collaboratively learning holistic-with-regional depth distributions. 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13273-13282, 2023. 2, 3, 6, 8" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 204, + 288, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 204, + 288, + 237 + ], + "spans": [ + { + "bbox": [ + 53, + 204, + 288, + 237 + ], + "type": "text", + "content": "[3] Iro Armeni, Sasha Sax, Amir Roshan Zamir, and Silvio Savarese. Joint 2d-3d-semantic data for indoor scene understanding. CoRR, abs/1702.01105, 2017. 2, 5, 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 239, + 287, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 239, + 287, + 293 + ], + "spans": [ + { + "bbox": [ + 53, + 239, + 287, + 293 + ], + "type": "text", + "content": "[4] Zidong Cao, Hao Ai, Yan Cao, Ying Shan, Xiaohu Qie, and Lin Wang. Omnizoomer: Learning to move and zoom in on sphere at high-resolution. 2023 IEEE/CVF International Conference on Computer Vision (ICCV), pages 12851-12861, 2023. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 295, + 287, + 350 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 295, + 287, + 350 + ], + "spans": [ + { + "bbox": [ + 53, + 295, + 287, + 350 + ], + "type": "text", + "content": "[5] Angel X. Chang, Angela Dai, Thomas A. Funkhouser, Maciej Halber, Matthias Nießner, Manolis Savva, Shuran Song, Andy Zeng, and Yinda Zhang. Matterport3d: Learning from RGB-D data in indoor environments. In 3DV, pages 667-676. IEEE Computer Society, 2017. 1, 2, 5, 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 352, + 287, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 352, + 287, + 407 + ], + "spans": [ + { + "bbox": [ + 53, + 352, + 287, + 407 + ], + "type": "text", + "content": "[6] Zehui Chen, Zhenyu Li, Shiquan Zhang, Liangji Fang, Qinghong Jiang, Feng Zhao, Bolei Zhou, and Hang Zhao. Autoalign: Pixel-instance feature aggregation for multimodal 3d object detection. In International Joint Conference on Artificial Intelligence, 2022. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 408, + 287, + 463 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 408, + 287, + 463 + ], + "spans": [ + { + "bbox": [ + 53, + 408, + 287, + 463 + ], + "type": "text", + "content": "[7] Hsien-Tzu Cheng, Chun-Hung Chao, Jin-Dong Dong, Hao-Kai Wen, Tyng-Luh Liu, and Min Sun. Cube padding for weakly-supervised saliency prediction in " + }, + { + "bbox": [ + 53, + 408, + 287, + 463 + ], + "type": "inline_equation", + "content": "360^{\\circ}" + }, + { + "bbox": [ + 53, + 408, + 287, + 463 + ], + "type": "text", + "content": " videos. 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1420-1429, 2018. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 465, + 287, + 509 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 465, + 287, + 509 + ], + "spans": [ + { + "bbox": [ + 53, + 465, + 287, + 509 + ], + "type": "text", + "content": "[8] Xinjing Cheng, Peng Wang, Yanqi Zhou, Chenye Guan, and Ruigang Yang. Omnidirectional depth extension networks. In 2020 IEEE International Conference on Robotics and Automation (ICRA), pages 589-595, 2020. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 510, + 287, + 565 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 510, + 287, + 565 + ], + "spans": [ + { + "bbox": [ + 53, + 510, + 287, + 565 + ], + "type": "text", + "content": "[9] Yanhua Cheng, Rui Cai, Zhiwei Li, Xin Zhao, and Kaiqi Huang. Locality-sensitive deconvolution networks with gated fusion for rgb-d indoor semantic segmentation. 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 1475-1483, 2017. 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 567, + 287, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 567, + 287, + 610 + ], + "spans": [ + { + "bbox": [ + 48, + 567, + 287, + 610 + ], + "type": "text", + "content": "[10] Taco Cohen, Maurice Weiler, Berkay Kicanaoglu, and Max Welling. Gauge equivariant convolutional networks and the icosahedral cnn. In International Conference on Machine Learning, 2019. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 613, + 287, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 613, + 287, + 656 + ], + "spans": [ + { + "bbox": [ + 48, + 613, + 287, + 656 + ], + "type": "text", + "content": "[11] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. computer vision and pattern recognition, 2009. 2, 4, 7, 8" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 658, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 287, + 713 + ], + "type": "text", + "content": "[12] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "text", + "content": "scale. In International Conference on Learning Representations, 2021. 3, 5" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 308, + 95, + 545, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 95, + 545, + 139 + ], + "spans": [ + { + "bbox": [ + 308, + 95, + 545, + 139 + ], + "type": "text", + "content": "[13] Marc Eder, Mykhailo Shvets, John Lim, and Jan-Michael Frahm. Tangent images for mitigating spherical distortion. 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12423-12431, 2019. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 140, + 545, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 140, + 545, + 183 + ], + "spans": [ + { + "bbox": [ + 308, + 140, + 545, + 183 + ], + "type": "text", + "content": "[14] Kaiming He, X. Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 770-778, 2015. 2, 4, 5, 6, 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 184, + 545, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 184, + 545, + 239 + ], + "spans": [ + { + "bbox": [ + 308, + 184, + 545, + 239 + ], + "type": "text", + "content": "[15] ChiyuMax Jiang, Jingwei Huang, Karthik Kashinath, Prabhat Prabhat, Philip Marcus, and Matthias Niessner. Spherical cnns on unstructured grids. International Conference on Learning Representations, International Conference on Learning Representations, 2019. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 239, + 545, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 239, + 545, + 282 + ], + "spans": [ + { + "bbox": [ + 308, + 239, + 545, + 282 + ], + "type": "text", + "content": "[16] Hualie Jiang, Zhe Sheng, Siyu Zhu, Zilong Dong, and Rui Huang. Unifuse: Unidirectional fusion for " + }, + { + "bbox": [ + 308, + 239, + 545, + 282 + ], + "type": "inline_equation", + "content": "360^{\\circ}" + }, + { + "bbox": [ + 308, + 239, + 545, + 282 + ], + "type": "text", + "content": " panorama depth estimation. IEEE Robotics and Automation Letters, 6: 1519-1526, 2021. 2, 3, 4, 6, 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 284, + 545, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 284, + 545, + 327 + ], + "spans": [ + { + "bbox": [ + 308, + 284, + 545, + 327 + ], + "type": "text", + "content": "[17] Jiayu Jiao, Yu-Ming Tang, Kun-Yu Lin, Yipeng Gao, Jinhua Ma, Yaowei Wang, and Wei-Shi Zheng. Dilateformer: Multi-scale dilated transformer for visual recognition. IEEE Transactions on Multimedia, pages 1-14, 2023. 5, 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 327, + 545, + 349 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 327, + 545, + 349 + ], + "spans": [ + { + "bbox": [ + 308, + 327, + 545, + 349 + ], + "type": "text", + "content": "[18] Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. CoRR, abs/1412.6980, 2014. 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 350, + 545, + 414 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 350, + 545, + 414 + ], + "spans": [ + { + "bbox": [ + 308, + 350, + 545, + 414 + ], + "type": "text", + "content": "[19] Varun Ravi Kumar, Senthil Kumar Yogamani, Hazem Rashed, Ganesh Sitsu, Christian Witt, Isabelle Leang, Stefan Milz, and Patrick Mäder. Omnidet: Surround view cameras based multi-task visual perception network for autonomous driving. IEEE Robotics and Automation Letters, 6:2830-2837, 2021. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 415, + 545, + 458 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 415, + 545, + 458 + ], + "spans": [ + { + "bbox": [ + 308, + 415, + 545, + 458 + ], + "type": "text", + "content": "[20] Iro Laina, C. Rupprecht, Vasileios Belagiannis, Federico Tombari, and Nassir Navab. Deeper depth prediction with fully convolutional residual networks. 3DV 2016, pages 239-248, 2016. 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 460, + 545, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 460, + 545, + 514 + ], + "spans": [ + { + "bbox": [ + 308, + 460, + 545, + 514 + ], + "type": "text", + "content": "[21] Yeonkun Lee, Jaeseok Jeong, Jong Seob Yun, Wonjune Cho, and Kuk jin Yoon. Spherephd: Applying cnns on a spherical polyhedron representation of " + }, + { + "bbox": [ + 308, + 460, + 545, + 514 + ], + "type": "inline_equation", + "content": "360^{\\circ}" + }, + { + "bbox": [ + 308, + 460, + 545, + 514 + ], + "type": "text", + "content": " images. 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9173-9181, 2018. 2, 3, 4" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 515, + 545, + 558 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 515, + 545, + 558 + ], + "spans": [ + { + "bbox": [ + 308, + 515, + 545, + 558 + ], + "type": "text", + "content": "[22] Meng Li, Senbo Wang, Weihao Yuan, Weichao Shen, Zhe Sheng, and Zilong Dong. " + }, + { + "bbox": [ + 308, + 515, + 545, + 558 + ], + "type": "inline_equation", + "content": "S^2" + }, + { + "bbox": [ + 308, + 515, + 545, + 558 + ], + "type": "text", + "content": " net: Accurate panorama depth estimation on spherical surface. IEEE Robotics and Automation Letters, 8:1053-1060, 2023. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 559, + 545, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 559, + 545, + 613 + ], + "spans": [ + { + "bbox": [ + 308, + 559, + 545, + 613 + ], + "type": "text", + "content": "[23] Yunhao Li, Wei Shen, Zhongpai Gao, Yucheng Zhu, Guangtao Zhai, and Guodong Guo. Looking here or there? gaze following in 360-degree images. 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 3722-3731, 2021. 1" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 614, + 545, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 614, + 545, + 669 + ], + "spans": [ + { + "bbox": [ + 308, + 614, + 545, + 669 + ], + "type": "text", + "content": "[24] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 9992-10002, 2021. 2, 3, 4, 5, 7" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 670, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 670, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 670, + 545, + 713 + ], + "type": "text", + "content": "[25] Yunze Man, Liangyan Gui, and Yu-Xiong Wang. Bev-guided multi-modality fusion for driving perception. 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 21960-21969, 2023. 3" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "9934" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 715 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "text", + "content": "[26] Daniel Martin, Ana Serrano, Alexander W. Bergman, Gordon Wetzstein, and Belén Masiá. Scangan360: A generative model of realistic scanpaths for " + }, + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "inline_equation", + "content": "360^{\\circ}" + }, + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "text", + "content": " images. IEEE Transactions on Visualization and Computer Graphics, 28:2003-2013, 2021. 1" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 128, + 288, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 128, + 288, + 172 + ], + "spans": [ + { + "bbox": [ + 48, + 128, + 288, + 172 + ], + "type": "text", + "content": "[27] Rafael Monroy, Sebastian Lutz, Tejo Chalasani, and Aljoscha Smolic. Salnet360: Saliency maps for omnidirectional images with cnn. Signal Process. Image Commun., 69:26-34, 2017. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 172, + 288, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 172, + 288, + 228 + ], + "spans": [ + { + "bbox": [ + 48, + 172, + 288, + 228 + ], + "type": "text", + "content": "[28] Giovanni Pintore, Eva Almansa, and Jens Schneider. Slicenet: deep dense depth estimation from a single indoor panorama using a slice-based representation. 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11531-11540, 2021. 1, 3, 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 228, + 288, + 272 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 228, + 288, + 272 + ], + "spans": [ + { + "bbox": [ + 48, + 228, + 288, + 272 + ], + "type": "text", + "content": "[29] Manuel Rey-Area, Mingze Yuan, and Christian Richardt. 360monodepth: High-resolution " + }, + { + "bbox": [ + 48, + 228, + 288, + 272 + ], + "type": "inline_equation", + "content": "360^{\\circ}" + }, + { + "bbox": [ + 48, + 228, + 288, + 272 + ], + "type": "text", + "content": " monocular depth estimation. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3752-3762, 2022. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 272, + 288, + 305 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 272, + 288, + 305 + ], + "spans": [ + { + "bbox": [ + 48, + 272, + 288, + 305 + ], + "type": "text", + "content": "[30] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In MICCAI (3), pages 234-241. Springer, 2015. 3, 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 305, + 288, + 350 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 305, + 288, + 350 + ], + "spans": [ + { + "bbox": [ + 48, + 305, + 288, + 350 + ], + "type": "text", + "content": "[31] Mehran Shakerinava and Siamak Ravanbakhsh. Equivariant networks for pixelized spheres. Proceedings of the 38th International Conference on Machine Learning, ICML, abs/2106.06662, 2021. 3, 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 350, + 288, + 393 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 350, + 288, + 393 + ], + "spans": [ + { + "bbox": [ + 48, + 350, + 288, + 393 + ], + "type": "text", + "content": "[32] Zhijie Shen, Chunyu Lin, Kang Liao, Lang Nie, Zishuo Zheng, and Yao Zhao. Panofformer: Panorama transformer for indoor " + }, + { + "bbox": [ + 48, + 350, + 288, + 393 + ], + "type": "inline_equation", + "content": "360^{\\circ}" + }, + { + "bbox": [ + 48, + 350, + 288, + 393 + ], + "type": "text", + "content": " depth estimation. In European Conference on Computer Vision, 2022. 2, 3, 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 393, + 288, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 393, + 288, + 437 + ], + "spans": [ + { + "bbox": [ + 48, + 393, + 288, + 437 + ], + "type": "text", + "content": "[33] Yu-Chuan Su and Kristen Grauman. Kernel transformer networks for compact spherical convolution. 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9434-9443, 2018. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 437, + 288, + 482 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 437, + 288, + 482 + ], + "spans": [ + { + "bbox": [ + 48, + 437, + 288, + 482 + ], + "type": "text", + "content": "[34] Cheng Sun, Min Sun, and Hwann-Tzong Chen. Hohonet: 360 indoor holistic understanding with latent horizontal features. 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2573-2582, 2020. 1, 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 481, + 288, + 515 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 481, + 288, + 515 + ], + "spans": [ + { + "bbox": [ + 48, + 481, + 288, + 515 + ], + "type": "text", + "content": "[35] Mingxing Tan and Quoc V. Le. Efficientnet: Rethinking model scaling for convolutional neural networks. In ICML, pages 6105-6114. PMLR, 2019. 4, 5, 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 515, + 288, + 559 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 515, + 288, + 559 + ], + "spans": [ + { + "bbox": [ + 48, + 515, + 288, + 559 + ], + "type": "text", + "content": "[36] Keisuke Tateno, Nassir Navab, and Federico Tombari. Distortion-aware convolutional filters for dense prediction in panoramic images. In European Conference on Computer Vision, 2018. 1, 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 559, + 288, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 559, + 288, + 613 + ], + "spans": [ + { + "bbox": [ + 48, + 559, + 288, + 613 + ], + "type": "text", + "content": "[37] Fu-En Wang, Yu-Hsuan Yeh, Min Sun, Wei-Chen Chiu, and Yi-Hsuan Tsai. Bifuse: Monocular 360 depth estimation via bi-projection fusion. 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 459–468, 2020. 2, 3, 4, 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 614, + 288, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 614, + 288, + 669 + ], + "spans": [ + { + "bbox": [ + 48, + 614, + 288, + 669 + ], + "type": "text", + "content": "[38] Fu-En Wang, Yu-Hsuan Yeh, Yi-Hsuan Tsai, Wei-Chen Chiu, and Min Sun. Bifuse++: Self-supervised and efficient bi-projection fusion for " + }, + { + "bbox": [ + 48, + 614, + 288, + 669 + ], + "type": "inline_equation", + "content": "360^{\\circ}" + }, + { + "bbox": [ + 48, + 614, + 288, + 669 + ], + "type": "text", + "content": " depth estimation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45: 5448-5460, 2022. 2, 3, 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 669, + 288, + 715 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 669, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 48, + 669, + 288, + 715 + ], + "type": "text", + "content": "[39] Sirui Wang, Di Liang, Jian Song, Yuntao Li, and Wei Wu. DABERT: dual attention enhanced BERT for semantic matching. In COLING, pages 1645-1654. International Committee on Computational Linguistics, 2022. 6" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 72, + 546, + 698 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 307, + 72, + 546, + 118 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 72, + 546, + 118 + ], + "spans": [ + { + "bbox": [ + 307, + 72, + 546, + 118 + ], + "type": "text", + "content": "[40] X. Wang, Ross B. Girshick, Abhinav Kumar Gupta, and Kaiming He. Non-local neural networks. 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7794-7803, 2017. 4" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 118, + 546, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 546, + 173 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 546, + 173 + ], + "type": "text", + "content": "[41] Yu yang Li, Yuliang Guo, Zhixin Yan, Xinyu Huang, Ye Duan, and Liu Ren. Omnifusion: 360 monocular depth estimation via geometry-aware fusion. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2791-2800, 2022. 2, 3, 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 173, + 546, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 173, + 546, + 239 + ], + "spans": [ + { + "bbox": [ + 307, + 173, + 546, + 239 + ], + "type": "text", + "content": "[42] Xiaoqing Ye, Mao Shu, Hanyu Li, Yifeng Shi, Yingying Li, Guan-Sheng Wang, Xiao Tan, and Errui Ding. Rope3d: The roadside perception dataset for autonomous driving and monocular 3d object detection task. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 21309-21318, 2022. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 239, + 546, + 306 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 239, + 546, + 306 + ], + "spans": [ + { + "bbox": [ + 307, + 239, + 546, + 306 + ], + "type": "text", + "content": "[43] Youngho Yoon, Inchul Chung, Lin Wang, and Kuk-Jin Yoon. Spheresr: " + }, + { + "bbox": [ + 307, + 239, + 546, + 306 + ], + "type": "inline_equation", + "content": "360^{\\circ}" + }, + { + "bbox": [ + 307, + 239, + 546, + 306 + ], + "type": "text", + "content": " image super-resolution with arbitrary projection via continuous spherical image representation. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5667-5676, 2021. 1, 2, 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 306, + 546, + 362 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 306, + 546, + 362 + ], + "spans": [ + { + "bbox": [ + 307, + 306, + 546, + 362 + ], + "type": "text", + "content": "[44] Haozheng Yu, Lu He, Bing Jian, Weiwei Feng, and Shanghua Liu. Panelnet: Understanding 360 indoor environment via panel representation. 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 878-887, 2023. 1, 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 362, + 546, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 362, + 546, + 407 + ], + "spans": [ + { + "bbox": [ + 307, + 362, + 546, + 407 + ], + "type": "text", + "content": "[45] Ilwi Yun, Chan-Yong Shin, Hyunku Lee, Hyuk-Jae Lee, and Chae-Eun Rhee. Egformer: Equirectangular geometry-biased transformer for 360 depth estimation. ArXiv, abs/2304.07803, 2023. 2, 3, 4, 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 407, + 546, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 407, + 546, + 453 + ], + "spans": [ + { + "bbox": [ + 307, + 407, + 546, + 453 + ], + "type": "text", + "content": "[46] Chao Zhang, Stephan Liwicki, William Smith, and Roberto Cipolla. Orientation-aware semantic segmentation on icosahedron spheres. 2019 IEEE/CVF International Conference on Computer Vision (ICCV), pages 3532-3540, 2019. 2, 4" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 453, + 546, + 496 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 453, + 546, + 496 + ], + "spans": [ + { + "bbox": [ + 307, + 453, + 546, + 496 + ], + "type": "text", + "content": "[47] Hengshuang Zhao, Li Jiang, Jiaya Jia, Philip H. S. Torr, and Vladlen Koltun. Point transformer. 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 16239-16248, 2020. 4, 7" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 496, + 546, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 496, + 546, + 542 + ], + "spans": [ + { + "bbox": [ + 307, + 496, + 546, + 542 + ], + "type": "text", + "content": "[48] Jia Zheng, Junfei Zhang, Jing Li, Rui Tang, Shenghua Gao, and Zihan Zhou. Structured3d: A large photo-realistic dataset for structured 3d modeling. In European Conference on Computer Vision, 2019. 5, 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 541, + 546, + 608 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 541, + 546, + 608 + ], + "spans": [ + { + "bbox": [ + 307, + 541, + 546, + 608 + ], + "type": "text", + "content": "[49] Xu Zheng, Jinjing Zhu, Ye-Peng Liu, Zidong Cao, Chong Fu, and Lin Wang. Both style and distortion matter: Dualpath unsupervised domain adaptation for panoramic semantic segmentation. 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1285-1295, 2023. 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 608, + 546, + 653 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 608, + 546, + 653 + ], + "spans": [ + { + "bbox": [ + 307, + 608, + 546, + 653 + ], + "type": "text", + "content": "[50] Chuanqing Zhuang, Zhengda Lu, Yiqun Wang, Jun Xiao, and Ying Wang. Acdnet: Adaptively combined dilated convolution for monocular panorama depth estimation. CoRR, abs/2112.14440, 2021. 1, 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 653, + 546, + 698 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 653, + 546, + 698 + ], + "spans": [ + { + "bbox": [ + 307, + 653, + 546, + 698 + ], + "type": "text", + "content": "[51] Nikolaos Zioulis, Antonis Karakottas, Dimitrios Zarpalas, and Petros Daras. Omndepth: Dense depth estimation for indoors spherical panoramicas. In ECCV (6), pages 453-471. Springer, 2018. 1, 2" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "9935" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Embodied Multi-Modal Agent trained by an LLM from a Parallel TextWorld/911fda92-445d-4ff5-8ef4-3ff0a7daf4bf_content_list.json b/2024/Embodied Multi-Modal Agent trained by an LLM from a Parallel TextWorld/911fda92-445d-4ff5-8ef4-3ff0a7daf4bf_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..1723c0fb86a534621a5038e11bc481814934e57e --- /dev/null +++ b/2024/Embodied Multi-Modal Agent trained by an LLM from a Parallel TextWorld/911fda92-445d-4ff5-8ef4-3ff0a7daf4bf_content_list.json @@ -0,0 +1,1533 @@ +[ + { + "type": "text", + "text": "Embodied Multi-Modal Agent trained by an LLM from a Parallel TextWorld", + "text_level": 1, + "bbox": [ + 96, + 130, + 875, + 154 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yijun Yang $^{1,5,4}$ Tianyi Zhou $^{2}$ Kanxue Li $^{3}$ Dapeng Tao $^{3}$ Lusong Li $^{4}$", + "bbox": [ + 181, + 178, + 787, + 199 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Li Shen $^{4,*}$ Xiaodong He $^{4}$ Jing Jiang $^{5}$ Yuhui Shi $^{1,*}$", + "bbox": [ + 271, + 198, + 723, + 217 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Southern University of Science and Technology $^{2}$ University of Maryland, College Park", + "bbox": [ + 127, + 217, + 841, + 234 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{3}$ Yunnan University $^{4}$ JD Explore Academy $^{5}$ University of Technology Sydney", + "bbox": [ + 156, + 234, + 810, + 253 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "https://github.com/stevenyangyj/Emma-Alfworld", + "bbox": [ + 279, + 256, + 684, + 270 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/dd386abecb87701148aa9f7c71f19256df3bf2e1083b59acc935a55ebaca7ade.jpg", + "image_caption": [ + "Instruction: put the gold vase into safe." + ], + "image_footnote": [], + "bbox": [ + 99, + 277, + 310, + 455 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/c81a71c3b7afa35b3b05f6787e1a509271383280f73a66b5aae128bd44839227.jpg", + "image_caption": [ + "Action succeeded" + ], + "image_footnote": [], + "bbox": [ + 107, + 478, + 179, + 493 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/5ff5616d4cb5341f71d5f08bc9ee781a70d2fd4702a88a64235b65077caba097.jpg", + "image_caption": [ + "Action failed due to hallucination" + ], + "image_footnote": [], + "bbox": [ + 106, + 496, + 124, + 512 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/18db195502585bb1f6a854aa47c535506f0c4862931e8d7447df405b1bfd1f64.jpg", + "image_caption": [ + "Action failed due to invalid grammar" + ], + "image_footnote": [], + "bbox": [ + 106, + 513, + 122, + 527 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/a33d8ecb5f073cfb94e2660f00a0ff46b8e2329182a418cf5e037e211d913d8e.jpg", + "image_caption": [ + "Figure 1: Comparison of three VLM-based agents in the visual environment from ALFWorld. Given the task instruction and the current-step observation as inputs, a VLM agent is expected to predict an action, e.g., \"go to cabinet 1\", towards completing the task. (a) GPT-4V(ison). (b) InstructBLIP finetuned by behavior cloning on a static dataset of demonstrations by a rule-based expert in the visual simulator. (c) Our Embodied Multi-Modal Agent (EMMA) trained via cross-modality interactive imitation learning from an LLM expert." + ], + "image_footnote": [], + "bbox": [ + 106, + 527, + 124, + 541 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/5ec5a05645cce7c2d08386c06e6bff99f89aa78dcc783f8772b75793136bcdff.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 316, + 280, + 442, + 364 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/3da630c5985c6d006562dd50b1a04300008e6b4a4c614da84e587230d60e8a74.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 444, + 282, + 547, + 364 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/67f671790556931a529d157a1a77a160e4f69365c4fe995c70425a0bd77ce609.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 550, + 282, + 656, + 363 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/b88796c2e9a6ae5a18343a0592fbad2eae4e6b26402607e92f072b2e91e02ddb.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 656, + 282, + 761, + 363 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/b9fdd5cbef6e4a55ad6a26e2c8e64e59ac3e9b74c2a8c9ef4d27fdb47d0c1eb5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 763, + 282, + 870, + 364 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/deecaff978c6d4646e43fefaadabf18ae535fc4e530aeaa4227e7b38e7c09699.jpg", + "image_caption": [ + "Target object" + ], + "image_footnote": [], + "bbox": [ + 316, + 371, + 442, + 454 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/8513336054da0b8916b21dca282448e7ffe438b335339d1fac6f88408253abf2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 444, + 372, + 549, + 454 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/7f74171183cc4f69c9be4b9cf20c7f8df12aa3e3d167f88a659fbf378d17cee7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 550, + 372, + 656, + 454 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/e087309f22d275658538b2441862eef16281351cb4b9236dcfabfa470f74ad6e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 656, + 372, + 761, + 454 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/9c709fbdc18013f5515b605389e83cce378de6df6c586faba909cdd37cfa5716.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 763, + 371, + 870, + 454 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/2f9b1fa27e890aac9551f12c31b27b2cb39693185efb876ef5b75bea74251b20.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 316, + 460, + 442, + 542 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/ee9caca2ddadb3d500ccacefb84ba6e158538645928f9b772dc658c44a1fcf61.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 444, + 460, + 549, + 542 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/54c99cb417df9b2afbb13914625e3d4624da4f78e3978a8c4b9a72c4693c1ef4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 550, + 460, + 656, + 542 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/b2c9a46a911a56759e786f03de43e4965fe1249e747e7780c07da6019aa5d744.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 658, + 460, + 763, + 542 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/8634dd00e81493d7777976ff6ecee3afd1a33173aaebbda5cf5004f665dcb0ff.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 764, + 460, + 870, + 542 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 631, + 313, + 647 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "While large language models (LLMs) excel in a simulated world of texts, they struggle to interact with the more realistic world without perceptions of other modalities such as visual or audio signals. Although vision-language models (VLMs) integrate LLM modules (1) aligned with static image features, and (2) may possess prior knowledge of world dynamics (as demonstrated in the text world), they have not been trained in an embodied visual world and thus cannot align with its dynamics. On the other hand, training an embodied agent in a noisy visual world without expert guidance is often challenging and inefficient. In this paper, we train a VLM agent living in a visual world using an LLM agent excelling in a parallel text world. Specifically, we distill LLM's reflection outcomes (improved actions by analyzing mistakes) in", + "bbox": [ + 75, + 662, + 472, + 876 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "a text world's tasks to finetune the VLM on the same tasks of the visual world, resulting in an Embodied Multi-Modal Agent (EMMA) quickly adapting to the visual world dynamics. Such cross-modality imitation learning between the two parallel worlds is achieved by a novel DAgger-DPO algorithm, enabling EMMA to generalize to a broad scope of new tasks without any further guidance from the LLM expert. Extensive evaluations on the ALFWorld benchmark's diverse tasks highlight EMMA's superior performance to SOTA VLM-based agents, e.g., $20\\% -70\\%$ improvement in the success rate.", + "bbox": [ + 496, + 633, + 893, + 799 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 500, + 814, + 632, + 830 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Embodied multi-modal agents have been acknowledged as a pivotal stride towards achieving Artificial General Intelligence (AGI), as they encompass the potential for a broad scope of intelligent activities [27, 63]. The rising of founda", + "bbox": [ + 496, + 839, + 895, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author", + "bbox": [ + 94, + 887, + 220, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "26275", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "tion models brings a glimmer of hope for constructing such agents [11, 32, 34, 35, 38, 71, 76], and notable efforts from the community have tried to harness them in many decision-making scenarios, e.g., autonomous driving [15, 66], daily household robots [6, 13, 77], and complex manipulation tasks [25, 29, 42, 57].", + "bbox": [ + 75, + 90, + 470, + 181 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "LLMs can perform as reflex agents interacting with a text world via verbalized descriptions of the world states and textual actions. In addition, they exhibit great potential in planning, reflection, and reward shaping. This is mainly due to their prior knowledge and semantic abstraction of the world. However, LLM-based agents cannot be directly applied in a visual world. Although current vision-language models (VLMs) try to align LLMs with the visual modality, their pretraining only focuses on static alignment between image-text pairs so the resulting agents have not been aligned well with the dynamics of the visual world. As shown in Fig. 1 (a), even the privileged VLM, e.g., GPT-4V [43], fails to accomplish tasks in an embodied ALFWorld environment [56]. In such a zero-shot setting, GPT-4V tends to mainly rely on its language prior of these objects detected in the current step, rather than the alignment between the visual input and the environment dynamics conditioned on the task instruction.", + "bbox": [ + 75, + 184, + 473, + 455 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we study how to train a VLM towards an embodied agent in a visual world by aligning it with the visual world dynamics and distilling the skills of an LLM agent in a parallel text world. Our overarching goal is to build such an Embodied Multi-Modal Agent (EMMA) that can take a textual task instruction (e.g., from human users) and pixel observations of the state per step to produce a sequence of actions leading to the efficient completion of the task. This is a challenging problem due to (1) the sparsity of task reward [2-4, 14], (2) noisy visual representations, (3) the hallucination of VLMs [60], and (4) the misalignment of VLM's static representations to the visual world dynamics. While offline distillation and imitation from a powerful LLM agent in a parallel text world can potentially overcome the former two challenges [75], effectively mitigating the remaining challenges necessitates online finetuning of the VLM agent within an interactive and visual world [18, 31, 70].", + "bbox": [ + 75, + 459, + 473, + 717 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To this end, we finetune a VLM agent by imitation learning from an LLM expert (e.g., built on ChatGPT) launched in a parallel text world on the same tasks. Specifically, in each step of EMMA interacting with the visual world, we convert its visual observation into an equivalent textual description sent to the LLM agent, which produces an action for EMMA to imitate. Such cross-modality interactive imitation learning is based on DAgger [51], which overcomes the cumulative errors and distribution shifts caused by behavior cloning (BC). As depicted in Fig. 1 (b), an InstructBLIP [11] agent finetuned by BC on 170K expert demonstrations produced by a rule-based expert in the", + "bbox": [ + 75, + 719, + 473, + 902 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/d391294bac87c81adc36437d75c4a211507e4811d908ce2f9080248a19a5a35e.jpg", + "image_caption": [ + "Figure 2: An example of tasks generated for the two parallel worlds. A VLM agent in the visual world and an LLM agent in the text world as household robots instructed to clean an apple and then put it into the fridge. Zoom in for more details." + ], + "image_footnote": [], + "bbox": [ + 503, + 88, + 890, + 282 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "visual world still fails to take correct actions based on visual observations. We further improve the DAgger's objective to be the direct preference optimization (DPO) [48], which maximizes the preference of LLM-expert's action (positive) over VLM-student's action (negative) in each interaction step. To collect better teaching signals retrospective to the VLM student's actions, the LLM expert is composed of an LLM actor prompted to output expert actions, and an LLM critic prompted for reflection feedback on the VLM agent's historical trajectories. We maintain a long-term memory storing the feedback, which is then used to induce the LLM actor to improve actions for imitation in future episodes.", + "bbox": [ + 496, + 377, + 892, + 558 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We evaluate EMMA and compare it with vision-only agents, LLM agents, and VLM agents deployed to the ALF-World benchmark [56], which includes numerous tasks in both visual and textual environments. Extensive evaluations highlight that EMMA substantially outperforms state-of-the-art (SOTA) VLM agents in visual-only environments by $20\\% - 70\\%$ in terms of success rate. In addition, EMMA is the only VLM agent that can generalize to open-vocabulary and free-form test tasks, shedding novel insights on using LLM feedback to train more versatile and generalizable embodied agents in multi-modality environments.", + "bbox": [ + 496, + 559, + 893, + 726 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Embodied Multi-Modal Agent", + "text_level": 1, + "bbox": [ + 500, + 738, + 779, + 757 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Fig. 3 illustrates the main idea of our \"Embodied MultiModal Agent (EMMA)\", whose detailed training procedures are given in Alg. 1. The agent is built upon a modularized VLM, which can follow instructions and interact with the environment through pixel observations and textual actions. To overcome these challenges associated with training EMMA, such as sparse reward or distribution shift, we explore the construction of an LLM expert from a parallel text world (Sec. 2.2) for providing EMMA with step-by-step guidance.", + "bbox": [ + 496, + 763, + 895, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "26276", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/b51fafabd31cf2d615efc5a6fe5a3c8a11df5156f42a4bdb0de996fe16c19afe.jpg", + "image_caption": [ + "Figure 3: Embodied Multi-Modal Agent (EMMA) trained by an LLM expert via cross-modality imitation learning. EMMA takes a textual task instruction and pixel observations as its input state per step to generate a sequence of actions using a VLM. Then, we convert each pixel observation into a textual equivalent as the context of an LLM expert to produce improved actions for EMMA to imitate." + ], + "image_footnote": [], + "bbox": [ + 99, + 93, + 872, + 311 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Lastly, Sec. 2.3 further discusses how to harness the LLM expert to train EMMA via cross-modality imitation learning.", + "bbox": [ + 75, + 397, + 470, + 429 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.1. EMMA in the Visual World", + "text_level": 1, + "bbox": [ + 76, + 439, + 326, + 455 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In visual environments, EMMA $\\pi_{\\theta}$ is designed to process a task instruction $x_{\\mathrm{task}}$ (e.g., provided by human users) and the pixel observation $s_v^t$ at each interaction step $t$ . It is expected to generate a sequence of high-level textual actions $\\{x_a^t \\sim \\pi_\\theta(\\cdot | x_{\\mathrm{task}}, s_v^t)\\}_{t=0}^T$ towards efficient completion of the task. To achieve this, we draw inspiration from recent advances of large pretrained VLMs [11, 32, 34, 38, 76], and modularize EMMA's architecture into four components: (1) a ViT to encode $s_v$ into visual embeddings, (2) a querying transformer (Q-Former) tailored to extract the most relevant visual features via the cross-attention between the visual embeddings and query tokens, (3) a linear projection layer to align visual features to text embeddings, (4) an LLM decoder taking the concatenation of the instruction tokens and the output of the linear projection layer to autoregressively generate the action $x_a$ . In order to reduce computational overhead and prevent catastrophic forgetting [37], we adopt the pretrained ViT, Q-Former, and LLM from InstructBLIP [11] and keep them frozen at the finetuning stage. We only update the linear projection layer, as illustrated in Fig. 3. Such a modularized architecture enables EMMA to integrate any existing pretrained vision models and LLMs in a flexible and computationally-efficient way.", + "bbox": [ + 75, + 465, + 470, + 811 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "However, deploying EMMA into a complex visual world remains challenging. One of the main obstacles is that the direct use of any pretrained VLM is suboptimal because existing pretraining only focuses on static alignment between", + "bbox": [ + 75, + 811, + 470, + 875 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "image-text pairs [11, 32, 34], so that the resulting agent may struggle to reason about the dynamics of the world. As discussed in Sec. 1, even the SOTA VLM, i.e., GPT-4V, fails to accomplish tasks in an embodied ALFWorld environment [56]. In such a zero-shot setting, GPT-4V tends to rely on the linguistic prior of the currently detected objects, rather than the given task instruction and the underlying dynamics of the environment, to guide the interactions. Moreover, finetuning a pretrained VLM on a pre-collected demonstration dataset is also suboptimal due to the diversity of environments and tasks [77], the lack of large-scale expert annotations [44] as well as the challenges posed by the distribution shift issue [31]. A seemingly natural solution to the above challenges is reinforcement learning from environmental feedback (RLEF) [68], in which reward signals rely on decomposing a task into a sequence of reasonable sub-goals and checking their completion. However, in real-world scenarios, most sub-goals cannot be defined or described precisely, so the reward is sparse; hence, we do not expect RLEF to be effective.", + "bbox": [ + 496, + 397, + 893, + 698 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To this end, we propose to leverage interactive imitation learning (IL) to align EMMA with the dynamics of any environment, which however results in two critical algorithmic challenges: (1) How to obtain a high-quality, accessible, and scalable expert that EMMA can query during IL (Sec. 2.2)? (2) Designing an effective strategy to train EMMA using this expert in complex, diverse, and potentially open-ended environments (Sec. 2.3).", + "bbox": [ + 496, + 699, + 893, + 820 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. LLM Expert from a Parallel TextWorld", + "text_level": 1, + "bbox": [ + 500, + 830, + 841, + 847 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Thanks to a series of prompting techniques in the realm of in-context learning, such as chain-of-thought [62], tree-of-thought [72], ReAct [73], and Reflexion [54], pretrained", + "bbox": [ + 496, + 854, + 893, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_footnote", + "text": "For brevity, we omit $x_{\\mathrm{task}}$ in the rest of this paper.", + "bbox": [ + 94, + 886, + 364, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "26277", + "bbox": [ + 478, + 944, + 517, + 957 + ], + "page_idx": 2 + }, + { + "type": "code", + "sub_type": "algorithm", + "code_caption": [ + "Algorithm 1 DAgger-DPO with a single task instruction" + ], + "code_body": "1: initialize: $i = 0$ $\\mathcal{D}\\gets \\emptyset$ , EMMA $\\pi_{\\theta}$ , LLM actor $M_a$ with a FIFO memory pool $\\mathcal{P}\\gets \\emptyset$ , LLM critic $M_c$ \n2: input: max trials $I$ , training epochs $I_{e}$ , visual and textual envs $E_{v},E_{l}$ , instruction $x_{\\mathrm{task}}$ , reference agent $\\pi_{\\mathrm{ref}}$ \n3: Initialize $\\pi_{\\theta}$ to $\\pi_{\\mathrm{ref}}$ Behavior Cloning Initialization \n4: while task not completed or $i < I$ do \n5: Get $\\tau_v^i = [x_{\\mathrm{task}},s_v^0,x_a^0,\\dots,s_v^T,x_a^T]$ via $E_{v}$ with $\\pi_{\\theta}$ \n6: Get $\\tau_l^i = [x_{\\mathrm{task}},s_l^0,x_a^0,\\dots,s_l^T,x_a^T]$ via $E_{l}$ with $\\tau_v^i$ \n7: Generate retrospective feedback $\\mathcal{P}_i = M_c(\\tau_l^i)$ \n8: Update $\\mathcal{P}$ with $\\mathcal{P}_i$ (i.e., $\\mathcal{P}\\leftarrow \\mathcal{P}\\cup \\mathcal{P}_i$ \n9: for $t = 0$ to $T$ do Dataset Aggregation \n10: $x_{a}^{*} = M_{a}(\\mathcal{P},x_{\\mathrm{task}},\\ldots ,x_{a}^{t - 1},s_{l}^{t})$ \n11: $\\mathcal{D}\\gets \\mathcal{D}\\cup \\{(x_{\\mathrm{task}},s_v^t,x_a^t,x_a^*)\\}$ \n12: for $j = 0$ to $I_{e} - 1$ do Gradient Descent on $\\theta$ \n13: Sample a mini-batch $\\tau$ from $\\mathcal{D}$ \n14: Update $\\theta$ by minimizing Eq. (2) with $\\pi_{\\mathrm{ref}}$ and $\\tau$ \n15: $i\\gets i + 1$ \n16: output: $\\pi_{\\theta^{*}}$", + "bbox": [ + 81, + 111, + 468, + 393 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "LLMs have demonstrated impressive zero-shot performance across many decision-making scenarios [6, 13, 15, 25, 29, 42, 57, 66, 77]. Despite the great potential in serving as high-quality and scalable experts, they are only able to interact with the environment via textual descriptions of the states, rather than using raw pixel observations like EMMA. To bridge this gap, we convert each pixel observation $s_v$ into a textual equivalent by extracting its metadata from the simulator [30], which is composed of attributes such as Observed Objects, Observed Relations, Inventory, and Locations. We then employ the Planning Domain Definition Language (PDDL) [1] to describe this metadata and create an equivalent textual description/state $s_l$ using the TextWorld engine [10]. Additional details are available in Appendix 7 and an example of this process is illustrated in Fig. 2. This methodology enables the utilization of any pretrained LLM agent that generates a sequence of actions facilitating the training of EMMA through cross-modality IL between the two agents.", + "bbox": [ + 75, + 422, + 472, + 710 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.3. Training EMMA via Cross-Modality Imitation", + "text_level": 1, + "bbox": [ + 76, + 720, + 468, + 736 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Given an LLM expert from the parallel TextWorld, we aim to train a VLM agent $\\pi_{\\theta}$ in the visual world to imitate its behaviors closely. This is equal to minimizing the following objective under the distribution of states induced by $\\pi_{\\theta}$ .", + "bbox": [ + 75, + 744, + 468, + 806 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\theta^ {*} = \\underset {\\theta \\in \\Theta} {\\arg \\min } \\mathbb {E} _ {\\pi_ {\\theta}} \\left[ \\mathcal {L} _ {\\mathrm {i m i t}} \\left(\\pi_ {\\theta} \\left(x _ {a} \\mid s _ {v}\\right), x _ {a} ^ {*}\\right)\\right) ], \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 133, + 818, + 468, + 842 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "in which the choice of loss function $\\mathcal{L}_{\\mathrm{imit}}$ is dependent on specific scenarios. For instance, it may be the expected cross-entropy loss for the discrete action space, or the expected", + "bbox": [ + 75, + 854, + 472, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "MSE loss for the continuous action space. In our case, we select DPO [48] loss due to its proven superior performance to the cross-entropy on aligning models with expert preferences within the discrete language space. Hence, Eq. (1) can be extended to the formulation below.", + "bbox": [ + 496, + 90, + 893, + 167 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\theta^ {*} = \\underset {\\theta \\in \\Theta} {\\arg \\min } - \\mathbb {E} _ {\\pi_ {\\theta}} \\left[ \\mathcal {L} _ {\\operatorname {i m i t}} \\left(\\pi_ {\\theta}, \\pi_ {\\text {r e f}}, s _ {v}, x _ {a}, x _ {a} ^ {*}\\right) \\right], \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 504, + 178, + 892, + 203 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {i m i t}} (\\cdot) \\triangleq \\log \\sigma \\left(\\beta \\log \\frac {\\pi_ {\\theta} \\left(x _ {a} ^ {*} \\mid s _ {v}\\right)}{\\pi_ {\\mathrm {r e f}} \\left(x _ {a} ^ {*} \\mid s _ {v}\\right)} - \\beta \\log \\frac {\\pi_ {\\theta} \\left(x _ {a} \\mid s _ {v}\\right)}{\\pi_ {\\mathrm {r e f}} \\left(x _ {a} \\mid s _ {v}\\right)}\\right)\n$$\n", + "text_format": "latex", + "bbox": [ + 506, + 205, + 883, + 241 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $x_{a}^{*}$ is the action given by an expert while $x_{a}$ is the action given by $\\pi_{\\theta}$ , $\\sigma$ is the logistic function, and $\\beta$ is a hyperparameter controlling the deviation from $\\pi_{\\mathrm{ref}}$ , i.e. the reference agent obtained by behavior cloning on a demonstration dataset produced by a rule-based expert in the visual world (see Appendix 9 for complete details). This regularization is essential, as it prevents the agent from deviating too far from the distribution on which the expert is accurate, as well as maintaining the generation diversity and avoiding premature convergence to some easy tasks [48]. In practice, the VLM agent $\\pi_{\\theta}$ is also initialized to $\\pi_{\\mathrm{ref}}$ for stabilizing the training process. Since environmental dynamics is both unknown and complex, we cannot compute the distribution of states visited by $\\pi_{\\theta}$ and can only sample it by rolling out the agent. Hence, Eq. (2) is a non-i.i.d. supervised learning problem due to the dependence of the state distribution on $\\pi_{\\theta}$ itself, in which naive behavior cloning faces issues like cumulative error and distribution shift [26]. To address this, we employ an interactive IL algorithm, DAgger [51], which provably converges to the optimal agent $\\pi_{\\theta^{*}}$ .", + "bbox": [ + 496, + 250, + 893, + 551 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As discussed in Sec. 2.2 and Sec. 2.3, we can harness an LLM expert to generate a sequence of actions, serving as $x_{a}^{*}$ in Eq. (2). However, these actions may be suboptimal due to sparse environmental feedback [54, 68] or defective in-context instructions [9]. To collect better teaching signals, we introduce a \"retrospective LLM expert\" that is composed of two specialized models: an actor $(M_{a})$ , built upon an API LLM and prompted to generate actions based on the task instruction and state observations; and a critic $(M_{c})$ , also based on the same LLM, but designed to analyze EMMA's historical interactions and provide reflective feedback. A long-term memory $\\mathcal{P}$ is maintained to store the feedback generated by $M_{c}$ , which is then used to prompt $M_{a}$ for improved actions. The complete procedure is detailed in Line 7-10 of Alg. 1, and all prompts are provided in Sec. 6 of the Appendix.", + "bbox": [ + 496, + 551, + 893, + 792 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3. Experiments", + "text_level": 1, + "bbox": [ + 500, + 806, + 633, + 824 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1. Experimental Setup", + "text_level": 1, + "bbox": [ + 500, + 832, + 689, + 848 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "**Environments.** We base our experiments on the ALFWorld benchmark [56], a cross-modality simulation platform that encompasses a wide range of embodied household tasks. For", + "bbox": [ + 496, + 854, + 893, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "26278", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/da5c7f948991768971c6b06a25303425486e734e314fb500c279df7962cdf677.jpg", + "image_caption": [ + "Figure 4: Comparison of success rate between EMMA and the retrospective LLM expert. As the number of trials increases, the LLM expert progressively improves its performance through retrospective processes, and also the gap between the two agents decreases, indicating the effectiveness of cross-modality imitation. We separately plot the curve per task type in Fig. 9 of Appendix." + ], + "image_footnote": [], + "bbox": [ + 91, + 88, + 352, + 251 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/9ca5b87bd0e46e319a5481efb7c4bd2a7caa6919802e4bf6c3c360919664de97.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
AgentVisual Env.Textual Env.Success Rate with Template Task Instruction
Avg.PickCleanHeatCoolLookPick2
VMsResNet-18* [56]✓X0.06(-)------
MCNN-FPN* [56]✓X0.05(-)------
LMsBUTLER* [56]X✓0.26(-)0.31(-)0.41(-)0.60(-)0.27(-)0.12(-)0.29(-)
GPT-BUTLER [41]X✓0.69(18.8)0.62(18.4)0.81(18.4)0.85(12.7)0.78(15.6)0.50(24.4)0.47(26.6)
ReAct [73]X✓0.54(20.6)0.71(18.1)0.65(18.8)0.62(18.2)0.44(23.2)0.28(23.7)0.35(25.5)
Reflexion [54]X✓0.91(18.7)0.96(17.4)1.00(17.0)0.81(19.4)0.83(21.6)0.94(16.9)0.88(21.6)
DEPS* [61]X✓0.76(-)0.93(-)0.50(-)0.80(-)1.00(-)1.00(-)0.00(-)
AutoGen* [64]X✓0.77(-)0.92(-)0.74(-)0.78(-)0.86(-)0.83(-)0.41(-)
VLMsMiniGPT-4 [76]✓X0.16(26.9)0.04(29.0)0.00(30.0)0.19(26.3)0.17(26.7)0.67(17.7)0.06(28.9)
BLIP-2 [34]✓X0.04(29.5)0.00(30.0)0.06(29.3)0.04(29.9)0.11(28.2)0.06(29.2)0.00(30.0)
LLaMA-Adapter [17]✓X0.13(27.5)0.17(26.4)0.10(28.6)0.27(24.2)0.22(26.7)0.00(30.0)0.00(30.0)
InstructBLIP [11]✓X0.22(26.2)0.50(21.5)0.26(25.0)0.23(27.2)0.06(28.9)0.17(26.8)0.00(30.0)
EMMA (Ours)✓X0.82(19.5)0.71(19.3)0.94(17.5)0.85(19.6)0.83(19.9)0.88(19.6)0.67(22.4)
Human Performance* [55]✓X0.91(-)------
", + "bbox": [ + 370, + 88, + 883, + 250 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Table 1: Comparison with the state of the arts. $*$ -reported in previous work. VMs = vision models, LMs = language models, VLMs = vision-language models. \"Visual Env.\" and \"Textual Env.\" refer to the visual and the parallel textual environments from ALFWorld [56], respectively. $\\sqrt{} / X$ denotes the corresponding environment used/not used to deploy the agent. The highest scores for each task in the same type of environment are highlighted in bold. The average interaction steps are given in the (parentheses). EMMA substantially outperforms other SOTA VLM agents in the visual environments, and its success also directs a promising way to achieve human-level performance in ALFWorld.", + "bbox": [ + 367, + 258, + 885, + 385 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "each task, ALFWorld integrates a visual environment, rendered by the Ai2Thor simulator [30], with a corresponding textual environment. Textual environments employ the Planning Domain Definition Language (PDDL) [1] to translate each pixel observation from the simulator into an equivalent text-based observation, and then construct an interactive world using the TextWorld engine [10]. Fig. 2 provides an illustrative example of the tasks featured in ALFWorld. The tasks within the benchmark are categorized into 6 types: Pick & Place, Clean & Place, Heat & Place, Cool & Place, Look in Light, and Pick Two Objects & Place. Each task requires an agent to execute a series of text-based actions, such as \"go to safe 1\", \"open safe 1\", or \"heat egg 1 with microwave 1\", following a predefined instruction. These actions involve navigating and interacting with the environment. To provide a comprehensive understanding, we have visualized an example of each task type in Fig. 8 of the Appendix. A task in this benchmark may involve interactions with over 10 objects and require more than 30 steps for a human expert to solve, thus challenging an agent's capabilities in long-horizon planning, instruction following, and the utilization of commonsense knowledge. For a fair comparison, we follow the same setting as prior work [41, 54-56, 73] and evaluate all baselines using 134 out-of-distribution (OOD) tasks.", + "bbox": [ + 75, + 411, + 472, + 773 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Baselines. To verify the effectiveness of cross-modality imitation learning, we compare our EMMA with several baselines and SOTA agents using the ALFWorld benchmark with both visual and textual environments. The compared agents can be divided into three categories: vision models, language models, and vision-language models. Concretely, vision models, including ResNet-18 [20] and MCNN-FPN [21], utilize pretrained vision encoders to extract salient features", + "bbox": [ + 75, + 779, + 472, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "from each pixel observation. The extracted features then serve as input for a Multi-Layer Perceptron (MLP) policy, which is trained by behavior cloning on a pre-collected demonstration dataset. Unlike vision models performing in the visual environment, language models complete the same tasks but in a parallel, text-based environment. BUTLER [56] employs a transformer seq2seq model enhanced with a pointer softmax mechanism [19]. This architecture aggregates previous observations as input to generate text-based actions token-by-token. GPT-BUTLER [41], a variant of the GPT-2 model [45], is initially pretrained on a static demonstration dataset and further finetuned using data collected online. ReAct [73] takes a novel approach by utilizing LLMs to generate reasoning traces and task-specific actions in an interleaved manner. This method aids the agent in developing, tracking, and updating its action plans interactively. Reflexion [54] similarly employs an LLM, but it focuses on reflecting upon environmental feedback. It maintains this reflective text in an episodic memory buffer, enhancing the agent's ability to improve actions in subsequent trials. Similar to Reflexion, a concurrent work, DEPS [61], also corrects errors in previous LLM-generated actions by integrating descriptions of the action execution process and providing self-explanations for the feedback. Moreover, beyond the single-agent framework, AutoGen [64] exhibits the potential to accomplish a broad spectrum of tasks through the cooperation of multiple LLM agents. Finally, we consider a range of vision-language models, such as MiniGPT-4 [76], BLIP-2 [34], LLaMA-Adaptor [17], and InstructBLIP [11], as agents to interact with the visual environment. Unlike pure vision or language models, VLMs are designed to process and integrate both visual and textual data, offering a", + "bbox": [ + 496, + 411, + 895, + 897 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "26279", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "more holistic understanding of the environment. To align these agents with the specific requirements of the ALFWorld benchmark, we finetune them on a pre-collected demonstration dataset. This finetuning process is crucial as it enables the agent to comprehend and adhere to ALFWorld's unique grammar and to develop a basic \"gamesense\".", + "bbox": [ + 75, + 90, + 470, + 183 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.2. Training Details", + "text_level": 1, + "bbox": [ + 76, + 193, + 238, + 208 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The architectural design of EMMA is depicted in Fig. 3. At its core, EMMA employs a Query Transformer (Q-Former) to process visual data. This Q-Former extracts features using a frozen ViT encoder. Its output consists of 32 visual tokens, which are then passed through a linear projection layer before being fed to a frozen LLM decoder. Similar to other VLM agents, EMMA is also finetuned on a pre-collected demonstration dataset, aligning its basic ability with the ALFWorld benchmark. In order to align EMMA with the dynamics of ALFWorld, we train it by imitating an LLM expert (see Alg. 1). We choose text-davinci-003 developed by OpenAI as our LLM expert because of its established capabilities in reasoning and planning [54, 61, 64, 73]. In this setup, text-davinci-003 serves dual roles: it serves as an actor, providing EMMA with expert actions, and as a critic, analyzing EMMA's historical trajectories. This analysis generates retrospective feedback, which is then incorporated into the actor's long-term memory, leading to improved actions in future trials. Further details about the hyperparameters and prompts used in our training procedure are available in Table 2 of the Appendix.", + "bbox": [ + 75, + 215, + 470, + 534 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.3. Comparison with State of the Art", + "text_level": 1, + "bbox": [ + 76, + 544, + 370, + 559 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "EMMA sets new SOTA performance in visual ALFWorld.", + "text_level": 1, + "bbox": [ + 75, + 568, + 470, + 582 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this section, we compare EMMA with 12 other representative agents on the ALFWorld benchmark, spanning both visual and textual environments (refer to Table 1). We assess two key metrics: the success rate, which is the percentage of trials completed successfully, and the average number of interaction steps required for task completion, with a lower number indicating higher efficiency. EMMA demonstrates superior performance in both metrics, significantly outperforming all VLM agents in visual environments. This achievement underscores the effectiveness of our cross-modality imitation learning approach, as depicted in the learning curve shown in Fig. 4. Furthermore, EMMA's performance markedly exceeds that of VM agents, highlighting the crucial role of the prior knowledge embedded in VLMs. Intriguingly, EMMA's performance is comparable with LLM agents that operate using perfectly semantic descriptions of visual observations. This is largely attributed to EMMA's training strategy of imitating an expert LLM agent, proving to be more efficient than learning from scratch in a purely visual setting. As a result, EMMA stands out as the only VLM agent that substantially surpasses SOTA LLM agents, such", + "bbox": [ + 75, + 583, + 470, + 901 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/661817e8d2e11ec4bf36b68541811cd20f5933ab66c201456828618cbbf7bb2f.jpg", + "image_caption": [ + "Figure 5: Comparison of robustness between VLM and LLM agents. Given a specific noise rate, we crop a random portion of the pixel observation and resize it to a given size as the new observation. Similarly, we randomly replace some tokens in the textual observation with arbitrary ones." + ], + "image_footnote": [], + "bbox": [ + 517, + 87, + 874, + 306 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "as AutoGen [64] and DEPS [61], in these environments. And its success also directs a potential way to achieve human-level performance in the visual environments of ALFWorld. EMMA is more robust to noisy observations than LLM agents. While LLM agents exhibit a higher success rate with fewer interaction steps in textual environments, as indicated in Table 1, we hypothesize that this superior performance largely relies on their precise semantic abstraction of the environment. However, such an abstraction might not be feasible in real-world applications. To verify this assumption, we set up a more practical scenario where observations are deliberately perturbed at a specific noise rate. We then compare the robustness of EMMA and a SOTA LLM agent, Reflexion, under these noisy observations. To generate noisy observations, a random portion of the visual observation is cropped, resized, and then used to replace the original observation. Similarly, in the textual observation, random tokens are substituted with arbitrary ones. As illustrated in Fig. 5, with the noise rate increasing, EMMA's performance remains significantly more robust than Reflexion. This could be attributed to the vision encoder in the VLM, which is adept at filtering out visual noises. On the other hand, textual noises are directly processed by the LLM, which can substantially impair the performance of LLM-based agents. This finding highlights the potential advantages of VLM agents like EMMA in visual scenarios, in which data is often imperfect and noisy.", + "bbox": [ + 496, + 412, + 893, + 821 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.4. Ablation Study", + "text_level": 1, + "bbox": [ + 500, + 830, + 653, + 847 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Retrospection improves EMMA over time. To assess the importance of the retrospective LLM expert, we present the average success rate of EMMA after each trial and compare it", + "bbox": [ + 498, + 854, + 893, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "26280", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/905e47f392e5add8df7e5b62799313afdd903e44e94ef900117cf7d6cc0b2ebf.jpg", + "image_caption": [ + "Figure 6: Ablation study. left: The ability of LLM expert to reflect on and learn from past actions appears to be a key factor in achieving the impressive results we observed. middle: The \"w/o BC initialization\" variant uses a pretrained VLM instead of the reference agent $\\pi_{\\mathrm{ref}}$ to initialize EMMA. right: The 'w/ CE Loss' variant replaces the DPO loss (described in Eq. 2) with a token-level cross-entropy (CE) loss. Please refer to Sec. 3.4 for a thorough discussion." + ], + "image_footnote": [], + "bbox": [ + 83, + 87, + 349, + 242 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/338e765a491bd9953ea2d10a8a5b075ecbc0ea7e8758ec8632f97b4e55254ff1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 351, + 87, + 614, + 242 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/764c580ded6530e9ca9a4cb531669fb436a44f13801208a5de2f71d579b0407c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 619, + 88, + 883, + 242 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "with a key variation: \"EMMA w/o Retrospection\", as shown in Fig. 6 (left). This variant of EMMA is trained using the same procedure as the original EMMA but removes the retrospective process. Instead, it relies solely on a plain LLM actor to provide relabeled actions. The results show that EMMA with the retrospective mechanism significantly outperforms its counterpart. This finding is crucial as it indicates that the retrospective process is not just a supplementary feature but a fundamental component of EMMA's architecture that contributes substantially to its enhanced performance.", + "bbox": [ + 75, + 335, + 470, + 487 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "EMMA benefits from BC initialization. Through an ablation study, we evaluate the impact of behavior cloning (BC) initialization, a process described in line 3 of Alg. 1. Fig. 6 (middle) demonstrates that EMMA, when deprived of BC initialization, experiences a slight reduction in the average success rate across 134 unseen tasks compared to its original setup. Despite this decrease, EMMA without BC initialization still outperforms other VLM agents, as clearly shown when compared with the results in Table 1. Furthermore, Fig. 10 in the Appendix breaks down the result by task type. It reveals a consistent but slight drop in performance across 5 out of the 6 task types. These results reflect that while BC initialization contributes positively to EMMA's overall performance, it is not critical for achieving the notable results we have reported.", + "bbox": [ + 75, + 489, + 470, + 715 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "DPO enables more effective imitation learning. To evaluate the effectiveness of using DPO loss in Eq. (1), we conducted an ablation study with an alternative version of EMMA, referred to \"EMMA w/ Cross Entropy (CE) Loss\". In this variant, EMMA is optimized using the token-level CE loss, a common objective for finetuning VLMs. The results, as depicted in Fig. 6 (right), reveal that EMMA w/ CE Loss does not achieve the same high success rate as the original EMMA with DPO loss, suggesting that DPO loss contributes to enhancing the upper performance bound of EMMA. In addition, we noted that EMMA w/ CE Loss exhibits faster convergence in the initial training stages than the original", + "bbox": [ + 75, + 719, + 473, + 901 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "EMMA. This premature convergence leads to the agent paying attention to the expert actions from easier tasks, usually addressed in the first few training epochs. This can suppress EMMA's exploration and learning on more complex tasks, resulting in worse performance.", + "bbox": [ + 496, + 335, + 893, + 411 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.5. Generalization to Free-form Task Instructions", + "text_level": 1, + "bbox": [ + 498, + 417, + 890, + 431 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The ability of AI agents to precisely follow task instructions given in open-vocabulary and free-form text is crucial for real-world problems. To assess this, we conducted an additional experiment focusing on the generalization capabilities of EMMA, an agent trained with template task instructions. In this experiment, we re-evaluated the performance of EMMA and other baseline agents on 134 unseen tasks, using human-annotated instructions instead of template ones. These human-annotated instructions include a large amount of OOD verbs and objects, presenting a more realistic and challenging scenario for the agents to address. To underscore this challenge, we compared the vocabulary distribution between the template and human-annotated instructions, as shown in Fig. 7 (right). Moreover, we provide a comprehensive analysis of the vocabulary used across both instruction types in Fig. 11 of the Appendix. In Fig. 7 (left), EMMA demonstrates a slight performance decline, while a significant degradation is observed in other baselines. We also note that Reflexion, a SOTA LLM agent, exhibits exceptional generalization to those OOD instructions. According to these empirical results in Fig. 7, we have the following conclusions: (1) EMMA obtains and benefits from the generalization capabilities inherent in the SOTA LLM agent through cross-modality imitation learning; (2) Our work sheds novel insights on using LLM feedback to train more versatile and generalizable embodied agents.", + "bbox": [ + 496, + 439, + 893, + 834 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4. Related Work", + "text_level": 1, + "bbox": [ + 500, + 844, + 640, + 859 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Agents based on Foundation Models. Recent research has increasingly focused on harnessing the capabilities of large", + "bbox": [ + 500, + 869, + 892, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "26281", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/6867d7eea91c5b3cc9d4ec3847cf76ee936850c1bfe521c48aa225110338c3b4.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
AgentVisual Env.Textual Env.Success Rate with Free-form Task Instruction
Avg.PickCleanHeatCoolLookPick2
LMsBUTLER* [56]X✓0.03 (-)0.10 (-)0.22 (-)0.05 (-)0.17 (-)0.00 (-)0.00 (-)
GPT-BUTLER [41]X✓0.31 (24.7)0.25 (25.4)0.42 (23.5)0.46 (20.9)0.44 (21.8)0.00 (30.0)0.12 (28.9)
ReAct [73]X✓0.37 (23.6)0.46 (22.0)0.39 (23.8)0.50 (20.5)0.28 (24.6)0.22 (26.4)0.29 (25.9)
Reflexion [54]X✓0.78 (17.0)0.84 (15.8)0.74 (18.4)0.77 (17.2)0.61 (19.4)0.94 (13.8)0.82 (16.7)
VLMsMiniGPT-4 [76]✓X0.00 (30.0)0.00 (30.0)0.00 (30.0)0.00 (30.0)0.00 (30.0)0.00 (30.0)0.00 (30.0)
BLIP-2 [34]✓X0.01 (29.7)0.04 (29.1)0.03 (29.5)0.00 (30.0)0.00 (30.0)0.00 (30.0)0.00 (30.0)
LLaMA-Adapter [17]✓X0.02 (29.6)0.04 (29.3)0.03 (29.4)0.04 (29.3)0.00 (30.0)0.00 (30.0)0.00 (30.0)
InstructBLIP [11]✓X0.01 (29.8)0.00 (30.0)0.03 (29.3)0.00 (30.0)0.00 (30.0)0.00 (30.0)0.00 (30.0)
EMMA (Ours)✓X0.68 (22.0)0.72 (21.7)0.65 (22.7)0.72 (22.0)0.56 (23.9)0.67 (23.3)0.76 (18.1)
", + "bbox": [ + 80, + 89, + 645, + 218 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/429e678f6864656b282ff8d85ee02fb2cc44319c20b241abdf6e173f0fae4930.jpg", + "image_caption": [ + "Figure 7: Generalization to open-vocabulary, free-form task instructions. Left: These instructions are provided by different human annotators using Amazon Mechanical Turk [55]. Right: Frequency distribution of verbs for human-annotated and templated task instructions." + ], + "image_footnote": [], + "bbox": [ + 661, + 88, + 890, + 220 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "pre-trained foundation models to build AI agents [65, 69]. These models (e.g., LLMs), benefiting from their commonsense knowledge inherited from Internet-scale pretraining, are able to reason actions according to descriptions of the external environments. For example, given a set of task instructions, LLMs can be elaborately prompted to perform as agents generating high-level step-by-step plans [7, 24, 54, 64, 73], and each step can be parsed into a sequence of robotic actions that are executed via pretrained policies or available APIs [6, 36, 59]. Furthermore, by using VLMs [32], plans can also be conditioned on visual inputs that are transformed into language descriptions [7, 16, 25, 57] or token embeddings aligned with LLMs [13, 42, 68, 77]. However, existing foundation models are usually pretrained on static text or text-image datasets and thus may struggle to align with the dynamics of the world. To bridge this gap, we study how to finetune a VLM to be an embodied agent dynamically aligned with the world by distilling the cross-modality knowledge from an LLM expert. The work most closely related to ours is EUREKA [40], which also explores using source information provided by the simulator as the context of an LLM to aid agent training. Instead of directly mimicking the output of the LLM as we did, EUREKA harnesses the coding LLM to generate a desired reward function for a given task and optimizes a policy against the reward function using RL, leading to a more complex and expensive training procedure [47, 52, 53].", + "bbox": [ + 75, + 285, + 472, + 694 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Imitation Learning. Imitation learning is the study of algorithms that improve performance by mimicking an expert's decisions and behaviors. We summarize three main categories of existing methods in the following: (1) behavior cloning (BC), (2) inverse reinforcement learning (IRL), and (3) the combination of imitation and reinforcement learning. The naive BC [5] ignores the change in distribution and simply trains a policy that only performs well under the distribution of states visited by the expert. Following works, such as dataset aggregation [51] or policy aggregation [28, 49], have been proposed to address the limitations of BC. Another line of work, IRL, is a more complicated algorithm framework that learns the reward function from", + "bbox": [ + 75, + 703, + 470, + 901 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "expert demonstrations and then improves the policy using RL with the learned reward. A representative method in this category is generative adversarial imitation learning (GAIL) [23], in which a policy and a discriminator compete with each other in order to maximize the likelihood of the policy's behavior matching the expert. The third category of methods usually leverages an IL policy to initialize RL and continues to boost its performance via online collected data from RL [8, 50, 58]. This simple combination significantly improves RL's sample efficiency and IL's upper performance bound constrained by the expert. Nevertheless, all of the above methods assume that the expert and the imitator understand the world in the same modality, and thus overlook the fact that the complementary knowledge from other modalities often boosts the model's accuracy and generalization dramatically [46, 67].", + "bbox": [ + 496, + 285, + 893, + 527 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 541, + 619, + 556 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We create EMMA, an Embodied Multi-Modal Agent, by finetuning a VLM in an embodied visual world with interactive imitation learning from an LLM expert in a parallel text world, who produces better actions and retrospective feedback to VLM's trajectories. Such imitation learning exhibits substantial advantages over vision or VLM policies directly finetuned in the visual world or finetuned by behavior cloning of a rule-based expert, and SOTA API VLMs such as GPT-4V(ison). As a result, EMMA achieves a comparable success rate and much better robustness in the noisy visual world than its LLM teacher in the easier text world. Furthermore, EMMA shows powerful generalization to open-vocabulary and free-form task instructions, highlighting its potential in real-world scenarios.", + "bbox": [ + 496, + 566, + 893, + 779 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgment", + "text_level": 1, + "bbox": [ + 500, + 792, + 651, + 809 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This work is partially supported by the National Key R&D Program of China under the Grant No. (2023YFE0106300) and (2017YFC0804002), Shenzhen Fundamental Research Program under the Grant No. (JCYJ20200109141235597), and Program for Guangdong Introducing Innovative and Entrepreneurial Teams under Grant No. (2017ZT07X386).", + "bbox": [ + 496, + 816, + 893, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "26282", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 119, + 173, + 136 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Constructions Aeronautiques, Adele Howe, Craig Knoblock, ISI Drew McDermott, Ashwin Ram, Manuela Veloso, Daniel Weld, David Wilkins SRI, Anthony Barrett, Dave Christianson, et al. PDDL: the planning domain definition language. Technical Report, Tech. Rep., 1998. 4, 5, 3", + "[2] Marcin Andrychowicz, Filip Wolski, Alex Ray, Jonas Schneider, Rachel Fong, Peter Welinder, Bob McGrew, Josh Tobin, OpenAI Pieter Abbeel, and Wojciech Zaremba. Hindsight experience replay. 2017. 2", + "[3] Shuang Ao, Tianyi Zhou, Guodong Long, Qinghua Lu, Liming Zhu, and Jing Jiang. Co-pilot: Collaborative planning and reinforcement learning on sub-task curriculum. 2021.", + "[4] Shuang Ao, Tianyi Zhou, Jing Jiang, Guodong Long, Xuan Song, and Chengqi Zhang. Eat-c: environment-adversarial sub-task curriculum for efficient reinforcement learning. In ICML, 2022. 2", + "[5] Michael Bain and Claude Sammut. A framework for behavioural cloning. In Machine Intelligence, 1995. 8", + "[6] Anthony Brohan, Noah Brown, Justice Carbajal, Yevgen Chebotar, Joseph Dabis, Chelsea Finn, Keerthana Gopalakrishnan, Karol Hausman, Alex Herzog, Jasmine Hsu, et al. Rt-1: Robotics transformer for real-world control at scale. arXiv preprint arXiv:2212.06817, 2022. 2, 4, 8", + "[7] Anthony Brohan, Yevgen Chebotar, Chelsea Finn, Karol Hausman, Alexander Herzog, Daniel Ho, Julian Ibarz, Alex Irpan, Eric Jang, Ryan Julian, et al. Do as i can, not as i say: Grounding language in robotic affordances. In CoRL, 2023. 8", + "[8] Kai-Wei Chang, Akshay Krishnamurthy, Alekh Agarwal, Hal Daumé III, and John Langford. Learning to search better than your teacher. In ICML, 2015. 8", + "[9] Lichang Chen, Jiuhai Chen, Tom Goldstein, Heng Huang, and Tianyi Zhou. Instructzero: Efficient instruction optimization for black-box large language models. arXiv preprint arXiv:2306.03082, 2023. 4", + "[10] Marc-Alexandre Côté, Akos Kádár, Xingdi Yuan, Ben Kybartas, Tavian Barnes, Emery Fine, James Moore, Matthew Hausknecht, Layla El Asri, Mahmoud Adada, et al. Textworld: A learning environment for text-based games. In Computer Games Workshop, IJCAI, 2019. 4, 5", + "[11] Wenliang Dai, Junnan Li, Dongxu Li, Anthony Tiong, Junqi Zhao, Weisheng Wang, Boyang Li, Pascale Fung, and Steven Hoi. InstructBLIP: Towards general-purpose vision-language models with instruction tuning. In NeurIPS, 2023. 2, 3, 5, 8", + "[12] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. BERT: pre-training of deep bidirectional transformers for language understanding. In NAACL-HLT, 2019. 3", + "[13] Danny Driess, Fei Xia, Mehdi SM Sajjadi, Corey Lynch, Aakanksha Chowdhery, Brian Ichter, Ayzaan Wahid, Jonathan Thompson, Quan Vuong, Tianhe Yu, et al. Palm-e: An embodied multimodal language model. arXiv preprint arXiv:2303.03378, 2023. 2, 4, 8" + ], + "bbox": [ + 78, + 146, + 472, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[14] Meng Fang, Tianyi Zhou, Yali Du, Lei Han, and Zhengyou Zhang. Curriculum-guided hindsight experience replay. NeurIPS, 2019. 2", + "[15] Daocheng Fu, Xin Li, Licheng Wen, Min Dou, Pinlong Cai, Botian Shi, and Yu Qiao. Drive like a human: Rethinking autonomous driving with large language models. arXiv preprint arXiv:2307.07162, 2023. 2, 4", + "[16] Jensen Gao, Bidipta Sarkar, Fei Xia, Ted Xiao, Jiajun Wu, Brian Ichter, Anirudha Majumdar, and Dorsa Sadigh. Physically grounded vision-language models for robotic manipulation. arXiv preprint arXiv:2309.02561, 2023. 8", + "[17] Peng Gao, Jiaming Han, Renrui Zhang, Ziyi Lin, Shijie Geng, Aojun Zhou, Wei Zhang, Pan Lu, Conghui He, Xiangyu Yue, et al. Llama-adapter v2: Parameter-efficient visual instruction model. arXiv preprint arXiv:2304.15010, 2023. 5, 8, 3", + "[18] Jiayi Guan, Guang Chen, Jiaming Ji, Long Yang, Zhijun Li, et al. Voce: Variational optimization with conservative estimation for offline safe reinforcement learning. NeurIPS, 36, 2023. 2", + "[19] Caglar Gülcehre, Sungjin Ahn, Ramesh Nallapati, Bowen Zhou, and Yoshua Bengio. Pointing the unknown words. In ACL, 2016. 5", + "[20] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, 2016. 5", + "[21] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross B. Girshick. Mask R-CNN. In ICCV, 2017. 5", + "[22] Malte Helmert. The fast downward planning system. Journal of Artificial Intelligence Research, 2006. 3", + "[23] Jonathan Ho and Stefano Ermon. Generative adversarial imitation learning. In NeurIPS, 2016. 8", + "[24] Wenlong Huang, Pieter Abbeel, Deepak Pathak, and Igor Mordatch. Language models as zero-shot planners: Extracting actionable knowledge for embodied agents. In ICML, 2022. 8", + "[25] Wenlong Huang, Chen Wang, Ruohan Zhang, Yunzhu Li, Jiajun Wu, and Li Fei-Fei. Voxposer: Composable 3d value maps for robotic manipulation with language models. arXiv preprint arXiv:2307.05973, 2023. 2, 4, 8", + "[26] Ahmed Hussein, Mohamed Medhat Gaber, Eyad Elyan, and Chrisina Jayne. Imitation learning: A survey of learning methods. ACM Computing Surveys (CSUR), 50(2):1-35, 2017. 4", + "[27] Marcus Hutter. Universal artificial intelligence: Sequential decisions based on algorithmic probability. Springer Science & Business Media, 2004. 1", + "[28] Hal Daumé III, John Langford, and Daniel Marcu. Search-based structured prediction. Mach. Learn., 2009. 8", + "[29] Yunfan Jiang, Agrim Gupta, Zichen Zhang, Guanzhi Wang, Yongqiang Dou, Yanjun Chen, Li Fei-Fei, Anima Anandkumar, Yuke Zhu, and Linxi Fan. VIMA: robot manipulation with multimodal prompts. In ICML, 2023. 2, 4", + "[30] Eric Kolve, Roozbeh Mottaghi, Winson Han, Eli VanderBilt, Luca Weihs, Alvaro Herrasti, Matt Deitke, Kiana Ehsani, Daniel Gordon, Yuke Zhu, et al. Ai2-thor: An interactive 3d environment for visual ai. arXiv preprint arXiv:1712.05474, 2017. 4, 5, 1" + ], + "bbox": [ + 501, + 92, + 893, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "26283", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[31] Sergey Levine, Aviral Kumar, George Tucker, and Justin Fu. Offline reinforcement learning: Tutorial, review, and perspectives on open problems. arXiv preprint arXiv:2005.01643, 2020. 2, 3", + "[32] Chunyuan Li, Zhe Gan, Zhengyuan Yang, Jianwei Yang, Linjie Li, Lijuan Wang, and Jianfeng Gao. Multimodal foundation models: From specialists to general-purpose assistants. arXiv preprint arXiv:2309.10020, 2023. 2, 3, 8", + "[33] Dongxu Li, Junnan Li, Hung Le, Guangsen Wang, Silvio Savarese, and Steven C.H. Hoi. LAVIS: A one-stop library for language-vision intelligence. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 3: System Demonstrations), 2023. 3", + "[34] Junnan Li, Dongxu Li, Silvio Savarese, and Steven C. H. Hoi. BLIP-2: bootstrapping language-image pre-training with frozen image encoders and large language models. In ICML, 2023. 2, 3, 5, 8", + "[35] Chen Liang, Jiahui Yu, Ming-Hsuan Yang, Matthew Brown, Yin Cui, Tuo Zhao, Boqing Gong, and Tianyi Zhou. Modulewise adaptive distillation for multimodality foundation models. 2024. 2", + "[36] Jacky Liang, Wenlong Huang, Fei Xia, Peng Xu, Karol Hausman, Brian Ichter, Pete Florence, and Andy Zeng. Code as policies: Language model programs for embodied control. In ICRA, 2023. 8", + "[37] Yong Lin, Lu Tan, Hangyu Lin, Zeming Zheng, Renjie Pi, Jipeng Zhang, Shizhe Diao, Haoxiang Wang, Han Zhao, Yuan Yao, et al. Speciality vs generality: An empirical study on catastrophic forgetting in fine-tuning foundation models. arXiv preprint arXiv:2309.06256, 2023. 3", + "[38] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. arXiv preprint arXiv:2304.08485, 2023. 2, 3", + "[39] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 3", + "[40] Yecheng Jason Ma, William Liang, Guanzhi Wang, De-An Huang, Osbert Bastani, Dinesh Jayaraman, Yuke Zhu, Linxi Fan, and Anima Anandkumar. Eureka: Human-level reward design via coding large language models. arXiv preprint arXiv:2310.12931, 2023. 8", + "[41] Vincent Micheli and François Fleuret. Language models are few-shot butlers. In EMNLP, 2021. 5, 8", + "[42] Yao Mu, Qinglong Zhang, Mengkang Hu, Wenhai Wang, Mingyu Ding, Jun Jin, Bin Wang, Jifeng Dai, Yu Qiao, and Ping Luo. Embodiedgpt: Vision-language pre-training via embodied chain of thought. arXiv preprint arXiv:2305.15021, 2023. 2, 4, 8", + "[43] OpenAI. GPT-4V(ison) system card. https://cdn.openai.com/papers/GPTV_System_Card.pdf, 2023.2", + "[44] Abhishek Padalkar, Acorn Pooley, Ajinkya Jain, Alex Bewley, Alex Herzog, Alex Irpan, Alexander Khazatsky, Anant Rai, Anikait Singh, Anthony Brohan, et al. Open x-embediment: Robotic learning datasets and rt-x models, 2023. 3", + "[45] Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, Ilya Sutskever, et al. Language models are unsupervised multitask learners. OpenAI blog, 2019. 5" + ], + "bbox": [ + 78, + 90, + 470, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[46] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, 2021. 8, 3", + "[47] Rafael Rafailov, Archit Sharma, Eric Mitchell, Stefano Ermon, Christopher D Manning, and Chelsea Finn. Direct preference optimization: Your language model is secretly a reward model. arXiv preprint arXiv:2305.18290, 2023. 8", + "[48] Rafael Rafailov, Archit Sharma, Eric Mitchell, Christopher D Manning, Stefano Ermon, and Chelsea Finn. Direct preference optimization: Your language model is secretly a reward model. NeurIPS, 2024. 2, 4", + "[49] Stéphane Ross and Drew Bagnell. Efficient reductions for imitation learning. In AISTATS, 2010. 8", + "[50] Stephane Ross and J Andrew Bagnell. Reinforcement and imitation learning via interactive no-regret learning. arXiv preprint arXiv:1406.5979, 2014. 8", + "[51] Stéphane Ross, Geoffrey J. Gordon, and Drew Bagnell. A reduction of imitation learning and structured prediction to no-regret online learning. In AISTATS, 2011. 2, 4, 8", + "[52] John Schulman, Sergey Levine, Pieter Abbeel, Michael I. Jordan, and Philipp Moritz. Trust region policy optimization. In ICML, 2015. 8", + "[53] John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Radford, and Oleg Klimov. Proximal policy optimization algorithms. arXiv preprint arXiv:1707.06347, 2017. 8", + "[54] Noah Shinn, Federico Cassano, Ashwin Gopinath, Karthik R Narasimhan, and Shunyu Yao. Reflexion: Language agents with verbal reinforcement learning. In NeurIPS, 2023. 3, 4, 5, 6, 8", + "[55] Mohit Shridhar, Jesse Thomason, Daniel Gordon, Yonatan Bisk, Winson Han, Roozbeh Mottaghi, Luke Zettlemoyer, and Dieter Fox. ALFRED: A benchmark for interpreting grounded instructions for everyday tasks. In CVPR, 2020. 5, 8, 3, 6", + "[56] Mohit Shridhar, Xingdi Yuan, Marc-Alexandre Côté, Yonatan Bisk, Adam Trischler, and Matthew J. Hausknecht. Alfworld: Aligning text and embodied environments for interactive learning. In ICLR, 2021. 2, 3, 4, 5, 8, 1", + "[57] Theodore Sumers, Kenneth Marino, Arun Ahuja, Rob Fergus, and Ishita Dasgupta. Distilling internet-scale vision-language models into embodied agents. In ICML, 2023. 2, 4, 8", + "[58] Wen Sun, J. Andrew Bagnell, and Byron Boots. Truncated horizon policy search: Combining reinforcement learning & imitation learning. In ICLR, 2018. 8", + "[59] Guanzhi Wang, Yuqi Xie, Yunfan Jiang, Ajay Mandlekar, Chaowei Xiao, Yuke Zhu, Linxi Fan, and Anima Anandkumar. Voyager: An open-ended embodied agent with large language models. arXiv preprint arXiv:2305.16291, 2023. 8", + "[60] Xiyao Wang, Yuhang Zhou, Xiaoyu Liu, Hongjin Lu, Yuancheng Xu, Feihong He, Jaehong Yoon, Taixi Lu, Gedas Bertasius, Mohit Bansal, et al. Mementos: A comprehensive benchmark for multimodal large language model reasoning over image sequences. arXiv preprint arXiv:2401.10529, 2024. 2" + ], + "bbox": [ + 501, + 92, + 893, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "26284", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[61] Zihao Wang, Shaofei Cai, Anji Liu, Xiaojian Ma, and Yitao Liang. Describe, explain, plan and select: Interactive planning with large language models enables open-world multi-task agents. arXiv preprint arXiv:2302.01560, 2023. 5, 6", + "[62] Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Brian Ichter, Fei Xia, Ed H. Chi, Quoc V. Le, and Denny Zhou. Chain-of-thought prompting elicits reasoning in large language models. In NeurIPS, 2022. 3", + "[63] Michael Wooldridge and Nicholas R Jennings. Intelligent agents: Theory and practice. The knowledge engineering review, 10(2):115-152, 1995. 1", + "[64] Qingyun Wu, Gagan Bansal, Jieyu Zhang, Yiran Wu, Shaokun Zhang, Erkang Zhu, Beibin Li, Li Jiang, Xiaoyun Zhang, and Chi Wang. Autogen: Enabling next-gen llm applications via multi-agent conversation framework. arXiv preprint arXiv:2308.08155, 2023. 5, 6, 8", + "[65] Zhiheng Xi, Wenxiang Chen, Xin Guo, Wei He, Yiwen Ding, Boyang Hong, Ming Zhang, Junzhe Wang, Senjie Jin, Enyu Zhou, et al. The rise and potential of large language model based agents: A survey. arXiv preprint arXiv:2309.07864, 2023. 8", + "[66] Zhenhua Xu, Yujia Zhang, Enze Xie, Zhen Zhao, Yong Guo, Kenneth KY Wong, Zhenguo Li, and Hengshuang Zhao. Drivegpt4: Interpretable end-to-end autonomous driving via large language model. arXiv preprint arXiv:2310.01412, 2023. 2, 4", + "[67] Zihui Xue, Zhengqi Gao, Sucheng Ren, and Hang Zhao. The modality focusing hypothesis: Towards understanding cross-modal knowledge distillation. In ICLR, 2023. 8", + "[68] Jingkang Yang, Yuhao Dong, Shuai Liu, Bo Li, Ziyue Wang, Chencheng Jiang, Haoran Tan, Jiamu Kang, Yuanhan Zhang, Kaiyang Zhou, et al. Octopus: Embodied vision-language programmer from environmental feedback. arXiv preprint arXiv:2310.08588, 2023. 3, 4, 8", + "[69] Sherry Yang, Ofir Nachum, Yilun Du, Jason Wei, Pieter Abbeel, and Dale Schuurmans. Foundation models for decision making: Problems, methods, and opportunities. arXiv preprint arXiv:2303.04129, 2023. 8", + "[70] Yijun Yang, Jing Jiang, Tianyi Zhou, Jie Ma, and Yuhui Shi. Pareto policy pool for model-based offline reinforcement learning. In ICLR, 2022. 2", + "[71] Yijun Yang, Tianyi Zhou, Jing Jiang, Guodong Long, and Yuhui Shi. Continual task allocation in meta-policy network via sparse prompting. In ICML, 2023. 2", + "[72] Shunyu Yao, Dian Yu, Jeffrey Zhao, Izhak Shafran, Thomas L Griffiths, Yuan Cao, and Karthik Narasimhan. Tree of thoughts: Deliberate problem solving with large language models. arXiv preprint arXiv:2305.10601, 2023. 3", + "[73] Shunyu Yao, Jeffrey Zhao, Dian Yu, Nan Du, Izhak Shafran, Karthik R. Narasimhan, and Yuan Cao. React: Synergizing reasoning and acting in language models. In ICLR, 2023. 3, 5, 6, 8, 2", + "[74] Lianmin Zheng, Wei-Lin Chiang, Ying Sheng, Siyuan Zhuang, Zhanghao Wu, Yonghao Zhuang, Zi Lin, Zhuohan Li, Dacheng Li, Eric Xing, et al. Judging llm-as-a-judge with mt-bench and chatbot arena. arXiv preprint arXiv:2306.05685, 2023. 3" + ], + "bbox": [ + 78, + 90, + 470, + 898 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[75] Victor Zhong, Dipendra Misra, Xingdi Yuan, and Marc-Alexandre Côté. Policy improvement using language feedback models. arXiv preprint arXiv:2402.07876, 2024. 2", + "[76] Deyao Zhu, Jun Chen, Xiaogian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models. arXiv preprint arXiv:2304.10592, 2023. 2, 3, 5, 8", + "[77] Brianna Zitkovich, Tianhe Yu, Sichun Xu, Peng Xu, Ted Xiao, Fei Xia, Jialin Wu, Paul Wohlhart, Stefan Welker, Ayzaan Wahid, et al. Rt-2: Vision-language-action models transfer web knowledge to robotic control. In CoRL, 2023. 2, 3, 4, 8" + ], + "bbox": [ + 501, + 92, + 893, + 246 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "26285", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2024/Embodied Multi-Modal Agent trained by an LLM from a Parallel TextWorld/911fda92-445d-4ff5-8ef4-3ff0a7daf4bf_model.json b/2024/Embodied Multi-Modal Agent trained by an LLM from a Parallel TextWorld/911fda92-445d-4ff5-8ef4-3ff0a7daf4bf_model.json new file mode 100644 index 0000000000000000000000000000000000000000..b4d544113e2d976b04cc542b7e40cff33afe8233 --- /dev/null +++ b/2024/Embodied Multi-Modal Agent trained by an LLM from a Parallel TextWorld/911fda92-445d-4ff5-8ef4-3ff0a7daf4bf_model.json @@ -0,0 +1,2345 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.098, + 0.131, + 0.877, + 0.155 + ], + "angle": 0, + "content": "Embodied Multi-Modal Agent trained by an LLM from a Parallel TextWorld" + }, + { + "type": "text", + "bbox": [ + 0.182, + 0.179, + 0.788, + 0.2 + ], + "angle": 0, + "content": "Yijun Yang\\(^{1,5,4}\\) Tianyi Zhou\\(^{2}\\) Kanxue Li\\(^{3}\\) Dapeng Tao\\(^{3}\\) Lusong Li\\(^{4}\\)" + }, + { + "type": "text", + "bbox": [ + 0.272, + 0.199, + 0.725, + 0.218 + ], + "angle": 0, + "content": "Li Shen\\(^{4,*}\\) Xiaodong He\\(^{4}\\) Jing Jiang\\(^{5}\\) Yuhui Shi\\(^{1,*}\\)" + }, + { + "type": "text", + "bbox": [ + 0.129, + 0.218, + 0.842, + 0.236 + ], + "angle": 0, + "content": "\\(^{1}\\)Southern University of Science and Technology \\(^{2}\\)University of Maryland, College Park" + }, + { + "type": "text", + "bbox": [ + 0.158, + 0.236, + 0.812, + 0.255 + ], + "angle": 0, + "content": "\\(^{3}\\)Yunnan University \\(^{4}\\)JD Explore Academy \\(^{5}\\)University of Technology Sydney" + }, + { + "type": "text", + "bbox": [ + 0.281, + 0.257, + 0.685, + 0.271 + ], + "angle": 0, + "content": "https://github.com/stevenyangyj/Emma-Alfworld" + }, + { + "type": "image", + "bbox": [ + 0.1, + 0.279, + 0.311, + 0.457 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.108, + 0.46, + 0.303, + 0.476 + ], + "angle": 0, + "content": "Instruction: put the gold vase into safe." + }, + { + "type": "image", + "bbox": [ + 0.108, + 0.479, + 0.181, + 0.494 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.197, + 0.48, + 0.282, + 0.493 + ], + "angle": 0, + "content": "Target object" + }, + { + "type": "image", + "bbox": [ + 0.107, + 0.497, + 0.125, + 0.513 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.129, + 0.499, + 0.212, + 0.51 + ], + "angle": 0, + "content": "Action succeeded" + }, + { + "type": "image", + "bbox": [ + 0.107, + 0.514, + 0.124, + 0.529 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.129, + 0.514, + 0.277, + 0.525 + ], + "angle": 0, + "content": "Action failed due to hallucination" + }, + { + "type": "image", + "bbox": [ + 0.107, + 0.529, + 0.125, + 0.542 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.129, + 0.528, + 0.293, + 0.54 + ], + "angle": 0, + "content": "Action failed due to invalid grammar" + }, + { + "type": "image", + "bbox": [ + 0.318, + 0.281, + 0.443, + 0.366 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.445, + 0.284, + 0.549, + 0.365 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.552, + 0.284, + 0.657, + 0.364 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.658, + 0.284, + 0.762, + 0.364 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.764, + 0.284, + 0.871, + 0.366 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.317, + 0.372, + 0.443, + 0.455 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.445, + 0.373, + 0.55, + 0.455 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.552, + 0.373, + 0.657, + 0.455 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.658, + 0.373, + 0.763, + 0.455 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.764, + 0.372, + 0.871, + 0.455 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.318, + 0.461, + 0.443, + 0.544 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.445, + 0.462, + 0.55, + 0.544 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.552, + 0.462, + 0.658, + 0.544 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.659, + 0.462, + 0.764, + 0.544 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.765, + 0.462, + 0.871, + 0.544 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.56, + 0.895, + 0.618 + ], + "angle": 0, + "content": "Figure 1: Comparison of three VLM-based agents in the visual environment from ALFWorld. Given the task instruction and the current-step observation as inputs, a VLM agent is expected to predict an action, e.g., \"go to cabinet 1\", towards completing the task. (a) GPT-4V(ison). (b) InstructBLIP finetuned by behavior cloning on a static dataset of demonstrations by a rule-based expert in the visual simulator. (c) Our Embodied Multi-Modal Agent (EMMA) trained via cross-modality interactive imitation learning from an LLM expert." + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.632, + 0.314, + 0.648 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.664, + 0.473, + 0.877 + ], + "angle": 0, + "content": "While large language models (LLMs) excel in a simulated world of texts, they struggle to interact with the more realistic world without perceptions of other modalities such as visual or audio signals. Although vision-language models (VLMs) integrate LLM modules (1) aligned with static image features, and (2) may possess prior knowledge of world dynamics (as demonstrated in the text world), they have not been trained in an embodied visual world and thus cannot align with its dynamics. On the other hand, training an embodied agent in a noisy visual world without expert guidance is often challenging and inefficient. In this paper, we train a VLM agent living in a visual world using an LLM agent excelling in a parallel text world. Specifically, we distill LLM's reflection outcomes (improved actions by analyzing mistakes) in" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.634, + 0.895, + 0.8 + ], + "angle": 0, + "content": "a text world's tasks to finetune the VLM on the same tasks of the visual world, resulting in an Embodied Multi-Modal Agent (EMMA) quickly adapting to the visual world dynamics. Such cross-modality imitation learning between the two parallel worlds is achieved by a novel DAgger-DPO algorithm, enabling EMMA to generalize to a broad scope of new tasks without any further guidance from the LLM expert. Extensive evaluations on the ALFWorld benchmark's diverse tasks highlight EMMA's superior performance to SOTA VLM-based agents, e.g., \\(20\\% -70\\%\\) improvement in the success rate." + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.815, + 0.633, + 0.831 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.896, + 0.903 + ], + "angle": 0, + "content": "Embodied multi-modal agents have been acknowledged as a pivotal stride towards achieving Artificial General Intelligence (AGI), as they encompass the potential for a broad scope of intelligent activities [27, 63]. The rising of founda" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.888, + 0.221, + 0.9 + ], + "angle": 0, + "content": "*Corresponding author" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "26275" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.471, + 0.182 + ], + "angle": 0, + "content": "tion models brings a glimmer of hope for constructing such agents [11, 32, 34, 35, 38, 71, 76], and notable efforts from the community have tried to harness them in many decision-making scenarios, e.g., autonomous driving [15, 66], daily household robots [6, 13, 77], and complex manipulation tasks [25, 29, 42, 57]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.185, + 0.474, + 0.456 + ], + "angle": 0, + "content": "LLMs can perform as reflex agents interacting with a text world via verbalized descriptions of the world states and textual actions. In addition, they exhibit great potential in planning, reflection, and reward shaping. This is mainly due to their prior knowledge and semantic abstraction of the world. However, LLM-based agents cannot be directly applied in a visual world. Although current vision-language models (VLMs) try to align LLMs with the visual modality, their pretraining only focuses on static alignment between image-text pairs so the resulting agents have not been aligned well with the dynamics of the visual world. As shown in Fig. 1 (a), even the privileged VLM, e.g., GPT-4V [43], fails to accomplish tasks in an embodied ALFWorld environment [56]. In such a zero-shot setting, GPT-4V tends to mainly rely on its language prior of these objects detected in the current step, rather than the alignment between the visual input and the environment dynamics conditioned on the task instruction." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.46, + 0.474, + 0.718 + ], + "angle": 0, + "content": "In this paper, we study how to train a VLM towards an embodied agent in a visual world by aligning it with the visual world dynamics and distilling the skills of an LLM agent in a parallel text world. Our overarching goal is to build such an Embodied Multi-Modal Agent (EMMA) that can take a textual task instruction (e.g., from human users) and pixel observations of the state per step to produce a sequence of actions leading to the efficient completion of the task. This is a challenging problem due to (1) the sparsity of task reward [2-4, 14], (2) noisy visual representations, (3) the hallucination of VLMs [60], and (4) the misalignment of VLM's static representations to the visual world dynamics. While offline distillation and imitation from a powerful LLM agent in a parallel text world can potentially overcome the former two challenges [75], effectively mitigating the remaining challenges necessitates online finetuning of the VLM agent within an interactive and visual world [18, 31, 70]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.72, + 0.474, + 0.903 + ], + "angle": 0, + "content": "To this end, we finetune a VLM agent by imitation learning from an LLM expert (e.g., built on ChatGPT) launched in a parallel text world on the same tasks. Specifically, in each step of EMMA interacting with the visual world, we convert its visual observation into an equivalent textual description sent to the LLM agent, which produces an action for EMMA to imitate. Such cross-modality interactive imitation learning is based on DAgger [51], which overcomes the cumulative errors and distribution shifts caused by behavior cloning (BC). As depicted in Fig. 1 (b), an InstructBLIP [11] agent finetuned by BC on 170K expert demonstrations produced by a rule-based expert in the" + }, + { + "type": "image", + "bbox": [ + 0.504, + 0.089, + 0.891, + 0.284 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.296, + 0.893, + 0.352 + ], + "angle": 0, + "content": "Figure 2: An example of tasks generated for the two parallel worlds. A VLM agent in the visual world and an LLM agent in the text world as household robots instructed to clean an apple and then put it into the fridge. Zoom in for more details." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.378, + 0.893, + 0.559 + ], + "angle": 0, + "content": "visual world still fails to take correct actions based on visual observations. We further improve the DAgger's objective to be the direct preference optimization (DPO) [48], which maximizes the preference of LLM-expert's action (positive) over VLM-student's action (negative) in each interaction step. To collect better teaching signals retrospective to the VLM student's actions, the LLM expert is composed of an LLM actor prompted to output expert actions, and an LLM critic prompted for reflection feedback on the VLM agent's historical trajectories. We maintain a long-term memory storing the feedback, which is then used to induce the LLM actor to improve actions for imitation in future episodes." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.56, + 0.895, + 0.727 + ], + "angle": 0, + "content": "We evaluate EMMA and compare it with vision-only agents, LLM agents, and VLM agents deployed to the ALF-World benchmark [56], which includes numerous tasks in both visual and textual environments. Extensive evaluations highlight that EMMA substantially outperforms state-of-the-art (SOTA) VLM agents in visual-only environments by \\(20\\% - 70\\%\\) in terms of success rate. In addition, EMMA is the only VLM agent that can generalize to open-vocabulary and free-form test tasks, shedding novel insights on using LLM feedback to train more versatile and generalizable embodied agents in multi-modality environments." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.739, + 0.78, + 0.758 + ], + "angle": 0, + "content": "2. Embodied Multi-Modal Agent" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.897, + 0.903 + ], + "angle": 0, + "content": "Fig. 3 illustrates the main idea of our \"Embodied MultiModal Agent (EMMA)\", whose detailed training procedures are given in Alg. 1. The agent is built upon a modularized VLM, which can follow instructions and interact with the environment through pixel observations and textual actions. To overcome these challenges associated with training EMMA, such as sparse reward or distribution shift, we explore the construction of an LLM expert from a parallel text world (Sec. 2.2) for providing EMMA with step-by-step guidance." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "26276" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.101, + 0.094, + 0.873, + 0.313 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.329, + 0.895, + 0.375 + ], + "angle": 0, + "content": "Figure 3: Embodied Multi-Modal Agent (EMMA) trained by an LLM expert via cross-modality imitation learning. EMMA takes a textual task instruction and pixel observations as its input state per step to generate a sequence of actions using a VLM. Then, we convert each pixel observation into a textual equivalent as the context of an LLM expert to produce improved actions for EMMA to imitate." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.398, + 0.471, + 0.43 + ], + "angle": 0, + "content": "Lastly, Sec. 2.3 further discusses how to harness the LLM expert to train EMMA via cross-modality imitation learning." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.44, + 0.327, + 0.456 + ], + "angle": 0, + "content": "2.1. EMMA in the Visual World" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.466, + 0.471, + 0.812 + ], + "angle": 0, + "content": "In visual environments, EMMA \\(\\pi_{\\theta}\\) is designed to process a task instruction \\(x_{\\mathrm{task}}\\) (e.g., provided by human users) and the pixel observation \\(s_v^t\\) at each interaction step \\(t\\). It is expected to generate a sequence of high-level textual actions \\(\\{x_a^t \\sim \\pi_\\theta(\\cdot | x_{\\mathrm{task}}, s_v^t)\\}_{t=0}^T\\) towards efficient completion of the task. To achieve this, we draw inspiration from recent advances of large pretrained VLMs [11, 32, 34, 38, 76], and modularize EMMA's architecture into four components: (1) a ViT to encode \\(s_v\\) into visual embeddings, (2) a querying transformer (Q-Former) tailored to extract the most relevant visual features via the cross-attention between the visual embeddings and query tokens, (3) a linear projection layer to align visual features to text embeddings, (4) an LLM decoder taking the concatenation of the instruction tokens and the output of the linear projection layer to autoregressively generate the action \\(x_a\\). In order to reduce computational overhead and prevent catastrophic forgetting [37], we adopt the pretrained ViT, Q-Former, and LLM from InstructBLIP [11] and keep them frozen at the finetuning stage. We only update the linear projection layer, as illustrated in Fig. 3. Such a modularized architecture enables EMMA to integrate any existing pretrained vision models and LLMs in a flexible and computationally-efficient way." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.813, + 0.471, + 0.875 + ], + "angle": 0, + "content": "However, deploying EMMA into a complex visual world remains challenging. One of the main obstacles is that the direct use of any pretrained VLM is suboptimal because existing pretraining only focuses on static alignment between" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.398, + 0.895, + 0.699 + ], + "angle": 0, + "content": "image-text pairs [11, 32, 34], so that the resulting agent may struggle to reason about the dynamics of the world. As discussed in Sec. 1, even the SOTA VLM, i.e., GPT-4V, fails to accomplish tasks in an embodied ALFWorld environment [56]. In such a zero-shot setting, GPT-4V tends to rely on the linguistic prior of the currently detected objects, rather than the given task instruction and the underlying dynamics of the environment, to guide the interactions. Moreover, finetuning a pretrained VLM on a pre-collected demonstration dataset is also suboptimal due to the diversity of environments and tasks [77], the lack of large-scale expert annotations [44] as well as the challenges posed by the distribution shift issue [31]. A seemingly natural solution to the above challenges is reinforcement learning from environmental feedback (RLEF) [68], in which reward signals rely on decomposing a task into a sequence of reasonable sub-goals and checking their completion. However, in real-world scenarios, most sub-goals cannot be defined or described precisely, so the reward is sparse; hence, we do not expect RLEF to be effective." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.7, + 0.895, + 0.821 + ], + "angle": 0, + "content": "To this end, we propose to leverage interactive imitation learning (IL) to align EMMA with the dynamics of any environment, which however results in two critical algorithmic challenges: (1) How to obtain a high-quality, accessible, and scalable expert that EMMA can query during IL (Sec. 2.2)? (2) Designing an effective strategy to train EMMA using this expert in complex, diverse, and potentially open-ended environments (Sec. 2.3)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.832, + 0.842, + 0.848 + ], + "angle": 0, + "content": "2.2. LLM Expert from a Parallel TextWorld" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Thanks to a series of prompting techniques in the realm of in-context learning, such as chain-of-thought [62], tree-of-thought [72], ReAct [73], and Reflexion [54], pretrained" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.887, + 0.365, + 0.902 + ], + "angle": 0, + "content": "For brevity, we omit \\( x_{\\mathrm{task}} \\) in the rest of this paper." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.958 + ], + "angle": 0, + "content": "26277" + } + ], + [ + { + "type": "code_caption", + "bbox": [ + 0.08, + 0.095, + 0.451, + 0.11 + ], + "angle": 0, + "content": "Algorithm 1 DAgger-DPO with a single task instruction" + }, + { + "type": "algorithm", + "bbox": [ + 0.083, + 0.112, + 0.47, + 0.395 + ], + "angle": 0, + "content": "1: initialize: \\(i = 0\\) \\(\\mathcal{D}\\gets \\emptyset\\) , EMMA \\(\\pi_{\\theta}\\) , LLM actor \\(M_a\\) with a FIFO memory pool \\(\\mathcal{P}\\gets \\emptyset\\) , LLM critic \\(M_c\\) \n2: input: max trials \\(I\\) , training epochs \\(I_{e}\\) , visual and textual envs \\(E_{v},E_{l}\\) , instruction \\(x_{\\mathrm{task}}\\) , reference agent \\(\\pi_{\\mathrm{ref}}\\) \n3: Initialize \\(\\pi_{\\theta}\\) to \\(\\pi_{\\mathrm{ref}}\\) Behavior Cloning Initialization \n4: while task not completed or \\(i < I\\) do \n5: Get \\(\\tau_v^i = [x_{\\mathrm{task}},s_v^0,x_a^0,\\dots,s_v^T,x_a^T]\\) via \\(E_{v}\\) with \\(\\pi_{\\theta}\\) \n6: Get \\(\\tau_l^i = [x_{\\mathrm{task}},s_l^0,x_a^0,\\dots,s_l^T,x_a^T]\\) via \\(E_{l}\\) with \\(\\tau_v^i\\) \n7: Generate retrospective feedback \\(\\mathcal{P}_i = M_c(\\tau_l^i)\\) \n8: Update \\(\\mathcal{P}\\) with \\(\\mathcal{P}_i\\) (i.e., \\(\\mathcal{P}\\leftarrow \\mathcal{P}\\cup \\mathcal{P}_i\\) \n9: for \\(t = 0\\) to \\(T\\) do Dataset Aggregation \n10: \\(x_{a}^{*} = M_{a}(\\mathcal{P},x_{\\mathrm{task}},\\ldots ,x_{a}^{t - 1},s_{l}^{t})\\) \n11: \\(\\mathcal{D}\\gets \\mathcal{D}\\cup \\{(x_{\\mathrm{task}},s_v^t,x_a^t,x_a^*)\\}\\) \n12: for \\(j = 0\\) to \\(I_{e} - 1\\) do Gradient Descent on \\(\\theta\\) \n13: Sample a mini-batch \\(\\tau\\) from \\(\\mathcal{D}\\) \n14: Update \\(\\theta\\) by minimizing Eq. (2) with \\(\\pi_{\\mathrm{ref}}\\) and \\(\\tau\\) \n15: \\(i\\gets i + 1\\) \n16: output: \\(\\pi_{\\theta^{*}}\\)" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.424, + 0.473, + 0.712 + ], + "angle": 0, + "content": "LLMs have demonstrated impressive zero-shot performance across many decision-making scenarios [6, 13, 15, 25, 29, 42, 57, 66, 77]. Despite the great potential in serving as high-quality and scalable experts, they are only able to interact with the environment via textual descriptions of the states, rather than using raw pixel observations like EMMA. To bridge this gap, we convert each pixel observation \\( s_v \\) into a textual equivalent by extracting its metadata from the simulator [30], which is composed of attributes such as Observed Objects, Observed Relations, Inventory, and Locations. We then employ the Planning Domain Definition Language (PDDL) [1] to describe this metadata and create an equivalent textual description/state \\( s_l \\) using the TextWorld engine [10]. Additional details are available in Appendix 7 and an example of this process is illustrated in Fig. 2. This methodology enables the utilization of any pretrained LLM agent that generates a sequence of actions facilitating the training of EMMA through cross-modality IL between the two agents." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.722, + 0.47, + 0.737 + ], + "angle": 0, + "content": "2.3. Training EMMA via Cross-Modality Imitation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.745, + 0.47, + 0.807 + ], + "angle": 0, + "content": "Given an LLM expert from the parallel TextWorld, we aim to train a VLM agent \\(\\pi_{\\theta}\\) in the visual world to imitate its behaviors closely. This is equal to minimizing the following objective under the distribution of states induced by \\(\\pi_{\\theta}\\)." + }, + { + "type": "equation", + "bbox": [ + 0.134, + 0.819, + 0.47, + 0.843 + ], + "angle": 0, + "content": "\\[\n\\theta^ {*} = \\underset {\\theta \\in \\Theta} {\\arg \\min } \\mathbb {E} _ {\\pi_ {\\theta}} \\left[ \\mathcal {L} _ {\\mathrm {i m i t}} \\left(\\pi_ {\\theta} \\left(x _ {a} \\mid s _ {v}\\right), x _ {a} ^ {*}\\right)\\right) ], \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.473, + 0.902 + ], + "angle": 0, + "content": "in which the choice of loss function \\(\\mathcal{L}_{\\mathrm{imit}}\\) is dependent on specific scenarios. For instance, it may be the expected cross-entropy loss for the discrete action space, or the expected" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.168 + ], + "angle": 0, + "content": "MSE loss for the continuous action space. In our case, we select DPO [48] loss due to its proven superior performance to the cross-entropy on aligning models with expert preferences within the discrete language space. Hence, Eq. (1) can be extended to the formulation below." + }, + { + "type": "equation", + "bbox": [ + 0.506, + 0.179, + 0.893, + 0.204 + ], + "angle": 0, + "content": "\\[\n\\theta^ {*} = \\underset {\\theta \\in \\Theta} {\\arg \\min } - \\mathbb {E} _ {\\pi_ {\\theta}} \\left[ \\mathcal {L} _ {\\operatorname {i m i t}} \\left(\\pi_ {\\theta}, \\pi_ {\\text {r e f}}, s _ {v}, x _ {a}, x _ {a} ^ {*}\\right) \\right], \\tag {2}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.507, + 0.207, + 0.884, + 0.242 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {i m i t}} (\\cdot) \\triangleq \\log \\sigma \\left(\\beta \\log \\frac {\\pi_ {\\theta} \\left(x _ {a} ^ {*} \\mid s _ {v}\\right)}{\\pi_ {\\mathrm {r e f}} \\left(x _ {a} ^ {*} \\mid s _ {v}\\right)} - \\beta \\log \\frac {\\pi_ {\\theta} \\left(x _ {a} \\mid s _ {v}\\right)}{\\pi_ {\\mathrm {r e f}} \\left(x _ {a} \\mid s _ {v}\\right)}\\right)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.25, + 0.895, + 0.552 + ], + "angle": 0, + "content": "where \\( x_{a}^{*} \\) is the action given by an expert while \\( x_{a} \\) is the action given by \\( \\pi_{\\theta} \\), \\( \\sigma \\) is the logistic function, and \\( \\beta \\) is a hyperparameter controlling the deviation from \\( \\pi_{\\mathrm{ref}} \\), i.e. the reference agent obtained by behavior cloning on a demonstration dataset produced by a rule-based expert in the visual world (see Appendix 9 for complete details). This regularization is essential, as it prevents the agent from deviating too far from the distribution on which the expert is accurate, as well as maintaining the generation diversity and avoiding premature convergence to some easy tasks [48]. In practice, the VLM agent \\( \\pi_{\\theta} \\) is also initialized to \\( \\pi_{\\mathrm{ref}} \\) for stabilizing the training process. Since environmental dynamics is both unknown and complex, we cannot compute the distribution of states visited by \\( \\pi_{\\theta} \\) and can only sample it by rolling out the agent. Hence, Eq. (2) is a non-i.i.d. supervised learning problem due to the dependence of the state distribution on \\( \\pi_{\\theta} \\) itself, in which naive behavior cloning faces issues like cumulative error and distribution shift [26]. To address this, we employ an interactive IL algorithm, DAgger [51], which provably converges to the optimal agent \\( \\pi_{\\theta^{*}} \\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.553, + 0.895, + 0.794 + ], + "angle": 0, + "content": "As discussed in Sec. 2.2 and Sec. 2.3, we can harness an LLM expert to generate a sequence of actions, serving as \\( x_{a}^{*} \\) in Eq. (2). However, these actions may be suboptimal due to sparse environmental feedback [54, 68] or defective in-context instructions [9]. To collect better teaching signals, we introduce a \"retrospective LLM expert\" that is composed of two specialized models: an actor \\( (M_{a}) \\), built upon an API LLM and prompted to generate actions based on the task instruction and state observations; and a critic \\( (M_{c}) \\), also based on the same LLM, but designed to analyze EMMA's historical interactions and provide reflective feedback. A long-term memory \\( \\mathcal{P} \\) is maintained to store the feedback generated by \\( M_{c} \\), which is then used to prompt \\( M_{a} \\) for improved actions. The complete procedure is detailed in Line 7-10 of Alg. 1, and all prompts are provided in Sec. 6 of the Appendix." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.808, + 0.634, + 0.825 + ], + "angle": 0, + "content": "3. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.833, + 0.691, + 0.849 + ], + "angle": 0, + "content": "3.1. Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.894, + 0.901 + ], + "angle": 0, + "content": "**Environments.** We base our experiments on the ALFWorld benchmark [56], a cross-modality simulation platform that encompasses a wide range of embodied household tasks. For" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "26278" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.093, + 0.089, + 0.353, + 0.252 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.084, + 0.26, + 0.361, + 0.388 + ], + "angle": 0, + "content": "Figure 4: Comparison of success rate between EMMA and the retrospective LLM expert. As the number of trials increases, the LLM expert progressively improves its performance through retrospective processes, and also the gap between the two agents decreases, indicating the effectiveness of cross-modality imitation. We separately plot the curve per task type in Fig. 9 of Appendix." + }, + { + "type": "table", + "bbox": [ + 0.372, + 0.089, + 0.885, + 0.251 + ], + "angle": 0, + "content": "
AgentVisual Env.Textual Env.Success Rate with Template Task Instruction
Avg.PickCleanHeatCoolLookPick2
VMsResNet-18* [56]✓X0.06(-)------
MCNN-FPN* [56]✓X0.05(-)------
LMsBUTLER* [56]X✓0.26(-)0.31(-)0.41(-)0.60(-)0.27(-)0.12(-)0.29(-)
GPT-BUTLER [41]X✓0.69(18.8)0.62(18.4)0.81(18.4)0.85(12.7)0.78(15.6)0.50(24.4)0.47(26.6)
ReAct [73]X✓0.54(20.6)0.71(18.1)0.65(18.8)0.62(18.2)0.44(23.2)0.28(23.7)0.35(25.5)
Reflexion [54]X✓0.91(18.7)0.96(17.4)1.00(17.0)0.81(19.4)0.83(21.6)0.94(16.9)0.88(21.6)
DEPS* [61]X✓0.76(-)0.93(-)0.50(-)0.80(-)1.00(-)1.00(-)0.00(-)
AutoGen* [64]X✓0.77(-)0.92(-)0.74(-)0.78(-)0.86(-)0.83(-)0.41(-)
VLMsMiniGPT-4 [76]✓X0.16(26.9)0.04(29.0)0.00(30.0)0.19(26.3)0.17(26.7)0.67(17.7)0.06(28.9)
BLIP-2 [34]✓X0.04(29.5)0.00(30.0)0.06(29.3)0.04(29.9)0.11(28.2)0.06(29.2)0.00(30.0)
LLaMA-Adapter [17]✓X0.13(27.5)0.17(26.4)0.10(28.6)0.27(24.2)0.22(26.7)0.00(30.0)0.00(30.0)
InstructBLIP [11]✓X0.22(26.2)0.50(21.5)0.26(25.0)0.23(27.2)0.06(28.9)0.17(26.8)0.00(30.0)
EMMA (Ours)✓X0.82(19.5)0.71(19.3)0.94(17.5)0.85(19.6)0.83(19.9)0.88(19.6)0.67(22.4)
Human Performance* [55]✓X0.91(-)------
" + }, + { + "type": "table_caption", + "bbox": [ + 0.368, + 0.26, + 0.887, + 0.386 + ], + "angle": 0, + "content": "Table 1: Comparison with the state of the arts. \\(*\\)-reported in previous work. VMs = vision models, LMs = language models, VLMs = vision-language models. \"Visual Env.\" and \"Textual Env.\" refer to the visual and the parallel textual environments from ALFWorld [56], respectively. \\(\\sqrt{} / X\\) denotes the corresponding environment used/not used to deploy the agent. The highest scores for each task in the same type of environment are highlighted in bold. The average interaction steps are given in the (parentheses). EMMA substantially outperforms other SOTA VLM agents in the visual environments, and its success also directs a promising way to achieve human-level performance in ALFWorld." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.412, + 0.473, + 0.775 + ], + "angle": 0, + "content": "each task, ALFWorld integrates a visual environment, rendered by the Ai2Thor simulator [30], with a corresponding textual environment. Textual environments employ the Planning Domain Definition Language (PDDL) [1] to translate each pixel observation from the simulator into an equivalent text-based observation, and then construct an interactive world using the TextWorld engine [10]. Fig. 2 provides an illustrative example of the tasks featured in ALFWorld. The tasks within the benchmark are categorized into 6 types: Pick & Place, Clean & Place, Heat & Place, Cool & Place, Look in Light, and Pick Two Objects & Place. Each task requires an agent to execute a series of text-based actions, such as \"go to safe 1\", \"open safe 1\", or \"heat egg 1 with microwave 1\", following a predefined instruction. These actions involve navigating and interacting with the environment. To provide a comprehensive understanding, we have visualized an example of each task type in Fig. 8 of the Appendix. A task in this benchmark may involve interactions with over 10 objects and require more than 30 steps for a human expert to solve, thus challenging an agent's capabilities in long-horizon planning, instruction following, and the utilization of commonsense knowledge. For a fair comparison, we follow the same setting as prior work [41, 54-56, 73] and evaluate all baselines using 134 out-of-distribution (OOD) tasks." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.78, + 0.473, + 0.902 + ], + "angle": 0, + "content": "Baselines. To verify the effectiveness of cross-modality imitation learning, we compare our EMMA with several baselines and SOTA agents using the ALFWorld benchmark with both visual and textual environments. The compared agents can be divided into three categories: vision models, language models, and vision-language models. Concretely, vision models, including ResNet-18 [20] and MCNN-FPN [21], utilize pretrained vision encoders to extract salient features" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.412, + 0.896, + 0.898 + ], + "angle": 0, + "content": "from each pixel observation. The extracted features then serve as input for a Multi-Layer Perceptron (MLP) policy, which is trained by behavior cloning on a pre-collected demonstration dataset. Unlike vision models performing in the visual environment, language models complete the same tasks but in a parallel, text-based environment. BUTLER [56] employs a transformer seq2seq model enhanced with a pointer softmax mechanism [19]. This architecture aggregates previous observations as input to generate text-based actions token-by-token. GPT-BUTLER [41], a variant of the GPT-2 model [45], is initially pretrained on a static demonstration dataset and further finetuned using data collected online. ReAct [73] takes a novel approach by utilizing LLMs to generate reasoning traces and task-specific actions in an interleaved manner. This method aids the agent in developing, tracking, and updating its action plans interactively. Reflexion [54] similarly employs an LLM, but it focuses on reflecting upon environmental feedback. It maintains this reflective text in an episodic memory buffer, enhancing the agent's ability to improve actions in subsequent trials. Similar to Reflexion, a concurrent work, DEPS [61], also corrects errors in previous LLM-generated actions by integrating descriptions of the action execution process and providing self-explanations for the feedback. Moreover, beyond the single-agent framework, AutoGen [64] exhibits the potential to accomplish a broad spectrum of tasks through the cooperation of multiple LLM agents. Finally, we consider a range of vision-language models, such as MiniGPT-4 [76], BLIP-2 [34], LLaMA-Adaptor [17], and InstructBLIP [11], as agents to interact with the visual environment. Unlike pure vision or language models, VLMs are designed to process and integrate both visual and textual data, offering a" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "26279" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.471, + 0.184 + ], + "angle": 0, + "content": "more holistic understanding of the environment. To align these agents with the specific requirements of the ALFWorld benchmark, we finetune them on a pre-collected demonstration dataset. This finetuning process is crucial as it enables the agent to comprehend and adhere to ALFWorld's unique grammar and to develop a basic \"gamesense\"." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.194, + 0.239, + 0.209 + ], + "angle": 0, + "content": "3.2. Training Details" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.217, + 0.471, + 0.535 + ], + "angle": 0, + "content": "The architectural design of EMMA is depicted in Fig. 3. At its core, EMMA employs a Query Transformer (Q-Former) to process visual data. This Q-Former extracts features using a frozen ViT encoder. Its output consists of 32 visual tokens, which are then passed through a linear projection layer before being fed to a frozen LLM decoder. Similar to other VLM agents, EMMA is also finetuned on a pre-collected demonstration dataset, aligning its basic ability with the ALFWorld benchmark. In order to align EMMA with the dynamics of ALFWorld, we train it by imitating an LLM expert (see Alg. 1). We choose text-davinci-003 developed by OpenAI as our LLM expert because of its established capabilities in reasoning and planning [54, 61, 64, 73]. In this setup, text-davinci-003 serves dual roles: it serves as an actor, providing EMMA with expert actions, and as a critic, analyzing EMMA's historical trajectories. This analysis generates retrospective feedback, which is then incorporated into the actor's long-term memory, leading to improved actions in future trials. Further details about the hyperparameters and prompts used in our training procedure are available in Table 2 of the Appendix." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.545, + 0.371, + 0.56 + ], + "angle": 0, + "content": "3.3. Comparison with State of the Art" + }, + { + "type": "title", + "bbox": [ + 0.076, + 0.569, + 0.472, + 0.583 + ], + "angle": 0, + "content": "EMMA sets new SOTA performance in visual ALFWorld." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.584, + 0.472, + 0.902 + ], + "angle": 0, + "content": "In this section, we compare EMMA with 12 other representative agents on the ALFWorld benchmark, spanning both visual and textual environments (refer to Table 1). We assess two key metrics: the success rate, which is the percentage of trials completed successfully, and the average number of interaction steps required for task completion, with a lower number indicating higher efficiency. EMMA demonstrates superior performance in both metrics, significantly outperforming all VLM agents in visual environments. This achievement underscores the effectiveness of our cross-modality imitation learning approach, as depicted in the learning curve shown in Fig. 4. Furthermore, EMMA's performance markedly exceeds that of VM agents, highlighting the crucial role of the prior knowledge embedded in VLMs. Intriguingly, EMMA's performance is comparable with LLM agents that operate using perfectly semantic descriptions of visual observations. This is largely attributed to EMMA's training strategy of imitating an expert LLM agent, proving to be more efficient than learning from scratch in a purely visual setting. As a result, EMMA stands out as the only VLM agent that substantially surpasses SOTA LLM agents, such" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.088, + 0.875, + 0.308 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.318, + 0.895, + 0.388 + ], + "angle": 0, + "content": "Figure 5: Comparison of robustness between VLM and LLM agents. Given a specific noise rate, we crop a random portion of the pixel observation and resize it to a given size as the new observation. Similarly, we randomly replace some tokens in the textual observation with arbitrary ones." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.414, + 0.895, + 0.822 + ], + "angle": 0, + "content": "as AutoGen [64] and DEPS [61], in these environments. And its success also directs a potential way to achieve human-level performance in the visual environments of ALFWorld. EMMA is more robust to noisy observations than LLM agents. While LLM agents exhibit a higher success rate with fewer interaction steps in textual environments, as indicated in Table 1, we hypothesize that this superior performance largely relies on their precise semantic abstraction of the environment. However, such an abstraction might not be feasible in real-world applications. To verify this assumption, we set up a more practical scenario where observations are deliberately perturbed at a specific noise rate. We then compare the robustness of EMMA and a SOTA LLM agent, Reflexion, under these noisy observations. To generate noisy observations, a random portion of the visual observation is cropped, resized, and then used to replace the original observation. Similarly, in the textual observation, random tokens are substituted with arbitrary ones. As illustrated in Fig. 5, with the noise rate increasing, EMMA's performance remains significantly more robust than Reflexion. This could be attributed to the vision encoder in the VLM, which is adept at filtering out visual noises. On the other hand, textual noises are directly processed by the LLM, which can substantially impair the performance of LLM-based agents. This finding highlights the potential advantages of VLM agents like EMMA in visual scenarios, in which data is often imperfect and noisy." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.832, + 0.655, + 0.848 + ], + "angle": 0, + "content": "3.4. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.856, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Retrospection improves EMMA over time. To assess the importance of the retrospective LLM expert, we present the average success rate of EMMA after each trial and compare it" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "26280" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.084, + 0.088, + 0.35, + 0.243 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.352, + 0.088, + 0.616, + 0.243 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.62, + 0.089, + 0.885, + 0.243 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.254, + 0.895, + 0.31 + ], + "angle": 0, + "content": "Figure 6: Ablation study. left: The ability of LLM expert to reflect on and learn from past actions appears to be a key factor in achieving the impressive results we observed. middle: The \"w/o BC initialization\" variant uses a pretrained VLM instead of the reference agent \\(\\pi_{\\mathrm{ref}}\\) to initialize EMMA. right: The 'w/ CE Loss' variant replaces the DPO loss (described in Eq. 2) with a token-level cross-entropy (CE) loss. Please refer to Sec. 3.4 for a thorough discussion." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.336, + 0.471, + 0.488 + ], + "angle": 0, + "content": "with a key variation: \"EMMA w/o Retrospection\", as shown in Fig. 6 (left). This variant of EMMA is trained using the same procedure as the original EMMA but removes the retrospective process. Instead, it relies solely on a plain LLM actor to provide relabeled actions. The results show that EMMA with the retrospective mechanism significantly outperforms its counterpart. This finding is crucial as it indicates that the retrospective process is not just a supplementary feature but a fundamental component of EMMA's architecture that contributes substantially to its enhanced performance." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.49, + 0.471, + 0.717 + ], + "angle": 0, + "content": "EMMA benefits from BC initialization. Through an ablation study, we evaluate the impact of behavior cloning (BC) initialization, a process described in line 3 of Alg. 1. Fig. 6 (middle) demonstrates that EMMA, when deprived of BC initialization, experiences a slight reduction in the average success rate across 134 unseen tasks compared to its original setup. Despite this decrease, EMMA without BC initialization still outperforms other VLM agents, as clearly shown when compared with the results in Table 1. Furthermore, Fig. 10 in the Appendix breaks down the result by task type. It reveals a consistent but slight drop in performance across 5 out of the 6 task types. These results reflect that while BC initialization contributes positively to EMMA's overall performance, it is not critical for achieving the notable results we have reported." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.72, + 0.474, + 0.902 + ], + "angle": 0, + "content": "DPO enables more effective imitation learning. To evaluate the effectiveness of using DPO loss in Eq. (1), we conducted an ablation study with an alternative version of EMMA, referred to \"EMMA w/ Cross Entropy (CE) Loss\". In this variant, EMMA is optimized using the token-level CE loss, a common objective for finetuning VLMs. The results, as depicted in Fig. 6 (right), reveal that EMMA w/ CE Loss does not achieve the same high success rate as the original EMMA with DPO loss, suggesting that DPO loss contributes to enhancing the upper performance bound of EMMA. In addition, we noted that EMMA w/ CE Loss exhibits faster convergence in the initial training stages than the original" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.336, + 0.895, + 0.412 + ], + "angle": 0, + "content": "EMMA. This premature convergence leads to the agent paying attention to the expert actions from easier tasks, usually addressed in the first few training epochs. This can suppress EMMA's exploration and learning on more complex tasks, resulting in worse performance." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.419, + 0.891, + 0.433 + ], + "angle": 0, + "content": "3.5. Generalization to Free-form Task Instructions" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.44, + 0.895, + 0.835 + ], + "angle": 0, + "content": "The ability of AI agents to precisely follow task instructions given in open-vocabulary and free-form text is crucial for real-world problems. To assess this, we conducted an additional experiment focusing on the generalization capabilities of EMMA, an agent trained with template task instructions. In this experiment, we re-evaluated the performance of EMMA and other baseline agents on 134 unseen tasks, using human-annotated instructions instead of template ones. These human-annotated instructions include a large amount of OOD verbs and objects, presenting a more realistic and challenging scenario for the agents to address. To underscore this challenge, we compared the vocabulary distribution between the template and human-annotated instructions, as shown in Fig. 7 (right). Moreover, we provide a comprehensive analysis of the vocabulary used across both instruction types in Fig. 11 of the Appendix. In Fig. 7 (left), EMMA demonstrates a slight performance decline, while a significant degradation is observed in other baselines. We also note that Reflexion, a SOTA LLM agent, exhibits exceptional generalization to those OOD instructions. According to these empirical results in Fig. 7, we have the following conclusions: (1) EMMA obtains and benefits from the generalization capabilities inherent in the SOTA LLM agent through cross-modality imitation learning; (2) Our work sheds novel insights on using LLM feedback to train more versatile and generalizable embodied agents." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.845, + 0.642, + 0.861 + ], + "angle": 0, + "content": "4. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.893, + 0.901 + ], + "angle": 0, + "content": "Agents based on Foundation Models. Recent research has increasingly focused on harnessing the capabilities of large" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.518, + 0.957 + ], + "angle": 0, + "content": "26281" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.081, + 0.09, + 0.647, + 0.219 + ], + "angle": 0, + "content": "
AgentVisual Env.Textual Env.Success Rate with Free-form Task Instruction
Avg.PickCleanHeatCoolLookPick2
LMsBUTLER* [56]X✓0.03 (-)0.10 (-)0.22 (-)0.05 (-)0.17 (-)0.00 (-)0.00 (-)
GPT-BUTLER [41]X✓0.31 (24.7)0.25 (25.4)0.42 (23.5)0.46 (20.9)0.44 (21.8)0.00 (30.0)0.12 (28.9)
ReAct [73]X✓0.37 (23.6)0.46 (22.0)0.39 (23.8)0.50 (20.5)0.28 (24.6)0.22 (26.4)0.29 (25.9)
Reflexion [54]X✓0.78 (17.0)0.84 (15.8)0.74 (18.4)0.77 (17.2)0.61 (19.4)0.94 (13.8)0.82 (16.7)
VLMsMiniGPT-4 [76]✓X0.00 (30.0)0.00 (30.0)0.00 (30.0)0.00 (30.0)0.00 (30.0)0.00 (30.0)0.00 (30.0)
BLIP-2 [34]✓X0.01 (29.7)0.04 (29.1)0.03 (29.5)0.00 (30.0)0.00 (30.0)0.00 (30.0)0.00 (30.0)
LLaMA-Adapter [17]✓X0.02 (29.6)0.04 (29.3)0.03 (29.4)0.04 (29.3)0.00 (30.0)0.00 (30.0)0.00 (30.0)
InstructBLIP [11]✓X0.01 (29.8)0.00 (30.0)0.03 (29.3)0.00 (30.0)0.00 (30.0)0.00 (30.0)0.00 (30.0)
EMMA (Ours)✓X0.68 (22.0)0.72 (21.7)0.65 (22.7)0.72 (22.0)0.56 (23.9)0.67 (23.3)0.76 (18.1)
" + }, + { + "type": "image", + "bbox": [ + 0.662, + 0.089, + 0.892, + 0.221 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.231, + 0.893, + 0.26 + ], + "angle": 0, + "content": "Figure 7: Generalization to open-vocabulary, free-form task instructions. Left: These instructions are provided by different human annotators using Amazon Mechanical Turk [55]. Right: Frequency distribution of verbs for human-annotated and templated task instructions." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.286, + 0.473, + 0.695 + ], + "angle": 0, + "content": "pre-trained foundation models to build AI agents [65, 69]. These models (e.g., LLMs), benefiting from their commonsense knowledge inherited from Internet-scale pretraining, are able to reason actions according to descriptions of the external environments. For example, given a set of task instructions, LLMs can be elaborately prompted to perform as agents generating high-level step-by-step plans [7, 24, 54, 64, 73], and each step can be parsed into a sequence of robotic actions that are executed via pretrained policies or available APIs [6, 36, 59]. Furthermore, by using VLMs [32], plans can also be conditioned on visual inputs that are transformed into language descriptions [7, 16, 25, 57] or token embeddings aligned with LLMs [13, 42, 68, 77]. However, existing foundation models are usually pretrained on static text or text-image datasets and thus may struggle to align with the dynamics of the world. To bridge this gap, we study how to finetune a VLM to be an embodied agent dynamically aligned with the world by distilling the cross-modality knowledge from an LLM expert. The work most closely related to ours is EUREKA [40], which also explores using source information provided by the simulator as the context of an LLM to aid agent training. Instead of directly mimicking the output of the LLM as we did, EUREKA harnesses the coding LLM to generate a desired reward function for a given task and optimizes a policy against the reward function using RL, leading to a more complex and expensive training procedure [47, 52, 53]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.704, + 0.472, + 0.902 + ], + "angle": 0, + "content": "Imitation Learning. Imitation learning is the study of algorithms that improve performance by mimicking an expert's decisions and behaviors. We summarize three main categories of existing methods in the following: (1) behavior cloning (BC), (2) inverse reinforcement learning (IRL), and (3) the combination of imitation and reinforcement learning. The naive BC [5] ignores the change in distribution and simply trains a policy that only performs well under the distribution of states visited by the expert. Following works, such as dataset aggregation [51] or policy aggregation [28, 49], have been proposed to address the limitations of BC. Another line of work, IRL, is a more complicated algorithm framework that learns the reward function from" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.286, + 0.895, + 0.528 + ], + "angle": 0, + "content": "expert demonstrations and then improves the policy using RL with the learned reward. A representative method in this category is generative adversarial imitation learning (GAIL) [23], in which a policy and a discriminator compete with each other in order to maximize the likelihood of the policy's behavior matching the expert. The third category of methods usually leverages an IL policy to initialize RL and continues to boost its performance via online collected data from RL [8, 50, 58]. This simple combination significantly improves RL's sample efficiency and IL's upper performance bound constrained by the expert. Nevertheless, all of the above methods assume that the expert and the imitator understand the world in the same modality, and thus overlook the fact that the complementary knowledge from other modalities often boosts the model's accuracy and generalization dramatically [46, 67]." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.542, + 0.62, + 0.558 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.567, + 0.895, + 0.78 + ], + "angle": 0, + "content": "We create EMMA, an Embodied Multi-Modal Agent, by finetuning a VLM in an embodied visual world with interactive imitation learning from an LLM expert in a parallel text world, who produces better actions and retrospective feedback to VLM's trajectories. Such imitation learning exhibits substantial advantages over vision or VLM policies directly finetuned in the visual world or finetuned by behavior cloning of a rule-based expert, and SOTA API VLMs such as GPT-4V(ison). As a result, EMMA achieves a comparable success rate and much better robustness in the noisy visual world than its LLM teacher in the easier text world. Furthermore, EMMA shows powerful generalization to open-vocabulary and free-form task instructions, highlighting its potential in real-world scenarios." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.793, + 0.652, + 0.81 + ], + "angle": 0, + "content": "Acknowledgment" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.817, + 0.895, + 0.903 + ], + "angle": 0, + "content": "This work is partially supported by the National Key R&D Program of China under the Grant No. (2023YFE0106300) and (2017YFC0804002), Shenzhen Fundamental Research Program under the Grant No. (JCYJ20200109141235597), and Program for Guangdong Introducing Innovative and Entrepreneurial Teams under Grant No. (2017ZT07X386)." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "26282" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.121, + 0.174, + 0.137 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.147, + 0.473, + 0.216 + ], + "angle": 0, + "content": "[1] Constructions Aeronautiques, Adele Howe, Craig Knoblock, ISI Drew McDermott, Ashwin Ram, Manuela Veloso, Daniel Weld, David Wilkins SRI, Anthony Barrett, Dave Christianson, et al. PDDL: the planning domain definition language. Technical Report, Tech. Rep., 1998. 4, 5, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.218, + 0.472, + 0.273 + ], + "angle": 0, + "content": "[2] Marcin Andrychowicz, Filip Wolski, Alex Ray, Jonas Schneider, Rachel Fong, Peter Welinder, Bob McGrew, Josh Tobin, OpenAI Pieter Abbeel, and Wojciech Zaremba. Hindsight experience replay. 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.275, + 0.472, + 0.316 + ], + "angle": 0, + "content": "[3] Shuang Ao, Tianyi Zhou, Guodong Long, Qinghua Lu, Liming Zhu, and Jing Jiang. Co-pilot: Collaborative planning and reinforcement learning on sub-task curriculum. 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.318, + 0.472, + 0.373 + ], + "angle": 0, + "content": "[4] Shuang Ao, Tianyi Zhou, Jing Jiang, Guodong Long, Xuan Song, and Chengqi Zhang. Eat-c: environment-adversarial sub-task curriculum for efficient reinforcement learning. In ICML, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.375, + 0.472, + 0.403 + ], + "angle": 0, + "content": "[5] Michael Bain and Claude Sammut. A framework for behavioural cloning. In Machine Intelligence, 1995. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.405, + 0.472, + 0.473 + ], + "angle": 0, + "content": "[6] Anthony Brohan, Noah Brown, Justice Carbajal, Yevgen Chebotar, Joseph Dabis, Chelsea Finn, Keerthana Gopalakrishnan, Karol Hausman, Alex Herzog, Jasmine Hsu, et al. Rt-1: Robotics transformer for real-world control at scale. arXiv preprint arXiv:2212.06817, 2022. 2, 4, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.476, + 0.472, + 0.544 + ], + "angle": 0, + "content": "[7] Anthony Brohan, Yevgen Chebotar, Chelsea Finn, Karol Hausman, Alexander Herzog, Daniel Ho, Julian Ibarz, Alex Irpan, Eric Jang, Ryan Julian, et al. Do as i can, not as i say: Grounding language in robotic affordances. In CoRL, 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.547, + 0.472, + 0.587 + ], + "angle": 0, + "content": "[8] Kai-Wei Chang, Akshay Krishnamurthy, Alekh Agarwal, Hal Daumé III, and John Langford. Learning to search better than your teacher. In ICML, 2015. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.59, + 0.472, + 0.644 + ], + "angle": 0, + "content": "[9] Lichang Chen, Jiuhai Chen, Tom Goldstein, Heng Huang, and Tianyi Zhou. Instructzero: Efficient instruction optimization for black-box large language models. arXiv preprint arXiv:2306.03082, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.647, + 0.472, + 0.715 + ], + "angle": 0, + "content": "[10] Marc-Alexandre Côté, Akos Kádár, Xingdi Yuan, Ben Kybartas, Tavian Barnes, Emery Fine, James Moore, Matthew Hausknecht, Layla El Asri, Mahmoud Adada, et al. Textworld: A learning environment for text-based games. In Computer Games Workshop, IJCAI, 2019. 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.718, + 0.472, + 0.773 + ], + "angle": 0, + "content": "[11] Wenliang Dai, Junnan Li, Dongxu Li, Anthony Tiong, Junqi Zhao, Weisheng Wang, Boyang Li, Pascale Fung, and Steven Hoi. InstructBLIP: Towards general-purpose vision-language models with instruction tuning. In NeurIPS, 2023. 2, 3, 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.775, + 0.472, + 0.829 + ], + "angle": 0, + "content": "[12] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. BERT: pre-training of deep bidirectional transformers for language understanding. In NAACL-HLT, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.832, + 0.472, + 0.9 + ], + "angle": 0, + "content": "[13] Danny Driess, Fei Xia, Mehdi SM Sajjadi, Corey Lynch, Aakanksha Chowdhery, Brian Ichter, Ayzaan Wahid, Jonathan Thompson, Quan Vuong, Tianhe Yu, et al. Palm-e: An embodied multimodal language model. arXiv preprint arXiv:2303.03378, 2023. 2, 4, 8" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.147, + 0.473, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.134 + ], + "angle": 0, + "content": "[14] Meng Fang, Tianyi Zhou, Yali Du, Lei Han, and Zhengyou Zhang. Curriculum-guided hindsight experience replay. NeurIPS, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.137, + 0.894, + 0.191 + ], + "angle": 0, + "content": "[15] Daocheng Fu, Xin Li, Licheng Wen, Min Dou, Pinlong Cai, Botian Shi, and Yu Qiao. Drive like a human: Rethinking autonomous driving with large language models. arXiv preprint arXiv:2307.07162, 2023. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.194, + 0.894, + 0.248 + ], + "angle": 0, + "content": "[16] Jensen Gao, Bidipta Sarkar, Fei Xia, Ted Xiao, Jiajun Wu, Brian Ichter, Anirudha Majumdar, and Dorsa Sadigh. Physically grounded vision-language models for robotic manipulation. arXiv preprint arXiv:2309.02561, 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.251, + 0.894, + 0.307 + ], + "angle": 0, + "content": "[17] Peng Gao, Jiaming Han, Renrui Zhang, Ziyi Lin, Shijie Geng, Aojun Zhou, Wei Zhang, Pan Lu, Conghui He, Xiangyu Yue, et al. Llama-adapter v2: Parameter-efficient visual instruction model. arXiv preprint arXiv:2304.15010, 2023. 5, 8, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.309, + 0.894, + 0.363 + ], + "angle": 0, + "content": "[18] Jiayi Guan, Guang Chen, Jiaming Ji, Long Yang, Zhijun Li, et al. Voce: Variational optimization with conservative estimation for offline safe reinforcement learning. NeurIPS, 36, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.366, + 0.893, + 0.407 + ], + "angle": 0, + "content": "[19] Caglar Gülcehre, Sungjin Ahn, Ramesh Nallapati, Bowen Zhou, and Yoshua Bengio. Pointing the unknown words. In ACL, 2016. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.41, + 0.894, + 0.45 + ], + "angle": 0, + "content": "[20] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, 2016. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.453, + 0.894, + 0.48 + ], + "angle": 0, + "content": "[21] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross B. Girshick. Mask R-CNN. In ICCV, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.483, + 0.893, + 0.511 + ], + "angle": 0, + "content": "[22] Malte Helmert. The fast downward planning system. Journal of Artificial Intelligence Research, 2006. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.513, + 0.893, + 0.54 + ], + "angle": 0, + "content": "[23] Jonathan Ho and Stefano Ermon. Generative adversarial imitation learning. In NeurIPS, 2016. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.543, + 0.894, + 0.596 + ], + "angle": 0, + "content": "[24] Wenlong Huang, Pieter Abbeel, Deepak Pathak, and Igor Mordatch. Language models as zero-shot planners: Extracting actionable knowledge for embodied agents. In ICML, 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.6, + 0.893, + 0.655 + ], + "angle": 0, + "content": "[25] Wenlong Huang, Chen Wang, Ruohan Zhang, Yunzhu Li, Jiajun Wu, and Li Fei-Fei. Voxposer: Composable 3d value maps for robotic manipulation with language models. arXiv preprint arXiv:2307.05973, 2023. 2, 4, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.657, + 0.894, + 0.699 + ], + "angle": 0, + "content": "[26] Ahmed Hussein, Mohamed Medhat Gaber, Eyad Elyan, and Chrisina Jayne. Imitation learning: A survey of learning methods. ACM Computing Surveys (CSUR), 50(2):1-35, 2017. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.702, + 0.893, + 0.741 + ], + "angle": 0, + "content": "[27] Marcus Hutter. Universal artificial intelligence: Sequential decisions based on algorithmic probability. Springer Science & Business Media, 2004. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.744, + 0.894, + 0.772 + ], + "angle": 0, + "content": "[28] Hal Daumé III, John Langford, and Daniel Marcu. Search-based structured prediction. Mach. Learn., 2009. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.775, + 0.894, + 0.829 + ], + "angle": 0, + "content": "[29] Yunfan Jiang, Agrim Gupta, Zichen Zhang, Guanzhi Wang, Yongqiang Dou, Yanjun Chen, Li Fei-Fei, Anima Anandkumar, Yuke Zhu, and Linxi Fan. VIMA: robot manipulation with multimodal prompts. In ICML, 2023. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.832, + 0.894, + 0.9 + ], + "angle": 0, + "content": "[30] Eric Kolve, Roozbeh Mottaghi, Winson Han, Eli VanderBilt, Luca Weihs, Alvaro Herrasti, Matt Deitke, Kiana Ehsani, Daniel Gordon, Yuke Zhu, et al. Ai2-thor: An interactive 3d environment for visual ai. arXiv preprint arXiv:1712.05474, 2017. 4, 5, 1" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "26283" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.147 + ], + "angle": 0, + "content": "[31] Sergey Levine, Aviral Kumar, George Tucker, and Justin Fu. Offline reinforcement learning: Tutorial, review, and perspectives on open problems. arXiv preprint arXiv:2005.01643, 2020. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.149, + 0.472, + 0.205 + ], + "angle": 0, + "content": "[32] Chunyuan Li, Zhe Gan, Zhengyuan Yang, Jianwei Yang, Linjie Li, Lijuan Wang, and Jianfeng Gao. Multimodal foundation models: From specialists to general-purpose assistants. arXiv preprint arXiv:2309.10020, 2023. 2, 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.207, + 0.472, + 0.275 + ], + "angle": 0, + "content": "[33] Dongxu Li, Junnan Li, Hung Le, Guangsen Wang, Silvio Savarese, and Steven C.H. Hoi. LAVIS: A one-stop library for language-vision intelligence. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 3: System Demonstrations), 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.277, + 0.472, + 0.332 + ], + "angle": 0, + "content": "[34] Junnan Li, Dongxu Li, Silvio Savarese, and Steven C. H. Hoi. BLIP-2: bootstrapping language-image pre-training with frozen image encoders and large language models. In ICML, 2023. 2, 3, 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.334, + 0.472, + 0.387 + ], + "angle": 0, + "content": "[35] Chen Liang, Jiahui Yu, Ming-Hsuan Yang, Matthew Brown, Yin Cui, Tuo Zhao, Boqing Gong, and Tianyi Zhou. Modulewise adaptive distillation for multimodality foundation models. 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.39, + 0.472, + 0.445 + ], + "angle": 0, + "content": "[36] Jacky Liang, Wenlong Huang, Fei Xia, Peng Xu, Karol Hausman, Brian Ichter, Pete Florence, and Andy Zeng. Code as policies: Language model programs for embodied control. In ICRA, 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.447, + 0.472, + 0.516 + ], + "angle": 0, + "content": "[37] Yong Lin, Lu Tan, Hangyu Lin, Zeming Zheng, Renjie Pi, Jipeng Zhang, Shizhe Diao, Haoxiang Wang, Han Zhao, Yuan Yao, et al. Speciality vs generality: An empirical study on catastrophic forgetting in fine-tuning foundation models. arXiv preprint arXiv:2309.06256, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.518, + 0.472, + 0.558 + ], + "angle": 0, + "content": "[38] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. arXiv preprint arXiv:2304.08485, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.561, + 0.472, + 0.588 + ], + "angle": 0, + "content": "[39] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.59, + 0.472, + 0.658 + ], + "angle": 0, + "content": "[40] Yecheng Jason Ma, William Liang, Guanzhi Wang, De-An Huang, Osbert Bastani, Dinesh Jayaraman, Yuke Zhu, Linxi Fan, and Anima Anandkumar. Eureka: Human-level reward design via coding large language models. arXiv preprint arXiv:2310.12931, 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.66, + 0.472, + 0.687 + ], + "angle": 0, + "content": "[41] Vincent Micheli and François Fleuret. Language models are few-shot butlers. In EMNLP, 2021. 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.689, + 0.472, + 0.757 + ], + "angle": 0, + "content": "[42] Yao Mu, Qinglong Zhang, Mengkang Hu, Wenhai Wang, Mingyu Ding, Jun Jin, Bin Wang, Jifeng Dai, Yu Qiao, and Ping Luo. Embodiedgpt: Vision-language pre-training via embodied chain of thought. arXiv preprint arXiv:2305.15021, 2023. 2, 4, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.76, + 0.472, + 0.8 + ], + "angle": 0, + "content": "[43] OpenAI. GPT-4V(ison) system card. https://cdn.openai.com/papers/GPTV_System_Card.pdf, 2023.2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.803, + 0.472, + 0.858 + ], + "angle": 0, + "content": "[44] Abhishek Padalkar, Acorn Pooley, Ajinkya Jain, Alex Bewley, Alex Herzog, Alex Irpan, Alexander Khazatsky, Anant Rai, Anikait Singh, Anthony Brohan, et al. Open x-embediment: Robotic learning datasets and rt-x models, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.86, + 0.472, + 0.901 + ], + "angle": 0, + "content": "[45] Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, Ilya Sutskever, et al. Language models are unsupervised multitask learners. OpenAI blog, 2019. 5" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.161 + ], + "angle": 0, + "content": "[46] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, 2021. 8, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.164, + 0.894, + 0.22 + ], + "angle": 0, + "content": "[47] Rafael Rafailov, Archit Sharma, Eric Mitchell, Stefano Ermon, Christopher D Manning, and Chelsea Finn. Direct preference optimization: Your language model is secretly a reward model. arXiv preprint arXiv:2305.18290, 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.222, + 0.894, + 0.276 + ], + "angle": 0, + "content": "[48] Rafael Rafailov, Archit Sharma, Eric Mitchell, Christopher D Manning, Stefano Ermon, and Chelsea Finn. Direct preference optimization: Your language model is secretly a reward model. NeurIPS, 2024. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.279, + 0.892, + 0.308 + ], + "angle": 0, + "content": "[49] Stéphane Ross and Drew Bagnell. Efficient reductions for imitation learning. In AISTATS, 2010. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.31, + 0.892, + 0.352 + ], + "angle": 0, + "content": "[50] Stephane Ross and J Andrew Bagnell. Reinforcement and imitation learning via interactive no-regret learning. arXiv preprint arXiv:1406.5979, 2014. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.354, + 0.892, + 0.395 + ], + "angle": 0, + "content": "[51] Stéphane Ross, Geoffrey J. Gordon, and Drew Bagnell. A reduction of imitation learning and structured prediction to no-regret online learning. In AISTATS, 2011. 2, 4, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.398, + 0.894, + 0.438 + ], + "angle": 0, + "content": "[52] John Schulman, Sergey Levine, Pieter Abbeel, Michael I. Jordan, and Philipp Moritz. Trust region policy optimization. In ICML, 2015. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.442, + 0.894, + 0.483 + ], + "angle": 0, + "content": "[53] John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Radford, and Oleg Klimov. Proximal policy optimization algorithms. arXiv preprint arXiv:1707.06347, 2017. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.485, + 0.894, + 0.54 + ], + "angle": 0, + "content": "[54] Noah Shinn, Federico Cassano, Ashwin Gopinath, Karthik R Narasimhan, and Shunyu Yao. Reflexion: Language agents with verbal reinforcement learning. In NeurIPS, 2023. 3, 4, 5, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.543, + 0.894, + 0.611 + ], + "angle": 0, + "content": "[55] Mohit Shridhar, Jesse Thomason, Daniel Gordon, Yonatan Bisk, Winson Han, Roozbeh Mottaghi, Luke Zettlemoyer, and Dieter Fox. ALFRED: A benchmark for interpreting grounded instructions for everyday tasks. In CVPR, 2020. 5, 8, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.615, + 0.894, + 0.67 + ], + "angle": 0, + "content": "[56] Mohit Shridhar, Xingdi Yuan, Marc-Alexandre Côté, Yonatan Bisk, Adam Trischler, and Matthew J. Hausknecht. Alfworld: Aligning text and embodied environments for interactive learning. In ICLR, 2021. 2, 3, 4, 5, 8, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.672, + 0.894, + 0.714 + ], + "angle": 0, + "content": "[57] Theodore Sumers, Kenneth Marino, Arun Ahuja, Rob Fergus, and Ishita Dasgupta. Distilling internet-scale vision-language models into embodied agents. In ICML, 2023. 2, 4, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.716, + 0.894, + 0.758 + ], + "angle": 0, + "content": "[58] Wen Sun, J. Andrew Bagnell, and Byron Boots. Truncated horizon policy search: Combining reinforcement learning & imitation learning. In ICLR, 2018. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.76, + 0.894, + 0.815 + ], + "angle": 0, + "content": "[59] Guanzhi Wang, Yuqi Xie, Yunfan Jiang, Ajay Mandlekar, Chaowei Xiao, Yuke Zhu, Linxi Fan, and Anima Anandkumar. Voyager: An open-ended embodied agent with large language models. arXiv preprint arXiv:2305.16291, 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.818, + 0.894, + 0.899 + ], + "angle": 0, + "content": "[60] Xiyao Wang, Yuhang Zhou, Xiaoyu Liu, Hongjin Lu, Yuancheng Xu, Feihong He, Jaehong Yoon, Taixi Lu, Gedas Bertasius, Mohit Bansal, et al. Mementos: A comprehensive benchmark for multimodal large language model reasoning over image sequences. arXiv preprint arXiv:2401.10529, 2024. 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "26284" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.147 + ], + "angle": 0, + "content": "[61] Zihao Wang, Shaofei Cai, Anji Liu, Xiaojian Ma, and Yitao Liang. Describe, explain, plan and select: Interactive planning with large language models enables open-world multi-task agents. arXiv preprint arXiv:2302.01560, 2023. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.15, + 0.471, + 0.205 + ], + "angle": 0, + "content": "[62] Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Brian Ichter, Fei Xia, Ed H. Chi, Quoc V. Le, and Denny Zhou. Chain-of-thought prompting elicits reasoning in large language models. In NeurIPS, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.207, + 0.469, + 0.248 + ], + "angle": 0, + "content": "[63] Michael Wooldridge and Nicholas R Jennings. Intelligent agents: Theory and practice. The knowledge engineering review, 10(2):115-152, 1995. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.25, + 0.47, + 0.318 + ], + "angle": 0, + "content": "[64] Qingyun Wu, Gagan Bansal, Jieyu Zhang, Yiran Wu, Shaokun Zhang, Erkang Zhu, Beibin Li, Li Jiang, Xiaoyun Zhang, and Chi Wang. Autogen: Enabling next-gen llm applications via multi-agent conversation framework. arXiv preprint arXiv:2308.08155, 2023. 5, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.321, + 0.471, + 0.387 + ], + "angle": 0, + "content": "[65] Zhiheng Xi, Wenxiang Chen, Xin Guo, Wei He, Yiwen Ding, Boyang Hong, Ming Zhang, Junzhe Wang, Senjie Jin, Enyu Zhou, et al. The rise and potential of large language model based agents: A survey. arXiv preprint arXiv:2309.07864, 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.391, + 0.471, + 0.459 + ], + "angle": 0, + "content": "[66] Zhenhua Xu, Yujia Zhang, Enze Xie, Zhen Zhao, Yong Guo, Kenneth KY Wong, Zhenguo Li, and Hengshuang Zhao. Drivegpt4: Interpretable end-to-end autonomous driving via large language model. arXiv preprint arXiv:2310.01412, 2023. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.462, + 0.47, + 0.503 + ], + "angle": 0, + "content": "[67] Zihui Xue, Zhengqi Gao, Sucheng Ren, and Hang Zhao. The modality focusing hypothesis: Towards understanding cross-modal knowledge distillation. In ICLR, 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.505, + 0.471, + 0.573 + ], + "angle": 0, + "content": "[68] Jingkang Yang, Yuhao Dong, Shuai Liu, Bo Li, Ziyue Wang, Chencheng Jiang, Haoran Tan, Jiamu Kang, Yuanhan Zhang, Kaiyang Zhou, et al. Octopus: Embodied vision-language programmer from environmental feedback. arXiv preprint arXiv:2310.08588, 2023. 3, 4, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.576, + 0.47, + 0.63 + ], + "angle": 0, + "content": "[69] Sherry Yang, Ofir Nachum, Yilun Du, Jason Wei, Pieter Abbeel, and Dale Schuurmans. Foundation models for decision making: Problems, methods, and opportunities. arXiv preprint arXiv:2303.04129, 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.633, + 0.469, + 0.673 + ], + "angle": 0, + "content": "[70] Yijun Yang, Jing Jiang, Tianyi Zhou, Jie Ma, and Yuhui Shi. Pareto policy pool for model-based offline reinforcement learning. In ICLR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.676, + 0.469, + 0.716 + ], + "angle": 0, + "content": "[71] Yijun Yang, Tianyi Zhou, Jing Jiang, Guodong Long, and Yuhui Shi. Continual task allocation in meta-policy network via sparse prompting. In ICML, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.719, + 0.47, + 0.773 + ], + "angle": 0, + "content": "[72] Shunyu Yao, Dian Yu, Jeffrey Zhao, Izhak Shafran, Thomas L Griffiths, Yuan Cao, and Karthik Narasimhan. Tree of thoughts: Deliberate problem solving with large language models. arXiv preprint arXiv:2305.10601, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.776, + 0.47, + 0.829 + ], + "angle": 0, + "content": "[73] Shunyu Yao, Jeffrey Zhao, Dian Yu, Nan Du, Izhak Shafran, Karthik R. Narasimhan, and Yuan Cao. React: Synergizing reasoning and acting in language models. In ICLR, 2023. 3, 5, 6, 8, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.833, + 0.47, + 0.899 + ], + "angle": 0, + "content": "[74] Lianmin Zheng, Wei-Lin Chiang, Ying Sheng, Siyuan Zhuang, Zhanghao Wu, Yonghao Zhuang, Zi Lin, Zhuohan Li, Dacheng Li, Eric Xing, et al. Judging llm-as-a-judge with mt-bench and chatbot arena. arXiv preprint arXiv:2306.05685, 2023. 3" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.471, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.134 + ], + "angle": 0, + "content": "[75] Victor Zhong, Dipendra Misra, Xingdi Yuan, and Marc-Alexandre Côté. Policy improvement using language feedback models. arXiv preprint arXiv:2402.07876, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.137, + 0.894, + 0.19 + ], + "angle": 0, + "content": "[76] Deyao Zhu, Jun Chen, Xiaogian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models. arXiv preprint arXiv:2304.10592, 2023. 2, 3, 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.193, + 0.894, + 0.247 + ], + "angle": 0, + "content": "[77] Brianna Zitkovich, Tianhe Yu, Sichun Xu, Peng Xu, Ted Xiao, Fei Xia, Jialin Wu, Paul Wohlhart, Stefan Welker, Ayzaan Wahid, et al. Rt-2: Vision-language-action models transfer web knowledge to robotic control. In CoRL, 2023. 2, 3, 4, 8" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.247 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.518, + 0.956 + ], + "angle": 0, + "content": "26285" + } + ] +] \ No newline at end of file diff --git a/2024/Embodied Multi-Modal Agent trained by an LLM from a Parallel TextWorld/911fda92-445d-4ff5-8ef4-3ff0a7daf4bf_origin.pdf b/2024/Embodied Multi-Modal Agent trained by an LLM from a Parallel TextWorld/911fda92-445d-4ff5-8ef4-3ff0a7daf4bf_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..557ab54169d5786b10c34656302a385363932665 --- /dev/null +++ b/2024/Embodied Multi-Modal Agent trained by an LLM from a Parallel TextWorld/911fda92-445d-4ff5-8ef4-3ff0a7daf4bf_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0ebe054fab2af18430b9ceaeea3a07827ae4006d08a43b35ae01613f87225886 +size 2680437 diff --git a/2024/Embodied Multi-Modal Agent trained by an LLM from a Parallel TextWorld/full.md b/2024/Embodied Multi-Modal Agent trained by an LLM from a Parallel TextWorld/full.md new file mode 100644 index 0000000000000000000000000000000000000000..00bf797f76fe4f3287518ade571091ee3b061709 --- /dev/null +++ b/2024/Embodied Multi-Modal Agent trained by an LLM from a Parallel TextWorld/full.md @@ -0,0 +1,316 @@ +# Embodied Multi-Modal Agent trained by an LLM from a Parallel TextWorld + +Yijun Yang $^{1,5,4}$ Tianyi Zhou $^{2}$ Kanxue Li $^{3}$ Dapeng Tao $^{3}$ Lusong Li $^{4}$ + +Li Shen $^{4,*}$ Xiaodong He $^{4}$ Jing Jiang $^{5}$ Yuhui Shi $^{1,*}$ + +$^{1}$ Southern University of Science and Technology $^{2}$ University of Maryland, College Park + +$^{3}$ Yunnan University $^{4}$ JD Explore Academy $^{5}$ University of Technology Sydney + +https://github.com/stevenyangyj/Emma-Alfworld + +![](images/dd386abecb87701148aa9f7c71f19256df3bf2e1083b59acc935a55ebaca7ade.jpg) +Instruction: put the gold vase into safe. + +![](images/c81a71c3b7afa35b3b05f6787e1a509271383280f73a66b5aae128bd44839227.jpg) +Action succeeded + +![](images/5ff5616d4cb5341f71d5f08bc9ee781a70d2fd4702a88a64235b65077caba097.jpg) +Action failed due to hallucination + +![](images/18db195502585bb1f6a854aa47c535506f0c4862931e8d7447df405b1bfd1f64.jpg) +Action failed due to invalid grammar + +![](images/a33d8ecb5f073cfb94e2660f00a0ff46b8e2329182a418cf5e037e211d913d8e.jpg) +Figure 1: Comparison of three VLM-based agents in the visual environment from ALFWorld. Given the task instruction and the current-step observation as inputs, a VLM agent is expected to predict an action, e.g., "go to cabinet 1", towards completing the task. (a) GPT-4V(ison). (b) InstructBLIP finetuned by behavior cloning on a static dataset of demonstrations by a rule-based expert in the visual simulator. (c) Our Embodied Multi-Modal Agent (EMMA) trained via cross-modality interactive imitation learning from an LLM expert. + +![](images/5ec5a05645cce7c2d08386c06e6bff99f89aa78dcc783f8772b75793136bcdff.jpg) + +![](images/3da630c5985c6d006562dd50b1a04300008e6b4a4c614da84e587230d60e8a74.jpg) + +![](images/67f671790556931a529d157a1a77a160e4f69365c4fe995c70425a0bd77ce609.jpg) + +![](images/b88796c2e9a6ae5a18343a0592fbad2eae4e6b26402607e92f072b2e91e02ddb.jpg) + +![](images/b9fdd5cbef6e4a55ad6a26e2c8e64e59ac3e9b74c2a8c9ef4d27fdb47d0c1eb5.jpg) + +![](images/deecaff978c6d4646e43fefaadabf18ae535fc4e530aeaa4227e7b38e7c09699.jpg) +Target object + +![](images/8513336054da0b8916b21dca282448e7ffe438b335339d1fac6f88408253abf2.jpg) + +![](images/7f74171183cc4f69c9be4b9cf20c7f8df12aa3e3d167f88a659fbf378d17cee7.jpg) + +![](images/e087309f22d275658538b2441862eef16281351cb4b9236dcfabfa470f74ad6e.jpg) + +![](images/9c709fbdc18013f5515b605389e83cce378de6df6c586faba909cdd37cfa5716.jpg) + +![](images/2f9b1fa27e890aac9551f12c31b27b2cb39693185efb876ef5b75bea74251b20.jpg) + +![](images/ee9caca2ddadb3d500ccacefb84ba6e158538645928f9b772dc658c44a1fcf61.jpg) + +![](images/54c99cb417df9b2afbb13914625e3d4624da4f78e3978a8c4b9a72c4693c1ef4.jpg) + +![](images/b2c9a46a911a56759e786f03de43e4965fe1249e747e7780c07da6019aa5d744.jpg) + +![](images/8634dd00e81493d7777976ff6ecee3afd1a33173aaebbda5cf5004f665dcb0ff.jpg) + +# Abstract + +While large language models (LLMs) excel in a simulated world of texts, they struggle to interact with the more realistic world without perceptions of other modalities such as visual or audio signals. Although vision-language models (VLMs) integrate LLM modules (1) aligned with static image features, and (2) may possess prior knowledge of world dynamics (as demonstrated in the text world), they have not been trained in an embodied visual world and thus cannot align with its dynamics. On the other hand, training an embodied agent in a noisy visual world without expert guidance is often challenging and inefficient. In this paper, we train a VLM agent living in a visual world using an LLM agent excelling in a parallel text world. Specifically, we distill LLM's reflection outcomes (improved actions by analyzing mistakes) in + +a text world's tasks to finetune the VLM on the same tasks of the visual world, resulting in an Embodied Multi-Modal Agent (EMMA) quickly adapting to the visual world dynamics. Such cross-modality imitation learning between the two parallel worlds is achieved by a novel DAgger-DPO algorithm, enabling EMMA to generalize to a broad scope of new tasks without any further guidance from the LLM expert. Extensive evaluations on the ALFWorld benchmark's diverse tasks highlight EMMA's superior performance to SOTA VLM-based agents, e.g., $20\% -70\%$ improvement in the success rate. + +# 1. Introduction + +Embodied multi-modal agents have been acknowledged as a pivotal stride towards achieving Artificial General Intelligence (AGI), as they encompass the potential for a broad scope of intelligent activities [27, 63]. The rising of founda + +tion models brings a glimmer of hope for constructing such agents [11, 32, 34, 35, 38, 71, 76], and notable efforts from the community have tried to harness them in many decision-making scenarios, e.g., autonomous driving [15, 66], daily household robots [6, 13, 77], and complex manipulation tasks [25, 29, 42, 57]. + +LLMs can perform as reflex agents interacting with a text world via verbalized descriptions of the world states and textual actions. In addition, they exhibit great potential in planning, reflection, and reward shaping. This is mainly due to their prior knowledge and semantic abstraction of the world. However, LLM-based agents cannot be directly applied in a visual world. Although current vision-language models (VLMs) try to align LLMs with the visual modality, their pretraining only focuses on static alignment between image-text pairs so the resulting agents have not been aligned well with the dynamics of the visual world. As shown in Fig. 1 (a), even the privileged VLM, e.g., GPT-4V [43], fails to accomplish tasks in an embodied ALFWorld environment [56]. In such a zero-shot setting, GPT-4V tends to mainly rely on its language prior of these objects detected in the current step, rather than the alignment between the visual input and the environment dynamics conditioned on the task instruction. + +In this paper, we study how to train a VLM towards an embodied agent in a visual world by aligning it with the visual world dynamics and distilling the skills of an LLM agent in a parallel text world. Our overarching goal is to build such an Embodied Multi-Modal Agent (EMMA) that can take a textual task instruction (e.g., from human users) and pixel observations of the state per step to produce a sequence of actions leading to the efficient completion of the task. This is a challenging problem due to (1) the sparsity of task reward [2-4, 14], (2) noisy visual representations, (3) the hallucination of VLMs [60], and (4) the misalignment of VLM's static representations to the visual world dynamics. While offline distillation and imitation from a powerful LLM agent in a parallel text world can potentially overcome the former two challenges [75], effectively mitigating the remaining challenges necessitates online finetuning of the VLM agent within an interactive and visual world [18, 31, 70]. + +To this end, we finetune a VLM agent by imitation learning from an LLM expert (e.g., built on ChatGPT) launched in a parallel text world on the same tasks. Specifically, in each step of EMMA interacting with the visual world, we convert its visual observation into an equivalent textual description sent to the LLM agent, which produces an action for EMMA to imitate. Such cross-modality interactive imitation learning is based on DAgger [51], which overcomes the cumulative errors and distribution shifts caused by behavior cloning (BC). As depicted in Fig. 1 (b), an InstructBLIP [11] agent finetuned by BC on 170K expert demonstrations produced by a rule-based expert in the + +![](images/d391294bac87c81adc36437d75c4a211507e4811d908ce2f9080248a19a5a35e.jpg) +Figure 2: An example of tasks generated for the two parallel worlds. A VLM agent in the visual world and an LLM agent in the text world as household robots instructed to clean an apple and then put it into the fridge. Zoom in for more details. + +visual world still fails to take correct actions based on visual observations. We further improve the DAgger's objective to be the direct preference optimization (DPO) [48], which maximizes the preference of LLM-expert's action (positive) over VLM-student's action (negative) in each interaction step. To collect better teaching signals retrospective to the VLM student's actions, the LLM expert is composed of an LLM actor prompted to output expert actions, and an LLM critic prompted for reflection feedback on the VLM agent's historical trajectories. We maintain a long-term memory storing the feedback, which is then used to induce the LLM actor to improve actions for imitation in future episodes. + +We evaluate EMMA and compare it with vision-only agents, LLM agents, and VLM agents deployed to the ALF-World benchmark [56], which includes numerous tasks in both visual and textual environments. Extensive evaluations highlight that EMMA substantially outperforms state-of-the-art (SOTA) VLM agents in visual-only environments by $20\% - 70\%$ in terms of success rate. In addition, EMMA is the only VLM agent that can generalize to open-vocabulary and free-form test tasks, shedding novel insights on using LLM feedback to train more versatile and generalizable embodied agents in multi-modality environments. + +# 2. Embodied Multi-Modal Agent + +Fig. 3 illustrates the main idea of our "Embodied MultiModal Agent (EMMA)", whose detailed training procedures are given in Alg. 1. The agent is built upon a modularized VLM, which can follow instructions and interact with the environment through pixel observations and textual actions. To overcome these challenges associated with training EMMA, such as sparse reward or distribution shift, we explore the construction of an LLM expert from a parallel text world (Sec. 2.2) for providing EMMA with step-by-step guidance. + +![](images/b51fafabd31cf2d615efc5a6fe5a3c8a11df5156f42a4bdb0de996fe16c19afe.jpg) +Figure 3: Embodied Multi-Modal Agent (EMMA) trained by an LLM expert via cross-modality imitation learning. EMMA takes a textual task instruction and pixel observations as its input state per step to generate a sequence of actions using a VLM. Then, we convert each pixel observation into a textual equivalent as the context of an LLM expert to produce improved actions for EMMA to imitate. + +Lastly, Sec. 2.3 further discusses how to harness the LLM expert to train EMMA via cross-modality imitation learning. + +# 2.1. EMMA in the Visual World + +In visual environments, EMMA $\pi_{\theta}$ is designed to process a task instruction $x_{\mathrm{task}}$ (e.g., provided by human users) and the pixel observation $s_v^t$ at each interaction step $t$ . It is expected to generate a sequence of high-level textual actions $\{x_a^t \sim \pi_\theta(\cdot | x_{\mathrm{task}}, s_v^t)\}_{t=0}^T$ towards efficient completion of the task. To achieve this, we draw inspiration from recent advances of large pretrained VLMs [11, 32, 34, 38, 76], and modularize EMMA's architecture into four components: (1) a ViT to encode $s_v$ into visual embeddings, (2) a querying transformer (Q-Former) tailored to extract the most relevant visual features via the cross-attention between the visual embeddings and query tokens, (3) a linear projection layer to align visual features to text embeddings, (4) an LLM decoder taking the concatenation of the instruction tokens and the output of the linear projection layer to autoregressively generate the action $x_a$ . In order to reduce computational overhead and prevent catastrophic forgetting [37], we adopt the pretrained ViT, Q-Former, and LLM from InstructBLIP [11] and keep them frozen at the finetuning stage. We only update the linear projection layer, as illustrated in Fig. 3. Such a modularized architecture enables EMMA to integrate any existing pretrained vision models and LLMs in a flexible and computationally-efficient way. + +However, deploying EMMA into a complex visual world remains challenging. One of the main obstacles is that the direct use of any pretrained VLM is suboptimal because existing pretraining only focuses on static alignment between + +image-text pairs [11, 32, 34], so that the resulting agent may struggle to reason about the dynamics of the world. As discussed in Sec. 1, even the SOTA VLM, i.e., GPT-4V, fails to accomplish tasks in an embodied ALFWorld environment [56]. In such a zero-shot setting, GPT-4V tends to rely on the linguistic prior of the currently detected objects, rather than the given task instruction and the underlying dynamics of the environment, to guide the interactions. Moreover, finetuning a pretrained VLM on a pre-collected demonstration dataset is also suboptimal due to the diversity of environments and tasks [77], the lack of large-scale expert annotations [44] as well as the challenges posed by the distribution shift issue [31]. A seemingly natural solution to the above challenges is reinforcement learning from environmental feedback (RLEF) [68], in which reward signals rely on decomposing a task into a sequence of reasonable sub-goals and checking their completion. However, in real-world scenarios, most sub-goals cannot be defined or described precisely, so the reward is sparse; hence, we do not expect RLEF to be effective. + +To this end, we propose to leverage interactive imitation learning (IL) to align EMMA with the dynamics of any environment, which however results in two critical algorithmic challenges: (1) How to obtain a high-quality, accessible, and scalable expert that EMMA can query during IL (Sec. 2.2)? (2) Designing an effective strategy to train EMMA using this expert in complex, diverse, and potentially open-ended environments (Sec. 2.3). + +# 2.2. LLM Expert from a Parallel TextWorld + +Thanks to a series of prompting techniques in the realm of in-context learning, such as chain-of-thought [62], tree-of-thought [72], ReAct [73], and Reflexion [54], pretrained + +Algorithm 1 DAgger-DPO with a single task instruction +1: initialize: $i = 0$ $\mathcal{D}\gets \emptyset$ , EMMA $\pi_{\theta}$ , LLM actor $M_a$ with a FIFO memory pool $\mathcal{P}\gets \emptyset$ , LLM critic $M_c$ +2: input: max trials $I$ , training epochs $I_{e}$ , visual and textual envs $E_{v},E_{l}$ , instruction $x_{\mathrm{task}}$ , reference agent $\pi_{\mathrm{ref}}$ +3: Initialize $\pi_{\theta}$ to $\pi_{\mathrm{ref}}$ Behavior Cloning Initialization +4: while task not completed or $i < I$ do +5: Get $\tau_v^i = [x_{\mathrm{task}},s_v^0,x_a^0,\dots,s_v^T,x_a^T]$ via $E_{v}$ with $\pi_{\theta}$ +6: Get $\tau_l^i = [x_{\mathrm{task}},s_l^0,x_a^0,\dots,s_l^T,x_a^T]$ via $E_{l}$ with $\tau_v^i$ +7: Generate retrospective feedback $\mathcal{P}_i = M_c(\tau_l^i)$ +8: Update $\mathcal{P}$ with $\mathcal{P}_i$ (i.e., $\mathcal{P}\leftarrow \mathcal{P}\cup \mathcal{P}_i$ +9: for $t = 0$ to $T$ do Dataset Aggregation +10: $x_{a}^{*} = M_{a}(\mathcal{P},x_{\mathrm{task}},\ldots ,x_{a}^{t - 1},s_{l}^{t})$ +11: $\mathcal{D}\gets \mathcal{D}\cup \{(x_{\mathrm{task}},s_v^t,x_a^t,x_a^*)\}$ +12: for $j = 0$ to $I_{e} - 1$ do Gradient Descent on $\theta$ +13: Sample a mini-batch $\tau$ from $\mathcal{D}$ +14: Update $\theta$ by minimizing Eq. (2) with $\pi_{\mathrm{ref}}$ and $\tau$ +15: $i\gets i + 1$ +16: output: $\pi_{\theta^{*}}$ + +LLMs have demonstrated impressive zero-shot performance across many decision-making scenarios [6, 13, 15, 25, 29, 42, 57, 66, 77]. Despite the great potential in serving as high-quality and scalable experts, they are only able to interact with the environment via textual descriptions of the states, rather than using raw pixel observations like EMMA. To bridge this gap, we convert each pixel observation $s_v$ into a textual equivalent by extracting its metadata from the simulator [30], which is composed of attributes such as Observed Objects, Observed Relations, Inventory, and Locations. We then employ the Planning Domain Definition Language (PDDL) [1] to describe this metadata and create an equivalent textual description/state $s_l$ using the TextWorld engine [10]. Additional details are available in Appendix 7 and an example of this process is illustrated in Fig. 2. This methodology enables the utilization of any pretrained LLM agent that generates a sequence of actions facilitating the training of EMMA through cross-modality IL between the two agents. + +# 2.3. Training EMMA via Cross-Modality Imitation + +Given an LLM expert from the parallel TextWorld, we aim to train a VLM agent $\pi_{\theta}$ in the visual world to imitate its behaviors closely. This is equal to minimizing the following objective under the distribution of states induced by $\pi_{\theta}$ . + +$$ +\theta^ {*} = \underset {\theta \in \Theta} {\arg \min } \mathbb {E} _ {\pi_ {\theta}} \left[ \mathcal {L} _ {\mathrm {i m i t}} \left(\pi_ {\theta} \left(x _ {a} \mid s _ {v}\right), x _ {a} ^ {*}\right)\right) ], \tag {1} +$$ + +in which the choice of loss function $\mathcal{L}_{\mathrm{imit}}$ is dependent on specific scenarios. For instance, it may be the expected cross-entropy loss for the discrete action space, or the expected + +MSE loss for the continuous action space. In our case, we select DPO [48] loss due to its proven superior performance to the cross-entropy on aligning models with expert preferences within the discrete language space. Hence, Eq. (1) can be extended to the formulation below. + +$$ +\theta^ {*} = \underset {\theta \in \Theta} {\arg \min } - \mathbb {E} _ {\pi_ {\theta}} \left[ \mathcal {L} _ {\operatorname {i m i t}} \left(\pi_ {\theta}, \pi_ {\text {r e f}}, s _ {v}, x _ {a}, x _ {a} ^ {*}\right) \right], \tag {2} +$$ + +$$ +\mathcal {L} _ {\mathrm {i m i t}} (\cdot) \triangleq \log \sigma \left(\beta \log \frac {\pi_ {\theta} \left(x _ {a} ^ {*} \mid s _ {v}\right)}{\pi_ {\mathrm {r e f}} \left(x _ {a} ^ {*} \mid s _ {v}\right)} - \beta \log \frac {\pi_ {\theta} \left(x _ {a} \mid s _ {v}\right)}{\pi_ {\mathrm {r e f}} \left(x _ {a} \mid s _ {v}\right)}\right) +$$ + +where $x_{a}^{*}$ is the action given by an expert while $x_{a}$ is the action given by $\pi_{\theta}$ , $\sigma$ is the logistic function, and $\beta$ is a hyperparameter controlling the deviation from $\pi_{\mathrm{ref}}$ , i.e. the reference agent obtained by behavior cloning on a demonstration dataset produced by a rule-based expert in the visual world (see Appendix 9 for complete details). This regularization is essential, as it prevents the agent from deviating too far from the distribution on which the expert is accurate, as well as maintaining the generation diversity and avoiding premature convergence to some easy tasks [48]. In practice, the VLM agent $\pi_{\theta}$ is also initialized to $\pi_{\mathrm{ref}}$ for stabilizing the training process. Since environmental dynamics is both unknown and complex, we cannot compute the distribution of states visited by $\pi_{\theta}$ and can only sample it by rolling out the agent. Hence, Eq. (2) is a non-i.i.d. supervised learning problem due to the dependence of the state distribution on $\pi_{\theta}$ itself, in which naive behavior cloning faces issues like cumulative error and distribution shift [26]. To address this, we employ an interactive IL algorithm, DAgger [51], which provably converges to the optimal agent $\pi_{\theta^{*}}$ . + +As discussed in Sec. 2.2 and Sec. 2.3, we can harness an LLM expert to generate a sequence of actions, serving as $x_{a}^{*}$ in Eq. (2). However, these actions may be suboptimal due to sparse environmental feedback [54, 68] or defective in-context instructions [9]. To collect better teaching signals, we introduce a "retrospective LLM expert" that is composed of two specialized models: an actor $(M_{a})$ , built upon an API LLM and prompted to generate actions based on the task instruction and state observations; and a critic $(M_{c})$ , also based on the same LLM, but designed to analyze EMMA's historical interactions and provide reflective feedback. A long-term memory $\mathcal{P}$ is maintained to store the feedback generated by $M_{c}$ , which is then used to prompt $M_{a}$ for improved actions. The complete procedure is detailed in Line 7-10 of Alg. 1, and all prompts are provided in Sec. 6 of the Appendix. + +# 3. Experiments + +# 3.1. Experimental Setup + +**Environments.** We base our experiments on the ALFWorld benchmark [56], a cross-modality simulation platform that encompasses a wide range of embodied household tasks. For + +![](images/da5c7f948991768971c6b06a25303425486e734e314fb500c279df7962cdf677.jpg) +Figure 4: Comparison of success rate between EMMA and the retrospective LLM expert. As the number of trials increases, the LLM expert progressively improves its performance through retrospective processes, and also the gap between the two agents decreases, indicating the effectiveness of cross-modality imitation. We separately plot the curve per task type in Fig. 9 of Appendix. + +
AgentVisual Env.Textual Env.Success Rate with Template Task Instruction
Avg.PickCleanHeatCoolLookPick2
VMsResNet-18* [56]✓X0.06(-)------
MCNN-FPN* [56]✓X0.05(-)------
LMsBUTLER* [56]X✓0.26(-)0.31(-)0.41(-)0.60(-)0.27(-)0.12(-)0.29(-)
GPT-BUTLER [41]X✓0.69(18.8)0.62(18.4)0.81(18.4)0.85(12.7)0.78(15.6)0.50(24.4)0.47(26.6)
ReAct [73]X✓0.54(20.6)0.71(18.1)0.65(18.8)0.62(18.2)0.44(23.2)0.28(23.7)0.35(25.5)
Reflexion [54]X✓0.91(18.7)0.96(17.4)1.00(17.0)0.81(19.4)0.83(21.6)0.94(16.9)0.88(21.6)
DEPS* [61]X✓0.76(-)0.93(-)0.50(-)0.80(-)1.00(-)1.00(-)0.00(-)
AutoGen* [64]X✓0.77(-)0.92(-)0.74(-)0.78(-)0.86(-)0.83(-)0.41(-)
VLMsMiniGPT-4 [76]✓X0.16(26.9)0.04(29.0)0.00(30.0)0.19(26.3)0.17(26.7)0.67(17.7)0.06(28.9)
BLIP-2 [34]✓X0.04(29.5)0.00(30.0)0.06(29.3)0.04(29.9)0.11(28.2)0.06(29.2)0.00(30.0)
LLaMA-Adapter [17]✓X0.13(27.5)0.17(26.4)0.10(28.6)0.27(24.2)0.22(26.7)0.00(30.0)0.00(30.0)
InstructBLIP [11]✓X0.22(26.2)0.50(21.5)0.26(25.0)0.23(27.2)0.06(28.9)0.17(26.8)0.00(30.0)
EMMA (Ours)✓X0.82(19.5)0.71(19.3)0.94(17.5)0.85(19.6)0.83(19.9)0.88(19.6)0.67(22.4)
Human Performance* [55]✓X0.91(-)------
+ +Table 1: Comparison with the state of the arts. $*$ -reported in previous work. VMs = vision models, LMs = language models, VLMs = vision-language models. "Visual Env." and "Textual Env." refer to the visual and the parallel textual environments from ALFWorld [56], respectively. $\sqrt{} / X$ denotes the corresponding environment used/not used to deploy the agent. The highest scores for each task in the same type of environment are highlighted in bold. The average interaction steps are given in the (parentheses). EMMA substantially outperforms other SOTA VLM agents in the visual environments, and its success also directs a promising way to achieve human-level performance in ALFWorld. + +each task, ALFWorld integrates a visual environment, rendered by the Ai2Thor simulator [30], with a corresponding textual environment. Textual environments employ the Planning Domain Definition Language (PDDL) [1] to translate each pixel observation from the simulator into an equivalent text-based observation, and then construct an interactive world using the TextWorld engine [10]. Fig. 2 provides an illustrative example of the tasks featured in ALFWorld. The tasks within the benchmark are categorized into 6 types: Pick & Place, Clean & Place, Heat & Place, Cool & Place, Look in Light, and Pick Two Objects & Place. Each task requires an agent to execute a series of text-based actions, such as "go to safe 1", "open safe 1", or "heat egg 1 with microwave 1", following a predefined instruction. These actions involve navigating and interacting with the environment. To provide a comprehensive understanding, we have visualized an example of each task type in Fig. 8 of the Appendix. A task in this benchmark may involve interactions with over 10 objects and require more than 30 steps for a human expert to solve, thus challenging an agent's capabilities in long-horizon planning, instruction following, and the utilization of commonsense knowledge. For a fair comparison, we follow the same setting as prior work [41, 54-56, 73] and evaluate all baselines using 134 out-of-distribution (OOD) tasks. + +Baselines. To verify the effectiveness of cross-modality imitation learning, we compare our EMMA with several baselines and SOTA agents using the ALFWorld benchmark with both visual and textual environments. The compared agents can be divided into three categories: vision models, language models, and vision-language models. Concretely, vision models, including ResNet-18 [20] and MCNN-FPN [21], utilize pretrained vision encoders to extract salient features + +from each pixel observation. The extracted features then serve as input for a Multi-Layer Perceptron (MLP) policy, which is trained by behavior cloning on a pre-collected demonstration dataset. Unlike vision models performing in the visual environment, language models complete the same tasks but in a parallel, text-based environment. BUTLER [56] employs a transformer seq2seq model enhanced with a pointer softmax mechanism [19]. This architecture aggregates previous observations as input to generate text-based actions token-by-token. GPT-BUTLER [41], a variant of the GPT-2 model [45], is initially pretrained on a static demonstration dataset and further finetuned using data collected online. ReAct [73] takes a novel approach by utilizing LLMs to generate reasoning traces and task-specific actions in an interleaved manner. This method aids the agent in developing, tracking, and updating its action plans interactively. Reflexion [54] similarly employs an LLM, but it focuses on reflecting upon environmental feedback. It maintains this reflective text in an episodic memory buffer, enhancing the agent's ability to improve actions in subsequent trials. Similar to Reflexion, a concurrent work, DEPS [61], also corrects errors in previous LLM-generated actions by integrating descriptions of the action execution process and providing self-explanations for the feedback. Moreover, beyond the single-agent framework, AutoGen [64] exhibits the potential to accomplish a broad spectrum of tasks through the cooperation of multiple LLM agents. Finally, we consider a range of vision-language models, such as MiniGPT-4 [76], BLIP-2 [34], LLaMA-Adaptor [17], and InstructBLIP [11], as agents to interact with the visual environment. Unlike pure vision or language models, VLMs are designed to process and integrate both visual and textual data, offering a + +more holistic understanding of the environment. To align these agents with the specific requirements of the ALFWorld benchmark, we finetune them on a pre-collected demonstration dataset. This finetuning process is crucial as it enables the agent to comprehend and adhere to ALFWorld's unique grammar and to develop a basic "gamesense". + +# 3.2. Training Details + +The architectural design of EMMA is depicted in Fig. 3. At its core, EMMA employs a Query Transformer (Q-Former) to process visual data. This Q-Former extracts features using a frozen ViT encoder. Its output consists of 32 visual tokens, which are then passed through a linear projection layer before being fed to a frozen LLM decoder. Similar to other VLM agents, EMMA is also finetuned on a pre-collected demonstration dataset, aligning its basic ability with the ALFWorld benchmark. In order to align EMMA with the dynamics of ALFWorld, we train it by imitating an LLM expert (see Alg. 1). We choose text-davinci-003 developed by OpenAI as our LLM expert because of its established capabilities in reasoning and planning [54, 61, 64, 73]. In this setup, text-davinci-003 serves dual roles: it serves as an actor, providing EMMA with expert actions, and as a critic, analyzing EMMA's historical trajectories. This analysis generates retrospective feedback, which is then incorporated into the actor's long-term memory, leading to improved actions in future trials. Further details about the hyperparameters and prompts used in our training procedure are available in Table 2 of the Appendix. + +# 3.3. Comparison with State of the Art + +# EMMA sets new SOTA performance in visual ALFWorld. + +In this section, we compare EMMA with 12 other representative agents on the ALFWorld benchmark, spanning both visual and textual environments (refer to Table 1). We assess two key metrics: the success rate, which is the percentage of trials completed successfully, and the average number of interaction steps required for task completion, with a lower number indicating higher efficiency. EMMA demonstrates superior performance in both metrics, significantly outperforming all VLM agents in visual environments. This achievement underscores the effectiveness of our cross-modality imitation learning approach, as depicted in the learning curve shown in Fig. 4. Furthermore, EMMA's performance markedly exceeds that of VM agents, highlighting the crucial role of the prior knowledge embedded in VLMs. Intriguingly, EMMA's performance is comparable with LLM agents that operate using perfectly semantic descriptions of visual observations. This is largely attributed to EMMA's training strategy of imitating an expert LLM agent, proving to be more efficient than learning from scratch in a purely visual setting. As a result, EMMA stands out as the only VLM agent that substantially surpasses SOTA LLM agents, such + +![](images/661817e8d2e11ec4bf36b68541811cd20f5933ab66c201456828618cbbf7bb2f.jpg) +Figure 5: Comparison of robustness between VLM and LLM agents. Given a specific noise rate, we crop a random portion of the pixel observation and resize it to a given size as the new observation. Similarly, we randomly replace some tokens in the textual observation with arbitrary ones. + +as AutoGen [64] and DEPS [61], in these environments. And its success also directs a potential way to achieve human-level performance in the visual environments of ALFWorld. EMMA is more robust to noisy observations than LLM agents. While LLM agents exhibit a higher success rate with fewer interaction steps in textual environments, as indicated in Table 1, we hypothesize that this superior performance largely relies on their precise semantic abstraction of the environment. However, such an abstraction might not be feasible in real-world applications. To verify this assumption, we set up a more practical scenario where observations are deliberately perturbed at a specific noise rate. We then compare the robustness of EMMA and a SOTA LLM agent, Reflexion, under these noisy observations. To generate noisy observations, a random portion of the visual observation is cropped, resized, and then used to replace the original observation. Similarly, in the textual observation, random tokens are substituted with arbitrary ones. As illustrated in Fig. 5, with the noise rate increasing, EMMA's performance remains significantly more robust than Reflexion. This could be attributed to the vision encoder in the VLM, which is adept at filtering out visual noises. On the other hand, textual noises are directly processed by the LLM, which can substantially impair the performance of LLM-based agents. This finding highlights the potential advantages of VLM agents like EMMA in visual scenarios, in which data is often imperfect and noisy. + +# 3.4. Ablation Study + +Retrospection improves EMMA over time. To assess the importance of the retrospective LLM expert, we present the average success rate of EMMA after each trial and compare it + +![](images/905e47f392e5add8df7e5b62799313afdd903e44e94ef900117cf7d6cc0b2ebf.jpg) +Figure 6: Ablation study. left: The ability of LLM expert to reflect on and learn from past actions appears to be a key factor in achieving the impressive results we observed. middle: The "w/o BC initialization" variant uses a pretrained VLM instead of the reference agent $\pi_{\mathrm{ref}}$ to initialize EMMA. right: The 'w/ CE Loss' variant replaces the DPO loss (described in Eq. 2) with a token-level cross-entropy (CE) loss. Please refer to Sec. 3.4 for a thorough discussion. + +![](images/338e765a491bd9953ea2d10a8a5b075ecbc0ea7e8758ec8632f97b4e55254ff1.jpg) + +![](images/764c580ded6530e9ca9a4cb531669fb436a44f13801208a5de2f71d579b0407c.jpg) + +with a key variation: "EMMA w/o Retrospection", as shown in Fig. 6 (left). This variant of EMMA is trained using the same procedure as the original EMMA but removes the retrospective process. Instead, it relies solely on a plain LLM actor to provide relabeled actions. The results show that EMMA with the retrospective mechanism significantly outperforms its counterpart. This finding is crucial as it indicates that the retrospective process is not just a supplementary feature but a fundamental component of EMMA's architecture that contributes substantially to its enhanced performance. + +EMMA benefits from BC initialization. Through an ablation study, we evaluate the impact of behavior cloning (BC) initialization, a process described in line 3 of Alg. 1. Fig. 6 (middle) demonstrates that EMMA, when deprived of BC initialization, experiences a slight reduction in the average success rate across 134 unseen tasks compared to its original setup. Despite this decrease, EMMA without BC initialization still outperforms other VLM agents, as clearly shown when compared with the results in Table 1. Furthermore, Fig. 10 in the Appendix breaks down the result by task type. It reveals a consistent but slight drop in performance across 5 out of the 6 task types. These results reflect that while BC initialization contributes positively to EMMA's overall performance, it is not critical for achieving the notable results we have reported. + +DPO enables more effective imitation learning. To evaluate the effectiveness of using DPO loss in Eq. (1), we conducted an ablation study with an alternative version of EMMA, referred to "EMMA w/ Cross Entropy (CE) Loss". In this variant, EMMA is optimized using the token-level CE loss, a common objective for finetuning VLMs. The results, as depicted in Fig. 6 (right), reveal that EMMA w/ CE Loss does not achieve the same high success rate as the original EMMA with DPO loss, suggesting that DPO loss contributes to enhancing the upper performance bound of EMMA. In addition, we noted that EMMA w/ CE Loss exhibits faster convergence in the initial training stages than the original + +EMMA. This premature convergence leads to the agent paying attention to the expert actions from easier tasks, usually addressed in the first few training epochs. This can suppress EMMA's exploration and learning on more complex tasks, resulting in worse performance. + +# 3.5. Generalization to Free-form Task Instructions + +The ability of AI agents to precisely follow task instructions given in open-vocabulary and free-form text is crucial for real-world problems. To assess this, we conducted an additional experiment focusing on the generalization capabilities of EMMA, an agent trained with template task instructions. In this experiment, we re-evaluated the performance of EMMA and other baseline agents on 134 unseen tasks, using human-annotated instructions instead of template ones. These human-annotated instructions include a large amount of OOD verbs and objects, presenting a more realistic and challenging scenario for the agents to address. To underscore this challenge, we compared the vocabulary distribution between the template and human-annotated instructions, as shown in Fig. 7 (right). Moreover, we provide a comprehensive analysis of the vocabulary used across both instruction types in Fig. 11 of the Appendix. In Fig. 7 (left), EMMA demonstrates a slight performance decline, while a significant degradation is observed in other baselines. We also note that Reflexion, a SOTA LLM agent, exhibits exceptional generalization to those OOD instructions. According to these empirical results in Fig. 7, we have the following conclusions: (1) EMMA obtains and benefits from the generalization capabilities inherent in the SOTA LLM agent through cross-modality imitation learning; (2) Our work sheds novel insights on using LLM feedback to train more versatile and generalizable embodied agents. + +# 4. Related Work + +Agents based on Foundation Models. Recent research has increasingly focused on harnessing the capabilities of large + +
AgentVisual Env.Textual Env.Success Rate with Free-form Task Instruction
Avg.PickCleanHeatCoolLookPick2
LMsBUTLER* [56]X✓0.03 (-)0.10 (-)0.22 (-)0.05 (-)0.17 (-)0.00 (-)0.00 (-)
GPT-BUTLER [41]X✓0.31 (24.7)0.25 (25.4)0.42 (23.5)0.46 (20.9)0.44 (21.8)0.00 (30.0)0.12 (28.9)
ReAct [73]X✓0.37 (23.6)0.46 (22.0)0.39 (23.8)0.50 (20.5)0.28 (24.6)0.22 (26.4)0.29 (25.9)
Reflexion [54]X✓0.78 (17.0)0.84 (15.8)0.74 (18.4)0.77 (17.2)0.61 (19.4)0.94 (13.8)0.82 (16.7)
VLMsMiniGPT-4 [76]✓X0.00 (30.0)0.00 (30.0)0.00 (30.0)0.00 (30.0)0.00 (30.0)0.00 (30.0)0.00 (30.0)
BLIP-2 [34]✓X0.01 (29.7)0.04 (29.1)0.03 (29.5)0.00 (30.0)0.00 (30.0)0.00 (30.0)0.00 (30.0)
LLaMA-Adapter [17]✓X0.02 (29.6)0.04 (29.3)0.03 (29.4)0.04 (29.3)0.00 (30.0)0.00 (30.0)0.00 (30.0)
InstructBLIP [11]✓X0.01 (29.8)0.00 (30.0)0.03 (29.3)0.00 (30.0)0.00 (30.0)0.00 (30.0)0.00 (30.0)
EMMA (Ours)✓X0.68 (22.0)0.72 (21.7)0.65 (22.7)0.72 (22.0)0.56 (23.9)0.67 (23.3)0.76 (18.1)
+ +![](images/429e678f6864656b282ff8d85ee02fb2cc44319c20b241abdf6e173f0fae4930.jpg) +Figure 7: Generalization to open-vocabulary, free-form task instructions. Left: These instructions are provided by different human annotators using Amazon Mechanical Turk [55]. Right: Frequency distribution of verbs for human-annotated and templated task instructions. + +pre-trained foundation models to build AI agents [65, 69]. These models (e.g., LLMs), benefiting from their commonsense knowledge inherited from Internet-scale pretraining, are able to reason actions according to descriptions of the external environments. For example, given a set of task instructions, LLMs can be elaborately prompted to perform as agents generating high-level step-by-step plans [7, 24, 54, 64, 73], and each step can be parsed into a sequence of robotic actions that are executed via pretrained policies or available APIs [6, 36, 59]. Furthermore, by using VLMs [32], plans can also be conditioned on visual inputs that are transformed into language descriptions [7, 16, 25, 57] or token embeddings aligned with LLMs [13, 42, 68, 77]. However, existing foundation models are usually pretrained on static text or text-image datasets and thus may struggle to align with the dynamics of the world. To bridge this gap, we study how to finetune a VLM to be an embodied agent dynamically aligned with the world by distilling the cross-modality knowledge from an LLM expert. The work most closely related to ours is EUREKA [40], which also explores using source information provided by the simulator as the context of an LLM to aid agent training. Instead of directly mimicking the output of the LLM as we did, EUREKA harnesses the coding LLM to generate a desired reward function for a given task and optimizes a policy against the reward function using RL, leading to a more complex and expensive training procedure [47, 52, 53]. + +Imitation Learning. Imitation learning is the study of algorithms that improve performance by mimicking an expert's decisions and behaviors. We summarize three main categories of existing methods in the following: (1) behavior cloning (BC), (2) inverse reinforcement learning (IRL), and (3) the combination of imitation and reinforcement learning. The naive BC [5] ignores the change in distribution and simply trains a policy that only performs well under the distribution of states visited by the expert. Following works, such as dataset aggregation [51] or policy aggregation [28, 49], have been proposed to address the limitations of BC. Another line of work, IRL, is a more complicated algorithm framework that learns the reward function from + +expert demonstrations and then improves the policy using RL with the learned reward. A representative method in this category is generative adversarial imitation learning (GAIL) [23], in which a policy and a discriminator compete with each other in order to maximize the likelihood of the policy's behavior matching the expert. The third category of methods usually leverages an IL policy to initialize RL and continues to boost its performance via online collected data from RL [8, 50, 58]. This simple combination significantly improves RL's sample efficiency and IL's upper performance bound constrained by the expert. Nevertheless, all of the above methods assume that the expert and the imitator understand the world in the same modality, and thus overlook the fact that the complementary knowledge from other modalities often boosts the model's accuracy and generalization dramatically [46, 67]. + +# 5. Conclusion + +We create EMMA, an Embodied Multi-Modal Agent, by finetuning a VLM in an embodied visual world with interactive imitation learning from an LLM expert in a parallel text world, who produces better actions and retrospective feedback to VLM's trajectories. Such imitation learning exhibits substantial advantages over vision or VLM policies directly finetuned in the visual world or finetuned by behavior cloning of a rule-based expert, and SOTA API VLMs such as GPT-4V(ison). As a result, EMMA achieves a comparable success rate and much better robustness in the noisy visual world than its LLM teacher in the easier text world. Furthermore, EMMA shows powerful generalization to open-vocabulary and free-form task instructions, highlighting its potential in real-world scenarios. + +# Acknowledgment + +This work is partially supported by the National Key R&D Program of China under the Grant No. (2023YFE0106300) and (2017YFC0804002), Shenzhen Fundamental Research Program under the Grant No. (JCYJ20200109141235597), and Program for Guangdong Introducing Innovative and Entrepreneurial Teams under Grant No. (2017ZT07X386). + +# References + +[1] Constructions Aeronautiques, Adele Howe, Craig Knoblock, ISI Drew McDermott, Ashwin Ram, Manuela Veloso, Daniel Weld, David Wilkins SRI, Anthony Barrett, Dave Christianson, et al. PDDL: the planning domain definition language. Technical Report, Tech. Rep., 1998. 4, 5, 3 +[2] Marcin Andrychowicz, Filip Wolski, Alex Ray, Jonas Schneider, Rachel Fong, Peter Welinder, Bob McGrew, Josh Tobin, OpenAI Pieter Abbeel, and Wojciech Zaremba. Hindsight experience replay. 2017. 2 +[3] Shuang Ao, Tianyi Zhou, Guodong Long, Qinghua Lu, Liming Zhu, and Jing Jiang. Co-pilot: Collaborative planning and reinforcement learning on sub-task curriculum. 2021. +[4] Shuang Ao, Tianyi Zhou, Jing Jiang, Guodong Long, Xuan Song, and Chengqi Zhang. Eat-c: environment-adversarial sub-task curriculum for efficient reinforcement learning. In ICML, 2022. 2 +[5] Michael Bain and Claude Sammut. A framework for behavioural cloning. In Machine Intelligence, 1995. 8 +[6] Anthony Brohan, Noah Brown, Justice Carbajal, Yevgen Chebotar, Joseph Dabis, Chelsea Finn, Keerthana Gopalakrishnan, Karol Hausman, Alex Herzog, Jasmine Hsu, et al. Rt-1: Robotics transformer for real-world control at scale. arXiv preprint arXiv:2212.06817, 2022. 2, 4, 8 +[7] Anthony Brohan, Yevgen Chebotar, Chelsea Finn, Karol Hausman, Alexander Herzog, Daniel Ho, Julian Ibarz, Alex Irpan, Eric Jang, Ryan Julian, et al. Do as i can, not as i say: Grounding language in robotic affordances. In CoRL, 2023. 8 +[8] Kai-Wei Chang, Akshay Krishnamurthy, Alekh Agarwal, Hal Daumé III, and John Langford. Learning to search better than your teacher. In ICML, 2015. 8 +[9] Lichang Chen, Jiuhai Chen, Tom Goldstein, Heng Huang, and Tianyi Zhou. Instructzero: Efficient instruction optimization for black-box large language models. arXiv preprint arXiv:2306.03082, 2023. 4 +[10] Marc-Alexandre Côté, Akos Kádár, Xingdi Yuan, Ben Kybartas, Tavian Barnes, Emery Fine, James Moore, Matthew Hausknecht, Layla El Asri, Mahmoud Adada, et al. Textworld: A learning environment for text-based games. In Computer Games Workshop, IJCAI, 2019. 4, 5 +[11] Wenliang Dai, Junnan Li, Dongxu Li, Anthony Tiong, Junqi Zhao, Weisheng Wang, Boyang Li, Pascale Fung, and Steven Hoi. InstructBLIP: Towards general-purpose vision-language models with instruction tuning. In NeurIPS, 2023. 2, 3, 5, 8 +[12] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. BERT: pre-training of deep bidirectional transformers for language understanding. In NAACL-HLT, 2019. 3 +[13] Danny Driess, Fei Xia, Mehdi SM Sajjadi, Corey Lynch, Aakanksha Chowdhery, Brian Ichter, Ayzaan Wahid, Jonathan Thompson, Quan Vuong, Tianhe Yu, et al. Palm-e: An embodied multimodal language model. arXiv preprint arXiv:2303.03378, 2023. 2, 4, 8 + +[14] Meng Fang, Tianyi Zhou, Yali Du, Lei Han, and Zhengyou Zhang. Curriculum-guided hindsight experience replay. NeurIPS, 2019. 2 +[15] Daocheng Fu, Xin Li, Licheng Wen, Min Dou, Pinlong Cai, Botian Shi, and Yu Qiao. Drive like a human: Rethinking autonomous driving with large language models. arXiv preprint arXiv:2307.07162, 2023. 2, 4 +[16] Jensen Gao, Bidipta Sarkar, Fei Xia, Ted Xiao, Jiajun Wu, Brian Ichter, Anirudha Majumdar, and Dorsa Sadigh. Physically grounded vision-language models for robotic manipulation. arXiv preprint arXiv:2309.02561, 2023. 8 +[17] Peng Gao, Jiaming Han, Renrui Zhang, Ziyi Lin, Shijie Geng, Aojun Zhou, Wei Zhang, Pan Lu, Conghui He, Xiangyu Yue, et al. Llama-adapter v2: Parameter-efficient visual instruction model. arXiv preprint arXiv:2304.15010, 2023. 5, 8, 3 +[18] Jiayi Guan, Guang Chen, Jiaming Ji, Long Yang, Zhijun Li, et al. Voce: Variational optimization with conservative estimation for offline safe reinforcement learning. NeurIPS, 36, 2023. 2 +[19] Caglar Gülcehre, Sungjin Ahn, Ramesh Nallapati, Bowen Zhou, and Yoshua Bengio. Pointing the unknown words. In ACL, 2016. 5 +[20] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, 2016. 5 +[21] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross B. Girshick. Mask R-CNN. In ICCV, 2017. 5 +[22] Malte Helmert. The fast downward planning system. Journal of Artificial Intelligence Research, 2006. 3 +[23] Jonathan Ho and Stefano Ermon. Generative adversarial imitation learning. In NeurIPS, 2016. 8 +[24] Wenlong Huang, Pieter Abbeel, Deepak Pathak, and Igor Mordatch. Language models as zero-shot planners: Extracting actionable knowledge for embodied agents. In ICML, 2022. 8 +[25] Wenlong Huang, Chen Wang, Ruohan Zhang, Yunzhu Li, Jiajun Wu, and Li Fei-Fei. Voxposer: Composable 3d value maps for robotic manipulation with language models. arXiv preprint arXiv:2307.05973, 2023. 2, 4, 8 +[26] Ahmed Hussein, Mohamed Medhat Gaber, Eyad Elyan, and Chrisina Jayne. Imitation learning: A survey of learning methods. ACM Computing Surveys (CSUR), 50(2):1-35, 2017. 4 +[27] Marcus Hutter. Universal artificial intelligence: Sequential decisions based on algorithmic probability. Springer Science & Business Media, 2004. 1 +[28] Hal Daumé III, John Langford, and Daniel Marcu. Search-based structured prediction. Mach. Learn., 2009. 8 +[29] Yunfan Jiang, Agrim Gupta, Zichen Zhang, Guanzhi Wang, Yongqiang Dou, Yanjun Chen, Li Fei-Fei, Anima Anandkumar, Yuke Zhu, and Linxi Fan. VIMA: robot manipulation with multimodal prompts. In ICML, 2023. 2, 4 +[30] Eric Kolve, Roozbeh Mottaghi, Winson Han, Eli VanderBilt, Luca Weihs, Alvaro Herrasti, Matt Deitke, Kiana Ehsani, Daniel Gordon, Yuke Zhu, et al. Ai2-thor: An interactive 3d environment for visual ai. arXiv preprint arXiv:1712.05474, 2017. 4, 5, 1 + +[31] Sergey Levine, Aviral Kumar, George Tucker, and Justin Fu. Offline reinforcement learning: Tutorial, review, and perspectives on open problems. arXiv preprint arXiv:2005.01643, 2020. 2, 3 +[32] Chunyuan Li, Zhe Gan, Zhengyuan Yang, Jianwei Yang, Linjie Li, Lijuan Wang, and Jianfeng Gao. Multimodal foundation models: From specialists to general-purpose assistants. arXiv preprint arXiv:2309.10020, 2023. 2, 3, 8 +[33] Dongxu Li, Junnan Li, Hung Le, Guangsen Wang, Silvio Savarese, and Steven C.H. Hoi. LAVIS: A one-stop library for language-vision intelligence. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 3: System Demonstrations), 2023. 3 +[34] Junnan Li, Dongxu Li, Silvio Savarese, and Steven C. H. Hoi. BLIP-2: bootstrapping language-image pre-training with frozen image encoders and large language models. In ICML, 2023. 2, 3, 5, 8 +[35] Chen Liang, Jiahui Yu, Ming-Hsuan Yang, Matthew Brown, Yin Cui, Tuo Zhao, Boqing Gong, and Tianyi Zhou. Modulewise adaptive distillation for multimodality foundation models. 2024. 2 +[36] Jacky Liang, Wenlong Huang, Fei Xia, Peng Xu, Karol Hausman, Brian Ichter, Pete Florence, and Andy Zeng. Code as policies: Language model programs for embodied control. In ICRA, 2023. 8 +[37] Yong Lin, Lu Tan, Hangyu Lin, Zeming Zheng, Renjie Pi, Jipeng Zhang, Shizhe Diao, Haoxiang Wang, Han Zhao, Yuan Yao, et al. Speciality vs generality: An empirical study on catastrophic forgetting in fine-tuning foundation models. arXiv preprint arXiv:2309.06256, 2023. 3 +[38] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. arXiv preprint arXiv:2304.08485, 2023. 2, 3 +[39] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 3 +[40] Yecheng Jason Ma, William Liang, Guanzhi Wang, De-An Huang, Osbert Bastani, Dinesh Jayaraman, Yuke Zhu, Linxi Fan, and Anima Anandkumar. Eureka: Human-level reward design via coding large language models. arXiv preprint arXiv:2310.12931, 2023. 8 +[41] Vincent Micheli and François Fleuret. Language models are few-shot butlers. In EMNLP, 2021. 5, 8 +[42] Yao Mu, Qinglong Zhang, Mengkang Hu, Wenhai Wang, Mingyu Ding, Jun Jin, Bin Wang, Jifeng Dai, Yu Qiao, and Ping Luo. Embodiedgpt: Vision-language pre-training via embodied chain of thought. arXiv preprint arXiv:2305.15021, 2023. 2, 4, 8 +[43] OpenAI. GPT-4V(ison) system card. https://cdn.openai.com/papers/GPTV_System_Card.pdf, 2023.2 +[44] Abhishek Padalkar, Acorn Pooley, Ajinkya Jain, Alex Bewley, Alex Herzog, Alex Irpan, Alexander Khazatsky, Anant Rai, Anikait Singh, Anthony Brohan, et al. Open x-embediment: Robotic learning datasets and rt-x models, 2023. 3 +[45] Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, Ilya Sutskever, et al. Language models are unsupervised multitask learners. OpenAI blog, 2019. 5 + +[46] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, 2021. 8, 3 +[47] Rafael Rafailov, Archit Sharma, Eric Mitchell, Stefano Ermon, Christopher D Manning, and Chelsea Finn. Direct preference optimization: Your language model is secretly a reward model. arXiv preprint arXiv:2305.18290, 2023. 8 +[48] Rafael Rafailov, Archit Sharma, Eric Mitchell, Christopher D Manning, Stefano Ermon, and Chelsea Finn. Direct preference optimization: Your language model is secretly a reward model. NeurIPS, 2024. 2, 4 +[49] Stéphane Ross and Drew Bagnell. Efficient reductions for imitation learning. In AISTATS, 2010. 8 +[50] Stephane Ross and J Andrew Bagnell. Reinforcement and imitation learning via interactive no-regret learning. arXiv preprint arXiv:1406.5979, 2014. 8 +[51] Stéphane Ross, Geoffrey J. Gordon, and Drew Bagnell. A reduction of imitation learning and structured prediction to no-regret online learning. In AISTATS, 2011. 2, 4, 8 +[52] John Schulman, Sergey Levine, Pieter Abbeel, Michael I. Jordan, and Philipp Moritz. Trust region policy optimization. In ICML, 2015. 8 +[53] John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Radford, and Oleg Klimov. Proximal policy optimization algorithms. arXiv preprint arXiv:1707.06347, 2017. 8 +[54] Noah Shinn, Federico Cassano, Ashwin Gopinath, Karthik R Narasimhan, and Shunyu Yao. Reflexion: Language agents with verbal reinforcement learning. In NeurIPS, 2023. 3, 4, 5, 6, 8 +[55] Mohit Shridhar, Jesse Thomason, Daniel Gordon, Yonatan Bisk, Winson Han, Roozbeh Mottaghi, Luke Zettlemoyer, and Dieter Fox. ALFRED: A benchmark for interpreting grounded instructions for everyday tasks. In CVPR, 2020. 5, 8, 3, 6 +[56] Mohit Shridhar, Xingdi Yuan, Marc-Alexandre Côté, Yonatan Bisk, Adam Trischler, and Matthew J. Hausknecht. Alfworld: Aligning text and embodied environments for interactive learning. In ICLR, 2021. 2, 3, 4, 5, 8, 1 +[57] Theodore Sumers, Kenneth Marino, Arun Ahuja, Rob Fergus, and Ishita Dasgupta. Distilling internet-scale vision-language models into embodied agents. In ICML, 2023. 2, 4, 8 +[58] Wen Sun, J. Andrew Bagnell, and Byron Boots. Truncated horizon policy search: Combining reinforcement learning & imitation learning. In ICLR, 2018. 8 +[59] Guanzhi Wang, Yuqi Xie, Yunfan Jiang, Ajay Mandlekar, Chaowei Xiao, Yuke Zhu, Linxi Fan, and Anima Anandkumar. Voyager: An open-ended embodied agent with large language models. arXiv preprint arXiv:2305.16291, 2023. 8 +[60] Xiyao Wang, Yuhang Zhou, Xiaoyu Liu, Hongjin Lu, Yuancheng Xu, Feihong He, Jaehong Yoon, Taixi Lu, Gedas Bertasius, Mohit Bansal, et al. Mementos: A comprehensive benchmark for multimodal large language model reasoning over image sequences. arXiv preprint arXiv:2401.10529, 2024. 2 + +[61] Zihao Wang, Shaofei Cai, Anji Liu, Xiaojian Ma, and Yitao Liang. Describe, explain, plan and select: Interactive planning with large language models enables open-world multi-task agents. arXiv preprint arXiv:2302.01560, 2023. 5, 6 +[62] Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Brian Ichter, Fei Xia, Ed H. Chi, Quoc V. Le, and Denny Zhou. Chain-of-thought prompting elicits reasoning in large language models. In NeurIPS, 2022. 3 +[63] Michael Wooldridge and Nicholas R Jennings. Intelligent agents: Theory and practice. The knowledge engineering review, 10(2):115-152, 1995. 1 +[64] Qingyun Wu, Gagan Bansal, Jieyu Zhang, Yiran Wu, Shaokun Zhang, Erkang Zhu, Beibin Li, Li Jiang, Xiaoyun Zhang, and Chi Wang. Autogen: Enabling next-gen llm applications via multi-agent conversation framework. arXiv preprint arXiv:2308.08155, 2023. 5, 6, 8 +[65] Zhiheng Xi, Wenxiang Chen, Xin Guo, Wei He, Yiwen Ding, Boyang Hong, Ming Zhang, Junzhe Wang, Senjie Jin, Enyu Zhou, et al. The rise and potential of large language model based agents: A survey. arXiv preprint arXiv:2309.07864, 2023. 8 +[66] Zhenhua Xu, Yujia Zhang, Enze Xie, Zhen Zhao, Yong Guo, Kenneth KY Wong, Zhenguo Li, and Hengshuang Zhao. Drivegpt4: Interpretable end-to-end autonomous driving via large language model. arXiv preprint arXiv:2310.01412, 2023. 2, 4 +[67] Zihui Xue, Zhengqi Gao, Sucheng Ren, and Hang Zhao. The modality focusing hypothesis: Towards understanding cross-modal knowledge distillation. In ICLR, 2023. 8 +[68] Jingkang Yang, Yuhao Dong, Shuai Liu, Bo Li, Ziyue Wang, Chencheng Jiang, Haoran Tan, Jiamu Kang, Yuanhan Zhang, Kaiyang Zhou, et al. Octopus: Embodied vision-language programmer from environmental feedback. arXiv preprint arXiv:2310.08588, 2023. 3, 4, 8 +[69] Sherry Yang, Ofir Nachum, Yilun Du, Jason Wei, Pieter Abbeel, and Dale Schuurmans. Foundation models for decision making: Problems, methods, and opportunities. arXiv preprint arXiv:2303.04129, 2023. 8 +[70] Yijun Yang, Jing Jiang, Tianyi Zhou, Jie Ma, and Yuhui Shi. Pareto policy pool for model-based offline reinforcement learning. In ICLR, 2022. 2 +[71] Yijun Yang, Tianyi Zhou, Jing Jiang, Guodong Long, and Yuhui Shi. Continual task allocation in meta-policy network via sparse prompting. In ICML, 2023. 2 +[72] Shunyu Yao, Dian Yu, Jeffrey Zhao, Izhak Shafran, Thomas L Griffiths, Yuan Cao, and Karthik Narasimhan. Tree of thoughts: Deliberate problem solving with large language models. arXiv preprint arXiv:2305.10601, 2023. 3 +[73] Shunyu Yao, Jeffrey Zhao, Dian Yu, Nan Du, Izhak Shafran, Karthik R. Narasimhan, and Yuan Cao. React: Synergizing reasoning and acting in language models. In ICLR, 2023. 3, 5, 6, 8, 2 +[74] Lianmin Zheng, Wei-Lin Chiang, Ying Sheng, Siyuan Zhuang, Zhanghao Wu, Yonghao Zhuang, Zi Lin, Zhuohan Li, Dacheng Li, Eric Xing, et al. Judging llm-as-a-judge with mt-bench and chatbot arena. arXiv preprint arXiv:2306.05685, 2023. 3 + +[75] Victor Zhong, Dipendra Misra, Xingdi Yuan, and Marc-Alexandre Côté. Policy improvement using language feedback models. arXiv preprint arXiv:2402.07876, 2024. 2 +[76] Deyao Zhu, Jun Chen, Xiaogian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models. arXiv preprint arXiv:2304.10592, 2023. 2, 3, 5, 8 +[77] Brianna Zitkovich, Tianhe Yu, Sichun Xu, Peng Xu, Ted Xiao, Fei Xia, Jialin Wu, Paul Wohlhart, Stefan Welker, Ayzaan Wahid, et al. Rt-2: Vision-language-action models transfer web knowledge to robotic control. In CoRL, 2023. 2, 3, 4, 8 \ No newline at end of file diff --git a/2024/Embodied Multi-Modal Agent trained by an LLM from a Parallel TextWorld/images.zip b/2024/Embodied Multi-Modal Agent trained by an LLM from a Parallel TextWorld/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..d3b7f182ea53a3d6d819e459e93aeb93cd3b4a90 --- /dev/null +++ b/2024/Embodied Multi-Modal Agent trained by an LLM from a Parallel TextWorld/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:fd50ac56a7c7968250dd805bfef1d7332407bbde517e3e9862479e9f086e0346 +size 585119 diff --git a/2024/Embodied Multi-Modal Agent trained by an LLM from a Parallel TextWorld/layout.json b/2024/Embodied Multi-Modal Agent trained by an LLM from a Parallel TextWorld/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..f53b3ba1d3f37f653fd6c7103e1fefb828eda84c --- /dev/null +++ b/2024/Embodied Multi-Modal Agent trained by an LLM from a Parallel TextWorld/layout.json @@ -0,0 +1,9034 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 59, + 103, + 536, + 122 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 103, + 536, + 122 + ], + "spans": [ + { + "bbox": [ + 59, + 103, + 536, + 122 + ], + "type": "text", + "content": "Embodied Multi-Modal Agent trained by an LLM from a Parallel TextWorld" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 111, + 141, + 482, + 158 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 111, + 141, + 482, + 158 + ], + "spans": [ + { + "bbox": [ + 111, + 141, + 482, + 158 + ], + "type": "text", + "content": "Yijun Yang" + }, + { + "bbox": [ + 111, + 141, + 482, + 158 + ], + "type": "inline_equation", + "content": "^{1,5,4}" + }, + { + "bbox": [ + 111, + 141, + 482, + 158 + ], + "type": "text", + "content": " Tianyi Zhou" + }, + { + "bbox": [ + 111, + 141, + 482, + 158 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 111, + 141, + 482, + 158 + ], + "type": "text", + "content": " Kanxue Li" + }, + { + "bbox": [ + 111, + 141, + 482, + 158 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 111, + 141, + 482, + 158 + ], + "type": "text", + "content": " Dapeng Tao" + }, + { + "bbox": [ + 111, + 141, + 482, + 158 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 111, + 141, + 482, + 158 + ], + "type": "text", + "content": " Lusong Li" + }, + { + "bbox": [ + 111, + 141, + 482, + 158 + ], + "type": "inline_equation", + "content": "^{4}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 166, + 157, + 443, + 172 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 157, + 443, + 172 + ], + "spans": [ + { + "bbox": [ + 166, + 157, + 443, + 172 + ], + "type": "text", + "content": "Li Shen" + }, + { + "bbox": [ + 166, + 157, + 443, + 172 + ], + "type": "inline_equation", + "content": "^{4,*}" + }, + { + "bbox": [ + 166, + 157, + 443, + 172 + ], + "type": "text", + "content": " Xiaodong He" + }, + { + "bbox": [ + 166, + 157, + 443, + 172 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 166, + 157, + 443, + 172 + ], + "type": "text", + "content": " Jing Jiang" + }, + { + "bbox": [ + 166, + 157, + 443, + 172 + ], + "type": "inline_equation", + "content": "^{5}" + }, + { + "bbox": [ + 166, + 157, + 443, + 172 + ], + "type": "text", + "content": " Yuhui Shi" + }, + { + "bbox": [ + 166, + 157, + 443, + 172 + ], + "type": "inline_equation", + "content": "^{1,*}" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 78, + 172, + 515, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 78, + 172, + 515, + 186 + ], + "spans": [ + { + "bbox": [ + 78, + 172, + 515, + 186 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 78, + 172, + 515, + 186 + ], + "type": "text", + "content": "Southern University of Science and Technology " + }, + { + "bbox": [ + 78, + 172, + 515, + 186 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 78, + 172, + 515, + 186 + ], + "type": "text", + "content": "University of Maryland, College Park" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 96, + 186, + 496, + 201 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 96, + 186, + 496, + 201 + ], + "spans": [ + { + "bbox": [ + 96, + 186, + 496, + 201 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 96, + 186, + 496, + 201 + ], + "type": "text", + "content": "Yunnan University " + }, + { + "bbox": [ + 96, + 186, + 496, + 201 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 96, + 186, + 496, + 201 + ], + "type": "text", + "content": "JD Explore Academy " + }, + { + "bbox": [ + 96, + 186, + 496, + 201 + ], + "type": "inline_equation", + "content": "^{5}" + }, + { + "bbox": [ + 96, + 186, + 496, + 201 + ], + "type": "text", + "content": "University of Technology Sydney" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 171, + 203, + 419, + 214 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 171, + 203, + 419, + 214 + ], + "spans": [ + { + "bbox": [ + 171, + 203, + 419, + 214 + ], + "type": "text", + "content": "https://github.com/stevenyangyj/Emma-Alfworld" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 61, + 220, + 190, + 361 + ], + "blocks": [ + { + "bbox": [ + 61, + 220, + 190, + 361 + ], + "lines": [ + { + "bbox": [ + 61, + 220, + 190, + 361 + ], + "spans": [ + { + "bbox": [ + 61, + 220, + 190, + 361 + ], + "type": "image", + "image_path": "dd386abecb87701148aa9f7c71f19256df3bf2e1083b59acc935a55ebaca7ade.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 66, + 364, + 185, + 376 + ], + "lines": [ + { + "bbox": [ + 66, + 364, + 185, + 376 + ], + "spans": [ + { + "bbox": [ + 66, + 364, + 185, + 376 + ], + "type": "text", + "content": "Instruction: put the gold vase into safe." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 66, + 379, + 110, + 391 + ], + "blocks": [ + { + "bbox": [ + 66, + 379, + 110, + 391 + ], + "lines": [ + { + "bbox": [ + 66, + 379, + 110, + 391 + ], + "spans": [ + { + "bbox": [ + 66, + 379, + 110, + 391 + ], + "type": "image", + "image_path": "c81a71c3b7afa35b3b05f6787e1a509271383280f73a66b5aae128bd44839227.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 78, + 395, + 129, + 403 + ], + "lines": [ + { + "bbox": [ + 78, + 395, + 129, + 403 + ], + "spans": [ + { + "bbox": [ + 78, + 395, + 129, + 403 + ], + "type": "text", + "content": "Action succeeded" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 65, + 393, + 76, + 406 + ], + "blocks": [ + { + "bbox": [ + 65, + 393, + 76, + 406 + ], + "lines": [ + { + "bbox": [ + 65, + 393, + 76, + 406 + ], + "spans": [ + { + "bbox": [ + 65, + 393, + 76, + 406 + ], + "type": "image", + "image_path": "5ff5616d4cb5341f71d5f08bc9ee781a70d2fd4702a88a64235b65077caba097.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 78, + 407, + 169, + 415 + ], + "lines": [ + { + "bbox": [ + 78, + 407, + 169, + 415 + ], + "spans": [ + { + "bbox": [ + 78, + 407, + 169, + 415 + ], + "type": "text", + "content": "Action failed due to hallucination" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 65, + 407, + 75, + 418 + ], + "blocks": [ + { + "bbox": [ + 65, + 407, + 75, + 418 + ], + "lines": [ + { + "bbox": [ + 65, + 407, + 75, + 418 + ], + "spans": [ + { + "bbox": [ + 65, + 407, + 75, + 418 + ], + "type": "image", + "image_path": "18db195502585bb1f6a854aa47c535506f0c4862931e8d7447df405b1bfd1f64.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 78, + 418, + 179, + 427 + ], + "lines": [ + { + "bbox": [ + 78, + 418, + 179, + 427 + ], + "spans": [ + { + "bbox": [ + 78, + 418, + 179, + 427 + ], + "type": "text", + "content": "Action failed due to invalid grammar" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 65, + 418, + 76, + 429 + ], + "blocks": [ + { + "bbox": [ + 65, + 418, + 76, + 429 + ], + "lines": [ + { + "bbox": [ + 65, + 418, + 76, + 429 + ], + "spans": [ + { + "bbox": [ + 65, + 418, + 76, + 429 + ], + "type": "image", + "image_path": "a33d8ecb5f073cfb94e2660f00a0ff46b8e2329182a418cf5e037e211d913d8e.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 443, + 547, + 489 + ], + "lines": [ + { + "bbox": [ + 46, + 443, + 547, + 489 + ], + "spans": [ + { + "bbox": [ + 46, + 443, + 547, + 489 + ], + "type": "text", + "content": "Figure 1: Comparison of three VLM-based agents in the visual environment from ALFWorld. Given the task instruction and the current-step observation as inputs, a VLM agent is expected to predict an action, e.g., \"go to cabinet 1\", towards completing the task. (a) GPT-4V(ison). (b) InstructBLIP finetuned by behavior cloning on a static dataset of demonstrations by a rule-based expert in the visual simulator. (c) Our Embodied Multi-Modal Agent (EMMA) trained via cross-modality interactive imitation learning from an LLM expert." + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 194, + 222, + 271, + 289 + ], + "blocks": [ + { + "bbox": [ + 194, + 222, + 271, + 289 + ], + "lines": [ + { + "bbox": [ + 194, + 222, + 271, + 289 + ], + "spans": [ + { + "bbox": [ + 194, + 222, + 271, + 289 + ], + "type": "image", + "image_path": "5ec5a05645cce7c2d08386c06e6bff99f89aa78dcc783f8772b75793136bcdff.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 272, + 224, + 335, + 289 + ], + "blocks": [ + { + "bbox": [ + 272, + 224, + 335, + 289 + ], + "lines": [ + { + "bbox": [ + 272, + 224, + 335, + 289 + ], + "spans": [ + { + "bbox": [ + 272, + 224, + 335, + 289 + ], + "type": "image", + "image_path": "3da630c5985c6d006562dd50b1a04300008e6b4a4c614da84e587230d60e8a74.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 337, + 224, + 402, + 288 + ], + "blocks": [ + { + "bbox": [ + 337, + 224, + 402, + 288 + ], + "lines": [ + { + "bbox": [ + 337, + 224, + 402, + 288 + ], + "spans": [ + { + "bbox": [ + 337, + 224, + 402, + 288 + ], + "type": "image", + "image_path": "67f671790556931a529d157a1a77a160e4f69365c4fe995c70425a0bd77ce609.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 402, + 224, + 466, + 288 + ], + "blocks": [ + { + "bbox": [ + 402, + 224, + 466, + 288 + ], + "lines": [ + { + "bbox": [ + 402, + 224, + 466, + 288 + ], + "spans": [ + { + "bbox": [ + 402, + 224, + 466, + 288 + ], + "type": "image", + "image_path": "b88796c2e9a6ae5a18343a0592fbad2eae4e6b26402607e92f072b2e91e02ddb.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 467, + 224, + 533, + 289 + ], + "blocks": [ + { + "bbox": [ + 467, + 224, + 533, + 289 + ], + "lines": [ + { + "bbox": [ + 467, + 224, + 533, + 289 + ], + "spans": [ + { + "bbox": [ + 467, + 224, + 533, + 289 + ], + "type": "image", + "image_path": "b9fdd5cbef6e4a55ad6a26e2c8e64e59ac3e9b74c2a8c9ef4d27fdb47d0c1eb5.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 194, + 294, + 271, + 360 + ], + "blocks": [ + { + "bbox": [ + 120, + 380, + 172, + 390 + ], + "lines": [ + { + "bbox": [ + 120, + 380, + 172, + 390 + ], + "spans": [ + { + "bbox": [ + 120, + 380, + 172, + 390 + ], + "type": "text", + "content": "Target object" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 194, + 294, + 271, + 360 + ], + "lines": [ + { + "bbox": [ + 194, + 294, + 271, + 360 + ], + "spans": [ + { + "bbox": [ + 194, + 294, + 271, + 360 + ], + "type": "image", + "image_path": "deecaff978c6d4646e43fefaadabf18ae535fc4e530aeaa4227e7b38e7c09699.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 272, + 295, + 336, + 360 + ], + "blocks": [ + { + "bbox": [ + 272, + 295, + 336, + 360 + ], + "lines": [ + { + "bbox": [ + 272, + 295, + 336, + 360 + ], + "spans": [ + { + "bbox": [ + 272, + 295, + 336, + 360 + ], + "type": "image", + "image_path": "8513336054da0b8916b21dca282448e7ffe438b335339d1fac6f88408253abf2.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 337, + 295, + 402, + 360 + ], + "blocks": [ + { + "bbox": [ + 337, + 295, + 402, + 360 + ], + "lines": [ + { + "bbox": [ + 337, + 295, + 402, + 360 + ], + "spans": [ + { + "bbox": [ + 337, + 295, + 402, + 360 + ], + "type": "image", + "image_path": "7f74171183cc4f69c9be4b9cf20c7f8df12aa3e3d167f88a659fbf378d17cee7.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 402, + 295, + 466, + 360 + ], + "blocks": [ + { + "bbox": [ + 402, + 295, + 466, + 360 + ], + "lines": [ + { + "bbox": [ + 402, + 295, + 466, + 360 + ], + "spans": [ + { + "bbox": [ + 402, + 295, + 466, + 360 + ], + "type": "image", + "image_path": "e087309f22d275658538b2441862eef16281351cb4b9236dcfabfa470f74ad6e.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 467, + 294, + 533, + 360 + ], + "blocks": [ + { + "bbox": [ + 467, + 294, + 533, + 360 + ], + "lines": [ + { + "bbox": [ + 467, + 294, + 533, + 360 + ], + "spans": [ + { + "bbox": [ + 467, + 294, + 533, + 360 + ], + "type": "image", + "image_path": "9c709fbdc18013f5515b605389e83cce378de6df6c586faba909cdd37cfa5716.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_body" + } + ], + "index": 27 + }, + { + "type": "image", + "bbox": [ + 194, + 365, + 271, + 430 + ], + "blocks": [ + { + "bbox": [ + 194, + 365, + 271, + 430 + ], + "lines": [ + { + "bbox": [ + 194, + 365, + 271, + 430 + ], + "spans": [ + { + "bbox": [ + 194, + 365, + 271, + 430 + ], + "type": "image", + "image_path": "2f9b1fa27e890aac9551f12c31b27b2cb39693185efb876ef5b75bea74251b20.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + } + ], + "index": 28 + }, + { + "type": "image", + "bbox": [ + 272, + 365, + 336, + 430 + ], + "blocks": [ + { + "bbox": [ + 272, + 365, + 336, + 430 + ], + "lines": [ + { + "bbox": [ + 272, + 365, + 336, + 430 + ], + "spans": [ + { + "bbox": [ + 272, + 365, + 336, + 430 + ], + "type": "image", + "image_path": "ee9caca2ddadb3d500ccacefb84ba6e158538645928f9b772dc658c44a1fcf61.jpg" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_body" + } + ], + "index": 29 + }, + { + "type": "image", + "bbox": [ + 337, + 365, + 402, + 430 + ], + "blocks": [ + { + "bbox": [ + 337, + 365, + 402, + 430 + ], + "lines": [ + { + "bbox": [ + 337, + 365, + 402, + 430 + ], + "spans": [ + { + "bbox": [ + 337, + 365, + 402, + 430 + ], + "type": "image", + "image_path": "54c99cb417df9b2afbb13914625e3d4624da4f78e3978a8c4b9a72c4693c1ef4.jpg" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_body" + } + ], + "index": 30 + }, + { + "type": "image", + "bbox": [ + 403, + 365, + 467, + 430 + ], + "blocks": [ + { + "bbox": [ + 403, + 365, + 467, + 430 + ], + "lines": [ + { + "bbox": [ + 403, + 365, + 467, + 430 + ], + "spans": [ + { + "bbox": [ + 403, + 365, + 467, + 430 + ], + "type": "image", + "image_path": "b2c9a46a911a56759e786f03de43e4965fe1249e747e7780c07da6019aa5d744.jpg" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_body" + } + ], + "index": 31 + }, + { + "type": "image", + "bbox": [ + 468, + 365, + 533, + 430 + ], + "blocks": [ + { + "bbox": [ + 468, + 365, + 533, + 430 + ], + "lines": [ + { + "bbox": [ + 468, + 365, + 533, + 430 + ], + "spans": [ + { + "bbox": [ + 468, + 365, + 533, + 430 + ], + "type": "image", + "image_path": "8634dd00e81493d7777976ff6ecee3afd1a33173aaebbda5cf5004f665dcb0ff.jpg" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_body" + } + ], + "index": 32 + }, + { + "bbox": [ + 143, + 500, + 192, + 513 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 500, + 192, + 513 + ], + "spans": [ + { + "bbox": [ + 143, + 500, + 192, + 513 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 46, + 525, + 289, + 694 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 525, + 289, + 694 + ], + "spans": [ + { + "bbox": [ + 46, + 525, + 289, + 694 + ], + "type": "text", + "content": "While large language models (LLMs) excel in a simulated world of texts, they struggle to interact with the more realistic world without perceptions of other modalities such as visual or audio signals. Although vision-language models (VLMs) integrate LLM modules (1) aligned with static image features, and (2) may possess prior knowledge of world dynamics (as demonstrated in the text world), they have not been trained in an embodied visual world and thus cannot align with its dynamics. On the other hand, training an embodied agent in a noisy visual world without expert guidance is often challenging and inefficient. In this paper, we train a VLM agent living in a visual world using an LLM agent excelling in a parallel text world. Specifically, we distill LLM's reflection outcomes (improved actions by analyzing mistakes) in" + } + ] + } + ], + "index": 35 + }, + { + "bbox": [ + 304, + 502, + 547, + 633 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 502, + 547, + 633 + ], + "spans": [ + { + "bbox": [ + 304, + 502, + 547, + 633 + ], + "type": "text", + "content": "a text world's tasks to finetune the VLM on the same tasks of the visual world, resulting in an Embodied Multi-Modal Agent (EMMA) quickly adapting to the visual world dynamics. Such cross-modality imitation learning between the two parallel worlds is achieved by a novel DAgger-DPO algorithm, enabling EMMA to generalize to a broad scope of new tasks without any further guidance from the LLM expert. Extensive evaluations on the ALFWorld benchmark's diverse tasks highlight EMMA's superior performance to SOTA VLM-based agents, e.g., " + }, + { + "bbox": [ + 304, + 502, + 547, + 633 + ], + "type": "inline_equation", + "content": "20\\% -70\\%" + }, + { + "bbox": [ + 304, + 502, + 547, + 633 + ], + "type": "text", + "content": " improvement in the success rate." + } + ] + } + ], + "index": 36 + }, + { + "bbox": [ + 306, + 645, + 387, + 658 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 645, + 387, + 658 + ], + "spans": [ + { + "bbox": [ + 306, + 645, + 387, + 658 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 37 + }, + { + "bbox": [ + 304, + 665, + 548, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 548, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 548, + 715 + ], + "type": "text", + "content": "Embodied multi-modal agents have been acknowledged as a pivotal stride towards achieving Artificial General Intelligence (AGI), as they encompass the potential for a broad scope of intelligent activities [27, 63]. The rising of founda" + } + ] + } + ], + "index": 38 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 703, + 135, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 703, + 135, + 712 + ], + "spans": [ + { + "bbox": [ + 58, + 703, + 135, + 712 + ], + "type": "text", + "content": "*Corresponding author" + } + ] + } + ], + "index": 39 + }, + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "26275" + } + ] + } + ], + "index": 40 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 288, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 288, + 144 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 288, + 144 + ], + "type": "text", + "content": "tion models brings a glimmer of hope for constructing such agents [11, 32, 34, 35, 38, 71, 76], and notable efforts from the community have tried to harness them in many decision-making scenarios, e.g., autonomous driving [15, 66], daily household robots [6, 13, 77], and complex manipulation tasks [25, 29, 42, 57]." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 146, + 290, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 146, + 290, + 361 + ], + "spans": [ + { + "bbox": [ + 46, + 146, + 290, + 361 + ], + "type": "text", + "content": "LLMs can perform as reflex agents interacting with a text world via verbalized descriptions of the world states and textual actions. In addition, they exhibit great potential in planning, reflection, and reward shaping. This is mainly due to their prior knowledge and semantic abstraction of the world. However, LLM-based agents cannot be directly applied in a visual world. Although current vision-language models (VLMs) try to align LLMs with the visual modality, their pretraining only focuses on static alignment between image-text pairs so the resulting agents have not been aligned well with the dynamics of the visual world. As shown in Fig. 1 (a), even the privileged VLM, e.g., GPT-4V [43], fails to accomplish tasks in an embodied ALFWorld environment [56]. In such a zero-shot setting, GPT-4V tends to mainly rely on its language prior of these objects detected in the current step, rather than the alignment between the visual input and the environment dynamics conditioned on the task instruction." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 364, + 290, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 364, + 290, + 568 + ], + "spans": [ + { + "bbox": [ + 46, + 364, + 290, + 568 + ], + "type": "text", + "content": "In this paper, we study how to train a VLM towards an embodied agent in a visual world by aligning it with the visual world dynamics and distilling the skills of an LLM agent in a parallel text world. Our overarching goal is to build such an Embodied Multi-Modal Agent (EMMA) that can take a textual task instruction (e.g., from human users) and pixel observations of the state per step to produce a sequence of actions leading to the efficient completion of the task. This is a challenging problem due to (1) the sparsity of task reward [2-4, 14], (2) noisy visual representations, (3) the hallucination of VLMs [60], and (4) the misalignment of VLM's static representations to the visual world dynamics. While offline distillation and imitation from a powerful LLM agent in a parallel text world can potentially overcome the former two challenges [75], effectively mitigating the remaining challenges necessitates online finetuning of the VLM agent within an interactive and visual world [18, 31, 70]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 570, + 290, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 570, + 290, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 570, + 290, + 715 + ], + "type": "text", + "content": "To this end, we finetune a VLM agent by imitation learning from an LLM expert (e.g., built on ChatGPT) launched in a parallel text world on the same tasks. Specifically, in each step of EMMA interacting with the visual world, we convert its visual observation into an equivalent textual description sent to the LLM agent, which produces an action for EMMA to imitate. Such cross-modality interactive imitation learning is based on DAgger [51], which overcomes the cumulative errors and distribution shifts caused by behavior cloning (BC). As depicted in Fig. 1 (b), an InstructBLIP [11] agent finetuned by BC on 170K expert demonstrations produced by a rule-based expert in the" + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 308, + 70, + 545, + 224 + ], + "blocks": [ + { + "bbox": [ + 308, + 70, + 545, + 224 + ], + "lines": [ + { + "bbox": [ + 308, + 70, + 545, + 224 + ], + "spans": [ + { + "bbox": [ + 308, + 70, + 545, + 224 + ], + "type": "image", + "image_path": "d391294bac87c81adc36437d75c4a211507e4811d908ce2f9080248a19a5a35e.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 234, + 546, + 278 + ], + "lines": [ + { + "bbox": [ + 305, + 234, + 546, + 278 + ], + "spans": [ + { + "bbox": [ + 305, + 234, + 546, + 278 + ], + "type": "text", + "content": "Figure 2: An example of tasks generated for the two parallel worlds. A VLM agent in the visual world and an LLM agent in the text world as household robots instructed to clean an apple and then put it into the fridge. Zoom in for more details." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 299, + 546, + 442 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 299, + 546, + 442 + ], + "spans": [ + { + "bbox": [ + 304, + 299, + 546, + 442 + ], + "type": "text", + "content": "visual world still fails to take correct actions based on visual observations. We further improve the DAgger's objective to be the direct preference optimization (DPO) [48], which maximizes the preference of LLM-expert's action (positive) over VLM-student's action (negative) in each interaction step. To collect better teaching signals retrospective to the VLM student's actions, the LLM expert is composed of an LLM actor prompted to output expert actions, and an LLM critic prompted for reflection feedback on the VLM agent's historical trajectories. We maintain a long-term memory storing the feedback, which is then used to induce the LLM actor to improve actions for imitation in future episodes." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 443, + 547, + 575 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 443, + 547, + 575 + ], + "spans": [ + { + "bbox": [ + 304, + 443, + 547, + 575 + ], + "type": "text", + "content": "We evaluate EMMA and compare it with vision-only agents, LLM agents, and VLM agents deployed to the ALF-World benchmark [56], which includes numerous tasks in both visual and textual environments. Extensive evaluations highlight that EMMA substantially outperforms state-of-the-art (SOTA) VLM agents in visual-only environments by " + }, + { + "bbox": [ + 304, + 443, + 547, + 575 + ], + "type": "inline_equation", + "content": "20\\% - 70\\%" + }, + { + "bbox": [ + 304, + 443, + 547, + 575 + ], + "type": "text", + "content": " in terms of success rate. In addition, EMMA is the only VLM agent that can generalize to open-vocabulary and free-form test tasks, shedding novel insights on using LLM feedback to train more versatile and generalizable embodied agents in multi-modality environments." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 585, + 477, + 600 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 585, + 477, + 600 + ], + "spans": [ + { + "bbox": [ + 306, + 585, + 477, + 600 + ], + "type": "text", + "content": "2. Embodied Multi-Modal Agent" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 605, + 548, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 548, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 548, + 715 + ], + "type": "text", + "content": "Fig. 3 illustrates the main idea of our \"Embodied MultiModal Agent (EMMA)\", whose detailed training procedures are given in Alg. 1. The agent is built upon a modularized VLM, which can follow instructions and interact with the environment through pixel observations and textual actions. To overcome these challenges associated with training EMMA, such as sparse reward or distribution shift, we explore the construction of an LLM expert from a parallel text world (Sec. 2.2) for providing EMMA with step-by-step guidance." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "26276" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 61, + 74, + 534, + 247 + ], + "blocks": [ + { + "bbox": [ + 61, + 74, + 534, + 247 + ], + "lines": [ + { + "bbox": [ + 61, + 74, + 534, + 247 + ], + "spans": [ + { + "bbox": [ + 61, + 74, + 534, + 247 + ], + "type": "image", + "image_path": "b51fafabd31cf2d615efc5a6fe5a3c8a11df5156f42a4bdb0de996fe16c19afe.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 260, + 547, + 297 + ], + "lines": [ + { + "bbox": [ + 46, + 260, + 547, + 297 + ], + "spans": [ + { + "bbox": [ + 46, + 260, + 547, + 297 + ], + "type": "text", + "content": "Figure 3: Embodied Multi-Modal Agent (EMMA) trained by an LLM expert via cross-modality imitation learning. EMMA takes a textual task instruction and pixel observations as its input state per step to generate a sequence of actions using a VLM. Then, we convert each pixel observation into a textual equivalent as the context of an LLM expert to produce improved actions for EMMA to imitate." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 315, + 288, + 340 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 315, + 288, + 340 + ], + "spans": [ + { + "bbox": [ + 46, + 315, + 288, + 340 + ], + "type": "text", + "content": "Lastly, Sec. 2.3 further discusses how to harness the LLM expert to train EMMA via cross-modality imitation learning." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 348, + 200, + 361 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 348, + 200, + 361 + ], + "spans": [ + { + "bbox": [ + 47, + 348, + 200, + 361 + ], + "type": "text", + "content": "2.1. EMMA in the Visual World" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 369, + 288, + 643 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 369, + 288, + 643 + ], + "spans": [ + { + "bbox": [ + 46, + 369, + 288, + 643 + ], + "type": "text", + "content": "In visual environments, EMMA " + }, + { + "bbox": [ + 46, + 369, + 288, + 643 + ], + "type": "inline_equation", + "content": "\\pi_{\\theta}" + }, + { + "bbox": [ + 46, + 369, + 288, + 643 + ], + "type": "text", + "content": " is designed to process a task instruction " + }, + { + "bbox": [ + 46, + 369, + 288, + 643 + ], + "type": "inline_equation", + "content": "x_{\\mathrm{task}}" + }, + { + "bbox": [ + 46, + 369, + 288, + 643 + ], + "type": "text", + "content": " (e.g., provided by human users) and the pixel observation " + }, + { + "bbox": [ + 46, + 369, + 288, + 643 + ], + "type": "inline_equation", + "content": "s_v^t" + }, + { + "bbox": [ + 46, + 369, + 288, + 643 + ], + "type": "text", + "content": " at each interaction step " + }, + { + "bbox": [ + 46, + 369, + 288, + 643 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 369, + 288, + 643 + ], + "type": "text", + "content": ". It is expected to generate a sequence of high-level textual actions " + }, + { + "bbox": [ + 46, + 369, + 288, + 643 + ], + "type": "inline_equation", + "content": "\\{x_a^t \\sim \\pi_\\theta(\\cdot | x_{\\mathrm{task}}, s_v^t)\\}_{t=0}^T" + }, + { + "bbox": [ + 46, + 369, + 288, + 643 + ], + "type": "text", + "content": " towards efficient completion of the task. To achieve this, we draw inspiration from recent advances of large pretrained VLMs [11, 32, 34, 38, 76], and modularize EMMA's architecture into four components: (1) a ViT to encode " + }, + { + "bbox": [ + 46, + 369, + 288, + 643 + ], + "type": "inline_equation", + "content": "s_v" + }, + { + "bbox": [ + 46, + 369, + 288, + 643 + ], + "type": "text", + "content": " into visual embeddings, (2) a querying transformer (Q-Former) tailored to extract the most relevant visual features via the cross-attention between the visual embeddings and query tokens, (3) a linear projection layer to align visual features to text embeddings, (4) an LLM decoder taking the concatenation of the instruction tokens and the output of the linear projection layer to autoregressively generate the action " + }, + { + "bbox": [ + 46, + 369, + 288, + 643 + ], + "type": "inline_equation", + "content": "x_a" + }, + { + "bbox": [ + 46, + 369, + 288, + 643 + ], + "type": "text", + "content": ". In order to reduce computational overhead and prevent catastrophic forgetting [37], we adopt the pretrained ViT, Q-Former, and LLM from InstructBLIP [11] and keep them frozen at the finetuning stage. We only update the linear projection layer, as illustrated in Fig. 3. Such a modularized architecture enables EMMA to integrate any existing pretrained vision models and LLMs in a flexible and computationally-efficient way." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 643, + 288, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 643, + 288, + 693 + ], + "spans": [ + { + "bbox": [ + 46, + 643, + 288, + 693 + ], + "type": "text", + "content": "However, deploying EMMA into a complex visual world remains challenging. One of the main obstacles is that the direct use of any pretrained VLM is suboptimal because existing pretraining only focuses on static alignment between" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 315, + 547, + 553 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 315, + 547, + 553 + ], + "spans": [ + { + "bbox": [ + 304, + 315, + 547, + 553 + ], + "type": "text", + "content": "image-text pairs [11, 32, 34], so that the resulting agent may struggle to reason about the dynamics of the world. As discussed in Sec. 1, even the SOTA VLM, i.e., GPT-4V, fails to accomplish tasks in an embodied ALFWorld environment [56]. In such a zero-shot setting, GPT-4V tends to rely on the linguistic prior of the currently detected objects, rather than the given task instruction and the underlying dynamics of the environment, to guide the interactions. Moreover, finetuning a pretrained VLM on a pre-collected demonstration dataset is also suboptimal due to the diversity of environments and tasks [77], the lack of large-scale expert annotations [44] as well as the challenges posed by the distribution shift issue [31]. A seemingly natural solution to the above challenges is reinforcement learning from environmental feedback (RLEF) [68], in which reward signals rely on decomposing a task into a sequence of reasonable sub-goals and checking their completion. However, in real-world scenarios, most sub-goals cannot be defined or described precisely, so the reward is sparse; hence, we do not expect RLEF to be effective." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 554, + 547, + 650 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 554, + 547, + 650 + ], + "spans": [ + { + "bbox": [ + 304, + 554, + 547, + 650 + ], + "type": "text", + "content": "To this end, we propose to leverage interactive imitation learning (IL) to align EMMA with the dynamics of any environment, which however results in two critical algorithmic challenges: (1) How to obtain a high-quality, accessible, and scalable expert that EMMA can query during IL (Sec. 2.2)? (2) Designing an effective strategy to train EMMA using this expert in complex, diverse, and potentially open-ended environments (Sec. 2.3)." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 658, + 515, + 671 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 658, + 515, + 671 + ], + "spans": [ + { + "bbox": [ + 306, + 658, + 515, + 671 + ], + "type": "text", + "content": "2.2. LLM Expert from a Parallel TextWorld" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 677, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 547, + 715 + ], + "type": "text", + "content": "Thanks to a series of prompting techniques in the realm of in-context learning, such as chain-of-thought [62], tree-of-thought [72], ReAct [73], and Reflexion [54], pretrained" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 58, + 702, + 223, + 714 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 702, + 223, + 714 + ], + "spans": [ + { + "bbox": [ + 58, + 702, + 223, + 714 + ], + "type": "text", + "content": "For brevity, we omit " + }, + { + "bbox": [ + 58, + 702, + 223, + 714 + ], + "type": "inline_equation", + "content": "x_{\\mathrm{task}}" + }, + { + "bbox": [ + 58, + 702, + 223, + 714 + ], + "type": "text", + "content": " in the rest of this paper." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "type": "text", + "content": "26277" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "code", + "bbox": [ + 50, + 88, + 287, + 312 + ], + "blocks": [ + { + "bbox": [ + 48, + 75, + 276, + 87 + ], + "lines": [ + { + "bbox": [ + 48, + 75, + 276, + 87 + ], + "spans": [ + { + "bbox": [ + 48, + 75, + 276, + 87 + ], + "type": "text", + "content": "Algorithm 1 DAgger-DPO with a single task instruction" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "code_caption" + }, + { + "bbox": [ + 50, + 88, + 287, + 312 + ], + "lines": [ + { + "bbox": [ + 50, + 88, + 287, + 312 + ], + "spans": [ + { + "bbox": [ + 50, + 88, + 287, + 312 + ], + "type": "text", + "content": "1: initialize: " + }, + { + "bbox": [ + 50, + 88, + 287, + 312 + ], + "type": "inline_equation", + "content": "i = 0" + }, + { + "bbox": [ + 50, + 88, + 287, + 312 + ], + "type": "inline_equation", + "content": "\\mathcal{D}\\gets \\emptyset" + }, + { + "bbox": [ + 50, + 88, + 287, + 312 + ], + "type": "text", + "content": " , EMMA " + }, + { + "bbox": [ + 50, + 88, + 287, + 312 + ], + "type": "inline_equation", + "content": "\\pi_{\\theta}" + }, + { + "bbox": [ + 50, + 88, + 287, + 312 + ], + "type": "text", + "content": " , LLM actor " + }, + { + "bbox": [ + 50, + 88, + 287, + 312 + ], + "type": "inline_equation", + "content": "M_a" + }, + { + "bbox": [ + 50, + 88, + 287, + 312 + ], + "type": "text", + "content": " with a FIFO memory pool " + }, + { + "bbox": [ + 50, + 88, + 287, + 312 + ], + "type": "inline_equation", + "content": "\\mathcal{P}\\gets \\emptyset" + }, + { + "bbox": [ + 50, + 88, + 287, + 312 + ], + "type": "text", + "content": " , LLM critic " + }, + { + "bbox": [ + 50, + 88, + 287, + 312 + ], + "type": "inline_equation", + "content": "M_c" + }, + { + "bbox": [ + 50, + 88, + 287, + 312 + ], + "type": "text", + "content": " \n2: input: max trials " + }, + { + "bbox": [ + 50, + 88, + 287, + 312 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 50, + 88, + 287, + 312 + ], + "type": "text", + "content": " , training epochs " + }, + { + "bbox": [ + 50, + 88, + 287, + 312 + ], + "type": "inline_equation", + "content": "I_{e}" + }, + { + "bbox": [ + 50, + 88, + 287, + 312 + ], + "type": "text", + "content": " , visual and textual envs " + }, + { + "bbox": [ + 50, + 88, + 287, + 312 + ], + "type": "inline_equation", + "content": "E_{v},E_{l}" + }, + { + "bbox": [ + 50, + 88, + 287, + 312 + ], + "type": "text", + "content": " , instruction " + }, + { + "bbox": [ + 50, + 88, + 287, + 312 + ], + "type": "inline_equation", + "content": "x_{\\mathrm{task}}" + }, + { + "bbox": [ + 50, + 88, + 287, + 312 + ], + "type": "text", + "content": " , reference agent " + }, + { + "bbox": [ + 50, + 88, + 287, + 312 + ], + "type": "inline_equation", + "content": "\\pi_{\\mathrm{ref}}" + }, + { + "bbox": [ + 50, + 88, + 287, + 312 + ], + "type": "text", + "content": " \n3: Initialize " + }, + { + "bbox": [ + 50, + 88, + 287, + 312 + ], + "type": "inline_equation", + "content": "\\pi_{\\theta}" + }, + { + "bbox": [ + 50, + 88, + 287, + 312 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 50, + 88, + 287, + 312 + ], + "type": "inline_equation", + "content": "\\pi_{\\mathrm{ref}}" + }, + { + "bbox": [ + 50, + 88, + 287, + 312 + ], + "type": "text", + "content": " Behavior Cloning Initialization \n4: while task not completed or " + }, + { + "bbox": [ + 50, + 88, + 287, + 312 + ], + "type": "inline_equation", + "content": "i < I" + }, + { + "bbox": [ + 50, + 88, + 287, + 312 + ], + "type": "text", + "content": " do \n5: Get " + }, + { + "bbox": [ + 50, + 88, + 287, + 312 + ], + "type": "inline_equation", + "content": "\\tau_v^i = [x_{\\mathrm{task}},s_v^0,x_a^0,\\dots,s_v^T,x_a^T]" + }, + { + "bbox": [ + 50, + 88, + 287, + 312 + ], + "type": "text", + "content": " via " + }, + { + "bbox": [ + 50, + 88, + 287, + 312 + ], + "type": "inline_equation", + "content": "E_{v}" + }, + { + "bbox": [ + 50, + 88, + 287, + 312 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 50, + 88, + 287, + 312 + ], + "type": "inline_equation", + "content": "\\pi_{\\theta}" + }, + { + "bbox": [ + 50, + 88, + 287, + 312 + ], + "type": "text", + "content": " \n6: Get " + }, + { + "bbox": [ + 50, + 88, + 287, + 312 + ], + "type": "inline_equation", + "content": "\\tau_l^i = [x_{\\mathrm{task}},s_l^0,x_a^0,\\dots,s_l^T,x_a^T]" + }, + { + "bbox": [ + 50, + 88, + 287, + 312 + ], + "type": "text", + "content": " via " + }, + { + "bbox": [ + 50, + 88, + 287, + 312 + ], + "type": "inline_equation", + "content": "E_{l}" + }, + { + "bbox": [ + 50, + 88, + 287, + 312 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 50, + 88, + 287, + 312 + ], + "type": "inline_equation", + "content": "\\tau_v^i" + }, + { + "bbox": [ + 50, + 88, + 287, + 312 + ], + "type": "text", + "content": " \n7: Generate retrospective feedback " + }, + { + "bbox": [ + 50, + 88, + 287, + 312 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_i = M_c(\\tau_l^i)" + }, + { + "bbox": [ + 50, + 88, + 287, + 312 + ], + "type": "text", + "content": " \n8: Update " + }, + { + "bbox": [ + 50, + 88, + 287, + 312 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 50, + 88, + 287, + 312 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 50, + 88, + 287, + 312 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_i" + }, + { + "bbox": [ + 50, + 88, + 287, + 312 + ], + "type": "text", + "content": " (i.e., " + }, + { + "bbox": [ + 50, + 88, + 287, + 312 + ], + "type": "inline_equation", + "content": "\\mathcal{P}\\leftarrow \\mathcal{P}\\cup \\mathcal{P}_i" + }, + { + "bbox": [ + 50, + 88, + 287, + 312 + ], + "type": "text", + "content": " \n9: for " + }, + { + "bbox": [ + 50, + 88, + 287, + 312 + ], + "type": "inline_equation", + "content": "t = 0" + }, + { + "bbox": [ + 50, + 88, + 287, + 312 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 50, + 88, + 287, + 312 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 50, + 88, + 287, + 312 + ], + "type": "text", + "content": " do Dataset Aggregation \n10: " + }, + { + "bbox": [ + 50, + 88, + 287, + 312 + ], + "type": "inline_equation", + "content": "x_{a}^{*} = M_{a}(\\mathcal{P},x_{\\mathrm{task}},\\ldots ,x_{a}^{t - 1},s_{l}^{t})" + }, + { + "bbox": [ + 50, + 88, + 287, + 312 + ], + "type": "text", + "content": " \n11: " + }, + { + "bbox": [ + 50, + 88, + 287, + 312 + ], + "type": "inline_equation", + "content": "\\mathcal{D}\\gets \\mathcal{D}\\cup \\{(x_{\\mathrm{task}},s_v^t,x_a^t,x_a^*)\\}" + }, + { + "bbox": [ + 50, + 88, + 287, + 312 + ], + "type": "text", + "content": " \n12: for " + }, + { + "bbox": [ + 50, + 88, + 287, + 312 + ], + "type": "inline_equation", + "content": "j = 0" + }, + { + "bbox": [ + 50, + 88, + 287, + 312 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 50, + 88, + 287, + 312 + ], + "type": "inline_equation", + "content": "I_{e} - 1" + }, + { + "bbox": [ + 50, + 88, + 287, + 312 + ], + "type": "text", + "content": " do Gradient Descent on " + }, + { + "bbox": [ + 50, + 88, + 287, + 312 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 50, + 88, + 287, + 312 + ], + "type": "text", + "content": " \n13: Sample a mini-batch " + }, + { + "bbox": [ + 50, + 88, + 287, + 312 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 50, + 88, + 287, + 312 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 50, + 88, + 287, + 312 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 50, + 88, + 287, + 312 + ], + "type": "text", + "content": " \n14: Update " + }, + { + "bbox": [ + 50, + 88, + 287, + 312 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 50, + 88, + 287, + 312 + ], + "type": "text", + "content": " by minimizing Eq. (2) with " + }, + { + "bbox": [ + 50, + 88, + 287, + 312 + ], + "type": "inline_equation", + "content": "\\pi_{\\mathrm{ref}}" + }, + { + "bbox": [ + 50, + 88, + 287, + 312 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 50, + 88, + 287, + 312 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 50, + 88, + 287, + 312 + ], + "type": "text", + "content": " \n15: " + }, + { + "bbox": [ + 50, + 88, + 287, + 312 + ], + "type": "inline_equation", + "content": "i\\gets i + 1" + }, + { + "bbox": [ + 50, + 88, + 287, + 312 + ], + "type": "text", + "content": " \n16: output: " + }, + { + "bbox": [ + 50, + 88, + 287, + 312 + ], + "type": "inline_equation", + "content": "\\pi_{\\theta^{*}}" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "code_body" + } + ], + "index": 1, + "sub_type": "algorithm" + }, + { + "bbox": [ + 46, + 335, + 289, + 563 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 335, + 289, + 563 + ], + "spans": [ + { + "bbox": [ + 46, + 335, + 289, + 563 + ], + "type": "text", + "content": "LLMs have demonstrated impressive zero-shot performance across many decision-making scenarios [6, 13, 15, 25, 29, 42, 57, 66, 77]. Despite the great potential in serving as high-quality and scalable experts, they are only able to interact with the environment via textual descriptions of the states, rather than using raw pixel observations like EMMA. To bridge this gap, we convert each pixel observation " + }, + { + "bbox": [ + 46, + 335, + 289, + 563 + ], + "type": "inline_equation", + "content": "s_v" + }, + { + "bbox": [ + 46, + 335, + 289, + 563 + ], + "type": "text", + "content": " into a textual equivalent by extracting its metadata from the simulator [30], which is composed of attributes such as Observed Objects, Observed Relations, Inventory, and Locations. We then employ the Planning Domain Definition Language (PDDL) [1] to describe this metadata and create an equivalent textual description/state " + }, + { + "bbox": [ + 46, + 335, + 289, + 563 + ], + "type": "inline_equation", + "content": "s_l" + }, + { + "bbox": [ + 46, + 335, + 289, + 563 + ], + "type": "text", + "content": " using the TextWorld engine [10]. Additional details are available in Appendix 7 and an example of this process is illustrated in Fig. 2. This methodology enables the utilization of any pretrained LLM agent that generates a sequence of actions facilitating the training of EMMA through cross-modality IL between the two agents." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 571, + 287, + 583 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 571, + 287, + 583 + ], + "spans": [ + { + "bbox": [ + 47, + 571, + 287, + 583 + ], + "type": "text", + "content": "2.3. Training EMMA via Cross-Modality Imitation" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 590, + 287, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 590, + 287, + 639 + ], + "spans": [ + { + "bbox": [ + 46, + 590, + 287, + 639 + ], + "type": "text", + "content": "Given an LLM expert from the parallel TextWorld, we aim to train a VLM agent " + }, + { + "bbox": [ + 46, + 590, + 287, + 639 + ], + "type": "inline_equation", + "content": "\\pi_{\\theta}" + }, + { + "bbox": [ + 46, + 590, + 287, + 639 + ], + "type": "text", + "content": " in the visual world to imitate its behaviors closely. This is equal to minimizing the following objective under the distribution of states induced by " + }, + { + "bbox": [ + 46, + 590, + 287, + 639 + ], + "type": "inline_equation", + "content": "\\pi_{\\theta}" + }, + { + "bbox": [ + 46, + 590, + 287, + 639 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 82, + 648, + 287, + 667 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 82, + 648, + 287, + 667 + ], + "spans": [ + { + "bbox": [ + 82, + 648, + 287, + 667 + ], + "type": "interline_equation", + "content": "\\theta^ {*} = \\underset {\\theta \\in \\Theta} {\\arg \\min } \\mathbb {E} _ {\\pi_ {\\theta}} \\left[ \\mathcal {L} _ {\\mathrm {i m i t}} \\left(\\pi_ {\\theta} \\left(x _ {a} \\mid s _ {v}\\right), x _ {a} ^ {*}\\right)\\right) ], \\tag {1}", + "image_path": "4d5a5ec3d2d631dd97f02c1254f6a1f468c028b47d4a7b100029ef2cd1d084f0.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 677, + 289, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 289, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 289, + 714 + ], + "type": "text", + "content": "in which the choice of loss function " + }, + { + "bbox": [ + 46, + 677, + 289, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{imit}}" + }, + { + "bbox": [ + 46, + 677, + 289, + 714 + ], + "type": "text", + "content": " is dependent on specific scenarios. For instance, it may be the expected cross-entropy loss for the discrete action space, or the expected" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 72, + 547, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 133 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 133 + ], + "type": "text", + "content": "MSE loss for the continuous action space. In our case, we select DPO [48] loss due to its proven superior performance to the cross-entropy on aligning models with expert preferences within the discrete language space. Hence, Eq. (1) can be extended to the formulation below." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 309, + 141, + 546, + 161 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 309, + 141, + 546, + 161 + ], + "spans": [ + { + "bbox": [ + 309, + 141, + 546, + 161 + ], + "type": "interline_equation", + "content": "\\theta^ {*} = \\underset {\\theta \\in \\Theta} {\\arg \\min } - \\mathbb {E} _ {\\pi_ {\\theta}} \\left[ \\mathcal {L} _ {\\operatorname {i m i t}} \\left(\\pi_ {\\theta}, \\pi_ {\\text {r e f}}, s _ {v}, x _ {a}, x _ {a} ^ {*}\\right) \\right], \\tag {2}", + "image_path": "5fcd5ca277aae1df8dcb704e3b1c3e067544e89ead003f22b0a3dee2db7b1b98.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 310, + 163, + 541, + 191 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 310, + 163, + 541, + 191 + ], + "spans": [ + { + "bbox": [ + 310, + 163, + 541, + 191 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {i m i t}} (\\cdot) \\triangleq \\log \\sigma \\left(\\beta \\log \\frac {\\pi_ {\\theta} \\left(x _ {a} ^ {*} \\mid s _ {v}\\right)}{\\pi_ {\\mathrm {r e f}} \\left(x _ {a} ^ {*} \\mid s _ {v}\\right)} - \\beta \\log \\frac {\\pi_ {\\theta} \\left(x _ {a} \\mid s _ {v}\\right)}{\\pi_ {\\mathrm {r e f}} \\left(x _ {a} \\mid s _ {v}\\right)}\\right)", + "image_path": "ef5d448421fb8d56d1f9b309960cb3942eca7982436139595ecb2aa8458cbf5d.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 198, + 547, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 198, + 547, + 437 + ], + "spans": [ + { + "bbox": [ + 304, + 198, + 547, + 437 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 198, + 547, + 437 + ], + "type": "inline_equation", + "content": "x_{a}^{*}" + }, + { + "bbox": [ + 304, + 198, + 547, + 437 + ], + "type": "text", + "content": " is the action given by an expert while " + }, + { + "bbox": [ + 304, + 198, + 547, + 437 + ], + "type": "inline_equation", + "content": "x_{a}" + }, + { + "bbox": [ + 304, + 198, + 547, + 437 + ], + "type": "text", + "content": " is the action given by " + }, + { + "bbox": [ + 304, + 198, + 547, + 437 + ], + "type": "inline_equation", + "content": "\\pi_{\\theta}" + }, + { + "bbox": [ + 304, + 198, + 547, + 437 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 198, + 547, + 437 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 304, + 198, + 547, + 437 + ], + "type": "text", + "content": " is the logistic function, and " + }, + { + "bbox": [ + 304, + 198, + 547, + 437 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 304, + 198, + 547, + 437 + ], + "type": "text", + "content": " is a hyperparameter controlling the deviation from " + }, + { + "bbox": [ + 304, + 198, + 547, + 437 + ], + "type": "inline_equation", + "content": "\\pi_{\\mathrm{ref}}" + }, + { + "bbox": [ + 304, + 198, + 547, + 437 + ], + "type": "text", + "content": ", i.e. the reference agent obtained by behavior cloning on a demonstration dataset produced by a rule-based expert in the visual world (see Appendix 9 for complete details). This regularization is essential, as it prevents the agent from deviating too far from the distribution on which the expert is accurate, as well as maintaining the generation diversity and avoiding premature convergence to some easy tasks [48]. In practice, the VLM agent " + }, + { + "bbox": [ + 304, + 198, + 547, + 437 + ], + "type": "inline_equation", + "content": "\\pi_{\\theta}" + }, + { + "bbox": [ + 304, + 198, + 547, + 437 + ], + "type": "text", + "content": " is also initialized to " + }, + { + "bbox": [ + 304, + 198, + 547, + 437 + ], + "type": "inline_equation", + "content": "\\pi_{\\mathrm{ref}}" + }, + { + "bbox": [ + 304, + 198, + 547, + 437 + ], + "type": "text", + "content": " for stabilizing the training process. Since environmental dynamics is both unknown and complex, we cannot compute the distribution of states visited by " + }, + { + "bbox": [ + 304, + 198, + 547, + 437 + ], + "type": "inline_equation", + "content": "\\pi_{\\theta}" + }, + { + "bbox": [ + 304, + 198, + 547, + 437 + ], + "type": "text", + "content": " and can only sample it by rolling out the agent. Hence, Eq. (2) is a non-i.i.d. supervised learning problem due to the dependence of the state distribution on " + }, + { + "bbox": [ + 304, + 198, + 547, + 437 + ], + "type": "inline_equation", + "content": "\\pi_{\\theta}" + }, + { + "bbox": [ + 304, + 198, + 547, + 437 + ], + "type": "text", + "content": " itself, in which naive behavior cloning faces issues like cumulative error and distribution shift [26]. To address this, we employ an interactive IL algorithm, DAgger [51], which provably converges to the optimal agent " + }, + { + "bbox": [ + 304, + 198, + 547, + 437 + ], + "type": "inline_equation", + "content": "\\pi_{\\theta^{*}}" + }, + { + "bbox": [ + 304, + 198, + 547, + 437 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 437, + 547, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 437, + 547, + 628 + ], + "spans": [ + { + "bbox": [ + 304, + 437, + 547, + 628 + ], + "type": "text", + "content": "As discussed in Sec. 2.2 and Sec. 2.3, we can harness an LLM expert to generate a sequence of actions, serving as " + }, + { + "bbox": [ + 304, + 437, + 547, + 628 + ], + "type": "inline_equation", + "content": "x_{a}^{*}" + }, + { + "bbox": [ + 304, + 437, + 547, + 628 + ], + "type": "text", + "content": " in Eq. (2). However, these actions may be suboptimal due to sparse environmental feedback [54, 68] or defective in-context instructions [9]. To collect better teaching signals, we introduce a \"retrospective LLM expert\" that is composed of two specialized models: an actor " + }, + { + "bbox": [ + 304, + 437, + 547, + 628 + ], + "type": "inline_equation", + "content": "(M_{a})" + }, + { + "bbox": [ + 304, + 437, + 547, + 628 + ], + "type": "text", + "content": ", built upon an API LLM and prompted to generate actions based on the task instruction and state observations; and a critic " + }, + { + "bbox": [ + 304, + 437, + 547, + 628 + ], + "type": "inline_equation", + "content": "(M_{c})" + }, + { + "bbox": [ + 304, + 437, + 547, + 628 + ], + "type": "text", + "content": ", also based on the same LLM, but designed to analyze EMMA's historical interactions and provide reflective feedback. A long-term memory " + }, + { + "bbox": [ + 304, + 437, + 547, + 628 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 304, + 437, + 547, + 628 + ], + "type": "text", + "content": " is maintained to store the feedback generated by " + }, + { + "bbox": [ + 304, + 437, + 547, + 628 + ], + "type": "inline_equation", + "content": "M_{c}" + }, + { + "bbox": [ + 304, + 437, + 547, + 628 + ], + "type": "text", + "content": ", which is then used to prompt " + }, + { + "bbox": [ + 304, + 437, + 547, + 628 + ], + "type": "inline_equation", + "content": "M_{a}" + }, + { + "bbox": [ + 304, + 437, + 547, + 628 + ], + "type": "text", + "content": " for improved actions. The complete procedure is detailed in Line 7-10 of Alg. 1, and all prompts are provided in Sec. 6 of the Appendix." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 639, + 388, + 653 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 639, + 388, + 653 + ], + "spans": [ + { + "bbox": [ + 306, + 639, + 388, + 653 + ], + "type": "text", + "content": "3. Experiments" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 659, + 422, + 672 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 659, + 422, + 672 + ], + "spans": [ + { + "bbox": [ + 306, + 659, + 422, + 672 + ], + "type": "text", + "content": "3.1. Experimental Setup" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 677, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 547, + 713 + ], + "type": "text", + "content": "**Environments.** We base our experiments on the ALFWorld benchmark [56], a cross-modality simulation platform that encompasses a wide range of embodied household tasks. For" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "26278" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 70, + 216, + 199 + ], + "blocks": [ + { + "bbox": [ + 56, + 70, + 216, + 199 + ], + "lines": [ + { + "bbox": [ + 56, + 70, + 216, + 199 + ], + "spans": [ + { + "bbox": [ + 56, + 70, + 216, + 199 + ], + "type": "image", + "image_path": "da5c7f948991768971c6b06a25303425486e734e314fb500c279df7962cdf677.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 51, + 205, + 220, + 307 + ], + "lines": [ + { + "bbox": [ + 51, + 205, + 220, + 307 + ], + "spans": [ + { + "bbox": [ + 51, + 205, + 220, + 307 + ], + "type": "text", + "content": "Figure 4: Comparison of success rate between EMMA and the retrospective LLM expert. As the number of trials increases, the LLM expert progressively improves its performance through retrospective processes, and also the gap between the two agents decreases, indicating the effectiveness of cross-modality imitation. We separately plot the curve per task type in Fig. 9 of Appendix." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 227, + 70, + 541, + 198 + ], + "blocks": [ + { + "bbox": [ + 227, + 70, + 541, + 198 + ], + "lines": [ + { + "bbox": [ + 227, + 70, + 541, + 198 + ], + "spans": [ + { + "bbox": [ + 227, + 70, + 541, + 198 + ], + "type": "table", + "html": "
AgentVisual Env.Textual Env.Success Rate with Template Task Instruction
Avg.PickCleanHeatCoolLookPick2
VMsResNet-18* [56]✓X0.06(-)------
MCNN-FPN* [56]✓X0.05(-)------
LMsBUTLER* [56]X✓0.26(-)0.31(-)0.41(-)0.60(-)0.27(-)0.12(-)0.29(-)
GPT-BUTLER [41]X✓0.69(18.8)0.62(18.4)0.81(18.4)0.85(12.7)0.78(15.6)0.50(24.4)0.47(26.6)
ReAct [73]X✓0.54(20.6)0.71(18.1)0.65(18.8)0.62(18.2)0.44(23.2)0.28(23.7)0.35(25.5)
Reflexion [54]X✓0.91(18.7)0.96(17.4)1.00(17.0)0.81(19.4)0.83(21.6)0.94(16.9)0.88(21.6)
DEPS* [61]X✓0.76(-)0.93(-)0.50(-)0.80(-)1.00(-)1.00(-)0.00(-)
AutoGen* [64]X✓0.77(-)0.92(-)0.74(-)0.78(-)0.86(-)0.83(-)0.41(-)
VLMsMiniGPT-4 [76]✓X0.16(26.9)0.04(29.0)0.00(30.0)0.19(26.3)0.17(26.7)0.67(17.7)0.06(28.9)
BLIP-2 [34]✓X0.04(29.5)0.00(30.0)0.06(29.3)0.04(29.9)0.11(28.2)0.06(29.2)0.00(30.0)
LLaMA-Adapter [17]✓X0.13(27.5)0.17(26.4)0.10(28.6)0.27(24.2)0.22(26.7)0.00(30.0)0.00(30.0)
InstructBLIP [11]✓X0.22(26.2)0.50(21.5)0.26(25.0)0.23(27.2)0.06(28.9)0.17(26.8)0.00(30.0)
EMMA (Ours)✓X0.82(19.5)0.71(19.3)0.94(17.5)0.85(19.6)0.83(19.9)0.88(19.6)0.67(22.4)
Human Performance* [55]✓X0.91(-)------
", + "image_path": "9ca5b87bd0e46e319a5481efb7c4bd2a7caa6919802e4bf6c3c360919664de97.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 225, + 205, + 542, + 305 + ], + "lines": [ + { + "bbox": [ + 225, + 205, + 542, + 305 + ], + "spans": [ + { + "bbox": [ + 225, + 205, + 542, + 305 + ], + "type": "text", + "content": "Table 1: Comparison with the state of the arts. " + }, + { + "bbox": [ + 225, + 205, + 542, + 305 + ], + "type": "inline_equation", + "content": "*" + }, + { + "bbox": [ + 225, + 205, + 542, + 305 + ], + "type": "text", + "content": "-reported in previous work. VMs = vision models, LMs = language models, VLMs = vision-language models. \"Visual Env.\" and \"Textual Env.\" refer to the visual and the parallel textual environments from ALFWorld [56], respectively. " + }, + { + "bbox": [ + 225, + 205, + 542, + 305 + ], + "type": "inline_equation", + "content": "\\sqrt{} / X" + }, + { + "bbox": [ + 225, + 205, + 542, + 305 + ], + "type": "text", + "content": " denotes the corresponding environment used/not used to deploy the agent. The highest scores for each task in the same type of environment are highlighted in bold. The average interaction steps are given in the (parentheses). EMMA substantially outperforms other SOTA VLM agents in the visual environments, and its success also directs a promising way to achieve human-level performance in ALFWorld." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 326, + 289, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 326, + 289, + 613 + ], + "spans": [ + { + "bbox": [ + 46, + 326, + 289, + 613 + ], + "type": "text", + "content": "each task, ALFWorld integrates a visual environment, rendered by the Ai2Thor simulator [30], with a corresponding textual environment. Textual environments employ the Planning Domain Definition Language (PDDL) [1] to translate each pixel observation from the simulator into an equivalent text-based observation, and then construct an interactive world using the TextWorld engine [10]. Fig. 2 provides an illustrative example of the tasks featured in ALFWorld. The tasks within the benchmark are categorized into 6 types: Pick & Place, Clean & Place, Heat & Place, Cool & Place, Look in Light, and Pick Two Objects & Place. Each task requires an agent to execute a series of text-based actions, such as \"go to safe 1\", \"open safe 1\", or \"heat egg 1 with microwave 1\", following a predefined instruction. These actions involve navigating and interacting with the environment. To provide a comprehensive understanding, we have visualized an example of each task type in Fig. 8 of the Appendix. A task in this benchmark may involve interactions with over 10 objects and require more than 30 steps for a human expert to solve, thus challenging an agent's capabilities in long-horizon planning, instruction following, and the utilization of commonsense knowledge. For a fair comparison, we follow the same setting as prior work [41, 54-56, 73] and evaluate all baselines using 134 out-of-distribution (OOD) tasks." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 617, + 289, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 617, + 289, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 617, + 289, + 714 + ], + "type": "text", + "content": "Baselines. To verify the effectiveness of cross-modality imitation learning, we compare our EMMA with several baselines and SOTA agents using the ALFWorld benchmark with both visual and textual environments. The compared agents can be divided into three categories: vision models, language models, and vision-language models. Concretely, vision models, including ResNet-18 [20] and MCNN-FPN [21], utilize pretrained vision encoders to extract salient features" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 326, + 548, + 711 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 326, + 548, + 711 + ], + "spans": [ + { + "bbox": [ + 304, + 326, + 548, + 711 + ], + "type": "text", + "content": "from each pixel observation. The extracted features then serve as input for a Multi-Layer Perceptron (MLP) policy, which is trained by behavior cloning on a pre-collected demonstration dataset. Unlike vision models performing in the visual environment, language models complete the same tasks but in a parallel, text-based environment. BUTLER [56] employs a transformer seq2seq model enhanced with a pointer softmax mechanism [19]. This architecture aggregates previous observations as input to generate text-based actions token-by-token. GPT-BUTLER [41], a variant of the GPT-2 model [45], is initially pretrained on a static demonstration dataset and further finetuned using data collected online. ReAct [73] takes a novel approach by utilizing LLMs to generate reasoning traces and task-specific actions in an interleaved manner. This method aids the agent in developing, tracking, and updating its action plans interactively. Reflexion [54] similarly employs an LLM, but it focuses on reflecting upon environmental feedback. It maintains this reflective text in an episodic memory buffer, enhancing the agent's ability to improve actions in subsequent trials. Similar to Reflexion, a concurrent work, DEPS [61], also corrects errors in previous LLM-generated actions by integrating descriptions of the action execution process and providing self-explanations for the feedback. Moreover, beyond the single-agent framework, AutoGen [64] exhibits the potential to accomplish a broad spectrum of tasks through the cooperation of multiple LLM agents. Finally, we consider a range of vision-language models, such as MiniGPT-4 [76], BLIP-2 [34], LLaMA-Adaptor [17], and InstructBLIP [11], as agents to interact with the visual environment. Unlike pure vision or language models, VLMs are designed to process and integrate both visual and textual data, offering a" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "26279" + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 288, + 145 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 288, + 145 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 288, + 145 + ], + "type": "text", + "content": "more holistic understanding of the environment. To align these agents with the specific requirements of the ALFWorld benchmark, we finetune them on a pre-collected demonstration dataset. This finetuning process is crucial as it enables the agent to comprehend and adhere to ALFWorld's unique grammar and to develop a basic \"gamesense\"." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 153, + 146, + 165 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 153, + 146, + 165 + ], + "spans": [ + { + "bbox": [ + 47, + 153, + 146, + 165 + ], + "type": "text", + "content": "3.2. Training Details" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 171, + 288, + 423 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 171, + 288, + 423 + ], + "spans": [ + { + "bbox": [ + 46, + 171, + 288, + 423 + ], + "type": "text", + "content": "The architectural design of EMMA is depicted in Fig. 3. At its core, EMMA employs a Query Transformer (Q-Former) to process visual data. This Q-Former extracts features using a frozen ViT encoder. Its output consists of 32 visual tokens, which are then passed through a linear projection layer before being fed to a frozen LLM decoder. Similar to other VLM agents, EMMA is also finetuned on a pre-collected demonstration dataset, aligning its basic ability with the ALFWorld benchmark. In order to align EMMA with the dynamics of ALFWorld, we train it by imitating an LLM expert (see Alg. 1). We choose text-davinci-003 developed by OpenAI as our LLM expert because of its established capabilities in reasoning and planning [54, 61, 64, 73]. In this setup, text-davinci-003 serves dual roles: it serves as an actor, providing EMMA with expert actions, and as a critic, analyzing EMMA's historical trajectories. This analysis generates retrospective feedback, which is then incorporated into the actor's long-term memory, leading to improved actions in future trials. Further details about the hyperparameters and prompts used in our training procedure are available in Table 2 of the Appendix." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 431, + 227, + 443 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 431, + 227, + 443 + ], + "spans": [ + { + "bbox": [ + 47, + 431, + 227, + 443 + ], + "type": "text", + "content": "3.3. Comparison with State of the Art" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 450, + 288, + 461 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 450, + 288, + 461 + ], + "spans": [ + { + "bbox": [ + 46, + 450, + 288, + 461 + ], + "type": "text", + "content": "EMMA sets new SOTA performance in visual ALFWorld." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 462, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 462, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 462, + 288, + 714 + ], + "type": "text", + "content": "In this section, we compare EMMA with 12 other representative agents on the ALFWorld benchmark, spanning both visual and textual environments (refer to Table 1). We assess two key metrics: the success rate, which is the percentage of trials completed successfully, and the average number of interaction steps required for task completion, with a lower number indicating higher efficiency. EMMA demonstrates superior performance in both metrics, significantly outperforming all VLM agents in visual environments. This achievement underscores the effectiveness of our cross-modality imitation learning approach, as depicted in the learning curve shown in Fig. 4. Furthermore, EMMA's performance markedly exceeds that of VM agents, highlighting the crucial role of the prior knowledge embedded in VLMs. Intriguingly, EMMA's performance is comparable with LLM agents that operate using perfectly semantic descriptions of visual observations. This is largely attributed to EMMA's training strategy of imitating an expert LLM agent, proving to be more efficient than learning from scratch in a purely visual setting. As a result, EMMA stands out as the only VLM agent that substantially surpasses SOTA LLM agents, such" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 317, + 69, + 535, + 243 + ], + "blocks": [ + { + "bbox": [ + 317, + 69, + 535, + 243 + ], + "lines": [ + { + "bbox": [ + 317, + 69, + 535, + 243 + ], + "spans": [ + { + "bbox": [ + 317, + 69, + 535, + 243 + ], + "type": "image", + "image_path": "661817e8d2e11ec4bf36b68541811cd20f5933ab66c201456828618cbbf7bb2f.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 251, + 547, + 307 + ], + "lines": [ + { + "bbox": [ + 305, + 251, + 547, + 307 + ], + "spans": [ + { + "bbox": [ + 305, + 251, + 547, + 307 + ], + "type": "text", + "content": "Figure 5: Comparison of robustness between VLM and LLM agents. Given a specific noise rate, we crop a random portion of the pixel observation and resize it to a given size as the new observation. Similarly, we randomly replace some tokens in the textual observation with arbitrary ones." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 327, + 547, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 327, + 547, + 651 + ], + "spans": [ + { + "bbox": [ + 304, + 327, + 547, + 651 + ], + "type": "text", + "content": "as AutoGen [64] and DEPS [61], in these environments. And its success also directs a potential way to achieve human-level performance in the visual environments of ALFWorld. EMMA is more robust to noisy observations than LLM agents. While LLM agents exhibit a higher success rate with fewer interaction steps in textual environments, as indicated in Table 1, we hypothesize that this superior performance largely relies on their precise semantic abstraction of the environment. However, such an abstraction might not be feasible in real-world applications. To verify this assumption, we set up a more practical scenario where observations are deliberately perturbed at a specific noise rate. We then compare the robustness of EMMA and a SOTA LLM agent, Reflexion, under these noisy observations. To generate noisy observations, a random portion of the visual observation is cropped, resized, and then used to replace the original observation. Similarly, in the textual observation, random tokens are substituted with arbitrary ones. As illustrated in Fig. 5, with the noise rate increasing, EMMA's performance remains significantly more robust than Reflexion. This could be attributed to the vision encoder in the VLM, which is adept at filtering out visual noises. On the other hand, textual noises are directly processed by the LLM, which can substantially impair the performance of LLM-based agents. This finding highlights the potential advantages of VLM agents like EMMA in visual scenarios, in which data is often imperfect and noisy." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 658, + 400, + 671 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 658, + 400, + 671 + ], + "spans": [ + { + "bbox": [ + 306, + 658, + 400, + 671 + ], + "type": "text", + "content": "3.4. Ablation Study" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 677, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 677, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 305, + 677, + 547, + 715 + ], + "type": "text", + "content": "Retrospection improves EMMA over time. To assess the importance of the retrospective LLM expert, we present the average success rate of EMMA after each trial and compare it" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "26280" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 51, + 69, + 214, + 192 + ], + "blocks": [ + { + "bbox": [ + 51, + 69, + 214, + 192 + ], + "lines": [ + { + "bbox": [ + 51, + 69, + 214, + 192 + ], + "spans": [ + { + "bbox": [ + 51, + 69, + 214, + 192 + ], + "type": "image", + "image_path": "905e47f392e5add8df7e5b62799313afdd903e44e94ef900117cf7d6cc0b2ebf.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 201, + 547, + 245 + ], + "lines": [ + { + "bbox": [ + 46, + 201, + 547, + 245 + ], + "spans": [ + { + "bbox": [ + 46, + 201, + 547, + 245 + ], + "type": "text", + "content": "Figure 6: Ablation study. left: The ability of LLM expert to reflect on and learn from past actions appears to be a key factor in achieving the impressive results we observed. middle: The \"w/o BC initialization\" variant uses a pretrained VLM instead of the reference agent " + }, + { + "bbox": [ + 46, + 201, + 547, + 245 + ], + "type": "inline_equation", + "content": "\\pi_{\\mathrm{ref}}" + }, + { + "bbox": [ + 46, + 201, + 547, + 245 + ], + "type": "text", + "content": " to initialize EMMA. right: The 'w/ CE Loss' variant replaces the DPO loss (described in Eq. 2) with a token-level cross-entropy (CE) loss. Please refer to Sec. 3.4 for a thorough discussion." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 215, + 69, + 376, + 192 + ], + "blocks": [ + { + "bbox": [ + 215, + 69, + 376, + 192 + ], + "lines": [ + { + "bbox": [ + 215, + 69, + 376, + 192 + ], + "spans": [ + { + "bbox": [ + 215, + 69, + 376, + 192 + ], + "type": "image", + "image_path": "338e765a491bd9953ea2d10a8a5b075ecbc0ea7e8758ec8632f97b4e55254ff1.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 379, + 70, + 541, + 192 + ], + "blocks": [ + { + "bbox": [ + 379, + 70, + 541, + 192 + ], + "lines": [ + { + "bbox": [ + 379, + 70, + 541, + 192 + ], + "spans": [ + { + "bbox": [ + 379, + 70, + 541, + 192 + ], + "type": "image", + "image_path": "764c580ded6530e9ca9a4cb531669fb436a44f13801208a5de2f71d579b0407c.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 266, + 288, + 386 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 266, + 288, + 386 + ], + "spans": [ + { + "bbox": [ + 46, + 266, + 288, + 386 + ], + "type": "text", + "content": "with a key variation: \"EMMA w/o Retrospection\", as shown in Fig. 6 (left). This variant of EMMA is trained using the same procedure as the original EMMA but removes the retrospective process. Instead, it relies solely on a plain LLM actor to provide relabeled actions. The results show that EMMA with the retrospective mechanism significantly outperforms its counterpart. This finding is crucial as it indicates that the retrospective process is not just a supplementary feature but a fundamental component of EMMA's architecture that contributes substantially to its enhanced performance." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 388, + 288, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 388, + 288, + 567 + ], + "spans": [ + { + "bbox": [ + 46, + 388, + 288, + 567 + ], + "type": "text", + "content": "EMMA benefits from BC initialization. Through an ablation study, we evaluate the impact of behavior cloning (BC) initialization, a process described in line 3 of Alg. 1. Fig. 6 (middle) demonstrates that EMMA, when deprived of BC initialization, experiences a slight reduction in the average success rate across 134 unseen tasks compared to its original setup. Despite this decrease, EMMA without BC initialization still outperforms other VLM agents, as clearly shown when compared with the results in Table 1. Furthermore, Fig. 10 in the Appendix breaks down the result by task type. It reveals a consistent but slight drop in performance across 5 out of the 6 task types. These results reflect that while BC initialization contributes positively to EMMA's overall performance, it is not critical for achieving the notable results we have reported." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 570, + 290, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 570, + 290, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 570, + 290, + 714 + ], + "type": "text", + "content": "DPO enables more effective imitation learning. To evaluate the effectiveness of using DPO loss in Eq. (1), we conducted an ablation study with an alternative version of EMMA, referred to \"EMMA w/ Cross Entropy (CE) Loss\". In this variant, EMMA is optimized using the token-level CE loss, a common objective for finetuning VLMs. The results, as depicted in Fig. 6 (right), reveal that EMMA w/ CE Loss does not achieve the same high success rate as the original EMMA with DPO loss, suggesting that DPO loss contributes to enhancing the upper performance bound of EMMA. In addition, we noted that EMMA w/ CE Loss exhibits faster convergence in the initial training stages than the original" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 266, + 547, + 326 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 266, + 547, + 326 + ], + "spans": [ + { + "bbox": [ + 304, + 266, + 547, + 326 + ], + "type": "text", + "content": "EMMA. This premature convergence leads to the agent paying attention to the expert actions from easier tasks, usually addressed in the first few training epochs. This can suppress EMMA's exploration and learning on more complex tasks, resulting in worse performance." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 331, + 545, + 342 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 331, + 545, + 342 + ], + "spans": [ + { + "bbox": [ + 305, + 331, + 545, + 342 + ], + "type": "text", + "content": "3.5. Generalization to Free-form Task Instructions" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 348, + 547, + 661 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 348, + 547, + 661 + ], + "spans": [ + { + "bbox": [ + 304, + 348, + 547, + 661 + ], + "type": "text", + "content": "The ability of AI agents to precisely follow task instructions given in open-vocabulary and free-form text is crucial for real-world problems. To assess this, we conducted an additional experiment focusing on the generalization capabilities of EMMA, an agent trained with template task instructions. In this experiment, we re-evaluated the performance of EMMA and other baseline agents on 134 unseen tasks, using human-annotated instructions instead of template ones. These human-annotated instructions include a large amount of OOD verbs and objects, presenting a more realistic and challenging scenario for the agents to address. To underscore this challenge, we compared the vocabulary distribution between the template and human-annotated instructions, as shown in Fig. 7 (right). Moreover, we provide a comprehensive analysis of the vocabulary used across both instruction types in Fig. 11 of the Appendix. In Fig. 7 (left), EMMA demonstrates a slight performance decline, while a significant degradation is observed in other baselines. We also note that Reflexion, a SOTA LLM agent, exhibits exceptional generalization to those OOD instructions. According to these empirical results in Fig. 7, we have the following conclusions: (1) EMMA obtains and benefits from the generalization capabilities inherent in the SOTA LLM agent through cross-modality imitation learning; (2) Our work sheds novel insights on using LLM feedback to train more versatile and generalizable embodied agents." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 669, + 392, + 681 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 669, + 392, + 681 + ], + "spans": [ + { + "bbox": [ + 306, + 669, + 392, + 681 + ], + "type": "text", + "content": "4. Related Work" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 689, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 546, + 713 + ], + "type": "text", + "content": "Agents based on Foundation Models. Recent research has increasingly focused on harnessing the capabilities of large" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "26281" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 49, + 71, + 395, + 173 + ], + "blocks": [ + { + "bbox": [ + 49, + 71, + 395, + 173 + ], + "lines": [ + { + "bbox": [ + 49, + 71, + 395, + 173 + ], + "spans": [ + { + "bbox": [ + 49, + 71, + 395, + 173 + ], + "type": "table", + "html": "
AgentVisual Env.Textual Env.Success Rate with Free-form Task Instruction
Avg.PickCleanHeatCoolLookPick2
LMsBUTLER* [56]X✓0.03 (-)0.10 (-)0.22 (-)0.05 (-)0.17 (-)0.00 (-)0.00 (-)
GPT-BUTLER [41]X✓0.31 (24.7)0.25 (25.4)0.42 (23.5)0.46 (20.9)0.44 (21.8)0.00 (30.0)0.12 (28.9)
ReAct [73]X✓0.37 (23.6)0.46 (22.0)0.39 (23.8)0.50 (20.5)0.28 (24.6)0.22 (26.4)0.29 (25.9)
Reflexion [54]X✓0.78 (17.0)0.84 (15.8)0.74 (18.4)0.77 (17.2)0.61 (19.4)0.94 (13.8)0.82 (16.7)
VLMsMiniGPT-4 [76]✓X0.00 (30.0)0.00 (30.0)0.00 (30.0)0.00 (30.0)0.00 (30.0)0.00 (30.0)0.00 (30.0)
BLIP-2 [34]✓X0.01 (29.7)0.04 (29.1)0.03 (29.5)0.00 (30.0)0.00 (30.0)0.00 (30.0)0.00 (30.0)
LLaMA-Adapter [17]✓X0.02 (29.6)0.04 (29.3)0.03 (29.4)0.04 (29.3)0.00 (30.0)0.00 (30.0)0.00 (30.0)
InstructBLIP [11]✓X0.01 (29.8)0.00 (30.0)0.03 (29.3)0.00 (30.0)0.00 (30.0)0.00 (30.0)0.00 (30.0)
EMMA (Ours)✓X0.68 (22.0)0.72 (21.7)0.65 (22.7)0.72 (22.0)0.56 (23.9)0.67 (23.3)0.76 (18.1)
", + "image_path": "6867d7eea91c5b3cc9d4ec3847cf76ee936850c1bfe521c48aa225110338c3b4.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 405, + 70, + 545, + 175 + ], + "blocks": [ + { + "bbox": [ + 405, + 70, + 545, + 175 + ], + "lines": [ + { + "bbox": [ + 405, + 70, + 545, + 175 + ], + "spans": [ + { + "bbox": [ + 405, + 70, + 545, + 175 + ], + "type": "image", + "image_path": "429e678f6864656b282ff8d85ee02fb2cc44319c20b241abdf6e173f0fae4930.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 182, + 546, + 205 + ], + "lines": [ + { + "bbox": [ + 46, + 182, + 546, + 205 + ], + "spans": [ + { + "bbox": [ + 46, + 182, + 546, + 205 + ], + "type": "text", + "content": "Figure 7: Generalization to open-vocabulary, free-form task instructions. Left: These instructions are provided by different human annotators using Amazon Mechanical Turk [55]. Right: Frequency distribution of verbs for human-annotated and templated task instructions." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 226, + 289, + 550 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 226, + 289, + 550 + ], + "spans": [ + { + "bbox": [ + 46, + 226, + 289, + 550 + ], + "type": "text", + "content": "pre-trained foundation models to build AI agents [65, 69]. These models (e.g., LLMs), benefiting from their commonsense knowledge inherited from Internet-scale pretraining, are able to reason actions according to descriptions of the external environments. For example, given a set of task instructions, LLMs can be elaborately prompted to perform as agents generating high-level step-by-step plans [7, 24, 54, 64, 73], and each step can be parsed into a sequence of robotic actions that are executed via pretrained policies or available APIs [6, 36, 59]. Furthermore, by using VLMs [32], plans can also be conditioned on visual inputs that are transformed into language descriptions [7, 16, 25, 57] or token embeddings aligned with LLMs [13, 42, 68, 77]. However, existing foundation models are usually pretrained on static text or text-image datasets and thus may struggle to align with the dynamics of the world. To bridge this gap, we study how to finetune a VLM to be an embodied agent dynamically aligned with the world by distilling the cross-modality knowledge from an LLM expert. The work most closely related to ours is EUREKA [40], which also explores using source information provided by the simulator as the context of an LLM to aid agent training. Instead of directly mimicking the output of the LLM as we did, EUREKA harnesses the coding LLM to generate a desired reward function for a given task and optimizes a policy against the reward function using RL, leading to a more complex and expensive training procedure [47, 52, 53]." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 557, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 557, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 557, + 288, + 714 + ], + "type": "text", + "content": "Imitation Learning. Imitation learning is the study of algorithms that improve performance by mimicking an expert's decisions and behaviors. We summarize three main categories of existing methods in the following: (1) behavior cloning (BC), (2) inverse reinforcement learning (IRL), and (3) the combination of imitation and reinforcement learning. The naive BC [5] ignores the change in distribution and simply trains a policy that only performs well under the distribution of states visited by the expert. Following works, such as dataset aggregation [51] or policy aggregation [28, 49], have been proposed to address the limitations of BC. Another line of work, IRL, is a more complicated algorithm framework that learns the reward function from" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 226, + 547, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 226, + 547, + 418 + ], + "spans": [ + { + "bbox": [ + 304, + 226, + 547, + 418 + ], + "type": "text", + "content": "expert demonstrations and then improves the policy using RL with the learned reward. A representative method in this category is generative adversarial imitation learning (GAIL) [23], in which a policy and a discriminator compete with each other in order to maximize the likelihood of the policy's behavior matching the expert. The third category of methods usually leverages an IL policy to initialize RL and continues to boost its performance via online collected data from RL [8, 50, 58]. This simple combination significantly improves RL's sample efficiency and IL's upper performance bound constrained by the expert. Nevertheless, all of the above methods assume that the expert and the imitator understand the world in the same modality, and thus overlook the fact that the complementary knowledge from other modalities often boosts the model's accuracy and generalization dramatically [46, 67]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 429, + 379, + 441 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 429, + 379, + 441 + ], + "spans": [ + { + "bbox": [ + 306, + 429, + 379, + 441 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 449, + 547, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 449, + 547, + 617 + ], + "spans": [ + { + "bbox": [ + 304, + 449, + 547, + 617 + ], + "type": "text", + "content": "We create EMMA, an Embodied Multi-Modal Agent, by finetuning a VLM in an embodied visual world with interactive imitation learning from an LLM expert in a parallel text world, who produces better actions and retrospective feedback to VLM's trajectories. Such imitation learning exhibits substantial advantages over vision or VLM policies directly finetuned in the visual world or finetuned by behavior cloning of a rule-based expert, and SOTA API VLMs such as GPT-4V(ison). As a result, EMMA achieves a comparable success rate and much better robustness in the noisy visual world than its LLM teacher in the easier text world. Furthermore, EMMA shows powerful generalization to open-vocabulary and free-form task instructions, highlighting its potential in real-world scenarios." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 628, + 399, + 641 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 628, + 399, + 641 + ], + "spans": [ + { + "bbox": [ + 306, + 628, + 399, + 641 + ], + "type": "text", + "content": "Acknowledgment" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 647, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 647, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 647, + 547, + 715 + ], + "type": "text", + "content": "This work is partially supported by the National Key R&D Program of China under the Grant No. (2023YFE0106300) and (2017YFC0804002), Shenzhen Fundamental Research Program under the Grant No. (JCYJ20200109141235597), and Program for Guangdong Introducing Innovative and Entrepreneurial Teams under Grant No. (2017ZT07X386)." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "26282" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 95, + 106, + 108 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 95, + 106, + 108 + ], + "spans": [ + { + "bbox": [ + 48, + 95, + 106, + 108 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 116, + 289, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 53, + 116, + 289, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 116, + 289, + 171 + ], + "spans": [ + { + "bbox": [ + 53, + 116, + 289, + 171 + ], + "type": "text", + "content": "[1] Constructions Aeronautiques, Adele Howe, Craig Knoblock, ISI Drew McDermott, Ashwin Ram, Manuela Veloso, Daniel Weld, David Wilkins SRI, Anthony Barrett, Dave Christianson, et al. PDDL: the planning domain definition language. Technical Report, Tech. Rep., 1998. 4, 5, 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 172, + 288, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 172, + 288, + 216 + ], + "spans": [ + { + "bbox": [ + 53, + 172, + 288, + 216 + ], + "type": "text", + "content": "[2] Marcin Andrychowicz, Filip Wolski, Alex Ray, Jonas Schneider, Rachel Fong, Peter Welinder, Bob McGrew, Josh Tobin, OpenAI Pieter Abbeel, and Wojciech Zaremba. Hindsight experience replay. 2017. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 217, + 288, + 250 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 217, + 288, + 250 + ], + "spans": [ + { + "bbox": [ + 53, + 217, + 288, + 250 + ], + "type": "text", + "content": "[3] Shuang Ao, Tianyi Zhou, Guodong Long, Qinghua Lu, Liming Zhu, and Jing Jiang. Co-pilot: Collaborative planning and reinforcement learning on sub-task curriculum. 2021." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 251, + 288, + 295 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 251, + 288, + 295 + ], + "spans": [ + { + "bbox": [ + 53, + 251, + 288, + 295 + ], + "type": "text", + "content": "[4] Shuang Ao, Tianyi Zhou, Jing Jiang, Guodong Long, Xuan Song, and Chengqi Zhang. Eat-c: environment-adversarial sub-task curriculum for efficient reinforcement learning. In ICML, 2022. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 297, + 288, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 297, + 288, + 319 + ], + "spans": [ + { + "bbox": [ + 53, + 297, + 288, + 319 + ], + "type": "text", + "content": "[5] Michael Bain and Claude Sammut. A framework for behavioural cloning. In Machine Intelligence, 1995. 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 320, + 288, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 320, + 288, + 374 + ], + "spans": [ + { + "bbox": [ + 53, + 320, + 288, + 374 + ], + "type": "text", + "content": "[6] Anthony Brohan, Noah Brown, Justice Carbajal, Yevgen Chebotar, Joseph Dabis, Chelsea Finn, Keerthana Gopalakrishnan, Karol Hausman, Alex Herzog, Jasmine Hsu, et al. Rt-1: Robotics transformer for real-world control at scale. arXiv preprint arXiv:2212.06817, 2022. 2, 4, 8" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 376, + 288, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 376, + 288, + 430 + ], + "spans": [ + { + "bbox": [ + 53, + 376, + 288, + 430 + ], + "type": "text", + "content": "[7] Anthony Brohan, Yevgen Chebotar, Chelsea Finn, Karol Hausman, Alexander Herzog, Daniel Ho, Julian Ibarz, Alex Irpan, Eric Jang, Ryan Julian, et al. Do as i can, not as i say: Grounding language in robotic affordances. In CoRL, 2023. 8" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 433, + 288, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 433, + 288, + 464 + ], + "spans": [ + { + "bbox": [ + 53, + 433, + 288, + 464 + ], + "type": "text", + "content": "[8] Kai-Wei Chang, Akshay Krishnamurthy, Alekh Agarwal, Hal Daumé III, and John Langford. Learning to search better than your teacher. In ICML, 2015. 8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 467, + 288, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 467, + 288, + 510 + ], + "spans": [ + { + "bbox": [ + 53, + 467, + 288, + 510 + ], + "type": "text", + "content": "[9] Lichang Chen, Jiuhai Chen, Tom Goldstein, Heng Huang, and Tianyi Zhou. Instructzero: Efficient instruction optimization for black-box large language models. arXiv preprint arXiv:2306.03082, 2023. 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 512, + 288, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 512, + 288, + 566 + ], + "spans": [ + { + "bbox": [ + 48, + 512, + 288, + 566 + ], + "type": "text", + "content": "[10] Marc-Alexandre Côté, Akos Kádár, Xingdi Yuan, Ben Kybartas, Tavian Barnes, Emery Fine, James Moore, Matthew Hausknecht, Layla El Asri, Mahmoud Adada, et al. Textworld: A learning environment for text-based games. In Computer Games Workshop, IJCAI, 2019. 4, 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 568, + 288, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 568, + 288, + 612 + ], + "spans": [ + { + "bbox": [ + 48, + 568, + 288, + 612 + ], + "type": "text", + "content": "[11] Wenliang Dai, Junnan Li, Dongxu Li, Anthony Tiong, Junqi Zhao, Weisheng Wang, Boyang Li, Pascale Fung, and Steven Hoi. InstructBLIP: Towards general-purpose vision-language models with instruction tuning. In NeurIPS, 2023. 2, 3, 5, 8" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 613, + 288, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 613, + 288, + 656 + ], + "spans": [ + { + "bbox": [ + 48, + 613, + 288, + 656 + ], + "type": "text", + "content": "[12] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. BERT: pre-training of deep bidirectional transformers for language understanding. In NAACL-HLT, 2019. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 658, + 288, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 288, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 288, + 712 + ], + "type": "text", + "content": "[13] Danny Driess, Fei Xia, Mehdi SM Sajjadi, Corey Lynch, Aakanksha Chowdhery, Brian Ichter, Ayzaan Wahid, Jonathan Thompson, Quan Vuong, Tianhe Yu, et al. Palm-e: An embodied multimodal language model. arXiv preprint arXiv:2303.03378, 2023. 2, 4, 8" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 712 + ], + "type": "list", + "angle": 0, + "index": 32, + "blocks": [ + { + "bbox": [ + 307, + 73, + 547, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 547, + 106 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 547, + 106 + ], + "type": "text", + "content": "[14] Meng Fang, Tianyi Zhou, Yali Du, Lei Han, and Zhengyou Zhang. Curriculum-guided hindsight experience replay. NeurIPS, 2019. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 108, + 547, + 151 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 108, + 547, + 151 + ], + "spans": [ + { + "bbox": [ + 307, + 108, + 547, + 151 + ], + "type": "text", + "content": "[15] Daocheng Fu, Xin Li, Licheng Wen, Min Dou, Pinlong Cai, Botian Shi, and Yu Qiao. Drive like a human: Rethinking autonomous driving with large language models. arXiv preprint arXiv:2307.07162, 2023. 2, 4" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 153, + 547, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 153, + 547, + 196 + ], + "spans": [ + { + "bbox": [ + 307, + 153, + 547, + 196 + ], + "type": "text", + "content": "[16] Jensen Gao, Bidipta Sarkar, Fei Xia, Ted Xiao, Jiajun Wu, Brian Ichter, Anirudha Majumdar, and Dorsa Sadigh. Physically grounded vision-language models for robotic manipulation. arXiv preprint arXiv:2309.02561, 2023. 8" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 198, + 547, + 243 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 198, + 547, + 243 + ], + "spans": [ + { + "bbox": [ + 307, + 198, + 547, + 243 + ], + "type": "text", + "content": "[17] Peng Gao, Jiaming Han, Renrui Zhang, Ziyi Lin, Shijie Geng, Aojun Zhou, Wei Zhang, Pan Lu, Conghui He, Xiangyu Yue, et al. Llama-adapter v2: Parameter-efficient visual instruction model. arXiv preprint arXiv:2304.15010, 2023. 5, 8, 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 244, + 547, + 287 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 244, + 547, + 287 + ], + "spans": [ + { + "bbox": [ + 307, + 244, + 547, + 287 + ], + "type": "text", + "content": "[18] Jiayi Guan, Guang Chen, Jiaming Ji, Long Yang, Zhijun Li, et al. Voce: Variational optimization with conservative estimation for offline safe reinforcement learning. NeurIPS, 36, 2023. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 289, + 546, + 322 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 289, + 546, + 322 + ], + "spans": [ + { + "bbox": [ + 307, + 289, + 546, + 322 + ], + "type": "text", + "content": "[19] Caglar Gülcehre, Sungjin Ahn, Ramesh Nallapati, Bowen Zhou, and Yoshua Bengio. Pointing the unknown words. In ACL, 2016. 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 324, + 547, + 356 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 324, + 547, + 356 + ], + "spans": [ + { + "bbox": [ + 307, + 324, + 547, + 356 + ], + "type": "text", + "content": "[20] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, 2016. 5" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 358, + 547, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 358, + 547, + 380 + ], + "spans": [ + { + "bbox": [ + 307, + 358, + 547, + 380 + ], + "type": "text", + "content": "[21] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross B. Girshick. Mask R-CNN. In ICCV, 2017. 5" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 382, + 546, + 404 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 382, + 546, + 404 + ], + "spans": [ + { + "bbox": [ + 307, + 382, + 546, + 404 + ], + "type": "text", + "content": "[22] Malte Helmert. The fast downward planning system. Journal of Artificial Intelligence Research, 2006. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 406, + 546, + 427 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 406, + 546, + 427 + ], + "spans": [ + { + "bbox": [ + 307, + 406, + 546, + 427 + ], + "type": "text", + "content": "[23] Jonathan Ho and Stefano Ermon. Generative adversarial imitation learning. In NeurIPS, 2016. 8" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 430, + 547, + 472 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 430, + 547, + 472 + ], + "spans": [ + { + "bbox": [ + 307, + 430, + 547, + 472 + ], + "type": "text", + "content": "[24] Wenlong Huang, Pieter Abbeel, Deepak Pathak, and Igor Mordatch. Language models as zero-shot planners: Extracting actionable knowledge for embodied agents. In ICML, 2022. 8" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 475, + 546, + 518 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 475, + 546, + 518 + ], + "spans": [ + { + "bbox": [ + 307, + 475, + 546, + 518 + ], + "type": "text", + "content": "[25] Wenlong Huang, Chen Wang, Ruohan Zhang, Yunzhu Li, Jiajun Wu, and Li Fei-Fei. Voxposer: Composable 3d value maps for robotic manipulation with language models. arXiv preprint arXiv:2307.05973, 2023. 2, 4, 8" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 520, + 547, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 520, + 547, + 553 + ], + "spans": [ + { + "bbox": [ + 307, + 520, + 547, + 553 + ], + "type": "text", + "content": "[26] Ahmed Hussein, Mohamed Medhat Gaber, Eyad Elyan, and Chrisina Jayne. Imitation learning: A survey of learning methods. ACM Computing Surveys (CSUR), 50(2):1-35, 2017. 4" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 555, + 546, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 555, + 546, + 586 + ], + "spans": [ + { + "bbox": [ + 307, + 555, + 546, + 586 + ], + "type": "text", + "content": "[27] Marcus Hutter. Universal artificial intelligence: Sequential decisions based on algorithmic probability. Springer Science & Business Media, 2004. 1" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 589, + 547, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 589, + 547, + 611 + ], + "spans": [ + { + "bbox": [ + 307, + 589, + 547, + 611 + ], + "type": "text", + "content": "[28] Hal Daumé III, John Langford, and Daniel Marcu. Search-based structured prediction. Mach. Learn., 2009. 8" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 613, + 547, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 613, + 547, + 656 + ], + "spans": [ + { + "bbox": [ + 307, + 613, + 547, + 656 + ], + "type": "text", + "content": "[29] Yunfan Jiang, Agrim Gupta, Zichen Zhang, Guanzhi Wang, Yongqiang Dou, Yanjun Chen, Li Fei-Fei, Anima Anandkumar, Yuke Zhu, and Linxi Fan. VIMA: robot manipulation with multimodal prompts. In ICML, 2023. 2, 4" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 307, + 658, + 547, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 658, + 547, + 712 + ], + "spans": [ + { + "bbox": [ + 307, + 658, + 547, + 712 + ], + "type": "text", + "content": "[30] Eric Kolve, Roozbeh Mottaghi, Winson Han, Eli VanderBilt, Luca Weihs, Alvaro Herrasti, Matt Deitke, Kiana Ehsani, Daniel Gordon, Yuke Zhu, et al. Ai2-thor: An interactive 3d environment for visual ai. arXiv preprint arXiv:1712.05474, 2017. 4, 5, 1" + } + ] + } + ], + "index": 31 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "26283" + } + ] + } + ], + "index": 33 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 288, + 116 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 288, + 116 + ], + "type": "text", + "content": "[31] Sergey Levine, Aviral Kumar, George Tucker, and Justin Fu. Offline reinforcement learning: Tutorial, review, and perspectives on open problems. arXiv preprint arXiv:2005.01643, 2020. 2, 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 118, + 288, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 118, + 288, + 162 + ], + "spans": [ + { + "bbox": [ + 49, + 118, + 288, + 162 + ], + "type": "text", + "content": "[32] Chunyuan Li, Zhe Gan, Zhengyuan Yang, Jianwei Yang, Linjie Li, Lijuan Wang, and Jianfeng Gao. Multimodal foundation models: From specialists to general-purpose assistants. arXiv preprint arXiv:2309.10020, 2023. 2, 3, 8" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 163, + 288, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 163, + 288, + 217 + ], + "spans": [ + { + "bbox": [ + 48, + 163, + 288, + 217 + ], + "type": "text", + "content": "[33] Dongxu Li, Junnan Li, Hung Le, Guangsen Wang, Silvio Savarese, and Steven C.H. Hoi. LAVIS: A one-stop library for language-vision intelligence. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 3: System Demonstrations), 2023. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 219, + 288, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 219, + 288, + 262 + ], + "spans": [ + { + "bbox": [ + 48, + 219, + 288, + 262 + ], + "type": "text", + "content": "[34] Junnan Li, Dongxu Li, Silvio Savarese, and Steven C. H. Hoi. BLIP-2: bootstrapping language-image pre-training with frozen image encoders and large language models. In ICML, 2023. 2, 3, 5, 8" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 264, + 288, + 306 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 264, + 288, + 306 + ], + "spans": [ + { + "bbox": [ + 48, + 264, + 288, + 306 + ], + "type": "text", + "content": "[35] Chen Liang, Jiahui Yu, Ming-Hsuan Yang, Matthew Brown, Yin Cui, Tuo Zhao, Boqing Gong, and Tianyi Zhou. Modulewise adaptive distillation for multimodality foundation models. 2024. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 308, + 288, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 308, + 288, + 352 + ], + "spans": [ + { + "bbox": [ + 48, + 308, + 288, + 352 + ], + "type": "text", + "content": "[36] Jacky Liang, Wenlong Huang, Fei Xia, Peng Xu, Karol Hausman, Brian Ichter, Pete Florence, and Andy Zeng. Code as policies: Language model programs for embodied control. In ICRA, 2023. 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 354, + 288, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 354, + 288, + 408 + ], + "spans": [ + { + "bbox": [ + 48, + 354, + 288, + 408 + ], + "type": "text", + "content": "[37] Yong Lin, Lu Tan, Hangyu Lin, Zeming Zheng, Renjie Pi, Jipeng Zhang, Shizhe Diao, Haoxiang Wang, Han Zhao, Yuan Yao, et al. Speciality vs generality: An empirical study on catastrophic forgetting in fine-tuning foundation models. arXiv preprint arXiv:2309.06256, 2023. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 410, + 288, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 410, + 288, + 441 + ], + "spans": [ + { + "bbox": [ + 48, + 410, + 288, + 441 + ], + "type": "text", + "content": "[38] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. arXiv preprint arXiv:2304.08485, 2023. 2, 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 444, + 288, + 465 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 444, + 288, + 465 + ], + "spans": [ + { + "bbox": [ + 48, + 444, + 288, + 465 + ], + "type": "text", + "content": "[39] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 467, + 288, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 467, + 288, + 521 + ], + "spans": [ + { + "bbox": [ + 48, + 467, + 288, + 521 + ], + "type": "text", + "content": "[40] Yecheng Jason Ma, William Liang, Guanzhi Wang, De-An Huang, Osbert Bastani, Dinesh Jayaraman, Yuke Zhu, Linxi Fan, and Anima Anandkumar. Eureka: Human-level reward design via coding large language models. arXiv preprint arXiv:2310.12931, 2023. 8" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 522, + 288, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 522, + 288, + 544 + ], + "spans": [ + { + "bbox": [ + 48, + 522, + 288, + 544 + ], + "type": "text", + "content": "[41] Vincent Micheli and François Fleuret. Language models are few-shot butlers. In EMNLP, 2021. 5, 8" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 545, + 288, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 545, + 288, + 599 + ], + "spans": [ + { + "bbox": [ + 48, + 545, + 288, + 599 + ], + "type": "text", + "content": "[42] Yao Mu, Qinglong Zhang, Mengkang Hu, Wenhai Wang, Mingyu Ding, Jun Jin, Bin Wang, Jifeng Dai, Yu Qiao, and Ping Luo. Embodiedgpt: Vision-language pre-training via embodied chain of thought. arXiv preprint arXiv:2305.15021, 2023. 2, 4, 8" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 601, + 288, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 601, + 288, + 633 + ], + "spans": [ + { + "bbox": [ + 48, + 601, + 288, + 633 + ], + "type": "text", + "content": "[43] OpenAI. GPT-4V(ison) system card. https://cdn.openai.com/papers/GPTV_System_Card.pdf, 2023.2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 635, + 288, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 635, + 288, + 679 + ], + "spans": [ + { + "bbox": [ + 48, + 635, + 288, + 679 + ], + "type": "text", + "content": "[44] Abhishek Padalkar, Acorn Pooley, Ajinkya Jain, Alex Bewley, Alex Herzog, Alex Irpan, Alexander Khazatsky, Anant Rai, Anikait Singh, Anthony Brohan, et al. Open x-embediment: Robotic learning datasets and rt-x models, 2023. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 681, + 288, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 681, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 681, + 288, + 713 + ], + "type": "text", + "content": "[45] Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, Ilya Sutskever, et al. Language models are unsupervised multitask learners. OpenAI blog, 2019. 5" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 712 + ], + "type": "list", + "angle": 0, + "index": 31, + "blocks": [ + { + "bbox": [ + 307, + 73, + 547, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 547, + 127 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 547, + 127 + ], + "type": "text", + "content": "[46] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, 2021. 8, 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 129, + 547, + 174 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 129, + 547, + 174 + ], + "spans": [ + { + "bbox": [ + 307, + 129, + 547, + 174 + ], + "type": "text", + "content": "[47] Rafael Rafailov, Archit Sharma, Eric Mitchell, Stefano Ermon, Christopher D Manning, and Chelsea Finn. Direct preference optimization: Your language model is secretly a reward model. arXiv preprint arXiv:2305.18290, 2023. 8" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 175, + 547, + 218 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 175, + 547, + 218 + ], + "spans": [ + { + "bbox": [ + 307, + 175, + 547, + 218 + ], + "type": "text", + "content": "[48] Rafael Rafailov, Archit Sharma, Eric Mitchell, Christopher D Manning, Stefano Ermon, and Chelsea Finn. Direct preference optimization: Your language model is secretly a reward model. NeurIPS, 2024. 2, 4" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 220, + 545, + 243 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 220, + 545, + 243 + ], + "spans": [ + { + "bbox": [ + 307, + 220, + 545, + 243 + ], + "type": "text", + "content": "[49] Stéphane Ross and Drew Bagnell. Efficient reductions for imitation learning. In AISTATS, 2010. 8" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 245, + 545, + 278 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 245, + 545, + 278 + ], + "spans": [ + { + "bbox": [ + 307, + 245, + 545, + 278 + ], + "type": "text", + "content": "[50] Stephane Ross and J Andrew Bagnell. Reinforcement and imitation learning via interactive no-regret learning. arXiv preprint arXiv:1406.5979, 2014. 8" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 280, + 545, + 312 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 280, + 545, + 312 + ], + "spans": [ + { + "bbox": [ + 307, + 280, + 545, + 312 + ], + "type": "text", + "content": "[51] Stéphane Ross, Geoffrey J. Gordon, and Drew Bagnell. A reduction of imitation learning and structured prediction to no-regret online learning. In AISTATS, 2011. 2, 4, 8" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 315, + 547, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 315, + 547, + 346 + ], + "spans": [ + { + "bbox": [ + 307, + 315, + 547, + 346 + ], + "type": "text", + "content": "[52] John Schulman, Sergey Levine, Pieter Abbeel, Michael I. Jordan, and Philipp Moritz. Trust region policy optimization. In ICML, 2015. 8" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 350, + 547, + 382 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 350, + 547, + 382 + ], + "spans": [ + { + "bbox": [ + 307, + 350, + 547, + 382 + ], + "type": "text", + "content": "[53] John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Radford, and Oleg Klimov. Proximal policy optimization algorithms. arXiv preprint arXiv:1707.06347, 2017. 8" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 384, + 547, + 427 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 384, + 547, + 427 + ], + "spans": [ + { + "bbox": [ + 307, + 384, + 547, + 427 + ], + "type": "text", + "content": "[54] Noah Shinn, Federico Cassano, Ashwin Gopinath, Karthik R Narasimhan, and Shunyu Yao. Reflexion: Language agents with verbal reinforcement learning. In NeurIPS, 2023. 3, 4, 5, 6, 8" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 430, + 547, + 483 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 430, + 547, + 483 + ], + "spans": [ + { + "bbox": [ + 307, + 430, + 547, + 483 + ], + "type": "text", + "content": "[55] Mohit Shridhar, Jesse Thomason, Daniel Gordon, Yonatan Bisk, Winson Han, Roozbeh Mottaghi, Luke Zettlemoyer, and Dieter Fox. ALFRED: A benchmark for interpreting grounded instructions for everyday tasks. In CVPR, 2020. 5, 8, 3, 6" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 487, + 547, + 530 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 487, + 547, + 530 + ], + "spans": [ + { + "bbox": [ + 307, + 487, + 547, + 530 + ], + "type": "text", + "content": "[56] Mohit Shridhar, Xingdi Yuan, Marc-Alexandre Côté, Yonatan Bisk, Adam Trischler, and Matthew J. Hausknecht. Alfworld: Aligning text and embodied environments for interactive learning. In ICLR, 2021. 2, 3, 4, 5, 8, 1" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 532, + 547, + 565 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 532, + 547, + 565 + ], + "spans": [ + { + "bbox": [ + 307, + 532, + 547, + 565 + ], + "type": "text", + "content": "[57] Theodore Sumers, Kenneth Marino, Arun Ahuja, Rob Fergus, and Ishita Dasgupta. Distilling internet-scale vision-language models into embodied agents. In ICML, 2023. 2, 4, 8" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 567, + 547, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 567, + 547, + 600 + ], + "spans": [ + { + "bbox": [ + 307, + 567, + 547, + 600 + ], + "type": "text", + "content": "[58] Wen Sun, J. Andrew Bagnell, and Byron Boots. Truncated horizon policy search: Combining reinforcement learning & imitation learning. In ICLR, 2018. 8" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 601, + 547, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 601, + 547, + 645 + ], + "spans": [ + { + "bbox": [ + 307, + 601, + 547, + 645 + ], + "type": "text", + "content": "[59] Guanzhi Wang, Yuqi Xie, Yunfan Jiang, Ajay Mandlekar, Chaowei Xiao, Yuke Zhu, Linxi Fan, and Anima Anandkumar. Voyager: An open-ended embodied agent with large language models. arXiv preprint arXiv:2305.16291, 2023. 8" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 647, + 547, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 647, + 547, + 712 + ], + "spans": [ + { + "bbox": [ + 307, + 647, + 547, + 712 + ], + "type": "text", + "content": "[60] Xiyao Wang, Yuhang Zhou, Xiaoyu Liu, Hongjin Lu, Yuancheng Xu, Feihong He, Jaehong Yoon, Taixi Lu, Gedas Bertasius, Mohit Bansal, et al. Mementos: A comprehensive benchmark for multimodal large language model reasoning over image sequences. arXiv preprint arXiv:2401.10529, 2024. 2" + } + ] + } + ], + "index": 30 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "26284" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "text", + "content": "[61] Zihao Wang, Shaofei Cai, Anji Liu, Xiaojian Ma, and Yitao Liang. Describe, explain, plan and select: Interactive planning with large language models enables open-world multi-task agents. arXiv preprint arXiv:2302.01560, 2023. 5, 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 118, + 288, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 118, + 288, + 162 + ], + "spans": [ + { + "bbox": [ + 49, + 118, + 288, + 162 + ], + "type": "text", + "content": "[62] Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Brian Ichter, Fei Xia, Ed H. Chi, Quoc V. Le, and Denny Zhou. Chain-of-thought prompting elicits reasoning in large language models. In NeurIPS, 2022. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 163, + 287, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 163, + 287, + 196 + ], + "spans": [ + { + "bbox": [ + 49, + 163, + 287, + 196 + ], + "type": "text", + "content": "[63] Michael Wooldridge and Nicholas R Jennings. Intelligent agents: Theory and practice. The knowledge engineering review, 10(2):115-152, 1995. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 198, + 287, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 198, + 287, + 251 + ], + "spans": [ + { + "bbox": [ + 49, + 198, + 287, + 251 + ], + "type": "text", + "content": "[64] Qingyun Wu, Gagan Bansal, Jieyu Zhang, Yiran Wu, Shaokun Zhang, Erkang Zhu, Beibin Li, Li Jiang, Xiaoyun Zhang, and Chi Wang. Autogen: Enabling next-gen llm applications via multi-agent conversation framework. arXiv preprint arXiv:2308.08155, 2023. 5, 6, 8" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 49, + 254, + 288, + 306 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 254, + 288, + 306 + ], + "spans": [ + { + "bbox": [ + 49, + 254, + 288, + 306 + ], + "type": "text", + "content": "[65] Zhiheng Xi, Wenxiang Chen, Xin Guo, Wei He, Yiwen Ding, Boyang Hong, Ming Zhang, Junzhe Wang, Senjie Jin, Enyu Zhou, et al. The rise and potential of large language model based agents: A survey. arXiv preprint arXiv:2309.07864, 2023. 8" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 49, + 309, + 288, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 309, + 288, + 363 + ], + "spans": [ + { + "bbox": [ + 49, + 309, + 288, + 363 + ], + "type": "text", + "content": "[66] Zhenhua Xu, Yujia Zhang, Enze Xie, Zhen Zhao, Yong Guo, Kenneth KY Wong, Zhenguo Li, and Hengshuang Zhao. Drivegpt4: Interpretable end-to-end autonomous driving via large language model. arXiv preprint arXiv:2310.01412, 2023. 2, 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 49, + 365, + 287, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 365, + 287, + 398 + ], + "spans": [ + { + "bbox": [ + 49, + 365, + 287, + 398 + ], + "type": "text", + "content": "[67] Zihui Xue, Zhengqi Gao, Sucheng Ren, and Hang Zhao. The modality focusing hypothesis: Towards understanding cross-modal knowledge distillation. In ICLR, 2023. 8" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 49, + 399, + 288, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 399, + 288, + 453 + ], + "spans": [ + { + "bbox": [ + 49, + 399, + 288, + 453 + ], + "type": "text", + "content": "[68] Jingkang Yang, Yuhao Dong, Shuai Liu, Bo Li, Ziyue Wang, Chencheng Jiang, Haoran Tan, Jiamu Kang, Yuanhan Zhang, Kaiyang Zhou, et al. Octopus: Embodied vision-language programmer from environmental feedback. arXiv preprint arXiv:2310.08588, 2023. 3, 4, 8" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 49, + 456, + 287, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 456, + 287, + 498 + ], + "spans": [ + { + "bbox": [ + 49, + 456, + 287, + 498 + ], + "type": "text", + "content": "[69] Sherry Yang, Ofir Nachum, Yilun Du, Jason Wei, Pieter Abbeel, and Dale Schuurmans. Foundation models for decision making: Problems, methods, and opportunities. arXiv preprint arXiv:2303.04129, 2023. 8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 49, + 501, + 287, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 501, + 287, + 533 + ], + "spans": [ + { + "bbox": [ + 49, + 501, + 287, + 533 + ], + "type": "text", + "content": "[70] Yijun Yang, Jing Jiang, Tianyi Zhou, Jie Ma, and Yuhui Shi. Pareto policy pool for model-based offline reinforcement learning. In ICLR, 2022. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 49, + 535, + 287, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 535, + 287, + 567 + ], + "spans": [ + { + "bbox": [ + 49, + 535, + 287, + 567 + ], + "type": "text", + "content": "[71] Yijun Yang, Tianyi Zhou, Jing Jiang, Guodong Long, and Yuhui Shi. Continual task allocation in meta-policy network via sparse prompting. In ICML, 2023. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 49, + 569, + 287, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 569, + 287, + 612 + ], + "spans": [ + { + "bbox": [ + 49, + 569, + 287, + 612 + ], + "type": "text", + "content": "[72] Shunyu Yao, Dian Yu, Jeffrey Zhao, Izhak Shafran, Thomas L Griffiths, Yuan Cao, and Karthik Narasimhan. Tree of thoughts: Deliberate problem solving with large language models. arXiv preprint arXiv:2305.10601, 2023. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 49, + 614, + 287, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 614, + 287, + 656 + ], + "spans": [ + { + "bbox": [ + 49, + 614, + 287, + 656 + ], + "type": "text", + "content": "[73] Shunyu Yao, Jeffrey Zhao, Dian Yu, Nan Du, Izhak Shafran, Karthik R. Narasimhan, and Yuan Cao. React: Synergizing reasoning and acting in language models. In ICLR, 2023. 3, 5, 6, 8, 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 49, + 659, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 659, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 49, + 659, + 287, + 712 + ], + "type": "text", + "content": "[74] Lianmin Zheng, Wei-Lin Chiang, Ying Sheng, Siyuan Zhuang, Zhanghao Wu, Yonghao Zhuang, Zi Lin, Zhuohan Li, Dacheng Li, Eric Xing, et al. Judging llm-as-a-judge with mt-bench and chatbot arena. arXiv preprint arXiv:2306.05685, 2023. 3" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 195 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 307, + 73, + 547, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 547, + 106 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 547, + 106 + ], + "type": "text", + "content": "[75] Victor Zhong, Dipendra Misra, Xingdi Yuan, and Marc-Alexandre Côté. Policy improvement using language feedback models. arXiv preprint arXiv:2402.07876, 2024. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 108, + 547, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 108, + 547, + 150 + ], + "spans": [ + { + "bbox": [ + 307, + 108, + 547, + 150 + ], + "type": "text", + "content": "[76] Deyao Zhu, Jun Chen, Xiaogian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models. arXiv preprint arXiv:2304.10592, 2023. 2, 3, 5, 8" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 152, + 547, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 152, + 547, + 195 + ], + "spans": [ + { + "bbox": [ + 307, + 152, + 547, + 195 + ], + "type": "text", + "content": "[77] Brianna Zitkovich, Tianhe Yu, Sichun Xu, Peng Xu, Ted Xiao, Fei Xia, Jialin Wu, Paul Wohlhart, Stefan Welker, Ayzaan Wahid, et al. Rt-2: Vision-language-action models transfer web knowledge to robotic control. In CoRL, 2023. 2, 3, 4, 8" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "26285" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/EmbodiedScan_ A Holistic Multi-Modal 3D Perception Suite Towards Embodied AI/ab5e0a2a-2a5f-41d4-a08d-01fd4564b83b_content_list.json b/2024/EmbodiedScan_ A Holistic Multi-Modal 3D Perception Suite Towards Embodied AI/ab5e0a2a-2a5f-41d4-a08d-01fd4564b83b_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..a28780e81e4092ccdb389ea507c518f98d2e220a --- /dev/null +++ b/2024/EmbodiedScan_ A Holistic Multi-Modal 3D Perception Suite Towards Embodied AI/ab5e0a2a-2a5f-41d4-a08d-01fd4564b83b_content_list.json @@ -0,0 +1,1537 @@ +[ + { + "type": "text", + "text": "EmbodiedScan: A Holistic Multi-Modal 3D Perception Suite Towards Embodied AI", + "text_level": 1, + "bbox": [ + 179, + 119, + 790, + 162 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Tai Wang $^{1*}$ , Xiaohan Mao $^{1,2*}$ , Chenming Zhu $^{1,3*}$ , Runsen Xu $^{1,4}$ , Ruiyuan Lyu $^{1,5}$ , Peisen Li $^{1,5}$ , Xiao Chen $^{1,4}$ , Wenwei Zhang $^{1}$ , Kai Chen $^{1}$ , Tianfan Xue $^{4}$ , Xihui Liu $^{3}$ , Cewu Lu $^{2}$ , Dahua Lin $^{1,4}$ , Jiangmiao Pang $^{1\\boxtimes}$", + "bbox": [ + 112, + 176, + 854, + 232 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Shanghai AI Laboratory $^{2}$ Shanghai Jiao Tong University $^{3}$ The University of Hong Kong $^{4}$ The Chinese University of Hong Kong $^{5}$ Tsinghua University *Equal contribution Corresponding author", + "bbox": [ + 165, + 239, + 787, + 292 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "EmbodiedScan: 5k scans, 1M ego-centric images, 1M language prompts, 160k instances, 760+ categories", + "text_level": 1, + "bbox": [ + 145, + 305, + 815, + 320 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/c52f8b1370f793eaae7fde5751da68ce5caf52fd8989aae25eeed199e0e26615.jpg", + "image_caption": [ + "Figure 1. EmbodiedScan provides a multi-modal, ego-centric 3D perception dataset with massive real-scanned data and rich annotations for indoor scenes. It benchmarks language-grounded holistic 3D scene understanding capabilities for real-world embodied agents." + ], + "image_footnote": [], + "bbox": [ + 119, + 327, + 839, + 523 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 571, + 313, + 585 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In the realm of computer vision and robotics, embodied agents are expected to explore their environment and carry out human instructions. This necessitates the ability to fully understand 3D scenes given their first-person observations and contextualize them into language for interaction. However, traditional research focuses more on scene-level input and output setups from a global view. To address the gap, we introduce EmbodiedScan, a multi-modal, ego-centric 3D perception dataset and benchmark for holistic 3D scene understanding. It encompasses over 5k scans encapsulating 1M ego-centric RGB-D views, 1M language prompts, 160k 3D-oriented boxes spanning over 760 categories, some of which partially align with LVIS, and dense semantic occupancy with 80 common categories. Building upon this database, we introduce a baseline framework named Embodied Perceptron. It is capable of processing an arbitrary number of multi-modal inputs and demonstrates remarkable 3D perception capabilities, both within the two series of benchmarks we set up, i.e., fundamental 3D perception tasks and language-grounded tasks, and in the wild.", + "bbox": [ + 75, + 592, + 473, + 893 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 500, + 571, + 630, + 585 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Consider an embodied agent operating in an indoor environment. It commences its journey devoid of any prior knowledge about the scene, guided only by an initial instruction. As it begins to explore, it recognizes objects in context and acts with goals along with language interaction. In this process, a commonly needed, fundamental perception capability is to establish a holistic 3D scene understanding given ego-centric observations. This understanding operates at the scene level, covers both object semantics and scene geometry, and can be grounded in language descriptions.", + "bbox": [ + 496, + 597, + 890, + 748 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Nonetheless, subtle but significant discrepancies exist between this expectation and research problems examined within the computer vision community. Most previous studies have primarily revolved around scene-level input and output problems from a global view [13, 34, 40], i.e., taking reconstructed 3D point clouds or meshes as inputs and predicting 3D object bounding boxes or segmenting point clouds. Regarding data, earlier datasets targeting egocentric RGB-D inputs are either too small [12, 45] or lack comprehensive annotations [6, 51] to support the aforementioned results.", + "bbox": [ + 496, + 750, + 892, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "19757", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "table", + "img_path": "images/2f94649cea2c5a1165bd5000545b54bd205ce5c0f3536917ebf879b17f86a911.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Dataset#Scans#ImgS#Objs#Cats#PromptsEgo Capture3D Annotations
Replica [46]35----XX
NYU v2 [12]4641.4k35k14-✓X
SUN RGB-D [45]-10k-37-Mono.Box
ScanNet [13, 39]1513264k36k1852k [8]✓Seg., Lang.
Matterport3D [6]2056194k51k40-Multi-ViewSeg.
3RScan [51]1482363k---✓Seg.
ArkitScenes [2]5047450k51k17-✓Box
HyperSim [38]46177k-40+-Mono. & Syn.Box
EmbodiedScan5185890k160k762970k✓Box, Occ., Lang.
", + "bbox": [ + 78, + 85, + 625, + 204 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Table 1. Comparison with other 3D indoor scene datasets. \"Cats\" refers to the categories with box annotations for the 3D detection benchmark. EmbodiedScan features more than $10 \\times$ categories, prompts, and the most diverse annotations. The numbers are still scaling up with further annotations. Mono./Syn./Lang. means Monocular/Synthetic/Language.", + "bbox": [ + 75, + 214, + 619, + 272 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/71464ca534167399ffc23364e742f6ac43003909e34525eb95af974672f29b08.jpg", + "image_caption": [ + "Figure 2. Dataset composition. Embodied-Scan is composed of three data sources and has similar scans, images, objects, and categories in each of them." + ], + "image_footnote": [], + "bbox": [ + 635, + 84, + 903, + 203 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "tioned research. It is also not feasible to generate such realistic views by rendering from the existing imperfect meshes. On the other hand, since we cannot trivially obtain the reconstruction of a new environment, models trained with scene-level input are not directly applicable in practice.", + "bbox": [ + 75, + 281, + 468, + 356 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To bridge this divide, we introduce a multi-modal, ego-centric 3D perception dataset and benchmark for holistic 3D scene understanding, termed EmbodiedScan, aimed at facilitating real-world embodied AI applications (Fig. 1). This dataset exploits existing large-scale 3D scene datasets [6, 13, 51] but re-purposes them for continuous scene-level perception from the first-view RGB-D streams. Unlike previous works that offer only point segmentation labels with limited semantics, we employ a SAM-assisted [22] pipeline to annotate objects with oriented 3D bounding boxes and generate language prompts on top. Consequently, EmbodiedScan provides more than 5k scans, nearly 1M ego-centric RGB-D images, and multi-modality annotations, covering 3D oriented boxes with more than 160k instances spanning over 760 categories, dense semantic occupancy with 80 common categories, and 1M language descriptions focusing on spatial relationships among objects.", + "bbox": [ + 75, + 357, + 470, + 613 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Built upon this dataset, we devise a baseline framework for ego-centric 3D perception, Embodied Perceptron. It accepts RGB-D sequences and texts as inputs and manifests scalability to any number of views input with encoders shared across different tasks. With the encoded 2D and 3D features, we employ dense fusion and isomorphic multilevel fusion across them guided by the perspective projection to produce 3D scene and object representations, which are further processed to decode occupancy and 3D box predictions. The derived 3D representations can be further integrated with text embeddings for 3D visual grounding, thus supporting language-grounded applications.", + "bbox": [ + 75, + 614, + 468, + 794 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We establish two series of benchmarks on Embodied-Scan: 1) fundamental 3D perception benchmarks focusing on traditional tasks, including 3D detection and semantic occupancy prediction under different input settings, and 2) a language-grounded scene understanding benchmark with 3D visual grounding as a preliminary exploration. Experimental results validate the effectiveness of our baseline", + "bbox": [ + 75, + 795, + 470, + 898 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "model on EmbodiedScan and demonstrate its generalization ability in the wild. Detailed analysis further underscores the value of EmbodiedScan and highlights the primary challenges posed by this new setup.", + "bbox": [ + 496, + 281, + 892, + 342 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 358, + 640, + 375 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3D Scene Datasets. The development of 3D scene understanding has benefited from large-scale, high-quality datasets like KITTI [16] and SUN RGB-D [45]. These foundational datasets have paved the way for subsequent larger and more diverse collections targeting indoor [6, 13, 38, 51] and driving scenes [4, 7, 32, 47]. However, compared to autonomous driving datasets, those meant for indoor scenes still lack variety in terms of scenes and object diversity (Tab. 1). In contrast, EmbodiedScan provides a large amount of multi-modal data with much richer annotations. Furthermore, it differs by placing an emphasis on the ego-centric perspective within its setup, a feature often overlooked in previous works [2, 13].", + "bbox": [ + 496, + 383, + 892, + 580 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Except for these conventional dataset works, Omni3D [3] integrates urban and indoor datasets for monocular 3D detection. Our focus, however, lies in indoor scenes due to their unique challenges but has a larger amount of data and annotations, e.g., more than $3 \\times$ images and categories with more than 10 instances. In addition, we offer a comprehensive exploration of more general problems for ego-centric 3D perception, such as continuous perception and visual grounding. Other embodied AI datasets like HM3D [36, 56] and HSSD [21] provide ample interaction opportunities but can suffer from poor transferability to real-world scenarios due to their imperfect meshes or synthetic data. Conversely, EmbodiedScan is based on real-scanned RGB-D images, offering a more realistic playground for model training.", + "bbox": [ + 496, + 583, + 893, + 808 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3D Object Detection & Occupancy Prediction. 3D detection and occupancy prediction, as fundamental tasks in 3D perception, focus on different aspects of 3D scene understanding. The former focuses on recognizing foreground objects through a sparse and efficient representation - a set of 3D cuboids corresponding to instances of interest -", + "bbox": [ + 496, + 810, + 893, + 898 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "19758", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/c59cb8467a1dc5e78484e2bdbf91952e7c49bc91d6bb0c8fc83917520d78d2c8.jpg", + "image_caption": [ + "(a) SAM-Assisted Oriented 3D Bounding Boxes Annotation." + ], + "image_footnote": [], + "bbox": [ + 83, + 85, + 472, + 207 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/1439bf27f5175c1f831d53a0f34eb65edeb4c7e241438e29929d9ae590aa5b03.jpg", + "image_caption": [ + "(c) Instance Statistics (Increase w.r.t. ScanNet)." + ], + "image_footnote": [], + "bbox": [ + 81, + 233, + 475, + 344 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/da6ddacfbcc39a354bafab126b48f1dc899c1767819bab09815df3e84b347e9e.jpg", + "image_caption": [ + "(b) 3D Boxes and Language Prompt Statistics." + ], + "image_footnote": [], + "bbox": [ + 491, + 87, + 890, + 212 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/eec26b8dd828e242d2f14d5e57e2d4ced8f4a0d0803213ab6f86bc14ea5e5f0a.jpg", + "image_caption": [ + "(d) Occupancy Statistics.", + "Figure 3. EmbodiedScan annotation and statistics. (a) UI for 3D box annotation. We select keyframes and generate their SAM masks with corresponding axis-aligned boxes. With simple clicks, annotators can create 3D boxes for target objects and further adjust them with reference in three orthogonal views and images. (b) Small boxes $(< 1m^3)$ increase more & prompt statistics. obj./avg./des. refer to objects/average/descriptions. (c) We show the number of instances per category (300 classes). For categories that exist in ScanNet, we plot the absolute increase and observe a significant improvement. (d) We plot the occupancy distribution for each category and see a different word cloud distribution. These two clouds show different aspects, occupied space vs. number of instances, of this dataset." + ], + "image_footnote": [], + "bbox": [ + 491, + 233, + 883, + 343 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "while the latter offers a dense, structured pattern that benefits downstream planning. The research community has developed solutions ranging from single-modality, such as LiDAR-based [23, 29, 34, 40, 42, 57, 59] or camera-only approaches [28, 41, 44, 52-54], to multi-modality techniques [26, 30, 35, 50, 58]. Recently, occupancy as a representation, due to its potential in handling unknown semantics and irregular object shapes, has gained more attention [5, 19, 25, 43, 49, 55]. Given their distinctive focuses, we selected these two tasks to form the fundamental 3D perception track on EmbodiedScan.", + "bbox": [ + 75, + 460, + 468, + 626 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Previous works on indoor scenes mainly centered around 3D detection with limitations in object orientations, semantic categories, and input format [29, 34, 40]. In practice, a model is expected to perceive the environment during egocentric exploration, ultimately providing a holistic understanding inclusive of rich semantics, scene geometry, and object poses. To this end, EmbodiedScan and our proposed framework, Embodied Perceptron, provide this necessary data foundation and baseline methodology.", + "bbox": [ + 75, + 628, + 468, + 763 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Language-Grounded 3D Scene Understanding. Language plays a crucial role in human-computer interaction, heightened by recent advances in Large Language Models (LLMs). Its integration with 3D scene understanding is vital for future embodied agents. Past research first explored 3D visual grounding [1, 8, 18, 20] and established new benchmarks including 3D dense captioning [9, 10], open-vocabulary 3D segmentation [15, 33, 48] and detection [31, 61]. This paper focuses on 3D visual grounding", + "bbox": [ + 75, + 763, + 468, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "first, with plans to expand language annotations and benchmarks in the future. Our visual grounding benchmark aligns with the multi-view setting of the basic 3D perception track, taking multiple ego-centric RGB-D images as input, and includes tenfold more complex prompts.", + "bbox": [ + 496, + 460, + 890, + 536 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Dataset", + "text_level": 1, + "bbox": [ + 500, + 545, + 588, + 560 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "This section presents the dataset construction, including data processing and annotation, and shows the statistics.", + "bbox": [ + 496, + 566, + 890, + 598 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Data Collection & Processing", + "text_level": 1, + "bbox": [ + 500, + 604, + 759, + 622 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Ego-Centric Sensor Data Collection. Considering there have been readily available 3D indoor scene scans from existing datasets, we start with integrating those providing ego-centric RGB-D captures with the corresponding camera poses. Given the compatibility of ScanNet [13], 3RScan [51], and Matterport3D [6], we select the high-quality part with complete regions and necessary annotations to form the initial version of EmbodiedScan (Fig. 2). ARKitScenes [2], possessing different data organization, sensors, and annotations, is considered for future inclusion.", + "bbox": [ + 496, + 628, + 890, + 777 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Frame Selection & Scene Division. Although these datasets all have RGB-D data, the data format, sampling frequency, and relationships among viewpoints are different. See more details in the appendix. We first unified the format into a general multi-view case to fit Matterport3D by adding randomness when loading images but maintaining sequential continuity for ScanNet and 3RScan during inference. Our model can thus handle both temporal and", + "bbox": [ + 496, + 779, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "19759", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "randomly captured multi-view images. Additionally, we divided building-scale scenes of Matterport3D into regions based on official annotation, selecting corresponding images with depth points falling into the region. As for different sampling rates of images in ScanNet and 3RScan videos, we sample one keyframe per 10 frames for ScanNet and keep all the images for 3RScan. The uniform sampling is generally in line with the actual situation.", + "bbox": [ + 75, + 90, + 472, + 210 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Global Coordinate System. A global coordinate system is necessary to aggregate multi-view observations and serve as a reference for outputs. We follow the convention of ScanNet, deriving a system with the origin around the center of the scene, the horizontal plane lying on the floor and axes aligning with walls [34]. This post-processing harmonizes the data distribution, slightly improving performance on benchmarks. Practical applications may not have such a prior global system or vary according to observations, posing another interesting problem for future exploration.", + "bbox": [ + 75, + 212, + 472, + 364 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Annotation", + "text_level": 1, + "bbox": [ + 76, + 375, + 202, + 390 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We provide three types of annotations - 3D bounding boxes, semantic occupancy, and language descriptions - each serving to enrich different aspects of scene understanding.", + "bbox": [ + 75, + 398, + 468, + 445 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3D Bounding Boxes. Following standard definitions [2, 3], a cuboid is defined by its 3D center, size, and ZXY Euler angle orientation. We used the Segment Anything Model (SAM) [22] and a customized annotation tool based on [24] (Fig. 3a) to address limitations in existing 3D box annotations, i.e., lack of orientation and small object annotations. It supports the conventional functionality of annotating 3D boxes with orientation in three orthographic views. Furthermore, we sample several keyframes with clear imaging according to the camera pose changes and ensure they cover non-overlap regions and most objects to generate SAM masks and axis-aligned boxes for further adjustment. We work with an annotation team and check the quality of all the labels in the end. Each scene takes around 10-30 minutes to annotate, varying with the scene complexity.", + "bbox": [ + 75, + 446, + 472, + 672 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Semantic Occupancy. Semantic occupancy necessitates accurate boundaries across semantic regions without considering object pose or recalling all the objects, so the original point cloud segmentation annotations were more suitable to be used for deriving occupancy ground truth. For each voxel, we assigned the category with the most points as the semantic label for that cell. A compromise between perception granularity and computational efficiency resulted in $40 \\times 40 \\times 16$ occupancy maps in the perception range $[-3.2m \\sim 3.2m, -3.2m \\sim 3.2m, -0.78m \\sim 1.78m]$ along the X-Y (horizontal) plane and Z (vertical) axis.", + "bbox": [ + 75, + 672, + 472, + 839 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Language Descriptions. Given updated 3D bounding boxes annotated with orientations, we derive the language prompts that describe the spatial relationships among objects following SR3D [1]. They serve as the prompt input", + "bbox": [ + 75, + 839, + 472, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "to the language-grounded perception models for performing 3D visual grounding. Due to increased object density after annotation, identifying unique objects became more challenging. To overcome this, we combined multiple spatial relationship prompts to exclusively ground objects. See more samples in the appendix.", + "bbox": [ + 496, + 90, + 893, + 183 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Statistics", + "text_level": 1, + "bbox": [ + 500, + 191, + 607, + 208 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Vocabulary Construction. During labeling, we ask annotators to write semantic categories in an open-vocabulary manner. This was efficient and suited the complex, large-vocabulary dataset and can provide natural annotations for future open-world research. To sort out these labels, we used Sentence-BERT [37] to cluster similar categories with text embeddings, match them to WordNet nodes, and finally revise and merge them manually. The vocabulary shares common categories with COCO (50/69 indoor classes) and LVIS (550/1203).", + "bbox": [ + 496, + 215, + 893, + 367 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Instance Statistics. We first show the instances of different categories in Fig. 3c. Our dataset contains over 760 categories, covering common objects in our daily life. More than 288 categories have over 10 instances, and around 400 categories have more than 5 instances. These numbers are $20 \\times$ higher than most previous works with 3D box annotations and $3 \\times$ higher than ScanNet instance segmentation annotations with more than 5 instances. There is also a notable increase in object numbers of small boxes and different categories (Fig. 3b and 3c). We remove four categories, {wall, ceiling, floor, object} in our 3D detection benchmark and divide the remaining 284 categories into three splits, {head, common, tail} with {90, 94, 100} classes.", + "bbox": [ + 496, + 368, + 893, + 564 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Occupancy Statistics. Semantic occupancy statistics (Fig. 3d) reveal the space occupied by different categories, relevant for navigation and motion planning. We chose the first 80 categories for our occupancy prediction benchmark based on distribution and significance in downstream tasks.", + "bbox": [ + 496, + 565, + 893, + 641 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Language Prompts Statistics. Generated language prompts following SR3D fall into five types of spatial object-to-object relations: Horizontal Proximity, Vertical Proximity, Support, Allocentric, and Between. If a scene has between 2 and 6 instances of a certain category, these categories are considered valid target categories. If a scene has a single instance of a category, it is selected as the anchor category. The training/validation set contains 801711/168322 language prompts, nearly 10 times larger than the original SR3D datasets (Fig. 3b and Tab. 1).", + "bbox": [ + 496, + 641, + 893, + 792 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Embodied Perceptron", + "text_level": 1, + "bbox": [ + 498, + 801, + 709, + 819 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Given this dataset, we can take multi-modality input, including RGB images, point clouds derived from depth maps as well as language prompts, to extract multi-modal representations and perform different downstream tasks. This section provides a baseline, namely Embodied Perceptron,", + "bbox": [ + 496, + 825, + 893, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "19760", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/3ccc529bbff2fad98adfb99c9489e207499db1022647684bd5ad2043567f73e0.jpg", + "image_caption": [ + "Figure 4. Embodied Perceptron accepts RGB-D sequence with any number of views along with texts as multi-modal input. It uses classical encoders to extract features for each modality and adopts dense and isomorphic sparse fusion with corresponding decoders for different predictions. The 3D features integrated with the text feature can be further used for language-grounded understanding." + ], + "image_footnote": [], + "bbox": [ + 135, + 89, + 841, + 333 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "with a unified framework and customized design for holistic 3D scene understanding from ego-centric views.", + "bbox": [ + 76, + 390, + 468, + 420 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Framework Overview. The framework includes a multimodal 3D encoder to extract object & scene representations and sparse & dense decoders for various downstream tasks. In addition, we customize the output's parameterization and training objectives to fit the formulation of oriented 3D bounding boxes in the sparse decoder.", + "bbox": [ + 76, + 421, + 468, + 511 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Multi-Modal 3D Encoder", + "text_level": 1, + "bbox": [ + 76, + 517, + 308, + 532 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As shown in Fig. 4, the multi-modal 3D encoder first has separate encoders for different modalities - ResNet50 [17] and FPN [27] (optional) for 2D images, Minkowski ResNet34 [11] for point clouds, and BERT [14] for texts. After extracting these features, we further fuse and process them into sparse or dense features for different downstream tasks. Next, we first present how we aggregate multi-view inputs and then introduce different fusion approaches for dense and sparse feature extraction.", + "bbox": [ + 75, + 537, + 468, + 672 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Scalability for Input Views. Contrasting with prior works, our framework can accept any number of RGB-D views, making it adaptable and generalizable to varying input orders and quantities. We conveniently aggregate different depth map views by transforming the point clouds into a global coordinate system, downsampling as needed. For multiple images, we query corresponding 2D features using perspective projection from 3D points, averaging them to maintain permutation invariance. This technique allows consistent feature updates during ego-centric exploration. Theoretically, voxel features could be updated by merging the volume feature at frame $t$ with the incremental feature from RGB-D input at frame $t + 1$ . In practice, we accommodate any number of views as batch-wise samples for accelerating training and evaluation. Our model demonstrates no", + "bbox": [ + 75, + 674, + 470, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "table scalability, where fewer views (e.g., 20) may be used for memory efficiency during training, while more views (e.g., 50) can enhance performance during inference.", + "bbox": [ + 496, + 390, + 890, + 436 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Dense Fusion. Previous works typically integrate the color and coordinates of points at the input stage, like \"painting\" [50], or form multi-modality dense BEV features for concatenated fusion [26, 30]. The latter way suits our occupancy prediction baseline and thus we adopt the straightforward dense fusion on the pre-defined grid, which shares the same resolution with the ground truth. We construct feature volume by projecting grid points onto a single 2D feature map post-FPN, then consolidate it with 3D features densified from sparse voxel features. For object detection, we argue that concurrent multi-modality fusion across multiple feature levels is more effective.", + "bbox": [ + 496, + 439, + 892, + 619 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Isomorphic Multi-Level Multi-Modality Fusion. Formally, the input aggregated points $P \\in \\mathcal{R}^{N_p \\times 3}$ (first voxelized) and $N_i$ images as $I \\in \\mathcal{R}^{N_i \\times H \\times W}$ are processed via a Minkowski ResNet and a shared 2D ResNet respectively. This extracts multi-level sparse voxel features $V_k \\in \\mathcal{R}^{C_k \\times N_{V_k}}$ on $K$ levels and image features $F_s \\in \\mathcal{R}^{C_s \\times H_s \\times W_s}$ on $S$ levels. In practice, these two ResNets produce 4 levels of features, for both point clouds and images, denoted as isomorphic multi-modality encoders.", + "bbox": [ + 496, + 625, + 890, + 761 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In dense fusion, we filter $F_{s}$ with an upsampling FPN to derive a feature map $F_{up}$ with stride $= 4$ and use it to construct the feature volume for fusing with $V_{4}$ . For the sparse case, we use multi-level features as seeds instead of a single dense feature map to predict 3D objects. The initial attempt of still query features from $F_{up}$ or raw images $I$ for these seeds is unstable due to inconsistent features for fusion and confusing gradients back-propagation. Thus, we leverage the isomorphic architecture for level-based projec", + "bbox": [ + 496, + 763, + 892, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "19761", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/96e4746e2ddf03338c208d0e23107d66ac299cb495a7a1b532dda8d1932d859d.jpg", + "table_caption": [ + "Table 2. Continuous and multi-view 3D object detection benchmark on EmbodiedScan (split by the double line)." + ], + "table_footnote": [], + "table_body": "
MethodsInputLarge-VocabularyHeadCommonTail
AP25AR25AP50AR50AP25AR25AP25AR25AP25AR25
Camera-OnlyRGB12.8034.614.2513.0717.4044.797.6424.220.033.09
Depth-OnlyDepth17.1651.4010.5225.7521.3961.1413.2741.582.7420.91
Multi-ModalityRGB-D19.0751.5611.5728.1523.5460.2315.8044.991.2417.74
ImVoxelNet [41]RGB6.1520.392.416.3110.9634.294.1215.402.639.21
VoteNet [34]Depth3.206.110.381.226.3112.261.813.341.001.83
FCAF3D [40]Depth9.0744.234.1120.2216.5461.386.7342.772.6724.83
+our decoderDepth14.8051.188.7727.4625.9867.1210.8550.085.7232.85
+paintingRGB-D15.1051.328.6426.6626.2367.5311.3950.645.8032.13
OursRGB-D16.8551.079.7728.2128.6567.5112.8350.467.0931.52
", + "bbox": [ + 119, + 109, + 846, + 229 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "tion and feature fusion, i.e., $V_{k}$ queries the corresponding image features of $F_{k}$ , which empirically shows a better and more stable performance. This method enables multi-level multi-modality feature fusion compared to the \"painting\" approach and ensures the consistency of features and gradients across different network levels and modalities.", + "bbox": [ + 75, + 239, + 468, + 328 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Vision-Language (VL) Fusion. Given the multi-level sparse visual features $F_{k}^{S}$ and text features from the text encoder, we use a multi-modal fusion transformer model [20, 61] for vision-language information interactions. Each transformer layer uses a self-attention block to refine sparse visual features and exploit spatial relationships. Then visual and text features interact in cross-modal attention blocks. This interaction guides updated sparse grounding features $F^{G}$ to be context-aware for subsequent prediction.", + "bbox": [ + 75, + 329, + 470, + 465 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Sparse & Dense Decoder", + "text_level": 1, + "bbox": [ + 76, + 474, + 303, + 489 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Given multi-modal features from typical encoders, we employ separate fusion streams for sparse and dense tasks. This results in four levels of sparse voxel features $F_{k}^{S}$ from isomorphic sparse fusion and a single dense feature $F^{D}$ for decoding and predictions. These are then processed to obtain 3D box and occupancy predictions.", + "bbox": [ + 75, + 497, + 468, + 587 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Sparse Decoder for 3D Boxes Prediction. Using the multi-level fused features $F_{k}^{S}$ , we upsample them as in FCAF3D, appending classification, regression, and centerness prediction heads for 3D object detection. In particular, to fit the oriented 3D box output, we add a 6D rotation representation [60] into original regression targets, ultimately decoded as 3D centers c, 3D sizes l, and Euler angles Θ. Training objectives include the original classification loss, centerness loss, and a disentangled Chamfer Distance (CD) loss for eight corners [3, 44]. Specifically, we use one of three groups of decoded predictions, {3D centers, 3D sizes, and Euler angles}, while setting the other two with ground truths to compute three corner losses. For example, given 3D sizes and Euler angles ground truth, we can derive the corner loss between the predicted B and the ground truth box B yielded by 3D center prediction errors:", + "bbox": [ + 75, + 588, + 470, + 829 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nL _ {\\mathbf {c}} = L _ {C D} (\\mathbf {B} (\\mathbf {c}, \\hat {\\mathbf {l}}, \\hat {\\boldsymbol {\\Theta}}), \\hat {\\mathbf {B}}) \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 181, + 833, + 468, + 851 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Together with the corner loss derived by the overall predicted bounding boxes $L_{pred}$ , We balance these losses with preset weights and use them to replace the original box loss:", + "bbox": [ + 75, + 854, + 468, + 901 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nL _ {l o c} = \\lambda_ {\\mathbf {c}} L _ {\\mathbf {c}} + \\lambda_ {\\mathbf {l}} L _ {\\mathbf {l}} + \\lambda_ {\\boldsymbol {\\Theta}} L _ {\\boldsymbol {\\Theta}} + \\lambda_ {p r e d} L _ {p r e d} \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 544, + 248, + 890, + 265 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We set $\\lambda_{\\mathbf{c}} = \\lambda_{\\mathbf{l}} = \\lambda_{\\Theta} = 0.2$ and $\\lambda_{pred} = 0.4$ to highlight the importance of the overall prediction, which performs well empirically. The target assignment strategy and post-processing during inference also follow FCAF3D [40].", + "bbox": [ + 496, + 271, + 890, + 330 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Dense Decoder for Occupancy Prediction. With the dense feature $F^D$ , we use a 3D FPN [41] to aggregate multi-level features and produce multi-scale occupancy predictions. Since the task requires more powerful low-level features for fine detail understanding, predictions at each scale are thus supervised with decayed half weights from high to low resolution [5]. We use cross-entropy loss and scene-class affinity loss [55] for training. During inference, we only use the high-resolution output for prediction.", + "bbox": [ + 496, + 332, + 890, + 468 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Sparse Decoder for 3D Visual Grounding. Grounding features $F^G$ updated after each transformer layer are fed into the prediction heads sharing the same architecture as those used for 3D detection. All prediction head outputs in each layer are supervised during training for stability and improved performance. An additional contrastive loss aligns the visual feature with target text prompts, ensuring the features of a target text token are closer to corresponding visual features and further from other visual or text tokens.", + "bbox": [ + 496, + 469, + 890, + 604 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Benchmark", + "text_level": 1, + "bbox": [ + 500, + 622, + 624, + 638 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Our benchmark has three categories based on data samples: scene-based, view-based, and prompt-based. Scene-based benchmarks mean the samples are based on different scenes, covering continuous and multi-view perception. View-based benchmarks use ego-centric views for tasks like monocular 3D detection. Lastly, samples of 3D visual grounding are based on constructed language prompts. Detailed splits will be discussed in each benchmark.", + "bbox": [ + 496, + 648, + 890, + 768 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For metrics, we use the 3D IoU-based average precision (AP) with thresholds of 0.25 and 0.5 for 3D detection and visual grounding. We also provide average recall (AR) for reference. For occupancy prediction, we employ the mean Intersection of Union (mIoU) as a performance measure. Due to the space limitation, please refer to the appendix for implementation details of different baselines, and more quantitative and qualitative results including an \"in-the-wild\" evaluation demo.", + "bbox": [ + 496, + 770, + 890, + 904 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "19762", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/12da896c12782fda0da8dcbde1eff7c3a407658a1028bd8d12ba93eb48cd3777.jpg", + "table_caption": [ + "Table 3. Continuous and multi-view occupancy prediction benchmark (split by the double line). \"refri.\" means \"refrigerator\"." + ], + "table_footnote": [], + "table_body": "
MethodsInputmIOUemptyfloorwallchaircabinetdoortablecouchshelfwindowbedcurtainrefri.plantstairstoilet
Camera-OnlyRGB10.4339.0934.1030.2426.469.4925.5341.6035.1916.2220.4524.4619.8026.0117.261.8329.78
Depth-OnlyDepth14.4473.9166.2256.1349.9615.7024.3756.8455.3530.5526.6642.8130.8133.0121.716.2145.35
Multi-ModalityRGB-D20.7973.5063.6462.3054.6019.9648.9961.1069.7639.8634.6254.8354.4548.9041.227.9763.52
OccNet [43]RGB8.0737.1546.9025.6320.9413.1718.4026.8122.8613.5913.4926.7522.9217.1517.074.7733.60
SurroundOcc [55]RGB9.1038.5446.1723.5523.0413.6019.1527.7922.8813.1113.7224.3218.8913.5814.777.8334.71
Camera-OnlyRGB10.4840.4541.2527.1926.1615.5020.3030.8226.7015.0114.3329.1723.3016.9915.986.1742.57
Depth-OnlyDepth15.5669.9260.5251.7449.4423.0824.3345.7743.5229.7423.0239.0441.2217.4219.5825.7960.45
Multi-ModalityRGB-D19.9771.2164.9255.0052.0427.3533.9747.9346.2631.8727.9846.5846.5624.0539.0124.4067.79
", + "bbox": [ + 80, + 106, + 888, + 205 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/2daeefa63262ca3e272f3adc8a0cf2a2ea2f61bbd314b9167e1eb400d7950398.jpg", + "table_caption": [ + "Table 4. Monocular 3D object detection benchmark on EmbodiedScan." + ], + "table_footnote": [], + "table_body": "
MethodsInput20 Common Classeschaircabinettablebincouchbedbathtubtoilet
AP25AR25AP50AR50
FCOS3D [52]RGB8.9327.960.915.0027.151.146.2110.239.4718.386.3140.51
ImVoxelNet [41]RGB18.9552.741.817.1046.704.6318.1017.8220.3941.5110.1465.70
VoteNet [34]Depth14.3031.441.685.1454.002.4119.5314.7221.8045.5813.4968.16
ImVoteNet [35]RGB-D19.6334.323.888.8256.722.8829.0021.9627.7756.9437.5674.08
FCAF3D [40]Depth25.7078.535.7320.2665.916.4726.6434.9322.5053.6826.3871.90
+our decoderDepth28.1684.504.9220.6963.856.6232.3438.9631.6160.3338.1775.57
+paintingRGB-D30.1983.935.7421.9066.397.4133.6642.8632.2460.0441.3177.59
OursRGB-D34.2885.0312.6132.2569.4710.0137.2945.1731.6763.2750.6380.39
", + "bbox": [ + 94, + 227, + 870, + 333 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.1. Fundamental 3D Perception Benchmarks", + "text_level": 1, + "bbox": [ + 76, + 342, + 429, + 358 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Continuous 3D Perception. As opposed to driving scenarios, indoor scene understanding is typically in an enclosed space, making it important to fully leverage multi-view cues formed by RGB-D sequence and continuously maintain an overall scene-level representation. Thus, we design this new benchmark involving sequential views for perceiving covered 3D regions. Models are trained and evaluated scene-wise with 3930/703/552 scans allocated for training/validation/testing. To accelerate the training and evaluation, we construct $N$ data samples with $1 \\sim N$ views from $N$ sampled views per scan. Here, $N = 10$ during training with random view sampling, while in evaluation, $N = 50$ with fixed views. Corresponding instances and occupancy truths are obtained by combining pre-computed visible instance IDs and occupancy masks of selected views. If a category lacks instances, it is removed when calculating mAP and mIoU. Given this new setup, we primarily offer three baselines with different input modalities (Tab. 2 and 3).", + "bbox": [ + 75, + 364, + 468, + 636 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Continuous 3D Object Detection. As anticipated, both RGB and depth features significantly impact this task, leading to superior results of our RGB-D approach (Tab. 2). The performance of the depth-only model closely mirrors the multimodality approach, indicating depth's dominance in 3D perception. Our method of constructing multi-modal features based on sparse voxel features also aligns with this intuition. Low performance on tail categories suggests dataset size influences performance, warranting future enhancement.", + "bbox": [ + 75, + 640, + 468, + 775 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Continuous Semantic Occupancy Prediction. This benchmark offers comprehensive results including mIoU and IoU for common classes. Unlike the detection benchmark, there is a notable gap between the depth-only and RGB-D baseline. This might be due to the former's limited semantic understanding capability, especially evident in categories like door and curtain, which are similar to walls in shape. On such a task that requires more fine-grained understanding,", + "bbox": [ + 75, + 779, + 470, + 901 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/984d3a1af9b88e4bde83a69374418b06db0f86322102738c657daa7e43a743a3.jpg", + "table_caption": [ + "Table 5. Multi-view 3D visual grounding benchmark. \"Indep/Dep\" refer to \"View-Independent/Dependent\". Easy/Hard and Indep/Dep have a ratio of $80\\% /20\\%$ and $78\\% /22\\%$ ." + ], + "table_footnote": [], + "table_body": "
MethodsInputOverallEasyHardIndepDep
AP25AP25AP25AP25AP25
ScanRefer [8]RGB-D12.8513.789.1213.4410.77
BUTD-DETR [20]RGB-D22.1423.1218.2322.4720.98
L3Det [61]RGB-D23.0724.0118.3423.5921.22
OursRGB-D25.7227.1120.1226.3723.42
", + "bbox": [ + 501, + 388, + 906, + 454 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "the depth sensor's weakness is enlarged. Meanwhile, depth plays a crucial role in predicting empty space, floor, and wall, while RGB information substantially improves prediction for most categories.", + "bbox": [ + 496, + 460, + 890, + 521 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Multi-View 3D Perception. Unlike continuous settings, multi-view 3D perception does not predefine the order of views but provides all views to the model for scene-level results. This setting was studied previously [41], so we first reproduce common methods on our benchmark.", + "bbox": [ + 496, + 522, + 890, + 597 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Multi-View 3D Object Detection. We implement baselines including ImVoxelNet [41] with RGB-only input and VoteNet [34] and FCAF3D [40] with depth-only input (Tab. 2). Additional dimensions are added to predict Euler angles with a simple L1 loss on their cosine values, but it yields overwhelming results. Substituting this with our decoder design markedly improves performance. Further using point cloud input painted for FCAF3D with RGB-D input slightly underperforms our baseline. Nevertheless, all models have substantial potential for improvement, demonstrating the challenges of this new dataset and setup.", + "bbox": [ + 496, + 598, + 890, + 763 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Multi-View Semantic Occupancy Prediction. We implement two popular baselines from autonomous driving benchmarks, OccNet [43] and SurroundOcc [55] (similar to TPV-Former [19]). Their performance slightly lags behind our camera-only baseline. Variants of our baselines exhibit a performance trend akin to embodied benchmarks.", + "bbox": [ + 496, + 763, + 890, + 854 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Monocular 3D Perception. Finally, the basic ego-centric setting is monocular 3D perception, specifically 3D detection, where each data sample comprises a single RGB-D", + "bbox": [ + 496, + 854, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "19763", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/ad9063d0ac94eb16e13005cfc28ebf740f3e68e494fd8a61e14e650f174e6e19.jpg", + "table_caption": [ + "Table 6. Ablation with conventional settings." + ], + "table_footnote": [], + "table_body": "
OrientedMulti-ViewAP25AR25AP50AR50
XX70.1790.4654.5875.66
✓X61.8790.3147.3073.93
✓✓59.9587.9243.3369.95
", + "bbox": [ + 107, + 107, + 433, + 152 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/84d0e86953a09da5a112a8d1256ac16b08e03037868cdd63d0128bac978e562a.jpg", + "table_caption": [ + "Table 7. Real vs. rendered images on ScanNet." + ], + "table_footnote": [], + "table_body": "
TrainValOverallHeadCommonTail
RenderRender22.1133.0116.446.74
RenderReal18.7227.0214.856.25
RealReal21.9832.9117.185.05
", + "bbox": [ + 114, + 174, + 426, + 219 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "frame and corresponding visible 3D boxes. Scan splits are used to extract frames as data samples, resulting in $689\\mathrm{k} / 115\\mathrm{k} / 86\\mathrm{k}$ images for the training/validation/testing.", + "bbox": [ + 75, + 226, + 467, + 270 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Monocular 3D Object Detection. This is more challenging than multi-view due to the absence of stereo geometric cues and truncated object views in indoor scenes, so the performance is significantly reduced in large-vocabulary settings. Hence, we first create a benchmark for 20 common categories (Tab. 4), observing a larger AP-AR gap for top methods because of difficulties predicting accurate 3D boxes from partial views. Similarly, our method outperforms others, providing a solid baseline for future studies.", + "bbox": [ + 75, + 271, + 467, + 406 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.2. Language-Grounded Benchmark", + "text_level": 1, + "bbox": [ + 76, + 414, + 367, + 429 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Multi-View 3D Visual Grounding. Our benchmark introduces language into the perception loop to foster interactive 3D scene representation learning. With comprehensive instance annotations, our benchmark presents more complex prompts and grounding cases than previous works. As an initial step, this setup takes multi-view RGB-D images as input without considering differing prompt timestamps. The goal is to ground the object described by the language prompt in the scene using information from different egocentric views. Ground-truth detection boxes are not provided as candidates for grounding during evaluation, which can better validate end-to-end 3D visual grounding ability than the original SR3D [1]. Data sample splits align with previous benchmarks' 3D scan splits.", + "bbox": [ + 75, + 434, + 467, + 645 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We reimplement classic methods like ScanRefer [8], BUTD-DETR [20], and L3Det [61] (Tab. 5). Our baseline outperforms all due to the strong multi-modal encoder. However, the performance remains much lower than previous works, partly due to changes in input format and annotations, which we will analyze next. Further challenges arise from handling more categories and small objects, making the grounding task more complex in parsing input prompts and predictions. Addressing these new challenges in this classical task would be promising for future research.", + "bbox": [ + 75, + 646, + 467, + 796 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.3. Analysis", + "text_level": 1, + "bbox": [ + 76, + 804, + 179, + 819 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Finally, we make further analysis to connect EmbodiedScan to current progress in computer vision.", + "bbox": [ + 75, + 825, + 467, + 854 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Axis-aligned vs. Oriented Boxes. We start with the 18-class detection performance of FCAF3D on ScanNet (Tab. 6). First, we change the annotations to oriented 3D", + "bbox": [ + 76, + 854, + 467, + 900 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/e23737867c5a0a4bf9d8689637b9c8d393df91da5c477bcc4e4a2bf46028a6f4.jpg", + "table_caption": [ + "Table 8. Benefits from training with EmbodiedScan." + ], + "table_footnote": [], + "table_body": "
TrainValOverallHeadCommonTail
ScanNetScanNet20.2829.8115.576.40
OursScanNet23.0233.8218.096.57
ScanNetOurs10.9221.108.061.78
OursOurs16.8528.6512.837.09
", + "bbox": [ + 532, + 109, + 854, + 164 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "boxes and adapt with our decoder. We find a significant drop in performance, indicating that the orientation estimation makes this task more challenging. We need to explore a better method to represent and predict the object pose.", + "bbox": [ + 496, + 167, + 890, + 227 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Reconstructed Point Cloud vs. Multi-View RGB-D. Subsequently, replacing the reconstructed point clouds with the aggregated ones from multi-view depth maps has minor effects on $AP_{25}$ but heavily impacts $AP_{50}$ (Tab. 6), implying that the accuracy of reconstructed point clouds is superior to raw depth maps. Therefore, integrating reconstruction techniques in perception loops shows potential.", + "bbox": [ + 496, + 228, + 890, + 334 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Next, we study the gap between the real and rendered images, and the benefits from training with EmbodiedScan. We do the comparison with our multi-modality baseline on the large-vocabulary multi-view 3D detection benchmark.", + "bbox": [ + 496, + 335, + 890, + 393 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Real Capture vs. Rendering. As shown in Tab. 7, apart from the significant visual difference between real and rendered images (Fig. 1), the model's performance also has a remarkable decrease when transferring models trained with rendered images to the real world, particularly when the annotations are sufficient (5.99% AP drop on head categories and 5.89% AP lower than models trained with real images.)", + "bbox": [ + 496, + 396, + 890, + 501 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Benefits from EmbodiedScan. Finally, we also quantitatively evaluate the benefits of training models with our large-scale EmbodiedScan (Tab. 8). As expected, when training our models with EmbodiedScan, we observed a significant improvement in both ScanNet (2.74% AP) and the overall validation split (5.93% AP), particularly 4.01% AP and 7.55% AP increase on head categories.", + "bbox": [ + 496, + 502, + 890, + 607 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 618, + 617, + 632 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This paper introduces EmbodiedScan, a multi-modal perception suite aiming for language-grounded holistic 3D scene understanding from ego-centric views. We construct a large-scale dataset with diverse sensor data and multimodal annotations, including 3D oriented boxes, semantic occupancy and language descriptions. Based on this dataset, we propose a baseline framework capable of handling any number of views input, using a unified multimodal encoder and task-specific decoders. We establish benchmarks for basic and language-grounded 3D perception. Experiment results highlight our work's value and reveal new challenges in this setup. We believe EmbodiedScan can bring opportunities in embodied 3D perception and may also have a broader impact in related fields with the massive data and rich annotations.", + "bbox": [ + 496, + 640, + 890, + 866 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements. This work is supported by Shanghai Artificial Intelligence Laboratory.", + "bbox": [ + 500, + 869, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "19764", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 174, + 106 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Panos Achlioptas, Ahmed Abdelreehm, Fei Xia, Mohamed Elhoseiny, and Leonidas Guibas. Referit3d: Neural listeners for fine-grained 3d object identification in real-world scenes. In European conference on computer vision, 2020. 3, 4, 8", + "[2] Gilad Baruch, Zhuoyuan Chen, Afshin Dehghan, Tal Dimry, Yuri Feigin, Peter Fu, Thomas Gebauer, Brandon Joffe, Daniel Kurz, Arik Schwartz, and Elad Shulman. ARK-scenes - a diverse real-world dataset for 3d indoor scene understanding using mobile RGB-d data. In Thirty-fifth Conference on Neural Information Processing Systems Datasets and Benchmarks Track (Round 1), 2021. 2, 3, 4", + "[3] Garrick Brazil, Abhinav Kumar, Julian Straub, Nikhila Ravi, Justin Johnson, and Georgia Gkioxari. Omni3D: A large benchmark and model for 3D object detection in the wild. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2023. 2, 4, 6", + "[4] Holger Caesar, Varun Bankiti, Alex H. Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuscenes: A multimodal dataset for autonomous driving. CoRR, abs/1903.11027, 2019. 2", + "[5] Anh-Quan Cao and Raoul de Charette. Monoscene: Monocular 3d semantic scene completion. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2022. 3, 6", + "[6] Angel Chang, Angela Dai, Thomas Funkhouser, Maciej Halber, Matthias Niessner, Manolis Savva, Shuran Song, Andy Zeng, and Yinda Zhang. Matterport3D: Learning from RGB-D data in indoor environments. International Conference on 3D Vision (3DV), 2017. 1, 2, 3", + "[7] Ming-Fang Chang, John Lambert, Patsorn Sangkloy, Jagjeet Singh, Slawomir Bak, Andrew Hartnett, De Wang, Peter Carr, Simon Lucey, Deva Ramanan, et al. Argoverse: 3d tracking and forecasting with rich maps. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2019. 2", + "[8] Dave Zhenyu Chen, Angel X Chang, and Matthias Nießner. Scanrefer: 3d object localization in rgb-d scans using natural language. In European conference on computer vision, 2020. 2, 3, 7, 8", + "[9] Zhenyu Chen, Ali Gholami, Matthias Nießner, and Angel X Chang. Scan2cap: Context-aware dense captioning in rgb-d scans. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3193-3203, 2021. 3", + "[10] Zhenyu Chen, Ronghang Hu, Xinlei Chen, Matthias Nießner, and Angel X Chang. Unit3d: A unified transformer for 3d dense captioning and visual grounding. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2023. 3", + "[11] Christopher Choy, JunYoung Gwak, and Silvio Savarese. 4d spatio-temporal convnets: Minkowski convolutional neural networks. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2019. 5", + "[12] Camille Couprie, Clément Farabet, Laurent Najman, and" + ], + "bbox": [ + 78, + 114, + 470, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Yann LeCun. Indoor semantic segmentation using depth information. arXiv preprint arXiv:1301.3572, 2013. 1, 2", + "[13] Angela Dai, Angel X Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2017. 1, 2, 3", + "[14] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018. 5", + "[15] Runyu Ding, Jihan Yang, Chuhui Xue, Wenqing Zhang, Song Bai, and Xiaojuan Qi. Pla: Language-driven open-vocabulary 3d scene understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023. 3", + "[16] Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for autonomous driving? the kitti vision benchmark suite. In IEEE Conference on Computer Vision and Pattern Recognition, 2012. 2", + "[17] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In IEEE Conference on Computer Vision and Pattern Recognition, 2016. 5", + "[18] Shijia Huang, Yilun Chen, Jiaya Jia, and Liwei Wang. Multiview transformer for 3d visual grounding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022. 3", + "[19] Yuanhui Huang, Wenzhao Zheng, Yunpeng Zhang, Jie Zhou, and Jiwen Lu. Tri-perspective view for vision-based 3d semantic occupancy prediction. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2023. 3, 7", + "[20] Ayush Jain, Nikolaos Gkanatsios, Ishita Mediratta, and Katerina Fragkiadaki. Bottom up top down detection transformers for language grounding in images and point clouds. In European Conference on Computer Vision, 2022. 3, 6, 7, 8", + "[21] Mukul Khanna*, Yongsen Mao*, Hanxiao Jiang, Sanjay Haresh, Brennan Shacklett, Dhruv Batra, Alexander Clegg, Eric Undersander, Angel X. Chang, and Manolis Savva. Habitat Synthetic Scenes Dataset (HSSD-200): An Analysis of 3D Scene Scale and Realism Tradeoffs for ObjectGoal Navigation. arXiv preprint, 2023. 2", + "[22] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C. Berg, Wan-Yen Lo, Piotr Dollar, and Ross Girshick. Segment anything. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2023. 2, 4", + "[23] Alex H. Lang, Sourabh Vora, Holger Caesar, Lubing Zhou, Jiong Yang, and Oscar Beijbom. Pointpillars: Fast encoders for object detection from point clouds. In IEEE Conference on Computer Vision and Pattern Recognition, 2019. 3", + "[24] E Li, Shuaijun Wang, Chengyang Li, Dachuan Li, Xiangbin Wu, and Qi Hao. Sustech points: A portable 3d point cloud interactive annotation platform system. In 2020 IEEE Intelligent Vehicles Symposium (IV), 2020. 4" + ], + "bbox": [ + 501, + 92, + 893, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "19765", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[25] Yiming Li, Zhiding Yu, Christopher Choy, Chaowei Xiao, Jose M Alvarez, Sanja Fidler, Chen Feng, and Anima Anandkumar. Voxformer: Sparse voxel transformer for camera-based 3d semantic scene completion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023. 3", + "[26] Tingting Liang, Hongwei Xie, Kaicheng Yu, Zhongyu Xia, Zhiwei Lin, Yongtao Wang, Tao Tang, Bing Wang, and Zhi Tang. Bevfusion: A simple and robust lidar-camera fusion framework. Advances in Neural Information Processing Systems, 2022. 3, 5", + "[27] Tsung-Yi Lin, Piotr Dólar, Ross Girshick, Kaiming He, Bharath Hariharan, and Serge Belongie. Feature pyramid networks for object detection. In IEEE Conference on Computer Vision and Pattern Recognition, 2017. 5", + "[28] Zechen Liu, Zizhang Wu, and Roland Toth. Smoke: Single-stage monocular 3d object detection via keypoint estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, pages 996-997, 2020. 3", + "[29] Ze Liu, Zheng Zhang, Yue Cao, Han Hu, and Xin Tong. Group-free 3d object detection via transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2021. 3", + "[30] Zhijian Liu, Haotian Tang, Alexander Amini, Xinyu Yang, Huizi Mao, Daniela L Rus, and Song Han. Bevfusion: Multi-task multi-sensor fusion with unified bird's-eye view representation. In 2023 IEEE International Conference on Robotics and Automation (ICRA), 2023. 3, 5", + "[31] Yuheng Lu, Chenfeng Xu, Xiaobao Wei, Xiaodong Xie, Masayoshi Tomizuka, Kurt Keutzer, and Shanghang Zhang. Open-vocabulary 3d detection via image-level class and de-biased cross-modal contrastive learning. arXiv preprint arXiv:2207.01987, 2022. 3", + "[32] Jiageng Mao, Minzhe Niu, Chenhan Jiang, Hanxue Liang, Jingheng Chen, Xiaodan Liang, Yamin Li, Chaoqiang Ye, Wei Zhang, Zhenguo Li, et al. One million scenes for autonomous driving: Once dataset. arXiv preprint arXiv:2106.11037, 2021. 2", + "[33] Songyou Peng, Kyle Genova, Chiyu Jiang, Andrea Tagliasacchi, Marc Pollefeys, Thomas Funkhouser, et al. Openscene: 3d scene understanding with open vocabularies. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023. 3", + "[34] Charles R Qi, Or Litany, Kaiming He, and Leonidas J Guibas. Deep hough voting for 3d object detection in point clouds. In proceedings of the IEEE/CVF International Conference on Computer Vision, 2019. 1, 3, 4, 6, 7", + "[35] Charles R Qi, Xinlei Chen, Or Litany, and Leonidas J Guibas. Imvotenet: Boosting 3d object detection in point clouds with image votes. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2020. 3, 7", + "[36] Santhosh K Ramakrishnan, Aaron Gokaslan, Erik Wijmans, Oleksandr Maksymets, Alex Clegg, John Turner, Eric Undersander, Wojciech Galuba, Andrew Westbury, Angel X Chang, et al. Habitat-matterport 3d dataset (hm3d): 1000" + ], + "bbox": [ + 78, + 90, + 470, + 901 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "large-scale 3d environments for embodied ai. arXiv preprint arXiv:2109.08238, 2021. 2", + "[37] Nils Reimers and Iryna Gurevych. Sentence-bert: Sentence embeddings using siamese bert-networks. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing, 2019. 4", + "[38] Mike Roberts, Jason Ramapuram, Anurag Ranjan, Atulit Kumar, Miguel Angel Bautista, Nathan Paczan, Russ Webb, and Joshua M Susskind. Hypersim: A photorealistic synthetic dataset for holistic indoor scene understanding. In Proceedings of the IEEE/CVF international conference on computer vision, 2021. 2", + "[39] David Rozenberszki, Or Litany, and Angela Dai. Language-grounded indoor 3d semantic segmentation in the wild. In Proceedings of the European Conference on Computer Vision (ECCV), 2022. 2", + "[40] Danila Rukhovich, Anna Vorontsova, and Anton Konushin. Fcaf3d: Fully convolutional anchor-free 3d object detection. In European Conference on Computer Vision, 2022. 1, 3, 6, 7", + "[41] Danila Rukhovich, Anna Vorontsova, and Anton Konushin. Imvoxelnet: Image to voxels projection for monocular and multi-view general-purpose 3d object detection. In WACV, pages 2397-2406, 2022. 3, 6, 7", + "[42] Shaoshuai Shi, Xiaogang Wang, and Hongsheng Li. Poincaré: 3d object proposal generation and detection from point cloud. In IEEE Conference on Computer Vision and Pattern Recognition, 2019. 3", + "[43] Chonghao Sima, Wenwen Tong, Tai Wang, Li Chen, Silei Wu, Hanming Deng, Yi Gu, Lewei Lu, Ping Luo, Dahua Lin, and Hongyang Li. Scene as occupancy. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2023. 3, 7", + "[44] Andrea Simonelli, Samuel Rota Rota Bulò, Lorenzo Porzi, Manuel López-Antequera, and Peter Kontschieder. Disentangling monocular 3d object detection. In IEEE International Conference on Computer Vision, 2019. 3, 6", + "[45] Shuran Song, Samuel P Lichtenberg, and Jianxiong Xiao. Sun rgb-d: A rgb-d scene understanding benchmark suite. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2015. 1, 2", + "[46] Julian Straub, Thomas Whelan, Lingni Ma, Yufan Chen, Erik Wijmans, Simon Green, Jakob J Engel, Raul Mur-Artal, Carl Ren, Shobhit Verma, et al. The replica dataset: A digital replica of indoor spaces. arXiv preprint arXiv:1906.05797, 2019. 2", + "[47] Pei Sun, Henrik Kretzschmar, Xerxes Dotiwalla, Aurelien Chouard, Vijaysai Patnaik, Paul Tsui, James Guo, Yin Zhou, Yuning Chai, Benjamin Caine, et al. Scalability in perception for autonomous driving: Waymo open dataset. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2020. 2", + "[48] Ayca Takmaz, Elisabetta Fedele, Robert W Sumner, Marc Pollefeys, Federico Tombari, and Francis Engelmann. Openmask3d: Open-vocabulary 3d instance segmentation. arXiv preprint arXiv:2306.13631, 2023. 3" + ], + "bbox": [ + 501, + 92, + 893, + 901 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "19766", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[49] Xiaoyu Tian, Tao Jiang, Longfei Yun, Yue Wang, Yilun Wang, and Hang Zhao. Occ3d: A large-scale 3d occupancy prediction benchmark for autonomous driving. arXiv preprint arXiv:2304.14365, 2023. 3", + "[50] Sourabh Vora, Alex H Lang, Bassam Helou, and Oscar Beijbom. Pointpainting: Sequential fusion for 3d object detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2020. 3, 5", + "[51] Johanna Wald, Armen Avetisyan, Nassir Navab, Federico Tombari, and Matthias Nießner. Rio: 3d object instance re-localization in changing indoor environments. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2019. 1, 2, 3", + "[52] Tai Wang, Xinge Zhu, Jiangmiao Pang, and Dahua Lin. FCOS3D: Fully convolutional one-stage monocular 3d object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) Workshops, 2021. 3, 7", + "[53] Tai Wang, Jiangmiao Pang, and Dahua Lin. Monocular 3d object detection with depth from motion. In European Conference on Computer Vision (ECCV), 2022.", + "[54] Tai Wang, ZHU Xinge, Jiangmiao Pang, and Dahua Lin. Probabilistic and geometric depth: Detecting objects in perspective. In Conference on Robot Learning, pages 1475-1485. PMLR, 2022. 3", + "[55] Yi Wei, Linqing Zhao, Wenzhao Zheng, Zheng Zhu, Jie Zhou, and Jiwen Lu. Surroundocc: Multi-camera 3d occupancy prediction for autonomous driving. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2023. 3, 6, 7", + "[56] Karmesh Yadav, Ram Ramrakhya, Santhosh Kumar Ramakrishnan, Theo Gervet, John Turner, Aaron Gokaslan, Noah Maestre, Angel Xuan Chang, Dhruv Batra, Manolis Savva, et al. Habitat-matterport 3d semantics dataset. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023. 2", + "[57] Yan Yan, Yuxing Mao, and Bo Li. Second: Sparsely embedded convolutional detection. Sensors, 18(10), 2018. 3", + "[58] Wenwei Zhang, Zhe Wang, and Chen Change Loy. Exploring data augmentation for multi-modality 3d object detection. arXiv preprint arXiv:2012.12741, 2020. 3", + "[59] Yin Zhou and Oncel Tuzel. Voxelnet: End-to-end learning for point cloud based 3d object detection. In IEEE Conference on Computer Vision and Pattern Recognition, 2018. 3", + "[60] Yi Zhou, Connelly Barnes, Jingwan Lu, Jimei Yang, and Hao Li. On the continuity of rotation representations in neural networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2019. 6", + "[61] Chenming Zhu, Wenwei Zhang, Tai Wang, Xihui Liu, and Kai Chen. Object2scene: Putting objects in context for open-vocabulary 3d detection. arXiv preprint arXiv:2309.09456, 2023. 3, 6, 7, 8" + ], + "bbox": [ + 78, + 90, + 468, + 825 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "19767", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2024/EmbodiedScan_ A Holistic Multi-Modal 3D Perception Suite Towards Embodied AI/ab5e0a2a-2a5f-41d4-a08d-01fd4564b83b_model.json b/2024/EmbodiedScan_ A Holistic Multi-Modal 3D Perception Suite Towards Embodied AI/ab5e0a2a-2a5f-41d4-a08d-01fd4564b83b_model.json new file mode 100644 index 0000000000000000000000000000000000000000..56ca31159de09924f400301c708582897a936870 --- /dev/null +++ b/2024/EmbodiedScan_ A Holistic Multi-Modal 3D Perception Suite Towards Embodied AI/ab5e0a2a-2a5f-41d4-a08d-01fd4564b83b_model.json @@ -0,0 +1,2257 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.18, + 0.121, + 0.791, + 0.164 + ], + "angle": 0, + "content": "EmbodiedScan: A Holistic Multi-Modal 3D Perception Suite Towards Embodied AI" + }, + { + "type": "text", + "bbox": [ + 0.114, + 0.178, + 0.856, + 0.233 + ], + "angle": 0, + "content": "Tai Wang\\(^{1*}\\), Xiaohan Mao\\(^{1,2*}\\), Chenming Zhu\\(^{1,3*}\\), Runsen Xu\\(^{1,4}\\), Ruiyuan Lyu\\(^{1,5}\\), Peisen Li\\(^{1,5}\\), Xiao Chen\\(^{1,4}\\), Wenwei Zhang\\(^{1}\\), Kai Chen\\(^{1}\\), Tianfan Xue\\(^{4}\\), Xihui Liu\\(^{3}\\), Cewu Lu\\(^{2}\\), Dahua Lin\\(^{1,4}\\), Jiangmiao Pang\\(^{1\\boxtimes}\\)" + }, + { + "type": "text", + "bbox": [ + 0.166, + 0.24, + 0.788, + 0.293 + ], + "angle": 0, + "content": "\\(^{1}\\)Shanghai AI Laboratory \\(^{2}\\)Shanghai Jiao Tong University \\(^{3}\\)The University of Hong Kong \\(^{4}\\)The Chinese University of Hong Kong \\(^{5}\\)Tsinghua University *Equal contribution Corresponding author" + }, + { + "type": "title", + "bbox": [ + 0.147, + 0.306, + 0.816, + 0.321 + ], + "angle": 0, + "content": "EmbodiedScan: 5k scans, 1M ego-centric images, 1M language prompts, 160k instances, 760+ categories" + }, + { + "type": "image", + "bbox": [ + 0.12, + 0.328, + 0.84, + 0.525 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.531, + 0.893, + 0.561 + ], + "angle": 0, + "content": "Figure 1. EmbodiedScan provides a multi-modal, ego-centric 3D perception dataset with massive real-scanned data and rich annotations for indoor scenes. It benchmarks language-grounded holistic 3D scene understanding capabilities for real-world embodied agents." + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.572, + 0.314, + 0.587 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.593, + 0.474, + 0.895 + ], + "angle": 0, + "content": "In the realm of computer vision and robotics, embodied agents are expected to explore their environment and carry out human instructions. This necessitates the ability to fully understand 3D scenes given their first-person observations and contextualize them into language for interaction. However, traditional research focuses more on scene-level input and output setups from a global view. To address the gap, we introduce EmbodiedScan, a multi-modal, ego-centric 3D perception dataset and benchmark for holistic 3D scene understanding. It encompasses over 5k scans encapsulating 1M ego-centric RGB-D views, 1M language prompts, 160k 3D-oriented boxes spanning over 760 categories, some of which partially align with LVIS, and dense semantic occupancy with 80 common categories. Building upon this database, we introduce a baseline framework named Embodied Perceptron. It is capable of processing an arbitrary number of multi-modal inputs and demonstrates remarkable 3D perception capabilities, both within the two series of benchmarks we set up, i.e., fundamental 3D perception tasks and language-grounded tasks, and in the wild." + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.572, + 0.631, + 0.587 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.598, + 0.892, + 0.749 + ], + "angle": 0, + "content": "Consider an embodied agent operating in an indoor environment. It commences its journey devoid of any prior knowledge about the scene, guided only by an initial instruction. As it begins to explore, it recognizes objects in context and acts with goals along with language interaction. In this process, a commonly needed, fundamental perception capability is to establish a holistic 3D scene understanding given ego-centric observations. This understanding operates at the scene level, covers both object semantics and scene geometry, and can be grounded in language descriptions." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.893, + 0.901 + ], + "angle": 0, + "content": "Nonetheless, subtle but significant discrepancies exist between this expectation and research problems examined within the computer vision community. Most previous studies have primarily revolved around scene-level input and output problems from a global view [13, 34, 40], i.e., taking reconstructed 3D point clouds or meshes as inputs and predicting 3D object bounding boxes or segmenting point clouds. Regarding data, earlier datasets targeting egocentric RGB-D inputs are either too small [12, 45] or lack comprehensive annotations [6, 51] to support the aforementioned results." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19757" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.079, + 0.087, + 0.627, + 0.205 + ], + "angle": 0, + "content": "
Dataset#Scans#ImgS#Objs#Cats#PromptsEgo Capture3D Annotations
Replica [46]35----XX
NYU v2 [12]4641.4k35k14-✓X
SUN RGB-D [45]-10k-37-Mono.Box
ScanNet [13, 39]1513264k36k1852k [8]✓Seg., Lang.
Matterport3D [6]2056194k51k40-Multi-ViewSeg.
3RScan [51]1482363k---✓Seg.
ArkitScenes [2]5047450k51k17-✓Box
HyperSim [38]46177k-40+-Mono. & Syn.Box
EmbodiedScan5185890k160k762970k✓Box, Occ., Lang.
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.215, + 0.62, + 0.273 + ], + "angle": 0, + "content": "Table 1. Comparison with other 3D indoor scene datasets. \"Cats\" refers to the categories with box annotations for the 3D detection benchmark. EmbodiedScan features more than \\(10 \\times\\) categories, prompts, and the most diverse annotations. The numbers are still scaling up with further annotations. Mono./Syn./Lang. means Monocular/Synthetic/Language." + }, + { + "type": "image", + "bbox": [ + 0.637, + 0.085, + 0.905, + 0.204 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.631, + 0.215, + 0.899, + 0.271 + ], + "angle": 0, + "content": "Figure 2. Dataset composition. Embodied-Scan is composed of three data sources and has similar scans, images, objects, and categories in each of them." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.282, + 0.47, + 0.357 + ], + "angle": 0, + "content": "tioned research. It is also not feasible to generate such realistic views by rendering from the existing imperfect meshes. On the other hand, since we cannot trivially obtain the reconstruction of a new environment, models trained with scene-level input are not directly applicable in practice." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.358, + 0.471, + 0.614 + ], + "angle": 0, + "content": "To bridge this divide, we introduce a multi-modal, ego-centric 3D perception dataset and benchmark for holistic 3D scene understanding, termed EmbodiedScan, aimed at facilitating real-world embodied AI applications (Fig. 1). This dataset exploits existing large-scale 3D scene datasets [6, 13, 51] but re-purposes them for continuous scene-level perception from the first-view RGB-D streams. Unlike previous works that offer only point segmentation labels with limited semantics, we employ a SAM-assisted [22] pipeline to annotate objects with oriented 3D bounding boxes and generate language prompts on top. Consequently, EmbodiedScan provides more than 5k scans, nearly 1M ego-centric RGB-D images, and multi-modality annotations, covering 3D oriented boxes with more than 160k instances spanning over 760 categories, dense semantic occupancy with 80 common categories, and 1M language descriptions focusing on spatial relationships among objects." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.615, + 0.47, + 0.795 + ], + "angle": 0, + "content": "Built upon this dataset, we devise a baseline framework for ego-centric 3D perception, Embodied Perceptron. It accepts RGB-D sequences and texts as inputs and manifests scalability to any number of views input with encoders shared across different tasks. With the encoded 2D and 3D features, we employ dense fusion and isomorphic multilevel fusion across them guided by the perspective projection to produce 3D scene and object representations, which are further processed to decode occupancy and 3D box predictions. The derived 3D representations can be further integrated with text embeddings for 3D visual grounding, thus supporting language-grounded applications." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.796, + 0.471, + 0.9 + ], + "angle": 0, + "content": "We establish two series of benchmarks on Embodied-Scan: 1) fundamental 3D perception benchmarks focusing on traditional tasks, including 3D detection and semantic occupancy prediction under different input settings, and 2) a language-grounded scene understanding benchmark with 3D visual grounding as a preliminary exploration. Experimental results validate the effectiveness of our baseline" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.282, + 0.893, + 0.343 + ], + "angle": 0, + "content": "model on EmbodiedScan and demonstrate its generalization ability in the wild. Detailed analysis further underscores the value of EmbodiedScan and highlights the primary challenges posed by this new setup." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.359, + 0.642, + 0.375 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.385, + 0.893, + 0.581 + ], + "angle": 0, + "content": "3D Scene Datasets. The development of 3D scene understanding has benefited from large-scale, high-quality datasets like KITTI [16] and SUN RGB-D [45]. These foundational datasets have paved the way for subsequent larger and more diverse collections targeting indoor [6, 13, 38, 51] and driving scenes [4, 7, 32, 47]. However, compared to autonomous driving datasets, those meant for indoor scenes still lack variety in terms of scenes and object diversity (Tab. 1). In contrast, EmbodiedScan provides a large amount of multi-modal data with much richer annotations. Furthermore, it differs by placing an emphasis on the ego-centric perspective within its setup, a feature often overlooked in previous works [2, 13]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.584, + 0.894, + 0.809 + ], + "angle": 0, + "content": "Except for these conventional dataset works, Omni3D [3] integrates urban and indoor datasets for monocular 3D detection. Our focus, however, lies in indoor scenes due to their unique challenges but has a larger amount of data and annotations, e.g., more than \\(3 \\times\\) images and categories with more than 10 instances. In addition, we offer a comprehensive exploration of more general problems for ego-centric 3D perception, such as continuous perception and visual grounding. Other embodied AI datasets like HM3D [36, 56] and HSSD [21] provide ample interaction opportunities but can suffer from poor transferability to real-world scenarios due to their imperfect meshes or synthetic data. Conversely, EmbodiedScan is based on real-scanned RGB-D images, offering a more realistic playground for model training." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.811, + 0.894, + 0.9 + ], + "angle": 0, + "content": "3D Object Detection & Occupancy Prediction. 3D detection and occupancy prediction, as fundamental tasks in 3D perception, focus on different aspects of 3D scene understanding. The former focuses on recognizing foreground objects through a sparse and efficient representation - a set of 3D cuboids corresponding to instances of interest -" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19758" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.084, + 0.086, + 0.473, + 0.208 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.119, + 0.215, + 0.439, + 0.228 + ], + "angle": 0, + "content": "(a) SAM-Assisted Oriented 3D Bounding Boxes Annotation." + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.234, + 0.476, + 0.345 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.155, + 0.348, + 0.404, + 0.359 + ], + "angle": 0, + "content": "(c) Instance Statistics (Increase w.r.t. ScanNet)." + }, + { + "type": "image", + "bbox": [ + 0.493, + 0.088, + 0.891, + 0.213 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.569, + 0.215, + 0.813, + 0.228 + ], + "angle": 0, + "content": "(b) 3D Boxes and Language Prompt Statistics." + }, + { + "type": "image", + "bbox": [ + 0.492, + 0.234, + 0.885, + 0.344 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.621, + 0.347, + 0.753, + 0.359 + ], + "angle": 0, + "content": "(d) Occupancy Statistics." + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.369, + 0.892, + 0.453 + ], + "angle": 0, + "content": "Figure 3. EmbodiedScan annotation and statistics. (a) UI for 3D box annotation. We select keyframes and generate their SAM masks with corresponding axis-aligned boxes. With simple clicks, annotators can create 3D boxes for target objects and further adjust them with reference in three orthogonal views and images. (b) Small boxes \\((< 1m^3)\\) increase more & prompt statistics. obj./avg./des. refer to objects/average/descriptions. (c) We show the number of instances per category (300 classes). For categories that exist in ScanNet, we plot the absolute increase and observe a significant improvement. (d) We plot the occupancy distribution for each category and see a different word cloud distribution. These two clouds show different aspects, occupied space vs. number of instances, of this dataset." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.462, + 0.469, + 0.627 + ], + "angle": 0, + "content": "while the latter offers a dense, structured pattern that benefits downstream planning. The research community has developed solutions ranging from single-modality, such as LiDAR-based [23, 29, 34, 40, 42, 57, 59] or camera-only approaches [28, 41, 44, 52-54], to multi-modality techniques [26, 30, 35, 50, 58]. Recently, occupancy as a representation, due to its potential in handling unknown semantics and irregular object shapes, has gained more attention [5, 19, 25, 43, 49, 55]. Given their distinctive focuses, we selected these two tasks to form the fundamental 3D perception track on EmbodiedScan." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.629, + 0.469, + 0.764 + ], + "angle": 0, + "content": "Previous works on indoor scenes mainly centered around 3D detection with limitations in object orientations, semantic categories, and input format [29, 34, 40]. In practice, a model is expected to perceive the environment during egocentric exploration, ultimately providing a holistic understanding inclusive of rich semantics, scene geometry, and object poses. To this end, EmbodiedScan and our proposed framework, Embodied Perceptron, provide this necessary data foundation and baseline methodology." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.469, + 0.902 + ], + "angle": 0, + "content": "Language-Grounded 3D Scene Understanding. Language plays a crucial role in human-computer interaction, heightened by recent advances in Large Language Models (LLMs). Its integration with 3D scene understanding is vital for future embodied agents. Past research first explored 3D visual grounding [1, 8, 18, 20] and established new benchmarks including 3D dense captioning [9, 10], open-vocabulary 3D segmentation [15, 33, 48] and detection [31, 61]. This paper focuses on 3D visual grounding" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.462, + 0.892, + 0.537 + ], + "angle": 0, + "content": "first, with plans to expand language annotations and benchmarks in the future. Our visual grounding benchmark aligns with the multi-view setting of the basic 3D perception track, taking multiple ego-centric RGB-D images as input, and includes tenfold more complex prompts." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.546, + 0.589, + 0.561 + ], + "angle": 0, + "content": "3. Dataset" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.568, + 0.891, + 0.599 + ], + "angle": 0, + "content": "This section presents the dataset construction, including data processing and annotation, and shows the statistics." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.606, + 0.761, + 0.623 + ], + "angle": 0, + "content": "3.1. Data Collection & Processing" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.629, + 0.892, + 0.779 + ], + "angle": 0, + "content": "Ego-Centric Sensor Data Collection. Considering there have been readily available 3D indoor scene scans from existing datasets, we start with integrating those providing ego-centric RGB-D captures with the corresponding camera poses. Given the compatibility of ScanNet [13], 3RScan [51], and Matterport3D [6], we select the high-quality part with complete regions and necessary annotations to form the initial version of EmbodiedScan (Fig. 2). ARKitScenes [2], possessing different data organization, sensors, and annotations, is considered for future inclusion." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.78, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Frame Selection & Scene Division. Although these datasets all have RGB-D data, the data format, sampling frequency, and relationships among viewpoints are different. See more details in the appendix. We first unified the format into a general multi-view case to fit Matterport3D by adding randomness when loading images but maintaining sequential continuity for ScanNet and 3RScan during inference. Our model can thus handle both temporal and" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19759" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.212 + ], + "angle": 0, + "content": "randomly captured multi-view images. Additionally, we divided building-scale scenes of Matterport3D into regions based on official annotation, selecting corresponding images with depth points falling into the region. As for different sampling rates of images in ScanNet and 3RScan videos, we sample one keyframe per 10 frames for ScanNet and keep all the images for 3RScan. The uniform sampling is generally in line with the actual situation." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.213, + 0.473, + 0.365 + ], + "angle": 0, + "content": "Global Coordinate System. A global coordinate system is necessary to aggregate multi-view observations and serve as a reference for outputs. We follow the convention of ScanNet, deriving a system with the origin around the center of the scene, the horizontal plane lying on the floor and axes aligning with walls [34]. This post-processing harmonizes the data distribution, slightly improving performance on benchmarks. Practical applications may not have such a prior global system or vary according to observations, posing another interesting problem for future exploration." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.376, + 0.203, + 0.391 + ], + "angle": 0, + "content": "3.2. Annotation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.4, + 0.47, + 0.446 + ], + "angle": 0, + "content": "We provide three types of annotations - 3D bounding boxes, semantic occupancy, and language descriptions - each serving to enrich different aspects of scene understanding." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.447, + 0.473, + 0.674 + ], + "angle": 0, + "content": "3D Bounding Boxes. Following standard definitions [2, 3], a cuboid is defined by its 3D center, size, and ZXY Euler angle orientation. We used the Segment Anything Model (SAM) [22] and a customized annotation tool based on [24] (Fig. 3a) to address limitations in existing 3D box annotations, i.e., lack of orientation and small object annotations. It supports the conventional functionality of annotating 3D boxes with orientation in three orthographic views. Furthermore, we sample several keyframes with clear imaging according to the camera pose changes and ensure they cover non-overlap regions and most objects to generate SAM masks and axis-aligned boxes for further adjustment. We work with an annotation team and check the quality of all the labels in the end. Each scene takes around 10-30 minutes to annotate, varying with the scene complexity." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.674, + 0.473, + 0.84 + ], + "angle": 0, + "content": "Semantic Occupancy. Semantic occupancy necessitates accurate boundaries across semantic regions without considering object pose or recalling all the objects, so the original point cloud segmentation annotations were more suitable to be used for deriving occupancy ground truth. For each voxel, we assigned the category with the most points as the semantic label for that cell. A compromise between perception granularity and computational efficiency resulted in \\(40 \\times 40 \\times 16\\) occupancy maps in the perception range \\([-3.2m \\sim 3.2m, -3.2m \\sim 3.2m, -0.78m \\sim 1.78m]\\) along the X-Y (horizontal) plane and Z (vertical) axis." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.473, + 0.903 + ], + "angle": 0, + "content": "Language Descriptions. Given updated 3D bounding boxes annotated with orientations, we derive the language prompts that describe the spatial relationships among objects following SR3D [1]. They serve as the prompt input" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.184 + ], + "angle": 0, + "content": "to the language-grounded perception models for performing 3D visual grounding. Due to increased object density after annotation, identifying unique objects became more challenging. To overcome this, we combined multiple spatial relationship prompts to exclusively ground objects. See more samples in the appendix." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.193, + 0.608, + 0.209 + ], + "angle": 0, + "content": "3.3. Statistics" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.217, + 0.895, + 0.368 + ], + "angle": 0, + "content": "Vocabulary Construction. During labeling, we ask annotators to write semantic categories in an open-vocabulary manner. This was efficient and suited the complex, large-vocabulary dataset and can provide natural annotations for future open-world research. To sort out these labels, we used Sentence-BERT [37] to cluster similar categories with text embeddings, match them to WordNet nodes, and finally revise and merge them manually. The vocabulary shares common categories with COCO (50/69 indoor classes) and LVIS (550/1203)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.369, + 0.895, + 0.565 + ], + "angle": 0, + "content": "Instance Statistics. We first show the instances of different categories in Fig. 3c. Our dataset contains over 760 categories, covering common objects in our daily life. More than 288 categories have over 10 instances, and around 400 categories have more than 5 instances. These numbers are \\(20 \\times\\) higher than most previous works with 3D box annotations and \\(3 \\times\\) higher than ScanNet instance segmentation annotations with more than 5 instances. There is also a notable increase in object numbers of small boxes and different categories (Fig. 3b and 3c). We remove four categories, {wall, ceiling, floor, object} in our 3D detection benchmark and divide the remaining 284 categories into three splits, {head, common, tail} with {90, 94, 100} classes." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.566, + 0.895, + 0.642 + ], + "angle": 0, + "content": "Occupancy Statistics. Semantic occupancy statistics (Fig. 3d) reveal the space occupied by different categories, relevant for navigation and motion planning. We chose the first 80 categories for our occupancy prediction benchmark based on distribution and significance in downstream tasks." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.642, + 0.895, + 0.793 + ], + "angle": 0, + "content": "Language Prompts Statistics. Generated language prompts following SR3D fall into five types of spatial object-to-object relations: Horizontal Proximity, Vertical Proximity, Support, Allocentric, and Between. If a scene has between 2 and 6 instances of a certain category, these categories are considered valid target categories. If a scene has a single instance of a category, it is selected as the anchor category. The training/validation set contains 801711/168322 language prompts, nearly 10 times larger than the original SR3D datasets (Fig. 3b and Tab. 1)." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.803, + 0.71, + 0.82 + ], + "angle": 0, + "content": "4. Embodied Perceptron" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.826, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Given this dataset, we can take multi-modality input, including RGB images, point clouds derived from depth maps as well as language prompts, to extract multi-modal representations and perform different downstream tasks. This section provides a baseline, namely Embodied Perceptron," + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "19760" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.137, + 0.09, + 0.842, + 0.334 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.339, + 0.895, + 0.385 + ], + "angle": 0, + "content": "Figure 4. Embodied Perceptron accepts RGB-D sequence with any number of views along with texts as multi-modal input. It uses classical encoders to extract features for each modality and adopts dense and isomorphic sparse fusion with corresponding decoders for different predictions. The 3D features integrated with the text feature can be further used for language-grounded understanding." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.391, + 0.469, + 0.421 + ], + "angle": 0, + "content": "with a unified framework and customized design for holistic 3D scene understanding from ego-centric views." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.422, + 0.47, + 0.512 + ], + "angle": 0, + "content": "Framework Overview. The framework includes a multimodal 3D encoder to extract object & scene representations and sparse & dense decoders for various downstream tasks. In addition, we customize the output's parameterization and training objectives to fit the formulation of oriented 3D bounding boxes in the sparse decoder." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.518, + 0.31, + 0.533 + ], + "angle": 0, + "content": "4.1. Multi-Modal 3D Encoder" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.539, + 0.47, + 0.674 + ], + "angle": 0, + "content": "As shown in Fig. 4, the multi-modal 3D encoder first has separate encoders for different modalities - ResNet50 [17] and FPN [27] (optional) for 2D images, Minkowski ResNet34 [11] for point clouds, and BERT [14] for texts. After extracting these features, we further fuse and process them into sparse or dense features for different downstream tasks. Next, we first present how we aggregate multi-view inputs and then introduce different fusion approaches for dense and sparse feature extraction." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.675, + 0.471, + 0.901 + ], + "angle": 0, + "content": "Scalability for Input Views. Contrasting with prior works, our framework can accept any number of RGB-D views, making it adaptable and generalizable to varying input orders and quantities. We conveniently aggregate different depth map views by transforming the point clouds into a global coordinate system, downsampling as needed. For multiple images, we query corresponding 2D features using perspective projection from 3D points, averaging them to maintain permutation invariance. This technique allows consistent feature updates during ego-centric exploration. Theoretically, voxel features could be updated by merging the volume feature at frame \\( t \\) with the incremental feature from RGB-D input at frame \\( t + 1 \\). In practice, we accommodate any number of views as batch-wise samples for accelerating training and evaluation. Our model demonstrates no" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.391, + 0.892, + 0.437 + ], + "angle": 0, + "content": "table scalability, where fewer views (e.g., 20) may be used for memory efficiency during training, while more views (e.g., 50) can enhance performance during inference." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.44, + 0.893, + 0.621 + ], + "angle": 0, + "content": "Dense Fusion. Previous works typically integrate the color and coordinates of points at the input stage, like \"painting\" [50], or form multi-modality dense BEV features for concatenated fusion [26, 30]. The latter way suits our occupancy prediction baseline and thus we adopt the straightforward dense fusion on the pre-defined grid, which shares the same resolution with the ground truth. We construct feature volume by projecting grid points onto a single 2D feature map post-FPN, then consolidate it with 3D features densified from sparse voxel features. For object detection, we argue that concurrent multi-modality fusion across multiple feature levels is more effective." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.625, + 0.892, + 0.762 + ], + "angle": 0, + "content": "Isomorphic Multi-Level Multi-Modality Fusion. Formally, the input aggregated points \\( P \\in \\mathcal{R}^{N_p \\times 3} \\) (first voxelized) and \\( N_i \\) images as \\( I \\in \\mathcal{R}^{N_i \\times H \\times W} \\) are processed via a Minkowski ResNet and a shared 2D ResNet respectively. This extracts multi-level sparse voxel features \\( V_k \\in \\mathcal{R}^{C_k \\times N_{V_k}} \\) on \\( K \\) levels and image features \\( F_s \\in \\mathcal{R}^{C_s \\times H_s \\times W_s} \\) on \\( S \\) levels. In practice, these two ResNets produce 4 levels of features, for both point clouds and images, denoted as isomorphic multi-modality encoders." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.893, + 0.902 + ], + "angle": 0, + "content": "In dense fusion, we filter \\( F_{s} \\) with an upsampling FPN to derive a feature map \\( F_{up} \\) with stride \\( = 4 \\) and use it to construct the feature volume for fusing with \\( V_{4} \\). For the sparse case, we use multi-level features as seeds instead of a single dense feature map to predict 3D objects. The initial attempt of still query features from \\( F_{up} \\) or raw images \\( I \\) for these seeds is unstable due to inconsistent features for fusion and confusing gradients back-propagation. Thus, we leverage the isomorphic architecture for level-based projec" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "19761" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.152, + 0.09, + 0.817, + 0.104 + ], + "angle": 0, + "content": "Table 2. Continuous and multi-view 3D object detection benchmark on EmbodiedScan (split by the double line)." + }, + { + "type": "table", + "bbox": [ + 0.12, + 0.11, + 0.847, + 0.23 + ], + "angle": 0, + "content": "
MethodsInputLarge-VocabularyHeadCommonTail
AP25AR25AP50AR50AP25AR25AP25AR25AP25AR25
Camera-OnlyRGB12.8034.614.2513.0717.4044.797.6424.220.033.09
Depth-OnlyDepth17.1651.4010.5225.7521.3961.1413.2741.582.7420.91
Multi-ModalityRGB-D19.0751.5611.5728.1523.5460.2315.8044.991.2417.74
ImVoxelNet [41]RGB6.1520.392.416.3110.9634.294.1215.402.639.21
VoteNet [34]Depth3.206.110.381.226.3112.261.813.341.001.83
FCAF3D [40]Depth9.0744.234.1120.2216.5461.386.7342.772.6724.83
+our decoderDepth14.8051.188.7727.4625.9867.1210.8550.085.7232.85
+paintingRGB-D15.1051.328.6426.6626.2367.5311.3950.645.8032.13
OursRGB-D16.8551.079.7728.2128.6567.5112.8350.467.0931.52
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.24, + 0.47, + 0.329 + ], + "angle": 0, + "content": "tion and feature fusion, i.e., \\( V_{k} \\) queries the corresponding image features of \\( F_{k} \\), which empirically shows a better and more stable performance. This method enables multi-level multi-modality feature fusion compared to the \"painting\" approach and ensures the consistency of features and gradients across different network levels and modalities." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.33, + 0.471, + 0.466 + ], + "angle": 0, + "content": "Vision-Language (VL) Fusion. Given the multi-level sparse visual features \\( F_{k}^{S} \\) and text features from the text encoder, we use a multi-modal fusion transformer model [20, 61] for vision-language information interactions. Each transformer layer uses a self-attention block to refine sparse visual features and exploit spatial relationships. Then visual and text features interact in cross-modal attention blocks. This interaction guides updated sparse grounding features \\( F^{G} \\) to be context-aware for subsequent prediction." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.475, + 0.305, + 0.491 + ], + "angle": 0, + "content": "4.2. Sparse & Dense Decoder" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.498, + 0.47, + 0.588 + ], + "angle": 0, + "content": "Given multi-modal features from typical encoders, we employ separate fusion streams for sparse and dense tasks. This results in four levels of sparse voxel features \\( F_{k}^{S} \\) from isomorphic sparse fusion and a single dense feature \\( F^{D} \\) for decoding and predictions. These are then processed to obtain 3D box and occupancy predictions." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.589, + 0.471, + 0.83 + ], + "angle": 0, + "content": "Sparse Decoder for 3D Boxes Prediction. Using the multi-level fused features \\( F_{k}^{S} \\), we upsample them as in FCAF3D, appending classification, regression, and centerness prediction heads for 3D object detection. In particular, to fit the oriented 3D box output, we add a 6D rotation representation [60] into original regression targets, ultimately decoded as 3D centers c, 3D sizes l, and Euler angles Θ. Training objectives include the original classification loss, centerness loss, and a disentangled Chamfer Distance (CD) loss for eight corners [3, 44]. Specifically, we use one of three groups of decoded predictions, {3D centers, 3D sizes, and Euler angles}, while setting the other two with ground truths to compute three corner losses. For example, given 3D sizes and Euler angles ground truth, we can derive the corner loss between the predicted B and the ground truth box B yielded by 3D center prediction errors:" + }, + { + "type": "equation", + "bbox": [ + 0.182, + 0.834, + 0.469, + 0.852 + ], + "angle": 0, + "content": "\\[\nL _ {\\mathbf {c}} = L _ {C D} (\\mathbf {B} (\\mathbf {c}, \\hat {\\mathbf {l}}, \\hat {\\boldsymbol {\\Theta}}), \\hat {\\mathbf {B}}) \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.47, + 0.902 + ], + "angle": 0, + "content": "Together with the corner loss derived by the overall predicted bounding boxes \\( L_{pred} \\), We balance these losses with preset weights and use them to replace the original box loss:" + }, + { + "type": "equation", + "bbox": [ + 0.545, + 0.249, + 0.892, + 0.266 + ], + "angle": 0, + "content": "\\[\nL _ {l o c} = \\lambda_ {\\mathbf {c}} L _ {\\mathbf {c}} + \\lambda_ {\\mathbf {l}} L _ {\\mathbf {l}} + \\lambda_ {\\boldsymbol {\\Theta}} L _ {\\boldsymbol {\\Theta}} + \\lambda_ {p r e d} L _ {p r e d} \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.272, + 0.892, + 0.332 + ], + "angle": 0, + "content": "We set \\(\\lambda_{\\mathbf{c}} = \\lambda_{\\mathbf{l}} = \\lambda_{\\Theta} = 0.2\\) and \\(\\lambda_{pred} = 0.4\\) to highlight the importance of the overall prediction, which performs well empirically. The target assignment strategy and post-processing during inference also follow FCAF3D [40]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.333, + 0.892, + 0.469 + ], + "angle": 0, + "content": "Dense Decoder for Occupancy Prediction. With the dense feature \\( F^D \\), we use a 3D FPN [41] to aggregate multi-level features and produce multi-scale occupancy predictions. Since the task requires more powerful low-level features for fine detail understanding, predictions at each scale are thus supervised with decayed half weights from high to low resolution [5]. We use cross-entropy loss and scene-class affinity loss [55] for training. During inference, we only use the high-resolution output for prediction." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.47, + 0.892, + 0.606 + ], + "angle": 0, + "content": "Sparse Decoder for 3D Visual Grounding. Grounding features \\( F^G \\) updated after each transformer layer are fed into the prediction heads sharing the same architecture as those used for 3D detection. All prediction head outputs in each layer are supervised during training for stability and improved performance. An additional contrastive loss aligns the visual feature with target text prompts, ensuring the features of a target text token are closer to corresponding visual features and further from other visual or text tokens." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.623, + 0.625, + 0.639 + ], + "angle": 0, + "content": "5. Benchmark" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.649, + 0.892, + 0.77 + ], + "angle": 0, + "content": "Our benchmark has three categories based on data samples: scene-based, view-based, and prompt-based. Scene-based benchmarks mean the samples are based on different scenes, covering continuous and multi-view perception. View-based benchmarks use ego-centric views for tasks like monocular 3D detection. Lastly, samples of 3D visual grounding are based on constructed language prompts. Detailed splits will be discussed in each benchmark." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.771, + 0.892, + 0.905 + ], + "angle": 0, + "content": "For metrics, we use the 3D IoU-based average precision (AP) with thresholds of 0.25 and 0.5 for 3D detection and visual grounding. We also provide average recall (AR) for reference. For occupancy prediction, we employ the mean Intersection of Union (mIoU) as a performance measure. Due to the space limitation, please refer to the appendix for implementation details of different baselines, and more quantitative and qualitative results including an \"in-the-wild\" evaluation demo." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "19762" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.115, + 0.09, + 0.854, + 0.104 + ], + "angle": 0, + "content": "Table 3. Continuous and multi-view occupancy prediction benchmark (split by the double line). \"refri.\" means \"refrigerator\"." + }, + { + "type": "table", + "bbox": [ + 0.081, + 0.107, + 0.889, + 0.206 + ], + "angle": 0, + "content": "
MethodsInputmIOUemptyfloorwallchaircabinetdoortablecouchshelfwindowbedcurtainrefri.plantstairstoilet
Camera-OnlyRGB10.4339.0934.1030.2426.469.4925.5341.6035.1916.2220.4524.4619.8026.0117.261.8329.78
Depth-OnlyDepth14.4473.9166.2256.1349.9615.7024.3756.8455.3530.5526.6642.8130.8133.0121.716.2145.35
Multi-ModalityRGB-D20.7973.5063.6462.3054.6019.9648.9961.1069.7639.8634.6254.8354.4548.9041.227.9763.52
OccNet [43]RGB8.0737.1546.9025.6320.9413.1718.4026.8122.8613.5913.4926.7522.9217.1517.074.7733.60
SurroundOcc [55]RGB9.1038.5446.1723.5523.0413.6019.1527.7922.8813.1113.7224.3218.8913.5814.777.8334.71
Camera-OnlyRGB10.4840.4541.2527.1926.1615.5020.3030.8226.7015.0114.3329.1723.3016.9915.986.1742.57
Depth-OnlyDepth15.5669.9260.5251.7449.4423.0824.3345.7743.5229.7423.0239.0441.2217.4219.5825.7960.45
Multi-ModalityRGB-D19.9771.2164.9255.0052.0427.3533.9747.9346.2631.8727.9846.5846.5624.0539.0124.4067.79
" + }, + { + "type": "table_caption", + "bbox": [ + 0.273, + 0.21, + 0.696, + 0.223 + ], + "angle": 0, + "content": "Table 4. Monocular 3D object detection benchmark on EmbodiedScan." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.228, + 0.872, + 0.334 + ], + "angle": 0, + "content": "
MethodsInput20 Common Classeschaircabinettablebincouchbedbathtubtoilet
AP25AR25AP50AR50
FCOS3D [52]RGB8.9327.960.915.0027.151.146.2110.239.4718.386.3140.51
ImVoxelNet [41]RGB18.9552.741.817.1046.704.6318.1017.8220.3941.5110.1465.70
VoteNet [34]Depth14.3031.441.685.1454.002.4119.5314.7221.8045.5813.4968.16
ImVoteNet [35]RGB-D19.6334.323.888.8256.722.8829.0021.9627.7756.9437.5674.08
FCAF3D [40]Depth25.7078.535.7320.2665.916.4726.6434.9322.5053.6826.3871.90
+our decoderDepth28.1684.504.9220.6963.856.6232.3438.9631.6160.3338.1775.57
+paintingRGB-D30.1983.935.7421.9066.397.4133.6642.8632.2460.0441.3177.59
OursRGB-D34.2885.0312.6132.2569.4710.0137.2945.1731.6763.2750.6380.39
" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.343, + 0.431, + 0.359 + ], + "angle": 0, + "content": "5.1. Fundamental 3D Perception Benchmarks" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.366, + 0.47, + 0.637 + ], + "angle": 0, + "content": "Continuous 3D Perception. As opposed to driving scenarios, indoor scene understanding is typically in an enclosed space, making it important to fully leverage multi-view cues formed by RGB-D sequence and continuously maintain an overall scene-level representation. Thus, we design this new benchmark involving sequential views for perceiving covered 3D regions. Models are trained and evaluated scene-wise with 3930/703/552 scans allocated for training/validation/testing. To accelerate the training and evaluation, we construct \\(N\\) data samples with \\(1 \\sim N\\) views from \\(N\\) sampled views per scan. Here, \\(N = 10\\) during training with random view sampling, while in evaluation, \\(N = 50\\) with fixed views. Corresponding instances and occupancy truths are obtained by combining pre-computed visible instance IDs and occupancy masks of selected views. If a category lacks instances, it is removed when calculating mAP and mIoU. Given this new setup, we primarily offer three baselines with different input modalities (Tab. 2 and 3)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.641, + 0.47, + 0.776 + ], + "angle": 0, + "content": "Continuous 3D Object Detection. As anticipated, both RGB and depth features significantly impact this task, leading to superior results of our RGB-D approach (Tab. 2). The performance of the depth-only model closely mirrors the multimodality approach, indicating depth's dominance in 3D perception. Our method of constructing multi-modal features based on sparse voxel features also aligns with this intuition. Low performance on tail categories suggests dataset size influences performance, warranting future enhancement." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.78, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Continuous Semantic Occupancy Prediction. This benchmark offers comprehensive results including mIoU and IoU for common classes. Unlike the detection benchmark, there is a notable gap between the depth-only and RGB-D baseline. This might be due to the former's limited semantic understanding capability, especially evident in categories like door and curtain, which are similar to walls in shape. On such a task that requires more fine-grained understanding," + }, + { + "type": "table_caption", + "bbox": [ + 0.498, + 0.342, + 0.892, + 0.384 + ], + "angle": 0, + "content": "Table 5. Multi-view 3D visual grounding benchmark. \"Indep/Dep\" refer to \"View-Independent/Dependent\". Easy/Hard and Indep/Dep have a ratio of \\(80\\% /20\\%\\) and \\(78\\% /22\\%\\)." + }, + { + "type": "table", + "bbox": [ + 0.502, + 0.389, + 0.907, + 0.455 + ], + "angle": 0, + "content": "
MethodsInputOverallEasyHardIndepDep
AP25AP25AP25AP25AP25
ScanRefer [8]RGB-D12.8513.789.1213.4410.77
BUTD-DETR [20]RGB-D22.1423.1218.2322.4720.98
L3Det [61]RGB-D23.0724.0118.3423.5921.22
OursRGB-D25.7227.1120.1226.3723.42
" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.462, + 0.892, + 0.522 + ], + "angle": 0, + "content": "the depth sensor's weakness is enlarged. Meanwhile, depth plays a crucial role in predicting empty space, floor, and wall, while RGB information substantially improves prediction for most categories." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.523, + 0.892, + 0.598 + ], + "angle": 0, + "content": "Multi-View 3D Perception. Unlike continuous settings, multi-view 3D perception does not predefine the order of views but provides all views to the model for scene-level results. This setting was studied previously [41], so we first reproduce common methods on our benchmark." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.599, + 0.892, + 0.765 + ], + "angle": 0, + "content": "Multi-View 3D Object Detection. We implement baselines including ImVoxelNet [41] with RGB-only input and VoteNet [34] and FCAF3D [40] with depth-only input (Tab. 2). Additional dimensions are added to predict Euler angles with a simple L1 loss on their cosine values, but it yields overwhelming results. Substituting this with our decoder design markedly improves performance. Further using point cloud input painted for FCAF3D with RGB-D input slightly underperforms our baseline. Nevertheless, all models have substantial potential for improvement, demonstrating the challenges of this new dataset and setup." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.892, + 0.855 + ], + "angle": 0, + "content": "Multi-View Semantic Occupancy Prediction. We implement two popular baselines from autonomous driving benchmarks, OccNet [43] and SurroundOcc [55] (similar to TPV-Former [19]). Their performance slightly lags behind our camera-only baseline. Variants of our baselines exhibit a performance trend akin to embodied benchmarks." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Monocular 3D Perception. Finally, the basic ego-centric setting is monocular 3D perception, specifically 3D detection, where each data sample comprises a single RGB-D" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19763" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.14, + 0.09, + 0.406, + 0.104 + ], + "angle": 0, + "content": "Table 6. Ablation with conventional settings." + }, + { + "type": "table", + "bbox": [ + 0.109, + 0.108, + 0.434, + 0.153 + ], + "angle": 0, + "content": "
OrientedMulti-ViewAP25AR25AP50AR50
XX70.1790.4654.5875.66
✓X61.8790.3147.3073.93
✓✓59.9587.9243.3369.95
" + }, + { + "type": "table_caption", + "bbox": [ + 0.134, + 0.157, + 0.412, + 0.171 + ], + "angle": 0, + "content": "Table 7. Real vs. rendered images on ScanNet." + }, + { + "type": "table", + "bbox": [ + 0.116, + 0.175, + 0.427, + 0.22 + ], + "angle": 0, + "content": "
TrainValOverallHeadCommonTail
RenderRender22.1133.0116.446.74
RenderReal18.7227.0214.856.25
RealReal21.9832.9117.185.05
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.227, + 0.468, + 0.271 + ], + "angle": 0, + "content": "frame and corresponding visible 3D boxes. Scan splits are used to extract frames as data samples, resulting in \\(689\\mathrm{k} / 115\\mathrm{k} / 86\\mathrm{k}\\) images for the training/validation/testing." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.272, + 0.468, + 0.407 + ], + "angle": 0, + "content": "Monocular 3D Object Detection. This is more challenging than multi-view due to the absence of stereo geometric cues and truncated object views in indoor scenes, so the performance is significantly reduced in large-vocabulary settings. Hence, we first create a benchmark for 20 common categories (Tab. 4), observing a larger AP-AR gap for top methods because of difficulties predicting accurate 3D boxes from partial views. Similarly, our method outperforms others, providing a solid baseline for future studies." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.415, + 0.369, + 0.43 + ], + "angle": 0, + "content": "5.2. Language-Grounded Benchmark" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.435, + 0.468, + 0.646 + ], + "angle": 0, + "content": "Multi-View 3D Visual Grounding. Our benchmark introduces language into the perception loop to foster interactive 3D scene representation learning. With comprehensive instance annotations, our benchmark presents more complex prompts and grounding cases than previous works. As an initial step, this setup takes multi-view RGB-D images as input without considering differing prompt timestamps. The goal is to ground the object described by the language prompt in the scene using information from different egocentric views. Ground-truth detection boxes are not provided as candidates for grounding during evaluation, which can better validate end-to-end 3D visual grounding ability than the original SR3D [1]. Data sample splits align with previous benchmarks' 3D scan splits." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.647, + 0.468, + 0.797 + ], + "angle": 0, + "content": "We reimplement classic methods like ScanRefer [8], BUTD-DETR [20], and L3Det [61] (Tab. 5). Our baseline outperforms all due to the strong multi-modal encoder. However, the performance remains much lower than previous works, partly due to changes in input format and annotations, which we will analyze next. Further challenges arise from handling more categories and small objects, making the grounding task more complex in parsing input prompts and predictions. Addressing these new challenges in this classical task would be promising for future research." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.805, + 0.18, + 0.82 + ], + "angle": 0, + "content": "5.3. Analysis" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.826, + 0.468, + 0.855 + ], + "angle": 0, + "content": "Finally, we make further analysis to connect EmbodiedScan to current progress in computer vision." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.856, + 0.468, + 0.901 + ], + "angle": 0, + "content": "Axis-aligned vs. Oriented Boxes. We start with the 18-class detection performance of FCAF3D on ScanNet (Tab. 6). First, we change the annotations to oriented 3D" + }, + { + "type": "table_caption", + "bbox": [ + 0.54, + 0.09, + 0.851, + 0.104 + ], + "angle": 0, + "content": "Table 8. Benefits from training with EmbodiedScan." + }, + { + "type": "table", + "bbox": [ + 0.533, + 0.11, + 0.855, + 0.165 + ], + "angle": 0, + "content": "
TrainValOverallHeadCommonTail
ScanNetScanNet20.2829.8115.576.40
OursScanNet23.0233.8218.096.57
ScanNetOurs10.9221.108.061.78
OursOurs16.8528.6512.837.09
" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.168, + 0.892, + 0.228 + ], + "angle": 0, + "content": "boxes and adapt with our decoder. We find a significant drop in performance, indicating that the orientation estimation makes this task more challenging. We need to explore a better method to represent and predict the object pose." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.229, + 0.892, + 0.335 + ], + "angle": 0, + "content": "Reconstructed Point Cloud vs. Multi-View RGB-D. Subsequently, replacing the reconstructed point clouds with the aggregated ones from multi-view depth maps has minor effects on \\(AP_{25}\\) but heavily impacts \\(AP_{50}\\) (Tab. 6), implying that the accuracy of reconstructed point clouds is superior to raw depth maps. Therefore, integrating reconstruction techniques in perception loops shows potential." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.336, + 0.892, + 0.395 + ], + "angle": 0, + "content": "Next, we study the gap between the real and rendered images, and the benefits from training with EmbodiedScan. We do the comparison with our multi-modality baseline on the large-vocabulary multi-view 3D detection benchmark." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.397, + 0.892, + 0.502 + ], + "angle": 0, + "content": "Real Capture vs. Rendering. As shown in Tab. 7, apart from the significant visual difference between real and rendered images (Fig. 1), the model's performance also has a remarkable decrease when transferring models trained with rendered images to the real world, particularly when the annotations are sufficient (5.99% AP drop on head categories and 5.89% AP lower than models trained with real images.)" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.503, + 0.892, + 0.608 + ], + "angle": 0, + "content": "Benefits from EmbodiedScan. Finally, we also quantitatively evaluate the benefits of training models with our large-scale EmbodiedScan (Tab. 8). As expected, when training our models with EmbodiedScan, we observed a significant improvement in both ScanNet (2.74% AP) and the overall validation split (5.93% AP), particularly 4.01% AP and 7.55% AP increase on head categories." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.619, + 0.618, + 0.633 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.641, + 0.892, + 0.867 + ], + "angle": 0, + "content": "This paper introduces EmbodiedScan, a multi-modal perception suite aiming for language-grounded holistic 3D scene understanding from ego-centric views. We construct a large-scale dataset with diverse sensor data and multimodal annotations, including 3D oriented boxes, semantic occupancy and language descriptions. Based on this dataset, we propose a baseline framework capable of handling any number of views input, using a unified multimodal encoder and task-specific decoders. We establish benchmarks for basic and language-grounded 3D perception. Experiment results highlight our work's value and reveal new challenges in this setup. We believe EmbodiedScan can bring opportunities in embodied 3D perception and may also have a broader impact in related fields with the massive data and rich annotations." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Acknowledgements. This work is supported by Shanghai Artificial Intelligence Laboratory." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19764" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.079, + 0.09, + 0.176, + 0.107 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.171 + ], + "angle": 0, + "content": "[1] Panos Achlioptas, Ahmed Abdelreehm, Fei Xia, Mohamed Elhoseiny, and Leonidas Guibas. Referit3d: Neural listeners for fine-grained 3d object identification in real-world scenes. In European conference on computer vision, 2020. 3, 4, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.173, + 0.472, + 0.27 + ], + "angle": 0, + "content": "[2] Gilad Baruch, Zhuoyuan Chen, Afshin Dehghan, Tal Dimry, Yuri Feigin, Peter Fu, Thomas Gebauer, Brandon Joffe, Daniel Kurz, Arik Schwartz, and Elad Shulman. ARK-scenes - a diverse real-world dataset for 3d indoor scene understanding using mobile RGB-d data. In Thirty-fifth Conference on Neural Information Processing Systems Datasets and Benchmarks Track (Round 1), 2021. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.272, + 0.47, + 0.342 + ], + "angle": 0, + "content": "[3] Garrick Brazil, Abhinav Kumar, Julian Straub, Nikhila Ravi, Justin Johnson, and Georgia Gkioxari. Omni3D: A large benchmark and model for 3D object detection in the wild. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2023. 2, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.343, + 0.47, + 0.412 + ], + "angle": 0, + "content": "[4] Holger Caesar, Varun Bankiti, Alex H. Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuscenes: A multimodal dataset for autonomous driving. CoRR, abs/1903.11027, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.415, + 0.47, + 0.469 + ], + "angle": 0, + "content": "[5] Anh-Quan Cao and Raoul de Charette. Monoscene: Monocular 3d semantic scene completion. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2022. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.472, + 0.47, + 0.541 + ], + "angle": 0, + "content": "[6] Angel Chang, Angela Dai, Thomas Funkhouser, Maciej Halber, Matthias Niessner, Manolis Savva, Shuran Song, Andy Zeng, and Yinda Zhang. Matterport3D: Learning from RGB-D data in indoor environments. International Conference on 3D Vision (3DV), 2017. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.543, + 0.47, + 0.627 + ], + "angle": 0, + "content": "[7] Ming-Fang Chang, John Lambert, Patsorn Sangkloy, Jagjeet Singh, Slawomir Bak, Andrew Hartnett, De Wang, Peter Carr, Simon Lucey, Deva Ramanan, et al. Argoverse: 3d tracking and forecasting with rich maps. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.629, + 0.47, + 0.684 + ], + "angle": 0, + "content": "[8] Dave Zhenyu Chen, Angel X Chang, and Matthias Nießner. Scanrefer: 3d object localization in rgb-d scans using natural language. In European conference on computer vision, 2020. 2, 3, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.686, + 0.47, + 0.755 + ], + "angle": 0, + "content": "[9] Zhenyu Chen, Ali Gholami, Matthias Nießner, and Angel X Chang. Scan2cap: Context-aware dense captioning in rgb-d scans. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3193-3203, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.757, + 0.47, + 0.827 + ], + "angle": 0, + "content": "[10] Zhenyu Chen, Ronghang Hu, Xinlei Chen, Matthias Nießner, and Angel X Chang. Unit3d: A unified transformer for 3d dense captioning and visual grounding. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.829, + 0.47, + 0.885 + ], + "angle": 0, + "content": "[11] Christopher Choy, JunYoung Gwak, and Silvio Savarese. 4d spatio-temporal convnets: Minkowski convolutional neural networks. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.887, + 0.47, + 0.902 + ], + "angle": 0, + "content": "[12] Camille Couprie, Clément Farabet, Laurent Najman, and" + }, + { + "type": "list", + "bbox": [ + 0.079, + 0.115, + 0.472, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "Yann LeCun. Indoor semantic segmentation using depth information. arXiv preprint arXiv:1301.3572, 2013. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.122, + 0.894, + 0.191 + ], + "angle": 0, + "content": "[13] Angela Dai, Angel X Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2017. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.193, + 0.894, + 0.247 + ], + "angle": 0, + "content": "[14] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.25, + 0.895, + 0.319 + ], + "angle": 0, + "content": "[15] Runyu Ding, Jihan Yang, Chuhui Xue, Wenqing Zhang, Song Bai, and Xiaojuan Qi. Pla: Language-driven open-vocabulary 3d scene understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.321, + 0.894, + 0.375 + ], + "angle": 0, + "content": "[16] Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for autonomous driving? the kitti vision benchmark suite. In IEEE Conference on Computer Vision and Pattern Recognition, 2012. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.378, + 0.892, + 0.432 + ], + "angle": 0, + "content": "[17] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In IEEE Conference on Computer Vision and Pattern Recognition, 2016. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.435, + 0.895, + 0.49 + ], + "angle": 0, + "content": "[18] Shijia Huang, Yilun Chen, Jiaya Jia, and Liwei Wang. Multiview transformer for 3d visual grounding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.492, + 0.895, + 0.56 + ], + "angle": 0, + "content": "[19] Yuanhui Huang, Wenzhao Zheng, Yunpeng Zhang, Jie Zhou, and Jiwen Lu. Tri-perspective view for vision-based 3d semantic occupancy prediction. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2023. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.563, + 0.892, + 0.618 + ], + "angle": 0, + "content": "[20] Ayush Jain, Nikolaos Gkanatsios, Ishita Mediratta, and Katerina Fragkiadaki. Bottom up top down detection transformers for language grounding in images and point clouds. In European Conference on Computer Vision, 2022. 3, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.62, + 0.892, + 0.702 + ], + "angle": 0, + "content": "[21] Mukul Khanna*, Yongsen Mao*, Hanxiao Jiang, Sanjay Haresh, Brennan Shacklett, Dhruv Batra, Alexander Clegg, Eric Undersander, Angel X. Chang, and Manolis Savva. Habitat Synthetic Scenes Dataset (HSSD-200): An Analysis of 3D Scene Scale and Realism Tradeoffs for ObjectGoal Navigation. arXiv preprint, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.704, + 0.892, + 0.786 + ], + "angle": 0, + "content": "[22] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C. Berg, Wan-Yen Lo, Piotr Dollar, and Ross Girshick. Segment anything. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2023. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.789, + 0.892, + 0.844 + ], + "angle": 0, + "content": "[23] Alex H. Lang, Sourabh Vora, Holger Caesar, Lubing Zhou, Jiong Yang, and Oscar Beijbom. Pointpillars: Fast encoders for object detection from point clouds. In IEEE Conference on Computer Vision and Pattern Recognition, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.846, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[24] E Li, Shuaijun Wang, Chengyang Li, Dachuan Li, Xiangbin Wu, and Qi Hao. Sustech points: A portable 3d point cloud interactive annotation platform system. In 2020 IEEE Intelligent Vehicles Symposium (IV), 2020. 4" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.895, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "19765" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.176 + ], + "angle": 0, + "content": "[25] Yiming Li, Zhiding Yu, Christopher Choy, Chaowei Xiao, Jose M Alvarez, Sanja Fidler, Chen Feng, and Anima Anandkumar. Voxformer: Sparse voxel transformer for camera-based 3d semantic scene completion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.178, + 0.472, + 0.247 + ], + "angle": 0, + "content": "[26] Tingting Liang, Hongwei Xie, Kaicheng Yu, Zhongyu Xia, Zhiwei Lin, Yongtao Wang, Tao Tang, Bing Wang, and Zhi Tang. Bevfusion: A simple and robust lidar-camera fusion framework. Advances in Neural Information Processing Systems, 2022. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.249, + 0.472, + 0.305 + ], + "angle": 0, + "content": "[27] Tsung-Yi Lin, Piotr Dólar, Ross Girshick, Kaiming He, Bharath Hariharan, and Serge Belongie. Feature pyramid networks for object detection. In IEEE Conference on Computer Vision and Pattern Recognition, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.306, + 0.472, + 0.375 + ], + "angle": 0, + "content": "[28] Zechen Liu, Zizhang Wu, and Roland Toth. Smoke: Single-stage monocular 3d object detection via keypoint estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, pages 996-997, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.377, + 0.472, + 0.433 + ], + "angle": 0, + "content": "[29] Ze Liu, Zheng Zhang, Yue Cao, Han Hu, and Xin Tong. Group-free 3d object detection via transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.434, + 0.472, + 0.503 + ], + "angle": 0, + "content": "[30] Zhijian Liu, Haotian Tang, Alexander Amini, Xinyu Yang, Huizi Mao, Daniela L Rus, and Song Han. Bevfusion: Multi-task multi-sensor fusion with unified bird's-eye view representation. In 2023 IEEE International Conference on Robotics and Automation (ICRA), 2023. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.504, + 0.472, + 0.574 + ], + "angle": 0, + "content": "[31] Yuheng Lu, Chenfeng Xu, Xiaobao Wei, Xiaodong Xie, Masayoshi Tomizuka, Kurt Keutzer, and Shanghang Zhang. Open-vocabulary 3d detection via image-level class and de-biased cross-modal contrastive learning. arXiv preprint arXiv:2207.01987, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.576, + 0.472, + 0.645 + ], + "angle": 0, + "content": "[32] Jiageng Mao, Minzhe Niu, Chenhan Jiang, Hanxue Liang, Jingheng Chen, Xiaodan Liang, Yamin Li, Chaoqiang Ye, Wei Zhang, Zhenguo Li, et al. One million scenes for autonomous driving: Once dataset. arXiv preprint arXiv:2106.11037, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.647, + 0.472, + 0.716 + ], + "angle": 0, + "content": "[33] Songyou Peng, Kyle Genova, Chiyu Jiang, Andrea Tagliasacchi, Marc Pollefeys, Thomas Funkhouser, et al. Openscene: 3d scene understanding with open vocabularies. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.718, + 0.472, + 0.773 + ], + "angle": 0, + "content": "[34] Charles R Qi, Or Litany, Kaiming He, and Leonidas J Guibas. Deep hough voting for 3d object detection in point clouds. In proceedings of the IEEE/CVF International Conference on Computer Vision, 2019. 1, 3, 4, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.775, + 0.472, + 0.843 + ], + "angle": 0, + "content": "[35] Charles R Qi, Xinlei Chen, Or Litany, and Leonidas J Guibas. Imvotenet: Boosting 3d object detection in point clouds with image votes. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2020. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.845, + 0.472, + 0.902 + ], + "angle": 0, + "content": "[36] Santhosh K Ramakrishnan, Aaron Gokaslan, Erik Wijmans, Oleksandr Maksymets, Alex Clegg, John Turner, Eric Undersander, Wojciech Galuba, Andrew Westbury, Angel X Chang, et al. Habitat-matterport 3d dataset (hm3d): 1000" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "large-scale 3d environments for embodied ai. arXiv preprint arXiv:2109.08238, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.123, + 0.894, + 0.179 + ], + "angle": 0, + "content": "[37] Nils Reimers and Iryna Gurevych. Sentence-bert: Sentence embeddings using siamese bert-networks. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing, 2019. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.181, + 0.892, + 0.265 + ], + "angle": 0, + "content": "[38] Mike Roberts, Jason Ramapuram, Anurag Ranjan, Atulit Kumar, Miguel Angel Bautista, Nathan Paczan, Russ Webb, and Joshua M Susskind. Hypersim: A photorealistic synthetic dataset for holistic indoor scene understanding. In Proceedings of the IEEE/CVF international conference on computer vision, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.267, + 0.892, + 0.323 + ], + "angle": 0, + "content": "[39] David Rozenberszki, Or Litany, and Angela Dai. Language-grounded indoor 3d semantic segmentation in the wild. In Proceedings of the European Conference on Computer Vision (ECCV), 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.325, + 0.892, + 0.379 + ], + "angle": 0, + "content": "[40] Danila Rukhovich, Anna Vorontsova, and Anton Konushin. Fcaf3d: Fully convolutional anchor-free 3d object detection. In European Conference on Computer Vision, 2022. 1, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.383, + 0.892, + 0.439 + ], + "angle": 0, + "content": "[41] Danila Rukhovich, Anna Vorontsova, and Anton Konushin. Imvoxelnet: Image to voxels projection for monocular and multi-view general-purpose 3d object detection. In WACV, pages 2397-2406, 2022. 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.441, + 0.892, + 0.497 + ], + "angle": 0, + "content": "[42] Shaoshuai Shi, Xiaogang Wang, and Hongsheng Li. Poincaré: 3d object proposal generation and detection from point cloud. In IEEE Conference on Computer Vision and Pattern Recognition, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.499, + 0.892, + 0.569 + ], + "angle": 0, + "content": "[43] Chonghao Sima, Wenwen Tong, Tai Wang, Li Chen, Silei Wu, Hanming Deng, Yi Gu, Lewei Lu, Ping Luo, Dahua Lin, and Hongyang Li. Scene as occupancy. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2023. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.571, + 0.892, + 0.627 + ], + "angle": 0, + "content": "[44] Andrea Simonelli, Samuel Rota Rota Bulò, Lorenzo Porzi, Manuel López-Antequera, and Peter Kontschieder. Disentangling monocular 3d object detection. In IEEE International Conference on Computer Vision, 2019. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.629, + 0.892, + 0.686 + ], + "angle": 0, + "content": "[45] Shuran Song, Samuel P Lichtenberg, and Jianxiong Xiao. Sun rgb-d: A rgb-d scene understanding benchmark suite. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2015. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.688, + 0.892, + 0.757 + ], + "angle": 0, + "content": "[46] Julian Straub, Thomas Whelan, Lingni Ma, Yufan Chen, Erik Wijmans, Simon Green, Jakob J Engel, Raul Mur-Artal, Carl Ren, Shobhit Verma, et al. The replica dataset: A digital replica of indoor spaces. arXiv preprint arXiv:1906.05797, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.759, + 0.892, + 0.843 + ], + "angle": 0, + "content": "[47] Pei Sun, Henrik Kretzschmar, Xerxes Dotiwalla, Aurelien Chouard, Vijaysai Patnaik, Paul Tsui, James Guo, Yin Zhou, Yuning Chai, Benjamin Caine, et al. Scalability in perception for autonomous driving: Waymo open dataset. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.845, + 0.892, + 0.902 + ], + "angle": 0, + "content": "[48] Ayca Takmaz, Elisabetta Fedele, Robert W Sumner, Marc Pollefeys, Federico Tombari, and Francis Engelmann. Openmask3d: Open-vocabulary 3d instance segmentation. arXiv preprint arXiv:2306.13631, 2023. 3" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "19766" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.147 + ], + "angle": 0, + "content": "[49] Xiaoyu Tian, Tao Jiang, Longfei Yun, Yue Wang, Yilun Wang, and Hang Zhao. Occ3d: A large-scale 3d occupancy prediction benchmark for autonomous driving. arXiv preprint arXiv:2304.14365, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.149, + 0.47, + 0.205 + ], + "angle": 0, + "content": "[50] Sourabh Vora, Alex H Lang, Bassam Helou, and Oscar Beijbom. Pointpainting: Sequential fusion for 3d object detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2020. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.207, + 0.469, + 0.274 + ], + "angle": 0, + "content": "[51] Johanna Wald, Armen Avetisyan, Nassir Navab, Federico Tombari, and Matthias Nießner. Rio: 3d object instance re-localization in changing indoor environments. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2019. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.277, + 0.469, + 0.343 + ], + "angle": 0, + "content": "[52] Tai Wang, Xinge Zhu, Jiangmiao Pang, and Dahua Lin. FCOS3D: Fully convolutional one-stage monocular 3d object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) Workshops, 2021. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.346, + 0.469, + 0.388 + ], + "angle": 0, + "content": "[53] Tai Wang, Jiangmiao Pang, and Dahua Lin. Monocular 3d object detection with depth from motion. In European Conference on Computer Vision (ECCV), 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.39, + 0.469, + 0.443 + ], + "angle": 0, + "content": "[54] Tai Wang, ZHU Xinge, Jiangmiao Pang, and Dahua Lin. Probabilistic and geometric depth: Detecting objects in perspective. In Conference on Robot Learning, pages 1475-1485. PMLR, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.446, + 0.469, + 0.514 + ], + "angle": 0, + "content": "[55] Yi Wei, Linqing Zhao, Wenzhao Zheng, Zheng Zhu, Jie Zhou, and Jiwen Lu. Surroundocc: Multi-camera 3d occupancy prediction for autonomous driving. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2023. 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.517, + 0.469, + 0.599 + ], + "angle": 0, + "content": "[56] Karmesh Yadav, Ram Ramrakhya, Santhosh Kumar Ramakrishnan, Theo Gervet, John Turner, Aaron Gokaslan, Noah Maestre, Angel Xuan Chang, Dhruv Batra, Manolis Savva, et al. Habitat-matterport 3d semantics dataset. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.602, + 0.469, + 0.627 + ], + "angle": 0, + "content": "[57] Yan Yan, Yuxing Mao, and Bo Li. Second: Sparsely embedded convolutional detection. Sensors, 18(10), 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.63, + 0.469, + 0.671 + ], + "angle": 0, + "content": "[58] Wenwei Zhang, Zhe Wang, and Chen Change Loy. Exploring data augmentation for multi-modality 3d object detection. arXiv preprint arXiv:2012.12741, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.673, + 0.469, + 0.713 + ], + "angle": 0, + "content": "[59] Yin Zhou and Oncel Tuzel. Voxelnet: End-to-end learning for point cloud based 3d object detection. In IEEE Conference on Computer Vision and Pattern Recognition, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.716, + 0.469, + 0.77 + ], + "angle": 0, + "content": "[60] Yi Zhou, Connelly Barnes, Jingwan Lu, Jimei Yang, and Hao Li. On the continuity of rotation representations in neural networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.772, + 0.469, + 0.826 + ], + "angle": 0, + "content": "[61] Chenming Zhu, Wenwei Zhang, Tai Wang, Xihui Liu, and Kai Chen. Object2scene: Putting objects in context for open-vocabulary 3d detection. arXiv preprint arXiv:2309.09456, 2023. 3, 6, 7, 8" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.826 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.956 + ], + "angle": 0, + "content": "19767" + } + ] +] \ No newline at end of file diff --git a/2024/EmbodiedScan_ A Holistic Multi-Modal 3D Perception Suite Towards Embodied AI/ab5e0a2a-2a5f-41d4-a08d-01fd4564b83b_origin.pdf b/2024/EmbodiedScan_ A Holistic Multi-Modal 3D Perception Suite Towards Embodied AI/ab5e0a2a-2a5f-41d4-a08d-01fd4564b83b_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..74a25a4314e760e7a83d01e410d64467b2e2dae6 --- /dev/null +++ b/2024/EmbodiedScan_ A Holistic Multi-Modal 3D Perception Suite Towards Embodied AI/ab5e0a2a-2a5f-41d4-a08d-01fd4564b83b_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0d644a18ea0512bacf10a60d281f3eba3039919298ff27dc5fb049c65af1ff53 +size 5253167 diff --git a/2024/EmbodiedScan_ A Holistic Multi-Modal 3D Perception Suite Towards Embodied AI/full.md b/2024/EmbodiedScan_ A Holistic Multi-Modal 3D Perception Suite Towards Embodied AI/full.md new file mode 100644 index 0000000000000000000000000000000000000000..11f5dd0c4644a1b55bebb790c928a3024a09c25b --- /dev/null +++ b/2024/EmbodiedScan_ A Holistic Multi-Modal 3D Perception Suite Towards Embodied AI/full.md @@ -0,0 +1,307 @@ +# EmbodiedScan: A Holistic Multi-Modal 3D Perception Suite Towards Embodied AI + +Tai Wang $^{1*}$ , Xiaohan Mao $^{1,2*}$ , Chenming Zhu $^{1,3*}$ , Runsen Xu $^{1,4}$ , Ruiyuan Lyu $^{1,5}$ , Peisen Li $^{1,5}$ , Xiao Chen $^{1,4}$ , Wenwei Zhang $^{1}$ , Kai Chen $^{1}$ , Tianfan Xue $^{4}$ , Xihui Liu $^{3}$ , Cewu Lu $^{2}$ , Dahua Lin $^{1,4}$ , Jiangmiao Pang $^{1\boxtimes}$ + +$^{1}$ Shanghai AI Laboratory $^{2}$ Shanghai Jiao Tong University $^{3}$ The University of Hong Kong $^{4}$ The Chinese University of Hong Kong $^{5}$ Tsinghua University *Equal contribution Corresponding author + +# EmbodiedScan: 5k scans, 1M ego-centric images, 1M language prompts, 160k instances, 760+ categories + +![](images/c52f8b1370f793eaae7fde5751da68ce5caf52fd8989aae25eeed199e0e26615.jpg) +Figure 1. EmbodiedScan provides a multi-modal, ego-centric 3D perception dataset with massive real-scanned data and rich annotations for indoor scenes. It benchmarks language-grounded holistic 3D scene understanding capabilities for real-world embodied agents. + +# Abstract + +In the realm of computer vision and robotics, embodied agents are expected to explore their environment and carry out human instructions. This necessitates the ability to fully understand 3D scenes given their first-person observations and contextualize them into language for interaction. However, traditional research focuses more on scene-level input and output setups from a global view. To address the gap, we introduce EmbodiedScan, a multi-modal, ego-centric 3D perception dataset and benchmark for holistic 3D scene understanding. It encompasses over 5k scans encapsulating 1M ego-centric RGB-D views, 1M language prompts, 160k 3D-oriented boxes spanning over 760 categories, some of which partially align with LVIS, and dense semantic occupancy with 80 common categories. Building upon this database, we introduce a baseline framework named Embodied Perceptron. It is capable of processing an arbitrary number of multi-modal inputs and demonstrates remarkable 3D perception capabilities, both within the two series of benchmarks we set up, i.e., fundamental 3D perception tasks and language-grounded tasks, and in the wild. + +# 1. Introduction + +Consider an embodied agent operating in an indoor environment. It commences its journey devoid of any prior knowledge about the scene, guided only by an initial instruction. As it begins to explore, it recognizes objects in context and acts with goals along with language interaction. In this process, a commonly needed, fundamental perception capability is to establish a holistic 3D scene understanding given ego-centric observations. This understanding operates at the scene level, covers both object semantics and scene geometry, and can be grounded in language descriptions. + +Nonetheless, subtle but significant discrepancies exist between this expectation and research problems examined within the computer vision community. Most previous studies have primarily revolved around scene-level input and output problems from a global view [13, 34, 40], i.e., taking reconstructed 3D point clouds or meshes as inputs and predicting 3D object bounding boxes or segmenting point clouds. Regarding data, earlier datasets targeting egocentric RGB-D inputs are either too small [12, 45] or lack comprehensive annotations [6, 51] to support the aforementioned results. + +
Dataset#Scans#ImgS#Objs#Cats#PromptsEgo Capture3D Annotations
Replica [46]35----XX
NYU v2 [12]4641.4k35k14-✓X
SUN RGB-D [45]-10k-37-Mono.Box
ScanNet [13, 39]1513264k36k1852k [8]✓Seg., Lang.
Matterport3D [6]2056194k51k40-Multi-ViewSeg.
3RScan [51]1482363k---✓Seg.
ArkitScenes [2]5047450k51k17-✓Box
HyperSim [38]46177k-40+-Mono. & Syn.Box
EmbodiedScan5185890k160k762970k✓Box, Occ., Lang.
+ +Table 1. Comparison with other 3D indoor scene datasets. "Cats" refers to the categories with box annotations for the 3D detection benchmark. EmbodiedScan features more than $10 \times$ categories, prompts, and the most diverse annotations. The numbers are still scaling up with further annotations. Mono./Syn./Lang. means Monocular/Synthetic/Language. + +![](images/71464ca534167399ffc23364e742f6ac43003909e34525eb95af974672f29b08.jpg) +Figure 2. Dataset composition. Embodied-Scan is composed of three data sources and has similar scans, images, objects, and categories in each of them. + +tioned research. It is also not feasible to generate such realistic views by rendering from the existing imperfect meshes. On the other hand, since we cannot trivially obtain the reconstruction of a new environment, models trained with scene-level input are not directly applicable in practice. + +To bridge this divide, we introduce a multi-modal, ego-centric 3D perception dataset and benchmark for holistic 3D scene understanding, termed EmbodiedScan, aimed at facilitating real-world embodied AI applications (Fig. 1). This dataset exploits existing large-scale 3D scene datasets [6, 13, 51] but re-purposes them for continuous scene-level perception from the first-view RGB-D streams. Unlike previous works that offer only point segmentation labels with limited semantics, we employ a SAM-assisted [22] pipeline to annotate objects with oriented 3D bounding boxes and generate language prompts on top. Consequently, EmbodiedScan provides more than 5k scans, nearly 1M ego-centric RGB-D images, and multi-modality annotations, covering 3D oriented boxes with more than 160k instances spanning over 760 categories, dense semantic occupancy with 80 common categories, and 1M language descriptions focusing on spatial relationships among objects. + +Built upon this dataset, we devise a baseline framework for ego-centric 3D perception, Embodied Perceptron. It accepts RGB-D sequences and texts as inputs and manifests scalability to any number of views input with encoders shared across different tasks. With the encoded 2D and 3D features, we employ dense fusion and isomorphic multilevel fusion across them guided by the perspective projection to produce 3D scene and object representations, which are further processed to decode occupancy and 3D box predictions. The derived 3D representations can be further integrated with text embeddings for 3D visual grounding, thus supporting language-grounded applications. + +We establish two series of benchmarks on Embodied-Scan: 1) fundamental 3D perception benchmarks focusing on traditional tasks, including 3D detection and semantic occupancy prediction under different input settings, and 2) a language-grounded scene understanding benchmark with 3D visual grounding as a preliminary exploration. Experimental results validate the effectiveness of our baseline + +model on EmbodiedScan and demonstrate its generalization ability in the wild. Detailed analysis further underscores the value of EmbodiedScan and highlights the primary challenges posed by this new setup. + +# 2. Related Work + +3D Scene Datasets. The development of 3D scene understanding has benefited from large-scale, high-quality datasets like KITTI [16] and SUN RGB-D [45]. These foundational datasets have paved the way for subsequent larger and more diverse collections targeting indoor [6, 13, 38, 51] and driving scenes [4, 7, 32, 47]. However, compared to autonomous driving datasets, those meant for indoor scenes still lack variety in terms of scenes and object diversity (Tab. 1). In contrast, EmbodiedScan provides a large amount of multi-modal data with much richer annotations. Furthermore, it differs by placing an emphasis on the ego-centric perspective within its setup, a feature often overlooked in previous works [2, 13]. + +Except for these conventional dataset works, Omni3D [3] integrates urban and indoor datasets for monocular 3D detection. Our focus, however, lies in indoor scenes due to their unique challenges but has a larger amount of data and annotations, e.g., more than $3 \times$ images and categories with more than 10 instances. In addition, we offer a comprehensive exploration of more general problems for ego-centric 3D perception, such as continuous perception and visual grounding. Other embodied AI datasets like HM3D [36, 56] and HSSD [21] provide ample interaction opportunities but can suffer from poor transferability to real-world scenarios due to their imperfect meshes or synthetic data. Conversely, EmbodiedScan is based on real-scanned RGB-D images, offering a more realistic playground for model training. + +3D Object Detection & Occupancy Prediction. 3D detection and occupancy prediction, as fundamental tasks in 3D perception, focus on different aspects of 3D scene understanding. The former focuses on recognizing foreground objects through a sparse and efficient representation - a set of 3D cuboids corresponding to instances of interest - + +![](images/c59cb8467a1dc5e78484e2bdbf91952e7c49bc91d6bb0c8fc83917520d78d2c8.jpg) +(a) SAM-Assisted Oriented 3D Bounding Boxes Annotation. + +![](images/1439bf27f5175c1f831d53a0f34eb65edeb4c7e241438e29929d9ae590aa5b03.jpg) +(c) Instance Statistics (Increase w.r.t. ScanNet). + +![](images/da6ddacfbcc39a354bafab126b48f1dc899c1767819bab09815df3e84b347e9e.jpg) +(b) 3D Boxes and Language Prompt Statistics. + +![](images/eec26b8dd828e242d2f14d5e57e2d4ced8f4a0d0803213ab6f86bc14ea5e5f0a.jpg) +(d) Occupancy Statistics. +Figure 3. EmbodiedScan annotation and statistics. (a) UI for 3D box annotation. We select keyframes and generate their SAM masks with corresponding axis-aligned boxes. With simple clicks, annotators can create 3D boxes for target objects and further adjust them with reference in three orthogonal views and images. (b) Small boxes $(< 1m^3)$ increase more & prompt statistics. obj./avg./des. refer to objects/average/descriptions. (c) We show the number of instances per category (300 classes). For categories that exist in ScanNet, we plot the absolute increase and observe a significant improvement. (d) We plot the occupancy distribution for each category and see a different word cloud distribution. These two clouds show different aspects, occupied space vs. number of instances, of this dataset. + +while the latter offers a dense, structured pattern that benefits downstream planning. The research community has developed solutions ranging from single-modality, such as LiDAR-based [23, 29, 34, 40, 42, 57, 59] or camera-only approaches [28, 41, 44, 52-54], to multi-modality techniques [26, 30, 35, 50, 58]. Recently, occupancy as a representation, due to its potential in handling unknown semantics and irregular object shapes, has gained more attention [5, 19, 25, 43, 49, 55]. Given their distinctive focuses, we selected these two tasks to form the fundamental 3D perception track on EmbodiedScan. + +Previous works on indoor scenes mainly centered around 3D detection with limitations in object orientations, semantic categories, and input format [29, 34, 40]. In practice, a model is expected to perceive the environment during egocentric exploration, ultimately providing a holistic understanding inclusive of rich semantics, scene geometry, and object poses. To this end, EmbodiedScan and our proposed framework, Embodied Perceptron, provide this necessary data foundation and baseline methodology. + +Language-Grounded 3D Scene Understanding. Language plays a crucial role in human-computer interaction, heightened by recent advances in Large Language Models (LLMs). Its integration with 3D scene understanding is vital for future embodied agents. Past research first explored 3D visual grounding [1, 8, 18, 20] and established new benchmarks including 3D dense captioning [9, 10], open-vocabulary 3D segmentation [15, 33, 48] and detection [31, 61]. This paper focuses on 3D visual grounding + +first, with plans to expand language annotations and benchmarks in the future. Our visual grounding benchmark aligns with the multi-view setting of the basic 3D perception track, taking multiple ego-centric RGB-D images as input, and includes tenfold more complex prompts. + +# 3. Dataset + +This section presents the dataset construction, including data processing and annotation, and shows the statistics. + +# 3.1. Data Collection & Processing + +Ego-Centric Sensor Data Collection. Considering there have been readily available 3D indoor scene scans from existing datasets, we start with integrating those providing ego-centric RGB-D captures with the corresponding camera poses. Given the compatibility of ScanNet [13], 3RScan [51], and Matterport3D [6], we select the high-quality part with complete regions and necessary annotations to form the initial version of EmbodiedScan (Fig. 2). ARKitScenes [2], possessing different data organization, sensors, and annotations, is considered for future inclusion. + +Frame Selection & Scene Division. Although these datasets all have RGB-D data, the data format, sampling frequency, and relationships among viewpoints are different. See more details in the appendix. We first unified the format into a general multi-view case to fit Matterport3D by adding randomness when loading images but maintaining sequential continuity for ScanNet and 3RScan during inference. Our model can thus handle both temporal and + +randomly captured multi-view images. Additionally, we divided building-scale scenes of Matterport3D into regions based on official annotation, selecting corresponding images with depth points falling into the region. As for different sampling rates of images in ScanNet and 3RScan videos, we sample one keyframe per 10 frames for ScanNet and keep all the images for 3RScan. The uniform sampling is generally in line with the actual situation. + +Global Coordinate System. A global coordinate system is necessary to aggregate multi-view observations and serve as a reference for outputs. We follow the convention of ScanNet, deriving a system with the origin around the center of the scene, the horizontal plane lying on the floor and axes aligning with walls [34]. This post-processing harmonizes the data distribution, slightly improving performance on benchmarks. Practical applications may not have such a prior global system or vary according to observations, posing another interesting problem for future exploration. + +# 3.2. Annotation + +We provide three types of annotations - 3D bounding boxes, semantic occupancy, and language descriptions - each serving to enrich different aspects of scene understanding. + +3D Bounding Boxes. Following standard definitions [2, 3], a cuboid is defined by its 3D center, size, and ZXY Euler angle orientation. We used the Segment Anything Model (SAM) [22] and a customized annotation tool based on [24] (Fig. 3a) to address limitations in existing 3D box annotations, i.e., lack of orientation and small object annotations. It supports the conventional functionality of annotating 3D boxes with orientation in three orthographic views. Furthermore, we sample several keyframes with clear imaging according to the camera pose changes and ensure they cover non-overlap regions and most objects to generate SAM masks and axis-aligned boxes for further adjustment. We work with an annotation team and check the quality of all the labels in the end. Each scene takes around 10-30 minutes to annotate, varying with the scene complexity. + +Semantic Occupancy. Semantic occupancy necessitates accurate boundaries across semantic regions without considering object pose or recalling all the objects, so the original point cloud segmentation annotations were more suitable to be used for deriving occupancy ground truth. For each voxel, we assigned the category with the most points as the semantic label for that cell. A compromise between perception granularity and computational efficiency resulted in $40 \times 40 \times 16$ occupancy maps in the perception range $[-3.2m \sim 3.2m, -3.2m \sim 3.2m, -0.78m \sim 1.78m]$ along the X-Y (horizontal) plane and Z (vertical) axis. + +Language Descriptions. Given updated 3D bounding boxes annotated with orientations, we derive the language prompts that describe the spatial relationships among objects following SR3D [1]. They serve as the prompt input + +to the language-grounded perception models for performing 3D visual grounding. Due to increased object density after annotation, identifying unique objects became more challenging. To overcome this, we combined multiple spatial relationship prompts to exclusively ground objects. See more samples in the appendix. + +# 3.3. Statistics + +Vocabulary Construction. During labeling, we ask annotators to write semantic categories in an open-vocabulary manner. This was efficient and suited the complex, large-vocabulary dataset and can provide natural annotations for future open-world research. To sort out these labels, we used Sentence-BERT [37] to cluster similar categories with text embeddings, match them to WordNet nodes, and finally revise and merge them manually. The vocabulary shares common categories with COCO (50/69 indoor classes) and LVIS (550/1203). + +Instance Statistics. We first show the instances of different categories in Fig. 3c. Our dataset contains over 760 categories, covering common objects in our daily life. More than 288 categories have over 10 instances, and around 400 categories have more than 5 instances. These numbers are $20 \times$ higher than most previous works with 3D box annotations and $3 \times$ higher than ScanNet instance segmentation annotations with more than 5 instances. There is also a notable increase in object numbers of small boxes and different categories (Fig. 3b and 3c). We remove four categories, {wall, ceiling, floor, object} in our 3D detection benchmark and divide the remaining 284 categories into three splits, {head, common, tail} with {90, 94, 100} classes. + +Occupancy Statistics. Semantic occupancy statistics (Fig. 3d) reveal the space occupied by different categories, relevant for navigation and motion planning. We chose the first 80 categories for our occupancy prediction benchmark based on distribution and significance in downstream tasks. + +Language Prompts Statistics. Generated language prompts following SR3D fall into five types of spatial object-to-object relations: Horizontal Proximity, Vertical Proximity, Support, Allocentric, and Between. If a scene has between 2 and 6 instances of a certain category, these categories are considered valid target categories. If a scene has a single instance of a category, it is selected as the anchor category. The training/validation set contains 801711/168322 language prompts, nearly 10 times larger than the original SR3D datasets (Fig. 3b and Tab. 1). + +# 4. Embodied Perceptron + +Given this dataset, we can take multi-modality input, including RGB images, point clouds derived from depth maps as well as language prompts, to extract multi-modal representations and perform different downstream tasks. This section provides a baseline, namely Embodied Perceptron, + +![](images/3ccc529bbff2fad98adfb99c9489e207499db1022647684bd5ad2043567f73e0.jpg) +Figure 4. Embodied Perceptron accepts RGB-D sequence with any number of views along with texts as multi-modal input. It uses classical encoders to extract features for each modality and adopts dense and isomorphic sparse fusion with corresponding decoders for different predictions. The 3D features integrated with the text feature can be further used for language-grounded understanding. + +with a unified framework and customized design for holistic 3D scene understanding from ego-centric views. + +Framework Overview. The framework includes a multimodal 3D encoder to extract object & scene representations and sparse & dense decoders for various downstream tasks. In addition, we customize the output's parameterization and training objectives to fit the formulation of oriented 3D bounding boxes in the sparse decoder. + +# 4.1. Multi-Modal 3D Encoder + +As shown in Fig. 4, the multi-modal 3D encoder first has separate encoders for different modalities - ResNet50 [17] and FPN [27] (optional) for 2D images, Minkowski ResNet34 [11] for point clouds, and BERT [14] for texts. After extracting these features, we further fuse and process them into sparse or dense features for different downstream tasks. Next, we first present how we aggregate multi-view inputs and then introduce different fusion approaches for dense and sparse feature extraction. + +Scalability for Input Views. Contrasting with prior works, our framework can accept any number of RGB-D views, making it adaptable and generalizable to varying input orders and quantities. We conveniently aggregate different depth map views by transforming the point clouds into a global coordinate system, downsampling as needed. For multiple images, we query corresponding 2D features using perspective projection from 3D points, averaging them to maintain permutation invariance. This technique allows consistent feature updates during ego-centric exploration. Theoretically, voxel features could be updated by merging the volume feature at frame $t$ with the incremental feature from RGB-D input at frame $t + 1$ . In practice, we accommodate any number of views as batch-wise samples for accelerating training and evaluation. Our model demonstrates no + +table scalability, where fewer views (e.g., 20) may be used for memory efficiency during training, while more views (e.g., 50) can enhance performance during inference. + +Dense Fusion. Previous works typically integrate the color and coordinates of points at the input stage, like "painting" [50], or form multi-modality dense BEV features for concatenated fusion [26, 30]. The latter way suits our occupancy prediction baseline and thus we adopt the straightforward dense fusion on the pre-defined grid, which shares the same resolution with the ground truth. We construct feature volume by projecting grid points onto a single 2D feature map post-FPN, then consolidate it with 3D features densified from sparse voxel features. For object detection, we argue that concurrent multi-modality fusion across multiple feature levels is more effective. + +Isomorphic Multi-Level Multi-Modality Fusion. Formally, the input aggregated points $P \in \mathcal{R}^{N_p \times 3}$ (first voxelized) and $N_i$ images as $I \in \mathcal{R}^{N_i \times H \times W}$ are processed via a Minkowski ResNet and a shared 2D ResNet respectively. This extracts multi-level sparse voxel features $V_k \in \mathcal{R}^{C_k \times N_{V_k}}$ on $K$ levels and image features $F_s \in \mathcal{R}^{C_s \times H_s \times W_s}$ on $S$ levels. In practice, these two ResNets produce 4 levels of features, for both point clouds and images, denoted as isomorphic multi-modality encoders. + +In dense fusion, we filter $F_{s}$ with an upsampling FPN to derive a feature map $F_{up}$ with stride $= 4$ and use it to construct the feature volume for fusing with $V_{4}$ . For the sparse case, we use multi-level features as seeds instead of a single dense feature map to predict 3D objects. The initial attempt of still query features from $F_{up}$ or raw images $I$ for these seeds is unstable due to inconsistent features for fusion and confusing gradients back-propagation. Thus, we leverage the isomorphic architecture for level-based projec + +Table 2. Continuous and multi-view 3D object detection benchmark on EmbodiedScan (split by the double line). + +
MethodsInputLarge-VocabularyHeadCommonTail
AP25AR25AP50AR50AP25AR25AP25AR25AP25AR25
Camera-OnlyRGB12.8034.614.2513.0717.4044.797.6424.220.033.09
Depth-OnlyDepth17.1651.4010.5225.7521.3961.1413.2741.582.7420.91
Multi-ModalityRGB-D19.0751.5611.5728.1523.5460.2315.8044.991.2417.74
ImVoxelNet [41]RGB6.1520.392.416.3110.9634.294.1215.402.639.21
VoteNet [34]Depth3.206.110.381.226.3112.261.813.341.001.83
FCAF3D [40]Depth9.0744.234.1120.2216.5461.386.7342.772.6724.83
+our decoderDepth14.8051.188.7727.4625.9867.1210.8550.085.7232.85
+paintingRGB-D15.1051.328.6426.6626.2367.5311.3950.645.8032.13
OursRGB-D16.8551.079.7728.2128.6567.5112.8350.467.0931.52
+ +tion and feature fusion, i.e., $V_{k}$ queries the corresponding image features of $F_{k}$ , which empirically shows a better and more stable performance. This method enables multi-level multi-modality feature fusion compared to the "painting" approach and ensures the consistency of features and gradients across different network levels and modalities. + +Vision-Language (VL) Fusion. Given the multi-level sparse visual features $F_{k}^{S}$ and text features from the text encoder, we use a multi-modal fusion transformer model [20, 61] for vision-language information interactions. Each transformer layer uses a self-attention block to refine sparse visual features and exploit spatial relationships. Then visual and text features interact in cross-modal attention blocks. This interaction guides updated sparse grounding features $F^{G}$ to be context-aware for subsequent prediction. + +# 4.2. Sparse & Dense Decoder + +Given multi-modal features from typical encoders, we employ separate fusion streams for sparse and dense tasks. This results in four levels of sparse voxel features $F_{k}^{S}$ from isomorphic sparse fusion and a single dense feature $F^{D}$ for decoding and predictions. These are then processed to obtain 3D box and occupancy predictions. + +Sparse Decoder for 3D Boxes Prediction. Using the multi-level fused features $F_{k}^{S}$ , we upsample them as in FCAF3D, appending classification, regression, and centerness prediction heads for 3D object detection. In particular, to fit the oriented 3D box output, we add a 6D rotation representation [60] into original regression targets, ultimately decoded as 3D centers c, 3D sizes l, and Euler angles Θ. Training objectives include the original classification loss, centerness loss, and a disentangled Chamfer Distance (CD) loss for eight corners [3, 44]. Specifically, we use one of three groups of decoded predictions, {3D centers, 3D sizes, and Euler angles}, while setting the other two with ground truths to compute three corner losses. For example, given 3D sizes and Euler angles ground truth, we can derive the corner loss between the predicted B and the ground truth box B yielded by 3D center prediction errors: + +$$ +L _ {\mathbf {c}} = L _ {C D} (\mathbf {B} (\mathbf {c}, \hat {\mathbf {l}}, \hat {\boldsymbol {\Theta}}), \hat {\mathbf {B}}) \tag {1} +$$ + +Together with the corner loss derived by the overall predicted bounding boxes $L_{pred}$ , We balance these losses with preset weights and use them to replace the original box loss: + +$$ +L _ {l o c} = \lambda_ {\mathbf {c}} L _ {\mathbf {c}} + \lambda_ {\mathbf {l}} L _ {\mathbf {l}} + \lambda_ {\boldsymbol {\Theta}} L _ {\boldsymbol {\Theta}} + \lambda_ {p r e d} L _ {p r e d} \tag {2} +$$ + +We set $\lambda_{\mathbf{c}} = \lambda_{\mathbf{l}} = \lambda_{\Theta} = 0.2$ and $\lambda_{pred} = 0.4$ to highlight the importance of the overall prediction, which performs well empirically. The target assignment strategy and post-processing during inference also follow FCAF3D [40]. + +Dense Decoder for Occupancy Prediction. With the dense feature $F^D$ , we use a 3D FPN [41] to aggregate multi-level features and produce multi-scale occupancy predictions. Since the task requires more powerful low-level features for fine detail understanding, predictions at each scale are thus supervised with decayed half weights from high to low resolution [5]. We use cross-entropy loss and scene-class affinity loss [55] for training. During inference, we only use the high-resolution output for prediction. + +Sparse Decoder for 3D Visual Grounding. Grounding features $F^G$ updated after each transformer layer are fed into the prediction heads sharing the same architecture as those used for 3D detection. All prediction head outputs in each layer are supervised during training for stability and improved performance. An additional contrastive loss aligns the visual feature with target text prompts, ensuring the features of a target text token are closer to corresponding visual features and further from other visual or text tokens. + +# 5. Benchmark + +Our benchmark has three categories based on data samples: scene-based, view-based, and prompt-based. Scene-based benchmarks mean the samples are based on different scenes, covering continuous and multi-view perception. View-based benchmarks use ego-centric views for tasks like monocular 3D detection. Lastly, samples of 3D visual grounding are based on constructed language prompts. Detailed splits will be discussed in each benchmark. + +For metrics, we use the 3D IoU-based average precision (AP) with thresholds of 0.25 and 0.5 for 3D detection and visual grounding. We also provide average recall (AR) for reference. For occupancy prediction, we employ the mean Intersection of Union (mIoU) as a performance measure. Due to the space limitation, please refer to the appendix for implementation details of different baselines, and more quantitative and qualitative results including an "in-the-wild" evaluation demo. + +Table 3. Continuous and multi-view occupancy prediction benchmark (split by the double line). "refri." means "refrigerator". + +
MethodsInputmIOUemptyfloorwallchaircabinetdoortablecouchshelfwindowbedcurtainrefri.plantstairstoilet
Camera-OnlyRGB10.4339.0934.1030.2426.469.4925.5341.6035.1916.2220.4524.4619.8026.0117.261.8329.78
Depth-OnlyDepth14.4473.9166.2256.1349.9615.7024.3756.8455.3530.5526.6642.8130.8133.0121.716.2145.35
Multi-ModalityRGB-D20.7973.5063.6462.3054.6019.9648.9961.1069.7639.8634.6254.8354.4548.9041.227.9763.52
OccNet [43]RGB8.0737.1546.9025.6320.9413.1718.4026.8122.8613.5913.4926.7522.9217.1517.074.7733.60
SurroundOcc [55]RGB9.1038.5446.1723.5523.0413.6019.1527.7922.8813.1113.7224.3218.8913.5814.777.8334.71
Camera-OnlyRGB10.4840.4541.2527.1926.1615.5020.3030.8226.7015.0114.3329.1723.3016.9915.986.1742.57
Depth-OnlyDepth15.5669.9260.5251.7449.4423.0824.3345.7743.5229.7423.0239.0441.2217.4219.5825.7960.45
Multi-ModalityRGB-D19.9771.2164.9255.0052.0427.3533.9747.9346.2631.8727.9846.5846.5624.0539.0124.4067.79
+ +Table 4. Monocular 3D object detection benchmark on EmbodiedScan. + +
MethodsInput20 Common Classeschaircabinettablebincouchbedbathtubtoilet
AP25AR25AP50AR50
FCOS3D [52]RGB8.9327.960.915.0027.151.146.2110.239.4718.386.3140.51
ImVoxelNet [41]RGB18.9552.741.817.1046.704.6318.1017.8220.3941.5110.1465.70
VoteNet [34]Depth14.3031.441.685.1454.002.4119.5314.7221.8045.5813.4968.16
ImVoteNet [35]RGB-D19.6334.323.888.8256.722.8829.0021.9627.7756.9437.5674.08
FCAF3D [40]Depth25.7078.535.7320.2665.916.4726.6434.9322.5053.6826.3871.90
+our decoderDepth28.1684.504.9220.6963.856.6232.3438.9631.6160.3338.1775.57
+paintingRGB-D30.1983.935.7421.9066.397.4133.6642.8632.2460.0441.3177.59
OursRGB-D34.2885.0312.6132.2569.4710.0137.2945.1731.6763.2750.6380.39
+ +# 5.1. Fundamental 3D Perception Benchmarks + +Continuous 3D Perception. As opposed to driving scenarios, indoor scene understanding is typically in an enclosed space, making it important to fully leverage multi-view cues formed by RGB-D sequence and continuously maintain an overall scene-level representation. Thus, we design this new benchmark involving sequential views for perceiving covered 3D regions. Models are trained and evaluated scene-wise with 3930/703/552 scans allocated for training/validation/testing. To accelerate the training and evaluation, we construct $N$ data samples with $1 \sim N$ views from $N$ sampled views per scan. Here, $N = 10$ during training with random view sampling, while in evaluation, $N = 50$ with fixed views. Corresponding instances and occupancy truths are obtained by combining pre-computed visible instance IDs and occupancy masks of selected views. If a category lacks instances, it is removed when calculating mAP and mIoU. Given this new setup, we primarily offer three baselines with different input modalities (Tab. 2 and 3). + +Continuous 3D Object Detection. As anticipated, both RGB and depth features significantly impact this task, leading to superior results of our RGB-D approach (Tab. 2). The performance of the depth-only model closely mirrors the multimodality approach, indicating depth's dominance in 3D perception. Our method of constructing multi-modal features based on sparse voxel features also aligns with this intuition. Low performance on tail categories suggests dataset size influences performance, warranting future enhancement. + +Continuous Semantic Occupancy Prediction. This benchmark offers comprehensive results including mIoU and IoU for common classes. Unlike the detection benchmark, there is a notable gap between the depth-only and RGB-D baseline. This might be due to the former's limited semantic understanding capability, especially evident in categories like door and curtain, which are similar to walls in shape. On such a task that requires more fine-grained understanding, + +Table 5. Multi-view 3D visual grounding benchmark. "Indep/Dep" refer to "View-Independent/Dependent". Easy/Hard and Indep/Dep have a ratio of $80\% /20\%$ and $78\% /22\%$ . + +
MethodsInputOverallEasyHardIndepDep
AP25AP25AP25AP25AP25
ScanRefer [8]RGB-D12.8513.789.1213.4410.77
BUTD-DETR [20]RGB-D22.1423.1218.2322.4720.98
L3Det [61]RGB-D23.0724.0118.3423.5921.22
OursRGB-D25.7227.1120.1226.3723.42
+ +the depth sensor's weakness is enlarged. Meanwhile, depth plays a crucial role in predicting empty space, floor, and wall, while RGB information substantially improves prediction for most categories. + +Multi-View 3D Perception. Unlike continuous settings, multi-view 3D perception does not predefine the order of views but provides all views to the model for scene-level results. This setting was studied previously [41], so we first reproduce common methods on our benchmark. + +Multi-View 3D Object Detection. We implement baselines including ImVoxelNet [41] with RGB-only input and VoteNet [34] and FCAF3D [40] with depth-only input (Tab. 2). Additional dimensions are added to predict Euler angles with a simple L1 loss on their cosine values, but it yields overwhelming results. Substituting this with our decoder design markedly improves performance. Further using point cloud input painted for FCAF3D with RGB-D input slightly underperforms our baseline. Nevertheless, all models have substantial potential for improvement, demonstrating the challenges of this new dataset and setup. + +Multi-View Semantic Occupancy Prediction. We implement two popular baselines from autonomous driving benchmarks, OccNet [43] and SurroundOcc [55] (similar to TPV-Former [19]). Their performance slightly lags behind our camera-only baseline. Variants of our baselines exhibit a performance trend akin to embodied benchmarks. + +Monocular 3D Perception. Finally, the basic ego-centric setting is monocular 3D perception, specifically 3D detection, where each data sample comprises a single RGB-D + +Table 6. Ablation with conventional settings. + +
OrientedMulti-ViewAP25AR25AP50AR50
XX70.1790.4654.5875.66
✓X61.8790.3147.3073.93
✓✓59.9587.9243.3369.95
+ +Table 7. Real vs. rendered images on ScanNet. + +
TrainValOverallHeadCommonTail
RenderRender22.1133.0116.446.74
RenderReal18.7227.0214.856.25
RealReal21.9832.9117.185.05
+ +frame and corresponding visible 3D boxes. Scan splits are used to extract frames as data samples, resulting in $689\mathrm{k} / 115\mathrm{k} / 86\mathrm{k}$ images for the training/validation/testing. + +Monocular 3D Object Detection. This is more challenging than multi-view due to the absence of stereo geometric cues and truncated object views in indoor scenes, so the performance is significantly reduced in large-vocabulary settings. Hence, we first create a benchmark for 20 common categories (Tab. 4), observing a larger AP-AR gap for top methods because of difficulties predicting accurate 3D boxes from partial views. Similarly, our method outperforms others, providing a solid baseline for future studies. + +# 5.2. Language-Grounded Benchmark + +Multi-View 3D Visual Grounding. Our benchmark introduces language into the perception loop to foster interactive 3D scene representation learning. With comprehensive instance annotations, our benchmark presents more complex prompts and grounding cases than previous works. As an initial step, this setup takes multi-view RGB-D images as input without considering differing prompt timestamps. The goal is to ground the object described by the language prompt in the scene using information from different egocentric views. Ground-truth detection boxes are not provided as candidates for grounding during evaluation, which can better validate end-to-end 3D visual grounding ability than the original SR3D [1]. Data sample splits align with previous benchmarks' 3D scan splits. + +We reimplement classic methods like ScanRefer [8], BUTD-DETR [20], and L3Det [61] (Tab. 5). Our baseline outperforms all due to the strong multi-modal encoder. However, the performance remains much lower than previous works, partly due to changes in input format and annotations, which we will analyze next. Further challenges arise from handling more categories and small objects, making the grounding task more complex in parsing input prompts and predictions. Addressing these new challenges in this classical task would be promising for future research. + +# 5.3. Analysis + +Finally, we make further analysis to connect EmbodiedScan to current progress in computer vision. + +Axis-aligned vs. Oriented Boxes. We start with the 18-class detection performance of FCAF3D on ScanNet (Tab. 6). First, we change the annotations to oriented 3D + +Table 8. Benefits from training with EmbodiedScan. + +
TrainValOverallHeadCommonTail
ScanNetScanNet20.2829.8115.576.40
OursScanNet23.0233.8218.096.57
ScanNetOurs10.9221.108.061.78
OursOurs16.8528.6512.837.09
+ +boxes and adapt with our decoder. We find a significant drop in performance, indicating that the orientation estimation makes this task more challenging. We need to explore a better method to represent and predict the object pose. + +Reconstructed Point Cloud vs. Multi-View RGB-D. Subsequently, replacing the reconstructed point clouds with the aggregated ones from multi-view depth maps has minor effects on $AP_{25}$ but heavily impacts $AP_{50}$ (Tab. 6), implying that the accuracy of reconstructed point clouds is superior to raw depth maps. Therefore, integrating reconstruction techniques in perception loops shows potential. + +Next, we study the gap between the real and rendered images, and the benefits from training with EmbodiedScan. We do the comparison with our multi-modality baseline on the large-vocabulary multi-view 3D detection benchmark. + +Real Capture vs. Rendering. As shown in Tab. 7, apart from the significant visual difference between real and rendered images (Fig. 1), the model's performance also has a remarkable decrease when transferring models trained with rendered images to the real world, particularly when the annotations are sufficient (5.99% AP drop on head categories and 5.89% AP lower than models trained with real images.) + +Benefits from EmbodiedScan. Finally, we also quantitatively evaluate the benefits of training models with our large-scale EmbodiedScan (Tab. 8). As expected, when training our models with EmbodiedScan, we observed a significant improvement in both ScanNet (2.74% AP) and the overall validation split (5.93% AP), particularly 4.01% AP and 7.55% AP increase on head categories. + +# 6. Conclusion + +This paper introduces EmbodiedScan, a multi-modal perception suite aiming for language-grounded holistic 3D scene understanding from ego-centric views. We construct a large-scale dataset with diverse sensor data and multimodal annotations, including 3D oriented boxes, semantic occupancy and language descriptions. Based on this dataset, we propose a baseline framework capable of handling any number of views input, using a unified multimodal encoder and task-specific decoders. We establish benchmarks for basic and language-grounded 3D perception. Experiment results highlight our work's value and reveal new challenges in this setup. We believe EmbodiedScan can bring opportunities in embodied 3D perception and may also have a broader impact in related fields with the massive data and rich annotations. + +Acknowledgements. This work is supported by Shanghai Artificial Intelligence Laboratory. + +# References + +[1] Panos Achlioptas, Ahmed Abdelreehm, Fei Xia, Mohamed Elhoseiny, and Leonidas Guibas. Referit3d: Neural listeners for fine-grained 3d object identification in real-world scenes. In European conference on computer vision, 2020. 3, 4, 8 +[2] Gilad Baruch, Zhuoyuan Chen, Afshin Dehghan, Tal Dimry, Yuri Feigin, Peter Fu, Thomas Gebauer, Brandon Joffe, Daniel Kurz, Arik Schwartz, and Elad Shulman. ARK-scenes - a diverse real-world dataset for 3d indoor scene understanding using mobile RGB-d data. In Thirty-fifth Conference on Neural Information Processing Systems Datasets and Benchmarks Track (Round 1), 2021. 2, 3, 4 +[3] Garrick Brazil, Abhinav Kumar, Julian Straub, Nikhila Ravi, Justin Johnson, and Georgia Gkioxari. Omni3D: A large benchmark and model for 3D object detection in the wild. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2023. 2, 4, 6 +[4] Holger Caesar, Varun Bankiti, Alex H. Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuscenes: A multimodal dataset for autonomous driving. CoRR, abs/1903.11027, 2019. 2 +[5] Anh-Quan Cao and Raoul de Charette. Monoscene: Monocular 3d semantic scene completion. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2022. 3, 6 +[6] Angel Chang, Angela Dai, Thomas Funkhouser, Maciej Halber, Matthias Niessner, Manolis Savva, Shuran Song, Andy Zeng, and Yinda Zhang. Matterport3D: Learning from RGB-D data in indoor environments. International Conference on 3D Vision (3DV), 2017. 1, 2, 3 +[7] Ming-Fang Chang, John Lambert, Patsorn Sangkloy, Jagjeet Singh, Slawomir Bak, Andrew Hartnett, De Wang, Peter Carr, Simon Lucey, Deva Ramanan, et al. Argoverse: 3d tracking and forecasting with rich maps. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2019. 2 +[8] Dave Zhenyu Chen, Angel X Chang, and Matthias Nießner. Scanrefer: 3d object localization in rgb-d scans using natural language. In European conference on computer vision, 2020. 2, 3, 7, 8 +[9] Zhenyu Chen, Ali Gholami, Matthias Nießner, and Angel X Chang. Scan2cap: Context-aware dense captioning in rgb-d scans. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3193-3203, 2021. 3 +[10] Zhenyu Chen, Ronghang Hu, Xinlei Chen, Matthias Nießner, and Angel X Chang. Unit3d: A unified transformer for 3d dense captioning and visual grounding. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2023. 3 +[11] Christopher Choy, JunYoung Gwak, and Silvio Savarese. 4d spatio-temporal convnets: Minkowski convolutional neural networks. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2019. 5 +[12] Camille Couprie, Clément Farabet, Laurent Najman, and + +Yann LeCun. Indoor semantic segmentation using depth information. arXiv preprint arXiv:1301.3572, 2013. 1, 2 +[13] Angela Dai, Angel X Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2017. 1, 2, 3 +[14] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018. 5 +[15] Runyu Ding, Jihan Yang, Chuhui Xue, Wenqing Zhang, Song Bai, and Xiaojuan Qi. Pla: Language-driven open-vocabulary 3d scene understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023. 3 +[16] Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for autonomous driving? the kitti vision benchmark suite. In IEEE Conference on Computer Vision and Pattern Recognition, 2012. 2 +[17] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In IEEE Conference on Computer Vision and Pattern Recognition, 2016. 5 +[18] Shijia Huang, Yilun Chen, Jiaya Jia, and Liwei Wang. Multiview transformer for 3d visual grounding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022. 3 +[19] Yuanhui Huang, Wenzhao Zheng, Yunpeng Zhang, Jie Zhou, and Jiwen Lu. Tri-perspective view for vision-based 3d semantic occupancy prediction. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2023. 3, 7 +[20] Ayush Jain, Nikolaos Gkanatsios, Ishita Mediratta, and Katerina Fragkiadaki. Bottom up top down detection transformers for language grounding in images and point clouds. In European Conference on Computer Vision, 2022. 3, 6, 7, 8 +[21] Mukul Khanna*, Yongsen Mao*, Hanxiao Jiang, Sanjay Haresh, Brennan Shacklett, Dhruv Batra, Alexander Clegg, Eric Undersander, Angel X. Chang, and Manolis Savva. Habitat Synthetic Scenes Dataset (HSSD-200): An Analysis of 3D Scene Scale and Realism Tradeoffs for ObjectGoal Navigation. arXiv preprint, 2023. 2 +[22] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C. Berg, Wan-Yen Lo, Piotr Dollar, and Ross Girshick. Segment anything. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2023. 2, 4 +[23] Alex H. Lang, Sourabh Vora, Holger Caesar, Lubing Zhou, Jiong Yang, and Oscar Beijbom. Pointpillars: Fast encoders for object detection from point clouds. In IEEE Conference on Computer Vision and Pattern Recognition, 2019. 3 +[24] E Li, Shuaijun Wang, Chengyang Li, Dachuan Li, Xiangbin Wu, and Qi Hao. Sustech points: A portable 3d point cloud interactive annotation platform system. In 2020 IEEE Intelligent Vehicles Symposium (IV), 2020. 4 + +[25] Yiming Li, Zhiding Yu, Christopher Choy, Chaowei Xiao, Jose M Alvarez, Sanja Fidler, Chen Feng, and Anima Anandkumar. Voxformer: Sparse voxel transformer for camera-based 3d semantic scene completion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023. 3 +[26] Tingting Liang, Hongwei Xie, Kaicheng Yu, Zhongyu Xia, Zhiwei Lin, Yongtao Wang, Tao Tang, Bing Wang, and Zhi Tang. Bevfusion: A simple and robust lidar-camera fusion framework. Advances in Neural Information Processing Systems, 2022. 3, 5 +[27] Tsung-Yi Lin, Piotr Dólar, Ross Girshick, Kaiming He, Bharath Hariharan, and Serge Belongie. Feature pyramid networks for object detection. In IEEE Conference on Computer Vision and Pattern Recognition, 2017. 5 +[28] Zechen Liu, Zizhang Wu, and Roland Toth. Smoke: Single-stage monocular 3d object detection via keypoint estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, pages 996-997, 2020. 3 +[29] Ze Liu, Zheng Zhang, Yue Cao, Han Hu, and Xin Tong. Group-free 3d object detection via transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2021. 3 +[30] Zhijian Liu, Haotian Tang, Alexander Amini, Xinyu Yang, Huizi Mao, Daniela L Rus, and Song Han. Bevfusion: Multi-task multi-sensor fusion with unified bird's-eye view representation. In 2023 IEEE International Conference on Robotics and Automation (ICRA), 2023. 3, 5 +[31] Yuheng Lu, Chenfeng Xu, Xiaobao Wei, Xiaodong Xie, Masayoshi Tomizuka, Kurt Keutzer, and Shanghang Zhang. Open-vocabulary 3d detection via image-level class and de-biased cross-modal contrastive learning. arXiv preprint arXiv:2207.01987, 2022. 3 +[32] Jiageng Mao, Minzhe Niu, Chenhan Jiang, Hanxue Liang, Jingheng Chen, Xiaodan Liang, Yamin Li, Chaoqiang Ye, Wei Zhang, Zhenguo Li, et al. One million scenes for autonomous driving: Once dataset. arXiv preprint arXiv:2106.11037, 2021. 2 +[33] Songyou Peng, Kyle Genova, Chiyu Jiang, Andrea Tagliasacchi, Marc Pollefeys, Thomas Funkhouser, et al. Openscene: 3d scene understanding with open vocabularies. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023. 3 +[34] Charles R Qi, Or Litany, Kaiming He, and Leonidas J Guibas. Deep hough voting for 3d object detection in point clouds. In proceedings of the IEEE/CVF International Conference on Computer Vision, 2019. 1, 3, 4, 6, 7 +[35] Charles R Qi, Xinlei Chen, Or Litany, and Leonidas J Guibas. Imvotenet: Boosting 3d object detection in point clouds with image votes. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2020. 3, 7 +[36] Santhosh K Ramakrishnan, Aaron Gokaslan, Erik Wijmans, Oleksandr Maksymets, Alex Clegg, John Turner, Eric Undersander, Wojciech Galuba, Andrew Westbury, Angel X Chang, et al. Habitat-matterport 3d dataset (hm3d): 1000 + +large-scale 3d environments for embodied ai. arXiv preprint arXiv:2109.08238, 2021. 2 +[37] Nils Reimers and Iryna Gurevych. Sentence-bert: Sentence embeddings using siamese bert-networks. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing, 2019. 4 +[38] Mike Roberts, Jason Ramapuram, Anurag Ranjan, Atulit Kumar, Miguel Angel Bautista, Nathan Paczan, Russ Webb, and Joshua M Susskind. Hypersim: A photorealistic synthetic dataset for holistic indoor scene understanding. In Proceedings of the IEEE/CVF international conference on computer vision, 2021. 2 +[39] David Rozenberszki, Or Litany, and Angela Dai. Language-grounded indoor 3d semantic segmentation in the wild. In Proceedings of the European Conference on Computer Vision (ECCV), 2022. 2 +[40] Danila Rukhovich, Anna Vorontsova, and Anton Konushin. Fcaf3d: Fully convolutional anchor-free 3d object detection. In European Conference on Computer Vision, 2022. 1, 3, 6, 7 +[41] Danila Rukhovich, Anna Vorontsova, and Anton Konushin. Imvoxelnet: Image to voxels projection for monocular and multi-view general-purpose 3d object detection. In WACV, pages 2397-2406, 2022. 3, 6, 7 +[42] Shaoshuai Shi, Xiaogang Wang, and Hongsheng Li. Poincaré: 3d object proposal generation and detection from point cloud. In IEEE Conference on Computer Vision and Pattern Recognition, 2019. 3 +[43] Chonghao Sima, Wenwen Tong, Tai Wang, Li Chen, Silei Wu, Hanming Deng, Yi Gu, Lewei Lu, Ping Luo, Dahua Lin, and Hongyang Li. Scene as occupancy. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2023. 3, 7 +[44] Andrea Simonelli, Samuel Rota Rota Bulò, Lorenzo Porzi, Manuel López-Antequera, and Peter Kontschieder. Disentangling monocular 3d object detection. In IEEE International Conference on Computer Vision, 2019. 3, 6 +[45] Shuran Song, Samuel P Lichtenberg, and Jianxiong Xiao. Sun rgb-d: A rgb-d scene understanding benchmark suite. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2015. 1, 2 +[46] Julian Straub, Thomas Whelan, Lingni Ma, Yufan Chen, Erik Wijmans, Simon Green, Jakob J Engel, Raul Mur-Artal, Carl Ren, Shobhit Verma, et al. The replica dataset: A digital replica of indoor spaces. arXiv preprint arXiv:1906.05797, 2019. 2 +[47] Pei Sun, Henrik Kretzschmar, Xerxes Dotiwalla, Aurelien Chouard, Vijaysai Patnaik, Paul Tsui, James Guo, Yin Zhou, Yuning Chai, Benjamin Caine, et al. Scalability in perception for autonomous driving: Waymo open dataset. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2020. 2 +[48] Ayca Takmaz, Elisabetta Fedele, Robert W Sumner, Marc Pollefeys, Federico Tombari, and Francis Engelmann. Openmask3d: Open-vocabulary 3d instance segmentation. arXiv preprint arXiv:2306.13631, 2023. 3 + +[49] Xiaoyu Tian, Tao Jiang, Longfei Yun, Yue Wang, Yilun Wang, and Hang Zhao. Occ3d: A large-scale 3d occupancy prediction benchmark for autonomous driving. arXiv preprint arXiv:2304.14365, 2023. 3 +[50] Sourabh Vora, Alex H Lang, Bassam Helou, and Oscar Beijbom. Pointpainting: Sequential fusion for 3d object detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2020. 3, 5 +[51] Johanna Wald, Armen Avetisyan, Nassir Navab, Federico Tombari, and Matthias Nießner. Rio: 3d object instance re-localization in changing indoor environments. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2019. 1, 2, 3 +[52] Tai Wang, Xinge Zhu, Jiangmiao Pang, and Dahua Lin. FCOS3D: Fully convolutional one-stage monocular 3d object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) Workshops, 2021. 3, 7 +[53] Tai Wang, Jiangmiao Pang, and Dahua Lin. Monocular 3d object detection with depth from motion. In European Conference on Computer Vision (ECCV), 2022. +[54] Tai Wang, ZHU Xinge, Jiangmiao Pang, and Dahua Lin. Probabilistic and geometric depth: Detecting objects in perspective. In Conference on Robot Learning, pages 1475-1485. PMLR, 2022. 3 +[55] Yi Wei, Linqing Zhao, Wenzhao Zheng, Zheng Zhu, Jie Zhou, and Jiwen Lu. Surroundocc: Multi-camera 3d occupancy prediction for autonomous driving. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2023. 3, 6, 7 +[56] Karmesh Yadav, Ram Ramrakhya, Santhosh Kumar Ramakrishnan, Theo Gervet, John Turner, Aaron Gokaslan, Noah Maestre, Angel Xuan Chang, Dhruv Batra, Manolis Savva, et al. Habitat-matterport 3d semantics dataset. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023. 2 +[57] Yan Yan, Yuxing Mao, and Bo Li. Second: Sparsely embedded convolutional detection. Sensors, 18(10), 2018. 3 +[58] Wenwei Zhang, Zhe Wang, and Chen Change Loy. Exploring data augmentation for multi-modality 3d object detection. arXiv preprint arXiv:2012.12741, 2020. 3 +[59] Yin Zhou and Oncel Tuzel. Voxelnet: End-to-end learning for point cloud based 3d object detection. In IEEE Conference on Computer Vision and Pattern Recognition, 2018. 3 +[60] Yi Zhou, Connelly Barnes, Jingwan Lu, Jimei Yang, and Hao Li. On the continuity of rotation representations in neural networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2019. 6 +[61] Chenming Zhu, Wenwei Zhang, Tai Wang, Xihui Liu, and Kai Chen. Object2scene: Putting objects in context for open-vocabulary 3d detection. arXiv preprint arXiv:2309.09456, 2023. 3, 6, 7, 8 \ No newline at end of file diff --git a/2024/EmbodiedScan_ A Holistic Multi-Modal 3D Perception Suite Towards Embodied AI/images.zip b/2024/EmbodiedScan_ A Holistic Multi-Modal 3D Perception Suite Towards Embodied AI/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..1ec2df43f810583fbd4d672ccd93d66028d9365a --- /dev/null +++ b/2024/EmbodiedScan_ A Holistic Multi-Modal 3D Perception Suite Towards Embodied AI/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:fb66700985e37a98644372a3ae6ff07da852d6293dec9acf0f229a4532fc1f71 +size 689336 diff --git a/2024/EmbodiedScan_ A Holistic Multi-Modal 3D Perception Suite Towards Embodied AI/layout.json b/2024/EmbodiedScan_ A Holistic Multi-Modal 3D Perception Suite Towards Embodied AI/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..b4630273924a7830051ced98cebfffe3919c5ad3 --- /dev/null +++ b/2024/EmbodiedScan_ A Holistic Multi-Modal 3D Perception Suite Towards Embodied AI/layout.json @@ -0,0 +1,8161 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 110, + 95, + 484, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 110, + 95, + 484, + 129 + ], + "spans": [ + { + "bbox": [ + 110, + 95, + 484, + 129 + ], + "type": "text", + "content": "EmbodiedScan: A Holistic Multi-Modal 3D Perception Suite Towards Embodied AI" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 69, + 140, + 523, + 184 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 140, + 523, + 184 + ], + "spans": [ + { + "bbox": [ + 69, + 140, + 523, + 184 + ], + "type": "text", + "content": "Tai Wang" + }, + { + "bbox": [ + 69, + 140, + 523, + 184 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 69, + 140, + 523, + 184 + ], + "type": "text", + "content": ", Xiaohan Mao" + }, + { + "bbox": [ + 69, + 140, + 523, + 184 + ], + "type": "inline_equation", + "content": "^{1,2*}" + }, + { + "bbox": [ + 69, + 140, + 523, + 184 + ], + "type": "text", + "content": ", Chenming Zhu" + }, + { + "bbox": [ + 69, + 140, + 523, + 184 + ], + "type": "inline_equation", + "content": "^{1,3*}" + }, + { + "bbox": [ + 69, + 140, + 523, + 184 + ], + "type": "text", + "content": ", Runsen Xu" + }, + { + "bbox": [ + 69, + 140, + 523, + 184 + ], + "type": "inline_equation", + "content": "^{1,4}" + }, + { + "bbox": [ + 69, + 140, + 523, + 184 + ], + "type": "text", + "content": ", Ruiyuan Lyu" + }, + { + "bbox": [ + 69, + 140, + 523, + 184 + ], + "type": "inline_equation", + "content": "^{1,5}" + }, + { + "bbox": [ + 69, + 140, + 523, + 184 + ], + "type": "text", + "content": ", Peisen Li" + }, + { + "bbox": [ + 69, + 140, + 523, + 184 + ], + "type": "inline_equation", + "content": "^{1,5}" + }, + { + "bbox": [ + 69, + 140, + 523, + 184 + ], + "type": "text", + "content": ", Xiao Chen" + }, + { + "bbox": [ + 69, + 140, + 523, + 184 + ], + "type": "inline_equation", + "content": "^{1,4}" + }, + { + "bbox": [ + 69, + 140, + 523, + 184 + ], + "type": "text", + "content": ", Wenwei Zhang" + }, + { + "bbox": [ + 69, + 140, + 523, + 184 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 69, + 140, + 523, + 184 + ], + "type": "text", + "content": ", Kai Chen" + }, + { + "bbox": [ + 69, + 140, + 523, + 184 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 69, + 140, + 523, + 184 + ], + "type": "text", + "content": ", Tianfan Xue" + }, + { + "bbox": [ + 69, + 140, + 523, + 184 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 69, + 140, + 523, + 184 + ], + "type": "text", + "content": ", Xihui Liu" + }, + { + "bbox": [ + 69, + 140, + 523, + 184 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 69, + 140, + 523, + 184 + ], + "type": "text", + "content": ", Cewu Lu" + }, + { + "bbox": [ + 69, + 140, + 523, + 184 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 69, + 140, + 523, + 184 + ], + "type": "text", + "content": ", Dahua Lin" + }, + { + "bbox": [ + 69, + 140, + 523, + 184 + ], + "type": "inline_equation", + "content": "^{1,4}" + }, + { + "bbox": [ + 69, + 140, + 523, + 184 + ], + "type": "text", + "content": ", Jiangmiao Pang" + }, + { + "bbox": [ + 69, + 140, + 523, + 184 + ], + "type": "inline_equation", + "content": "^{1\\boxtimes}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 101, + 190, + 482, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 101, + 190, + 482, + 232 + ], + "spans": [ + { + "bbox": [ + 101, + 190, + 482, + 232 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 101, + 190, + 482, + 232 + ], + "type": "text", + "content": "Shanghai AI Laboratory " + }, + { + "bbox": [ + 101, + 190, + 482, + 232 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 101, + 190, + 482, + 232 + ], + "type": "text", + "content": "Shanghai Jiao Tong University " + }, + { + "bbox": [ + 101, + 190, + 482, + 232 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 101, + 190, + 482, + 232 + ], + "type": "text", + "content": "The University of Hong Kong " + }, + { + "bbox": [ + 101, + 190, + 482, + 232 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 101, + 190, + 482, + 232 + ], + "type": "text", + "content": "The Chinese University of Hong Kong " + }, + { + "bbox": [ + 101, + 190, + 482, + 232 + ], + "type": "inline_equation", + "content": "^{5}" + }, + { + "bbox": [ + 101, + 190, + 482, + 232 + ], + "type": "text", + "content": "Tsinghua University *Equal contribution Corresponding author" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 89, + 242, + 499, + 254 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 242, + 499, + 254 + ], + "spans": [ + { + "bbox": [ + 89, + 242, + 499, + 254 + ], + "type": "text", + "content": "EmbodiedScan: 5k scans, 1M ego-centric images, 1M language prompts, 160k instances, 760+ categories" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 73, + 259, + 514, + 415 + ], + "blocks": [ + { + "bbox": [ + 73, + 259, + 514, + 415 + ], + "lines": [ + { + "bbox": [ + 73, + 259, + 514, + 415 + ], + "spans": [ + { + "bbox": [ + 73, + 259, + 514, + 415 + ], + "type": "image", + "image_path": "c52f8b1370f793eaae7fde5751da68ce5caf52fd8989aae25eeed199e0e26615.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 420, + 546, + 444 + ], + "lines": [ + { + "bbox": [ + 46, + 420, + 546, + 444 + ], + "spans": [ + { + "bbox": [ + 46, + 420, + 546, + 444 + ], + "type": "text", + "content": "Figure 1. EmbodiedScan provides a multi-modal, ego-centric 3D perception dataset with massive real-scanned data and rich annotations for indoor scenes. It benchmarks language-grounded holistic 3D scene understanding capabilities for real-world embodied agents." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 143, + 453, + 192, + 464 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 453, + 192, + 464 + ], + "spans": [ + { + "bbox": [ + 143, + 453, + 192, + 464 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 469, + 290, + 708 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 469, + 290, + 708 + ], + "spans": [ + { + "bbox": [ + 46, + 469, + 290, + 708 + ], + "type": "text", + "content": "In the realm of computer vision and robotics, embodied agents are expected to explore their environment and carry out human instructions. This necessitates the ability to fully understand 3D scenes given their first-person observations and contextualize them into language for interaction. However, traditional research focuses more on scene-level input and output setups from a global view. To address the gap, we introduce EmbodiedScan, a multi-modal, ego-centric 3D perception dataset and benchmark for holistic 3D scene understanding. It encompasses over 5k scans encapsulating 1M ego-centric RGB-D views, 1M language prompts, 160k 3D-oriented boxes spanning over 760 categories, some of which partially align with LVIS, and dense semantic occupancy with 80 common categories. Building upon this database, we introduce a baseline framework named Embodied Perceptron. It is capable of processing an arbitrary number of multi-modal inputs and demonstrates remarkable 3D perception capabilities, both within the two series of benchmarks we set up, i.e., fundamental 3D perception tasks and language-grounded tasks, and in the wild." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 453, + 386, + 464 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 453, + 386, + 464 + ], + "spans": [ + { + "bbox": [ + 306, + 453, + 386, + 464 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 473, + 545, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 473, + 545, + 593 + ], + "spans": [ + { + "bbox": [ + 304, + 473, + 545, + 593 + ], + "type": "text", + "content": "Consider an embodied agent operating in an indoor environment. It commences its journey devoid of any prior knowledge about the scene, guided only by an initial instruction. As it begins to explore, it recognizes objects in context and acts with goals along with language interaction. In this process, a commonly needed, fundamental perception capability is to establish a holistic 3D scene understanding given ego-centric observations. This understanding operates at the scene level, covers both object semantics and scene geometry, and can be grounded in language descriptions." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 594, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 546, + 713 + ], + "type": "text", + "content": "Nonetheless, subtle but significant discrepancies exist between this expectation and research problems examined within the computer vision community. Most previous studies have primarily revolved around scene-level input and output problems from a global view [13, 34, 40], i.e., taking reconstructed 3D point clouds or meshes as inputs and predicting 3D object bounding boxes or segmenting point clouds. Regarding data, earlier datasets targeting egocentric RGB-D inputs are either too small [12, 45] or lack comprehensive annotations [6, 51] to support the aforementioned results." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19757" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 48, + 68, + 383, + 162 + ], + "blocks": [ + { + "bbox": [ + 48, + 68, + 383, + 162 + ], + "lines": [ + { + "bbox": [ + 48, + 68, + 383, + 162 + ], + "spans": [ + { + "bbox": [ + 48, + 68, + 383, + 162 + ], + "type": "table", + "html": "
Dataset#Scans#ImgS#Objs#Cats#PromptsEgo Capture3D Annotations
Replica [46]35----XX
NYU v2 [12]4641.4k35k14-✓X
SUN RGB-D [45]-10k-37-Mono.Box
ScanNet [13, 39]1513264k36k1852k [8]✓Seg., Lang.
Matterport3D [6]2056194k51k40-Multi-ViewSeg.
3RScan [51]1482363k---✓Seg.
ArkitScenes [2]5047450k51k17-✓Box
HyperSim [38]46177k-40+-Mono. & Syn.Box
EmbodiedScan5185890k160k762970k✓Box, Occ., Lang.
", + "image_path": "2f94649cea2c5a1165bd5000545b54bd205ce5c0f3536917ebf879b17f86a911.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 170, + 379, + 216 + ], + "lines": [ + { + "bbox": [ + 46, + 170, + 379, + 216 + ], + "spans": [ + { + "bbox": [ + 46, + 170, + 379, + 216 + ], + "type": "text", + "content": "Table 1. Comparison with other 3D indoor scene datasets. \"Cats\" refers to the categories with box annotations for the 3D detection benchmark. EmbodiedScan features more than " + }, + { + "bbox": [ + 46, + 170, + 379, + 216 + ], + "type": "inline_equation", + "content": "10 \\times" + }, + { + "bbox": [ + 46, + 170, + 379, + 216 + ], + "type": "text", + "content": " categories, prompts, and the most diverse annotations. The numbers are still scaling up with further annotations. Mono./Syn./Lang. means Monocular/Synthetic/Language." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 389, + 67, + 553, + 161 + ], + "blocks": [ + { + "bbox": [ + 389, + 67, + 553, + 161 + ], + "lines": [ + { + "bbox": [ + 389, + 67, + 553, + 161 + ], + "spans": [ + { + "bbox": [ + 389, + 67, + 553, + 161 + ], + "type": "image", + "image_path": "71464ca534167399ffc23364e742f6ac43003909e34525eb95af974672f29b08.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 386, + 170, + 550, + 214 + ], + "lines": [ + { + "bbox": [ + 386, + 170, + 550, + 214 + ], + "spans": [ + { + "bbox": [ + 386, + 170, + 550, + 214 + ], + "type": "text", + "content": "Figure 2. Dataset composition. Embodied-Scan is composed of three data sources and has similar scans, images, objects, and categories in each of them." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 223, + 287, + 282 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 223, + 287, + 282 + ], + "spans": [ + { + "bbox": [ + 46, + 223, + 287, + 282 + ], + "type": "text", + "content": "tioned research. It is also not feasible to generate such realistic views by rendering from the existing imperfect meshes. On the other hand, since we cannot trivially obtain the reconstruction of a new environment, models trained with scene-level input are not directly applicable in practice." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 283, + 288, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 283, + 288, + 486 + ], + "spans": [ + { + "bbox": [ + 46, + 283, + 288, + 486 + ], + "type": "text", + "content": "To bridge this divide, we introduce a multi-modal, ego-centric 3D perception dataset and benchmark for holistic 3D scene understanding, termed EmbodiedScan, aimed at facilitating real-world embodied AI applications (Fig. 1). This dataset exploits existing large-scale 3D scene datasets [6, 13, 51] but re-purposes them for continuous scene-level perception from the first-view RGB-D streams. Unlike previous works that offer only point segmentation labels with limited semantics, we employ a SAM-assisted [22] pipeline to annotate objects with oriented 3D bounding boxes and generate language prompts on top. Consequently, EmbodiedScan provides more than 5k scans, nearly 1M ego-centric RGB-D images, and multi-modality annotations, covering 3D oriented boxes with more than 160k instances spanning over 760 categories, dense semantic occupancy with 80 common categories, and 1M language descriptions focusing on spatial relationships among objects." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 487, + 287, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 487, + 287, + 629 + ], + "spans": [ + { + "bbox": [ + 46, + 487, + 287, + 629 + ], + "type": "text", + "content": "Built upon this dataset, we devise a baseline framework for ego-centric 3D perception, Embodied Perceptron. It accepts RGB-D sequences and texts as inputs and manifests scalability to any number of views input with encoders shared across different tasks. With the encoded 2D and 3D features, we employ dense fusion and isomorphic multilevel fusion across them guided by the perspective projection to produce 3D scene and object representations, which are further processed to decode occupancy and 3D box predictions. The derived 3D representations can be further integrated with text embeddings for 3D visual grounding, thus supporting language-grounded applications." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 630, + 288, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 630, + 288, + 712 + ], + "spans": [ + { + "bbox": [ + 46, + 630, + 288, + 712 + ], + "type": "text", + "content": "We establish two series of benchmarks on Embodied-Scan: 1) fundamental 3D perception benchmarks focusing on traditional tasks, including 3D detection and semantic occupancy prediction under different input settings, and 2) a language-grounded scene understanding benchmark with 3D visual grounding as a preliminary exploration. Experimental results validate the effectiveness of our baseline" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 223, + 546, + 271 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 223, + 546, + 271 + ], + "spans": [ + { + "bbox": [ + 304, + 223, + 546, + 271 + ], + "type": "text", + "content": "model on EmbodiedScan and demonstrate its generalization ability in the wild. Detailed analysis further underscores the value of EmbodiedScan and highlights the primary challenges posed by this new setup." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 284, + 392, + 297 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 284, + 392, + 297 + ], + "spans": [ + { + "bbox": [ + 306, + 284, + 392, + 297 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 304, + 546, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 304, + 546, + 460 + ], + "spans": [ + { + "bbox": [ + 304, + 304, + 546, + 460 + ], + "type": "text", + "content": "3D Scene Datasets. The development of 3D scene understanding has benefited from large-scale, high-quality datasets like KITTI [16] and SUN RGB-D [45]. These foundational datasets have paved the way for subsequent larger and more diverse collections targeting indoor [6, 13, 38, 51] and driving scenes [4, 7, 32, 47]. However, compared to autonomous driving datasets, those meant for indoor scenes still lack variety in terms of scenes and object diversity (Tab. 1). In contrast, EmbodiedScan provides a large amount of multi-modal data with much richer annotations. Furthermore, it differs by placing an emphasis on the ego-centric perspective within its setup, a feature often overlooked in previous works [2, 13]." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 462, + 547, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 462, + 547, + 640 + ], + "spans": [ + { + "bbox": [ + 304, + 462, + 547, + 640 + ], + "type": "text", + "content": "Except for these conventional dataset works, Omni3D [3] integrates urban and indoor datasets for monocular 3D detection. Our focus, however, lies in indoor scenes due to their unique challenges but has a larger amount of data and annotations, e.g., more than " + }, + { + "bbox": [ + 304, + 462, + 547, + 640 + ], + "type": "inline_equation", + "content": "3 \\times" + }, + { + "bbox": [ + 304, + 462, + 547, + 640 + ], + "type": "text", + "content": " images and categories with more than 10 instances. In addition, we offer a comprehensive exploration of more general problems for ego-centric 3D perception, such as continuous perception and visual grounding. Other embodied AI datasets like HM3D [36, 56] and HSSD [21] provide ample interaction opportunities but can suffer from poor transferability to real-world scenarios due to their imperfect meshes or synthetic data. Conversely, EmbodiedScan is based on real-scanned RGB-D images, offering a more realistic playground for model training." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 642, + 547, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 642, + 547, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 642, + 547, + 712 + ], + "type": "text", + "content": "3D Object Detection & Occupancy Prediction. 3D detection and occupancy prediction, as fundamental tasks in 3D perception, focus on different aspects of 3D scene understanding. The former focuses on recognizing foreground objects through a sparse and efficient representation - a set of 3D cuboids corresponding to instances of interest -" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19758" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 51, + 68, + 289, + 164 + ], + "blocks": [ + { + "bbox": [ + 51, + 68, + 289, + 164 + ], + "lines": [ + { + "bbox": [ + 51, + 68, + 289, + 164 + ], + "spans": [ + { + "bbox": [ + 51, + 68, + 289, + 164 + ], + "type": "image", + "image_path": "c59cb8467a1dc5e78484e2bdbf91952e7c49bc91d6bb0c8fc83917520d78d2c8.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 72, + 170, + 268, + 180 + ], + "lines": [ + { + "bbox": [ + 72, + 170, + 268, + 180 + ], + "spans": [ + { + "bbox": [ + 72, + 170, + 268, + 180 + ], + "type": "text", + "content": "(a) SAM-Assisted Oriented 3D Bounding Boxes Annotation." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 50, + 185, + 291, + 273 + ], + "blocks": [ + { + "bbox": [ + 50, + 185, + 291, + 273 + ], + "lines": [ + { + "bbox": [ + 50, + 185, + 291, + 273 + ], + "spans": [ + { + "bbox": [ + 50, + 185, + 291, + 273 + ], + "type": "image", + "image_path": "1439bf27f5175c1f831d53a0f34eb65edeb4c7e241438e29929d9ae590aa5b03.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 94, + 275, + 247, + 284 + ], + "lines": [ + { + "bbox": [ + 94, + 275, + 247, + 284 + ], + "spans": [ + { + "bbox": [ + 94, + 275, + 247, + 284 + ], + "type": "text", + "content": "(c) Instance Statistics (Increase w.r.t. ScanNet)." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 301, + 69, + 545, + 168 + ], + "blocks": [ + { + "bbox": [ + 301, + 69, + 545, + 168 + ], + "lines": [ + { + "bbox": [ + 301, + 69, + 545, + 168 + ], + "spans": [ + { + "bbox": [ + 301, + 69, + 545, + 168 + ], + "type": "image", + "image_path": "da6ddacfbcc39a354bafab126b48f1dc899c1767819bab09815df3e84b347e9e.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 348, + 170, + 497, + 180 + ], + "lines": [ + { + "bbox": [ + 348, + 170, + 497, + 180 + ], + "spans": [ + { + "bbox": [ + 348, + 170, + 497, + 180 + ], + "type": "text", + "content": "(b) 3D Boxes and Language Prompt Statistics." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 301, + 185, + 541, + 272 + ], + "blocks": [ + { + "bbox": [ + 301, + 185, + 541, + 272 + ], + "lines": [ + { + "bbox": [ + 301, + 185, + 541, + 272 + ], + "spans": [ + { + "bbox": [ + 301, + 185, + 541, + 272 + ], + "type": "image", + "image_path": "eec26b8dd828e242d2f14d5e57e2d4ced8f4a0d0803213ab6f86bc14ea5e5f0a.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 380, + 274, + 460, + 284 + ], + "lines": [ + { + "bbox": [ + 380, + 274, + 460, + 284 + ], + "spans": [ + { + "bbox": [ + 380, + 274, + 460, + 284 + ], + "type": "text", + "content": "(d) Occupancy Statistics." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 292, + 545, + 358 + ], + "lines": [ + { + "bbox": [ + 46, + 292, + 545, + 358 + ], + "spans": [ + { + "bbox": [ + 46, + 292, + 545, + 358 + ], + "type": "text", + "content": "Figure 3. EmbodiedScan annotation and statistics. (a) UI for 3D box annotation. We select keyframes and generate their SAM masks with corresponding axis-aligned boxes. With simple clicks, annotators can create 3D boxes for target objects and further adjust them with reference in three orthogonal views and images. (b) Small boxes " + }, + { + "bbox": [ + 46, + 292, + 545, + 358 + ], + "type": "inline_equation", + "content": "(< 1m^3)" + }, + { + "bbox": [ + 46, + 292, + 545, + 358 + ], + "type": "text", + "content": " increase more & prompt statistics. obj./avg./des. refer to objects/average/descriptions. (c) We show the number of instances per category (300 classes). For categories that exist in ScanNet, we plot the absolute increase and observe a significant improvement. (d) We plot the occupancy distribution for each category and see a different word cloud distribution. These two clouds show different aspects, occupied space vs. number of instances, of this dataset." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 365, + 287, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 365, + 287, + 496 + ], + "spans": [ + { + "bbox": [ + 46, + 365, + 287, + 496 + ], + "type": "text", + "content": "while the latter offers a dense, structured pattern that benefits downstream planning. The research community has developed solutions ranging from single-modality, such as LiDAR-based [23, 29, 34, 40, 42, 57, 59] or camera-only approaches [28, 41, 44, 52-54], to multi-modality techniques [26, 30, 35, 50, 58]. Recently, occupancy as a representation, due to its potential in handling unknown semantics and irregular object shapes, has gained more attention [5, 19, 25, 43, 49, 55]. Given their distinctive focuses, we selected these two tasks to form the fundamental 3D perception track on EmbodiedScan." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 498, + 287, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 498, + 287, + 605 + ], + "spans": [ + { + "bbox": [ + 46, + 498, + 287, + 605 + ], + "type": "text", + "content": "Previous works on indoor scenes mainly centered around 3D detection with limitations in object orientations, semantic categories, and input format [29, 34, 40]. In practice, a model is expected to perceive the environment during egocentric exploration, ultimately providing a holistic understanding inclusive of rich semantics, scene geometry, and object poses. To this end, EmbodiedScan and our proposed framework, Embodied Perceptron, provide this necessary data foundation and baseline methodology." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 605, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 287, + 714 + ], + "type": "text", + "content": "Language-Grounded 3D Scene Understanding. Language plays a crucial role in human-computer interaction, heightened by recent advances in Large Language Models (LLMs). Its integration with 3D scene understanding is vital for future embodied agents. Past research first explored 3D visual grounding [1, 8, 18, 20] and established new benchmarks including 3D dense captioning [9, 10], open-vocabulary 3D segmentation [15, 33, 48] and detection [31, 61]. This paper focuses on 3D visual grounding" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 365, + 545, + 425 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 365, + 545, + 425 + ], + "spans": [ + { + "bbox": [ + 304, + 365, + 545, + 425 + ], + "type": "text", + "content": "first, with plans to expand language annotations and benchmarks in the future. Our visual grounding benchmark aligns with the multi-view setting of the basic 3D perception track, taking multiple ego-centric RGB-D images as input, and includes tenfold more complex prompts." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 432, + 360, + 444 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 432, + 360, + 444 + ], + "spans": [ + { + "bbox": [ + 306, + 432, + 360, + 444 + ], + "type": "text", + "content": "3. Dataset" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 449, + 545, + 474 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 449, + 545, + 474 + ], + "spans": [ + { + "bbox": [ + 304, + 449, + 545, + 474 + ], + "type": "text", + "content": "This section presents the dataset construction, including data processing and annotation, and shows the statistics." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 479, + 465, + 493 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 479, + 465, + 493 + ], + "spans": [ + { + "bbox": [ + 306, + 479, + 465, + 493 + ], + "type": "text", + "content": "3.1. Data Collection & Processing" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 498, + 545, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 498, + 545, + 616 + ], + "spans": [ + { + "bbox": [ + 304, + 498, + 545, + 616 + ], + "type": "text", + "content": "Ego-Centric Sensor Data Collection. Considering there have been readily available 3D indoor scene scans from existing datasets, we start with integrating those providing ego-centric RGB-D captures with the corresponding camera poses. Given the compatibility of ScanNet [13], 3RScan [51], and Matterport3D [6], we select the high-quality part with complete regions and necessary annotations to form the initial version of EmbodiedScan (Fig. 2). ARKitScenes [2], possessing different data organization, sensors, and annotations, is considered for future inclusion." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "content": "Frame Selection & Scene Division. Although these datasets all have RGB-D data, the data format, sampling frequency, and relationships among viewpoints are different. See more details in the appendix. We first unified the format into a general multi-view case to fit Matterport3D by adding randomness when loading images but maintaining sequential continuity for ScanNet and 3RScan during inference. Our model can thus handle both temporal and" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19759" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 167 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 167 + ], + "type": "text", + "content": "randomly captured multi-view images. Additionally, we divided building-scale scenes of Matterport3D into regions based on official annotation, selecting corresponding images with depth points falling into the region. As for different sampling rates of images in ScanNet and 3RScan videos, we sample one keyframe per 10 frames for ScanNet and keep all the images for 3RScan. The uniform sampling is generally in line with the actual situation." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 168, + 289, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 168, + 289, + 289 + ], + "spans": [ + { + "bbox": [ + 46, + 168, + 289, + 289 + ], + "type": "text", + "content": "Global Coordinate System. A global coordinate system is necessary to aggregate multi-view observations and serve as a reference for outputs. We follow the convention of ScanNet, deriving a system with the origin around the center of the scene, the horizontal plane lying on the floor and axes aligning with walls [34]. This post-processing harmonizes the data distribution, slightly improving performance on benchmarks. Practical applications may not have such a prior global system or vary according to observations, posing another interesting problem for future exploration." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 297, + 124, + 309 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 297, + 124, + 309 + ], + "spans": [ + { + "bbox": [ + 47, + 297, + 124, + 309 + ], + "type": "text", + "content": "3.2. Annotation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 316, + 287, + 353 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 316, + 287, + 353 + ], + "spans": [ + { + "bbox": [ + 46, + 316, + 287, + 353 + ], + "type": "text", + "content": "We provide three types of annotations - 3D bounding boxes, semantic occupancy, and language descriptions - each serving to enrich different aspects of scene understanding." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 354, + 289, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 354, + 289, + 533 + ], + "spans": [ + { + "bbox": [ + 46, + 354, + 289, + 533 + ], + "type": "text", + "content": "3D Bounding Boxes. Following standard definitions [2, 3], a cuboid is defined by its 3D center, size, and ZXY Euler angle orientation. We used the Segment Anything Model (SAM) [22] and a customized annotation tool based on [24] (Fig. 3a) to address limitations in existing 3D box annotations, i.e., lack of orientation and small object annotations. It supports the conventional functionality of annotating 3D boxes with orientation in three orthographic views. Furthermore, we sample several keyframes with clear imaging according to the camera pose changes and ensure they cover non-overlap regions and most objects to generate SAM masks and axis-aligned boxes for further adjustment. We work with an annotation team and check the quality of all the labels in the end. Each scene takes around 10-30 minutes to annotate, varying with the scene complexity." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 533, + 289, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 533, + 289, + 665 + ], + "spans": [ + { + "bbox": [ + 46, + 533, + 289, + 665 + ], + "type": "text", + "content": "Semantic Occupancy. Semantic occupancy necessitates accurate boundaries across semantic regions without considering object pose or recalling all the objects, so the original point cloud segmentation annotations were more suitable to be used for deriving occupancy ground truth. For each voxel, we assigned the category with the most points as the semantic label for that cell. A compromise between perception granularity and computational efficiency resulted in " + }, + { + "bbox": [ + 46, + 533, + 289, + 665 + ], + "type": "inline_equation", + "content": "40 \\times 40 \\times 16" + }, + { + "bbox": [ + 46, + 533, + 289, + 665 + ], + "type": "text", + "content": " occupancy maps in the perception range " + }, + { + "bbox": [ + 46, + 533, + 289, + 665 + ], + "type": "inline_equation", + "content": "[-3.2m \\sim 3.2m, -3.2m \\sim 3.2m, -0.78m \\sim 1.78m]" + }, + { + "bbox": [ + 46, + 533, + 289, + 665 + ], + "type": "text", + "content": " along the X-Y (horizontal) plane and Z (vertical) axis." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 665, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 289, + 715 + ], + "type": "text", + "content": "Language Descriptions. Given updated 3D bounding boxes annotated with orientations, we derive the language prompts that describe the spatial relationships among objects following SR3D [1]. They serve as the prompt input" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 72, + 547, + 145 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 145 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 145 + ], + "type": "text", + "content": "to the language-grounded perception models for performing 3D visual grounding. Due to increased object density after annotation, identifying unique objects became more challenging. To overcome this, we combined multiple spatial relationship prompts to exclusively ground objects. See more samples in the appendix." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 152, + 372, + 165 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 152, + 372, + 165 + ], + "spans": [ + { + "bbox": [ + 306, + 152, + 372, + 165 + ], + "type": "text", + "content": "3.3. Statistics" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 171, + 547, + 291 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 171, + 547, + 291 + ], + "spans": [ + { + "bbox": [ + 304, + 171, + 547, + 291 + ], + "type": "text", + "content": "Vocabulary Construction. During labeling, we ask annotators to write semantic categories in an open-vocabulary manner. This was efficient and suited the complex, large-vocabulary dataset and can provide natural annotations for future open-world research. To sort out these labels, we used Sentence-BERT [37] to cluster similar categories with text embeddings, match them to WordNet nodes, and finally revise and merge them manually. The vocabulary shares common categories with COCO (50/69 indoor classes) and LVIS (550/1203)." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 292, + 547, + 447 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 292, + 547, + 447 + ], + "spans": [ + { + "bbox": [ + 304, + 292, + 547, + 447 + ], + "type": "text", + "content": "Instance Statistics. We first show the instances of different categories in Fig. 3c. Our dataset contains over 760 categories, covering common objects in our daily life. More than 288 categories have over 10 instances, and around 400 categories have more than 5 instances. These numbers are " + }, + { + "bbox": [ + 304, + 292, + 547, + 447 + ], + "type": "inline_equation", + "content": "20 \\times" + }, + { + "bbox": [ + 304, + 292, + 547, + 447 + ], + "type": "text", + "content": " higher than most previous works with 3D box annotations and " + }, + { + "bbox": [ + 304, + 292, + 547, + 447 + ], + "type": "inline_equation", + "content": "3 \\times" + }, + { + "bbox": [ + 304, + 292, + 547, + 447 + ], + "type": "text", + "content": " higher than ScanNet instance segmentation annotations with more than 5 instances. There is also a notable increase in object numbers of small boxes and different categories (Fig. 3b and 3c). We remove four categories, {wall, ceiling, floor, object} in our 3D detection benchmark and divide the remaining 284 categories into three splits, {head, common, tail} with {90, 94, 100} classes." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 448, + 547, + 508 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 448, + 547, + 508 + ], + "spans": [ + { + "bbox": [ + 304, + 448, + 547, + 508 + ], + "type": "text", + "content": "Occupancy Statistics. Semantic occupancy statistics (Fig. 3d) reveal the space occupied by different categories, relevant for navigation and motion planning. We chose the first 80 categories for our occupancy prediction benchmark based on distribution and significance in downstream tasks." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 508, + 547, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 508, + 547, + 628 + ], + "spans": [ + { + "bbox": [ + 304, + 508, + 547, + 628 + ], + "type": "text", + "content": "Language Prompts Statistics. Generated language prompts following SR3D fall into five types of spatial object-to-object relations: Horizontal Proximity, Vertical Proximity, Support, Allocentric, and Between. If a scene has between 2 and 6 instances of a certain category, these categories are considered valid target categories. If a scene has a single instance of a category, it is selected as the anchor category. The training/validation set contains 801711/168322 language prompts, nearly 10 times larger than the original SR3D datasets (Fig. 3b and Tab. 1)." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 635, + 434, + 649 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 635, + 434, + 649 + ], + "spans": [ + { + "bbox": [ + 305, + 635, + 434, + 649 + ], + "type": "text", + "content": "4. Embodied Perceptron" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 654, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 547, + 715 + ], + "type": "text", + "content": "Given this dataset, we can take multi-modality input, including RGB images, point clouds derived from depth maps as well as language prompts, to extract multi-modal representations and perform different downstream tasks. This section provides a baseline, namely Embodied Perceptron," + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "19760" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 83, + 71, + 515, + 264 + ], + "blocks": [ + { + "bbox": [ + 83, + 71, + 515, + 264 + ], + "lines": [ + { + "bbox": [ + 83, + 71, + 515, + 264 + ], + "spans": [ + { + "bbox": [ + 83, + 71, + 515, + 264 + ], + "type": "image", + "image_path": "3ccc529bbff2fad98adfb99c9489e207499db1022647684bd5ad2043567f73e0.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 268, + 547, + 304 + ], + "lines": [ + { + "bbox": [ + 46, + 268, + 547, + 304 + ], + "spans": [ + { + "bbox": [ + 46, + 268, + 547, + 304 + ], + "type": "text", + "content": "Figure 4. Embodied Perceptron accepts RGB-D sequence with any number of views along with texts as multi-modal input. It uses classical encoders to extract features for each modality and adopts dense and isomorphic sparse fusion with corresponding decoders for different predictions. The 3D features integrated with the text feature can be further used for language-grounded understanding." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 309, + 287, + 333 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 309, + 287, + 333 + ], + "spans": [ + { + "bbox": [ + 47, + 309, + 287, + 333 + ], + "type": "text", + "content": "with a unified framework and customized design for holistic 3D scene understanding from ego-centric views." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 334, + 287, + 405 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 334, + 287, + 405 + ], + "spans": [ + { + "bbox": [ + 47, + 334, + 287, + 405 + ], + "type": "text", + "content": "Framework Overview. The framework includes a multimodal 3D encoder to extract object & scene representations and sparse & dense decoders for various downstream tasks. In addition, we customize the output's parameterization and training objectives to fit the formulation of oriented 3D bounding boxes in the sparse decoder." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 410, + 189, + 422 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 410, + 189, + 422 + ], + "spans": [ + { + "bbox": [ + 47, + 410, + 189, + 422 + ], + "type": "text", + "content": "4.1. Multi-Modal 3D Encoder" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 426, + 287, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 426, + 287, + 533 + ], + "spans": [ + { + "bbox": [ + 46, + 426, + 287, + 533 + ], + "type": "text", + "content": "As shown in Fig. 4, the multi-modal 3D encoder first has separate encoders for different modalities - ResNet50 [17] and FPN [27] (optional) for 2D images, Minkowski ResNet34 [11] for point clouds, and BERT [14] for texts. After extracting these features, we further fuse and process them into sparse or dense features for different downstream tasks. Next, we first present how we aggregate multi-view inputs and then introduce different fusion approaches for dense and sparse feature extraction." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 534, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 534, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 534, + 288, + 713 + ], + "type": "text", + "content": "Scalability for Input Views. Contrasting with prior works, our framework can accept any number of RGB-D views, making it adaptable and generalizable to varying input orders and quantities. We conveniently aggregate different depth map views by transforming the point clouds into a global coordinate system, downsampling as needed. For multiple images, we query corresponding 2D features using perspective projection from 3D points, averaging them to maintain permutation invariance. This technique allows consistent feature updates during ego-centric exploration. Theoretically, voxel features could be updated by merging the volume feature at frame " + }, + { + "bbox": [ + 46, + 534, + 288, + 713 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 534, + 288, + 713 + ], + "type": "text", + "content": " with the incremental feature from RGB-D input at frame " + }, + { + "bbox": [ + 46, + 534, + 288, + 713 + ], + "type": "inline_equation", + "content": "t + 1" + }, + { + "bbox": [ + 46, + 534, + 288, + 713 + ], + "type": "text", + "content": ". In practice, we accommodate any number of views as batch-wise samples for accelerating training and evaluation. Our model demonstrates no" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 309, + 545, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 309, + 545, + 346 + ], + "spans": [ + { + "bbox": [ + 304, + 309, + 545, + 346 + ], + "type": "text", + "content": "table scalability, where fewer views (e.g., 20) may be used for memory efficiency during training, while more views (e.g., 50) can enhance performance during inference." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 348, + 546, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 348, + 546, + 491 + ], + "spans": [ + { + "bbox": [ + 304, + 348, + 546, + 491 + ], + "type": "text", + "content": "Dense Fusion. Previous works typically integrate the color and coordinates of points at the input stage, like \"painting\" [50], or form multi-modality dense BEV features for concatenated fusion [26, 30]. The latter way suits our occupancy prediction baseline and thus we adopt the straightforward dense fusion on the pre-defined grid, which shares the same resolution with the ground truth. We construct feature volume by projecting grid points onto a single 2D feature map post-FPN, then consolidate it with 3D features densified from sparse voxel features. For object detection, we argue that concurrent multi-modality fusion across multiple feature levels is more effective." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 495, + 545, + 603 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 495, + 545, + 603 + ], + "spans": [ + { + "bbox": [ + 304, + 495, + 545, + 603 + ], + "type": "text", + "content": "Isomorphic Multi-Level Multi-Modality Fusion. Formally, the input aggregated points " + }, + { + "bbox": [ + 304, + 495, + 545, + 603 + ], + "type": "inline_equation", + "content": "P \\in \\mathcal{R}^{N_p \\times 3}" + }, + { + "bbox": [ + 304, + 495, + 545, + 603 + ], + "type": "text", + "content": " (first voxelized) and " + }, + { + "bbox": [ + 304, + 495, + 545, + 603 + ], + "type": "inline_equation", + "content": "N_i" + }, + { + "bbox": [ + 304, + 495, + 545, + 603 + ], + "type": "text", + "content": " images as " + }, + { + "bbox": [ + 304, + 495, + 545, + 603 + ], + "type": "inline_equation", + "content": "I \\in \\mathcal{R}^{N_i \\times H \\times W}" + }, + { + "bbox": [ + 304, + 495, + 545, + 603 + ], + "type": "text", + "content": " are processed via a Minkowski ResNet and a shared 2D ResNet respectively. This extracts multi-level sparse voxel features " + }, + { + "bbox": [ + 304, + 495, + 545, + 603 + ], + "type": "inline_equation", + "content": "V_k \\in \\mathcal{R}^{C_k \\times N_{V_k}}" + }, + { + "bbox": [ + 304, + 495, + 545, + 603 + ], + "type": "text", + "content": " on " + }, + { + "bbox": [ + 304, + 495, + 545, + 603 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 304, + 495, + 545, + 603 + ], + "type": "text", + "content": " levels and image features " + }, + { + "bbox": [ + 304, + 495, + 545, + 603 + ], + "type": "inline_equation", + "content": "F_s \\in \\mathcal{R}^{C_s \\times H_s \\times W_s}" + }, + { + "bbox": [ + 304, + 495, + 545, + 603 + ], + "type": "text", + "content": " on " + }, + { + "bbox": [ + 304, + 495, + 545, + 603 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 304, + 495, + 545, + 603 + ], + "type": "text", + "content": " levels. In practice, these two ResNets produce 4 levels of features, for both point clouds and images, denoted as isomorphic multi-modality encoders." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 605, + 546, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 546, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 546, + 714 + ], + "type": "text", + "content": "In dense fusion, we filter " + }, + { + "bbox": [ + 304, + 605, + 546, + 714 + ], + "type": "inline_equation", + "content": "F_{s}" + }, + { + "bbox": [ + 304, + 605, + 546, + 714 + ], + "type": "text", + "content": " with an upsampling FPN to derive a feature map " + }, + { + "bbox": [ + 304, + 605, + 546, + 714 + ], + "type": "inline_equation", + "content": "F_{up}" + }, + { + "bbox": [ + 304, + 605, + 546, + 714 + ], + "type": "text", + "content": " with stride " + }, + { + "bbox": [ + 304, + 605, + 546, + 714 + ], + "type": "inline_equation", + "content": "= 4" + }, + { + "bbox": [ + 304, + 605, + 546, + 714 + ], + "type": "text", + "content": " and use it to construct the feature volume for fusing with " + }, + { + "bbox": [ + 304, + 605, + 546, + 714 + ], + "type": "inline_equation", + "content": "V_{4}" + }, + { + "bbox": [ + 304, + 605, + 546, + 714 + ], + "type": "text", + "content": ". For the sparse case, we use multi-level features as seeds instead of a single dense feature map to predict 3D objects. The initial attempt of still query features from " + }, + { + "bbox": [ + 304, + 605, + 546, + 714 + ], + "type": "inline_equation", + "content": "F_{up}" + }, + { + "bbox": [ + 304, + 605, + 546, + 714 + ], + "type": "text", + "content": " or raw images " + }, + { + "bbox": [ + 304, + 605, + 546, + 714 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 304, + 605, + 546, + 714 + ], + "type": "text", + "content": " for these seeds is unstable due to inconsistent features for fusion and confusing gradients back-propagation. Thus, we leverage the isomorphic architecture for level-based projec" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19761" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 73, + 87, + 518, + 182 + ], + "blocks": [ + { + "bbox": [ + 93, + 71, + 500, + 82 + ], + "lines": [ + { + "bbox": [ + 93, + 71, + 500, + 82 + ], + "spans": [ + { + "bbox": [ + 93, + 71, + 500, + 82 + ], + "type": "text", + "content": "Table 2. Continuous and multi-view 3D object detection benchmark on EmbodiedScan (split by the double line)." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 73, + 87, + 518, + 182 + ], + "lines": [ + { + "bbox": [ + 73, + 87, + 518, + 182 + ], + "spans": [ + { + "bbox": [ + 73, + 87, + 518, + 182 + ], + "type": "table", + "html": "
MethodsInputLarge-VocabularyHeadCommonTail
AP25AR25AP50AR50AP25AR25AP25AR25AP25AR25
Camera-OnlyRGB12.8034.614.2513.0717.4044.797.6424.220.033.09
Depth-OnlyDepth17.1651.4010.5225.7521.3961.1413.2741.582.7420.91
Multi-ModalityRGB-D19.0751.5611.5728.1523.5460.2315.8044.991.2417.74
ImVoxelNet [41]RGB6.1520.392.416.3110.9634.294.1215.402.639.21
VoteNet [34]Depth3.206.110.381.226.3112.261.813.341.001.83
FCAF3D [40]Depth9.0744.234.1120.2216.5461.386.7342.772.6724.83
+our decoderDepth14.8051.188.7727.4625.9867.1210.8550.085.7232.85
+paintingRGB-D15.1051.328.6426.6626.2367.5311.3950.645.8032.13
OursRGB-D16.8551.079.7728.2128.6567.5112.8350.467.0931.52
", + "image_path": "96e4746e2ddf03338c208d0e23107d66ac299cb495a7a1b532dda8d1932d859d.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 190, + 287, + 260 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 190, + 287, + 260 + ], + "spans": [ + { + "bbox": [ + 46, + 190, + 287, + 260 + ], + "type": "text", + "content": "tion and feature fusion, i.e., " + }, + { + "bbox": [ + 46, + 190, + 287, + 260 + ], + "type": "inline_equation", + "content": "V_{k}" + }, + { + "bbox": [ + 46, + 190, + 287, + 260 + ], + "type": "text", + "content": " queries the corresponding image features of " + }, + { + "bbox": [ + 46, + 190, + 287, + 260 + ], + "type": "inline_equation", + "content": "F_{k}" + }, + { + "bbox": [ + 46, + 190, + 287, + 260 + ], + "type": "text", + "content": ", which empirically shows a better and more stable performance. This method enables multi-level multi-modality feature fusion compared to the \"painting\" approach and ensures the consistency of features and gradients across different network levels and modalities." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 261, + 288, + 369 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 261, + 288, + 369 + ], + "spans": [ + { + "bbox": [ + 46, + 261, + 288, + 369 + ], + "type": "text", + "content": "Vision-Language (VL) Fusion. Given the multi-level sparse visual features " + }, + { + "bbox": [ + 46, + 261, + 288, + 369 + ], + "type": "inline_equation", + "content": "F_{k}^{S}" + }, + { + "bbox": [ + 46, + 261, + 288, + 369 + ], + "type": "text", + "content": " and text features from the text encoder, we use a multi-modal fusion transformer model [20, 61] for vision-language information interactions. Each transformer layer uses a self-attention block to refine sparse visual features and exploit spatial relationships. Then visual and text features interact in cross-modal attention blocks. This interaction guides updated sparse grounding features " + }, + { + "bbox": [ + 46, + 261, + 288, + 369 + ], + "type": "inline_equation", + "content": "F^{G}" + }, + { + "bbox": [ + 46, + 261, + 288, + 369 + ], + "type": "text", + "content": " to be context-aware for subsequent prediction." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 376, + 186, + 388 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 376, + 186, + 388 + ], + "spans": [ + { + "bbox": [ + 47, + 376, + 186, + 388 + ], + "type": "text", + "content": "4.2. Sparse & Dense Decoder" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 394, + 287, + 465 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 394, + 287, + 465 + ], + "spans": [ + { + "bbox": [ + 46, + 394, + 287, + 465 + ], + "type": "text", + "content": "Given multi-modal features from typical encoders, we employ separate fusion streams for sparse and dense tasks. This results in four levels of sparse voxel features " + }, + { + "bbox": [ + 46, + 394, + 287, + 465 + ], + "type": "inline_equation", + "content": "F_{k}^{S}" + }, + { + "bbox": [ + 46, + 394, + 287, + 465 + ], + "type": "text", + "content": " from isomorphic sparse fusion and a single dense feature " + }, + { + "bbox": [ + 46, + 394, + 287, + 465 + ], + "type": "inline_equation", + "content": "F^{D}" + }, + { + "bbox": [ + 46, + 394, + 287, + 465 + ], + "type": "text", + "content": " for decoding and predictions. These are then processed to obtain 3D box and occupancy predictions." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 466, + 288, + 657 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 466, + 288, + 657 + ], + "spans": [ + { + "bbox": [ + 46, + 466, + 288, + 657 + ], + "type": "text", + "content": "Sparse Decoder for 3D Boxes Prediction. Using the multi-level fused features " + }, + { + "bbox": [ + 46, + 466, + 288, + 657 + ], + "type": "inline_equation", + "content": "F_{k}^{S}" + }, + { + "bbox": [ + 46, + 466, + 288, + 657 + ], + "type": "text", + "content": ", we upsample them as in FCAF3D, appending classification, regression, and centerness prediction heads for 3D object detection. In particular, to fit the oriented 3D box output, we add a 6D rotation representation [60] into original regression targets, ultimately decoded as 3D centers c, 3D sizes l, and Euler angles Θ. Training objectives include the original classification loss, centerness loss, and a disentangled Chamfer Distance (CD) loss for eight corners [3, 44]. Specifically, we use one of three groups of decoded predictions, {3D centers, 3D sizes, and Euler angles}, while setting the other two with ground truths to compute three corner losses. For example, given 3D sizes and Euler angles ground truth, we can derive the corner loss between the predicted B and the ground truth box B yielded by 3D center prediction errors:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 111, + 660, + 287, + 674 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 111, + 660, + 287, + 674 + ], + "spans": [ + { + "bbox": [ + 111, + 660, + 287, + 674 + ], + "type": "interline_equation", + "content": "L _ {\\mathbf {c}} = L _ {C D} (\\mathbf {B} (\\mathbf {c}, \\hat {\\mathbf {l}}, \\hat {\\boldsymbol {\\Theta}}), \\hat {\\mathbf {B}}) \\tag {1}", + "image_path": "d3ff7889da745762068df531cb4a1c6aedc1cf374a0e75c8424a597c3ff2a411.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 677, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 287, + 714 + ], + "type": "text", + "content": "Together with the corner loss derived by the overall predicted bounding boxes " + }, + { + "bbox": [ + 46, + 677, + 287, + 714 + ], + "type": "inline_equation", + "content": "L_{pred}" + }, + { + "bbox": [ + 46, + 677, + 287, + 714 + ], + "type": "text", + "content": ", We balance these losses with preset weights and use them to replace the original box loss:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 333, + 197, + 545, + 210 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 197, + 545, + 210 + ], + "spans": [ + { + "bbox": [ + 333, + 197, + 545, + 210 + ], + "type": "interline_equation", + "content": "L _ {l o c} = \\lambda_ {\\mathbf {c}} L _ {\\mathbf {c}} + \\lambda_ {\\mathbf {l}} L _ {\\mathbf {l}} + \\lambda_ {\\boldsymbol {\\Theta}} L _ {\\boldsymbol {\\Theta}} + \\lambda_ {p r e d} L _ {p r e d} \\tag {2}", + "image_path": "841eb2c23ef7841155c63db1428dbc4d092d7812cefcfa1313198961e905a48f.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 215, + 545, + 262 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 215, + 545, + 262 + ], + "spans": [ + { + "bbox": [ + 304, + 215, + 545, + 262 + ], + "type": "text", + "content": "We set " + }, + { + "bbox": [ + 304, + 215, + 545, + 262 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathbf{c}} = \\lambda_{\\mathbf{l}} = \\lambda_{\\Theta} = 0.2" + }, + { + "bbox": [ + 304, + 215, + 545, + 262 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 215, + 545, + 262 + ], + "type": "inline_equation", + "content": "\\lambda_{pred} = 0.4" + }, + { + "bbox": [ + 304, + 215, + 545, + 262 + ], + "type": "text", + "content": " to highlight the importance of the overall prediction, which performs well empirically. The target assignment strategy and post-processing during inference also follow FCAF3D [40]." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 263, + 545, + 371 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 263, + 545, + 371 + ], + "spans": [ + { + "bbox": [ + 304, + 263, + 545, + 371 + ], + "type": "text", + "content": "Dense Decoder for Occupancy Prediction. With the dense feature " + }, + { + "bbox": [ + 304, + 263, + 545, + 371 + ], + "type": "inline_equation", + "content": "F^D" + }, + { + "bbox": [ + 304, + 263, + 545, + 371 + ], + "type": "text", + "content": ", we use a 3D FPN [41] to aggregate multi-level features and produce multi-scale occupancy predictions. Since the task requires more powerful low-level features for fine detail understanding, predictions at each scale are thus supervised with decayed half weights from high to low resolution [5]. We use cross-entropy loss and scene-class affinity loss [55] for training. During inference, we only use the high-resolution output for prediction." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 372, + 545, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 372, + 545, + 479 + ], + "spans": [ + { + "bbox": [ + 304, + 372, + 545, + 479 + ], + "type": "text", + "content": "Sparse Decoder for 3D Visual Grounding. Grounding features " + }, + { + "bbox": [ + 304, + 372, + 545, + 479 + ], + "type": "inline_equation", + "content": "F^G" + }, + { + "bbox": [ + 304, + 372, + 545, + 479 + ], + "type": "text", + "content": " updated after each transformer layer are fed into the prediction heads sharing the same architecture as those used for 3D detection. All prediction head outputs in each layer are supervised during training for stability and improved performance. An additional contrastive loss aligns the visual feature with target text prompts, ensuring the features of a target text token are closer to corresponding visual features and further from other visual or text tokens." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 493, + 382, + 506 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 493, + 382, + 506 + ], + "spans": [ + { + "bbox": [ + 306, + 493, + 382, + 506 + ], + "type": "text", + "content": "5. Benchmark" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 514, + 545, + 609 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 514, + 545, + 609 + ], + "spans": [ + { + "bbox": [ + 304, + 514, + 545, + 609 + ], + "type": "text", + "content": "Our benchmark has three categories based on data samples: scene-based, view-based, and prompt-based. Scene-based benchmarks mean the samples are based on different scenes, covering continuous and multi-view perception. View-based benchmarks use ego-centric views for tasks like monocular 3D detection. Lastly, samples of 3D visual grounding are based on constructed language prompts. Detailed splits will be discussed in each benchmark." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 610, + 545, + 716 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 610, + 545, + 716 + ], + "spans": [ + { + "bbox": [ + 304, + 610, + 545, + 716 + ], + "type": "text", + "content": "For metrics, we use the 3D IoU-based average precision (AP) with thresholds of 0.25 and 0.5 for 3D detection and visual grounding. We also provide average recall (AR) for reference. For occupancy prediction, we employ the mean Intersection of Union (mIoU) as a performance measure. Due to the space limitation, please refer to the appendix for implementation details of different baselines, and more quantitative and qualitative results including an \"in-the-wild\" evaluation demo." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "19762" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 49, + 84, + 544, + 163 + ], + "blocks": [ + { + "bbox": [ + 70, + 71, + 522, + 82 + ], + "lines": [ + { + "bbox": [ + 70, + 71, + 522, + 82 + ], + "spans": [ + { + "bbox": [ + 70, + 71, + 522, + 82 + ], + "type": "text", + "content": "Table 3. Continuous and multi-view occupancy prediction benchmark (split by the double line). \"refri.\" means \"refrigerator\"." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 49, + 84, + 544, + 163 + ], + "lines": [ + { + "bbox": [ + 49, + 84, + 544, + 163 + ], + "spans": [ + { + "bbox": [ + 49, + 84, + 544, + 163 + ], + "type": "table", + "html": "
MethodsInputmIOUemptyfloorwallchaircabinetdoortablecouchshelfwindowbedcurtainrefri.plantstairstoilet
Camera-OnlyRGB10.4339.0934.1030.2426.469.4925.5341.6035.1916.2220.4524.4619.8026.0117.261.8329.78
Depth-OnlyDepth14.4473.9166.2256.1349.9615.7024.3756.8455.3530.5526.6642.8130.8133.0121.716.2145.35
Multi-ModalityRGB-D20.7973.5063.6462.3054.6019.9648.9961.1069.7639.8634.6254.8354.4548.9041.227.9763.52
OccNet [43]RGB8.0737.1546.9025.6320.9413.1718.4026.8122.8613.5913.4926.7522.9217.1517.074.7733.60
SurroundOcc [55]RGB9.1038.5446.1723.5523.0413.6019.1527.7922.8813.1113.7224.3218.8913.5814.777.8334.71
Camera-OnlyRGB10.4840.4541.2527.1926.1615.5020.3030.8226.7015.0114.3329.1723.3016.9915.986.1742.57
Depth-OnlyDepth15.5669.9260.5251.7449.4423.0824.3345.7743.5229.7423.0239.0441.2217.4219.5825.7960.45
Multi-ModalityRGB-D19.9771.2164.9255.0052.0427.3533.9747.9346.2631.8727.9846.5846.5624.0539.0124.4067.79
", + "image_path": "12da896c12782fda0da8dcbde1eff7c3a407658a1028bd8d12ba93eb48cd3777.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 58, + 180, + 533, + 264 + ], + "blocks": [ + { + "bbox": [ + 167, + 166, + 425, + 176 + ], + "lines": [ + { + "bbox": [ + 167, + 166, + 425, + 176 + ], + "spans": [ + { + "bbox": [ + 167, + 166, + 425, + 176 + ], + "type": "text", + "content": "Table 4. Monocular 3D object detection benchmark on EmbodiedScan." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 180, + 533, + 264 + ], + "lines": [ + { + "bbox": [ + 58, + 180, + 533, + 264 + ], + "spans": [ + { + "bbox": [ + 58, + 180, + 533, + 264 + ], + "type": "table", + "html": "
MethodsInput20 Common Classeschaircabinettablebincouchbedbathtubtoilet
AP25AR25AP50AR50
FCOS3D [52]RGB8.9327.960.915.0027.151.146.2110.239.4718.386.3140.51
ImVoxelNet [41]RGB18.9552.741.817.1046.704.6318.1017.8220.3941.5110.1465.70
VoteNet [34]Depth14.3031.441.685.1454.002.4119.5314.7221.8045.5813.4968.16
ImVoteNet [35]RGB-D19.6334.323.888.8256.722.8829.0021.9627.7756.9437.5674.08
FCAF3D [40]Depth25.7078.535.7320.2665.916.4726.6434.9322.5053.6826.3871.90
+our decoderDepth28.1684.504.9220.6963.856.6232.3438.9631.6160.3338.1775.57
+paintingRGB-D30.1983.935.7421.9066.397.4133.6642.8632.2460.0441.3177.59
OursRGB-D34.2885.0312.6132.2569.4710.0137.2945.1731.6763.2750.6380.39
", + "image_path": "2daeefa63262ca3e272f3adc8a0cf2a2ea2f61bbd314b9167e1eb400d7950398.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 271, + 263, + 284 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 271, + 263, + 284 + ], + "spans": [ + { + "bbox": [ + 47, + 271, + 263, + 284 + ], + "type": "text", + "content": "5.1. Fundamental 3D Perception Benchmarks" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 289, + 287, + 504 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 289, + 287, + 504 + ], + "spans": [ + { + "bbox": [ + 46, + 289, + 287, + 504 + ], + "type": "text", + "content": "Continuous 3D Perception. As opposed to driving scenarios, indoor scene understanding is typically in an enclosed space, making it important to fully leverage multi-view cues formed by RGB-D sequence and continuously maintain an overall scene-level representation. Thus, we design this new benchmark involving sequential views for perceiving covered 3D regions. Models are trained and evaluated scene-wise with 3930/703/552 scans allocated for training/validation/testing. To accelerate the training and evaluation, we construct " + }, + { + "bbox": [ + 46, + 289, + 287, + 504 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 289, + 287, + 504 + ], + "type": "text", + "content": " data samples with " + }, + { + "bbox": [ + 46, + 289, + 287, + 504 + ], + "type": "inline_equation", + "content": "1 \\sim N" + }, + { + "bbox": [ + 46, + 289, + 287, + 504 + ], + "type": "text", + "content": " views from " + }, + { + "bbox": [ + 46, + 289, + 287, + 504 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 289, + 287, + 504 + ], + "type": "text", + "content": " sampled views per scan. Here, " + }, + { + "bbox": [ + 46, + 289, + 287, + 504 + ], + "type": "inline_equation", + "content": "N = 10" + }, + { + "bbox": [ + 46, + 289, + 287, + 504 + ], + "type": "text", + "content": " during training with random view sampling, while in evaluation, " + }, + { + "bbox": [ + 46, + 289, + 287, + 504 + ], + "type": "inline_equation", + "content": "N = 50" + }, + { + "bbox": [ + 46, + 289, + 287, + 504 + ], + "type": "text", + "content": " with fixed views. Corresponding instances and occupancy truths are obtained by combining pre-computed visible instance IDs and occupancy masks of selected views. If a category lacks instances, it is removed when calculating mAP and mIoU. Given this new setup, we primarily offer three baselines with different input modalities (Tab. 2 and 3)." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 507, + 287, + 614 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 507, + 287, + 614 + ], + "spans": [ + { + "bbox": [ + 46, + 507, + 287, + 614 + ], + "type": "text", + "content": "Continuous 3D Object Detection. As anticipated, both RGB and depth features significantly impact this task, leading to superior results of our RGB-D approach (Tab. 2). The performance of the depth-only model closely mirrors the multimodality approach, indicating depth's dominance in 3D perception. Our method of constructing multi-modal features based on sparse voxel features also aligns with this intuition. Low performance on tail categories suggests dataset size influences performance, warranting future enhancement." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 617, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 617, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 617, + 288, + 714 + ], + "type": "text", + "content": "Continuous Semantic Occupancy Prediction. This benchmark offers comprehensive results including mIoU and IoU for common classes. Unlike the detection benchmark, there is a notable gap between the depth-only and RGB-D baseline. This might be due to the former's limited semantic understanding capability, especially evident in categories like door and curtain, which are similar to walls in shape. On such a task that requires more fine-grained understanding," + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 307, + 308, + 555, + 360 + ], + "blocks": [ + { + "bbox": [ + 304, + 270, + 545, + 304 + ], + "lines": [ + { + "bbox": [ + 304, + 270, + 545, + 304 + ], + "spans": [ + { + "bbox": [ + 304, + 270, + 545, + 304 + ], + "type": "text", + "content": "Table 5. Multi-view 3D visual grounding benchmark. \"Indep/Dep\" refer to \"View-Independent/Dependent\". Easy/Hard and Indep/Dep have a ratio of " + }, + { + "bbox": [ + 304, + 270, + 545, + 304 + ], + "type": "inline_equation", + "content": "80\\% /20\\%" + }, + { + "bbox": [ + 304, + 270, + 545, + 304 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 270, + 545, + 304 + ], + "type": "inline_equation", + "content": "78\\% /22\\%" + }, + { + "bbox": [ + 304, + 270, + 545, + 304 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 307, + 308, + 555, + 360 + ], + "lines": [ + { + "bbox": [ + 307, + 308, + 555, + 360 + ], + "spans": [ + { + "bbox": [ + 307, + 308, + 555, + 360 + ], + "type": "table", + "html": "
MethodsInputOverallEasyHardIndepDep
AP25AP25AP25AP25AP25
ScanRefer [8]RGB-D12.8513.789.1213.4410.77
BUTD-DETR [20]RGB-D22.1423.1218.2322.4720.98
L3Det [61]RGB-D23.0724.0118.3423.5921.22
OursRGB-D25.7227.1120.1226.3723.42
", + "image_path": "984d3a1af9b88e4bde83a69374418b06db0f86322102738c657daa7e43a743a3.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 365, + 545, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 365, + 545, + 413 + ], + "spans": [ + { + "bbox": [ + 304, + 365, + 545, + 413 + ], + "type": "text", + "content": "the depth sensor's weakness is enlarged. Meanwhile, depth plays a crucial role in predicting empty space, floor, and wall, while RGB information substantially improves prediction for most categories." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 414, + 545, + 473 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 414, + 545, + 473 + ], + "spans": [ + { + "bbox": [ + 304, + 414, + 545, + 473 + ], + "type": "text", + "content": "Multi-View 3D Perception. Unlike continuous settings, multi-view 3D perception does not predefine the order of views but provides all views to the model for scene-level results. This setting was studied previously [41], so we first reproduce common methods on our benchmark." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 474, + 545, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 474, + 545, + 605 + ], + "spans": [ + { + "bbox": [ + 304, + 474, + 545, + 605 + ], + "type": "text", + "content": "Multi-View 3D Object Detection. We implement baselines including ImVoxelNet [41] with RGB-only input and VoteNet [34] and FCAF3D [40] with depth-only input (Tab. 2). Additional dimensions are added to predict Euler angles with a simple L1 loss on their cosine values, but it yields overwhelming results. Substituting this with our decoder design markedly improves performance. Further using point cloud input painted for FCAF3D with RGB-D input slightly underperforms our baseline. Nevertheless, all models have substantial potential for improvement, demonstrating the challenges of this new dataset and setup." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 605, + 545, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 545, + 677 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 545, + 677 + ], + "type": "text", + "content": "Multi-View Semantic Occupancy Prediction. We implement two popular baselines from autonomous driving benchmarks, OccNet [43] and SurroundOcc [55] (similar to TPV-Former [19]). Their performance slightly lags behind our camera-only baseline. Variants of our baselines exhibit a performance trend akin to embodied benchmarks." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "content": "Monocular 3D Perception. Finally, the basic ego-centric setting is monocular 3D perception, specifically 3D detection, where each data sample comprises a single RGB-D" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19763" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 66, + 85, + 265, + 121 + ], + "blocks": [ + { + "bbox": [ + 85, + 71, + 248, + 82 + ], + "lines": [ + { + "bbox": [ + 85, + 71, + 248, + 82 + ], + "spans": [ + { + "bbox": [ + 85, + 71, + 248, + 82 + ], + "type": "text", + "content": "Table 6. Ablation with conventional settings." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 66, + 85, + 265, + 121 + ], + "lines": [ + { + "bbox": [ + 66, + 85, + 265, + 121 + ], + "spans": [ + { + "bbox": [ + 66, + 85, + 265, + 121 + ], + "type": "table", + "html": "
OrientedMulti-ViewAP25AR25AP50AR50
XX70.1790.4654.5875.66
✓X61.8790.3147.3073.93
✓✓59.9587.9243.3369.95
", + "image_path": "ad9063d0ac94eb16e13005cfc28ebf740f3e68e494fd8a61e14e650f174e6e19.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 70, + 138, + 261, + 174 + ], + "blocks": [ + { + "bbox": [ + 82, + 124, + 252, + 135 + ], + "lines": [ + { + "bbox": [ + 82, + 124, + 252, + 135 + ], + "spans": [ + { + "bbox": [ + 82, + 124, + 252, + 135 + ], + "type": "text", + "content": "Table 7. Real vs. rendered images on ScanNet." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 70, + 138, + 261, + 174 + ], + "lines": [ + { + "bbox": [ + 70, + 138, + 261, + 174 + ], + "spans": [ + { + "bbox": [ + 70, + 138, + 261, + 174 + ], + "type": "table", + "html": "
TrainValOverallHeadCommonTail
RenderRender22.1133.0116.446.74
RenderReal18.7227.0214.856.25
RealReal21.9832.9117.185.05
", + "image_path": "84d0e86953a09da5a112a8d1256ac16b08e03037868cdd63d0128bac978e562a.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 179, + 286, + 214 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 179, + 286, + 214 + ], + "spans": [ + { + "bbox": [ + 46, + 179, + 286, + 214 + ], + "type": "text", + "content": "frame and corresponding visible 3D boxes. Scan splits are used to extract frames as data samples, resulting in " + }, + { + "bbox": [ + 46, + 179, + 286, + 214 + ], + "type": "inline_equation", + "content": "689\\mathrm{k} / 115\\mathrm{k} / 86\\mathrm{k}" + }, + { + "bbox": [ + 46, + 179, + 286, + 214 + ], + "type": "text", + "content": " images for the training/validation/testing." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 215, + 286, + 322 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 215, + 286, + 322 + ], + "spans": [ + { + "bbox": [ + 46, + 215, + 286, + 322 + ], + "type": "text", + "content": "Monocular 3D Object Detection. This is more challenging than multi-view due to the absence of stereo geometric cues and truncated object views in indoor scenes, so the performance is significantly reduced in large-vocabulary settings. Hence, we first create a benchmark for 20 common categories (Tab. 4), observing a larger AP-AR gap for top methods because of difficulties predicting accurate 3D boxes from partial views. Similarly, our method outperforms others, providing a solid baseline for future studies." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 328, + 225, + 340 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 328, + 225, + 340 + ], + "spans": [ + { + "bbox": [ + 47, + 328, + 225, + 340 + ], + "type": "text", + "content": "5.2. Language-Grounded Benchmark" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 344, + 286, + 511 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 344, + 286, + 511 + ], + "spans": [ + { + "bbox": [ + 46, + 344, + 286, + 511 + ], + "type": "text", + "content": "Multi-View 3D Visual Grounding. Our benchmark introduces language into the perception loop to foster interactive 3D scene representation learning. With comprehensive instance annotations, our benchmark presents more complex prompts and grounding cases than previous works. As an initial step, this setup takes multi-view RGB-D images as input without considering differing prompt timestamps. The goal is to ground the object described by the language prompt in the scene using information from different egocentric views. Ground-truth detection boxes are not provided as candidates for grounding during evaluation, which can better validate end-to-end 3D visual grounding ability than the original SR3D [1]. Data sample splits align with previous benchmarks' 3D scan splits." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 512, + 286, + 631 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 512, + 286, + 631 + ], + "spans": [ + { + "bbox": [ + 46, + 512, + 286, + 631 + ], + "type": "text", + "content": "We reimplement classic methods like ScanRefer [8], BUTD-DETR [20], and L3Det [61] (Tab. 5). Our baseline outperforms all due to the strong multi-modal encoder. However, the performance remains much lower than previous works, partly due to changes in input format and annotations, which we will analyze next. Further challenges arise from handling more categories and small objects, making the grounding task more complex in parsing input prompts and predictions. Addressing these new challenges in this classical task would be promising for future research." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 637, + 110, + 649 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 637, + 110, + 649 + ], + "spans": [ + { + "bbox": [ + 47, + 637, + 110, + 649 + ], + "type": "text", + "content": "5.3. Analysis" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 654, + 286, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 654, + 286, + 677 + ], + "spans": [ + { + "bbox": [ + 46, + 654, + 286, + 677 + ], + "type": "text", + "content": "Finally, we make further analysis to connect EmbodiedScan to current progress in computer vision." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 677, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 677, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 677, + 286, + 713 + ], + "type": "text", + "content": "Axis-aligned vs. Oriented Boxes. We start with the 18-class detection performance of FCAF3D on ScanNet (Tab. 6). First, we change the annotations to oriented 3D" + } + ] + } + ], + "index": 11 + }, + { + "type": "table", + "bbox": [ + 326, + 87, + 523, + 130 + ], + "blocks": [ + { + "bbox": [ + 330, + 71, + 520, + 82 + ], + "lines": [ + { + "bbox": [ + 330, + 71, + 520, + 82 + ], + "spans": [ + { + "bbox": [ + 330, + 71, + 520, + 82 + ], + "type": "text", + "content": "Table 8. Benefits from training with EmbodiedScan." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 326, + 87, + 523, + 130 + ], + "lines": [ + { + "bbox": [ + 326, + 87, + 523, + 130 + ], + "spans": [ + { + "bbox": [ + 326, + 87, + 523, + 130 + ], + "type": "table", + "html": "
TrainValOverallHeadCommonTail
ScanNetScanNet20.2829.8115.576.40
OursScanNet23.0233.8218.096.57
ScanNetOurs10.9221.108.061.78
OursOurs16.8528.6512.837.09
", + "image_path": "e23737867c5a0a4bf9d8689637b9c8d393df91da5c477bcc4e4a2bf46028a6f4.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_body" + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 133, + 545, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 133, + 545, + 180 + ], + "spans": [ + { + "bbox": [ + 304, + 133, + 545, + 180 + ], + "type": "text", + "content": "boxes and adapt with our decoder. We find a significant drop in performance, indicating that the orientation estimation makes this task more challenging. We need to explore a better method to represent and predict the object pose." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 181, + 545, + 265 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 181, + 545, + 265 + ], + "spans": [ + { + "bbox": [ + 304, + 181, + 545, + 265 + ], + "type": "text", + "content": "Reconstructed Point Cloud vs. Multi-View RGB-D. Subsequently, replacing the reconstructed point clouds with the aggregated ones from multi-view depth maps has minor effects on " + }, + { + "bbox": [ + 304, + 181, + 545, + 265 + ], + "type": "inline_equation", + "content": "AP_{25}" + }, + { + "bbox": [ + 304, + 181, + 545, + 265 + ], + "type": "text", + "content": " but heavily impacts " + }, + { + "bbox": [ + 304, + 181, + 545, + 265 + ], + "type": "inline_equation", + "content": "AP_{50}" + }, + { + "bbox": [ + 304, + 181, + 545, + 265 + ], + "type": "text", + "content": " (Tab. 6), implying that the accuracy of reconstructed point clouds is superior to raw depth maps. Therefore, integrating reconstruction techniques in perception loops shows potential." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 266, + 545, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 266, + 545, + 312 + ], + "spans": [ + { + "bbox": [ + 304, + 266, + 545, + 312 + ], + "type": "text", + "content": "Next, we study the gap between the real and rendered images, and the benefits from training with EmbodiedScan. We do the comparison with our multi-modality baseline on the large-vocabulary multi-view 3D detection benchmark." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 314, + 545, + 397 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 314, + 545, + 397 + ], + "spans": [ + { + "bbox": [ + 304, + 314, + 545, + 397 + ], + "type": "text", + "content": "Real Capture vs. Rendering. As shown in Tab. 7, apart from the significant visual difference between real and rendered images (Fig. 1), the model's performance also has a remarkable decrease when transferring models trained with rendered images to the real world, particularly when the annotations are sufficient (5.99% AP drop on head categories and 5.89% AP lower than models trained with real images.)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 398, + 545, + 481 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 398, + 545, + 481 + ], + "spans": [ + { + "bbox": [ + 304, + 398, + 545, + 481 + ], + "type": "text", + "content": "Benefits from EmbodiedScan. Finally, we also quantitatively evaluate the benefits of training models with our large-scale EmbodiedScan (Tab. 8). As expected, when training our models with EmbodiedScan, we observed a significant improvement in both ScanNet (2.74% AP) and the overall validation split (5.93% AP), particularly 4.01% AP and 7.55% AP increase on head categories." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 306, + 490, + 378, + 501 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 490, + 378, + 501 + ], + "spans": [ + { + "bbox": [ + 306, + 490, + 378, + 501 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 507, + 545, + 686 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 507, + 545, + 686 + ], + "spans": [ + { + "bbox": [ + 304, + 507, + 545, + 686 + ], + "type": "text", + "content": "This paper introduces EmbodiedScan, a multi-modal perception suite aiming for language-grounded holistic 3D scene understanding from ego-centric views. We construct a large-scale dataset with diverse sensor data and multimodal annotations, including 3D oriented boxes, semantic occupancy and language descriptions. Based on this dataset, we propose a baseline framework capable of handling any number of views input, using a unified multimodal encoder and task-specific decoders. We establish benchmarks for basic and language-grounded 3D perception. Experiment results highlight our work's value and reveal new challenges in this setup. We believe EmbodiedScan can bring opportunities in embodied 3D perception and may also have a broader impact in related fields with the massive data and rich annotations." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "content": "Acknowledgements. This work is supported by Shanghai Artificial Intelligence Laboratory." + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19764" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 714 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 135 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 135 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 135 + ], + "type": "text", + "content": "[1] Panos Achlioptas, Ahmed Abdelreehm, Fei Xia, Mohamed Elhoseiny, and Leonidas Guibas. Referit3d: Neural listeners for fine-grained 3d object identification in real-world scenes. In European conference on computer vision, 2020. 3, 4, 8" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 137, + 288, + 213 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 137, + 288, + 213 + ], + "spans": [ + { + "bbox": [ + 53, + 137, + 288, + 213 + ], + "type": "text", + "content": "[2] Gilad Baruch, Zhuoyuan Chen, Afshin Dehghan, Tal Dimry, Yuri Feigin, Peter Fu, Thomas Gebauer, Brandon Joffe, Daniel Kurz, Arik Schwartz, and Elad Shulman. ARK-scenes - a diverse real-world dataset for 3d indoor scene understanding using mobile RGB-d data. In Thirty-fifth Conference on Neural Information Processing Systems Datasets and Benchmarks Track (Round 1), 2021. 2, 3, 4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 215, + 287, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 215, + 287, + 270 + ], + "spans": [ + { + "bbox": [ + 53, + 215, + 287, + 270 + ], + "type": "text", + "content": "[3] Garrick Brazil, Abhinav Kumar, Julian Straub, Nikhila Ravi, Justin Johnson, and Georgia Gkioxari. Omni3D: A large benchmark and model for 3D object detection in the wild. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2023. 2, 4, 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 271, + 287, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 271, + 287, + 326 + ], + "spans": [ + { + "bbox": [ + 53, + 271, + 287, + 326 + ], + "type": "text", + "content": "[4] Holger Caesar, Varun Bankiti, Alex H. Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuscenes: A multimodal dataset for autonomous driving. CoRR, abs/1903.11027, 2019. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 328, + 287, + 371 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 328, + 287, + 371 + ], + "spans": [ + { + "bbox": [ + 53, + 328, + 287, + 371 + ], + "type": "text", + "content": "[5] Anh-Quan Cao and Raoul de Charette. Monoscene: Monocular 3d semantic scene completion. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2022. 3, 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 373, + 287, + 428 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 373, + 287, + 428 + ], + "spans": [ + { + "bbox": [ + 53, + 373, + 287, + 428 + ], + "type": "text", + "content": "[6] Angel Chang, Angela Dai, Thomas Funkhouser, Maciej Halber, Matthias Niessner, Manolis Savva, Shuran Song, Andy Zeng, and Yinda Zhang. Matterport3D: Learning from RGB-D data in indoor environments. International Conference on 3D Vision (3DV), 2017. 1, 2, 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 430, + 287, + 496 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 430, + 287, + 496 + ], + "spans": [ + { + "bbox": [ + 53, + 430, + 287, + 496 + ], + "type": "text", + "content": "[7] Ming-Fang Chang, John Lambert, Patsorn Sangkloy, Jagjeet Singh, Slawomir Bak, Andrew Hartnett, De Wang, Peter Carr, Simon Lucey, Deva Ramanan, et al. Argoverse: 3d tracking and forecasting with rich maps. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2019. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 498, + 287, + 541 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 498, + 287, + 541 + ], + "spans": [ + { + "bbox": [ + 53, + 498, + 287, + 541 + ], + "type": "text", + "content": "[8] Dave Zhenyu Chen, Angel X Chang, and Matthias Nießner. Scanrefer: 3d object localization in rgb-d scans using natural language. In European conference on computer vision, 2020. 2, 3, 7, 8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 543, + 287, + 597 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 543, + 287, + 597 + ], + "spans": [ + { + "bbox": [ + 53, + 543, + 287, + 597 + ], + "type": "text", + "content": "[9] Zhenyu Chen, Ali Gholami, Matthias Nießner, and Angel X Chang. Scan2cap: Context-aware dense captioning in rgb-d scans. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3193-3203, 2021. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 599, + 287, + 654 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 599, + 287, + 654 + ], + "spans": [ + { + "bbox": [ + 48, + 599, + 287, + 654 + ], + "type": "text", + "content": "[10] Zhenyu Chen, Ronghang Hu, Xinlei Chen, Matthias Nießner, and Angel X Chang. Unit3d: A unified transformer for 3d dense captioning and visual grounding. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2023. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 656, + 287, + 700 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 656, + 287, + 700 + ], + "spans": [ + { + "bbox": [ + 48, + 656, + 287, + 700 + ], + "type": "text", + "content": "[11] Christopher Choy, JunYoung Gwak, and Silvio Savarese. 4d spatio-temporal convnets: Minkowski convolutional neural networks. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2019. 5" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 702, + 287, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 702, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 48, + 702, + 287, + 714 + ], + "type": "text", + "content": "[12] Camille Couprie, Clément Farabet, Laurent Najman, and" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 713 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "text", + "content": "Yann LeCun. Indoor semantic segmentation using depth information. arXiv preprint arXiv:1301.3572, 2013. 1, 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 96, + 547, + 151 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 96, + 547, + 151 + ], + "spans": [ + { + "bbox": [ + 307, + 96, + 547, + 151 + ], + "type": "text", + "content": "[13] Angela Dai, Angel X Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2017. 1, 2, 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 152, + 547, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 152, + 547, + 195 + ], + "spans": [ + { + "bbox": [ + 307, + 152, + 547, + 195 + ], + "type": "text", + "content": "[14] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018. 5" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 198, + 547, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 198, + 547, + 252 + ], + "spans": [ + { + "bbox": [ + 307, + 198, + 547, + 252 + ], + "type": "text", + "content": "[15] Runyu Ding, Jihan Yang, Chuhui Xue, Wenqing Zhang, Song Bai, and Xiaojuan Qi. Pla: Language-driven open-vocabulary 3d scene understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 254, + 547, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 254, + 547, + 297 + ], + "spans": [ + { + "bbox": [ + 307, + 254, + 547, + 297 + ], + "type": "text", + "content": "[16] Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for autonomous driving? the kitti vision benchmark suite. In IEEE Conference on Computer Vision and Pattern Recognition, 2012. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 299, + 545, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 299, + 545, + 342 + ], + "spans": [ + { + "bbox": [ + 307, + 299, + 545, + 342 + ], + "type": "text", + "content": "[17] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In IEEE Conference on Computer Vision and Pattern Recognition, 2016. 5" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 344, + 547, + 388 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 344, + 547, + 388 + ], + "spans": [ + { + "bbox": [ + 307, + 344, + 547, + 388 + ], + "type": "text", + "content": "[18] Shijia Huang, Yilun Chen, Jiaya Jia, and Liwei Wang. Multiview transformer for 3d visual grounding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 389, + 547, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 389, + 547, + 443 + ], + "spans": [ + { + "bbox": [ + 307, + 389, + 547, + 443 + ], + "type": "text", + "content": "[19] Yuanhui Huang, Wenzhao Zheng, Yunpeng Zhang, Jie Zhou, and Jiwen Lu. Tri-perspective view for vision-based 3d semantic occupancy prediction. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2023. 3, 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 445, + 545, + 489 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 445, + 545, + 489 + ], + "spans": [ + { + "bbox": [ + 307, + 445, + 545, + 489 + ], + "type": "text", + "content": "[20] Ayush Jain, Nikolaos Gkanatsios, Ishita Mediratta, and Katerina Fragkiadaki. Bottom up top down detection transformers for language grounding in images and point clouds. In European Conference on Computer Vision, 2022. 3, 6, 7, 8" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 491, + 545, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 491, + 545, + 555 + ], + "spans": [ + { + "bbox": [ + 307, + 491, + 545, + 555 + ], + "type": "text", + "content": "[21] Mukul Khanna*, Yongsen Mao*, Hanxiao Jiang, Sanjay Haresh, Brennan Shacklett, Dhruv Batra, Alexander Clegg, Eric Undersander, Angel X. Chang, and Manolis Savva. Habitat Synthetic Scenes Dataset (HSSD-200): An Analysis of 3D Scene Scale and Realism Tradeoffs for ObjectGoal Navigation. arXiv preprint, 2023. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 557, + 545, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 557, + 545, + 622 + ], + "spans": [ + { + "bbox": [ + 307, + 557, + 545, + 622 + ], + "type": "text", + "content": "[22] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C. Berg, Wan-Yen Lo, Piotr Dollar, and Ross Girshick. Segment anything. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2023. 2, 4" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 624, + 545, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 624, + 545, + 668 + ], + "spans": [ + { + "bbox": [ + 307, + 624, + 545, + 668 + ], + "type": "text", + "content": "[23] Alex H. Lang, Sourabh Vora, Holger Caesar, Lubing Zhou, Jiong Yang, and Oscar Beijbom. Pointpillars: Fast encoders for object detection from point clouds. In IEEE Conference on Computer Vision and Pattern Recognition, 2019. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 670, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 670, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 670, + 545, + 713 + ], + "type": "text", + "content": "[24] E Li, Shuaijun Wang, Chengyang Li, Dachuan Li, Xiangbin Wu, and Qi Hao. Sustech points: A portable 3d point cloud interactive annotation platform system. In 2020 IEEE Intelligent Vehicles Symposium (IV), 2020. 4" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "19765" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 714 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 288, + 139 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 288, + 139 + ], + "type": "text", + "content": "[25] Yiming Li, Zhiding Yu, Christopher Choy, Chaowei Xiao, Jose M Alvarez, Sanja Fidler, Chen Feng, and Anima Anandkumar. Voxformer: Sparse voxel transformer for camera-based 3d semantic scene completion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 140, + 288, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 140, + 288, + 195 + ], + "spans": [ + { + "bbox": [ + 48, + 140, + 288, + 195 + ], + "type": "text", + "content": "[26] Tingting Liang, Hongwei Xie, Kaicheng Yu, Zhongyu Xia, Zhiwei Lin, Yongtao Wang, Tao Tang, Bing Wang, and Zhi Tang. Bevfusion: A simple and robust lidar-camera fusion framework. Advances in Neural Information Processing Systems, 2022. 3, 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 197, + 288, + 241 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 197, + 288, + 241 + ], + "spans": [ + { + "bbox": [ + 48, + 197, + 288, + 241 + ], + "type": "text", + "content": "[27] Tsung-Yi Lin, Piotr Dólar, Ross Girshick, Kaiming He, Bharath Hariharan, and Serge Belongie. Feature pyramid networks for object detection. In IEEE Conference on Computer Vision and Pattern Recognition, 2017. 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 242, + 288, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 242, + 288, + 297 + ], + "spans": [ + { + "bbox": [ + 48, + 242, + 288, + 297 + ], + "type": "text", + "content": "[28] Zechen Liu, Zizhang Wu, and Roland Toth. Smoke: Single-stage monocular 3d object detection via keypoint estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, pages 996-997, 2020. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 298, + 288, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 298, + 288, + 342 + ], + "spans": [ + { + "bbox": [ + 48, + 298, + 288, + 342 + ], + "type": "text", + "content": "[29] Ze Liu, Zheng Zhang, Yue Cao, Han Hu, and Xin Tong. Group-free 3d object detection via transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2021. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 343, + 288, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 343, + 288, + 398 + ], + "spans": [ + { + "bbox": [ + 48, + 343, + 288, + 398 + ], + "type": "text", + "content": "[30] Zhijian Liu, Haotian Tang, Alexander Amini, Xinyu Yang, Huizi Mao, Daniela L Rus, and Song Han. Bevfusion: Multi-task multi-sensor fusion with unified bird's-eye view representation. In 2023 IEEE International Conference on Robotics and Automation (ICRA), 2023. 3, 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 399, + 288, + 454 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 399, + 288, + 454 + ], + "spans": [ + { + "bbox": [ + 48, + 399, + 288, + 454 + ], + "type": "text", + "content": "[31] Yuheng Lu, Chenfeng Xu, Xiaobao Wei, Xiaodong Xie, Masayoshi Tomizuka, Kurt Keutzer, and Shanghang Zhang. Open-vocabulary 3d detection via image-level class and de-biased cross-modal contrastive learning. arXiv preprint arXiv:2207.01987, 2022. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 456, + 288, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 456, + 288, + 510 + ], + "spans": [ + { + "bbox": [ + 48, + 456, + 288, + 510 + ], + "type": "text", + "content": "[32] Jiageng Mao, Minzhe Niu, Chenhan Jiang, Hanxue Liang, Jingheng Chen, Xiaodan Liang, Yamin Li, Chaoqiang Ye, Wei Zhang, Zhenguo Li, et al. One million scenes for autonomous driving: Once dataset. arXiv preprint arXiv:2106.11037, 2021. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 512, + 288, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 512, + 288, + 567 + ], + "spans": [ + { + "bbox": [ + 48, + 512, + 288, + 567 + ], + "type": "text", + "content": "[33] Songyou Peng, Kyle Genova, Chiyu Jiang, Andrea Tagliasacchi, Marc Pollefeys, Thomas Funkhouser, et al. Openscene: 3d scene understanding with open vocabularies. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 568, + 288, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 568, + 288, + 612 + ], + "spans": [ + { + "bbox": [ + 48, + 568, + 288, + 612 + ], + "type": "text", + "content": "[34] Charles R Qi, Or Litany, Kaiming He, and Leonidas J Guibas. Deep hough voting for 3d object detection in point clouds. In proceedings of the IEEE/CVF International Conference on Computer Vision, 2019. 1, 3, 4, 6, 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 613, + 288, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 613, + 288, + 667 + ], + "spans": [ + { + "bbox": [ + 48, + 613, + 288, + 667 + ], + "type": "text", + "content": "[35] Charles R Qi, Xinlei Chen, Or Litany, and Leonidas J Guibas. Imvotenet: Boosting 3d object detection in point clouds with image votes. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2020. 3, 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 669, + 288, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 669, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 48, + 669, + 288, + 714 + ], + "type": "text", + "content": "[36] Santhosh K Ramakrishnan, Aaron Gokaslan, Erik Wijmans, Oleksandr Maksymets, Alex Clegg, John Turner, Eric Undersander, Wojciech Galuba, Andrew Westbury, Angel X Chang, et al. Habitat-matterport 3d dataset (hm3d): 1000" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 714 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "type": "text", + "content": "large-scale 3d environments for embodied ai. arXiv preprint arXiv:2109.08238, 2021. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 307, + 97, + 547, + 141 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 97, + 547, + 141 + ], + "spans": [ + { + "bbox": [ + 307, + 97, + 547, + 141 + ], + "type": "text", + "content": "[37] Nils Reimers and Iryna Gurevych. Sentence-bert: Sentence embeddings using siamese bert-networks. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing, 2019. 4" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 143, + 545, + 209 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 143, + 545, + 209 + ], + "spans": [ + { + "bbox": [ + 307, + 143, + 545, + 209 + ], + "type": "text", + "content": "[38] Mike Roberts, Jason Ramapuram, Anurag Ranjan, Atulit Kumar, Miguel Angel Bautista, Nathan Paczan, Russ Webb, and Joshua M Susskind. Hypersim: A photorealistic synthetic dataset for holistic indoor scene understanding. In Proceedings of the IEEE/CVF international conference on computer vision, 2021. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 211, + 545, + 255 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 211, + 545, + 255 + ], + "spans": [ + { + "bbox": [ + 307, + 211, + 545, + 255 + ], + "type": "text", + "content": "[39] David Rozenberszki, Or Litany, and Angela Dai. Language-grounded indoor 3d semantic segmentation in the wild. In Proceedings of the European Conference on Computer Vision (ECCV), 2022. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 257, + 545, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 257, + 545, + 300 + ], + "spans": [ + { + "bbox": [ + 307, + 257, + 545, + 300 + ], + "type": "text", + "content": "[40] Danila Rukhovich, Anna Vorontsova, and Anton Konushin. Fcaf3d: Fully convolutional anchor-free 3d object detection. In European Conference on Computer Vision, 2022. 1, 3, 6, 7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 303, + 545, + 347 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 303, + 545, + 347 + ], + "spans": [ + { + "bbox": [ + 307, + 303, + 545, + 347 + ], + "type": "text", + "content": "[41] Danila Rukhovich, Anna Vorontsova, and Anton Konushin. Imvoxelnet: Image to voxels projection for monocular and multi-view general-purpose 3d object detection. In WACV, pages 2397-2406, 2022. 3, 6, 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 349, + 545, + 393 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 349, + 545, + 393 + ], + "spans": [ + { + "bbox": [ + 307, + 349, + 545, + 393 + ], + "type": "text", + "content": "[42] Shaoshuai Shi, Xiaogang Wang, and Hongsheng Li. Poincaré: 3d object proposal generation and detection from point cloud. In IEEE Conference on Computer Vision and Pattern Recognition, 2019. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 395, + 545, + 450 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 395, + 545, + 450 + ], + "spans": [ + { + "bbox": [ + 307, + 395, + 545, + 450 + ], + "type": "text", + "content": "[43] Chonghao Sima, Wenwen Tong, Tai Wang, Li Chen, Silei Wu, Hanming Deng, Yi Gu, Lewei Lu, Ping Luo, Dahua Lin, and Hongyang Li. Scene as occupancy. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2023. 3, 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 452, + 545, + 496 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 452, + 545, + 496 + ], + "spans": [ + { + "bbox": [ + 307, + 452, + 545, + 496 + ], + "type": "text", + "content": "[44] Andrea Simonelli, Samuel Rota Rota Bulò, Lorenzo Porzi, Manuel López-Antequera, and Peter Kontschieder. Disentangling monocular 3d object detection. In IEEE International Conference on Computer Vision, 2019. 3, 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 498, + 545, + 543 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 498, + 545, + 543 + ], + "spans": [ + { + "bbox": [ + 307, + 498, + 545, + 543 + ], + "type": "text", + "content": "[45] Shuran Song, Samuel P Lichtenberg, and Jianxiong Xiao. Sun rgb-d: A rgb-d scene understanding benchmark suite. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2015. 1, 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 544, + 545, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 544, + 545, + 599 + ], + "spans": [ + { + "bbox": [ + 307, + 544, + 545, + 599 + ], + "type": "text", + "content": "[46] Julian Straub, Thomas Whelan, Lingni Ma, Yufan Chen, Erik Wijmans, Simon Green, Jakob J Engel, Raul Mur-Artal, Carl Ren, Shobhit Verma, et al. The replica dataset: A digital replica of indoor spaces. arXiv preprint arXiv:1906.05797, 2019. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 601, + 545, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 601, + 545, + 667 + ], + "spans": [ + { + "bbox": [ + 307, + 601, + 545, + 667 + ], + "type": "text", + "content": "[47] Pei Sun, Henrik Kretzschmar, Xerxes Dotiwalla, Aurelien Chouard, Vijaysai Patnaik, Paul Tsui, James Guo, Yin Zhou, Yuning Chai, Benjamin Caine, et al. Scalability in perception for autonomous driving: Waymo open dataset. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2020. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 669, + 545, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 669, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 307, + 669, + 545, + 714 + ], + "type": "text", + "content": "[48] Ayca Takmaz, Elisabetta Fedele, Robert W Sumner, Marc Pollefeys, Federico Tombari, and Francis Engelmann. Openmask3d: Open-vocabulary 3d instance segmentation. arXiv preprint arXiv:2306.13631, 2023. 3" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "19766" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 654 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "text", + "content": "[49] Xiaoyu Tian, Tao Jiang, Longfei Yun, Yue Wang, Yilun Wang, and Hang Zhao. Occ3d: A large-scale 3d occupancy prediction benchmark for autonomous driving. arXiv preprint arXiv:2304.14365, 2023. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 118, + 287, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 118, + 287, + 162 + ], + "spans": [ + { + "bbox": [ + 49, + 118, + 287, + 162 + ], + "type": "text", + "content": "[50] Sourabh Vora, Alex H Lang, Bassam Helou, and Oscar Beijbom. Pointpainting: Sequential fusion for 3d object detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2020. 3, 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 163, + 287, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 163, + 287, + 217 + ], + "spans": [ + { + "bbox": [ + 49, + 163, + 287, + 217 + ], + "type": "text", + "content": "[51] Johanna Wald, Armen Avetisyan, Nassir Navab, Federico Tombari, and Matthias Nießner. Rio: 3d object instance re-localization in changing indoor environments. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2019. 1, 2, 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 219, + 287, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 219, + 287, + 271 + ], + "spans": [ + { + "bbox": [ + 49, + 219, + 287, + 271 + ], + "type": "text", + "content": "[52] Tai Wang, Xinge Zhu, Jiangmiao Pang, and Dahua Lin. FCOS3D: Fully convolutional one-stage monocular 3d object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) Workshops, 2021. 3, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 49, + 274, + 287, + 307 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 274, + 287, + 307 + ], + "spans": [ + { + "bbox": [ + 49, + 274, + 287, + 307 + ], + "type": "text", + "content": "[53] Tai Wang, Jiangmiao Pang, and Dahua Lin. Monocular 3d object detection with depth from motion. In European Conference on Computer Vision (ECCV), 2022." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 49, + 308, + 287, + 350 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 308, + 287, + 350 + ], + "spans": [ + { + "bbox": [ + 49, + 308, + 287, + 350 + ], + "type": "text", + "content": "[54] Tai Wang, ZHU Xinge, Jiangmiao Pang, and Dahua Lin. Probabilistic and geometric depth: Detecting objects in perspective. In Conference on Robot Learning, pages 1475-1485. PMLR, 2022. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 49, + 353, + 287, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 353, + 287, + 407 + ], + "spans": [ + { + "bbox": [ + 49, + 353, + 287, + 407 + ], + "type": "text", + "content": "[55] Yi Wei, Linqing Zhao, Wenzhao Zheng, Zheng Zhu, Jie Zhou, and Jiwen Lu. Surroundocc: Multi-camera 3d occupancy prediction for autonomous driving. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2023. 3, 6, 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 49, + 409, + 287, + 474 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 409, + 287, + 474 + ], + "spans": [ + { + "bbox": [ + 49, + 409, + 287, + 474 + ], + "type": "text", + "content": "[56] Karmesh Yadav, Ram Ramrakhya, Santhosh Kumar Ramakrishnan, Theo Gervet, John Turner, Aaron Gokaslan, Noah Maestre, Angel Xuan Chang, Dhruv Batra, Manolis Savva, et al. Habitat-matterport 3d semantics dataset. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 49, + 476, + 287, + 496 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 476, + 287, + 496 + ], + "spans": [ + { + "bbox": [ + 49, + 476, + 287, + 496 + ], + "type": "text", + "content": "[57] Yan Yan, Yuxing Mao, and Bo Li. Second: Sparsely embedded convolutional detection. Sensors, 18(10), 2018. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 49, + 498, + 287, + 531 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 498, + 287, + 531 + ], + "spans": [ + { + "bbox": [ + 49, + 498, + 287, + 531 + ], + "type": "text", + "content": "[58] Wenwei Zhang, Zhe Wang, and Chen Change Loy. Exploring data augmentation for multi-modality 3d object detection. arXiv preprint arXiv:2012.12741, 2020. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 49, + 533, + 287, + 564 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 533, + 287, + 564 + ], + "spans": [ + { + "bbox": [ + 49, + 533, + 287, + 564 + ], + "type": "text", + "content": "[59] Yin Zhou and Oncel Tuzel. Voxelnet: End-to-end learning for point cloud based 3d object detection. In IEEE Conference on Computer Vision and Pattern Recognition, 2018. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 49, + 567, + 287, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 567, + 287, + 609 + ], + "spans": [ + { + "bbox": [ + 49, + 567, + 287, + 609 + ], + "type": "text", + "content": "[60] Yi Zhou, Connelly Barnes, Jingwan Lu, Jimei Yang, and Hao Li. On the continuity of rotation representations in neural networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2019. 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 49, + 611, + 287, + 654 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 611, + 287, + 654 + ], + "spans": [ + { + "bbox": [ + 49, + 611, + 287, + 654 + ], + "type": "text", + "content": "[61] Chenming Zhu, Wenwei Zhang, Tai Wang, Xihui Liu, and Kai Chen. Object2scene: Putting objects in context for open-vocabulary 3d detection. arXiv preprint arXiv:2309.09456, 2023. 3, 6, 7, 8" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19767" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Embracing Unimodal Aleatoric Uncertainty for Robust Multimodal Fusion/b864c4f0-5e7f-441f-a2a0-028202223b38_content_list.json b/2024/Embracing Unimodal Aleatoric Uncertainty for Robust Multimodal Fusion/b864c4f0-5e7f-441f-a2a0-028202223b38_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..f021527167c33a5742d49429494a93318dea2780 --- /dev/null +++ b/2024/Embracing Unimodal Aleatoric Uncertainty for Robust Multimodal Fusion/b864c4f0-5e7f-441f-a2a0-028202223b38_content_list.json @@ -0,0 +1,1536 @@ +[ + { + "type": "text", + "text": "Embracing Unimodal Aleatoric Uncertainty for Robust Multimodal Fusion", + "text_level": 1, + "bbox": [ + 107, + 130, + 864, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Zixian Gao $^{1*}$ , Xun Jiang $^{1*}$ , Xing Xu $^{1,2\\dagger}$ , Fumin Shen $^{1}$ , Yujie Li $^{3}$ , Heng Tao Shen $^{1,2}$", + "bbox": [ + 165, + 179, + 803, + 199 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1Center for Future Media & School of Computer Science and Engineering, University of Electronic Science and Technology of China, China", + "bbox": [ + 186, + 199, + 776, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ College of Electronic and Information Engineering, Tongji University, China", + "bbox": [ + 178, + 233, + 790, + 251 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{3}$ Kyushu Institute of Technology, Japan", + "bbox": [ + 326, + 251, + 638, + 268 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 304, + 312, + 319 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "As a fundamental problem in multimodal learning, multimodal fusion aims to compensate for the inherent limitations of a single modality. One challenge of multimodal fusion is that the unimodal data in their unique embedding space mostly contains potential noise, which leads to corrupted cross-modal interactions. However, in this paper, we show that the potential noise in unimodal data could be well quantified and further employed to enhance more stable unimodal embeddings via contrastive learning. Specifically, we propose a novel generic and robust multimodal fusion strategy, termed Embracing Aleatoric Uncertainty (EAU), which is simple and can be applied to kinds of modalities. It consists of two key steps: (1) the Stable Unimodal Feature Augmentation (SUFA) that learns a stable unimodal representation by incorporating the aleatoric uncertainty into self-supervised contrastive learning. (2) Robust Multimodal Feature Integration (RMFI) leveraging an information-theoretic strategy to learn a robust compact joint representation. We evaluate our proposed EAU method on five multimodal datasets, where the video, RGB image, text, audio, and depth image are involved. Extensive experiments demonstrate the EAU method is more noise-resistant than existing multimodal fusion strategies and establishes new state-of-the-art on several benchmarks.", + "bbox": [ + 75, + 335, + 473, + 700 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 729, + 209, + 744 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "By exploring the complementary information from different modalities, multimodal learning has achieved impressive success in kinds of artificial intelligence applications, such as multimodal image classification in Internet applications [1-5] and sentiment analysis in intelligent robotics [6-9]. It has been widely proved that fusing information from different modalities appropriately is helpful to gain", + "bbox": [ + 75, + 756, + 468, + 863 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/e36592c270ca40b909a9bc8637326a72e660b6515055de806100dfa06f72618b.jpg", + "image_caption": [ + "(a) Uncertainties from fuzzy semantics in multimodal data" + ], + "image_footnote": [], + "bbox": [ + 504, + 303, + 890, + 395 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/5c9d5c6423030f4a3c756a370d0827832c115c907a91f48c646f6e863cca5ff2.jpg", + "image_caption": [ + "(b) Multivariate Gaussian distributional representations for fuzzy semantics", + "Figure 1. An illustration of the aleatoric uncertainty in multimodal datasets and the distributional representations: (a) As the semantics are ambiguous, the multimodal data are prone to introducing noisy data. (b) We adopt a multivariate Gaussian distribution to represent the fuzzy semantics in a noisy latent space." + ], + "image_footnote": [], + "bbox": [ + 519, + 412, + 887, + 515 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "better performance. Dedicated to this, multimodal fusion has become an emerging challenge in multimodal learning, which integrates different modalities into a unified representation with powerful neural networks.", + "bbox": [ + 496, + 626, + 892, + 686 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Nevertheless, with the studies developing, researchers have recognized that the multimodal data in different modal forms may be unreliable due to the unique noise in their own modal space. Recently, several works [5, 10-13] show that previous multimodal fusion methods have overlooked the unreliable quality of multimodal data. Specifically, the widely adopted fusion strategies may fail on noisy multimodal data because the cross-modal interactions may obtain limited influence by the uncertainty in data. Here we illustrate typical noise in multimodal data in Fig. 1. Due to the semantics of \"Positive\" being fuzzy and the labels being judged by humans subjectively, the given image-text pairs are all recognized as positive in sentiment analysis, even though there is obvious noise in both image and text", + "bbox": [ + 496, + 689, + 893, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*These authors contributed equally to this work.", + "bbox": [ + 94, + 875, + 351, + 887 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "† Corresponding author.", + "bbox": [ + 96, + 887, + 220, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "26876", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "modalities. In general, we can say the introduced noise comes from the aleatoric uncertainty [14], which undermines the effectiveness of exploiting multimodal data and demonstrates multimodal learning is not a free lunch.", + "bbox": [ + 75, + 90, + 468, + 150 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Hence, with the emerging challenge of uncertainty in multimodal data, we raise the first fundamental question: Can we quantify the uncertainty in multimodal data? Inspired by probability Distributional Representation [15-18], we naturally adopt the Gaussian distributions to work out this problem. We assume that each instance can be represented as a multivariate normal distribution, where the variance can be regarded as the intrinsic aleatoric uncertainty. With the quantified uncertainty, we are allowed to take a closer look at the multimodal fusion and raise the second question: Is it appropriate to completely drop the intrinsic uncertainty? From the examples in Fig. 1(a), we can observe that even if image-text pairs reveal similar semantics, the aleatoric uncertainty is still unavoidable because of the domain shift, extra descriptions, or image quality, etc. To this end, as is illustrated in Fig. 1(b), we argue that a multivariate normal distribution that considers aleatoric uncertainty can be regarded as a fuzzy representation for semantics, where the semantic-related data are within similar distributions, even if they are in different modalities. Motivated by the two assumptions, we develop a novel multimodal fusion strategy in this paper, i.e., Embracing Aleatoric Uncertainty (EAU).", + "bbox": [ + 75, + 152, + 472, + 496 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Specifically, our EAU method consists of the following two processes: (1) The Stable Unimodal Feature Augmentation (SUFA) quantifies the intrinsic aleatoric uncertainty in each modality by representing samples as multivariate normal distributions. Then, we further conduct self-supervised contrastive learning with these distributional representations to learn stable unimodal embeddings with better semantical consistency. (2) Robust Multimodal Feature Integration (RMFI) that dynamically fuses the stable unimodal embeddings into a joint representation. Particularly, considering that SUFA focuses on semantic consistency only and ignores the problem of information redundancy, we employ an information-theoretic strategy, i.e., Variational Information Bottleneck [19, 20], to learn a compact joint representation with less redundancy. We evaluate our method on five multimodal benchmark datasets and demonstrate it outperforms existing state-of-the-art methods on both Multimodal Sentiment Analysis and Multimodal Image Classification tasks. Moreover, our proposed EAU method also reveals better robustness on noisy datasets compared with the counterpart methods [2, 3, 5].", + "bbox": [ + 75, + 498, + 472, + 815 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Overall, our contributions in this paper can be summarized as three-fold:", + "bbox": [ + 76, + 816, + 468, + 845 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "- We propose a novel multimodal fusion method termed Embracing Aleatoric Uncertainty (EAU), which quantifies the intrinsic aleatoric uncertainty and leverages it", + "bbox": [ + 94, + 854, + 468, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "to learn stable and robust joint representations.", + "bbox": [ + 531, + 90, + 841, + 107 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We devise the Stable Unimodal Feature Augmentation (SUFA) module, which quantifies aleatoric uncertainty and learns stable unimodal representations with better semantical consistency.", + "- We design the Robust Multimodal Feature Integration (RMFI) module to learn compact joint representations with less redundancy, which further improves the robustness of multimodal fusion in our method." + ], + "bbox": [ + 517, + 114, + 890, + 243 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 256, + 640, + 272 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Multimodal Fusion. Multimodal fusion, which aims at learning stronger representations from different modalities, has become an essential part of a spectrum of computer vision research, such as vision-audio learning [21,22], vision-language learning [23, 24], image retrieval [25-27], and video understanding [28-30]. Generally, multimodal fusion can be categorized into three types: early fusion, intermediate fusion, and late fusion. The previous multimodal fusion methods can be categorized into early fusion and late fusion in terms of their feature-level or decision-level fusion operations. In the last decades, massive works with deep learning suggest that intermediate fusion [13, 29, 31], which learns unified embeddings for multimodal data in the hidden layers of neural networks, could benefit representation learning. Although these multimodal fusion methods achieve remarkable improvements in kinds of multimodal learning tasks, most of them ignore the uncertainty in multimodal learning. Recently, several works [5, 10-13] empirically or theoretically demonstrated that conventional multimodal fusion methods had limited performance and robustness on noisy or corrupted multimodal data. Inspired by these pioneering works, we devised a novel intermediate multimodal fusion method, which could quantify the uncertainty and learn more robust joint representations.", + "bbox": [ + 496, + 281, + 890, + 643 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Uncertainty in Deep Learning. In general, the uncertainties in deep learning can be categorized into epistemic uncertainty and aleatoric uncertainty [14]. The former aims at capturing the noise of the parameters in deep neural networks, while the latter measures the noise inherent in given training data. To improve the robustness and generalization of open-world scenarios, many researchers incorporate the uncertainty estimation into the deep learning models for computer vision tasks, such as face recognition [15, 16], semantic segmentation [32], and action localization [33, 34]. There is also a list of works that incorporate uncertainty estimation into multimodal learning tasks [5, 17, 22, 35]. However, most of them only quantified uncertainties to learn probability distribution representations that are used for cross-modal interactions, but ignore the value of intrinsic uncertainties and hardly take a further analysis on the robustness of multimodal fusion. Following this observation,", + "bbox": [ + 496, + 643, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "26877", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "we develop this work to learn more robust compact joint representations for multimodal learning, which embraces the uncertainties via self-supervised contrastive learning.", + "bbox": [ + 75, + 90, + 468, + 136 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Information Bottleneck Theory. The Information Bottleneck Theory [19] was initially formulated in the context of signal processing, which is proposed to discover a more concise representation of a signal while retaining its utmost informative content. Alemi et al. [36] proposed the Variational Information Bottleneck (VIB) to bridge the gap between Information Bottleneck Theory and deep learning, which approximates the information bottleneck constraints and enables it in deep learning. Based on the VIB, massive researchers introduce the Information Bottleneck Theory in deep learning models to tackle kinds of computer vision tasks, such as object detection [37-39] or image classification [37,38]. Furthermore, as the VIB aims at learning compact minimal representations, it also attracts wide attention from the multimodal learning field [28, 31]. For example, Mai et al. [28] devised a multimodal information bottleneck to learn the minimal sufficient unimodal and multimodal representations. Inspired by their successes, we also devise an multimodal integration strategy based on VIB to reduce the redundancy in joint representations, which further improves the robustness and effectiveness.", + "bbox": [ + 75, + 138, + 470, + 455 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Proposed Method", + "text_level": 1, + "bbox": [ + 76, + 479, + 250, + 497 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Preliminaries", + "text_level": 1, + "bbox": [ + 76, + 508, + 217, + 523 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Uncertainty Estimation in Deep Learning Models. In terms of [14], the uncertainties in deep learning models can be categorized into aleatoric uncertainty and epistemic uncertainty. The former refers to the uncertainty inherent in the observations and can not be explained away with more data, while the latter only accounts for uncertainty in the model and can be explained away given enough data. Typically, given a deep learning model $f_{\\theta}(\\cdot)$ , it conducts the mapping $\\mathcal{X} \\rightarrow \\mathcal{Y}$ , where $\\mathcal{X}$ and $\\mathcal{Y}$ represent input data and observed label space respectively. The aleatoric uncertainty in $\\mathcal{X}$ will lead to a corrupted predicted results in $\\mathcal{Y}$ . To estimate the uncertainties in deep learning models, a widely adopted strategy is sampling $N$ weights $\\theta \\sim p(\\theta | x, y)$ , $x \\in \\mathcal{X}, y \\in \\mathcal{Y}$ via Dropout operations for different predictions for mean $\\mu$ and variance $\\sigma^2$ , and adopting a Gaussian posterior $p(y|x) \\sim \\mathcal{N}(\\mu_{*}(x), \\sigma_{*}^{2}(x))$ . Here we can get the predictive variance and extract the aleatoric uncertainty which represents the ambiguity in $x$ :", + "bbox": [ + 75, + 535, + 468, + 808 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mu_ {*} ^ {2} (x) = \\frac {1}{N} \\sum_ {i} \\mu_ {i} ^ {2} (x), \\\\ \\sigma_ {*} ^ {2} (x) = \\frac {1}{N} \\sum_ {i} \\sigma_ {i} ^ {2} (x) + \\frac {1}{N} \\sum_ {i} \\mu_ {i} ^ {2} (x) - \\mu_ {*} ^ {2} (x) \\tag {1} \\\\ = \\mathbb {E} _ {i} \\left[ \\sigma_ {i} ^ {2} (x) \\right] + \\operatorname {V a r} _ {i} [ \\mu_ {i} (x) ], \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 96, + 824, + 468, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mathbb{E}_i\\left[\\sigma_i^2 (x)\\right]$ represents the aleatoric uncertainty, while $\\operatorname {Var}_i[\\mu_i(x)]$ is the epistemic uncertainty, which attributes to the model $f_{\\theta}(\\cdot)$ rather than input data $\\mathcal{X}$ . In our work, we pay more attention to the aleatoric uncertainty since our goal is to quantify the intrinsic noise in multimodal data and leverage it for learning better joint representations.", + "bbox": [ + 496, + 89, + 890, + 181 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Variational Information Bottleneck. The Variational Information Bottleneck (VIB) [20] is an information-theoretic strategy widely adopted in deep learning models, which aims to maintain maximal discriminative feature representations with minimal redundancy. Specifically, given the input variable $\\mathbf{x}$ with noise or redundant information and the target variable $\\mathbf{y}$ , the VIB is to learn the compressed latent variable $\\mathbf{z}$ , and $\\mathbf{z}$ is maximally discriminative about the target variable $\\mathbf{y}$ . Moreover, since the input $\\mathbf{x}$ is noisy and redundant, the VIB also requires $\\mathbf{z}$ to be minimally discriminative about the original variable $\\mathbf{x}$ . In our work, we leverage the VIB to learn compact joint representations, which overcome the redundancy brought by highly aligned distributional multimodal representations to improve the robustness of multimodal fusion.", + "bbox": [ + 496, + 181, + 892, + 407 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/7235b48e125f760e7c4f14f3b2d7e2e6e2174341886e5a08e358ffdf17457e53.jpg", + "image_caption": [ + "Figure 2. An illustration of the Unimodal Distributional Representation process. For clarity, we visualize the multivariate Gaussian distributions with two variables only." + ], + "image_footnote": [], + "bbox": [ + 501, + 419, + 890, + 522 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Stable Unimodal Feature Augmentation", + "text_level": 1, + "bbox": [ + 500, + 590, + 841, + 606 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "As the data in different modalities contain unique noise in their own modal space, we first propose the Stable Unimodal Feature Augmentation (SUFA) module to quantify their intrinsic aleatoric uncertainty. Following the argument that the uncertainty is attributed to the fuzzy representation of semantics, we further leverage the aleatoric uncertainty via self-supervised contrastive learning.", + "bbox": [ + 496, + 613, + 890, + 718 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Unimodal Distributional Representation. Given the multimodal samples $\\mathbf{x}_i^m$ , $m \\in \\mathcal{M}$ where $\\mathcal{M}$ is the modality set including images, audio, text, etc., we learn the distributional representations to quantify the aleatoric uncertainty in each modality. According to the deviation listed in Eq. 1, we can observe that the aleatoric uncertainty can be predicted with deep learning models $f_{\\theta}(\\cdot)$ as the variance $\\sigma_i^{m2}$ directly if the epistemic uncertainty is not considered. To this end, we first employ the corresponding feature extractor to learn the preliminary embeddings for each modality, then deploy two additional fully connected layers to learn a mean vector $\\mu_i^{m2} \\in \\mathbb{R}^d$ and a variance vector $\\sigma_i^{m2} \\in \\mathbb{R}^d$ .", + "bbox": [ + 496, + 719, + 892, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "26878", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Furtherly, we define the representation $\\mathbf{z}_i^m$ in latent space of each sample $\\mathbf{x}_i^m$ as a multivariate Gaussian distribution with $d$ variable, which can be represented as:", + "bbox": [ + 76, + 90, + 468, + 137 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} p (\\mathbf {z} _ {i} ^ {m} | \\mathbf {x} _ {i} ^ {m}) \\sim \\mathcal {N} (\\boldsymbol {\\mu} _ {i} ^ {m}, \\boldsymbol {\\sigma} _ {i} ^ {m 2} \\mathbf {I}), \\\\ \\pmb {\\mu} _ {i} ^ {m} = f _ {\\theta_ {1} ^ {m}} \\left(\\mathbf {x} _ {i} ^ {m}\\right), \\pmb {\\sigma} _ {i} ^ {m} = f _ {\\theta_ {2} ^ {m}} \\left(\\mathbf {x} _ {i} ^ {m}\\right), \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 151, + 147, + 395, + 186 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $f_{\\theta_1^m}(\\cdot)$ and $f_{\\theta_2^m}(\\cdot)$ represent the two different fully connected layers for mean and variance respectively. $\\mathbf{I}$ is the identity matrix. To maintain the consistency of semantics in different modalities, we further align the distributional representations of a multimodal sample with the Kullback-Leibler divergence:", + "bbox": [ + 76, + 196, + 468, + 286 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} L _ {\\mathrm {U D R}} = \\sum_ {m _ {1} \\neq m _ {2}} ^ {m _ {1}, m _ {2} \\in \\mathcal {M}} K L \\left(p \\left(\\mathbf {z} _ {i} ^ {m _ {1}} \\mid \\mathbf {x} _ {i} ^ {m _ {1}}\\right) \\| p \\left(\\mathbf {z} _ {i} ^ {m _ {2}} \\mid \\mathbf {x} _ {i} ^ {m _ {2}}\\right)\\right) \\tag {3} \\\\ + \\sum_ {m} ^ {m \\in \\mathcal {M}} K L \\left(p \\left(\\mathbf {z} _ {i} ^ {m} \\mid \\mathbf {x} _ {i} ^ {m}\\right) \\| \\mathcal {N} (0, \\mathbf {I})\\right). \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 91, + 297, + 467, + 385 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We illustrate the Unimodal Distributional Representation process in Fig. 2 with bi-modal input for clarity. It can be observed the $L_{\\mathrm{UDR}}$ will push two multimodal distributional representations with similar semantics closer. In this way, the representation of each multimodal sample is not limited within a deterministic point embedding, but a consistent fuzzy representation on several multivariate Gaussian distributions. Particularly, the variance $\\sigma_i^m$ reveals the aleatoric uncertainty in $m$ modality and the mean $\\mu_i^m$ is the corresponding stable representation.", + "bbox": [ + 75, + 396, + 468, + 547 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/0607f458b577d9e9813f032a20f946587a98b8f103e6fd7a5e4317dfc22208b4.jpg", + "image_caption": [ + "Figure 3. Illustration of the Uncertainty-based Representation Augmentation process. Here we show the distributional representations of image modality only for clarity." + ], + "image_footnote": [], + "bbox": [ + 86, + 559, + 460, + 718 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Uncertainty-based Representation Augmentation. With the quantified aleatoric uncertainty, we consider the second question we raised in Sec.1: should we drop the aleatoric uncertainty in multimodal data? Intuitively, the aleatoric uncertainty in multimodal data is unavoidable since the natural ambiguity of semantics. However, it also results in the diversity of unimodal data in different modalities. To this end, we leverage the aleatoric uncertainty to generate the", + "bbox": [ + 75, + 780, + 468, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "unseen samples such that learned unimodal representations are insensitive to diverse unimodal input with similar semantics. Take the image modality as an example, we illustrate the Uncertainty-based Representation Augmentation process in Fig. 3. Specifically, given a unimodal distributional representation $p(\\mathbf{z}_i^m | \\mathbf{x}_i^m) \\sim \\mathcal{N}(\\boldsymbol{\\mu}_i^m, \\boldsymbol{\\sigma}_i^{m2}\\mathbf{I})$ , we first random sample an anchor point $\\tilde{\\mathbf{z}}_i^m$ and an augmented point $\\mathbf{z}_i^m$ from the multivariate Gaussian distribution as matching pairs. Moreover, we randomly sample a set of negative points from other distributional representations and devise a self-supervised contrastive learning mechanism as follows:", + "bbox": [ + 496, + 90, + 890, + 257 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nL _ {\\mathrm {U R A}} = - \\log \\frac {e ^ {\\operatorname* {s i m} \\left(\\tilde {\\mathbf {z}} _ {i} ^ {m} , \\mathbf {z} _ {i} ^ {m}\\right) / \\tau}}{\\sum_ {j \\neq i} \\left(e ^ {\\operatorname* {s i m} \\left(\\tilde {\\mathbf {z}} _ {i} ^ {m} , \\mathbf {z} _ {i} ^ {m}\\right) / \\tau} + e ^ {\\operatorname* {s i m} \\left(\\tilde {\\mathbf {z}} _ {i} ^ {m} , \\mathbf {z} _ {j} ^ {m}\\right) / \\tau}\\right)}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 506, + 265, + 890, + 323 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\tau$ is a temperature factor and sim is cosine similarity calculation. Here we adopt the re-parameterization trick [40] to conduct the sampling operations, which can be formulated as follows:", + "bbox": [ + 496, + 323, + 890, + 382 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {z} _ {i} ^ {m} = \\boldsymbol {\\mu} _ {i} ^ {m} + \\epsilon \\sigma_ {i} ^ {m}, \\epsilon \\sim \\mathcal {N} (0, \\mathbf {I}). \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 584, + 393, + 890, + 411 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Robust Multimodal Feature Integration", + "text_level": 1, + "bbox": [ + 500, + 420, + 841, + 436 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "With the SUFA, we obtain stable unimodal representations with consistent semantics. However, we only consider the consistency in each modality but ignore the redundancy caused by duplicated representations. To this end, we proposed the Robust Multimodal Feature Integration (RMFI) module in this section. The overview of the RMFI module is illustrated in Fig. 4.", + "bbox": [ + 496, + 443, + 890, + 547 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Dynamic Multimodal Integration. Inspired by dynamic multimodal fusion [5], we assume that different modalities have unequal contributions to the joint representations in terms of the observed label space. To this end, we first apply a Dynamic Multimodal Integration strategy based on an attention mechanism. Specifically, given the stable unimodal representations $\\pmb{\\mu}_i^m$ , we calculate the attentive weights in the distributional representations across modalities according to their quantified uncertainties $\\sigma_i^m$ , and apply them for multimodal integration:", + "bbox": [ + 496, + 550, + 890, + 700 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathbf {x}} _ {i} = \\sum_ {m \\in \\mathcal {M}} \\alpha_ {i} ^ {m} \\boldsymbol {\\mu} _ {i} ^ {m}, \\quad \\alpha_ {i} ^ {m} = \\frac {e ^ {\\frac {1}{\\overline {{\\sigma}} _ {i} ^ {m}}}}{\\sum_ {m \\in \\mathcal {M}} e ^ {\\frac {1}{\\overline {{\\sigma}} _ {i} ^ {m}}}}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 550, + 710, + 890, + 756 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\hat{\\mathbf{x}}_i, \\overline{\\sigma}_i^m$ represent the integrated joint representations and the average variance of the multivariate Gaussian distribution of $m$ modality. In this way, we preliminarily integrate stable unimodal representations into joint representations, where the contributions of different modalities are estimated dynamically.", + "bbox": [ + 496, + 763, + 890, + 853 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Joint Representation Compression. In the SUFA module, we fully consider the consistency of semantics in different modalities to avoid noise in unimodal data. However,", + "bbox": [ + 496, + 854, + 890, + 898 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "26879", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/a08fd6e1048c2c148313340cc28584caf926f1655f346014537314ac9584d642.jpg", + "image_caption": [ + "Figure 4. An illustration of the RMFI module. We provide Cross-Entropy loss and $\\mathrm{L}_1$ loss for classification-based and regression-based downstream tasks respectively." + ], + "image_footnote": [], + "bbox": [ + 81, + 89, + 467, + 198 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "with similar multivariate Gaussian distributions among different modalities, redundant duplicated information will be introduced into joint representations. Hence, we devise a joint representation compression with the Variational Information Bottleneck (VIB) [20]. Specifically, given the preliminary joint representation $\\hat{\\mathbf{x}}_i$ and the target observations $\\mathbf{y}$ in label space, we learn the compressed joint representation $\\hat{\\mathbf{z}}$ in a latent space:", + "bbox": [ + 75, + 263, + 467, + 383 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\np \\left(\\hat {\\mathbf {z}} _ {i} \\mid \\hat {\\mathbf {x}} _ {i}\\right) \\sim \\mathcal {N} \\left(\\hat {\\boldsymbol {\\mu}} _ {i}, \\hat {\\boldsymbol {\\sigma}} _ {i} ^ {2} \\mathbf {I}\\right), \\quad \\hat {\\boldsymbol {\\mu}} _ {i} = f _ {\\hat {\\theta} _ {1}} (\\hat {\\mathbf {x}} _ {i}), \\quad \\hat {\\boldsymbol {\\sigma}} _ {i} = f _ {\\hat {\\theta} _ {2}} (\\hat {\\mathbf {x}} _ {i}). \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 75, + 395, + 467, + 426 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $f_{\\hat{\\theta}_1}$ and $f_{\\hat{\\theta}_2}$ are two fully connected layers. Similarly, we also adopt the re-parameterization tricks [40] for the final compressed joint representations:", + "bbox": [ + 76, + 426, + 467, + 472 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathbf {z}} _ {i} = \\hat {\\boldsymbol {\\mu}} _ {i} + \\epsilon \\hat {\\boldsymbol {\\sigma}} _ {i}, \\epsilon \\sim \\mathcal {N} (0, \\mathbf {I}). \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 171, + 484, + 467, + 500 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As our proposed method can be applied to different downstream tasks, here we provide two training objectives for the joint representation compression. Specifically, we employ the cross-entropy for the classification-based task:", + "bbox": [ + 75, + 512, + 467, + 571 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nL _ {\\mathrm {J R C}} = - \\mathbf {y} ^ {\\top} \\log \\delta \\left(f _ {\\theta} \\left(\\hat {\\mathbf {z}} _ {i}\\right)\\right) + \\lambda K L \\left(p \\left(\\hat {\\mathbf {z}} _ {i} \\mid \\hat {\\mathbf {x}} _ {i}\\right) \\| \\mathcal {N} (0, \\mathbf {I})\\right), \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 83, + 583, + 467, + 614 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\delta (\\cdot)$ represents softmax function and $f_{\\theta}(\\cdot)$ is the deep learning models for classification, $\\lambda$ is a hyperparameter. As for the regression-based task, we employ the mean square error as the training objective:", + "bbox": [ + 75, + 614, + 467, + 675 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nL _ {\\mathrm {J R C}} = \\left[ y - f _ {\\theta} \\left(\\hat {\\mathbf {z}} _ {i}\\right) \\right] ^ {2} + K L \\left(p \\left(\\hat {\\mathbf {z}} _ {i} \\mid \\hat {\\mathbf {x}} _ {i}\\right) \\| \\mathcal {N} (0, \\mathbf {I})\\right). \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 94, + 686, + 467, + 704 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 76, + 715, + 207, + 733 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Experimental Settings", + "text_level": 1, + "bbox": [ + 76, + 741, + 282, + 758 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Datasets. We evaluate our method on five multimodal datasets, including two tri-modal datasets, i.e., CMU-MOSI [41] and CMU-MOSEI [42] and three bi-modal datasets, i.e., MVSA-Single [43], UPMC Food101 [1], and NYU Depth v2 [44] datasets. The CMU-MOSI [41] and CMU-MOSEI [42] datasets are widely employed video datasets for Multimodal Sentiment Analysis (MSA) task. The former encompasses 2199 short videos, each accompanied by a sentiment strength score ranging from -3 to 3, offering", + "bbox": [ + 75, + 763, + 467, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "a nuanced measure of emotional intensity. The latter, designed specifically for emotion analysis tasks, is a larger dataset comprising 22,856 movie review clips. The bimodal dataset, i.e., MVSA-Single [43] is also used for MSA task but it only contains image-text pairs obtained from social media for sentiment classification. Following [2-5], we also employ the UPMC Food101 [1] and the NYU Depth v2 [44] datasets for Multimodal Image Classification (MIC) task. The UPMC Food101 dataset contain 101 categories, respectively, these datasets are specifically designed for bimodal classification tasks. The UPMC FOOD101 dataset contains images with 101 categories obtained by Google Image Search and their corresponding textual descriptions. The NYU Depth v2 dataset [44] is utilized for scene recognition and comprises depth and RGB images. Following [5], we also adopt the commonly used 9 out of the 27 scene categories and the remaining categories as \"Others\".", + "bbox": [ + 496, + 90, + 890, + 347 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Implementation Details. In alignment with alternative methodologies, for our trimodal video dataset, we employ FACET, COVAREP, and BERT as feature extractors for visual, audio, and text modalities, respectively. For the bimodal dataset, we utilize ResNet-152 as the feature extractor for RGB and depth images, coupled with BERT as the feature extractor for text. As for the hyperparameters, we set the temperature factor $\\tau$ and the balance factor $\\lambda$ to 0.5 and 1e-3 respectively. We employed the Adam optimizer with learning rate of 1e-5,and adopt the Reduce-on-Plateau learning rate adjustment strategy to train our EAU method.", + "bbox": [ + 496, + 348, + 890, + 513 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Evaluation Metrics. For the regression-based MSA task based on CMU-MOSI and CMU-MOSEI, we follow the previous work [6, 7, 9, 45] and adopt Acc7, F1 score, and Pearson correlation coefficient as our metrics. As for the classification-based MSA task on the MSVA-Single dataset and MIC task on the UPMC FOOD101 and The NYU Depth v2 datasets, we reported commonly used metrics, including accuracy and F1 score.", + "bbox": [ + 496, + 515, + 890, + 636 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Comparisons with State-Of-The-Arts", + "text_level": 1, + "bbox": [ + 500, + 648, + 821, + 665 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To show the superiority of our proposed EAU method on multimodal fusion, we compared our EAU method with existing state-of-the-art methods of the MSA task on the CMU-MOSI and CMU-MOSEI datasets, including MIB [28], HMA [6], MIM [7], GCNet [46], ConFEDE [8], DiC-MoR [45], and DMD [9]. Moreover, we also make fair comparisons with the widely adopted simple multimodal fusion strategies i.e., Concat and Late Fusion, as well as recent well-designed multimodal fusion strategies on the bimodal datasets, including TMC [3], ITIN [47], MMBT [2], PMF [4], MVCN [48], and QMF [5]. Particularly, similar to the counterpart method QMF [5], we evaluate our method on noisy datasets to observe the model robustness.", + "bbox": [ + 496, + 672, + 890, + 868 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Multimodal Fusion on Classification Task. We report the performance of classification tasks, including both MSA", + "bbox": [ + 500, + 869, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "26880", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/abc54406763ec0bed9548d3d49bda7a6f3821aed706d8ad8e547843f9a4beca5.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodCMU-MOSICMU-MOSEI
Acc7F1CorrAcc7F1Corr
MIB [28] (2022)48.685.30.79854.186.20.790
HMA [6] (2023)45.385.60.78252.885.40.787
MIM [7] (2023)47.085.90.80552.586.30.792
GCNet [46] (2023)44.985.1-51.585.2-
ConFEDE [8] (2023)42.385.50.78454.985.80.780
DiCMoR [45] (2023)45.385.6-53.485.1-
DMD [9] (2023)45.686.0-54.586.6-
EAU (Ours)48.886.20.80954.886.90.816
", + "bbox": [ + 83, + 88, + 488, + 224 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/79e18773d8283195a9616a82475616370b4b444929398cad45c75c729a5fc75e.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodMVSA-SingleFood 101NYU Depth v2
AccF1AccF1AccF1
Concat65.5965.4388.2088.1970.3069.82
Late Fusion76.8875.7290.6990.7769.1468.32
MMBT [2] (2020)78.50-91.5291.2871.04-
TMC [3] (2021)76.0674.5589.8689.8071.0669.83
ITIN [47] (2022)75.1974.97----
PMF [4] (2023)--91.68---
MVCN [48] (2023)76.0674.55----
QMF [5] (2023)78.0777.1892.9292.9370.0968.65
EAU (Ours)79.1578.3693.2093.1872.0570.63
", + "bbox": [ + 506, + 89, + 887, + 224 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "and MIC tasks on bi-modal datasets in Table 1, where the best results are marked in bold. From the experimental results, we can observe that our proposed EAU method outperforms all three datasets. Particularly, compared with the recent state-of-the-art method QMF method [5], our EAU method achieves at least $1\\%$ absolute improvements on the MVSA-Single and NYU Depth v2 datasets. These results demonstrate that our EAU approach conducts better multimodal fusion on text, depth, and RGB images, affirming the effectiveness of leveraging aleatoric uncertainty in multimodal data. However, we also note that the improvements on UPMC Food 101 are not as remarkable as the other two benchmark datasets. We speculate that one probable reason is the images and texts in this dataset are clear compared with the others. Hence, our EAU method obtains fewer improvements since the augmentation is limited.", + "bbox": [ + 75, + 300, + 472, + 542 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Multimodal Fusion on Regression Task. We evaluate our method on the regression task on the CMU-MOSEI and CMU-MOSI datasets, where the MSA task is conducted by predicting sentiment strengths. From the experimental results in Table 1, we can observe that: For the CMU-MOSI dataset, our proposed EAU method achieves state-of-the-art performance across all evaluation metrics, with a particularly significant improvement noted in Acc7. On the CMU-MOSEI dataset, our method has also demonstrated substantial improvements on most evaluation metrics. These results demonstrate the superiority of our proposed EAU, which benefits from high-quality multimodal fusion via learning stable and robust joint representations. Simultaneously, these results indicate that our approach can effortlessly extend to other modalities and can be seamlessly transferred to frameworks accommodating kinds of modalities.", + "bbox": [ + 75, + 547, + 468, + 789 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Model Robustness on Noisy Multimodal Datasets. To validate the effectiveness of our model in handling data noise, we conducted more evaluation on the noisy datasets following a recent multimodal fusion strategy QMF [5]. Concretely, we consider different intensities of Gaussian and Salt-Pepper noise on the MVSA-Single and NYU Depth v2 datasets. For fair comparisons, we conducted 10", + "bbox": [ + 75, + 795, + 470, + 902 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/197cb90e57e853e37c1dd3fa9ab7e35c8ffcb684b367d4b1a5b2cc6a54fc5ebc.jpg", + "table_caption": [ + "Table 1. Comparisons with state-of-the-art multimodal fusion methods on MSA and MIC tasks. The CMU-MOSI and CMU-MOSEI datasets contain videos, audi, and texts. The MVSA-Single and Food-101 datasets consist of texts and RGB images. The NYU Depth v2 contains RGB and depth images. Note that CMU-MOSI and CMU-MOSEI are used for regression-based MSA tasks, while the others are used for classification-based MSA or MIC tasks." + ], + "table_footnote": [], + "table_body": "
Noisy MVSA-Single
MethodCleanSalt-Pepper NoiseGaussian Noise
ε = 0ε = 5ε = 10ε = 5ε = 10
Concat65.5958.6951.1650.7046.12
Late Fusion76.8867.8855.4363.4655.16
MMBT [2] (2020)78.5074.0751.2671.9955.34
TMC [3] (2021)74.8768.0256.6266.7260.35
QMF [5] (2023)78.0773.9060.4173.8561.28
EAU (Ours)79.1574.8161.0473.8962.04
", + "bbox": [ + 504, + 297, + 888, + 420 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/2387cdab7f44efbc90c0d17f441fcb498aafbc1c3d891286b006f5d58133445e.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Noisy NYU Depth v2
MethodCleanSalt-Pepper NoiseGaussian Noise
ε = 0ε = 5ε = 10ε = 5ε = 10
Concat70.4457.9844.5159.9753.20
Late fusion69.1656.2741.2259.6351.99
MMTM [2] (2020)71.0459.4544.5960.3752.28
TMC [3] (2021)71.0659.3444.6561.0453.36
QMF [5] (2023)70.0958.5045.6961.6255.60
EAU (Ours)72.0559.8346.8563.3358.85
", + "bbox": [ + 504, + 421, + 888, + 541 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 2. Comparisons with state-of-the-arts concerning model performance on noisy MVSA-Single and NYU Depth v2 datasets.", + "bbox": [ + 500, + 547, + 890, + 575 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "experiments with different random seeds and reported the mean results as our final performance. According to the experimental results listed in Table 2, we can observe that our method has achieved state-of-the-art performance under various types and intensities of noise. Moreover, with the intensity of noise increasing, the margin of performance between our method and existing multimodal fusion strategies goes larger. Particularly, our method achieves more than $3\\%$ improvement on the Noisy NYU Depth v2 dataset over the QMF method [5]. Such a result demonstrates our method reveals better robustness against data noise. The reason is our model pioneeringly leverages the intrinsic aleatoric uncertainty in the training data to enhance the stable feature representation, which is not fully considered in existing counterpart methods [2, 3, 5].", + "bbox": [ + 496, + 590, + 890, + 818 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Further Analysis", + "text_level": 1, + "bbox": [ + 500, + 830, + 668, + 847 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Analysis on Model Structure. To explore the impact of different model structures, we decomposed our proposed method into two components, SUFA and RMFI, and evalu", + "bbox": [ + 498, + 854, + 893, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "26881", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/4fadb6373ceae741e89bb3c819aef9c6508ac910406e925570672f59108b7fd1.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodCleanSalt-Pepper NoiseGaussian Noise
Acc@ε=0Acc@ε=5Acc@ε=10Acc@ε=5Acc@ε=10
Naive Backbone + Late Fusion76.88 ± 1.3067.88 ± 1.8755.43 ± 1.9463.46 ± 3.4655.16 ± 3.60
Naive Backbone + Concat74.53 ± 0.9762.24 ± 2.4753.14 ± 2.8956.08 ± 1.8949.21 ± 2.81
Naive Backbone + RMFI77.23 ± 0.7470.99 ± 2.7759.11 ± 1.9266.53 ± 2.2652.02 ± 3.45
SUFA + Late Fusion77.11 ± 1.0570.05 ± 1.5055.93 ± 2.2562.27 ± 4.2055.90 ± 3.67
SUFA + Concat78.18 ± 1.0873.10 ± 1.0760.67 ± 1.3871.40 ± 1.4859.11 ± 1.60
BERT (NAACL'19) [49]75.61 ± 0.5369.50 ± 1.5047.41 ± 0.7969.50 ± 1.5047.41 ± 0.79
MMBT (arXiv'19) [2]78.50 ± 0.4074.07 ± 1.1251.26 ± 5.6571.99 ± 1.0455.34 ± 2.84
TMC (ICLR'21) [3]74.87 ± 2.2468.02 ± 3.0756.62 ± 3.6766.72 ± 4.5560.35 ± 2.79
QMF (ICML'23) [5]78.07 ± 1.1073.90 ± 1.8960.41 ± 2.6373.85 ± 1.4261.28 ± 2.12
SUFA + RMFI (EAU)79.15 ± 0.6074.81 ± 1.5961.04 ± 1.3173.89 ± 0.9662.04 ± 1.26
", + "bbox": [ + 99, + 89, + 665, + 232 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/73d9ef011638ff007667783360b4eba5edcf3d14cc9a4d1cdff9740a24e2d1dc.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModalityAccF1Corr
T44.183.70.785
V16.745.30.072
A15.848.70.099
T+A46.784.20.794
T+V46.384.60.786
A+V20.255.30.117
T+A+V48.886.20.809
T76.3075.90-
V63.5863.35-
T+V79.1578.36-
", + "bbox": [ + 669, + 89, + 867, + 232 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 3. Analysis concerning key model structures in our EAU method on the Noisy MSVA-Single (Left), and analysis concerning modalities on the CMU-MOSI (Right Upper) and MVSA-Single (Right Lower) datasets. Note that BERT [49] adopts text modality only.", + "bbox": [ + 75, + 238, + 890, + 267 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "ated their effectiveness in combination with various model structures on the Noisy MVSA-Single dataset. Specifically, we employ two alternative designs for SUFA and RMFI respectively: (1) Naive Backbone: deploying the feature extractors directly, where the distributional representation and stable feature augmentation in the SUFA module are disabled. (2) Late Fusion: the unimodal representations are used for downstream tasks first and the fusion operations are conducted in probability space. To observe the stability of our method, we conduct 10 experiments with different random seeds and report the mean and variance in Table 3. Here we also make fair comparisons with several counterpart methods [2, 3, 5, 49] to furtherly show our superiority.", + "bbox": [ + 75, + 277, + 468, + 474 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "According to the experimental results presented in Table 3, we can observe that: (1) Our proposed SUFA and RMFI modules consistently improve performance on all metrics. Particularly, with the SUFA module, the results of fusion strategy Concat are boosted with more than $10\\%$ absolute performance on the noisy data. Moreover, compared with Late Fusion, our proposed RMFI module can also significantly boost the classification accuracy on noisy data with the SUFA module. These results prove again that our method shows great superiority in the effectiveness and robustness of multimodal fusion. (2) It also explicates that our proposed EAU method remarkably outperforms compared with the counterpart methods [2, 3, 5]. Particularly, compared with the recent state-of-the-art method QMF [5], our complete model reveals better performance and stability, demonstrating the capability of our EAU method again.", + "bbox": [ + 75, + 477, + 468, + 718 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Analysis on the Effectiveness of Multimodal Fusion. To verify the effectiveness of multimodal fusion in our EAU method, we also conduct ablation studies concerning different modalities on the CMU-MOSI datasets, which consist of video (V), text (T), and audio (A) three modalities. By observing the experimental results illustrated in Table 3, we have drawn the following conclusions: (1) Our fusion strategy demonstrated significant effectiveness across different modal combinations. As the number of modalities increased, our approach facilitated the integration of multimodal information, resulting in a significant improvement across all evaluation metrics. (2) We also note that the text", + "bbox": [ + 75, + 719, + 468, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "modality performs an essential role in the MSA task on the CMU-MOSI dataset. However, combined with the other two modalities via our proposed EAU method, the results are also boosted significantly, demonstrating the superiority of our EAU method again. Such a phenomenon also proves the rationality of our Dynamic Multimodal Integration process in the RMFI module.", + "bbox": [ + 496, + 277, + 890, + 383 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/8668a04804c4b09a944523bf3b075fa005aa77bbd40a05d1a1e70698757000bc.jpg", + "image_caption": [ + "Figure 5. Analysis for robustness evaluated by over 10 times random experiments under different noisy data." + ], + "image_footnote": [], + "bbox": [ + 500, + 391, + 696, + 526 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/71ff691682e7c3f72c18382f828a43e21e7dd09a6ae48544210d4add733b0410.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 696, + 391, + 890, + 526 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Analysis on Model Robustness. To verify the robustness and stability of our proposed EAU method, we conduct more ablation studies on the MVSA-Single and NYU Depth v2 datasets under different noise levels and observe the sparsity of final accuracy. Specifically, we conduct more than 10 experiments randomly with complete EAU (denoted as Complete) and EAU w/o SUFA (denoted as w/o SUFA), and show the statistics in Fig. 5. By observing the experimental results, we can find that: (1) The SUFA module consistently shows its effectiveness on both two datasets under different noise strengths. Particularly, when introducing more noise to the data, the ablated model shows strong fluctuations. Contrastively, the complete EAU model reveals remarkably reduced fluctuations, which demonstrates the superiority of model robustness and stability. (2) We also note that the improvements on the NYU Depth v2 are slightly lower than the MVSA-Single dataset, though both two datasets have significant results. We speculate this is caused by the difference between text modality and depth images. Concretely, the depth images can be regarded as another view of spatial information, which has less complementary information for RGB images compared with texts.", + "bbox": [ + 496, + 568, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "26882", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To this end, the introduced noise will show more impact on the depth-RGB than the image-text compositions.", + "bbox": [ + 76, + 90, + 468, + 121 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/726e724da2f8f1cecd8658f2394ba73f8127abb5c4412c3104b23edd9751a0ac.jpg", + "image_caption": [ + "Figure 6. Analysis concerning the training convergence and performance fluctuations during training." + ], + "image_footnote": [], + "bbox": [ + 76, + 132, + 468, + 256 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Analysis on Training Process. We also illustrate the training process of our proposed EAU method on the MVSA-Single in Fig 6 to observe the training convergence and performance fluctuations. According to the illustrated experimental results, we can observe that the complete EAU model exhibits a smoother training process and converges more rapidly with a remarkably better performance. Particularly, compared with the ablated model w/o RMFI, the other two models show much better stability during training. This is because the fused features learned by the ablated model w/o RMFI have redundant information caused by cross-modal distributional alignment in the SUFA module. It proves again the rationality of our proposed RMFI module, which deploys an information-theoretic strategy to address the limitation.", + "bbox": [ + 75, + 297, + 470, + 523 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/a0a44e1bc2b030f65ccba3a62e40971d0ae6ad4b7430367fd66c5b355106b151.jpg", + "image_caption": [ + "(a) Concat" + ], + "image_footnote": [], + "bbox": [ + 81, + 534, + 254, + 635 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/b006de730ac2a0fa72ffc5821911aa703b6e4f3a889149c36ef3c7fdec7b7191.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 279, + 532, + 468, + 635 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/4025742e48aae512ebe097fe777828d2f04e9687ae3b56ed36661b85da848df5.jpg", + "image_caption": [ + "(c) EAU w/o RMFI" + ], + "image_footnote": [], + "bbox": [ + 81, + 651, + 254, + 751 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/68fa03c3a5c26ed034bbc01e245e8d35c8dc77ad8493aae4bf3f6b40f7e00693.jpg", + "image_caption": [ + "(b) EAU w/o SUFA", + "(d) EAU", + "Figure 7. Visualizations of joint representations by t-SNE [50] on MVSA-Single dataset. Here we adopt Concat for the ablated model w/o RMFI." + ], + "image_footnote": [], + "bbox": [ + 276, + 650, + 468, + 750 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Visualizations of Joint Representations. To further verify the superiority of our proposed EAU method, we also employ the t-SNE [50] to visualize the learned joint representations. As illustrated in Fig. 7, it can be observed that the features generated by our method exhibit a more", + "bbox": [ + 75, + 824, + 468, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "compact and discriminative distribution. In contrast, when the SUFA or RMFI module is removed, more sparse points appear around the feature distribution which leads to corrupted classification results at last. Such results indicate that our method is more capable of learning a representative joint representation.", + "bbox": [ + 496, + 90, + 890, + 181 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/9ffa48bd060d471068ad61f4ee8c613e060afdd37867e1a17166aeaa5ee2981b.jpg", + "image_caption": [ + "Figure 8. Visualizations of test cases selected from the MSVA-Single datasets. It can be observed that our EAU method reveals better robustness against the noise." + ], + "image_footnote": [], + "bbox": [ + 501, + 188, + 890, + 349 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Qualitative Analysis. Moreover, we also illustrate several typical test cases in Fig. 8 and make comparisons with the recent counterpart method QMF [5]. Under different noise settings, our proposed EAU method consistently classifies the sentiment of the given image-text input precisely, while the QMF method returns incorrect results. It explicates that our proposed EAU method benefits from the well-designed SUFA and RMFI module, which is effective in learning more stable and robust joint representations for noise-resistant performance with higher accuracy.", + "bbox": [ + 496, + 407, + 890, + 559 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 574, + 617, + 589 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we proposed a novel multimodal fusion method, namely, Embracing Aleatoric Uncertainty (EAU). It achieves more discriminative joint representations by fully considering the aleatoric uncertainty in multimodal data. Specifically, with the well-designed Stable Unimodal Feature Augmentation (SUFA) and Robust Multimodal Feature Integration (RMFI) modules, our proposed EAU could learn compact and robust joint representations. We evaluated our proposed EAU method on five multimodal benchmark datasets for both classification and regression tasks and demonstrated its superiority in fusion performance and robustness. For future work, we will further explore the uncertainties in multimodal learning.", + "bbox": [ + 496, + 599, + 890, + 797 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Acknowledgement", + "text_level": 1, + "bbox": [ + 500, + 814, + 679, + 830 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This work was sponsored in part by the National Natural Science Foundation of China under Grants (No. 62222203 and No. 62072080) and the New Cornerstone Science Foundation through the XPLORER PRIZE.", + "bbox": [ + 496, + 839, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "26883", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 174, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Lukas Bossard, Matthieu Guillaumin, and Luc Van Gool. Food-101-mining discriminative components with random forests. In European Conference on Computer Vision, pages 446-461, 2014. 1, 5", + "[2] Douwe Kiela, Suvrat Bhooshan, Hamed Firooz, Ethan Perez, and Davide Testuggine. Supervised multimodal bittransformers for classifying images and text. arXiv preprint arXiv:1909.02950, 2019. 1, 2, 5, 6, 7", + "[3] Zongbo Han, Changqing Zhang, Huazhu Fu, and Joey Tianyi Zhou. Trusted multi-view classification. arXiv preprint arXiv:2102.02051, 2021. 1, 2, 5, 6, 7", + "[4] Yaowei Li, Ruijie Quan, Linchao Zhu, and Yi Yang. Efficient multimodal fusion via interactive prompting. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 2604-2613, 2023. 1, 5, 6", + "[5] Qingyang Zhang, Haitao Wu, Changqing Zhang, Qinghua Hu, Huazhu Fu, Joey Tianyi Zhou, and Xi Peng. Provable dynamic fusion for low-quality multimodal data. arXiv preprint arXiv:2306.02050, 2023. 1, 2, 4, 5, 6, 7, 8", + "[6] Ronghao Lin and Haifeng Hu. Dynamically shifting multimodal representations via hybrid-modal attention for multimodal sentiment analysis. IEEE Transactions on Multimedia, 2023. 1, 5, 6", + "[7] Ying Zeng, Sijie Mai, Wenjun Yan, and Haifeng Hu. Multimodal reaction: Information modulation for cross-modal representation learning. IEEE Transactions on Multimedia, 2023. 1, 5, 6", + "[8] Jiuding Yang, Yakun Yu, Di Niu, Weidong Guo, and Yu Xu. Confede: Contrastive feature decomposition for multimodal sentiment analysis. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics, pages 7617-7630, 2023. 1, 5, 6", + "[9] Yong Li, Yuanzhi Wang, and Zhen Cui. Decoupled multimodal distilling for emotion recognition. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 6631-6640, 2023. 1, 5, 6", + "[10] Weiyao Wang, Du Tran, and Matt Feiszli. What makes training multi-modal classification networks hard? In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 12695-12705, 2020. 1, 2", + "[11] Xiaokang Peng, Yake Wei, Andong Deng, Dong Wang, and Di Hu. Balanced multimodal learning via on-the-fly gradient modulation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 8238-8247, 2022. 1, 2", + "[12] Yu Huang, Junyang Lin, Chang Zhou, Hongxia Yang, and Longbo Huang. Modality competition: What makes joint training of multi-modal network fail in deep learning? (provably). In International Conference on Machine Learning, pages 9226-9259. PMLR, 2022. 1, 2", + "[13] Yunfeng Fan, Wenchao Xu, Haozhao Wang, Junxiao Wang, and Song Guo. PMR: prototypical modal rebalance for multimodal learning. In Proceedings of the IEEE Conference" + ], + "bbox": [ + 78, + 114, + 470, + 887 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "on Computer Vision and Pattern Recognition, pages 20029-20038. IEEE, 2023. 1, 2", + "[14] Alex Kendall and Yarin Gal. What uncertainties do we need in bayesian deep learning for computer vision? Advances in neural information processing systems, 30, 2017. 2, 3", + "[15] Jie Chang, Zhonghao Lan, Changmao Cheng, and Yichen Wei. Data uncertainty learning in face recognition. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 5710-5719, 2020. 2", + "[16] Yichun Shi and Anil K Jain. Probabilistic face embeddings. In Proceedings of the IEEE International Conference on Computer Vision, pages 6902-6911, 2019. 2", + "[17] Yatai Ji, Junjie Wang, Yuan Gong, Lin Zhang, Yanru Zhu, Hongfa Wang, Jiaxing Zhang, Tetsuya Sakai, and Yujiu Yang. Map: Multimodal uncertainty-aware vision-language pre-training model. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 23262-23271, 2023. 2", + "[18] Luke Vilnis and Andrew McCallum. Word representations via gaussian embedding. arXiv preprint arXiv:1412.6623, 2014. 2", + "[19] Naftali Tishby, Fernando C Pereira, and William Bialek. The information bottleneck method. arXiv preprint physics/0004057, 2000. 2, 3", + "[20] Alexander A Alemi, Ian Fischer, Joshua V Dillon, and Kevin Murphy. Deep variational information bottleneck. arXiv preprint arXiv:1612.00410, 2016. 2, 3, 5", + "[21] Xun Jiang, Xing Xu, Zhiguo Chen, Jingran Zhang, Jingkuan Song, Fumin Shen, Huimin Lu, and Heng Tao Shen. DHHN: dual hierarchical hybrid network for weakly-supervised audio-visual video parsing. In ACM International Conference on Multimedia, pages 719–727, 2022. 2", + "[22] Junyu Gao, Mengyuan Chen, and Changsheng Xu. Collecting cross-modal presence-absence evidence for weakly-supervised audio-visual event perception. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 18827–18836, 2023. 2", + "[23] Zhiguo Chen, Xun Jiang, Xing Xu, Zuo Cao, Yijun Mo, and Heng Tao Shen. Joint searching and grounding: Multigranularity video content retrieval. In Proceedings of the 31st ACM International Conference on Multimedia, pages 975-983, 2023. 2", + "[24] Shenshen Li, Xing Xu, Xun Jiang, Fumin Shen, Xin Liu, and Heng Tao Shen. Multi-grained attention network with mutual exclusion for composed query-based image retrieval. IEEE Transactions on Circuits and Systems for Video Technology, 2023. 2", + "[25] Shenshen Li, Xing Xu, Xun Jiang, Fumin Shen, Zhe Sun, and Andrzej Cichocki. Cross-modal attention preservation with self-contrastive learning for composed query-based image retrieval. ACM Transactions on Multimedia Computing, Communications, and Applications, 20(6), 2024. 2" + ], + "bbox": [ + 503, + 92, + 890, + 872 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "26884", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[26] Xing Xu, Kaiyi Lin, Yang Yang, Alan Hanjalic, and Heng Tao Shen. Joint feature synthesis and embedding: Adversarial cross-modal retrieval revisited. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(6):3030-3047, 2022. 2", + "[27] Shenshen Li, Xing Xu, Xun Jiang, Fumin Shen, Xin Liu, and Heng Tao Shen. Multi-grained attention network with mutual exclusion for composed query-based image retrieval. IEEE Transactions on Circuits and Systems for Video Technology, pages 1-1, 2023. 2", + "[28] Sijie Mai, Ying Zeng, and Haifeng Hu. Multimodal information bottleneck: Learning minimal sufficient unimodal and multimodal representations. IEEE Transactions on Multimedia, 2022. 2, 3, 5, 6", + "[29] Xun Jiang, Xing Xu, Jingran Zhang, Fumin Shen, Zuo Cao, and Heng Tao Shen. Sdn: Semantic decoupling network for temporal language grounding. IEEE Transactions on Neural Networks and Learning Systems, pages 1-15, 2022. 2", + "[30] Xun Jiang, Xing Xu, Jingran Zhang, Fumin Shen, Zuo Cao, and Heng Tao Shen. Semi-supervised video paragraph grounding with contrastive encoder. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 2456-2465. IEEE, 2022. 2", + "[31] Arsha Nagrani, Shan Yang, Anurag Arnab, Aren Jansen, Cordelia Schmid, and Chen Sun. Attention bottlenecks for multimodal fusion. In Advances in Neural Information Processing Systems, pages 14200-14213, 2021. 2, 3", + "[32] Guanglei Yang, Enrico Fini, Dan Xu, Paolo Rota, Mingli Ding, Moin Nabi, Xavier Alameda-Pineda, and Elisa Ricci. Uncertainty-aware contrastive distillation for incremental semantic segmentation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(2):2567-2581, 2022. 2", + "[33] Mengyuan Chen, Junyu Gao, Shicai Yang, and Changsheng Xu. Dual-evidential learning for weakly-supervised temporal action localization. In European Conference on Computer Vision, pages 192-208. Springer, 2022. 2", + "[34] Yifei Huang, Lijin Yang, and Yoichi Sato. Weakly supervised temporal sentence grounding with uncertainty-guided self-training. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 18908-18918, 2023. 2", + "[35] Shenshen Li, Xing Xu, Yang Yang, Fumin Shen, Yijun Mo, Yujie Li, and Heng Tao Shen. DCEL: deep cross-modal evidential learning for text-based person retrieval. In Proceedings of the 31st ACM International Conference on Multimedia, pages 6292-6300, 2023. 2", + "[36] Alexander A Alemi, Ian Fischer, Joshua V Dillon, and Kevin Murphy. Deep variational information bottleneck. arXiv preprint arXiv:1612.00410, 2016. 3", + "[37] Yuchen Shen, Dong Zhang, Zhihao Song, Xuesong Jiang, and Qiaolin Ye. Learning to reduce information bottleneck for object detection in aerial images. IEEE Geoscience and Remote Sensing Letters, 2023. 3" + ], + "bbox": [ + 78, + 90, + 468, + 878 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[38] Among Wu and Cheng Deng. Tib: Detecting unknown objects via two-stream information bottleneck. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023. 3", + "[39] Sungtae An, Nataraj Jammalamadaka, and Eunji Chong. Maximum entropy information bottleneck for uncertainty-aware stochastic embedding. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 3808-3817, 2023. 3", + "[40] Diederik P Kingma and Max Welling. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114, 2013. 4, 5", + "[41] Amir Zadeh, Rowan Zellers, Eli Pincus, and Louis-Philippe Morency. Mosi: multimodal corpus of sentiment intensity and subjectivity analysis in online opinion videos. arXiv preprint arXiv:1606.06259, 2016. 5", + "[42] AmirAli Bagher Zadeh, Paul Pu Liang, Soujanya Poria, Erik Cambria, and Louis-Philippe Morency. Multimodal language analysis in the wild: Cmu-mosei dataset and interpretable dynamic fusion graph. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics, pages 2236–2246, 2018. 5", + "[43] Teng Niu, Shiai Zhu, Lei Pang, and Abdulmotaleb El Saddik. Sentiment analysis on multi-view social data. In International Conference on Multimedia Modeling, pages 15-27, 2016. 5", + "[44] Nathan Silberman, Derek Hoiem, Pushmeet Kohli, and Rob Fergus. Indoor segmentation and support inference from rgbd images. In European Conference on Computer Vision, pages 746-760, 2012. 5", + "[45] Yuanzhi Wang, Zhen Cui, and Yong Li. Distribution-consistent modal recovering for incomplete multimodal learning. In Proceedings of the IEEE International Conference on Computer Vision, pages 22025-22034, 2023. 5, 6", + "[46] Zheng Lian, Lan Chen, Licai Sun, Bin Liu, and Jianhua Tao. Gcnet: graph completion network for incomplete multimodal learning in conversation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023. 5, 6", + "[47] Tong Zhu, Leida Li, Jufeng Yang, Sicheng Zhao, Hantao Liu, and Jiansheng Qian. Multimodal sentiment analysis with image-text interaction network. IEEE Transactions on Multimedia, 2022. 5, 6", + "[48] Yiwei Wei, Shaozu Yuan, Ruosong Yang, Lei Shen, Zhangmeizhi Li, Longbiao Wang, and Meng Chen. Tackling modality heterogeneity with multi-view calibration network for multimodal sentiment detection. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics, pages 5240-5252, 2023. 5, 6", + "[49] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018. 7", + "[50] Laurens Van der Maaten and Geoffrey Hinton. Visualizing data using t-sne. Journal of machine learning research, 9(11), 2008. 8" + ], + "bbox": [ + 501, + 92, + 890, + 883 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "26885", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2024/Embracing Unimodal Aleatoric Uncertainty for Robust Multimodal Fusion/b864c4f0-5e7f-441f-a2a0-028202223b38_model.json b/2024/Embracing Unimodal Aleatoric Uncertainty for Robust Multimodal Fusion/b864c4f0-5e7f-441f-a2a0-028202223b38_model.json new file mode 100644 index 0000000000000000000000000000000000000000..7e79022513e0378d0d979aaa8979d06715f92e99 --- /dev/null +++ b/2024/Embracing Unimodal Aleatoric Uncertainty for Robust Multimodal Fusion/b864c4f0-5e7f-441f-a2a0-028202223b38_model.json @@ -0,0 +1,2156 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.108, + 0.131, + 0.865, + 0.154 + ], + "angle": 0, + "content": "Embracing Unimodal Aleatoric Uncertainty for Robust Multimodal Fusion" + }, + { + "type": "text", + "bbox": [ + 0.166, + 0.18, + 0.804, + 0.2 + ], + "angle": 0, + "content": "Zixian Gao\\(^{1*}\\), Xun Jiang\\(^{1*}\\), Xing Xu\\(^{1,2\\dagger}\\), Fumin Shen\\(^{1}\\), Yujie Li\\(^{3}\\), Heng Tao Shen\\(^{1,2}\\)" + }, + { + "type": "text", + "bbox": [ + 0.187, + 0.2, + 0.777, + 0.234 + ], + "angle": 0, + "content": "1Center for Future Media & School of Computer Science and Engineering, University of Electronic Science and Technology of China, China" + }, + { + "type": "text", + "bbox": [ + 0.179, + 0.234, + 0.792, + 0.252 + ], + "angle": 0, + "content": "\\(^{2}\\)College of Electronic and Information Engineering, Tongji University, China" + }, + { + "type": "text", + "bbox": [ + 0.327, + 0.252, + 0.639, + 0.27 + ], + "angle": 0, + "content": "\\(^{3}\\)Kyushu Institute of Technology, Japan" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.305, + 0.313, + 0.32 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.337, + 0.474, + 0.701 + ], + "angle": 0, + "content": "As a fundamental problem in multimodal learning, multimodal fusion aims to compensate for the inherent limitations of a single modality. One challenge of multimodal fusion is that the unimodal data in their unique embedding space mostly contains potential noise, which leads to corrupted cross-modal interactions. However, in this paper, we show that the potential noise in unimodal data could be well quantified and further employed to enhance more stable unimodal embeddings via contrastive learning. Specifically, we propose a novel generic and robust multimodal fusion strategy, termed Embracing Aleatoric Uncertainty (EAU), which is simple and can be applied to kinds of modalities. It consists of two key steps: (1) the Stable Unimodal Feature Augmentation (SUFA) that learns a stable unimodal representation by incorporating the aleatoric uncertainty into self-supervised contrastive learning. (2) Robust Multimodal Feature Integration (RMFI) leveraging an information-theoretic strategy to learn a robust compact joint representation. We evaluate our proposed EAU method on five multimodal datasets, where the video, RGB image, text, audio, and depth image are involved. Extensive experiments demonstrate the EAU method is more noise-resistant than existing multimodal fusion strategies and establishes new state-of-the-art on several benchmarks." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.731, + 0.21, + 0.746 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.757, + 0.47, + 0.864 + ], + "angle": 0, + "content": "By exploring the complementary information from different modalities, multimodal learning has achieved impressive success in kinds of artificial intelligence applications, such as multimodal image classification in Internet applications [1-5] and sentiment analysis in intelligent robotics [6-9]. It has been widely proved that fusing information from different modalities appropriately is helpful to gain" + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.304, + 0.891, + 0.396 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.577, + 0.4, + 0.82, + 0.41 + ], + "angle": 0, + "content": "(a) Uncertainties from fuzzy semantics in multimodal data" + }, + { + "type": "image", + "bbox": [ + 0.52, + 0.414, + 0.888, + 0.516 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.535, + 0.519, + 0.85, + 0.529 + ], + "angle": 0, + "content": "(b) Multivariate Gaussian distributional representations for fuzzy semantics" + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.541, + 0.895, + 0.612 + ], + "angle": 0, + "content": "Figure 1. An illustration of the aleatoric uncertainty in multimodal datasets and the distributional representations: (a) As the semantics are ambiguous, the multimodal data are prone to introducing noisy data. (b) We adopt a multivariate Gaussian distribution to represent the fuzzy semantics in a noisy latent space." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.627, + 0.893, + 0.687 + ], + "angle": 0, + "content": "better performance. Dedicated to this, multimodal fusion has become an emerging challenge in multimodal learning, which integrates different modalities into a unified representation with powerful neural networks." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.69, + 0.895, + 0.901 + ], + "angle": 0, + "content": "Nevertheless, with the studies developing, researchers have recognized that the multimodal data in different modal forms may be unreliable due to the unique noise in their own modal space. Recently, several works [5, 10-13] show that previous multimodal fusion methods have overlooked the unreliable quality of multimodal data. Specifically, the widely adopted fusion strategies may fail on noisy multimodal data because the cross-modal interactions may obtain limited influence by the uncertainty in data. Here we illustrate typical noise in multimodal data in Fig. 1. Due to the semantics of \"Positive\" being fuzzy and the labels being judged by humans subjectively, the given image-text pairs are all recognized as positive in sentiment analysis, even though there is obvious noise in both image and text" + }, + { + "type": "page_footnote", + "bbox": [ + 0.095, + 0.875, + 0.352, + 0.888 + ], + "angle": 0, + "content": "*These authors contributed equally to this work." + }, + { + "type": "page_footnote", + "bbox": [ + 0.098, + 0.888, + 0.222, + 0.9 + ], + "angle": 0, + "content": "† Corresponding author." + }, + { + "type": "list", + "bbox": [ + 0.095, + 0.875, + 0.352, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "26876" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.151 + ], + "angle": 0, + "content": "modalities. In general, we can say the introduced noise comes from the aleatoric uncertainty [14], which undermines the effectiveness of exploiting multimodal data and demonstrates multimodal learning is not a free lunch." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.153, + 0.473, + 0.497 + ], + "angle": 0, + "content": "Hence, with the emerging challenge of uncertainty in multimodal data, we raise the first fundamental question: Can we quantify the uncertainty in multimodal data? Inspired by probability Distributional Representation [15-18], we naturally adopt the Gaussian distributions to work out this problem. We assume that each instance can be represented as a multivariate normal distribution, where the variance can be regarded as the intrinsic aleatoric uncertainty. With the quantified uncertainty, we are allowed to take a closer look at the multimodal fusion and raise the second question: Is it appropriate to completely drop the intrinsic uncertainty? From the examples in Fig. 1(a), we can observe that even if image-text pairs reveal similar semantics, the aleatoric uncertainty is still unavoidable because of the domain shift, extra descriptions, or image quality, etc. To this end, as is illustrated in Fig. 1(b), we argue that a multivariate normal distribution that considers aleatoric uncertainty can be regarded as a fuzzy representation for semantics, where the semantic-related data are within similar distributions, even if they are in different modalities. Motivated by the two assumptions, we develop a novel multimodal fusion strategy in this paper, i.e., Embracing Aleatoric Uncertainty (EAU)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.499, + 0.473, + 0.816 + ], + "angle": 0, + "content": "Specifically, our EAU method consists of the following two processes: (1) The Stable Unimodal Feature Augmentation (SUFA) quantifies the intrinsic aleatoric uncertainty in each modality by representing samples as multivariate normal distributions. Then, we further conduct self-supervised contrastive learning with these distributional representations to learn stable unimodal embeddings with better semantical consistency. (2) Robust Multimodal Feature Integration (RMFI) that dynamically fuses the stable unimodal embeddings into a joint representation. Particularly, considering that SUFA focuses on semantic consistency only and ignores the problem of information redundancy, we employ an information-theoretic strategy, i.e., Variational Information Bottleneck [19, 20], to learn a compact joint representation with less redundancy. We evaluate our method on five multimodal benchmark datasets and demonstrate it outperforms existing state-of-the-art methods on both Multimodal Sentiment Analysis and Multimodal Image Classification tasks. Moreover, our proposed EAU method also reveals better robustness on noisy datasets compared with the counterpart methods [2, 3, 5]." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.817, + 0.47, + 0.846 + ], + "angle": 0, + "content": "Overall, our contributions in this paper can be summarized as three-fold:" + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.856, + 0.47, + 0.901 + ], + "angle": 0, + "content": "- We propose a novel multimodal fusion method termed Embracing Aleatoric Uncertainty (EAU), which quantifies the intrinsic aleatoric uncertainty and leverages it" + }, + { + "type": "text", + "bbox": [ + 0.532, + 0.092, + 0.842, + 0.108 + ], + "angle": 0, + "content": "to learn stable and robust joint representations." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.115, + 0.892, + 0.176 + ], + "angle": 0, + "content": "- We devise the Stable Unimodal Feature Augmentation (SUFA) module, which quantifies aleatoric uncertainty and learns stable unimodal representations with better semantical consistency." + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.184, + 0.892, + 0.244 + ], + "angle": 0, + "content": "- We design the Robust Multimodal Feature Integration (RMFI) module to learn compact joint representations with less redundancy, which further improves the robustness of multimodal fusion in our method." + }, + { + "type": "list", + "bbox": [ + 0.518, + 0.115, + 0.892, + 0.244 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.257, + 0.642, + 0.273 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.282, + 0.892, + 0.644 + ], + "angle": 0, + "content": "Multimodal Fusion. Multimodal fusion, which aims at learning stronger representations from different modalities, has become an essential part of a spectrum of computer vision research, such as vision-audio learning [21,22], vision-language learning [23, 24], image retrieval [25-27], and video understanding [28-30]. Generally, multimodal fusion can be categorized into three types: early fusion, intermediate fusion, and late fusion. The previous multimodal fusion methods can be categorized into early fusion and late fusion in terms of their feature-level or decision-level fusion operations. In the last decades, massive works with deep learning suggest that intermediate fusion [13, 29, 31], which learns unified embeddings for multimodal data in the hidden layers of neural networks, could benefit representation learning. Although these multimodal fusion methods achieve remarkable improvements in kinds of multimodal learning tasks, most of them ignore the uncertainty in multimodal learning. Recently, several works [5, 10-13] empirically or theoretically demonstrated that conventional multimodal fusion methods had limited performance and robustness on noisy or corrupted multimodal data. Inspired by these pioneering works, we devised a novel intermediate multimodal fusion method, which could quantify the uncertainty and learn more robust joint representations." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.645, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Uncertainty in Deep Learning. In general, the uncertainties in deep learning can be categorized into epistemic uncertainty and aleatoric uncertainty [14]. The former aims at capturing the noise of the parameters in deep neural networks, while the latter measures the noise inherent in given training data. To improve the robustness and generalization of open-world scenarios, many researchers incorporate the uncertainty estimation into the deep learning models for computer vision tasks, such as face recognition [15, 16], semantic segmentation [32], and action localization [33, 34]. There is also a list of works that incorporate uncertainty estimation into multimodal learning tasks [5, 17, 22, 35]. However, most of them only quantified uncertainties to learn probability distribution representations that are used for cross-modal interactions, but ignore the value of intrinsic uncertainties and hardly take a further analysis on the robustness of multimodal fusion. Following this observation," + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "26877" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.137 + ], + "angle": 0, + "content": "we develop this work to learn more robust compact joint representations for multimodal learning, which embraces the uncertainties via self-supervised contrastive learning." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.14, + 0.471, + 0.457 + ], + "angle": 0, + "content": "Information Bottleneck Theory. The Information Bottleneck Theory [19] was initially formulated in the context of signal processing, which is proposed to discover a more concise representation of a signal while retaining its utmost informative content. Alemi et al. [36] proposed the Variational Information Bottleneck (VIB) to bridge the gap between Information Bottleneck Theory and deep learning, which approximates the information bottleneck constraints and enables it in deep learning. Based on the VIB, massive researchers introduce the Information Bottleneck Theory in deep learning models to tackle kinds of computer vision tasks, such as object detection [37-39] or image classification [37,38]. Furthermore, as the VIB aims at learning compact minimal representations, it also attracts wide attention from the multimodal learning field [28, 31]. For example, Mai et al. [28] devised a multimodal information bottleneck to learn the minimal sufficient unimodal and multimodal representations. Inspired by their successes, we also devise an multimodal integration strategy based on VIB to reduce the redundancy in joint representations, which further improves the robustness and effectiveness." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.481, + 0.251, + 0.498 + ], + "angle": 0, + "content": "3. Proposed Method" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.509, + 0.218, + 0.525 + ], + "angle": 0, + "content": "3.1. Preliminaries" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.536, + 0.47, + 0.809 + ], + "angle": 0, + "content": "Uncertainty Estimation in Deep Learning Models. In terms of [14], the uncertainties in deep learning models can be categorized into aleatoric uncertainty and epistemic uncertainty. The former refers to the uncertainty inherent in the observations and can not be explained away with more data, while the latter only accounts for uncertainty in the model and can be explained away given enough data. Typically, given a deep learning model \\( f_{\\theta}(\\cdot) \\), it conducts the mapping \\( \\mathcal{X} \\rightarrow \\mathcal{Y} \\), where \\( \\mathcal{X} \\) and \\( \\mathcal{Y} \\) represent input data and observed label space respectively. The aleatoric uncertainty in \\( \\mathcal{X} \\) will lead to a corrupted predicted results in \\( \\mathcal{Y} \\). To estimate the uncertainties in deep learning models, a widely adopted strategy is sampling \\( N \\) weights \\( \\theta \\sim p(\\theta | x, y) \\), \\( x \\in \\mathcal{X}, y \\in \\mathcal{Y} \\) via Dropout operations for different predictions for mean \\( \\mu \\) and variance \\( \\sigma^2 \\), and adopting a Gaussian posterior \\( p(y|x) \\sim \\mathcal{N}(\\mu_{*}(x), \\sigma_{*}^{2}(x)) \\). Here we can get the predictive variance and extract the aleatoric uncertainty which represents the ambiguity in \\( x \\):" + }, + { + "type": "equation", + "bbox": [ + 0.098, + 0.825, + 0.469, + 0.904 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mu_ {*} ^ {2} (x) = \\frac {1}{N} \\sum_ {i} \\mu_ {i} ^ {2} (x), \\\\ \\sigma_ {*} ^ {2} (x) = \\frac {1}{N} \\sum_ {i} \\sigma_ {i} ^ {2} (x) + \\frac {1}{N} \\sum_ {i} \\mu_ {i} ^ {2} (x) - \\mu_ {*} ^ {2} (x) \\tag {1} \\\\ = \\mathbb {E} _ {i} \\left[ \\sigma_ {i} ^ {2} (x) \\right] + \\operatorname {V a r} _ {i} [ \\mu_ {i} (x) ], \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.09, + 0.892, + 0.182 + ], + "angle": 0, + "content": "where \\(\\mathbb{E}_i\\left[\\sigma_i^2 (x)\\right]\\) represents the aleatoric uncertainty, while \\(\\operatorname {Var}_i[\\mu_i(x)]\\) is the epistemic uncertainty, which attributes to the model \\(f_{\\theta}(\\cdot)\\) rather than input data \\(\\mathcal{X}\\). In our work, we pay more attention to the aleatoric uncertainty since our goal is to quantify the intrinsic noise in multimodal data and leverage it for learning better joint representations." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.183, + 0.893, + 0.408 + ], + "angle": 0, + "content": "Variational Information Bottleneck. The Variational Information Bottleneck (VIB) [20] is an information-theoretic strategy widely adopted in deep learning models, which aims to maintain maximal discriminative feature representations with minimal redundancy. Specifically, given the input variable \\( \\mathbf{x} \\) with noise or redundant information and the target variable \\( \\mathbf{y} \\), the VIB is to learn the compressed latent variable \\( \\mathbf{z} \\), and \\( \\mathbf{z} \\) is maximally discriminative about the target variable \\( \\mathbf{y} \\). Moreover, since the input \\( \\mathbf{x} \\) is noisy and redundant, the VIB also requires \\( \\mathbf{z} \\) to be minimally discriminative about the original variable \\( \\mathbf{x} \\). In our work, we leverage the VIB to learn compact joint representations, which overcome the redundancy brought by highly aligned distributional multimodal representations to improve the robustness of multimodal fusion." + }, + { + "type": "image", + "bbox": [ + 0.502, + 0.42, + 0.892, + 0.523 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.535, + 0.892, + 0.577 + ], + "angle": 0, + "content": "Figure 2. An illustration of the Unimodal Distributional Representation process. For clarity, we visualize the multivariate Gaussian distributions with two variables only." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.591, + 0.842, + 0.607 + ], + "angle": 0, + "content": "3.2. Stable Unimodal Feature Augmentation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.614, + 0.892, + 0.719 + ], + "angle": 0, + "content": "As the data in different modalities contain unique noise in their own modal space, we first propose the Stable Unimodal Feature Augmentation (SUFA) module to quantify their intrinsic aleatoric uncertainty. Following the argument that the uncertainty is attributed to the fuzzy representation of semantics, we further leverage the aleatoric uncertainty via self-supervised contrastive learning." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.893, + 0.902 + ], + "angle": 0, + "content": "Unimodal Distributional Representation. Given the multimodal samples \\(\\mathbf{x}_i^m\\), \\(m \\in \\mathcal{M}\\) where \\(\\mathcal{M}\\) is the modality set including images, audio, text, etc., we learn the distributional representations to quantify the aleatoric uncertainty in each modality. According to the deviation listed in Eq. 1, we can observe that the aleatoric uncertainty can be predicted with deep learning models \\(f_{\\theta}(\\cdot)\\) as the variance \\(\\sigma_i^{m2}\\) directly if the epistemic uncertainty is not considered. To this end, we first employ the corresponding feature extractor to learn the preliminary embeddings for each modality, then deploy two additional fully connected layers to learn a mean vector \\(\\mu_i^{m2} \\in \\mathbb{R}^d\\) and a variance vector \\(\\sigma_i^{m2} \\in \\mathbb{R}^d\\)." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "26878" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.47, + 0.138 + ], + "angle": 0, + "content": "Furtherly, we define the representation \\(\\mathbf{z}_i^m\\) in latent space of each sample \\(\\mathbf{x}_i^m\\) as a multivariate Gaussian distribution with \\(d\\) variable, which can be represented as:" + }, + { + "type": "equation", + "bbox": [ + 0.152, + 0.148, + 0.396, + 0.187 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} p (\\mathbf {z} _ {i} ^ {m} | \\mathbf {x} _ {i} ^ {m}) \\sim \\mathcal {N} (\\boldsymbol {\\mu} _ {i} ^ {m}, \\boldsymbol {\\sigma} _ {i} ^ {m 2} \\mathbf {I}), \\\\ \\pmb {\\mu} _ {i} ^ {m} = f _ {\\theta_ {1} ^ {m}} \\left(\\mathbf {x} _ {i} ^ {m}\\right), \\pmb {\\sigma} _ {i} ^ {m} = f _ {\\theta_ {2} ^ {m}} \\left(\\mathbf {x} _ {i} ^ {m}\\right), \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.197, + 0.469, + 0.287 + ], + "angle": 0, + "content": "where \\(f_{\\theta_1^m}(\\cdot)\\) and \\(f_{\\theta_2^m}(\\cdot)\\) represent the two different fully connected layers for mean and variance respectively. \\(\\mathbf{I}\\) is the identity matrix. To maintain the consistency of semantics in different modalities, we further align the distributional representations of a multimodal sample with the Kullback-Leibler divergence:" + }, + { + "type": "equation", + "bbox": [ + 0.093, + 0.299, + 0.468, + 0.386 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} L _ {\\mathrm {U D R}} = \\sum_ {m _ {1} \\neq m _ {2}} ^ {m _ {1}, m _ {2} \\in \\mathcal {M}} K L \\left(p \\left(\\mathbf {z} _ {i} ^ {m _ {1}} \\mid \\mathbf {x} _ {i} ^ {m _ {1}}\\right) \\| p \\left(\\mathbf {z} _ {i} ^ {m _ {2}} \\mid \\mathbf {x} _ {i} ^ {m _ {2}}\\right)\\right) \\tag {3} \\\\ + \\sum_ {m} ^ {m \\in \\mathcal {M}} K L \\left(p \\left(\\mathbf {z} _ {i} ^ {m} \\mid \\mathbf {x} _ {i} ^ {m}\\right) \\| \\mathcal {N} (0, \\mathbf {I})\\right). \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.397, + 0.469, + 0.548 + ], + "angle": 0, + "content": "We illustrate the Unimodal Distributional Representation process in Fig. 2 with bi-modal input for clarity. It can be observed the \\(L_{\\mathrm{UDR}}\\) will push two multimodal distributional representations with similar semantics closer. In this way, the representation of each multimodal sample is not limited within a deterministic point embedding, but a consistent fuzzy representation on several multivariate Gaussian distributions. Particularly, the variance \\(\\sigma_i^m\\) reveals the aleatoric uncertainty in \\(m\\) modality and the mean \\(\\mu_i^m\\) is the corresponding stable representation." + }, + { + "type": "image", + "bbox": [ + 0.087, + 0.56, + 0.462, + 0.719 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.729, + 0.469, + 0.771 + ], + "angle": 0, + "content": "Figure 3. Illustration of the Uncertainty-based Representation Augmentation process. Here we show the distributional representations of image modality only for clarity." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.781, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Uncertainty-based Representation Augmentation. With the quantified aleatoric uncertainty, we consider the second question we raised in Sec.1: should we drop the aleatoric uncertainty in multimodal data? Intuitively, the aleatoric uncertainty in multimodal data is unavoidable since the natural ambiguity of semantics. However, it also results in the diversity of unimodal data in different modalities. To this end, we leverage the aleatoric uncertainty to generate the" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.258 + ], + "angle": 0, + "content": "unseen samples such that learned unimodal representations are insensitive to diverse unimodal input with similar semantics. Take the image modality as an example, we illustrate the Uncertainty-based Representation Augmentation process in Fig. 3. Specifically, given a unimodal distributional representation \\( p(\\mathbf{z}_i^m | \\mathbf{x}_i^m) \\sim \\mathcal{N}(\\boldsymbol{\\mu}_i^m, \\boldsymbol{\\sigma}_i^{m2}\\mathbf{I}) \\), we first random sample an anchor point \\( \\tilde{\\mathbf{z}}_i^m \\) and an augmented point \\( \\mathbf{z}_i^m \\) from the multivariate Gaussian distribution as matching pairs. Moreover, we randomly sample a set of negative points from other distributional representations and devise a self-supervised contrastive learning mechanism as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.507, + 0.266, + 0.892, + 0.324 + ], + "angle": 0, + "content": "\\[\nL _ {\\mathrm {U R A}} = - \\log \\frac {e ^ {\\operatorname* {s i m} \\left(\\tilde {\\mathbf {z}} _ {i} ^ {m} , \\mathbf {z} _ {i} ^ {m}\\right) / \\tau}}{\\sum_ {j \\neq i} \\left(e ^ {\\operatorname* {s i m} \\left(\\tilde {\\mathbf {z}} _ {i} ^ {m} , \\mathbf {z} _ {i} ^ {m}\\right) / \\tau} + e ^ {\\operatorname* {s i m} \\left(\\tilde {\\mathbf {z}} _ {i} ^ {m} , \\mathbf {z} _ {j} ^ {m}\\right) / \\tau}\\right)}, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.324, + 0.892, + 0.383 + ], + "angle": 0, + "content": "where \\(\\tau\\) is a temperature factor and sim is cosine similarity calculation. Here we adopt the re-parameterization trick [40] to conduct the sampling operations, which can be formulated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.585, + 0.395, + 0.891, + 0.412 + ], + "angle": 0, + "content": "\\[\n\\mathbf {z} _ {i} ^ {m} = \\boldsymbol {\\mu} _ {i} ^ {m} + \\epsilon \\sigma_ {i} ^ {m}, \\epsilon \\sim \\mathcal {N} (0, \\mathbf {I}). \\tag {5}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.421, + 0.842, + 0.437 + ], + "angle": 0, + "content": "3.3. Robust Multimodal Feature Integration" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.444, + 0.892, + 0.549 + ], + "angle": 0, + "content": "With the SUFA, we obtain stable unimodal representations with consistent semantics. However, we only consider the consistency in each modality but ignore the redundancy caused by duplicated representations. To this end, we proposed the Robust Multimodal Feature Integration (RMFI) module in this section. The overview of the RMFI module is illustrated in Fig. 4." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.551, + 0.892, + 0.701 + ], + "angle": 0, + "content": "Dynamic Multimodal Integration. Inspired by dynamic multimodal fusion [5], we assume that different modalities have unequal contributions to the joint representations in terms of the observed label space. To this end, we first apply a Dynamic Multimodal Integration strategy based on an attention mechanism. Specifically, given the stable unimodal representations \\(\\pmb{\\mu}_i^m\\), we calculate the attentive weights in the distributional representations across modalities according to their quantified uncertainties \\(\\sigma_i^m\\), and apply them for multimodal integration:" + }, + { + "type": "equation", + "bbox": [ + 0.551, + 0.711, + 0.892, + 0.757 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathbf {x}} _ {i} = \\sum_ {m \\in \\mathcal {M}} \\alpha_ {i} ^ {m} \\boldsymbol {\\mu} _ {i} ^ {m}, \\quad \\alpha_ {i} ^ {m} = \\frac {e ^ {\\frac {1}{\\overline {{\\sigma}} _ {i} ^ {m}}}}{\\sum_ {m \\in \\mathcal {M}} e ^ {\\frac {1}{\\overline {{\\sigma}} _ {i} ^ {m}}}}, \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.892, + 0.854 + ], + "angle": 0, + "content": "where \\(\\hat{\\mathbf{x}}_i, \\overline{\\sigma}_i^m\\) represent the integrated joint representations and the average variance of the multivariate Gaussian distribution of \\(m\\) modality. In this way, we preliminarily integrate stable unimodal representations into joint representations, where the contributions of different modalities are estimated dynamically." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.892, + 0.9 + ], + "angle": 0, + "content": "Joint Representation Compression. In the SUFA module, we fully consider the consistency of semantics in different modalities to avoid noise in unimodal data. However," + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "26879" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.09, + 0.468, + 0.199 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.21, + 0.47, + 0.253 + ], + "angle": 0, + "content": "Figure 4. An illustration of the RMFI module. We provide Cross-Entropy loss and \\( \\mathrm{L}_1 \\) loss for classification-based and regression-based downstream tasks respectively." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.264, + 0.468, + 0.384 + ], + "angle": 0, + "content": "with similar multivariate Gaussian distributions among different modalities, redundant duplicated information will be introduced into joint representations. Hence, we devise a joint representation compression with the Variational Information Bottleneck (VIB) [20]. Specifically, given the preliminary joint representation \\(\\hat{\\mathbf{x}}_i\\) and the target observations \\(\\mathbf{y}\\) in label space, we learn the compressed joint representation \\(\\hat{\\mathbf{z}}\\) in a latent space:" + }, + { + "type": "equation", + "bbox": [ + 0.076, + 0.396, + 0.468, + 0.427 + ], + "angle": 0, + "content": "\\[\np \\left(\\hat {\\mathbf {z}} _ {i} \\mid \\hat {\\mathbf {x}} _ {i}\\right) \\sim \\mathcal {N} \\left(\\hat {\\boldsymbol {\\mu}} _ {i}, \\hat {\\boldsymbol {\\sigma}} _ {i} ^ {2} \\mathbf {I}\\right), \\quad \\hat {\\boldsymbol {\\mu}} _ {i} = f _ {\\hat {\\theta} _ {1}} (\\hat {\\mathbf {x}} _ {i}), \\quad \\hat {\\boldsymbol {\\sigma}} _ {i} = f _ {\\hat {\\theta} _ {2}} (\\hat {\\mathbf {x}} _ {i}). \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.428, + 0.468, + 0.473 + ], + "angle": 0, + "content": "where \\( f_{\\hat{\\theta}_1} \\) and \\( f_{\\hat{\\theta}_2} \\) are two fully connected layers. Similarly, we also adopt the re-parameterization tricks [40] for the final compressed joint representations:" + }, + { + "type": "equation", + "bbox": [ + 0.173, + 0.485, + 0.468, + 0.501 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathbf {z}} _ {i} = \\hat {\\boldsymbol {\\mu}} _ {i} + \\epsilon \\hat {\\boldsymbol {\\sigma}} _ {i}, \\epsilon \\sim \\mathcal {N} (0, \\mathbf {I}). \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.513, + 0.468, + 0.573 + ], + "angle": 0, + "content": "As our proposed method can be applied to different downstream tasks, here we provide two training objectives for the joint representation compression. Specifically, we employ the cross-entropy for the classification-based task:" + }, + { + "type": "equation", + "bbox": [ + 0.084, + 0.584, + 0.468, + 0.615 + ], + "angle": 0, + "content": "\\[\nL _ {\\mathrm {J R C}} = - \\mathbf {y} ^ {\\top} \\log \\delta \\left(f _ {\\theta} \\left(\\hat {\\mathbf {z}} _ {i}\\right)\\right) + \\lambda K L \\left(p \\left(\\hat {\\mathbf {z}} _ {i} \\mid \\hat {\\mathbf {x}} _ {i}\\right) \\| \\mathcal {N} (0, \\mathbf {I})\\right), \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.616, + 0.468, + 0.676 + ], + "angle": 0, + "content": "where \\(\\delta (\\cdot)\\) represents softmax function and \\(f_{\\theta}(\\cdot)\\) is the deep learning models for classification, \\(\\lambda\\) is a hyperparameter. As for the regression-based task, we employ the mean square error as the training objective:" + }, + { + "type": "equation", + "bbox": [ + 0.095, + 0.687, + 0.468, + 0.705 + ], + "angle": 0, + "content": "\\[\nL _ {\\mathrm {J R C}} = \\left[ y - f _ {\\theta} \\left(\\hat {\\mathbf {z}} _ {i}\\right) \\right] ^ {2} + K L \\left(p \\left(\\hat {\\mathbf {z}} _ {i} \\mid \\hat {\\mathbf {x}} _ {i}\\right) \\| \\mathcal {N} (0, \\mathbf {I})\\right). \\tag {10}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.717, + 0.209, + 0.734 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.742, + 0.283, + 0.759 + ], + "angle": 0, + "content": "4.1. Experimental Settings" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.468, + 0.901 + ], + "angle": 0, + "content": "Datasets. We evaluate our method on five multimodal datasets, including two tri-modal datasets, i.e., CMU-MOSI [41] and CMU-MOSEI [42] and three bi-modal datasets, i.e., MVSA-Single [43], UPMC Food101 [1], and NYU Depth v2 [44] datasets. The CMU-MOSI [41] and CMU-MOSEI [42] datasets are widely employed video datasets for Multimodal Sentiment Analysis (MSA) task. The former encompasses 2199 short videos, each accompanied by a sentiment strength score ranging from -3 to 3, offering" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.348 + ], + "angle": 0, + "content": "a nuanced measure of emotional intensity. The latter, designed specifically for emotion analysis tasks, is a larger dataset comprising 22,856 movie review clips. The bimodal dataset, i.e., MVSA-Single [43] is also used for MSA task but it only contains image-text pairs obtained from social media for sentiment classification. Following [2-5], we also employ the UPMC Food101 [1] and the NYU Depth v2 [44] datasets for Multimodal Image Classification (MIC) task. The UPMC Food101 dataset contain 101 categories, respectively, these datasets are specifically designed for bimodal classification tasks. The UPMC FOOD101 dataset contains images with 101 categories obtained by Google Image Search and their corresponding textual descriptions. The NYU Depth v2 dataset [44] is utilized for scene recognition and comprises depth and RGB images. Following [5], we also adopt the commonly used 9 out of the 27 scene categories and the remaining categories as \"Others\"." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.349, + 0.892, + 0.515 + ], + "angle": 0, + "content": "Implementation Details. In alignment with alternative methodologies, for our trimodal video dataset, we employ FACET, COVAREP, and BERT as feature extractors for visual, audio, and text modalities, respectively. For the bimodal dataset, we utilize ResNet-152 as the feature extractor for RGB and depth images, coupled with BERT as the feature extractor for text. As for the hyperparameters, we set the temperature factor \\(\\tau\\) and the balance factor \\(\\lambda\\) to 0.5 and 1e-3 respectively. We employed the Adam optimizer with learning rate of 1e-5,and adopt the Reduce-on-Plateau learning rate adjustment strategy to train our EAU method." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.516, + 0.892, + 0.637 + ], + "angle": 0, + "content": "Evaluation Metrics. For the regression-based MSA task based on CMU-MOSI and CMU-MOSEI, we follow the previous work [6, 7, 9, 45] and adopt Acc7, F1 score, and Pearson correlation coefficient as our metrics. As for the classification-based MSA task on the MSVA-Single dataset and MIC task on the UPMC FOOD101 and The NYU Depth v2 datasets, we reported commonly used metrics, including accuracy and F1 score." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.65, + 0.822, + 0.666 + ], + "angle": 0, + "content": "4.2. Comparisons with State-Of-The-Arts" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.674, + 0.892, + 0.869 + ], + "angle": 0, + "content": "To show the superiority of our proposed EAU method on multimodal fusion, we compared our EAU method with existing state-of-the-art methods of the MSA task on the CMU-MOSI and CMU-MOSEI datasets, including MIB [28], HMA [6], MIM [7], GCNet [46], ConFEDE [8], DiC-MoR [45], and DMD [9]. Moreover, we also make fair comparisons with the widely adopted simple multimodal fusion strategies i.e., Concat and Late Fusion, as well as recent well-designed multimodal fusion strategies on the bimodal datasets, including TMC [3], ITIN [47], MMBT [2], PMF [4], MVCN [48], and QMF [5]. Particularly, similar to the counterpart method QMF [5], we evaluate our method on noisy datasets to observe the model robustness." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Multimodal Fusion on Classification Task. We report the performance of classification tasks, including both MSA" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "26880" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.084, + 0.089, + 0.49, + 0.226 + ], + "angle": 0, + "content": "
MethodCMU-MOSICMU-MOSEI
Acc7F1CorrAcc7F1Corr
MIB [28] (2022)48.685.30.79854.186.20.790
HMA [6] (2023)45.385.60.78252.885.40.787
MIM [7] (2023)47.085.90.80552.586.30.792
GCNet [46] (2023)44.985.1-51.585.2-
ConFEDE [8] (2023)42.385.50.78454.985.80.780
DiCMoR [45] (2023)45.385.6-53.485.1-
DMD [9] (2023)45.686.0-54.586.6-
EAU (Ours)48.886.20.80954.886.90.816
" + }, + { + "type": "table", + "bbox": [ + 0.508, + 0.09, + 0.888, + 0.226 + ], + "angle": 0, + "content": "
MethodMVSA-SingleFood 101NYU Depth v2
AccF1AccF1AccF1
Concat65.5965.4388.2088.1970.3069.82
Late Fusion76.8875.7290.6990.7769.1468.32
MMBT [2] (2020)78.50-91.5291.2871.04-
TMC [3] (2021)76.0674.5589.8689.8071.0669.83
ITIN [47] (2022)75.1974.97----
PMF [4] (2023)--91.68---
MVCN [48] (2023)76.0674.55----
QMF [5] (2023)78.0777.1892.9292.9370.0968.65
EAU (Ours)79.1578.3693.2093.1872.0570.63
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.232, + 0.895, + 0.288 + ], + "angle": 0, + "content": "Table 1. Comparisons with state-of-the-art multimodal fusion methods on MSA and MIC tasks. The CMU-MOSI and CMU-MOSEI datasets contain videos, audi, and texts. The MVSA-Single and Food-101 datasets consist of texts and RGB images. The NYU Depth v2 contains RGB and depth images. Note that CMU-MOSI and CMU-MOSEI are used for regression-based MSA tasks, while the others are used for classification-based MSA or MIC tasks." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.301, + 0.473, + 0.543 + ], + "angle": 0, + "content": "and MIC tasks on bi-modal datasets in Table 1, where the best results are marked in bold. From the experimental results, we can observe that our proposed EAU method outperforms all three datasets. Particularly, compared with the recent state-of-the-art method QMF method [5], our EAU method achieves at least \\(1\\%\\) absolute improvements on the MVSA-Single and NYU Depth v2 datasets. These results demonstrate that our EAU approach conducts better multimodal fusion on text, depth, and RGB images, affirming the effectiveness of leveraging aleatoric uncertainty in multimodal data. However, we also note that the improvements on UPMC Food 101 are not as remarkable as the other two benchmark datasets. We speculate that one probable reason is the images and texts in this dataset are clear compared with the others. Hence, our EAU method obtains fewer improvements since the augmentation is limited." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.548, + 0.47, + 0.79 + ], + "angle": 0, + "content": "Multimodal Fusion on Regression Task. We evaluate our method on the regression task on the CMU-MOSEI and CMU-MOSI datasets, where the MSA task is conducted by predicting sentiment strengths. From the experimental results in Table 1, we can observe that: For the CMU-MOSI dataset, our proposed EAU method achieves state-of-the-art performance across all evaluation metrics, with a particularly significant improvement noted in Acc7. On the CMU-MOSEI dataset, our method has also demonstrated substantial improvements on most evaluation metrics. These results demonstrate the superiority of our proposed EAU, which benefits from high-quality multimodal fusion via learning stable and robust joint representations. Simultaneously, these results indicate that our approach can effortlessly extend to other modalities and can be seamlessly transferred to frameworks accommodating kinds of modalities." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.796, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Model Robustness on Noisy Multimodal Datasets. To validate the effectiveness of our model in handling data noise, we conducted more evaluation on the noisy datasets following a recent multimodal fusion strategy QMF [5]. Concretely, we consider different intensities of Gaussian and Salt-Pepper noise on the MVSA-Single and NYU Depth v2 datasets. For fair comparisons, we conducted 10" + }, + { + "type": "table", + "bbox": [ + 0.505, + 0.298, + 0.889, + 0.421 + ], + "angle": 0, + "content": "
Noisy MVSA-Single
MethodCleanSalt-Pepper NoiseGaussian Noise
ε = 0ε = 5ε = 10ε = 5ε = 10
Concat65.5958.6951.1650.7046.12
Late Fusion76.8867.8855.4363.4655.16
MMBT [2] (2020)78.5074.0751.2671.9955.34
TMC [3] (2021)74.8768.0256.6266.7260.35
QMF [5] (2023)78.0773.9060.4173.8561.28
EAU (Ours)79.1574.8161.0473.8962.04
" + }, + { + "type": "table", + "bbox": [ + 0.505, + 0.422, + 0.889, + 0.542 + ], + "angle": 0, + "content": "
Noisy NYU Depth v2
MethodCleanSalt-Pepper NoiseGaussian Noise
ε = 0ε = 5ε = 10ε = 5ε = 10
Concat70.4457.9844.5159.9753.20
Late fusion69.1656.2741.2259.6351.99
MMTM [2] (2020)71.0459.4544.5960.3752.28
TMC [3] (2021)71.0659.3444.6561.0453.36
QMF [5] (2023)70.0958.5045.6961.6255.60
EAU (Ours)72.0559.8346.8563.3358.85
" + }, + { + "type": "table_caption", + "bbox": [ + 0.5, + 0.548, + 0.892, + 0.577 + ], + "angle": 0, + "content": "Table 2. Comparisons with state-of-the-arts concerning model performance on noisy MVSA-Single and NYU Depth v2 datasets." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.592, + 0.892, + 0.819 + ], + "angle": 0, + "content": "experiments with different random seeds and reported the mean results as our final performance. According to the experimental results listed in Table 2, we can observe that our method has achieved state-of-the-art performance under various types and intensities of noise. Moreover, with the intensity of noise increasing, the margin of performance between our method and existing multimodal fusion strategies goes larger. Particularly, our method achieves more than \\(3\\%\\) improvement on the Noisy NYU Depth v2 dataset over the QMF method [5]. Such a result demonstrates our method reveals better robustness against data noise. The reason is our model pioneeringly leverages the intrinsic aleatoric uncertainty in the training data to enhance the stable feature representation, which is not fully considered in existing counterpart methods [2, 3, 5]." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.831, + 0.669, + 0.848 + ], + "angle": 0, + "content": "4.3. Further Analysis" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.856, + 0.894, + 0.901 + ], + "angle": 0, + "content": "Analysis on Model Structure. To explore the impact of different model structures, we decomposed our proposed method into two components, SUFA and RMFI, and evalu" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "26881" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.1, + 0.09, + 0.666, + 0.233 + ], + "angle": 0, + "content": "
MethodCleanSalt-Pepper NoiseGaussian Noise
Acc@ε=0Acc@ε=5Acc@ε=10Acc@ε=5Acc@ε=10
Naive Backbone + Late Fusion76.88 ± 1.3067.88 ± 1.8755.43 ± 1.9463.46 ± 3.4655.16 ± 3.60
Naive Backbone + Concat74.53 ± 0.9762.24 ± 2.4753.14 ± 2.8956.08 ± 1.8949.21 ± 2.81
Naive Backbone + RMFI77.23 ± 0.7470.99 ± 2.7759.11 ± 1.9266.53 ± 2.2652.02 ± 3.45
SUFA + Late Fusion77.11 ± 1.0570.05 ± 1.5055.93 ± 2.2562.27 ± 4.2055.90 ± 3.67
SUFA + Concat78.18 ± 1.0873.10 ± 1.0760.67 ± 1.3871.40 ± 1.4859.11 ± 1.60
BERT (NAACL'19) [49]75.61 ± 0.5369.50 ± 1.5047.41 ± 0.7969.50 ± 1.5047.41 ± 0.79
MMBT (arXiv'19) [2]78.50 ± 0.4074.07 ± 1.1251.26 ± 5.6571.99 ± 1.0455.34 ± 2.84
TMC (ICLR'21) [3]74.87 ± 2.2468.02 ± 3.0756.62 ± 3.6766.72 ± 4.5560.35 ± 2.79
QMF (ICML'23) [5]78.07 ± 1.1073.90 ± 1.8960.41 ± 2.6373.85 ± 1.4261.28 ± 2.12
SUFA + RMFI (EAU)79.15 ± 0.6074.81 ± 1.5961.04 ± 1.3173.89 ± 0.9662.04 ± 1.26
" + }, + { + "type": "table", + "bbox": [ + 0.671, + 0.09, + 0.868, + 0.233 + ], + "angle": 0, + "content": "
ModalityAccF1Corr
T44.183.70.785
V16.745.30.072
A15.848.70.099
T+A46.784.20.794
T+V46.384.60.786
A+V20.255.30.117
T+A+V48.886.20.809
T76.3075.90-
V63.5863.35-
T+V79.1578.36-
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.239, + 0.892, + 0.268 + ], + "angle": 0, + "content": "Table 3. Analysis concerning key model structures in our EAU method on the Noisy MSVA-Single (Left), and analysis concerning modalities on the CMU-MOSI (Right Upper) and MVSA-Single (Right Lower) datasets. Note that BERT [49] adopts text modality only." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.279, + 0.47, + 0.476 + ], + "angle": 0, + "content": "ated their effectiveness in combination with various model structures on the Noisy MVSA-Single dataset. Specifically, we employ two alternative designs for SUFA and RMFI respectively: (1) Naive Backbone: deploying the feature extractors directly, where the distributional representation and stable feature augmentation in the SUFA module are disabled. (2) Late Fusion: the unimodal representations are used for downstream tasks first and the fusion operations are conducted in probability space. To observe the stability of our method, we conduct 10 experiments with different random seeds and report the mean and variance in Table 3. Here we also make fair comparisons with several counterpart methods [2, 3, 5, 49] to furtherly show our superiority." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.478, + 0.47, + 0.719 + ], + "angle": 0, + "content": "According to the experimental results presented in Table 3, we can observe that: (1) Our proposed SUFA and RMFI modules consistently improve performance on all metrics. Particularly, with the SUFA module, the results of fusion strategy Concat are boosted with more than \\(10\\%\\) absolute performance on the noisy data. Moreover, compared with Late Fusion, our proposed RMFI module can also significantly boost the classification accuracy on noisy data with the SUFA module. These results prove again that our method shows great superiority in the effectiveness and robustness of multimodal fusion. (2) It also explicates that our proposed EAU method remarkably outperforms compared with the counterpart methods [2, 3, 5]. Particularly, compared with the recent state-of-the-art method QMF [5], our complete model reveals better performance and stability, demonstrating the capability of our EAU method again." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.72, + 0.47, + 0.901 + ], + "angle": 0, + "content": "Analysis on the Effectiveness of Multimodal Fusion. To verify the effectiveness of multimodal fusion in our EAU method, we also conduct ablation studies concerning different modalities on the CMU-MOSI datasets, which consist of video (V), text (T), and audio (A) three modalities. By observing the experimental results illustrated in Table 3, we have drawn the following conclusions: (1) Our fusion strategy demonstrated significant effectiveness across different modal combinations. As the number of modalities increased, our approach facilitated the integration of multimodal information, resulting in a significant improvement across all evaluation metrics. (2) We also note that the text" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.279, + 0.892, + 0.384 + ], + "angle": 0, + "content": "modality performs an essential role in the MSA task on the CMU-MOSI dataset. However, combined with the other two modalities via our proposed EAU method, the results are also boosted significantly, demonstrating the superiority of our EAU method again. Such a phenomenon also proves the rationality of our Dynamic Multimodal Integration process in the RMFI module." + }, + { + "type": "image", + "bbox": [ + 0.501, + 0.392, + 0.697, + 0.527 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.697, + 0.392, + 0.892, + 0.527 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.538, + 0.892, + 0.566 + ], + "angle": 0, + "content": "Figure 5. Analysis for robustness evaluated by over 10 times random experiments under different noisy data." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.569, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Analysis on Model Robustness. To verify the robustness and stability of our proposed EAU method, we conduct more ablation studies on the MVSA-Single and NYU Depth v2 datasets under different noise levels and observe the sparsity of final accuracy. Specifically, we conduct more than 10 experiments randomly with complete EAU (denoted as Complete) and EAU w/o SUFA (denoted as w/o SUFA), and show the statistics in Fig. 5. By observing the experimental results, we can find that: (1) The SUFA module consistently shows its effectiveness on both two datasets under different noise strengths. Particularly, when introducing more noise to the data, the ablated model shows strong fluctuations. Contrastively, the complete EAU model reveals remarkably reduced fluctuations, which demonstrates the superiority of model robustness and stability. (2) We also note that the improvements on the NYU Depth v2 are slightly lower than the MVSA-Single dataset, though both two datasets have significant results. We speculate this is caused by the difference between text modality and depth images. Concretely, the depth images can be regarded as another view of spatial information, which has less complementary information for RGB images compared with texts." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "26882" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.47, + 0.122 + ], + "angle": 0, + "content": "To this end, the introduced noise will show more impact on the depth-RGB than the image-text compositions." + }, + { + "type": "image", + "bbox": [ + 0.078, + 0.133, + 0.47, + 0.257 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.262, + 0.468, + 0.289 + ], + "angle": 0, + "content": "Figure 6. Analysis concerning the training convergence and performance fluctuations during training." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.298, + 0.471, + 0.525 + ], + "angle": 0, + "content": "Analysis on Training Process. We also illustrate the training process of our proposed EAU method on the MVSA-Single in Fig 6 to observe the training convergence and performance fluctuations. According to the illustrated experimental results, we can observe that the complete EAU model exhibits a smoother training process and converges more rapidly with a remarkably better performance. Particularly, compared with the ablated model w/o RMFI, the other two models show much better stability during training. This is because the fused features learned by the ablated model w/o RMFI have redundant information caused by cross-modal distributional alignment in the SUFA module. It proves again the rationality of our proposed RMFI module, which deploys an information-theoretic strategy to address the limitation." + }, + { + "type": "image", + "bbox": [ + 0.083, + 0.535, + 0.255, + 0.636 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.147, + 0.639, + 0.196, + 0.649 + ], + "angle": 0, + "content": "(a) Concat" + }, + { + "type": "image", + "bbox": [ + 0.28, + 0.534, + 0.47, + 0.636 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.352, + 0.639, + 0.438, + 0.65 + ], + "angle": 0, + "content": "(b) EAU w/o SUFA" + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.652, + 0.256, + 0.752 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.131, + 0.755, + 0.213, + 0.765 + ], + "angle": 0, + "content": "(c) EAU w/o RMFI" + }, + { + "type": "image", + "bbox": [ + 0.277, + 0.651, + 0.47, + 0.751 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.372, + 0.754, + 0.41, + 0.765 + ], + "angle": 0, + "content": "(d) EAU" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.776, + 0.47, + 0.817 + ], + "angle": 0, + "content": "Figure 7. Visualizations of joint representations by t-SNE [50] on MVSA-Single dataset. Here we adopt Concat for the ablated model w/o RMFI." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.47, + 0.901 + ], + "angle": 0, + "content": "Visualizations of Joint Representations. To further verify the superiority of our proposed EAU method, we also employ the t-SNE [50] to visualize the learned joint representations. As illustrated in Fig. 7, it can be observed that the features generated by our method exhibit a more" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.182 + ], + "angle": 0, + "content": "compact and discriminative distribution. In contrast, when the SUFA or RMFI module is removed, more sparse points appear around the feature distribution which leads to corrupted classification results at last. Such results indicate that our method is more capable of learning a representative joint representation." + }, + { + "type": "image", + "bbox": [ + 0.502, + 0.189, + 0.891, + 0.351 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.361, + 0.892, + 0.403 + ], + "angle": 0, + "content": "Figure 8. Visualizations of test cases selected from the MSVA-Single datasets. It can be observed that our EAU method reveals better robustness against the noise." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.408, + 0.892, + 0.56 + ], + "angle": 0, + "content": "Qualitative Analysis. Moreover, we also illustrate several typical test cases in Fig. 8 and make comparisons with the recent counterpart method QMF [5]. Under different noise settings, our proposed EAU method consistently classifies the sentiment of the given image-text input precisely, while the QMF method returns incorrect results. It explicates that our proposed EAU method benefits from the well-designed SUFA and RMFI module, which is effective in learning more stable and robust joint representations for noise-resistant performance with higher accuracy." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.575, + 0.619, + 0.59 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.601, + 0.892, + 0.798 + ], + "angle": 0, + "content": "In this paper, we proposed a novel multimodal fusion method, namely, Embracing Aleatoric Uncertainty (EAU). It achieves more discriminative joint representations by fully considering the aleatoric uncertainty in multimodal data. Specifically, with the well-designed Stable Unimodal Feature Augmentation (SUFA) and Robust Multimodal Feature Integration (RMFI) modules, our proposed EAU could learn compact and robust joint representations. We evaluated our proposed EAU method on five multimodal benchmark datasets for both classification and regression tasks and demonstrated its superiority in fusion performance and robustness. For future work, we will further explore the uncertainties in multimodal learning." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.815, + 0.68, + 0.832 + ], + "angle": 0, + "content": "6. Acknowledgement" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.892, + 0.901 + ], + "angle": 0, + "content": "This work was sponsored in part by the National Natural Science Foundation of China under Grants (No. 62222203 and No. 62072080) and the New Cornerstone Science Foundation through the XPLORER PRIZE." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "26883" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.175, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.17 + ], + "angle": 0, + "content": "[1] Lukas Bossard, Matthieu Guillaumin, and Luc Van Gool. Food-101-mining discriminative components with random forests. In European Conference on Computer Vision, pages 446-461, 2014. 1, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.173, + 0.472, + 0.229 + ], + "angle": 0, + "content": "[2] Douwe Kiela, Suvrat Bhooshan, Hamed Firooz, Ethan Perez, and Davide Testuggine. Supervised multimodal bittransformers for classifying images and text. arXiv preprint arXiv:1909.02950, 2019. 1, 2, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.232, + 0.471, + 0.274 + ], + "angle": 0, + "content": "[3] Zongbo Han, Changqing Zhang, Huazhu Fu, and Joey Tianyi Zhou. Trusted multi-view classification. arXiv preprint arXiv:2102.02051, 2021. 1, 2, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.277, + 0.471, + 0.332 + ], + "angle": 0, + "content": "[4] Yaowei Li, Ruijie Quan, Linchao Zhu, and Yi Yang. Efficient multimodal fusion via interactive prompting. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 2604-2613, 2023. 1, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.335, + 0.47, + 0.392 + ], + "angle": 0, + "content": "[5] Qingyang Zhang, Haitao Wu, Changqing Zhang, Qinghua Hu, Huazhu Fu, Joey Tianyi Zhou, and Xi Peng. Provable dynamic fusion for low-quality multimodal data. arXiv preprint arXiv:2306.02050, 2023. 1, 2, 4, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.394, + 0.47, + 0.449 + ], + "angle": 0, + "content": "[6] Ronghao Lin and Haifeng Hu. Dynamically shifting multimodal representations via hybrid-modal attention for multimodal sentiment analysis. IEEE Transactions on Multimedia, 2023. 1, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.452, + 0.469, + 0.507 + ], + "angle": 0, + "content": "[7] Ying Zeng, Sijie Mai, Wenjun Yan, and Haifeng Hu. Multimodal reaction: Information modulation for cross-modal representation learning. IEEE Transactions on Multimedia, 2023. 1, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.511, + 0.469, + 0.58 + ], + "angle": 0, + "content": "[8] Jiuding Yang, Yakun Yu, Di Niu, Weidong Guo, and Yu Xu. Confede: Contrastive feature decomposition for multimodal sentiment analysis. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics, pages 7617-7630, 2023. 1, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.583, + 0.469, + 0.639 + ], + "angle": 0, + "content": "[9] Yong Li, Yuanzhi Wang, and Zhen Cui. Decoupled multimodal distilling for emotion recognition. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 6631-6640, 2023. 1, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.642, + 0.469, + 0.698 + ], + "angle": 0, + "content": "[10] Weiyao Wang, Du Tran, and Matt Feiszli. What makes training multi-modal classification networks hard? In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 12695-12705, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.701, + 0.469, + 0.769 + ], + "angle": 0, + "content": "[11] Xiaokang Peng, Yake Wei, Andong Deng, Dong Wang, and Di Hu. Balanced multimodal learning via on-the-fly gradient modulation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 8238-8247, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.773, + 0.469, + 0.843 + ], + "angle": 0, + "content": "[12] Yu Huang, Junyang Lin, Chang Zhou, Hongxia Yang, and Longbo Huang. Modality competition: What makes joint training of multi-modal network fail in deep learning? (provably). In International Conference on Machine Learning, pages 9226-9259. PMLR, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.846, + 0.469, + 0.888 + ], + "angle": 0, + "content": "[13] Yunfeng Fan, Wenchao Xu, Haozhao Wang, Junxiao Wang, and Song Guo. PMR: prototypical modal rebalance for multimodal learning. In Proceedings of the IEEE Conference" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.472, + 0.888 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "on Computer Vision and Pattern Recognition, pages 20029-20038. IEEE, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.125, + 0.892, + 0.168 + ], + "angle": 0, + "content": "[14] Alex Kendall and Yarin Gal. What uncertainties do we need in bayesian deep learning for computer vision? Advances in neural information processing systems, 30, 2017. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.172, + 0.892, + 0.227 + ], + "angle": 0, + "content": "[15] Jie Chang, Zhonghao Lan, Changmao Cheng, and Yichen Wei. Data uncertainty learning in face recognition. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 5710-5719, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.232, + 0.892, + 0.274 + ], + "angle": 0, + "content": "[16] Yichun Shi and Anil K Jain. Probabilistic face embeddings. In Proceedings of the IEEE International Conference on Computer Vision, pages 6902-6911, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.278, + 0.892, + 0.361 + ], + "angle": 0, + "content": "[17] Yatai Ji, Junjie Wang, Yuan Gong, Lin Zhang, Yanru Zhu, Hongfa Wang, Jiaxing Zhang, Tetsuya Sakai, and Yujiu Yang. Map: Multimodal uncertainty-aware vision-language pre-training model. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 23262-23271, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.366, + 0.892, + 0.407 + ], + "angle": 0, + "content": "[18] Luke Vilnis and Andrew McCallum. Word representations via gaussian embedding. arXiv preprint arXiv:1412.6623, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.413, + 0.892, + 0.455 + ], + "angle": 0, + "content": "[19] Naftali Tishby, Fernando C Pereira, and William Bialek. The information bottleneck method. arXiv preprint physics/0004057, 2000. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.46, + 0.892, + 0.502 + ], + "angle": 0, + "content": "[20] Alexander A Alemi, Ian Fischer, Joshua V Dillon, and Kevin Murphy. Deep variational information bottleneck. arXiv preprint arXiv:1612.00410, 2016. 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.506, + 0.892, + 0.576 + ], + "angle": 0, + "content": "[21] Xun Jiang, Xing Xu, Zhiguo Chen, Jingran Zhang, Jingkuan Song, Fumin Shen, Huimin Lu, and Heng Tao Shen. DHHN: dual hierarchical hybrid network for weakly-supervised audio-visual video parsing. In ACM International Conference on Multimedia, pages 719–727, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.581, + 0.892, + 0.65 + ], + "angle": 0, + "content": "[22] Junyu Gao, Mengyuan Chen, and Changsheng Xu. Collecting cross-modal presence-absence evidence for weakly-supervised audio-visual event perception. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 18827–18836, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.655, + 0.892, + 0.723 + ], + "angle": 0, + "content": "[23] Zhiguo Chen, Xun Jiang, Xing Xu, Zuo Cao, Yijun Mo, and Heng Tao Shen. Joint searching and grounding: Multigranularity video content retrieval. In Proceedings of the 31st ACM International Conference on Multimedia, pages 975-983, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.729, + 0.892, + 0.799 + ], + "angle": 0, + "content": "[24] Shenshen Li, Xing Xu, Xun Jiang, Fumin Shen, Xin Liu, and Heng Tao Shen. Multi-grained attention network with mutual exclusion for composed query-based image retrieval. IEEE Transactions on Circuits and Systems for Video Technology, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.803, + 0.892, + 0.873 + ], + "angle": 0, + "content": "[25] Shenshen Li, Xing Xu, Xun Jiang, Fumin Shen, Zhe Sun, and Andrzej Cichocki. Cross-modal attention preservation with self-contrastive learning for composed query-based image retrieval. ACM Transactions on Multimedia Computing, Communications, and Applications, 20(6), 2024. 2" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.892, + 0.873 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "26884" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.161 + ], + "angle": 0, + "content": "[26] Xing Xu, Kaiyi Lin, Yang Yang, Alan Hanjalic, and Heng Tao Shen. Joint feature synthesis and embedding: Adversarial cross-modal retrieval revisited. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(6):3030-3047, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.167, + 0.47, + 0.236 + ], + "angle": 0, + "content": "[27] Shenshen Li, Xing Xu, Xun Jiang, Fumin Shen, Xin Liu, and Heng Tao Shen. Multi-grained attention network with mutual exclusion for composed query-based image retrieval. IEEE Transactions on Circuits and Systems for Video Technology, pages 1-1, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.241, + 0.469, + 0.297 + ], + "angle": 0, + "content": "[28] Sijie Mai, Ying Zeng, and Haifeng Hu. Multimodal information bottleneck: Learning minimal sufficient unimodal and multimodal representations. IEEE Transactions on Multimedia, 2022. 2, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.301, + 0.469, + 0.357 + ], + "angle": 0, + "content": "[29] Xun Jiang, Xing Xu, Jingran Zhang, Fumin Shen, Zuo Cao, and Heng Tao Shen. Sdn: Semantic decoupling network for temporal language grounding. IEEE Transactions on Neural Networks and Learning Systems, pages 1-15, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.362, + 0.469, + 0.431 + ], + "angle": 0, + "content": "[30] Xun Jiang, Xing Xu, Jingran Zhang, Fumin Shen, Zuo Cao, and Heng Tao Shen. Semi-supervised video paragraph grounding with contrastive encoder. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 2456-2465. IEEE, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.436, + 0.469, + 0.491 + ], + "angle": 0, + "content": "[31] Arsha Nagrani, Shan Yang, Anurag Arnab, Aren Jansen, Cordelia Schmid, and Chen Sun. Attention bottlenecks for multimodal fusion. In Advances in Neural Information Processing Systems, pages 14200-14213, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.497, + 0.469, + 0.566 + ], + "angle": 0, + "content": "[32] Guanglei Yang, Enrico Fini, Dan Xu, Paolo Rota, Mingli Ding, Moin Nabi, Xavier Alameda-Pineda, and Elisa Ricci. Uncertainty-aware contrastive distillation for incremental semantic segmentation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(2):2567-2581, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.571, + 0.469, + 0.627 + ], + "angle": 0, + "content": "[33] Mengyuan Chen, Junyu Gao, Shicai Yang, and Changsheng Xu. Dual-evidential learning for weakly-supervised temporal action localization. In European Conference on Computer Vision, pages 192-208. Springer, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.631, + 0.469, + 0.699 + ], + "angle": 0, + "content": "[34] Yifei Huang, Lijin Yang, and Yoichi Sato. Weakly supervised temporal sentence grounding with uncertainty-guided self-training. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 18908-18918, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.705, + 0.469, + 0.774 + ], + "angle": 0, + "content": "[35] Shenshen Li, Xing Xu, Yang Yang, Fumin Shen, Yijun Mo, Yujie Li, and Heng Tao Shen. DCEL: deep cross-modal evidential learning for text-based person retrieval. In Proceedings of the 31st ACM International Conference on Multimedia, pages 6292-6300, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.78, + 0.469, + 0.821 + ], + "angle": 0, + "content": "[36] Alexander A Alemi, Ian Fischer, Joshua V Dillon, and Kevin Murphy. Deep variational information bottleneck. arXiv preprint arXiv:1612.00410, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.826, + 0.469, + 0.88 + ], + "angle": 0, + "content": "[37] Yuchen Shen, Dong Zhang, Zhihao Song, Xuesong Jiang, and Qiaolin Ye. Learning to reduce information bottleneck for object detection in aerial images. IEEE Geoscience and Remote Sensing Letters, 2023. 3" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.88 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.135 + ], + "angle": 0, + "content": "[38] Among Wu and Cheng Deng. Tib: Detecting unknown objects via two-stream information bottleneck. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.139, + 0.892, + 0.209 + ], + "angle": 0, + "content": "[39] Sungtae An, Nataraj Jammalamadaka, and Eunji Chong. Maximum entropy information bottleneck for uncertainty-aware stochastic embedding. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 3808-3817, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.214, + 0.892, + 0.241 + ], + "angle": 0, + "content": "[40] Diederik P Kingma and Max Welling. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114, 2013. 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.246, + 0.892, + 0.302 + ], + "angle": 0, + "content": "[41] Amir Zadeh, Rowan Zellers, Eli Pincus, and Louis-Philippe Morency. Mosi: multimodal corpus of sentiment intensity and subjectivity analysis in online opinion videos. arXiv preprint arXiv:1606.06259, 2016. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.307, + 0.892, + 0.389 + ], + "angle": 0, + "content": "[42] AmirAli Bagher Zadeh, Paul Pu Liang, Soujanya Poria, Erik Cambria, and Louis-Philippe Morency. Multimodal language analysis in the wild: Cmu-mosei dataset and interpretable dynamic fusion graph. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics, pages 2236–2246, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.395, + 0.892, + 0.449 + ], + "angle": 0, + "content": "[43] Teng Niu, Shiai Zhu, Lei Pang, and Abdulmotaleb El Saddik. Sentiment analysis on multi-view social data. In International Conference on Multimedia Modeling, pages 15-27, 2016. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.455, + 0.892, + 0.511 + ], + "angle": 0, + "content": "[44] Nathan Silberman, Derek Hoiem, Pushmeet Kohli, and Rob Fergus. Indoor segmentation and support inference from rgbd images. In European Conference on Computer Vision, pages 746-760, 2012. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.516, + 0.892, + 0.571 + ], + "angle": 0, + "content": "[45] Yuanzhi Wang, Zhen Cui, and Yong Li. Distribution-consistent modal recovering for incomplete multimodal learning. In Proceedings of the IEEE International Conference on Computer Vision, pages 22025-22034, 2023. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.576, + 0.892, + 0.631 + ], + "angle": 0, + "content": "[46] Zheng Lian, Lan Chen, Licai Sun, Bin Liu, and Jianhua Tao. Gcnet: graph completion network for incomplete multimodal learning in conversation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.636, + 0.892, + 0.691 + ], + "angle": 0, + "content": "[47] Tong Zhu, Leida Li, Jufeng Yang, Sicheng Zhao, Hantao Liu, and Jiansheng Qian. Multimodal sentiment analysis with image-text interaction network. IEEE Transactions on Multimedia, 2022. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.697, + 0.892, + 0.78 + ], + "angle": 0, + "content": "[48] Yiwei Wei, Shaozu Yuan, Ruosong Yang, Lei Shen, Zhangmeizhi Li, Longbiao Wang, and Meng Chen. Tackling modality heterogeneity with multi-view calibration network for multimodal sentiment detection. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics, pages 5240-5252, 2023. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.785, + 0.892, + 0.839 + ], + "angle": 0, + "content": "[49] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.845, + 0.892, + 0.885 + ], + "angle": 0, + "content": "[50] Laurens Van der Maaten and Geoffrey Hinton. Visualizing data using t-sne. Journal of machine learning research, 9(11), 2008. 8" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.885 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "26885" + } + ] +] \ No newline at end of file diff --git a/2024/Embracing Unimodal Aleatoric Uncertainty for Robust Multimodal Fusion/b864c4f0-5e7f-441f-a2a0-028202223b38_origin.pdf b/2024/Embracing Unimodal Aleatoric Uncertainty for Robust Multimodal Fusion/b864c4f0-5e7f-441f-a2a0-028202223b38_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..2f13e2203120b6c0661bb7c0b50ed7169495cc64 --- /dev/null +++ b/2024/Embracing Unimodal Aleatoric Uncertainty for Robust Multimodal Fusion/b864c4f0-5e7f-441f-a2a0-028202223b38_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:95e22bd408bb7f630f917d1aed93ccc38e0fb6021c1387ee454d6eeba2fde091 +size 6037405 diff --git a/2024/Embracing Unimodal Aleatoric Uncertainty for Robust Multimodal Fusion/full.md b/2024/Embracing Unimodal Aleatoric Uncertainty for Robust Multimodal Fusion/full.md new file mode 100644 index 0000000000000000000000000000000000000000..34578524ff33fdb08f586a6bdf2b60d7a5f9611e --- /dev/null +++ b/2024/Embracing Unimodal Aleatoric Uncertainty for Robust Multimodal Fusion/full.md @@ -0,0 +1,305 @@ +# Embracing Unimodal Aleatoric Uncertainty for Robust Multimodal Fusion + +Zixian Gao $^{1*}$ , Xun Jiang $^{1*}$ , Xing Xu $^{1,2\dagger}$ , Fumin Shen $^{1}$ , Yujie Li $^{3}$ , Heng Tao Shen $^{1,2}$ + +1Center for Future Media & School of Computer Science and Engineering, University of Electronic Science and Technology of China, China + +$^{2}$ College of Electronic and Information Engineering, Tongji University, China + +$^{3}$ Kyushu Institute of Technology, Japan + +# Abstract + +As a fundamental problem in multimodal learning, multimodal fusion aims to compensate for the inherent limitations of a single modality. One challenge of multimodal fusion is that the unimodal data in their unique embedding space mostly contains potential noise, which leads to corrupted cross-modal interactions. However, in this paper, we show that the potential noise in unimodal data could be well quantified and further employed to enhance more stable unimodal embeddings via contrastive learning. Specifically, we propose a novel generic and robust multimodal fusion strategy, termed Embracing Aleatoric Uncertainty (EAU), which is simple and can be applied to kinds of modalities. It consists of two key steps: (1) the Stable Unimodal Feature Augmentation (SUFA) that learns a stable unimodal representation by incorporating the aleatoric uncertainty into self-supervised contrastive learning. (2) Robust Multimodal Feature Integration (RMFI) leveraging an information-theoretic strategy to learn a robust compact joint representation. We evaluate our proposed EAU method on five multimodal datasets, where the video, RGB image, text, audio, and depth image are involved. Extensive experiments demonstrate the EAU method is more noise-resistant than existing multimodal fusion strategies and establishes new state-of-the-art on several benchmarks. + +# 1. Introduction + +By exploring the complementary information from different modalities, multimodal learning has achieved impressive success in kinds of artificial intelligence applications, such as multimodal image classification in Internet applications [1-5] and sentiment analysis in intelligent robotics [6-9]. It has been widely proved that fusing information from different modalities appropriately is helpful to gain + +![](images/e36592c270ca40b909a9bc8637326a72e660b6515055de806100dfa06f72618b.jpg) +(a) Uncertainties from fuzzy semantics in multimodal data + +![](images/5c9d5c6423030f4a3c756a370d0827832c115c907a91f48c646f6e863cca5ff2.jpg) +(b) Multivariate Gaussian distributional representations for fuzzy semantics +Figure 1. An illustration of the aleatoric uncertainty in multimodal datasets and the distributional representations: (a) As the semantics are ambiguous, the multimodal data are prone to introducing noisy data. (b) We adopt a multivariate Gaussian distribution to represent the fuzzy semantics in a noisy latent space. + +better performance. Dedicated to this, multimodal fusion has become an emerging challenge in multimodal learning, which integrates different modalities into a unified representation with powerful neural networks. + +Nevertheless, with the studies developing, researchers have recognized that the multimodal data in different modal forms may be unreliable due to the unique noise in their own modal space. Recently, several works [5, 10-13] show that previous multimodal fusion methods have overlooked the unreliable quality of multimodal data. Specifically, the widely adopted fusion strategies may fail on noisy multimodal data because the cross-modal interactions may obtain limited influence by the uncertainty in data. Here we illustrate typical noise in multimodal data in Fig. 1. Due to the semantics of "Positive" being fuzzy and the labels being judged by humans subjectively, the given image-text pairs are all recognized as positive in sentiment analysis, even though there is obvious noise in both image and text + +modalities. In general, we can say the introduced noise comes from the aleatoric uncertainty [14], which undermines the effectiveness of exploiting multimodal data and demonstrates multimodal learning is not a free lunch. + +Hence, with the emerging challenge of uncertainty in multimodal data, we raise the first fundamental question: Can we quantify the uncertainty in multimodal data? Inspired by probability Distributional Representation [15-18], we naturally adopt the Gaussian distributions to work out this problem. We assume that each instance can be represented as a multivariate normal distribution, where the variance can be regarded as the intrinsic aleatoric uncertainty. With the quantified uncertainty, we are allowed to take a closer look at the multimodal fusion and raise the second question: Is it appropriate to completely drop the intrinsic uncertainty? From the examples in Fig. 1(a), we can observe that even if image-text pairs reveal similar semantics, the aleatoric uncertainty is still unavoidable because of the domain shift, extra descriptions, or image quality, etc. To this end, as is illustrated in Fig. 1(b), we argue that a multivariate normal distribution that considers aleatoric uncertainty can be regarded as a fuzzy representation for semantics, where the semantic-related data are within similar distributions, even if they are in different modalities. Motivated by the two assumptions, we develop a novel multimodal fusion strategy in this paper, i.e., Embracing Aleatoric Uncertainty (EAU). + +Specifically, our EAU method consists of the following two processes: (1) The Stable Unimodal Feature Augmentation (SUFA) quantifies the intrinsic aleatoric uncertainty in each modality by representing samples as multivariate normal distributions. Then, we further conduct self-supervised contrastive learning with these distributional representations to learn stable unimodal embeddings with better semantical consistency. (2) Robust Multimodal Feature Integration (RMFI) that dynamically fuses the stable unimodal embeddings into a joint representation. Particularly, considering that SUFA focuses on semantic consistency only and ignores the problem of information redundancy, we employ an information-theoretic strategy, i.e., Variational Information Bottleneck [19, 20], to learn a compact joint representation with less redundancy. We evaluate our method on five multimodal benchmark datasets and demonstrate it outperforms existing state-of-the-art methods on both Multimodal Sentiment Analysis and Multimodal Image Classification tasks. Moreover, our proposed EAU method also reveals better robustness on noisy datasets compared with the counterpart methods [2, 3, 5]. + +Overall, our contributions in this paper can be summarized as three-fold: + +- We propose a novel multimodal fusion method termed Embracing Aleatoric Uncertainty (EAU), which quantifies the intrinsic aleatoric uncertainty and leverages it + +to learn stable and robust joint representations. + +- We devise the Stable Unimodal Feature Augmentation (SUFA) module, which quantifies aleatoric uncertainty and learns stable unimodal representations with better semantical consistency. +- We design the Robust Multimodal Feature Integration (RMFI) module to learn compact joint representations with less redundancy, which further improves the robustness of multimodal fusion in our method. + +# 2. Related Work + +Multimodal Fusion. Multimodal fusion, which aims at learning stronger representations from different modalities, has become an essential part of a spectrum of computer vision research, such as vision-audio learning [21,22], vision-language learning [23, 24], image retrieval [25-27], and video understanding [28-30]. Generally, multimodal fusion can be categorized into three types: early fusion, intermediate fusion, and late fusion. The previous multimodal fusion methods can be categorized into early fusion and late fusion in terms of their feature-level or decision-level fusion operations. In the last decades, massive works with deep learning suggest that intermediate fusion [13, 29, 31], which learns unified embeddings for multimodal data in the hidden layers of neural networks, could benefit representation learning. Although these multimodal fusion methods achieve remarkable improvements in kinds of multimodal learning tasks, most of them ignore the uncertainty in multimodal learning. Recently, several works [5, 10-13] empirically or theoretically demonstrated that conventional multimodal fusion methods had limited performance and robustness on noisy or corrupted multimodal data. Inspired by these pioneering works, we devised a novel intermediate multimodal fusion method, which could quantify the uncertainty and learn more robust joint representations. + +Uncertainty in Deep Learning. In general, the uncertainties in deep learning can be categorized into epistemic uncertainty and aleatoric uncertainty [14]. The former aims at capturing the noise of the parameters in deep neural networks, while the latter measures the noise inherent in given training data. To improve the robustness and generalization of open-world scenarios, many researchers incorporate the uncertainty estimation into the deep learning models for computer vision tasks, such as face recognition [15, 16], semantic segmentation [32], and action localization [33, 34]. There is also a list of works that incorporate uncertainty estimation into multimodal learning tasks [5, 17, 22, 35]. However, most of them only quantified uncertainties to learn probability distribution representations that are used for cross-modal interactions, but ignore the value of intrinsic uncertainties and hardly take a further analysis on the robustness of multimodal fusion. Following this observation, + +we develop this work to learn more robust compact joint representations for multimodal learning, which embraces the uncertainties via self-supervised contrastive learning. + +Information Bottleneck Theory. The Information Bottleneck Theory [19] was initially formulated in the context of signal processing, which is proposed to discover a more concise representation of a signal while retaining its utmost informative content. Alemi et al. [36] proposed the Variational Information Bottleneck (VIB) to bridge the gap between Information Bottleneck Theory and deep learning, which approximates the information bottleneck constraints and enables it in deep learning. Based on the VIB, massive researchers introduce the Information Bottleneck Theory in deep learning models to tackle kinds of computer vision tasks, such as object detection [37-39] or image classification [37,38]. Furthermore, as the VIB aims at learning compact minimal representations, it also attracts wide attention from the multimodal learning field [28, 31]. For example, Mai et al. [28] devised a multimodal information bottleneck to learn the minimal sufficient unimodal and multimodal representations. Inspired by their successes, we also devise an multimodal integration strategy based on VIB to reduce the redundancy in joint representations, which further improves the robustness and effectiveness. + +# 3. Proposed Method + +# 3.1. Preliminaries + +Uncertainty Estimation in Deep Learning Models. In terms of [14], the uncertainties in deep learning models can be categorized into aleatoric uncertainty and epistemic uncertainty. The former refers to the uncertainty inherent in the observations and can not be explained away with more data, while the latter only accounts for uncertainty in the model and can be explained away given enough data. Typically, given a deep learning model $f_{\theta}(\cdot)$ , it conducts the mapping $\mathcal{X} \rightarrow \mathcal{Y}$ , where $\mathcal{X}$ and $\mathcal{Y}$ represent input data and observed label space respectively. The aleatoric uncertainty in $\mathcal{X}$ will lead to a corrupted predicted results in $\mathcal{Y}$ . To estimate the uncertainties in deep learning models, a widely adopted strategy is sampling $N$ weights $\theta \sim p(\theta | x, y)$ , $x \in \mathcal{X}, y \in \mathcal{Y}$ via Dropout operations for different predictions for mean $\mu$ and variance $\sigma^2$ , and adopting a Gaussian posterior $p(y|x) \sim \mathcal{N}(\mu_{*}(x), \sigma_{*}^{2}(x))$ . Here we can get the predictive variance and extract the aleatoric uncertainty which represents the ambiguity in $x$ : + +$$ +\begin{array}{l} \mu_ {*} ^ {2} (x) = \frac {1}{N} \sum_ {i} \mu_ {i} ^ {2} (x), \\ \sigma_ {*} ^ {2} (x) = \frac {1}{N} \sum_ {i} \sigma_ {i} ^ {2} (x) + \frac {1}{N} \sum_ {i} \mu_ {i} ^ {2} (x) - \mu_ {*} ^ {2} (x) \tag {1} \\ = \mathbb {E} _ {i} \left[ \sigma_ {i} ^ {2} (x) \right] + \operatorname {V a r} _ {i} [ \mu_ {i} (x) ], \\ \end{array} +$$ + +where $\mathbb{E}_i\left[\sigma_i^2 (x)\right]$ represents the aleatoric uncertainty, while $\operatorname {Var}_i[\mu_i(x)]$ is the epistemic uncertainty, which attributes to the model $f_{\theta}(\cdot)$ rather than input data $\mathcal{X}$ . In our work, we pay more attention to the aleatoric uncertainty since our goal is to quantify the intrinsic noise in multimodal data and leverage it for learning better joint representations. + +Variational Information Bottleneck. The Variational Information Bottleneck (VIB) [20] is an information-theoretic strategy widely adopted in deep learning models, which aims to maintain maximal discriminative feature representations with minimal redundancy. Specifically, given the input variable $\mathbf{x}$ with noise or redundant information and the target variable $\mathbf{y}$ , the VIB is to learn the compressed latent variable $\mathbf{z}$ , and $\mathbf{z}$ is maximally discriminative about the target variable $\mathbf{y}$ . Moreover, since the input $\mathbf{x}$ is noisy and redundant, the VIB also requires $\mathbf{z}$ to be minimally discriminative about the original variable $\mathbf{x}$ . In our work, we leverage the VIB to learn compact joint representations, which overcome the redundancy brought by highly aligned distributional multimodal representations to improve the robustness of multimodal fusion. + +![](images/7235b48e125f760e7c4f14f3b2d7e2e6e2174341886e5a08e358ffdf17457e53.jpg) +Figure 2. An illustration of the Unimodal Distributional Representation process. For clarity, we visualize the multivariate Gaussian distributions with two variables only. + +# 3.2. Stable Unimodal Feature Augmentation + +As the data in different modalities contain unique noise in their own modal space, we first propose the Stable Unimodal Feature Augmentation (SUFA) module to quantify their intrinsic aleatoric uncertainty. Following the argument that the uncertainty is attributed to the fuzzy representation of semantics, we further leverage the aleatoric uncertainty via self-supervised contrastive learning. + +Unimodal Distributional Representation. Given the multimodal samples $\mathbf{x}_i^m$ , $m \in \mathcal{M}$ where $\mathcal{M}$ is the modality set including images, audio, text, etc., we learn the distributional representations to quantify the aleatoric uncertainty in each modality. According to the deviation listed in Eq. 1, we can observe that the aleatoric uncertainty can be predicted with deep learning models $f_{\theta}(\cdot)$ as the variance $\sigma_i^{m2}$ directly if the epistemic uncertainty is not considered. To this end, we first employ the corresponding feature extractor to learn the preliminary embeddings for each modality, then deploy two additional fully connected layers to learn a mean vector $\mu_i^{m2} \in \mathbb{R}^d$ and a variance vector $\sigma_i^{m2} \in \mathbb{R}^d$ . + +Furtherly, we define the representation $\mathbf{z}_i^m$ in latent space of each sample $\mathbf{x}_i^m$ as a multivariate Gaussian distribution with $d$ variable, which can be represented as: + +$$ +\begin{array}{l} p (\mathbf {z} _ {i} ^ {m} | \mathbf {x} _ {i} ^ {m}) \sim \mathcal {N} (\boldsymbol {\mu} _ {i} ^ {m}, \boldsymbol {\sigma} _ {i} ^ {m 2} \mathbf {I}), \\ \pmb {\mu} _ {i} ^ {m} = f _ {\theta_ {1} ^ {m}} \left(\mathbf {x} _ {i} ^ {m}\right), \pmb {\sigma} _ {i} ^ {m} = f _ {\theta_ {2} ^ {m}} \left(\mathbf {x} _ {i} ^ {m}\right), \\ \end{array} +$$ + +where $f_{\theta_1^m}(\cdot)$ and $f_{\theta_2^m}(\cdot)$ represent the two different fully connected layers for mean and variance respectively. $\mathbf{I}$ is the identity matrix. To maintain the consistency of semantics in different modalities, we further align the distributional representations of a multimodal sample with the Kullback-Leibler divergence: + +$$ +\begin{array}{l} L _ {\mathrm {U D R}} = \sum_ {m _ {1} \neq m _ {2}} ^ {m _ {1}, m _ {2} \in \mathcal {M}} K L \left(p \left(\mathbf {z} _ {i} ^ {m _ {1}} \mid \mathbf {x} _ {i} ^ {m _ {1}}\right) \| p \left(\mathbf {z} _ {i} ^ {m _ {2}} \mid \mathbf {x} _ {i} ^ {m _ {2}}\right)\right) \tag {3} \\ + \sum_ {m} ^ {m \in \mathcal {M}} K L \left(p \left(\mathbf {z} _ {i} ^ {m} \mid \mathbf {x} _ {i} ^ {m}\right) \| \mathcal {N} (0, \mathbf {I})\right). \\ \end{array} +$$ + +We illustrate the Unimodal Distributional Representation process in Fig. 2 with bi-modal input for clarity. It can be observed the $L_{\mathrm{UDR}}$ will push two multimodal distributional representations with similar semantics closer. In this way, the representation of each multimodal sample is not limited within a deterministic point embedding, but a consistent fuzzy representation on several multivariate Gaussian distributions. Particularly, the variance $\sigma_i^m$ reveals the aleatoric uncertainty in $m$ modality and the mean $\mu_i^m$ is the corresponding stable representation. + +![](images/0607f458b577d9e9813f032a20f946587a98b8f103e6fd7a5e4317dfc22208b4.jpg) +Figure 3. Illustration of the Uncertainty-based Representation Augmentation process. Here we show the distributional representations of image modality only for clarity. + +Uncertainty-based Representation Augmentation. With the quantified aleatoric uncertainty, we consider the second question we raised in Sec.1: should we drop the aleatoric uncertainty in multimodal data? Intuitively, the aleatoric uncertainty in multimodal data is unavoidable since the natural ambiguity of semantics. However, it also results in the diversity of unimodal data in different modalities. To this end, we leverage the aleatoric uncertainty to generate the + +unseen samples such that learned unimodal representations are insensitive to diverse unimodal input with similar semantics. Take the image modality as an example, we illustrate the Uncertainty-based Representation Augmentation process in Fig. 3. Specifically, given a unimodal distributional representation $p(\mathbf{z}_i^m | \mathbf{x}_i^m) \sim \mathcal{N}(\boldsymbol{\mu}_i^m, \boldsymbol{\sigma}_i^{m2}\mathbf{I})$ , we first random sample an anchor point $\tilde{\mathbf{z}}_i^m$ and an augmented point $\mathbf{z}_i^m$ from the multivariate Gaussian distribution as matching pairs. Moreover, we randomly sample a set of negative points from other distributional representations and devise a self-supervised contrastive learning mechanism as follows: + +$$ +L _ {\mathrm {U R A}} = - \log \frac {e ^ {\operatorname* {s i m} \left(\tilde {\mathbf {z}} _ {i} ^ {m} , \mathbf {z} _ {i} ^ {m}\right) / \tau}}{\sum_ {j \neq i} \left(e ^ {\operatorname* {s i m} \left(\tilde {\mathbf {z}} _ {i} ^ {m} , \mathbf {z} _ {i} ^ {m}\right) / \tau} + e ^ {\operatorname* {s i m} \left(\tilde {\mathbf {z}} _ {i} ^ {m} , \mathbf {z} _ {j} ^ {m}\right) / \tau}\right)}, \tag {4} +$$ + +where $\tau$ is a temperature factor and sim is cosine similarity calculation. Here we adopt the re-parameterization trick [40] to conduct the sampling operations, which can be formulated as follows: + +$$ +\mathbf {z} _ {i} ^ {m} = \boldsymbol {\mu} _ {i} ^ {m} + \epsilon \sigma_ {i} ^ {m}, \epsilon \sim \mathcal {N} (0, \mathbf {I}). \tag {5} +$$ + +# 3.3. Robust Multimodal Feature Integration + +With the SUFA, we obtain stable unimodal representations with consistent semantics. However, we only consider the consistency in each modality but ignore the redundancy caused by duplicated representations. To this end, we proposed the Robust Multimodal Feature Integration (RMFI) module in this section. The overview of the RMFI module is illustrated in Fig. 4. + +Dynamic Multimodal Integration. Inspired by dynamic multimodal fusion [5], we assume that different modalities have unequal contributions to the joint representations in terms of the observed label space. To this end, we first apply a Dynamic Multimodal Integration strategy based on an attention mechanism. Specifically, given the stable unimodal representations $\pmb{\mu}_i^m$ , we calculate the attentive weights in the distributional representations across modalities according to their quantified uncertainties $\sigma_i^m$ , and apply them for multimodal integration: + +$$ +\hat {\mathbf {x}} _ {i} = \sum_ {m \in \mathcal {M}} \alpha_ {i} ^ {m} \boldsymbol {\mu} _ {i} ^ {m}, \quad \alpha_ {i} ^ {m} = \frac {e ^ {\frac {1}{\overline {{\sigma}} _ {i} ^ {m}}}}{\sum_ {m \in \mathcal {M}} e ^ {\frac {1}{\overline {{\sigma}} _ {i} ^ {m}}}}, \tag {6} +$$ + +where $\hat{\mathbf{x}}_i, \overline{\sigma}_i^m$ represent the integrated joint representations and the average variance of the multivariate Gaussian distribution of $m$ modality. In this way, we preliminarily integrate stable unimodal representations into joint representations, where the contributions of different modalities are estimated dynamically. + +Joint Representation Compression. In the SUFA module, we fully consider the consistency of semantics in different modalities to avoid noise in unimodal data. However, + +![](images/a08fd6e1048c2c148313340cc28584caf926f1655f346014537314ac9584d642.jpg) +Figure 4. An illustration of the RMFI module. We provide Cross-Entropy loss and $\mathrm{L}_1$ loss for classification-based and regression-based downstream tasks respectively. + +with similar multivariate Gaussian distributions among different modalities, redundant duplicated information will be introduced into joint representations. Hence, we devise a joint representation compression with the Variational Information Bottleneck (VIB) [20]. Specifically, given the preliminary joint representation $\hat{\mathbf{x}}_i$ and the target observations $\mathbf{y}$ in label space, we learn the compressed joint representation $\hat{\mathbf{z}}$ in a latent space: + +$$ +p \left(\hat {\mathbf {z}} _ {i} \mid \hat {\mathbf {x}} _ {i}\right) \sim \mathcal {N} \left(\hat {\boldsymbol {\mu}} _ {i}, \hat {\boldsymbol {\sigma}} _ {i} ^ {2} \mathbf {I}\right), \quad \hat {\boldsymbol {\mu}} _ {i} = f _ {\hat {\theta} _ {1}} (\hat {\mathbf {x}} _ {i}), \quad \hat {\boldsymbol {\sigma}} _ {i} = f _ {\hat {\theta} _ {2}} (\hat {\mathbf {x}} _ {i}). \tag {7} +$$ + +where $f_{\hat{\theta}_1}$ and $f_{\hat{\theta}_2}$ are two fully connected layers. Similarly, we also adopt the re-parameterization tricks [40] for the final compressed joint representations: + +$$ +\hat {\mathbf {z}} _ {i} = \hat {\boldsymbol {\mu}} _ {i} + \epsilon \hat {\boldsymbol {\sigma}} _ {i}, \epsilon \sim \mathcal {N} (0, \mathbf {I}). \tag {8} +$$ + +As our proposed method can be applied to different downstream tasks, here we provide two training objectives for the joint representation compression. Specifically, we employ the cross-entropy for the classification-based task: + +$$ +L _ {\mathrm {J R C}} = - \mathbf {y} ^ {\top} \log \delta \left(f _ {\theta} \left(\hat {\mathbf {z}} _ {i}\right)\right) + \lambda K L \left(p \left(\hat {\mathbf {z}} _ {i} \mid \hat {\mathbf {x}} _ {i}\right) \| \mathcal {N} (0, \mathbf {I})\right), \tag {9} +$$ + +where $\delta (\cdot)$ represents softmax function and $f_{\theta}(\cdot)$ is the deep learning models for classification, $\lambda$ is a hyperparameter. As for the regression-based task, we employ the mean square error as the training objective: + +$$ +L _ {\mathrm {J R C}} = \left[ y - f _ {\theta} \left(\hat {\mathbf {z}} _ {i}\right) \right] ^ {2} + K L \left(p \left(\hat {\mathbf {z}} _ {i} \mid \hat {\mathbf {x}} _ {i}\right) \| \mathcal {N} (0, \mathbf {I})\right). \tag {10} +$$ + +# 4. Experiments + +# 4.1. Experimental Settings + +Datasets. We evaluate our method on five multimodal datasets, including two tri-modal datasets, i.e., CMU-MOSI [41] and CMU-MOSEI [42] and three bi-modal datasets, i.e., MVSA-Single [43], UPMC Food101 [1], and NYU Depth v2 [44] datasets. The CMU-MOSI [41] and CMU-MOSEI [42] datasets are widely employed video datasets for Multimodal Sentiment Analysis (MSA) task. The former encompasses 2199 short videos, each accompanied by a sentiment strength score ranging from -3 to 3, offering + +a nuanced measure of emotional intensity. The latter, designed specifically for emotion analysis tasks, is a larger dataset comprising 22,856 movie review clips. The bimodal dataset, i.e., MVSA-Single [43] is also used for MSA task but it only contains image-text pairs obtained from social media for sentiment classification. Following [2-5], we also employ the UPMC Food101 [1] and the NYU Depth v2 [44] datasets for Multimodal Image Classification (MIC) task. The UPMC Food101 dataset contain 101 categories, respectively, these datasets are specifically designed for bimodal classification tasks. The UPMC FOOD101 dataset contains images with 101 categories obtained by Google Image Search and their corresponding textual descriptions. The NYU Depth v2 dataset [44] is utilized for scene recognition and comprises depth and RGB images. Following [5], we also adopt the commonly used 9 out of the 27 scene categories and the remaining categories as "Others". + +Implementation Details. In alignment with alternative methodologies, for our trimodal video dataset, we employ FACET, COVAREP, and BERT as feature extractors for visual, audio, and text modalities, respectively. For the bimodal dataset, we utilize ResNet-152 as the feature extractor for RGB and depth images, coupled with BERT as the feature extractor for text. As for the hyperparameters, we set the temperature factor $\tau$ and the balance factor $\lambda$ to 0.5 and 1e-3 respectively. We employed the Adam optimizer with learning rate of 1e-5,and adopt the Reduce-on-Plateau learning rate adjustment strategy to train our EAU method. + +Evaluation Metrics. For the regression-based MSA task based on CMU-MOSI and CMU-MOSEI, we follow the previous work [6, 7, 9, 45] and adopt Acc7, F1 score, and Pearson correlation coefficient as our metrics. As for the classification-based MSA task on the MSVA-Single dataset and MIC task on the UPMC FOOD101 and The NYU Depth v2 datasets, we reported commonly used metrics, including accuracy and F1 score. + +# 4.2. Comparisons with State-Of-The-Arts + +To show the superiority of our proposed EAU method on multimodal fusion, we compared our EAU method with existing state-of-the-art methods of the MSA task on the CMU-MOSI and CMU-MOSEI datasets, including MIB [28], HMA [6], MIM [7], GCNet [46], ConFEDE [8], DiC-MoR [45], and DMD [9]. Moreover, we also make fair comparisons with the widely adopted simple multimodal fusion strategies i.e., Concat and Late Fusion, as well as recent well-designed multimodal fusion strategies on the bimodal datasets, including TMC [3], ITIN [47], MMBT [2], PMF [4], MVCN [48], and QMF [5]. Particularly, similar to the counterpart method QMF [5], we evaluate our method on noisy datasets to observe the model robustness. + +Multimodal Fusion on Classification Task. We report the performance of classification tasks, including both MSA + +
MethodCMU-MOSICMU-MOSEI
Acc7F1CorrAcc7F1Corr
MIB [28] (2022)48.685.30.79854.186.20.790
HMA [6] (2023)45.385.60.78252.885.40.787
MIM [7] (2023)47.085.90.80552.586.30.792
GCNet [46] (2023)44.985.1-51.585.2-
ConFEDE [8] (2023)42.385.50.78454.985.80.780
DiCMoR [45] (2023)45.385.6-53.485.1-
DMD [9] (2023)45.686.0-54.586.6-
EAU (Ours)48.886.20.80954.886.90.816
+ +
MethodMVSA-SingleFood 101NYU Depth v2
AccF1AccF1AccF1
Concat65.5965.4388.2088.1970.3069.82
Late Fusion76.8875.7290.6990.7769.1468.32
MMBT [2] (2020)78.50-91.5291.2871.04-
TMC [3] (2021)76.0674.5589.8689.8071.0669.83
ITIN [47] (2022)75.1974.97----
PMF [4] (2023)--91.68---
MVCN [48] (2023)76.0674.55----
QMF [5] (2023)78.0777.1892.9292.9370.0968.65
EAU (Ours)79.1578.3693.2093.1872.0570.63
+ +and MIC tasks on bi-modal datasets in Table 1, where the best results are marked in bold. From the experimental results, we can observe that our proposed EAU method outperforms all three datasets. Particularly, compared with the recent state-of-the-art method QMF method [5], our EAU method achieves at least $1\%$ absolute improvements on the MVSA-Single and NYU Depth v2 datasets. These results demonstrate that our EAU approach conducts better multimodal fusion on text, depth, and RGB images, affirming the effectiveness of leveraging aleatoric uncertainty in multimodal data. However, we also note that the improvements on UPMC Food 101 are not as remarkable as the other two benchmark datasets. We speculate that one probable reason is the images and texts in this dataset are clear compared with the others. Hence, our EAU method obtains fewer improvements since the augmentation is limited. + +Multimodal Fusion on Regression Task. We evaluate our method on the regression task on the CMU-MOSEI and CMU-MOSI datasets, where the MSA task is conducted by predicting sentiment strengths. From the experimental results in Table 1, we can observe that: For the CMU-MOSI dataset, our proposed EAU method achieves state-of-the-art performance across all evaluation metrics, with a particularly significant improvement noted in Acc7. On the CMU-MOSEI dataset, our method has also demonstrated substantial improvements on most evaluation metrics. These results demonstrate the superiority of our proposed EAU, which benefits from high-quality multimodal fusion via learning stable and robust joint representations. Simultaneously, these results indicate that our approach can effortlessly extend to other modalities and can be seamlessly transferred to frameworks accommodating kinds of modalities. + +Model Robustness on Noisy Multimodal Datasets. To validate the effectiveness of our model in handling data noise, we conducted more evaluation on the noisy datasets following a recent multimodal fusion strategy QMF [5]. Concretely, we consider different intensities of Gaussian and Salt-Pepper noise on the MVSA-Single and NYU Depth v2 datasets. For fair comparisons, we conducted 10 + +Table 1. Comparisons with state-of-the-art multimodal fusion methods on MSA and MIC tasks. The CMU-MOSI and CMU-MOSEI datasets contain videos, audi, and texts. The MVSA-Single and Food-101 datasets consist of texts and RGB images. The NYU Depth v2 contains RGB and depth images. Note that CMU-MOSI and CMU-MOSEI are used for regression-based MSA tasks, while the others are used for classification-based MSA or MIC tasks. + +
Noisy MVSA-Single
MethodCleanSalt-Pepper NoiseGaussian Noise
ε = 0ε = 5ε = 10ε = 5ε = 10
Concat65.5958.6951.1650.7046.12
Late Fusion76.8867.8855.4363.4655.16
MMBT [2] (2020)78.5074.0751.2671.9955.34
TMC [3] (2021)74.8768.0256.6266.7260.35
QMF [5] (2023)78.0773.9060.4173.8561.28
EAU (Ours)79.1574.8161.0473.8962.04
+ +
Noisy NYU Depth v2
MethodCleanSalt-Pepper NoiseGaussian Noise
ε = 0ε = 5ε = 10ε = 5ε = 10
Concat70.4457.9844.5159.9753.20
Late fusion69.1656.2741.2259.6351.99
MMTM [2] (2020)71.0459.4544.5960.3752.28
TMC [3] (2021)71.0659.3444.6561.0453.36
QMF [5] (2023)70.0958.5045.6961.6255.60
EAU (Ours)72.0559.8346.8563.3358.85
+ +Table 2. Comparisons with state-of-the-arts concerning model performance on noisy MVSA-Single and NYU Depth v2 datasets. + +experiments with different random seeds and reported the mean results as our final performance. According to the experimental results listed in Table 2, we can observe that our method has achieved state-of-the-art performance under various types and intensities of noise. Moreover, with the intensity of noise increasing, the margin of performance between our method and existing multimodal fusion strategies goes larger. Particularly, our method achieves more than $3\%$ improvement on the Noisy NYU Depth v2 dataset over the QMF method [5]. Such a result demonstrates our method reveals better robustness against data noise. The reason is our model pioneeringly leverages the intrinsic aleatoric uncertainty in the training data to enhance the stable feature representation, which is not fully considered in existing counterpart methods [2, 3, 5]. + +# 4.3. Further Analysis + +Analysis on Model Structure. To explore the impact of different model structures, we decomposed our proposed method into two components, SUFA and RMFI, and evalu + +
MethodCleanSalt-Pepper NoiseGaussian Noise
Acc@ε=0Acc@ε=5Acc@ε=10Acc@ε=5Acc@ε=10
Naive Backbone + Late Fusion76.88 ± 1.3067.88 ± 1.8755.43 ± 1.9463.46 ± 3.4655.16 ± 3.60
Naive Backbone + Concat74.53 ± 0.9762.24 ± 2.4753.14 ± 2.8956.08 ± 1.8949.21 ± 2.81
Naive Backbone + RMFI77.23 ± 0.7470.99 ± 2.7759.11 ± 1.9266.53 ± 2.2652.02 ± 3.45
SUFA + Late Fusion77.11 ± 1.0570.05 ± 1.5055.93 ± 2.2562.27 ± 4.2055.90 ± 3.67
SUFA + Concat78.18 ± 1.0873.10 ± 1.0760.67 ± 1.3871.40 ± 1.4859.11 ± 1.60
BERT (NAACL'19) [49]75.61 ± 0.5369.50 ± 1.5047.41 ± 0.7969.50 ± 1.5047.41 ± 0.79
MMBT (arXiv'19) [2]78.50 ± 0.4074.07 ± 1.1251.26 ± 5.6571.99 ± 1.0455.34 ± 2.84
TMC (ICLR'21) [3]74.87 ± 2.2468.02 ± 3.0756.62 ± 3.6766.72 ± 4.5560.35 ± 2.79
QMF (ICML'23) [5]78.07 ± 1.1073.90 ± 1.8960.41 ± 2.6373.85 ± 1.4261.28 ± 2.12
SUFA + RMFI (EAU)79.15 ± 0.6074.81 ± 1.5961.04 ± 1.3173.89 ± 0.9662.04 ± 1.26
+ +
ModalityAccF1Corr
T44.183.70.785
V16.745.30.072
A15.848.70.099
T+A46.784.20.794
T+V46.384.60.786
A+V20.255.30.117
T+A+V48.886.20.809
T76.3075.90-
V63.5863.35-
T+V79.1578.36-
+ +Table 3. Analysis concerning key model structures in our EAU method on the Noisy MSVA-Single (Left), and analysis concerning modalities on the CMU-MOSI (Right Upper) and MVSA-Single (Right Lower) datasets. Note that BERT [49] adopts text modality only. + +ated their effectiveness in combination with various model structures on the Noisy MVSA-Single dataset. Specifically, we employ two alternative designs for SUFA and RMFI respectively: (1) Naive Backbone: deploying the feature extractors directly, where the distributional representation and stable feature augmentation in the SUFA module are disabled. (2) Late Fusion: the unimodal representations are used for downstream tasks first and the fusion operations are conducted in probability space. To observe the stability of our method, we conduct 10 experiments with different random seeds and report the mean and variance in Table 3. Here we also make fair comparisons with several counterpart methods [2, 3, 5, 49] to furtherly show our superiority. + +According to the experimental results presented in Table 3, we can observe that: (1) Our proposed SUFA and RMFI modules consistently improve performance on all metrics. Particularly, with the SUFA module, the results of fusion strategy Concat are boosted with more than $10\%$ absolute performance on the noisy data. Moreover, compared with Late Fusion, our proposed RMFI module can also significantly boost the classification accuracy on noisy data with the SUFA module. These results prove again that our method shows great superiority in the effectiveness and robustness of multimodal fusion. (2) It also explicates that our proposed EAU method remarkably outperforms compared with the counterpart methods [2, 3, 5]. Particularly, compared with the recent state-of-the-art method QMF [5], our complete model reveals better performance and stability, demonstrating the capability of our EAU method again. + +Analysis on the Effectiveness of Multimodal Fusion. To verify the effectiveness of multimodal fusion in our EAU method, we also conduct ablation studies concerning different modalities on the CMU-MOSI datasets, which consist of video (V), text (T), and audio (A) three modalities. By observing the experimental results illustrated in Table 3, we have drawn the following conclusions: (1) Our fusion strategy demonstrated significant effectiveness across different modal combinations. As the number of modalities increased, our approach facilitated the integration of multimodal information, resulting in a significant improvement across all evaluation metrics. (2) We also note that the text + +modality performs an essential role in the MSA task on the CMU-MOSI dataset. However, combined with the other two modalities via our proposed EAU method, the results are also boosted significantly, demonstrating the superiority of our EAU method again. Such a phenomenon also proves the rationality of our Dynamic Multimodal Integration process in the RMFI module. + +![](images/8668a04804c4b09a944523bf3b075fa005aa77bbd40a05d1a1e70698757000bc.jpg) +Figure 5. Analysis for robustness evaluated by over 10 times random experiments under different noisy data. + +![](images/71ff691682e7c3f72c18382f828a43e21e7dd09a6ae48544210d4add733b0410.jpg) + +Analysis on Model Robustness. To verify the robustness and stability of our proposed EAU method, we conduct more ablation studies on the MVSA-Single and NYU Depth v2 datasets under different noise levels and observe the sparsity of final accuracy. Specifically, we conduct more than 10 experiments randomly with complete EAU (denoted as Complete) and EAU w/o SUFA (denoted as w/o SUFA), and show the statistics in Fig. 5. By observing the experimental results, we can find that: (1) The SUFA module consistently shows its effectiveness on both two datasets under different noise strengths. Particularly, when introducing more noise to the data, the ablated model shows strong fluctuations. Contrastively, the complete EAU model reveals remarkably reduced fluctuations, which demonstrates the superiority of model robustness and stability. (2) We also note that the improvements on the NYU Depth v2 are slightly lower than the MVSA-Single dataset, though both two datasets have significant results. We speculate this is caused by the difference between text modality and depth images. Concretely, the depth images can be regarded as another view of spatial information, which has less complementary information for RGB images compared with texts. + +To this end, the introduced noise will show more impact on the depth-RGB than the image-text compositions. + +![](images/726e724da2f8f1cecd8658f2394ba73f8127abb5c4412c3104b23edd9751a0ac.jpg) +Figure 6. Analysis concerning the training convergence and performance fluctuations during training. + +Analysis on Training Process. We also illustrate the training process of our proposed EAU method on the MVSA-Single in Fig 6 to observe the training convergence and performance fluctuations. According to the illustrated experimental results, we can observe that the complete EAU model exhibits a smoother training process and converges more rapidly with a remarkably better performance. Particularly, compared with the ablated model w/o RMFI, the other two models show much better stability during training. This is because the fused features learned by the ablated model w/o RMFI have redundant information caused by cross-modal distributional alignment in the SUFA module. It proves again the rationality of our proposed RMFI module, which deploys an information-theoretic strategy to address the limitation. + +![](images/a0a44e1bc2b030f65ccba3a62e40971d0ae6ad4b7430367fd66c5b355106b151.jpg) +(a) Concat + +![](images/b006de730ac2a0fa72ffc5821911aa703b6e4f3a889149c36ef3c7fdec7b7191.jpg) + +![](images/4025742e48aae512ebe097fe777828d2f04e9687ae3b56ed36661b85da848df5.jpg) +(c) EAU w/o RMFI + +![](images/68fa03c3a5c26ed034bbc01e245e8d35c8dc77ad8493aae4bf3f6b40f7e00693.jpg) +(b) EAU w/o SUFA +(d) EAU +Figure 7. Visualizations of joint representations by t-SNE [50] on MVSA-Single dataset. Here we adopt Concat for the ablated model w/o RMFI. + +Visualizations of Joint Representations. To further verify the superiority of our proposed EAU method, we also employ the t-SNE [50] to visualize the learned joint representations. As illustrated in Fig. 7, it can be observed that the features generated by our method exhibit a more + +compact and discriminative distribution. In contrast, when the SUFA or RMFI module is removed, more sparse points appear around the feature distribution which leads to corrupted classification results at last. Such results indicate that our method is more capable of learning a representative joint representation. + +![](images/9ffa48bd060d471068ad61f4ee8c613e060afdd37867e1a17166aeaa5ee2981b.jpg) +Figure 8. Visualizations of test cases selected from the MSVA-Single datasets. It can be observed that our EAU method reveals better robustness against the noise. + +Qualitative Analysis. Moreover, we also illustrate several typical test cases in Fig. 8 and make comparisons with the recent counterpart method QMF [5]. Under different noise settings, our proposed EAU method consistently classifies the sentiment of the given image-text input precisely, while the QMF method returns incorrect results. It explicates that our proposed EAU method benefits from the well-designed SUFA and RMFI module, which is effective in learning more stable and robust joint representations for noise-resistant performance with higher accuracy. + +# 5. Conclusion + +In this paper, we proposed a novel multimodal fusion method, namely, Embracing Aleatoric Uncertainty (EAU). It achieves more discriminative joint representations by fully considering the aleatoric uncertainty in multimodal data. Specifically, with the well-designed Stable Unimodal Feature Augmentation (SUFA) and Robust Multimodal Feature Integration (RMFI) modules, our proposed EAU could learn compact and robust joint representations. We evaluated our proposed EAU method on five multimodal benchmark datasets for both classification and regression tasks and demonstrated its superiority in fusion performance and robustness. For future work, we will further explore the uncertainties in multimodal learning. + +# 6. Acknowledgement + +This work was sponsored in part by the National Natural Science Foundation of China under Grants (No. 62222203 and No. 62072080) and the New Cornerstone Science Foundation through the XPLORER PRIZE. + +# References + +[1] Lukas Bossard, Matthieu Guillaumin, and Luc Van Gool. Food-101-mining discriminative components with random forests. In European Conference on Computer Vision, pages 446-461, 2014. 1, 5 +[2] Douwe Kiela, Suvrat Bhooshan, Hamed Firooz, Ethan Perez, and Davide Testuggine. Supervised multimodal bittransformers for classifying images and text. arXiv preprint arXiv:1909.02950, 2019. 1, 2, 5, 6, 7 +[3] Zongbo Han, Changqing Zhang, Huazhu Fu, and Joey Tianyi Zhou. Trusted multi-view classification. arXiv preprint arXiv:2102.02051, 2021. 1, 2, 5, 6, 7 +[4] Yaowei Li, Ruijie Quan, Linchao Zhu, and Yi Yang. Efficient multimodal fusion via interactive prompting. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 2604-2613, 2023. 1, 5, 6 +[5] Qingyang Zhang, Haitao Wu, Changqing Zhang, Qinghua Hu, Huazhu Fu, Joey Tianyi Zhou, and Xi Peng. Provable dynamic fusion for low-quality multimodal data. arXiv preprint arXiv:2306.02050, 2023. 1, 2, 4, 5, 6, 7, 8 +[6] Ronghao Lin and Haifeng Hu. Dynamically shifting multimodal representations via hybrid-modal attention for multimodal sentiment analysis. IEEE Transactions on Multimedia, 2023. 1, 5, 6 +[7] Ying Zeng, Sijie Mai, Wenjun Yan, and Haifeng Hu. Multimodal reaction: Information modulation for cross-modal representation learning. IEEE Transactions on Multimedia, 2023. 1, 5, 6 +[8] Jiuding Yang, Yakun Yu, Di Niu, Weidong Guo, and Yu Xu. Confede: Contrastive feature decomposition for multimodal sentiment analysis. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics, pages 7617-7630, 2023. 1, 5, 6 +[9] Yong Li, Yuanzhi Wang, and Zhen Cui. Decoupled multimodal distilling for emotion recognition. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 6631-6640, 2023. 1, 5, 6 +[10] Weiyao Wang, Du Tran, and Matt Feiszli. What makes training multi-modal classification networks hard? In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 12695-12705, 2020. 1, 2 +[11] Xiaokang Peng, Yake Wei, Andong Deng, Dong Wang, and Di Hu. Balanced multimodal learning via on-the-fly gradient modulation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 8238-8247, 2022. 1, 2 +[12] Yu Huang, Junyang Lin, Chang Zhou, Hongxia Yang, and Longbo Huang. Modality competition: What makes joint training of multi-modal network fail in deep learning? (provably). In International Conference on Machine Learning, pages 9226-9259. PMLR, 2022. 1, 2 +[13] Yunfeng Fan, Wenchao Xu, Haozhao Wang, Junxiao Wang, and Song Guo. PMR: prototypical modal rebalance for multimodal learning. In Proceedings of the IEEE Conference + +on Computer Vision and Pattern Recognition, pages 20029-20038. IEEE, 2023. 1, 2 +[14] Alex Kendall and Yarin Gal. What uncertainties do we need in bayesian deep learning for computer vision? Advances in neural information processing systems, 30, 2017. 2, 3 +[15] Jie Chang, Zhonghao Lan, Changmao Cheng, and Yichen Wei. Data uncertainty learning in face recognition. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 5710-5719, 2020. 2 +[16] Yichun Shi and Anil K Jain. Probabilistic face embeddings. In Proceedings of the IEEE International Conference on Computer Vision, pages 6902-6911, 2019. 2 +[17] Yatai Ji, Junjie Wang, Yuan Gong, Lin Zhang, Yanru Zhu, Hongfa Wang, Jiaxing Zhang, Tetsuya Sakai, and Yujiu Yang. Map: Multimodal uncertainty-aware vision-language pre-training model. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 23262-23271, 2023. 2 +[18] Luke Vilnis and Andrew McCallum. Word representations via gaussian embedding. arXiv preprint arXiv:1412.6623, 2014. 2 +[19] Naftali Tishby, Fernando C Pereira, and William Bialek. The information bottleneck method. arXiv preprint physics/0004057, 2000. 2, 3 +[20] Alexander A Alemi, Ian Fischer, Joshua V Dillon, and Kevin Murphy. Deep variational information bottleneck. arXiv preprint arXiv:1612.00410, 2016. 2, 3, 5 +[21] Xun Jiang, Xing Xu, Zhiguo Chen, Jingran Zhang, Jingkuan Song, Fumin Shen, Huimin Lu, and Heng Tao Shen. DHHN: dual hierarchical hybrid network for weakly-supervised audio-visual video parsing. In ACM International Conference on Multimedia, pages 719–727, 2022. 2 +[22] Junyu Gao, Mengyuan Chen, and Changsheng Xu. Collecting cross-modal presence-absence evidence for weakly-supervised audio-visual event perception. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 18827–18836, 2023. 2 +[23] Zhiguo Chen, Xun Jiang, Xing Xu, Zuo Cao, Yijun Mo, and Heng Tao Shen. Joint searching and grounding: Multigranularity video content retrieval. In Proceedings of the 31st ACM International Conference on Multimedia, pages 975-983, 2023. 2 +[24] Shenshen Li, Xing Xu, Xun Jiang, Fumin Shen, Xin Liu, and Heng Tao Shen. Multi-grained attention network with mutual exclusion for composed query-based image retrieval. IEEE Transactions on Circuits and Systems for Video Technology, 2023. 2 +[25] Shenshen Li, Xing Xu, Xun Jiang, Fumin Shen, Zhe Sun, and Andrzej Cichocki. Cross-modal attention preservation with self-contrastive learning for composed query-based image retrieval. ACM Transactions on Multimedia Computing, Communications, and Applications, 20(6), 2024. 2 + +[26] Xing Xu, Kaiyi Lin, Yang Yang, Alan Hanjalic, and Heng Tao Shen. Joint feature synthesis and embedding: Adversarial cross-modal retrieval revisited. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(6):3030-3047, 2022. 2 +[27] Shenshen Li, Xing Xu, Xun Jiang, Fumin Shen, Xin Liu, and Heng Tao Shen. Multi-grained attention network with mutual exclusion for composed query-based image retrieval. IEEE Transactions on Circuits and Systems for Video Technology, pages 1-1, 2023. 2 +[28] Sijie Mai, Ying Zeng, and Haifeng Hu. Multimodal information bottleneck: Learning minimal sufficient unimodal and multimodal representations. IEEE Transactions on Multimedia, 2022. 2, 3, 5, 6 +[29] Xun Jiang, Xing Xu, Jingran Zhang, Fumin Shen, Zuo Cao, and Heng Tao Shen. Sdn: Semantic decoupling network for temporal language grounding. IEEE Transactions on Neural Networks and Learning Systems, pages 1-15, 2022. 2 +[30] Xun Jiang, Xing Xu, Jingran Zhang, Fumin Shen, Zuo Cao, and Heng Tao Shen. Semi-supervised video paragraph grounding with contrastive encoder. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 2456-2465. IEEE, 2022. 2 +[31] Arsha Nagrani, Shan Yang, Anurag Arnab, Aren Jansen, Cordelia Schmid, and Chen Sun. Attention bottlenecks for multimodal fusion. In Advances in Neural Information Processing Systems, pages 14200-14213, 2021. 2, 3 +[32] Guanglei Yang, Enrico Fini, Dan Xu, Paolo Rota, Mingli Ding, Moin Nabi, Xavier Alameda-Pineda, and Elisa Ricci. Uncertainty-aware contrastive distillation for incremental semantic segmentation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(2):2567-2581, 2022. 2 +[33] Mengyuan Chen, Junyu Gao, Shicai Yang, and Changsheng Xu. Dual-evidential learning for weakly-supervised temporal action localization. In European Conference on Computer Vision, pages 192-208. Springer, 2022. 2 +[34] Yifei Huang, Lijin Yang, and Yoichi Sato. Weakly supervised temporal sentence grounding with uncertainty-guided self-training. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 18908-18918, 2023. 2 +[35] Shenshen Li, Xing Xu, Yang Yang, Fumin Shen, Yijun Mo, Yujie Li, and Heng Tao Shen. DCEL: deep cross-modal evidential learning for text-based person retrieval. In Proceedings of the 31st ACM International Conference on Multimedia, pages 6292-6300, 2023. 2 +[36] Alexander A Alemi, Ian Fischer, Joshua V Dillon, and Kevin Murphy. Deep variational information bottleneck. arXiv preprint arXiv:1612.00410, 2016. 3 +[37] Yuchen Shen, Dong Zhang, Zhihao Song, Xuesong Jiang, and Qiaolin Ye. Learning to reduce information bottleneck for object detection in aerial images. IEEE Geoscience and Remote Sensing Letters, 2023. 3 + +[38] Among Wu and Cheng Deng. Tib: Detecting unknown objects via two-stream information bottleneck. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023. 3 +[39] Sungtae An, Nataraj Jammalamadaka, and Eunji Chong. Maximum entropy information bottleneck for uncertainty-aware stochastic embedding. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 3808-3817, 2023. 3 +[40] Diederik P Kingma and Max Welling. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114, 2013. 4, 5 +[41] Amir Zadeh, Rowan Zellers, Eli Pincus, and Louis-Philippe Morency. Mosi: multimodal corpus of sentiment intensity and subjectivity analysis in online opinion videos. arXiv preprint arXiv:1606.06259, 2016. 5 +[42] AmirAli Bagher Zadeh, Paul Pu Liang, Soujanya Poria, Erik Cambria, and Louis-Philippe Morency. Multimodal language analysis in the wild: Cmu-mosei dataset and interpretable dynamic fusion graph. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics, pages 2236–2246, 2018. 5 +[43] Teng Niu, Shiai Zhu, Lei Pang, and Abdulmotaleb El Saddik. Sentiment analysis on multi-view social data. In International Conference on Multimedia Modeling, pages 15-27, 2016. 5 +[44] Nathan Silberman, Derek Hoiem, Pushmeet Kohli, and Rob Fergus. Indoor segmentation and support inference from rgbd images. In European Conference on Computer Vision, pages 746-760, 2012. 5 +[45] Yuanzhi Wang, Zhen Cui, and Yong Li. Distribution-consistent modal recovering for incomplete multimodal learning. In Proceedings of the IEEE International Conference on Computer Vision, pages 22025-22034, 2023. 5, 6 +[46] Zheng Lian, Lan Chen, Licai Sun, Bin Liu, and Jianhua Tao. Gcnet: graph completion network for incomplete multimodal learning in conversation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023. 5, 6 +[47] Tong Zhu, Leida Li, Jufeng Yang, Sicheng Zhao, Hantao Liu, and Jiansheng Qian. Multimodal sentiment analysis with image-text interaction network. IEEE Transactions on Multimedia, 2022. 5, 6 +[48] Yiwei Wei, Shaozu Yuan, Ruosong Yang, Lei Shen, Zhangmeizhi Li, Longbiao Wang, and Meng Chen. Tackling modality heterogeneity with multi-view calibration network for multimodal sentiment detection. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics, pages 5240-5252, 2023. 5, 6 +[49] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018. 7 +[50] Laurens Van der Maaten and Geoffrey Hinton. Visualizing data using t-sne. Journal of machine learning research, 9(11), 2008. 8 \ No newline at end of file diff --git a/2024/Embracing Unimodal Aleatoric Uncertainty for Robust Multimodal Fusion/images.zip b/2024/Embracing Unimodal Aleatoric Uncertainty for Robust Multimodal Fusion/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..6c741e2717f6322f75878b2829ff2b487713c6c1 --- /dev/null +++ b/2024/Embracing Unimodal Aleatoric Uncertainty for Robust Multimodal Fusion/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:25d9223edef387135d16cb4160a494370b7eadd6de6a74431610ba2d5d0d3beb +size 660773 diff --git a/2024/Embracing Unimodal Aleatoric Uncertainty for Robust Multimodal Fusion/layout.json b/2024/Embracing Unimodal Aleatoric Uncertainty for Robust Multimodal Fusion/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..472f30012b47b16e51a0806bcfa272e0eca29e40 --- /dev/null +++ b/2024/Embracing Unimodal Aleatoric Uncertainty for Robust Multimodal Fusion/layout.json @@ -0,0 +1,8044 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 66, + 103, + 529, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 103, + 529, + 121 + ], + "spans": [ + { + "bbox": [ + 66, + 103, + 529, + 121 + ], + "type": "text", + "content": "Embracing Unimodal Aleatoric Uncertainty for Robust Multimodal Fusion" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 101, + 142, + 492, + 158 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 101, + 142, + 492, + 158 + ], + "spans": [ + { + "bbox": [ + 101, + 142, + 492, + 158 + ], + "type": "text", + "content": "Zixian Gao" + }, + { + "bbox": [ + 101, + 142, + 492, + 158 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 101, + 142, + 492, + 158 + ], + "type": "text", + "content": ", Xun Jiang" + }, + { + "bbox": [ + 101, + 142, + 492, + 158 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 101, + 142, + 492, + 158 + ], + "type": "text", + "content": ", Xing Xu" + }, + { + "bbox": [ + 101, + 142, + 492, + 158 + ], + "type": "inline_equation", + "content": "^{1,2\\dagger}" + }, + { + "bbox": [ + 101, + 142, + 492, + 158 + ], + "type": "text", + "content": ", Fumin Shen" + }, + { + "bbox": [ + 101, + 142, + 492, + 158 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 101, + 142, + 492, + 158 + ], + "type": "text", + "content": ", Yujie Li" + }, + { + "bbox": [ + 101, + 142, + 492, + 158 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 101, + 142, + 492, + 158 + ], + "type": "text", + "content": ", Heng Tao Shen" + }, + { + "bbox": [ + 101, + 142, + 492, + 158 + ], + "type": "inline_equation", + "content": "^{1,2}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 114, + 158, + 475, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 114, + 158, + 475, + 185 + ], + "spans": [ + { + "bbox": [ + 114, + 158, + 475, + 185 + ], + "type": "text", + "content": "1Center for Future Media & School of Computer Science and Engineering, University of Electronic Science and Technology of China, China" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 109, + 185, + 484, + 199 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 109, + 185, + 484, + 199 + ], + "spans": [ + { + "bbox": [ + 109, + 185, + 484, + 199 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 109, + 185, + 484, + 199 + ], + "type": "text", + "content": "College of Electronic and Information Engineering, Tongji University, China" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 200, + 199, + 391, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 200, + 199, + 391, + 213 + ], + "spans": [ + { + "bbox": [ + 200, + 199, + 391, + 213 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 200, + 199, + 391, + 213 + ], + "type": "text", + "content": "Kyushu Institute of Technology, Japan" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 143, + 241, + 191, + 253 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 241, + 191, + 253 + ], + "spans": [ + { + "bbox": [ + 143, + 241, + 191, + 253 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 266, + 290, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 266, + 290, + 555 + ], + "spans": [ + { + "bbox": [ + 46, + 266, + 290, + 555 + ], + "type": "text", + "content": "As a fundamental problem in multimodal learning, multimodal fusion aims to compensate for the inherent limitations of a single modality. One challenge of multimodal fusion is that the unimodal data in their unique embedding space mostly contains potential noise, which leads to corrupted cross-modal interactions. However, in this paper, we show that the potential noise in unimodal data could be well quantified and further employed to enhance more stable unimodal embeddings via contrastive learning. Specifically, we propose a novel generic and robust multimodal fusion strategy, termed Embracing Aleatoric Uncertainty (EAU), which is simple and can be applied to kinds of modalities. It consists of two key steps: (1) the Stable Unimodal Feature Augmentation (SUFA) that learns a stable unimodal representation by incorporating the aleatoric uncertainty into self-supervised contrastive learning. (2) Robust Multimodal Feature Integration (RMFI) leveraging an information-theoretic strategy to learn a robust compact joint representation. We evaluate our proposed EAU method on five multimodal datasets, where the video, RGB image, text, audio, and depth image are involved. Extensive experiments demonstrate the EAU method is more noise-resistant than existing multimodal fusion strategies and establishes new state-of-the-art on several benchmarks." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 578, + 128, + 590 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 578, + 128, + 590 + ], + "spans": [ + { + "bbox": [ + 47, + 578, + 128, + 590 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 599, + 287, + 684 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 599, + 287, + 684 + ], + "spans": [ + { + "bbox": [ + 46, + 599, + 287, + 684 + ], + "type": "text", + "content": "By exploring the complementary information from different modalities, multimodal learning has achieved impressive success in kinds of artificial intelligence applications, such as multimodal image classification in Internet applications [1-5] and sentiment analysis in intelligent robotics [6-9]. It has been widely proved that fusing information from different modalities appropriately is helpful to gain" + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 309, + 240, + 545, + 313 + ], + "blocks": [ + { + "bbox": [ + 309, + 240, + 545, + 313 + ], + "lines": [ + { + "bbox": [ + 309, + 240, + 545, + 313 + ], + "spans": [ + { + "bbox": [ + 309, + 240, + 545, + 313 + ], + "type": "image", + "image_path": "e36592c270ca40b909a9bc8637326a72e660b6515055de806100dfa06f72618b.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 353, + 316, + 501, + 324 + ], + "lines": [ + { + "bbox": [ + 353, + 316, + 501, + 324 + ], + "spans": [ + { + "bbox": [ + 353, + 316, + 501, + 324 + ], + "type": "text", + "content": "(a) Uncertainties from fuzzy semantics in multimodal data" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 318, + 327, + 543, + 408 + ], + "blocks": [ + { + "bbox": [ + 318, + 327, + 543, + 408 + ], + "lines": [ + { + "bbox": [ + 318, + 327, + 543, + 408 + ], + "spans": [ + { + "bbox": [ + 318, + 327, + 543, + 408 + ], + "type": "image", + "image_path": "5c9d5c6423030f4a3c756a370d0827832c115c907a91f48c646f6e863cca5ff2.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 327, + 411, + 520, + 418 + ], + "lines": [ + { + "bbox": [ + 327, + 411, + 520, + 418 + ], + "spans": [ + { + "bbox": [ + 327, + 411, + 520, + 418 + ], + "type": "text", + "content": "(b) Multivariate Gaussian distributional representations for fuzzy semantics" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 305, + 428, + 547, + 484 + ], + "lines": [ + { + "bbox": [ + 305, + 428, + 547, + 484 + ], + "spans": [ + { + "bbox": [ + 305, + 428, + 547, + 484 + ], + "type": "text", + "content": "Figure 1. An illustration of the aleatoric uncertainty in multimodal datasets and the distributional representations: (a) As the semantics are ambiguous, the multimodal data are prone to introducing noisy data. (b) We adopt a multivariate Gaussian distribution to represent the fuzzy semantics in a noisy latent space." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 496, + 546, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 496, + 546, + 544 + ], + "spans": [ + { + "bbox": [ + 304, + 496, + 546, + 544 + ], + "type": "text", + "content": "better performance. Dedicated to this, multimodal fusion has become an emerging challenge in multimodal learning, which integrates different modalities into a unified representation with powerful neural networks." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 546, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 546, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 546, + 547, + 713 + ], + "type": "text", + "content": "Nevertheless, with the studies developing, researchers have recognized that the multimodal data in different modal forms may be unreliable due to the unique noise in their own modal space. Recently, several works [5, 10-13] show that previous multimodal fusion methods have overlooked the unreliable quality of multimodal data. Specifically, the widely adopted fusion strategies may fail on noisy multimodal data because the cross-modal interactions may obtain limited influence by the uncertainty in data. Here we illustrate typical noise in multimodal data in Fig. 1. Due to the semantics of \"Positive\" being fuzzy and the labels being judged by humans subjectively, the given image-text pairs are all recognized as positive in sentiment analysis, even though there is obvious noise in both image and text" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 693, + 215, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 693, + 215, + 703 + ], + "spans": [ + { + "bbox": [ + 58, + 693, + 215, + 703 + ], + "type": "text", + "content": "*These authors contributed equally to this work." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 59, + 703, + 135, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 703, + 135, + 712 + ], + "spans": [ + { + "bbox": [ + 59, + 703, + 135, + 712 + ], + "type": "text", + "content": "† Corresponding author." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "26876" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 119 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 119 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 119 + ], + "type": "text", + "content": "modalities. In general, we can say the introduced noise comes from the aleatoric uncertainty [14], which undermines the effectiveness of exploiting multimodal data and demonstrates multimodal learning is not a free lunch." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 121, + 289, + 393 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 121, + 289, + 393 + ], + "spans": [ + { + "bbox": [ + 46, + 121, + 289, + 393 + ], + "type": "text", + "content": "Hence, with the emerging challenge of uncertainty in multimodal data, we raise the first fundamental question: Can we quantify the uncertainty in multimodal data? Inspired by probability Distributional Representation [15-18], we naturally adopt the Gaussian distributions to work out this problem. We assume that each instance can be represented as a multivariate normal distribution, where the variance can be regarded as the intrinsic aleatoric uncertainty. With the quantified uncertainty, we are allowed to take a closer look at the multimodal fusion and raise the second question: Is it appropriate to completely drop the intrinsic uncertainty? From the examples in Fig. 1(a), we can observe that even if image-text pairs reveal similar semantics, the aleatoric uncertainty is still unavoidable because of the domain shift, extra descriptions, or image quality, etc. To this end, as is illustrated in Fig. 1(b), we argue that a multivariate normal distribution that considers aleatoric uncertainty can be regarded as a fuzzy representation for semantics, where the semantic-related data are within similar distributions, even if they are in different modalities. Motivated by the two assumptions, we develop a novel multimodal fusion strategy in this paper, i.e., Embracing Aleatoric Uncertainty (EAU)." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 395, + 289, + 646 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 395, + 289, + 646 + ], + "spans": [ + { + "bbox": [ + 46, + 395, + 289, + 646 + ], + "type": "text", + "content": "Specifically, our EAU method consists of the following two processes: (1) The Stable Unimodal Feature Augmentation (SUFA) quantifies the intrinsic aleatoric uncertainty in each modality by representing samples as multivariate normal distributions. Then, we further conduct self-supervised contrastive learning with these distributional representations to learn stable unimodal embeddings with better semantical consistency. (2) Robust Multimodal Feature Integration (RMFI) that dynamically fuses the stable unimodal embeddings into a joint representation. Particularly, considering that SUFA focuses on semantic consistency only and ignores the problem of information redundancy, we employ an information-theoretic strategy, i.e., Variational Information Bottleneck [19, 20], to learn a compact joint representation with less redundancy. We evaluate our method on five multimodal benchmark datasets and demonstrate it outperforms existing state-of-the-art methods on both Multimodal Sentiment Analysis and Multimodal Image Classification tasks. Moreover, our proposed EAU method also reveals better robustness on noisy datasets compared with the counterpart methods [2, 3, 5]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 647, + 287, + 670 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 647, + 287, + 670 + ], + "spans": [ + { + "bbox": [ + 47, + 647, + 287, + 670 + ], + "type": "text", + "content": "Overall, our contributions in this paper can be summarized as three-fold:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 58, + 677, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 677, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 58, + 677, + 287, + 713 + ], + "type": "text", + "content": "- We propose a novel multimodal fusion method termed Embracing Aleatoric Uncertainty (EAU), which quantifies the intrinsic aleatoric uncertainty and leverages it" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 325, + 72, + 515, + 85 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 72, + 515, + 85 + ], + "spans": [ + { + "bbox": [ + 325, + 72, + 515, + 85 + ], + "type": "text", + "content": "to learn stable and robust joint representations." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 317, + 91, + 545, + 193 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 317, + 91, + 545, + 139 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 91, + 545, + 139 + ], + "spans": [ + { + "bbox": [ + 317, + 91, + 545, + 139 + ], + "type": "text", + "content": "- We devise the Stable Unimodal Feature Augmentation (SUFA) module, which quantifies aleatoric uncertainty and learns stable unimodal representations with better semantical consistency." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 317, + 145, + 545, + 193 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 145, + 545, + 193 + ], + "spans": [ + { + "bbox": [ + 317, + 145, + 545, + 193 + ], + "type": "text", + "content": "- We design the Robust Multimodal Feature Integration (RMFI) module to learn compact joint representations with less redundancy, which further improves the robustness of multimodal fusion in our method." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 203, + 392, + 216 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 203, + 392, + 216 + ], + "spans": [ + { + "bbox": [ + 306, + 203, + 392, + 216 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 223, + 545, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 223, + 545, + 510 + ], + "spans": [ + { + "bbox": [ + 304, + 223, + 545, + 510 + ], + "type": "text", + "content": "Multimodal Fusion. Multimodal fusion, which aims at learning stronger representations from different modalities, has become an essential part of a spectrum of computer vision research, such as vision-audio learning [21,22], vision-language learning [23, 24], image retrieval [25-27], and video understanding [28-30]. Generally, multimodal fusion can be categorized into three types: early fusion, intermediate fusion, and late fusion. The previous multimodal fusion methods can be categorized into early fusion and late fusion in terms of their feature-level or decision-level fusion operations. In the last decades, massive works with deep learning suggest that intermediate fusion [13, 29, 31], which learns unified embeddings for multimodal data in the hidden layers of neural networks, could benefit representation learning. Although these multimodal fusion methods achieve remarkable improvements in kinds of multimodal learning tasks, most of them ignore the uncertainty in multimodal learning. Recently, several works [5, 10-13] empirically or theoretically demonstrated that conventional multimodal fusion methods had limited performance and robustness on noisy or corrupted multimodal data. Inspired by these pioneering works, we devised a novel intermediate multimodal fusion method, which could quantify the uncertainty and learn more robust joint representations." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "text", + "content": "Uncertainty in Deep Learning. In general, the uncertainties in deep learning can be categorized into epistemic uncertainty and aleatoric uncertainty [14]. The former aims at capturing the noise of the parameters in deep neural networks, while the latter measures the noise inherent in given training data. To improve the robustness and generalization of open-world scenarios, many researchers incorporate the uncertainty estimation into the deep learning models for computer vision tasks, such as face recognition [15, 16], semantic segmentation [32], and action localization [33, 34]. There is also a list of works that incorporate uncertainty estimation into multimodal learning tasks [5, 17, 22, 35]. However, most of them only quantified uncertainties to learn probability distribution representations that are used for cross-modal interactions, but ignore the value of intrinsic uncertainties and hardly take a further analysis on the robustness of multimodal fusion. Following this observation," + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "26877" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 108 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 108 + ], + "type": "text", + "content": "we develop this work to learn more robust compact joint representations for multimodal learning, which embraces the uncertainties via self-supervised contrastive learning." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 110, + 288, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 110, + 288, + 361 + ], + "spans": [ + { + "bbox": [ + 46, + 110, + 288, + 361 + ], + "type": "text", + "content": "Information Bottleneck Theory. The Information Bottleneck Theory [19] was initially formulated in the context of signal processing, which is proposed to discover a more concise representation of a signal while retaining its utmost informative content. Alemi et al. [36] proposed the Variational Information Bottleneck (VIB) to bridge the gap between Information Bottleneck Theory and deep learning, which approximates the information bottleneck constraints and enables it in deep learning. Based on the VIB, massive researchers introduce the Information Bottleneck Theory in deep learning models to tackle kinds of computer vision tasks, such as object detection [37-39] or image classification [37,38]. Furthermore, as the VIB aims at learning compact minimal representations, it also attracts wide attention from the multimodal learning field [28, 31]. For example, Mai et al. [28] devised a multimodal information bottleneck to learn the minimal sufficient unimodal and multimodal representations. Inspired by their successes, we also devise an multimodal integration strategy based on VIB to reduce the redundancy in joint representations, which further improves the robustness and effectiveness." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 380, + 153, + 394 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 380, + 153, + 394 + ], + "spans": [ + { + "bbox": [ + 47, + 380, + 153, + 394 + ], + "type": "text", + "content": "3. Proposed Method" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 403, + 133, + 415 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 403, + 133, + 415 + ], + "spans": [ + { + "bbox": [ + 47, + 403, + 133, + 415 + ], + "type": "text", + "content": "3.1. Preliminaries" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 424, + 287, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 424, + 287, + 640 + ], + "spans": [ + { + "bbox": [ + 46, + 424, + 287, + 640 + ], + "type": "text", + "content": "Uncertainty Estimation in Deep Learning Models. In terms of [14], the uncertainties in deep learning models can be categorized into aleatoric uncertainty and epistemic uncertainty. The former refers to the uncertainty inherent in the observations and can not be explained away with more data, while the latter only accounts for uncertainty in the model and can be explained away given enough data. Typically, given a deep learning model " + }, + { + "bbox": [ + 46, + 424, + 287, + 640 + ], + "type": "inline_equation", + "content": "f_{\\theta}(\\cdot)" + }, + { + "bbox": [ + 46, + 424, + 287, + 640 + ], + "type": "text", + "content": ", it conducts the mapping " + }, + { + "bbox": [ + 46, + 424, + 287, + 640 + ], + "type": "inline_equation", + "content": "\\mathcal{X} \\rightarrow \\mathcal{Y}" + }, + { + "bbox": [ + 46, + 424, + 287, + 640 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 424, + 287, + 640 + ], + "type": "inline_equation", + "content": "\\mathcal{X}" + }, + { + "bbox": [ + 46, + 424, + 287, + 640 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 424, + 287, + 640 + ], + "type": "inline_equation", + "content": "\\mathcal{Y}" + }, + { + "bbox": [ + 46, + 424, + 287, + 640 + ], + "type": "text", + "content": " represent input data and observed label space respectively. The aleatoric uncertainty in " + }, + { + "bbox": [ + 46, + 424, + 287, + 640 + ], + "type": "inline_equation", + "content": "\\mathcal{X}" + }, + { + "bbox": [ + 46, + 424, + 287, + 640 + ], + "type": "text", + "content": " will lead to a corrupted predicted results in " + }, + { + "bbox": [ + 46, + 424, + 287, + 640 + ], + "type": "inline_equation", + "content": "\\mathcal{Y}" + }, + { + "bbox": [ + 46, + 424, + 287, + 640 + ], + "type": "text", + "content": ". To estimate the uncertainties in deep learning models, a widely adopted strategy is sampling " + }, + { + "bbox": [ + 46, + 424, + 287, + 640 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 424, + 287, + 640 + ], + "type": "text", + "content": " weights " + }, + { + "bbox": [ + 46, + 424, + 287, + 640 + ], + "type": "inline_equation", + "content": "\\theta \\sim p(\\theta | x, y)" + }, + { + "bbox": [ + 46, + 424, + 287, + 640 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 424, + 287, + 640 + ], + "type": "inline_equation", + "content": "x \\in \\mathcal{X}, y \\in \\mathcal{Y}" + }, + { + "bbox": [ + 46, + 424, + 287, + 640 + ], + "type": "text", + "content": " via Dropout operations for different predictions for mean " + }, + { + "bbox": [ + 46, + 424, + 287, + 640 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 46, + 424, + 287, + 640 + ], + "type": "text", + "content": " and variance " + }, + { + "bbox": [ + 46, + 424, + 287, + 640 + ], + "type": "inline_equation", + "content": "\\sigma^2" + }, + { + "bbox": [ + 46, + 424, + 287, + 640 + ], + "type": "text", + "content": ", and adopting a Gaussian posterior " + }, + { + "bbox": [ + 46, + 424, + 287, + 640 + ], + "type": "inline_equation", + "content": "p(y|x) \\sim \\mathcal{N}(\\mu_{*}(x), \\sigma_{*}^{2}(x))" + }, + { + "bbox": [ + 46, + 424, + 287, + 640 + ], + "type": "text", + "content": ". Here we can get the predictive variance and extract the aleatoric uncertainty which represents the ambiguity in " + }, + { + "bbox": [ + 46, + 424, + 287, + 640 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 46, + 424, + 287, + 640 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 59, + 653, + 287, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 653, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 59, + 653, + 287, + 715 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mu_ {*} ^ {2} (x) = \\frac {1}{N} \\sum_ {i} \\mu_ {i} ^ {2} (x), \\\\ \\sigma_ {*} ^ {2} (x) = \\frac {1}{N} \\sum_ {i} \\sigma_ {i} ^ {2} (x) + \\frac {1}{N} \\sum_ {i} \\mu_ {i} ^ {2} (x) - \\mu_ {*} ^ {2} (x) \\tag {1} \\\\ = \\mathbb {E} _ {i} \\left[ \\sigma_ {i} ^ {2} (x) \\right] + \\operatorname {V a r} _ {i} [ \\mu_ {i} (x) ], \\\\ \\end{array}", + "image_path": "21c47761dac33bf7928afba9ca142f3c4f509c92ab4768d682b70209838e3288.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 71, + 545, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 71, + 545, + 144 + ], + "spans": [ + { + "bbox": [ + 304, + 71, + 545, + 144 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 71, + 545, + 144 + ], + "type": "inline_equation", + "content": "\\mathbb{E}_i\\left[\\sigma_i^2 (x)\\right]" + }, + { + "bbox": [ + 304, + 71, + 545, + 144 + ], + "type": "text", + "content": " represents the aleatoric uncertainty, while " + }, + { + "bbox": [ + 304, + 71, + 545, + 144 + ], + "type": "inline_equation", + "content": "\\operatorname {Var}_i[\\mu_i(x)]" + }, + { + "bbox": [ + 304, + 71, + 545, + 144 + ], + "type": "text", + "content": " is the epistemic uncertainty, which attributes to the model " + }, + { + "bbox": [ + 304, + 71, + 545, + 144 + ], + "type": "inline_equation", + "content": "f_{\\theta}(\\cdot)" + }, + { + "bbox": [ + 304, + 71, + 545, + 144 + ], + "type": "text", + "content": " rather than input data " + }, + { + "bbox": [ + 304, + 71, + 545, + 144 + ], + "type": "inline_equation", + "content": "\\mathcal{X}" + }, + { + "bbox": [ + 304, + 71, + 545, + 144 + ], + "type": "text", + "content": ". In our work, we pay more attention to the aleatoric uncertainty since our goal is to quantify the intrinsic noise in multimodal data and leverage it for learning better joint representations." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 144, + 546, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 144, + 546, + 323 + ], + "spans": [ + { + "bbox": [ + 304, + 144, + 546, + 323 + ], + "type": "text", + "content": "Variational Information Bottleneck. The Variational Information Bottleneck (VIB) [20] is an information-theoretic strategy widely adopted in deep learning models, which aims to maintain maximal discriminative feature representations with minimal redundancy. Specifically, given the input variable " + }, + { + "bbox": [ + 304, + 144, + 546, + 323 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 304, + 144, + 546, + 323 + ], + "type": "text", + "content": " with noise or redundant information and the target variable " + }, + { + "bbox": [ + 304, + 144, + 546, + 323 + ], + "type": "inline_equation", + "content": "\\mathbf{y}" + }, + { + "bbox": [ + 304, + 144, + 546, + 323 + ], + "type": "text", + "content": ", the VIB is to learn the compressed latent variable " + }, + { + "bbox": [ + 304, + 144, + 546, + 323 + ], + "type": "inline_equation", + "content": "\\mathbf{z}" + }, + { + "bbox": [ + 304, + 144, + 546, + 323 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 144, + 546, + 323 + ], + "type": "inline_equation", + "content": "\\mathbf{z}" + }, + { + "bbox": [ + 304, + 144, + 546, + 323 + ], + "type": "text", + "content": " is maximally discriminative about the target variable " + }, + { + "bbox": [ + 304, + 144, + 546, + 323 + ], + "type": "inline_equation", + "content": "\\mathbf{y}" + }, + { + "bbox": [ + 304, + 144, + 546, + 323 + ], + "type": "text", + "content": ". Moreover, since the input " + }, + { + "bbox": [ + 304, + 144, + 546, + 323 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 304, + 144, + 546, + 323 + ], + "type": "text", + "content": " is noisy and redundant, the VIB also requires " + }, + { + "bbox": [ + 304, + 144, + 546, + 323 + ], + "type": "inline_equation", + "content": "\\mathbf{z}" + }, + { + "bbox": [ + 304, + 144, + 546, + 323 + ], + "type": "text", + "content": " to be minimally discriminative about the original variable " + }, + { + "bbox": [ + 304, + 144, + 546, + 323 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 304, + 144, + 546, + 323 + ], + "type": "text", + "content": ". In our work, we leverage the VIB to learn compact joint representations, which overcome the redundancy brought by highly aligned distributional multimodal representations to improve the robustness of multimodal fusion." + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 307, + 332, + 545, + 414 + ], + "blocks": [ + { + "bbox": [ + 307, + 332, + 545, + 414 + ], + "lines": [ + { + "bbox": [ + 307, + 332, + 545, + 414 + ], + "spans": [ + { + "bbox": [ + 307, + 332, + 545, + 414 + ], + "type": "image", + "image_path": "7235b48e125f760e7c4f14f3b2d7e2e6e2174341886e5a08e358ffdf17457e53.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 423, + 545, + 456 + ], + "lines": [ + { + "bbox": [ + 304, + 423, + 545, + 456 + ], + "spans": [ + { + "bbox": [ + 304, + 423, + 545, + 456 + ], + "type": "text", + "content": "Figure 2. An illustration of the Unimodal Distributional Representation process. For clarity, we visualize the multivariate Gaussian distributions with two variables only." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 468, + 515, + 480 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 468, + 515, + 480 + ], + "spans": [ + { + "bbox": [ + 306, + 468, + 515, + 480 + ], + "type": "text", + "content": "3.2. Stable Unimodal Feature Augmentation" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 486, + 545, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 486, + 545, + 569 + ], + "spans": [ + { + "bbox": [ + 304, + 486, + 545, + 569 + ], + "type": "text", + "content": "As the data in different modalities contain unique noise in their own modal space, we first propose the Stable Unimodal Feature Augmentation (SUFA) module to quantify their intrinsic aleatoric uncertainty. Following the argument that the uncertainty is attributed to the fuzzy representation of semantics, we further leverage the aleatoric uncertainty via self-supervised contrastive learning." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 570, + 546, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 546, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 546, + 714 + ], + "type": "text", + "content": "Unimodal Distributional Representation. Given the multimodal samples " + }, + { + "bbox": [ + 304, + 570, + 546, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_i^m" + }, + { + "bbox": [ + 304, + 570, + 546, + 714 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 570, + 546, + 714 + ], + "type": "inline_equation", + "content": "m \\in \\mathcal{M}" + }, + { + "bbox": [ + 304, + 570, + 546, + 714 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 304, + 570, + 546, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 304, + 570, + 546, + 714 + ], + "type": "text", + "content": " is the modality set including images, audio, text, etc., we learn the distributional representations to quantify the aleatoric uncertainty in each modality. According to the deviation listed in Eq. 1, we can observe that the aleatoric uncertainty can be predicted with deep learning models " + }, + { + "bbox": [ + 304, + 570, + 546, + 714 + ], + "type": "inline_equation", + "content": "f_{\\theta}(\\cdot)" + }, + { + "bbox": [ + 304, + 570, + 546, + 714 + ], + "type": "text", + "content": " as the variance " + }, + { + "bbox": [ + 304, + 570, + 546, + 714 + ], + "type": "inline_equation", + "content": "\\sigma_i^{m2}" + }, + { + "bbox": [ + 304, + 570, + 546, + 714 + ], + "type": "text", + "content": " directly if the epistemic uncertainty is not considered. To this end, we first employ the corresponding feature extractor to learn the preliminary embeddings for each modality, then deploy two additional fully connected layers to learn a mean vector " + }, + { + "bbox": [ + 304, + 570, + 546, + 714 + ], + "type": "inline_equation", + "content": "\\mu_i^{m2} \\in \\mathbb{R}^d" + }, + { + "bbox": [ + 304, + 570, + 546, + 714 + ], + "type": "text", + "content": " and a variance vector " + }, + { + "bbox": [ + 304, + 570, + 546, + 714 + ], + "type": "inline_equation", + "content": "\\sigma_i^{m2} \\in \\mathbb{R}^d" + }, + { + "bbox": [ + 304, + 570, + 546, + 714 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "26878" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 287, + 109 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 287, + 109 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 287, + 109 + ], + "type": "text", + "content": "Furtherly, we define the representation " + }, + { + "bbox": [ + 47, + 72, + 287, + 109 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_i^m" + }, + { + "bbox": [ + 47, + 72, + 287, + 109 + ], + "type": "text", + "content": " in latent space of each sample " + }, + { + "bbox": [ + 47, + 72, + 287, + 109 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_i^m" + }, + { + "bbox": [ + 47, + 72, + 287, + 109 + ], + "type": "text", + "content": " as a multivariate Gaussian distribution with " + }, + { + "bbox": [ + 47, + 72, + 287, + 109 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 47, + 72, + 287, + 109 + ], + "type": "text", + "content": " variable, which can be represented as:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 93, + 117, + 242, + 148 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 93, + 117, + 242, + 148 + ], + "spans": [ + { + "bbox": [ + 93, + 117, + 242, + 148 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} p (\\mathbf {z} _ {i} ^ {m} | \\mathbf {x} _ {i} ^ {m}) \\sim \\mathcal {N} (\\boldsymbol {\\mu} _ {i} ^ {m}, \\boldsymbol {\\sigma} _ {i} ^ {m 2} \\mathbf {I}), \\\\ \\pmb {\\mu} _ {i} ^ {m} = f _ {\\theta_ {1} ^ {m}} \\left(\\mathbf {x} _ {i} ^ {m}\\right), \\pmb {\\sigma} _ {i} ^ {m} = f _ {\\theta_ {2} ^ {m}} \\left(\\mathbf {x} _ {i} ^ {m}\\right), \\\\ \\end{array}", + "image_path": "f299298f650d3c450b53d2aa631fadb5ac06437077e1a49ce5256d951bbff01f.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 156, + 287, + 227 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 156, + 287, + 227 + ], + "spans": [ + { + "bbox": [ + 47, + 156, + 287, + 227 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 156, + 287, + 227 + ], + "type": "inline_equation", + "content": "f_{\\theta_1^m}(\\cdot)" + }, + { + "bbox": [ + 47, + 156, + 287, + 227 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 156, + 287, + 227 + ], + "type": "inline_equation", + "content": "f_{\\theta_2^m}(\\cdot)" + }, + { + "bbox": [ + 47, + 156, + 287, + 227 + ], + "type": "text", + "content": " represent the two different fully connected layers for mean and variance respectively. " + }, + { + "bbox": [ + 47, + 156, + 287, + 227 + ], + "type": "inline_equation", + "content": "\\mathbf{I}" + }, + { + "bbox": [ + 47, + 156, + 287, + 227 + ], + "type": "text", + "content": " is the identity matrix. To maintain the consistency of semantics in different modalities, we further align the distributional representations of a multimodal sample with the Kullback-Leibler divergence:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 236, + 286, + 305 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 236, + 286, + 305 + ], + "spans": [ + { + "bbox": [ + 56, + 236, + 286, + 305 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} L _ {\\mathrm {U D R}} = \\sum_ {m _ {1} \\neq m _ {2}} ^ {m _ {1}, m _ {2} \\in \\mathcal {M}} K L \\left(p \\left(\\mathbf {z} _ {i} ^ {m _ {1}} \\mid \\mathbf {x} _ {i} ^ {m _ {1}}\\right) \\| p \\left(\\mathbf {z} _ {i} ^ {m _ {2}} \\mid \\mathbf {x} _ {i} ^ {m _ {2}}\\right)\\right) \\tag {3} \\\\ + \\sum_ {m} ^ {m \\in \\mathcal {M}} K L \\left(p \\left(\\mathbf {z} _ {i} ^ {m} \\mid \\mathbf {x} _ {i} ^ {m}\\right) \\| \\mathcal {N} (0, \\mathbf {I})\\right). \\\\ \\end{array}", + "image_path": "1fdcc0c5b079701a531574282faf7074d785bcef189b3d659b0866985a871f77.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 314, + 287, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 314, + 287, + 434 + ], + "spans": [ + { + "bbox": [ + 46, + 314, + 287, + 434 + ], + "type": "text", + "content": "We illustrate the Unimodal Distributional Representation process in Fig. 2 with bi-modal input for clarity. It can be observed the " + }, + { + "bbox": [ + 46, + 314, + 287, + 434 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{UDR}}" + }, + { + "bbox": [ + 46, + 314, + 287, + 434 + ], + "type": "text", + "content": " will push two multimodal distributional representations with similar semantics closer. In this way, the representation of each multimodal sample is not limited within a deterministic point embedding, but a consistent fuzzy representation on several multivariate Gaussian distributions. Particularly, the variance " + }, + { + "bbox": [ + 46, + 314, + 287, + 434 + ], + "type": "inline_equation", + "content": "\\sigma_i^m" + }, + { + "bbox": [ + 46, + 314, + 287, + 434 + ], + "type": "text", + "content": " reveals the aleatoric uncertainty in " + }, + { + "bbox": [ + 46, + 314, + 287, + 434 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 46, + 314, + 287, + 434 + ], + "type": "text", + "content": " modality and the mean " + }, + { + "bbox": [ + 46, + 314, + 287, + 434 + ], + "type": "inline_equation", + "content": "\\mu_i^m" + }, + { + "bbox": [ + 46, + 314, + 287, + 434 + ], + "type": "text", + "content": " is the corresponding stable representation." + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 53, + 443, + 282, + 569 + ], + "blocks": [ + { + "bbox": [ + 53, + 443, + 282, + 569 + ], + "lines": [ + { + "bbox": [ + 53, + 443, + 282, + 569 + ], + "spans": [ + { + "bbox": [ + 53, + 443, + 282, + 569 + ], + "type": "image", + "image_path": "0607f458b577d9e9813f032a20f946587a98b8f103e6fd7a5e4317dfc22208b4.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 577, + 287, + 610 + ], + "lines": [ + { + "bbox": [ + 46, + 577, + 287, + 610 + ], + "spans": [ + { + "bbox": [ + 46, + 577, + 287, + 610 + ], + "type": "text", + "content": "Figure 3. Illustration of the Uncertainty-based Representation Augmentation process. Here we show the distributional representations of image modality only for clarity." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 618, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 618, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 618, + 287, + 713 + ], + "type": "text", + "content": "Uncertainty-based Representation Augmentation. With the quantified aleatoric uncertainty, we consider the second question we raised in Sec.1: should we drop the aleatoric uncertainty in multimodal data? Intuitively, the aleatoric uncertainty in multimodal data is unavoidable since the natural ambiguity of semantics. However, it also results in the diversity of unimodal data in different modalities. To this end, we leverage the aleatoric uncertainty to generate the" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 72, + 545, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 204 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 204 + ], + "type": "text", + "content": "unseen samples such that learned unimodal representations are insensitive to diverse unimodal input with similar semantics. Take the image modality as an example, we illustrate the Uncertainty-based Representation Augmentation process in Fig. 3. Specifically, given a unimodal distributional representation " + }, + { + "bbox": [ + 304, + 72, + 545, + 204 + ], + "type": "inline_equation", + "content": "p(\\mathbf{z}_i^m | \\mathbf{x}_i^m) \\sim \\mathcal{N}(\\boldsymbol{\\mu}_i^m, \\boldsymbol{\\sigma}_i^{m2}\\mathbf{I})" + }, + { + "bbox": [ + 304, + 72, + 545, + 204 + ], + "type": "text", + "content": ", we first random sample an anchor point " + }, + { + "bbox": [ + 304, + 72, + 545, + 204 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{z}}_i^m" + }, + { + "bbox": [ + 304, + 72, + 545, + 204 + ], + "type": "text", + "content": " and an augmented point " + }, + { + "bbox": [ + 304, + 72, + 545, + 204 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_i^m" + }, + { + "bbox": [ + 304, + 72, + 545, + 204 + ], + "type": "text", + "content": " from the multivariate Gaussian distribution as matching pairs. Moreover, we randomly sample a set of negative points from other distributional representations and devise a self-supervised contrastive learning mechanism as follows:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 310, + 210, + 545, + 256 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 310, + 210, + 545, + 256 + ], + "spans": [ + { + "bbox": [ + 310, + 210, + 545, + 256 + ], + "type": "interline_equation", + "content": "L _ {\\mathrm {U R A}} = - \\log \\frac {e ^ {\\operatorname* {s i m} \\left(\\tilde {\\mathbf {z}} _ {i} ^ {m} , \\mathbf {z} _ {i} ^ {m}\\right) / \\tau}}{\\sum_ {j \\neq i} \\left(e ^ {\\operatorname* {s i m} \\left(\\tilde {\\mathbf {z}} _ {i} ^ {m} , \\mathbf {z} _ {i} ^ {m}\\right) / \\tau} + e ^ {\\operatorname* {s i m} \\left(\\tilde {\\mathbf {z}} _ {i} ^ {m} , \\mathbf {z} _ {j} ^ {m}\\right) / \\tau}\\right)}, \\tag {4}", + "image_path": "6304966733bfd1e5ab24ea901d866a75f8ee76ddd7b763967226c861edfb7ef2.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 256, + 545, + 303 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 256, + 545, + 303 + ], + "spans": [ + { + "bbox": [ + 304, + 256, + 545, + 303 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 256, + 545, + 303 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 304, + 256, + 545, + 303 + ], + "type": "text", + "content": " is a temperature factor and sim is cosine similarity calculation. Here we adopt the re-parameterization trick [40] to conduct the sampling operations, which can be formulated as follows:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 358, + 312, + 545, + 326 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 358, + 312, + 545, + 326 + ], + "spans": [ + { + "bbox": [ + 358, + 312, + 545, + 326 + ], + "type": "interline_equation", + "content": "\\mathbf {z} _ {i} ^ {m} = \\boldsymbol {\\mu} _ {i} ^ {m} + \\epsilon \\sigma_ {i} ^ {m}, \\epsilon \\sim \\mathcal {N} (0, \\mathbf {I}). \\tag {5}", + "image_path": "247c21331f06031fe3690608b3829fd0c18f871635875c71700186769400b076.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 333, + 515, + 346 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 333, + 515, + 346 + ], + "spans": [ + { + "bbox": [ + 306, + 333, + 515, + 346 + ], + "type": "text", + "content": "3.3. Robust Multimodal Feature Integration" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 351, + 545, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 351, + 545, + 434 + ], + "spans": [ + { + "bbox": [ + 304, + 351, + 545, + 434 + ], + "type": "text", + "content": "With the SUFA, we obtain stable unimodal representations with consistent semantics. However, we only consider the consistency in each modality but ignore the redundancy caused by duplicated representations. To this end, we proposed the Robust Multimodal Feature Integration (RMFI) module in this section. The overview of the RMFI module is illustrated in Fig. 4." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 436, + 545, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 436, + 545, + 555 + ], + "spans": [ + { + "bbox": [ + 304, + 436, + 545, + 555 + ], + "type": "text", + "content": "Dynamic Multimodal Integration. Inspired by dynamic multimodal fusion [5], we assume that different modalities have unequal contributions to the joint representations in terms of the observed label space. To this end, we first apply a Dynamic Multimodal Integration strategy based on an attention mechanism. Specifically, given the stable unimodal representations " + }, + { + "bbox": [ + 304, + 436, + 545, + 555 + ], + "type": "inline_equation", + "content": "\\pmb{\\mu}_i^m" + }, + { + "bbox": [ + 304, + 436, + 545, + 555 + ], + "type": "text", + "content": ", we calculate the attentive weights in the distributional representations across modalities according to their quantified uncertainties " + }, + { + "bbox": [ + 304, + 436, + 545, + 555 + ], + "type": "inline_equation", + "content": "\\sigma_i^m" + }, + { + "bbox": [ + 304, + 436, + 545, + 555 + ], + "type": "text", + "content": ", and apply them for multimodal integration:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 337, + 563, + 545, + 599 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 337, + 563, + 545, + 599 + ], + "spans": [ + { + "bbox": [ + 337, + 563, + 545, + 599 + ], + "type": "interline_equation", + "content": "\\hat {\\mathbf {x}} _ {i} = \\sum_ {m \\in \\mathcal {M}} \\alpha_ {i} ^ {m} \\boldsymbol {\\mu} _ {i} ^ {m}, \\quad \\alpha_ {i} ^ {m} = \\frac {e ^ {\\frac {1}{\\overline {{\\sigma}} _ {i} ^ {m}}}}{\\sum_ {m \\in \\mathcal {M}} e ^ {\\frac {1}{\\overline {{\\sigma}} _ {i} ^ {m}}}}, \\tag {6}", + "image_path": "58071a46090657d9cc3dd4909a3c8895bd10c83e3431beb84d1d5445da052810.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 605, + 545, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 545, + 676 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 545, + 676 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 605, + 545, + 676 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{x}}_i, \\overline{\\sigma}_i^m" + }, + { + "bbox": [ + 304, + 605, + 545, + 676 + ], + "type": "text", + "content": " represent the integrated joint representations and the average variance of the multivariate Gaussian distribution of " + }, + { + "bbox": [ + 304, + 605, + 545, + 676 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 304, + 605, + 545, + 676 + ], + "type": "text", + "content": " modality. In this way, we preliminarily integrate stable unimodal representations into joint representations, where the contributions of different modalities are estimated dynamically." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 677, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 545, + 712 + ], + "type": "text", + "content": "Joint Representation Compression. In the SUFA module, we fully consider the consistency of semantics in different modalities to avoid noise in unimodal data. However," + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "26879" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 71, + 286, + 157 + ], + "blocks": [ + { + "bbox": [ + 50, + 71, + 286, + 157 + ], + "lines": [ + { + "bbox": [ + 50, + 71, + 286, + 157 + ], + "spans": [ + { + "bbox": [ + 50, + 71, + 286, + 157 + ], + "type": "image", + "image_path": "a08fd6e1048c2c148313340cc28584caf926f1655f346014537314ac9584d642.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 166, + 287, + 200 + ], + "lines": [ + { + "bbox": [ + 46, + 166, + 287, + 200 + ], + "spans": [ + { + "bbox": [ + 46, + 166, + 287, + 200 + ], + "type": "text", + "content": "Figure 4. An illustration of the RMFI module. We provide Cross-Entropy loss and " + }, + { + "bbox": [ + 46, + 166, + 287, + 200 + ], + "type": "inline_equation", + "content": "\\mathrm{L}_1" + }, + { + "bbox": [ + 46, + 166, + 287, + 200 + ], + "type": "text", + "content": " loss for classification-based and regression-based downstream tasks respectively." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 209, + 286, + 304 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 209, + 286, + 304 + ], + "spans": [ + { + "bbox": [ + 46, + 209, + 286, + 304 + ], + "type": "text", + "content": "with similar multivariate Gaussian distributions among different modalities, redundant duplicated information will be introduced into joint representations. Hence, we devise a joint representation compression with the Variational Information Bottleneck (VIB) [20]. Specifically, given the preliminary joint representation " + }, + { + "bbox": [ + 46, + 209, + 286, + 304 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{x}}_i" + }, + { + "bbox": [ + 46, + 209, + 286, + 304 + ], + "type": "text", + "content": " and the target observations " + }, + { + "bbox": [ + 46, + 209, + 286, + 304 + ], + "type": "inline_equation", + "content": "\\mathbf{y}" + }, + { + "bbox": [ + 46, + 209, + 286, + 304 + ], + "type": "text", + "content": " in label space, we learn the compressed joint representation " + }, + { + "bbox": [ + 46, + 209, + 286, + 304 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{z}}" + }, + { + "bbox": [ + 46, + 209, + 286, + 304 + ], + "type": "text", + "content": " in a latent space:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 313, + 286, + 338 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 313, + 286, + 338 + ], + "spans": [ + { + "bbox": [ + 46, + 313, + 286, + 338 + ], + "type": "interline_equation", + "content": "p \\left(\\hat {\\mathbf {z}} _ {i} \\mid \\hat {\\mathbf {x}} _ {i}\\right) \\sim \\mathcal {N} \\left(\\hat {\\boldsymbol {\\mu}} _ {i}, \\hat {\\boldsymbol {\\sigma}} _ {i} ^ {2} \\mathbf {I}\\right), \\quad \\hat {\\boldsymbol {\\mu}} _ {i} = f _ {\\hat {\\theta} _ {1}} (\\hat {\\mathbf {x}} _ {i}), \\quad \\hat {\\boldsymbol {\\sigma}} _ {i} = f _ {\\hat {\\theta} _ {2}} (\\hat {\\mathbf {x}} _ {i}). \\tag {7}", + "image_path": "b83454b17018d69692f013a4dabd791ea7914a8355f51588d93eb30b581868d9.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 338, + 286, + 374 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 338, + 286, + 374 + ], + "spans": [ + { + "bbox": [ + 47, + 338, + 286, + 374 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 338, + 286, + 374 + ], + "type": "inline_equation", + "content": "f_{\\hat{\\theta}_1}" + }, + { + "bbox": [ + 47, + 338, + 286, + 374 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 338, + 286, + 374 + ], + "type": "inline_equation", + "content": "f_{\\hat{\\theta}_2}" + }, + { + "bbox": [ + 47, + 338, + 286, + 374 + ], + "type": "text", + "content": " are two fully connected layers. Similarly, we also adopt the re-parameterization tricks [40] for the final compressed joint representations:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 384, + 286, + 396 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 105, + 384, + 286, + 396 + ], + "spans": [ + { + "bbox": [ + 105, + 384, + 286, + 396 + ], + "type": "interline_equation", + "content": "\\hat {\\mathbf {z}} _ {i} = \\hat {\\boldsymbol {\\mu}} _ {i} + \\epsilon \\hat {\\boldsymbol {\\sigma}} _ {i}, \\epsilon \\sim \\mathcal {N} (0, \\mathbf {I}). \\tag {8}", + "image_path": "35af5270608b4ba851d5a9070b7d57d1dbcc208a6cdaa6c7171d350350de44bc.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 406, + 286, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 406, + 286, + 453 + ], + "spans": [ + { + "bbox": [ + 46, + 406, + 286, + 453 + ], + "type": "text", + "content": "As our proposed method can be applied to different downstream tasks, here we provide two training objectives for the joint representation compression. Specifically, we employ the cross-entropy for the classification-based task:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 51, + 462, + 286, + 487 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 51, + 462, + 286, + 487 + ], + "spans": [ + { + "bbox": [ + 51, + 462, + 286, + 487 + ], + "type": "interline_equation", + "content": "L _ {\\mathrm {J R C}} = - \\mathbf {y} ^ {\\top} \\log \\delta \\left(f _ {\\theta} \\left(\\hat {\\mathbf {z}} _ {i}\\right)\\right) + \\lambda K L \\left(p \\left(\\hat {\\mathbf {z}} _ {i} \\mid \\hat {\\mathbf {x}} _ {i}\\right) \\| \\mathcal {N} (0, \\mathbf {I})\\right), \\tag {9}", + "image_path": "7d2e6d9daa0bd4497b62c833c8ee5a47096d61802b94dd4d1da012cd2a7b3343.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 487, + 286, + 535 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 487, + 286, + 535 + ], + "spans": [ + { + "bbox": [ + 46, + 487, + 286, + 535 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 487, + 286, + 535 + ], + "type": "inline_equation", + "content": "\\delta (\\cdot)" + }, + { + "bbox": [ + 46, + 487, + 286, + 535 + ], + "type": "text", + "content": " represents softmax function and " + }, + { + "bbox": [ + 46, + 487, + 286, + 535 + ], + "type": "inline_equation", + "content": "f_{\\theta}(\\cdot)" + }, + { + "bbox": [ + 46, + 487, + 286, + 535 + ], + "type": "text", + "content": " is the deep learning models for classification, " + }, + { + "bbox": [ + 46, + 487, + 286, + 535 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 46, + 487, + 286, + 535 + ], + "type": "text", + "content": " is a hyperparameter. As for the regression-based task, we employ the mean square error as the training objective:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 58, + 544, + 286, + 558 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 544, + 286, + 558 + ], + "spans": [ + { + "bbox": [ + 58, + 544, + 286, + 558 + ], + "type": "interline_equation", + "content": "L _ {\\mathrm {J R C}} = \\left[ y - f _ {\\theta} \\left(\\hat {\\mathbf {z}} _ {i}\\right) \\right] ^ {2} + K L \\left(p \\left(\\hat {\\mathbf {z}} _ {i} \\mid \\hat {\\mathbf {x}} _ {i}\\right) \\| \\mathcal {N} (0, \\mathbf {I})\\right). \\tag {10}", + "image_path": "7939389d79cc9c5f1d7923d249b6b029fe2fe2f8853a0e84b8314fa1dbae3229.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 567, + 127, + 581 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 567, + 127, + 581 + ], + "spans": [ + { + "bbox": [ + 47, + 567, + 127, + 581 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 587, + 173, + 601 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 587, + 173, + 601 + ], + "spans": [ + { + "bbox": [ + 47, + 587, + 173, + 601 + ], + "type": "text", + "content": "4.1. Experimental Settings" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 605, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 286, + 713 + ], + "type": "text", + "content": "Datasets. We evaluate our method on five multimodal datasets, including two tri-modal datasets, i.e., CMU-MOSI [41] and CMU-MOSEI [42] and three bi-modal datasets, i.e., MVSA-Single [43], UPMC Food101 [1], and NYU Depth v2 [44] datasets. The CMU-MOSI [41] and CMU-MOSEI [42] datasets are widely employed video datasets for Multimodal Sentiment Analysis (MSA) task. The former encompasses 2199 short videos, each accompanied by a sentiment strength score ranging from -3 to 3, offering" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 72, + 545, + 275 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 275 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 275 + ], + "type": "text", + "content": "a nuanced measure of emotional intensity. The latter, designed specifically for emotion analysis tasks, is a larger dataset comprising 22,856 movie review clips. The bimodal dataset, i.e., MVSA-Single [43] is also used for MSA task but it only contains image-text pairs obtained from social media for sentiment classification. Following [2-5], we also employ the UPMC Food101 [1] and the NYU Depth v2 [44] datasets for Multimodal Image Classification (MIC) task. The UPMC Food101 dataset contain 101 categories, respectively, these datasets are specifically designed for bimodal classification tasks. The UPMC FOOD101 dataset contains images with 101 categories obtained by Google Image Search and their corresponding textual descriptions. The NYU Depth v2 dataset [44] is utilized for scene recognition and comprises depth and RGB images. Following [5], we also adopt the commonly used 9 out of the 27 scene categories and the remaining categories as \"Others\"." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 276, + 545, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 276, + 545, + 407 + ], + "spans": [ + { + "bbox": [ + 304, + 276, + 545, + 407 + ], + "type": "text", + "content": "Implementation Details. In alignment with alternative methodologies, for our trimodal video dataset, we employ FACET, COVAREP, and BERT as feature extractors for visual, audio, and text modalities, respectively. For the bimodal dataset, we utilize ResNet-152 as the feature extractor for RGB and depth images, coupled with BERT as the feature extractor for text. As for the hyperparameters, we set the temperature factor " + }, + { + "bbox": [ + 304, + 276, + 545, + 407 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 304, + 276, + 545, + 407 + ], + "type": "text", + "content": " and the balance factor " + }, + { + "bbox": [ + 304, + 276, + 545, + 407 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 304, + 276, + 545, + 407 + ], + "type": "text", + "content": " to 0.5 and 1e-3 respectively. We employed the Adam optimizer with learning rate of 1e-5,and adopt the Reduce-on-Plateau learning rate adjustment strategy to train our EAU method." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 408, + 545, + 504 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 408, + 545, + 504 + ], + "spans": [ + { + "bbox": [ + 304, + 408, + 545, + 504 + ], + "type": "text", + "content": "Evaluation Metrics. For the regression-based MSA task based on CMU-MOSI and CMU-MOSEI, we follow the previous work [6, 7, 9, 45] and adopt Acc7, F1 score, and Pearson correlation coefficient as our metrics. As for the classification-based MSA task on the MSVA-Single dataset and MIC task on the UPMC FOOD101 and The NYU Depth v2 datasets, we reported commonly used metrics, including accuracy and F1 score." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 514, + 503, + 527 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 514, + 503, + 527 + ], + "spans": [ + { + "bbox": [ + 306, + 514, + 503, + 527 + ], + "type": "text", + "content": "4.2. Comparisons with State-Of-The-Arts" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 533, + 545, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 533, + 545, + 688 + ], + "spans": [ + { + "bbox": [ + 304, + 533, + 545, + 688 + ], + "type": "text", + "content": "To show the superiority of our proposed EAU method on multimodal fusion, we compared our EAU method with existing state-of-the-art methods of the MSA task on the CMU-MOSI and CMU-MOSEI datasets, including MIB [28], HMA [6], MIM [7], GCNet [46], ConFEDE [8], DiC-MoR [45], and DMD [9]. Moreover, we also make fair comparisons with the widely adopted simple multimodal fusion strategies i.e., Concat and Late Fusion, as well as recent well-designed multimodal fusion strategies on the bimodal datasets, including TMC [3], ITIN [47], MMBT [2], PMF [4], MVCN [48], and QMF [5]. Particularly, similar to the counterpart method QMF [5], we evaluate our method on noisy datasets to observe the model robustness." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "content": "Multimodal Fusion on Classification Task. We report the performance of classification tasks, including both MSA" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "26880" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 51, + 70, + 299, + 178 + ], + "blocks": [ + { + "bbox": [ + 51, + 70, + 299, + 178 + ], + "lines": [ + { + "bbox": [ + 51, + 70, + 299, + 178 + ], + "spans": [ + { + "bbox": [ + 51, + 70, + 299, + 178 + ], + "type": "table", + "html": "
MethodCMU-MOSICMU-MOSEI
Acc7F1CorrAcc7F1Corr
MIB [28] (2022)48.685.30.79854.186.20.790
HMA [6] (2023)45.385.60.78252.885.40.787
MIM [7] (2023)47.085.90.80552.586.30.792
GCNet [46] (2023)44.985.1-51.585.2-
ConFEDE [8] (2023)42.385.50.78454.985.80.780
DiCMoR [45] (2023)45.385.6-53.485.1-
DMD [9] (2023)45.686.0-54.586.6-
EAU (Ours)48.886.20.80954.886.90.816
", + "image_path": "abc54406763ec0bed9548d3d49bda7a6f3821aed706d8ad8e547843f9a4beca5.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 310, + 71, + 543, + 178 + ], + "blocks": [ + { + "bbox": [ + 310, + 71, + 543, + 178 + ], + "lines": [ + { + "bbox": [ + 310, + 71, + 543, + 178 + ], + "spans": [ + { + "bbox": [ + 310, + 71, + 543, + 178 + ], + "type": "table", + "html": "
MethodMVSA-SingleFood 101NYU Depth v2
AccF1AccF1AccF1
Concat65.5965.4388.2088.1970.3069.82
Late Fusion76.8875.7290.6990.7769.1468.32
MMBT [2] (2020)78.50-91.5291.2871.04-
TMC [3] (2021)76.0674.5589.8689.8071.0669.83
ITIN [47] (2022)75.1974.97----
PMF [4] (2023)--91.68---
MVCN [48] (2023)76.0674.55----
QMF [5] (2023)78.0777.1892.9292.9370.0968.65
EAU (Ours)79.1578.3693.2093.1872.0570.63
", + "image_path": "79e18773d8283195a9616a82475616370b4b444929398cad45c75c729a5fc75e.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 238, + 289, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 238, + 289, + 430 + ], + "spans": [ + { + "bbox": [ + 46, + 238, + 289, + 430 + ], + "type": "text", + "content": "and MIC tasks on bi-modal datasets in Table 1, where the best results are marked in bold. From the experimental results, we can observe that our proposed EAU method outperforms all three datasets. Particularly, compared with the recent state-of-the-art method QMF method [5], our EAU method achieves at least " + }, + { + "bbox": [ + 46, + 238, + 289, + 430 + ], + "type": "inline_equation", + "content": "1\\%" + }, + { + "bbox": [ + 46, + 238, + 289, + 430 + ], + "type": "text", + "content": " absolute improvements on the MVSA-Single and NYU Depth v2 datasets. These results demonstrate that our EAU approach conducts better multimodal fusion on text, depth, and RGB images, affirming the effectiveness of leveraging aleatoric uncertainty in multimodal data. However, we also note that the improvements on UPMC Food 101 are not as remarkable as the other two benchmark datasets. We speculate that one probable reason is the images and texts in this dataset are clear compared with the others. Hence, our EAU method obtains fewer improvements since the augmentation is limited." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 434, + 287, + 625 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 434, + 287, + 625 + ], + "spans": [ + { + "bbox": [ + 46, + 434, + 287, + 625 + ], + "type": "text", + "content": "Multimodal Fusion on Regression Task. We evaluate our method on the regression task on the CMU-MOSEI and CMU-MOSI datasets, where the MSA task is conducted by predicting sentiment strengths. From the experimental results in Table 1, we can observe that: For the CMU-MOSI dataset, our proposed EAU method achieves state-of-the-art performance across all evaluation metrics, with a particularly significant improvement noted in Acc7. On the CMU-MOSEI dataset, our method has also demonstrated substantial improvements on most evaluation metrics. These results demonstrate the superiority of our proposed EAU, which benefits from high-quality multimodal fusion via learning stable and robust joint representations. Simultaneously, these results indicate that our approach can effortlessly extend to other modalities and can be seamlessly transferred to frameworks accommodating kinds of modalities." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 630, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 630, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 630, + 288, + 715 + ], + "type": "text", + "content": "Model Robustness on Noisy Multimodal Datasets. To validate the effectiveness of our model in handling data noise, we conducted more evaluation on the noisy datasets following a recent multimodal fusion strategy QMF [5]. Concretely, we consider different intensities of Gaussian and Salt-Pepper noise on the MVSA-Single and NYU Depth v2 datasets. For fair comparisons, we conducted 10" + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 309, + 236, + 544, + 333 + ], + "blocks": [ + { + "bbox": [ + 46, + 183, + 547, + 228 + ], + "lines": [ + { + "bbox": [ + 46, + 183, + 547, + 228 + ], + "spans": [ + { + "bbox": [ + 46, + 183, + 547, + 228 + ], + "type": "text", + "content": "Table 1. Comparisons with state-of-the-art multimodal fusion methods on MSA and MIC tasks. The CMU-MOSI and CMU-MOSEI datasets contain videos, audi, and texts. The MVSA-Single and Food-101 datasets consist of texts and RGB images. The NYU Depth v2 contains RGB and depth images. Note that CMU-MOSI and CMU-MOSEI are used for regression-based MSA tasks, while the others are used for classification-based MSA or MIC tasks." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 309, + 236, + 544, + 333 + ], + "lines": [ + { + "bbox": [ + 309, + 236, + 544, + 333 + ], + "spans": [ + { + "bbox": [ + 309, + 236, + 544, + 333 + ], + "type": "table", + "html": "
Noisy MVSA-Single
MethodCleanSalt-Pepper NoiseGaussian Noise
ε = 0ε = 5ε = 10ε = 5ε = 10
Concat65.5958.6951.1650.7046.12
Late Fusion76.8867.8855.4363.4655.16
MMBT [2] (2020)78.5074.0751.2671.9955.34
TMC [3] (2021)74.8768.0256.6266.7260.35
QMF [5] (2023)78.0773.9060.4173.8561.28
EAU (Ours)79.1574.8161.0473.8962.04
", + "image_path": "197cb90e57e853e37c1dd3fa9ab7e35c8ffcb684b367d4b1a5b2cc6a54fc5ebc.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 309, + 334, + 544, + 429 + ], + "blocks": [ + { + "bbox": [ + 309, + 334, + 544, + 429 + ], + "lines": [ + { + "bbox": [ + 309, + 334, + 544, + 429 + ], + "spans": [ + { + "bbox": [ + 309, + 334, + 544, + 429 + ], + "type": "table", + "html": "
Noisy NYU Depth v2
MethodCleanSalt-Pepper NoiseGaussian Noise
ε = 0ε = 5ε = 10ε = 5ε = 10
Concat70.4457.9844.5159.9753.20
Late fusion69.1656.2741.2259.6351.99
MMTM [2] (2020)71.0459.4544.5960.3752.28
TMC [3] (2021)71.0659.3444.6561.0453.36
QMF [5] (2023)70.0958.5045.6961.6255.60
EAU (Ours)72.0559.8346.8563.3358.85
", + "image_path": "2387cdab7f44efbc90c0d17f441fcb498aafbc1c3d891286b006f5d58133445e.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 434, + 545, + 456 + ], + "lines": [ + { + "bbox": [ + 306, + 434, + 545, + 456 + ], + "spans": [ + { + "bbox": [ + 306, + 434, + 545, + 456 + ], + "type": "text", + "content": "Table 2. Comparisons with state-of-the-arts concerning model performance on noisy MVSA-Single and NYU Depth v2 datasets." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 468, + 545, + 648 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 468, + 545, + 648 + ], + "spans": [ + { + "bbox": [ + 304, + 468, + 545, + 648 + ], + "type": "text", + "content": "experiments with different random seeds and reported the mean results as our final performance. According to the experimental results listed in Table 2, we can observe that our method has achieved state-of-the-art performance under various types and intensities of noise. Moreover, with the intensity of noise increasing, the margin of performance between our method and existing multimodal fusion strategies goes larger. Particularly, our method achieves more than " + }, + { + "bbox": [ + 304, + 468, + 545, + 648 + ], + "type": "inline_equation", + "content": "3\\%" + }, + { + "bbox": [ + 304, + 468, + 545, + 648 + ], + "type": "text", + "content": " improvement on the Noisy NYU Depth v2 dataset over the QMF method [5]. Such a result demonstrates our method reveals better robustness against data noise. The reason is our model pioneeringly leverages the intrinsic aleatoric uncertainty in the training data to enhance the stable feature representation, which is not fully considered in existing counterpart methods [2, 3, 5]." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 658, + 409, + 671 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 658, + 409, + 671 + ], + "spans": [ + { + "bbox": [ + 306, + 658, + 409, + 671 + ], + "type": "text", + "content": "4.3. Further Analysis" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 677, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 677, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 677, + 547, + 713 + ], + "type": "text", + "content": "Analysis on Model Structure. To explore the impact of different model structures, we decomposed our proposed method into two components, SUFA and RMFI, and evalu" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "26881" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 61, + 71, + 407, + 184 + ], + "blocks": [ + { + "bbox": [ + 61, + 71, + 407, + 184 + ], + "lines": [ + { + "bbox": [ + 61, + 71, + 407, + 184 + ], + "spans": [ + { + "bbox": [ + 61, + 71, + 407, + 184 + ], + "type": "table", + "html": "
MethodCleanSalt-Pepper NoiseGaussian Noise
Acc@ε=0Acc@ε=5Acc@ε=10Acc@ε=5Acc@ε=10
Naive Backbone + Late Fusion76.88 ± 1.3067.88 ± 1.8755.43 ± 1.9463.46 ± 3.4655.16 ± 3.60
Naive Backbone + Concat74.53 ± 0.9762.24 ± 2.4753.14 ± 2.8956.08 ± 1.8949.21 ± 2.81
Naive Backbone + RMFI77.23 ± 0.7470.99 ± 2.7759.11 ± 1.9266.53 ± 2.2652.02 ± 3.45
SUFA + Late Fusion77.11 ± 1.0570.05 ± 1.5055.93 ± 2.2562.27 ± 4.2055.90 ± 3.67
SUFA + Concat78.18 ± 1.0873.10 ± 1.0760.67 ± 1.3871.40 ± 1.4859.11 ± 1.60
BERT (NAACL'19) [49]75.61 ± 0.5369.50 ± 1.5047.41 ± 0.7969.50 ± 1.5047.41 ± 0.79
MMBT (arXiv'19) [2]78.50 ± 0.4074.07 ± 1.1251.26 ± 5.6571.99 ± 1.0455.34 ± 2.84
TMC (ICLR'21) [3]74.87 ± 2.2468.02 ± 3.0756.62 ± 3.6766.72 ± 4.5560.35 ± 2.79
QMF (ICML'23) [5]78.07 ± 1.1073.90 ± 1.8960.41 ± 2.6373.85 ± 1.4261.28 ± 2.12
SUFA + RMFI (EAU)79.15 ± 0.6074.81 ± 1.5961.04 ± 1.3173.89 ± 0.9662.04 ± 1.26
", + "image_path": "4fadb6373ceae741e89bb3c819aef9c6508ac910406e925570672f59108b7fd1.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 410, + 71, + 531, + 184 + ], + "blocks": [ + { + "bbox": [ + 410, + 71, + 531, + 184 + ], + "lines": [ + { + "bbox": [ + 410, + 71, + 531, + 184 + ], + "spans": [ + { + "bbox": [ + 410, + 71, + 531, + 184 + ], + "type": "table", + "html": "
ModalityAccF1Corr
T44.183.70.785
V16.745.30.072
A15.848.70.099
T+A46.784.20.794
T+V46.384.60.786
A+V20.255.30.117
T+A+V48.886.20.809
T76.3075.90-
V63.5863.35-
T+V79.1578.36-
", + "image_path": "73d9ef011638ff007667783360b4eba5edcf3d14cc9a4d1cdff9740a24e2d1dc.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 189, + 545, + 212 + ], + "lines": [ + { + "bbox": [ + 46, + 189, + 545, + 212 + ], + "spans": [ + { + "bbox": [ + 46, + 189, + 545, + 212 + ], + "type": "text", + "content": "Table 3. Analysis concerning key model structures in our EAU method on the Noisy MSVA-Single (Left), and analysis concerning modalities on the CMU-MOSI (Right Upper) and MVSA-Single (Right Lower) datasets. Note that BERT [49] adopts text modality only." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 220, + 287, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 220, + 287, + 376 + ], + "spans": [ + { + "bbox": [ + 46, + 220, + 287, + 376 + ], + "type": "text", + "content": "ated their effectiveness in combination with various model structures on the Noisy MVSA-Single dataset. Specifically, we employ two alternative designs for SUFA and RMFI respectively: (1) Naive Backbone: deploying the feature extractors directly, where the distributional representation and stable feature augmentation in the SUFA module are disabled. (2) Late Fusion: the unimodal representations are used for downstream tasks first and the fusion operations are conducted in probability space. To observe the stability of our method, we conduct 10 experiments with different random seeds and report the mean and variance in Table 3. Here we also make fair comparisons with several counterpart methods [2, 3, 5, 49] to furtherly show our superiority." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 378, + 287, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 378, + 287, + 569 + ], + "spans": [ + { + "bbox": [ + 46, + 378, + 287, + 569 + ], + "type": "text", + "content": "According to the experimental results presented in Table 3, we can observe that: (1) Our proposed SUFA and RMFI modules consistently improve performance on all metrics. Particularly, with the SUFA module, the results of fusion strategy Concat are boosted with more than " + }, + { + "bbox": [ + 46, + 378, + 287, + 569 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 46, + 378, + 287, + 569 + ], + "type": "text", + "content": " absolute performance on the noisy data. Moreover, compared with Late Fusion, our proposed RMFI module can also significantly boost the classification accuracy on noisy data with the SUFA module. These results prove again that our method shows great superiority in the effectiveness and robustness of multimodal fusion. (2) It also explicates that our proposed EAU method remarkably outperforms compared with the counterpart methods [2, 3, 5]. Particularly, compared with the recent state-of-the-art method QMF [5], our complete model reveals better performance and stability, demonstrating the capability of our EAU method again." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 570, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 570, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 570, + 287, + 713 + ], + "type": "text", + "content": "Analysis on the Effectiveness of Multimodal Fusion. To verify the effectiveness of multimodal fusion in our EAU method, we also conduct ablation studies concerning different modalities on the CMU-MOSI datasets, which consist of video (V), text (T), and audio (A) three modalities. By observing the experimental results illustrated in Table 3, we have drawn the following conclusions: (1) Our fusion strategy demonstrated significant effectiveness across different modal combinations. As the number of modalities increased, our approach facilitated the integration of multimodal information, resulting in a significant improvement across all evaluation metrics. (2) We also note that the text" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 220, + 545, + 304 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 220, + 545, + 304 + ], + "spans": [ + { + "bbox": [ + 304, + 220, + 545, + 304 + ], + "type": "text", + "content": "modality performs an essential role in the MSA task on the CMU-MOSI dataset. However, combined with the other two modalities via our proposed EAU method, the results are also boosted significantly, demonstrating the superiority of our EAU method again. Such a phenomenon also proves the rationality of our Dynamic Multimodal Integration process in the RMFI module." + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 306, + 310, + 426, + 417 + ], + "blocks": [ + { + "bbox": [ + 306, + 310, + 426, + 417 + ], + "lines": [ + { + "bbox": [ + 306, + 310, + 426, + 417 + ], + "spans": [ + { + "bbox": [ + 306, + 310, + 426, + 417 + ], + "type": "image", + "image_path": "8668a04804c4b09a944523bf3b075fa005aa77bbd40a05d1a1e70698757000bc.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 426, + 545, + 448 + ], + "lines": [ + { + "bbox": [ + 304, + 426, + 545, + 448 + ], + "spans": [ + { + "bbox": [ + 304, + 426, + 545, + 448 + ], + "type": "text", + "content": "Figure 5. Analysis for robustness evaluated by over 10 times random experiments under different noisy data." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 426, + 310, + 545, + 417 + ], + "blocks": [ + { + "bbox": [ + 426, + 310, + 545, + 417 + ], + "lines": [ + { + "bbox": [ + 426, + 310, + 545, + 417 + ], + "spans": [ + { + "bbox": [ + 426, + 310, + 545, + 417 + ], + "type": "image", + "image_path": "71ff691682e7c3f72c18382f828a43e21e7dd09a6ae48544210d4add733b0410.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 450, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 450, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 450, + 545, + 713 + ], + "type": "text", + "content": "Analysis on Model Robustness. To verify the robustness and stability of our proposed EAU method, we conduct more ablation studies on the MVSA-Single and NYU Depth v2 datasets under different noise levels and observe the sparsity of final accuracy. Specifically, we conduct more than 10 experiments randomly with complete EAU (denoted as Complete) and EAU w/o SUFA (denoted as w/o SUFA), and show the statistics in Fig. 5. By observing the experimental results, we can find that: (1) The SUFA module consistently shows its effectiveness on both two datasets under different noise strengths. Particularly, when introducing more noise to the data, the ablated model shows strong fluctuations. Contrastively, the complete EAU model reveals remarkably reduced fluctuations, which demonstrates the superiority of model robustness and stability. (2) We also note that the improvements on the NYU Depth v2 are slightly lower than the MVSA-Single dataset, though both two datasets have significant results. We speculate this is caused by the difference between text modality and depth images. Concretely, the depth images can be regarded as another view of spatial information, which has less complementary information for RGB images compared with texts." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "26882" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 287, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 287, + 96 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 287, + 96 + ], + "type": "text", + "content": "To this end, the introduced noise will show more impact on the depth-RGB than the image-text compositions." + } + ] + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 47, + 105, + 287, + 203 + ], + "blocks": [ + { + "bbox": [ + 47, + 105, + 287, + 203 + ], + "lines": [ + { + "bbox": [ + 47, + 105, + 287, + 203 + ], + "spans": [ + { + "bbox": [ + 47, + 105, + 287, + 203 + ], + "type": "image", + "image_path": "726e724da2f8f1cecd8658f2394ba73f8127abb5c4412c3104b23edd9751a0ac.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 47, + 207, + 286, + 228 + ], + "lines": [ + { + "bbox": [ + 47, + 207, + 286, + 228 + ], + "spans": [ + { + "bbox": [ + 47, + 207, + 286, + 228 + ], + "type": "text", + "content": "Figure 6. Analysis concerning the training convergence and performance fluctuations during training." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 236, + 288, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 236, + 288, + 415 + ], + "spans": [ + { + "bbox": [ + 46, + 236, + 288, + 415 + ], + "type": "text", + "content": "Analysis on Training Process. We also illustrate the training process of our proposed EAU method on the MVSA-Single in Fig 6 to observe the training convergence and performance fluctuations. According to the illustrated experimental results, we can observe that the complete EAU model exhibits a smoother training process and converges more rapidly with a remarkably better performance. Particularly, compared with the ablated model w/o RMFI, the other two models show much better stability during training. This is because the fused features learned by the ablated model w/o RMFI have redundant information caused by cross-modal distributional alignment in the SUFA module. It proves again the rationality of our proposed RMFI module, which deploys an information-theoretic strategy to address the limitation." + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 50, + 423, + 156, + 503 + ], + "blocks": [ + { + "bbox": [ + 50, + 423, + 156, + 503 + ], + "lines": [ + { + "bbox": [ + 50, + 423, + 156, + 503 + ], + "spans": [ + { + "bbox": [ + 50, + 423, + 156, + 503 + ], + "type": "image", + "image_path": "a0a44e1bc2b030f65ccba3a62e40971d0ae6ad4b7430367fd66c5b355106b151.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 89, + 506, + 119, + 514 + ], + "lines": [ + { + "bbox": [ + 89, + 506, + 119, + 514 + ], + "spans": [ + { + "bbox": [ + 89, + 506, + 119, + 514 + ], + "type": "text", + "content": "(a) Concat" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 171, + 422, + 287, + 503 + ], + "blocks": [ + { + "bbox": [ + 171, + 422, + 287, + 503 + ], + "lines": [ + { + "bbox": [ + 171, + 422, + 287, + 503 + ], + "spans": [ + { + "bbox": [ + 171, + 422, + 287, + 503 + ], + "type": "image", + "image_path": "b006de730ac2a0fa72ffc5821911aa703b6e4f3a889149c36ef3c7fdec7b7191.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 50, + 516, + 156, + 595 + ], + "blocks": [ + { + "bbox": [ + 50, + 516, + 156, + 595 + ], + "lines": [ + { + "bbox": [ + 50, + 516, + 156, + 595 + ], + "spans": [ + { + "bbox": [ + 50, + 516, + 156, + 595 + ], + "type": "image", + "image_path": "4025742e48aae512ebe097fe777828d2f04e9687ae3b56ed36661b85da848df5.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 80, + 597, + 130, + 605 + ], + "lines": [ + { + "bbox": [ + 80, + 597, + 130, + 605 + ], + "spans": [ + { + "bbox": [ + 80, + 597, + 130, + 605 + ], + "type": "text", + "content": "(c) EAU w/o RMFI" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 169, + 515, + 287, + 594 + ], + "blocks": [ + { + "bbox": [ + 215, + 506, + 268, + 514 + ], + "lines": [ + { + "bbox": [ + 215, + 506, + 268, + 514 + ], + "spans": [ + { + "bbox": [ + 215, + 506, + 268, + 514 + ], + "type": "text", + "content": "(b) EAU w/o SUFA" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 169, + 515, + 287, + 594 + ], + "lines": [ + { + "bbox": [ + 169, + 515, + 287, + 594 + ], + "spans": [ + { + "bbox": [ + 169, + 515, + 287, + 594 + ], + "type": "image", + "image_path": "68fa03c3a5c26ed034bbc01e245e8d35c8dc77ad8493aae4bf3f6b40f7e00693.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 227, + 597, + 250, + 605 + ], + "lines": [ + { + "bbox": [ + 227, + 597, + 250, + 605 + ], + "spans": [ + { + "bbox": [ + 227, + 597, + 250, + 605 + ], + "type": "text", + "content": "(d) EAU" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 614, + 287, + 647 + ], + "lines": [ + { + "bbox": [ + 46, + 614, + 287, + 647 + ], + "spans": [ + { + "bbox": [ + 46, + 614, + 287, + 647 + ], + "type": "text", + "content": "Figure 7. Visualizations of joint representations by t-SNE [50] on MVSA-Single dataset. Here we adopt Concat for the ablated model w/o RMFI." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 653, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 287, + 713 + ], + "type": "text", + "content": "Visualizations of Joint Representations. To further verify the superiority of our proposed EAU method, we also employ the t-SNE [50] to visualize the learned joint representations. As illustrated in Fig. 7, it can be observed that the features generated by our method exhibit a more" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "text", + "content": "compact and discriminative distribution. In contrast, when the SUFA or RMFI module is removed, more sparse points appear around the feature distribution which leads to corrupted classification results at last. Such results indicate that our method is more capable of learning a representative joint representation." + } + ] + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 307, + 149, + 545, + 277 + ], + "blocks": [ + { + "bbox": [ + 307, + 149, + 545, + 277 + ], + "lines": [ + { + "bbox": [ + 307, + 149, + 545, + 277 + ], + "spans": [ + { + "bbox": [ + 307, + 149, + 545, + 277 + ], + "type": "image", + "image_path": "9ffa48bd060d471068ad61f4ee8c613e060afdd37867e1a17166aeaa5ee2981b.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 285, + 545, + 319 + ], + "lines": [ + { + "bbox": [ + 304, + 285, + 545, + 319 + ], + "spans": [ + { + "bbox": [ + 304, + 285, + 545, + 319 + ], + "type": "text", + "content": "Figure 8. Visualizations of test cases selected from the MSVA-Single datasets. It can be observed that our EAU method reveals better robustness against the noise." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 323, + 545, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 323, + 545, + 443 + ], + "spans": [ + { + "bbox": [ + 304, + 323, + 545, + 443 + ], + "type": "text", + "content": "Qualitative Analysis. Moreover, we also illustrate several typical test cases in Fig. 8 and make comparisons with the recent counterpart method QMF [5]. Under different noise settings, our proposed EAU method consistently classifies the sentiment of the given image-text input precisely, while the QMF method returns incorrect results. It explicates that our proposed EAU method benefits from the well-designed SUFA and RMFI module, which is effective in learning more stable and robust joint representations for noise-resistant performance with higher accuracy." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 306, + 455, + 378, + 467 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 455, + 378, + 467 + ], + "spans": [ + { + "bbox": [ + 306, + 455, + 378, + 467 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 475, + 545, + 632 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 475, + 545, + 632 + ], + "spans": [ + { + "bbox": [ + 304, + 475, + 545, + 632 + ], + "type": "text", + "content": "In this paper, we proposed a novel multimodal fusion method, namely, Embracing Aleatoric Uncertainty (EAU). It achieves more discriminative joint representations by fully considering the aleatoric uncertainty in multimodal data. Specifically, with the well-designed Stable Unimodal Feature Augmentation (SUFA) and Robust Multimodal Feature Integration (RMFI) modules, our proposed EAU could learn compact and robust joint representations. We evaluated our proposed EAU method on five multimodal benchmark datasets for both classification and regression tasks and demonstrated its superiority in fusion performance and robustness. For future work, we will further explore the uncertainties in multimodal learning." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 306, + 645, + 416, + 658 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 645, + 416, + 658 + ], + "spans": [ + { + "bbox": [ + 306, + 645, + 416, + 658 + ], + "type": "text", + "content": "6. Acknowledgement" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "content": "This work was sponsored in part by the National Natural Science Foundation of China under Grants (No. 62222203 and No. 62072080) and the New Cornerstone Science Foundation through the XPLORER PRIZE." + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "26883" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 703 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 134 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 134 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 134 + ], + "type": "text", + "content": "[1] Lukas Bossard, Matthieu Guillaumin, and Luc Van Gool. Food-101-mining discriminative components with random forests. In European Conference on Computer Vision, pages 446-461, 2014. 1, 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 137, + 288, + 181 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 137, + 288, + 181 + ], + "spans": [ + { + "bbox": [ + 53, + 137, + 288, + 181 + ], + "type": "text", + "content": "[2] Douwe Kiela, Suvrat Bhooshan, Hamed Firooz, Ethan Perez, and Davide Testuggine. Supervised multimodal bittransformers for classifying images and text. arXiv preprint arXiv:1909.02950, 2019. 1, 2, 5, 6, 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 183, + 288, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 183, + 288, + 217 + ], + "spans": [ + { + "bbox": [ + 53, + 183, + 288, + 217 + ], + "type": "text", + "content": "[3] Zongbo Han, Changqing Zhang, Huazhu Fu, and Joey Tianyi Zhou. Trusted multi-view classification. arXiv preprint arXiv:2102.02051, 2021. 1, 2, 5, 6, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 219, + 288, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 219, + 288, + 262 + ], + "spans": [ + { + "bbox": [ + 53, + 219, + 288, + 262 + ], + "type": "text", + "content": "[4] Yaowei Li, Ruijie Quan, Linchao Zhu, and Yi Yang. Efficient multimodal fusion via interactive prompting. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 2604-2613, 2023. 1, 5, 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 265, + 287, + 310 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 265, + 287, + 310 + ], + "spans": [ + { + "bbox": [ + 53, + 265, + 287, + 310 + ], + "type": "text", + "content": "[5] Qingyang Zhang, Haitao Wu, Changqing Zhang, Qinghua Hu, Huazhu Fu, Joey Tianyi Zhou, and Xi Peng. Provable dynamic fusion for low-quality multimodal data. arXiv preprint arXiv:2306.02050, 2023. 1, 2, 4, 5, 6, 7, 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 312, + 287, + 355 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 312, + 287, + 355 + ], + "spans": [ + { + "bbox": [ + 53, + 312, + 287, + 355 + ], + "type": "text", + "content": "[6] Ronghao Lin and Haifeng Hu. Dynamically shifting multimodal representations via hybrid-modal attention for multimodal sentiment analysis. IEEE Transactions on Multimedia, 2023. 1, 5, 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 357, + 287, + 401 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 357, + 287, + 401 + ], + "spans": [ + { + "bbox": [ + 53, + 357, + 287, + 401 + ], + "type": "text", + "content": "[7] Ying Zeng, Sijie Mai, Wenjun Yan, and Haifeng Hu. Multimodal reaction: Information modulation for cross-modal representation learning. IEEE Transactions on Multimedia, 2023. 1, 5, 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 404, + 287, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 404, + 287, + 459 + ], + "spans": [ + { + "bbox": [ + 53, + 404, + 287, + 459 + ], + "type": "text", + "content": "[8] Jiuding Yang, Yakun Yu, Di Niu, Weidong Guo, and Yu Xu. Confede: Contrastive feature decomposition for multimodal sentiment analysis. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics, pages 7617-7630, 2023. 1, 5, 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 461, + 287, + 506 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 461, + 287, + 506 + ], + "spans": [ + { + "bbox": [ + 53, + 461, + 287, + 506 + ], + "type": "text", + "content": "[9] Yong Li, Yuanzhi Wang, and Zhen Cui. Decoupled multimodal distilling for emotion recognition. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 6631-6640, 2023. 1, 5, 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 508, + 287, + 552 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 508, + 287, + 552 + ], + "spans": [ + { + "bbox": [ + 48, + 508, + 287, + 552 + ], + "type": "text", + "content": "[10] Weiyao Wang, Du Tran, and Matt Feiszli. What makes training multi-modal classification networks hard? In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 12695-12705, 2020. 1, 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 555, + 287, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 555, + 287, + 609 + ], + "spans": [ + { + "bbox": [ + 48, + 555, + 287, + 609 + ], + "type": "text", + "content": "[11] Xiaokang Peng, Yake Wei, Andong Deng, Dong Wang, and Di Hu. Balanced multimodal learning via on-the-fly gradient modulation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 8238-8247, 2022. 1, 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 612, + 287, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 612, + 287, + 667 + ], + "spans": [ + { + "bbox": [ + 48, + 612, + 287, + 667 + ], + "type": "text", + "content": "[12] Yu Huang, Junyang Lin, Chang Zhou, Hongxia Yang, and Longbo Huang. Modality competition: What makes joint training of multi-modal network fail in deep learning? (provably). In International Conference on Machine Learning, pages 9226-9259. PMLR, 2022. 1, 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 670, + 287, + 703 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 670, + 287, + 703 + ], + "spans": [ + { + "bbox": [ + 48, + 670, + 287, + 703 + ], + "type": "text", + "content": "[13] Yunfeng Fan, Wenchao Xu, Haozhao Wang, Junxiao Wang, and Song Guo. PMR: prototypical modal rebalance for multimodal learning. In Proceedings of the IEEE Conference" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 545, + 691 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "type": "text", + "content": "on Computer Vision and Pattern Recognition, pages 20029-20038. IEEE, 2023. 1, 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 99, + 545, + 133 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 99, + 545, + 133 + ], + "spans": [ + { + "bbox": [ + 308, + 99, + 545, + 133 + ], + "type": "text", + "content": "[14] Alex Kendall and Yarin Gal. What uncertainties do we need in bayesian deep learning for computer vision? Advances in neural information processing systems, 30, 2017. 2, 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 136, + 545, + 179 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 136, + 545, + 179 + ], + "spans": [ + { + "bbox": [ + 308, + 136, + 545, + 179 + ], + "type": "text", + "content": "[15] Jie Chang, Zhonghao Lan, Changmao Cheng, and Yichen Wei. Data uncertainty learning in face recognition. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 5710-5719, 2020. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 183, + 545, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 183, + 545, + 217 + ], + "spans": [ + { + "bbox": [ + 308, + 183, + 545, + 217 + ], + "type": "text", + "content": "[16] Yichun Shi and Anil K Jain. Probabilistic face embeddings. In Proceedings of the IEEE International Conference on Computer Vision, pages 6902-6911, 2019. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 220, + 545, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 220, + 545, + 285 + ], + "spans": [ + { + "bbox": [ + 308, + 220, + 545, + 285 + ], + "type": "text", + "content": "[17] Yatai Ji, Junjie Wang, Yuan Gong, Lin Zhang, Yanru Zhu, Hongfa Wang, Jiaxing Zhang, Tetsuya Sakai, and Yujiu Yang. Map: Multimodal uncertainty-aware vision-language pre-training model. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 23262-23271, 2023. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 289, + 545, + 322 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 289, + 545, + 322 + ], + "spans": [ + { + "bbox": [ + 308, + 289, + 545, + 322 + ], + "type": "text", + "content": "[18] Luke Vilnis and Andrew McCallum. Word representations via gaussian embedding. arXiv preprint arXiv:1412.6623, 2014. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 327, + 545, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 327, + 545, + 360 + ], + "spans": [ + { + "bbox": [ + 308, + 327, + 545, + 360 + ], + "type": "text", + "content": "[19] Naftali Tishby, Fernando C Pereira, and William Bialek. The information bottleneck method. arXiv preprint physics/0004057, 2000. 2, 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 364, + 545, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 364, + 545, + 397 + ], + "spans": [ + { + "bbox": [ + 308, + 364, + 545, + 397 + ], + "type": "text", + "content": "[20] Alexander A Alemi, Ian Fischer, Joshua V Dillon, and Kevin Murphy. Deep variational information bottleneck. arXiv preprint arXiv:1612.00410, 2016. 2, 3, 5" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 400, + 545, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 400, + 545, + 456 + ], + "spans": [ + { + "bbox": [ + 308, + 400, + 545, + 456 + ], + "type": "text", + "content": "[21] Xun Jiang, Xing Xu, Zhiguo Chen, Jingran Zhang, Jingkuan Song, Fumin Shen, Huimin Lu, and Heng Tao Shen. DHHN: dual hierarchical hybrid network for weakly-supervised audio-visual video parsing. In ACM International Conference on Multimedia, pages 719–727, 2022. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 460, + 545, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 460, + 545, + 514 + ], + "spans": [ + { + "bbox": [ + 308, + 460, + 545, + 514 + ], + "type": "text", + "content": "[22] Junyu Gao, Mengyuan Chen, and Changsheng Xu. Collecting cross-modal presence-absence evidence for weakly-supervised audio-visual event perception. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 18827–18836, 2023. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 518, + 545, + 572 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 518, + 545, + 572 + ], + "spans": [ + { + "bbox": [ + 308, + 518, + 545, + 572 + ], + "type": "text", + "content": "[23] Zhiguo Chen, Xun Jiang, Xing Xu, Zuo Cao, Yijun Mo, and Heng Tao Shen. Joint searching and grounding: Multigranularity video content retrieval. In Proceedings of the 31st ACM International Conference on Multimedia, pages 975-983, 2023. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 577, + 545, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 577, + 545, + 632 + ], + "spans": [ + { + "bbox": [ + 308, + 577, + 545, + 632 + ], + "type": "text", + "content": "[24] Shenshen Li, Xing Xu, Xun Jiang, Fumin Shen, Xin Liu, and Heng Tao Shen. Multi-grained attention network with mutual exclusion for composed query-based image retrieval. IEEE Transactions on Circuits and Systems for Video Technology, 2023. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 635, + 545, + 691 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 635, + 545, + 691 + ], + "spans": [ + { + "bbox": [ + 308, + 635, + 545, + 691 + ], + "type": "text", + "content": "[25] Shenshen Li, Xing Xu, Xun Jiang, Fumin Shen, Zhe Sun, and Andrzej Cichocki. Cross-modal attention preservation with self-contrastive learning for composed query-based image retrieval. ACM Transactions on Multimedia Computing, Communications, and Applications, 20(6), 2024. 2" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "26884" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 696 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "text", + "content": "[26] Xing Xu, Kaiyi Lin, Yang Yang, Alan Hanjalic, and Heng Tao Shen. Joint feature synthesis and embedding: Adversarial cross-modal retrieval revisited. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(6):3030-3047, 2022. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 132, + 287, + 186 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 132, + 287, + 186 + ], + "spans": [ + { + "bbox": [ + 48, + 132, + 287, + 186 + ], + "type": "text", + "content": "[27] Shenshen Li, Xing Xu, Xun Jiang, Fumin Shen, Xin Liu, and Heng Tao Shen. Multi-grained attention network with mutual exclusion for composed query-based image retrieval. IEEE Transactions on Circuits and Systems for Video Technology, pages 1-1, 2023. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 190, + 287, + 235 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 190, + 287, + 235 + ], + "spans": [ + { + "bbox": [ + 48, + 190, + 287, + 235 + ], + "type": "text", + "content": "[28] Sijie Mai, Ying Zeng, and Haifeng Hu. Multimodal information bottleneck: Learning minimal sufficient unimodal and multimodal representations. IEEE Transactions on Multimedia, 2022. 2, 3, 5, 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 238, + 287, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 238, + 287, + 282 + ], + "spans": [ + { + "bbox": [ + 48, + 238, + 287, + 282 + ], + "type": "text", + "content": "[29] Xun Jiang, Xing Xu, Jingran Zhang, Fumin Shen, Zuo Cao, and Heng Tao Shen. Sdn: Semantic decoupling network for temporal language grounding. IEEE Transactions on Neural Networks and Learning Systems, pages 1-15, 2022. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 286, + 287, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 286, + 287, + 341 + ], + "spans": [ + { + "bbox": [ + 48, + 286, + 287, + 341 + ], + "type": "text", + "content": "[30] Xun Jiang, Xing Xu, Jingran Zhang, Fumin Shen, Zuo Cao, and Heng Tao Shen. Semi-supervised video paragraph grounding with contrastive encoder. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 2456-2465. IEEE, 2022. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 345, + 287, + 388 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 345, + 287, + 388 + ], + "spans": [ + { + "bbox": [ + 48, + 345, + 287, + 388 + ], + "type": "text", + "content": "[31] Arsha Nagrani, Shan Yang, Anurag Arnab, Aren Jansen, Cordelia Schmid, and Chen Sun. Attention bottlenecks for multimodal fusion. In Advances in Neural Information Processing Systems, pages 14200-14213, 2021. 2, 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 393, + 287, + 448 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 393, + 287, + 448 + ], + "spans": [ + { + "bbox": [ + 48, + 393, + 287, + 448 + ], + "type": "text", + "content": "[32] Guanglei Yang, Enrico Fini, Dan Xu, Paolo Rota, Mingli Ding, Moin Nabi, Xavier Alameda-Pineda, and Elisa Ricci. Uncertainty-aware contrastive distillation for incremental semantic segmentation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(2):2567-2581, 2022. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 452, + 287, + 496 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 452, + 287, + 496 + ], + "spans": [ + { + "bbox": [ + 48, + 452, + 287, + 496 + ], + "type": "text", + "content": "[33] Mengyuan Chen, Junyu Gao, Shicai Yang, and Changsheng Xu. Dual-evidential learning for weakly-supervised temporal action localization. In European Conference on Computer Vision, pages 192-208. Springer, 2022. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 499, + 287, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 499, + 287, + 553 + ], + "spans": [ + { + "bbox": [ + 48, + 499, + 287, + 553 + ], + "type": "text", + "content": "[34] Yifei Huang, Lijin Yang, and Yoichi Sato. Weakly supervised temporal sentence grounding with uncertainty-guided self-training. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 18908-18918, 2023. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 558, + 287, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 558, + 287, + 613 + ], + "spans": [ + { + "bbox": [ + 48, + 558, + 287, + 613 + ], + "type": "text", + "content": "[35] Shenshen Li, Xing Xu, Yang Yang, Fumin Shen, Yijun Mo, Yujie Li, and Heng Tao Shen. DCEL: deep cross-modal evidential learning for text-based person retrieval. In Proceedings of the 31st ACM International Conference on Multimedia, pages 6292-6300, 2023. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 617, + 287, + 650 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 617, + 287, + 650 + ], + "spans": [ + { + "bbox": [ + 48, + 617, + 287, + 650 + ], + "type": "text", + "content": "[36] Alexander A Alemi, Ian Fischer, Joshua V Dillon, and Kevin Murphy. Deep variational information bottleneck. arXiv preprint arXiv:1612.00410, 2016. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 654, + 287, + 696 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 654, + 287, + 696 + ], + "spans": [ + { + "bbox": [ + 48, + 654, + 287, + 696 + ], + "type": "text", + "content": "[37] Yuchen Shen, Dong Zhang, Zhihao Song, Xuesong Jiang, and Qiaolin Ye. Learning to reduce information bottleneck for object detection in aerial images. IEEE Geoscience and Remote Sensing Letters, 2023. 3" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 700 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "text", + "content": "[38] Among Wu and Cheng Deng. Tib: Detecting unknown objects via two-stream information bottleneck. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 307, + 110, + 545, + 165 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 110, + 545, + 165 + ], + "spans": [ + { + "bbox": [ + 307, + 110, + 545, + 165 + ], + "type": "text", + "content": "[39] Sungtae An, Nataraj Jammalamadaka, and Eunji Chong. Maximum entropy information bottleneck for uncertainty-aware stochastic embedding. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 3808-3817, 2023. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 169, + 545, + 190 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 169, + 545, + 190 + ], + "spans": [ + { + "bbox": [ + 307, + 169, + 545, + 190 + ], + "type": "text", + "content": "[40] Diederik P Kingma and Max Welling. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114, 2013. 4, 5" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 194, + 545, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 194, + 545, + 239 + ], + "spans": [ + { + "bbox": [ + 307, + 194, + 545, + 239 + ], + "type": "text", + "content": "[41] Amir Zadeh, Rowan Zellers, Eli Pincus, and Louis-Philippe Morency. Mosi: multimodal corpus of sentiment intensity and subjectivity analysis in online opinion videos. arXiv preprint arXiv:1606.06259, 2016. 5" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 243, + 545, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 243, + 545, + 308 + ], + "spans": [ + { + "bbox": [ + 307, + 243, + 545, + 308 + ], + "type": "text", + "content": "[42] AmirAli Bagher Zadeh, Paul Pu Liang, Soujanya Poria, Erik Cambria, and Louis-Philippe Morency. Multimodal language analysis in the wild: Cmu-mosei dataset and interpretable dynamic fusion graph. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics, pages 2236–2246, 2018. 5" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 312, + 545, + 355 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 312, + 545, + 355 + ], + "spans": [ + { + "bbox": [ + 307, + 312, + 545, + 355 + ], + "type": "text", + "content": "[43] Teng Niu, Shiai Zhu, Lei Pang, and Abdulmotaleb El Saddik. Sentiment analysis on multi-view social data. In International Conference on Multimedia Modeling, pages 15-27, 2016. 5" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 360, + 545, + 404 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 360, + 545, + 404 + ], + "spans": [ + { + "bbox": [ + 307, + 360, + 545, + 404 + ], + "type": "text", + "content": "[44] Nathan Silberman, Derek Hoiem, Pushmeet Kohli, and Rob Fergus. Indoor segmentation and support inference from rgbd images. In European Conference on Computer Vision, pages 746-760, 2012. 5" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 408, + 545, + 452 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 408, + 545, + 452 + ], + "spans": [ + { + "bbox": [ + 307, + 408, + 545, + 452 + ], + "type": "text", + "content": "[45] Yuanzhi Wang, Zhen Cui, and Yong Li. Distribution-consistent modal recovering for incomplete multimodal learning. In Proceedings of the IEEE International Conference on Computer Vision, pages 22025-22034, 2023. 5, 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 456, + 545, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 456, + 545, + 499 + ], + "spans": [ + { + "bbox": [ + 307, + 456, + 545, + 499 + ], + "type": "text", + "content": "[46] Zheng Lian, Lan Chen, Licai Sun, Bin Liu, and Jianhua Tao. Gcnet: graph completion network for incomplete multimodal learning in conversation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023. 5, 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 503, + 545, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 503, + 545, + 547 + ], + "spans": [ + { + "bbox": [ + 307, + 503, + 545, + 547 + ], + "type": "text", + "content": "[47] Tong Zhu, Leida Li, Jufeng Yang, Sicheng Zhao, Hantao Liu, and Jiansheng Qian. Multimodal sentiment analysis with image-text interaction network. IEEE Transactions on Multimedia, 2022. 5, 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 552, + 545, + 617 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 552, + 545, + 617 + ], + "spans": [ + { + "bbox": [ + 307, + 552, + 545, + 617 + ], + "type": "text", + "content": "[48] Yiwei Wei, Shaozu Yuan, Ruosong Yang, Lei Shen, Zhangmeizhi Li, Longbiao Wang, and Meng Chen. Tackling modality heterogeneity with multi-view calibration network for multimodal sentiment detection. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics, pages 5240-5252, 2023. 5, 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 621, + 545, + 664 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 621, + 545, + 664 + ], + "spans": [ + { + "bbox": [ + 307, + 621, + 545, + 664 + ], + "type": "text", + "content": "[49] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018. 7" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 669, + 545, + 700 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 669, + 545, + 700 + ], + "spans": [ + { + "bbox": [ + 307, + 669, + 545, + 700 + ], + "type": "text", + "content": "[50] Laurens Van der Maaten and Geoffrey Hinton. Visualizing data using t-sne. Journal of machine learning research, 9(11), 2008. 8" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "26885" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Emergent Open-Vocabulary Semantic Segmentation from Off-the-shelf Vision-Language Models/bf254b56-9ba7-4039-9f52-56365463bb00_content_list.json b/2024/Emergent Open-Vocabulary Semantic Segmentation from Off-the-shelf Vision-Language Models/bf254b56-9ba7-4039-9f52-56365463bb00_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..cf68da05006c79f07ddb583b6c1481b289ea7eac --- /dev/null +++ b/2024/Emergent Open-Vocabulary Semantic Segmentation from Off-the-shelf Vision-Language Models/bf254b56-9ba7-4039-9f52-56365463bb00_content_list.json @@ -0,0 +1,1947 @@ +[ + { + "type": "text", + "text": "Emergent Open-Vocabulary Semantic Segmentation from Off-the-shelf Vision-Language Models", + "text_level": 1, + "bbox": [ + 127, + 130, + 843, + 174 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Jiayun Luo $^{\\flat}$ , Siddhesh Khandelwal $^{\\sharp}$ , Leonid Sigal $^{\\sharp}$ , and Boyang Li $^{\\flat}$", + "bbox": [ + 212, + 202, + 750, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{b}$ Nanyang Technological University, Singapore", + "bbox": [ + 297, + 222, + 669, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{\\ddagger}$ University of British Columbia, Vector Institute for AI, Canada", + "bbox": [ + 230, + 239, + 738, + 256 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{luoj0028,boyang.li}@ntu.edu.sg，{skhandel,lsignal}@cs.ubc.ca", + "bbox": [ + 210, + 260, + 754, + 273 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/1ff679333025df281c2c264dce0007ffa1e9b4392c0c8e79e9805419be2db86c.jpg", + "image_caption": [ + "Figure 1. Qualitative Results of PnP-OVSS + BLIP. Images are from Pascal Context and COCO stuff. The right columns and bottom rows show the ground-truth (GT); the rest are our results. Note accurate results even on complex (trees) and small objects (last column)." + ], + "image_footnote": [], + "bbox": [ + 81, + 310, + 207, + 444 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/90f2b88f1e18d0f77460141d9566fdb4af766fb5e89e26f13e48fad8f438a73e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 209, + 310, + 334, + 444 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/58858b884765120250fddbe6138fde61d893d84686039a0ddc10e510760a8523.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 336, + 310, + 491, + 444 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/bac028cc5228ca6a1d842d476f8201b7e0a41dcc5aae3f16c20b371a4c03b5d2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 493, + 309, + 643, + 444 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/3c969cb703bd95a7da9929f65b58dca7a5b2f2d6b2fcccff5b2c7e717b2ce8bf.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 645, + 309, + 754, + 444 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/7c6356ae1fb78a8926503660df0845ab899375d8b32c5d739407f4ec2c6c41a2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 759, + 310, + 890, + 444 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 489, + 313, + 505 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "From image-text pairs, large-scale vision-language models (VLMs) learn to implicitly associate image regions with words, which prove effective for tasks like visual question answering. However, leveraging the learned association for open-vocabulary semantic segmentation remains a challenge. In this paper, we propose a simple, yet extremely effective, training-free technique, Plug-and-Play Open-Vocabulary Semantic Segmentation (PnP-OVSS) for this task. PnP-OVSS leverages a VLM with direct text-to-image cross-attention and an image-text matching loss. To balance between over-segmentation and under-segmentation, we introduce Salience Dropout; by iteratively dropping patches that the model is most attentive to, we are able to better resolve the entire extent of the segmentation mask. PnP-OVSS does not require any neural network training and performs hyperparameter tuning without the need for any segmentation annotations, even for a validation set. PnP-OVSS demonstrates substantial improvements over comparable baselines (+29.4% mIoU on Pascal VOC, +13.2% mIoU on Pascal Context, +14.0% mIoU on MS COCO, +2.4% mIoU on COCO Stuff) and even outperforms most baselines that conduct additional network training on top of pretrained VLMs. Our codebase is at https://github.com/letitiabana/PnP-OVSS.", + "bbox": [ + 73, + 521, + 472, + 883 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 501, + 489, + 630, + 505 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The classic task of semantic segmentation [20, 22] aims to classify pixels to their object types. Traditional supervised methods require dense pixel-level annotations and are restricted to recognizing a predefined set of objects. To relax these constraints, open-vocabulary semantic segmentation [8, 17, 21, 41, 44, 46, 48, 74, 76, 77, 80, 84] aspires to identify arbitrary object categories, whereas weakly supervised techniques [9, 11, 15, 25, 31, 36, 53, 55, 65, 75, 90] can acquire pixel-level localization capabilities from coarse supervision, e.g., image labels or boxes.", + "bbox": [ + 496, + 517, + 890, + 667 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Large-scale vision-language models (VLMs) pretrained on image-text pairs [1, 37, 38, 40, 60, 78, 85] achieve unprecedented performance on multimodal tasks, such as describing arbitrary images and answering free-form, open-ended questions about them (either with [37, 38, 60] or without finetuning [1, 19, 62, 70]). These tasks apparently involve some ability to localize objects. For example, to answer the question \"what objects appear on the table?\", the model would have to first localize the table in the image and identify the objects on it. Hence, it is reasonable to conjecture that the VLM network learns to perform open-vocabulary localization from image-text pretraining. However, distilling the localization capability from the VLMs remains an open challenge.", + "bbox": [ + 496, + 671, + 892, + 883 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Most existing methods for open-vocabulary semantic", + "bbox": [ + 517, + 885, + 890, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 44 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "4029", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/121ad0ade58b246149c5186512c220dafbf9ccfde40adf06e04955f94076089f.jpg", + "image_caption": [ + "(a) Original image" + ], + "image_footnote": [], + "bbox": [ + 80, + 88, + 176, + 238 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/0cd3e661d89d18175c63de5f8e1f53b307d27537740551d5d8b239eea3fb8172.jpg", + "image_caption": [ + "(b) Cross-attention", + "Figure 2. Segmentation masks for elephant and train using (a) off-the-shelf cross-attention, (b) cross-attention + GradCAM, and (c) cross-attention + GradCam + Salience DropOut (\\$3.2). The naive cross-attention masks are too inclusive whereas GradCAM is too exclusive." + ], + "image_footnote": [], + "bbox": [ + 176, + 88, + 272, + 239 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/389d49edd686ef6e886871752288378ae63044a0c67c4a3cf04429552cd44238.jpg", + "image_caption": [ + "(c) GradCAM" + ], + "image_footnote": [], + "bbox": [ + 274, + 88, + 370, + 239 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/6b6386f24e5d894b02003cba93d289f6a972b1d99c05955596708b16cc7751a1.jpg", + "image_caption": [ + "(d) Salience Dropout" + ], + "image_footnote": [], + "bbox": [ + 370, + 88, + 470, + 239 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "segmentation (OVSS) from VLMs usually obtain single vector encodings for the visual and text inputs respectively [43, 44, 51, 52, 73, 74]. However, pooling every token into a single vector likely discards information about detailed positions of objects and words. We investigate the use of pretrained cross-attention layers for OVSS, which retain finer-grained correspondence between text and image patches.", + "bbox": [ + 75, + 364, + 468, + 470 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Nevertheless, a naive application of cross-attention between the class name and the image patches lead to overly broad segmentation masks that include irrelevant parts of the image (i.e., over-segmentation, see Fig. 2(b)). To alleviate this, [62] employs gradient information from the imagetext matching loss, GradCAM[57]-style, to sharpen the attention-based masks for the purpose of guiding caption generation. However, this results in masks that capture only the most discriminative regions of an object, such as the head of an elephant (i.e., under-segmentation, Fig. 2(c)). To acquire complete object masks, we propose Salience DropOut, which iteratively drops the image patches with high GradCAM attention scores, forcing the model to attend to less discriminative but relevant object parts.", + "bbox": [ + 75, + 474, + 470, + 686 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Another important consideration is the cross-attention layer and the attention head to extract the masks from. These hyperparameters have enormous influence on the final results and are traditionally tuned on a validation set with pixel-level mask annotations. To eliminate the need for dense annotations, we propose an weakly-supervised reward function based on CLIP [49]. On a validation set with images as well as object class names, the technique contrasts the extracted object regions with a blank image. If, according to CLIP, the former is more similar to the corresponding class name than the latter, we increment the reward. All hyperparameter tuning of our technique is performed with a simple random search with this reward, leading to high performance.", + "bbox": [ + 75, + 689, + 468, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In summary, we propose Plug-and-Play Open-vocabulary Semantic Segmentation (PnP-OVSS), an extremely simple and training-free framework to extract semantic segmentations from VLMs. At zero extra training cost, PnP-OVSS can be used with any pretrained VLM with text-to-image cross attention layer and an image-text matching loss. It has zero reliance on pixel-level annotations, including a validation set for hyperparameter tuning. At the same time, PnP-OVSS delivers excellent performance. It not only beats the training-free baseline with remarkable margins $(+29.4\\%)$ mIoU on Pascal VOC, $+13.2\\%$ on Pascal Context, $+14.0\\%$ on MS COCO, $+2.4\\%$ on COCO Stuff), but also outperforms most recent techniques within the past two years that require extensive finetuning on top of the VLM pretraining.", + "bbox": [ + 496, + 90, + 893, + 301 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "With this paper, we make three contributions:", + "bbox": [ + 517, + 303, + 821, + 318 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose to combine text-to-image attention, GradCAM, and Salience DropOut to iteratively acquire accurate segmentation of arbitrary classes from pretrained VLM.", + "- We replace the densely annotated validation set for hyperparameter tuning, which is needed by most existing methods, with a contrastive reward function based on CLIP. This reward function, coupled with random search, finds a good set of hyperparameters for OVSS.", + "- The proposed method, PnP-OVSS, is simple to use, requires no extra finetuning, and delivers high performance. Its success hints at a new direction for open-vocabulary segmentation tasks leveraging large VLMs." + ], + "bbox": [ + 500, + 319, + 900, + 501 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 517, + 640, + 532 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Large-Scale Vision-Language Model", + "text_level": 1, + "bbox": [ + 500, + 542, + 815, + 559 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Large-scale vision-language models (VLMs), trained on millions of image-text pairs, have become the foundation for many multimodal tasks. Architecturally, the straightforward approach to training such methods involves aligning visual and textual latent representations via a simple dot product [24, 49]. However, this is insufficient for complex structured tasks like visual question answering or image captioning, which require specialized approaches that employ separate encoders before cross-attention between modalities [33, 35, 37, 78, 79] or self-attention networks over all tokens from both modalities [10, 27, 39, 40, 60]. Another design dimension when training VLMs is the loss function. Commonly used losses include image-text contrastive learning [34, 35, 37, 50, 60, 81], image-text matching (ITM) [27, 34, 35, 37, 60, 72, 78, 81, 82], prediction of masked tokens or patches [34, 60, 78, 81], and language modeling [1, 7, 34, 37].", + "bbox": [ + 496, + 566, + 893, + 824 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "This work utilizes models with unimodal encoders followed by cross-attention fusion [34, 35, 37, 72, 78, 81, 82], as they work with high-level features, and accurately attend to the appropriate image patches. Additionally, we utilize the gradient from the ITM loss in the GradCAM step (§3.1)", + "bbox": [ + 496, + 825, + 893, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "4030", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "table", + "img_path": "images/cc54eecb1e6ce837da52ebc603041c25c6481b63abd560afabe7988417314cd5.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodPT Data SizeFT data size
Requires finetuning on image-text pairs
OVSegmentor [74]-4.3M
Vil-Seg [43]400M412M
GroupVit* [73]-3.4M
GroupVit [73]-26M
CLIPpy [51]-134M
SegClip [44]400M3.4M
ViewCo [52]-26M
TCL [6]400M15M
PACL [46]400M30M
Requires finetuning but not image-text pairs
MaskClip w/ ST [89]400M1.2M
ZeroSeg* [8]400M3.4M
ZeroSeg [8]400M1.2M
Requires no finetuning
MaskClip[89]400M0
Reco[58]400M0
PnP-OVSS (Ours)
+ BLIP129M0
+ BridgeTower400M+ 4M0
", + "bbox": [ + 114, + 88, + 433, + 409 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Table 1. Training data of current Zero-shot semantic segmentation methods with only text supervision. PT stands for pretraining with image-caption data, FT stands for fine-tuning with image-caption data. ST stands for self training. We list only the image-caption data used for pretraining and all type of data for finetuning. For hyperparameter tuning, our method uses CLIP-L/14 pretrained with 400M data to calculate the reward.", + "bbox": [ + 75, + 417, + 470, + 518 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "to sharpen segmentation masks.", + "bbox": [ + 76, + 545, + 290, + 560 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. Zero-shot Semantic Segmentation", + "text_level": 1, + "bbox": [ + 76, + 571, + 372, + 588 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Zero-shot semantic segmentation predicts a dense segmentation mask for any object class described by a given text prompt, with only prior exposure to class-agnostic image-level supervision. This contrasts with weakly supervised semantic segmentation [3, 13, 23, 26, 28, 30, 54, 61, 69, 69, 83, 86, 87], which relies on class-specific annotations, and unsupervised object discovery [12, 14, 56, 59, 63, 64, 67, 68], which identifies the sole object in the foreground.", + "bbox": [ + 75, + 595, + 468, + 717 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Traditional methods [4, 18, 47, 71] train a classifier to distinguish between seen and unseen visual features, wherein the unseen visual features are obtained from a generative model, trained on pairs of seen class image-text embeddings. Recently, methods additionally leverage knowledge from VLMs to attain better matching of visual and textual features [17, 32, 41, 76, 89]. Concretely, they train the segmentation network with dense annotations, while replacing a part of the framework with components from VLMs.", + "bbox": [ + 75, + 718, + 468, + 854 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Recently, methods have explored the use of additional supervision to further reduce the need for pixel level annotations. Prior work is compared in Table 1.", + "bbox": [ + 75, + 854, + 468, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Models Finetuned on Image-Text Pairs. Methods proposed in [6, 43, 44, 46, 51, 52, 73, 74] require paired image-text annotations to adapt to the zero-shot segmentation task. Specifically, approaches in [43, 44, 51, 52, 73, 74] cluster semantically similar pixels by contrasting grouped region embeddings with textual embeddings, usually via the use of contrastive and self-supervised losses. PACL [46] modifies the constrastive loss to operate over aggregated patch embeddings (instead of image embeddings) to encourage better alignment between image patches and text. TCL [6] achieves a similar patch-text alignment by introducing an additional module to extract text-grounded image regions.", + "bbox": [ + 496, + 90, + 890, + 272 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Models Finetuned on Pseudo-labels. Methods in [8, 89] do not require additional image-text supervision, but involve additional fine-tuning to adapt to the segmentation task. ZeroSeg [8] attempts to match its own group embeddings with multi-scaled segment embeddings obtained from CLIP [49]. MaskClip w/ ST [89] modifies the global attention pooling layer within CLIP to output segmentation masks, which are used as pseudo-labels to train a segmentation network.", + "bbox": [ + 496, + 277, + 890, + 412 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In comparison, our proposed PnP-OVSS does not require any fine-tuning or additional paired image-text annotations. It can directly distill high quality open vocabulary semantic segmentations from any VLM with direct text-to-image cross-attention and an image-text matching loss [34, 35, 37, 72, 78, 81, 82]. Compared to approaches in [58, 89] that perform zero-shot open vocabulary semantic segmentation under a similar no training and no additional annotations paradigm, PnP-OVSS achieves considerably superior performance.", + "bbox": [ + 496, + 415, + 890, + 566 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 500, + 584, + 589, + 599 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "PnP-OVSS has four major steps. First, we extract a cross-attention salience map per object class from a VLM. Second, we sharpen the salience map by weighing it with the ITM gradient in the style of GradCAM. Third, we apply Salience DropOut that iteratively completes the salience maps. Fourth, we apply Dense CRF [29] for fine-grained adjustment. The process is illustrated in Fig. 3 and Fig. 4.", + "bbox": [ + 496, + 609, + 890, + 717 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In the first step, we feed an image and a text prompt to the pretrained VLM and extract cross-attention maps. The text prompt is \"A picture of [class 1] [class 2] ... [class K]\", which includes all $K$ class names of interest in the dataset. The image is divided into $P \\times P$ patches. The text prompt and the image first go through the modality-specific encoders respectively, followed by a cross-attention fusion module. In the cross-attention layers, the text encodings serve as query vectors and the image patche encodings are as key and value vectors.", + "bbox": [ + 496, + 718, + 890, + 869 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We extract an attention map for each text token to the image patches from a particular cross-attention layer and an", + "bbox": [ + 500, + 869, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "4031", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/c09ee7dc9b163aecd63adce78dc64440bff4ac145ac5613a803270cd95bde287.jpg", + "image_caption": [ + "Figure 3. The first iteration with cross-attention + GradCAM + Salience DropOut. The text prompt contains $K$ class names and the image contains $P \\times P$ patches. From a cross-attention layer and an attention head in the pretrained VLM, we obtain $K$ attention score maps of size $P \\times P$ , which are sharpened by GradCAM using gradients from the image-text-matching (ITM) loss. To get more complete predictions, we perform Salience Dropout, which repeatedly zero out image patches of the highest average scores and feeds the remaining patches to the image encoder again, forcing the model to attend to other less discriminative patches. We show example salience maps from all iterations in Fig. 4." + ], + "image_footnote": [], + "bbox": [ + 101, + 88, + 880, + 397 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "attention head. Note that different layers and heads lead to drastic performance differences, and the choices are hyperparameters, tuned using the procedure in §3.4. We exclude attention maps for the first three tokens \"A\", \"picture\", \"of\", which do not describe semantic classes. If a class name contains two or more tokens, we take the mean attention map. This procedure yields an attention tensor of size $K \\times P \\times P$ .", + "bbox": [ + 75, + 522, + 468, + 627 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "With the correct layer and attention head, we observe that this attention map, when normalized by softmax along the first dimension $K$ , can provide passable semantic segmentations. However, it tends to include many patches unrelated to the class being segmented, leading to over-segmentation. Hence, we introduce two refinement steps, GradCAM and Salience DropOut, explained in §3.1 and §3.2. After these steps, we acquire aggregate salience maps for every class. Finally, we conduct local polish to the resultant salience maps using Dense CRF, as described in §3.3.", + "bbox": [ + 75, + 628, + 470, + 779 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1. Map Sharpening with GradCAM", + "text_level": 1, + "bbox": [ + 76, + 786, + 369, + 801 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The off-the-shelf attention maps tend to also cover many patches unrelated to the class name (See Fig. 2 (b)). Prior works [35, 62] further sharpen the attention maps and focus them on class-discriminative regions using a variant of GradCAM [57], which is originally proposed for convolutional networks, but applicable to attention maps. It is worth", + "bbox": [ + 75, + 810, + 470, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "noting that [35, 62] use the technique for purposes other than semantic segmentation.", + "bbox": [ + 496, + 522, + 890, + 551 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The GradCAM method requires a gradient. Here we leverage the image-text matching (ITM) loss, which trains the VLM to classify if an image-text pair match each other or not. Computing the ITM loss requires a label. We use \"matching\" (as opposed to \"not matching\") as the label and compute the gradient of the loss with respect to the attention score. This is equivalent to asking: which attention scores contribute the most to the decision that the image-text pair is matching? Formally, we denote a $P \\times P$ attention map for class $k$ as $M^{(k)}$ and the ITM loss as $\\mathcal{L}_{\\mathrm{ITM}}$ . The GradCAM class salience map is computed as", + "bbox": [ + 496, + 553, + 892, + 719 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\widetilde {M} ^ {(k)} = \\max \\left(0, \\frac {\\partial \\mathcal {L} _ {\\mathrm {I T M}}}{\\partial M ^ {(k)}}\\right) \\otimes M ^ {(k)}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 573, + 729, + 890, + 763 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\otimes$ denote the component-wise multiplication and $\\max (\\cdot)$ is also applied component-wise.", + "bbox": [ + 496, + 775, + 890, + 806 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Salience Dropout", + "text_level": 1, + "bbox": [ + 500, + 816, + 671, + 832 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As illustrated in Fig. 2 (c), segmentations generated by the GradCAM-style re-weighting of cross-attention are often narrowly focused on the most discriminative regions for a given class. However, the less discriminative regions are", + "bbox": [ + 496, + 839, + 892, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4032", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/d3a733835318b2e7e0db8bcce74e07c5564f79f60a3435f7f61dc0c279b0cea1.jpg", + "image_caption": [ + "GradCAM Class Salience Map" + ], + "image_footnote": [], + "bbox": [ + 83, + 109, + 163, + 167 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/32656e85b9ec5cf0849b5761447164fe7ad2eebf1d1ac65da6597c7dd2442342.jpg", + "image_caption": [ + "$1^{\\mathrm{st}}$ Salience Drop" + ], + "image_footnote": [], + "bbox": [ + 174, + 109, + 259, + 167 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/744a6862317d00d58f745171488df677d32103b6c246ac1b7cc3214e7d9ebd8f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 269, + 109, + 354, + 167 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/73bb98368c603e8ff1de7b26f553ba5d08e4a8b3b06f34ef34f3eedfc247b85f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 362, + 109, + 449, + 167 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/238cdefc313455f4837881828fad3a25e3f4d543b91cad0b118a075cad197d40.jpg", + "image_caption": [ + "$2^{\\text{nd}}$ Salience Drop" + ], + "image_footnote": [], + "bbox": [ + 81, + 188, + 166, + 247 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/74884e061dfabf8267a7181b21c527458224149a23deacfee289b1e7596c6ea4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 179, + 188, + 264, + 246 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/7322f1594217095b3fdd8220e094f3ac5ff1fe9df76a83c16e1654366d684284.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 269, + 188, + 354, + 246 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/edfb58038c4d94ee7569afbb4dadbdf199a4daf88bfeefc766b8b51160d8e613.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 362, + 188, + 447, + 246 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/d3f610134af2c85a76beac4c6dc4a4159c4171cb2ba8d7495325d7cdad7ed4c9.jpg", + "image_caption": [ + "$3^{\\text{rd}}$ Salience Drop" + ], + "image_footnote": [], + "bbox": [ + 81, + 266, + 165, + 324 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/7d7cef942ce057cb60a00774c0a66d4a2014fae01734de2c4a9236efa962ab35.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 179, + 266, + 261, + 321 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/599dd1680e78322b71cf4320a1fbabef424abcf2dcdc92cf320d90fb9eae7357.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 269, + 266, + 354, + 323 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/1ab3915806d0c865ae4022e60b1548c2f6b798a20cbcbc5e2618c2ac9be3b6cb.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 362, + 266, + 450, + 323 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/574affc845379e36ae8702dd04be0f2832554dd1dbd960ddc799e5ed2e4962a8.jpg", + "image_caption": [ + "$4^{\\text{th}}$ Salience Drop" + ], + "image_footnote": [], + "bbox": [ + 83, + 344, + 168, + 402 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/8dd8cae2b3a507aba4606c3f75e8484f63a74de2ebb0238f04b22a132dfb59cf.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 176, + 344, + 261, + 402 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/a0e00ed28d66eb4d9aff998fcf6aa08e57dac8ccd3f011a844301dc7d6463c52.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 269, + 344, + 354, + 402 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/2d5cc0608631b234445baed22a92557952a986e72343a55c1560eed6638d0a7e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 362, + 344, + 450, + 402 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/4c8c4c001c6c33acf2f5c3cc255a13ccfc0fc59c4a7df3b9456ddbfbcb3e2a1e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 83, + 422, + 169, + 482 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/ada2590fe21afaeef9cf5a094d9e93eb679042206ded94d0f56620920dfb9831.jpg", + "image_caption": [ + "Thresholding" + ], + "image_footnote": [], + "bbox": [ + 176, + 422, + 259, + 482 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/8b5d38c510f3efa6df5c28d36d86aa7a70ba01e01594fcb7527ea928a0eddfdf.jpg", + "image_caption": [ + "Gaussian Blur" + ], + "image_footnote": [], + "bbox": [ + 269, + 422, + 352, + 482 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/74dccded38d1fd993f7f9dc49dafac7db0251247dffd40af1cb61aec99dbc44e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 362, + 422, + 450, + 482 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/c25ee5c291fab16c727494f080edec705ac1602c98abb8c60ef1bc61f2fdfc4b.jpg", + "image_caption": [ + "Figure 4. An illustration of Salience DropOut, showing GradCAM salience values after each iteration. Black squares in the images indicate dropped patches. We obtain the final result by summing the salience maps from all iterations and applying thresholding, Gaussian blur, and Dense CRF." + ], + "image_footnote": [], + "bbox": [ + 83, + 556, + 168, + 618 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/d15c7bcad4504082e1f2f8e7978b68e10b08dfb951ba47e8fe4d4c9723144036.jpg", + "image_caption": [ + "Dense CRF" + ], + "image_footnote": [], + "bbox": [ + 174, + 556, + 259, + 618 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/71cda47ac3f80029bb03cc0bb58cac64d88f81c1227bad5c911df2592ffd00c3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 267, + 556, + 354, + 618 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/e53037d5bc68fc58c941803f94421208306468a2776e930d7028434eb79bc6ed.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 362, + 556, + 450, + 618 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "still important for the completeness of masks. To compel the VLM to attend to these regions, we propose an iterative technique called Salience DropOut.", + "bbox": [ + 75, + 728, + 468, + 773 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Since our task is zero-shot and open-vocabulary, we do not have prior knowledge of the classes present in the image. Hence, we sum up the salience maps $\\widetilde{M}^{(k)}$ over all classes $k$ , yielding a class-agnostic salience map $U^{(t)}$ for the $t^{\\text{th}}$ Salience DropOut iteration, $U^{(t)} = \\sum_{k=1}^{K} \\widetilde{M}^{(k,t)}$ , where $\\widetilde{M}^{(k,t)}$ is the GradCAM salience map for class $k$ after the $t^{\\text{th}}$ iteration. Next, we zero out the $50\\%$ of image patches with the highest values in $U^{(t)}$ . On the remaining", + "bbox": [ + 75, + 773, + 470, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "image, we compute the GradCAM salience maps $\\widetilde{M}^{(k,t + 1)}$ and their sum, $U^{(t + 1)}$ . Any image patch previously zeroed out will always receive zero salience in later iterations.", + "bbox": [ + 498, + 88, + 890, + 135 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Formally, the set of remaining image patches $S^{(t)} \\subseteq \\{1, \\dots, P\\}^2$ after the $t^{\\text{th}}$ dropout iteration is defined as", + "bbox": [ + 500, + 136, + 890, + 167 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nS ^ {(t)} = S ^ {(t - 1)} \\backslash \\{(i, j) \\mid U _ {i j} ^ {(t)} \\geq \\eta \\}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 547, + 179, + 890, + 199 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\eta = \\operatorname {m e d i a n} \\left(\\left\\{U _ {i j} ^ {(t)} \\mid (i, j) \\in S ^ {(t - 1)} \\right\\}\\right), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 565, + 202, + 890, + 227 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $U_{i,j}^{(t)}$ denotes the aggregate salience value of the image patch at row $i$ and column $j$ . Additionally, note that the input to the first dropout iteration $S^0 = \\{1,\\dots ,P\\} ^2$ is the set of all image patches.", + "bbox": [ + 498, + 241, + 890, + 301 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We stop at four rounds of dropout, as almost all $(93.75\\%)$ patches are removed beyond that point. The final output for each class $k$ is the sum over all salience maps across the four dropout iterations, $\\hat{M}^{(k)} = \\sum_{t=1}^{4} \\widetilde{M}^{(k,t)}$ .", + "bbox": [ + 498, + 303, + 890, + 366 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Gaussian Blur and Dense CRF", + "text_level": 1, + "bbox": [ + 500, + 375, + 769, + 388 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The salience dropout procedure generates $k$ continuous-valued salience maps, one per object class. To filter out small random noise in the salience values, we subsequent apply a straightforward thresholding operation at a predefined value $T$ on the salience values and obtain binary segmentation masks. Nonetheless, the hard thresholding creates jagged segmentations with sharp edges that often do not coincide with object boundaries. One common strategy in zero-shot segmentation is to apply Dense Conditional Random Field (CRF) [29], which makes fine-grained adjustments to the estimated masks by enforcing consistency between nearby image pixels with similar colors.", + "bbox": [ + 496, + 398, + 890, + 579 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "However, we find that the hard 0/1 labels in the binary masks do not work well as pixel unary potentials for Dense CRF. Hence, we smooth them using a Gaussian kernel with a preset variance $\\sigma$ , which results in a better initialization for unary terms and accounts for uncertainty of exact segmentation boundary along the patch boundaries.", + "bbox": [ + 496, + 580, + 890, + 670 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Hyperparameter Tuning", + "text_level": 1, + "bbox": [ + 500, + 680, + 725, + 696 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Three hyperparameters in PnP-OVSS have the strongest influence on the result, the cross-attention layer $L$ , the attention head $H$ , and the binary threshold $T$ . Traditionally, tuning these hyperparameters requires a validation set with pixel-level labels. However, since our goal is to perform zero-shot open-vocabulary semantic segmentation, this requirement could potentially limit the applicability of the technique. Instead, we propose a weakly supervised reward function for hyperparameter tuning, which only requires a set of images and the class names appearing in each image.", + "bbox": [ + 496, + 703, + 890, + 854 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The reward for an image $I$ is calculated as follows. We start with a set of classes present in the image, denoted as $\\mathcal{K}(I)$ . For each class $k\\in \\mathcal{K}(I)$ , we obtain a segmentation", + "bbox": [ + 498, + 854, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "4033", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "mask $M^{(k)}$ (which can be the GradCAM mask, the Salience DropOut mask, or the Dense CRF mask). Next, we apply the mask to the image $I$ and input the extracted regions $M^{(k)} \\otimes I$ into a pretrained neural network $f$ , which takes an image and a textual class name as input and produces a similarity score. We calculate the normalized probability that the masked image $M^{(k)} \\otimes I$ belongs to the ground-truth class $k$ and contrast with a completely black image $\\mathbf{0}$ .", + "bbox": [ + 75, + 89, + 472, + 212 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {R e w a r d} = \\sum_ {k \\in \\mathcal {K} (I)} \\mathbb {1} \\left[ P r \\left(M ^ {(k)} \\otimes I, k\\right) > P r (\\mathbf {0}, k) \\right], \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 96, + 217, + 468, + 251 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nP r (I, k) = \\frac {\\exp (f (I , k))}{\\sum_ {k ^ {\\prime} \\in \\mathcal {K} (I)} \\exp (f (I , k ^ {\\prime}))}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 89, + 253, + 468, + 289 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\mathbb{1}(\\cdot)$ is the indicator function. Intuitively, a reward of 1 is assigned if and only if the image features pulled with the estimated mask for class $k$ bears higher similarity to the class name of $k$ than a black image (which can be interpreted as the prior probability of class $k$ ).", + "bbox": [ + 75, + 294, + 468, + 369 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We sum up the reward for all validation images as the total reward. The best hyperparameters, including the cross-attention layer, the attention head, the threshold $T$ , and the variance in the Gaussian blur kernel, are determined using a simple random search.", + "bbox": [ + 75, + 369, + 468, + 445 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 76, + 455, + 209, + 473 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Datasets and Implementation Details", + "text_level": 1, + "bbox": [ + 76, + 481, + 397, + 498 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Following the previous work for zero-shot semantic segmentation, we adopt validation sets of Pascal-VOC 2012 [16], Pascal Context [45], COCO Object [42], COCO Stuff [5], and ADE20K [88] that contain, respectively, 20 object classes, 59 object and stuff classes, 80 object classes, 171 object and stuff classes, and 150 object and stuff classes to evaluate our framework. To verify its versatility, we apply PnP-OVSSto two high-performance VLMs, BLIP [37] and BridgeTower [78], which have the ITM loss and text-to-image cross-attention. More details are in the supplementary material.", + "bbox": [ + 75, + 503, + 468, + 671 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Hyperparameter tuning. We use CLIP VIT-L/14 to calculate the reward. The input resolution is $336 \\times 336$ . The search spaces of the random search and results are shown in Tab. 2. For computation efficiency, we tune the layer, the head, and the threshold with GradCAM masks. With the first three hyperparameters fixed, we tune the Gaussian variance on masks before Dense CRF. We directly adopt Dense CRF hyperparameters from CutLER [66] without tuning.", + "bbox": [ + 75, + 679, + 468, + 801 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Baselines. We adopt recent papers on zero-shot open-vocabulary semantic segmentation as baselines. The only baselines strictly comparable to our work are MaskClip [89] and Reco [58], which do not perform any finetuning on pretrained VLMs and do not perform hyperparameter tuning on dense annotations. We call these Group", + "bbox": [ + 75, + 809, + 468, + 902 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/8b50861a9067b7f49de6e8b30457b616c45d27227634fbdbf78e91a159253ac1.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
HyperparametersStartEndStepSolution
BLIP
Layer11218
Head112110
Attention Threshold0.050.50.10.15
BridgeTower
Layer1612
Head11618
Attention Threshold0.050.50.10.15
Gaussian Blur
Standard Deviation0.010.110.020.05
", + "bbox": [ + 531, + 88, + 864, + 271 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 2. Search space for hyperparameters", + "bbox": [ + 568, + 280, + 823, + 295 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3. To further expand our scope, we also include two other groups of baselines. Group 2 finetunes VLMs but does not require image-text pairs, including MaskClip with self-training (ST) [89] and ZeroSeg [8]. We include ZeroSeg, trained with ImageNet1K, and ZeroSeg*, trained with CC3M+COCO. Group 1 contains baselines that require training on image-text pairs. For completeness, we also compare against supervised techniques since 2019. To maintain the zero-shot setting, we test them on classes not observed during training. For more details, we refer readers to the supplementary material.", + "bbox": [ + 496, + 321, + 890, + 488 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Main Results", + "text_level": 1, + "bbox": [ + 500, + 498, + 638, + 513 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We show the main results in Tab. 3. As input resolution may influence results and cause unfair comparisons, we label the resolution used by each method. PACL [46] uses an 224 resolution but changes the stride for image patchification from 16 to 4, hence introducing overlapping patches.", + "bbox": [ + 496, + 521, + 890, + 598 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "PnP-OVSS exhibits excellent performance. Comparing with MaskClip [89] and Reco [58], the two methods that require no additional training and ground truth for hyperparameter tuning, on an equal-resolution basis, we attain $+29.4\\%$ mIoU on Pascal VOC, $+13.2\\%$ mIoU on Pascal Context, $+14.0\\%$ mIoU on COCO Object, and $+11.4\\%$ mIoU on ADE-20K. Further, PnP-OVSS surpasses all baselines in Group 2. On an equal-resolution basis, we achieve $+13.7\\%$ mIoU on Pascal Voc, $+9.6\\%$ on Pascal Context, $+9.5\\%$ on COCO Object, $+11.6\\%$ on COCO Stuff, and $+12.8\\%$ on ADE-20K.", + "bbox": [ + 496, + 598, + 890, + 763 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "When compared to Group 1, PnP-OVSS still outperforms most. On the Pascal datasets, under equal resolutions, PnP-OVSS + BLIPFlickr outperforms 6 out of 10 baselines on Pascal VOC, and 8 out of 9 baselines on Pascal Context. For the COCO datasets, under equal resolutions, PnP-OVSS + BLIPFlickr beats 5 out of 8 baselines on COCO Object, and all baselines except PACL [46] on COCO Stuff. The two Pascal datasets share many images, and the two COCO datasets use exactly the same images, but they have", + "bbox": [ + 496, + 763, + 890, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "4034", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/55f789a89ab373b01f375bad2003af82fa7db4d7763d03bed43635ff49253f50.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodFinetuning VLMsHT on Dense LabelsShort-side ResolutionPascal VOC-20Pascal Context-59COCO Object-80COCO Stuff-171ADE 20K-150
Group 1: Methods that require weakly supervised finetuning on image-text data
ViL-Seg†[43]✓✓-37.318.9-18.0-
CLIPpy [51]✓✓22452.2-32.025.5*13.5
SegClip [44]✓✓22452.624.726.5--
GroupVit (by [51])✓✓22428.114.812.9-6.2
GroupVit (by [6])✓✓44850.418.727.515.39.2
GroupVit [73]✓✓44852.322.424.3--
ViewCo [52]✓✓44852.423.023.5--
OVSegmentor [74]✓✓44853.820.425.1--
TCL [6] +PAMR [2]✓✓44855.030.431.622.417.1
PACL [46]✓✓224×472.350.1-38.831.4
Group 2: Methods that require finetuning but not real image-text data
MaskClip w/ ST [89]✓✓336-31.1-18.0-
MaskClip w/ ST (by [6])✓✓44838.823.620.616.49.8
ZeroSeg [8]✓✓44840.820.420.2--
Group 3: Methods that require no finetuning
MaskClip (by [51])××22422.1-13.88.16.8
MaskClip [89]××336-25.5-14.6-
Reco [58]××320-27.2-27.2-
Reco (by [6])××44825.119.915.714.811.2
PnP-OVSS with different VLMs
BLIPFlickr××22447.836.424.825.818.2
BLIPFlickr××32051.740.428.029.621.3
BLIPFlickr××33652.540.728.229.621.9
BLIPFlickr××44854.542.229.731.522.6
BLIPFlickr××76854.142.831.832.523.5
BLIPCOCO××76855.741.933.832.623.2
BridgeTower××32236.432.324.227.618.6
BridgeTower××33635.332.424.227.618.0
BridgeTower××77035.232.424.528.019.0
", + "bbox": [ + 84, + 87, + 885, + 592 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/720db4477a50564d77d47e1dd7384735bf23d99d4612b0de3ec53df734d61d27.jpg", + "table_caption": [ + "Table 3. Zero-shot semantic segmentation performance in mIoU. Group 3 contains the most similar baselines that serve as fair comparisons to PnP-OVSS. Groups 1 and 2 benefit from additional training, extra image-text data, and hyperparameter tuning on dense labels. We use the word \"by\" followed by a paper citation to indicate results of the same technique reported by different papers. $\\star$ CLIPpy tests on 133 categories of COCO Stuff while we test all 171 classes of COCO Stuff. ViL-Seg† is tested on subset of classes on the three datasets, as detailed in the supplementary." + ], + "table_footnote": [], + "table_body": "
MethodDense \nLabelsHT on Dense \nLabelsPascal \nVOC-20Pascal \nContext-59COCO \nStuff-171
PnP-OVSS + BLIPFlickr (Ours)××53.653.839.8
SPNet+ST [71]✓✓25.8-26.9
ZS3Net+ST [4]✓✓21.220.710.6
CaGNet+ST [18]✓✓30.3-13.4
STRICT [47]✓✓35.6-30.3
LSeg [32]✓✓41.0--
SimBase [76]✓✓72.5-36.3
MaskCLIP+ w/ ST [89]✓✓86.166.754.7
", + "bbox": [ + 194, + 684, + 777, + 849 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 4. Comparison of zero-shot semantic segmentation performance on unseen categories with methods trained with dense annotation.", + "bbox": [ + 78, + 859, + 887, + 875 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "4035", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/202c3313600172d50dec906994c14ece22a2b25bc916fbf50f6c3b84c4c57976.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Ablated ModelPascal ContextCOCO Stuff
BLIPFlickr19.814.5
+ GradCam21.617.5
+ Drop 125.119.8
+ Drop 226.520.6
+ Drop 327.020.9
+ Drop 427.220.9
+ Drop 4 + Blur36.828.6
+ Drop 4 + Dense CRF35.331.8
+ Drop 4 + Blur + Dense CRF42.832.5
", + "bbox": [ + 107, + 88, + 439, + 257 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "vastly different lists of classes to segment. We observe that the advantage of PnP-OVSS over methods in Group 1 becomes more pronounced as the number of classes on the same image increases.", + "bbox": [ + 75, + 324, + 468, + 383 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "When applied to BridgeTower [78], PnP-OVSS still surpasses all methods in Group 3 by $10.2\\%$ on Pascal VOC, $5.1\\%$ on Pascal Context, $8.5\\%$ on COCO Object, $0.4\\%$ on COCO Stuff, and $6.8\\%$ on ADE-20K. This showcases the plug-and-play ability of PnP-OVSS, which excels with different base networks.", + "bbox": [ + 75, + 383, + 468, + 474 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We report comparisons against supervised methods in Tab 4. PnP-OVSS + BLIPFlickr outperforms 5 out of 7 methods on Pascal VOC, as well as every baseline except MaskCLIP+ on Pascal Context and COCO stuff. As these baselines benefit from dense supervision, the results further demonstrate the strengths of PnP-OVSS.", + "bbox": [ + 75, + 476, + 468, + 568 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.3. Ablation Study", + "text_level": 1, + "bbox": [ + 76, + 579, + 230, + 595 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We perform gradual ablation of the components of $\\mathrm{PnP-OVSS}$ on $\\mathrm{BLIP}_{\\mathrm{Flickr}}$ and report the results in Tab. 5. Each component, including GradCAM, all Salience DropOut iterations, Gaussian blur, and Dense CRF, contribute positively to the final performance. In particular, the first iteration of Salience DropOut has much larger impact (+3.5/2.3) than the second iteration (+1.4/0.8), which in turn is more important than the rest. Interestingly, Gaussian blur by itself attains good performance (+9.6/7.7) whereas Dense CRF only works well when combined with blur. Dense CRF alone is worse than Gaussian blur by 1.5 mIoU on Pascal Context. This is likely caused by the fact that hard 0/1 labels resulted from thresholding are not informative unary potentials that can be leveraged by CRF effectively.", + "bbox": [ + 75, + 603, + 468, + 815 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4. Hyperparameter Sensitivity", + "text_level": 1, + "bbox": [ + 76, + 825, + 328, + 843 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The choice of hyperparameters often exerts substantial influence on segmentation performance. Here we quantitatively examine how the choice of cross-attention layers", + "bbox": [ + 75, + 854, + 468, + 902 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/59e3c2a8cb356f32db262a9e7028beca7a840fbb873e7f211f09427967049036.jpg", + "table_caption": [ + "Table 5. An ablation study of PnP-OVSS + BLIPFlickr with resolution 768 on Pascal Context and COCO Stuff." + ], + "table_footnote": [], + "table_body": "
Layer Mean123456
mIoU11.612.311.311.011.612.9
Layer Mean789101112
mIoU13.725.325.823.412.55.8
Head in Layer 8123456
mIoU16.63.88.620.019.85.2
Head in Layer 8789101112
mIoU11.16.03.829.617.38.0
", + "bbox": [ + 509, + 88, + 883, + 257 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 6. Semantic segmentation performance using cross-attention maps averaged across all heads in a layer and separate heads in Layer 8. Results are attained with PnP-OVSS+BLIPFlickr on COCO Stuff and resolution 336.", + "bbox": [ + 498, + 267, + 890, + 323 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "and attention heads may change the segmentation mIoU on COCO Stuff. Tab. 6 shows the results obtained from the average cross-attention maps over all heads in each layer and those from different attention heads.", + "bbox": [ + 496, + 354, + 890, + 415 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We make the following observations. First, different layers and heads have drastic performance differences. The best-to-worst difference among all layers is 20, and that among heads in Layer 8 is 23.8, underscoring the importance of hyperparameter tuning. Second, the random search using the proposed reward function correctly identifies the best layer-head combination, even though Layer 8 is not the best layer based on average head performance. This indicates the effectiveness of our method.", + "bbox": [ + 496, + 417, + 892, + 554 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusions", + "text_level": 1, + "bbox": [ + 500, + 575, + 627, + 592 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We propose PnP-OVSS, which extracts the ability of semantic segmentation from opaque VLMs. PnP-OVSS is simple to use, requires no extra finetuning, and delivers high performance, exceeding not only all baselines that requires no finetuning, but also all baselines that do not use image-text pairs in finetuning. Its success hints at a new direction for open-vocabulary segmentation tasks leveraging large VLMs.", + "bbox": [ + 496, + 608, + 890, + 729 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Acknowledgments", + "text_level": 1, + "bbox": [ + 500, + 751, + 679, + 768 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This work has been supported by the Nanyang Associate Professorship and the National Research Foundation Fellowship (NRF- NRFF13-2021-0006), Singapore, and by the Vector Institute for AI, Canada CIFAR AI Chair, and NSERC Discovery Grant. Any opinions, findings, conclusions, or recommendations expressed in this material are those of the authors and do not reflect the views of the funding agencies.", + "bbox": [ + 496, + 779, + 890, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "4036", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 174, + 106 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, Roman Ring, Eliza Rutherford, Serkan Cabi, Tengda Han, Zhitao Gong, Sina Samangooei, Marianne Monteiro, Jacob L Menick, Sebastian Borgeaud, Andy Brock, Aida Nematzadeh, Sahand Sharifzadeh, Mikol aj Binkowski, Ricardo Barreira, Oriol Vinyals, Andrew Zisserman, and Karén Simonyan. Flamingo: a visual language model for few-shot learning. In Advances in Neural Information Processing Systems, 2022. 1, 2", + "[2] Nikita Araslanov and Stefan Roth. Single-stage semantic segmentation from image labels. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4253-4262, 2020. 7", + "[3] Loris Bazzani, Alessandra Bergamo, Dragomir Anguelov, and Lorenzo Torresani. Self-taught object localization with deep networks. In 2016 IEEE winter conference on applications of computer vision (WACV), pages 1-9. IEEE, 2016. 3", + "[4] Maxime Bucher, Tuan-Hung Vu, Matthieu Cord, and Patrick Pérez. Zero-shot semantic segmentation. Advances in Neural Information Processing Systems, 32, 2019. 3, 7", + "[5] Holger Caesar, Jasper Uijlings, and Vittorio Ferrari. Cocosstuff: Thing and stuff classes in context. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1209-1218, 2018. 6", + "[6] Junbum Cha, Jonghwan Mun, and Byungseok Roh. Learning to generate text-grounded mask for open-world semantic segmentation from only image-text pairs. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11165-11174, 2023. 3, 7", + "[7] Jun Chen, Han Guo, Kai Yi, Boyang Li, and Mohamed Elhoseiny. Visualgpt: Data-efficient adaptation of pretrained language models for image captioning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18030-18040, 2022. 2", + "[8] Jun Chen, Deyao Zhu, Guocheng Qian, Bernard Ghanem, Zhicheng Yan, Chenchen Zhu, Fanyi Xiao, Sean Chang Culatana, and Mohamed Elhoseiny. Exploring open-vocabulary semantic segmentation from clip vision encoder distillation only. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 699-710, 2023. 1, 3, 6, 7", + "[9] Qi Chen, Lingxiao Yang, Jian-Huang Lai, and Xiaohua Xie. Self-supervised image-specific prototype exploration for weakly supervised semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4288-4298, 2022. 1", + "[10] Yen-Chun Chen, Linjie Li, Licheng Yu, Ahmed El Kholy, Faisal Ahmed, Zhe Gan, Yu Cheng, and Jingjing Liu. Uniter: Universal image-text representation learning. In Proceedings of the 16th European Conference on Computer Vision, 2020. 2", + "[11] Zhaozheng Chen, Tan Wang, Xiongwei Wu, Xian-Sheng Hua, Hanwang Zhang, and Qianru Sun. Class re-activation" + ], + "bbox": [ + 78, + 114, + 470, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "maps for weakly-supervised semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 969-978, 2022. 1", + "[12] Minsu Cho, Suha Kwak, Cordelia Schmid, and Jean Ponce. Unsupervised object discovery and localization in the wild: Part-based matching with bottom-up region proposals. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1201-1210, 2015. 3", + "[13] Ramazan Gokberk Cinbis, Jakob Verbeek, and Cordelia Schmid. Weakly supervised object localization with multifold multiple instance learning. IEEE transactions on pattern analysis and machine intelligence, 39(1):189-203, 2016. 3", + "[14] Timothée Darcet, Maxime Oquab, Julien Mairal, and Piotr Bojanowski. Vision transformers need registers. arXiv preprint arXiv:2309.16588, 2023. 3", + "[15] Ye Du, Zehua Fu, Qingjie Liu, and Yunhong Wang. Weakly supervised semantic segmentation by pixel-to-prototype contrast. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4320–4329, 2022. 1", + "[16] Mark Everingham, Luc Van Gool, Christopher KI Williams, John Winn, and Andrew Zisserman. The pascal visual object classes (voc) challenge. International journal of computer vision, 88:303-338, 2010. 6", + "[17] Golnaz Ghiasi, Xiuye Gu, Yin Cui, and Tsung-Yi Lin. Scaling open-vocabulary image segmentation with image-level labels. In European Conference on Computer Vision, pages 540-557. Springer, 2022. 1, 3", + "[18] Zhangxuan Gu, Siyuan Zhou, Li Niu, Zihan Zhao, and Liqing Zhang. Context-aware feature generation for zero-shot semantic segmentation. In Proceedings of the 28th ACM International Conference on Multimedia, pages 1921-1929, 2020. 3, 7", + "[19] Jiaxian Guo, Junnan Li, Dongxu Li, Anthony Meng Huat Tiong, Boyang Li, Dacheng Tao, and Steven Hoi. From images to textual prompts: Zero-shot visual question answering with frozen large language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10867-10877, 2023. 1", + "[20] Yanming Guo, Yu Liu, Theodoros Georgiou, and Michael S Lew. A review of semantic segmentation using deep neural networks. International journal of multimedia information retrieval, 7:87-93, 2018. 1", + "[21] Kunyang Han, Yong Liu, Jun Hao Liew, Henghui Ding, Jiajun Liu, Yitong Wang, Yansong Tang, Yujiu Yang, Jiashi Feng, Yao Zhao, et al. Global knowledge calibration for fast open-vocabulary segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 797-807, 2023. 1", + "[22] Shijie Hao, Yuan Zhou, and Yanrong Guo. A brief survey on semantic segmentation with deep learning. Neurocomputing, 406:302-321, 2020. 1", + "[23] Qibin Hou, PengTao Jiang, Yunchao Wei, and Ming-Ming Cheng. Self-erasing network for integral object attention. Advances in Neural Information Processing Systems, 31, 2018. 3", + "[24] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom" + ], + "bbox": [ + 503, + 93, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "4037", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In Proceedings of the 38th International Conference on Machine Learning, 2021. 2", + "[25] Peng-Tao Jiang, Yuqi Yang, Qibin Hou, and Yunchao Wei. L2g: A simple local-to-global knowledge transfer framework for weakly supervised semantic segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16886-16896, 2022. 1", + "[26] Tsung-Wei Ke, Jyh-Jing Hwang, and Stella X Yu. Universal weakly supervised segmentation by pixel-to-segment contrastive learning. arXiv preprint arXiv:2105.00957, 2021. 3", + "[27] Wonjae Kim, Bokyung Son, and Ildoo Kim. Vilt: Vision-and-language transformer without convolution or region supervision. In Proceedings of the 38th International Conference on Machine Learning, 2021. 2", + "[28] Alexander Kolesnikov and Christoph H Lampert. Seed, expand and constrain: Three principles for weakly-supervised image segmentation. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part IV 14, pages 695-711. Springer, 2016. 3", + "[29] Philipp Krahenbuhl and Vladlen Koltun. Efficient inference in fully connected crfs with gaussian edge potentials. Advances in neural information processing systems, 24, 2011. 3, 5", + "[30] Krishna Kumar Singh and Yong Jae Lee. Hide-and-seek: Forcing a network to be meticulous for weakly-supervised object and action localization. In Proceedings of the IEEE International Conference on Computer Vision, pages 3524-3533, 2017. 3", + "[31] Jungbeom Lee, Seong Joon Oh, Sangdoo Yun, Junsuk Choe, Eunji Kim, and Sungroh Yoon. Weakly supervised semantic segmentation using out-of-distribution data. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16897-16906, 2022. 1", + "[32] Boyi Li, Kilian Q. Weinberger, Serge Belongie, Vladlen Koltun, and René Ranftl. Language-driven semantic segmentation, 2022. 3, 7", + "[33] Chenliang Li, Haiyang Xu, Junfeng Tian, Wei Wang, Ming Yan, Bin Bi, Jiabo Ye, He Chen, Guohai Xu, Zheng Cao, Ji Zhang, Songfang Huang, Fei Huang, Jingren Zhou, and Luo Si. mPLUG: Effective and efficient vision-language learning by cross-modal skip-connections. In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, 2022. 2", + "[34] Chenliang Li, Haiyang Xu, Junfeng Tian, Wei Wang, Ming Yan, Bin Bi, Jiabo Ye, Hehong Chen, Guohai Xu, Zheng Cao, et al. mplug: Effective and efficient vision-language learning by cross-modal skip-connections. arXiv preprint arXiv:2205.12005, 2022. 2, 3", + "[35] Junnan Li, Ramprasaath R. Selvaraju, Akhilesh Deepak Gotmare, Shafiq Joty, Caiming Xiong, and Steven Hoi. Align before fuse: Vision and language representation learning with momentum distillation. In NeurIPS, 2021. 2, 3, 4", + "[36] Jinlong Li, Zequn Jie, Xu Wang, Xiaolin Wei, and Lin Ma. Expansion and shrinkage of localization for weakly-" + ], + "bbox": [ + 78, + 92, + 468, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "supervised semantic segmentation. Advances in Neural Information Processing Systems, 35:16037-16051, 2022. 1", + "[37] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In International Conference on Machine Learning, pages 12888-12900. PMLR, 2022. 1, 2, 3, 6", + "[38] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In ICML, 2023. 1", + "[39] Wei Li, Can Gao, Guocheng Niu, Xinyan Xiao, Hao Liu, Jiachen Liu, Hua Wu, and Haifeng Wang. UNIMO: Towards unified-modal understanding and generation via cross-modal contrastive learning. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), 2021. 2", + "[40] Xiujun Li, Xi Yin, Chunyuan Li, Pengchuan Zhang, Xiaowei Hu, Lei Zhang, Lijuan Wang, Houdong Hu, Li Dong, Furu Wei, Yejin Choi, and Jianfeng Gao. Oscar: Object-semantics aligned pre-training for vision-language tasks. In ECCV 2020, 2020. 1, 2", + "[41] Feng Liang, Bichen Wu, Xiaoliang Dai, Kunpeng Li, Yinan Zhao, Hang Zhang, Peizhao Zhang, Peter Vajda, and Diana Marculescu. Open-vocabulary semantic segmentation with mask-adapted clip. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7061-7070, 2023. 1, 3", + "[42] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 6", + "[43] Quande Liu, Youpeng Wen, Jianhua Han, Chunjing Xu, Hang Xu, and Xiaodan Liang. Open-world semantic segmentation via contrasting and clustering vision-language embedding. In European Conference on Computer Vision, pages 275-292. Springer, 2022. 2, 3, 7", + "[44] Huaishao Luo, Junwei Bao, Youzheng Wu, Xiaodong He, and Tianrui Li. Segclip: Patch aggregation with learnable centers for open-vocabulary semantic segmentation. In International Conference on Machine Learning, pages 23033-23044. PMLR, 2023. 1, 2, 3, 7", + "[45] Roozbeh Mottaghi, Xianjie Chen, Xiaobai Liu, Nam-Gyu Cho, Seong-Whan Lee, Sanja Fidler, Raquel Urtasun, and Alan Yuille. The role of context for object detection and semantic segmentation in the wild. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 891–898, 2014. 6", + "[46] Jishnu Mukhoti, Tsung-Yu Lin, Omid Poursaeed, Rui Wang, Ashish Shah, Philip HS Torr, and Ser-Nam Lim. Open vocabulary semantic segmentation with patch aligned contrastive learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19413-19423, 2023. 1, 3, 6, 7" + ], + "bbox": [ + 503, + 92, + 890, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "4038", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[47] Giuseppe Pastore, Fabio Cermelli, Yongqin Xian, Massimiliano Mancini, Zeynep Akata, and Barbara Caputo. A closer look at self-training for zero-label semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2693-2702, 2021. 3, 7", + "[48] Jie Qin, Jie Wu, Pengxiang Yan, Ming Li, Ren Yuxi, Xuefeng Xiao, Yitong Wang, Rui Wang, Shilei Wen, Xin Pan, et al. Freeseg: Unified, universal and open-vocabulary image segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19446-19455, 2023. 1", + "[49] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In Proceedings of the 38th International Conference on Machine Learning, 2021. 2, 3", + "[50] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 2", + "[51] Kanchana Ranasinghe, Brandon McKinzie, Sachin Ravi, Yinfei Yang, Alexander Toshev, and Jonathon Shlens. Perceptual grouping in contrastive vision-language models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5571-5584, 2023. 2, 3, 7", + "[52] Pengzhen Ren, Changlin Li, Hang Xu, Yi Zhu, Guangrun Wang, Jianzhuang Liu, Xiaojun Chang, and Xiaodan Liang. Viewco: Discovering text-supervised segmentation masks via multi-view semantic consistency. arXiv preprint arXiv:2302.10307, 2023. 2, 3, 7", + "[53] Lixiang Ru, Yibing Zhan, Baosheng Yu, and Bo Du. Learning affinity from attention: End-to-end weakly-supervised semantic segmentation with transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16846-16855, 2022. 1", + "[54] Lixiang Ru, Heliang Zheng, Yibing Zhan, and Bo Du. Token contrast for weakly-supervised semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3093-3102, 2023. 3", + "[55] Lixiang Ru, Heliang Zheng, Yibing Zhan, and Bo Du. Token contrast for weakly-supervised semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3093-3102, 2023. 1", + "[56] Michael Rubinstein, Armand Joulin, Johannes Kopf, and Ce Liu. Unsupervised joint object discovery and segmentation in internet images. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1939-1946, 2013. 3", + "[57] Ramprasaath R. Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. Grad-cam: Visual explanations from deep networks via gradient-based localization. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2017. 2, 4" + ], + "bbox": [ + 78, + 90, + 470, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[58] Gyungin Shin, Weidi Xie, and Samuel Albanie. Reco: Retrieve and co-segment for zero-shot transfer. Advances in Neural Information Processing Systems, 35:33754-33767, 2022. 3, 6, 7", + "[59] Oriane Simeoni, Gilles Puy, Huy V Vo, Simon Roburin, Spyros Gidaris, Andrei Bursuc, Patrick Pérez, Renaud Marlet, and Jean Ponce. Localizing objects with self-supervised transformers and no labels. arXiv preprint arXiv:2109.14279, 2021. 3", + "[60] Amanpreet Singh, Ronghang Hu, Vedanuj Goswami, Guillaume Couairon, Wojciech Galuba, Marcus Rohrbach, and Douwe Kiela. Flava: A foundational language and vision alignment model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 15638-15650, 2022. 1, 2", + "[61] Meng Tang, Federico Perazzi, Abdelaziz Djelouah, Ismail Ben Ayed, Christopher Schroers, and Yuri Boykov. On regularized losses for weakly-supervised cnn segmentation. In Proceedings of the European Conference on Computer Vision (ECCV), pages 507-522, 2018. 3", + "[62] Anthony Meng Huat Tiong, Junnan Li, Boyang Li, Silvio Savarese, and Steven C.H. Hoi. Plug-and-play vqa: Zero-shot vqa by conjoining large pretrained models with zero training. In Findings of the Conference on Empirical Methods in Natural Language Processing (Findings of EMNLP), 2022. 1, 2, 4", + "[63] Huy V Vo, Francis Bach, Minsu Cho, Kai Han, Yann LeCun, Patrick Pérez, and Jean Ponce. Unsupervised image matching and object discovery as optimization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8287-8296, 2019. 3", + "[64] Van Huy Vo, Elena Sizikova, Cordelia Schmid, Patrick Pérez, and Jean Ponce. Large-scale unsupervised object discovery. Advances in Neural Information Processing Systems, 34:16764-16778, 2021. 3", + "[65] Wenguan Wang, Guolei Sun, and Luc Van Gool. Looking beyond single images for weakly supervised semantic segmentation learning. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2022. 1", + "[66] Xudong Wang, Rohit Girdhar, Stella X Yu, and Ishan Misra. Cut and learn for unsupervised object detection and instance segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3124-3134, 2023. 6", + "[67] Yangtao Wang, Xi Shen, Shell Xu Hu, Yuan Yuan, James L Crowley, and Dominique Vaufreydaz. Self-supervised transformers for unsupervised object discovery using normalized cut. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14543-14553, 2022. 3", + "[68] Xiu-Shen Wei, Chen-Lin Zhang, Jianxin Wu, Chunhua Shen, and Zhi-Hua Zhou. Unsupervised object discovery and co-localization by deep descriptor transformation. Pattern Recognition, 88:113-126, 2019. 3", + "[69] Yunchao Wei, Jiashi Feng, Xiaodan Liang, Ming-Ming Cheng, Yao Zhao, and Shuicheng Yan. Object region mining with adversarial erasing: A simple classification to semantic" + ], + "bbox": [ + 501, + 92, + 893, + 900 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "4039", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "segmentation approach. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1568-1576, 2017. 3", + "[70] Alexandros Xenos, Themos Stafylakis, Ioannis Patras, and Georgios Tzimiropoulos. A simple baseline for knowledge-based visual question answering. In EMNLP, 2023. 1", + "[71] Yongqin Xian, Subhabrata Choudhury, Yang He, Bernt Schiele, and Zeynep Akata. Semantic projection network for zero-and few-label semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8256-8265, 2019. 3, 7", + "[72] Haiyang Xu, Qinghao Ye, Ming Yan, Yaya Shi, Jiabo Ye, Yuanhong Xu, Chenliang Li, Bin Bi, Qi Qian, Wei Wang, et al. mplug-2: A modularized multi-modal foundation model across text, image and video. arXiv preprint arXiv:2302.00402, 2023. 2, 3", + "[73] Jiarui Xu, Shalini De Mello, Sifei Liu, Wonmin Byeon, Thomas Breuel, Jan Kautz, and Xiaolong Wang. Groupvit: Semantic segmentation emerges from text supervision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18134-18144, 2022. 2, 3, 7", + "[74] Jilan Xu, Junlin Hou, Yuejie Zhang, Rui Feng, Yi Wang, Yu Qiao, and Weidi Xie. Learning open-vocabulary semantic segmentation models from natural language supervision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2935-2944, 2023. 1, 2, 3, 7", + "[75] Lian Xu, Wanli Ouyang, Mohammed Bennamoun, Farid Boussaid, and Dan Xu. Multi-class token transformer for weakly supervised semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4310-4319, 2022. 1", + "[76] Mengde Xu, Zheng Zhang, Fangyun Wei, Yutong Lin, Yue Cao, Han Hu, and Xiang Bai. A simple baseline for open-vocabulary semantic segmentation with pre-trained vision-language model. In European Conference on Computer Vision, pages 736-753. Springer, 2022. 1, 3, 7", + "[77] Mengde Xu, Zheng Zhang, Fangyun Wei, Han Hu, and Xiang Bai. Side adapter network for open-vocabulary semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2945-2954, 2023. 1", + "[78] Xiao Xu, Chenfei Wu, Shachar Rosenman, Vasudev Lal, Wanxiang Che, and Nan Duan. Bridgetower: Building bridges between encoders in vision-language representation learning. arXiv preprint arXiv:2206.08657, 2022. 1, 2, 3, 6, 8", + "[79] Jiahui Yu, Zirui Wang, Vijay Vasudevan, Legg Yeung, Mojtaba Seyedhosseini, and Yonghui Wu. CoCa: Contrastive captioners are image-text foundation models. arXiv Preprint 2205.01917, 2022. 2", + "[80] Nir Zabari and Yedid Hoshen. Open-vocabulary semantic segmentation using test-time distillation. In European Conference on Computer Vision, pages 56-72. Springer, 2022. 1" + ], + "bbox": [ + 78, + 90, + 470, + 898 + ], + "page_idx": 11 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[81] Yan Zeng, Xinsong Zhang, and Hang Li. Multi-grained vision language pre-training: Aligning texts with visual concepts. arXiv preprint arXiv:2111.08276, 2021. 2, 3", + "[82] Yan Zeng, Xinsong Zhang, Hang Li, Jiawei Wang, Jipeng Zhang, and Wangchunshu Zhou. $X^2$ -vlm: All-in-one pretrained model for vision-language tasks. arXiv preprint arXiv:2211.12402, 2022. 2, 3", + "[83] Bingfeng Zhang, Jimin Xiao, Yunchao Wei, Mingjie Sun, and Kaizhu Huang. Reliability does matter: An end-to-end weakly supervised semantic segmentation approach. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 12765-12772, 2020. 3", + "[84] Hao Zhang, Feng Li, Xueyan Zou, Shilong Liu, Chunyuan Li, Jianwei Yang, and Lei Zhang. A simple framework for open-vocabulary segmentation and detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1020-1031, 2023. 1", + "[85] Pengchuan Zhang, Xiujun Li, Xiaowei Hu, Jianwei Yang, Lei Zhang, Lijuan Wang, Yejin Choi, and Jianfeng Gao. Vinvl: Revisiting visual representations in vision-language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5579-5588, 2021. 1", + "[86] Xiaolin Zhang, Yunchao Wei, Jiashi Feng, Yi Yang, and Thomas S Huang. Adversarial complementary learning for weakly supervised object localization. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1325-1334, 2018. 3", + "[87] Bolei Zhou, Aditya Khosla, Agata Lapedriza, Aude Oliva, and Antonio Torralba. Object detectors emerge in deep scene cnns. arXiv preprint arXiv:1412.6856, 2014. 3", + "[88] Bolei Zhou, Hang Zhao, Xavier Puig, Sanja Fidler, Adela Barriuso, and Antonio Torralba. Scene parsing through ade20k dataset. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 633-641, 2017. 6", + "[89] Chong Zhou, Chen Change Loy, and Bo Dai. Extract free dense labels from clip. In European Conference on Computer Vision, pages 696-712. Springer, 2022. 3, 6, 7", + "[90] Tianfei Zhou, Meijie Zhang, Fang Zhao, and Jianwu Li. Regional semantic contrast and aggregation for weakly supervised semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4299-4309, 2022. 1" + ], + "bbox": [ + 501, + 92, + 890, + 712 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "4040", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 11 + } +] \ No newline at end of file diff --git a/2024/Emergent Open-Vocabulary Semantic Segmentation from Off-the-shelf Vision-Language Models/bf254b56-9ba7-4039-9f52-56365463bb00_model.json b/2024/Emergent Open-Vocabulary Semantic Segmentation from Off-the-shelf Vision-Language Models/bf254b56-9ba7-4039-9f52-56365463bb00_model.json new file mode 100644 index 0000000000000000000000000000000000000000..3f8a98577fea14ed4a7b5077ba87917376ee75d9 --- /dev/null +++ b/2024/Emergent Open-Vocabulary Semantic Segmentation from Off-the-shelf Vision-Language Models/bf254b56-9ba7-4039-9f52-56365463bb00_model.json @@ -0,0 +1,2974 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.0, + 0.812, + 0.045 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.128, + 0.131, + 0.844, + 0.175 + ], + "angle": 0, + "content": "Emergent Open-Vocabulary Semantic Segmentation from Off-the-shelf Vision-Language Models" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.203, + 0.751, + 0.222 + ], + "angle": 0, + "content": "Jiayun Luo\\(^{\\flat}\\), Siddhesh Khandelwal\\(^{\\sharp}\\), Leonid Sigal\\(^{\\sharp}\\), and Boyang Li\\(^{\\flat}\\)" + }, + { + "type": "text", + "bbox": [ + 0.299, + 0.223, + 0.671, + 0.24 + ], + "angle": 0, + "content": "\\(^{b}\\)Nanyang Technological University, Singapore" + }, + { + "type": "text", + "bbox": [ + 0.232, + 0.24, + 0.739, + 0.257 + ], + "angle": 0, + "content": "\\(^{\\ddagger}\\)University of British Columbia, Vector Institute for AI, Canada" + }, + { + "type": "text", + "bbox": [ + 0.211, + 0.261, + 0.756, + 0.275 + ], + "angle": 0, + "content": "{luoj0028,boyang.li}@ntu.edu.sg，{skhandel,lsignal}@cs.ubc.ca" + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.311, + 0.209, + 0.445 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.21, + 0.311, + 0.335, + 0.445 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.338, + 0.311, + 0.493, + 0.445 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.495, + 0.31, + 0.645, + 0.445 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.647, + 0.31, + 0.755, + 0.445 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.76, + 0.311, + 0.891, + 0.445 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.449, + 0.893, + 0.478 + ], + "angle": 0, + "content": "Figure 1. Qualitative Results of PnP-OVSS + BLIP. Images are from Pascal Context and COCO stuff. The right columns and bottom rows show the ground-truth (GT); the rest are our results. Note accurate results even on complex (trees) and small objects (last column)." + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.49, + 0.314, + 0.506 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.522, + 0.473, + 0.884 + ], + "angle": 0, + "content": "From image-text pairs, large-scale vision-language models (VLMs) learn to implicitly associate image regions with words, which prove effective for tasks like visual question answering. However, leveraging the learned association for open-vocabulary semantic segmentation remains a challenge. In this paper, we propose a simple, yet extremely effective, training-free technique, Plug-and-Play Open-Vocabulary Semantic Segmentation (PnP-OVSS) for this task. PnP-OVSS leverages a VLM with direct text-to-image cross-attention and an image-text matching loss. To balance between over-segmentation and under-segmentation, we introduce Salience Dropout; by iteratively dropping patches that the model is most attentive to, we are able to better resolve the entire extent of the segmentation mask. PnP-OVSS does not require any neural network training and performs hyperparameter tuning without the need for any segmentation annotations, even for a validation set. PnP-OVSS demonstrates substantial improvements over comparable baselines (+29.4% mIoU on Pascal VOC, +13.2% mIoU on Pascal Context, +14.0% mIoU on MS COCO, +2.4% mIoU on COCO Stuff) and even outperforms most baselines that conduct additional network training on top of pretrained VLMs. Our codebase is at https://github.com/letitiabana/PnP-OVSS." + }, + { + "type": "title", + "bbox": [ + 0.502, + 0.49, + 0.631, + 0.506 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.518, + 0.892, + 0.669 + ], + "angle": 0, + "content": "The classic task of semantic segmentation [20, 22] aims to classify pixels to their object types. Traditional supervised methods require dense pixel-level annotations and are restricted to recognizing a predefined set of objects. To relax these constraints, open-vocabulary semantic segmentation [8, 17, 21, 41, 44, 46, 48, 74, 76, 77, 80, 84] aspires to identify arbitrary object categories, whereas weakly supervised techniques [9, 11, 15, 25, 31, 36, 53, 55, 65, 75, 90] can acquire pixel-level localization capabilities from coarse supervision, e.g., image labels or boxes." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.672, + 0.893, + 0.884 + ], + "angle": 0, + "content": "Large-scale vision-language models (VLMs) pretrained on image-text pairs [1, 37, 38, 40, 60, 78, 85] achieve unprecedented performance on multimodal tasks, such as describing arbitrary images and answering free-form, open-ended questions about them (either with [37, 38, 60] or without finetuning [1, 19, 62, 70]). These tasks apparently involve some ability to localize objects. For example, to answer the question \"what objects appear on the table?\", the model would have to first localize the table in the image and identify the objects on it. Hence, it is reasonable to conjecture that the VLM network learns to perform open-vocabulary localization from image-text pretraining. However, distilling the localization capability from the VLMs remains an open challenge." + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.886, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Most existing methods for open-vocabulary semantic" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4029" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.081, + 0.089, + 0.177, + 0.239 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.097, + 0.241, + 0.162, + 0.265 + ], + "angle": 0, + "content": "(a) Original image" + }, + { + "type": "image", + "bbox": [ + 0.178, + 0.089, + 0.274, + 0.24 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.199, + 0.241, + 0.266, + 0.265 + ], + "angle": 0, + "content": "(b) Cross-attention" + }, + { + "type": "image", + "bbox": [ + 0.276, + 0.089, + 0.371, + 0.24 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.285, + 0.242, + 0.361, + 0.253 + ], + "angle": 0, + "content": "(c) GradCAM" + }, + { + "type": "image", + "bbox": [ + 0.371, + 0.089, + 0.471, + 0.24 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.388, + 0.241, + 0.455, + 0.266 + ], + "angle": 0, + "content": "(d) Salience Dropout" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.277, + 0.471, + 0.346 + ], + "angle": 0, + "content": "Figure 2. Segmentation masks for elephant and train using (a) off-the-shelf cross-attention, (b) cross-attention + GradCAM, and (c) cross-attention + GradCam + Salience DropOut (\\$3.2). The naive cross-attention masks are too inclusive whereas GradCAM is too exclusive." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.366, + 0.47, + 0.472 + ], + "angle": 0, + "content": "segmentation (OVSS) from VLMs usually obtain single vector encodings for the visual and text inputs respectively [43, 44, 51, 52, 73, 74]. However, pooling every token into a single vector likely discards information about detailed positions of objects and words. We investigate the use of pretrained cross-attention layers for OVSS, which retain finer-grained correspondence between text and image patches." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.476, + 0.471, + 0.687 + ], + "angle": 0, + "content": "Nevertheless, a naive application of cross-attention between the class name and the image patches lead to overly broad segmentation masks that include irrelevant parts of the image (i.e., over-segmentation, see Fig. 2(b)). To alleviate this, [62] employs gradient information from the imagetext matching loss, GradCAM[57]-style, to sharpen the attention-based masks for the purpose of guiding caption generation. However, this results in masks that capture only the most discriminative regions of an object, such as the head of an elephant (i.e., under-segmentation, Fig. 2(c)). To acquire complete object masks, we propose Salience DropOut, which iteratively drops the image patches with high GradCAM attention scores, forcing the model to attend to less discriminative but relevant object parts." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.69, + 0.47, + 0.902 + ], + "angle": 0, + "content": "Another important consideration is the cross-attention layer and the attention head to extract the masks from. These hyperparameters have enormous influence on the final results and are traditionally tuned on a validation set with pixel-level mask annotations. To eliminate the need for dense annotations, we propose an weakly-supervised reward function based on CLIP [49]. On a validation set with images as well as object class names, the technique contrasts the extracted object regions with a blank image. If, according to CLIP, the former is more similar to the corresponding class name than the latter, we increment the reward. All hyperparameter tuning of our technique is performed with a simple random search with this reward, leading to high performance." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.303 + ], + "angle": 0, + "content": "In summary, we propose Plug-and-Play Open-vocabulary Semantic Segmentation (PnP-OVSS), an extremely simple and training-free framework to extract semantic segmentations from VLMs. At zero extra training cost, PnP-OVSS can be used with any pretrained VLM with text-to-image cross attention layer and an image-text matching loss. It has zero reliance on pixel-level annotations, including a validation set for hyperparameter tuning. At the same time, PnP-OVSS delivers excellent performance. It not only beats the training-free baseline with remarkable margins \\((+29.4\\%)\\) mIoU on Pascal VOC, \\(+13.2\\%\\) on Pascal Context, \\(+14.0\\%\\) on MS COCO, \\(+2.4\\%\\) on COCO Stuff), but also outperforms most recent techniques within the past two years that require extensive finetuning on top of the VLM pretraining." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.304, + 0.822, + 0.319 + ], + "angle": 0, + "content": "With this paper, we make three contributions:" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.32, + 0.901, + 0.365 + ], + "angle": 0, + "content": "- We propose to combine text-to-image attention, GradCAM, and Salience DropOut to iteratively acquire accurate segmentation of arbitrary classes from pretrained VLM." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.366, + 0.892, + 0.44 + ], + "angle": 0, + "content": "- We replace the densely annotated validation set for hyperparameter tuning, which is needed by most existing methods, with a contrastive reward function based on CLIP. This reward function, coupled with random search, finds a good set of hyperparameters for OVSS." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.441, + 0.892, + 0.502 + ], + "angle": 0, + "content": "- The proposed method, PnP-OVSS, is simple to use, requires no extra finetuning, and delivers high performance. Its success hints at a new direction for open-vocabulary segmentation tasks leveraging large VLMs." + }, + { + "type": "list", + "bbox": [ + 0.5, + 0.32, + 0.901, + 0.502 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.518, + 0.642, + 0.533 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.543, + 0.816, + 0.56 + ], + "angle": 0, + "content": "2.1. Large-Scale Vision-Language Model" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.567, + 0.895, + 0.825 + ], + "angle": 0, + "content": "Large-scale vision-language models (VLMs), trained on millions of image-text pairs, have become the foundation for many multimodal tasks. Architecturally, the straightforward approach to training such methods involves aligning visual and textual latent representations via a simple dot product [24, 49]. However, this is insufficient for complex structured tasks like visual question answering or image captioning, which require specialized approaches that employ separate encoders before cross-attention between modalities [33, 35, 37, 78, 79] or self-attention networks over all tokens from both modalities [10, 27, 39, 40, 60]. Another design dimension when training VLMs is the loss function. Commonly used losses include image-text contrastive learning [34, 35, 37, 50, 60, 81], image-text matching (ITM) [27, 34, 35, 37, 60, 72, 78, 81, 82], prediction of masked tokens or patches [34, 60, 78, 81], and language modeling [1, 7, 34, 37]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.826, + 0.894, + 0.903 + ], + "angle": 0, + "content": "This work utilizes models with unimodal encoders followed by cross-attention fusion [34, 35, 37, 72, 78, 81, 82], as they work with high-level features, and accurately attend to the appropriate image patches. Additionally, we utilize the gradient from the ITM loss in the GradCAM step (§3.1)" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4030" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.115, + 0.089, + 0.434, + 0.41 + ], + "angle": 0, + "content": "
MethodPT Data SizeFT data size
Requires finetuning on image-text pairs
OVSegmentor [74]-4.3M
Vil-Seg [43]400M412M
GroupVit* [73]-3.4M
GroupVit [73]-26M
CLIPpy [51]-134M
SegClip [44]400M3.4M
ViewCo [52]-26M
TCL [6]400M15M
PACL [46]400M30M
Requires finetuning but not image-text pairs
MaskClip w/ ST [89]400M1.2M
ZeroSeg* [8]400M3.4M
ZeroSeg [8]400M1.2M
Requires no finetuning
MaskClip[89]400M0
Reco[58]400M0
PnP-OVSS (Ours)
+ BLIP129M0
+ BridgeTower400M+ 4M0
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.419, + 0.472, + 0.519 + ], + "angle": 0, + "content": "Table 1. Training data of current Zero-shot semantic segmentation methods with only text supervision. PT stands for pretraining with image-caption data, FT stands for fine-tuning with image-caption data. ST stands for self training. We list only the image-caption data used for pretraining and all type of data for finetuning. For hyperparameter tuning, our method uses CLIP-L/14 pretrained with 400M data to calculate the reward." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.546, + 0.291, + 0.561 + ], + "angle": 0, + "content": "to sharpen segmentation masks." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.573, + 0.373, + 0.589 + ], + "angle": 0, + "content": "2.2. Zero-shot Semantic Segmentation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.597, + 0.47, + 0.718 + ], + "angle": 0, + "content": "Zero-shot semantic segmentation predicts a dense segmentation mask for any object class described by a given text prompt, with only prior exposure to class-agnostic image-level supervision. This contrasts with weakly supervised semantic segmentation [3, 13, 23, 26, 28, 30, 54, 61, 69, 69, 83, 86, 87], which relies on class-specific annotations, and unsupervised object discovery [12, 14, 56, 59, 63, 64, 67, 68], which identifies the sole object in the foreground." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.719, + 0.47, + 0.856 + ], + "angle": 0, + "content": "Traditional methods [4, 18, 47, 71] train a classifier to distinguish between seen and unseen visual features, wherein the unseen visual features are obtained from a generative model, trained on pairs of seen class image-text embeddings. Recently, methods additionally leverage knowledge from VLMs to attain better matching of visual and textual features [17, 32, 41, 76, 89]. Concretely, they train the segmentation network with dense annotations, while replacing a part of the framework with components from VLMs." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.47, + 0.901 + ], + "angle": 0, + "content": "Recently, methods have explored the use of additional supervision to further reduce the need for pixel level annotations. Prior work is compared in Table 1." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.273 + ], + "angle": 0, + "content": "Models Finetuned on Image-Text Pairs. Methods proposed in [6, 43, 44, 46, 51, 52, 73, 74] require paired image-text annotations to adapt to the zero-shot segmentation task. Specifically, approaches in [43, 44, 51, 52, 73, 74] cluster semantically similar pixels by contrasting grouped region embeddings with textual embeddings, usually via the use of contrastive and self-supervised losses. PACL [46] modifies the constrastive loss to operate over aggregated patch embeddings (instead of image embeddings) to encourage better alignment between image patches and text. TCL [6] achieves a similar patch-text alignment by introducing an additional module to extract text-grounded image regions." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.278, + 0.892, + 0.414 + ], + "angle": 0, + "content": "Models Finetuned on Pseudo-labels. Methods in [8, 89] do not require additional image-text supervision, but involve additional fine-tuning to adapt to the segmentation task. ZeroSeg [8] attempts to match its own group embeddings with multi-scaled segment embeddings obtained from CLIP [49]. MaskClip w/ ST [89] modifies the global attention pooling layer within CLIP to output segmentation masks, which are used as pseudo-labels to train a segmentation network." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.416, + 0.892, + 0.568 + ], + "angle": 0, + "content": "In comparison, our proposed PnP-OVSS does not require any fine-tuning or additional paired image-text annotations. It can directly distill high quality open vocabulary semantic segmentations from any VLM with direct text-to-image cross-attention and an image-text matching loss [34, 35, 37, 72, 78, 81, 82]. Compared to approaches in [58, 89] that perform zero-shot open vocabulary semantic segmentation under a similar no training and no additional annotations paradigm, PnP-OVSS achieves considerably superior performance." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.585, + 0.591, + 0.601 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.611, + 0.892, + 0.718 + ], + "angle": 0, + "content": "PnP-OVSS has four major steps. First, we extract a cross-attention salience map per object class from a VLM. Second, we sharpen the salience map by weighing it with the ITM gradient in the style of GradCAM. Third, we apply Salience DropOut that iteratively completes the salience maps. Fourth, we apply Dense CRF [29] for fine-grained adjustment. The process is illustrated in Fig. 3 and Fig. 4." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.719, + 0.892, + 0.87 + ], + "angle": 0, + "content": "In the first step, we feed an image and a text prompt to the pretrained VLM and extract cross-attention maps. The text prompt is \"A picture of [class 1] [class 2] ... [class K]\", which includes all \\(K\\) class names of interest in the dataset. The image is divided into \\(P \\times P\\) patches. The text prompt and the image first go through the modality-specific encoders respectively, followed by a cross-attention fusion module. In the cross-attention layers, the text encodings serve as query vectors and the image patche encodings are as key and value vectors." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.892, + 0.901 + ], + "angle": 0, + "content": "We extract an attention map for each text token to the image patches from a particular cross-attention layer and an" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "4031" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.102, + 0.089, + 0.881, + 0.398 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.413, + 0.896, + 0.496 + ], + "angle": 0, + "content": "Figure 3. The first iteration with cross-attention + GradCAM + Salience DropOut. The text prompt contains \\( K \\) class names and the image contains \\( P \\times P \\) patches. From a cross-attention layer and an attention head in the pretrained VLM, we obtain \\( K \\) attention score maps of size \\( P \\times P \\), which are sharpened by GradCAM using gradients from the image-text-matching (ITM) loss. To get more complete predictions, we perform Salience Dropout, which repeatedly zero out image patches of the highest average scores and feeds the remaining patches to the image encoder again, forcing the model to attend to other less discriminative patches. We show example salience maps from all iterations in Fig. 4." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.523, + 0.47, + 0.628 + ], + "angle": 0, + "content": "attention head. Note that different layers and heads lead to drastic performance differences, and the choices are hyperparameters, tuned using the procedure in §3.4. We exclude attention maps for the first three tokens \"A\", \"picture\", \"of\", which do not describe semantic classes. If a class name contains two or more tokens, we take the mean attention map. This procedure yields an attention tensor of size \\( K \\times P \\times P \\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.63, + 0.471, + 0.78 + ], + "angle": 0, + "content": "With the correct layer and attention head, we observe that this attention map, when normalized by softmax along the first dimension \\( K \\), can provide passable semantic segmentations. However, it tends to include many patches unrelated to the class being segmented, leading to over-segmentation. Hence, we introduce two refinement steps, GradCAM and Salience DropOut, explained in §3.1 and §3.2. After these steps, we acquire aggregate salience maps for every class. Finally, we conduct local polish to the resultant salience maps using Dense CRF, as described in §3.3." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.787, + 0.37, + 0.803 + ], + "angle": 0, + "content": "3.1. Map Sharpening with GradCAM" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.811, + 0.471, + 0.902 + ], + "angle": 0, + "content": "The off-the-shelf attention maps tend to also cover many patches unrelated to the class name (See Fig. 2 (b)). Prior works [35, 62] further sharpen the attention maps and focus them on class-discriminative regions using a variant of GradCAM [57], which is originally proposed for convolutional networks, but applicable to attention maps. It is worth" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.523, + 0.892, + 0.553 + ], + "angle": 0, + "content": "noting that [35, 62] use the technique for purposes other than semantic segmentation." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.554, + 0.893, + 0.72 + ], + "angle": 0, + "content": "The GradCAM method requires a gradient. Here we leverage the image-text matching (ITM) loss, which trains the VLM to classify if an image-text pair match each other or not. Computing the ITM loss requires a label. We use \"matching\" (as opposed to \"not matching\") as the label and compute the gradient of the loss with respect to the attention score. This is equivalent to asking: which attention scores contribute the most to the decision that the image-text pair is matching? Formally, we denote a \\( P \\times P \\) attention map for class \\( k \\) as \\( M^{(k)} \\) and the ITM loss as \\( \\mathcal{L}_{\\mathrm{ITM}} \\). The GradCAM class salience map is computed as" + }, + { + "type": "equation", + "bbox": [ + 0.574, + 0.731, + 0.892, + 0.765 + ], + "angle": 0, + "content": "\\[\n\\widetilde {M} ^ {(k)} = \\max \\left(0, \\frac {\\partial \\mathcal {L} _ {\\mathrm {I T M}}}{\\partial M ^ {(k)}}\\right) \\otimes M ^ {(k)}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.776, + 0.892, + 0.807 + ], + "angle": 0, + "content": "where \\(\\otimes\\) denote the component-wise multiplication and \\(\\max (\\cdot)\\) is also applied component-wise." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.817, + 0.673, + 0.833 + ], + "angle": 0, + "content": "3.2. Salience Dropout" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.893, + 0.902 + ], + "angle": 0, + "content": "As illustrated in Fig. 2 (c), segmentations generated by the GradCAM-style re-weighting of cross-attention are often narrowly focused on the most discriminative regions for a given class. However, the less discriminative regions are" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4032" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.092, + 0.096, + 0.277, + 0.108 + ], + "angle": 0, + "content": "GradCAM Class Salience Map" + }, + { + "type": "image", + "bbox": [ + 0.084, + 0.111, + 0.165, + 0.168 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.175, + 0.11, + 0.261, + 0.169 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.27, + 0.111, + 0.355, + 0.169 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.363, + 0.111, + 0.45, + 0.169 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.102, + 0.172, + 0.209, + 0.185 + ], + "angle": 0, + "content": "\\(1^{\\mathrm{st}}\\) Salience Drop" + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.189, + 0.168, + 0.248 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.18, + 0.189, + 0.265, + 0.247 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.27, + 0.189, + 0.355, + 0.247 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.363, + 0.189, + 0.449, + 0.247 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.1, + 0.251, + 0.21, + 0.263 + ], + "angle": 0, + "content": "\\(2^{\\text{nd}}\\) Salience Drop" + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.267, + 0.166, + 0.325 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.18, + 0.267, + 0.262, + 0.323 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.27, + 0.267, + 0.355, + 0.324 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.363, + 0.267, + 0.452, + 0.324 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.102, + 0.328, + 0.209, + 0.34 + ], + "angle": 0, + "content": "\\(3^{\\text{rd}}\\) Salience Drop" + }, + { + "type": "image", + "bbox": [ + 0.084, + 0.345, + 0.169, + 0.404 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.177, + 0.345, + 0.262, + 0.404 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.27, + 0.345, + 0.355, + 0.404 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.364, + 0.345, + 0.452, + 0.404 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.101, + 0.407, + 0.209, + 0.42 + ], + "angle": 0, + "content": "\\(4^{\\text{th}}\\) Salience Drop" + }, + { + "type": "image", + "bbox": [ + 0.084, + 0.424, + 0.17, + 0.483 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.177, + 0.424, + 0.261, + 0.483 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.27, + 0.424, + 0.354, + 0.483 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.364, + 0.424, + 0.451, + 0.483 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.227, + 0.49, + 0.311, + 0.503 + ], + "angle": 0, + "content": "Thresholding" + }, + { + "type": "image_caption", + "bbox": [ + 0.225, + 0.514, + 0.313, + 0.526 + ], + "angle": 0, + "content": "Gaussian Blur" + }, + { + "type": "image_caption", + "bbox": [ + 0.234, + 0.538, + 0.303, + 0.55 + ], + "angle": 0, + "content": "Dense CRF" + }, + { + "type": "image", + "bbox": [ + 0.084, + 0.558, + 0.169, + 0.619 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.176, + 0.558, + 0.261, + 0.619 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.269, + 0.558, + 0.355, + 0.619 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.363, + 0.558, + 0.451, + 0.619 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.631, + 0.47, + 0.701 + ], + "angle": 0, + "content": "Figure 4. An illustration of Salience DropOut, showing GradCAM salience values after each iteration. Black squares in the images indicate dropped patches. We obtain the final result by summing the salience maps from all iterations and applying thresholding, Gaussian blur, and Dense CRF." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.729, + 0.47, + 0.774 + ], + "angle": 0, + "content": "still important for the completeness of masks. To compel the VLM to attend to these regions, we propose an iterative technique called Salience DropOut." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.775, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Since our task is zero-shot and open-vocabulary, we do not have prior knowledge of the classes present in the image. Hence, we sum up the salience maps \\(\\widetilde{M}^{(k)}\\) over all classes \\(k\\), yielding a class-agnostic salience map \\(U^{(t)}\\) for the \\(t^{\\text{th}}\\) Salience DropOut iteration, \\(U^{(t)} = \\sum_{k=1}^{K} \\widetilde{M}^{(k,t)}\\), where \\(\\widetilde{M}^{(k,t)}\\) is the GradCAM salience map for class \\(k\\) after the \\(t^{\\text{th}}\\) iteration. Next, we zero out the \\(50\\%\\) of image patches with the highest values in \\(U^{(t)}\\). On the remaining" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.089, + 0.892, + 0.136 + ], + "angle": 0, + "content": "image, we compute the GradCAM salience maps \\(\\widetilde{M}^{(k,t + 1)}\\) and their sum, \\(U^{(t + 1)}\\). Any image patch previously zeroed out will always receive zero salience in later iterations." + }, + { + "type": "text", + "bbox": [ + 0.501, + 0.137, + 0.892, + 0.168 + ], + "angle": 0, + "content": "Formally, the set of remaining image patches \\( S^{(t)} \\subseteq \\{1, \\dots, P\\}^2 \\) after the \\( t^{\\text{th}} \\) dropout iteration is defined as" + }, + { + "type": "equation", + "bbox": [ + 0.548, + 0.18, + 0.892, + 0.2 + ], + "angle": 0, + "content": "\\[\nS ^ {(t)} = S ^ {(t - 1)} \\backslash \\{(i, j) \\mid U _ {i j} ^ {(t)} \\geq \\eta \\}, \\tag {2}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.566, + 0.203, + 0.892, + 0.228 + ], + "angle": 0, + "content": "\\[\n\\eta = \\operatorname {m e d i a n} \\left(\\left\\{U _ {i j} ^ {(t)} \\mid (i, j) \\in S ^ {(t - 1)} \\right\\}\\right), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.242, + 0.892, + 0.303 + ], + "angle": 0, + "content": "where \\( U_{i,j}^{(t)} \\) denotes the aggregate salience value of the image patch at row \\( i \\) and column \\( j \\). Additionally, note that the input to the first dropout iteration \\( S^0 = \\{1,\\dots ,P\\} ^2 \\) is the set of all image patches." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.304, + 0.892, + 0.367 + ], + "angle": 0, + "content": "We stop at four rounds of dropout, as almost all \\((93.75\\%)\\) patches are removed beyond that point. The final output for each class \\(k\\) is the sum over all salience maps across the four dropout iterations, \\(\\hat{M}^{(k)} = \\sum_{t=1}^{4} \\widetilde{M}^{(k,t)}\\)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.375, + 0.771, + 0.39 + ], + "angle": 0, + "content": "3.3. Gaussian Blur and Dense CRF" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.399, + 0.892, + 0.58 + ], + "angle": 0, + "content": "The salience dropout procedure generates \\( k \\) continuous-valued salience maps, one per object class. To filter out small random noise in the salience values, we subsequent apply a straightforward thresholding operation at a predefined value \\( T \\) on the salience values and obtain binary segmentation masks. Nonetheless, the hard thresholding creates jagged segmentations with sharp edges that often do not coincide with object boundaries. One common strategy in zero-shot segmentation is to apply Dense Conditional Random Field (CRF) [29], which makes fine-grained adjustments to the estimated masks by enforcing consistency between nearby image pixels with similar colors." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.581, + 0.892, + 0.671 + ], + "angle": 0, + "content": "However, we find that the hard 0/1 labels in the binary masks do not work well as pixel unary potentials for Dense CRF. Hence, we smooth them using a Gaussian kernel with a preset variance \\(\\sigma\\), which results in a better initialization for unary terms and accounts for uncertainty of exact segmentation boundary along the patch boundaries." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.681, + 0.726, + 0.698 + ], + "angle": 0, + "content": "3.4. Hyperparameter Tuning" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.704, + 0.892, + 0.856 + ], + "angle": 0, + "content": "Three hyperparameters in PnP-OVSS have the strongest influence on the result, the cross-attention layer \\( L \\), the attention head \\( H \\), and the binary threshold \\( T \\). Traditionally, tuning these hyperparameters requires a validation set with pixel-level labels. However, since our goal is to perform zero-shot open-vocabulary semantic segmentation, this requirement could potentially limit the applicability of the technique. Instead, we propose a weakly supervised reward function for hyperparameter tuning, which only requires a set of images and the class names appearing in each image." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.856, + 0.892, + 0.901 + ], + "angle": 0, + "content": "The reward for an image \\(I\\) is calculated as follows. We start with a set of classes present in the image, denoted as \\(\\mathcal{K}(I)\\). For each class \\(k\\in \\mathcal{K}(I)\\), we obtain a segmentation" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4033" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.09, + 0.473, + 0.213 + ], + "angle": 0, + "content": "mask \\( M^{(k)} \\) (which can be the GradCAM mask, the Salience DropOut mask, or the Dense CRF mask). Next, we apply the mask to the image \\( I \\) and input the extracted regions \\( M^{(k)} \\otimes I \\) into a pretrained neural network \\( f \\), which takes an image and a textual class name as input and produces a similarity score. We calculate the normalized probability that the masked image \\( M^{(k)} \\otimes I \\) belongs to the ground-truth class \\( k \\) and contrast with a completely black image \\( \\mathbf{0} \\)." + }, + { + "type": "equation", + "bbox": [ + 0.097, + 0.218, + 0.47, + 0.252 + ], + "angle": 0, + "content": "\\[\n\\operatorname {R e w a r d} = \\sum_ {k \\in \\mathcal {K} (I)} \\mathbb {1} \\left[ P r \\left(M ^ {(k)} \\otimes I, k\\right) > P r (\\mathbf {0}, k) \\right], \\tag {4}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.091, + 0.255, + 0.47, + 0.29 + ], + "angle": 0, + "content": "\\[\nP r (I, k) = \\frac {\\exp (f (I , k))}{\\sum_ {k ^ {\\prime} \\in \\mathcal {K} (I)} \\exp (f (I , k ^ {\\prime}))}, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.295, + 0.47, + 0.37 + ], + "angle": 0, + "content": "where \\(\\mathbb{1}(\\cdot)\\) is the indicator function. Intuitively, a reward of 1 is assigned if and only if the image features pulled with the estimated mask for class \\(k\\) bears higher similarity to the class name of \\(k\\) than a black image (which can be interpreted as the prior probability of class \\(k\\))." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.371, + 0.47, + 0.446 + ], + "angle": 0, + "content": "We sum up the reward for all validation images as the total reward. The best hyperparameters, including the cross-attention layer, the attention head, the threshold \\( T \\), and the variance in the Gaussian blur kernel, are determined using a simple random search." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.457, + 0.21, + 0.474 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.482, + 0.398, + 0.499 + ], + "angle": 0, + "content": "4.1. Datasets and Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.505, + 0.47, + 0.672 + ], + "angle": 0, + "content": "Following the previous work for zero-shot semantic segmentation, we adopt validation sets of Pascal-VOC 2012 [16], Pascal Context [45], COCO Object [42], COCO Stuff [5], and ADE20K [88] that contain, respectively, 20 object classes, 59 object and stuff classes, 80 object classes, 171 object and stuff classes, and 150 object and stuff classes to evaluate our framework. To verify its versatility, we apply PnP-OVSSto two high-performance VLMs, BLIP [37] and BridgeTower [78], which have the ITM loss and text-to-image cross-attention. More details are in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.68, + 0.469, + 0.803 + ], + "angle": 0, + "content": "Hyperparameter tuning. We use CLIP VIT-L/14 to calculate the reward. The input resolution is \\(336 \\times 336\\). The search spaces of the random search and results are shown in Tab. 2. For computation efficiency, we tune the layer, the head, and the threshold with GradCAM masks. With the first three hyperparameters fixed, we tune the Gaussian variance on masks before Dense CRF. We directly adopt Dense CRF hyperparameters from CutLER [66] without tuning." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.81, + 0.47, + 0.903 + ], + "angle": 0, + "content": "Baselines. We adopt recent papers on zero-shot open-vocabulary semantic segmentation as baselines. The only baselines strictly comparable to our work are MaskClip [89] and Reco [58], which do not perform any finetuning on pretrained VLMs and do not perform hyperparameter tuning on dense annotations. We call these Group" + }, + { + "type": "table", + "bbox": [ + 0.532, + 0.089, + 0.865, + 0.272 + ], + "angle": 0, + "content": "
HyperparametersStartEndStepSolution
BLIP
Layer11218
Head112110
Attention Threshold0.050.50.10.15
BridgeTower
Layer1612
Head11618
Attention Threshold0.050.50.10.15
Gaussian Blur
Standard Deviation0.010.110.020.05
" + }, + { + "type": "table_caption", + "bbox": [ + 0.569, + 0.281, + 0.824, + 0.296 + ], + "angle": 0, + "content": "Table 2. Search space for hyperparameters" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.322, + 0.892, + 0.489 + ], + "angle": 0, + "content": "3. To further expand our scope, we also include two other groups of baselines. Group 2 finetunes VLMs but does not require image-text pairs, including MaskClip with self-training (ST) [89] and ZeroSeg [8]. We include ZeroSeg, trained with ImageNet1K, and ZeroSeg*, trained with CC3M+COCO. Group 1 contains baselines that require training on image-text pairs. For completeness, we also compare against supervised techniques since 2019. To maintain the zero-shot setting, we test them on classes not observed during training. For more details, we refer readers to the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.499, + 0.639, + 0.515 + ], + "angle": 0, + "content": "4.2. Main Results" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.522, + 0.892, + 0.599 + ], + "angle": 0, + "content": "We show the main results in Tab. 3. As input resolution may influence results and cause unfair comparisons, we label the resolution used by each method. PACL [46] uses an 224 resolution but changes the stride for image patchification from 16 to 4, hence introducing overlapping patches." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.599, + 0.892, + 0.764 + ], + "angle": 0, + "content": "PnP-OVSS exhibits excellent performance. Comparing with MaskClip [89] and Reco [58], the two methods that require no additional training and ground truth for hyperparameter tuning, on an equal-resolution basis, we attain \\(+29.4\\%\\) mIoU on Pascal VOC, \\(+13.2\\%\\) mIoU on Pascal Context, \\(+14.0\\%\\) mIoU on COCO Object, and \\(+11.4\\%\\) mIoU on ADE-20K. Further, PnP-OVSS surpasses all baselines in Group 2. On an equal-resolution basis, we achieve \\(+13.7\\%\\) mIoU on Pascal Voc, \\(+9.6\\%\\) on Pascal Context, \\(+9.5\\%\\) on COCO Object, \\(+11.6\\%\\) on COCO Stuff, and \\(+12.8\\%\\) on ADE-20K." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.892, + 0.902 + ], + "angle": 0, + "content": "When compared to Group 1, PnP-OVSS still outperforms most. On the Pascal datasets, under equal resolutions, PnP-OVSS + BLIPFlickr outperforms 6 out of 10 baselines on Pascal VOC, and 8 out of 9 baselines on Pascal Context. For the COCO datasets, under equal resolutions, PnP-OVSS + BLIPFlickr beats 5 out of 8 baselines on COCO Object, and all baselines except PACL [46] on COCO Stuff. The two Pascal datasets share many images, and the two COCO datasets use exactly the same images, but they have" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4034" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.086, + 0.088, + 0.887, + 0.593 + ], + "angle": 0, + "content": "
MethodFinetuning VLMsHT on Dense LabelsShort-side ResolutionPascal VOC-20Pascal Context-59COCO Object-80COCO Stuff-171ADE 20K-150
Group 1: Methods that require weakly supervised finetuning on image-text data
ViL-Seg†[43]✓✓-37.318.9-18.0-
CLIPpy [51]✓✓22452.2-32.025.5*13.5
SegClip [44]✓✓22452.624.726.5--
GroupVit (by [51])✓✓22428.114.812.9-6.2
GroupVit (by [6])✓✓44850.418.727.515.39.2
GroupVit [73]✓✓44852.322.424.3--
ViewCo [52]✓✓44852.423.023.5--
OVSegmentor [74]✓✓44853.820.425.1--
TCL [6] +PAMR [2]✓✓44855.030.431.622.417.1
PACL [46]✓✓224×472.350.1-38.831.4
Group 2: Methods that require finetuning but not real image-text data
MaskClip w/ ST [89]✓✓336-31.1-18.0-
MaskClip w/ ST (by [6])✓✓44838.823.620.616.49.8
ZeroSeg [8]✓✓44840.820.420.2--
Group 3: Methods that require no finetuning
MaskClip (by [51])××22422.1-13.88.16.8
MaskClip [89]××336-25.5-14.6-
Reco [58]××320-27.2-27.2-
Reco (by [6])××44825.119.915.714.811.2
PnP-OVSS with different VLMs
BLIPFlickr××22447.836.424.825.818.2
BLIPFlickr××32051.740.428.029.621.3
BLIPFlickr××33652.540.728.229.621.9
BLIPFlickr××44854.542.229.731.522.6
BLIPFlickr××76854.142.831.832.523.5
BLIPCOCO××76855.741.933.832.623.2
BridgeTower××32236.432.324.227.618.6
BridgeTower××33635.332.424.227.618.0
BridgeTower××77035.232.424.528.019.0
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.603, + 0.895, + 0.675 + ], + "angle": 0, + "content": "Table 3. Zero-shot semantic segmentation performance in mIoU. Group 3 contains the most similar baselines that serve as fair comparisons to PnP-OVSS. Groups 1 and 2 benefit from additional training, extra image-text data, and hyperparameter tuning on dense labels. We use the word \"by\" followed by a paper citation to indicate results of the same technique reported by different papers. \\(\\star\\) CLIPpy tests on 133 categories of COCO Stuff while we test all 171 classes of COCO Stuff. ViL-Seg† is tested on subset of classes on the three datasets, as detailed in the supplementary." + }, + { + "type": "table", + "bbox": [ + 0.195, + 0.685, + 0.778, + 0.851 + ], + "angle": 0, + "content": "
MethodDense \nLabelsHT on Dense \nLabelsPascal \nVOC-20Pascal \nContext-59COCO \nStuff-171
PnP-OVSS + BLIPFlickr (Ours)××53.653.839.8
SPNet+ST [71]✓✓25.8-26.9
ZS3Net+ST [4]✓✓21.220.710.6
CaGNet+ST [18]✓✓30.3-13.4
STRICT [47]✓✓35.6-30.3
LSeg [32]✓✓41.0--
SimBase [76]✓✓72.5-36.3
MaskCLIP+ w/ ST [89]✓✓86.166.754.7
" + }, + { + "type": "table_caption", + "bbox": [ + 0.08, + 0.86, + 0.888, + 0.875 + ], + "angle": 0, + "content": "Table 4. Comparison of zero-shot semantic segmentation performance on unseen categories with methods trained with dense annotation." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4035" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.109, + 0.089, + 0.44, + 0.258 + ], + "angle": 0, + "content": "
Ablated ModelPascal ContextCOCO Stuff
BLIPFlickr19.814.5
+ GradCam21.617.5
+ Drop 125.119.8
+ Drop 226.520.6
+ Drop 327.020.9
+ Drop 427.220.9
+ Drop 4 + Blur36.828.6
+ Drop 4 + Dense CRF35.331.8
+ Drop 4 + Blur + Dense CRF42.832.5
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.268, + 0.47, + 0.297 + ], + "angle": 0, + "content": "Table 5. An ablation study of PnP-OVSS + BLIPFlickr with resolution 768 on Pascal Context and COCO Stuff." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.325, + 0.47, + 0.385 + ], + "angle": 0, + "content": "vastly different lists of classes to segment. We observe that the advantage of PnP-OVSS over methods in Group 1 becomes more pronounced as the number of classes on the same image increases." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.385, + 0.469, + 0.476 + ], + "angle": 0, + "content": "When applied to BridgeTower [78], PnP-OVSS still surpasses all methods in Group 3 by \\(10.2\\%\\) on Pascal VOC, \\(5.1\\%\\) on Pascal Context, \\(8.5\\%\\) on COCO Object, \\(0.4\\%\\) on COCO Stuff, and \\(6.8\\%\\) on ADE-20K. This showcases the plug-and-play ability of PnP-OVSS, which excels with different base networks." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.477, + 0.47, + 0.569 + ], + "angle": 0, + "content": "We report comparisons against supervised methods in Tab 4. PnP-OVSS + BLIPFlickr outperforms 5 out of 7 methods on Pascal VOC, as well as every baseline except MaskCLIP+ on Pascal Context and COCO stuff. As these baselines benefit from dense supervision, the results further demonstrate the strengths of PnP-OVSS." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.58, + 0.231, + 0.596 + ], + "angle": 0, + "content": "4.3. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.604, + 0.47, + 0.816 + ], + "angle": 0, + "content": "We perform gradual ablation of the components of \\(\\mathrm{PnP-OVSS}\\) on \\(\\mathrm{BLIP}_{\\mathrm{Flickr}}\\) and report the results in Tab. 5. Each component, including GradCAM, all Salience DropOut iterations, Gaussian blur, and Dense CRF, contribute positively to the final performance. In particular, the first iteration of Salience DropOut has much larger impact (+3.5/2.3) than the second iteration (+1.4/0.8), which in turn is more important than the rest. Interestingly, Gaussian blur by itself attains good performance (+9.6/7.7) whereas Dense CRF only works well when combined with blur. Dense CRF alone is worse than Gaussian blur by 1.5 mIoU on Pascal Context. This is likely caused by the fact that hard 0/1 labels resulted from thresholding are not informative unary potentials that can be leveraged by CRF effectively." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.827, + 0.329, + 0.844 + ], + "angle": 0, + "content": "4.4. Hyperparameter Sensitivity" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.47, + 0.903 + ], + "angle": 0, + "content": "The choice of hyperparameters often exerts substantial influence on segmentation performance. Here we quantitatively examine how the choice of cross-attention layers" + }, + { + "type": "table", + "bbox": [ + 0.511, + 0.089, + 0.885, + 0.258 + ], + "angle": 0, + "content": "
Layer Mean123456
mIoU11.612.311.311.011.612.9
Layer Mean789101112
mIoU13.725.325.823.412.55.8
Head in Layer 8123456
mIoU16.63.88.620.019.85.2
Head in Layer 8789101112
mIoU11.16.03.829.617.38.0
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.268, + 0.892, + 0.324 + ], + "angle": 0, + "content": "Table 6. Semantic segmentation performance using cross-attention maps averaged across all heads in a layer and separate heads in Layer 8. Results are attained with PnP-OVSS+BLIPFlickr on COCO Stuff and resolution 336." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.356, + 0.892, + 0.416 + ], + "angle": 0, + "content": "and attention heads may change the segmentation mIoU on COCO Stuff. Tab. 6 shows the results obtained from the average cross-attention maps over all heads in each layer and those from different attention heads." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.419, + 0.893, + 0.555 + ], + "angle": 0, + "content": "We make the following observations. First, different layers and heads have drastic performance differences. The best-to-worst difference among all layers is 20, and that among heads in Layer 8 is 23.8, underscoring the importance of hyperparameter tuning. Second, the random search using the proposed reward function correctly identifies the best layer-head combination, even though Layer 8 is not the best layer based on average head performance. This indicates the effectiveness of our method." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.577, + 0.628, + 0.593 + ], + "angle": 0, + "content": "5. Conclusions" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.609, + 0.892, + 0.731 + ], + "angle": 0, + "content": "We propose PnP-OVSS, which extracts the ability of semantic segmentation from opaque VLMs. PnP-OVSS is simple to use, requires no extra finetuning, and delivers high performance, exceeding not only all baselines that requires no finetuning, but also all baselines that do not use image-text pairs in finetuning. Its success hints at a new direction for open-vocabulary segmentation tasks leveraging large VLMs." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.752, + 0.681, + 0.77 + ], + "angle": 0, + "content": "6. Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.78, + 0.892, + 0.903 + ], + "angle": 0, + "content": "This work has been supported by the Nanyang Associate Professorship and the National Research Foundation Fellowship (NRF- NRFF13-2021-0006), Singapore, and by the Vector Institute for AI, Canada CIFAR AI Chair, and NSERC Discovery Grant. Any opinions, findings, conclusions, or recommendations expressed in this material are those of the authors and do not reflect the views of the funding agencies." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.957 + ], + "angle": 0, + "content": "4036" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.079, + 0.09, + 0.176, + 0.107 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.266 + ], + "angle": 0, + "content": "[1] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, Roman Ring, Eliza Rutherford, Serkan Cabi, Tengda Han, Zhitao Gong, Sina Samangooei, Marianne Monteiro, Jacob L Menick, Sebastian Borgeaud, Andy Brock, Aida Nematzadeh, Sahand Sharifzadeh, Mikol aj Binkowski, Ricardo Barreira, Oriol Vinyals, Andrew Zisserman, and Karén Simonyan. Flamingo: a visual language model for few-shot learning. In Advances in Neural Information Processing Systems, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.269, + 0.472, + 0.323 + ], + "angle": 0, + "content": "[2] Nikita Araslanov and Stefan Roth. Single-stage semantic segmentation from image labels. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4253-4262, 2020. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.326, + 0.47, + 0.393 + ], + "angle": 0, + "content": "[3] Loris Bazzani, Alessandra Bergamo, Dragomir Anguelov, and Lorenzo Torresani. Self-taught object localization with deep networks. In 2016 IEEE winter conference on applications of computer vision (WACV), pages 1-9. IEEE, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.396, + 0.47, + 0.436 + ], + "angle": 0, + "content": "[4] Maxime Bucher, Tuan-Hung Vu, Matthieu Cord, and Patrick Pérez. Zero-shot semantic segmentation. Advances in Neural Information Processing Systems, 32, 2019. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.439, + 0.471, + 0.492 + ], + "angle": 0, + "content": "[5] Holger Caesar, Jasper Uijlings, and Vittorio Ferrari. Cocosstuff: Thing and stuff classes in context. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1209-1218, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.495, + 0.471, + 0.563 + ], + "angle": 0, + "content": "[6] Junbum Cha, Jonghwan Mun, and Byungseok Roh. Learning to generate text-grounded mask for open-world semantic segmentation from only image-text pairs. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11165-11174, 2023. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.565, + 0.471, + 0.633 + ], + "angle": 0, + "content": "[7] Jun Chen, Han Guo, Kai Yi, Boyang Li, and Mohamed Elhoseiny. Visualgpt: Data-efficient adaptation of pretrained language models for image captioning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18030-18040, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.635, + 0.47, + 0.73 + ], + "angle": 0, + "content": "[8] Jun Chen, Deyao Zhu, Guocheng Qian, Bernard Ghanem, Zhicheng Yan, Chenchen Zhu, Fanyi Xiao, Sean Chang Culatana, and Mohamed Elhoseiny. Exploring open-vocabulary semantic segmentation from clip vision encoder distillation only. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 699-710, 2023. 1, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.733, + 0.47, + 0.801 + ], + "angle": 0, + "content": "[9] Qi Chen, Lingxiao Yang, Jian-Huang Lai, and Xiaohua Xie. Self-supervised image-specific prototype exploration for weakly supervised semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4288-4298, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.804, + 0.47, + 0.87 + ], + "angle": 0, + "content": "[10] Yen-Chun Chen, Linjie Li, Licheng Yu, Ahmed El Kholy, Faisal Ahmed, Zhe Gan, Yu Cheng, and Jingjing Liu. Uniter: Universal image-text representation learning. In Proceedings of the 16th European Conference on Computer Vision, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.47, + 0.901 + ], + "angle": 0, + "content": "[11] Zhaozheng Chen, Tan Wang, Xiongwei Wu, Xian-Sheng Hua, Hanwang Zhang, and Qianru Sun. Class re-activation" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.094, + 0.892, + 0.134 + ], + "angle": 0, + "content": "maps for weakly-supervised semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 969-978, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.136, + 0.892, + 0.204 + ], + "angle": 0, + "content": "[12] Minsu Cho, Suha Kwak, Cordelia Schmid, and Jean Ponce. Unsupervised object discovery and localization in the wild: Part-based matching with bottom-up region proposals. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1201-1210, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.205, + 0.892, + 0.259 + ], + "angle": 0, + "content": "[13] Ramazan Gokberk Cinbis, Jakob Verbeek, and Cordelia Schmid. Weakly supervised object localization with multifold multiple instance learning. IEEE transactions on pattern analysis and machine intelligence, 39(1):189-203, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.261, + 0.892, + 0.301 + ], + "angle": 0, + "content": "[14] Timothée Darcet, Maxime Oquab, Julien Mairal, and Piotr Bojanowski. Vision transformers need registers. arXiv preprint arXiv:2309.16588, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.303, + 0.892, + 0.37 + ], + "angle": 0, + "content": "[15] Ye Du, Zehua Fu, Qingjie Liu, and Yunhong Wang. Weakly supervised semantic segmentation by pixel-to-prototype contrast. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4320–4329, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.372, + 0.892, + 0.426 + ], + "angle": 0, + "content": "[16] Mark Everingham, Luc Van Gool, Christopher KI Williams, John Winn, and Andrew Zisserman. The pascal visual object classes (voc) challenge. International journal of computer vision, 88:303-338, 2010. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.428, + 0.892, + 0.482 + ], + "angle": 0, + "content": "[17] Golnaz Ghiasi, Xiuye Gu, Yin Cui, and Tsung-Yi Lin. Scaling open-vocabulary image segmentation with image-level labels. In European Conference on Computer Vision, pages 540-557. Springer, 2022. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.484, + 0.892, + 0.55 + ], + "angle": 0, + "content": "[18] Zhangxuan Gu, Siyuan Zhou, Li Niu, Zihan Zhao, and Liqing Zhang. Context-aware feature generation for zero-shot semantic segmentation. In Proceedings of the 28th ACM International Conference on Multimedia, pages 1921-1929, 2020. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.553, + 0.892, + 0.635 + ], + "angle": 0, + "content": "[19] Jiaxian Guo, Junnan Li, Dongxu Li, Anthony Meng Huat Tiong, Boyang Li, Dacheng Tao, and Steven Hoi. From images to textual prompts: Zero-shot visual question answering with frozen large language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10867-10877, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.637, + 0.892, + 0.691 + ], + "angle": 0, + "content": "[20] Yanming Guo, Yu Liu, Theodoros Georgiou, and Michael S Lew. A review of semantic segmentation using deep neural networks. International journal of multimedia information retrieval, 7:87-93, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.693, + 0.892, + 0.774 + ], + "angle": 0, + "content": "[21] Kunyang Han, Yong Liu, Jun Hao Liew, Henghui Ding, Jiajun Liu, Yitong Wang, Yansong Tang, Yujiu Yang, Jiashi Feng, Yao Zhao, et al. Global knowledge calibration for fast open-vocabulary segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 797-807, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.776, + 0.892, + 0.815 + ], + "angle": 0, + "content": "[22] Shijie Hao, Yuan Zhou, and Yanrong Guo. A brief survey on semantic segmentation with deep learning. Neurocomputing, 406:302-321, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.818, + 0.892, + 0.87 + ], + "angle": 0, + "content": "[23] Qibin Hou, PengTao Jiang, Yunchao Wei, and Ming-Ming Cheng. Self-erasing network for integral object attention. Advances in Neural Information Processing Systems, 31, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.873, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[24] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.094, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4037" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.11, + 0.093, + 0.469, + 0.146 + ], + "angle": 0, + "content": "Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In Proceedings of the 38th International Conference on Machine Learning, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.149, + 0.469, + 0.218 + ], + "angle": 0, + "content": "[25] Peng-Tao Jiang, Yuqi Yang, Qibin Hou, and Yunchao Wei. L2g: A simple local-to-global knowledge transfer framework for weakly supervised semantic segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16886-16896, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.219, + 0.469, + 0.272 + ], + "angle": 0, + "content": "[26] Tsung-Wei Ke, Jyh-Jing Hwang, and Stella X Yu. Universal weakly supervised segmentation by pixel-to-segment contrastive learning. arXiv preprint arXiv:2105.00957, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.274, + 0.469, + 0.328 + ], + "angle": 0, + "content": "[27] Wonjae Kim, Bokyung Son, and Ildoo Kim. Vilt: Vision-and-language transformer without convolution or region supervision. In Proceedings of the 38th International Conference on Machine Learning, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.33, + 0.469, + 0.412 + ], + "angle": 0, + "content": "[28] Alexander Kolesnikov and Christoph H Lampert. Seed, expand and constrain: Three principles for weakly-supervised image segmentation. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part IV 14, pages 695-711. Springer, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.414, + 0.469, + 0.467 + ], + "angle": 0, + "content": "[29] Philipp Krahenbuhl and Vladlen Koltun. Efficient inference in fully connected crfs with gaussian edge potentials. Advances in neural information processing systems, 24, 2011. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.469, + 0.469, + 0.537 + ], + "angle": 0, + "content": "[30] Krishna Kumar Singh and Yong Jae Lee. Hide-and-seek: Forcing a network to be meticulous for weakly-supervised object and action localization. In Proceedings of the IEEE International Conference on Computer Vision, pages 3524-3533, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.539, + 0.469, + 0.607 + ], + "angle": 0, + "content": "[31] Jungbeom Lee, Seong Joon Oh, Sangdoo Yun, Junsuk Choe, Eunji Kim, and Sungroh Yoon. Weakly supervised semantic segmentation using out-of-distribution data. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16897-16906, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.608, + 0.469, + 0.649 + ], + "angle": 0, + "content": "[32] Boyi Li, Kilian Q. Weinberger, Serge Belongie, Vladlen Koltun, and René Ranftl. Language-driven semantic segmentation, 2022. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.65, + 0.469, + 0.747 + ], + "angle": 0, + "content": "[33] Chenliang Li, Haiyang Xu, Junfeng Tian, Wei Wang, Ming Yan, Bin Bi, Jiabo Ye, He Chen, Guohai Xu, Zheng Cao, Ji Zhang, Songfang Huang, Fei Huang, Jingren Zhou, and Luo Si. mPLUG: Effective and efficient vision-language learning by cross-modal skip-connections. In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.748, + 0.469, + 0.815 + ], + "angle": 0, + "content": "[34] Chenliang Li, Haiyang Xu, Junfeng Tian, Wei Wang, Ming Yan, Bin Bi, Jiabo Ye, Hehong Chen, Guohai Xu, Zheng Cao, et al. mplug: Effective and efficient vision-language learning by cross-modal skip-connections. arXiv preprint arXiv:2205.12005, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.818, + 0.469, + 0.871 + ], + "angle": 0, + "content": "[35] Junnan Li, Ramprasaath R. Selvaraju, Akhilesh Deepak Gotmare, Shafiq Joty, Caiming Xiong, and Steven Hoi. Align before fuse: Vision and language representation learning with momentum distillation. In NeurIPS, 2021. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[36] Jinlong Li, Zequn Jie, Xu Wang, Xiaolin Wei, and Lin Ma. Expansion and shrinkage of localization for weakly-" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.093, + 0.469, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "supervised semantic segmentation. Advances in Neural Information Processing Systems, 35:16037-16051, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.123, + 0.892, + 0.19 + ], + "angle": 0, + "content": "[37] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In International Conference on Machine Learning, pages 12888-12900. PMLR, 2022. 1, 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.193, + 0.892, + 0.247 + ], + "angle": 0, + "content": "[38] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In ICML, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.251, + 0.892, + 0.348 + ], + "angle": 0, + "content": "[39] Wei Li, Can Gao, Guocheng Niu, Xinyan Xiao, Hao Liu, Jiachen Liu, Hua Wu, and Haifeng Wang. UNIMO: Towards unified-modal understanding and generation via cross-modal contrastive learning. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.35, + 0.892, + 0.417 + ], + "angle": 0, + "content": "[40] Xiujun Li, Xi Yin, Chunyuan Li, Pengchuan Zhang, Xiaowei Hu, Lei Zhang, Lijuan Wang, Houdong Hu, Li Dong, Furu Wei, Yejin Choi, and Jianfeng Gao. Oscar: Object-semantics aligned pre-training for vision-language tasks. In ECCV 2020, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.421, + 0.892, + 0.502 + ], + "angle": 0, + "content": "[41] Feng Liang, Bichen Wu, Xiaoliang Dai, Kunpeng Li, Yinan Zhao, Hang Zhang, Peizhao Zhang, Peter Vajda, and Diana Marculescu. Open-vocabulary semantic segmentation with mask-adapted clip. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7061-7070, 2023. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.506, + 0.892, + 0.589 + ], + "angle": 0, + "content": "[42] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.591, + 0.892, + 0.659 + ], + "angle": 0, + "content": "[43] Quande Liu, Youpeng Wen, Jianhua Han, Chunjing Xu, Hang Xu, and Xiaodan Liang. Open-world semantic segmentation via contrasting and clustering vision-language embedding. In European Conference on Computer Vision, pages 275-292. Springer, 2022. 2, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.662, + 0.892, + 0.73 + ], + "angle": 0, + "content": "[44] Huaishao Luo, Junwei Bao, Youzheng Wu, Xiaodong He, and Tianrui Li. Segclip: Patch aggregation with learnable centers for open-vocabulary semantic segmentation. In International Conference on Machine Learning, pages 23033-23044. PMLR, 2023. 1, 2, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.733, + 0.892, + 0.815 + ], + "angle": 0, + "content": "[45] Roozbeh Mottaghi, Xianjie Chen, Xiaobai Liu, Nam-Gyu Cho, Seong-Whan Lee, Sanja Fidler, Raquel Urtasun, and Alan Yuille. The role of context for object detection and semantic segmentation in the wild. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 891–898, 2014. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.818, + 0.892, + 0.899 + ], + "angle": 0, + "content": "[46] Jishnu Mukhoti, Tsung-Yu Lin, Omid Poursaeed, Rui Wang, Ashish Shah, Philip HS Torr, and Ser-Nam Lim. Open vocabulary semantic segmentation with patch aligned contrastive learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19413-19423, 2023. 1, 3, 6, 7" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.892, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.956 + ], + "angle": 0, + "content": "4038" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.174 + ], + "angle": 0, + "content": "[47] Giuseppe Pastore, Fabio Cermelli, Yongqin Xian, Massimiliano Mancini, Zeynep Akata, and Barbara Caputo. A closer look at self-training for zero-label semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2693-2702, 2021. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.176, + 0.472, + 0.259 + ], + "angle": 0, + "content": "[48] Jie Qin, Jie Wu, Pengxiang Yan, Ming Li, Ren Yuxi, Xuefeng Xiao, Yitong Wang, Rui Wang, Shilei Wen, Xin Pan, et al. Freeseg: Unified, universal and open-vocabulary image segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19446-19455, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.26, + 0.472, + 0.356 + ], + "angle": 0, + "content": "[49] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In Proceedings of the 38th International Conference on Machine Learning, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.358, + 0.472, + 0.44 + ], + "angle": 0, + "content": "[50] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.442, + 0.472, + 0.51 + ], + "angle": 0, + "content": "[51] Kanchana Ranasinghe, Brandon McKinzie, Sachin Ravi, Yinfei Yang, Alexander Toshev, and Jonathon Shlens. Perceptual grouping in contrastive vision-language models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5571-5584, 2023. 2, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.511, + 0.472, + 0.58 + ], + "angle": 0, + "content": "[52] Pengzhen Ren, Changlin Li, Hang Xu, Yi Zhu, Guangrun Wang, Jianzhuang Liu, Xiaojun Chang, and Xiaodan Liang. Viewco: Discovering text-supervised segmentation masks via multi-view semantic consistency. arXiv preprint arXiv:2302.10307, 2023. 2, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.58, + 0.472, + 0.649 + ], + "angle": 0, + "content": "[53] Lixiang Ru, Yibing Zhan, Baosheng Yu, and Bo Du. Learning affinity from attention: End-to-end weakly-supervised semantic segmentation with transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16846-16855, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.65, + 0.472, + 0.705 + ], + "angle": 0, + "content": "[54] Lixiang Ru, Heliang Zheng, Yibing Zhan, and Bo Du. Token contrast for weakly-supervised semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3093-3102, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.707, + 0.472, + 0.762 + ], + "angle": 0, + "content": "[55] Lixiang Ru, Heliang Zheng, Yibing Zhan, and Bo Du. Token contrast for weakly-supervised semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3093-3102, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.763, + 0.472, + 0.83 + ], + "angle": 0, + "content": "[56] Michael Rubinstein, Armand Joulin, Johannes Kopf, and Ce Liu. Unsupervised joint object discovery and segmentation in internet images. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1939-1946, 2013. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.832, + 0.472, + 0.901 + ], + "angle": 0, + "content": "[57] Ramprasaath R. Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. Grad-cam: Visual explanations from deep networks via gradient-based localization. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2017. 2, 4" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.893, + 0.147 + ], + "angle": 0, + "content": "[58] Gyungin Shin, Weidi Xie, and Samuel Albanie. Reco: Retrieve and co-segment for zero-shot transfer. Advances in Neural Information Processing Systems, 35:33754-33767, 2022. 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.149, + 0.894, + 0.219 + ], + "angle": 0, + "content": "[59] Oriane Simeoni, Gilles Puy, Huy V Vo, Simon Roburin, Spyros Gidaris, Andrei Bursuc, Patrick Pérez, Renaud Marlet, and Jean Ponce. Localizing objects with self-supervised transformers and no labels. arXiv preprint arXiv:2109.14279, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.221, + 0.893, + 0.304 + ], + "angle": 0, + "content": "[60] Amanpreet Singh, Ronghang Hu, Vedanuj Goswami, Guillaume Couairon, Wojciech Galuba, Marcus Rohrbach, and Douwe Kiela. Flava: A foundational language and vision alignment model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 15638-15650, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.305, + 0.893, + 0.375 + ], + "angle": 0, + "content": "[61] Meng Tang, Federico Perazzi, Abdelaziz Djelouah, Ismail Ben Ayed, Christopher Schroers, and Yuri Boykov. On regularized losses for weakly-supervised cnn segmentation. In Proceedings of the European Conference on Computer Vision (ECCV), pages 507-522, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.377, + 0.893, + 0.459 + ], + "angle": 0, + "content": "[62] Anthony Meng Huat Tiong, Junnan Li, Boyang Li, Silvio Savarese, and Steven C.H. Hoi. Plug-and-play vqa: Zero-shot vqa by conjoining large pretrained models with zero training. In Findings of the Conference on Empirical Methods in Natural Language Processing (Findings of EMNLP), 2022. 1, 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.461, + 0.894, + 0.532 + ], + "angle": 0, + "content": "[63] Huy V Vo, Francis Bach, Minsu Cho, Kai Han, Yann LeCun, Patrick Pérez, and Jean Ponce. Unsupervised image matching and object discovery as optimization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8287-8296, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.533, + 0.893, + 0.587 + ], + "angle": 0, + "content": "[64] Van Huy Vo, Elena Sizikova, Cordelia Schmid, Patrick Pérez, and Jean Ponce. Large-scale unsupervised object discovery. Advances in Neural Information Processing Systems, 34:16764-16778, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.589, + 0.893, + 0.645 + ], + "angle": 0, + "content": "[65] Wenguan Wang, Guolei Sun, and Luc Van Gool. Looking beyond single images for weakly supervised semantic segmentation learning. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.647, + 0.893, + 0.715 + ], + "angle": 0, + "content": "[66] Xudong Wang, Rohit Girdhar, Stella X Yu, and Ishan Misra. Cut and learn for unsupervised object detection and instance segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3124-3134, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.717, + 0.893, + 0.799 + ], + "angle": 0, + "content": "[67] Yangtao Wang, Xi Shen, Shell Xu Hu, Yuan Yuan, James L Crowley, and Dominique Vaufreydaz. Self-supervised transformers for unsupervised object discovery using normalized cut. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14543-14553, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.802, + 0.893, + 0.858 + ], + "angle": 0, + "content": "[68] Xiu-Shen Wei, Chen-Lin Zhang, Jianxin Wu, Chunhua Shen, and Zhi-Hua Zhou. Unsupervised object discovery and co-localization by deep descriptor transformation. Pattern Recognition, 88:113-126, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.86, + 0.893, + 0.901 + ], + "angle": 0, + "content": "[69] Yunchao Wei, Jiashi Feng, Xiaodan Liang, Ming-Ming Cheng, Yao Zhao, and Shuicheng Yan. Object region mining with adversarial erasing: A simple classification to semantic" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "4039" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.092, + 0.47, + 0.134 + ], + "angle": 0, + "content": "segmentation approach. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1568-1576, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.137, + 0.47, + 0.18 + ], + "angle": 0, + "content": "[70] Alexandros Xenos, Themos Stafylakis, Ioannis Patras, and Georgios Tzimiropoulos. A simple baseline for knowledge-based visual question answering. In EMNLP, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.182, + 0.471, + 0.252 + ], + "angle": 0, + "content": "[71] Yongqin Xian, Subhabrata Choudhury, Yang He, Bernt Schiele, and Zeynep Akata. Semantic projection network for zero-and few-label semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8256-8265, 2019. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.254, + 0.471, + 0.323 + ], + "angle": 0, + "content": "[72] Haiyang Xu, Qinghao Ye, Ming Yan, Yaya Shi, Jiabo Ye, Yuanhong Xu, Chenliang Li, Bin Bi, Qi Qian, Wei Wang, et al. mplug-2: A modularized multi-modal foundation model across text, image and video. arXiv preprint arXiv:2302.00402, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.326, + 0.471, + 0.409 + ], + "angle": 0, + "content": "[73] Jiarui Xu, Shalini De Mello, Sifei Liu, Wonmin Byeon, Thomas Breuel, Jan Kautz, and Xiaolong Wang. Groupvit: Semantic segmentation emerges from text supervision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18134-18144, 2022. 2, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.412, + 0.471, + 0.495 + ], + "angle": 0, + "content": "[74] Jilan Xu, Junlin Hou, Yuejie Zhang, Rui Feng, Yi Wang, Yu Qiao, and Weidi Xie. Learning open-vocabulary semantic segmentation models from natural language supervision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2935-2944, 2023. 1, 2, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.498, + 0.471, + 0.568 + ], + "angle": 0, + "content": "[75] Lian Xu, Wanli Ouyang, Mohammed Bennamoun, Farid Boussaid, and Dan Xu. Multi-class token transformer for weakly supervised semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4310-4319, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.571, + 0.471, + 0.641 + ], + "angle": 0, + "content": "[76] Mengde Xu, Zheng Zhang, Fangyun Wei, Yutong Lin, Yue Cao, Han Hu, and Xiang Bai. A simple baseline for open-vocabulary semantic segmentation with pre-trained vision-language model. In European Conference on Computer Vision, pages 736-753. Springer, 2022. 1, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.643, + 0.471, + 0.712 + ], + "angle": 0, + "content": "[77] Mengde Xu, Zheng Zhang, Fangyun Wei, Han Hu, and Xiang Bai. Side adapter network for open-vocabulary semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2945-2954, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.715, + 0.471, + 0.784 + ], + "angle": 0, + "content": "[78] Xiao Xu, Chenfei Wu, Shachar Rosenman, Vasudev Lal, Wanxiang Che, and Nan Duan. Bridgetower: Building bridges between encoders in vision-language representation learning. arXiv preprint arXiv:2206.08657, 2022. 1, 2, 3, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.787, + 0.471, + 0.843 + ], + "angle": 0, + "content": "[79] Jiahui Yu, Zirui Wang, Vijay Vasudevan, Legg Yeung, Mojtaba Seyedhosseini, and Yonghui Wu. CoCa: Contrastive captioners are image-text foundation models. arXiv Preprint 2205.01917, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.846, + 0.471, + 0.9 + ], + "angle": 0, + "content": "[80] Nir Zabari and Yedid Hoshen. Open-vocabulary semantic segmentation using test-time distillation. In European Conference on Computer Vision, pages 56-72. Springer, 2022. 1" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.471, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.135 + ], + "angle": 0, + "content": "[81] Yan Zeng, Xinsong Zhang, and Hang Li. Multi-grained vision language pre-training: Aligning texts with visual concepts. arXiv preprint arXiv:2111.08276, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.137, + 0.892, + 0.19 + ], + "angle": 0, + "content": "[82] Yan Zeng, Xinsong Zhang, Hang Li, Jiawei Wang, Jipeng Zhang, and Wangchunshu Zhou. \\(X^2\\)-vlm: All-in-one pretrained model for vision-language tasks. arXiv preprint arXiv:2211.12402, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.193, + 0.892, + 0.261 + ], + "angle": 0, + "content": "[83] Bingfeng Zhang, Jimin Xiao, Yunchao Wei, Mingjie Sun, and Kaizhu Huang. Reliability does matter: An end-to-end weakly supervised semantic segmentation approach. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 12765-12772, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.263, + 0.892, + 0.332 + ], + "angle": 0, + "content": "[84] Hao Zhang, Feng Li, Xueyan Zou, Shilong Liu, Chunyuan Li, Jianwei Yang, and Lei Zhang. A simple framework for open-vocabulary segmentation and detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1020-1031, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.334, + 0.892, + 0.415 + ], + "angle": 0, + "content": "[85] Pengchuan Zhang, Xiujun Li, Xiaowei Hu, Jianwei Yang, Lei Zhang, Lijuan Wang, Yejin Choi, and Jianfeng Gao. Vinvl: Revisiting visual representations in vision-language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5579-5588, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.418, + 0.892, + 0.487 + ], + "angle": 0, + "content": "[86] Xiaolin Zhang, Yunchao Wei, Jiashi Feng, Yi Yang, and Thomas S Huang. Adversarial complementary learning for weakly supervised object localization. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1325-1334, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.489, + 0.892, + 0.53 + ], + "angle": 0, + "content": "[87] Bolei Zhou, Aditya Khosla, Agata Lapedriza, Aude Oliva, and Antonio Torralba. Object detectors emerge in deep scene cnns. arXiv preprint arXiv:1412.6856, 2014. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.531, + 0.892, + 0.599 + ], + "angle": 0, + "content": "[88] Bolei Zhou, Hang Zhao, Xavier Puig, Sanja Fidler, Adela Barriuso, and Antonio Torralba. Scene parsing through ade20k dataset. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 633-641, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.601, + 0.892, + 0.643 + ], + "angle": 0, + "content": "[89] Chong Zhou, Chen Change Loy, and Bo Dai. Extract free dense labels from clip. In European Conference on Computer Vision, pages 696-712. Springer, 2022. 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.644, + 0.892, + 0.713 + ], + "angle": 0, + "content": "[90] Tianfei Zhou, Meijie Zhang, Fang Zhao, and Jianwu Li. Regional semantic contrast and aggregation for weakly supervised semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4299-4309, 2022. 1" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.713 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.957 + ], + "angle": 0, + "content": "4040" + } + ] +] \ No newline at end of file diff --git a/2024/Emergent Open-Vocabulary Semantic Segmentation from Off-the-shelf Vision-Language Models/bf254b56-9ba7-4039-9f52-56365463bb00_origin.pdf b/2024/Emergent Open-Vocabulary Semantic Segmentation from Off-the-shelf Vision-Language Models/bf254b56-9ba7-4039-9f52-56365463bb00_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..585c3474aea6997f18cab88ff06b0d59a3d112e5 --- /dev/null +++ b/2024/Emergent Open-Vocabulary Semantic Segmentation from Off-the-shelf Vision-Language Models/bf254b56-9ba7-4039-9f52-56365463bb00_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:519daff26fd667b4e074050bc488dea9bfb0ca501f54d30b5c81d1675aaf1329 +size 3885401 diff --git a/2024/Emergent Open-Vocabulary Semantic Segmentation from Off-the-shelf Vision-Language Models/full.md b/2024/Emergent Open-Vocabulary Semantic Segmentation from Off-the-shelf Vision-Language Models/full.md new file mode 100644 index 0000000000000000000000000000000000000000..696c9c602164edabc9e6c45d55c05dac87f9a459 --- /dev/null +++ b/2024/Emergent Open-Vocabulary Semantic Segmentation from Off-the-shelf Vision-Language Models/full.md @@ -0,0 +1,400 @@ +# Emergent Open-Vocabulary Semantic Segmentation from Off-the-shelf Vision-Language Models + +Jiayun Luo $^{\flat}$ , Siddhesh Khandelwal $^{\sharp}$ , Leonid Sigal $^{\sharp}$ , and Boyang Li $^{\flat}$ + +$^{b}$ Nanyang Technological University, Singapore + +$^{\ddagger}$ University of British Columbia, Vector Institute for AI, Canada + +{luoj0028,boyang.li}@ntu.edu.sg，{skhandel,lsignal}@cs.ubc.ca + +![](images/1ff679333025df281c2c264dce0007ffa1e9b4392c0c8e79e9805419be2db86c.jpg) +Figure 1. Qualitative Results of PnP-OVSS + BLIP. Images are from Pascal Context and COCO stuff. The right columns and bottom rows show the ground-truth (GT); the rest are our results. Note accurate results even on complex (trees) and small objects (last column). + +![](images/90f2b88f1e18d0f77460141d9566fdb4af766fb5e89e26f13e48fad8f438a73e.jpg) + +![](images/58858b884765120250fddbe6138fde61d893d84686039a0ddc10e510760a8523.jpg) + +![](images/bac028cc5228ca6a1d842d476f8201b7e0a41dcc5aae3f16c20b371a4c03b5d2.jpg) + +![](images/3c969cb703bd95a7da9929f65b58dca7a5b2f2d6b2fcccff5b2c7e717b2ce8bf.jpg) + +![](images/7c6356ae1fb78a8926503660df0845ab899375d8b32c5d739407f4ec2c6c41a2.jpg) + +# Abstract + +From image-text pairs, large-scale vision-language models (VLMs) learn to implicitly associate image regions with words, which prove effective for tasks like visual question answering. However, leveraging the learned association for open-vocabulary semantic segmentation remains a challenge. In this paper, we propose a simple, yet extremely effective, training-free technique, Plug-and-Play Open-Vocabulary Semantic Segmentation (PnP-OVSS) for this task. PnP-OVSS leverages a VLM with direct text-to-image cross-attention and an image-text matching loss. To balance between over-segmentation and under-segmentation, we introduce Salience Dropout; by iteratively dropping patches that the model is most attentive to, we are able to better resolve the entire extent of the segmentation mask. PnP-OVSS does not require any neural network training and performs hyperparameter tuning without the need for any segmentation annotations, even for a validation set. PnP-OVSS demonstrates substantial improvements over comparable baselines (+29.4% mIoU on Pascal VOC, +13.2% mIoU on Pascal Context, +14.0% mIoU on MS COCO, +2.4% mIoU on COCO Stuff) and even outperforms most baselines that conduct additional network training on top of pretrained VLMs. Our codebase is at https://github.com/letitiabana/PnP-OVSS. + +# 1. Introduction + +The classic task of semantic segmentation [20, 22] aims to classify pixels to their object types. Traditional supervised methods require dense pixel-level annotations and are restricted to recognizing a predefined set of objects. To relax these constraints, open-vocabulary semantic segmentation [8, 17, 21, 41, 44, 46, 48, 74, 76, 77, 80, 84] aspires to identify arbitrary object categories, whereas weakly supervised techniques [9, 11, 15, 25, 31, 36, 53, 55, 65, 75, 90] can acquire pixel-level localization capabilities from coarse supervision, e.g., image labels or boxes. + +Large-scale vision-language models (VLMs) pretrained on image-text pairs [1, 37, 38, 40, 60, 78, 85] achieve unprecedented performance on multimodal tasks, such as describing arbitrary images and answering free-form, open-ended questions about them (either with [37, 38, 60] or without finetuning [1, 19, 62, 70]). These tasks apparently involve some ability to localize objects. For example, to answer the question "what objects appear on the table?", the model would have to first localize the table in the image and identify the objects on it. Hence, it is reasonable to conjecture that the VLM network learns to perform open-vocabulary localization from image-text pretraining. However, distilling the localization capability from the VLMs remains an open challenge. + +Most existing methods for open-vocabulary semantic + +![](images/121ad0ade58b246149c5186512c220dafbf9ccfde40adf06e04955f94076089f.jpg) +(a) Original image + +![](images/0cd3e661d89d18175c63de5f8e1f53b307d27537740551d5d8b239eea3fb8172.jpg) +(b) Cross-attention +Figure 2. Segmentation masks for elephant and train using (a) off-the-shelf cross-attention, (b) cross-attention + GradCAM, and (c) cross-attention + GradCam + Salience DropOut (\$3.2). The naive cross-attention masks are too inclusive whereas GradCAM is too exclusive. + +![](images/389d49edd686ef6e886871752288378ae63044a0c67c4a3cf04429552cd44238.jpg) +(c) GradCAM + +![](images/6b6386f24e5d894b02003cba93d289f6a972b1d99c05955596708b16cc7751a1.jpg) +(d) Salience Dropout + +segmentation (OVSS) from VLMs usually obtain single vector encodings for the visual and text inputs respectively [43, 44, 51, 52, 73, 74]. However, pooling every token into a single vector likely discards information about detailed positions of objects and words. We investigate the use of pretrained cross-attention layers for OVSS, which retain finer-grained correspondence between text and image patches. + +Nevertheless, a naive application of cross-attention between the class name and the image patches lead to overly broad segmentation masks that include irrelevant parts of the image (i.e., over-segmentation, see Fig. 2(b)). To alleviate this, [62] employs gradient information from the imagetext matching loss, GradCAM[57]-style, to sharpen the attention-based masks for the purpose of guiding caption generation. However, this results in masks that capture only the most discriminative regions of an object, such as the head of an elephant (i.e., under-segmentation, Fig. 2(c)). To acquire complete object masks, we propose Salience DropOut, which iteratively drops the image patches with high GradCAM attention scores, forcing the model to attend to less discriminative but relevant object parts. + +Another important consideration is the cross-attention layer and the attention head to extract the masks from. These hyperparameters have enormous influence on the final results and are traditionally tuned on a validation set with pixel-level mask annotations. To eliminate the need for dense annotations, we propose an weakly-supervised reward function based on CLIP [49]. On a validation set with images as well as object class names, the technique contrasts the extracted object regions with a blank image. If, according to CLIP, the former is more similar to the corresponding class name than the latter, we increment the reward. All hyperparameter tuning of our technique is performed with a simple random search with this reward, leading to high performance. + +In summary, we propose Plug-and-Play Open-vocabulary Semantic Segmentation (PnP-OVSS), an extremely simple and training-free framework to extract semantic segmentations from VLMs. At zero extra training cost, PnP-OVSS can be used with any pretrained VLM with text-to-image cross attention layer and an image-text matching loss. It has zero reliance on pixel-level annotations, including a validation set for hyperparameter tuning. At the same time, PnP-OVSS delivers excellent performance. It not only beats the training-free baseline with remarkable margins $(+29.4\%)$ mIoU on Pascal VOC, $+13.2\%$ on Pascal Context, $+14.0\%$ on MS COCO, $+2.4\%$ on COCO Stuff), but also outperforms most recent techniques within the past two years that require extensive finetuning on top of the VLM pretraining. + +With this paper, we make three contributions: + +- We propose to combine text-to-image attention, GradCAM, and Salience DropOut to iteratively acquire accurate segmentation of arbitrary classes from pretrained VLM. +- We replace the densely annotated validation set for hyperparameter tuning, which is needed by most existing methods, with a contrastive reward function based on CLIP. This reward function, coupled with random search, finds a good set of hyperparameters for OVSS. +- The proposed method, PnP-OVSS, is simple to use, requires no extra finetuning, and delivers high performance. Its success hints at a new direction for open-vocabulary segmentation tasks leveraging large VLMs. + +# 2. Related Work + +# 2.1. Large-Scale Vision-Language Model + +Large-scale vision-language models (VLMs), trained on millions of image-text pairs, have become the foundation for many multimodal tasks. Architecturally, the straightforward approach to training such methods involves aligning visual and textual latent representations via a simple dot product [24, 49]. However, this is insufficient for complex structured tasks like visual question answering or image captioning, which require specialized approaches that employ separate encoders before cross-attention between modalities [33, 35, 37, 78, 79] or self-attention networks over all tokens from both modalities [10, 27, 39, 40, 60]. Another design dimension when training VLMs is the loss function. Commonly used losses include image-text contrastive learning [34, 35, 37, 50, 60, 81], image-text matching (ITM) [27, 34, 35, 37, 60, 72, 78, 81, 82], prediction of masked tokens or patches [34, 60, 78, 81], and language modeling [1, 7, 34, 37]. + +This work utilizes models with unimodal encoders followed by cross-attention fusion [34, 35, 37, 72, 78, 81, 82], as they work with high-level features, and accurately attend to the appropriate image patches. Additionally, we utilize the gradient from the ITM loss in the GradCAM step (§3.1) + +
MethodPT Data SizeFT data size
Requires finetuning on image-text pairs
OVSegmentor [74]-4.3M
Vil-Seg [43]400M412M
GroupVit* [73]-3.4M
GroupVit [73]-26M
CLIPpy [51]-134M
SegClip [44]400M3.4M
ViewCo [52]-26M
TCL [6]400M15M
PACL [46]400M30M
Requires finetuning but not image-text pairs
MaskClip w/ ST [89]400M1.2M
ZeroSeg* [8]400M3.4M
ZeroSeg [8]400M1.2M
Requires no finetuning
MaskClip[89]400M0
Reco[58]400M0
PnP-OVSS (Ours)
+ BLIP129M0
+ BridgeTower400M+ 4M0
+ +Table 1. Training data of current Zero-shot semantic segmentation methods with only text supervision. PT stands for pretraining with image-caption data, FT stands for fine-tuning with image-caption data. ST stands for self training. We list only the image-caption data used for pretraining and all type of data for finetuning. For hyperparameter tuning, our method uses CLIP-L/14 pretrained with 400M data to calculate the reward. + +to sharpen segmentation masks. + +# 2.2. Zero-shot Semantic Segmentation + +Zero-shot semantic segmentation predicts a dense segmentation mask for any object class described by a given text prompt, with only prior exposure to class-agnostic image-level supervision. This contrasts with weakly supervised semantic segmentation [3, 13, 23, 26, 28, 30, 54, 61, 69, 69, 83, 86, 87], which relies on class-specific annotations, and unsupervised object discovery [12, 14, 56, 59, 63, 64, 67, 68], which identifies the sole object in the foreground. + +Traditional methods [4, 18, 47, 71] train a classifier to distinguish between seen and unseen visual features, wherein the unseen visual features are obtained from a generative model, trained on pairs of seen class image-text embeddings. Recently, methods additionally leverage knowledge from VLMs to attain better matching of visual and textual features [17, 32, 41, 76, 89]. Concretely, they train the segmentation network with dense annotations, while replacing a part of the framework with components from VLMs. + +Recently, methods have explored the use of additional supervision to further reduce the need for pixel level annotations. Prior work is compared in Table 1. + +Models Finetuned on Image-Text Pairs. Methods proposed in [6, 43, 44, 46, 51, 52, 73, 74] require paired image-text annotations to adapt to the zero-shot segmentation task. Specifically, approaches in [43, 44, 51, 52, 73, 74] cluster semantically similar pixels by contrasting grouped region embeddings with textual embeddings, usually via the use of contrastive and self-supervised losses. PACL [46] modifies the constrastive loss to operate over aggregated patch embeddings (instead of image embeddings) to encourage better alignment between image patches and text. TCL [6] achieves a similar patch-text alignment by introducing an additional module to extract text-grounded image regions. + +Models Finetuned on Pseudo-labels. Methods in [8, 89] do not require additional image-text supervision, but involve additional fine-tuning to adapt to the segmentation task. ZeroSeg [8] attempts to match its own group embeddings with multi-scaled segment embeddings obtained from CLIP [49]. MaskClip w/ ST [89] modifies the global attention pooling layer within CLIP to output segmentation masks, which are used as pseudo-labels to train a segmentation network. + +In comparison, our proposed PnP-OVSS does not require any fine-tuning or additional paired image-text annotations. It can directly distill high quality open vocabulary semantic segmentations from any VLM with direct text-to-image cross-attention and an image-text matching loss [34, 35, 37, 72, 78, 81, 82]. Compared to approaches in [58, 89] that perform zero-shot open vocabulary semantic segmentation under a similar no training and no additional annotations paradigm, PnP-OVSS achieves considerably superior performance. + +# 3. Method + +PnP-OVSS has four major steps. First, we extract a cross-attention salience map per object class from a VLM. Second, we sharpen the salience map by weighing it with the ITM gradient in the style of GradCAM. Third, we apply Salience DropOut that iteratively completes the salience maps. Fourth, we apply Dense CRF [29] for fine-grained adjustment. The process is illustrated in Fig. 3 and Fig. 4. + +In the first step, we feed an image and a text prompt to the pretrained VLM and extract cross-attention maps. The text prompt is "A picture of [class 1] [class 2] ... [class K]", which includes all $K$ class names of interest in the dataset. The image is divided into $P \times P$ patches. The text prompt and the image first go through the modality-specific encoders respectively, followed by a cross-attention fusion module. In the cross-attention layers, the text encodings serve as query vectors and the image patche encodings are as key and value vectors. + +We extract an attention map for each text token to the image patches from a particular cross-attention layer and an + +![](images/c09ee7dc9b163aecd63adce78dc64440bff4ac145ac5613a803270cd95bde287.jpg) +Figure 3. The first iteration with cross-attention + GradCAM + Salience DropOut. The text prompt contains $K$ class names and the image contains $P \times P$ patches. From a cross-attention layer and an attention head in the pretrained VLM, we obtain $K$ attention score maps of size $P \times P$ , which are sharpened by GradCAM using gradients from the image-text-matching (ITM) loss. To get more complete predictions, we perform Salience Dropout, which repeatedly zero out image patches of the highest average scores and feeds the remaining patches to the image encoder again, forcing the model to attend to other less discriminative patches. We show example salience maps from all iterations in Fig. 4. + +attention head. Note that different layers and heads lead to drastic performance differences, and the choices are hyperparameters, tuned using the procedure in §3.4. We exclude attention maps for the first three tokens "A", "picture", "of", which do not describe semantic classes. If a class name contains two or more tokens, we take the mean attention map. This procedure yields an attention tensor of size $K \times P \times P$ . + +With the correct layer and attention head, we observe that this attention map, when normalized by softmax along the first dimension $K$ , can provide passable semantic segmentations. However, it tends to include many patches unrelated to the class being segmented, leading to over-segmentation. Hence, we introduce two refinement steps, GradCAM and Salience DropOut, explained in §3.1 and §3.2. After these steps, we acquire aggregate salience maps for every class. Finally, we conduct local polish to the resultant salience maps using Dense CRF, as described in §3.3. + +# 3.1. Map Sharpening with GradCAM + +The off-the-shelf attention maps tend to also cover many patches unrelated to the class name (See Fig. 2 (b)). Prior works [35, 62] further sharpen the attention maps and focus them on class-discriminative regions using a variant of GradCAM [57], which is originally proposed for convolutional networks, but applicable to attention maps. It is worth + +noting that [35, 62] use the technique for purposes other than semantic segmentation. + +The GradCAM method requires a gradient. Here we leverage the image-text matching (ITM) loss, which trains the VLM to classify if an image-text pair match each other or not. Computing the ITM loss requires a label. We use "matching" (as opposed to "not matching") as the label and compute the gradient of the loss with respect to the attention score. This is equivalent to asking: which attention scores contribute the most to the decision that the image-text pair is matching? Formally, we denote a $P \times P$ attention map for class $k$ as $M^{(k)}$ and the ITM loss as $\mathcal{L}_{\mathrm{ITM}}$ . The GradCAM class salience map is computed as + +$$ +\widetilde {M} ^ {(k)} = \max \left(0, \frac {\partial \mathcal {L} _ {\mathrm {I T M}}}{\partial M ^ {(k)}}\right) \otimes M ^ {(k)}, \tag {1} +$$ + +where $\otimes$ denote the component-wise multiplication and $\max (\cdot)$ is also applied component-wise. + +# 3.2. Salience Dropout + +As illustrated in Fig. 2 (c), segmentations generated by the GradCAM-style re-weighting of cross-attention are often narrowly focused on the most discriminative regions for a given class. However, the less discriminative regions are + +![](images/d3a733835318b2e7e0db8bcce74e07c5564f79f60a3435f7f61dc0c279b0cea1.jpg) +GradCAM Class Salience Map + +![](images/32656e85b9ec5cf0849b5761447164fe7ad2eebf1d1ac65da6597c7dd2442342.jpg) +$1^{\mathrm{st}}$ Salience Drop + +![](images/744a6862317d00d58f745171488df677d32103b6c246ac1b7cc3214e7d9ebd8f.jpg) + +![](images/73bb98368c603e8ff1de7b26f553ba5d08e4a8b3b06f34ef34f3eedfc247b85f.jpg) + +![](images/238cdefc313455f4837881828fad3a25e3f4d543b91cad0b118a075cad197d40.jpg) +$2^{\text{nd}}$ Salience Drop + +![](images/74884e061dfabf8267a7181b21c527458224149a23deacfee289b1e7596c6ea4.jpg) + +![](images/7322f1594217095b3fdd8220e094f3ac5ff1fe9df76a83c16e1654366d684284.jpg) + +![](images/edfb58038c4d94ee7569afbb4dadbdf199a4daf88bfeefc766b8b51160d8e613.jpg) + +![](images/d3f610134af2c85a76beac4c6dc4a4159c4171cb2ba8d7495325d7cdad7ed4c9.jpg) +$3^{\text{rd}}$ Salience Drop + +![](images/7d7cef942ce057cb60a00774c0a66d4a2014fae01734de2c4a9236efa962ab35.jpg) + +![](images/599dd1680e78322b71cf4320a1fbabef424abcf2dcdc92cf320d90fb9eae7357.jpg) + +![](images/1ab3915806d0c865ae4022e60b1548c2f6b798a20cbcbc5e2618c2ac9be3b6cb.jpg) + +![](images/574affc845379e36ae8702dd04be0f2832554dd1dbd960ddc799e5ed2e4962a8.jpg) +$4^{\text{th}}$ Salience Drop + +![](images/8dd8cae2b3a507aba4606c3f75e8484f63a74de2ebb0238f04b22a132dfb59cf.jpg) + +![](images/a0e00ed28d66eb4d9aff998fcf6aa08e57dac8ccd3f011a844301dc7d6463c52.jpg) + +![](images/2d5cc0608631b234445baed22a92557952a986e72343a55c1560eed6638d0a7e.jpg) + +![](images/4c8c4c001c6c33acf2f5c3cc255a13ccfc0fc59c4a7df3b9456ddbfbcb3e2a1e.jpg) + +![](images/ada2590fe21afaeef9cf5a094d9e93eb679042206ded94d0f56620920dfb9831.jpg) +Thresholding + +![](images/8b5d38c510f3efa6df5c28d36d86aa7a70ba01e01594fcb7527ea928a0eddfdf.jpg) +Gaussian Blur + +![](images/74dccded38d1fd993f7f9dc49dafac7db0251247dffd40af1cb61aec99dbc44e.jpg) + +![](images/c25ee5c291fab16c727494f080edec705ac1602c98abb8c60ef1bc61f2fdfc4b.jpg) +Figure 4. An illustration of Salience DropOut, showing GradCAM salience values after each iteration. Black squares in the images indicate dropped patches. We obtain the final result by summing the salience maps from all iterations and applying thresholding, Gaussian blur, and Dense CRF. + +![](images/d15c7bcad4504082e1f2f8e7978b68e10b08dfb951ba47e8fe4d4c9723144036.jpg) +Dense CRF + +![](images/71cda47ac3f80029bb03cc0bb58cac64d88f81c1227bad5c911df2592ffd00c3.jpg) + +![](images/e53037d5bc68fc58c941803f94421208306468a2776e930d7028434eb79bc6ed.jpg) + +still important for the completeness of masks. To compel the VLM to attend to these regions, we propose an iterative technique called Salience DropOut. + +Since our task is zero-shot and open-vocabulary, we do not have prior knowledge of the classes present in the image. Hence, we sum up the salience maps $\widetilde{M}^{(k)}$ over all classes $k$ , yielding a class-agnostic salience map $U^{(t)}$ for the $t^{\text{th}}$ Salience DropOut iteration, $U^{(t)} = \sum_{k=1}^{K} \widetilde{M}^{(k,t)}$ , where $\widetilde{M}^{(k,t)}$ is the GradCAM salience map for class $k$ after the $t^{\text{th}}$ iteration. Next, we zero out the $50\%$ of image patches with the highest values in $U^{(t)}$ . On the remaining + +image, we compute the GradCAM salience maps $\widetilde{M}^{(k,t + 1)}$ and their sum, $U^{(t + 1)}$ . Any image patch previously zeroed out will always receive zero salience in later iterations. + +Formally, the set of remaining image patches $S^{(t)} \subseteq \{1, \dots, P\}^2$ after the $t^{\text{th}}$ dropout iteration is defined as + +$$ +S ^ {(t)} = S ^ {(t - 1)} \backslash \{(i, j) \mid U _ {i j} ^ {(t)} \geq \eta \}, \tag {2} +$$ + +$$ +\eta = \operatorname {m e d i a n} \left(\left\{U _ {i j} ^ {(t)} \mid (i, j) \in S ^ {(t - 1)} \right\}\right), \tag {3} +$$ + +where $U_{i,j}^{(t)}$ denotes the aggregate salience value of the image patch at row $i$ and column $j$ . Additionally, note that the input to the first dropout iteration $S^0 = \{1,\dots ,P\} ^2$ is the set of all image patches. + +We stop at four rounds of dropout, as almost all $(93.75\%)$ patches are removed beyond that point. The final output for each class $k$ is the sum over all salience maps across the four dropout iterations, $\hat{M}^{(k)} = \sum_{t=1}^{4} \widetilde{M}^{(k,t)}$ . + +# 3.3. Gaussian Blur and Dense CRF + +The salience dropout procedure generates $k$ continuous-valued salience maps, one per object class. To filter out small random noise in the salience values, we subsequent apply a straightforward thresholding operation at a predefined value $T$ on the salience values and obtain binary segmentation masks. Nonetheless, the hard thresholding creates jagged segmentations with sharp edges that often do not coincide with object boundaries. One common strategy in zero-shot segmentation is to apply Dense Conditional Random Field (CRF) [29], which makes fine-grained adjustments to the estimated masks by enforcing consistency between nearby image pixels with similar colors. + +However, we find that the hard 0/1 labels in the binary masks do not work well as pixel unary potentials for Dense CRF. Hence, we smooth them using a Gaussian kernel with a preset variance $\sigma$ , which results in a better initialization for unary terms and accounts for uncertainty of exact segmentation boundary along the patch boundaries. + +# 3.4. Hyperparameter Tuning + +Three hyperparameters in PnP-OVSS have the strongest influence on the result, the cross-attention layer $L$ , the attention head $H$ , and the binary threshold $T$ . Traditionally, tuning these hyperparameters requires a validation set with pixel-level labels. However, since our goal is to perform zero-shot open-vocabulary semantic segmentation, this requirement could potentially limit the applicability of the technique. Instead, we propose a weakly supervised reward function for hyperparameter tuning, which only requires a set of images and the class names appearing in each image. + +The reward for an image $I$ is calculated as follows. We start with a set of classes present in the image, denoted as $\mathcal{K}(I)$ . For each class $k\in \mathcal{K}(I)$ , we obtain a segmentation + +mask $M^{(k)}$ (which can be the GradCAM mask, the Salience DropOut mask, or the Dense CRF mask). Next, we apply the mask to the image $I$ and input the extracted regions $M^{(k)} \otimes I$ into a pretrained neural network $f$ , which takes an image and a textual class name as input and produces a similarity score. We calculate the normalized probability that the masked image $M^{(k)} \otimes I$ belongs to the ground-truth class $k$ and contrast with a completely black image $\mathbf{0}$ . + +$$ +\operatorname {R e w a r d} = \sum_ {k \in \mathcal {K} (I)} \mathbb {1} \left[ P r \left(M ^ {(k)} \otimes I, k\right) > P r (\mathbf {0}, k) \right], \tag {4} +$$ + +$$ +P r (I, k) = \frac {\exp (f (I , k))}{\sum_ {k ^ {\prime} \in \mathcal {K} (I)} \exp (f (I , k ^ {\prime}))}, \tag {5} +$$ + +where $\mathbb{1}(\cdot)$ is the indicator function. Intuitively, a reward of 1 is assigned if and only if the image features pulled with the estimated mask for class $k$ bears higher similarity to the class name of $k$ than a black image (which can be interpreted as the prior probability of class $k$ ). + +We sum up the reward for all validation images as the total reward. The best hyperparameters, including the cross-attention layer, the attention head, the threshold $T$ , and the variance in the Gaussian blur kernel, are determined using a simple random search. + +# 4. Experiments + +# 4.1. Datasets and Implementation Details + +Following the previous work for zero-shot semantic segmentation, we adopt validation sets of Pascal-VOC 2012 [16], Pascal Context [45], COCO Object [42], COCO Stuff [5], and ADE20K [88] that contain, respectively, 20 object classes, 59 object and stuff classes, 80 object classes, 171 object and stuff classes, and 150 object and stuff classes to evaluate our framework. To verify its versatility, we apply PnP-OVSSto two high-performance VLMs, BLIP [37] and BridgeTower [78], which have the ITM loss and text-to-image cross-attention. More details are in the supplementary material. + +Hyperparameter tuning. We use CLIP VIT-L/14 to calculate the reward. The input resolution is $336 \times 336$ . The search spaces of the random search and results are shown in Tab. 2. For computation efficiency, we tune the layer, the head, and the threshold with GradCAM masks. With the first three hyperparameters fixed, we tune the Gaussian variance on masks before Dense CRF. We directly adopt Dense CRF hyperparameters from CutLER [66] without tuning. + +Baselines. We adopt recent papers on zero-shot open-vocabulary semantic segmentation as baselines. The only baselines strictly comparable to our work are MaskClip [89] and Reco [58], which do not perform any finetuning on pretrained VLMs and do not perform hyperparameter tuning on dense annotations. We call these Group + +
HyperparametersStartEndStepSolution
BLIP
Layer11218
Head112110
Attention Threshold0.050.50.10.15
BridgeTower
Layer1612
Head11618
Attention Threshold0.050.50.10.15
Gaussian Blur
Standard Deviation0.010.110.020.05
+ +Table 2. Search space for hyperparameters + +3. To further expand our scope, we also include two other groups of baselines. Group 2 finetunes VLMs but does not require image-text pairs, including MaskClip with self-training (ST) [89] and ZeroSeg [8]. We include ZeroSeg, trained with ImageNet1K, and ZeroSeg*, trained with CC3M+COCO. Group 1 contains baselines that require training on image-text pairs. For completeness, we also compare against supervised techniques since 2019. To maintain the zero-shot setting, we test them on classes not observed during training. For more details, we refer readers to the supplementary material. + +# 4.2. Main Results + +We show the main results in Tab. 3. As input resolution may influence results and cause unfair comparisons, we label the resolution used by each method. PACL [46] uses an 224 resolution but changes the stride for image patchification from 16 to 4, hence introducing overlapping patches. + +PnP-OVSS exhibits excellent performance. Comparing with MaskClip [89] and Reco [58], the two methods that require no additional training and ground truth for hyperparameter tuning, on an equal-resolution basis, we attain $+29.4\%$ mIoU on Pascal VOC, $+13.2\%$ mIoU on Pascal Context, $+14.0\%$ mIoU on COCO Object, and $+11.4\%$ mIoU on ADE-20K. Further, PnP-OVSS surpasses all baselines in Group 2. On an equal-resolution basis, we achieve $+13.7\%$ mIoU on Pascal Voc, $+9.6\%$ on Pascal Context, $+9.5\%$ on COCO Object, $+11.6\%$ on COCO Stuff, and $+12.8\%$ on ADE-20K. + +When compared to Group 1, PnP-OVSS still outperforms most. On the Pascal datasets, under equal resolutions, PnP-OVSS + BLIPFlickr outperforms 6 out of 10 baselines on Pascal VOC, and 8 out of 9 baselines on Pascal Context. For the COCO datasets, under equal resolutions, PnP-OVSS + BLIPFlickr beats 5 out of 8 baselines on COCO Object, and all baselines except PACL [46] on COCO Stuff. The two Pascal datasets share many images, and the two COCO datasets use exactly the same images, but they have + +
MethodFinetuning VLMsHT on Dense LabelsShort-side ResolutionPascal VOC-20Pascal Context-59COCO Object-80COCO Stuff-171ADE 20K-150
Group 1: Methods that require weakly supervised finetuning on image-text data
ViL-Seg†[43]✓✓-37.318.9-18.0-
CLIPpy [51]✓✓22452.2-32.025.5*13.5
SegClip [44]✓✓22452.624.726.5--
GroupVit (by [51])✓✓22428.114.812.9-6.2
GroupVit (by [6])✓✓44850.418.727.515.39.2
GroupVit [73]✓✓44852.322.424.3--
ViewCo [52]✓✓44852.423.023.5--
OVSegmentor [74]✓✓44853.820.425.1--
TCL [6] +PAMR [2]✓✓44855.030.431.622.417.1
PACL [46]✓✓224×472.350.1-38.831.4
Group 2: Methods that require finetuning but not real image-text data
MaskClip w/ ST [89]✓✓336-31.1-18.0-
MaskClip w/ ST (by [6])✓✓44838.823.620.616.49.8
ZeroSeg [8]✓✓44840.820.420.2--
Group 3: Methods that require no finetuning
MaskClip (by [51])××22422.1-13.88.16.8
MaskClip [89]××336-25.5-14.6-
Reco [58]××320-27.2-27.2-
Reco (by [6])××44825.119.915.714.811.2
PnP-OVSS with different VLMs
BLIPFlickr××22447.836.424.825.818.2
BLIPFlickr××32051.740.428.029.621.3
BLIPFlickr××33652.540.728.229.621.9
BLIPFlickr××44854.542.229.731.522.6
BLIPFlickr××76854.142.831.832.523.5
BLIPCOCO××76855.741.933.832.623.2
BridgeTower××32236.432.324.227.618.6
BridgeTower××33635.332.424.227.618.0
BridgeTower××77035.232.424.528.019.0
+ +Table 3. Zero-shot semantic segmentation performance in mIoU. Group 3 contains the most similar baselines that serve as fair comparisons to PnP-OVSS. Groups 1 and 2 benefit from additional training, extra image-text data, and hyperparameter tuning on dense labels. We use the word "by" followed by a paper citation to indicate results of the same technique reported by different papers. $\star$ CLIPpy tests on 133 categories of COCO Stuff while we test all 171 classes of COCO Stuff. ViL-Seg† is tested on subset of classes on the three datasets, as detailed in the supplementary. + +
MethodDense +LabelsHT on Dense +LabelsPascal +VOC-20Pascal +Context-59COCO +Stuff-171
PnP-OVSS + BLIPFlickr (Ours)××53.653.839.8
SPNet+ST [71]✓✓25.8-26.9
ZS3Net+ST [4]✓✓21.220.710.6
CaGNet+ST [18]✓✓30.3-13.4
STRICT [47]✓✓35.6-30.3
LSeg [32]✓✓41.0--
SimBase [76]✓✓72.5-36.3
MaskCLIP+ w/ ST [89]✓✓86.166.754.7
+ +Table 4. Comparison of zero-shot semantic segmentation performance on unseen categories with methods trained with dense annotation. + +
Ablated ModelPascal ContextCOCO Stuff
BLIPFlickr19.814.5
+ GradCam21.617.5
+ Drop 125.119.8
+ Drop 226.520.6
+ Drop 327.020.9
+ Drop 427.220.9
+ Drop 4 + Blur36.828.6
+ Drop 4 + Dense CRF35.331.8
+ Drop 4 + Blur + Dense CRF42.832.5
+ +vastly different lists of classes to segment. We observe that the advantage of PnP-OVSS over methods in Group 1 becomes more pronounced as the number of classes on the same image increases. + +When applied to BridgeTower [78], PnP-OVSS still surpasses all methods in Group 3 by $10.2\%$ on Pascal VOC, $5.1\%$ on Pascal Context, $8.5\%$ on COCO Object, $0.4\%$ on COCO Stuff, and $6.8\%$ on ADE-20K. This showcases the plug-and-play ability of PnP-OVSS, which excels with different base networks. + +We report comparisons against supervised methods in Tab 4. PnP-OVSS + BLIPFlickr outperforms 5 out of 7 methods on Pascal VOC, as well as every baseline except MaskCLIP+ on Pascal Context and COCO stuff. As these baselines benefit from dense supervision, the results further demonstrate the strengths of PnP-OVSS. + +# 4.3. Ablation Study + +We perform gradual ablation of the components of $\mathrm{PnP-OVSS}$ on $\mathrm{BLIP}_{\mathrm{Flickr}}$ and report the results in Tab. 5. Each component, including GradCAM, all Salience DropOut iterations, Gaussian blur, and Dense CRF, contribute positively to the final performance. In particular, the first iteration of Salience DropOut has much larger impact (+3.5/2.3) than the second iteration (+1.4/0.8), which in turn is more important than the rest. Interestingly, Gaussian blur by itself attains good performance (+9.6/7.7) whereas Dense CRF only works well when combined with blur. Dense CRF alone is worse than Gaussian blur by 1.5 mIoU on Pascal Context. This is likely caused by the fact that hard 0/1 labels resulted from thresholding are not informative unary potentials that can be leveraged by CRF effectively. + +# 4.4. Hyperparameter Sensitivity + +The choice of hyperparameters often exerts substantial influence on segmentation performance. Here we quantitatively examine how the choice of cross-attention layers + +Table 5. An ablation study of PnP-OVSS + BLIPFlickr with resolution 768 on Pascal Context and COCO Stuff. + +
Layer Mean123456
mIoU11.612.311.311.011.612.9
Layer Mean789101112
mIoU13.725.325.823.412.55.8
Head in Layer 8123456
mIoU16.63.88.620.019.85.2
Head in Layer 8789101112
mIoU11.16.03.829.617.38.0
+ +Table 6. Semantic segmentation performance using cross-attention maps averaged across all heads in a layer and separate heads in Layer 8. Results are attained with PnP-OVSS+BLIPFlickr on COCO Stuff and resolution 336. + +and attention heads may change the segmentation mIoU on COCO Stuff. Tab. 6 shows the results obtained from the average cross-attention maps over all heads in each layer and those from different attention heads. + +We make the following observations. First, different layers and heads have drastic performance differences. The best-to-worst difference among all layers is 20, and that among heads in Layer 8 is 23.8, underscoring the importance of hyperparameter tuning. Second, the random search using the proposed reward function correctly identifies the best layer-head combination, even though Layer 8 is not the best layer based on average head performance. This indicates the effectiveness of our method. + +# 5. Conclusions + +We propose PnP-OVSS, which extracts the ability of semantic segmentation from opaque VLMs. PnP-OVSS is simple to use, requires no extra finetuning, and delivers high performance, exceeding not only all baselines that requires no finetuning, but also all baselines that do not use image-text pairs in finetuning. Its success hints at a new direction for open-vocabulary segmentation tasks leveraging large VLMs. + +# 6. Acknowledgments + +This work has been supported by the Nanyang Associate Professorship and the National Research Foundation Fellowship (NRF- NRFF13-2021-0006), Singapore, and by the Vector Institute for AI, Canada CIFAR AI Chair, and NSERC Discovery Grant. Any opinions, findings, conclusions, or recommendations expressed in this material are those of the authors and do not reflect the views of the funding agencies. + +# References + +[1] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, Roman Ring, Eliza Rutherford, Serkan Cabi, Tengda Han, Zhitao Gong, Sina Samangooei, Marianne Monteiro, Jacob L Menick, Sebastian Borgeaud, Andy Brock, Aida Nematzadeh, Sahand Sharifzadeh, Mikol aj Binkowski, Ricardo Barreira, Oriol Vinyals, Andrew Zisserman, and Karén Simonyan. Flamingo: a visual language model for few-shot learning. In Advances in Neural Information Processing Systems, 2022. 1, 2 +[2] Nikita Araslanov and Stefan Roth. Single-stage semantic segmentation from image labels. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4253-4262, 2020. 7 +[3] Loris Bazzani, Alessandra Bergamo, Dragomir Anguelov, and Lorenzo Torresani. Self-taught object localization with deep networks. In 2016 IEEE winter conference on applications of computer vision (WACV), pages 1-9. IEEE, 2016. 3 +[4] Maxime Bucher, Tuan-Hung Vu, Matthieu Cord, and Patrick Pérez. Zero-shot semantic segmentation. Advances in Neural Information Processing Systems, 32, 2019. 3, 7 +[5] Holger Caesar, Jasper Uijlings, and Vittorio Ferrari. Cocosstuff: Thing and stuff classes in context. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1209-1218, 2018. 6 +[6] Junbum Cha, Jonghwan Mun, and Byungseok Roh. Learning to generate text-grounded mask for open-world semantic segmentation from only image-text pairs. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11165-11174, 2023. 3, 7 +[7] Jun Chen, Han Guo, Kai Yi, Boyang Li, and Mohamed Elhoseiny. Visualgpt: Data-efficient adaptation of pretrained language models for image captioning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18030-18040, 2022. 2 +[8] Jun Chen, Deyao Zhu, Guocheng Qian, Bernard Ghanem, Zhicheng Yan, Chenchen Zhu, Fanyi Xiao, Sean Chang Culatana, and Mohamed Elhoseiny. Exploring open-vocabulary semantic segmentation from clip vision encoder distillation only. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 699-710, 2023. 1, 3, 6, 7 +[9] Qi Chen, Lingxiao Yang, Jian-Huang Lai, and Xiaohua Xie. Self-supervised image-specific prototype exploration for weakly supervised semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4288-4298, 2022. 1 +[10] Yen-Chun Chen, Linjie Li, Licheng Yu, Ahmed El Kholy, Faisal Ahmed, Zhe Gan, Yu Cheng, and Jingjing Liu. Uniter: Universal image-text representation learning. In Proceedings of the 16th European Conference on Computer Vision, 2020. 2 +[11] Zhaozheng Chen, Tan Wang, Xiongwei Wu, Xian-Sheng Hua, Hanwang Zhang, and Qianru Sun. Class re-activation + +maps for weakly-supervised semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 969-978, 2022. 1 +[12] Minsu Cho, Suha Kwak, Cordelia Schmid, and Jean Ponce. Unsupervised object discovery and localization in the wild: Part-based matching with bottom-up region proposals. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1201-1210, 2015. 3 +[13] Ramazan Gokberk Cinbis, Jakob Verbeek, and Cordelia Schmid. Weakly supervised object localization with multifold multiple instance learning. IEEE transactions on pattern analysis and machine intelligence, 39(1):189-203, 2016. 3 +[14] Timothée Darcet, Maxime Oquab, Julien Mairal, and Piotr Bojanowski. Vision transformers need registers. arXiv preprint arXiv:2309.16588, 2023. 3 +[15] Ye Du, Zehua Fu, Qingjie Liu, and Yunhong Wang. Weakly supervised semantic segmentation by pixel-to-prototype contrast. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4320–4329, 2022. 1 +[16] Mark Everingham, Luc Van Gool, Christopher KI Williams, John Winn, and Andrew Zisserman. The pascal visual object classes (voc) challenge. International journal of computer vision, 88:303-338, 2010. 6 +[17] Golnaz Ghiasi, Xiuye Gu, Yin Cui, and Tsung-Yi Lin. Scaling open-vocabulary image segmentation with image-level labels. In European Conference on Computer Vision, pages 540-557. Springer, 2022. 1, 3 +[18] Zhangxuan Gu, Siyuan Zhou, Li Niu, Zihan Zhao, and Liqing Zhang. Context-aware feature generation for zero-shot semantic segmentation. In Proceedings of the 28th ACM International Conference on Multimedia, pages 1921-1929, 2020. 3, 7 +[19] Jiaxian Guo, Junnan Li, Dongxu Li, Anthony Meng Huat Tiong, Boyang Li, Dacheng Tao, and Steven Hoi. From images to textual prompts: Zero-shot visual question answering with frozen large language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10867-10877, 2023. 1 +[20] Yanming Guo, Yu Liu, Theodoros Georgiou, and Michael S Lew. A review of semantic segmentation using deep neural networks. International journal of multimedia information retrieval, 7:87-93, 2018. 1 +[21] Kunyang Han, Yong Liu, Jun Hao Liew, Henghui Ding, Jiajun Liu, Yitong Wang, Yansong Tang, Yujiu Yang, Jiashi Feng, Yao Zhao, et al. Global knowledge calibration for fast open-vocabulary segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 797-807, 2023. 1 +[22] Shijie Hao, Yuan Zhou, and Yanrong Guo. A brief survey on semantic segmentation with deep learning. Neurocomputing, 406:302-321, 2020. 1 +[23] Qibin Hou, PengTao Jiang, Yunchao Wei, and Ming-Ming Cheng. Self-erasing network for integral object attention. Advances in Neural Information Processing Systems, 31, 2018. 3 +[24] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom + +Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In Proceedings of the 38th International Conference on Machine Learning, 2021. 2 +[25] Peng-Tao Jiang, Yuqi Yang, Qibin Hou, and Yunchao Wei. L2g: A simple local-to-global knowledge transfer framework for weakly supervised semantic segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16886-16896, 2022. 1 +[26] Tsung-Wei Ke, Jyh-Jing Hwang, and Stella X Yu. Universal weakly supervised segmentation by pixel-to-segment contrastive learning. arXiv preprint arXiv:2105.00957, 2021. 3 +[27] Wonjae Kim, Bokyung Son, and Ildoo Kim. Vilt: Vision-and-language transformer without convolution or region supervision. In Proceedings of the 38th International Conference on Machine Learning, 2021. 2 +[28] Alexander Kolesnikov and Christoph H Lampert. Seed, expand and constrain: Three principles for weakly-supervised image segmentation. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part IV 14, pages 695-711. Springer, 2016. 3 +[29] Philipp Krahenbuhl and Vladlen Koltun. Efficient inference in fully connected crfs with gaussian edge potentials. Advances in neural information processing systems, 24, 2011. 3, 5 +[30] Krishna Kumar Singh and Yong Jae Lee. Hide-and-seek: Forcing a network to be meticulous for weakly-supervised object and action localization. In Proceedings of the IEEE International Conference on Computer Vision, pages 3524-3533, 2017. 3 +[31] Jungbeom Lee, Seong Joon Oh, Sangdoo Yun, Junsuk Choe, Eunji Kim, and Sungroh Yoon. Weakly supervised semantic segmentation using out-of-distribution data. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16897-16906, 2022. 1 +[32] Boyi Li, Kilian Q. Weinberger, Serge Belongie, Vladlen Koltun, and René Ranftl. Language-driven semantic segmentation, 2022. 3, 7 +[33] Chenliang Li, Haiyang Xu, Junfeng Tian, Wei Wang, Ming Yan, Bin Bi, Jiabo Ye, He Chen, Guohai Xu, Zheng Cao, Ji Zhang, Songfang Huang, Fei Huang, Jingren Zhou, and Luo Si. mPLUG: Effective and efficient vision-language learning by cross-modal skip-connections. In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, 2022. 2 +[34] Chenliang Li, Haiyang Xu, Junfeng Tian, Wei Wang, Ming Yan, Bin Bi, Jiabo Ye, Hehong Chen, Guohai Xu, Zheng Cao, et al. mplug: Effective and efficient vision-language learning by cross-modal skip-connections. arXiv preprint arXiv:2205.12005, 2022. 2, 3 +[35] Junnan Li, Ramprasaath R. Selvaraju, Akhilesh Deepak Gotmare, Shafiq Joty, Caiming Xiong, and Steven Hoi. Align before fuse: Vision and language representation learning with momentum distillation. In NeurIPS, 2021. 2, 3, 4 +[36] Jinlong Li, Zequn Jie, Xu Wang, Xiaolin Wei, and Lin Ma. Expansion and shrinkage of localization for weakly- + +supervised semantic segmentation. Advances in Neural Information Processing Systems, 35:16037-16051, 2022. 1 +[37] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In International Conference on Machine Learning, pages 12888-12900. PMLR, 2022. 1, 2, 3, 6 +[38] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In ICML, 2023. 1 +[39] Wei Li, Can Gao, Guocheng Niu, Xinyan Xiao, Hao Liu, Jiachen Liu, Hua Wu, and Haifeng Wang. UNIMO: Towards unified-modal understanding and generation via cross-modal contrastive learning. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), 2021. 2 +[40] Xiujun Li, Xi Yin, Chunyuan Li, Pengchuan Zhang, Xiaowei Hu, Lei Zhang, Lijuan Wang, Houdong Hu, Li Dong, Furu Wei, Yejin Choi, and Jianfeng Gao. Oscar: Object-semantics aligned pre-training for vision-language tasks. In ECCV 2020, 2020. 1, 2 +[41] Feng Liang, Bichen Wu, Xiaoliang Dai, Kunpeng Li, Yinan Zhao, Hang Zhang, Peizhao Zhang, Peter Vajda, and Diana Marculescu. Open-vocabulary semantic segmentation with mask-adapted clip. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7061-7070, 2023. 1, 3 +[42] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 6 +[43] Quande Liu, Youpeng Wen, Jianhua Han, Chunjing Xu, Hang Xu, and Xiaodan Liang. Open-world semantic segmentation via contrasting and clustering vision-language embedding. In European Conference on Computer Vision, pages 275-292. Springer, 2022. 2, 3, 7 +[44] Huaishao Luo, Junwei Bao, Youzheng Wu, Xiaodong He, and Tianrui Li. Segclip: Patch aggregation with learnable centers for open-vocabulary semantic segmentation. In International Conference on Machine Learning, pages 23033-23044. PMLR, 2023. 1, 2, 3, 7 +[45] Roozbeh Mottaghi, Xianjie Chen, Xiaobai Liu, Nam-Gyu Cho, Seong-Whan Lee, Sanja Fidler, Raquel Urtasun, and Alan Yuille. The role of context for object detection and semantic segmentation in the wild. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 891–898, 2014. 6 +[46] Jishnu Mukhoti, Tsung-Yu Lin, Omid Poursaeed, Rui Wang, Ashish Shah, Philip HS Torr, and Ser-Nam Lim. Open vocabulary semantic segmentation with patch aligned contrastive learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19413-19423, 2023. 1, 3, 6, 7 + +[47] Giuseppe Pastore, Fabio Cermelli, Yongqin Xian, Massimiliano Mancini, Zeynep Akata, and Barbara Caputo. A closer look at self-training for zero-label semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2693-2702, 2021. 3, 7 +[48] Jie Qin, Jie Wu, Pengxiang Yan, Ming Li, Ren Yuxi, Xuefeng Xiao, Yitong Wang, Rui Wang, Shilei Wen, Xin Pan, et al. Freeseg: Unified, universal and open-vocabulary image segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19446-19455, 2023. 1 +[49] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In Proceedings of the 38th International Conference on Machine Learning, 2021. 2, 3 +[50] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 2 +[51] Kanchana Ranasinghe, Brandon McKinzie, Sachin Ravi, Yinfei Yang, Alexander Toshev, and Jonathon Shlens. Perceptual grouping in contrastive vision-language models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5571-5584, 2023. 2, 3, 7 +[52] Pengzhen Ren, Changlin Li, Hang Xu, Yi Zhu, Guangrun Wang, Jianzhuang Liu, Xiaojun Chang, and Xiaodan Liang. Viewco: Discovering text-supervised segmentation masks via multi-view semantic consistency. arXiv preprint arXiv:2302.10307, 2023. 2, 3, 7 +[53] Lixiang Ru, Yibing Zhan, Baosheng Yu, and Bo Du. Learning affinity from attention: End-to-end weakly-supervised semantic segmentation with transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16846-16855, 2022. 1 +[54] Lixiang Ru, Heliang Zheng, Yibing Zhan, and Bo Du. Token contrast for weakly-supervised semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3093-3102, 2023. 3 +[55] Lixiang Ru, Heliang Zheng, Yibing Zhan, and Bo Du. Token contrast for weakly-supervised semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3093-3102, 2023. 1 +[56] Michael Rubinstein, Armand Joulin, Johannes Kopf, and Ce Liu. Unsupervised joint object discovery and segmentation in internet images. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1939-1946, 2013. 3 +[57] Ramprasaath R. Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. Grad-cam: Visual explanations from deep networks via gradient-based localization. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2017. 2, 4 + +[58] Gyungin Shin, Weidi Xie, and Samuel Albanie. Reco: Retrieve and co-segment for zero-shot transfer. Advances in Neural Information Processing Systems, 35:33754-33767, 2022. 3, 6, 7 +[59] Oriane Simeoni, Gilles Puy, Huy V Vo, Simon Roburin, Spyros Gidaris, Andrei Bursuc, Patrick Pérez, Renaud Marlet, and Jean Ponce. Localizing objects with self-supervised transformers and no labels. arXiv preprint arXiv:2109.14279, 2021. 3 +[60] Amanpreet Singh, Ronghang Hu, Vedanuj Goswami, Guillaume Couairon, Wojciech Galuba, Marcus Rohrbach, and Douwe Kiela. Flava: A foundational language and vision alignment model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 15638-15650, 2022. 1, 2 +[61] Meng Tang, Federico Perazzi, Abdelaziz Djelouah, Ismail Ben Ayed, Christopher Schroers, and Yuri Boykov. On regularized losses for weakly-supervised cnn segmentation. In Proceedings of the European Conference on Computer Vision (ECCV), pages 507-522, 2018. 3 +[62] Anthony Meng Huat Tiong, Junnan Li, Boyang Li, Silvio Savarese, and Steven C.H. Hoi. Plug-and-play vqa: Zero-shot vqa by conjoining large pretrained models with zero training. In Findings of the Conference on Empirical Methods in Natural Language Processing (Findings of EMNLP), 2022. 1, 2, 4 +[63] Huy V Vo, Francis Bach, Minsu Cho, Kai Han, Yann LeCun, Patrick Pérez, and Jean Ponce. Unsupervised image matching and object discovery as optimization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8287-8296, 2019. 3 +[64] Van Huy Vo, Elena Sizikova, Cordelia Schmid, Patrick Pérez, and Jean Ponce. Large-scale unsupervised object discovery. Advances in Neural Information Processing Systems, 34:16764-16778, 2021. 3 +[65] Wenguan Wang, Guolei Sun, and Luc Van Gool. Looking beyond single images for weakly supervised semantic segmentation learning. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2022. 1 +[66] Xudong Wang, Rohit Girdhar, Stella X Yu, and Ishan Misra. Cut and learn for unsupervised object detection and instance segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3124-3134, 2023. 6 +[67] Yangtao Wang, Xi Shen, Shell Xu Hu, Yuan Yuan, James L Crowley, and Dominique Vaufreydaz. Self-supervised transformers for unsupervised object discovery using normalized cut. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14543-14553, 2022. 3 +[68] Xiu-Shen Wei, Chen-Lin Zhang, Jianxin Wu, Chunhua Shen, and Zhi-Hua Zhou. Unsupervised object discovery and co-localization by deep descriptor transformation. Pattern Recognition, 88:113-126, 2019. 3 +[69] Yunchao Wei, Jiashi Feng, Xiaodan Liang, Ming-Ming Cheng, Yao Zhao, and Shuicheng Yan. Object region mining with adversarial erasing: A simple classification to semantic + +segmentation approach. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1568-1576, 2017. 3 +[70] Alexandros Xenos, Themos Stafylakis, Ioannis Patras, and Georgios Tzimiropoulos. A simple baseline for knowledge-based visual question answering. In EMNLP, 2023. 1 +[71] Yongqin Xian, Subhabrata Choudhury, Yang He, Bernt Schiele, and Zeynep Akata. Semantic projection network for zero-and few-label semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8256-8265, 2019. 3, 7 +[72] Haiyang Xu, Qinghao Ye, Ming Yan, Yaya Shi, Jiabo Ye, Yuanhong Xu, Chenliang Li, Bin Bi, Qi Qian, Wei Wang, et al. mplug-2: A modularized multi-modal foundation model across text, image and video. arXiv preprint arXiv:2302.00402, 2023. 2, 3 +[73] Jiarui Xu, Shalini De Mello, Sifei Liu, Wonmin Byeon, Thomas Breuel, Jan Kautz, and Xiaolong Wang. Groupvit: Semantic segmentation emerges from text supervision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18134-18144, 2022. 2, 3, 7 +[74] Jilan Xu, Junlin Hou, Yuejie Zhang, Rui Feng, Yi Wang, Yu Qiao, and Weidi Xie. Learning open-vocabulary semantic segmentation models from natural language supervision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2935-2944, 2023. 1, 2, 3, 7 +[75] Lian Xu, Wanli Ouyang, Mohammed Bennamoun, Farid Boussaid, and Dan Xu. Multi-class token transformer for weakly supervised semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4310-4319, 2022. 1 +[76] Mengde Xu, Zheng Zhang, Fangyun Wei, Yutong Lin, Yue Cao, Han Hu, and Xiang Bai. A simple baseline for open-vocabulary semantic segmentation with pre-trained vision-language model. In European Conference on Computer Vision, pages 736-753. Springer, 2022. 1, 3, 7 +[77] Mengde Xu, Zheng Zhang, Fangyun Wei, Han Hu, and Xiang Bai. Side adapter network for open-vocabulary semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2945-2954, 2023. 1 +[78] Xiao Xu, Chenfei Wu, Shachar Rosenman, Vasudev Lal, Wanxiang Che, and Nan Duan. Bridgetower: Building bridges between encoders in vision-language representation learning. arXiv preprint arXiv:2206.08657, 2022. 1, 2, 3, 6, 8 +[79] Jiahui Yu, Zirui Wang, Vijay Vasudevan, Legg Yeung, Mojtaba Seyedhosseini, and Yonghui Wu. CoCa: Contrastive captioners are image-text foundation models. arXiv Preprint 2205.01917, 2022. 2 +[80] Nir Zabari and Yedid Hoshen. Open-vocabulary semantic segmentation using test-time distillation. In European Conference on Computer Vision, pages 56-72. Springer, 2022. 1 + +[81] Yan Zeng, Xinsong Zhang, and Hang Li. Multi-grained vision language pre-training: Aligning texts with visual concepts. arXiv preprint arXiv:2111.08276, 2021. 2, 3 +[82] Yan Zeng, Xinsong Zhang, Hang Li, Jiawei Wang, Jipeng Zhang, and Wangchunshu Zhou. $X^2$ -vlm: All-in-one pretrained model for vision-language tasks. arXiv preprint arXiv:2211.12402, 2022. 2, 3 +[83] Bingfeng Zhang, Jimin Xiao, Yunchao Wei, Mingjie Sun, and Kaizhu Huang. Reliability does matter: An end-to-end weakly supervised semantic segmentation approach. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 12765-12772, 2020. 3 +[84] Hao Zhang, Feng Li, Xueyan Zou, Shilong Liu, Chunyuan Li, Jianwei Yang, and Lei Zhang. A simple framework for open-vocabulary segmentation and detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1020-1031, 2023. 1 +[85] Pengchuan Zhang, Xiujun Li, Xiaowei Hu, Jianwei Yang, Lei Zhang, Lijuan Wang, Yejin Choi, and Jianfeng Gao. Vinvl: Revisiting visual representations in vision-language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5579-5588, 2021. 1 +[86] Xiaolin Zhang, Yunchao Wei, Jiashi Feng, Yi Yang, and Thomas S Huang. Adversarial complementary learning for weakly supervised object localization. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1325-1334, 2018. 3 +[87] Bolei Zhou, Aditya Khosla, Agata Lapedriza, Aude Oliva, and Antonio Torralba. Object detectors emerge in deep scene cnns. arXiv preprint arXiv:1412.6856, 2014. 3 +[88] Bolei Zhou, Hang Zhao, Xavier Puig, Sanja Fidler, Adela Barriuso, and Antonio Torralba. Scene parsing through ade20k dataset. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 633-641, 2017. 6 +[89] Chong Zhou, Chen Change Loy, and Bo Dai. Extract free dense labels from clip. In European Conference on Computer Vision, pages 696-712. Springer, 2022. 3, 6, 7 +[90] Tianfei Zhou, Meijie Zhang, Fang Zhao, and Jianwu Li. Regional semantic contrast and aggregation for weakly supervised semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4299-4309, 2022. 1 \ No newline at end of file diff --git a/2024/Emergent Open-Vocabulary Semantic Segmentation from Off-the-shelf Vision-Language Models/images.zip b/2024/Emergent Open-Vocabulary Semantic Segmentation from Off-the-shelf Vision-Language Models/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..56c53bc1355737154788eb3e7a23b445358a65bd --- /dev/null +++ b/2024/Emergent Open-Vocabulary Semantic Segmentation from Off-the-shelf Vision-Language Models/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a06a792e8e2f4b9aa9e3b540ee7e0332e5480d934d215aa00ef1680d69dbcc1a +size 719619 diff --git a/2024/Emergent Open-Vocabulary Semantic Segmentation from Off-the-shelf Vision-Language Models/layout.json b/2024/Emergent Open-Vocabulary Semantic Segmentation from Off-the-shelf Vision-Language Models/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..1ac4d0dad3d931114a7a832c1bcd7dc51a2a57f1 --- /dev/null +++ b/2024/Emergent Open-Vocabulary Semantic Segmentation from Off-the-shelf Vision-Language Models/layout.json @@ -0,0 +1,11024 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 78, + 103, + 516, + 138 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 78, + 103, + 516, + 138 + ], + "spans": [ + { + "bbox": [ + 78, + 103, + 516, + 138 + ], + "type": "text", + "content": "Emergent Open-Vocabulary Semantic Segmentation from Off-the-shelf Vision-Language Models" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 160, + 459, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 160, + 459, + 175 + ], + "spans": [ + { + "bbox": [ + 130, + 160, + 459, + 175 + ], + "type": "text", + "content": "Jiayun Luo" + }, + { + "bbox": [ + 130, + 160, + 459, + 175 + ], + "type": "inline_equation", + "content": "^{\\flat}" + }, + { + "bbox": [ + 130, + 160, + 459, + 175 + ], + "type": "text", + "content": ", Siddhesh Khandelwal" + }, + { + "bbox": [ + 130, + 160, + 459, + 175 + ], + "type": "inline_equation", + "content": "^{\\sharp}" + }, + { + "bbox": [ + 130, + 160, + 459, + 175 + ], + "type": "text", + "content": ", Leonid Sigal" + }, + { + "bbox": [ + 130, + 160, + 459, + 175 + ], + "type": "inline_equation", + "content": "^{\\sharp}" + }, + { + "bbox": [ + 130, + 160, + 459, + 175 + ], + "type": "text", + "content": ", and Boyang Li" + }, + { + "bbox": [ + 130, + 160, + 459, + 175 + ], + "type": "inline_equation", + "content": "^{\\flat}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 182, + 176, + 410, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 182, + 176, + 410, + 190 + ], + "spans": [ + { + "bbox": [ + 182, + 176, + 410, + 190 + ], + "type": "inline_equation", + "content": "^{b}" + }, + { + "bbox": [ + 182, + 176, + 410, + 190 + ], + "type": "text", + "content": "Nanyang Technological University, Singapore" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 141, + 190, + 452, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 190, + 452, + 203 + ], + "spans": [ + { + "bbox": [ + 141, + 190, + 452, + 203 + ], + "type": "inline_equation", + "content": "^{\\ddagger}" + }, + { + "bbox": [ + 141, + 190, + 452, + 203 + ], + "type": "text", + "content": "University of British Columbia, Vector Institute for AI, Canada" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 129, + 206, + 462, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 129, + 206, + 462, + 217 + ], + "spans": [ + { + "bbox": [ + 129, + 206, + 462, + 217 + ], + "type": "text", + "content": "{luoj0028,boyang.li}@ntu.edu.sg，{skhandel,lsignal}@cs.ubc.ca" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 50, + 246, + 127, + 352 + ], + "blocks": [ + { + "bbox": [ + 50, + 246, + 127, + 352 + ], + "lines": [ + { + "bbox": [ + 50, + 246, + 127, + 352 + ], + "spans": [ + { + "bbox": [ + 50, + 246, + 127, + 352 + ], + "type": "image", + "image_path": "1ff679333025df281c2c264dce0007ffa1e9b4392c0c8e79e9805419be2db86c.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 355, + 546, + 378 + ], + "lines": [ + { + "bbox": [ + 46, + 355, + 546, + 378 + ], + "spans": [ + { + "bbox": [ + 46, + 355, + 546, + 378 + ], + "type": "text", + "content": "Figure 1. Qualitative Results of PnP-OVSS + BLIP. Images are from Pascal Context and COCO stuff. The right columns and bottom rows show the ground-truth (GT); the rest are our results. Note accurate results even on complex (trees) and small objects (last column)." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 128, + 246, + 205, + 352 + ], + "blocks": [ + { + "bbox": [ + 128, + 246, + 205, + 352 + ], + "lines": [ + { + "bbox": [ + 128, + 246, + 205, + 352 + ], + "spans": [ + { + "bbox": [ + 128, + 246, + 205, + 352 + ], + "type": "image", + "image_path": "90f2b88f1e18d0f77460141d9566fdb4af766fb5e89e26f13e48fad8f438a73e.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 206, + 246, + 301, + 352 + ], + "blocks": [ + { + "bbox": [ + 206, + 246, + 301, + 352 + ], + "lines": [ + { + "bbox": [ + 206, + 246, + 301, + 352 + ], + "spans": [ + { + "bbox": [ + 206, + 246, + 301, + 352 + ], + "type": "image", + "image_path": "58858b884765120250fddbe6138fde61d893d84686039a0ddc10e510760a8523.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 302, + 245, + 394, + 352 + ], + "blocks": [ + { + "bbox": [ + 302, + 245, + 394, + 352 + ], + "lines": [ + { + "bbox": [ + 302, + 245, + 394, + 352 + ], + "spans": [ + { + "bbox": [ + 302, + 245, + 394, + 352 + ], + "type": "image", + "image_path": "bac028cc5228ca6a1d842d476f8201b7e0a41dcc5aae3f16c20b371a4c03b5d2.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 395, + 245, + 462, + 352 + ], + "blocks": [ + { + "bbox": [ + 395, + 245, + 462, + 352 + ], + "lines": [ + { + "bbox": [ + 395, + 245, + 462, + 352 + ], + "spans": [ + { + "bbox": [ + 395, + 245, + 462, + 352 + ], + "type": "image", + "image_path": "3c969cb703bd95a7da9929f65b58dca7a5b2f2d6b2fcccff5b2c7e717b2ce8bf.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 465, + 246, + 545, + 352 + ], + "blocks": [ + { + "bbox": [ + 465, + 246, + 545, + 352 + ], + "lines": [ + { + "bbox": [ + 465, + 246, + 545, + 352 + ], + "spans": [ + { + "bbox": [ + 465, + 246, + 545, + 352 + ], + "type": "image", + "image_path": "7c6356ae1fb78a8926503660df0845ab899375d8b32c5d739407f4ec2c6c41a2.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "bbox": [ + 143, + 388, + 192, + 400 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 388, + 192, + 400 + ], + "spans": [ + { + "bbox": [ + 143, + 388, + 192, + 400 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 45, + 413, + 289, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 413, + 289, + 700 + ], + "spans": [ + { + "bbox": [ + 45, + 413, + 289, + 700 + ], + "type": "text", + "content": "From image-text pairs, large-scale vision-language models (VLMs) learn to implicitly associate image regions with words, which prove effective for tasks like visual question answering. However, leveraging the learned association for open-vocabulary semantic segmentation remains a challenge. In this paper, we propose a simple, yet extremely effective, training-free technique, Plug-and-Play Open-Vocabulary Semantic Segmentation (PnP-OVSS) for this task. PnP-OVSS leverages a VLM with direct text-to-image cross-attention and an image-text matching loss. To balance between over-segmentation and under-segmentation, we introduce Salience Dropout; by iteratively dropping patches that the model is most attentive to, we are able to better resolve the entire extent of the segmentation mask. PnP-OVSS does not require any neural network training and performs hyperparameter tuning without the need for any segmentation annotations, even for a validation set. PnP-OVSS demonstrates substantial improvements over comparable baselines (+29.4% mIoU on Pascal VOC, +13.2% mIoU on Pascal Context, +14.0% mIoU on MS COCO, +2.4% mIoU on COCO Stuff) and even outperforms most baselines that conduct additional network training on top of pretrained VLMs. Our codebase is at https://github.com/letitiabana/PnP-OVSS." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 388, + 386, + 400 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 388, + 386, + 400 + ], + "spans": [ + { + "bbox": [ + 307, + 388, + 386, + 400 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 410, + 545, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 410, + 545, + 529 + ], + "spans": [ + { + "bbox": [ + 304, + 410, + 545, + 529 + ], + "type": "text", + "content": "The classic task of semantic segmentation [20, 22] aims to classify pixels to their object types. Traditional supervised methods require dense pixel-level annotations and are restricted to recognizing a predefined set of objects. To relax these constraints, open-vocabulary semantic segmentation [8, 17, 21, 41, 44, 46, 48, 74, 76, 77, 80, 84] aspires to identify arbitrary object categories, whereas weakly supervised techniques [9, 11, 15, 25, 31, 36, 53, 55, 65, 75, 90] can acquire pixel-level localization capabilities from coarse supervision, e.g., image labels or boxes." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 532, + 546, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 532, + 546, + 700 + ], + "spans": [ + { + "bbox": [ + 304, + 532, + 546, + 700 + ], + "type": "text", + "content": "Large-scale vision-language models (VLMs) pretrained on image-text pairs [1, 37, 38, 40, 60, 78, 85] achieve unprecedented performance on multimodal tasks, such as describing arbitrary images and answering free-form, open-ended questions about them (either with [37, 38, 60] or without finetuning [1, 19, 62, 70]). These tasks apparently involve some ability to localize objects. For example, to answer the question \"what objects appear on the table?\", the model would have to first localize the table in the image and identify the objects on it. Hence, it is reasonable to conjecture that the VLM network learns to perform open-vocabulary localization from image-text pretraining. However, distilling the localization capability from the VLMs remains an open challenge." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 701, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 701, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 317, + 701, + 545, + 713 + ], + "type": "text", + "content": "Most existing methods for open-vocabulary semantic" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4029" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 49, + 70, + 108, + 189 + ], + "blocks": [ + { + "bbox": [ + 49, + 70, + 108, + 189 + ], + "lines": [ + { + "bbox": [ + 49, + 70, + 108, + 189 + ], + "spans": [ + { + "bbox": [ + 49, + 70, + 108, + 189 + ], + "type": "image", + "image_path": "121ad0ade58b246149c5186512c220dafbf9ccfde40adf06e04955f94076089f.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 59, + 190, + 99, + 209 + ], + "lines": [ + { + "bbox": [ + 59, + 190, + 99, + 209 + ], + "spans": [ + { + "bbox": [ + 59, + 190, + 99, + 209 + ], + "type": "text", + "content": "(a) Original image" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 108, + 70, + 167, + 190 + ], + "blocks": [ + { + "bbox": [ + 108, + 70, + 167, + 190 + ], + "lines": [ + { + "bbox": [ + 108, + 70, + 167, + 190 + ], + "spans": [ + { + "bbox": [ + 108, + 70, + 167, + 190 + ], + "type": "image", + "image_path": "0cd3e661d89d18175c63de5f8e1f53b307d27537740551d5d8b239eea3fb8172.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 121, + 190, + 162, + 209 + ], + "lines": [ + { + "bbox": [ + 121, + 190, + 162, + 209 + ], + "spans": [ + { + "bbox": [ + 121, + 190, + 162, + 209 + ], + "type": "text", + "content": "(b) Cross-attention" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 219, + 288, + 274 + ], + "lines": [ + { + "bbox": [ + 46, + 219, + 288, + 274 + ], + "spans": [ + { + "bbox": [ + 46, + 219, + 288, + 274 + ], + "type": "text", + "content": "Figure 2. Segmentation masks for elephant and train using (a) off-the-shelf cross-attention, (b) cross-attention + GradCAM, and (c) cross-attention + GradCam + Salience DropOut (\\$3.2). The naive cross-attention masks are too inclusive whereas GradCAM is too exclusive." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 168, + 70, + 227, + 190 + ], + "blocks": [ + { + "bbox": [ + 168, + 70, + 227, + 190 + ], + "lines": [ + { + "bbox": [ + 168, + 70, + 227, + 190 + ], + "spans": [ + { + "bbox": [ + 168, + 70, + 227, + 190 + ], + "type": "image", + "image_path": "389d49edd686ef6e886871752288378ae63044a0c67c4a3cf04429552cd44238.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 174, + 191, + 220, + 200 + ], + "lines": [ + { + "bbox": [ + 174, + 191, + 220, + 200 + ], + "spans": [ + { + "bbox": [ + 174, + 191, + 220, + 200 + ], + "type": "text", + "content": "(c) GradCAM" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 227, + 70, + 288, + 190 + ], + "blocks": [ + { + "bbox": [ + 227, + 70, + 288, + 190 + ], + "lines": [ + { + "bbox": [ + 227, + 70, + 288, + 190 + ], + "spans": [ + { + "bbox": [ + 227, + 70, + 288, + 190 + ], + "type": "image", + "image_path": "6b6386f24e5d894b02003cba93d289f6a972b1d99c05955596708b16cc7751a1.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 237, + 190, + 278, + 210 + ], + "lines": [ + { + "bbox": [ + 237, + 190, + 278, + 210 + ], + "spans": [ + { + "bbox": [ + 237, + 190, + 278, + 210 + ], + "type": "text", + "content": "(d) Salience Dropout" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 289, + 287, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 289, + 287, + 373 + ], + "spans": [ + { + "bbox": [ + 46, + 289, + 287, + 373 + ], + "type": "text", + "content": "segmentation (OVSS) from VLMs usually obtain single vector encodings for the visual and text inputs respectively [43, 44, 51, 52, 73, 74]. However, pooling every token into a single vector likely discards information about detailed positions of objects and words. We investigate the use of pretrained cross-attention layers for OVSS, which retain finer-grained correspondence between text and image patches." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 376, + 288, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 376, + 288, + 544 + ], + "spans": [ + { + "bbox": [ + 46, + 376, + 288, + 544 + ], + "type": "text", + "content": "Nevertheless, a naive application of cross-attention between the class name and the image patches lead to overly broad segmentation masks that include irrelevant parts of the image (i.e., over-segmentation, see Fig. 2(b)). To alleviate this, [62] employs gradient information from the imagetext matching loss, GradCAM[57]-style, to sharpen the attention-based masks for the purpose of guiding caption generation. However, this results in masks that capture only the most discriminative regions of an object, such as the head of an elephant (i.e., under-segmentation, Fig. 2(c)). To acquire complete object masks, we propose Salience DropOut, which iteratively drops the image patches with high GradCAM attention scores, forcing the model to attend to less discriminative but relevant object parts." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 546, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 546, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 546, + 287, + 714 + ], + "type": "text", + "content": "Another important consideration is the cross-attention layer and the attention head to extract the masks from. These hyperparameters have enormous influence on the final results and are traditionally tuned on a validation set with pixel-level mask annotations. To eliminate the need for dense annotations, we propose an weakly-supervised reward function based on CLIP [49]. On a validation set with images as well as object class names, the technique contrasts the extracted object regions with a blank image. If, according to CLIP, the former is more similar to the corresponding class name than the latter, we increment the reward. All hyperparameter tuning of our technique is performed with a simple random search with this reward, leading to high performance." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 72, + 547, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 239 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 239 + ], + "type": "text", + "content": "In summary, we propose Plug-and-Play Open-vocabulary Semantic Segmentation (PnP-OVSS), an extremely simple and training-free framework to extract semantic segmentations from VLMs. At zero extra training cost, PnP-OVSS can be used with any pretrained VLM with text-to-image cross attention layer and an image-text matching loss. It has zero reliance on pixel-level annotations, including a validation set for hyperparameter tuning. At the same time, PnP-OVSS delivers excellent performance. It not only beats the training-free baseline with remarkable margins " + }, + { + "bbox": [ + 304, + 72, + 547, + 239 + ], + "type": "inline_equation", + "content": "(+29.4\\%)" + }, + { + "bbox": [ + 304, + 72, + 547, + 239 + ], + "type": "text", + "content": " mIoU on Pascal VOC, " + }, + { + "bbox": [ + 304, + 72, + 547, + 239 + ], + "type": "inline_equation", + "content": "+13.2\\%" + }, + { + "bbox": [ + 304, + 72, + 547, + 239 + ], + "type": "text", + "content": " on Pascal Context, " + }, + { + "bbox": [ + 304, + 72, + 547, + 239 + ], + "type": "inline_equation", + "content": "+14.0\\%" + }, + { + "bbox": [ + 304, + 72, + 547, + 239 + ], + "type": "text", + "content": " on MS COCO, " + }, + { + "bbox": [ + 304, + 72, + 547, + 239 + ], + "type": "inline_equation", + "content": "+2.4\\%" + }, + { + "bbox": [ + 304, + 72, + 547, + 239 + ], + "type": "text", + "content": " on COCO Stuff), but also outperforms most recent techniques within the past two years that require extensive finetuning on top of the VLM pretraining." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 317, + 240, + 503, + 252 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 240, + 503, + 252 + ], + "spans": [ + { + "bbox": [ + 317, + 240, + 503, + 252 + ], + "type": "text", + "content": "With this paper, we make three contributions:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 253, + 551, + 397 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 306, + 253, + 551, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 253, + 551, + 289 + ], + "spans": [ + { + "bbox": [ + 306, + 253, + 551, + 289 + ], + "type": "text", + "content": "- We propose to combine text-to-image attention, GradCAM, and Salience DropOut to iteratively acquire accurate segmentation of arbitrary classes from pretrained VLM." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 289, + 545, + 348 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 289, + 545, + 348 + ], + "spans": [ + { + "bbox": [ + 306, + 289, + 545, + 348 + ], + "type": "text", + "content": "- We replace the densely annotated validation set for hyperparameter tuning, which is needed by most existing methods, with a contrastive reward function based on CLIP. This reward function, coupled with random search, finds a good set of hyperparameters for OVSS." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 349, + 545, + 397 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 349, + 545, + 397 + ], + "spans": [ + { + "bbox": [ + 306, + 349, + 545, + 397 + ], + "type": "text", + "content": "- The proposed method, PnP-OVSS, is simple to use, requires no extra finetuning, and delivers high performance. Its success hints at a new direction for open-vocabulary segmentation tasks leveraging large VLMs." + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 410, + 392, + 422 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 410, + 392, + 422 + ], + "spans": [ + { + "bbox": [ + 306, + 410, + 392, + 422 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 306, + 430, + 499, + 443 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 430, + 499, + 443 + ], + "spans": [ + { + "bbox": [ + 306, + 430, + 499, + 443 + ], + "type": "text", + "content": "2.1. Large-Scale Vision-Language Model" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 449, + 547, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 449, + 547, + 653 + ], + "spans": [ + { + "bbox": [ + 304, + 449, + 547, + 653 + ], + "type": "text", + "content": "Large-scale vision-language models (VLMs), trained on millions of image-text pairs, have become the foundation for many multimodal tasks. Architecturally, the straightforward approach to training such methods involves aligning visual and textual latent representations via a simple dot product [24, 49]. However, this is insufficient for complex structured tasks like visual question answering or image captioning, which require specialized approaches that employ separate encoders before cross-attention between modalities [33, 35, 37, 78, 79] or self-attention networks over all tokens from both modalities [10, 27, 39, 40, 60]. Another design dimension when training VLMs is the loss function. Commonly used losses include image-text contrastive learning [34, 35, 37, 50, 60, 81], image-text matching (ITM) [27, 34, 35, 37, 60, 72, 78, 81, 82], prediction of masked tokens or patches [34, 60, 78, 81], and language modeling [1, 7, 34, 37]." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 654, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 547, + 715 + ], + "type": "text", + "content": "This work utilizes models with unimodal encoders followed by cross-attention fusion [34, 35, 37, 72, 78, 81, 82], as they work with high-level features, and accurately attend to the appropriate image patches. Additionally, we utilize the gradient from the ITM loss in the GradCAM step (§3.1)" + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4030" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 70, + 70, + 265, + 324 + ], + "blocks": [ + { + "bbox": [ + 70, + 70, + 265, + 324 + ], + "lines": [ + { + "bbox": [ + 70, + 70, + 265, + 324 + ], + "spans": [ + { + "bbox": [ + 70, + 70, + 265, + 324 + ], + "type": "table", + "html": "
MethodPT Data SizeFT data size
Requires finetuning on image-text pairs
OVSegmentor [74]-4.3M
Vil-Seg [43]400M412M
GroupVit* [73]-3.4M
GroupVit [73]-26M
CLIPpy [51]-134M
SegClip [44]400M3.4M
ViewCo [52]-26M
TCL [6]400M15M
PACL [46]400M30M
Requires finetuning but not image-text pairs
MaskClip w/ ST [89]400M1.2M
ZeroSeg* [8]400M3.4M
ZeroSeg [8]400M1.2M
Requires no finetuning
MaskClip[89]400M0
Reco[58]400M0
PnP-OVSS (Ours)
+ BLIP129M0
+ BridgeTower400M+ 4M0
", + "image_path": "cc54eecb1e6ce837da52ebc603041c25c6481b63abd560afabe7988417314cd5.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 331, + 288, + 411 + ], + "lines": [ + { + "bbox": [ + 46, + 331, + 288, + 411 + ], + "spans": [ + { + "bbox": [ + 46, + 331, + 288, + 411 + ], + "type": "text", + "content": "Table 1. Training data of current Zero-shot semantic segmentation methods with only text supervision. PT stands for pretraining with image-caption data, FT stands for fine-tuning with image-caption data. ST stands for self training. We list only the image-caption data used for pretraining and all type of data for finetuning. For hyperparameter tuning, our method uses CLIP-L/14 pretrained with 400M data to calculate the reward." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 47, + 432, + 178, + 444 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 432, + 178, + 444 + ], + "spans": [ + { + "bbox": [ + 47, + 432, + 178, + 444 + ], + "type": "text", + "content": "to sharpen segmentation masks." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 453, + 228, + 466 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 453, + 228, + 466 + ], + "spans": [ + { + "bbox": [ + 47, + 453, + 228, + 466 + ], + "type": "text", + "content": "2.2. Zero-shot Semantic Segmentation" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 472, + 287, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 472, + 287, + 568 + ], + "spans": [ + { + "bbox": [ + 46, + 472, + 287, + 568 + ], + "type": "text", + "content": "Zero-shot semantic segmentation predicts a dense segmentation mask for any object class described by a given text prompt, with only prior exposure to class-agnostic image-level supervision. This contrasts with weakly supervised semantic segmentation [3, 13, 23, 26, 28, 30, 54, 61, 69, 69, 83, 86, 87], which relies on class-specific annotations, and unsupervised object discovery [12, 14, 56, 59, 63, 64, 67, 68], which identifies the sole object in the foreground." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 569, + 287, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 569, + 287, + 677 + ], + "spans": [ + { + "bbox": [ + 46, + 569, + 287, + 677 + ], + "type": "text", + "content": "Traditional methods [4, 18, 47, 71] train a classifier to distinguish between seen and unseen visual features, wherein the unseen visual features are obtained from a generative model, trained on pairs of seen class image-text embeddings. Recently, methods additionally leverage knowledge from VLMs to attain better matching of visual and textual features [17, 32, 41, 76, 89]. Concretely, they train the segmentation network with dense annotations, while replacing a part of the framework with components from VLMs." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "text", + "content": "Recently, methods have explored the use of additional supervision to further reduce the need for pixel level annotations. Prior work is compared in Table 1." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 72, + 545, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 216 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 216 + ], + "type": "text", + "content": "Models Finetuned on Image-Text Pairs. Methods proposed in [6, 43, 44, 46, 51, 52, 73, 74] require paired image-text annotations to adapt to the zero-shot segmentation task. Specifically, approaches in [43, 44, 51, 52, 73, 74] cluster semantically similar pixels by contrasting grouped region embeddings with textual embeddings, usually via the use of contrastive and self-supervised losses. PACL [46] modifies the constrastive loss to operate over aggregated patch embeddings (instead of image embeddings) to encourage better alignment between image patches and text. TCL [6] achieves a similar patch-text alignment by introducing an additional module to extract text-grounded image regions." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 220, + 545, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 220, + 545, + 327 + ], + "spans": [ + { + "bbox": [ + 304, + 220, + 545, + 327 + ], + "type": "text", + "content": "Models Finetuned on Pseudo-labels. Methods in [8, 89] do not require additional image-text supervision, but involve additional fine-tuning to adapt to the segmentation task. ZeroSeg [8] attempts to match its own group embeddings with multi-scaled segment embeddings obtained from CLIP [49]. MaskClip w/ ST [89] modifies the global attention pooling layer within CLIP to output segmentation masks, which are used as pseudo-labels to train a segmentation network." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 329, + 545, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 329, + 545, + 449 + ], + "spans": [ + { + "bbox": [ + 304, + 329, + 545, + 449 + ], + "type": "text", + "content": "In comparison, our proposed PnP-OVSS does not require any fine-tuning or additional paired image-text annotations. It can directly distill high quality open vocabulary semantic segmentations from any VLM with direct text-to-image cross-attention and an image-text matching loss [34, 35, 37, 72, 78, 81, 82]. Compared to approaches in [58, 89] that perform zero-shot open vocabulary semantic segmentation under a similar no training and no additional annotations paradigm, PnP-OVSS achieves considerably superior performance." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 463, + 361, + 475 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 463, + 361, + 475 + ], + "spans": [ + { + "bbox": [ + 306, + 463, + 361, + 475 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 483, + 545, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 483, + 545, + 568 + ], + "spans": [ + { + "bbox": [ + 304, + 483, + 545, + 568 + ], + "type": "text", + "content": "PnP-OVSS has four major steps. First, we extract a cross-attention salience map per object class from a VLM. Second, we sharpen the salience map by weighing it with the ITM gradient in the style of GradCAM. Third, we apply Salience DropOut that iteratively completes the salience maps. Fourth, we apply Dense CRF [29] for fine-grained adjustment. The process is illustrated in Fig. 3 and Fig. 4." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 569, + 545, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 569, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 304, + 569, + 545, + 689 + ], + "type": "text", + "content": "In the first step, we feed an image and a text prompt to the pretrained VLM and extract cross-attention maps. The text prompt is \"A picture of [class 1] [class 2] ... [class K]\", which includes all " + }, + { + "bbox": [ + 304, + 569, + 545, + 689 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 304, + 569, + 545, + 689 + ], + "type": "text", + "content": " class names of interest in the dataset. The image is divided into " + }, + { + "bbox": [ + 304, + 569, + 545, + 689 + ], + "type": "inline_equation", + "content": "P \\times P" + }, + { + "bbox": [ + 304, + 569, + 545, + 689 + ], + "type": "text", + "content": " patches. The text prompt and the image first go through the modality-specific encoders respectively, followed by a cross-attention fusion module. In the cross-attention layers, the text encodings serve as query vectors and the image patche encodings are as key and value vectors." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "content": "We extract an attention map for each text token to the image patches from a particular cross-attention layer and an" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "4031" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 62, + 70, + 539, + 315 + ], + "blocks": [ + { + "bbox": [ + 62, + 70, + 539, + 315 + ], + "lines": [ + { + "bbox": [ + 62, + 70, + 539, + 315 + ], + "spans": [ + { + "bbox": [ + 62, + 70, + 539, + 315 + ], + "type": "image", + "image_path": "c09ee7dc9b163aecd63adce78dc64440bff4ac145ac5613a803270cd95bde287.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 327, + 548, + 392 + ], + "lines": [ + { + "bbox": [ + 46, + 327, + 548, + 392 + ], + "spans": [ + { + "bbox": [ + 46, + 327, + 548, + 392 + ], + "type": "text", + "content": "Figure 3. The first iteration with cross-attention + GradCAM + Salience DropOut. The text prompt contains " + }, + { + "bbox": [ + 46, + 327, + 548, + 392 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 46, + 327, + 548, + 392 + ], + "type": "text", + "content": " class names and the image contains " + }, + { + "bbox": [ + 46, + 327, + 548, + 392 + ], + "type": "inline_equation", + "content": "P \\times P" + }, + { + "bbox": [ + 46, + 327, + 548, + 392 + ], + "type": "text", + "content": " patches. From a cross-attention layer and an attention head in the pretrained VLM, we obtain " + }, + { + "bbox": [ + 46, + 327, + 548, + 392 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 46, + 327, + 548, + 392 + ], + "type": "text", + "content": " attention score maps of size " + }, + { + "bbox": [ + 46, + 327, + 548, + 392 + ], + "type": "inline_equation", + "content": "P \\times P" + }, + { + "bbox": [ + 46, + 327, + 548, + 392 + ], + "type": "text", + "content": ", which are sharpened by GradCAM using gradients from the image-text-matching (ITM) loss. To get more complete predictions, we perform Salience Dropout, which repeatedly zero out image patches of the highest average scores and feeds the remaining patches to the image encoder again, forcing the model to attend to other less discriminative patches. We show example salience maps from all iterations in Fig. 4." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 414, + 287, + 497 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 414, + 287, + 497 + ], + "spans": [ + { + "bbox": [ + 46, + 414, + 287, + 497 + ], + "type": "text", + "content": "attention head. Note that different layers and heads lead to drastic performance differences, and the choices are hyperparameters, tuned using the procedure in §3.4. We exclude attention maps for the first three tokens \"A\", \"picture\", \"of\", which do not describe semantic classes. If a class name contains two or more tokens, we take the mean attention map. This procedure yields an attention tensor of size " + }, + { + "bbox": [ + 46, + 414, + 287, + 497 + ], + "type": "inline_equation", + "content": "K \\times P \\times P" + }, + { + "bbox": [ + 46, + 414, + 287, + 497 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 498, + 288, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 498, + 288, + 617 + ], + "spans": [ + { + "bbox": [ + 46, + 498, + 288, + 617 + ], + "type": "text", + "content": "With the correct layer and attention head, we observe that this attention map, when normalized by softmax along the first dimension " + }, + { + "bbox": [ + 46, + 498, + 288, + 617 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 46, + 498, + 288, + 617 + ], + "type": "text", + "content": ", can provide passable semantic segmentations. However, it tends to include many patches unrelated to the class being segmented, leading to over-segmentation. Hence, we introduce two refinement steps, GradCAM and Salience DropOut, explained in §3.1 and §3.2. After these steps, we acquire aggregate salience maps for every class. Finally, we conduct local polish to the resultant salience maps using Dense CRF, as described in §3.3." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 623, + 226, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 623, + 226, + 635 + ], + "spans": [ + { + "bbox": [ + 47, + 623, + 226, + 635 + ], + "type": "text", + "content": "3.1. Map Sharpening with GradCAM" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 642, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 642, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 642, + 288, + 714 + ], + "type": "text", + "content": "The off-the-shelf attention maps tend to also cover many patches unrelated to the class name (See Fig. 2 (b)). Prior works [35, 62] further sharpen the attention maps and focus them on class-discriminative regions using a variant of GradCAM [57], which is originally proposed for convolutional networks, but applicable to attention maps. It is worth" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 414, + 545, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 414, + 545, + 437 + ], + "spans": [ + { + "bbox": [ + 304, + 414, + 545, + 437 + ], + "type": "text", + "content": "noting that [35, 62] use the technique for purposes other than semantic segmentation." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 438, + 546, + 570 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 438, + 546, + 570 + ], + "spans": [ + { + "bbox": [ + 304, + 438, + 546, + 570 + ], + "type": "text", + "content": "The GradCAM method requires a gradient. Here we leverage the image-text matching (ITM) loss, which trains the VLM to classify if an image-text pair match each other or not. Computing the ITM loss requires a label. We use \"matching\" (as opposed to \"not matching\") as the label and compute the gradient of the loss with respect to the attention score. This is equivalent to asking: which attention scores contribute the most to the decision that the image-text pair is matching? Formally, we denote a " + }, + { + "bbox": [ + 304, + 438, + 546, + 570 + ], + "type": "inline_equation", + "content": "P \\times P" + }, + { + "bbox": [ + 304, + 438, + 546, + 570 + ], + "type": "text", + "content": " attention map for class " + }, + { + "bbox": [ + 304, + 438, + 546, + 570 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 438, + 546, + 570 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 304, + 438, + 546, + 570 + ], + "type": "inline_equation", + "content": "M^{(k)}" + }, + { + "bbox": [ + 304, + 438, + 546, + 570 + ], + "type": "text", + "content": " and the ITM loss as " + }, + { + "bbox": [ + 304, + 438, + 546, + 570 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{ITM}}" + }, + { + "bbox": [ + 304, + 438, + 546, + 570 + ], + "type": "text", + "content": ". The GradCAM class salience map is computed as" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 351, + 578, + 545, + 605 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 351, + 578, + 545, + 605 + ], + "spans": [ + { + "bbox": [ + 351, + 578, + 545, + 605 + ], + "type": "interline_equation", + "content": "\\widetilde {M} ^ {(k)} = \\max \\left(0, \\frac {\\partial \\mathcal {L} _ {\\mathrm {I T M}}}{\\partial M ^ {(k)}}\\right) \\otimes M ^ {(k)}, \\tag {1}", + "image_path": "3fa0052c49ebc4b38e3848efa5d151fa7e7fb41c72387c656428680ee5eb59e0.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 614, + 545, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 614, + 545, + 639 + ], + "spans": [ + { + "bbox": [ + 304, + 614, + 545, + 639 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 614, + 545, + 639 + ], + "type": "inline_equation", + "content": "\\otimes" + }, + { + "bbox": [ + 304, + 614, + 545, + 639 + ], + "type": "text", + "content": " denote the component-wise multiplication and " + }, + { + "bbox": [ + 304, + 614, + 545, + 639 + ], + "type": "inline_equation", + "content": "\\max (\\cdot)" + }, + { + "bbox": [ + 304, + 614, + 545, + 639 + ], + "type": "text", + "content": " is also applied component-wise." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 647, + 411, + 659 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 647, + 411, + 659 + ], + "spans": [ + { + "bbox": [ + 306, + 647, + 411, + 659 + ], + "type": "text", + "content": "3.2. Salience Dropout" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 665, + 546, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 546, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 546, + 714 + ], + "type": "text", + "content": "As illustrated in Fig. 2 (c), segmentations generated by the GradCAM-style re-weighting of cross-attention are often narrowly focused on the most discriminative regions for a given class. However, the less discriminative regions are" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4032" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 51, + 87, + 100, + 133 + ], + "blocks": [ + { + "bbox": [ + 56, + 76, + 169, + 85 + ], + "lines": [ + { + "bbox": [ + 56, + 76, + 169, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 76, + 169, + 85 + ], + "type": "text", + "content": "GradCAM Class Salience Map" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 51, + 87, + 100, + 133 + ], + "lines": [ + { + "bbox": [ + 51, + 87, + 100, + 133 + ], + "spans": [ + { + "bbox": [ + 51, + 87, + 100, + 133 + ], + "type": "image", + "image_path": "d3a733835318b2e7e0db8bcce74e07c5564f79f60a3435f7f61dc0c279b0cea1.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 107, + 87, + 159, + 133 + ], + "blocks": [ + { + "bbox": [ + 107, + 87, + 159, + 133 + ], + "lines": [ + { + "bbox": [ + 107, + 87, + 159, + 133 + ], + "spans": [ + { + "bbox": [ + 107, + 87, + 159, + 133 + ], + "type": "image", + "image_path": "32656e85b9ec5cf0849b5761447164fe7ad2eebf1d1ac65da6597c7dd2442342.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 62, + 136, + 127, + 146 + ], + "lines": [ + { + "bbox": [ + 62, + 136, + 127, + 146 + ], + "spans": [ + { + "bbox": [ + 62, + 136, + 127, + 146 + ], + "type": "inline_equation", + "content": "1^{\\mathrm{st}}" + }, + { + "bbox": [ + 62, + 136, + 127, + 146 + ], + "type": "text", + "content": " Salience Drop" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 165, + 87, + 217, + 133 + ], + "blocks": [ + { + "bbox": [ + 165, + 87, + 217, + 133 + ], + "lines": [ + { + "bbox": [ + 165, + 87, + 217, + 133 + ], + "spans": [ + { + "bbox": [ + 165, + 87, + 217, + 133 + ], + "type": "image", + "image_path": "744a6862317d00d58f745171488df677d32103b6c246ac1b7cc3214e7d9ebd8f.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 222, + 87, + 275, + 133 + ], + "blocks": [ + { + "bbox": [ + 222, + 87, + 275, + 133 + ], + "lines": [ + { + "bbox": [ + 222, + 87, + 275, + 133 + ], + "spans": [ + { + "bbox": [ + 222, + 87, + 275, + 133 + ], + "type": "image", + "image_path": "73bb98368c603e8ff1de7b26f553ba5d08e4a8b3b06f34ef34f3eedfc247b85f.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 50, + 149, + 102, + 196 + ], + "blocks": [ + { + "bbox": [ + 50, + 149, + 102, + 196 + ], + "lines": [ + { + "bbox": [ + 50, + 149, + 102, + 196 + ], + "spans": [ + { + "bbox": [ + 50, + 149, + 102, + 196 + ], + "type": "image", + "image_path": "238cdefc313455f4837881828fad3a25e3f4d543b91cad0b118a075cad197d40.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 61, + 198, + 128, + 208 + ], + "lines": [ + { + "bbox": [ + 61, + 198, + 128, + 208 + ], + "spans": [ + { + "bbox": [ + 61, + 198, + 128, + 208 + ], + "type": "inline_equation", + "content": "2^{\\text{nd}}" + }, + { + "bbox": [ + 61, + 198, + 128, + 208 + ], + "type": "text", + "content": " Salience Drop" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 110, + 149, + 162, + 195 + ], + "blocks": [ + { + "bbox": [ + 110, + 149, + 162, + 195 + ], + "lines": [ + { + "bbox": [ + 110, + 149, + 162, + 195 + ], + "spans": [ + { + "bbox": [ + 110, + 149, + 162, + 195 + ], + "type": "image", + "image_path": "74884e061dfabf8267a7181b21c527458224149a23deacfee289b1e7596c6ea4.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 165, + 149, + 217, + 195 + ], + "blocks": [ + { + "bbox": [ + 165, + 149, + 217, + 195 + ], + "lines": [ + { + "bbox": [ + 165, + 149, + 217, + 195 + ], + "spans": [ + { + "bbox": [ + 165, + 149, + 217, + 195 + ], + "type": "image", + "image_path": "7322f1594217095b3fdd8220e094f3ac5ff1fe9df76a83c16e1654366d684284.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 222, + 149, + 274, + 195 + ], + "blocks": [ + { + "bbox": [ + 222, + 149, + 274, + 195 + ], + "lines": [ + { + "bbox": [ + 222, + 149, + 274, + 195 + ], + "spans": [ + { + "bbox": [ + 222, + 149, + 274, + 195 + ], + "type": "image", + "image_path": "edfb58038c4d94ee7569afbb4dadbdf199a4daf88bfeefc766b8b51160d8e613.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 50, + 211, + 101, + 257 + ], + "blocks": [ + { + "bbox": [ + 50, + 211, + 101, + 257 + ], + "lines": [ + { + "bbox": [ + 50, + 211, + 101, + 257 + ], + "spans": [ + { + "bbox": [ + 50, + 211, + 101, + 257 + ], + "type": "image", + "image_path": "d3f610134af2c85a76beac4c6dc4a4159c4171cb2ba8d7495325d7cdad7ed4c9.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 62, + 259, + 127, + 269 + ], + "lines": [ + { + "bbox": [ + 62, + 259, + 127, + 269 + ], + "spans": [ + { + "bbox": [ + 62, + 259, + 127, + 269 + ], + "type": "inline_equation", + "content": "3^{\\text{rd}}" + }, + { + "bbox": [ + 62, + 259, + 127, + 269 + ], + "type": "text", + "content": " Salience Drop" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 110, + 211, + 160, + 255 + ], + "blocks": [ + { + "bbox": [ + 110, + 211, + 160, + 255 + ], + "lines": [ + { + "bbox": [ + 110, + 211, + 160, + 255 + ], + "spans": [ + { + "bbox": [ + 110, + 211, + 160, + 255 + ], + "type": "image", + "image_path": "7d7cef942ce057cb60a00774c0a66d4a2014fae01734de2c4a9236efa962ab35.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 165, + 211, + 217, + 256 + ], + "blocks": [ + { + "bbox": [ + 165, + 211, + 217, + 256 + ], + "lines": [ + { + "bbox": [ + 165, + 211, + 217, + 256 + ], + "spans": [ + { + "bbox": [ + 165, + 211, + 217, + 256 + ], + "type": "image", + "image_path": "599dd1680e78322b71cf4320a1fbabef424abcf2dcdc92cf320d90fb9eae7357.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 222, + 211, + 276, + 256 + ], + "blocks": [ + { + "bbox": [ + 222, + 211, + 276, + 256 + ], + "lines": [ + { + "bbox": [ + 222, + 211, + 276, + 256 + ], + "spans": [ + { + "bbox": [ + 222, + 211, + 276, + 256 + ], + "type": "image", + "image_path": "1ab3915806d0c865ae4022e60b1548c2f6b798a20cbcbc5e2618c2ac9be3b6cb.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 51, + 273, + 103, + 319 + ], + "blocks": [ + { + "bbox": [ + 51, + 273, + 103, + 319 + ], + "lines": [ + { + "bbox": [ + 51, + 273, + 103, + 319 + ], + "spans": [ + { + "bbox": [ + 51, + 273, + 103, + 319 + ], + "type": "image", + "image_path": "574affc845379e36ae8702dd04be0f2832554dd1dbd960ddc799e5ed2e4962a8.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 61, + 322, + 127, + 332 + ], + "lines": [ + { + "bbox": [ + 61, + 322, + 127, + 332 + ], + "spans": [ + { + "bbox": [ + 61, + 322, + 127, + 332 + ], + "type": "inline_equation", + "content": "4^{\\text{th}}" + }, + { + "bbox": [ + 61, + 322, + 127, + 332 + ], + "type": "text", + "content": " Salience Drop" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 108, + 273, + 160, + 319 + ], + "blocks": [ + { + "bbox": [ + 108, + 273, + 160, + 319 + ], + "lines": [ + { + "bbox": [ + 108, + 273, + 160, + 319 + ], + "spans": [ + { + "bbox": [ + 108, + 273, + 160, + 319 + ], + "type": "image", + "image_path": "8dd8cae2b3a507aba4606c3f75e8484f63a74de2ebb0238f04b22a132dfb59cf.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 165, + 273, + 217, + 319 + ], + "blocks": [ + { + "bbox": [ + 165, + 273, + 217, + 319 + ], + "lines": [ + { + "bbox": [ + 165, + 273, + 217, + 319 + ], + "spans": [ + { + "bbox": [ + 165, + 273, + 217, + 319 + ], + "type": "image", + "image_path": "a0e00ed28d66eb4d9aff998fcf6aa08e57dac8ccd3f011a844301dc7d6463c52.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 222, + 273, + 276, + 319 + ], + "blocks": [ + { + "bbox": [ + 222, + 273, + 276, + 319 + ], + "lines": [ + { + "bbox": [ + 222, + 273, + 276, + 319 + ], + "spans": [ + { + "bbox": [ + 222, + 273, + 276, + 319 + ], + "type": "image", + "image_path": "2d5cc0608631b234445baed22a92557952a986e72343a55c1560eed6638d0a7e.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 51, + 335, + 104, + 382 + ], + "blocks": [ + { + "bbox": [ + 51, + 335, + 104, + 382 + ], + "lines": [ + { + "bbox": [ + 51, + 335, + 104, + 382 + ], + "spans": [ + { + "bbox": [ + 51, + 335, + 104, + 382 + ], + "type": "image", + "image_path": "4c8c4c001c6c33acf2f5c3cc255a13ccfc0fc59c4a7df3b9456ddbfbcb3e2a1e.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 108, + 335, + 159, + 382 + ], + "blocks": [ + { + "bbox": [ + 108, + 335, + 159, + 382 + ], + "lines": [ + { + "bbox": [ + 108, + 335, + 159, + 382 + ], + "spans": [ + { + "bbox": [ + 108, + 335, + 159, + 382 + ], + "type": "image", + "image_path": "ada2590fe21afaeef9cf5a094d9e93eb679042206ded94d0f56620920dfb9831.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 138, + 388, + 190, + 398 + ], + "lines": [ + { + "bbox": [ + 138, + 388, + 190, + 398 + ], + "spans": [ + { + "bbox": [ + 138, + 388, + 190, + 398 + ], + "type": "text", + "content": "Thresholding" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 165, + 335, + 216, + 382 + ], + "blocks": [ + { + "bbox": [ + 165, + 335, + 216, + 382 + ], + "lines": [ + { + "bbox": [ + 165, + 335, + 216, + 382 + ], + "spans": [ + { + "bbox": [ + 165, + 335, + 216, + 382 + ], + "type": "image", + "image_path": "8b5d38c510f3efa6df5c28d36d86aa7a70ba01e01594fcb7527ea928a0eddfdf.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 137, + 407, + 191, + 416 + ], + "lines": [ + { + "bbox": [ + 137, + 407, + 191, + 416 + ], + "spans": [ + { + "bbox": [ + 137, + 407, + 191, + 416 + ], + "type": "text", + "content": "Gaussian Blur" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_caption" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 222, + 335, + 276, + 382 + ], + "blocks": [ + { + "bbox": [ + 222, + 335, + 276, + 382 + ], + "lines": [ + { + "bbox": [ + 222, + 335, + 276, + 382 + ], + "spans": [ + { + "bbox": [ + 222, + 335, + 276, + 382 + ], + "type": "image", + "image_path": "74dccded38d1fd993f7f9dc49dafac7db0251247dffd40af1cb61aec99dbc44e.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 51, + 441, + 103, + 490 + ], + "blocks": [ + { + "bbox": [ + 51, + 441, + 103, + 490 + ], + "lines": [ + { + "bbox": [ + 51, + 441, + 103, + 490 + ], + "spans": [ + { + "bbox": [ + 51, + 441, + 103, + 490 + ], + "type": "image", + "image_path": "c25ee5c291fab16c727494f080edec705ac1602c98abb8c60ef1bc61f2fdfc4b.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 499, + 287, + 555 + ], + "lines": [ + { + "bbox": [ + 46, + 499, + 287, + 555 + ], + "spans": [ + { + "bbox": [ + 46, + 499, + 287, + 555 + ], + "type": "text", + "content": "Figure 4. An illustration of Salience DropOut, showing GradCAM salience values after each iteration. Black squares in the images indicate dropped patches. We obtain the final result by summing the salience maps from all iterations and applying thresholding, Gaussian blur, and Dense CRF." + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_caption" + } + ], + "index": 28 + }, + { + "type": "image", + "bbox": [ + 107, + 441, + 159, + 490 + ], + "blocks": [ + { + "bbox": [ + 143, + 426, + 185, + 435 + ], + "lines": [ + { + "bbox": [ + 143, + 426, + 185, + 435 + ], + "spans": [ + { + "bbox": [ + 143, + 426, + 185, + 435 + ], + "type": "text", + "content": "Dense CRF" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 107, + 441, + 159, + 490 + ], + "lines": [ + { + "bbox": [ + 107, + 441, + 159, + 490 + ], + "spans": [ + { + "bbox": [ + 107, + 441, + 159, + 490 + ], + "type": "image", + "image_path": "d15c7bcad4504082e1f2f8e7978b68e10b08dfb951ba47e8fe4d4c9723144036.jpg" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_body" + } + ], + "index": 29 + }, + { + "type": "image", + "bbox": [ + 164, + 441, + 217, + 490 + ], + "blocks": [ + { + "bbox": [ + 164, + 441, + 217, + 490 + ], + "lines": [ + { + "bbox": [ + 164, + 441, + 217, + 490 + ], + "spans": [ + { + "bbox": [ + 164, + 441, + 217, + 490 + ], + "type": "image", + "image_path": "71cda47ac3f80029bb03cc0bb58cac64d88f81c1227bad5c911df2592ffd00c3.jpg" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_body" + } + ], + "index": 30 + }, + { + "type": "image", + "bbox": [ + 222, + 441, + 276, + 490 + ], + "blocks": [ + { + "bbox": [ + 222, + 441, + 276, + 490 + ], + "lines": [ + { + "bbox": [ + 222, + 441, + 276, + 490 + ], + "spans": [ + { + "bbox": [ + 222, + 441, + 276, + 490 + ], + "type": "image", + "image_path": "e53037d5bc68fc58c941803f94421208306468a2776e930d7028434eb79bc6ed.jpg" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_body" + } + ], + "index": 31 + }, + { + "bbox": [ + 46, + 577, + 287, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 577, + 287, + 613 + ], + "spans": [ + { + "bbox": [ + 46, + 577, + 287, + 613 + ], + "type": "text", + "content": "still important for the completeness of masks. To compel the VLM to attend to these regions, we propose an iterative technique called Salience DropOut." + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 46, + 613, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 613, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 613, + 288, + 714 + ], + "type": "text", + "content": "Since our task is zero-shot and open-vocabulary, we do not have prior knowledge of the classes present in the image. Hence, we sum up the salience maps " + }, + { + "bbox": [ + 46, + 613, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\widetilde{M}^{(k)}" + }, + { + "bbox": [ + 46, + 613, + 288, + 714 + ], + "type": "text", + "content": " over all classes " + }, + { + "bbox": [ + 46, + 613, + 288, + 714 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 613, + 288, + 714 + ], + "type": "text", + "content": ", yielding a class-agnostic salience map " + }, + { + "bbox": [ + 46, + 613, + 288, + 714 + ], + "type": "inline_equation", + "content": "U^{(t)}" + }, + { + "bbox": [ + 46, + 613, + 288, + 714 + ], + "type": "text", + "content": " for the " + }, + { + "bbox": [ + 46, + 613, + 288, + 714 + ], + "type": "inline_equation", + "content": "t^{\\text{th}}" + }, + { + "bbox": [ + 46, + 613, + 288, + 714 + ], + "type": "text", + "content": " Salience DropOut iteration, " + }, + { + "bbox": [ + 46, + 613, + 288, + 714 + ], + "type": "inline_equation", + "content": "U^{(t)} = \\sum_{k=1}^{K} \\widetilde{M}^{(k,t)}" + }, + { + "bbox": [ + 46, + 613, + 288, + 714 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 613, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\widetilde{M}^{(k,t)}" + }, + { + "bbox": [ + 46, + 613, + 288, + 714 + ], + "type": "text", + "content": " is the GradCAM salience map for class " + }, + { + "bbox": [ + 46, + 613, + 288, + 714 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 613, + 288, + 714 + ], + "type": "text", + "content": " after the " + }, + { + "bbox": [ + 46, + 613, + 288, + 714 + ], + "type": "inline_equation", + "content": "t^{\\text{th}}" + }, + { + "bbox": [ + 46, + 613, + 288, + 714 + ], + "type": "text", + "content": " iteration. Next, we zero out the " + }, + { + "bbox": [ + 46, + 613, + 288, + 714 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 46, + 613, + 288, + 714 + ], + "type": "text", + "content": " of image patches with the highest values in " + }, + { + "bbox": [ + 46, + 613, + 288, + 714 + ], + "type": "inline_equation", + "content": "U^{(t)}" + }, + { + "bbox": [ + 46, + 613, + 288, + 714 + ], + "type": "text", + "content": ". On the remaining" + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 305, + 70, + 545, + 107 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 70, + 545, + 107 + ], + "spans": [ + { + "bbox": [ + 305, + 70, + 545, + 107 + ], + "type": "text", + "content": "image, we compute the GradCAM salience maps " + }, + { + "bbox": [ + 305, + 70, + 545, + 107 + ], + "type": "inline_equation", + "content": "\\widetilde{M}^{(k,t + 1)}" + }, + { + "bbox": [ + 305, + 70, + 545, + 107 + ], + "type": "text", + "content": " and their sum, " + }, + { + "bbox": [ + 305, + 70, + 545, + 107 + ], + "type": "inline_equation", + "content": "U^{(t + 1)}" + }, + { + "bbox": [ + 305, + 70, + 545, + 107 + ], + "type": "text", + "content": ". Any image patch previously zeroed out will always receive zero salience in later iterations." + } + ] + } + ], + "index": 35 + }, + { + "bbox": [ + 306, + 108, + 545, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 108, + 545, + 133 + ], + "spans": [ + { + "bbox": [ + 306, + 108, + 545, + 133 + ], + "type": "text", + "content": "Formally, the set of remaining image patches " + }, + { + "bbox": [ + 306, + 108, + 545, + 133 + ], + "type": "inline_equation", + "content": "S^{(t)} \\subseteq \\{1, \\dots, P\\}^2" + }, + { + "bbox": [ + 306, + 108, + 545, + 133 + ], + "type": "text", + "content": " after the " + }, + { + "bbox": [ + 306, + 108, + 545, + 133 + ], + "type": "inline_equation", + "content": "t^{\\text{th}}" + }, + { + "bbox": [ + 306, + 108, + 545, + 133 + ], + "type": "text", + "content": " dropout iteration is defined as" + } + ] + } + ], + "index": 36 + }, + { + "bbox": [ + 335, + 142, + 545, + 158 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 142, + 545, + 158 + ], + "spans": [ + { + "bbox": [ + 335, + 142, + 545, + 158 + ], + "type": "interline_equation", + "content": "S ^ {(t)} = S ^ {(t - 1)} \\backslash \\{(i, j) \\mid U _ {i j} ^ {(t)} \\geq \\eta \\}, \\tag {2}", + "image_path": "5b8309311840cff0a7c7646817e7e84642d6a1ab66ddd000d04ee7dd1fdfce8b.jpg" + } + ] + } + ], + "index": 37 + }, + { + "bbox": [ + 346, + 160, + 545, + 180 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 346, + 160, + 545, + 180 + ], + "spans": [ + { + "bbox": [ + 346, + 160, + 545, + 180 + ], + "type": "interline_equation", + "content": "\\eta = \\operatorname {m e d i a n} \\left(\\left\\{U _ {i j} ^ {(t)} \\mid (i, j) \\in S ^ {(t - 1)} \\right\\}\\right), \\tag {3}", + "image_path": "9801e82caecf7c26a46cf44a59e4f4dbdc92e96f7fca8f97adf37e81f05c867b.jpg" + } + ] + } + ], + "index": 38 + }, + { + "bbox": [ + 305, + 191, + 545, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 191, + 545, + 239 + ], + "spans": [ + { + "bbox": [ + 305, + 191, + 545, + 239 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 191, + 545, + 239 + ], + "type": "inline_equation", + "content": "U_{i,j}^{(t)}" + }, + { + "bbox": [ + 305, + 191, + 545, + 239 + ], + "type": "text", + "content": " denotes the aggregate salience value of the image patch at row " + }, + { + "bbox": [ + 305, + 191, + 545, + 239 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 305, + 191, + 545, + 239 + ], + "type": "text", + "content": " and column " + }, + { + "bbox": [ + 305, + 191, + 545, + 239 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 305, + 191, + 545, + 239 + ], + "type": "text", + "content": ". Additionally, note that the input to the first dropout iteration " + }, + { + "bbox": [ + 305, + 191, + 545, + 239 + ], + "type": "inline_equation", + "content": "S^0 = \\{1,\\dots ,P\\} ^2" + }, + { + "bbox": [ + 305, + 191, + 545, + 239 + ], + "type": "text", + "content": " is the set of all image patches." + } + ] + } + ], + "index": 39 + }, + { + "bbox": [ + 305, + 240, + 545, + 290 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 240, + 545, + 290 + ], + "spans": [ + { + "bbox": [ + 305, + 240, + 545, + 290 + ], + "type": "text", + "content": "We stop at four rounds of dropout, as almost all " + }, + { + "bbox": [ + 305, + 240, + 545, + 290 + ], + "type": "inline_equation", + "content": "(93.75\\%)" + }, + { + "bbox": [ + 305, + 240, + 545, + 290 + ], + "type": "text", + "content": " patches are removed beyond that point. The final output for each class " + }, + { + "bbox": [ + 305, + 240, + 545, + 290 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 305, + 240, + 545, + 290 + ], + "type": "text", + "content": " is the sum over all salience maps across the four dropout iterations, " + }, + { + "bbox": [ + 305, + 240, + 545, + 290 + ], + "type": "inline_equation", + "content": "\\hat{M}^{(k)} = \\sum_{t=1}^{4} \\widetilde{M}^{(k,t)}" + }, + { + "bbox": [ + 305, + 240, + 545, + 290 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 40 + }, + { + "bbox": [ + 306, + 297, + 471, + 308 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 297, + 471, + 308 + ], + "spans": [ + { + "bbox": [ + 306, + 297, + 471, + 308 + ], + "type": "text", + "content": "3.3. Gaussian Blur and Dense CRF" + } + ] + } + ], + "index": 41 + }, + { + "bbox": [ + 304, + 316, + 545, + 459 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 316, + 545, + 459 + ], + "spans": [ + { + "bbox": [ + 304, + 316, + 545, + 459 + ], + "type": "text", + "content": "The salience dropout procedure generates " + }, + { + "bbox": [ + 304, + 316, + 545, + 459 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 316, + 545, + 459 + ], + "type": "text", + "content": " continuous-valued salience maps, one per object class. To filter out small random noise in the salience values, we subsequent apply a straightforward thresholding operation at a predefined value " + }, + { + "bbox": [ + 304, + 316, + 545, + 459 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 304, + 316, + 545, + 459 + ], + "type": "text", + "content": " on the salience values and obtain binary segmentation masks. Nonetheless, the hard thresholding creates jagged segmentations with sharp edges that often do not coincide with object boundaries. One common strategy in zero-shot segmentation is to apply Dense Conditional Random Field (CRF) [29], which makes fine-grained adjustments to the estimated masks by enforcing consistency between nearby image pixels with similar colors." + } + ] + } + ], + "index": 42 + }, + { + "bbox": [ + 304, + 460, + 545, + 531 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 460, + 545, + 531 + ], + "spans": [ + { + "bbox": [ + 304, + 460, + 545, + 531 + ], + "type": "text", + "content": "However, we find that the hard 0/1 labels in the binary masks do not work well as pixel unary potentials for Dense CRF. Hence, we smooth them using a Gaussian kernel with a preset variance " + }, + { + "bbox": [ + 304, + 460, + 545, + 531 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 304, + 460, + 545, + 531 + ], + "type": "text", + "content": ", which results in a better initialization for unary terms and accounts for uncertainty of exact segmentation boundary along the patch boundaries." + } + ] + } + ], + "index": 43 + }, + { + "bbox": [ + 306, + 539, + 444, + 552 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 539, + 444, + 552 + ], + "spans": [ + { + "bbox": [ + 306, + 539, + 444, + 552 + ], + "type": "text", + "content": "3.4. Hyperparameter Tuning" + } + ] + } + ], + "index": 44 + }, + { + "bbox": [ + 304, + 557, + 545, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 557, + 545, + 677 + ], + "spans": [ + { + "bbox": [ + 304, + 557, + 545, + 677 + ], + "type": "text", + "content": "Three hyperparameters in PnP-OVSS have the strongest influence on the result, the cross-attention layer " + }, + { + "bbox": [ + 304, + 557, + 545, + 677 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 304, + 557, + 545, + 677 + ], + "type": "text", + "content": ", the attention head " + }, + { + "bbox": [ + 304, + 557, + 545, + 677 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 304, + 557, + 545, + 677 + ], + "type": "text", + "content": ", and the binary threshold " + }, + { + "bbox": [ + 304, + 557, + 545, + 677 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 304, + 557, + 545, + 677 + ], + "type": "text", + "content": ". Traditionally, tuning these hyperparameters requires a validation set with pixel-level labels. However, since our goal is to perform zero-shot open-vocabulary semantic segmentation, this requirement could potentially limit the applicability of the technique. Instead, we propose a weakly supervised reward function for hyperparameter tuning, which only requires a set of images and the class names appearing in each image." + } + ] + } + ], + "index": 45 + }, + { + "bbox": [ + 305, + 677, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 677, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 677, + 545, + 713 + ], + "type": "text", + "content": "The reward for an image " + }, + { + "bbox": [ + 305, + 677, + 545, + 713 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 305, + 677, + 545, + 713 + ], + "type": "text", + "content": " is calculated as follows. We start with a set of classes present in the image, denoted as " + }, + { + "bbox": [ + 305, + 677, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{K}(I)" + }, + { + "bbox": [ + 305, + 677, + 545, + 713 + ], + "type": "text", + "content": ". For each class " + }, + { + "bbox": [ + 305, + 677, + 545, + 713 + ], + "type": "inline_equation", + "content": "k\\in \\mathcal{K}(I)" + }, + { + "bbox": [ + 305, + 677, + 545, + 713 + ], + "type": "text", + "content": ", we obtain a segmentation" + } + ] + } + ], + "index": 46 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4033" + } + ] + } + ], + "index": 47 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 71, + 289, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 71, + 289, + 168 + ], + "spans": [ + { + "bbox": [ + 46, + 71, + 289, + 168 + ], + "type": "text", + "content": "mask " + }, + { + "bbox": [ + 46, + 71, + 289, + 168 + ], + "type": "inline_equation", + "content": "M^{(k)}" + }, + { + "bbox": [ + 46, + 71, + 289, + 168 + ], + "type": "text", + "content": " (which can be the GradCAM mask, the Salience DropOut mask, or the Dense CRF mask). Next, we apply the mask to the image " + }, + { + "bbox": [ + 46, + 71, + 289, + 168 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 46, + 71, + 289, + 168 + ], + "type": "text", + "content": " and input the extracted regions " + }, + { + "bbox": [ + 46, + 71, + 289, + 168 + ], + "type": "inline_equation", + "content": "M^{(k)} \\otimes I" + }, + { + "bbox": [ + 46, + 71, + 289, + 168 + ], + "type": "text", + "content": " into a pretrained neural network " + }, + { + "bbox": [ + 46, + 71, + 289, + 168 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 46, + 71, + 289, + 168 + ], + "type": "text", + "content": ", which takes an image and a textual class name as input and produces a similarity score. We calculate the normalized probability that the masked image " + }, + { + "bbox": [ + 46, + 71, + 289, + 168 + ], + "type": "inline_equation", + "content": "M^{(k)} \\otimes I" + }, + { + "bbox": [ + 46, + 71, + 289, + 168 + ], + "type": "text", + "content": " belongs to the ground-truth class " + }, + { + "bbox": [ + 46, + 71, + 289, + 168 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 71, + 289, + 168 + ], + "type": "text", + "content": " and contrast with a completely black image " + }, + { + "bbox": [ + 46, + 71, + 289, + 168 + ], + "type": "inline_equation", + "content": "\\mathbf{0}" + }, + { + "bbox": [ + 46, + 71, + 289, + 168 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 59, + 172, + 287, + 199 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 172, + 287, + 199 + ], + "spans": [ + { + "bbox": [ + 59, + 172, + 287, + 199 + ], + "type": "interline_equation", + "content": "\\operatorname {R e w a r d} = \\sum_ {k \\in \\mathcal {K} (I)} \\mathbb {1} \\left[ P r \\left(M ^ {(k)} \\otimes I, k\\right) > P r (\\mathbf {0}, k) \\right], \\tag {4}", + "image_path": "e5ee12fc2675cb7b790815e0d21390162f795d5ffa6c75aecef6e02a74376db4.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 201, + 287, + 229 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 201, + 287, + 229 + ], + "spans": [ + { + "bbox": [ + 55, + 201, + 287, + 229 + ], + "type": "interline_equation", + "content": "P r (I, k) = \\frac {\\exp (f (I , k))}{\\sum_ {k ^ {\\prime} \\in \\mathcal {K} (I)} \\exp (f (I , k ^ {\\prime}))}, \\tag {5}", + "image_path": "765fd676a97929bececc678f7b79c4f272eea17f84c5032a4207039c52bcf16b.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 233, + 287, + 293 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 233, + 287, + 293 + ], + "spans": [ + { + "bbox": [ + 46, + 233, + 287, + 293 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 233, + 287, + 293 + ], + "type": "inline_equation", + "content": "\\mathbb{1}(\\cdot)" + }, + { + "bbox": [ + 46, + 233, + 287, + 293 + ], + "type": "text", + "content": " is the indicator function. Intuitively, a reward of 1 is assigned if and only if the image features pulled with the estimated mask for class " + }, + { + "bbox": [ + 46, + 233, + 287, + 293 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 233, + 287, + 293 + ], + "type": "text", + "content": " bears higher similarity to the class name of " + }, + { + "bbox": [ + 46, + 233, + 287, + 293 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 233, + 287, + 293 + ], + "type": "text", + "content": " than a black image (which can be interpreted as the prior probability of class " + }, + { + "bbox": [ + 46, + 233, + 287, + 293 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 233, + 287, + 293 + ], + "type": "text", + "content": ")." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 293, + 287, + 353 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 293, + 287, + 353 + ], + "spans": [ + { + "bbox": [ + 46, + 293, + 287, + 353 + ], + "type": "text", + "content": "We sum up the reward for all validation images as the total reward. The best hyperparameters, including the cross-attention layer, the attention head, the threshold " + }, + { + "bbox": [ + 46, + 293, + 287, + 353 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 46, + 293, + 287, + 353 + ], + "type": "text", + "content": ", and the variance in the Gaussian blur kernel, are determined using a simple random search." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 361, + 128, + 375 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 361, + 128, + 375 + ], + "spans": [ + { + "bbox": [ + 47, + 361, + 128, + 375 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 381, + 243, + 395 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 381, + 243, + 395 + ], + "spans": [ + { + "bbox": [ + 47, + 381, + 243, + 395 + ], + "type": "text", + "content": "4.1. Datasets and Implementation Details" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 399, + 287, + 532 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 399, + 287, + 532 + ], + "spans": [ + { + "bbox": [ + 46, + 399, + 287, + 532 + ], + "type": "text", + "content": "Following the previous work for zero-shot semantic segmentation, we adopt validation sets of Pascal-VOC 2012 [16], Pascal Context [45], COCO Object [42], COCO Stuff [5], and ADE20K [88] that contain, respectively, 20 object classes, 59 object and stuff classes, 80 object classes, 171 object and stuff classes, and 150 object and stuff classes to evaluate our framework. To verify its versatility, we apply PnP-OVSSto two high-performance VLMs, BLIP [37] and BridgeTower [78], which have the ITM loss and text-to-image cross-attention. More details are in the supplementary material." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 538, + 287, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 538, + 287, + 635 + ], + "spans": [ + { + "bbox": [ + 46, + 538, + 287, + 635 + ], + "type": "text", + "content": "Hyperparameter tuning. We use CLIP VIT-L/14 to calculate the reward. The input resolution is " + }, + { + "bbox": [ + 46, + 538, + 287, + 635 + ], + "type": "inline_equation", + "content": "336 \\times 336" + }, + { + "bbox": [ + 46, + 538, + 287, + 635 + ], + "type": "text", + "content": ". The search spaces of the random search and results are shown in Tab. 2. For computation efficiency, we tune the layer, the head, and the threshold with GradCAM masks. With the first three hyperparameters fixed, we tune the Gaussian variance on masks before Dense CRF. We directly adopt Dense CRF hyperparameters from CutLER [66] without tuning." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 641, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 641, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 641, + 287, + 715 + ], + "type": "text", + "content": "Baselines. We adopt recent papers on zero-shot open-vocabulary semantic segmentation as baselines. The only baselines strictly comparable to our work are MaskClip [89] and Reco [58], which do not perform any finetuning on pretrained VLMs and do not perform hyperparameter tuning on dense annotations. We call these Group" + } + ] + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 325, + 70, + 529, + 215 + ], + "blocks": [ + { + "bbox": [ + 325, + 70, + 529, + 215 + ], + "lines": [ + { + "bbox": [ + 325, + 70, + 529, + 215 + ], + "spans": [ + { + "bbox": [ + 325, + 70, + 529, + 215 + ], + "type": "table", + "html": "
HyperparametersStartEndStepSolution
BLIP
Layer11218
Head112110
Attention Threshold0.050.50.10.15
BridgeTower
Layer1612
Head11618
Attention Threshold0.050.50.10.15
Gaussian Blur
Standard Deviation0.010.110.020.05
", + "image_path": "8b50861a9067b7f49de6e8b30457b616c45d27227634fbdbf78e91a159253ac1.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 348, + 222, + 504, + 234 + ], + "lines": [ + { + "bbox": [ + 348, + 222, + 504, + 234 + ], + "spans": [ + { + "bbox": [ + 348, + 222, + 504, + 234 + ], + "type": "text", + "content": "Table 2. Search space for hyperparameters" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 255, + 545, + 387 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 255, + 545, + 387 + ], + "spans": [ + { + "bbox": [ + 304, + 255, + 545, + 387 + ], + "type": "text", + "content": "3. To further expand our scope, we also include two other groups of baselines. Group 2 finetunes VLMs but does not require image-text pairs, including MaskClip with self-training (ST) [89] and ZeroSeg [8]. We include ZeroSeg, trained with ImageNet1K, and ZeroSeg*, trained with CC3M+COCO. Group 1 contains baselines that require training on image-text pairs. For completeness, we also compare against supervised techniques since 2019. To maintain the zero-shot setting, we test them on classes not observed during training. For more details, we refer readers to the supplementary material." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 395, + 391, + 407 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 395, + 391, + 407 + ], + "spans": [ + { + "bbox": [ + 306, + 395, + 391, + 407 + ], + "type": "text", + "content": "4.2. Main Results" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 413, + 545, + 474 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 413, + 545, + 474 + ], + "spans": [ + { + "bbox": [ + 304, + 413, + 545, + 474 + ], + "type": "text", + "content": "We show the main results in Tab. 3. As input resolution may influence results and cause unfair comparisons, we label the resolution used by each method. PACL [46] uses an 224 resolution but changes the stride for image patchification from 16 to 4, hence introducing overlapping patches." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 474, + 545, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 474, + 545, + 605 + ], + "spans": [ + { + "bbox": [ + 304, + 474, + 545, + 605 + ], + "type": "text", + "content": "PnP-OVSS exhibits excellent performance. Comparing with MaskClip [89] and Reco [58], the two methods that require no additional training and ground truth for hyperparameter tuning, on an equal-resolution basis, we attain " + }, + { + "bbox": [ + 304, + 474, + 545, + 605 + ], + "type": "inline_equation", + "content": "+29.4\\%" + }, + { + "bbox": [ + 304, + 474, + 545, + 605 + ], + "type": "text", + "content": " mIoU on Pascal VOC, " + }, + { + "bbox": [ + 304, + 474, + 545, + 605 + ], + "type": "inline_equation", + "content": "+13.2\\%" + }, + { + "bbox": [ + 304, + 474, + 545, + 605 + ], + "type": "text", + "content": " mIoU on Pascal Context, " + }, + { + "bbox": [ + 304, + 474, + 545, + 605 + ], + "type": "inline_equation", + "content": "+14.0\\%" + }, + { + "bbox": [ + 304, + 474, + 545, + 605 + ], + "type": "text", + "content": " mIoU on COCO Object, and " + }, + { + "bbox": [ + 304, + 474, + 545, + 605 + ], + "type": "inline_equation", + "content": "+11.4\\%" + }, + { + "bbox": [ + 304, + 474, + 545, + 605 + ], + "type": "text", + "content": " mIoU on ADE-20K. Further, PnP-OVSS surpasses all baselines in Group 2. On an equal-resolution basis, we achieve " + }, + { + "bbox": [ + 304, + 474, + 545, + 605 + ], + "type": "inline_equation", + "content": "+13.7\\%" + }, + { + "bbox": [ + 304, + 474, + 545, + 605 + ], + "type": "text", + "content": " mIoU on Pascal Voc, " + }, + { + "bbox": [ + 304, + 474, + 545, + 605 + ], + "type": "inline_equation", + "content": "+9.6\\%" + }, + { + "bbox": [ + 304, + 474, + 545, + 605 + ], + "type": "text", + "content": " on Pascal Context, " + }, + { + "bbox": [ + 304, + 474, + 545, + 605 + ], + "type": "inline_equation", + "content": "+9.5\\%" + }, + { + "bbox": [ + 304, + 474, + 545, + 605 + ], + "type": "text", + "content": " on COCO Object, " + }, + { + "bbox": [ + 304, + 474, + 545, + 605 + ], + "type": "inline_equation", + "content": "+11.6\\%" + }, + { + "bbox": [ + 304, + 474, + 545, + 605 + ], + "type": "text", + "content": " on COCO Stuff, and " + }, + { + "bbox": [ + 304, + 474, + 545, + 605 + ], + "type": "inline_equation", + "content": "+12.8\\%" + }, + { + "bbox": [ + 304, + 474, + 545, + 605 + ], + "type": "text", + "content": " on ADE-20K." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 605, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 545, + 714 + ], + "type": "text", + "content": "When compared to Group 1, PnP-OVSS still outperforms most. On the Pascal datasets, under equal resolutions, PnP-OVSS + BLIPFlickr outperforms 6 out of 10 baselines on Pascal VOC, and 8 out of 9 baselines on Pascal Context. For the COCO datasets, under equal resolutions, PnP-OVSS + BLIPFlickr beats 5 out of 8 baselines on COCO Object, and all baselines except PACL [46] on COCO Stuff. The two Pascal datasets share many images, and the two COCO datasets use exactly the same images, but they have" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4034" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 52, + 69, + 542, + 469 + ], + "blocks": [ + { + "bbox": [ + 52, + 69, + 542, + 469 + ], + "lines": [ + { + "bbox": [ + 52, + 69, + 542, + 469 + ], + "spans": [ + { + "bbox": [ + 52, + 69, + 542, + 469 + ], + "type": "table", + "html": "
MethodFinetuning VLMsHT on Dense LabelsShort-side ResolutionPascal VOC-20Pascal Context-59COCO Object-80COCO Stuff-171ADE 20K-150
Group 1: Methods that require weakly supervised finetuning on image-text data
ViL-Seg†[43]✓✓-37.318.9-18.0-
CLIPpy [51]✓✓22452.2-32.025.5*13.5
SegClip [44]✓✓22452.624.726.5--
GroupVit (by [51])✓✓22428.114.812.9-6.2
GroupVit (by [6])✓✓44850.418.727.515.39.2
GroupVit [73]✓✓44852.322.424.3--
ViewCo [52]✓✓44852.423.023.5--
OVSegmentor [74]✓✓44853.820.425.1--
TCL [6] +PAMR [2]✓✓44855.030.431.622.417.1
PACL [46]✓✓224×472.350.1-38.831.4
Group 2: Methods that require finetuning but not real image-text data
MaskClip w/ ST [89]✓✓336-31.1-18.0-
MaskClip w/ ST (by [6])✓✓44838.823.620.616.49.8
ZeroSeg [8]✓✓44840.820.420.2--
Group 3: Methods that require no finetuning
MaskClip (by [51])××22422.1-13.88.16.8
MaskClip [89]××336-25.5-14.6-
Reco [58]××320-27.2-27.2-
Reco (by [6])××44825.119.915.714.811.2
PnP-OVSS with different VLMs
BLIPFlickr××22447.836.424.825.818.2
BLIPFlickr××32051.740.428.029.621.3
BLIPFlickr××33652.540.728.229.621.9
BLIPFlickr××44854.542.229.731.522.6
BLIPFlickr××76854.142.831.832.523.5
BLIPCOCO××76855.741.933.832.623.2
BridgeTower××32236.432.324.227.618.6
BridgeTower××33635.332.424.227.618.0
BridgeTower××77035.232.424.528.019.0
", + "image_path": "55f789a89ab373b01f375bad2003af82fa7db4d7763d03bed43635ff49253f50.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 119, + 542, + 476, + 673 + ], + "blocks": [ + { + "bbox": [ + 46, + 477, + 547, + 534 + ], + "lines": [ + { + "bbox": [ + 46, + 477, + 547, + 534 + ], + "spans": [ + { + "bbox": [ + 46, + 477, + 547, + 534 + ], + "type": "text", + "content": "Table 3. Zero-shot semantic segmentation performance in mIoU. Group 3 contains the most similar baselines that serve as fair comparisons to PnP-OVSS. Groups 1 and 2 benefit from additional training, extra image-text data, and hyperparameter tuning on dense labels. We use the word \"by\" followed by a paper citation to indicate results of the same technique reported by different papers. " + }, + { + "bbox": [ + 46, + 477, + 547, + 534 + ], + "type": "inline_equation", + "content": "\\star" + }, + { + "bbox": [ + 46, + 477, + 547, + 534 + ], + "type": "text", + "content": " CLIPpy tests on 133 categories of COCO Stuff while we test all 171 classes of COCO Stuff. ViL-Seg† is tested on subset of classes on the three datasets, as detailed in the supplementary." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 119, + 542, + 476, + 673 + ], + "lines": [ + { + "bbox": [ + 119, + 542, + 476, + 673 + ], + "spans": [ + { + "bbox": [ + 119, + 542, + 476, + 673 + ], + "type": "table", + "html": "
MethodDense \nLabelsHT on Dense \nLabelsPascal \nVOC-20Pascal \nContext-59COCO \nStuff-171
PnP-OVSS + BLIPFlickr (Ours)××53.653.839.8
SPNet+ST [71]✓✓25.8-26.9
ZS3Net+ST [4]✓✓21.220.710.6
CaGNet+ST [18]✓✓30.3-13.4
STRICT [47]✓✓35.6-30.3
LSeg [32]✓✓41.0--
SimBase [76]✓✓72.5-36.3
MaskCLIP+ w/ ST [89]✓✓86.166.754.7
", + "image_path": "720db4477a50564d77d47e1dd7384735bf23d99d4612b0de3ec53df734d61d27.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 681, + 543, + 693 + ], + "lines": [ + { + "bbox": [ + 48, + 681, + 543, + 693 + ], + "spans": [ + { + "bbox": [ + 48, + 681, + 543, + 693 + ], + "type": "text", + "content": "Table 4. Comparison of zero-shot semantic segmentation performance on unseen categories with methods trained with dense annotation." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4035" + } + ] + } + ], + "index": 4 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 66, + 70, + 269, + 204 + ], + "blocks": [ + { + "bbox": [ + 66, + 70, + 269, + 204 + ], + "lines": [ + { + "bbox": [ + 66, + 70, + 269, + 204 + ], + "spans": [ + { + "bbox": [ + 66, + 70, + 269, + 204 + ], + "type": "table", + "html": "
Ablated ModelPascal ContextCOCO Stuff
BLIPFlickr19.814.5
+ GradCam21.617.5
+ Drop 125.119.8
+ Drop 226.520.6
+ Drop 327.020.9
+ Drop 427.220.9
+ Drop 4 + Blur36.828.6
+ Drop 4 + Dense CRF35.331.8
+ Drop 4 + Blur + Dense CRF42.832.5
", + "image_path": "202c3313600172d50dec906994c14ece22a2b25bc916fbf50f6c3b84c4c57976.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 257, + 287, + 304 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 257, + 287, + 304 + ], + "spans": [ + { + "bbox": [ + 46, + 257, + 287, + 304 + ], + "type": "text", + "content": "vastly different lists of classes to segment. We observe that the advantage of PnP-OVSS over methods in Group 1 becomes more pronounced as the number of classes on the same image increases." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 304, + 287, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 304, + 287, + 376 + ], + "spans": [ + { + "bbox": [ + 46, + 304, + 287, + 376 + ], + "type": "text", + "content": "When applied to BridgeTower [78], PnP-OVSS still surpasses all methods in Group 3 by " + }, + { + "bbox": [ + 46, + 304, + 287, + 376 + ], + "type": "inline_equation", + "content": "10.2\\%" + }, + { + "bbox": [ + 46, + 304, + 287, + 376 + ], + "type": "text", + "content": " on Pascal VOC, " + }, + { + "bbox": [ + 46, + 304, + 287, + 376 + ], + "type": "inline_equation", + "content": "5.1\\%" + }, + { + "bbox": [ + 46, + 304, + 287, + 376 + ], + "type": "text", + "content": " on Pascal Context, " + }, + { + "bbox": [ + 46, + 304, + 287, + 376 + ], + "type": "inline_equation", + "content": "8.5\\%" + }, + { + "bbox": [ + 46, + 304, + 287, + 376 + ], + "type": "text", + "content": " on COCO Object, " + }, + { + "bbox": [ + 46, + 304, + 287, + 376 + ], + "type": "inline_equation", + "content": "0.4\\%" + }, + { + "bbox": [ + 46, + 304, + 287, + 376 + ], + "type": "text", + "content": " on COCO Stuff, and " + }, + { + "bbox": [ + 46, + 304, + 287, + 376 + ], + "type": "inline_equation", + "content": "6.8\\%" + }, + { + "bbox": [ + 46, + 304, + 287, + 376 + ], + "type": "text", + "content": " on ADE-20K. This showcases the plug-and-play ability of PnP-OVSS, which excels with different base networks." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 377, + 287, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 377, + 287, + 450 + ], + "spans": [ + { + "bbox": [ + 46, + 377, + 287, + 450 + ], + "type": "text", + "content": "We report comparisons against supervised methods in Tab 4. PnP-OVSS + BLIPFlickr outperforms 5 out of 7 methods on Pascal VOC, as well as every baseline except MaskCLIP+ on Pascal Context and COCO stuff. As these baselines benefit from dense supervision, the results further demonstrate the strengths of PnP-OVSS." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 459, + 141, + 472 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 459, + 141, + 472 + ], + "spans": [ + { + "bbox": [ + 47, + 459, + 141, + 472 + ], + "type": "text", + "content": "4.3. Ablation Study" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 478, + 287, + 646 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 478, + 287, + 646 + ], + "spans": [ + { + "bbox": [ + 46, + 478, + 287, + 646 + ], + "type": "text", + "content": "We perform gradual ablation of the components of " + }, + { + "bbox": [ + 46, + 478, + 287, + 646 + ], + "type": "inline_equation", + "content": "\\mathrm{PnP-OVSS}" + }, + { + "bbox": [ + 46, + 478, + 287, + 646 + ], + "type": "text", + "content": " on " + }, + { + "bbox": [ + 46, + 478, + 287, + 646 + ], + "type": "inline_equation", + "content": "\\mathrm{BLIP}_{\\mathrm{Flickr}}" + }, + { + "bbox": [ + 46, + 478, + 287, + 646 + ], + "type": "text", + "content": " and report the results in Tab. 5. Each component, including GradCAM, all Salience DropOut iterations, Gaussian blur, and Dense CRF, contribute positively to the final performance. In particular, the first iteration of Salience DropOut has much larger impact (+3.5/2.3) than the second iteration (+1.4/0.8), which in turn is more important than the rest. Interestingly, Gaussian blur by itself attains good performance (+9.6/7.7) whereas Dense CRF only works well when combined with blur. Dense CRF alone is worse than Gaussian blur by 1.5 mIoU on Pascal Context. This is likely caused by the fact that hard 0/1 labels resulted from thresholding are not informative unary potentials that can be leveraged by CRF effectively." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 654, + 201, + 668 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 654, + 201, + 668 + ], + "spans": [ + { + "bbox": [ + 47, + 654, + 201, + 668 + ], + "type": "text", + "content": "4.4. Hyperparameter Sensitivity" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 677, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 287, + 715 + ], + "type": "text", + "content": "The choice of hyperparameters often exerts substantial influence on segmentation performance. Here we quantitatively examine how the choice of cross-attention layers" + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 312, + 70, + 541, + 204 + ], + "blocks": [ + { + "bbox": [ + 47, + 212, + 287, + 235 + ], + "lines": [ + { + "bbox": [ + 47, + 212, + 287, + 235 + ], + "spans": [ + { + "bbox": [ + 47, + 212, + 287, + 235 + ], + "type": "text", + "content": "Table 5. An ablation study of PnP-OVSS + BLIPFlickr with resolution 768 on Pascal Context and COCO Stuff." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 312, + 70, + 541, + 204 + ], + "lines": [ + { + "bbox": [ + 312, + 70, + 541, + 204 + ], + "spans": [ + { + "bbox": [ + 312, + 70, + 541, + 204 + ], + "type": "table", + "html": "
Layer Mean123456
mIoU11.612.311.311.011.612.9
Layer Mean789101112
mIoU13.725.325.823.412.55.8
Head in Layer 8123456
mIoU16.63.88.620.019.85.2
Head in Layer 8789101112
mIoU11.16.03.829.617.38.0
", + "image_path": "59e3c2a8cb356f32db262a9e7028beca7a840fbb873e7f211f09427967049036.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 212, + 545, + 256 + ], + "lines": [ + { + "bbox": [ + 305, + 212, + 545, + 256 + ], + "spans": [ + { + "bbox": [ + 305, + 212, + 545, + 256 + ], + "type": "text", + "content": "Table 6. Semantic segmentation performance using cross-attention maps averaged across all heads in a layer and separate heads in Layer 8. Results are attained with PnP-OVSS+BLIPFlickr on COCO Stuff and resolution 336." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 281, + 545, + 329 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 281, + 545, + 329 + ], + "spans": [ + { + "bbox": [ + 304, + 281, + 545, + 329 + ], + "type": "text", + "content": "and attention heads may change the segmentation mIoU on COCO Stuff. Tab. 6 shows the results obtained from the average cross-attention maps over all heads in each layer and those from different attention heads." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 331, + 546, + 439 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 331, + 546, + 439 + ], + "spans": [ + { + "bbox": [ + 304, + 331, + 546, + 439 + ], + "type": "text", + "content": "We make the following observations. First, different layers and heads have drastic performance differences. The best-to-worst difference among all layers is 20, and that among heads in Layer 8 is 23.8, underscoring the importance of hyperparameter tuning. Second, the random search using the proposed reward function correctly identifies the best layer-head combination, even though Layer 8 is not the best layer based on average head performance. This indicates the effectiveness of our method." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 456, + 384, + 469 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 456, + 384, + 469 + ], + "spans": [ + { + "bbox": [ + 306, + 456, + 384, + 469 + ], + "type": "text", + "content": "5. Conclusions" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 482, + 545, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 482, + 545, + 578 + ], + "spans": [ + { + "bbox": [ + 304, + 482, + 545, + 578 + ], + "type": "text", + "content": "We propose PnP-OVSS, which extracts the ability of semantic segmentation from opaque VLMs. PnP-OVSS is simple to use, requires no extra finetuning, and delivers high performance, exceeding not only all baselines that requires no finetuning, but also all baselines that do not use image-text pairs in finetuning. Its success hints at a new direction for open-vocabulary segmentation tasks leveraging large VLMs." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 595, + 416, + 609 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 595, + 416, + 609 + ], + "spans": [ + { + "bbox": [ + 306, + 595, + 416, + 609 + ], + "type": "text", + "content": "6. Acknowledgments" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 617, + 545, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 617, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 617, + 545, + 715 + ], + "type": "text", + "content": "This work has been supported by the Nanyang Associate Professorship and the National Research Foundation Fellowship (NRF- NRFF13-2021-0006), Singapore, and by the Vector Institute for AI, Canada CIFAR AI Chair, and NSERC Discovery Grant. Any opinions, findings, conclusions, or recommendations expressed in this material are those of the authors and do not reflect the views of the funding agencies." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "4036" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 210 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 210 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 210 + ], + "type": "text", + "content": "[1] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, Roman Ring, Eliza Rutherford, Serkan Cabi, Tengda Han, Zhitao Gong, Sina Samangooei, Marianne Monteiro, Jacob L Menick, Sebastian Borgeaud, Andy Brock, Aida Nematzadeh, Sahand Sharifzadeh, Mikol aj Binkowski, Ricardo Barreira, Oriol Vinyals, Andrew Zisserman, and Karén Simonyan. Flamingo: a visual language model for few-shot learning. In Advances in Neural Information Processing Systems, 2022. 1, 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 213, + 288, + 255 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 213, + 288, + 255 + ], + "spans": [ + { + "bbox": [ + 53, + 213, + 288, + 255 + ], + "type": "text", + "content": "[2] Nikita Araslanov and Stefan Roth. Single-stage semantic segmentation from image labels. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4253-4262, 2020. 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 258, + 287, + 311 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 258, + 287, + 311 + ], + "spans": [ + { + "bbox": [ + 53, + 258, + 287, + 311 + ], + "type": "text", + "content": "[3] Loris Bazzani, Alessandra Bergamo, Dragomir Anguelov, and Lorenzo Torresani. Self-taught object localization with deep networks. In 2016 IEEE winter conference on applications of computer vision (WACV), pages 1-9. IEEE, 2016. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 313, + 287, + 345 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 313, + 287, + 345 + ], + "spans": [ + { + "bbox": [ + 53, + 313, + 287, + 345 + ], + "type": "text", + "content": "[4] Maxime Bucher, Tuan-Hung Vu, Matthieu Cord, and Patrick Pérez. Zero-shot semantic segmentation. Advances in Neural Information Processing Systems, 32, 2019. 3, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 347, + 288, + 389 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 347, + 288, + 389 + ], + "spans": [ + { + "bbox": [ + 53, + 347, + 288, + 389 + ], + "type": "text", + "content": "[5] Holger Caesar, Jasper Uijlings, and Vittorio Ferrari. Cocosstuff: Thing and stuff classes in context. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1209-1218, 2018. 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 392, + 288, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 392, + 288, + 445 + ], + "spans": [ + { + "bbox": [ + 53, + 392, + 288, + 445 + ], + "type": "text", + "content": "[6] Junbum Cha, Jonghwan Mun, and Byungseok Roh. Learning to generate text-grounded mask for open-world semantic segmentation from only image-text pairs. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11165-11174, 2023. 3, 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 447, + 288, + 501 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 447, + 288, + 501 + ], + "spans": [ + { + "bbox": [ + 53, + 447, + 288, + 501 + ], + "type": "text", + "content": "[7] Jun Chen, Han Guo, Kai Yi, Boyang Li, and Mohamed Elhoseiny. Visualgpt: Data-efficient adaptation of pretrained language models for image captioning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18030-18040, 2022. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 502, + 287, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 502, + 287, + 578 + ], + "spans": [ + { + "bbox": [ + 53, + 502, + 287, + 578 + ], + "type": "text", + "content": "[8] Jun Chen, Deyao Zhu, Guocheng Qian, Bernard Ghanem, Zhicheng Yan, Chenchen Zhu, Fanyi Xiao, Sean Chang Culatana, and Mohamed Elhoseiny. Exploring open-vocabulary semantic segmentation from clip vision encoder distillation only. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 699-710, 2023. 1, 3, 6, 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 580, + 287, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 580, + 287, + 634 + ], + "spans": [ + { + "bbox": [ + 53, + 580, + 287, + 634 + ], + "type": "text", + "content": "[9] Qi Chen, Lingxiao Yang, Jian-Huang Lai, and Xiaohua Xie. Self-supervised image-specific prototype exploration for weakly supervised semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4288-4298, 2022. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 636, + 287, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 636, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 636, + 287, + 689 + ], + "type": "text", + "content": "[10] Yen-Chun Chen, Linjie Li, Licheng Yu, Ahmed El Kholy, Faisal Ahmed, Zhe Gan, Yu Cheng, and Jingjing Liu. Uniter: Universal image-text representation learning. In Proceedings of the 16th European Conference on Computer Vision, 2020. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "type": "text", + "content": "[11] Zhaozheng Chen, Tan Wang, Xiongwei Wu, Xian-Sheng Hua, Hanwang Zhang, and Qianru Sun. Class re-activation" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 74, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 326, + 74, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 74, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 326, + 74, + 545, + 106 + ], + "type": "text", + "content": "maps for weakly-supervised semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 969-978, 2022. 1" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 308, + 107, + 545, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 107, + 545, + 161 + ], + "spans": [ + { + "bbox": [ + 308, + 107, + 545, + 161 + ], + "type": "text", + "content": "[12] Minsu Cho, Suha Kwak, Cordelia Schmid, and Jean Ponce. Unsupervised object discovery and localization in the wild: Part-based matching with bottom-up region proposals. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1201-1210, 2015. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 308, + 162, + 545, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 162, + 545, + 205 + ], + "spans": [ + { + "bbox": [ + 308, + 162, + 545, + 205 + ], + "type": "text", + "content": "[13] Ramazan Gokberk Cinbis, Jakob Verbeek, and Cordelia Schmid. Weakly supervised object localization with multifold multiple instance learning. IEEE transactions on pattern analysis and machine intelligence, 39(1):189-203, 2016. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 206, + 545, + 238 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 206, + 545, + 238 + ], + "spans": [ + { + "bbox": [ + 308, + 206, + 545, + 238 + ], + "type": "text", + "content": "[14] Timothée Darcet, Maxime Oquab, Julien Mairal, and Piotr Bojanowski. Vision transformers need registers. arXiv preprint arXiv:2309.16588, 2023. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 239, + 545, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 239, + 545, + 293 + ], + "spans": [ + { + "bbox": [ + 308, + 239, + 545, + 293 + ], + "type": "text", + "content": "[15] Ye Du, Zehua Fu, Qingjie Liu, and Yunhong Wang. Weakly supervised semantic segmentation by pixel-to-prototype contrast. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4320–4329, 2022. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 294, + 545, + 337 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 294, + 545, + 337 + ], + "spans": [ + { + "bbox": [ + 308, + 294, + 545, + 337 + ], + "type": "text", + "content": "[16] Mark Everingham, Luc Van Gool, Christopher KI Williams, John Winn, and Andrew Zisserman. The pascal visual object classes (voc) challenge. International journal of computer vision, 88:303-338, 2010. 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 338, + 545, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 338, + 545, + 381 + ], + "spans": [ + { + "bbox": [ + 308, + 338, + 545, + 381 + ], + "type": "text", + "content": "[17] Golnaz Ghiasi, Xiuye Gu, Yin Cui, and Tsung-Yi Lin. Scaling open-vocabulary image segmentation with image-level labels. In European Conference on Computer Vision, pages 540-557. Springer, 2022. 1, 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 383, + 545, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 383, + 545, + 435 + ], + "spans": [ + { + "bbox": [ + 308, + 383, + 545, + 435 + ], + "type": "text", + "content": "[18] Zhangxuan Gu, Siyuan Zhou, Li Niu, Zihan Zhao, and Liqing Zhang. Context-aware feature generation for zero-shot semantic segmentation. In Proceedings of the 28th ACM International Conference on Multimedia, pages 1921-1929, 2020. 3, 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 437, + 545, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 437, + 545, + 502 + ], + "spans": [ + { + "bbox": [ + 308, + 437, + 545, + 502 + ], + "type": "text", + "content": "[19] Jiaxian Guo, Junnan Li, Dongxu Li, Anthony Meng Huat Tiong, Boyang Li, Dacheng Tao, and Steven Hoi. From images to textual prompts: Zero-shot visual question answering with frozen large language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10867-10877, 2023. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 504, + 545, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 504, + 545, + 547 + ], + "spans": [ + { + "bbox": [ + 308, + 504, + 545, + 547 + ], + "type": "text", + "content": "[20] Yanming Guo, Yu Liu, Theodoros Georgiou, and Michael S Lew. A review of semantic segmentation using deep neural networks. International journal of multimedia information retrieval, 7:87-93, 2018. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 548, + 545, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 548, + 545, + 613 + ], + "spans": [ + { + "bbox": [ + 308, + 548, + 545, + 613 + ], + "type": "text", + "content": "[21] Kunyang Han, Yong Liu, Jun Hao Liew, Henghui Ding, Jiajun Liu, Yitong Wang, Yansong Tang, Yujiu Yang, Jiashi Feng, Yao Zhao, et al. Global knowledge calibration for fast open-vocabulary segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 797-807, 2023. 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 614, + 545, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 614, + 545, + 645 + ], + "spans": [ + { + "bbox": [ + 308, + 614, + 545, + 645 + ], + "type": "text", + "content": "[22] Shijie Hao, Yuan Zhou, and Yanrong Guo. A brief survey on semantic segmentation with deep learning. Neurocomputing, 406:302-321, 2020. 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 647, + 545, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 647, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 308, + 647, + 545, + 689 + ], + "type": "text", + "content": "[23] Qibin Hou, PengTao Jiang, Yunchao Wei, and Ming-Ming Cheng. Self-erasing network for integral object attention. Advances in Neural Information Processing Systems, 31, 2018. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 691, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 691, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 691, + 545, + 713 + ], + "type": "text", + "content": "[24] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4037" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 73, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 67, + 73, + 287, + 115 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 73, + 287, + 115 + ], + "spans": [ + { + "bbox": [ + 67, + 73, + 287, + 115 + ], + "type": "text", + "content": "Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In Proceedings of the 38th International Conference on Machine Learning, 2021. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 118, + 287, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 118, + 287, + 172 + ], + "spans": [ + { + "bbox": [ + 48, + 118, + 287, + 172 + ], + "type": "text", + "content": "[25] Peng-Tao Jiang, Yuqi Yang, Qibin Hou, and Yunchao Wei. L2g: A simple local-to-global knowledge transfer framework for weakly supervised semantic segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16886-16896, 2022. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 173, + 287, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 173, + 287, + 215 + ], + "spans": [ + { + "bbox": [ + 49, + 173, + 287, + 215 + ], + "type": "text", + "content": "[26] Tsung-Wei Ke, Jyh-Jing Hwang, and Stella X Yu. Universal weakly supervised segmentation by pixel-to-segment contrastive learning. arXiv preprint arXiv:2105.00957, 2021. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 217, + 287, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 217, + 287, + 259 + ], + "spans": [ + { + "bbox": [ + 49, + 217, + 287, + 259 + ], + "type": "text", + "content": "[27] Wonjae Kim, Bokyung Son, and Ildoo Kim. Vilt: Vision-and-language transformer without convolution or region supervision. In Proceedings of the 38th International Conference on Machine Learning, 2021. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 49, + 261, + 287, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 261, + 287, + 326 + ], + "spans": [ + { + "bbox": [ + 49, + 261, + 287, + 326 + ], + "type": "text", + "content": "[28] Alexander Kolesnikov and Christoph H Lampert. Seed, expand and constrain: Three principles for weakly-supervised image segmentation. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part IV 14, pages 695-711. Springer, 2016. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 49, + 327, + 287, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 327, + 287, + 369 + ], + "spans": [ + { + "bbox": [ + 49, + 327, + 287, + 369 + ], + "type": "text", + "content": "[29] Philipp Krahenbuhl and Vladlen Koltun. Efficient inference in fully connected crfs with gaussian edge potentials. Advances in neural information processing systems, 24, 2011. 3, 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 371, + 287, + 425 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 371, + 287, + 425 + ], + "spans": [ + { + "bbox": [ + 48, + 371, + 287, + 425 + ], + "type": "text", + "content": "[30] Krishna Kumar Singh and Yong Jae Lee. Hide-and-seek: Forcing a network to be meticulous for weakly-supervised object and action localization. In Proceedings of the IEEE International Conference on Computer Vision, pages 3524-3533, 2017. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 426, + 287, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 426, + 287, + 480 + ], + "spans": [ + { + "bbox": [ + 48, + 426, + 287, + 480 + ], + "type": "text", + "content": "[31] Jungbeom Lee, Seong Joon Oh, Sangdoo Yun, Junsuk Choe, Eunji Kim, and Sungroh Yoon. Weakly supervised semantic segmentation using out-of-distribution data. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16897-16906, 2022. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 481, + 287, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 481, + 287, + 514 + ], + "spans": [ + { + "bbox": [ + 48, + 481, + 287, + 514 + ], + "type": "text", + "content": "[32] Boyi Li, Kilian Q. Weinberger, Serge Belongie, Vladlen Koltun, and René Ranftl. Language-driven semantic segmentation, 2022. 3, 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 514, + 287, + 591 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 514, + 287, + 591 + ], + "spans": [ + { + "bbox": [ + 48, + 514, + 287, + 591 + ], + "type": "text", + "content": "[33] Chenliang Li, Haiyang Xu, Junfeng Tian, Wei Wang, Ming Yan, Bin Bi, Jiabo Ye, He Chen, Guohai Xu, Zheng Cao, Ji Zhang, Songfang Huang, Fei Huang, Jingren Zhou, and Luo Si. mPLUG: Effective and efficient vision-language learning by cross-modal skip-connections. In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, 2022. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 592, + 287, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 592, + 287, + 645 + ], + "spans": [ + { + "bbox": [ + 48, + 592, + 287, + 645 + ], + "type": "text", + "content": "[34] Chenliang Li, Haiyang Xu, Junfeng Tian, Wei Wang, Ming Yan, Bin Bi, Jiabo Ye, Hehong Chen, Guohai Xu, Zheng Cao, et al. mplug: Effective and efficient vision-language learning by cross-modal skip-connections. arXiv preprint arXiv:2205.12005, 2022. 2, 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 647, + 287, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 647, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 647, + 287, + 689 + ], + "type": "text", + "content": "[35] Junnan Li, Ramprasaath R. Selvaraju, Akhilesh Deepak Gotmare, Shafiq Joty, Caiming Xiong, and Steven Hoi. Align before fuse: Vision and language representation learning with momentum distillation. In NeurIPS, 2021. 2, 3, 4" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "type": "text", + "content": "[36] Jinlong Li, Zequn Jie, Xu Wang, Xiaolin Wei, and Lin Ma. Expansion and shrinkage of localization for weakly-" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 25, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "text", + "content": "supervised semantic segmentation. Advances in Neural Information Processing Systems, 35:16037-16051, 2022. 1" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 308, + 97, + 545, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 97, + 545, + 150 + ], + "spans": [ + { + "bbox": [ + 308, + 97, + 545, + 150 + ], + "type": "text", + "content": "[37] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In International Conference on Machine Learning, pages 12888-12900. PMLR, 2022. 1, 2, 3, 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 152, + 545, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 152, + 545, + 195 + ], + "spans": [ + { + "bbox": [ + 308, + 152, + 545, + 195 + ], + "type": "text", + "content": "[38] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In ICML, 2023. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 198, + 545, + 275 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 198, + 545, + 275 + ], + "spans": [ + { + "bbox": [ + 308, + 198, + 545, + 275 + ], + "type": "text", + "content": "[39] Wei Li, Can Gao, Guocheng Niu, Xinyan Xiao, Hao Liu, Jiachen Liu, Hua Wu, and Haifeng Wang. UNIMO: Towards unified-modal understanding and generation via cross-modal contrastive learning. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), 2021. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 277, + 545, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 277, + 545, + 330 + ], + "spans": [ + { + "bbox": [ + 308, + 277, + 545, + 330 + ], + "type": "text", + "content": "[40] Xiujun Li, Xi Yin, Chunyuan Li, Pengchuan Zhang, Xiaowei Hu, Lei Zhang, Lijuan Wang, Houdong Hu, Li Dong, Furu Wei, Yejin Choi, and Jianfeng Gao. Oscar: Object-semantics aligned pre-training for vision-language tasks. In ECCV 2020, 2020. 1, 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 333, + 545, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 333, + 545, + 397 + ], + "spans": [ + { + "bbox": [ + 308, + 333, + 545, + 397 + ], + "type": "text", + "content": "[41] Feng Liang, Bichen Wu, Xiaoliang Dai, Kunpeng Li, Yinan Zhao, Hang Zhang, Peizhao Zhang, Peter Vajda, and Diana Marculescu. Open-vocabulary semantic segmentation with mask-adapted clip. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7061-7070, 2023. 1, 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 400, + 545, + 466 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 400, + 545, + 466 + ], + "spans": [ + { + "bbox": [ + 308, + 400, + 545, + 466 + ], + "type": "text", + "content": "[42] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 468, + 545, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 468, + 545, + 521 + ], + "spans": [ + { + "bbox": [ + 308, + 468, + 545, + 521 + ], + "type": "text", + "content": "[43] Quande Liu, Youpeng Wen, Jianhua Han, Chunjing Xu, Hang Xu, and Xiaodan Liang. Open-world semantic segmentation via contrasting and clustering vision-language embedding. In European Conference on Computer Vision, pages 275-292. Springer, 2022. 2, 3, 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 524, + 545, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 524, + 545, + 578 + ], + "spans": [ + { + "bbox": [ + 308, + 524, + 545, + 578 + ], + "type": "text", + "content": "[44] Huaishao Luo, Junwei Bao, Youzheng Wu, Xiaodong He, and Tianrui Li. Segclip: Patch aggregation with learnable centers for open-vocabulary semantic segmentation. In International Conference on Machine Learning, pages 23033-23044. PMLR, 2023. 1, 2, 3, 7" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 580, + 545, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 580, + 545, + 645 + ], + "spans": [ + { + "bbox": [ + 308, + 580, + 545, + 645 + ], + "type": "text", + "content": "[45] Roozbeh Mottaghi, Xianjie Chen, Xiaobai Liu, Nam-Gyu Cho, Seong-Whan Lee, Sanja Fidler, Raquel Urtasun, and Alan Yuille. The role of context for object detection and semantic segmentation in the wild. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 891–898, 2014. 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 647, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 647, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 308, + 647, + 545, + 712 + ], + "type": "text", + "content": "[46] Jishnu Mukhoti, Tsung-Yu Lin, Omid Poursaeed, Rui Wang, Ashish Shah, Philip HS Torr, and Ser-Nam Lim. Open vocabulary semantic segmentation with patch aligned contrastive learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19413-19423, 2023. 1, 3, 6, 7" + } + ] + } + ], + "index": 24 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "4038" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 137 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 137 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 137 + ], + "type": "text", + "content": "[47] Giuseppe Pastore, Fabio Cermelli, Yongqin Xian, Massimiliano Mancini, Zeynep Akata, and Barbara Caputo. A closer look at self-training for zero-label semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2693-2702, 2021. 3, 7" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 139, + 288, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 139, + 288, + 205 + ], + "spans": [ + { + "bbox": [ + 48, + 139, + 288, + 205 + ], + "type": "text", + "content": "[48] Jie Qin, Jie Wu, Pengxiang Yan, Ming Li, Ren Yuxi, Xuefeng Xiao, Yitong Wang, Rui Wang, Shilei Wen, Xin Pan, et al. Freeseg: Unified, universal and open-vocabulary image segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19446-19455, 2023. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 205, + 288, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 205, + 288, + 281 + ], + "spans": [ + { + "bbox": [ + 48, + 205, + 288, + 281 + ], + "type": "text", + "content": "[49] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In Proceedings of the 38th International Conference on Machine Learning, 2021. 2, 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 283, + 288, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 283, + 288, + 348 + ], + "spans": [ + { + "bbox": [ + 48, + 283, + 288, + 348 + ], + "type": "text", + "content": "[50] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 350, + 288, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 350, + 288, + 403 + ], + "spans": [ + { + "bbox": [ + 48, + 350, + 288, + 403 + ], + "type": "text", + "content": "[51] Kanchana Ranasinghe, Brandon McKinzie, Sachin Ravi, Yinfei Yang, Alexander Toshev, and Jonathon Shlens. Perceptual grouping in contrastive vision-language models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5571-5584, 2023. 2, 3, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 404, + 288, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 404, + 288, + 459 + ], + "spans": [ + { + "bbox": [ + 48, + 404, + 288, + 459 + ], + "type": "text", + "content": "[52] Pengzhen Ren, Changlin Li, Hang Xu, Yi Zhu, Guangrun Wang, Jianzhuang Liu, Xiaojun Chang, and Xiaodan Liang. Viewco: Discovering text-supervised segmentation masks via multi-view semantic consistency. arXiv preprint arXiv:2302.10307, 2023. 2, 3, 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 459, + 288, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 459, + 288, + 514 + ], + "spans": [ + { + "bbox": [ + 48, + 459, + 288, + 514 + ], + "type": "text", + "content": "[53] Lixiang Ru, Yibing Zhan, Baosheng Yu, and Bo Du. Learning affinity from attention: End-to-end weakly-supervised semantic segmentation with transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16846-16855, 2022. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 514, + 288, + 558 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 514, + 288, + 558 + ], + "spans": [ + { + "bbox": [ + 48, + 514, + 288, + 558 + ], + "type": "text", + "content": "[54] Lixiang Ru, Heliang Zheng, Yibing Zhan, and Bo Du. Token contrast for weakly-supervised semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3093-3102, 2023. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 559, + 288, + 603 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 559, + 288, + 603 + ], + "spans": [ + { + "bbox": [ + 48, + 559, + 288, + 603 + ], + "type": "text", + "content": "[55] Lixiang Ru, Heliang Zheng, Yibing Zhan, and Bo Du. Token contrast for weakly-supervised semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3093-3102, 2023. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 604, + 288, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 604, + 288, + 657 + ], + "spans": [ + { + "bbox": [ + 48, + 604, + 288, + 657 + ], + "type": "text", + "content": "[56] Michael Rubinstein, Armand Joulin, Johannes Kopf, and Ce Liu. Unsupervised joint object discovery and segmentation in internet images. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1939-1946, 2013. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 658, + 288, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 288, + 713 + ], + "type": "text", + "content": "[57] Ramprasaath R. Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. Grad-cam: Visual explanations from deep networks via gradient-based localization. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2017. 2, 4" + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 713 + ], + "type": "list", + "angle": 0, + "index": 24, + "blocks": [ + { + "bbox": [ + 307, + 73, + 546, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 546, + 116 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 546, + 116 + ], + "type": "text", + "content": "[58] Gyungin Shin, Weidi Xie, and Samuel Albanie. Reco: Retrieve and co-segment for zero-shot transfer. Advances in Neural Information Processing Systems, 35:33754-33767, 2022. 3, 6, 7" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 307, + 118, + 547, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 547, + 173 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 547, + 173 + ], + "type": "text", + "content": "[59] Oriane Simeoni, Gilles Puy, Huy V Vo, Simon Roburin, Spyros Gidaris, Andrei Bursuc, Patrick Pérez, Renaud Marlet, and Jean Ponce. Localizing objects with self-supervised transformers and no labels. arXiv preprint arXiv:2109.14279, 2021. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 307, + 175, + 546, + 240 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 175, + 546, + 240 + ], + "spans": [ + { + "bbox": [ + 307, + 175, + 546, + 240 + ], + "type": "text", + "content": "[60] Amanpreet Singh, Ronghang Hu, Vedanuj Goswami, Guillaume Couairon, Wojciech Galuba, Marcus Rohrbach, and Douwe Kiela. Flava: A foundational language and vision alignment model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 15638-15650, 2022. 1, 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 241, + 546, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 241, + 546, + 297 + ], + "spans": [ + { + "bbox": [ + 307, + 241, + 546, + 297 + ], + "type": "text", + "content": "[61] Meng Tang, Federico Perazzi, Abdelaziz Djelouah, Ismail Ben Ayed, Christopher Schroers, and Yuri Boykov. On regularized losses for weakly-supervised cnn segmentation. In Proceedings of the European Conference on Computer Vision (ECCV), pages 507-522, 2018. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 298, + 546, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 298, + 546, + 363 + ], + "spans": [ + { + "bbox": [ + 307, + 298, + 546, + 363 + ], + "type": "text", + "content": "[62] Anthony Meng Huat Tiong, Junnan Li, Boyang Li, Silvio Savarese, and Steven C.H. Hoi. Plug-and-play vqa: Zero-shot vqa by conjoining large pretrained models with zero training. In Findings of the Conference on Empirical Methods in Natural Language Processing (Findings of EMNLP), 2022. 1, 2, 4" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 365, + 547, + 421 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 365, + 547, + 421 + ], + "spans": [ + { + "bbox": [ + 307, + 365, + 547, + 421 + ], + "type": "text", + "content": "[63] Huy V Vo, Francis Bach, Minsu Cho, Kai Han, Yann LeCun, Patrick Pérez, and Jean Ponce. Unsupervised image matching and object discovery as optimization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8287-8296, 2019. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 422, + 546, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 422, + 546, + 464 + ], + "spans": [ + { + "bbox": [ + 307, + 422, + 546, + 464 + ], + "type": "text", + "content": "[64] Van Huy Vo, Elena Sizikova, Cordelia Schmid, Patrick Pérez, and Jean Ponce. Large-scale unsupervised object discovery. Advances in Neural Information Processing Systems, 34:16764-16778, 2021. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 466, + 546, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 466, + 546, + 510 + ], + "spans": [ + { + "bbox": [ + 307, + 466, + 546, + 510 + ], + "type": "text", + "content": "[65] Wenguan Wang, Guolei Sun, and Luc Van Gool. Looking beyond single images for weakly supervised semantic segmentation learning. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2022. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 512, + 546, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 512, + 546, + 566 + ], + "spans": [ + { + "bbox": [ + 307, + 512, + 546, + 566 + ], + "type": "text", + "content": "[66] Xudong Wang, Rohit Girdhar, Stella X Yu, and Ishan Misra. Cut and learn for unsupervised object detection and instance segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3124-3134, 2023. 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 567, + 546, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 567, + 546, + 632 + ], + "spans": [ + { + "bbox": [ + 307, + 567, + 546, + 632 + ], + "type": "text", + "content": "[67] Yangtao Wang, Xi Shen, Shell Xu Hu, Yuan Yuan, James L Crowley, and Dominique Vaufreydaz. Self-supervised transformers for unsupervised object discovery using normalized cut. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14543-14553, 2022. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 635, + 546, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 635, + 546, + 679 + ], + "spans": [ + { + "bbox": [ + 307, + 635, + 546, + 679 + ], + "type": "text", + "content": "[68] Xiu-Shen Wei, Chen-Lin Zhang, Jianxin Wu, Chunhua Shen, and Zhi-Hua Zhou. Unsupervised object discovery and co-localization by deep descriptor transformation. Pattern Recognition, 88:113-126, 2019. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 681, + 546, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 681, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 681, + 546, + 713 + ], + "type": "text", + "content": "[69] Yunchao Wei, Jiashi Feng, Xiaodan Liang, Ming-Ming Cheng, Yao Zhao, and Shuicheng Yan. Object region mining with adversarial erasing: A simple classification to semantic" + } + ] + } + ], + "index": 23 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4039" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 66, + 72, + 287, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 72, + 287, + 106 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 287, + 106 + ], + "type": "text", + "content": "segmentation approach. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1568-1576, 2017. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 108, + 287, + 142 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 108, + 287, + 142 + ], + "spans": [ + { + "bbox": [ + 48, + 108, + 287, + 142 + ], + "type": "text", + "content": "[70] Alexandros Xenos, Themos Stafylakis, Ioannis Patras, and Georgios Tzimiropoulos. A simple baseline for knowledge-based visual question answering. In EMNLP, 2023. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 144, + 288, + 199 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 144, + 288, + 199 + ], + "spans": [ + { + "bbox": [ + 48, + 144, + 288, + 199 + ], + "type": "text", + "content": "[71] Yongqin Xian, Subhabrata Choudhury, Yang He, Bernt Schiele, and Zeynep Akata. Semantic projection network for zero-and few-label semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8256-8265, 2019. 3, 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 201, + 288, + 255 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 201, + 288, + 255 + ], + "spans": [ + { + "bbox": [ + 48, + 201, + 288, + 255 + ], + "type": "text", + "content": "[72] Haiyang Xu, Qinghao Ye, Ming Yan, Yaya Shi, Jiabo Ye, Yuanhong Xu, Chenliang Li, Bin Bi, Qi Qian, Wei Wang, et al. mplug-2: A modularized multi-modal foundation model across text, image and video. arXiv preprint arXiv:2302.00402, 2023. 2, 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 258, + 288, + 323 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 258, + 288, + 323 + ], + "spans": [ + { + "bbox": [ + 48, + 258, + 288, + 323 + ], + "type": "text", + "content": "[73] Jiarui Xu, Shalini De Mello, Sifei Liu, Wonmin Byeon, Thomas Breuel, Jan Kautz, and Xiaolong Wang. Groupvit: Semantic segmentation emerges from text supervision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18134-18144, 2022. 2, 3, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 326, + 288, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 326, + 288, + 392 + ], + "spans": [ + { + "bbox": [ + 48, + 326, + 288, + 392 + ], + "type": "text", + "content": "[74] Jilan Xu, Junlin Hou, Yuejie Zhang, Rui Feng, Yi Wang, Yu Qiao, and Weidi Xie. Learning open-vocabulary semantic segmentation models from natural language supervision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2935-2944, 2023. 1, 2, 3, 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 394, + 288, + 449 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 394, + 288, + 449 + ], + "spans": [ + { + "bbox": [ + 48, + 394, + 288, + 449 + ], + "type": "text", + "content": "[75] Lian Xu, Wanli Ouyang, Mohammed Bennamoun, Farid Boussaid, and Dan Xu. Multi-class token transformer for weakly supervised semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4310-4319, 2022. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 452, + 288, + 507 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 452, + 288, + 507 + ], + "spans": [ + { + "bbox": [ + 48, + 452, + 288, + 507 + ], + "type": "text", + "content": "[76] Mengde Xu, Zheng Zhang, Fangyun Wei, Yutong Lin, Yue Cao, Han Hu, and Xiang Bai. A simple baseline for open-vocabulary semantic segmentation with pre-trained vision-language model. In European Conference on Computer Vision, pages 736-753. Springer, 2022. 1, 3, 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 509, + 288, + 563 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 509, + 288, + 563 + ], + "spans": [ + { + "bbox": [ + 48, + 509, + 288, + 563 + ], + "type": "text", + "content": "[77] Mengde Xu, Zheng Zhang, Fangyun Wei, Han Hu, and Xiang Bai. Side adapter network for open-vocabulary semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2945-2954, 2023. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 566, + 288, + 620 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 566, + 288, + 620 + ], + "spans": [ + { + "bbox": [ + 48, + 566, + 288, + 620 + ], + "type": "text", + "content": "[78] Xiao Xu, Chenfei Wu, Shachar Rosenman, Vasudev Lal, Wanxiang Che, and Nan Duan. Bridgetower: Building bridges between encoders in vision-language representation learning. arXiv preprint arXiv:2206.08657, 2022. 1, 2, 3, 6, 8" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 623, + 288, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 623, + 288, + 667 + ], + "spans": [ + { + "bbox": [ + 48, + 623, + 288, + 667 + ], + "type": "text", + "content": "[79] Jiahui Yu, Zirui Wang, Vijay Vasudevan, Legg Yeung, Mojtaba Seyedhosseini, and Yonghui Wu. CoCa: Contrastive captioners are image-text foundation models. arXiv Preprint 2205.01917, 2022. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 670, + 288, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 670, + 288, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 670, + 288, + 712 + ], + "type": "text", + "content": "[80] Nir Zabari and Yedid Hoshen. Open-vocabulary semantic segmentation using test-time distillation. In European Conference on Computer Vision, pages 56-72. Springer, 2022. 1" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 564 + ], + "type": "list", + "angle": 0, + "index": 23, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "text", + "content": "[81] Yan Zeng, Xinsong Zhang, and Hang Li. Multi-grained vision language pre-training: Aligning texts with visual concepts. arXiv preprint arXiv:2111.08276, 2021. 2, 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 307, + 108, + 545, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 108, + 545, + 150 + ], + "spans": [ + { + "bbox": [ + 307, + 108, + 545, + 150 + ], + "type": "text", + "content": "[82] Yan Zeng, Xinsong Zhang, Hang Li, Jiawei Wang, Jipeng Zhang, and Wangchunshu Zhou. " + }, + { + "bbox": [ + 307, + 108, + 545, + 150 + ], + "type": "inline_equation", + "content": "X^2" + }, + { + "bbox": [ + 307, + 108, + 545, + 150 + ], + "type": "text", + "content": "-vlm: All-in-one pretrained model for vision-language tasks. arXiv preprint arXiv:2211.12402, 2022. 2, 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 152, + 545, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 152, + 545, + 206 + ], + "spans": [ + { + "bbox": [ + 307, + 152, + 545, + 206 + ], + "type": "text", + "content": "[83] Bingfeng Zhang, Jimin Xiao, Yunchao Wei, Mingjie Sun, and Kaizhu Huang. Reliability does matter: An end-to-end weakly supervised semantic segmentation approach. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 12765-12772, 2020. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 208, + 545, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 208, + 545, + 262 + ], + "spans": [ + { + "bbox": [ + 307, + 208, + 545, + 262 + ], + "type": "text", + "content": "[84] Hao Zhang, Feng Li, Xueyan Zou, Shilong Liu, Chunyuan Li, Jianwei Yang, and Lei Zhang. A simple framework for open-vocabulary segmentation and detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1020-1031, 2023. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 264, + 545, + 328 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 264, + 545, + 328 + ], + "spans": [ + { + "bbox": [ + 308, + 264, + 545, + 328 + ], + "type": "text", + "content": "[85] Pengchuan Zhang, Xiujun Li, Xiaowei Hu, Jianwei Yang, Lei Zhang, Lijuan Wang, Yejin Choi, and Jianfeng Gao. Vinvl: Revisiting visual representations in vision-language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5579-5588, 2021. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 331, + 545, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 331, + 545, + 385 + ], + "spans": [ + { + "bbox": [ + 308, + 331, + 545, + 385 + ], + "type": "text", + "content": "[86] Xiaolin Zhang, Yunchao Wei, Jiashi Feng, Yi Yang, and Thomas S Huang. Adversarial complementary learning for weakly supervised object localization. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1325-1334, 2018. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 387, + 545, + 419 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 387, + 545, + 419 + ], + "spans": [ + { + "bbox": [ + 308, + 387, + 545, + 419 + ], + "type": "text", + "content": "[87] Bolei Zhou, Aditya Khosla, Agata Lapedriza, Aude Oliva, and Antonio Torralba. Object detectors emerge in deep scene cnns. arXiv preprint arXiv:1412.6856, 2014. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 420, + 545, + 474 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 420, + 545, + 474 + ], + "spans": [ + { + "bbox": [ + 308, + 420, + 545, + 474 + ], + "type": "text", + "content": "[88] Bolei Zhou, Hang Zhao, Xavier Puig, Sanja Fidler, Adela Barriuso, and Antonio Torralba. Scene parsing through ade20k dataset. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 633-641, 2017. 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 475, + 545, + 509 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 475, + 545, + 509 + ], + "spans": [ + { + "bbox": [ + 308, + 475, + 545, + 509 + ], + "type": "text", + "content": "[89] Chong Zhou, Chen Change Loy, and Bo Dai. Extract free dense labels from clip. In European Conference on Computer Vision, pages 696-712. Springer, 2022. 3, 6, 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 510, + 545, + 564 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 510, + 545, + 564 + ], + "spans": [ + { + "bbox": [ + 308, + 510, + 545, + 564 + ], + "type": "text", + "content": "[90] Tianfei Zhou, Meijie Zhang, Fang Zhao, and Jianwu Li. Regional semantic contrast and aggregation for weakly supervised semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4299-4309, 2022. 1" + } + ] + } + ], + "index": 22 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "4040" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/EmoGen_ Emotional Image Content Generation with Text-to-Image Diffusion Models/a0b8035c-47d5-4f15-8dd7-134ed07d94b6_content_list.json b/2024/EmoGen_ Emotional Image Content Generation with Text-to-Image Diffusion Models/a0b8035c-47d5-4f15-8dd7-134ed07d94b6_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..d22c26a78e32cae470c75fcd395fbd52fcf8acc6 --- /dev/null +++ b/2024/EmoGen_ Emotional Image Content Generation with Text-to-Image Diffusion Models/a0b8035c-47d5-4f15-8dd7-134ed07d94b6_content_list.json @@ -0,0 +1,1461 @@ +[ + { + "type": "text", + "text": "EmoGen: Emotional Image Content Generation with Text-to-Image Diffusion Models", + "text_level": 1, + "bbox": [ + 217, + 130, + 754, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Jingyuan Yang, Jiawei Feng, Hui Huang*", + "bbox": [ + 323, + 203, + 651, + 222 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Shenzhen University", + "bbox": [ + 401, + 223, + 566, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{jingyuanyang.jyy, fengjiawei0909, hhzhiyan}@gmail.com", + "bbox": [ + 243, + 241, + 720, + 257 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/3fc12973e2e4509d7d55d6a4667bb7fd1e11c39a8516f2e5b365b787e9af60b1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 106, + 263, + 287, + 412 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/69d9a7ffc9eaf994bd51691e7c3dcdf7393ac08e185033f3146b3dac0706c75b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 299, + 263, + 478, + 411 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/1a3ffcf1c2f079935bf3944a83119a81baa91e5ecba26c7991bfda597ff61615.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 491, + 263, + 669, + 411 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/91b5b344aa006292105d35ca67e41db0f7a85238bd932ee239522ff8cd48e43c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 684, + 265, + 864, + 411 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/9c3314b79d1222a4519c490cb43c5d562ff6ac46e47638505ce2bbe17c1b0b7a.jpg", + "image_caption": [ + "Figure 1. Emotional Image Content Generation (EICG). Given an emotion category, our network produces images that exhibit unambiguous meanings (semantic-clear), reflect the intended emotion (emotion-faithful) and incorporate varied semantics (semantic-diverse)." + ], + "image_footnote": [], + "bbox": [ + 106, + 417, + 287, + 566 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/158a4ac8a6ec0c14f6519a00a14c99a86ac1e012f548eab7c2971d580ad69c94.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 299, + 417, + 478, + 566 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/97ce0c88ceeb33497b4ceb32c5b00b94fe8967464248edd70559c9c7e1acf3fb.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 491, + 417, + 669, + 565 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/aec73d73247d35cafd079ad62262717cc1ac515a3e4cc52c8657dbd1dc9d78e9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 684, + 417, + 864, + 565 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 619, + 313, + 636 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recent years have witnessed remarkable progress in image generation task, where users can create visually astonishing images with high-quality. However, existing text-to-image diffusion models are proficient in generating concrete concepts (dogs) but encounter challenges with more abstract ones (emotions). Several efforts have been made to modify image emotions with color and style adjustments, facing limitations in effectively conveying emotions with fixed image contents. In this work, we introduce Emotional Image Content Generation (EICG), a new task to generate semantic-clear and emotion-faithful images given emotion categories. Specifically, we propose an emotion space and construct a mapping network to align it with the powerful Contrastive Language-Image Pre-training (CLIP) space, providing a concrete interpretation of abstract emo", + "bbox": [ + 73, + 642, + 473, + 869 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "tions. Attribute loss and emotion confidence are further proposed to ensure the semantic diversity and emotion fidelity of the generated images. Our method outperforms the state-of-the-art text-to-image approaches both quantitatively and qualitatively, where we derive three custom metrics, i.e., emotion accuracy, semantic clarity and semantic diversity. In addition to generation, our method can help emotion understanding and inspire emotional art design. Project page: https://vcc.tech/research/2024/EmoGen.", + "bbox": [ + 500, + 622, + 893, + 758 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 500, + 782, + 632, + 797 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "\"What I cannot create, I do not understand.\"", + "bbox": [ + 547, + 806, + 844, + 821 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Richard Feynman", + "bbox": [ + 761, + 821, + 890, + 837 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Emotions, often elusive yet profoundly influential, shape our actions, foster connections, and spark passions. With the prevalence of social medias, users tend to share specially crafted images to express their feelings. Aiming to find out", + "bbox": [ + 496, + 839, + 893, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author", + "bbox": [ + 94, + 887, + 220, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "6358", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "people's emotional responses towards different stimuli, Visual Emotion Analysis (VEA) is an intriguing yet challenging task in computer vision [34, 50, 51]. Recent years have witnessed rapid development in this field, bringing potential applications such as opinion mining [48], market advertising [6] and mental healthcare [17].", + "bbox": [ + 75, + 90, + 468, + 181 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Thanks to the advent of diffusion models [7, 16, 38], unprecedented progress has been made in text-to-image generation, where users can generate high-quality images with crafted prompts or personalized objects [10, 39, 56]. Existing text-to-image diffusion models, are often excel in generating concrete concepts (e.g., cat, house, mountain) but face limitations when tasked with more abstract ones (e.g., amusement, anger, sadness). In reality, however, photographic works are not necessarily targeted on specific entities, but are often composed to convey certain feelings.", + "bbox": [ + 75, + 184, + 467, + 335 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "A natural question arises: What if machines could create images that not only please our eyes but also touch our hearts? Generating emotions is very challenging. Emotions are abstract while images are concrete, leaving the affective gap [13] hard to surmount. To bridge the gap, several efforts have been made to modify visual emotions by adjusting colors and styles, i.e., image emotion transfer [30, 41, 47]. These methods, however, meet difficulties in evoking emotions correctly and significantly, i.e., $29\\%$ emotion accuracy [47], as fixed image contents limit emotional variations. Moreover, we cannot generate emotional images solely from colors and styles. What truly triggers emotion? Psychological studies show that visual emotions are often evoked by specific semantics [1, 3, 4].", + "bbox": [ + 75, + 337, + 467, + 547 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we propose Emotional Image Content Generation (EICG), a new task to generate semantically clear and emotionally faithful visual contents conditioned on a given emotion category, as shown in Figure 1. Semantic clarity demands an unambiguous representation of visual contents, while emotion faithfulness entails generating images evoke the intended emotions. Contrastive Language-Image Pre-training (CLIP) [31] is a large-scale vision-language model with rich semantics. However, we observe in Figure 2 that CLIP space can not well capture emotional relationships. Therefore, we introduce an emotion space, which groups similar emotions together while keeping dissimilar ones apart. While emotion space excels in representing emotions, CLIP space exhibits a powerful semantic structure. To align emotion space with CLIP space, we propose a mapping network, interpreting abstract emotions with concrete semantics.", + "bbox": [ + 75, + 550, + 467, + 806 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "EmoSet [53] is a recently proposed large-scale visual emotion dataset with rich attributes. The Latent Diffusion Model (LDM) loss [38] is often utilized to optimize concrete entities with single and explicit semantics, posing a challenge in capturing the diversity within each emotion. To address this, we introduce an attribute loss to ensure seman", + "bbox": [ + 75, + 810, + 467, + 898 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/a4ce53800202aec8a914937794b0b8f5d576aaaca2c2400d8161dd98a0244f6b.jpg", + "image_caption": [ + "(a) Emotion Space", + "Figure 2. Despite (b) CLIP space demonstrates a powerful semantic structure, it struggles to effectively capture emotional relationships within (a), the proposed emotion space." + ], + "image_footnote": [], + "bbox": [ + 506, + 90, + 694, + 237 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/d05b48c7b7f51571c138f86093c71efeb0a5d825959e0721fb579b67e46b6299.jpg", + "image_caption": [ + "(b)CLIP Space" + ], + "image_footnote": [], + "bbox": [ + 697, + 90, + 885, + 237 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "tic clarity and diversity, by leveraging the attribute labels in EmoSet. Recognizing that not all objects are affective, emotion confidence is further proposed to ensure the emotion fidelity of the generated contents.", + "bbox": [ + 496, + 324, + 890, + 383 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To estimate the generation quality of EICG, three evaluation metrics are specially designed: emotion accuracy, semantic clarity and semantic diversity. As EICG aims to create emotional contents, we design emotion accuracy to measure the alignment between intended and perceived emotions in the generated images. People are prone to evoke emotions only when the contents are easily recognizable. Thus we propose semantic clarity to assess the unambiguity of the generated image content. Additionally, in view of the assorted emotion stimuli, we devise semantic diversity to quantify the content richness under each emotion. We evaluate our method through both qualitative and quantitative analyses, surpassing the state-of-the-art text-to-image generation approaches across five metrics. Ablation studies are performed to verify the network design, and user studies are conducted to resonate our method with human viewers. Besides generation task, our method can also be applied to decompose emotion concepts, transfer emotional contents and fuse different emotions, which may be helpful to understand emotions and create emotional art design.", + "bbox": [ + 496, + 388, + 892, + 691 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In summary, our contributions are:", + "bbox": [ + 517, + 696, + 750, + 710 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We introduce Emotional Image Content Generation, a novel task to generate emotion-faithful and semantic-clear image contents. We also derive three custom metrics to estimate the generation performance.", + "- We develop a mapping network to align the proposed emotion space to the powerful CLIP space, where attribute loss and emotion confidence are further designed to ensure the semantic richness and emotion fidelity.", + "- We evaluate our method against the state-of-the-art text-to-image approaches and demonstrate our superiority. Potential applications are exhibited for emotion understanding and emotional art design." + ], + "bbox": [ + 500, + 715, + 890, + 897 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "6359", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related work", + "text_level": 1, + "bbox": [ + 76, + 90, + 215, + 104 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.1. Visual Emotion Analysis", + "text_level": 1, + "bbox": [ + 76, + 114, + 302, + 131 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Researchers have been involved in VEA for over two decades, ranging from early traditional approaches [2, 23, 26] to recent deep learning ones [35, 51, 52, 58]. Given the inherent abstractness and complexity of visual emotion, researchers aim to identify the most influential elements, which range from low-level features like color, texture and style [23, 26, 35, 58] to high-level semantics [2, 35, 51, 52, 58]. Lee et al. [23] propose a scheme to evaluate emotional response from color images by reasoning the prototypical color for each emotion and the input images. As a milestone, Machajdik et al. [26] extract representative low-level features in composition, including color and texture, to predict visual emotions. Besides low-level features, Borth et al. [2] propose Adjective-Noun Pair (ANP) and build a visual concept detector named Sentibank. With the help of deep learning techniques, Rao et al. [35] construct MldrNet to extract emotional clues from pixel-level, aesthetic-level and semantic-level. To form a more discriminative emotional representation, Zhang et al. [58] integrate high-level contents and low-level styles. Yang et al. propose network to mine emotions from multiple objects [52] as well as object-scene correlations [51]. Existing work often treat VEA as a classification task, i.e., input an image and predict the emotion within it. Can we reverse this process? In other words, can we generate an image targeting on the given emotion word? Only by creating emotional images can we demonstrate the significance of visual elements, leading to a better understanding of emotions.", + "bbox": [ + 76, + 138, + 472, + 563 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. Text-to-Image Generation", + "text_level": 1, + "bbox": [ + 76, + 574, + 313, + 589 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Text-to-image generation aims to convert textual descriptions into corresponding realistic images. Existing generative models can be grouped into GANs [12, 24, 59], VAEs [9, 20, 55], flow-based [37], energy-based [22] and diffusion-based [7, 16, 38, 39, 56]. Diffusion models are witnessed impressive and rapid progress in recent years, where methods like GLIDE [28], DALLE2 [32], Image [40] can generate diverse, photo-realistic and high-quality images. Notably, Stable diffusion [38] is one of the most popular diffusion models, owing to its stable training and the capability for fine-grained control, supported by an active user community. For customized generation, several diffusion-based text-to-image methods are introduced, where methods vary from learning a new embedding [8, 10] and finetuning the network parameters [21, 39, 46]. Textual inversion [10] and DreamArtist [8] learn new concepts with a few user-provided images in the word embedding space, without further training on diffusion models. While DreamBooth [39] finetunes all the parameters to learn a new concept, Custom diffusion [21] only updates the key", + "bbox": [ + 76, + 599, + 472, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "and value parameters in the cross attention layers. Further, ELITE [46] speeds up the running time with accurate generation results by adopting a global and local mapping network. Existing text-to-images models are capable of generating concrete concepts [7, 24, 56], or personalized ones [10, 21, 39], but face difficulties in generating more abstract ones. In reality, photographic works are not necessarily composed of targeted concepts, but often aim to convey specific feelings. Thus, how to generate emotion-evoking images remains a pressing and critical challenge.", + "bbox": [ + 496, + 90, + 893, + 244 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.3. Image Emotion Transfer", + "text_level": 1, + "bbox": [ + 498, + 253, + 725, + 268 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Image style transfer [11] aims to render the semantic content under different styles, producing visually stunning results [19, 33, 45, 54]. Similarly, image color transfer [36] seeks to adjust and harmonize the color characteristics of one image to match another [18, 29]. Specifically, color and style choices can strongly influence the emotions of an image [27]. By adjusting low-level visual elements, image emotion transfer aims to modify the emotional tone of the input image, including the color-based methods [5, 25, 30, 44, 49, 60] and the style-based ones [41, 47]. Yang and Peng et al. [49] makes the first attempt to transfer image colors. Wang et al. [44] present a system to modify the image color according to a given emotion word, and Liu et al. [25] further advance it with deep learning techniques. Peng et al. [30] introduce a new approach to alter the emotion of an input image by guiding its color and texture under the target image. More recently, to reflect emotions in styles, Sun et al. [41] and Weng et al. [47] bring promising results on emotion-aware image style transfer. Nevertheless, the alteration of visual emotions through colors and styles is limited due to fixed content, resulting in subtle emotional changes, i.e., $29\\%$ emotion accuracy in [47]. Psychological studies suggest that visual emotions can be elicited by specific semantics [3]. Thus, we propose a novel method to generate emotional image contents with clear semantics.", + "bbox": [ + 496, + 277, + 893, + 655 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 500, + 670, + 591, + 686 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Emotion Representation", + "text_level": 1, + "bbox": [ + 500, + 696, + 723, + 710 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Emotion Space EICG is a challenging task, which requires both semantic clarity and emotion fidelity. How to generate an image with distinct and emotional semantics? CLIP [31] is developed to align image and text modalities, where semantically related features are located in close proximity to each other. While CLIP shows impressive semantic representation capabilities, it struggles to effectively capture emotional relationships. As demonstrated in Figure 2, we can observe that sample points with emotional similarities are distantly separated within the CLIP space due to their differing semantics, e.g., toy, amusement park and Christmas tree. To better depict emotional relation", + "bbox": [ + 496, + 719, + 893, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "6360", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/55fe37c14368b504bde71e5a8fc55fb3d6f580460f8ec9755dcc558656664520.jpg", + "image_caption": [ + "Figure 3. Training process of our network. Emotion representation (stage 1) learns a well-behaved emotion space and emotion content generation (stage 2) maps this space to CLIP space, aiming to generate image contents with emotion fidelity, semantic clarity and diversity." + ], + "image_footnote": [], + "bbox": [ + 127, + 87, + 841, + 255 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "ships, we introduce the emotion space, a latent space that clusters similar emotions together while keeping dissimilar ones apart. EmoSet [53] is a large-scale dataset with rich attributes, where each image is labeled with an emotion. Using aligned image-emotion pairs, we construct an encoder $\\varphi$ with ResNet-50 [14] to capture emotion representations. To train the encoder, we devise an emotion loss by implementing the widely-used Cross-Entropy (CE) loss, following the previous work [50, 51]:", + "bbox": [ + 75, + 304, + 472, + 441 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {e m o} = - \\sum_ {i = 1} ^ {C} y _ {e m o} \\log \\frac {\\exp (\\varphi (x , i))}{\\sum_ {i = 1} ^ {C} \\exp (\\varphi (x , i))}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 122, + 446, + 468, + 487 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $x$ represents the input image, $y_{emo}$ denotes the emotion label and $C$ stands for the total number of emotion categories. Once the loss function converges, emotion space is established. Parameters in emotion encoder remain fixed in the following emotional content generation process.", + "bbox": [ + 75, + 493, + 468, + 568 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "During inference, each emotion cluster is represented by a Gaussian distribution with learned parameters, i.e., mean and standard deviation. For example, when taking amusement as input, we randomly sample a data point from corresponding Gaussian distribution to serve as its emotion representation, as shown in Figure 3. We have confirmed that Gaussian distribution suits emotion clusters well and the random sampling process induces diversity to EICG.", + "bbox": [ + 75, + 568, + 470, + 690 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Emotional Content Generation", + "text_level": 1, + "bbox": [ + 76, + 696, + 352, + 710 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Mapping Network While emotion space is emotionally separable, CLIP space captures rich semantics. Existing text-to-image models entail clear and specific semantics as input, making CLIP space indispensable in the generation process. Consequently, establishing the mapping between emotion space and CLIP space becomes a crucial challenge. Intuitively, we attempt to build the mapping network using fully connected layers, following previous work [33, 43].", + "bbox": [ + 75, + 719, + 468, + 839 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "However, as depicted in Figure 2, clustered feature points in the emotion space are expected to disperse in the CLIP space to capture diverse semantics. Therefore, we utilize a Multilayer Perceptron (MLP) to build the mapping", + "bbox": [ + 75, + 839, + 470, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "network, incorporating non-linear operations, i.e., RELU, to facilitate the separation process. The non-linear projection $F$ is succeeded by a CLIP text transformer $t_{\\theta}$ , yielding textual embedding for U-Net. The end-token embedding of the transformer's output is passed through a fully-connected layer, producing the CLIP text feature. Particularly, to better preserve the prior knowledge in the CLIP space, parameters in the transformer and linear projection are kept frozen, while parameters in non-linear projection are learned, as depicted in Figure 3.", + "bbox": [ + 496, + 304, + 893, + 455 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Attribute Loss Existing text-to-image diffusion models often employ Latent Diffusion Model (LDM) loss [38] for optimization process [10, 39, 56]:", + "bbox": [ + 496, + 472, + 893, + 518 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {L D M} = \\mathbb {E} _ {z, x, \\epsilon , t} \\left[ \\| \\epsilon - \\epsilon_ {\\theta} (z _ {t}, t, t _ {\\theta} (F (\\varphi (x)))) \\| _ {2} ^ {2} \\right], \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 514, + 522, + 890, + 549 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\epsilon$ represents the added noise, $\\epsilon_{\\theta}$ denotes the denoising network and $z_{t}$ indicates the latent noised to time $t$ .", + "bbox": [ + 496, + 553, + 893, + 583 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In these cases, target concepts typically involve concrete entities (e.g., dog, car, flower) or personalized objects (e.g., someone's corgi). These concepts often exhibit consistency on semantic level and share certain similarities on pixel level. However, emotions are abstract concepts, where multiple semantics coexist under one specific category. Learning emotions solely with LDM loss may pose some challenges. For one thing, each emotion might collapse to a specific semantic point, e.g., amusement collapsing to amusement park, losing intra-class diversity. In reality, semantics within one emotion are diverse, where single point cannot fully capture. Moreover, since LDM loss is designed to reconstruct the input image, it primarily focuses on learning and preserving pixel-level commonalities such as color and texture. In Figure 4 (a), with LDM loss alone, CLIP embedding for amusement is prone to be colorful, without exhibiting explicit and diverse semantics. We can conclude that it is hard to achieve robust emotion representations in CLIP space by implementing LDM loss alone.", + "bbox": [ + 496, + 583, + 892, + 869 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In the pursuit of clear and diverse contents, semantics guidance is essential for the generation process. Thanks to", + "bbox": [ + 498, + 869, + 893, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "6361", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/26e2189a5806e3c24af1b9000267e94f9341a5a03dda4f02a746e96920ad2454.jpg", + "image_caption": [ + "Figure 4. Motivation for loss function design. Compare to (a) LDM loss alone, (b) attribute loss enhances semantic clarity while (c) emotion confidence ensures emotion accuracy." + ], + "image_footnote": [], + "bbox": [ + 96, + 89, + 447, + 262 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "the rich attribute annotations in EmoSet, we select the mid-level attributes, i.e., object class and scene type to guide the generation process. With this semantic guidance, we formulate an attribute loss to guarantee that the generated image contents possess clear and diverse semantics. For clarity, emotions are easily triggered in people only when visual contents are represented in an unambiguous manner. Considering the varied emotional stimuli in reality, attribute loss guides the network to learn multiple semantics under one specific emotion. Our attribute loss is devised on CLIP space, by calculating the cosine similarities $f(\\cdot)$ and optimizing a symmetric CE loss over the similarity scores [31]:", + "bbox": [ + 75, + 323, + 472, + 506 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {a t t r} = - \\sum_ {j = 1} ^ {K} y _ {a t t r} \\log \\frac {\\exp (f (v _ {e m o} , \\tau_ {\\theta} (a _ {j})))}{\\sum_ {j = 1} ^ {K} \\exp (f (v _ {e m o} , \\tau_ {\\theta} (a _ {j})))}, \\quad (3)\n$$\n", + "text_format": "latex", + "bbox": [ + 89, + 516, + 468, + 561 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nf (p, q) = \\frac {p \\cdot q}{\\| p \\| \\| q \\|}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 204, + 583, + 468, + 616 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $a_{j}$ denotes the $j$ member in the attribute set, $\\tau_{\\theta}$ represents the text encoder, $v_{emo}$ implies the learned CLIP embedding and $K$ indicates the total number of the attributes. With the attribute loss, each sample point is converging towards the correct semantic and distancing itself from the incorrect ones. Through the combination of attribute loss and LDM loss, we can effectively map each emotion to clear and diverse semantics, as demonstrated in Figure 4 (b).", + "bbox": [ + 75, + 625, + 468, + 747 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Emotion Confidence However, it is worth noting that some of the semantics in Figure 4 (b) appear emotionally neutral, e.g., plant and tree. Since attributes are annotated objectively, not all the attributes in EmoSet are emotional. Therefore, we propose emotion confidence to measure the correlations between emotions and semantic attributes. Initially, we gather all images associated with attribute $j$ in EmoSet and send them to a pre-trained emotion classifier. Each image is predicted as an emotion vector $p(\\cdot)$ and we", + "bbox": [ + 75, + 763, + 470, + 902 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/28e43b6a69be182719839923a26db0c9befa21c76c8cabb777821c2a4a8dfddc.jpg", + "image_caption": [ + "Figure 5. Illustration of emotion confidence. Each (a) attribute is represented by (b) a distribution of confidence on eight emotions." + ], + "image_footnote": [], + "bbox": [ + 519, + 88, + 870, + 207 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "sum all images up to get the emotional distribution $d_{j}$ for attribute $j$ . Each emotion $i$ within this distribution is assigned a corresponding emotion confidence $\\alpha_{ij}$ :", + "bbox": [ + 498, + 251, + 890, + 297 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\alpha_ {i j} = \\frac {1}{N _ {j}} \\sum_ {n = 1} ^ {N _ {j}} p (x _ {n}, i), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 612, + 305, + 890, + 348 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $x_{n}$ represents the input image and $N_{j}$ denotes the total image number in attribute $j$ . We further illustrate the above process in Figure 5 with visual representations. When mountain snowy appears, people are more likely to experience awe and cemetery often elicits sadness. In contrast, the presence of tree in every emotion category suggests its lack of emotional specificity. Some attributes are emotion-related while others are not, which can be beneficial for generating emotional contents. We then use emotion confidence to balance between LDM loss and attribute loss:", + "bbox": [ + 496, + 354, + 890, + 505 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\left(1 - \\alpha_ {i j}\\right) \\mathcal {L} _ {L D M} + \\alpha_ {i j} \\mathcal {L} _ {\\text {a t t r}}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 578, + 513, + 890, + 531 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $i$ represents the emotion category $y_{emo}$ and $j$ denotes the attribute type $y_{att}$ . The greater the emotion confidence $\\alpha_{ij}$ is, the stronger the impact attribute $j$ has on the specific emotion $i$ . Low confidence suggests a weak connection between the attribute and emotion, signaling that the network should learn more from the pixel-wise LDM loss. When higher confidence occurs, the network should prioritize the semantic meaning of the image, i.e., the attribute loss. With this design, our network can adapt to a wide range of cases, generating image contents that are both semantically explicit and emotionally faithful, as shown in Figure 4 (c).", + "bbox": [ + 496, + 539, + 892, + 705 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 500, + 715, + 633, + 733 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Dataset and Evaluation", + "text_level": 1, + "bbox": [ + 500, + 741, + 717, + 757 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Dataset EmoSet [53] is a large-scale visual emotion dataset with rich attributes, comprising a total of 118,102 images. To investigate the connections between emotions and specific contents, we create a subset from EmoSet by preserving images with object/scene labels. Each image is labeled with both emotion and attribute labels, guiding the optimization process of emotion loss and attribute loss. Notably, the wide range of attribute labels assures for learning diverse and representative emotional contents.", + "bbox": [ + 496, + 763, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "6362", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/6f2b27521357d6835976cc477fb41cce39cc1ad564a1057d9b5736ac9efb940d.jpg", + "image_caption": [ + "Figure 6. Qualitative comparisons with the state-of-the-art text-to-image generation approaches and ablation studies of our method." + ], + "image_footnote": [], + "bbox": [ + 83, + 88, + 893, + 470 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation Metrics To comprehensively evaluate the performance of different methods on EICG task, we utilize commonly used metrics (FID, LPIPS) and design some specific ones (Emo-A, Sem-C, Sem-D). 1) FID: Frechet Inception Distance (FID) [15] quantifies the distribution distance between generated and real images, providing an estimate of image fidelity. 2) LPIPS: Similar to [42], we employ LPIPS [57] to assess the overall image diversity, with higher values indicating better performance. 3) Emo-A: Since EICG aims at creating emotion-evoking images, we design emotion accuracy to assess the emotional alignment between the targeted emotions and the generated images. 4) Sem-C: People are easily to evoke emotions under recognizable contents. We thus introduce semantic clarity to assess the explicitness of generated image contents. 5) Sem-D: Emotions are complex, where each can be triggered by multiple factors. To cover a diverse range of potential scenes or objects, we derive semantic diversity to estimate the content richness associated with each emotion. For more details, please refer to the supplementary materials.", + "bbox": [ + 76, + 498, + 472, + 801 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Comparisons", + "text_level": 1, + "bbox": [ + 76, + 815, + 217, + 830 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "As our method is the first attempt in EICG, we compare it with the most relevant and state-of-the-art text-to-image generation techniques: Stable diffusion [38], Textual inversion [10] and Dreambooth [39]. While Stable diffusion is", + "bbox": [ + 75, + 839, + 468, + 902 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/fd04e9eedb303e1f29ac501c24a40fbcd9a662f66fbfc42e32f7cae6e2c8d68f.jpg", + "table_caption": [ + "Table 1. Comparisons with the state-of-the-art methods and ablation studies on emotion generation task, involving five metrics." + ], + "table_footnote": [], + "table_body": "
MethodFID ↓LPIPS ↑Emo-A ↑Sem-C ↑Sem-D ↑
Stable Diffusion [38]44.050.68770.77%0.6080.0199
Textual Inversion [10]50.510.70274.87%0.6050.0282
DreamBooth [39]46.890.66170.50%0.6140.0178
Ours41.600.71776.25%0.6330.0335
w/o F57.540.71371.12%0.6150.0261
w/o Lattr51.130.70765.75%0.5920.0270
w/o αij43.300.71474.88%0.5910.0263
", + "bbox": [ + 504, + 529, + 883, + 648 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "a general image generation pipeline, Textual inversion and Dreambooth specialize in customized image generation.", + "bbox": [ + 498, + 654, + 890, + 686 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Qualitative Comparisons In Figure 6, our method is qualitatively compared with the state-of-the-art methods across three emotion categories, i.e., awe, anger and contentment. Generation results of the rest five emotions can be found in the supplementary materials. Take awe as an example, all the three compared methods tend to produce images with dense textures and dim colors, which suggests that representations for each emotion may collapse to a single feature point. For anger and contentment, both Stable diffusion and Dreambooth distort the visual representations, e.g., tiger and bicycle, and generate some contents with ambiguous semantics. Though Textual inversion preserves some semantic fidelity, it generates emotion-agnostic", + "bbox": [ + 496, + 703, + 892, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6363", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/b87abdfa5aca660f3ce77f0748e7127db7f78133fb42e6bf163494c608a017a7.jpg", + "table_caption": [ + "Table 2. User preference study. The numbers indicate the percentage of participants who prefer our results over those compared methods, given the same emotion category as input." + ], + "table_footnote": [], + "table_body": "
MethodImage fidelity ↑Emotion faithfulness ↑Semantic diversity ↑
Stable Diffusion67.86±15.08%73.66±11.80%87.88±9.64%
Textual Inversion79.91±16.92%72.75±16.90%85.66±10.51%
DreamBooth77.23±14.00%80.79±8.64%81.68±17.06%
", + "bbox": [ + 78, + 136, + 468, + 200 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "contents such as shoes and cars. Since these methods are crafted to learn customized concepts, challenges may arise when handling complex and diverse emotional images. Rather than generating plants and trees, our method can provide diverse and emotion-evoking image contents for awe through lakes, oceans, valleys and snow-covered mountains. In anger, our approach extends beyond mere beasts, encompassing flags, posters, and guns. Owing to attribute loss and emotion confidence, our method can effectively capture the rich and varied semantics while maintaining emotion faithfulness in EmoSet.", + "bbox": [ + 75, + 209, + 472, + 375 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Quantitative Comparisons As shown in Table 1, the proposed method surpasses the compared methods across all five evaluation metrics. Particularly, better performance on FID and LPIPS indicates our method can generate images with higher fidelity and diversity, effectively capturing the characteristics of the training data. All methods achieve comparable results on emotion accuracy. From Figure 6, we observe that comparison methods are prone to fall into singular or incorrect emotion representations. Even such generation results are still separable in eight classes, they do not conform to human emotional cognition. This suggests that relying solely on Emo-A may be insufficient for EICG task. Therefore, we additionally introduce Sem-C and Sem-D to estimate the content clarity and diversity, where our method exhibits a clear advantage over other methods.", + "bbox": [ + 75, + 396, + 472, + 623 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "User Study Besides qualitative and quantitative comparisons, we also conduct a user study to determine whether our method is preferred by humans and to understand how people perceive emotions. We invite 14 participants from different social backgrounds and each test session lasts about 30 minutes. In the first part, generation results are evaluated on three dimensions: image fidelity, emotion faithfulness and semantic diversity. Each question presented to the participants includes two sets of images conveying the same emotion, drawn from our method and one of the comparison methods. The participants are then asked: which group is more realistic? which group evokes a stronger sense of [emotion type]? which group is more diverse? As illustrated in Table 2, our method attains the top rankings compared to the other three methods, particularly excelling in semantic diversity. We aim to explore the factors influencing visual emotions in the second part.", + "bbox": [ + 75, + 643, + 472, + 901 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/9cee6fd645a6a66e7296391c0540e7c64975abbb174d334b372a99b66aa45f47.jpg", + "image_caption": [ + "Figure 7. Emotion decomposition. Each emotion word is broken down into a set of (a) emotional concepts, reflecting the semantics in (b) generated images." + ], + "image_footnote": [], + "bbox": [ + 504, + 88, + 893, + 262 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/c2b88aade77b397846f0c19948209b76776aaaf18d95b909ebae15a5782c92e8.jpg", + "table_caption": [ + "Table 3. Comparisons with the state-of-the-art methods on emotion transfer task, involving three metrics." + ], + "table_footnote": [], + "table_body": "
MethodEmo-A↑CLIP-img ↑CLIP-text ↑
amusementfearamusementfearamusementfear
Stable Diffusion51.54%56.67%0.9290.8250.2570.251
Textual Inversion60.82%40.00%0.9020.7920.2700.259
Ours72.16%63.33%0.9130.8410.2760.270
", + "bbox": [ + 501, + 366, + 890, + 444 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Participants are shown an emotional image generated by our method and are asked: which emotion best describes the image? why do you feel such emotion? Compared to the $76.25\\%$ machine predicted one in Table 1, $82.14\\%$ emotion accuracy is achieved by user voting, where generated images are more emotion-evoking towards human participants. Additionally, $88.39\\%$ of the responses indicate that emotions are predominantly triggered by the content/semantic. This underscores how our task, EICG, is closely aligned with human cognition.", + "bbox": [ + 496, + 470, + 890, + 623 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Ablation Study", + "text_level": 1, + "bbox": [ + 500, + 645, + 653, + 662 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We examine the efficacy of each network design, encompassing the non-linear mapping network $F$ , the attribute loss $\\mathcal{L}_{attr}$ and the emotion confidence $\\alpha_{ij}$ . In Table 1, without nonlinear mapping network, emotion representations are aggregated, which fails to restore the real image set (high FID) and lacks semantic diversity (low Sem-D). Attribute loss is introduced to enhance semantic clarity and diversity, whose absence leads to performance drops in Sem-C and Sem-D. Besides, as shown in Figure 6, generated images exhibit semantic distortions when attribute loss is absent (w/o $\\mathcal{L}_{attr}$ ) and display explicit contents with attribute loss (w/o $\\alpha_{ij}$ ). While image contents become clear and diverse with attribute loss, it is only with emotion confidence that we can effectively filter out emotion-agnostic semantics and generate images that evoke specific emotions (Ours).", + "bbox": [ + 496, + 674, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "6364", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/bfcb88e00f1b1601d7354f7fc0da355a52539c18e645964841cebd940b05f1bb.jpg", + "image_caption": [ + "(a) Emotion Transfer" + ], + "image_footnote": [], + "bbox": [ + 102, + 90, + 500, + 239 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/cb74f6f382a80c6fff396e98e9705b86e9e055e9777d1bd656cbfcfc83f2cb27.jpg", + "image_caption": [ + "(b) Emotion Fusion", + "Figure 8. Emotion creation. (a) transfers emotion representations (i.e., amusement, fear) to a series of neutral contents while (b) fuse two emotions (i.e., amusement-awe, amusement-fear) together, which may be helpful for emotional art design." + ], + "image_footnote": [], + "bbox": [ + 500, + 90, + 866, + 239 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4. Applications", + "text_level": 1, + "bbox": [ + 76, + 286, + 210, + 301 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Emotion Decomposition Emotions, serving as abstract concepts, pose a challenge for generative models to understand. Our method provides an opportunity to comprehend visual emotions by identifying the most relevant semantic contents for each emotion. To be specific, we visualize the semantics that are most closely aligned with our emotion representations in CLIP space. Each concept in Figure 7 (a), such as surfboard, bicycle and athletic field, is very likely to elicit excitement, where the corresponding images are presented in Figure 7 (b). Upon viewing such images, we identify the semantics and instinctively link them to specific emotions. These emotional concepts exhibit diversity, explicitness, and a strong capacity to evoke emotions. By decomposing visual emotions, we can not only generate emotional images with various semantics but also gain a deeper understanding of emotion evocation process. The results reveal the close relationship between emotions and semantics, in accordance with the psychological studies [3].", + "bbox": [ + 75, + 304, + 473, + 575 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Emotion Transfer Once we identify emotional contents, the next step is to explore how we can use it to create meaningful and compelling designs. In addition to emotional content, there are also neutral ones. As shown in Figure 8 (a), we combine the common neutral objects/scenes with emotional representations learned by our method. Surprisingly, we find that these representations effectively preserve emotional semantics and seamlessly integrate them with new concepts. Taking amusement as an example, it preserves several semantics including amusement park, picnic, princess, balloon and beautiful lanterns. In Table 3, our method is quantitatively compared with the state-of-the-art methods on emotion transfer task, specializing in room, where our method can well-preserve semantics and effectively elicit emotions. Crucially, these creations can evoke explicit and strong emotions across various neutral semantics, suggesting the potential of our method in image editing, image transfer and emotional art design.", + "bbox": [ + 75, + 587, + 470, + 859 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Emotion Fusion Additionally, we explore the possibilities of combining different emotion representations to evoke", + "bbox": [ + 76, + 869, + 470, + 902 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "multiple emotions. In Figure 8 (b), we combine amusement and awe (positive-positive) as well as amusement and fear (positive-negative), bringing some intriguing observations. In the combination of amusement and awe, we observe elements associated with amusement, such as toys, balloons, and ice-creams, alongside awe-inspiring elements like the blue sky, mountains, ocean, and city views. Particularly, one may feel both fear and amusement when viewing the funny and horrible face. When we fuse emotions, we are essentially combining their corresponding visual contents.", + "bbox": [ + 496, + 286, + 893, + 438 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 445, + 619, + 460 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Discussion In this paper, we introduce a new task named EICG and derive three specially designed metrics. We propose an emotion space and align it with the CLIP space, incorporating attribute loss and emotion confidence to ensure semantic clarity, semantic diversity and emotion fidelity. Experimental results indicate that our method surpasses the state-of-the-art text-to-image diffusion models both qualitatively and quantitatively, where user study confirms its superiority. Additionally, we outline potential applications for EICG and present some initial but promising results.", + "bbox": [ + 496, + 464, + 893, + 616 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Limitations Emotions can be evoked by various visual factors such as color, style and content. In this paper, we focus on investigating the most influential factor, i.e., contents. Moreover, the relationships between emotions and content is not strictly binary. In this paper, we simplify this connection by assuming content to be either emotional or emotion-agnostic. However, in reality, it is hard to assign rose to a single emotion category. White rose may evoke sadness while red rose can elicit amusement, making it hard to decide whether rose is emotional or not.", + "bbox": [ + 496, + 628, + 893, + 777 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments: This work was supported in parts by NSFC (62302312, U21B2023), Guangdong Science and Technology Program (2023A1515011440), DEGP Innovation Team (2022KCXTD025), Shenzhen Science and Technology Program (KQTD20210811090044003, RCJC20200714114435012), Guangdong Laboratory of Artificial Intelligence and Digital Economy (SZ) and Scientific Development Funds of Shenzhen University.", + "bbox": [ + 496, + 780, + 893, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "6365", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 174, + 106 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Damian Borth, Tao Chen, Rongrong Ji, and Shih-Fu Chang. Sentibank: large-scale ontology and classifiers for detecting sentiment and emotions in visual content. In Proceedings of the 21st ACM International Conference on Multimedia, pages 459-460, 2013. 2", + "[2] Damian Borth, Rongrong Ji, Tao Chen, Thomas Breuel, and Shih-Fu Chang. Large-scale visual sentiment ontology and detectors using adjective noun pairs. In Proceedings of the 21st ACM International Conference on Multimedia, pages 223-232, 2013. 3", + "[3] Tobias Brosch, Gilles Pourtois, and David Sander. The perception and categorisation of emotional stimuli: A review. Cognition and Emotion, 24(3):377-400, 2010. 2, 3, 8", + "[4] Linda Camras. Emotion: a psychoevolutionary synthesis, 1980. 2", + "[5] Tianlang Chen, Wei Xiong, Haitian Zheng, and Jiebo Luo. Image sentiment transfer. In Proceedings of the 28th ACM International Conference on Multimedia, pages 4407-4415, 2020. 3", + "[6] Domenico Consoli. A new concept of marketing: The emotional marketing. _BRAND_. Broad Research in Accounting, Negotiation, and Distribution, 1(1):52-59, 2010. 2", + "[7] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in Neural Information Processing Systems, 34:8780-8794, 2021. 2, 3", + "[8] Ziyi Dong, Pengxu Wei, and Liang Lin. Dreamartist: Towards controllable one-shot text-to-image generation via contrastive prompt-tuning. arXiv preprint arXiv:2211.11337, 2022. 3", + "[9] Oran Gafni, Adam Polyak, Oron Ashual, Shelly Sheynin, Devi Parikh, and Yaniv Taigman. Make-a-scene: Scene-based text-to-image generation with human priors. In European Conference on Computer Vision, pages 89-106. Springer, 2022. 3", + "[10] Rinon Gal, Yuval Alaluf, Yuval Atzmon, Or Patashnik, Amit H Bermano, Gal Chechik, and Daniel Cohen-Or. An image is worth one word: Personalizing text-to-image generation using textual inversion. arXiv preprint arXiv:2208.01618, 2022. 2, 3, 4, 6", + "[11] Leon A Gatys, Alexander S Ecker, and Matthias Bethge. Image style transfer using convolutional neural networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2414-2423, 2016. 3", + "[12] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial networks. Communications of the ACM, 63(11):139–144, 2020. 3", + "[13] Alan Hanjalic. Extracting moods from pictures and sounds: Towards truly personalized tv. IEEE Signal Processing Magazine, 23(2):90-100, 2006. 2", + "[14] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 770-778, 2016. 4", + "[15] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a" + ], + "bbox": [ + 76, + 114, + 470, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "two time-scale update rule converge to a local nash equilibrium. Advances in Neural Information Processing Systems, 30, 2017. 6", + "[16] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in Neural Information Processing Systems, 33:6840-6851, 2020. 2, 3", + "[17] Elaine Hsieh and Brenda Nicodemus. Conceptualizing emotion in healthcare interpreting: A normative approach to interpreters' emotion work. Patient Education and Counseling, 98(12):1474-1481, 2015. 2", + "[18] Yifei Huang, Sheng Qiu, Changbo Wang, and Chenhui Li. Learning representations for high-dynamic-range image color transfer in a self-supervised way. IEEE Transactions on Multimedia, 23:176-188, 2020. 3", + "[19] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4401-4410, 2019. 3", + "[20] Diederik P Kingma and Max Welling. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114, 2013. 3", + "[21] Nupur Kumari, Bingliang Zhang, Richard Zhang, Eli Shechtman, and Jun-Yan Zhu. Multi-concept customization of text-to-image diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1931-1941, 2023. 3", + "[22] Yann LeCun, Sumit Chopra, Raia Hadsell, M Ranzato, and Fujie Huang. A tutorial on energy-based learning. Predicting Structured Data, 1(0), 2006. 3", + "[23] Joonwooan Lee and EunJong Park. Fuzzy similarity-based emotional classification of color images. IEEE Transactions on Multimedia, 13(5):1031-1039, 2011. 3", + "[24] Wentong Liao, Kai Hu, Michael Ying Yang, and Bodo Rosenhahn. Text to image generation with semantic-spatial aware gan. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18187-18196, 2022. 3", + "[25] Da Liu, Yaxi Jiang, Min Pei, and Shiguang Liu. Emotional image color transfer via deep learning. Pattern Recognition Letters, 110:16-22, 2018. 3", + "[26] Jana Machajdik and Allan Hanbury. Affective image classification using features inspired by psychology and art theory. In Proceedings of the 18th ACM International Conference on Multimedia, pages 83-92, 2010. 3", + "[27] Saif Mohammad and Svetlana Kiritchenko. Wikiart emotions: An annotated dataset of emotions evoked by art. In Proceedings of the eleventh International Conference on Language Resources and Evaluation, 2018. 3", + "[28] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741, 2021. 3", + "[29] Magnus Oskarsson. Robust image-to-image color transfer using optimal inlier maximization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 786-795, 2021. 3" + ], + "bbox": [ + 501, + 92, + 893, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "6366", + "bbox": [ + 482, + 945, + 516, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[30] Kuan-Chuan Peng, Tsuhan Chen, Amir Sadovnik, and Andrew C Gallagher. A mixed bag of emotions: Model, predict, and transfer emotion distributions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 860-868, 2015. 2, 3", + "[31] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763, 2021. 2, 3, 5", + "[32] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 1 (2):3, 2022. 3", + "[33] Harsh Rangwani, Lavish Bansal, Kartik Sharma, Tejan Karmali, Varun Jampani, and R Venkatesh Babu. Noisytwins: Class-consistent and diverse image generation through stylegans. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5987-5996, 2023. 3, 4", + "[34] Tianrong Rao, Xiaoxu Li, and Min Xu. Learning multi-level deep representations for image emotion classification. Neural Processing Letters, pages 1-19, 2016. 2", + "[35] Tianrong Rao, Xiaoxu Li, and Min Xu. Learning multi-level deep representations for image emotion classification. Neural Processing Letters, 51:2043-2061, 2020. 3", + "[36] Erik Reinhard, Michael Adhikhmin, Bruce Gooch, and Peter Shirley. Color transfer between images. IEEE Computer Graphics and Applications, 21(5):34-41, 2001. 3", + "[37] Danilo Rezende and Shakir Mohamed. Variational inference with normalizing flows. In International Conference on Machine Learning, pages 1530-1538, 2015. 3", + "[38] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10684-10695, 2022. 2, 3, 4, 6", + "[39] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22500-22510, 2023. 2, 3, 4, 6", + "[40] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in Neural Information Processing Systems, 35:36479-36494, 2022. 3", + "[41] Shikun Sun, Jia Jia, Haozhe Wu, Zijie Ye, and Junliang Xing. Msnet: A deep architecture using multi-sentiment semantics for sentiment-aware image style transfer. In ICASSP 2023-2023 IEEE International Conference on Acoustics, Speech and Signal Processing, pages 1-5. IEEE, 2023. 2, 3", + "[42] Zhentao Tan, Mengelei Chai, Dongdong Chen, Jing Liao, Qi Chu, Bin Liu, Gang Hua, and Nenghai Yu. Diverse semantic image synthesis via probability distribution modeling. In" + ], + "bbox": [ + 78, + 90, + 468, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7962-7971, 2021. 6", + "[43] Guy Tevet, Brian Gordon, Amir Hertz, Amit H Bermano, and Daniel Cohen-Or. Motionclip: Exposing human motion generation to clip space. In European Conference on Computer Vision, pages 358–374. Springer, 2022. 4", + "[44] Xiaohui Wang, Jia Jia, and Lianhong Cai. Affective image adjustment with a single word. The Visual Computer, 29: 1121-1133, 2013. 3", + "[45] Zhizhong Wang, Lei Zhao, and Wei Xing. *Styloid diffusion: Controllable disentangled style transfer via diffusion models*. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7677-7689, 2023. 3", + "[46] Yuxiang Wei, Yabo Zhang, Zhilong Ji, Jinfeng Bai, Lei Zhang, and Wangmeng Zuo. Elite: Encoding visual concepts into textual embeddings for customized text-to-image generation. arXiv preprint arXiv:2302.13848, 2023. 3", + "[47] Shuchen Weng, Peixuan Zhang, Zheng Chang, Xinlong Wang, Si Li, and Boxin Shi. Affective image filter: Reflecting emotions from text to images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10810-10819, 2023. 2, 3", + "[48] Ali Yadollahi, Ameneh Gholipour Shahraki, and Osmar R Zaiane. Current state of text sentiment analysis from opinion to emotion mining. ACM Computing Surveys (CSUR), 50(2): 1-33, 2017. 2", + "[49] Chuan-Kai Yang and Li-Kai Peng. Automatic mood-transferring between color images. IEEE Computer Graphics and Applications, 28(2):52-61, 2008. 3", + "[50] Jufeng Yang, Dongyu She, Yu-Kun Lai, Paul L Rosin, and Ming-Hsuan Yang. Weakly supervised coupled networks for visual sentiment analysis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7584-7592, 2018. 2, 4", + "[51] Jingyuan Yang, Xinbo Gao, Leida Li, Xiumei Wang, and Jinshan Ding. Solver: Scene-object interrelated visual emotion reasoning network. IEEE Transactions on Image Processing, 30:8686-8701, 2021. 2, 3, 4", + "[52] Jingyuan Yang, Jie Li, Xiumei Wang, Yuxuan Ding, and Xinbo Gao. Stimuli-aware visual emotion analysis. IEEE Transactions on Image Processing, 30:7432-7445, 2021. 3", + "[53] Jingyuan Yang, Qirui Huang, Tingting Ding, Dani Lischinski, Danny Cohen-Or, and Hui Huang. Emoset: A large-scale visual emotion dataset with rich attributes. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 20383–20394, 2023. 2, 4, 5", + "[54] Shuai Yang, Liming Jiang, Ziwei Liu, and Chen Change Loy. Pastiche master: Exemplar-based high-resolution portrait style transfer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7693-7702, 2022. 3", + "[55] Chenrui Zhang and Yuxin Peng. Stacking vae and gan for context-aware text-to-image generation. In 2018 IEEE Fourth International Conference on Multimedia Big Data (BigMM), pages 1-5. IEEE, 2018. 3", + "[56] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In" + ], + "bbox": [ + 503, + 92, + 890, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "6367", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3836-3847, 2023. 2, 3, 4", + "[57] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 586-595, 2018. 6", + "[58] Wei Zhang, Xuanyu He, and Weizhi Lu. Exploring discriminative representations for image emotion recognition with cnns. IEEE Transactions on Multimedia, 22(2):515-523, 2019. 3", + "[59] Minfeng Zhu, Pingbo Pan, Wei Chen, and Yi Yang. Dm-gan: Dynamic memory generative adversarial networks for text-to-image synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5802-5810, 2019. 3", + "[60] Siqi Zhu, Chunmei Qing, Canqiang Chen, and Xiangmin Xu. Emotional generative adversarial network for image emotion transfer. Expert Systems with Applications, 216:119485, 2023. 3" + ], + "bbox": [ + 78, + 90, + 470, + 372 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "6368", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2024/EmoGen_ Emotional Image Content Generation with Text-to-Image Diffusion Models/a0b8035c-47d5-4f15-8dd7-134ed07d94b6_model.json b/2024/EmoGen_ Emotional Image Content Generation with Text-to-Image Diffusion Models/a0b8035c-47d5-4f15-8dd7-134ed07d94b6_model.json new file mode 100644 index 0000000000000000000000000000000000000000..1cf0e8c890ad554c1950491bc8ca6c0bfd96e280 --- /dev/null +++ b/2024/EmoGen_ Emotional Image Content Generation with Text-to-Image Diffusion Models/a0b8035c-47d5-4f15-8dd7-134ed07d94b6_model.json @@ -0,0 +1,2202 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.0, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.131, + 0.756, + 0.177 + ], + "angle": 0, + "content": "EmoGen: Emotional Image Content Generation with Text-to-Image Diffusion Models" + }, + { + "type": "text", + "bbox": [ + 0.324, + 0.204, + 0.652, + 0.223 + ], + "angle": 0, + "content": "Jingyuan Yang, Jiawei Feng, Hui Huang*" + }, + { + "type": "text", + "bbox": [ + 0.402, + 0.224, + 0.568, + 0.24 + ], + "angle": 0, + "content": "Shenzhen University" + }, + { + "type": "text", + "bbox": [ + 0.244, + 0.242, + 0.721, + 0.258 + ], + "angle": 0, + "content": "{jingyuanyang.jyy, fengjiawei0909, hhzhiyan}@gmail.com" + }, + { + "type": "image", + "bbox": [ + 0.107, + 0.265, + 0.289, + 0.414 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.3, + 0.265, + 0.48, + 0.412 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.492, + 0.265, + 0.671, + 0.412 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.685, + 0.266, + 0.865, + 0.412 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.107, + 0.419, + 0.289, + 0.567 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.3, + 0.419, + 0.48, + 0.567 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.492, + 0.419, + 0.671, + 0.566 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.685, + 0.419, + 0.865, + 0.566 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.58, + 0.894, + 0.609 + ], + "angle": 0, + "content": "Figure 1. Emotional Image Content Generation (EICG). Given an emotion category, our network produces images that exhibit unambiguous meanings (semantic-clear), reflect the intended emotion (emotion-faithful) and incorporate varied semantics (semantic-diverse)." + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.621, + 0.314, + 0.637 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.643, + 0.474, + 0.87 + ], + "angle": 0, + "content": "Recent years have witnessed remarkable progress in image generation task, where users can create visually astonishing images with high-quality. However, existing text-to-image diffusion models are proficient in generating concrete concepts (dogs) but encounter challenges with more abstract ones (emotions). Several efforts have been made to modify image emotions with color and style adjustments, facing limitations in effectively conveying emotions with fixed image contents. In this work, we introduce Emotional Image Content Generation (EICG), a new task to generate semantic-clear and emotion-faithful images given emotion categories. Specifically, we propose an emotion space and construct a mapping network to align it with the powerful Contrastive Language-Image Pre-training (CLIP) space, providing a concrete interpretation of abstract emo" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.623, + 0.895, + 0.759 + ], + "angle": 0, + "content": "tions. Attribute loss and emotion confidence are further proposed to ensure the semantic diversity and emotion fidelity of the generated images. Our method outperforms the state-of-the-art text-to-image approaches both quantitatively and qualitatively, where we derive three custom metrics, i.e., emotion accuracy, semantic clarity and semantic diversity. In addition to generation, our method can help emotion understanding and inspire emotional art design. Project page: https://vcc.tech/research/2024/EmoGen." + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.783, + 0.633, + 0.799 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.549, + 0.808, + 0.845, + 0.822 + ], + "angle": 0, + "content": "\"What I cannot create, I do not understand.\"" + }, + { + "type": "text", + "bbox": [ + 0.763, + 0.823, + 0.892, + 0.838 + ], + "angle": 0, + "content": "Richard Feynman" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Emotions, often elusive yet profoundly influential, shape our actions, foster connections, and spark passions. With the prevalence of social medias, users tend to share specially crafted images to express their feelings. Aiming to find out" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.888, + 0.221, + 0.9 + ], + "angle": 0, + "content": "*Corresponding author" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6358" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.182 + ], + "angle": 0, + "content": "people's emotional responses towards different stimuli, Visual Emotion Analysis (VEA) is an intriguing yet challenging task in computer vision [34, 50, 51]. Recent years have witnessed rapid development in this field, bringing potential applications such as opinion mining [48], market advertising [6] and mental healthcare [17]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.185, + 0.468, + 0.336 + ], + "angle": 0, + "content": "Thanks to the advent of diffusion models [7, 16, 38], unprecedented progress has been made in text-to-image generation, where users can generate high-quality images with crafted prompts or personalized objects [10, 39, 56]. Existing text-to-image diffusion models, are often excel in generating concrete concepts (e.g., cat, house, mountain) but face limitations when tasked with more abstract ones (e.g., amusement, anger, sadness). In reality, however, photographic works are not necessarily targeted on specific entities, but are often composed to convey certain feelings." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.338, + 0.468, + 0.549 + ], + "angle": 0, + "content": "A natural question arises: What if machines could create images that not only please our eyes but also touch our hearts? Generating emotions is very challenging. Emotions are abstract while images are concrete, leaving the affective gap [13] hard to surmount. To bridge the gap, several efforts have been made to modify visual emotions by adjusting colors and styles, i.e., image emotion transfer [30, 41, 47]. These methods, however, meet difficulties in evoking emotions correctly and significantly, i.e., \\(29\\%\\) emotion accuracy [47], as fixed image contents limit emotional variations. Moreover, we cannot generate emotional images solely from colors and styles. What truly triggers emotion? Psychological studies show that visual emotions are often evoked by specific semantics [1, 3, 4]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.551, + 0.468, + 0.807 + ], + "angle": 0, + "content": "In this paper, we propose Emotional Image Content Generation (EICG), a new task to generate semantically clear and emotionally faithful visual contents conditioned on a given emotion category, as shown in Figure 1. Semantic clarity demands an unambiguous representation of visual contents, while emotion faithfulness entails generating images evoke the intended emotions. Contrastive Language-Image Pre-training (CLIP) [31] is a large-scale vision-language model with rich semantics. However, we observe in Figure 2 that CLIP space can not well capture emotional relationships. Therefore, we introduce an emotion space, which groups similar emotions together while keeping dissimilar ones apart. While emotion space excels in representing emotions, CLIP space exhibits a powerful semantic structure. To align emotion space with CLIP space, we propose a mapping network, interpreting abstract emotions with concrete semantics." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.811, + 0.468, + 0.899 + ], + "angle": 0, + "content": "EmoSet [53] is a recently proposed large-scale visual emotion dataset with rich attributes. The Latent Diffusion Model (LDM) loss [38] is often utilized to optimize concrete entities with single and explicit semantics, posing a challenge in capturing the diversity within each emotion. To address this, we introduce an attribute loss to ensure seman" + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.091, + 0.696, + 0.238 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.558, + 0.24, + 0.645, + 0.25 + ], + "angle": 0, + "content": "(a) Emotion Space" + }, + { + "type": "image", + "bbox": [ + 0.699, + 0.091, + 0.887, + 0.238 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.758, + 0.24, + 0.829, + 0.249 + ], + "angle": 0, + "content": "(b)CLIP Space" + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.266, + 0.892, + 0.308 + ], + "angle": 0, + "content": "Figure 2. Despite (b) CLIP space demonstrates a powerful semantic structure, it struggles to effectively capture emotional relationships within (a), the proposed emotion space." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.325, + 0.892, + 0.385 + ], + "angle": 0, + "content": "tic clarity and diversity, by leveraging the attribute labels in EmoSet. Recognizing that not all objects are affective, emotion confidence is further proposed to ensure the emotion fidelity of the generated contents." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.39, + 0.893, + 0.692 + ], + "angle": 0, + "content": "To estimate the generation quality of EICG, three evaluation metrics are specially designed: emotion accuracy, semantic clarity and semantic diversity. As EICG aims to create emotional contents, we design emotion accuracy to measure the alignment between intended and perceived emotions in the generated images. People are prone to evoke emotions only when the contents are easily recognizable. Thus we propose semantic clarity to assess the unambiguity of the generated image content. Additionally, in view of the assorted emotion stimuli, we devise semantic diversity to quantify the content richness under each emotion. We evaluate our method through both qualitative and quantitative analyses, surpassing the state-of-the-art text-to-image generation approaches across five metrics. Ablation studies are performed to verify the network design, and user studies are conducted to resonate our method with human viewers. Besides generation task, our method can also be applied to decompose emotion concepts, transfer emotional contents and fuse different emotions, which may be helpful to understand emotions and create emotional art design." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.698, + 0.75, + 0.712 + ], + "angle": 0, + "content": "In summary, our contributions are:" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.717, + 0.892, + 0.776 + ], + "angle": 0, + "content": "- We introduce Emotional Image Content Generation, a novel task to generate emotion-faithful and semantic-clear image contents. We also derive three custom metrics to estimate the generation performance." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.778, + 0.892, + 0.837 + ], + "angle": 0, + "content": "- We develop a mapping network to align the proposed emotion space to the powerful CLIP space, where attribute loss and emotion confidence are further designed to ensure the semantic richness and emotion fidelity." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.838, + 0.892, + 0.898 + ], + "angle": 0, + "content": "- We evaluate our method against the state-of-the-art text-to-image approaches and demonstrate our superiority. Potential applications are exhibited for emotion understanding and emotional art design." + }, + { + "type": "list", + "bbox": [ + 0.5, + 0.717, + 0.892, + 0.898 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.515, + 0.956 + ], + "angle": 0, + "content": "6359" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.078, + 0.091, + 0.216, + 0.105 + ], + "angle": 0, + "content": "2. Related work" + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.115, + 0.303, + 0.132 + ], + "angle": 0, + "content": "2.1. Visual Emotion Analysis" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.14, + 0.473, + 0.564 + ], + "angle": 0, + "content": "Researchers have been involved in VEA for over two decades, ranging from early traditional approaches [2, 23, 26] to recent deep learning ones [35, 51, 52, 58]. Given the inherent abstractness and complexity of visual emotion, researchers aim to identify the most influential elements, which range from low-level features like color, texture and style [23, 26, 35, 58] to high-level semantics [2, 35, 51, 52, 58]. Lee et al. [23] propose a scheme to evaluate emotional response from color images by reasoning the prototypical color for each emotion and the input images. As a milestone, Machajdik et al. [26] extract representative low-level features in composition, including color and texture, to predict visual emotions. Besides low-level features, Borth et al. [2] propose Adjective-Noun Pair (ANP) and build a visual concept detector named Sentibank. With the help of deep learning techniques, Rao et al. [35] construct MldrNet to extract emotional clues from pixel-level, aesthetic-level and semantic-level. To form a more discriminative emotional representation, Zhang et al. [58] integrate high-level contents and low-level styles. Yang et al. propose network to mine emotions from multiple objects [52] as well as object-scene correlations [51]. Existing work often treat VEA as a classification task, i.e., input an image and predict the emotion within it. Can we reverse this process? In other words, can we generate an image targeting on the given emotion word? Only by creating emotional images can we demonstrate the significance of visual elements, leading to a better understanding of emotions." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.575, + 0.315, + 0.59 + ], + "angle": 0, + "content": "2.2. Text-to-Image Generation" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.6, + 0.473, + 0.903 + ], + "angle": 0, + "content": "Text-to-image generation aims to convert textual descriptions into corresponding realistic images. Existing generative models can be grouped into GANs [12, 24, 59], VAEs [9, 20, 55], flow-based [37], energy-based [22] and diffusion-based [7, 16, 38, 39, 56]. Diffusion models are witnessed impressive and rapid progress in recent years, where methods like GLIDE [28], DALLE2 [32], Image [40] can generate diverse, photo-realistic and high-quality images. Notably, Stable diffusion [38] is one of the most popular diffusion models, owing to its stable training and the capability for fine-grained control, supported by an active user community. For customized generation, several diffusion-based text-to-image methods are introduced, where methods vary from learning a new embedding [8, 10] and finetuning the network parameters [21, 39, 46]. Textual inversion [10] and DreamArtist [8] learn new concepts with a few user-provided images in the word embedding space, without further training on diffusion models. While DreamBooth [39] finetunes all the parameters to learn a new concept, Custom diffusion [21] only updates the key" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.245 + ], + "angle": 0, + "content": "and value parameters in the cross attention layers. Further, ELITE [46] speeds up the running time with accurate generation results by adopting a global and local mapping network. Existing text-to-images models are capable of generating concrete concepts [7, 24, 56], or personalized ones [10, 21, 39], but face difficulties in generating more abstract ones. In reality, photographic works are not necessarily composed of targeted concepts, but often aim to convey specific feelings. Thus, how to generate emotion-evoking images remains a pressing and critical challenge." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.254, + 0.727, + 0.27 + ], + "angle": 0, + "content": "2.3. Image Emotion Transfer" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.278, + 0.895, + 0.656 + ], + "angle": 0, + "content": "Image style transfer [11] aims to render the semantic content under different styles, producing visually stunning results [19, 33, 45, 54]. Similarly, image color transfer [36] seeks to adjust and harmonize the color characteristics of one image to match another [18, 29]. Specifically, color and style choices can strongly influence the emotions of an image [27]. By adjusting low-level visual elements, image emotion transfer aims to modify the emotional tone of the input image, including the color-based methods [5, 25, 30, 44, 49, 60] and the style-based ones [41, 47]. Yang and Peng et al. [49] makes the first attempt to transfer image colors. Wang et al. [44] present a system to modify the image color according to a given emotion word, and Liu et al. [25] further advance it with deep learning techniques. Peng et al. [30] introduce a new approach to alter the emotion of an input image by guiding its color and texture under the target image. More recently, to reflect emotions in styles, Sun et al. [41] and Weng et al. [47] bring promising results on emotion-aware image style transfer. Nevertheless, the alteration of visual emotions through colors and styles is limited due to fixed content, resulting in subtle emotional changes, i.e., \\(29\\%\\) emotion accuracy in [47]. Psychological studies suggest that visual emotions can be elicited by specific semantics [3]. Thus, we propose a novel method to generate emotional image contents with clear semantics." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.671, + 0.593, + 0.687 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.697, + 0.724, + 0.712 + ], + "angle": 0, + "content": "3.1. Emotion Representation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.895, + 0.901 + ], + "angle": 0, + "content": "Emotion Space EICG is a challenging task, which requires both semantic clarity and emotion fidelity. How to generate an image with distinct and emotional semantics? CLIP [31] is developed to align image and text modalities, where semantically related features are located in close proximity to each other. While CLIP shows impressive semantic representation capabilities, it struggles to effectively capture emotional relationships. As demonstrated in Figure 2, we can observe that sample points with emotional similarities are distantly separated within the CLIP space due to their differing semantics, e.g., toy, amusement park and Christmas tree. To better depict emotional relation" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.957 + ], + "angle": 0, + "content": "6360" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.129, + 0.088, + 0.843, + 0.256 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.262, + 0.895, + 0.293 + ], + "angle": 0, + "content": "Figure 3. Training process of our network. Emotion representation (stage 1) learns a well-behaved emotion space and emotion content generation (stage 2) maps this space to CLIP space, aiming to generate image contents with emotion fidelity, semantic clarity and diversity." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.305, + 0.473, + 0.442 + ], + "angle": 0, + "content": "ships, we introduce the emotion space, a latent space that clusters similar emotions together while keeping dissimilar ones apart. EmoSet [53] is a large-scale dataset with rich attributes, where each image is labeled with an emotion. Using aligned image-emotion pairs, we construct an encoder \\(\\varphi\\) with ResNet-50 [14] to capture emotion representations. To train the encoder, we devise an emotion loss by implementing the widely-used Cross-Entropy (CE) loss, following the previous work [50, 51]:" + }, + { + "type": "equation", + "bbox": [ + 0.124, + 0.447, + 0.47, + 0.488 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {e m o} = - \\sum_ {i = 1} ^ {C} y _ {e m o} \\log \\frac {\\exp (\\varphi (x , i))}{\\sum_ {i = 1} ^ {C} \\exp (\\varphi (x , i))}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.494, + 0.47, + 0.569 + ], + "angle": 0, + "content": "where \\( x \\) represents the input image, \\( y_{emo} \\) denotes the emotion label and \\( C \\) stands for the total number of emotion categories. Once the loss function converges, emotion space is established. Parameters in emotion encoder remain fixed in the following emotional content generation process." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.569, + 0.471, + 0.691 + ], + "angle": 0, + "content": "During inference, each emotion cluster is represented by a Gaussian distribution with learned parameters, i.e., mean and standard deviation. For example, when taking amusement as input, we randomly sample a data point from corresponding Gaussian distribution to serve as its emotion representation, as shown in Figure 3. We have confirmed that Gaussian distribution suits emotion clusters well and the random sampling process induces diversity to EICG." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.697, + 0.353, + 0.712 + ], + "angle": 0, + "content": "3.2. Emotional Content Generation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.72, + 0.47, + 0.84 + ], + "angle": 0, + "content": "Mapping Network While emotion space is emotionally separable, CLIP space captures rich semantics. Existing text-to-image models entail clear and specific semantics as input, making CLIP space indispensable in the generation process. Consequently, establishing the mapping between emotion space and CLIP space becomes a crucial challenge. Intuitively, we attempt to build the mapping network using fully connected layers, following previous work [33, 43]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.471, + 0.902 + ], + "angle": 0, + "content": "However, as depicted in Figure 2, clustered feature points in the emotion space are expected to disperse in the CLIP space to capture diverse semantics. Therefore, we utilize a Multilayer Perceptron (MLP) to build the mapping" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.305, + 0.895, + 0.457 + ], + "angle": 0, + "content": "network, incorporating non-linear operations, i.e., RELU, to facilitate the separation process. The non-linear projection \\( F \\) is succeeded by a CLIP text transformer \\( t_{\\theta} \\), yielding textual embedding for U-Net. The end-token embedding of the transformer's output is passed through a fully-connected layer, producing the CLIP text feature. Particularly, to better preserve the prior knowledge in the CLIP space, parameters in the transformer and linear projection are kept frozen, while parameters in non-linear projection are learned, as depicted in Figure 3." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.473, + 0.895, + 0.519 + ], + "angle": 0, + "content": "Attribute Loss Existing text-to-image diffusion models often employ Latent Diffusion Model (LDM) loss [38] for optimization process [10, 39, 56]:" + }, + { + "type": "equation", + "bbox": [ + 0.515, + 0.523, + 0.892, + 0.55 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {L D M} = \\mathbb {E} _ {z, x, \\epsilon , t} \\left[ \\| \\epsilon - \\epsilon_ {\\theta} (z _ {t}, t, t _ {\\theta} (F (\\varphi (x)))) \\| _ {2} ^ {2} \\right], \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.554, + 0.894, + 0.584 + ], + "angle": 0, + "content": "where \\(\\epsilon\\) represents the added noise, \\(\\epsilon_{\\theta}\\) denotes the denoising network and \\(z_{t}\\) indicates the latent noised to time \\(t\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.584, + 0.893, + 0.87 + ], + "angle": 0, + "content": "In these cases, target concepts typically involve concrete entities (e.g., dog, car, flower) or personalized objects (e.g., someone's corgi). These concepts often exhibit consistency on semantic level and share certain similarities on pixel level. However, emotions are abstract concepts, where multiple semantics coexist under one specific category. Learning emotions solely with LDM loss may pose some challenges. For one thing, each emotion might collapse to a specific semantic point, e.g., amusement collapsing to amusement park, losing intra-class diversity. In reality, semantics within one emotion are diverse, where single point cannot fully capture. Moreover, since LDM loss is designed to reconstruct the input image, it primarily focuses on learning and preserving pixel-level commonalities such as color and texture. In Figure 4 (a), with LDM loss alone, CLIP embedding for amusement is prone to be colorful, without exhibiting explicit and diverse semantics. We can conclude that it is hard to achieve robust emotion representations in CLIP space by implementing LDM loss alone." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.871, + 0.894, + 0.902 + ], + "angle": 0, + "content": "In the pursuit of clear and diverse contents, semantics guidance is essential for the generation process. Thanks to" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "6361" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.097, + 0.09, + 0.448, + 0.263 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.268, + 0.47, + 0.31 + ], + "angle": 0, + "content": "Figure 4. Motivation for loss function design. Compare to (a) LDM loss alone, (b) attribute loss enhances semantic clarity while (c) emotion confidence ensures emotion accuracy." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.324, + 0.473, + 0.507 + ], + "angle": 0, + "content": "the rich attribute annotations in EmoSet, we select the mid-level attributes, i.e., object class and scene type to guide the generation process. With this semantic guidance, we formulate an attribute loss to guarantee that the generated image contents possess clear and diverse semantics. For clarity, emotions are easily triggered in people only when visual contents are represented in an unambiguous manner. Considering the varied emotional stimuli in reality, attribute loss guides the network to learn multiple semantics under one specific emotion. Our attribute loss is devised on CLIP space, by calculating the cosine similarities \\( f(\\cdot) \\) and optimizing a symmetric CE loss over the similarity scores [31]:" + }, + { + "type": "equation", + "bbox": [ + 0.091, + 0.517, + 0.47, + 0.562 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {a t t r} = - \\sum_ {j = 1} ^ {K} y _ {a t t r} \\log \\frac {\\exp (f (v _ {e m o} , \\tau_ {\\theta} (a _ {j})))}{\\sum_ {j = 1} ^ {K} \\exp (f (v _ {e m o} , \\tau_ {\\theta} (a _ {j})))}, \\quad (3)\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.205, + 0.584, + 0.47, + 0.617 + ], + "angle": 0, + "content": "\\[\nf (p, q) = \\frac {p \\cdot q}{\\| p \\| \\| q \\|}, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.625, + 0.47, + 0.748 + ], + "angle": 0, + "content": "where \\( a_{j} \\) denotes the \\( j \\) member in the attribute set, \\( \\tau_{\\theta} \\) represents the text encoder, \\( v_{emo} \\) implies the learned CLIP embedding and \\( K \\) indicates the total number of the attributes. With the attribute loss, each sample point is converging towards the correct semantic and distancing itself from the incorrect ones. Through the combination of attribute loss and LDM loss, we can effectively map each emotion to clear and diverse semantics, as demonstrated in Figure 4 (b)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Emotion Confidence However, it is worth noting that some of the semantics in Figure 4 (b) appear emotionally neutral, e.g., plant and tree. Since attributes are annotated objectively, not all the attributes in EmoSet are emotional. Therefore, we propose emotion confidence to measure the correlations between emotions and semantic attributes. Initially, we gather all images associated with attribute \\( j \\) in EmoSet and send them to a pre-trained emotion classifier. Each image is predicted as an emotion vector \\( p(\\cdot) \\) and we" + }, + { + "type": "image", + "bbox": [ + 0.521, + 0.089, + 0.872, + 0.208 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.213, + 0.892, + 0.242 + ], + "angle": 0, + "content": "Figure 5. Illustration of emotion confidence. Each (a) attribute is represented by (b) a distribution of confidence on eight emotions." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.252, + 0.892, + 0.299 + ], + "angle": 0, + "content": "sum all images up to get the emotional distribution \\(d_{j}\\) for attribute \\(j\\). Each emotion \\(i\\) within this distribution is assigned a corresponding emotion confidence \\(\\alpha_{ij}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.614, + 0.306, + 0.892, + 0.349 + ], + "angle": 0, + "content": "\\[\n\\alpha_ {i j} = \\frac {1}{N _ {j}} \\sum_ {n = 1} ^ {N _ {j}} p (x _ {n}, i), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.355, + 0.892, + 0.506 + ], + "angle": 0, + "content": "where \\( x_{n} \\) represents the input image and \\( N_{j} \\) denotes the total image number in attribute \\( j \\). We further illustrate the above process in Figure 5 with visual representations. When mountain snowy appears, people are more likely to experience awe and cemetery often elicits sadness. In contrast, the presence of tree in every emotion category suggests its lack of emotional specificity. Some attributes are emotion-related while others are not, which can be beneficial for generating emotional contents. We then use emotion confidence to balance between LDM loss and attribute loss:" + }, + { + "type": "equation", + "bbox": [ + 0.58, + 0.515, + 0.891, + 0.532 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\left(1 - \\alpha_ {i j}\\right) \\mathcal {L} _ {L D M} + \\alpha_ {i j} \\mathcal {L} _ {\\text {a t t r}}, \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.54, + 0.893, + 0.706 + ], + "angle": 0, + "content": "where \\(i\\) represents the emotion category \\(y_{emo}\\) and \\(j\\) denotes the attribute type \\(y_{att}\\). The greater the emotion confidence \\(\\alpha_{ij}\\) is, the stronger the impact attribute \\(j\\) has on the specific emotion \\(i\\). Low confidence suggests a weak connection between the attribute and emotion, signaling that the network should learn more from the pixel-wise LDM loss. When higher confidence occurs, the network should prioritize the semantic meaning of the image, i.e., the attribute loss. With this design, our network can adapt to a wide range of cases, generating image contents that are both semantically explicit and emotionally faithful, as shown in Figure 4 (c)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.717, + 0.634, + 0.734 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.742, + 0.718, + 0.758 + ], + "angle": 0, + "content": "4.1. Dataset and Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Dataset EmoSet [53] is a large-scale visual emotion dataset with rich attributes, comprising a total of 118,102 images. To investigate the connections between emotions and specific contents, we create a subset from EmoSet by preserving images with object/scene labels. Each image is labeled with both emotion and attribute labels, guiding the optimization process of emotion loss and attribute loss. Notably, the wide range of attribute labels assures for learning diverse and representative emotional contents." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6362" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.084, + 0.089, + 0.895, + 0.472 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.097, + 0.477, + 0.871, + 0.492 + ], + "angle": 0, + "content": "Figure 6. Qualitative comparisons with the state-of-the-art text-to-image generation approaches and ablation studies of our method." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.499, + 0.473, + 0.803 + ], + "angle": 0, + "content": "Evaluation Metrics To comprehensively evaluate the performance of different methods on EICG task, we utilize commonly used metrics (FID, LPIPS) and design some specific ones (Emo-A, Sem-C, Sem-D). 1) FID: Frechet Inception Distance (FID) [15] quantifies the distribution distance between generated and real images, providing an estimate of image fidelity. 2) LPIPS: Similar to [42], we employ LPIPS [57] to assess the overall image diversity, with higher values indicating better performance. 3) Emo-A: Since EICG aims at creating emotion-evoking images, we design emotion accuracy to assess the emotional alignment between the targeted emotions and the generated images. 4) Sem-C: People are easily to evoke emotions under recognizable contents. We thus introduce semantic clarity to assess the explicitness of generated image contents. 5) Sem-D: Emotions are complex, where each can be triggered by multiple factors. To cover a diverse range of potential scenes or objects, we derive semantic diversity to estimate the content richness associated with each emotion. For more details, please refer to the supplementary materials." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.816, + 0.218, + 0.832 + ], + "angle": 0, + "content": "4.2. Comparisons" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.47, + 0.903 + ], + "angle": 0, + "content": "As our method is the first attempt in EICG, we compare it with the most relevant and state-of-the-art text-to-image generation techniques: Stable diffusion [38], Textual inversion [10] and Dreambooth [39]. While Stable diffusion is" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.498, + 0.892, + 0.525 + ], + "angle": 0, + "content": "Table 1. Comparisons with the state-of-the-art methods and ablation studies on emotion generation task, involving five metrics." + }, + { + "type": "table", + "bbox": [ + 0.506, + 0.53, + 0.885, + 0.649 + ], + "angle": 0, + "content": "
MethodFID ↓LPIPS ↑Emo-A ↑Sem-C ↑Sem-D ↑
Stable Diffusion [38]44.050.68770.77%0.6080.0199
Textual Inversion [10]50.510.70274.87%0.6050.0282
DreamBooth [39]46.890.66170.50%0.6140.0178
Ours41.600.71776.25%0.6330.0335
w/o F57.540.71371.12%0.6150.0261
w/o Lattr51.130.70765.75%0.5920.0270
w/o αij43.300.71474.88%0.5910.0263
" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.655, + 0.892, + 0.687 + ], + "angle": 0, + "content": "a general image generation pipeline, Textual inversion and Dreambooth specialize in customized image generation." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.704, + 0.893, + 0.901 + ], + "angle": 0, + "content": "Qualitative Comparisons In Figure 6, our method is qualitatively compared with the state-of-the-art methods across three emotion categories, i.e., awe, anger and contentment. Generation results of the rest five emotions can be found in the supplementary materials. Take awe as an example, all the three compared methods tend to produce images with dense textures and dim colors, which suggests that representations for each emotion may collapse to a single feature point. For anger and contentment, both Stable diffusion and Dreambooth distort the visual representations, e.g., tiger and bicycle, and generate some contents with ambiguous semantics. Though Textual inversion preserves some semantic fidelity, it generates emotion-agnostic" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6363" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.089, + 0.473, + 0.132 + ], + "angle": 0, + "content": "Table 2. User preference study. The numbers indicate the percentage of participants who prefer our results over those compared methods, given the same emotion category as input." + }, + { + "type": "table", + "bbox": [ + 0.08, + 0.137, + 0.47, + 0.202 + ], + "angle": 0, + "content": "
MethodImage fidelity ↑Emotion faithfulness ↑Semantic diversity ↑
Stable Diffusion67.86±15.08%73.66±11.80%87.88±9.64%
Textual Inversion79.91±16.92%72.75±16.90%85.66±10.51%
DreamBooth77.23±14.00%80.79±8.64%81.68±17.06%
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.21, + 0.473, + 0.376 + ], + "angle": 0, + "content": "contents such as shoes and cars. Since these methods are crafted to learn customized concepts, challenges may arise when handling complex and diverse emotional images. Rather than generating plants and trees, our method can provide diverse and emotion-evoking image contents for awe through lakes, oceans, valleys and snow-covered mountains. In anger, our approach extends beyond mere beasts, encompassing flags, posters, and guns. Owing to attribute loss and emotion confidence, our method can effectively capture the rich and varied semantics while maintaining emotion faithfulness in EmoSet." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.397, + 0.473, + 0.624 + ], + "angle": 0, + "content": "Quantitative Comparisons As shown in Table 1, the proposed method surpasses the compared methods across all five evaluation metrics. Particularly, better performance on FID and LPIPS indicates our method can generate images with higher fidelity and diversity, effectively capturing the characteristics of the training data. All methods achieve comparable results on emotion accuracy. From Figure 6, we observe that comparison methods are prone to fall into singular or incorrect emotion representations. Even such generation results are still separable in eight classes, they do not conform to human emotional cognition. This suggests that relying solely on Emo-A may be insufficient for EICG task. Therefore, we additionally introduce Sem-C and Sem-D to estimate the content clarity and diversity, where our method exhibits a clear advantage over other methods." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.644, + 0.473, + 0.902 + ], + "angle": 0, + "content": "User Study Besides qualitative and quantitative comparisons, we also conduct a user study to determine whether our method is preferred by humans and to understand how people perceive emotions. We invite 14 participants from different social backgrounds and each test session lasts about 30 minutes. In the first part, generation results are evaluated on three dimensions: image fidelity, emotion faithfulness and semantic diversity. Each question presented to the participants includes two sets of images conveying the same emotion, drawn from our method and one of the comparison methods. The participants are then asked: which group is more realistic? which group evokes a stronger sense of [emotion type]? which group is more diverse? As illustrated in Table 2, our method attains the top rankings compared to the other three methods, particularly excelling in semantic diversity. We aim to explore the factors influencing visual emotions in the second part." + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.089, + 0.895, + 0.263 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.28, + 0.895, + 0.323 + ], + "angle": 0, + "content": "Figure 7. Emotion decomposition. Each emotion word is broken down into a set of (a) emotional concepts, reflecting the semantics in (b) generated images." + }, + { + "type": "table_caption", + "bbox": [ + 0.5, + 0.334, + 0.892, + 0.362 + ], + "angle": 0, + "content": "Table 3. Comparisons with the state-of-the-art methods on emotion transfer task, involving three metrics." + }, + { + "type": "table", + "bbox": [ + 0.502, + 0.367, + 0.891, + 0.445 + ], + "angle": 0, + "content": "
MethodEmo-A↑CLIP-img ↑CLIP-text ↑
amusementfearamusementfearamusementfear
Stable Diffusion51.54%56.67%0.9290.8250.2570.251
Textual Inversion60.82%40.00%0.9020.7920.2700.259
Ours72.16%63.33%0.9130.8410.2760.270
" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.472, + 0.892, + 0.624 + ], + "angle": 0, + "content": "Participants are shown an emotional image generated by our method and are asked: which emotion best describes the image? why do you feel such emotion? Compared to the \\(76.25\\%\\) machine predicted one in Table 1, \\(82.14\\%\\) emotion accuracy is achieved by user voting, where generated images are more emotion-evoking towards human participants. Additionally, \\(88.39\\%\\) of the responses indicate that emotions are predominantly triggered by the content/semantic. This underscores how our task, EICG, is closely aligned with human cognition." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.646, + 0.655, + 0.663 + ], + "angle": 0, + "content": "4.3. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.675, + 0.892, + 0.901 + ], + "angle": 0, + "content": "We examine the efficacy of each network design, encompassing the non-linear mapping network \\( F \\), the attribute loss \\( \\mathcal{L}_{attr} \\) and the emotion confidence \\( \\alpha_{ij} \\). In Table 1, without nonlinear mapping network, emotion representations are aggregated, which fails to restore the real image set (high FID) and lacks semantic diversity (low Sem-D). Attribute loss is introduced to enhance semantic clarity and diversity, whose absence leads to performance drops in Sem-C and Sem-D. Besides, as shown in Figure 6, generated images exhibit semantic distortions when attribute loss is absent (w/o \\( \\mathcal{L}_{attr} \\)) and display explicit contents with attribute loss (w/o \\( \\alpha_{ij} \\)). While image contents become clear and diverse with attribute loss, it is only with emotion confidence that we can effectively filter out emotion-agnostic semantics and generate images that evoke specific emotions (Ours)." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6364" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.104, + 0.091, + 0.5, + 0.24 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.264, + 0.242, + 0.361, + 0.252 + ], + "angle": 0, + "content": "(a) Emotion Transfer" + }, + { + "type": "image", + "bbox": [ + 0.501, + 0.091, + 0.867, + 0.24 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.638, + 0.242, + 0.728, + 0.252 + ], + "angle": 0, + "content": "(b) Emotion Fusion" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.252, + 0.892, + 0.28 + ], + "angle": 0, + "content": "Figure 8. Emotion creation. (a) transfers emotion representations (i.e., amusement, fear) to a series of neutral contents while (b) fuse two emotions (i.e., amusement-awe, amusement-fear) together, which may be helpful for emotional art design." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.287, + 0.212, + 0.302 + ], + "angle": 0, + "content": "4.4. Applications" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.305, + 0.474, + 0.577 + ], + "angle": 0, + "content": "Emotion Decomposition Emotions, serving as abstract concepts, pose a challenge for generative models to understand. Our method provides an opportunity to comprehend visual emotions by identifying the most relevant semantic contents for each emotion. To be specific, we visualize the semantics that are most closely aligned with our emotion representations in CLIP space. Each concept in Figure 7 (a), such as surfboard, bicycle and athletic field, is very likely to elicit excitement, where the corresponding images are presented in Figure 7 (b). Upon viewing such images, we identify the semantics and instinctively link them to specific emotions. These emotional concepts exhibit diversity, explicitness, and a strong capacity to evoke emotions. By decomposing visual emotions, we can not only generate emotional images with various semantics but also gain a deeper understanding of emotion evocation process. The results reveal the close relationship between emotions and semantics, in accordance with the psychological studies [3]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.588, + 0.471, + 0.86 + ], + "angle": 0, + "content": "Emotion Transfer Once we identify emotional contents, the next step is to explore how we can use it to create meaningful and compelling designs. In addition to emotional content, there are also neutral ones. As shown in Figure 8 (a), we combine the common neutral objects/scenes with emotional representations learned by our method. Surprisingly, we find that these representations effectively preserve emotional semantics and seamlessly integrate them with new concepts. Taking amusement as an example, it preserves several semantics including amusement park, picnic, princess, balloon and beautiful lanterns. In Table 3, our method is quantitatively compared with the state-of-the-art methods on emotion transfer task, specializing in room, where our method can well-preserve semantics and effectively elicit emotions. Crucially, these creations can evoke explicit and strong emotions across various neutral semantics, suggesting the potential of our method in image editing, image transfer and emotional art design." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Emotion Fusion Additionally, we explore the possibilities of combining different emotion representations to evoke" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.287, + 0.895, + 0.439 + ], + "angle": 0, + "content": "multiple emotions. In Figure 8 (b), we combine amusement and awe (positive-positive) as well as amusement and fear (positive-negative), bringing some intriguing observations. In the combination of amusement and awe, we observe elements associated with amusement, such as toys, balloons, and ice-creams, alongside awe-inspiring elements like the blue sky, mountains, ocean, and city views. Particularly, one may feel both fear and amusement when viewing the funny and horrible face. When we fuse emotions, we are essentially combining their corresponding visual contents." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.446, + 0.62, + 0.461 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.465, + 0.895, + 0.617 + ], + "angle": 0, + "content": "Discussion In this paper, we introduce a new task named EICG and derive three specially designed metrics. We propose an emotion space and align it with the CLIP space, incorporating attribute loss and emotion confidence to ensure semantic clarity, semantic diversity and emotion fidelity. Experimental results indicate that our method surpasses the state-of-the-art text-to-image diffusion models both qualitatively and quantitatively, where user study confirms its superiority. Additionally, we outline potential applications for EICG and present some initial but promising results." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.629, + 0.894, + 0.779 + ], + "angle": 0, + "content": "Limitations Emotions can be evoked by various visual factors such as color, style and content. In this paper, we focus on investigating the most influential factor, i.e., contents. Moreover, the relationships between emotions and content is not strictly binary. In this paper, we simplify this connection by assuming content to be either emotional or emotion-agnostic. However, in reality, it is hard to assign rose to a single emotion category. White rose may evoke sadness while red rose can elicit amusement, making it hard to decide whether rose is emotional or not." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.781, + 0.894, + 0.901 + ], + "angle": 0, + "content": "Acknowledgments: This work was supported in parts by NSFC (62302312, U21B2023), Guangdong Science and Technology Program (2023A1515011440), DEGP Innovation Team (2022KCXTD025), Shenzhen Science and Technology Program (KQTD20210811090044003, RCJC20200714114435012), Guangdong Laboratory of Artificial Intelligence and Digital Economy (SZ) and Scientific Development Funds of Shenzhen University." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6365" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.079, + 0.09, + 0.176, + 0.107 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.184 + ], + "angle": 0, + "content": "[1] Damian Borth, Tao Chen, Rongrong Ji, and Shih-Fu Chang. Sentibank: large-scale ontology and classifiers for detecting sentiment and emotions in visual content. In Proceedings of the 21st ACM International Conference on Multimedia, pages 459-460, 2013. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.185, + 0.472, + 0.253 + ], + "angle": 0, + "content": "[2] Damian Borth, Rongrong Ji, Tao Chen, Thomas Breuel, and Shih-Fu Chang. Large-scale visual sentiment ontology and detectors using adjective noun pairs. In Proceedings of the 21st ACM International Conference on Multimedia, pages 223-232, 2013. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.255, + 0.471, + 0.297 + ], + "angle": 0, + "content": "[3] Tobias Brosch, Gilles Pourtois, and David Sander. The perception and categorisation of emotional stimuli: A review. Cognition and Emotion, 24(3):377-400, 2010. 2, 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.298, + 0.471, + 0.325 + ], + "angle": 0, + "content": "[4] Linda Camras. Emotion: a psychoevolutionary synthesis, 1980. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.326, + 0.471, + 0.38 + ], + "angle": 0, + "content": "[5] Tianlang Chen, Wei Xiong, Haitian Zheng, and Jiebo Luo. Image sentiment transfer. In Proceedings of the 28th ACM International Conference on Multimedia, pages 4407-4415, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.382, + 0.471, + 0.424 + ], + "angle": 0, + "content": "[6] Domenico Consoli. A new concept of marketing: The emotional marketing. _BRAND_. Broad Research in Accounting, Negotiation, and Distribution, 1(1):52-59, 2010. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.425, + 0.471, + 0.465 + ], + "angle": 0, + "content": "[7] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in Neural Information Processing Systems, 34:8780-8794, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.467, + 0.471, + 0.521 + ], + "angle": 0, + "content": "[8] Ziyi Dong, Pengxu Wei, and Liang Lin. Dreamartist: Towards controllable one-shot text-to-image generation via contrastive prompt-tuning. arXiv preprint arXiv:2211.11337, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.523, + 0.471, + 0.591 + ], + "angle": 0, + "content": "[9] Oran Gafni, Adam Polyak, Oron Ashual, Shelly Sheynin, Devi Parikh, and Yaniv Taigman. Make-a-scene: Scene-based text-to-image generation with human priors. In European Conference on Computer Vision, pages 89-106. Springer, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.078, + 0.593, + 0.471, + 0.661 + ], + "angle": 0, + "content": "[10] Rinon Gal, Yuval Alaluf, Yuval Atzmon, Or Patashnik, Amit H Bermano, Gal Chechik, and Daniel Cohen-Or. An image is worth one word: Personalizing text-to-image generation using textual inversion. arXiv preprint arXiv:2208.01618, 2022. 2, 3, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.078, + 0.662, + 0.471, + 0.718 + ], + "angle": 0, + "content": "[11] Leon A Gatys, Alexander S Ecker, and Matthias Bethge. Image style transfer using convolutional neural networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2414-2423, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.078, + 0.719, + 0.471, + 0.773 + ], + "angle": 0, + "content": "[12] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial networks. Communications of the ACM, 63(11):139–144, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.078, + 0.775, + 0.471, + 0.816 + ], + "angle": 0, + "content": "[13] Alan Hanjalic. Extracting moods from pictures and sounds: Towards truly personalized tv. IEEE Signal Processing Magazine, 23(2):90-100, 2006. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.078, + 0.817, + 0.471, + 0.872 + ], + "angle": 0, + "content": "[14] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 770-778, 2016. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.078, + 0.873, + 0.471, + 0.902 + ], + "angle": 0, + "content": "[15] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a" + }, + { + "type": "list", + "bbox": [ + 0.078, + 0.115, + 0.472, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.134 + ], + "angle": 0, + "content": "two time-scale update rule converge to a local nash equilibrium. Advances in Neural Information Processing Systems, 30, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.136, + 0.892, + 0.178 + ], + "angle": 0, + "content": "[16] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in Neural Information Processing Systems, 33:6840-6851, 2020. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.18, + 0.892, + 0.235 + ], + "angle": 0, + "content": "[17] Elaine Hsieh and Brenda Nicodemus. Conceptualizing emotion in healthcare interpreting: A normative approach to interpreters' emotion work. Patient Education and Counseling, 98(12):1474-1481, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.238, + 0.892, + 0.293 + ], + "angle": 0, + "content": "[18] Yifei Huang, Sheng Qiu, Changbo Wang, and Chenhui Li. Learning representations for high-dynamic-range image color transfer in a self-supervised way. IEEE Transactions on Multimedia, 23:176-188, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.296, + 0.892, + 0.352 + ], + "angle": 0, + "content": "[19] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4401-4410, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.354, + 0.892, + 0.382 + ], + "angle": 0, + "content": "[20] Diederik P Kingma and Max Welling. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114, 2013. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.384, + 0.892, + 0.453 + ], + "angle": 0, + "content": "[21] Nupur Kumari, Bingliang Zhang, Richard Zhang, Eli Shechtman, and Jun-Yan Zhu. Multi-concept customization of text-to-image diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1931-1941, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.455, + 0.892, + 0.496 + ], + "angle": 0, + "content": "[22] Yann LeCun, Sumit Chopra, Raia Hadsell, M Ranzato, and Fujie Huang. A tutorial on energy-based learning. Predicting Structured Data, 1(0), 2006. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.499, + 0.892, + 0.54 + ], + "angle": 0, + "content": "[23] Joonwooan Lee and EunJong Park. Fuzzy similarity-based emotional classification of color images. IEEE Transactions on Multimedia, 13(5):1031-1039, 2011. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.542, + 0.892, + 0.611 + ], + "angle": 0, + "content": "[24] Wentong Liao, Kai Hu, Michael Ying Yang, and Bodo Rosenhahn. Text to image generation with semantic-spatial aware gan. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18187-18196, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.614, + 0.892, + 0.655 + ], + "angle": 0, + "content": "[25] Da Liu, Yaxi Jiang, Min Pei, and Shiguang Liu. Emotional image color transfer via deep learning. Pattern Recognition Letters, 110:16-22, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.657, + 0.892, + 0.714 + ], + "angle": 0, + "content": "[26] Jana Machajdik and Allan Hanbury. Affective image classification using features inspired by psychology and art theory. In Proceedings of the 18th ACM International Conference on Multimedia, pages 83-92, 2010. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.716, + 0.892, + 0.771 + ], + "angle": 0, + "content": "[27] Saif Mohammad and Svetlana Kiritchenko. Wikiart emotions: An annotated dataset of emotions evoked by art. In Proceedings of the eleventh International Conference on Language Resources and Evaluation, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.773, + 0.892, + 0.843 + ], + "angle": 0, + "content": "[28] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.845, + 0.895, + 0.901 + ], + "angle": 0, + "content": "[29] Magnus Oskarsson. Robust image-to-image color transfer using optimal inlier maximization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 786-795, 2021. 3" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.895, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.517, + 0.957 + ], + "angle": 0, + "content": "6366" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.161 + ], + "angle": 0, + "content": "[30] Kuan-Chuan Peng, Tsuhan Chen, Amir Sadovnik, and Andrew C Gallagher. A mixed bag of emotions: Model, predict, and transfer emotion distributions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 860-868, 2015. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.162, + 0.47, + 0.245 + ], + "angle": 0, + "content": "[31] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763, 2021. 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.246, + 0.469, + 0.3 + ], + "angle": 0, + "content": "[32] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 1 (2):3, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.301, + 0.469, + 0.383 + ], + "angle": 0, + "content": "[33] Harsh Rangwani, Lavish Bansal, Kartik Sharma, Tejan Karmali, Varun Jampani, and R Venkatesh Babu. Noisytwins: Class-consistent and diverse image generation through stylegans. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5987-5996, 2023. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.385, + 0.469, + 0.426 + ], + "angle": 0, + "content": "[34] Tianrong Rao, Xiaoxu Li, and Min Xu. Learning multi-level deep representations for image emotion classification. Neural Processing Letters, pages 1-19, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.427, + 0.469, + 0.468 + ], + "angle": 0, + "content": "[35] Tianrong Rao, Xiaoxu Li, and Min Xu. Learning multi-level deep representations for image emotion classification. Neural Processing Letters, 51:2043-2061, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.469, + 0.469, + 0.511 + ], + "angle": 0, + "content": "[36] Erik Reinhard, Michael Adhikhmin, Bruce Gooch, and Peter Shirley. Color transfer between images. IEEE Computer Graphics and Applications, 21(5):34-41, 2001. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.511, + 0.469, + 0.552 + ], + "angle": 0, + "content": "[37] Danilo Rezende and Shakir Mohamed. Variational inference with normalizing flows. In International Conference on Machine Learning, pages 1530-1538, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.553, + 0.47, + 0.621 + ], + "angle": 0, + "content": "[38] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10684-10695, 2022. 2, 3, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.622, + 0.469, + 0.704 + ], + "angle": 0, + "content": "[39] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22500-22510, 2023. 2, 3, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.706, + 0.469, + 0.788 + ], + "angle": 0, + "content": "[40] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in Neural Information Processing Systems, 35:36479-36494, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.789, + 0.469, + 0.859 + ], + "angle": 0, + "content": "[41] Shikun Sun, Jia Jia, Haozhe Wu, Zijie Ye, and Junliang Xing. Msnet: A deep architecture using multi-sentiment semantics for sentiment-aware image style transfer. In ICASSP 2023-2023 IEEE International Conference on Acoustics, Speech and Signal Processing, pages 1-5. IEEE, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.86, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[42] Zhentao Tan, Mengelei Chai, Dongdong Chen, Jing Liao, Qi Chu, Bin Liu, Gang Hua, and Nenghai Yu. Diverse semantic image synthesis via probability distribution modeling. In" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.892, + 0.121 + ], + "angle": 0, + "content": "Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7962-7971, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.122, + 0.892, + 0.178 + ], + "angle": 0, + "content": "[43] Guy Tevet, Brian Gordon, Amir Hertz, Amit H Bermano, and Daniel Cohen-Or. Motionclip: Exposing human motion generation to clip space. In European Conference on Computer Vision, pages 358–374. Springer, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.179, + 0.892, + 0.219 + ], + "angle": 0, + "content": "[44] Xiaohui Wang, Jia Jia, and Lianhong Cai. Affective image adjustment with a single word. The Visual Computer, 29: 1121-1133, 2013. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.221, + 0.892, + 0.277 + ], + "angle": 0, + "content": "[45] Zhizhong Wang, Lei Zhao, and Wei Xing. *Styloid diffusion: Controllable disentangled style transfer via diffusion models*. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7677-7689, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.278, + 0.892, + 0.333 + ], + "angle": 0, + "content": "[46] Yuxiang Wei, Yabo Zhang, Zhilong Ji, Jinfeng Bai, Lei Zhang, and Wangmeng Zuo. Elite: Encoding visual concepts into textual embeddings for customized text-to-image generation. arXiv preprint arXiv:2302.13848, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.334, + 0.892, + 0.404 + ], + "angle": 0, + "content": "[47] Shuchen Weng, Peixuan Zhang, Zheng Chang, Xinlong Wang, Si Li, and Boxin Shi. Affective image filter: Reflecting emotions from text to images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10810-10819, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.405, + 0.892, + 0.46 + ], + "angle": 0, + "content": "[48] Ali Yadollahi, Ameneh Gholipour Shahraki, and Osmar R Zaiane. Current state of text sentiment analysis from opinion to emotion mining. ACM Computing Surveys (CSUR), 50(2): 1-33, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.461, + 0.892, + 0.504 + ], + "angle": 0, + "content": "[49] Chuan-Kai Yang and Li-Kai Peng. Automatic mood-transferring between color images. IEEE Computer Graphics and Applications, 28(2):52-61, 2008. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.505, + 0.892, + 0.574 + ], + "angle": 0, + "content": "[50] Jufeng Yang, Dongyu She, Yu-Kun Lai, Paul L Rosin, and Ming-Hsuan Yang. Weakly supervised coupled networks for visual sentiment analysis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7584-7592, 2018. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.576, + 0.892, + 0.63 + ], + "angle": 0, + "content": "[51] Jingyuan Yang, Xinbo Gao, Leida Li, Xiumei Wang, and Jinshan Ding. Solver: Scene-object interrelated visual emotion reasoning network. IEEE Transactions on Image Processing, 30:8686-8701, 2021. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.632, + 0.892, + 0.674 + ], + "angle": 0, + "content": "[52] Jingyuan Yang, Jie Li, Xiumei Wang, Yuxuan Ding, and Xinbo Gao. Stimuli-aware visual emotion analysis. IEEE Transactions on Image Processing, 30:7432-7445, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.675, + 0.892, + 0.745 + ], + "angle": 0, + "content": "[53] Jingyuan Yang, Qirui Huang, Tingting Ding, Dani Lischinski, Danny Cohen-Or, and Hui Huang. Emoset: A large-scale visual emotion dataset with rich attributes. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 20383–20394, 2023. 2, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.746, + 0.892, + 0.814 + ], + "angle": 0, + "content": "[54] Shuai Yang, Liming Jiang, Ziwei Liu, and Chen Change Loy. Pastiche master: Exemplar-based high-resolution portrait style transfer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7693-7702, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.816, + 0.892, + 0.872 + ], + "angle": 0, + "content": "[55] Chenrui Zhang and Yuxin Peng. Stacking vae and gan for context-aware text-to-image generation. In 2018 IEEE Fourth International Conference on Multimedia Big Data (BigMM), pages 1-5. IEEE, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.873, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[56] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "6367" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.11, + 0.092, + 0.469, + 0.12 + ], + "angle": 0, + "content": "Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3836-3847, 2023. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.122, + 0.472, + 0.19 + ], + "angle": 0, + "content": "[57] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 586-595, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.192, + 0.471, + 0.247 + ], + "angle": 0, + "content": "[58] Wei Zhang, Xuanyu He, and Weizhi Lu. Exploring discriminative representations for image emotion recognition with cnns. IEEE Transactions on Multimedia, 22(2):515-523, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.249, + 0.471, + 0.317 + ], + "angle": 0, + "content": "[59] Minfeng Zhu, Pingbo Pan, Wei Chen, and Yi Yang. Dm-gan: Dynamic memory generative adversarial networks for text-to-image synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5802-5810, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.319, + 0.471, + 0.373 + ], + "angle": 0, + "content": "[60] Siqi Zhu, Chunmei Qing, Canqiang Chen, and Xiangmin Xu. Emotional generative adversarial network for image emotion transfer. Expert Systems with Applications, 216:119485, 2023. 3" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.373 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.956 + ], + "angle": 0, + "content": "6368" + } + ] +] \ No newline at end of file diff --git a/2024/EmoGen_ Emotional Image Content Generation with Text-to-Image Diffusion Models/a0b8035c-47d5-4f15-8dd7-134ed07d94b6_origin.pdf b/2024/EmoGen_ Emotional Image Content Generation with Text-to-Image Diffusion Models/a0b8035c-47d5-4f15-8dd7-134ed07d94b6_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..7b552f7028455dc08eb883ccc130c26de82cc919 --- /dev/null +++ b/2024/EmoGen_ Emotional Image Content Generation with Text-to-Image Diffusion Models/a0b8035c-47d5-4f15-8dd7-134ed07d94b6_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ba4fa4ad66f1e7a79449af26025ba12be0dcadc3087ad891cbdd6a7cb6d7eee6 +size 4183982 diff --git a/2024/EmoGen_ Emotional Image Content Generation with Text-to-Image Diffusion Models/full.md b/2024/EmoGen_ Emotional Image Content Generation with Text-to-Image Diffusion Models/full.md new file mode 100644 index 0000000000000000000000000000000000000000..28799c5a775d3117e719c0b31666fbe4907ee55c --- /dev/null +++ b/2024/EmoGen_ Emotional Image Content Generation with Text-to-Image Diffusion Models/full.md @@ -0,0 +1,295 @@ +# EmoGen: Emotional Image Content Generation with Text-to-Image Diffusion Models + +Jingyuan Yang, Jiawei Feng, Hui Huang* + +Shenzhen University + +{jingyuanyang.jyy, fengjiawei0909, hhzhiyan}@gmail.com + +![](images/3fc12973e2e4509d7d55d6a4667bb7fd1e11c39a8516f2e5b365b787e9af60b1.jpg) + +![](images/69d9a7ffc9eaf994bd51691e7c3dcdf7393ac08e185033f3146b3dac0706c75b.jpg) + +![](images/1a3ffcf1c2f079935bf3944a83119a81baa91e5ecba26c7991bfda597ff61615.jpg) + +![](images/91b5b344aa006292105d35ca67e41db0f7a85238bd932ee239522ff8cd48e43c.jpg) + +![](images/9c3314b79d1222a4519c490cb43c5d562ff6ac46e47638505ce2bbe17c1b0b7a.jpg) +Figure 1. Emotional Image Content Generation (EICG). Given an emotion category, our network produces images that exhibit unambiguous meanings (semantic-clear), reflect the intended emotion (emotion-faithful) and incorporate varied semantics (semantic-diverse). + +![](images/158a4ac8a6ec0c14f6519a00a14c99a86ac1e012f548eab7c2971d580ad69c94.jpg) + +![](images/97ce0c88ceeb33497b4ceb32c5b00b94fe8967464248edd70559c9c7e1acf3fb.jpg) + +![](images/aec73d73247d35cafd079ad62262717cc1ac515a3e4cc52c8657dbd1dc9d78e9.jpg) + +# Abstract + +Recent years have witnessed remarkable progress in image generation task, where users can create visually astonishing images with high-quality. However, existing text-to-image diffusion models are proficient in generating concrete concepts (dogs) but encounter challenges with more abstract ones (emotions). Several efforts have been made to modify image emotions with color and style adjustments, facing limitations in effectively conveying emotions with fixed image contents. In this work, we introduce Emotional Image Content Generation (EICG), a new task to generate semantic-clear and emotion-faithful images given emotion categories. Specifically, we propose an emotion space and construct a mapping network to align it with the powerful Contrastive Language-Image Pre-training (CLIP) space, providing a concrete interpretation of abstract emo + +tions. Attribute loss and emotion confidence are further proposed to ensure the semantic diversity and emotion fidelity of the generated images. Our method outperforms the state-of-the-art text-to-image approaches both quantitatively and qualitatively, where we derive three custom metrics, i.e., emotion accuracy, semantic clarity and semantic diversity. In addition to generation, our method can help emotion understanding and inspire emotional art design. Project page: https://vcc.tech/research/2024/EmoGen. + +# 1. Introduction + +"What I cannot create, I do not understand." + +Richard Feynman + +Emotions, often elusive yet profoundly influential, shape our actions, foster connections, and spark passions. With the prevalence of social medias, users tend to share specially crafted images to express their feelings. Aiming to find out + +people's emotional responses towards different stimuli, Visual Emotion Analysis (VEA) is an intriguing yet challenging task in computer vision [34, 50, 51]. Recent years have witnessed rapid development in this field, bringing potential applications such as opinion mining [48], market advertising [6] and mental healthcare [17]. + +Thanks to the advent of diffusion models [7, 16, 38], unprecedented progress has been made in text-to-image generation, where users can generate high-quality images with crafted prompts or personalized objects [10, 39, 56]. Existing text-to-image diffusion models, are often excel in generating concrete concepts (e.g., cat, house, mountain) but face limitations when tasked with more abstract ones (e.g., amusement, anger, sadness). In reality, however, photographic works are not necessarily targeted on specific entities, but are often composed to convey certain feelings. + +A natural question arises: What if machines could create images that not only please our eyes but also touch our hearts? Generating emotions is very challenging. Emotions are abstract while images are concrete, leaving the affective gap [13] hard to surmount. To bridge the gap, several efforts have been made to modify visual emotions by adjusting colors and styles, i.e., image emotion transfer [30, 41, 47]. These methods, however, meet difficulties in evoking emotions correctly and significantly, i.e., $29\%$ emotion accuracy [47], as fixed image contents limit emotional variations. Moreover, we cannot generate emotional images solely from colors and styles. What truly triggers emotion? Psychological studies show that visual emotions are often evoked by specific semantics [1, 3, 4]. + +In this paper, we propose Emotional Image Content Generation (EICG), a new task to generate semantically clear and emotionally faithful visual contents conditioned on a given emotion category, as shown in Figure 1. Semantic clarity demands an unambiguous representation of visual contents, while emotion faithfulness entails generating images evoke the intended emotions. Contrastive Language-Image Pre-training (CLIP) [31] is a large-scale vision-language model with rich semantics. However, we observe in Figure 2 that CLIP space can not well capture emotional relationships. Therefore, we introduce an emotion space, which groups similar emotions together while keeping dissimilar ones apart. While emotion space excels in representing emotions, CLIP space exhibits a powerful semantic structure. To align emotion space with CLIP space, we propose a mapping network, interpreting abstract emotions with concrete semantics. + +EmoSet [53] is a recently proposed large-scale visual emotion dataset with rich attributes. The Latent Diffusion Model (LDM) loss [38] is often utilized to optimize concrete entities with single and explicit semantics, posing a challenge in capturing the diversity within each emotion. To address this, we introduce an attribute loss to ensure seman + +![](images/a4ce53800202aec8a914937794b0b8f5d576aaaca2c2400d8161dd98a0244f6b.jpg) +(a) Emotion Space +Figure 2. Despite (b) CLIP space demonstrates a powerful semantic structure, it struggles to effectively capture emotional relationships within (a), the proposed emotion space. + +![](images/d05b48c7b7f51571c138f86093c71efeb0a5d825959e0721fb579b67e46b6299.jpg) +(b)CLIP Space + +tic clarity and diversity, by leveraging the attribute labels in EmoSet. Recognizing that not all objects are affective, emotion confidence is further proposed to ensure the emotion fidelity of the generated contents. + +To estimate the generation quality of EICG, three evaluation metrics are specially designed: emotion accuracy, semantic clarity and semantic diversity. As EICG aims to create emotional contents, we design emotion accuracy to measure the alignment between intended and perceived emotions in the generated images. People are prone to evoke emotions only when the contents are easily recognizable. Thus we propose semantic clarity to assess the unambiguity of the generated image content. Additionally, in view of the assorted emotion stimuli, we devise semantic diversity to quantify the content richness under each emotion. We evaluate our method through both qualitative and quantitative analyses, surpassing the state-of-the-art text-to-image generation approaches across five metrics. Ablation studies are performed to verify the network design, and user studies are conducted to resonate our method with human viewers. Besides generation task, our method can also be applied to decompose emotion concepts, transfer emotional contents and fuse different emotions, which may be helpful to understand emotions and create emotional art design. + +In summary, our contributions are: + +- We introduce Emotional Image Content Generation, a novel task to generate emotion-faithful and semantic-clear image contents. We also derive three custom metrics to estimate the generation performance. +- We develop a mapping network to align the proposed emotion space to the powerful CLIP space, where attribute loss and emotion confidence are further designed to ensure the semantic richness and emotion fidelity. +- We evaluate our method against the state-of-the-art text-to-image approaches and demonstrate our superiority. Potential applications are exhibited for emotion understanding and emotional art design. + +# 2. Related work + +# 2.1. Visual Emotion Analysis + +Researchers have been involved in VEA for over two decades, ranging from early traditional approaches [2, 23, 26] to recent deep learning ones [35, 51, 52, 58]. Given the inherent abstractness and complexity of visual emotion, researchers aim to identify the most influential elements, which range from low-level features like color, texture and style [23, 26, 35, 58] to high-level semantics [2, 35, 51, 52, 58]. Lee et al. [23] propose a scheme to evaluate emotional response from color images by reasoning the prototypical color for each emotion and the input images. As a milestone, Machajdik et al. [26] extract representative low-level features in composition, including color and texture, to predict visual emotions. Besides low-level features, Borth et al. [2] propose Adjective-Noun Pair (ANP) and build a visual concept detector named Sentibank. With the help of deep learning techniques, Rao et al. [35] construct MldrNet to extract emotional clues from pixel-level, aesthetic-level and semantic-level. To form a more discriminative emotional representation, Zhang et al. [58] integrate high-level contents and low-level styles. Yang et al. propose network to mine emotions from multiple objects [52] as well as object-scene correlations [51]. Existing work often treat VEA as a classification task, i.e., input an image and predict the emotion within it. Can we reverse this process? In other words, can we generate an image targeting on the given emotion word? Only by creating emotional images can we demonstrate the significance of visual elements, leading to a better understanding of emotions. + +# 2.2. Text-to-Image Generation + +Text-to-image generation aims to convert textual descriptions into corresponding realistic images. Existing generative models can be grouped into GANs [12, 24, 59], VAEs [9, 20, 55], flow-based [37], energy-based [22] and diffusion-based [7, 16, 38, 39, 56]. Diffusion models are witnessed impressive and rapid progress in recent years, where methods like GLIDE [28], DALLE2 [32], Image [40] can generate diverse, photo-realistic and high-quality images. Notably, Stable diffusion [38] is one of the most popular diffusion models, owing to its stable training and the capability for fine-grained control, supported by an active user community. For customized generation, several diffusion-based text-to-image methods are introduced, where methods vary from learning a new embedding [8, 10] and finetuning the network parameters [21, 39, 46]. Textual inversion [10] and DreamArtist [8] learn new concepts with a few user-provided images in the word embedding space, without further training on diffusion models. While DreamBooth [39] finetunes all the parameters to learn a new concept, Custom diffusion [21] only updates the key + +and value parameters in the cross attention layers. Further, ELITE [46] speeds up the running time with accurate generation results by adopting a global and local mapping network. Existing text-to-images models are capable of generating concrete concepts [7, 24, 56], or personalized ones [10, 21, 39], but face difficulties in generating more abstract ones. In reality, photographic works are not necessarily composed of targeted concepts, but often aim to convey specific feelings. Thus, how to generate emotion-evoking images remains a pressing and critical challenge. + +# 2.3. Image Emotion Transfer + +Image style transfer [11] aims to render the semantic content under different styles, producing visually stunning results [19, 33, 45, 54]. Similarly, image color transfer [36] seeks to adjust and harmonize the color characteristics of one image to match another [18, 29]. Specifically, color and style choices can strongly influence the emotions of an image [27]. By adjusting low-level visual elements, image emotion transfer aims to modify the emotional tone of the input image, including the color-based methods [5, 25, 30, 44, 49, 60] and the style-based ones [41, 47]. Yang and Peng et al. [49] makes the first attempt to transfer image colors. Wang et al. [44] present a system to modify the image color according to a given emotion word, and Liu et al. [25] further advance it with deep learning techniques. Peng et al. [30] introduce a new approach to alter the emotion of an input image by guiding its color and texture under the target image. More recently, to reflect emotions in styles, Sun et al. [41] and Weng et al. [47] bring promising results on emotion-aware image style transfer. Nevertheless, the alteration of visual emotions through colors and styles is limited due to fixed content, resulting in subtle emotional changes, i.e., $29\%$ emotion accuracy in [47]. Psychological studies suggest that visual emotions can be elicited by specific semantics [3]. Thus, we propose a novel method to generate emotional image contents with clear semantics. + +# 3. Method + +# 3.1. Emotion Representation + +Emotion Space EICG is a challenging task, which requires both semantic clarity and emotion fidelity. How to generate an image with distinct and emotional semantics? CLIP [31] is developed to align image and text modalities, where semantically related features are located in close proximity to each other. While CLIP shows impressive semantic representation capabilities, it struggles to effectively capture emotional relationships. As demonstrated in Figure 2, we can observe that sample points with emotional similarities are distantly separated within the CLIP space due to their differing semantics, e.g., toy, amusement park and Christmas tree. To better depict emotional relation + +![](images/55fe37c14368b504bde71e5a8fc55fb3d6f580460f8ec9755dcc558656664520.jpg) +Figure 3. Training process of our network. Emotion representation (stage 1) learns a well-behaved emotion space and emotion content generation (stage 2) maps this space to CLIP space, aiming to generate image contents with emotion fidelity, semantic clarity and diversity. + +ships, we introduce the emotion space, a latent space that clusters similar emotions together while keeping dissimilar ones apart. EmoSet [53] is a large-scale dataset with rich attributes, where each image is labeled with an emotion. Using aligned image-emotion pairs, we construct an encoder $\varphi$ with ResNet-50 [14] to capture emotion representations. To train the encoder, we devise an emotion loss by implementing the widely-used Cross-Entropy (CE) loss, following the previous work [50, 51]: + +$$ +\mathcal {L} _ {e m o} = - \sum_ {i = 1} ^ {C} y _ {e m o} \log \frac {\exp (\varphi (x , i))}{\sum_ {i = 1} ^ {C} \exp (\varphi (x , i))}, \tag {1} +$$ + +where $x$ represents the input image, $y_{emo}$ denotes the emotion label and $C$ stands for the total number of emotion categories. Once the loss function converges, emotion space is established. Parameters in emotion encoder remain fixed in the following emotional content generation process. + +During inference, each emotion cluster is represented by a Gaussian distribution with learned parameters, i.e., mean and standard deviation. For example, when taking amusement as input, we randomly sample a data point from corresponding Gaussian distribution to serve as its emotion representation, as shown in Figure 3. We have confirmed that Gaussian distribution suits emotion clusters well and the random sampling process induces diversity to EICG. + +# 3.2. Emotional Content Generation + +Mapping Network While emotion space is emotionally separable, CLIP space captures rich semantics. Existing text-to-image models entail clear and specific semantics as input, making CLIP space indispensable in the generation process. Consequently, establishing the mapping between emotion space and CLIP space becomes a crucial challenge. Intuitively, we attempt to build the mapping network using fully connected layers, following previous work [33, 43]. + +However, as depicted in Figure 2, clustered feature points in the emotion space are expected to disperse in the CLIP space to capture diverse semantics. Therefore, we utilize a Multilayer Perceptron (MLP) to build the mapping + +network, incorporating non-linear operations, i.e., RELU, to facilitate the separation process. The non-linear projection $F$ is succeeded by a CLIP text transformer $t_{\theta}$ , yielding textual embedding for U-Net. The end-token embedding of the transformer's output is passed through a fully-connected layer, producing the CLIP text feature. Particularly, to better preserve the prior knowledge in the CLIP space, parameters in the transformer and linear projection are kept frozen, while parameters in non-linear projection are learned, as depicted in Figure 3. + +Attribute Loss Existing text-to-image diffusion models often employ Latent Diffusion Model (LDM) loss [38] for optimization process [10, 39, 56]: + +$$ +\mathcal {L} _ {L D M} = \mathbb {E} _ {z, x, \epsilon , t} \left[ \| \epsilon - \epsilon_ {\theta} (z _ {t}, t, t _ {\theta} (F (\varphi (x)))) \| _ {2} ^ {2} \right], \tag {2} +$$ + +where $\epsilon$ represents the added noise, $\epsilon_{\theta}$ denotes the denoising network and $z_{t}$ indicates the latent noised to time $t$ . + +In these cases, target concepts typically involve concrete entities (e.g., dog, car, flower) or personalized objects (e.g., someone's corgi). These concepts often exhibit consistency on semantic level and share certain similarities on pixel level. However, emotions are abstract concepts, where multiple semantics coexist under one specific category. Learning emotions solely with LDM loss may pose some challenges. For one thing, each emotion might collapse to a specific semantic point, e.g., amusement collapsing to amusement park, losing intra-class diversity. In reality, semantics within one emotion are diverse, where single point cannot fully capture. Moreover, since LDM loss is designed to reconstruct the input image, it primarily focuses on learning and preserving pixel-level commonalities such as color and texture. In Figure 4 (a), with LDM loss alone, CLIP embedding for amusement is prone to be colorful, without exhibiting explicit and diverse semantics. We can conclude that it is hard to achieve robust emotion representations in CLIP space by implementing LDM loss alone. + +In the pursuit of clear and diverse contents, semantics guidance is essential for the generation process. Thanks to + +![](images/26e2189a5806e3c24af1b9000267e94f9341a5a03dda4f02a746e96920ad2454.jpg) +Figure 4. Motivation for loss function design. Compare to (a) LDM loss alone, (b) attribute loss enhances semantic clarity while (c) emotion confidence ensures emotion accuracy. + +the rich attribute annotations in EmoSet, we select the mid-level attributes, i.e., object class and scene type to guide the generation process. With this semantic guidance, we formulate an attribute loss to guarantee that the generated image contents possess clear and diverse semantics. For clarity, emotions are easily triggered in people only when visual contents are represented in an unambiguous manner. Considering the varied emotional stimuli in reality, attribute loss guides the network to learn multiple semantics under one specific emotion. Our attribute loss is devised on CLIP space, by calculating the cosine similarities $f(\cdot)$ and optimizing a symmetric CE loss over the similarity scores [31]: + +$$ +\mathcal {L} _ {a t t r} = - \sum_ {j = 1} ^ {K} y _ {a t t r} \log \frac {\exp (f (v _ {e m o} , \tau_ {\theta} (a _ {j})))}{\sum_ {j = 1} ^ {K} \exp (f (v _ {e m o} , \tau_ {\theta} (a _ {j})))}, \quad (3) +$$ + +$$ +f (p, q) = \frac {p \cdot q}{\| p \| \| q \|}, \tag {4} +$$ + +where $a_{j}$ denotes the $j$ member in the attribute set, $\tau_{\theta}$ represents the text encoder, $v_{emo}$ implies the learned CLIP embedding and $K$ indicates the total number of the attributes. With the attribute loss, each sample point is converging towards the correct semantic and distancing itself from the incorrect ones. Through the combination of attribute loss and LDM loss, we can effectively map each emotion to clear and diverse semantics, as demonstrated in Figure 4 (b). + +Emotion Confidence However, it is worth noting that some of the semantics in Figure 4 (b) appear emotionally neutral, e.g., plant and tree. Since attributes are annotated objectively, not all the attributes in EmoSet are emotional. Therefore, we propose emotion confidence to measure the correlations between emotions and semantic attributes. Initially, we gather all images associated with attribute $j$ in EmoSet and send them to a pre-trained emotion classifier. Each image is predicted as an emotion vector $p(\cdot)$ and we + +![](images/28e43b6a69be182719839923a26db0c9befa21c76c8cabb777821c2a4a8dfddc.jpg) +Figure 5. Illustration of emotion confidence. Each (a) attribute is represented by (b) a distribution of confidence on eight emotions. + +sum all images up to get the emotional distribution $d_{j}$ for attribute $j$ . Each emotion $i$ within this distribution is assigned a corresponding emotion confidence $\alpha_{ij}$ : + +$$ +\alpha_ {i j} = \frac {1}{N _ {j}} \sum_ {n = 1} ^ {N _ {j}} p (x _ {n}, i), \tag {5} +$$ + +where $x_{n}$ represents the input image and $N_{j}$ denotes the total image number in attribute $j$ . We further illustrate the above process in Figure 5 with visual representations. When mountain snowy appears, people are more likely to experience awe and cemetery often elicits sadness. In contrast, the presence of tree in every emotion category suggests its lack of emotional specificity. Some attributes are emotion-related while others are not, which can be beneficial for generating emotional contents. We then use emotion confidence to balance between LDM loss and attribute loss: + +$$ +\mathcal {L} = \left(1 - \alpha_ {i j}\right) \mathcal {L} _ {L D M} + \alpha_ {i j} \mathcal {L} _ {\text {a t t r}}, \tag {6} +$$ + +where $i$ represents the emotion category $y_{emo}$ and $j$ denotes the attribute type $y_{att}$ . The greater the emotion confidence $\alpha_{ij}$ is, the stronger the impact attribute $j$ has on the specific emotion $i$ . Low confidence suggests a weak connection between the attribute and emotion, signaling that the network should learn more from the pixel-wise LDM loss. When higher confidence occurs, the network should prioritize the semantic meaning of the image, i.e., the attribute loss. With this design, our network can adapt to a wide range of cases, generating image contents that are both semantically explicit and emotionally faithful, as shown in Figure 4 (c). + +# 4. Experiments + +# 4.1. Dataset and Evaluation + +Dataset EmoSet [53] is a large-scale visual emotion dataset with rich attributes, comprising a total of 118,102 images. To investigate the connections between emotions and specific contents, we create a subset from EmoSet by preserving images with object/scene labels. Each image is labeled with both emotion and attribute labels, guiding the optimization process of emotion loss and attribute loss. Notably, the wide range of attribute labels assures for learning diverse and representative emotional contents. + +![](images/6f2b27521357d6835976cc477fb41cce39cc1ad564a1057d9b5736ac9efb940d.jpg) +Figure 6. Qualitative comparisons with the state-of-the-art text-to-image generation approaches and ablation studies of our method. + +Evaluation Metrics To comprehensively evaluate the performance of different methods on EICG task, we utilize commonly used metrics (FID, LPIPS) and design some specific ones (Emo-A, Sem-C, Sem-D). 1) FID: Frechet Inception Distance (FID) [15] quantifies the distribution distance between generated and real images, providing an estimate of image fidelity. 2) LPIPS: Similar to [42], we employ LPIPS [57] to assess the overall image diversity, with higher values indicating better performance. 3) Emo-A: Since EICG aims at creating emotion-evoking images, we design emotion accuracy to assess the emotional alignment between the targeted emotions and the generated images. 4) Sem-C: People are easily to evoke emotions under recognizable contents. We thus introduce semantic clarity to assess the explicitness of generated image contents. 5) Sem-D: Emotions are complex, where each can be triggered by multiple factors. To cover a diverse range of potential scenes or objects, we derive semantic diversity to estimate the content richness associated with each emotion. For more details, please refer to the supplementary materials. + +# 4.2. Comparisons + +As our method is the first attempt in EICG, we compare it with the most relevant and state-of-the-art text-to-image generation techniques: Stable diffusion [38], Textual inversion [10] and Dreambooth [39]. While Stable diffusion is + +Table 1. Comparisons with the state-of-the-art methods and ablation studies on emotion generation task, involving five metrics. + +
MethodFID ↓LPIPS ↑Emo-A ↑Sem-C ↑Sem-D ↑
Stable Diffusion [38]44.050.68770.77%0.6080.0199
Textual Inversion [10]50.510.70274.87%0.6050.0282
DreamBooth [39]46.890.66170.50%0.6140.0178
Ours41.600.71776.25%0.6330.0335
w/o F57.540.71371.12%0.6150.0261
w/o Lattr51.130.70765.75%0.5920.0270
w/o αij43.300.71474.88%0.5910.0263
+ +a general image generation pipeline, Textual inversion and Dreambooth specialize in customized image generation. + +Qualitative Comparisons In Figure 6, our method is qualitatively compared with the state-of-the-art methods across three emotion categories, i.e., awe, anger and contentment. Generation results of the rest five emotions can be found in the supplementary materials. Take awe as an example, all the three compared methods tend to produce images with dense textures and dim colors, which suggests that representations for each emotion may collapse to a single feature point. For anger and contentment, both Stable diffusion and Dreambooth distort the visual representations, e.g., tiger and bicycle, and generate some contents with ambiguous semantics. Though Textual inversion preserves some semantic fidelity, it generates emotion-agnostic + +Table 2. User preference study. The numbers indicate the percentage of participants who prefer our results over those compared methods, given the same emotion category as input. + +
MethodImage fidelity ↑Emotion faithfulness ↑Semantic diversity ↑
Stable Diffusion67.86±15.08%73.66±11.80%87.88±9.64%
Textual Inversion79.91±16.92%72.75±16.90%85.66±10.51%
DreamBooth77.23±14.00%80.79±8.64%81.68±17.06%
+ +contents such as shoes and cars. Since these methods are crafted to learn customized concepts, challenges may arise when handling complex and diverse emotional images. Rather than generating plants and trees, our method can provide diverse and emotion-evoking image contents for awe through lakes, oceans, valleys and snow-covered mountains. In anger, our approach extends beyond mere beasts, encompassing flags, posters, and guns. Owing to attribute loss and emotion confidence, our method can effectively capture the rich and varied semantics while maintaining emotion faithfulness in EmoSet. + +Quantitative Comparisons As shown in Table 1, the proposed method surpasses the compared methods across all five evaluation metrics. Particularly, better performance on FID and LPIPS indicates our method can generate images with higher fidelity and diversity, effectively capturing the characteristics of the training data. All methods achieve comparable results on emotion accuracy. From Figure 6, we observe that comparison methods are prone to fall into singular or incorrect emotion representations. Even such generation results are still separable in eight classes, they do not conform to human emotional cognition. This suggests that relying solely on Emo-A may be insufficient for EICG task. Therefore, we additionally introduce Sem-C and Sem-D to estimate the content clarity and diversity, where our method exhibits a clear advantage over other methods. + +User Study Besides qualitative and quantitative comparisons, we also conduct a user study to determine whether our method is preferred by humans and to understand how people perceive emotions. We invite 14 participants from different social backgrounds and each test session lasts about 30 minutes. In the first part, generation results are evaluated on three dimensions: image fidelity, emotion faithfulness and semantic diversity. Each question presented to the participants includes two sets of images conveying the same emotion, drawn from our method and one of the comparison methods. The participants are then asked: which group is more realistic? which group evokes a stronger sense of [emotion type]? which group is more diverse? As illustrated in Table 2, our method attains the top rankings compared to the other three methods, particularly excelling in semantic diversity. We aim to explore the factors influencing visual emotions in the second part. + +![](images/9cee6fd645a6a66e7296391c0540e7c64975abbb174d334b372a99b66aa45f47.jpg) +Figure 7. Emotion decomposition. Each emotion word is broken down into a set of (a) emotional concepts, reflecting the semantics in (b) generated images. + +Table 3. Comparisons with the state-of-the-art methods on emotion transfer task, involving three metrics. + +
MethodEmo-A↑CLIP-img ↑CLIP-text ↑
amusementfearamusementfearamusementfear
Stable Diffusion51.54%56.67%0.9290.8250.2570.251
Textual Inversion60.82%40.00%0.9020.7920.2700.259
Ours72.16%63.33%0.9130.8410.2760.270
+ +Participants are shown an emotional image generated by our method and are asked: which emotion best describes the image? why do you feel such emotion? Compared to the $76.25\%$ machine predicted one in Table 1, $82.14\%$ emotion accuracy is achieved by user voting, where generated images are more emotion-evoking towards human participants. Additionally, $88.39\%$ of the responses indicate that emotions are predominantly triggered by the content/semantic. This underscores how our task, EICG, is closely aligned with human cognition. + +# 4.3. Ablation Study + +We examine the efficacy of each network design, encompassing the non-linear mapping network $F$ , the attribute loss $\mathcal{L}_{attr}$ and the emotion confidence $\alpha_{ij}$ . In Table 1, without nonlinear mapping network, emotion representations are aggregated, which fails to restore the real image set (high FID) and lacks semantic diversity (low Sem-D). Attribute loss is introduced to enhance semantic clarity and diversity, whose absence leads to performance drops in Sem-C and Sem-D. Besides, as shown in Figure 6, generated images exhibit semantic distortions when attribute loss is absent (w/o $\mathcal{L}_{attr}$ ) and display explicit contents with attribute loss (w/o $\alpha_{ij}$ ). While image contents become clear and diverse with attribute loss, it is only with emotion confidence that we can effectively filter out emotion-agnostic semantics and generate images that evoke specific emotions (Ours). + +![](images/bfcb88e00f1b1601d7354f7fc0da355a52539c18e645964841cebd940b05f1bb.jpg) +(a) Emotion Transfer + +![](images/cb74f6f382a80c6fff396e98e9705b86e9e055e9777d1bd656cbfcfc83f2cb27.jpg) +(b) Emotion Fusion +Figure 8. Emotion creation. (a) transfers emotion representations (i.e., amusement, fear) to a series of neutral contents while (b) fuse two emotions (i.e., amusement-awe, amusement-fear) together, which may be helpful for emotional art design. + +# 4.4. Applications + +Emotion Decomposition Emotions, serving as abstract concepts, pose a challenge for generative models to understand. Our method provides an opportunity to comprehend visual emotions by identifying the most relevant semantic contents for each emotion. To be specific, we visualize the semantics that are most closely aligned with our emotion representations in CLIP space. Each concept in Figure 7 (a), such as surfboard, bicycle and athletic field, is very likely to elicit excitement, where the corresponding images are presented in Figure 7 (b). Upon viewing such images, we identify the semantics and instinctively link them to specific emotions. These emotional concepts exhibit diversity, explicitness, and a strong capacity to evoke emotions. By decomposing visual emotions, we can not only generate emotional images with various semantics but also gain a deeper understanding of emotion evocation process. The results reveal the close relationship between emotions and semantics, in accordance with the psychological studies [3]. + +Emotion Transfer Once we identify emotional contents, the next step is to explore how we can use it to create meaningful and compelling designs. In addition to emotional content, there are also neutral ones. As shown in Figure 8 (a), we combine the common neutral objects/scenes with emotional representations learned by our method. Surprisingly, we find that these representations effectively preserve emotional semantics and seamlessly integrate them with new concepts. Taking amusement as an example, it preserves several semantics including amusement park, picnic, princess, balloon and beautiful lanterns. In Table 3, our method is quantitatively compared with the state-of-the-art methods on emotion transfer task, specializing in room, where our method can well-preserve semantics and effectively elicit emotions. Crucially, these creations can evoke explicit and strong emotions across various neutral semantics, suggesting the potential of our method in image editing, image transfer and emotional art design. + +Emotion Fusion Additionally, we explore the possibilities of combining different emotion representations to evoke + +multiple emotions. In Figure 8 (b), we combine amusement and awe (positive-positive) as well as amusement and fear (positive-negative), bringing some intriguing observations. In the combination of amusement and awe, we observe elements associated with amusement, such as toys, balloons, and ice-creams, alongside awe-inspiring elements like the blue sky, mountains, ocean, and city views. Particularly, one may feel both fear and amusement when viewing the funny and horrible face. When we fuse emotions, we are essentially combining their corresponding visual contents. + +# 5. Conclusion + +Discussion In this paper, we introduce a new task named EICG and derive three specially designed metrics. We propose an emotion space and align it with the CLIP space, incorporating attribute loss and emotion confidence to ensure semantic clarity, semantic diversity and emotion fidelity. Experimental results indicate that our method surpasses the state-of-the-art text-to-image diffusion models both qualitatively and quantitatively, where user study confirms its superiority. Additionally, we outline potential applications for EICG and present some initial but promising results. + +Limitations Emotions can be evoked by various visual factors such as color, style and content. In this paper, we focus on investigating the most influential factor, i.e., contents. Moreover, the relationships between emotions and content is not strictly binary. In this paper, we simplify this connection by assuming content to be either emotional or emotion-agnostic. However, in reality, it is hard to assign rose to a single emotion category. White rose may evoke sadness while red rose can elicit amusement, making it hard to decide whether rose is emotional or not. + +Acknowledgments: This work was supported in parts by NSFC (62302312, U21B2023), Guangdong Science and Technology Program (2023A1515011440), DEGP Innovation Team (2022KCXTD025), Shenzhen Science and Technology Program (KQTD20210811090044003, RCJC20200714114435012), Guangdong Laboratory of Artificial Intelligence and Digital Economy (SZ) and Scientific Development Funds of Shenzhen University. + +# References + +[1] Damian Borth, Tao Chen, Rongrong Ji, and Shih-Fu Chang. Sentibank: large-scale ontology and classifiers for detecting sentiment and emotions in visual content. In Proceedings of the 21st ACM International Conference on Multimedia, pages 459-460, 2013. 2 +[2] Damian Borth, Rongrong Ji, Tao Chen, Thomas Breuel, and Shih-Fu Chang. Large-scale visual sentiment ontology and detectors using adjective noun pairs. In Proceedings of the 21st ACM International Conference on Multimedia, pages 223-232, 2013. 3 +[3] Tobias Brosch, Gilles Pourtois, and David Sander. The perception and categorisation of emotional stimuli: A review. Cognition and Emotion, 24(3):377-400, 2010. 2, 3, 8 +[4] Linda Camras. Emotion: a psychoevolutionary synthesis, 1980. 2 +[5] Tianlang Chen, Wei Xiong, Haitian Zheng, and Jiebo Luo. Image sentiment transfer. In Proceedings of the 28th ACM International Conference on Multimedia, pages 4407-4415, 2020. 3 +[6] Domenico Consoli. A new concept of marketing: The emotional marketing. _BRAND_. Broad Research in Accounting, Negotiation, and Distribution, 1(1):52-59, 2010. 2 +[7] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in Neural Information Processing Systems, 34:8780-8794, 2021. 2, 3 +[8] Ziyi Dong, Pengxu Wei, and Liang Lin. Dreamartist: Towards controllable one-shot text-to-image generation via contrastive prompt-tuning. arXiv preprint arXiv:2211.11337, 2022. 3 +[9] Oran Gafni, Adam Polyak, Oron Ashual, Shelly Sheynin, Devi Parikh, and Yaniv Taigman. Make-a-scene: Scene-based text-to-image generation with human priors. In European Conference on Computer Vision, pages 89-106. Springer, 2022. 3 +[10] Rinon Gal, Yuval Alaluf, Yuval Atzmon, Or Patashnik, Amit H Bermano, Gal Chechik, and Daniel Cohen-Or. An image is worth one word: Personalizing text-to-image generation using textual inversion. arXiv preprint arXiv:2208.01618, 2022. 2, 3, 4, 6 +[11] Leon A Gatys, Alexander S Ecker, and Matthias Bethge. Image style transfer using convolutional neural networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2414-2423, 2016. 3 +[12] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial networks. Communications of the ACM, 63(11):139–144, 2020. 3 +[13] Alan Hanjalic. Extracting moods from pictures and sounds: Towards truly personalized tv. IEEE Signal Processing Magazine, 23(2):90-100, 2006. 2 +[14] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 770-778, 2016. 4 +[15] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a + +two time-scale update rule converge to a local nash equilibrium. Advances in Neural Information Processing Systems, 30, 2017. 6 +[16] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in Neural Information Processing Systems, 33:6840-6851, 2020. 2, 3 +[17] Elaine Hsieh and Brenda Nicodemus. Conceptualizing emotion in healthcare interpreting: A normative approach to interpreters' emotion work. Patient Education and Counseling, 98(12):1474-1481, 2015. 2 +[18] Yifei Huang, Sheng Qiu, Changbo Wang, and Chenhui Li. Learning representations for high-dynamic-range image color transfer in a self-supervised way. IEEE Transactions on Multimedia, 23:176-188, 2020. 3 +[19] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4401-4410, 2019. 3 +[20] Diederik P Kingma and Max Welling. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114, 2013. 3 +[21] Nupur Kumari, Bingliang Zhang, Richard Zhang, Eli Shechtman, and Jun-Yan Zhu. Multi-concept customization of text-to-image diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1931-1941, 2023. 3 +[22] Yann LeCun, Sumit Chopra, Raia Hadsell, M Ranzato, and Fujie Huang. A tutorial on energy-based learning. Predicting Structured Data, 1(0), 2006. 3 +[23] Joonwooan Lee and EunJong Park. Fuzzy similarity-based emotional classification of color images. IEEE Transactions on Multimedia, 13(5):1031-1039, 2011. 3 +[24] Wentong Liao, Kai Hu, Michael Ying Yang, and Bodo Rosenhahn. Text to image generation with semantic-spatial aware gan. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18187-18196, 2022. 3 +[25] Da Liu, Yaxi Jiang, Min Pei, and Shiguang Liu. Emotional image color transfer via deep learning. Pattern Recognition Letters, 110:16-22, 2018. 3 +[26] Jana Machajdik and Allan Hanbury. Affective image classification using features inspired by psychology and art theory. In Proceedings of the 18th ACM International Conference on Multimedia, pages 83-92, 2010. 3 +[27] Saif Mohammad and Svetlana Kiritchenko. Wikiart emotions: An annotated dataset of emotions evoked by art. In Proceedings of the eleventh International Conference on Language Resources and Evaluation, 2018. 3 +[28] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741, 2021. 3 +[29] Magnus Oskarsson. Robust image-to-image color transfer using optimal inlier maximization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 786-795, 2021. 3 + +[30] Kuan-Chuan Peng, Tsuhan Chen, Amir Sadovnik, and Andrew C Gallagher. A mixed bag of emotions: Model, predict, and transfer emotion distributions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 860-868, 2015. 2, 3 +[31] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763, 2021. 2, 3, 5 +[32] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 1 (2):3, 2022. 3 +[33] Harsh Rangwani, Lavish Bansal, Kartik Sharma, Tejan Karmali, Varun Jampani, and R Venkatesh Babu. Noisytwins: Class-consistent and diverse image generation through stylegans. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5987-5996, 2023. 3, 4 +[34] Tianrong Rao, Xiaoxu Li, and Min Xu. Learning multi-level deep representations for image emotion classification. Neural Processing Letters, pages 1-19, 2016. 2 +[35] Tianrong Rao, Xiaoxu Li, and Min Xu. Learning multi-level deep representations for image emotion classification. Neural Processing Letters, 51:2043-2061, 2020. 3 +[36] Erik Reinhard, Michael Adhikhmin, Bruce Gooch, and Peter Shirley. Color transfer between images. IEEE Computer Graphics and Applications, 21(5):34-41, 2001. 3 +[37] Danilo Rezende and Shakir Mohamed. Variational inference with normalizing flows. In International Conference on Machine Learning, pages 1530-1538, 2015. 3 +[38] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10684-10695, 2022. 2, 3, 4, 6 +[39] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22500-22510, 2023. 2, 3, 4, 6 +[40] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in Neural Information Processing Systems, 35:36479-36494, 2022. 3 +[41] Shikun Sun, Jia Jia, Haozhe Wu, Zijie Ye, and Junliang Xing. Msnet: A deep architecture using multi-sentiment semantics for sentiment-aware image style transfer. In ICASSP 2023-2023 IEEE International Conference on Acoustics, Speech and Signal Processing, pages 1-5. IEEE, 2023. 2, 3 +[42] Zhentao Tan, Mengelei Chai, Dongdong Chen, Jing Liao, Qi Chu, Bin Liu, Gang Hua, and Nenghai Yu. Diverse semantic image synthesis via probability distribution modeling. In + +Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7962-7971, 2021. 6 +[43] Guy Tevet, Brian Gordon, Amir Hertz, Amit H Bermano, and Daniel Cohen-Or. Motionclip: Exposing human motion generation to clip space. In European Conference on Computer Vision, pages 358–374. Springer, 2022. 4 +[44] Xiaohui Wang, Jia Jia, and Lianhong Cai. Affective image adjustment with a single word. The Visual Computer, 29: 1121-1133, 2013. 3 +[45] Zhizhong Wang, Lei Zhao, and Wei Xing. *Styloid diffusion: Controllable disentangled style transfer via diffusion models*. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7677-7689, 2023. 3 +[46] Yuxiang Wei, Yabo Zhang, Zhilong Ji, Jinfeng Bai, Lei Zhang, and Wangmeng Zuo. Elite: Encoding visual concepts into textual embeddings for customized text-to-image generation. arXiv preprint arXiv:2302.13848, 2023. 3 +[47] Shuchen Weng, Peixuan Zhang, Zheng Chang, Xinlong Wang, Si Li, and Boxin Shi. Affective image filter: Reflecting emotions from text to images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10810-10819, 2023. 2, 3 +[48] Ali Yadollahi, Ameneh Gholipour Shahraki, and Osmar R Zaiane. Current state of text sentiment analysis from opinion to emotion mining. ACM Computing Surveys (CSUR), 50(2): 1-33, 2017. 2 +[49] Chuan-Kai Yang and Li-Kai Peng. Automatic mood-transferring between color images. IEEE Computer Graphics and Applications, 28(2):52-61, 2008. 3 +[50] Jufeng Yang, Dongyu She, Yu-Kun Lai, Paul L Rosin, and Ming-Hsuan Yang. Weakly supervised coupled networks for visual sentiment analysis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7584-7592, 2018. 2, 4 +[51] Jingyuan Yang, Xinbo Gao, Leida Li, Xiumei Wang, and Jinshan Ding. Solver: Scene-object interrelated visual emotion reasoning network. IEEE Transactions on Image Processing, 30:8686-8701, 2021. 2, 3, 4 +[52] Jingyuan Yang, Jie Li, Xiumei Wang, Yuxuan Ding, and Xinbo Gao. Stimuli-aware visual emotion analysis. IEEE Transactions on Image Processing, 30:7432-7445, 2021. 3 +[53] Jingyuan Yang, Qirui Huang, Tingting Ding, Dani Lischinski, Danny Cohen-Or, and Hui Huang. Emoset: A large-scale visual emotion dataset with rich attributes. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 20383–20394, 2023. 2, 4, 5 +[54] Shuai Yang, Liming Jiang, Ziwei Liu, and Chen Change Loy. Pastiche master: Exemplar-based high-resolution portrait style transfer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7693-7702, 2022. 3 +[55] Chenrui Zhang and Yuxin Peng. Stacking vae and gan for context-aware text-to-image generation. In 2018 IEEE Fourth International Conference on Multimedia Big Data (BigMM), pages 1-5. IEEE, 2018. 3 +[56] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In + +Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3836-3847, 2023. 2, 3, 4 +[57] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 586-595, 2018. 6 +[58] Wei Zhang, Xuanyu He, and Weizhi Lu. Exploring discriminative representations for image emotion recognition with cnns. IEEE Transactions on Multimedia, 22(2):515-523, 2019. 3 +[59] Minfeng Zhu, Pingbo Pan, Wei Chen, and Yi Yang. Dm-gan: Dynamic memory generative adversarial networks for text-to-image synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5802-5810, 2019. 3 +[60] Siqi Zhu, Chunmei Qing, Canqiang Chen, and Xiangmin Xu. Emotional generative adversarial network for image emotion transfer. Expert Systems with Applications, 216:119485, 2023. 3 \ No newline at end of file diff --git a/2024/EmoGen_ Emotional Image Content Generation with Text-to-Image Diffusion Models/images.zip b/2024/EmoGen_ Emotional Image Content Generation with Text-to-Image Diffusion Models/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..e31964a5efef2947e6aa030b994993061f5e8746 --- /dev/null +++ b/2024/EmoGen_ Emotional Image Content Generation with Text-to-Image Diffusion Models/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:cbabb49470ebeff00db19eaa967c5a98d883fd8fddf2947a14a34ba637b6cbe7 +size 976790 diff --git a/2024/EmoGen_ Emotional Image Content Generation with Text-to-Image Diffusion Models/layout.json b/2024/EmoGen_ Emotional Image Content Generation with Text-to-Image Diffusion Models/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..2dc2d46816c7c1477b9edf50fb0577ff63b0d90b --- /dev/null +++ b/2024/EmoGen_ Emotional Image Content Generation with Text-to-Image Diffusion Models/layout.json @@ -0,0 +1,7636 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 133, + 103, + 462, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 103, + 462, + 140 + ], + "spans": [ + { + "bbox": [ + 133, + 103, + 462, + 140 + ], + "type": "text", + "content": "EmoGen: Emotional Image Content Generation with Text-to-Image Diffusion Models" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 198, + 161, + 399, + 176 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 198, + 161, + 399, + 176 + ], + "spans": [ + { + "bbox": [ + 198, + 161, + 399, + 176 + ], + "type": "text", + "content": "Jingyuan Yang, Jiawei Feng, Hui Huang*" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 246, + 177, + 347, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 246, + 177, + 347, + 190 + ], + "spans": [ + { + "bbox": [ + 246, + 177, + 347, + 190 + ], + "type": "text", + "content": "Shenzhen University" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 149, + 191, + 441, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 149, + 191, + 441, + 204 + ], + "spans": [ + { + "bbox": [ + 149, + 191, + 441, + 204 + ], + "type": "text", + "content": "{jingyuanyang.jyy, fengjiawei0909, hhzhiyan}@gmail.com" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 65, + 209, + 176, + 327 + ], + "blocks": [ + { + "bbox": [ + 65, + 209, + 176, + 327 + ], + "lines": [ + { + "bbox": [ + 65, + 209, + 176, + 327 + ], + "spans": [ + { + "bbox": [ + 65, + 209, + 176, + 327 + ], + "type": "image", + "image_path": "3fc12973e2e4509d7d55d6a4667bb7fd1e11c39a8516f2e5b365b787e9af60b1.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 183, + 209, + 293, + 326 + ], + "blocks": [ + { + "bbox": [ + 183, + 209, + 293, + 326 + ], + "lines": [ + { + "bbox": [ + 183, + 209, + 293, + 326 + ], + "spans": [ + { + "bbox": [ + 183, + 209, + 293, + 326 + ], + "type": "image", + "image_path": "69d9a7ffc9eaf994bd51691e7c3dcdf7393ac08e185033f3146b3dac0706c75b.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 301, + 209, + 410, + 326 + ], + "blocks": [ + { + "bbox": [ + 301, + 209, + 410, + 326 + ], + "lines": [ + { + "bbox": [ + 301, + 209, + 410, + 326 + ], + "spans": [ + { + "bbox": [ + 301, + 209, + 410, + 326 + ], + "type": "image", + "image_path": "1a3ffcf1c2f079935bf3944a83119a81baa91e5ecba26c7991bfda597ff61615.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 419, + 210, + 529, + 326 + ], + "blocks": [ + { + "bbox": [ + 419, + 210, + 529, + 326 + ], + "lines": [ + { + "bbox": [ + 419, + 210, + 529, + 326 + ], + "spans": [ + { + "bbox": [ + 419, + 210, + 529, + 326 + ], + "type": "image", + "image_path": "91b5b344aa006292105d35ca67e41db0f7a85238bd932ee239522ff8cd48e43c.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 65, + 331, + 176, + 449 + ], + "blocks": [ + { + "bbox": [ + 65, + 331, + 176, + 449 + ], + "lines": [ + { + "bbox": [ + 65, + 331, + 176, + 449 + ], + "spans": [ + { + "bbox": [ + 65, + 331, + 176, + 449 + ], + "type": "image", + "image_path": "9c3314b79d1222a4519c490cb43c5d562ff6ac46e47638505ce2bbe17c1b0b7a.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 459, + 547, + 482 + ], + "lines": [ + { + "bbox": [ + 46, + 459, + 547, + 482 + ], + "spans": [ + { + "bbox": [ + 46, + 459, + 547, + 482 + ], + "type": "text", + "content": "Figure 1. Emotional Image Content Generation (EICG). Given an emotion category, our network produces images that exhibit unambiguous meanings (semantic-clear), reflect the intended emotion (emotion-faithful) and incorporate varied semantics (semantic-diverse)." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 183, + 331, + 293, + 449 + ], + "blocks": [ + { + "bbox": [ + 183, + 331, + 293, + 449 + ], + "lines": [ + { + "bbox": [ + 183, + 331, + 293, + 449 + ], + "spans": [ + { + "bbox": [ + 183, + 331, + 293, + 449 + ], + "type": "image", + "image_path": "158a4ac8a6ec0c14f6519a00a14c99a86ac1e012f548eab7c2971d580ad69c94.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 301, + 331, + 410, + 448 + ], + "blocks": [ + { + "bbox": [ + 301, + 331, + 410, + 448 + ], + "lines": [ + { + "bbox": [ + 301, + 331, + 410, + 448 + ], + "spans": [ + { + "bbox": [ + 301, + 331, + 410, + 448 + ], + "type": "image", + "image_path": "97ce0c88ceeb33497b4ceb32c5b00b94fe8967464248edd70559c9c7e1acf3fb.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 419, + 331, + 529, + 448 + ], + "blocks": [ + { + "bbox": [ + 419, + 331, + 529, + 448 + ], + "lines": [ + { + "bbox": [ + 419, + 331, + 529, + 448 + ], + "spans": [ + { + "bbox": [ + 419, + 331, + 529, + 448 + ], + "type": "image", + "image_path": "aec73d73247d35cafd079ad62262717cc1ac515a3e4cc52c8657dbd1dc9d78e9.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "bbox": [ + 143, + 491, + 192, + 504 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 491, + 192, + 504 + ], + "spans": [ + { + "bbox": [ + 143, + 491, + 192, + 504 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 45, + 509, + 290, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 509, + 290, + 689 + ], + "spans": [ + { + "bbox": [ + 45, + 509, + 290, + 689 + ], + "type": "text", + "content": "Recent years have witnessed remarkable progress in image generation task, where users can create visually astonishing images with high-quality. However, existing text-to-image diffusion models are proficient in generating concrete concepts (dogs) but encounter challenges with more abstract ones (emotions). Several efforts have been made to modify image emotions with color and style adjustments, facing limitations in effectively conveying emotions with fixed image contents. In this work, we introduce Emotional Image Content Generation (EICG), a new task to generate semantic-clear and emotion-faithful images given emotion categories. Specifically, we propose an emotion space and construct a mapping network to align it with the powerful Contrastive Language-Image Pre-training (CLIP) space, providing a concrete interpretation of abstract emo" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 493, + 547, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 493, + 547, + 601 + ], + "spans": [ + { + "bbox": [ + 306, + 493, + 547, + 601 + ], + "type": "text", + "content": "tions. Attribute loss and emotion confidence are further proposed to ensure the semantic diversity and emotion fidelity of the generated images. Our method outperforms the state-of-the-art text-to-image approaches both quantitatively and qualitatively, where we derive three custom metrics, i.e., emotion accuracy, semantic clarity and semantic diversity. In addition to generation, our method can help emotion understanding and inspire emotional art design. Project page: https://vcc.tech/research/2024/EmoGen." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 306, + 620, + 387, + 632 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 620, + 387, + 632 + ], + "spans": [ + { + "bbox": [ + 306, + 620, + 387, + 632 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 335, + 639, + 517, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 639, + 517, + 651 + ], + "spans": [ + { + "bbox": [ + 335, + 639, + 517, + 651 + ], + "type": "text", + "content": "\"What I cannot create, I do not understand.\"" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 466, + 651, + 545, + 663 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 466, + 651, + 545, + 663 + ], + "spans": [ + { + "bbox": [ + 466, + 651, + 545, + 663 + ], + "type": "text", + "content": "Richard Feynman" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 665, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 547, + 715 + ], + "type": "text", + "content": "Emotions, often elusive yet profoundly influential, shape our actions, foster connections, and spark passions. With the prevalence of social medias, users tend to share specially crafted images to express their feelings. Aiming to find out" + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 703, + 135, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 703, + 135, + 712 + ], + "spans": [ + { + "bbox": [ + 58, + 703, + 135, + 712 + ], + "type": "text", + "content": "*Corresponding author" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6358" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 144 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 144 + ], + "type": "text", + "content": "people's emotional responses towards different stimuli, Visual Emotion Analysis (VEA) is an intriguing yet challenging task in computer vision [34, 50, 51]. Recent years have witnessed rapid development in this field, bringing potential applications such as opinion mining [48], market advertising [6] and mental healthcare [17]." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 146, + 286, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 146, + 286, + 266 + ], + "spans": [ + { + "bbox": [ + 46, + 146, + 286, + 266 + ], + "type": "text", + "content": "Thanks to the advent of diffusion models [7, 16, 38], unprecedented progress has been made in text-to-image generation, where users can generate high-quality images with crafted prompts or personalized objects [10, 39, 56]. Existing text-to-image diffusion models, are often excel in generating concrete concepts (e.g., cat, house, mountain) but face limitations when tasked with more abstract ones (e.g., amusement, anger, sadness). In reality, however, photographic works are not necessarily targeted on specific entities, but are often composed to convey certain feelings." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 267, + 286, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 267, + 286, + 434 + ], + "spans": [ + { + "bbox": [ + 46, + 267, + 286, + 434 + ], + "type": "text", + "content": "A natural question arises: What if machines could create images that not only please our eyes but also touch our hearts? Generating emotions is very challenging. Emotions are abstract while images are concrete, leaving the affective gap [13] hard to surmount. To bridge the gap, several efforts have been made to modify visual emotions by adjusting colors and styles, i.e., image emotion transfer [30, 41, 47]. These methods, however, meet difficulties in evoking emotions correctly and significantly, i.e., " + }, + { + "bbox": [ + 46, + 267, + 286, + 434 + ], + "type": "inline_equation", + "content": "29\\%" + }, + { + "bbox": [ + 46, + 267, + 286, + 434 + ], + "type": "text", + "content": " emotion accuracy [47], as fixed image contents limit emotional variations. Moreover, we cannot generate emotional images solely from colors and styles. What truly triggers emotion? Psychological studies show that visual emotions are often evoked by specific semantics [1, 3, 4]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 436, + 286, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 436, + 286, + 639 + ], + "spans": [ + { + "bbox": [ + 46, + 436, + 286, + 639 + ], + "type": "text", + "content": "In this paper, we propose Emotional Image Content Generation (EICG), a new task to generate semantically clear and emotionally faithful visual contents conditioned on a given emotion category, as shown in Figure 1. Semantic clarity demands an unambiguous representation of visual contents, while emotion faithfulness entails generating images evoke the intended emotions. Contrastive Language-Image Pre-training (CLIP) [31] is a large-scale vision-language model with rich semantics. However, we observe in Figure 2 that CLIP space can not well capture emotional relationships. Therefore, we introduce an emotion space, which groups similar emotions together while keeping dissimilar ones apart. While emotion space excels in representing emotions, CLIP space exhibits a powerful semantic structure. To align emotion space with CLIP space, we propose a mapping network, interpreting abstract emotions with concrete semantics." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 642, + 286, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 642, + 286, + 712 + ], + "spans": [ + { + "bbox": [ + 46, + 642, + 286, + 712 + ], + "type": "text", + "content": "EmoSet [53] is a recently proposed large-scale visual emotion dataset with rich attributes. The Latent Diffusion Model (LDM) loss [38] is often utilized to optimize concrete entities with single and explicit semantics, posing a challenge in capturing the diversity within each emotion. To address this, we introduce an attribute loss to ensure seman" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 310, + 72, + 425, + 188 + ], + "blocks": [ + { + "bbox": [ + 310, + 72, + 425, + 188 + ], + "lines": [ + { + "bbox": [ + 310, + 72, + 425, + 188 + ], + "spans": [ + { + "bbox": [ + 310, + 72, + 425, + 188 + ], + "type": "image", + "image_path": "a4ce53800202aec8a914937794b0b8f5d576aaaca2c2400d8161dd98a0244f6b.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 341, + 190, + 394, + 198 + ], + "lines": [ + { + "bbox": [ + 341, + 190, + 394, + 198 + ], + "spans": [ + { + "bbox": [ + 341, + 190, + 394, + 198 + ], + "type": "text", + "content": "(a) Emotion Space" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 305, + 210, + 545, + 243 + ], + "lines": [ + { + "bbox": [ + 305, + 210, + 545, + 243 + ], + "spans": [ + { + "bbox": [ + 305, + 210, + 545, + 243 + ], + "type": "text", + "content": "Figure 2. Despite (b) CLIP space demonstrates a powerful semantic structure, it struggles to effectively capture emotional relationships within (a), the proposed emotion space." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 427, + 72, + 542, + 188 + ], + "blocks": [ + { + "bbox": [ + 427, + 72, + 542, + 188 + ], + "lines": [ + { + "bbox": [ + 427, + 72, + 542, + 188 + ], + "spans": [ + { + "bbox": [ + 427, + 72, + 542, + 188 + ], + "type": "image", + "image_path": "d05b48c7b7f51571c138f86093c71efeb0a5d825959e0721fb579b67e46b6299.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 463, + 190, + 507, + 197 + ], + "lines": [ + { + "bbox": [ + 463, + 190, + 507, + 197 + ], + "spans": [ + { + "bbox": [ + 463, + 190, + 507, + 197 + ], + "type": "text", + "content": "(b)CLIP Space" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 257, + 545, + 304 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 257, + 545, + 304 + ], + "spans": [ + { + "bbox": [ + 304, + 257, + 545, + 304 + ], + "type": "text", + "content": "tic clarity and diversity, by leveraging the attribute labels in EmoSet. Recognizing that not all objects are affective, emotion confidence is further proposed to ensure the emotion fidelity of the generated contents." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 308, + 546, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 308, + 546, + 548 + ], + "spans": [ + { + "bbox": [ + 304, + 308, + 546, + 548 + ], + "type": "text", + "content": "To estimate the generation quality of EICG, three evaluation metrics are specially designed: emotion accuracy, semantic clarity and semantic diversity. As EICG aims to create emotional contents, we design emotion accuracy to measure the alignment between intended and perceived emotions in the generated images. People are prone to evoke emotions only when the contents are easily recognizable. Thus we propose semantic clarity to assess the unambiguity of the generated image content. Additionally, in view of the assorted emotion stimuli, we devise semantic diversity to quantify the content richness under each emotion. We evaluate our method through both qualitative and quantitative analyses, surpassing the state-of-the-art text-to-image generation approaches across five metrics. Ablation studies are performed to verify the network design, and user studies are conducted to resonate our method with human viewers. Besides generation task, our method can also be applied to decompose emotion concepts, transfer emotional contents and fuse different emotions, which may be helpful to understand emotions and create emotional art design." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 317, + 552, + 459, + 563 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 552, + 459, + 563 + ], + "spans": [ + { + "bbox": [ + 317, + 552, + 459, + 563 + ], + "type": "text", + "content": "In summary, our contributions are:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 567, + 545, + 711 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 306, + 567, + 545, + 614 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 567, + 545, + 614 + ], + "spans": [ + { + "bbox": [ + 306, + 567, + 545, + 614 + ], + "type": "text", + "content": "- We introduce Emotional Image Content Generation, a novel task to generate emotion-faithful and semantic-clear image contents. We also derive three custom metrics to estimate the generation performance." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 616, + 545, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 616, + 545, + 662 + ], + "spans": [ + { + "bbox": [ + 306, + 616, + 545, + 662 + ], + "type": "text", + "content": "- We develop a mapping network to align the proposed emotion space to the powerful CLIP space, where attribute loss and emotion confidence are further designed to ensure the semantic richness and emotion fidelity." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 663, + 545, + 711 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 663, + 545, + 711 + ], + "spans": [ + { + "bbox": [ + 306, + 663, + 545, + 711 + ], + "type": "text", + "content": "- We evaluate our method against the state-of-the-art text-to-image approaches and demonstrate our superiority. Potential applications are exhibited for emotion understanding and emotional art design." + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "6359" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 132, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 132, + 83 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 132, + 83 + ], + "type": "text", + "content": "2. Related work" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 91, + 185, + 104 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 91, + 185, + 104 + ], + "spans": [ + { + "bbox": [ + 47, + 91, + 185, + 104 + ], + "type": "text", + "content": "2.1. Visual Emotion Analysis" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 110, + 289, + 446 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 110, + 289, + 446 + ], + "spans": [ + { + "bbox": [ + 47, + 110, + 289, + 446 + ], + "type": "text", + "content": "Researchers have been involved in VEA for over two decades, ranging from early traditional approaches [2, 23, 26] to recent deep learning ones [35, 51, 52, 58]. Given the inherent abstractness and complexity of visual emotion, researchers aim to identify the most influential elements, which range from low-level features like color, texture and style [23, 26, 35, 58] to high-level semantics [2, 35, 51, 52, 58]. Lee et al. [23] propose a scheme to evaluate emotional response from color images by reasoning the prototypical color for each emotion and the input images. As a milestone, Machajdik et al. [26] extract representative low-level features in composition, including color and texture, to predict visual emotions. Besides low-level features, Borth et al. [2] propose Adjective-Noun Pair (ANP) and build a visual concept detector named Sentibank. With the help of deep learning techniques, Rao et al. [35] construct MldrNet to extract emotional clues from pixel-level, aesthetic-level and semantic-level. To form a more discriminative emotional representation, Zhang et al. [58] integrate high-level contents and low-level styles. Yang et al. propose network to mine emotions from multiple objects [52] as well as object-scene correlations [51]. Existing work often treat VEA as a classification task, i.e., input an image and predict the emotion within it. Can we reverse this process? In other words, can we generate an image targeting on the given emotion word? Only by creating emotional images can we demonstrate the significance of visual elements, leading to a better understanding of emotions." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 455, + 192, + 467 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 455, + 192, + 467 + ], + "spans": [ + { + "bbox": [ + 47, + 455, + 192, + 467 + ], + "type": "text", + "content": "2.2. Text-to-Image Generation" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 475, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 475, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 47, + 475, + 289, + 715 + ], + "type": "text", + "content": "Text-to-image generation aims to convert textual descriptions into corresponding realistic images. Existing generative models can be grouped into GANs [12, 24, 59], VAEs [9, 20, 55], flow-based [37], energy-based [22] and diffusion-based [7, 16, 38, 39, 56]. Diffusion models are witnessed impressive and rapid progress in recent years, where methods like GLIDE [28], DALLE2 [32], Image [40] can generate diverse, photo-realistic and high-quality images. Notably, Stable diffusion [38] is one of the most popular diffusion models, owing to its stable training and the capability for fine-grained control, supported by an active user community. For customized generation, several diffusion-based text-to-image methods are introduced, where methods vary from learning a new embedding [8, 10] and finetuning the network parameters [21, 39, 46]. Textual inversion [10] and DreamArtist [8] learn new concepts with a few user-provided images in the word embedding space, without further training on diffusion models. While DreamBooth [39] finetunes all the parameters to learn a new concept, Custom diffusion [21] only updates the key" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 72, + 547, + 194 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 194 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 194 + ], + "type": "text", + "content": "and value parameters in the cross attention layers. Further, ELITE [46] speeds up the running time with accurate generation results by adopting a global and local mapping network. Existing text-to-images models are capable of generating concrete concepts [7, 24, 56], or personalized ones [10, 21, 39], but face difficulties in generating more abstract ones. In reality, photographic works are not necessarily composed of targeted concepts, but often aim to convey specific feelings. Thus, how to generate emotion-evoking images remains a pressing and critical challenge." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 305, + 201, + 444, + 213 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 201, + 444, + 213 + ], + "spans": [ + { + "bbox": [ + 305, + 201, + 444, + 213 + ], + "type": "text", + "content": "2.3. Image Emotion Transfer" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 220, + 547, + 519 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 220, + 547, + 519 + ], + "spans": [ + { + "bbox": [ + 304, + 220, + 547, + 519 + ], + "type": "text", + "content": "Image style transfer [11] aims to render the semantic content under different styles, producing visually stunning results [19, 33, 45, 54]. Similarly, image color transfer [36] seeks to adjust and harmonize the color characteristics of one image to match another [18, 29]. Specifically, color and style choices can strongly influence the emotions of an image [27]. By adjusting low-level visual elements, image emotion transfer aims to modify the emotional tone of the input image, including the color-based methods [5, 25, 30, 44, 49, 60] and the style-based ones [41, 47]. Yang and Peng et al. [49] makes the first attempt to transfer image colors. Wang et al. [44] present a system to modify the image color according to a given emotion word, and Liu et al. [25] further advance it with deep learning techniques. Peng et al. [30] introduce a new approach to alter the emotion of an input image by guiding its color and texture under the target image. More recently, to reflect emotions in styles, Sun et al. [41] and Weng et al. [47] bring promising results on emotion-aware image style transfer. Nevertheless, the alteration of visual emotions through colors and styles is limited due to fixed content, resulting in subtle emotional changes, i.e., " + }, + { + "bbox": [ + 304, + 220, + 547, + 519 + ], + "type": "inline_equation", + "content": "29\\%" + }, + { + "bbox": [ + 304, + 220, + 547, + 519 + ], + "type": "text", + "content": " emotion accuracy in [47]. Psychological studies suggest that visual emotions can be elicited by specific semantics [3]. Thus, we propose a novel method to generate emotional image contents with clear semantics." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 531, + 362, + 544 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 531, + 362, + 544 + ], + "spans": [ + { + "bbox": [ + 306, + 531, + 362, + 544 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 552, + 443, + 563 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 552, + 443, + 563 + ], + "spans": [ + { + "bbox": [ + 306, + 552, + 443, + 563 + ], + "type": "text", + "content": "3.1. Emotion Representation" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 570, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 547, + 713 + ], + "type": "text", + "content": "Emotion Space EICG is a challenging task, which requires both semantic clarity and emotion fidelity. How to generate an image with distinct and emotional semantics? CLIP [31] is developed to align image and text modalities, where semantically related features are located in close proximity to each other. While CLIP shows impressive semantic representation capabilities, it struggles to effectively capture emotional relationships. As demonstrated in Figure 2, we can observe that sample points with emotional similarities are distantly separated within the CLIP space due to their differing semantics, e.g., toy, amusement park and Christmas tree. To better depict emotional relation" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "6360" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 78, + 69, + 515, + 202 + ], + "blocks": [ + { + "bbox": [ + 78, + 69, + 515, + 202 + ], + "lines": [ + { + "bbox": [ + 78, + 69, + 515, + 202 + ], + "spans": [ + { + "bbox": [ + 78, + 69, + 515, + 202 + ], + "type": "image", + "image_path": "55fe37c14368b504bde71e5a8fc55fb3d6f580460f8ec9755dcc558656664520.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 207, + 547, + 232 + ], + "lines": [ + { + "bbox": [ + 46, + 207, + 547, + 232 + ], + "spans": [ + { + "bbox": [ + 46, + 207, + 547, + 232 + ], + "type": "text", + "content": "Figure 3. Training process of our network. Emotion representation (stage 1) learns a well-behaved emotion space and emotion content generation (stage 2) maps this space to CLIP space, aiming to generate image contents with emotion fidelity, semantic clarity and diversity." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 241, + 289, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 241, + 289, + 350 + ], + "spans": [ + { + "bbox": [ + 46, + 241, + 289, + 350 + ], + "type": "text", + "content": "ships, we introduce the emotion space, a latent space that clusters similar emotions together while keeping dissimilar ones apart. EmoSet [53] is a large-scale dataset with rich attributes, where each image is labeled with an emotion. Using aligned image-emotion pairs, we construct an encoder " + }, + { + "bbox": [ + 46, + 241, + 289, + 350 + ], + "type": "inline_equation", + "content": "\\varphi" + }, + { + "bbox": [ + 46, + 241, + 289, + 350 + ], + "type": "text", + "content": " with ResNet-50 [14] to capture emotion representations. To train the encoder, we devise an emotion loss by implementing the widely-used Cross-Entropy (CE) loss, following the previous work [50, 51]:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 75, + 354, + 287, + 386 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 354, + 287, + 386 + ], + "spans": [ + { + "bbox": [ + 75, + 354, + 287, + 386 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {e m o} = - \\sum_ {i = 1} ^ {C} y _ {e m o} \\log \\frac {\\exp (\\varphi (x , i))}{\\sum_ {i = 1} ^ {C} \\exp (\\varphi (x , i))}, \\tag {1}", + "image_path": "df118bd0de3236a595919ab3e5df8613e15daa82070eecaa354c8ef0b58ccf70.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 391, + 287, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 391, + 287, + 450 + ], + "spans": [ + { + "bbox": [ + 46, + 391, + 287, + 450 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 391, + 287, + 450 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 46, + 391, + 287, + 450 + ], + "type": "text", + "content": " represents the input image, " + }, + { + "bbox": [ + 46, + 391, + 287, + 450 + ], + "type": "inline_equation", + "content": "y_{emo}" + }, + { + "bbox": [ + 46, + 391, + 287, + 450 + ], + "type": "text", + "content": " denotes the emotion label and " + }, + { + "bbox": [ + 46, + 391, + 287, + 450 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 46, + 391, + 287, + 450 + ], + "type": "text", + "content": " stands for the total number of emotion categories. Once the loss function converges, emotion space is established. Parameters in emotion encoder remain fixed in the following emotional content generation process." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 450, + 288, + 547 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 450, + 288, + 547 + ], + "spans": [ + { + "bbox": [ + 46, + 450, + 288, + 547 + ], + "type": "text", + "content": "During inference, each emotion cluster is represented by a Gaussian distribution with learned parameters, i.e., mean and standard deviation. For example, when taking amusement as input, we randomly sample a data point from corresponding Gaussian distribution to serve as its emotion representation, as shown in Figure 3. We have confirmed that Gaussian distribution suits emotion clusters well and the random sampling process induces diversity to EICG." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 552, + 216, + 563 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 552, + 216, + 563 + ], + "spans": [ + { + "bbox": [ + 47, + 552, + 216, + 563 + ], + "type": "text", + "content": "3.2. Emotional Content Generation" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 570, + 287, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 570, + 287, + 665 + ], + "spans": [ + { + "bbox": [ + 46, + 570, + 287, + 665 + ], + "type": "text", + "content": "Mapping Network While emotion space is emotionally separable, CLIP space captures rich semantics. Existing text-to-image models entail clear and specific semantics as input, making CLIP space indispensable in the generation process. Consequently, establishing the mapping between emotion space and CLIP space becomes a crucial challenge. Intuitively, we attempt to build the mapping network using fully connected layers, following previous work [33, 43]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 665, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 288, + 714 + ], + "type": "text", + "content": "However, as depicted in Figure 2, clustered feature points in the emotion space are expected to disperse in the CLIP space to capture diverse semantics. Therefore, we utilize a Multilayer Perceptron (MLP) to build the mapping" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 241, + 547, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 241, + 547, + 361 + ], + "spans": [ + { + "bbox": [ + 304, + 241, + 547, + 361 + ], + "type": "text", + "content": "network, incorporating non-linear operations, i.e., RELU, to facilitate the separation process. The non-linear projection " + }, + { + "bbox": [ + 304, + 241, + 547, + 361 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 304, + 241, + 547, + 361 + ], + "type": "text", + "content": " is succeeded by a CLIP text transformer " + }, + { + "bbox": [ + 304, + 241, + 547, + 361 + ], + "type": "inline_equation", + "content": "t_{\\theta}" + }, + { + "bbox": [ + 304, + 241, + 547, + 361 + ], + "type": "text", + "content": ", yielding textual embedding for U-Net. The end-token embedding of the transformer's output is passed through a fully-connected layer, producing the CLIP text feature. Particularly, to better preserve the prior knowledge in the CLIP space, parameters in the transformer and linear projection are kept frozen, while parameters in non-linear projection are learned, as depicted in Figure 3." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 374, + 547, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 374, + 547, + 411 + ], + "spans": [ + { + "bbox": [ + 304, + 374, + 547, + 411 + ], + "type": "text", + "content": "Attribute Loss Existing text-to-image diffusion models often employ Latent Diffusion Model (LDM) loss [38] for optimization process [10, 39, 56]:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 315, + 414, + 545, + 435 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 414, + 545, + 435 + ], + "spans": [ + { + "bbox": [ + 315, + 414, + 545, + 435 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {L D M} = \\mathbb {E} _ {z, x, \\epsilon , t} \\left[ \\| \\epsilon - \\epsilon_ {\\theta} (z _ {t}, t, t _ {\\theta} (F (\\varphi (x)))) \\| _ {2} ^ {2} \\right], \\tag {2}", + "image_path": "a1ae452cec5bdb33eeadb1271fc13a5bd2f90b226d190eeef2197096d9579712.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 438, + 547, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 438, + 547, + 462 + ], + "spans": [ + { + "bbox": [ + 304, + 438, + 547, + 462 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 438, + 547, + 462 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 304, + 438, + 547, + 462 + ], + "type": "text", + "content": " represents the added noise, " + }, + { + "bbox": [ + 304, + 438, + 547, + 462 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}" + }, + { + "bbox": [ + 304, + 438, + 547, + 462 + ], + "type": "text", + "content": " denotes the denoising network and " + }, + { + "bbox": [ + 304, + 438, + 547, + 462 + ], + "type": "inline_equation", + "content": "z_{t}" + }, + { + "bbox": [ + 304, + 438, + 547, + 462 + ], + "type": "text", + "content": " indicates the latent noised to time " + }, + { + "bbox": [ + 304, + 438, + 547, + 462 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 438, + 547, + 462 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 462, + 546, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 462, + 546, + 689 + ], + "spans": [ + { + "bbox": [ + 304, + 462, + 546, + 689 + ], + "type": "text", + "content": "In these cases, target concepts typically involve concrete entities (e.g., dog, car, flower) or personalized objects (e.g., someone's corgi). These concepts often exhibit consistency on semantic level and share certain similarities on pixel level. However, emotions are abstract concepts, where multiple semantics coexist under one specific category. Learning emotions solely with LDM loss may pose some challenges. For one thing, each emotion might collapse to a specific semantic point, e.g., amusement collapsing to amusement park, losing intra-class diversity. In reality, semantics within one emotion are diverse, where single point cannot fully capture. Moreover, since LDM loss is designed to reconstruct the input image, it primarily focuses on learning and preserving pixel-level commonalities such as color and texture. In Figure 4 (a), with LDM loss alone, CLIP embedding for amusement is prone to be colorful, without exhibiting explicit and diverse semantics. We can conclude that it is hard to achieve robust emotion representations in CLIP space by implementing LDM loss alone." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 689, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 689, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 305, + 689, + 547, + 714 + ], + "type": "text", + "content": "In the pursuit of clear and diverse contents, semantics guidance is essential for the generation process. Thanks to" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "6361" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 71, + 274, + 208 + ], + "blocks": [ + { + "bbox": [ + 59, + 71, + 274, + 208 + ], + "lines": [ + { + "bbox": [ + 59, + 71, + 274, + 208 + ], + "spans": [ + { + "bbox": [ + 59, + 71, + 274, + 208 + ], + "type": "image", + "image_path": "26e2189a5806e3c24af1b9000267e94f9341a5a03dda4f02a746e96920ad2454.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 212, + 287, + 245 + ], + "lines": [ + { + "bbox": [ + 46, + 212, + 287, + 245 + ], + "spans": [ + { + "bbox": [ + 46, + 212, + 287, + 245 + ], + "type": "text", + "content": "Figure 4. Motivation for loss function design. Compare to (a) LDM loss alone, (b) attribute loss enhances semantic clarity while (c) emotion confidence ensures emotion accuracy." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 256, + 289, + 401 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 256, + 289, + 401 + ], + "spans": [ + { + "bbox": [ + 46, + 256, + 289, + 401 + ], + "type": "text", + "content": "the rich attribute annotations in EmoSet, we select the mid-level attributes, i.e., object class and scene type to guide the generation process. With this semantic guidance, we formulate an attribute loss to guarantee that the generated image contents possess clear and diverse semantics. For clarity, emotions are easily triggered in people only when visual contents are represented in an unambiguous manner. Considering the varied emotional stimuli in reality, attribute loss guides the network to learn multiple semantics under one specific emotion. Our attribute loss is devised on CLIP space, by calculating the cosine similarities " + }, + { + "bbox": [ + 46, + 256, + 289, + 401 + ], + "type": "inline_equation", + "content": "f(\\cdot)" + }, + { + "bbox": [ + 46, + 256, + 289, + 401 + ], + "type": "text", + "content": " and optimizing a symmetric CE loss over the similarity scores [31]:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 409, + 287, + 445 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 409, + 287, + 445 + ], + "spans": [ + { + "bbox": [ + 55, + 409, + 287, + 445 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {a t t r} = - \\sum_ {j = 1} ^ {K} y _ {a t t r} \\log \\frac {\\exp (f (v _ {e m o} , \\tau_ {\\theta} (a _ {j})))}{\\sum_ {j = 1} ^ {K} \\exp (f (v _ {e m o} , \\tau_ {\\theta} (a _ {j})))}, \\quad (3)", + "image_path": "0f1440985b5c81623302f36482e9d81174584483b37580ae8937940044f415be.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 125, + 462, + 287, + 488 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 125, + 462, + 287, + 488 + ], + "spans": [ + { + "bbox": [ + 125, + 462, + 287, + 488 + ], + "type": "interline_equation", + "content": "f (p, q) = \\frac {p \\cdot q}{\\| p \\| \\| q \\|}, \\tag {4}", + "image_path": "4601e3164eec6c0d1f4d33bb0a80e9017d0b63a1db1ad4b392c9dae6b5e47232.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 495, + 287, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 495, + 287, + 592 + ], + "spans": [ + { + "bbox": [ + 46, + 495, + 287, + 592 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 495, + 287, + 592 + ], + "type": "inline_equation", + "content": "a_{j}" + }, + { + "bbox": [ + 46, + 495, + 287, + 592 + ], + "type": "text", + "content": " denotes the " + }, + { + "bbox": [ + 46, + 495, + 287, + 592 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 46, + 495, + 287, + 592 + ], + "type": "text", + "content": " member in the attribute set, " + }, + { + "bbox": [ + 46, + 495, + 287, + 592 + ], + "type": "inline_equation", + "content": "\\tau_{\\theta}" + }, + { + "bbox": [ + 46, + 495, + 287, + 592 + ], + "type": "text", + "content": " represents the text encoder, " + }, + { + "bbox": [ + 46, + 495, + 287, + 592 + ], + "type": "inline_equation", + "content": "v_{emo}" + }, + { + "bbox": [ + 46, + 495, + 287, + 592 + ], + "type": "text", + "content": " implies the learned CLIP embedding and " + }, + { + "bbox": [ + 46, + 495, + 287, + 592 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 46, + 495, + 287, + 592 + ], + "type": "text", + "content": " indicates the total number of the attributes. With the attribute loss, each sample point is converging towards the correct semantic and distancing itself from the incorrect ones. Through the combination of attribute loss and LDM loss, we can effectively map each emotion to clear and diverse semantics, as demonstrated in Figure 4 (b)." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "text", + "content": "Emotion Confidence However, it is worth noting that some of the semantics in Figure 4 (b) appear emotionally neutral, e.g., plant and tree. Since attributes are annotated objectively, not all the attributes in EmoSet are emotional. Therefore, we propose emotion confidence to measure the correlations between emotions and semantic attributes. Initially, we gather all images associated with attribute " + }, + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "text", + "content": " in EmoSet and send them to a pre-trained emotion classifier. Each image is predicted as an emotion vector " + }, + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "inline_equation", + "content": "p(\\cdot)" + }, + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "text", + "content": " and we" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 318, + 70, + 533, + 164 + ], + "blocks": [ + { + "bbox": [ + 318, + 70, + 533, + 164 + ], + "lines": [ + { + "bbox": [ + 318, + 70, + 533, + 164 + ], + "spans": [ + { + "bbox": [ + 318, + 70, + 533, + 164 + ], + "type": "image", + "image_path": "28e43b6a69be182719839923a26db0c9befa21c76c8cabb777821c2a4a8dfddc.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 168, + 545, + 191 + ], + "lines": [ + { + "bbox": [ + 305, + 168, + 545, + 191 + ], + "spans": [ + { + "bbox": [ + 305, + 168, + 545, + 191 + ], + "type": "text", + "content": "Figure 5. Illustration of emotion confidence. Each (a) attribute is represented by (b) a distribution of confidence on eight emotions." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 199, + 545, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 199, + 545, + 236 + ], + "spans": [ + { + "bbox": [ + 305, + 199, + 545, + 236 + ], + "type": "text", + "content": "sum all images up to get the emotional distribution " + }, + { + "bbox": [ + 305, + 199, + 545, + 236 + ], + "type": "inline_equation", + "content": "d_{j}" + }, + { + "bbox": [ + 305, + 199, + 545, + 236 + ], + "type": "text", + "content": " for attribute " + }, + { + "bbox": [ + 305, + 199, + 545, + 236 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 305, + 199, + 545, + 236 + ], + "type": "text", + "content": ". Each emotion " + }, + { + "bbox": [ + 305, + 199, + 545, + 236 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 305, + 199, + 545, + 236 + ], + "type": "text", + "content": " within this distribution is assigned a corresponding emotion confidence " + }, + { + "bbox": [ + 305, + 199, + 545, + 236 + ], + "type": "inline_equation", + "content": "\\alpha_{ij}" + }, + { + "bbox": [ + 305, + 199, + 545, + 236 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 375, + 242, + 545, + 276 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 375, + 242, + 545, + 276 + ], + "spans": [ + { + "bbox": [ + 375, + 242, + 545, + 276 + ], + "type": "interline_equation", + "content": "\\alpha_ {i j} = \\frac {1}{N _ {j}} \\sum_ {n = 1} ^ {N _ {j}} p (x _ {n}, i), \\tag {5}", + "image_path": "0188fbabcc5be0b98d3285fa3709457a3d597b929e7733483ce65942bc3c498e.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 281, + 545, + 400 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 281, + 545, + 400 + ], + "spans": [ + { + "bbox": [ + 304, + 281, + 545, + 400 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 281, + 545, + 400 + ], + "type": "inline_equation", + "content": "x_{n}" + }, + { + "bbox": [ + 304, + 281, + 545, + 400 + ], + "type": "text", + "content": " represents the input image and " + }, + { + "bbox": [ + 304, + 281, + 545, + 400 + ], + "type": "inline_equation", + "content": "N_{j}" + }, + { + "bbox": [ + 304, + 281, + 545, + 400 + ], + "type": "text", + "content": " denotes the total image number in attribute " + }, + { + "bbox": [ + 304, + 281, + 545, + 400 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 304, + 281, + 545, + 400 + ], + "type": "text", + "content": ". We further illustrate the above process in Figure 5 with visual representations. When mountain snowy appears, people are more likely to experience awe and cemetery often elicits sadness. In contrast, the presence of tree in every emotion category suggests its lack of emotional specificity. Some attributes are emotion-related while others are not, which can be beneficial for generating emotional contents. We then use emotion confidence to balance between LDM loss and attribute loss:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 354, + 407, + 545, + 421 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 354, + 407, + 545, + 421 + ], + "spans": [ + { + "bbox": [ + 354, + 407, + 545, + 421 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\left(1 - \\alpha_ {i j}\\right) \\mathcal {L} _ {L D M} + \\alpha_ {i j} \\mathcal {L} _ {\\text {a t t r}}, \\tag {6}", + "image_path": "cdadbdd226564f9a46948ba7840cc8709423c799df27b85c3dce2455526adced.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 427, + 546, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 427, + 546, + 559 + ], + "spans": [ + { + "bbox": [ + 304, + 427, + 546, + 559 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 427, + 546, + 559 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 427, + 546, + 559 + ], + "type": "text", + "content": " represents the emotion category " + }, + { + "bbox": [ + 304, + 427, + 546, + 559 + ], + "type": "inline_equation", + "content": "y_{emo}" + }, + { + "bbox": [ + 304, + 427, + 546, + 559 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 427, + 546, + 559 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 304, + 427, + 546, + 559 + ], + "type": "text", + "content": " denotes the attribute type " + }, + { + "bbox": [ + 304, + 427, + 546, + 559 + ], + "type": "inline_equation", + "content": "y_{att}" + }, + { + "bbox": [ + 304, + 427, + 546, + 559 + ], + "type": "text", + "content": ". The greater the emotion confidence " + }, + { + "bbox": [ + 304, + 427, + 546, + 559 + ], + "type": "inline_equation", + "content": "\\alpha_{ij}" + }, + { + "bbox": [ + 304, + 427, + 546, + 559 + ], + "type": "text", + "content": " is, the stronger the impact attribute " + }, + { + "bbox": [ + 304, + 427, + 546, + 559 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 304, + 427, + 546, + 559 + ], + "type": "text", + "content": " has on the specific emotion " + }, + { + "bbox": [ + 304, + 427, + 546, + 559 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 427, + 546, + 559 + ], + "type": "text", + "content": ". Low confidence suggests a weak connection between the attribute and emotion, signaling that the network should learn more from the pixel-wise LDM loss. When higher confidence occurs, the network should prioritize the semantic meaning of the image, i.e., the attribute loss. With this design, our network can adapt to a wide range of cases, generating image contents that are both semantically explicit and emotionally faithful, as shown in Figure 4 (c)." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 567, + 388, + 581 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 567, + 388, + 581 + ], + "spans": [ + { + "bbox": [ + 306, + 567, + 388, + 581 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 587, + 439, + 600 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 587, + 439, + 600 + ], + "spans": [ + { + "bbox": [ + 306, + 587, + 439, + 600 + ], + "type": "text", + "content": "4.1. Dataset and Evaluation" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": "Dataset EmoSet [53] is a large-scale visual emotion dataset with rich attributes, comprising a total of 118,102 images. To investigate the connections between emotions and specific contents, we create a subset from EmoSet by preserving images with object/scene labels. Each image is labeled with both emotion and attribute labels, guiding the optimization process of emotion loss and attribute loss. Notably, the wide range of attribute labels assures for learning diverse and representative emotional contents." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6362" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 51, + 70, + 547, + 373 + ], + "blocks": [ + { + "bbox": [ + 51, + 70, + 547, + 373 + ], + "lines": [ + { + "bbox": [ + 51, + 70, + 547, + 373 + ], + "spans": [ + { + "bbox": [ + 51, + 70, + 547, + 373 + ], + "type": "image", + "image_path": "6f2b27521357d6835976cc477fb41cce39cc1ad564a1057d9b5736ac9efb940d.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 59, + 377, + 533, + 389 + ], + "lines": [ + { + "bbox": [ + 59, + 377, + 533, + 389 + ], + "spans": [ + { + "bbox": [ + 59, + 377, + 533, + 389 + ], + "type": "text", + "content": "Figure 6. Qualitative comparisons with the state-of-the-art text-to-image generation approaches and ablation studies of our method." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 395, + 289, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 395, + 289, + 635 + ], + "spans": [ + { + "bbox": [ + 47, + 395, + 289, + 635 + ], + "type": "text", + "content": "Evaluation Metrics To comprehensively evaluate the performance of different methods on EICG task, we utilize commonly used metrics (FID, LPIPS) and design some specific ones (Emo-A, Sem-C, Sem-D). 1) FID: Frechet Inception Distance (FID) [15] quantifies the distribution distance between generated and real images, providing an estimate of image fidelity. 2) LPIPS: Similar to [42], we employ LPIPS [57] to assess the overall image diversity, with higher values indicating better performance. 3) Emo-A: Since EICG aims at creating emotion-evoking images, we design emotion accuracy to assess the emotional alignment between the targeted emotions and the generated images. 4) Sem-C: People are easily to evoke emotions under recognizable contents. We thus introduce semantic clarity to assess the explicitness of generated image contents. 5) Sem-D: Emotions are complex, where each can be triggered by multiple factors. To cover a diverse range of potential scenes or objects, we derive semantic diversity to estimate the content richness associated with each emotion. For more details, please refer to the supplementary materials." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 646, + 133, + 658 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 646, + 133, + 658 + ], + "spans": [ + { + "bbox": [ + 47, + 646, + 133, + 658 + ], + "type": "text", + "content": "4.2. Comparisons" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 665, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 287, + 715 + ], + "type": "text", + "content": "As our method is the first attempt in EICG, we compare it with the most relevant and state-of-the-art text-to-image generation techniques: Stable diffusion [38], Textual inversion [10] and Dreambooth [39]. While Stable diffusion is" + } + ] + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 309, + 419, + 541, + 514 + ], + "blocks": [ + { + "bbox": [ + 305, + 394, + 545, + 415 + ], + "lines": [ + { + "bbox": [ + 305, + 394, + 545, + 415 + ], + "spans": [ + { + "bbox": [ + 305, + 394, + 545, + 415 + ], + "type": "text", + "content": "Table 1. Comparisons with the state-of-the-art methods and ablation studies on emotion generation task, involving five metrics." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 309, + 419, + 541, + 514 + ], + "lines": [ + { + "bbox": [ + 309, + 419, + 541, + 514 + ], + "spans": [ + { + "bbox": [ + 309, + 419, + 541, + 514 + ], + "type": "table", + "html": "
MethodFID ↓LPIPS ↑Emo-A ↑Sem-C ↑Sem-D ↑
Stable Diffusion [38]44.050.68770.77%0.6080.0199
Textual Inversion [10]50.510.70274.87%0.6050.0282
DreamBooth [39]46.890.66170.50%0.6140.0178
Ours41.600.71776.25%0.6330.0335
w/o F57.540.71371.12%0.6150.0261
w/o Lattr51.130.70765.75%0.5920.0270
w/o αij43.300.71474.88%0.5910.0263
", + "image_path": "fd04e9eedb303e1f29ac501c24a40fbcd9a662f66fbfc42e32f7cae6e2c8d68f.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 518, + 545, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 518, + 545, + 544 + ], + "spans": [ + { + "bbox": [ + 305, + 518, + 545, + 544 + ], + "type": "text", + "content": "a general image generation pipeline, Textual inversion and Dreambooth specialize in customized image generation." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 557, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 557, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 557, + 546, + 713 + ], + "type": "text", + "content": "Qualitative Comparisons In Figure 6, our method is qualitatively compared with the state-of-the-art methods across three emotion categories, i.e., awe, anger and contentment. Generation results of the rest five emotions can be found in the supplementary materials. Take awe as an example, all the three compared methods tend to produce images with dense textures and dim colors, which suggests that representations for each emotion may collapse to a single feature point. For anger and contentment, both Stable diffusion and Dreambooth distort the visual representations, e.g., tiger and bicycle, and generate some contents with ambiguous semantics. Though Textual inversion preserves some semantic fidelity, it generates emotion-agnostic" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6363" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 48, + 108, + 287, + 159 + ], + "blocks": [ + { + "bbox": [ + 47, + 70, + 289, + 104 + ], + "lines": [ + { + "bbox": [ + 47, + 70, + 289, + 104 + ], + "spans": [ + { + "bbox": [ + 47, + 70, + 289, + 104 + ], + "type": "text", + "content": "Table 2. User preference study. The numbers indicate the percentage of participants who prefer our results over those compared methods, given the same emotion category as input." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 48, + 108, + 287, + 159 + ], + "lines": [ + { + "bbox": [ + 48, + 108, + 287, + 159 + ], + "spans": [ + { + "bbox": [ + 48, + 108, + 287, + 159 + ], + "type": "table", + "html": "
MethodImage fidelity ↑Emotion faithfulness ↑Semantic diversity ↑
Stable Diffusion67.86±15.08%73.66±11.80%87.88±9.64%
Textual Inversion79.91±16.92%72.75±16.90%85.66±10.51%
DreamBooth77.23±14.00%80.79±8.64%81.68±17.06%
", + "image_path": "b87abdfa5aca660f3ce77f0748e7127db7f78133fb42e6bf163494c608a017a7.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 166, + 289, + 297 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 166, + 289, + 297 + ], + "spans": [ + { + "bbox": [ + 46, + 166, + 289, + 297 + ], + "type": "text", + "content": "contents such as shoes and cars. Since these methods are crafted to learn customized concepts, challenges may arise when handling complex and diverse emotional images. Rather than generating plants and trees, our method can provide diverse and emotion-evoking image contents for awe through lakes, oceans, valleys and snow-covered mountains. In anger, our approach extends beyond mere beasts, encompassing flags, posters, and guns. Owing to attribute loss and emotion confidence, our method can effectively capture the rich and varied semantics while maintaining emotion faithfulness in EmoSet." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 314, + 289, + 494 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 314, + 289, + 494 + ], + "spans": [ + { + "bbox": [ + 46, + 314, + 289, + 494 + ], + "type": "text", + "content": "Quantitative Comparisons As shown in Table 1, the proposed method surpasses the compared methods across all five evaluation metrics. Particularly, better performance on FID and LPIPS indicates our method can generate images with higher fidelity and diversity, effectively capturing the characteristics of the training data. All methods achieve comparable results on emotion accuracy. From Figure 6, we observe that comparison methods are prone to fall into singular or incorrect emotion representations. Even such generation results are still separable in eight classes, they do not conform to human emotional cognition. This suggests that relying solely on Emo-A may be insufficient for EICG task. Therefore, we additionally introduce Sem-C and Sem-D to estimate the content clarity and diversity, where our method exhibits a clear advantage over other methods." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 510, + 289, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 510, + 289, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 510, + 289, + 714 + ], + "type": "text", + "content": "User Study Besides qualitative and quantitative comparisons, we also conduct a user study to determine whether our method is preferred by humans and to understand how people perceive emotions. We invite 14 participants from different social backgrounds and each test session lasts about 30 minutes. In the first part, generation results are evaluated on three dimensions: image fidelity, emotion faithfulness and semantic diversity. Each question presented to the participants includes two sets of images conveying the same emotion, drawn from our method and one of the comparison methods. The participants are then asked: which group is more realistic? which group evokes a stronger sense of [emotion type]? which group is more diverse? As illustrated in Table 2, our method attains the top rankings compared to the other three methods, particularly excelling in semantic diversity. We aim to explore the factors influencing visual emotions in the second part." + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 309, + 70, + 547, + 208 + ], + "blocks": [ + { + "bbox": [ + 309, + 70, + 547, + 208 + ], + "lines": [ + { + "bbox": [ + 309, + 70, + 547, + 208 + ], + "spans": [ + { + "bbox": [ + 309, + 70, + 547, + 208 + ], + "type": "image", + "image_path": "9cee6fd645a6a66e7296391c0540e7c64975abbb174d334b372a99b66aa45f47.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 221, + 547, + 255 + ], + "lines": [ + { + "bbox": [ + 305, + 221, + 547, + 255 + ], + "spans": [ + { + "bbox": [ + 305, + 221, + 547, + 255 + ], + "type": "text", + "content": "Figure 7. Emotion decomposition. Each emotion word is broken down into a set of (a) emotional concepts, reflecting the semantics in (b) generated images." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 307, + 290, + 545, + 352 + ], + "blocks": [ + { + "bbox": [ + 306, + 264, + 545, + 286 + ], + "lines": [ + { + "bbox": [ + 306, + 264, + 545, + 286 + ], + "spans": [ + { + "bbox": [ + 306, + 264, + 545, + 286 + ], + "type": "text", + "content": "Table 3. Comparisons with the state-of-the-art methods on emotion transfer task, involving three metrics." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 307, + 290, + 545, + 352 + ], + "lines": [ + { + "bbox": [ + 307, + 290, + 545, + 352 + ], + "spans": [ + { + "bbox": [ + 307, + 290, + 545, + 352 + ], + "type": "table", + "html": "
MethodEmo-A↑CLIP-img ↑CLIP-text ↑
amusementfearamusementfearamusementfear
Stable Diffusion51.54%56.67%0.9290.8250.2570.251
Textual Inversion60.82%40.00%0.9020.7920.2700.259
Ours72.16%63.33%0.9130.8410.2760.270
", + "image_path": "c2b88aade77b397846f0c19948209b76776aaaf18d95b909ebae15a5782c92e8.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 373, + 545, + 494 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 373, + 545, + 494 + ], + "spans": [ + { + "bbox": [ + 304, + 373, + 545, + 494 + ], + "type": "text", + "content": "Participants are shown an emotional image generated by our method and are asked: which emotion best describes the image? why do you feel such emotion? Compared to the " + }, + { + "bbox": [ + 304, + 373, + 545, + 494 + ], + "type": "inline_equation", + "content": "76.25\\%" + }, + { + "bbox": [ + 304, + 373, + 545, + 494 + ], + "type": "text", + "content": " machine predicted one in Table 1, " + }, + { + "bbox": [ + 304, + 373, + 545, + 494 + ], + "type": "inline_equation", + "content": "82.14\\%" + }, + { + "bbox": [ + 304, + 373, + 545, + 494 + ], + "type": "text", + "content": " emotion accuracy is achieved by user voting, where generated images are more emotion-evoking towards human participants. Additionally, " + }, + { + "bbox": [ + 304, + 373, + 545, + 494 + ], + "type": "inline_equation", + "content": "88.39\\%" + }, + { + "bbox": [ + 304, + 373, + 545, + 494 + ], + "type": "text", + "content": " of the responses indicate that emotions are predominantly triggered by the content/semantic. This underscores how our task, EICG, is closely aligned with human cognition." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 511, + 400, + 525 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 511, + 400, + 525 + ], + "spans": [ + { + "bbox": [ + 306, + 511, + 400, + 525 + ], + "type": "text", + "content": "4.3. Ablation Study" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 534, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 534, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 534, + 545, + 713 + ], + "type": "text", + "content": "We examine the efficacy of each network design, encompassing the non-linear mapping network " + }, + { + "bbox": [ + 304, + 534, + 545, + 713 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 304, + 534, + 545, + 713 + ], + "type": "text", + "content": ", the attribute loss " + }, + { + "bbox": [ + 304, + 534, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{attr}" + }, + { + "bbox": [ + 304, + 534, + 545, + 713 + ], + "type": "text", + "content": " and the emotion confidence " + }, + { + "bbox": [ + 304, + 534, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\alpha_{ij}" + }, + { + "bbox": [ + 304, + 534, + 545, + 713 + ], + "type": "text", + "content": ". In Table 1, without nonlinear mapping network, emotion representations are aggregated, which fails to restore the real image set (high FID) and lacks semantic diversity (low Sem-D). Attribute loss is introduced to enhance semantic clarity and diversity, whose absence leads to performance drops in Sem-C and Sem-D. Besides, as shown in Figure 6, generated images exhibit semantic distortions when attribute loss is absent (w/o " + }, + { + "bbox": [ + 304, + 534, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{attr}" + }, + { + "bbox": [ + 304, + 534, + 545, + 713 + ], + "type": "text", + "content": ") and display explicit contents with attribute loss (w/o " + }, + { + "bbox": [ + 304, + 534, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\alpha_{ij}" + }, + { + "bbox": [ + 304, + 534, + 545, + 713 + ], + "type": "text", + "content": "). While image contents become clear and diverse with attribute loss, it is only with emotion confidence that we can effectively filter out emotion-agnostic semantics and generate images that evoke specific emotions (Ours)." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6364" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 63, + 72, + 306, + 190 + ], + "blocks": [ + { + "bbox": [ + 63, + 72, + 306, + 190 + ], + "lines": [ + { + "bbox": [ + 63, + 72, + 306, + 190 + ], + "spans": [ + { + "bbox": [ + 63, + 72, + 306, + 190 + ], + "type": "image", + "image_path": "bfcb88e00f1b1601d7354f7fc0da355a52539c18e645964841cebd940b05f1bb.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 161, + 191, + 220, + 199 + ], + "lines": [ + { + "bbox": [ + 161, + 191, + 220, + 199 + ], + "spans": [ + { + "bbox": [ + 161, + 191, + 220, + 199 + ], + "type": "text", + "content": "(a) Emotion Transfer" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 306, + 72, + 530, + 190 + ], + "blocks": [ + { + "bbox": [ + 306, + 72, + 530, + 190 + ], + "lines": [ + { + "bbox": [ + 306, + 72, + 530, + 190 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 530, + 190 + ], + "type": "image", + "image_path": "cb74f6f382a80c6fff396e98e9705b86e9e055e9777d1bd656cbfcfc83f2cb27.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 390, + 191, + 445, + 199 + ], + "lines": [ + { + "bbox": [ + 390, + 191, + 445, + 199 + ], + "spans": [ + { + "bbox": [ + 390, + 191, + 445, + 199 + ], + "type": "text", + "content": "(b) Emotion Fusion" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 199, + 545, + 221 + ], + "lines": [ + { + "bbox": [ + 46, + 199, + 545, + 221 + ], + "spans": [ + { + "bbox": [ + 46, + 199, + 545, + 221 + ], + "type": "text", + "content": "Figure 8. Emotion creation. (a) transfers emotion representations (i.e., amusement, fear) to a series of neutral contents while (b) fuse two emotions (i.e., amusement-awe, amusement-fear) together, which may be helpful for emotional art design." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 227, + 129, + 239 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 227, + 129, + 239 + ], + "spans": [ + { + "bbox": [ + 47, + 227, + 129, + 239 + ], + "type": "text", + "content": "4.4. Applications" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 241, + 290, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 241, + 290, + 456 + ], + "spans": [ + { + "bbox": [ + 46, + 241, + 290, + 456 + ], + "type": "text", + "content": "Emotion Decomposition Emotions, serving as abstract concepts, pose a challenge for generative models to understand. Our method provides an opportunity to comprehend visual emotions by identifying the most relevant semantic contents for each emotion. To be specific, we visualize the semantics that are most closely aligned with our emotion representations in CLIP space. Each concept in Figure 7 (a), such as surfboard, bicycle and athletic field, is very likely to elicit excitement, where the corresponding images are presented in Figure 7 (b). Upon viewing such images, we identify the semantics and instinctively link them to specific emotions. These emotional concepts exhibit diversity, explicitness, and a strong capacity to evoke emotions. By decomposing visual emotions, we can not only generate emotional images with various semantics but also gain a deeper understanding of emotion evocation process. The results reveal the close relationship between emotions and semantics, in accordance with the psychological studies [3]." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 465, + 288, + 681 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 465, + 288, + 681 + ], + "spans": [ + { + "bbox": [ + 46, + 465, + 288, + 681 + ], + "type": "text", + "content": "Emotion Transfer Once we identify emotional contents, the next step is to explore how we can use it to create meaningful and compelling designs. In addition to emotional content, there are also neutral ones. As shown in Figure 8 (a), we combine the common neutral objects/scenes with emotional representations learned by our method. Surprisingly, we find that these representations effectively preserve emotional semantics and seamlessly integrate them with new concepts. Taking amusement as an example, it preserves several semantics including amusement park, picnic, princess, balloon and beautiful lanterns. In Table 3, our method is quantitatively compared with the state-of-the-art methods on emotion transfer task, specializing in room, where our method can well-preserve semantics and effectively elicit emotions. Crucially, these creations can evoke explicit and strong emotions across various neutral semantics, suggesting the potential of our method in image editing, image transfer and emotional art design." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 689, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 288, + 715 + ], + "type": "text", + "content": "Emotion Fusion Additionally, we explore the possibilities of combining different emotion representations to evoke" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 227, + 547, + 347 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 227, + 547, + 347 + ], + "spans": [ + { + "bbox": [ + 304, + 227, + 547, + 347 + ], + "type": "text", + "content": "multiple emotions. In Figure 8 (b), we combine amusement and awe (positive-positive) as well as amusement and fear (positive-negative), bringing some intriguing observations. In the combination of amusement and awe, we observe elements associated with amusement, such as toys, balloons, and ice-creams, alongside awe-inspiring elements like the blue sky, mountains, ocean, and city views. Particularly, one may feel both fear and amusement when viewing the funny and horrible face. When we fuse emotions, we are essentially combining their corresponding visual contents." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 353, + 379, + 365 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 353, + 379, + 365 + ], + "spans": [ + { + "bbox": [ + 306, + 353, + 379, + 365 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 368, + 547, + 488 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 368, + 547, + 488 + ], + "spans": [ + { + "bbox": [ + 304, + 368, + 547, + 488 + ], + "type": "text", + "content": "Discussion In this paper, we introduce a new task named EICG and derive three specially designed metrics. We propose an emotion space and align it with the CLIP space, incorporating attribute loss and emotion confidence to ensure semantic clarity, semantic diversity and emotion fidelity. Experimental results indicate that our method surpasses the state-of-the-art text-to-image diffusion models both qualitatively and quantitatively, where user study confirms its superiority. Additionally, we outline potential applications for EICG and present some initial but promising results." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 498, + 547, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 498, + 547, + 616 + ], + "spans": [ + { + "bbox": [ + 304, + 498, + 547, + 616 + ], + "type": "text", + "content": "Limitations Emotions can be evoked by various visual factors such as color, style and content. In this paper, we focus on investigating the most influential factor, i.e., contents. Moreover, the relationships between emotions and content is not strictly binary. In this paper, we simplify this connection by assuming content to be either emotional or emotion-agnostic. However, in reality, it is hard to assign rose to a single emotion category. White rose may evoke sadness while red rose can elicit amusement, making it hard to decide whether rose is emotional or not." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 618, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 618, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 618, + 547, + 713 + ], + "type": "text", + "content": "Acknowledgments: This work was supported in parts by NSFC (62302312, U21B2023), Guangdong Science and Technology Program (2023A1515011440), DEGP Innovation Team (2022KCXTD025), Shenzhen Science and Technology Program (KQTD20210811090044003, RCJC20200714114435012), Guangdong Laboratory of Artificial Intelligence and Digital Economy (SZ) and Scientific Development Funds of Shenzhen University." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6365" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 91, + 288, + 714 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "type": "text", + "content": "[1] Damian Borth, Tao Chen, Rongrong Ji, and Shih-Fu Chang. Sentibank: large-scale ontology and classifiers for detecting sentiment and emotions in visual content. In Proceedings of the 21st ACM International Conference on Multimedia, pages 459-460, 2013. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 146, + 288, + 200 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 146, + 288, + 200 + ], + "spans": [ + { + "bbox": [ + 53, + 146, + 288, + 200 + ], + "type": "text", + "content": "[2] Damian Borth, Rongrong Ji, Tao Chen, Thomas Breuel, and Shih-Fu Chang. Large-scale visual sentiment ontology and detectors using adjective noun pairs. In Proceedings of the 21st ACM International Conference on Multimedia, pages 223-232, 2013. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 201, + 288, + 235 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 201, + 288, + 235 + ], + "spans": [ + { + "bbox": [ + 53, + 201, + 288, + 235 + ], + "type": "text", + "content": "[3] Tobias Brosch, Gilles Pourtois, and David Sander. The perception and categorisation of emotional stimuli: A review. Cognition and Emotion, 24(3):377-400, 2010. 2, 3, 8" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 236, + 288, + 257 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 236, + 288, + 257 + ], + "spans": [ + { + "bbox": [ + 53, + 236, + 288, + 257 + ], + "type": "text", + "content": "[4] Linda Camras. Emotion: a psychoevolutionary synthesis, 1980. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 258, + 288, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 258, + 288, + 300 + ], + "spans": [ + { + "bbox": [ + 53, + 258, + 288, + 300 + ], + "type": "text", + "content": "[5] Tianlang Chen, Wei Xiong, Haitian Zheng, and Jiebo Luo. Image sentiment transfer. In Proceedings of the 28th ACM International Conference on Multimedia, pages 4407-4415, 2020. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 302, + 288, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 302, + 288, + 335 + ], + "spans": [ + { + "bbox": [ + 53, + 302, + 288, + 335 + ], + "type": "text", + "content": "[6] Domenico Consoli. A new concept of marketing: The emotional marketing. _BRAND_. Broad Research in Accounting, Negotiation, and Distribution, 1(1):52-59, 2010. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 336, + 288, + 368 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 336, + 288, + 368 + ], + "spans": [ + { + "bbox": [ + 53, + 336, + 288, + 368 + ], + "type": "text", + "content": "[7] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in Neural Information Processing Systems, 34:8780-8794, 2021. 2, 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 369, + 288, + 412 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 369, + 288, + 412 + ], + "spans": [ + { + "bbox": [ + 53, + 369, + 288, + 412 + ], + "type": "text", + "content": "[8] Ziyi Dong, Pengxu Wei, and Liang Lin. Dreamartist: Towards controllable one-shot text-to-image generation via contrastive prompt-tuning. arXiv preprint arXiv:2211.11337, 2022. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 414, + 288, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 414, + 288, + 468 + ], + "spans": [ + { + "bbox": [ + 53, + 414, + 288, + 468 + ], + "type": "text", + "content": "[9] Oran Gafni, Adam Polyak, Oron Ashual, Shelly Sheynin, Devi Parikh, and Yaniv Taigman. Make-a-scene: Scene-based text-to-image generation with human priors. In European Conference on Computer Vision, pages 89-106. Springer, 2022. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 469, + 288, + 523 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 469, + 288, + 523 + ], + "spans": [ + { + "bbox": [ + 47, + 469, + 288, + 523 + ], + "type": "text", + "content": "[10] Rinon Gal, Yuval Alaluf, Yuval Atzmon, Or Patashnik, Amit H Bermano, Gal Chechik, and Daniel Cohen-Or. An image is worth one word: Personalizing text-to-image generation using textual inversion. arXiv preprint arXiv:2208.01618, 2022. 2, 3, 4, 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 524, + 288, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 524, + 288, + 568 + ], + "spans": [ + { + "bbox": [ + 47, + 524, + 288, + 568 + ], + "type": "text", + "content": "[11] Leon A Gatys, Alexander S Ecker, and Matthias Bethge. Image style transfer using convolutional neural networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2414-2423, 2016. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 47, + 569, + 288, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 569, + 288, + 612 + ], + "spans": [ + { + "bbox": [ + 47, + 569, + 288, + 612 + ], + "type": "text", + "content": "[12] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial networks. Communications of the ACM, 63(11):139–144, 2020. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 47, + 613, + 288, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 613, + 288, + 646 + ], + "spans": [ + { + "bbox": [ + 47, + 613, + 288, + 646 + ], + "type": "text", + "content": "[13] Alan Hanjalic. Extracting moods from pictures and sounds: Towards truly personalized tv. IEEE Signal Processing Magazine, 23(2):90-100, 2006. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 47, + 647, + 288, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 647, + 288, + 690 + ], + "spans": [ + { + "bbox": [ + 47, + 647, + 288, + 690 + ], + "type": "text", + "content": "[14] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 770-778, 2016. 4" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 47, + 691, + 288, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 691, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 691, + 288, + 714 + ], + "type": "text", + "content": "[15] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 713 + ], + "type": "list", + "angle": 0, + "index": 32, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 106 + ], + "type": "text", + "content": "two time-scale update rule converge to a local nash equilibrium. Advances in Neural Information Processing Systems, 30, 2017. 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 107, + 545, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 107, + 545, + 140 + ], + "spans": [ + { + "bbox": [ + 307, + 107, + 545, + 140 + ], + "type": "text", + "content": "[16] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in Neural Information Processing Systems, 33:6840-6851, 2020. 2, 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 142, + 545, + 186 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 142, + 545, + 186 + ], + "spans": [ + { + "bbox": [ + 307, + 142, + 545, + 186 + ], + "type": "text", + "content": "[17] Elaine Hsieh and Brenda Nicodemus. Conceptualizing emotion in healthcare interpreting: A normative approach to interpreters' emotion work. Patient Education and Counseling, 98(12):1474-1481, 2015. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 188, + 545, + 232 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 188, + 545, + 232 + ], + "spans": [ + { + "bbox": [ + 307, + 188, + 545, + 232 + ], + "type": "text", + "content": "[18] Yifei Huang, Sheng Qiu, Changbo Wang, and Chenhui Li. Learning representations for high-dynamic-range image color transfer in a self-supervised way. IEEE Transactions on Multimedia, 23:176-188, 2020. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 234, + 545, + 278 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 234, + 545, + 278 + ], + "spans": [ + { + "bbox": [ + 307, + 234, + 545, + 278 + ], + "type": "text", + "content": "[19] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4401-4410, 2019. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 280, + 545, + 302 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 280, + 545, + 302 + ], + "spans": [ + { + "bbox": [ + 307, + 280, + 545, + 302 + ], + "type": "text", + "content": "[20] Diederik P Kingma and Max Welling. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114, 2013. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 304, + 545, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 304, + 545, + 358 + ], + "spans": [ + { + "bbox": [ + 307, + 304, + 545, + 358 + ], + "type": "text", + "content": "[21] Nupur Kumari, Bingliang Zhang, Richard Zhang, Eli Shechtman, and Jun-Yan Zhu. Multi-concept customization of text-to-image diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1931-1941, 2023. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 360, + 545, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 360, + 545, + 392 + ], + "spans": [ + { + "bbox": [ + 307, + 360, + 545, + 392 + ], + "type": "text", + "content": "[22] Yann LeCun, Sumit Chopra, Raia Hadsell, M Ranzato, and Fujie Huang. A tutorial on energy-based learning. Predicting Structured Data, 1(0), 2006. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 395, + 545, + 427 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 395, + 545, + 427 + ], + "spans": [ + { + "bbox": [ + 307, + 395, + 545, + 427 + ], + "type": "text", + "content": "[23] Joonwooan Lee and EunJong Park. Fuzzy similarity-based emotional classification of color images. IEEE Transactions on Multimedia, 13(5):1031-1039, 2011. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 429, + 545, + 483 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 429, + 545, + 483 + ], + "spans": [ + { + "bbox": [ + 307, + 429, + 545, + 483 + ], + "type": "text", + "content": "[24] Wentong Liao, Kai Hu, Michael Ying Yang, and Bodo Rosenhahn. Text to image generation with semantic-spatial aware gan. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18187-18196, 2022. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 486, + 545, + 518 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 486, + 545, + 518 + ], + "spans": [ + { + "bbox": [ + 307, + 486, + 545, + 518 + ], + "type": "text", + "content": "[25] Da Liu, Yaxi Jiang, Min Pei, and Shiguang Liu. Emotional image color transfer via deep learning. Pattern Recognition Letters, 110:16-22, 2018. 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 520, + 545, + 565 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 520, + 545, + 565 + ], + "spans": [ + { + "bbox": [ + 307, + 520, + 545, + 565 + ], + "type": "text", + "content": "[26] Jana Machajdik and Allan Hanbury. Affective image classification using features inspired by psychology and art theory. In Proceedings of the 18th ACM International Conference on Multimedia, pages 83-92, 2010. 3" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 567, + 545, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 567, + 545, + 610 + ], + "spans": [ + { + "bbox": [ + 307, + 567, + 545, + 610 + ], + "type": "text", + "content": "[27] Saif Mohammad and Svetlana Kiritchenko. Wikiart emotions: An annotated dataset of emotions evoked by art. In Proceedings of the eleventh International Conference on Language Resources and Evaluation, 2018. 3" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 612, + 545, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 612, + 545, + 667 + ], + "spans": [ + { + "bbox": [ + 307, + 612, + 545, + 667 + ], + "type": "text", + "content": "[28] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741, 2021. 3" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 307, + 669, + 547, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 669, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 669, + 547, + 713 + ], + "type": "text", + "content": "[29] Magnus Oskarsson. Robust image-to-image color transfer using optimal inlier maximization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 786-795, 2021. 3" + } + ] + } + ], + "index": 31 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 316, + 757 + ], + "type": "text", + "content": "6366" + } + ] + } + ], + "index": 33 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "text", + "content": "[30] Kuan-Chuan Peng, Tsuhan Chen, Amir Sadovnik, and Andrew C Gallagher. A mixed bag of emotions: Model, predict, and transfer emotion distributions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 860-868, 2015. 2, 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 128, + 287, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 128, + 287, + 194 + ], + "spans": [ + { + "bbox": [ + 48, + 128, + 287, + 194 + ], + "type": "text", + "content": "[31] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763, 2021. 2, 3, 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 194, + 287, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 194, + 287, + 237 + ], + "spans": [ + { + "bbox": [ + 48, + 194, + 287, + 237 + ], + "type": "text", + "content": "[32] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 1 (2):3, 2022. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 238, + 287, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 238, + 287, + 303 + ], + "spans": [ + { + "bbox": [ + 48, + 238, + 287, + 303 + ], + "type": "text", + "content": "[33] Harsh Rangwani, Lavish Bansal, Kartik Sharma, Tejan Karmali, Varun Jampani, and R Venkatesh Babu. Noisytwins: Class-consistent and diverse image generation through stylegans. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5987-5996, 2023. 3, 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 304, + 287, + 337 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 304, + 287, + 337 + ], + "spans": [ + { + "bbox": [ + 48, + 304, + 287, + 337 + ], + "type": "text", + "content": "[34] Tianrong Rao, Xiaoxu Li, and Min Xu. Learning multi-level deep representations for image emotion classification. Neural Processing Letters, pages 1-19, 2016. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 338, + 287, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 338, + 287, + 370 + ], + "spans": [ + { + "bbox": [ + 48, + 338, + 287, + 370 + ], + "type": "text", + "content": "[35] Tianrong Rao, Xiaoxu Li, and Min Xu. Learning multi-level deep representations for image emotion classification. Neural Processing Letters, 51:2043-2061, 2020. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 371, + 287, + 404 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 371, + 287, + 404 + ], + "spans": [ + { + "bbox": [ + 48, + 371, + 287, + 404 + ], + "type": "text", + "content": "[36] Erik Reinhard, Michael Adhikhmin, Bruce Gooch, and Peter Shirley. Color transfer between images. IEEE Computer Graphics and Applications, 21(5):34-41, 2001. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 404, + 287, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 404, + 287, + 437 + ], + "spans": [ + { + "bbox": [ + 48, + 404, + 287, + 437 + ], + "type": "text", + "content": "[37] Danilo Rezende and Shakir Mohamed. Variational inference with normalizing flows. In International Conference on Machine Learning, pages 1530-1538, 2015. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 437, + 287, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 437, + 287, + 491 + ], + "spans": [ + { + "bbox": [ + 48, + 437, + 287, + 491 + ], + "type": "text", + "content": "[38] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10684-10695, 2022. 2, 3, 4, 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 492, + 287, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 492, + 287, + 557 + ], + "spans": [ + { + "bbox": [ + 48, + 492, + 287, + 557 + ], + "type": "text", + "content": "[39] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22500-22510, 2023. 2, 3, 4, 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 559, + 287, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 559, + 287, + 624 + ], + "spans": [ + { + "bbox": [ + 48, + 559, + 287, + 624 + ], + "type": "text", + "content": "[40] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in Neural Information Processing Systems, 35:36479-36494, 2022. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 624, + 287, + 680 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 624, + 287, + 680 + ], + "spans": [ + { + "bbox": [ + 48, + 624, + 287, + 680 + ], + "type": "text", + "content": "[41] Shikun Sun, Jia Jia, Haozhe Wu, Zijie Ye, and Junliang Xing. Msnet: A deep architecture using multi-sentiment semantics for sentiment-aware image style transfer. In ICASSP 2023-2023 IEEE International Conference on Acoustics, Speech and Signal Processing, pages 1-5. IEEE, 2023. 2, 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 681, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 681, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 681, + 287, + 713 + ], + "type": "text", + "content": "[42] Zhentao Tan, Mengelei Chai, Dongdong Chen, Jing Liao, Qi Chu, Bin Liu, Gang Hua, and Nenghai Yu. Diverse semantic image synthesis via probability distribution modeling. In" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "type": "text", + "content": "Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7962-7971, 2021. 6" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 308, + 96, + 545, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 96, + 545, + 140 + ], + "spans": [ + { + "bbox": [ + 308, + 96, + 545, + 140 + ], + "type": "text", + "content": "[43] Guy Tevet, Brian Gordon, Amir Hertz, Amit H Bermano, and Daniel Cohen-Or. Motionclip: Exposing human motion generation to clip space. In European Conference on Computer Vision, pages 358–374. Springer, 2022. 4" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 141, + 545, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 141, + 545, + 173 + ], + "spans": [ + { + "bbox": [ + 308, + 141, + 545, + 173 + ], + "type": "text", + "content": "[44] Xiaohui Wang, Jia Jia, and Lianhong Cai. Affective image adjustment with a single word. The Visual Computer, 29: 1121-1133, 2013. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 175, + 545, + 219 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 175, + 545, + 219 + ], + "spans": [ + { + "bbox": [ + 308, + 175, + 545, + 219 + ], + "type": "text", + "content": "[45] Zhizhong Wang, Lei Zhao, and Wei Xing. *Styloid diffusion: Controllable disentangled style transfer via diffusion models*. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7677-7689, 2023. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 220, + 545, + 263 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 220, + 545, + 263 + ], + "spans": [ + { + "bbox": [ + 308, + 220, + 545, + 263 + ], + "type": "text", + "content": "[46] Yuxiang Wei, Yabo Zhang, Zhilong Ji, Jinfeng Bai, Lei Zhang, and Wangmeng Zuo. Elite: Encoding visual concepts into textual embeddings for customized text-to-image generation. arXiv preprint arXiv:2302.13848, 2023. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 264, + 545, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 264, + 545, + 319 + ], + "spans": [ + { + "bbox": [ + 308, + 264, + 545, + 319 + ], + "type": "text", + "content": "[47] Shuchen Weng, Peixuan Zhang, Zheng Chang, Xinlong Wang, Si Li, and Boxin Shi. Affective image filter: Reflecting emotions from text to images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10810-10819, 2023. 2, 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 320, + 545, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 320, + 545, + 364 + ], + "spans": [ + { + "bbox": [ + 308, + 320, + 545, + 364 + ], + "type": "text", + "content": "[48] Ali Yadollahi, Ameneh Gholipour Shahraki, and Osmar R Zaiane. Current state of text sentiment analysis from opinion to emotion mining. ACM Computing Surveys (CSUR), 50(2): 1-33, 2017. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 365, + 545, + 399 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 365, + 545, + 399 + ], + "spans": [ + { + "bbox": [ + 308, + 365, + 545, + 399 + ], + "type": "text", + "content": "[49] Chuan-Kai Yang and Li-Kai Peng. Automatic mood-transferring between color images. IEEE Computer Graphics and Applications, 28(2):52-61, 2008. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 399, + 545, + 454 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 399, + 545, + 454 + ], + "spans": [ + { + "bbox": [ + 308, + 399, + 545, + 454 + ], + "type": "text", + "content": "[50] Jufeng Yang, Dongyu She, Yu-Kun Lai, Paul L Rosin, and Ming-Hsuan Yang. Weakly supervised coupled networks for visual sentiment analysis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7584-7592, 2018. 2, 4" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 456, + 545, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 456, + 545, + 498 + ], + "spans": [ + { + "bbox": [ + 308, + 456, + 545, + 498 + ], + "type": "text", + "content": "[51] Jingyuan Yang, Xinbo Gao, Leida Li, Xiumei Wang, and Jinshan Ding. Solver: Scene-object interrelated visual emotion reasoning network. IEEE Transactions on Image Processing, 30:8686-8701, 2021. 2, 3, 4" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 500, + 545, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 500, + 545, + 533 + ], + "spans": [ + { + "bbox": [ + 308, + 500, + 545, + 533 + ], + "type": "text", + "content": "[52] Jingyuan Yang, Jie Li, Xiumei Wang, Yuxuan Ding, and Xinbo Gao. Stimuli-aware visual emotion analysis. IEEE Transactions on Image Processing, 30:7432-7445, 2021. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 534, + 545, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 534, + 545, + 590 + ], + "spans": [ + { + "bbox": [ + 308, + 534, + 545, + 590 + ], + "type": "text", + "content": "[53] Jingyuan Yang, Qirui Huang, Tingting Ding, Dani Lischinski, Danny Cohen-Or, and Hui Huang. Emoset: A large-scale visual emotion dataset with rich attributes. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 20383–20394, 2023. 2, 4, 5" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 590, + 545, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 590, + 545, + 644 + ], + "spans": [ + { + "bbox": [ + 308, + 590, + 545, + 644 + ], + "type": "text", + "content": "[54] Shuai Yang, Liming Jiang, Ziwei Liu, and Chen Change Loy. Pastiche master: Exemplar-based high-resolution portrait style transfer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7693-7702, 2022. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 646, + 545, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 646, + 545, + 690 + ], + "spans": [ + { + "bbox": [ + 308, + 646, + 545, + 690 + ], + "type": "text", + "content": "[55] Chenrui Zhang and Yuxin Peng. Stacking vae and gan for context-aware text-to-image generation. In 2018 IEEE Fourth International Conference on Multimedia Big Data (BigMM), pages 1-5. IEEE, 2018. 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 691, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 691, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 691, + 545, + 713 + ], + "type": "text", + "content": "[56] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6367" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 295 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 67, + 72, + 287, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 72, + 287, + 95 + ], + "spans": [ + { + "bbox": [ + 67, + 72, + 287, + 95 + ], + "type": "text", + "content": "Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3836-3847, 2023. 2, 3, 4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 96, + 288, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 96, + 288, + 150 + ], + "spans": [ + { + "bbox": [ + 48, + 96, + 288, + 150 + ], + "type": "text", + "content": "[57] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 586-595, 2018. 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 152, + 288, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 152, + 288, + 195 + ], + "spans": [ + { + "bbox": [ + 48, + 152, + 288, + 195 + ], + "type": "text", + "content": "[58] Wei Zhang, Xuanyu He, and Weizhi Lu. Exploring discriminative representations for image emotion recognition with cnns. IEEE Transactions on Multimedia, 22(2):515-523, 2019. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 197, + 288, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 197, + 288, + 251 + ], + "spans": [ + { + "bbox": [ + 48, + 197, + 288, + 251 + ], + "type": "text", + "content": "[59] Minfeng Zhu, Pingbo Pan, Wei Chen, and Yi Yang. Dm-gan: Dynamic memory generative adversarial networks for text-to-image synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5802-5810, 2019. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 252, + 288, + 295 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 252, + 288, + 295 + ], + "spans": [ + { + "bbox": [ + 48, + 252, + 288, + 295 + ], + "type": "text", + "content": "[60] Siqi Zhu, Chunmei Qing, Canqiang Chen, and Xiangmin Xu. Emotional generative adversarial network for image emotion transfer. Expert Systems with Applications, 216:119485, 2023. 3" + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "6368" + } + ] + } + ], + "index": 6 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/EmoVIT_ Revolutionizing Emotion Insights with Visual Instruction Tuning/337f29a8-74df-4960-89a4-6eff2797f1cc_content_list.json b/2024/EmoVIT_ Revolutionizing Emotion Insights with Visual Instruction Tuning/337f29a8-74df-4960-89a4-6eff2797f1cc_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..434ec2a8eaba8478054b3554866095416aeb4979 --- /dev/null +++ b/2024/EmoVIT_ Revolutionizing Emotion Insights with Visual Instruction Tuning/337f29a8-74df-4960-89a4-6eff2797f1cc_content_list.json @@ -0,0 +1,1821 @@ +[ + { + "type": "text", + "text": "EmoVIT: Revolutionizing Emotion Insights with Visual Instruction Tuning", + "text_level": 1, + "bbox": [ + 107, + 130, + 864, + 154 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Hongxia Xie1, Chu-Jun Peng2, Yu-Wen Tseng2, Hung-Jen Chen3, Chan-Feng Hsu3, Hong-Han Shuai3, Wen-Huang Cheng2 \n1Jilin University \n2National Taiwan University \n3National Yang Ming Chiao Tung University \nhongxiaie@jlu.edu.cn, {d12922035, ywt, wenhuang}@csie.ntu.edu.tw, {hjc.eed07g, cfhsu311510211.ee11, hhshuai}@nytu.edu.tw", + "bbox": [ + 187, + 179, + 779, + 305 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 339, + 313, + 354 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Visual Instruction Tuning represents a novel learning paradigm involving the fine-tuning of pre-trained language models using task-specific instructions. This paradigm shows promising zero-shot results in various natural language processing tasks but is still unexplored in vision emotion understanding. In this work, we focus on enhancing the model's proficiency in understanding and adhering to instructions related to emotional contexts. Initially, we identify key visual clues critical to visual emotion recognition. Subsequently, we introduce a novel GPT-assisted pipeline for generating emotion visual instruction data, effectively addressing the scarcity of annotated instruction data in this domain. Expanding on the groundwork established by InstructBLIP, our proposed EmoVIT architecture incorporates emotion-specific instruction data, leveraging the powerful capabilities of Large Language Models to enhance performance. Through extensive experiments, our model showcases its proficiency in emotion classification, adeptness in affective reasoning, and competence in comprehending humor. The comparative analysis provides a robust benchmark for Emotion Visual Instruction Tuning in the era of LLMs, providing valuable insights and opening avenues for future exploration in this domain. Our code is available at https://github.com/aimmemotion/EmoVIT.", + "bbox": [ + 75, + 372, + 473, + 734 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 782, + 209, + 797 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Visual emotion recognition, a key area within artificial intelligence and computer vision, aims to predict human emotions based on visual cues such as facial expressions and body language. This technology is essential in bridging the gap between human affective states and machine understanding. Its diverse applications [19, 20, 23, 24], spanning", + "bbox": [ + 75, + 809, + 468, + 902 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/74183f96b9146668203aeb0e0e4c75d5fc3d1e5936bcb0e72888514819342a07.jpg", + "image_caption": [ + "Figure 1. Illustration of the importance of instruction-following ability in visual emotion understanding." + ], + "image_footnote": [], + "bbox": [ + 501, + 338, + 910, + 529 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "from improving human-computer interaction to aiding in mental health assessment, underscore its significance. Accurate emotion recognition is vital for enhancing user experience and ensuring information security, as it helps prevent emotional manipulation and misinformation [13]. Developing robust emotion recognition models is not only a technical challenge but also a step towards more empathetic and intuitive AI systems, paving the way for more efficient and natural human-computer interactions.", + "bbox": [ + 496, + 582, + 893, + 718 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The AI community has recently shown a growing interest in developing foundational vision models, e.g., Flamingo [1], LLaVA [8], BLIP2 [5]. These models excel in open-world visual understanding, tackling several vision tasks such as classification, detection, segmentation, and captioning. In contrast, current large-scale multimodal models are still in its infancy when it comes to emotion perception [7]. As illustrated in Fig. 1, when directly query the GPT-4 [11] about the emotional category of an image, the model tends to provide incorrect responses. However, the model delivers accurate responses when provided with revised instructions. To fully leverage the potential of ex", + "bbox": [ + 496, + 719, + 895, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "26596", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "isting vision-based large models, our approach is based on the concept of Instruction Tuning. This effective strategy is aimed at teaching language models to follow natural language instructions, a technique proven to enhance their generalization performance across unseen tasks [2, 8, 18].", + "bbox": [ + 75, + 90, + 468, + 167 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this work, we focus on developing the model's proficiency in understanding and following instructions related to emotional contexts. This approach highlights the importance of fine-tuning the model's instruction-following capabilities, enabling it to interpret and respond to emotional content effectively. This is achieved by leveraging its pre-existing knowledge base, thereby eliminating the necessity for an emotion-specific architectural framework.", + "bbox": [ + 75, + 169, + 468, + 289 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address the notable challenges encountered in Instruction Tuning for visual emotion recognition, especially the lack of specific instruction data, we introduce a novel self-generation pipeline explicitly crafted for visual emotion recognition by using GPT-4 [11]. This innovative pipeline excels in generating a diverse array of (image, instruction, output) instances, thereby notably enhancing the dataset with a more extensive and task-oriented variety of examples. This approach not only overcomes the challenge of limited data availability but also reduces the dependence on human labor. Therefore, it streamlines the process, enabling more efficient and effective emotion recognition.", + "bbox": [ + 75, + 291, + 470, + 472 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Additionally, Instruction Tuning has been criticized for its emphasis on surface-level features like output patterns and styles, rather than achieving a profound comprehension and assimilation of tasks [26]. To tackle this issue and enhance the diversity and creativity of instruction data, our dataset includes instructions that demand complex reasoning, going beyond basic question-and-answer formats. This is further enriched by incorporating visual cues such as brightness, colorfulness, scene type, object class, facial expressions, and human actions. These aspects are pivotal in fostering a nuanced comprehension of visual emotions, thus allowing the model to generate more precise and contextually appropriate interpretations [24].", + "bbox": [ + 75, + 474, + 470, + 671 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "After generating the emotion visual instruction data, we propose an Emotion Visual Instruction Tuning (EmoVIT) framework, leveraging the foundation of InstructBLIP [2]. This framework incorporates an emotion-centric, instruction-aware module that proficiently guides Large Language Models (LLMs) in assimilating the nuances of emotion instructions. Our work signifies a paradigm shift, presenting a new era of instruction-based learning for visual emotion understanding that relies less on explicit training data. Remarkably, as shown in Fig. 2, our approach requires almost $50\\%$ of the training data typically needed yet exceeds the performance of previous visual emotion recognition methods and popular Visual Instruction Tuning methods.", + "bbox": [ + 75, + 672, + 468, + 883 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our contributions can be summarized as follows:", + "bbox": [ + 96, + 885, + 421, + 898 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/ca388d39393f94f7229363cdad32b2a024c9d1f0313d6164e0808e710feac5e7.jpg", + "image_caption": [ + "Figure 2. Performance comparison on EmoSet test set [24] (Accuracy $\\%$" + ], + "image_footnote": [], + "bbox": [ + 504, + 87, + 915, + 348 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We explore the potential of the Visual Instruction Tuning paradigm for emotion comprehension and introduce the concept of Emotion Visual Instruction Tuning.", + "- After thoroughly considering the unique characteristics of visual emotion recognition, we develop a novel GPT-assisted pipeline for generating emotion visual instruction data. This approach effectively bridges the gap in available annotated instruction data within this specific domain.", + "- Building upon the foundation of InstructBLIP, our EmoVIT architecture integrates emotion domain-specific instruction data, harnessing the robust capabilities of LLMs to boost performance. The extensive experiments demonstrate our model's proficiency in emotion classification, affective reasoning, and comprehension of humour." + ], + "bbox": [ + 500, + 416, + 890, + 656 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 671, + 640, + 686 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Visual Emotion Recognition", + "text_level": 1, + "bbox": [ + 500, + 696, + 751, + 712 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "A key challenge in visual emotion recognition is bridging the gap between an image's visual cues and the emotions it portrays [12, 21, 25]. While traditional efforts, e.g., Xu et al.'s multi-level dependent attention network [21], focus on visual models for emotional feature learning, recent advancements like EmoSet [24] offer rich emotion-laden datasets with 3.3 million images. The rise of multimodal models, such as the GPT series [11], has further propelled Vision-Language Recognition. However, fully leveraging these models in emotion recognition is an area ripe for exploration. Our work leads the way in utilizing large-scale models for Emotion Visual Instruction Tuning.", + "bbox": [ + 496, + 719, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "26597", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/9365186dfb4f66d5b67d49c60155a0e09792131e67ecb988d89722371788fb2f.jpg", + "image_caption": [ + "Figure 3. The comparison of different visual tuning paradigms." + ], + "image_footnote": [], + "bbox": [ + 127, + 85, + 843, + 324 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. Visual Instruction Tuning", + "text_level": 1, + "bbox": [ + 76, + 358, + 312, + 375 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Current Large Language Models (LLMs) have extensive knowledge bases, but their effectiveness depends on accurately interpreting human instructions due to a mismatch between training goals and user expectations. LLMs are trained to minimize prediction errors, whereas users expect helpful and safe instruction-following. Instruction Tuning addresses this by teaching models to follow natural language instructions, enhancing generalization to new tasks. FLAN [18] demonstrated that training a large model on instruction-based datasets improves zero-shot performance. This approach has extended to vision-language tasks, with BLIP2 [5] and LLaVA [8] adapting instruction-tuned LLMs for visual inputs. InstructBLIP [2] introduces instruction-aware visual feature extraction and the Q-Former, enabling more flexible, instruction-driven feature extraction.", + "bbox": [ + 75, + 382, + 470, + 608 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "As a novel area, visual emotion instruction tuning lacks benchmarks or guidelines for creating emotion instruction data. Our work pioneers the use of large-scale models to develop an emotion instruction data pipeline, overcoming the limitations of manual annotation.", + "bbox": [ + 75, + 609, + 468, + 684 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 76, + 699, + 168, + 714 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Preliminary of Visual Instruction Tuning", + "text_level": 1, + "bbox": [ + 76, + 724, + 428, + 742 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In the deep learning era, visual tuning has experienced significant paradigm shifts, as depicted in Fig. 3.", + "bbox": [ + 75, + 750, + 468, + 779 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In Fig. 3(a), conventional tuning methodologies encompass Full fine-tuning, Head-oriented, and Backbone-oriented techniques, capitalizing on large-scale pre-trained models. Predominantly, thoroughly fine-tuning these models for specific tasks, conducted end-to-end, is recognized as a highly effective strategy. However, this method requires maintaining separate copies of the backbone parameters for each distinct task, posing challenges in storage and deploy-", + "bbox": [ + 75, + 780, + 470, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "ment.", + "bbox": [ + 500, + 361, + 540, + 372 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Alternatively, Visual Prompt Tuning (VPT) [3], presents an efficient substitute for full fine-tuning within large-scale vision Transformer models. It achieves this by employing a minimal fraction of trainable parameters in the input space while maintaining a frozen backbone model. The objective function for Visual Prompt Tuning is given by:", + "bbox": [ + 496, + 375, + 890, + 465 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {\\theta_ {\\mathrm {P}}} \\mathcal {L} (f (X, P; \\theta_ {\\mathrm {P}}), Y) \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 616, + 479, + 890, + 503 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\min_{\\theta_{\\mathrm{P}}}$ is the minimization over the prompt parameters $P$ , $\\mathcal{L}$ is the loss function, $f$ represents the model function with input image $X$ , prompt parameters $P$ , and learnable model parameters $\\theta_{\\mathrm{P}}$ as input, and $Y$ is the target output.", + "bbox": [ + 496, + 516, + 890, + 590 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Visual Prompt Tuning focuses on optimizing LLMs using a small set of parameters, whereas Visual Instruction Tuning (VIT) aims to improve the model's comprehension of instructions, thereby addressing the model's shortcomings in specific domains. This type of method aims to enhance the model's proficiency in following instructions, leveraging the capabilities of the latest foundation models, e.g., Llama [17], and BLIP2 [5]. Instructions serve as guiding constraints, shaping the model's outputs to conform to specific response characteristics and domain-relevant knowledge. This approach enables human monitoring of the model's behavior, thereby assuring alignment with the desired outcomes. Moreover, Instruction Tuning is computationally efficient, allowing LLMs to swiftly adapt to particular domains without extensive retraining or architectural alterations.", + "bbox": [ + 496, + 593, + 892, + 833 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The objective function for Visual Instruction Tuning is given by:", + "bbox": [ + 496, + 835, + 890, + 866 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {\\theta_ {\\text {t u n a b l e}}} \\mathcal {L} (g (X, I, C; \\theta_ {\\text {t u n a b l e}}), Y) \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 593, + 880, + 890, + 904 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "26598", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/9342f2ab39a50f8859b88420ccd7949deef2992314bd34406459d0f268e2bd0e.jpg", + "image_caption": [ + "Emotion class", + "Contentment", + "Brightness", + "0.8", + "Colorfulness", + "0.8", + "Scene type" + ], + "image_footnote": [], + "bbox": [ + 83, + 123, + 223, + 200 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/10f7bfb5bdea883c73faa2db931822742351b94eebfa64006c7229a06b6ca39d.jpg", + "image_caption": [ + "Beach", + "Object class", + "Sky", + "Facial", + "Contentment", + "Human", + "Laughing.", + "Ludwig", + "(a) Emotion Visual Instruction Data Generation", + "Figure 4. The overall architecture of our proposed method. The Emotion Instruction data generated by (a) will be used for Emotion Visual Instruction Tuning in (b). During Emotion Visual Instruction Tuning, given an input image, the frozen Image Encoder initiates the process by extracting visual features. Emotion Instruction generated by (a) are subsequently interacting with Queries embedding through the learnable Q-Former. This interaction is key to drawing out image features that are relevant to the task at hand. As a result, the frozen LLM receives visual information conducive to instruction following." + ], + "image_footnote": [], + "bbox": [ + 230, + 104, + 390, + 330 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/831fecbf7d7448ec656e1d747dd2c4ffab7b3c61068499a8d6d2a04fdcc61957.jpg", + "image_caption": [ + "(b) Emotion Visual Instruction Tuning Architecture" + ], + "image_footnote": [], + "bbox": [ + 395, + 97, + 666, + 329 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/ea875063c75ecf06e8fc35841680bf9a653c54e0f47e94ec3f979194722849c3.jpg", + "image_caption": [ + "(c) The Details of Q-Former Module" + ], + "image_footnote": [], + "bbox": [ + 694, + 125, + 893, + 315 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\min_{\\theta_{\\mathrm{tunable}}}$ denotes the minimization over the tunable parameters $\\theta_{\\mathrm{tunable}}$ in the Instruction Tuning Module, $\\mathcal{L}$ is the loss function, $g$ is the model function with instruction $I$ , image $X$ , other contexts $C$ , and tunable parameters $\\theta_{\\mathrm{tunable}}$ and $Y$ denotes the target output. The optional context $C$ is not just raw data; it encompasses descriptive or directive information guiding the model on how to process input or which task to execute, e.g., image caption. It's integral to the model's understanding and execution of tasks based on specific instructions or guidelines.", + "bbox": [ + 75, + 446, + 470, + 598 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. GPT-assisted Emotion Visual Instruction Data Generation", + "text_level": 1, + "bbox": [ + 76, + 604, + 468, + 635 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Previous methodologies commonly employed a consistent template-based set of instructions for every image within a dataset across various specific tasks [2]. For instance, a standard instruction such as \"Briefly describe the content of the image\" was employed uniformly across all images for Image Captioning. In this way, the model may not be able to adequately capture the unique characteristics of each image. Moreover, this one-size-fits-all approach often leads to suboptimal performance in emotion recognition tasks that require nuanced perception and differentiation of ambiguous emotion classes.", + "bbox": [ + 75, + 643, + 468, + 808 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Since the topic of Emotion Visual Instruction Tuning is still in its infancy, no benchmarks or guidelines have been proposed so far for constructing emotion instruction data. Based on the recent successes of machine-generated instructions demonstrated in LLaVA [8], our work pioneers the use of existing LLMs to create a pipeline for", + "bbox": [ + 75, + 810, + 468, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "self-generating emotion instructions. Different from previous template-based and one-size-fits-all instruction data, we propose an instance-wise and LLM-assisted visual emotion instruction data pipeline. This methodology transcends the constraints of manual annotation by employing GPT-4 [11] to generate instance-wise, tailored instruction data that dynamically corresponds to visual content.", + "bbox": [ + 496, + 446, + 890, + 551 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Prior to the development of instructional data for the visual emotion recognition task, it is imperative to confront a fundamental academic problem: What types of visual clues are pivotal in identifying emotions? This necessitates a careful consideration of the unique characteristics inherent to the task, along with a comprehensive understanding of the potential visual cues associated with human emotions. In this work, we propose a novel visual instruction data mechanism to remove the inherent subjectivity and ambiguity in emotional interpretation. Specifically, we integrate a broad spectrum of emotion attributes across multiple levels: low-level attributes (e.g., brightness, colorfulness), mid-level attributes (e.g., scene type and object class), and high-level attributes (e.g., facial expressions and human actions), building upon insights from previous work [24]. This comprehensive strategy not only aligns with the intricate nature of emotions but also significantly enhances the model's capability to interpret and understand visual emotional cues more accurately and holistically.", + "bbox": [ + 496, + 553, + 890, + 839 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The overall pipeline of our proposed emotion visual instruction data is shown in Fig. 4 (a). For an image $X_{\\mathrm{img}}$ , three types of image-related contexts are essential for GPT-4 to generate emotion instruction data: (i) a caption $X_{c}$ , (ii)", + "bbox": [ + 496, + 839, + 890, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "26599", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "an emotion attribute list $X_{\\mathrm{attr}}$ , which includes emotion class, brightness, colorfulness, scene type, object class, facial expression, and human action, and (iii) the system prompt, designed to enable GPT-4 to comprehend the specific task requirement1.", + "bbox": [ + 75, + 90, + 468, + 165 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We first manually design a few examples which are used as seed examples for in-context learning to query GPT-4. This operation leverages the model's ability to extrapolate from given examples, enhancing its understanding and response accuracy based on the principles of few-shot learning [8]. Our generated emotion instruction data includes three types: Categorical, Conversation, and Reasoning. Building upon previous research [8], our generated instruction data adheres to the dialogue format, exemplified in Fig. 5.", + "bbox": [ + 75, + 167, + 470, + 316 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our strategy for generating emotion instruction data adopts a progressive approach from simple to complex. Initially, for the Categorical data, we transform the associated emotion class of the image into a structured format. This process serves as the foundational component of our emotion instruction data.", + "bbox": [ + 75, + 318, + 468, + 407 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For the Conversation data, our framework is designed to create dialogues in which the GPT assistant interacts with an inquirer, focusing on the emotion attributes of the image. In this setup, the assistant's responses are tailored to interpret and describe the image as though it were within its own visual field, thereby providing insights from an observational viewpoint. The scope of questions posed is comprehensive, encompassing the types of objects depicted, their actions, and the dynamics of their interrelationships. The dialogues we generate fall into two categories: (i) Basic Interaction, focusing on the provided emotion attribute list with simple, direct characteristics, and (ii) Advanced Interaction, which builds on the first type to reach greater conversational complexity and sophistication.", + "bbox": [ + 75, + 409, + 468, + 619 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For the Reasoning data, our approach extends beyond mere visual content, prompting the model to generate in-depth reasoning questions. To enhance the dialogue's credibility and structure, detailed examples are incorporated alongside logical reasoning steps, ensuring that the discourse convincingly captures the intricacies of the visual content.", + "bbox": [ + 75, + 619, + 468, + 724 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Emotion Visual Instruction Tuning", + "text_level": 1, + "bbox": [ + 76, + 734, + 382, + 752 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "After acquiring the emotion visual instruction data as detailed in Sec. 3.2, our goal is to employ this data in enhancing the existing Visual Instruction Tuning model. This enhancement aims to align the LLMs' existing knowledge with the emotion understanding domain.", + "bbox": [ + 75, + 758, + 468, + 834 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As shown in Fig. 4 b, we have developed an Emotion Visual Instruction Tuning (EmoVIT) architecture based on", + "bbox": [ + 76, + 835, + 468, + 864 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our Emotion Instruction Data Sample", + "text_level": 1, + "bbox": [ + 609, + 93, + 803, + 104 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/2f4119e5872a968c563b86bb5c30d6c9e26c9413d6245d39614774b22b57e4bd.jpg", + "image_caption": [ + "Figure 5. The sample of our generated visual emotion instruction data." + ], + "image_footnote": [], + "bbox": [ + 522, + 114, + 635, + 178 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Context type 1: Caption:", + "bbox": [ + 640, + 118, + 746, + 127 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Two young women dancing in a kitchen.", + "bbox": [ + 640, + 128, + 803, + 136 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Context type 2: Attribute:", + "bbox": [ + 640, + 141, + 753, + 148 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Emotion: contentment Human action: stretching arm", + "bbox": [ + 640, + 150, + 856, + 157 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Brightness: 0.7 Object: ['Cupboard', 'Food', 'Countertop']", + "bbox": [ + 640, + 159, + 890, + 167 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Colorfulness: 0.4 Facial expression: happy", + "bbox": [ + 640, + 169, + 834, + 176 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Categorical", + "text_level": 1, + "bbox": [ + 545, + 198, + 604, + 208 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Question: Please select the emotion closest to the image from the following options: amusement, anger, awe, contentment, disgust, excitement, fear and sadness.", + "bbox": [ + 524, + 210, + 877, + 242 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Answer: Predict emotion: contentment.", + "bbox": [ + 524, + 243, + 712, + 253 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Conversation", + "text_level": 1, + "bbox": [ + 542, + 261, + 607, + 270 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Basic Interaction", + "text_level": 1, + "bbox": [ + 637, + 261, + 717, + 270 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Question: Can you describe the activity happening in the image?", + "bbox": [ + 524, + 273, + 826, + 284 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Answer: The image depicts two young women enjoying themselves by dancing in a kitchen. One of the women is seen stretching her arm, most likely as part of their dance routine.", + "bbox": [ + 524, + 284, + 859, + 316 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Conversation", + "text_level": 1, + "bbox": [ + 542, + 321, + 607, + 330 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Advanced Interaction", + "text_level": 1, + "bbox": [ + 625, + 320, + 728, + 330 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Question: Based on the given emotion of contentment and the activities in the image, can you guess about the possible context or story of this scenario?", + "bbox": [ + 524, + 334, + 887, + 354 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Answer: These women are good friends or roommates, spending quality time together. The act of dancing, especially in an unconventional place like a kitchen, suggests they are comfortable with each other and are enjoying a break from routine.", + "bbox": [ + 524, + 354, + 885, + 396 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Reasoning", + "text_level": 1, + "bbox": [ + 547, + 404, + 601, + 415 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Question: Based on the setting and activities of the two young women, what might have prompted this impromptu dance session in the middle of a kitchen? Answer: The impromptu dance could have been prompted by a favorite song playing, or simple spontaneous fun amidst preparing a meal or organizing the kitchen.", + "bbox": [ + 522, + 417, + 890, + 470 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "InstructBLIP [2]. This architecture specifically leverages its Instruction-aware Q-Former Module, as depicted in Fig. 4 c, for emotion-centric instructional tasks.", + "bbox": [ + 498, + 535, + 890, + 580 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Specifically, the Instruction-aware Q-Former Module takes in the emotion instruction tokens, queries, and image embeddings as input. The image embeddings are extracted by a frozen image encoder. The learnable queries are initially produced by the pre-trained Q-Former of InstructBLIP. During training, the Instruction-aware module enhances task-specific feature extraction. It does this by integrating emotion instruction and query embeddings within self-attention layers, aligning visual information with the LLM's instruction-following requirements. Our approach adopts cross-entropy loss, tailoring it to the intricacies of visual emotion recognition tasks, thus ensuring precise and contextually relevant model training outcomes.", + "bbox": [ + 496, + 582, + 892, + 777 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We note that the data generated by our approach is not confined to a single model but can also be applied to other Visual Instruction Tuning models, such as LLaVA [17]. Notably, when LLaVA is fine-tuned with our data, it exhibits a significant enhancement in emotion recognition capabilities, as detailed in Sec. 4.2. In this way, we demonstrate not only the effectiveness but also the transferability of our generated data, showing its broad applicability and impact.", + "bbox": [ + 496, + 779, + 893, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_footnote", + "text": "1A detailed description of the system prompt is provided in the supplementary materials.", + "bbox": [ + 76, + 875, + 468, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "26600", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experimental Results", + "text_level": 1, + "bbox": [ + 76, + 89, + 282, + 107 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Implemental Details", + "text_level": 1, + "bbox": [ + 76, + 114, + 269, + 131 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Our implementation is based on the LAVIS library [4]. Our EmoVIT starts with a pre-trained InstructBLIP baseline and proceeds to fine-tune exclusively the Q-Former module, whilst keeping both the image encoder and the language model frozen. The parameters for our training adhere to the default settings established by InstructBLIP.", + "bbox": [ + 75, + 137, + 467, + 227 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Datasets. We evaluate our framework on ten benchmark datasets annotated under different scenarios and class number, namely EmoSet [24], WEBEmo [12], Emotion6 [14], the Flickr and Instagram (FI) [25], Artphoto [9], IAPS [10], Abstract [9], EmotionROI [15], UnbiasedEmo [12], and OxfordTVG-HIC [6].", + "bbox": [ + 75, + 228, + 467, + 318 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Held-in Pretraining. Following previous work [2], we divide our dataset into two categories: held-in for pretraining and held-out for evaluation2. Considering the EmoSSet dataset's comprehensive inclusion of emotion attributes for each image, it has been chosen as the primary resource for our held-in pretraining phase. Simultaneously, for a broader assessment, we perform held-out evaluations using the test sets from various other datasets.", + "bbox": [ + 75, + 319, + 467, + 439 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For the generation of emotion visual instruction data, we initially employ the BLIP2 model for image captioning, followed by leveraging the GPT-4 API to generate emotion instruction data. In total, our collection comprises Categorical, Conversation, and Reasoning instruction data, derived from 51,200 unique images. This represents less than $50\\%$ of the entire EmoSet.", + "bbox": [ + 75, + 441, + 467, + 546 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Held-out Evaluation", + "text_level": 1, + "bbox": [ + 76, + 556, + 269, + 571 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "As shown in Tab. 1, our proposed methodology exhibits a marked superiority in performance relative to the burgeoning Visual Instruction Tuning Methods. Since they have been pre-trained on dozens of large-scale datasets, it is evident that our generated emotion visual instruction data is particularly effective for emotional understanding. Our results signify a paradigm shift, heralding a new era of model training that relies less on explicit supervision and more on the robustness of emotion instruction-driven learning.", + "bbox": [ + 75, + 580, + 467, + 715 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The Effectiveness of Our Proposed Emotion Visual Instruction Data. As the first to introduce the concept of emotion visual instruction data, our study seeks to evaluate the generalizability of this newly generated instruction data. Our goal is to test its efficacy not only with InstructBLIP but also across other Visual Instruction Tuning model, to understand its broader applicability. As depicted in Fig. 6, we employ two Visual Instruction Tuning models, LLaVA and InstructBLIP, which were fine-tuned on our specially gen", + "bbox": [ + 75, + 715, + 467, + 852 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/ef9f5da6de3e6f86a45bf5ddf42b29074a440d3f8afefe5d124bd6be0d190963.jpg", + "image_caption": [ + "Accuracy Improvements of FT Models Over Base Models", + "Figure 6. The improvement of our proposed emotion visual instruction tuning data tuning on LLaVA [8] and InstructBLIP [2]." + ], + "image_footnote": [], + "bbox": [ + 506, + 104, + 908, + 321 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "erated emotion visual instruction data. Subsequent testing across five distinct datasets reveals notable improvements in both models, substantiating the efficacy of our generated data. Notably, InstructBLIP demonstrated a more substantial overall enhancement compared to LLaVA. This can be attributed to InstructBLIP's specialized Instruction-aware Q-Former Module, which adeptly extracts the salient features of our emotion instructions and synergizes them effectively with the corresponding images, thereby yielding improved performance.", + "bbox": [ + 496, + 382, + 890, + 532 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Effectiveness of Different Instruction Data", + "text_level": 1, + "bbox": [ + 500, + 547, + 859, + 561 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3.1 Ablation Study of Different Instruction Data", + "text_level": 1, + "bbox": [ + 500, + 571, + 862, + 585 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The ablation study outlined in Tab. 2 provides a comprehensive analysis of the impact that different instructional data types have on model performance, specifically concerning accuracy metrics on the EmoSet test set. Initially, the model, referred to as InstructBLIP [2], operates without the integration of the three types of instructional data and attains a baseline accuracy of $42.20\\%$ . This foundational performance is significantly enhanced with the inclusion of Categorical data, which alone contributes to a substantial increase in accuracy. The introduction of Conversation data further amplifies this effect, underscoring the value of conversational context in improving the model's predictive capabilities. The addition of Reasoning data notably boosts performance, achieving a peak accuracy of $83.36\\%$ . This indicates that the model significantly benefits from the nuanced cues in reasoning, aiding in understanding complex emotional instructions. The gradual improvements with each data type support the idea that a diverse approach to instructional data markedly enhances model comprehension and performance.", + "bbox": [ + 496, + 598, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_footnote", + "text": "2Unlike the setup in InstructBLIP, our dataset exclusively comprises emotion-related content. Consequently, our held-out evaluation does not constitute a strict zero-shot evaluation in the conventional sense.", + "bbox": [ + 75, + 862, + 467, + 898 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "26601", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/e2cb0d9137810c97b65195d666a489ec30f622606c4ebfa7ee8bc62e12385832.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodWebEmoFIEmotion6AbstractArtPhotoIAPSaEmotionROIEmoSet
Number of Classes258688868
Flanmingo [1]9.3614.9121.673.5717.510.1321.7229.59
LLaVA [8]12.5556.0449.4419.5436.2542.4346.4644.03
BLIP2 [5]20.1057.7250.0028.5736.2539.2450.5146.79
InstructBLIP [2]12.8037.9746.1121.4226.2534.1846.1342.20
Ours*21.1268.0957.8132.3444.9044.1353.8783.36
", + "bbox": [ + 78, + 88, + 893, + 218 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/4da102b85b62cd03c01a5fe65c5818e416fc9c3626968cbd5bf01f6aa70a21e3.jpg", + "table_caption": [ + "Table 1. Held-out performance comparison on visual emotion datasets $(\\%)$" + ], + "table_footnote": [ + "Table 2. Ablation study of three types of instruction data. Accuracy (%) on EmoSet test set." + ], + "table_body": "
CategoricalConversationReasoningAccuracy (%)
---42.20
✓--80.90 (+38.70)
✓✓-81.95 (+39.75)
✓✓✓83.36 (+41.16)
", + "bbox": [ + 78, + 265, + 478, + 339 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3.2 Instruction Sensitivity", + "text_level": 1, + "bbox": [ + 76, + 388, + 284, + 402 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "This work is dedicated to the creation of a varied corpus of visual emotion instruction data, alongside the development of a robust instruction-based model. Our objective is for the model to demonstrate stability, producing consistent results in the face of minor variations in instruction phrasing, provided the core objective of the task persists unchanged. To this end, we employ the Sensitivity evaluation metric, as introduced by [22], to assess the model's fidelity in generating uniform outcomes irrespective of instructional nuances.", + "bbox": [ + 75, + 412, + 468, + 547 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We employ two semantically similar instructions as input prompts for the model, testing their impact on the Sensitivity score across three visual emotion datasets for different Visual Instruction Tuning models. The first instruction is: \"From the given options: cls_1, cls_2, cls_3, etc., identify the emotion that most accurately reflects the image. Ensure your selection is confined to the listed options. Respond in the format: Predicted emotion.\" The second one states: \"Please choose the emotion that best corresponds to the image from the following options: cls_1, cls_2, cls_3, etc. (Do not provide answers beyond the provided candidates.) Please reply in the following format: Predict emotion.\"", + "bbox": [ + 75, + 549, + 468, + 743 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "As illustrated in Fig. 7, our approach, along with BLIP2, exhibited exceptionally low Sensitivity values, demonstrating robustness in understanding the instructions. Conversely, Flamingo and InstructBLIP displayed a higher degree of sensitivity, indicating a relative susceptibility to variations in instruction wording.", + "bbox": [ + 75, + 744, + 468, + 835 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Robustness", + "text_level": 1, + "bbox": [ + 76, + 845, + 200, + 859 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Given that current emotion recognition datasets often exhibit category imbalances and labeling biases, our aim is", + "bbox": [ + 75, + 869, + 468, + 900 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/9310629969a8e41f19348f83511eba7915ed94c2d9fea133652a4458ae950dc1.jpg", + "image_caption": [ + "Figure 7. The sensitivity score comparison (the lower the better)." + ], + "image_footnote": [], + "bbox": [ + 506, + 270, + 915, + 474 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "to evaluate the generalization ability of various learning strategies more impartially. Hence, we selected the Un-BiasedEmo test set [12], which is uniquely suited for recognizing intricate emotions, such as those associated with identical objects or scenes, e.g., landscapes, crowds, families, babies, and animals, where the emotional undertones can be particularly subtle and complex.", + "bbox": [ + 496, + 537, + 890, + 643 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "As depicted in Tab. 3, our proposed methodology demonstrates superior performance when benchmarked against conventional supervised emotion recognition techniques, thereby underscoring the efficacy of our approach in more accurately discerning complex emotional contexts.", + "bbox": [ + 496, + 647, + 890, + 723 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/2ee8702a65ad3d1a4fb54341b8105000760c282855125751b118dec9f4c63f07.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodAccuracy (%)
Direct Learning [12]71.64
Self-Directed Learning [12]72.45
Joint Learning [12]71.64
Curriculum Learning [12]74.27
Ours*74.72
", + "bbox": [ + 535, + 744, + 857, + 851 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 3. Performance comparison on UnbiasedEmo dataset.", + "bbox": [ + 517, + 862, + 872, + 876 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "26602", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/109bb11314538a214cd82aed3b6d0b5421a081e885cde55b680079b58077f669.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 120, + 104, + 243, + 159 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/053e3769043cbb64654c8c3e03aab6d5a184737816f4645df6871152ef833fd0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 321, + 104, + 421, + 162 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "BECAUSE", + "bbox": [ + 161, + 193, + 215, + 203 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/b0abd87cf84e26d5614dc0f16917a8986399d1404f505f75f291a50d6a3b49a2.jpg", + "image_caption": [ + "Predicted emotion: anger", + "Predicted emotion: sadness" + ], + "image_footnote": [], + "bbox": [ + 246, + 170, + 315, + 222 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "BECAUSE", + "bbox": [ + 344, + 195, + 398, + 204 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Reason: The image depicts a group of people with their arms up in the air, which is a common gesture associated with anger.", + "bbox": [ + 91, + 224, + 272, + 268 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Additionally, the smoke billowing from the building in the background adds to the sense of chaos and unrest.", + "bbox": [ + 91, + 268, + 277, + 301 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Reason: The image is a black and white photograph of an old, abandoned house in the middle of a field.", + "bbox": [ + 289, + 224, + 475, + 257 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The scene is desolate and eerie, which suggests a sense of sadness or melancholy.", + "bbox": [ + 290, + 268, + 475, + 301 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4.1 Affective Reasoning", + "text_level": 1, + "bbox": [ + 76, + 349, + 269, + 366 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In the domain of visual emotion recognition, where ambiguity and subjectivity are pervasive, the advent of an interpretable model is of considerable value. Such a model elucidates its cognitive processes, enhancing its trustworthiness and practicality in scenarios requiring a delicate grasp of emotional subtleties.", + "bbox": [ + 75, + 375, + 468, + 463 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Leveraging Visual Instruction Tuning, our model transcends mere categorization of emotions; it articulates the underlying rationale for its classifications. The executing commands for identifying emotions and elucidating the decision basis is illustrated below:", + "bbox": [ + 75, + 465, + 468, + 539 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Predicted emotion: [emotion].", + "bbox": [ + 135, + 553, + 419, + 566 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Reason: [explanation].", + "bbox": [ + 135, + 568, + 352, + 580 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Our model delineates the visual features influencing its determinations, thereby addressing the complexities inherent in discerning and explaining emotion-related nuances.", + "bbox": [ + 75, + 594, + 468, + 638 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The explanations provide us with visual clues contained within the images, as exemplified in Fig. 8. It provides interpretable visual indicators that inform the model's outputs, as demonstrated in our example, by disambiguating the often abstract emotional categories.", + "bbox": [ + 75, + 640, + 468, + 715 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.5. Scaling Law", + "text_level": 1, + "bbox": [ + 76, + 723, + 207, + 739 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Pretraining data. As demonstrated in Tab. 4, there is a clear correlation between the size of the pre-training dataset and improved performance. Consequently, we anticipate that an increase in training data in the future could enhance the effectiveness of Emotion Visual Instruction Tuning.", + "bbox": [ + 75, + 747, + 468, + 823 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.6. Humour Caption Generation", + "text_level": 1, + "bbox": [ + 76, + 832, + 336, + 848 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The comprehension of humor is intricately linked to the understanding of emotions. Leveraging our generative language model, we conduct a caption generation task without", + "bbox": [ + 75, + 854, + 468, + 900 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/298fb83e44b71d807ad55c49350025bb3a95c37916be87523ba01a98e6609571.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
5%10%30%50%
79.0081.0079.3483.36
", + "bbox": [ + 588, + 88, + 803, + 119 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 4. Ablation study of different portion of pre-training data. Accuracy $(\\%)$ on EmoSet test set.", + "bbox": [ + 500, + 130, + 888, + 159 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/eb9cb349f1ef4b50b1a208cf1b5bd1b0aa7d233078c483d280120546ab7f8105.jpg", + "image_caption": [ + "Figure 8. The sample of our generated explanation.", + "Figure 9. The sample of our generated humour caption vs human writing humour caption from OxfordTVG-HIC." + ], + "image_footnote": [], + "bbox": [ + 524, + 174, + 885, + 378 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "modifying the model's architecture, specifically testing the model's proficiency in generating humorous captions. For this purpose, we select 50 images from the OxfordTVGHIC dataset [6] and generate corresponding captions using our model. Subsequently, the captions produced by our model are compared with manually annotated captions from the dataset in a user study. Thirty participants were asked to vote on which captions were more humorous. Our model-generated captions receive $60\\%$ of the votes, demonstrating its effective humor generation capabilities. One sample is visualized in Fig. 9.", + "bbox": [ + 496, + 420, + 890, + 587 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 601, + 617, + 616 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In our study, drawing upon the distinctive visual cues key to visual emotion recognition, we present a GPT-assisted pipeline specifically designed for generating emotion visual instruction data. The developed EmoVIT model incorporates emotion-specific instructions, leveraging LLMs for enhanced performance. Our comprehensive experiments validate its effectiveness in emotion classification, affective reasoning, and humor understanding. This comparative analysis sets a benchmark for Emotion Visual Instruction Tuning with LLMs, providing valuable insights and directions for future research in this field.", + "bbox": [ + 496, + 626, + 890, + 791 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgement", + "text_level": 1, + "bbox": [ + 500, + 806, + 658, + 821 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This work is partially supported by the National Science and Technology Council, Taiwan under Grant, NSTC-112-2628-E-002-033-MY4, NSTC-112-2634-F-002-002-MBK, NSTC-112-26234-F-002-005, NSTC-112-2221-E-A49-059-MY3, and NSTC-112-2221-E-A49-094-MY3.", + "bbox": [ + 498, + 829, + 890, + 898 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "26603", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. In Advances in Neural Information Processing Systems, pages 23716-23736, 2022.", + "[2] Wenliang Dai, Junnan Li, Dongxu Li, Anthony Meng Huat Tiong, Junqi Zhao, Weisheng Wang, Boyang Li, Pascale Fung, and Steven Hoi. Instructclip: Towards general-purpose vision-language models with instruction tuning. In Advances in Neural Information Processing Systems, 2023.", + "[3] Menglin Jia, Luming Tang, Bor-Chun Chen, Claire Cardie, Serge Belongie, Bharath Hariharan, and Ser-Nam Lim. Visual prompt tuning. In Proceedings of the European Conference on Computer Vision, pages 709-727. Springer, 2022.", + "[4] Dongxu Li, Junnan Li, Hung Le, Guangsen Wang, Silvio Savarese, and Steven CH Hoi. Lavis: A library for language-vision intelligence. arXiv preprint arXiv:2209.09019, 2022.", + "[5] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In Proceedings of the International Conference on Machine Learning, 2023.", + "[6] Runjia Li, Shuyang Sun, Mohamed Elhoseiny, and Philip Torr. Oxfordtvg-hic: Can machine make humorous captions from images? In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 20293-20303, 2023.", + "[7] Zheng Lian, Licai Sun, Mingyu Xu, Haiyang Sun, Ke Xu, Zhuofan Wen, Shun Chen, Bin Liu, and Jianhua Tao. Explainable multimodal emotion reasoning. arXiv preprint arXiv:2306.15401, 2023.", + "[8] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. In Advances in Neural Information Processing Systems, 2023.", + "[9] Jana Machajdik and Allan Hanbury. Affective image classification using features inspired by psychology and art theory. In Proceedings of the 18th ACM international conference on Multimedia, pages 83-92, 2010.", + "[10] Joseph A Mikels, Barbara L Fredrickson, Gregory R Larkin, Casey M Lindberg, Sam J Maglio, and Patricia A Reuter-Lorenz. Emotional category data on images from the international affective picture system. Behavior research methods, 37:626-630, 2005.", + "[11] OpenAI. Gpt-4 technical report. Technical report, 2023.", + "[12] Rameswar Panda, Jianming Zhang, Haoxiang Li, Joon-Young Lee, Xin Lu, and Amit K Roy-Chowdhury. Contemplating visual emotions: Understanding and overcoming dataset bias. In Proceedings of the European Conference on Computer Vision, pages 579–595, 2018.", + "[13] EU Parliament. Eu ai act: first regulation on artificial intelligence. Accessed June, 25:2023, 2023.", + "[14] Kuan-Chuan Peng, Tsuhan Chen, Amir Sadovnik, and Andrew C Gallagher. A mixed bag of emotions: Model, predict, and transfer emotion distributions. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 860-868, 2015." + ], + "bbox": [ + 78, + 114, + 470, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[15] Kuan-Chuan Peng, Amir Sadovnik, Andrew Gallagher, and Tsuhan Chen. Where do emotions come from? predicting the emotion stimuli map. In 2016 IEEE international conference on image processing (ICIP), pages 614-618. IEEE, 2016.", + "[16] Dongyu She, Jufeng Yang, Ming-Ming Cheng, Yu-Kun Lai, Paul L Rosin, and Liang Wang. Wscnet: Weakly supervised coupled networks for visual sentiment classification and detection. IEEE Transactions on Multimedia, 22(5): 1358-1371, 2019.", + "[17] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023.", + "[18] Jason Wei, Maarten Bosma, Vincent Zhao, Kelvin Guu, Adams Wei Yu, Brian Lester, Nan Du, Andrew M Dai, and Quoc V Le. Finetuned language models are zero-shot learners. In International Conference on Learning Representations.", + "[19] Hongxia Xie, Hua Chung, Hong-Han Shuai, and Wen-Huang Cheng. Learning to prompt for vision-language emotion recognition. In 2023 11th International Conference on Affective Computing and Intelligent Interaction Workshops and Demos (ACIIW), pages 1-4, 2023.", + "[20] Hongxia Xie, Ming-Xian Lee, Tzu-Jui Chen, Hung-Jen Chen, Hou-I Liu, Hong-Han Shuai, and Wen-Huang Cheng. Most important person-guided dual-branch cross-patch attention for group affect recognition. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 20598–20608, 2023.", + "[21] Liwen Xu, Zhengtao Wang, Bin Wu, and Simon Lui. Mdan: Multi-level dependent attention network for visual emotion analysis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9479-9488, 2022.", + "[22] Zhiyang Xu, Ying Shen, and Lifu Huang. Multiinstruct: Improving multi-modal zero-shot learning via instruction tuning. arXiv preprint arXiv:2212.10773, 2022.", + "[23] Jingyuan Yang, Jie Li, Xiumei Wang, Yuxuan Ding, and Xinbo Gao. Stimuli-aware visual emotion analysis. IEEE Transactions on Image Processing, 30:7432-7445, 2021.", + "[24] Jingyuan Yang, Qirui Huang, Tingting Ding, Dani Lischinski, Danny Cohen-Or, and Hui Huang. Emoset: A large-scale visual emotion dataset with rich attributes. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 20383-20394, 2023.", + "[25] Quanzeng You, Jiebo Luo, Hailin Jin, and Jianchao Yang. Building a large scale dataset for image emotion recognition: The fine print and the benchmark. In Proceedings of the AAAI conference on artificial intelligence, 2016.", + "[26] Shengyu Zhang, Linfeng Dong, Xiaoya Li, Sen Zhang, Xiaofei Sun, Shuhe Wang, Jiwei Li, Runyi Hu, Tianwei Zhang, Fei Wu, et al. Instruction tuning for large language models: A survey. arXiv preprint arXiv:2308.10792, 2023.", + "[27] Wei Zhang, Xuanyu He, and Weizhi Lu. Exploring discriminative representations for image emotion recognition with" + ], + "bbox": [ + 501, + 92, + 890, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "26604", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "cnns. IEEE Transactions on Multimedia, 22(2):515-523, 2019.", + "[28] Sicheng Zhao, Zizhou Jia, Hui Chen, Leida Li, Guiguang Ding, and Kurt Keutzer. Pd Janet: Polarity-consistent deep attention network for fine-grained visual emotion regression. In Proceedings of the 27th ACM International Conference on Multimedia, pages 192-201, 2019." + ], + "bbox": [ + 78, + 90, + 467, + 190 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "26605", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2024/EmoVIT_ Revolutionizing Emotion Insights with Visual Instruction Tuning/337f29a8-74df-4960-89a4-6eff2797f1cc_model.json b/2024/EmoVIT_ Revolutionizing Emotion Insights with Visual Instruction Tuning/337f29a8-74df-4960-89a4-6eff2797f1cc_model.json new file mode 100644 index 0000000000000000000000000000000000000000..6027b6cc01d7b0c7df95b6c58a4ef5ab28c48538 --- /dev/null +++ b/2024/EmoVIT_ Revolutionizing Emotion Insights with Visual Instruction Tuning/337f29a8-74df-4960-89a4-6eff2797f1cc_model.json @@ -0,0 +1,2387 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.109, + 0.131, + 0.865, + 0.155 + ], + "angle": 0, + "content": "EmoVIT: Revolutionizing Emotion Insights with Visual Instruction Tuning" + }, + { + "type": "text", + "bbox": [ + 0.189, + 0.18, + 0.781, + 0.306 + ], + "angle": 0, + "content": "Hongxia Xie1, Chu-Jun Peng2, Yu-Wen Tseng2, Hung-Jen Chen3, Chan-Feng Hsu3, Hong-Han Shuai3, Wen-Huang Cheng2 \n1Jilin University \n2National Taiwan University \n3National Yang Ming Chiao Tung University \nhongxiaie@jlu.edu.cn, {d12922035, ywt, wenhuang}@csie.ntu.edu.tw, {hjc.eed07g, cfhsu311510211.ee11, hhshuai}@nytu.edu.tw" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.34, + 0.314, + 0.356 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.373, + 0.474, + 0.736 + ], + "angle": 0, + "content": "Visual Instruction Tuning represents a novel learning paradigm involving the fine-tuning of pre-trained language models using task-specific instructions. This paradigm shows promising zero-shot results in various natural language processing tasks but is still unexplored in vision emotion understanding. In this work, we focus on enhancing the model's proficiency in understanding and adhering to instructions related to emotional contexts. Initially, we identify key visual clues critical to visual emotion recognition. Subsequently, we introduce a novel GPT-assisted pipeline for generating emotion visual instruction data, effectively addressing the scarcity of annotated instruction data in this domain. Expanding on the groundwork established by InstructBLIP, our proposed EmoVIT architecture incorporates emotion-specific instruction data, leveraging the powerful capabilities of Large Language Models to enhance performance. Through extensive experiments, our model showcases its proficiency in emotion classification, adeptness in affective reasoning, and competence in comprehending humor. The comparative analysis provides a robust benchmark for Emotion Visual Instruction Tuning in the era of LLMs, providing valuable insights and opening avenues for future exploration in this domain. Our code is available at https://github.com/aimmemotion/EmoVIT." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.784, + 0.21, + 0.799 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.81, + 0.47, + 0.903 + ], + "angle": 0, + "content": "Visual emotion recognition, a key area within artificial intelligence and computer vision, aims to predict human emotions based on visual cues such as facial expressions and body language. This technology is essential in bridging the gap between human affective states and machine understanding. Its diverse applications [19, 20, 23, 24], spanning" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.339, + 0.911, + 0.53 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.54, + 0.91, + 0.569 + ], + "angle": 0, + "content": "Figure 1. Illustration of the importance of instruction-following ability in visual emotion understanding." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.583, + 0.895, + 0.719 + ], + "angle": 0, + "content": "from improving human-computer interaction to aiding in mental health assessment, underscore its significance. Accurate emotion recognition is vital for enhancing user experience and ensuring information security, as it helps prevent emotional manipulation and misinformation [13]. Developing robust emotion recognition models is not only a technical challenge but also a step towards more empathetic and intuitive AI systems, paving the way for more efficient and natural human-computer interactions." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.896, + 0.902 + ], + "angle": 0, + "content": "The AI community has recently shown a growing interest in developing foundational vision models, e.g., Flamingo [1], LLaVA [8], BLIP2 [5]. These models excel in open-world visual understanding, tackling several vision tasks such as classification, detection, segmentation, and captioning. In contrast, current large-scale multimodal models are still in its infancy when it comes to emotion perception [7]. As illustrated in Fig. 1, when directly query the GPT-4 [11] about the emotional category of an image, the model tends to provide incorrect responses. However, the model delivers accurate responses when provided with revised instructions. To fully leverage the potential of ex" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "26596" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.168 + ], + "angle": 0, + "content": "isting vision-based large models, our approach is based on the concept of Instruction Tuning. This effective strategy is aimed at teaching language models to follow natural language instructions, a technique proven to enhance their generalization performance across unseen tasks [2, 8, 18]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.17, + 0.47, + 0.29 + ], + "angle": 0, + "content": "In this work, we focus on developing the model's proficiency in understanding and following instructions related to emotional contexts. This approach highlights the importance of fine-tuning the model's instruction-following capabilities, enabling it to interpret and respond to emotional content effectively. This is achieved by leveraging its pre-existing knowledge base, thereby eliminating the necessity for an emotion-specific architectural framework." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.292, + 0.471, + 0.473 + ], + "angle": 0, + "content": "To address the notable challenges encountered in Instruction Tuning for visual emotion recognition, especially the lack of specific instruction data, we introduce a novel self-generation pipeline explicitly crafted for visual emotion recognition by using GPT-4 [11]. This innovative pipeline excels in generating a diverse array of (image, instruction, output) instances, thereby notably enhancing the dataset with a more extensive and task-oriented variety of examples. This approach not only overcomes the challenge of limited data availability but also reduces the dependence on human labor. Therefore, it streamlines the process, enabling more efficient and effective emotion recognition." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.475, + 0.471, + 0.672 + ], + "angle": 0, + "content": "Additionally, Instruction Tuning has been criticized for its emphasis on surface-level features like output patterns and styles, rather than achieving a profound comprehension and assimilation of tasks [26]. To tackle this issue and enhance the diversity and creativity of instruction data, our dataset includes instructions that demand complex reasoning, going beyond basic question-and-answer formats. This is further enriched by incorporating visual cues such as brightness, colorfulness, scene type, object class, facial expressions, and human actions. These aspects are pivotal in fostering a nuanced comprehension of visual emotions, thus allowing the model to generate more precise and contextually appropriate interpretations [24]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.673, + 0.47, + 0.884 + ], + "angle": 0, + "content": "After generating the emotion visual instruction data, we propose an Emotion Visual Instruction Tuning (EmoVIT) framework, leveraging the foundation of InstructBLIP [2]. This framework incorporates an emotion-centric, instruction-aware module that proficiently guides Large Language Models (LLMs) in assimilating the nuances of emotion instructions. Our work signifies a paradigm shift, presenting a new era of instruction-based learning for visual emotion understanding that relies less on explicit training data. Remarkably, as shown in Fig. 2, our approach requires almost \\(50\\%\\) of the training data typically needed yet exceeds the performance of previous visual emotion recognition methods and popular Visual Instruction Tuning methods." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.886, + 0.422, + 0.9 + ], + "angle": 0, + "content": "Our contributions can be summarized as follows:" + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.088, + 0.916, + 0.349 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.364, + 0.892, + 0.392 + ], + "angle": 0, + "content": "Figure 2. Performance comparison on EmoSet test set [24] (Accuracy \\(\\%\\)" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.417, + 0.892, + 0.463 + ], + "angle": 0, + "content": "- We explore the potential of the Visual Instruction Tuning paradigm for emotion comprehension and introduce the concept of Emotion Visual Instruction Tuning." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.463, + 0.892, + 0.551 + ], + "angle": 0, + "content": "- After thoroughly considering the unique characteristics of visual emotion recognition, we develop a novel GPT-assisted pipeline for generating emotion visual instruction data. This approach effectively bridges the gap in available annotated instruction data within this specific domain." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.553, + 0.892, + 0.657 + ], + "angle": 0, + "content": "- Building upon the foundation of InstructBLIP, our EmoVIT architecture integrates emotion domain-specific instruction data, harnessing the robust capabilities of LLMs to boost performance. The extensive experiments demonstrate our model's proficiency in emotion classification, affective reasoning, and comprehension of humour." + }, + { + "type": "list", + "bbox": [ + 0.5, + 0.417, + 0.892, + 0.657 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.672, + 0.642, + 0.687 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.697, + 0.753, + 0.713 + ], + "angle": 0, + "content": "2.1. Visual Emotion Recognition" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.892, + 0.901 + ], + "angle": 0, + "content": "A key challenge in visual emotion recognition is bridging the gap between an image's visual cues and the emotions it portrays [12, 21, 25]. While traditional efforts, e.g., Xu et al.'s multi-level dependent attention network [21], focus on visual models for emotional feature learning, recent advancements like EmoSet [24] offer rich emotion-laden datasets with 3.3 million images. The rise of multimodal models, such as the GPT series [11], has further propelled Vision-Language Recognition. However, fully leveraging these models in emotion recognition is an area ripe for exploration. Our work leads the way in utilizing large-scale models for Emotion Visual Instruction Tuning." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "26597" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.129, + 0.087, + 0.844, + 0.325 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.296, + 0.331, + 0.673, + 0.347 + ], + "angle": 0, + "content": "Figure 3. The comparison of different visual tuning paradigms." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.359, + 0.313, + 0.376 + ], + "angle": 0, + "content": "2.2. Visual Instruction Tuning" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.383, + 0.471, + 0.609 + ], + "angle": 0, + "content": "Current Large Language Models (LLMs) have extensive knowledge bases, but their effectiveness depends on accurately interpreting human instructions due to a mismatch between training goals and user expectations. LLMs are trained to minimize prediction errors, whereas users expect helpful and safe instruction-following. Instruction Tuning addresses this by teaching models to follow natural language instructions, enhancing generalization to new tasks. FLAN [18] demonstrated that training a large model on instruction-based datasets improves zero-shot performance. This approach has extended to vision-language tasks, with BLIP2 [5] and LLaVA [8] adapting instruction-tuned LLMs for visual inputs. InstructBLIP [2] introduces instruction-aware visual feature extraction and the Q-Former, enabling more flexible, instruction-driven feature extraction." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.61, + 0.47, + 0.685 + ], + "angle": 0, + "content": "As a novel area, visual emotion instruction tuning lacks benchmarks or guidelines for creating emotion instruction data. Our work pioneers the use of large-scale models to develop an emotion instruction data pipeline, overcoming the limitations of manual annotation." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.7, + 0.169, + 0.715 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.726, + 0.429, + 0.743 + ], + "angle": 0, + "content": "3.1. Preliminary of Visual Instruction Tuning" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.47, + 0.78 + ], + "angle": 0, + "content": "In the deep learning era, visual tuning has experienced significant paradigm shifts, as depicted in Fig. 3." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.781, + 0.471, + 0.902 + ], + "angle": 0, + "content": "In Fig. 3(a), conventional tuning methodologies encompass Full fine-tuning, Head-oriented, and Backbone-oriented techniques, capitalizing on large-scale pre-trained models. Predominantly, thoroughly fine-tuning these models for specific tasks, conducted end-to-end, is recognized as a highly effective strategy. However, this method requires maintaining separate copies of the backbone parameters for each distinct task, posing challenges in storage and deploy-" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.362, + 0.542, + 0.373 + ], + "angle": 0, + "content": "ment." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.376, + 0.892, + 0.467 + ], + "angle": 0, + "content": "Alternatively, Visual Prompt Tuning (VPT) [3], presents an efficient substitute for full fine-tuning within large-scale vision Transformer models. It achieves this by employing a minimal fraction of trainable parameters in the input space while maintaining a frozen backbone model. The objective function for Visual Prompt Tuning is given by:" + }, + { + "type": "equation", + "bbox": [ + 0.617, + 0.481, + 0.892, + 0.505 + ], + "angle": 0, + "content": "\\[\n\\min _ {\\theta_ {\\mathrm {P}}} \\mathcal {L} (f (X, P; \\theta_ {\\mathrm {P}}), Y) \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.517, + 0.892, + 0.592 + ], + "angle": 0, + "content": "where \\(\\min_{\\theta_{\\mathrm{P}}}\\) is the minimization over the prompt parameters \\(P\\), \\(\\mathcal{L}\\) is the loss function, \\(f\\) represents the model function with input image \\(X\\), prompt parameters \\(P\\), and learnable model parameters \\(\\theta_{\\mathrm{P}}\\) as input, and \\(Y\\) is the target output." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.594, + 0.893, + 0.834 + ], + "angle": 0, + "content": "Visual Prompt Tuning focuses on optimizing LLMs using a small set of parameters, whereas Visual Instruction Tuning (VIT) aims to improve the model's comprehension of instructions, thereby addressing the model's shortcomings in specific domains. This type of method aims to enhance the model's proficiency in following instructions, leveraging the capabilities of the latest foundation models, e.g., Llama [17], and BLIP2 [5]. Instructions serve as guiding constraints, shaping the model's outputs to conform to specific response characteristics and domain-relevant knowledge. This approach enables human monitoring of the model's behavior, thereby assuring alignment with the desired outcomes. Moreover, Instruction Tuning is computationally efficient, allowing LLMs to swiftly adapt to particular domains without extensive retraining or architectural alterations." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.837, + 0.892, + 0.867 + ], + "angle": 0, + "content": "The objective function for Visual Instruction Tuning is given by:" + }, + { + "type": "equation", + "bbox": [ + 0.594, + 0.881, + 0.892, + 0.905 + ], + "angle": 0, + "content": "\\[\n\\min _ {\\theta_ {\\text {t u n a b l e}}} \\mathcal {L} (g (X, I, C; \\theta_ {\\text {t u n a b l e}}), Y) \\tag {2}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "26598" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.084, + 0.124, + 0.225, + 0.201 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.099, + 0.206, + 0.157, + 0.214 + ], + "angle": 0, + "content": "Emotion class" + }, + { + "type": "image_caption", + "bbox": [ + 0.165, + 0.206, + 0.216, + 0.213 + ], + "angle": 0, + "content": "Contentment" + }, + { + "type": "image_caption", + "bbox": [ + 0.099, + 0.221, + 0.149, + 0.229 + ], + "angle": 0, + "content": "Brightness" + }, + { + "type": "image_caption", + "bbox": [ + 0.166, + 0.221, + 0.18, + 0.228 + ], + "angle": 0, + "content": "0.8" + }, + { + "type": "image_caption", + "bbox": [ + 0.099, + 0.237, + 0.152, + 0.244 + ], + "angle": 0, + "content": "Colorfulness" + }, + { + "type": "image_caption", + "bbox": [ + 0.166, + 0.237, + 0.18, + 0.244 + ], + "angle": 0, + "content": "0.8" + }, + { + "type": "image_caption", + "bbox": [ + 0.099, + 0.252, + 0.149, + 0.26 + ], + "angle": 0, + "content": "Scene type" + }, + { + "type": "image_caption", + "bbox": [ + 0.166, + 0.252, + 0.194, + 0.259 + ], + "angle": 0, + "content": "Beach" + }, + { + "type": "image_caption", + "bbox": [ + 0.099, + 0.267, + 0.149, + 0.275 + ], + "angle": 0, + "content": "Object class" + }, + { + "type": "image_caption", + "bbox": [ + 0.166, + 0.267, + 0.18, + 0.274 + ], + "angle": 0, + "content": "Sky" + }, + { + "type": "image_caption", + "bbox": [ + 0.099, + 0.283, + 0.149, + 0.29 + ], + "angle": 0, + "content": "Facial" + }, + { + "type": "image_caption", + "bbox": [ + 0.166, + 0.284, + 0.216, + 0.291 + ], + "angle": 0, + "content": "Contentment" + }, + { + "type": "image_caption", + "bbox": [ + 0.099, + 0.301, + 0.14, + 0.308 + ], + "angle": 0, + "content": "Human" + }, + { + "type": "image_caption", + "bbox": [ + 0.166, + 0.302, + 0.205, + 0.309 + ], + "angle": 0, + "content": "Laughing." + }, + { + "type": "image_caption", + "bbox": [ + 0.166, + 0.309, + 0.205, + 0.316 + ], + "angle": 0, + "content": "Ludwig" + }, + { + "type": "image", + "bbox": [ + 0.231, + 0.105, + 0.391, + 0.332 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.111, + 0.334, + 0.357, + 0.346 + ], + "angle": 0, + "content": "(a) Emotion Visual Instruction Data Generation" + }, + { + "type": "image", + "bbox": [ + 0.396, + 0.098, + 0.668, + 0.33 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.396, + 0.334, + 0.663, + 0.346 + ], + "angle": 0, + "content": "(b) Emotion Visual Instruction Tuning Architecture" + }, + { + "type": "image", + "bbox": [ + 0.695, + 0.125, + 0.895, + 0.316 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.696, + 0.334, + 0.886, + 0.346 + ], + "angle": 0, + "content": "(c) The Details of Q-Former Module" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.364, + 0.892, + 0.435 + ], + "angle": 0, + "content": "Figure 4. The overall architecture of our proposed method. The Emotion Instruction data generated by (a) will be used for Emotion Visual Instruction Tuning in (b). During Emotion Visual Instruction Tuning, given an input image, the frozen Image Encoder initiates the process by extracting visual features. Emotion Instruction generated by (a) are subsequently interacting with Queries embedding through the learnable Q-Former. This interaction is key to drawing out image features that are relevant to the task at hand. As a result, the frozen LLM receives visual information conducive to instruction following." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.448, + 0.471, + 0.599 + ], + "angle": 0, + "content": "where \\(\\min_{\\theta_{\\mathrm{tunable}}}\\) denotes the minimization over the tunable parameters \\(\\theta_{\\mathrm{tunable}}\\) in the Instruction Tuning Module, \\(\\mathcal{L}\\) is the loss function, \\(g\\) is the model function with instruction \\(I\\), image \\(X\\), other contexts \\(C\\), and tunable parameters \\(\\theta_{\\mathrm{tunable}}\\) and \\(Y\\) denotes the target output. The optional context \\(C\\) is not just raw data; it encompasses descriptive or directive information guiding the model on how to process input or which task to execute, e.g., image caption. It's integral to the model's understanding and execution of tasks based on specific instructions or guidelines." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.606, + 0.469, + 0.636 + ], + "angle": 0, + "content": "3.2. GPT-assisted Emotion Visual Instruction Data Generation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.645, + 0.47, + 0.809 + ], + "angle": 0, + "content": "Previous methodologies commonly employed a consistent template-based set of instructions for every image within a dataset across various specific tasks [2]. For instance, a standard instruction such as \"Briefly describe the content of the image\" was employed uniformly across all images for Image Captioning. In this way, the model may not be able to adequately capture the unique characteristics of each image. Moreover, this one-size-fits-all approach often leads to suboptimal performance in emotion recognition tasks that require nuanced perception and differentiation of ambiguous emotion classes." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.811, + 0.469, + 0.902 + ], + "angle": 0, + "content": "Since the topic of Emotion Visual Instruction Tuning is still in its infancy, no benchmarks or guidelines have been proposed so far for constructing emotion instruction data. Based on the recent successes of machine-generated instructions demonstrated in LLaVA [8], our work pioneers the use of existing LLMs to create a pipeline for" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.448, + 0.892, + 0.553 + ], + "angle": 0, + "content": "self-generating emotion instructions. Different from previous template-based and one-size-fits-all instruction data, we propose an instance-wise and LLM-assisted visual emotion instruction data pipeline. This methodology transcends the constraints of manual annotation by employing GPT-4 [11] to generate instance-wise, tailored instruction data that dynamically corresponds to visual content." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.554, + 0.892, + 0.84 + ], + "angle": 0, + "content": "Prior to the development of instructional data for the visual emotion recognition task, it is imperative to confront a fundamental academic problem: What types of visual clues are pivotal in identifying emotions? This necessitates a careful consideration of the unique characteristics inherent to the task, along with a comprehensive understanding of the potential visual cues associated with human emotions. In this work, we propose a novel visual instruction data mechanism to remove the inherent subjectivity and ambiguity in emotional interpretation. Specifically, we integrate a broad spectrum of emotion attributes across multiple levels: low-level attributes (e.g., brightness, colorfulness), mid-level attributes (e.g., scene type and object class), and high-level attributes (e.g., facial expressions and human actions), building upon insights from previous work [24]. This comprehensive strategy not only aligns with the intricate nature of emotions but also significantly enhances the model's capability to interpret and understand visual emotional cues more accurately and holistically." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.892, + 0.902 + ], + "angle": 0, + "content": "The overall pipeline of our proposed emotion visual instruction data is shown in Fig. 4 (a). For an image \\( X_{\\mathrm{img}} \\), three types of image-related contexts are essential for GPT-4 to generate emotion instruction data: (i) a caption \\( X_{c} \\), (ii)" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "26599" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.166 + ], + "angle": 0, + "content": "an emotion attribute list \\( X_{\\mathrm{attr}} \\), which includes emotion class, brightness, colorfulness, scene type, object class, facial expression, and human action, and (iii) the system prompt, designed to enable GPT-4 to comprehend the specific task requirement1." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.168, + 0.471, + 0.317 + ], + "angle": 0, + "content": "We first manually design a few examples which are used as seed examples for in-context learning to query GPT-4. This operation leverages the model's ability to extrapolate from given examples, enhancing its understanding and response accuracy based on the principles of few-shot learning [8]. Our generated emotion instruction data includes three types: Categorical, Conversation, and Reasoning. Building upon previous research [8], our generated instruction data adheres to the dialogue format, exemplified in Fig. 5." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.319, + 0.47, + 0.408 + ], + "angle": 0, + "content": "Our strategy for generating emotion instruction data adopts a progressive approach from simple to complex. Initially, for the Categorical data, we transform the associated emotion class of the image into a structured format. This process serves as the foundational component of our emotion instruction data." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.41, + 0.47, + 0.62 + ], + "angle": 0, + "content": "For the Conversation data, our framework is designed to create dialogues in which the GPT assistant interacts with an inquirer, focusing on the emotion attributes of the image. In this setup, the assistant's responses are tailored to interpret and describe the image as though it were within its own visual field, thereby providing insights from an observational viewpoint. The scope of questions posed is comprehensive, encompassing the types of objects depicted, their actions, and the dynamics of their interrelationships. The dialogues we generate fall into two categories: (i) Basic Interaction, focusing on the provided emotion attribute list with simple, direct characteristics, and (ii) Advanced Interaction, which builds on the first type to reach greater conversational complexity and sophistication." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.621, + 0.47, + 0.725 + ], + "angle": 0, + "content": "For the Reasoning data, our approach extends beyond mere visual content, prompting the model to generate in-depth reasoning questions. To enhance the dialogue's credibility and structure, detailed examples are incorporated alongside logical reasoning steps, ensuring that the discourse convincingly captures the intricacies of the visual content." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.736, + 0.383, + 0.753 + ], + "angle": 0, + "content": "3.3. Emotion Visual Instruction Tuning" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.759, + 0.47, + 0.835 + ], + "angle": 0, + "content": "After acquiring the emotion visual instruction data as detailed in Sec. 3.2, our goal is to employ this data in enhancing the existing Visual Instruction Tuning model. This enhancement aims to align the LLMs' existing knowledge with the emotion understanding domain." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.836, + 0.47, + 0.866 + ], + "angle": 0, + "content": "As shown in Fig. 4 b, we have developed an Emotion Visual Instruction Tuning (EmoVIT) architecture based on" + }, + { + "type": "page_footnote", + "bbox": [ + 0.077, + 0.875, + 0.469, + 0.901 + ], + "angle": 0, + "content": "1A detailed description of the system prompt is provided in the supplementary materials." + }, + { + "type": "title", + "bbox": [ + 0.61, + 0.094, + 0.804, + 0.106 + ], + "angle": 0, + "content": "Our Emotion Instruction Data Sample" + }, + { + "type": "image", + "bbox": [ + 0.524, + 0.115, + 0.636, + 0.179 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.641, + 0.119, + 0.748, + 0.128 + ], + "angle": 0, + "content": "Context type 1: Caption:" + }, + { + "type": "text", + "bbox": [ + 0.641, + 0.129, + 0.805, + 0.137 + ], + "angle": 0, + "content": "Two young women dancing in a kitchen." + }, + { + "type": "text", + "bbox": [ + 0.641, + 0.142, + 0.754, + 0.15 + ], + "angle": 0, + "content": "Context type 2: Attribute:" + }, + { + "type": "text", + "bbox": [ + 0.642, + 0.151, + 0.857, + 0.159 + ], + "angle": 0, + "content": "Emotion: contentment Human action: stretching arm" + }, + { + "type": "text", + "bbox": [ + 0.642, + 0.16, + 0.892, + 0.168 + ], + "angle": 0, + "content": "Brightness: 0.7 Object: ['Cupboard', 'Food', 'Countertop']" + }, + { + "type": "text", + "bbox": [ + 0.642, + 0.17, + 0.836, + 0.178 + ], + "angle": 0, + "content": "Colorfulness: 0.4 Facial expression: happy" + }, + { + "type": "title", + "bbox": [ + 0.547, + 0.199, + 0.605, + 0.209 + ], + "angle": 0, + "content": "Categorical" + }, + { + "type": "text", + "bbox": [ + 0.525, + 0.212, + 0.878, + 0.243 + ], + "angle": 0, + "content": "Question: Please select the emotion closest to the image from the following options: amusement, anger, awe, contentment, disgust, excitement, fear and sadness." + }, + { + "type": "text", + "bbox": [ + 0.526, + 0.244, + 0.714, + 0.254 + ], + "angle": 0, + "content": "Answer: Predict emotion: contentment." + }, + { + "type": "title", + "bbox": [ + 0.543, + 0.262, + 0.608, + 0.271 + ], + "angle": 0, + "content": "Conversation" + }, + { + "type": "title", + "bbox": [ + 0.638, + 0.262, + 0.718, + 0.271 + ], + "angle": 0, + "content": "Basic Interaction" + }, + { + "type": "text", + "bbox": [ + 0.526, + 0.274, + 0.828, + 0.285 + ], + "angle": 0, + "content": "Question: Can you describe the activity happening in the image?" + }, + { + "type": "text", + "bbox": [ + 0.526, + 0.285, + 0.861, + 0.317 + ], + "angle": 0, + "content": "Answer: The image depicts two young women enjoying themselves by dancing in a kitchen. One of the women is seen stretching her arm, most likely as part of their dance routine." + }, + { + "type": "title", + "bbox": [ + 0.543, + 0.322, + 0.608, + 0.332 + ], + "angle": 0, + "content": "Conversation" + }, + { + "type": "title", + "bbox": [ + 0.627, + 0.321, + 0.729, + 0.332 + ], + "angle": 0, + "content": "Advanced Interaction" + }, + { + "type": "text", + "bbox": [ + 0.525, + 0.335, + 0.888, + 0.356 + ], + "angle": 0, + "content": "Question: Based on the given emotion of contentment and the activities in the image, can you guess about the possible context or story of this scenario?" + }, + { + "type": "text", + "bbox": [ + 0.525, + 0.356, + 0.886, + 0.397 + ], + "angle": 0, + "content": "Answer: These women are good friends or roommates, spending quality time together. The act of dancing, especially in an unconventional place like a kitchen, suggests they are comfortable with each other and are enjoying a break from routine." + }, + { + "type": "title", + "bbox": [ + 0.549, + 0.405, + 0.602, + 0.416 + ], + "angle": 0, + "content": "Reasoning" + }, + { + "type": "text", + "bbox": [ + 0.524, + 0.418, + 0.892, + 0.471 + ], + "angle": 0, + "content": "Question: Based on the setting and activities of the two young women, what might have prompted this impromptu dance session in the middle of a kitchen? Answer: The impromptu dance could have been prompted by a favorite song playing, or simple spontaneous fun amidst preparing a meal or organizing the kitchen." + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.492, + 0.91, + 0.519 + ], + "angle": 0, + "content": "Figure 5. The sample of our generated visual emotion instruction data." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.536, + 0.892, + 0.581 + ], + "angle": 0, + "content": "InstructBLIP [2]. This architecture specifically leverages its Instruction-aware Q-Former Module, as depicted in Fig. 4 c, for emotion-centric instructional tasks." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.583, + 0.893, + 0.779 + ], + "angle": 0, + "content": "Specifically, the Instruction-aware Q-Former Module takes in the emotion instruction tokens, queries, and image embeddings as input. The image embeddings are extracted by a frozen image encoder. The learnable queries are initially produced by the pre-trained Q-Former of InstructBLIP. During training, the Instruction-aware module enhances task-specific feature extraction. It does this by integrating emotion instruction and query embeddings within self-attention layers, aligning visual information with the LLM's instruction-following requirements. Our approach adopts cross-entropy loss, tailoring it to the intricacies of visual emotion recognition tasks, thus ensuring precise and contextually relevant model training outcomes." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.78, + 0.894, + 0.902 + ], + "angle": 0, + "content": "We note that the data generated by our approach is not confined to a single model but can also be applied to other Visual Instruction Tuning models, such as LLaVA [17]. Notably, when LLaVA is fine-tuned with our data, it exhibits a significant enhancement in emotion recognition capabilities, as detailed in Sec. 4.2. In this way, we demonstrate not only the effectiveness but also the transferability of our generated data, showing its broad applicability and impact." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "26600" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.077, + 0.09, + 0.283, + 0.108 + ], + "angle": 0, + "content": "4. Experimental Results" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.115, + 0.27, + 0.132 + ], + "angle": 0, + "content": "4.1. Implemental Details" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.138, + 0.468, + 0.228 + ], + "angle": 0, + "content": "Our implementation is based on the LAVIS library [4]. Our EmoVIT starts with a pre-trained InstructBLIP baseline and proceeds to fine-tune exclusively the Q-Former module, whilst keeping both the image encoder and the language model frozen. The parameters for our training adhere to the default settings established by InstructBLIP." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.229, + 0.468, + 0.319 + ], + "angle": 0, + "content": "Datasets. We evaluate our framework on ten benchmark datasets annotated under different scenarios and class number, namely EmoSet [24], WEBEmo [12], Emotion6 [14], the Flickr and Instagram (FI) [25], Artphoto [9], IAPS [10], Abstract [9], EmotionROI [15], UnbiasedEmo [12], and OxfordTVG-HIC [6]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.32, + 0.468, + 0.44 + ], + "angle": 0, + "content": "Held-in Pretraining. Following previous work [2], we divide our dataset into two categories: held-in for pretraining and held-out for evaluation2. Considering the EmoSSet dataset's comprehensive inclusion of emotion attributes for each image, it has been chosen as the primary resource for our held-in pretraining phase. Simultaneously, for a broader assessment, we perform held-out evaluations using the test sets from various other datasets." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.442, + 0.468, + 0.547 + ], + "angle": 0, + "content": "For the generation of emotion visual instruction data, we initially employ the BLIP2 model for image captioning, followed by leveraging the GPT-4 API to generate emotion instruction data. In total, our collection comprises Categorical, Conversation, and Reasoning instruction data, derived from 51,200 unique images. This represents less than \\(50\\%\\) of the entire EmoSet." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.558, + 0.27, + 0.573 + ], + "angle": 0, + "content": "4.2. Held-out Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.581, + 0.468, + 0.716 + ], + "angle": 0, + "content": "As shown in Tab. 1, our proposed methodology exhibits a marked superiority in performance relative to the burgeoning Visual Instruction Tuning Methods. Since they have been pre-trained on dozens of large-scale datasets, it is evident that our generated emotion visual instruction data is particularly effective for emotional understanding. Our results signify a paradigm shift, heralding a new era of model training that relies less on explicit supervision and more on the robustness of emotion instruction-driven learning." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.717, + 0.468, + 0.853 + ], + "angle": 0, + "content": "The Effectiveness of Our Proposed Emotion Visual Instruction Data. As the first to introduce the concept of emotion visual instruction data, our study seeks to evaluate the generalizability of this newly generated instruction data. Our goal is to test its efficacy not only with InstructBLIP but also across other Visual Instruction Tuning model, to understand its broader applicability. As depicted in Fig. 6, we employ two Visual Instruction Tuning models, LLaVA and InstructBLIP, which were fine-tuned on our specially gen" + }, + { + "type": "image_caption", + "bbox": [ + 0.568, + 0.094, + 0.876, + 0.105 + ], + "angle": 0, + "content": "Accuracy Improvements of FT Models Over Base Models" + }, + { + "type": "image", + "bbox": [ + 0.507, + 0.106, + 0.909, + 0.323 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.338, + 0.91, + 0.366 + ], + "angle": 0, + "content": "Figure 6. The improvement of our proposed emotion visual instruction tuning data tuning on LLaVA [8] and InstructBLIP [2]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.383, + 0.892, + 0.534 + ], + "angle": 0, + "content": "erated emotion visual instruction data. Subsequent testing across five distinct datasets reveals notable improvements in both models, substantiating the efficacy of our generated data. Notably, InstructBLIP demonstrated a more substantial overall enhancement compared to LLaVA. This can be attributed to InstructBLIP's specialized Instruction-aware Q-Former Module, which adeptly extracts the salient features of our emotion instructions and synergizes them effectively with the corresponding images, thereby yielding improved performance." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.548, + 0.861, + 0.563 + ], + "angle": 0, + "content": "4.3. Effectiveness of Different Instruction Data" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.573, + 0.863, + 0.587 + ], + "angle": 0, + "content": "4.3.1 Ablation Study of Different Instruction Data" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.599, + 0.892, + 0.901 + ], + "angle": 0, + "content": "The ablation study outlined in Tab. 2 provides a comprehensive analysis of the impact that different instructional data types have on model performance, specifically concerning accuracy metrics on the EmoSet test set. Initially, the model, referred to as InstructBLIP [2], operates without the integration of the three types of instructional data and attains a baseline accuracy of \\(42.20\\%\\). This foundational performance is significantly enhanced with the inclusion of Categorical data, which alone contributes to a substantial increase in accuracy. The introduction of Conversation data further amplifies this effect, underscoring the value of conversational context in improving the model's predictive capabilities. The addition of Reasoning data notably boosts performance, achieving a peak accuracy of \\(83.36\\%\\). This indicates that the model significantly benefits from the nuanced cues in reasoning, aiding in understanding complex emotional instructions. The gradual improvements with each data type support the idea that a diverse approach to instructional data markedly enhances model comprehension and performance." + }, + { + "type": "page_footnote", + "bbox": [ + 0.076, + 0.863, + 0.468, + 0.9 + ], + "angle": 0, + "content": "2Unlike the setup in InstructBLIP, our dataset exclusively comprises emotion-related content. Consequently, our held-out evaluation does not constitute a strict zero-shot evaluation in the conventional sense." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "26601" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.08, + 0.089, + 0.895, + 0.219 + ], + "angle": 0, + "content": "
MethodWebEmoFIEmotion6AbstractArtPhotoIAPSaEmotionROIEmoSet
Number of Classes258688868
Flanmingo [1]9.3614.9121.673.5717.510.1321.7229.59
LLaVA [8]12.5556.0449.4419.5436.2542.4346.4644.03
BLIP2 [5]20.1057.7250.0028.5736.2539.2450.5146.79
InstructBLIP [2]12.8037.9746.1121.4226.2534.1846.1342.20
Ours*21.1268.0957.8132.3444.9044.1353.8783.36
" + }, + { + "type": "table_caption", + "bbox": [ + 0.261, + 0.228, + 0.707, + 0.242 + ], + "angle": 0, + "content": "Table 1. Held-out performance comparison on visual emotion datasets \\((\\%)\\)" + }, + { + "type": "table", + "bbox": [ + 0.079, + 0.266, + 0.48, + 0.34 + ], + "angle": 0, + "content": "
CategoricalConversationReasoningAccuracy (%)
---42.20
✓--80.90 (+38.70)
✓✓-81.95 (+39.75)
✓✓✓83.36 (+41.16)
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.076, + 0.351, + 0.47, + 0.379 + ], + "angle": 0, + "content": "Table 2. Ablation study of three types of instruction data. Accuracy (%) on EmoSet test set." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.389, + 0.285, + 0.403 + ], + "angle": 0, + "content": "4.3.2 Instruction Sensitivity" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.413, + 0.47, + 0.548 + ], + "angle": 0, + "content": "This work is dedicated to the creation of a varied corpus of visual emotion instruction data, alongside the development of a robust instruction-based model. Our objective is for the model to demonstrate stability, producing consistent results in the face of minor variations in instruction phrasing, provided the core objective of the task persists unchanged. To this end, we employ the Sensitivity evaluation metric, as introduced by [22], to assess the model's fidelity in generating uniform outcomes irrespective of instructional nuances." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.55, + 0.47, + 0.744 + ], + "angle": 0, + "content": "We employ two semantically similar instructions as input prompts for the model, testing their impact on the Sensitivity score across three visual emotion datasets for different Visual Instruction Tuning models. The first instruction is: \"From the given options: cls_1, cls_2, cls_3, etc., identify the emotion that most accurately reflects the image. Ensure your selection is confined to the listed options. Respond in the format: Predicted emotion.\" The second one states: \"Please choose the emotion that best corresponds to the image from the following options: cls_1, cls_2, cls_3, etc. (Do not provide answers beyond the provided candidates.) Please reply in the following format: Predict emotion.\"" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.746, + 0.47, + 0.837 + ], + "angle": 0, + "content": "As illustrated in Fig. 7, our approach, along with BLIP2, exhibited exceptionally low Sensitivity values, demonstrating robustness in understanding the instructions. Conversely, Flamingo and InstructBLIP displayed a higher degree of sensitivity, indicating a relative susceptibility to variations in instruction wording." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.847, + 0.202, + 0.861 + ], + "angle": 0, + "content": "4.4. Robustness" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.871, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Given that current emotion recognition datasets often exhibit category imbalances and labeling biases, our aim is" + }, + { + "type": "image", + "bbox": [ + 0.507, + 0.271, + 0.916, + 0.476 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.511, + 0.493, + 0.898, + 0.507 + ], + "angle": 0, + "content": "Figure 7. The sensitivity score comparison (the lower the better)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.538, + 0.892, + 0.644 + ], + "angle": 0, + "content": "to evaluate the generalization ability of various learning strategies more impartially. Hence, we selected the Un-BiasedEmo test set [12], which is uniquely suited for recognizing intricate emotions, such as those associated with identical objects or scenes, e.g., landscapes, crowds, families, babies, and animals, where the emotional undertones can be particularly subtle and complex." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.648, + 0.892, + 0.724 + ], + "angle": 0, + "content": "As depicted in Tab. 3, our proposed methodology demonstrates superior performance when benchmarked against conventional supervised emotion recognition techniques, thereby underscoring the efficacy of our approach in more accurately discerning complex emotional contexts." + }, + { + "type": "table", + "bbox": [ + 0.536, + 0.745, + 0.858, + 0.852 + ], + "angle": 0, + "content": "
MethodAccuracy (%)
Direct Learning [12]71.64
Self-Directed Learning [12]72.45
Joint Learning [12]71.64
Curriculum Learning [12]74.27
Ours*74.72
" + }, + { + "type": "table_caption", + "bbox": [ + 0.518, + 0.863, + 0.874, + 0.877 + ], + "angle": 0, + "content": "Table 3. Performance comparison on UnbiasedEmo dataset." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "26602" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.122, + 0.105, + 0.244, + 0.16 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.323, + 0.105, + 0.423, + 0.164 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.096, + 0.166, + 0.248, + 0.18 + ], + "angle": 0, + "content": "Predicted emotion: anger" + }, + { + "type": "text", + "bbox": [ + 0.162, + 0.194, + 0.217, + 0.204 + ], + "angle": 0, + "content": "BECAUSE" + }, + { + "type": "image", + "bbox": [ + 0.248, + 0.171, + 0.316, + 0.223 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.313, + 0.168, + 0.475, + 0.179 + ], + "angle": 0, + "content": "Predicted emotion: sadness" + }, + { + "type": "text", + "bbox": [ + 0.345, + 0.196, + 0.4, + 0.205 + ], + "angle": 0, + "content": "BECAUSE" + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.226, + 0.273, + 0.269 + ], + "angle": 0, + "content": "Reason: The image depicts a group of people with their arms up in the air, which is a common gesture associated with anger." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.27, + 0.278, + 0.302 + ], + "angle": 0, + "content": "Additionally, the smoke billowing from the building in the background adds to the sense of chaos and unrest." + }, + { + "type": "text", + "bbox": [ + 0.29, + 0.226, + 0.476, + 0.258 + ], + "angle": 0, + "content": "Reason: The image is a black and white photograph of an old, abandoned house in the middle of a field." + }, + { + "type": "text", + "bbox": [ + 0.291, + 0.269, + 0.476, + 0.302 + ], + "angle": 0, + "content": "The scene is desolate and eerie, which suggests a sense of sadness or melancholy." + }, + { + "type": "image_caption", + "bbox": [ + 0.131, + 0.334, + 0.434, + 0.348 + ], + "angle": 0, + "content": "Figure 8. The sample of our generated explanation." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.351, + 0.271, + 0.367 + ], + "angle": 0, + "content": "4.4.1 Affective Reasoning" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.375, + 0.469, + 0.464 + ], + "angle": 0, + "content": "In the domain of visual emotion recognition, where ambiguity and subjectivity are pervasive, the advent of an interpretable model is of considerable value. Such a model elucidates its cognitive processes, enhancing its trustworthiness and practicality in scenarios requiring a delicate grasp of emotional subtleties." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.466, + 0.469, + 0.54 + ], + "angle": 0, + "content": "Leveraging Visual Instruction Tuning, our model transcends mere categorization of emotions; it articulates the underlying rationale for its classifications. The executing commands for identifying emotions and elucidating the decision basis is illustrated below:" + }, + { + "type": "text", + "bbox": [ + 0.136, + 0.554, + 0.421, + 0.568 + ], + "angle": 0, + "content": "Predicted emotion: [emotion]." + }, + { + "type": "text", + "bbox": [ + 0.137, + 0.569, + 0.353, + 0.582 + ], + "angle": 0, + "content": "Reason: [explanation]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.595, + 0.469, + 0.64 + ], + "angle": 0, + "content": "Our model delineates the visual features influencing its determinations, thereby addressing the complexities inherent in discerning and explaining emotion-related nuances." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.641, + 0.469, + 0.716 + ], + "angle": 0, + "content": "The explanations provide us with visual clues contained within the images, as exemplified in Fig. 8. It provides interpretable visual indicators that inform the model's outputs, as demonstrated in our example, by disambiguating the often abstract emotional categories." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.724, + 0.208, + 0.74 + ], + "angle": 0, + "content": "4.5. Scaling Law" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.748, + 0.469, + 0.824 + ], + "angle": 0, + "content": "Pretraining data. As demonstrated in Tab. 4, there is a clear correlation between the size of the pre-training dataset and improved performance. Consequently, we anticipate that an increase in training data in the future could enhance the effectiveness of Emotion Visual Instruction Tuning." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.833, + 0.337, + 0.849 + ], + "angle": 0, + "content": "4.6. Humour Caption Generation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.469, + 0.901 + ], + "angle": 0, + "content": "The comprehension of humor is intricately linked to the understanding of emotions. Leveraging our generative language model, we conduct a caption generation task without" + }, + { + "type": "table", + "bbox": [ + 0.589, + 0.089, + 0.804, + 0.12 + ], + "angle": 0, + "content": "
5%10%30%50%
79.0081.0079.3483.36
" + }, + { + "type": "table_caption", + "bbox": [ + 0.5, + 0.131, + 0.89, + 0.16 + ], + "angle": 0, + "content": "Table 4. Ablation study of different portion of pre-training data. Accuracy \\((\\%)\\) on EmoSet test set." + }, + { + "type": "image", + "bbox": [ + 0.526, + 0.175, + 0.887, + 0.38 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.39, + 0.91, + 0.419 + ], + "angle": 0, + "content": "Figure 9. The sample of our generated humour caption vs human writing humour caption from OxfordTVG-HIC." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.421, + 0.892, + 0.588 + ], + "angle": 0, + "content": "modifying the model's architecture, specifically testing the model's proficiency in generating humorous captions. For this purpose, we select 50 images from the OxfordTVGHIC dataset [6] and generate corresponding captions using our model. Subsequently, the captions produced by our model are compared with manually annotated captions from the dataset in a user study. Thirty participants were asked to vote on which captions were more humorous. Our model-generated captions receive \\(60\\%\\) of the votes, demonstrating its effective humor generation capabilities. One sample is visualized in Fig. 9." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.602, + 0.619, + 0.617 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.627, + 0.892, + 0.792 + ], + "angle": 0, + "content": "In our study, drawing upon the distinctive visual cues key to visual emotion recognition, we present a GPT-assisted pipeline specifically designed for generating emotion visual instruction data. The developed EmoVIT model incorporates emotion-specific instructions, leveraging LLMs for enhanced performance. Our comprehensive experiments validate its effectiveness in emotion classification, affective reasoning, and humor understanding. This comparative analysis sets a benchmark for Emotion Visual Instruction Tuning with LLMs, providing valuable insights and directions for future research in this field." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.807, + 0.66, + 0.823 + ], + "angle": 0, + "content": "Acknowledgement" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.83, + 0.892, + 0.9 + ], + "angle": 0, + "content": "This work is partially supported by the National Science and Technology Council, Taiwan under Grant, NSTC-112-2628-E-002-033-MY4, NSTC-112-2634-F-002-002-MBK, NSTC-112-26234-F-002-005, NSTC-112-2221-E-A49-059-MY3, and NSTC-112-2221-E-A49-094-MY3." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "26603" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.197 + ], + "angle": 0, + "content": "[1] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. In Advances in Neural Information Processing Systems, pages 23716-23736, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.2, + 0.472, + 0.268 + ], + "angle": 0, + "content": "[2] Wenliang Dai, Junnan Li, Dongxu Li, Anthony Meng Huat Tiong, Junqi Zhao, Weisheng Wang, Boyang Li, Pascale Fung, and Steven Hoi. Instructclip: Towards general-purpose vision-language models with instruction tuning. In Advances in Neural Information Processing Systems, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.27, + 0.468, + 0.324 + ], + "angle": 0, + "content": "[3] Menglin Jia, Luming Tang, Bor-Chun Chen, Claire Cardie, Serge Belongie, Bharath Hariharan, and Ser-Nam Lim. Visual prompt tuning. In Proceedings of the European Conference on Computer Vision, pages 709-727. Springer, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.326, + 0.468, + 0.367 + ], + "angle": 0, + "content": "[4] Dongxu Li, Junnan Li, Hung Le, Guangsen Wang, Silvio Savarese, and Steven CH Hoi. Lavis: A library for language-vision intelligence. arXiv preprint arXiv:2209.09019, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.368, + 0.468, + 0.436 + ], + "angle": 0, + "content": "[5] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In Proceedings of the International Conference on Machine Learning, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.438, + 0.468, + 0.492 + ], + "angle": 0, + "content": "[6] Runjia Li, Shuyang Sun, Mohamed Elhoseiny, and Philip Torr. Oxfordtvg-hic: Can machine make humorous captions from images? In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 20293-20303, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.494, + 0.468, + 0.548 + ], + "angle": 0, + "content": "[7] Zheng Lian, Licai Sun, Mingyu Xu, Haiyang Sun, Ke Xu, Zhuofan Wen, Shun Chen, Bin Liu, and Jianhua Tao. Explainable multimodal emotion reasoning. arXiv preprint arXiv:2306.15401, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.551, + 0.468, + 0.591 + ], + "angle": 0, + "content": "[8] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. In Advances in Neural Information Processing Systems, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.593, + 0.468, + 0.647 + ], + "angle": 0, + "content": "[9] Jana Machajdik and Allan Hanbury. Affective image classification using features inspired by psychology and art theory. In Proceedings of the 18th ACM international conference on Multimedia, pages 83-92, 2010." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.649, + 0.468, + 0.716 + ], + "angle": 0, + "content": "[10] Joseph A Mikels, Barbara L Fredrickson, Gregory R Larkin, Casey M Lindberg, Sam J Maglio, and Patricia A Reuter-Lorenz. Emotional category data on images from the international affective picture system. Behavior research methods, 37:626-630, 2005." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.719, + 0.445, + 0.733 + ], + "angle": 0, + "content": "[11] OpenAI. Gpt-4 technical report. Technical report, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.735, + 0.468, + 0.802 + ], + "angle": 0, + "content": "[12] Rameswar Panda, Jianming Zhang, Haoxiang Li, Joon-Young Lee, Xin Lu, and Amit K Roy-Chowdhury. Contemplating visual emotions: Understanding and overcoming dataset bias. In Proceedings of the European Conference on Computer Vision, pages 579–595, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.804, + 0.468, + 0.83 + ], + "angle": 0, + "content": "[13] EU Parliament. Eu ai act: first regulation on artificial intelligence. Accessed June, 25:2023, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.832, + 0.468, + 0.9 + ], + "angle": 0, + "content": "[14] Kuan-Chuan Peng, Tsuhan Chen, Amir Sadovnik, and Andrew C Gallagher. A mixed bag of emotions: Model, predict, and transfer emotion distributions. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 860-868, 2015." + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.472, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.16 + ], + "angle": 0, + "content": "[15] Kuan-Chuan Peng, Amir Sadovnik, Andrew Gallagher, and Tsuhan Chen. Where do emotions come from? predicting the emotion stimuli map. In 2016 IEEE international conference on image processing (ICIP), pages 614-618. IEEE, 2016." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.164, + 0.892, + 0.232 + ], + "angle": 0, + "content": "[16] Dongyu She, Jufeng Yang, Ming-Ming Cheng, Yu-Kun Lai, Paul L Rosin, and Liang Wang. Wscnet: Weakly supervised coupled networks for visual sentiment classification and detection. IEEE Transactions on Multimedia, 22(5): 1358-1371, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.234, + 0.892, + 0.303 + ], + "angle": 0, + "content": "[17] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.305, + 0.892, + 0.373 + ], + "angle": 0, + "content": "[18] Jason Wei, Maarten Bosma, Vincent Zhao, Kelvin Guu, Adams Wei Yu, Brian Lester, Nan Du, Andrew M Dai, and Quoc V Le. Finetuned language models are zero-shot learners. In International Conference on Learning Representations." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.376, + 0.892, + 0.445 + ], + "angle": 0, + "content": "[19] Hongxia Xie, Hua Chung, Hong-Han Shuai, and Wen-Huang Cheng. Learning to prompt for vision-language emotion recognition. In 2023 11th International Conference on Affective Computing and Intelligent Interaction Workshops and Demos (ACIIW), pages 1-4, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.447, + 0.892, + 0.529 + ], + "angle": 0, + "content": "[20] Hongxia Xie, Ming-Xian Lee, Tzu-Jui Chen, Hung-Jen Chen, Hou-I Liu, Hong-Han Shuai, and Wen-Huang Cheng. Most important person-guided dual-branch cross-patch attention for group affect recognition. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 20598–20608, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.532, + 0.892, + 0.599 + ], + "angle": 0, + "content": "[21] Liwen Xu, Zhengtao Wang, Bin Wu, and Simon Lui. Mdan: Multi-level dependent attention network for visual emotion analysis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9479-9488, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.602, + 0.892, + 0.643 + ], + "angle": 0, + "content": "[22] Zhiyang Xu, Ying Shen, and Lifu Huang. Multiinstruct: Improving multi-modal zero-shot learning via instruction tuning. arXiv preprint arXiv:2212.10773, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.645, + 0.892, + 0.686 + ], + "angle": 0, + "content": "[23] Jingyuan Yang, Jie Li, Xiumei Wang, Yuxuan Ding, and Xinbo Gao. Stimuli-aware visual emotion analysis. IEEE Transactions on Image Processing, 30:7432-7445, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.688, + 0.892, + 0.757 + ], + "angle": 0, + "content": "[24] Jingyuan Yang, Qirui Huang, Tingting Ding, Dani Lischinski, Danny Cohen-Or, and Hui Huang. Emoset: A large-scale visual emotion dataset with rich attributes. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 20383-20394, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.759, + 0.892, + 0.814 + ], + "angle": 0, + "content": "[25] Quanzeng You, Jiebo Luo, Hailin Jin, and Jianchao Yang. Building a large scale dataset for image emotion recognition: The fine print and the benchmark. In Proceedings of the AAAI conference on artificial intelligence, 2016." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.817, + 0.892, + 0.871 + ], + "angle": 0, + "content": "[26] Shengyu Zhang, Linfeng Dong, Xiaoya Li, Sen Zhang, Xiaofei Sun, Shuhe Wang, Jiwei Li, Runyi Hu, Tianwei Zhang, Fei Wu, et al. Instruction tuning for large language models: A survey. arXiv preprint arXiv:2308.10792, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.873, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[27] Wei Zhang, Xuanyu He, and Weizhi Lu. Exploring discriminative representations for image emotion recognition with" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "26604" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.11, + 0.092, + 0.468, + 0.119 + ], + "angle": 0, + "content": "cnns. IEEE Transactions on Multimedia, 22(2):515-523, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.122, + 0.468, + 0.191 + ], + "angle": 0, + "content": "[28] Sicheng Zhao, Zizhou Jia, Hui Chen, Leida Li, Guiguang Ding, and Kurt Keutzer. Pd Janet: Polarity-consistent deep attention network for fine-grained visual emotion regression. In Proceedings of the 27th ACM International Conference on Multimedia, pages 192-201, 2019." + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.468, + 0.191 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.518, + 0.956 + ], + "angle": 0, + "content": "26605" + } + ] +] \ No newline at end of file diff --git a/2024/EmoVIT_ Revolutionizing Emotion Insights with Visual Instruction Tuning/337f29a8-74df-4960-89a4-6eff2797f1cc_origin.pdf b/2024/EmoVIT_ Revolutionizing Emotion Insights with Visual Instruction Tuning/337f29a8-74df-4960-89a4-6eff2797f1cc_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..3f3adb81a9bfaa182c77c6699a5b4f1d00f734bc --- /dev/null +++ b/2024/EmoVIT_ Revolutionizing Emotion Insights with Visual Instruction Tuning/337f29a8-74df-4960-89a4-6eff2797f1cc_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3ed53b0c56c1477faa1262e4f7b65a01cf0cedc96ecf6001dff49e74a4f8ab01 +size 6458292 diff --git a/2024/EmoVIT_ Revolutionizing Emotion Insights with Visual Instruction Tuning/full.md b/2024/EmoVIT_ Revolutionizing Emotion Insights with Visual Instruction Tuning/full.md new file mode 100644 index 0000000000000000000000000000000000000000..877c0dbaf5e3bf88778330ff908e309a56da4f07 --- /dev/null +++ b/2024/EmoVIT_ Revolutionizing Emotion Insights with Visual Instruction Tuning/full.md @@ -0,0 +1,343 @@ +# EmoVIT: Revolutionizing Emotion Insights with Visual Instruction Tuning + +Hongxia Xie1, Chu-Jun Peng2, Yu-Wen Tseng2, Hung-Jen Chen3, Chan-Feng Hsu3, Hong-Han Shuai3, Wen-Huang Cheng2 +1Jilin University +2National Taiwan University +3National Yang Ming Chiao Tung University +hongxiaie@jlu.edu.cn, {d12922035, ywt, wenhuang}@csie.ntu.edu.tw, {hjc.eed07g, cfhsu311510211.ee11, hhshuai}@nytu.edu.tw + +# Abstract + +Visual Instruction Tuning represents a novel learning paradigm involving the fine-tuning of pre-trained language models using task-specific instructions. This paradigm shows promising zero-shot results in various natural language processing tasks but is still unexplored in vision emotion understanding. In this work, we focus on enhancing the model's proficiency in understanding and adhering to instructions related to emotional contexts. Initially, we identify key visual clues critical to visual emotion recognition. Subsequently, we introduce a novel GPT-assisted pipeline for generating emotion visual instruction data, effectively addressing the scarcity of annotated instruction data in this domain. Expanding on the groundwork established by InstructBLIP, our proposed EmoVIT architecture incorporates emotion-specific instruction data, leveraging the powerful capabilities of Large Language Models to enhance performance. Through extensive experiments, our model showcases its proficiency in emotion classification, adeptness in affective reasoning, and competence in comprehending humor. The comparative analysis provides a robust benchmark for Emotion Visual Instruction Tuning in the era of LLMs, providing valuable insights and opening avenues for future exploration in this domain. Our code is available at https://github.com/aimmemotion/EmoVIT. + +# 1. Introduction + +Visual emotion recognition, a key area within artificial intelligence and computer vision, aims to predict human emotions based on visual cues such as facial expressions and body language. This technology is essential in bridging the gap between human affective states and machine understanding. Its diverse applications [19, 20, 23, 24], spanning + +![](images/74183f96b9146668203aeb0e0e4c75d5fc3d1e5936bcb0e72888514819342a07.jpg) +Figure 1. Illustration of the importance of instruction-following ability in visual emotion understanding. + +from improving human-computer interaction to aiding in mental health assessment, underscore its significance. Accurate emotion recognition is vital for enhancing user experience and ensuring information security, as it helps prevent emotional manipulation and misinformation [13]. Developing robust emotion recognition models is not only a technical challenge but also a step towards more empathetic and intuitive AI systems, paving the way for more efficient and natural human-computer interactions. + +The AI community has recently shown a growing interest in developing foundational vision models, e.g., Flamingo [1], LLaVA [8], BLIP2 [5]. These models excel in open-world visual understanding, tackling several vision tasks such as classification, detection, segmentation, and captioning. In contrast, current large-scale multimodal models are still in its infancy when it comes to emotion perception [7]. As illustrated in Fig. 1, when directly query the GPT-4 [11] about the emotional category of an image, the model tends to provide incorrect responses. However, the model delivers accurate responses when provided with revised instructions. To fully leverage the potential of ex + +isting vision-based large models, our approach is based on the concept of Instruction Tuning. This effective strategy is aimed at teaching language models to follow natural language instructions, a technique proven to enhance their generalization performance across unseen tasks [2, 8, 18]. + +In this work, we focus on developing the model's proficiency in understanding and following instructions related to emotional contexts. This approach highlights the importance of fine-tuning the model's instruction-following capabilities, enabling it to interpret and respond to emotional content effectively. This is achieved by leveraging its pre-existing knowledge base, thereby eliminating the necessity for an emotion-specific architectural framework. + +To address the notable challenges encountered in Instruction Tuning for visual emotion recognition, especially the lack of specific instruction data, we introduce a novel self-generation pipeline explicitly crafted for visual emotion recognition by using GPT-4 [11]. This innovative pipeline excels in generating a diverse array of (image, instruction, output) instances, thereby notably enhancing the dataset with a more extensive and task-oriented variety of examples. This approach not only overcomes the challenge of limited data availability but also reduces the dependence on human labor. Therefore, it streamlines the process, enabling more efficient and effective emotion recognition. + +Additionally, Instruction Tuning has been criticized for its emphasis on surface-level features like output patterns and styles, rather than achieving a profound comprehension and assimilation of tasks [26]. To tackle this issue and enhance the diversity and creativity of instruction data, our dataset includes instructions that demand complex reasoning, going beyond basic question-and-answer formats. This is further enriched by incorporating visual cues such as brightness, colorfulness, scene type, object class, facial expressions, and human actions. These aspects are pivotal in fostering a nuanced comprehension of visual emotions, thus allowing the model to generate more precise and contextually appropriate interpretations [24]. + +After generating the emotion visual instruction data, we propose an Emotion Visual Instruction Tuning (EmoVIT) framework, leveraging the foundation of InstructBLIP [2]. This framework incorporates an emotion-centric, instruction-aware module that proficiently guides Large Language Models (LLMs) in assimilating the nuances of emotion instructions. Our work signifies a paradigm shift, presenting a new era of instruction-based learning for visual emotion understanding that relies less on explicit training data. Remarkably, as shown in Fig. 2, our approach requires almost $50\%$ of the training data typically needed yet exceeds the performance of previous visual emotion recognition methods and popular Visual Instruction Tuning methods. + +Our contributions can be summarized as follows: + +![](images/ca388d39393f94f7229363cdad32b2a024c9d1f0313d6164e0808e710feac5e7.jpg) +Figure 2. Performance comparison on EmoSet test set [24] (Accuracy $\%$ + +- We explore the potential of the Visual Instruction Tuning paradigm for emotion comprehension and introduce the concept of Emotion Visual Instruction Tuning. +- After thoroughly considering the unique characteristics of visual emotion recognition, we develop a novel GPT-assisted pipeline for generating emotion visual instruction data. This approach effectively bridges the gap in available annotated instruction data within this specific domain. +- Building upon the foundation of InstructBLIP, our EmoVIT architecture integrates emotion domain-specific instruction data, harnessing the robust capabilities of LLMs to boost performance. The extensive experiments demonstrate our model's proficiency in emotion classification, affective reasoning, and comprehension of humour. + +# 2. Related Work + +# 2.1. Visual Emotion Recognition + +A key challenge in visual emotion recognition is bridging the gap between an image's visual cues and the emotions it portrays [12, 21, 25]. While traditional efforts, e.g., Xu et al.'s multi-level dependent attention network [21], focus on visual models for emotional feature learning, recent advancements like EmoSet [24] offer rich emotion-laden datasets with 3.3 million images. The rise of multimodal models, such as the GPT series [11], has further propelled Vision-Language Recognition. However, fully leveraging these models in emotion recognition is an area ripe for exploration. Our work leads the way in utilizing large-scale models for Emotion Visual Instruction Tuning. + +![](images/9365186dfb4f66d5b67d49c60155a0e09792131e67ecb988d89722371788fb2f.jpg) +Figure 3. The comparison of different visual tuning paradigms. + +# 2.2. Visual Instruction Tuning + +Current Large Language Models (LLMs) have extensive knowledge bases, but their effectiveness depends on accurately interpreting human instructions due to a mismatch between training goals and user expectations. LLMs are trained to minimize prediction errors, whereas users expect helpful and safe instruction-following. Instruction Tuning addresses this by teaching models to follow natural language instructions, enhancing generalization to new tasks. FLAN [18] demonstrated that training a large model on instruction-based datasets improves zero-shot performance. This approach has extended to vision-language tasks, with BLIP2 [5] and LLaVA [8] adapting instruction-tuned LLMs for visual inputs. InstructBLIP [2] introduces instruction-aware visual feature extraction and the Q-Former, enabling more flexible, instruction-driven feature extraction. + +As a novel area, visual emotion instruction tuning lacks benchmarks or guidelines for creating emotion instruction data. Our work pioneers the use of large-scale models to develop an emotion instruction data pipeline, overcoming the limitations of manual annotation. + +# 3. Method + +# 3.1. Preliminary of Visual Instruction Tuning + +In the deep learning era, visual tuning has experienced significant paradigm shifts, as depicted in Fig. 3. + +In Fig. 3(a), conventional tuning methodologies encompass Full fine-tuning, Head-oriented, and Backbone-oriented techniques, capitalizing on large-scale pre-trained models. Predominantly, thoroughly fine-tuning these models for specific tasks, conducted end-to-end, is recognized as a highly effective strategy. However, this method requires maintaining separate copies of the backbone parameters for each distinct task, posing challenges in storage and deploy- + +ment. + +Alternatively, Visual Prompt Tuning (VPT) [3], presents an efficient substitute for full fine-tuning within large-scale vision Transformer models. It achieves this by employing a minimal fraction of trainable parameters in the input space while maintaining a frozen backbone model. The objective function for Visual Prompt Tuning is given by: + +$$ +\min _ {\theta_ {\mathrm {P}}} \mathcal {L} (f (X, P; \theta_ {\mathrm {P}}), Y) \tag {1} +$$ + +where $\min_{\theta_{\mathrm{P}}}$ is the minimization over the prompt parameters $P$ , $\mathcal{L}$ is the loss function, $f$ represents the model function with input image $X$ , prompt parameters $P$ , and learnable model parameters $\theta_{\mathrm{P}}$ as input, and $Y$ is the target output. + +Visual Prompt Tuning focuses on optimizing LLMs using a small set of parameters, whereas Visual Instruction Tuning (VIT) aims to improve the model's comprehension of instructions, thereby addressing the model's shortcomings in specific domains. This type of method aims to enhance the model's proficiency in following instructions, leveraging the capabilities of the latest foundation models, e.g., Llama [17], and BLIP2 [5]. Instructions serve as guiding constraints, shaping the model's outputs to conform to specific response characteristics and domain-relevant knowledge. This approach enables human monitoring of the model's behavior, thereby assuring alignment with the desired outcomes. Moreover, Instruction Tuning is computationally efficient, allowing LLMs to swiftly adapt to particular domains without extensive retraining or architectural alterations. + +The objective function for Visual Instruction Tuning is given by: + +$$ +\min _ {\theta_ {\text {t u n a b l e}}} \mathcal {L} (g (X, I, C; \theta_ {\text {t u n a b l e}}), Y) \tag {2} +$$ + +![](images/9342f2ab39a50f8859b88420ccd7949deef2992314bd34406459d0f268e2bd0e.jpg) +Emotion class +Contentment +Brightness +0.8 +Colorfulness +0.8 +Scene type + +![](images/10f7bfb5bdea883c73faa2db931822742351b94eebfa64006c7229a06b6ca39d.jpg) +Beach +Object class +Sky +Facial +Contentment +Human +Laughing. +Ludwig +(a) Emotion Visual Instruction Data Generation +Figure 4. The overall architecture of our proposed method. The Emotion Instruction data generated by (a) will be used for Emotion Visual Instruction Tuning in (b). During Emotion Visual Instruction Tuning, given an input image, the frozen Image Encoder initiates the process by extracting visual features. Emotion Instruction generated by (a) are subsequently interacting with Queries embedding through the learnable Q-Former. This interaction is key to drawing out image features that are relevant to the task at hand. As a result, the frozen LLM receives visual information conducive to instruction following. + +![](images/831fecbf7d7448ec656e1d747dd2c4ffab7b3c61068499a8d6d2a04fdcc61957.jpg) +(b) Emotion Visual Instruction Tuning Architecture + +![](images/ea875063c75ecf06e8fc35841680bf9a653c54e0f47e94ec3f979194722849c3.jpg) +(c) The Details of Q-Former Module + +where $\min_{\theta_{\mathrm{tunable}}}$ denotes the minimization over the tunable parameters $\theta_{\mathrm{tunable}}$ in the Instruction Tuning Module, $\mathcal{L}$ is the loss function, $g$ is the model function with instruction $I$ , image $X$ , other contexts $C$ , and tunable parameters $\theta_{\mathrm{tunable}}$ and $Y$ denotes the target output. The optional context $C$ is not just raw data; it encompasses descriptive or directive information guiding the model on how to process input or which task to execute, e.g., image caption. It's integral to the model's understanding and execution of tasks based on specific instructions or guidelines. + +# 3.2. GPT-assisted Emotion Visual Instruction Data Generation + +Previous methodologies commonly employed a consistent template-based set of instructions for every image within a dataset across various specific tasks [2]. For instance, a standard instruction such as "Briefly describe the content of the image" was employed uniformly across all images for Image Captioning. In this way, the model may not be able to adequately capture the unique characteristics of each image. Moreover, this one-size-fits-all approach often leads to suboptimal performance in emotion recognition tasks that require nuanced perception and differentiation of ambiguous emotion classes. + +Since the topic of Emotion Visual Instruction Tuning is still in its infancy, no benchmarks or guidelines have been proposed so far for constructing emotion instruction data. Based on the recent successes of machine-generated instructions demonstrated in LLaVA [8], our work pioneers the use of existing LLMs to create a pipeline for + +self-generating emotion instructions. Different from previous template-based and one-size-fits-all instruction data, we propose an instance-wise and LLM-assisted visual emotion instruction data pipeline. This methodology transcends the constraints of manual annotation by employing GPT-4 [11] to generate instance-wise, tailored instruction data that dynamically corresponds to visual content. + +Prior to the development of instructional data for the visual emotion recognition task, it is imperative to confront a fundamental academic problem: What types of visual clues are pivotal in identifying emotions? This necessitates a careful consideration of the unique characteristics inherent to the task, along with a comprehensive understanding of the potential visual cues associated with human emotions. In this work, we propose a novel visual instruction data mechanism to remove the inherent subjectivity and ambiguity in emotional interpretation. Specifically, we integrate a broad spectrum of emotion attributes across multiple levels: low-level attributes (e.g., brightness, colorfulness), mid-level attributes (e.g., scene type and object class), and high-level attributes (e.g., facial expressions and human actions), building upon insights from previous work [24]. This comprehensive strategy not only aligns with the intricate nature of emotions but also significantly enhances the model's capability to interpret and understand visual emotional cues more accurately and holistically. + +The overall pipeline of our proposed emotion visual instruction data is shown in Fig. 4 (a). For an image $X_{\mathrm{img}}$ , three types of image-related contexts are essential for GPT-4 to generate emotion instruction data: (i) a caption $X_{c}$ , (ii) + +an emotion attribute list $X_{\mathrm{attr}}$ , which includes emotion class, brightness, colorfulness, scene type, object class, facial expression, and human action, and (iii) the system prompt, designed to enable GPT-4 to comprehend the specific task requirement1. + +We first manually design a few examples which are used as seed examples for in-context learning to query GPT-4. This operation leverages the model's ability to extrapolate from given examples, enhancing its understanding and response accuracy based on the principles of few-shot learning [8]. Our generated emotion instruction data includes three types: Categorical, Conversation, and Reasoning. Building upon previous research [8], our generated instruction data adheres to the dialogue format, exemplified in Fig. 5. + +Our strategy for generating emotion instruction data adopts a progressive approach from simple to complex. Initially, for the Categorical data, we transform the associated emotion class of the image into a structured format. This process serves as the foundational component of our emotion instruction data. + +For the Conversation data, our framework is designed to create dialogues in which the GPT assistant interacts with an inquirer, focusing on the emotion attributes of the image. In this setup, the assistant's responses are tailored to interpret and describe the image as though it were within its own visual field, thereby providing insights from an observational viewpoint. The scope of questions posed is comprehensive, encompassing the types of objects depicted, their actions, and the dynamics of their interrelationships. The dialogues we generate fall into two categories: (i) Basic Interaction, focusing on the provided emotion attribute list with simple, direct characteristics, and (ii) Advanced Interaction, which builds on the first type to reach greater conversational complexity and sophistication. + +For the Reasoning data, our approach extends beyond mere visual content, prompting the model to generate in-depth reasoning questions. To enhance the dialogue's credibility and structure, detailed examples are incorporated alongside logical reasoning steps, ensuring that the discourse convincingly captures the intricacies of the visual content. + +# 3.3. Emotion Visual Instruction Tuning + +After acquiring the emotion visual instruction data as detailed in Sec. 3.2, our goal is to employ this data in enhancing the existing Visual Instruction Tuning model. This enhancement aims to align the LLMs' existing knowledge with the emotion understanding domain. + +As shown in Fig. 4 b, we have developed an Emotion Visual Instruction Tuning (EmoVIT) architecture based on + +# Our Emotion Instruction Data Sample + +![](images/2f4119e5872a968c563b86bb5c30d6c9e26c9413d6245d39614774b22b57e4bd.jpg) +Figure 5. The sample of our generated visual emotion instruction data. + +Context type 1: Caption: + +Two young women dancing in a kitchen. + +Context type 2: Attribute: + +Emotion: contentment Human action: stretching arm + +Brightness: 0.7 Object: ['Cupboard', 'Food', 'Countertop'] + +Colorfulness: 0.4 Facial expression: happy + +# Categorical + +Question: Please select the emotion closest to the image from the following options: amusement, anger, awe, contentment, disgust, excitement, fear and sadness. + +Answer: Predict emotion: contentment. + +# Conversation + +# Basic Interaction + +Question: Can you describe the activity happening in the image? + +Answer: The image depicts two young women enjoying themselves by dancing in a kitchen. One of the women is seen stretching her arm, most likely as part of their dance routine. + +# Conversation + +# Advanced Interaction + +Question: Based on the given emotion of contentment and the activities in the image, can you guess about the possible context or story of this scenario? + +Answer: These women are good friends or roommates, spending quality time together. The act of dancing, especially in an unconventional place like a kitchen, suggests they are comfortable with each other and are enjoying a break from routine. + +# Reasoning + +Question: Based on the setting and activities of the two young women, what might have prompted this impromptu dance session in the middle of a kitchen? Answer: The impromptu dance could have been prompted by a favorite song playing, or simple spontaneous fun amidst preparing a meal or organizing the kitchen. + +InstructBLIP [2]. This architecture specifically leverages its Instruction-aware Q-Former Module, as depicted in Fig. 4 c, for emotion-centric instructional tasks. + +Specifically, the Instruction-aware Q-Former Module takes in the emotion instruction tokens, queries, and image embeddings as input. The image embeddings are extracted by a frozen image encoder. The learnable queries are initially produced by the pre-trained Q-Former of InstructBLIP. During training, the Instruction-aware module enhances task-specific feature extraction. It does this by integrating emotion instruction and query embeddings within self-attention layers, aligning visual information with the LLM's instruction-following requirements. Our approach adopts cross-entropy loss, tailoring it to the intricacies of visual emotion recognition tasks, thus ensuring precise and contextually relevant model training outcomes. + +We note that the data generated by our approach is not confined to a single model but can also be applied to other Visual Instruction Tuning models, such as LLaVA [17]. Notably, when LLaVA is fine-tuned with our data, it exhibits a significant enhancement in emotion recognition capabilities, as detailed in Sec. 4.2. In this way, we demonstrate not only the effectiveness but also the transferability of our generated data, showing its broad applicability and impact. + +# 4. Experimental Results + +# 4.1. Implemental Details + +Our implementation is based on the LAVIS library [4]. Our EmoVIT starts with a pre-trained InstructBLIP baseline and proceeds to fine-tune exclusively the Q-Former module, whilst keeping both the image encoder and the language model frozen. The parameters for our training adhere to the default settings established by InstructBLIP. + +Datasets. We evaluate our framework on ten benchmark datasets annotated under different scenarios and class number, namely EmoSet [24], WEBEmo [12], Emotion6 [14], the Flickr and Instagram (FI) [25], Artphoto [9], IAPS [10], Abstract [9], EmotionROI [15], UnbiasedEmo [12], and OxfordTVG-HIC [6]. + +Held-in Pretraining. Following previous work [2], we divide our dataset into two categories: held-in for pretraining and held-out for evaluation2. Considering the EmoSSet dataset's comprehensive inclusion of emotion attributes for each image, it has been chosen as the primary resource for our held-in pretraining phase. Simultaneously, for a broader assessment, we perform held-out evaluations using the test sets from various other datasets. + +For the generation of emotion visual instruction data, we initially employ the BLIP2 model for image captioning, followed by leveraging the GPT-4 API to generate emotion instruction data. In total, our collection comprises Categorical, Conversation, and Reasoning instruction data, derived from 51,200 unique images. This represents less than $50\%$ of the entire EmoSet. + +# 4.2. Held-out Evaluation + +As shown in Tab. 1, our proposed methodology exhibits a marked superiority in performance relative to the burgeoning Visual Instruction Tuning Methods. Since they have been pre-trained on dozens of large-scale datasets, it is evident that our generated emotion visual instruction data is particularly effective for emotional understanding. Our results signify a paradigm shift, heralding a new era of model training that relies less on explicit supervision and more on the robustness of emotion instruction-driven learning. + +The Effectiveness of Our Proposed Emotion Visual Instruction Data. As the first to introduce the concept of emotion visual instruction data, our study seeks to evaluate the generalizability of this newly generated instruction data. Our goal is to test its efficacy not only with InstructBLIP but also across other Visual Instruction Tuning model, to understand its broader applicability. As depicted in Fig. 6, we employ two Visual Instruction Tuning models, LLaVA and InstructBLIP, which were fine-tuned on our specially gen + +![](images/ef9f5da6de3e6f86a45bf5ddf42b29074a440d3f8afefe5d124bd6be0d190963.jpg) +Accuracy Improvements of FT Models Over Base Models +Figure 6. The improvement of our proposed emotion visual instruction tuning data tuning on LLaVA [8] and InstructBLIP [2]. + +erated emotion visual instruction data. Subsequent testing across five distinct datasets reveals notable improvements in both models, substantiating the efficacy of our generated data. Notably, InstructBLIP demonstrated a more substantial overall enhancement compared to LLaVA. This can be attributed to InstructBLIP's specialized Instruction-aware Q-Former Module, which adeptly extracts the salient features of our emotion instructions and synergizes them effectively with the corresponding images, thereby yielding improved performance. + +# 4.3. Effectiveness of Different Instruction Data + +# 4.3.1 Ablation Study of Different Instruction Data + +The ablation study outlined in Tab. 2 provides a comprehensive analysis of the impact that different instructional data types have on model performance, specifically concerning accuracy metrics on the EmoSet test set. Initially, the model, referred to as InstructBLIP [2], operates without the integration of the three types of instructional data and attains a baseline accuracy of $42.20\%$ . This foundational performance is significantly enhanced with the inclusion of Categorical data, which alone contributes to a substantial increase in accuracy. The introduction of Conversation data further amplifies this effect, underscoring the value of conversational context in improving the model's predictive capabilities. The addition of Reasoning data notably boosts performance, achieving a peak accuracy of $83.36\%$ . This indicates that the model significantly benefits from the nuanced cues in reasoning, aiding in understanding complex emotional instructions. The gradual improvements with each data type support the idea that a diverse approach to instructional data markedly enhances model comprehension and performance. + +
MethodWebEmoFIEmotion6AbstractArtPhotoIAPSaEmotionROIEmoSet
Number of Classes258688868
Flanmingo [1]9.3614.9121.673.5717.510.1321.7229.59
LLaVA [8]12.5556.0449.4419.5436.2542.4346.4644.03
BLIP2 [5]20.1057.7250.0028.5736.2539.2450.5146.79
InstructBLIP [2]12.8037.9746.1121.4226.2534.1846.1342.20
Ours*21.1268.0957.8132.3444.9044.1353.8783.36
+ +Table 1. Held-out performance comparison on visual emotion datasets $(\%)$ + +
CategoricalConversationReasoningAccuracy (%)
---42.20
✓--80.90 (+38.70)
✓✓-81.95 (+39.75)
✓✓✓83.36 (+41.16)
+ +Table 2. Ablation study of three types of instruction data. Accuracy (%) on EmoSet test set. + +# 4.3.2 Instruction Sensitivity + +This work is dedicated to the creation of a varied corpus of visual emotion instruction data, alongside the development of a robust instruction-based model. Our objective is for the model to demonstrate stability, producing consistent results in the face of minor variations in instruction phrasing, provided the core objective of the task persists unchanged. To this end, we employ the Sensitivity evaluation metric, as introduced by [22], to assess the model's fidelity in generating uniform outcomes irrespective of instructional nuances. + +We employ two semantically similar instructions as input prompts for the model, testing their impact on the Sensitivity score across three visual emotion datasets for different Visual Instruction Tuning models. The first instruction is: "From the given options: cls_1, cls_2, cls_3, etc., identify the emotion that most accurately reflects the image. Ensure your selection is confined to the listed options. Respond in the format: Predicted emotion." The second one states: "Please choose the emotion that best corresponds to the image from the following options: cls_1, cls_2, cls_3, etc. (Do not provide answers beyond the provided candidates.) Please reply in the following format: Predict emotion." + +As illustrated in Fig. 7, our approach, along with BLIP2, exhibited exceptionally low Sensitivity values, demonstrating robustness in understanding the instructions. Conversely, Flamingo and InstructBLIP displayed a higher degree of sensitivity, indicating a relative susceptibility to variations in instruction wording. + +# 4.4. Robustness + +Given that current emotion recognition datasets often exhibit category imbalances and labeling biases, our aim is + +![](images/9310629969a8e41f19348f83511eba7915ed94c2d9fea133652a4458ae950dc1.jpg) +Figure 7. The sensitivity score comparison (the lower the better). + +to evaluate the generalization ability of various learning strategies more impartially. Hence, we selected the Un-BiasedEmo test set [12], which is uniquely suited for recognizing intricate emotions, such as those associated with identical objects or scenes, e.g., landscapes, crowds, families, babies, and animals, where the emotional undertones can be particularly subtle and complex. + +As depicted in Tab. 3, our proposed methodology demonstrates superior performance when benchmarked against conventional supervised emotion recognition techniques, thereby underscoring the efficacy of our approach in more accurately discerning complex emotional contexts. + +
MethodAccuracy (%)
Direct Learning [12]71.64
Self-Directed Learning [12]72.45
Joint Learning [12]71.64
Curriculum Learning [12]74.27
Ours*74.72
+ +Table 3. Performance comparison on UnbiasedEmo dataset. + +![](images/109bb11314538a214cd82aed3b6d0b5421a081e885cde55b680079b58077f669.jpg) + +![](images/053e3769043cbb64654c8c3e03aab6d5a184737816f4645df6871152ef833fd0.jpg) + +BECAUSE + +![](images/b0abd87cf84e26d5614dc0f16917a8986399d1404f505f75f291a50d6a3b49a2.jpg) +Predicted emotion: anger +Predicted emotion: sadness + +BECAUSE + +Reason: The image depicts a group of people with their arms up in the air, which is a common gesture associated with anger. + +Additionally, the smoke billowing from the building in the background adds to the sense of chaos and unrest. + +Reason: The image is a black and white photograph of an old, abandoned house in the middle of a field. + +The scene is desolate and eerie, which suggests a sense of sadness or melancholy. + +# 4.4.1 Affective Reasoning + +In the domain of visual emotion recognition, where ambiguity and subjectivity are pervasive, the advent of an interpretable model is of considerable value. Such a model elucidates its cognitive processes, enhancing its trustworthiness and practicality in scenarios requiring a delicate grasp of emotional subtleties. + +Leveraging Visual Instruction Tuning, our model transcends mere categorization of emotions; it articulates the underlying rationale for its classifications. The executing commands for identifying emotions and elucidating the decision basis is illustrated below: + +Predicted emotion: [emotion]. + +Reason: [explanation]. + +Our model delineates the visual features influencing its determinations, thereby addressing the complexities inherent in discerning and explaining emotion-related nuances. + +The explanations provide us with visual clues contained within the images, as exemplified in Fig. 8. It provides interpretable visual indicators that inform the model's outputs, as demonstrated in our example, by disambiguating the often abstract emotional categories. + +# 4.5. Scaling Law + +Pretraining data. As demonstrated in Tab. 4, there is a clear correlation between the size of the pre-training dataset and improved performance. Consequently, we anticipate that an increase in training data in the future could enhance the effectiveness of Emotion Visual Instruction Tuning. + +# 4.6. Humour Caption Generation + +The comprehension of humor is intricately linked to the understanding of emotions. Leveraging our generative language model, we conduct a caption generation task without + +
5%10%30%50%
79.0081.0079.3483.36
+ +Table 4. Ablation study of different portion of pre-training data. Accuracy $(\%)$ on EmoSet test set. + +![](images/eb9cb349f1ef4b50b1a208cf1b5bd1b0aa7d233078c483d280120546ab7f8105.jpg) +Figure 8. The sample of our generated explanation. +Figure 9. The sample of our generated humour caption vs human writing humour caption from OxfordTVG-HIC. + +modifying the model's architecture, specifically testing the model's proficiency in generating humorous captions. For this purpose, we select 50 images from the OxfordTVGHIC dataset [6] and generate corresponding captions using our model. Subsequently, the captions produced by our model are compared with manually annotated captions from the dataset in a user study. Thirty participants were asked to vote on which captions were more humorous. Our model-generated captions receive $60\%$ of the votes, demonstrating its effective humor generation capabilities. One sample is visualized in Fig. 9. + +# 5. Conclusion + +In our study, drawing upon the distinctive visual cues key to visual emotion recognition, we present a GPT-assisted pipeline specifically designed for generating emotion visual instruction data. The developed EmoVIT model incorporates emotion-specific instructions, leveraging LLMs for enhanced performance. Our comprehensive experiments validate its effectiveness in emotion classification, affective reasoning, and humor understanding. This comparative analysis sets a benchmark for Emotion Visual Instruction Tuning with LLMs, providing valuable insights and directions for future research in this field. + +# Acknowledgement + +This work is partially supported by the National Science and Technology Council, Taiwan under Grant, NSTC-112-2628-E-002-033-MY4, NSTC-112-2634-F-002-002-MBK, NSTC-112-26234-F-002-005, NSTC-112-2221-E-A49-059-MY3, and NSTC-112-2221-E-A49-094-MY3. + +# References + +[1] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. In Advances in Neural Information Processing Systems, pages 23716-23736, 2022. +[2] Wenliang Dai, Junnan Li, Dongxu Li, Anthony Meng Huat Tiong, Junqi Zhao, Weisheng Wang, Boyang Li, Pascale Fung, and Steven Hoi. Instructclip: Towards general-purpose vision-language models with instruction tuning. In Advances in Neural Information Processing Systems, 2023. +[3] Menglin Jia, Luming Tang, Bor-Chun Chen, Claire Cardie, Serge Belongie, Bharath Hariharan, and Ser-Nam Lim. Visual prompt tuning. In Proceedings of the European Conference on Computer Vision, pages 709-727. Springer, 2022. +[4] Dongxu Li, Junnan Li, Hung Le, Guangsen Wang, Silvio Savarese, and Steven CH Hoi. Lavis: A library for language-vision intelligence. arXiv preprint arXiv:2209.09019, 2022. +[5] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In Proceedings of the International Conference on Machine Learning, 2023. +[6] Runjia Li, Shuyang Sun, Mohamed Elhoseiny, and Philip Torr. Oxfordtvg-hic: Can machine make humorous captions from images? In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 20293-20303, 2023. +[7] Zheng Lian, Licai Sun, Mingyu Xu, Haiyang Sun, Ke Xu, Zhuofan Wen, Shun Chen, Bin Liu, and Jianhua Tao. Explainable multimodal emotion reasoning. arXiv preprint arXiv:2306.15401, 2023. +[8] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. In Advances in Neural Information Processing Systems, 2023. +[9] Jana Machajdik and Allan Hanbury. Affective image classification using features inspired by psychology and art theory. In Proceedings of the 18th ACM international conference on Multimedia, pages 83-92, 2010. +[10] Joseph A Mikels, Barbara L Fredrickson, Gregory R Larkin, Casey M Lindberg, Sam J Maglio, and Patricia A Reuter-Lorenz. Emotional category data on images from the international affective picture system. Behavior research methods, 37:626-630, 2005. +[11] OpenAI. Gpt-4 technical report. Technical report, 2023. +[12] Rameswar Panda, Jianming Zhang, Haoxiang Li, Joon-Young Lee, Xin Lu, and Amit K Roy-Chowdhury. Contemplating visual emotions: Understanding and overcoming dataset bias. In Proceedings of the European Conference on Computer Vision, pages 579–595, 2018. +[13] EU Parliament. Eu ai act: first regulation on artificial intelligence. Accessed June, 25:2023, 2023. +[14] Kuan-Chuan Peng, Tsuhan Chen, Amir Sadovnik, and Andrew C Gallagher. A mixed bag of emotions: Model, predict, and transfer emotion distributions. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 860-868, 2015. + +[15] Kuan-Chuan Peng, Amir Sadovnik, Andrew Gallagher, and Tsuhan Chen. Where do emotions come from? predicting the emotion stimuli map. In 2016 IEEE international conference on image processing (ICIP), pages 614-618. IEEE, 2016. +[16] Dongyu She, Jufeng Yang, Ming-Ming Cheng, Yu-Kun Lai, Paul L Rosin, and Liang Wang. Wscnet: Weakly supervised coupled networks for visual sentiment classification and detection. IEEE Transactions on Multimedia, 22(5): 1358-1371, 2019. +[17] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. +[18] Jason Wei, Maarten Bosma, Vincent Zhao, Kelvin Guu, Adams Wei Yu, Brian Lester, Nan Du, Andrew M Dai, and Quoc V Le. Finetuned language models are zero-shot learners. In International Conference on Learning Representations. +[19] Hongxia Xie, Hua Chung, Hong-Han Shuai, and Wen-Huang Cheng. Learning to prompt for vision-language emotion recognition. In 2023 11th International Conference on Affective Computing and Intelligent Interaction Workshops and Demos (ACIIW), pages 1-4, 2023. +[20] Hongxia Xie, Ming-Xian Lee, Tzu-Jui Chen, Hung-Jen Chen, Hou-I Liu, Hong-Han Shuai, and Wen-Huang Cheng. Most important person-guided dual-branch cross-patch attention for group affect recognition. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 20598–20608, 2023. +[21] Liwen Xu, Zhengtao Wang, Bin Wu, and Simon Lui. Mdan: Multi-level dependent attention network for visual emotion analysis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9479-9488, 2022. +[22] Zhiyang Xu, Ying Shen, and Lifu Huang. Multiinstruct: Improving multi-modal zero-shot learning via instruction tuning. arXiv preprint arXiv:2212.10773, 2022. +[23] Jingyuan Yang, Jie Li, Xiumei Wang, Yuxuan Ding, and Xinbo Gao. Stimuli-aware visual emotion analysis. IEEE Transactions on Image Processing, 30:7432-7445, 2021. +[24] Jingyuan Yang, Qirui Huang, Tingting Ding, Dani Lischinski, Danny Cohen-Or, and Hui Huang. Emoset: A large-scale visual emotion dataset with rich attributes. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 20383-20394, 2023. +[25] Quanzeng You, Jiebo Luo, Hailin Jin, and Jianchao Yang. Building a large scale dataset for image emotion recognition: The fine print and the benchmark. In Proceedings of the AAAI conference on artificial intelligence, 2016. +[26] Shengyu Zhang, Linfeng Dong, Xiaoya Li, Sen Zhang, Xiaofei Sun, Shuhe Wang, Jiwei Li, Runyi Hu, Tianwei Zhang, Fei Wu, et al. Instruction tuning for large language models: A survey. arXiv preprint arXiv:2308.10792, 2023. +[27] Wei Zhang, Xuanyu He, and Weizhi Lu. Exploring discriminative representations for image emotion recognition with + +cnns. IEEE Transactions on Multimedia, 22(2):515-523, 2019. +[28] Sicheng Zhao, Zizhou Jia, Hui Chen, Leida Li, Guiguang Ding, and Kurt Keutzer. Pd Janet: Polarity-consistent deep attention network for fine-grained visual emotion regression. In Proceedings of the 27th ACM International Conference on Multimedia, pages 192-201, 2019. \ No newline at end of file diff --git a/2024/EmoVIT_ Revolutionizing Emotion Insights with Visual Instruction Tuning/images.zip b/2024/EmoVIT_ Revolutionizing Emotion Insights with Visual Instruction Tuning/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..513aff481b3ee0fd784cdf152d2fc5f06ae8522a --- /dev/null +++ b/2024/EmoVIT_ Revolutionizing Emotion Insights with Visual Instruction Tuning/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:585d9618a6db4c3a9f2f7ee460f8975ed887d52fcb9d2ce05492b5febf997d74 +size 466437 diff --git a/2024/EmoVIT_ Revolutionizing Emotion Insights with Visual Instruction Tuning/layout.json b/2024/EmoVIT_ Revolutionizing Emotion Insights with Visual Instruction Tuning/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..c63f09e8768f3c2e81c42a89f5d8528146637dab --- /dev/null +++ b/2024/EmoVIT_ Revolutionizing Emotion Insights with Visual Instruction Tuning/layout.json @@ -0,0 +1,7917 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 66, + 103, + 529, + 122 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 103, + 529, + 122 + ], + "spans": [ + { + "bbox": [ + 66, + 103, + 529, + 122 + ], + "type": "text", + "content": "EmoVIT: Revolutionizing Emotion Insights with Visual Instruction Tuning" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 115, + 142, + 477, + 242 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 115, + 142, + 477, + 242 + ], + "spans": [ + { + "bbox": [ + 115, + 142, + 477, + 242 + ], + "type": "text", + "content": "Hongxia Xie1, Chu-Jun Peng2, Yu-Wen Tseng2, Hung-Jen Chen3, Chan-Feng Hsu3, Hong-Han Shuai3, Wen-Huang Cheng2 \n1Jilin University \n2National Taiwan University \n3National Yang Ming Chiao Tung University \nhongxiaie@jlu.edu.cn, {d12922035, ywt, wenhuang}@csie.ntu.edu.tw, {hjc.eed07g, cfhsu311510211.ee11, hhshuai}@nytu.edu.tw" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 143, + 269, + 192, + 281 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 269, + 192, + 281 + ], + "spans": [ + { + "bbox": [ + 143, + 269, + 192, + 281 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 295, + 290, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 295, + 290, + 582 + ], + "spans": [ + { + "bbox": [ + 46, + 295, + 290, + 582 + ], + "type": "text", + "content": "Visual Instruction Tuning represents a novel learning paradigm involving the fine-tuning of pre-trained language models using task-specific instructions. This paradigm shows promising zero-shot results in various natural language processing tasks but is still unexplored in vision emotion understanding. In this work, we focus on enhancing the model's proficiency in understanding and adhering to instructions related to emotional contexts. Initially, we identify key visual clues critical to visual emotion recognition. Subsequently, we introduce a novel GPT-assisted pipeline for generating emotion visual instruction data, effectively addressing the scarcity of annotated instruction data in this domain. Expanding on the groundwork established by InstructBLIP, our proposed EmoVIT architecture incorporates emotion-specific instruction data, leveraging the powerful capabilities of Large Language Models to enhance performance. Through extensive experiments, our model showcases its proficiency in emotion classification, adeptness in affective reasoning, and competence in comprehending humor. The comparative analysis provides a robust benchmark for Emotion Visual Instruction Tuning in the era of LLMs, providing valuable insights and opening avenues for future exploration in this domain. Our code is available at https://github.com/aimmemotion/EmoVIT." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 620, + 128, + 632 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 620, + 128, + 632 + ], + "spans": [ + { + "bbox": [ + 47, + 620, + 128, + 632 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 641, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 641, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 641, + 287, + 715 + ], + "type": "text", + "content": "Visual emotion recognition, a key area within artificial intelligence and computer vision, aims to predict human emotions based on visual cues such as facial expressions and body language. This technology is essential in bridging the gap between human affective states and machine understanding. Its diverse applications [19, 20, 23, 24], spanning" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 307, + 268, + 557, + 419 + ], + "blocks": [ + { + "bbox": [ + 307, + 268, + 557, + 419 + ], + "lines": [ + { + "bbox": [ + 307, + 268, + 557, + 419 + ], + "spans": [ + { + "bbox": [ + 307, + 268, + 557, + 419 + ], + "type": "image", + "image_path": "74183f96b9146668203aeb0e0e4c75d5fc3d1e5936bcb0e72888514819342a07.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 427, + 556, + 450 + ], + "lines": [ + { + "bbox": [ + 305, + 427, + 556, + 450 + ], + "spans": [ + { + "bbox": [ + 305, + 427, + 556, + 450 + ], + "type": "text", + "content": "Figure 1. Illustration of the importance of instruction-following ability in visual emotion understanding." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 461, + 547, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 461, + 547, + 569 + ], + "spans": [ + { + "bbox": [ + 304, + 461, + 547, + 569 + ], + "type": "text", + "content": "from improving human-computer interaction to aiding in mental health assessment, underscore its significance. Accurate emotion recognition is vital for enhancing user experience and ensuring information security, as it helps prevent emotional manipulation and misinformation [13]. Developing robust emotion recognition models is not only a technical challenge but also a step towards more empathetic and intuitive AI systems, paving the way for more efficient and natural human-computer interactions." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 570, + 548, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 548, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 548, + 714 + ], + "type": "text", + "content": "The AI community has recently shown a growing interest in developing foundational vision models, e.g., Flamingo [1], LLaVA [8], BLIP2 [5]. These models excel in open-world visual understanding, tackling several vision tasks such as classification, detection, segmentation, and captioning. In contrast, current large-scale multimodal models are still in its infancy when it comes to emotion perception [7]. As illustrated in Fig. 1, when directly query the GPT-4 [11] about the emotional category of an image, the model tends to provide incorrect responses. However, the model delivers accurate responses when provided with revised instructions. To fully leverage the potential of ex" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "26596" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 133 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 133 + ], + "type": "text", + "content": "isting vision-based large models, our approach is based on the concept of Instruction Tuning. This effective strategy is aimed at teaching language models to follow natural language instructions, a technique proven to enhance their generalization performance across unseen tasks [2, 8, 18]." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 134, + 287, + 229 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 134, + 287, + 229 + ], + "spans": [ + { + "bbox": [ + 46, + 134, + 287, + 229 + ], + "type": "text", + "content": "In this work, we focus on developing the model's proficiency in understanding and following instructions related to emotional contexts. This approach highlights the importance of fine-tuning the model's instruction-following capabilities, enabling it to interpret and respond to emotional content effectively. This is achieved by leveraging its pre-existing knowledge base, thereby eliminating the necessity for an emotion-specific architectural framework." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 231, + 288, + 374 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 231, + 288, + 374 + ], + "spans": [ + { + "bbox": [ + 46, + 231, + 288, + 374 + ], + "type": "text", + "content": "To address the notable challenges encountered in Instruction Tuning for visual emotion recognition, especially the lack of specific instruction data, we introduce a novel self-generation pipeline explicitly crafted for visual emotion recognition by using GPT-4 [11]. This innovative pipeline excels in generating a diverse array of (image, instruction, output) instances, thereby notably enhancing the dataset with a more extensive and task-oriented variety of examples. This approach not only overcomes the challenge of limited data availability but also reduces the dependence on human labor. Therefore, it streamlines the process, enabling more efficient and effective emotion recognition." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 376, + 288, + 532 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 376, + 288, + 532 + ], + "spans": [ + { + "bbox": [ + 46, + 376, + 288, + 532 + ], + "type": "text", + "content": "Additionally, Instruction Tuning has been criticized for its emphasis on surface-level features like output patterns and styles, rather than achieving a profound comprehension and assimilation of tasks [26]. To tackle this issue and enhance the diversity and creativity of instruction data, our dataset includes instructions that demand complex reasoning, going beyond basic question-and-answer formats. This is further enriched by incorporating visual cues such as brightness, colorfulness, scene type, object class, facial expressions, and human actions. These aspects are pivotal in fostering a nuanced comprehension of visual emotions, thus allowing the model to generate more precise and contextually appropriate interpretations [24]." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 533, + 287, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 533, + 287, + 700 + ], + "spans": [ + { + "bbox": [ + 46, + 533, + 287, + 700 + ], + "type": "text", + "content": "After generating the emotion visual instruction data, we propose an Emotion Visual Instruction Tuning (EmoVIT) framework, leveraging the foundation of InstructBLIP [2]. This framework incorporates an emotion-centric, instruction-aware module that proficiently guides Large Language Models (LLMs) in assimilating the nuances of emotion instructions. Our work signifies a paradigm shift, presenting a new era of instruction-based learning for visual emotion understanding that relies less on explicit training data. Remarkably, as shown in Fig. 2, our approach requires almost " + }, + { + "bbox": [ + 46, + 533, + 287, + 700 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 46, + 533, + 287, + 700 + ], + "type": "text", + "content": " of the training data typically needed yet exceeds the performance of previous visual emotion recognition methods and popular Visual Instruction Tuning methods." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 59, + 701, + 258, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 701, + 258, + 712 + ], + "spans": [ + { + "bbox": [ + 59, + 701, + 258, + 712 + ], + "type": "text", + "content": "Our contributions can be summarized as follows:" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 309, + 69, + 560, + 276 + ], + "blocks": [ + { + "bbox": [ + 309, + 69, + 560, + 276 + ], + "lines": [ + { + "bbox": [ + 309, + 69, + 560, + 276 + ], + "spans": [ + { + "bbox": [ + 309, + 69, + 560, + 276 + ], + "type": "image", + "image_path": "ca388d39393f94f7229363cdad32b2a024c9d1f0313d6164e0808e710feac5e7.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 288, + 545, + 310 + ], + "lines": [ + { + "bbox": [ + 305, + 288, + 545, + 310 + ], + "spans": [ + { + "bbox": [ + 305, + 288, + 545, + 310 + ], + "type": "text", + "content": "Figure 2. Performance comparison on EmoSet test set [24] (Accuracy " + }, + { + "bbox": [ + 305, + 288, + 545, + 310 + ], + "type": "inline_equation", + "content": "\\%" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 330, + 545, + 520 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 306, + 330, + 545, + 366 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 330, + 545, + 366 + ], + "spans": [ + { + "bbox": [ + 306, + 330, + 545, + 366 + ], + "type": "text", + "content": "- We explore the potential of the Visual Instruction Tuning paradigm for emotion comprehension and introduce the concept of Emotion Visual Instruction Tuning." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 366, + 545, + 436 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 366, + 545, + 436 + ], + "spans": [ + { + "bbox": [ + 306, + 366, + 545, + 436 + ], + "type": "text", + "content": "- After thoroughly considering the unique characteristics of visual emotion recognition, we develop a novel GPT-assisted pipeline for generating emotion visual instruction data. This approach effectively bridges the gap in available annotated instruction data within this specific domain." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 437, + 545, + 520 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 437, + 545, + 520 + ], + "spans": [ + { + "bbox": [ + 306, + 437, + 545, + 520 + ], + "type": "text", + "content": "- Building upon the foundation of InstructBLIP, our EmoVIT architecture integrates emotion domain-specific instruction data, harnessing the robust capabilities of LLMs to boost performance. The extensive experiments demonstrate our model's proficiency in emotion classification, affective reasoning, and comprehension of humour." + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 532, + 392, + 544 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 532, + 392, + 544 + ], + "spans": [ + { + "bbox": [ + 306, + 532, + 392, + 544 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 552, + 460, + 564 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 552, + 460, + 564 + ], + "spans": [ + { + "bbox": [ + 306, + 552, + 460, + 564 + ], + "type": "text", + "content": "2.1. Visual Emotion Recognition" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "text", + "content": "A key challenge in visual emotion recognition is bridging the gap between an image's visual cues and the emotions it portrays [12, 21, 25]. While traditional efforts, e.g., Xu et al.'s multi-level dependent attention network [21], focus on visual models for emotional feature learning, recent advancements like EmoSet [24] offer rich emotion-laden datasets with 3.3 million images. The rise of multimodal models, such as the GPT series [11], has further propelled Vision-Language Recognition. However, fully leveraging these models in emotion recognition is an area ripe for exploration. Our work leads the way in utilizing large-scale models for Emotion Visual Instruction Tuning." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "26597" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 78, + 68, + 516, + 257 + ], + "blocks": [ + { + "bbox": [ + 78, + 68, + 516, + 257 + ], + "lines": [ + { + "bbox": [ + 78, + 68, + 516, + 257 + ], + "spans": [ + { + "bbox": [ + 78, + 68, + 516, + 257 + ], + "type": "image", + "image_path": "9365186dfb4f66d5b67d49c60155a0e09792131e67ecb988d89722371788fb2f.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 181, + 262, + 411, + 274 + ], + "lines": [ + { + "bbox": [ + 181, + 262, + 411, + 274 + ], + "spans": [ + { + "bbox": [ + 181, + 262, + 411, + 274 + ], + "type": "text", + "content": "Figure 3. The comparison of different visual tuning paradigms." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 284, + 191, + 297 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 284, + 191, + 297 + ], + "spans": [ + { + "bbox": [ + 47, + 284, + 191, + 297 + ], + "type": "text", + "content": "2.2. Visual Instruction Tuning" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 303, + 288, + 482 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 303, + 288, + 482 + ], + "spans": [ + { + "bbox": [ + 46, + 303, + 288, + 482 + ], + "type": "text", + "content": "Current Large Language Models (LLMs) have extensive knowledge bases, but their effectiveness depends on accurately interpreting human instructions due to a mismatch between training goals and user expectations. LLMs are trained to minimize prediction errors, whereas users expect helpful and safe instruction-following. Instruction Tuning addresses this by teaching models to follow natural language instructions, enhancing generalization to new tasks. FLAN [18] demonstrated that training a large model on instruction-based datasets improves zero-shot performance. This approach has extended to vision-language tasks, with BLIP2 [5] and LLaVA [8] adapting instruction-tuned LLMs for visual inputs. InstructBLIP [2] introduces instruction-aware visual feature extraction and the Q-Former, enabling more flexible, instruction-driven feature extraction." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 483, + 287, + 542 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 483, + 287, + 542 + ], + "spans": [ + { + "bbox": [ + 46, + 483, + 287, + 542 + ], + "type": "text", + "content": "As a novel area, visual emotion instruction tuning lacks benchmarks or guidelines for creating emotion instruction data. Our work pioneers the use of large-scale models to develop an emotion instruction data pipeline, overcoming the limitations of manual annotation." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 554, + 103, + 566 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 554, + 103, + 566 + ], + "spans": [ + { + "bbox": [ + 47, + 554, + 103, + 566 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 574, + 262, + 588 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 574, + 262, + 588 + ], + "spans": [ + { + "bbox": [ + 47, + 574, + 262, + 588 + ], + "type": "text", + "content": "3.1. Preliminary of Visual Instruction Tuning" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 594, + 287, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 287, + 617 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 287, + 617 + ], + "type": "text", + "content": "In the deep learning era, visual tuning has experienced significant paradigm shifts, as depicted in Fig. 3." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 618, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 618, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 618, + 288, + 714 + ], + "type": "text", + "content": "In Fig. 3(a), conventional tuning methodologies encompass Full fine-tuning, Head-oriented, and Backbone-oriented techniques, capitalizing on large-scale pre-trained models. Predominantly, thoroughly fine-tuning these models for specific tasks, conducted end-to-end, is recognized as a highly effective strategy. However, this method requires maintaining separate copies of the backbone parameters for each distinct task, posing challenges in storage and deploy-" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 286, + 331, + 295 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 286, + 331, + 295 + ], + "spans": [ + { + "bbox": [ + 306, + 286, + 331, + 295 + ], + "type": "text", + "content": "ment." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 297, + 545, + 369 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 297, + 545, + 369 + ], + "spans": [ + { + "bbox": [ + 304, + 297, + 545, + 369 + ], + "type": "text", + "content": "Alternatively, Visual Prompt Tuning (VPT) [3], presents an efficient substitute for full fine-tuning within large-scale vision Transformer models. It achieves this by employing a minimal fraction of trainable parameters in the input space while maintaining a frozen backbone model. The objective function for Visual Prompt Tuning is given by:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 377, + 380, + 545, + 399 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 377, + 380, + 545, + 399 + ], + "spans": [ + { + "bbox": [ + 377, + 380, + 545, + 399 + ], + "type": "interline_equation", + "content": "\\min _ {\\theta_ {\\mathrm {P}}} \\mathcal {L} (f (X, P; \\theta_ {\\mathrm {P}}), Y) \\tag {1}", + "image_path": "e710a08d9d13201821e62fa77f5a2c6a92b2d78b1d869a9d317340949a5a5b0d.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 409, + 545, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 409, + 545, + 468 + ], + "spans": [ + { + "bbox": [ + 304, + 409, + 545, + 468 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 409, + 545, + 468 + ], + "type": "inline_equation", + "content": "\\min_{\\theta_{\\mathrm{P}}}" + }, + { + "bbox": [ + 304, + 409, + 545, + 468 + ], + "type": "text", + "content": " is the minimization over the prompt parameters " + }, + { + "bbox": [ + 304, + 409, + 545, + 468 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 304, + 409, + 545, + 468 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 409, + 545, + 468 + ], + "type": "inline_equation", + "content": "\\mathcal{L}" + }, + { + "bbox": [ + 304, + 409, + 545, + 468 + ], + "type": "text", + "content": " is the loss function, " + }, + { + "bbox": [ + 304, + 409, + 545, + 468 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 304, + 409, + 545, + 468 + ], + "type": "text", + "content": " represents the model function with input image " + }, + { + "bbox": [ + 304, + 409, + 545, + 468 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 304, + 409, + 545, + 468 + ], + "type": "text", + "content": ", prompt parameters " + }, + { + "bbox": [ + 304, + 409, + 545, + 468 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 304, + 409, + 545, + 468 + ], + "type": "text", + "content": ", and learnable model parameters " + }, + { + "bbox": [ + 304, + 409, + 545, + 468 + ], + "type": "inline_equation", + "content": "\\theta_{\\mathrm{P}}" + }, + { + "bbox": [ + 304, + 409, + 545, + 468 + ], + "type": "text", + "content": " as input, and " + }, + { + "bbox": [ + 304, + 409, + 545, + 468 + ], + "type": "inline_equation", + "content": "Y" + }, + { + "bbox": [ + 304, + 409, + 545, + 468 + ], + "type": "text", + "content": " is the target output." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 470, + 546, + 660 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 470, + 546, + 660 + ], + "spans": [ + { + "bbox": [ + 304, + 470, + 546, + 660 + ], + "type": "text", + "content": "Visual Prompt Tuning focuses on optimizing LLMs using a small set of parameters, whereas Visual Instruction Tuning (VIT) aims to improve the model's comprehension of instructions, thereby addressing the model's shortcomings in specific domains. This type of method aims to enhance the model's proficiency in following instructions, leveraging the capabilities of the latest foundation models, e.g., Llama [17], and BLIP2 [5]. Instructions serve as guiding constraints, shaping the model's outputs to conform to specific response characteristics and domain-relevant knowledge. This approach enables human monitoring of the model's behavior, thereby assuring alignment with the desired outcomes. Moreover, Instruction Tuning is computationally efficient, allowing LLMs to swiftly adapt to particular domains without extensive retraining or architectural alterations." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 662, + 545, + 686 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 662, + 545, + 686 + ], + "spans": [ + { + "bbox": [ + 304, + 662, + 545, + 686 + ], + "type": "text", + "content": "The objective function for Visual Instruction Tuning is given by:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 363, + 697, + 545, + 716 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 363, + 697, + 545, + 716 + ], + "spans": [ + { + "bbox": [ + 363, + 697, + 545, + 716 + ], + "type": "interline_equation", + "content": "\\min _ {\\theta_ {\\text {t u n a b l e}}} \\mathcal {L} (g (X, I, C; \\theta_ {\\text {t u n a b l e}}), Y) \\tag {2}", + "image_path": "788c011b6f296af4d0d4eb0b5f0c99c0fdbb9689f22bc78de7008641e019706f.jpg" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "26598" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 51, + 98, + 137, + 159 + ], + "blocks": [ + { + "bbox": [ + 51, + 98, + 137, + 159 + ], + "lines": [ + { + "bbox": [ + 51, + 98, + 137, + 159 + ], + "spans": [ + { + "bbox": [ + 51, + 98, + 137, + 159 + ], + "type": "image", + "image_path": "9342f2ab39a50f8859b88420ccd7949deef2992314bd34406459d0f268e2bd0e.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 60, + 163, + 96, + 169 + ], + "lines": [ + { + "bbox": [ + 60, + 163, + 96, + 169 + ], + "spans": [ + { + "bbox": [ + 60, + 163, + 96, + 169 + ], + "type": "text", + "content": "Emotion class" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 100, + 163, + 132, + 168 + ], + "lines": [ + { + "bbox": [ + 100, + 163, + 132, + 168 + ], + "spans": [ + { + "bbox": [ + 100, + 163, + 132, + 168 + ], + "type": "text", + "content": "Contentment" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 60, + 175, + 91, + 181 + ], + "lines": [ + { + "bbox": [ + 60, + 175, + 91, + 181 + ], + "spans": [ + { + "bbox": [ + 60, + 175, + 91, + 181 + ], + "type": "text", + "content": "Brightness" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 101, + 175, + 110, + 180 + ], + "lines": [ + { + "bbox": [ + 101, + 175, + 110, + 180 + ], + "spans": [ + { + "bbox": [ + 101, + 175, + 110, + 180 + ], + "type": "text", + "content": "0.8" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 60, + 187, + 93, + 193 + ], + "lines": [ + { + "bbox": [ + 60, + 187, + 93, + 193 + ], + "spans": [ + { + "bbox": [ + 60, + 187, + 93, + 193 + ], + "type": "text", + "content": "Colorfulness" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 101, + 187, + 110, + 193 + ], + "lines": [ + { + "bbox": [ + 101, + 187, + 110, + 193 + ], + "spans": [ + { + "bbox": [ + 101, + 187, + 110, + 193 + ], + "type": "text", + "content": "0.8" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 60, + 199, + 91, + 205 + ], + "lines": [ + { + "bbox": [ + 60, + 199, + 91, + 205 + ], + "spans": [ + { + "bbox": [ + 60, + 199, + 91, + 205 + ], + "type": "text", + "content": "Scene type" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 141, + 83, + 239, + 262 + ], + "blocks": [ + { + "bbox": [ + 101, + 199, + 118, + 205 + ], + "lines": [ + { + "bbox": [ + 101, + 199, + 118, + 205 + ], + "spans": [ + { + "bbox": [ + 101, + 199, + 118, + 205 + ], + "type": "text", + "content": "Beach" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 60, + 211, + 91, + 217 + ], + "lines": [ + { + "bbox": [ + 60, + 211, + 91, + 217 + ], + "spans": [ + { + "bbox": [ + 60, + 211, + 91, + 217 + ], + "type": "text", + "content": "Object class" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 101, + 211, + 110, + 217 + ], + "lines": [ + { + "bbox": [ + 101, + 211, + 110, + 217 + ], + "spans": [ + { + "bbox": [ + 101, + 211, + 110, + 217 + ], + "type": "text", + "content": "Sky" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 60, + 224, + 91, + 229 + ], + "lines": [ + { + "bbox": [ + 60, + 224, + 91, + 229 + ], + "spans": [ + { + "bbox": [ + 60, + 224, + 91, + 229 + ], + "type": "text", + "content": "Facial" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 101, + 224, + 132, + 230 + ], + "lines": [ + { + "bbox": [ + 101, + 224, + 132, + 230 + ], + "spans": [ + { + "bbox": [ + 101, + 224, + 132, + 230 + ], + "type": "text", + "content": "Contentment" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 60, + 238, + 85, + 243 + ], + "lines": [ + { + "bbox": [ + 60, + 238, + 85, + 243 + ], + "spans": [ + { + "bbox": [ + 60, + 238, + 85, + 243 + ], + "type": "text", + "content": "Human" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 101, + 239, + 125, + 244 + ], + "lines": [ + { + "bbox": [ + 101, + 239, + 125, + 244 + ], + "spans": [ + { + "bbox": [ + 101, + 239, + 125, + 244 + ], + "type": "text", + "content": "Laughing." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 101, + 244, + 125, + 250 + ], + "lines": [ + { + "bbox": [ + 101, + 244, + 125, + 250 + ], + "spans": [ + { + "bbox": [ + 101, + 244, + 125, + 250 + ], + "type": "text", + "content": "Ludwig" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 141, + 83, + 239, + 262 + ], + "lines": [ + { + "bbox": [ + 141, + 83, + 239, + 262 + ], + "spans": [ + { + "bbox": [ + 141, + 83, + 239, + 262 + ], + "type": "image", + "image_path": "10f7bfb5bdea883c73faa2db931822742351b94eebfa64006c7229a06b6ca39d.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 67, + 264, + 218, + 274 + ], + "lines": [ + { + "bbox": [ + 67, + 264, + 218, + 274 + ], + "spans": [ + { + "bbox": [ + 67, + 264, + 218, + 274 + ], + "type": "text", + "content": "(a) Emotion Visual Instruction Data Generation" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 288, + 545, + 344 + ], + "lines": [ + { + "bbox": [ + 46, + 288, + 545, + 344 + ], + "spans": [ + { + "bbox": [ + 46, + 288, + 545, + 344 + ], + "type": "text", + "content": "Figure 4. The overall architecture of our proposed method. The Emotion Instruction data generated by (a) will be used for Emotion Visual Instruction Tuning in (b). During Emotion Visual Instruction Tuning, given an input image, the frozen Image Encoder initiates the process by extracting visual features. Emotion Instruction generated by (a) are subsequently interacting with Queries embedding through the learnable Q-Former. This interaction is key to drawing out image features that are relevant to the task at hand. As a result, the frozen LLM receives visual information conducive to instruction following." + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 242, + 77, + 408, + 261 + ], + "blocks": [ + { + "bbox": [ + 242, + 77, + 408, + 261 + ], + "lines": [ + { + "bbox": [ + 242, + 77, + 408, + 261 + ], + "spans": [ + { + "bbox": [ + 242, + 77, + 408, + 261 + ], + "type": "image", + "image_path": "831fecbf7d7448ec656e1d747dd2c4ffab7b3c61068499a8d6d2a04fdcc61957.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 242, + 264, + 405, + 274 + ], + "lines": [ + { + "bbox": [ + 242, + 264, + 405, + 274 + ], + "spans": [ + { + "bbox": [ + 242, + 264, + 405, + 274 + ], + "type": "text", + "content": "(b) Emotion Visual Instruction Tuning Architecture" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 425, + 99, + 547, + 250 + ], + "blocks": [ + { + "bbox": [ + 425, + 99, + 547, + 250 + ], + "lines": [ + { + "bbox": [ + 425, + 99, + 547, + 250 + ], + "spans": [ + { + "bbox": [ + 425, + 99, + 547, + 250 + ], + "type": "image", + "image_path": "ea875063c75ecf06e8fc35841680bf9a653c54e0f47e94ec3f979194722849c3.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 425, + 264, + 542, + 274 + ], + "lines": [ + { + "bbox": [ + 425, + 264, + 542, + 274 + ], + "spans": [ + { + "bbox": [ + 425, + 264, + 542, + 274 + ], + "type": "text", + "content": "(c) The Details of Q-Former Module" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 20 + }, + { + "bbox": [ + 46, + 354, + 288, + 474 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 354, + 288, + 474 + ], + "spans": [ + { + "bbox": [ + 46, + 354, + 288, + 474 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 354, + 288, + 474 + ], + "type": "inline_equation", + "content": "\\min_{\\theta_{\\mathrm{tunable}}}" + }, + { + "bbox": [ + 46, + 354, + 288, + 474 + ], + "type": "text", + "content": " denotes the minimization over the tunable parameters " + }, + { + "bbox": [ + 46, + 354, + 288, + 474 + ], + "type": "inline_equation", + "content": "\\theta_{\\mathrm{tunable}}" + }, + { + "bbox": [ + 46, + 354, + 288, + 474 + ], + "type": "text", + "content": " in the Instruction Tuning Module, " + }, + { + "bbox": [ + 46, + 354, + 288, + 474 + ], + "type": "inline_equation", + "content": "\\mathcal{L}" + }, + { + "bbox": [ + 46, + 354, + 288, + 474 + ], + "type": "text", + "content": " is the loss function, " + }, + { + "bbox": [ + 46, + 354, + 288, + 474 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 46, + 354, + 288, + 474 + ], + "type": "text", + "content": " is the model function with instruction " + }, + { + "bbox": [ + 46, + 354, + 288, + 474 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 46, + 354, + 288, + 474 + ], + "type": "text", + "content": ", image " + }, + { + "bbox": [ + 46, + 354, + 288, + 474 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 46, + 354, + 288, + 474 + ], + "type": "text", + "content": ", other contexts " + }, + { + "bbox": [ + 46, + 354, + 288, + 474 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 46, + 354, + 288, + 474 + ], + "type": "text", + "content": ", and tunable parameters " + }, + { + "bbox": [ + 46, + 354, + 288, + 474 + ], + "type": "inline_equation", + "content": "\\theta_{\\mathrm{tunable}}" + }, + { + "bbox": [ + 46, + 354, + 288, + 474 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 354, + 288, + 474 + ], + "type": "inline_equation", + "content": "Y" + }, + { + "bbox": [ + 46, + 354, + 288, + 474 + ], + "type": "text", + "content": " denotes the target output. The optional context " + }, + { + "bbox": [ + 46, + 354, + 288, + 474 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 46, + 354, + 288, + 474 + ], + "type": "text", + "content": " is not just raw data; it encompasses descriptive or directive information guiding the model on how to process input or which task to execute, e.g., image caption. It's integral to the model's understanding and execution of tasks based on specific instructions or guidelines." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 47, + 479, + 287, + 503 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 479, + 287, + 503 + ], + "spans": [ + { + "bbox": [ + 47, + 479, + 287, + 503 + ], + "type": "text", + "content": "3.2. GPT-assisted Emotion Visual Instruction Data Generation" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 46, + 510, + 287, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 510, + 287, + 640 + ], + "spans": [ + { + "bbox": [ + 46, + 510, + 287, + 640 + ], + "type": "text", + "content": "Previous methodologies commonly employed a consistent template-based set of instructions for every image within a dataset across various specific tasks [2]. For instance, a standard instruction such as \"Briefly describe the content of the image\" was employed uniformly across all images for Image Captioning. In this way, the model may not be able to adequately capture the unique characteristics of each image. Moreover, this one-size-fits-all approach often leads to suboptimal performance in emotion recognition tasks that require nuanced perception and differentiation of ambiguous emotion classes." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 46, + 642, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 642, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 642, + 287, + 714 + ], + "type": "text", + "content": "Since the topic of Emotion Visual Instruction Tuning is still in its infancy, no benchmarks or guidelines have been proposed so far for constructing emotion instruction data. Based on the recent successes of machine-generated instructions demonstrated in LLaVA [8], our work pioneers the use of existing LLMs to create a pipeline for" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 304, + 354, + 545, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 354, + 545, + 437 + ], + "spans": [ + { + "bbox": [ + 304, + 354, + 545, + 437 + ], + "type": "text", + "content": "self-generating emotion instructions. Different from previous template-based and one-size-fits-all instruction data, we propose an instance-wise and LLM-assisted visual emotion instruction data pipeline. This methodology transcends the constraints of manual annotation by employing GPT-4 [11] to generate instance-wise, tailored instruction data that dynamically corresponds to visual content." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 304, + 438, + 545, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 438, + 545, + 665 + ], + "spans": [ + { + "bbox": [ + 304, + 438, + 545, + 665 + ], + "type": "text", + "content": "Prior to the development of instructional data for the visual emotion recognition task, it is imperative to confront a fundamental academic problem: What types of visual clues are pivotal in identifying emotions? This necessitates a careful consideration of the unique characteristics inherent to the task, along with a comprehensive understanding of the potential visual cues associated with human emotions. In this work, we propose a novel visual instruction data mechanism to remove the inherent subjectivity and ambiguity in emotional interpretation. Specifically, we integrate a broad spectrum of emotion attributes across multiple levels: low-level attributes (e.g., brightness, colorfulness), mid-level attributes (e.g., scene type and object class), and high-level attributes (e.g., facial expressions and human actions), building upon insights from previous work [24]. This comprehensive strategy not only aligns with the intricate nature of emotions but also significantly enhances the model's capability to interpret and understand visual emotional cues more accurately and holistically." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 304, + 665, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 545, + 714 + ], + "type": "text", + "content": "The overall pipeline of our proposed emotion visual instruction data is shown in Fig. 4 (a). For an image " + }, + { + "bbox": [ + 304, + 665, + 545, + 714 + ], + "type": "inline_equation", + "content": "X_{\\mathrm{img}}" + }, + { + "bbox": [ + 304, + 665, + 545, + 714 + ], + "type": "text", + "content": ", three types of image-related contexts are essential for GPT-4 to generate emotion instruction data: (i) a caption " + }, + { + "bbox": [ + 304, + 665, + 545, + 714 + ], + "type": "inline_equation", + "content": "X_{c}" + }, + { + "bbox": [ + 304, + 665, + 545, + 714 + ], + "type": "text", + "content": ", (ii)" + } + ] + } + ], + "index": 29 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "26599" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 131 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 131 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 131 + ], + "type": "text", + "content": "an emotion attribute list " + }, + { + "bbox": [ + 46, + 72, + 287, + 131 + ], + "type": "inline_equation", + "content": "X_{\\mathrm{attr}}" + }, + { + "bbox": [ + 46, + 72, + 287, + 131 + ], + "type": "text", + "content": ", which includes emotion class, brightness, colorfulness, scene type, object class, facial expression, and human action, and (iii) the system prompt, designed to enable GPT-4 to comprehend the specific task requirement1." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 133, + 288, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 133, + 288, + 251 + ], + "spans": [ + { + "bbox": [ + 46, + 133, + 288, + 251 + ], + "type": "text", + "content": "We first manually design a few examples which are used as seed examples for in-context learning to query GPT-4. This operation leverages the model's ability to extrapolate from given examples, enhancing its understanding and response accuracy based on the principles of few-shot learning [8]. Our generated emotion instruction data includes three types: Categorical, Conversation, and Reasoning. Building upon previous research [8], our generated instruction data adheres to the dialogue format, exemplified in Fig. 5." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 252, + 287, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 252, + 287, + 323 + ], + "spans": [ + { + "bbox": [ + 46, + 252, + 287, + 323 + ], + "type": "text", + "content": "Our strategy for generating emotion instruction data adopts a progressive approach from simple to complex. Initially, for the Categorical data, we transform the associated emotion class of the image into a structured format. This process serves as the foundational component of our emotion instruction data." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 324, + 287, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 324, + 287, + 491 + ], + "spans": [ + { + "bbox": [ + 46, + 324, + 287, + 491 + ], + "type": "text", + "content": "For the Conversation data, our framework is designed to create dialogues in which the GPT assistant interacts with an inquirer, focusing on the emotion attributes of the image. In this setup, the assistant's responses are tailored to interpret and describe the image as though it were within its own visual field, thereby providing insights from an observational viewpoint. The scope of questions posed is comprehensive, encompassing the types of objects depicted, their actions, and the dynamics of their interrelationships. The dialogues we generate fall into two categories: (i) Basic Interaction, focusing on the provided emotion attribute list with simple, direct characteristics, and (ii) Advanced Interaction, which builds on the first type to reach greater conversational complexity and sophistication." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 491, + 287, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 491, + 287, + 574 + ], + "spans": [ + { + "bbox": [ + 46, + 491, + 287, + 574 + ], + "type": "text", + "content": "For the Reasoning data, our approach extends beyond mere visual content, prompting the model to generate in-depth reasoning questions. To enhance the dialogue's credibility and structure, detailed examples are incorporated alongside logical reasoning steps, ensuring that the discourse convincingly captures the intricacies of the visual content." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 582, + 234, + 596 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 582, + 234, + 596 + ], + "spans": [ + { + "bbox": [ + 47, + 582, + 234, + 596 + ], + "type": "text", + "content": "3.3. Emotion Visual Instruction Tuning" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 601, + 287, + 661 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 601, + 287, + 661 + ], + "spans": [ + { + "bbox": [ + 46, + 601, + 287, + 661 + ], + "type": "text", + "content": "After acquiring the emotion visual instruction data as detailed in Sec. 3.2, our goal is to employ this data in enhancing the existing Visual Instruction Tuning model. This enhancement aims to align the LLMs' existing knowledge with the emotion understanding domain." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 662, + 287, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 662, + 287, + 685 + ], + "spans": [ + { + "bbox": [ + 47, + 662, + 287, + 685 + ], + "type": "text", + "content": "As shown in Fig. 4 b, we have developed an Emotion Visual Instruction Tuning (EmoVIT) architecture based on" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 373, + 74, + 492, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 373, + 74, + 492, + 83 + ], + "spans": [ + { + "bbox": [ + 373, + 74, + 492, + 83 + ], + "type": "text", + "content": "Our Emotion Instruction Data Sample" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 320, + 91, + 389, + 141 + ], + "blocks": [ + { + "bbox": [ + 320, + 91, + 389, + 141 + ], + "lines": [ + { + "bbox": [ + 320, + 91, + 389, + 141 + ], + "spans": [ + { + "bbox": [ + 320, + 91, + 389, + 141 + ], + "type": "image", + "image_path": "2f4119e5872a968c563b86bb5c30d6c9e26c9413d6245d39614774b22b57e4bd.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 389, + 556, + 411 + ], + "lines": [ + { + "bbox": [ + 305, + 389, + 556, + 411 + ], + "spans": [ + { + "bbox": [ + 305, + 389, + 556, + 411 + ], + "type": "text", + "content": "Figure 5. The sample of our generated visual emotion instruction data." + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 392, + 94, + 457, + 101 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 392, + 94, + 457, + 101 + ], + "spans": [ + { + "bbox": [ + 392, + 94, + 457, + 101 + ], + "type": "text", + "content": "Context type 1: Caption:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 392, + 102, + 492, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 392, + 102, + 492, + 108 + ], + "spans": [ + { + "bbox": [ + 392, + 102, + 492, + 108 + ], + "type": "text", + "content": "Two young women dancing in a kitchen." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 392, + 112, + 461, + 118 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 392, + 112, + 461, + 118 + ], + "spans": [ + { + "bbox": [ + 392, + 112, + 461, + 118 + ], + "type": "text", + "content": "Context type 2: Attribute:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 392, + 119, + 524, + 125 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 392, + 119, + 524, + 125 + ], + "spans": [ + { + "bbox": [ + 392, + 119, + 524, + 125 + ], + "type": "text", + "content": "Emotion: contentment Human action: stretching arm" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 392, + 126, + 545, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 392, + 126, + 545, + 133 + ], + "spans": [ + { + "bbox": [ + 392, + 126, + 545, + 133 + ], + "type": "text", + "content": "Brightness: 0.7 Object: ['Cupboard', 'Food', 'Countertop']" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 392, + 134, + 511, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 392, + 134, + 511, + 140 + ], + "spans": [ + { + "bbox": [ + 392, + 134, + 511, + 140 + ], + "type": "text", + "content": "Colorfulness: 0.4 Facial expression: happy" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 334, + 157, + 370, + 165 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 334, + 157, + 370, + 165 + ], + "spans": [ + { + "bbox": [ + 334, + 157, + 370, + 165 + ], + "type": "text", + "content": "Categorical" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 321, + 167, + 537, + 192 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 167, + 537, + 192 + ], + "spans": [ + { + "bbox": [ + 321, + 167, + 537, + 192 + ], + "type": "text", + "content": "Question: Please select the emotion closest to the image from the following options: amusement, anger, awe, contentment, disgust, excitement, fear and sadness." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 321, + 193, + 436, + 201 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 193, + 436, + 201 + ], + "spans": [ + { + "bbox": [ + 321, + 193, + 436, + 201 + ], + "type": "text", + "content": "Answer: Predict emotion: contentment." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 332, + 207, + 372, + 214 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 332, + 207, + 372, + 214 + ], + "spans": [ + { + "bbox": [ + 332, + 207, + 372, + 214 + ], + "type": "text", + "content": "Conversation" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 390, + 207, + 439, + 214 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 390, + 207, + 439, + 214 + ], + "spans": [ + { + "bbox": [ + 390, + 207, + 439, + 214 + ], + "type": "text", + "content": "Basic Interaction" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 321, + 217, + 506, + 225 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 217, + 506, + 225 + ], + "spans": [ + { + "bbox": [ + 321, + 217, + 506, + 225 + ], + "type": "text", + "content": "Question: Can you describe the activity happening in the image?" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 321, + 225, + 526, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 225, + 526, + 251 + ], + "spans": [ + { + "bbox": [ + 321, + 225, + 526, + 251 + ], + "type": "text", + "content": "Answer: The image depicts two young women enjoying themselves by dancing in a kitchen. One of the women is seen stretching her arm, most likely as part of their dance routine." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 332, + 255, + 372, + 262 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 332, + 255, + 372, + 262 + ], + "spans": [ + { + "bbox": [ + 332, + 255, + 372, + 262 + ], + "type": "text", + "content": "Conversation" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 383, + 254, + 446, + 262 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 383, + 254, + 446, + 262 + ], + "spans": [ + { + "bbox": [ + 383, + 254, + 446, + 262 + ], + "type": "text", + "content": "Advanced Interaction" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 321, + 265, + 543, + 281 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 265, + 543, + 281 + ], + "spans": [ + { + "bbox": [ + 321, + 265, + 543, + 281 + ], + "type": "text", + "content": "Question: Based on the given emotion of contentment and the activities in the image, can you guess about the possible context or story of this scenario?" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 321, + 281, + 542, + 314 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 281, + 542, + 314 + ], + "spans": [ + { + "bbox": [ + 321, + 281, + 542, + 314 + ], + "type": "text", + "content": "Answer: These women are good friends or roommates, spending quality time together. The act of dancing, especially in an unconventional place like a kitchen, suggests they are comfortable with each other and are enjoying a break from routine." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 335, + 320, + 368, + 329 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 320, + 368, + 329 + ], + "spans": [ + { + "bbox": [ + 335, + 320, + 368, + 329 + ], + "type": "text", + "content": "Reasoning" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 320, + 331, + 545, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 331, + 545, + 373 + ], + "spans": [ + { + "bbox": [ + 320, + 331, + 545, + 373 + ], + "type": "text", + "content": "Question: Based on the setting and activities of the two young women, what might have prompted this impromptu dance session in the middle of a kitchen? Answer: The impromptu dance could have been prompted by a favorite song playing, or simple spontaneous fun amidst preparing a meal or organizing the kitchen." + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 305, + 424, + 545, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 424, + 545, + 460 + ], + "spans": [ + { + "bbox": [ + 305, + 424, + 545, + 460 + ], + "type": "text", + "content": "InstructBLIP [2]. This architecture specifically leverages its Instruction-aware Q-Former Module, as depicted in Fig. 4 c, for emotion-centric instructional tasks." + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 304, + 461, + 546, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 461, + 546, + 616 + ], + "spans": [ + { + "bbox": [ + 304, + 461, + 546, + 616 + ], + "type": "text", + "content": "Specifically, the Instruction-aware Q-Former Module takes in the emotion instruction tokens, queries, and image embeddings as input. The image embeddings are extracted by a frozen image encoder. The learnable queries are initially produced by the pre-trained Q-Former of InstructBLIP. During training, the Instruction-aware module enhances task-specific feature extraction. It does this by integrating emotion instruction and query embeddings within self-attention layers, aligning visual information with the LLM's instruction-following requirements. Our approach adopts cross-entropy loss, tailoring it to the intricacies of visual emotion recognition tasks, thus ensuring precise and contextually relevant model training outcomes." + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 304, + 617, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 617, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 617, + 547, + 714 + ], + "type": "text", + "content": "We note that the data generated by our approach is not confined to a single model but can also be applied to other Visual Instruction Tuning models, such as LLaVA [17]. Notably, when LLaVA is fine-tuned with our data, it exhibits a significant enhancement in emotion recognition capabilities, as detailed in Sec. 4.2. In this way, we demonstrate not only the effectiveness but also the transferability of our generated data, showing its broad applicability and impact." + } + ] + } + ], + "index": 33 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 47, + 693, + 287, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 693, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 693, + 287, + 713 + ], + "type": "text", + "content": "1A detailed description of the system prompt is provided in the supplementary materials." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "26600" + } + ] + } + ], + "index": 34 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 71, + 173, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 71, + 173, + 85 + ], + "spans": [ + { + "bbox": [ + 47, + 71, + 173, + 85 + ], + "type": "text", + "content": "4. Experimental Results" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 91, + 165, + 104 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 91, + 165, + 104 + ], + "spans": [ + { + "bbox": [ + 47, + 91, + 165, + 104 + ], + "type": "text", + "content": "4.1. Implemental Details" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 109, + 286, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 109, + 286, + 180 + ], + "spans": [ + { + "bbox": [ + 46, + 109, + 286, + 180 + ], + "type": "text", + "content": "Our implementation is based on the LAVIS library [4]. Our EmoVIT starts with a pre-trained InstructBLIP baseline and proceeds to fine-tune exclusively the Q-Former module, whilst keeping both the image encoder and the language model frozen. The parameters for our training adhere to the default settings established by InstructBLIP." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 181, + 286, + 252 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 181, + 286, + 252 + ], + "spans": [ + { + "bbox": [ + 46, + 181, + 286, + 252 + ], + "type": "text", + "content": "Datasets. We evaluate our framework on ten benchmark datasets annotated under different scenarios and class number, namely EmoSet [24], WEBEmo [12], Emotion6 [14], the Flickr and Instagram (FI) [25], Artphoto [9], IAPS [10], Abstract [9], EmotionROI [15], UnbiasedEmo [12], and OxfordTVG-HIC [6]." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 253, + 286, + 348 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 253, + 286, + 348 + ], + "spans": [ + { + "bbox": [ + 46, + 253, + 286, + 348 + ], + "type": "text", + "content": "Held-in Pretraining. Following previous work [2], we divide our dataset into two categories: held-in for pretraining and held-out for evaluation2. Considering the EmoSSet dataset's comprehensive inclusion of emotion attributes for each image, it has been chosen as the primary resource for our held-in pretraining phase. Simultaneously, for a broader assessment, we perform held-out evaluations using the test sets from various other datasets." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 350, + 286, + 433 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 350, + 286, + 433 + ], + "spans": [ + { + "bbox": [ + 46, + 350, + 286, + 433 + ], + "type": "text", + "content": "For the generation of emotion visual instruction data, we initially employ the BLIP2 model for image captioning, followed by leveraging the GPT-4 API to generate emotion instruction data. In total, our collection comprises Categorical, Conversation, and Reasoning instruction data, derived from 51,200 unique images. This represents less than " + }, + { + "bbox": [ + 46, + 350, + 286, + 433 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 46, + 350, + 286, + 433 + ], + "type": "text", + "content": " of the entire EmoSet." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 441, + 165, + 453 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 441, + 165, + 453 + ], + "spans": [ + { + "bbox": [ + 47, + 441, + 165, + 453 + ], + "type": "text", + "content": "4.2. Held-out Evaluation" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 460, + 286, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 460, + 286, + 567 + ], + "spans": [ + { + "bbox": [ + 46, + 460, + 286, + 567 + ], + "type": "text", + "content": "As shown in Tab. 1, our proposed methodology exhibits a marked superiority in performance relative to the burgeoning Visual Instruction Tuning Methods. Since they have been pre-trained on dozens of large-scale datasets, it is evident that our generated emotion visual instruction data is particularly effective for emotional understanding. Our results signify a paradigm shift, heralding a new era of model training that relies less on explicit supervision and more on the robustness of emotion instruction-driven learning." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 567, + 286, + 675 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 567, + 286, + 675 + ], + "spans": [ + { + "bbox": [ + 46, + 567, + 286, + 675 + ], + "type": "text", + "content": "The Effectiveness of Our Proposed Emotion Visual Instruction Data. As the first to introduce the concept of emotion visual instruction data, our study seeks to evaluate the generalizability of this newly generated instruction data. Our goal is to test its efficacy not only with InstructBLIP but also across other Visual Instruction Tuning model, to understand its broader applicability. As depicted in Fig. 6, we employ two Visual Instruction Tuning models, LLaVA and InstructBLIP, which were fine-tuned on our specially gen" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 310, + 83, + 556, + 255 + ], + "blocks": [ + { + "bbox": [ + 347, + 74, + 536, + 83 + ], + "lines": [ + { + "bbox": [ + 347, + 74, + 536, + 83 + ], + "spans": [ + { + "bbox": [ + 347, + 74, + 536, + 83 + ], + "type": "text", + "content": "Accuracy Improvements of FT Models Over Base Models" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 310, + 83, + 556, + 255 + ], + "lines": [ + { + "bbox": [ + 310, + 83, + 556, + 255 + ], + "spans": [ + { + "bbox": [ + 310, + 83, + 556, + 255 + ], + "type": "image", + "image_path": "ef9f5da6de3e6f86a45bf5ddf42b29074a440d3f8afefe5d124bd6be0d190963.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 267, + 556, + 289 + ], + "lines": [ + { + "bbox": [ + 305, + 267, + 556, + 289 + ], + "spans": [ + { + "bbox": [ + 305, + 267, + 556, + 289 + ], + "type": "text", + "content": "Figure 6. The improvement of our proposed emotion visual instruction tuning data tuning on LLaVA [8] and InstructBLIP [2]." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 303, + 545, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 303, + 545, + 422 + ], + "spans": [ + { + "bbox": [ + 304, + 303, + 545, + 422 + ], + "type": "text", + "content": "erated emotion visual instruction data. Subsequent testing across five distinct datasets reveals notable improvements in both models, substantiating the efficacy of our generated data. Notably, InstructBLIP demonstrated a more substantial overall enhancement compared to LLaVA. This can be attributed to InstructBLIP's specialized Instruction-aware Q-Former Module, which adeptly extracts the salient features of our emotion instructions and synergizes them effectively with the corresponding images, thereby yielding improved performance." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 434, + 526, + 445 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 434, + 526, + 445 + ], + "spans": [ + { + "bbox": [ + 306, + 434, + 526, + 445 + ], + "type": "text", + "content": "4.3. Effectiveness of Different Instruction Data" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 453, + 528, + 464 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 453, + 528, + 464 + ], + "spans": [ + { + "bbox": [ + 306, + 453, + 528, + 464 + ], + "type": "text", + "content": "4.3.1 Ablation Study of Different Instruction Data" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 474, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 474, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 474, + 545, + 713 + ], + "type": "text", + "content": "The ablation study outlined in Tab. 2 provides a comprehensive analysis of the impact that different instructional data types have on model performance, specifically concerning accuracy metrics on the EmoSet test set. Initially, the model, referred to as InstructBLIP [2], operates without the integration of the three types of instructional data and attains a baseline accuracy of " + }, + { + "bbox": [ + 304, + 474, + 545, + 713 + ], + "type": "inline_equation", + "content": "42.20\\%" + }, + { + "bbox": [ + 304, + 474, + 545, + 713 + ], + "type": "text", + "content": ". This foundational performance is significantly enhanced with the inclusion of Categorical data, which alone contributes to a substantial increase in accuracy. The introduction of Conversation data further amplifies this effect, underscoring the value of conversational context in improving the model's predictive capabilities. The addition of Reasoning data notably boosts performance, achieving a peak accuracy of " + }, + { + "bbox": [ + 304, + 474, + 545, + 713 + ], + "type": "inline_equation", + "content": "83.36\\%" + }, + { + "bbox": [ + 304, + 474, + 545, + 713 + ], + "type": "text", + "content": ". This indicates that the model significantly benefits from the nuanced cues in reasoning, aiding in understanding complex emotional instructions. The gradual improvements with each data type support the idea that a diverse approach to instructional data markedly enhances model comprehension and performance." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 46, + 683, + 286, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 683, + 286, + 712 + ], + "spans": [ + { + "bbox": [ + 46, + 683, + 286, + 712 + ], + "type": "text", + "content": "2Unlike the setup in InstructBLIP, our dataset exclusively comprises emotion-related content. Consequently, our held-out evaluation does not constitute a strict zero-shot evaluation in the conventional sense." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "26601" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 48, + 70, + 547, + 173 + ], + "blocks": [ + { + "bbox": [ + 48, + 70, + 547, + 173 + ], + "lines": [ + { + "bbox": [ + 48, + 70, + 547, + 173 + ], + "spans": [ + { + "bbox": [ + 48, + 70, + 547, + 173 + ], + "type": "table", + "html": "
MethodWebEmoFIEmotion6AbstractArtPhotoIAPSaEmotionROIEmoSet
Number of Classes258688868
Flanmingo [1]9.3614.9121.673.5717.510.1321.7229.59
LLaVA [8]12.5556.0449.4419.5436.2542.4346.4644.03
BLIP2 [5]20.1057.7250.0028.5736.2539.2450.5146.79
InstructBLIP [2]12.8037.9746.1121.4226.2534.1846.1342.20
Ours*21.1268.0957.8132.3444.9044.1353.8783.36
", + "image_path": "e2cb0d9137810c97b65195d666a489ec30f622606c4ebfa7ee8bc62e12385832.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 48, + 210, + 293, + 269 + ], + "blocks": [ + { + "bbox": [ + 159, + 180, + 432, + 191 + ], + "lines": [ + { + "bbox": [ + 159, + 180, + 432, + 191 + ], + "spans": [ + { + "bbox": [ + 159, + 180, + 432, + 191 + ], + "type": "text", + "content": "Table 1. Held-out performance comparison on visual emotion datasets " + }, + { + "bbox": [ + 159, + 180, + 432, + 191 + ], + "type": "inline_equation", + "content": "(\\%)" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 48, + 210, + 293, + 269 + ], + "lines": [ + { + "bbox": [ + 48, + 210, + 293, + 269 + ], + "spans": [ + { + "bbox": [ + 48, + 210, + 293, + 269 + ], + "type": "table", + "html": "
CategoricalConversationReasoningAccuracy (%)
---42.20
✓--80.90 (+38.70)
✓✓-81.95 (+39.75)
✓✓✓83.36 (+41.16)
", + "image_path": "4da102b85b62cd03c01a5fe65c5818e416fc9c3626968cbd5bf01f6aa70a21e3.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 46, + 277, + 287, + 300 + ], + "lines": [ + { + "bbox": [ + 46, + 277, + 287, + 300 + ], + "spans": [ + { + "bbox": [ + 46, + 277, + 287, + 300 + ], + "type": "text", + "content": "Table 2. Ablation study of three types of instruction data. Accuracy (%) on EmoSet test set." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 308, + 174, + 319 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 308, + 174, + 319 + ], + "spans": [ + { + "bbox": [ + 47, + 308, + 174, + 319 + ], + "type": "text", + "content": "4.3.2 Instruction Sensitivity" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 327, + 287, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 327, + 287, + 434 + ], + "spans": [ + { + "bbox": [ + 46, + 327, + 287, + 434 + ], + "type": "text", + "content": "This work is dedicated to the creation of a varied corpus of visual emotion instruction data, alongside the development of a robust instruction-based model. Our objective is for the model to demonstrate stability, producing consistent results in the face of minor variations in instruction phrasing, provided the core objective of the task persists unchanged. To this end, we employ the Sensitivity evaluation metric, as introduced by [22], to assess the model's fidelity in generating uniform outcomes irrespective of instructional nuances." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 435, + 287, + 589 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 435, + 287, + 589 + ], + "spans": [ + { + "bbox": [ + 46, + 435, + 287, + 589 + ], + "type": "text", + "content": "We employ two semantically similar instructions as input prompts for the model, testing their impact on the Sensitivity score across three visual emotion datasets for different Visual Instruction Tuning models. The first instruction is: \"From the given options: cls_1, cls_2, cls_3, etc., identify the emotion that most accurately reflects the image. Ensure your selection is confined to the listed options. Respond in the format: Predicted emotion.\" The second one states: \"Please choose the emotion that best corresponds to the image from the following options: cls_1, cls_2, cls_3, etc. (Do not provide answers beyond the provided candidates.) Please reply in the following format: Predict emotion.\"" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 590, + 287, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 590, + 287, + 662 + ], + "spans": [ + { + "bbox": [ + 46, + 590, + 287, + 662 + ], + "type": "text", + "content": "As illustrated in Fig. 7, our approach, along with BLIP2, exhibited exceptionally low Sensitivity values, demonstrating robustness in understanding the instructions. Conversely, Flamingo and InstructBLIP displayed a higher degree of sensitivity, indicating a relative susceptibility to variations in instruction wording." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 670, + 123, + 681 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 670, + 123, + 681 + ], + "spans": [ + { + "bbox": [ + 47, + 670, + 123, + 681 + ], + "type": "text", + "content": "4.4. Robustness" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 689, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 689, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 689, + 287, + 713 + ], + "type": "text", + "content": "Given that current emotion recognition datasets often exhibit category imbalances and labeling biases, our aim is" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 310, + 214, + 560, + 376 + ], + "blocks": [ + { + "bbox": [ + 310, + 214, + 560, + 376 + ], + "lines": [ + { + "bbox": [ + 310, + 214, + 560, + 376 + ], + "spans": [ + { + "bbox": [ + 310, + 214, + 560, + 376 + ], + "type": "image", + "image_path": "9310629969a8e41f19348f83511eba7915ed94c2d9fea133652a4458ae950dc1.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 312, + 390, + 549, + 401 + ], + "lines": [ + { + "bbox": [ + 312, + 390, + 549, + 401 + ], + "spans": [ + { + "bbox": [ + 312, + 390, + 549, + 401 + ], + "type": "text", + "content": "Figure 7. The sensitivity score comparison (the lower the better)." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 426, + 545, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 426, + 545, + 510 + ], + "spans": [ + { + "bbox": [ + 304, + 426, + 545, + 510 + ], + "type": "text", + "content": "to evaluate the generalization ability of various learning strategies more impartially. Hence, we selected the Un-BiasedEmo test set [12], which is uniquely suited for recognizing intricate emotions, such as those associated with identical objects or scenes, e.g., landscapes, crowds, families, babies, and animals, where the emotional undertones can be particularly subtle and complex." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 513, + 545, + 573 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 513, + 545, + 573 + ], + "spans": [ + { + "bbox": [ + 304, + 513, + 545, + 573 + ], + "type": "text", + "content": "As depicted in Tab. 3, our proposed methodology demonstrates superior performance when benchmarked against conventional supervised emotion recognition techniques, thereby underscoring the efficacy of our approach in more accurately discerning complex emotional contexts." + } + ] + } + ], + "index": 13 + }, + { + "type": "table", + "bbox": [ + 328, + 590, + 525, + 674 + ], + "blocks": [ + { + "bbox": [ + 328, + 590, + 525, + 674 + ], + "lines": [ + { + "bbox": [ + 328, + 590, + 525, + 674 + ], + "spans": [ + { + "bbox": [ + 328, + 590, + 525, + 674 + ], + "type": "table", + "html": "
MethodAccuracy (%)
Direct Learning [12]71.64
Self-Directed Learning [12]72.45
Joint Learning [12]71.64
Curriculum Learning [12]74.27
Ours*74.72
", + "image_path": "2ee8702a65ad3d1a4fb54341b8105000760c282855125751b118dec9f4c63f07.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "table_body" + } + ], + "index": 14 + }, + { + "bbox": [ + 317, + 683, + 534, + 694 + ], + "lines": [ + { + "bbox": [ + 317, + 683, + 534, + 694 + ], + "spans": [ + { + "bbox": [ + 317, + 683, + 534, + 694 + ], + "type": "text", + "content": "Table 3. Performance comparison on UnbiasedEmo dataset." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "26602" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 74, + 83, + 149, + 126 + ], + "blocks": [ + { + "bbox": [ + 74, + 83, + 149, + 126 + ], + "lines": [ + { + "bbox": [ + 74, + 83, + 149, + 126 + ], + "spans": [ + { + "bbox": [ + 74, + 83, + 149, + 126 + ], + "type": "image", + "image_path": "109bb11314538a214cd82aed3b6d0b5421a081e885cde55b680079b58077f669.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 197, + 83, + 258, + 129 + ], + "blocks": [ + { + "bbox": [ + 197, + 83, + 258, + 129 + ], + "lines": [ + { + "bbox": [ + 197, + 83, + 258, + 129 + ], + "spans": [ + { + "bbox": [ + 197, + 83, + 258, + 129 + ], + "type": "image", + "image_path": "053e3769043cbb64654c8c3e03aab6d5a184737816f4645df6871152ef833fd0.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 99, + 153, + 132, + 161 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 99, + 153, + 132, + 161 + ], + "spans": [ + { + "bbox": [ + 99, + 153, + 132, + 161 + ], + "type": "text", + "content": "BECAUSE" + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 151, + 135, + 193, + 176 + ], + "blocks": [ + { + "bbox": [ + 58, + 131, + 151, + 142 + ], + "lines": [ + { + "bbox": [ + 58, + 131, + 151, + 142 + ], + "spans": [ + { + "bbox": [ + 58, + 131, + 151, + 142 + ], + "type": "text", + "content": "Predicted emotion: anger" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 151, + 135, + 193, + 176 + ], + "lines": [ + { + "bbox": [ + 151, + 135, + 193, + 176 + ], + "spans": [ + { + "bbox": [ + 151, + 135, + 193, + 176 + ], + "type": "image", + "image_path": "b0abd87cf84e26d5614dc0f16917a8986399d1404f505f75f291a50d6a3b49a2.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 191, + 133, + 290, + 141 + ], + "lines": [ + { + "bbox": [ + 191, + 133, + 290, + 141 + ], + "spans": [ + { + "bbox": [ + 191, + 133, + 290, + 141 + ], + "type": "text", + "content": "Predicted emotion: sadness" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 211, + 155, + 244, + 162 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 211, + 155, + 244, + 162 + ], + "spans": [ + { + "bbox": [ + 211, + 155, + 244, + 162 + ], + "type": "text", + "content": "BECAUSE" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 178, + 167, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 178, + 167, + 213 + ], + "spans": [ + { + "bbox": [ + 56, + 178, + 167, + 213 + ], + "type": "text", + "content": "Reason: The image depicts a group of people with their arms up in the air, which is a common gesture associated with anger." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 213, + 170, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 213, + 170, + 239 + ], + "spans": [ + { + "bbox": [ + 56, + 213, + 170, + 239 + ], + "type": "text", + "content": "Additionally, the smoke billowing from the building in the background adds to the sense of chaos and unrest." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 177, + 178, + 291, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 177, + 178, + 291, + 204 + ], + "spans": [ + { + "bbox": [ + 177, + 178, + 291, + 204 + ], + "type": "text", + "content": "Reason: The image is a black and white photograph of an old, abandoned house in the middle of a field." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 178, + 213, + 291, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 178, + 213, + 291, + 239 + ], + "spans": [ + { + "bbox": [ + 178, + 213, + 291, + 239 + ], + "type": "text", + "content": "The scene is desolate and eerie, which suggests a sense of sadness or melancholy." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 277, + 165, + 290 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 277, + 165, + 290 + ], + "spans": [ + { + "bbox": [ + 47, + 277, + 165, + 290 + ], + "type": "text", + "content": "4.4.1 Affective Reasoning" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 46, + 297, + 287, + 367 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 297, + 287, + 367 + ], + "spans": [ + { + "bbox": [ + 46, + 297, + 287, + 367 + ], + "type": "text", + "content": "In the domain of visual emotion recognition, where ambiguity and subjectivity are pervasive, the advent of an interpretable model is of considerable value. Such a model elucidates its cognitive processes, enhancing its trustworthiness and practicality in scenarios requiring a delicate grasp of emotional subtleties." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 46, + 369, + 287, + 427 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 369, + 287, + 427 + ], + "spans": [ + { + "bbox": [ + 46, + 369, + 287, + 427 + ], + "type": "text", + "content": "Leveraging Visual Instruction Tuning, our model transcends mere categorization of emotions; it articulates the underlying rationale for its classifications. The executing commands for identifying emotions and elucidating the decision basis is illustrated below:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 83, + 438, + 257, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 83, + 438, + 257, + 449 + ], + "spans": [ + { + "bbox": [ + 83, + 438, + 257, + 449 + ], + "type": "text", + "content": "Predicted emotion: [emotion]." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 83, + 450, + 216, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 83, + 450, + 216, + 460 + ], + "spans": [ + { + "bbox": [ + 83, + 450, + 216, + 460 + ], + "type": "text", + "content": "Reason: [explanation]." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 46, + 471, + 287, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 471, + 287, + 506 + ], + "spans": [ + { + "bbox": [ + 46, + 471, + 287, + 506 + ], + "type": "text", + "content": "Our model delineates the visual features influencing its determinations, thereby addressing the complexities inherent in discerning and explaining emotion-related nuances." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 46, + 507, + 287, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 507, + 287, + 567 + ], + "spans": [ + { + "bbox": [ + 46, + 507, + 287, + 567 + ], + "type": "text", + "content": "The explanations provide us with visual clues contained within the images, as exemplified in Fig. 8. It provides interpretable visual indicators that inform the model's outputs, as demonstrated in our example, by disambiguating the often abstract emotional categories." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 47, + 573, + 127, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 573, + 127, + 586 + ], + "spans": [ + { + "bbox": [ + 47, + 573, + 127, + 586 + ], + "type": "text", + "content": "4.5. Scaling Law" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 46, + 592, + 287, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 592, + 287, + 652 + ], + "spans": [ + { + "bbox": [ + 46, + 592, + 287, + 652 + ], + "type": "text", + "content": "Pretraining data. As demonstrated in Tab. 4, there is a clear correlation between the size of the pre-training dataset and improved performance. Consequently, we anticipate that an increase in training data in the future could enhance the effectiveness of Emotion Visual Instruction Tuning." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 47, + 659, + 206, + 672 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 659, + 206, + 672 + ], + "spans": [ + { + "bbox": [ + 47, + 659, + 206, + 672 + ], + "type": "text", + "content": "4.6. Humour Caption Generation" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "text", + "content": "The comprehension of humor is intricately linked to the understanding of emotions. Leveraging our generative language model, we conduct a caption generation task without" + } + ] + } + ], + "index": 22 + }, + { + "type": "table", + "bbox": [ + 360, + 70, + 492, + 95 + ], + "blocks": [ + { + "bbox": [ + 360, + 70, + 492, + 95 + ], + "lines": [ + { + "bbox": [ + 360, + 70, + 492, + 95 + ], + "spans": [ + { + "bbox": [ + 360, + 70, + 492, + 95 + ], + "type": "table", + "html": "
5%10%30%50%
79.0081.0079.3483.36
", + "image_path": "298fb83e44b71d807ad55c49350025bb3a95c37916be87523ba01a98e6609571.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "table_body" + } + ], + "index": 23 + }, + { + "bbox": [ + 306, + 103, + 544, + 126 + ], + "lines": [ + { + "bbox": [ + 306, + 103, + 544, + 126 + ], + "spans": [ + { + "bbox": [ + 306, + 103, + 544, + 126 + ], + "type": "text", + "content": "Table 4. Ablation study of different portion of pre-training data. Accuracy " + }, + { + "bbox": [ + 306, + 103, + 544, + 126 + ], + "type": "inline_equation", + "content": "(\\%)" + }, + { + "bbox": [ + 306, + 103, + 544, + 126 + ], + "type": "text", + "content": " on EmoSet test set." + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 321, + 138, + 542, + 300 + ], + "blocks": [ + { + "bbox": [ + 80, + 264, + 265, + 275 + ], + "lines": [ + { + "bbox": [ + 80, + 264, + 265, + 275 + ], + "spans": [ + { + "bbox": [ + 80, + 264, + 265, + 275 + ], + "type": "text", + "content": "Figure 8. The sample of our generated explanation." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 321, + 138, + 542, + 300 + ], + "lines": [ + { + "bbox": [ + 321, + 138, + 542, + 300 + ], + "spans": [ + { + "bbox": [ + 321, + 138, + 542, + 300 + ], + "type": "image", + "image_path": "eb9cb349f1ef4b50b1a208cf1b5bd1b0aa7d233078c483d280120546ab7f8105.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 308, + 556, + 331 + ], + "lines": [ + { + "bbox": [ + 305, + 308, + 556, + 331 + ], + "spans": [ + { + "bbox": [ + 305, + 308, + 556, + 331 + ], + "type": "text", + "content": "Figure 9. The sample of our generated humour caption vs human writing humour caption from OxfordTVG-HIC." + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_caption" + } + ], + "index": 25 + }, + { + "bbox": [ + 304, + 333, + 545, + 465 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 333, + 545, + 465 + ], + "spans": [ + { + "bbox": [ + 304, + 333, + 545, + 465 + ], + "type": "text", + "content": "modifying the model's architecture, specifically testing the model's proficiency in generating humorous captions. For this purpose, we select 50 images from the OxfordTVGHIC dataset [6] and generate corresponding captions using our model. Subsequently, the captions produced by our model are compared with manually annotated captions from the dataset in a user study. Thirty participants were asked to vote on which captions were more humorous. Our model-generated captions receive " + }, + { + "bbox": [ + 304, + 333, + 545, + 465 + ], + "type": "inline_equation", + "content": "60\\%" + }, + { + "bbox": [ + 304, + 333, + 545, + 465 + ], + "type": "text", + "content": " of the votes, demonstrating its effective humor generation capabilities. One sample is visualized in Fig. 9." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 306, + 476, + 378, + 488 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 476, + 378, + 488 + ], + "spans": [ + { + "bbox": [ + 306, + 476, + 378, + 488 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 304, + 496, + 545, + 627 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 496, + 545, + 627 + ], + "spans": [ + { + "bbox": [ + 304, + 496, + 545, + 627 + ], + "type": "text", + "content": "In our study, drawing upon the distinctive visual cues key to visual emotion recognition, we present a GPT-assisted pipeline specifically designed for generating emotion visual instruction data. The developed EmoVIT model incorporates emotion-specific instructions, leveraging LLMs for enhanced performance. Our comprehensive experiments validate its effectiveness in emotion classification, affective reasoning, and humor understanding. This comparative analysis sets a benchmark for Emotion Visual Instruction Tuning with LLMs, providing valuable insights and directions for future research in this field." + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 306, + 639, + 403, + 651 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 639, + 403, + 651 + ], + "spans": [ + { + "bbox": [ + 306, + 639, + 403, + 651 + ], + "type": "text", + "content": "Acknowledgement" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 305, + 657, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 657, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 305, + 657, + 545, + 712 + ], + "type": "text", + "content": "This work is partially supported by the National Science and Technology Council, Taiwan under Grant, NSTC-112-2628-E-002-033-MY4, NSTC-112-2634-F-002-002-MBK, NSTC-112-26234-F-002-005, NSTC-112-2221-E-A49-059-MY3, and NSTC-112-2221-E-A49-094-MY3." + } + ] + } + ], + "index": 31 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "26603" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 156 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 156 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 156 + ], + "type": "text", + "content": "[1] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. In Advances in Neural Information Processing Systems, pages 23716-23736, 2022." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 158, + 288, + 212 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 158, + 288, + 212 + ], + "spans": [ + { + "bbox": [ + 53, + 158, + 288, + 212 + ], + "type": "text", + "content": "[2] Wenliang Dai, Junnan Li, Dongxu Li, Anthony Meng Huat Tiong, Junqi Zhao, Weisheng Wang, Boyang Li, Pascale Fung, and Steven Hoi. Instructclip: Towards general-purpose vision-language models with instruction tuning. In Advances in Neural Information Processing Systems, 2023." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 213, + 286, + 256 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 213, + 286, + 256 + ], + "spans": [ + { + "bbox": [ + 53, + 213, + 286, + 256 + ], + "type": "text", + "content": "[3] Menglin Jia, Luming Tang, Bor-Chun Chen, Claire Cardie, Serge Belongie, Bharath Hariharan, and Ser-Nam Lim. Visual prompt tuning. In Proceedings of the European Conference on Computer Vision, pages 709-727. Springer, 2022." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 258, + 286, + 290 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 258, + 286, + 290 + ], + "spans": [ + { + "bbox": [ + 53, + 258, + 286, + 290 + ], + "type": "text", + "content": "[4] Dongxu Li, Junnan Li, Hung Le, Guangsen Wang, Silvio Savarese, and Steven CH Hoi. Lavis: A library for language-vision intelligence. arXiv preprint arXiv:2209.09019, 2022." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 291, + 286, + 345 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 291, + 286, + 345 + ], + "spans": [ + { + "bbox": [ + 53, + 291, + 286, + 345 + ], + "type": "text", + "content": "[5] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In Proceedings of the International Conference on Machine Learning, 2023." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 346, + 286, + 389 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 346, + 286, + 389 + ], + "spans": [ + { + "bbox": [ + 53, + 346, + 286, + 389 + ], + "type": "text", + "content": "[6] Runjia Li, Shuyang Sun, Mohamed Elhoseiny, and Philip Torr. Oxfordtvg-hic: Can machine make humorous captions from images? In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 20293-20303, 2023." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 391, + 286, + 434 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 391, + 286, + 434 + ], + "spans": [ + { + "bbox": [ + 53, + 391, + 286, + 434 + ], + "type": "text", + "content": "[7] Zheng Lian, Licai Sun, Mingyu Xu, Haiyang Sun, Ke Xu, Zhuofan Wen, Shun Chen, Bin Liu, and Jianhua Tao. Explainable multimodal emotion reasoning. arXiv preprint arXiv:2306.15401, 2023." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 436, + 286, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 436, + 286, + 468 + ], + "spans": [ + { + "bbox": [ + 53, + 436, + 286, + 468 + ], + "type": "text", + "content": "[8] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. In Advances in Neural Information Processing Systems, 2023." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 469, + 286, + 512 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 469, + 286, + 512 + ], + "spans": [ + { + "bbox": [ + 53, + 469, + 286, + 512 + ], + "type": "text", + "content": "[9] Jana Machajdik and Allan Hanbury. Affective image classification using features inspired by psychology and art theory. In Proceedings of the 18th ACM international conference on Multimedia, pages 83-92, 2010." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 514, + 286, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 514, + 286, + 567 + ], + "spans": [ + { + "bbox": [ + 48, + 514, + 286, + 567 + ], + "type": "text", + "content": "[10] Joseph A Mikels, Barbara L Fredrickson, Gregory R Larkin, Casey M Lindberg, Sam J Maglio, and Patricia A Reuter-Lorenz. Emotional category data on images from the international affective picture system. Behavior research methods, 37:626-630, 2005." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 569, + 272, + 580 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 569, + 272, + 580 + ], + "spans": [ + { + "bbox": [ + 48, + 569, + 272, + 580 + ], + "type": "text", + "content": "[11] OpenAI. Gpt-4 technical report. Technical report, 2023." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 582, + 286, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 582, + 286, + 635 + ], + "spans": [ + { + "bbox": [ + 48, + 582, + 286, + 635 + ], + "type": "text", + "content": "[12] Rameswar Panda, Jianming Zhang, Haoxiang Li, Joon-Young Lee, Xin Lu, and Amit K Roy-Chowdhury. Contemplating visual emotions: Understanding and overcoming dataset bias. In Proceedings of the European Conference on Computer Vision, pages 579–595, 2018." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 636, + 286, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 636, + 286, + 657 + ], + "spans": [ + { + "bbox": [ + 48, + 636, + 286, + 657 + ], + "type": "text", + "content": "[13] EU Parliament. Eu ai act: first regulation on artificial intelligence. Accessed June, 25:2023, 2023." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 658, + 286, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 286, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 286, + 712 + ], + "type": "text", + "content": "[14] Kuan-Chuan Peng, Tsuhan Chen, Amir Sadovnik, and Andrew C Gallagher. A mixed bag of emotions: Model, predict, and transfer emotion distributions. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 860-868, 2015." + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 126 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 126 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 126 + ], + "type": "text", + "content": "[15] Kuan-Chuan Peng, Amir Sadovnik, Andrew Gallagher, and Tsuhan Chen. Where do emotions come from? predicting the emotion stimuli map. In 2016 IEEE international conference on image processing (ICIP), pages 614-618. IEEE, 2016." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 129, + 545, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 129, + 545, + 183 + ], + "spans": [ + { + "bbox": [ + 307, + 129, + 545, + 183 + ], + "type": "text", + "content": "[16] Dongyu She, Jufeng Yang, Ming-Ming Cheng, Yu-Kun Lai, Paul L Rosin, and Liang Wang. Wscnet: Weakly supervised coupled networks for visual sentiment classification and detection. IEEE Transactions on Multimedia, 22(5): 1358-1371, 2019." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 185, + 545, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 185, + 545, + 239 + ], + "spans": [ + { + "bbox": [ + 307, + 185, + 545, + 239 + ], + "type": "text", + "content": "[17] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 241, + 545, + 295 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 241, + 545, + 295 + ], + "spans": [ + { + "bbox": [ + 307, + 241, + 545, + 295 + ], + "type": "text", + "content": "[18] Jason Wei, Maarten Bosma, Vincent Zhao, Kelvin Guu, Adams Wei Yu, Brian Lester, Nan Du, Andrew M Dai, and Quoc V Le. Finetuned language models are zero-shot learners. In International Conference on Learning Representations." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 297, + 545, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 297, + 545, + 352 + ], + "spans": [ + { + "bbox": [ + 307, + 297, + 545, + 352 + ], + "type": "text", + "content": "[19] Hongxia Xie, Hua Chung, Hong-Han Shuai, and Wen-Huang Cheng. Learning to prompt for vision-language emotion recognition. In 2023 11th International Conference on Affective Computing and Intelligent Interaction Workshops and Demos (ACIIW), pages 1-4, 2023." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 354, + 545, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 354, + 545, + 418 + ], + "spans": [ + { + "bbox": [ + 307, + 354, + 545, + 418 + ], + "type": "text", + "content": "[20] Hongxia Xie, Ming-Xian Lee, Tzu-Jui Chen, Hung-Jen Chen, Hou-I Liu, Hong-Han Shuai, and Wen-Huang Cheng. Most important person-guided dual-branch cross-patch attention for group affect recognition. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 20598–20608, 2023." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 421, + 545, + 474 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 421, + 545, + 474 + ], + "spans": [ + { + "bbox": [ + 307, + 421, + 545, + 474 + ], + "type": "text", + "content": "[21] Liwen Xu, Zhengtao Wang, Bin Wu, and Simon Lui. Mdan: Multi-level dependent attention network for visual emotion analysis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9479-9488, 2022." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 476, + 545, + 509 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 476, + 545, + 509 + ], + "spans": [ + { + "bbox": [ + 307, + 476, + 545, + 509 + ], + "type": "text", + "content": "[22] Zhiyang Xu, Ying Shen, and Lifu Huang. Multiinstruct: Improving multi-modal zero-shot learning via instruction tuning. arXiv preprint arXiv:2212.10773, 2022." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 510, + 545, + 543 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 510, + 545, + 543 + ], + "spans": [ + { + "bbox": [ + 307, + 510, + 545, + 543 + ], + "type": "text", + "content": "[23] Jingyuan Yang, Jie Li, Xiumei Wang, Yuxuan Ding, and Xinbo Gao. Stimuli-aware visual emotion analysis. IEEE Transactions on Image Processing, 30:7432-7445, 2021." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 544, + 545, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 544, + 545, + 599 + ], + "spans": [ + { + "bbox": [ + 307, + 544, + 545, + 599 + ], + "type": "text", + "content": "[24] Jingyuan Yang, Qirui Huang, Tingting Ding, Dani Lischinski, Danny Cohen-Or, and Hui Huang. Emoset: A large-scale visual emotion dataset with rich attributes. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 20383-20394, 2023." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 601, + 545, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 601, + 545, + 644 + ], + "spans": [ + { + "bbox": [ + 307, + 601, + 545, + 644 + ], + "type": "text", + "content": "[25] Quanzeng You, Jiebo Luo, Hailin Jin, and Jianchao Yang. Building a large scale dataset for image emotion recognition: The fine print and the benchmark. In Proceedings of the AAAI conference on artificial intelligence, 2016." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 647, + 545, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 647, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 307, + 647, + 545, + 689 + ], + "type": "text", + "content": "[26] Shengyu Zhang, Linfeng Dong, Xiaoya Li, Sen Zhang, Xiaofei Sun, Shuhe Wang, Jiwei Li, Runyi Hu, Tianwei Zhang, Fei Wu, et al. Instruction tuning for large language models: A survey. arXiv preprint arXiv:2308.10792, 2023." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 691, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 691, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 307, + 691, + 545, + 712 + ], + "type": "text", + "content": "[27] Wei Zhang, Xuanyu He, and Weizhi Lu. Exploring discriminative representations for image emotion recognition with" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "26604" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 286, + 151 + ], + "type": "list", + "angle": 0, + "index": 2, + "blocks": [ + { + "bbox": [ + 67, + 72, + 286, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 72, + 286, + 94 + ], + "spans": [ + { + "bbox": [ + 67, + 72, + 286, + 94 + ], + "type": "text", + "content": "cnns. IEEE Transactions on Multimedia, 22(2):515-523, 2019." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 96, + 286, + 151 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 96, + 286, + 151 + ], + "spans": [ + { + "bbox": [ + 48, + 96, + 286, + 151 + ], + "type": "text", + "content": "[28] Sicheng Zhao, Zizhou Jia, Hui Chen, Leida Li, Guiguang Ding, and Kurt Keutzer. Pd Janet: Polarity-consistent deep attention network for fine-grained visual emotion regression. In Proceedings of the 27th ACM International Conference on Multimedia, pages 192-201, 2019." + } + ] + } + ], + "index": 1 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "26605" + } + ] + } + ], + "index": 3 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Emotional Speech-driven 3D Body Animation via Disentangled Latent Diffusion/8fe0dd5b-5118-4e0a-ac09-f70da9816cdc_content_list.json b/2024/Emotional Speech-driven 3D Body Animation via Disentangled Latent Diffusion/8fe0dd5b-5118-4e0a-ac09-f70da9816cdc_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..ced1e56b1f4a588cf8c230bd8dbc2b0e7d689aef --- /dev/null +++ b/2024/Emotional Speech-driven 3D Body Animation via Disentangled Latent Diffusion/8fe0dd5b-5118-4e0a-ac09-f70da9816cdc_content_list.json @@ -0,0 +1,1605 @@ +[ + { + "type": "text", + "text": "Emotional Speech-driven 3D Body Animation via Disentangled Latent Diffusion", + "text_level": 1, + "bbox": [ + 81, + 130, + 887, + 151 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Kiran Chhatre", + "bbox": [ + 250, + 179, + 372, + 196 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Radek Danček2", + "bbox": [ + 408, + 180, + 542, + 196 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Nikos Athanasiou2", + "bbox": [ + 578, + 180, + 728, + 196 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Giorgio Becherini²", + "bbox": [ + 153, + 198, + 305, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Christopher Peters1", + "bbox": [ + 341, + 198, + 495, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Michael J. Black", + "bbox": [ + 532, + 198, + 674, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Timo Bolkart $^{2*}$", + "bbox": [ + 710, + 198, + 833, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ KTH Royal Institute of Technology, Sweden $^{2}$ Max Planck Institute for Intelligent Systems, Germany", + "bbox": [ + 89, + 215, + 887, + 234 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 268, + 312, + 284 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Existing methods for synthesizing 3D human gestures from speech have shown promising results, but they do not explicitly model the impact of emotions on the generated gestures. Instead, these methods directly output animations from speech without control over the expressed emotion. To address this limitation, we present AMUSE, an emotional speech-driven body animation model based on latent diffusion. Our observation is that content (i.e., gestures related to speech rhythm and word utterances), emotion, and personal style are separable. To account for this, AMUSE maps the driving audio to three disentangled latent vectors: one for content, one for emotion, and one for personal style. A latent diffusion model, trained to generate gesture motion sequences, is then conditioned on these latent vectors. Once trained, AMUSE synthesizes 3D human gestures directly from speech with control over the expressed emotions and style by combining the content from the driving speech with the emotion and style of another speech sequence. Randomly sampling the noise of the diffusion model further generates variations of the gesture with the same emotional expressivity. Qualitative, quantitative, and perceptual evaluations demonstrate that AMUSE outputs realistic gesture sequences. Compared to the state of the art, the generated gestures are better synchronized with the speech content, and better represent the emotion expressed by the input speech. Our code is available at amuse.is.tue.mpg.de.", + "bbox": [ + 73, + 301, + 473, + 694 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 726, + 207, + 739 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Animating 3D bodies from speech has a wide range of applications, such as telepresence in AR/VR, avatar animation in games and movies, and to embody interactive digital assistants. While methods for speech-driven 3D body animation have recently shown great progress [5, 7, 31, 56, 101], existing methods do not adequately address one crucial factor: the impact of emotion from the driving speech signal on the generated gestures. Emotions and their expressions play", + "bbox": [ + 75, + 752, + 468, + 875 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/4414e5a2abd995e7a359daea2dc2751ecdd1f332f098d2c2ca660461c2179747.jpg", + "image_caption": [ + "... choose a major that is easy to find a good job in the future..." + ], + "image_footnote": [], + "bbox": [ + 527, + 268, + 864, + 388 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/44e4e54140317ad646e9f8780ec5ed0aebb5535002a5208ee58ef257be33ee5d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 527, + 405, + 861, + 473 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/c127b0d6bd325b8ed5eee9e914bcb897ac31ee4e9c89592aeb40f1f0d40ef7c2.jpg", + "image_caption": [ + "Figure 1. Goal. AMUSE generates realistic emotional 3D body gestures directly from a speech sequence (top). It provides user control over the generated emotion by combining the driving speech sequence with a different emotional audio (bottom)." + ], + "image_footnote": [], + "bbox": [ + 503, + 474, + 887, + 571 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "a fundamental role in human communication [29, 35, 65] and have become an important consideration when designing computer systems that interact with humans in a natural manner [78, 79]. They are of central concern when synthesizing human animations for a wide variety of application contexts, such as Socially Interactive Agents [61]. Because of this, speech-driven animation systems must not only align movement with the rhythm of the speech, but should also be capable of generating gestures that are perceived as expressing the suitable emotion.", + "bbox": [ + 496, + 671, + 890, + 821 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Many factors contribute to the perception of emotion and personal idiosyncrasies, such as facial expressions [19], gaze and eye contact [42], physiological responses [47], tone of voice [87], body language [66], and gestures [39]. When it comes to 3D animation, the most relevant factors", + "bbox": [ + 496, + 824, + 890, + 898 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 807, + 44 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Now at Google.", + "bbox": [ + 94, + 887, + 187, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "1942", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "are facial expressions, gestures, and body language [95]. While emotional speech-driven animation methods have recently been proposed for 3D faces [18, 74, 90, 107], animating emotional bodies from speech remains under-explored.", + "bbox": [ + 75, + 90, + 467, + 151 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Generating gestures solely from speech with emotional control is a difficult task. First, the mapping from audio to body motion is a non-deterministic many-to-many mapping, which is difficult to model. Gestures across subjects can vary when uttering the same sentence, and a single individual's motions can change significantly across repetitions. Second, factoring out the impact of emotional state on the body motion from other, unknown factors, is difficult. This requires disentangling the effects of three different factors on the generated motion, namely content-based (i.e., gestures related to speech rhythm and word utterances), emotion-based, and those based on personal style. AMUSE addresses this by separating a speech sequence into content, emotion, and style latent vectors, which are then used to condition a latent diffusion model. Specifically, AMUSE consists of three main components: (1) an audio autoencoder trained to produce disentangled vectors of content, emotion, and style, (2) a 3D body motion prior in the form of a temporal variational autoencoder (VAE) to generate smooth and realistic gestures, and (3) a latent diffusion model, which generates 3D body motion given the input content, emotion, and style latent vectors.", + "bbox": [ + 75, + 152, + 468, + 484 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Training such a model requires a speech-to-3D body dataset of sufficient scale, which is rich and diverse in speakers and emotions. BEAT [55] is a good candidate because it provides a large set of 3D gestures associated with single-person monologues. Unfortunately, the bodies are represented as skeletons, and it lacks face mocap markers and FLAME expressions. Instead, to produce realistic body animations, we require articulated 3D body surfaces. To overcome this, we convert BEAT sequences to SMPL-X [73] format using MoSh++ [62] and use the SMPL-X parameters for training. See [56] for comparison.", + "bbox": [ + 75, + 484, + 467, + 651 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our contributions are: (1) We present a framework to synthesize emotional 3D body gestures directly from speech. (2) We factor an input audio into disentangled content, emotion and style vectors, which enables us to separately control emotion in generated gestures. (3) We adapt temporal latent diffusion for multiple target conditions.", + "bbox": [ + 75, + 652, + 467, + 742 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 76, + 758, + 217, + 773 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. 3D Conditional Human Motion Generation", + "text_level": 1, + "bbox": [ + 76, + 785, + 442, + 800 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Early works focus mostly on predicting [10, 16, 33, 41, 57, 64, 70, 86, 106, 109] or generating human motion [30, 49], but do not consider multi-modal control. Recently, conditional motion generation through other modalities, such as text [2, 8, 9, 17, 22, 28, 77], music [50, 68, 94], speech [32], or action labels [27, 75], has gained more attention. Be", + "bbox": [ + 75, + 809, + 467, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "low, we focus on speech-driven motion generation methods, since they are the most relevant to our work.", + "bbox": [ + 498, + 90, + 888, + 119 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Gesture Generation from Speech", + "text_level": 1, + "bbox": [ + 500, + 133, + 787, + 150 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Rule-based gesture synthesis. Embodied conversational agents (ECA) are designed to interact and communicate with humans. Using the Behavior Markup Language (BML) [44] one can build rule-based systems for humanoids based on predefined behaviors [80]. This is used for completion of a storytelling task in an expressive manner [45]. The BEAT rule-based toolkit [14] enables adding non-verbal behavior on top of a pre-animated figure. Thiebaux et al. [92] develop an ECA by using procedural animation techniques and keyframe interpolation. Marsella et al. [63] design a generalized rule-based agent to generate expressions, eye gaze, and gestures from speech. Each of these approaches are based on non-trainable, rule-based techniques that may require substantial manual modelling effort to adapt to new tasks.", + "bbox": [ + 496, + 159, + 890, + 385 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Data-driven gesture synthesis. More recently, data-driven methods have superseded rule-based systems. Yoon et al. [104] use a fusion of text, audio and upper body gestures to learn an upper body gesture avatar, but can only control the style of individual speakers by sampling from their latent space. SpeechGestureMatching [32] generates 3D facial meshes and 3D keypoints of the body and hands from speech, but the outputs are separated and the method does not provide control over the generations. QPGesture [98] uses phase to better align the generated 3D skeleton-based gesturing avatars with the audio input. Ginosar et al. [23] and Diverse-3D-Hand-Gesture-Prediction [82] generate hand and arm motions only. Audio2Gestures [48] encode motion and audio to a low-dimensional latent space and generate gestures. SEEG [53] aims to generate gestures that align well with the semantics of the speech. DiffTTSG [67] regresses speech and gestures at the same time, joining the two modalities in a single system. DiffGAN [3] retargets gestures across speakers in a low-resource setting. The GENEA challenge [105] tackles gesticulation from speech alone using the Talking-with-Hands dataset [46]. Gesture2Vec [99] uses a machine translation model to translate text into gesture chunks and output full sequences using such quantized representations. TalkSHOW [101] uses a VQ-VAE to generate 3D human bodies gesturing with facial expressions from speech segments, but in an uncontrolled manner. Similarly, Co-speech gesture [60] uses an RQ-VAE to generate different gestures from speech. Alternative gesture generation from speech methods have been proposed such as reinforcement learning [91], self-supervised pretraining [40], and diffusion [67, 110]. BodyFormer [71] introduces a dataset of pseudo-groundtruth and a transformer-based method for generating gestures from speech. However, none of these methods provide explicit emotional con", + "bbox": [ + 496, + 387, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "1943", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "trol over the generated motion.", + "bbox": [ + 76, + 90, + 282, + 104 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "For controllable generation, GestureDiffuCLIP [7] incorporates multiple conditions including CLIP [83] text features, video, or motion prompts via AdaIn [37] layers to generate gestures from speech, however, it does not allow explicit control over the emotion conveyed by the driving audio. ListenDenoiseAction [5] combines conformers and the DiffWave [43] architecture to generate gestures that can be controlled by a style vector, RhythmicGesticulator [6] disentangles the latent space into a vector related to the semantics of the gesture and one related to the subtle variations, while DisCo [54] models content and rhythm. StyleGestures [4] adapts MoGlow [34], demonstrating limited control over some motion attributes like the speed and expressiveness of gestures. DiffuseStyleGesture [97] uses diffusion to generate diverse gestures from speech.", + "bbox": [ + 75, + 107, + 470, + 333 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.3. Emotion Control", + "text_level": 1, + "bbox": [ + 76, + 347, + 243, + 361 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Emotion classification and control has been little studied in 3D human motion generation with only a few methods using skeletal motion in multi-class classification. Ghaleb et al. [20] employ a spatio-temporal graph convolution network to classify gestures into four classes: preparation, stroke, retraction, and neutral. Li et al. [52], on the other hand, use hidden Markov models for emotion classification of human movement mocap data. Karras et al. [38] learn face animations of a single actor, and test their method on different tasks by modifying the latent vectors. However, there is no disentanglement mechanism, and they do not model the synchronization of the emotion with the facial motions. Recently, EmoTalk [74], animates emotional 3D faces from speech input with control over the emotion intensity and EMOTE [18] disentangles emotion and speech to allow emotion editing at test time. However, models solely intended for facial tasks like lipSyncing and capturing expressions might not smoothly adapt to the complexity of whole-body movements and distinct articulation. Regarding emotion-conditioned motion generation, Aberman et al. [1] show style-transfer from video data to motion and provide some style-based control, but do not address speech-driven emotional gestures. Similarly, the ZeroEGGs [21] dataset contains some emotional gesture controls but also includes more generic styles of motion. The method requires the input of arbitrary frames of desired motion to encode a style, thereby relying on motions and speech as conditions during inference. Text-driven emotional gesticulation, as explored by Bhattacharya et al. [11, 12], emphasizes the generation of gestures based on textual cues, incorporating additional conditions such as speech, speaker ID, seed poses, as well as valence, arousal, and dominance triplets. However, these approaches do not provide the means to distill explicit emotion features, limiting free control over the generated gestures. Closer to", + "bbox": [ + 75, + 371, + 472, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "our work, EMoG [102] incorporates emotion cues from the BEAT dataset [55] to generate improved gesture quality without explicit emotion control. EmotionGesture [81] uses a TED Emotion Dataset and BEAT to incorporate emotion features in gesture generation and generate emotional gestures. Although they can generate emotional gestures, their method is not end-to-end and has no explicit motion control. Specifically, it uses an emotion-conditioned VAE after training to acquire diverse emotion features that are used to generate gestures without guarantees and control over emotion types. Wu et al. [96] introduce the first multi-cultural gesture dataset containing 200 individuals of 10 different cultures. In contrast to prior work, we explicitly control the emotions conveyed by the generated gestures solely through emotional speech without relying on additional conditions.", + "bbox": [ + 496, + 90, + 893, + 319 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 500, + 329, + 591, + 345 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The AMUSE pipeline consists of two separately trained networks. The audio disentanglement module, which encodes input speech into latent vectors for content, emotion, and style is described in Sec. 3.2. The main architecture is described in Sec. 3.3. It consists of a 3D human motion prior coupled with a latent diffusion model. It takes random noise (or partially denoised latent vectors) on the input and outputs a human motion sequence. We introduce broader applications in gesture editing in Sec. 3.4.", + "bbox": [ + 496, + 354, + 893, + 491 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Preliminary: Expressive 3D Body Model", + "text_level": 1, + "bbox": [ + 498, + 498, + 848, + 513 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "SMPL-X [73] is a 3D model of the body surface. SMPL-X is defined as function $M(\\boldsymbol{\\beta}, \\boldsymbol{\\theta}, \\boldsymbol{\\psi})$ that produces a 3D body mesh. It is parameterized by identity shape $\\boldsymbol{\\beta} \\in \\mathbb{R}^{300}$ , pose $\\boldsymbol{\\theta} \\in \\mathbb{R}^{J \\times 3}$ including finger articulation for rotations around $J$ joints, and facial expression $\\boldsymbol{\\psi} \\in \\mathbb{R}^{100}$ . We adopt the continuous 6D rotation representation for training following Zhou et al. [108], making $\\boldsymbol{\\theta} \\in \\mathbb{R}^{J \\times 6}$ . Given pose parameters and any shape parameter, we can obtain body mesh vertices $V$ using the differentiable SMPL-X layer [73]. As the focus of our paper is on synthesizing body gestures and not locomotion, we disregard 8 joints that correspond those of the lower body joint poses, leaving $J = 47$ . Further, we omit the facial expression parameters, i.e., set $\\boldsymbol{\\psi} = \\mathbf{0}$ .", + "bbox": [ + 496, + 521, + 893, + 719 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Speech Disentanglement Model", + "text_level": 1, + "bbox": [ + 498, + 726, + 777, + 742 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Architecture. The goal of the this model is to factor an input speech into three disentangled latent representations, one for content (i.e., the words spoken), one for emotion, and one for personal style. To do so, we devise a specialized encoder-decoder architecture with three separate encoders, one for each latent space. We denote the encoders as: $E_{c}(a) = c$ , $E_{e}(a) = e$ , $E_{s}(a) = s$ , where $a$ is the input filterbank, $c$ , $e$ and $s$ denote the latent vectors for content, emotion and style and $E_{c}$ , $E_{e}$ and $E_{s}$ are their encoders. The architecture of the three encoders follows the", + "bbox": [ + 496, + 750, + 893, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "1944", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/f5a86adb69535c00ac5d1290edf434c200ec423010af30604e43ba061845a680.jpg", + "image_caption": [ + "Figure 2. Training. We train the motion prior $(\\mathcal{P}_E, \\mathcal{P}_D)$ and the latent denoiser $\\Delta$ jointly, while keeping the audio encoding networks frozen. In the forward pass, we take an input audio $a^{1:T}$ and pose sequence $m^{1:T}$ . Firstly, we do a forward pass of $m^{1:T}$ through $\\mathcal{P}_E$ and $\\mathcal{P}_D$ and compute $\\mathcal{L}_{rec}$ , $\\mathcal{L}_{Vrec}$ , and $\\mathcal{L}_{KL}$ . Then, we apply the diffusion process to a gradient-detached sg $[z_m]$ obtaining the noisy $z_m^{(D)}$ which is then denoised with $\\Delta$ and $\\mathcal{L}_{LD}$ is computed. Finally, we use $\\Delta$ to fully denoise $z_n$ into gradient-detached sg $[z_{\\tilde{m}}]$ , further decode $\\tilde{m}^{1:T}$ using $\\mathcal{P}_D$ , and compute $\\mathcal{L}_{align}$ and $\\mathcal{L}_{Valign}$ ." + ], + "image_footnote": [], + "bbox": [ + 101, + 87, + 893, + 263 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "design by Gong et al. [25, 26] (i.e., leveraging the DeiT visual transformer [93] adapted for processing filterbank images extracted from the input audio). The decoder takes the three latent vectors and produces a reconstructed filterbank. Formally $D(c, e, s) = \\widehat{a}$ , where $\\widehat{a}$ denotes the reconstructed filterbank. The decoder architecture consists of a fusion module and transformer-encoder layers.", + "bbox": [ + 75, + 369, + 468, + 474 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Training. The audio module is trained with a multiple loss terms that ensure that the three latent spaces are properly disentangled. In addition to the standard autoencoder reconstruction loss, we also employ three cross-reconstruction losses, in which we enforce the correct reconstruction of the audio signal where we modify one of the content, style or emotion latents. Additionally, we employ three loss terms on the latent vector predictions – namely emotion and style classification losses over $e$ and $s$ , and a content similarity loss between pairs of two content latent vectors extracted from audios that have the same spoken content. For a detailed description of the encoder-decoder architecture, a formal definition of the loss functions and a detailed description of the training process please refer to the Sup. Mat.", + "bbox": [ + 75, + 476, + 470, + 690 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Gesture Generation Model", + "text_level": 1, + "bbox": [ + 76, + 696, + 318, + 710 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Motion prior. Similar to [15, 76], our motion prior network is a VAE transformer architecture with encoder $\\mathcal{P}_E$ and decoder $\\mathcal{P}_D$ . Specifically, both $\\mathcal{P}_E$ and $\\mathcal{P}_D$ follow a U-Net-like [85] structure with skip connections between transformer blocks (see Sup. Mat. for details). The positional embeddings are learnable and injected into each multi-head attention layer, following the design of Carion et al. [13]. Formally, the encoder takes a sequence of $T$ frames of the SMPL-X pose vectors $m^{1:T} \\in \\mathbb{R}^{6J \\times T}$ and the first two tokens of its output, $\\mu \\in \\mathbb{R}^{d_m}$ and $\\Sigma \\in \\mathbb{R}^{d_m \\times d_m}$ are used to extract the motion latent $z_m \\in \\mathbb{R}^{d_m}$ via the reparametrization trick. The decoder takes zero positional encodings as", + "bbox": [ + 75, + 719, + 470, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "query input and the motion latent is fed as memory to every cross-attention transformer layer, producing the reconstructed motion $\\hat{m}^{1:T}$ .", + "bbox": [ + 496, + 369, + 890, + 414 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Diffusion process. The forward diffusion process is similar to [36, 69]. We employ fixed variance and linearly scaled noise scheduler. We add noise to the motion latent $z_{m}$ for $D$ diffusion timesteps to obtain $z^{(D)}$ following:", + "bbox": [ + 496, + 415, + 890, + 476 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nq \\left(z _ {m} ^ {\\left(t _ {d}\\right)} \\mid z _ {m} ^ {(0)}\\right) = \\mathcal {N} \\left(z _ {m} ^ {\\left(t _ {d}\\right)}; \\sqrt {\\bar {\\alpha} _ {t _ {d}}} z _ {m} ^ {(0)}, (1 - \\bar {\\alpha} _ {t _ {d}}) \\mathbf {I}\\right),\n$$\n", + "text_format": "latex", + "bbox": [ + 527, + 484, + 861, + 503 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "with $\\alpha_{t_d} = 1 - \\beta_{t_d}$ , $\\bar{\\alpha}_{t_d} = \\prod_{s=1}^{t_d} \\alpha_s$ , and $\\beta_{t_d}$ denotes diffusion process variance.", + "bbox": [ + 496, + 513, + 890, + 544 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Conditional denoising process. The denoising process consists of iteratively denoising a conditioned noisy motion latent vector to obtain the denoised motion latent $z_{\\tilde{m}^{1:T}}$ . Our denoiser $\\Delta$ is a latent variable model [84] and its architecture is similar to the U-Net-like structure of the motion prior encoder $\\mathcal{P}_E$ . The input of the model is a concatenation of: $z_m^{(t_d)}$ , $\\mathrm{SE}(t_d)$ , $c, e, s \\in \\mathbb{R}^{256}$ , where $\\mathrm{SE}(t_d)$ is a sinusoidal positional encoding of diffusion timestep $t_d$ as defined in [36]. $\\Delta$ iteratively denoises through each reversed diffusion step:", + "bbox": [ + 496, + 545, + 890, + 699 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nz _ {m} ^ {(t _ {d} - 1)} = \\Delta \\left(\\left[ z _ {m} ^ {(t _ {d})}, \\operatorname {S E} \\left(t _ {d}\\right), c, e, s \\right]\\right).\n$$\n", + "text_format": "latex", + "bbox": [ + 573, + 705, + 815, + 726 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Training. We optimize the motion prior and the latent denoiser jointly to ensure audio-motion latent code alignment during conditional fusion in the denoising process using a 3-step forward pass through the gesture generation model. First, following standard VAE practice, we reconstruct $\\hat{m}^{1:T}$ by the motion prior forward pass. As shown in Fig. 2, we then disable gradient calculation in $\\mathcal{P}_E$ to infer the intermediate motion latent $\\mathrm{sg}[z_m]$ , which serves as input to the denoiser. At this stage, we obtain the denoiser noise prediction, $\\delta$ and use to compute the diffusion model gradients. Finally, in the third step we com", + "bbox": [ + 496, + 734, + 890, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "1945", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "pute $\\tilde{m}^{1:T} = \\mathcal{P}_D(\\mathrm{sg}[z_{\\tilde{m}}])$ , where $z_{\\tilde{m}}$ is obtained by iteratively using the $\\Delta$ to obtain a fully denoised latent from $z_n^{(t_D)} \\sim \\mathcal{N}(\\mathbf{0}, \\mathbf{I})$ . We indicate computations done without gradients with a stop-gradient operation $\\mathrm{sg}[\\cdot]$ .", + "bbox": [ + 76, + 89, + 467, + 152 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Losses. To train the motion prior, we include the standard VAE losses, namely the reconstruction loss on pose parameters $\\mathcal{L}_{rec}$ and on vertex coordinates $\\mathcal{L}_{Vrec}$ using the smooth L1 metric introduced in [24], which we denote as $L_1^s$ :", + "bbox": [ + 76, + 154, + 467, + 214 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {r e c} = L _ {1} ^ {s} (m ^ {1: T}, \\hat {m} ^ {1: T}), \\quad \\mathcal {L} _ {V r e c} = L _ {1} ^ {s} (V ^ {1: T}, \\hat {V} ^ {1: T}),\n$$\n", + "text_format": "latex", + "bbox": [ + 91, + 220, + 454, + 242 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where the root-centered vertices $V$ are obtained by feeding in pose parameters $m$ to a differentiable SMPL-X layer (without learnable parameters) and a mean shape $\\beta = \\vec{0}$ . The KL divergence loss of the motion prior is:", + "bbox": [ + 76, + 250, + 467, + 310 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {K L} = \\frac {1}{2} \\left[ \\sum_ {i = 1} ^ {z} \\left(\\mu_ {i} ^ {2} + \\sigma_ {i} ^ {2}\\right) - \\sum_ {i = 1} ^ {z} \\left(l o g \\left(\\sigma_ {i} ^ {2}\\right) + 1\\right) \\right].\n$$\n", + "text_format": "latex", + "bbox": [ + 106, + 319, + 437, + 361 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To ensure the alignment of the diffusion-generated motions and the input audio, we apply the alignment reconstruction loss on the inferred motion pose parameters and the vertex coordinates:", + "bbox": [ + 76, + 369, + 467, + 428 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {a l i g n} = L _ {1} ^ {s} (m ^ {1: T}, \\tilde {m} ^ {1: T}), \\quad \\mathcal {L} _ {V a l i g n} = L _ {1} ^ {s} (V ^ {1: T}, \\tilde {V} ^ {1: T}).\n$$\n", + "text_format": "latex", + "bbox": [ + 78, + 436, + 465, + 458 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Finally, we utilize the objective similar to [15, 36, 84] to supervise the denoiser:", + "bbox": [ + 76, + 465, + 467, + 496 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {L D} = \\left\\| \\delta^ {(t _ {d})} - \\Delta (z _ {m} ^ {(t _ {d})}, \\mathrm {S E} (t _ {d}), c, e, s) \\right\\| _ {2} ^ {2},\n$$\n", + "text_format": "latex", + "bbox": [ + 124, + 505, + 419, + 534 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\delta^{(t_d)}$ is the noise vector sampled from $\\mathcal{N}(\\mathbf{0},\\mathbf{I})$ in the corresponding diffusion step $t_d$ . The combined gesture model loss is:", + "bbox": [ + 76, + 544, + 467, + 588 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {g e s} = \\mathcal {L} _ {r e c} + \\mathcal {L} _ {V r e c} + \\mathcal {L} _ {K L} + \\mathcal {L} _ {a l i g n} + \\mathcal {L} _ {V a l i g n} + \\mathcal {L} _ {L D}\n$$\n", + "text_format": "latex", + "bbox": [ + 76, + 601, + 467, + 618 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Inference. We employ DDIM [89] to infer high quality conditional motion samples with a small number of denoising timesteps. During inference we draw a sample vector from $\\mathcal{N}(\\mathbf{0},\\mathbf{I})$ to iteratively denoise in reversed timesteps. The denoised sample is then passed through the decoder $\\mathcal{P}_D(z_{\\tilde{m}^{1:T}})$ to obtain motion $\\tilde{m}^{1:T}$ .", + "bbox": [ + 76, + 627, + 467, + 717 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Gesture Editing", + "text_level": 1, + "bbox": [ + 76, + 726, + 235, + 742 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Due to the disentangling of the inputs, AMUSE achieves semantic gesticulation control using two driving input audios. Specifically, given two input audio signals $a_1$ and $a_2$ , we extract their latent representations of content $c_1, c_2$ , emotion $e_1, e_2$ , and style $s_1, s_2$ . Then, we simply initialize the denoising procedure of $\\Delta$ with the triplet $(c_1, e_2, s_1)$ , generating the gesture with the content and style of $a_1$ but the emotion of input audio $a_2$ . Similarly, instead of emotion we can also change the gesticulation style to that of the speaker of $a_2$ by initializing with $(c_1, e_1, s_2)$ .", + "bbox": [ + 76, + 750, + 467, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Implementation Details", + "text_level": 1, + "bbox": [ + 500, + 89, + 720, + 106 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "MoCap data preparation. The BEAT [55] mocap sequences, captured in a Vicon system at $120\\mathrm{Hz}$ , are downsampled to $30\\mathrm{Hz}$ and processed using $\\mathrm{MoSh}++$ [58, 62] to obtain SMPL-X parameters. Given a sequence of 3D mocap marker positions, we jointly optimize SMPL-X shape and pose parameters, 3D body translation, and embedding of the mocap markers in the SMPL-X surface. Once processed, the sequences are then divided according to the emotion annotations in the BEAT dataset. We use sequences of English speaking subjects in monologue speaking style for training and evaluating AMUSE. For each sequence we draw $m^{1:L}$ at 30 FPS and concatenate with audio content $c$ , emotion $e$ , and style $s$ latent vectors. Then, we segment it to 10-sec windows $T$ , beginning from the timestamp 0 and discarding additional unaligned information at the end. This preprocessing choice allows us to train transformer networks without masking. We provide additional data processing information in the Sup. Mat.", + "bbox": [ + 496, + 114, + 890, + 387 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Audio preprocessing. We use audio sequences belonging to eight categorical emotion labels (neutral, happy, angry, sad, contempt, surprise, fear, and disgust). Each audio chunk of 10s is converted into a filter bank with 128 melfrequency bins with a 25ms Hamming frame window and 10ms frame shift. We mask each sample with a maximum length of 24 in the frequency domain and a maximum length of 96 in the time domain, employing Park et al. [72]. Following [25, 26], we standardize the filter bank and augment it via noise injection and circular shifting. Before feeding in our speech disentanglement model, each filter bank is split into a sequence of fixed 1209 patches of $16 \\times 16$ each having 6 units overlap in frequency and time domain.", + "bbox": [ + 496, + 388, + 890, + 584 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Motion prior. The motion prior is a VAE encoder-decoder with 9 layers and 4 heads, following Chen et al. [15]. The encoder-decoder is a U-Net-like transformer with residual connections. Learnable positional embeddings are injected in each multi-head attention layer. We have a linear projection at the start and the end of our motion prior network. The KL divergence term is weighted with a factor of $1e - 4$ . Denoiser. The denoiser follows the same network architecture as our prior encoder. The hidden dimension of all transformer layers is 1024. We use 1000 diffusion steps $D$ during training and 50 during inference. Noise betas are in range $[0.00085, 0.012]$ . We jointly optimize the prior and denoiser networks for 5000 epochs with batch size of 64, learning rate 0.0001, and the AdamW optimizer [59].", + "bbox": [ + 496, + 585, + 890, + 797 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 500, + 814, + 630, + 830 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Speech disentanglement model. We evaluate the performance of the speech disentanglement model quantitatively using classification accuracy and F1 scores on emotion and style. The accuracy is computed as average scores", + "bbox": [ + 496, + 839, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "1946", + "bbox": [ + 483, + 945, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "for all 8 emotion as well style categories that are part of the test dataset. The emotion and style accuracy is $91.53\\%$ and $96.06\\%$ , respectively. The emotion F1 score and style F1 scores are 0.914 and 0.960, respectively. See the Sup. Mat. for ablations and a detailed metric analysis.", + "bbox": [ + 75, + 90, + 467, + 166 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Gesture generation model. We evaluate the performance of our gesture generation model quantitatively, qualitatively, and perceptually against following methods: TalkSHOW [101] and the re-implementation of Habibie et al. [31] provided by the TalkSHOW authors in the official TalkSHOW release [100], DiffuseStyleGesture (DSG) [97], MoGlow [34], and CaMN [55]. Additionally, we adapt TalkSHOW to include categorical emotion labels as input along with the existing architecture that only allows one-hot encodings of personal style. We then retrain it on our training data. We refer to it as TalkSHOW-BEAT. There are some concurrent works [5, 7, 56], which introduce methods for gesture generation from speech, however, direct comparison is hindered by the unavailability of released code our task. Refer to the Sup. Mat. for the ablation experiments, the emotion and style editing experiments, and their quantitative evaluation.", + "bbox": [ + 75, + 167, + 468, + 422 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1. Quantitative Evaluation", + "text_level": 1, + "bbox": [ + 76, + 434, + 297, + 450 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To quantitatively evaluate our method's gesture generations and edited gesture generations, we train a transformer-based encoder architecture (denoted as $M$ ) similar to Petrovich et al. [76] in an autoencoder setting, where we append a CLS token at the beginning of the motion sequence. $M$ is trained with a cross-entropy emotion classification objective applied to the output CLS token. We train $M$ on the BEAT training dataset and use its features to compute the following metrics: (1) Fréchet gesture distance (FGD): We", + "bbox": [ + 75, + 457, + 467, + 592 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/efd0dbe3c6ed65df87582b36cc38c98560ad69b7038fee6f302b3ac43cd5fcab.jpg", + "table_caption": [], + "table_footnote": [ + "a GA is average of all 8 emotions.", + "b GA for these are average accuracy for all generations with 7 edited audio sequences." + ], + "table_body": "
MethodSRGR↑BA↑FGD↓Div→GAa↑
GT—0.83—27.8364.04
Ours0.360.81388.6325.0646.76
Ours-EmoEditb—0.79792.5824.6834.18
TalkSHOW-BEAT0.310.64808.9924.1622.71
TalkSHOW [101]0.300.60762.1523.1929.41
DSG [97]0.230.40763.1019.7722.70
Habibie et al. [31]0.230.39809.1721.3416.67
MoGlow [34]0.210.351097.0319.5016.62
CaMN [55]0.210.391063.8718.9014.17
", + "bbox": [ + 81, + 616, + 467, + 760 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1. Gesture quantitative results. We compare our methods against several SOTA methods using metrics explained in Sec. 5.1. We observe that AMUSE outperforms in all scores compared to baseline methods. Additionally, AMUSE-EmoEdit outperforms in Beat Align, Diversity, and Gesture Emotion Accuracy scores compared to the baseline methods.", + "bbox": [ + 75, + 813, + 467, + 895 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "follow [88, 103, 104] to compute the feature distance between generated and ground truth motion features. (2) Gesture diversity (Div): Similarly to Chen et al. [15], we compute variance across generated features. (3) Gesture emotion accuracy (GA): We report top-1 emotion classification accuracy predicted by a classifier trained on the motion $M$ -predicted latents. (4) Beat align (BA): We follow [51, 55], to evaluate the motion-speech correlation in terms of the similarity between the kinematic motion beats and speech audio beats. The kinematic motion beats are directly computed from the generated motion sequences. (5) Semantic-Relevant Gesture Recall (SRGR): We follow Liu et al. [55], to evaluate the semantic relevancy of gestures with GT motion. We use the ground truth semantic scores to compute this metric. The scores are obtained from the BEAT authors, representing a continuous score on a scale 0-1 per gesture style for 4 gesture semantic categories: beat, deictic, iconic, and metaphoric. While comparing with methods that output coarse skeletal data (DSG [97], MoGlow [34], and CaMN [55]), we convert the skeleton motion data into the SMPL-X axis angle representation. For details on the architectures and training of $M$ , and the losses, please refer to the Sup. Mat.", + "bbox": [ + 496, + 90, + 890, + 438 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We prepare the evaluation data by randomly selecting 72 unique motion sequences each of length 10s and comprising 8 emotions across test subjects and compute the aforementioned metrics. We use 9 sequences for each emotion per subject. The results are reported in Tab. 1. All best scores are highlighted in green and second best in blue. AMUSE outperforms the baseline methods in all given metrics. To validate the performance of gesture emotion editing, we also report the same metrics for the emotion editing task (Ours-EmoEdit). During inference, the input style and content latents are extracted from neutral-emotion audio, while the emotion latent comes from a different audio of different emotion. These emotional edits offer numerous possibilities, allowing for transitions from any to any emotion. Tab. 1 shows the average for editing from neutral to other emotions. Since we require the GT gesture semantics score to compute SRGR metric, it is not possible to compute the SRGR for the synthetic edited-emotion gestures as they are not part of the original BEAT dataset. Ours-EmoEdit outperforms the baseline methods in BA, Div, and GA metrics. This demonstrates the capability of our model to maintain highly discriminative cues when switching between different emotions. TalkSHOW-BEAT has the second best score for SRGR whereas TalkSHOW demonstrates second best FGD score. Although, our model and ours-EmoEdit show improvements over the baseline methods, GT motions have higher diversity, Beat alignment score, and are easier to classify than generations of AMUSE, highlighting the challenging nature of the problem.", + "bbox": [ + 496, + 462, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "1947", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/fc60d13af2bb6a9cafc31879ce92d3d0be33118b096d7346506181c227888a85.jpg", + "image_caption": [ + "Figure 3. Qualitative comparison across all emotions. We evaluate generation on different test audios. AMUSE exhibits well-synchronized beat gestures and consistently produces gestures that accurately convey the emotional content expressed in the input speech." + ], + "image_footnote": [], + "bbox": [ + 76, + 85, + 885, + 234 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/10f8b2f70747ff519b934e9e660b3e1def7176179d1c1813988bf47314fe3b3b.jpg", + "image_caption": [ + "Figure 4. Qualitative comparison with baseline methods. The speech segment describes intense angry speech." + ], + "image_footnote": [], + "bbox": [ + 78, + 287, + 472, + 593 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.2. Qualitative Evaluation", + "text_level": 1, + "bbox": [ + 76, + 662, + 287, + 678 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Comparison with baseline methods. In Fig. 4, we demonstrate comparison with baseline methods that output a 3D body mesh: Habibie et al. [31], TalkSHOW [101], TalkSHOW-BEAT, and the BEAT ground truth (GT) [55]. We observe that AMUSE generates gestures that are semantically closer to the speech content and produces expressive emotional gestures closer to the perceived emotion. For example, the GT motion exhibits anger when saying \"put them into detention\". AMUSE demonstrates tense posture and aggressive movements comparable with the ground truth data and accurate synchronization with the spoken words. TalkSHOW [101] and Habibie et al. [31] exhibit limited movement and display inferior and static gestures on test audios as seen in the last two rows", + "bbox": [ + 75, + 688, + 472, + 900 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/7359634e0462ace41c9f62be4167596525d8f96ebbde7384a8a2de69b67e5475.jpg", + "image_caption": [ + "Figure 5. Qualitative evaluation of diverse generations. Multiple generations overlayed." + ], + "image_footnote": [], + "bbox": [ + 594, + 289, + 830, + 364 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "of Fig. 4. TalkSHOW-BEAT slightly outperforms other baseline methods by demonstrating enhanced synchronized gestures, but it still does not perform as well as AMUSE.", + "bbox": [ + 496, + 430, + 890, + 474 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Diverse emotional gestures. In Fig. 5, our probabilistic model can generate diverse gestures for same input audio.", + "bbox": [ + 498, + 476, + 890, + 506 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Emotional gesture generation. In Fig. 3 AMUSE demonstrates strong correlation with the spoken utterances as well as different emotions. We observe that our model is able to correlate semantic words to associated gestures. For example, gestures demonstrate forceful actions and tense stance with angry audio \"normally get angry\" whereas it generates lowered and calm hand positions for sad audio \"I went to collect\". Similarly, our generations show hands that are closer to body for fearful audio \"in a dangerous situation\" while widely open expressing astonishment for happy and surprised audio \"people pass by\" and \"Wow! You are here\".", + "bbox": [ + 496, + 507, + 892, + 672 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Emotion editing. We use two audio streams of a female subject for neutral and sad emotion. This experiment edits the subject's gesture style from moderately controlled hand movements to a sad style with lethargic posture conveying a sense of heaviness, as seen in Fig. 7 (top).", + "bbox": [ + 496, + 672, + 890, + 750 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Gesture style editing. We use audio streams of two male subjects for the happy (ID - 13) and angry (ID - 2) emotion. With the emotion, style and content latent fusion mechanism from two driving audio streams, AMUSE is able to adapt the male (ID - 13) subject's body gestures from being close to their body to more open with squared tightened shoulders, expressing a shift from happy to angry emotions of a different subject (ID - 2), as shown in Fig. 7 (bottom). Please refer to the supplemental video for qualitative results and comparisons to additional gesture genera", + "bbox": [ + 496, + 750, + 892, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "1948", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/dcd48010663c371765d28e509f35149e52387aa74e2c0ede6cd216ce8325d5ce.jpg", + "image_caption": [ + "Figure 6. The perceptual study results for gesture emotion preference (left) and synchronization with speech (right). The number of attentive participants that passed the catch trials is indicated on the right and the reported results only consider these participants." + ], + "image_footnote": [], + "bbox": [ + 60, + 90, + 890, + 155 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/790684fdaf80e92c2b5d8b5dae1f520bf3f102717544d6328c5f80a8af4ec7ac.jpg", + "image_caption": [ + "Figure 7. Gesture editing. Top: We modify style from being neutral (left) to being sad (right) by combining the emotion latent from sad audio with the content latent from neutral audio. Bottom: We transform the style from Subject 13 being happy (left) to being angry (right) by merging the content latent from happy audio with the style and emotion latents from an angry audio of Subject 2." + ], + "image_footnote": [], + "bbox": [ + 88, + 210, + 472, + 366 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "tion methods [4, 55, 110] trained on coarse skeletal data.", + "bbox": [ + 76, + 486, + 450, + 501 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.3. Perceptual Study", + "text_level": 1, + "bbox": [ + 76, + 513, + 243, + 529 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Design. Our perceptual study is designed as a side-by-side comparison of two gesture videos generated with the same audio as input but by two different methods (AMUSE and another model or GT). The participants are asked to rate their preference of the methods on a five-point Likert scale for \"synchronization with speech\" and \"gesture emotion appropriateness\" given the GT emotion label of the input audio. We recruit 25 participants per method-to-method comparison on Amazon Mechanical Turk. Each participant is shown 24 pairs of randomly selected test set animations, 3 per emotion (neutral, happy, angry, sad, disgust, fear, surprise, and contempt). To allow the participant to get used to the task, we discard the answers of the first three comparisons and repeat these at the end. We incorporate three catch trials and responses from participants that fail on more than one are filtered out, as shown in Fig. 6 (right).", + "bbox": [ + 75, + 537, + 468, + 777 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Results. The results of the study are shown in Fig. 6. AMUSE outperforms all competing methods by a considerable margin on both tasks, suggesting that AMUSE's generations are more appropriate for both the content of the input speech and its emotion compared to the baselines. However, it must be noted that there is still a significant gap between AMUSE and the GT. Please refer to the Sup. Mat. for details about the perceptual study.", + "bbox": [ + 75, + 780, + 470, + 901 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.4. Discussion and Future Work", + "text_level": 1, + "bbox": [ + 500, + 210, + 756, + 226 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Upper-body motion. We focus on the smooth coordination between the pelvis and upper body animation for side-by-side comparisons with other methods, as all other methods primarily focus on upper body movements. Future work should include lower-body motion and locomotion as these impact the perceived emotional state of a sequence.", + "bbox": [ + 496, + 234, + 890, + 325 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Semantics. While the generated gestures, synchronized with the driving speech sequence, do not account for semantics such as deictic and metaphoric gestures, incorporating the text/language modality could help further improve in this direction.", + "bbox": [ + 496, + 325, + 890, + 398 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Facial expressions. While emotional speech-driven face animation methods [18, 74] can be combined with bodies generated from AMUSE, jointly learning to generate emotional 3D bodies from speech is a topic that needs attention. End-to-end training. Joint audio-gesture training may enhance results but requires careful loss term balancing and increased GPU memory. Therefore, we opted for separate training.", + "bbox": [ + 496, + 401, + 890, + 523 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 537, + 617, + 551 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We present AMUSE, a framework to generate emotional body gestures from speech. The emotions and personal styles of the synthesized gestures can be controlled, thanks to the disentanglement of content, emotion, and style directly from the speech. The latent diffusion-based framework can further generate variations of the same gesture with the same emotion. Our quantitative evaluations show that AMUSE achieves state of the art performance on a variety of metrics: diversity, gesture emotion classification accuracy, Frechet gesture distance, beat alignment score, and semantic relevant gesture recall. Finally, our perceptual study demonstrates that AMUSE generates motions that are better synchronized and better match the emotion expressed of the input speech than previous state of the art.", + "bbox": [ + 496, + 561, + 890, + 773 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments. We thank A Cseke, T McConnell and T Alexiadis for help with the perceptual study, B Pellkofer, J Piles-Contreras, and E Fritzler for IT support, and P Kulits, M Petrovich, S Kokane, and S Zojaji for proofreading and insightful discussions. This project has received funding from the European Union's Horizon 2020 research and innovation programme under the Marie Skłodowska-Curie grant agreement No. 860768.", + "bbox": [ + 496, + 787, + 890, + 886 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Disclosure. https://download.is.tue.mpg.de/amuse/Disclosure.txt", + "bbox": [ + 500, + 887, + 885, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "1949", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Kfir Aberman, Yijia Weng, Dani Lischinski, Daniel Cohen-Or, and Baoquan Chen. Unpaired motion style transfer from video to animation. Transactions on Graphics (TOG), 2020. 3", + "[2] Chaitanya Ahuja and Louis-Philippe Morency. Language2Pose: Natural language grounded pose forecasting. International Conference on 3D Vision (3DV), 2019. 2", + "[3] Chaitanya Ahuja, Dong Won Lee, and Louis-Philippe Morency. Low-resource adaptation for personalized cospeech gesture generation. In Computer Vision and Pattern Recognition (CVPR), pages 20566-20576, 2022. 2", + "[4] Simon Alexanderson, Gustav Eje Henter, Taras Kucherenko, and Jonas Beskow. Style-controllable speech-driven gesture synthesis using normalising flows. Computer Graphics Forum (CGF), 39(2):487-496, 2020. 3, 8", + "[5] Simon Alexanderson, Rajmund Nagy, Jonas Beskow, and Gustav Eje Henter. Listen, Denoise, Action! Audio-driven motion synthesis with diffusion models. Transactions on Graphics (TOG), 42(4):1-20, 2023. 1, 3, 6", + "[6] Tenglong Ao, Qingzhe Gao, Yuke Lou, Baoquan Chen, and Libin Liu. Rhythmic gesticulator. Transactions on Graphics (TOG), 41(6):1-19, 2022. 3", + "[7] Tenglong Ao, Zeyi Zhang, and Libin Liu. GestureDiffu-CLIP: Gesture diffusion model with CLIP latents. Transactions on Graphics (TOG), 2023. 1, 3, 6", + "[8] Nikos Athanasiou, Mathis Petrovich, Michael J. Black, and Gül Varol. TEACH: Temporal Action Compositions for 3D Humans. In International Conference on 3D Vision (3DV), 2022. 2", + "[9] Nikos Athanasiou, Mathis Petrovich, Michael J. Black, and Gül Varol. SINC: Spatial composition of 3D human motions for simultaneous action generation. International Conference on Computer Vision (ICCV), 2023. 2", + "[10] Emad Barsoum, John Kender, and Zicheng Liu. HPGAN: probabilistic 3d human motion prediction via GAN. In Computer Vision and Pattern Recognition Workshops (CVPRW), 2017. 2", + "[11] Uttaran Bhattacharya, Elizabeth Childs, Nicholas Rewkowski, and Dinesh Manocha. Speech2affectivegestures: Synthesizing co-speech gestures with generative adversarial affective expression learning. In ACM International Conference on Multimedia (MM), 2021. 3", + "[12] Uttaran Bhattacharya, Nicholas Rewkowski, Abhishek Banerjee, Pooja Guhan, Aniket Bera, and Dinesh Manocha. Text2gestures: A transformer-based network for generating emotive body gestures for virtual agents. In IEEE Virtual Reality and 3D User Interfaces, VR, 2021. 3", + "[13] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In European Conference on Computer Vision (ECCV), pages 213-229. Springer, 2020. 4", + "[14] Justine Cassell, Hannes Hogni Vilhjalmsson, and Timo" + ], + "bbox": [ + 86, + 114, + 468, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "thy Bickmore. BEAT: The Behavior Expression Animation Toolkit, pages 163-185. Springer, 2004. 2", + "[15] Xin Chen, Biao Jiang, Wen Liu, Zilong Huang, Bin Fu, Tao Chen, and Gang Yu. Executing your commands via motion diffusion in latent space. In Computer Vision and Pattern Recognition (CVPR), pages 18000-18010, 2023. 4, 5, 6", + "[16] Enric Corona, Albert Pumarola, G. Alenyà, and F. Moreno-Noguer. Context-aware human motion prediction. In Computer Vision and Pattern Recognition (CVPR), pages 6990–6999, 2020. 2", + "[17] Rishabh Dabral, Muhammad Hamza Mughal, Vladislav Golyanik, and Christian Theobalt. Mofusion: A framework for denoising-diffusion-based motion synthesis. In Computer Vision and Pattern Recognition (CVPR), 2023. 2", + "[18] Radek Daneček, Kiran Chhatre, Shashank Tripathi, Yandong Wen, Michael Black, and Timo Bolkart. Emotional speech-driven animation with content-emotion disentangle-ment. In International Conference on Computer Graphics and Interactive Techniques in Asia (SIGGRAPH ASIA), 2023. 2, 3, 8", + "[19] Paul Ekman. Facial expression and emotion. American Psychologist, 48(4):384-392, 1993. 1", + "[20] Esam Ghaleb, Ilya Burenko, Marlou Rasenberg, Wim Pouw, Peter Uhrig, Judith Holler, Ivan Toni, Asli Özyurek, and Raquel Fernández. Co-speech gesture detection through multi-phase sequence labeling. In Winter Conference on Applications of Computer Vision (WACV), 2024. 3", + "[21] Saeed Ghorbani, Ylva Ferstl, Daniel Holden, Nikolaus F. Troje, and Marc-Andre Carbonneau. Zeroeggs: Zero-shot example-based gesture generation from speech. Computer Graphics Forum (CGF), 42(1):206-216, 2023. 3", + "[22] Anindita Ghosh, Noshaba Cheema, Cennet Oguz, Christian Theobalt, and Philipp Slusallek. Synthesis of compositional animations from textual descriptions. In International Conference on Computer Vision (ICCV), 2021. 2", + "[23] Shiry Ginosar, Amir Bar, Gefen Kohavi, Caroline Chan, Andrew Owens, and Jitendra Malik. Learning individual styles of conversational gesture. In Computer Vision and Pattern Recognition (CVPR), pages 3497-3506, 2019. 2", + "[24] Ross B. Girshick. Fast R-CNN. In International Conference on Computer Vision (ICCV), pages 1440-1448, 2015. 5", + "[25] Yuan Gong, Yu-An Chung, and James R. Glass. AST: audio spectrogram transformer. In *Interspeech* 2021, pages 571-575, 2021. 4, 5", + "[26] Yuan Gong, Cheng-I Lai, Yu-An Chung, and James R. Glass. SSAST: self-supervised audio spectrogram transformer. In AAAI Conference on Artificial Intelligence, pages 10699-10709, 2022. 4, 5", + "[27] Chuan Guo, Xinxin Zuo, Sen Wang, Shihao Zou, Qingyao Sun, Annan Deng, Minglun Gong, and Li Cheng. Action2Motion: Conditioned generation of 3d human motions. In ACM International Conference on Multimedia (MM), 2020. 2", + "[28] Chuan Guo, Shihao Zou, Xinxin Zuo, Sen Wang, Wei Ji, Xingyu Li, and Li Cheng. Generating diverse and natural 3D human motions from text. In Computer Vision and Pattern Recognition (CVPR), 2022. 2" + ], + "bbox": [ + 509, + 93, + 890, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "1950", + "bbox": [ + 483, + 945, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[29] Nejla Gürefe. The role of gestures in mathematical discourse of hard-hearing students: Prism example. Acta Destructica Napocensia, 11:125-140, 2018. 1", + "[30] I. Habibie, Daniel Holden, Jonathan Schwarz, J. Yearsley, and T. Komura. A recurrent variational autoencoder for human motion synthesis. In *British Machine Vision Conference (BMVC)*, 2017. 2", + "[31] Ikhsanul Habibie, Weipeng Xu, Dushyant Mehta, Lingjie Liu, Hans-Peter Seidel, Gerard Pons-Moll, Mohamed Elgharib, and Christian Theobalt. Learning speech-driven 3d conversational gestures from video. In ACM International Conference on Intelligent Virtual Agents (IVA), 2021. 1, 6, 7", + "[32] Ikhsanul Habibie, Mohamed A. Elgharib, Kripasindhu Sarkar, Ahsan Abdullah, Simbarashe Linval Nyatsanga, Michael Neff, and Christian Theobalt. A motion matching-based framework for controllable gesture synthesis from speech. International Conference on Computer Graphics and Interactive Techniques (SIGGRAPH), 2022. 2", + "[33] Félix G. Harvey, Mike Yurick, Derek Nowrouzezahrai, and Christopher Joseph Pal. Robust motion in-between. Transactions on Graphics (TOG), 2020. 2", + "[34] Gustav Eje Henter, Simon Alexanderson, and Jonas Beskow. MoGlow: Probabilistic and controllable motion synthesis using normalising flows. Transactions on Graphics (TOG), 39(4):236:1-236:14, 2020. 3, 6", + "[35] Sandra Herbert. Gesture types for functions. Mathematics Education Research Group of Australasia, pages 322-329, 2012. 1", + "[36] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In Conference on Neural Information Processing Systems (NeurIPS), 2020. 4, 5", + "[37] Xun Huang and Serge J. Belongie. Arbitrary style transfer in real-time with adaptive instance normalization. In International Conference on Computer Vision (ICCV), pages 1510-1519, 2017. 3", + "[38] Tero Karras, Timo Aila, Samuli Laine, Antti Herva, and Jaakko Lehtinen. Audio-driven facial animation by joint end-to-end learning of pose and emotion. Transactions on Graphics (TOG), 2017. 3", + "[39] Adam Kendon. Gesture: Visible action as utterance. Cambridge University Press, 2004. 1", + "[40] Gwantae Kim, Seonghyeok Noh, Insung Ham, and Hanseok Ko. MPE4G: Multimodal pretrained encoder for co-speech gesture generation. In International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 1-5, 2023. 2", + "[41] Jihoon Kim, Taehyun Byun, Seungyoung Shin, Jung-dam Won, and Sungjoon Choi. Conditional motion in-between. Pattern Recognition, 2022. 2", + "[42] Chris L Kleinke. Gaze and eye contact: A research review. Psychological Bulletin, 100(1):78-100, 1986. 1", + "[43] Zhifeng Kong, Wei Ping, Jiaji Huang, Kexin Zhao, and Bryan Catanzaro. DiffWave: A versatile diffusion model for audio synthesis. In International Conference on Learning Representations (ICLR), 2021. 3" + ], + "bbox": [ + 86, + 92, + 470, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[44] Stefan Kopp, Brigitte Krenn, Stacy Marsella, Andrew N. Marshall, Catherine Pelachaud, Hannes Pirker, Kristinn R. Thorisson, and Hannes Vilhjalmsson. Towards a common framework for multimodal generation: The behavior markup language. In Intelligent Virtual Agents (IVA), pages 205-217. Springer, 2006. 2", + "[45] Quoc Anh Le, Souheil Hanoune, and Catherine Pelachaud. Design and implementation of an expressive gesture model for a humanoid robot. In IEEE-RAS International Conference on Humanoid Robots, pages 134-140, 2011. 2", + "[46] Gilwoo Lee, Zhiwei Deng, Shugao Ma, Takaaki Shiratori, Siddhartha S. Srinivasa, and Yaser Sheikh. Talking With Hands 16.2M: A large-scale dataset of synchronized body-finger motion and audio for conversational motion analysis and synthesis. In International Conference on Computer Vision (ICCV), pages 763-772, 2019. 2", + "[47] Robert W Levenson. Blood, sweat, and fears: The autonomic architecture of emotion. Annals of the New York Academy of Sciences, 1000:348-366, 2003. 1", + "[48] Jing Li, Di Kang, Wenjie Pei, Xuefei Zhe, Ying Zhang, Zhenyu He, and Linchao Bao. Audio2gestures: Generating diverse gestures from speech audio with conditional variational autoencoders. In Computer Vision and Pattern Recognition (CVPR), pages 11293-11302, 2021. 2", + "[49] Peizhuo Li, Kfir Aberman, Zihan Zhang, Rana Hanocka, and Olga Sorkine-Hornung. GANimator: Neural motion synthesis from a single sequence. Transactions on Graphics (TOG), 2022. 2", + "[50] Ruilong Li, Shan Yang, D. A. Ross, and Angjoo Kanazawa. AI choreographer: Music conditioned 3D dance generation with AIST++. In International Conference on Computer Vision (ICCV), 2021. 2", + "[51] Ruilong Li, Shan Yang, David A. Ross, and Angjoo Kanazawa. AI choreographer: Music conditioned 3d dance generation with AIST++. In International Conference on Computer Vision (ICCV), pages 13381-13392, 2021. 6", + "[52] William Li and Philippe Pasquier. Automatic affect classification of human motion capture sequences in the valence-arousal model. In International Symposium on Movement and Computing, 2016. 3", + "[53] Yuanzhi Liang, Qianyu Feng, Linchao Zhu, Li Hu, Pan Pan, and Yi Yang. Seeg: Semantic energized co-speech gesture generation. In Computer Vision and Pattern Recognition (CVPR), 2022. 2", + "[54] Haiyang Liu, Naoya Iwamoto, Zihao Zhu, Zhengqing Li, You Zhou, Elif Bozkurt, and Bo Zheng. DisCo: Disentangled implicit content and rhythm learning for diverse co-speech gestures synthesis. In ACM International Conference on Multimedia (MM), pages 3764-3773, 2022. 3", + "[55] Haiyang Liu, Zihao Zhu, Naoya Iwamoto, Yichen Peng, Zhengqing Li, You Zhou, Elif Bozkurt, and Bo Zheng. BEAT: A large-scale semantic and emotional multi-modal dataset for conversational gestures synthesis. European Conference on Computer Vision (ECCV), 2022. 2, 3, 5, 6, 7, 8", + "[56] Haiyang Liu, Zihao Zhu, Giorgio Becherini, Yichen Peng, Mingyang Su, You Zhou, Xuefei Zhe, Naoya Iwamoto, Bo" + ], + "bbox": [ + 509, + 92, + 890, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "1951", + "bbox": [ + 483, + 944, + 513, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Zheng, and Michael J. Black. EMAGE: Towards unified holistic co-speech gesture generation via expressive masked audio gesture modeling. In Computer Vision and Pattern Recognition (CVPR), 2024. 1, 2, 6", + "[57] Yuejiang Liu, Riccardo Cadei, Jonas Schweizer, Sherwin Bahmani, and Alexandre Alahi. Towards robust and adaptive motion forecasting: A causal representation perspective. In Computer Vision and Pattern Recognition (CVPR), 2022. 2", + "[58] Matthew M. Loper, Naureen Mahmood, and Michael J. Black. MoSh: Motion and shape capture from sparse markers. Transactions on Graphics (TOG), 33(6):220:1-220:13, 2014. 5", + "[59] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In International Conference on Learning Representations (ICLR), 2019. 5", + "[60] Shuhong Lu, Youngwoo Yoon, and Andrew W. Feng. Cospeech gesture synthesis using discrete gesture token learning. International Conference on Intelligent Robots and Systems (IROS), pages 9808-9815, 2023. 2", + "[61] Birgit Lugarin, Catherine Pelachaud, and David Traum, editors. The Handbook on Socially Interactive Agents: 20 Years of Research on Embodied Conversational Agents, Intelligent Virtual Agents, and Social Robotics Volume 1: Methods, Behavior, Cognition. 2021. 1", + "[62] Naureen Mahmood, Nima Ghorbani, Nikolaus F. Troje, Gerard Pons-Moll, and Michael J. Black. Amass: Archive of motion capture as surface shapes. In International Conference on Computer Vision (ICCV), 2019. 2, 5", + "[63] Stacy Marsella, Yuyu Xu, Margaux Lhommet, Andrew Feng, Stefan Scherer, and Ari Shapiro. Virtual character performance from speech. In SIGGRAPH/Eurographics Symposium on Computer Animation (SCA), page 25-35, 2013. 2", + "[64] Julieta Martinez, Michael J. Black, and Javier Romero. On human motion prediction using recurrent neural networks. In Computer Vision and Pattern Recognition (CVPR), 2017. 2", + "[65] David Mcneill. Hand and mind: What gestures reveal about thought. Bibliovault OAI Repository, the University of Chicago Press, 27, 1994. 1", + "[66] Albert Mehrabian. Nonverbal communication. Aldine-Atherton Chicago, 1972. 1", + "[67] Shivam Mehta, Siyang Wang, Simon Alexanderson, Jonas Beskow, Éva Székely, and Gustav Eje Henter. Diff-TTSG: Denoising probabilistic integrated speech and gesture synthesis. In Proc. ISCA Speech Synthesis Workshop (SSW), pages 150–156, 2023. 2", + "[68] Davide Moltisanti, Jinyi Wu, Bo Dai, and Chen Change Loy. BRACE: the breakdancing competition dataset for dance motion synthesis. In European Conference on Computer Vision (ECCV), pages 329-344, 2022. 2", + "[69] Alexander Quinn Nichol and Prafulla Dhariwal. Improved denoising diffusion probabilistic models. In International Conference on Machine Learning (ICML), pages 8162-8171, 2021. 4" + ], + "bbox": [ + 86, + 90, + 470, + 898 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[70] Dirk Ormoneit, Michael J. Black, T. Hastie, and H. Kjellström. Representing cyclic human motion using functional analysis. Image and Vision Computing (IVC), 2005. 2", + "[71] Kunkun Pang, Dafei Qin, Yingruo Fan, Julian Habekost, Takaaki Shiratori, Junichi Yamagishi, and Taku Komura. BodyFormer: Semantics-guided 3d body gesture synthesis with transformer. Transactions on Graphics (TOG), 42(4), 2023. 2", + "[72] Daniel S. Park, William Chan, Yu Zhang, Chung-Cheng Chiu, Barret Zoph, Ekin D. Cubuk, and Quoc V. Le. SpecAugment: A simple data augmentation method for automatic speech recognition. In Interspeech 2019. ISCA, 2019. 5", + "[73] Georgios Pavlakos, Vasileios Choutas, Nima Ghorbani, Timo Bolkart, Ahmed A. A. Osman, Dimitrios Tzionas, and Michael J. Black. Expressive body capture: 3D hands, face, and body from a single image. In Computer Vision and Pattern Recognition (CVPR), pages 10975-10985, 2019. 2, 3", + "[74] Ziqiao Peng, Haoyu Wu, Zhenbo Song, Hao Xu, Xiangyu Zhu, Jun He, Hongyan Liu, and Zhaoxin Fan. EmoTalk: Speech-driven emotional disentanglement for 3d face animation. In International Conference on Computer Vision (ICCV), pages 20687-20697, 2023. 2, 3, 8", + "[75] Mathis Petrovich, Michael J. Black, and Gül Varol. Action-conditioned 3D human motion synthesis with transformer VAE. In International Conference on Computer Vision (ICCV), 2021. 2", + "[76] Mathis Petrovich, Michael J. Black, and Gül Varol. Action-conditioned 3D human motion synthesis with transformer VAE. In International Conference on Computer Vision (ICCV), 2021. 4, 6", + "[77] Mathis Petrovich, Michael J. Black, and Gül Varol. TEMOS: Generating diverse human motions from textual descriptions. In European Conference on Computer Vision (ECCV), 2022. 2", + "[78] Paolo Petta, Catherine Pelachaud, and Roddy Cowie. Emotion-Oriented Systems: The Humaine Handbook. Springer Publishing Company, Incorporated, 2011. 1", + "[79] Rosalind W. Picard. Affective Computing. MIT Press, 1997. 1", + "[80] I. Poggi, C. Pelachaud, F. de Rosis, V. Carofiglio, and B. De Carolis. Greta. A Believable Embodied Conversational Agent, pages 3-25. Springer, 2005. 2", + "[81] Xingqun Qi, Chen Liu, Lincheng Li, Jie Hou, Haoran Xin, and Xin Yu. EmotionGesture: Audio-driven diverse emotional co-speech 3d gesture generation, 2023. 3", + "[82] Xingqun Qi, Chen Liu, Muyi Sun, Lincheng Li, Changjie Fan, and Xin Yu. Diverse 3D hand gesture prediction from body dynamics by bilateral hand disentanglement. In Computer Vision and Pattern Recognition (CVPR), pages 4616-4626, 2023. 2", + "[83] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning (ICML), pages 8748-8763. PMLR, 2021. 3" + ], + "bbox": [ + 509, + 90, + 890, + 898 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "1952", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[84] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Computer Vision and Pattern Recognition (CVPR), pages 10674-10685, 2022. 4, 5", + "[85] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-Net: Convolutional networks for biomedical image segmentation. In Medical Image Computing and Computer-Assisted Intervention - MICCAI, pages 234-241, 2015. 4", + "[86] Tim Salzmann, Marco Pavone, and Markus Ryll. Motron: Multimodal Probabilistic Human Motion Forecasting. In Computer Vision and Pattern Recognition (CVPR), 2022. 2", + "[87] Klaus R Scherer. Vocal communication of emotion: A review of research paradigms. Speech Communication, 40 (1-2):227-256, 2003. 1", + "[88] Maximilian Seitzer. pytorch-fid: FID Score for PyTorch. https://github.com/mseitzer/pytorch-fid, 2020. Version 0.3.0.6", + "[89] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In International Conference on Learning Representations (ICLR), 2021. 5", + "[90] Kim Sung-Bin, Lee Hyun, Da Hye Hong, Suekyeong Nam, Janghoon Ju, and Tae-Hyun Oh. Laughtalk: Expressive 3d talking head generation with laughter. In Winter Conference on Applications of Computer Vision (WACV), pages 6404-6413, 2024. 2", + "[91] Mingyang Sun†, Mengchen Zhao†, Yaqing Hou*, Minglei Li, Huang Xu, Songcen Xu, and Jianye Hao. Cospeech gesture synthesis by reinforcement learning with contrastive pre-trained rewards. In Computer Vision and Pattern Recognition (CVPR), 2023. 2", + "[92] Marcus Thiébaux, Stacy Marsella, Andrew N. Marshall, and Marcelo Kallmann. SmartBody: behavior realization for embodied conversational agents. In Adaptive Agents and Multi-Agent Systems (AAMAS), 2008. 2", + "[93] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Hervé Jégou. Training data-efficient image transformers & distillation through attention. In International Conference on Machine Learning (ICML), pages 10347-10357, 2021. 4", + "[94] Jonathan Tseng, Rodrigo Castellon, and Karen Liu. Edge: Editable dance generation from music. In Computer Vision and Pattern Recognition (CVPR), 2023. 2", + "[95] R. Williams. The Animator's Survival Kit: A Manual of Methods, Principles and Formulas for Classical, Computer Games, Stop Motion and Internet Animators. Farrar, Straus and Giroux, 2012. 2", + "[96] Jingyu Wu, Shi Chen, Shuyu Gan, Weijun Li, Changyuan Yang, and Lingyun Sun. Cultural self-adaptive multimodal gesture generation based on multiple culture gesture dataset. In ACM International Conference on Multimedia (MM), page 3538-3549, 2023. 3", + "[97] Sicheng Yang, Zhiyong Wu, Minglei Li, Zhensong Zhang, Lei Hao, Weihong Bao, Ming Cheng, and Long Xiao. Diffusestylegusture: Stylized audio-driven co-speech gesture generation with diffusion models. In International Joint Conference on Artificial Intelligence (IJCAI), 2023. 3, 6" + ], + "bbox": [ + 86, + 90, + 470, + 900 + ], + "page_idx": 11 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[98] Sicheng Yang, Zhiyong Wu, Minglei Li, Zhensong Zhang, Lei Hao, Weihong Bao, and Haolin Zhuang. Qpgesture: Quantization-based and phase-guided motion matching for natural speech-driven gesture generation. In Computer Vision and Pattern Recognition (CVPR), pages 2321-2330, 2023. 2", + "[99] Payam Jome Yazdian, Mo Chen, and Angelica Lim. Gesture2vec: Clustering gestures using representation learning methods for co-speech gesture generation. In International Conference on Intelligent Robots and Systems (IROS), pages 3100-3107, 2022. 2", + "[100] Hongwei Yi. Show github. https://github.com/yhw-yhw/SHOW, 2023.6", + "[101] Hongwei Yi, Hualin Liang, Yifei Liu, Qiong Cao, Yandong Wen, Timo Bolkart, Dacheng Tao, and Michael J Black. Generating holistic 3d human motion from speech. In Computer Vision and Pattern Recognition (CVPR), pages 469-480, 2023. 1, 2, 6, 7", + "[102] Li-Ping Yin, Yijun Wang, Tianyu He, Jinming Liu, Wei Zhao, Bohan Li, Xin Jin, and Jianxin Lin. Emog: Synthesizing emotive co-speech 3d gesture with diffusion model. ArXiv, abs/2306.11496, 2023. 3", + "[103] Youngwoo Yoon, Woo-Ri Ko, Minsu Jang, Jaeyeon Lee, Jaehong Kim, and Geehyuk Lee. Robots learn social skills: End-to-end learning of co-speech gesture generation for humanoid robots. In International Conference on Robotics and Automation (ICRA), 2019. 6", + "[104] Youngwoo Yoon, Bok Cha, Joo-Haeng Lee, Minsu Jang, Jaeyeon Lee, Jaehong Kim, and Geehyuk Lee. Speech gesture generation from the trimodal context of text, audio, and speaker identity. Transactions on Graphics (TOG), 39(6), 2020. 2, 6", + "[105] Youngwoo Yoon, Pieter Wolfert, Taras Kucherenko, Carla Viegas, Teodor Nikolov, Mihail Tsakov, and Gustav Eje Henter. The genea challenge 2022: A large evaluation of data-driven co-speech gesture generation. In International Conference on Multimodal Interaction, 2022. 2", + "[106] Ye Yuan and Kris M. Kitani. DLow: Diversifying latent flows for diverse human motion prediction. In European Conference on Computer Vision (ECCV), 2020. 2", + "[107] Yicheng Zhong, Huawei Wei, Peiji Yang, and Zhisheng Wang. Expclip: Bridging text and facial expressions via semantic alignment. In AAAI Conference on Artificial Intelligence, pages 7614-7622, 2024. 2", + "[108] Yi Zhou, Connelly Barnes, Jingwan Lu, Jimei Yang, and Hao Li. On the continuity of rotation representations in neural networks. In Computer Vision and Pattern Recognition (CVPR), pages 5745-5753, 2019. 3", + "[109] Yi Zhou, Jingwan Lu, Connelly Barnes, Jimei Yang, Sitao Xiang, and Hao Li. Generative tinkering: Long-term inbetweening of 3d human motions. arXiv:2005.08891, 2020. 2", + "[110] Lingting Zhu, Xian Liu, Xuanyu Liu, Rui Qian, Ziwei Liu, and Lequan Yu. Taming diffusion models for audio-driven co-speech gesture generation. In Computer Vision and Pattern Recognition (CVPR), pages 10544–10553, 2023. 2, 8" + ], + "bbox": [ + 503, + 90, + 890, + 882 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "1953", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 11 + } +] \ No newline at end of file diff --git a/2024/Emotional Speech-driven 3D Body Animation via Disentangled Latent Diffusion/8fe0dd5b-5118-4e0a-ac09-f70da9816cdc_model.json b/2024/Emotional Speech-driven 3D Body Animation via Disentangled Latent Diffusion/8fe0dd5b-5118-4e0a-ac09-f70da9816cdc_model.json new file mode 100644 index 0000000000000000000000000000000000000000..dffac2d118719be2602129016a4e3c2cc1a9638b --- /dev/null +++ b/2024/Emotional Speech-driven 3D Body Animation via Disentangled Latent Diffusion/8fe0dd5b-5118-4e0a-ac09-f70da9816cdc_model.json @@ -0,0 +1,2787 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.808, + 0.045 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.082, + 0.131, + 0.888, + 0.152 + ], + "angle": 0, + "content": "Emotional Speech-driven 3D Body Animation via Disentangled Latent Diffusion" + }, + { + "type": "text", + "bbox": [ + 0.25, + 0.18, + 0.373, + 0.198 + ], + "angle": 0, + "content": "Kiran Chhatre" + }, + { + "type": "text", + "bbox": [ + 0.41, + 0.181, + 0.543, + 0.198 + ], + "angle": 0, + "content": "Radek Danček2" + }, + { + "type": "text", + "bbox": [ + 0.579, + 0.181, + 0.73, + 0.198 + ], + "angle": 0, + "content": "Nikos Athanasiou2" + }, + { + "type": "text", + "bbox": [ + 0.155, + 0.199, + 0.306, + 0.216 + ], + "angle": 0, + "content": "Giorgio Becherini²" + }, + { + "type": "text", + "bbox": [ + 0.343, + 0.199, + 0.496, + 0.216 + ], + "angle": 0, + "content": "Christopher Peters1" + }, + { + "type": "text", + "bbox": [ + 0.534, + 0.199, + 0.676, + 0.216 + ], + "angle": 0, + "content": "Michael J. Black" + }, + { + "type": "text", + "bbox": [ + 0.712, + 0.199, + 0.834, + 0.216 + ], + "angle": 0, + "content": "Timo Bolkart\\(^{2*}\\)" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.217, + 0.888, + 0.235 + ], + "angle": 0, + "content": "\\(^{1}\\)KTH Royal Institute of Technology, Sweden \\(^{2}\\)Max Planck Institute for Intelligent Systems, Germany" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.269, + 0.313, + 0.285 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.302, + 0.474, + 0.695 + ], + "angle": 0, + "content": "Existing methods for synthesizing 3D human gestures from speech have shown promising results, but they do not explicitly model the impact of emotions on the generated gestures. Instead, these methods directly output animations from speech without control over the expressed emotion. To address this limitation, we present AMUSE, an emotional speech-driven body animation model based on latent diffusion. Our observation is that content (i.e., gestures related to speech rhythm and word utterances), emotion, and personal style are separable. To account for this, AMUSE maps the driving audio to three disentangled latent vectors: one for content, one for emotion, and one for personal style. A latent diffusion model, trained to generate gesture motion sequences, is then conditioned on these latent vectors. Once trained, AMUSE synthesizes 3D human gestures directly from speech with control over the expressed emotions and style by combining the content from the driving speech with the emotion and style of another speech sequence. Randomly sampling the noise of the diffusion model further generates variations of the gesture with the same emotional expressivity. Qualitative, quantitative, and perceptual evaluations demonstrate that AMUSE outputs realistic gesture sequences. Compared to the state of the art, the generated gestures are better synchronized with the speech content, and better represent the emotion expressed by the input speech. Our code is available at amuse.is.tue.mpg.de." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.727, + 0.208, + 0.741 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.753, + 0.47, + 0.875 + ], + "angle": 0, + "content": "Animating 3D bodies from speech has a wide range of applications, such as telepresence in AR/VR, avatar animation in games and movies, and to embody interactive digital assistants. While methods for speech-driven 3D body animation have recently shown great progress [5, 7, 31, 56, 101], existing methods do not adequately address one crucial factor: the impact of emotion from the driving speech signal on the generated gestures. Emotions and their expressions play" + }, + { + "type": "image", + "bbox": [ + 0.529, + 0.269, + 0.865, + 0.39 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.528, + 0.39, + 0.871, + 0.4 + ], + "angle": 0, + "content": "... choose a major that is easy to find a good job in the future..." + }, + { + "type": "image", + "bbox": [ + 0.528, + 0.406, + 0.862, + 0.474 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.504, + 0.476, + 0.888, + 0.572 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.584, + 0.892, + 0.64 + ], + "angle": 0, + "content": "Figure 1. Goal. AMUSE generates realistic emotional 3D body gestures directly from a speech sequence (top). It provides user control over the generated emotion by combining the driving speech sequence with a different emotional audio (bottom)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.672, + 0.892, + 0.822 + ], + "angle": 0, + "content": "a fundamental role in human communication [29, 35, 65] and have become an important consideration when designing computer systems that interact with humans in a natural manner [78, 79]. They are of central concern when synthesizing human animations for a wide variety of application contexts, such as Socially Interactive Agents [61]. Because of this, speech-driven animation systems must not only align movement with the rhythm of the speech, but should also be capable of generating gestures that are perceived as expressing the suitable emotion." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.825, + 0.892, + 0.9 + ], + "angle": 0, + "content": "Many factors contribute to the perception of emotion and personal idiosyncrasies, such as facial expressions [19], gaze and eye contact [42], physiological responses [47], tone of voice [87], body language [66], and gestures [39]. When it comes to 3D animation, the most relevant factors" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.888, + 0.188, + 0.9 + ], + "angle": 0, + "content": "*Now at Google." + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1942" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.468, + 0.152 + ], + "angle": 0, + "content": "are facial expressions, gestures, and body language [95]. While emotional speech-driven animation methods have recently been proposed for 3D faces [18, 74, 90, 107], animating emotional bodies from speech remains under-explored." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.153, + 0.47, + 0.485 + ], + "angle": 0, + "content": "Generating gestures solely from speech with emotional control is a difficult task. First, the mapping from audio to body motion is a non-deterministic many-to-many mapping, which is difficult to model. Gestures across subjects can vary when uttering the same sentence, and a single individual's motions can change significantly across repetitions. Second, factoring out the impact of emotional state on the body motion from other, unknown factors, is difficult. This requires disentangling the effects of three different factors on the generated motion, namely content-based (i.e., gestures related to speech rhythm and word utterances), emotion-based, and those based on personal style. AMUSE addresses this by separating a speech sequence into content, emotion, and style latent vectors, which are then used to condition a latent diffusion model. Specifically, AMUSE consists of three main components: (1) an audio autoencoder trained to produce disentangled vectors of content, emotion, and style, (2) a 3D body motion prior in the form of a temporal variational autoencoder (VAE) to generate smooth and realistic gestures, and (3) a latent diffusion model, which generates 3D body motion given the input content, emotion, and style latent vectors." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.486, + 0.468, + 0.652 + ], + "angle": 0, + "content": "Training such a model requires a speech-to-3D body dataset of sufficient scale, which is rich and diverse in speakers and emotions. BEAT [55] is a good candidate because it provides a large set of 3D gestures associated with single-person monologues. Unfortunately, the bodies are represented as skeletons, and it lacks face mocap markers and FLAME expressions. Instead, to produce realistic body animations, we require articulated 3D body surfaces. To overcome this, we convert BEAT sequences to SMPL-X [73] format using MoSh++ [62] and use the SMPL-X parameters for training. See [56] for comparison." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.653, + 0.468, + 0.743 + ], + "angle": 0, + "content": "Our contributions are: (1) We present a framework to synthesize emotional 3D body gestures directly from speech. (2) We factor an input audio into disentangled content, emotion and style vectors, which enables us to separately control emotion in generated gestures. (3) We adapt temporal latent diffusion for multiple target conditions." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.76, + 0.218, + 0.775 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.786, + 0.444, + 0.801 + ], + "angle": 0, + "content": "2.1. 3D Conditional Human Motion Generation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.81, + 0.468, + 0.901 + ], + "angle": 0, + "content": "Early works focus mostly on predicting [10, 16, 33, 41, 57, 64, 70, 86, 106, 109] or generating human motion [30, 49], but do not consider multi-modal control. Recently, conditional motion generation through other modalities, such as text [2, 8, 9, 17, 22, 28, 77], music [50, 68, 94], speech [32], or action labels [27, 75], has gained more attention. Be" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.89, + 0.121 + ], + "angle": 0, + "content": "low, we focus on speech-driven motion generation methods, since they are the most relevant to our work." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.135, + 0.788, + 0.151 + ], + "angle": 0, + "content": "2.2. Gesture Generation from Speech" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.16, + 0.892, + 0.386 + ], + "angle": 0, + "content": "Rule-based gesture synthesis. Embodied conversational agents (ECA) are designed to interact and communicate with humans. Using the Behavior Markup Language (BML) [44] one can build rule-based systems for humanoids based on predefined behaviors [80]. This is used for completion of a storytelling task in an expressive manner [45]. The BEAT rule-based toolkit [14] enables adding non-verbal behavior on top of a pre-animated figure. Thiebaux et al. [92] develop an ECA by using procedural animation techniques and keyframe interpolation. Marsella et al. [63] design a generalized rule-based agent to generate expressions, eye gaze, and gestures from speech. Each of these approaches are based on non-trainable, rule-based techniques that may require substantial manual modelling effort to adapt to new tasks." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.388, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Data-driven gesture synthesis. More recently, data-driven methods have superseded rule-based systems. Yoon et al. [104] use a fusion of text, audio and upper body gestures to learn an upper body gesture avatar, but can only control the style of individual speakers by sampling from their latent space. SpeechGestureMatching [32] generates 3D facial meshes and 3D keypoints of the body and hands from speech, but the outputs are separated and the method does not provide control over the generations. QPGesture [98] uses phase to better align the generated 3D skeleton-based gesturing avatars with the audio input. Ginosar et al. [23] and Diverse-3D-Hand-Gesture-Prediction [82] generate hand and arm motions only. Audio2Gestures [48] encode motion and audio to a low-dimensional latent space and generate gestures. SEEG [53] aims to generate gestures that align well with the semantics of the speech. DiffTTSG [67] regresses speech and gestures at the same time, joining the two modalities in a single system. DiffGAN [3] retargets gestures across speakers in a low-resource setting. The GENEA challenge [105] tackles gesticulation from speech alone using the Talking-with-Hands dataset [46]. Gesture2Vec [99] uses a machine translation model to translate text into gesture chunks and output full sequences using such quantized representations. TalkSHOW [101] uses a VQ-VAE to generate 3D human bodies gesturing with facial expressions from speech segments, but in an uncontrolled manner. Similarly, Co-speech gesture [60] uses an RQ-VAE to generate different gestures from speech. Alternative gesture generation from speech methods have been proposed such as reinforcement learning [91], self-supervised pretraining [40], and diffusion [67, 110]. BodyFormer [71] introduces a dataset of pseudo-groundtruth and a transformer-based method for generating gestures from speech. However, none of these methods provide explicit emotional con" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1943" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.283, + 0.106 + ], + "angle": 0, + "content": "trol over the generated motion." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.108, + 0.471, + 0.334 + ], + "angle": 0, + "content": "For controllable generation, GestureDiffuCLIP [7] incorporates multiple conditions including CLIP [83] text features, video, or motion prompts via AdaIn [37] layers to generate gestures from speech, however, it does not allow explicit control over the emotion conveyed by the driving audio. ListenDenoiseAction [5] combines conformers and the DiffWave [43] architecture to generate gestures that can be controlled by a style vector, RhythmicGesticulator [6] disentangles the latent space into a vector related to the semantics of the gesture and one related to the subtle variations, while DisCo [54] models content and rhythm. StyleGestures [4] adapts MoGlow [34], demonstrating limited control over some motion attributes like the speed and expressiveness of gestures. DiffuseStyleGesture [97] uses diffusion to generate diverse gestures from speech." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.348, + 0.245, + 0.362 + ], + "angle": 0, + "content": "2.3. Emotion Control" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.372, + 0.473, + 0.904 + ], + "angle": 0, + "content": "Emotion classification and control has been little studied in 3D human motion generation with only a few methods using skeletal motion in multi-class classification. Ghaleb et al. [20] employ a spatio-temporal graph convolution network to classify gestures into four classes: preparation, stroke, retraction, and neutral. Li et al. [52], on the other hand, use hidden Markov models for emotion classification of human movement mocap data. Karras et al. [38] learn face animations of a single actor, and test their method on different tasks by modifying the latent vectors. However, there is no disentanglement mechanism, and they do not model the synchronization of the emotion with the facial motions. Recently, EmoTalk [74], animates emotional 3D faces from speech input with control over the emotion intensity and EMOTE [18] disentangles emotion and speech to allow emotion editing at test time. However, models solely intended for facial tasks like lipSyncing and capturing expressions might not smoothly adapt to the complexity of whole-body movements and distinct articulation. Regarding emotion-conditioned motion generation, Aberman et al. [1] show style-transfer from video data to motion and provide some style-based control, but do not address speech-driven emotional gestures. Similarly, the ZeroEGGs [21] dataset contains some emotional gesture controls but also includes more generic styles of motion. The method requires the input of arbitrary frames of desired motion to encode a style, thereby relying on motions and speech as conditions during inference. Text-driven emotional gesticulation, as explored by Bhattacharya et al. [11, 12], emphasizes the generation of gestures based on textual cues, incorporating additional conditions such as speech, speaker ID, seed poses, as well as valence, arousal, and dominance triplets. However, these approaches do not provide the means to distill explicit emotion features, limiting free control over the generated gestures. Closer to" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.32 + ], + "angle": 0, + "content": "our work, EMoG [102] incorporates emotion cues from the BEAT dataset [55] to generate improved gesture quality without explicit emotion control. EmotionGesture [81] uses a TED Emotion Dataset and BEAT to incorporate emotion features in gesture generation and generate emotional gestures. Although they can generate emotional gestures, their method is not end-to-end and has no explicit motion control. Specifically, it uses an emotion-conditioned VAE after training to acquire diverse emotion features that are used to generate gestures without guarantees and control over emotion types. Wu et al. [96] introduce the first multi-cultural gesture dataset containing 200 individuals of 10 different cultures. In contrast to prior work, we explicitly control the emotions conveyed by the generated gestures solely through emotional speech without relying on additional conditions." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.33, + 0.593, + 0.346 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.355, + 0.895, + 0.492 + ], + "angle": 0, + "content": "The AMUSE pipeline consists of two separately trained networks. The audio disentanglement module, which encodes input speech into latent vectors for content, emotion, and style is described in Sec. 3.2. The main architecture is described in Sec. 3.3. It consists of a 3D human motion prior coupled with a latent diffusion model. It takes random noise (or partially denoised latent vectors) on the input and outputs a human motion sequence. We introduce broader applications in gesture editing in Sec. 3.4." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.499, + 0.849, + 0.515 + ], + "angle": 0, + "content": "3.1. Preliminary: Expressive 3D Body Model" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.522, + 0.895, + 0.72 + ], + "angle": 0, + "content": "SMPL-X [73] is a 3D model of the body surface. SMPL-X is defined as function \\( M(\\boldsymbol{\\beta}, \\boldsymbol{\\theta}, \\boldsymbol{\\psi}) \\) that produces a 3D body mesh. It is parameterized by identity shape \\( \\boldsymbol{\\beta} \\in \\mathbb{R}^{300} \\), pose \\( \\boldsymbol{\\theta} \\in \\mathbb{R}^{J \\times 3} \\) including finger articulation for rotations around \\( J \\) joints, and facial expression \\( \\boldsymbol{\\psi} \\in \\mathbb{R}^{100} \\). We adopt the continuous 6D rotation representation for training following Zhou et al. [108], making \\( \\boldsymbol{\\theta} \\in \\mathbb{R}^{J \\times 6} \\). Given pose parameters and any shape parameter, we can obtain body mesh vertices \\( V \\) using the differentiable SMPL-X layer [73]. As the focus of our paper is on synthesizing body gestures and not locomotion, we disregard 8 joints that correspond those of the lower body joint poses, leaving \\( J = 47 \\). Further, we omit the facial expression parameters, i.e., set \\( \\boldsymbol{\\psi} = \\mathbf{0} \\)." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.727, + 0.779, + 0.743 + ], + "angle": 0, + "content": "3.2. Speech Disentanglement Model" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.895, + 0.901 + ], + "angle": 0, + "content": "Architecture. The goal of the this model is to factor an input speech into three disentangled latent representations, one for content (i.e., the words spoken), one for emotion, and one for personal style. To do so, we devise a specialized encoder-decoder architecture with three separate encoders, one for each latent space. We denote the encoders as: \\( E_{c}(a) = c \\), \\( E_{e}(a) = e \\), \\( E_{s}(a) = s \\), where \\( a \\) is the input filterbank, \\( c \\), \\( e \\) and \\( s \\) denote the latent vectors for content, emotion and style and \\( E_{c} \\), \\( E_{e} \\) and \\( E_{s} \\) are their encoders. The architecture of the three encoders follows the" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1944" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.102, + 0.088, + 0.895, + 0.265 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.273, + 0.895, + 0.349 + ], + "angle": 0, + "content": "Figure 2. Training. We train the motion prior \\((\\mathcal{P}_E, \\mathcal{P}_D)\\) and the latent denoiser \\(\\Delta\\) jointly, while keeping the audio encoding networks frozen. In the forward pass, we take an input audio \\(a^{1:T}\\) and pose sequence \\(m^{1:T}\\). Firstly, we do a forward pass of \\(m^{1:T}\\) through \\(\\mathcal{P}_E\\) and \\(\\mathcal{P}_D\\) and compute \\(\\mathcal{L}_{rec}\\), \\(\\mathcal{L}_{Vrec}\\), and \\(\\mathcal{L}_{KL}\\). Then, we apply the diffusion process to a gradient-detached sg \\([z_m]\\) obtaining the noisy \\(z_m^{(D)}\\) which is then denoised with \\(\\Delta\\) and \\(\\mathcal{L}_{LD}\\) is computed. Finally, we use \\(\\Delta\\) to fully denoise \\(z_n\\) into gradient-detached sg \\([z_{\\tilde{m}}]\\), further decode \\(\\tilde{m}^{1:T}\\) using \\(\\mathcal{P}_D\\), and compute \\(\\mathcal{L}_{align}\\) and \\(\\mathcal{L}_{Valign}\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.371, + 0.47, + 0.476 + ], + "angle": 0, + "content": "design by Gong et al. [25, 26] (i.e., leveraging the DeiT visual transformer [93] adapted for processing filterbank images extracted from the input audio). The decoder takes the three latent vectors and produces a reconstructed filterbank. Formally \\( D(c, e, s) = \\widehat{a} \\), where \\( \\widehat{a} \\) denotes the reconstructed filterbank. The decoder architecture consists of a fusion module and transformer-encoder layers." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.477, + 0.471, + 0.691 + ], + "angle": 0, + "content": "Training. The audio module is trained with a multiple loss terms that ensure that the three latent spaces are properly disentangled. In addition to the standard autoencoder reconstruction loss, we also employ three cross-reconstruction losses, in which we enforce the correct reconstruction of the audio signal where we modify one of the content, style or emotion latents. Additionally, we employ three loss terms on the latent vector predictions – namely emotion and style classification losses over \\( e \\) and \\( s \\), and a content similarity loss between pairs of two content latent vectors extracted from audios that have the same spoken content. For a detailed description of the encoder-decoder architecture, a formal definition of the loss functions and a detailed description of the training process please refer to the Sup. Mat." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.697, + 0.32, + 0.712 + ], + "angle": 0, + "content": "3.3. Gesture Generation Model" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.72, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Motion prior. Similar to [15, 76], our motion prior network is a VAE transformer architecture with encoder \\(\\mathcal{P}_E\\) and decoder \\(\\mathcal{P}_D\\). Specifically, both \\(\\mathcal{P}_E\\) and \\(\\mathcal{P}_D\\) follow a U-Net-like [85] structure with skip connections between transformer blocks (see Sup. Mat. for details). The positional embeddings are learnable and injected into each multi-head attention layer, following the design of Carion et al. [13]. Formally, the encoder takes a sequence of \\(T\\) frames of the SMPL-X pose vectors \\(m^{1:T} \\in \\mathbb{R}^{6J \\times T}\\) and the first two tokens of its output, \\(\\mu \\in \\mathbb{R}^{d_m}\\) and \\(\\Sigma \\in \\mathbb{R}^{d_m \\times d_m}\\) are used to extract the motion latent \\(z_m \\in \\mathbb{R}^{d_m}\\) via the reparametrization trick. The decoder takes zero positional encodings as" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.371, + 0.892, + 0.415 + ], + "angle": 0, + "content": "query input and the motion latent is fed as memory to every cross-attention transformer layer, producing the reconstructed motion \\(\\hat{m}^{1:T}\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.416, + 0.892, + 0.477 + ], + "angle": 0, + "content": "Diffusion process. The forward diffusion process is similar to [36, 69]. We employ fixed variance and linearly scaled noise scheduler. We add noise to the motion latent \\( z_{m} \\) for \\( D \\) diffusion timesteps to obtain \\( z^{(D)} \\) following:" + }, + { + "type": "equation", + "bbox": [ + 0.529, + 0.485, + 0.862, + 0.505 + ], + "angle": 0, + "content": "\\[\nq \\left(z _ {m} ^ {\\left(t _ {d}\\right)} \\mid z _ {m} ^ {(0)}\\right) = \\mathcal {N} \\left(z _ {m} ^ {\\left(t _ {d}\\right)}; \\sqrt {\\bar {\\alpha} _ {t _ {d}}} z _ {m} ^ {(0)}, (1 - \\bar {\\alpha} _ {t _ {d}}) \\mathbf {I}\\right),\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.515, + 0.891, + 0.545 + ], + "angle": 0, + "content": "with \\(\\alpha_{t_d} = 1 - \\beta_{t_d}\\), \\(\\bar{\\alpha}_{t_d} = \\prod_{s=1}^{t_d} \\alpha_s\\), and \\(\\beta_{t_d}\\) denotes diffusion process variance." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.546, + 0.892, + 0.7 + ], + "angle": 0, + "content": "Conditional denoising process. The denoising process consists of iteratively denoising a conditioned noisy motion latent vector to obtain the denoised motion latent \\( z_{\\tilde{m}^{1:T}} \\). Our denoiser \\( \\Delta \\) is a latent variable model [84] and its architecture is similar to the U-Net-like structure of the motion prior encoder \\( \\mathcal{P}_E \\). The input of the model is a concatenation of: \\( z_m^{(t_d)} \\), \\( \\mathrm{SE}(t_d) \\), \\( c, e, s \\in \\mathbb{R}^{256} \\), where \\( \\mathrm{SE}(t_d) \\) is a sinusoidal positional encoding of diffusion timestep \\( t_d \\) as defined in [36]. \\( \\Delta \\) iteratively denoises through each reversed diffusion step:" + }, + { + "type": "equation", + "bbox": [ + 0.575, + 0.707, + 0.816, + 0.727 + ], + "angle": 0, + "content": "\\[\nz _ {m} ^ {(t _ {d} - 1)} = \\Delta \\left(\\left[ z _ {m} ^ {(t _ {d})}, \\operatorname {S E} \\left(t _ {d}\\right), c, e, s \\right]\\right).\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.735, + 0.892, + 0.902 + ], + "angle": 0, + "content": "Training. We optimize the motion prior and the latent denoiser jointly to ensure audio-motion latent code alignment during conditional fusion in the denoising process using a 3-step forward pass through the gesture generation model. First, following standard VAE practice, we reconstruct \\(\\hat{m}^{1:T}\\) by the motion prior forward pass. As shown in Fig. 2, we then disable gradient calculation in \\(\\mathcal{P}_E\\) to infer the intermediate motion latent \\(\\mathrm{sg}[z_m]\\), which serves as input to the denoiser. At this stage, we obtain the denoiser noise prediction, \\(\\delta\\) and use to compute the diffusion model gradients. Finally, in the third step we com" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1945" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.09, + 0.468, + 0.154 + ], + "angle": 0, + "content": "pute \\(\\tilde{m}^{1:T} = \\mathcal{P}_D(\\mathrm{sg}[z_{\\tilde{m}}])\\), where \\(z_{\\tilde{m}}\\) is obtained by iteratively using the \\(\\Delta\\) to obtain a fully denoised latent from \\(z_n^{(t_D)} \\sim \\mathcal{N}(\\mathbf{0}, \\mathbf{I})\\). We indicate computations done without gradients with a stop-gradient operation \\(\\mathrm{sg}[\\cdot]\\)." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.155, + 0.468, + 0.215 + ], + "angle": 0, + "content": "Losses. To train the motion prior, we include the standard VAE losses, namely the reconstruction loss on pose parameters \\(\\mathcal{L}_{rec}\\) and on vertex coordinates \\(\\mathcal{L}_{Vrec}\\) using the smooth L1 metric introduced in [24], which we denote as \\(L_1^s\\):" + }, + { + "type": "equation", + "bbox": [ + 0.092, + 0.222, + 0.455, + 0.243 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {r e c} = L _ {1} ^ {s} (m ^ {1: T}, \\hat {m} ^ {1: T}), \\quad \\mathcal {L} _ {V r e c} = L _ {1} ^ {s} (V ^ {1: T}, \\hat {V} ^ {1: T}),\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.251, + 0.468, + 0.311 + ], + "angle": 0, + "content": "where the root-centered vertices \\(V\\) are obtained by feeding in pose parameters \\(m\\) to a differentiable SMPL-X layer (without learnable parameters) and a mean shape \\(\\beta = \\vec{0}\\). The KL divergence loss of the motion prior is:" + }, + { + "type": "equation", + "bbox": [ + 0.107, + 0.32, + 0.438, + 0.362 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {K L} = \\frac {1}{2} \\left[ \\sum_ {i = 1} ^ {z} \\left(\\mu_ {i} ^ {2} + \\sigma_ {i} ^ {2}\\right) - \\sum_ {i = 1} ^ {z} \\left(l o g \\left(\\sigma_ {i} ^ {2}\\right) + 1\\right) \\right].\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.37, + 0.468, + 0.429 + ], + "angle": 0, + "content": "To ensure the alignment of the diffusion-generated motions and the input audio, we apply the alignment reconstruction loss on the inferred motion pose parameters and the vertex coordinates:" + }, + { + "type": "equation", + "bbox": [ + 0.08, + 0.438, + 0.466, + 0.459 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {a l i g n} = L _ {1} ^ {s} (m ^ {1: T}, \\tilde {m} ^ {1: T}), \\quad \\mathcal {L} _ {V a l i g n} = L _ {1} ^ {s} (V ^ {1: T}, \\tilde {V} ^ {1: T}).\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.467, + 0.468, + 0.497 + ], + "angle": 0, + "content": "Finally, we utilize the objective similar to [15, 36, 84] to supervise the denoiser:" + }, + { + "type": "equation", + "bbox": [ + 0.125, + 0.506, + 0.421, + 0.535 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {L D} = \\left\\| \\delta^ {(t _ {d})} - \\Delta (z _ {m} ^ {(t _ {d})}, \\mathrm {S E} (t _ {d}), c, e, s) \\right\\| _ {2} ^ {2},\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.545, + 0.468, + 0.589 + ], + "angle": 0, + "content": "where \\(\\delta^{(t_d)}\\) is the noise vector sampled from \\(\\mathcal{N}(\\mathbf{0},\\mathbf{I})\\) in the corresponding diffusion step \\(t_d\\). The combined gesture model loss is:" + }, + { + "type": "equation", + "bbox": [ + 0.078, + 0.602, + 0.468, + 0.619 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {g e s} = \\mathcal {L} _ {r e c} + \\mathcal {L} _ {V r e c} + \\mathcal {L} _ {K L} + \\mathcal {L} _ {a l i g n} + \\mathcal {L} _ {V a l i g n} + \\mathcal {L} _ {L D}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.628, + 0.468, + 0.718 + ], + "angle": 0, + "content": "Inference. We employ DDIM [89] to infer high quality conditional motion samples with a small number of denoising timesteps. During inference we draw a sample vector from \\(\\mathcal{N}(\\mathbf{0},\\mathbf{I})\\) to iteratively denoise in reversed timesteps. The denoised sample is then passed through the decoder \\(\\mathcal{P}_D(z_{\\tilde{m}^{1:T}})\\) to obtain motion \\(\\tilde{m}^{1:T}\\)." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.727, + 0.236, + 0.743 + ], + "angle": 0, + "content": "3.4. Gesture Editing" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.75, + 0.468, + 0.901 + ], + "angle": 0, + "content": "Due to the disentangling of the inputs, AMUSE achieves semantic gesticulation control using two driving input audios. Specifically, given two input audio signals \\( a_1 \\) and \\( a_2 \\), we extract their latent representations of content \\( c_1, c_2 \\), emotion \\( e_1, e_2 \\), and style \\( s_1, s_2 \\). Then, we simply initialize the denoising procedure of \\( \\Delta \\) with the triplet \\( (c_1, e_2, s_1) \\), generating the gesture with the content and style of \\( a_1 \\) but the emotion of input audio \\( a_2 \\). Similarly, instead of emotion we can also change the gesticulation style to that of the speaker of \\( a_2 \\) by initializing with \\( (c_1, e_1, s_2) \\)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.09, + 0.722, + 0.107 + ], + "angle": 0, + "content": "4. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.116, + 0.892, + 0.388 + ], + "angle": 0, + "content": "MoCap data preparation. The BEAT [55] mocap sequences, captured in a Vicon system at \\(120\\mathrm{Hz}\\), are downsampled to \\(30\\mathrm{Hz}\\) and processed using \\(\\mathrm{MoSh}++\\) [58, 62] to obtain SMPL-X parameters. Given a sequence of 3D mocap marker positions, we jointly optimize SMPL-X shape and pose parameters, 3D body translation, and embedding of the mocap markers in the SMPL-X surface. Once processed, the sequences are then divided according to the emotion annotations in the BEAT dataset. We use sequences of English speaking subjects in monologue speaking style for training and evaluating AMUSE. For each sequence we draw \\(m^{1:L}\\) at 30 FPS and concatenate with audio content \\(c\\), emotion \\(e\\), and style \\(s\\) latent vectors. Then, we segment it to 10-sec windows \\(T\\), beginning from the timestamp 0 and discarding additional unaligned information at the end. This preprocessing choice allows us to train transformer networks without masking. We provide additional data processing information in the Sup. Mat." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.39, + 0.892, + 0.585 + ], + "angle": 0, + "content": "Audio preprocessing. We use audio sequences belonging to eight categorical emotion labels (neutral, happy, angry, sad, contempt, surprise, fear, and disgust). Each audio chunk of 10s is converted into a filter bank with 128 melfrequency bins with a 25ms Hamming frame window and 10ms frame shift. We mask each sample with a maximum length of 24 in the frequency domain and a maximum length of 96 in the time domain, employing Park et al. [72]. Following [25, 26], we standardize the filter bank and augment it via noise injection and circular shifting. Before feeding in our speech disentanglement model, each filter bank is split into a sequence of fixed 1209 patches of \\(16 \\times 16\\) each having 6 units overlap in frequency and time domain." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.586, + 0.892, + 0.799 + ], + "angle": 0, + "content": "Motion prior. The motion prior is a VAE encoder-decoder with 9 layers and 4 heads, following Chen et al. [15]. The encoder-decoder is a U-Net-like transformer with residual connections. Learnable positional embeddings are injected in each multi-head attention layer. We have a linear projection at the start and the end of our motion prior network. The KL divergence term is weighted with a factor of \\(1e - 4\\). Denoiser. The denoiser follows the same network architecture as our prior encoder. The hidden dimension of all transformer layers is 1024. We use 1000 diffusion steps \\(D\\) during training and 50 during inference. Noise betas are in range \\([0.00085, 0.012]\\). We jointly optimize the prior and denoiser networks for 5000 epochs with batch size of 64, learning rate 0.0001, and the AdamW optimizer [59]." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.815, + 0.632, + 0.831 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Speech disentanglement model. We evaluate the performance of the speech disentanglement model quantitatively using classification accuracy and F1 scores on emotion and style. The accuracy is computed as average scores" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1946" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.468, + 0.167 + ], + "angle": 0, + "content": "for all 8 emotion as well style categories that are part of the test dataset. The emotion and style accuracy is \\(91.53\\%\\) and \\(96.06\\%\\), respectively. The emotion F1 score and style F1 scores are 0.914 and 0.960, respectively. See the Sup. Mat. for ablations and a detailed metric analysis." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.168, + 0.47, + 0.424 + ], + "angle": 0, + "content": "Gesture generation model. We evaluate the performance of our gesture generation model quantitatively, qualitatively, and perceptually against following methods: TalkSHOW [101] and the re-implementation of Habibie et al. [31] provided by the TalkSHOW authors in the official TalkSHOW release [100], DiffuseStyleGesture (DSG) [97], MoGlow [34], and CaMN [55]. Additionally, we adapt TalkSHOW to include categorical emotion labels as input along with the existing architecture that only allows one-hot encodings of personal style. We then retrain it on our training data. We refer to it as TalkSHOW-BEAT. There are some concurrent works [5, 7, 56], which introduce methods for gesture generation from speech, however, direct comparison is hindered by the unavailability of released code our task. Refer to the Sup. Mat. for the ablation experiments, the emotion and style editing experiments, and their quantitative evaluation." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.435, + 0.298, + 0.451 + ], + "angle": 0, + "content": "5.1. Quantitative Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.458, + 0.468, + 0.593 + ], + "angle": 0, + "content": "To quantitatively evaluate our method's gesture generations and edited gesture generations, we train a transformer-based encoder architecture (denoted as \\(M\\)) similar to Petrovich et al. [76] in an autoencoder setting, where we append a CLS token at the beginning of the motion sequence. \\(M\\) is trained with a cross-entropy emotion classification objective applied to the output CLS token. We train \\(M\\) on the BEAT training dataset and use its features to compute the following metrics: (1) Fréchet gesture distance (FGD): We" + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.617, + 0.468, + 0.761 + ], + "angle": 0, + "content": "
MethodSRGR↑BA↑FGD↓Div→GAa↑
GT—0.83—27.8364.04
Ours0.360.81388.6325.0646.76
Ours-EmoEditb—0.79792.5824.6834.18
TalkSHOW-BEAT0.310.64808.9924.1622.71
TalkSHOW [101]0.300.60762.1523.1929.41
DSG [97]0.230.40763.1019.7722.70
Habibie et al. [31]0.230.39809.1721.3416.67
MoGlow [34]0.210.351097.0319.5016.62
CaMN [55]0.210.391063.8718.9014.17
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.081, + 0.764, + 0.273, + 0.776 + ], + "angle": 0, + "content": "a GA is average of all 8 emotions." + }, + { + "type": "table_footnote", + "bbox": [ + 0.083, + 0.777, + 0.468, + 0.803 + ], + "angle": 0, + "content": "b GA for these are average accuracy for all generations with 7 edited audio sequences." + }, + { + "type": "list", + "bbox": [ + 0.081, + 0.764, + 0.468, + 0.803 + ], + "angle": 0, + "content": null + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.814, + 0.468, + 0.896 + ], + "angle": 0, + "content": "Table 1. Gesture quantitative results. We compare our methods against several SOTA methods using metrics explained in Sec. 5.1. We observe that AMUSE outperforms in all scores compared to baseline methods. Additionally, AMUSE-EmoEdit outperforms in Beat Align, Diversity, and Gesture Emotion Accuracy scores compared to the baseline methods." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.439 + ], + "angle": 0, + "content": "follow [88, 103, 104] to compute the feature distance between generated and ground truth motion features. (2) Gesture diversity (Div): Similarly to Chen et al. [15], we compute variance across generated features. (3) Gesture emotion accuracy (GA): We report top-1 emotion classification accuracy predicted by a classifier trained on the motion \\( M \\)-predicted latents. (4) Beat align (BA): We follow [51, 55], to evaluate the motion-speech correlation in terms of the similarity between the kinematic motion beats and speech audio beats. The kinematic motion beats are directly computed from the generated motion sequences. (5) Semantic-Relevant Gesture Recall (SRGR): We follow Liu et al. [55], to evaluate the semantic relevancy of gestures with GT motion. We use the ground truth semantic scores to compute this metric. The scores are obtained from the BEAT authors, representing a continuous score on a scale 0-1 per gesture style for 4 gesture semantic categories: beat, deictic, iconic, and metaphoric. While comparing with methods that output coarse skeletal data (DSG [97], MoGlow [34], and CaMN [55]), we convert the skeleton motion data into the SMPL-X axis angle representation. For details on the architectures and training of \\( M \\), and the losses, please refer to the Sup. Mat." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.463, + 0.892, + 0.901 + ], + "angle": 0, + "content": "We prepare the evaluation data by randomly selecting 72 unique motion sequences each of length 10s and comprising 8 emotions across test subjects and compute the aforementioned metrics. We use 9 sequences for each emotion per subject. The results are reported in Tab. 1. All best scores are highlighted in green and second best in blue. AMUSE outperforms the baseline methods in all given metrics. To validate the performance of gesture emotion editing, we also report the same metrics for the emotion editing task (Ours-EmoEdit). During inference, the input style and content latents are extracted from neutral-emotion audio, while the emotion latent comes from a different audio of different emotion. These emotional edits offer numerous possibilities, allowing for transitions from any to any emotion. Tab. 1 shows the average for editing from neutral to other emotions. Since we require the GT gesture semantics score to compute SRGR metric, it is not possible to compute the SRGR for the synthetic edited-emotion gestures as they are not part of the original BEAT dataset. Ours-EmoEdit outperforms the baseline methods in BA, Div, and GA metrics. This demonstrates the capability of our model to maintain highly discriminative cues when switching between different emotions. TalkSHOW-BEAT has the second best score for SRGR whereas TalkSHOW demonstrates second best FGD score. Although, our model and ours-EmoEdit show improvements over the baseline methods, GT motions have higher diversity, Beat alignment score, and are easier to classify than generations of AMUSE, highlighting the challenging nature of the problem." + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "1947" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.078, + 0.087, + 0.887, + 0.236 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.239, + 0.893, + 0.269 + ], + "angle": 0, + "content": "Figure 3. Qualitative comparison across all emotions. We evaluate generation on different test audios. AMUSE exhibits well-synchronized beat gestures and consistently produces gestures that accurately convey the emotional content expressed in the input speech." + }, + { + "type": "image", + "bbox": [ + 0.079, + 0.289, + 0.473, + 0.594 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.602, + 0.47, + 0.631 + ], + "angle": 0, + "content": "Figure 4. Qualitative comparison with baseline methods. The speech segment describes intense angry speech." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.663, + 0.289, + 0.679 + ], + "angle": 0, + "content": "5.2. Qualitative Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.689, + 0.473, + 0.901 + ], + "angle": 0, + "content": "Comparison with baseline methods. In Fig. 4, we demonstrate comparison with baseline methods that output a 3D body mesh: Habibie et al. [31], TalkSHOW [101], TalkSHOW-BEAT, and the BEAT ground truth (GT) [55]. We observe that AMUSE generates gestures that are semantically closer to the speech content and produces expressive emotional gestures closer to the perceived emotion. For example, the GT motion exhibits anger when saying \"put them into detention\". AMUSE demonstrates tense posture and aggressive movements comparable with the ground truth data and accurate synchronization with the spoken words. TalkSHOW [101] and Habibie et al. [31] exhibit limited movement and display inferior and static gestures on test audios as seen in the last two rows" + }, + { + "type": "image", + "bbox": [ + 0.596, + 0.29, + 0.831, + 0.365 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.375, + 0.892, + 0.404 + ], + "angle": 0, + "content": "Figure 5. Qualitative evaluation of diverse generations. Multiple generations overlayed." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.431, + 0.892, + 0.476 + ], + "angle": 0, + "content": "of Fig. 4. TalkSHOW-BEAT slightly outperforms other baseline methods by demonstrating enhanced synchronized gestures, but it still does not perform as well as AMUSE." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.477, + 0.892, + 0.507 + ], + "angle": 0, + "content": "Diverse emotional gestures. In Fig. 5, our probabilistic model can generate diverse gestures for same input audio." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.508, + 0.893, + 0.674 + ], + "angle": 0, + "content": "Emotional gesture generation. In Fig. 3 AMUSE demonstrates strong correlation with the spoken utterances as well as different emotions. We observe that our model is able to correlate semantic words to associated gestures. For example, gestures demonstrate forceful actions and tense stance with angry audio \"normally get angry\" whereas it generates lowered and calm hand positions for sad audio \"I went to collect\". Similarly, our generations show hands that are closer to body for fearful audio \"in a dangerous situation\" while widely open expressing astonishment for happy and surprised audio \"people pass by\" and \"Wow! You are here\"." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.674, + 0.892, + 0.75 + ], + "angle": 0, + "content": "Emotion editing. We use two audio streams of a female subject for neutral and sad emotion. This experiment edits the subject's gesture style from moderately controlled hand movements to a sad style with lethargic posture conveying a sense of heaviness, as seen in Fig. 7 (top)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.893, + 0.901 + ], + "angle": 0, + "content": "Gesture style editing. We use audio streams of two male subjects for the happy (ID - 13) and angry (ID - 2) emotion. With the emotion, style and content latent fusion mechanism from two driving audio streams, AMUSE is able to adapt the male (ID - 13) subject's body gestures from being close to their body to more open with squared tightened shoulders, expressing a shift from happy to angry emotions of a different subject (ID - 2), as shown in Fig. 7 (bottom). Please refer to the supplemental video for qualitative results and comparisons to additional gesture genera" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1948" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.062, + 0.092, + 0.891, + 0.156 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.158, + 0.894, + 0.187 + ], + "angle": 0, + "content": "Figure 6. The perceptual study results for gesture emotion preference (left) and synchronization with speech (right). The number of attentive participants that passed the catch trials is indicated on the right and the reported results only consider these participants." + }, + { + "type": "image", + "bbox": [ + 0.089, + 0.211, + 0.473, + 0.367 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.375, + 0.471, + 0.459 + ], + "angle": 0, + "content": "Figure 7. Gesture editing. Top: We modify style from being neutral (left) to being sad (right) by combining the emotion latent from sad audio with the content latent from neutral audio. Bottom: We transform the style from Subject 13 being happy (left) to being angry (right) by merging the content latent from happy audio with the style and emotion latents from an angry audio of Subject 2." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.487, + 0.452, + 0.502 + ], + "angle": 0, + "content": "tion methods [4, 55, 110] trained on coarse skeletal data." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.514, + 0.245, + 0.53 + ], + "angle": 0, + "content": "5.3. Perceptual Study" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.538, + 0.47, + 0.779 + ], + "angle": 0, + "content": "Design. Our perceptual study is designed as a side-by-side comparison of two gesture videos generated with the same audio as input but by two different methods (AMUSE and another model or GT). The participants are asked to rate their preference of the methods on a five-point Likert scale for \"synchronization with speech\" and \"gesture emotion appropriateness\" given the GT emotion label of the input audio. We recruit 25 participants per method-to-method comparison on Amazon Mechanical Turk. Each participant is shown 24 pairs of randomly selected test set animations, 3 per emotion (neutral, happy, angry, sad, disgust, fear, surprise, and contempt). To allow the participant to get used to the task, we discard the answers of the first three comparisons and repeat these at the end. We incorporate three catch trials and responses from participants that fail on more than one are filtered out, as shown in Fig. 6 (right)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.781, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Results. The results of the study are shown in Fig. 6. AMUSE outperforms all competing methods by a considerable margin on both tasks, suggesting that AMUSE's generations are more appropriate for both the content of the input speech and its emotion compared to the baselines. However, it must be noted that there is still a significant gap between AMUSE and the GT. Please refer to the Sup. Mat. for details about the perceptual study." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.212, + 0.757, + 0.227 + ], + "angle": 0, + "content": "5.4. Discussion and Future Work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.235, + 0.892, + 0.326 + ], + "angle": 0, + "content": "Upper-body motion. We focus on the smooth coordination between the pelvis and upper body animation for side-by-side comparisons with other methods, as all other methods primarily focus on upper body movements. Future work should include lower-body motion and locomotion as these impact the perceived emotional state of a sequence." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.326, + 0.892, + 0.4 + ], + "angle": 0, + "content": "Semantics. While the generated gestures, synchronized with the driving speech sequence, do not account for semantics such as deictic and metaphoric gestures, incorporating the text/language modality could help further improve in this direction." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.402, + 0.892, + 0.524 + ], + "angle": 0, + "content": "Facial expressions. While emotional speech-driven face animation methods [18, 74] can be combined with bodies generated from AMUSE, jointly learning to generate emotional 3D bodies from speech is a topic that needs attention. End-to-end training. Joint audio-gesture training may enhance results but requires careful loss term balancing and increased GPU memory. Therefore, we opted for separate training." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.538, + 0.619, + 0.552 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.563, + 0.892, + 0.775 + ], + "angle": 0, + "content": "We present AMUSE, a framework to generate emotional body gestures from speech. The emotions and personal styles of the synthesized gestures can be controlled, thanks to the disentanglement of content, emotion, and style directly from the speech. The latent diffusion-based framework can further generate variations of the same gesture with the same emotion. Our quantitative evaluations show that AMUSE achieves state of the art performance on a variety of metrics: diversity, gesture emotion classification accuracy, Frechet gesture distance, beat alignment score, and semantic relevant gesture recall. Finally, our perceptual study demonstrates that AMUSE generates motions that are better synchronized and better match the emotion expressed of the input speech than previous state of the art." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.789, + 0.892, + 0.887 + ], + "angle": 0, + "content": "Acknowledgments. We thank A Cseke, T McConnell and T Alexiadis for help with the perceptual study, B Pellkofer, J Piles-Contreras, and E Fritzler for IT support, and P Kulits, M Petrovich, S Kokane, and S Zojaji for proofreading and insightful discussions. This project has received funding from the European Union's Horizon 2020 research and innovation programme under the Marie Skłodowska-Curie grant agreement No. 860768." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.888, + 0.886, + 0.901 + ], + "angle": 0, + "content": "Disclosure. https://download.is.tue.mpg.de/amuse/Disclosure.txt" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1949" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.105 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.096, + 0.116, + 0.469, + 0.169 + ], + "angle": 0, + "content": "[1] Kfir Aberman, Yijia Weng, Dani Lischinski, Daniel Cohen-Or, and Baoquan Chen. Unpaired motion style transfer from video to animation. Transactions on Graphics (TOG), 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.096, + 0.173, + 0.468, + 0.214 + ], + "angle": 0, + "content": "[2] Chaitanya Ahuja and Louis-Philippe Morency. Language2Pose: Natural language grounded pose forecasting. International Conference on 3D Vision (3DV), 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.096, + 0.217, + 0.468, + 0.271 + ], + "angle": 0, + "content": "[3] Chaitanya Ahuja, Dong Won Lee, and Louis-Philippe Morency. Low-resource adaptation for personalized cospeech gesture generation. In Computer Vision and Pattern Recognition (CVPR), pages 20566-20576, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.096, + 0.274, + 0.468, + 0.341 + ], + "angle": 0, + "content": "[4] Simon Alexanderson, Gustav Eje Henter, Taras Kucherenko, and Jonas Beskow. Style-controllable speech-driven gesture synthesis using normalising flows. Computer Graphics Forum (CGF), 39(2):487-496, 2020. 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.096, + 0.345, + 0.468, + 0.399 + ], + "angle": 0, + "content": "[5] Simon Alexanderson, Rajmund Nagy, Jonas Beskow, and Gustav Eje Henter. Listen, Denoise, Action! Audio-driven motion synthesis with diffusion models. Transactions on Graphics (TOG), 42(4):1-20, 2023. 1, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.096, + 0.402, + 0.468, + 0.442 + ], + "angle": 0, + "content": "[6] Tenglong Ao, Qingzhe Gao, Yuke Lou, Baoquan Chen, and Libin Liu. Rhythmic gesticulator. Transactions on Graphics (TOG), 41(6):1-19, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.096, + 0.446, + 0.468, + 0.486 + ], + "angle": 0, + "content": "[7] Tenglong Ao, Zeyi Zhang, and Libin Liu. GestureDiffu-CLIP: Gesture diffusion model with CLIP latents. Transactions on Graphics (TOG), 2023. 1, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.096, + 0.489, + 0.468, + 0.542 + ], + "angle": 0, + "content": "[8] Nikos Athanasiou, Mathis Petrovich, Michael J. Black, and Gül Varol. TEACH: Temporal Action Compositions for 3D Humans. In International Conference on 3D Vision (3DV), 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.096, + 0.546, + 0.468, + 0.6 + ], + "angle": 0, + "content": "[9] Nikos Athanasiou, Mathis Petrovich, Michael J. Black, and Gül Varol. SINC: Spatial composition of 3D human motions for simultaneous action generation. International Conference on Computer Vision (ICCV), 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.604, + 0.468, + 0.657 + ], + "angle": 0, + "content": "[10] Emad Barsoum, John Kender, and Zicheng Liu. HPGAN: probabilistic 3d human motion prediction via GAN. In Computer Vision and Pattern Recognition Workshops (CVPRW), 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.661, + 0.468, + 0.741 + ], + "angle": 0, + "content": "[11] Uttaran Bhattacharya, Elizabeth Childs, Nicholas Rewkowski, and Dinesh Manocha. Speech2affectivegestures: Synthesizing co-speech gestures with generative adversarial affective expression learning. In ACM International Conference on Multimedia (MM), 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.746, + 0.468, + 0.813 + ], + "angle": 0, + "content": "[12] Uttaran Bhattacharya, Nicholas Rewkowski, Abhishek Banerjee, Pooja Guhan, Aniket Bera, and Dinesh Manocha. Text2gestures: A transformer-based network for generating emotive body gestures for virtual agents. In IEEE Virtual Reality and 3D User Interfaces, VR, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.817, + 0.468, + 0.884 + ], + "angle": 0, + "content": "[13] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In European Conference on Computer Vision (ECCV), pages 213-229. Springer, 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.888, + 0.468, + 0.901 + ], + "angle": 0, + "content": "[14] Justine Cassell, Hannes Hogni Vilhjalmsson, and Timo" + }, + { + "type": "list", + "bbox": [ + 0.088, + 0.116, + 0.469, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.541, + 0.094, + 0.892, + 0.12 + ], + "angle": 0, + "content": "thy Bickmore. BEAT: The Behavior Expression Animation Toolkit, pages 163-185. Springer, 2004. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.122, + 0.892, + 0.176 + ], + "angle": 0, + "content": "[15] Xin Chen, Biao Jiang, Wen Liu, Zilong Huang, Bin Fu, Tao Chen, and Gang Yu. Executing your commands via motion diffusion in latent space. In Computer Vision and Pattern Recognition (CVPR), pages 18000-18010, 2023. 4, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.178, + 0.892, + 0.23 + ], + "angle": 0, + "content": "[16] Enric Corona, Albert Pumarola, G. Alenyà, and F. Moreno-Noguer. Context-aware human motion prediction. In Computer Vision and Pattern Recognition (CVPR), pages 6990–6999, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.233, + 0.892, + 0.287 + ], + "angle": 0, + "content": "[17] Rishabh Dabral, Muhammad Hamza Mughal, Vladislav Golyanik, and Christian Theobalt. Mofusion: A framework for denoising-diffusion-based motion synthesis. In Computer Vision and Pattern Recognition (CVPR), 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.289, + 0.892, + 0.37 + ], + "angle": 0, + "content": "[18] Radek Daneček, Kiran Chhatre, Shashank Tripathi, Yandong Wen, Michael Black, and Timo Bolkart. Emotional speech-driven animation with content-emotion disentangle-ment. In International Conference on Computer Graphics and Interactive Techniques in Asia (SIGGRAPH ASIA), 2023. 2, 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.373, + 0.892, + 0.398 + ], + "angle": 0, + "content": "[19] Paul Ekman. Facial expression and emotion. American Psychologist, 48(4):384-392, 1993. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.401, + 0.892, + 0.468 + ], + "angle": 0, + "content": "[20] Esam Ghaleb, Ilya Burenko, Marlou Rasenberg, Wim Pouw, Peter Uhrig, Judith Holler, Ivan Toni, Asli Özyurek, and Raquel Fernández. Co-speech gesture detection through multi-phase sequence labeling. In Winter Conference on Applications of Computer Vision (WACV), 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.47, + 0.892, + 0.524 + ], + "angle": 0, + "content": "[21] Saeed Ghorbani, Ylva Ferstl, Daniel Holden, Nikolaus F. Troje, and Marc-Andre Carbonneau. Zeroeggs: Zero-shot example-based gesture generation from speech. Computer Graphics Forum (CGF), 42(1):206-216, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.526, + 0.892, + 0.58 + ], + "angle": 0, + "content": "[22] Anindita Ghosh, Noshaba Cheema, Cennet Oguz, Christian Theobalt, and Philipp Slusallek. Synthesis of compositional animations from textual descriptions. In International Conference on Computer Vision (ICCV), 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.582, + 0.892, + 0.636 + ], + "angle": 0, + "content": "[23] Shiry Ginosar, Amir Bar, Gefen Kohavi, Caroline Chan, Andrew Owens, and Jitendra Malik. Learning individual styles of conversational gesture. In Computer Vision and Pattern Recognition (CVPR), pages 3497-3506, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.638, + 0.892, + 0.675 + ], + "angle": 0, + "content": "[24] Ross B. Girshick. Fast R-CNN. In International Conference on Computer Vision (ICCV), pages 1440-1448, 2015. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.679, + 0.892, + 0.718 + ], + "angle": 0, + "content": "[25] Yuan Gong, Yu-An Chung, and James R. Glass. AST: audio spectrogram transformer. In *Interspeech* 2021, pages 571-575, 2021. 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.721, + 0.892, + 0.775 + ], + "angle": 0, + "content": "[26] Yuan Gong, Cheng-I Lai, Yu-An Chung, and James R. Glass. SSAST: self-supervised audio spectrogram transformer. In AAAI Conference on Artificial Intelligence, pages 10699-10709, 2022. 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.778, + 0.892, + 0.844 + ], + "angle": 0, + "content": "[27] Chuan Guo, Xinxin Zuo, Sen Wang, Shihao Zou, Qingyao Sun, Annan Deng, Minglun Gong, and Li Cheng. Action2Motion: Conditioned generation of 3d human motions. In ACM International Conference on Multimedia (MM), 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.847, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[28] Chuan Guo, Shihao Zou, Xinxin Zuo, Sen Wang, Wei Ji, Xingyu Li, and Li Cheng. Generating diverse and natural 3D human motions from text. In Computer Vision and Pattern Recognition (CVPR), 2022. 2" + }, + { + "type": "list", + "bbox": [ + 0.511, + 0.094, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1950" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.093, + 0.47, + 0.134 + ], + "angle": 0, + "content": "[29] Nejla Gürefe. The role of gestures in mathematical discourse of hard-hearing students: Prism example. Acta Destructica Napocensia, 11:125-140, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.136, + 0.471, + 0.192 + ], + "angle": 0, + "content": "[30] I. Habibie, Daniel Holden, Jonathan Schwarz, J. Yearsley, and T. Komura. A recurrent variational autoencoder for human motion synthesis. In *British Machine Vision Conference (BMVC)*, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.194, + 0.472, + 0.276 + ], + "angle": 0, + "content": "[31] Ikhsanul Habibie, Weipeng Xu, Dushyant Mehta, Lingjie Liu, Hans-Peter Seidel, Gerard Pons-Moll, Mohamed Elgharib, and Christian Theobalt. Learning speech-driven 3d conversational gestures from video. In ACM International Conference on Intelligent Virtual Agents (IVA), 2021. 1, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.279, + 0.472, + 0.363 + ], + "angle": 0, + "content": "[32] Ikhsanul Habibie, Mohamed A. Elgharib, Kripasindhu Sarkar, Ahsan Abdullah, Simbarashe Linval Nyatsanga, Michael Neff, and Christian Theobalt. A motion matching-based framework for controllable gesture synthesis from speech. International Conference on Computer Graphics and Interactive Techniques (SIGGRAPH), 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.365, + 0.47, + 0.407 + ], + "angle": 0, + "content": "[33] Félix G. Harvey, Mike Yurick, Derek Nowrouzezahrai, and Christopher Joseph Pal. Robust motion in-between. Transactions on Graphics (TOG), 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.409, + 0.47, + 0.464 + ], + "angle": 0, + "content": "[34] Gustav Eje Henter, Simon Alexanderson, and Jonas Beskow. MoGlow: Probabilistic and controllable motion synthesis using normalising flows. Transactions on Graphics (TOG), 39(4):236:1-236:14, 2020. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.467, + 0.47, + 0.508 + ], + "angle": 0, + "content": "[35] Sandra Herbert. Gesture types for functions. Mathematics Education Research Group of Australasia, pages 322-329, 2012. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.511, + 0.47, + 0.552 + ], + "angle": 0, + "content": "[36] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In Conference on Neural Information Processing Systems (NeurIPS), 2020. 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.555, + 0.47, + 0.609 + ], + "angle": 0, + "content": "[37] Xun Huang and Serge J. Belongie. Arbitrary style transfer in real-time with adaptive instance normalization. In International Conference on Computer Vision (ICCV), pages 1510-1519, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.612, + 0.47, + 0.668 + ], + "angle": 0, + "content": "[38] Tero Karras, Timo Aila, Samuli Laine, Antti Herva, and Jaakko Lehtinen. Audio-driven facial animation by joint end-to-end learning of pose and emotion. Transactions on Graphics (TOG), 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.67, + 0.47, + 0.697 + ], + "angle": 0, + "content": "[39] Adam Kendon. Gesture: Visible action as utterance. Cambridge University Press, 2004. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.7, + 0.47, + 0.769 + ], + "angle": 0, + "content": "[40] Gwantae Kim, Seonghyeok Noh, Insung Ham, and Hanseok Ko. MPE4G: Multimodal pretrained encoder for co-speech gesture generation. In International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 1-5, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.772, + 0.47, + 0.813 + ], + "angle": 0, + "content": "[41] Jihoon Kim, Taehyun Byun, Seungyoung Shin, Jung-dam Won, and Sungjoon Choi. Conditional motion in-between. Pattern Recognition, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.816, + 0.47, + 0.843 + ], + "angle": 0, + "content": "[42] Chris L Kleinke. Gaze and eye contact: A research review. Psychological Bulletin, 100(1):78-100, 1986. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.846, + 0.47, + 0.901 + ], + "angle": 0, + "content": "[43] Zhifeng Kong, Wei Ping, Jiaji Huang, Kexin Zhao, and Bryan Catanzaro. DiffWave: A versatile diffusion model for audio synthesis. In International Conference on Learning Representations (ICLR), 2021. 3" + }, + { + "type": "list", + "bbox": [ + 0.087, + 0.093, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.093, + 0.892, + 0.176 + ], + "angle": 0, + "content": "[44] Stefan Kopp, Brigitte Krenn, Stacy Marsella, Andrew N. Marshall, Catherine Pelachaud, Hannes Pirker, Kristinn R. Thorisson, and Hannes Vilhjalmsson. Towards a common framework for multimodal generation: The behavior markup language. In Intelligent Virtual Agents (IVA), pages 205-217. Springer, 2006. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.178, + 0.892, + 0.233 + ], + "angle": 0, + "content": "[45] Quoc Anh Le, Souheil Hanoune, and Catherine Pelachaud. Design and implementation of an expressive gesture model for a humanoid robot. In IEEE-RAS International Conference on Humanoid Robots, pages 134-140, 2011. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.235, + 0.892, + 0.317 + ], + "angle": 0, + "content": "[46] Gilwoo Lee, Zhiwei Deng, Shugao Ma, Takaaki Shiratori, Siddhartha S. Srinivasa, and Yaser Sheikh. Talking With Hands 16.2M: A large-scale dataset of synchronized body-finger motion and audio for conversational motion analysis and synthesis. In International Conference on Computer Vision (ICCV), pages 763-772, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.319, + 0.892, + 0.36 + ], + "angle": 0, + "content": "[47] Robert W Levenson. Blood, sweat, and fears: The autonomic architecture of emotion. Annals of the New York Academy of Sciences, 1000:348-366, 2003. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.362, + 0.892, + 0.431 + ], + "angle": 0, + "content": "[48] Jing Li, Di Kang, Wenjie Pei, Xuefei Zhe, Ying Zhang, Zhenyu He, and Linchao Bao. Audio2gestures: Generating diverse gestures from speech audio with conditional variational autoencoders. In Computer Vision and Pattern Recognition (CVPR), pages 11293-11302, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.433, + 0.892, + 0.488 + ], + "angle": 0, + "content": "[49] Peizhuo Li, Kfir Aberman, Zihan Zhang, Rana Hanocka, and Olga Sorkine-Hornung. GANimator: Neural motion synthesis from a single sequence. Transactions on Graphics (TOG), 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.49, + 0.892, + 0.545 + ], + "angle": 0, + "content": "[50] Ruilong Li, Shan Yang, D. A. Ross, and Angjoo Kanazawa. AI choreographer: Music conditioned 3D dance generation with AIST++. In International Conference on Computer Vision (ICCV), 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.547, + 0.892, + 0.603 + ], + "angle": 0, + "content": "[51] Ruilong Li, Shan Yang, David A. Ross, and Angjoo Kanazawa. AI choreographer: Music conditioned 3d dance generation with AIST++. In International Conference on Computer Vision (ICCV), pages 13381-13392, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.604, + 0.892, + 0.659 + ], + "angle": 0, + "content": "[52] William Li and Philippe Pasquier. Automatic affect classification of human motion capture sequences in the valence-arousal model. In International Symposium on Movement and Computing, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.661, + 0.892, + 0.716 + ], + "angle": 0, + "content": "[53] Yuanzhi Liang, Qianyu Feng, Linchao Zhu, Li Hu, Pan Pan, and Yi Yang. Seeg: Semantic energized co-speech gesture generation. In Computer Vision and Pattern Recognition (CVPR), 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.718, + 0.892, + 0.787 + ], + "angle": 0, + "content": "[54] Haiyang Liu, Naoya Iwamoto, Zihao Zhu, Zhengqing Li, You Zhou, Elif Bozkurt, and Bo Zheng. DisCo: Disentangled implicit content and rhythm learning for diverse co-speech gestures synthesis. In ACM International Conference on Multimedia (MM), pages 3764-3773, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.789, + 0.892, + 0.871 + ], + "angle": 0, + "content": "[55] Haiyang Liu, Zihao Zhu, Naoya Iwamoto, Yichen Peng, Zhengqing Li, You Zhou, Elif Bozkurt, and Bo Zheng. BEAT: A large-scale semantic and emotional multi-modal dataset for conversational gestures synthesis. European Conference on Computer Vision (ECCV), 2022. 2, 3, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.873, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[56] Haiyang Liu, Zihao Zhu, Giorgio Becherini, Yichen Peng, Mingyang Su, You Zhou, Xuefei Zhe, Naoya Iwamoto, Bo" + }, + { + "type": "list", + "bbox": [ + 0.51, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "1951" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.118, + 0.092, + 0.47, + 0.148 + ], + "angle": 0, + "content": "Zheng, and Michael J. Black. EMAGE: Towards unified holistic co-speech gesture generation via expressive masked audio gesture modeling. In Computer Vision and Pattern Recognition (CVPR), 2024. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.151, + 0.472, + 0.219 + ], + "angle": 0, + "content": "[57] Yuejiang Liu, Riccardo Cadei, Jonas Schweizer, Sherwin Bahmani, and Alexandre Alahi. Towards robust and adaptive motion forecasting: A causal representation perspective. In Computer Vision and Pattern Recognition (CVPR), 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.222, + 0.47, + 0.276 + ], + "angle": 0, + "content": "[58] Matthew M. Loper, Naureen Mahmood, and Michael J. Black. MoSh: Motion and shape capture from sparse markers. Transactions on Graphics (TOG), 33(6):220:1-220:13, 2014. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.28, + 0.47, + 0.322 + ], + "angle": 0, + "content": "[59] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In International Conference on Learning Representations (ICLR), 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.324, + 0.47, + 0.38 + ], + "angle": 0, + "content": "[60] Shuhong Lu, Youngwoo Yoon, and Andrew W. Feng. Cospeech gesture synthesis using discrete gesture token learning. International Conference on Intelligent Robots and Systems (IROS), pages 9808-9815, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.382, + 0.47, + 0.451 + ], + "angle": 0, + "content": "[61] Birgit Lugarin, Catherine Pelachaud, and David Traum, editors. The Handbook on Socially Interactive Agents: 20 Years of Research on Embodied Conversational Agents, Intelligent Virtual Agents, and Social Robotics Volume 1: Methods, Behavior, Cognition. 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.453, + 0.47, + 0.509 + ], + "angle": 0, + "content": "[62] Naureen Mahmood, Nima Ghorbani, Nikolaus F. Troje, Gerard Pons-Moll, and Michael J. Black. Amass: Archive of motion capture as surface shapes. In International Conference on Computer Vision (ICCV), 2019. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.511, + 0.47, + 0.58 + ], + "angle": 0, + "content": "[63] Stacy Marsella, Yuyu Xu, Margaux Lhommet, Andrew Feng, Stefan Scherer, and Ari Shapiro. Virtual character performance from speech. In SIGGRAPH/Eurographics Symposium on Computer Animation (SCA), page 25-35, 2013. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.583, + 0.47, + 0.638 + ], + "angle": 0, + "content": "[64] Julieta Martinez, Michael J. Black, and Javier Romero. On human motion prediction using recurrent neural networks. In Computer Vision and Pattern Recognition (CVPR), 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.641, + 0.47, + 0.683 + ], + "angle": 0, + "content": "[65] David Mcneill. Hand and mind: What gestures reveal about thought. Bibliovault OAI Repository, the University of Chicago Press, 27, 1994. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.686, + 0.47, + 0.713 + ], + "angle": 0, + "content": "[66] Albert Mehrabian. Nonverbal communication. Aldine-Atherton Chicago, 1972. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.715, + 0.47, + 0.785 + ], + "angle": 0, + "content": "[67] Shivam Mehta, Siyang Wang, Simon Alexanderson, Jonas Beskow, Éva Székely, and Gustav Eje Henter. Diff-TTSG: Denoising probabilistic integrated speech and gesture synthesis. In Proc. ISCA Speech Synthesis Workshop (SSW), pages 150–156, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.787, + 0.47, + 0.843 + ], + "angle": 0, + "content": "[68] Davide Moltisanti, Jinyi Wu, Bo Dai, and Chen Change Loy. BRACE: the breakdancing competition dataset for dance motion synthesis. In European Conference on Computer Vision (ECCV), pages 329-344, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.845, + 0.47, + 0.9 + ], + "angle": 0, + "content": "[69] Alexander Quinn Nichol and Prafulla Dhariwal. Improved denoising diffusion probabilistic models. In International Conference on Machine Learning (ICML), pages 8162-8171, 2021. 4" + }, + { + "type": "list", + "bbox": [ + 0.088, + 0.092, + 0.472, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.092, + 0.892, + 0.134 + ], + "angle": 0, + "content": "[70] Dirk Ormoneit, Michael J. Black, T. Hastie, and H. Kjellström. Representing cyclic human motion using functional analysis. Image and Vision Computing (IVC), 2005. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.136, + 0.892, + 0.203 + ], + "angle": 0, + "content": "[71] Kunkun Pang, Dafei Qin, Yingruo Fan, Julian Habekost, Takaaki Shiratori, Junichi Yamagishi, and Taku Komura. BodyFormer: Semantics-guided 3d body gesture synthesis with transformer. Transactions on Graphics (TOG), 42(4), 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.205, + 0.892, + 0.272 + ], + "angle": 0, + "content": "[72] Daniel S. Park, William Chan, Yu Zhang, Chung-Cheng Chiu, Barret Zoph, Ekin D. Cubuk, and Quoc V. Le. SpecAugment: A simple data augmentation method for automatic speech recognition. In Interspeech 2019. ISCA, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.274, + 0.892, + 0.343 + ], + "angle": 0, + "content": "[73] Georgios Pavlakos, Vasileios Choutas, Nima Ghorbani, Timo Bolkart, Ahmed A. A. Osman, Dimitrios Tzionas, and Michael J. Black. Expressive body capture: 3D hands, face, and body from a single image. In Computer Vision and Pattern Recognition (CVPR), pages 10975-10985, 2019. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.344, + 0.892, + 0.413 + ], + "angle": 0, + "content": "[74] Ziqiao Peng, Haoyu Wu, Zhenbo Song, Hao Xu, Xiangyu Zhu, Jun He, Hongyan Liu, and Zhaoxin Fan. EmoTalk: Speech-driven emotional disentanglement for 3d face animation. In International Conference on Computer Vision (ICCV), pages 20687-20697, 2023. 2, 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.414, + 0.892, + 0.468 + ], + "angle": 0, + "content": "[75] Mathis Petrovich, Michael J. Black, and Gül Varol. Action-conditioned 3D human motion synthesis with transformer VAE. In International Conference on Computer Vision (ICCV), 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.469, + 0.892, + 0.524 + ], + "angle": 0, + "content": "[76] Mathis Petrovich, Michael J. Black, and Gül Varol. Action-conditioned 3D human motion synthesis with transformer VAE. In International Conference on Computer Vision (ICCV), 2021. 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.525, + 0.892, + 0.579 + ], + "angle": 0, + "content": "[77] Mathis Petrovich, Michael J. Black, and Gül Varol. TEMOS: Generating diverse human motions from textual descriptions. In European Conference on Computer Vision (ECCV), 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.581, + 0.892, + 0.622 + ], + "angle": 0, + "content": "[78] Paolo Petta, Catherine Pelachaud, and Roddy Cowie. Emotion-Oriented Systems: The Humaine Handbook. Springer Publishing Company, Incorporated, 2011. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.623, + 0.892, + 0.648 + ], + "angle": 0, + "content": "[79] Rosalind W. Picard. Affective Computing. MIT Press, 1997. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.65, + 0.892, + 0.693 + ], + "angle": 0, + "content": "[80] I. Poggi, C. Pelachaud, F. de Rosis, V. Carofiglio, and B. De Carolis. Greta. A Believable Embodied Conversational Agent, pages 3-25. Springer, 2005. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.693, + 0.892, + 0.734 + ], + "angle": 0, + "content": "[81] Xingqun Qi, Chen Liu, Lincheng Li, Jie Hou, Haoran Xin, and Xin Yu. EmotionGesture: Audio-driven diverse emotional co-speech 3d gesture generation, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.735, + 0.892, + 0.802 + ], + "angle": 0, + "content": "[82] Xingqun Qi, Chen Liu, Muyi Sun, Lincheng Li, Changjie Fan, and Xin Yu. Diverse 3D hand gesture prediction from body dynamics by bilateral hand disentanglement. In Computer Vision and Pattern Recognition (CVPR), pages 4616-4626, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.804, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[83] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning (ICML), pages 8748-8763. PMLR, 2021. 3" + }, + { + "type": "list", + "bbox": [ + 0.51, + 0.092, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1952" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.092, + 0.47, + 0.161 + ], + "angle": 0, + "content": "[84] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Computer Vision and Pattern Recognition (CVPR), pages 10674-10685, 2022. 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.163, + 0.472, + 0.219 + ], + "angle": 0, + "content": "[85] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-Net: Convolutional networks for biomedical image segmentation. In Medical Image Computing and Computer-Assisted Intervention - MICCAI, pages 234-241, 2015. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.221, + 0.471, + 0.262 + ], + "angle": 0, + "content": "[86] Tim Salzmann, Marco Pavone, and Markus Ryll. Motron: Multimodal Probabilistic Human Motion Forecasting. In Computer Vision and Pattern Recognition (CVPR), 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.264, + 0.471, + 0.305 + ], + "angle": 0, + "content": "[87] Klaus R Scherer. Vocal communication of emotion: A review of research paradigms. Speech Communication, 40 (1-2):227-256, 2003. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.307, + 0.471, + 0.347 + ], + "angle": 0, + "content": "[88] Maximilian Seitzer. pytorch-fid: FID Score for PyTorch. https://github.com/mseitzer/pytorch-fid, 2020. Version 0.3.0.6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.349, + 0.471, + 0.391 + ], + "angle": 0, + "content": "[89] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In International Conference on Learning Representations (ICLR), 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.393, + 0.471, + 0.461 + ], + "angle": 0, + "content": "[90] Kim Sung-Bin, Lee Hyun, Da Hye Hong, Suekyeong Nam, Janghoon Ju, and Tae-Hyun Oh. Laughtalk: Expressive 3d talking head generation with laughter. In Winter Conference on Applications of Computer Vision (WACV), pages 6404-6413, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.463, + 0.471, + 0.532 + ], + "angle": 0, + "content": "[91] Mingyang Sun†, Mengchen Zhao†, Yaqing Hou*, Minglei Li, Huang Xu, Songcen Xu, and Jianye Hao. Cospeech gesture synthesis by reinforcement learning with contrastive pre-trained rewards. In Computer Vision and Pattern Recognition (CVPR), 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.534, + 0.471, + 0.589 + ], + "angle": 0, + "content": "[92] Marcus Thiébaux, Stacy Marsella, Andrew N. Marshall, and Marcelo Kallmann. SmartBody: behavior realization for embodied conversational agents. In Adaptive Agents and Multi-Agent Systems (AAMAS), 2008. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.591, + 0.471, + 0.659 + ], + "angle": 0, + "content": "[93] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Hervé Jégou. Training data-efficient image transformers & distillation through attention. In International Conference on Machine Learning (ICML), pages 10347-10357, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.661, + 0.471, + 0.703 + ], + "angle": 0, + "content": "[94] Jonathan Tseng, Rodrigo Castellon, and Karen Liu. Edge: Editable dance generation from music. In Computer Vision and Pattern Recognition (CVPR), 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.705, + 0.471, + 0.759 + ], + "angle": 0, + "content": "[95] R. Williams. The Animator's Survival Kit: A Manual of Methods, Principles and Formulas for Classical, Computer Games, Stop Motion and Internet Animators. Farrar, Straus and Giroux, 2012. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.762, + 0.471, + 0.831 + ], + "angle": 0, + "content": "[96] Jingyu Wu, Shi Chen, Shuyu Gan, Weijun Li, Changyuan Yang, and Lingyun Sun. Cultural self-adaptive multimodal gesture generation based on multiple culture gesture dataset. In ACM International Conference on Multimedia (MM), page 3538-3549, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.832, + 0.471, + 0.901 + ], + "angle": 0, + "content": "[97] Sicheng Yang, Zhiyong Wu, Minglei Li, Zhensong Zhang, Lei Hao, Weihong Bao, Ming Cheng, and Long Xiao. Diffusestylegusture: Stylized audio-driven co-speech gesture generation with diffusion models. In International Joint Conference on Artificial Intelligence (IJCAI), 2023. 3, 6" + }, + { + "type": "list", + "bbox": [ + 0.087, + 0.092, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.092, + 0.892, + 0.175 + ], + "angle": 0, + "content": "[98] Sicheng Yang, Zhiyong Wu, Minglei Li, Zhensong Zhang, Lei Hao, Weihong Bao, and Haolin Zhuang. Qpgesture: Quantization-based and phase-guided motion matching for natural speech-driven gesture generation. In Computer Vision and Pattern Recognition (CVPR), pages 2321-2330, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.178, + 0.892, + 0.245 + ], + "angle": 0, + "content": "[99] Payam Jome Yazdian, Mo Chen, and Angelica Lim. Gesture2vec: Clustering gestures using representation learning methods for co-speech gesture generation. In International Conference on Intelligent Robots and Systems (IROS), pages 3100-3107, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.248, + 0.892, + 0.275 + ], + "angle": 0, + "content": "[100] Hongwei Yi. Show github. https://github.com/yhw-yhw/SHOW, 2023.6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.277, + 0.892, + 0.344 + ], + "angle": 0, + "content": "[101] Hongwei Yi, Hualin Liang, Yifei Liu, Qiong Cao, Yandong Wen, Timo Bolkart, Dacheng Tao, and Michael J Black. Generating holistic 3d human motion from speech. In Computer Vision and Pattern Recognition (CVPR), pages 469-480, 2023. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.347, + 0.892, + 0.401 + ], + "angle": 0, + "content": "[102] Li-Ping Yin, Yijun Wang, Tianyu He, Jinming Liu, Wei Zhao, Bohan Li, Xin Jin, and Jianxin Lin. Emog: Synthesizing emotive co-speech 3d gesture with diffusion model. ArXiv, abs/2306.11496, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.404, + 0.892, + 0.471 + ], + "angle": 0, + "content": "[103] Youngwoo Yoon, Woo-Ri Ko, Minsu Jang, Jaeyeon Lee, Jaehong Kim, and Geehyuk Lee. Robots learn social skills: End-to-end learning of co-speech gesture generation for humanoid robots. In International Conference on Robotics and Automation (ICRA), 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.474, + 0.892, + 0.543 + ], + "angle": 0, + "content": "[104] Youngwoo Yoon, Bok Cha, Joo-Haeng Lee, Minsu Jang, Jaeyeon Lee, Jaehong Kim, and Geehyuk Lee. Speech gesture generation from the trimodal context of text, audio, and speaker identity. Transactions on Graphics (TOG), 39(6), 2020. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.545, + 0.892, + 0.613 + ], + "angle": 0, + "content": "[105] Youngwoo Yoon, Pieter Wolfert, Taras Kucherenko, Carla Viegas, Teodor Nikolov, Mihail Tsakov, and Gustav Eje Henter. The genea challenge 2022: A large evaluation of data-driven co-speech gesture generation. In International Conference on Multimodal Interaction, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.615, + 0.892, + 0.656 + ], + "angle": 0, + "content": "[106] Ye Yuan and Kris M. Kitani. DLow: Diversifying latent flows for diverse human motion prediction. In European Conference on Computer Vision (ECCV), 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.657, + 0.892, + 0.713 + ], + "angle": 0, + "content": "[107] Yicheng Zhong, Huawei Wei, Peiji Yang, and Zhisheng Wang. Expclip: Bridging text and facial expressions via semantic alignment. In AAAI Conference on Artificial Intelligence, pages 7614-7622, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.715, + 0.892, + 0.769 + ], + "angle": 0, + "content": "[108] Yi Zhou, Connelly Barnes, Jingwan Lu, Jimei Yang, and Hao Li. On the continuity of rotation representations in neural networks. In Computer Vision and Pattern Recognition (CVPR), pages 5745-5753, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.771, + 0.892, + 0.825 + ], + "angle": 0, + "content": "[109] Yi Zhou, Jingwan Lu, Connelly Barnes, Jimei Yang, Sitao Xiang, and Hao Li. Generative tinkering: Long-term inbetweening of 3d human motions. arXiv:2005.08891, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.828, + 0.892, + 0.883 + ], + "angle": 0, + "content": "[110] Lingting Zhu, Xian Liu, Xuanyu Liu, Rui Qian, Ziwei Liu, and Lequan Yu. Taming diffusion models for audio-driven co-speech gesture generation. In Computer Vision and Pattern Recognition (CVPR), pages 10544–10553, 2023. 2, 8" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.092, + 0.892, + 0.883 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1953" + } + ] +] \ No newline at end of file diff --git a/2024/Emotional Speech-driven 3D Body Animation via Disentangled Latent Diffusion/8fe0dd5b-5118-4e0a-ac09-f70da9816cdc_origin.pdf b/2024/Emotional Speech-driven 3D Body Animation via Disentangled Latent Diffusion/8fe0dd5b-5118-4e0a-ac09-f70da9816cdc_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..1ed581210184567165bba7c4ac0ce03ce86d35c9 --- /dev/null +++ b/2024/Emotional Speech-driven 3D Body Animation via Disentangled Latent Diffusion/8fe0dd5b-5118-4e0a-ac09-f70da9816cdc_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d08f5c6034b5fdc3e2adf0baa80ca4a413b65c51e41509624702ea4af95e6c29 +size 693512 diff --git a/2024/Emotional Speech-driven 3D Body Animation via Disentangled Latent Diffusion/full.md b/2024/Emotional Speech-driven 3D Body Animation via Disentangled Latent Diffusion/full.md new file mode 100644 index 0000000000000000000000000000000000000000..ff698446b7ecf60383298ea18c7f62905cdd0b98 --- /dev/null +++ b/2024/Emotional Speech-driven 3D Body Animation via Disentangled Latent Diffusion/full.md @@ -0,0 +1,354 @@ +# Emotional Speech-driven 3D Body Animation via Disentangled Latent Diffusion + +Kiran Chhatre + +Radek Danček2 + +Nikos Athanasiou2 + +Giorgio Becherini² + +Christopher Peters1 + +Michael J. Black + +Timo Bolkart $^{2*}$ + +$^{1}$ KTH Royal Institute of Technology, Sweden $^{2}$ Max Planck Institute for Intelligent Systems, Germany + +# Abstract + +Existing methods for synthesizing 3D human gestures from speech have shown promising results, but they do not explicitly model the impact of emotions on the generated gestures. Instead, these methods directly output animations from speech without control over the expressed emotion. To address this limitation, we present AMUSE, an emotional speech-driven body animation model based on latent diffusion. Our observation is that content (i.e., gestures related to speech rhythm and word utterances), emotion, and personal style are separable. To account for this, AMUSE maps the driving audio to three disentangled latent vectors: one for content, one for emotion, and one for personal style. A latent diffusion model, trained to generate gesture motion sequences, is then conditioned on these latent vectors. Once trained, AMUSE synthesizes 3D human gestures directly from speech with control over the expressed emotions and style by combining the content from the driving speech with the emotion and style of another speech sequence. Randomly sampling the noise of the diffusion model further generates variations of the gesture with the same emotional expressivity. Qualitative, quantitative, and perceptual evaluations demonstrate that AMUSE outputs realistic gesture sequences. Compared to the state of the art, the generated gestures are better synchronized with the speech content, and better represent the emotion expressed by the input speech. Our code is available at amuse.is.tue.mpg.de. + +# 1. Introduction + +Animating 3D bodies from speech has a wide range of applications, such as telepresence in AR/VR, avatar animation in games and movies, and to embody interactive digital assistants. While methods for speech-driven 3D body animation have recently shown great progress [5, 7, 31, 56, 101], existing methods do not adequately address one crucial factor: the impact of emotion from the driving speech signal on the generated gestures. Emotions and their expressions play + +![](images/4414e5a2abd995e7a359daea2dc2751ecdd1f332f098d2c2ca660461c2179747.jpg) +... choose a major that is easy to find a good job in the future... + +![](images/44e4e54140317ad646e9f8780ec5ed0aebb5535002a5208ee58ef257be33ee5d.jpg) + +![](images/c127b0d6bd325b8ed5eee9e914bcb897ac31ee4e9c89592aeb40f1f0d40ef7c2.jpg) +Figure 1. Goal. AMUSE generates realistic emotional 3D body gestures directly from a speech sequence (top). It provides user control over the generated emotion by combining the driving speech sequence with a different emotional audio (bottom). + +a fundamental role in human communication [29, 35, 65] and have become an important consideration when designing computer systems that interact with humans in a natural manner [78, 79]. They are of central concern when synthesizing human animations for a wide variety of application contexts, such as Socially Interactive Agents [61]. Because of this, speech-driven animation systems must not only align movement with the rhythm of the speech, but should also be capable of generating gestures that are perceived as expressing the suitable emotion. + +Many factors contribute to the perception of emotion and personal idiosyncrasies, such as facial expressions [19], gaze and eye contact [42], physiological responses [47], tone of voice [87], body language [66], and gestures [39]. When it comes to 3D animation, the most relevant factors + +are facial expressions, gestures, and body language [95]. While emotional speech-driven animation methods have recently been proposed for 3D faces [18, 74, 90, 107], animating emotional bodies from speech remains under-explored. + +Generating gestures solely from speech with emotional control is a difficult task. First, the mapping from audio to body motion is a non-deterministic many-to-many mapping, which is difficult to model. Gestures across subjects can vary when uttering the same sentence, and a single individual's motions can change significantly across repetitions. Second, factoring out the impact of emotional state on the body motion from other, unknown factors, is difficult. This requires disentangling the effects of three different factors on the generated motion, namely content-based (i.e., gestures related to speech rhythm and word utterances), emotion-based, and those based on personal style. AMUSE addresses this by separating a speech sequence into content, emotion, and style latent vectors, which are then used to condition a latent diffusion model. Specifically, AMUSE consists of three main components: (1) an audio autoencoder trained to produce disentangled vectors of content, emotion, and style, (2) a 3D body motion prior in the form of a temporal variational autoencoder (VAE) to generate smooth and realistic gestures, and (3) a latent diffusion model, which generates 3D body motion given the input content, emotion, and style latent vectors. + +Training such a model requires a speech-to-3D body dataset of sufficient scale, which is rich and diverse in speakers and emotions. BEAT [55] is a good candidate because it provides a large set of 3D gestures associated with single-person monologues. Unfortunately, the bodies are represented as skeletons, and it lacks face mocap markers and FLAME expressions. Instead, to produce realistic body animations, we require articulated 3D body surfaces. To overcome this, we convert BEAT sequences to SMPL-X [73] format using MoSh++ [62] and use the SMPL-X parameters for training. See [56] for comparison. + +Our contributions are: (1) We present a framework to synthesize emotional 3D body gestures directly from speech. (2) We factor an input audio into disentangled content, emotion and style vectors, which enables us to separately control emotion in generated gestures. (3) We adapt temporal latent diffusion for multiple target conditions. + +# 2. Related Work + +# 2.1. 3D Conditional Human Motion Generation + +Early works focus mostly on predicting [10, 16, 33, 41, 57, 64, 70, 86, 106, 109] or generating human motion [30, 49], but do not consider multi-modal control. Recently, conditional motion generation through other modalities, such as text [2, 8, 9, 17, 22, 28, 77], music [50, 68, 94], speech [32], or action labels [27, 75], has gained more attention. Be + +low, we focus on speech-driven motion generation methods, since they are the most relevant to our work. + +# 2.2. Gesture Generation from Speech + +Rule-based gesture synthesis. Embodied conversational agents (ECA) are designed to interact and communicate with humans. Using the Behavior Markup Language (BML) [44] one can build rule-based systems for humanoids based on predefined behaviors [80]. This is used for completion of a storytelling task in an expressive manner [45]. The BEAT rule-based toolkit [14] enables adding non-verbal behavior on top of a pre-animated figure. Thiebaux et al. [92] develop an ECA by using procedural animation techniques and keyframe interpolation. Marsella et al. [63] design a generalized rule-based agent to generate expressions, eye gaze, and gestures from speech. Each of these approaches are based on non-trainable, rule-based techniques that may require substantial manual modelling effort to adapt to new tasks. + +Data-driven gesture synthesis. More recently, data-driven methods have superseded rule-based systems. Yoon et al. [104] use a fusion of text, audio and upper body gestures to learn an upper body gesture avatar, but can only control the style of individual speakers by sampling from their latent space. SpeechGestureMatching [32] generates 3D facial meshes and 3D keypoints of the body and hands from speech, but the outputs are separated and the method does not provide control over the generations. QPGesture [98] uses phase to better align the generated 3D skeleton-based gesturing avatars with the audio input. Ginosar et al. [23] and Diverse-3D-Hand-Gesture-Prediction [82] generate hand and arm motions only. Audio2Gestures [48] encode motion and audio to a low-dimensional latent space and generate gestures. SEEG [53] aims to generate gestures that align well with the semantics of the speech. DiffTTSG [67] regresses speech and gestures at the same time, joining the two modalities in a single system. DiffGAN [3] retargets gestures across speakers in a low-resource setting. The GENEA challenge [105] tackles gesticulation from speech alone using the Talking-with-Hands dataset [46]. Gesture2Vec [99] uses a machine translation model to translate text into gesture chunks and output full sequences using such quantized representations. TalkSHOW [101] uses a VQ-VAE to generate 3D human bodies gesturing with facial expressions from speech segments, but in an uncontrolled manner. Similarly, Co-speech gesture [60] uses an RQ-VAE to generate different gestures from speech. Alternative gesture generation from speech methods have been proposed such as reinforcement learning [91], self-supervised pretraining [40], and diffusion [67, 110]. BodyFormer [71] introduces a dataset of pseudo-groundtruth and a transformer-based method for generating gestures from speech. However, none of these methods provide explicit emotional con + +trol over the generated motion. + +For controllable generation, GestureDiffuCLIP [7] incorporates multiple conditions including CLIP [83] text features, video, or motion prompts via AdaIn [37] layers to generate gestures from speech, however, it does not allow explicit control over the emotion conveyed by the driving audio. ListenDenoiseAction [5] combines conformers and the DiffWave [43] architecture to generate gestures that can be controlled by a style vector, RhythmicGesticulator [6] disentangles the latent space into a vector related to the semantics of the gesture and one related to the subtle variations, while DisCo [54] models content and rhythm. StyleGestures [4] adapts MoGlow [34], demonstrating limited control over some motion attributes like the speed and expressiveness of gestures. DiffuseStyleGesture [97] uses diffusion to generate diverse gestures from speech. + +# 2.3. Emotion Control + +Emotion classification and control has been little studied in 3D human motion generation with only a few methods using skeletal motion in multi-class classification. Ghaleb et al. [20] employ a spatio-temporal graph convolution network to classify gestures into four classes: preparation, stroke, retraction, and neutral. Li et al. [52], on the other hand, use hidden Markov models for emotion classification of human movement mocap data. Karras et al. [38] learn face animations of a single actor, and test their method on different tasks by modifying the latent vectors. However, there is no disentanglement mechanism, and they do not model the synchronization of the emotion with the facial motions. Recently, EmoTalk [74], animates emotional 3D faces from speech input with control over the emotion intensity and EMOTE [18] disentangles emotion and speech to allow emotion editing at test time. However, models solely intended for facial tasks like lipSyncing and capturing expressions might not smoothly adapt to the complexity of whole-body movements and distinct articulation. Regarding emotion-conditioned motion generation, Aberman et al. [1] show style-transfer from video data to motion and provide some style-based control, but do not address speech-driven emotional gestures. Similarly, the ZeroEGGs [21] dataset contains some emotional gesture controls but also includes more generic styles of motion. The method requires the input of arbitrary frames of desired motion to encode a style, thereby relying on motions and speech as conditions during inference. Text-driven emotional gesticulation, as explored by Bhattacharya et al. [11, 12], emphasizes the generation of gestures based on textual cues, incorporating additional conditions such as speech, speaker ID, seed poses, as well as valence, arousal, and dominance triplets. However, these approaches do not provide the means to distill explicit emotion features, limiting free control over the generated gestures. Closer to + +our work, EMoG [102] incorporates emotion cues from the BEAT dataset [55] to generate improved gesture quality without explicit emotion control. EmotionGesture [81] uses a TED Emotion Dataset and BEAT to incorporate emotion features in gesture generation and generate emotional gestures. Although they can generate emotional gestures, their method is not end-to-end and has no explicit motion control. Specifically, it uses an emotion-conditioned VAE after training to acquire diverse emotion features that are used to generate gestures without guarantees and control over emotion types. Wu et al. [96] introduce the first multi-cultural gesture dataset containing 200 individuals of 10 different cultures. In contrast to prior work, we explicitly control the emotions conveyed by the generated gestures solely through emotional speech without relying on additional conditions. + +# 3. Method + +The AMUSE pipeline consists of two separately trained networks. The audio disentanglement module, which encodes input speech into latent vectors for content, emotion, and style is described in Sec. 3.2. The main architecture is described in Sec. 3.3. It consists of a 3D human motion prior coupled with a latent diffusion model. It takes random noise (or partially denoised latent vectors) on the input and outputs a human motion sequence. We introduce broader applications in gesture editing in Sec. 3.4. + +# 3.1. Preliminary: Expressive 3D Body Model + +SMPL-X [73] is a 3D model of the body surface. SMPL-X is defined as function $M(\boldsymbol{\beta}, \boldsymbol{\theta}, \boldsymbol{\psi})$ that produces a 3D body mesh. It is parameterized by identity shape $\boldsymbol{\beta} \in \mathbb{R}^{300}$ , pose $\boldsymbol{\theta} \in \mathbb{R}^{J \times 3}$ including finger articulation for rotations around $J$ joints, and facial expression $\boldsymbol{\psi} \in \mathbb{R}^{100}$ . We adopt the continuous 6D rotation representation for training following Zhou et al. [108], making $\boldsymbol{\theta} \in \mathbb{R}^{J \times 6}$ . Given pose parameters and any shape parameter, we can obtain body mesh vertices $V$ using the differentiable SMPL-X layer [73]. As the focus of our paper is on synthesizing body gestures and not locomotion, we disregard 8 joints that correspond those of the lower body joint poses, leaving $J = 47$ . Further, we omit the facial expression parameters, i.e., set $\boldsymbol{\psi} = \mathbf{0}$ . + +# 3.2. Speech Disentanglement Model + +Architecture. The goal of the this model is to factor an input speech into three disentangled latent representations, one for content (i.e., the words spoken), one for emotion, and one for personal style. To do so, we devise a specialized encoder-decoder architecture with three separate encoders, one for each latent space. We denote the encoders as: $E_{c}(a) = c$ , $E_{e}(a) = e$ , $E_{s}(a) = s$ , where $a$ is the input filterbank, $c$ , $e$ and $s$ denote the latent vectors for content, emotion and style and $E_{c}$ , $E_{e}$ and $E_{s}$ are their encoders. The architecture of the three encoders follows the + +![](images/f5a86adb69535c00ac5d1290edf434c200ec423010af30604e43ba061845a680.jpg) +Figure 2. Training. We train the motion prior $(\mathcal{P}_E, \mathcal{P}_D)$ and the latent denoiser $\Delta$ jointly, while keeping the audio encoding networks frozen. In the forward pass, we take an input audio $a^{1:T}$ and pose sequence $m^{1:T}$ . Firstly, we do a forward pass of $m^{1:T}$ through $\mathcal{P}_E$ and $\mathcal{P}_D$ and compute $\mathcal{L}_{rec}$ , $\mathcal{L}_{Vrec}$ , and $\mathcal{L}_{KL}$ . Then, we apply the diffusion process to a gradient-detached sg $[z_m]$ obtaining the noisy $z_m^{(D)}$ which is then denoised with $\Delta$ and $\mathcal{L}_{LD}$ is computed. Finally, we use $\Delta$ to fully denoise $z_n$ into gradient-detached sg $[z_{\tilde{m}}]$ , further decode $\tilde{m}^{1:T}$ using $\mathcal{P}_D$ , and compute $\mathcal{L}_{align}$ and $\mathcal{L}_{Valign}$ . + +design by Gong et al. [25, 26] (i.e., leveraging the DeiT visual transformer [93] adapted for processing filterbank images extracted from the input audio). The decoder takes the three latent vectors and produces a reconstructed filterbank. Formally $D(c, e, s) = \widehat{a}$ , where $\widehat{a}$ denotes the reconstructed filterbank. The decoder architecture consists of a fusion module and transformer-encoder layers. + +Training. The audio module is trained with a multiple loss terms that ensure that the three latent spaces are properly disentangled. In addition to the standard autoencoder reconstruction loss, we also employ three cross-reconstruction losses, in which we enforce the correct reconstruction of the audio signal where we modify one of the content, style or emotion latents. Additionally, we employ three loss terms on the latent vector predictions – namely emotion and style classification losses over $e$ and $s$ , and a content similarity loss between pairs of two content latent vectors extracted from audios that have the same spoken content. For a detailed description of the encoder-decoder architecture, a formal definition of the loss functions and a detailed description of the training process please refer to the Sup. Mat. + +# 3.3. Gesture Generation Model + +Motion prior. Similar to [15, 76], our motion prior network is a VAE transformer architecture with encoder $\mathcal{P}_E$ and decoder $\mathcal{P}_D$ . Specifically, both $\mathcal{P}_E$ and $\mathcal{P}_D$ follow a U-Net-like [85] structure with skip connections between transformer blocks (see Sup. Mat. for details). The positional embeddings are learnable and injected into each multi-head attention layer, following the design of Carion et al. [13]. Formally, the encoder takes a sequence of $T$ frames of the SMPL-X pose vectors $m^{1:T} \in \mathbb{R}^{6J \times T}$ and the first two tokens of its output, $\mu \in \mathbb{R}^{d_m}$ and $\Sigma \in \mathbb{R}^{d_m \times d_m}$ are used to extract the motion latent $z_m \in \mathbb{R}^{d_m}$ via the reparametrization trick. The decoder takes zero positional encodings as + +query input and the motion latent is fed as memory to every cross-attention transformer layer, producing the reconstructed motion $\hat{m}^{1:T}$ . + +Diffusion process. The forward diffusion process is similar to [36, 69]. We employ fixed variance and linearly scaled noise scheduler. We add noise to the motion latent $z_{m}$ for $D$ diffusion timesteps to obtain $z^{(D)}$ following: + +$$ +q \left(z _ {m} ^ {\left(t _ {d}\right)} \mid z _ {m} ^ {(0)}\right) = \mathcal {N} \left(z _ {m} ^ {\left(t _ {d}\right)}; \sqrt {\bar {\alpha} _ {t _ {d}}} z _ {m} ^ {(0)}, (1 - \bar {\alpha} _ {t _ {d}}) \mathbf {I}\right), +$$ + +with $\alpha_{t_d} = 1 - \beta_{t_d}$ , $\bar{\alpha}_{t_d} = \prod_{s=1}^{t_d} \alpha_s$ , and $\beta_{t_d}$ denotes diffusion process variance. + +Conditional denoising process. The denoising process consists of iteratively denoising a conditioned noisy motion latent vector to obtain the denoised motion latent $z_{\tilde{m}^{1:T}}$ . Our denoiser $\Delta$ is a latent variable model [84] and its architecture is similar to the U-Net-like structure of the motion prior encoder $\mathcal{P}_E$ . The input of the model is a concatenation of: $z_m^{(t_d)}$ , $\mathrm{SE}(t_d)$ , $c, e, s \in \mathbb{R}^{256}$ , where $\mathrm{SE}(t_d)$ is a sinusoidal positional encoding of diffusion timestep $t_d$ as defined in [36]. $\Delta$ iteratively denoises through each reversed diffusion step: + +$$ +z _ {m} ^ {(t _ {d} - 1)} = \Delta \left(\left[ z _ {m} ^ {(t _ {d})}, \operatorname {S E} \left(t _ {d}\right), c, e, s \right]\right). +$$ + +Training. We optimize the motion prior and the latent denoiser jointly to ensure audio-motion latent code alignment during conditional fusion in the denoising process using a 3-step forward pass through the gesture generation model. First, following standard VAE practice, we reconstruct $\hat{m}^{1:T}$ by the motion prior forward pass. As shown in Fig. 2, we then disable gradient calculation in $\mathcal{P}_E$ to infer the intermediate motion latent $\mathrm{sg}[z_m]$ , which serves as input to the denoiser. At this stage, we obtain the denoiser noise prediction, $\delta$ and use to compute the diffusion model gradients. Finally, in the third step we com + +pute $\tilde{m}^{1:T} = \mathcal{P}_D(\mathrm{sg}[z_{\tilde{m}}])$ , where $z_{\tilde{m}}$ is obtained by iteratively using the $\Delta$ to obtain a fully denoised latent from $z_n^{(t_D)} \sim \mathcal{N}(\mathbf{0}, \mathbf{I})$ . We indicate computations done without gradients with a stop-gradient operation $\mathrm{sg}[\cdot]$ . + +Losses. To train the motion prior, we include the standard VAE losses, namely the reconstruction loss on pose parameters $\mathcal{L}_{rec}$ and on vertex coordinates $\mathcal{L}_{Vrec}$ using the smooth L1 metric introduced in [24], which we denote as $L_1^s$ : + +$$ +\mathcal {L} _ {r e c} = L _ {1} ^ {s} (m ^ {1: T}, \hat {m} ^ {1: T}), \quad \mathcal {L} _ {V r e c} = L _ {1} ^ {s} (V ^ {1: T}, \hat {V} ^ {1: T}), +$$ + +where the root-centered vertices $V$ are obtained by feeding in pose parameters $m$ to a differentiable SMPL-X layer (without learnable parameters) and a mean shape $\beta = \vec{0}$ . The KL divergence loss of the motion prior is: + +$$ +\mathcal {L} _ {K L} = \frac {1}{2} \left[ \sum_ {i = 1} ^ {z} \left(\mu_ {i} ^ {2} + \sigma_ {i} ^ {2}\right) - \sum_ {i = 1} ^ {z} \left(l o g \left(\sigma_ {i} ^ {2}\right) + 1\right) \right]. +$$ + +To ensure the alignment of the diffusion-generated motions and the input audio, we apply the alignment reconstruction loss on the inferred motion pose parameters and the vertex coordinates: + +$$ +\mathcal {L} _ {a l i g n} = L _ {1} ^ {s} (m ^ {1: T}, \tilde {m} ^ {1: T}), \quad \mathcal {L} _ {V a l i g n} = L _ {1} ^ {s} (V ^ {1: T}, \tilde {V} ^ {1: T}). +$$ + +Finally, we utilize the objective similar to [15, 36, 84] to supervise the denoiser: + +$$ +\mathcal {L} _ {L D} = \left\| \delta^ {(t _ {d})} - \Delta (z _ {m} ^ {(t _ {d})}, \mathrm {S E} (t _ {d}), c, e, s) \right\| _ {2} ^ {2}, +$$ + +where $\delta^{(t_d)}$ is the noise vector sampled from $\mathcal{N}(\mathbf{0},\mathbf{I})$ in the corresponding diffusion step $t_d$ . The combined gesture model loss is: + +$$ +\mathcal {L} _ {g e s} = \mathcal {L} _ {r e c} + \mathcal {L} _ {V r e c} + \mathcal {L} _ {K L} + \mathcal {L} _ {a l i g n} + \mathcal {L} _ {V a l i g n} + \mathcal {L} _ {L D} +$$ + +Inference. We employ DDIM [89] to infer high quality conditional motion samples with a small number of denoising timesteps. During inference we draw a sample vector from $\mathcal{N}(\mathbf{0},\mathbf{I})$ to iteratively denoise in reversed timesteps. The denoised sample is then passed through the decoder $\mathcal{P}_D(z_{\tilde{m}^{1:T}})$ to obtain motion $\tilde{m}^{1:T}$ . + +# 3.4. Gesture Editing + +Due to the disentangling of the inputs, AMUSE achieves semantic gesticulation control using two driving input audios. Specifically, given two input audio signals $a_1$ and $a_2$ , we extract their latent representations of content $c_1, c_2$ , emotion $e_1, e_2$ , and style $s_1, s_2$ . Then, we simply initialize the denoising procedure of $\Delta$ with the triplet $(c_1, e_2, s_1)$ , generating the gesture with the content and style of $a_1$ but the emotion of input audio $a_2$ . Similarly, instead of emotion we can also change the gesticulation style to that of the speaker of $a_2$ by initializing with $(c_1, e_1, s_2)$ . + +# 4. Implementation Details + +MoCap data preparation. The BEAT [55] mocap sequences, captured in a Vicon system at $120\mathrm{Hz}$ , are downsampled to $30\mathrm{Hz}$ and processed using $\mathrm{MoSh}++$ [58, 62] to obtain SMPL-X parameters. Given a sequence of 3D mocap marker positions, we jointly optimize SMPL-X shape and pose parameters, 3D body translation, and embedding of the mocap markers in the SMPL-X surface. Once processed, the sequences are then divided according to the emotion annotations in the BEAT dataset. We use sequences of English speaking subjects in monologue speaking style for training and evaluating AMUSE. For each sequence we draw $m^{1:L}$ at 30 FPS and concatenate with audio content $c$ , emotion $e$ , and style $s$ latent vectors. Then, we segment it to 10-sec windows $T$ , beginning from the timestamp 0 and discarding additional unaligned information at the end. This preprocessing choice allows us to train transformer networks without masking. We provide additional data processing information in the Sup. Mat. + +Audio preprocessing. We use audio sequences belonging to eight categorical emotion labels (neutral, happy, angry, sad, contempt, surprise, fear, and disgust). Each audio chunk of 10s is converted into a filter bank with 128 melfrequency bins with a 25ms Hamming frame window and 10ms frame shift. We mask each sample with a maximum length of 24 in the frequency domain and a maximum length of 96 in the time domain, employing Park et al. [72]. Following [25, 26], we standardize the filter bank and augment it via noise injection and circular shifting. Before feeding in our speech disentanglement model, each filter bank is split into a sequence of fixed 1209 patches of $16 \times 16$ each having 6 units overlap in frequency and time domain. + +Motion prior. The motion prior is a VAE encoder-decoder with 9 layers and 4 heads, following Chen et al. [15]. The encoder-decoder is a U-Net-like transformer with residual connections. Learnable positional embeddings are injected in each multi-head attention layer. We have a linear projection at the start and the end of our motion prior network. The KL divergence term is weighted with a factor of $1e - 4$ . Denoiser. The denoiser follows the same network architecture as our prior encoder. The hidden dimension of all transformer layers is 1024. We use 1000 diffusion steps $D$ during training and 50 during inference. Noise betas are in range $[0.00085, 0.012]$ . We jointly optimize the prior and denoiser networks for 5000 epochs with batch size of 64, learning rate 0.0001, and the AdamW optimizer [59]. + +# 5. Experiments + +Speech disentanglement model. We evaluate the performance of the speech disentanglement model quantitatively using classification accuracy and F1 scores on emotion and style. The accuracy is computed as average scores + +for all 8 emotion as well style categories that are part of the test dataset. The emotion and style accuracy is $91.53\%$ and $96.06\%$ , respectively. The emotion F1 score and style F1 scores are 0.914 and 0.960, respectively. See the Sup. Mat. for ablations and a detailed metric analysis. + +Gesture generation model. We evaluate the performance of our gesture generation model quantitatively, qualitatively, and perceptually against following methods: TalkSHOW [101] and the re-implementation of Habibie et al. [31] provided by the TalkSHOW authors in the official TalkSHOW release [100], DiffuseStyleGesture (DSG) [97], MoGlow [34], and CaMN [55]. Additionally, we adapt TalkSHOW to include categorical emotion labels as input along with the existing architecture that only allows one-hot encodings of personal style. We then retrain it on our training data. We refer to it as TalkSHOW-BEAT. There are some concurrent works [5, 7, 56], which introduce methods for gesture generation from speech, however, direct comparison is hindered by the unavailability of released code our task. Refer to the Sup. Mat. for the ablation experiments, the emotion and style editing experiments, and their quantitative evaluation. + +# 5.1. Quantitative Evaluation + +To quantitatively evaluate our method's gesture generations and edited gesture generations, we train a transformer-based encoder architecture (denoted as $M$ ) similar to Petrovich et al. [76] in an autoencoder setting, where we append a CLS token at the beginning of the motion sequence. $M$ is trained with a cross-entropy emotion classification objective applied to the output CLS token. We train $M$ on the BEAT training dataset and use its features to compute the following metrics: (1) Fréchet gesture distance (FGD): We + +
MethodSRGR↑BA↑FGD↓Div→GAa↑
GT—0.83—27.8364.04
Ours0.360.81388.6325.0646.76
Ours-EmoEditb—0.79792.5824.6834.18
TalkSHOW-BEAT0.310.64808.9924.1622.71
TalkSHOW [101]0.300.60762.1523.1929.41
DSG [97]0.230.40763.1019.7722.70
Habibie et al. [31]0.230.39809.1721.3416.67
MoGlow [34]0.210.351097.0319.5016.62
CaMN [55]0.210.391063.8718.9014.17
+ +a GA is average of all 8 emotions. +b GA for these are average accuracy for all generations with 7 edited audio sequences. + +Table 1. Gesture quantitative results. We compare our methods against several SOTA methods using metrics explained in Sec. 5.1. We observe that AMUSE outperforms in all scores compared to baseline methods. Additionally, AMUSE-EmoEdit outperforms in Beat Align, Diversity, and Gesture Emotion Accuracy scores compared to the baseline methods. + +follow [88, 103, 104] to compute the feature distance between generated and ground truth motion features. (2) Gesture diversity (Div): Similarly to Chen et al. [15], we compute variance across generated features. (3) Gesture emotion accuracy (GA): We report top-1 emotion classification accuracy predicted by a classifier trained on the motion $M$ -predicted latents. (4) Beat align (BA): We follow [51, 55], to evaluate the motion-speech correlation in terms of the similarity between the kinematic motion beats and speech audio beats. The kinematic motion beats are directly computed from the generated motion sequences. (5) Semantic-Relevant Gesture Recall (SRGR): We follow Liu et al. [55], to evaluate the semantic relevancy of gestures with GT motion. We use the ground truth semantic scores to compute this metric. The scores are obtained from the BEAT authors, representing a continuous score on a scale 0-1 per gesture style for 4 gesture semantic categories: beat, deictic, iconic, and metaphoric. While comparing with methods that output coarse skeletal data (DSG [97], MoGlow [34], and CaMN [55]), we convert the skeleton motion data into the SMPL-X axis angle representation. For details on the architectures and training of $M$ , and the losses, please refer to the Sup. Mat. + +We prepare the evaluation data by randomly selecting 72 unique motion sequences each of length 10s and comprising 8 emotions across test subjects and compute the aforementioned metrics. We use 9 sequences for each emotion per subject. The results are reported in Tab. 1. All best scores are highlighted in green and second best in blue. AMUSE outperforms the baseline methods in all given metrics. To validate the performance of gesture emotion editing, we also report the same metrics for the emotion editing task (Ours-EmoEdit). During inference, the input style and content latents are extracted from neutral-emotion audio, while the emotion latent comes from a different audio of different emotion. These emotional edits offer numerous possibilities, allowing for transitions from any to any emotion. Tab. 1 shows the average for editing from neutral to other emotions. Since we require the GT gesture semantics score to compute SRGR metric, it is not possible to compute the SRGR for the synthetic edited-emotion gestures as they are not part of the original BEAT dataset. Ours-EmoEdit outperforms the baseline methods in BA, Div, and GA metrics. This demonstrates the capability of our model to maintain highly discriminative cues when switching between different emotions. TalkSHOW-BEAT has the second best score for SRGR whereas TalkSHOW demonstrates second best FGD score. Although, our model and ours-EmoEdit show improvements over the baseline methods, GT motions have higher diversity, Beat alignment score, and are easier to classify than generations of AMUSE, highlighting the challenging nature of the problem. + +![](images/fc60d13af2bb6a9cafc31879ce92d3d0be33118b096d7346506181c227888a85.jpg) +Figure 3. Qualitative comparison across all emotions. We evaluate generation on different test audios. AMUSE exhibits well-synchronized beat gestures and consistently produces gestures that accurately convey the emotional content expressed in the input speech. + +![](images/10f8b2f70747ff519b934e9e660b3e1def7176179d1c1813988bf47314fe3b3b.jpg) +Figure 4. Qualitative comparison with baseline methods. The speech segment describes intense angry speech. + +# 5.2. Qualitative Evaluation + +Comparison with baseline methods. In Fig. 4, we demonstrate comparison with baseline methods that output a 3D body mesh: Habibie et al. [31], TalkSHOW [101], TalkSHOW-BEAT, and the BEAT ground truth (GT) [55]. We observe that AMUSE generates gestures that are semantically closer to the speech content and produces expressive emotional gestures closer to the perceived emotion. For example, the GT motion exhibits anger when saying "put them into detention". AMUSE demonstrates tense posture and aggressive movements comparable with the ground truth data and accurate synchronization with the spoken words. TalkSHOW [101] and Habibie et al. [31] exhibit limited movement and display inferior and static gestures on test audios as seen in the last two rows + +![](images/7359634e0462ace41c9f62be4167596525d8f96ebbde7384a8a2de69b67e5475.jpg) +Figure 5. Qualitative evaluation of diverse generations. Multiple generations overlayed. + +of Fig. 4. TalkSHOW-BEAT slightly outperforms other baseline methods by demonstrating enhanced synchronized gestures, but it still does not perform as well as AMUSE. + +Diverse emotional gestures. In Fig. 5, our probabilistic model can generate diverse gestures for same input audio. + +Emotional gesture generation. In Fig. 3 AMUSE demonstrates strong correlation with the spoken utterances as well as different emotions. We observe that our model is able to correlate semantic words to associated gestures. For example, gestures demonstrate forceful actions and tense stance with angry audio "normally get angry" whereas it generates lowered and calm hand positions for sad audio "I went to collect". Similarly, our generations show hands that are closer to body for fearful audio "in a dangerous situation" while widely open expressing astonishment for happy and surprised audio "people pass by" and "Wow! You are here". + +Emotion editing. We use two audio streams of a female subject for neutral and sad emotion. This experiment edits the subject's gesture style from moderately controlled hand movements to a sad style with lethargic posture conveying a sense of heaviness, as seen in Fig. 7 (top). + +Gesture style editing. We use audio streams of two male subjects for the happy (ID - 13) and angry (ID - 2) emotion. With the emotion, style and content latent fusion mechanism from two driving audio streams, AMUSE is able to adapt the male (ID - 13) subject's body gestures from being close to their body to more open with squared tightened shoulders, expressing a shift from happy to angry emotions of a different subject (ID - 2), as shown in Fig. 7 (bottom). Please refer to the supplemental video for qualitative results and comparisons to additional gesture genera + +![](images/dcd48010663c371765d28e509f35149e52387aa74e2c0ede6cd216ce8325d5ce.jpg) +Figure 6. The perceptual study results for gesture emotion preference (left) and synchronization with speech (right). The number of attentive participants that passed the catch trials is indicated on the right and the reported results only consider these participants. + +![](images/790684fdaf80e92c2b5d8b5dae1f520bf3f102717544d6328c5f80a8af4ec7ac.jpg) +Figure 7. Gesture editing. Top: We modify style from being neutral (left) to being sad (right) by combining the emotion latent from sad audio with the content latent from neutral audio. Bottom: We transform the style from Subject 13 being happy (left) to being angry (right) by merging the content latent from happy audio with the style and emotion latents from an angry audio of Subject 2. + +tion methods [4, 55, 110] trained on coarse skeletal data. + +# 5.3. Perceptual Study + +Design. Our perceptual study is designed as a side-by-side comparison of two gesture videos generated with the same audio as input but by two different methods (AMUSE and another model or GT). The participants are asked to rate their preference of the methods on a five-point Likert scale for "synchronization with speech" and "gesture emotion appropriateness" given the GT emotion label of the input audio. We recruit 25 participants per method-to-method comparison on Amazon Mechanical Turk. Each participant is shown 24 pairs of randomly selected test set animations, 3 per emotion (neutral, happy, angry, sad, disgust, fear, surprise, and contempt). To allow the participant to get used to the task, we discard the answers of the first three comparisons and repeat these at the end. We incorporate three catch trials and responses from participants that fail on more than one are filtered out, as shown in Fig. 6 (right). + +Results. The results of the study are shown in Fig. 6. AMUSE outperforms all competing methods by a considerable margin on both tasks, suggesting that AMUSE's generations are more appropriate for both the content of the input speech and its emotion compared to the baselines. However, it must be noted that there is still a significant gap between AMUSE and the GT. Please refer to the Sup. Mat. for details about the perceptual study. + +# 5.4. Discussion and Future Work + +Upper-body motion. We focus on the smooth coordination between the pelvis and upper body animation for side-by-side comparisons with other methods, as all other methods primarily focus on upper body movements. Future work should include lower-body motion and locomotion as these impact the perceived emotional state of a sequence. + +Semantics. While the generated gestures, synchronized with the driving speech sequence, do not account for semantics such as deictic and metaphoric gestures, incorporating the text/language modality could help further improve in this direction. + +Facial expressions. While emotional speech-driven face animation methods [18, 74] can be combined with bodies generated from AMUSE, jointly learning to generate emotional 3D bodies from speech is a topic that needs attention. End-to-end training. Joint audio-gesture training may enhance results but requires careful loss term balancing and increased GPU memory. Therefore, we opted for separate training. + +# 6. Conclusion + +We present AMUSE, a framework to generate emotional body gestures from speech. The emotions and personal styles of the synthesized gestures can be controlled, thanks to the disentanglement of content, emotion, and style directly from the speech. The latent diffusion-based framework can further generate variations of the same gesture with the same emotion. Our quantitative evaluations show that AMUSE achieves state of the art performance on a variety of metrics: diversity, gesture emotion classification accuracy, Frechet gesture distance, beat alignment score, and semantic relevant gesture recall. Finally, our perceptual study demonstrates that AMUSE generates motions that are better synchronized and better match the emotion expressed of the input speech than previous state of the art. + +Acknowledgments. We thank A Cseke, T McConnell and T Alexiadis for help with the perceptual study, B Pellkofer, J Piles-Contreras, and E Fritzler for IT support, and P Kulits, M Petrovich, S Kokane, and S Zojaji for proofreading and insightful discussions. This project has received funding from the European Union's Horizon 2020 research and innovation programme under the Marie Skłodowska-Curie grant agreement No. 860768. + +Disclosure. https://download.is.tue.mpg.de/amuse/Disclosure.txt + +# References + +[1] Kfir Aberman, Yijia Weng, Dani Lischinski, Daniel Cohen-Or, and Baoquan Chen. Unpaired motion style transfer from video to animation. Transactions on Graphics (TOG), 2020. 3 +[2] Chaitanya Ahuja and Louis-Philippe Morency. Language2Pose: Natural language grounded pose forecasting. International Conference on 3D Vision (3DV), 2019. 2 +[3] Chaitanya Ahuja, Dong Won Lee, and Louis-Philippe Morency. Low-resource adaptation for personalized cospeech gesture generation. In Computer Vision and Pattern Recognition (CVPR), pages 20566-20576, 2022. 2 +[4] Simon Alexanderson, Gustav Eje Henter, Taras Kucherenko, and Jonas Beskow. Style-controllable speech-driven gesture synthesis using normalising flows. Computer Graphics Forum (CGF), 39(2):487-496, 2020. 3, 8 +[5] Simon Alexanderson, Rajmund Nagy, Jonas Beskow, and Gustav Eje Henter. Listen, Denoise, Action! Audio-driven motion synthesis with diffusion models. Transactions on Graphics (TOG), 42(4):1-20, 2023. 1, 3, 6 +[6] Tenglong Ao, Qingzhe Gao, Yuke Lou, Baoquan Chen, and Libin Liu. Rhythmic gesticulator. Transactions on Graphics (TOG), 41(6):1-19, 2022. 3 +[7] Tenglong Ao, Zeyi Zhang, and Libin Liu. GestureDiffu-CLIP: Gesture diffusion model with CLIP latents. Transactions on Graphics (TOG), 2023. 1, 3, 6 +[8] Nikos Athanasiou, Mathis Petrovich, Michael J. Black, and Gül Varol. TEACH: Temporal Action Compositions for 3D Humans. In International Conference on 3D Vision (3DV), 2022. 2 +[9] Nikos Athanasiou, Mathis Petrovich, Michael J. Black, and Gül Varol. SINC: Spatial composition of 3D human motions for simultaneous action generation. International Conference on Computer Vision (ICCV), 2023. 2 +[10] Emad Barsoum, John Kender, and Zicheng Liu. HPGAN: probabilistic 3d human motion prediction via GAN. In Computer Vision and Pattern Recognition Workshops (CVPRW), 2017. 2 +[11] Uttaran Bhattacharya, Elizabeth Childs, Nicholas Rewkowski, and Dinesh Manocha. Speech2affectivegestures: Synthesizing co-speech gestures with generative adversarial affective expression learning. In ACM International Conference on Multimedia (MM), 2021. 3 +[12] Uttaran Bhattacharya, Nicholas Rewkowski, Abhishek Banerjee, Pooja Guhan, Aniket Bera, and Dinesh Manocha. Text2gestures: A transformer-based network for generating emotive body gestures for virtual agents. In IEEE Virtual Reality and 3D User Interfaces, VR, 2021. 3 +[13] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In European Conference on Computer Vision (ECCV), pages 213-229. Springer, 2020. 4 +[14] Justine Cassell, Hannes Hogni Vilhjalmsson, and Timo + +thy Bickmore. BEAT: The Behavior Expression Animation Toolkit, pages 163-185. Springer, 2004. 2 +[15] Xin Chen, Biao Jiang, Wen Liu, Zilong Huang, Bin Fu, Tao Chen, and Gang Yu. Executing your commands via motion diffusion in latent space. In Computer Vision and Pattern Recognition (CVPR), pages 18000-18010, 2023. 4, 5, 6 +[16] Enric Corona, Albert Pumarola, G. Alenyà, and F. Moreno-Noguer. Context-aware human motion prediction. In Computer Vision and Pattern Recognition (CVPR), pages 6990–6999, 2020. 2 +[17] Rishabh Dabral, Muhammad Hamza Mughal, Vladislav Golyanik, and Christian Theobalt. Mofusion: A framework for denoising-diffusion-based motion synthesis. In Computer Vision and Pattern Recognition (CVPR), 2023. 2 +[18] Radek Daneček, Kiran Chhatre, Shashank Tripathi, Yandong Wen, Michael Black, and Timo Bolkart. Emotional speech-driven animation with content-emotion disentangle-ment. In International Conference on Computer Graphics and Interactive Techniques in Asia (SIGGRAPH ASIA), 2023. 2, 3, 8 +[19] Paul Ekman. Facial expression and emotion. American Psychologist, 48(4):384-392, 1993. 1 +[20] Esam Ghaleb, Ilya Burenko, Marlou Rasenberg, Wim Pouw, Peter Uhrig, Judith Holler, Ivan Toni, Asli Özyurek, and Raquel Fernández. Co-speech gesture detection through multi-phase sequence labeling. In Winter Conference on Applications of Computer Vision (WACV), 2024. 3 +[21] Saeed Ghorbani, Ylva Ferstl, Daniel Holden, Nikolaus F. Troje, and Marc-Andre Carbonneau. Zeroeggs: Zero-shot example-based gesture generation from speech. Computer Graphics Forum (CGF), 42(1):206-216, 2023. 3 +[22] Anindita Ghosh, Noshaba Cheema, Cennet Oguz, Christian Theobalt, and Philipp Slusallek. Synthesis of compositional animations from textual descriptions. In International Conference on Computer Vision (ICCV), 2021. 2 +[23] Shiry Ginosar, Amir Bar, Gefen Kohavi, Caroline Chan, Andrew Owens, and Jitendra Malik. Learning individual styles of conversational gesture. In Computer Vision and Pattern Recognition (CVPR), pages 3497-3506, 2019. 2 +[24] Ross B. Girshick. Fast R-CNN. In International Conference on Computer Vision (ICCV), pages 1440-1448, 2015. 5 +[25] Yuan Gong, Yu-An Chung, and James R. Glass. AST: audio spectrogram transformer. In *Interspeech* 2021, pages 571-575, 2021. 4, 5 +[26] Yuan Gong, Cheng-I Lai, Yu-An Chung, and James R. Glass. SSAST: self-supervised audio spectrogram transformer. In AAAI Conference on Artificial Intelligence, pages 10699-10709, 2022. 4, 5 +[27] Chuan Guo, Xinxin Zuo, Sen Wang, Shihao Zou, Qingyao Sun, Annan Deng, Minglun Gong, and Li Cheng. Action2Motion: Conditioned generation of 3d human motions. In ACM International Conference on Multimedia (MM), 2020. 2 +[28] Chuan Guo, Shihao Zou, Xinxin Zuo, Sen Wang, Wei Ji, Xingyu Li, and Li Cheng. Generating diverse and natural 3D human motions from text. In Computer Vision and Pattern Recognition (CVPR), 2022. 2 + +[29] Nejla Gürefe. The role of gestures in mathematical discourse of hard-hearing students: Prism example. Acta Destructica Napocensia, 11:125-140, 2018. 1 +[30] I. Habibie, Daniel Holden, Jonathan Schwarz, J. Yearsley, and T. Komura. A recurrent variational autoencoder for human motion synthesis. In *British Machine Vision Conference (BMVC)*, 2017. 2 +[31] Ikhsanul Habibie, Weipeng Xu, Dushyant Mehta, Lingjie Liu, Hans-Peter Seidel, Gerard Pons-Moll, Mohamed Elgharib, and Christian Theobalt. Learning speech-driven 3d conversational gestures from video. In ACM International Conference on Intelligent Virtual Agents (IVA), 2021. 1, 6, 7 +[32] Ikhsanul Habibie, Mohamed A. Elgharib, Kripasindhu Sarkar, Ahsan Abdullah, Simbarashe Linval Nyatsanga, Michael Neff, and Christian Theobalt. A motion matching-based framework for controllable gesture synthesis from speech. International Conference on Computer Graphics and Interactive Techniques (SIGGRAPH), 2022. 2 +[33] Félix G. Harvey, Mike Yurick, Derek Nowrouzezahrai, and Christopher Joseph Pal. Robust motion in-between. Transactions on Graphics (TOG), 2020. 2 +[34] Gustav Eje Henter, Simon Alexanderson, and Jonas Beskow. MoGlow: Probabilistic and controllable motion synthesis using normalising flows. Transactions on Graphics (TOG), 39(4):236:1-236:14, 2020. 3, 6 +[35] Sandra Herbert. Gesture types for functions. Mathematics Education Research Group of Australasia, pages 322-329, 2012. 1 +[36] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In Conference on Neural Information Processing Systems (NeurIPS), 2020. 4, 5 +[37] Xun Huang and Serge J. Belongie. Arbitrary style transfer in real-time with adaptive instance normalization. In International Conference on Computer Vision (ICCV), pages 1510-1519, 2017. 3 +[38] Tero Karras, Timo Aila, Samuli Laine, Antti Herva, and Jaakko Lehtinen. Audio-driven facial animation by joint end-to-end learning of pose and emotion. Transactions on Graphics (TOG), 2017. 3 +[39] Adam Kendon. Gesture: Visible action as utterance. Cambridge University Press, 2004. 1 +[40] Gwantae Kim, Seonghyeok Noh, Insung Ham, and Hanseok Ko. MPE4G: Multimodal pretrained encoder for co-speech gesture generation. In International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 1-5, 2023. 2 +[41] Jihoon Kim, Taehyun Byun, Seungyoung Shin, Jung-dam Won, and Sungjoon Choi. Conditional motion in-between. Pattern Recognition, 2022. 2 +[42] Chris L Kleinke. Gaze and eye contact: A research review. Psychological Bulletin, 100(1):78-100, 1986. 1 +[43] Zhifeng Kong, Wei Ping, Jiaji Huang, Kexin Zhao, and Bryan Catanzaro. DiffWave: A versatile diffusion model for audio synthesis. In International Conference on Learning Representations (ICLR), 2021. 3 + +[44] Stefan Kopp, Brigitte Krenn, Stacy Marsella, Andrew N. Marshall, Catherine Pelachaud, Hannes Pirker, Kristinn R. Thorisson, and Hannes Vilhjalmsson. Towards a common framework for multimodal generation: The behavior markup language. In Intelligent Virtual Agents (IVA), pages 205-217. Springer, 2006. 2 +[45] Quoc Anh Le, Souheil Hanoune, and Catherine Pelachaud. Design and implementation of an expressive gesture model for a humanoid robot. In IEEE-RAS International Conference on Humanoid Robots, pages 134-140, 2011. 2 +[46] Gilwoo Lee, Zhiwei Deng, Shugao Ma, Takaaki Shiratori, Siddhartha S. Srinivasa, and Yaser Sheikh. Talking With Hands 16.2M: A large-scale dataset of synchronized body-finger motion and audio for conversational motion analysis and synthesis. In International Conference on Computer Vision (ICCV), pages 763-772, 2019. 2 +[47] Robert W Levenson. Blood, sweat, and fears: The autonomic architecture of emotion. Annals of the New York Academy of Sciences, 1000:348-366, 2003. 1 +[48] Jing Li, Di Kang, Wenjie Pei, Xuefei Zhe, Ying Zhang, Zhenyu He, and Linchao Bao. Audio2gestures: Generating diverse gestures from speech audio with conditional variational autoencoders. In Computer Vision and Pattern Recognition (CVPR), pages 11293-11302, 2021. 2 +[49] Peizhuo Li, Kfir Aberman, Zihan Zhang, Rana Hanocka, and Olga Sorkine-Hornung. GANimator: Neural motion synthesis from a single sequence. Transactions on Graphics (TOG), 2022. 2 +[50] Ruilong Li, Shan Yang, D. A. Ross, and Angjoo Kanazawa. AI choreographer: Music conditioned 3D dance generation with AIST++. In International Conference on Computer Vision (ICCV), 2021. 2 +[51] Ruilong Li, Shan Yang, David A. Ross, and Angjoo Kanazawa. AI choreographer: Music conditioned 3d dance generation with AIST++. In International Conference on Computer Vision (ICCV), pages 13381-13392, 2021. 6 +[52] William Li and Philippe Pasquier. Automatic affect classification of human motion capture sequences in the valence-arousal model. In International Symposium on Movement and Computing, 2016. 3 +[53] Yuanzhi Liang, Qianyu Feng, Linchao Zhu, Li Hu, Pan Pan, and Yi Yang. Seeg: Semantic energized co-speech gesture generation. In Computer Vision and Pattern Recognition (CVPR), 2022. 2 +[54] Haiyang Liu, Naoya Iwamoto, Zihao Zhu, Zhengqing Li, You Zhou, Elif Bozkurt, and Bo Zheng. DisCo: Disentangled implicit content and rhythm learning for diverse co-speech gestures synthesis. In ACM International Conference on Multimedia (MM), pages 3764-3773, 2022. 3 +[55] Haiyang Liu, Zihao Zhu, Naoya Iwamoto, Yichen Peng, Zhengqing Li, You Zhou, Elif Bozkurt, and Bo Zheng. BEAT: A large-scale semantic and emotional multi-modal dataset for conversational gestures synthesis. European Conference on Computer Vision (ECCV), 2022. 2, 3, 5, 6, 7, 8 +[56] Haiyang Liu, Zihao Zhu, Giorgio Becherini, Yichen Peng, Mingyang Su, You Zhou, Xuefei Zhe, Naoya Iwamoto, Bo + +Zheng, and Michael J. Black. EMAGE: Towards unified holistic co-speech gesture generation via expressive masked audio gesture modeling. In Computer Vision and Pattern Recognition (CVPR), 2024. 1, 2, 6 +[57] Yuejiang Liu, Riccardo Cadei, Jonas Schweizer, Sherwin Bahmani, and Alexandre Alahi. Towards robust and adaptive motion forecasting: A causal representation perspective. In Computer Vision and Pattern Recognition (CVPR), 2022. 2 +[58] Matthew M. Loper, Naureen Mahmood, and Michael J. Black. MoSh: Motion and shape capture from sparse markers. Transactions on Graphics (TOG), 33(6):220:1-220:13, 2014. 5 +[59] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In International Conference on Learning Representations (ICLR), 2019. 5 +[60] Shuhong Lu, Youngwoo Yoon, and Andrew W. Feng. Cospeech gesture synthesis using discrete gesture token learning. International Conference on Intelligent Robots and Systems (IROS), pages 9808-9815, 2023. 2 +[61] Birgit Lugarin, Catherine Pelachaud, and David Traum, editors. The Handbook on Socially Interactive Agents: 20 Years of Research on Embodied Conversational Agents, Intelligent Virtual Agents, and Social Robotics Volume 1: Methods, Behavior, Cognition. 2021. 1 +[62] Naureen Mahmood, Nima Ghorbani, Nikolaus F. Troje, Gerard Pons-Moll, and Michael J. Black. Amass: Archive of motion capture as surface shapes. In International Conference on Computer Vision (ICCV), 2019. 2, 5 +[63] Stacy Marsella, Yuyu Xu, Margaux Lhommet, Andrew Feng, Stefan Scherer, and Ari Shapiro. Virtual character performance from speech. In SIGGRAPH/Eurographics Symposium on Computer Animation (SCA), page 25-35, 2013. 2 +[64] Julieta Martinez, Michael J. Black, and Javier Romero. On human motion prediction using recurrent neural networks. In Computer Vision and Pattern Recognition (CVPR), 2017. 2 +[65] David Mcneill. Hand and mind: What gestures reveal about thought. Bibliovault OAI Repository, the University of Chicago Press, 27, 1994. 1 +[66] Albert Mehrabian. Nonverbal communication. Aldine-Atherton Chicago, 1972. 1 +[67] Shivam Mehta, Siyang Wang, Simon Alexanderson, Jonas Beskow, Éva Székely, and Gustav Eje Henter. Diff-TTSG: Denoising probabilistic integrated speech and gesture synthesis. In Proc. ISCA Speech Synthesis Workshop (SSW), pages 150–156, 2023. 2 +[68] Davide Moltisanti, Jinyi Wu, Bo Dai, and Chen Change Loy. BRACE: the breakdancing competition dataset for dance motion synthesis. In European Conference on Computer Vision (ECCV), pages 329-344, 2022. 2 +[69] Alexander Quinn Nichol and Prafulla Dhariwal. Improved denoising diffusion probabilistic models. In International Conference on Machine Learning (ICML), pages 8162-8171, 2021. 4 + +[70] Dirk Ormoneit, Michael J. Black, T. Hastie, and H. Kjellström. Representing cyclic human motion using functional analysis. Image and Vision Computing (IVC), 2005. 2 +[71] Kunkun Pang, Dafei Qin, Yingruo Fan, Julian Habekost, Takaaki Shiratori, Junichi Yamagishi, and Taku Komura. BodyFormer: Semantics-guided 3d body gesture synthesis with transformer. Transactions on Graphics (TOG), 42(4), 2023. 2 +[72] Daniel S. Park, William Chan, Yu Zhang, Chung-Cheng Chiu, Barret Zoph, Ekin D. Cubuk, and Quoc V. Le. SpecAugment: A simple data augmentation method for automatic speech recognition. In Interspeech 2019. ISCA, 2019. 5 +[73] Georgios Pavlakos, Vasileios Choutas, Nima Ghorbani, Timo Bolkart, Ahmed A. A. Osman, Dimitrios Tzionas, and Michael J. Black. Expressive body capture: 3D hands, face, and body from a single image. In Computer Vision and Pattern Recognition (CVPR), pages 10975-10985, 2019. 2, 3 +[74] Ziqiao Peng, Haoyu Wu, Zhenbo Song, Hao Xu, Xiangyu Zhu, Jun He, Hongyan Liu, and Zhaoxin Fan. EmoTalk: Speech-driven emotional disentanglement for 3d face animation. In International Conference on Computer Vision (ICCV), pages 20687-20697, 2023. 2, 3, 8 +[75] Mathis Petrovich, Michael J. Black, and Gül Varol. Action-conditioned 3D human motion synthesis with transformer VAE. In International Conference on Computer Vision (ICCV), 2021. 2 +[76] Mathis Petrovich, Michael J. Black, and Gül Varol. Action-conditioned 3D human motion synthesis with transformer VAE. In International Conference on Computer Vision (ICCV), 2021. 4, 6 +[77] Mathis Petrovich, Michael J. Black, and Gül Varol. TEMOS: Generating diverse human motions from textual descriptions. In European Conference on Computer Vision (ECCV), 2022. 2 +[78] Paolo Petta, Catherine Pelachaud, and Roddy Cowie. Emotion-Oriented Systems: The Humaine Handbook. Springer Publishing Company, Incorporated, 2011. 1 +[79] Rosalind W. Picard. Affective Computing. MIT Press, 1997. 1 +[80] I. Poggi, C. Pelachaud, F. de Rosis, V. Carofiglio, and B. De Carolis. Greta. A Believable Embodied Conversational Agent, pages 3-25. Springer, 2005. 2 +[81] Xingqun Qi, Chen Liu, Lincheng Li, Jie Hou, Haoran Xin, and Xin Yu. EmotionGesture: Audio-driven diverse emotional co-speech 3d gesture generation, 2023. 3 +[82] Xingqun Qi, Chen Liu, Muyi Sun, Lincheng Li, Changjie Fan, and Xin Yu. Diverse 3D hand gesture prediction from body dynamics by bilateral hand disentanglement. In Computer Vision and Pattern Recognition (CVPR), pages 4616-4626, 2023. 2 +[83] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning (ICML), pages 8748-8763. PMLR, 2021. 3 + +[84] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Computer Vision and Pattern Recognition (CVPR), pages 10674-10685, 2022. 4, 5 +[85] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-Net: Convolutional networks for biomedical image segmentation. In Medical Image Computing and Computer-Assisted Intervention - MICCAI, pages 234-241, 2015. 4 +[86] Tim Salzmann, Marco Pavone, and Markus Ryll. Motron: Multimodal Probabilistic Human Motion Forecasting. In Computer Vision and Pattern Recognition (CVPR), 2022. 2 +[87] Klaus R Scherer. Vocal communication of emotion: A review of research paradigms. Speech Communication, 40 (1-2):227-256, 2003. 1 +[88] Maximilian Seitzer. pytorch-fid: FID Score for PyTorch. https://github.com/mseitzer/pytorch-fid, 2020. Version 0.3.0.6 +[89] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In International Conference on Learning Representations (ICLR), 2021. 5 +[90] Kim Sung-Bin, Lee Hyun, Da Hye Hong, Suekyeong Nam, Janghoon Ju, and Tae-Hyun Oh. Laughtalk: Expressive 3d talking head generation with laughter. In Winter Conference on Applications of Computer Vision (WACV), pages 6404-6413, 2024. 2 +[91] Mingyang Sun†, Mengchen Zhao†, Yaqing Hou*, Minglei Li, Huang Xu, Songcen Xu, and Jianye Hao. Cospeech gesture synthesis by reinforcement learning with contrastive pre-trained rewards. In Computer Vision and Pattern Recognition (CVPR), 2023. 2 +[92] Marcus Thiébaux, Stacy Marsella, Andrew N. Marshall, and Marcelo Kallmann. SmartBody: behavior realization for embodied conversational agents. In Adaptive Agents and Multi-Agent Systems (AAMAS), 2008. 2 +[93] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Hervé Jégou. Training data-efficient image transformers & distillation through attention. In International Conference on Machine Learning (ICML), pages 10347-10357, 2021. 4 +[94] Jonathan Tseng, Rodrigo Castellon, and Karen Liu. Edge: Editable dance generation from music. In Computer Vision and Pattern Recognition (CVPR), 2023. 2 +[95] R. Williams. The Animator's Survival Kit: A Manual of Methods, Principles and Formulas for Classical, Computer Games, Stop Motion and Internet Animators. Farrar, Straus and Giroux, 2012. 2 +[96] Jingyu Wu, Shi Chen, Shuyu Gan, Weijun Li, Changyuan Yang, and Lingyun Sun. Cultural self-adaptive multimodal gesture generation based on multiple culture gesture dataset. In ACM International Conference on Multimedia (MM), page 3538-3549, 2023. 3 +[97] Sicheng Yang, Zhiyong Wu, Minglei Li, Zhensong Zhang, Lei Hao, Weihong Bao, Ming Cheng, and Long Xiao. Diffusestylegusture: Stylized audio-driven co-speech gesture generation with diffusion models. In International Joint Conference on Artificial Intelligence (IJCAI), 2023. 3, 6 + +[98] Sicheng Yang, Zhiyong Wu, Minglei Li, Zhensong Zhang, Lei Hao, Weihong Bao, and Haolin Zhuang. Qpgesture: Quantization-based and phase-guided motion matching for natural speech-driven gesture generation. In Computer Vision and Pattern Recognition (CVPR), pages 2321-2330, 2023. 2 +[99] Payam Jome Yazdian, Mo Chen, and Angelica Lim. Gesture2vec: Clustering gestures using representation learning methods for co-speech gesture generation. In International Conference on Intelligent Robots and Systems (IROS), pages 3100-3107, 2022. 2 +[100] Hongwei Yi. Show github. https://github.com/yhw-yhw/SHOW, 2023.6 +[101] Hongwei Yi, Hualin Liang, Yifei Liu, Qiong Cao, Yandong Wen, Timo Bolkart, Dacheng Tao, and Michael J Black. Generating holistic 3d human motion from speech. In Computer Vision and Pattern Recognition (CVPR), pages 469-480, 2023. 1, 2, 6, 7 +[102] Li-Ping Yin, Yijun Wang, Tianyu He, Jinming Liu, Wei Zhao, Bohan Li, Xin Jin, and Jianxin Lin. Emog: Synthesizing emotive co-speech 3d gesture with diffusion model. ArXiv, abs/2306.11496, 2023. 3 +[103] Youngwoo Yoon, Woo-Ri Ko, Minsu Jang, Jaeyeon Lee, Jaehong Kim, and Geehyuk Lee. Robots learn social skills: End-to-end learning of co-speech gesture generation for humanoid robots. In International Conference on Robotics and Automation (ICRA), 2019. 6 +[104] Youngwoo Yoon, Bok Cha, Joo-Haeng Lee, Minsu Jang, Jaeyeon Lee, Jaehong Kim, and Geehyuk Lee. Speech gesture generation from the trimodal context of text, audio, and speaker identity. Transactions on Graphics (TOG), 39(6), 2020. 2, 6 +[105] Youngwoo Yoon, Pieter Wolfert, Taras Kucherenko, Carla Viegas, Teodor Nikolov, Mihail Tsakov, and Gustav Eje Henter. The genea challenge 2022: A large evaluation of data-driven co-speech gesture generation. In International Conference on Multimodal Interaction, 2022. 2 +[106] Ye Yuan and Kris M. Kitani. DLow: Diversifying latent flows for diverse human motion prediction. In European Conference on Computer Vision (ECCV), 2020. 2 +[107] Yicheng Zhong, Huawei Wei, Peiji Yang, and Zhisheng Wang. Expclip: Bridging text and facial expressions via semantic alignment. In AAAI Conference on Artificial Intelligence, pages 7614-7622, 2024. 2 +[108] Yi Zhou, Connelly Barnes, Jingwan Lu, Jimei Yang, and Hao Li. On the continuity of rotation representations in neural networks. In Computer Vision and Pattern Recognition (CVPR), pages 5745-5753, 2019. 3 +[109] Yi Zhou, Jingwan Lu, Connelly Barnes, Jimei Yang, Sitao Xiang, and Hao Li. Generative tinkering: Long-term inbetweening of 3d human motions. arXiv:2005.08891, 2020. 2 +[110] Lingting Zhu, Xian Liu, Xuanyu Liu, Rui Qian, Ziwei Liu, and Lequan Yu. Taming diffusion models for audio-driven co-speech gesture generation. In Computer Vision and Pattern Recognition (CVPR), pages 10544–10553, 2023. 2, 8 \ No newline at end of file diff --git a/2024/Emotional Speech-driven 3D Body Animation via Disentangled Latent Diffusion/images.zip b/2024/Emotional Speech-driven 3D Body Animation via Disentangled Latent Diffusion/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..b39ba3d0b9351c92a3abfbb23131213da2914348 --- /dev/null +++ b/2024/Emotional Speech-driven 3D Body Animation via Disentangled Latent Diffusion/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2bce64c590aa148e9a2425d52cf1b433bbb7029822ef60f7607bfa52b2ff1106 +size 402881 diff --git a/2024/Emotional Speech-driven 3D Body Animation via Disentangled Latent Diffusion/layout.json b/2024/Emotional Speech-driven 3D Body Animation via Disentangled Latent Diffusion/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..84b0507a990603d12956bd3db0ab0ed8a828243e --- /dev/null +++ b/2024/Emotional Speech-driven 3D Body Animation via Disentangled Latent Diffusion/layout.json @@ -0,0 +1,10804 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 50, + 103, + 543, + 120 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 50, + 103, + 543, + 120 + ], + "spans": [ + { + "bbox": [ + 50, + 103, + 543, + 120 + ], + "type": "text", + "content": "Emotional Speech-driven 3D Body Animation via Disentangled Latent Diffusion" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 153, + 142, + 228, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 153, + 142, + 228, + 156 + ], + "spans": [ + { + "bbox": [ + 153, + 142, + 228, + 156 + ], + "type": "text", + "content": "Kiran Chhatre" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 250, + 143, + 332, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 250, + 143, + 332, + 156 + ], + "spans": [ + { + "bbox": [ + 250, + 143, + 332, + 156 + ], + "type": "text", + "content": "Radek Danček2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 354, + 143, + 446, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 354, + 143, + 446, + 156 + ], + "spans": [ + { + "bbox": [ + 354, + 143, + 446, + 156 + ], + "type": "text", + "content": "Nikos Athanasiou2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 94, + 157, + 187, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 94, + 157, + 187, + 171 + ], + "spans": [ + { + "bbox": [ + 94, + 157, + 187, + 171 + ], + "type": "text", + "content": "Giorgio Becherini²" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 209, + 157, + 303, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 209, + 157, + 303, + 171 + ], + "spans": [ + { + "bbox": [ + 209, + 157, + 303, + 171 + ], + "type": "text", + "content": "Christopher Peters1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 326, + 157, + 413, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 157, + 413, + 171 + ], + "spans": [ + { + "bbox": [ + 326, + 157, + 413, + 171 + ], + "type": "text", + "content": "Michael J. Black" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 435, + 157, + 510, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 435, + 157, + 510, + 171 + ], + "spans": [ + { + "bbox": [ + 435, + 157, + 510, + 171 + ], + "type": "text", + "content": "Timo Bolkart" + }, + { + "bbox": [ + 435, + 157, + 510, + 171 + ], + "type": "inline_equation", + "content": "^{2*}" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 171, + 543, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 171, + 543, + 186 + ], + "spans": [ + { + "bbox": [ + 55, + 171, + 543, + 186 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 55, + 171, + 543, + 186 + ], + "type": "text", + "content": "KTH Royal Institute of Technology, Sweden " + }, + { + "bbox": [ + 55, + 171, + 543, + 186 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 55, + 171, + 543, + 186 + ], + "type": "text", + "content": "Max Planck Institute for Intelligent Systems, Germany" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 143, + 213, + 191, + 225 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 213, + 191, + 225 + ], + "spans": [ + { + "bbox": [ + 143, + 213, + 191, + 225 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 45, + 239, + 290, + 550 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 239, + 290, + 550 + ], + "spans": [ + { + "bbox": [ + 45, + 239, + 290, + 550 + ], + "type": "text", + "content": "Existing methods for synthesizing 3D human gestures from speech have shown promising results, but they do not explicitly model the impact of emotions on the generated gestures. Instead, these methods directly output animations from speech without control over the expressed emotion. To address this limitation, we present AMUSE, an emotional speech-driven body animation model based on latent diffusion. Our observation is that content (i.e., gestures related to speech rhythm and word utterances), emotion, and personal style are separable. To account for this, AMUSE maps the driving audio to three disentangled latent vectors: one for content, one for emotion, and one for personal style. A latent diffusion model, trained to generate gesture motion sequences, is then conditioned on these latent vectors. Once trained, AMUSE synthesizes 3D human gestures directly from speech with control over the expressed emotions and style by combining the content from the driving speech with the emotion and style of another speech sequence. Randomly sampling the noise of the diffusion model further generates variations of the gesture with the same emotional expressivity. Qualitative, quantitative, and perceptual evaluations demonstrate that AMUSE outputs realistic gesture sequences. Compared to the state of the art, the generated gestures are better synchronized with the speech content, and better represent the emotion expressed by the input speech. Our code is available at amuse.is.tue.mpg.de." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 47, + 575, + 127, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 575, + 127, + 586 + ], + "spans": [ + { + "bbox": [ + 47, + 575, + 127, + 586 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 46, + 596, + 287, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 596, + 287, + 693 + ], + "spans": [ + { + "bbox": [ + 46, + 596, + 287, + 693 + ], + "type": "text", + "content": "Animating 3D bodies from speech has a wide range of applications, such as telepresence in AR/VR, avatar animation in games and movies, and to embody interactive digital assistants. While methods for speech-driven 3D body animation have recently shown great progress [5, 7, 31, 56, 101], existing methods do not adequately address one crucial factor: the impact of emotion from the driving speech signal on the generated gestures. Emotions and their expressions play" + } + ] + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 323, + 213, + 529, + 308 + ], + "blocks": [ + { + "bbox": [ + 323, + 213, + 529, + 308 + ], + "lines": [ + { + "bbox": [ + 323, + 213, + 529, + 308 + ], + "spans": [ + { + "bbox": [ + 323, + 213, + 529, + 308 + ], + "type": "image", + "image_path": "4414e5a2abd995e7a359daea2dc2751ecdd1f332f098d2c2ca660461c2179747.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 323, + 308, + 533, + 316 + ], + "lines": [ + { + "bbox": [ + 323, + 308, + 533, + 316 + ], + "spans": [ + { + "bbox": [ + 323, + 308, + 533, + 316 + ], + "type": "text", + "content": "... choose a major that is easy to find a good job in the future..." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 323, + 321, + 527, + 375 + ], + "blocks": [ + { + "bbox": [ + 323, + 321, + 527, + 375 + ], + "lines": [ + { + "bbox": [ + 323, + 321, + 527, + 375 + ], + "spans": [ + { + "bbox": [ + 323, + 321, + 527, + 375 + ], + "type": "image", + "image_path": "44e4e54140317ad646e9f8780ec5ed0aebb5535002a5208ee58ef257be33ee5d.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 308, + 376, + 543, + 453 + ], + "blocks": [ + { + "bbox": [ + 308, + 376, + 543, + 453 + ], + "lines": [ + { + "bbox": [ + 308, + 376, + 543, + 453 + ], + "spans": [ + { + "bbox": [ + 308, + 376, + 543, + 453 + ], + "type": "image", + "image_path": "c127b0d6bd325b8ed5eee9e914bcb897ac31ee4e9c89592aeb40f1f0d40ef7c2.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 462, + 545, + 506 + ], + "lines": [ + { + "bbox": [ + 304, + 462, + 545, + 506 + ], + "spans": [ + { + "bbox": [ + 304, + 462, + 545, + 506 + ], + "type": "text", + "content": "Figure 1. Goal. AMUSE generates realistic emotional 3D body gestures directly from a speech sequence (top). It provides user control over the generated emotion by combining the driving speech sequence with a different emotional audio (bottom)." + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 532, + 545, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 532, + 545, + 651 + ], + "spans": [ + { + "bbox": [ + 304, + 532, + 545, + 651 + ], + "type": "text", + "content": "a fundamental role in human communication [29, 35, 65] and have become an important consideration when designing computer systems that interact with humans in a natural manner [78, 79]. They are of central concern when synthesizing human animations for a wide variety of application contexts, such as Socially Interactive Agents [61]. Because of this, speech-driven animation systems must not only align movement with the rhythm of the speech, but should also be capable of generating gestures that are perceived as expressing the suitable emotion." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 653, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 653, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 653, + 545, + 712 + ], + "type": "text", + "content": "Many factors contribute to the perception of emotion and personal idiosyncrasies, such as facial expressions [19], gaze and eye contact [42], physiological responses [47], tone of voice [87], body language [66], and gestures [39]. When it comes to 3D animation, the most relevant factors" + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 494, + 35 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 494, + 35 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 494, + 35 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 703, + 115, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 703, + 115, + 712 + ], + "spans": [ + { + "bbox": [ + 58, + 703, + 115, + 712 + ], + "type": "text", + "content": "*Now at Google." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1942" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 286, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 286, + 120 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 286, + 120 + ], + "type": "text", + "content": "are facial expressions, gestures, and body language [95]. While emotional speech-driven animation methods have recently been proposed for 3D faces [18, 74, 90, 107], animating emotional bodies from speech remains under-explored." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 121, + 287, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 121, + 287, + 384 + ], + "spans": [ + { + "bbox": [ + 46, + 121, + 287, + 384 + ], + "type": "text", + "content": "Generating gestures solely from speech with emotional control is a difficult task. First, the mapping from audio to body motion is a non-deterministic many-to-many mapping, which is difficult to model. Gestures across subjects can vary when uttering the same sentence, and a single individual's motions can change significantly across repetitions. Second, factoring out the impact of emotional state on the body motion from other, unknown factors, is difficult. This requires disentangling the effects of three different factors on the generated motion, namely content-based (i.e., gestures related to speech rhythm and word utterances), emotion-based, and those based on personal style. AMUSE addresses this by separating a speech sequence into content, emotion, and style latent vectors, which are then used to condition a latent diffusion model. Specifically, AMUSE consists of three main components: (1) an audio autoencoder trained to produce disentangled vectors of content, emotion, and style, (2) a 3D body motion prior in the form of a temporal variational autoencoder (VAE) to generate smooth and realistic gestures, and (3) a latent diffusion model, which generates 3D body motion given the input content, emotion, and style latent vectors." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 384, + 286, + 516 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 384, + 286, + 516 + ], + "spans": [ + { + "bbox": [ + 46, + 384, + 286, + 516 + ], + "type": "text", + "content": "Training such a model requires a speech-to-3D body dataset of sufficient scale, which is rich and diverse in speakers and emotions. BEAT [55] is a good candidate because it provides a large set of 3D gestures associated with single-person monologues. Unfortunately, the bodies are represented as skeletons, and it lacks face mocap markers and FLAME expressions. Instead, to produce realistic body animations, we require articulated 3D body surfaces. To overcome this, we convert BEAT sequences to SMPL-X [73] format using MoSh++ [62] and use the SMPL-X parameters for training. See [56] for comparison." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 517, + 286, + 588 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 517, + 286, + 588 + ], + "spans": [ + { + "bbox": [ + 46, + 517, + 286, + 588 + ], + "type": "text", + "content": "Our contributions are: (1) We present a framework to synthesize emotional 3D body gestures directly from speech. (2) We factor an input audio into disentangled content, emotion and style vectors, which enables us to separately control emotion in generated gestures. (3) We adapt temporal latent diffusion for multiple target conditions." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 601, + 133, + 613 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 601, + 133, + 613 + ], + "spans": [ + { + "bbox": [ + 47, + 601, + 133, + 613 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 622, + 271, + 634 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 622, + 271, + 634 + ], + "spans": [ + { + "bbox": [ + 47, + 622, + 271, + 634 + ], + "type": "text", + "content": "2.1. 3D Conditional Human Motion Generation" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 641, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 641, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 641, + 286, + 713 + ], + "type": "text", + "content": "Early works focus mostly on predicting [10, 16, 33, 41, 57, 64, 70, 86, 106, 109] or generating human motion [30, 49], but do not consider multi-modal control. Recently, conditional motion generation through other modalities, such as text [2, 8, 9, 17, 22, 28, 77], music [50, 68, 94], speech [32], or action labels [27, 75], has gained more attention. Be" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 72, + 544, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 544, + 95 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 544, + 95 + ], + "type": "text", + "content": "low, we focus on speech-driven motion generation methods, since they are the most relevant to our work." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 106, + 482, + 119 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 106, + 482, + 119 + ], + "spans": [ + { + "bbox": [ + 306, + 106, + 482, + 119 + ], + "type": "text", + "content": "2.2. Gesture Generation from Speech" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 126, + 545, + 305 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 126, + 545, + 305 + ], + "spans": [ + { + "bbox": [ + 304, + 126, + 545, + 305 + ], + "type": "text", + "content": "Rule-based gesture synthesis. Embodied conversational agents (ECA) are designed to interact and communicate with humans. Using the Behavior Markup Language (BML) [44] one can build rule-based systems for humanoids based on predefined behaviors [80]. This is used for completion of a storytelling task in an expressive manner [45]. The BEAT rule-based toolkit [14] enables adding non-verbal behavior on top of a pre-animated figure. Thiebaux et al. [92] develop an ECA by using procedural animation techniques and keyframe interpolation. Marsella et al. [63] design a generalized rule-based agent to generate expressions, eye gaze, and gestures from speech. Each of these approaches are based on non-trainable, rule-based techniques that may require substantial manual modelling effort to adapt to new tasks." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 307, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 307, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 307, + 545, + 713 + ], + "type": "text", + "content": "Data-driven gesture synthesis. More recently, data-driven methods have superseded rule-based systems. Yoon et al. [104] use a fusion of text, audio and upper body gestures to learn an upper body gesture avatar, but can only control the style of individual speakers by sampling from their latent space. SpeechGestureMatching [32] generates 3D facial meshes and 3D keypoints of the body and hands from speech, but the outputs are separated and the method does not provide control over the generations. QPGesture [98] uses phase to better align the generated 3D skeleton-based gesturing avatars with the audio input. Ginosar et al. [23] and Diverse-3D-Hand-Gesture-Prediction [82] generate hand and arm motions only. Audio2Gestures [48] encode motion and audio to a low-dimensional latent space and generate gestures. SEEG [53] aims to generate gestures that align well with the semantics of the speech. DiffTTSG [67] regresses speech and gestures at the same time, joining the two modalities in a single system. DiffGAN [3] retargets gestures across speakers in a low-resource setting. The GENEA challenge [105] tackles gesticulation from speech alone using the Talking-with-Hands dataset [46]. Gesture2Vec [99] uses a machine translation model to translate text into gesture chunks and output full sequences using such quantized representations. TalkSHOW [101] uses a VQ-VAE to generate 3D human bodies gesturing with facial expressions from speech segments, but in an uncontrolled manner. Similarly, Co-speech gesture [60] uses an RQ-VAE to generate different gestures from speech. Alternative gesture generation from speech methods have been proposed such as reinforcement learning [91], self-supervised pretraining [40], and diffusion [67, 110]. BodyFormer [71] introduces a dataset of pseudo-groundtruth and a transformer-based method for generating gestures from speech. However, none of these methods provide explicit emotional con" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1943" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 173, + 83 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 173, + 83 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 173, + 83 + ], + "type": "text", + "content": "trol over the generated motion." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 85, + 288, + 264 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 85, + 288, + 264 + ], + "spans": [ + { + "bbox": [ + 46, + 85, + 288, + 264 + ], + "type": "text", + "content": "For controllable generation, GestureDiffuCLIP [7] incorporates multiple conditions including CLIP [83] text features, video, or motion prompts via AdaIn [37] layers to generate gestures from speech, however, it does not allow explicit control over the emotion conveyed by the driving audio. ListenDenoiseAction [5] combines conformers and the DiffWave [43] architecture to generate gestures that can be controlled by a style vector, RhythmicGesticulator [6] disentangles the latent space into a vector related to the semantics of the gesture and one related to the subtle variations, while DisCo [54] models content and rhythm. StyleGestures [4] adapts MoGlow [34], demonstrating limited control over some motion attributes like the speed and expressiveness of gestures. DiffuseStyleGesture [97] uses diffusion to generate diverse gestures from speech." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 275, + 149, + 286 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 275, + 149, + 286 + ], + "spans": [ + { + "bbox": [ + 47, + 275, + 149, + 286 + ], + "type": "text", + "content": "2.3. Emotion Control" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 294, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 294, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 294, + 289, + 715 + ], + "type": "text", + "content": "Emotion classification and control has been little studied in 3D human motion generation with only a few methods using skeletal motion in multi-class classification. Ghaleb et al. [20] employ a spatio-temporal graph convolution network to classify gestures into four classes: preparation, stroke, retraction, and neutral. Li et al. [52], on the other hand, use hidden Markov models for emotion classification of human movement mocap data. Karras et al. [38] learn face animations of a single actor, and test their method on different tasks by modifying the latent vectors. However, there is no disentanglement mechanism, and they do not model the synchronization of the emotion with the facial motions. Recently, EmoTalk [74], animates emotional 3D faces from speech input with control over the emotion intensity and EMOTE [18] disentangles emotion and speech to allow emotion editing at test time. However, models solely intended for facial tasks like lipSyncing and capturing expressions might not smoothly adapt to the complexity of whole-body movements and distinct articulation. Regarding emotion-conditioned motion generation, Aberman et al. [1] show style-transfer from video data to motion and provide some style-based control, but do not address speech-driven emotional gestures. Similarly, the ZeroEGGs [21] dataset contains some emotional gesture controls but also includes more generic styles of motion. The method requires the input of arbitrary frames of desired motion to encode a style, thereby relying on motions and speech as conditions during inference. Text-driven emotional gesticulation, as explored by Bhattacharya et al. [11, 12], emphasizes the generation of gestures based on textual cues, incorporating additional conditions such as speech, speaker ID, seed poses, as well as valence, arousal, and dominance triplets. However, these approaches do not provide the means to distill explicit emotion features, limiting free control over the generated gestures. Closer to" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 304, + 72, + 547, + 253 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 253 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 253 + ], + "type": "text", + "content": "our work, EMoG [102] incorporates emotion cues from the BEAT dataset [55] to generate improved gesture quality without explicit emotion control. EmotionGesture [81] uses a TED Emotion Dataset and BEAT to incorporate emotion features in gesture generation and generate emotional gestures. Although they can generate emotional gestures, their method is not end-to-end and has no explicit motion control. Specifically, it uses an emotion-conditioned VAE after training to acquire diverse emotion features that are used to generate gestures without guarantees and control over emotion types. Wu et al. [96] introduce the first multi-cultural gesture dataset containing 200 individuals of 10 different cultures. In contrast to prior work, we explicitly control the emotions conveyed by the generated gestures solely through emotional speech without relying on additional conditions." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 306, + 261, + 362, + 274 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 261, + 362, + 274 + ], + "spans": [ + { + "bbox": [ + 306, + 261, + 362, + 274 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 281, + 547, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 281, + 547, + 389 + ], + "spans": [ + { + "bbox": [ + 304, + 281, + 547, + 389 + ], + "type": "text", + "content": "The AMUSE pipeline consists of two separately trained networks. The audio disentanglement module, which encodes input speech into latent vectors for content, emotion, and style is described in Sec. 3.2. The main architecture is described in Sec. 3.3. It consists of a 3D human motion prior coupled with a latent diffusion model. It takes random noise (or partially denoised latent vectors) on the input and outputs a human motion sequence. We introduce broader applications in gesture editing in Sec. 3.4." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 395, + 519, + 407 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 395, + 519, + 407 + ], + "spans": [ + { + "bbox": [ + 305, + 395, + 519, + 407 + ], + "type": "text", + "content": "3.1. Preliminary: Expressive 3D Body Model" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 413, + 547, + 570 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 413, + 547, + 570 + ], + "spans": [ + { + "bbox": [ + 304, + 413, + 547, + 570 + ], + "type": "text", + "content": "SMPL-X [73] is a 3D model of the body surface. SMPL-X is defined as function " + }, + { + "bbox": [ + 304, + 413, + 547, + 570 + ], + "type": "inline_equation", + "content": "M(\\boldsymbol{\\beta}, \\boldsymbol{\\theta}, \\boldsymbol{\\psi})" + }, + { + "bbox": [ + 304, + 413, + 547, + 570 + ], + "type": "text", + "content": " that produces a 3D body mesh. It is parameterized by identity shape " + }, + { + "bbox": [ + 304, + 413, + 547, + 570 + ], + "type": "inline_equation", + "content": "\\boldsymbol{\\beta} \\in \\mathbb{R}^{300}" + }, + { + "bbox": [ + 304, + 413, + 547, + 570 + ], + "type": "text", + "content": ", pose " + }, + { + "bbox": [ + 304, + 413, + 547, + 570 + ], + "type": "inline_equation", + "content": "\\boldsymbol{\\theta} \\in \\mathbb{R}^{J \\times 3}" + }, + { + "bbox": [ + 304, + 413, + 547, + 570 + ], + "type": "text", + "content": " including finger articulation for rotations around " + }, + { + "bbox": [ + 304, + 413, + 547, + 570 + ], + "type": "inline_equation", + "content": "J" + }, + { + "bbox": [ + 304, + 413, + 547, + 570 + ], + "type": "text", + "content": " joints, and facial expression " + }, + { + "bbox": [ + 304, + 413, + 547, + 570 + ], + "type": "inline_equation", + "content": "\\boldsymbol{\\psi} \\in \\mathbb{R}^{100}" + }, + { + "bbox": [ + 304, + 413, + 547, + 570 + ], + "type": "text", + "content": ". We adopt the continuous 6D rotation representation for training following Zhou et al. [108], making " + }, + { + "bbox": [ + 304, + 413, + 547, + 570 + ], + "type": "inline_equation", + "content": "\\boldsymbol{\\theta} \\in \\mathbb{R}^{J \\times 6}" + }, + { + "bbox": [ + 304, + 413, + 547, + 570 + ], + "type": "text", + "content": ". Given pose parameters and any shape parameter, we can obtain body mesh vertices " + }, + { + "bbox": [ + 304, + 413, + 547, + 570 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 304, + 413, + 547, + 570 + ], + "type": "text", + "content": " using the differentiable SMPL-X layer [73]. As the focus of our paper is on synthesizing body gestures and not locomotion, we disregard 8 joints that correspond those of the lower body joint poses, leaving " + }, + { + "bbox": [ + 304, + 413, + 547, + 570 + ], + "type": "inline_equation", + "content": "J = 47" + }, + { + "bbox": [ + 304, + 413, + 547, + 570 + ], + "type": "text", + "content": ". Further, we omit the facial expression parameters, i.e., set " + }, + { + "bbox": [ + 304, + 413, + 547, + 570 + ], + "type": "inline_equation", + "content": "\\boldsymbol{\\psi} = \\mathbf{0}" + }, + { + "bbox": [ + 304, + 413, + 547, + 570 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 575, + 476, + 588 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 575, + 476, + 588 + ], + "spans": [ + { + "bbox": [ + 305, + 575, + 476, + 588 + ], + "type": "text", + "content": "3.2. Speech Disentanglement Model" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 594, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 547, + 713 + ], + "type": "text", + "content": "Architecture. The goal of the this model is to factor an input speech into three disentangled latent representations, one for content (i.e., the words spoken), one for emotion, and one for personal style. To do so, we devise a specialized encoder-decoder architecture with three separate encoders, one for each latent space. We denote the encoders as: " + }, + { + "bbox": [ + 304, + 594, + 547, + 713 + ], + "type": "inline_equation", + "content": "E_{c}(a) = c" + }, + { + "bbox": [ + 304, + 594, + 547, + 713 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 594, + 547, + 713 + ], + "type": "inline_equation", + "content": "E_{e}(a) = e" + }, + { + "bbox": [ + 304, + 594, + 547, + 713 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 594, + 547, + 713 + ], + "type": "inline_equation", + "content": "E_{s}(a) = s" + }, + { + "bbox": [ + 304, + 594, + 547, + 713 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 594, + 547, + 713 + ], + "type": "inline_equation", + "content": "a" + }, + { + "bbox": [ + 304, + 594, + 547, + 713 + ], + "type": "text", + "content": " is the input filterbank, " + }, + { + "bbox": [ + 304, + 594, + 547, + 713 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 304, + 594, + 547, + 713 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 594, + 547, + 713 + ], + "type": "inline_equation", + "content": "e" + }, + { + "bbox": [ + 304, + 594, + 547, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 594, + 547, + 713 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 304, + 594, + 547, + 713 + ], + "type": "text", + "content": " denote the latent vectors for content, emotion and style and " + }, + { + "bbox": [ + 304, + 594, + 547, + 713 + ], + "type": "inline_equation", + "content": "E_{c}" + }, + { + "bbox": [ + 304, + 594, + 547, + 713 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 594, + 547, + 713 + ], + "type": "inline_equation", + "content": "E_{e}" + }, + { + "bbox": [ + 304, + 594, + 547, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 594, + 547, + 713 + ], + "type": "inline_equation", + "content": "E_{s}" + }, + { + "bbox": [ + 304, + 594, + 547, + 713 + ], + "type": "text", + "content": " are their encoders. The architecture of the three encoders follows the" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1944" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 62, + 69, + 547, + 209 + ], + "blocks": [ + { + "bbox": [ + 62, + 69, + 547, + 209 + ], + "lines": [ + { + "bbox": [ + 62, + 69, + 547, + 209 + ], + "spans": [ + { + "bbox": [ + 62, + 69, + 547, + 209 + ], + "type": "image", + "image_path": "f5a86adb69535c00ac5d1290edf434c200ec423010af30604e43ba061845a680.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 216, + 547, + 276 + ], + "lines": [ + { + "bbox": [ + 46, + 216, + 547, + 276 + ], + "spans": [ + { + "bbox": [ + 46, + 216, + 547, + 276 + ], + "type": "text", + "content": "Figure 2. Training. We train the motion prior " + }, + { + "bbox": [ + 46, + 216, + 547, + 276 + ], + "type": "inline_equation", + "content": "(\\mathcal{P}_E, \\mathcal{P}_D)" + }, + { + "bbox": [ + 46, + 216, + 547, + 276 + ], + "type": "text", + "content": " and the latent denoiser " + }, + { + "bbox": [ + 46, + 216, + 547, + 276 + ], + "type": "inline_equation", + "content": "\\Delta" + }, + { + "bbox": [ + 46, + 216, + 547, + 276 + ], + "type": "text", + "content": " jointly, while keeping the audio encoding networks frozen. In the forward pass, we take an input audio " + }, + { + "bbox": [ + 46, + 216, + 547, + 276 + ], + "type": "inline_equation", + "content": "a^{1:T}" + }, + { + "bbox": [ + 46, + 216, + 547, + 276 + ], + "type": "text", + "content": " and pose sequence " + }, + { + "bbox": [ + 46, + 216, + 547, + 276 + ], + "type": "inline_equation", + "content": "m^{1:T}" + }, + { + "bbox": [ + 46, + 216, + 547, + 276 + ], + "type": "text", + "content": ". Firstly, we do a forward pass of " + }, + { + "bbox": [ + 46, + 216, + 547, + 276 + ], + "type": "inline_equation", + "content": "m^{1:T}" + }, + { + "bbox": [ + 46, + 216, + 547, + 276 + ], + "type": "text", + "content": " through " + }, + { + "bbox": [ + 46, + 216, + 547, + 276 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_E" + }, + { + "bbox": [ + 46, + 216, + 547, + 276 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 216, + 547, + 276 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_D" + }, + { + "bbox": [ + 46, + 216, + 547, + 276 + ], + "type": "text", + "content": " and compute " + }, + { + "bbox": [ + 46, + 216, + 547, + 276 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{rec}" + }, + { + "bbox": [ + 46, + 216, + 547, + 276 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 216, + 547, + 276 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{Vrec}" + }, + { + "bbox": [ + 46, + 216, + 547, + 276 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 216, + 547, + 276 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{KL}" + }, + { + "bbox": [ + 46, + 216, + 547, + 276 + ], + "type": "text", + "content": ". Then, we apply the diffusion process to a gradient-detached sg " + }, + { + "bbox": [ + 46, + 216, + 547, + 276 + ], + "type": "inline_equation", + "content": "[z_m]" + }, + { + "bbox": [ + 46, + 216, + 547, + 276 + ], + "type": "text", + "content": " obtaining the noisy " + }, + { + "bbox": [ + 46, + 216, + 547, + 276 + ], + "type": "inline_equation", + "content": "z_m^{(D)}" + }, + { + "bbox": [ + 46, + 216, + 547, + 276 + ], + "type": "text", + "content": " which is then denoised with " + }, + { + "bbox": [ + 46, + 216, + 547, + 276 + ], + "type": "inline_equation", + "content": "\\Delta" + }, + { + "bbox": [ + 46, + 216, + 547, + 276 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 216, + 547, + 276 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{LD}" + }, + { + "bbox": [ + 46, + 216, + 547, + 276 + ], + "type": "text", + "content": " is computed. Finally, we use " + }, + { + "bbox": [ + 46, + 216, + 547, + 276 + ], + "type": "inline_equation", + "content": "\\Delta" + }, + { + "bbox": [ + 46, + 216, + 547, + 276 + ], + "type": "text", + "content": " to fully denoise " + }, + { + "bbox": [ + 46, + 216, + 547, + 276 + ], + "type": "inline_equation", + "content": "z_n" + }, + { + "bbox": [ + 46, + 216, + 547, + 276 + ], + "type": "text", + "content": " into gradient-detached sg " + }, + { + "bbox": [ + 46, + 216, + 547, + 276 + ], + "type": "inline_equation", + "content": "[z_{\\tilde{m}}]" + }, + { + "bbox": [ + 46, + 216, + 547, + 276 + ], + "type": "text", + "content": ", further decode " + }, + { + "bbox": [ + 46, + 216, + 547, + 276 + ], + "type": "inline_equation", + "content": "\\tilde{m}^{1:T}" + }, + { + "bbox": [ + 46, + 216, + 547, + 276 + ], + "type": "text", + "content": " using " + }, + { + "bbox": [ + 46, + 216, + 547, + 276 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_D" + }, + { + "bbox": [ + 46, + 216, + 547, + 276 + ], + "type": "text", + "content": ", and compute " + }, + { + "bbox": [ + 46, + 216, + 547, + 276 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{align}" + }, + { + "bbox": [ + 46, + 216, + 547, + 276 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 216, + 547, + 276 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{Valign}" + }, + { + "bbox": [ + 46, + 216, + 547, + 276 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 293, + 287, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 293, + 287, + 376 + ], + "spans": [ + { + "bbox": [ + 46, + 293, + 287, + 376 + ], + "type": "text", + "content": "design by Gong et al. [25, 26] (i.e., leveraging the DeiT visual transformer [93] adapted for processing filterbank images extracted from the input audio). The decoder takes the three latent vectors and produces a reconstructed filterbank. Formally " + }, + { + "bbox": [ + 46, + 293, + 287, + 376 + ], + "type": "inline_equation", + "content": "D(c, e, s) = \\widehat{a}" + }, + { + "bbox": [ + 46, + 293, + 287, + 376 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 293, + 287, + 376 + ], + "type": "inline_equation", + "content": "\\widehat{a}" + }, + { + "bbox": [ + 46, + 293, + 287, + 376 + ], + "type": "text", + "content": " denotes the reconstructed filterbank. The decoder architecture consists of a fusion module and transformer-encoder layers." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 377, + 288, + 547 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 377, + 288, + 547 + ], + "spans": [ + { + "bbox": [ + 46, + 377, + 288, + 547 + ], + "type": "text", + "content": "Training. The audio module is trained with a multiple loss terms that ensure that the three latent spaces are properly disentangled. In addition to the standard autoencoder reconstruction loss, we also employ three cross-reconstruction losses, in which we enforce the correct reconstruction of the audio signal where we modify one of the content, style or emotion latents. Additionally, we employ three loss terms on the latent vector predictions – namely emotion and style classification losses over " + }, + { + "bbox": [ + 46, + 377, + 288, + 547 + ], + "type": "inline_equation", + "content": "e" + }, + { + "bbox": [ + 46, + 377, + 288, + 547 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 377, + 288, + 547 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 46, + 377, + 288, + 547 + ], + "type": "text", + "content": ", and a content similarity loss between pairs of two content latent vectors extracted from audios that have the same spoken content. For a detailed description of the encoder-decoder architecture, a formal definition of the loss functions and a detailed description of the training process please refer to the Sup. Mat." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 552, + 195, + 563 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 552, + 195, + 563 + ], + "spans": [ + { + "bbox": [ + 47, + 552, + 195, + 563 + ], + "type": "text", + "content": "3.3. Gesture Generation Model" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "type": "text", + "content": "Motion prior. Similar to [15, 76], our motion prior network is a VAE transformer architecture with encoder " + }, + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_E" + }, + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "type": "text", + "content": " and decoder " + }, + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_D" + }, + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "type": "text", + "content": ". Specifically, both " + }, + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_E" + }, + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_D" + }, + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "type": "text", + "content": " follow a U-Net-like [85] structure with skip connections between transformer blocks (see Sup. Mat. for details). The positional embeddings are learnable and injected into each multi-head attention layer, following the design of Carion et al. [13]. Formally, the encoder takes a sequence of " + }, + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "type": "text", + "content": " frames of the SMPL-X pose vectors " + }, + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "type": "inline_equation", + "content": "m^{1:T} \\in \\mathbb{R}^{6J \\times T}" + }, + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "type": "text", + "content": " and the first two tokens of its output, " + }, + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\mu \\in \\mathbb{R}^{d_m}" + }, + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\Sigma \\in \\mathbb{R}^{d_m \\times d_m}" + }, + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "type": "text", + "content": " are used to extract the motion latent " + }, + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "type": "inline_equation", + "content": "z_m \\in \\mathbb{R}^{d_m}" + }, + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "type": "text", + "content": " via the reparametrization trick. The decoder takes zero positional encodings as" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 293, + 545, + 328 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 293, + 545, + 328 + ], + "spans": [ + { + "bbox": [ + 304, + 293, + 545, + 328 + ], + "type": "text", + "content": "query input and the motion latent is fed as memory to every cross-attention transformer layer, producing the reconstructed motion " + }, + { + "bbox": [ + 304, + 293, + 545, + 328 + ], + "type": "inline_equation", + "content": "\\hat{m}^{1:T}" + }, + { + "bbox": [ + 304, + 293, + 545, + 328 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 329, + 545, + 377 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 329, + 545, + 377 + ], + "spans": [ + { + "bbox": [ + 304, + 329, + 545, + 377 + ], + "type": "text", + "content": "Diffusion process. The forward diffusion process is similar to [36, 69]. We employ fixed variance and linearly scaled noise scheduler. We add noise to the motion latent " + }, + { + "bbox": [ + 304, + 329, + 545, + 377 + ], + "type": "inline_equation", + "content": "z_{m}" + }, + { + "bbox": [ + 304, + 329, + 545, + 377 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 304, + 329, + 545, + 377 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 304, + 329, + 545, + 377 + ], + "type": "text", + "content": " diffusion timesteps to obtain " + }, + { + "bbox": [ + 304, + 329, + 545, + 377 + ], + "type": "inline_equation", + "content": "z^{(D)}" + }, + { + "bbox": [ + 304, + 329, + 545, + 377 + ], + "type": "text", + "content": " following:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 323, + 384, + 527, + 399 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 323, + 384, + 527, + 399 + ], + "spans": [ + { + "bbox": [ + 323, + 384, + 527, + 399 + ], + "type": "interline_equation", + "content": "q \\left(z _ {m} ^ {\\left(t _ {d}\\right)} \\mid z _ {m} ^ {(0)}\\right) = \\mathcal {N} \\left(z _ {m} ^ {\\left(t _ {d}\\right)}; \\sqrt {\\bar {\\alpha} _ {t _ {d}}} z _ {m} ^ {(0)}, (1 - \\bar {\\alpha} _ {t _ {d}}) \\mathbf {I}\\right),", + "image_path": "196dae283600bd0a2000835776d64594fe54dbebd37af20e1af5ac7641064957.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 407, + 545, + 431 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 407, + 545, + 431 + ], + "spans": [ + { + "bbox": [ + 304, + 407, + 545, + 431 + ], + "type": "text", + "content": "with " + }, + { + "bbox": [ + 304, + 407, + 545, + 431 + ], + "type": "inline_equation", + "content": "\\alpha_{t_d} = 1 - \\beta_{t_d}" + }, + { + "bbox": [ + 304, + 407, + 545, + 431 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 407, + 545, + 431 + ], + "type": "inline_equation", + "content": "\\bar{\\alpha}_{t_d} = \\prod_{s=1}^{t_d} \\alpha_s" + }, + { + "bbox": [ + 304, + 407, + 545, + 431 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 407, + 545, + 431 + ], + "type": "inline_equation", + "content": "\\beta_{t_d}" + }, + { + "bbox": [ + 304, + 407, + 545, + 431 + ], + "type": "text", + "content": " denotes diffusion process variance." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 432, + 545, + 554 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 432, + 545, + 554 + ], + "spans": [ + { + "bbox": [ + 304, + 432, + 545, + 554 + ], + "type": "text", + "content": "Conditional denoising process. The denoising process consists of iteratively denoising a conditioned noisy motion latent vector to obtain the denoised motion latent " + }, + { + "bbox": [ + 304, + 432, + 545, + 554 + ], + "type": "inline_equation", + "content": "z_{\\tilde{m}^{1:T}}" + }, + { + "bbox": [ + 304, + 432, + 545, + 554 + ], + "type": "text", + "content": ". Our denoiser " + }, + { + "bbox": [ + 304, + 432, + 545, + 554 + ], + "type": "inline_equation", + "content": "\\Delta" + }, + { + "bbox": [ + 304, + 432, + 545, + 554 + ], + "type": "text", + "content": " is a latent variable model [84] and its architecture is similar to the U-Net-like structure of the motion prior encoder " + }, + { + "bbox": [ + 304, + 432, + 545, + 554 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_E" + }, + { + "bbox": [ + 304, + 432, + 545, + 554 + ], + "type": "text", + "content": ". The input of the model is a concatenation of: " + }, + { + "bbox": [ + 304, + 432, + 545, + 554 + ], + "type": "inline_equation", + "content": "z_m^{(t_d)}" + }, + { + "bbox": [ + 304, + 432, + 545, + 554 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 432, + 545, + 554 + ], + "type": "inline_equation", + "content": "\\mathrm{SE}(t_d)" + }, + { + "bbox": [ + 304, + 432, + 545, + 554 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 432, + 545, + 554 + ], + "type": "inline_equation", + "content": "c, e, s \\in \\mathbb{R}^{256}" + }, + { + "bbox": [ + 304, + 432, + 545, + 554 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 432, + 545, + 554 + ], + "type": "inline_equation", + "content": "\\mathrm{SE}(t_d)" + }, + { + "bbox": [ + 304, + 432, + 545, + 554 + ], + "type": "text", + "content": " is a sinusoidal positional encoding of diffusion timestep " + }, + { + "bbox": [ + 304, + 432, + 545, + 554 + ], + "type": "inline_equation", + "content": "t_d" + }, + { + "bbox": [ + 304, + 432, + 545, + 554 + ], + "type": "text", + "content": " as defined in [36]. " + }, + { + "bbox": [ + 304, + 432, + 545, + 554 + ], + "type": "inline_equation", + "content": "\\Delta" + }, + { + "bbox": [ + 304, + 432, + 545, + 554 + ], + "type": "text", + "content": " iteratively denoises through each reversed diffusion step:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 351, + 559, + 499, + 575 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 351, + 559, + 499, + 575 + ], + "spans": [ + { + "bbox": [ + 351, + 559, + 499, + 575 + ], + "type": "interline_equation", + "content": "z _ {m} ^ {(t _ {d} - 1)} = \\Delta \\left(\\left[ z _ {m} ^ {(t _ {d})}, \\operatorname {S E} \\left(t _ {d}\\right), c, e, s \\right]\\right).", + "image_path": "c3e2f4162c659ee630a360076702f6972bc487720230c64b065602fc19aad82b.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 582, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 582, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 582, + 545, + 714 + ], + "type": "text", + "content": "Training. We optimize the motion prior and the latent denoiser jointly to ensure audio-motion latent code alignment during conditional fusion in the denoising process using a 3-step forward pass through the gesture generation model. First, following standard VAE practice, we reconstruct " + }, + { + "bbox": [ + 304, + 582, + 545, + 714 + ], + "type": "inline_equation", + "content": "\\hat{m}^{1:T}" + }, + { + "bbox": [ + 304, + 582, + 545, + 714 + ], + "type": "text", + "content": " by the motion prior forward pass. As shown in Fig. 2, we then disable gradient calculation in " + }, + { + "bbox": [ + 304, + 582, + 545, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_E" + }, + { + "bbox": [ + 304, + 582, + 545, + 714 + ], + "type": "text", + "content": " to infer the intermediate motion latent " + }, + { + "bbox": [ + 304, + 582, + 545, + 714 + ], + "type": "inline_equation", + "content": "\\mathrm{sg}[z_m]" + }, + { + "bbox": [ + 304, + 582, + 545, + 714 + ], + "type": "text", + "content": ", which serves as input to the denoiser. At this stage, we obtain the denoiser noise prediction, " + }, + { + "bbox": [ + 304, + 582, + 545, + 714 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 304, + 582, + 545, + 714 + ], + "type": "text", + "content": " and use to compute the diffusion model gradients. Finally, in the third step we com" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1945" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 71, + 286, + 121 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 71, + 286, + 121 + ], + "spans": [ + { + "bbox": [ + 47, + 71, + 286, + 121 + ], + "type": "text", + "content": "pute " + }, + { + "bbox": [ + 47, + 71, + 286, + 121 + ], + "type": "inline_equation", + "content": "\\tilde{m}^{1:T} = \\mathcal{P}_D(\\mathrm{sg}[z_{\\tilde{m}}])" + }, + { + "bbox": [ + 47, + 71, + 286, + 121 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 47, + 71, + 286, + 121 + ], + "type": "inline_equation", + "content": "z_{\\tilde{m}}" + }, + { + "bbox": [ + 47, + 71, + 286, + 121 + ], + "type": "text", + "content": " is obtained by iteratively using the " + }, + { + "bbox": [ + 47, + 71, + 286, + 121 + ], + "type": "inline_equation", + "content": "\\Delta" + }, + { + "bbox": [ + 47, + 71, + 286, + 121 + ], + "type": "text", + "content": " to obtain a fully denoised latent from " + }, + { + "bbox": [ + 47, + 71, + 286, + 121 + ], + "type": "inline_equation", + "content": "z_n^{(t_D)} \\sim \\mathcal{N}(\\mathbf{0}, \\mathbf{I})" + }, + { + "bbox": [ + 47, + 71, + 286, + 121 + ], + "type": "text", + "content": ". We indicate computations done without gradients with a stop-gradient operation " + }, + { + "bbox": [ + 47, + 71, + 286, + 121 + ], + "type": "inline_equation", + "content": "\\mathrm{sg}[\\cdot]" + }, + { + "bbox": [ + 47, + 71, + 286, + 121 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 122, + 286, + 170 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 122, + 286, + 170 + ], + "spans": [ + { + "bbox": [ + 47, + 122, + 286, + 170 + ], + "type": "text", + "content": "Losses. To train the motion prior, we include the standard VAE losses, namely the reconstruction loss on pose parameters " + }, + { + "bbox": [ + 47, + 122, + 286, + 170 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{rec}" + }, + { + "bbox": [ + 47, + 122, + 286, + 170 + ], + "type": "text", + "content": " and on vertex coordinates " + }, + { + "bbox": [ + 47, + 122, + 286, + 170 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{Vrec}" + }, + { + "bbox": [ + 47, + 122, + 286, + 170 + ], + "type": "text", + "content": " using the smooth L1 metric introduced in [24], which we denote as " + }, + { + "bbox": [ + 47, + 122, + 286, + 170 + ], + "type": "inline_equation", + "content": "L_1^s" + }, + { + "bbox": [ + 47, + 122, + 286, + 170 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 175, + 278, + 192 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 175, + 278, + 192 + ], + "spans": [ + { + "bbox": [ + 56, + 175, + 278, + 192 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {r e c} = L _ {1} ^ {s} (m ^ {1: T}, \\hat {m} ^ {1: T}), \\quad \\mathcal {L} _ {V r e c} = L _ {1} ^ {s} (V ^ {1: T}, \\hat {V} ^ {1: T}),", + "image_path": "412209d83cf5c05894e417319ba5b1f493250c6d1935226595e01018e91d7a42.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 198, + 286, + 246 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 198, + 286, + 246 + ], + "spans": [ + { + "bbox": [ + 47, + 198, + 286, + 246 + ], + "type": "text", + "content": "where the root-centered vertices " + }, + { + "bbox": [ + 47, + 198, + 286, + 246 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 47, + 198, + 286, + 246 + ], + "type": "text", + "content": " are obtained by feeding in pose parameters " + }, + { + "bbox": [ + 47, + 198, + 286, + 246 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 47, + 198, + 286, + 246 + ], + "type": "text", + "content": " to a differentiable SMPL-X layer (without learnable parameters) and a mean shape " + }, + { + "bbox": [ + 47, + 198, + 286, + 246 + ], + "type": "inline_equation", + "content": "\\beta = \\vec{0}" + }, + { + "bbox": [ + 47, + 198, + 286, + 246 + ], + "type": "text", + "content": ". The KL divergence loss of the motion prior is:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 65, + 253, + 268, + 286 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 253, + 268, + 286 + ], + "spans": [ + { + "bbox": [ + 65, + 253, + 268, + 286 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {K L} = \\frac {1}{2} \\left[ \\sum_ {i = 1} ^ {z} \\left(\\mu_ {i} ^ {2} + \\sigma_ {i} ^ {2}\\right) - \\sum_ {i = 1} ^ {z} \\left(l o g \\left(\\sigma_ {i} ^ {2}\\right) + 1\\right) \\right].", + "image_path": "b98fa1dbe208e9a998d3fa7dee4e69a1f1fffe75b2540ed27b3aaa21b0c113eb.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 293, + 286, + 339 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 293, + 286, + 339 + ], + "spans": [ + { + "bbox": [ + 47, + 293, + 286, + 339 + ], + "type": "text", + "content": "To ensure the alignment of the diffusion-generated motions and the input audio, we apply the alignment reconstruction loss on the inferred motion pose parameters and the vertex coordinates:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 346, + 285, + 363 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 346, + 285, + 363 + ], + "spans": [ + { + "bbox": [ + 48, + 346, + 285, + 363 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {a l i g n} = L _ {1} ^ {s} (m ^ {1: T}, \\tilde {m} ^ {1: T}), \\quad \\mathcal {L} _ {V a l i g n} = L _ {1} ^ {s} (V ^ {1: T}, \\tilde {V} ^ {1: T}).", + "image_path": "52a753eef5f21a74644ad7c9429e2b185f5a1851e4da83b783a2ed30d5428464.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 369, + 286, + 393 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 369, + 286, + 393 + ], + "spans": [ + { + "bbox": [ + 47, + 369, + 286, + 393 + ], + "type": "text", + "content": "Finally, we utilize the objective similar to [15, 36, 84] to supervise the denoiser:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 76, + 400, + 257, + 423 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 400, + 257, + 423 + ], + "spans": [ + { + "bbox": [ + 76, + 400, + 257, + 423 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {L D} = \\left\\| \\delta^ {(t _ {d})} - \\Delta (z _ {m} ^ {(t _ {d})}, \\mathrm {S E} (t _ {d}), c, e, s) \\right\\| _ {2} ^ {2},", + "image_path": "01057aaaa0c824f8d7d27af1790f5ca47f0b4f44d30a24b9d83bf63ef8e9e006.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 431, + 286, + 466 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 431, + 286, + 466 + ], + "spans": [ + { + "bbox": [ + 47, + 431, + 286, + 466 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 431, + 286, + 466 + ], + "type": "inline_equation", + "content": "\\delta^{(t_d)}" + }, + { + "bbox": [ + 47, + 431, + 286, + 466 + ], + "type": "text", + "content": " is the noise vector sampled from " + }, + { + "bbox": [ + 47, + 431, + 286, + 466 + ], + "type": "inline_equation", + "content": "\\mathcal{N}(\\mathbf{0},\\mathbf{I})" + }, + { + "bbox": [ + 47, + 431, + 286, + 466 + ], + "type": "text", + "content": " in the corresponding diffusion step " + }, + { + "bbox": [ + 47, + 431, + 286, + 466 + ], + "type": "inline_equation", + "content": "t_d" + }, + { + "bbox": [ + 47, + 431, + 286, + 466 + ], + "type": "text", + "content": ". The combined gesture model loss is:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 476, + 286, + 490 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 476, + 286, + 490 + ], + "spans": [ + { + "bbox": [ + 47, + 476, + 286, + 490 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {g e s} = \\mathcal {L} _ {r e c} + \\mathcal {L} _ {V r e c} + \\mathcal {L} _ {K L} + \\mathcal {L} _ {a l i g n} + \\mathcal {L} _ {V a l i g n} + \\mathcal {L} _ {L D}", + "image_path": "07de0fc761c29b6ef64395b03f38f4670006e3ad13d9206a1a02941b828ddb67.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 497, + 286, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 497, + 286, + 568 + ], + "spans": [ + { + "bbox": [ + 47, + 497, + 286, + 568 + ], + "type": "text", + "content": "Inference. We employ DDIM [89] to infer high quality conditional motion samples with a small number of denoising timesteps. During inference we draw a sample vector from " + }, + { + "bbox": [ + 47, + 497, + 286, + 568 + ], + "type": "inline_equation", + "content": "\\mathcal{N}(\\mathbf{0},\\mathbf{I})" + }, + { + "bbox": [ + 47, + 497, + 286, + 568 + ], + "type": "text", + "content": " to iteratively denoise in reversed timesteps. The denoised sample is then passed through the decoder " + }, + { + "bbox": [ + 47, + 497, + 286, + 568 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_D(z_{\\tilde{m}^{1:T}})" + }, + { + "bbox": [ + 47, + 497, + 286, + 568 + ], + "type": "text", + "content": " to obtain motion " + }, + { + "bbox": [ + 47, + 497, + 286, + 568 + ], + "type": "inline_equation", + "content": "\\tilde{m}^{1:T}" + }, + { + "bbox": [ + 47, + 497, + 286, + 568 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 47, + 575, + 144, + 588 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 575, + 144, + 588 + ], + "spans": [ + { + "bbox": [ + 47, + 575, + 144, + 588 + ], + "type": "text", + "content": "3.4. Gesture Editing" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 47, + 594, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 594, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 594, + 286, + 713 + ], + "type": "text", + "content": "Due to the disentangling of the inputs, AMUSE achieves semantic gesticulation control using two driving input audios. Specifically, given two input audio signals " + }, + { + "bbox": [ + 47, + 594, + 286, + 713 + ], + "type": "inline_equation", + "content": "a_1" + }, + { + "bbox": [ + 47, + 594, + 286, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 594, + 286, + 713 + ], + "type": "inline_equation", + "content": "a_2" + }, + { + "bbox": [ + 47, + 594, + 286, + 713 + ], + "type": "text", + "content": ", we extract their latent representations of content " + }, + { + "bbox": [ + 47, + 594, + 286, + 713 + ], + "type": "inline_equation", + "content": "c_1, c_2" + }, + { + "bbox": [ + 47, + 594, + 286, + 713 + ], + "type": "text", + "content": ", emotion " + }, + { + "bbox": [ + 47, + 594, + 286, + 713 + ], + "type": "inline_equation", + "content": "e_1, e_2" + }, + { + "bbox": [ + 47, + 594, + 286, + 713 + ], + "type": "text", + "content": ", and style " + }, + { + "bbox": [ + 47, + 594, + 286, + 713 + ], + "type": "inline_equation", + "content": "s_1, s_2" + }, + { + "bbox": [ + 47, + 594, + 286, + 713 + ], + "type": "text", + "content": ". Then, we simply initialize the denoising procedure of " + }, + { + "bbox": [ + 47, + 594, + 286, + 713 + ], + "type": "inline_equation", + "content": "\\Delta" + }, + { + "bbox": [ + 47, + 594, + 286, + 713 + ], + "type": "text", + "content": " with the triplet " + }, + { + "bbox": [ + 47, + 594, + 286, + 713 + ], + "type": "inline_equation", + "content": "(c_1, e_2, s_1)" + }, + { + "bbox": [ + 47, + 594, + 286, + 713 + ], + "type": "text", + "content": ", generating the gesture with the content and style of " + }, + { + "bbox": [ + 47, + 594, + 286, + 713 + ], + "type": "inline_equation", + "content": "a_1" + }, + { + "bbox": [ + 47, + 594, + 286, + 713 + ], + "type": "text", + "content": " but the emotion of input audio " + }, + { + "bbox": [ + 47, + 594, + 286, + 713 + ], + "type": "inline_equation", + "content": "a_2" + }, + { + "bbox": [ + 47, + 594, + 286, + 713 + ], + "type": "text", + "content": ". Similarly, instead of emotion we can also change the gesticulation style to that of the speaker of " + }, + { + "bbox": [ + 47, + 594, + 286, + 713 + ], + "type": "inline_equation", + "content": "a_2" + }, + { + "bbox": [ + 47, + 594, + 286, + 713 + ], + "type": "text", + "content": " by initializing with " + }, + { + "bbox": [ + 47, + 594, + 286, + 713 + ], + "type": "inline_equation", + "content": "(c_1, e_1, s_2)" + }, + { + "bbox": [ + 47, + 594, + 286, + 713 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 71, + 441, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 71, + 441, + 84 + ], + "spans": [ + { + "bbox": [ + 306, + 71, + 441, + 84 + ], + "type": "text", + "content": "4. Implementation Details" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 91, + 545, + 307 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 91, + 545, + 307 + ], + "spans": [ + { + "bbox": [ + 304, + 91, + 545, + 307 + ], + "type": "text", + "content": "MoCap data preparation. The BEAT [55] mocap sequences, captured in a Vicon system at " + }, + { + "bbox": [ + 304, + 91, + 545, + 307 + ], + "type": "inline_equation", + "content": "120\\mathrm{Hz}" + }, + { + "bbox": [ + 304, + 91, + 545, + 307 + ], + "type": "text", + "content": ", are downsampled to " + }, + { + "bbox": [ + 304, + 91, + 545, + 307 + ], + "type": "inline_equation", + "content": "30\\mathrm{Hz}" + }, + { + "bbox": [ + 304, + 91, + 545, + 307 + ], + "type": "text", + "content": " and processed using " + }, + { + "bbox": [ + 304, + 91, + 545, + 307 + ], + "type": "inline_equation", + "content": "\\mathrm{MoSh}++" + }, + { + "bbox": [ + 304, + 91, + 545, + 307 + ], + "type": "text", + "content": " [58, 62] to obtain SMPL-X parameters. Given a sequence of 3D mocap marker positions, we jointly optimize SMPL-X shape and pose parameters, 3D body translation, and embedding of the mocap markers in the SMPL-X surface. Once processed, the sequences are then divided according to the emotion annotations in the BEAT dataset. We use sequences of English speaking subjects in monologue speaking style for training and evaluating AMUSE. For each sequence we draw " + }, + { + "bbox": [ + 304, + 91, + 545, + 307 + ], + "type": "inline_equation", + "content": "m^{1:L}" + }, + { + "bbox": [ + 304, + 91, + 545, + 307 + ], + "type": "text", + "content": " at 30 FPS and concatenate with audio content " + }, + { + "bbox": [ + 304, + 91, + 545, + 307 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 304, + 91, + 545, + 307 + ], + "type": "text", + "content": ", emotion " + }, + { + "bbox": [ + 304, + 91, + 545, + 307 + ], + "type": "inline_equation", + "content": "e" + }, + { + "bbox": [ + 304, + 91, + 545, + 307 + ], + "type": "text", + "content": ", and style " + }, + { + "bbox": [ + 304, + 91, + 545, + 307 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 304, + 91, + 545, + 307 + ], + "type": "text", + "content": " latent vectors. Then, we segment it to 10-sec windows " + }, + { + "bbox": [ + 304, + 91, + 545, + 307 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 304, + 91, + 545, + 307 + ], + "type": "text", + "content": ", beginning from the timestamp 0 and discarding additional unaligned information at the end. This preprocessing choice allows us to train transformer networks without masking. We provide additional data processing information in the Sup. Mat." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 308, + 545, + 463 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 308, + 545, + 463 + ], + "spans": [ + { + "bbox": [ + 304, + 308, + 545, + 463 + ], + "type": "text", + "content": "Audio preprocessing. We use audio sequences belonging to eight categorical emotion labels (neutral, happy, angry, sad, contempt, surprise, fear, and disgust). Each audio chunk of 10s is converted into a filter bank with 128 melfrequency bins with a 25ms Hamming frame window and 10ms frame shift. We mask each sample with a maximum length of 24 in the frequency domain and a maximum length of 96 in the time domain, employing Park et al. [72]. Following [25, 26], we standardize the filter bank and augment it via noise injection and circular shifting. Before feeding in our speech disentanglement model, each filter bank is split into a sequence of fixed 1209 patches of " + }, + { + "bbox": [ + 304, + 308, + 545, + 463 + ], + "type": "inline_equation", + "content": "16 \\times 16" + }, + { + "bbox": [ + 304, + 308, + 545, + 463 + ], + "type": "text", + "content": " each having 6 units overlap in frequency and time domain." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 464, + 545, + 632 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 464, + 545, + 632 + ], + "spans": [ + { + "bbox": [ + 304, + 464, + 545, + 632 + ], + "type": "text", + "content": "Motion prior. The motion prior is a VAE encoder-decoder with 9 layers and 4 heads, following Chen et al. [15]. The encoder-decoder is a U-Net-like transformer with residual connections. Learnable positional embeddings are injected in each multi-head attention layer. We have a linear projection at the start and the end of our motion prior network. The KL divergence term is weighted with a factor of " + }, + { + "bbox": [ + 304, + 464, + 545, + 632 + ], + "type": "inline_equation", + "content": "1e - 4" + }, + { + "bbox": [ + 304, + 464, + 545, + 632 + ], + "type": "text", + "content": ". Denoiser. The denoiser follows the same network architecture as our prior encoder. The hidden dimension of all transformer layers is 1024. We use 1000 diffusion steps " + }, + { + "bbox": [ + 304, + 464, + 545, + 632 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 304, + 464, + 545, + 632 + ], + "type": "text", + "content": " during training and 50 during inference. Noise betas are in range " + }, + { + "bbox": [ + 304, + 464, + 545, + 632 + ], + "type": "inline_equation", + "content": "[0.00085, 0.012]" + }, + { + "bbox": [ + 304, + 464, + 545, + 632 + ], + "type": "text", + "content": ". We jointly optimize the prior and denoiser networks for 5000 epochs with batch size of 64, learning rate 0.0001, and the AdamW optimizer [59]." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 306, + 645, + 386, + 658 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 645, + 386, + 658 + ], + "spans": [ + { + "bbox": [ + 306, + 645, + 386, + 658 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "content": "Speech disentanglement model. We evaluate the performance of the speech disentanglement model quantitatively using classification accuracy and F1 scores on emotion and style. The accuracy is computed as average scores" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "type": "text", + "content": "1946" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 286, + 132 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 286, + 132 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 286, + 132 + ], + "type": "text", + "content": "for all 8 emotion as well style categories that are part of the test dataset. The emotion and style accuracy is " + }, + { + "bbox": [ + 46, + 72, + 286, + 132 + ], + "type": "inline_equation", + "content": "91.53\\%" + }, + { + "bbox": [ + 46, + 72, + 286, + 132 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 72, + 286, + 132 + ], + "type": "inline_equation", + "content": "96.06\\%" + }, + { + "bbox": [ + 46, + 72, + 286, + 132 + ], + "type": "text", + "content": ", respectively. The emotion F1 score and style F1 scores are 0.914 and 0.960, respectively. See the Sup. Mat. for ablations and a detailed metric analysis." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 133, + 287, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 133, + 287, + 335 + ], + "spans": [ + { + "bbox": [ + 46, + 133, + 287, + 335 + ], + "type": "text", + "content": "Gesture generation model. We evaluate the performance of our gesture generation model quantitatively, qualitatively, and perceptually against following methods: TalkSHOW [101] and the re-implementation of Habibie et al. [31] provided by the TalkSHOW authors in the official TalkSHOW release [100], DiffuseStyleGesture (DSG) [97], MoGlow [34], and CaMN [55]. Additionally, we adapt TalkSHOW to include categorical emotion labels as input along with the existing architecture that only allows one-hot encodings of personal style. We then retrain it on our training data. We refer to it as TalkSHOW-BEAT. There are some concurrent works [5, 7, 56], which introduce methods for gesture generation from speech, however, direct comparison is hindered by the unavailability of released code our task. Refer to the Sup. Mat. for the ablation experiments, the emotion and style editing experiments, and their quantitative evaluation." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 344, + 182, + 357 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 344, + 182, + 357 + ], + "spans": [ + { + "bbox": [ + 47, + 344, + 182, + 357 + ], + "type": "text", + "content": "5.1. Quantitative Evaluation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 362, + 286, + 469 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 362, + 286, + 469 + ], + "spans": [ + { + "bbox": [ + 46, + 362, + 286, + 469 + ], + "type": "text", + "content": "To quantitatively evaluate our method's gesture generations and edited gesture generations, we train a transformer-based encoder architecture (denoted as " + }, + { + "bbox": [ + 46, + 362, + 286, + 469 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 46, + 362, + 286, + 469 + ], + "type": "text", + "content": ") similar to Petrovich et al. [76] in an autoencoder setting, where we append a CLS token at the beginning of the motion sequence. " + }, + { + "bbox": [ + 46, + 362, + 286, + 469 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 46, + 362, + 286, + 469 + ], + "type": "text", + "content": " is trained with a cross-entropy emotion classification objective applied to the output CLS token. We train " + }, + { + "bbox": [ + 46, + 362, + 286, + 469 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 46, + 362, + 286, + 469 + ], + "type": "text", + "content": " on the BEAT training dataset and use its features to compute the following metrics: (1) Fréchet gesture distance (FGD): We" + } + ] + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 50, + 488, + 286, + 602 + ], + "blocks": [ + { + "bbox": [ + 50, + 488, + 286, + 602 + ], + "lines": [ + { + "bbox": [ + 50, + 488, + 286, + 602 + ], + "spans": [ + { + "bbox": [ + 50, + 488, + 286, + 602 + ], + "type": "table", + "html": "
MethodSRGR↑BA↑FGD↓Div→GAa↑
GT—0.83—27.8364.04
Ours0.360.81388.6325.0646.76
Ours-EmoEditb—0.79792.5824.6834.18
TalkSHOW-BEAT0.310.64808.9924.1622.71
TalkSHOW [101]0.300.60762.1523.1929.41
DSG [97]0.230.40763.1019.7722.70
Habibie et al. [31]0.230.39809.1721.3416.67
MoGlow [34]0.210.351097.0319.5016.62
CaMN [55]0.210.391063.8718.9014.17
", + "image_path": "efd0dbe3c6ed65df87582b36cc38c98560ad69b7038fee6f302b3ac43cd5fcab.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 49, + 605, + 167, + 614 + ], + "lines": [ + { + "bbox": [ + 49, + 605, + 167, + 614 + ], + "spans": [ + { + "bbox": [ + 49, + 605, + 167, + 614 + ], + "type": "text", + "content": "a GA is average of all 8 emotions." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_footnote" + }, + { + "bbox": [ + 50, + 615, + 286, + 635 + ], + "lines": [ + { + "bbox": [ + 50, + 615, + 286, + 635 + ], + "spans": [ + { + "bbox": [ + 50, + 615, + 286, + 635 + ], + "type": "text", + "content": "b GA for these are average accuracy for all generations with 7 edited audio sequences." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 644, + 286, + 709 + ], + "lines": [ + { + "bbox": [ + 46, + 644, + 286, + 709 + ], + "spans": [ + { + "bbox": [ + 46, + 644, + 286, + 709 + ], + "type": "text", + "content": "Table 1. Gesture quantitative results. We compare our methods against several SOTA methods using metrics explained in Sec. 5.1. We observe that AMUSE outperforms in all scores compared to baseline methods. Additionally, AMUSE-EmoEdit outperforms in Beat Align, Diversity, and Gesture Emotion Accuracy scores compared to the baseline methods." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 72, + 545, + 347 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 347 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 347 + ], + "type": "text", + "content": "follow [88, 103, 104] to compute the feature distance between generated and ground truth motion features. (2) Gesture diversity (Div): Similarly to Chen et al. [15], we compute variance across generated features. (3) Gesture emotion accuracy (GA): We report top-1 emotion classification accuracy predicted by a classifier trained on the motion " + }, + { + "bbox": [ + 304, + 72, + 545, + 347 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 304, + 72, + 545, + 347 + ], + "type": "text", + "content": "-predicted latents. (4) Beat align (BA): We follow [51, 55], to evaluate the motion-speech correlation in terms of the similarity between the kinematic motion beats and speech audio beats. The kinematic motion beats are directly computed from the generated motion sequences. (5) Semantic-Relevant Gesture Recall (SRGR): We follow Liu et al. [55], to evaluate the semantic relevancy of gestures with GT motion. We use the ground truth semantic scores to compute this metric. The scores are obtained from the BEAT authors, representing a continuous score on a scale 0-1 per gesture style for 4 gesture semantic categories: beat, deictic, iconic, and metaphoric. While comparing with methods that output coarse skeletal data (DSG [97], MoGlow [34], and CaMN [55]), we convert the skeleton motion data into the SMPL-X axis angle representation. For details on the architectures and training of " + }, + { + "bbox": [ + 304, + 72, + 545, + 347 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 304, + 72, + 545, + 347 + ], + "type": "text", + "content": ", and the losses, please refer to the Sup. Mat." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 366, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 366, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 366, + 545, + 713 + ], + "type": "text", + "content": "We prepare the evaluation data by randomly selecting 72 unique motion sequences each of length 10s and comprising 8 emotions across test subjects and compute the aforementioned metrics. We use 9 sequences for each emotion per subject. The results are reported in Tab. 1. All best scores are highlighted in green and second best in blue. AMUSE outperforms the baseline methods in all given metrics. To validate the performance of gesture emotion editing, we also report the same metrics for the emotion editing task (Ours-EmoEdit). During inference, the input style and content latents are extracted from neutral-emotion audio, while the emotion latent comes from a different audio of different emotion. These emotional edits offer numerous possibilities, allowing for transitions from any to any emotion. Tab. 1 shows the average for editing from neutral to other emotions. Since we require the GT gesture semantics score to compute SRGR metric, it is not possible to compute the SRGR for the synthetic edited-emotion gestures as they are not part of the original BEAT dataset. Ours-EmoEdit outperforms the baseline methods in BA, Div, and GA metrics. This demonstrates the capability of our model to maintain highly discriminative cues when switching between different emotions. TalkSHOW-BEAT has the second best score for SRGR whereas TalkSHOW demonstrates second best FGD score. Although, our model and ours-EmoEdit show improvements over the baseline methods, GT motions have higher diversity, Beat alignment score, and are easier to classify than generations of AMUSE, highlighting the challenging nature of the problem." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1947" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 47, + 68, + 542, + 186 + ], + "blocks": [ + { + "bbox": [ + 47, + 68, + 542, + 186 + ], + "lines": [ + { + "bbox": [ + 47, + 68, + 542, + 186 + ], + "spans": [ + { + "bbox": [ + 47, + 68, + 542, + 186 + ], + "type": "image", + "image_path": "fc60d13af2bb6a9cafc31879ce92d3d0be33118b096d7346506181c227888a85.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 189, + 546, + 213 + ], + "lines": [ + { + "bbox": [ + 46, + 189, + 546, + 213 + ], + "spans": [ + { + "bbox": [ + 46, + 189, + 546, + 213 + ], + "type": "text", + "content": "Figure 3. Qualitative comparison across all emotions. We evaluate generation on different test audios. AMUSE exhibits well-synchronized beat gestures and consistently produces gestures that accurately convey the emotional content expressed in the input speech." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 48, + 228, + 289, + 470 + ], + "blocks": [ + { + "bbox": [ + 48, + 228, + 289, + 470 + ], + "lines": [ + { + "bbox": [ + 48, + 228, + 289, + 470 + ], + "spans": [ + { + "bbox": [ + 48, + 228, + 289, + 470 + ], + "type": "image", + "image_path": "10f8b2f70747ff519b934e9e660b3e1def7176179d1c1813988bf47314fe3b3b.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 476, + 287, + 499 + ], + "lines": [ + { + "bbox": [ + 46, + 476, + 287, + 499 + ], + "spans": [ + { + "bbox": [ + 46, + 476, + 287, + 499 + ], + "type": "text", + "content": "Figure 4. Qualitative comparison with baseline methods. The speech segment describes intense angry speech." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 525, + 176, + 537 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 525, + 176, + 537 + ], + "spans": [ + { + "bbox": [ + 47, + 525, + 176, + 537 + ], + "type": "text", + "content": "5.2. Qualitative Evaluation" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 545, + 289, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 545, + 289, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 545, + 289, + 713 + ], + "type": "text", + "content": "Comparison with baseline methods. In Fig. 4, we demonstrate comparison with baseline methods that output a 3D body mesh: Habibie et al. [31], TalkSHOW [101], TalkSHOW-BEAT, and the BEAT ground truth (GT) [55]. We observe that AMUSE generates gestures that are semantically closer to the speech content and produces expressive emotional gestures closer to the perceived emotion. For example, the GT motion exhibits anger when saying \"put them into detention\". AMUSE demonstrates tense posture and aggressive movements comparable with the ground truth data and accurate synchronization with the spoken words. TalkSHOW [101] and Habibie et al. [31] exhibit limited movement and display inferior and static gestures on test audios as seen in the last two rows" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 364, + 229, + 508, + 289 + ], + "blocks": [ + { + "bbox": [ + 364, + 229, + 508, + 289 + ], + "lines": [ + { + "bbox": [ + 364, + 229, + 508, + 289 + ], + "spans": [ + { + "bbox": [ + 364, + 229, + 508, + 289 + ], + "type": "image", + "image_path": "7359634e0462ace41c9f62be4167596525d8f96ebbde7384a8a2de69b67e5475.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 297, + 545, + 319 + ], + "lines": [ + { + "bbox": [ + 305, + 297, + 545, + 319 + ], + "spans": [ + { + "bbox": [ + 305, + 297, + 545, + 319 + ], + "type": "text", + "content": "Figure 5. Qualitative evaluation of diverse generations. Multiple generations overlayed." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 341, + 545, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 341, + 545, + 376 + ], + "spans": [ + { + "bbox": [ + 304, + 341, + 545, + 376 + ], + "type": "text", + "content": "of Fig. 4. TalkSHOW-BEAT slightly outperforms other baseline methods by demonstrating enhanced synchronized gestures, but it still does not perform as well as AMUSE." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 377, + 545, + 401 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 377, + 545, + 401 + ], + "spans": [ + { + "bbox": [ + 305, + 377, + 545, + 401 + ], + "type": "text", + "content": "Diverse emotional gestures. In Fig. 5, our probabilistic model can generate diverse gestures for same input audio." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 402, + 546, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 402, + 546, + 533 + ], + "spans": [ + { + "bbox": [ + 304, + 402, + 546, + 533 + ], + "type": "text", + "content": "Emotional gesture generation. In Fig. 3 AMUSE demonstrates strong correlation with the spoken utterances as well as different emotions. We observe that our model is able to correlate semantic words to associated gestures. For example, gestures demonstrate forceful actions and tense stance with angry audio \"normally get angry\" whereas it generates lowered and calm hand positions for sad audio \"I went to collect\". Similarly, our generations show hands that are closer to body for fearful audio \"in a dangerous situation\" while widely open expressing astonishment for happy and surprised audio \"people pass by\" and \"Wow! You are here\"." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 533, + 545, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 533, + 545, + 594 + ], + "spans": [ + { + "bbox": [ + 304, + 533, + 545, + 594 + ], + "type": "text", + "content": "Emotion editing. We use two audio streams of a female subject for neutral and sad emotion. This experiment edits the subject's gesture style from moderately controlled hand movements to a sad style with lethargic posture conveying a sense of heaviness, as seen in Fig. 7 (top)." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 594, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 546, + 713 + ], + "type": "text", + "content": "Gesture style editing. We use audio streams of two male subjects for the happy (ID - 13) and angry (ID - 2) emotion. With the emotion, style and content latent fusion mechanism from two driving audio streams, AMUSE is able to adapt the male (ID - 13) subject's body gestures from being close to their body to more open with squared tightened shoulders, expressing a shift from happy to angry emotions of a different subject (ID - 2), as shown in Fig. 7 (bottom). Please refer to the supplemental video for qualitative results and comparisons to additional gesture genera" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1948" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 37, + 72, + 545, + 123 + ], + "blocks": [ + { + "bbox": [ + 37, + 72, + 545, + 123 + ], + "lines": [ + { + "bbox": [ + 37, + 72, + 545, + 123 + ], + "spans": [ + { + "bbox": [ + 37, + 72, + 545, + 123 + ], + "type": "image", + "image_path": "dcd48010663c371765d28e509f35149e52387aa74e2c0ede6cd216ce8325d5ce.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 125, + 547, + 148 + ], + "lines": [ + { + "bbox": [ + 46, + 125, + 547, + 148 + ], + "spans": [ + { + "bbox": [ + 46, + 125, + 547, + 148 + ], + "type": "text", + "content": "Figure 6. The perceptual study results for gesture emotion preference (left) and synchronization with speech (right). The number of attentive participants that passed the catch trials is indicated on the right and the reported results only consider these participants." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 54, + 167, + 289, + 290 + ], + "blocks": [ + { + "bbox": [ + 54, + 167, + 289, + 290 + ], + "lines": [ + { + "bbox": [ + 54, + 167, + 289, + 290 + ], + "spans": [ + { + "bbox": [ + 54, + 167, + 289, + 290 + ], + "type": "image", + "image_path": "790684fdaf80e92c2b5d8b5dae1f520bf3f102717544d6328c5f80a8af4ec7ac.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 297, + 288, + 363 + ], + "lines": [ + { + "bbox": [ + 46, + 297, + 288, + 363 + ], + "spans": [ + { + "bbox": [ + 46, + 297, + 288, + 363 + ], + "type": "text", + "content": "Figure 7. Gesture editing. Top: We modify style from being neutral (left) to being sad (right) by combining the emotion latent from sad audio with the content latent from neutral audio. Bottom: We transform the style from Subject 13 being happy (left) to being angry (right) by merging the content latent from happy audio with the style and emotion latents from an angry audio of Subject 2." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 385, + 276, + 397 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 385, + 276, + 397 + ], + "spans": [ + { + "bbox": [ + 47, + 385, + 276, + 397 + ], + "type": "text", + "content": "tion methods [4, 55, 110] trained on coarse skeletal data." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 407, + 149, + 419 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 407, + 149, + 419 + ], + "spans": [ + { + "bbox": [ + 47, + 407, + 149, + 419 + ], + "type": "text", + "content": "5.3. Perceptual Study" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 426, + 287, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 426, + 287, + 616 + ], + "spans": [ + { + "bbox": [ + 46, + 426, + 287, + 616 + ], + "type": "text", + "content": "Design. Our perceptual study is designed as a side-by-side comparison of two gesture videos generated with the same audio as input but by two different methods (AMUSE and another model or GT). The participants are asked to rate their preference of the methods on a five-point Likert scale for \"synchronization with speech\" and \"gesture emotion appropriateness\" given the GT emotion label of the input audio. We recruit 25 participants per method-to-method comparison on Amazon Mechanical Turk. Each participant is shown 24 pairs of randomly selected test set animations, 3 per emotion (neutral, happy, angry, sad, disgust, fear, surprise, and contempt). To allow the participant to get used to the task, we discard the answers of the first three comparisons and repeat these at the end. We incorporate three catch trials and responses from participants that fail on more than one are filtered out, as shown in Fig. 6 (right)." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 618, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 618, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 618, + 288, + 714 + ], + "type": "text", + "content": "Results. The results of the study are shown in Fig. 6. AMUSE outperforms all competing methods by a considerable margin on both tasks, suggesting that AMUSE's generations are more appropriate for both the content of the input speech and its emotion compared to the baselines. However, it must be noted that there is still a significant gap between AMUSE and the GT. Please refer to the Sup. Mat. for details about the perceptual study." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 167, + 463, + 179 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 167, + 463, + 179 + ], + "spans": [ + { + "bbox": [ + 306, + 167, + 463, + 179 + ], + "type": "text", + "content": "5.4. Discussion and Future Work" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 186, + 545, + 258 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 186, + 545, + 258 + ], + "spans": [ + { + "bbox": [ + 304, + 186, + 545, + 258 + ], + "type": "text", + "content": "Upper-body motion. We focus on the smooth coordination between the pelvis and upper body animation for side-by-side comparisons with other methods, as all other methods primarily focus on upper body movements. Future work should include lower-body motion and locomotion as these impact the perceived emotional state of a sequence." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 258, + 545, + 316 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 258, + 545, + 316 + ], + "spans": [ + { + "bbox": [ + 304, + 258, + 545, + 316 + ], + "type": "text", + "content": "Semantics. While the generated gestures, synchronized with the driving speech sequence, do not account for semantics such as deictic and metaphoric gestures, incorporating the text/language modality could help further improve in this direction." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 318, + 545, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 318, + 545, + 415 + ], + "spans": [ + { + "bbox": [ + 304, + 318, + 545, + 415 + ], + "type": "text", + "content": "Facial expressions. While emotional speech-driven face animation methods [18, 74] can be combined with bodies generated from AMUSE, jointly learning to generate emotional 3D bodies from speech is a topic that needs attention. End-to-end training. Joint audio-gesture training may enhance results but requires careful loss term balancing and increased GPU memory. Therefore, we opted for separate training." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 426, + 378, + 437 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 426, + 378, + 437 + ], + "spans": [ + { + "bbox": [ + 306, + 426, + 378, + 437 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 445, + 545, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 445, + 545, + 613 + ], + "spans": [ + { + "bbox": [ + 304, + 445, + 545, + 613 + ], + "type": "text", + "content": "We present AMUSE, a framework to generate emotional body gestures from speech. The emotions and personal styles of the synthesized gestures can be controlled, thanks to the disentanglement of content, emotion, and style directly from the speech. The latent diffusion-based framework can further generate variations of the same gesture with the same emotion. Our quantitative evaluations show that AMUSE achieves state of the art performance on a variety of metrics: diversity, gesture emotion classification accuracy, Frechet gesture distance, beat alignment score, and semantic relevant gesture recall. Finally, our perceptual study demonstrates that AMUSE generates motions that are better synchronized and better match the emotion expressed of the input speech than previous state of the art." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 624, + 545, + 702 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 624, + 545, + 702 + ], + "spans": [ + { + "bbox": [ + 304, + 624, + 545, + 702 + ], + "type": "text", + "content": "Acknowledgments. We thank A Cseke, T McConnell and T Alexiadis for help with the perceptual study, B Pellkofer, J Piles-Contreras, and E Fritzler for IT support, and P Kulits, M Petrovich, S Kokane, and S Zojaji for proofreading and insightful discussions. This project has received funding from the European Union's Horizon 2020 research and innovation programme under the Marie Skłodowska-Curie grant agreement No. 860768." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 703, + 542, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 703, + 542, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 703, + 542, + 713 + ], + "type": "text", + "content": "Disclosure. https://download.is.tue.mpg.de/amuse/Disclosure.txt" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1949" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 53, + 91, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 58, + 91, + 287, + 133 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 91, + 287, + 133 + ], + "spans": [ + { + "bbox": [ + 58, + 91, + 287, + 133 + ], + "type": "text", + "content": "[1] Kfir Aberman, Yijia Weng, Dani Lischinski, Daniel Cohen-Or, and Baoquan Chen. Unpaired motion style transfer from video to animation. Transactions on Graphics (TOG), 2020. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 137, + 286, + 169 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 137, + 286, + 169 + ], + "spans": [ + { + "bbox": [ + 58, + 137, + 286, + 169 + ], + "type": "text", + "content": "[2] Chaitanya Ahuja and Louis-Philippe Morency. Language2Pose: Natural language grounded pose forecasting. International Conference on 3D Vision (3DV), 2019. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 58, + 171, + 286, + 214 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 171, + 286, + 214 + ], + "spans": [ + { + "bbox": [ + 58, + 171, + 286, + 214 + ], + "type": "text", + "content": "[3] Chaitanya Ahuja, Dong Won Lee, and Louis-Philippe Morency. Low-resource adaptation for personalized cospeech gesture generation. In Computer Vision and Pattern Recognition (CVPR), pages 20566-20576, 2022. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 58, + 217, + 286, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 217, + 286, + 270 + ], + "spans": [ + { + "bbox": [ + 58, + 217, + 286, + 270 + ], + "type": "text", + "content": "[4] Simon Alexanderson, Gustav Eje Henter, Taras Kucherenko, and Jonas Beskow. Style-controllable speech-driven gesture synthesis using normalising flows. Computer Graphics Forum (CGF), 39(2):487-496, 2020. 3, 8" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 58, + 273, + 286, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 273, + 286, + 316 + ], + "spans": [ + { + "bbox": [ + 58, + 273, + 286, + 316 + ], + "type": "text", + "content": "[5] Simon Alexanderson, Rajmund Nagy, Jonas Beskow, and Gustav Eje Henter. Listen, Denoise, Action! Audio-driven motion synthesis with diffusion models. Transactions on Graphics (TOG), 42(4):1-20, 2023. 1, 3, 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 58, + 318, + 286, + 350 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 318, + 286, + 350 + ], + "spans": [ + { + "bbox": [ + 58, + 318, + 286, + 350 + ], + "type": "text", + "content": "[6] Tenglong Ao, Qingzhe Gao, Yuke Lou, Baoquan Chen, and Libin Liu. Rhythmic gesticulator. Transactions on Graphics (TOG), 41(6):1-19, 2022. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 58, + 353, + 286, + 384 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 353, + 286, + 384 + ], + "spans": [ + { + "bbox": [ + 58, + 353, + 286, + 384 + ], + "type": "text", + "content": "[7] Tenglong Ao, Zeyi Zhang, and Libin Liu. GestureDiffu-CLIP: Gesture diffusion model with CLIP latents. Transactions on Graphics (TOG), 2023. 1, 3, 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 58, + 387, + 286, + 429 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 387, + 286, + 429 + ], + "spans": [ + { + "bbox": [ + 58, + 387, + 286, + 429 + ], + "type": "text", + "content": "[8] Nikos Athanasiou, Mathis Petrovich, Michael J. Black, and Gül Varol. TEACH: Temporal Action Compositions for 3D Humans. In International Conference on 3D Vision (3DV), 2022. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 58, + 432, + 286, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 432, + 286, + 475 + ], + "spans": [ + { + "bbox": [ + 58, + 432, + 286, + 475 + ], + "type": "text", + "content": "[9] Nikos Athanasiou, Mathis Petrovich, Michael J. Black, and Gül Varol. SINC: Spatial composition of 3D human motions for simultaneous action generation. International Conference on Computer Vision (ICCV), 2023. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 53, + 478, + 286, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 478, + 286, + 520 + ], + "spans": [ + { + "bbox": [ + 53, + 478, + 286, + 520 + ], + "type": "text", + "content": "[10] Emad Barsoum, John Kender, and Zicheng Liu. HPGAN: probabilistic 3d human motion prediction via GAN. In Computer Vision and Pattern Recognition Workshops (CVPRW), 2017. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 53, + 523, + 286, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 523, + 286, + 586 + ], + "spans": [ + { + "bbox": [ + 53, + 523, + 286, + 586 + ], + "type": "text", + "content": "[11] Uttaran Bhattacharya, Elizabeth Childs, Nicholas Rewkowski, and Dinesh Manocha. Speech2affectivegestures: Synthesizing co-speech gestures with generative adversarial affective expression learning. In ACM International Conference on Multimedia (MM), 2021. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 53, + 590, + 286, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 590, + 286, + 643 + ], + "spans": [ + { + "bbox": [ + 53, + 590, + 286, + 643 + ], + "type": "text", + "content": "[12] Uttaran Bhattacharya, Nicholas Rewkowski, Abhishek Banerjee, Pooja Guhan, Aniket Bera, and Dinesh Manocha. Text2gestures: A transformer-based network for generating emotive body gestures for virtual agents. In IEEE Virtual Reality and 3D User Interfaces, VR, 2021. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 53, + 647, + 286, + 700 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 647, + 286, + 700 + ], + "spans": [ + { + "bbox": [ + 53, + 647, + 286, + 700 + ], + "type": "text", + "content": "[13] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In European Conference on Computer Vision (ECCV), pages 213-229. Springer, 2020. 4" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 53, + 703, + 286, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 703, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 53, + 703, + 286, + 713 + ], + "type": "text", + "content": "[14] Justine Cassell, Hannes Hogni Vilhjalmsson, and Timo" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 312, + 74, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 31, + "blocks": [ + { + "bbox": [ + 331, + 74, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 331, + 74, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 331, + 74, + 545, + 95 + ], + "type": "text", + "content": "thy Bickmore. BEAT: The Behavior Expression Animation Toolkit, pages 163-185. Springer, 2004. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 312, + 96, + 545, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 96, + 545, + 139 + ], + "spans": [ + { + "bbox": [ + 312, + 96, + 545, + 139 + ], + "type": "text", + "content": "[15] Xin Chen, Biao Jiang, Wen Liu, Zilong Huang, Bin Fu, Tao Chen, and Gang Yu. Executing your commands via motion diffusion in latent space. In Computer Vision and Pattern Recognition (CVPR), pages 18000-18010, 2023. 4, 5, 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 312, + 140, + 545, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 140, + 545, + 182 + ], + "spans": [ + { + "bbox": [ + 312, + 140, + 545, + 182 + ], + "type": "text", + "content": "[16] Enric Corona, Albert Pumarola, G. Alenyà, and F. Moreno-Noguer. Context-aware human motion prediction. In Computer Vision and Pattern Recognition (CVPR), pages 6990–6999, 2020. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 312, + 184, + 545, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 184, + 545, + 227 + ], + "spans": [ + { + "bbox": [ + 312, + 184, + 545, + 227 + ], + "type": "text", + "content": "[17] Rishabh Dabral, Muhammad Hamza Mughal, Vladislav Golyanik, and Christian Theobalt. Mofusion: A framework for denoising-diffusion-based motion synthesis. In Computer Vision and Pattern Recognition (CVPR), 2023. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 312, + 228, + 545, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 228, + 545, + 293 + ], + "spans": [ + { + "bbox": [ + 312, + 228, + 545, + 293 + ], + "type": "text", + "content": "[18] Radek Daneček, Kiran Chhatre, Shashank Tripathi, Yandong Wen, Michael Black, and Timo Bolkart. Emotional speech-driven animation with content-emotion disentangle-ment. In International Conference on Computer Graphics and Interactive Techniques in Asia (SIGGRAPH ASIA), 2023. 2, 3, 8" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 312, + 295, + 545, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 295, + 545, + 315 + ], + "spans": [ + { + "bbox": [ + 312, + 295, + 545, + 315 + ], + "type": "text", + "content": "[19] Paul Ekman. Facial expression and emotion. American Psychologist, 48(4):384-392, 1993. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 312, + 317, + 545, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 317, + 545, + 370 + ], + "spans": [ + { + "bbox": [ + 312, + 317, + 545, + 370 + ], + "type": "text", + "content": "[20] Esam Ghaleb, Ilya Burenko, Marlou Rasenberg, Wim Pouw, Peter Uhrig, Judith Holler, Ivan Toni, Asli Özyurek, and Raquel Fernández. Co-speech gesture detection through multi-phase sequence labeling. In Winter Conference on Applications of Computer Vision (WACV), 2024. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 312, + 372, + 545, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 372, + 545, + 415 + ], + "spans": [ + { + "bbox": [ + 312, + 372, + 545, + 415 + ], + "type": "text", + "content": "[21] Saeed Ghorbani, Ylva Ferstl, Daniel Holden, Nikolaus F. Troje, and Marc-Andre Carbonneau. Zeroeggs: Zero-shot example-based gesture generation from speech. Computer Graphics Forum (CGF), 42(1):206-216, 2023. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 312, + 416, + 545, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 416, + 545, + 459 + ], + "spans": [ + { + "bbox": [ + 312, + 416, + 545, + 459 + ], + "type": "text", + "content": "[22] Anindita Ghosh, Noshaba Cheema, Cennet Oguz, Christian Theobalt, and Philipp Slusallek. Synthesis of compositional animations from textual descriptions. In International Conference on Computer Vision (ICCV), 2021. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 312, + 460, + 545, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 460, + 545, + 503 + ], + "spans": [ + { + "bbox": [ + 312, + 460, + 545, + 503 + ], + "type": "text", + "content": "[23] Shiry Ginosar, Amir Bar, Gefen Kohavi, Caroline Chan, Andrew Owens, and Jitendra Malik. Learning individual styles of conversational gesture. In Computer Vision and Pattern Recognition (CVPR), pages 3497-3506, 2019. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 312, + 505, + 545, + 534 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 505, + 545, + 534 + ], + "spans": [ + { + "bbox": [ + 312, + 505, + 545, + 534 + ], + "type": "text", + "content": "[24] Ross B. Girshick. Fast R-CNN. In International Conference on Computer Vision (ICCV), pages 1440-1448, 2015. 5" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 312, + 537, + 545, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 537, + 545, + 568 + ], + "spans": [ + { + "bbox": [ + 312, + 537, + 545, + 568 + ], + "type": "text", + "content": "[25] Yuan Gong, Yu-An Chung, and James R. Glass. AST: audio spectrogram transformer. In *Interspeech* 2021, pages 571-575, 2021. 4, 5" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 312, + 571, + 545, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 571, + 545, + 613 + ], + "spans": [ + { + "bbox": [ + 312, + 571, + 545, + 613 + ], + "type": "text", + "content": "[26] Yuan Gong, Cheng-I Lai, Yu-An Chung, and James R. Glass. SSAST: self-supervised audio spectrogram transformer. In AAAI Conference on Artificial Intelligence, pages 10699-10709, 2022. 4, 5" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 312, + 616, + 545, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 616, + 545, + 668 + ], + "spans": [ + { + "bbox": [ + 312, + 616, + 545, + 668 + ], + "type": "text", + "content": "[27] Chuan Guo, Xinxin Zuo, Sen Wang, Shihao Zou, Qingyao Sun, Annan Deng, Minglun Gong, and Li Cheng. Action2Motion: Conditioned generation of 3d human motions. In ACM International Conference on Multimedia (MM), 2020. 2" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 312, + 670, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 670, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 312, + 670, + 545, + 712 + ], + "type": "text", + "content": "[28] Chuan Guo, Shihao Zou, Xinxin Zuo, Sen Wang, Wei Ji, Xingyu Li, and Li Cheng. Generating diverse and natural 3D human motions from text. In Computer Vision and Pattern Recognition (CVPR), 2022. 2" + } + ] + } + ], + "index": 30 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "type": "text", + "content": "1950" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 53, + 73, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 53, + 73, + 287, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 73, + 287, + 106 + ], + "spans": [ + { + "bbox": [ + 53, + 73, + 287, + 106 + ], + "type": "text", + "content": "[29] Nejla Gürefe. The role of gestures in mathematical discourse of hard-hearing students: Prism example. Acta Destructica Napocensia, 11:125-140, 2018. 1" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 53, + 107, + 288, + 152 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 107, + 288, + 152 + ], + "spans": [ + { + "bbox": [ + 53, + 107, + 288, + 152 + ], + "type": "text", + "content": "[30] I. Habibie, Daniel Holden, Jonathan Schwarz, J. Yearsley, and T. Komura. A recurrent variational autoencoder for human motion synthesis. In *British Machine Vision Conference (BMVC)*, 2017. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 153, + 288, + 218 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 153, + 288, + 218 + ], + "spans": [ + { + "bbox": [ + 53, + 153, + 288, + 218 + ], + "type": "text", + "content": "[31] Ikhsanul Habibie, Weipeng Xu, Dushyant Mehta, Lingjie Liu, Hans-Peter Seidel, Gerard Pons-Moll, Mohamed Elgharib, and Christian Theobalt. Learning speech-driven 3d conversational gestures from video. In ACM International Conference on Intelligent Virtual Agents (IVA), 2021. 1, 6, 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 220, + 288, + 287 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 220, + 288, + 287 + ], + "spans": [ + { + "bbox": [ + 53, + 220, + 288, + 287 + ], + "type": "text", + "content": "[32] Ikhsanul Habibie, Mohamed A. Elgharib, Kripasindhu Sarkar, Ahsan Abdullah, Simbarashe Linval Nyatsanga, Michael Neff, and Christian Theobalt. A motion matching-based framework for controllable gesture synthesis from speech. International Conference on Computer Graphics and Interactive Techniques (SIGGRAPH), 2022. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 289, + 287, + 322 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 289, + 287, + 322 + ], + "spans": [ + { + "bbox": [ + 53, + 289, + 287, + 322 + ], + "type": "text", + "content": "[33] Félix G. Harvey, Mike Yurick, Derek Nowrouzezahrai, and Christopher Joseph Pal. Robust motion in-between. Transactions on Graphics (TOG), 2020. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 323, + 287, + 367 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 323, + 287, + 367 + ], + "spans": [ + { + "bbox": [ + 53, + 323, + 287, + 367 + ], + "type": "text", + "content": "[34] Gustav Eje Henter, Simon Alexanderson, and Jonas Beskow. MoGlow: Probabilistic and controllable motion synthesis using normalising flows. Transactions on Graphics (TOG), 39(4):236:1-236:14, 2020. 3, 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 369, + 287, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 369, + 287, + 402 + ], + "spans": [ + { + "bbox": [ + 53, + 369, + 287, + 402 + ], + "type": "text", + "content": "[35] Sandra Herbert. Gesture types for functions. Mathematics Education Research Group of Australasia, pages 322-329, 2012. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 404, + 287, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 404, + 287, + 437 + ], + "spans": [ + { + "bbox": [ + 53, + 404, + 287, + 437 + ], + "type": "text", + "content": "[36] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In Conference on Neural Information Processing Systems (NeurIPS), 2020. 4, 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 439, + 287, + 482 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 439, + 287, + 482 + ], + "spans": [ + { + "bbox": [ + 53, + 439, + 287, + 482 + ], + "type": "text", + "content": "[37] Xun Huang and Serge J. Belongie. Arbitrary style transfer in real-time with adaptive instance normalization. In International Conference on Computer Vision (ICCV), pages 1510-1519, 2017. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 484, + 287, + 529 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 484, + 287, + 529 + ], + "spans": [ + { + "bbox": [ + 53, + 484, + 287, + 529 + ], + "type": "text", + "content": "[38] Tero Karras, Timo Aila, Samuli Laine, Antti Herva, and Jaakko Lehtinen. Audio-driven facial animation by joint end-to-end learning of pose and emotion. Transactions on Graphics (TOG), 2017. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 53, + 530, + 287, + 552 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 530, + 287, + 552 + ], + "spans": [ + { + "bbox": [ + 53, + 530, + 287, + 552 + ], + "type": "text", + "content": "[39] Adam Kendon. Gesture: Visible action as utterance. Cambridge University Press, 2004. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 53, + 554, + 287, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 554, + 287, + 609 + ], + "spans": [ + { + "bbox": [ + 53, + 554, + 287, + 609 + ], + "type": "text", + "content": "[40] Gwantae Kim, Seonghyeok Noh, Insung Ham, and Hanseok Ko. MPE4G: Multimodal pretrained encoder for co-speech gesture generation. In International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 1-5, 2023. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 53, + 611, + 287, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 611, + 287, + 643 + ], + "spans": [ + { + "bbox": [ + 53, + 611, + 287, + 643 + ], + "type": "text", + "content": "[41] Jihoon Kim, Taehyun Byun, Seungyoung Shin, Jung-dam Won, and Sungjoon Choi. Conditional motion in-between. Pattern Recognition, 2022. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 53, + 646, + 287, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 646, + 287, + 667 + ], + "spans": [ + { + "bbox": [ + 53, + 646, + 287, + 667 + ], + "type": "text", + "content": "[42] Chris L Kleinke. Gaze and eye contact: A research review. Psychological Bulletin, 100(1):78-100, 1986. 1" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 53, + 670, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 670, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 53, + 670, + 287, + 713 + ], + "type": "text", + "content": "[43] Zhifeng Kong, Wei Ping, Jiaji Huang, Kexin Zhao, and Bryan Catanzaro. DiffWave: A versatile diffusion model for audio synthesis. In International Conference on Learning Representations (ICLR), 2021. 3" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 312, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 312, + 73, + 545, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 73, + 545, + 139 + ], + "spans": [ + { + "bbox": [ + 312, + 73, + 545, + 139 + ], + "type": "text", + "content": "[44] Stefan Kopp, Brigitte Krenn, Stacy Marsella, Andrew N. Marshall, Catherine Pelachaud, Hannes Pirker, Kristinn R. Thorisson, and Hannes Vilhjalmsson. Towards a common framework for multimodal generation: The behavior markup language. In Intelligent Virtual Agents (IVA), pages 205-217. Springer, 2006. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 312, + 140, + 545, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 140, + 545, + 184 + ], + "spans": [ + { + "bbox": [ + 312, + 140, + 545, + 184 + ], + "type": "text", + "content": "[45] Quoc Anh Le, Souheil Hanoune, and Catherine Pelachaud. Design and implementation of an expressive gesture model for a humanoid robot. In IEEE-RAS International Conference on Humanoid Robots, pages 134-140, 2011. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 312, + 186, + 545, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 186, + 545, + 251 + ], + "spans": [ + { + "bbox": [ + 312, + 186, + 545, + 251 + ], + "type": "text", + "content": "[46] Gilwoo Lee, Zhiwei Deng, Shugao Ma, Takaaki Shiratori, Siddhartha S. Srinivasa, and Yaser Sheikh. Talking With Hands 16.2M: A large-scale dataset of synchronized body-finger motion and audio for conversational motion analysis and synthesis. In International Conference on Computer Vision (ICCV), pages 763-772, 2019. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 312, + 252, + 545, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 252, + 545, + 285 + ], + "spans": [ + { + "bbox": [ + 312, + 252, + 545, + 285 + ], + "type": "text", + "content": "[47] Robert W Levenson. Blood, sweat, and fears: The autonomic architecture of emotion. Annals of the New York Academy of Sciences, 1000:348-366, 2003. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 312, + 286, + 545, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 286, + 545, + 341 + ], + "spans": [ + { + "bbox": [ + 312, + 286, + 545, + 341 + ], + "type": "text", + "content": "[48] Jing Li, Di Kang, Wenjie Pei, Xuefei Zhe, Ying Zhang, Zhenyu He, and Linchao Bao. Audio2gestures: Generating diverse gestures from speech audio with conditional variational autoencoders. In Computer Vision and Pattern Recognition (CVPR), pages 11293-11302, 2021. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 312, + 342, + 545, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 342, + 545, + 386 + ], + "spans": [ + { + "bbox": [ + 312, + 342, + 545, + 386 + ], + "type": "text", + "content": "[49] Peizhuo Li, Kfir Aberman, Zihan Zhang, Rana Hanocka, and Olga Sorkine-Hornung. GANimator: Neural motion synthesis from a single sequence. Transactions on Graphics (TOG), 2022. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 312, + 388, + 545, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 388, + 545, + 431 + ], + "spans": [ + { + "bbox": [ + 312, + 388, + 545, + 431 + ], + "type": "text", + "content": "[50] Ruilong Li, Shan Yang, D. A. Ross, and Angjoo Kanazawa. AI choreographer: Music conditioned 3D dance generation with AIST++. In International Conference on Computer Vision (ICCV), 2021. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 312, + 433, + 545, + 477 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 433, + 545, + 477 + ], + "spans": [ + { + "bbox": [ + 312, + 433, + 545, + 477 + ], + "type": "text", + "content": "[51] Ruilong Li, Shan Yang, David A. Ross, and Angjoo Kanazawa. AI choreographer: Music conditioned 3d dance generation with AIST++. In International Conference on Computer Vision (ICCV), pages 13381-13392, 2021. 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 312, + 478, + 545, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 478, + 545, + 521 + ], + "spans": [ + { + "bbox": [ + 312, + 478, + 545, + 521 + ], + "type": "text", + "content": "[52] William Li and Philippe Pasquier. Automatic affect classification of human motion capture sequences in the valence-arousal model. In International Symposium on Movement and Computing, 2016. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 312, + 523, + 545, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 523, + 545, + 567 + ], + "spans": [ + { + "bbox": [ + 312, + 523, + 545, + 567 + ], + "type": "text", + "content": "[53] Yuanzhi Liang, Qianyu Feng, Linchao Zhu, Li Hu, Pan Pan, and Yi Yang. Seeg: Semantic energized co-speech gesture generation. In Computer Vision and Pattern Recognition (CVPR), 2022. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 312, + 568, + 545, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 568, + 545, + 623 + ], + "spans": [ + { + "bbox": [ + 312, + 568, + 545, + 623 + ], + "type": "text", + "content": "[54] Haiyang Liu, Naoya Iwamoto, Zihao Zhu, Zhengqing Li, You Zhou, Elif Bozkurt, and Bo Zheng. DisCo: Disentangled implicit content and rhythm learning for diverse co-speech gestures synthesis. In ACM International Conference on Multimedia (MM), pages 3764-3773, 2022. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 312, + 624, + 545, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 624, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 312, + 624, + 545, + 689 + ], + "type": "text", + "content": "[55] Haiyang Liu, Zihao Zhu, Naoya Iwamoto, Yichen Peng, Zhengqing Li, You Zhou, Elif Bozkurt, and Bo Zheng. BEAT: A large-scale semantic and emotional multi-modal dataset for conversational gestures synthesis. European Conference on Computer Vision (ECCV), 2022. 2, 3, 5, 6, 7, 8" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 312, + 691, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 691, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 312, + 691, + 545, + 713 + ], + "type": "text", + "content": "[56] Haiyang Liu, Zihao Zhu, Giorgio Becherini, Yichen Peng, Mingyang Su, You Zhou, Xuefei Zhe, Naoya Iwamoto, Bo" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 314, + 757 + ], + "type": "text", + "content": "1951" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 53, + 72, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 72, + 72, + 287, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 72, + 72, + 287, + 117 + ], + "spans": [ + { + "bbox": [ + 72, + 72, + 287, + 117 + ], + "type": "text", + "content": "Zheng, and Michael J. Black. EMAGE: Towards unified holistic co-speech gesture generation via expressive masked audio gesture modeling. In Computer Vision and Pattern Recognition (CVPR), 2024. 1, 2, 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 53, + 119, + 288, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 119, + 288, + 173 + ], + "spans": [ + { + "bbox": [ + 53, + 119, + 288, + 173 + ], + "type": "text", + "content": "[57] Yuejiang Liu, Riccardo Cadei, Jonas Schweizer, Sherwin Bahmani, and Alexandre Alahi. Towards robust and adaptive motion forecasting: A causal representation perspective. In Computer Vision and Pattern Recognition (CVPR), 2022. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 175, + 287, + 218 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 175, + 287, + 218 + ], + "spans": [ + { + "bbox": [ + 53, + 175, + 287, + 218 + ], + "type": "text", + "content": "[58] Matthew M. Loper, Naureen Mahmood, and Michael J. Black. MoSh: Motion and shape capture from sparse markers. Transactions on Graphics (TOG), 33(6):220:1-220:13, 2014. 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 221, + 287, + 255 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 221, + 287, + 255 + ], + "spans": [ + { + "bbox": [ + 53, + 221, + 287, + 255 + ], + "type": "text", + "content": "[59] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In International Conference on Learning Representations (ICLR), 2019. 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 256, + 287, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 256, + 287, + 300 + ], + "spans": [ + { + "bbox": [ + 53, + 256, + 287, + 300 + ], + "type": "text", + "content": "[60] Shuhong Lu, Youngwoo Yoon, and Andrew W. Feng. Cospeech gesture synthesis using discrete gesture token learning. International Conference on Intelligent Robots and Systems (IROS), pages 9808-9815, 2023. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 302, + 287, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 302, + 287, + 357 + ], + "spans": [ + { + "bbox": [ + 53, + 302, + 287, + 357 + ], + "type": "text", + "content": "[61] Birgit Lugarin, Catherine Pelachaud, and David Traum, editors. The Handbook on Socially Interactive Agents: 20 Years of Research on Embodied Conversational Agents, Intelligent Virtual Agents, and Social Robotics Volume 1: Methods, Behavior, Cognition. 2021. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 358, + 287, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 358, + 287, + 403 + ], + "spans": [ + { + "bbox": [ + 53, + 358, + 287, + 403 + ], + "type": "text", + "content": "[62] Naureen Mahmood, Nima Ghorbani, Nikolaus F. Troje, Gerard Pons-Moll, and Michael J. Black. Amass: Archive of motion capture as surface shapes. In International Conference on Computer Vision (ICCV), 2019. 2, 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 404, + 287, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 404, + 287, + 459 + ], + "spans": [ + { + "bbox": [ + 53, + 404, + 287, + 459 + ], + "type": "text", + "content": "[63] Stacy Marsella, Yuyu Xu, Margaux Lhommet, Andrew Feng, Stefan Scherer, and Ari Shapiro. Virtual character performance from speech. In SIGGRAPH/Eurographics Symposium on Computer Animation (SCA), page 25-35, 2013. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 461, + 287, + 505 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 461, + 287, + 505 + ], + "spans": [ + { + "bbox": [ + 53, + 461, + 287, + 505 + ], + "type": "text", + "content": "[64] Julieta Martinez, Michael J. Black, and Javier Romero. On human motion prediction using recurrent neural networks. In Computer Vision and Pattern Recognition (CVPR), 2017. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 507, + 287, + 540 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 507, + 287, + 540 + ], + "spans": [ + { + "bbox": [ + 53, + 507, + 287, + 540 + ], + "type": "text", + "content": "[65] David Mcneill. Hand and mind: What gestures reveal about thought. Bibliovault OAI Repository, the University of Chicago Press, 27, 1994. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 53, + 543, + 287, + 564 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 543, + 287, + 564 + ], + "spans": [ + { + "bbox": [ + 53, + 543, + 287, + 564 + ], + "type": "text", + "content": "[66] Albert Mehrabian. Nonverbal communication. Aldine-Atherton Chicago, 1972. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 53, + 566, + 287, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 566, + 287, + 621 + ], + "spans": [ + { + "bbox": [ + 53, + 566, + 287, + 621 + ], + "type": "text", + "content": "[67] Shivam Mehta, Siyang Wang, Simon Alexanderson, Jonas Beskow, Éva Székely, and Gustav Eje Henter. Diff-TTSG: Denoising probabilistic integrated speech and gesture synthesis. In Proc. ISCA Speech Synthesis Workshop (SSW), pages 150–156, 2023. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 53, + 623, + 287, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 623, + 287, + 667 + ], + "spans": [ + { + "bbox": [ + 53, + 623, + 287, + 667 + ], + "type": "text", + "content": "[68] Davide Moltisanti, Jinyi Wu, Bo Dai, and Chen Change Loy. BRACE: the breakdancing competition dataset for dance motion synthesis. In European Conference on Computer Vision (ECCV), pages 329-344, 2022. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 53, + 669, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 669, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 53, + 669, + 287, + 712 + ], + "type": "text", + "content": "[69] Alexander Quinn Nichol and Prafulla Dhariwal. Improved denoising diffusion probabilistic models. In International Conference on Machine Learning (ICML), pages 8162-8171, 2021. 4" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 312, + 72, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 312, + 72, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 72, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 312, + 72, + 545, + 106 + ], + "type": "text", + "content": "[70] Dirk Ormoneit, Michael J. Black, T. Hastie, and H. Kjellström. Representing cyclic human motion using functional analysis. Image and Vision Computing (IVC), 2005. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 312, + 107, + 545, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 107, + 545, + 160 + ], + "spans": [ + { + "bbox": [ + 312, + 107, + 545, + 160 + ], + "type": "text", + "content": "[71] Kunkun Pang, Dafei Qin, Yingruo Fan, Julian Habekost, Takaaki Shiratori, Junichi Yamagishi, and Taku Komura. BodyFormer: Semantics-guided 3d body gesture synthesis with transformer. Transactions on Graphics (TOG), 42(4), 2023. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 312, + 162, + 545, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 162, + 545, + 215 + ], + "spans": [ + { + "bbox": [ + 312, + 162, + 545, + 215 + ], + "type": "text", + "content": "[72] Daniel S. Park, William Chan, Yu Zhang, Chung-Cheng Chiu, Barret Zoph, Ekin D. Cubuk, and Quoc V. Le. SpecAugment: A simple data augmentation method for automatic speech recognition. In Interspeech 2019. ISCA, 2019. 5" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 312, + 217, + 545, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 217, + 545, + 271 + ], + "spans": [ + { + "bbox": [ + 312, + 217, + 545, + 271 + ], + "type": "text", + "content": "[73] Georgios Pavlakos, Vasileios Choutas, Nima Ghorbani, Timo Bolkart, Ahmed A. A. Osman, Dimitrios Tzionas, and Michael J. Black. Expressive body capture: 3D hands, face, and body from a single image. In Computer Vision and Pattern Recognition (CVPR), pages 10975-10985, 2019. 2, 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 312, + 272, + 545, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 272, + 545, + 327 + ], + "spans": [ + { + "bbox": [ + 312, + 272, + 545, + 327 + ], + "type": "text", + "content": "[74] Ziqiao Peng, Haoyu Wu, Zhenbo Song, Hao Xu, Xiangyu Zhu, Jun He, Hongyan Liu, and Zhaoxin Fan. EmoTalk: Speech-driven emotional disentanglement for 3d face animation. In International Conference on Computer Vision (ICCV), pages 20687-20697, 2023. 2, 3, 8" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 312, + 327, + 545, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 327, + 545, + 370 + ], + "spans": [ + { + "bbox": [ + 312, + 327, + 545, + 370 + ], + "type": "text", + "content": "[75] Mathis Petrovich, Michael J. Black, and Gül Varol. Action-conditioned 3D human motion synthesis with transformer VAE. In International Conference on Computer Vision (ICCV), 2021. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 312, + 371, + 545, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 371, + 545, + 415 + ], + "spans": [ + { + "bbox": [ + 312, + 371, + 545, + 415 + ], + "type": "text", + "content": "[76] Mathis Petrovich, Michael J. Black, and Gül Varol. Action-conditioned 3D human motion synthesis with transformer VAE. In International Conference on Computer Vision (ICCV), 2021. 4, 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 312, + 415, + 545, + 458 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 415, + 545, + 458 + ], + "spans": [ + { + "bbox": [ + 312, + 415, + 545, + 458 + ], + "type": "text", + "content": "[77] Mathis Petrovich, Michael J. Black, and Gül Varol. TEMOS: Generating diverse human motions from textual descriptions. In European Conference on Computer Vision (ECCV), 2022. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 312, + 460, + 545, + 492 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 460, + 545, + 492 + ], + "spans": [ + { + "bbox": [ + 312, + 460, + 545, + 492 + ], + "type": "text", + "content": "[78] Paolo Petta, Catherine Pelachaud, and Roddy Cowie. Emotion-Oriented Systems: The Humaine Handbook. Springer Publishing Company, Incorporated, 2011. 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 312, + 493, + 545, + 513 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 493, + 545, + 513 + ], + "spans": [ + { + "bbox": [ + 312, + 493, + 545, + 513 + ], + "type": "text", + "content": "[79] Rosalind W. Picard. Affective Computing. MIT Press, 1997. 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 312, + 514, + 545, + 548 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 514, + 545, + 548 + ], + "spans": [ + { + "bbox": [ + 312, + 514, + 545, + 548 + ], + "type": "text", + "content": "[80] I. Poggi, C. Pelachaud, F. de Rosis, V. Carofiglio, and B. De Carolis. Greta. A Believable Embodied Conversational Agent, pages 3-25. Springer, 2005. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 312, + 548, + 545, + 581 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 548, + 545, + 581 + ], + "spans": [ + { + "bbox": [ + 312, + 548, + 545, + 581 + ], + "type": "text", + "content": "[81] Xingqun Qi, Chen Liu, Lincheng Li, Jie Hou, Haoran Xin, and Xin Yu. EmotionGesture: Audio-driven diverse emotional co-speech 3d gesture generation, 2023. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 312, + 582, + 545, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 582, + 545, + 635 + ], + "spans": [ + { + "bbox": [ + 312, + 582, + 545, + 635 + ], + "type": "text", + "content": "[82] Xingqun Qi, Chen Liu, Muyi Sun, Lincheng Li, Changjie Fan, and Xin Yu. Diverse 3D hand gesture prediction from body dynamics by bilateral hand disentanglement. In Computer Vision and Pattern Recognition (CVPR), pages 4616-4626, 2023. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 312, + 636, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 636, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 312, + 636, + 545, + 712 + ], + "type": "text", + "content": "[83] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning (ICML), pages 8748-8763. PMLR, 2021. 3" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1952" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 53, + 72, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 53, + 72, + 287, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 72, + 287, + 127 + ], + "spans": [ + { + "bbox": [ + 53, + 72, + 287, + 127 + ], + "type": "text", + "content": "[84] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Computer Vision and Pattern Recognition (CVPR), pages 10674-10685, 2022. 4, 5" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 53, + 129, + 288, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 129, + 288, + 173 + ], + "spans": [ + { + "bbox": [ + 53, + 129, + 288, + 173 + ], + "type": "text", + "content": "[85] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-Net: Convolutional networks for biomedical image segmentation. In Medical Image Computing and Computer-Assisted Intervention - MICCAI, pages 234-241, 2015. 4" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 175, + 288, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 175, + 288, + 207 + ], + "spans": [ + { + "bbox": [ + 53, + 175, + 288, + 207 + ], + "type": "text", + "content": "[86] Tim Salzmann, Marco Pavone, and Markus Ryll. Motron: Multimodal Probabilistic Human Motion Forecasting. In Computer Vision and Pattern Recognition (CVPR), 2022. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 209, + 288, + 241 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 209, + 288, + 241 + ], + "spans": [ + { + "bbox": [ + 53, + 209, + 288, + 241 + ], + "type": "text", + "content": "[87] Klaus R Scherer. Vocal communication of emotion: A review of research paradigms. Speech Communication, 40 (1-2):227-256, 2003. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 243, + 288, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 243, + 288, + 274 + ], + "spans": [ + { + "bbox": [ + 53, + 243, + 288, + 274 + ], + "type": "text", + "content": "[88] Maximilian Seitzer. pytorch-fid: FID Score for PyTorch. https://github.com/mseitzer/pytorch-fid, 2020. Version 0.3.0.6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 276, + 288, + 309 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 276, + 288, + 309 + ], + "spans": [ + { + "bbox": [ + 53, + 276, + 288, + 309 + ], + "type": "text", + "content": "[89] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In International Conference on Learning Representations (ICLR), 2021. 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 311, + 288, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 311, + 288, + 365 + ], + "spans": [ + { + "bbox": [ + 53, + 311, + 288, + 365 + ], + "type": "text", + "content": "[90] Kim Sung-Bin, Lee Hyun, Da Hye Hong, Suekyeong Nam, Janghoon Ju, and Tae-Hyun Oh. Laughtalk: Expressive 3d talking head generation with laughter. In Winter Conference on Applications of Computer Vision (WACV), pages 6404-6413, 2024. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 366, + 288, + 421 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 366, + 288, + 421 + ], + "spans": [ + { + "bbox": [ + 53, + 366, + 288, + 421 + ], + "type": "text", + "content": "[91] Mingyang Sun†, Mengchen Zhao†, Yaqing Hou*, Minglei Li, Huang Xu, Songcen Xu, and Jianye Hao. Cospeech gesture synthesis by reinforcement learning with contrastive pre-trained rewards. In Computer Vision and Pattern Recognition (CVPR), 2023. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 422, + 288, + 466 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 422, + 288, + 466 + ], + "spans": [ + { + "bbox": [ + 53, + 422, + 288, + 466 + ], + "type": "text", + "content": "[92] Marcus Thiébaux, Stacy Marsella, Andrew N. Marshall, and Marcelo Kallmann. SmartBody: behavior realization for embodied conversational agents. In Adaptive Agents and Multi-Agent Systems (AAMAS), 2008. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 468, + 288, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 468, + 288, + 521 + ], + "spans": [ + { + "bbox": [ + 53, + 468, + 288, + 521 + ], + "type": "text", + "content": "[93] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Hervé Jégou. Training data-efficient image transformers & distillation through attention. In International Conference on Machine Learning (ICML), pages 10347-10357, 2021. 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 53, + 523, + 288, + 556 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 523, + 288, + 556 + ], + "spans": [ + { + "bbox": [ + 53, + 523, + 288, + 556 + ], + "type": "text", + "content": "[94] Jonathan Tseng, Rodrigo Castellon, and Karen Liu. Edge: Editable dance generation from music. In Computer Vision and Pattern Recognition (CVPR), 2023. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 53, + 558, + 288, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 558, + 288, + 601 + ], + "spans": [ + { + "bbox": [ + 53, + 558, + 288, + 601 + ], + "type": "text", + "content": "[95] R. Williams. The Animator's Survival Kit: A Manual of Methods, Principles and Formulas for Classical, Computer Games, Stop Motion and Internet Animators. Farrar, Straus and Giroux, 2012. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 53, + 603, + 288, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 603, + 288, + 658 + ], + "spans": [ + { + "bbox": [ + 53, + 603, + 288, + 658 + ], + "type": "text", + "content": "[96] Jingyu Wu, Shi Chen, Shuyu Gan, Weijun Li, Changyuan Yang, and Lingyun Sun. Cultural self-adaptive multimodal gesture generation based on multiple culture gesture dataset. In ACM International Conference on Multimedia (MM), page 3538-3549, 2023. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 53, + 658, + 288, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 658, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 53, + 658, + 288, + 713 + ], + "type": "text", + "content": "[97] Sicheng Yang, Zhiyong Wu, Minglei Li, Zhensong Zhang, Lei Hao, Weihong Bao, Ming Cheng, and Long Xiao. Diffusestylegusture: Stylized audio-driven co-speech gesture generation with diffusion models. In International Joint Conference on Artificial Intelligence (IJCAI), 2023. 3, 6" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 72, + 545, + 699 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 312, + 72, + 545, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 72, + 545, + 138 + ], + "spans": [ + { + "bbox": [ + 312, + 72, + 545, + 138 + ], + "type": "text", + "content": "[98] Sicheng Yang, Zhiyong Wu, Minglei Li, Zhensong Zhang, Lei Hao, Weihong Bao, and Haolin Zhuang. Qpgesture: Quantization-based and phase-guided motion matching for natural speech-driven gesture generation. In Computer Vision and Pattern Recognition (CVPR), pages 2321-2330, 2023. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 312, + 140, + 545, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 140, + 545, + 194 + ], + "spans": [ + { + "bbox": [ + 312, + 140, + 545, + 194 + ], + "type": "text", + "content": "[99] Payam Jome Yazdian, Mo Chen, and Angelica Lim. Gesture2vec: Clustering gestures using representation learning methods for co-speech gesture generation. In International Conference on Intelligent Robots and Systems (IROS), pages 3100-3107, 2022. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 196, + 545, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 196, + 545, + 217 + ], + "spans": [ + { + "bbox": [ + 308, + 196, + 545, + 217 + ], + "type": "text", + "content": "[100] Hongwei Yi. Show github. https://github.com/yhw-yhw/SHOW, 2023.6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 219, + 545, + 272 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 219, + 545, + 272 + ], + "spans": [ + { + "bbox": [ + 308, + 219, + 545, + 272 + ], + "type": "text", + "content": "[101] Hongwei Yi, Hualin Liang, Yifei Liu, Qiong Cao, Yandong Wen, Timo Bolkart, Dacheng Tao, and Michael J Black. Generating holistic 3d human motion from speech. In Computer Vision and Pattern Recognition (CVPR), pages 469-480, 2023. 1, 2, 6, 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 274, + 545, + 317 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 274, + 545, + 317 + ], + "spans": [ + { + "bbox": [ + 308, + 274, + 545, + 317 + ], + "type": "text", + "content": "[102] Li-Ping Yin, Yijun Wang, Tianyu He, Jinming Liu, Wei Zhao, Bohan Li, Xin Jin, and Jianxin Lin. Emog: Synthesizing emotive co-speech 3d gesture with diffusion model. ArXiv, abs/2306.11496, 2023. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 319, + 545, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 319, + 545, + 373 + ], + "spans": [ + { + "bbox": [ + 308, + 319, + 545, + 373 + ], + "type": "text", + "content": "[103] Youngwoo Yoon, Woo-Ri Ko, Minsu Jang, Jaeyeon Lee, Jaehong Kim, and Geehyuk Lee. Robots learn social skills: End-to-end learning of co-speech gesture generation for humanoid robots. In International Conference on Robotics and Automation (ICRA), 2019. 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 375, + 545, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 375, + 545, + 430 + ], + "spans": [ + { + "bbox": [ + 308, + 375, + 545, + 430 + ], + "type": "text", + "content": "[104] Youngwoo Yoon, Bok Cha, Joo-Haeng Lee, Minsu Jang, Jaeyeon Lee, Jaehong Kim, and Geehyuk Lee. Speech gesture generation from the trimodal context of text, audio, and speaker identity. Transactions on Graphics (TOG), 39(6), 2020. 2, 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 431, + 545, + 485 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 431, + 545, + 485 + ], + "spans": [ + { + "bbox": [ + 308, + 431, + 545, + 485 + ], + "type": "text", + "content": "[105] Youngwoo Yoon, Pieter Wolfert, Taras Kucherenko, Carla Viegas, Teodor Nikolov, Mihail Tsakov, and Gustav Eje Henter. The genea challenge 2022: A large evaluation of data-driven co-speech gesture generation. In International Conference on Multimodal Interaction, 2022. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 487, + 545, + 519 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 487, + 545, + 519 + ], + "spans": [ + { + "bbox": [ + 308, + 487, + 545, + 519 + ], + "type": "text", + "content": "[106] Ye Yuan and Kris M. Kitani. DLow: Diversifying latent flows for diverse human motion prediction. In European Conference on Computer Vision (ECCV), 2020. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 520, + 545, + 564 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 520, + 545, + 564 + ], + "spans": [ + { + "bbox": [ + 308, + 520, + 545, + 564 + ], + "type": "text", + "content": "[107] Yicheng Zhong, Huawei Wei, Peiji Yang, and Zhisheng Wang. Expclip: Bridging text and facial expressions via semantic alignment. In AAAI Conference on Artificial Intelligence, pages 7614-7622, 2024. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 566, + 545, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 566, + 545, + 609 + ], + "spans": [ + { + "bbox": [ + 308, + 566, + 545, + 609 + ], + "type": "text", + "content": "[108] Yi Zhou, Connelly Barnes, Jingwan Lu, Jimei Yang, and Hao Li. On the continuity of rotation representations in neural networks. In Computer Vision and Pattern Recognition (CVPR), pages 5745-5753, 2019. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 610, + 545, + 653 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 610, + 545, + 653 + ], + "spans": [ + { + "bbox": [ + 308, + 610, + 545, + 653 + ], + "type": "text", + "content": "[109] Yi Zhou, Jingwan Lu, Connelly Barnes, Jimei Yang, Sitao Xiang, and Hao Li. Generative tinkering: Long-term inbetweening of 3d human motions. arXiv:2005.08891, 2020. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 655, + 545, + 699 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 655, + 545, + 699 + ], + "spans": [ + { + "bbox": [ + 308, + 655, + 545, + 699 + ], + "type": "text", + "content": "[110] Lingting Zhu, Xian Liu, Xuanyu Liu, Rui Qian, Ziwei Liu, and Lequan Yu. Taming diffusion models for audio-driven co-speech gesture generation. In Computer Vision and Pattern Recognition (CVPR), pages 10544–10553, 2023. 2, 8" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1953" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Empowering Resampling Operation for Ultra-High-Definition Image Enhancement with Model-Aware Guidance/5e31bf7f-5e48-4030-90cc-b526f724e932_content_list.json b/2024/Empowering Resampling Operation for Ultra-High-Definition Image Enhancement with Model-Aware Guidance/5e31bf7f-5e48-4030-90cc-b526f724e932_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..8b0ea1d26200c832f0e51721e96f0236decdd121 --- /dev/null +++ b/2024/Empowering Resampling Operation for Ultra-High-Definition Image Enhancement with Model-Aware Guidance/5e31bf7f-5e48-4030-90cc-b526f724e932_content_list.json @@ -0,0 +1,2076 @@ +[ + { + "type": "text", + "text": "Empowering Resampling Operation for Ultra-High-Definition Image Enhancement with Model-Aware Guidance", + "text_level": 1, + "bbox": [ + 138, + 130, + 831, + 174 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Wei Yu,\\* Jie Huang,\\* Bing Li, Kaiwen Zheng, Qi Zhu, Man Zhou, Feng Zhao† \nUniversity of Science and Technology of China", + "bbox": [ + 204, + 203, + 810, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{patrick914y, hj0117,bing0123, kezh,zqcrafts, manman}@mail.ustc.edu.cn，fzhao956@ustc.edu.cn", + "bbox": [ + 86, + 241, + 916, + 258 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/1a5bd2092423b406acafa244151d833ba75749badadefc7ba13a96ee075c0614.jpg", + "image_caption": [ + "(a) Down-Enhancement-Up" + ], + "image_footnote": [], + "bbox": [ + 78, + 285, + 184, + 351 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/588349239bdf26e4b63c890585679a3a59f70d6928bd7c396e4bdd5ad9ddec95.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 183, + 286, + 300, + 351 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/57ad2f614867adbb59a1d277191baa8e845b8e1a394d49956ee0640f4361b690.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 303, + 286, + 408, + 351 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/ec1ce28b3efdcc0349ecac11d3cb237058ee621b36a018375d4d579e8492afcf.jpg", + "image_caption": [ + "(b) Ours" + ], + "image_footnote": [], + "bbox": [ + 80, + 367, + 408, + 450 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/e47c2c900f580925422897fe14d1eeca5fb205c328a80ddc29835e93d702e8a8.jpg", + "image_caption": [ + "PSNR: 25.59" + ], + "image_footnote": [], + "bbox": [ + 428, + 286, + 532, + 349 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/5c97497c3d726a06d80ee57c6a2eb657c0ae9765837a8d269c3ee8d352959be1.jpg", + "image_caption": [ + "PSNR: 28.51" + ], + "image_footnote": [], + "bbox": [ + 545, + 286, + 650, + 349 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/af1759eaa900b0b71b03a975c18a833fe40ab7afa420db4539b6ededada66984.jpg", + "image_caption": [ + "PSNR: 30.78" + ], + "image_footnote": [], + "bbox": [ + 661, + 286, + 764, + 349 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/5065405d10d298974acbfb18cd7764d5be0d0d4d01b836e55174c755e44829ae.jpg", + "image_caption": [ + "PSNR: 31.05" + ], + "image_footnote": [], + "bbox": [ + 777, + 286, + 882, + 351 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/9856432f41090743975fd662f70cc6401d7d0eb86905334b297cdae64fef9f13.jpg", + "image_caption": [ + "PSNR: 28.09", + "PSNR: 29.56", + "PSNR: 32.24", + "PSNR: 33.16", + "Figure 1. Comparison between previous Downsampling-Enhancement-Upsampling paradigm and our proposed method. (a) Previous works treat resampling operators and inner models as separate components, while (b) our method integrates them via empowering model-aware resampling. As shown on the right side, our approach achieves significant performance gains over various resampling scales." + ], + "image_footnote": [], + "bbox": [ + 428, + 364, + 890, + 450 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 536, + 313, + 551 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Image enhancement algorithms have made remarkable advancements in recent years, but directly applying them to Ultra-high-definition (UHD) images presents intractable computational overheads. Therefore, previous straightforward solutions employ resampling techniques to reduce the resolution by adopting a \"Downsampling-Enhancement-Upsampling\" processing paradigm. However, this paradigm disentangles the resampling operators and inner enhancement algorithms, which results in the loss of information that is favored by the model, further leading to sub-optimal outcomes. In this paper, we propose a novel method of Learning Model-Aware Resampling (LMAR), which learns to customize resampling by extracting model-aware information from the UHD input image, under the guidance of model knowledge. Specifically, our method consists of two core designs, namely compensatory kernel estimation and steganographic resampling. At the first stage, we dynamically predict compensatory kernels tailored to the specific input and resampling scales. At", + "bbox": [ + 73, + 570, + 472, + 858 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "the second stage, the image-wise compensatory information is derived with the compensatory kernels and embedded into the rescaled input images. This promotes the representation of the newly derived downscaled inputs to be more consistent with the full-resolution UHD inputs, as perceived by the model. Our LMAR enables model-aware and model-favored resampling while maintaining compatibility with existing resampling operators. Extensive experiments on multiple UHD image enhancement datasets and different backbones have shown consistent performance gains after correlating resizer and enhancer, e.g., up to 1.2dB PSNR gain for $\\times 1.8$ resampling scale on UHD-LOL4K. The code is available at https://github.com/YPPatrickW/LMAR.", + "bbox": [ + 500, + 537, + 893, + 734 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 500, + 739, + 632, + 755 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Over time, learning-based image enhancement algorithms have achieved progressive performance improvements. However, the majority of existing methods are not compatible with UHD images due to the heavy computational burdens imposed by their megapixel counts.", + "bbox": [ + 496, + 765, + 890, + 840 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To address this, previous works have sought to alleviate computation overheads by employing downsampling to reduce the resolution, as shown in Figure 1. Concretely, UHD input images are often resampled to smaller sizes for", + "bbox": [ + 498, + 840, + 893, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Both authors contributed equally to this research.", + "bbox": [ + 93, + 875, + 359, + 888 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "† Corresponding author.", + "bbox": [ + 96, + 888, + 220, + 901 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "25722", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/4b768982962e32cc832c9c561b31d4e6530d9dd28d6965b8a93c755ab20382b3.jpg", + "image_caption": [ + "Figure 2. The influence of the front-end resampler. We choose the cubic resampling operator as the baseline and introduce perturbations to its kernel function. Interestingly, manipulating the resampling function can lead to volatile outcomes, with the image quality either being improved or degraded. We also present two more comparisons: the resampling residual of the UHD input image and the feature maps generated by different resizers. It is indicated that the response of the model is highly sensitive to the sampling strategies. This highlights the importance of customizing the resampling process in a model-aware manner." + ], + "image_footnote": [], + "bbox": [ + 78, + 88, + 472, + 233 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "enhancement, and then subsequently resampled back to the original resolution[22, 39]. Surprisingly, the majority of research efforts have been paid to developing inner processing algorithms, with little attention devoted to the foremost and endmost resampling operators.", + "bbox": [ + 75, + 397, + 468, + 473 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recently, a line of work has proposed the replacement of the conventional interpolation-based resampling operators (e.g., nearest, bilinear, bicubic) with learnable resampling operators [12, 20, 28, 30]. Attributing to the superior modeling capability of deep neural networks, these learnable resampling operators have demonstrated remarkable performance gains. However, the practical application of these learnable resampling operators is hindered by their high computational complexity and reduced efficiency, resulting in a continued preference for interpolation-based resampling operators.", + "bbox": [ + 75, + 474, + 468, + 641 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Revisiting the above UHD image enhancement pipeline: Downsampling-Enhancement-Upsampling, with the internal enhancement algorithm remaining unchanged, a fundamental question arises: 'How significant is the impact of these two resampling operations?' The answer to this question is twofold. On the one hand, the quality of the upsampled enhanced results exhibits a decreasing trend as the degree of downsampling aggravates. On the other hand, manipulating the interpolation kernel functions yields a diverse range of outcomes, as illustrated in Figure 2. We attribute this phenomenon to the isolation between the resampling operators and the enhancement algorithm, where both components are disentangled from each other.", + "bbox": [ + 75, + 642, + 468, + 838 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To this end, the focus of this paper lies in correlating the resampling operators within the internal enhancement algorithms and unleashing the potential of the resampling operation to make benefits for the aforementioned framework,", + "bbox": [ + 75, + 839, + 468, + 898 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "without modifying the inner processing methods (e.g., pretrained networks). To accomplish this goal, we propose a novel method termed Learning Model-Aware Resampling (LMAR). LMAR correlates the resizer and the enhancer through a compensatory information learning process.", + "bbox": [ + 498, + 90, + 890, + 167 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We begin by delving into the impact of resampling from the viewpoint of the enhancer, where the resampling operations directly undermine the representation ability of intermediate features. Therefore, the intermediate features of the UHD image offer valuable insights on how to perform resampling in a model-aware manner. Building upon this recognition, we formulate our method into two steps: Compensatory Kernel Estimation (CKE) and Steganographic Resampling (SR). In the CKE step, we generate imagespecific convolution kernels for the input UHD image via implicit neural representation. These kernels are specifically customized for each input and resampling scale. Then, we can obtain compensatory information produced by these predicted filters. Subsequently, in the SR step, the compensatory information is embedded in the downscaled input image, resulting in a newly learnable downscaled input. By utilizing the compensatory information as a medium, we establish the correlation between the resampling operators and the inner enhancer. Finally, the learnable input is passed to the enhancer and optimized to encourage that the intermediate features align more consistently with those of the UHD image. Thanks to these designs, our method not only maintains compatibility with existing interpolation-based resampling operators but also incorporates them into a part of enhancement algorithms under the guidance of model knowledge. Our contributions are summarized as:", + "bbox": [ + 496, + 194, + 892, + 585 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We rethink the disentanglement between the resizer and enhancement model and propose to perform the resampling in a model-aware manner.", + "- We introduce LMAR, a novel method that facilitates the collaboration between resizer and enhancer via customizing resampling under the guidance of model knowledge, where the representative capabilities of low-resolution images are largely improved.", + "- Our LMAR delivers two core designs: compensatory kernel estimation and steganographic resampling, which models the interplay between responder and enhancer by learning embeddable information. Our method is compatible with any interpolation resamplers.", + "- Extensive experiments are conducted to verify the superiority of correlating resizer and enhancer. Remarkably, our algorithm significantly improves the performance at low resolutions while maintaining equivalent results at the original resolution, without requiring retraining of the enhancer." + ], + "bbox": [ + 500, + 613, + 890, + 898 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "25723", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/a999215e919be8d6ab60f20228b1e4e496f972e2b3d45178b34cf3393b9edf1b.jpg", + "image_caption": [ + "Figure 3. The inspiration for our method design. We replace the compressed bottleneck features of the downsampled input with features extracted from the UHD image (The green arrow). Large-margin performance gains can be achieved compared with the vested paradigm (The gray arrow). This observation highlights the potential role of feature consistency between UHD input and any downscaled input, which can play as a bridge for establishing correlations between resampling operators and enhancers." + ], + "image_footnote": [], + "bbox": [ + 78, + 85, + 472, + 251 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 76, + 382, + 218, + 398 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.1. Image Resampling", + "text_level": 1, + "bbox": [ + 76, + 407, + 256, + 424 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Image resampling has emerged as one of the most commonly employed techniques for altering the resolution of an image. The key to image resampling lies in effectively preserving the quality of image content while maintaining efficiency.", + "bbox": [ + 75, + 430, + 468, + 505 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Interpolation operators, such as bilinear, and bicubic [1, 8], have been the popular choice for image resampling for many years. These operators are capable of effectively resampling images at any scale, yet they disregard the spatial variations of distinct image patches, leading to inadequate preservation of the local structure [26].", + "bbox": [ + 75, + 506, + 468, + 595 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In recent years, deep neural networks have facilitated advancements in various tasks [5, 10, 13-19, 31, 32, 36, 42-48], including image resampling [4, 11, 12, 20, 28, 30, 34, 35]. Task-aware image downsampling (TAD) [20] presents an encoder-decoder framework that aims to enhance the quality of both downscaling and upscaling processes in a collaborative manner. Though these methods achieve excellent results, none of them focus on correlating interpolation resampler with a pre-trained model.", + "bbox": [ + 75, + 597, + 468, + 732 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. UHD Image Processing", + "text_level": 1, + "bbox": [ + 76, + 741, + 292, + 758 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "A line of studies that specifically focus on high-resolution image processing has been proposed [2, 6, 23, 33, 40, 41]. These methods share a common paradigm: learning from downscaled images to alleviate the computational burdens. In this learning paradigm, the quality of the results is directly determined by the resampled image representation, highlighting the crucial role of resampling strategies. However, they often overlook the impact of resampling approaches and lack in-depth investigation of various resam", + "bbox": [ + 75, + 763, + 468, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "pling operators. In this work, our objective is to explore the interplay between enhancer and resampling operators and establish a collaborative relationship for them to improve performance.", + "bbox": [ + 496, + 90, + 890, + 151 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Methodology", + "text_level": 1, + "bbox": [ + 500, + 162, + 633, + 180 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we will present a comprehensive illustration of our proposed LMAR. We begin by introducing the motivation behind our approach and describing our settings. Subsequently, we will overview the whole processing pipeline and delve into the design of the Compensatory Kernel Estimation (CKE) and Steganographic Resampling (SR). Finally, we will discuss the optimization strategies employed to enable model-aware resampling.", + "bbox": [ + 496, + 188, + 890, + 309 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Setting and Motivation", + "text_level": 1, + "bbox": [ + 500, + 316, + 712, + 332 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our setting still follows the Downsampling-Enhancement-Upsampling paradigm, employing the interpolation-based resampling operators instead of introducing additional burdens with learnable resampling operators. Notably, we will first train an enhancement model for each dataset in our experimental setup and keep unchanged. The main difference lies in the collaboration process between the resizing and enhancer.", + "bbox": [ + 496, + 339, + 890, + 459 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The main inspiration stems from observing the impact of resampling on the representative feature extraction process of the enhancer. As shown in Figure 3, we inject the full-resolution bottleneck features into the enhancement process of the downsampled input images, where the final results significantly outperform the vested paradigm. It proves that front-most resampling degrades the representative ability of bottleneck features and impacts the following processes. Based on the insight, we correlate the resizer and enhancer through a compensatory information-learning process to achieve consistent feature alignment.", + "bbox": [ + 496, + 460, + 890, + 626 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Overview", + "text_level": 1, + "bbox": [ + 500, + 633, + 609, + 648 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The Downsampling-Enhancement-Upsampling processing can be mathematically formulated as:", + "bbox": [ + 496, + 657, + 890, + 686 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nx _ {d} = D (x; [ h, w ]),\n$$\n", + "text_format": "latex", + "bbox": [ + 625, + 694, + 754, + 710 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {y} = F \\left(x _ {d}\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 635, + 713, + 890, + 729 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\ny = U (\\tilde {y}; [ H, W ]),\n$$\n", + "text_format": "latex", + "bbox": [ + 635, + 732, + 764, + 750 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $D(\\cdot)$ is the down sampler that resamples the UHD input image $x \\in \\mathbb{R}^{H \\times W \\times 3}$ to lower resolution of $x_{d} \\in \\mathbb{R}^{h \\times w \\times 3}$ , $F(\\cdot)$ is the pre-trained enhancement model, and $U(\\cdot)$ denotes the upsampler which resamples the low-resolution enhanced result $\\tilde{y}$ back to the UHD resolution of $H \\times W$ , termed as $y$ . In LMAR, we rebuild the above paradigm, given as:", + "bbox": [ + 496, + 755, + 890, + 861 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {y} = F \\left(x _ {d}, \\Theta (x; D, U, F)\\right), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 602, + 867, + 890, + 891 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\ny = U (\\tilde {y}; [ H, W ]).\n$$\n", + "text_format": "latex", + "bbox": [ + 604, + 888, + 730, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "25724", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/01ad7bbe2229e30ed77c4945b98e2edd3255d39b3569939d760f107929f0459f.jpg", + "image_caption": [ + "1. Training Stage" + ], + "image_footnote": [], + "bbox": [ + 91, + 112, + 532, + 253 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/a5164a9d7600e4e3125278c151df765d3ecf52d49d6eb923f406ce623047717c.jpg", + "image_caption": [ + "2. Inference Stage", + "(b)" + ], + "image_footnote": [], + "bbox": [ + 547, + 112, + 885, + 247 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/d3ee7d979e37be6c5a7a1570df5a9130f3d38ec79b0c57b80f132495d3d286ac.jpg", + "image_caption": [ + "(a)" + ], + "image_footnote": [], + "bbox": [ + 81, + 271, + 679, + 422 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/900be3f80c5813429b86065c569562b523d9eb1fed3b0b29657625088a47802e.jpg", + "image_caption": [ + "Figure 4. Overview of the proposed LMAR. The sub-graph (a) depicts the training phase of LMAR, which encourages the compensated low-resolution input to maintain representation consistency with the full-resolution UHD input, as perceived by the enhancer. The subgraph (b) demonstrates the inference pipeline of our LMAR, where the compensated low-resolution input is directly fed into the enhancer and then upsampled to the UHD result. The sub-graph (c) illustrates how LMAR works, where the core lies in estimating compensatory kernels under the guidance of model knowledge to make up for the resampling process." + ], + "image_footnote": [], + "bbox": [ + 683, + 272, + 890, + 422 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Here, we introduce additional compensatory information as a medium to facilitate model-aware resampling, denoted as $\\Theta (x;D,U,F)$ . Our method is presented in Figure 4, where the model-aware resampling is achieved via two core designs, compensatory kernel estimation, and steganographic resampling.", + "bbox": [ + 75, + 517, + 470, + 609 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Compensatory Kernel Estimation", + "text_level": 1, + "bbox": [ + 76, + 628, + 370, + 645 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Interpolation-based image resampling process can be decomposed as projected grid calculation and weighted pixel aggregation [29]. To this end, we estimate the compensatory kernel using a representation that is related to the scale. Inspired by LIFF [4], we employ two scale features, the relative coordinate grid and the pixel cell. The relative coordinate grid depicts the pixel location shift in the resampling processing. We first compute the uniform coordinate of the downsampling scale $C_d \\in \\mathbb{R}^{h \\times w \\times 2}$ and the UHD scale $C_u \\in \\mathbb{R}^{H \\times W \\times 2}$ , which are normalized between $[-1, 1]$ . The calculations of the above grids are given as:", + "bbox": [ + 75, + 657, + 468, + 824 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nC _ {d} (i, j) = (- 1 + \\frac {(2 i + 1)}{h}, - 1 + \\frac {(2 j + 1)}{w}), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 119, + 842, + 468, + 878 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nC _ {u} (i, j) = (- 1 + \\frac {(2 i + 1)}{H}, - 1 + \\frac {(2 j + 1)}{W}),\n$$\n", + "text_format": "latex", + "bbox": [ + 120, + 875, + 424, + 905 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $i \\in [0, h - 1] \\vee [0, H - 1], j \\in [0, w - 1] \\vee [0, W - 1]$ are the position indexes for the height and width dimensions. As the sizes of the two coordinates are different, we first project them $C_d$ to $C_u$ and obtain the relative grid shift at the downsampling scale, which is expressed as:", + "bbox": [ + 500, + 516, + 890, + 593 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {C} _ {d} = \\operatorname {g r i d} \\text {- s a m p l e} \\left(C _ {d}, C _ {u}\\right),\n$$\n", + "text_format": "latex", + "bbox": [ + 589, + 598, + 779, + 616 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {C} _ {r} = C _ {u} - \\tilde {C} _ {d}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 591, + 619, + 890, + 638 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nC _ {r} = \\left(\\tilde {C} _ {r} (i) \\times h, \\tilde {C} _ {r} (j) \\times w\\right),\n$$\n", + "text_format": "latex", + "bbox": [ + 591, + 641, + 799, + 659 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where the grid_sample is a remapping function that remaps the $C_d$ into $\\tilde{C}_d$ of size $H \\times W$ , $\\tilde{C}_r$ represents the relative shift and $C_r$ denotes the relative coordinate grid measured at downsampling scale. The pixel cell represents the ratio of pixel area changes between downsampling and upsampling, denoted as:", + "bbox": [ + 496, + 664, + 890, + 753 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nP _ {c} (i, j) = \\left(2 * \\frac {h}{H}, 2 * \\frac {w}{W}\\right), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 599, + 758, + 890, + 789 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $P_{c}\\in \\mathbb{R}^{H\\times W\\times 2}$ , and all pixel locations share the same scale ratio.", + "bbox": [ + 496, + 794, + 890, + 823 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The aforementioned scale-related representations are instance-same at a given scale, leading to the generation of identical kernels without considering the distinctions between images. Thus, we propose the utilization of backprojection [9] as a constraint to capture the interrelation", + "bbox": [ + 496, + 825, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "25725", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/5a71842bcd4ce590e6d4f429e2a04ff9705d6a101bbcdba719ff25e08352c242.jpg", + "image_caption": [ + "(a) $x_{d}$" + ], + "image_footnote": [], + "bbox": [ + 80, + 88, + 202, + 219 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/4d7591c13ae2781df508dd572e0804e77da7711bce344a22fecfe30ca143ca3f.jpg", + "image_caption": [ + "(b) $x_{c}$" + ], + "image_footnote": [], + "bbox": [ + 212, + 88, + 336, + 219 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/a7e30423361ebdce19e2d1e497ea1d2867d803195ef9465e17cbcca7cb10beae.jpg", + "image_caption": [ + "(c) $|x_{c} - x_{d}|$", + "Figure 5. Comparison of different downscaled input images. (a) The downscaled input images by the cubic resampling operator, (b) images derived from the cubic resampling operator in a model-aware manner, and (c) the residual between (a) and (b), where obvious differences can be witnessed in high-frequency areas." + ], + "image_footnote": [], + "bbox": [ + 346, + 88, + 470, + 219 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "between each downscaled input and its corresponding UHD counterpart, formulated as:", + "bbox": [ + 75, + 335, + 468, + 367 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {x} = \\operatorname {C o n v} (x - U (x _ {d}; [ H, W ]))), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 158, + 378, + 468, + 395 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\tilde{x} \\in \\mathbb{R}^{H \\times W \\times 3}$ is the back-projection of UHD input image $x$ .", + "bbox": [ + 75, + 406, + 468, + 436 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Considering the pixel-wise aggregation nature of the resampling process, we model the generation of compensatory information by predicting the convolution kernels for the UHD input image instead of predicting pixel values for specific locations. Specifically, we flatten the relative coordinate $C_r$ , pixel cell $P_c$ , and back-projection $\\tilde{x}$ then concatenate them. Afterward, we feed these concatenated features into a multi-layer perception (MLP) to predict compensatory kernels as follows:", + "bbox": [ + 75, + 438, + 468, + 574 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nK = M L P _ {\\phi} \\left(\\left[ C _ {r}; P _ {c}; \\tilde {x} \\right]\\right), \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 184, + 585, + 468, + 603 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $[\\cdot ]$ denotes the channel-wise concatenation, and $K\\in$ $\\mathbb{R}^{H\\times W\\times s}$ represents the predicted compensatory kernels. The kernel for each pixel location is represented in the channel dimension $s$ , consisting of the multiplication of the input channel, out channel, and kernel size, with all being three in our case.", + "bbox": [ + 76, + 614, + 468, + 704 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Steganographic Resampling", + "text_level": 1, + "bbox": [ + 76, + 715, + 328, + 732 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "With the estimated compensatory kernel from the above process, we calculate the compensatory information for the input UHD image as follows:", + "bbox": [ + 75, + 738, + 468, + 784 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\widehat {x} = K \\otimes x, \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 228, + 796, + 468, + 811 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\widehat{x} \\in \\mathbb{R}^{H \\times W \\times 3}$ is the generated UHD compensatory embedding, and $\\otimes$ denotes the convolution operation.", + "bbox": [ + 75, + 823, + 468, + 854 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To further maintain the intrinsic characteristics of the downscaled input images, we downscale the compensatory information and perform compensatory steganography in", + "bbox": [ + 75, + 854, + 468, + 901 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/cbccefd1c5f2b8da13056c8bfa704158bdd5f1e19ec2f65febf874bfa689d51f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 501, + 88, + 627, + 152 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/d41d07451a105764a98e9d2340ea2da8e7b97f63b1bc72e83552cab7d39f231c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 635, + 88, + 759, + 152 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/5ff6d4209e997f5cd29523555b0c59a245220e7a1a123c4e00efac79befb38fb.jpg", + "image_caption": [ + "Errors" + ], + "image_footnote": [], + "bbox": [ + 767, + 88, + 893, + 152 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/c4912282f7391e97ff6a710a8d743beec2d9b9cad447ed23be325295a1d5bec0.jpg", + "image_caption": [ + "$x_{d}$", + "$x_{c}$" + ], + "image_footnote": [], + "bbox": [ + 501, + 166, + 627, + 229 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/a0c3bbed12608d45557a1c1c7fb4a9fbf66f25cdd46d16835f2722094bdf7203.jpg", + "image_caption": [ + "$E(x_{d})$", + "$E(x_{c})$" + ], + "image_footnote": [], + "bbox": [ + 635, + 166, + 759, + 229 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/410630a60511c85aaf12b17e1a05d907be443a8873753c2af0423aeb3ebbcc3e.jpg", + "image_caption": [ + "Errors", + "Figure 6. Feature comparison of different downscaled input images. The second column represents the bottleneck features of the downscaled inputs. As observed in the error maps with the UHD features in the third column, downscaled input produced in a model-aware manner $(x_{c})$ maintains a higher level of representative consistency with the UHD input, as perceived by the model." + ], + "image_footnote": [], + "bbox": [ + 767, + 166, + 893, + 231 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "the low-resolution space, which is computed as:", + "bbox": [ + 500, + 354, + 816, + 371 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nx _ {c} = \\operatorname {C o n v} ([ x _ {d}; D (\\widehat {x}; [ h, w ]) ]). \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 586, + 382, + 890, + 400 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Here, we concatenate the downscaled UHD image and downscaled compensatory information, then perform channel reduction by convolution layer. In this way, the newly derived $x_{c}$ can be optimized under the guidance of model knowledge, where we utilize the compensatory information to bridge the resizer and enhancer.", + "bbox": [ + 498, + 410, + 890, + 501 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.5. Optimization Object", + "text_level": 1, + "bbox": [ + 500, + 510, + 694, + 527 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Given the pre-trained enhancer, the bottleneck features of the UHD input image possess the most powerful representational capability compared to all of its downscaled counterparts. Therefore, we encourage the consistency of bottleneck features, which is referred to as resampling consistency loss:", + "bbox": [ + 498, + 534, + 890, + 625 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {r c} = \\operatorname {S m o o t h L 1} (E (x) - U (E \\left(x _ {c}\\right))), \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 566, + 636, + 890, + 652 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where the $E(x)$ and $E(x_{c})$ represent the corresponding bottleneck features of the UHD input $x$ and learnable down-scaled input $x_{c}$ . $U(\\cdot)$ is used for size adjustment. In addition, we introduce attention loss [37] for these two bottleneck features, given as:", + "bbox": [ + 498, + 664, + 890, + 739 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {f a} = \\left\\| \\frac {E (x)}{\\| E (x) \\| _ {2}} - \\frac {U \\left(E \\left(x _ {c}\\right)\\right)}{\\| U \\left(E \\left(x _ {c}\\right)\\right) \\| _ {2}} \\right\\| _ {2}, \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 565, + 750, + 890, + 785 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\| \\cdot \\| _2$ is the $\\mathcal{L}_2$ normalization and this loss encourages the attention transfer from $E(x)$ to $U(E(x_{c}))$ .", + "bbox": [ + 498, + 794, + 888, + 824 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The aforementioned losses are designed to incorporate model-favored information into the learnable downscaled input image $x_{c}$ but neglect the distribution consistency between the naive downscaled images $x_{d}$ and newly derived images $x_{c}$ . Thus, we need to impose restrictions on the", + "bbox": [ + 498, + 825, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "25726", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/f43d798d3b19c208acf55c398c581c0b6a109303373bb87722b141483baa0b11.jpg", + "table_caption": [ + "Table 1. Quantitative results on the UHD-LOL4K datasets with CNN-I as the backbone. The results with LMAR are shown in gray with better results highlighted in bold." + ], + "table_footnote": [], + "table_body": "
MethodCNN-I (UHD-LOL4K)
Scales(2160, 3840)(1440, 2560)(1080, 1920)(1200, 1600)(720, 1280)(540, 960)(432, 768)
Nearest31.92 / 0.984328.20 / 0.981527.01 / 0.979226.66 / 0.976325.86 / 0.976123.99 / 0.961523.06 / 0.9445
31.86 / 0.984128.46 / 0.982427.69 / 0.981327.42 / 0.979026.60 / 0.980624.38 / 0.971523.98 / 0.9608
Bilinear31.92 / 0.984331.07 / 0.983929.75 / 0.982829.16 / 0.982027.01 / 0.976725.90 / 0.974424.03 / 0.9448
31.88 / 0.984131.28 / 0.984030.17 / 0.983529.85 / 0.983227.94 / 0.981426.27 / 0.979524.85 / 0.9613
Bicubic31.92 / 0.984331.19 / 0.984029.82 / 0.983429.26 / 0.982227.16 / 0.976725.40 / 0.969424.04 / 0.9450
31.89 / 0.984231.42 / 0.984130.20 / 0.983929.96 / 0.983627.97 / 0.981526.36 / 0.979224.88 / 0.9617
Lanczos231.92 / 0.984331.04 / 0.984029.83 / 0.983429.22 / 0.981926.97 / 0.976125.39 / 0.969424.03 / 0.9450
31.88 / 0.984131.39 / 0.984130.19 / 0.983829.97 / 0.983427.90 / 0.981526.29 / 0.978924.87 / 0.9617
Lanczos331.92 / 0.984330.86 / 0.983829.08 / 0.981528.88 / 0.981327.10 / 0.976425.04 / 0.966323.86 / 0.9438
31.90 / 0.984231.59 / 0.984230.24 / 0.983630.13 / 0.983427.93 / 0.981725.79 / 0.971824.66 / 0.9611
Time (ms)9.79.613.313.414.815.415.0
10.710.213.313.815.115.615.4
", + "bbox": [ + 76, + 127, + 898, + 327 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/d58b29bd97dcbeff9ae8c5a7a15c91f5034b1e344eaa0ded7731dedb54039240.jpg", + "image_caption": [ + "(a) Input", + "Figure 7. Qualitative results on the UHD-LOL4K dataset of different scales. The top row is obtained from the cubic operator without LMAR, while the bottom row is with LMAR. Please zoom in for details." + ], + "image_footnote": [], + "bbox": [ + 80, + 339, + 176, + 479 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/37ddaff8a6b67033470bf78130bd2dd64710b9de7bbc10d618b31f1c9eaedb44.jpg", + "image_caption": [ + "(b) $\\times 5$" + ], + "image_footnote": [], + "bbox": [ + 181, + 339, + 279, + 479 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/68ae2f9151fd7ac58b63a14b8ad90d1fd2411921787a80bba76d8ca97719715b.jpg", + "image_caption": [ + "(c) $\\times 4$" + ], + "image_footnote": [], + "bbox": [ + 282, + 339, + 382, + 479 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/ea6548538423be7c55c9315ad3ac97411de436e82b619618d29f76b2961bf71b.jpg", + "image_caption": [ + "(d) $\\times 3$" + ], + "image_footnote": [], + "bbox": [ + 383, + 339, + 483, + 479 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/356762e0677b95d010cadcd0860f5044d4346594890b06e5bd6efb7548d1a57b.jpg", + "image_caption": [ + "(e) $\\times 2$" + ], + "image_footnote": [], + "bbox": [ + 486, + 339, + 586, + 479 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/cab9aafb8a7e31bdf52c03983d868d623942a532c24bd881ebc21b421ba21552.jpg", + "image_caption": [ + "$(\\mathrm{f})\\times 1.5$" + ], + "image_footnote": [], + "bbox": [ + 588, + 339, + 687, + 479 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/187165f31835c1d6584b9866bad3bf3172e6de7eca1597eea320e74caccd91d7.jpg", + "image_caption": [ + "$(\\mathrm{g})\\times 1$" + ], + "image_footnote": [], + "bbox": [ + 691, + 339, + 790, + 479 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/7fa57e3df1902bb6b1cef9e7657a5975611274c15f331eb810567e8c4215a284.jpg", + "image_caption": [ + "(h) GT" + ], + "image_footnote": [], + "bbox": [ + 792, + 339, + 890, + 479 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "$x_{c}$ to ensure its similarity to $x_{d}$ . However, measuring the distances in the pixel space deteriorates the learning of the above process. Inspired by [7], we add a discriminator and impose $x_{c}$ and $x_{d}$ to confound it, known as adversarial loss, formulated as:", + "bbox": [ + 75, + 561, + 468, + 637 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {d} = \\log \\left(D \\left(x _ {d}\\right)\\right) + \\log \\left(1 - D \\left(x _ {c}\\right)\\right), \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 145, + 645, + 468, + 662 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $D(\\cdot)$ is the discriminator, which ensures the interaction between the resizer and enhancer while maintaining distribution consistency between $x_{c}$ and $x_{d}$ . In addition, to account for the introduction of the discriminator, a new generation loss is incorporated to constrain $x_{c}$ , computed as:", + "bbox": [ + 75, + 667, + 468, + 744 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {g} = \\log \\left(D \\left(x _ {c}\\right)\\right). \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 209, + 752, + 468, + 768 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Overall, our optimization objects are twofold. The first optimization focus is on correlating the resizer and enhancer:", + "bbox": [ + 75, + 773, + 468, + 818 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {c} = \\mathcal {L} _ {r c} + \\alpha \\mathcal {L} _ {f a} + \\beta \\mathcal {L} _ {g}, \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 181, + 820, + 468, + 837 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\alpha$ and $\\beta$ are adjustable parameters to balance different losses, which are set empirically. The second optimization object is $\\mathcal{L}_d$ to restrict the distribution. These two optimization objects are alternatively optimized. As can be", + "bbox": [ + 75, + 839, + 468, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "seen in Figures 5 and 6, our downscaled images not only maintain consistency in terms of perceptual distribution but also exhibit a more consistent representation, as perceived by the model.", + "bbox": [ + 496, + 561, + 892, + 625 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 500, + 638, + 633, + 656 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Experimental Setup", + "text_level": 1, + "bbox": [ + 500, + 664, + 689, + 681 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Datasets. To evaluate the significance of correlating the enhancer and resizer, experiments are conducted on two UHD low-light enhancement datasets, including the UHD-LOL4K dataset proposed by [33] and the 4KIL dataset proposed by [24]. More details about datasets are provided in the supplementary materials.", + "bbox": [ + 496, + 688, + 890, + 779 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Pre-training. We begin by training multiple enhancers $F(\\cdot)$ to verify the effectiveness of our method. Specifically, we employ the widely used encoder-decoder architecture in image restoration tasks as the backbone. Regarding the UHD-LOL4K dataset, we leverage the invertible block proposed by [25] as the basic unit for processing, termed CNN-I. For the 4KIL dataset, we replace the basic processing unit with the half instance normalization block proposed", + "bbox": [ + 496, + 779, + 892, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "25727", + "bbox": [ + 478, + 944, + 517, + 957 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/a6c09525226d53e9ba60f76f9a8eab0b9f99dcb52c074f9f96f501ce2dfee28b.jpg", + "table_caption": [ + "Table 2. Quantitative results on the 4KIL datasets with CNN-H as the backbone. The results with LMAR are shown in gray with better results highlighted in bold." + ], + "table_footnote": [], + "table_body": "
MethodCNN-H (4KIL)
Scales(2160, 3840)(1440, 2560)(1080, 1920)(1200, 1600)(720, 1280)(540, 960)(432, 768)
Nearest29.80 / 0.980528.36 / 0.977927.71 / 0.975427.67 / 0.974226.98 / 0.971925.92 / 0.961625.37 / 0.9428
29.77 / 0.980128.53 / 0.978928.11 / 0.977927.93 / 0.976227.21 / 0.974526.16 / 0.967225.56 / 0.9493
Bilinear29.80 / 0.980529.12 / 0.979628.47 / 0.978528.39 / 0.977927.30 / 0.972626.38 / 0.965925.56 / 0.9439
29.79 / 0.980429.14 / 0.980428.55 / 0.979428.51 / 0.979027.52 / 0.975026.57 / 0.968725.62 / 0.9464
Bicubic29.80 / 0.980529.25 / 0.979728.60 / 0.978728.47 / 0.978027.30 / 0.972626.36 / 0.965025.58 / 0.9444
29.80 / 0.980429.35 / 0.980428.74 / 0.979728.96 / 0.979327.53 / 0.975226.57 / 0.968525.79 / 0.9509
Lanczos229.80 / 0.980529.24 / 0.979728.59 / 0.978728.46 / 0.978027.29 / 0.972626.35 / 0.964925.56 / 0.9444
29.80 / 0.980329.34 / 0.980528.72 / 0.979628.63 / 0.979227.54 / 0.975226.57 / 0.968425.78 / 0.9509
Lanczos329.80 / 0.980529.32/ 0.979828.69 / 0.979828.54 / 0.978127.34 / 0.973026.34 / 0.964525.58 / 0.9457
29.79 / 0.980329.48 / 0.980428.89 / 0.979928.76 / 0.979427.60 / 0.975626.58 / 0.968625.80 / 0.9523
Time (ms)5.05.05.86.36.56.96.9
5.95.76.16.56.88.58.0
", + "bbox": [ + 76, + 128, + 898, + 327 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/60b002ea2323e7f5640a5347f72f0a4f34effb0971bc47f066d92c7ace3dccb0.jpg", + "table_caption": [ + "Table 3. Quantitative results on the 4KIL datasets with Restormer as the backbone. The results with LMAR are shown in gray with better results highlighted in bold." + ], + "table_footnote": [], + "table_body": "
MethodRestromer (4KIL)
Scales(2160, 3840)(1440, 2560)(1080, 1920)(1200, 1600)(720, 1280)(540, 960)(432, 768)
Nearest28.86 / 0.977327.81 / 0.975027.57 / 0.973827.45 / 0.972026.99 / 0.971425.91 / 0.960825.31 / 0.9470
28.78 / 0.976828.36 / 0.976527.67 / 0.974227.80 / 0.973927.27 / 0.973325.98 / 0.961925.46 / 0.9527
Bilinear28.86 / 0.977328.63 / 0.977028.50 / 0.976828.49 / 0.976727.56 / 0.971926.71 / 0.968625.71 / 0.9459
28.83 / 0.977028.75 / 0.977128.52 / 0.976928.64 / 0.977027.85 / 0.973926.77 / 0.968825.90 / 0.9531
Bicubic28.86 / 0.977328.83 / 0.977128.64 / 0.977028.53 / 0.976627.57 / 0.971926.65 / 0.966825.79 / 0.9477
28.84 / 0.977228.84 / 0.977228.69 / 0.977128.76 / 0.977127.82 / 0.973826.71 / 0.967825.94 / 0.9533
Lanczos228.86 / 0.977328.83 / 0.977028.63 / 0.976928.52 / 0.976627.56 / 0.971926.65 / 0.966825.78 / 0.9476
28.83 / 0.977228.84 / 0.977128.67 / 0.977028.74 / 0.977327.81 / 0.973926.70 / 0.967825.93 / 0.9533
Lanczos328.86 / 0.977328.84 / 0.977028.68 / 0.976828.55 / 0.976527.56 / 0.971926.56 / 0.965225.72 / 0.9468
28.85 / 0.977228.85 / 0.977328.84 / 0.977128.85 / 0.977227.84 / 0.974026.58 / 0.966525.89 / 0.9528
Time (ms)11.911.512.511.712.011.712.3
12.812.912.913.012.212.612.5
", + "bbox": [ + 76, + 378, + 898, + 580 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "by [3] to demonstrate the robustness of our method, termed CNN-H. In addition to the above CNN models, we also train a transformer-based enhancer conditioned on Restormer, as proposed by [38], on the 4KIL dataset to further validate the scalability of our method. More details about pre-training are provided in the supplementary materials.", + "bbox": [ + 75, + 604, + 470, + 696 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. Collaborative Training", + "text_level": 1, + "bbox": [ + 76, + 709, + 290, + 726 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In this stage, we keep the pre-trained enhancers fixed and train to correlate the resizer and enhancer. Considering the randomness of scale variability in the resampling process, we incorporate random scale training. Specifically, we begin by cropping a $H_{p} \\times W_{p}$ patch from the UHD input image. Next, we randomly sample a scale factor $r$ from a uniform distribution $\\mathcal{U}(1,4)$ . Finally, we downsample the high-resolution $H_{p} \\times W_{p}$ patch by the scale factor $r$ to generate a low-resolution $h_{p} \\times w_{p}$ patch. The optimization objectives mentioned earlier are applied to these two resolution patches.", + "bbox": [ + 75, + 734, + 468, + 902 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/321516f72cba637bdb49fcd3bf3ba59c6b74f80bfc1fd1fb895bea603cf31ab9.jpg", + "table_caption": [ + "Table 4. Comparison of model parameters. Baseline models with LMAR are shown in gray, where the extra parameters are trainable parameters possessed by LMAR." + ], + "table_footnote": [], + "table_body": "
ModelCNN-ICNN-HRestormer
#Params1.009MB1.128MB1.660MB
1.314MB1.495MB1.965MB
", + "bbox": [ + 527, + 654, + 859, + 705 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Implementation Details. We conduct experiments based on the PyTorch framework along with the tiny-cuda-nn [27] library with one NVIDIA 3090 GPU. During the collaborative training process, the Adam [21] optimizer with $\\beta_{1} = 0.9$ and $\\beta_{2} = 0.999$ is used in our experiments. A batch size of 1 is utilized for all experiments. The cropped patch size $(H_{p} \\times W_{p})$ is set to $1024 \\times 1024$ for the CNN enhancers, and $768 \\times 768$ for the transformer enhancer. For the CNN-I and CNN-H, the collaborative training epoch is set to 12 and 24, respectively. The initial learning rate is $4e^{-4}$ , which decays by a factor value of 0.75 every 4 epochs and 8 epochs for the CNN-I and CNN-H enhancers. As for", + "bbox": [ + 496, + 719, + 893, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "25728", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/cf7df67d306a02489cf38f1f11027c5268559e0eb87dbf8eaf3e7cb9b339103a.jpg", + "image_caption": [ + "(a) Input" + ], + "image_footnote": [], + "bbox": [ + 80, + 88, + 176, + 227 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/8003faa17f193bf480bfd2eb50764ec53f8d809021b577077222fd50af8ea5f1.jpg", + "image_caption": [ + "(b) $\\times 5$" + ], + "image_footnote": [], + "bbox": [ + 181, + 88, + 277, + 228 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/fc46698ed5f88479ff83cb874e01e3c43b2b24584b29f55951d74a7f51413217.jpg", + "image_caption": [ + "(c)×4" + ], + "image_footnote": [], + "bbox": [ + 282, + 88, + 380, + 228 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/fe6f5a25ef97440ebef75c8ef255903de2ebc32af2cc0db0d644274d3f96cb8a.jpg", + "image_caption": [ + "(d)×3", + "Figure 8. Qualitative results on the 4KIL dataset of different scales. The top row is obtained from the lanzcos3 operator without LMAR, while the bottom row is with LMAR. Please zoom in for details." + ], + "image_footnote": [], + "bbox": [ + 385, + 88, + 482, + 228 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/e832857fd3e6fe0638274e575c4e26ac7c16315857300943311848a7e13b667f.jpg", + "image_caption": [ + "(e)×2" + ], + "image_footnote": [], + "bbox": [ + 486, + 88, + 584, + 228 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/d95033fe1b9124f124285d799ef4a1d875c09e97ebae88c15244f2e4bf3549f6.jpg", + "image_caption": [ + "(f) $\\times 1.5$" + ], + "image_footnote": [], + "bbox": [ + 589, + 88, + 687, + 228 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/a6720055a11619e8d182c78e8f4135edec30d3a85ffe6b9e772710176f5a3be7.jpg", + "image_caption": [ + "(g)×1" + ], + "image_footnote": [], + "bbox": [ + 692, + 88, + 789, + 228 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/48d6835db7bd1a3fdc5c394d18204752e7918d0972153b5833127b49be6aff88.jpg", + "image_caption": [ + "(h) GT" + ], + "image_footnote": [], + "bbox": [ + 792, + 88, + 890, + 228 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "the transformer enhancer, the number of collaborative training epochs is 60, with a $4e^{-4}$ initial learning rate, which decays by a factor of 0.75 every 20 epochs. The discriminator is composed of several convolution layers, whose initial learning rate is $2e^{-4}$ and shares the same training strategies as mentioned above.", + "bbox": [ + 75, + 295, + 467, + 385 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.3. Performance Comparison", + "text_level": 1, + "bbox": [ + 76, + 393, + 310, + 409 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Quantitative Evaluation. We utilize two metrics, the Peak Signal to Noise Ratio (PSNR, dB), and the Structural Similarity (SSIM), to assess the performance gains. To verify the universality of our methods, we adopt five types of resizers and test them on various commonly used scales, including both in-distribution and out-of-distribution scales. As shown in Tables 1, 2, and 3, our method consistently achieves performance improvements across different resolutions, regardless of the choice of resizer or the category of the backbone. Notably, the performance with LMAR at UHD resolution is analogous to the original UHD input, which illustrates the superior continuous modeling capability of LMAR. Additionally, our approach maintains comparable effectiveness with the baselines, where a minor increase in parameters, as shown in Table 4. More importantly, our algorithm only needs to be trained on one type of resizer, which can then be applied to any other resizers.", + "bbox": [ + 75, + 416, + 467, + 672 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Qualitative Evaluation. Figures 7 and 8 present the visual comparison on the UHD-LOL4K dataset and 4KIL dataset respectively. These comparisons demonstrate that enabling model-aware resampling leads to enhanced results with improved global lightness and local structural details, across various resampling scales. It verifies that coupling the resize-ers and enhancers could generate better visual results. More visual results are presented in the supplementary materials.", + "bbox": [ + 75, + 672, + 467, + 792 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4. Ablation Studies", + "text_level": 1, + "bbox": [ + 76, + 801, + 240, + 815 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We conduct ablation studies on the UHD-LOL4K dataset to investigate the rationality of our designs. The ablation results are shown in Table 5. First, we replace the bottleneck features (Full) with the relatively front features $(F_{f})$ and relatively end features $(F_{e})$ to apply feature constraints,", + "bbox": [ + 75, + 824, + 467, + 898 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where remarkable performance drops can be observed at 4 times downsampling. The findings indicate that the bottleneck features exhibit the most representative capability in guiding model-aware resampling. Second, we replace the discriminator (Full) with the perceptual loss (VGG) to conduct the distribution restriction. The utilization of discriminator encourages more consistent feature alignment. Lastly, even with sub-optimal constraints, our method outperforms the baseline, emphasizing the effectiveness of coupling resizers and enhancers.", + "bbox": [ + 496, + 295, + 890, + 445 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/8f33bfd143f070a8bdb6e8c9e5641285ea70daab5c6164d6322289cd4e12d7ff.jpg", + "table_caption": [ + "Table 5. Ablation studies on the constraints employed in our designs. We present the PSNR results here." + ], + "table_footnote": [], + "table_body": "
ScalesBaselineFfFeVGGFull
x229.8229.9830.0530.0330.20
x425.4025.8625.5025.4726.36
", + "bbox": [ + 511, + 497, + 877, + 547 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 565, + 617, + 580 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we introduce a novel approach for effectively enhancing the UHD images at different downsampling scales by a model-aware resampling strategy. Our method customizes resampling under the guidance of model knowledge, integrating the resizers and enhancers through compensatory information learning. Notably, our algorithm is fully compatible with existing interpolation resamplers and maintains the efficiency. Moreover, our design promotes performances at low resolutions remarkably while keeping comparable results at UHD resolution, without the need to retrain the enhancer. Extensive experiments have validated the effectiveness and scalability of our method.", + "bbox": [ + 496, + 590, + 890, + 771 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Acknowledgment", + "text_level": 1, + "bbox": [ + 500, + 784, + 669, + 801 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This work was supported by the JKW Research Funds under Grant 20-163-14-LZ-001-004-01, and the Anhui Provincial Natural Science Foundation under Grant 2108085UD12. We acknowledge the support of GPU cluster built by MCC Lab of Information Science and Technology Institution, USTC.", + "bbox": [ + 496, + 809, + 890, + 898 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "25729", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Kenneth R Castleman. Digital image processing. Prentice Hall Press, 1996. 3", + "[2] Jiawen Chen, Andrew Adams, Neal Wadhwa, and Samuel W Hasinoff. Bilateral guided upsampling. ACM Transactions on Graphics (TOG), 35(6):1-8, 2016. 3", + "[3] Liangyu Chen, Xin Lu, Jie Zhang, Xiaojie Chu, and Chengpeng Chen. Hinet: Half instance normalization network for image restoration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, pages 182-192, 2021. 7", + "[4] Yinbo Chen, Sifei Liu, and Xiaolong Wang. Learning continuous image representation with local implicit image function. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8628-8638, 2021. 3, 4", + "[5] Weitao Feng, Deyi Ji, Yiru Wang, Shuorong Chang, Hansheng Ren, and Weihao Gan. Challenges on large scale surveillance video analysis. In Proceedings of the IEEE conference on computer vision and pattern recognition workshops, pages 69-76, 2018. 3", + "[6] Michael Gharbi, Jiawen Chen, Jonathan T Barron, Samuel W Hasinoff, and Frédo Durand. Deep bilateral learning for real-time image enhancement. ACM Transactions on Graphics (TOG), 36(4):1-12, 2017. 3", + "[7] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial networks. Communications of the ACM, 63(11):139-144, 2020. 6", + "[8] Dianyuan Han. Comparison of commonly used image interpolation methods. In Conference of the 2nd International Conference on Computer Science and Electronics Engineering (ICCSEE 2013), pages 1556-1559. Atlantis Press, 2013. 3", + "[9] Muhammad Haris, Gregory Shakhnarovich, and Norimichi Ukita. Deep back-projection networks for super-resolution. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1664-1673, 2018. 4", + "[10] Hanzhe Hu, Deyi Ji, Weihao Gan, Shuai Bai, Wei Wu, and Junjie Yan. Class-wise dynamic graph convolution for semantic segmentation. In European Conference on Computer Vision, pages 1-17. Springer, 2020. 3", + "[11] Xuecai Hu, Haoyuan Mu, Xiangyu Zhang, Zilei Wang, Tieniu Tan, and Jian Sun. Meta-sr: A magnification-arbitrary network for super-resolution. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1575-1584, 2019. 3", + "[12] Max Jaderberg, Karen Simonyan, Andrew Zisserman, et al. Spatial transformer networks. Advances in neural information processing systems, 28, 2015. 2, 3", + "[13] Deyi Ji, Hongtao Lu, and Tongzhen Zhang. End to end multiscale convolutional neural network for crowd counting. In Eleventh international conference on machine vision, pages 761-766. SPIE, 2019. 3", + "[14] Deyi Ji, Haoran Wang, Hanzhe Hu, Weihao Gan, Wei Wu, and Junjie Yan. Context-aware graph convolution network" + ], + "bbox": [ + 78, + 114, + 468, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "for target re-identification. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1646–1654, 2021.", + "[15] Deyi Ji, Haoran Wang, Mingyuan Tao, Jianqiang Huang, Xian-Sheng Hua, and Hongtao Lu. Structural and statistical texture knowledge distillation for semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16876-16885, 2022.", + "[16] Deyi Ji, Siqi Gao, Mingyuan Tao, Hongtao Lu, and Feng Zhao. Changenet: Multi-temporal asymmetric change detection dataset. arXiv preprint arXiv:2312.17428, 2023.", + "[17] Deyi Ji, Feng Zhao, and Hongtao Lu. Guided patch-grouping wavelet transformer with spatial congruence for ultra-high resolution segmentation. International Joint Conference on Artificial Intelligence, 2023.", + "[18] Deyi Ji, Feng Zhao, Hongtao Lu, Mingyuan Tao, and Jieping Ye. Ultra-high resolution segmentation with ultra-rich context: A novel benchmark. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23621-23630, 2023.", + "[19] Deyi Ji, Siqi Gao, Lanyun Zhu, Yiru Zhao, Peng Xu, Hong-tao Lu, and Feng Zhao. View-centric multi-object tracking with homographic matching in moving UAV. arXiv preprint, 2024. 3", + "[20] Heewon Kim, Myungsub Choi, Bee Lim, and Kyoung Mu Lee. Task-aware image downscaling. In Proceedings of the European conference on computer vision (ECCV), pages 399-414, 2018. 2, 3", + "[21] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014.7", + "[22] Yue Li, Dong Liu, Houqiang Li, Li Li, Zhu Li, and Feng Wu. Learning a convolutional neural network for image compact-resolution. IEEE Transactions on Image Processing, 28(3): 1092-1107, 2018. 2", + "[23] Jie Liang, Hui Zeng, and Lei Zhang. High-resolution photorealistic image translation in real-time: A laplacian pyramid translation network. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9392-9400, 2021. 3", + "[24] Qiaowanni Lin, Zhuoran Zheng, and Xiuyi Jia. Uhd low-light image enhancement via interpretable bilateral learning. Information Sciences, 2022. 6", + "[25] Yang Liu, Zhenyue Qin, Saeed Anwar, Pan Ji, Dongwoo Kim, Sabrina Caldwell, and Tom Gedeon. Invertible denoising network: A light solution for real noise removal. arXiv preprint arXiv:2104.10546, 2021. 6", + "[26] Don P Mitchell and Arun N Netravali. Reconstruction filters in computer-graphics. ACM Siggraph Computer Graphics, 22(4):221-228, 1988. 3", + "[27] Thomas Müller. tiny-cuda-nn, 2021. 7", + "[28] Adria Recasens, Petr Kellnhofer, Simon Stent, Wojciech Matusik, and Antonio Torralba. Learning to zoom: a saliency-based sampling layer for neural networks. In Proceedings of the European conference on computer vision (ECCV), pages 51-66, 2018. 2, 3", + "[29] Assaf Shocher, Ben Feinstein, Niv Haim, and Michal Irani. From discrete to continuous convolution layers. arXiv preprint arXiv:2006.11120, 2020.4" + ], + "bbox": [ + 503, + 92, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "25730", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[30] Wanjie Sun and Zhenzhong Chen. Learned image downscaling for upscaling using content adaptive resampler. IEEE Transactions on Image Processing, 29:4027-4040, 2020. 2, 3", + "[31] Haoran Wang, Licheng Jiao, Fang Liu, Lingling Li, Xu Liu, Deyi Ji, and Weihao Gan. Ipgn: Interactiveness proposal graph network for human-object interaction detection. IEEE Transactions on Image Processing, 30:6583-6593, 2021. 3", + "[32] Haoran Wang, Licheng Jiao, Fang Liu, Lingling Li, Xu Liu, Deyi Ji, and Weihao Gan. Learning social spatio-temporal relation graph in the wild and a video benchmark. IEEE Transactions on Neural Networks and Learning Systems, 2021. 3", + "[33] Tao Wang, Kaihao Zhang, Tianrun Shen, Wenhan Luo, Bjorn Stenger, and Tong Lu. Ultra-high-definition low-light image enhancement: A benchmark and transformer-based method. arXiv preprint arXiv:2212.11548, 2022. 3, 6", + "[34] Mingqing Xiao, Shuxin Zheng, Chang Liu, Yaolong Wang, Di He, Guolin Ke, Jiang Bian, Zhouchen Lin, and Tie-Yan Liu. Invertible image rescaling. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part I 16, pages 126-144. Springer, 2020. 3", + "[35] Jinbo Xing, Wenbo Hu, Menghan Xia, and Tien-Tsin Wong. Scale-arbitrary invertible image downscaling. IEEE Transactions on Image Processing, 2023. 3", + "[36] Wei Yu, Qi Zhu, Naishan Zheng, Jie Huang, Man Zhou, and Feng Zhao. Learning non-uniform-sampling for ultra-high-definition image enhancement. In Proceedings of the 31st ACM International Conference on Multimedia, pages 1412-1421, 2023. 3", + "[37] Sergey Zagoruyko and Nikos Komodakis. Paying more attention to attention: Improving the performance of convolutional neural networks via attention transfer. arXiv preprint arXiv:1612.03928, 2016. 5", + "[38] Syed Waqas Zamir, Aditya Arora, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, and Ming-Hsuan Yang. Restormer: Efficient transformer for high-resolution image restoration. In CVPR, 2022. 7", + "[39] Yongbing Zhang, Debin Zhao, Jian Zhang, Ruiqin Xiong, and Wen Gao. Interpolation-dependent image downsampling. IEEE Transactions on Image Processing, 20(11): 3291-3296, 2011. 2", + "[40] Zhuoran Zheng, Wenqi Ren, Xiaochun Cao, Xiaobin Hu, Tao Wang, Fenglong Song, and Xiuyi Jia. Ultra-high-definition image dehazing via multi-guided bilateral learning. In 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 16180-16189. IEEE, 2021. 3", + "[41] Zhuoran Zheng, Wenqi Ren, Xiaochun Cao, Tao Wang, and Xiuyi Jia. Ultra-high-definition image hdr reconstruction via collaborative bilateral learning. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4449-4458, 2021. 3", + "[42] Lanyun Zhu, Deyi Ji, Shiping Zhu, Weihao Gan, Wei Wu, and Junjie Yan. Learning statistical texture for semantic segmentation. In Proceedings of the IEEE/CVF Conference" + ], + "bbox": [ + 78, + 90, + 468, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "on Computer Vision and Pattern Recognition (CVPR), pages 12537-12546, 2021. 3", + "[43] Lanyun Zhu, Tianrun Chen, Deyi Ji, Jieping Ye, and Jun Liu. Llafs: When large-language models meet few-shot segmentation. arXiv preprint arXiv:2311.16926, 2023.", + "[44] Lanyun Zhu, Deyi Ji, Tianrun Chen, Peng Xu, Jieping Ye, and Jun Liu. Ibid: Alleviating hallucinations in large vision-language models via image-biased decoding. arXiv preprint arXiv:2402.18476, 2024.", + "[45] Qi Zhu, Zeyu Xiao, Jie Huang, and Feng Zhao. Dast-net: Depth-aware spatio-temporal network for video deblurring. In 2022 IEEE International Conference on Multimedia and Expo (ICME), pages 1-6. IEEE, 2022.", + "[46] Qi Zhu, Naishan Zheng, Jie Huang, Man Zhou, Jinghao Zhang, and Feng Zhao. Learning spatio-temporal sharpness map for video deblurring. IEEE Transactions on Circuits and Systems for Video Technology, 2023.", + "[47] Qi Zhu, Man Zhou, Naishan Zheng, Chongyi Li, Jie Huang, and Feng Zhao. Exploring temporal frequency spectrum in deep video deblurring. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12428-12437, 2023.", + "[48] Qi Zhu, Jie Huang, Naishan Zheng, Hongzhi Gao, Chongyi Li, Yuan Xu, Feng Zhao, et al. Fouridown: Factoring down-sampling into shuffling and superposing. Advances in Neural Information Processing Systems, 36, 2024. 3" + ], + "bbox": [ + 501, + 92, + 890, + 459 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "25731", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2024/Empowering Resampling Operation for Ultra-High-Definition Image Enhancement with Model-Aware Guidance/5e31bf7f-5e48-4030-90cc-b526f724e932_model.json b/2024/Empowering Resampling Operation for Ultra-High-Definition Image Enhancement with Model-Aware Guidance/5e31bf7f-5e48-4030-90cc-b526f724e932_model.json new file mode 100644 index 0000000000000000000000000000000000000000..2bbd0a58fd3aa1943545ad1f669e9d3eb30f78c6 --- /dev/null +++ b/2024/Empowering Resampling Operation for Ultra-High-Definition Image Enhancement with Model-Aware Guidance/5e31bf7f-5e48-4030-90cc-b526f724e932_model.json @@ -0,0 +1,2981 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.139, + 0.131, + 0.833, + 0.175 + ], + "angle": 0, + "content": "Empowering Resampling Operation for Ultra-High-Definition Image Enhancement with Model-Aware Guidance" + }, + { + "type": "text", + "bbox": [ + 0.205, + 0.204, + 0.812, + 0.24 + ], + "angle": 0, + "content": "Wei Yu,\\* Jie Huang,\\* Bing Li, Kaiwen Zheng, Qi Zhu, Man Zhou, Feng Zhao† \nUniversity of Science and Technology of China" + }, + { + "type": "text", + "bbox": [ + 0.088, + 0.242, + 0.918, + 0.259 + ], + "angle": 0, + "content": "{patrick914y, hj0117,bing0123, kezh,zqcrafts, manman}@mail.ustc.edu.cn，fzhao956@ustc.edu.cn" + }, + { + "type": "image", + "bbox": [ + 0.08, + 0.286, + 0.185, + 0.352 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.184, + 0.287, + 0.302, + 0.352 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.304, + 0.287, + 0.409, + 0.352 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.177, + 0.355, + 0.308, + 0.366 + ], + "angle": 0, + "content": "(a) Down-Enhancement-Up" + }, + { + "type": "image", + "bbox": [ + 0.081, + 0.368, + 0.409, + 0.452 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.223, + 0.453, + 0.264, + 0.463 + ], + "angle": 0, + "content": "(b) Ours" + }, + { + "type": "image", + "bbox": [ + 0.429, + 0.287, + 0.534, + 0.351 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.445, + 0.352, + 0.507, + 0.362 + ], + "angle": 0, + "content": "PSNR: 25.59" + }, + { + "type": "image", + "bbox": [ + 0.546, + 0.287, + 0.651, + 0.351 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.562, + 0.352, + 0.622, + 0.362 + ], + "angle": 0, + "content": "PSNR: 28.51" + }, + { + "type": "image", + "bbox": [ + 0.662, + 0.287, + 0.766, + 0.351 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.681, + 0.352, + 0.743, + 0.361 + ], + "angle": 0, + "content": "PSNR: 30.78" + }, + { + "type": "image", + "bbox": [ + 0.778, + 0.287, + 0.883, + 0.352 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.795, + 0.352, + 0.856, + 0.361 + ], + "angle": 0, + "content": "PSNR: 31.05" + }, + { + "type": "image", + "bbox": [ + 0.429, + 0.365, + 0.892, + 0.452 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.445, + 0.453, + 0.507, + 0.463 + ], + "angle": 0, + "content": "PSNR: 28.09" + }, + { + "type": "image_caption", + "bbox": [ + 0.562, + 0.453, + 0.624, + 0.463 + ], + "angle": 0, + "content": "PSNR: 29.56" + }, + { + "type": "image_caption", + "bbox": [ + 0.681, + 0.453, + 0.743, + 0.463 + ], + "angle": 0, + "content": "PSNR: 32.24" + }, + { + "type": "image_caption", + "bbox": [ + 0.795, + 0.453, + 0.856, + 0.463 + ], + "angle": 0, + "content": "PSNR: 33.16" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.474, + 0.894, + 0.517 + ], + "angle": 0, + "content": "Figure 1. Comparison between previous Downsampling-Enhancement-Upsampling paradigm and our proposed method. (a) Previous works treat resampling operators and inner models as separate components, while (b) our method integrates them via empowering model-aware resampling. As shown on the right side, our approach achieves significant performance gains over various resampling scales." + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.537, + 0.314, + 0.552 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.571, + 0.473, + 0.859 + ], + "angle": 0, + "content": "Image enhancement algorithms have made remarkable advancements in recent years, but directly applying them to Ultra-high-definition (UHD) images presents intractable computational overheads. Therefore, previous straightforward solutions employ resampling techniques to reduce the resolution by adopting a \"Downsampling-Enhancement-Upsampling\" processing paradigm. However, this paradigm disentangles the resampling operators and inner enhancement algorithms, which results in the loss of information that is favored by the model, further leading to sub-optimal outcomes. In this paper, we propose a novel method of Learning Model-Aware Resampling (LMAR), which learns to customize resampling by extracting model-aware information from the UHD input image, under the guidance of model knowledge. Specifically, our method consists of two core designs, namely compensatory kernel estimation and steganographic resampling. At the first stage, we dynamically predict compensatory kernels tailored to the specific input and resampling scales. At" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.539, + 0.895, + 0.735 + ], + "angle": 0, + "content": "the second stage, the image-wise compensatory information is derived with the compensatory kernels and embedded into the rescaled input images. This promotes the representation of the newly derived downscaled inputs to be more consistent with the full-resolution UHD inputs, as perceived by the model. Our LMAR enables model-aware and model-favored resampling while maintaining compatibility with existing resampling operators. Extensive experiments on multiple UHD image enhancement datasets and different backbones have shown consistent performance gains after correlating resizer and enhancer, e.g., up to 1.2dB PSNR gain for \\(\\times 1.8\\) resampling scale on UHD-LOL4K. The code is available at https://github.com/YPPatrickW/LMAR." + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.74, + 0.633, + 0.756 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.766, + 0.892, + 0.842 + ], + "angle": 0, + "content": "Over time, learning-based image enhancement algorithms have achieved progressive performance improvements. However, the majority of existing methods are not compatible with UHD images due to the heavy computational burdens imposed by their megapixel counts." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.842, + 0.894, + 0.903 + ], + "angle": 0, + "content": "To address this, previous works have sought to alleviate computation overheads by employing downsampling to reduce the resolution, as shown in Figure 1. Concretely, UHD input images are often resampled to smaller sizes for" + }, + { + "type": "page_footnote", + "bbox": [ + 0.094, + 0.876, + 0.361, + 0.889 + ], + "angle": 0, + "content": "*Both authors contributed equally to this research." + }, + { + "type": "page_footnote", + "bbox": [ + 0.098, + 0.889, + 0.222, + 0.902 + ], + "angle": 0, + "content": "† Corresponding author." + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.876, + 0.361, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "25722" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.08, + 0.089, + 0.473, + 0.234 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.245, + 0.47, + 0.384 + ], + "angle": 0, + "content": "Figure 2. The influence of the front-end resampler. We choose the cubic resampling operator as the baseline and introduce perturbations to its kernel function. Interestingly, manipulating the resampling function can lead to volatile outcomes, with the image quality either being improved or degraded. We also present two more comparisons: the resampling residual of the UHD input image and the feature maps generated by different resizers. It is indicated that the response of the model is highly sensitive to the sampling strategies. This highlights the importance of customizing the resampling process in a model-aware manner." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.398, + 0.469, + 0.474 + ], + "angle": 0, + "content": "enhancement, and then subsequently resampled back to the original resolution[22, 39]. Surprisingly, the majority of research efforts have been paid to developing inner processing algorithms, with little attention devoted to the foremost and endmost resampling operators." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.476, + 0.469, + 0.642 + ], + "angle": 0, + "content": "Recently, a line of work has proposed the replacement of the conventional interpolation-based resampling operators (e.g., nearest, bilinear, bicubic) with learnable resampling operators [12, 20, 28, 30]. Attributing to the superior modeling capability of deep neural networks, these learnable resampling operators have demonstrated remarkable performance gains. However, the practical application of these learnable resampling operators is hindered by their high computational complexity and reduced efficiency, resulting in a continued preference for interpolation-based resampling operators." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.643, + 0.469, + 0.839 + ], + "angle": 0, + "content": "Revisiting the above UHD image enhancement pipeline: Downsampling-Enhancement-Upsampling, with the internal enhancement algorithm remaining unchanged, a fundamental question arises: 'How significant is the impact of these two resampling operations?' The answer to this question is twofold. On the one hand, the quality of the upsampled enhanced results exhibits a decreasing trend as the degree of downsampling aggravates. On the other hand, manipulating the interpolation kernel functions yields a diverse range of outcomes, as illustrated in Figure 2. We attribute this phenomenon to the isolation between the resampling operators and the enhancement algorithm, where both components are disentangled from each other." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.469, + 0.9 + ], + "angle": 0, + "content": "To this end, the focus of this paper lies in correlating the resampling operators within the internal enhancement algorithms and unleashing the potential of the resampling operation to make benefits for the aforementioned framework," + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.892, + 0.168 + ], + "angle": 0, + "content": "without modifying the inner processing methods (e.g., pretrained networks). To accomplish this goal, we propose a novel method termed Learning Model-Aware Resampling (LMAR). LMAR correlates the resizer and the enhancer through a compensatory information learning process." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.195, + 0.893, + 0.587 + ], + "angle": 0, + "content": "We begin by delving into the impact of resampling from the viewpoint of the enhancer, where the resampling operations directly undermine the representation ability of intermediate features. Therefore, the intermediate features of the UHD image offer valuable insights on how to perform resampling in a model-aware manner. Building upon this recognition, we formulate our method into two steps: Compensatory Kernel Estimation (CKE) and Steganographic Resampling (SR). In the CKE step, we generate imagespecific convolution kernels for the input UHD image via implicit neural representation. These kernels are specifically customized for each input and resampling scale. Then, we can obtain compensatory information produced by these predicted filters. Subsequently, in the SR step, the compensatory information is embedded in the downscaled input image, resulting in a newly learnable downscaled input. By utilizing the compensatory information as a medium, we establish the correlation between the resampling operators and the inner enhancer. Finally, the learnable input is passed to the enhancer and optimized to encourage that the intermediate features align more consistently with those of the UHD image. Thanks to these designs, our method not only maintains compatibility with existing interpolation-based resampling operators but also incorporates them into a part of enhancement algorithms under the guidance of model knowledge. Our contributions are summarized as:" + }, + { + "type": "text", + "bbox": [ + 0.501, + 0.614, + 0.892, + 0.658 + ], + "angle": 0, + "content": "- We rethink the disentanglement between the resizer and enhancement model and propose to perform the resampling in a model-aware manner." + }, + { + "type": "text", + "bbox": [ + 0.501, + 0.66, + 0.892, + 0.735 + ], + "angle": 0, + "content": "- We introduce LMAR, a novel method that facilitates the collaboration between resizer and enhancer via customizing resampling under the guidance of model knowledge, where the representative capabilities of low-resolution images are largely improved." + }, + { + "type": "text", + "bbox": [ + 0.501, + 0.736, + 0.892, + 0.81 + ], + "angle": 0, + "content": "- Our LMAR delivers two core designs: compensatory kernel estimation and steganographic resampling, which models the interplay between responder and enhancer by learning embeddable information. Our method is compatible with any interpolation resamplers." + }, + { + "type": "text", + "bbox": [ + 0.501, + 0.811, + 0.892, + 0.9 + ], + "angle": 0, + "content": "- Extensive experiments are conducted to verify the superiority of correlating resizer and enhancer. Remarkably, our algorithm significantly improves the performance at low resolutions while maintaining equivalent results at the original resolution, without requiring retraining of the enhancer." + }, + { + "type": "list", + "bbox": [ + 0.501, + 0.614, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "25723" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.08, + 0.087, + 0.473, + 0.252 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.261, + 0.47, + 0.371 + ], + "angle": 0, + "content": "Figure 3. The inspiration for our method design. We replace the compressed bottleneck features of the downsampled input with features extracted from the UHD image (The green arrow). Large-margin performance gains can be achieved compared with the vested paradigm (The gray arrow). This observation highlights the potential role of feature consistency between UHD input and any downscaled input, which can play as a bridge for establishing correlations between resampling operators and enhancers." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.383, + 0.219, + 0.399 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.408, + 0.257, + 0.425 + ], + "angle": 0, + "content": "2.1. Image Resampling" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.431, + 0.47, + 0.506 + ], + "angle": 0, + "content": "Image resampling has emerged as one of the most commonly employed techniques for altering the resolution of an image. The key to image resampling lies in effectively preserving the quality of image content while maintaining efficiency." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.507, + 0.469, + 0.597 + ], + "angle": 0, + "content": "Interpolation operators, such as bilinear, and bicubic [1, 8], have been the popular choice for image resampling for many years. These operators are capable of effectively resampling images at any scale, yet they disregard the spatial variations of distinct image patches, leading to inadequate preservation of the local structure [26]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.598, + 0.469, + 0.733 + ], + "angle": 0, + "content": "In recent years, deep neural networks have facilitated advancements in various tasks [5, 10, 13-19, 31, 32, 36, 42-48], including image resampling [4, 11, 12, 20, 28, 30, 34, 35]. Task-aware image downsampling (TAD) [20] presents an encoder-decoder framework that aims to enhance the quality of both downscaling and upscaling processes in a collaborative manner. Though these methods achieve excellent results, none of them focus on correlating interpolation resampler with a pre-trained model." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.742, + 0.294, + 0.759 + ], + "angle": 0, + "content": "2.2. UHD Image Processing" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.469, + 0.901 + ], + "angle": 0, + "content": "A line of studies that specifically focus on high-resolution image processing has been proposed [2, 6, 23, 33, 40, 41]. These methods share a common paradigm: learning from downscaled images to alleviate the computational burdens. In this learning paradigm, the quality of the results is directly determined by the resampled image representation, highlighting the crucial role of resampling strategies. However, they often overlook the impact of resampling approaches and lack in-depth investigation of various resam" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.152 + ], + "angle": 0, + "content": "pling operators. In this work, our objective is to explore the interplay between enhancer and resampling operators and establish a collaborative relationship for them to improve performance." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.164, + 0.635, + 0.181 + ], + "angle": 0, + "content": "3. Methodology" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.189, + 0.892, + 0.31 + ], + "angle": 0, + "content": "In this section, we will present a comprehensive illustration of our proposed LMAR. We begin by introducing the motivation behind our approach and describing our settings. Subsequently, we will overview the whole processing pipeline and delve into the design of the Compensatory Kernel Estimation (CKE) and Steganographic Resampling (SR). Finally, we will discuss the optimization strategies employed to enable model-aware resampling." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.317, + 0.714, + 0.333 + ], + "angle": 0, + "content": "3.1. Setting and Motivation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.34, + 0.892, + 0.46 + ], + "angle": 0, + "content": "Our setting still follows the Downsampling-Enhancement-Upsampling paradigm, employing the interpolation-based resampling operators instead of introducing additional burdens with learnable resampling operators. Notably, we will first train an enhancement model for each dataset in our experimental setup and keep unchanged. The main difference lies in the collaboration process between the resizing and enhancer." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.462, + 0.892, + 0.627 + ], + "angle": 0, + "content": "The main inspiration stems from observing the impact of resampling on the representative feature extraction process of the enhancer. As shown in Figure 3, we inject the full-resolution bottleneck features into the enhancement process of the downsampled input images, where the final results significantly outperform the vested paradigm. It proves that front-most resampling degrades the representative ability of bottleneck features and impacts the following processes. Based on the insight, we correlate the resizer and enhancer through a compensatory information-learning process to achieve consistent feature alignment." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.635, + 0.611, + 0.649 + ], + "angle": 0, + "content": "3.2. Overview" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.658, + 0.892, + 0.688 + ], + "angle": 0, + "content": "The Downsampling-Enhancement-Upsampling processing can be mathematically formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.627, + 0.695, + 0.756, + 0.711 + ], + "angle": 0, + "content": "\\[\nx _ {d} = D (x; [ h, w ]),\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.636, + 0.714, + 0.891, + 0.73 + ], + "angle": 0, + "content": "\\[\n\\tilde {y} = F \\left(x _ {d}\\right), \\tag {1}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.636, + 0.733, + 0.765, + 0.75 + ], + "angle": 0, + "content": "\\[\ny = U (\\tilde {y}; [ H, W ]),\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.756, + 0.892, + 0.862 + ], + "angle": 0, + "content": "where \\(D(\\cdot)\\) is the down sampler that resamples the UHD input image \\(x \\in \\mathbb{R}^{H \\times W \\times 3}\\) to lower resolution of \\(x_{d} \\in \\mathbb{R}^{h \\times w \\times 3}\\), \\(F(\\cdot)\\) is the pre-trained enhancement model, and \\(U(\\cdot)\\) denotes the upsampler which resamples the low-resolution enhanced result \\(\\tilde{y}\\) back to the UHD resolution of \\(H \\times W\\), termed as \\(y\\). In LMAR, we rebuild the above paradigm, given as:" + }, + { + "type": "equation", + "bbox": [ + 0.603, + 0.868, + 0.891, + 0.892 + ], + "angle": 0, + "content": "\\[\n\\tilde {y} = F \\left(x _ {d}, \\Theta (x; D, U, F)\\right), \\tag {2}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.606, + 0.89, + 0.732, + 0.904 + ], + "angle": 0, + "content": "\\[\ny = U (\\tilde {y}; [ H, W ]).\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "25724" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.094, + 0.096, + 0.2, + 0.11 + ], + "angle": 0, + "content": "1. Training Stage" + }, + { + "type": "image", + "bbox": [ + 0.093, + 0.113, + 0.534, + 0.254 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.32, + 0.256, + 0.338, + 0.27 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image_caption", + "bbox": [ + 0.549, + 0.097, + 0.669, + 0.11 + ], + "angle": 0, + "content": "2. Inference Stage" + }, + { + "type": "image", + "bbox": [ + 0.549, + 0.113, + 0.887, + 0.248 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.711, + 0.254, + 0.73, + 0.267 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.272, + 0.681, + 0.424 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.684, + 0.273, + 0.891, + 0.424 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.437, + 0.894, + 0.507 + ], + "angle": 0, + "content": "Figure 4. Overview of the proposed LMAR. The sub-graph (a) depicts the training phase of LMAR, which encourages the compensated low-resolution input to maintain representation consistency with the full-resolution UHD input, as perceived by the enhancer. The subgraph (b) demonstrates the inference pipeline of our LMAR, where the compensated low-resolution input is directly fed into the enhancer and then upsampled to the UHD result. The sub-graph (c) illustrates how LMAR works, where the core lies in estimating compensatory kernels under the guidance of model knowledge to make up for the resampling process." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.518, + 0.471, + 0.61 + ], + "angle": 0, + "content": "Here, we introduce additional compensatory information as a medium to facilitate model-aware resampling, denoted as \\(\\Theta (x;D,U,F)\\). Our method is presented in Figure 4, where the model-aware resampling is achieved via two core designs, compensatory kernel estimation, and steganographic resampling." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.63, + 0.371, + 0.646 + ], + "angle": 0, + "content": "3.3. Compensatory Kernel Estimation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.658, + 0.47, + 0.825 + ], + "angle": 0, + "content": "Interpolation-based image resampling process can be decomposed as projected grid calculation and weighted pixel aggregation [29]. To this end, we estimate the compensatory kernel using a representation that is related to the scale. Inspired by LIFF [4], we employ two scale features, the relative coordinate grid and the pixel cell. The relative coordinate grid depicts the pixel location shift in the resampling processing. We first compute the uniform coordinate of the downsampling scale \\( C_d \\in \\mathbb{R}^{h \\times w \\times 2} \\) and the UHD scale \\( C_u \\in \\mathbb{R}^{H \\times W \\times 2} \\), which are normalized between \\([-1, 1]\\). The calculations of the above grids are given as:" + }, + { + "type": "equation", + "bbox": [ + 0.12, + 0.843, + 0.47, + 0.88 + ], + "angle": 0, + "content": "\\[\nC _ {d} (i, j) = (- 1 + \\frac {(2 i + 1)}{h}, - 1 + \\frac {(2 j + 1)}{w}), \\tag {3}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.122, + 0.875, + 0.425, + 0.906 + ], + "angle": 0, + "content": "\\[\nC _ {u} (i, j) = (- 1 + \\frac {(2 i + 1)}{H}, - 1 + \\frac {(2 j + 1)}{W}),\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.517, + 0.892, + 0.594 + ], + "angle": 0, + "content": "where \\( i \\in [0, h - 1] \\vee [0, H - 1], j \\in [0, w - 1] \\vee [0, W - 1] \\) are the position indexes for the height and width dimensions. As the sizes of the two coordinates are different, we first project them \\( C_d \\) to \\( C_u \\) and obtain the relative grid shift at the downsampling scale, which is expressed as:" + }, + { + "type": "equation", + "bbox": [ + 0.591, + 0.599, + 0.78, + 0.617 + ], + "angle": 0, + "content": "\\[\n\\tilde {C} _ {d} = \\operatorname {g r i d} \\text {- s a m p l e} \\left(C _ {d}, C _ {u}\\right),\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.593, + 0.62, + 0.892, + 0.639 + ], + "angle": 0, + "content": "\\[\n\\tilde {C} _ {r} = C _ {u} - \\tilde {C} _ {d}, \\tag {4}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.593, + 0.642, + 0.8, + 0.66 + ], + "angle": 0, + "content": "\\[\nC _ {r} = \\left(\\tilde {C} _ {r} (i) \\times h, \\tilde {C} _ {r} (j) \\times w\\right),\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.665, + 0.892, + 0.755 + ], + "angle": 0, + "content": "where the grid_sample is a remapping function that remaps the \\( C_d \\) into \\( \\tilde{C}_d \\) of size \\( H \\times W \\), \\( \\tilde{C}_r \\) represents the relative shift and \\( C_r \\) denotes the relative coordinate grid measured at downsampling scale. The pixel cell represents the ratio of pixel area changes between downsampling and upsampling, denoted as:" + }, + { + "type": "equation", + "bbox": [ + 0.601, + 0.759, + 0.892, + 0.79 + ], + "angle": 0, + "content": "\\[\nP _ {c} (i, j) = \\left(2 * \\frac {h}{H}, 2 * \\frac {w}{W}\\right), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.795, + 0.891, + 0.824 + ], + "angle": 0, + "content": "where \\(P_{c}\\in \\mathbb{R}^{H\\times W\\times 2}\\), and all pixel locations share the same scale ratio." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.826, + 0.892, + 0.901 + ], + "angle": 0, + "content": "The aforementioned scale-related representations are instance-same at a given scale, leading to the generation of identical kernels without considering the distinctions between images. Thus, we propose the utilization of backprojection [9] as a constraint to capture the interrelation" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "25725" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.081, + 0.089, + 0.203, + 0.22 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.121, + 0.223, + 0.16, + 0.236 + ], + "angle": 0, + "content": "(a) \\(x_{d}\\)" + }, + { + "type": "image", + "bbox": [ + 0.214, + 0.089, + 0.338, + 0.22 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.254, + 0.222, + 0.293, + 0.236 + ], + "angle": 0, + "content": "(b) \\(x_{c}\\)" + }, + { + "type": "image", + "bbox": [ + 0.347, + 0.089, + 0.471, + 0.22 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.365, + 0.222, + 0.448, + 0.236 + ], + "angle": 0, + "content": "(c) \\( |x_{c} - x_{d}| \\)" + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.254, + 0.47, + 0.325 + ], + "angle": 0, + "content": "Figure 5. Comparison of different downscaled input images. (a) The downscaled input images by the cubic resampling operator, (b) images derived from the cubic resampling operator in a model-aware manner, and (c) the residual between (a) and (b), where obvious differences can be witnessed in high-frequency areas." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.337, + 0.47, + 0.368 + ], + "angle": 0, + "content": "between each downscaled input and its corresponding UHD counterpart, formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.16, + 0.379, + 0.469, + 0.396 + ], + "angle": 0, + "content": "\\[\n\\tilde {x} = \\operatorname {C o n v} (x - U (x _ {d}; [ H, W ]))), \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.407, + 0.469, + 0.437 + ], + "angle": 0, + "content": "where \\(\\tilde{x} \\in \\mathbb{R}^{H \\times W \\times 3}\\) is the back-projection of UHD input image \\(x\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.439, + 0.469, + 0.575 + ], + "angle": 0, + "content": "Considering the pixel-wise aggregation nature of the resampling process, we model the generation of compensatory information by predicting the convolution kernels for the UHD input image instead of predicting pixel values for specific locations. Specifically, we flatten the relative coordinate \\( C_r \\), pixel cell \\( P_c \\), and back-projection \\( \\tilde{x} \\) then concatenate them. Afterward, we feed these concatenated features into a multi-layer perception (MLP) to predict compensatory kernels as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.186, + 0.586, + 0.469, + 0.604 + ], + "angle": 0, + "content": "\\[\nK = M L P _ {\\phi} \\left(\\left[ C _ {r}; P _ {c}; \\tilde {x} \\right]\\right), \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.615, + 0.469, + 0.705 + ], + "angle": 0, + "content": "where \\([\\cdot ]\\) denotes the channel-wise concatenation, and \\(K\\in\\) \\(\\mathbb{R}^{H\\times W\\times s}\\) represents the predicted compensatory kernels. The kernel for each pixel location is represented in the channel dimension \\(s\\), consisting of the multiplication of the input channel, out channel, and kernel size, with all being three in our case." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.716, + 0.33, + 0.733 + ], + "angle": 0, + "content": "3.4. Steganographic Resampling" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.739, + 0.469, + 0.785 + ], + "angle": 0, + "content": "With the estimated compensatory kernel from the above process, we calculate the compensatory information for the input UHD image as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.23, + 0.797, + 0.469, + 0.813 + ], + "angle": 0, + "content": "\\[\n\\widehat {x} = K \\otimes x, \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.824, + 0.469, + 0.856 + ], + "angle": 0, + "content": "where \\(\\widehat{x} \\in \\mathbb{R}^{H \\times W \\times 3}\\) is the generated UHD compensatory embedding, and \\(\\otimes\\) denotes the convolution operation." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.469, + 0.902 + ], + "angle": 0, + "content": "To further maintain the intrinsic characteristics of the downscaled input images, we downscale the compensatory information and perform compensatory steganography in" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.089, + 0.628, + 0.154 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.554, + 0.156, + 0.573, + 0.167 + ], + "angle": 0, + "content": "\\(x_{d}\\)" + }, + { + "type": "image", + "bbox": [ + 0.636, + 0.089, + 0.761, + 0.154 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.678, + 0.155, + 0.715, + 0.167 + ], + "angle": 0, + "content": "\\(E(x_{d})\\)" + }, + { + "type": "image", + "bbox": [ + 0.768, + 0.089, + 0.894, + 0.154 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.811, + 0.155, + 0.85, + 0.166 + ], + "angle": 0, + "content": "Errors" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.167, + 0.629, + 0.23 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.554, + 0.232, + 0.572, + 0.244 + ], + "angle": 0, + "content": "\\(x_{c}\\)" + }, + { + "type": "image", + "bbox": [ + 0.636, + 0.167, + 0.761, + 0.23 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.678, + 0.231, + 0.714, + 0.244 + ], + "angle": 0, + "content": "\\(E(x_{c})\\)" + }, + { + "type": "image", + "bbox": [ + 0.768, + 0.167, + 0.894, + 0.232 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.811, + 0.232, + 0.85, + 0.242 + ], + "angle": 0, + "content": "Errors" + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.26, + 0.894, + 0.344 + ], + "angle": 0, + "content": "Figure 6. Feature comparison of different downscaled input images. The second column represents the bottleneck features of the downscaled inputs. As observed in the error maps with the UHD features in the third column, downscaled input produced in a model-aware manner \\((x_{c})\\) maintains a higher level of representative consistency with the UHD input, as perceived by the model." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.356, + 0.817, + 0.372 + ], + "angle": 0, + "content": "the low-resolution space, which is computed as:" + }, + { + "type": "equation", + "bbox": [ + 0.588, + 0.383, + 0.891, + 0.401 + ], + "angle": 0, + "content": "\\[\nx _ {c} = \\operatorname {C o n v} ([ x _ {d}; D (\\widehat {x}; [ h, w ]) ]). \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.411, + 0.892, + 0.502 + ], + "angle": 0, + "content": "Here, we concatenate the downscaled UHD image and downscaled compensatory information, then perform channel reduction by convolution layer. In this way, the newly derived \\( x_{c} \\) can be optimized under the guidance of model knowledge, where we utilize the compensatory information to bridge the resizer and enhancer." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.511, + 0.695, + 0.528 + ], + "angle": 0, + "content": "3.5. Optimization Object" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.535, + 0.892, + 0.626 + ], + "angle": 0, + "content": "Given the pre-trained enhancer, the bottleneck features of the UHD input image possess the most powerful representational capability compared to all of its downscaled counterparts. Therefore, we encourage the consistency of bottleneck features, which is referred to as resampling consistency loss:" + }, + { + "type": "equation", + "bbox": [ + 0.568, + 0.637, + 0.891, + 0.654 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {r c} = \\operatorname {S m o o t h L 1} (E (x) - U (E \\left(x _ {c}\\right))), \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.665, + 0.891, + 0.741 + ], + "angle": 0, + "content": "where the \\( E(x) \\) and \\( E(x_{c}) \\) represent the corresponding bottleneck features of the UHD input \\( x \\) and learnable down-scaled input \\( x_{c} \\). \\( U(\\cdot) \\) is used for size adjustment. In addition, we introduce attention loss [37] for these two bottleneck features, given as:" + }, + { + "type": "equation", + "bbox": [ + 0.566, + 0.751, + 0.891, + 0.786 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {f a} = \\left\\| \\frac {E (x)}{\\| E (x) \\| _ {2}} - \\frac {U \\left(E \\left(x _ {c}\\right)\\right)}{\\| U \\left(E \\left(x _ {c}\\right)\\right) \\| _ {2}} \\right\\| _ {2}, \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.795, + 0.89, + 0.825 + ], + "angle": 0, + "content": "where \\(\\| \\cdot \\| _2\\) is the \\(\\mathcal{L}_2\\) normalization and this loss encourages the attention transfer from \\(E(x)\\) to \\(U(E(x_{c}))\\)." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.826, + 0.892, + 0.901 + ], + "angle": 0, + "content": "The aforementioned losses are designed to incorporate model-favored information into the learnable downscaled input image \\( x_{c} \\) but neglect the distribution consistency between the naive downscaled images \\( x_{d} \\) and newly derived images \\( x_{c} \\). Thus, we need to impose restrictions on the" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "25726" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.089, + 0.895, + 0.119 + ], + "angle": 0, + "content": "Table 1. Quantitative results on the UHD-LOL4K datasets with CNN-I as the backbone. The results with LMAR are shown in gray with better results highlighted in bold." + }, + { + "type": "table", + "bbox": [ + 0.078, + 0.128, + 0.899, + 0.328 + ], + "angle": 0, + "content": "
MethodCNN-I (UHD-LOL4K)
Scales(2160, 3840)(1440, 2560)(1080, 1920)(1200, 1600)(720, 1280)(540, 960)(432, 768)
Nearest31.92 / 0.984328.20 / 0.981527.01 / 0.979226.66 / 0.976325.86 / 0.976123.99 / 0.961523.06 / 0.9445
31.86 / 0.984128.46 / 0.982427.69 / 0.981327.42 / 0.979026.60 / 0.980624.38 / 0.971523.98 / 0.9608
Bilinear31.92 / 0.984331.07 / 0.983929.75 / 0.982829.16 / 0.982027.01 / 0.976725.90 / 0.974424.03 / 0.9448
31.88 / 0.984131.28 / 0.984030.17 / 0.983529.85 / 0.983227.94 / 0.981426.27 / 0.979524.85 / 0.9613
Bicubic31.92 / 0.984331.19 / 0.984029.82 / 0.983429.26 / 0.982227.16 / 0.976725.40 / 0.969424.04 / 0.9450
31.89 / 0.984231.42 / 0.984130.20 / 0.983929.96 / 0.983627.97 / 0.981526.36 / 0.979224.88 / 0.9617
Lanczos231.92 / 0.984331.04 / 0.984029.83 / 0.983429.22 / 0.981926.97 / 0.976125.39 / 0.969424.03 / 0.9450
31.88 / 0.984131.39 / 0.984130.19 / 0.983829.97 / 0.983427.90 / 0.981526.29 / 0.978924.87 / 0.9617
Lanczos331.92 / 0.984330.86 / 0.983829.08 / 0.981528.88 / 0.981327.10 / 0.976425.04 / 0.966323.86 / 0.9438
31.90 / 0.984231.59 / 0.984230.24 / 0.983630.13 / 0.983427.93 / 0.981725.79 / 0.971824.66 / 0.9611
Time (ms)9.79.613.313.414.815.415.0
10.710.213.313.815.115.615.4
" + }, + { + "type": "image", + "bbox": [ + 0.081, + 0.34, + 0.177, + 0.48 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.107, + 0.481, + 0.151, + 0.491 + ], + "angle": 0, + "content": "(a) Input" + }, + { + "type": "image", + "bbox": [ + 0.182, + 0.34, + 0.281, + 0.48 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.212, + 0.481, + 0.25, + 0.491 + ], + "angle": 0, + "content": "(b) \\(\\times 5\\)" + }, + { + "type": "image", + "bbox": [ + 0.284, + 0.34, + 0.383, + 0.48 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.315, + 0.481, + 0.352, + 0.491 + ], + "angle": 0, + "content": "(c) \\(\\times 4\\)" + }, + { + "type": "image", + "bbox": [ + 0.385, + 0.34, + 0.484, + 0.48 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.416, + 0.481, + 0.454, + 0.491 + ], + "angle": 0, + "content": "(d) \\(\\times 3\\)" + }, + { + "type": "image", + "bbox": [ + 0.488, + 0.34, + 0.587, + 0.48 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.519, + 0.481, + 0.556, + 0.491 + ], + "angle": 0, + "content": "(e) \\(\\times 2\\)" + }, + { + "type": "image", + "bbox": [ + 0.589, + 0.34, + 0.688, + 0.48 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.617, + 0.481, + 0.662, + 0.491 + ], + "angle": 0, + "content": "\\((\\mathrm{f})\\times 1.5\\)" + }, + { + "type": "image", + "bbox": [ + 0.692, + 0.34, + 0.791, + 0.48 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.724, + 0.481, + 0.76, + 0.491 + ], + "angle": 0, + "content": "\\((\\mathrm{g})\\times 1\\)" + }, + { + "type": "image", + "bbox": [ + 0.793, + 0.34, + 0.892, + 0.48 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.823, + 0.481, + 0.858, + 0.491 + ], + "angle": 0, + "content": "(h) GT" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.509, + 0.895, + 0.537 + ], + "angle": 0, + "content": "Figure 7. Qualitative results on the UHD-LOL4K dataset of different scales. The top row is obtained from the cubic operator without LMAR, while the bottom row is with LMAR. Please zoom in for details." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.563, + 0.47, + 0.638 + ], + "angle": 0, + "content": "\\(x_{c}\\) to ensure its similarity to \\(x_{d}\\). However, measuring the distances in the pixel space deteriorates the learning of the above process. Inspired by [7], we add a discriminator and impose \\(x_{c}\\) and \\(x_{d}\\) to confound it, known as adversarial loss, formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.147, + 0.646, + 0.469, + 0.663 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {d} = \\log \\left(D \\left(x _ {d}\\right)\\right) + \\log \\left(1 - D \\left(x _ {c}\\right)\\right), \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.669, + 0.47, + 0.746 + ], + "angle": 0, + "content": "where \\(D(\\cdot)\\) is the discriminator, which ensures the interaction between the resizer and enhancer while maintaining distribution consistency between \\(x_{c}\\) and \\(x_{d}\\). In addition, to account for the introduction of the discriminator, a new generation loss is incorporated to constrain \\(x_{c}\\), computed as:" + }, + { + "type": "equation", + "bbox": [ + 0.21, + 0.753, + 0.469, + 0.77 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {g} = \\log \\left(D \\left(x _ {c}\\right)\\right). \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.775, + 0.47, + 0.819 + ], + "angle": 0, + "content": "Overall, our optimization objects are twofold. The first optimization focus is on correlating the resizer and enhancer:" + }, + { + "type": "equation", + "bbox": [ + 0.183, + 0.821, + 0.469, + 0.838 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {c} = \\mathcal {L} _ {r c} + \\alpha \\mathcal {L} _ {f a} + \\beta \\mathcal {L} _ {g}, \\tag {14}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.47, + 0.902 + ], + "angle": 0, + "content": "where \\(\\alpha\\) and \\(\\beta\\) are adjustable parameters to balance different losses, which are set empirically. The second optimization object is \\(\\mathcal{L}_d\\) to restrict the distribution. These two optimization objects are alternatively optimized. As can be" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.563, + 0.893, + 0.625 + ], + "angle": 0, + "content": "seen in Figures 5 and 6, our downscaled images not only maintain consistency in terms of perceptual distribution but also exhibit a more consistent representation, as perceived by the model." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.639, + 0.634, + 0.657 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.665, + 0.691, + 0.682 + ], + "angle": 0, + "content": "4.1. Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.689, + 0.892, + 0.78 + ], + "angle": 0, + "content": "Datasets. To evaluate the significance of correlating the enhancer and resizer, experiments are conducted on two UHD low-light enhancement datasets, including the UHD-LOL4K dataset proposed by [33] and the 4KIL dataset proposed by [24]. More details about datasets are provided in the supplementary materials." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.78, + 0.893, + 0.902 + ], + "angle": 0, + "content": "Pre-training. We begin by training multiple enhancers \\( F(\\cdot) \\) to verify the effectiveness of our method. Specifically, we employ the widely used encoder-decoder architecture in image restoration tasks as the backbone. Regarding the UHD-LOL4K dataset, we leverage the invertible block proposed by [25] as the basic unit for processing, termed CNN-I. For the 4KIL dataset, we replace the basic processing unit with the half instance normalization block proposed" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.958 + ], + "angle": 0, + "content": "25727" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.089, + 0.895, + 0.119 + ], + "angle": 0, + "content": "Table 2. Quantitative results on the 4KIL datasets with CNN-H as the backbone. The results with LMAR are shown in gray with better results highlighted in bold." + }, + { + "type": "table", + "bbox": [ + 0.078, + 0.129, + 0.899, + 0.328 + ], + "angle": 0, + "content": "
MethodCNN-H (4KIL)
Scales(2160, 3840)(1440, 2560)(1080, 1920)(1200, 1600)(720, 1280)(540, 960)(432, 768)
Nearest29.80 / 0.980528.36 / 0.977927.71 / 0.975427.67 / 0.974226.98 / 0.971925.92 / 0.961625.37 / 0.9428
29.77 / 0.980128.53 / 0.978928.11 / 0.977927.93 / 0.976227.21 / 0.974526.16 / 0.967225.56 / 0.9493
Bilinear29.80 / 0.980529.12 / 0.979628.47 / 0.978528.39 / 0.977927.30 / 0.972626.38 / 0.965925.56 / 0.9439
29.79 / 0.980429.14 / 0.980428.55 / 0.979428.51 / 0.979027.52 / 0.975026.57 / 0.968725.62 / 0.9464
Bicubic29.80 / 0.980529.25 / 0.979728.60 / 0.978728.47 / 0.978027.30 / 0.972626.36 / 0.965025.58 / 0.9444
29.80 / 0.980429.35 / 0.980428.74 / 0.979728.96 / 0.979327.53 / 0.975226.57 / 0.968525.79 / 0.9509
Lanczos229.80 / 0.980529.24 / 0.979728.59 / 0.978728.46 / 0.978027.29 / 0.972626.35 / 0.964925.56 / 0.9444
29.80 / 0.980329.34 / 0.980528.72 / 0.979628.63 / 0.979227.54 / 0.975226.57 / 0.968425.78 / 0.9509
Lanczos329.80 / 0.980529.32/ 0.979828.69 / 0.979828.54 / 0.978127.34 / 0.973026.34 / 0.964525.58 / 0.9457
29.79 / 0.980329.48 / 0.980428.89 / 0.979928.76 / 0.979427.60 / 0.975626.58 / 0.968625.80 / 0.9523
Time (ms)5.05.05.86.36.56.96.9
5.95.76.16.56.88.58.0
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.341, + 0.895, + 0.371 + ], + "angle": 0, + "content": "Table 3. Quantitative results on the 4KIL datasets with Restormer as the backbone. The results with LMAR are shown in gray with better results highlighted in bold." + }, + { + "type": "table", + "bbox": [ + 0.078, + 0.38, + 0.899, + 0.581 + ], + "angle": 0, + "content": "
MethodRestromer (4KIL)
Scales(2160, 3840)(1440, 2560)(1080, 1920)(1200, 1600)(720, 1280)(540, 960)(432, 768)
Nearest28.86 / 0.977327.81 / 0.975027.57 / 0.973827.45 / 0.972026.99 / 0.971425.91 / 0.960825.31 / 0.9470
28.78 / 0.976828.36 / 0.976527.67 / 0.974227.80 / 0.973927.27 / 0.973325.98 / 0.961925.46 / 0.9527
Bilinear28.86 / 0.977328.63 / 0.977028.50 / 0.976828.49 / 0.976727.56 / 0.971926.71 / 0.968625.71 / 0.9459
28.83 / 0.977028.75 / 0.977128.52 / 0.976928.64 / 0.977027.85 / 0.973926.77 / 0.968825.90 / 0.9531
Bicubic28.86 / 0.977328.83 / 0.977128.64 / 0.977028.53 / 0.976627.57 / 0.971926.65 / 0.966825.79 / 0.9477
28.84 / 0.977228.84 / 0.977228.69 / 0.977128.76 / 0.977127.82 / 0.973826.71 / 0.967825.94 / 0.9533
Lanczos228.86 / 0.977328.83 / 0.977028.63 / 0.976928.52 / 0.976627.56 / 0.971926.65 / 0.966825.78 / 0.9476
28.83 / 0.977228.84 / 0.977128.67 / 0.977028.74 / 0.977327.81 / 0.973926.70 / 0.967825.93 / 0.9533
Lanczos328.86 / 0.977328.84 / 0.977028.68 / 0.976828.55 / 0.976527.56 / 0.971926.56 / 0.965225.72 / 0.9468
28.85 / 0.977228.85 / 0.977328.84 / 0.977128.85 / 0.977227.84 / 0.974026.58 / 0.966525.89 / 0.9528
Time (ms)11.911.512.511.712.011.712.3
12.812.912.913.012.212.612.5
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.605, + 0.471, + 0.697 + ], + "angle": 0, + "content": "by [3] to demonstrate the robustness of our method, termed CNN-H. In addition to the above CNN models, we also train a transformer-based enhancer conditioned on Restormer, as proposed by [38], on the 4KIL dataset to further validate the scalability of our method. More details about pre-training are provided in the supplementary materials." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.71, + 0.291, + 0.727 + ], + "angle": 0, + "content": "4.2. Collaborative Training" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.735, + 0.47, + 0.903 + ], + "angle": 0, + "content": "In this stage, we keep the pre-trained enhancers fixed and train to correlate the resizer and enhancer. Considering the randomness of scale variability in the resampling process, we incorporate random scale training. Specifically, we begin by cropping a \\(H_{p} \\times W_{p}\\) patch from the UHD input image. Next, we randomly sample a scale factor \\(r\\) from a uniform distribution \\(\\mathcal{U}(1,4)\\). Finally, we downsample the high-resolution \\(H_{p} \\times W_{p}\\) patch by the scale factor \\(r\\) to generate a low-resolution \\(h_{p} \\times w_{p}\\) patch. The optimization objectives mentioned earlier are applied to these two resolution patches." + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.603, + 0.895, + 0.646 + ], + "angle": 0, + "content": "Table 4. Comparison of model parameters. Baseline models with LMAR are shown in gray, where the extra parameters are trainable parameters possessed by LMAR." + }, + { + "type": "table", + "bbox": [ + 0.529, + 0.655, + 0.86, + 0.707 + ], + "angle": 0, + "content": "
ModelCNN-ICNN-HRestormer
#Params1.009MB1.128MB1.660MB
1.314MB1.495MB1.965MB
" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Implementation Details. We conduct experiments based on the PyTorch framework along with the tiny-cuda-nn [27] library with one NVIDIA 3090 GPU. During the collaborative training process, the Adam [21] optimizer with \\(\\beta_{1} = 0.9\\) and \\(\\beta_{2} = 0.999\\) is used in our experiments. A batch size of 1 is utilized for all experiments. The cropped patch size \\((H_{p} \\times W_{p})\\) is set to \\(1024 \\times 1024\\) for the CNN enhancers, and \\(768 \\times 768\\) for the transformer enhancer. For the CNN-I and CNN-H, the collaborative training epoch is set to 12 and 24, respectively. The initial learning rate is \\(4e^{-4}\\), which decays by a factor value of 0.75 every 4 epochs and 8 epochs for the CNN-I and CNN-H enhancers. As for" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "25728" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.081, + 0.089, + 0.177, + 0.228 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.107, + 0.23, + 0.151, + 0.24 + ], + "angle": 0, + "content": "(a) Input" + }, + { + "type": "image", + "bbox": [ + 0.182, + 0.089, + 0.279, + 0.229 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.212, + 0.23, + 0.25, + 0.24 + ], + "angle": 0, + "content": "(b) \\(\\times 5\\)" + }, + { + "type": "image", + "bbox": [ + 0.284, + 0.089, + 0.382, + 0.229 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.315, + 0.229, + 0.352, + 0.239 + ], + "angle": 0, + "content": "(c)×4" + }, + { + "type": "image", + "bbox": [ + 0.386, + 0.089, + 0.483, + 0.229 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.417, + 0.229, + 0.454, + 0.239 + ], + "angle": 0, + "content": "(d)×3" + }, + { + "type": "image", + "bbox": [ + 0.488, + 0.089, + 0.586, + 0.229 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.519, + 0.229, + 0.556, + 0.239 + ], + "angle": 0, + "content": "(e)×2" + }, + { + "type": "image", + "bbox": [ + 0.59, + 0.089, + 0.688, + 0.229 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.618, + 0.229, + 0.662, + 0.239 + ], + "angle": 0, + "content": "(f) \\(\\times 1.5\\)" + }, + { + "type": "image", + "bbox": [ + 0.693, + 0.089, + 0.79, + 0.229 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.724, + 0.229, + 0.76, + 0.239 + ], + "angle": 0, + "content": "(g)×1" + }, + { + "type": "image", + "bbox": [ + 0.793, + 0.089, + 0.891, + 0.229 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.825, + 0.229, + 0.86, + 0.239 + ], + "angle": 0, + "content": "(h) GT" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.257, + 0.892, + 0.284 + ], + "angle": 0, + "content": "Figure 8. Qualitative results on the 4KIL dataset of different scales. The top row is obtained from the lanzcos3 operator without LMAR, while the bottom row is with LMAR. Please zoom in for details." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.296, + 0.468, + 0.386 + ], + "angle": 0, + "content": "the transformer enhancer, the number of collaborative training epochs is 60, with a \\(4e^{-4}\\) initial learning rate, which decays by a factor of 0.75 every 20 epochs. The discriminator is composed of several convolution layers, whose initial learning rate is \\(2e^{-4}\\) and shares the same training strategies as mentioned above." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.395, + 0.312, + 0.41 + ], + "angle": 0, + "content": "4.3. Performance Comparison" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.417, + 0.468, + 0.674 + ], + "angle": 0, + "content": "Quantitative Evaluation. We utilize two metrics, the Peak Signal to Noise Ratio (PSNR, dB), and the Structural Similarity (SSIM), to assess the performance gains. To verify the universality of our methods, we adopt five types of resizers and test them on various commonly used scales, including both in-distribution and out-of-distribution scales. As shown in Tables 1, 2, and 3, our method consistently achieves performance improvements across different resolutions, regardless of the choice of resizer or the category of the backbone. Notably, the performance with LMAR at UHD resolution is analogous to the original UHD input, which illustrates the superior continuous modeling capability of LMAR. Additionally, our approach maintains comparable effectiveness with the baselines, where a minor increase in parameters, as shown in Table 4. More importantly, our algorithm only needs to be trained on one type of resizer, which can then be applied to any other resizers." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.674, + 0.468, + 0.794 + ], + "angle": 0, + "content": "Qualitative Evaluation. Figures 7 and 8 present the visual comparison on the UHD-LOL4K dataset and 4KIL dataset respectively. These comparisons demonstrate that enabling model-aware resampling leads to enhanced results with improved global lightness and local structural details, across various resampling scales. It verifies that coupling the resize-ers and enhancers could generate better visual results. More visual results are presented in the supplementary materials." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.802, + 0.241, + 0.816 + ], + "angle": 0, + "content": "4.4. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.468, + 0.9 + ], + "angle": 0, + "content": "We conduct ablation studies on the UHD-LOL4K dataset to investigate the rationality of our designs. The ablation results are shown in Table 5. First, we replace the bottleneck features (Full) with the relatively front features \\((F_{f})\\) and relatively end features \\((F_{e})\\) to apply feature constraints," + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.296, + 0.892, + 0.446 + ], + "angle": 0, + "content": "where remarkable performance drops can be observed at 4 times downsampling. The findings indicate that the bottleneck features exhibit the most representative capability in guiding model-aware resampling. Second, we replace the discriminator (Full) with the perceptual loss (VGG) to conduct the distribution restriction. The utilization of discriminator encourages more consistent feature alignment. Lastly, even with sub-optimal constraints, our method outperforms the baseline, emphasizing the effectiveness of coupling resizers and enhancers." + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.458, + 0.891, + 0.486 + ], + "angle": 0, + "content": "Table 5. Ablation studies on the constraints employed in our designs. We present the PSNR results here." + }, + { + "type": "table", + "bbox": [ + 0.512, + 0.498, + 0.878, + 0.548 + ], + "angle": 0, + "content": "
ScalesBaselineFfFeVGGFull
x229.8229.9830.0530.0330.20
x425.4025.8625.5025.4726.36
" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.566, + 0.618, + 0.581 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.591, + 0.892, + 0.772 + ], + "angle": 0, + "content": "In this paper, we introduce a novel approach for effectively enhancing the UHD images at different downsampling scales by a model-aware resampling strategy. Our method customizes resampling under the guidance of model knowledge, integrating the resizers and enhancers through compensatory information learning. Notably, our algorithm is fully compatible with existing interpolation resamplers and maintains the efficiency. Moreover, our design promotes performances at low resolutions remarkably while keeping comparable results at UHD resolution, without the need to retrain the enhancer. Extensive experiments have validated the effectiveness and scalability of our method." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.785, + 0.671, + 0.802 + ], + "angle": 0, + "content": "6. Acknowledgment" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.81, + 0.892, + 0.899 + ], + "angle": 0, + "content": "This work was supported by the JKW Research Funds under Grant 20-163-14-LZ-001-004-01, and the Anhui Provincial Natural Science Foundation under Grant 2108085UD12. We acknowledge the support of GPU cluster built by MCC Lab of Information Science and Technology Institution, USTC." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "25729" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.469, + 0.142 + ], + "angle": 0, + "content": "[1] Kenneth R Castleman. Digital image processing. Prentice Hall Press, 1996. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.145, + 0.469, + 0.186 + ], + "angle": 0, + "content": "[2] Jiawen Chen, Andrew Adams, Neal Wadhwa, and Samuel W Hasinoff. Bilateral guided upsampling. ACM Transactions on Graphics (TOG), 35(6):1-8, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.188, + 0.469, + 0.257 + ], + "angle": 0, + "content": "[3] Liangyu Chen, Xin Lu, Jie Zhang, Xiaojie Chu, and Chengpeng Chen. Hinet: Half instance normalization network for image restoration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, pages 182-192, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.259, + 0.469, + 0.327 + ], + "angle": 0, + "content": "[4] Yinbo Chen, Sifei Liu, and Xiaolong Wang. Learning continuous image representation with local implicit image function. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8628-8638, 2021. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.33, + 0.469, + 0.4 + ], + "angle": 0, + "content": "[5] Weitao Feng, Deyi Ji, Yiru Wang, Shuorong Chang, Hansheng Ren, and Weihao Gan. Challenges on large scale surveillance video analysis. In Proceedings of the IEEE conference on computer vision and pattern recognition workshops, pages 69-76, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.402, + 0.469, + 0.457 + ], + "angle": 0, + "content": "[6] Michael Gharbi, Jiawen Chen, Jonathan T Barron, Samuel W Hasinoff, and Frédo Durand. Deep bilateral learning for real-time image enhancement. ACM Transactions on Graphics (TOG), 36(4):1-12, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.459, + 0.469, + 0.514 + ], + "angle": 0, + "content": "[7] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial networks. Communications of the ACM, 63(11):139-144, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.517, + 0.469, + 0.584 + ], + "angle": 0, + "content": "[8] Dianyuan Han. Comparison of commonly used image interpolation methods. In Conference of the 2nd International Conference on Computer Science and Electronics Engineering (ICCSEE 2013), pages 1556-1559. Atlantis Press, 2013. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.587, + 0.469, + 0.642 + ], + "angle": 0, + "content": "[9] Muhammad Haris, Gregory Shakhnarovich, and Norimichi Ukita. Deep back-projection networks for super-resolution. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1664-1673, 2018. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.645, + 0.469, + 0.699 + ], + "angle": 0, + "content": "[10] Hanzhe Hu, Deyi Ji, Weihao Gan, Shuai Bai, Wei Wu, and Junjie Yan. Class-wise dynamic graph convolution for semantic segmentation. In European Conference on Computer Vision, pages 1-17. Springer, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.702, + 0.47, + 0.77 + ], + "angle": 0, + "content": "[11] Xuecai Hu, Haoyuan Mu, Xiangyu Zhang, Zilei Wang, Tieniu Tan, and Jian Sun. Meta-sr: A magnification-arbitrary network for super-resolution. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1575-1584, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.773, + 0.469, + 0.814 + ], + "angle": 0, + "content": "[12] Max Jaderberg, Karen Simonyan, Andrew Zisserman, et al. Spatial transformer networks. Advances in neural information processing systems, 28, 2015. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.817, + 0.469, + 0.87 + ], + "angle": 0, + "content": "[13] Deyi Ji, Hongtao Lu, and Tongzhen Zhang. End to end multiscale convolutional neural network for crowd counting. In Eleventh international conference on machine vision, pages 761-766. SPIE, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[14] Deyi Ji, Haoran Wang, Hanzhe Hu, Weihao Gan, Wei Wu, and Junjie Yan. Context-aware graph convolution network" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.47, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "for target re-identification. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1646–1654, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.122, + 0.892, + 0.189 + ], + "angle": 0, + "content": "[15] Deyi Ji, Haoran Wang, Mingyuan Tao, Jianqiang Huang, Xian-Sheng Hua, and Hongtao Lu. Structural and statistical texture knowledge distillation for semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16876-16885, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.191, + 0.892, + 0.232 + ], + "angle": 0, + "content": "[16] Deyi Ji, Siqi Gao, Mingyuan Tao, Hongtao Lu, and Feng Zhao. Changenet: Multi-temporal asymmetric change detection dataset. arXiv preprint arXiv:2312.17428, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.234, + 0.892, + 0.287 + ], + "angle": 0, + "content": "[17] Deyi Ji, Feng Zhao, and Hongtao Lu. Guided patch-grouping wavelet transformer with spatial congruence for ultra-high resolution segmentation. International Joint Conference on Artificial Intelligence, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.289, + 0.892, + 0.356 + ], + "angle": 0, + "content": "[18] Deyi Ji, Feng Zhao, Hongtao Lu, Mingyuan Tao, and Jieping Ye. Ultra-high resolution segmentation with ultra-rich context: A novel benchmark. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23621-23630, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.358, + 0.892, + 0.411 + ], + "angle": 0, + "content": "[19] Deyi Ji, Siqi Gao, Lanyun Zhu, Yiru Zhao, Peng Xu, Hong-tao Lu, and Feng Zhao. View-centric multi-object tracking with homographic matching in moving UAV. arXiv preprint, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.414, + 0.892, + 0.467 + ], + "angle": 0, + "content": "[20] Heewon Kim, Myungsub Choi, Bee Lim, and Kyoung Mu Lee. Task-aware image downscaling. In Proceedings of the European conference on computer vision (ECCV), pages 399-414, 2018. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.47, + 0.892, + 0.508 + ], + "angle": 0, + "content": "[21] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014.7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.511, + 0.892, + 0.565 + ], + "angle": 0, + "content": "[22] Yue Li, Dong Liu, Houqiang Li, Li Li, Zhu Li, and Feng Wu. Learning a convolutional neural network for image compact-resolution. IEEE Transactions on Image Processing, 28(3): 1092-1107, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.568, + 0.892, + 0.634 + ], + "angle": 0, + "content": "[23] Jie Liang, Hui Zeng, and Lei Zhang. High-resolution photorealistic image translation in real-time: A laplacian pyramid translation network. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9392-9400, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.637, + 0.892, + 0.677 + ], + "angle": 0, + "content": "[24] Qiaowanni Lin, Zhuoran Zheng, and Xiuyi Jia. Uhd low-light image enhancement via interpretable bilateral learning. Information Sciences, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.679, + 0.892, + 0.733 + ], + "angle": 0, + "content": "[25] Yang Liu, Zhenyue Qin, Saeed Anwar, Pan Ji, Dongwoo Kim, Sabrina Caldwell, and Tom Gedeon. Invertible denoising network: A light solution for real noise removal. arXiv preprint arXiv:2104.10546, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.735, + 0.892, + 0.774 + ], + "angle": 0, + "content": "[26] Don P Mitchell and Arun N Netravali. Reconstruction filters in computer-graphics. ACM Siggraph Computer Graphics, 22(4):221-228, 1988. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.777, + 0.764, + 0.789 + ], + "angle": 0, + "content": "[27] Thomas Müller. tiny-cuda-nn, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.791, + 0.892, + 0.857 + ], + "angle": 0, + "content": "[28] Adria Recasens, Petr Kellnhofer, Simon Stent, Wojciech Matusik, and Antonio Torralba. Learning to zoom: a saliency-based sampling layer for neural networks. In Proceedings of the European conference on computer vision (ECCV), pages 51-66, 2018. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.86, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[29] Assaf Shocher, Ben Feinstein, Niv Haim, and Michal Irani. From discrete to continuous convolution layers. arXiv preprint arXiv:2006.11120, 2020.4" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "25730" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.146 + ], + "angle": 0, + "content": "[30] Wanjie Sun and Zhenzhong Chen. Learned image downscaling for upscaling using content adaptive resampler. IEEE Transactions on Image Processing, 29:4027-4040, 2020. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.149, + 0.47, + 0.205 + ], + "angle": 0, + "content": "[31] Haoran Wang, Licheng Jiao, Fang Liu, Lingling Li, Xu Liu, Deyi Ji, and Weihao Gan. Ipgn: Interactiveness proposal graph network for human-object interaction detection. IEEE Transactions on Image Processing, 30:6583-6593, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.207, + 0.469, + 0.275 + ], + "angle": 0, + "content": "[32] Haoran Wang, Licheng Jiao, Fang Liu, Lingling Li, Xu Liu, Deyi Ji, and Weihao Gan. Learning social spatio-temporal relation graph in the wild and a video benchmark. IEEE Transactions on Neural Networks and Learning Systems, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.278, + 0.469, + 0.332 + ], + "angle": 0, + "content": "[33] Tao Wang, Kaihao Zhang, Tianrun Shen, Wenhan Luo, Bjorn Stenger, and Tong Lu. Ultra-high-definition low-light image enhancement: A benchmark and transformer-based method. arXiv preprint arXiv:2212.11548, 2022. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.334, + 0.469, + 0.417 + ], + "angle": 0, + "content": "[34] Mingqing Xiao, Shuxin Zheng, Chang Liu, Yaolong Wang, Di He, Guolin Ke, Jiang Bian, Zhouchen Lin, and Tie-Yan Liu. Invertible image rescaling. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part I 16, pages 126-144. Springer, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.419, + 0.469, + 0.46 + ], + "angle": 0, + "content": "[35] Jinbo Xing, Wenbo Hu, Menghan Xia, and Tien-Tsin Wong. Scale-arbitrary invertible image downscaling. IEEE Transactions on Image Processing, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.462, + 0.469, + 0.53 + ], + "angle": 0, + "content": "[36] Wei Yu, Qi Zhu, Naishan Zheng, Jie Huang, Man Zhou, and Feng Zhao. Learning non-uniform-sampling for ultra-high-definition image enhancement. In Proceedings of the 31st ACM International Conference on Multimedia, pages 1412-1421, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.533, + 0.469, + 0.587 + ], + "angle": 0, + "content": "[37] Sergey Zagoruyko and Nikos Komodakis. Paying more attention to attention: Improving the performance of convolutional neural networks via attention transfer. arXiv preprint arXiv:1612.03928, 2016. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.59, + 0.469, + 0.645 + ], + "angle": 0, + "content": "[38] Syed Waqas Zamir, Aditya Arora, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, and Ming-Hsuan Yang. Restormer: Efficient transformer for high-resolution image restoration. In CVPR, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.647, + 0.469, + 0.701 + ], + "angle": 0, + "content": "[39] Yongbing Zhang, Debin Zhao, Jian Zhang, Ruiqin Xiong, and Wen Gao. Interpolation-dependent image downsampling. IEEE Transactions on Image Processing, 20(11): 3291-3296, 2011. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.704, + 0.469, + 0.785 + ], + "angle": 0, + "content": "[40] Zhuoran Zheng, Wenqi Ren, Xiaochun Cao, Xiaobin Hu, Tao Wang, Fenglong Song, and Xiuyi Jia. Ultra-high-definition image dehazing via multi-guided bilateral learning. In 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 16180-16189. IEEE, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.788, + 0.469, + 0.857 + ], + "angle": 0, + "content": "[41] Zhuoran Zheng, Wenqi Ren, Xiaochun Cao, Tao Wang, and Xiuyi Jia. Ultra-high-definition image hdr reconstruction via collaborative bilateral learning. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4449-4458, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.859, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[42] Lanyun Zhu, Deyi Ji, Shiping Zhu, Weihao Gan, Wei Wu, and Junjie Yan. Learning statistical texture for semantic segmentation. In Proceedings of the IEEE/CVF Conference" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "on Computer Vision and Pattern Recognition (CVPR), pages 12537-12546, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.122, + 0.892, + 0.163 + ], + "angle": 0, + "content": "[43] Lanyun Zhu, Tianrun Chen, Deyi Ji, Jieping Ye, and Jun Liu. Llafs: When large-language models meet few-shot segmentation. arXiv preprint arXiv:2311.16926, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.165, + 0.892, + 0.219 + ], + "angle": 0, + "content": "[44] Lanyun Zhu, Deyi Ji, Tianrun Chen, Peng Xu, Jieping Ye, and Jun Liu. Ibid: Alleviating hallucinations in large vision-language models via image-biased decoding. arXiv preprint arXiv:2402.18476, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.221, + 0.892, + 0.276 + ], + "angle": 0, + "content": "[45] Qi Zhu, Zeyu Xiao, Jie Huang, and Feng Zhao. Dast-net: Depth-aware spatio-temporal network for video deblurring. In 2022 IEEE International Conference on Multimedia and Expo (ICME), pages 1-6. IEEE, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.278, + 0.892, + 0.332 + ], + "angle": 0, + "content": "[46] Qi Zhu, Naishan Zheng, Jie Huang, Man Zhou, Jinghao Zhang, and Feng Zhao. Learning spatio-temporal sharpness map for video deblurring. IEEE Transactions on Circuits and Systems for Video Technology, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.334, + 0.892, + 0.403 + ], + "angle": 0, + "content": "[47] Qi Zhu, Man Zhou, Naishan Zheng, Chongyi Li, Jie Huang, and Feng Zhao. Exploring temporal frequency spectrum in deep video deblurring. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12428-12437, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.405, + 0.892, + 0.46 + ], + "angle": 0, + "content": "[48] Qi Zhu, Jie Huang, Naishan Zheng, Hongzhi Gao, Chongyi Li, Yuan Xu, Feng Zhao, et al. Fouridown: Factoring down-sampling into shuffling and superposing. Advances in Neural Information Processing Systems, 36, 2024. 3" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.46 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.518, + 0.957 + ], + "angle": 0, + "content": "25731" + } + ] +] \ No newline at end of file diff --git a/2024/Empowering Resampling Operation for Ultra-High-Definition Image Enhancement with Model-Aware Guidance/5e31bf7f-5e48-4030-90cc-b526f724e932_origin.pdf b/2024/Empowering Resampling Operation for Ultra-High-Definition Image Enhancement with Model-Aware Guidance/5e31bf7f-5e48-4030-90cc-b526f724e932_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..305e4c1651d3e81b0e8332a4343fca76b1af156c --- /dev/null +++ b/2024/Empowering Resampling Operation for Ultra-High-Definition Image Enhancement with Model-Aware Guidance/5e31bf7f-5e48-4030-90cc-b526f724e932_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:cb7788319aa5669f2f92a7f6cb5a0a5714f582d388590177e44cc709f13c2d73 +size 1383848 diff --git a/2024/Empowering Resampling Operation for Ultra-High-Definition Image Enhancement with Model-Aware Guidance/full.md b/2024/Empowering Resampling Operation for Ultra-High-Definition Image Enhancement with Model-Aware Guidance/full.md new file mode 100644 index 0000000000000000000000000000000000000000..4ab2406d59b7308c55c013d6303dd8c439a992f4 --- /dev/null +++ b/2024/Empowering Resampling Operation for Ultra-High-Definition Image Enhancement with Model-Aware Guidance/full.md @@ -0,0 +1,442 @@ +# Empowering Resampling Operation for Ultra-High-Definition Image Enhancement with Model-Aware Guidance + +Wei Yu,\* Jie Huang,\* Bing Li, Kaiwen Zheng, Qi Zhu, Man Zhou, Feng Zhao† +University of Science and Technology of China + +{patrick914y, hj0117,bing0123, kezh,zqcrafts, manman}@mail.ustc.edu.cn，fzhao956@ustc.edu.cn + +![](images/1a5bd2092423b406acafa244151d833ba75749badadefc7ba13a96ee075c0614.jpg) +(a) Down-Enhancement-Up + +![](images/588349239bdf26e4b63c890585679a3a59f70d6928bd7c396e4bdd5ad9ddec95.jpg) + +![](images/57ad2f614867adbb59a1d277191baa8e845b8e1a394d49956ee0640f4361b690.jpg) + +![](images/ec1ce28b3efdcc0349ecac11d3cb237058ee621b36a018375d4d579e8492afcf.jpg) +(b) Ours + +![](images/e47c2c900f580925422897fe14d1eeca5fb205c328a80ddc29835e93d702e8a8.jpg) +PSNR: 25.59 + +![](images/5c97497c3d726a06d80ee57c6a2eb657c0ae9765837a8d269c3ee8d352959be1.jpg) +PSNR: 28.51 + +![](images/af1759eaa900b0b71b03a975c18a833fe40ab7afa420db4539b6ededada66984.jpg) +PSNR: 30.78 + +![](images/5065405d10d298974acbfb18cd7764d5be0d0d4d01b836e55174c755e44829ae.jpg) +PSNR: 31.05 + +![](images/9856432f41090743975fd662f70cc6401d7d0eb86905334b297cdae64fef9f13.jpg) +PSNR: 28.09 +PSNR: 29.56 +PSNR: 32.24 +PSNR: 33.16 +Figure 1. Comparison between previous Downsampling-Enhancement-Upsampling paradigm and our proposed method. (a) Previous works treat resampling operators and inner models as separate components, while (b) our method integrates them via empowering model-aware resampling. As shown on the right side, our approach achieves significant performance gains over various resampling scales. + +# Abstract + +Image enhancement algorithms have made remarkable advancements in recent years, but directly applying them to Ultra-high-definition (UHD) images presents intractable computational overheads. Therefore, previous straightforward solutions employ resampling techniques to reduce the resolution by adopting a "Downsampling-Enhancement-Upsampling" processing paradigm. However, this paradigm disentangles the resampling operators and inner enhancement algorithms, which results in the loss of information that is favored by the model, further leading to sub-optimal outcomes. In this paper, we propose a novel method of Learning Model-Aware Resampling (LMAR), which learns to customize resampling by extracting model-aware information from the UHD input image, under the guidance of model knowledge. Specifically, our method consists of two core designs, namely compensatory kernel estimation and steganographic resampling. At the first stage, we dynamically predict compensatory kernels tailored to the specific input and resampling scales. At + +the second stage, the image-wise compensatory information is derived with the compensatory kernels and embedded into the rescaled input images. This promotes the representation of the newly derived downscaled inputs to be more consistent with the full-resolution UHD inputs, as perceived by the model. Our LMAR enables model-aware and model-favored resampling while maintaining compatibility with existing resampling operators. Extensive experiments on multiple UHD image enhancement datasets and different backbones have shown consistent performance gains after correlating resizer and enhancer, e.g., up to 1.2dB PSNR gain for $\times 1.8$ resampling scale on UHD-LOL4K. The code is available at https://github.com/YPPatrickW/LMAR. + +# 1. Introduction + +Over time, learning-based image enhancement algorithms have achieved progressive performance improvements. However, the majority of existing methods are not compatible with UHD images due to the heavy computational burdens imposed by their megapixel counts. + +To address this, previous works have sought to alleviate computation overheads by employing downsampling to reduce the resolution, as shown in Figure 1. Concretely, UHD input images are often resampled to smaller sizes for + +![](images/4b768982962e32cc832c9c561b31d4e6530d9dd28d6965b8a93c755ab20382b3.jpg) +Figure 2. The influence of the front-end resampler. We choose the cubic resampling operator as the baseline and introduce perturbations to its kernel function. Interestingly, manipulating the resampling function can lead to volatile outcomes, with the image quality either being improved or degraded. We also present two more comparisons: the resampling residual of the UHD input image and the feature maps generated by different resizers. It is indicated that the response of the model is highly sensitive to the sampling strategies. This highlights the importance of customizing the resampling process in a model-aware manner. + +enhancement, and then subsequently resampled back to the original resolution[22, 39]. Surprisingly, the majority of research efforts have been paid to developing inner processing algorithms, with little attention devoted to the foremost and endmost resampling operators. + +Recently, a line of work has proposed the replacement of the conventional interpolation-based resampling operators (e.g., nearest, bilinear, bicubic) with learnable resampling operators [12, 20, 28, 30]. Attributing to the superior modeling capability of deep neural networks, these learnable resampling operators have demonstrated remarkable performance gains. However, the practical application of these learnable resampling operators is hindered by their high computational complexity and reduced efficiency, resulting in a continued preference for interpolation-based resampling operators. + +Revisiting the above UHD image enhancement pipeline: Downsampling-Enhancement-Upsampling, with the internal enhancement algorithm remaining unchanged, a fundamental question arises: 'How significant is the impact of these two resampling operations?' The answer to this question is twofold. On the one hand, the quality of the upsampled enhanced results exhibits a decreasing trend as the degree of downsampling aggravates. On the other hand, manipulating the interpolation kernel functions yields a diverse range of outcomes, as illustrated in Figure 2. We attribute this phenomenon to the isolation between the resampling operators and the enhancement algorithm, where both components are disentangled from each other. + +To this end, the focus of this paper lies in correlating the resampling operators within the internal enhancement algorithms and unleashing the potential of the resampling operation to make benefits for the aforementioned framework, + +without modifying the inner processing methods (e.g., pretrained networks). To accomplish this goal, we propose a novel method termed Learning Model-Aware Resampling (LMAR). LMAR correlates the resizer and the enhancer through a compensatory information learning process. + +We begin by delving into the impact of resampling from the viewpoint of the enhancer, where the resampling operations directly undermine the representation ability of intermediate features. Therefore, the intermediate features of the UHD image offer valuable insights on how to perform resampling in a model-aware manner. Building upon this recognition, we formulate our method into two steps: Compensatory Kernel Estimation (CKE) and Steganographic Resampling (SR). In the CKE step, we generate imagespecific convolution kernels for the input UHD image via implicit neural representation. These kernels are specifically customized for each input and resampling scale. Then, we can obtain compensatory information produced by these predicted filters. Subsequently, in the SR step, the compensatory information is embedded in the downscaled input image, resulting in a newly learnable downscaled input. By utilizing the compensatory information as a medium, we establish the correlation between the resampling operators and the inner enhancer. Finally, the learnable input is passed to the enhancer and optimized to encourage that the intermediate features align more consistently with those of the UHD image. Thanks to these designs, our method not only maintains compatibility with existing interpolation-based resampling operators but also incorporates them into a part of enhancement algorithms under the guidance of model knowledge. Our contributions are summarized as: + +- We rethink the disentanglement between the resizer and enhancement model and propose to perform the resampling in a model-aware manner. +- We introduce LMAR, a novel method that facilitates the collaboration between resizer and enhancer via customizing resampling under the guidance of model knowledge, where the representative capabilities of low-resolution images are largely improved. +- Our LMAR delivers two core designs: compensatory kernel estimation and steganographic resampling, which models the interplay between responder and enhancer by learning embeddable information. Our method is compatible with any interpolation resamplers. +- Extensive experiments are conducted to verify the superiority of correlating resizer and enhancer. Remarkably, our algorithm significantly improves the performance at low resolutions while maintaining equivalent results at the original resolution, without requiring retraining of the enhancer. + +![](images/a999215e919be8d6ab60f20228b1e4e496f972e2b3d45178b34cf3393b9edf1b.jpg) +Figure 3. The inspiration for our method design. We replace the compressed bottleneck features of the downsampled input with features extracted from the UHD image (The green arrow). Large-margin performance gains can be achieved compared with the vested paradigm (The gray arrow). This observation highlights the potential role of feature consistency between UHD input and any downscaled input, which can play as a bridge for establishing correlations between resampling operators and enhancers. + +# 2. Related Work + +# 2.1. Image Resampling + +Image resampling has emerged as one of the most commonly employed techniques for altering the resolution of an image. The key to image resampling lies in effectively preserving the quality of image content while maintaining efficiency. + +Interpolation operators, such as bilinear, and bicubic [1, 8], have been the popular choice for image resampling for many years. These operators are capable of effectively resampling images at any scale, yet they disregard the spatial variations of distinct image patches, leading to inadequate preservation of the local structure [26]. + +In recent years, deep neural networks have facilitated advancements in various tasks [5, 10, 13-19, 31, 32, 36, 42-48], including image resampling [4, 11, 12, 20, 28, 30, 34, 35]. Task-aware image downsampling (TAD) [20] presents an encoder-decoder framework that aims to enhance the quality of both downscaling and upscaling processes in a collaborative manner. Though these methods achieve excellent results, none of them focus on correlating interpolation resampler with a pre-trained model. + +# 2.2. UHD Image Processing + +A line of studies that specifically focus on high-resolution image processing has been proposed [2, 6, 23, 33, 40, 41]. These methods share a common paradigm: learning from downscaled images to alleviate the computational burdens. In this learning paradigm, the quality of the results is directly determined by the resampled image representation, highlighting the crucial role of resampling strategies. However, they often overlook the impact of resampling approaches and lack in-depth investigation of various resam + +pling operators. In this work, our objective is to explore the interplay between enhancer and resampling operators and establish a collaborative relationship for them to improve performance. + +# 3. Methodology + +In this section, we will present a comprehensive illustration of our proposed LMAR. We begin by introducing the motivation behind our approach and describing our settings. Subsequently, we will overview the whole processing pipeline and delve into the design of the Compensatory Kernel Estimation (CKE) and Steganographic Resampling (SR). Finally, we will discuss the optimization strategies employed to enable model-aware resampling. + +# 3.1. Setting and Motivation + +Our setting still follows the Downsampling-Enhancement-Upsampling paradigm, employing the interpolation-based resampling operators instead of introducing additional burdens with learnable resampling operators. Notably, we will first train an enhancement model for each dataset in our experimental setup and keep unchanged. The main difference lies in the collaboration process between the resizing and enhancer. + +The main inspiration stems from observing the impact of resampling on the representative feature extraction process of the enhancer. As shown in Figure 3, we inject the full-resolution bottleneck features into the enhancement process of the downsampled input images, where the final results significantly outperform the vested paradigm. It proves that front-most resampling degrades the representative ability of bottleneck features and impacts the following processes. Based on the insight, we correlate the resizer and enhancer through a compensatory information-learning process to achieve consistent feature alignment. + +# 3.2. Overview + +The Downsampling-Enhancement-Upsampling processing can be mathematically formulated as: + +$$ +x _ {d} = D (x; [ h, w ]), +$$ + +$$ +\tilde {y} = F \left(x _ {d}\right), \tag {1} +$$ + +$$ +y = U (\tilde {y}; [ H, W ]), +$$ + +where $D(\cdot)$ is the down sampler that resamples the UHD input image $x \in \mathbb{R}^{H \times W \times 3}$ to lower resolution of $x_{d} \in \mathbb{R}^{h \times w \times 3}$ , $F(\cdot)$ is the pre-trained enhancement model, and $U(\cdot)$ denotes the upsampler which resamples the low-resolution enhanced result $\tilde{y}$ back to the UHD resolution of $H \times W$ , termed as $y$ . In LMAR, we rebuild the above paradigm, given as: + +$$ +\tilde {y} = F \left(x _ {d}, \Theta (x; D, U, F)\right), \tag {2} +$$ + +$$ +y = U (\tilde {y}; [ H, W ]). +$$ + +![](images/01ad7bbe2229e30ed77c4945b98e2edd3255d39b3569939d760f107929f0459f.jpg) +1. Training Stage + +![](images/a5164a9d7600e4e3125278c151df765d3ecf52d49d6eb923f406ce623047717c.jpg) +2. Inference Stage +(b) + +![](images/d3ee7d979e37be6c5a7a1570df5a9130f3d38ec79b0c57b80f132495d3d286ac.jpg) +(a) + +![](images/900be3f80c5813429b86065c569562b523d9eb1fed3b0b29657625088a47802e.jpg) +Figure 4. Overview of the proposed LMAR. The sub-graph (a) depicts the training phase of LMAR, which encourages the compensated low-resolution input to maintain representation consistency with the full-resolution UHD input, as perceived by the enhancer. The subgraph (b) demonstrates the inference pipeline of our LMAR, where the compensated low-resolution input is directly fed into the enhancer and then upsampled to the UHD result. The sub-graph (c) illustrates how LMAR works, where the core lies in estimating compensatory kernels under the guidance of model knowledge to make up for the resampling process. + +Here, we introduce additional compensatory information as a medium to facilitate model-aware resampling, denoted as $\Theta (x;D,U,F)$ . Our method is presented in Figure 4, where the model-aware resampling is achieved via two core designs, compensatory kernel estimation, and steganographic resampling. + +# 3.3. Compensatory Kernel Estimation + +Interpolation-based image resampling process can be decomposed as projected grid calculation and weighted pixel aggregation [29]. To this end, we estimate the compensatory kernel using a representation that is related to the scale. Inspired by LIFF [4], we employ two scale features, the relative coordinate grid and the pixel cell. The relative coordinate grid depicts the pixel location shift in the resampling processing. We first compute the uniform coordinate of the downsampling scale $C_d \in \mathbb{R}^{h \times w \times 2}$ and the UHD scale $C_u \in \mathbb{R}^{H \times W \times 2}$ , which are normalized between $[-1, 1]$ . The calculations of the above grids are given as: + +$$ +C _ {d} (i, j) = (- 1 + \frac {(2 i + 1)}{h}, - 1 + \frac {(2 j + 1)}{w}), \tag {3} +$$ + +$$ +C _ {u} (i, j) = (- 1 + \frac {(2 i + 1)}{H}, - 1 + \frac {(2 j + 1)}{W}), +$$ + +where $i \in [0, h - 1] \vee [0, H - 1], j \in [0, w - 1] \vee [0, W - 1]$ are the position indexes for the height and width dimensions. As the sizes of the two coordinates are different, we first project them $C_d$ to $C_u$ and obtain the relative grid shift at the downsampling scale, which is expressed as: + +$$ +\tilde {C} _ {d} = \operatorname {g r i d} \text {- s a m p l e} \left(C _ {d}, C _ {u}\right), +$$ + +$$ +\tilde {C} _ {r} = C _ {u} - \tilde {C} _ {d}, \tag {4} +$$ + +$$ +C _ {r} = \left(\tilde {C} _ {r} (i) \times h, \tilde {C} _ {r} (j) \times w\right), +$$ + +where the grid_sample is a remapping function that remaps the $C_d$ into $\tilde{C}_d$ of size $H \times W$ , $\tilde{C}_r$ represents the relative shift and $C_r$ denotes the relative coordinate grid measured at downsampling scale. The pixel cell represents the ratio of pixel area changes between downsampling and upsampling, denoted as: + +$$ +P _ {c} (i, j) = \left(2 * \frac {h}{H}, 2 * \frac {w}{W}\right), \tag {5} +$$ + +where $P_{c}\in \mathbb{R}^{H\times W\times 2}$ , and all pixel locations share the same scale ratio. + +The aforementioned scale-related representations are instance-same at a given scale, leading to the generation of identical kernels without considering the distinctions between images. Thus, we propose the utilization of backprojection [9] as a constraint to capture the interrelation + +![](images/5a71842bcd4ce590e6d4f429e2a04ff9705d6a101bbcdba719ff25e08352c242.jpg) +(a) $x_{d}$ + +![](images/4d7591c13ae2781df508dd572e0804e77da7711bce344a22fecfe30ca143ca3f.jpg) +(b) $x_{c}$ + +![](images/a7e30423361ebdce19e2d1e497ea1d2867d803195ef9465e17cbcca7cb10beae.jpg) +(c) $|x_{c} - x_{d}|$ +Figure 5. Comparison of different downscaled input images. (a) The downscaled input images by the cubic resampling operator, (b) images derived from the cubic resampling operator in a model-aware manner, and (c) the residual between (a) and (b), where obvious differences can be witnessed in high-frequency areas. + +between each downscaled input and its corresponding UHD counterpart, formulated as: + +$$ +\tilde {x} = \operatorname {C o n v} (x - U (x _ {d}; [ H, W ]))), \tag {6} +$$ + +where $\tilde{x} \in \mathbb{R}^{H \times W \times 3}$ is the back-projection of UHD input image $x$ . + +Considering the pixel-wise aggregation nature of the resampling process, we model the generation of compensatory information by predicting the convolution kernels for the UHD input image instead of predicting pixel values for specific locations. Specifically, we flatten the relative coordinate $C_r$ , pixel cell $P_c$ , and back-projection $\tilde{x}$ then concatenate them. Afterward, we feed these concatenated features into a multi-layer perception (MLP) to predict compensatory kernels as follows: + +$$ +K = M L P _ {\phi} \left(\left[ C _ {r}; P _ {c}; \tilde {x} \right]\right), \tag {7} +$$ + +where $[\cdot ]$ denotes the channel-wise concatenation, and $K\in$ $\mathbb{R}^{H\times W\times s}$ represents the predicted compensatory kernels. The kernel for each pixel location is represented in the channel dimension $s$ , consisting of the multiplication of the input channel, out channel, and kernel size, with all being three in our case. + +# 3.4. Steganographic Resampling + +With the estimated compensatory kernel from the above process, we calculate the compensatory information for the input UHD image as follows: + +$$ +\widehat {x} = K \otimes x, \tag {8} +$$ + +where $\widehat{x} \in \mathbb{R}^{H \times W \times 3}$ is the generated UHD compensatory embedding, and $\otimes$ denotes the convolution operation. + +To further maintain the intrinsic characteristics of the downscaled input images, we downscale the compensatory information and perform compensatory steganography in + +![](images/cbccefd1c5f2b8da13056c8bfa704158bdd5f1e19ec2f65febf874bfa689d51f.jpg) + +![](images/d41d07451a105764a98e9d2340ea2da8e7b97f63b1bc72e83552cab7d39f231c.jpg) + +![](images/5ff6d4209e997f5cd29523555b0c59a245220e7a1a123c4e00efac79befb38fb.jpg) +Errors + +![](images/c4912282f7391e97ff6a710a8d743beec2d9b9cad447ed23be325295a1d5bec0.jpg) +$x_{d}$ +$x_{c}$ + +![](images/a0c3bbed12608d45557a1c1c7fb4a9fbf66f25cdd46d16835f2722094bdf7203.jpg) +$E(x_{d})$ +$E(x_{c})$ + +![](images/410630a60511c85aaf12b17e1a05d907be443a8873753c2af0423aeb3ebbcc3e.jpg) +Errors +Figure 6. Feature comparison of different downscaled input images. The second column represents the bottleneck features of the downscaled inputs. As observed in the error maps with the UHD features in the third column, downscaled input produced in a model-aware manner $(x_{c})$ maintains a higher level of representative consistency with the UHD input, as perceived by the model. + +the low-resolution space, which is computed as: + +$$ +x _ {c} = \operatorname {C o n v} ([ x _ {d}; D (\widehat {x}; [ h, w ]) ]). \tag {9} +$$ + +Here, we concatenate the downscaled UHD image and downscaled compensatory information, then perform channel reduction by convolution layer. In this way, the newly derived $x_{c}$ can be optimized under the guidance of model knowledge, where we utilize the compensatory information to bridge the resizer and enhancer. + +# 3.5. Optimization Object + +Given the pre-trained enhancer, the bottleneck features of the UHD input image possess the most powerful representational capability compared to all of its downscaled counterparts. Therefore, we encourage the consistency of bottleneck features, which is referred to as resampling consistency loss: + +$$ +\mathcal {L} _ {r c} = \operatorname {S m o o t h L 1} (E (x) - U (E \left(x _ {c}\right))), \tag {10} +$$ + +where the $E(x)$ and $E(x_{c})$ represent the corresponding bottleneck features of the UHD input $x$ and learnable down-scaled input $x_{c}$ . $U(\cdot)$ is used for size adjustment. In addition, we introduce attention loss [37] for these two bottleneck features, given as: + +$$ +\mathcal {L} _ {f a} = \left\| \frac {E (x)}{\| E (x) \| _ {2}} - \frac {U \left(E \left(x _ {c}\right)\right)}{\| U \left(E \left(x _ {c}\right)\right) \| _ {2}} \right\| _ {2}, \tag {11} +$$ + +where $\| \cdot \| _2$ is the $\mathcal{L}_2$ normalization and this loss encourages the attention transfer from $E(x)$ to $U(E(x_{c}))$ . + +The aforementioned losses are designed to incorporate model-favored information into the learnable downscaled input image $x_{c}$ but neglect the distribution consistency between the naive downscaled images $x_{d}$ and newly derived images $x_{c}$ . Thus, we need to impose restrictions on the + +Table 1. Quantitative results on the UHD-LOL4K datasets with CNN-I as the backbone. The results with LMAR are shown in gray with better results highlighted in bold. + +
MethodCNN-I (UHD-LOL4K)
Scales(2160, 3840)(1440, 2560)(1080, 1920)(1200, 1600)(720, 1280)(540, 960)(432, 768)
Nearest31.92 / 0.984328.20 / 0.981527.01 / 0.979226.66 / 0.976325.86 / 0.976123.99 / 0.961523.06 / 0.9445
31.86 / 0.984128.46 / 0.982427.69 / 0.981327.42 / 0.979026.60 / 0.980624.38 / 0.971523.98 / 0.9608
Bilinear31.92 / 0.984331.07 / 0.983929.75 / 0.982829.16 / 0.982027.01 / 0.976725.90 / 0.974424.03 / 0.9448
31.88 / 0.984131.28 / 0.984030.17 / 0.983529.85 / 0.983227.94 / 0.981426.27 / 0.979524.85 / 0.9613
Bicubic31.92 / 0.984331.19 / 0.984029.82 / 0.983429.26 / 0.982227.16 / 0.976725.40 / 0.969424.04 / 0.9450
31.89 / 0.984231.42 / 0.984130.20 / 0.983929.96 / 0.983627.97 / 0.981526.36 / 0.979224.88 / 0.9617
Lanczos231.92 / 0.984331.04 / 0.984029.83 / 0.983429.22 / 0.981926.97 / 0.976125.39 / 0.969424.03 / 0.9450
31.88 / 0.984131.39 / 0.984130.19 / 0.983829.97 / 0.983427.90 / 0.981526.29 / 0.978924.87 / 0.9617
Lanczos331.92 / 0.984330.86 / 0.983829.08 / 0.981528.88 / 0.981327.10 / 0.976425.04 / 0.966323.86 / 0.9438
31.90 / 0.984231.59 / 0.984230.24 / 0.983630.13 / 0.983427.93 / 0.981725.79 / 0.971824.66 / 0.9611
Time (ms)9.79.613.313.414.815.415.0
10.710.213.313.815.115.615.4
+ +![](images/d58b29bd97dcbeff9ae8c5a7a15c91f5034b1e344eaa0ded7731dedb54039240.jpg) +(a) Input +Figure 7. Qualitative results on the UHD-LOL4K dataset of different scales. The top row is obtained from the cubic operator without LMAR, while the bottom row is with LMAR. Please zoom in for details. + +![](images/37ddaff8a6b67033470bf78130bd2dd64710b9de7bbc10d618b31f1c9eaedb44.jpg) +(b) $\times 5$ + +![](images/68ae2f9151fd7ac58b63a14b8ad90d1fd2411921787a80bba76d8ca97719715b.jpg) +(c) $\times 4$ + +![](images/ea6548538423be7c55c9315ad3ac97411de436e82b619618d29f76b2961bf71b.jpg) +(d) $\times 3$ + +![](images/356762e0677b95d010cadcd0860f5044d4346594890b06e5bd6efb7548d1a57b.jpg) +(e) $\times 2$ + +![](images/cab9aafb8a7e31bdf52c03983d868d623942a532c24bd881ebc21b421ba21552.jpg) +$(\mathrm{f})\times 1.5$ + +![](images/187165f31835c1d6584b9866bad3bf3172e6de7eca1597eea320e74caccd91d7.jpg) +$(\mathrm{g})\times 1$ + +![](images/7fa57e3df1902bb6b1cef9e7657a5975611274c15f331eb810567e8c4215a284.jpg) +(h) GT + +$x_{c}$ to ensure its similarity to $x_{d}$ . However, measuring the distances in the pixel space deteriorates the learning of the above process. Inspired by [7], we add a discriminator and impose $x_{c}$ and $x_{d}$ to confound it, known as adversarial loss, formulated as: + +$$ +\mathcal {L} _ {d} = \log \left(D \left(x _ {d}\right)\right) + \log \left(1 - D \left(x _ {c}\right)\right), \tag {12} +$$ + +where $D(\cdot)$ is the discriminator, which ensures the interaction between the resizer and enhancer while maintaining distribution consistency between $x_{c}$ and $x_{d}$ . In addition, to account for the introduction of the discriminator, a new generation loss is incorporated to constrain $x_{c}$ , computed as: + +$$ +\mathcal {L} _ {g} = \log \left(D \left(x _ {c}\right)\right). \tag {13} +$$ + +Overall, our optimization objects are twofold. The first optimization focus is on correlating the resizer and enhancer: + +$$ +\mathcal {L} _ {c} = \mathcal {L} _ {r c} + \alpha \mathcal {L} _ {f a} + \beta \mathcal {L} _ {g}, \tag {14} +$$ + +where $\alpha$ and $\beta$ are adjustable parameters to balance different losses, which are set empirically. The second optimization object is $\mathcal{L}_d$ to restrict the distribution. These two optimization objects are alternatively optimized. As can be + +seen in Figures 5 and 6, our downscaled images not only maintain consistency in terms of perceptual distribution but also exhibit a more consistent representation, as perceived by the model. + +# 4. Experiments + +# 4.1. Experimental Setup + +Datasets. To evaluate the significance of correlating the enhancer and resizer, experiments are conducted on two UHD low-light enhancement datasets, including the UHD-LOL4K dataset proposed by [33] and the 4KIL dataset proposed by [24]. More details about datasets are provided in the supplementary materials. + +Pre-training. We begin by training multiple enhancers $F(\cdot)$ to verify the effectiveness of our method. Specifically, we employ the widely used encoder-decoder architecture in image restoration tasks as the backbone. Regarding the UHD-LOL4K dataset, we leverage the invertible block proposed by [25] as the basic unit for processing, termed CNN-I. For the 4KIL dataset, we replace the basic processing unit with the half instance normalization block proposed + +Table 2. Quantitative results on the 4KIL datasets with CNN-H as the backbone. The results with LMAR are shown in gray with better results highlighted in bold. + +
MethodCNN-H (4KIL)
Scales(2160, 3840)(1440, 2560)(1080, 1920)(1200, 1600)(720, 1280)(540, 960)(432, 768)
Nearest29.80 / 0.980528.36 / 0.977927.71 / 0.975427.67 / 0.974226.98 / 0.971925.92 / 0.961625.37 / 0.9428
29.77 / 0.980128.53 / 0.978928.11 / 0.977927.93 / 0.976227.21 / 0.974526.16 / 0.967225.56 / 0.9493
Bilinear29.80 / 0.980529.12 / 0.979628.47 / 0.978528.39 / 0.977927.30 / 0.972626.38 / 0.965925.56 / 0.9439
29.79 / 0.980429.14 / 0.980428.55 / 0.979428.51 / 0.979027.52 / 0.975026.57 / 0.968725.62 / 0.9464
Bicubic29.80 / 0.980529.25 / 0.979728.60 / 0.978728.47 / 0.978027.30 / 0.972626.36 / 0.965025.58 / 0.9444
29.80 / 0.980429.35 / 0.980428.74 / 0.979728.96 / 0.979327.53 / 0.975226.57 / 0.968525.79 / 0.9509
Lanczos229.80 / 0.980529.24 / 0.979728.59 / 0.978728.46 / 0.978027.29 / 0.972626.35 / 0.964925.56 / 0.9444
29.80 / 0.980329.34 / 0.980528.72 / 0.979628.63 / 0.979227.54 / 0.975226.57 / 0.968425.78 / 0.9509
Lanczos329.80 / 0.980529.32/ 0.979828.69 / 0.979828.54 / 0.978127.34 / 0.973026.34 / 0.964525.58 / 0.9457
29.79 / 0.980329.48 / 0.980428.89 / 0.979928.76 / 0.979427.60 / 0.975626.58 / 0.968625.80 / 0.9523
Time (ms)5.05.05.86.36.56.96.9
5.95.76.16.56.88.58.0
+ +Table 3. Quantitative results on the 4KIL datasets with Restormer as the backbone. The results with LMAR are shown in gray with better results highlighted in bold. + +
MethodRestromer (4KIL)
Scales(2160, 3840)(1440, 2560)(1080, 1920)(1200, 1600)(720, 1280)(540, 960)(432, 768)
Nearest28.86 / 0.977327.81 / 0.975027.57 / 0.973827.45 / 0.972026.99 / 0.971425.91 / 0.960825.31 / 0.9470
28.78 / 0.976828.36 / 0.976527.67 / 0.974227.80 / 0.973927.27 / 0.973325.98 / 0.961925.46 / 0.9527
Bilinear28.86 / 0.977328.63 / 0.977028.50 / 0.976828.49 / 0.976727.56 / 0.971926.71 / 0.968625.71 / 0.9459
28.83 / 0.977028.75 / 0.977128.52 / 0.976928.64 / 0.977027.85 / 0.973926.77 / 0.968825.90 / 0.9531
Bicubic28.86 / 0.977328.83 / 0.977128.64 / 0.977028.53 / 0.976627.57 / 0.971926.65 / 0.966825.79 / 0.9477
28.84 / 0.977228.84 / 0.977228.69 / 0.977128.76 / 0.977127.82 / 0.973826.71 / 0.967825.94 / 0.9533
Lanczos228.86 / 0.977328.83 / 0.977028.63 / 0.976928.52 / 0.976627.56 / 0.971926.65 / 0.966825.78 / 0.9476
28.83 / 0.977228.84 / 0.977128.67 / 0.977028.74 / 0.977327.81 / 0.973926.70 / 0.967825.93 / 0.9533
Lanczos328.86 / 0.977328.84 / 0.977028.68 / 0.976828.55 / 0.976527.56 / 0.971926.56 / 0.965225.72 / 0.9468
28.85 / 0.977228.85 / 0.977328.84 / 0.977128.85 / 0.977227.84 / 0.974026.58 / 0.966525.89 / 0.9528
Time (ms)11.911.512.511.712.011.712.3
12.812.912.913.012.212.612.5
+ +by [3] to demonstrate the robustness of our method, termed CNN-H. In addition to the above CNN models, we also train a transformer-based enhancer conditioned on Restormer, as proposed by [38], on the 4KIL dataset to further validate the scalability of our method. More details about pre-training are provided in the supplementary materials. + +# 4.2. Collaborative Training + +In this stage, we keep the pre-trained enhancers fixed and train to correlate the resizer and enhancer. Considering the randomness of scale variability in the resampling process, we incorporate random scale training. Specifically, we begin by cropping a $H_{p} \times W_{p}$ patch from the UHD input image. Next, we randomly sample a scale factor $r$ from a uniform distribution $\mathcal{U}(1,4)$ . Finally, we downsample the high-resolution $H_{p} \times W_{p}$ patch by the scale factor $r$ to generate a low-resolution $h_{p} \times w_{p}$ patch. The optimization objectives mentioned earlier are applied to these two resolution patches. + +Table 4. Comparison of model parameters. Baseline models with LMAR are shown in gray, where the extra parameters are trainable parameters possessed by LMAR. + +
ModelCNN-ICNN-HRestormer
#Params1.009MB1.128MB1.660MB
1.314MB1.495MB1.965MB
+ +Implementation Details. We conduct experiments based on the PyTorch framework along with the tiny-cuda-nn [27] library with one NVIDIA 3090 GPU. During the collaborative training process, the Adam [21] optimizer with $\beta_{1} = 0.9$ and $\beta_{2} = 0.999$ is used in our experiments. A batch size of 1 is utilized for all experiments. The cropped patch size $(H_{p} \times W_{p})$ is set to $1024 \times 1024$ for the CNN enhancers, and $768 \times 768$ for the transformer enhancer. For the CNN-I and CNN-H, the collaborative training epoch is set to 12 and 24, respectively. The initial learning rate is $4e^{-4}$ , which decays by a factor value of 0.75 every 4 epochs and 8 epochs for the CNN-I and CNN-H enhancers. As for + +![](images/cf7df67d306a02489cf38f1f11027c5268559e0eb87dbf8eaf3e7cb9b339103a.jpg) +(a) Input + +![](images/8003faa17f193bf480bfd2eb50764ec53f8d809021b577077222fd50af8ea5f1.jpg) +(b) $\times 5$ + +![](images/fc46698ed5f88479ff83cb874e01e3c43b2b24584b29f55951d74a7f51413217.jpg) +(c)×4 + +![](images/fe6f5a25ef97440ebef75c8ef255903de2ebc32af2cc0db0d644274d3f96cb8a.jpg) +(d)×3 +Figure 8. Qualitative results on the 4KIL dataset of different scales. The top row is obtained from the lanzcos3 operator without LMAR, while the bottom row is with LMAR. Please zoom in for details. + +![](images/e832857fd3e6fe0638274e575c4e26ac7c16315857300943311848a7e13b667f.jpg) +(e)×2 + +![](images/d95033fe1b9124f124285d799ef4a1d875c09e97ebae88c15244f2e4bf3549f6.jpg) +(f) $\times 1.5$ + +![](images/a6720055a11619e8d182c78e8f4135edec30d3a85ffe6b9e772710176f5a3be7.jpg) +(g)×1 + +![](images/48d6835db7bd1a3fdc5c394d18204752e7918d0972153b5833127b49be6aff88.jpg) +(h) GT + +the transformer enhancer, the number of collaborative training epochs is 60, with a $4e^{-4}$ initial learning rate, which decays by a factor of 0.75 every 20 epochs. The discriminator is composed of several convolution layers, whose initial learning rate is $2e^{-4}$ and shares the same training strategies as mentioned above. + +# 4.3. Performance Comparison + +Quantitative Evaluation. We utilize two metrics, the Peak Signal to Noise Ratio (PSNR, dB), and the Structural Similarity (SSIM), to assess the performance gains. To verify the universality of our methods, we adopt five types of resizers and test them on various commonly used scales, including both in-distribution and out-of-distribution scales. As shown in Tables 1, 2, and 3, our method consistently achieves performance improvements across different resolutions, regardless of the choice of resizer or the category of the backbone. Notably, the performance with LMAR at UHD resolution is analogous to the original UHD input, which illustrates the superior continuous modeling capability of LMAR. Additionally, our approach maintains comparable effectiveness with the baselines, where a minor increase in parameters, as shown in Table 4. More importantly, our algorithm only needs to be trained on one type of resizer, which can then be applied to any other resizers. + +Qualitative Evaluation. Figures 7 and 8 present the visual comparison on the UHD-LOL4K dataset and 4KIL dataset respectively. These comparisons demonstrate that enabling model-aware resampling leads to enhanced results with improved global lightness and local structural details, across various resampling scales. It verifies that coupling the resize-ers and enhancers could generate better visual results. More visual results are presented in the supplementary materials. + +# 4.4. Ablation Studies + +We conduct ablation studies on the UHD-LOL4K dataset to investigate the rationality of our designs. The ablation results are shown in Table 5. First, we replace the bottleneck features (Full) with the relatively front features $(F_{f})$ and relatively end features $(F_{e})$ to apply feature constraints, + +where remarkable performance drops can be observed at 4 times downsampling. The findings indicate that the bottleneck features exhibit the most representative capability in guiding model-aware resampling. Second, we replace the discriminator (Full) with the perceptual loss (VGG) to conduct the distribution restriction. The utilization of discriminator encourages more consistent feature alignment. Lastly, even with sub-optimal constraints, our method outperforms the baseline, emphasizing the effectiveness of coupling resizers and enhancers. + +Table 5. Ablation studies on the constraints employed in our designs. We present the PSNR results here. + +
ScalesBaselineFfFeVGGFull
x229.8229.9830.0530.0330.20
x425.4025.8625.5025.4726.36
+ +# 5. Conclusion + +In this paper, we introduce a novel approach for effectively enhancing the UHD images at different downsampling scales by a model-aware resampling strategy. Our method customizes resampling under the guidance of model knowledge, integrating the resizers and enhancers through compensatory information learning. Notably, our algorithm is fully compatible with existing interpolation resamplers and maintains the efficiency. Moreover, our design promotes performances at low resolutions remarkably while keeping comparable results at UHD resolution, without the need to retrain the enhancer. Extensive experiments have validated the effectiveness and scalability of our method. + +# 6. Acknowledgment + +This work was supported by the JKW Research Funds under Grant 20-163-14-LZ-001-004-01, and the Anhui Provincial Natural Science Foundation under Grant 2108085UD12. We acknowledge the support of GPU cluster built by MCC Lab of Information Science and Technology Institution, USTC. + +# References + +[1] Kenneth R Castleman. Digital image processing. Prentice Hall Press, 1996. 3 +[2] Jiawen Chen, Andrew Adams, Neal Wadhwa, and Samuel W Hasinoff. Bilateral guided upsampling. ACM Transactions on Graphics (TOG), 35(6):1-8, 2016. 3 +[3] Liangyu Chen, Xin Lu, Jie Zhang, Xiaojie Chu, and Chengpeng Chen. Hinet: Half instance normalization network for image restoration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, pages 182-192, 2021. 7 +[4] Yinbo Chen, Sifei Liu, and Xiaolong Wang. Learning continuous image representation with local implicit image function. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8628-8638, 2021. 3, 4 +[5] Weitao Feng, Deyi Ji, Yiru Wang, Shuorong Chang, Hansheng Ren, and Weihao Gan. Challenges on large scale surveillance video analysis. In Proceedings of the IEEE conference on computer vision and pattern recognition workshops, pages 69-76, 2018. 3 +[6] Michael Gharbi, Jiawen Chen, Jonathan T Barron, Samuel W Hasinoff, and Frédo Durand. Deep bilateral learning for real-time image enhancement. ACM Transactions on Graphics (TOG), 36(4):1-12, 2017. 3 +[7] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial networks. Communications of the ACM, 63(11):139-144, 2020. 6 +[8] Dianyuan Han. Comparison of commonly used image interpolation methods. In Conference of the 2nd International Conference on Computer Science and Electronics Engineering (ICCSEE 2013), pages 1556-1559. Atlantis Press, 2013. 3 +[9] Muhammad Haris, Gregory Shakhnarovich, and Norimichi Ukita. Deep back-projection networks for super-resolution. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1664-1673, 2018. 4 +[10] Hanzhe Hu, Deyi Ji, Weihao Gan, Shuai Bai, Wei Wu, and Junjie Yan. Class-wise dynamic graph convolution for semantic segmentation. In European Conference on Computer Vision, pages 1-17. Springer, 2020. 3 +[11] Xuecai Hu, Haoyuan Mu, Xiangyu Zhang, Zilei Wang, Tieniu Tan, and Jian Sun. Meta-sr: A magnification-arbitrary network for super-resolution. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1575-1584, 2019. 3 +[12] Max Jaderberg, Karen Simonyan, Andrew Zisserman, et al. Spatial transformer networks. Advances in neural information processing systems, 28, 2015. 2, 3 +[13] Deyi Ji, Hongtao Lu, and Tongzhen Zhang. End to end multiscale convolutional neural network for crowd counting. In Eleventh international conference on machine vision, pages 761-766. SPIE, 2019. 3 +[14] Deyi Ji, Haoran Wang, Hanzhe Hu, Weihao Gan, Wei Wu, and Junjie Yan. Context-aware graph convolution network + +for target re-identification. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1646–1654, 2021. +[15] Deyi Ji, Haoran Wang, Mingyuan Tao, Jianqiang Huang, Xian-Sheng Hua, and Hongtao Lu. Structural and statistical texture knowledge distillation for semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16876-16885, 2022. +[16] Deyi Ji, Siqi Gao, Mingyuan Tao, Hongtao Lu, and Feng Zhao. Changenet: Multi-temporal asymmetric change detection dataset. arXiv preprint arXiv:2312.17428, 2023. +[17] Deyi Ji, Feng Zhao, and Hongtao Lu. Guided patch-grouping wavelet transformer with spatial congruence for ultra-high resolution segmentation. International Joint Conference on Artificial Intelligence, 2023. +[18] Deyi Ji, Feng Zhao, Hongtao Lu, Mingyuan Tao, and Jieping Ye. Ultra-high resolution segmentation with ultra-rich context: A novel benchmark. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23621-23630, 2023. +[19] Deyi Ji, Siqi Gao, Lanyun Zhu, Yiru Zhao, Peng Xu, Hong-tao Lu, and Feng Zhao. View-centric multi-object tracking with homographic matching in moving UAV. arXiv preprint, 2024. 3 +[20] Heewon Kim, Myungsub Choi, Bee Lim, and Kyoung Mu Lee. Task-aware image downscaling. In Proceedings of the European conference on computer vision (ECCV), pages 399-414, 2018. 2, 3 +[21] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014.7 +[22] Yue Li, Dong Liu, Houqiang Li, Li Li, Zhu Li, and Feng Wu. Learning a convolutional neural network for image compact-resolution. IEEE Transactions on Image Processing, 28(3): 1092-1107, 2018. 2 +[23] Jie Liang, Hui Zeng, and Lei Zhang. High-resolution photorealistic image translation in real-time: A laplacian pyramid translation network. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9392-9400, 2021. 3 +[24] Qiaowanni Lin, Zhuoran Zheng, and Xiuyi Jia. Uhd low-light image enhancement via interpretable bilateral learning. Information Sciences, 2022. 6 +[25] Yang Liu, Zhenyue Qin, Saeed Anwar, Pan Ji, Dongwoo Kim, Sabrina Caldwell, and Tom Gedeon. Invertible denoising network: A light solution for real noise removal. arXiv preprint arXiv:2104.10546, 2021. 6 +[26] Don P Mitchell and Arun N Netravali. Reconstruction filters in computer-graphics. ACM Siggraph Computer Graphics, 22(4):221-228, 1988. 3 +[27] Thomas Müller. tiny-cuda-nn, 2021. 7 +[28] Adria Recasens, Petr Kellnhofer, Simon Stent, Wojciech Matusik, and Antonio Torralba. Learning to zoom: a saliency-based sampling layer for neural networks. In Proceedings of the European conference on computer vision (ECCV), pages 51-66, 2018. 2, 3 +[29] Assaf Shocher, Ben Feinstein, Niv Haim, and Michal Irani. From discrete to continuous convolution layers. arXiv preprint arXiv:2006.11120, 2020.4 + +[30] Wanjie Sun and Zhenzhong Chen. Learned image downscaling for upscaling using content adaptive resampler. IEEE Transactions on Image Processing, 29:4027-4040, 2020. 2, 3 +[31] Haoran Wang, Licheng Jiao, Fang Liu, Lingling Li, Xu Liu, Deyi Ji, and Weihao Gan. Ipgn: Interactiveness proposal graph network for human-object interaction detection. IEEE Transactions on Image Processing, 30:6583-6593, 2021. 3 +[32] Haoran Wang, Licheng Jiao, Fang Liu, Lingling Li, Xu Liu, Deyi Ji, and Weihao Gan. Learning social spatio-temporal relation graph in the wild and a video benchmark. IEEE Transactions on Neural Networks and Learning Systems, 2021. 3 +[33] Tao Wang, Kaihao Zhang, Tianrun Shen, Wenhan Luo, Bjorn Stenger, and Tong Lu. Ultra-high-definition low-light image enhancement: A benchmark and transformer-based method. arXiv preprint arXiv:2212.11548, 2022. 3, 6 +[34] Mingqing Xiao, Shuxin Zheng, Chang Liu, Yaolong Wang, Di He, Guolin Ke, Jiang Bian, Zhouchen Lin, and Tie-Yan Liu. Invertible image rescaling. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part I 16, pages 126-144. Springer, 2020. 3 +[35] Jinbo Xing, Wenbo Hu, Menghan Xia, and Tien-Tsin Wong. Scale-arbitrary invertible image downscaling. IEEE Transactions on Image Processing, 2023. 3 +[36] Wei Yu, Qi Zhu, Naishan Zheng, Jie Huang, Man Zhou, and Feng Zhao. Learning non-uniform-sampling for ultra-high-definition image enhancement. In Proceedings of the 31st ACM International Conference on Multimedia, pages 1412-1421, 2023. 3 +[37] Sergey Zagoruyko and Nikos Komodakis. Paying more attention to attention: Improving the performance of convolutional neural networks via attention transfer. arXiv preprint arXiv:1612.03928, 2016. 5 +[38] Syed Waqas Zamir, Aditya Arora, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, and Ming-Hsuan Yang. Restormer: Efficient transformer for high-resolution image restoration. In CVPR, 2022. 7 +[39] Yongbing Zhang, Debin Zhao, Jian Zhang, Ruiqin Xiong, and Wen Gao. Interpolation-dependent image downsampling. IEEE Transactions on Image Processing, 20(11): 3291-3296, 2011. 2 +[40] Zhuoran Zheng, Wenqi Ren, Xiaochun Cao, Xiaobin Hu, Tao Wang, Fenglong Song, and Xiuyi Jia. Ultra-high-definition image dehazing via multi-guided bilateral learning. In 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 16180-16189. IEEE, 2021. 3 +[41] Zhuoran Zheng, Wenqi Ren, Xiaochun Cao, Tao Wang, and Xiuyi Jia. Ultra-high-definition image hdr reconstruction via collaborative bilateral learning. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4449-4458, 2021. 3 +[42] Lanyun Zhu, Deyi Ji, Shiping Zhu, Weihao Gan, Wei Wu, and Junjie Yan. Learning statistical texture for semantic segmentation. In Proceedings of the IEEE/CVF Conference + +on Computer Vision and Pattern Recognition (CVPR), pages 12537-12546, 2021. 3 +[43] Lanyun Zhu, Tianrun Chen, Deyi Ji, Jieping Ye, and Jun Liu. Llafs: When large-language models meet few-shot segmentation. arXiv preprint arXiv:2311.16926, 2023. +[44] Lanyun Zhu, Deyi Ji, Tianrun Chen, Peng Xu, Jieping Ye, and Jun Liu. Ibid: Alleviating hallucinations in large vision-language models via image-biased decoding. arXiv preprint arXiv:2402.18476, 2024. +[45] Qi Zhu, Zeyu Xiao, Jie Huang, and Feng Zhao. Dast-net: Depth-aware spatio-temporal network for video deblurring. In 2022 IEEE International Conference on Multimedia and Expo (ICME), pages 1-6. IEEE, 2022. +[46] Qi Zhu, Naishan Zheng, Jie Huang, Man Zhou, Jinghao Zhang, and Feng Zhao. Learning spatio-temporal sharpness map for video deblurring. IEEE Transactions on Circuits and Systems for Video Technology, 2023. +[47] Qi Zhu, Man Zhou, Naishan Zheng, Chongyi Li, Jie Huang, and Feng Zhao. Exploring temporal frequency spectrum in deep video deblurring. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12428-12437, 2023. +[48] Qi Zhu, Jie Huang, Naishan Zheng, Hongzhi Gao, Chongyi Li, Yuan Xu, Feng Zhao, et al. Fouridown: Factoring down-sampling into shuffling and superposing. Advances in Neural Information Processing Systems, 36, 2024. 3 \ No newline at end of file diff --git a/2024/Empowering Resampling Operation for Ultra-High-Definition Image Enhancement with Model-Aware Guidance/images.zip b/2024/Empowering Resampling Operation for Ultra-High-Definition Image Enhancement with Model-Aware Guidance/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..5c6c63f7b11601965eade338f5720fb5c6c0af7c --- /dev/null +++ b/2024/Empowering Resampling Operation for Ultra-High-Definition Image Enhancement with Model-Aware Guidance/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:368ab7d874eab061d17f1b5ec6c6f9ba08ba18fcc750f30d14da9cda0051d486 +size 1083165 diff --git a/2024/Empowering Resampling Operation for Ultra-High-Definition Image Enhancement with Model-Aware Guidance/layout.json b/2024/Empowering Resampling Operation for Ultra-High-Definition Image Enhancement with Model-Aware Guidance/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..71d82f6dda097889f8de88321ee2ad4a7b5e27ff --- /dev/null +++ b/2024/Empowering Resampling Operation for Ultra-High-Definition Image Enhancement with Model-Aware Guidance/layout.json @@ -0,0 +1,11000 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 85, + 103, + 509, + 138 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 85, + 103, + 509, + 138 + ], + "spans": [ + { + "bbox": [ + 85, + 103, + 509, + 138 + ], + "type": "text", + "content": "Empowering Resampling Operation for Ultra-High-Definition Image Enhancement with Model-Aware Guidance" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 125, + 161, + 496, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 125, + 161, + 496, + 190 + ], + "spans": [ + { + "bbox": [ + 125, + 161, + 496, + 190 + ], + "type": "text", + "content": "Wei Yu,\\* Jie Huang,\\* Bing Li, Kaiwen Zheng, Qi Zhu, Man Zhou, Feng Zhao† \nUniversity of Science and Technology of China" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 191, + 561, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 191, + 561, + 205 + ], + "spans": [ + { + "bbox": [ + 53, + 191, + 561, + 205 + ], + "type": "text", + "content": "{patrick914y, hj0117,bing0123, kezh,zqcrafts, manman}@mail.ustc.edu.cn，fzhao956@ustc.edu.cn" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 48, + 226, + 113, + 278 + ], + "blocks": [ + { + "bbox": [ + 48, + 226, + 113, + 278 + ], + "lines": [ + { + "bbox": [ + 48, + 226, + 113, + 278 + ], + "spans": [ + { + "bbox": [ + 48, + 226, + 113, + 278 + ], + "type": "image", + "image_path": "1a5bd2092423b406acafa244151d833ba75749badadefc7ba13a96ee075c0614.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 108, + 281, + 188, + 289 + ], + "lines": [ + { + "bbox": [ + 108, + 281, + 188, + 289 + ], + "spans": [ + { + "bbox": [ + 108, + 281, + 188, + 289 + ], + "type": "text", + "content": "(a) Down-Enhancement-Up" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 112, + 227, + 184, + 278 + ], + "blocks": [ + { + "bbox": [ + 112, + 227, + 184, + 278 + ], + "lines": [ + { + "bbox": [ + 112, + 227, + 184, + 278 + ], + "spans": [ + { + "bbox": [ + 112, + 227, + 184, + 278 + ], + "type": "image", + "image_path": "588349239bdf26e4b63c890585679a3a59f70d6928bd7c396e4bdd5ad9ddec95.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 186, + 227, + 250, + 278 + ], + "blocks": [ + { + "bbox": [ + 186, + 227, + 250, + 278 + ], + "lines": [ + { + "bbox": [ + 186, + 227, + 250, + 278 + ], + "spans": [ + { + "bbox": [ + 186, + 227, + 250, + 278 + ], + "type": "image", + "image_path": "57ad2f614867adbb59a1d277191baa8e845b8e1a394d49956ee0640f4361b690.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 49, + 291, + 250, + 357 + ], + "blocks": [ + { + "bbox": [ + 49, + 291, + 250, + 357 + ], + "lines": [ + { + "bbox": [ + 49, + 291, + 250, + 357 + ], + "spans": [ + { + "bbox": [ + 49, + 291, + 250, + 357 + ], + "type": "image", + "image_path": "ec1ce28b3efdcc0349ecac11d3cb237058ee621b36a018375d4d579e8492afcf.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 136, + 358, + 161, + 366 + ], + "lines": [ + { + "bbox": [ + 136, + 358, + 161, + 366 + ], + "spans": [ + { + "bbox": [ + 136, + 358, + 161, + 366 + ], + "type": "text", + "content": "(b) Ours" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 262, + 227, + 326, + 277 + ], + "blocks": [ + { + "bbox": [ + 262, + 227, + 326, + 277 + ], + "lines": [ + { + "bbox": [ + 262, + 227, + 326, + 277 + ], + "spans": [ + { + "bbox": [ + 262, + 227, + 326, + 277 + ], + "type": "image", + "image_path": "e47c2c900f580925422897fe14d1eeca5fb205c328a80ddc29835e93d702e8a8.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 272, + 278, + 310, + 286 + ], + "lines": [ + { + "bbox": [ + 272, + 278, + 310, + 286 + ], + "spans": [ + { + "bbox": [ + 272, + 278, + 310, + 286 + ], + "type": "text", + "content": "PSNR: 25.59" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 334, + 227, + 398, + 277 + ], + "blocks": [ + { + "bbox": [ + 334, + 227, + 398, + 277 + ], + "lines": [ + { + "bbox": [ + 334, + 227, + 398, + 277 + ], + "spans": [ + { + "bbox": [ + 334, + 227, + 398, + 277 + ], + "type": "image", + "image_path": "5c97497c3d726a06d80ee57c6a2eb657c0ae9765837a8d269c3ee8d352959be1.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 343, + 278, + 380, + 286 + ], + "lines": [ + { + "bbox": [ + 343, + 278, + 380, + 286 + ], + "spans": [ + { + "bbox": [ + 343, + 278, + 380, + 286 + ], + "type": "text", + "content": "PSNR: 28.51" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 405, + 227, + 468, + 277 + ], + "blocks": [ + { + "bbox": [ + 405, + 227, + 468, + 277 + ], + "lines": [ + { + "bbox": [ + 405, + 227, + 468, + 277 + ], + "spans": [ + { + "bbox": [ + 405, + 227, + 468, + 277 + ], + "type": "image", + "image_path": "af1759eaa900b0b71b03a975c18a833fe40ab7afa420db4539b6ededada66984.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 416, + 278, + 454, + 285 + ], + "lines": [ + { + "bbox": [ + 416, + 278, + 454, + 285 + ], + "spans": [ + { + "bbox": [ + 416, + 278, + 454, + 285 + ], + "type": "text", + "content": "PSNR: 30.78" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 476, + 227, + 540, + 278 + ], + "blocks": [ + { + "bbox": [ + 476, + 227, + 540, + 278 + ], + "lines": [ + { + "bbox": [ + 476, + 227, + 540, + 278 + ], + "spans": [ + { + "bbox": [ + 476, + 227, + 540, + 278 + ], + "type": "image", + "image_path": "5065405d10d298974acbfb18cd7764d5be0d0d4d01b836e55174c755e44829ae.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 486, + 278, + 523, + 285 + ], + "lines": [ + { + "bbox": [ + 486, + 278, + 523, + 285 + ], + "spans": [ + { + "bbox": [ + 486, + 278, + 523, + 285 + ], + "type": "text", + "content": "PSNR: 31.05" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 262, + 289, + 545, + 357 + ], + "blocks": [ + { + "bbox": [ + 262, + 289, + 545, + 357 + ], + "lines": [ + { + "bbox": [ + 262, + 289, + 545, + 357 + ], + "spans": [ + { + "bbox": [ + 262, + 289, + 545, + 357 + ], + "type": "image", + "image_path": "9856432f41090743975fd662f70cc6401d7d0eb86905334b297cdae64fef9f13.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 272, + 358, + 310, + 366 + ], + "lines": [ + { + "bbox": [ + 272, + 358, + 310, + 366 + ], + "spans": [ + { + "bbox": [ + 272, + 358, + 310, + 366 + ], + "type": "text", + "content": "PSNR: 28.09" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 343, + 358, + 381, + 366 + ], + "lines": [ + { + "bbox": [ + 343, + 358, + 381, + 366 + ], + "spans": [ + { + "bbox": [ + 343, + 358, + 381, + 366 + ], + "type": "text", + "content": "PSNR: 29.56" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 416, + 358, + 454, + 366 + ], + "lines": [ + { + "bbox": [ + 416, + 358, + 454, + 366 + ], + "spans": [ + { + "bbox": [ + 416, + 358, + 454, + 366 + ], + "type": "text", + "content": "PSNR: 32.24" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 486, + 358, + 523, + 366 + ], + "lines": [ + { + "bbox": [ + 486, + 358, + 523, + 366 + ], + "spans": [ + { + "bbox": [ + 486, + 358, + 523, + 366 + ], + "type": "text", + "content": "PSNR: 33.16" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 375, + 547, + 409 + ], + "lines": [ + { + "bbox": [ + 46, + 375, + 547, + 409 + ], + "spans": [ + { + "bbox": [ + 46, + 375, + 547, + 409 + ], + "type": "text", + "content": "Figure 1. Comparison between previous Downsampling-Enhancement-Upsampling paradigm and our proposed method. (a) Previous works treat resampling operators and inner models as separate components, while (b) our method integrates them via empowering model-aware resampling. As shown on the right side, our approach achieves significant performance gains over various resampling scales." + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "bbox": [ + 143, + 425, + 192, + 437 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 425, + 192, + 437 + ], + "spans": [ + { + "bbox": [ + 143, + 425, + 192, + 437 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 45, + 452, + 289, + 680 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 452, + 289, + 680 + ], + "spans": [ + { + "bbox": [ + 45, + 452, + 289, + 680 + ], + "type": "text", + "content": "Image enhancement algorithms have made remarkable advancements in recent years, but directly applying them to Ultra-high-definition (UHD) images presents intractable computational overheads. Therefore, previous straightforward solutions employ resampling techniques to reduce the resolution by adopting a \"Downsampling-Enhancement-Upsampling\" processing paradigm. However, this paradigm disentangles the resampling operators and inner enhancement algorithms, which results in the loss of information that is favored by the model, further leading to sub-optimal outcomes. In this paper, we propose a novel method of Learning Model-Aware Resampling (LMAR), which learns to customize resampling by extracting model-aware information from the UHD input image, under the guidance of model knowledge. Specifically, our method consists of two core designs, namely compensatory kernel estimation and steganographic resampling. At the first stage, we dynamically predict compensatory kernels tailored to the specific input and resampling scales. At" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 306, + 426, + 547, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 426, + 547, + 582 + ], + "spans": [ + { + "bbox": [ + 306, + 426, + 547, + 582 + ], + "type": "text", + "content": "the second stage, the image-wise compensatory information is derived with the compensatory kernels and embedded into the rescaled input images. This promotes the representation of the newly derived downscaled inputs to be more consistent with the full-resolution UHD inputs, as perceived by the model. Our LMAR enables model-aware and model-favored resampling while maintaining compatibility with existing resampling operators. Extensive experiments on multiple UHD image enhancement datasets and different backbones have shown consistent performance gains after correlating resizer and enhancer, e.g., up to 1.2dB PSNR gain for " + }, + { + "bbox": [ + 306, + 426, + 547, + 582 + ], + "type": "inline_equation", + "content": "\\times 1.8" + }, + { + "bbox": [ + 306, + 426, + 547, + 582 + ], + "type": "text", + "content": " resampling scale on UHD-LOL4K. The code is available at https://github.com/YPPatrickW/LMAR." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 306, + 586, + 387, + 598 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 586, + 387, + 598 + ], + "spans": [ + { + "bbox": [ + 306, + 586, + 387, + 598 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 304, + 606, + 545, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 606, + 545, + 666 + ], + "spans": [ + { + "bbox": [ + 304, + 606, + 545, + 666 + ], + "type": "text", + "content": "Over time, learning-based image enhancement algorithms have achieved progressive performance improvements. However, the majority of existing methods are not compatible with UHD images due to the heavy computational burdens imposed by their megapixel counts." + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 305, + 666, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 666, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 305, + 666, + 547, + 715 + ], + "type": "text", + "content": "To address this, previous works have sought to alleviate computation overheads by employing downsampling to reduce the resolution, as shown in Figure 1. Concretely, UHD input images are often resampled to smaller sizes for" + } + ] + } + ], + "index": 30 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 57, + 693, + 220, + 704 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 693, + 220, + 704 + ], + "spans": [ + { + "bbox": [ + 57, + 693, + 220, + 704 + ], + "type": "text", + "content": "*Both authors contributed equally to this research." + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 59, + 704, + 135, + 714 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 704, + 135, + 714 + ], + "spans": [ + { + "bbox": [ + 59, + 704, + 135, + 714 + ], + "type": "text", + "content": "† Corresponding author." + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "25722" + } + ] + } + ], + "index": 34 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 48, + 70, + 289, + 185 + ], + "blocks": [ + { + "bbox": [ + 48, + 70, + 289, + 185 + ], + "lines": [ + { + "bbox": [ + 48, + 70, + 289, + 185 + ], + "spans": [ + { + "bbox": [ + 48, + 70, + 289, + 185 + ], + "type": "image", + "image_path": "4b768982962e32cc832c9c561b31d4e6530d9dd28d6965b8a93c755ab20382b3.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 194, + 287, + 304 + ], + "lines": [ + { + "bbox": [ + 46, + 194, + 287, + 304 + ], + "spans": [ + { + "bbox": [ + 46, + 194, + 287, + 304 + ], + "type": "text", + "content": "Figure 2. The influence of the front-end resampler. We choose the cubic resampling operator as the baseline and introduce perturbations to its kernel function. Interestingly, manipulating the resampling function can lead to volatile outcomes, with the image quality either being improved or degraded. We also present two more comparisons: the resampling residual of the UHD input image and the feature maps generated by different resizers. It is indicated that the response of the model is highly sensitive to the sampling strategies. This highlights the importance of customizing the resampling process in a model-aware manner." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 315, + 287, + 375 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 315, + 287, + 375 + ], + "spans": [ + { + "bbox": [ + 46, + 315, + 287, + 375 + ], + "type": "text", + "content": "enhancement, and then subsequently resampled back to the original resolution[22, 39]. Surprisingly, the majority of research efforts have been paid to developing inner processing algorithms, with little attention devoted to the foremost and endmost resampling operators." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 376, + 287, + 508 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 376, + 287, + 508 + ], + "spans": [ + { + "bbox": [ + 46, + 376, + 287, + 508 + ], + "type": "text", + "content": "Recently, a line of work has proposed the replacement of the conventional interpolation-based resampling operators (e.g., nearest, bilinear, bicubic) with learnable resampling operators [12, 20, 28, 30]. Attributing to the superior modeling capability of deep neural networks, these learnable resampling operators have demonstrated remarkable performance gains. However, the practical application of these learnable resampling operators is hindered by their high computational complexity and reduced efficiency, resulting in a continued preference for interpolation-based resampling operators." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 509, + 287, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 509, + 287, + 664 + ], + "spans": [ + { + "bbox": [ + 46, + 509, + 287, + 664 + ], + "type": "text", + "content": "Revisiting the above UHD image enhancement pipeline: Downsampling-Enhancement-Upsampling, with the internal enhancement algorithm remaining unchanged, a fundamental question arises: 'How significant is the impact of these two resampling operations?' The answer to this question is twofold. On the one hand, the quality of the upsampled enhanced results exhibits a decreasing trend as the degree of downsampling aggravates. On the other hand, manipulating the interpolation kernel functions yields a diverse range of outcomes, as illustrated in Figure 2. We attribute this phenomenon to the isolation between the resampling operators and the enhancement algorithm, where both components are disentangled from each other." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 665, + 287, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 287, + 712 + ], + "type": "text", + "content": "To this end, the focus of this paper lies in correlating the resampling operators within the internal enhancement algorithms and unleashing the potential of the resampling operation to make benefits for the aforementioned framework," + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 305, + 72, + 545, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 133 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 133 + ], + "type": "text", + "content": "without modifying the inner processing methods (e.g., pretrained networks). To accomplish this goal, we propose a novel method termed Learning Model-Aware Resampling (LMAR). LMAR correlates the resizer and the enhancer through a compensatory information learning process." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 154, + 546, + 464 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 154, + 546, + 464 + ], + "spans": [ + { + "bbox": [ + 304, + 154, + 546, + 464 + ], + "type": "text", + "content": "We begin by delving into the impact of resampling from the viewpoint of the enhancer, where the resampling operations directly undermine the representation ability of intermediate features. Therefore, the intermediate features of the UHD image offer valuable insights on how to perform resampling in a model-aware manner. Building upon this recognition, we formulate our method into two steps: Compensatory Kernel Estimation (CKE) and Steganographic Resampling (SR). In the CKE step, we generate imagespecific convolution kernels for the input UHD image via implicit neural representation. These kernels are specifically customized for each input and resampling scale. Then, we can obtain compensatory information produced by these predicted filters. Subsequently, in the SR step, the compensatory information is embedded in the downscaled input image, resulting in a newly learnable downscaled input. By utilizing the compensatory information as a medium, we establish the correlation between the resampling operators and the inner enhancer. Finally, the learnable input is passed to the enhancer and optimized to encourage that the intermediate features align more consistently with those of the UHD image. Thanks to these designs, our method not only maintains compatibility with existing interpolation-based resampling operators but also incorporates them into a part of enhancement algorithms under the guidance of model knowledge. Our contributions are summarized as:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 486, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 306, + 486, + 545, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 486, + 545, + 521 + ], + "spans": [ + { + "bbox": [ + 306, + 486, + 545, + 521 + ], + "type": "text", + "content": "- We rethink the disentanglement between the resizer and enhancement model and propose to perform the resampling in a model-aware manner." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 522, + 545, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 522, + 545, + 582 + ], + "spans": [ + { + "bbox": [ + 306, + 522, + 545, + 582 + ], + "type": "text", + "content": "- We introduce LMAR, a novel method that facilitates the collaboration between resizer and enhancer via customizing resampling under the guidance of model knowledge, where the representative capabilities of low-resolution images are largely improved." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 582, + 545, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 582, + 545, + 641 + ], + "spans": [ + { + "bbox": [ + 306, + 582, + 545, + 641 + ], + "type": "text", + "content": "- Our LMAR delivers two core designs: compensatory kernel estimation and steganographic resampling, which models the interplay between responder and enhancer by learning embeddable information. Our method is compatible with any interpolation resamplers." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 642, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 642, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 306, + 642, + 545, + 712 + ], + "type": "text", + "content": "- Extensive experiments are conducted to verify the superiority of correlating resizer and enhancer. Remarkably, our algorithm significantly improves the performance at low resolutions while maintaining equivalent results at the original resolution, without requiring retraining of the enhancer." + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "25723" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 48, + 68, + 289, + 199 + ], + "blocks": [ + { + "bbox": [ + 48, + 68, + 289, + 199 + ], + "lines": [ + { + "bbox": [ + 48, + 68, + 289, + 199 + ], + "spans": [ + { + "bbox": [ + 48, + 68, + 289, + 199 + ], + "type": "image", + "image_path": "a999215e919be8d6ab60f20228b1e4e496f972e2b3d45178b34cf3393b9edf1b.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 206, + 287, + 293 + ], + "lines": [ + { + "bbox": [ + 46, + 206, + 287, + 293 + ], + "spans": [ + { + "bbox": [ + 46, + 206, + 287, + 293 + ], + "type": "text", + "content": "Figure 3. The inspiration for our method design. We replace the compressed bottleneck features of the downsampled input with features extracted from the UHD image (The green arrow). Large-margin performance gains can be achieved compared with the vested paradigm (The gray arrow). This observation highlights the potential role of feature consistency between UHD input and any downscaled input, which can play as a bridge for establishing correlations between resampling operators and enhancers." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 303, + 134, + 316 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 303, + 134, + 316 + ], + "spans": [ + { + "bbox": [ + 47, + 303, + 134, + 316 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 323, + 157, + 336 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 323, + 157, + 336 + ], + "spans": [ + { + "bbox": [ + 47, + 323, + 157, + 336 + ], + "type": "text", + "content": "2.1. Image Resampling" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 341, + 287, + 400 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 341, + 287, + 400 + ], + "spans": [ + { + "bbox": [ + 46, + 341, + 287, + 400 + ], + "type": "text", + "content": "Image resampling has emerged as one of the most commonly employed techniques for altering the resolution of an image. The key to image resampling lies in effectively preserving the quality of image content while maintaining efficiency." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 401, + 287, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 401, + 287, + 472 + ], + "spans": [ + { + "bbox": [ + 46, + 401, + 287, + 472 + ], + "type": "text", + "content": "Interpolation operators, such as bilinear, and bicubic [1, 8], have been the popular choice for image resampling for many years. These operators are capable of effectively resampling images at any scale, yet they disregard the spatial variations of distinct image patches, leading to inadequate preservation of the local structure [26]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 473, + 287, + 580 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 473, + 287, + 580 + ], + "spans": [ + { + "bbox": [ + 46, + 473, + 287, + 580 + ], + "type": "text", + "content": "In recent years, deep neural networks have facilitated advancements in various tasks [5, 10, 13-19, 31, 32, 36, 42-48], including image resampling [4, 11, 12, 20, 28, 30, 34, 35]. Task-aware image downsampling (TAD) [20] presents an encoder-decoder framework that aims to enhance the quality of both downscaling and upscaling processes in a collaborative manner. Though these methods achieve excellent results, none of them focus on correlating interpolation resampler with a pre-trained model." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 587, + 179, + 601 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 587, + 179, + 601 + ], + "spans": [ + { + "bbox": [ + 47, + 587, + 179, + 601 + ], + "type": "text", + "content": "2.2. UHD Image Processing" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "type": "text", + "content": "A line of studies that specifically focus on high-resolution image processing has been proposed [2, 6, 23, 33, 40, 41]. These methods share a common paradigm: learning from downscaled images to alleviate the computational burdens. In this learning paradigm, the quality of the results is directly determined by the resampled image representation, highlighting the crucial role of resampling strategies. However, they often overlook the impact of resampling approaches and lack in-depth investigation of various resam" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "text", + "content": "pling operators. In this work, our objective is to explore the interplay between enhancer and resampling operators and establish a collaborative relationship for them to improve performance." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 129, + 388, + 143 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 129, + 388, + 143 + ], + "spans": [ + { + "bbox": [ + 306, + 129, + 388, + 143 + ], + "type": "text", + "content": "3. Methodology" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 149, + 545, + 245 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 149, + 545, + 245 + ], + "spans": [ + { + "bbox": [ + 304, + 149, + 545, + 245 + ], + "type": "text", + "content": "In this section, we will present a comprehensive illustration of our proposed LMAR. We begin by introducing the motivation behind our approach and describing our settings. Subsequently, we will overview the whole processing pipeline and delve into the design of the Compensatory Kernel Estimation (CKE) and Steganographic Resampling (SR). Finally, we will discuss the optimization strategies employed to enable model-aware resampling." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 251, + 436, + 263 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 251, + 436, + 263 + ], + "spans": [ + { + "bbox": [ + 306, + 251, + 436, + 263 + ], + "type": "text", + "content": "3.1. Setting and Motivation" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 269, + 545, + 364 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 269, + 545, + 364 + ], + "spans": [ + { + "bbox": [ + 304, + 269, + 545, + 364 + ], + "type": "text", + "content": "Our setting still follows the Downsampling-Enhancement-Upsampling paradigm, employing the interpolation-based resampling operators instead of introducing additional burdens with learnable resampling operators. Notably, we will first train an enhancement model for each dataset in our experimental setup and keep unchanged. The main difference lies in the collaboration process between the resizing and enhancer." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 365, + 545, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 365, + 545, + 496 + ], + "spans": [ + { + "bbox": [ + 304, + 365, + 545, + 496 + ], + "type": "text", + "content": "The main inspiration stems from observing the impact of resampling on the representative feature extraction process of the enhancer. As shown in Figure 3, we inject the full-resolution bottleneck features into the enhancement process of the downsampled input images, where the final results significantly outperform the vested paradigm. It proves that front-most resampling degrades the representative ability of bottleneck features and impacts the following processes. Based on the insight, we correlate the resizer and enhancer through a compensatory information-learning process to achieve consistent feature alignment." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 502, + 373, + 514 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 502, + 373, + 514 + ], + "spans": [ + { + "bbox": [ + 306, + 502, + 373, + 514 + ], + "type": "text", + "content": "3.2. Overview" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 521, + 545, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 521, + 545, + 544 + ], + "spans": [ + { + "bbox": [ + 304, + 521, + 545, + 544 + ], + "type": "text", + "content": "The Downsampling-Enhancement-Upsampling processing can be mathematically formulated as:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 383, + 550, + 462, + 563 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 383, + 550, + 462, + 563 + ], + "spans": [ + { + "bbox": [ + 383, + 550, + 462, + 563 + ], + "type": "interline_equation", + "content": "x _ {d} = D (x; [ h, w ]),", + "image_path": "77a7d24f3fa3817e04159400af97dd19c743ddf15fd2fbc45af553238837da7f.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 389, + 565, + 545, + 578 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 389, + 565, + 545, + 578 + ], + "spans": [ + { + "bbox": [ + 389, + 565, + 545, + 578 + ], + "type": "interline_equation", + "content": "\\tilde {y} = F \\left(x _ {d}\\right), \\tag {1}", + "image_path": "53fb964f7a883820de71c8bc0516124ed565d57bc32abfe2ab81793ae9936144.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 389, + 580, + 468, + 594 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 389, + 580, + 468, + 594 + ], + "spans": [ + { + "bbox": [ + 389, + 580, + 468, + 594 + ], + "type": "interline_equation", + "content": "y = U (\\tilde {y}; [ H, W ]),", + "image_path": "c2d795b122d37bc0cfde09ae2908cd6c218f6b98f458dce1481bebe400805bd7.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 598, + 545, + 682 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 598, + 545, + 682 + ], + "spans": [ + { + "bbox": [ + 304, + 598, + 545, + 682 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 598, + 545, + 682 + ], + "type": "inline_equation", + "content": "D(\\cdot)" + }, + { + "bbox": [ + 304, + 598, + 545, + 682 + ], + "type": "text", + "content": " is the down sampler that resamples the UHD input image " + }, + { + "bbox": [ + 304, + 598, + 545, + 682 + ], + "type": "inline_equation", + "content": "x \\in \\mathbb{R}^{H \\times W \\times 3}" + }, + { + "bbox": [ + 304, + 598, + 545, + 682 + ], + "type": "text", + "content": " to lower resolution of " + }, + { + "bbox": [ + 304, + 598, + 545, + 682 + ], + "type": "inline_equation", + "content": "x_{d} \\in \\mathbb{R}^{h \\times w \\times 3}" + }, + { + "bbox": [ + 304, + 598, + 545, + 682 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 598, + 545, + 682 + ], + "type": "inline_equation", + "content": "F(\\cdot)" + }, + { + "bbox": [ + 304, + 598, + 545, + 682 + ], + "type": "text", + "content": " is the pre-trained enhancement model, and " + }, + { + "bbox": [ + 304, + 598, + 545, + 682 + ], + "type": "inline_equation", + "content": "U(\\cdot)" + }, + { + "bbox": [ + 304, + 598, + 545, + 682 + ], + "type": "text", + "content": " denotes the upsampler which resamples the low-resolution enhanced result " + }, + { + "bbox": [ + 304, + 598, + 545, + 682 + ], + "type": "inline_equation", + "content": "\\tilde{y}" + }, + { + "bbox": [ + 304, + 598, + 545, + 682 + ], + "type": "text", + "content": " back to the UHD resolution of " + }, + { + "bbox": [ + 304, + 598, + 545, + 682 + ], + "type": "inline_equation", + "content": "H \\times W" + }, + { + "bbox": [ + 304, + 598, + 545, + 682 + ], + "type": "text", + "content": ", termed as " + }, + { + "bbox": [ + 304, + 598, + 545, + 682 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 304, + 598, + 545, + 682 + ], + "type": "text", + "content": ". In LMAR, we rebuild the above paradigm, given as:" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 369, + 687, + 545, + 706 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 369, + 687, + 545, + 706 + ], + "spans": [ + { + "bbox": [ + 369, + 687, + 545, + 706 + ], + "type": "interline_equation", + "content": "\\tilde {y} = F \\left(x _ {d}, \\Theta (x; D, U, F)\\right), \\tag {2}", + "image_path": "00db4395e436e7d55638112bbc03a7061f43dc69695fe6661ac0845c97567b3e.jpg" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 370, + 704, + 447, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 370, + 704, + 447, + 715 + ], + "spans": [ + { + "bbox": [ + 370, + 704, + 447, + 715 + ], + "type": "interline_equation", + "content": "y = U (\\tilde {y}; [ H, W ]).", + "image_path": "cf7caced65f6c5aa89cbe6ec6f739d65c4c59ef5827d6184132996e99ae0a3a9.jpg" + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "25724" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 89, + 326, + 201 + ], + "blocks": [ + { + "bbox": [ + 57, + 76, + 122, + 87 + ], + "lines": [ + { + "bbox": [ + 57, + 76, + 122, + 87 + ], + "spans": [ + { + "bbox": [ + 57, + 76, + 122, + 87 + ], + "type": "text", + "content": "1. Training Stage" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 56, + 89, + 326, + 201 + ], + "lines": [ + { + "bbox": [ + 56, + 89, + 326, + 201 + ], + "spans": [ + { + "bbox": [ + 56, + 89, + 326, + 201 + ], + "type": "image", + "image_path": "01ad7bbe2229e30ed77c4945b98e2edd3255d39b3569939d760f107929f0459f.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 335, + 89, + 542, + 196 + ], + "blocks": [ + { + "bbox": [ + 335, + 76, + 409, + 87 + ], + "lines": [ + { + "bbox": [ + 335, + 76, + 409, + 87 + ], + "spans": [ + { + "bbox": [ + 335, + 76, + 409, + 87 + ], + "type": "text", + "content": "2. Inference Stage" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 335, + 89, + 542, + 196 + ], + "lines": [ + { + "bbox": [ + 335, + 89, + 542, + 196 + ], + "spans": [ + { + "bbox": [ + 335, + 89, + 542, + 196 + ], + "type": "image", + "image_path": "a5164a9d7600e4e3125278c151df765d3ecf52d49d6eb923f406ce623047717c.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 435, + 201, + 446, + 211 + ], + "lines": [ + { + "bbox": [ + 435, + 201, + 446, + 211 + ], + "spans": [ + { + "bbox": [ + 435, + 201, + 446, + 211 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 50, + 215, + 416, + 335 + ], + "blocks": [ + { + "bbox": [ + 195, + 202, + 206, + 213 + ], + "lines": [ + { + "bbox": [ + 195, + 202, + 206, + 213 + ], + "spans": [ + { + "bbox": [ + 195, + 202, + 206, + 213 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 50, + 215, + 416, + 335 + ], + "lines": [ + { + "bbox": [ + 50, + 215, + 416, + 335 + ], + "spans": [ + { + "bbox": [ + 50, + 215, + 416, + 335 + ], + "type": "image", + "image_path": "d3ee7d979e37be6c5a7a1570df5a9130f3d38ec79b0c57b80f132495d3d286ac.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 418, + 216, + 545, + 335 + ], + "blocks": [ + { + "bbox": [ + 418, + 216, + 545, + 335 + ], + "lines": [ + { + "bbox": [ + 418, + 216, + 545, + 335 + ], + "spans": [ + { + "bbox": [ + 418, + 216, + 545, + 335 + ], + "type": "image", + "image_path": "900be3f80c5813429b86065c569562b523d9eb1fed3b0b29657625088a47802e.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 346, + 547, + 401 + ], + "lines": [ + { + "bbox": [ + 46, + 346, + 547, + 401 + ], + "spans": [ + { + "bbox": [ + 46, + 346, + 547, + 401 + ], + "type": "text", + "content": "Figure 4. Overview of the proposed LMAR. The sub-graph (a) depicts the training phase of LMAR, which encourages the compensated low-resolution input to maintain representation consistency with the full-resolution UHD input, as perceived by the enhancer. The subgraph (b) demonstrates the inference pipeline of our LMAR, where the compensated low-resolution input is directly fed into the enhancer and then upsampled to the UHD result. The sub-graph (c) illustrates how LMAR works, where the core lies in estimating compensatory kernels under the guidance of model knowledge to make up for the resampling process." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 410, + 288, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 410, + 288, + 483 + ], + "spans": [ + { + "bbox": [ + 46, + 410, + 288, + 483 + ], + "type": "text", + "content": "Here, we introduce additional compensatory information as a medium to facilitate model-aware resampling, denoted as " + }, + { + "bbox": [ + 46, + 410, + 288, + 483 + ], + "type": "inline_equation", + "content": "\\Theta (x;D,U,F)" + }, + { + "bbox": [ + 46, + 410, + 288, + 483 + ], + "type": "text", + "content": ". Our method is presented in Figure 4, where the model-aware resampling is achieved via two core designs, compensatory kernel estimation, and steganographic resampling." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 498, + 227, + 511 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 498, + 227, + 511 + ], + "spans": [ + { + "bbox": [ + 47, + 498, + 227, + 511 + ], + "type": "text", + "content": "3.3. Compensatory Kernel Estimation" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 521, + 287, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 521, + 287, + 653 + ], + "spans": [ + { + "bbox": [ + 46, + 521, + 287, + 653 + ], + "type": "text", + "content": "Interpolation-based image resampling process can be decomposed as projected grid calculation and weighted pixel aggregation [29]. To this end, we estimate the compensatory kernel using a representation that is related to the scale. Inspired by LIFF [4], we employ two scale features, the relative coordinate grid and the pixel cell. The relative coordinate grid depicts the pixel location shift in the resampling processing. We first compute the uniform coordinate of the downsampling scale " + }, + { + "bbox": [ + 46, + 521, + 287, + 653 + ], + "type": "inline_equation", + "content": "C_d \\in \\mathbb{R}^{h \\times w \\times 2}" + }, + { + "bbox": [ + 46, + 521, + 287, + 653 + ], + "type": "text", + "content": " and the UHD scale " + }, + { + "bbox": [ + 46, + 521, + 287, + 653 + ], + "type": "inline_equation", + "content": "C_u \\in \\mathbb{R}^{H \\times W \\times 2}" + }, + { + "bbox": [ + 46, + 521, + 287, + 653 + ], + "type": "text", + "content": ", which are normalized between " + }, + { + "bbox": [ + 46, + 521, + 287, + 653 + ], + "type": "inline_equation", + "content": "[-1, 1]" + }, + { + "bbox": [ + 46, + 521, + 287, + 653 + ], + "type": "text", + "content": ". The calculations of the above grids are given as:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 73, + 667, + 287, + 696 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 73, + 667, + 287, + 696 + ], + "spans": [ + { + "bbox": [ + 73, + 667, + 287, + 696 + ], + "type": "interline_equation", + "content": "C _ {d} (i, j) = (- 1 + \\frac {(2 i + 1)}{h}, - 1 + \\frac {(2 j + 1)}{w}), \\tag {3}", + "image_path": "4b212b9dc07de7b06bd0a5d7339d6d7f947247f914e7d6b7e9acc960b4c67cd0.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 74, + 693, + 260, + 717 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 74, + 693, + 260, + 717 + ], + "spans": [ + { + "bbox": [ + 74, + 693, + 260, + 717 + ], + "type": "interline_equation", + "content": "C _ {u} (i, j) = (- 1 + \\frac {(2 i + 1)}{H}, - 1 + \\frac {(2 j + 1)}{W}),", + "image_path": "8b1b3602b23af3d3129f994d8206f3832a93e823f397316c1b8e21aa2baa9169.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 409, + 545, + 470 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 409, + 545, + 470 + ], + "spans": [ + { + "bbox": [ + 306, + 409, + 545, + 470 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 306, + 409, + 545, + 470 + ], + "type": "inline_equation", + "content": "i \\in [0, h - 1] \\vee [0, H - 1], j \\in [0, w - 1] \\vee [0, W - 1]" + }, + { + "bbox": [ + 306, + 409, + 545, + 470 + ], + "type": "text", + "content": " are the position indexes for the height and width dimensions. As the sizes of the two coordinates are different, we first project them " + }, + { + "bbox": [ + 306, + 409, + 545, + 470 + ], + "type": "inline_equation", + "content": "C_d" + }, + { + "bbox": [ + 306, + 409, + 545, + 470 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 306, + 409, + 545, + 470 + ], + "type": "inline_equation", + "content": "C_u" + }, + { + "bbox": [ + 306, + 409, + 545, + 470 + ], + "type": "text", + "content": " and obtain the relative grid shift at the downsampling scale, which is expressed as:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 361, + 474, + 477, + 488 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 361, + 474, + 477, + 488 + ], + "spans": [ + { + "bbox": [ + 361, + 474, + 477, + 488 + ], + "type": "interline_equation", + "content": "\\tilde {C} _ {d} = \\operatorname {g r i d} \\text {- s a m p l e} \\left(C _ {d}, C _ {u}\\right),", + "image_path": "5df2ea7ce5f29aaf3a321a0cc3905194fcf9cf43e3cadfeeb262daa9420ea70d.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 362, + 491, + 545, + 506 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 362, + 491, + 545, + 506 + ], + "spans": [ + { + "bbox": [ + 362, + 491, + 545, + 506 + ], + "type": "interline_equation", + "content": "\\tilde {C} _ {r} = C _ {u} - \\tilde {C} _ {d}, \\tag {4}", + "image_path": "7cebe993cafc01c3c214d2509c2f415084388a31cc8805186cbea1426fe7a5a8.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 362, + 508, + 489, + 522 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 362, + 508, + 489, + 522 + ], + "spans": [ + { + "bbox": [ + 362, + 508, + 489, + 522 + ], + "type": "interline_equation", + "content": "C _ {r} = \\left(\\tilde {C} _ {r} (i) \\times h, \\tilde {C} _ {r} (j) \\times w\\right),", + "image_path": "56278454931cb15490488deb611f79983a8252103d38d19a612354650135abc2.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 526, + 545, + 597 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 526, + 545, + 597 + ], + "spans": [ + { + "bbox": [ + 304, + 526, + 545, + 597 + ], + "type": "text", + "content": "where the grid_sample is a remapping function that remaps the " + }, + { + "bbox": [ + 304, + 526, + 545, + 597 + ], + "type": "inline_equation", + "content": "C_d" + }, + { + "bbox": [ + 304, + 526, + 545, + 597 + ], + "type": "text", + "content": " into " + }, + { + "bbox": [ + 304, + 526, + 545, + 597 + ], + "type": "inline_equation", + "content": "\\tilde{C}_d" + }, + { + "bbox": [ + 304, + 526, + 545, + 597 + ], + "type": "text", + "content": " of size " + }, + { + "bbox": [ + 304, + 526, + 545, + 597 + ], + "type": "inline_equation", + "content": "H \\times W" + }, + { + "bbox": [ + 304, + 526, + 545, + 597 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 526, + 545, + 597 + ], + "type": "inline_equation", + "content": "\\tilde{C}_r" + }, + { + "bbox": [ + 304, + 526, + 545, + 597 + ], + "type": "text", + "content": " represents the relative shift and " + }, + { + "bbox": [ + 304, + 526, + 545, + 597 + ], + "type": "inline_equation", + "content": "C_r" + }, + { + "bbox": [ + 304, + 526, + 545, + 597 + ], + "type": "text", + "content": " denotes the relative coordinate grid measured at downsampling scale. The pixel cell represents the ratio of pixel area changes between downsampling and upsampling, denoted as:" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 367, + 601, + 545, + 625 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 367, + 601, + 545, + 625 + ], + "spans": [ + { + "bbox": [ + 367, + 601, + 545, + 625 + ], + "type": "interline_equation", + "content": "P _ {c} (i, j) = \\left(2 * \\frac {h}{H}, 2 * \\frac {w}{W}\\right), \\tag {5}", + "image_path": "1f32201e04ddd2ed6ce0d452131d13a582f96a1d28083469344b7a89d28dc45a.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 629, + 545, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 629, + 545, + 652 + ], + "spans": [ + { + "bbox": [ + 304, + 629, + 545, + 652 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 629, + 545, + 652 + ], + "type": "inline_equation", + "content": "P_{c}\\in \\mathbb{R}^{H\\times W\\times 2}" + }, + { + "bbox": [ + 304, + 629, + 545, + 652 + ], + "type": "text", + "content": ", and all pixel locations share the same scale ratio." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "content": "The aforementioned scale-related representations are instance-same at a given scale, leading to the generation of identical kernels without considering the distinctions between images. Thus, we propose the utilization of backprojection [9] as a constraint to capture the interrelation" + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "25725" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 49, + 70, + 124, + 174 + ], + "blocks": [ + { + "bbox": [ + 49, + 70, + 124, + 174 + ], + "lines": [ + { + "bbox": [ + 49, + 70, + 124, + 174 + ], + "spans": [ + { + "bbox": [ + 49, + 70, + 124, + 174 + ], + "type": "image", + "image_path": "5a71842bcd4ce590e6d4f429e2a04ff9705d6a101bbcdba719ff25e08352c242.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 74, + 176, + 97, + 186 + ], + "lines": [ + { + "bbox": [ + 74, + 176, + 97, + 186 + ], + "spans": [ + { + "bbox": [ + 74, + 176, + 97, + 186 + ], + "type": "text", + "content": "(a) " + }, + { + "bbox": [ + 74, + 176, + 97, + 186 + ], + "type": "inline_equation", + "content": "x_{d}" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 130, + 70, + 206, + 174 + ], + "blocks": [ + { + "bbox": [ + 130, + 70, + 206, + 174 + ], + "lines": [ + { + "bbox": [ + 130, + 70, + 206, + 174 + ], + "spans": [ + { + "bbox": [ + 130, + 70, + 206, + 174 + ], + "type": "image", + "image_path": "4d7591c13ae2781df508dd572e0804e77da7711bce344a22fecfe30ca143ca3f.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 155, + 175, + 179, + 186 + ], + "lines": [ + { + "bbox": [ + 155, + 175, + 179, + 186 + ], + "spans": [ + { + "bbox": [ + 155, + 175, + 179, + 186 + ], + "type": "text", + "content": "(b) " + }, + { + "bbox": [ + 155, + 175, + 179, + 186 + ], + "type": "inline_equation", + "content": "x_{c}" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 212, + 70, + 288, + 174 + ], + "blocks": [ + { + "bbox": [ + 212, + 70, + 288, + 174 + ], + "lines": [ + { + "bbox": [ + 212, + 70, + 288, + 174 + ], + "spans": [ + { + "bbox": [ + 212, + 70, + 288, + 174 + ], + "type": "image", + "image_path": "a7e30423361ebdce19e2d1e497ea1d2867d803195ef9465e17cbcca7cb10beae.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 223, + 175, + 274, + 186 + ], + "lines": [ + { + "bbox": [ + 223, + 175, + 274, + 186 + ], + "spans": [ + { + "bbox": [ + 223, + 175, + 274, + 186 + ], + "type": "text", + "content": "(c) " + }, + { + "bbox": [ + 223, + 175, + 274, + 186 + ], + "type": "inline_equation", + "content": "|x_{c} - x_{d}|" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 47, + 201, + 287, + 257 + ], + "lines": [ + { + "bbox": [ + 47, + 201, + 287, + 257 + ], + "spans": [ + { + "bbox": [ + 47, + 201, + 287, + 257 + ], + "type": "text", + "content": "Figure 5. Comparison of different downscaled input images. (a) The downscaled input images by the cubic resampling operator, (b) images derived from the cubic resampling operator in a model-aware manner, and (c) the residual between (a) and (b), where obvious differences can be witnessed in high-frequency areas." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 266, + 287, + 291 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 266, + 287, + 291 + ], + "spans": [ + { + "bbox": [ + 46, + 266, + 287, + 291 + ], + "type": "text", + "content": "between each downscaled input and its corresponding UHD counterpart, formulated as:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 97, + 300, + 287, + 313 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 97, + 300, + 287, + 313 + ], + "spans": [ + { + "bbox": [ + 97, + 300, + 287, + 313 + ], + "type": "interline_equation", + "content": "\\tilde {x} = \\operatorname {C o n v} (x - U (x _ {d}; [ H, W ]))), \\tag {6}", + "image_path": "b0fc9c8af4a12b5bea0c1f95357fbaf5174086a4572f3e9c814adf8999b64d98.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 322, + 287, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 322, + 287, + 346 + ], + "spans": [ + { + "bbox": [ + 46, + 322, + 287, + 346 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 322, + 287, + 346 + ], + "type": "inline_equation", + "content": "\\tilde{x} \\in \\mathbb{R}^{H \\times W \\times 3}" + }, + { + "bbox": [ + 46, + 322, + 287, + 346 + ], + "type": "text", + "content": " is the back-projection of UHD input image " + }, + { + "bbox": [ + 46, + 322, + 287, + 346 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 46, + 322, + 287, + 346 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 347, + 287, + 455 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 347, + 287, + 455 + ], + "spans": [ + { + "bbox": [ + 46, + 347, + 287, + 455 + ], + "type": "text", + "content": "Considering the pixel-wise aggregation nature of the resampling process, we model the generation of compensatory information by predicting the convolution kernels for the UHD input image instead of predicting pixel values for specific locations. Specifically, we flatten the relative coordinate " + }, + { + "bbox": [ + 46, + 347, + 287, + 455 + ], + "type": "inline_equation", + "content": "C_r" + }, + { + "bbox": [ + 46, + 347, + 287, + 455 + ], + "type": "text", + "content": ", pixel cell " + }, + { + "bbox": [ + 46, + 347, + 287, + 455 + ], + "type": "inline_equation", + "content": "P_c" + }, + { + "bbox": [ + 46, + 347, + 287, + 455 + ], + "type": "text", + "content": ", and back-projection " + }, + { + "bbox": [ + 46, + 347, + 287, + 455 + ], + "type": "inline_equation", + "content": "\\tilde{x}" + }, + { + "bbox": [ + 46, + 347, + 287, + 455 + ], + "type": "text", + "content": " then concatenate them. Afterward, we feed these concatenated features into a multi-layer perception (MLP) to predict compensatory kernels as follows:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 113, + 464, + 287, + 478 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 113, + 464, + 287, + 478 + ], + "spans": [ + { + "bbox": [ + 113, + 464, + 287, + 478 + ], + "type": "interline_equation", + "content": "K = M L P _ {\\phi} \\left(\\left[ C _ {r}; P _ {c}; \\tilde {x} \\right]\\right), \\tag {7}", + "image_path": "c2fcaf5586257549b9f68ae4c3d86390d903971956a3b62afd8b544c0289fb8f.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 47, + 487, + 287, + 558 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 487, + 287, + 558 + ], + "spans": [ + { + "bbox": [ + 47, + 487, + 287, + 558 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 487, + 287, + 558 + ], + "type": "inline_equation", + "content": "[\\cdot ]" + }, + { + "bbox": [ + 47, + 487, + 287, + 558 + ], + "type": "text", + "content": " denotes the channel-wise concatenation, and " + }, + { + "bbox": [ + 47, + 487, + 287, + 558 + ], + "type": "inline_equation", + "content": "K\\in" + }, + { + "bbox": [ + 47, + 487, + 287, + 558 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^{H\\times W\\times s}" + }, + { + "bbox": [ + 47, + 487, + 287, + 558 + ], + "type": "text", + "content": " represents the predicted compensatory kernels. The kernel for each pixel location is represented in the channel dimension " + }, + { + "bbox": [ + 47, + 487, + 287, + 558 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 47, + 487, + 287, + 558 + ], + "type": "text", + "content": ", consisting of the multiplication of the input channel, out channel, and kernel size, with all being three in our case." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 47, + 567, + 201, + 580 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 567, + 201, + 580 + ], + "spans": [ + { + "bbox": [ + 47, + 567, + 201, + 580 + ], + "type": "text", + "content": "3.4. Steganographic Resampling" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 46, + 585, + 287, + 621 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 585, + 287, + 621 + ], + "spans": [ + { + "bbox": [ + 46, + 585, + 287, + 621 + ], + "type": "text", + "content": "With the estimated compensatory kernel from the above process, we calculate the compensatory information for the input UHD image as follows:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 140, + 631, + 287, + 643 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 140, + 631, + 287, + 643 + ], + "spans": [ + { + "bbox": [ + 140, + 631, + 287, + 643 + ], + "type": "interline_equation", + "content": "\\widehat {x} = K \\otimes x, \\tag {8}", + "image_path": "ef01a70d134dc173670b36c2874d0e5b053797417e296be987912ee4b477a268.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 46, + 652, + 287, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 652, + 287, + 677 + ], + "spans": [ + { + "bbox": [ + 46, + 652, + 287, + 677 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 652, + 287, + 677 + ], + "type": "inline_equation", + "content": "\\widehat{x} \\in \\mathbb{R}^{H \\times W \\times 3}" + }, + { + "bbox": [ + 46, + 652, + 287, + 677 + ], + "type": "text", + "content": " is the generated UHD compensatory embedding, and " + }, + { + "bbox": [ + 46, + 652, + 287, + 677 + ], + "type": "inline_equation", + "content": "\\otimes" + }, + { + "bbox": [ + 46, + 652, + 287, + 677 + ], + "type": "text", + "content": " denotes the convolution operation." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 46, + 677, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 287, + 714 + ], + "type": "text", + "content": "To further maintain the intrinsic characteristics of the downscaled input images, we downscale the compensatory information and perform compensatory steganography in" + } + ] + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 307, + 70, + 384, + 121 + ], + "blocks": [ + { + "bbox": [ + 307, + 70, + 384, + 121 + ], + "lines": [ + { + "bbox": [ + 307, + 70, + 384, + 121 + ], + "spans": [ + { + "bbox": [ + 307, + 70, + 384, + 121 + ], + "type": "image", + "image_path": "cbccefd1c5f2b8da13056c8bfa704158bdd5f1e19ec2f65febf874bfa689d51f.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 389, + 70, + 465, + 121 + ], + "blocks": [ + { + "bbox": [ + 389, + 70, + 465, + 121 + ], + "lines": [ + { + "bbox": [ + 389, + 70, + 465, + 121 + ], + "spans": [ + { + "bbox": [ + 389, + 70, + 465, + 121 + ], + "type": "image", + "image_path": "d41d07451a105764a98e9d2340ea2da8e7b97f63b1bc72e83552cab7d39f231c.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 470, + 70, + 547, + 121 + ], + "blocks": [ + { + "bbox": [ + 470, + 70, + 547, + 121 + ], + "lines": [ + { + "bbox": [ + 470, + 70, + 547, + 121 + ], + "spans": [ + { + "bbox": [ + 470, + 70, + 547, + 121 + ], + "type": "image", + "image_path": "5ff6d4209e997f5cd29523555b0c59a245220e7a1a123c4e00efac79befb38fb.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 496, + 122, + 520, + 131 + ], + "lines": [ + { + "bbox": [ + 496, + 122, + 520, + 131 + ], + "spans": [ + { + "bbox": [ + 496, + 122, + 520, + 131 + ], + "type": "text", + "content": "Errors" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 307, + 132, + 384, + 182 + ], + "blocks": [ + { + "bbox": [ + 339, + 123, + 350, + 132 + ], + "lines": [ + { + "bbox": [ + 339, + 123, + 350, + 132 + ], + "spans": [ + { + "bbox": [ + 339, + 123, + 350, + 132 + ], + "type": "inline_equation", + "content": "x_{d}" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 307, + 132, + 384, + 182 + ], + "lines": [ + { + "bbox": [ + 307, + 132, + 384, + 182 + ], + "spans": [ + { + "bbox": [ + 307, + 132, + 384, + 182 + ], + "type": "image", + "image_path": "c4912282f7391e97ff6a710a8d743beec2d9b9cad447ed23be325295a1d5bec0.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 339, + 183, + 350, + 193 + ], + "lines": [ + { + "bbox": [ + 339, + 183, + 350, + 193 + ], + "spans": [ + { + "bbox": [ + 339, + 183, + 350, + 193 + ], + "type": "inline_equation", + "content": "x_{c}" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 389, + 132, + 465, + 182 + ], + "blocks": [ + { + "bbox": [ + 414, + 122, + 437, + 132 + ], + "lines": [ + { + "bbox": [ + 414, + 122, + 437, + 132 + ], + "spans": [ + { + "bbox": [ + 414, + 122, + 437, + 132 + ], + "type": "inline_equation", + "content": "E(x_{d})" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 389, + 132, + 465, + 182 + ], + "lines": [ + { + "bbox": [ + 389, + 132, + 465, + 182 + ], + "spans": [ + { + "bbox": [ + 389, + 132, + 465, + 182 + ], + "type": "image", + "image_path": "a0c3bbed12608d45557a1c1c7fb4a9fbf66f25cdd46d16835f2722094bdf7203.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 414, + 182, + 436, + 193 + ], + "lines": [ + { + "bbox": [ + 414, + 182, + 436, + 193 + ], + "spans": [ + { + "bbox": [ + 414, + 182, + 436, + 193 + ], + "type": "inline_equation", + "content": "E(x_{c})" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_caption" + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 470, + 132, + 547, + 183 + ], + "blocks": [ + { + "bbox": [ + 470, + 132, + 547, + 183 + ], + "lines": [ + { + "bbox": [ + 470, + 132, + 547, + 183 + ], + "spans": [ + { + "bbox": [ + 470, + 132, + 547, + 183 + ], + "type": "image", + "image_path": "410630a60511c85aaf12b17e1a05d907be443a8873753c2af0423aeb3ebbcc3e.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 496, + 183, + 520, + 191 + ], + "lines": [ + { + "bbox": [ + 496, + 183, + 520, + 191 + ], + "spans": [ + { + "bbox": [ + 496, + 183, + 520, + 191 + ], + "type": "text", + "content": "Errors" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 305, + 205, + 547, + 272 + ], + "lines": [ + { + "bbox": [ + 305, + 205, + 547, + 272 + ], + "spans": [ + { + "bbox": [ + 305, + 205, + 547, + 272 + ], + "type": "text", + "content": "Figure 6. Feature comparison of different downscaled input images. The second column represents the bottleneck features of the downscaled inputs. As observed in the error maps with the UHD features in the third column, downscaled input produced in a model-aware manner " + }, + { + "bbox": [ + 305, + 205, + 547, + 272 + ], + "type": "inline_equation", + "content": "(x_{c})" + }, + { + "bbox": [ + 305, + 205, + 547, + 272 + ], + "type": "text", + "content": " maintains a higher level of representative consistency with the UHD input, as perceived by the model." + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_caption" + } + ], + "index": 28 + }, + { + "bbox": [ + 306, + 281, + 500, + 294 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 281, + 500, + 294 + ], + "spans": [ + { + "bbox": [ + 306, + 281, + 500, + 294 + ], + "type": "text", + "content": "the low-resolution space, which is computed as:" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 359, + 303, + 545, + 317 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 359, + 303, + 545, + 317 + ], + "spans": [ + { + "bbox": [ + 359, + 303, + 545, + 317 + ], + "type": "interline_equation", + "content": "x _ {c} = \\operatorname {C o n v} ([ x _ {d}; D (\\widehat {x}; [ h, w ]) ]). \\tag {9}", + "image_path": "2ae3697e86339ebc88eeef006469558a81b2b021888d8433857ff46155f4da44.jpg" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 305, + 325, + 545, + 397 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 325, + 545, + 397 + ], + "spans": [ + { + "bbox": [ + 305, + 325, + 545, + 397 + ], + "type": "text", + "content": "Here, we concatenate the downscaled UHD image and downscaled compensatory information, then perform channel reduction by convolution layer. In this way, the newly derived " + }, + { + "bbox": [ + 305, + 325, + 545, + 397 + ], + "type": "inline_equation", + "content": "x_{c}" + }, + { + "bbox": [ + 305, + 325, + 545, + 397 + ], + "type": "text", + "content": " can be optimized under the guidance of model knowledge, where we utilize the compensatory information to bridge the resizer and enhancer." + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 306, + 404, + 425, + 418 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 404, + 425, + 418 + ], + "spans": [ + { + "bbox": [ + 306, + 404, + 425, + 418 + ], + "type": "text", + "content": "3.5. Optimization Object" + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 305, + 423, + 545, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 423, + 545, + 495 + ], + "spans": [ + { + "bbox": [ + 305, + 423, + 545, + 495 + ], + "type": "text", + "content": "Given the pre-trained enhancer, the bottleneck features of the UHD input image possess the most powerful representational capability compared to all of its downscaled counterparts. Therefore, we encourage the consistency of bottleneck features, which is referred to as resampling consistency loss:" + } + ] + } + ], + "index": 35 + }, + { + "bbox": [ + 347, + 504, + 545, + 517 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 347, + 504, + 545, + 517 + ], + "spans": [ + { + "bbox": [ + 347, + 504, + 545, + 517 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {r c} = \\operatorname {S m o o t h L 1} (E (x) - U (E \\left(x _ {c}\\right))), \\tag {10}", + "image_path": "75c6c89da93980b3b630b4467129624ab2f2fecb4c87610b549ac22488c2c05b.jpg" + } + ] + } + ], + "index": 36 + }, + { + "bbox": [ + 305, + 526, + 545, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 526, + 545, + 586 + ], + "spans": [ + { + "bbox": [ + 305, + 526, + 545, + 586 + ], + "type": "text", + "content": "where the " + }, + { + "bbox": [ + 305, + 526, + 545, + 586 + ], + "type": "inline_equation", + "content": "E(x)" + }, + { + "bbox": [ + 305, + 526, + 545, + 586 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 526, + 545, + 586 + ], + "type": "inline_equation", + "content": "E(x_{c})" + }, + { + "bbox": [ + 305, + 526, + 545, + 586 + ], + "type": "text", + "content": " represent the corresponding bottleneck features of the UHD input " + }, + { + "bbox": [ + 305, + 526, + 545, + 586 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 305, + 526, + 545, + 586 + ], + "type": "text", + "content": " and learnable down-scaled input " + }, + { + "bbox": [ + 305, + 526, + 545, + 586 + ], + "type": "inline_equation", + "content": "x_{c}" + }, + { + "bbox": [ + 305, + 526, + 545, + 586 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 305, + 526, + 545, + 586 + ], + "type": "inline_equation", + "content": "U(\\cdot)" + }, + { + "bbox": [ + 305, + 526, + 545, + 586 + ], + "type": "text", + "content": " is used for size adjustment. In addition, we introduce attention loss [37] for these two bottleneck features, given as:" + } + ] + } + ], + "index": 37 + }, + { + "bbox": [ + 346, + 594, + 545, + 622 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 346, + 594, + 545, + 622 + ], + "spans": [ + { + "bbox": [ + 346, + 594, + 545, + 622 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {f a} = \\left\\| \\frac {E (x)}{\\| E (x) \\| _ {2}} - \\frac {U \\left(E \\left(x _ {c}\\right)\\right)}{\\| U \\left(E \\left(x _ {c}\\right)\\right) \\| _ {2}} \\right\\| _ {2}, \\tag {11}", + "image_path": "bf3c533ff30e5133a748f357c30c2c57d66a6b9ce5d17564b6365f45dd8237dd.jpg" + } + ] + } + ], + "index": 38 + }, + { + "bbox": [ + 305, + 629, + 544, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 629, + 544, + 653 + ], + "spans": [ + { + "bbox": [ + 305, + 629, + 544, + 653 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 629, + 544, + 653 + ], + "type": "inline_equation", + "content": "\\| \\cdot \\| _2" + }, + { + "bbox": [ + 305, + 629, + 544, + 653 + ], + "type": "text", + "content": " is the " + }, + { + "bbox": [ + 305, + 629, + 544, + 653 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_2" + }, + { + "bbox": [ + 305, + 629, + 544, + 653 + ], + "type": "text", + "content": " normalization and this loss encourages the attention transfer from " + }, + { + "bbox": [ + 305, + 629, + 544, + 653 + ], + "type": "inline_equation", + "content": "E(x)" + }, + { + "bbox": [ + 305, + 629, + 544, + 653 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 305, + 629, + 544, + 653 + ], + "type": "inline_equation", + "content": "U(E(x_{c}))" + }, + { + "bbox": [ + 305, + 629, + 544, + 653 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 39 + }, + { + "bbox": [ + 305, + 654, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 654, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 654, + 545, + 713 + ], + "type": "text", + "content": "The aforementioned losses are designed to incorporate model-favored information into the learnable downscaled input image " + }, + { + "bbox": [ + 305, + 654, + 545, + 713 + ], + "type": "inline_equation", + "content": "x_{c}" + }, + { + "bbox": [ + 305, + 654, + 545, + 713 + ], + "type": "text", + "content": " but neglect the distribution consistency between the naive downscaled images " + }, + { + "bbox": [ + 305, + 654, + 545, + 713 + ], + "type": "inline_equation", + "content": "x_{d}" + }, + { + "bbox": [ + 305, + 654, + 545, + 713 + ], + "type": "text", + "content": " and newly derived images " + }, + { + "bbox": [ + 305, + 654, + 545, + 713 + ], + "type": "inline_equation", + "content": "x_{c}" + }, + { + "bbox": [ + 305, + 654, + 545, + 713 + ], + "type": "text", + "content": ". Thus, we need to impose restrictions on the" + } + ] + } + ], + "index": 40 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "25726" + } + ] + } + ], + "index": 41 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 47, + 101, + 550, + 259 + ], + "blocks": [ + { + "bbox": [ + 47, + 70, + 547, + 94 + ], + "lines": [ + { + "bbox": [ + 47, + 70, + 547, + 94 + ], + "spans": [ + { + "bbox": [ + 47, + 70, + 547, + 94 + ], + "type": "text", + "content": "Table 1. Quantitative results on the UHD-LOL4K datasets with CNN-I as the backbone. The results with LMAR are shown in gray with better results highlighted in bold." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 47, + 101, + 550, + 259 + ], + "lines": [ + { + "bbox": [ + 47, + 101, + 550, + 259 + ], + "spans": [ + { + "bbox": [ + 47, + 101, + 550, + 259 + ], + "type": "table", + "html": "
MethodCNN-I (UHD-LOL4K)
Scales(2160, 3840)(1440, 2560)(1080, 1920)(1200, 1600)(720, 1280)(540, 960)(432, 768)
Nearest31.92 / 0.984328.20 / 0.981527.01 / 0.979226.66 / 0.976325.86 / 0.976123.99 / 0.961523.06 / 0.9445
31.86 / 0.984128.46 / 0.982427.69 / 0.981327.42 / 0.979026.60 / 0.980624.38 / 0.971523.98 / 0.9608
Bilinear31.92 / 0.984331.07 / 0.983929.75 / 0.982829.16 / 0.982027.01 / 0.976725.90 / 0.974424.03 / 0.9448
31.88 / 0.984131.28 / 0.984030.17 / 0.983529.85 / 0.983227.94 / 0.981426.27 / 0.979524.85 / 0.9613
Bicubic31.92 / 0.984331.19 / 0.984029.82 / 0.983429.26 / 0.982227.16 / 0.976725.40 / 0.969424.04 / 0.9450
31.89 / 0.984231.42 / 0.984130.20 / 0.983929.96 / 0.983627.97 / 0.981526.36 / 0.979224.88 / 0.9617
Lanczos231.92 / 0.984331.04 / 0.984029.83 / 0.983429.22 / 0.981926.97 / 0.976125.39 / 0.969424.03 / 0.9450
31.88 / 0.984131.39 / 0.984130.19 / 0.983829.97 / 0.983427.90 / 0.981526.29 / 0.978924.87 / 0.9617
Lanczos331.92 / 0.984330.86 / 0.983829.08 / 0.981528.88 / 0.981327.10 / 0.976425.04 / 0.966323.86 / 0.9438
31.90 / 0.984231.59 / 0.984230.24 / 0.983630.13 / 0.983427.93 / 0.981725.79 / 0.971824.66 / 0.9611
Time (ms)9.79.613.313.414.815.415.0
10.710.213.313.815.115.615.4
", + "image_path": "f43d798d3b19c208acf55c398c581c0b6a109303373bb87722b141483baa0b11.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 49, + 269, + 108, + 380 + ], + "blocks": [ + { + "bbox": [ + 49, + 269, + 108, + 380 + ], + "lines": [ + { + "bbox": [ + 49, + 269, + 108, + 380 + ], + "spans": [ + { + "bbox": [ + 49, + 269, + 108, + 380 + ], + "type": "image", + "image_path": "d58b29bd97dcbeff9ae8c5a7a15c91f5034b1e344eaa0ded7731dedb54039240.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 65, + 380, + 92, + 388 + ], + "lines": [ + { + "bbox": [ + 65, + 380, + 92, + 388 + ], + "spans": [ + { + "bbox": [ + 65, + 380, + 92, + 388 + ], + "type": "text", + "content": "(a) Input" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 403, + 547, + 425 + ], + "lines": [ + { + "bbox": [ + 46, + 403, + 547, + 425 + ], + "spans": [ + { + "bbox": [ + 46, + 403, + 547, + 425 + ], + "type": "text", + "content": "Figure 7. Qualitative results on the UHD-LOL4K dataset of different scales. The top row is obtained from the cubic operator without LMAR, while the bottom row is with LMAR. Please zoom in for details." + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 111, + 269, + 171, + 380 + ], + "blocks": [ + { + "bbox": [ + 111, + 269, + 171, + 380 + ], + "lines": [ + { + "bbox": [ + 111, + 269, + 171, + 380 + ], + "spans": [ + { + "bbox": [ + 111, + 269, + 171, + 380 + ], + "type": "image", + "image_path": "37ddaff8a6b67033470bf78130bd2dd64710b9de7bbc10d618b31f1c9eaedb44.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 129, + 380, + 153, + 388 + ], + "lines": [ + { + "bbox": [ + 129, + 380, + 153, + 388 + ], + "spans": [ + { + "bbox": [ + 129, + 380, + 153, + 388 + ], + "type": "text", + "content": "(b) " + }, + { + "bbox": [ + 129, + 380, + 153, + 388 + ], + "type": "inline_equation", + "content": "\\times 5" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 173, + 269, + 234, + 380 + ], + "blocks": [ + { + "bbox": [ + 173, + 269, + 234, + 380 + ], + "lines": [ + { + "bbox": [ + 173, + 269, + 234, + 380 + ], + "spans": [ + { + "bbox": [ + 173, + 269, + 234, + 380 + ], + "type": "image", + "image_path": "68ae2f9151fd7ac58b63a14b8ad90d1fd2411921787a80bba76d8ca97719715b.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 192, + 380, + 215, + 388 + ], + "lines": [ + { + "bbox": [ + 192, + 380, + 215, + 388 + ], + "spans": [ + { + "bbox": [ + 192, + 380, + 215, + 388 + ], + "type": "text", + "content": "(c) " + }, + { + "bbox": [ + 192, + 380, + 215, + 388 + ], + "type": "inline_equation", + "content": "\\times 4" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 235, + 269, + 296, + 380 + ], + "blocks": [ + { + "bbox": [ + 235, + 269, + 296, + 380 + ], + "lines": [ + { + "bbox": [ + 235, + 269, + 296, + 380 + ], + "spans": [ + { + "bbox": [ + 235, + 269, + 296, + 380 + ], + "type": "image", + "image_path": "ea6548538423be7c55c9315ad3ac97411de436e82b619618d29f76b2961bf71b.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 254, + 380, + 277, + 388 + ], + "lines": [ + { + "bbox": [ + 254, + 380, + 277, + 388 + ], + "spans": [ + { + "bbox": [ + 254, + 380, + 277, + 388 + ], + "type": "text", + "content": "(d) " + }, + { + "bbox": [ + 254, + 380, + 277, + 388 + ], + "type": "inline_equation", + "content": "\\times 3" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 298, + 269, + 359, + 380 + ], + "blocks": [ + { + "bbox": [ + 298, + 269, + 359, + 380 + ], + "lines": [ + { + "bbox": [ + 298, + 269, + 359, + 380 + ], + "spans": [ + { + "bbox": [ + 298, + 269, + 359, + 380 + ], + "type": "image", + "image_path": "356762e0677b95d010cadcd0860f5044d4346594890b06e5bd6efb7548d1a57b.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 317, + 380, + 340, + 388 + ], + "lines": [ + { + "bbox": [ + 317, + 380, + 340, + 388 + ], + "spans": [ + { + "bbox": [ + 317, + 380, + 340, + 388 + ], + "type": "text", + "content": "(e) " + }, + { + "bbox": [ + 317, + 380, + 340, + 388 + ], + "type": "inline_equation", + "content": "\\times 2" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 360, + 269, + 421, + 380 + ], + "blocks": [ + { + "bbox": [ + 360, + 269, + 421, + 380 + ], + "lines": [ + { + "bbox": [ + 360, + 269, + 421, + 380 + ], + "spans": [ + { + "bbox": [ + 360, + 269, + 421, + 380 + ], + "type": "image", + "image_path": "cab9aafb8a7e31bdf52c03983d868d623942a532c24bd881ebc21b421ba21552.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 377, + 380, + 405, + 388 + ], + "lines": [ + { + "bbox": [ + 377, + 380, + 405, + 388 + ], + "spans": [ + { + "bbox": [ + 377, + 380, + 405, + 388 + ], + "type": "inline_equation", + "content": "(\\mathrm{f})\\times 1.5" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 423, + 269, + 484, + 380 + ], + "blocks": [ + { + "bbox": [ + 423, + 269, + 484, + 380 + ], + "lines": [ + { + "bbox": [ + 423, + 269, + 484, + 380 + ], + "spans": [ + { + "bbox": [ + 423, + 269, + 484, + 380 + ], + "type": "image", + "image_path": "187165f31835c1d6584b9866bad3bf3172e6de7eca1597eea320e74caccd91d7.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 443, + 380, + 465, + 388 + ], + "lines": [ + { + "bbox": [ + 443, + 380, + 465, + 388 + ], + "spans": [ + { + "bbox": [ + 443, + 380, + 465, + 388 + ], + "type": "inline_equation", + "content": "(\\mathrm{g})\\times 1" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 485, + 269, + 545, + 380 + ], + "blocks": [ + { + "bbox": [ + 485, + 269, + 545, + 380 + ], + "lines": [ + { + "bbox": [ + 485, + 269, + 545, + 380 + ], + "spans": [ + { + "bbox": [ + 485, + 269, + 545, + 380 + ], + "type": "image", + "image_path": "7fa57e3df1902bb6b1cef9e7657a5975611274c15f331eb810567e8c4215a284.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 503, + 380, + 525, + 388 + ], + "lines": [ + { + "bbox": [ + 503, + 380, + 525, + 388 + ], + "spans": [ + { + "bbox": [ + 503, + 380, + 525, + 388 + ], + "type": "text", + "content": "(h) GT" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "bbox": [ + 46, + 445, + 287, + 505 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 445, + 287, + 505 + ], + "spans": [ + { + "bbox": [ + 46, + 445, + 287, + 505 + ], + "type": "inline_equation", + "content": "x_{c}" + }, + { + "bbox": [ + 46, + 445, + 287, + 505 + ], + "type": "text", + "content": " to ensure its similarity to " + }, + { + "bbox": [ + 46, + 445, + 287, + 505 + ], + "type": "inline_equation", + "content": "x_{d}" + }, + { + "bbox": [ + 46, + 445, + 287, + 505 + ], + "type": "text", + "content": ". However, measuring the distances in the pixel space deteriorates the learning of the above process. Inspired by [7], we add a discriminator and impose " + }, + { + "bbox": [ + 46, + 445, + 287, + 505 + ], + "type": "inline_equation", + "content": "x_{c}" + }, + { + "bbox": [ + 46, + 445, + 287, + 505 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 445, + 287, + 505 + ], + "type": "inline_equation", + "content": "x_{d}" + }, + { + "bbox": [ + 46, + 445, + 287, + 505 + ], + "type": "text", + "content": " to confound it, known as adversarial loss, formulated as:" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 89, + 511, + 287, + 525 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 511, + 287, + 525 + ], + "spans": [ + { + "bbox": [ + 89, + 511, + 287, + 525 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {d} = \\log \\left(D \\left(x _ {d}\\right)\\right) + \\log \\left(1 - D \\left(x _ {c}\\right)\\right), \\tag {12}", + "image_path": "bf96aeb548086a2fb774c3eb18a421e9f4bc0e530d1b0120733bbee89ccf15a9.jpg" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 46, + 529, + 287, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 529, + 287, + 590 + ], + "spans": [ + { + "bbox": [ + 46, + 529, + 287, + 590 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 529, + 287, + 590 + ], + "type": "inline_equation", + "content": "D(\\cdot)" + }, + { + "bbox": [ + 46, + 529, + 287, + 590 + ], + "type": "text", + "content": " is the discriminator, which ensures the interaction between the resizer and enhancer while maintaining distribution consistency between " + }, + { + "bbox": [ + 46, + 529, + 287, + 590 + ], + "type": "inline_equation", + "content": "x_{c}" + }, + { + "bbox": [ + 46, + 529, + 287, + 590 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 529, + 287, + 590 + ], + "type": "inline_equation", + "content": "x_{d}" + }, + { + "bbox": [ + 46, + 529, + 287, + 590 + ], + "type": "text", + "content": ". In addition, to account for the introduction of the discriminator, a new generation loss is incorporated to constrain " + }, + { + "bbox": [ + 46, + 529, + 287, + 590 + ], + "type": "inline_equation", + "content": "x_{c}" + }, + { + "bbox": [ + 46, + 529, + 287, + 590 + ], + "type": "text", + "content": ", computed as:" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 128, + 596, + 287, + 609 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 128, + 596, + 287, + 609 + ], + "spans": [ + { + "bbox": [ + 128, + 596, + 287, + 609 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {g} = \\log \\left(D \\left(x _ {c}\\right)\\right). \\tag {13}", + "image_path": "3648ef7701c642e4bea390c3a0fba86c60896267113ee8019f58aa9c9aa08612.jpg" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 46, + 613, + 287, + 648 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 613, + 287, + 648 + ], + "spans": [ + { + "bbox": [ + 46, + 613, + 287, + 648 + ], + "type": "text", + "content": "Overall, our optimization objects are twofold. The first optimization focus is on correlating the resizer and enhancer:" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 111, + 650, + 287, + 663 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 111, + 650, + 287, + 663 + ], + "spans": [ + { + "bbox": [ + 111, + 650, + 287, + 663 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {c} = \\mathcal {L} _ {r c} + \\alpha \\mathcal {L} _ {f a} + \\beta \\mathcal {L} _ {g}, \\tag {14}", + "image_path": "8467cadef0d52c3cdc7ec764fb5f293c6e5957040371a7b8b6ca03b681c7857b.jpg" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "text", + "content": " are adjustable parameters to balance different losses, which are set empirically. The second optimization object is " + }, + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_d" + }, + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "text", + "content": " to restrict the distribution. These two optimization objects are alternatively optimized. As can be" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 304, + 445, + 546, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 445, + 546, + 495 + ], + "spans": [ + { + "bbox": [ + 304, + 445, + 546, + 495 + ], + "type": "text", + "content": "seen in Figures 5 and 6, our downscaled images not only maintain consistency in terms of perceptual distribution but also exhibit a more consistent representation, as perceived by the model." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 306, + 506, + 388, + 520 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 506, + 388, + 520 + ], + "spans": [ + { + "bbox": [ + 306, + 506, + 388, + 520 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 306, + 526, + 422, + 540 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 526, + 422, + 540 + ], + "spans": [ + { + "bbox": [ + 306, + 526, + 422, + 540 + ], + "type": "text", + "content": "4.1. Experimental Setup" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 304, + 545, + 545, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 545, + 545, + 617 + ], + "spans": [ + { + "bbox": [ + 304, + 545, + 545, + 617 + ], + "type": "text", + "content": "Datasets. To evaluate the significance of correlating the enhancer and resizer, experiments are conducted on two UHD low-light enhancement datasets, including the UHD-LOL4K dataset proposed by [33] and the 4KIL dataset proposed by [24]. More details about datasets are provided in the supplementary materials." + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 304, + 617, + 546, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 617, + 546, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 617, + 546, + 714 + ], + "type": "text", + "content": "Pre-training. We begin by training multiple enhancers " + }, + { + "bbox": [ + 304, + 617, + 546, + 714 + ], + "type": "inline_equation", + "content": "F(\\cdot)" + }, + { + "bbox": [ + 304, + 617, + 546, + 714 + ], + "type": "text", + "content": " to verify the effectiveness of our method. Specifically, we employ the widely used encoder-decoder architecture in image restoration tasks as the backbone. Regarding the UHD-LOL4K dataset, we leverage the invertible block proposed by [25] as the basic unit for processing, termed CNN-I. For the 4KIL dataset, we replace the basic processing unit with the half instance normalization block proposed" + } + ] + } + ], + "index": 30 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "type": "text", + "content": "25727" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 47, + 102, + 550, + 259 + ], + "blocks": [ + { + "bbox": [ + 46, + 70, + 547, + 94 + ], + "lines": [ + { + "bbox": [ + 46, + 70, + 547, + 94 + ], + "spans": [ + { + "bbox": [ + 46, + 70, + 547, + 94 + ], + "type": "text", + "content": "Table 2. Quantitative results on the 4KIL datasets with CNN-H as the backbone. The results with LMAR are shown in gray with better results highlighted in bold." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 47, + 102, + 550, + 259 + ], + "lines": [ + { + "bbox": [ + 47, + 102, + 550, + 259 + ], + "spans": [ + { + "bbox": [ + 47, + 102, + 550, + 259 + ], + "type": "table", + "html": "
MethodCNN-H (4KIL)
Scales(2160, 3840)(1440, 2560)(1080, 1920)(1200, 1600)(720, 1280)(540, 960)(432, 768)
Nearest29.80 / 0.980528.36 / 0.977927.71 / 0.975427.67 / 0.974226.98 / 0.971925.92 / 0.961625.37 / 0.9428
29.77 / 0.980128.53 / 0.978928.11 / 0.977927.93 / 0.976227.21 / 0.974526.16 / 0.967225.56 / 0.9493
Bilinear29.80 / 0.980529.12 / 0.979628.47 / 0.978528.39 / 0.977927.30 / 0.972626.38 / 0.965925.56 / 0.9439
29.79 / 0.980429.14 / 0.980428.55 / 0.979428.51 / 0.979027.52 / 0.975026.57 / 0.968725.62 / 0.9464
Bicubic29.80 / 0.980529.25 / 0.979728.60 / 0.978728.47 / 0.978027.30 / 0.972626.36 / 0.965025.58 / 0.9444
29.80 / 0.980429.35 / 0.980428.74 / 0.979728.96 / 0.979327.53 / 0.975226.57 / 0.968525.79 / 0.9509
Lanczos229.80 / 0.980529.24 / 0.979728.59 / 0.978728.46 / 0.978027.29 / 0.972626.35 / 0.964925.56 / 0.9444
29.80 / 0.980329.34 / 0.980528.72 / 0.979628.63 / 0.979227.54 / 0.975226.57 / 0.968425.78 / 0.9509
Lanczos329.80 / 0.980529.32/ 0.979828.69 / 0.979828.54 / 0.978127.34 / 0.973026.34 / 0.964525.58 / 0.9457
29.79 / 0.980329.48 / 0.980428.89 / 0.979928.76 / 0.979427.60 / 0.975626.58 / 0.968625.80 / 0.9523
Time (ms)5.05.05.86.36.56.96.9
5.95.76.16.56.88.58.0
", + "image_path": "a6c09525226d53e9ba60f76f9a8eab0b9f99dcb52c074f9f96f501ce2dfee28b.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 47, + 300, + 550, + 460 + ], + "blocks": [ + { + "bbox": [ + 46, + 270, + 547, + 293 + ], + "lines": [ + { + "bbox": [ + 46, + 270, + 547, + 293 + ], + "spans": [ + { + "bbox": [ + 46, + 270, + 547, + 293 + ], + "type": "text", + "content": "Table 3. Quantitative results on the 4KIL datasets with Restormer as the backbone. The results with LMAR are shown in gray with better results highlighted in bold." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 47, + 300, + 550, + 460 + ], + "lines": [ + { + "bbox": [ + 47, + 300, + 550, + 460 + ], + "spans": [ + { + "bbox": [ + 47, + 300, + 550, + 460 + ], + "type": "table", + "html": "
MethodRestromer (4KIL)
Scales(2160, 3840)(1440, 2560)(1080, 1920)(1200, 1600)(720, 1280)(540, 960)(432, 768)
Nearest28.86 / 0.977327.81 / 0.975027.57 / 0.973827.45 / 0.972026.99 / 0.971425.91 / 0.960825.31 / 0.9470
28.78 / 0.976828.36 / 0.976527.67 / 0.974227.80 / 0.973927.27 / 0.973325.98 / 0.961925.46 / 0.9527
Bilinear28.86 / 0.977328.63 / 0.977028.50 / 0.976828.49 / 0.976727.56 / 0.971926.71 / 0.968625.71 / 0.9459
28.83 / 0.977028.75 / 0.977128.52 / 0.976928.64 / 0.977027.85 / 0.973926.77 / 0.968825.90 / 0.9531
Bicubic28.86 / 0.977328.83 / 0.977128.64 / 0.977028.53 / 0.976627.57 / 0.971926.65 / 0.966825.79 / 0.9477
28.84 / 0.977228.84 / 0.977228.69 / 0.977128.76 / 0.977127.82 / 0.973826.71 / 0.967825.94 / 0.9533
Lanczos228.86 / 0.977328.83 / 0.977028.63 / 0.976928.52 / 0.976627.56 / 0.971926.65 / 0.966825.78 / 0.9476
28.83 / 0.977228.84 / 0.977128.67 / 0.977028.74 / 0.977327.81 / 0.973926.70 / 0.967825.93 / 0.9533
Lanczos328.86 / 0.977328.84 / 0.977028.68 / 0.976828.55 / 0.976527.56 / 0.971926.56 / 0.965225.72 / 0.9468
28.85 / 0.977228.85 / 0.977328.84 / 0.977128.85 / 0.977227.84 / 0.974026.58 / 0.966525.89 / 0.9528
Time (ms)11.911.512.511.712.011.712.3
12.812.912.913.012.212.612.5
", + "image_path": "60b002ea2323e7f5640a5347f72f0a4f34effb0971bc47f066d92c7ace3dccb0.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 479, + 288, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 479, + 288, + 552 + ], + "spans": [ + { + "bbox": [ + 46, + 479, + 288, + 552 + ], + "type": "text", + "content": "by [3] to demonstrate the robustness of our method, termed CNN-H. In addition to the above CNN models, we also train a transformer-based enhancer conditioned on Restormer, as proposed by [38], on the 4KIL dataset to further validate the scalability of our method. More details about pre-training are provided in the supplementary materials." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 562, + 178, + 575 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 562, + 178, + 575 + ], + "spans": [ + { + "bbox": [ + 47, + 562, + 178, + 575 + ], + "type": "text", + "content": "4.2. Collaborative Training" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 582, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 582, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 582, + 287, + 715 + ], + "type": "text", + "content": "In this stage, we keep the pre-trained enhancers fixed and train to correlate the resizer and enhancer. Considering the randomness of scale variability in the resampling process, we incorporate random scale training. Specifically, we begin by cropping a " + }, + { + "bbox": [ + 46, + 582, + 287, + 715 + ], + "type": "inline_equation", + "content": "H_{p} \\times W_{p}" + }, + { + "bbox": [ + 46, + 582, + 287, + 715 + ], + "type": "text", + "content": " patch from the UHD input image. Next, we randomly sample a scale factor " + }, + { + "bbox": [ + 46, + 582, + 287, + 715 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 46, + 582, + 287, + 715 + ], + "type": "text", + "content": " from a uniform distribution " + }, + { + "bbox": [ + 46, + 582, + 287, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{U}(1,4)" + }, + { + "bbox": [ + 46, + 582, + 287, + 715 + ], + "type": "text", + "content": ". Finally, we downsample the high-resolution " + }, + { + "bbox": [ + 46, + 582, + 287, + 715 + ], + "type": "inline_equation", + "content": "H_{p} \\times W_{p}" + }, + { + "bbox": [ + 46, + 582, + 287, + 715 + ], + "type": "text", + "content": " patch by the scale factor " + }, + { + "bbox": [ + 46, + 582, + 287, + 715 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 46, + 582, + 287, + 715 + ], + "type": "text", + "content": " to generate a low-resolution " + }, + { + "bbox": [ + 46, + 582, + 287, + 715 + ], + "type": "inline_equation", + "content": "h_{p} \\times w_{p}" + }, + { + "bbox": [ + 46, + 582, + 287, + 715 + ], + "type": "text", + "content": " patch. The optimization objectives mentioned earlier are applied to these two resolution patches." + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 323, + 518, + 526, + 559 + ], + "blocks": [ + { + "bbox": [ + 305, + 477, + 547, + 511 + ], + "lines": [ + { + "bbox": [ + 305, + 477, + 547, + 511 + ], + "spans": [ + { + "bbox": [ + 305, + 477, + 547, + 511 + ], + "type": "text", + "content": "Table 4. Comparison of model parameters. Baseline models with LMAR are shown in gray, where the extra parameters are trainable parameters possessed by LMAR." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 323, + 518, + 526, + 559 + ], + "lines": [ + { + "bbox": [ + 323, + 518, + 526, + 559 + ], + "spans": [ + { + "bbox": [ + 323, + 518, + 526, + 559 + ], + "type": "table", + "html": "
ModelCNN-ICNN-HRestormer
#Params1.009MB1.128MB1.660MB
1.314MB1.495MB1.965MB
", + "image_path": "321516f72cba637bdb49fcd3bf3ba59c6b74f80bfc1fd1fb895bea603cf31ab9.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 570, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 547, + 715 + ], + "type": "text", + "content": "Implementation Details. We conduct experiments based on the PyTorch framework along with the tiny-cuda-nn [27] library with one NVIDIA 3090 GPU. During the collaborative training process, the Adam [21] optimizer with " + }, + { + "bbox": [ + 304, + 570, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\beta_{1} = 0.9" + }, + { + "bbox": [ + 304, + 570, + 547, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 570, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\beta_{2} = 0.999" + }, + { + "bbox": [ + 304, + 570, + 547, + 715 + ], + "type": "text", + "content": " is used in our experiments. A batch size of 1 is utilized for all experiments. The cropped patch size " + }, + { + "bbox": [ + 304, + 570, + 547, + 715 + ], + "type": "inline_equation", + "content": "(H_{p} \\times W_{p})" + }, + { + "bbox": [ + 304, + 570, + 547, + 715 + ], + "type": "text", + "content": " is set to " + }, + { + "bbox": [ + 304, + 570, + 547, + 715 + ], + "type": "inline_equation", + "content": "1024 \\times 1024" + }, + { + "bbox": [ + 304, + 570, + 547, + 715 + ], + "type": "text", + "content": " for the CNN enhancers, and " + }, + { + "bbox": [ + 304, + 570, + 547, + 715 + ], + "type": "inline_equation", + "content": "768 \\times 768" + }, + { + "bbox": [ + 304, + 570, + 547, + 715 + ], + "type": "text", + "content": " for the transformer enhancer. For the CNN-I and CNN-H, the collaborative training epoch is set to 12 and 24, respectively. The initial learning rate is " + }, + { + "bbox": [ + 304, + 570, + 547, + 715 + ], + "type": "inline_equation", + "content": "4e^{-4}" + }, + { + "bbox": [ + 304, + 570, + 547, + 715 + ], + "type": "text", + "content": ", which decays by a factor value of 0.75 every 4 epochs and 8 epochs for the CNN-I and CNN-H enhancers. As for" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "25728" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 49, + 70, + 108, + 180 + ], + "blocks": [ + { + "bbox": [ + 49, + 70, + 108, + 180 + ], + "lines": [ + { + "bbox": [ + 49, + 70, + 108, + 180 + ], + "spans": [ + { + "bbox": [ + 49, + 70, + 108, + 180 + ], + "type": "image", + "image_path": "cf7df67d306a02489cf38f1f11027c5268559e0eb87dbf8eaf3e7cb9b339103a.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 65, + 182, + 92, + 190 + ], + "lines": [ + { + "bbox": [ + 65, + 182, + 92, + 190 + ], + "spans": [ + { + "bbox": [ + 65, + 182, + 92, + 190 + ], + "type": "text", + "content": "(a) Input" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 111, + 70, + 170, + 181 + ], + "blocks": [ + { + "bbox": [ + 111, + 70, + 170, + 181 + ], + "lines": [ + { + "bbox": [ + 111, + 70, + 170, + 181 + ], + "spans": [ + { + "bbox": [ + 111, + 70, + 170, + 181 + ], + "type": "image", + "image_path": "8003faa17f193bf480bfd2eb50764ec53f8d809021b577077222fd50af8ea5f1.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 129, + 182, + 153, + 190 + ], + "lines": [ + { + "bbox": [ + 129, + 182, + 153, + 190 + ], + "spans": [ + { + "bbox": [ + 129, + 182, + 153, + 190 + ], + "type": "text", + "content": "(b) " + }, + { + "bbox": [ + 129, + 182, + 153, + 190 + ], + "type": "inline_equation", + "content": "\\times 5" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 173, + 70, + 233, + 181 + ], + "blocks": [ + { + "bbox": [ + 173, + 70, + 233, + 181 + ], + "lines": [ + { + "bbox": [ + 173, + 70, + 233, + 181 + ], + "spans": [ + { + "bbox": [ + 173, + 70, + 233, + 181 + ], + "type": "image", + "image_path": "fc46698ed5f88479ff83cb874e01e3c43b2b24584b29f55951d74a7f51413217.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 192, + 181, + 215, + 189 + ], + "lines": [ + { + "bbox": [ + 192, + 181, + 215, + 189 + ], + "spans": [ + { + "bbox": [ + 192, + 181, + 215, + 189 + ], + "type": "text", + "content": "(c)×4" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 236, + 70, + 295, + 181 + ], + "blocks": [ + { + "bbox": [ + 236, + 70, + 295, + 181 + ], + "lines": [ + { + "bbox": [ + 236, + 70, + 295, + 181 + ], + "spans": [ + { + "bbox": [ + 236, + 70, + 295, + 181 + ], + "type": "image", + "image_path": "fe6f5a25ef97440ebef75c8ef255903de2ebc32af2cc0db0d644274d3f96cb8a.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 255, + 181, + 277, + 189 + ], + "lines": [ + { + "bbox": [ + 255, + 181, + 277, + 189 + ], + "spans": [ + { + "bbox": [ + 255, + 181, + 277, + 189 + ], + "type": "text", + "content": "(d)×3" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 203, + 545, + 224 + ], + "lines": [ + { + "bbox": [ + 46, + 203, + 545, + 224 + ], + "spans": [ + { + "bbox": [ + 46, + 203, + 545, + 224 + ], + "type": "text", + "content": "Figure 8. Qualitative results on the 4KIL dataset of different scales. The top row is obtained from the lanzcos3 operator without LMAR, while the bottom row is with LMAR. Please zoom in for details." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 298, + 70, + 358, + 181 + ], + "blocks": [ + { + "bbox": [ + 298, + 70, + 358, + 181 + ], + "lines": [ + { + "bbox": [ + 298, + 70, + 358, + 181 + ], + "spans": [ + { + "bbox": [ + 298, + 70, + 358, + 181 + ], + "type": "image", + "image_path": "e832857fd3e6fe0638274e575c4e26ac7c16315857300943311848a7e13b667f.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 317, + 181, + 340, + 189 + ], + "lines": [ + { + "bbox": [ + 317, + 181, + 340, + 189 + ], + "spans": [ + { + "bbox": [ + 317, + 181, + 340, + 189 + ], + "type": "text", + "content": "(e)×2" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 361, + 70, + 421, + 181 + ], + "blocks": [ + { + "bbox": [ + 361, + 70, + 421, + 181 + ], + "lines": [ + { + "bbox": [ + 361, + 70, + 421, + 181 + ], + "spans": [ + { + "bbox": [ + 361, + 70, + 421, + 181 + ], + "type": "image", + "image_path": "d95033fe1b9124f124285d799ef4a1d875c09e97ebae88c15244f2e4bf3549f6.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 378, + 181, + 405, + 189 + ], + "lines": [ + { + "bbox": [ + 378, + 181, + 405, + 189 + ], + "spans": [ + { + "bbox": [ + 378, + 181, + 405, + 189 + ], + "type": "text", + "content": "(f) " + }, + { + "bbox": [ + 378, + 181, + 405, + 189 + ], + "type": "inline_equation", + "content": "\\times 1.5" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 424, + 70, + 483, + 181 + ], + "blocks": [ + { + "bbox": [ + 424, + 70, + 483, + 181 + ], + "lines": [ + { + "bbox": [ + 424, + 70, + 483, + 181 + ], + "spans": [ + { + "bbox": [ + 424, + 70, + 483, + 181 + ], + "type": "image", + "image_path": "a6720055a11619e8d182c78e8f4135edec30d3a85ffe6b9e772710176f5a3be7.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 443, + 181, + 465, + 189 + ], + "lines": [ + { + "bbox": [ + 443, + 181, + 465, + 189 + ], + "spans": [ + { + "bbox": [ + 443, + 181, + 465, + 189 + ], + "type": "text", + "content": "(g)×1" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 485, + 70, + 545, + 181 + ], + "blocks": [ + { + "bbox": [ + 485, + 70, + 545, + 181 + ], + "lines": [ + { + "bbox": [ + 485, + 70, + 545, + 181 + ], + "spans": [ + { + "bbox": [ + 485, + 70, + 545, + 181 + ], + "type": "image", + "image_path": "48d6835db7bd1a3fdc5c394d18204752e7918d0972153b5833127b49be6aff88.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 504, + 181, + 526, + 189 + ], + "lines": [ + { + "bbox": [ + 504, + 181, + 526, + 189 + ], + "spans": [ + { + "bbox": [ + 504, + 181, + 526, + 189 + ], + "type": "text", + "content": "(h) GT" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "bbox": [ + 46, + 234, + 286, + 305 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 234, + 286, + 305 + ], + "spans": [ + { + "bbox": [ + 46, + 234, + 286, + 305 + ], + "type": "text", + "content": "the transformer enhancer, the number of collaborative training epochs is 60, with a " + }, + { + "bbox": [ + 46, + 234, + 286, + 305 + ], + "type": "inline_equation", + "content": "4e^{-4}" + }, + { + "bbox": [ + 46, + 234, + 286, + 305 + ], + "type": "text", + "content": " initial learning rate, which decays by a factor of 0.75 every 20 epochs. The discriminator is composed of several convolution layers, whose initial learning rate is " + }, + { + "bbox": [ + 46, + 234, + 286, + 305 + ], + "type": "inline_equation", + "content": "2e^{-4}" + }, + { + "bbox": [ + 46, + 234, + 286, + 305 + ], + "type": "text", + "content": " and shares the same training strategies as mentioned above." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 47, + 312, + 190, + 324 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 312, + 190, + 324 + ], + "spans": [ + { + "bbox": [ + 47, + 312, + 190, + 324 + ], + "type": "text", + "content": "4.3. Performance Comparison" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 46, + 330, + 286, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 330, + 286, + 533 + ], + "spans": [ + { + "bbox": [ + 46, + 330, + 286, + 533 + ], + "type": "text", + "content": "Quantitative Evaluation. We utilize two metrics, the Peak Signal to Noise Ratio (PSNR, dB), and the Structural Similarity (SSIM), to assess the performance gains. To verify the universality of our methods, we adopt five types of resizers and test them on various commonly used scales, including both in-distribution and out-of-distribution scales. As shown in Tables 1, 2, and 3, our method consistently achieves performance improvements across different resolutions, regardless of the choice of resizer or the category of the backbone. Notably, the performance with LMAR at UHD resolution is analogous to the original UHD input, which illustrates the superior continuous modeling capability of LMAR. Additionally, our approach maintains comparable effectiveness with the baselines, where a minor increase in parameters, as shown in Table 4. More importantly, our algorithm only needs to be trained on one type of resizer, which can then be applied to any other resizers." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 46, + 533, + 286, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 533, + 286, + 628 + ], + "spans": [ + { + "bbox": [ + 46, + 533, + 286, + 628 + ], + "type": "text", + "content": "Qualitative Evaluation. Figures 7 and 8 present the visual comparison on the UHD-LOL4K dataset and 4KIL dataset respectively. These comparisons demonstrate that enabling model-aware resampling leads to enhanced results with improved global lightness and local structural details, across various resampling scales. It verifies that coupling the resize-ers and enhancers could generate better visual results. More visual results are presented in the supplementary materials." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 47, + 635, + 147, + 646 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 635, + 147, + 646 + ], + "spans": [ + { + "bbox": [ + 47, + 635, + 147, + 646 + ], + "type": "text", + "content": "4.4. Ablation Studies" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 46, + 653, + 286, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 286, + 712 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 286, + 712 + ], + "type": "text", + "content": "We conduct ablation studies on the UHD-LOL4K dataset to investigate the rationality of our designs. The ablation results are shown in Table 5. First, we replace the bottleneck features (Full) with the relatively front features " + }, + { + "bbox": [ + 46, + 653, + 286, + 712 + ], + "type": "inline_equation", + "content": "(F_{f})" + }, + { + "bbox": [ + 46, + 653, + 286, + 712 + ], + "type": "text", + "content": " and relatively end features " + }, + { + "bbox": [ + 46, + 653, + 286, + 712 + ], + "type": "inline_equation", + "content": "(F_{e})" + }, + { + "bbox": [ + 46, + 653, + 286, + 712 + ], + "type": "text", + "content": " to apply feature constraints," + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 304, + 234, + 545, + 353 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 234, + 545, + 353 + ], + "spans": [ + { + "bbox": [ + 304, + 234, + 545, + 353 + ], + "type": "text", + "content": "where remarkable performance drops can be observed at 4 times downsampling. The findings indicate that the bottleneck features exhibit the most representative capability in guiding model-aware resampling. Second, we replace the discriminator (Full) with the perceptual loss (VGG) to conduct the distribution restriction. The utilization of discriminator encourages more consistent feature alignment. Lastly, even with sub-optimal constraints, our method outperforms the baseline, emphasizing the effectiveness of coupling resizers and enhancers." + } + ] + } + ], + "index": 23 + }, + { + "type": "table", + "bbox": [ + 313, + 394, + 537, + 434 + ], + "blocks": [ + { + "bbox": [ + 305, + 362, + 545, + 384 + ], + "lines": [ + { + "bbox": [ + 305, + 362, + 545, + 384 + ], + "spans": [ + { + "bbox": [ + 305, + 362, + 545, + 384 + ], + "type": "text", + "content": "Table 5. Ablation studies on the constraints employed in our designs. We present the PSNR results here." + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 313, + 394, + 537, + 434 + ], + "lines": [ + { + "bbox": [ + 313, + 394, + 537, + 434 + ], + "spans": [ + { + "bbox": [ + 313, + 394, + 537, + 434 + ], + "type": "table", + "html": "
ScalesBaselineFfFeVGGFull
x229.8229.9830.0530.0330.20
x425.4025.8625.5025.4726.36
", + "image_path": "8f33bfd143f070a8bdb6e8c9e5641285ea70daab5c6164d6322289cd4e12d7ff.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "table_body" + } + ], + "index": 25 + }, + { + "bbox": [ + 306, + 448, + 378, + 460 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 448, + 378, + 460 + ], + "spans": [ + { + "bbox": [ + 306, + 448, + 378, + 460 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 304, + 468, + 545, + 611 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 468, + 545, + 611 + ], + "spans": [ + { + "bbox": [ + 304, + 468, + 545, + 611 + ], + "type": "text", + "content": "In this paper, we introduce a novel approach for effectively enhancing the UHD images at different downsampling scales by a model-aware resampling strategy. Our method customizes resampling under the guidance of model knowledge, integrating the resizers and enhancers through compensatory information learning. Notably, our algorithm is fully compatible with existing interpolation resamplers and maintains the efficiency. Moreover, our design promotes performances at low resolutions remarkably while keeping comparable results at UHD resolution, without the need to retrain the enhancer. Extensive experiments have validated the effectiveness and scalability of our method." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 306, + 621, + 410, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 621, + 410, + 635 + ], + "spans": [ + { + "bbox": [ + 306, + 621, + 410, + 635 + ], + "type": "text", + "content": "6. Acknowledgment" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 304, + 641, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 641, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 641, + 545, + 712 + ], + "type": "text", + "content": "This work was supported by the JKW Research Funds under Grant 20-163-14-LZ-001-004-01, and the Anhui Provincial Natural Science Foundation under Grant 2108085UD12. We acknowledge the support of GPU cluster built by MCC Lab of Information Science and Technology Institution, USTC." + } + ] + } + ], + "index": 29 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "25729" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 112 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 112 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 112 + ], + "type": "text", + "content": "[1] Kenneth R Castleman. Digital image processing. Prentice Hall Press, 1996. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 114, + 287, + 147 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 114, + 287, + 147 + ], + "spans": [ + { + "bbox": [ + 53, + 114, + 287, + 147 + ], + "type": "text", + "content": "[2] Jiawen Chen, Andrew Adams, Neal Wadhwa, and Samuel W Hasinoff. Bilateral guided upsampling. ACM Transactions on Graphics (TOG), 35(6):1-8, 2016. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 148, + 287, + 203 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 148, + 287, + 203 + ], + "spans": [ + { + "bbox": [ + 53, + 148, + 287, + 203 + ], + "type": "text", + "content": "[3] Liangyu Chen, Xin Lu, Jie Zhang, Xiaojie Chu, and Chengpeng Chen. Hinet: Half instance normalization network for image restoration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, pages 182-192, 2021. 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 205, + 287, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 205, + 287, + 258 + ], + "spans": [ + { + "bbox": [ + 53, + 205, + 287, + 258 + ], + "type": "text", + "content": "[4] Yinbo Chen, Sifei Liu, and Xiaolong Wang. Learning continuous image representation with local implicit image function. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8628-8638, 2021. 3, 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 261, + 287, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 261, + 287, + 316 + ], + "spans": [ + { + "bbox": [ + 53, + 261, + 287, + 316 + ], + "type": "text", + "content": "[5] Weitao Feng, Deyi Ji, Yiru Wang, Shuorong Chang, Hansheng Ren, and Weihao Gan. Challenges on large scale surveillance video analysis. In Proceedings of the IEEE conference on computer vision and pattern recognition workshops, pages 69-76, 2018. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 318, + 287, + 361 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 318, + 287, + 361 + ], + "spans": [ + { + "bbox": [ + 53, + 318, + 287, + 361 + ], + "type": "text", + "content": "[6] Michael Gharbi, Jiawen Chen, Jonathan T Barron, Samuel W Hasinoff, and Frédo Durand. Deep bilateral learning for real-time image enhancement. ACM Transactions on Graphics (TOG), 36(4):1-12, 2017. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 363, + 287, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 363, + 287, + 407 + ], + "spans": [ + { + "bbox": [ + 53, + 363, + 287, + 407 + ], + "type": "text", + "content": "[7] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial networks. Communications of the ACM, 63(11):139-144, 2020. 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 409, + 287, + 462 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 409, + 287, + 462 + ], + "spans": [ + { + "bbox": [ + 53, + 409, + 287, + 462 + ], + "type": "text", + "content": "[8] Dianyuan Han. Comparison of commonly used image interpolation methods. In Conference of the 2nd International Conference on Computer Science and Electronics Engineering (ICCSEE 2013), pages 1556-1559. Atlantis Press, 2013. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 464, + 287, + 508 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 464, + 287, + 508 + ], + "spans": [ + { + "bbox": [ + 53, + 464, + 287, + 508 + ], + "type": "text", + "content": "[9] Muhammad Haris, Gregory Shakhnarovich, and Norimichi Ukita. Deep back-projection networks for super-resolution. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1664-1673, 2018. 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 510, + 287, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 510, + 287, + 553 + ], + "spans": [ + { + "bbox": [ + 48, + 510, + 287, + 553 + ], + "type": "text", + "content": "[10] Hanzhe Hu, Deyi Ji, Weihao Gan, Shuai Bai, Wei Wu, and Junjie Yan. Class-wise dynamic graph convolution for semantic segmentation. In European Conference on Computer Vision, pages 1-17. Springer, 2020. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 555, + 287, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 555, + 287, + 609 + ], + "spans": [ + { + "bbox": [ + 48, + 555, + 287, + 609 + ], + "type": "text", + "content": "[11] Xuecai Hu, Haoyuan Mu, Xiangyu Zhang, Zilei Wang, Tieniu Tan, and Jian Sun. Meta-sr: A magnification-arbitrary network for super-resolution. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1575-1584, 2019. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 612, + 287, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 612, + 287, + 644 + ], + "spans": [ + { + "bbox": [ + 48, + 612, + 287, + 644 + ], + "type": "text", + "content": "[12] Max Jaderberg, Karen Simonyan, Andrew Zisserman, et al. Spatial transformer networks. Advances in neural information processing systems, 28, 2015. 2, 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 647, + 287, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 647, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 647, + 287, + 689 + ], + "type": "text", + "content": "[13] Deyi Ji, Hongtao Lu, and Tongzhen Zhang. End to end multiscale convolutional neural network for crowd counting. In Eleventh international conference on machine vision, pages 761-766. SPIE, 2019. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "type": "text", + "content": "[14] Deyi Ji, Haoran Wang, Hanzhe Hu, Weihao Gan, Wei Wu, and Junjie Yan. Context-aware graph convolution network" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 32, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "text", + "content": "for target re-identification. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1646–1654, 2021." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 96, + 545, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 96, + 545, + 149 + ], + "spans": [ + { + "bbox": [ + 308, + 96, + 545, + 149 + ], + "type": "text", + "content": "[15] Deyi Ji, Haoran Wang, Mingyuan Tao, Jianqiang Huang, Xian-Sheng Hua, and Hongtao Lu. Structural and statistical texture knowledge distillation for semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16876-16885, 2022." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 151, + 545, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 151, + 545, + 183 + ], + "spans": [ + { + "bbox": [ + 308, + 151, + 545, + 183 + ], + "type": "text", + "content": "[16] Deyi Ji, Siqi Gao, Mingyuan Tao, Hongtao Lu, and Feng Zhao. Changenet: Multi-temporal asymmetric change detection dataset. arXiv preprint arXiv:2312.17428, 2023." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 185, + 545, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 185, + 545, + 227 + ], + "spans": [ + { + "bbox": [ + 308, + 185, + 545, + 227 + ], + "type": "text", + "content": "[17] Deyi Ji, Feng Zhao, and Hongtao Lu. Guided patch-grouping wavelet transformer with spatial congruence for ultra-high resolution segmentation. International Joint Conference on Artificial Intelligence, 2023." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 228, + 545, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 228, + 545, + 281 + ], + "spans": [ + { + "bbox": [ + 308, + 228, + 545, + 281 + ], + "type": "text", + "content": "[18] Deyi Ji, Feng Zhao, Hongtao Lu, Mingyuan Tao, and Jieping Ye. Ultra-high resolution segmentation with ultra-rich context: A novel benchmark. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23621-23630, 2023." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 283, + 545, + 325 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 283, + 545, + 325 + ], + "spans": [ + { + "bbox": [ + 308, + 283, + 545, + 325 + ], + "type": "text", + "content": "[19] Deyi Ji, Siqi Gao, Lanyun Zhu, Yiru Zhao, Peng Xu, Hong-tao Lu, and Feng Zhao. View-centric multi-object tracking with homographic matching in moving UAV. arXiv preprint, 2024. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 327, + 545, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 327, + 545, + 369 + ], + "spans": [ + { + "bbox": [ + 308, + 327, + 545, + 369 + ], + "type": "text", + "content": "[20] Heewon Kim, Myungsub Choi, Bee Lim, and Kyoung Mu Lee. Task-aware image downscaling. In Proceedings of the European conference on computer vision (ECCV), pages 399-414, 2018. 2, 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 372, + 545, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 372, + 545, + 402 + ], + "spans": [ + { + "bbox": [ + 308, + 372, + 545, + 402 + ], + "type": "text", + "content": "[21] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014.7" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 404, + 545, + 447 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 404, + 545, + 447 + ], + "spans": [ + { + "bbox": [ + 308, + 404, + 545, + 447 + ], + "type": "text", + "content": "[22] Yue Li, Dong Liu, Houqiang Li, Li Li, Zhu Li, and Feng Wu. Learning a convolutional neural network for image compact-resolution. IEEE Transactions on Image Processing, 28(3): 1092-1107, 2018. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 449, + 545, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 449, + 545, + 502 + ], + "spans": [ + { + "bbox": [ + 308, + 449, + 545, + 502 + ], + "type": "text", + "content": "[23] Jie Liang, Hui Zeng, and Lei Zhang. High-resolution photorealistic image translation in real-time: A laplacian pyramid translation network. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9392-9400, 2021. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 504, + 545, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 504, + 545, + 536 + ], + "spans": [ + { + "bbox": [ + 308, + 504, + 545, + 536 + ], + "type": "text", + "content": "[24] Qiaowanni Lin, Zhuoran Zheng, and Xiuyi Jia. Uhd low-light image enhancement via interpretable bilateral learning. Information Sciences, 2022. 6" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 537, + 545, + 580 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 537, + 545, + 580 + ], + "spans": [ + { + "bbox": [ + 308, + 537, + 545, + 580 + ], + "type": "text", + "content": "[25] Yang Liu, Zhenyue Qin, Saeed Anwar, Pan Ji, Dongwoo Kim, Sabrina Caldwell, and Tom Gedeon. Invertible denoising network: A light solution for real noise removal. arXiv preprint arXiv:2104.10546, 2021. 6" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 582, + 545, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 582, + 545, + 613 + ], + "spans": [ + { + "bbox": [ + 308, + 582, + 545, + 613 + ], + "type": "text", + "content": "[26] Don P Mitchell and Arun N Netravali. Reconstruction filters in computer-graphics. ACM Siggraph Computer Graphics, 22(4):221-228, 1988. 3" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 615, + 467, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 615, + 467, + 624 + ], + "spans": [ + { + "bbox": [ + 308, + 615, + 467, + 624 + ], + "type": "text", + "content": "[27] Thomas Müller. tiny-cuda-nn, 2021. 7" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 308, + 626, + 545, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 626, + 545, + 678 + ], + "spans": [ + { + "bbox": [ + 308, + 626, + 545, + 678 + ], + "type": "text", + "content": "[28] Adria Recasens, Petr Kellnhofer, Simon Stent, Wojciech Matusik, and Antonio Torralba. Learning to zoom: a saliency-based sampling layer for neural networks. In Proceedings of the European conference on computer vision (ECCV), pages 51-66, 2018. 2, 3" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 308, + 681, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 681, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 681, + 545, + 713 + ], + "type": "text", + "content": "[29] Assaf Shocher, Ben Feinstein, Niv Haim, and Michal Irani. From discrete to continuous convolution layers. arXiv preprint arXiv:2006.11120, 2020.4" + } + ] + } + ], + "index": 31 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "25730" + } + ] + } + ], + "index": 33 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 115 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 115 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 115 + ], + "type": "text", + "content": "[30] Wanjie Sun and Zhenzhong Chen. Learned image downscaling for upscaling using content adaptive resampler. IEEE Transactions on Image Processing, 29:4027-4040, 2020. 2, 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 118, + 287, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 118, + 287, + 162 + ], + "spans": [ + { + "bbox": [ + 49, + 118, + 287, + 162 + ], + "type": "text", + "content": "[31] Haoran Wang, Licheng Jiao, Fang Liu, Lingling Li, Xu Liu, Deyi Ji, and Weihao Gan. Ipgn: Interactiveness proposal graph network for human-object interaction detection. IEEE Transactions on Image Processing, 30:6583-6593, 2021. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 163, + 287, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 163, + 287, + 217 + ], + "spans": [ + { + "bbox": [ + 48, + 163, + 287, + 217 + ], + "type": "text", + "content": "[32] Haoran Wang, Licheng Jiao, Fang Liu, Lingling Li, Xu Liu, Deyi Ji, and Weihao Gan. Learning social spatio-temporal relation graph in the wild and a video benchmark. IEEE Transactions on Neural Networks and Learning Systems, 2021. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 220, + 287, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 220, + 287, + 262 + ], + "spans": [ + { + "bbox": [ + 48, + 220, + 287, + 262 + ], + "type": "text", + "content": "[33] Tao Wang, Kaihao Zhang, Tianrun Shen, Wenhan Luo, Bjorn Stenger, and Tong Lu. Ultra-high-definition low-light image enhancement: A benchmark and transformer-based method. arXiv preprint arXiv:2212.11548, 2022. 3, 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 264, + 287, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 264, + 287, + 330 + ], + "spans": [ + { + "bbox": [ + 48, + 264, + 287, + 330 + ], + "type": "text", + "content": "[34] Mingqing Xiao, Shuxin Zheng, Chang Liu, Yaolong Wang, Di He, Guolin Ke, Jiang Bian, Zhouchen Lin, and Tie-Yan Liu. Invertible image rescaling. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part I 16, pages 126-144. Springer, 2020. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 331, + 287, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 331, + 287, + 364 + ], + "spans": [ + { + "bbox": [ + 48, + 331, + 287, + 364 + ], + "type": "text", + "content": "[35] Jinbo Xing, Wenbo Hu, Menghan Xia, and Tien-Tsin Wong. Scale-arbitrary invertible image downscaling. IEEE Transactions on Image Processing, 2023. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 365, + 287, + 419 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 365, + 287, + 419 + ], + "spans": [ + { + "bbox": [ + 48, + 365, + 287, + 419 + ], + "type": "text", + "content": "[36] Wei Yu, Qi Zhu, Naishan Zheng, Jie Huang, Man Zhou, and Feng Zhao. Learning non-uniform-sampling for ultra-high-definition image enhancement. In Proceedings of the 31st ACM International Conference on Multimedia, pages 1412-1421, 2023. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 422, + 287, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 422, + 287, + 464 + ], + "spans": [ + { + "bbox": [ + 48, + 422, + 287, + 464 + ], + "type": "text", + "content": "[37] Sergey Zagoruyko and Nikos Komodakis. Paying more attention to attention: Improving the performance of convolutional neural networks via attention transfer. arXiv preprint arXiv:1612.03928, 2016. 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 467, + 287, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 467, + 287, + 510 + ], + "spans": [ + { + "bbox": [ + 48, + 467, + 287, + 510 + ], + "type": "text", + "content": "[38] Syed Waqas Zamir, Aditya Arora, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, and Ming-Hsuan Yang. Restormer: Efficient transformer for high-resolution image restoration. In CVPR, 2022. 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 512, + 287, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 512, + 287, + 555 + ], + "spans": [ + { + "bbox": [ + 48, + 512, + 287, + 555 + ], + "type": "text", + "content": "[39] Yongbing Zhang, Debin Zhao, Jian Zhang, Ruiqin Xiong, and Wen Gao. Interpolation-dependent image downsampling. IEEE Transactions on Image Processing, 20(11): 3291-3296, 2011. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 557, + 287, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 557, + 287, + 621 + ], + "spans": [ + { + "bbox": [ + 48, + 557, + 287, + 621 + ], + "type": "text", + "content": "[40] Zhuoran Zheng, Wenqi Ren, Xiaochun Cao, Xiaobin Hu, Tao Wang, Fenglong Song, and Xiuyi Jia. Ultra-high-definition image dehazing via multi-guided bilateral learning. In 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 16180-16189. IEEE, 2021. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 624, + 287, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 624, + 287, + 678 + ], + "spans": [ + { + "bbox": [ + 48, + 624, + 287, + 678 + ], + "type": "text", + "content": "[41] Zhuoran Zheng, Wenqi Ren, Xiaochun Cao, Tao Wang, and Xiuyi Jia. Ultra-high-definition image hdr reconstruction via collaborative bilateral learning. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4449-4458, 2021. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "type": "text", + "content": "[42] Lanyun Zhu, Deyi Ji, Shiping Zhu, Weihao Gan, Wei Wu, and Junjie Yan. Learning statistical texture for semantic segmentation. In Proceedings of the IEEE/CVF Conference" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 364 + ], + "type": "list", + "angle": 0, + "index": 21, + "blocks": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "type": "text", + "content": "on Computer Vision and Pattern Recognition (CVPR), pages 12537-12546, 2021. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 96, + 545, + 129 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 96, + 545, + 129 + ], + "spans": [ + { + "bbox": [ + 307, + 96, + 545, + 129 + ], + "type": "text", + "content": "[43] Lanyun Zhu, Tianrun Chen, Deyi Ji, Jieping Ye, and Jun Liu. Llafs: When large-language models meet few-shot segmentation. arXiv preprint arXiv:2311.16926, 2023." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 130, + 545, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 130, + 545, + 173 + ], + "spans": [ + { + "bbox": [ + 307, + 130, + 545, + 173 + ], + "type": "text", + "content": "[44] Lanyun Zhu, Deyi Ji, Tianrun Chen, Peng Xu, Jieping Ye, and Jun Liu. Ibid: Alleviating hallucinations in large vision-language models via image-biased decoding. arXiv preprint arXiv:2402.18476, 2024." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 175, + 545, + 218 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 175, + 545, + 218 + ], + "spans": [ + { + "bbox": [ + 307, + 175, + 545, + 218 + ], + "type": "text", + "content": "[45] Qi Zhu, Zeyu Xiao, Jie Huang, and Feng Zhao. Dast-net: Depth-aware spatio-temporal network for video deblurring. In 2022 IEEE International Conference on Multimedia and Expo (ICME), pages 1-6. IEEE, 2022." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 220, + 545, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 220, + 545, + 262 + ], + "spans": [ + { + "bbox": [ + 307, + 220, + 545, + 262 + ], + "type": "text", + "content": "[46] Qi Zhu, Naishan Zheng, Jie Huang, Man Zhou, Jinghao Zhang, and Feng Zhao. Learning spatio-temporal sharpness map for video deblurring. IEEE Transactions on Circuits and Systems for Video Technology, 2023." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 264, + 545, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 264, + 545, + 319 + ], + "spans": [ + { + "bbox": [ + 307, + 264, + 545, + 319 + ], + "type": "text", + "content": "[47] Qi Zhu, Man Zhou, Naishan Zheng, Chongyi Li, Jie Huang, and Feng Zhao. Exploring temporal frequency spectrum in deep video deblurring. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12428-12437, 2023." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 320, + 545, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 320, + 545, + 364 + ], + "spans": [ + { + "bbox": [ + 307, + 320, + 545, + 364 + ], + "type": "text", + "content": "[48] Qi Zhu, Jie Huang, Naishan Zheng, Hongzhi Gao, Chongyi Li, Yuan Xu, Feng Zhao, et al. Fouridown: Factoring down-sampling into shuffling and superposing. Advances in Neural Information Processing Systems, 36, 2024. 3" + } + ] + } + ], + "index": 20 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "25731" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Emu Edit_ Precise Image Editing via Recognition and Generation Tasks/b6371d78-25ca-4b11-84b1-a566ed9d62f2_content_list.json b/2024/Emu Edit_ Precise Image Editing via Recognition and Generation Tasks/b6371d78-25ca-4b11-84b1-a566ed9d62f2_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..41234cc17017b607bd564f66b5ba2090597179a2 --- /dev/null +++ b/2024/Emu Edit_ Precise Image Editing via Recognition and Generation Tasks/b6371d78-25ca-4b11-84b1-a566ed9d62f2_content_list.json @@ -0,0 +1,1442 @@ +[ + { + "type": "text", + "text": "Emu Edit: Precise Image Editing via Recognition and Generation Tasks", + "text_level": 1, + "bbox": [ + 122, + 130, + 846, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Shelly Sheynin*, Adam Polyak*, Uriel Singer*, Yuval Kirstain*, Amit Zohar*, Oron Ashual, Devi Parikh and Yaniv Taigman", + "bbox": [ + 122, + 179, + 844, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "GenAI,Meta", + "bbox": [ + 429, + 233, + 540, + 250 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/1daaa348ca8afee9b9522231b19578495255b61c69bb7fa381085ebff184eec9.jpg", + "image_caption": [ + "Dress the emu with a fireman outfit" + ], + "image_footnote": [], + "bbox": [ + 83, + 273, + 349, + 377 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/121a82bdf17aea58535197abd2870754df60ef1528d39a5ce16c55d569d0f221.jpg", + "image_caption": [ + "Let's see it graduating", + "Figure 1. Emu Edit is a multi-tasking model that combines various editing (left, middle) and vision (right) tasks for precise image editing." + ], + "image_footnote": [], + "bbox": [ + 352, + 273, + 617, + 377 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/0967a632fde260268fe3b3f4fc15dccddf079cf81e5f8fac9b18c6cde487e995.jpg", + "image_caption": [ + "Mark the drinks" + ], + "image_footnote": [], + "bbox": [ + 619, + 273, + 887, + 377 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 426, + 313, + 441 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Instruction-based image editing holds immense potential for a variety of applications, as it enables users to perform any editing operation using a natural language instruction. However, current models in this domain often struggle with accurately executing user instructions. We present Emu Edit, a multi-task image editing model which sets state-of-the-art results in instruction-based image editing. To develop Emu Edit we train it to multi-task across an unprecedented range of tasks, such as region-based editing, free-form editing, and Computer Vision tasks, all of which are formulated as generative tasks. Additionally, to enhance Emu Edit's multi-task learning abilities, we provide it with learned task embeddings which guide the generation process towards the correct edit type. Both these elements are essential for Emu Edit's outstanding performance. Furthermore, we show that Emu Edit can generalize to new tasks, such as image inpainting, super-resolution, and compositions of editing tasks, with just a few labeled examples. This capability offers a significant advantage in scenarios where high-quality samples are scarce. Lastly, to facilitate a more rigorous and informed assessment of instructable image editing models, we release a new challenging and versatile benchmark that includes seven different image editing tasks.", + "bbox": [ + 75, + 457, + 473, + 819 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 501, + 426, + 630, + 441 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Image editing is a widely-used application that millions engage with every day. Popular image editing tools, however, either demand considerable expertise and are time-consuming to use, or are quite limited, providing only a predefined set of editing operations, such as specific filters. Instruction-based image editing [2, 25] attempts to resolve these limitations by allowing users to effortlessly describe their editing goals using natural language instructions. For instance, a user can provide a model with an image and instruct it to \"Dress the emu with a fireman outfit\" or \"Let's see it graduating\" (see Fig. 1). Nevertheless, while instruction-based image editing models like InstructPix2Pix [2] are designed to process any given instruction, they often struggle to accurately interpret and execute such instructions. Moreover, their generalization is limited, often falling short on tasks that deviate slightly from those they were trained on (see Fig. 3). To address these gaps, we introduce Emu Edit, the first image editing model trained on an extensive and diverse set of tasks, including both image editing and computer vision tasks. Emu Edit provides a substantial improvement in both compliance with the edit instruction and preservation of the visual fidelity of the original image. As we show through both automatic metrics [14] and human judgments on two benchmarks [25], Emu Edit achieves state-of-the-art results in instruction-based image editing. The success of Emu Edit stems from two key contributions. First, we train our model to multi-task across sixteen distinct image editing tasks. These tasks span region-based editing tasks, free-form editing tasks and computer", + "bbox": [ + 496, + 462, + 893, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "1Project Page: https://emu-edit.metademolab.com/", + "bbox": [ + 94, + 875, + 436, + 887 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Equal contribution.", + "bbox": [ + 96, + 887, + 205, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "8871", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/f6daee2dd1ad5774950c595167a32770ca7e13f1189f15de497229100214adba.jpg", + "image_caption": [ + "(a) A cat" + ], + "image_footnote": [], + "bbox": [ + 80, + 102, + 241, + 224 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/e20d245dddf075ebb2f9a85e3577ad0e4ad463efa803748455a964ab0c108ffd.jpg", + "image_caption": [ + "(b) Remove the tail" + ], + "image_footnote": [], + "bbox": [ + 241, + 102, + 403, + 224 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/53da06711e6d3d8423eacac630c92eea95a427e99482ac28db72a69654f67fc4.jpg", + "image_caption": [ + "(c) Add a pink jacket" + ], + "image_footnote": [], + "bbox": [ + 403, + 102, + 565, + 224 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/3625b84b9371fb02d492e67157add04fc9d29f97c7717af61ecbcdf2f791e658.jpg", + "image_caption": [ + "(d)Make it rainy" + ], + "image_footnote": [], + "bbox": [ + 566, + 102, + 727, + 224 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/506af08233d4e36e2f8bbce59603fe18b2c59cfb79e7660fc10d64ab3bfbdf70.jpg", + "image_caption": [ + "(e) Have the cat look shocked" + ], + "image_footnote": [], + "bbox": [ + 727, + 102, + 890, + 224 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/2f4c3ffb2eee89852984d434d8896b910ed6d6dd66675ea26147c60a85839bc1.jpg", + "image_caption": [ + "(f) Extract the depth map" + ], + "image_footnote": [], + "bbox": [ + 80, + 241, + 241, + 364 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/4641d1eab2feee72b020ae720e2337bfaf683f0a908d5c765094e57532838c29.jpg", + "image_caption": [ + "(g) Generate a rainy day image of a hedgehog in a dress using the depth map" + ], + "image_footnote": [], + "bbox": [ + 241, + 241, + 413, + 364 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/c9f3cb108decacef2bbdc2436d64171e786f0940660fbc7b29b03667bd400d6b.jpg", + "image_caption": [ + "(h) Replace the dress with an astronaut outfit" + ], + "image_footnote": [], + "bbox": [ + 413, + 241, + 565, + 364 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/d8b340b4db0cf630a244006cf7460b8d8e3f058dd3c6e89b6c3f84e7ed43e8a3.jpg", + "image_caption": [ + "(i) Segment the spacesuit, and detect the hands", + "Figure 2. Instruction-based image generation. Each image originates from a previous one with a corresponding instruction. The prior image for (a) was set to zero." + ], + "image_footnote": [], + "bbox": [ + 566, + 241, + 727, + 364 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/421fc7fc2f04003c5f2645ecebf7d8aa0491255a119677dfbe47c1f0ac10e446.jpg", + "image_caption": [ + "(j) Add the text \"purple cat\" using a purple font" + ], + "image_footnote": [], + "bbox": [ + 728, + 241, + 890, + 364 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "vision tasks, all formulated as generative tasks. Unlike previous work, we develop a distinct data generation pipeline for each task, allowing us to gather a training set that is not only more diverse but also more precise in its examples. We find that training a single model on all tasks yields better results than training expert models on each task independently. Additionally, we show that as the number of training tasks increases, so does the performance of Emu Edit. Furthermore, we discover that surprisingly, computer vision tasks such as detection, segmentation, and others, significantly enhance editing performance, as validated both by human raters as well as quantitative measures. Second, to process this wide array of tasks effectively, we introduce the concept of learned task embeddings, which are used to steer the generation process toward the correct generative task. Concretely, for each task, we learn a unique task embedding vector, and integrate it into the model through cross-attention interactions, and by adding it to the timestep embeddings. We demonstrate that learned task embeddings significantly enhance the ability of our model to accurately infer the appropriate edit type from the user instruction and execute the correct edit. Equipped with a robust model trained across a broad spectrum of tasks and guided by learned task embeddings, we explore few-shot adaptation to unseen tasks via task inversion. In this process, we maintain the model weights untouched, and solely update a task embedding to fit the new task. Our experiments demonstrate that Emu Edit can swiftly adapt to new tasks, such as superresolution, contour detection, and others. Notably, for some tasks, fine-tuning the model on just a handful of examples", + "bbox": [ + 75, + 446, + 472, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "yields results that nearly match those of an expert model trained on one hundred thousand examples. This makes task inversion with Emu Edit particularly advantageous in scenarios where labeled examples are limited, or when the compute budget is low. Finally, to support future research for instruction-based image editing, we publicly release a diverse and challenging benchmark that includes seven different image editing operations, as well as our model's generations on this dataset. In summary, this work addresses the limitations of instruction-based image editing models in accurately following user instructions. We demonstrate that by employing multi-task training across a diverse array of tasks, including recognition, generation, and editing, we can enhance our model's performance. Furthermore, by incorporating learned task embeddings into our model's architecture, we not only improve its results but also enable efficient few-shot learning for new tasks. With these improvements, our model sets a new standard in the field, offering significantly more precise and robust instruction-based image editing capabilities than existing models.", + "bbox": [ + 496, + 446, + 893, + 750 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 768, + 640, + 784 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The emergence of high-performing text-to-image diffusion models [7, 16, 17, 19] facilitated the development of effective text-based image editing methods. Such methods usually employ aligned and detailed descriptions of the input and edited image to perform a specific edit. Prompt-to-Prompt (P2P) [8] injects the input caption attention maps to the target caption attentions maps. Null-Text Inver", + "bbox": [ + 496, + 794, + 893, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "8872", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "sion [12] inverts an input image using the null-text embedding to support editing of a real image. Plug-and-Play (PNP) [21] injects spatial features in addition to attention maps and obtains better performance at global editing. Imagic [9] finetunes the diffusion model to support complex textual instructions. EDICT [22] suggests an image inversion based on two noise vectors enabling better image reconstruction and textual faithfulness. Another class of image editing models, employs an input mask as additional input [1, 23, 24]. Blended Diffusion [1] modifies the diffusion step by blending the input image in the unmasked regions. Imagen Editor [23] and SmartBrush [24] finetune the text-to-image model to be conditioned on both the input image and mask. While the text-based image editing methods detailed above enable humans to edit images, they frequently exhibit inconsistent performance and require multiple inputs, such as aligned and detailed descriptions of both the input and target images, or at times, input masks.", + "bbox": [ + 75, + 90, + 472, + 362 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To offer a more intuitive and user-friendly interface, and significantly enhance ease of use for humans, InstructPix2Pix [2] introduced an instructable image editing model. They developed this model by utilizing both GPT-3 [3] and Prompt-to-Prompt [8], to generate a large synthetic dataset for instruction-based image editing, and employed the dataset to train an instructable image editing model. Unlike InstructPix2Pix which used a synthetic dataset, MagicBrush [25] developed a manually-annotated instruction-guided image editing dataset by requesting humans to use an online image editing tool [13]. Finetuning InstructPix2Pix on this dataset led to improved image editing capabilities. However, even though there has been progress and improvement in instruction-based image editing models, we show in Sec. 5 that state-of-the-art image editing models still struggle with accurately interpreting and precisely executing editing instructions.", + "bbox": [ + 75, + 363, + 472, + 619 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this paper, we drastically narrow such performance gaps by leveraging multi-task training and a matching architecture. Unlike prior work that solely focuses on image editing [2, 25], we train our model to perform various tasks and learn a very diverse set of capabilities. The quality and versatility of our training procedure and dataset, together with our improved architecture for multi-task learning, enables us to make a big leap in performance and differentiates us from prior work in the field. Fig. 3 include several challenging editing samples as examples.", + "bbox": [ + 75, + 619, + 472, + 772 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Multi-Task Dataset for Image Editing", + "text_level": 1, + "bbox": [ + 76, + 784, + 416, + 801 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Training a robust and accurate image editing model requires a highly diverse dataset of input images, editing instructions, and output edited images. However, manually collecting such examples is impractically time-consuming, existing sources on the web (e.g. communities and forums on social media) are limited in size, and publicly available syn-", + "bbox": [ + 75, + 809, + 470, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "1. Region-Based Editing", + "text_level": 1, + "bbox": [ + 500, + 93, + 674, + 108 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Local: Substituting one object for another, altering an object's attributes (e.g., \"make it smile\").", + "- Remove: Erasing an object from the image.", + "- Add: Inserting a new object into the image.", + "- Texture: Altering an object's visual characteristics without affecting its structure (e.g., painting over, filling or covering an object).", + "- Background: Changing the scene's background." + ], + "bbox": [ + 500, + 109, + 890, + 229 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2. Free-Form Editing", + "text_level": 1, + "bbox": [ + 500, + 236, + 653, + 250 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Global: Edit instructions that affect the entire image, or that can not be described using a mask (e.g., \"let's see it in the summer\").", + "- Style: Change the style of an image.", + "- Text Editing: This involves text-related editing tasks such as adding, removing, swapping text, and altering the text's font and color." + ], + "bbox": [ + 500, + 251, + 890, + 354 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Vision tasks", + "text_level": 1, + "bbox": [ + 500, + 363, + 606, + 376 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Detect: Identifying and marking a specific object within the image with a rectangle bounding box.", + "- Segment: Isolating and marking an object in the image.", + "Color: Color adjustments like sharpening and blurring.", + "- Image-to-Image Translation: Tasks that involve bidirectional image type conversion, such as sketch-to-image, depth map-to-image, normal map-to-image, pose-to-image, segmentation map-to-image, and so on." + ], + "bbox": [ + 500, + 378, + 890, + 498 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Table 1. Description of the tasks forming the Emu Edit dataset.", + "bbox": [ + 506, + 513, + 880, + 527 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "thetic datasets often lack in diversity or quality. Therefore, we construct a new dataset that encompasses sixteen distinct tasks and ten million examples. Each example $(c_{I}, c_{T}, x, i)$ in our dataset, contains an input image $c_{I}$ , a text instruction $c_{T}$ , a target image $x$ , and a task index $i$ (out of the sixteen). The following section outlines the process of the data construction. In Sec. 3.1 we describe the instruction generation process, and in Sec. 3.2 the image pairs $(c_{I}, x)$ generation and filtering.", + "bbox": [ + 496, + 539, + 890, + 675 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Task Categories. The dataset is composed of tasks which are divided into three main categories: region-based editing, free-form editing, and vision tasks. Tab. 1 includes the full list of tasks, and their distribution in the train set is visualized in Fig. ??", + "bbox": [ + 496, + 696, + 890, + 773 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Instruction Generation", + "text_level": 1, + "bbox": [ + 500, + 785, + 712, + 800 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To generate editing instructions, we leverage the dialogue-optimized 70 billion parameter Llama 2 variant [20]. We observed that using a single agent to generate the instructions for all tasks leads to a lack of diversity in the dataset. Notably, the LLM exhibits a bias towards particular tasks and instruction phrasings. To address this, we utilize in", + "bbox": [ + 496, + 809, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "8873", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "context learning to create a task-specific agent for each task. Concretely, we provide the LLM with a task description, a few task-specific exemplars, and a real image caption. To increase diversity we sample the exemplars and randomize their order. Given such input, we expect the LLM to output (1) an editing instruction, (2) an output caption for an ideal output image, and (3) which objects should be updated or added to the original image. We refer the readers to supplementary Figs. ??-?? for examples of our prompts. Further details on instruction generation are provided in the supplementary Sec. ??", + "bbox": [ + 75, + 90, + 472, + 257 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Image Pairs Generation", + "text_level": 1, + "bbox": [ + 76, + 270, + 295, + 286 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our aim is to generate pairs of input and edited images that adhere to the edit instructions and preserve image elements that should remain intact. A crucial prerequisite when creating a pair of input and edited images is to guarantee that the two images differ only in specific elements or locations, while remaining identical in all other aspects. Previous instruct-based image editing methods [2] rely on Prompt-to-Prompt (P2P) to build an image-editing dataset. P2P injects cross-attention maps from the input image generation to the edited image generation. To support local edits, P2P additionally approximates a mask of the edited part, based on the cross-attention maps and constrains the edit to this local area. P2P relies on word-to-word alignment between the input image caption and the edited image caption (e.g. \"a cat riding a bicycle\" and \"a cat riding a car\") to produce editing image pairs. However, when there is no word-to-word alignment, the resulting mask tends to be imprecise due to its reliance on cross-attention maps. Furthermore, as word-to-word alignment is not a practical assumption in most of the image editing tasks, this approach often fails to preserve structure and identity. To address this challenge, we propose a mask extraction method, which is applied before the editing process. Our approach involves: (i) identifying the edited areas from the editing instruction via an LLM and creating corresponding masks before image generation, and (ii) integrating these masks during the editing process to ensure seamless fusion of edited regions with the original image. Distinct editing challenges, such as adding or removing objects, require tailored solutions. We utilize various techniques, including dilation and Gaussian blurring, to refine the masks. Further description of the method can be found at supplementary Sec. ??", + "bbox": [ + 75, + 295, + 472, + 779 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Filtering. We employ a comprehensive filtering approach to ensure the fidelity of the dataset. This includes: (i) using the task predictor (Sec. 4.2) to reassign samples with instructions that should belong to another task, (ii) applying CLIP filtering metrics [2], (iii) employing structure preserving filtering based on the L1 distance between the depth map of the input image and the edited image, and (iv) applying image detectors to validate the presence (in Add task),", + "bbox": [ + 75, + 780, + 470, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "the absence (in Remove task) or replacement (in Local task) of elements, according to the objects specified in the instruction. This process filters out $70\\%$ of the data, resulting in a final dataset of ten million samples.", + "bbox": [ + 496, + 90, + 890, + 151 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Method", + "text_level": 1, + "bbox": [ + 498, + 165, + 591, + 181 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Emu Edit is a diffusion model designed to multi-task across a broad spectrum of editing tasks. As Emu Edit is trained on various tasks, a crucial aspect is the ability to identify the semantic edit (e.g., global/local/texture) that needs to be applied, based on the user instruction. To address this, we use task-specific embeddings integrated into the model, learned during training. After training, Emu Edit can quickly adapt to new tasks by learning a new task embedding through few-shot learning, without altering the rest of the model. We follow next with a detailed description of each part of our method.", + "bbox": [ + 496, + 190, + 890, + 354 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1. Architecture", + "text_level": 1, + "bbox": [ + 498, + 366, + 633, + 381 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our model builds upon the foundation set by Emu, which is outlined in [6]. The Emu model is a two-stage approach that begins with a pre-training phase and concludes with a quality fine-tuning stage. Emu adapted the latent diffusion model architecture [18] to support high-resolution image generation and incorporated a 16-channel autoencoder with encoder $E$ and decoder $D$ . In the following section, we adapt the notation of [2]. A large U-Net, $\\epsilon_{\\theta}$ , with 2.8 billion parameters, $\\theta$ , text embeddings from CLIP ViT-L [14] and T5-XXL [15], and a substantial pre-training dataset of 1.1 billion images facilitate the model's ability to learn complex semantics and finer details, with a noise-offset strategy contributing to high-contrast and aesthetically pleasing image generation. Given the encoded latent of an image $z = E(x)$ , the diffusion process generates a noisy latent $z_{t}$ where the noise level increases over timesteps $t \\in T$ . To convert Emu to an instruction-based image editing model, we condition it on the image to be modified $c_{I}$ and the instruction $c_{T}$ . Emu Edit is trained to minimize the following optimization problem,", + "bbox": [ + 496, + 388, + 890, + 691 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {\\theta} \\mathbb {E} _ {y, \\epsilon , t} \\left[ \\| \\epsilon - \\epsilon_ {\\theta} \\left(z _ {t}, t, E \\left(c _ {I}\\right), c _ {T}\\right) \\| _ {2} ^ {2} \\right] \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 563, + 694, + 890, + 715 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\epsilon \\in N(0,1)$ is the noise added by the diffusion process and $y = (c_{T},c_{I},x)$ is a triplet of instruction, input image and target image from the dataset. In practice, we initialize the weights of Emu Edit with the weights of Emu. To support the image conditioning, we follow [2] and increase the number of input channels. New weights are initialized to zero. During inference, we perform classifier-free guidance on both image and text conditions. In our experiments we use a scale of $\\gamma_I = 1.5$ for the image condition and $\\gamma_T = 5.0$ for the text condition. Furthermore, we apply a rescaling of the diffusion scheduler to achieve a zero signal-to-noise ratio (SNR) at the terminal timestamp, as suggested", + "bbox": [ + 496, + 719, + 893, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "8874", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/0119e04a5b0b1c102feed425ba9973d27d070e0a1684db23b73de9d48474c9ea.jpg", + "image_caption": [ + "Make it play a rainbow colored trumpet", + "Local. Baseslines struggle to perform intricate local edits.", + "Figure 3. Failure cases of baseline instruction-based image editing models.2" + ], + "image_footnote": [], + "bbox": [ + 225, + 90, + 769, + 445 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "in [10]. This is crucial in order to avoid any mismatch between the model's training and testing phases. For more implementation details see supplementary Sec. ??", + "bbox": [ + 75, + 488, + 470, + 532 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Learned Task Embeddings", + "text_level": 1, + "bbox": [ + 76, + 542, + 318, + 559 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To guide the generation process toward the correct task, we learn an embedding vector for each task in the dataset. During training, given a sample from our dataset, we use the task index, $i$ , to fetch the task's embedding vector, $v_{i}$ , from an embedding table, and optimize it jointly with the model weights. We do so by introducing the task embedding $v_{i}$ as an additional condition to the U-Net, $\\epsilon_{\\theta}$ . Concretely, we integrate the task embedding into the U-Net via crossattention interactions, and by adding it to the timestep embeddings. The optimization problem is updated to", + "bbox": [ + 75, + 566, + 468, + 718 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {\\theta , v _ {1}, \\dots , v _ {k}} \\mathbb {E} _ {\\hat {y}, \\epsilon , t} \\left[ \\| \\epsilon - \\epsilon_ {\\theta} \\left(z _ {t}, t, E \\left(c _ {I}\\right), c _ {T}, v _ {i}\\right) \\| _ {2} ^ {2} \\right] \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 116, + 720, + 468, + 744 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $k$ is the total number of tasks in our dataset and $\\hat{y} = (c_I, c_T, x, i)$ is a quadruplet of input image, input instruction text, target image, and task index from the dataset. Task-specific conditioning arises from the observation that models lacking such conditioning can become perplexed about the type of edit required, particularly when the instructions are complex or the edit type is ambiguous. For", + "bbox": [ + 75, + 747, + 468, + 853 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "instance, as visualized in Fig. 4, (1) a model without task conditioning might perform a global edit when a texture edit is required, (2) it might opt for segmentation when a global edit is necessary, and (3) it could implement a style edit in situations where a local edit would fit better. In the inference stage, we predict the task index. Specifically, we fine-tune a Flan-T5-XL model to identify the task at hand given the input instruction.", + "bbox": [ + 496, + 488, + 890, + 609 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.3. Task Inversion", + "text_level": 1, + "bbox": [ + 498, + 621, + 648, + 636 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To enable few-shot learning of new tasks without losing the general abilities of Emu Edit, we propose a method for adapting the model without changing the U-Net weights. Given a few examples of a new task, we learn a new task embedding, $v_{\\mathrm{new}}$ . We freeze the model weights, and adapt it to the task only through the task embedding. Thus, to fit a new task embedding we solve the following optimization problem:", + "bbox": [ + 496, + 645, + 890, + 765 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {v _ {\\text {n e w}}} \\mathbb {E} _ {y, \\epsilon , t} \\left[ \\| \\epsilon - \\epsilon_ {\\theta} \\left(z _ {t}, t, E \\left(c _ {I}\\right), c _ {T}, v _ {\\text {n e w}}\\right) \\| _ {2} ^ {2} \\right] \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 547, + 768, + 890, + 792 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $v_{\\mathrm{new}}$ is the learned task embedding. Note that during task inversion $y$ is a triplet belonging to the new task.", + "bbox": [ + 496, + 794, + 890, + 823 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The model can then be employed for the new task by conditioning it on the learned task embedding, and it can still handle its original tasks by relying on the initial task embeddings. In Sec. 5.3, we demonstrate that our model effectively generalizes to novel tasks using this method.", + "bbox": [ + 496, + 825, + 890, + 898 + ], + "page_idx": 4 + }, + { + "type": "page_footnote", + "text": "2The samples depicted in this caption were selected by the authors. They do not cover all scenarios, but are meant to represent some common scenarios the authors encountered.", + "bbox": [ + 75, + 862, + 468, + 898 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "8875", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/c92be4727be869365485f5df5ea2528035c7666e91eb0d08737d1f4c7224804c.jpg", + "image_caption": [ + "Figure 4. Task embeddings. Model trained without task embeddings may get confused about the edit type when the instructions are complicated or there is ambiguity regarding the edit type: (1) Global edit (instead of Texture), (2) Segmentation (instead of Global), (3) Style edit (instead of Local)." + ], + "image_footnote": [], + "bbox": [ + 71, + 88, + 467, + 375 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Sequential Edit Thresholding. We notice that applying the model repeatedly, in multi-turn editing scenarios, aggregates reconstruction and numerical errors, which translate to noticeable artifacts. To mitigate this, we add a per-pixel thresholding step after each edit-turn. At each step $s$ , we use the pixel value in the output image, $c_{I}^{s+1}$ , only if its alteration surpasses a specific threshold. Otherwise, we keep the pixel value from the input image, $c_{I}^{s}$ . Specifically, given an edit turn $s$ , we compute the absolute difference image $d = \\|c_{I}^{s+1} - c_{I}^{s}\\|_{1}$ over the RGB channel, and apply the following thresholding:", + "bbox": [ + 75, + 463, + 468, + 630 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nc _ {I} ^ {s + 1} = \\left\\{ \\begin{array}{l l} c _ {I} ^ {s} & \\text {i f} \\bar {d} < \\alpha , \\\\ c _ {I} ^ {s + 1} & \\text {o t h e r w i s e .} \\end{array} \\right. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 178, + 643, + 468, + 683 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where, $\\bar{d}$ is obtained after passing $d$ through a low pass filter, in order to smooth the transition between previous and current pixels. In practice, we choose $\\alpha = 0.03$ . Please refer to supplementary Sec. ?? for a qualitative comparison. In Fig. 2 we show examples of multi-turn editing.", + "bbox": [ + 75, + 691, + 468, + 768 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 76, + 782, + 209, + 800 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Our experiments evaluate the ability of Emu Edit to follow user instructions faithfully and preserve the visual fidelity of the original image. First, we evaluate the performance of our approach on instruction-based image editing tasks. Second, we conduct a comprehensive ablation study to assess the effectiveness of our different contributions. Specifically,", + "bbox": [ + 75, + 809, + 470, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "we ablate the contribution of the computer vision tasks to the model performance on image editing tasks, the importance of learned task embeddings, and the effect of multitask learning on instruction-based image editing. Further ablation on our data generation pipeline can be found in the supplementary Sec. ?? Finally, we demonstrate our model's ability to learn new tasks via few-shot learning.", + "bbox": [ + 496, + 90, + 892, + 196 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Measures. We employ two main measures in our evaluation: edit text alignment and image faithfulness. Specifically, for each pair of input image and editing instruction, we use the following automatic metrics: (i) CLIP [14] text-image direction similarity $(\\mathrm{CLIP}_{dir})$ - measuring agreement between change in captions and the change in images, (ii) CLIP image similarity $(\\mathrm{CLIP}_{img})$ - measuring change between edited and input image, (iii) CLIP output similarity $(\\mathrm{CLIP}_{out})$ - measuring edited image similarity with output caption, (iv) L1 pixel-distance between input and edit image, and (v) DINO [4] similarity between the DINO embeddings of input and edited images. With the exception of the L1 distance, where lower values indicate better performance, higher values in all other measures signify better results. A low L1 distance translates to small changes in image's pixel values. A high DINO and $\\mathrm{CLIP}_{img}$ similarity score, suggests semantic similarity between the images. For region-based edits, high image similarity scores indicate the edits were precise. For free-form edits, high similarity scores indicate image structure preservation. $\\mathrm{CLIP}_{dir}$ and $\\mathrm{CLIP}_{out}$ measure how well the model followed the instruction. In addition, we asked human raters to evaluate the text alignment and image faithfulness. In each evaluation scenario, raters are presented with two modified images alongside the original input image and instruction, and are presented with two questions: (i) Image Faithfulness: which image better preserves elements in the input image, and (ii) Text Alignment: which image best follows the instruction.", + "bbox": [ + 496, + 217, + 892, + 640 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1. Evaluation", + "text_level": 1, + "bbox": [ + 500, + 650, + 620, + 664 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Throughout the paper, we report results on the MagicBrush test set [25] and the Emu Edit benchmark. In the following section, we describe our motivation for creating this benchmark, and detail its curation process. To date, there are two main benchmarks for evaluating instruction-based image editing capabilities. First, the InstructPix2Pix benchmark [2], which is intrinsically biased due to its reliance on generated Stable Diffusion [17] input images, and GPT-3 [3] generated instructions. Consequently, it is unclear whether its results will truly mirror the performance on real input images, with genuine user instructions.", + "bbox": [ + 496, + 672, + 890, + 839 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Unlike InstructPix2Pix, the second benchmark, MagicBrush [25], uses a diverse set of authentic input images from the MS-COCO benchmark [5, 11], and annotator-defined instructions. Nonetheless, this dataset also suffers", + "bbox": [ + 496, + 839, + 890, + 898 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "8876", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/55413ddfd51fa795535859ef7be038787113fdfe1781e61c9ab6f56467788b82.jpg", + "table_caption": [ + "Table 2. Comparison with image-editing baselines evaluated on Emu Edit test set and MagicBrush test set. For each benchmark we report CLIP, L1, DINO metrics and human ratings. Human evaluation shows the percentage of raters that prefer the results of Emu Edit." + ], + "table_footnote": [], + "table_body": "
Emu Edit Test SetMagicBrush Test Set
MethodCLIPdir↑CLIPim↑CLIPout↑L1↓DINO↑Text align. faith.CLIPdir↑CLIPim↑CLIPout↑L1↓DINO↑Text align. faith.
InstructPix2Pix [2]0.0780.8340.2190.1210.76277.3376.710.1150.8370.2450.0930.767
MagicBrush [25]0.0900.8380.2220.1000.77674.5074.100.1230.8830.2610.0580.871
PnP [21]0.0280.5210.0890.3040.15398.9599.000.0250.5680.1010.2890.220
Null-Text Inv. [12]0.1010.7610.2360.0750.67881.6385.470.1210.7520.2630.0770.664
Our0.1090.8590.2310.0940.819--0.1350.8970.2610.0520.879
", + "bbox": [ + 84, + 127, + 883, + 252 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "from inherent bias. During data collection, annotators were directed to use the DALLE-2 image editing platform [13] to generate the edited images. Thus, this benchmark is biased towards editing instructions that the DALLE-2 editor can successfully follow, which may compromise both its diversity and complexity.", + "bbox": [ + 75, + 258, + 470, + 349 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Emu Edit Benchmark. To collect a dataset with reduced bias and of higher diversity, we take a different approach. We first define seven different categories of potential image editing operations: background alteration (Background), comprehensive image changes (Global), style alteration (Style), object removal (Remove), object addition (Add), localized modifications (Local), and color/texture alterations (Texture). Then, we utilize the diverse set of input images from the MagicBrush benchmark [25], and for each editing operation, we task crowd workers to devise relevant, creative, and challenging instructions. Moreover, to increase the quality of the collected examples, we apply a post-verification stage, in which crowd workers filter examples with irrelevant instructions. Finally, to support evaluation for methods that require input and output captions [8, 21], we additionally collect an input caption and output caption. When doing so, we ask annotators to ensure that the captions capture both important elements in the image, and elements that should change based on the instruction. See supplementary Sec. ?? for examples of our benchmark, which we release to support better evaluation.", + "bbox": [ + 75, + 351, + 470, + 667 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Baseline Comparisons. We compare our model against two instruction-based image editing baseline models: InstructPix2Pix [2], and MagicBrush [25], which is a variant of InstructPix2Pix that was fine-tuned on the MagicBrush dataset. Additionally, we compare our model against two text-based image editing methods: PNP [21] and Null-Text Inversion modification of P2P [8, 12]. Unlike instruction-based models, these works expect image descriptions. Therefore, we provide them with access to the input caption and output caption. Note however, providing these methods with access to the ground-truth captions could potentially offer an advantage over instruction-based models, since the automatic metrics also rely on these captions. Tab. 2 shows our results versus the baselines. The findings indicate that human raters consistently prefer Emu", + "bbox": [ + 75, + 672, + 468, + 901 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Edit over all baselines by a large margin. Furthermore, apart from Null-Text Inversion, which as explained above, utilizes the ground-truth captions during inference, our approach outperforms the existing baselines on the automatic metrics. We provide qualitative comparisons in Fig. 3. Additional comparisons are available in the supplementary: Fig.??-??, and Fig.??-??. For performance on vision tasks, see Sec. ??", + "bbox": [ + 496, + 258, + 890, + 380 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.2. Ablations", + "text_level": 1, + "bbox": [ + 500, + 388, + 609, + 402 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Computer Vision Tasks Enhance Image Editing Tasks. Here we demonstrate the importance of the vision tasks to Emu Edit performance on image editing tasks. For this, we trained two additional models on all tasks except: (i) detect and segment tasks, and (ii) image-to-image translation tasks. As we show in Tab. 4, adding the detection and segmentation tasks improves the model performance in region-based editing tasks. Additionally, we observe that image-to-image translation tasks improve the performance in free-form editing tasks. We hypothesize that the recognition tasks improve the model's recognition capabilities, leading to more accurate and precise localized modifications. Similarly, image-to-image tasks assist the model in understanding the entire image structure, thereby enhancing its capabilities for global operations.", + "bbox": [ + 496, + 411, + 890, + 638 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Contribution of Learned Task Embeddings. We compare three variants of Emu Edit: (i) conditioned on the ground-truth task embedding, (ii) conditioned on the task embedding, as predicted by the task predictor described in Sec. 4.2, and (iii) without conditioning on the task type.3 Tab. 3 shows the results on the validation set of our benchmark. As can be seen, conditioning on the task type boosts the model's performance. Furthermore, our task predictor closes the gap with the ground-truth conditioned model. Qualitatively, we observe that without conditioning on the task type the model may perform the wrong editing operation (Fig. 4). In supplementary Fig. ??, we demonstrate the effect of manipulating the task while keeping the instruction and input image fixed. As can be seen, changing the task embedding directly influences the task executed by the model.", + "bbox": [ + 496, + 638, + 890, + 878 + ], + "page_idx": 6 + }, + { + "type": "page_footnote", + "text": "$^{3}$ (iii) was trained without learned task embeddings.", + "bbox": [ + 514, + 886, + 787, + 898 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "8877", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/ebfc35572c3933b3b9c1758b5151d2d8f738422fda0528692d61750c3af3b47c.jpg", + "table_caption": [ + "Table 3. Learned task embeddings ablation on our validation set. We compare variations of Emu Edit: without task type condition, with predicted task type, and with ground-truth task type." + ], + "table_footnote": [], + "table_body": "
MethodCLIPdir↑CLIPim↑CLIPout↑L1↓DINO↑
w/o task emb.0.1040.8430.2270.1090.792
with pred. task0.1170.8500.2310.1030.809
with gt task0.1190.8520.2310.1000.811
", + "bbox": [ + 94, + 136, + 452, + 203 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/22fe06395b2e6e58c3c9786971b2175a3f3147a988a86fc14b95c9e96f9ae557.jpg", + "table_caption": [ + "Table 4. Contribution of computer vision tasks. Human evaluation is shown as a percentage of majority votes in favor of our model." + ], + "table_footnote": [], + "table_body": "
Region-BasedFree-form
MethodText align.Image faith.Text align.Image faith.
without detect/segment60.060.252.351.5
without im2im translation50.249.058.060.1
", + "bbox": [ + 117, + 241, + 433, + 327 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/39123bad4f47e58ef3229c1e7a2b69267c6105a096a1c393a5d87bbbf0acd560.jpg", + "image_caption": [ + "(i) Incorporate a bee into the bag's pattern and detect it" + ], + "image_footnote": [], + "bbox": [ + 174, + 339, + 292, + 431 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/550332dd0b724988cbe59bd88711563be1cac0be2f343cc947851469b953494d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 292, + 339, + 408, + 431 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/28a809ddf6b49196190fc05506c4b8702c52aba451060a4af490f8cb3a107582.jpg", + "image_caption": [ + "(ii) Mark the gift bags", + "Input", + "Figure 5. Generations on unseen tasks with task inversion. (i) composition of add and detect tasks, (ii) object contour detection." + ], + "image_footnote": [], + "bbox": [ + 173, + 445, + 290, + 537 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/92491dca5918554091ff7fed3d219e548fc2bb2af9abd1eb4159a98124759791.jpg", + "image_caption": [ + "Emu Edit" + ], + "image_footnote": [], + "bbox": [ + 292, + 445, + 408, + 537 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Multi-Task versus Expert Models. We show that training a single model on a diverse range of tasks, leads to enhanced performance in each individual task, outperforming models that are specifically trained for a single task. To validate this, we train an expert model for each task, and compare its performance to the single one (See supplementary Tab. ??).", + "bbox": [ + 75, + 594, + 468, + 684 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Influence of Number of Tasks. Here, we ablate the number of tasks participating in the multi-task training scheme. In supplementary Fig. ?? we report the average $\\mathrm{CLIP}_{dir}$ on the Style and Texture tasks when iteratively excluding other tasks, and training a model on the new tasks list. As can be seen, augmenting the model with additional tasks leads to improved performance, even in tasks which are not directly associated with the added ones.", + "bbox": [ + 75, + 684, + 468, + 805 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.3. Few-Shot Learning of New Tasks", + "text_level": 1, + "bbox": [ + 76, + 816, + 366, + 832 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Finally, we evaluate our model's ability to generalize by testing it in a few-shot scenario with previously unseen tasks. We test its performance across the following tasks: (i) super-resolution (x4), (ii) object contour detection, (iii)", + "bbox": [ + 75, + 839, + 468, + 901 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/f4e4f96d972ec08ff045d7ace8a5e94898ace017418c3deb72d8af9abf0a0aea.jpg", + "image_caption": [ + "(a) Inpainting", + "Figure 6. Few-shot performance for different tasks over 1, 10, and 100 samples. Each line represents a different training setting: 'Scratch' finetune (Blue, $\\bigcirc$ ), Emu Edit finetune (Orange, $\\square$ ), task inversion (Green, $\\triangle$ ), all compared to an upper-bound expert trained on 100k samples (Red dashed line, $\\diamond$ )." + ], + "image_footnote": [], + "bbox": [ + 488, + 107, + 674, + 212 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/d7be06768fe65bd0afeb7e4b6fe4566b2c930661aae31bb2c587ffc2c786595f.jpg", + "image_caption": [ + "(b) Add + Detect" + ], + "image_footnote": [], + "bbox": [ + 684, + 107, + 870, + 212 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "mask-based inpainting, and (iv) a composite task formed by combining two tasks from our dataset: add and detect. For each task, we assess the model's performance when trained with 0, 1, and 100 examples, with the 0-example case being equivalent to a zero-shot setting. We compare three baselines: (i) Scratch - Emu Edit initialized with Emu's weights, trained on the examples, (ii) Task Inversion - Emu Edit with task inversion (Sec. 4.3), and (iii) Finetune - Emu Edit where we finetune all of the model's weights on the examples. As an upper-bound expert, we train the first baseline on 100,000 examples. Note that, the \"Task Inversion\" and \"Finetune\" baselines were trained on the multitask dataset whereas the \"Scratch\" baseline was not. As can be seen in Fig. 6, fine-tuning with a single example is enough to significantly enhance the performance, while training from scratch results in overfitting. Moreover, utilizing 100 samples nearly achieves expert-level performance, implying that the model can effectively generalize to novel tasks. We also observe that task inversion performs similarly to complete fine-tuning, indicating that the model already contains essential information and can be prompted with a new task embedding to yield the desired result. For performance and generation examples on additional tasks see Fig. 5 and supplementary ????.", + "bbox": [ + 496, + 324, + 890, + 686 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 693, + 617, + 708 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Emu Edit presents a step change in instructable image editing capabilities, primarily due to its unique training on both recognition and generation tasks. This dual-focus approach significantly enhances the model's comprehension of natural language instructions, enabling it to accurately execute a wide array of editing operations. Its ability to generalize to new tasks with minimal examples further demonstrates its versatility and robustness. Furthermore, our framework has the potential for further integration with a multimodal LLM in future projects, particularly for complex editing tasks requiring detailed reasoning from the input image, such as counting objects or executing intricate tasks.", + "bbox": [ + 496, + 719, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8878", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 174, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Omri Avrahami, Dani Lischinski, and Ohad Fried. Blended diffusion for text-driven editing of natural images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18208-18218, 2022. 3", + "[2] Tim Brooks et al. Instructpix2pix: Learning to follow image editing instructions. In CVPR, 2023. 1, 3, 4, 6, 7", + "[3] Tom B. Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-Voss, Gretchen Krueger, T. J. Henighan, Rewon Child, Aditya Ramesh, Daniel M. Ziegler, Jeff Wu, Clemens Winter, Christopher Hesse, Mark Chen, Eric Sigler, Mateusz Litwin, Scott Gray, Benjamin Chess, Jack Clark, Christopher Berner, Sam McCandlish, Alec Radford, Ilya Sutskever, and Dario Amodei. Language models are few-shot learners. ArXiv, abs/2005.14165, 2020. 3, 6", + "[4] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9650-9660, 2021. 6", + "[5] Xinlei Chen, Hao Fang, Tsung-Yi Lin, Ramakrishna Vedantam, Saurabh Gupta, Piotr Dólar, and C. Lawrence Zittnick. Microsoft coco captions: Data collection and evaluation server. ArXiv, abs/1504.00325, 2015. 6", + "[6] Xiaoliang Dai, Ji Hou, Chih-Yao Ma, Sam Tsai, Jialiang Wang, Rui Wang, Peizhao Zhang, Simon Vandenhende, Xiaofang Wang, Abhimanyu Dubey, et al. Emu: Enhancing image generation models using photogenic needles in a haystack. arXiv preprint arXiv:2309.15807, 2023. 4", + "[7] Oran Gafni, Adam Polyak, Oron Ashual, Shelly Sheynin, Devi Parikh, and Yaniv Taigman. Make-a-scene: Scene-based text-to-image generation with human priors. In European Conference on Computer Vision, pages 89-106. Springer, 2022. 2", + "[8] Amir Hertz, Ron Mokady, Jay M. Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-prompt image editing with cross attention control. ArXiv, abs/2208.01626, 2022. 2, 3, 7", + "[9] Bahjat Kawar et al. Imagic: Text-based real image editing with diffusion models. 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6007-6017, 2022. 3", + "[10] Shanchuan Lin, Bingchen Liu, Jiashi Li, and Xiao Yang. Common diffusion noise schedules and sample steps are flawed. arXiv preprint arXiv:2305.08891, 2023. 5", + "[11] Tsung-Yi Lin, Michael Maire, Serge J. Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C. Lawrence Zitnick. Microsoft coco: Common objects in context. In ECCV, 2014. 6", + "[12] Ron Mokady, Amir Hertz, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Null-text inversion for editing real images using guided diffusion models. 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6038–6047, 2022. 3, 7" + ], + "bbox": [ + 78, + 114, + 470, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[13] OpenAI. Dall-e 2, 2022. https://openai.com/product/dall-e-2.3,7", + "[14] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 1, 4, 6", + "[15] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J. Liu. Exploring the limits of transfer learning with a unified text-to-text transformer. Journal of Machine Learning Research, 21(140):1-67, 2020. 4", + "[16] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022. 2", + "[17] Robin Rombach, A. Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10674-10685, 2021. 2, 6", + "[18] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models, 2021. 4", + "[19] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L. Denton, Seyed Kamyar Seyed Ghasemipour, Burcu Karagol Ayan, Seyedeh Sara Mahdavi, Raphael Gontijo Lopes, Tim Salimans, Jonathan Ho, David J. Fleet, and Mohammad Norouzi. Photorealistic text-to-image diffusion models with deep language understanding. ArXiv, abs/2205.11487, 2022. 2", + "[20] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288, 2023. 3", + "[21] Narek Tumanyan, Michal Geyer, Shai Bagon, and Tali Dekel. Plug-and-play diffusion features for text-driven image-to-image translation. 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1921-1930, 2022. 3, 7", + "[22] Bram Wallace, Akash Gokul, and Nikhil Vijay Naik. Edict: Exact diffusion inversion via coupled transformations. 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 22532-22541, 2022. 3", + "[23] Su Wang et al. Imagen editor and editbench: Advancing and evaluating text-guided image inpainting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18359-18369, 2023. 3", + "[24] Shaoan Xie, Zhifei Zhang, Zhe Lin, Tobias Hinz, and Kun Zhang. Smartbrush: Text and shape guided object inpainting with diffusion model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22428-22437, 2023. 3", + "[25] Kai Zhang et al. Magicbrush: A manually annotated dataset for instruction-guided image editing. arXiv:2306.10012, 2023. 1, 3, 6, 7" + ], + "bbox": [ + 501, + 92, + 890, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "8879", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + } +] \ No newline at end of file diff --git a/2024/Emu Edit_ Precise Image Editing via Recognition and Generation Tasks/b6371d78-25ca-4b11-84b1-a566ed9d62f2_model.json b/2024/Emu Edit_ Precise Image Editing via Recognition and Generation Tasks/b6371d78-25ca-4b11-84b1-a566ed9d62f2_model.json new file mode 100644 index 0000000000000000000000000000000000000000..7d3af0772125d23a2b052a95b2d783d01f3b0a35 --- /dev/null +++ b/2024/Emu Edit_ Precise Image Editing via Recognition and Generation Tasks/b6371d78-25ca-4b11-84b1-a566ed9d62f2_model.json @@ -0,0 +1,2033 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.0, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.124, + 0.131, + 0.848, + 0.154 + ], + "angle": 0, + "content": "Emu Edit: Precise Image Editing via Recognition and Generation Tasks" + }, + { + "type": "text", + "bbox": [ + 0.124, + 0.18, + 0.846, + 0.217 + ], + "angle": 0, + "content": "Shelly Sheynin*, Adam Polyak*, Uriel Singer*, Yuval Kirstain*, Amit Zohar*, Oron Ashual, Devi Parikh and Yaniv Taigman" + }, + { + "type": "text", + "bbox": [ + 0.43, + 0.234, + 0.542, + 0.251 + ], + "angle": 0, + "content": "GenAI,Meta" + }, + { + "type": "image", + "bbox": [ + 0.084, + 0.274, + 0.351, + 0.378 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.146, + 0.38, + 0.31, + 0.391 + ], + "angle": 0, + "content": "Dress the emu with a fireman outfit" + }, + { + "type": "image", + "bbox": [ + 0.353, + 0.274, + 0.618, + 0.378 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.446, + 0.38, + 0.551, + 0.391 + ], + "angle": 0, + "content": "Let's see it graduating" + }, + { + "type": "image", + "bbox": [ + 0.62, + 0.274, + 0.888, + 0.378 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.717, + 0.38, + 0.792, + 0.39 + ], + "angle": 0, + "content": "Mark the drinks" + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.401, + 0.89, + 0.417 + ], + "angle": 0, + "content": "Figure 1. Emu Edit is a multi-tasking model that combines various editing (left, middle) and vision (right) tasks for precise image editing." + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.428, + 0.314, + 0.443 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.458, + 0.474, + 0.82 + ], + "angle": 0, + "content": "Instruction-based image editing holds immense potential for a variety of applications, as it enables users to perform any editing operation using a natural language instruction. However, current models in this domain often struggle with accurately executing user instructions. We present Emu Edit, a multi-task image editing model which sets state-of-the-art results in instruction-based image editing. To develop Emu Edit we train it to multi-task across an unprecedented range of tasks, such as region-based editing, free-form editing, and Computer Vision tasks, all of which are formulated as generative tasks. Additionally, to enhance Emu Edit's multi-task learning abilities, we provide it with learned task embeddings which guide the generation process towards the correct edit type. Both these elements are essential for Emu Edit's outstanding performance. Furthermore, we show that Emu Edit can generalize to new tasks, such as image inpainting, super-resolution, and compositions of editing tasks, with just a few labeled examples. This capability offers a significant advantage in scenarios where high-quality samples are scarce. Lastly, to facilitate a more rigorous and informed assessment of instructable image editing models, we release a new challenging and versatile benchmark that includes seven different image editing tasks." + }, + { + "type": "title", + "bbox": [ + 0.502, + 0.428, + 0.632, + 0.443 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.463, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Image editing is a widely-used application that millions engage with every day. Popular image editing tools, however, either demand considerable expertise and are time-consuming to use, or are quite limited, providing only a predefined set of editing operations, such as specific filters. Instruction-based image editing [2, 25] attempts to resolve these limitations by allowing users to effortlessly describe their editing goals using natural language instructions. For instance, a user can provide a model with an image and instruct it to \"Dress the emu with a fireman outfit\" or \"Let's see it graduating\" (see Fig. 1). Nevertheless, while instruction-based image editing models like InstructPix2Pix [2] are designed to process any given instruction, they often struggle to accurately interpret and execute such instructions. Moreover, their generalization is limited, often falling short on tasks that deviate slightly from those they were trained on (see Fig. 3). To address these gaps, we introduce Emu Edit, the first image editing model trained on an extensive and diverse set of tasks, including both image editing and computer vision tasks. Emu Edit provides a substantial improvement in both compliance with the edit instruction and preservation of the visual fidelity of the original image. As we show through both automatic metrics [14] and human judgments on two benchmarks [25], Emu Edit achieves state-of-the-art results in instruction-based image editing. The success of Emu Edit stems from two key contributions. First, we train our model to multi-task across sixteen distinct image editing tasks. These tasks span region-based editing tasks, free-form editing tasks and computer" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.875, + 0.437, + 0.888 + ], + "angle": 0, + "content": "1Project Page: https://emu-edit.metademolab.com/" + }, + { + "type": "page_footnote", + "bbox": [ + 0.098, + 0.888, + 0.207, + 0.9 + ], + "angle": 0, + "content": "*Equal contribution." + }, + { + "type": "list", + "bbox": [ + 0.096, + 0.875, + 0.437, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "8871" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.081, + 0.103, + 0.242, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.138, + 0.228, + 0.186, + 0.239 + ], + "angle": 0, + "content": "(a) A cat" + }, + { + "type": "image", + "bbox": [ + 0.242, + 0.103, + 0.404, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.272, + 0.228, + 0.375, + 0.239 + ], + "angle": 0, + "content": "(b) Remove the tail" + }, + { + "type": "image", + "bbox": [ + 0.405, + 0.103, + 0.566, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.429, + 0.228, + 0.541, + 0.239 + ], + "angle": 0, + "content": "(c) Add a pink jacket" + }, + { + "type": "image", + "bbox": [ + 0.567, + 0.103, + 0.728, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.603, + 0.228, + 0.693, + 0.239 + ], + "angle": 0, + "content": "(d)Make it rainy" + }, + { + "type": "image", + "bbox": [ + 0.728, + 0.103, + 0.892, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.731, + 0.228, + 0.888, + 0.239 + ], + "angle": 0, + "content": "(e) Have the cat look shocked" + }, + { + "type": "image", + "bbox": [ + 0.081, + 0.242, + 0.242, + 0.366 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.095, + 0.367, + 0.229, + 0.379 + ], + "angle": 0, + "content": "(f) Extract the depth map" + }, + { + "type": "image", + "bbox": [ + 0.242, + 0.242, + 0.414, + 0.366 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.254, + 0.367, + 0.393, + 0.403 + ], + "angle": 0, + "content": "(g) Generate a rainy day image of a hedgehog in a dress using the depth map" + }, + { + "type": "image", + "bbox": [ + 0.415, + 0.242, + 0.566, + 0.366 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.408, + 0.367, + 0.564, + 0.39 + ], + "angle": 0, + "content": "(h) Replace the dress with an astronaut outfit" + }, + { + "type": "image", + "bbox": [ + 0.567, + 0.242, + 0.728, + 0.366 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.567, + 0.367, + 0.728, + 0.39 + ], + "angle": 0, + "content": "(i) Segment the spacesuit, and detect the hands" + }, + { + "type": "image", + "bbox": [ + 0.729, + 0.242, + 0.892, + 0.366 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.733, + 0.367, + 0.886, + 0.391 + ], + "angle": 0, + "content": "(j) Add the text \"purple cat\" using a purple font" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.415, + 0.893, + 0.443 + ], + "angle": 0, + "content": "Figure 2. Instruction-based image generation. Each image originates from a previous one with a corresponding instruction. The prior image for (a) was set to zero." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.448, + 0.473, + 0.903 + ], + "angle": 0, + "content": "vision tasks, all formulated as generative tasks. Unlike previous work, we develop a distinct data generation pipeline for each task, allowing us to gather a training set that is not only more diverse but also more precise in its examples. We find that training a single model on all tasks yields better results than training expert models on each task independently. Additionally, we show that as the number of training tasks increases, so does the performance of Emu Edit. Furthermore, we discover that surprisingly, computer vision tasks such as detection, segmentation, and others, significantly enhance editing performance, as validated both by human raters as well as quantitative measures. Second, to process this wide array of tasks effectively, we introduce the concept of learned task embeddings, which are used to steer the generation process toward the correct generative task. Concretely, for each task, we learn a unique task embedding vector, and integrate it into the model through cross-attention interactions, and by adding it to the timestep embeddings. We demonstrate that learned task embeddings significantly enhance the ability of our model to accurately infer the appropriate edit type from the user instruction and execute the correct edit. Equipped with a robust model trained across a broad spectrum of tasks and guided by learned task embeddings, we explore few-shot adaptation to unseen tasks via task inversion. In this process, we maintain the model weights untouched, and solely update a task embedding to fit the new task. Our experiments demonstrate that Emu Edit can swiftly adapt to new tasks, such as superresolution, contour detection, and others. Notably, for some tasks, fine-tuning the model on just a handful of examples" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.448, + 0.895, + 0.751 + ], + "angle": 0, + "content": "yields results that nearly match those of an expert model trained on one hundred thousand examples. This makes task inversion with Emu Edit particularly advantageous in scenarios where labeled examples are limited, or when the compute budget is low. Finally, to support future research for instruction-based image editing, we publicly release a diverse and challenging benchmark that includes seven different image editing operations, as well as our model's generations on this dataset. In summary, this work addresses the limitations of instruction-based image editing models in accurately following user instructions. We demonstrate that by employing multi-task training across a diverse array of tasks, including recognition, generation, and editing, we can enhance our model's performance. Furthermore, by incorporating learned task embeddings into our model's architecture, we not only improve its results but also enable efficient few-shot learning for new tasks. With these improvements, our model sets a new standard in the field, offering significantly more precise and robust instruction-based image editing capabilities than existing models." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.769, + 0.642, + 0.785 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.795, + 0.894, + 0.903 + ], + "angle": 0, + "content": "The emergence of high-performing text-to-image diffusion models [7, 16, 17, 19] facilitated the development of effective text-based image editing methods. Such methods usually employ aligned and detailed descriptions of the input and edited image to perform a specific edit. Prompt-to-Prompt (P2P) [8] injects the input caption attention maps to the target caption attentions maps. Null-Text Inver" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8872" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.091, + 0.473, + 0.363 + ], + "angle": 0, + "content": "sion [12] inverts an input image using the null-text embedding to support editing of a real image. Plug-and-Play (PNP) [21] injects spatial features in addition to attention maps and obtains better performance at global editing. Imagic [9] finetunes the diffusion model to support complex textual instructions. EDICT [22] suggests an image inversion based on two noise vectors enabling better image reconstruction and textual faithfulness. Another class of image editing models, employs an input mask as additional input [1, 23, 24]. Blended Diffusion [1] modifies the diffusion step by blending the input image in the unmasked regions. Imagen Editor [23] and SmartBrush [24] finetune the text-to-image model to be conditioned on both the input image and mask. While the text-based image editing methods detailed above enable humans to edit images, they frequently exhibit inconsistent performance and require multiple inputs, such as aligned and detailed descriptions of both the input and target images, or at times, input masks." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.364, + 0.473, + 0.62 + ], + "angle": 0, + "content": "To offer a more intuitive and user-friendly interface, and significantly enhance ease of use for humans, InstructPix2Pix [2] introduced an instructable image editing model. They developed this model by utilizing both GPT-3 [3] and Prompt-to-Prompt [8], to generate a large synthetic dataset for instruction-based image editing, and employed the dataset to train an instructable image editing model. Unlike InstructPix2Pix which used a synthetic dataset, MagicBrush [25] developed a manually-annotated instruction-guided image editing dataset by requesting humans to use an online image editing tool [13]. Finetuning InstructPix2Pix on this dataset led to improved image editing capabilities. However, even though there has been progress and improvement in instruction-based image editing models, we show in Sec. 5 that state-of-the-art image editing models still struggle with accurately interpreting and precisely executing editing instructions." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.62, + 0.473, + 0.773 + ], + "angle": 0, + "content": "In this paper, we drastically narrow such performance gaps by leveraging multi-task training and a matching architecture. Unlike prior work that solely focuses on image editing [2, 25], we train our model to perform various tasks and learn a very diverse set of capabilities. The quality and versatility of our training procedure and dataset, together with our improved architecture for multi-task learning, enables us to make a big leap in performance and differentiates us from prior work in the field. Fig. 3 include several challenging editing samples as examples." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.785, + 0.418, + 0.803 + ], + "angle": 0, + "content": "3. Multi-Task Dataset for Image Editing" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.81, + 0.472, + 0.903 + ], + "angle": 0, + "content": "Training a robust and accurate image editing model requires a highly diverse dataset of input images, editing instructions, and output edited images. However, manually collecting such examples is impractically time-consuming, existing sources on the web (e.g. communities and forums on social media) are limited in size, and publicly available syn-" + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.094, + 0.676, + 0.109 + ], + "angle": 0, + "content": "1. Region-Based Editing" + }, + { + "type": "text", + "bbox": [ + 0.501, + 0.11, + 0.891, + 0.139 + ], + "angle": 0, + "content": "- Local: Substituting one object for another, altering an object's attributes (e.g., \"make it smile\")." + }, + { + "type": "text", + "bbox": [ + 0.501, + 0.14, + 0.805, + 0.154 + ], + "angle": 0, + "content": "- Remove: Erasing an object from the image." + }, + { + "type": "text", + "bbox": [ + 0.501, + 0.155, + 0.803, + 0.169 + ], + "angle": 0, + "content": "- Add: Inserting a new object into the image." + }, + { + "type": "text", + "bbox": [ + 0.501, + 0.17, + 0.892, + 0.214 + ], + "angle": 0, + "content": "- Texture: Altering an object's visual characteristics without affecting its structure (e.g., painting over, filling or covering an object)." + }, + { + "type": "text", + "bbox": [ + 0.501, + 0.215, + 0.837, + 0.23 + ], + "angle": 0, + "content": "- Background: Changing the scene's background." + }, + { + "type": "list", + "bbox": [ + 0.501, + 0.11, + 0.892, + 0.23 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.237, + 0.655, + 0.251 + ], + "angle": 0, + "content": "2. Free-Form Editing" + }, + { + "type": "text", + "bbox": [ + 0.501, + 0.252, + 0.892, + 0.296 + ], + "angle": 0, + "content": "- Global: Edit instructions that affect the entire image, or that can not be described using a mask (e.g., \"let's see it in the summer\")." + }, + { + "type": "text", + "bbox": [ + 0.501, + 0.297, + 0.756, + 0.312 + ], + "angle": 0, + "content": "- Style: Change the style of an image." + }, + { + "type": "text", + "bbox": [ + 0.501, + 0.312, + 0.891, + 0.356 + ], + "angle": 0, + "content": "- Text Editing: This involves text-related editing tasks such as adding, removing, swapping text, and altering the text's font and color." + }, + { + "type": "list", + "bbox": [ + 0.501, + 0.252, + 0.892, + 0.356 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.364, + 0.607, + 0.377 + ], + "angle": 0, + "content": "3. Vision tasks" + }, + { + "type": "text", + "bbox": [ + 0.501, + 0.379, + 0.891, + 0.409 + ], + "angle": 0, + "content": "- Detect: Identifying and marking a specific object within the image with a rectangle bounding box." + }, + { + "type": "text", + "bbox": [ + 0.501, + 0.41, + 0.882, + 0.424 + ], + "angle": 0, + "content": "- Segment: Isolating and marking an object in the image." + }, + { + "type": "text", + "bbox": [ + 0.501, + 0.424, + 0.88, + 0.438 + ], + "angle": 0, + "content": "Color: Color adjustments like sharpening and blurring." + }, + { + "type": "text", + "bbox": [ + 0.502, + 0.44, + 0.891, + 0.499 + ], + "angle": 0, + "content": "- Image-to-Image Translation: Tasks that involve bidirectional image type conversion, such as sketch-to-image, depth map-to-image, normal map-to-image, pose-to-image, segmentation map-to-image, and so on." + }, + { + "type": "list", + "bbox": [ + 0.501, + 0.379, + 0.891, + 0.499 + ], + "angle": 0, + "content": null + }, + { + "type": "table_caption", + "bbox": [ + 0.508, + 0.515, + 0.882, + 0.529 + ], + "angle": 0, + "content": "Table 1. Description of the tasks forming the Emu Edit dataset." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.54, + 0.892, + 0.676 + ], + "angle": 0, + "content": "thetic datasets often lack in diversity or quality. Therefore, we construct a new dataset that encompasses sixteen distinct tasks and ten million examples. Each example \\((c_{I}, c_{T}, x, i)\\) in our dataset, contains an input image \\(c_{I}\\), a text instruction \\(c_{T}\\), a target image \\(x\\), and a task index \\(i\\) (out of the sixteen). The following section outlines the process of the data construction. In Sec. 3.1 we describe the instruction generation process, and in Sec. 3.2 the image pairs \\((c_{I}, x)\\) generation and filtering." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.698, + 0.892, + 0.775 + ], + "angle": 0, + "content": "Task Categories. The dataset is composed of tasks which are divided into three main categories: region-based editing, free-form editing, and vision tasks. Tab. 1 includes the full list of tasks, and their distribution in the train set is visualized in Fig. ??" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.786, + 0.714, + 0.801 + ], + "angle": 0, + "content": "3.1. Instruction Generation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.81, + 0.892, + 0.901 + ], + "angle": 0, + "content": "To generate editing instructions, we leverage the dialogue-optimized 70 billion parameter Llama 2 variant [20]. We observed that using a single agent to generate the instructions for all tasks leads to a lack of diversity in the dataset. Notably, the LLM exhibits a bias towards particular tasks and instruction phrasings. To address this, we utilize in" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8873" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.258 + ], + "angle": 0, + "content": "context learning to create a task-specific agent for each task. Concretely, we provide the LLM with a task description, a few task-specific exemplars, and a real image caption. To increase diversity we sample the exemplars and randomize their order. Given such input, we expect the LLM to output (1) an editing instruction, (2) an output caption for an ideal output image, and (3) which objects should be updated or added to the original image. We refer the readers to supplementary Figs. ??-?? for examples of our prompts. Further details on instruction generation are provided in the supplementary Sec. ??" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.271, + 0.296, + 0.287 + ], + "angle": 0, + "content": "3.2. Image Pairs Generation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.296, + 0.473, + 0.78 + ], + "angle": 0, + "content": "Our aim is to generate pairs of input and edited images that adhere to the edit instructions and preserve image elements that should remain intact. A crucial prerequisite when creating a pair of input and edited images is to guarantee that the two images differ only in specific elements or locations, while remaining identical in all other aspects. Previous instruct-based image editing methods [2] rely on Prompt-to-Prompt (P2P) to build an image-editing dataset. P2P injects cross-attention maps from the input image generation to the edited image generation. To support local edits, P2P additionally approximates a mask of the edited part, based on the cross-attention maps and constrains the edit to this local area. P2P relies on word-to-word alignment between the input image caption and the edited image caption (e.g. \"a cat riding a bicycle\" and \"a cat riding a car\") to produce editing image pairs. However, when there is no word-to-word alignment, the resulting mask tends to be imprecise due to its reliance on cross-attention maps. Furthermore, as word-to-word alignment is not a practical assumption in most of the image editing tasks, this approach often fails to preserve structure and identity. To address this challenge, we propose a mask extraction method, which is applied before the editing process. Our approach involves: (i) identifying the edited areas from the editing instruction via an LLM and creating corresponding masks before image generation, and (ii) integrating these masks during the editing process to ensure seamless fusion of edited regions with the original image. Distinct editing challenges, such as adding or removing objects, require tailored solutions. We utilize various techniques, including dilation and Gaussian blurring, to refine the masks. Further description of the method can be found at supplementary Sec. ??" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.781, + 0.472, + 0.903 + ], + "angle": 0, + "content": "Filtering. We employ a comprehensive filtering approach to ensure the fidelity of the dataset. This includes: (i) using the task predictor (Sec. 4.2) to reassign samples with instructions that should belong to another task, (ii) applying CLIP filtering metrics [2], (iii) employing structure preserving filtering based on the L1 distance between the depth map of the input image and the edited image, and (iv) applying image detectors to validate the presence (in Add task)," + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.152 + ], + "angle": 0, + "content": "the absence (in Remove task) or replacement (in Local task) of elements, according to the objects specified in the instruction. This process filters out \\(70\\%\\) of the data, resulting in a final dataset of ten million samples." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.166, + 0.593, + 0.182 + ], + "angle": 0, + "content": "4. Method" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.191, + 0.892, + 0.356 + ], + "angle": 0, + "content": "Emu Edit is a diffusion model designed to multi-task across a broad spectrum of editing tasks. As Emu Edit is trained on various tasks, a crucial aspect is the ability to identify the semantic edit (e.g., global/local/texture) that needs to be applied, based on the user instruction. To address this, we use task-specific embeddings integrated into the model, learned during training. After training, Emu Edit can quickly adapt to new tasks by learning a new task embedding through few-shot learning, without altering the rest of the model. We follow next with a detailed description of each part of our method." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.367, + 0.634, + 0.382 + ], + "angle": 0, + "content": "4.1. Architecture" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.39, + 0.892, + 0.692 + ], + "angle": 0, + "content": "Our model builds upon the foundation set by Emu, which is outlined in [6]. The Emu model is a two-stage approach that begins with a pre-training phase and concludes with a quality fine-tuning stage. Emu adapted the latent diffusion model architecture [18] to support high-resolution image generation and incorporated a 16-channel autoencoder with encoder \\( E \\) and decoder \\( D \\). In the following section, we adapt the notation of [2]. A large U-Net, \\( \\epsilon_{\\theta} \\), with 2.8 billion parameters, \\( \\theta \\), text embeddings from CLIP ViT-L [14] and T5-XXL [15], and a substantial pre-training dataset of 1.1 billion images facilitate the model's ability to learn complex semantics and finer details, with a noise-offset strategy contributing to high-contrast and aesthetically pleasing image generation. Given the encoded latent of an image \\( z = E(x) \\), the diffusion process generates a noisy latent \\( z_{t} \\) where the noise level increases over timesteps \\( t \\in T \\). To convert Emu to an instruction-based image editing model, we condition it on the image to be modified \\( c_{I} \\) and the instruction \\( c_{T} \\). Emu Edit is trained to minimize the following optimization problem," + }, + { + "type": "equation", + "bbox": [ + 0.565, + 0.695, + 0.892, + 0.717 + ], + "angle": 0, + "content": "\\[\n\\min _ {\\theta} \\mathbb {E} _ {y, \\epsilon , t} \\left[ \\| \\epsilon - \\epsilon_ {\\theta} \\left(z _ {t}, t, E \\left(c _ {I}\\right), c _ {T}\\right) \\| _ {2} ^ {2} \\right] \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.895, + 0.902 + ], + "angle": 0, + "content": "where \\(\\epsilon \\in N(0,1)\\) is the noise added by the diffusion process and \\(y = (c_{T},c_{I},x)\\) is a triplet of instruction, input image and target image from the dataset. In practice, we initialize the weights of Emu Edit with the weights of Emu. To support the image conditioning, we follow [2] and increase the number of input channels. New weights are initialized to zero. During inference, we perform classifier-free guidance on both image and text conditions. In our experiments we use a scale of \\(\\gamma_I = 1.5\\) for the image condition and \\(\\gamma_T = 5.0\\) for the text condition. Furthermore, we apply a rescaling of the diffusion scheduler to achieve a zero signal-to-noise ratio (SNR) at the terminal timestamp, as suggested" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8874" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.179, + 0.123, + 0.226, + 0.184 + ], + "angle": 0, + "content": "Make it play a rainbow colored trumpet" + }, + { + "type": "image", + "bbox": [ + 0.227, + 0.091, + 0.77, + 0.446 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.35, + 0.447, + 0.651, + 0.459 + ], + "angle": 0, + "content": "Local. Baseslines struggle to perform intricate local edits." + }, + { + "type": "image_caption", + "bbox": [ + 0.259, + 0.465, + 0.71, + 0.48 + ], + "angle": 0, + "content": "Figure 3. Failure cases of baseline instruction-based image editing models.2" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.489, + 0.471, + 0.534 + ], + "angle": 0, + "content": "in [10]. This is crucial in order to avoid any mismatch between the model's training and testing phases. For more implementation details see supplementary Sec. ??" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.544, + 0.32, + 0.56 + ], + "angle": 0, + "content": "4.2. Learned Task Embeddings" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.568, + 0.47, + 0.719 + ], + "angle": 0, + "content": "To guide the generation process toward the correct task, we learn an embedding vector for each task in the dataset. During training, given a sample from our dataset, we use the task index, \\(i\\), to fetch the task's embedding vector, \\(v_{i}\\), from an embedding table, and optimize it jointly with the model weights. We do so by introducing the task embedding \\(v_{i}\\) as an additional condition to the U-Net, \\(\\epsilon_{\\theta}\\). Concretely, we integrate the task embedding into the U-Net via crossattention interactions, and by adding it to the timestep embeddings. The optimization problem is updated to" + }, + { + "type": "equation", + "bbox": [ + 0.117, + 0.721, + 0.469, + 0.746 + ], + "angle": 0, + "content": "\\[\n\\min _ {\\theta , v _ {1}, \\dots , v _ {k}} \\mathbb {E} _ {\\hat {y}, \\epsilon , t} \\left[ \\| \\epsilon - \\epsilon_ {\\theta} \\left(z _ {t}, t, E \\left(c _ {I}\\right), c _ {T}, v _ {i}\\right) \\| _ {2} ^ {2} \\right] \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.748, + 0.47, + 0.854 + ], + "angle": 0, + "content": "where \\( k \\) is the total number of tasks in our dataset and \\( \\hat{y} = (c_I, c_T, x, i) \\) is a quadruplet of input image, input instruction text, target image, and task index from the dataset. Task-specific conditioning arises from the observation that models lacking such conditioning can become perplexed about the type of edit required, particularly when the instructions are complex or the edit type is ambiguous. For" + }, + { + "type": "page_footnote", + "bbox": [ + 0.076, + 0.863, + 0.469, + 0.9 + ], + "angle": 0, + "content": "2The samples depicted in this caption were selected by the authors. They do not cover all scenarios, but are meant to represent some common scenarios the authors encountered." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.489, + 0.892, + 0.61 + ], + "angle": 0, + "content": "instance, as visualized in Fig. 4, (1) a model without task conditioning might perform a global edit when a texture edit is required, (2) it might opt for segmentation when a global edit is necessary, and (3) it could implement a style edit in situations where a local edit would fit better. In the inference stage, we predict the task index. Specifically, we fine-tune a Flan-T5-XL model to identify the task at hand given the input instruction." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.622, + 0.65, + 0.637 + ], + "angle": 0, + "content": "4.3. Task Inversion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.646, + 0.892, + 0.766 + ], + "angle": 0, + "content": "To enable few-shot learning of new tasks without losing the general abilities of Emu Edit, we propose a method for adapting the model without changing the U-Net weights. Given a few examples of a new task, we learn a new task embedding, \\( v_{\\mathrm{new}} \\). We freeze the model weights, and adapt it to the task only through the task embedding. Thus, to fit a new task embedding we solve the following optimization problem:" + }, + { + "type": "equation", + "bbox": [ + 0.548, + 0.769, + 0.891, + 0.793 + ], + "angle": 0, + "content": "\\[\n\\min _ {v _ {\\text {n e w}}} \\mathbb {E} _ {y, \\epsilon , t} \\left[ \\| \\epsilon - \\epsilon_ {\\theta} \\left(z _ {t}, t, E \\left(c _ {I}\\right), c _ {T}, v _ {\\text {n e w}}\\right) \\| _ {2} ^ {2} \\right] \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.795, + 0.891, + 0.824 + ], + "angle": 0, + "content": "where \\( v_{\\mathrm{new}} \\) is the learned task embedding. Note that during task inversion \\( y \\) is a triplet belonging to the new task." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.826, + 0.892, + 0.9 + ], + "angle": 0, + "content": "The model can then be employed for the new task by conditioning it on the learned task embedding, and it can still handle its original tasks by relying on the initial task embeddings. In Sec. 5.3, we demonstrate that our model effectively generalizes to novel tasks using this method." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8875" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.073, + 0.089, + 0.468, + 0.376 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.388, + 0.472, + 0.458 + ], + "angle": 0, + "content": "Figure 4. Task embeddings. Model trained without task embeddings may get confused about the edit type when the instructions are complicated or there is ambiguity regarding the edit type: (1) Global edit (instead of Texture), (2) Segmentation (instead of Global), (3) Style edit (instead of Local)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.464, + 0.47, + 0.631 + ], + "angle": 0, + "content": "Sequential Edit Thresholding. We notice that applying the model repeatedly, in multi-turn editing scenarios, aggregates reconstruction and numerical errors, which translate to noticeable artifacts. To mitigate this, we add a per-pixel thresholding step after each edit-turn. At each step \\( s \\), we use the pixel value in the output image, \\( c_{I}^{s+1} \\), only if its alteration surpasses a specific threshold. Otherwise, we keep the pixel value from the input image, \\( c_{I}^{s} \\). Specifically, given an edit turn \\( s \\), we compute the absolute difference image \\( d = \\|c_{I}^{s+1} - c_{I}^{s}\\|_{1} \\) over the RGB channel, and apply the following thresholding:" + }, + { + "type": "equation", + "bbox": [ + 0.179, + 0.644, + 0.469, + 0.684 + ], + "angle": 0, + "content": "\\[\nc _ {I} ^ {s + 1} = \\left\\{ \\begin{array}{l l} c _ {I} ^ {s} & \\text {i f} \\bar {d} < \\alpha , \\\\ c _ {I} ^ {s + 1} & \\text {o t h e r w i s e .} \\end{array} \\right. \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.692, + 0.47, + 0.77 + ], + "angle": 0, + "content": "where, \\( \\bar{d} \\) is obtained after passing \\( d \\) through a low pass filter, in order to smooth the transition between previous and current pixels. In practice, we choose \\( \\alpha = 0.03 \\). Please refer to supplementary Sec. ?? for a qualitative comparison. In Fig. 2 we show examples of multi-turn editing." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.784, + 0.21, + 0.801 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.81, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Our experiments evaluate the ability of Emu Edit to follow user instructions faithfully and preserve the visual fidelity of the original image. First, we evaluate the performance of our approach on instruction-based image editing tasks. Second, we conduct a comprehensive ablation study to assess the effectiveness of our different contributions. Specifically," + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.893, + 0.198 + ], + "angle": 0, + "content": "we ablate the contribution of the computer vision tasks to the model performance on image editing tasks, the importance of learned task embeddings, and the effect of multitask learning on instruction-based image editing. Further ablation on our data generation pipeline can be found in the supplementary Sec. ?? Finally, we demonstrate our model's ability to learn new tasks via few-shot learning." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.218, + 0.893, + 0.641 + ], + "angle": 0, + "content": "Measures. We employ two main measures in our evaluation: edit text alignment and image faithfulness. Specifically, for each pair of input image and editing instruction, we use the following automatic metrics: (i) CLIP [14] text-image direction similarity \\((\\mathrm{CLIP}_{dir})\\) - measuring agreement between change in captions and the change in images, (ii) CLIP image similarity \\((\\mathrm{CLIP}_{img})\\) - measuring change between edited and input image, (iii) CLIP output similarity \\((\\mathrm{CLIP}_{out})\\) - measuring edited image similarity with output caption, (iv) L1 pixel-distance between input and edit image, and (v) DINO [4] similarity between the DINO embeddings of input and edited images. With the exception of the L1 distance, where lower values indicate better performance, higher values in all other measures signify better results. A low L1 distance translates to small changes in image's pixel values. A high DINO and \\(\\mathrm{CLIP}_{img}\\) similarity score, suggests semantic similarity between the images. For region-based edits, high image similarity scores indicate the edits were precise. For free-form edits, high similarity scores indicate image structure preservation. \\(\\mathrm{CLIP}_{dir}\\) and \\(\\mathrm{CLIP}_{out}\\) measure how well the model followed the instruction. In addition, we asked human raters to evaluate the text alignment and image faithfulness. In each evaluation scenario, raters are presented with two modified images alongside the original input image and instruction, and are presented with two questions: (i) Image Faithfulness: which image better preserves elements in the input image, and (ii) Text Alignment: which image best follows the instruction." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.651, + 0.622, + 0.665 + ], + "angle": 0, + "content": "5.1. Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.674, + 0.892, + 0.84 + ], + "angle": 0, + "content": "Throughout the paper, we report results on the MagicBrush test set [25] and the Emu Edit benchmark. In the following section, we describe our motivation for creating this benchmark, and detail its curation process. To date, there are two main benchmarks for evaluating instruction-based image editing capabilities. First, the InstructPix2Pix benchmark [2], which is intrinsically biased due to its reliance on generated Stable Diffusion [17] input images, and GPT-3 [3] generated instructions. Consequently, it is unclear whether its results will truly mirror the performance on real input images, with genuine user instructions." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.892, + 0.9 + ], + "angle": 0, + "content": "Unlike InstructPix2Pix, the second benchmark, MagicBrush [25], uses a diverse set of authentic input images from the MS-COCO benchmark [5, 11], and annotator-defined instructions. Nonetheless, this dataset also suffers" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8876" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.089, + 0.894, + 0.119 + ], + "angle": 0, + "content": "Table 2. Comparison with image-editing baselines evaluated on Emu Edit test set and MagicBrush test set. For each benchmark we report CLIP, L1, DINO metrics and human ratings. Human evaluation shows the percentage of raters that prefer the results of Emu Edit." + }, + { + "type": "table", + "bbox": [ + 0.085, + 0.128, + 0.885, + 0.253 + ], + "angle": 0, + "content": "
Emu Edit Test SetMagicBrush Test Set
MethodCLIPdir↑CLIPim↑CLIPout↑L1↓DINO↑Text align. faith.CLIPdir↑CLIPim↑CLIPout↑L1↓DINO↑Text align. faith.
InstructPix2Pix [2]0.0780.8340.2190.1210.76277.3376.710.1150.8370.2450.0930.767
MagicBrush [25]0.0900.8380.2220.1000.77674.5074.100.1230.8830.2610.0580.871
PnP [21]0.0280.5210.0890.3040.15398.9599.000.0250.5680.1010.2890.220
Null-Text Inv. [12]0.1010.7610.2360.0750.67881.6385.470.1210.7520.2630.0770.664
Our0.1090.8590.2310.0940.819--0.1350.8970.2610.0520.879
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.259, + 0.471, + 0.35 + ], + "angle": 0, + "content": "from inherent bias. During data collection, annotators were directed to use the DALLE-2 image editing platform [13] to generate the edited images. Thus, this benchmark is biased towards editing instructions that the DALLE-2 editor can successfully follow, which may compromise both its diversity and complexity." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.352, + 0.471, + 0.669 + ], + "angle": 0, + "content": "Emu Edit Benchmark. To collect a dataset with reduced bias and of higher diversity, we take a different approach. We first define seven different categories of potential image editing operations: background alteration (Background), comprehensive image changes (Global), style alteration (Style), object removal (Remove), object addition (Add), localized modifications (Local), and color/texture alterations (Texture). Then, we utilize the diverse set of input images from the MagicBrush benchmark [25], and for each editing operation, we task crowd workers to devise relevant, creative, and challenging instructions. Moreover, to increase the quality of the collected examples, we apply a post-verification stage, in which crowd workers filter examples with irrelevant instructions. Finally, to support evaluation for methods that require input and output captions [8, 21], we additionally collect an input caption and output caption. When doing so, we ask annotators to ensure that the captions capture both important elements in the image, and elements that should change based on the instruction. See supplementary Sec. ?? for examples of our benchmark, which we release to support better evaluation." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.674, + 0.47, + 0.902 + ], + "angle": 0, + "content": "Baseline Comparisons. We compare our model against two instruction-based image editing baseline models: InstructPix2Pix [2], and MagicBrush [25], which is a variant of InstructPix2Pix that was fine-tuned on the MagicBrush dataset. Additionally, we compare our model against two text-based image editing methods: PNP [21] and Null-Text Inversion modification of P2P [8, 12]. Unlike instruction-based models, these works expect image descriptions. Therefore, we provide them with access to the input caption and output caption. Note however, providing these methods with access to the ground-truth captions could potentially offer an advantage over instruction-based models, since the automatic metrics also rely on these captions. Tab. 2 shows our results versus the baselines. The findings indicate that human raters consistently prefer Emu" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.259, + 0.892, + 0.381 + ], + "angle": 0, + "content": "Edit over all baselines by a large margin. Furthermore, apart from Null-Text Inversion, which as explained above, utilizes the ground-truth captions during inference, our approach outperforms the existing baselines on the automatic metrics. We provide qualitative comparisons in Fig. 3. Additional comparisons are available in the supplementary: Fig.??-??, and Fig.??-??. For performance on vision tasks, see Sec. ??" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.389, + 0.611, + 0.404 + ], + "angle": 0, + "content": "5.2. Ablations" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.412, + 0.892, + 0.639 + ], + "angle": 0, + "content": "Computer Vision Tasks Enhance Image Editing Tasks. Here we demonstrate the importance of the vision tasks to Emu Edit performance on image editing tasks. For this, we trained two additional models on all tasks except: (i) detect and segment tasks, and (ii) image-to-image translation tasks. As we show in Tab. 4, adding the detection and segmentation tasks improves the model performance in region-based editing tasks. Additionally, we observe that image-to-image translation tasks improve the performance in free-form editing tasks. We hypothesize that the recognition tasks improve the model's recognition capabilities, leading to more accurate and precise localized modifications. Similarly, image-to-image tasks assist the model in understanding the entire image structure, thereby enhancing its capabilities for global operations." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.64, + 0.892, + 0.879 + ], + "angle": 0, + "content": "Contribution of Learned Task Embeddings. We compare three variants of Emu Edit: (i) conditioned on the ground-truth task embedding, (ii) conditioned on the task embedding, as predicted by the task predictor described in Sec. 4.2, and (iii) without conditioning on the task type.3 Tab. 3 shows the results on the validation set of our benchmark. As can be seen, conditioning on the task type boosts the model's performance. Furthermore, our task predictor closes the gap with the ground-truth conditioned model. Qualitatively, we observe that without conditioning on the task type the model may perform the wrong editing operation (Fig. 4). In supplementary Fig. ??, we demonstrate the effect of manipulating the task while keeping the instruction and input image fixed. As can be seen, changing the task embedding directly influences the task executed by the model." + }, + { + "type": "page_footnote", + "bbox": [ + 0.516, + 0.887, + 0.789, + 0.9 + ], + "angle": 0, + "content": "\\(^{3}\\) (iii) was trained without learned task embeddings." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8877" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.09, + 0.47, + 0.132 + ], + "angle": 0, + "content": "Table 3. Learned task embeddings ablation on our validation set. We compare variations of Emu Edit: without task type condition, with predicted task type, and with ground-truth task type." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.137, + 0.454, + 0.204 + ], + "angle": 0, + "content": "
MethodCLIPdir↑CLIPim↑CLIPout↑L1↓DINO↑
w/o task emb.0.1040.8430.2270.1090.792
with pred. task0.1170.8500.2310.1030.809
with gt task0.1190.8520.2310.1000.811
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.21, + 0.468, + 0.238 + ], + "angle": 0, + "content": "Table 4. Contribution of computer vision tasks. Human evaluation is shown as a percentage of majority votes in favor of our model." + }, + { + "type": "table", + "bbox": [ + 0.118, + 0.242, + 0.434, + 0.328 + ], + "angle": 0, + "content": "
Region-BasedFree-form
MethodText align.Image faith.Text align.Image faith.
without detect/segment60.060.252.351.5
without im2im translation50.249.058.060.1
" + }, + { + "type": "image_caption", + "bbox": [ + 0.116, + 0.34, + 0.177, + 0.416 + ], + "angle": 0, + "content": "(i) Incorporate a bee into the bag's pattern and detect it" + }, + { + "type": "image", + "bbox": [ + 0.176, + 0.34, + 0.293, + 0.433 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.293, + 0.34, + 0.41, + 0.433 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.123, + 0.466, + 0.171, + 0.513 + ], + "angle": 0, + "content": "(ii) Mark the gift bags" + }, + { + "type": "image", + "bbox": [ + 0.174, + 0.446, + 0.292, + 0.538 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.293, + 0.446, + 0.41, + 0.538 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.539, + 0.252, + 0.551 + ], + "angle": 0, + "content": "Input" + }, + { + "type": "image_caption", + "bbox": [ + 0.318, + 0.538, + 0.383, + 0.551 + ], + "angle": 0, + "content": "Emu Edit" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.557, + 0.469, + 0.586 + ], + "angle": 0, + "content": "Figure 5. Generations on unseen tasks with task inversion. (i) composition of add and detect tasks, (ii) object contour detection." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.595, + 0.469, + 0.685 + ], + "angle": 0, + "content": "Multi-Task versus Expert Models. We show that training a single model on a diverse range of tasks, leads to enhanced performance in each individual task, outperforming models that are specifically trained for a single task. To validate this, we train an expert model for each task, and compare its performance to the single one (See supplementary Tab. ??)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.685, + 0.469, + 0.806 + ], + "angle": 0, + "content": "Influence of Number of Tasks. Here, we ablate the number of tasks participating in the multi-task training scheme. In supplementary Fig. ?? we report the average \\(\\mathrm{CLIP}_{dir}\\) on the Style and Texture tasks when iteratively excluding other tasks, and training a model on the new tasks list. As can be seen, augmenting the model with additional tasks leads to improved performance, even in tasks which are not directly associated with the added ones." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.817, + 0.367, + 0.833 + ], + "angle": 0, + "content": "5.3. Few-Shot Learning of New Tasks" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.469, + 0.902 + ], + "angle": 0, + "content": "Finally, we evaluate our model's ability to generalize by testing it in a few-shot scenario with previously unseen tasks. We test its performance across the following tasks: (i) super-resolution (x4), (ii) object contour detection, (iii)" + }, + { + "type": "image", + "bbox": [ + 0.489, + 0.108, + 0.676, + 0.213 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.545, + 0.221, + 0.62, + 0.234 + ], + "angle": 0, + "content": "(a) Inpainting" + }, + { + "type": "image", + "bbox": [ + 0.686, + 0.108, + 0.871, + 0.213 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.733, + 0.221, + 0.825, + 0.233 + ], + "angle": 0, + "content": "(b) Add + Detect" + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.245, + 0.892, + 0.315 + ], + "angle": 0, + "content": "Figure 6. Few-shot performance for different tasks over 1, 10, and 100 samples. Each line represents a different training setting: 'Scratch' finetune (Blue, \\(\\bigcirc\\)), Emu Edit finetune (Orange, \\(\\square\\)), task inversion (Green, \\(\\triangle\\)), all compared to an upper-bound expert trained on 100k samples (Red dashed line, \\(\\diamond\\))." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.325, + 0.892, + 0.688 + ], + "angle": 0, + "content": "mask-based inpainting, and (iv) a composite task formed by combining two tasks from our dataset: add and detect. For each task, we assess the model's performance when trained with 0, 1, and 100 examples, with the 0-example case being equivalent to a zero-shot setting. We compare three baselines: (i) Scratch - Emu Edit initialized with Emu's weights, trained on the examples, (ii) Task Inversion - Emu Edit with task inversion (Sec. 4.3), and (iii) Finetune - Emu Edit where we finetune all of the model's weights on the examples. As an upper-bound expert, we train the first baseline on 100,000 examples. Note that, the \"Task Inversion\" and \"Finetune\" baselines were trained on the multitask dataset whereas the \"Scratch\" baseline was not. As can be seen in Fig. 6, fine-tuning with a single example is enough to significantly enhance the performance, while training from scratch results in overfitting. Moreover, utilizing 100 samples nearly achieves expert-level performance, implying that the model can effectively generalize to novel tasks. We also observe that task inversion performs similarly to complete fine-tuning, indicating that the model already contains essential information and can be prompted with a new task embedding to yield the desired result. For performance and generation examples on additional tasks see Fig. 5 and supplementary ????." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.694, + 0.619, + 0.709 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Emu Edit presents a step change in instructable image editing capabilities, primarily due to its unique training on both recognition and generation tasks. This dual-focus approach significantly enhances the model's comprehension of natural language instructions, enabling it to accurately execute a wide array of editing operations. Its ability to generalize to new tasks with minimal examples further demonstrates its versatility and robustness. Furthermore, our framework has the potential for further integration with a multimodal LLM in future projects, particularly for complex editing tasks requiring detailed reasoning from the input image, such as counting objects or executing intricate tasks." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8878" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.175, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.171 + ], + "angle": 0, + "content": "[1] Omri Avrahami, Dani Lischinski, and Ohad Fried. Blended diffusion for text-driven editing of natural images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18208-18218, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.173, + 0.47, + 0.2 + ], + "angle": 0, + "content": "[2] Tim Brooks et al. Instructpix2pix: Learning to follow image editing instructions. In CVPR, 2023. 1, 3, 4, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.203, + 0.471, + 0.342 + ], + "angle": 0, + "content": "[3] Tom B. Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-Voss, Gretchen Krueger, T. J. Henighan, Rewon Child, Aditya Ramesh, Daniel M. Ziegler, Jeff Wu, Clemens Winter, Christopher Hesse, Mark Chen, Eric Sigler, Mateusz Litwin, Scott Gray, Benjamin Chess, Jack Clark, Christopher Berner, Sam McCandlish, Alec Radford, Ilya Sutskever, and Dario Amodei. Language models are few-shot learners. ArXiv, abs/2005.14165, 2020. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.343, + 0.47, + 0.413 + ], + "angle": 0, + "content": "[4] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9650-9660, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.415, + 0.47, + 0.469 + ], + "angle": 0, + "content": "[5] Xinlei Chen, Hao Fang, Tsung-Yi Lin, Ramakrishna Vedantam, Saurabh Gupta, Piotr Dólar, and C. Lawrence Zittnick. Microsoft coco captions: Data collection and evaluation server. ArXiv, abs/1504.00325, 2015. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.472, + 0.47, + 0.543 + ], + "angle": 0, + "content": "[6] Xiaoliang Dai, Ji Hou, Chih-Yao Ma, Sam Tsai, Jialiang Wang, Rui Wang, Peizhao Zhang, Simon Vandenhende, Xiaofang Wang, Abhimanyu Dubey, et al. Emu: Enhancing image generation models using photogenic needles in a haystack. arXiv preprint arXiv:2309.15807, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.544, + 0.47, + 0.613 + ], + "angle": 0, + "content": "[7] Oran Gafni, Adam Polyak, Oron Ashual, Shelly Sheynin, Devi Parikh, and Yaniv Taigman. Make-a-scene: Scene-based text-to-image generation with human priors. In European Conference on Computer Vision, pages 89-106. Springer, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.615, + 0.47, + 0.671 + ], + "angle": 0, + "content": "[8] Amir Hertz, Ron Mokady, Jay M. Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-prompt image editing with cross attention control. ArXiv, abs/2208.01626, 2022. 2, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.673, + 0.47, + 0.728 + ], + "angle": 0, + "content": "[9] Bahjat Kawar et al. Imagic: Text-based real image editing with diffusion models. 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6007-6017, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.73, + 0.469, + 0.772 + ], + "angle": 0, + "content": "[10] Shanchuan Lin, Bingchen Liu, Jiashi Li, and Xiao Yang. Common diffusion noise schedules and sample steps are flawed. arXiv preprint arXiv:2305.08891, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.774, + 0.469, + 0.829 + ], + "angle": 0, + "content": "[11] Tsung-Yi Lin, Michael Maire, Serge J. Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C. Lawrence Zitnick. Microsoft coco: Common objects in context. In ECCV, 2014. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.831, + 0.469, + 0.9 + ], + "angle": 0, + "content": "[12] Ron Mokady, Amir Hertz, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Null-text inversion for editing real images using guided diffusion models. 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6038–6047, 2022. 3, 7" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.471, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.119 + ], + "angle": 0, + "content": "[13] OpenAI. Dall-e 2, 2022. https://openai.com/product/dall-e-2.3,7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.121, + 0.892, + 0.203 + ], + "angle": 0, + "content": "[14] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 1, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.204, + 0.892, + 0.273 + ], + "angle": 0, + "content": "[15] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J. Liu. Exploring the limits of transfer learning with a unified text-to-text transformer. Journal of Machine Learning Research, 21(140):1-67, 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.274, + 0.892, + 0.327 + ], + "angle": 0, + "content": "[16] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.329, + 0.892, + 0.398 + ], + "angle": 0, + "content": "[17] Robin Rombach, A. Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10674-10685, 2021. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.4, + 0.892, + 0.44 + ], + "angle": 0, + "content": "[18] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.442, + 0.892, + 0.537 + ], + "angle": 0, + "content": "[19] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L. Denton, Seyed Kamyar Seyed Ghasemipour, Burcu Karagol Ayan, Seyedeh Sara Mahdavi, Raphael Gontijo Lopes, Tim Salimans, Jonathan Ho, David J. Fleet, and Mohammad Norouzi. Photorealistic text-to-image diffusion models with deep language understanding. ArXiv, abs/2205.11487, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.539, + 0.892, + 0.607 + ], + "angle": 0, + "content": "[20] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.609, + 0.892, + 0.676 + ], + "angle": 0, + "content": "[21] Narek Tumanyan, Michal Geyer, Shai Bagon, and Tali Dekel. Plug-and-play diffusion features for text-driven image-to-image translation. 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1921-1930, 2022. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.678, + 0.892, + 0.734 + ], + "angle": 0, + "content": "[22] Bram Wallace, Akash Gokul, and Nikhil Vijay Naik. Edict: Exact diffusion inversion via coupled transformations. 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 22532-22541, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.735, + 0.892, + 0.789 + ], + "angle": 0, + "content": "[23] Su Wang et al. Imagen editor and editbench: Advancing and evaluating text-guided image inpainting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18359-18369, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.79, + 0.892, + 0.857 + ], + "angle": 0, + "content": "[24] Shaoan Xie, Zhifei Zhang, Zhe Lin, Tobias Hinz, and Kun Zhang. Smartbrush: Text and shape guided object inpainting with diffusion model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22428-22437, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.859, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[25] Kai Zhang et al. Magicbrush: A manually annotated dataset for instruction-guided image editing. arXiv:2306.10012, 2023. 1, 3, 6, 7" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "8879" + } + ] +] \ No newline at end of file diff --git a/2024/Emu Edit_ Precise Image Editing via Recognition and Generation Tasks/b6371d78-25ca-4b11-84b1-a566ed9d62f2_origin.pdf b/2024/Emu Edit_ Precise Image Editing via Recognition and Generation Tasks/b6371d78-25ca-4b11-84b1-a566ed9d62f2_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..e786de99ea0d01b10790ca1646c314ba95659acd --- /dev/null +++ b/2024/Emu Edit_ Precise Image Editing via Recognition and Generation Tasks/b6371d78-25ca-4b11-84b1-a566ed9d62f2_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:954ca34b935c3b7e1a1654af429f22f5d9eaa0a0a70c95db48147f498289630e +size 8472001 diff --git a/2024/Emu Edit_ Precise Image Editing via Recognition and Generation Tasks/full.md b/2024/Emu Edit_ Precise Image Editing via Recognition and Generation Tasks/full.md new file mode 100644 index 0000000000000000000000000000000000000000..7f56823404e0e90196a06d7c50096a227c221065 --- /dev/null +++ b/2024/Emu Edit_ Precise Image Editing via Recognition and Generation Tasks/full.md @@ -0,0 +1,272 @@ +# Emu Edit: Precise Image Editing via Recognition and Generation Tasks + +Shelly Sheynin*, Adam Polyak*, Uriel Singer*, Yuval Kirstain*, Amit Zohar*, Oron Ashual, Devi Parikh and Yaniv Taigman + +GenAI,Meta + +![](images/1daaa348ca8afee9b9522231b19578495255b61c69bb7fa381085ebff184eec9.jpg) +Dress the emu with a fireman outfit + +![](images/121a82bdf17aea58535197abd2870754df60ef1528d39a5ce16c55d569d0f221.jpg) +Let's see it graduating +Figure 1. Emu Edit is a multi-tasking model that combines various editing (left, middle) and vision (right) tasks for precise image editing. + +![](images/0967a632fde260268fe3b3f4fc15dccddf079cf81e5f8fac9b18c6cde487e995.jpg) +Mark the drinks + +# Abstract + +Instruction-based image editing holds immense potential for a variety of applications, as it enables users to perform any editing operation using a natural language instruction. However, current models in this domain often struggle with accurately executing user instructions. We present Emu Edit, a multi-task image editing model which sets state-of-the-art results in instruction-based image editing. To develop Emu Edit we train it to multi-task across an unprecedented range of tasks, such as region-based editing, free-form editing, and Computer Vision tasks, all of which are formulated as generative tasks. Additionally, to enhance Emu Edit's multi-task learning abilities, we provide it with learned task embeddings which guide the generation process towards the correct edit type. Both these elements are essential for Emu Edit's outstanding performance. Furthermore, we show that Emu Edit can generalize to new tasks, such as image inpainting, super-resolution, and compositions of editing tasks, with just a few labeled examples. This capability offers a significant advantage in scenarios where high-quality samples are scarce. Lastly, to facilitate a more rigorous and informed assessment of instructable image editing models, we release a new challenging and versatile benchmark that includes seven different image editing tasks. + +# 1. Introduction + +Image editing is a widely-used application that millions engage with every day. Popular image editing tools, however, either demand considerable expertise and are time-consuming to use, or are quite limited, providing only a predefined set of editing operations, such as specific filters. Instruction-based image editing [2, 25] attempts to resolve these limitations by allowing users to effortlessly describe their editing goals using natural language instructions. For instance, a user can provide a model with an image and instruct it to "Dress the emu with a fireman outfit" or "Let's see it graduating" (see Fig. 1). Nevertheless, while instruction-based image editing models like InstructPix2Pix [2] are designed to process any given instruction, they often struggle to accurately interpret and execute such instructions. Moreover, their generalization is limited, often falling short on tasks that deviate slightly from those they were trained on (see Fig. 3). To address these gaps, we introduce Emu Edit, the first image editing model trained on an extensive and diverse set of tasks, including both image editing and computer vision tasks. Emu Edit provides a substantial improvement in both compliance with the edit instruction and preservation of the visual fidelity of the original image. As we show through both automatic metrics [14] and human judgments on two benchmarks [25], Emu Edit achieves state-of-the-art results in instruction-based image editing. The success of Emu Edit stems from two key contributions. First, we train our model to multi-task across sixteen distinct image editing tasks. These tasks span region-based editing tasks, free-form editing tasks and computer + +![](images/f6daee2dd1ad5774950c595167a32770ca7e13f1189f15de497229100214adba.jpg) +(a) A cat + +![](images/e20d245dddf075ebb2f9a85e3577ad0e4ad463efa803748455a964ab0c108ffd.jpg) +(b) Remove the tail + +![](images/53da06711e6d3d8423eacac630c92eea95a427e99482ac28db72a69654f67fc4.jpg) +(c) Add a pink jacket + +![](images/3625b84b9371fb02d492e67157add04fc9d29f97c7717af61ecbcdf2f791e658.jpg) +(d)Make it rainy + +![](images/506af08233d4e36e2f8bbce59603fe18b2c59cfb79e7660fc10d64ab3bfbdf70.jpg) +(e) Have the cat look shocked + +![](images/2f4c3ffb2eee89852984d434d8896b910ed6d6dd66675ea26147c60a85839bc1.jpg) +(f) Extract the depth map + +![](images/4641d1eab2feee72b020ae720e2337bfaf683f0a908d5c765094e57532838c29.jpg) +(g) Generate a rainy day image of a hedgehog in a dress using the depth map + +![](images/c9f3cb108decacef2bbdc2436d64171e786f0940660fbc7b29b03667bd400d6b.jpg) +(h) Replace the dress with an astronaut outfit + +![](images/d8b340b4db0cf630a244006cf7460b8d8e3f058dd3c6e89b6c3f84e7ed43e8a3.jpg) +(i) Segment the spacesuit, and detect the hands +Figure 2. Instruction-based image generation. Each image originates from a previous one with a corresponding instruction. The prior image for (a) was set to zero. + +![](images/421fc7fc2f04003c5f2645ecebf7d8aa0491255a119677dfbe47c1f0ac10e446.jpg) +(j) Add the text "purple cat" using a purple font + +vision tasks, all formulated as generative tasks. Unlike previous work, we develop a distinct data generation pipeline for each task, allowing us to gather a training set that is not only more diverse but also more precise in its examples. We find that training a single model on all tasks yields better results than training expert models on each task independently. Additionally, we show that as the number of training tasks increases, so does the performance of Emu Edit. Furthermore, we discover that surprisingly, computer vision tasks such as detection, segmentation, and others, significantly enhance editing performance, as validated both by human raters as well as quantitative measures. Second, to process this wide array of tasks effectively, we introduce the concept of learned task embeddings, which are used to steer the generation process toward the correct generative task. Concretely, for each task, we learn a unique task embedding vector, and integrate it into the model through cross-attention interactions, and by adding it to the timestep embeddings. We demonstrate that learned task embeddings significantly enhance the ability of our model to accurately infer the appropriate edit type from the user instruction and execute the correct edit. Equipped with a robust model trained across a broad spectrum of tasks and guided by learned task embeddings, we explore few-shot adaptation to unseen tasks via task inversion. In this process, we maintain the model weights untouched, and solely update a task embedding to fit the new task. Our experiments demonstrate that Emu Edit can swiftly adapt to new tasks, such as superresolution, contour detection, and others. Notably, for some tasks, fine-tuning the model on just a handful of examples + +yields results that nearly match those of an expert model trained on one hundred thousand examples. This makes task inversion with Emu Edit particularly advantageous in scenarios where labeled examples are limited, or when the compute budget is low. Finally, to support future research for instruction-based image editing, we publicly release a diverse and challenging benchmark that includes seven different image editing operations, as well as our model's generations on this dataset. In summary, this work addresses the limitations of instruction-based image editing models in accurately following user instructions. We demonstrate that by employing multi-task training across a diverse array of tasks, including recognition, generation, and editing, we can enhance our model's performance. Furthermore, by incorporating learned task embeddings into our model's architecture, we not only improve its results but also enable efficient few-shot learning for new tasks. With these improvements, our model sets a new standard in the field, offering significantly more precise and robust instruction-based image editing capabilities than existing models. + +# 2. Related Work + +The emergence of high-performing text-to-image diffusion models [7, 16, 17, 19] facilitated the development of effective text-based image editing methods. Such methods usually employ aligned and detailed descriptions of the input and edited image to perform a specific edit. Prompt-to-Prompt (P2P) [8] injects the input caption attention maps to the target caption attentions maps. Null-Text Inver + +sion [12] inverts an input image using the null-text embedding to support editing of a real image. Plug-and-Play (PNP) [21] injects spatial features in addition to attention maps and obtains better performance at global editing. Imagic [9] finetunes the diffusion model to support complex textual instructions. EDICT [22] suggests an image inversion based on two noise vectors enabling better image reconstruction and textual faithfulness. Another class of image editing models, employs an input mask as additional input [1, 23, 24]. Blended Diffusion [1] modifies the diffusion step by blending the input image in the unmasked regions. Imagen Editor [23] and SmartBrush [24] finetune the text-to-image model to be conditioned on both the input image and mask. While the text-based image editing methods detailed above enable humans to edit images, they frequently exhibit inconsistent performance and require multiple inputs, such as aligned and detailed descriptions of both the input and target images, or at times, input masks. + +To offer a more intuitive and user-friendly interface, and significantly enhance ease of use for humans, InstructPix2Pix [2] introduced an instructable image editing model. They developed this model by utilizing both GPT-3 [3] and Prompt-to-Prompt [8], to generate a large synthetic dataset for instruction-based image editing, and employed the dataset to train an instructable image editing model. Unlike InstructPix2Pix which used a synthetic dataset, MagicBrush [25] developed a manually-annotated instruction-guided image editing dataset by requesting humans to use an online image editing tool [13]. Finetuning InstructPix2Pix on this dataset led to improved image editing capabilities. However, even though there has been progress and improvement in instruction-based image editing models, we show in Sec. 5 that state-of-the-art image editing models still struggle with accurately interpreting and precisely executing editing instructions. + +In this paper, we drastically narrow such performance gaps by leveraging multi-task training and a matching architecture. Unlike prior work that solely focuses on image editing [2, 25], we train our model to perform various tasks and learn a very diverse set of capabilities. The quality and versatility of our training procedure and dataset, together with our improved architecture for multi-task learning, enables us to make a big leap in performance and differentiates us from prior work in the field. Fig. 3 include several challenging editing samples as examples. + +# 3. Multi-Task Dataset for Image Editing + +Training a robust and accurate image editing model requires a highly diverse dataset of input images, editing instructions, and output edited images. However, manually collecting such examples is impractically time-consuming, existing sources on the web (e.g. communities and forums on social media) are limited in size, and publicly available syn- + +# 1. Region-Based Editing + +- Local: Substituting one object for another, altering an object's attributes (e.g., "make it smile"). +- Remove: Erasing an object from the image. +- Add: Inserting a new object into the image. +- Texture: Altering an object's visual characteristics without affecting its structure (e.g., painting over, filling or covering an object). +- Background: Changing the scene's background. + +# 2. Free-Form Editing + +- Global: Edit instructions that affect the entire image, or that can not be described using a mask (e.g., "let's see it in the summer"). +- Style: Change the style of an image. +- Text Editing: This involves text-related editing tasks such as adding, removing, swapping text, and altering the text's font and color. + +# 3. Vision tasks + +- Detect: Identifying and marking a specific object within the image with a rectangle bounding box. +- Segment: Isolating and marking an object in the image. +Color: Color adjustments like sharpening and blurring. +- Image-to-Image Translation: Tasks that involve bidirectional image type conversion, such as sketch-to-image, depth map-to-image, normal map-to-image, pose-to-image, segmentation map-to-image, and so on. + +Table 1. Description of the tasks forming the Emu Edit dataset. + +thetic datasets often lack in diversity or quality. Therefore, we construct a new dataset that encompasses sixteen distinct tasks and ten million examples. Each example $(c_{I}, c_{T}, x, i)$ in our dataset, contains an input image $c_{I}$ , a text instruction $c_{T}$ , a target image $x$ , and a task index $i$ (out of the sixteen). The following section outlines the process of the data construction. In Sec. 3.1 we describe the instruction generation process, and in Sec. 3.2 the image pairs $(c_{I}, x)$ generation and filtering. + +Task Categories. The dataset is composed of tasks which are divided into three main categories: region-based editing, free-form editing, and vision tasks. Tab. 1 includes the full list of tasks, and their distribution in the train set is visualized in Fig. ?? + +# 3.1. Instruction Generation + +To generate editing instructions, we leverage the dialogue-optimized 70 billion parameter Llama 2 variant [20]. We observed that using a single agent to generate the instructions for all tasks leads to a lack of diversity in the dataset. Notably, the LLM exhibits a bias towards particular tasks and instruction phrasings. To address this, we utilize in + +context learning to create a task-specific agent for each task. Concretely, we provide the LLM with a task description, a few task-specific exemplars, and a real image caption. To increase diversity we sample the exemplars and randomize their order. Given such input, we expect the LLM to output (1) an editing instruction, (2) an output caption for an ideal output image, and (3) which objects should be updated or added to the original image. We refer the readers to supplementary Figs. ??-?? for examples of our prompts. Further details on instruction generation are provided in the supplementary Sec. ?? + +# 3.2. Image Pairs Generation + +Our aim is to generate pairs of input and edited images that adhere to the edit instructions and preserve image elements that should remain intact. A crucial prerequisite when creating a pair of input and edited images is to guarantee that the two images differ only in specific elements or locations, while remaining identical in all other aspects. Previous instruct-based image editing methods [2] rely on Prompt-to-Prompt (P2P) to build an image-editing dataset. P2P injects cross-attention maps from the input image generation to the edited image generation. To support local edits, P2P additionally approximates a mask of the edited part, based on the cross-attention maps and constrains the edit to this local area. P2P relies on word-to-word alignment between the input image caption and the edited image caption (e.g. "a cat riding a bicycle" and "a cat riding a car") to produce editing image pairs. However, when there is no word-to-word alignment, the resulting mask tends to be imprecise due to its reliance on cross-attention maps. Furthermore, as word-to-word alignment is not a practical assumption in most of the image editing tasks, this approach often fails to preserve structure and identity. To address this challenge, we propose a mask extraction method, which is applied before the editing process. Our approach involves: (i) identifying the edited areas from the editing instruction via an LLM and creating corresponding masks before image generation, and (ii) integrating these masks during the editing process to ensure seamless fusion of edited regions with the original image. Distinct editing challenges, such as adding or removing objects, require tailored solutions. We utilize various techniques, including dilation and Gaussian blurring, to refine the masks. Further description of the method can be found at supplementary Sec. ?? + +Filtering. We employ a comprehensive filtering approach to ensure the fidelity of the dataset. This includes: (i) using the task predictor (Sec. 4.2) to reassign samples with instructions that should belong to another task, (ii) applying CLIP filtering metrics [2], (iii) employing structure preserving filtering based on the L1 distance between the depth map of the input image and the edited image, and (iv) applying image detectors to validate the presence (in Add task), + +the absence (in Remove task) or replacement (in Local task) of elements, according to the objects specified in the instruction. This process filters out $70\%$ of the data, resulting in a final dataset of ten million samples. + +# 4. Method + +Emu Edit is a diffusion model designed to multi-task across a broad spectrum of editing tasks. As Emu Edit is trained on various tasks, a crucial aspect is the ability to identify the semantic edit (e.g., global/local/texture) that needs to be applied, based on the user instruction. To address this, we use task-specific embeddings integrated into the model, learned during training. After training, Emu Edit can quickly adapt to new tasks by learning a new task embedding through few-shot learning, without altering the rest of the model. We follow next with a detailed description of each part of our method. + +# 4.1. Architecture + +Our model builds upon the foundation set by Emu, which is outlined in [6]. The Emu model is a two-stage approach that begins with a pre-training phase and concludes with a quality fine-tuning stage. Emu adapted the latent diffusion model architecture [18] to support high-resolution image generation and incorporated a 16-channel autoencoder with encoder $E$ and decoder $D$ . In the following section, we adapt the notation of [2]. A large U-Net, $\epsilon_{\theta}$ , with 2.8 billion parameters, $\theta$ , text embeddings from CLIP ViT-L [14] and T5-XXL [15], and a substantial pre-training dataset of 1.1 billion images facilitate the model's ability to learn complex semantics and finer details, with a noise-offset strategy contributing to high-contrast and aesthetically pleasing image generation. Given the encoded latent of an image $z = E(x)$ , the diffusion process generates a noisy latent $z_{t}$ where the noise level increases over timesteps $t \in T$ . To convert Emu to an instruction-based image editing model, we condition it on the image to be modified $c_{I}$ and the instruction $c_{T}$ . Emu Edit is trained to minimize the following optimization problem, + +$$ +\min _ {\theta} \mathbb {E} _ {y, \epsilon , t} \left[ \| \epsilon - \epsilon_ {\theta} \left(z _ {t}, t, E \left(c _ {I}\right), c _ {T}\right) \| _ {2} ^ {2} \right] \tag {1} +$$ + +where $\epsilon \in N(0,1)$ is the noise added by the diffusion process and $y = (c_{T},c_{I},x)$ is a triplet of instruction, input image and target image from the dataset. In practice, we initialize the weights of Emu Edit with the weights of Emu. To support the image conditioning, we follow [2] and increase the number of input channels. New weights are initialized to zero. During inference, we perform classifier-free guidance on both image and text conditions. In our experiments we use a scale of $\gamma_I = 1.5$ for the image condition and $\gamma_T = 5.0$ for the text condition. Furthermore, we apply a rescaling of the diffusion scheduler to achieve a zero signal-to-noise ratio (SNR) at the terminal timestamp, as suggested + +![](images/0119e04a5b0b1c102feed425ba9973d27d070e0a1684db23b73de9d48474c9ea.jpg) +Make it play a rainbow colored trumpet +Local. Baseslines struggle to perform intricate local edits. +Figure 3. Failure cases of baseline instruction-based image editing models.2 + +in [10]. This is crucial in order to avoid any mismatch between the model's training and testing phases. For more implementation details see supplementary Sec. ?? + +# 4.2. Learned Task Embeddings + +To guide the generation process toward the correct task, we learn an embedding vector for each task in the dataset. During training, given a sample from our dataset, we use the task index, $i$ , to fetch the task's embedding vector, $v_{i}$ , from an embedding table, and optimize it jointly with the model weights. We do so by introducing the task embedding $v_{i}$ as an additional condition to the U-Net, $\epsilon_{\theta}$ . Concretely, we integrate the task embedding into the U-Net via crossattention interactions, and by adding it to the timestep embeddings. The optimization problem is updated to + +$$ +\min _ {\theta , v _ {1}, \dots , v _ {k}} \mathbb {E} _ {\hat {y}, \epsilon , t} \left[ \| \epsilon - \epsilon_ {\theta} \left(z _ {t}, t, E \left(c _ {I}\right), c _ {T}, v _ {i}\right) \| _ {2} ^ {2} \right] \tag {2} +$$ + +where $k$ is the total number of tasks in our dataset and $\hat{y} = (c_I, c_T, x, i)$ is a quadruplet of input image, input instruction text, target image, and task index from the dataset. Task-specific conditioning arises from the observation that models lacking such conditioning can become perplexed about the type of edit required, particularly when the instructions are complex or the edit type is ambiguous. For + +instance, as visualized in Fig. 4, (1) a model without task conditioning might perform a global edit when a texture edit is required, (2) it might opt for segmentation when a global edit is necessary, and (3) it could implement a style edit in situations where a local edit would fit better. In the inference stage, we predict the task index. Specifically, we fine-tune a Flan-T5-XL model to identify the task at hand given the input instruction. + +# 4.3. Task Inversion + +To enable few-shot learning of new tasks without losing the general abilities of Emu Edit, we propose a method for adapting the model without changing the U-Net weights. Given a few examples of a new task, we learn a new task embedding, $v_{\mathrm{new}}$ . We freeze the model weights, and adapt it to the task only through the task embedding. Thus, to fit a new task embedding we solve the following optimization problem: + +$$ +\min _ {v _ {\text {n e w}}} \mathbb {E} _ {y, \epsilon , t} \left[ \| \epsilon - \epsilon_ {\theta} \left(z _ {t}, t, E \left(c _ {I}\right), c _ {T}, v _ {\text {n e w}}\right) \| _ {2} ^ {2} \right] \tag {3} +$$ + +where $v_{\mathrm{new}}$ is the learned task embedding. Note that during task inversion $y$ is a triplet belonging to the new task. + +The model can then be employed for the new task by conditioning it on the learned task embedding, and it can still handle its original tasks by relying on the initial task embeddings. In Sec. 5.3, we demonstrate that our model effectively generalizes to novel tasks using this method. + +![](images/c92be4727be869365485f5df5ea2528035c7666e91eb0d08737d1f4c7224804c.jpg) +Figure 4. Task embeddings. Model trained without task embeddings may get confused about the edit type when the instructions are complicated or there is ambiguity regarding the edit type: (1) Global edit (instead of Texture), (2) Segmentation (instead of Global), (3) Style edit (instead of Local). + +Sequential Edit Thresholding. We notice that applying the model repeatedly, in multi-turn editing scenarios, aggregates reconstruction and numerical errors, which translate to noticeable artifacts. To mitigate this, we add a per-pixel thresholding step after each edit-turn. At each step $s$ , we use the pixel value in the output image, $c_{I}^{s+1}$ , only if its alteration surpasses a specific threshold. Otherwise, we keep the pixel value from the input image, $c_{I}^{s}$ . Specifically, given an edit turn $s$ , we compute the absolute difference image $d = \|c_{I}^{s+1} - c_{I}^{s}\|_{1}$ over the RGB channel, and apply the following thresholding: + +$$ +c _ {I} ^ {s + 1} = \left\{ \begin{array}{l l} c _ {I} ^ {s} & \text {i f} \bar {d} < \alpha , \\ c _ {I} ^ {s + 1} & \text {o t h e r w i s e .} \end{array} \right. \tag {4} +$$ + +where, $\bar{d}$ is obtained after passing $d$ through a low pass filter, in order to smooth the transition between previous and current pixels. In practice, we choose $\alpha = 0.03$ . Please refer to supplementary Sec. ?? for a qualitative comparison. In Fig. 2 we show examples of multi-turn editing. + +# 5. Experiments + +Our experiments evaluate the ability of Emu Edit to follow user instructions faithfully and preserve the visual fidelity of the original image. First, we evaluate the performance of our approach on instruction-based image editing tasks. Second, we conduct a comprehensive ablation study to assess the effectiveness of our different contributions. Specifically, + +we ablate the contribution of the computer vision tasks to the model performance on image editing tasks, the importance of learned task embeddings, and the effect of multitask learning on instruction-based image editing. Further ablation on our data generation pipeline can be found in the supplementary Sec. ?? Finally, we demonstrate our model's ability to learn new tasks via few-shot learning. + +Measures. We employ two main measures in our evaluation: edit text alignment and image faithfulness. Specifically, for each pair of input image and editing instruction, we use the following automatic metrics: (i) CLIP [14] text-image direction similarity $(\mathrm{CLIP}_{dir})$ - measuring agreement between change in captions and the change in images, (ii) CLIP image similarity $(\mathrm{CLIP}_{img})$ - measuring change between edited and input image, (iii) CLIP output similarity $(\mathrm{CLIP}_{out})$ - measuring edited image similarity with output caption, (iv) L1 pixel-distance between input and edit image, and (v) DINO [4] similarity between the DINO embeddings of input and edited images. With the exception of the L1 distance, where lower values indicate better performance, higher values in all other measures signify better results. A low L1 distance translates to small changes in image's pixel values. A high DINO and $\mathrm{CLIP}_{img}$ similarity score, suggests semantic similarity between the images. For region-based edits, high image similarity scores indicate the edits were precise. For free-form edits, high similarity scores indicate image structure preservation. $\mathrm{CLIP}_{dir}$ and $\mathrm{CLIP}_{out}$ measure how well the model followed the instruction. In addition, we asked human raters to evaluate the text alignment and image faithfulness. In each evaluation scenario, raters are presented with two modified images alongside the original input image and instruction, and are presented with two questions: (i) Image Faithfulness: which image better preserves elements in the input image, and (ii) Text Alignment: which image best follows the instruction. + +# 5.1. Evaluation + +Throughout the paper, we report results on the MagicBrush test set [25] and the Emu Edit benchmark. In the following section, we describe our motivation for creating this benchmark, and detail its curation process. To date, there are two main benchmarks for evaluating instruction-based image editing capabilities. First, the InstructPix2Pix benchmark [2], which is intrinsically biased due to its reliance on generated Stable Diffusion [17] input images, and GPT-3 [3] generated instructions. Consequently, it is unclear whether its results will truly mirror the performance on real input images, with genuine user instructions. + +Unlike InstructPix2Pix, the second benchmark, MagicBrush [25], uses a diverse set of authentic input images from the MS-COCO benchmark [5, 11], and annotator-defined instructions. Nonetheless, this dataset also suffers + +Table 2. Comparison with image-editing baselines evaluated on Emu Edit test set and MagicBrush test set. For each benchmark we report CLIP, L1, DINO metrics and human ratings. Human evaluation shows the percentage of raters that prefer the results of Emu Edit. + +
Emu Edit Test SetMagicBrush Test Set
MethodCLIPdir↑CLIPim↑CLIPout↑L1↓DINO↑Text align. faith.CLIPdir↑CLIPim↑CLIPout↑L1↓DINO↑Text align. faith.
InstructPix2Pix [2]0.0780.8340.2190.1210.76277.3376.710.1150.8370.2450.0930.767
MagicBrush [25]0.0900.8380.2220.1000.77674.5074.100.1230.8830.2610.0580.871
PnP [21]0.0280.5210.0890.3040.15398.9599.000.0250.5680.1010.2890.220
Null-Text Inv. [12]0.1010.7610.2360.0750.67881.6385.470.1210.7520.2630.0770.664
Our0.1090.8590.2310.0940.819--0.1350.8970.2610.0520.879
+ +from inherent bias. During data collection, annotators were directed to use the DALLE-2 image editing platform [13] to generate the edited images. Thus, this benchmark is biased towards editing instructions that the DALLE-2 editor can successfully follow, which may compromise both its diversity and complexity. + +Emu Edit Benchmark. To collect a dataset with reduced bias and of higher diversity, we take a different approach. We first define seven different categories of potential image editing operations: background alteration (Background), comprehensive image changes (Global), style alteration (Style), object removal (Remove), object addition (Add), localized modifications (Local), and color/texture alterations (Texture). Then, we utilize the diverse set of input images from the MagicBrush benchmark [25], and for each editing operation, we task crowd workers to devise relevant, creative, and challenging instructions. Moreover, to increase the quality of the collected examples, we apply a post-verification stage, in which crowd workers filter examples with irrelevant instructions. Finally, to support evaluation for methods that require input and output captions [8, 21], we additionally collect an input caption and output caption. When doing so, we ask annotators to ensure that the captions capture both important elements in the image, and elements that should change based on the instruction. See supplementary Sec. ?? for examples of our benchmark, which we release to support better evaluation. + +Baseline Comparisons. We compare our model against two instruction-based image editing baseline models: InstructPix2Pix [2], and MagicBrush [25], which is a variant of InstructPix2Pix that was fine-tuned on the MagicBrush dataset. Additionally, we compare our model against two text-based image editing methods: PNP [21] and Null-Text Inversion modification of P2P [8, 12]. Unlike instruction-based models, these works expect image descriptions. Therefore, we provide them with access to the input caption and output caption. Note however, providing these methods with access to the ground-truth captions could potentially offer an advantage over instruction-based models, since the automatic metrics also rely on these captions. Tab. 2 shows our results versus the baselines. The findings indicate that human raters consistently prefer Emu + +Edit over all baselines by a large margin. Furthermore, apart from Null-Text Inversion, which as explained above, utilizes the ground-truth captions during inference, our approach outperforms the existing baselines on the automatic metrics. We provide qualitative comparisons in Fig. 3. Additional comparisons are available in the supplementary: Fig.??-??, and Fig.??-??. For performance on vision tasks, see Sec. ?? + +# 5.2. Ablations + +Computer Vision Tasks Enhance Image Editing Tasks. Here we demonstrate the importance of the vision tasks to Emu Edit performance on image editing tasks. For this, we trained two additional models on all tasks except: (i) detect and segment tasks, and (ii) image-to-image translation tasks. As we show in Tab. 4, adding the detection and segmentation tasks improves the model performance in region-based editing tasks. Additionally, we observe that image-to-image translation tasks improve the performance in free-form editing tasks. We hypothesize that the recognition tasks improve the model's recognition capabilities, leading to more accurate and precise localized modifications. Similarly, image-to-image tasks assist the model in understanding the entire image structure, thereby enhancing its capabilities for global operations. + +Contribution of Learned Task Embeddings. We compare three variants of Emu Edit: (i) conditioned on the ground-truth task embedding, (ii) conditioned on the task embedding, as predicted by the task predictor described in Sec. 4.2, and (iii) without conditioning on the task type.3 Tab. 3 shows the results on the validation set of our benchmark. As can be seen, conditioning on the task type boosts the model's performance. Furthermore, our task predictor closes the gap with the ground-truth conditioned model. Qualitatively, we observe that without conditioning on the task type the model may perform the wrong editing operation (Fig. 4). In supplementary Fig. ??, we demonstrate the effect of manipulating the task while keeping the instruction and input image fixed. As can be seen, changing the task embedding directly influences the task executed by the model. + +Table 3. Learned task embeddings ablation on our validation set. We compare variations of Emu Edit: without task type condition, with predicted task type, and with ground-truth task type. + +
MethodCLIPdir↑CLIPim↑CLIPout↑L1↓DINO↑
w/o task emb.0.1040.8430.2270.1090.792
with pred. task0.1170.8500.2310.1030.809
with gt task0.1190.8520.2310.1000.811
+ +Table 4. Contribution of computer vision tasks. Human evaluation is shown as a percentage of majority votes in favor of our model. + +
Region-BasedFree-form
MethodText align.Image faith.Text align.Image faith.
without detect/segment60.060.252.351.5
without im2im translation50.249.058.060.1
+ +![](images/39123bad4f47e58ef3229c1e7a2b69267c6105a096a1c393a5d87bbbf0acd560.jpg) +(i) Incorporate a bee into the bag's pattern and detect it + +![](images/550332dd0b724988cbe59bd88711563be1cac0be2f343cc947851469b953494d.jpg) + +![](images/28a809ddf6b49196190fc05506c4b8702c52aba451060a4af490f8cb3a107582.jpg) +(ii) Mark the gift bags +Input +Figure 5. Generations on unseen tasks with task inversion. (i) composition of add and detect tasks, (ii) object contour detection. + +![](images/92491dca5918554091ff7fed3d219e548fc2bb2af9abd1eb4159a98124759791.jpg) +Emu Edit + +Multi-Task versus Expert Models. We show that training a single model on a diverse range of tasks, leads to enhanced performance in each individual task, outperforming models that are specifically trained for a single task. To validate this, we train an expert model for each task, and compare its performance to the single one (See supplementary Tab. ??). + +Influence of Number of Tasks. Here, we ablate the number of tasks participating in the multi-task training scheme. In supplementary Fig. ?? we report the average $\mathrm{CLIP}_{dir}$ on the Style and Texture tasks when iteratively excluding other tasks, and training a model on the new tasks list. As can be seen, augmenting the model with additional tasks leads to improved performance, even in tasks which are not directly associated with the added ones. + +# 5.3. Few-Shot Learning of New Tasks + +Finally, we evaluate our model's ability to generalize by testing it in a few-shot scenario with previously unseen tasks. We test its performance across the following tasks: (i) super-resolution (x4), (ii) object contour detection, (iii) + +![](images/f4e4f96d972ec08ff045d7ace8a5e94898ace017418c3deb72d8af9abf0a0aea.jpg) +(a) Inpainting +Figure 6. Few-shot performance for different tasks over 1, 10, and 100 samples. Each line represents a different training setting: 'Scratch' finetune (Blue, $\bigcirc$ ), Emu Edit finetune (Orange, $\square$ ), task inversion (Green, $\triangle$ ), all compared to an upper-bound expert trained on 100k samples (Red dashed line, $\diamond$ ). + +![](images/d7be06768fe65bd0afeb7e4b6fe4566b2c930661aae31bb2c587ffc2c786595f.jpg) +(b) Add + Detect + +mask-based inpainting, and (iv) a composite task formed by combining two tasks from our dataset: add and detect. For each task, we assess the model's performance when trained with 0, 1, and 100 examples, with the 0-example case being equivalent to a zero-shot setting. We compare three baselines: (i) Scratch - Emu Edit initialized with Emu's weights, trained on the examples, (ii) Task Inversion - Emu Edit with task inversion (Sec. 4.3), and (iii) Finetune - Emu Edit where we finetune all of the model's weights on the examples. As an upper-bound expert, we train the first baseline on 100,000 examples. Note that, the "Task Inversion" and "Finetune" baselines were trained on the multitask dataset whereas the "Scratch" baseline was not. As can be seen in Fig. 6, fine-tuning with a single example is enough to significantly enhance the performance, while training from scratch results in overfitting. Moreover, utilizing 100 samples nearly achieves expert-level performance, implying that the model can effectively generalize to novel tasks. We also observe that task inversion performs similarly to complete fine-tuning, indicating that the model already contains essential information and can be prompted with a new task embedding to yield the desired result. For performance and generation examples on additional tasks see Fig. 5 and supplementary ????. + +# 6. Conclusion + +Emu Edit presents a step change in instructable image editing capabilities, primarily due to its unique training on both recognition and generation tasks. This dual-focus approach significantly enhances the model's comprehension of natural language instructions, enabling it to accurately execute a wide array of editing operations. Its ability to generalize to new tasks with minimal examples further demonstrates its versatility and robustness. Furthermore, our framework has the potential for further integration with a multimodal LLM in future projects, particularly for complex editing tasks requiring detailed reasoning from the input image, such as counting objects or executing intricate tasks. + +# References + +[1] Omri Avrahami, Dani Lischinski, and Ohad Fried. Blended diffusion for text-driven editing of natural images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18208-18218, 2022. 3 +[2] Tim Brooks et al. Instructpix2pix: Learning to follow image editing instructions. In CVPR, 2023. 1, 3, 4, 6, 7 +[3] Tom B. Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-Voss, Gretchen Krueger, T. J. Henighan, Rewon Child, Aditya Ramesh, Daniel M. Ziegler, Jeff Wu, Clemens Winter, Christopher Hesse, Mark Chen, Eric Sigler, Mateusz Litwin, Scott Gray, Benjamin Chess, Jack Clark, Christopher Berner, Sam McCandlish, Alec Radford, Ilya Sutskever, and Dario Amodei. Language models are few-shot learners. ArXiv, abs/2005.14165, 2020. 3, 6 +[4] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9650-9660, 2021. 6 +[5] Xinlei Chen, Hao Fang, Tsung-Yi Lin, Ramakrishna Vedantam, Saurabh Gupta, Piotr Dólar, and C. Lawrence Zittnick. Microsoft coco captions: Data collection and evaluation server. ArXiv, abs/1504.00325, 2015. 6 +[6] Xiaoliang Dai, Ji Hou, Chih-Yao Ma, Sam Tsai, Jialiang Wang, Rui Wang, Peizhao Zhang, Simon Vandenhende, Xiaofang Wang, Abhimanyu Dubey, et al. Emu: Enhancing image generation models using photogenic needles in a haystack. arXiv preprint arXiv:2309.15807, 2023. 4 +[7] Oran Gafni, Adam Polyak, Oron Ashual, Shelly Sheynin, Devi Parikh, and Yaniv Taigman. Make-a-scene: Scene-based text-to-image generation with human priors. In European Conference on Computer Vision, pages 89-106. Springer, 2022. 2 +[8] Amir Hertz, Ron Mokady, Jay M. Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-prompt image editing with cross attention control. ArXiv, abs/2208.01626, 2022. 2, 3, 7 +[9] Bahjat Kawar et al. Imagic: Text-based real image editing with diffusion models. 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6007-6017, 2022. 3 +[10] Shanchuan Lin, Bingchen Liu, Jiashi Li, and Xiao Yang. Common diffusion noise schedules and sample steps are flawed. arXiv preprint arXiv:2305.08891, 2023. 5 +[11] Tsung-Yi Lin, Michael Maire, Serge J. Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C. Lawrence Zitnick. Microsoft coco: Common objects in context. In ECCV, 2014. 6 +[12] Ron Mokady, Amir Hertz, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Null-text inversion for editing real images using guided diffusion models. 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6038–6047, 2022. 3, 7 + +[13] OpenAI. Dall-e 2, 2022. https://openai.com/product/dall-e-2.3,7 +[14] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 1, 4, 6 +[15] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J. Liu. Exploring the limits of transfer learning with a unified text-to-text transformer. Journal of Machine Learning Research, 21(140):1-67, 2020. 4 +[16] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022. 2 +[17] Robin Rombach, A. Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10674-10685, 2021. 2, 6 +[18] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models, 2021. 4 +[19] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L. Denton, Seyed Kamyar Seyed Ghasemipour, Burcu Karagol Ayan, Seyedeh Sara Mahdavi, Raphael Gontijo Lopes, Tim Salimans, Jonathan Ho, David J. Fleet, and Mohammad Norouzi. Photorealistic text-to-image diffusion models with deep language understanding. ArXiv, abs/2205.11487, 2022. 2 +[20] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288, 2023. 3 +[21] Narek Tumanyan, Michal Geyer, Shai Bagon, and Tali Dekel. Plug-and-play diffusion features for text-driven image-to-image translation. 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1921-1930, 2022. 3, 7 +[22] Bram Wallace, Akash Gokul, and Nikhil Vijay Naik. Edict: Exact diffusion inversion via coupled transformations. 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 22532-22541, 2022. 3 +[23] Su Wang et al. Imagen editor and editbench: Advancing and evaluating text-guided image inpainting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18359-18369, 2023. 3 +[24] Shaoan Xie, Zhifei Zhang, Zhe Lin, Tobias Hinz, and Kun Zhang. Smartbrush: Text and shape guided object inpainting with diffusion model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22428-22437, 2023. 3 +[25] Kai Zhang et al. Magicbrush: A manually annotated dataset for instruction-guided image editing. arXiv:2306.10012, 2023. 1, 3, 6, 7 \ No newline at end of file diff --git a/2024/Emu Edit_ Precise Image Editing via Recognition and Generation Tasks/images.zip b/2024/Emu Edit_ Precise Image Editing via Recognition and Generation Tasks/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..0569f4edeb4bce9b90df4769de64f9f776261eb7 --- /dev/null +++ b/2024/Emu Edit_ Precise Image Editing via Recognition and Generation Tasks/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b8d499be3157fe800080c5c9d13699460149c8892bff88ba63d37aab00f14ab7 +size 620786 diff --git a/2024/Emu Edit_ Precise Image Editing via Recognition and Generation Tasks/layout.json b/2024/Emu Edit_ Precise Image Editing via Recognition and Generation Tasks/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..73edf6d6180b16a7b15669a146944eb67fb1c45d --- /dev/null +++ b/2024/Emu Edit_ Precise Image Editing via Recognition and Generation Tasks/layout.json @@ -0,0 +1,7281 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 75, + 103, + 518, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 103, + 518, + 121 + ], + "spans": [ + { + "bbox": [ + 75, + 103, + 518, + 121 + ], + "type": "text", + "content": "Emu Edit: Precise Image Editing via Recognition and Generation Tasks" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 75, + 142, + 517, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 142, + 517, + 171 + ], + "spans": [ + { + "bbox": [ + 75, + 142, + 517, + 171 + ], + "type": "text", + "content": "Shelly Sheynin*, Adam Polyak*, Uriel Singer*, Yuval Kirstain*, Amit Zohar*, Oron Ashual, Devi Parikh and Yaniv Taigman" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 263, + 185, + 331, + 198 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 263, + 185, + 331, + 198 + ], + "spans": [ + { + "bbox": [ + 263, + 185, + 331, + 198 + ], + "type": "text", + "content": "GenAI,Meta" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 51, + 217, + 214, + 299 + ], + "blocks": [ + { + "bbox": [ + 51, + 217, + 214, + 299 + ], + "lines": [ + { + "bbox": [ + 51, + 217, + 214, + 299 + ], + "spans": [ + { + "bbox": [ + 51, + 217, + 214, + 299 + ], + "type": "image", + "image_path": "1daaa348ca8afee9b9522231b19578495255b61c69bb7fa381085ebff184eec9.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 89, + 300, + 189, + 309 + ], + "lines": [ + { + "bbox": [ + 89, + 300, + 189, + 309 + ], + "spans": [ + { + "bbox": [ + 89, + 300, + 189, + 309 + ], + "type": "text", + "content": "Dress the emu with a fireman outfit" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 216, + 217, + 378, + 299 + ], + "blocks": [ + { + "bbox": [ + 216, + 217, + 378, + 299 + ], + "lines": [ + { + "bbox": [ + 216, + 217, + 378, + 299 + ], + "spans": [ + { + "bbox": [ + 216, + 217, + 378, + 299 + ], + "type": "image", + "image_path": "121a82bdf17aea58535197abd2870754df60ef1528d39a5ce16c55d569d0f221.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 272, + 300, + 337, + 309 + ], + "lines": [ + { + "bbox": [ + 272, + 300, + 337, + 309 + ], + "spans": [ + { + "bbox": [ + 272, + 300, + 337, + 309 + ], + "type": "text", + "content": "Let's see it graduating" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 47, + 317, + 544, + 330 + ], + "lines": [ + { + "bbox": [ + 47, + 317, + 544, + 330 + ], + "spans": [ + { + "bbox": [ + 47, + 317, + 544, + 330 + ], + "type": "text", + "content": "Figure 1. Emu Edit is a multi-tasking model that combines various editing (left, middle) and vision (right) tasks for precise image editing." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 379, + 217, + 543, + 299 + ], + "blocks": [ + { + "bbox": [ + 379, + 217, + 543, + 299 + ], + "lines": [ + { + "bbox": [ + 379, + 217, + 543, + 299 + ], + "spans": [ + { + "bbox": [ + 379, + 217, + 543, + 299 + ], + "type": "image", + "image_path": "0967a632fde260268fe3b3f4fc15dccddf079cf81e5f8fac9b18c6cde487e995.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 438, + 300, + 484, + 308 + ], + "lines": [ + { + "bbox": [ + 438, + 300, + 484, + 308 + ], + "spans": [ + { + "bbox": [ + 438, + 300, + 484, + 308 + ], + "type": "text", + "content": "Mark the drinks" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 143, + 338, + 192, + 350 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 338, + 192, + 350 + ], + "spans": [ + { + "bbox": [ + 143, + 338, + 192, + 350 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 46, + 362, + 290, + 649 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 362, + 290, + 649 + ], + "spans": [ + { + "bbox": [ + 46, + 362, + 290, + 649 + ], + "type": "text", + "content": "Instruction-based image editing holds immense potential for a variety of applications, as it enables users to perform any editing operation using a natural language instruction. However, current models in this domain often struggle with accurately executing user instructions. We present Emu Edit, a multi-task image editing model which sets state-of-the-art results in instruction-based image editing. To develop Emu Edit we train it to multi-task across an unprecedented range of tasks, such as region-based editing, free-form editing, and Computer Vision tasks, all of which are formulated as generative tasks. Additionally, to enhance Emu Edit's multi-task learning abilities, we provide it with learned task embeddings which guide the generation process towards the correct edit type. Both these elements are essential for Emu Edit's outstanding performance. Furthermore, we show that Emu Edit can generalize to new tasks, such as image inpainting, super-resolution, and compositions of editing tasks, with just a few labeled examples. This capability offers a significant advantage in scenarios where high-quality samples are scarce. Lastly, to facilitate a more rigorous and informed assessment of instructable image editing models, we release a new challenging and versatile benchmark that includes seven different image editing tasks." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 307, + 338, + 386, + 350 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 338, + 386, + 350 + ], + "spans": [ + { + "bbox": [ + 307, + 338, + 386, + 350 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 366, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 366, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 366, + 547, + 715 + ], + "type": "text", + "content": "Image editing is a widely-used application that millions engage with every day. Popular image editing tools, however, either demand considerable expertise and are time-consuming to use, or are quite limited, providing only a predefined set of editing operations, such as specific filters. Instruction-based image editing [2, 25] attempts to resolve these limitations by allowing users to effortlessly describe their editing goals using natural language instructions. For instance, a user can provide a model with an image and instruct it to \"Dress the emu with a fireman outfit\" or \"Let's see it graduating\" (see Fig. 1). Nevertheless, while instruction-based image editing models like InstructPix2Pix [2] are designed to process any given instruction, they often struggle to accurately interpret and execute such instructions. Moreover, their generalization is limited, often falling short on tasks that deviate slightly from those they were trained on (see Fig. 3). To address these gaps, we introduce Emu Edit, the first image editing model trained on an extensive and diverse set of tasks, including both image editing and computer vision tasks. Emu Edit provides a substantial improvement in both compliance with the edit instruction and preservation of the visual fidelity of the original image. As we show through both automatic metrics [14] and human judgments on two benchmarks [25], Emu Edit achieves state-of-the-art results in instruction-based image editing. The success of Emu Edit stems from two key contributions. First, we train our model to multi-task across sixteen distinct image editing tasks. These tasks span region-based editing tasks, free-form editing tasks and computer" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 693, + 267, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 693, + 267, + 703 + ], + "spans": [ + { + "bbox": [ + 58, + 693, + 267, + 703 + ], + "type": "text", + "content": "1Project Page: https://emu-edit.metademolab.com/" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 59, + 703, + 126, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 703, + 126, + 712 + ], + "spans": [ + { + "bbox": [ + 59, + 703, + 126, + 712 + ], + "type": "text", + "content": "*Equal contribution." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "8871" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 49, + 81, + 148, + 178 + ], + "blocks": [ + { + "bbox": [ + 49, + 81, + 148, + 178 + ], + "lines": [ + { + "bbox": [ + 49, + 81, + 148, + 178 + ], + "spans": [ + { + "bbox": [ + 49, + 81, + 148, + 178 + ], + "type": "image", + "image_path": "f6daee2dd1ad5774950c595167a32770ca7e13f1189f15de497229100214adba.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 84, + 180, + 113, + 189 + ], + "lines": [ + { + "bbox": [ + 84, + 180, + 113, + 189 + ], + "spans": [ + { + "bbox": [ + 84, + 180, + 113, + 189 + ], + "type": "text", + "content": "(a) A cat" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 148, + 81, + 247, + 178 + ], + "blocks": [ + { + "bbox": [ + 148, + 81, + 247, + 178 + ], + "lines": [ + { + "bbox": [ + 148, + 81, + 247, + 178 + ], + "spans": [ + { + "bbox": [ + 148, + 81, + 247, + 178 + ], + "type": "image", + "image_path": "e20d245dddf075ebb2f9a85e3577ad0e4ad463efa803748455a964ab0c108ffd.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 166, + 180, + 229, + 189 + ], + "lines": [ + { + "bbox": [ + 166, + 180, + 229, + 189 + ], + "spans": [ + { + "bbox": [ + 166, + 180, + 229, + 189 + ], + "type": "text", + "content": "(b) Remove the tail" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 247, + 81, + 346, + 178 + ], + "blocks": [ + { + "bbox": [ + 247, + 81, + 346, + 178 + ], + "lines": [ + { + "bbox": [ + 247, + 81, + 346, + 178 + ], + "spans": [ + { + "bbox": [ + 247, + 81, + 346, + 178 + ], + "type": "image", + "image_path": "53da06711e6d3d8423eacac630c92eea95a427e99482ac28db72a69654f67fc4.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 262, + 180, + 331, + 189 + ], + "lines": [ + { + "bbox": [ + 262, + 180, + 331, + 189 + ], + "spans": [ + { + "bbox": [ + 262, + 180, + 331, + 189 + ], + "type": "text", + "content": "(c) Add a pink jacket" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 347, + 81, + 445, + 178 + ], + "blocks": [ + { + "bbox": [ + 347, + 81, + 445, + 178 + ], + "lines": [ + { + "bbox": [ + 347, + 81, + 445, + 178 + ], + "spans": [ + { + "bbox": [ + 347, + 81, + 445, + 178 + ], + "type": "image", + "image_path": "3625b84b9371fb02d492e67157add04fc9d29f97c7717af61ecbcdf2f791e658.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 369, + 180, + 424, + 189 + ], + "lines": [ + { + "bbox": [ + 369, + 180, + 424, + 189 + ], + "spans": [ + { + "bbox": [ + 369, + 180, + 424, + 189 + ], + "type": "text", + "content": "(d)Make it rainy" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 445, + 81, + 545, + 178 + ], + "blocks": [ + { + "bbox": [ + 445, + 81, + 545, + 178 + ], + "lines": [ + { + "bbox": [ + 445, + 81, + 545, + 178 + ], + "spans": [ + { + "bbox": [ + 445, + 81, + 545, + 178 + ], + "type": "image", + "image_path": "506af08233d4e36e2f8bbce59603fe18b2c59cfb79e7660fc10d64ab3bfbdf70.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 447, + 180, + 543, + 189 + ], + "lines": [ + { + "bbox": [ + 447, + 180, + 543, + 189 + ], + "spans": [ + { + "bbox": [ + 447, + 180, + 543, + 189 + ], + "type": "text", + "content": "(e) Have the cat look shocked" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 49, + 191, + 148, + 289 + ], + "blocks": [ + { + "bbox": [ + 49, + 191, + 148, + 289 + ], + "lines": [ + { + "bbox": [ + 49, + 191, + 148, + 289 + ], + "spans": [ + { + "bbox": [ + 49, + 191, + 148, + 289 + ], + "type": "image", + "image_path": "2f4c3ffb2eee89852984d434d8896b910ed6d6dd66675ea26147c60a85839bc1.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 58, + 290, + 140, + 300 + ], + "lines": [ + { + "bbox": [ + 58, + 290, + 140, + 300 + ], + "spans": [ + { + "bbox": [ + 58, + 290, + 140, + 300 + ], + "type": "text", + "content": "(f) Extract the depth map" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 148, + 191, + 253, + 289 + ], + "blocks": [ + { + "bbox": [ + 148, + 191, + 253, + 289 + ], + "lines": [ + { + "bbox": [ + 148, + 191, + 253, + 289 + ], + "spans": [ + { + "bbox": [ + 148, + 191, + 253, + 289 + ], + "type": "image", + "image_path": "4641d1eab2feee72b020ae720e2337bfaf683f0a908d5c765094e57532838c29.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 155, + 290, + 240, + 319 + ], + "lines": [ + { + "bbox": [ + 155, + 290, + 240, + 319 + ], + "spans": [ + { + "bbox": [ + 155, + 290, + 240, + 319 + ], + "type": "text", + "content": "(g) Generate a rainy day image of a hedgehog in a dress using the depth map" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 253, + 191, + 346, + 289 + ], + "blocks": [ + { + "bbox": [ + 253, + 191, + 346, + 289 + ], + "lines": [ + { + "bbox": [ + 253, + 191, + 346, + 289 + ], + "spans": [ + { + "bbox": [ + 253, + 191, + 346, + 289 + ], + "type": "image", + "image_path": "c9f3cb108decacef2bbdc2436d64171e786f0940660fbc7b29b03667bd400d6b.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 249, + 290, + 345, + 308 + ], + "lines": [ + { + "bbox": [ + 249, + 290, + 345, + 308 + ], + "spans": [ + { + "bbox": [ + 249, + 290, + 345, + 308 + ], + "type": "text", + "content": "(h) Replace the dress with an astronaut outfit" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 347, + 191, + 445, + 289 + ], + "blocks": [ + { + "bbox": [ + 347, + 191, + 445, + 289 + ], + "lines": [ + { + "bbox": [ + 347, + 191, + 445, + 289 + ], + "spans": [ + { + "bbox": [ + 347, + 191, + 445, + 289 + ], + "type": "image", + "image_path": "d8b340b4db0cf630a244006cf7460b8d8e3f058dd3c6e89b6c3f84e7ed43e8a3.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 347, + 290, + 445, + 308 + ], + "lines": [ + { + "bbox": [ + 347, + 290, + 445, + 308 + ], + "spans": [ + { + "bbox": [ + 347, + 290, + 445, + 308 + ], + "type": "text", + "content": "(i) Segment the spacesuit, and detect the hands" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 328, + 546, + 350 + ], + "lines": [ + { + "bbox": [ + 46, + 328, + 546, + 350 + ], + "spans": [ + { + "bbox": [ + 46, + 328, + 546, + 350 + ], + "type": "text", + "content": "Figure 2. Instruction-based image generation. Each image originates from a previous one with a corresponding instruction. The prior image for (a) was set to zero." + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 446, + 191, + 545, + 289 + ], + "blocks": [ + { + "bbox": [ + 446, + 191, + 545, + 289 + ], + "lines": [ + { + "bbox": [ + 446, + 191, + 545, + 289 + ], + "spans": [ + { + "bbox": [ + 446, + 191, + 545, + 289 + ], + "type": "image", + "image_path": "421fc7fc2f04003c5f2645ecebf7d8aa0491255a119677dfbe47c1f0ac10e446.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 448, + 290, + 542, + 309 + ], + "lines": [ + { + "bbox": [ + 448, + 290, + 542, + 309 + ], + "spans": [ + { + "bbox": [ + 448, + 290, + 542, + 309 + ], + "type": "text", + "content": "(j) Add the text \"purple cat\" using a purple font" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "bbox": [ + 46, + 354, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 354, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 354, + 289, + 715 + ], + "type": "text", + "content": "vision tasks, all formulated as generative tasks. Unlike previous work, we develop a distinct data generation pipeline for each task, allowing us to gather a training set that is not only more diverse but also more precise in its examples. We find that training a single model on all tasks yields better results than training expert models on each task independently. Additionally, we show that as the number of training tasks increases, so does the performance of Emu Edit. Furthermore, we discover that surprisingly, computer vision tasks such as detection, segmentation, and others, significantly enhance editing performance, as validated both by human raters as well as quantitative measures. Second, to process this wide array of tasks effectively, we introduce the concept of learned task embeddings, which are used to steer the generation process toward the correct generative task. Concretely, for each task, we learn a unique task embedding vector, and integrate it into the model through cross-attention interactions, and by adding it to the timestep embeddings. We demonstrate that learned task embeddings significantly enhance the ability of our model to accurately infer the appropriate edit type from the user instruction and execute the correct edit. Equipped with a robust model trained across a broad spectrum of tasks and guided by learned task embeddings, we explore few-shot adaptation to unseen tasks via task inversion. In this process, we maintain the model weights untouched, and solely update a task embedding to fit the new task. Our experiments demonstrate that Emu Edit can swiftly adapt to new tasks, such as superresolution, contour detection, and others. Notably, for some tasks, fine-tuning the model on just a handful of examples" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 304, + 354, + 547, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 354, + 547, + 594 + ], + "spans": [ + { + "bbox": [ + 304, + 354, + 547, + 594 + ], + "type": "text", + "content": "yields results that nearly match those of an expert model trained on one hundred thousand examples. This makes task inversion with Emu Edit particularly advantageous in scenarios where labeled examples are limited, or when the compute budget is low. Finally, to support future research for instruction-based image editing, we publicly release a diverse and challenging benchmark that includes seven different image editing operations, as well as our model's generations on this dataset. In summary, this work addresses the limitations of instruction-based image editing models in accurately following user instructions. We demonstrate that by employing multi-task training across a diverse array of tasks, including recognition, generation, and editing, we can enhance our model's performance. Furthermore, by incorporating learned task embeddings into our model's architecture, we not only improve its results but also enable efficient few-shot learning for new tasks. With these improvements, our model sets a new standard in the field, offering significantly more precise and robust instruction-based image editing capabilities than existing models." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 306, + 609, + 392, + 621 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 609, + 392, + 621 + ], + "spans": [ + { + "bbox": [ + 306, + 609, + 392, + 621 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 304, + 629, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 629, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 629, + 547, + 715 + ], + "type": "text", + "content": "The emergence of high-performing text-to-image diffusion models [7, 16, 17, 19] facilitated the development of effective text-based image editing methods. Such methods usually employ aligned and detailed descriptions of the input and edited image to perform a specific edit. Prompt-to-Prompt (P2P) [8] injects the input caption attention maps to the target caption attentions maps. Null-Text Inver" + } + ] + } + ], + "index": 24 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8872" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 287 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 287 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 287 + ], + "type": "text", + "content": "sion [12] inverts an input image using the null-text embedding to support editing of a real image. Plug-and-Play (PNP) [21] injects spatial features in addition to attention maps and obtains better performance at global editing. Imagic [9] finetunes the diffusion model to support complex textual instructions. EDICT [22] suggests an image inversion based on two noise vectors enabling better image reconstruction and textual faithfulness. Another class of image editing models, employs an input mask as additional input [1, 23, 24]. Blended Diffusion [1] modifies the diffusion step by blending the input image in the unmasked regions. Imagen Editor [23] and SmartBrush [24] finetune the text-to-image model to be conditioned on both the input image and mask. While the text-based image editing methods detailed above enable humans to edit images, they frequently exhibit inconsistent performance and require multiple inputs, such as aligned and detailed descriptions of both the input and target images, or at times, input masks." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 288, + 289, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 288, + 289, + 491 + ], + "spans": [ + { + "bbox": [ + 46, + 288, + 289, + 491 + ], + "type": "text", + "content": "To offer a more intuitive and user-friendly interface, and significantly enhance ease of use for humans, InstructPix2Pix [2] introduced an instructable image editing model. They developed this model by utilizing both GPT-3 [3] and Prompt-to-Prompt [8], to generate a large synthetic dataset for instruction-based image editing, and employed the dataset to train an instructable image editing model. Unlike InstructPix2Pix which used a synthetic dataset, MagicBrush [25] developed a manually-annotated instruction-guided image editing dataset by requesting humans to use an online image editing tool [13]. Finetuning InstructPix2Pix on this dataset led to improved image editing capabilities. However, even though there has been progress and improvement in instruction-based image editing models, we show in Sec. 5 that state-of-the-art image editing models still struggle with accurately interpreting and precisely executing editing instructions." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 491, + 289, + 612 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 491, + 289, + 612 + ], + "spans": [ + { + "bbox": [ + 46, + 491, + 289, + 612 + ], + "type": "text", + "content": "In this paper, we drastically narrow such performance gaps by leveraging multi-task training and a matching architecture. Unlike prior work that solely focuses on image editing [2, 25], we train our model to perform various tasks and learn a very diverse set of capabilities. The quality and versatility of our training procedure and dataset, together with our improved architecture for multi-task learning, enables us to make a big leap in performance and differentiates us from prior work in the field. Fig. 3 include several challenging editing samples as examples." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 621, + 255, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 621, + 255, + 635 + ], + "spans": [ + { + "bbox": [ + 47, + 621, + 255, + 635 + ], + "type": "text", + "content": "3. Multi-Task Dataset for Image Editing" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 641, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 641, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 641, + 288, + 715 + ], + "type": "text", + "content": "Training a robust and accurate image editing model requires a highly diverse dataset of input images, editing instructions, and output edited images. However, manually collecting such examples is impractically time-consuming, existing sources on the web (e.g. communities and forums on social media) are limited in size, and publicly available syn-" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 306, + 74, + 413, + 86 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 74, + 413, + 86 + ], + "spans": [ + { + "bbox": [ + 306, + 74, + 413, + 86 + ], + "type": "text", + "content": "1. Region-Based Editing" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 87, + 545, + 182 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 306, + 87, + 545, + 110 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 87, + 545, + 110 + ], + "spans": [ + { + "bbox": [ + 306, + 87, + 545, + 110 + ], + "type": "text", + "content": "- Local: Substituting one object for another, altering an object's attributes (e.g., \"make it smile\")." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 110, + 492, + 121 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 110, + 492, + 121 + ], + "spans": [ + { + "bbox": [ + 306, + 110, + 492, + 121 + ], + "type": "text", + "content": "- Remove: Erasing an object from the image." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 122, + 491, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 122, + 491, + 133 + ], + "spans": [ + { + "bbox": [ + 306, + 122, + 491, + 133 + ], + "type": "text", + "content": "- Add: Inserting a new object into the image." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 134, + 545, + 169 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 134, + 545, + 169 + ], + "spans": [ + { + "bbox": [ + 306, + 134, + 545, + 169 + ], + "type": "text", + "content": "- Texture: Altering an object's visual characteristics without affecting its structure (e.g., painting over, filling or covering an object)." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 170, + 512, + 182 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 170, + 512, + 182 + ], + "spans": [ + { + "bbox": [ + 306, + 170, + 512, + 182 + ], + "type": "text", + "content": "- Background: Changing the scene's background." + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 187, + 400, + 198 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 187, + 400, + 198 + ], + "spans": [ + { + "bbox": [ + 306, + 187, + 400, + 198 + ], + "type": "text", + "content": "2. Free-Form Editing" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 199, + 545, + 281 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 306, + 199, + 545, + 234 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 199, + 545, + 234 + ], + "spans": [ + { + "bbox": [ + 306, + 199, + 545, + 234 + ], + "type": "text", + "content": "- Global: Edit instructions that affect the entire image, or that can not be described using a mask (e.g., \"let's see it in the summer\")." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 235, + 462, + 247 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 235, + 462, + 247 + ], + "spans": [ + { + "bbox": [ + 306, + 235, + 462, + 247 + ], + "type": "text", + "content": "- Style: Change the style of an image." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 247, + 545, + 281 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 247, + 545, + 281 + ], + "spans": [ + { + "bbox": [ + 306, + 247, + 545, + 281 + ], + "type": "text", + "content": "- Text Editing: This involves text-related editing tasks such as adding, removing, swapping text, and altering the text's font and color." + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 288, + 371, + 298 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 288, + 371, + 298 + ], + "spans": [ + { + "bbox": [ + 306, + 288, + 371, + 298 + ], + "type": "text", + "content": "3. Vision tasks" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 306, + 300, + 545, + 395 + ], + "type": "list", + "angle": 0, + "index": 22, + "blocks": [ + { + "bbox": [ + 306, + 300, + 545, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 300, + 545, + 323 + ], + "spans": [ + { + "bbox": [ + 306, + 300, + 545, + 323 + ], + "type": "text", + "content": "- Detect: Identifying and marking a specific object within the image with a rectangle bounding box." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 306, + 324, + 539, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 324, + 539, + 335 + ], + "spans": [ + { + "bbox": [ + 306, + 324, + 539, + 335 + ], + "type": "text", + "content": "- Segment: Isolating and marking an object in the image." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 306, + 335, + 538, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 335, + 538, + 346 + ], + "spans": [ + { + "bbox": [ + 306, + 335, + 538, + 346 + ], + "type": "text", + "content": "Color: Color adjustments like sharpening and blurring." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 348, + 545, + 395 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 348, + 545, + 395 + ], + "spans": [ + { + "bbox": [ + 307, + 348, + 545, + 395 + ], + "type": "text", + "content": "- Image-to-Image Translation: Tasks that involve bidirectional image type conversion, such as sketch-to-image, depth map-to-image, normal map-to-image, pose-to-image, segmentation map-to-image, and so on." + } + ] + } + ], + "index": 21 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 310, + 407, + 539, + 418 + ], + "angle": 0, + "lines": [ + { + "bbox": [ + 310, + 407, + 539, + 418 + ], + "spans": [ + { + "bbox": [ + 310, + 407, + 539, + 418 + ], + "type": "text", + "content": "Table 1. Description of the tasks forming the Emu Edit dataset." + } + ] + } + ], + "index": 23, + "type": "text" + }, + { + "bbox": [ + 304, + 427, + 545, + 535 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 427, + 545, + 535 + ], + "spans": [ + { + "bbox": [ + 304, + 427, + 545, + 535 + ], + "type": "text", + "content": "thetic datasets often lack in diversity or quality. Therefore, we construct a new dataset that encompasses sixteen distinct tasks and ten million examples. Each example " + }, + { + "bbox": [ + 304, + 427, + 545, + 535 + ], + "type": "inline_equation", + "content": "(c_{I}, c_{T}, x, i)" + }, + { + "bbox": [ + 304, + 427, + 545, + 535 + ], + "type": "text", + "content": " in our dataset, contains an input image " + }, + { + "bbox": [ + 304, + 427, + 545, + 535 + ], + "type": "inline_equation", + "content": "c_{I}" + }, + { + "bbox": [ + 304, + 427, + 545, + 535 + ], + "type": "text", + "content": ", a text instruction " + }, + { + "bbox": [ + 304, + 427, + 545, + 535 + ], + "type": "inline_equation", + "content": "c_{T}" + }, + { + "bbox": [ + 304, + 427, + 545, + 535 + ], + "type": "text", + "content": ", a target image " + }, + { + "bbox": [ + 304, + 427, + 545, + 535 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 304, + 427, + 545, + 535 + ], + "type": "text", + "content": ", and a task index " + }, + { + "bbox": [ + 304, + 427, + 545, + 535 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 427, + 545, + 535 + ], + "type": "text", + "content": " (out of the sixteen). The following section outlines the process of the data construction. In Sec. 3.1 we describe the instruction generation process, and in Sec. 3.2 the image pairs " + }, + { + "bbox": [ + 304, + 427, + 545, + 535 + ], + "type": "inline_equation", + "content": "(c_{I}, x)" + }, + { + "bbox": [ + 304, + 427, + 545, + 535 + ], + "type": "text", + "content": " generation and filtering." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 304, + 552, + 545, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 552, + 545, + 613 + ], + "spans": [ + { + "bbox": [ + 304, + 552, + 545, + 613 + ], + "type": "text", + "content": "Task Categories. The dataset is composed of tasks which are divided into three main categories: region-based editing, free-form editing, and vision tasks. Tab. 1 includes the full list of tasks, and their distribution in the train set is visualized in Fig. ??" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 306, + 622, + 436, + 634 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 622, + 436, + 634 + ], + "spans": [ + { + "bbox": [ + 306, + 622, + 436, + 634 + ], + "type": "text", + "content": "3.1. Instruction Generation" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "type": "text", + "content": "To generate editing instructions, we leverage the dialogue-optimized 70 billion parameter Llama 2 variant [20]. We observed that using a single agent to generate the instructions for all tasks leads to a lack of diversity in the dataset. Notably, the LLM exhibits a bias towards particular tasks and instruction phrasings. To address this, we utilize in" + } + ] + } + ], + "index": 27 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8873" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 204 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 204 + ], + "type": "text", + "content": "context learning to create a task-specific agent for each task. Concretely, we provide the LLM with a task description, a few task-specific exemplars, and a real image caption. To increase diversity we sample the exemplars and randomize their order. Given such input, we expect the LLM to output (1) an editing instruction, (2) an output caption for an ideal output image, and (3) which objects should be updated or added to the original image. We refer the readers to supplementary Figs. ??-?? for examples of our prompts. Further details on instruction generation are provided in the supplementary Sec. ??" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 214, + 181, + 227 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 214, + 181, + 227 + ], + "spans": [ + { + "bbox": [ + 47, + 214, + 181, + 227 + ], + "type": "text", + "content": "3.2. Image Pairs Generation" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 234, + 289, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 234, + 289, + 617 + ], + "spans": [ + { + "bbox": [ + 46, + 234, + 289, + 617 + ], + "type": "text", + "content": "Our aim is to generate pairs of input and edited images that adhere to the edit instructions and preserve image elements that should remain intact. A crucial prerequisite when creating a pair of input and edited images is to guarantee that the two images differ only in specific elements or locations, while remaining identical in all other aspects. Previous instruct-based image editing methods [2] rely on Prompt-to-Prompt (P2P) to build an image-editing dataset. P2P injects cross-attention maps from the input image generation to the edited image generation. To support local edits, P2P additionally approximates a mask of the edited part, based on the cross-attention maps and constrains the edit to this local area. P2P relies on word-to-word alignment between the input image caption and the edited image caption (e.g. \"a cat riding a bicycle\" and \"a cat riding a car\") to produce editing image pairs. However, when there is no word-to-word alignment, the resulting mask tends to be imprecise due to its reliance on cross-attention maps. Furthermore, as word-to-word alignment is not a practical assumption in most of the image editing tasks, this approach often fails to preserve structure and identity. To address this challenge, we propose a mask extraction method, which is applied before the editing process. Our approach involves: (i) identifying the edited areas from the editing instruction via an LLM and creating corresponding masks before image generation, and (ii) integrating these masks during the editing process to ensure seamless fusion of edited regions with the original image. Distinct editing challenges, such as adding or removing objects, require tailored solutions. We utilize various techniques, including dilation and Gaussian blurring, to refine the masks. Further description of the method can be found at supplementary Sec. ??" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 618, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 618, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 618, + 288, + 715 + ], + "type": "text", + "content": "Filtering. We employ a comprehensive filtering approach to ensure the fidelity of the dataset. This includes: (i) using the task predictor (Sec. 4.2) to reassign samples with instructions that should belong to another task, (ii) applying CLIP filtering metrics [2], (iii) employing structure preserving filtering based on the L1 distance between the depth map of the input image and the edited image, and (iv) applying image detectors to validate the presence (in Add task)," + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "text", + "content": "the absence (in Remove task) or replacement (in Local task) of elements, according to the objects specified in the instruction. This process filters out " + }, + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "inline_equation", + "content": "70\\%" + }, + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "text", + "content": " of the data, resulting in a final dataset of ten million samples." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 305, + 131, + 362, + 144 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 131, + 362, + 144 + ], + "spans": [ + { + "bbox": [ + 305, + 131, + 362, + 144 + ], + "type": "text", + "content": "4. Method" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 151, + 545, + 281 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 151, + 545, + 281 + ], + "spans": [ + { + "bbox": [ + 304, + 151, + 545, + 281 + ], + "type": "text", + "content": "Emu Edit is a diffusion model designed to multi-task across a broad spectrum of editing tasks. As Emu Edit is trained on various tasks, a crucial aspect is the ability to identify the semantic edit (e.g., global/local/texture) that needs to be applied, based on the user instruction. To address this, we use task-specific embeddings integrated into the model, learned during training. After training, Emu Edit can quickly adapt to new tasks by learning a new task embedding through few-shot learning, without altering the rest of the model. We follow next with a detailed description of each part of our method." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 290, + 388, + 302 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 290, + 388, + 302 + ], + "spans": [ + { + "bbox": [ + 305, + 290, + 388, + 302 + ], + "type": "text", + "content": "4.1. Architecture" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 308, + 545, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 308, + 545, + 548 + ], + "spans": [ + { + "bbox": [ + 304, + 308, + 545, + 548 + ], + "type": "text", + "content": "Our model builds upon the foundation set by Emu, which is outlined in [6]. The Emu model is a two-stage approach that begins with a pre-training phase and concludes with a quality fine-tuning stage. Emu adapted the latent diffusion model architecture [18] to support high-resolution image generation and incorporated a 16-channel autoencoder with encoder " + }, + { + "bbox": [ + 304, + 308, + 545, + 548 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 304, + 308, + 545, + 548 + ], + "type": "text", + "content": " and decoder " + }, + { + "bbox": [ + 304, + 308, + 545, + 548 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 304, + 308, + 545, + 548 + ], + "type": "text", + "content": ". In the following section, we adapt the notation of [2]. A large U-Net, " + }, + { + "bbox": [ + 304, + 308, + 545, + 548 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}" + }, + { + "bbox": [ + 304, + 308, + 545, + 548 + ], + "type": "text", + "content": ", with 2.8 billion parameters, " + }, + { + "bbox": [ + 304, + 308, + 545, + 548 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 304, + 308, + 545, + 548 + ], + "type": "text", + "content": ", text embeddings from CLIP ViT-L [14] and T5-XXL [15], and a substantial pre-training dataset of 1.1 billion images facilitate the model's ability to learn complex semantics and finer details, with a noise-offset strategy contributing to high-contrast and aesthetically pleasing image generation. Given the encoded latent of an image " + }, + { + "bbox": [ + 304, + 308, + 545, + 548 + ], + "type": "inline_equation", + "content": "z = E(x)" + }, + { + "bbox": [ + 304, + 308, + 545, + 548 + ], + "type": "text", + "content": ", the diffusion process generates a noisy latent " + }, + { + "bbox": [ + 304, + 308, + 545, + 548 + ], + "type": "inline_equation", + "content": "z_{t}" + }, + { + "bbox": [ + 304, + 308, + 545, + 548 + ], + "type": "text", + "content": " where the noise level increases over timesteps " + }, + { + "bbox": [ + 304, + 308, + 545, + 548 + ], + "type": "inline_equation", + "content": "t \\in T" + }, + { + "bbox": [ + 304, + 308, + 545, + 548 + ], + "type": "text", + "content": ". To convert Emu to an instruction-based image editing model, we condition it on the image to be modified " + }, + { + "bbox": [ + 304, + 308, + 545, + 548 + ], + "type": "inline_equation", + "content": "c_{I}" + }, + { + "bbox": [ + 304, + 308, + 545, + 548 + ], + "type": "text", + "content": " and the instruction " + }, + { + "bbox": [ + 304, + 308, + 545, + 548 + ], + "type": "inline_equation", + "content": "c_{T}" + }, + { + "bbox": [ + 304, + 308, + 545, + 548 + ], + "type": "text", + "content": ". Emu Edit is trained to minimize the following optimization problem," + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 345, + 550, + 545, + 567 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 345, + 550, + 545, + 567 + ], + "spans": [ + { + "bbox": [ + 345, + 550, + 545, + 567 + ], + "type": "interline_equation", + "content": "\\min _ {\\theta} \\mathbb {E} _ {y, \\epsilon , t} \\left[ \\| \\epsilon - \\epsilon_ {\\theta} \\left(z _ {t}, t, E \\left(c _ {I}\\right), c _ {T}\\right) \\| _ {2} ^ {2} \\right] \\tag {1}", + "image_path": "866bf4fddb16fd43ce0df955d8a8bdeecd1b49ecf07415f5b70fd1f89277cfb3.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 570, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 547, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 570, + 547, + 714 + ], + "type": "inline_equation", + "content": "\\epsilon \\in N(0,1)" + }, + { + "bbox": [ + 304, + 570, + 547, + 714 + ], + "type": "text", + "content": " is the noise added by the diffusion process and " + }, + { + "bbox": [ + 304, + 570, + 547, + 714 + ], + "type": "inline_equation", + "content": "y = (c_{T},c_{I},x)" + }, + { + "bbox": [ + 304, + 570, + 547, + 714 + ], + "type": "text", + "content": " is a triplet of instruction, input image and target image from the dataset. In practice, we initialize the weights of Emu Edit with the weights of Emu. To support the image conditioning, we follow [2] and increase the number of input channels. New weights are initialized to zero. During inference, we perform classifier-free guidance on both image and text conditions. In our experiments we use a scale of " + }, + { + "bbox": [ + 304, + 570, + 547, + 714 + ], + "type": "inline_equation", + "content": "\\gamma_I = 1.5" + }, + { + "bbox": [ + 304, + 570, + 547, + 714 + ], + "type": "text", + "content": " for the image condition and " + }, + { + "bbox": [ + 304, + 570, + 547, + 714 + ], + "type": "inline_equation", + "content": "\\gamma_T = 5.0" + }, + { + "bbox": [ + 304, + 570, + 547, + 714 + ], + "type": "text", + "content": " for the text condition. Furthermore, we apply a rescaling of the diffusion scheduler to achieve a zero signal-to-noise ratio (SNR) at the terminal timestamp, as suggested" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8874" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 138, + 72, + 471, + 353 + ], + "blocks": [ + { + "bbox": [ + 109, + 97, + 138, + 145 + ], + "lines": [ + { + "bbox": [ + 109, + 97, + 138, + 145 + ], + "spans": [ + { + "bbox": [ + 109, + 97, + 138, + 145 + ], + "type": "text", + "content": "Make it play a rainbow colored trumpet" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 138, + 72, + 471, + 353 + ], + "lines": [ + { + "bbox": [ + 138, + 72, + 471, + 353 + ], + "spans": [ + { + "bbox": [ + 138, + 72, + 471, + 353 + ], + "type": "image", + "image_path": "0119e04a5b0b1c102feed425ba9973d27d070e0a1684db23b73de9d48474c9ea.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 214, + 354, + 398, + 363 + ], + "lines": [ + { + "bbox": [ + 214, + 354, + 398, + 363 + ], + "spans": [ + { + "bbox": [ + 214, + 354, + 398, + 363 + ], + "type": "text", + "content": "Local. Baseslines struggle to perform intricate local edits." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 158, + 368, + 434, + 380 + ], + "lines": [ + { + "bbox": [ + 158, + 368, + 434, + 380 + ], + "spans": [ + { + "bbox": [ + 158, + 368, + 434, + 380 + ], + "type": "text", + "content": "Figure 3. Failure cases of baseline instruction-based image editing models.2" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 387, + 288, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 387, + 288, + 422 + ], + "spans": [ + { + "bbox": [ + 46, + 387, + 288, + 422 + ], + "type": "text", + "content": "in [10]. This is crucial in order to avoid any mismatch between the model's training and testing phases. For more implementation details see supplementary Sec. ??" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 430, + 195, + 443 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 430, + 195, + 443 + ], + "spans": [ + { + "bbox": [ + 47, + 430, + 195, + 443 + ], + "type": "text", + "content": "4.2. Learned Task Embeddings" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 449, + 287, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 449, + 287, + 569 + ], + "spans": [ + { + "bbox": [ + 46, + 449, + 287, + 569 + ], + "type": "text", + "content": "To guide the generation process toward the correct task, we learn an embedding vector for each task in the dataset. During training, given a sample from our dataset, we use the task index, " + }, + { + "bbox": [ + 46, + 449, + 287, + 569 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 449, + 287, + 569 + ], + "type": "text", + "content": ", to fetch the task's embedding vector, " + }, + { + "bbox": [ + 46, + 449, + 287, + 569 + ], + "type": "inline_equation", + "content": "v_{i}" + }, + { + "bbox": [ + 46, + 449, + 287, + 569 + ], + "type": "text", + "content": ", from an embedding table, and optimize it jointly with the model weights. We do so by introducing the task embedding " + }, + { + "bbox": [ + 46, + 449, + 287, + 569 + ], + "type": "inline_equation", + "content": "v_{i}" + }, + { + "bbox": [ + 46, + 449, + 287, + 569 + ], + "type": "text", + "content": " as an additional condition to the U-Net, " + }, + { + "bbox": [ + 46, + 449, + 287, + 569 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}" + }, + { + "bbox": [ + 46, + 449, + 287, + 569 + ], + "type": "text", + "content": ". Concretely, we integrate the task embedding into the U-Net via crossattention interactions, and by adding it to the timestep embeddings. The optimization problem is updated to" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 71, + 571, + 287, + 590 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 71, + 571, + 287, + 590 + ], + "spans": [ + { + "bbox": [ + 71, + 571, + 287, + 590 + ], + "type": "interline_equation", + "content": "\\min _ {\\theta , v _ {1}, \\dots , v _ {k}} \\mathbb {E} _ {\\hat {y}, \\epsilon , t} \\left[ \\| \\epsilon - \\epsilon_ {\\theta} \\left(z _ {t}, t, E \\left(c _ {I}\\right), c _ {T}, v _ {i}\\right) \\| _ {2} ^ {2} \\right] \\tag {2}", + "image_path": "1f389d78618e933f854ec5048ff2f005be691e453b3a5120a43070dcfb3eb67a.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 592, + 287, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 592, + 287, + 676 + ], + "spans": [ + { + "bbox": [ + 46, + 592, + 287, + 676 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 592, + 287, + 676 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 592, + 287, + 676 + ], + "type": "text", + "content": " is the total number of tasks in our dataset and " + }, + { + "bbox": [ + 46, + 592, + 287, + 676 + ], + "type": "inline_equation", + "content": "\\hat{y} = (c_I, c_T, x, i)" + }, + { + "bbox": [ + 46, + 592, + 287, + 676 + ], + "type": "text", + "content": " is a quadruplet of input image, input instruction text, target image, and task index from the dataset. Task-specific conditioning arises from the observation that models lacking such conditioning can become perplexed about the type of edit required, particularly when the instructions are complex or the edit type is ambiguous. For" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 387, + 545, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 387, + 545, + 483 + ], + "spans": [ + { + "bbox": [ + 304, + 387, + 545, + 483 + ], + "type": "text", + "content": "instance, as visualized in Fig. 4, (1) a model without task conditioning might perform a global edit when a texture edit is required, (2) it might opt for segmentation when a global edit is necessary, and (3) it could implement a style edit in situations where a local edit would fit better. In the inference stage, we predict the task index. Specifically, we fine-tune a Flan-T5-XL model to identify the task at hand given the input instruction." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 492, + 397, + 504 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 492, + 397, + 504 + ], + "spans": [ + { + "bbox": [ + 305, + 492, + 397, + 504 + ], + "type": "text", + "content": "4.3. Task Inversion" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 511, + 545, + 606 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 511, + 545, + 606 + ], + "spans": [ + { + "bbox": [ + 304, + 511, + 545, + 606 + ], + "type": "text", + "content": "To enable few-shot learning of new tasks without losing the general abilities of Emu Edit, we propose a method for adapting the model without changing the U-Net weights. Given a few examples of a new task, we learn a new task embedding, " + }, + { + "bbox": [ + 304, + 511, + 545, + 606 + ], + "type": "inline_equation", + "content": "v_{\\mathrm{new}}" + }, + { + "bbox": [ + 304, + 511, + 545, + 606 + ], + "type": "text", + "content": ". We freeze the model weights, and adapt it to the task only through the task embedding. Thus, to fit a new task embedding we solve the following optimization problem:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 335, + 609, + 545, + 628 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 609, + 545, + 628 + ], + "spans": [ + { + "bbox": [ + 335, + 609, + 545, + 628 + ], + "type": "interline_equation", + "content": "\\min _ {v _ {\\text {n e w}}} \\mathbb {E} _ {y, \\epsilon , t} \\left[ \\| \\epsilon - \\epsilon_ {\\theta} \\left(z _ {t}, t, E \\left(c _ {I}\\right), c _ {T}, v _ {\\text {n e w}}\\right) \\| _ {2} ^ {2} \\right] \\tag {3}", + "image_path": "8173f83f8369ce1b908f9c0b91f1abecd6697fa5d4dfa41780077cdac788530c.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 629, + 545, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 629, + 545, + 652 + ], + "spans": [ + { + "bbox": [ + 304, + 629, + 545, + 652 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 629, + 545, + 652 + ], + "type": "inline_equation", + "content": "v_{\\mathrm{new}}" + }, + { + "bbox": [ + 304, + 629, + 545, + 652 + ], + "type": "text", + "content": " is the learned task embedding. Note that during task inversion " + }, + { + "bbox": [ + 304, + 629, + 545, + 652 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 304, + 629, + 545, + 652 + ], + "type": "text", + "content": " is a triplet belonging to the new task." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 654, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 545, + 712 + ], + "type": "text", + "content": "The model can then be employed for the new task by conditioning it on the learned task embedding, and it can still handle its original tasks by relying on the initial task embeddings. In Sec. 5.3, we demonstrate that our model effectively generalizes to novel tasks using this method." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 46, + 683, + 287, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 683, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 46, + 683, + 287, + 712 + ], + "type": "text", + "content": "2The samples depicted in this caption were selected by the authors. They do not cover all scenarios, but are meant to represent some common scenarios the authors encountered." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8875" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 44, + 70, + 286, + 297 + ], + "blocks": [ + { + "bbox": [ + 44, + 70, + 286, + 297 + ], + "lines": [ + { + "bbox": [ + 44, + 70, + 286, + 297 + ], + "spans": [ + { + "bbox": [ + 44, + 70, + 286, + 297 + ], + "type": "image", + "image_path": "c92be4727be869365485f5df5ea2528035c7666e91eb0d08737d1f4c7224804c.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 307, + 288, + 362 + ], + "lines": [ + { + "bbox": [ + 46, + 307, + 288, + 362 + ], + "spans": [ + { + "bbox": [ + 46, + 307, + 288, + 362 + ], + "type": "text", + "content": "Figure 4. Task embeddings. Model trained without task embeddings may get confused about the edit type when the instructions are complicated or there is ambiguity regarding the edit type: (1) Global edit (instead of Texture), (2) Segmentation (instead of Global), (3) Style edit (instead of Local)." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 367, + 287, + 499 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 367, + 287, + 499 + ], + "spans": [ + { + "bbox": [ + 46, + 367, + 287, + 499 + ], + "type": "text", + "content": "Sequential Edit Thresholding. We notice that applying the model repeatedly, in multi-turn editing scenarios, aggregates reconstruction and numerical errors, which translate to noticeable artifacts. To mitigate this, we add a per-pixel thresholding step after each edit-turn. At each step " + }, + { + "bbox": [ + 46, + 367, + 287, + 499 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 46, + 367, + 287, + 499 + ], + "type": "text", + "content": ", we use the pixel value in the output image, " + }, + { + "bbox": [ + 46, + 367, + 287, + 499 + ], + "type": "inline_equation", + "content": "c_{I}^{s+1}" + }, + { + "bbox": [ + 46, + 367, + 287, + 499 + ], + "type": "text", + "content": ", only if its alteration surpasses a specific threshold. Otherwise, we keep the pixel value from the input image, " + }, + { + "bbox": [ + 46, + 367, + 287, + 499 + ], + "type": "inline_equation", + "content": "c_{I}^{s}" + }, + { + "bbox": [ + 46, + 367, + 287, + 499 + ], + "type": "text", + "content": ". Specifically, given an edit turn " + }, + { + "bbox": [ + 46, + 367, + 287, + 499 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 46, + 367, + 287, + 499 + ], + "type": "text", + "content": ", we compute the absolute difference image " + }, + { + "bbox": [ + 46, + 367, + 287, + 499 + ], + "type": "inline_equation", + "content": "d = \\|c_{I}^{s+1} - c_{I}^{s}\\|_{1}" + }, + { + "bbox": [ + 46, + 367, + 287, + 499 + ], + "type": "text", + "content": " over the RGB channel, and apply the following thresholding:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 109, + 510, + 287, + 541 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 109, + 510, + 287, + 541 + ], + "spans": [ + { + "bbox": [ + 109, + 510, + 287, + 541 + ], + "type": "interline_equation", + "content": "c _ {I} ^ {s + 1} = \\left\\{ \\begin{array}{l l} c _ {I} ^ {s} & \\text {i f} \\bar {d} < \\alpha , \\\\ c _ {I} ^ {s + 1} & \\text {o t h e r w i s e .} \\end{array} \\right. \\tag {4}", + "image_path": "a26c349d27c7f50814537f4218e8f57c28f200c62150317a1e95b82859cc9fa4.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 548, + 287, + 609 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 548, + 287, + 609 + ], + "spans": [ + { + "bbox": [ + 46, + 548, + 287, + 609 + ], + "type": "text", + "content": "where, " + }, + { + "bbox": [ + 46, + 548, + 287, + 609 + ], + "type": "inline_equation", + "content": "\\bar{d}" + }, + { + "bbox": [ + 46, + 548, + 287, + 609 + ], + "type": "text", + "content": " is obtained after passing " + }, + { + "bbox": [ + 46, + 548, + 287, + 609 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 46, + 548, + 287, + 609 + ], + "type": "text", + "content": " through a low pass filter, in order to smooth the transition between previous and current pixels. In practice, we choose " + }, + { + "bbox": [ + 46, + 548, + 287, + 609 + ], + "type": "inline_equation", + "content": "\\alpha = 0.03" + }, + { + "bbox": [ + 46, + 548, + 287, + 609 + ], + "type": "text", + "content": ". Please refer to supplementary Sec. ?? for a qualitative comparison. In Fig. 2 we show examples of multi-turn editing." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 620, + 128, + 634 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 620, + 128, + 634 + ], + "spans": [ + { + "bbox": [ + 47, + 620, + 128, + 634 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 641, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 641, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 641, + 288, + 715 + ], + "type": "text", + "content": "Our experiments evaluate the ability of Emu Edit to follow user instructions faithfully and preserve the visual fidelity of the original image. First, we evaluate the performance of our approach on instruction-based image editing tasks. Second, we conduct a comprehensive ablation study to assess the effectiveness of our different contributions. Specifically," + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 72, + 546, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 546, + 156 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 546, + 156 + ], + "type": "text", + "content": "we ablate the contribution of the computer vision tasks to the model performance on image editing tasks, the importance of learned task embeddings, and the effect of multitask learning on instruction-based image editing. Further ablation on our data generation pipeline can be found in the supplementary Sec. ?? Finally, we demonstrate our model's ability to learn new tasks via few-shot learning." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 172, + 546, + 507 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 172, + 546, + 507 + ], + "spans": [ + { + "bbox": [ + 304, + 172, + 546, + 507 + ], + "type": "text", + "content": "Measures. We employ two main measures in our evaluation: edit text alignment and image faithfulness. Specifically, for each pair of input image and editing instruction, we use the following automatic metrics: (i) CLIP [14] text-image direction similarity " + }, + { + "bbox": [ + 304, + 172, + 546, + 507 + ], + "type": "inline_equation", + "content": "(\\mathrm{CLIP}_{dir})" + }, + { + "bbox": [ + 304, + 172, + 546, + 507 + ], + "type": "text", + "content": " - measuring agreement between change in captions and the change in images, (ii) CLIP image similarity " + }, + { + "bbox": [ + 304, + 172, + 546, + 507 + ], + "type": "inline_equation", + "content": "(\\mathrm{CLIP}_{img})" + }, + { + "bbox": [ + 304, + 172, + 546, + 507 + ], + "type": "text", + "content": " - measuring change between edited and input image, (iii) CLIP output similarity " + }, + { + "bbox": [ + 304, + 172, + 546, + 507 + ], + "type": "inline_equation", + "content": "(\\mathrm{CLIP}_{out})" + }, + { + "bbox": [ + 304, + 172, + 546, + 507 + ], + "type": "text", + "content": " - measuring edited image similarity with output caption, (iv) L1 pixel-distance between input and edit image, and (v) DINO [4] similarity between the DINO embeddings of input and edited images. With the exception of the L1 distance, where lower values indicate better performance, higher values in all other measures signify better results. A low L1 distance translates to small changes in image's pixel values. A high DINO and " + }, + { + "bbox": [ + 304, + 172, + 546, + 507 + ], + "type": "inline_equation", + "content": "\\mathrm{CLIP}_{img}" + }, + { + "bbox": [ + 304, + 172, + 546, + 507 + ], + "type": "text", + "content": " similarity score, suggests semantic similarity between the images. For region-based edits, high image similarity scores indicate the edits were precise. For free-form edits, high similarity scores indicate image structure preservation. " + }, + { + "bbox": [ + 304, + 172, + 546, + 507 + ], + "type": "inline_equation", + "content": "\\mathrm{CLIP}_{dir}" + }, + { + "bbox": [ + 304, + 172, + 546, + 507 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 172, + 546, + 507 + ], + "type": "inline_equation", + "content": "\\mathrm{CLIP}_{out}" + }, + { + "bbox": [ + 304, + 172, + 546, + 507 + ], + "type": "text", + "content": " measure how well the model followed the instruction. In addition, we asked human raters to evaluate the text alignment and image faithfulness. In each evaluation scenario, raters are presented with two modified images alongside the original input image and instruction, and are presented with two questions: (i) Image Faithfulness: which image better preserves elements in the input image, and (ii) Text Alignment: which image best follows the instruction." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 515, + 380, + 526 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 515, + 380, + 526 + ], + "spans": [ + { + "bbox": [ + 306, + 515, + 380, + 526 + ], + "type": "text", + "content": "5.1. Evaluation" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 533, + 545, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 533, + 545, + 665 + ], + "spans": [ + { + "bbox": [ + 304, + 533, + 545, + 665 + ], + "type": "text", + "content": "Throughout the paper, we report results on the MagicBrush test set [25] and the Emu Edit benchmark. In the following section, we describe our motivation for creating this benchmark, and detail its curation process. To date, there are two main benchmarks for evaluating instruction-based image editing capabilities. First, the InstructPix2Pix benchmark [2], which is intrinsically biased due to its reliance on generated Stable Diffusion [17] input images, and GPT-3 [3] generated instructions. Consequently, it is unclear whether its results will truly mirror the performance on real input images, with genuine user instructions." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 665, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 545, + 712 + ], + "type": "text", + "content": "Unlike InstructPix2Pix, the second benchmark, MagicBrush [25], uses a diverse set of authentic input images from the MS-COCO benchmark [5, 11], and annotator-defined instructions. Nonetheless, this dataset also suffers" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8876" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 52, + 101, + 541, + 200 + ], + "blocks": [ + { + "bbox": [ + 47, + 70, + 547, + 94 + ], + "lines": [ + { + "bbox": [ + 47, + 70, + 547, + 94 + ], + "spans": [ + { + "bbox": [ + 47, + 70, + 547, + 94 + ], + "type": "text", + "content": "Table 2. Comparison with image-editing baselines evaluated on Emu Edit test set and MagicBrush test set. For each benchmark we report CLIP, L1, DINO metrics and human ratings. Human evaluation shows the percentage of raters that prefer the results of Emu Edit." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 52, + 101, + 541, + 200 + ], + "lines": [ + { + "bbox": [ + 52, + 101, + 541, + 200 + ], + "spans": [ + { + "bbox": [ + 52, + 101, + 541, + 200 + ], + "type": "table", + "html": "
Emu Edit Test SetMagicBrush Test Set
MethodCLIPdir↑CLIPim↑CLIPout↑L1↓DINO↑Text align. faith.CLIPdir↑CLIPim↑CLIPout↑L1↓DINO↑Text align. faith.
InstructPix2Pix [2]0.0780.8340.2190.1210.76277.3376.710.1150.8370.2450.0930.767
MagicBrush [25]0.0900.8380.2220.1000.77674.5074.100.1230.8830.2610.0580.871
PnP [21]0.0280.5210.0890.3040.15398.9599.000.0250.5680.1010.2890.220
Null-Text Inv. [12]0.1010.7610.2360.0750.67881.6385.470.1210.7520.2630.0770.664
Our0.1090.8590.2310.0940.819--0.1350.8970.2610.0520.879
", + "image_path": "55413ddfd51fa795535859ef7be038787113fdfe1781e61c9ab6f56467788b82.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 205, + 288, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 205, + 288, + 277 + ], + "spans": [ + { + "bbox": [ + 46, + 205, + 288, + 277 + ], + "type": "text", + "content": "from inherent bias. During data collection, annotators were directed to use the DALLE-2 image editing platform [13] to generate the edited images. Thus, this benchmark is biased towards editing instructions that the DALLE-2 editor can successfully follow, which may compromise both its diversity and complexity." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 278, + 288, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 278, + 288, + 529 + ], + "spans": [ + { + "bbox": [ + 46, + 278, + 288, + 529 + ], + "type": "text", + "content": "Emu Edit Benchmark. To collect a dataset with reduced bias and of higher diversity, we take a different approach. We first define seven different categories of potential image editing operations: background alteration (Background), comprehensive image changes (Global), style alteration (Style), object removal (Remove), object addition (Add), localized modifications (Local), and color/texture alterations (Texture). Then, we utilize the diverse set of input images from the MagicBrush benchmark [25], and for each editing operation, we task crowd workers to devise relevant, creative, and challenging instructions. Moreover, to increase the quality of the collected examples, we apply a post-verification stage, in which crowd workers filter examples with irrelevant instructions. Finally, to support evaluation for methods that require input and output captions [8, 21], we additionally collect an input caption and output caption. When doing so, we ask annotators to ensure that the captions capture both important elements in the image, and elements that should change based on the instruction. See supplementary Sec. ?? for examples of our benchmark, which we release to support better evaluation." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 533, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 533, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 533, + 287, + 714 + ], + "type": "text", + "content": "Baseline Comparisons. We compare our model against two instruction-based image editing baseline models: InstructPix2Pix [2], and MagicBrush [25], which is a variant of InstructPix2Pix that was fine-tuned on the MagicBrush dataset. Additionally, we compare our model against two text-based image editing methods: PNP [21] and Null-Text Inversion modification of P2P [8, 12]. Unlike instruction-based models, these works expect image descriptions. Therefore, we provide them with access to the input caption and output caption. Note however, providing these methods with access to the ground-truth captions could potentially offer an advantage over instruction-based models, since the automatic metrics also rely on these captions. Tab. 2 shows our results versus the baselines. The findings indicate that human raters consistently prefer Emu" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 205, + 545, + 301 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 205, + 545, + 301 + ], + "spans": [ + { + "bbox": [ + 304, + 205, + 545, + 301 + ], + "type": "text", + "content": "Edit over all baselines by a large margin. Furthermore, apart from Null-Text Inversion, which as explained above, utilizes the ground-truth captions during inference, our approach outperforms the existing baselines on the automatic metrics. We provide qualitative comparisons in Fig. 3. Additional comparisons are available in the supplementary: Fig.??-??, and Fig.??-??. For performance on vision tasks, see Sec. ??" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 308, + 373, + 319 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 308, + 373, + 319 + ], + "spans": [ + { + "bbox": [ + 306, + 308, + 373, + 319 + ], + "type": "text", + "content": "5.2. Ablations" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 326, + 545, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 326, + 545, + 506 + ], + "spans": [ + { + "bbox": [ + 304, + 326, + 545, + 506 + ], + "type": "text", + "content": "Computer Vision Tasks Enhance Image Editing Tasks. Here we demonstrate the importance of the vision tasks to Emu Edit performance on image editing tasks. For this, we trained two additional models on all tasks except: (i) detect and segment tasks, and (ii) image-to-image translation tasks. As we show in Tab. 4, adding the detection and segmentation tasks improves the model performance in region-based editing tasks. Additionally, we observe that image-to-image translation tasks improve the performance in free-form editing tasks. We hypothesize that the recognition tasks improve the model's recognition capabilities, leading to more accurate and precise localized modifications. Similarly, image-to-image tasks assist the model in understanding the entire image structure, thereby enhancing its capabilities for global operations." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 506, + 545, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 506, + 545, + 696 + ], + "spans": [ + { + "bbox": [ + 304, + 506, + 545, + 696 + ], + "type": "text", + "content": "Contribution of Learned Task Embeddings. We compare three variants of Emu Edit: (i) conditioned on the ground-truth task embedding, (ii) conditioned on the task embedding, as predicted by the task predictor described in Sec. 4.2, and (iii) without conditioning on the task type.3 Tab. 3 shows the results on the validation set of our benchmark. As can be seen, conditioning on the task type boosts the model's performance. Furthermore, our task predictor closes the gap with the ground-truth conditioned model. Qualitatively, we observe that without conditioning on the task type the model may perform the wrong editing operation (Fig. 4). In supplementary Fig. ??, we demonstrate the effect of manipulating the task while keeping the instruction and input image fixed. As can be seen, changing the task embedding directly influences the task executed by the model." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 315, + 702, + 482, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 702, + 482, + 712 + ], + "spans": [ + { + "bbox": [ + 315, + 702, + 482, + 712 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 315, + 702, + 482, + 712 + ], + "type": "text", + "content": " (iii) was trained without learned task embeddings." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8877" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 58, + 108, + 277, + 161 + ], + "blocks": [ + { + "bbox": [ + 47, + 71, + 287, + 104 + ], + "lines": [ + { + "bbox": [ + 47, + 71, + 287, + 104 + ], + "spans": [ + { + "bbox": [ + 47, + 71, + 287, + 104 + ], + "type": "text", + "content": "Table 3. Learned task embeddings ablation on our validation set. We compare variations of Emu Edit: without task type condition, with predicted task type, and with ground-truth task type." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 108, + 277, + 161 + ], + "lines": [ + { + "bbox": [ + 58, + 108, + 277, + 161 + ], + "spans": [ + { + "bbox": [ + 58, + 108, + 277, + 161 + ], + "type": "table", + "html": "
MethodCLIPdir↑CLIPim↑CLIPout↑L1↓DINO↑
w/o task emb.0.1040.8430.2270.1090.792
with pred. task0.1170.8500.2310.1030.809
with gt task0.1190.8520.2310.1000.811
", + "image_path": "ebfc35572c3933b3b9c1758b5151d2d8f738422fda0528692d61750c3af3b47c.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 72, + 191, + 265, + 259 + ], + "blocks": [ + { + "bbox": [ + 47, + 166, + 286, + 188 + ], + "lines": [ + { + "bbox": [ + 47, + 166, + 286, + 188 + ], + "spans": [ + { + "bbox": [ + 47, + 166, + 286, + 188 + ], + "type": "text", + "content": "Table 4. Contribution of computer vision tasks. Human evaluation is shown as a percentage of majority votes in favor of our model." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 72, + 191, + 265, + 259 + ], + "lines": [ + { + "bbox": [ + 72, + 191, + 265, + 259 + ], + "spans": [ + { + "bbox": [ + 72, + 191, + 265, + 259 + ], + "type": "table", + "html": "
Region-BasedFree-form
MethodText align.Image faith.Text align.Image faith.
without detect/segment60.060.252.351.5
without im2im translation50.249.058.060.1
", + "image_path": "22fe06395b2e6e58c3c9786971b2175a3f3147a988a86fc14b95c9e96f9ae557.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 107, + 269, + 179, + 342 + ], + "blocks": [ + { + "bbox": [ + 70, + 269, + 108, + 329 + ], + "lines": [ + { + "bbox": [ + 70, + 269, + 108, + 329 + ], + "spans": [ + { + "bbox": [ + 70, + 269, + 108, + 329 + ], + "type": "text", + "content": "(i) Incorporate a bee into the bag's pattern and detect it" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 107, + 269, + 179, + 342 + ], + "lines": [ + { + "bbox": [ + 107, + 269, + 179, + 342 + ], + "spans": [ + { + "bbox": [ + 107, + 269, + 179, + 342 + ], + "type": "image", + "image_path": "39123bad4f47e58ef3229c1e7a2b69267c6105a096a1c393a5d87bbbf0acd560.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 179, + 269, + 250, + 342 + ], + "blocks": [ + { + "bbox": [ + 179, + 269, + 250, + 342 + ], + "lines": [ + { + "bbox": [ + 179, + 269, + 250, + 342 + ], + "spans": [ + { + "bbox": [ + 179, + 269, + 250, + 342 + ], + "type": "image", + "image_path": "550332dd0b724988cbe59bd88711563be1cac0be2f343cc947851469b953494d.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 106, + 353, + 178, + 426 + ], + "blocks": [ + { + "bbox": [ + 75, + 369, + 104, + 406 + ], + "lines": [ + { + "bbox": [ + 75, + 369, + 104, + 406 + ], + "spans": [ + { + "bbox": [ + 75, + 369, + 104, + 406 + ], + "type": "text", + "content": "(ii) Mark the gift bags" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 106, + 353, + 178, + 426 + ], + "lines": [ + { + "bbox": [ + 106, + 353, + 178, + 426 + ], + "spans": [ + { + "bbox": [ + 106, + 353, + 178, + 426 + ], + "type": "image", + "image_path": "28a809ddf6b49196190fc05506c4b8702c52aba451060a4af490f8cb3a107582.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 426, + 154, + 436 + ], + "lines": [ + { + "bbox": [ + 130, + 426, + 154, + 436 + ], + "spans": [ + { + "bbox": [ + 130, + 426, + 154, + 436 + ], + "type": "text", + "content": "Input" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 441, + 287, + 464 + ], + "lines": [ + { + "bbox": [ + 46, + 441, + 287, + 464 + ], + "spans": [ + { + "bbox": [ + 46, + 441, + 287, + 464 + ], + "type": "text", + "content": "Figure 5. Generations on unseen tasks with task inversion. (i) composition of add and detect tasks, (ii) object contour detection." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 179, + 353, + 250, + 426 + ], + "blocks": [ + { + "bbox": [ + 179, + 353, + 250, + 426 + ], + "lines": [ + { + "bbox": [ + 179, + 353, + 250, + 426 + ], + "spans": [ + { + "bbox": [ + 179, + 353, + 250, + 426 + ], + "type": "image", + "image_path": "92491dca5918554091ff7fed3d219e548fc2bb2af9abd1eb4159a98124759791.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 194, + 426, + 234, + 436 + ], + "lines": [ + { + "bbox": [ + 194, + 426, + 234, + 436 + ], + "spans": [ + { + "bbox": [ + 194, + 426, + 234, + 436 + ], + "type": "text", + "content": "Emu Edit" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 471, + 287, + 542 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 471, + 287, + 542 + ], + "spans": [ + { + "bbox": [ + 46, + 471, + 287, + 542 + ], + "type": "text", + "content": "Multi-Task versus Expert Models. We show that training a single model on a diverse range of tasks, leads to enhanced performance in each individual task, outperforming models that are specifically trained for a single task. To validate this, we train an expert model for each task, and compare its performance to the single one (See supplementary Tab. ??)." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 46, + 542, + 287, + 638 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 542, + 287, + 638 + ], + "spans": [ + { + "bbox": [ + 46, + 542, + 287, + 638 + ], + "type": "text", + "content": "Influence of Number of Tasks. Here, we ablate the number of tasks participating in the multi-task training scheme. In supplementary Fig. ?? we report the average " + }, + { + "bbox": [ + 46, + 542, + 287, + 638 + ], + "type": "inline_equation", + "content": "\\mathrm{CLIP}_{dir}" + }, + { + "bbox": [ + 46, + 542, + 287, + 638 + ], + "type": "text", + "content": " on the Style and Texture tasks when iteratively excluding other tasks, and training a model on the new tasks list. As can be seen, augmenting the model with additional tasks leads to improved performance, even in tasks which are not directly associated with the added ones." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 47, + 647, + 224, + 659 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 647, + 224, + 659 + ], + "spans": [ + { + "bbox": [ + 47, + 647, + 224, + 659 + ], + "type": "text", + "content": "5.3. Few-Shot Learning of New Tasks" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "text", + "content": "Finally, we evaluate our model's ability to generalize by testing it in a few-shot scenario with previously unseen tasks. We test its performance across the following tasks: (i) super-resolution (x4), (ii) object contour detection, (iii)" + } + ] + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 299, + 85, + 413, + 168 + ], + "blocks": [ + { + "bbox": [ + 299, + 85, + 413, + 168 + ], + "lines": [ + { + "bbox": [ + 299, + 85, + 413, + 168 + ], + "spans": [ + { + "bbox": [ + 299, + 85, + 413, + 168 + ], + "type": "image", + "image_path": "f4e4f96d972ec08ff045d7ace8a5e94898ace017418c3deb72d8af9abf0a0aea.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 333, + 175, + 379, + 185 + ], + "lines": [ + { + "bbox": [ + 333, + 175, + 379, + 185 + ], + "spans": [ + { + "bbox": [ + 333, + 175, + 379, + 185 + ], + "type": "text", + "content": "(a) Inpainting" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 304, + 194, + 545, + 249 + ], + "lines": [ + { + "bbox": [ + 304, + 194, + 545, + 249 + ], + "spans": [ + { + "bbox": [ + 304, + 194, + 545, + 249 + ], + "type": "text", + "content": "Figure 6. Few-shot performance for different tasks over 1, 10, and 100 samples. Each line represents a different training setting: 'Scratch' finetune (Blue, " + }, + { + "bbox": [ + 304, + 194, + 545, + 249 + ], + "type": "inline_equation", + "content": "\\bigcirc" + }, + { + "bbox": [ + 304, + 194, + 545, + 249 + ], + "type": "text", + "content": "), Emu Edit finetune (Orange, " + }, + { + "bbox": [ + 304, + 194, + 545, + 249 + ], + "type": "inline_equation", + "content": "\\square" + }, + { + "bbox": [ + 304, + 194, + 545, + 249 + ], + "type": "text", + "content": "), task inversion (Green, " + }, + { + "bbox": [ + 304, + 194, + 545, + 249 + ], + "type": "inline_equation", + "content": "\\triangle" + }, + { + "bbox": [ + 304, + 194, + 545, + 249 + ], + "type": "text", + "content": "), all compared to an upper-bound expert trained on 100k samples (Red dashed line, " + }, + { + "bbox": [ + 304, + 194, + 545, + 249 + ], + "type": "inline_equation", + "content": "\\diamond" + }, + { + "bbox": [ + 304, + 194, + 545, + 249 + ], + "type": "text", + "content": ")." + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 419, + 85, + 533, + 168 + ], + "blocks": [ + { + "bbox": [ + 419, + 85, + 533, + 168 + ], + "lines": [ + { + "bbox": [ + 419, + 85, + 533, + 168 + ], + "spans": [ + { + "bbox": [ + 419, + 85, + 533, + 168 + ], + "type": "image", + "image_path": "d7be06768fe65bd0afeb7e4b6fe4566b2c930661aae31bb2c587ffc2c786595f.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 448, + 175, + 504, + 184 + ], + "lines": [ + { + "bbox": [ + 448, + 175, + 504, + 184 + ], + "spans": [ + { + "bbox": [ + 448, + 175, + 504, + 184 + ], + "type": "text", + "content": "(b) Add + Detect" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 257, + 545, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 257, + 545, + 544 + ], + "spans": [ + { + "bbox": [ + 304, + 257, + 545, + 544 + ], + "type": "text", + "content": "mask-based inpainting, and (iv) a composite task formed by combining two tasks from our dataset: add and detect. For each task, we assess the model's performance when trained with 0, 1, and 100 examples, with the 0-example case being equivalent to a zero-shot setting. We compare three baselines: (i) Scratch - Emu Edit initialized with Emu's weights, trained on the examples, (ii) Task Inversion - Emu Edit with task inversion (Sec. 4.3), and (iii) Finetune - Emu Edit where we finetune all of the model's weights on the examples. As an upper-bound expert, we train the first baseline on 100,000 examples. Note that, the \"Task Inversion\" and \"Finetune\" baselines were trained on the multitask dataset whereas the \"Scratch\" baseline was not. As can be seen in Fig. 6, fine-tuning with a single example is enough to significantly enhance the performance, while training from scratch results in overfitting. Moreover, utilizing 100 samples nearly achieves expert-level performance, implying that the model can effectively generalize to novel tasks. We also observe that task inversion performs similarly to complete fine-tuning, indicating that the model already contains essential information and can be prompted with a new task embedding to yield the desired result. For performance and generation examples on additional tasks see Fig. 5 and supplementary ????." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 306, + 549, + 378, + 561 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 549, + 378, + 561 + ], + "spans": [ + { + "bbox": [ + 306, + 549, + 378, + 561 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "text", + "content": "Emu Edit presents a step change in instructable image editing capabilities, primarily due to its unique training on both recognition and generation tasks. This dual-focus approach significantly enhances the model's comprehension of natural language instructions, enabling it to accurately execute a wide array of editing operations. Its ability to generalize to new tasks with minimal examples further demonstrates its versatility and robustness. Furthermore, our framework has the potential for further integration with a multimodal LLM in future projects, particularly for complex editing tasks requiring detailed reasoning from the input image, such as counting objects or executing intricate tasks." + } + ] + } + ], + "index": 24 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8878" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 135 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 135 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 135 + ], + "type": "text", + "content": "[1] Omri Avrahami, Dani Lischinski, and Ohad Fried. Blended diffusion for text-driven editing of natural images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18208-18218, 2022. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 137, + 287, + 158 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 137, + 287, + 158 + ], + "spans": [ + { + "bbox": [ + 53, + 137, + 287, + 158 + ], + "type": "text", + "content": "[2] Tim Brooks et al. Instructpix2pix: Learning to follow image editing instructions. In CVPR, 2023. 1, 3, 4, 6, 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 160, + 288, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 160, + 288, + 270 + ], + "spans": [ + { + "bbox": [ + 53, + 160, + 288, + 270 + ], + "type": "text", + "content": "[3] Tom B. Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-Voss, Gretchen Krueger, T. J. Henighan, Rewon Child, Aditya Ramesh, Daniel M. Ziegler, Jeff Wu, Clemens Winter, Christopher Hesse, Mark Chen, Eric Sigler, Mateusz Litwin, Scott Gray, Benjamin Chess, Jack Clark, Christopher Berner, Sam McCandlish, Alec Radford, Ilya Sutskever, and Dario Amodei. Language models are few-shot learners. ArXiv, abs/2005.14165, 2020. 3, 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 271, + 287, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 271, + 287, + 327 + ], + "spans": [ + { + "bbox": [ + 53, + 271, + 287, + 327 + ], + "type": "text", + "content": "[4] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9650-9660, 2021. 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 328, + 287, + 371 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 328, + 287, + 371 + ], + "spans": [ + { + "bbox": [ + 53, + 328, + 287, + 371 + ], + "type": "text", + "content": "[5] Xinlei Chen, Hao Fang, Tsung-Yi Lin, Ramakrishna Vedantam, Saurabh Gupta, Piotr Dólar, and C. Lawrence Zittnick. Microsoft coco captions: Data collection and evaluation server. ArXiv, abs/1504.00325, 2015. 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 373, + 287, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 373, + 287, + 430 + ], + "spans": [ + { + "bbox": [ + 53, + 373, + 287, + 430 + ], + "type": "text", + "content": "[6] Xiaoliang Dai, Ji Hou, Chih-Yao Ma, Sam Tsai, Jialiang Wang, Rui Wang, Peizhao Zhang, Simon Vandenhende, Xiaofang Wang, Abhimanyu Dubey, et al. Emu: Enhancing image generation models using photogenic needles in a haystack. arXiv preprint arXiv:2309.15807, 2023. 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 430, + 287, + 485 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 430, + 287, + 485 + ], + "spans": [ + { + "bbox": [ + 53, + 430, + 287, + 485 + ], + "type": "text", + "content": "[7] Oran Gafni, Adam Polyak, Oron Ashual, Shelly Sheynin, Devi Parikh, and Yaniv Taigman. Make-a-scene: Scene-based text-to-image generation with human priors. In European Conference on Computer Vision, pages 89-106. Springer, 2022. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 487, + 287, + 531 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 487, + 287, + 531 + ], + "spans": [ + { + "bbox": [ + 53, + 487, + 287, + 531 + ], + "type": "text", + "content": "[8] Amir Hertz, Ron Mokady, Jay M. Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-prompt image editing with cross attention control. ArXiv, abs/2208.01626, 2022. 2, 3, 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 533, + 287, + 576 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 533, + 287, + 576 + ], + "spans": [ + { + "bbox": [ + 53, + 533, + 287, + 576 + ], + "type": "text", + "content": "[9] Bahjat Kawar et al. Imagic: Text-based real image editing with diffusion models. 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6007-6017, 2022. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 578, + 287, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 578, + 287, + 611 + ], + "spans": [ + { + "bbox": [ + 48, + 578, + 287, + 611 + ], + "type": "text", + "content": "[10] Shanchuan Lin, Bingchen Liu, Jiashi Li, and Xiao Yang. Common diffusion noise schedules and sample steps are flawed. arXiv preprint arXiv:2305.08891, 2023. 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 613, + 287, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 613, + 287, + 656 + ], + "spans": [ + { + "bbox": [ + 48, + 613, + 287, + 656 + ], + "type": "text", + "content": "[11] Tsung-Yi Lin, Michael Maire, Serge J. Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C. Lawrence Zitnick. Microsoft coco: Common objects in context. In ECCV, 2014. 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 658, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 287, + 712 + ], + "type": "text", + "content": "[12] Ron Mokady, Amir Hertz, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Null-text inversion for editing real images using guided diffusion models. 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6038–6047, 2022. 3, 7" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 94 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 94 + ], + "type": "text", + "content": "[13] OpenAI. Dall-e 2, 2022. https://openai.com/product/dall-e-2.3,7" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 95, + 545, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 95, + 545, + 160 + ], + "spans": [ + { + "bbox": [ + 307, + 95, + 545, + 160 + ], + "type": "text", + "content": "[14] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 1, 4, 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 161, + 545, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 161, + 545, + 216 + ], + "spans": [ + { + "bbox": [ + 308, + 161, + 545, + 216 + ], + "type": "text", + "content": "[15] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J. Liu. Exploring the limits of transfer learning with a unified text-to-text transformer. Journal of Machine Learning Research, 21(140):1-67, 2020. 4" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 217, + 545, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 217, + 545, + 258 + ], + "spans": [ + { + "bbox": [ + 308, + 217, + 545, + 258 + ], + "type": "text", + "content": "[16] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 260, + 545, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 260, + 545, + 315 + ], + "spans": [ + { + "bbox": [ + 308, + 260, + 545, + 315 + ], + "type": "text", + "content": "[17] Robin Rombach, A. Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10674-10685, 2021. 2, 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 316, + 545, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 316, + 545, + 348 + ], + "spans": [ + { + "bbox": [ + 308, + 316, + 545, + 348 + ], + "type": "text", + "content": "[18] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models, 2021. 4" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 350, + 545, + 425 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 350, + 545, + 425 + ], + "spans": [ + { + "bbox": [ + 308, + 350, + 545, + 425 + ], + "type": "text", + "content": "[19] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L. Denton, Seyed Kamyar Seyed Ghasemipour, Burcu Karagol Ayan, Seyedeh Sara Mahdavi, Raphael Gontijo Lopes, Tim Salimans, Jonathan Ho, David J. Fleet, and Mohammad Norouzi. Photorealistic text-to-image diffusion models with deep language understanding. ArXiv, abs/2205.11487, 2022. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 426, + 545, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 426, + 545, + 480 + ], + "spans": [ + { + "bbox": [ + 308, + 426, + 545, + 480 + ], + "type": "text", + "content": "[20] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288, 2023. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 482, + 545, + 535 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 482, + 545, + 535 + ], + "spans": [ + { + "bbox": [ + 308, + 482, + 545, + 535 + ], + "type": "text", + "content": "[21] Narek Tumanyan, Michal Geyer, Shai Bagon, and Tali Dekel. Plug-and-play diffusion features for text-driven image-to-image translation. 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1921-1930, 2022. 3, 7" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 536, + 545, + 581 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 536, + 545, + 581 + ], + "spans": [ + { + "bbox": [ + 308, + 536, + 545, + 581 + ], + "type": "text", + "content": "[22] Bram Wallace, Akash Gokul, and Nikhil Vijay Naik. Edict: Exact diffusion inversion via coupled transformations. 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 22532-22541, 2022. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 582, + 545, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 582, + 545, + 624 + ], + "spans": [ + { + "bbox": [ + 308, + 582, + 545, + 624 + ], + "type": "text", + "content": "[23] Su Wang et al. Imagen editor and editbench: Advancing and evaluating text-guided image inpainting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18359-18369, 2023. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 625, + 545, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 625, + 545, + 678 + ], + "spans": [ + { + "bbox": [ + 308, + 625, + 545, + 678 + ], + "type": "text", + "content": "[24] Shaoan Xie, Zhifei Zhang, Zhe Lin, Tobias Hinz, and Kun Zhang. Smartbrush: Text and shape guided object inpainting with diffusion model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22428-22437, 2023. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 680, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 680, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 308, + 680, + 545, + 712 + ], + "type": "text", + "content": "[25] Kai Zhang et al. Magicbrush: A manually annotated dataset for instruction-guided image editing. arXiv:2306.10012, 2023. 1, 3, 6, 7" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8879" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/En3D_ An Enhanced Generative Model for Sculpting 3D Humans from 2D Synthetic Data/cb7c8896-3bc3-4bc4-a3cb-b2aedebff579_content_list.json b/2024/En3D_ An Enhanced Generative Model for Sculpting 3D Humans from 2D Synthetic Data/cb7c8896-3bc3-4bc4-a3cb-b2aedebff579_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..0fa178e83d66b8a2fbbeaff5cc672c5acceaddb5 --- /dev/null +++ b/2024/En3D_ An Enhanced Generative Model for Sculpting 3D Humans from 2D Synthetic Data/cb7c8896-3bc3-4bc4-a3cb-b2aedebff579_content_list.json @@ -0,0 +1,1560 @@ +[ + { + "type": "text", + "text": "En3D: An Enhanced Generative Model for Sculpting 3D Humans from 2D Synthetic Data", + "text_level": 1, + "bbox": [ + 127, + 128, + 843, + 174 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yifang Men1, Biwen Lei1, Yuan Yao1, Miaomiao Cui1, Zhouhui Lian2, Xuansong Xie1", + "bbox": [ + 142, + 202, + 825, + 222 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Institute for Intelligent Computing, Alibaba Group", + "bbox": [ + 258, + 223, + 709, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ Wangxuan Institute of Computer Technology, Peking University, China", + "bbox": [ + 184, + 241, + 785, + 257 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "https://menyifang.github.io/projects/En3D/index.html", + "bbox": [ + 253, + 258, + 712, + 273 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/983670dcb4b85364dad9a3570384c18c010d456f295654df6337a5da0fd591ba.jpg", + "image_caption": [ + "(a)", + "(b)" + ], + "image_footnote": [], + "bbox": [ + 86, + 311, + 537, + 433 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "\"a 44-year-old Latino man with blonde side part haircut, wearing a bomber jacket with brooch\"", + "bbox": [ + 584, + 295, + 844, + 318 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/76e7ccab3e14de71b3795b0179e19c701cf0088e680e8c398646ea720bf08bde.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 552, + 321, + 874, + 431 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/402cb4a2fea8a39cad637b91faa73ad3935b124bcac012ab790e2cdccdbc4e8b.jpg", + "image_caption": [ + "(c)", + "Figure 1. Given random noises or guided texts, our generative scheme can synthesize high-fidelity 3D human avatars that are visually realistic and geometrically accurate. These avatars can be seamlessly animated and easily edited. Our model is trained on 2D synthetic data without directly relying on any pre-existing 3D or 2D collections." + ], + "image_footnote": [], + "bbox": [ + 99, + 435, + 488, + 547 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/b39a44ea2412c7c1054948a143823a1c704380e2114d9e8e16d8734ec5d15d7d.jpg", + "image_caption": [ + "(d)", + "(e)" + ], + "image_footnote": [], + "bbox": [ + 521, + 436, + 709, + 547 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/1f26002baed92f7862f31bd212cc40f3e1470aa0d9f7064e46d2a173cd05344f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 712, + 439, + 888, + 547 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 604, + 313, + 622 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We present En3D, an enhanced generative scheme for sculpting high-quality 3D human avatars. Unlike previous works that rely on scarce 3D datasets or limited 2D collections with imbalanced viewing angles and imprecise pose priors, our approach aims to develop a zero-shot 3D generative scheme capable of producing visually realistic, geometrically accurate and content-wise diverse 3D humans without directly relying on pre-existing 3D or 2D assets. To address this challenge, we introduce a meticulously crafted workflow that implements accurate physical modeling to learn the enhanced 3D generative model from synthetic 2D data. During inference, we integrate optimization modules to bridge the gap between realistic appearances and coarse 3D shapes. Specifically, En3D comprises three modules: a 3D generator that accurately models generalizable 3D humans with realistic appearance from synthesized balanced, diverse, and structured human images; a geometry sculptor", + "bbox": [ + 75, + 643, + 473, + 902 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "that enhances shape quality using multi-view normal constraints for intricate human structure; and a texturing module that disentangles explicit texture maps with fidelity and editability, leveraging semantical UV partitioning and a differentiable rasterizer. Experimental results show that our approach significantly outperforms prior works in terms of image quality, geometry accuracy and content diversity. We also showcase the applicability of our generated avatars for animation and editing, as well as the scalability of our approach for content-style free adaptation.", + "bbox": [ + 496, + 607, + 895, + 758 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 500, + 784, + 632, + 800 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3D human avatars play an important role in various applications of AR/VR such as video games, telepresence and virtual try-on. Realistic human modeling is an essential task, and many valuable efforts have been made by leveraging neural implicit fields to learn high-quality articulated avatars [8, 10, 41, 48]. However, these methods are directly", + "bbox": [ + 496, + 809, + 893, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "9981", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "learned from monocular videos or image sequences, where subjects are single individuals wearing specific garments, thus limiting their scalability.", + "bbox": [ + 75, + 90, + 468, + 136 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Generative models learn a shared 3D representation to synthesize clothed humans with varying identities, clothing and poses. Traditional methods are typically trained on 3D datasets, which are limited and expensive to acquire. This data scarcity limits the model's generalization ability and may lead to overfitting on small datasets. Recently, 3D-aware image synthesis methods [5, 18, 37] have demonstrated great potential in learning 3D generative models of rigid objects from 2D image collections. Follow-up works show the feasibility of learning articulated humans from image collections driven by SMPL-based deformations, but only in limited quality and resolution. EVA3D [15] represents humans as a composition of multiple parts with NeRF representations. AG3D [9] incorporates an efficient articulation module to capture both body shape and cloth deformation. Nevertheless, there remains a noticeable gap between generated and real humans in terms of appearance and geometry. Moreover, their results are limited to specific views (i.e., frontal angles) and lack diversity (i.e., fashion images in similar skin tone, body shape, and age).", + "bbox": [ + 75, + 138, + 470, + 441 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The aim of this paper is to propose a zero-shot 3D generative scheme that does not directly rely on any pre-existing 3D or 2D datasets, yet is capable of producing high-quality 3D humans that are visually realistic, geometrically accurate, and content-wise diverse. The generated avatars can be seamlessly animated and easily edited. An illustration is provided in Figure 1. To address this challenging task, our proposed method inherits from 3D-aware human image synthesis and exhibits substantial distinctions based on several key insights. Rethinking the nature of 3D-aware generative methods from 2D collections [5, 9, 15], they actually try to learn a generalizable and deformable 3D representation, whose 2D projections can meet the distribution of human images in corresponding views. Thereby, it is crucial for accurate physical modeling between 3D objects and 2D projections. However, previous works typically leverage pre-existing 2D human images to estimate physical parameters (i.e., camera and body poses), which are inaccurate because of imprecise SMPL priors for highly-articulated humans. This inaccuracy limits the synthesis ability for realistic multi-view renderings. Second, these methods solely rely on discriminating 2D renderings, which is ambiguous and loose to capture inherent 3D shapes in detail, especially for intricate human appearance.", + "bbox": [ + 75, + 444, + 468, + 806 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address these limitations, we propose a novel generative scheme with two core designs. Firstly, we introduce a meticulously-crafted workflow that implements accurate physical modeling to learn an enhanced 3D generative model from synthetic data. This is achieved by instantiating a 3D body scene and projecting the underlying", + "bbox": [ + 75, + 810, + 468, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3D skeletons into 2D pose images using explicit camera parameters. These 2D pose images act as conditions to control a 2D diffusion model, synthesizing realistic human images from specific viewpoints. By leveraging synthetic view-balanced, diverse and structured human images, along with known physical parameters, we employ a 3D generator equipped with an enhanced renderer and discriminator to learn realistic appearance modeling. Secondly, we improve the 3D shape quality by leveraging the gap between high-quality multi-view renderings and the coarse mesh produced by the 3D generative module. Specifically, we integrate an optimization module that utilizes multi-view normal constraints to rapidly refine geometry details under supervision. Additionally, we incorporate an explicit texturing module to ensure faithful UV texture maps. In contrast to previous works that rely on inaccurate physical settings and inadequate shape supervision, we rebuild the generative scheme from the ground up, resulting in comprehensive improvements in image quality, geometry accuracy, and content diversity. In summary, our contributions are threefold:", + "bbox": [ + 496, + 90, + 890, + 393 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We present a zero-shot generative scheme that efficiently synthesizes high-quality 3D human avatars with visual realism, geometric accuracy and content diversity. These avatars can be seamlessly animated and easily edited, offering greater flexibility in their applications.", + "- We develop a meticulously-crafted workflow to learn an enhanced generative model from synthesized human images that are balanced, diverse, and also possess known physical parameters. This leads to diverse 3D-aware human image synthesis with realistic appearance.", + "- We propose to integrate optimization modules into the 3D generator, leveraging multi-view guidance to enhance both shape quality and texture fidelity, thus achieving realistic 3D human assets." + ], + "bbox": [ + 500, + 393, + 890, + 603 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related work", + "text_level": 1, + "bbox": [ + 500, + 618, + 635, + 633 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3D Human Modeling. Parametric models [3, 19, 20, 29, 38] serve as a common representation for 3D human modeling, they allows for robust control by deforming a template mesh with a series of low-dimensional parameters, but can only generate naked 3D humans. Similar ideas have been extended to model clothed humans [2, 31], but geometric expressivity is restricted due to the fixed mesh topology. Subsequent works [6, 39, 39] further introduce implicit surfaces to produce complex non-linear deformations of 3D bodies. Unfortunately, the aforementioned approaches all require 3D scans of various human poses for model fitting, which are difficult to acquire. With the explosion of NeRF, valuable efforts have been made towards combining NeRF models with explicit human models [8, 10, 28, 41, 48]. Neural body [41] anchors a set of latent codes to the vertices of the SMPL model [29] and transforms the spatial locations of the codes to the volume in the observation space. Human-", + "bbox": [ + 496, + 643, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "9982", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/7d51faa41e042c861d61716c00ef486d05b132f800883da5657afb6ca6e7ccd4.jpg", + "image_caption": [ + "Figure 2. An overview of the proposed scheme, which consists of three modules: 3D generative modeling (3DGM), geometric sculpting (GS), and explicit texturing (ET). 3DGM employs synthesized diverse, balanced and structured human images with accurate camera parameters $\\varphi$ to learn generalizable 3D humans with the triplane-based architecture. GS is integrated as an optimization module by utilizing multi-view normal constraints to refine and carve geometry details. ET utilizes UV partitioning and a differentiable rasterizer to disentangle explicit UV texture maps. Not only multi-view renderings but also realistic 3D models can be acquired finally by our method." + ], + "image_footnote": [], + "bbox": [ + 83, + 94, + 890, + 303 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "NeRF [48] optimizes for a canonical, volumetric T-pose of the human with a motion field to map the non-rigid transformations. Text to 3D humans [16, 24] are introduced by combining SDS loss [42] with diffusion guidance. Nevertheless, these methods are learned directly from image sequences or guided by single text prompts, where subjects are fitting to single scenes, thus limiting their scalability.", + "bbox": [ + 75, + 402, + 470, + 508 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Generative 3D-aware Image Synthesis. Recently, 3D-aware image synthesis methods have lifted image generation with explicit view control by integrating the 2D generative models [21-23] with 3D representations, such as voxels [13, 34, 35, 49], meshes [27, 47] and points clouds [1, 26]. GRAF [46] and $\\pi$ -GAN[4] firstly integrate the implicit representation networks, i.e., NeRF [33], with differentiable volumetric rendering for 3D scene generation. However, they have difficulties in training on high-resolution images due to the costly rendering process. Subsequent works have sought to improve the efficiency and quality of such NeRF-based GANs, either by adopting a two-stage rendering process [5, 12, 36, 37, 51] or a smart sampling strategy [7, 55]. StyleSDF [37] combines a SDF-based volume renderer and a 2D StyleGAN network [22] for photorealistic image generation. EG3D [5] introduces a superior triplane representation to leverage 2D CNN-based feature generators for efficient generalization over 3D spaces. Although these methods demonstrate impressive quality in view-consistent image synthesis, they are limited to simplified rigid objects such as faces, cats and cars.", + "bbox": [ + 75, + 516, + 468, + 834 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To learn highly articulated humans from unstructured 2D images, recent works [9, 11, 15, 17, 52, 53] integrate the deformation field to learn non-rigid deformations based on the body prior of estimated SMPL parameters. EVA3D [15]", + "bbox": [ + 75, + 839, + 468, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "represents humans as a composition of multiple parts with NeRF representations. Instead of directly rendering the image from a 3D representation, 3DHumanGAN [52] uses an equivariant 2D generator modulated by 3D human body prior, which enables to establish one-to-many mapping from 3D geometry to synthesized textures from 2D images. AG3D [9] combines the 3D generator with an efficient articulation module to warp 3D objects from canonical space to posed space via a learned continuous deformation field. However, a gap still exists between the generated and real humans in terms of appearance, due to the imprecise priors from complex poses as well as the data biases from limited human poses and imbalanced viewing angles in the dataset.", + "bbox": [ + 496, + 402, + 892, + 601 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method Description", + "text_level": 1, + "bbox": [ + 500, + 616, + 692, + 633 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our goal is to develop a zero-shot 3D generative scheme that does not directly rely on any pre-existing 3D or 2D collections, yet is capable of producing high-quality 3D humans that are visually realistic, geometrically accurate and content-wise diverse to generalize to arbitrary humans.", + "bbox": [ + 496, + 642, + 890, + 718 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "An overview of the proposed scheme is illustrated in Figure 2. We build a sequential pipeline with the following three modules: 3D generative modeling (3DGM), geometric sculpting (GS), and explicit texturing (ET). The first module synthesizes view-balanced, structured and diverse human images with known camera parameters. Subsequently, it learns a 3D generative model from these synthetic data, focusing on realistic appearance modeling (Section 3.1). To overcome the inaccuracy of the 3D shape, the GS module is incorporated during the inference process. It optimizes a hybrid representation with multi-view normal constraints to carve intricate mesh details (Section 3.2).", + "bbox": [ + 496, + 719, + 892, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "9983", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Additionally, the ET module is employed to disentangle explicit texture by utilizing semantical UV partitioning and a differentiable rasterizer (Section 3.3). By combining these modules, we are able to synthesize high-quality and faithful 3D human avatars by incorporating random noises or guided texts/images (Section 3.4).", + "bbox": [ + 75, + 90, + 472, + 183 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1. 3D generative modeling", + "text_level": 1, + "bbox": [ + 76, + 194, + 295, + 209 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Without any 3D or 2D collections, we develop a synthesis-based flow to learn a 3D generative module from 2D synthetic data. We start by instantiating a 3D scene through the projection of underlying 3D skeletons onto 2D pose images, utilizing accurate physical parameters (i.e., camera parameters). Subsequently, the projected 2D pose images serve as conditions to control the 2D diffusion model [54] for synthesizing view-balanced, diverse, and lifelike human images. Finally, we employ a triplane-based generator with enhanced designs to learn a generalizable 3D representation from the synthetic data. Details are described as follows.", + "bbox": [ + 75, + 217, + 468, + 383 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3D instantiation. Starting with a template body mesh (e.g., SMPL-X [40]) positioned and posed in canonical space, we estimate the 3D joint locations $\\mathcal{P}_{3d}$ by regressing them from interpolated vertices. We then project $\\mathcal{P}_{3d}$ onto 2D poses $\\mathcal{P}_i, i = 1,\\dots,k$ from $\\kappa$ horizontally uniformly sampled viewpoints $\\varphi$ . In this way, paired 2D pose images and their corresponding camera parameters $\\{\\mathcal{P}_i,\\varphi_i\\}$ are formulated.", + "bbox": [ + 75, + 385, + 468, + 491 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Controlled 2D image synthesis. With the pose image $\\mathcal{P}_i$ , we feed it into off-the-shelf ControlNet [54] as the pose condition to guide diffusion models [44] to synthesize human images in desired poses (i.e., views). The text prompt $T$ is also used for diverse contents. Given a prompt $T$ , instead of generating a human image $\\mathcal{I}_s: \\mathcal{I}_s = \\mathcal{C}(\\mathcal{P}_i, T)$ independently for each view $\\varphi_i$ , we horizontally concatenate $\\mathcal{K}$ pose images $\\mathcal{P}_i \\in R^{H \\times W \\times 3}$ , resulting in $\\mathcal{P}_i' \\in R^{H \\times KW \\times 3}$ and feed $\\mathcal{P}_i'$ to $\\mathcal{C}$ , along with a prompt hint of 'multi-view' in $T$ . In this way, multi-view human images $\\mathcal{I}_s'$ are synthesized with roughly coherent appearance. We split $\\mathcal{I}_s'$ to single view images $\\mathcal{I}_{\\varphi}$ under specific views $\\varphi$ . This concatenation strategy facilitates the convergence of distributions in synthetic multi-views, thus easing the learning of common 3D representation meeting multi-view characteristics.", + "bbox": [ + 75, + 491, + 468, + 717 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Generalizable 3D representation learning. With synthetic data of paired $\\{\\mathcal{I}_{\\varphi}, \\varphi\\}$ , we learn the 3D generative module $\\mathcal{G}_{3d}$ from them to produce diverse 3D-aware human images with realistic appearance. Inspired by EG3D [5], we employ a triplane-based generator to produce a generalizable representation $\\mathcal{T}$ and introduce a patch-composed neural renderer to learn intricate human representation efficiently. Specifically, instead of uniformly sampling 2D pixels on the image $\\mathcal{I}$ , we decompose patches in the ROI region including human bodies, and only emit rays towards pixels in these patches. The rays are rendered into RGB color with opacity values via volume rendering. Based on", + "bbox": [ + 75, + 719, + 470, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "the decomposed rule, we decode rendered colors to multiple patches and re-combine these patches for full feature images. In this way, the representation is composed of effective human body parts, which directs the attention of the networks towards the human subject itself. This design facilitates fine-grained local human learning while maintaining computational efficiency.", + "bbox": [ + 496, + 90, + 890, + 196 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For the training process, we employ two discriminators, one for RGB images and another for silhouettes, which yields better disentanglement of foreground objects with global geometry. The training loss for this module $L_{3d}$ consists of the two adversarial terms:", + "bbox": [ + 496, + 196, + 890, + 272 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {3 d} = \\mathcal {L} _ {a d v} \\left(\\mathcal {D} _ {r g b}, \\mathcal {G} _ {3 d}\\right) + \\lambda_ {s} \\mathcal {L} _ {a d v} \\left(\\mathcal {D} _ {\\text {m a s k}}, \\mathcal {G} _ {3 d}\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 522, + 284, + 890, + 301 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\lambda_{s}$ denotes the weight of silhouette item. $\\mathcal{L}_{adv}$ is computed by the non-saturating GAN loss with R1 regularization [32].", + "bbox": [ + 496, + 311, + 890, + 357 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "With the trained $\\mathcal{G}_{3d}$ , we can synthesize 3D-aware human images $\\mathcal{T}_g^\\varphi$ with view control, and extract coarse 3D shapes $\\mathcal{M}_c$ from the density field of neural renderer using the Marching Cubes algorithm [30].", + "bbox": [ + 496, + 358, + 890, + 419 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Geometric sculpting", + "text_level": 1, + "bbox": [ + 500, + 428, + 692, + 445 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our 3D generative module can produce high-quality and 3D-consistent human images in view controls. However, its training solely relies on discriminations made using 2D renderings, which can result in inaccuracies in capturing the inherent geometry, especially for complex human bodies. Therefore, we integrate the geometric sculpting, an optimization module leveraging geometric information from high-quality multi-views to carve surface details. Combined with a hybrid 3D representation and a differentiable rasterizer, it can rapidly enhance the shape quality within seconds.", + "bbox": [ + 496, + 450, + 890, + 616 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "DMTET adaption. Owing to the expressive ability of arbitrary topologies and computational efficiency with direct shape optimization, we employ DMTET as our 3D representation in this module and adapt it to the coarse mesh $\\mathcal{M}_c$ via an initial fitting procedure. Specifically, we parameterize DMTET as an MLP network $\\Psi_g$ that learns to predict the SDF value $s(v_i)$ and the position offset $\\delta v_i$ for each vertex $v_i \\in VT$ of the tetrahedral grid $(VT,T)$ . A point set $P = \\{p_i \\in R^3\\}$ is randomly sampled near $\\mathcal{M}_c$ and their SDF values $SDF(p_i)$ can be pre-computed. We adapt the parameters $\\psi$ of $\\Psi_g$ by fitting it to the SDF of $\\mathcal{M}_c$ :", + "bbox": [ + 496, + 617, + 890, + 785 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {a d a} = \\sum_ {p _ {i} \\in P} | | s (p _ {i}; \\psi) - S D F (p _ {i}) | | _ {2}. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 565, + 796, + 890, + 830 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Geometry refinement. Using the adapted DMTET, we leverage the highly-detailed normal maps $\\mathcal{N}$ derived from realistic multi-view images as a guidance to refine local surfaces. To obtain the pseudo-GT normals $\\mathcal{N}_{\\varphi}$ , we extract", + "bbox": [ + 496, + 839, + 890, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "9984", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "them from $\\mathcal{I}_g^\\varphi$ using a pre-trained normal estimator [50]. For the rendered normals $\\hat{\\mathcal{N}}_{\\varphi}$ , we extract the triangular mesh $\\mathcal{M}_{tri}$ from $(VT,T)$ using the Marching Tetrahedra (MT) layer in our current DMTET. By rendering the generated mesh $\\mathcal{M}_{tri}$ with differentiable rasterization, we obtain the resulting normal map $\\hat{\\mathcal{N}}_{\\varphi}$ . To ensure holistic surface polishing that takes into account multi-view normals, we randomly sample camera poses $\\varphi$ that are uniformly distributed in space. We optimize the parameters of $\\Psi_g$ using the normal loss, which is defined as:", + "bbox": [ + 75, + 90, + 472, + 244 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {n o r m}} = \\left\\| \\hat {\\mathcal {N}} _ {\\varphi} - \\mathcal {N} _ {\\varphi} \\right\\| _ {2}. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 189, + 253, + 468, + 273 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "After rapid optimization, the final triangular mesh $\\mathcal{M}_{tri}$ can be easily extracted from the MT layer. If the hands exhibit noise, they can be optionally replaced with cleaner geometry hands from SMPL-X, benefiting from the alignment of the generated body in canonical space with the underlying template body.", + "bbox": [ + 75, + 281, + 468, + 375 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Explicit texturing", + "text_level": 1, + "bbox": [ + 76, + 381, + 250, + 398 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "With the final mesh, the explicit texturing module aims to disentangle a UV texture map from multi-view renderings $\\mathcal{I}_g^\\varphi$ . This intuitive module not only facilitates the incorporation of high-fidelity textures but also enables various editing applications, as verified in Section 4.4.", + "bbox": [ + 75, + 404, + 468, + 479 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Given the polished triangular mesh $\\mathcal{M}_{tri}$ and multi-views $\\mathcal{I}_g^\\varphi$ , we model the explicit texture map $T_{uv}$ of $\\mathcal{M}_{tri}$ with a semantic UV partition and optimize $T_{uv}$ using a differentiable rasterizer $\\mathcal{R}$ [25]. Specifically, leveraging the canonical properties of synthesized bodies, we semantically split $\\mathcal{M}_{tri}$ into $\\gamma$ components and rotate each component vertically, thus enabling effective UV projection for each component with cylinder unwarping. We then combine the texture partitions together for the full texture $T_{uv}$ . We optimize $T_{uv}$ from a randomly initialized scratch using the texture loss, which consists of a multi-view reconstruction term and a total-variation (tv) term:", + "bbox": [ + 75, + 479, + 470, + 660 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {t e x} = \\mathcal {L} _ {r e c} + \\lambda_ {t v} \\mathcal {L} _ {t v}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 192, + 672, + 468, + 688 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\lambda_{tv}$ denotes the weight of the tv loss.", + "bbox": [ + 75, + 698, + 367, + 713 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Multi-view guidance. To ensure comprehensive texturing in the 3D space, we render the color images $\\mathcal{R}(\\mathcal{M}_{tri},\\varphi)$ and silhouettes $S$ using $\\mathcal{R}$ and optimize $T_{uv}$ utilizing multi-view weighted guidance. Their pixel-alignment distances to the original multi-view renderings $\\mathcal{I}_g^\\varphi$ are defined as the reconstruction loss:", + "bbox": [ + 75, + 713, + 468, + 804 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {r e c} = \\sum_ {\\varphi \\in \\Omega} w _ {\\varphi} | | \\mathcal {R} \\left(\\mathcal {M} _ {t r i}, \\varphi\\right) \\cdot \\mathcal {S} - I _ {g} ^ {\\varphi} \\cdot \\mathcal {S} | | _ {2}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 120, + 811, + 468, + 844 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\Omega$ is the set of viewpoints $\\{\\varphi_i, i = 1, \\dots, k\\}$ and $w_{\\varphi}$ denotes weights of different views. $w_{\\varphi}$ equals to 1.0 for $\\varphi \\in \\{front, back\\}$ and 0.2 otherwise.", + "bbox": [ + 76, + 854, + 468, + 902 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/02980479c6aeb301dc5c2200602c629f3b2063c308353b0567074f2c275e80c5.jpg", + "image_caption": [ + "Figure 3. The visualized flowchart of our method that synthesize textured 3D human avatars from input noises, texts or images." + ], + "image_footnote": [], + "bbox": [ + 500, + 87, + 893, + 218 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Smooth constraint. To avoid abrupt variations and smooth the generated texture $T_{uv}$ , we utilize the total-variation loss $\\mathcal{L}_{tv}$ which is computed by:", + "bbox": [ + 498, + 282, + 890, + 329 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {t v} = \\frac {1}{h \\times w \\times c} | | \\nabla_ {x} \\left(T _ {u v}\\right) + \\nabla_ {y} \\left(T _ {u v}\\right) | |, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 553, + 337, + 890, + 367 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $x$ and $y$ denote horizontal and vertical directions.", + "bbox": [ + 500, + 376, + 867, + 391 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Inference", + "text_level": 1, + "bbox": [ + 500, + 400, + 609, + 415 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Built upon the above modules, we can generate high-quality 3D human avatars from either random noises or guided inputs such as texts or images. The flowchart for this process is shown in Figure 3. For input noises, we can easily obtain the final results by sequentially using the 3DGM, GS and ET modules. For text-guided synthesis, we first convert the text into a structured image using our controlled diffusion $\\mathcal{C}$ , and then inverse it to the latent space using PTI [43]. Specially, the GS and ET modules provide an interface that accurately reflects viewed modifications in the final 3D objects. As a result, we utilize the guided image to replace the corresponding view image, which results in improved fidelity in terms of geometry and texture. The same process is applied for input images as guided images.", + "bbox": [ + 496, + 422, + 890, + 636 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experimental Results", + "text_level": 1, + "bbox": [ + 500, + 648, + 705, + 666 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Implementation details. Our process begins by training the 3D generative module (3DGM) on synthetic data. During inference, we integrate the geometric sculpting (GS) and explicit texturing (ET) as optimization modules. For 3DGM, we normalize the template body to the $(0,1)$ space and place its center at the origin of the world coordinate system. We sample $7(\\mathcal{K} = 7)$ viewpoints uniformly from the horizontal plane, ranging from $0^{\\circ}$ to $180^{\\circ}$ (front to back), with a camera radius of 2.7. For each viewpoint, we generate $100K$ images using the corresponding pose image. To ensure diverse synthesis, we use detailed descriptions of age, gender, ethnicity, hairstyle, facial features, and clothing, leveraging a vast word bank. To cover $360^{\\circ}$ views, we horizontally flip the synthesized images and obtain 1.4 million human images at a resolution of $512^{2}$ in total. We train", + "bbox": [ + 496, + 674, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "9985", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/7aed5913e1c5aa7b1793cb911d0316caf918d886877e533dd8741aae16715624.jpg", + "image_caption": [ + "Figure 4. Results of synthesized 3D human avatars at $512^{2}$ ." + ], + "image_footnote": [], + "bbox": [ + 76, + 88, + 893, + 525 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "the 3DGM for about 2.5M iterations with a batch size of 32, using two discriminators with a learning rate of 0.002 and a generator learning rate of 0.0025. The training takes 8 days using 8 NVIDIA Tesla-V100 GPUs. For GS, we optimize $\\psi$ for 400 iterations for DMTET adaption and 100 iterations for surface carving (taking about 15s in total on 1 NVIDIA RTX 3090 GPU). For ET, we set $\\lambda_{uv} = 1$ and optimize $T_{uv}$ for 500 iterations (around 10 seconds). We split $\\mathcal{M}_{tri}$ into $5(\\gamma = 5)$ body parts (i.e., trunk, left/right arm/leg) with cylinder UV unwarping. We use the Adam optimizer with learning rates of 0.01 and 0.001 for $\\Psi_g$ and $T_{uv}$ , respectively. Detailed network architectures can be found in the supplemental materials (Suppl).", + "bbox": [ + 75, + 560, + 472, + 758 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. 3D human generation", + "text_level": 1, + "bbox": [ + 76, + 770, + 282, + 786 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Figure 4 showcases several 3D human avatars synthesized by our pipeline, highlighting the image quality, geometry accuracy, and diverse outputs achieved through our method. Additionally, we explore the interpolation of the latent conditions to yield smooth transitions in appearance, leveraging the smooth latent space learned by our generative model. For more synthesized examples and interpolation results,", + "bbox": [ + 75, + 794, + 468, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "please refer to the Suppl.", + "bbox": [ + 500, + 561, + 666, + 577 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Comparisons", + "text_level": 1, + "bbox": [ + 500, + 588, + 638, + 604 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Qualitative comparison. In Figure 5, we compare our method with three baselines: EVA3D [15] and AG3D [9], which are state-of-the-art methods for generating 3D humans from 2D images, and EG3D [5], which serves as the foundational backbone of our method. The results of first two methods are produced by directly using source codes and trained models released by authors. We train EG3D using our synthetic images with estimated cameras from scratch. As we can see, EVA3D fails to produce $360^{\\circ}$ humans with reasonable back inferring. AG3D and EG3D are able to generate $360^{\\circ}$ renderings but both struggle with photorealism and capturing detailed shapes. Our method synthesizes not only higher-quality, view-consistent $360^{\\circ}$ images but also higher-fidelity 3D geometry with intricate details, such as irregular dresses and haircuts.", + "bbox": [ + 496, + 612, + 890, + 838 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Quantitative comparison. Table 1 provides quantitative results comparing our method against the baselines. We measure image quality with Frechet Inception Distance (FID) [14] and Inception Score [45] for $360^{\\circ}$ views (IS-", + "bbox": [ + 496, + 839, + 892, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "9986", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/8eceb44aa9b924d4e0e2baa8eb42244150844cfa6809d3484dd1dae0082b5330.jpg", + "image_caption": [ + "Figure 5. Qualitative comparison with three state-of-the-art methods: EVA3D [15], AG3D [9] and EG3D [5]." + ], + "image_footnote": [], + "bbox": [ + 84, + 87, + 893, + 478 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/de17e76b97243efd18bb4d15c898b4c0f969df22da12a7b1bc9d4d3e7190697e.jpg", + "table_caption": [ + "Table 1. Quantitative evaluation using FID, IS-360, normal accuracy (Normal) and identity consistency (ID)." + ], + "table_footnote": [], + "table_body": "
MethodFID ↓IS-360 ↑Normal ↓ID↑
EVA3D [15]15.913.1930.810.72
AG3D [9]10.933.2820.830.69
EVA3D-syn6.273.259.570.72
AG3D-syn8.933.228.640.71
EG3D [5]7.483.2612.740.71
Ours2.733.435.620.74
", + "bbox": [ + 99, + 542, + 444, + 657 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "360). FID measures the visual similarity and distribution discrepancy between 50k generated images and all real images. IS-360 focuses on the self-realism of generated images in $360^{\\circ}$ views. For shape evaluation, we compute FID between rendered normals and pseudo-GT normal maps (Normal), following AG3D. The FID and Normal scores of EVA3D and AG3D are directly fetched from their reports. Additionally, we access the multi-view facial identity consistency using the ID metric introduced by EG3D. For a clearer comparison, we also provide metric results of EVA3D and AG3D trained on our synthetic data (-syn). Overall, our method demonstrates significant improvements in FID and Normal, bringing the generative human model to a new level of realistic $360^{\\circ}$ renderings with delicate ge", + "bbox": [ + 75, + 688, + 472, + 902 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/052fa6b1099cf9a07d05e0931f49fcea8fb28e3a918b9814592058e03da80542.jpg", + "table_caption": [ + "Table 2. Results of models trained by replacing physical parameters with estimated ones (w/o SYN-P) or removing patch-composed rendering (w/o PCR)." + ], + "table_footnote": [], + "table_body": "
OursOurs-w/o SYN-POurs-w/o PCR
FID ↓2.734.283.26
IS-360 ↑3.433.313.35
", + "bbox": [ + 519, + 556, + 870, + 616 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "ometry while maintaining state-of-the-art view consistency.", + "bbox": [ + 498, + 641, + 890, + 657 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Ablation study", + "text_level": 1, + "bbox": [ + 500, + 665, + 650, + 681 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Synthesis flow and patch-composed rendering. We assess the impact of our carefully designed synthesis flow by training a model with synthetic images but with camera and pose parameters estimated by SMPLify-X [40] (w/o SYN-P). As Table 2 shows, the model w/o SYN-P results in worse FID and IS-360 scores, indicating that the synthesis flow contributes to more accurate physical parameters for realistic appearance modeling. By utilizing patch-composed rendering (PCR), the networks focus more on the human region, leading to more realistic results.", + "bbox": [ + 496, + 688, + 890, + 839 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Geometry sculpting module (GS). We demonstrate the importance of this module by visualizing the meshes before and after its implementation. Figure 6 (b) shows that the preceding module yields a coarse mesh due to the complex", + "bbox": [ + 496, + 839, + 893, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "9987", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/8537eac5ff968ce7e4b86fb01af23539baaf0effb4de65a6e01d86a6a9a015d3.jpg", + "image_caption": [ + "(a) Rendering" + ], + "image_footnote": [], + "bbox": [ + 84, + 90, + 179, + 196 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/4cc72be7a7089040d22569f9920fbca696ab319f2661a67270d18c062954e5fd.jpg", + "image_caption": [ + "(b) w/o GS" + ], + "image_footnote": [], + "bbox": [ + 181, + 90, + 279, + 196 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/abf98b26aba27326a1c6e9a2e538f9aded98da542dda48a83a4a162b73f86b72.jpg", + "image_caption": [ + "(c) Full method" + ], + "image_footnote": [], + "bbox": [ + 282, + 90, + 379, + 196 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/93db33d01c4539a9f13305c65fc09867c17514a6f378dd8e40164170b106125f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 383, + 95, + 467, + 196 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/7c4797b81ef35a91dfd5910f7be2a8c9c8e618e521c93f2b21df1d388196c4a3.jpg", + "image_caption": [ + "Figure 6. Effects of the GS module to carve fine-grained surfaces.", + "(a) Guided image" + ], + "image_footnote": [], + "bbox": [ + 91, + 247, + 184, + 353 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/1b31d5535ae48e32a8971579db57f0de9c7157d9f719c0ca81ca92d16a5532a4.jpg", + "image_caption": [ + "(b) w/o ET", + "Figure 7. Effects of the ET module for guided synthesis." + ], + "image_footnote": [], + "bbox": [ + 189, + 247, + 279, + 354 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/73e7659150accc5a8ed134f271776f58a0194b43b6117cfbb3776dcc7eb462ca.jpg", + "image_caption": [ + "(c) Full method" + ], + "image_footnote": [], + "bbox": [ + 287, + 247, + 379, + 354 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/9afc8ce98f79de350a856978094f4cb2a80bdf96b59c8fa9cbfe672e5f779906.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 383, + 247, + 467, + 354 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "human anatomy and the challenges posed by decomposing ambiguous 3D shapes from 2D images. The GS module utilizes high-quality multi-view outputs and employs a more flexible hybrid representation to create expressive humans with arbitrary topologies. It learns from pixel-level surface supervision, leading to a significant improvement in shape quality, characterized by smooth surfaces and intricate outfits (Figure 6 (c)).", + "bbox": [ + 75, + 414, + 467, + 535 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Explicit texturing module (ET). This intuitive module not only extracts the explicit UV texture for complete 3D assets but also enables high-fidelity results for image guided synthesis. Following the flowchart in Figure 3, we compare the results produced with and without this module. Our method without ET directly generates implicit renderings through PTI inversion, as shown in Figure 7 (b). While it successfully preserves global identity, it struggles to synthesize highly faithful local textures (e.g., floral patterns). The ET module offers a convenient and efficient way to directly interact with the 3D representation, enabling the production of high-fidelity 3D humans with more consistent content including exquisite local patterns (Figure 7 (a, c)).", + "bbox": [ + 75, + 536, + 467, + 732 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4. Applications", + "text_level": 1, + "bbox": [ + 76, + 741, + 210, + 757 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Avatar animation. All avatars produced by our method are in a canonical body pose and aligned to an underlying 3D skeleton extracted from SMPL-X. This alignment allows for easy animation and the generation of motion videos, as demonstrated in Figure 1 and Suppl.", + "bbox": [ + 75, + 763, + 467, + 840 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Texture doodle and local editing. Our approach benefits from the explicit disentanglement of geometry and texture, enabling flexible editing capabilities. Following the flowchart of text or image guided synthesis (Section 3.4),", + "bbox": [ + 75, + 840, + 467, + 900 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/9a378752288b93e83b8a16f8a95e63c26c8f2fecb30cbf6c0cc9885649a09791.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 501, + 89, + 663, + 189 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/583a0aad33465cce86484fb0f9b15aa788adc4e18e73e6b916bac8029c63d208.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 663, + 90, + 803, + 188 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/70995fc2ada3f1fc26e4b2fe8b0e376047f51a38d3a01c66bb2473c40ac342e7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 805, + 90, + 890, + 188 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/d1ca738bad190fd3d970be65da1a0523155ff37b795d55f60f21ad9599b13c99.jpg", + "image_caption": [ + "Figure 8. Results synthesized by adapting our method to various styles (e.g., Disney cartoon characters) or contents (e.g., portrait heads)." + ], + "image_footnote": [], + "bbox": [ + 501, + 190, + 660, + 255 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/057816ebef0725a7559f2491d385fa0718a06e7d46476a0e286f818230fde632.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 660, + 190, + 803, + 255 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/c9c097e0e8aca241825753fbdde53c38c562a6cf475ee50049f556cde5b46803.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 805, + 190, + 890, + 255 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "users can paint any pattern or add text to a guided image. These modifications can be transferred to 3D human models by inputting modified views into the texture module (e.g., painting the text 'hey' on a jacket as shown in Figure 1 (d)). Our approach also allows for clothing editing by simultaneously injecting edited guide images with desired clothing into the GS and ET modules (e.g., changing a jacket and jeans to bodysuits in Figure 1 (e)). More results can be found in Suppl.", + "bbox": [ + 496, + 337, + 890, + 473 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Content-style free adaption. Our proposed scheme is versatile and can be extended to generate various types of contents (e.g., portrait heads) and styles (e.g., Disney cartoon characters). To achieve this, we fine-tune our model using synthetic images from these domains, allowing for flexible adaptation. We showcase the results in Figure 8. More results and other discussions (e.g., limitations, negative impact, etc.) can be found in the Suppl.", + "bbox": [ + 496, + 474, + 890, + 595 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 616, + 617, + 631 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We introduced En3D, a novel generative scheme for sculpting 3D humans from 2D synthetic data. This method overcomes limitations in existing 3D or 2D collections and significantly enhances the image quality, geometry accuracy, and content diversity of generated 3D humans. En3D comprises a 3D generative module that learns generalizable 3D humans from synthetic 2D data with accurate physical modeling, and two optimization modules to carve intricate shape details and disentangle explicit UV textures with high fidelity, respectively. Experimental results validated the superiority and effectiveness of our method. We also demonstrated the flexibility of our generated avatars for animation and editing, as well as the scalability of our approach for synthesizing portraits and Disney characters. We believe that our solution could provide invaluable human assets for the 3D vision community. Furthermore, it holds potential for use in common 3D object synthesis tasks.", + "bbox": [ + 496, + 643, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "9988", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Panos Achlioptas, Olga Diamanti, Ioannis Mitliagkas, and Leonidas Guibas. Learning representations and generative models for 3d point clouds. In International conference on machine learning, pages 40-49. PMLR, 2018. 3", + "[2] Thiemo Alldieck, Marcus Magnor, Weipeng Xu, Christian Theobalt, and Gerard Pons-Moll. Video based reconstruction of 3d people models. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 8387-8397, 2018. 2", + "[3] Dragomir Anguelov, Praveen Srinivasan, Daphne Koller, Sebastian Thrun, Jim Rodgers, and James Davis. Scape: shape completion and animation of people. In ACM SIGGRAPH 2005 Papers, pages 408-416. 2005. 2", + "[4] Eric R Chan, Marco Monteiro, Petr Kellnhofer, Jiajun Wu, and Gordon Wetzstein. pi-gan: Periodic implicit generative adversarial networks for 3d-aware image synthesis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5799-5809, 2021. 3", + "[5] Eric R Chan, Connor Z Lin, Matthew A Chan, Koki Nagano, Boxiao Pan, Shalini De Mello, Orazio Gallo, Leonidas J Guibas, Jonathan Tremblay, Sameh Khamis, et al. Efficient geometry-aware 3d generative adversarial networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16123-16133, 2022. 2, 3, 4, 6, 7", + "[6] Xu Chen, Tianjian Jiang, Jie Song, Jinlong Yang, Michael J Black, Andreas Geiger, and Otmar Hilliges. gdna: Towards generative detailed neural avatars. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20427-20437, 2022. 2", + "[7] Yu Deng, Jiaolong Yang, Jianfeng Xiang, and Xin Tong. Gram: Generative radiance manifolds for 3d-aware image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10673-10683, 2022. 3", + "[8] Zijian Dong, Chen Guo, Jie Song, Xu Chen, Andreas Geiger, and Otmar Hilliges. Pina: Learning a personalized implicit neural avatar from a single rgb-d video sequence. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20470-20480, 2022. 1, 2", + "[9] Zijian Dong, Xu Chen, Jinlong Yang, Michael J Black, Otmar Hilliges, and Andreas Geiger. Ag3d: Learning to generate 3d avatars from 2d image collections. arXiv preprint arXiv:2305.02312, 2023. 2, 3, 6, 7", + "[10] Yao Feng, Jinlong Yang, Marc Pollefeys, Michael J Black, and Timo Bolkart. Capturing and animation of body and clothing from monocular video. In SIGGRAPH Asia 2022 Conference Papers, pages 1-9, 2022. 1, 2", + "[11] Jianglin Fu, Shikai Li, Yuming Jiang, Kwan-Yee Lin, Chen Qian, Chen Change Loy, Wayne Wu, and Ziwei Liu. Stylegan-human: A data-centric odyssey of human generation. In European Conference on Computer Vision, pages 1-19. Springer, 2022. 3", + "[12] Jiatao Gu, Lingjie Liu, Peng Wang, and Christian Theobalt. Stylenerf: A style-based 3d-aware genera" + ], + "bbox": [ + 78, + 114, + 470, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "tor for high-resolution image synthesis. arXiv preprint arXiv:2110.08985, 2021.3", + "[13] Philipp Henzler, Niloy J Mitra, and Tobias Ritschel. Escaping plato's cave: 3d shape from adversarial rendering. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9984-9993, 2019. 3", + "[14] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 6", + "[15] Fangzhou Hong, Zhaoxi Chen, Yushi Lan, Liang Pan, and Ziwei Liu. Eva3d: Compositional 3d human generation from 2d image collections. arXiv preprint arXiv:2210.04888, 2022. 2, 3, 6, 7", + "[16] Xin Huang, Ruizhi Shao, Qi Zhang, Hongwen Zhang, Ying Feng, Yebin Liu, and Qing Wang. Humannorm: Learning normal diffusion model for high-quality and realistic 3d human generation. arXiv preprint arXiv:2310.01406, 2023. 3", + "[17] Suyi Jiang, Haoran Jiang, Ziyu Wang, Haimin Luo, Wenzheng Chen, and Lan Xu. Humangen: Generating human radiance fields with explicit priors. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12543-12554, 2023. 3", + "[18] Kyungmin Jo, Wonjoon Jin, Jaegul Choo, Hyunjoon Lee, and Sunghyun Cho. 3d-aware generative model for improved side-view image synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 22862-22872, 2023. 2", + "[19] Hanbyul Joo, Tomas Simon, and Yaser Sheikh. Total capture: A 3d deformation model for tracking faces, hands, and bodies. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 8320-8329, 2018. 2", + "[20] Angjoo Kanazawa, Michael J Black, David W Jacobs, and Jitendra Malik. End-to-end recovery of human shape and pose. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7122-7131, 2018. 2", + "[21] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4401-4410, 2019. 3", + "[22] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of stylegan. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8110-8119, 2020. 3", + "[23] Tero Karras, Miika Aittala, Samuli Laine, Erik Härkönen, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Alias-free generative adversarial networks. Advances in Neural Information Processing Systems, 34:852-863, 2021. 3", + "[24] Nikos Kolotouros, Thiemo Alldieck, Andrei Zanfir, Eduard Bazavan, Mihai Fieraru, and Cristian Sminchisescu. Dreamhuman: Animatable 3d avatars from text. Advances in Neural Information Processing Systems, 36, 2024. 3", + "[25] Samuli Laine, Janne Hellsten, Tero Karras, Yeongho Seol, Jaakko Lehtinen, and Timo Aila. Modular primitives for high-performance differentiable rendering. ACM Transactions on Graphics (TOG), 39(6):1-14, 2020. 5" + ], + "bbox": [ + 503, + 92, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9989", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[26] Ruihui Li, Xianzhi Li, Chi-Wing Fu, Daniel Cohen-Or, and Pheng-Ann Heng. Pu-gan: a point cloud upsampling adversarial network. In Proceedings of the IEEE/CVF international conference on computer vision, pages 7203–7212, 2019. 3", + "[27] Yiyi Liao, Katja Schwarz, Lars Mescheder, and Andreas Geiger. Towards unsupervised learning of generative models for 3d controllable image synthesis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5871-5880, 2020. 3", + "[28] Lingjie Liu, Marc Habermann, Viktor Rudnev, Kripasindhu Sarkar, Jiatao Gu, and Christian Theobalt. Neural actor: Neural free-view synthesis of human actors with pose control. ACM transactions on graphics (TOG), 40(6):1-16, 2021. 2", + "[29] Matthew Loper, Naureen Mahmood, Javier Romero, Gerard Pons-Moll, and Michael J Black. Smpl: A skinned multiperson linear model. ACM Transactions on Graphics, 34(6), 2015. 2", + "[30] William E Lorensen and Harvey E Cline. Marching cubes: A high resolution 3d surface construction algorithm. In Semin al graphics: pioneering efforts that shaped the field, pages 347-353. 1998. 4", + "[31] Qianli Ma, Jinlong Yang, Anurag Ranjan, Sergi Pujades, Gerard Pons-Moll, Siyu Tang, and Michael J Black. Learning to dress 3d people in generative clothing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6469-6478, 2020. 2", + "[32] Lars Mescheder, Andreas Geiger, and Sebastian Nowozin. Which training methods for gans do actually converge? In International conference on machine learning, pages 3481-3490. PMLR, 2018. 4", + "[33] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 3", + "[34] Thu Nguyen-Phuoc, Chuan Li, Lucas Theis, Christian Richardt, and Yong-Liang Yang. Hologan: Unsupervised learning of 3d representations from natural images. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7588-7597, 2019. 3", + "[35] Thu H Nguyen-Phuoc, Christian Richardt, Long Mai, Yongliang Yang, and Niloy Mitra. Blockgan: Learning 3d object-aware scene representations from unlabelled images. Advances in neural information processing systems, 33:6767-6778, 2020. 3", + "[36] Michael Niemeyer and Andreas Geiger. Giraffe: Representing scenes as compositional generative neural feature fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11453-11464, 2021. 3", + "[37] Roy Or-El, Xuan Luo, Mengyi Shan, Eli Shechtman, Jeong Joon Park, and Ira Kemelmacher-Shlizerman. Stylesdf: High-resolution 3d-consistent image and geometry generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13503–13513, 2022. 2, 3" + ], + "bbox": [ + 78, + 90, + 470, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[38] Ahmed AA Osman, Timo Bolkart, and Michael J Black. Star: Sparse trained articulated human body regressor. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part VI 16, pages 598-613. Springer, 2020. 2", + "[39] Pablo Palafox, Aljaž Božić, Justus Thies, Matthias Nießner, and Angela Dai. Npms: Neural parametric models for 3d deformable shapes. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12695-12705, 2021. 2", + "[40] Georgios Pavlakos, Vasileios Choutas, Nima Ghorbani, Timo Bolkart, Ahmed AA Osman, Dimitrios Tzionas, and Michael J Black. Expressive body capture: 3d hands, face, and body from a single image. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10975-10985, 2019. 4, 7", + "[41] Sida Peng, Yuanqing Zhang, Yinghao Xu, Qianqian Wang, Qing Shuai, Hujun Bao, and Xiaowei Zhou. Neural body: Implicit neural representations with structured latent codes for novel view synthesis of dynamic humans. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9054-9063, 2021. 1, 2", + "[42] Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. arXiv preprint arXiv:2209.14988, 2022. 3", + "[43] Daniel Roich, Ron Mokady, Amit H Bermano, and Daniel Cohen-Or. Pivotal tuning for latent-based editing of real images. ACM Transactions on graphics (TOG), 42(1):1-13, 2022. 5", + "[44] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 4", + "[45] Tim Salimans, Ian Goodfellow, Wojciech Zaremba, Vicki Cheung, Alec Radford, and Xi Chen. Improved techniques for training gans. Advances in neural information processing systems, 29, 2016. 6", + "[46] Katja Schwarz, Yiyi Liao, Michael Niemeyer, and Andreas Geiger. Graf: Generative radiance fields for 3d-aware image synthesis. Advances in Neural Information Processing Systems, 33:20154-20166, 2020. 3", + "[47] Attila Szabó, Givi Meishvili, and Paolo Favaro. Unsupervised generative 3d shape learning from natural images. arXiv preprint arXiv:1910.00287, 2019. 3", + "[48] Chung-Yi Weng, Brian Curless, Pratul P Srinivasan, Jonathan T Barron, and Ira Kemelmacher-Shlizerman. Humanerf: Free-viewpoint rendering of moving people from monocular video. In Proceedings of the IEEE/CVF conference on computer vision and pattern Recognition, pages 16210-16220, 2022. 1, 2, 3", + "[49] Jiajun Wu, Chengkai Zhang, Tianfan Xue, Bill Freeman, and Josh Tenenbaum. Learning a probabilistic latent space of object shapes via 3d generative-adversarial modeling. Advances in neural information processing systems, 29, 2016. 3", + "[50] Yuliang Xiu, Jinlong Yang, Dimitrios Tzionas, and Michael J Black. Icon: Implicit clothed humans obtained from nor" + ], + "bbox": [ + 501, + 92, + 892, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "9990", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "mals. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13286-13296. IEEE, 2022. 5", + "[51] Yang Xue, Yuheng Li, Krishna Kumar Singh, and Yong Jae Lee. Giraffe hd: A high-resolution 3d-aware generative model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18440-18449, 2022. 3", + "[52] Zhuoqian Yang, Shikai Li, Wayne Wu, and Bo Dai. 3dhumanan: 3d-aware human image generation with 3d pose mapping. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 23008-23019, 2023. 3", + "[53] Jianfeng Zhang, Zihang Jiang, Dingdong Yang, Hongyi Xu, Yichun Shi, Guoxian Song, Zhongcong Xu, Xinchao Wang, and Jiashi Feng. Avatargen: a 3d generative model for animatable human avatars. In European Conference on Computer Vision, pages 668-685. Springer, 2022. 3", + "[54] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3836-3847, 2023. 4", + "[55] Peng Zhou, Lingxi Xie, Bingbing Ni, and Qi Tian. Cips-3d: A 3d-aware generator of gans based on conditionally-independent pixel synthesis. arXiv preprint arXiv:2110.09788, 2021. 3" + ], + "bbox": [ + 78, + 90, + 468, + 457 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "9991", + "bbox": [ + 482, + 945, + 513, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2024/En3D_ An Enhanced Generative Model for Sculpting 3D Humans from 2D Synthetic Data/cb7c8896-3bc3-4bc4-a3cb-b2aedebff579_model.json b/2024/En3D_ An Enhanced Generative Model for Sculpting 3D Humans from 2D Synthetic Data/cb7c8896-3bc3-4bc4-a3cb-b2aedebff579_model.json new file mode 100644 index 0000000000000000000000000000000000000000..e7e203f9a88bed3b233316a562746812b3ac67ed --- /dev/null +++ b/2024/En3D_ An Enhanced Generative Model for Sculpting 3D Humans from 2D Synthetic Data/cb7c8896-3bc3-4bc4-a3cb-b2aedebff579_model.json @@ -0,0 +1,2290 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.128, + 0.13, + 0.844, + 0.175 + ], + "angle": 0, + "content": "En3D: An Enhanced Generative Model for Sculpting 3D Humans from 2D Synthetic Data" + }, + { + "type": "text", + "bbox": [ + 0.143, + 0.203, + 0.826, + 0.223 + ], + "angle": 0, + "content": "Yifang Men1, Biwen Lei1, Yuan Yao1, Miaomiao Cui1, Zhouhui Lian2, Xuansong Xie1" + }, + { + "type": "text", + "bbox": [ + 0.259, + 0.224, + 0.71, + 0.24 + ], + "angle": 0, + "content": "\\(^{1}\\)Institute for Intelligent Computing, Alibaba Group" + }, + { + "type": "text", + "bbox": [ + 0.185, + 0.242, + 0.786, + 0.258 + ], + "angle": 0, + "content": "\\(^{2}\\)Wangxuan Institute of Computer Technology, Peking University, China" + }, + { + "type": "text", + "bbox": [ + 0.254, + 0.26, + 0.714, + 0.275 + ], + "angle": 0, + "content": "https://menyifang.github.io/projects/En3D/index.html" + }, + { + "type": "image_caption", + "bbox": [ + 0.088, + 0.301, + 0.114, + 0.317 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.088, + 0.313, + 0.538, + 0.434 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.544, + 0.296, + 0.568, + 0.312 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "text", + "bbox": [ + 0.586, + 0.296, + 0.846, + 0.319 + ], + "angle": 0, + "content": "\"a 44-year-old Latino man with blonde side part haircut, wearing a bomber jacket with brooch\"" + }, + { + "type": "image", + "bbox": [ + 0.553, + 0.323, + 0.875, + 0.433 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.091, + 0.437, + 0.113, + 0.453 + ], + "angle": 0, + "content": "(c)" + }, + { + "type": "image", + "bbox": [ + 0.1, + 0.436, + 0.49, + 0.549 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.511, + 0.437, + 0.535, + 0.453 + ], + "angle": 0, + "content": "(d)" + }, + { + "type": "image", + "bbox": [ + 0.522, + 0.437, + 0.71, + 0.548 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.708, + 0.438, + 0.731, + 0.453 + ], + "angle": 0, + "content": "(e)" + }, + { + "type": "image", + "bbox": [ + 0.714, + 0.44, + 0.889, + 0.548 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.075, + 0.551, + 0.894, + 0.594 + ], + "angle": 0, + "content": "Figure 1. Given random noises or guided texts, our generative scheme can synthesize high-fidelity 3D human avatars that are visually realistic and geometrically accurate. These avatars can be seamlessly animated and easily edited. Our model is trained on 2D synthetic data without directly relying on any pre-existing 3D or 2D collections." + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.606, + 0.314, + 0.623 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.644, + 0.474, + 0.903 + ], + "angle": 0, + "content": "We present En3D, an enhanced generative scheme for sculpting high-quality 3D human avatars. Unlike previous works that rely on scarce 3D datasets or limited 2D collections with imbalanced viewing angles and imprecise pose priors, our approach aims to develop a zero-shot 3D generative scheme capable of producing visually realistic, geometrically accurate and content-wise diverse 3D humans without directly relying on pre-existing 3D or 2D assets. To address this challenge, we introduce a meticulously crafted workflow that implements accurate physical modeling to learn the enhanced 3D generative model from synthetic 2D data. During inference, we integrate optimization modules to bridge the gap between realistic appearances and coarse 3D shapes. Specifically, En3D comprises three modules: a 3D generator that accurately models generalizable 3D humans with realistic appearance from synthesized balanced, diverse, and structured human images; a geometry sculptor" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.608, + 0.896, + 0.76 + ], + "angle": 0, + "content": "that enhances shape quality using multi-view normal constraints for intricate human structure; and a texturing module that disentangles explicit texture maps with fidelity and editability, leveraging semantical UV partitioning and a differentiable rasterizer. Experimental results show that our approach significantly outperforms prior works in terms of image quality, geometry accuracy and content diversity. We also showcase the applicability of our generated avatars for animation and editing, as well as the scalability of our approach for content-style free adaptation." + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.785, + 0.633, + 0.801 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.81, + 0.894, + 0.903 + ], + "angle": 0, + "content": "3D human avatars play an important role in various applications of AR/VR such as video games, telepresence and virtual try-on. Realistic human modeling is an essential task, and many valuable efforts have been made by leveraging neural implicit fields to learn high-quality articulated avatars [8, 10, 41, 48]. However, these methods are directly" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "9981" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.137 + ], + "angle": 0, + "content": "learned from monocular videos or image sequences, where subjects are single individuals wearing specific garments, thus limiting their scalability." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.14, + 0.471, + 0.442 + ], + "angle": 0, + "content": "Generative models learn a shared 3D representation to synthesize clothed humans with varying identities, clothing and poses. Traditional methods are typically trained on 3D datasets, which are limited and expensive to acquire. This data scarcity limits the model's generalization ability and may lead to overfitting on small datasets. Recently, 3D-aware image synthesis methods [5, 18, 37] have demonstrated great potential in learning 3D generative models of rigid objects from 2D image collections. Follow-up works show the feasibility of learning articulated humans from image collections driven by SMPL-based deformations, but only in limited quality and resolution. EVA3D [15] represents humans as a composition of multiple parts with NeRF representations. AG3D [9] incorporates an efficient articulation module to capture both body shape and cloth deformation. Nevertheless, there remains a noticeable gap between generated and real humans in terms of appearance and geometry. Moreover, their results are limited to specific views (i.e., frontal angles) and lack diversity (i.e., fashion images in similar skin tone, body shape, and age)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.445, + 0.47, + 0.807 + ], + "angle": 0, + "content": "The aim of this paper is to propose a zero-shot 3D generative scheme that does not directly rely on any pre-existing 3D or 2D datasets, yet is capable of producing high-quality 3D humans that are visually realistic, geometrically accurate, and content-wise diverse. The generated avatars can be seamlessly animated and easily edited. An illustration is provided in Figure 1. To address this challenging task, our proposed method inherits from 3D-aware human image synthesis and exhibits substantial distinctions based on several key insights. Rethinking the nature of 3D-aware generative methods from 2D collections [5, 9, 15], they actually try to learn a generalizable and deformable 3D representation, whose 2D projections can meet the distribution of human images in corresponding views. Thereby, it is crucial for accurate physical modeling between 3D objects and 2D projections. However, previous works typically leverage pre-existing 2D human images to estimate physical parameters (i.e., camera and body poses), which are inaccurate because of imprecise SMPL priors for highly-articulated humans. This inaccuracy limits the synthesis ability for realistic multi-view renderings. Second, these methods solely rely on discriminating 2D renderings, which is ambiguous and loose to capture inherent 3D shapes in detail, especially for intricate human appearance." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.811, + 0.47, + 0.903 + ], + "angle": 0, + "content": "To address these limitations, we propose a novel generative scheme with two core designs. Firstly, we introduce a meticulously-crafted workflow that implements accurate physical modeling to learn an enhanced 3D generative model from synthetic data. This is achieved by instantiating a 3D body scene and projecting the underlying" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.394 + ], + "angle": 0, + "content": "3D skeletons into 2D pose images using explicit camera parameters. These 2D pose images act as conditions to control a 2D diffusion model, synthesizing realistic human images from specific viewpoints. By leveraging synthetic view-balanced, diverse and structured human images, along with known physical parameters, we employ a 3D generator equipped with an enhanced renderer and discriminator to learn realistic appearance modeling. Secondly, we improve the 3D shape quality by leveraging the gap between high-quality multi-view renderings and the coarse mesh produced by the 3D generative module. Specifically, we integrate an optimization module that utilizes multi-view normal constraints to rapidly refine geometry details under supervision. Additionally, we incorporate an explicit texturing module to ensure faithful UV texture maps. In contrast to previous works that rely on inaccurate physical settings and inadequate shape supervision, we rebuild the generative scheme from the ground up, resulting in comprehensive improvements in image quality, geometry accuracy, and content diversity. In summary, our contributions are threefold:" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.395, + 0.892, + 0.469 + ], + "angle": 0, + "content": "- We present a zero-shot generative scheme that efficiently synthesizes high-quality 3D human avatars with visual realism, geometric accuracy and content diversity. These avatars can be seamlessly animated and easily edited, offering greater flexibility in their applications." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.47, + 0.892, + 0.545 + ], + "angle": 0, + "content": "- We develop a meticulously-crafted workflow to learn an enhanced generative model from synthesized human images that are balanced, diverse, and also possess known physical parameters. This leads to diverse 3D-aware human image synthesis with realistic appearance." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.545, + 0.892, + 0.604 + ], + "angle": 0, + "content": "- We propose to integrate optimization modules into the 3D generator, leveraging multi-view guidance to enhance both shape quality and texture fidelity, thus achieving realistic 3D human assets." + }, + { + "type": "list", + "bbox": [ + 0.5, + 0.395, + 0.892, + 0.604 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.619, + 0.637, + 0.635 + ], + "angle": 0, + "content": "2. Related work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.644, + 0.892, + 0.901 + ], + "angle": 0, + "content": "3D Human Modeling. Parametric models [3, 19, 20, 29, 38] serve as a common representation for 3D human modeling, they allows for robust control by deforming a template mesh with a series of low-dimensional parameters, but can only generate naked 3D humans. Similar ideas have been extended to model clothed humans [2, 31], but geometric expressivity is restricted due to the fixed mesh topology. Subsequent works [6, 39, 39] further introduce implicit surfaces to produce complex non-linear deformations of 3D bodies. Unfortunately, the aforementioned approaches all require 3D scans of various human poses for model fitting, which are difficult to acquire. With the explosion of NeRF, valuable efforts have been made towards combining NeRF models with explicit human models [8, 10, 28, 41, 48]. Neural body [41] anchors a set of latent codes to the vertices of the SMPL model [29] and transforms the spatial locations of the codes to the volume in the observation space. Human-" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9982" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.084, + 0.095, + 0.891, + 0.304 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.306, + 0.893, + 0.377 + ], + "angle": 0, + "content": "Figure 2. An overview of the proposed scheme, which consists of three modules: 3D generative modeling (3DGM), geometric sculpting (GS), and explicit texturing (ET). 3DGM employs synthesized diverse, balanced and structured human images with accurate camera parameters \\(\\varphi\\) to learn generalizable 3D humans with the triplane-based architecture. GS is integrated as an optimization module by utilizing multi-view normal constraints to refine and carve geometry details. ET utilizes UV partitioning and a differentiable rasterizer to disentangle explicit UV texture maps. Not only multi-view renderings but also realistic 3D models can be acquired finally by our method." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.404, + 0.471, + 0.51 + ], + "angle": 0, + "content": "NeRF [48] optimizes for a canonical, volumetric T-pose of the human with a motion field to map the non-rigid transformations. Text to 3D humans [16, 24] are introduced by combining SDS loss [42] with diffusion guidance. Nevertheless, these methods are learned directly from image sequences or guided by single text prompts, where subjects are fitting to single scenes, thus limiting their scalability." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.517, + 0.47, + 0.835 + ], + "angle": 0, + "content": "Generative 3D-aware Image Synthesis. Recently, 3D-aware image synthesis methods have lifted image generation with explicit view control by integrating the 2D generative models [21-23] with 3D representations, such as voxels [13, 34, 35, 49], meshes [27, 47] and points clouds [1, 26]. GRAF [46] and \\(\\pi\\)-GAN[4] firstly integrate the implicit representation networks, i.e., NeRF [33], with differentiable volumetric rendering for 3D scene generation. However, they have difficulties in training on high-resolution images due to the costly rendering process. Subsequent works have sought to improve the efficiency and quality of such NeRF-based GANs, either by adopting a two-stage rendering process [5, 12, 36, 37, 51] or a smart sampling strategy [7, 55]. StyleSDF [37] combines a SDF-based volume renderer and a 2D StyleGAN network [22] for photorealistic image generation. EG3D [5] introduces a superior triplane representation to leverage 2D CNN-based feature generators for efficient generalization over 3D spaces. Although these methods demonstrate impressive quality in view-consistent image synthesis, they are limited to simplified rigid objects such as faces, cats and cars." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.47, + 0.903 + ], + "angle": 0, + "content": "To learn highly articulated humans from unstructured 2D images, recent works [9, 11, 15, 17, 52, 53] integrate the deformation field to learn non-rigid deformations based on the body prior of estimated SMPL parameters. EVA3D [15]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.404, + 0.893, + 0.602 + ], + "angle": 0, + "content": "represents humans as a composition of multiple parts with NeRF representations. Instead of directly rendering the image from a 3D representation, 3DHumanGAN [52] uses an equivariant 2D generator modulated by 3D human body prior, which enables to establish one-to-many mapping from 3D geometry to synthesized textures from 2D images. AG3D [9] combines the 3D generator with an efficient articulation module to warp 3D objects from canonical space to posed space via a learned continuous deformation field. However, a gap still exists between the generated and real humans in terms of appearance, due to the imprecise priors from complex poses as well as the data biases from limited human poses and imbalanced viewing angles in the dataset." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.617, + 0.693, + 0.635 + ], + "angle": 0, + "content": "3. Method Description" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.643, + 0.892, + 0.719 + ], + "angle": 0, + "content": "Our goal is to develop a zero-shot 3D generative scheme that does not directly rely on any pre-existing 3D or 2D collections, yet is capable of producing high-quality 3D humans that are visually realistic, geometrically accurate and content-wise diverse to generalize to arbitrary humans." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.893, + 0.901 + ], + "angle": 0, + "content": "An overview of the proposed scheme is illustrated in Figure 2. We build a sequential pipeline with the following three modules: 3D generative modeling (3DGM), geometric sculpting (GS), and explicit texturing (ET). The first module synthesizes view-balanced, structured and diverse human images with known camera parameters. Subsequently, it learns a 3D generative model from these synthetic data, focusing on realistic appearance modeling (Section 3.1). To overcome the inaccuracy of the 3D shape, the GS module is incorporated during the inference process. It optimizes a hybrid representation with multi-view normal constraints to carve intricate mesh details (Section 3.2)." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9983" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.184 + ], + "angle": 0, + "content": "Additionally, the ET module is employed to disentangle explicit texture by utilizing semantical UV partitioning and a differentiable rasterizer (Section 3.3). By combining these modules, we are able to synthesize high-quality and faithful 3D human avatars by incorporating random noises or guided texts/images (Section 3.4)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.195, + 0.296, + 0.21 + ], + "angle": 0, + "content": "3.1. 3D generative modeling" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.218, + 0.47, + 0.385 + ], + "angle": 0, + "content": "Without any 3D or 2D collections, we develop a synthesis-based flow to learn a 3D generative module from 2D synthetic data. We start by instantiating a 3D scene through the projection of underlying 3D skeletons onto 2D pose images, utilizing accurate physical parameters (i.e., camera parameters). Subsequently, the projected 2D pose images serve as conditions to control the 2D diffusion model [54] for synthesizing view-balanced, diverse, and lifelike human images. Finally, we employ a triplane-based generator with enhanced designs to learn a generalizable 3D representation from the synthetic data. Details are described as follows." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.386, + 0.47, + 0.492 + ], + "angle": 0, + "content": "3D instantiation. Starting with a template body mesh (e.g., SMPL-X [40]) positioned and posed in canonical space, we estimate the 3D joint locations \\(\\mathcal{P}_{3d}\\) by regressing them from interpolated vertices. We then project \\(\\mathcal{P}_{3d}\\) onto 2D poses \\(\\mathcal{P}_i, i = 1,\\dots,k\\) from \\(\\kappa\\) horizontally uniformly sampled viewpoints \\(\\varphi\\). In this way, paired 2D pose images and their corresponding camera parameters \\(\\{\\mathcal{P}_i,\\varphi_i\\}\\) are formulated." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.492, + 0.47, + 0.718 + ], + "angle": 0, + "content": "Controlled 2D image synthesis. With the pose image \\(\\mathcal{P}_i\\), we feed it into off-the-shelf ControlNet [54] as the pose condition to guide diffusion models [44] to synthesize human images in desired poses (i.e., views). The text prompt \\(T\\) is also used for diverse contents. Given a prompt \\(T\\), instead of generating a human image \\(\\mathcal{I}_s: \\mathcal{I}_s = \\mathcal{C}(\\mathcal{P}_i, T)\\) independently for each view \\(\\varphi_i\\), we horizontally concatenate \\(\\mathcal{K}\\) pose images \\(\\mathcal{P}_i \\in R^{H \\times W \\times 3}\\), resulting in \\(\\mathcal{P}_i' \\in R^{H \\times KW \\times 3}\\) and feed \\(\\mathcal{P}_i'\\) to \\(\\mathcal{C}\\), along with a prompt hint of 'multi-view' in \\(T\\). In this way, multi-view human images \\(\\mathcal{I}_s'\\) are synthesized with roughly coherent appearance. We split \\(\\mathcal{I}_s'\\) to single view images \\(\\mathcal{I}_{\\varphi}\\) under specific views \\(\\varphi\\). This concatenation strategy facilitates the convergence of distributions in synthetic multi-views, thus easing the learning of common 3D representation meeting multi-view characteristics." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.72, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Generalizable 3D representation learning. With synthetic data of paired \\(\\{\\mathcal{I}_{\\varphi}, \\varphi\\}\\), we learn the 3D generative module \\(\\mathcal{G}_{3d}\\) from them to produce diverse 3D-aware human images with realistic appearance. Inspired by EG3D [5], we employ a triplane-based generator to produce a generalizable representation \\(\\mathcal{T}\\) and introduce a patch-composed neural renderer to learn intricate human representation efficiently. Specifically, instead of uniformly sampling 2D pixels on the image \\(\\mathcal{I}\\), we decompose patches in the ROI region including human bodies, and only emit rays towards pixels in these patches. The rays are rendered into RGB color with opacity values via volume rendering. Based on" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.198 + ], + "angle": 0, + "content": "the decomposed rule, we decode rendered colors to multiple patches and re-combine these patches for full feature images. In this way, the representation is composed of effective human body parts, which directs the attention of the networks towards the human subject itself. This design facilitates fine-grained local human learning while maintaining computational efficiency." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.198, + 0.892, + 0.273 + ], + "angle": 0, + "content": "For the training process, we employ two discriminators, one for RGB images and another for silhouettes, which yields better disentanglement of foreground objects with global geometry. The training loss for this module \\( L_{3d} \\) consists of the two adversarial terms:" + }, + { + "type": "equation", + "bbox": [ + 0.524, + 0.285, + 0.892, + 0.302 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {3 d} = \\mathcal {L} _ {a d v} \\left(\\mathcal {D} _ {r g b}, \\mathcal {G} _ {3 d}\\right) + \\lambda_ {s} \\mathcal {L} _ {a d v} \\left(\\mathcal {D} _ {\\text {m a s k}}, \\mathcal {G} _ {3 d}\\right), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.313, + 0.892, + 0.358 + ], + "angle": 0, + "content": "where \\(\\lambda_{s}\\) denotes the weight of silhouette item. \\(\\mathcal{L}_{adv}\\) is computed by the non-saturating GAN loss with R1 regularization [32]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.359, + 0.892, + 0.42 + ], + "angle": 0, + "content": "With the trained \\(\\mathcal{G}_{3d}\\), we can synthesize 3D-aware human images \\(\\mathcal{T}_g^\\varphi\\) with view control, and extract coarse 3D shapes \\(\\mathcal{M}_c\\) from the density field of neural renderer using the Marching Cubes algorithm [30]." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.429, + 0.693, + 0.446 + ], + "angle": 0, + "content": "3.2. Geometric sculpting" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.452, + 0.892, + 0.617 + ], + "angle": 0, + "content": "Our 3D generative module can produce high-quality and 3D-consistent human images in view controls. However, its training solely relies on discriminations made using 2D renderings, which can result in inaccuracies in capturing the inherent geometry, especially for complex human bodies. Therefore, we integrate the geometric sculpting, an optimization module leveraging geometric information from high-quality multi-views to carve surface details. Combined with a hybrid 3D representation and a differentiable rasterizer, it can rapidly enhance the shape quality within seconds." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.618, + 0.892, + 0.786 + ], + "angle": 0, + "content": "DMTET adaption. Owing to the expressive ability of arbitrary topologies and computational efficiency with direct shape optimization, we employ DMTET as our 3D representation in this module and adapt it to the coarse mesh \\(\\mathcal{M}_c\\) via an initial fitting procedure. Specifically, we parameterize DMTET as an MLP network \\(\\Psi_g\\) that learns to predict the SDF value \\(s(v_i)\\) and the position offset \\(\\delta v_i\\) for each vertex \\(v_i \\in VT\\) of the tetrahedral grid \\((VT,T)\\). A point set \\(P = \\{p_i \\in R^3\\}\\) is randomly sampled near \\(\\mathcal{M}_c\\) and their SDF values \\(SDF(p_i)\\) can be pre-computed. We adapt the parameters \\(\\psi\\) of \\(\\Psi_g\\) by fitting it to the SDF of \\(\\mathcal{M}_c\\):" + }, + { + "type": "equation", + "bbox": [ + 0.566, + 0.797, + 0.892, + 0.831 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {a d a} = \\sum_ {p _ {i} \\in P} | | s (p _ {i}; \\psi) - S D F (p _ {i}) | | _ {2}. \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.892, + 0.902 + ], + "angle": 0, + "content": "Geometry refinement. Using the adapted DMTET, we leverage the highly-detailed normal maps \\(\\mathcal{N}\\) derived from realistic multi-view images as a guidance to refine local surfaces. To obtain the pseudo-GT normals \\(\\mathcal{N}_{\\varphi}\\), we extract" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9984" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.091, + 0.473, + 0.246 + ], + "angle": 0, + "content": "them from \\(\\mathcal{I}_g^\\varphi\\) using a pre-trained normal estimator [50]. For the rendered normals \\(\\hat{\\mathcal{N}}_{\\varphi}\\), we extract the triangular mesh \\(\\mathcal{M}_{tri}\\) from \\((VT,T)\\) using the Marching Tetrahedra (MT) layer in our current DMTET. By rendering the generated mesh \\(\\mathcal{M}_{tri}\\) with differentiable rasterization, we obtain the resulting normal map \\(\\hat{\\mathcal{N}}_{\\varphi}\\). To ensure holistic surface polishing that takes into account multi-view normals, we randomly sample camera poses \\(\\varphi\\) that are uniformly distributed in space. We optimize the parameters of \\(\\Psi_g\\) using the normal loss, which is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.19, + 0.255, + 0.47, + 0.274 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {n o r m}} = \\left\\| \\hat {\\mathcal {N}} _ {\\varphi} - \\mathcal {N} _ {\\varphi} \\right\\| _ {2}. \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.282, + 0.47, + 0.375 + ], + "angle": 0, + "content": "After rapid optimization, the final triangular mesh \\(\\mathcal{M}_{tri}\\) can be easily extracted from the MT layer. If the hands exhibit noise, they can be optionally replaced with cleaner geometry hands from SMPL-X, benefiting from the alignment of the generated body in canonical space with the underlying template body." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.382, + 0.251, + 0.399 + ], + "angle": 0, + "content": "3.3. Explicit texturing" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.405, + 0.47, + 0.48 + ], + "angle": 0, + "content": "With the final mesh, the explicit texturing module aims to disentangle a UV texture map from multi-view renderings \\(\\mathcal{I}_g^\\varphi\\). This intuitive module not only facilitates the incorporation of high-fidelity textures but also enables various editing applications, as verified in Section 4.4." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.481, + 0.471, + 0.661 + ], + "angle": 0, + "content": "Given the polished triangular mesh \\(\\mathcal{M}_{tri}\\) and multi-views \\(\\mathcal{I}_g^\\varphi\\), we model the explicit texture map \\(T_{uv}\\) of \\(\\mathcal{M}_{tri}\\) with a semantic UV partition and optimize \\(T_{uv}\\) using a differentiable rasterizer \\(\\mathcal{R}\\) [25]. Specifically, leveraging the canonical properties of synthesized bodies, we semantically split \\(\\mathcal{M}_{tri}\\) into \\(\\gamma\\) components and rotate each component vertically, thus enabling effective UV projection for each component with cylinder unwarping. We then combine the texture partitions together for the full texture \\(T_{uv}\\). We optimize \\(T_{uv}\\) from a randomly initialized scratch using the texture loss, which consists of a multi-view reconstruction term and a total-variation (tv) term:" + }, + { + "type": "equation", + "bbox": [ + 0.194, + 0.673, + 0.47, + 0.689 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {t e x} = \\mathcal {L} _ {r e c} + \\lambda_ {t v} \\mathcal {L} _ {t v}, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.699, + 0.368, + 0.714 + ], + "angle": 0, + "content": "where \\(\\lambda_{tv}\\) denotes the weight of the tv loss." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.714, + 0.469, + 0.805 + ], + "angle": 0, + "content": "Multi-view guidance. To ensure comprehensive texturing in the 3D space, we render the color images \\(\\mathcal{R}(\\mathcal{M}_{tri},\\varphi)\\) and silhouettes \\(S\\) using \\(\\mathcal{R}\\) and optimize \\(T_{uv}\\) utilizing multi-view weighted guidance. Their pixel-alignment distances to the original multi-view renderings \\(\\mathcal{I}_g^\\varphi\\) are defined as the reconstruction loss:" + }, + { + "type": "equation", + "bbox": [ + 0.122, + 0.812, + 0.47, + 0.845 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {r e c} = \\sum_ {\\varphi \\in \\Omega} w _ {\\varphi} | | \\mathcal {R} \\left(\\mathcal {M} _ {t r i}, \\varphi\\right) \\cdot \\mathcal {S} - I _ {g} ^ {\\varphi} \\cdot \\mathcal {S} | | _ {2}, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.855, + 0.47, + 0.903 + ], + "angle": 0, + "content": "where \\(\\Omega\\) is the set of viewpoints \\(\\{\\varphi_i, i = 1, \\dots, k\\}\\) and \\(w_{\\varphi}\\) denotes weights of different views. \\(w_{\\varphi}\\) equals to 1.0 for \\(\\varphi \\in \\{front, back\\}\\) and 0.2 otherwise." + }, + { + "type": "image", + "bbox": [ + 0.501, + 0.088, + 0.895, + 0.219 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.223, + 0.892, + 0.252 + ], + "angle": 0, + "content": "Figure 3. The visualized flowchart of our method that synthesize textured 3D human avatars from input noises, texts or images." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.283, + 0.892, + 0.33 + ], + "angle": 0, + "content": "Smooth constraint. To avoid abrupt variations and smooth the generated texture \\( T_{uv} \\), we utilize the total-variation loss \\( \\mathcal{L}_{tv} \\) which is computed by:" + }, + { + "type": "equation", + "bbox": [ + 0.554, + 0.338, + 0.892, + 0.368 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {t v} = \\frac {1}{h \\times w \\times c} | | \\nabla_ {x} \\left(T _ {u v}\\right) + \\nabla_ {y} \\left(T _ {u v}\\right) | |, \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.377, + 0.868, + 0.392 + ], + "angle": 0, + "content": "where \\(x\\) and \\(y\\) denote horizontal and vertical directions." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.401, + 0.611, + 0.416 + ], + "angle": 0, + "content": "3.4. Inference" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.424, + 0.892, + 0.637 + ], + "angle": 0, + "content": "Built upon the above modules, we can generate high-quality 3D human avatars from either random noises or guided inputs such as texts or images. The flowchart for this process is shown in Figure 3. For input noises, we can easily obtain the final results by sequentially using the 3DGM, GS and ET modules. For text-guided synthesis, we first convert the text into a structured image using our controlled diffusion \\(\\mathcal{C}\\), and then inverse it to the latent space using PTI [43]. Specially, the GS and ET modules provide an interface that accurately reflects viewed modifications in the final 3D objects. As a result, we utilize the guided image to replace the corresponding view image, which results in improved fidelity in terms of geometry and texture. The same process is applied for input images as guided images." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.649, + 0.706, + 0.667 + ], + "angle": 0, + "content": "4. Experimental Results" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.675, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Implementation details. Our process begins by training the 3D generative module (3DGM) on synthetic data. During inference, we integrate the geometric sculpting (GS) and explicit texturing (ET) as optimization modules. For 3DGM, we normalize the template body to the \\((0,1)\\) space and place its center at the origin of the world coordinate system. We sample \\(7(\\mathcal{K} = 7)\\) viewpoints uniformly from the horizontal plane, ranging from \\(0^{\\circ}\\) to \\(180^{\\circ}\\) (front to back), with a camera radius of 2.7. For each viewpoint, we generate \\(100K\\) images using the corresponding pose image. To ensure diverse synthesis, we use detailed descriptions of age, gender, ethnicity, hairstyle, facial features, and clothing, leveraging a vast word bank. To cover \\(360^{\\circ}\\) views, we horizontally flip the synthesized images and obtain 1.4 million human images at a resolution of \\(512^{2}\\) in total. We train" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9985" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.078, + 0.089, + 0.895, + 0.526 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.307, + 0.526, + 0.66, + 0.54 + ], + "angle": 0, + "content": "Figure 4. Results of synthesized 3D human avatars at \\(512^{2}\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.561, + 0.473, + 0.76 + ], + "angle": 0, + "content": "the 3DGM for about 2.5M iterations with a batch size of 32, using two discriminators with a learning rate of 0.002 and a generator learning rate of 0.0025. The training takes 8 days using 8 NVIDIA Tesla-V100 GPUs. For GS, we optimize \\(\\psi\\) for 400 iterations for DMTET adaption and 100 iterations for surface carving (taking about 15s in total on 1 NVIDIA RTX 3090 GPU). For ET, we set \\(\\lambda_{uv} = 1\\) and optimize \\(T_{uv}\\) for 500 iterations (around 10 seconds). We split \\(\\mathcal{M}_{tri}\\) into \\(5(\\gamma = 5)\\) body parts (i.e., trunk, left/right arm/leg) with cylinder UV unwarping. We use the Adam optimizer with learning rates of 0.01 and 0.001 for \\(\\Psi_g\\) and \\(T_{uv}\\), respectively. Detailed network architectures can be found in the supplemental materials (Suppl)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.771, + 0.283, + 0.787 + ], + "angle": 0, + "content": "4.1. 3D human generation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.795, + 0.47, + 0.903 + ], + "angle": 0, + "content": "Figure 4 showcases several 3D human avatars synthesized by our pipeline, highlighting the image quality, geometry accuracy, and diverse outputs achieved through our method. Additionally, we explore the interpolation of the latent conditions to yield smooth transitions in appearance, leveraging the smooth latent space learned by our generative model. For more synthesized examples and interpolation results," + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.562, + 0.667, + 0.578 + ], + "angle": 0, + "content": "please refer to the Suppl." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.589, + 0.64, + 0.606 + ], + "angle": 0, + "content": "4.2. Comparisons" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.613, + 0.892, + 0.839 + ], + "angle": 0, + "content": "Qualitative comparison. In Figure 5, we compare our method with three baselines: EVA3D [15] and AG3D [9], which are state-of-the-art methods for generating 3D humans from 2D images, and EG3D [5], which serves as the foundational backbone of our method. The results of first two methods are produced by directly using source codes and trained models released by authors. We train EG3D using our synthetic images with estimated cameras from scratch. As we can see, EVA3D fails to produce \\(360^{\\circ}\\) humans with reasonable back inferring. AG3D and EG3D are able to generate \\(360^{\\circ}\\) renderings but both struggle with photorealism and capturing detailed shapes. Our method synthesizes not only higher-quality, view-consistent \\(360^{\\circ}\\) images but also higher-fidelity 3D geometry with intricate details, such as irregular dresses and haircuts." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.893, + 0.901 + ], + "angle": 0, + "content": "Quantitative comparison. Table 1 provides quantitative results comparing our method against the baselines. We measure image quality with Frechet Inception Distance (FID) [14] and Inception Score [45] for \\(360^{\\circ}\\) views (IS-" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "9986" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.086, + 0.088, + 0.894, + 0.479 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.16, + 0.48, + 0.807, + 0.495 + ], + "angle": 0, + "content": "Figure 5. Qualitative comparison with three state-of-the-art methods: EVA3D [15], AG3D [9] and EG3D [5]." + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.513, + 0.47, + 0.542 + ], + "angle": 0, + "content": "Table 1. Quantitative evaluation using FID, IS-360, normal accuracy (Normal) and identity consistency (ID)." + }, + { + "type": "table", + "bbox": [ + 0.1, + 0.544, + 0.445, + 0.658 + ], + "angle": 0, + "content": "
MethodFID ↓IS-360 ↑Normal ↓ID↑
EVA3D [15]15.913.1930.810.72
AG3D [9]10.933.2820.830.69
EVA3D-syn6.273.259.570.72
AG3D-syn8.933.228.640.71
EG3D [5]7.483.2612.740.71
Ours2.733.435.620.74
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.689, + 0.473, + 0.903 + ], + "angle": 0, + "content": "360). FID measures the visual similarity and distribution discrepancy between 50k generated images and all real images. IS-360 focuses on the self-realism of generated images in \\(360^{\\circ}\\) views. For shape evaluation, we compute FID between rendered normals and pseudo-GT normal maps (Normal), following AG3D. The FID and Normal scores of EVA3D and AG3D are directly fetched from their reports. Additionally, we access the multi-view facial identity consistency using the ID metric introduced by EG3D. For a clearer comparison, we also provide metric results of EVA3D and AG3D trained on our synthetic data (-syn). Overall, our method demonstrates significant improvements in FID and Normal, bringing the generative human model to a new level of realistic \\(360^{\\circ}\\) renderings with delicate ge" + }, + { + "type": "table_caption", + "bbox": [ + 0.498, + 0.513, + 0.892, + 0.556 + ], + "angle": 0, + "content": "Table 2. Results of models trained by replacing physical parameters with estimated ones (w/o SYN-P) or removing patch-composed rendering (w/o PCR)." + }, + { + "type": "table", + "bbox": [ + 0.52, + 0.558, + 0.871, + 0.617 + ], + "angle": 0, + "content": "
OursOurs-w/o SYN-POurs-w/o PCR
FID ↓2.734.283.26
IS-360 ↑3.433.313.35
" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.642, + 0.892, + 0.658 + ], + "angle": 0, + "content": "ometry while maintaining state-of-the-art view consistency." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.666, + 0.651, + 0.683 + ], + "angle": 0, + "content": "4.3. Ablation study" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.689, + 0.892, + 0.84 + ], + "angle": 0, + "content": "Synthesis flow and patch-composed rendering. We assess the impact of our carefully designed synthesis flow by training a model with synthetic images but with camera and pose parameters estimated by SMPLify-X [40] (w/o SYN-P). As Table 2 shows, the model w/o SYN-P results in worse FID and IS-360 scores, indicating that the synthesis flow contributes to more accurate physical parameters for realistic appearance modeling. By utilizing patch-composed rendering (PCR), the networks focus more on the human region, leading to more realistic results." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.894, + 0.903 + ], + "angle": 0, + "content": "Geometry sculpting module (GS). We demonstrate the importance of this module by visualizing the meshes before and after its implementation. Figure 6 (b) shows that the preceding module yields a coarse mesh due to the complex" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9987" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.086, + 0.092, + 0.18, + 0.198 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.088, + 0.2, + 0.166, + 0.212 + ], + "angle": 0, + "content": "(a) Rendering" + }, + { + "type": "image", + "bbox": [ + 0.182, + 0.092, + 0.28, + 0.198 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.195, + 0.2, + 0.258, + 0.211 + ], + "angle": 0, + "content": "(b) w/o GS" + }, + { + "type": "image", + "bbox": [ + 0.283, + 0.092, + 0.38, + 0.198 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.384, + 0.096, + 0.468, + 0.198 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.328, + 0.2, + 0.415, + 0.212 + ], + "angle": 0, + "content": "(c) Full method" + }, + { + "type": "image_caption", + "bbox": [ + 0.078, + 0.217, + 0.468, + 0.231 + ], + "angle": 0, + "content": "Figure 6. Effects of the GS module to carve fine-grained surfaces." + }, + { + "type": "image", + "bbox": [ + 0.092, + 0.248, + 0.185, + 0.354 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.088, + 0.356, + 0.182, + 0.368 + ], + "angle": 0, + "content": "(a) Guided image" + }, + { + "type": "image", + "bbox": [ + 0.19, + 0.248, + 0.28, + 0.355 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.205, + 0.356, + 0.265, + 0.368 + ], + "angle": 0, + "content": "(b) w/o ET" + }, + { + "type": "image", + "bbox": [ + 0.288, + 0.248, + 0.38, + 0.355 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.334, + 0.356, + 0.418, + 0.368 + ], + "angle": 0, + "content": "(c) Full method" + }, + { + "type": "image", + "bbox": [ + 0.384, + 0.248, + 0.468, + 0.355 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.105, + 0.373, + 0.441, + 0.387 + ], + "angle": 0, + "content": "Figure 7. Effects of the ET module for guided synthesis." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.415, + 0.468, + 0.536 + ], + "angle": 0, + "content": "human anatomy and the challenges posed by decomposing ambiguous 3D shapes from 2D images. The GS module utilizes high-quality multi-view outputs and employs a more flexible hybrid representation to create expressive humans with arbitrary topologies. It learns from pixel-level surface supervision, leading to a significant improvement in shape quality, characterized by smooth surfaces and intricate outfits (Figure 6 (c))." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.537, + 0.468, + 0.733 + ], + "angle": 0, + "content": "Explicit texturing module (ET). This intuitive module not only extracts the explicit UV texture for complete 3D assets but also enables high-fidelity results for image guided synthesis. Following the flowchart in Figure 3, we compare the results produced with and without this module. Our method without ET directly generates implicit renderings through PTI inversion, as shown in Figure 7 (b). While it successfully preserves global identity, it struggles to synthesize highly faithful local textures (e.g., floral patterns). The ET module offers a convenient and efficient way to directly interact with the 3D representation, enabling the production of high-fidelity 3D humans with more consistent content including exquisite local patterns (Figure 7 (a, c))." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.742, + 0.212, + 0.758 + ], + "angle": 0, + "content": "4.4. Applications" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.468, + 0.841 + ], + "angle": 0, + "content": "Avatar animation. All avatars produced by our method are in a canonical body pose and aligned to an underlying 3D skeleton extracted from SMPL-X. This alignment allows for easy animation and the generation of motion videos, as demonstrated in Figure 1 and Suppl." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.841, + 0.468, + 0.901 + ], + "angle": 0, + "content": "Texture doodle and local editing. Our approach benefits from the explicit disentanglement of geometry and texture, enabling flexible editing capabilities. Following the flowchart of text or image guided synthesis (Section 3.4)," + }, + { + "type": "image", + "bbox": [ + 0.502, + 0.09, + 0.664, + 0.19 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.665, + 0.091, + 0.805, + 0.189 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.807, + 0.091, + 0.891, + 0.189 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.502, + 0.191, + 0.661, + 0.256 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.661, + 0.191, + 0.805, + 0.256 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.807, + 0.191, + 0.891, + 0.256 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.259, + 0.892, + 0.3 + ], + "angle": 0, + "content": "Figure 8. Results synthesized by adapting our method to various styles (e.g., Disney cartoon characters) or contents (e.g., portrait heads)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.338, + 0.892, + 0.474 + ], + "angle": 0, + "content": "users can paint any pattern or add text to a guided image. These modifications can be transferred to 3D human models by inputting modified views into the texture module (e.g., painting the text 'hey' on a jacket as shown in Figure 1 (d)). Our approach also allows for clothing editing by simultaneously injecting edited guide images with desired clothing into the GS and ET modules (e.g., changing a jacket and jeans to bodysuits in Figure 1 (e)). More results can be found in Suppl." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.476, + 0.892, + 0.596 + ], + "angle": 0, + "content": "Content-style free adaption. Our proposed scheme is versatile and can be extended to generate various types of contents (e.g., portrait heads) and styles (e.g., Disney cartoon characters). To achieve this, we fine-tune our model using synthetic images from these domains, allowing for flexible adaptation. We showcase the results in Figure 8. More results and other discussions (e.g., limitations, negative impact, etc.) can be found in the Suppl." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.617, + 0.619, + 0.632 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.644, + 0.892, + 0.901 + ], + "angle": 0, + "content": "We introduced En3D, a novel generative scheme for sculpting 3D humans from 2D synthetic data. This method overcomes limitations in existing 3D or 2D collections and significantly enhances the image quality, geometry accuracy, and content diversity of generated 3D humans. En3D comprises a 3D generative module that learns generalizable 3D humans from synthetic 2D data with accurate physical modeling, and two optimization modules to carve intricate shape details and disentangle explicit UV textures with high fidelity, respectively. Experimental results validated the superiority and effectiveness of our method. We also demonstrated the flexibility of our generated avatars for animation and editing, as well as the scalability of our approach for synthesizing portraits and Disney characters. We believe that our solution could provide invaluable human assets for the 3D vision community. Furthermore, it holds potential for use in common 3D object synthesis tasks." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9988" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.105 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.171 + ], + "angle": 0, + "content": "[1] Panos Achlioptas, Olga Diamanti, Ioannis Mitliagkas, and Leonidas Guibas. Learning representations and generative models for 3d point clouds. In International conference on machine learning, pages 40-49. PMLR, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.173, + 0.472, + 0.243 + ], + "angle": 0, + "content": "[2] Thiemo Alldieck, Marcus Magnor, Weipeng Xu, Christian Theobalt, and Gerard Pons-Moll. Video based reconstruction of 3d people models. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 8387-8397, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.245, + 0.472, + 0.3 + ], + "angle": 0, + "content": "[3] Dragomir Anguelov, Praveen Srinivasan, Daphne Koller, Sebastian Thrun, Jim Rodgers, and James Davis. Scape: shape completion and animation of people. In ACM SIGGRAPH 2005 Papers, pages 408-416. 2005. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.302, + 0.472, + 0.372 + ], + "angle": 0, + "content": "[4] Eric R Chan, Marco Monteiro, Petr Kellnhofer, Jiajun Wu, and Gordon Wetzstein. pi-gan: Periodic implicit generative adversarial networks for 3d-aware image synthesis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5799-5809, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.373, + 0.472, + 0.47 + ], + "angle": 0, + "content": "[5] Eric R Chan, Connor Z Lin, Matthew A Chan, Koki Nagano, Boxiao Pan, Shalini De Mello, Orazio Gallo, Leonidas J Guibas, Jonathan Tremblay, Sameh Khamis, et al. Efficient geometry-aware 3d generative adversarial networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16123-16133, 2022. 2, 3, 4, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.472, + 0.472, + 0.543 + ], + "angle": 0, + "content": "[6] Xu Chen, Tianjian Jiang, Jie Song, Jinlong Yang, Michael J Black, Andreas Geiger, and Otmar Hilliges. gdna: Towards generative detailed neural avatars. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20427-20437, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.545, + 0.472, + 0.613 + ], + "angle": 0, + "content": "[7] Yu Deng, Jiaolong Yang, Jianfeng Xiang, and Xin Tong. Gram: Generative radiance manifolds for 3d-aware image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10673-10683, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.615, + 0.472, + 0.685 + ], + "angle": 0, + "content": "[8] Zijian Dong, Chen Guo, Jie Song, Xu Chen, Andreas Geiger, and Otmar Hilliges. Pina: Learning a personalized implicit neural avatar from a single rgb-d video sequence. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20470-20480, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.687, + 0.472, + 0.743 + ], + "angle": 0, + "content": "[9] Zijian Dong, Xu Chen, Jinlong Yang, Michael J Black, Otmar Hilliges, and Andreas Geiger. Ag3d: Learning to generate 3d avatars from 2d image collections. arXiv preprint arXiv:2305.02312, 2023. 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.745, + 0.47, + 0.8 + ], + "angle": 0, + "content": "[10] Yao Feng, Jinlong Yang, Marc Pollefeys, Michael J Black, and Timo Bolkart. Capturing and animation of body and clothing from monocular video. In SIGGRAPH Asia 2022 Conference Papers, pages 1-9, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.802, + 0.47, + 0.871 + ], + "angle": 0, + "content": "[11] Jianglin Fu, Shikai Li, Yuming Jiang, Kwan-Yee Lin, Chen Qian, Chen Change Loy, Wayne Wu, and Ziwei Liu. Stylegan-human: A data-centric odyssey of human generation. In European Conference on Computer Vision, pages 1-19. Springer, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.47, + 0.902 + ], + "angle": 0, + "content": "[12] Jiatao Gu, Lingjie Liu, Peng Wang, and Christian Theobalt. Stylenerf: A style-based 3d-aware genera" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.472, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "tor for high-resolution image synthesis. arXiv preprint arXiv:2110.08985, 2021.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.122, + 0.892, + 0.176 + ], + "angle": 0, + "content": "[13] Philipp Henzler, Niloy J Mitra, and Tobias Ritschel. Escaping plato's cave: 3d shape from adversarial rendering. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9984-9993, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.178, + 0.892, + 0.245 + ], + "angle": 0, + "content": "[14] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.247, + 0.892, + 0.302 + ], + "angle": 0, + "content": "[15] Fangzhou Hong, Zhaoxi Chen, Yushi Lan, Liang Pan, and Ziwei Liu. Eva3d: Compositional 3d human generation from 2d image collections. arXiv preprint arXiv:2210.04888, 2022. 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.304, + 0.892, + 0.357 + ], + "angle": 0, + "content": "[16] Xin Huang, Ruizhi Shao, Qi Zhang, Hongwen Zhang, Ying Feng, Yebin Liu, and Qing Wang. Humannorm: Learning normal diffusion model for high-quality and realistic 3d human generation. arXiv preprint arXiv:2310.01406, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.359, + 0.892, + 0.428 + ], + "angle": 0, + "content": "[17] Suyi Jiang, Haoran Jiang, Ziyu Wang, Haimin Luo, Wenzheng Chen, and Lan Xu. Humangen: Generating human radiance fields with explicit priors. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12543-12554, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.429, + 0.892, + 0.496 + ], + "angle": 0, + "content": "[18] Kyungmin Jo, Wonjoon Jin, Jaegul Choo, Hyunjoon Lee, and Sunghyun Cho. 3d-aware generative model for improved side-view image synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 22862-22872, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.498, + 0.892, + 0.552 + ], + "angle": 0, + "content": "[19] Hanbyul Joo, Tomas Simon, and Yaser Sheikh. Total capture: A 3d deformation model for tracking faces, hands, and bodies. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 8320-8329, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.554, + 0.892, + 0.608 + ], + "angle": 0, + "content": "[20] Angjoo Kanazawa, Michael J Black, David W Jacobs, and Jitendra Malik. End-to-end recovery of human shape and pose. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7122-7131, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.61, + 0.892, + 0.664 + ], + "angle": 0, + "content": "[21] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4401-4410, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.665, + 0.892, + 0.734 + ], + "angle": 0, + "content": "[22] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of stylegan. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8110-8119, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.735, + 0.892, + 0.789 + ], + "angle": 0, + "content": "[23] Tero Karras, Miika Aittala, Samuli Laine, Erik Härkönen, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Alias-free generative adversarial networks. Advances in Neural Information Processing Systems, 34:852-863, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.79, + 0.892, + 0.845 + ], + "angle": 0, + "content": "[24] Nikos Kolotouros, Thiemo Alldieck, Andrei Zanfir, Eduard Bazavan, Mihai Fieraru, and Cristian Sminchisescu. Dreamhuman: Animatable 3d avatars from text. Advances in Neural Information Processing Systems, 36, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.847, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[25] Samuli Laine, Janne Hellsten, Tero Karras, Yeongho Seol, Jaakko Lehtinen, and Timo Aila. Modular primitives for high-performance differentiable rendering. ACM Transactions on Graphics (TOG), 39(6):1-14, 2020. 5" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.956 + ], + "angle": 0, + "content": "9989" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.16 + ], + "angle": 0, + "content": "[26] Ruihui Li, Xianzhi Li, Chi-Wing Fu, Daniel Cohen-Or, and Pheng-Ann Heng. Pu-gan: a point cloud upsampling adversarial network. In Proceedings of the IEEE/CVF international conference on computer vision, pages 7203–7212, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.163, + 0.472, + 0.232 + ], + "angle": 0, + "content": "[27] Yiyi Liao, Katja Schwarz, Lars Mescheder, and Andreas Geiger. Towards unsupervised learning of generative models for 3d controllable image synthesis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5871-5880, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.233, + 0.471, + 0.3 + ], + "angle": 0, + "content": "[28] Lingjie Liu, Marc Habermann, Viktor Rudnev, Kripasindhu Sarkar, Jiatao Gu, and Christian Theobalt. Neural actor: Neural free-view synthesis of human actors with pose control. ACM transactions on graphics (TOG), 40(6):1-16, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.302, + 0.47, + 0.356 + ], + "angle": 0, + "content": "[29] Matthew Loper, Naureen Mahmood, Javier Romero, Gerard Pons-Moll, and Michael J Black. Smpl: A skinned multiperson linear model. ACM Transactions on Graphics, 34(6), 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.358, + 0.47, + 0.412 + ], + "angle": 0, + "content": "[30] William E Lorensen and Harvey E Cline. Marching cubes: A high resolution 3d surface construction algorithm. In Semin al graphics: pioneering efforts that shaped the field, pages 347-353. 1998. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.414, + 0.469, + 0.483 + ], + "angle": 0, + "content": "[31] Qianli Ma, Jinlong Yang, Anurag Ranjan, Sergi Pujades, Gerard Pons-Moll, Siyu Tang, and Michael J Black. Learning to dress 3d people in generative clothing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6469-6478, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.484, + 0.469, + 0.538 + ], + "angle": 0, + "content": "[32] Lars Mescheder, Andreas Geiger, and Sebastian Nowozin. Which training methods for gans do actually converge? In International conference on machine learning, pages 3481-3490. PMLR, 2018. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.54, + 0.469, + 0.607 + ], + "angle": 0, + "content": "[33] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.609, + 0.469, + 0.678 + ], + "angle": 0, + "content": "[34] Thu Nguyen-Phuoc, Chuan Li, Lucas Theis, Christian Richardt, and Yong-Liang Yang. Hologan: Unsupervised learning of 3d representations from natural images. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7588-7597, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.679, + 0.469, + 0.747 + ], + "angle": 0, + "content": "[35] Thu H Nguyen-Phuoc, Christian Richardt, Long Mai, Yongliang Yang, and Niloy Mitra. Blockgan: Learning 3d object-aware scene representations from unlabelled images. Advances in neural information processing systems, 33:6767-6778, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.748, + 0.469, + 0.815 + ], + "angle": 0, + "content": "[36] Michael Niemeyer and Andreas Geiger. Giraffe: Representing scenes as compositional generative neural feature fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11453-11464, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.818, + 0.469, + 0.9 + ], + "angle": 0, + "content": "[37] Roy Or-El, Xuan Luo, Mengyi Shan, Eli Shechtman, Jeong Joon Park, and Ira Kemelmacher-Shlizerman. Stylesdf: High-resolution 3d-consistent image and geometry generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13503–13513, 2022. 2, 3" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.162 + ], + "angle": 0, + "content": "[38] Ahmed AA Osman, Timo Bolkart, and Michael J Black. Star: Sparse trained articulated human body regressor. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part VI 16, pages 598-613. Springer, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.163, + 0.892, + 0.231 + ], + "angle": 0, + "content": "[39] Pablo Palafox, Aljaž Božić, Justus Thies, Matthias Nießner, and Angela Dai. Npms: Neural parametric models for 3d deformable shapes. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12695-12705, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.232, + 0.893, + 0.315 + ], + "angle": 0, + "content": "[40] Georgios Pavlakos, Vasileios Choutas, Nima Ghorbani, Timo Bolkart, Ahmed AA Osman, Dimitrios Tzionas, and Michael J Black. Expressive body capture: 3d hands, face, and body from a single image. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10975-10985, 2019. 4, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.316, + 0.893, + 0.399 + ], + "angle": 0, + "content": "[41] Sida Peng, Yuanqing Zhang, Yinghao Xu, Qianqian Wang, Qing Shuai, Hujun Bao, and Xiaowei Zhou. Neural body: Implicit neural representations with structured latent codes for novel view synthesis of dynamic humans. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9054-9063, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.4, + 0.892, + 0.44 + ], + "angle": 0, + "content": "[42] Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. arXiv preprint arXiv:2209.14988, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.442, + 0.892, + 0.495 + ], + "angle": 0, + "content": "[43] Daniel Roich, Ron Mokady, Amit H Bermano, and Daniel Cohen-Or. Pivotal tuning for latent-based editing of real images. ACM Transactions on graphics (TOG), 42(1):1-13, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.497, + 0.893, + 0.566 + ], + "angle": 0, + "content": "[44] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.567, + 0.892, + 0.622 + ], + "angle": 0, + "content": "[45] Tim Salimans, Ian Goodfellow, Wojciech Zaremba, Vicki Cheung, Alec Radford, and Xi Chen. Improved techniques for training gans. Advances in neural information processing systems, 29, 2016. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.623, + 0.892, + 0.678 + ], + "angle": 0, + "content": "[46] Katja Schwarz, Yiyi Liao, Michael Niemeyer, and Andreas Geiger. Graf: Generative radiance fields for 3d-aware image synthesis. Advances in Neural Information Processing Systems, 33:20154-20166, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.679, + 0.892, + 0.719 + ], + "angle": 0, + "content": "[47] Attila Szabó, Givi Meishvili, and Paolo Favaro. Unsupervised generative 3d shape learning from natural images. arXiv preprint arXiv:1910.00287, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.721, + 0.892, + 0.802 + ], + "angle": 0, + "content": "[48] Chung-Yi Weng, Brian Curless, Pratul P Srinivasan, Jonathan T Barron, and Ira Kemelmacher-Shlizerman. Humanerf: Free-viewpoint rendering of moving people from monocular video. In Proceedings of the IEEE/CVF conference on computer vision and pattern Recognition, pages 16210-16220, 2022. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.804, + 0.892, + 0.871 + ], + "angle": 0, + "content": "[49] Jiajun Wu, Chengkai Zhang, Tianfan Xue, Bill Freeman, and Josh Tenenbaum. Learning a probabilistic latent space of object shapes via 3d generative-adversarial modeling. Advances in neural information processing systems, 29, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.873, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[50] Yuliang Xiu, Jinlong Yang, Dimitrios Tzionas, and Michael J Black. Icon: Implicit clothed humans obtained from nor" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.893, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.956 + ], + "angle": 0, + "content": "9990" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.092, + 0.468, + 0.133 + ], + "angle": 0, + "content": "mals. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13286-13296. IEEE, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.136, + 0.469, + 0.205 + ], + "angle": 0, + "content": "[51] Yang Xue, Yuheng Li, Krishna Kumar Singh, and Yong Jae Lee. Giraffe hd: A high-resolution 3d-aware generative model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18440-18449, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.207, + 0.469, + 0.274 + ], + "angle": 0, + "content": "[52] Zhuoqian Yang, Shikai Li, Wayne Wu, and Bo Dai. 3dhumanan: 3d-aware human image generation with 3d pose mapping. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 23008-23019, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.277, + 0.469, + 0.346 + ], + "angle": 0, + "content": "[53] Jianfeng Zhang, Zihang Jiang, Dingdong Yang, Hongyi Xu, Yichun Shi, Guoxian Song, Zhongcong Xu, Xinchao Wang, and Jiashi Feng. Avatargen: a 3d generative model for animatable human avatars. In European Conference on Computer Vision, pages 668-685. Springer, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.348, + 0.469, + 0.403 + ], + "angle": 0, + "content": "[54] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3836-3847, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.405, + 0.469, + 0.458 + ], + "angle": 0, + "content": "[55] Peng Zhou, Lingxi Xie, Bingbing Ni, and Qi Tian. Cips-3d: A 3d-aware generator of gans based on conditionally-independent pixel synthesis. arXiv preprint arXiv:2110.09788, 2021. 3" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.469, + 0.458 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.514, + 0.956 + ], + "angle": 0, + "content": "9991" + } + ] +] \ No newline at end of file diff --git a/2024/En3D_ An Enhanced Generative Model for Sculpting 3D Humans from 2D Synthetic Data/cb7c8896-3bc3-4bc4-a3cb-b2aedebff579_origin.pdf b/2024/En3D_ An Enhanced Generative Model for Sculpting 3D Humans from 2D Synthetic Data/cb7c8896-3bc3-4bc4-a3cb-b2aedebff579_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..92a9480314fd3d185452925292ddd2c0696f0e23 --- /dev/null +++ b/2024/En3D_ An Enhanced Generative Model for Sculpting 3D Humans from 2D Synthetic Data/cb7c8896-3bc3-4bc4-a3cb-b2aedebff579_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a3cc38595e8f3235bd03de471104ef36b86d6062c7786f24ee61e5585551683b +size 8417079 diff --git a/2024/En3D_ An Enhanced Generative Model for Sculpting 3D Humans from 2D Synthetic Data/full.md b/2024/En3D_ An Enhanced Generative Model for Sculpting 3D Humans from 2D Synthetic Data/full.md new file mode 100644 index 0000000000000000000000000000000000000000..dc69681479ec4ff9046d8cd783d55203994ed6eb --- /dev/null +++ b/2024/En3D_ An Enhanced Generative Model for Sculpting 3D Humans from 2D Synthetic Data/full.md @@ -0,0 +1,312 @@ +# En3D: An Enhanced Generative Model for Sculpting 3D Humans from 2D Synthetic Data + +Yifang Men1, Biwen Lei1, Yuan Yao1, Miaomiao Cui1, Zhouhui Lian2, Xuansong Xie1 + +$^{1}$ Institute for Intelligent Computing, Alibaba Group + +$^{2}$ Wangxuan Institute of Computer Technology, Peking University, China + +https://menyifang.github.io/projects/En3D/index.html + +![](images/983670dcb4b85364dad9a3570384c18c010d456f295654df6337a5da0fd591ba.jpg) +(a) +(b) + +"a 44-year-old Latino man with blonde side part haircut, wearing a bomber jacket with brooch" + +![](images/76e7ccab3e14de71b3795b0179e19c701cf0088e680e8c398646ea720bf08bde.jpg) + +![](images/402cb4a2fea8a39cad637b91faa73ad3935b124bcac012ab790e2cdccdbc4e8b.jpg) +(c) +Figure 1. Given random noises or guided texts, our generative scheme can synthesize high-fidelity 3D human avatars that are visually realistic and geometrically accurate. These avatars can be seamlessly animated and easily edited. Our model is trained on 2D synthetic data without directly relying on any pre-existing 3D or 2D collections. + +![](images/b39a44ea2412c7c1054948a143823a1c704380e2114d9e8e16d8734ec5d15d7d.jpg) +(d) +(e) + +![](images/1f26002baed92f7862f31bd212cc40f3e1470aa0d9f7064e46d2a173cd05344f.jpg) + +# Abstract + +We present En3D, an enhanced generative scheme for sculpting high-quality 3D human avatars. Unlike previous works that rely on scarce 3D datasets or limited 2D collections with imbalanced viewing angles and imprecise pose priors, our approach aims to develop a zero-shot 3D generative scheme capable of producing visually realistic, geometrically accurate and content-wise diverse 3D humans without directly relying on pre-existing 3D or 2D assets. To address this challenge, we introduce a meticulously crafted workflow that implements accurate physical modeling to learn the enhanced 3D generative model from synthetic 2D data. During inference, we integrate optimization modules to bridge the gap between realistic appearances and coarse 3D shapes. Specifically, En3D comprises three modules: a 3D generator that accurately models generalizable 3D humans with realistic appearance from synthesized balanced, diverse, and structured human images; a geometry sculptor + +that enhances shape quality using multi-view normal constraints for intricate human structure; and a texturing module that disentangles explicit texture maps with fidelity and editability, leveraging semantical UV partitioning and a differentiable rasterizer. Experimental results show that our approach significantly outperforms prior works in terms of image quality, geometry accuracy and content diversity. We also showcase the applicability of our generated avatars for animation and editing, as well as the scalability of our approach for content-style free adaptation. + +# 1. Introduction + +3D human avatars play an important role in various applications of AR/VR such as video games, telepresence and virtual try-on. Realistic human modeling is an essential task, and many valuable efforts have been made by leveraging neural implicit fields to learn high-quality articulated avatars [8, 10, 41, 48]. However, these methods are directly + +learned from monocular videos or image sequences, where subjects are single individuals wearing specific garments, thus limiting their scalability. + +Generative models learn a shared 3D representation to synthesize clothed humans with varying identities, clothing and poses. Traditional methods are typically trained on 3D datasets, which are limited and expensive to acquire. This data scarcity limits the model's generalization ability and may lead to overfitting on small datasets. Recently, 3D-aware image synthesis methods [5, 18, 37] have demonstrated great potential in learning 3D generative models of rigid objects from 2D image collections. Follow-up works show the feasibility of learning articulated humans from image collections driven by SMPL-based deformations, but only in limited quality and resolution. EVA3D [15] represents humans as a composition of multiple parts with NeRF representations. AG3D [9] incorporates an efficient articulation module to capture both body shape and cloth deformation. Nevertheless, there remains a noticeable gap between generated and real humans in terms of appearance and geometry. Moreover, their results are limited to specific views (i.e., frontal angles) and lack diversity (i.e., fashion images in similar skin tone, body shape, and age). + +The aim of this paper is to propose a zero-shot 3D generative scheme that does not directly rely on any pre-existing 3D or 2D datasets, yet is capable of producing high-quality 3D humans that are visually realistic, geometrically accurate, and content-wise diverse. The generated avatars can be seamlessly animated and easily edited. An illustration is provided in Figure 1. To address this challenging task, our proposed method inherits from 3D-aware human image synthesis and exhibits substantial distinctions based on several key insights. Rethinking the nature of 3D-aware generative methods from 2D collections [5, 9, 15], they actually try to learn a generalizable and deformable 3D representation, whose 2D projections can meet the distribution of human images in corresponding views. Thereby, it is crucial for accurate physical modeling between 3D objects and 2D projections. However, previous works typically leverage pre-existing 2D human images to estimate physical parameters (i.e., camera and body poses), which are inaccurate because of imprecise SMPL priors for highly-articulated humans. This inaccuracy limits the synthesis ability for realistic multi-view renderings. Second, these methods solely rely on discriminating 2D renderings, which is ambiguous and loose to capture inherent 3D shapes in detail, especially for intricate human appearance. + +To address these limitations, we propose a novel generative scheme with two core designs. Firstly, we introduce a meticulously-crafted workflow that implements accurate physical modeling to learn an enhanced 3D generative model from synthetic data. This is achieved by instantiating a 3D body scene and projecting the underlying + +3D skeletons into 2D pose images using explicit camera parameters. These 2D pose images act as conditions to control a 2D diffusion model, synthesizing realistic human images from specific viewpoints. By leveraging synthetic view-balanced, diverse and structured human images, along with known physical parameters, we employ a 3D generator equipped with an enhanced renderer and discriminator to learn realistic appearance modeling. Secondly, we improve the 3D shape quality by leveraging the gap between high-quality multi-view renderings and the coarse mesh produced by the 3D generative module. Specifically, we integrate an optimization module that utilizes multi-view normal constraints to rapidly refine geometry details under supervision. Additionally, we incorporate an explicit texturing module to ensure faithful UV texture maps. In contrast to previous works that rely on inaccurate physical settings and inadequate shape supervision, we rebuild the generative scheme from the ground up, resulting in comprehensive improvements in image quality, geometry accuracy, and content diversity. In summary, our contributions are threefold: + +- We present a zero-shot generative scheme that efficiently synthesizes high-quality 3D human avatars with visual realism, geometric accuracy and content diversity. These avatars can be seamlessly animated and easily edited, offering greater flexibility in their applications. +- We develop a meticulously-crafted workflow to learn an enhanced generative model from synthesized human images that are balanced, diverse, and also possess known physical parameters. This leads to diverse 3D-aware human image synthesis with realistic appearance. +- We propose to integrate optimization modules into the 3D generator, leveraging multi-view guidance to enhance both shape quality and texture fidelity, thus achieving realistic 3D human assets. + +# 2. Related work + +3D Human Modeling. Parametric models [3, 19, 20, 29, 38] serve as a common representation for 3D human modeling, they allows for robust control by deforming a template mesh with a series of low-dimensional parameters, but can only generate naked 3D humans. Similar ideas have been extended to model clothed humans [2, 31], but geometric expressivity is restricted due to the fixed mesh topology. Subsequent works [6, 39, 39] further introduce implicit surfaces to produce complex non-linear deformations of 3D bodies. Unfortunately, the aforementioned approaches all require 3D scans of various human poses for model fitting, which are difficult to acquire. With the explosion of NeRF, valuable efforts have been made towards combining NeRF models with explicit human models [8, 10, 28, 41, 48]. Neural body [41] anchors a set of latent codes to the vertices of the SMPL model [29] and transforms the spatial locations of the codes to the volume in the observation space. Human- + +![](images/7d51faa41e042c861d61716c00ef486d05b132f800883da5657afb6ca6e7ccd4.jpg) +Figure 2. An overview of the proposed scheme, which consists of three modules: 3D generative modeling (3DGM), geometric sculpting (GS), and explicit texturing (ET). 3DGM employs synthesized diverse, balanced and structured human images with accurate camera parameters $\varphi$ to learn generalizable 3D humans with the triplane-based architecture. GS is integrated as an optimization module by utilizing multi-view normal constraints to refine and carve geometry details. ET utilizes UV partitioning and a differentiable rasterizer to disentangle explicit UV texture maps. Not only multi-view renderings but also realistic 3D models can be acquired finally by our method. + +NeRF [48] optimizes for a canonical, volumetric T-pose of the human with a motion field to map the non-rigid transformations. Text to 3D humans [16, 24] are introduced by combining SDS loss [42] with diffusion guidance. Nevertheless, these methods are learned directly from image sequences or guided by single text prompts, where subjects are fitting to single scenes, thus limiting their scalability. + +Generative 3D-aware Image Synthesis. Recently, 3D-aware image synthesis methods have lifted image generation with explicit view control by integrating the 2D generative models [21-23] with 3D representations, such as voxels [13, 34, 35, 49], meshes [27, 47] and points clouds [1, 26]. GRAF [46] and $\pi$ -GAN[4] firstly integrate the implicit representation networks, i.e., NeRF [33], with differentiable volumetric rendering for 3D scene generation. However, they have difficulties in training on high-resolution images due to the costly rendering process. Subsequent works have sought to improve the efficiency and quality of such NeRF-based GANs, either by adopting a two-stage rendering process [5, 12, 36, 37, 51] or a smart sampling strategy [7, 55]. StyleSDF [37] combines a SDF-based volume renderer and a 2D StyleGAN network [22] for photorealistic image generation. EG3D [5] introduces a superior triplane representation to leverage 2D CNN-based feature generators for efficient generalization over 3D spaces. Although these methods demonstrate impressive quality in view-consistent image synthesis, they are limited to simplified rigid objects such as faces, cats and cars. + +To learn highly articulated humans from unstructured 2D images, recent works [9, 11, 15, 17, 52, 53] integrate the deformation field to learn non-rigid deformations based on the body prior of estimated SMPL parameters. EVA3D [15] + +represents humans as a composition of multiple parts with NeRF representations. Instead of directly rendering the image from a 3D representation, 3DHumanGAN [52] uses an equivariant 2D generator modulated by 3D human body prior, which enables to establish one-to-many mapping from 3D geometry to synthesized textures from 2D images. AG3D [9] combines the 3D generator with an efficient articulation module to warp 3D objects from canonical space to posed space via a learned continuous deformation field. However, a gap still exists between the generated and real humans in terms of appearance, due to the imprecise priors from complex poses as well as the data biases from limited human poses and imbalanced viewing angles in the dataset. + +# 3. Method Description + +Our goal is to develop a zero-shot 3D generative scheme that does not directly rely on any pre-existing 3D or 2D collections, yet is capable of producing high-quality 3D humans that are visually realistic, geometrically accurate and content-wise diverse to generalize to arbitrary humans. + +An overview of the proposed scheme is illustrated in Figure 2. We build a sequential pipeline with the following three modules: 3D generative modeling (3DGM), geometric sculpting (GS), and explicit texturing (ET). The first module synthesizes view-balanced, structured and diverse human images with known camera parameters. Subsequently, it learns a 3D generative model from these synthetic data, focusing on realistic appearance modeling (Section 3.1). To overcome the inaccuracy of the 3D shape, the GS module is incorporated during the inference process. It optimizes a hybrid representation with multi-view normal constraints to carve intricate mesh details (Section 3.2). + +Additionally, the ET module is employed to disentangle explicit texture by utilizing semantical UV partitioning and a differentiable rasterizer (Section 3.3). By combining these modules, we are able to synthesize high-quality and faithful 3D human avatars by incorporating random noises or guided texts/images (Section 3.4). + +# 3.1. 3D generative modeling + +Without any 3D or 2D collections, we develop a synthesis-based flow to learn a 3D generative module from 2D synthetic data. We start by instantiating a 3D scene through the projection of underlying 3D skeletons onto 2D pose images, utilizing accurate physical parameters (i.e., camera parameters). Subsequently, the projected 2D pose images serve as conditions to control the 2D diffusion model [54] for synthesizing view-balanced, diverse, and lifelike human images. Finally, we employ a triplane-based generator with enhanced designs to learn a generalizable 3D representation from the synthetic data. Details are described as follows. + +3D instantiation. Starting with a template body mesh (e.g., SMPL-X [40]) positioned and posed in canonical space, we estimate the 3D joint locations $\mathcal{P}_{3d}$ by regressing them from interpolated vertices. We then project $\mathcal{P}_{3d}$ onto 2D poses $\mathcal{P}_i, i = 1,\dots,k$ from $\kappa$ horizontally uniformly sampled viewpoints $\varphi$ . In this way, paired 2D pose images and their corresponding camera parameters $\{\mathcal{P}_i,\varphi_i\}$ are formulated. + +Controlled 2D image synthesis. With the pose image $\mathcal{P}_i$ , we feed it into off-the-shelf ControlNet [54] as the pose condition to guide diffusion models [44] to synthesize human images in desired poses (i.e., views). The text prompt $T$ is also used for diverse contents. Given a prompt $T$ , instead of generating a human image $\mathcal{I}_s: \mathcal{I}_s = \mathcal{C}(\mathcal{P}_i, T)$ independently for each view $\varphi_i$ , we horizontally concatenate $\mathcal{K}$ pose images $\mathcal{P}_i \in R^{H \times W \times 3}$ , resulting in $\mathcal{P}_i' \in R^{H \times KW \times 3}$ and feed $\mathcal{P}_i'$ to $\mathcal{C}$ , along with a prompt hint of 'multi-view' in $T$ . In this way, multi-view human images $\mathcal{I}_s'$ are synthesized with roughly coherent appearance. We split $\mathcal{I}_s'$ to single view images $\mathcal{I}_{\varphi}$ under specific views $\varphi$ . This concatenation strategy facilitates the convergence of distributions in synthetic multi-views, thus easing the learning of common 3D representation meeting multi-view characteristics. + +Generalizable 3D representation learning. With synthetic data of paired $\{\mathcal{I}_{\varphi}, \varphi\}$ , we learn the 3D generative module $\mathcal{G}_{3d}$ from them to produce diverse 3D-aware human images with realistic appearance. Inspired by EG3D [5], we employ a triplane-based generator to produce a generalizable representation $\mathcal{T}$ and introduce a patch-composed neural renderer to learn intricate human representation efficiently. Specifically, instead of uniformly sampling 2D pixels on the image $\mathcal{I}$ , we decompose patches in the ROI region including human bodies, and only emit rays towards pixels in these patches. The rays are rendered into RGB color with opacity values via volume rendering. Based on + +the decomposed rule, we decode rendered colors to multiple patches and re-combine these patches for full feature images. In this way, the representation is composed of effective human body parts, which directs the attention of the networks towards the human subject itself. This design facilitates fine-grained local human learning while maintaining computational efficiency. + +For the training process, we employ two discriminators, one for RGB images and another for silhouettes, which yields better disentanglement of foreground objects with global geometry. The training loss for this module $L_{3d}$ consists of the two adversarial terms: + +$$ +\mathcal {L} _ {3 d} = \mathcal {L} _ {a d v} \left(\mathcal {D} _ {r g b}, \mathcal {G} _ {3 d}\right) + \lambda_ {s} \mathcal {L} _ {a d v} \left(\mathcal {D} _ {\text {m a s k}}, \mathcal {G} _ {3 d}\right), \tag {1} +$$ + +where $\lambda_{s}$ denotes the weight of silhouette item. $\mathcal{L}_{adv}$ is computed by the non-saturating GAN loss with R1 regularization [32]. + +With the trained $\mathcal{G}_{3d}$ , we can synthesize 3D-aware human images $\mathcal{T}_g^\varphi$ with view control, and extract coarse 3D shapes $\mathcal{M}_c$ from the density field of neural renderer using the Marching Cubes algorithm [30]. + +# 3.2. Geometric sculpting + +Our 3D generative module can produce high-quality and 3D-consistent human images in view controls. However, its training solely relies on discriminations made using 2D renderings, which can result in inaccuracies in capturing the inherent geometry, especially for complex human bodies. Therefore, we integrate the geometric sculpting, an optimization module leveraging geometric information from high-quality multi-views to carve surface details. Combined with a hybrid 3D representation and a differentiable rasterizer, it can rapidly enhance the shape quality within seconds. + +DMTET adaption. Owing to the expressive ability of arbitrary topologies and computational efficiency with direct shape optimization, we employ DMTET as our 3D representation in this module and adapt it to the coarse mesh $\mathcal{M}_c$ via an initial fitting procedure. Specifically, we parameterize DMTET as an MLP network $\Psi_g$ that learns to predict the SDF value $s(v_i)$ and the position offset $\delta v_i$ for each vertex $v_i \in VT$ of the tetrahedral grid $(VT,T)$ . A point set $P = \{p_i \in R^3\}$ is randomly sampled near $\mathcal{M}_c$ and their SDF values $SDF(p_i)$ can be pre-computed. We adapt the parameters $\psi$ of $\Psi_g$ by fitting it to the SDF of $\mathcal{M}_c$ : + +$$ +\mathcal {L} _ {a d a} = \sum_ {p _ {i} \in P} | | s (p _ {i}; \psi) - S D F (p _ {i}) | | _ {2}. \tag {2} +$$ + +Geometry refinement. Using the adapted DMTET, we leverage the highly-detailed normal maps $\mathcal{N}$ derived from realistic multi-view images as a guidance to refine local surfaces. To obtain the pseudo-GT normals $\mathcal{N}_{\varphi}$ , we extract + +them from $\mathcal{I}_g^\varphi$ using a pre-trained normal estimator [50]. For the rendered normals $\hat{\mathcal{N}}_{\varphi}$ , we extract the triangular mesh $\mathcal{M}_{tri}$ from $(VT,T)$ using the Marching Tetrahedra (MT) layer in our current DMTET. By rendering the generated mesh $\mathcal{M}_{tri}$ with differentiable rasterization, we obtain the resulting normal map $\hat{\mathcal{N}}_{\varphi}$ . To ensure holistic surface polishing that takes into account multi-view normals, we randomly sample camera poses $\varphi$ that are uniformly distributed in space. We optimize the parameters of $\Psi_g$ using the normal loss, which is defined as: + +$$ +\mathcal {L} _ {\text {n o r m}} = \left\| \hat {\mathcal {N}} _ {\varphi} - \mathcal {N} _ {\varphi} \right\| _ {2}. \tag {3} +$$ + +After rapid optimization, the final triangular mesh $\mathcal{M}_{tri}$ can be easily extracted from the MT layer. If the hands exhibit noise, they can be optionally replaced with cleaner geometry hands from SMPL-X, benefiting from the alignment of the generated body in canonical space with the underlying template body. + +# 3.3. Explicit texturing + +With the final mesh, the explicit texturing module aims to disentangle a UV texture map from multi-view renderings $\mathcal{I}_g^\varphi$ . This intuitive module not only facilitates the incorporation of high-fidelity textures but also enables various editing applications, as verified in Section 4.4. + +Given the polished triangular mesh $\mathcal{M}_{tri}$ and multi-views $\mathcal{I}_g^\varphi$ , we model the explicit texture map $T_{uv}$ of $\mathcal{M}_{tri}$ with a semantic UV partition and optimize $T_{uv}$ using a differentiable rasterizer $\mathcal{R}$ [25]. Specifically, leveraging the canonical properties of synthesized bodies, we semantically split $\mathcal{M}_{tri}$ into $\gamma$ components and rotate each component vertically, thus enabling effective UV projection for each component with cylinder unwarping. We then combine the texture partitions together for the full texture $T_{uv}$ . We optimize $T_{uv}$ from a randomly initialized scratch using the texture loss, which consists of a multi-view reconstruction term and a total-variation (tv) term: + +$$ +\mathcal {L} _ {t e x} = \mathcal {L} _ {r e c} + \lambda_ {t v} \mathcal {L} _ {t v}, \tag {4} +$$ + +where $\lambda_{tv}$ denotes the weight of the tv loss. + +Multi-view guidance. To ensure comprehensive texturing in the 3D space, we render the color images $\mathcal{R}(\mathcal{M}_{tri},\varphi)$ and silhouettes $S$ using $\mathcal{R}$ and optimize $T_{uv}$ utilizing multi-view weighted guidance. Their pixel-alignment distances to the original multi-view renderings $\mathcal{I}_g^\varphi$ are defined as the reconstruction loss: + +$$ +\mathcal {L} _ {r e c} = \sum_ {\varphi \in \Omega} w _ {\varphi} | | \mathcal {R} \left(\mathcal {M} _ {t r i}, \varphi\right) \cdot \mathcal {S} - I _ {g} ^ {\varphi} \cdot \mathcal {S} | | _ {2}, \tag {5} +$$ + +where $\Omega$ is the set of viewpoints $\{\varphi_i, i = 1, \dots, k\}$ and $w_{\varphi}$ denotes weights of different views. $w_{\varphi}$ equals to 1.0 for $\varphi \in \{front, back\}$ and 0.2 otherwise. + +![](images/02980479c6aeb301dc5c2200602c629f3b2063c308353b0567074f2c275e80c5.jpg) +Figure 3. The visualized flowchart of our method that synthesize textured 3D human avatars from input noises, texts or images. + +Smooth constraint. To avoid abrupt variations and smooth the generated texture $T_{uv}$ , we utilize the total-variation loss $\mathcal{L}_{tv}$ which is computed by: + +$$ +\mathcal {L} _ {t v} = \frac {1}{h \times w \times c} | | \nabla_ {x} \left(T _ {u v}\right) + \nabla_ {y} \left(T _ {u v}\right) | |, \tag {6} +$$ + +where $x$ and $y$ denote horizontal and vertical directions. + +# 3.4. Inference + +Built upon the above modules, we can generate high-quality 3D human avatars from either random noises or guided inputs such as texts or images. The flowchart for this process is shown in Figure 3. For input noises, we can easily obtain the final results by sequentially using the 3DGM, GS and ET modules. For text-guided synthesis, we first convert the text into a structured image using our controlled diffusion $\mathcal{C}$ , and then inverse it to the latent space using PTI [43]. Specially, the GS and ET modules provide an interface that accurately reflects viewed modifications in the final 3D objects. As a result, we utilize the guided image to replace the corresponding view image, which results in improved fidelity in terms of geometry and texture. The same process is applied for input images as guided images. + +# 4. Experimental Results + +Implementation details. Our process begins by training the 3D generative module (3DGM) on synthetic data. During inference, we integrate the geometric sculpting (GS) and explicit texturing (ET) as optimization modules. For 3DGM, we normalize the template body to the $(0,1)$ space and place its center at the origin of the world coordinate system. We sample $7(\mathcal{K} = 7)$ viewpoints uniformly from the horizontal plane, ranging from $0^{\circ}$ to $180^{\circ}$ (front to back), with a camera radius of 2.7. For each viewpoint, we generate $100K$ images using the corresponding pose image. To ensure diverse synthesis, we use detailed descriptions of age, gender, ethnicity, hairstyle, facial features, and clothing, leveraging a vast word bank. To cover $360^{\circ}$ views, we horizontally flip the synthesized images and obtain 1.4 million human images at a resolution of $512^{2}$ in total. We train + +![](images/7aed5913e1c5aa7b1793cb911d0316caf918d886877e533dd8741aae16715624.jpg) +Figure 4. Results of synthesized 3D human avatars at $512^{2}$ . + +the 3DGM for about 2.5M iterations with a batch size of 32, using two discriminators with a learning rate of 0.002 and a generator learning rate of 0.0025. The training takes 8 days using 8 NVIDIA Tesla-V100 GPUs. For GS, we optimize $\psi$ for 400 iterations for DMTET adaption and 100 iterations for surface carving (taking about 15s in total on 1 NVIDIA RTX 3090 GPU). For ET, we set $\lambda_{uv} = 1$ and optimize $T_{uv}$ for 500 iterations (around 10 seconds). We split $\mathcal{M}_{tri}$ into $5(\gamma = 5)$ body parts (i.e., trunk, left/right arm/leg) with cylinder UV unwarping. We use the Adam optimizer with learning rates of 0.01 and 0.001 for $\Psi_g$ and $T_{uv}$ , respectively. Detailed network architectures can be found in the supplemental materials (Suppl). + +# 4.1. 3D human generation + +Figure 4 showcases several 3D human avatars synthesized by our pipeline, highlighting the image quality, geometry accuracy, and diverse outputs achieved through our method. Additionally, we explore the interpolation of the latent conditions to yield smooth transitions in appearance, leveraging the smooth latent space learned by our generative model. For more synthesized examples and interpolation results, + +please refer to the Suppl. + +# 4.2. Comparisons + +Qualitative comparison. In Figure 5, we compare our method with three baselines: EVA3D [15] and AG3D [9], which are state-of-the-art methods for generating 3D humans from 2D images, and EG3D [5], which serves as the foundational backbone of our method. The results of first two methods are produced by directly using source codes and trained models released by authors. We train EG3D using our synthetic images with estimated cameras from scratch. As we can see, EVA3D fails to produce $360^{\circ}$ humans with reasonable back inferring. AG3D and EG3D are able to generate $360^{\circ}$ renderings but both struggle with photorealism and capturing detailed shapes. Our method synthesizes not only higher-quality, view-consistent $360^{\circ}$ images but also higher-fidelity 3D geometry with intricate details, such as irregular dresses and haircuts. + +Quantitative comparison. Table 1 provides quantitative results comparing our method against the baselines. We measure image quality with Frechet Inception Distance (FID) [14] and Inception Score [45] for $360^{\circ}$ views (IS- + +![](images/8eceb44aa9b924d4e0e2baa8eb42244150844cfa6809d3484dd1dae0082b5330.jpg) +Figure 5. Qualitative comparison with three state-of-the-art methods: EVA3D [15], AG3D [9] and EG3D [5]. + +Table 1. Quantitative evaluation using FID, IS-360, normal accuracy (Normal) and identity consistency (ID). + +
MethodFID ↓IS-360 ↑Normal ↓ID↑
EVA3D [15]15.913.1930.810.72
AG3D [9]10.933.2820.830.69
EVA3D-syn6.273.259.570.72
AG3D-syn8.933.228.640.71
EG3D [5]7.483.2612.740.71
Ours2.733.435.620.74
+ +360). FID measures the visual similarity and distribution discrepancy between 50k generated images and all real images. IS-360 focuses on the self-realism of generated images in $360^{\circ}$ views. For shape evaluation, we compute FID between rendered normals and pseudo-GT normal maps (Normal), following AG3D. The FID and Normal scores of EVA3D and AG3D are directly fetched from their reports. Additionally, we access the multi-view facial identity consistency using the ID metric introduced by EG3D. For a clearer comparison, we also provide metric results of EVA3D and AG3D trained on our synthetic data (-syn). Overall, our method demonstrates significant improvements in FID and Normal, bringing the generative human model to a new level of realistic $360^{\circ}$ renderings with delicate ge + +Table 2. Results of models trained by replacing physical parameters with estimated ones (w/o SYN-P) or removing patch-composed rendering (w/o PCR). + +
OursOurs-w/o SYN-POurs-w/o PCR
FID ↓2.734.283.26
IS-360 ↑3.433.313.35
+ +ometry while maintaining state-of-the-art view consistency. + +# 4.3. Ablation study + +Synthesis flow and patch-composed rendering. We assess the impact of our carefully designed synthesis flow by training a model with synthetic images but with camera and pose parameters estimated by SMPLify-X [40] (w/o SYN-P). As Table 2 shows, the model w/o SYN-P results in worse FID and IS-360 scores, indicating that the synthesis flow contributes to more accurate physical parameters for realistic appearance modeling. By utilizing patch-composed rendering (PCR), the networks focus more on the human region, leading to more realistic results. + +Geometry sculpting module (GS). We demonstrate the importance of this module by visualizing the meshes before and after its implementation. Figure 6 (b) shows that the preceding module yields a coarse mesh due to the complex + +![](images/8537eac5ff968ce7e4b86fb01af23539baaf0effb4de65a6e01d86a6a9a015d3.jpg) +(a) Rendering + +![](images/4cc72be7a7089040d22569f9920fbca696ab319f2661a67270d18c062954e5fd.jpg) +(b) w/o GS + +![](images/abf98b26aba27326a1c6e9a2e538f9aded98da542dda48a83a4a162b73f86b72.jpg) +(c) Full method + +![](images/93db33d01c4539a9f13305c65fc09867c17514a6f378dd8e40164170b106125f.jpg) + +![](images/7c4797b81ef35a91dfd5910f7be2a8c9c8e618e521c93f2b21df1d388196c4a3.jpg) +Figure 6. Effects of the GS module to carve fine-grained surfaces. +(a) Guided image + +![](images/1b31d5535ae48e32a8971579db57f0de9c7157d9f719c0ca81ca92d16a5532a4.jpg) +(b) w/o ET +Figure 7. Effects of the ET module for guided synthesis. + +![](images/73e7659150accc5a8ed134f271776f58a0194b43b6117cfbb3776dcc7eb462ca.jpg) +(c) Full method + +![](images/9afc8ce98f79de350a856978094f4cb2a80bdf96b59c8fa9cbfe672e5f779906.jpg) + +human anatomy and the challenges posed by decomposing ambiguous 3D shapes from 2D images. The GS module utilizes high-quality multi-view outputs and employs a more flexible hybrid representation to create expressive humans with arbitrary topologies. It learns from pixel-level surface supervision, leading to a significant improvement in shape quality, characterized by smooth surfaces and intricate outfits (Figure 6 (c)). + +Explicit texturing module (ET). This intuitive module not only extracts the explicit UV texture for complete 3D assets but also enables high-fidelity results for image guided synthesis. Following the flowchart in Figure 3, we compare the results produced with and without this module. Our method without ET directly generates implicit renderings through PTI inversion, as shown in Figure 7 (b). While it successfully preserves global identity, it struggles to synthesize highly faithful local textures (e.g., floral patterns). The ET module offers a convenient and efficient way to directly interact with the 3D representation, enabling the production of high-fidelity 3D humans with more consistent content including exquisite local patterns (Figure 7 (a, c)). + +# 4.4. Applications + +Avatar animation. All avatars produced by our method are in a canonical body pose and aligned to an underlying 3D skeleton extracted from SMPL-X. This alignment allows for easy animation and the generation of motion videos, as demonstrated in Figure 1 and Suppl. + +Texture doodle and local editing. Our approach benefits from the explicit disentanglement of geometry and texture, enabling flexible editing capabilities. Following the flowchart of text or image guided synthesis (Section 3.4), + +![](images/9a378752288b93e83b8a16f8a95e63c26c8f2fecb30cbf6c0cc9885649a09791.jpg) + +![](images/583a0aad33465cce86484fb0f9b15aa788adc4e18e73e6b916bac8029c63d208.jpg) + +![](images/70995fc2ada3f1fc26e4b2fe8b0e376047f51a38d3a01c66bb2473c40ac342e7.jpg) + +![](images/d1ca738bad190fd3d970be65da1a0523155ff37b795d55f60f21ad9599b13c99.jpg) +Figure 8. Results synthesized by adapting our method to various styles (e.g., Disney cartoon characters) or contents (e.g., portrait heads). + +![](images/057816ebef0725a7559f2491d385fa0718a06e7d46476a0e286f818230fde632.jpg) + +![](images/c9c097e0e8aca241825753fbdde53c38c562a6cf475ee50049f556cde5b46803.jpg) + +users can paint any pattern or add text to a guided image. These modifications can be transferred to 3D human models by inputting modified views into the texture module (e.g., painting the text 'hey' on a jacket as shown in Figure 1 (d)). Our approach also allows for clothing editing by simultaneously injecting edited guide images with desired clothing into the GS and ET modules (e.g., changing a jacket and jeans to bodysuits in Figure 1 (e)). More results can be found in Suppl. + +Content-style free adaption. Our proposed scheme is versatile and can be extended to generate various types of contents (e.g., portrait heads) and styles (e.g., Disney cartoon characters). To achieve this, we fine-tune our model using synthetic images from these domains, allowing for flexible adaptation. We showcase the results in Figure 8. More results and other discussions (e.g., limitations, negative impact, etc.) can be found in the Suppl. + +# 5. Conclusion + +We introduced En3D, a novel generative scheme for sculpting 3D humans from 2D synthetic data. This method overcomes limitations in existing 3D or 2D collections and significantly enhances the image quality, geometry accuracy, and content diversity of generated 3D humans. En3D comprises a 3D generative module that learns generalizable 3D humans from synthetic 2D data with accurate physical modeling, and two optimization modules to carve intricate shape details and disentangle explicit UV textures with high fidelity, respectively. Experimental results validated the superiority and effectiveness of our method. We also demonstrated the flexibility of our generated avatars for animation and editing, as well as the scalability of our approach for synthesizing portraits and Disney characters. We believe that our solution could provide invaluable human assets for the 3D vision community. Furthermore, it holds potential for use in common 3D object synthesis tasks. + +# References + +[1] Panos Achlioptas, Olga Diamanti, Ioannis Mitliagkas, and Leonidas Guibas. Learning representations and generative models for 3d point clouds. In International conference on machine learning, pages 40-49. PMLR, 2018. 3 +[2] Thiemo Alldieck, Marcus Magnor, Weipeng Xu, Christian Theobalt, and Gerard Pons-Moll. Video based reconstruction of 3d people models. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 8387-8397, 2018. 2 +[3] Dragomir Anguelov, Praveen Srinivasan, Daphne Koller, Sebastian Thrun, Jim Rodgers, and James Davis. Scape: shape completion and animation of people. In ACM SIGGRAPH 2005 Papers, pages 408-416. 2005. 2 +[4] Eric R Chan, Marco Monteiro, Petr Kellnhofer, Jiajun Wu, and Gordon Wetzstein. pi-gan: Periodic implicit generative adversarial networks for 3d-aware image synthesis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5799-5809, 2021. 3 +[5] Eric R Chan, Connor Z Lin, Matthew A Chan, Koki Nagano, Boxiao Pan, Shalini De Mello, Orazio Gallo, Leonidas J Guibas, Jonathan Tremblay, Sameh Khamis, et al. Efficient geometry-aware 3d generative adversarial networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16123-16133, 2022. 2, 3, 4, 6, 7 +[6] Xu Chen, Tianjian Jiang, Jie Song, Jinlong Yang, Michael J Black, Andreas Geiger, and Otmar Hilliges. gdna: Towards generative detailed neural avatars. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20427-20437, 2022. 2 +[7] Yu Deng, Jiaolong Yang, Jianfeng Xiang, and Xin Tong. Gram: Generative radiance manifolds for 3d-aware image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10673-10683, 2022. 3 +[8] Zijian Dong, Chen Guo, Jie Song, Xu Chen, Andreas Geiger, and Otmar Hilliges. Pina: Learning a personalized implicit neural avatar from a single rgb-d video sequence. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20470-20480, 2022. 1, 2 +[9] Zijian Dong, Xu Chen, Jinlong Yang, Michael J Black, Otmar Hilliges, and Andreas Geiger. Ag3d: Learning to generate 3d avatars from 2d image collections. arXiv preprint arXiv:2305.02312, 2023. 2, 3, 6, 7 +[10] Yao Feng, Jinlong Yang, Marc Pollefeys, Michael J Black, and Timo Bolkart. Capturing and animation of body and clothing from monocular video. In SIGGRAPH Asia 2022 Conference Papers, pages 1-9, 2022. 1, 2 +[11] Jianglin Fu, Shikai Li, Yuming Jiang, Kwan-Yee Lin, Chen Qian, Chen Change Loy, Wayne Wu, and Ziwei Liu. Stylegan-human: A data-centric odyssey of human generation. In European Conference on Computer Vision, pages 1-19. Springer, 2022. 3 +[12] Jiatao Gu, Lingjie Liu, Peng Wang, and Christian Theobalt. Stylenerf: A style-based 3d-aware genera + +tor for high-resolution image synthesis. arXiv preprint arXiv:2110.08985, 2021.3 +[13] Philipp Henzler, Niloy J Mitra, and Tobias Ritschel. Escaping plato's cave: 3d shape from adversarial rendering. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9984-9993, 2019. 3 +[14] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 6 +[15] Fangzhou Hong, Zhaoxi Chen, Yushi Lan, Liang Pan, and Ziwei Liu. Eva3d: Compositional 3d human generation from 2d image collections. arXiv preprint arXiv:2210.04888, 2022. 2, 3, 6, 7 +[16] Xin Huang, Ruizhi Shao, Qi Zhang, Hongwen Zhang, Ying Feng, Yebin Liu, and Qing Wang. Humannorm: Learning normal diffusion model for high-quality and realistic 3d human generation. arXiv preprint arXiv:2310.01406, 2023. 3 +[17] Suyi Jiang, Haoran Jiang, Ziyu Wang, Haimin Luo, Wenzheng Chen, and Lan Xu. Humangen: Generating human radiance fields with explicit priors. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12543-12554, 2023. 3 +[18] Kyungmin Jo, Wonjoon Jin, Jaegul Choo, Hyunjoon Lee, and Sunghyun Cho. 3d-aware generative model for improved side-view image synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 22862-22872, 2023. 2 +[19] Hanbyul Joo, Tomas Simon, and Yaser Sheikh. Total capture: A 3d deformation model for tracking faces, hands, and bodies. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 8320-8329, 2018. 2 +[20] Angjoo Kanazawa, Michael J Black, David W Jacobs, and Jitendra Malik. End-to-end recovery of human shape and pose. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7122-7131, 2018. 2 +[21] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4401-4410, 2019. 3 +[22] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of stylegan. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8110-8119, 2020. 3 +[23] Tero Karras, Miika Aittala, Samuli Laine, Erik Härkönen, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Alias-free generative adversarial networks. Advances in Neural Information Processing Systems, 34:852-863, 2021. 3 +[24] Nikos Kolotouros, Thiemo Alldieck, Andrei Zanfir, Eduard Bazavan, Mihai Fieraru, and Cristian Sminchisescu. Dreamhuman: Animatable 3d avatars from text. Advances in Neural Information Processing Systems, 36, 2024. 3 +[25] Samuli Laine, Janne Hellsten, Tero Karras, Yeongho Seol, Jaakko Lehtinen, and Timo Aila. Modular primitives for high-performance differentiable rendering. ACM Transactions on Graphics (TOG), 39(6):1-14, 2020. 5 + +[26] Ruihui Li, Xianzhi Li, Chi-Wing Fu, Daniel Cohen-Or, and Pheng-Ann Heng. Pu-gan: a point cloud upsampling adversarial network. In Proceedings of the IEEE/CVF international conference on computer vision, pages 7203–7212, 2019. 3 +[27] Yiyi Liao, Katja Schwarz, Lars Mescheder, and Andreas Geiger. Towards unsupervised learning of generative models for 3d controllable image synthesis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5871-5880, 2020. 3 +[28] Lingjie Liu, Marc Habermann, Viktor Rudnev, Kripasindhu Sarkar, Jiatao Gu, and Christian Theobalt. Neural actor: Neural free-view synthesis of human actors with pose control. ACM transactions on graphics (TOG), 40(6):1-16, 2021. 2 +[29] Matthew Loper, Naureen Mahmood, Javier Romero, Gerard Pons-Moll, and Michael J Black. Smpl: A skinned multiperson linear model. ACM Transactions on Graphics, 34(6), 2015. 2 +[30] William E Lorensen and Harvey E Cline. Marching cubes: A high resolution 3d surface construction algorithm. In Semin al graphics: pioneering efforts that shaped the field, pages 347-353. 1998. 4 +[31] Qianli Ma, Jinlong Yang, Anurag Ranjan, Sergi Pujades, Gerard Pons-Moll, Siyu Tang, and Michael J Black. Learning to dress 3d people in generative clothing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6469-6478, 2020. 2 +[32] Lars Mescheder, Andreas Geiger, and Sebastian Nowozin. Which training methods for gans do actually converge? In International conference on machine learning, pages 3481-3490. PMLR, 2018. 4 +[33] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 3 +[34] Thu Nguyen-Phuoc, Chuan Li, Lucas Theis, Christian Richardt, and Yong-Liang Yang. Hologan: Unsupervised learning of 3d representations from natural images. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7588-7597, 2019. 3 +[35] Thu H Nguyen-Phuoc, Christian Richardt, Long Mai, Yongliang Yang, and Niloy Mitra. Blockgan: Learning 3d object-aware scene representations from unlabelled images. Advances in neural information processing systems, 33:6767-6778, 2020. 3 +[36] Michael Niemeyer and Andreas Geiger. Giraffe: Representing scenes as compositional generative neural feature fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11453-11464, 2021. 3 +[37] Roy Or-El, Xuan Luo, Mengyi Shan, Eli Shechtman, Jeong Joon Park, and Ira Kemelmacher-Shlizerman. Stylesdf: High-resolution 3d-consistent image and geometry generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13503–13513, 2022. 2, 3 + +[38] Ahmed AA Osman, Timo Bolkart, and Michael J Black. Star: Sparse trained articulated human body regressor. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part VI 16, pages 598-613. Springer, 2020. 2 +[39] Pablo Palafox, Aljaž Božić, Justus Thies, Matthias Nießner, and Angela Dai. Npms: Neural parametric models for 3d deformable shapes. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12695-12705, 2021. 2 +[40] Georgios Pavlakos, Vasileios Choutas, Nima Ghorbani, Timo Bolkart, Ahmed AA Osman, Dimitrios Tzionas, and Michael J Black. Expressive body capture: 3d hands, face, and body from a single image. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10975-10985, 2019. 4, 7 +[41] Sida Peng, Yuanqing Zhang, Yinghao Xu, Qianqian Wang, Qing Shuai, Hujun Bao, and Xiaowei Zhou. Neural body: Implicit neural representations with structured latent codes for novel view synthesis of dynamic humans. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9054-9063, 2021. 1, 2 +[42] Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. arXiv preprint arXiv:2209.14988, 2022. 3 +[43] Daniel Roich, Ron Mokady, Amit H Bermano, and Daniel Cohen-Or. Pivotal tuning for latent-based editing of real images. ACM Transactions on graphics (TOG), 42(1):1-13, 2022. 5 +[44] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 4 +[45] Tim Salimans, Ian Goodfellow, Wojciech Zaremba, Vicki Cheung, Alec Radford, and Xi Chen. Improved techniques for training gans. Advances in neural information processing systems, 29, 2016. 6 +[46] Katja Schwarz, Yiyi Liao, Michael Niemeyer, and Andreas Geiger. Graf: Generative radiance fields for 3d-aware image synthesis. Advances in Neural Information Processing Systems, 33:20154-20166, 2020. 3 +[47] Attila Szabó, Givi Meishvili, and Paolo Favaro. Unsupervised generative 3d shape learning from natural images. arXiv preprint arXiv:1910.00287, 2019. 3 +[48] Chung-Yi Weng, Brian Curless, Pratul P Srinivasan, Jonathan T Barron, and Ira Kemelmacher-Shlizerman. Humanerf: Free-viewpoint rendering of moving people from monocular video. In Proceedings of the IEEE/CVF conference on computer vision and pattern Recognition, pages 16210-16220, 2022. 1, 2, 3 +[49] Jiajun Wu, Chengkai Zhang, Tianfan Xue, Bill Freeman, and Josh Tenenbaum. Learning a probabilistic latent space of object shapes via 3d generative-adversarial modeling. Advances in neural information processing systems, 29, 2016. 3 +[50] Yuliang Xiu, Jinlong Yang, Dimitrios Tzionas, and Michael J Black. Icon: Implicit clothed humans obtained from nor + +mals. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13286-13296. IEEE, 2022. 5 +[51] Yang Xue, Yuheng Li, Krishna Kumar Singh, and Yong Jae Lee. Giraffe hd: A high-resolution 3d-aware generative model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18440-18449, 2022. 3 +[52] Zhuoqian Yang, Shikai Li, Wayne Wu, and Bo Dai. 3dhumanan: 3d-aware human image generation with 3d pose mapping. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 23008-23019, 2023. 3 +[53] Jianfeng Zhang, Zihang Jiang, Dingdong Yang, Hongyi Xu, Yichun Shi, Guoxian Song, Zhongcong Xu, Xinchao Wang, and Jiashi Feng. Avatargen: a 3d generative model for animatable human avatars. In European Conference on Computer Vision, pages 668-685. Springer, 2022. 3 +[54] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3836-3847, 2023. 4 +[55] Peng Zhou, Lingxi Xie, Bingbing Ni, and Qi Tian. Cips-3d: A 3d-aware generator of gans based on conditionally-independent pixel synthesis. arXiv preprint arXiv:2110.09788, 2021. 3 \ No newline at end of file diff --git a/2024/En3D_ An Enhanced Generative Model for Sculpting 3D Humans from 2D Synthetic Data/images.zip b/2024/En3D_ An Enhanced Generative Model for Sculpting 3D Humans from 2D Synthetic Data/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..2fbfff8d881b096f38283f70aab8ead1c501231a --- /dev/null +++ b/2024/En3D_ An Enhanced Generative Model for Sculpting 3D Humans from 2D Synthetic Data/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5a0c64027beac31e83f5b5c8d5ecaa9bba10573cd548b467740f41447ae3504f +size 535088 diff --git a/2024/En3D_ An Enhanced Generative Model for Sculpting 3D Humans from 2D Synthetic Data/layout.json b/2024/En3D_ An Enhanced Generative Model for Sculpting 3D Humans from 2D Synthetic Data/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..2f36c0ecc475e8ce4db6b3047acb30a3cdd0d4e2 --- /dev/null +++ b/2024/En3D_ An Enhanced Generative Model for Sculpting 3D Humans from 2D Synthetic Data/layout.json @@ -0,0 +1,9219 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 78, + 102, + 516, + 138 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 78, + 102, + 516, + 138 + ], + "spans": [ + { + "bbox": [ + 78, + 102, + 516, + 138 + ], + "type": "text", + "content": "En3D: An Enhanced Generative Model for Sculpting 3D Humans from 2D Synthetic Data" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 87, + 160, + 505, + 176 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 87, + 160, + 505, + 176 + ], + "spans": [ + { + "bbox": [ + 87, + 160, + 505, + 176 + ], + "type": "text", + "content": "Yifang Men1, Biwen Lei1, Yuan Yao1, Miaomiao Cui1, Zhouhui Lian2, Xuansong Xie1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 158, + 177, + 434, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 158, + 177, + 434, + 190 + ], + "spans": [ + { + "bbox": [ + 158, + 177, + 434, + 190 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 158, + 177, + 434, + 190 + ], + "type": "text", + "content": "Institute for Intelligent Computing, Alibaba Group" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 113, + 191, + 481, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 113, + 191, + 481, + 204 + ], + "spans": [ + { + "bbox": [ + 113, + 191, + 481, + 204 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 113, + 191, + 481, + 204 + ], + "type": "text", + "content": "Wangxuan Institute of Computer Technology, Peking University, China" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 155, + 205, + 436, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 155, + 205, + 436, + 217 + ], + "spans": [ + { + "bbox": [ + 155, + 205, + 436, + 217 + ], + "type": "text", + "content": "https://menyifang.github.io/projects/En3D/index.html" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 53, + 247, + 329, + 343 + ], + "blocks": [ + { + "bbox": [ + 53, + 238, + 69, + 251 + ], + "lines": [ + { + "bbox": [ + 53, + 238, + 69, + 251 + ], + "spans": [ + { + "bbox": [ + 53, + 238, + 69, + 251 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 53, + 247, + 329, + 343 + ], + "lines": [ + { + "bbox": [ + 53, + 247, + 329, + 343 + ], + "spans": [ + { + "bbox": [ + 53, + 247, + 329, + 343 + ], + "type": "image", + "image_path": "983670dcb4b85364dad9a3570384c18c010d456f295654df6337a5da0fd591ba.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 332, + 234, + 347, + 247 + ], + "lines": [ + { + "bbox": [ + 332, + 234, + 347, + 247 + ], + "spans": [ + { + "bbox": [ + 332, + 234, + 347, + 247 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 358, + 234, + 517, + 252 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 358, + 234, + 517, + 252 + ], + "spans": [ + { + "bbox": [ + 358, + 234, + 517, + 252 + ], + "type": "text", + "content": "\"a 44-year-old Latino man with blonde side part haircut, wearing a bomber jacket with brooch\"" + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 338, + 255, + 535, + 342 + ], + "blocks": [ + { + "bbox": [ + 338, + 255, + 535, + 342 + ], + "lines": [ + { + "bbox": [ + 338, + 255, + 535, + 342 + ], + "spans": [ + { + "bbox": [ + 338, + 255, + 535, + 342 + ], + "type": "image", + "image_path": "76e7ccab3e14de71b3795b0179e19c701cf0088e680e8c398646ea720bf08bde.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 61, + 345, + 299, + 434 + ], + "blocks": [ + { + "bbox": [ + 55, + 346, + 69, + 358 + ], + "lines": [ + { + "bbox": [ + 55, + 346, + 69, + 358 + ], + "spans": [ + { + "bbox": [ + 55, + 346, + 69, + 358 + ], + "type": "text", + "content": "(c)" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 61, + 345, + 299, + 434 + ], + "lines": [ + { + "bbox": [ + 61, + 345, + 299, + 434 + ], + "spans": [ + { + "bbox": [ + 61, + 345, + 299, + 434 + ], + "type": "image", + "image_path": "402cb4a2fea8a39cad637b91faa73ad3935b124bcac012ab790e2cdccdbc4e8b.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 45, + 436, + 547, + 470 + ], + "lines": [ + { + "bbox": [ + 45, + 436, + 547, + 470 + ], + "spans": [ + { + "bbox": [ + 45, + 436, + 547, + 470 + ], + "type": "text", + "content": "Figure 1. Given random noises or guided texts, our generative scheme can synthesize high-fidelity 3D human avatars that are visually realistic and geometrically accurate. These avatars can be seamlessly animated and easily edited. Our model is trained on 2D synthetic data without directly relying on any pre-existing 3D or 2D collections." + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 319, + 346, + 434, + 434 + ], + "blocks": [ + { + "bbox": [ + 312, + 346, + 327, + 358 + ], + "lines": [ + { + "bbox": [ + 312, + 346, + 327, + 358 + ], + "spans": [ + { + "bbox": [ + 312, + 346, + 327, + 358 + ], + "type": "text", + "content": "(d)" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 319, + 346, + 434, + 434 + ], + "lines": [ + { + "bbox": [ + 319, + 346, + 434, + 434 + ], + "spans": [ + { + "bbox": [ + 319, + 346, + 434, + 434 + ], + "type": "image", + "image_path": "b39a44ea2412c7c1054948a143823a1c704380e2114d9e8e16d8734ec5d15d7d.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 433, + 346, + 447, + 358 + ], + "lines": [ + { + "bbox": [ + 433, + 346, + 447, + 358 + ], + "spans": [ + { + "bbox": [ + 433, + 346, + 447, + 358 + ], + "type": "text", + "content": "(e)" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 436, + 348, + 544, + 434 + ], + "blocks": [ + { + "bbox": [ + 436, + 348, + 544, + 434 + ], + "lines": [ + { + "bbox": [ + 436, + 348, + 544, + 434 + ], + "spans": [ + { + "bbox": [ + 436, + 348, + 544, + 434 + ], + "type": "image", + "image_path": "1f26002baed92f7862f31bd212cc40f3e1470aa0d9f7064e46d2a173cd05344f.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "bbox": [ + 143, + 479, + 192, + 493 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 479, + 192, + 493 + ], + "spans": [ + { + "bbox": [ + 143, + 479, + 192, + 493 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 46, + 510, + 290, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 510, + 290, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 510, + 290, + 715 + ], + "type": "text", + "content": "We present En3D, an enhanced generative scheme for sculpting high-quality 3D human avatars. Unlike previous works that rely on scarce 3D datasets or limited 2D collections with imbalanced viewing angles and imprecise pose priors, our approach aims to develop a zero-shot 3D generative scheme capable of producing visually realistic, geometrically accurate and content-wise diverse 3D humans without directly relying on pre-existing 3D or 2D assets. To address this challenge, we introduce a meticulously crafted workflow that implements accurate physical modeling to learn the enhanced 3D generative model from synthetic 2D data. During inference, we integrate optimization modules to bridge the gap between realistic appearances and coarse 3D shapes. Specifically, En3D comprises three modules: a 3D generator that accurately models generalizable 3D humans with realistic appearance from synthesized balanced, diverse, and structured human images; a geometry sculptor" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 481, + 548, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 481, + 548, + 601 + ], + "spans": [ + { + "bbox": [ + 304, + 481, + 548, + 601 + ], + "type": "text", + "content": "that enhances shape quality using multi-view normal constraints for intricate human structure; and a texturing module that disentangles explicit texture maps with fidelity and editability, leveraging semantical UV partitioning and a differentiable rasterizer. Experimental results show that our approach significantly outperforms prior works in terms of image quality, geometry accuracy and content diversity. We also showcase the applicability of our generated avatars for animation and editing, as well as the scalability of our approach for content-style free adaptation." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 306, + 621, + 387, + 634 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 621, + 387, + 634 + ], + "spans": [ + { + "bbox": [ + 306, + 621, + 387, + 634 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 304, + 641, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 641, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 641, + 547, + 715 + ], + "type": "text", + "content": "3D human avatars play an important role in various applications of AR/VR such as video games, telepresence and virtual try-on. Realistic human modeling is an essential task, and many valuable efforts have been made by leveraging neural implicit fields to learn high-quality articulated avatars [8, 10, 41, 48]. However, these methods are directly" + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "9981" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 108 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 108 + ], + "type": "text", + "content": "learned from monocular videos or image sequences, where subjects are single individuals wearing specific garments, thus limiting their scalability." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 110, + 288, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 110, + 288, + 350 + ], + "spans": [ + { + "bbox": [ + 46, + 110, + 288, + 350 + ], + "type": "text", + "content": "Generative models learn a shared 3D representation to synthesize clothed humans with varying identities, clothing and poses. Traditional methods are typically trained on 3D datasets, which are limited and expensive to acquire. This data scarcity limits the model's generalization ability and may lead to overfitting on small datasets. Recently, 3D-aware image synthesis methods [5, 18, 37] have demonstrated great potential in learning 3D generative models of rigid objects from 2D image collections. Follow-up works show the feasibility of learning articulated humans from image collections driven by SMPL-based deformations, but only in limited quality and resolution. EVA3D [15] represents humans as a composition of multiple parts with NeRF representations. AG3D [9] incorporates an efficient articulation module to capture both body shape and cloth deformation. Nevertheless, there remains a noticeable gap between generated and real humans in terms of appearance and geometry. Moreover, their results are limited to specific views (i.e., frontal angles) and lack diversity (i.e., fashion images in similar skin tone, body shape, and age)." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 352, + 287, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 352, + 287, + 639 + ], + "spans": [ + { + "bbox": [ + 46, + 352, + 287, + 639 + ], + "type": "text", + "content": "The aim of this paper is to propose a zero-shot 3D generative scheme that does not directly rely on any pre-existing 3D or 2D datasets, yet is capable of producing high-quality 3D humans that are visually realistic, geometrically accurate, and content-wise diverse. The generated avatars can be seamlessly animated and easily edited. An illustration is provided in Figure 1. To address this challenging task, our proposed method inherits from 3D-aware human image synthesis and exhibits substantial distinctions based on several key insights. Rethinking the nature of 3D-aware generative methods from 2D collections [5, 9, 15], they actually try to learn a generalizable and deformable 3D representation, whose 2D projections can meet the distribution of human images in corresponding views. Thereby, it is crucial for accurate physical modeling between 3D objects and 2D projections. However, previous works typically leverage pre-existing 2D human images to estimate physical parameters (i.e., camera and body poses), which are inaccurate because of imprecise SMPL priors for highly-articulated humans. This inaccuracy limits the synthesis ability for realistic multi-view renderings. Second, these methods solely rely on discriminating 2D renderings, which is ambiguous and loose to capture inherent 3D shapes in detail, especially for intricate human appearance." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 642, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 642, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 642, + 287, + 715 + ], + "type": "text", + "content": "To address these limitations, we propose a novel generative scheme with two core designs. Firstly, we introduce a meticulously-crafted workflow that implements accurate physical modeling to learn an enhanced 3D generative model from synthetic data. This is achieved by instantiating a 3D body scene and projecting the underlying" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 304, + 72, + 545, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 312 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 312 + ], + "type": "text", + "content": "3D skeletons into 2D pose images using explicit camera parameters. These 2D pose images act as conditions to control a 2D diffusion model, synthesizing realistic human images from specific viewpoints. By leveraging synthetic view-balanced, diverse and structured human images, along with known physical parameters, we employ a 3D generator equipped with an enhanced renderer and discriminator to learn realistic appearance modeling. Secondly, we improve the 3D shape quality by leveraging the gap between high-quality multi-view renderings and the coarse mesh produced by the 3D generative module. Specifically, we integrate an optimization module that utilizes multi-view normal constraints to rapidly refine geometry details under supervision. Additionally, we incorporate an explicit texturing module to ensure faithful UV texture maps. In contrast to previous works that rely on inaccurate physical settings and inadequate shape supervision, we rebuild the generative scheme from the ground up, resulting in comprehensive improvements in image quality, geometry accuracy, and content diversity. In summary, our contributions are threefold:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 306, + 312, + 545, + 478 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 306, + 312, + 545, + 371 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 312, + 545, + 371 + ], + "spans": [ + { + "bbox": [ + 306, + 312, + 545, + 371 + ], + "type": "text", + "content": "- We present a zero-shot generative scheme that efficiently synthesizes high-quality 3D human avatars with visual realism, geometric accuracy and content diversity. These avatars can be seamlessly animated and easily edited, offering greater flexibility in their applications." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 372, + 545, + 431 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 372, + 545, + 431 + ], + "spans": [ + { + "bbox": [ + 306, + 372, + 545, + 431 + ], + "type": "text", + "content": "- We develop a meticulously-crafted workflow to learn an enhanced generative model from synthesized human images that are balanced, diverse, and also possess known physical parameters. This leads to diverse 3D-aware human image synthesis with realistic appearance." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 431, + 545, + 478 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 431, + 545, + 478 + ], + "spans": [ + { + "bbox": [ + 306, + 431, + 545, + 478 + ], + "type": "text", + "content": "- We propose to integrate optimization modules into the 3D generator, leveraging multi-view guidance to enhance both shape quality and texture fidelity, thus achieving realistic 3D human assets." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 490, + 389, + 502 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 490, + 389, + 502 + ], + "spans": [ + { + "bbox": [ + 306, + 490, + 389, + 502 + ], + "type": "text", + "content": "2. Related work" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "text", + "content": "3D Human Modeling. Parametric models [3, 19, 20, 29, 38] serve as a common representation for 3D human modeling, they allows for robust control by deforming a template mesh with a series of low-dimensional parameters, but can only generate naked 3D humans. Similar ideas have been extended to model clothed humans [2, 31], but geometric expressivity is restricted due to the fixed mesh topology. Subsequent works [6, 39, 39] further introduce implicit surfaces to produce complex non-linear deformations of 3D bodies. Unfortunately, the aforementioned approaches all require 3D scans of various human poses for model fitting, which are difficult to acquire. With the explosion of NeRF, valuable efforts have been made towards combining NeRF models with explicit human models [8, 10, 28, 41, 48]. Neural body [41] anchors a set of latent codes to the vertices of the SMPL model [29] and transforms the spatial locations of the codes to the volume in the observation space. Human-" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "9982" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 51, + 75, + 545, + 240 + ], + "blocks": [ + { + "bbox": [ + 51, + 75, + 545, + 240 + ], + "lines": [ + { + "bbox": [ + 51, + 75, + 545, + 240 + ], + "spans": [ + { + "bbox": [ + 51, + 75, + 545, + 240 + ], + "type": "image", + "image_path": "7d51faa41e042c861d61716c00ef486d05b132f800883da5657afb6ca6e7ccd4.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 242, + 546, + 298 + ], + "lines": [ + { + "bbox": [ + 46, + 242, + 546, + 298 + ], + "spans": [ + { + "bbox": [ + 46, + 242, + 546, + 298 + ], + "type": "text", + "content": "Figure 2. An overview of the proposed scheme, which consists of three modules: 3D generative modeling (3DGM), geometric sculpting (GS), and explicit texturing (ET). 3DGM employs synthesized diverse, balanced and structured human images with accurate camera parameters " + }, + { + "bbox": [ + 46, + 242, + 546, + 298 + ], + "type": "inline_equation", + "content": "\\varphi" + }, + { + "bbox": [ + 46, + 242, + 546, + 298 + ], + "type": "text", + "content": " to learn generalizable 3D humans with the triplane-based architecture. GS is integrated as an optimization module by utilizing multi-view normal constraints to refine and carve geometry details. ET utilizes UV partitioning and a differentiable rasterizer to disentangle explicit UV texture maps. Not only multi-view renderings but also realistic 3D models can be acquired finally by our method." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 319, + 288, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 319, + 288, + 403 + ], + "spans": [ + { + "bbox": [ + 46, + 319, + 288, + 403 + ], + "type": "text", + "content": "NeRF [48] optimizes for a canonical, volumetric T-pose of the human with a motion field to map the non-rigid transformations. Text to 3D humans [16, 24] are introduced by combining SDS loss [42] with diffusion guidance. Nevertheless, these methods are learned directly from image sequences or guided by single text prompts, where subjects are fitting to single scenes, thus limiting their scalability." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 409, + 287, + 661 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 409, + 287, + 661 + ], + "spans": [ + { + "bbox": [ + 46, + 409, + 287, + 661 + ], + "type": "text", + "content": "Generative 3D-aware Image Synthesis. Recently, 3D-aware image synthesis methods have lifted image generation with explicit view control by integrating the 2D generative models [21-23] with 3D representations, such as voxels [13, 34, 35, 49], meshes [27, 47] and points clouds [1, 26]. GRAF [46] and " + }, + { + "bbox": [ + 46, + 409, + 287, + 661 + ], + "type": "inline_equation", + "content": "\\pi" + }, + { + "bbox": [ + 46, + 409, + 287, + 661 + ], + "type": "text", + "content": "-GAN[4] firstly integrate the implicit representation networks, i.e., NeRF [33], with differentiable volumetric rendering for 3D scene generation. However, they have difficulties in training on high-resolution images due to the costly rendering process. Subsequent works have sought to improve the efficiency and quality of such NeRF-based GANs, either by adopting a two-stage rendering process [5, 12, 36, 37, 51] or a smart sampling strategy [7, 55]. StyleSDF [37] combines a SDF-based volume renderer and a 2D StyleGAN network [22] for photorealistic image generation. EG3D [5] introduces a superior triplane representation to leverage 2D CNN-based feature generators for efficient generalization over 3D spaces. Although these methods demonstrate impressive quality in view-consistent image synthesis, they are limited to simplified rigid objects such as faces, cats and cars." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 665, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 287, + 715 + ], + "type": "text", + "content": "To learn highly articulated humans from unstructured 2D images, recent works [9, 11, 15, 17, 52, 53] integrate the deformation field to learn non-rigid deformations based on the body prior of estimated SMPL parameters. EVA3D [15]" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 319, + 546, + 476 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 319, + 546, + 476 + ], + "spans": [ + { + "bbox": [ + 304, + 319, + 546, + 476 + ], + "type": "text", + "content": "represents humans as a composition of multiple parts with NeRF representations. Instead of directly rendering the image from a 3D representation, 3DHumanGAN [52] uses an equivariant 2D generator modulated by 3D human body prior, which enables to establish one-to-many mapping from 3D geometry to synthesized textures from 2D images. AG3D [9] combines the 3D generator with an efficient articulation module to warp 3D objects from canonical space to posed space via a learned continuous deformation field. However, a gap still exists between the generated and real humans in terms of appearance, due to the imprecise priors from complex poses as well as the data biases from limited human poses and imbalanced viewing angles in the dataset." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 488, + 424, + 502 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 488, + 424, + 502 + ], + "spans": [ + { + "bbox": [ + 306, + 488, + 424, + 502 + ], + "type": "text", + "content": "3. Method Description" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 509, + 545, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 509, + 545, + 569 + ], + "spans": [ + { + "bbox": [ + 304, + 509, + 545, + 569 + ], + "type": "text", + "content": "Our goal is to develop a zero-shot 3D generative scheme that does not directly rely on any pre-existing 3D or 2D collections, yet is capable of producing high-quality 3D humans that are visually realistic, geometrically accurate and content-wise diverse to generalize to arbitrary humans." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 570, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 546, + 713 + ], + "type": "text", + "content": "An overview of the proposed scheme is illustrated in Figure 2. We build a sequential pipeline with the following three modules: 3D generative modeling (3DGM), geometric sculpting (GS), and explicit texturing (ET). The first module synthesizes view-balanced, structured and diverse human images with known camera parameters. Subsequently, it learns a 3D generative model from these synthetic data, focusing on realistic appearance modeling (Section 3.1). To overcome the inaccuracy of the 3D shape, the GS module is incorporated during the inference process. It optimizes a hybrid representation with multi-view normal constraints to carve intricate mesh details (Section 3.2)." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "9983" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 145 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 145 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 145 + ], + "type": "text", + "content": "Additionally, the ET module is employed to disentangle explicit texture by utilizing semantical UV partitioning and a differentiable rasterizer (Section 3.3). By combining these modules, we are able to synthesize high-quality and faithful 3D human avatars by incorporating random noises or guided texts/images (Section 3.4)." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 154, + 181, + 166 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 154, + 181, + 166 + ], + "spans": [ + { + "bbox": [ + 47, + 154, + 181, + 166 + ], + "type": "text", + "content": "3.1. 3D generative modeling" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 172, + 287, + 304 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 172, + 287, + 304 + ], + "spans": [ + { + "bbox": [ + 46, + 172, + 287, + 304 + ], + "type": "text", + "content": "Without any 3D or 2D collections, we develop a synthesis-based flow to learn a 3D generative module from 2D synthetic data. We start by instantiating a 3D scene through the projection of underlying 3D skeletons onto 2D pose images, utilizing accurate physical parameters (i.e., camera parameters). Subsequently, the projected 2D pose images serve as conditions to control the 2D diffusion model [54] for synthesizing view-balanced, diverse, and lifelike human images. Finally, we employ a triplane-based generator with enhanced designs to learn a generalizable 3D representation from the synthetic data. Details are described as follows." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 305, + 287, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 305, + 287, + 389 + ], + "spans": [ + { + "bbox": [ + 46, + 305, + 287, + 389 + ], + "type": "text", + "content": "3D instantiation. Starting with a template body mesh (e.g., SMPL-X [40]) positioned and posed in canonical space, we estimate the 3D joint locations " + }, + { + "bbox": [ + 46, + 305, + 287, + 389 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_{3d}" + }, + { + "bbox": [ + 46, + 305, + 287, + 389 + ], + "type": "text", + "content": " by regressing them from interpolated vertices. We then project " + }, + { + "bbox": [ + 46, + 305, + 287, + 389 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_{3d}" + }, + { + "bbox": [ + 46, + 305, + 287, + 389 + ], + "type": "text", + "content": " onto 2D poses " + }, + { + "bbox": [ + 46, + 305, + 287, + 389 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_i, i = 1,\\dots,k" + }, + { + "bbox": [ + 46, + 305, + 287, + 389 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 46, + 305, + 287, + 389 + ], + "type": "inline_equation", + "content": "\\kappa" + }, + { + "bbox": [ + 46, + 305, + 287, + 389 + ], + "type": "text", + "content": " horizontally uniformly sampled viewpoints " + }, + { + "bbox": [ + 46, + 305, + 287, + 389 + ], + "type": "inline_equation", + "content": "\\varphi" + }, + { + "bbox": [ + 46, + 305, + 287, + 389 + ], + "type": "text", + "content": ". In this way, paired 2D pose images and their corresponding camera parameters " + }, + { + "bbox": [ + 46, + 305, + 287, + 389 + ], + "type": "inline_equation", + "content": "\\{\\mathcal{P}_i,\\varphi_i\\}" + }, + { + "bbox": [ + 46, + 305, + 287, + 389 + ], + "type": "text", + "content": " are formulated." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 389, + 287, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 389, + 287, + 568 + ], + "spans": [ + { + "bbox": [ + 46, + 389, + 287, + 568 + ], + "type": "text", + "content": "Controlled 2D image synthesis. With the pose image " + }, + { + "bbox": [ + 46, + 389, + 287, + 568 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_i" + }, + { + "bbox": [ + 46, + 389, + 287, + 568 + ], + "type": "text", + "content": ", we feed it into off-the-shelf ControlNet [54] as the pose condition to guide diffusion models [44] to synthesize human images in desired poses (i.e., views). The text prompt " + }, + { + "bbox": [ + 46, + 389, + 287, + 568 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 46, + 389, + 287, + 568 + ], + "type": "text", + "content": " is also used for diverse contents. Given a prompt " + }, + { + "bbox": [ + 46, + 389, + 287, + 568 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 46, + 389, + 287, + 568 + ], + "type": "text", + "content": ", instead of generating a human image " + }, + { + "bbox": [ + 46, + 389, + 287, + 568 + ], + "type": "inline_equation", + "content": "\\mathcal{I}_s: \\mathcal{I}_s = \\mathcal{C}(\\mathcal{P}_i, T)" + }, + { + "bbox": [ + 46, + 389, + 287, + 568 + ], + "type": "text", + "content": " independently for each view " + }, + { + "bbox": [ + 46, + 389, + 287, + 568 + ], + "type": "inline_equation", + "content": "\\varphi_i" + }, + { + "bbox": [ + 46, + 389, + 287, + 568 + ], + "type": "text", + "content": ", we horizontally concatenate " + }, + { + "bbox": [ + 46, + 389, + 287, + 568 + ], + "type": "inline_equation", + "content": "\\mathcal{K}" + }, + { + "bbox": [ + 46, + 389, + 287, + 568 + ], + "type": "text", + "content": " pose images " + }, + { + "bbox": [ + 46, + 389, + 287, + 568 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_i \\in R^{H \\times W \\times 3}" + }, + { + "bbox": [ + 46, + 389, + 287, + 568 + ], + "type": "text", + "content": ", resulting in " + }, + { + "bbox": [ + 46, + 389, + 287, + 568 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_i' \\in R^{H \\times KW \\times 3}" + }, + { + "bbox": [ + 46, + 389, + 287, + 568 + ], + "type": "text", + "content": " and feed " + }, + { + "bbox": [ + 46, + 389, + 287, + 568 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_i'" + }, + { + "bbox": [ + 46, + 389, + 287, + 568 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 46, + 389, + 287, + 568 + ], + "type": "inline_equation", + "content": "\\mathcal{C}" + }, + { + "bbox": [ + 46, + 389, + 287, + 568 + ], + "type": "text", + "content": ", along with a prompt hint of 'multi-view' in " + }, + { + "bbox": [ + 46, + 389, + 287, + 568 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 46, + 389, + 287, + 568 + ], + "type": "text", + "content": ". In this way, multi-view human images " + }, + { + "bbox": [ + 46, + 389, + 287, + 568 + ], + "type": "inline_equation", + "content": "\\mathcal{I}_s'" + }, + { + "bbox": [ + 46, + 389, + 287, + 568 + ], + "type": "text", + "content": " are synthesized with roughly coherent appearance. We split " + }, + { + "bbox": [ + 46, + 389, + 287, + 568 + ], + "type": "inline_equation", + "content": "\\mathcal{I}_s'" + }, + { + "bbox": [ + 46, + 389, + 287, + 568 + ], + "type": "text", + "content": " to single view images " + }, + { + "bbox": [ + 46, + 389, + 287, + 568 + ], + "type": "inline_equation", + "content": "\\mathcal{I}_{\\varphi}" + }, + { + "bbox": [ + 46, + 389, + 287, + 568 + ], + "type": "text", + "content": " under specific views " + }, + { + "bbox": [ + 46, + 389, + 287, + 568 + ], + "type": "inline_equation", + "content": "\\varphi" + }, + { + "bbox": [ + 46, + 389, + 287, + 568 + ], + "type": "text", + "content": ". This concatenation strategy facilitates the convergence of distributions in synthetic multi-views, thus easing the learning of common 3D representation meeting multi-view characteristics." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "type": "text", + "content": "Generalizable 3D representation learning. With synthetic data of paired " + }, + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\{\\mathcal{I}_{\\varphi}, \\varphi\\}" + }, + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "type": "text", + "content": ", we learn the 3D generative module " + }, + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_{3d}" + }, + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "type": "text", + "content": " from them to produce diverse 3D-aware human images with realistic appearance. Inspired by EG3D [5], we employ a triplane-based generator to produce a generalizable representation " + }, + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{T}" + }, + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "type": "text", + "content": " and introduce a patch-composed neural renderer to learn intricate human representation efficiently. Specifically, instead of uniformly sampling 2D pixels on the image " + }, + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{I}" + }, + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "type": "text", + "content": ", we decompose patches in the ROI region including human bodies, and only emit rays towards pixels in these patches. The rays are rendered into RGB color with opacity values via volume rendering. Based on" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "text", + "content": "the decomposed rule, we decode rendered colors to multiple patches and re-combine these patches for full feature images. In this way, the representation is composed of effective human body parts, which directs the attention of the networks towards the human subject itself. This design facilitates fine-grained local human learning while maintaining computational efficiency." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 156, + 545, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 156, + 545, + 216 + ], + "spans": [ + { + "bbox": [ + 304, + 156, + 545, + 216 + ], + "type": "text", + "content": "For the training process, we employ two discriminators, one for RGB images and another for silhouettes, which yields better disentanglement of foreground objects with global geometry. The training loss for this module " + }, + { + "bbox": [ + 304, + 156, + 545, + 216 + ], + "type": "inline_equation", + "content": "L_{3d}" + }, + { + "bbox": [ + 304, + 156, + 545, + 216 + ], + "type": "text", + "content": " consists of the two adversarial terms:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 320, + 225, + 545, + 239 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 225, + 545, + 239 + ], + "spans": [ + { + "bbox": [ + 320, + 225, + 545, + 239 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {3 d} = \\mathcal {L} _ {a d v} \\left(\\mathcal {D} _ {r g b}, \\mathcal {G} _ {3 d}\\right) + \\lambda_ {s} \\mathcal {L} _ {a d v} \\left(\\mathcal {D} _ {\\text {m a s k}}, \\mathcal {G} _ {3 d}\\right), \\tag {1}", + "image_path": "4623756c1cf39b8e4872a9953fe21cbc81a0d1f8655b0bd25fdfc00e87267327.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 247, + 545, + 283 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 247, + 545, + 283 + ], + "spans": [ + { + "bbox": [ + 304, + 247, + 545, + 283 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 247, + 545, + 283 + ], + "type": "inline_equation", + "content": "\\lambda_{s}" + }, + { + "bbox": [ + 304, + 247, + 545, + 283 + ], + "type": "text", + "content": " denotes the weight of silhouette item. " + }, + { + "bbox": [ + 304, + 247, + 545, + 283 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{adv}" + }, + { + "bbox": [ + 304, + 247, + 545, + 283 + ], + "type": "text", + "content": " is computed by the non-saturating GAN loss with R1 regularization [32]." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 284, + 545, + 332 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 284, + 545, + 332 + ], + "spans": [ + { + "bbox": [ + 304, + 284, + 545, + 332 + ], + "type": "text", + "content": "With the trained " + }, + { + "bbox": [ + 304, + 284, + 545, + 332 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_{3d}" + }, + { + "bbox": [ + 304, + 284, + 545, + 332 + ], + "type": "text", + "content": ", we can synthesize 3D-aware human images " + }, + { + "bbox": [ + 304, + 284, + 545, + 332 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_g^\\varphi" + }, + { + "bbox": [ + 304, + 284, + 545, + 332 + ], + "type": "text", + "content": " with view control, and extract coarse 3D shapes " + }, + { + "bbox": [ + 304, + 284, + 545, + 332 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_c" + }, + { + "bbox": [ + 304, + 284, + 545, + 332 + ], + "type": "text", + "content": " from the density field of neural renderer using the Marching Cubes algorithm [30]." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 339, + 424, + 353 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 339, + 424, + 353 + ], + "spans": [ + { + "bbox": [ + 306, + 339, + 424, + 353 + ], + "type": "text", + "content": "3.2. Geometric sculpting" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 357, + 545, + 488 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 357, + 545, + 488 + ], + "spans": [ + { + "bbox": [ + 304, + 357, + 545, + 488 + ], + "type": "text", + "content": "Our 3D generative module can produce high-quality and 3D-consistent human images in view controls. However, its training solely relies on discriminations made using 2D renderings, which can result in inaccuracies in capturing the inherent geometry, especially for complex human bodies. Therefore, we integrate the geometric sculpting, an optimization module leveraging geometric information from high-quality multi-views to carve surface details. Combined with a hybrid 3D representation and a differentiable rasterizer, it can rapidly enhance the shape quality within seconds." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 489, + 545, + 622 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 489, + 545, + 622 + ], + "spans": [ + { + "bbox": [ + 304, + 489, + 545, + 622 + ], + "type": "text", + "content": "DMTET adaption. Owing to the expressive ability of arbitrary topologies and computational efficiency with direct shape optimization, we employ DMTET as our 3D representation in this module and adapt it to the coarse mesh " + }, + { + "bbox": [ + 304, + 489, + 545, + 622 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_c" + }, + { + "bbox": [ + 304, + 489, + 545, + 622 + ], + "type": "text", + "content": " via an initial fitting procedure. Specifically, we parameterize DMTET as an MLP network " + }, + { + "bbox": [ + 304, + 489, + 545, + 622 + ], + "type": "inline_equation", + "content": "\\Psi_g" + }, + { + "bbox": [ + 304, + 489, + 545, + 622 + ], + "type": "text", + "content": " that learns to predict the SDF value " + }, + { + "bbox": [ + 304, + 489, + 545, + 622 + ], + "type": "inline_equation", + "content": "s(v_i)" + }, + { + "bbox": [ + 304, + 489, + 545, + 622 + ], + "type": "text", + "content": " and the position offset " + }, + { + "bbox": [ + 304, + 489, + 545, + 622 + ], + "type": "inline_equation", + "content": "\\delta v_i" + }, + { + "bbox": [ + 304, + 489, + 545, + 622 + ], + "type": "text", + "content": " for each vertex " + }, + { + "bbox": [ + 304, + 489, + 545, + 622 + ], + "type": "inline_equation", + "content": "v_i \\in VT" + }, + { + "bbox": [ + 304, + 489, + 545, + 622 + ], + "type": "text", + "content": " of the tetrahedral grid " + }, + { + "bbox": [ + 304, + 489, + 545, + 622 + ], + "type": "inline_equation", + "content": "(VT,T)" + }, + { + "bbox": [ + 304, + 489, + 545, + 622 + ], + "type": "text", + "content": ". A point set " + }, + { + "bbox": [ + 304, + 489, + 545, + 622 + ], + "type": "inline_equation", + "content": "P = \\{p_i \\in R^3\\}" + }, + { + "bbox": [ + 304, + 489, + 545, + 622 + ], + "type": "text", + "content": " is randomly sampled near " + }, + { + "bbox": [ + 304, + 489, + 545, + 622 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_c" + }, + { + "bbox": [ + 304, + 489, + 545, + 622 + ], + "type": "text", + "content": " and their SDF values " + }, + { + "bbox": [ + 304, + 489, + 545, + 622 + ], + "type": "inline_equation", + "content": "SDF(p_i)" + }, + { + "bbox": [ + 304, + 489, + 545, + 622 + ], + "type": "text", + "content": " can be pre-computed. We adapt the parameters " + }, + { + "bbox": [ + 304, + 489, + 545, + 622 + ], + "type": "inline_equation", + "content": "\\psi" + }, + { + "bbox": [ + 304, + 489, + 545, + 622 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 304, + 489, + 545, + 622 + ], + "type": "inline_equation", + "content": "\\Psi_g" + }, + { + "bbox": [ + 304, + 489, + 545, + 622 + ], + "type": "text", + "content": " by fitting it to the SDF of " + }, + { + "bbox": [ + 304, + 489, + 545, + 622 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_c" + }, + { + "bbox": [ + 304, + 489, + 545, + 622 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 346, + 631, + 545, + 658 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 346, + 631, + 545, + 658 + ], + "spans": [ + { + "bbox": [ + 346, + 631, + 545, + 658 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {a d a} = \\sum_ {p _ {i} \\in P} | | s (p _ {i}; \\psi) - S D F (p _ {i}) | | _ {2}. \\tag {2}", + "image_path": "39a236f5b6af958010c92b5c642fe909fd2a4bc76053f5f02ffd14c66b1dbcca.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 665, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 545, + 714 + ], + "type": "text", + "content": "Geometry refinement. Using the adapted DMTET, we leverage the highly-detailed normal maps " + }, + { + "bbox": [ + 304, + 665, + 545, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{N}" + }, + { + "bbox": [ + 304, + 665, + 545, + 714 + ], + "type": "text", + "content": " derived from realistic multi-view images as a guidance to refine local surfaces. To obtain the pseudo-GT normals " + }, + { + "bbox": [ + 304, + 665, + 545, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{N}_{\\varphi}" + }, + { + "bbox": [ + 304, + 665, + 545, + 714 + ], + "type": "text", + "content": ", we extract" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "9984" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 194 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 194 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 194 + ], + "type": "text", + "content": "them from " + }, + { + "bbox": [ + 46, + 72, + 289, + 194 + ], + "type": "inline_equation", + "content": "\\mathcal{I}_g^\\varphi" + }, + { + "bbox": [ + 46, + 72, + 289, + 194 + ], + "type": "text", + "content": " using a pre-trained normal estimator [50]. For the rendered normals " + }, + { + "bbox": [ + 46, + 72, + 289, + 194 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{N}}_{\\varphi}" + }, + { + "bbox": [ + 46, + 72, + 289, + 194 + ], + "type": "text", + "content": ", we extract the triangular mesh " + }, + { + "bbox": [ + 46, + 72, + 289, + 194 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_{tri}" + }, + { + "bbox": [ + 46, + 72, + 289, + 194 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 46, + 72, + 289, + 194 + ], + "type": "inline_equation", + "content": "(VT,T)" + }, + { + "bbox": [ + 46, + 72, + 289, + 194 + ], + "type": "text", + "content": " using the Marching Tetrahedra (MT) layer in our current DMTET. By rendering the generated mesh " + }, + { + "bbox": [ + 46, + 72, + 289, + 194 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_{tri}" + }, + { + "bbox": [ + 46, + 72, + 289, + 194 + ], + "type": "text", + "content": " with differentiable rasterization, we obtain the resulting normal map " + }, + { + "bbox": [ + 46, + 72, + 289, + 194 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{N}}_{\\varphi}" + }, + { + "bbox": [ + 46, + 72, + 289, + 194 + ], + "type": "text", + "content": ". To ensure holistic surface polishing that takes into account multi-view normals, we randomly sample camera poses " + }, + { + "bbox": [ + 46, + 72, + 289, + 194 + ], + "type": "inline_equation", + "content": "\\varphi" + }, + { + "bbox": [ + 46, + 72, + 289, + 194 + ], + "type": "text", + "content": " that are uniformly distributed in space. We optimize the parameters of " + }, + { + "bbox": [ + 46, + 72, + 289, + 194 + ], + "type": "inline_equation", + "content": "\\Psi_g" + }, + { + "bbox": [ + 46, + 72, + 289, + 194 + ], + "type": "text", + "content": " using the normal loss, which is defined as:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 116, + 201, + 287, + 217 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 116, + 201, + 287, + 217 + ], + "spans": [ + { + "bbox": [ + 116, + 201, + 287, + 217 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {n o r m}} = \\left\\| \\hat {\\mathcal {N}} _ {\\varphi} - \\mathcal {N} _ {\\varphi} \\right\\| _ {2}. \\tag {3}", + "image_path": "608512222db9380a47d5ccf65505970b69867f3456a08d4d1ddb872e819f6ed7.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 223, + 287, + 297 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 223, + 287, + 297 + ], + "spans": [ + { + "bbox": [ + 46, + 223, + 287, + 297 + ], + "type": "text", + "content": "After rapid optimization, the final triangular mesh " + }, + { + "bbox": [ + 46, + 223, + 287, + 297 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_{tri}" + }, + { + "bbox": [ + 46, + 223, + 287, + 297 + ], + "type": "text", + "content": " can be easily extracted from the MT layer. If the hands exhibit noise, they can be optionally replaced with cleaner geometry hands from SMPL-X, benefiting from the alignment of the generated body in canonical space with the underlying template body." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 302, + 153, + 316 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 302, + 153, + 316 + ], + "spans": [ + { + "bbox": [ + 47, + 302, + 153, + 316 + ], + "type": "text", + "content": "3.3. Explicit texturing" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 320, + 287, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 320, + 287, + 380 + ], + "spans": [ + { + "bbox": [ + 46, + 320, + 287, + 380 + ], + "type": "text", + "content": "With the final mesh, the explicit texturing module aims to disentangle a UV texture map from multi-view renderings " + }, + { + "bbox": [ + 46, + 320, + 287, + 380 + ], + "type": "inline_equation", + "content": "\\mathcal{I}_g^\\varphi" + }, + { + "bbox": [ + 46, + 320, + 287, + 380 + ], + "type": "text", + "content": ". This intuitive module not only facilitates the incorporation of high-fidelity textures but also enables various editing applications, as verified in Section 4.4." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 380, + 288, + 523 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 380, + 288, + 523 + ], + "spans": [ + { + "bbox": [ + 46, + 380, + 288, + 523 + ], + "type": "text", + "content": "Given the polished triangular mesh " + }, + { + "bbox": [ + 46, + 380, + 288, + 523 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_{tri}" + }, + { + "bbox": [ + 46, + 380, + 288, + 523 + ], + "type": "text", + "content": " and multi-views " + }, + { + "bbox": [ + 46, + 380, + 288, + 523 + ], + "type": "inline_equation", + "content": "\\mathcal{I}_g^\\varphi" + }, + { + "bbox": [ + 46, + 380, + 288, + 523 + ], + "type": "text", + "content": ", we model the explicit texture map " + }, + { + "bbox": [ + 46, + 380, + 288, + 523 + ], + "type": "inline_equation", + "content": "T_{uv}" + }, + { + "bbox": [ + 46, + 380, + 288, + 523 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 46, + 380, + 288, + 523 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_{tri}" + }, + { + "bbox": [ + 46, + 380, + 288, + 523 + ], + "type": "text", + "content": " with a semantic UV partition and optimize " + }, + { + "bbox": [ + 46, + 380, + 288, + 523 + ], + "type": "inline_equation", + "content": "T_{uv}" + }, + { + "bbox": [ + 46, + 380, + 288, + 523 + ], + "type": "text", + "content": " using a differentiable rasterizer " + }, + { + "bbox": [ + 46, + 380, + 288, + 523 + ], + "type": "inline_equation", + "content": "\\mathcal{R}" + }, + { + "bbox": [ + 46, + 380, + 288, + 523 + ], + "type": "text", + "content": " [25]. Specifically, leveraging the canonical properties of synthesized bodies, we semantically split " + }, + { + "bbox": [ + 46, + 380, + 288, + 523 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_{tri}" + }, + { + "bbox": [ + 46, + 380, + 288, + 523 + ], + "type": "text", + "content": " into " + }, + { + "bbox": [ + 46, + 380, + 288, + 523 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 46, + 380, + 288, + 523 + ], + "type": "text", + "content": " components and rotate each component vertically, thus enabling effective UV projection for each component with cylinder unwarping. We then combine the texture partitions together for the full texture " + }, + { + "bbox": [ + 46, + 380, + 288, + 523 + ], + "type": "inline_equation", + "content": "T_{uv}" + }, + { + "bbox": [ + 46, + 380, + 288, + 523 + ], + "type": "text", + "content": ". We optimize " + }, + { + "bbox": [ + 46, + 380, + 288, + 523 + ], + "type": "inline_equation", + "content": "T_{uv}" + }, + { + "bbox": [ + 46, + 380, + 288, + 523 + ], + "type": "text", + "content": " from a randomly initialized scratch using the texture loss, which consists of a multi-view reconstruction term and a total-variation (tv) term:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 118, + 533, + 287, + 545 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 118, + 533, + 287, + 545 + ], + "spans": [ + { + "bbox": [ + 118, + 533, + 287, + 545 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {t e x} = \\mathcal {L} _ {r e c} + \\lambda_ {t v} \\mathcal {L} _ {t v}, \\tag {4}", + "image_path": "4d84f9a61f9f4f7917aab48c2a828360f777eadbfb255f8e47f956515d264cbd.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 553, + 225, + 565 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 553, + 225, + 565 + ], + "spans": [ + { + "bbox": [ + 46, + 553, + 225, + 565 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 553, + 225, + 565 + ], + "type": "inline_equation", + "content": "\\lambda_{tv}" + }, + { + "bbox": [ + 46, + 553, + 225, + 565 + ], + "type": "text", + "content": " denotes the weight of the tv loss." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 565, + 287, + 637 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 565, + 287, + 637 + ], + "spans": [ + { + "bbox": [ + 46, + 565, + 287, + 637 + ], + "type": "text", + "content": "Multi-view guidance. To ensure comprehensive texturing in the 3D space, we render the color images " + }, + { + "bbox": [ + 46, + 565, + 287, + 637 + ], + "type": "inline_equation", + "content": "\\mathcal{R}(\\mathcal{M}_{tri},\\varphi)" + }, + { + "bbox": [ + 46, + 565, + 287, + 637 + ], + "type": "text", + "content": " and silhouettes " + }, + { + "bbox": [ + 46, + 565, + 287, + 637 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 46, + 565, + 287, + 637 + ], + "type": "text", + "content": " using " + }, + { + "bbox": [ + 46, + 565, + 287, + 637 + ], + "type": "inline_equation", + "content": "\\mathcal{R}" + }, + { + "bbox": [ + 46, + 565, + 287, + 637 + ], + "type": "text", + "content": " and optimize " + }, + { + "bbox": [ + 46, + 565, + 287, + 637 + ], + "type": "inline_equation", + "content": "T_{uv}" + }, + { + "bbox": [ + 46, + 565, + 287, + 637 + ], + "type": "text", + "content": " utilizing multi-view weighted guidance. Their pixel-alignment distances to the original multi-view renderings " + }, + { + "bbox": [ + 46, + 565, + 287, + 637 + ], + "type": "inline_equation", + "content": "\\mathcal{I}_g^\\varphi" + }, + { + "bbox": [ + 46, + 565, + 287, + 637 + ], + "type": "text", + "content": " are defined as the reconstruction loss:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 74, + 643, + 287, + 669 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 74, + 643, + 287, + 669 + ], + "spans": [ + { + "bbox": [ + 74, + 643, + 287, + 669 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {r e c} = \\sum_ {\\varphi \\in \\Omega} w _ {\\varphi} | | \\mathcal {R} \\left(\\mathcal {M} _ {t r i}, \\varphi\\right) \\cdot \\mathcal {S} - I _ {g} ^ {\\varphi} \\cdot \\mathcal {S} | | _ {2}, \\tag {5}", + "image_path": "79c0d6ec780101ae31f1fe9ef44a3bd50f6bd69e0d3fa68e08441ecbe7fb3cc1.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 677, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 677, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 47, + 677, + 287, + 715 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 677, + 287, + 715 + ], + "type": "inline_equation", + "content": "\\Omega" + }, + { + "bbox": [ + 47, + 677, + 287, + 715 + ], + "type": "text", + "content": " is the set of viewpoints " + }, + { + "bbox": [ + 47, + 677, + 287, + 715 + ], + "type": "inline_equation", + "content": "\\{\\varphi_i, i = 1, \\dots, k\\}" + }, + { + "bbox": [ + 47, + 677, + 287, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 677, + 287, + 715 + ], + "type": "inline_equation", + "content": "w_{\\varphi}" + }, + { + "bbox": [ + 47, + 677, + 287, + 715 + ], + "type": "text", + "content": " denotes weights of different views. " + }, + { + "bbox": [ + 47, + 677, + 287, + 715 + ], + "type": "inline_equation", + "content": "w_{\\varphi}" + }, + { + "bbox": [ + 47, + 677, + 287, + 715 + ], + "type": "text", + "content": " equals to 1.0 for " + }, + { + "bbox": [ + 47, + 677, + 287, + 715 + ], + "type": "inline_equation", + "content": "\\varphi \\in \\{front, back\\}" + }, + { + "bbox": [ + 47, + 677, + 287, + 715 + ], + "type": "text", + "content": " and 0.2 otherwise." + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 306, + 69, + 547, + 173 + ], + "blocks": [ + { + "bbox": [ + 306, + 69, + 547, + 173 + ], + "lines": [ + { + "bbox": [ + 306, + 69, + 547, + 173 + ], + "spans": [ + { + "bbox": [ + 306, + 69, + 547, + 173 + ], + "type": "image", + "image_path": "02980479c6aeb301dc5c2200602c629f3b2063c308353b0567074f2c275e80c5.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 176, + 545, + 199 + ], + "lines": [ + { + "bbox": [ + 305, + 176, + 545, + 199 + ], + "spans": [ + { + "bbox": [ + 305, + 176, + 545, + 199 + ], + "type": "text", + "content": "Figure 3. The visualized flowchart of our method that synthesize textured 3D human avatars from input noises, texts or images." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 224, + 545, + 261 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 224, + 545, + 261 + ], + "spans": [ + { + "bbox": [ + 305, + 224, + 545, + 261 + ], + "type": "text", + "content": "Smooth constraint. To avoid abrupt variations and smooth the generated texture " + }, + { + "bbox": [ + 305, + 224, + 545, + 261 + ], + "type": "inline_equation", + "content": "T_{uv}" + }, + { + "bbox": [ + 305, + 224, + 545, + 261 + ], + "type": "text", + "content": ", we utilize the total-variation loss " + }, + { + "bbox": [ + 305, + 224, + 545, + 261 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{tv}" + }, + { + "bbox": [ + 305, + 224, + 545, + 261 + ], + "type": "text", + "content": " which is computed by:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 339, + 267, + 545, + 291 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 339, + 267, + 545, + 291 + ], + "spans": [ + { + "bbox": [ + 339, + 267, + 545, + 291 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {t v} = \\frac {1}{h \\times w \\times c} | | \\nabla_ {x} \\left(T _ {u v}\\right) + \\nabla_ {y} \\left(T _ {u v}\\right) | |, \\tag {6}", + "image_path": "a9c81ea8b374c0edd8c9b4734063ea1777d9931eb2df1bfc777fa8818035c8b4.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 298, + 531, + 310 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 298, + 531, + 310 + ], + "spans": [ + { + "bbox": [ + 306, + 298, + 531, + 310 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 306, + 298, + 531, + 310 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 306, + 298, + 531, + 310 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 306, + 298, + 531, + 310 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 306, + 298, + 531, + 310 + ], + "type": "text", + "content": " denote horizontal and vertical directions." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 317, + 373, + 329 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 317, + 373, + 329 + ], + "spans": [ + { + "bbox": [ + 306, + 317, + 373, + 329 + ], + "type": "text", + "content": "3.4. Inference" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 335, + 545, + 504 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 335, + 545, + 504 + ], + "spans": [ + { + "bbox": [ + 304, + 335, + 545, + 504 + ], + "type": "text", + "content": "Built upon the above modules, we can generate high-quality 3D human avatars from either random noises or guided inputs such as texts or images. The flowchart for this process is shown in Figure 3. For input noises, we can easily obtain the final results by sequentially using the 3DGM, GS and ET modules. For text-guided synthesis, we first convert the text into a structured image using our controlled diffusion " + }, + { + "bbox": [ + 304, + 335, + 545, + 504 + ], + "type": "inline_equation", + "content": "\\mathcal{C}" + }, + { + "bbox": [ + 304, + 335, + 545, + 504 + ], + "type": "text", + "content": ", and then inverse it to the latent space using PTI [43]. Specially, the GS and ET modules provide an interface that accurately reflects viewed modifications in the final 3D objects. As a result, we utilize the guided image to replace the corresponding view image, which results in improved fidelity in terms of geometry and texture. The same process is applied for input images as guided images." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 306, + 514, + 432, + 528 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 514, + 432, + 528 + ], + "spans": [ + { + "bbox": [ + 306, + 514, + 432, + 528 + ], + "type": "text", + "content": "4. Experimental Results" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 534, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 534, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 534, + 545, + 713 + ], + "type": "text", + "content": "Implementation details. Our process begins by training the 3D generative module (3DGM) on synthetic data. During inference, we integrate the geometric sculpting (GS) and explicit texturing (ET) as optimization modules. For 3DGM, we normalize the template body to the " + }, + { + "bbox": [ + 304, + 534, + 545, + 713 + ], + "type": "inline_equation", + "content": "(0,1)" + }, + { + "bbox": [ + 304, + 534, + 545, + 713 + ], + "type": "text", + "content": " space and place its center at the origin of the world coordinate system. We sample " + }, + { + "bbox": [ + 304, + 534, + 545, + 713 + ], + "type": "inline_equation", + "content": "7(\\mathcal{K} = 7)" + }, + { + "bbox": [ + 304, + 534, + 545, + 713 + ], + "type": "text", + "content": " viewpoints uniformly from the horizontal plane, ranging from " + }, + { + "bbox": [ + 304, + 534, + 545, + 713 + ], + "type": "inline_equation", + "content": "0^{\\circ}" + }, + { + "bbox": [ + 304, + 534, + 545, + 713 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 304, + 534, + 545, + 713 + ], + "type": "inline_equation", + "content": "180^{\\circ}" + }, + { + "bbox": [ + 304, + 534, + 545, + 713 + ], + "type": "text", + "content": " (front to back), with a camera radius of 2.7. For each viewpoint, we generate " + }, + { + "bbox": [ + 304, + 534, + 545, + 713 + ], + "type": "inline_equation", + "content": "100K" + }, + { + "bbox": [ + 304, + 534, + 545, + 713 + ], + "type": "text", + "content": " images using the corresponding pose image. To ensure diverse synthesis, we use detailed descriptions of age, gender, ethnicity, hairstyle, facial features, and clothing, leveraging a vast word bank. To cover " + }, + { + "bbox": [ + 304, + 534, + 545, + 713 + ], + "type": "inline_equation", + "content": "360^{\\circ}" + }, + { + "bbox": [ + 304, + 534, + 545, + 713 + ], + "type": "text", + "content": " views, we horizontally flip the synthesized images and obtain 1.4 million human images at a resolution of " + }, + { + "bbox": [ + 304, + 534, + 545, + 713 + ], + "type": "inline_equation", + "content": "512^{2}" + }, + { + "bbox": [ + 304, + 534, + 545, + 713 + ], + "type": "text", + "content": " in total. We train" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "9985" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 47, + 70, + 547, + 416 + ], + "blocks": [ + { + "bbox": [ + 47, + 70, + 547, + 416 + ], + "lines": [ + { + "bbox": [ + 47, + 70, + 547, + 416 + ], + "spans": [ + { + "bbox": [ + 47, + 70, + 547, + 416 + ], + "type": "image", + "image_path": "7aed5913e1c5aa7b1793cb911d0316caf918d886877e533dd8741aae16715624.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 187, + 416, + 403, + 427 + ], + "lines": [ + { + "bbox": [ + 187, + 416, + 403, + 427 + ], + "spans": [ + { + "bbox": [ + 187, + 416, + 403, + 427 + ], + "type": "text", + "content": "Figure 4. Results of synthesized 3D human avatars at " + }, + { + "bbox": [ + 187, + 416, + 403, + 427 + ], + "type": "inline_equation", + "content": "512^{2}" + }, + { + "bbox": [ + 187, + 416, + 403, + 427 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 444, + 289, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 444, + 289, + 601 + ], + "spans": [ + { + "bbox": [ + 46, + 444, + 289, + 601 + ], + "type": "text", + "content": "the 3DGM for about 2.5M iterations with a batch size of 32, using two discriminators with a learning rate of 0.002 and a generator learning rate of 0.0025. The training takes 8 days using 8 NVIDIA Tesla-V100 GPUs. For GS, we optimize " + }, + { + "bbox": [ + 46, + 444, + 289, + 601 + ], + "type": "inline_equation", + "content": "\\psi" + }, + { + "bbox": [ + 46, + 444, + 289, + 601 + ], + "type": "text", + "content": " for 400 iterations for DMTET adaption and 100 iterations for surface carving (taking about 15s in total on 1 NVIDIA RTX 3090 GPU). For ET, we set " + }, + { + "bbox": [ + 46, + 444, + 289, + 601 + ], + "type": "inline_equation", + "content": "\\lambda_{uv} = 1" + }, + { + "bbox": [ + 46, + 444, + 289, + 601 + ], + "type": "text", + "content": " and optimize " + }, + { + "bbox": [ + 46, + 444, + 289, + 601 + ], + "type": "inline_equation", + "content": "T_{uv}" + }, + { + "bbox": [ + 46, + 444, + 289, + 601 + ], + "type": "text", + "content": " for 500 iterations (around 10 seconds). We split " + }, + { + "bbox": [ + 46, + 444, + 289, + 601 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_{tri}" + }, + { + "bbox": [ + 46, + 444, + 289, + 601 + ], + "type": "text", + "content": " into " + }, + { + "bbox": [ + 46, + 444, + 289, + 601 + ], + "type": "inline_equation", + "content": "5(\\gamma = 5)" + }, + { + "bbox": [ + 46, + 444, + 289, + 601 + ], + "type": "text", + "content": " body parts (i.e., trunk, left/right arm/leg) with cylinder UV unwarping. We use the Adam optimizer with learning rates of 0.01 and 0.001 for " + }, + { + "bbox": [ + 46, + 444, + 289, + 601 + ], + "type": "inline_equation", + "content": "\\Psi_g" + }, + { + "bbox": [ + 46, + 444, + 289, + 601 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 444, + 289, + 601 + ], + "type": "inline_equation", + "content": "T_{uv}" + }, + { + "bbox": [ + 46, + 444, + 289, + 601 + ], + "type": "text", + "content": ", respectively. Detailed network architectures can be found in the supplemental materials (Suppl)." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 610, + 173, + 623 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 610, + 173, + 623 + ], + "spans": [ + { + "bbox": [ + 47, + 610, + 173, + 623 + ], + "type": "text", + "content": "4.1. 3D human generation" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 629, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 629, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 629, + 287, + 715 + ], + "type": "text", + "content": "Figure 4 showcases several 3D human avatars synthesized by our pipeline, highlighting the image quality, geometry accuracy, and diverse outputs achieved through our method. Additionally, we explore the interpolation of the latent conditions to yield smooth transitions in appearance, leveraging the smooth latent space learned by our generative model. For more synthesized examples and interpolation results," + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 306, + 445, + 408, + 457 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 445, + 408, + 457 + ], + "spans": [ + { + "bbox": [ + 306, + 445, + 408, + 457 + ], + "type": "text", + "content": "please refer to the Suppl." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 466, + 391, + 479 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 466, + 391, + 479 + ], + "spans": [ + { + "bbox": [ + 306, + 466, + 391, + 479 + ], + "type": "text", + "content": "4.2. Comparisons" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 485, + 545, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 485, + 545, + 664 + ], + "spans": [ + { + "bbox": [ + 304, + 485, + 545, + 664 + ], + "type": "text", + "content": "Qualitative comparison. In Figure 5, we compare our method with three baselines: EVA3D [15] and AG3D [9], which are state-of-the-art methods for generating 3D humans from 2D images, and EG3D [5], which serves as the foundational backbone of our method. The results of first two methods are produced by directly using source codes and trained models released by authors. We train EG3D using our synthetic images with estimated cameras from scratch. As we can see, EVA3D fails to produce " + }, + { + "bbox": [ + 304, + 485, + 545, + 664 + ], + "type": "inline_equation", + "content": "360^{\\circ}" + }, + { + "bbox": [ + 304, + 485, + 545, + 664 + ], + "type": "text", + "content": " humans with reasonable back inferring. AG3D and EG3D are able to generate " + }, + { + "bbox": [ + 304, + 485, + 545, + 664 + ], + "type": "inline_equation", + "content": "360^{\\circ}" + }, + { + "bbox": [ + 304, + 485, + 545, + 664 + ], + "type": "text", + "content": " renderings but both struggle with photorealism and capturing detailed shapes. Our method synthesizes not only higher-quality, view-consistent " + }, + { + "bbox": [ + 304, + 485, + 545, + 664 + ], + "type": "inline_equation", + "content": "360^{\\circ}" + }, + { + "bbox": [ + 304, + 485, + 545, + 664 + ], + "type": "text", + "content": " images but also higher-fidelity 3D geometry with intricate details, such as irregular dresses and haircuts." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 665, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 546, + 713 + ], + "type": "text", + "content": "Quantitative comparison. Table 1 provides quantitative results comparing our method against the baselines. We measure image quality with Frechet Inception Distance (FID) [14] and Inception Score [45] for " + }, + { + "bbox": [ + 304, + 665, + 546, + 713 + ], + "type": "inline_equation", + "content": "360^{\\circ}" + }, + { + "bbox": [ + 304, + 665, + 546, + 713 + ], + "type": "text", + "content": " views (IS-" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "9986" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 52, + 69, + 547, + 379 + ], + "blocks": [ + { + "bbox": [ + 52, + 69, + 547, + 379 + ], + "lines": [ + { + "bbox": [ + 52, + 69, + 547, + 379 + ], + "spans": [ + { + "bbox": [ + 52, + 69, + 547, + 379 + ], + "type": "image", + "image_path": "8eceb44aa9b924d4e0e2baa8eb42244150844cfa6809d3484dd1dae0082b5330.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 97, + 380, + 493, + 392 + ], + "lines": [ + { + "bbox": [ + 97, + 380, + 493, + 392 + ], + "spans": [ + { + "bbox": [ + 97, + 380, + 493, + 392 + ], + "type": "text", + "content": "Figure 5. Qualitative comparison with three state-of-the-art methods: EVA3D [15], AG3D [9] and EG3D [5]." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 61, + 430, + 272, + 521 + ], + "blocks": [ + { + "bbox": [ + 46, + 406, + 287, + 429 + ], + "lines": [ + { + "bbox": [ + 46, + 406, + 287, + 429 + ], + "spans": [ + { + "bbox": [ + 46, + 406, + 287, + 429 + ], + "type": "text", + "content": "Table 1. Quantitative evaluation using FID, IS-360, normal accuracy (Normal) and identity consistency (ID)." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 61, + 430, + 272, + 521 + ], + "lines": [ + { + "bbox": [ + 61, + 430, + 272, + 521 + ], + "spans": [ + { + "bbox": [ + 61, + 430, + 272, + 521 + ], + "type": "table", + "html": "
MethodFID ↓IS-360 ↑Normal ↓ID↑
EVA3D [15]15.913.1930.810.72
AG3D [9]10.933.2820.830.69
EVA3D-syn6.273.259.570.72
AG3D-syn8.933.228.640.71
EG3D [5]7.483.2612.740.71
Ours2.733.435.620.74
", + "image_path": "de17e76b97243efd18bb4d15c898b4c0f969df22da12a7b1bc9d4d3e7190697e.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 545, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 545, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 545, + 289, + 715 + ], + "type": "text", + "content": "360). FID measures the visual similarity and distribution discrepancy between 50k generated images and all real images. IS-360 focuses on the self-realism of generated images in " + }, + { + "bbox": [ + 46, + 545, + 289, + 715 + ], + "type": "inline_equation", + "content": "360^{\\circ}" + }, + { + "bbox": [ + 46, + 545, + 289, + 715 + ], + "type": "text", + "content": " views. For shape evaluation, we compute FID between rendered normals and pseudo-GT normal maps (Normal), following AG3D. The FID and Normal scores of EVA3D and AG3D are directly fetched from their reports. Additionally, we access the multi-view facial identity consistency using the ID metric introduced by EG3D. For a clearer comparison, we also provide metric results of EVA3D and AG3D trained on our synthetic data (-syn). Overall, our method demonstrates significant improvements in FID and Normal, bringing the generative human model to a new level of realistic " + }, + { + "bbox": [ + 46, + 545, + 289, + 715 + ], + "type": "inline_equation", + "content": "360^{\\circ}" + }, + { + "bbox": [ + 46, + 545, + 289, + 715 + ], + "type": "text", + "content": " renderings with delicate ge" + } + ] + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 318, + 441, + 533, + 488 + ], + "blocks": [ + { + "bbox": [ + 304, + 406, + 545, + 440 + ], + "lines": [ + { + "bbox": [ + 304, + 406, + 545, + 440 + ], + "spans": [ + { + "bbox": [ + 304, + 406, + 545, + 440 + ], + "type": "text", + "content": "Table 2. Results of models trained by replacing physical parameters with estimated ones (w/o SYN-P) or removing patch-composed rendering (w/o PCR)." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 318, + 441, + 533, + 488 + ], + "lines": [ + { + "bbox": [ + 318, + 441, + 533, + 488 + ], + "spans": [ + { + "bbox": [ + 318, + 441, + 533, + 488 + ], + "type": "table", + "html": "
OursOurs-w/o SYN-POurs-w/o PCR
FID ↓2.734.283.26
IS-360 ↑3.433.313.35
", + "image_path": "052fa6b1099cf9a07d05e0931f49fcea8fb28e3a918b9814592058e03da80542.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 508, + 545, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 508, + 545, + 521 + ], + "spans": [ + { + "bbox": [ + 305, + 508, + 545, + 521 + ], + "type": "text", + "content": "ometry while maintaining state-of-the-art view consistency." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 527, + 398, + 540 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 527, + 398, + 540 + ], + "spans": [ + { + "bbox": [ + 306, + 527, + 398, + 540 + ], + "type": "text", + "content": "4.3. Ablation study" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 545, + 545, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 545, + 545, + 665 + ], + "spans": [ + { + "bbox": [ + 304, + 545, + 545, + 665 + ], + "type": "text", + "content": "Synthesis flow and patch-composed rendering. We assess the impact of our carefully designed synthesis flow by training a model with synthetic images but with camera and pose parameters estimated by SMPLify-X [40] (w/o SYN-P). As Table 2 shows, the model w/o SYN-P results in worse FID and IS-360 scores, indicating that the synthesis flow contributes to more accurate physical parameters for realistic appearance modeling. By utilizing patch-composed rendering (PCR), the networks focus more on the human region, leading to more realistic results." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 665, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 547, + 715 + ], + "type": "text", + "content": "Geometry sculpting module (GS). We demonstrate the importance of this module by visualizing the meshes before and after its implementation. Figure 6 (b) shows that the preceding module yields a coarse mesh due to the complex" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "9987" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 52, + 72, + 110, + 156 + ], + "blocks": [ + { + "bbox": [ + 52, + 72, + 110, + 156 + ], + "lines": [ + { + "bbox": [ + 52, + 72, + 110, + 156 + ], + "spans": [ + { + "bbox": [ + 52, + 72, + 110, + 156 + ], + "type": "image", + "image_path": "8537eac5ff968ce7e4b86fb01af23539baaf0effb4de65a6e01d86a6a9a015d3.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 53, + 158, + 101, + 167 + ], + "lines": [ + { + "bbox": [ + 53, + 158, + 101, + 167 + ], + "spans": [ + { + "bbox": [ + 53, + 158, + 101, + 167 + ], + "type": "text", + "content": "(a) Rendering" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 111, + 72, + 171, + 156 + ], + "blocks": [ + { + "bbox": [ + 111, + 72, + 171, + 156 + ], + "lines": [ + { + "bbox": [ + 111, + 72, + 171, + 156 + ], + "spans": [ + { + "bbox": [ + 111, + 72, + 171, + 156 + ], + "type": "image", + "image_path": "4cc72be7a7089040d22569f9920fbca696ab319f2661a67270d18c062954e5fd.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 119, + 158, + 157, + 167 + ], + "lines": [ + { + "bbox": [ + 119, + 158, + 157, + 167 + ], + "spans": [ + { + "bbox": [ + 119, + 158, + 157, + 167 + ], + "type": "text", + "content": "(b) w/o GS" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 173, + 72, + 232, + 156 + ], + "blocks": [ + { + "bbox": [ + 173, + 72, + 232, + 156 + ], + "lines": [ + { + "bbox": [ + 173, + 72, + 232, + 156 + ], + "spans": [ + { + "bbox": [ + 173, + 72, + 232, + 156 + ], + "type": "image", + "image_path": "abf98b26aba27326a1c6e9a2e538f9aded98da542dda48a83a4a162b73f86b72.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 200, + 158, + 253, + 167 + ], + "lines": [ + { + "bbox": [ + 200, + 158, + 253, + 167 + ], + "spans": [ + { + "bbox": [ + 200, + 158, + 253, + 167 + ], + "type": "text", + "content": "(c) Full method" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 235, + 76, + 286, + 156 + ], + "blocks": [ + { + "bbox": [ + 235, + 76, + 286, + 156 + ], + "lines": [ + { + "bbox": [ + 235, + 76, + 286, + 156 + ], + "spans": [ + { + "bbox": [ + 235, + 76, + 286, + 156 + ], + "type": "image", + "image_path": "93db33d01c4539a9f13305c65fc09867c17514a6f378dd8e40164170b106125f.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 56, + 196, + 113, + 280 + ], + "blocks": [ + { + "bbox": [ + 47, + 171, + 286, + 182 + ], + "lines": [ + { + "bbox": [ + 47, + 171, + 286, + 182 + ], + "spans": [ + { + "bbox": [ + 47, + 171, + 286, + 182 + ], + "type": "text", + "content": "Figure 6. Effects of the GS module to carve fine-grained surfaces." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 56, + 196, + 113, + 280 + ], + "lines": [ + { + "bbox": [ + 56, + 196, + 113, + 280 + ], + "spans": [ + { + "bbox": [ + 56, + 196, + 113, + 280 + ], + "type": "image", + "image_path": "7c4797b81ef35a91dfd5910f7be2a8c9c8e618e521c93f2b21df1d388196c4a3.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 53, + 281, + 111, + 291 + ], + "lines": [ + { + "bbox": [ + 53, + 281, + 111, + 291 + ], + "spans": [ + { + "bbox": [ + 53, + 281, + 111, + 291 + ], + "type": "text", + "content": "(a) Guided image" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 116, + 196, + 171, + 281 + ], + "blocks": [ + { + "bbox": [ + 116, + 196, + 171, + 281 + ], + "lines": [ + { + "bbox": [ + 116, + 196, + 171, + 281 + ], + "spans": [ + { + "bbox": [ + 116, + 196, + 171, + 281 + ], + "type": "image", + "image_path": "1b31d5535ae48e32a8971579db57f0de9c7157d9f719c0ca81ca92d16a5532a4.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 125, + 281, + 162, + 291 + ], + "lines": [ + { + "bbox": [ + 125, + 281, + 162, + 291 + ], + "spans": [ + { + "bbox": [ + 125, + 281, + 162, + 291 + ], + "type": "text", + "content": "(b) w/o ET" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 64, + 295, + 269, + 306 + ], + "lines": [ + { + "bbox": [ + 64, + 295, + 269, + 306 + ], + "spans": [ + { + "bbox": [ + 64, + 295, + 269, + 306 + ], + "type": "text", + "content": "Figure 7. Effects of the ET module for guided synthesis." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 176, + 196, + 232, + 281 + ], + "blocks": [ + { + "bbox": [ + 176, + 196, + 232, + 281 + ], + "lines": [ + { + "bbox": [ + 176, + 196, + 232, + 281 + ], + "spans": [ + { + "bbox": [ + 176, + 196, + 232, + 281 + ], + "type": "image", + "image_path": "73e7659150accc5a8ed134f271776f58a0194b43b6117cfbb3776dcc7eb462ca.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 204, + 281, + 255, + 291 + ], + "lines": [ + { + "bbox": [ + 204, + 281, + 255, + 291 + ], + "spans": [ + { + "bbox": [ + 204, + 281, + 255, + 291 + ], + "type": "text", + "content": "(c) Full method" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 235, + 196, + 286, + 281 + ], + "blocks": [ + { + "bbox": [ + 235, + 196, + 286, + 281 + ], + "lines": [ + { + "bbox": [ + 235, + 196, + 286, + 281 + ], + "spans": [ + { + "bbox": [ + 235, + 196, + 286, + 281 + ], + "type": "image", + "image_path": "9afc8ce98f79de350a856978094f4cb2a80bdf96b59c8fa9cbfe672e5f779906.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "bbox": [ + 46, + 328, + 286, + 424 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 328, + 286, + 424 + ], + "spans": [ + { + "bbox": [ + 46, + 328, + 286, + 424 + ], + "type": "text", + "content": "human anatomy and the challenges posed by decomposing ambiguous 3D shapes from 2D images. The GS module utilizes high-quality multi-view outputs and employs a more flexible hybrid representation to create expressive humans with arbitrary topologies. It learns from pixel-level surface supervision, leading to a significant improvement in shape quality, characterized by smooth surfaces and intricate outfits (Figure 6 (c))." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 46, + 425, + 286, + 580 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 425, + 286, + 580 + ], + "spans": [ + { + "bbox": [ + 46, + 425, + 286, + 580 + ], + "type": "text", + "content": "Explicit texturing module (ET). This intuitive module not only extracts the explicit UV texture for complete 3D assets but also enables high-fidelity results for image guided synthesis. Following the flowchart in Figure 3, we compare the results produced with and without this module. Our method without ET directly generates implicit renderings through PTI inversion, as shown in Figure 7 (b). While it successfully preserves global identity, it struggles to synthesize highly faithful local textures (e.g., floral patterns). The ET module offers a convenient and efficient way to directly interact with the 3D representation, enabling the production of high-fidelity 3D humans with more consistent content including exquisite local patterns (Figure 7 (a, c))." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 47, + 587, + 129, + 600 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 587, + 129, + 600 + ], + "spans": [ + { + "bbox": [ + 47, + 587, + 129, + 600 + ], + "type": "text", + "content": "4.4. Applications" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 46, + 605, + 286, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 286, + 666 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 286, + 666 + ], + "type": "text", + "content": "Avatar animation. All avatars produced by our method are in a canonical body pose and aligned to an underlying 3D skeleton extracted from SMPL-X. This alignment allows for easy animation and the generation of motion videos, as demonstrated in Figure 1 and Suppl." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 46, + 666, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 666, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 666, + 286, + 713 + ], + "type": "text", + "content": "Texture doodle and local editing. Our approach benefits from the explicit disentanglement of geometry and texture, enabling flexible editing capabilities. Following the flowchart of text or image guided synthesis (Section 3.4)," + } + ] + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 307, + 71, + 406, + 150 + ], + "blocks": [ + { + "bbox": [ + 307, + 71, + 406, + 150 + ], + "lines": [ + { + "bbox": [ + 307, + 71, + 406, + 150 + ], + "spans": [ + { + "bbox": [ + 307, + 71, + 406, + 150 + ], + "type": "image", + "image_path": "9a378752288b93e83b8a16f8a95e63c26c8f2fecb30cbf6c0cc9885649a09791.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 406, + 72, + 492, + 149 + ], + "blocks": [ + { + "bbox": [ + 406, + 72, + 492, + 149 + ], + "lines": [ + { + "bbox": [ + 406, + 72, + 492, + 149 + ], + "spans": [ + { + "bbox": [ + 406, + 72, + 492, + 149 + ], + "type": "image", + "image_path": "583a0aad33465cce86484fb0f9b15aa788adc4e18e73e6b916bac8029c63d208.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 493, + 72, + 545, + 149 + ], + "blocks": [ + { + "bbox": [ + 493, + 72, + 545, + 149 + ], + "lines": [ + { + "bbox": [ + 493, + 72, + 545, + 149 + ], + "spans": [ + { + "bbox": [ + 493, + 72, + 545, + 149 + ], + "type": "image", + "image_path": "70995fc2ada3f1fc26e4b2fe8b0e376047f51a38d3a01c66bb2473c40ac342e7.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 307, + 151, + 404, + 202 + ], + "blocks": [ + { + "bbox": [ + 307, + 151, + 404, + 202 + ], + "lines": [ + { + "bbox": [ + 307, + 151, + 404, + 202 + ], + "spans": [ + { + "bbox": [ + 307, + 151, + 404, + 202 + ], + "type": "image", + "image_path": "d1ca738bad190fd3d970be65da1a0523155ff37b795d55f60f21ad9599b13c99.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 205, + 545, + 237 + ], + "lines": [ + { + "bbox": [ + 305, + 205, + 545, + 237 + ], + "spans": [ + { + "bbox": [ + 305, + 205, + 545, + 237 + ], + "type": "text", + "content": "Figure 8. Results synthesized by adapting our method to various styles (e.g., Disney cartoon characters) or contents (e.g., portrait heads)." + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_caption" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 404, + 151, + 492, + 202 + ], + "blocks": [ + { + "bbox": [ + 404, + 151, + 492, + 202 + ], + "lines": [ + { + "bbox": [ + 404, + 151, + 492, + 202 + ], + "spans": [ + { + "bbox": [ + 404, + 151, + 492, + 202 + ], + "type": "image", + "image_path": "057816ebef0725a7559f2491d385fa0718a06e7d46476a0e286f818230fde632.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 493, + 151, + 545, + 202 + ], + "blocks": [ + { + "bbox": [ + 493, + 151, + 545, + 202 + ], + "lines": [ + { + "bbox": [ + 493, + 151, + 545, + 202 + ], + "spans": [ + { + "bbox": [ + 493, + 151, + 545, + 202 + ], + "type": "image", + "image_path": "c9c097e0e8aca241825753fbdde53c38c562a6cf475ee50049f556cde5b46803.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + } + ], + "index": 26 + }, + { + "bbox": [ + 304, + 267, + 545, + 375 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 267, + 545, + 375 + ], + "spans": [ + { + "bbox": [ + 304, + 267, + 545, + 375 + ], + "type": "text", + "content": "users can paint any pattern or add text to a guided image. These modifications can be transferred to 3D human models by inputting modified views into the texture module (e.g., painting the text 'hey' on a jacket as shown in Figure 1 (d)). Our approach also allows for clothing editing by simultaneously injecting edited guide images with desired clothing into the GS and ET modules (e.g., changing a jacket and jeans to bodysuits in Figure 1 (e)). More results can be found in Suppl." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 304, + 376, + 545, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 376, + 545, + 472 + ], + "spans": [ + { + "bbox": [ + 304, + 376, + 545, + 472 + ], + "type": "text", + "content": "Content-style free adaption. Our proposed scheme is versatile and can be extended to generate various types of contents (e.g., portrait heads) and styles (e.g., Disney cartoon characters). To achieve this, we fine-tune our model using synthetic images from these domains, allowing for flexible adaptation. We showcase the results in Figure 8. More results and other discussions (e.g., limitations, negative impact, etc.) can be found in the Suppl." + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 306, + 488, + 378, + 500 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 488, + 378, + 500 + ], + "spans": [ + { + "bbox": [ + 306, + 488, + 378, + 500 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "text", + "content": "We introduced En3D, a novel generative scheme for sculpting 3D humans from 2D synthetic data. This method overcomes limitations in existing 3D or 2D collections and significantly enhances the image quality, geometry accuracy, and content diversity of generated 3D humans. En3D comprises a 3D generative module that learns generalizable 3D humans from synthetic 2D data with accurate physical modeling, and two optimization modules to carve intricate shape details and disentangle explicit UV textures with high fidelity, respectively. Experimental results validated the superiority and effectiveness of our method. We also demonstrated the flexibility of our generated avatars for animation and editing, as well as the scalability of our approach for synthesizing portraits and Disney characters. We believe that our solution could provide invaluable human assets for the 3D vision community. Furthermore, it holds potential for use in common 3D object synthesis tasks." + } + ] + } + ], + "index": 31 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "9988" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 714 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 135 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 135 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 135 + ], + "type": "text", + "content": "[1] Panos Achlioptas, Olga Diamanti, Ioannis Mitliagkas, and Leonidas Guibas. Learning representations and generative models for 3d point clouds. In International conference on machine learning, pages 40-49. PMLR, 2018. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 137, + 288, + 192 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 137, + 288, + 192 + ], + "spans": [ + { + "bbox": [ + 53, + 137, + 288, + 192 + ], + "type": "text", + "content": "[2] Thiemo Alldieck, Marcus Magnor, Weipeng Xu, Christian Theobalt, and Gerard Pons-Moll. Video based reconstruction of 3d people models. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 8387-8397, 2018. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 194, + 288, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 194, + 288, + 237 + ], + "spans": [ + { + "bbox": [ + 53, + 194, + 288, + 237 + ], + "type": "text", + "content": "[3] Dragomir Anguelov, Praveen Srinivasan, Daphne Koller, Sebastian Thrun, Jim Rodgers, and James Davis. Scape: shape completion and animation of people. In ACM SIGGRAPH 2005 Papers, pages 408-416. 2005. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 239, + 288, + 294 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 239, + 288, + 294 + ], + "spans": [ + { + "bbox": [ + 53, + 239, + 288, + 294 + ], + "type": "text", + "content": "[4] Eric R Chan, Marco Monteiro, Petr Kellnhofer, Jiajun Wu, and Gordon Wetzstein. pi-gan: Periodic implicit generative adversarial networks for 3d-aware image synthesis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5799-5809, 2021. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 295, + 288, + 372 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 295, + 288, + 372 + ], + "spans": [ + { + "bbox": [ + 53, + 295, + 288, + 372 + ], + "type": "text", + "content": "[5] Eric R Chan, Connor Z Lin, Matthew A Chan, Koki Nagano, Boxiao Pan, Shalini De Mello, Orazio Gallo, Leonidas J Guibas, Jonathan Tremblay, Sameh Khamis, et al. Efficient geometry-aware 3d generative adversarial networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16123-16133, 2022. 2, 3, 4, 6, 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 373, + 288, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 373, + 288, + 430 + ], + "spans": [ + { + "bbox": [ + 53, + 373, + 288, + 430 + ], + "type": "text", + "content": "[6] Xu Chen, Tianjian Jiang, Jie Song, Jinlong Yang, Michael J Black, Andreas Geiger, and Otmar Hilliges. gdna: Towards generative detailed neural avatars. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20427-20437, 2022. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 431, + 288, + 485 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 431, + 288, + 485 + ], + "spans": [ + { + "bbox": [ + 53, + 431, + 288, + 485 + ], + "type": "text", + "content": "[7] Yu Deng, Jiaolong Yang, Jianfeng Xiang, and Xin Tong. Gram: Generative radiance manifolds for 3d-aware image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10673-10683, 2022. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 487, + 288, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 487, + 288, + 542 + ], + "spans": [ + { + "bbox": [ + 53, + 487, + 288, + 542 + ], + "type": "text", + "content": "[8] Zijian Dong, Chen Guo, Jie Song, Xu Chen, Andreas Geiger, and Otmar Hilliges. Pina: Learning a personalized implicit neural avatar from a single rgb-d video sequence. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20470-20480, 2022. 1, 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 544, + 288, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 544, + 288, + 588 + ], + "spans": [ + { + "bbox": [ + 53, + 544, + 288, + 588 + ], + "type": "text", + "content": "[9] Zijian Dong, Xu Chen, Jinlong Yang, Michael J Black, Otmar Hilliges, and Andreas Geiger. Ag3d: Learning to generate 3d avatars from 2d image collections. arXiv preprint arXiv:2305.02312, 2023. 2, 3, 6, 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 590, + 287, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 590, + 287, + 633 + ], + "spans": [ + { + "bbox": [ + 48, + 590, + 287, + 633 + ], + "type": "text", + "content": "[10] Yao Feng, Jinlong Yang, Marc Pollefeys, Michael J Black, and Timo Bolkart. Capturing and animation of body and clothing from monocular video. In SIGGRAPH Asia 2022 Conference Papers, pages 1-9, 2022. 1, 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 635, + 287, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 635, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 635, + 287, + 689 + ], + "type": "text", + "content": "[11] Jianglin Fu, Shikai Li, Yuming Jiang, Kwan-Yee Lin, Chen Qian, Chen Change Loy, Wayne Wu, and Ziwei Liu. Stylegan-human: A data-centric odyssey of human generation. In European Conference on Computer Vision, pages 1-19. Springer, 2022. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 691, + 287, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 287, + 714 + ], + "type": "text", + "content": "[12] Jiatao Gu, Lingjie Liu, Peng Wang, and Christian Theobalt. Stylenerf: A style-based 3d-aware genera" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "text", + "content": "tor for high-resolution image synthesis. arXiv preprint arXiv:2110.08985, 2021.3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 308, + 96, + 545, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 96, + 545, + 139 + ], + "spans": [ + { + "bbox": [ + 308, + 96, + 545, + 139 + ], + "type": "text", + "content": "[13] Philipp Henzler, Niloy J Mitra, and Tobias Ritschel. Escaping plato's cave: 3d shape from adversarial rendering. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9984-9993, 2019. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 140, + 545, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 140, + 545, + 194 + ], + "spans": [ + { + "bbox": [ + 308, + 140, + 545, + 194 + ], + "type": "text", + "content": "[14] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 195, + 545, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 195, + 545, + 239 + ], + "spans": [ + { + "bbox": [ + 308, + 195, + 545, + 239 + ], + "type": "text", + "content": "[15] Fangzhou Hong, Zhaoxi Chen, Yushi Lan, Liang Pan, and Ziwei Liu. Eva3d: Compositional 3d human generation from 2d image collections. arXiv preprint arXiv:2210.04888, 2022. 2, 3, 6, 7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 240, + 545, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 240, + 545, + 282 + ], + "spans": [ + { + "bbox": [ + 308, + 240, + 545, + 282 + ], + "type": "text", + "content": "[16] Xin Huang, Ruizhi Shao, Qi Zhang, Hongwen Zhang, Ying Feng, Yebin Liu, and Qing Wang. Humannorm: Learning normal diffusion model for high-quality and realistic 3d human generation. arXiv preprint arXiv:2310.01406, 2023. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 284, + 545, + 338 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 284, + 545, + 338 + ], + "spans": [ + { + "bbox": [ + 308, + 284, + 545, + 338 + ], + "type": "text", + "content": "[17] Suyi Jiang, Haoran Jiang, Ziyu Wang, Haimin Luo, Wenzheng Chen, and Lan Xu. Humangen: Generating human radiance fields with explicit priors. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12543-12554, 2023. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 339, + 545, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 339, + 545, + 392 + ], + "spans": [ + { + "bbox": [ + 308, + 339, + 545, + 392 + ], + "type": "text", + "content": "[18] Kyungmin Jo, Wonjoon Jin, Jaegul Choo, Hyunjoon Lee, and Sunghyun Cho. 3d-aware generative model for improved side-view image synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 22862-22872, 2023. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 394, + 545, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 394, + 545, + 437 + ], + "spans": [ + { + "bbox": [ + 308, + 394, + 545, + 437 + ], + "type": "text", + "content": "[19] Hanbyul Joo, Tomas Simon, and Yaser Sheikh. Total capture: A 3d deformation model for tracking faces, hands, and bodies. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 8320-8329, 2018. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 438, + 545, + 481 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 438, + 545, + 481 + ], + "spans": [ + { + "bbox": [ + 308, + 438, + 545, + 481 + ], + "type": "text", + "content": "[20] Angjoo Kanazawa, Michael J Black, David W Jacobs, and Jitendra Malik. End-to-end recovery of human shape and pose. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7122-7131, 2018. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 483, + 545, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 483, + 545, + 525 + ], + "spans": [ + { + "bbox": [ + 308, + 483, + 545, + 525 + ], + "type": "text", + "content": "[21] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4401-4410, 2019. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 526, + 545, + 581 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 526, + 545, + 581 + ], + "spans": [ + { + "bbox": [ + 308, + 526, + 545, + 581 + ], + "type": "text", + "content": "[22] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of stylegan. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8110-8119, 2020. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 582, + 545, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 582, + 545, + 624 + ], + "spans": [ + { + "bbox": [ + 308, + 582, + 545, + 624 + ], + "type": "text", + "content": "[23] Tero Karras, Miika Aittala, Samuli Laine, Erik Härkönen, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Alias-free generative adversarial networks. Advances in Neural Information Processing Systems, 34:852-863, 2021. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 625, + 545, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 625, + 545, + 669 + ], + "spans": [ + { + "bbox": [ + 308, + 625, + 545, + 669 + ], + "type": "text", + "content": "[24] Nikos Kolotouros, Thiemo Alldieck, Andrei Zanfir, Eduard Bazavan, Mihai Fieraru, and Cristian Sminchisescu. Dreamhuman: Animatable 3d avatars from text. Advances in Neural Information Processing Systems, 36, 2024. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 670, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 670, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 670, + 545, + 713 + ], + "type": "text", + "content": "[25] Samuli Laine, Janne Hellsten, Tero Karras, Yeongho Seol, Jaakko Lehtinen, and Timo Aila. Modular primitives for high-performance differentiable rendering. ACM Transactions on Graphics (TOG), 39(6):1-14, 2020. 5" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "9989" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 126 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 126 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 126 + ], + "type": "text", + "content": "[26] Ruihui Li, Xianzhi Li, Chi-Wing Fu, Daniel Cohen-Or, and Pheng-Ann Heng. Pu-gan: a point cloud upsampling adversarial network. In Proceedings of the IEEE/CVF international conference on computer vision, pages 7203–7212, 2019. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 129, + 288, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 129, + 288, + 183 + ], + "spans": [ + { + "bbox": [ + 48, + 129, + 288, + 183 + ], + "type": "text", + "content": "[27] Yiyi Liao, Katja Schwarz, Lars Mescheder, and Andreas Geiger. Towards unsupervised learning of generative models for 3d controllable image synthesis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5871-5880, 2020. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 184, + 288, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 184, + 288, + 237 + ], + "spans": [ + { + "bbox": [ + 48, + 184, + 288, + 237 + ], + "type": "text", + "content": "[28] Lingjie Liu, Marc Habermann, Viktor Rudnev, Kripasindhu Sarkar, Jiatao Gu, and Christian Theobalt. Neural actor: Neural free-view synthesis of human actors with pose control. ACM transactions on graphics (TOG), 40(6):1-16, 2021. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 239, + 287, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 239, + 287, + 281 + ], + "spans": [ + { + "bbox": [ + 48, + 239, + 287, + 281 + ], + "type": "text", + "content": "[29] Matthew Loper, Naureen Mahmood, Javier Romero, Gerard Pons-Moll, and Michael J Black. Smpl: A skinned multiperson linear model. ACM Transactions on Graphics, 34(6), 2015. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 283, + 287, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 283, + 287, + 326 + ], + "spans": [ + { + "bbox": [ + 48, + 283, + 287, + 326 + ], + "type": "text", + "content": "[30] William E Lorensen and Harvey E Cline. Marching cubes: A high resolution 3d surface construction algorithm. In Semin al graphics: pioneering efforts that shaped the field, pages 347-353. 1998. 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 327, + 287, + 382 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 327, + 287, + 382 + ], + "spans": [ + { + "bbox": [ + 48, + 327, + 287, + 382 + ], + "type": "text", + "content": "[31] Qianli Ma, Jinlong Yang, Anurag Ranjan, Sergi Pujades, Gerard Pons-Moll, Siyu Tang, and Michael J Black. Learning to dress 3d people in generative clothing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6469-6478, 2020. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 383, + 287, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 383, + 287, + 426 + ], + "spans": [ + { + "bbox": [ + 48, + 383, + 287, + 426 + ], + "type": "text", + "content": "[32] Lars Mescheder, Andreas Geiger, and Sebastian Nowozin. Which training methods for gans do actually converge? In International conference on machine learning, pages 3481-3490. PMLR, 2018. 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 427, + 287, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 427, + 287, + 480 + ], + "spans": [ + { + "bbox": [ + 48, + 427, + 287, + 480 + ], + "type": "text", + "content": "[33] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 482, + 287, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 482, + 287, + 536 + ], + "spans": [ + { + "bbox": [ + 48, + 482, + 287, + 536 + ], + "type": "text", + "content": "[34] Thu Nguyen-Phuoc, Chuan Li, Lucas Theis, Christian Richardt, and Yong-Liang Yang. Hologan: Unsupervised learning of 3d representations from natural images. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7588-7597, 2019. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 537, + 287, + 591 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 537, + 287, + 591 + ], + "spans": [ + { + "bbox": [ + 48, + 537, + 287, + 591 + ], + "type": "text", + "content": "[35] Thu H Nguyen-Phuoc, Christian Richardt, Long Mai, Yongliang Yang, and Niloy Mitra. Blockgan: Learning 3d object-aware scene representations from unlabelled images. Advances in neural information processing systems, 33:6767-6778, 2020. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 592, + 287, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 592, + 287, + 645 + ], + "spans": [ + { + "bbox": [ + 48, + 592, + 287, + 645 + ], + "type": "text", + "content": "[36] Michael Niemeyer and Andreas Geiger. Giraffe: Representing scenes as compositional generative neural feature fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11453-11464, 2021. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 647, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 647, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 647, + 287, + 712 + ], + "type": "text", + "content": "[37] Roy Or-El, Xuan Luo, Mengyi Shan, Eli Shechtman, Jeong Joon Park, and Ira Kemelmacher-Shlizerman. Stylesdf: High-resolution 3d-consistent image and geometry generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13503–13513, 2022. 2, 3" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 546, + 713 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 128 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 128 + ], + "type": "text", + "content": "[38] Ahmed AA Osman, Timo Bolkart, and Michael J Black. Star: Sparse trained articulated human body regressor. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part VI 16, pages 598-613. Springer, 2020. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 307, + 129, + 545, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 129, + 545, + 182 + ], + "spans": [ + { + "bbox": [ + 307, + 129, + 545, + 182 + ], + "type": "text", + "content": "[39] Pablo Palafox, Aljaž Božić, Justus Thies, Matthias Nießner, and Angela Dai. Npms: Neural parametric models for 3d deformable shapes. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12695-12705, 2021. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 308, + 183, + 546, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 183, + 546, + 249 + ], + "spans": [ + { + "bbox": [ + 308, + 183, + 546, + 249 + ], + "type": "text", + "content": "[40] Georgios Pavlakos, Vasileios Choutas, Nima Ghorbani, Timo Bolkart, Ahmed AA Osman, Dimitrios Tzionas, and Michael J Black. Expressive body capture: 3d hands, face, and body from a single image. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10975-10985, 2019. 4, 7" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 250, + 546, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 250, + 546, + 316 + ], + "spans": [ + { + "bbox": [ + 308, + 250, + 546, + 316 + ], + "type": "text", + "content": "[41] Sida Peng, Yuanqing Zhang, Yinghao Xu, Qianqian Wang, Qing Shuai, Hujun Bao, and Xiaowei Zhou. Neural body: Implicit neural representations with structured latent codes for novel view synthesis of dynamic humans. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9054-9063, 2021. 1, 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 316, + 545, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 316, + 545, + 348 + ], + "spans": [ + { + "bbox": [ + 308, + 316, + 545, + 348 + ], + "type": "text", + "content": "[42] Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. arXiv preprint arXiv:2209.14988, 2022. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 350, + 545, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 350, + 545, + 392 + ], + "spans": [ + { + "bbox": [ + 308, + 350, + 545, + 392 + ], + "type": "text", + "content": "[43] Daniel Roich, Ron Mokady, Amit H Bermano, and Daniel Cohen-Or. Pivotal tuning for latent-based editing of real images. ACM Transactions on graphics (TOG), 42(1):1-13, 2022. 5" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 393, + 546, + 448 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 393, + 546, + 448 + ], + "spans": [ + { + "bbox": [ + 308, + 393, + 546, + 448 + ], + "type": "text", + "content": "[44] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 4" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 449, + 545, + 492 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 449, + 545, + 492 + ], + "spans": [ + { + "bbox": [ + 308, + 449, + 545, + 492 + ], + "type": "text", + "content": "[45] Tim Salimans, Ian Goodfellow, Wojciech Zaremba, Vicki Cheung, Alec Radford, and Xi Chen. Improved techniques for training gans. Advances in neural information processing systems, 29, 2016. 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 493, + 545, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 493, + 545, + 536 + ], + "spans": [ + { + "bbox": [ + 308, + 493, + 545, + 536 + ], + "type": "text", + "content": "[46] Katja Schwarz, Yiyi Liao, Michael Niemeyer, and Andreas Geiger. Graf: Generative radiance fields for 3d-aware image synthesis. Advances in Neural Information Processing Systems, 33:20154-20166, 2020. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 537, + 545, + 569 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 537, + 545, + 569 + ], + "spans": [ + { + "bbox": [ + 308, + 537, + 545, + 569 + ], + "type": "text", + "content": "[47] Attila Szabó, Givi Meishvili, and Paolo Favaro. Unsupervised generative 3d shape learning from natural images. arXiv preprint arXiv:1910.00287, 2019. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 571, + 545, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 571, + 545, + 635 + ], + "spans": [ + { + "bbox": [ + 308, + 571, + 545, + 635 + ], + "type": "text", + "content": "[48] Chung-Yi Weng, Brian Curless, Pratul P Srinivasan, Jonathan T Barron, and Ira Kemelmacher-Shlizerman. Humanerf: Free-viewpoint rendering of moving people from monocular video. In Proceedings of the IEEE/CVF conference on computer vision and pattern Recognition, pages 16210-16220, 2022. 1, 2, 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 636, + 545, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 636, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 308, + 636, + 545, + 689 + ], + "type": "text", + "content": "[49] Jiajun Wu, Chengkai Zhang, Tianfan Xue, Bill Freeman, and Josh Tenenbaum. Learning a probabilistic latent space of object shapes via 3d generative-adversarial modeling. Advances in neural information processing systems, 29, 2016. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 691, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 691, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 691, + 545, + 713 + ], + "type": "text", + "content": "[50] Yuliang Xiu, Jinlong Yang, Dimitrios Tzionas, and Michael J Black. Icon: Implicit clothed humans obtained from nor" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "9990" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 362 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 66, + 72, + 286, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 72, + 286, + 105 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 286, + 105 + ], + "type": "text", + "content": "mals. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13286-13296. IEEE, 2022. 5" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 107, + 287, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 107, + 287, + 162 + ], + "spans": [ + { + "bbox": [ + 48, + 107, + 287, + 162 + ], + "type": "text", + "content": "[51] Yang Xue, Yuheng Li, Krishna Kumar Singh, and Yong Jae Lee. Giraffe hd: A high-resolution 3d-aware generative model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18440-18449, 2022. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 163, + 287, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 163, + 287, + 217 + ], + "spans": [ + { + "bbox": [ + 49, + 163, + 287, + 217 + ], + "type": "text", + "content": "[52] Zhuoqian Yang, Shikai Li, Wayne Wu, and Bo Dai. 3dhumanan: 3d-aware human image generation with 3d pose mapping. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 23008-23019, 2023. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 219, + 287, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 219, + 287, + 274 + ], + "spans": [ + { + "bbox": [ + 49, + 219, + 287, + 274 + ], + "type": "text", + "content": "[53] Jianfeng Zhang, Zihang Jiang, Dingdong Yang, Hongyi Xu, Yichun Shi, Guoxian Song, Zhongcong Xu, Xinchao Wang, and Jiashi Feng. Avatargen: a 3d generative model for animatable human avatars. In European Conference on Computer Vision, pages 668-685. Springer, 2022. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 49, + 275, + 287, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 275, + 287, + 319 + ], + "spans": [ + { + "bbox": [ + 49, + 275, + 287, + 319 + ], + "type": "text", + "content": "[54] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3836-3847, 2023. 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 49, + 320, + 287, + 362 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 320, + 287, + 362 + ], + "spans": [ + { + "bbox": [ + 49, + 320, + 287, + 362 + ], + "type": "text", + "content": "[55] Peng Zhou, Lingxi Xie, Bingbing Ni, and Qi Tian. Cips-3d: A 3d-aware generator of gans based on conditionally-independent pixel synthesis. arXiv preprint arXiv:2110.09788, 2021. 3" + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 314, + 757 + ], + "type": "text", + "content": "9991" + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/End-to-End Spatio-Temporal Action Localisation with Video Transformers/62e0dda5-0752-47db-aad5-cb1b83945b43_content_list.json b/2024/End-to-End Spatio-Temporal Action Localisation with Video Transformers/62e0dda5-0752-47db-aad5-cb1b83945b43_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..2fd91d9d253887ffc466281904528eb8f8d69831 --- /dev/null +++ b/2024/End-to-End Spatio-Temporal Action Localisation with Video Transformers/62e0dda5-0752-47db-aad5-cb1b83945b43_content_list.json @@ -0,0 +1,1593 @@ +[ + { + "type": "text", + "text": "End-to-End Spatio-Temporal Action Localisation with Video Transformers", + "text_level": 1, + "bbox": [ + 107, + 130, + 864, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Alexey A. Gritsenko Xuehan Xiong Josip Djolonga Mostafa Dehghani Chen Sun Mario Lučić Cordelia Schmid Anurag Arnab Google", + "bbox": [ + 176, + 180, + 794, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{agritsenko, aarnab}@google.com", + "bbox": [ + 339, + 234, + 617, + 252 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 272, + 313, + 287 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The most performant spatio-temporal action localisation models use external person proposals and complex external memory banks. We propose a fully end-to-end, purely-transformer based model that directly ingests an input video, and outputs tubelets - a sequence of bounding boxes and the action classes at each frame. Our flexible model can be trained with either sparse bounding-box supervision on individual frames, or full tubelet annotations. And in both cases, it predicts coherent tubelets as the output. Moreover, our end-to-end model requires no additional pre-processing in the form of proposals, or post-processing in terms of non-maximal suppression. We perform extensive ablation experiments, and significantly advance the state-of-the-art on five different spatio-temporal action localisation benchmarks with both sparse keyframes and full tubelet annotations.", + "bbox": [ + 75, + 304, + 473, + 545 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 573, + 209, + 588 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Spatio-temporal action localisation is an important problem with applications in advanced video search engines, robotics and security among others. It is typically formulated in one of two ways: Firstly, predicting the bounding boxes and actions performed by an actor at a single keyframe given neighbouring frames as spatio-temporal context [18, 28]. Or alternatively, predicting a sequence of bounding boxes and actions (i.e. \"tubes\"), for each actor at each frame in the video [21, 49].", + "bbox": [ + 75, + 598, + 468, + 733 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The most performant models [3, 15, 40, 62], particularly for the first, keyframe-based formulation of the problem, employ a two-stage pipeline inspired by the Fast-RCNN object detector [17]: They first run a separate person detector to obtain proposals. Features from these proposals are then aggregated and classified according to the actions of interest. These models have also been supplemented with memory banks containing long-term contextual information from other frames [40, 53, 61, 62], and/or detections of other potentially relevant objects [2, 53] to capture additional scene context, achieving state-of-the-art results.", + "bbox": [ + 75, + 734, + 470, + 901 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/248495399aef562cd7b0c1d860b436c8af9430ed0e3230775834ee9027aa8a88.jpg", + "image_caption": [ + "Figure 1. We propose an end-to-end Spatio-Temporal Action Recognition model named STAR. Our model is end-to-end in that it does not require any external region proposals to predict tubelets – sequences of bounding boxes associated with a given person in every frame and their corresponding action classes. Our model can be trained with either sparse box annotations on selected keyframes, or full tubelet supervision." + ], + "image_footnote": [], + "bbox": [ + 522, + 266, + 869, + 510 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "And whilst proposal-free algorithms, which do not require external person detectors, have been developed for detecting both at the keyframe-level [9, 25, 51] and tubelet-level [22, 66], their performance has typically lagged behind their proposal-based counterparts. Here, we show for the first time that an end-to-end trainable spatio-temporal model outperforms a two-stage approach.", + "bbox": [ + 496, + 625, + 893, + 731 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "As shown in Fig. 1, we propose our Spatio-Temporal Action TransformeR (STAR) that consists of a transformer architecture, and is based on the DETR [6] detection model. Our model is \"end-to-end\" in that it does not require preprocessing in the form of proposals, nor post-processing in the form of non-maximal suppression (NMS) in contrast to the majority of prior work. The initial stage of the model is a vision encoder. This is followed by a decoder that processes learned latent queries, which represent each actor in the video, into output tubelets - a sequence of bounding boxes and action classes at each time step of", + "bbox": [ + 496, + 734, + 895, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "18373", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "the input video clip. Our model is versatile in that we can train it with either fully-labeled tube annotations, or with sparse keyframe annotations (when only a limited number of keyframes are labelled). In the latter case, our network still predicts tubelets, and learns to associate detections of an actor, from one frame to the next, without explicit supervision. This behaviour is facilitated by our formulation of factorised queries, decoder architecture and tubelet matching in the loss which all contain temporal inductive biases.", + "bbox": [ + 75, + 90, + 468, + 226 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We conduct thorough ablation studies of these modelling choices, confirming the benefit of temporal inductive biases in our model design. Informed by these experiments, we achieve state-of-the-art on both keyframe-based action localisation datasets like AVA [18] and AVA-Kinetics [28], and also tubelet-based datasets like UCF101-24 [49], JHMDB [21] and MultiSports [32]. In particular, we achieve a Frame mAP of 45.1 on AVA-Kinetics, outperforming the best previous results achieved by a massive video foundation model [57]. Moreover, our state-of-the-art results are achieved with a single forward-pass through the model, using only a video clip as input, and without any separate external person detectors providing proposals [57, 60, 62], complex memory banks [40, 62, 66], or additional object detectors [2, 53], as used by the prior works. Furthermore, we outperform these complex, prior, state-of-the-art two-stage models whilst also having additional functionality in that our model predicts tubelets, that is, temporally consistent bounding boxes at each frame of the input video clip. This capability is demonstrated by our results on tube-based datasets like UCF101-24 where we surpass the prior work [66] by 13.2 points on Video AP50.", + "bbox": [ + 75, + 227, + 470, + 559 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 76, + 574, + 218, + 589 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Models for spatio-temporal action localisation have typically built upon advances in object detectors for images. The most performant methods for action localisation [3, 15, 40, 53, 62] are based on \"two-stage\" detectors like Fast-RCNN [17]. These models use external, pre-computed person detections, and use them to ROI-pool features which are then classified into action classes. Although these models are cumbersome in that they require an additional model and backbone to first detect people, and therefore additional detection training data as well, they are currently the leading approaches on datasets such as AVA [18]. Such models using external proposals are also particularly suited to datasets such as AVA [18] as each person is exhaustively labelled as performing an action, and therefore there are fewer false-positives from using action-agnostic person detections compared to datasets such as UCF101 [49].", + "bbox": [ + 75, + 597, + 467, + 838 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The performance of these two-stage models has further been improved by incorporating more contextual information using feature banks extracted from additional frames in the video [40, 53, 61, 62] or by utilising detections of ad", + "bbox": [ + 75, + 839, + 468, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "ditional objects in the scene [2, 5, 58, 64]. Both of these cases entail significant extra computation and complexity to train additional auxiliary models, and to precompute features from them that are then used during training and inference of the localisation model.", + "bbox": [ + 496, + 90, + 890, + 165 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our proposed method, in contrast, is end-to-end in that it directly produces detections without any additional inputs besides a video clip. Moreover, it outperforms these prior works without resorting to external proposals or memory banks, showing that a transformer backbone is sufficient to capture long-range dependencies in the input video. In addition, unlike previous two-stage methods, our method directly predicts tubelets: a sequence of bounding boxes and actions for each frame of the input video, and can do so even when we do not have full tubelet annotations available.", + "bbox": [ + 496, + 167, + 890, + 316 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "A number of proposal-free action localisation models have also been developed [9, 16, 22, 25, 51, 66]. These methods are based upon alternative object detection architectures such as SSD [35], CentreNet [67], YOLO [43], DETR [6] and Sparse-RCNN [52]. However, in contrast to our approach, they have been outperformed by their proposal-based counterparts. Moreover, some of these methods [16, 25, 51] also consist of separate network backbones for learning video feature representations and proposals for a keyframe, and are thus effectively two networks trained jointly, and cannot predict tubelets either.", + "bbox": [ + 496, + 318, + 890, + 484 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Among prior works that do not use external proposals, and also directly predict tubelets [22, 29, 31, 47, 48], our work is most similar to TubeR [66] given that our model is also based on DETR. Our model, however, is designed with additional temporal inductive biases which improves accuracy (without using external memory banks precomputed offline like [66]. And moreover, unlike TubeR, we also demonstrate how our model can predict tubelets (i.e. predictions at every frame of the input video), even when we only have sparse keyframe supervision (i.e. ground truth annotation for a limited number of frames) available.", + "bbox": [ + 496, + 484, + 890, + 650 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Finally, we note that DETR has also been extended as a proposal-free method to addressing other localisation tasks in video such as instance segmentation [59], temporal localisation [36, 39, 63] and moment retrieval [27].", + "bbox": [ + 496, + 651, + 890, + 712 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3. Spatio-Temporal Action Transformer", + "text_level": 1, + "bbox": [ + 498, + 724, + 836, + 742 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our proposed model ingests a sequence of video frames, and directly predicts tubelets (a sequence of bounding boxes and action labels). No external person detections [40, 54, 57], or memory banks [62, 66], are needed.", + "bbox": [ + 496, + 750, + 890, + 808 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "As summarised in Fig. 2, our model consists of a vision encoder (Sec. 3.1), followed by a decoder which processes learned query tokens into output tubelets (Sec. 3.2). We incorporate temporal inductive biases into our decoder to improve accuracy and tubelet prediction with weaker supervision. Our model is inspired by the DETR architecture [6]", + "bbox": [ + 496, + 810, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "18374", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/f91408e60bee52cb17a7ef82c064054ea15cf6fb90c4c239741b9f8ca8c0ebfd.jpg", + "image_caption": [ + "Figure 2. Our model processes a fixed-length video clip, and for each frame, outputs tubelets (i.e. linked bounding boxes with associated action class probabilities). It consists of vision encoder which outputs a video representation, $\\mathbf{x} \\in \\mathbb{R}^{T \\times h \\times w \\times d}$ . The video representation, along with learned queries, $\\mathbf{q}$ (which are factorised into spatial $\\mathbf{q}^s$ and temporal components $\\mathbf{q}^t$ ) are decoded into tubelets by a decoder of" + ], + "image_footnote": [], + "bbox": [ + 78, + 73, + 890, + 208 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "$L$ layers followed by shallow box and class prediction heads.", + "bbox": [ + 76, + 252, + 439, + 265 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "for object detection in images, and is also trained with a set-based loss and Hungarian matching. We detail our loss, and how we can train with either sparse keyframe supervision or full tubelet supervision, in Sec. 3.3.", + "bbox": [ + 75, + 277, + 468, + 339 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Vision Encoder", + "text_level": 1, + "bbox": [ + 76, + 351, + 230, + 366 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The vision backbone processes an input video, $\\mathbf{X} \\in \\mathbb{R}^{T \\times H \\times W \\times 3}$ to produce a feature representation of the input video $\\mathbf{x} \\in \\mathbb{R}^{t \\times h \\times w \\times d}$ . Here, $T$ , $H$ and $W$ are the original temporal-, height- and width-dimensions of the input video respectively, whilst $t$ , $h$ and $w$ are the spatio-temporal dimensions of their feature representation, and $d$ its latent dimension. When using a transformer backbone, these spatiotemporal dimensions depend on the patch size when tokenising the input, and when using a convolutional backbone, they depend on the overall stride. To retain spatiotemporal information, we remove the spatial- and temporal-aggregation steps at the end of the original backbone. And if the temporal patch size (or stride) is larger than 1, we bilinearly upsample the final feature map along the temporal axis to maintain the original temporal resolution.", + "bbox": [ + 75, + 376, + 472, + 603 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Tubelet Decoder", + "text_level": 1, + "bbox": [ + 76, + 616, + 238, + 631 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our decoder processes the visual features, $\\mathbf{x} \\in \\mathbb{R}^{T \\times h \\times w \\times c}$ , along with learned queries, $\\mathbf{q} \\in \\mathbb{R}^{T \\times S \\times d}$ , to output tubelets, $\\mathbf{y} = (\\mathbf{b}, \\mathbf{a})$ which are a sequence of bounding boxes, $b \\in \\mathbb{R}^{T \\times S \\times 4}$ and corresponding actions, $a \\in \\mathbb{R}^{T \\times S \\times C}$ . Here, $S$ denotes the maximum number of bounding boxes per frame (padded with \"background\" as necessary) and $C$ denotes the number of output classes.", + "bbox": [ + 75, + 638, + 468, + 747 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The idea of decoding learned queries into output detections using the transformer decoder architecture of Vaswani et al. [56] was used in DETR [6]. In summary, the decoder of [6, 56] consists of $L$ layers, each performing a series of self-attention operations on the queries, and cross-attention between the queries and encoder outputs.", + "bbox": [ + 75, + 748, + 468, + 839 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We modify the queries, self-attention and cross-attention operations for our spatio-temporal localisation scenario, as shown in Fig. 2 and 3 to include additional temporal inductive biases, and to improve accuracy as detailed below.", + "bbox": [ + 75, + 839, + 468, + 901 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/ab85d5469b56bb6f27c3dafa2a47896ec4d5bb7af2060b6907190866c9781041.jpg", + "image_caption": [ + "Figure 3. Our decoder layer consists of factorised self-attention (SA) (left) and cross-attention (CA) (right) operations designed to provide a spatio-temporal inductive bias and reduce computation. Both operations restrict attention to the same spatial and temporal slices as the query token, as illustrated by the receptive field (blue) for a given query token (magenta). Factorised SA consists of two operations, whilst in Factorised CA, there is one operation." + ], + "image_footnote": [], + "bbox": [ + 501, + 257, + 890, + 429 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Queries Queries, q, in DETR, are decoded using the encoded visual features, x, into bounding box predictions, and are analogous to the \"anchors\" used in other detection architectures such as Faster-RCNN [44].", + "bbox": [ + 496, + 545, + 890, + 606 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The most straightforward way to define queries is to randomly initialise $\\mathbf{q} \\in \\mathbb{R}^{T \\times S \\times d}$ , where there are $S$ bounding boxes at each of the $T$ input frames in the video clip.", + "bbox": [ + 496, + 608, + 890, + 654 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "However, we find it is more effective to factorise the queries into separate learned spatial, $\\mathbf{q}^s\\in \\mathbb{R}^{S\\times d}$ , and temporal, $\\mathbf{q}^{T\\times d}$ parameters. To obtain the final tubelet queries, we simply repeat the spatial queries across all frames, and add them to their corresponding temporal embedding at each location, as shown in Fig. 2. More concretely $\\mathbf{q}_{ij} = \\mathbf{q}_i^t +\\mathbf{q}_j^s$ where $i$ and $j$ denote the temporal and spatial indices respectively.", + "bbox": [ + 496, + 656, + 890, + 777 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The factorised query representation means that the same spatial embedding is used across all frames. Intuitively, this encourages the $i^{th}$ spatial query embedding, $\\mathbf{q}_i^s$ , to bind to the same location across different frames of the video, and since objects typically have small displacements from frame to frame, may help to associate bounding boxes within a tubelet together. We verify this intuition empirically in the experimental section.", + "bbox": [ + 496, + 780, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "18375", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Decoder layer The decoder layer in the original transformer [56] consists of self-attention on the queries, $\\mathbf{q}$ , followed by cross-attention between the queries and the outputs of the encoder, $\\mathbf{x}$ , and then a multilayer perceptron (MLP) layer [19, 56]:", + "bbox": [ + 76, + 90, + 468, + 167 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {u} ^ {\\ell} = \\operatorname {M H S A} \\left(\\mathbf {q} ^ {\\ell}\\right) + \\mathbf {q} ^ {\\ell}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 191, + 176, + 468, + 194 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {v} ^ {\\ell} = \\operatorname {C A} \\left(\\mathbf {u} ^ {\\ell}, \\mathbf {x}\\right) + \\mathbf {u} ^ {\\ell}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 192, + 196, + 468, + 215 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {z} ^ {\\ell} = \\operatorname {M L P} \\left(\\mathbf {v} ^ {\\ell}\\right) + \\mathbf {v} ^ {\\ell}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 194, + 218, + 468, + 236 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathbf{z}^{\\ell}$ is the output of the $\\ell^{th}$ decoder layer, $\\mathbf{u}$ and $\\mathbf{v}$ are intermediate variables, MHSA denotes multi-headed self-attention and CA cross-attention. Note that the inputs to the MLP, self- and cross-attention operations are layer-normalised [4], which we omit here for clarity.", + "bbox": [ + 76, + 248, + 468, + 324 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In our model, we factorise the self- and cross-attention layers across space and time respectively as shown in Fig. 3, to introduce a temporal locality inductive bias, and also to increase model efficiency. Concretely, when applying MHSA, we first compute the queries, keys and values, over which we attend twice: first independently at each time step with each frame, and then, independently along the time axis at each spatial location. Similarly, we modify the cross-attention operation so that only tubelet queries and backbone features from the same time index attend to each other.", + "bbox": [ + 75, + 325, + 468, + 474 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Localisation and classification heads We obtain the final predictions of the network, $\\mathbf{y} = (\\mathbf{b},\\mathbf{a})$ , by applying a small feed-forward network to the outputs to the decoder, $\\mathbf{z}$ , following DETR [6]. The sequence of bounding boxes, $\\mathbf{b}$ , is obtained with a 3-layer MLP, and is parameterised by the box center, width and height for each frame in the tubelet. A single-layer linear projection is used to obtain class logits, $\\mathbf{a}$ . As we predict a fixed number of $S$ bounding boxes per frame, and $S$ is more than the maximum number of ground truth instances in the frame, we also include an additional class label, $\\varnothing$ , which represents the \"background\" class which tubelets with no action class can be assigned to.", + "bbox": [ + 75, + 479, + 468, + 660 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Training objective", + "text_level": 1, + "bbox": [ + 76, + 667, + 253, + 686 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our model predicts bounding boxes and action classes at each frame of the input video. Many datasets, however, such as AVA [18], are only sparsely annotated at selected keyframes of the video. In order to leverage the available annotations, we compute our training loss, Eq. 4, only at the annotated frames of the video, after having matched the predictions to the ground truth:", + "bbox": [ + 75, + 691, + 468, + 797 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} (\\mathbf {y}, \\hat {\\mathbf {y}}) = \\frac {1}{| \\mathcal {T} |} \\sum_ {t \\in \\mathcal {T}} \\mathcal {L} _ {\\text {f r a m e}} (\\mathbf {y}, \\hat {\\mathbf {y}}), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 163, + 808, + 468, + 844 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathcal{T}$ is the set of labelled frames; $\\mathbf{y}$ and $\\hat{\\mathbf{y}}$ denote the ground truth and predicted tubelets after matching. Following DETR [6], our training loss at each frame, $\\mathcal{L}_{\\mathrm{frame}}$ , is a", + "bbox": [ + 76, + 854, + 468, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "sum of an $L_{1}$ regression loss on bounding boxes, the generalised IoU loss [45] on bounding boxes, and a cross-entropy loss on action labels:", + "bbox": [ + 498, + 90, + 890, + 135 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {\\mathrm {f r a m e}} \\left(\\mathbf {b} ^ {t}, \\hat {\\mathbf {b}} ^ {t}, \\mathbf {a} ^ {t}, \\hat {\\mathbf {a}} ^ {t}\\right) = \\sum_ {i} \\mathcal {L} _ {\\mathrm {b o x}} \\left(\\mathbf {b} _ {i} ^ {t}, \\hat {\\mathbf {b}} _ {i} ^ {t}\\right) + \\mathcal {L} _ {\\mathrm {i o u}} \\left(\\mathbf {b} _ {i} ^ {t}, \\hat {\\mathbf {b}} _ {i} ^ {t}\\right) \\\\ + \\mathcal {L} _ {\\text {c l a s s}} \\left(\\mathbf {a} _ {i} ^ {t}, \\hat {\\mathbf {a}} _ {i} ^ {t}\\right). \\tag {5} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 511, + 143, + 890, + 194 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Matching Set-based detection models such as DETR can make predictions in any order, which is why the predictions need to be matched to the ground truth before computing the training loss.", + "bbox": [ + 498, + 194, + 890, + 253 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The first form of matching that we consider is to independently perform bipartite matching at each frame to align the model's predictions to the ground truth (or the $\\varnothing$ background class) before computing the loss. In this case, we use the Hungarian algorithm [26] to obtain $T$ permutations of $S$ elements, $\\hat{\\pi}^t\\in \\Pi^t$ , at each frame, where the permutation at the $t^{th}$ frame minimises the per-frame loss,", + "bbox": [ + 498, + 253, + 890, + 359 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\pi} ^ {t} = \\underset {\\pi \\in \\Pi^ {t}} {\\arg \\min } \\mathcal {L} _ {\\text {f r a m e}} \\left(\\mathbf {y} ^ {t}, \\hat {\\mathbf {y}} _ {\\pi (i)} ^ {t}\\right). \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 589, + 369, + 890, + 396 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "An alternative is to perform tubelet matching, where all queries with the same spatial index, $\\mathbf{q}^s$ , must match to the same ground truth annotation across all frames of the input video. Here the permutation is obtained over $S$ elements as", + "bbox": [ + 498, + 405, + 890, + 465 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\pi} = \\underset {\\pi \\in \\Pi} {\\arg \\min } \\frac {1}{| \\mathcal {T} |} \\sum_ {t \\in \\mathcal {T}} \\mathcal {L} _ {\\text {f r a m e}} \\left(\\mathbf {y} ^ {t}, \\hat {\\mathbf {y}} _ {\\pi^ {t} (i)} ^ {t}\\right). \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 562, + 474, + 890, + 512 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Intuitively, tubelet matching provides stronger supervision when we have full tubelet annotations available. Note that regardless of the type of matching that we perform, the loss computation and the overall model architecture remains the same. Note that we do not weight terms in Eq. 5, for both matching and loss calculation, for simplicity, and to avoid having additional hyperparameters, as also done by [38].", + "bbox": [ + 498, + 520, + 890, + 627 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.4. Discussion", + "text_level": 1, + "bbox": [ + 500, + 635, + 617, + 650 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As our approach is based on DETR, it does not require external proposals nor non-maximal suppression for post-processing. The idea of using DETR for action localisation has also been explored by TubeR [66] and WOO [9]. There are, however, a number of key differences: WOO does not detect tubelets at all, but only actions at the center keyframe. We also factorise our queries in the spatial and temporal dimensions (Sec. 3.2) to provide inductive biases urging spatio-temporal association. Moreover, we predict action classes separately for each time step in the tubelet, meaning that each of our queries binds to an actor in the video. TubeR, in contrast, parameterises queries such that they are each associated with separate actions (features are average-pooled over the tubelet, and then linearly classified into a single action class). This choice also means that TubeR requires an additional \"action switch\" head to predict when", + "bbox": [ + 496, + 657, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "18376", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "tubelets start and end, which we do not require as different time steps in a tubelet can have different action classes in our model. Furthermore, we show experimentally (Tab. 1) that TubeR's parameterisation obtains lower accuracy. We also consider two types of matching in the loss computation (Sec. 3.3) unlike TubeR, with \"tubelet matching\" designed for predicting more temporally consistent tubelets. And in contrast to TubeR, we experimentally show how our decoder design allows our model to accurately predict tubelets even with weak, keyframe supervision.", + "bbox": [ + 75, + 90, + 472, + 241 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Finally, TubeR requires additional complexity in the form of a \"short-term context module\" [66] and the external memory bank of [61] which is computed offline using a separate model to achieve strong results. As we show experimentally in the next section, we outperform TubeR without any additional modules, meaning that our model does indeed produce tubelets in an end-to-end manner.", + "bbox": [ + 75, + 242, + 472, + 348 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experimental Evaluation", + "text_level": 1, + "bbox": [ + 76, + 358, + 312, + 375 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Experimental set-up", + "text_level": 1, + "bbox": [ + 76, + 380, + 269, + 396 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Datasets We evaluate on four spatio-temporal action localisation benchmarks. AVA and AVA-Kinetics contain sparse annotations at each keyframe, whereas UCF101-24 and JHMDB51-21 contain full tubelet annotations.", + "bbox": [ + 75, + 401, + 468, + 460 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "AVA [18] consists of 430, 15-minute video clips from movies. Keyframes are annotated at every second in the video, with about 210 000 labelled frames in the training set, and 57 000 in the validation set. There are 80 atomic actions labelled for every actor in the clip, of which 60 are used for evaluation [18]. Following standard practice, we report the Frame Average Precision (fAP) at an IoU threshold of 0.5 using the latest v2.2 annotations [18].", + "bbox": [ + 75, + 462, + 468, + 582 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "AVA-Kinetics [28] is a superset of AVA, and adds detection annotations following the AVA protocol, to a subset of Kinetics 700 [7] videos. Only a single keyframe in a 10-second Kinetics clip is labelled. In total, about 140 000 labelled keyframes are added to the training set, and 32 000 to the validation sets of AVA. Once again, we follow standard practice in reporting the Frame AP at a 0.5 IoU threshold.", + "bbox": [ + 75, + 583, + 468, + 688 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "UCF101-24 [49] is a subset of UCF101, and annotates 24 action classes with full spatio-temporal tubes in 3 207 untrimmed videos. Note that actions are not labelled exhaustively as in AVA, and there may be people present in the video who are not performing any labelled action. Following standard practice, we use the corrected annotations of [47]. We report both the Frame AP, which evaluates the predictions at each frame independently, and also the Video AP. The Video AP uses a 3D, spatio-temporal IoU to match predictions to targets. And since UCF101-24 videos are up to 900 frames long (median length of 164 frames), and our network typically processes $T = 32$ frames at a time, we link together tubelet predictions from our network into full-video-tubes using the same causal linking algo", + "bbox": [ + 75, + 688, + 470, + 901 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/dd017bc019ad0f7434ce23d0be46a97a8f2143dd16f36a5694be1ef18f094eaa.jpg", + "table_caption": [ + "Table 1. Comparison of detection architectures on AVA controlling for the same resolution (160p) and training settings. Binding each query to a person, rather than to an action (as done in TubeR [66]) yields solid improvements. We report the mean AP for both ViViT-B and CSN-152 backbones." + ], + "table_footnote": [], + "table_body": "
ViViT-BCSN-152
Query binds to action23.625.7
Ours, query binds to person26.727.8
", + "bbox": [ + 547, + 157, + 836, + 205 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "rithm as [22, 31] for fair comparison.", + "bbox": [ + 498, + 215, + 746, + 229 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "JHMDB51-21 [21] also contains full tube annotations in 928 trimmed videos. However, as the videos are shorter and at most 40 frames, we can process the entire clip with our network, and do not need to perform any linking.", + "bbox": [ + 496, + 229, + 890, + 290 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "MultiSports [32] is a recent dataset with full tube annotations for 66 fine-grained action classes in 2129 untrimmed videos, which are up to 3676 frames long. It is challenging, as it contains fast motion with multiple actors performing different concurrent actions, and there is a large variance in action durations. Evaluation is the same as for UCF101-24.", + "bbox": [ + 496, + 290, + 890, + 381 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Implementation details For our vision encoder backbone, we consider both transformer-based (ViViT Factorised Encoder [1]), and convolutional (CSN [55]) backbones. For ViViT, we use the \"Base\" and \"Large\" model sizes [11, 12], which are typically first pretrained on image datasets like ImageNet-21K [10] and then finetuned on video datasets like Kinetics [23]. We also use CSN-152 pretrained on Instagram65M [37] and then Kinetics following [66]. Our models process $T = 32$ frames unless otherwise specified, with $S = 64$ spatial queries per frame and latent decoder dimensionality of $d = 2048$ . Exhaustive implementation details and training hyperparameters are included in the supplement. We will also release all code and models upon acceptance.", + "bbox": [ + 496, + 383, + 892, + 595 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Ablation studies", + "text_level": 1, + "bbox": [ + 500, + 604, + 661, + 618 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We analyse the design choices in our model by conducting experiments on both AVA (with sparse per-frame supervision) and on UCF101-24 (where we can evaluate the quality of our predicted tubelets). Unless otherwise stated, our backbone is ViViT-Base pretrained on Kinetics 400, and the frame resolution is 160 pixels (160p) on the smaller side.", + "bbox": [ + 496, + 625, + 890, + 715 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Comparison of detection architectures Tab. 1 compares our model, where each query represents a person, and all of their actions (Sec. 3.2) to the approach of TubeR [66] (Sec. 3.4), where there is a separate query for each action being performed. We observe that our parameterisation has a substantial impact, with our method outperforming binding to actions by 3.1 points with a ViViT backbone, and 2.1 points with a CSN backbone on the AVA dataset, therefore motivating the design of our decoder. the supplement shows that this trend is consistent on UCF101-24 and JHMDB too.", + "bbox": [ + 496, + 719, + 890, + 869 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Another architectural baseline that we can compare to is that of a two-stage Fast-RCNN model using external person", + "bbox": [ + 498, + 869, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "18377", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/3a37ca940b5122466a807fe93fadab334d14b00ae27b1ec7a33d8101a1d7d105.jpg", + "table_caption": [ + "Table 2. Comparison of independent and factorised queries on the AVA, UCF101-24 and JHMDB51-21 datasets. Factorised queries are particularly beneficial for predicting tubelets, as shown by the VideoAP on UCF101-24 and JHMDB51-21 which has full tube annotations. Both models use tubelet matching in the loss." + ], + "table_footnote": [], + "table_body": "
QueryAVA fAPUCF101-24JHMDB51-21
fAPvAP20vAP50vAP50:95fAPvAP20vAP50
Independent25.285.686.359.528.985.088.585.2
Factorised26.386.587.463.429.886.989.588.2
", + "bbox": [ + 78, + 148, + 467, + 210 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/342c2361acb77237cbac5932fae57b28f2193969e0c40bc8c18413bb969638eb.jpg", + "table_caption": [ + "Table 3. Comparison of independent and tubelet matching for computing the loss on AVA, UCF101-24 and JHMD51-21. Tubelet matching helps for tube-level evaluation metrics like the Video AP (vAP) on UCF101-24 and JHMDB51-21. Note that tubelet matching is actually still possible on AVA as the annotations are at 1fps with actor identities." + ], + "table_footnote": [], + "table_body": "
QueryAVA fAPUCF101-24JHMDB51-21
fAPvAP20vAP50vAP50:95fAPvAP20vAP50
Per-frame matching26.788.285.763.529.486.088.386.0
Tubelet matching26.386.587.463.429.886.989.588.2
", + "bbox": [ + 80, + 299, + 467, + 356 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "detections from [61], as used by [3, 13, 15, 62]. This baseline using the same ViViT-B backbone achieved a mean AP of 25.2, which is still 1.5 points below our model, emphasising the promise of our end-to-end approach. Note that the proposals of [61] obtain an AP50 of 93.9 for person detection on the AVA validation set. They were obtained by first pretraining a Faster-RCNN [44] detector on COCO keypoints, and then finetuning on the person boxes from the AVA training set, using a resolution of 1333 on the longer side. Our model is end-to-end, and does not require any external proposals generated by a separate model at all.", + "bbox": [ + 75, + 369, + 467, + 535 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Comparison to TubeR The second row of Tab. 1 using a CSN-152 backbone corresponds to our reimplementation of TubeR. By keeping all other training hyperparameters constant, we observe that our query binding provides an improvement of $2.1\\mathrm{mAP}$ points in a fair comparison. Note that we could not use the public TubeR code [65], as it does not reproduce the paper's results: A higher resolution 256p model achieved only $20\\mathrm{mAP}$ when trained with the public code, whilst it is reported to achieve 31.1. Exhaustive details on our attempts to reproduce TubeR with the authors' public code is in the supplement.", + "bbox": [ + 75, + 540, + 467, + 707 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Query parameterisation Tab. 2 compares our independent and factorised query methods (Sec. 3.2) on AVA and UCF101-24. We observe that factorised queries consistently provide improvements on both the Frame AP and the Video AP across both datasets. As hypothesised in Sec. 3.2, we believe that this is due to the inductive bias present in this parameterisation. Note that we can only measure the Video AP on UCF101-24 as it has tubes labelled. We also show in the supplement that these observations are consistent on the JHMDB dataset too.", + "bbox": [ + 75, + 712, + 467, + 863 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Matching for loss calculation As described in Sec. 3.3, when matching the predictions to the ground truth for loss", + "bbox": [ + 76, + 869, + 467, + 900 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/f6787ce5ae778a37e3e5ae3bb282d58ad5ea08650cc3cd9f080acce266dce6db.jpg", + "table_caption": [ + "Table 4. Our model can predict tubelets even when the ground truth annotations are sparse. We show this by subsampling training annotations from the UCF101-24 dataset. Our model sees minimal performance deterioration even when using only $1/24$ or $4\\%$ of the annotated frames." + ], + "table_footnote": [], + "table_body": "
SamplingLabelled framesfAPvAP20vAP50vAP50:95
All frames458 81486.587.463.429.8
Every 1239 23785.287.263.029.3
Every 2420 24384.986.863.228.1
One per video2 28470.277.148.520.4
", + "bbox": [ + 504, + 148, + 888, + 217 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "computation, we can either independently match the outputs at each frame to the ground truths at each frame, or, we can match entire predicted tubelets to the ground truth tubelets. Tab. 3 shows that tubelet matching does indeed improve the quality of the predicted tubelets, as shown by the Video AP on UCF101-24. However, this comes at the cost of the quality of per-frame predictions, (i.e. Frame AP). This suggests that tubelet matching improves the association of bounding boxes predicted at different frames (hence higher Video AP), but may also impair the quality of the bounding boxes predicted at each frame (Frame AP). Note that it is technically possible for us to also perform tubelet matching on AVA, since AVA is annotated at 1fps with actor identities, and our model is input 32 frames at 12.5fps (therefore 2.56 seconds of temporal context) meaning that we have sparse tubelets with 2 or 3 annotated frames.", + "bbox": [ + 496, + 229, + 890, + 469 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "As tubelet matching helps with the overall Video AP, we use it for subsequent experiments on UCF101-24 and JHMDB51-21. For AVA, we use per-frame matching as the standard evaluation metric is the Frame AP, and annotations are sparse at 1fps.", + "bbox": [ + 496, + 470, + 890, + 546 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Weakly-supervised tubelet detection Our model can predict tubelets even when the ground truth annotations are sparse and only labelled at certain frames (such as the AVA dataset). We quantitatively measure this ability of our model on UCF101-24 which has full tube annotations. We do so by subsampling labels from the training set, and evaluating the full tubes on the validation set.", + "bbox": [ + 496, + 549, + 890, + 654 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "As shown in Tab. 4, we still obtain meaningful tube predictions, with a Video AP20 of 77.1, when using only a single frame of annotation from each UCF video clip. When retaining 1 frame of supervision for every 24 labelled frames (which is roughly 1fps and corresponds to the AVA dataset's annotations), we observe minimal deterioration with respect to the fully supervised model (all Video AP metrics are within 0.7 points). Retaining 1 frame of annotation for every 12 consecutive labelled frames also performs similarly to using all frames in the video clip. These results suggest that due to the redundancy in the data (motion between frames is often limited), and the inductive bias of our model, we do not require each frame in the tube to be labelled in order to predict accurate tubelets.", + "bbox": [ + 496, + 656, + 890, + 866 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Decoder design Tabs. 5 and 6 analyse the effect of the decoder depth and the type of attention in the decoder (de", + "bbox": [ + 500, + 869, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "18378", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/2b7a8967a6314c3128b9e78de868f8cd360cf22f2e5843f610ca0bb17a2aa49b.jpg", + "table_caption": [ + "Table 5. Effect of decoder depth on performance on the AVA dataset. Performance saturates at $L = 6$ layers." + ], + "table_footnote": [], + "table_body": "
Layers (L)01369
mAP ↑23.424.626.226.526.7
", + "bbox": [ + 135, + 109, + 408, + 147 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/b81ab6741550da5be0f14da3558662c29ec7bae98d3f70b9341e106037c1b897.jpg", + "table_caption": [ + "Table 6. Effect of the type of attention used in the decoder on AVA. Factorised attention is both more accurate and efficient (almost half of the GFLOPs per decoder layer)." + ], + "table_footnote": [], + "table_body": "
Decoder attentionmAPGFLOPs
Full26.410.5
Factorised26.75.3
", + "bbox": [ + 161, + 191, + 377, + 239 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/902d74a4ca5ba30c8635f75f16f02d9d99c3f0fcbf8721001d1f018f672b09b0.jpg", + "table_caption": [ + "Table 7. Increasing the image resolution on the AVA dataset leads to consistent accuracy improvements, primarily on small objects. APs, APm and API denote the AP at 0.5 IoU threshold on small, medium and large boxes respectively following the COCO protocol [33]. AVA videos have a median aspect ratio of 16:10, and we pad the larger side when the aspect ratio is different." + ], + "table_footnote": [], + "table_body": "
ResolutionmAPAPsAPmAPI
140 × 22425.47.211.227.8
160 × 25626.711.512.528.7
220 × 35228.812.015.130.7
260 × 41629.413.315.831.0
320 × 51230.017.516.032.0
", + "bbox": [ + 150, + 328, + 387, + 410 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/9f4bd6d036873a40974c2dee5b7e2a8183a5860166205d95f12a762c8caffa19.jpg", + "table_caption": [ + "Table 8. Comparison of pretraining for our models with ViViT-B and ViViT-L backbones on AVA using a resolution of $160 \\times 256$ . Larger models benefit more from additional initial pretraining." + ], + "table_footnote": [], + "table_body": "
PretrainSTAR/BSTAR/L
IN21K [10] → K400 [23]26.727.0
IN21K [10] → K700 [7]27.327.6
CLIP [42] → K700 [7]30.336.2
", + "bbox": [ + 135, + 452, + 405, + 510 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "scribed in Sec. 3.2). As seen in Tab. 5, detection accuracy on AVA increases with the number of decoder layers, plateauing at around 6 layers. It is possible to use no decoder layers too: In this case, instead of learning queries $\\mathbf{q}$ (Sec. 3.2), we simply interpret the outputs of the vision encoder (Sec. 3.1), $\\mathbf{x}$ , as our queries and apply the localisation and classification heads directly upon them. Using decoder layers, however, can provide a performance increase of up to $3.3\\mathrm{mAP}$ points ( $14\\%$ relative), emphasising their utility.", + "bbox": [ + 75, + 517, + 467, + 652 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Tab. 6 shows that factorised attention in the decoder is more accurate than standard, \"full\" attention between all queries and visual features. Moreover, it is more efficient too, using almost half of the GFLOPs at each decoder layer.", + "bbox": [ + 75, + 654, + 467, + 714 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Effect of resolution and pretraining Scaling up the image resolution is critical to achieving high performance for object detection in images [20, 46]. However, we are not aware of previous works studying this for video action localisation. Tab. 7 shows that we do indeed observe substantial improvements from higher resolution, improving by up to 4.6 points on AVA. As expected, higher resolutions help more for detection at small sizes, where we follow the COCO [33] convention of object sizes. Note that AVA videos have a median aspect ratio of 16:10, and we pad the larger side for videos with different aspect ratios.", + "bbox": [ + 75, + 719, + 467, + 883 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Similarly, Tab. 8 shows the effect of different pretraining", + "bbox": [ + 96, + 885, + 467, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "datasets. Video vision transformers are typically pretrained on an image dataset (like ImageNet-21K [10]), before being finetuned on a video dataset, such as Kinetics [23]. We find that the initial image checkpoint plays an important role, with CLIP [42] pretraining significantly outperforming supervised pretraining on ImageNet-21K [12, 50]. This improvement grows further when using a \"Large\" backbone. Our observations are consistent with prior works which have shown that CLIP-pretraining outperforms ImageNet-21K and even JFT pretraining [24, 30, 34, 41].", + "bbox": [ + 496, + 90, + 890, + 242 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Qualitative examples We include example result videos of our proposed model in the supplementary.", + "bbox": [ + 496, + 248, + 890, + 277 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Comparison to state-of-the-art", + "text_level": 1, + "bbox": [ + 498, + 290, + 769, + 306 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We compare our model to the state-of-the-art on datasets with both sparsely annotated keyframes (AVA and AVA-K), and full tubes (UCF101-24, JHMDB and MultiSports).", + "bbox": [ + 496, + 313, + 890, + 358 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "AVA and AVA-Kinetics Tab. 9 compares to prior work on AVA and AVA-Kinetics. The best previous methods relied on external proposals [3, 54, 60] and external memory banks [40, 62] which we outperform. There are fewer end-to-end approaches, and we outperform these by an even larger margin. Note that though TubeR [66] is a proposal-free approach, their best results are actually obtained with the external memory of [61]. Consequently, we have reported the end-to-end, and external-memory versions of TubeR (\"TubeR + LTC\") separately in Tab. 9. Furthermore, as detailed in the supplement, the public TubeR training code produces significantly lower performance (20.0 mAP). Hence, to compare to it, we report results with our reimplementation in addition to the paper's results. Observe that we outperform TubeR using the same CSN-152 backbone, and then improve further using larger transformer backbones.", + "bbox": [ + 496, + 364, + 890, + 606 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We achieve greater relative improvements on AVA-Kinetics, showing that our end-to-end approach can leverage larger datasets more effectively. To our knowledge, we surpass the best previous results on AVA-Kinetics, achieving a Frame AP of 45.1. Notably, we outperform InternVideo [60] and VideoMAE-v2 [57], which are two recent video foundation models using more powerful backbones and larger, private, web-scale video datasets. Note that InternVideo consists of two different encoders, one of which is also initialised from CLIP. And like [60], we achieve our best AVA results by training a model on AVA-Kinetics, and then evaluating it only on the AVA validation set. Finally, we do not use any test-time augmentation, unlike previous works that ensemble results over multiple resolutions and/or left/right flips as shown by the \"Views\" column.", + "bbox": [ + 496, + 607, + 890, + 833 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "UCF101-24 Tab. 10 shows that we outperform prior work on UCF101-24, both in terms of frame-level (Frame AP), and tube-level metrics (Video AP). We achieve state-of-the-art results with a CSN-152 backbone, and improve further", + "bbox": [ + 496, + 839, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "18379", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/9083e1d837640d73c685bf74f36d267bf0ca69ad9c12123ece5142adab4bc9fc.jpg", + "table_caption": [ + "Table 9. Comparison to the state-of-the-art (reported with mean Average Precision; mAP ↑) on AVA v2.2 [18] and AVA-Kinetics (AVA-K) [28]. Methods using external proposals (i.e. not end-to-end) are also trained on additional object detection and human pose data. Unless otherwise stated, separate models are trained for AVA and AVA-Kinetics. * denotes the model was trained on AVA-Kinetics and evaluated on AVA. “Res.” denotes the frame resolution of the shorter side. Web-scale foundational models are denoted in grey. † We also report our reimplemented results for TubeR, as the authors’ public code does not reproduce their reported results, as detailed in the supplement." + ], + "table_footnote": [], + "table_body": "
PretrainingViewsAVAAVA-KRes.BackboneEnd-to-end
MViT-B [13]K400127.3--MViTX
Unified [2]K400627.7-320SlowFastX
AIA [53]K7001832.3-320SlowFastX
ACAR [40]K700633.336.4320SlowFastX
TubeR [66] with LTC [61]IG65M [37]→K400, COCO233.6-256CSN-152X
MeMViT [62]K700-34.4-312MViT v2X
Co-finetuning [3]IN21K→K700, MiT, SSv2132.833.1320ViViT/LX
JFT,WTS→K700, MiT, SSv2136.136.2320ViViT/LX
VideoMAE [54]SSL K700 → Sup. K700.-39.3--ViViT/LX
InternVideo* [60]7 different datasets-41.042.5-Uniformer v2X
VideoMAE v2 [57]6 different datasets-42.643.9-ViViT/gX
Action Transformer [28]K4001-23.0400I3D✓
WOO [9]K600128.3-320SlowFast✓
TubeR [66]IG65M [37]→K400, COCO131.1-256CSN-152✓
TubeR reimplemented†IG65M [37]→K400129.533.6256CSN-152✓
STAR/CSN-152 (ours)IG65M→K400131.435.8256CSN-152✓
STAR/B (ours)IN21K→K400130.036.6320ViViT/B✓
CLIP→K700133.939.1320ViViT/B✓
STAR/L (ours)CLIP→K700139.244.5320ViViT/L✓
STAR/L (ours)*CLIP→K700142.545.1420ViViT/L✓
", + "bbox": [ + 96, + 148, + 870, + 417 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/8d0d423272809171f9c064297de9e2e11bbdc6d2b993f8cee2d973b8488af75b.jpg", + "table_caption": [ + "Table 10. State-of-the-art comparison on datasets with tubelet annotations, UCF101, JHMDB51 and MultiSports." + ], + "table_footnote": [], + "table_body": "
PretrainingUCF101-24JHMDB51-21MultiSportsBackbone
fAPvAP20vAP50vAP50:95fAPvAP20vAP50fAPvAP20vAP50
ACT [22]IN1K67.177.251.425.065.774.273.7---VGG
MOC [31]IN1K→COCO78.082.853.828.370.877.377.225.212.90.6DLA34
Unified [2]K60079.3---------SlowFast
WOO [9]K600----80.5-----SlowFast
TubeR [66]IG65M→K40083.283.358.428.9-87.482.3---CSN-152
HIT [14]K70084.888.874.3-83.889.788.133.327.88.8SlowFast
FBIL [8]K70086.086.969.1----40.830.39.9SlowFast
STAR/CSN-152IG65M→K40086.787.065.430.693.596.395.445.450.118.6CSN-152
STAR/BIN21K→K40087.388.268.631.786.989.588.246.351.625.6ViViT/B
STAR/LCLIP→K70090.389.873.435.892.193.192.659.362.036.2ViViT/L
", + "bbox": [ + 96, + 431, + 874, + 588 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "by scaling up to ViViT-Large, consistent with our results on AVA (Tab. 9). Moreover, note how we substantially outperform TubeR [66] using the same CSN-152 backbone. Our margin of improvement over TubeR grows from vAP20 (+3.7) to vAP50 (+7.0) with the same backbone, showing that our tubelets are more precise, and is in line with our visual observations in the supplement. To our knowledge, we outperform the best previous reported Video AP50 result by 13.2 points. Note that as UCF videos are up to 900 frames, and as our network processes $T = 32$ frames, we follow prior works and link together tubelets using the same causal algorithm as [2, 22, 31, 47] for fair comparison.", + "bbox": [ + 75, + 590, + 470, + 772 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "JHMDB51-21 Tab. 10 shows that we surpass the state-of-the-art on JHMDB. Once again, we significantly outperform TubeR [66], by 13.1 vAP50, with the same CSN-152 backbone. The CSN-152 backbone outperforms ViViT in this case, possibly because this is the smallest dataset and larger backbones can overfit more easily. The videos in this dataset are trimmed (meaning that labelled actions are being performed on each frame), and also shorter. Hence the", + "bbox": [ + 75, + 779, + 470, + 902 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Video AP is not as strict as it is on UCF. As videos are at most 40 frames, we set $T = 40$ in our model so that we process the entire clip at once without needing to link tubelets.", + "bbox": [ + 496, + 590, + 890, + 635 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "MultiSports Finally, Tab. 10 shows that we improve on prior state-of-the-art on MultiSports substantially on both Frame AP and Video AP. Although this is a recent benchmark with few methods to compare against, our improvement is significant due to the challenging nature of the dataset: it contains fast motion, many simultaneous actors and large variation in action duration.", + "bbox": [ + 496, + 636, + 890, + 741 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion and Future Work", + "text_level": 1, + "bbox": [ + 498, + 746, + 769, + 762 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We have presented STAR, an end-to-end spatio-temporal action localisation model that can output tubelets, when either sparse keyframe, or full tubelet annotation is available. Our approach achieves state-of-the-art results on four action localisation datasets for both frame-level and tubelet-level predictions (in particular, we obtain $45.1\\%$ mAP on the challenging AVA-Kinetics dataset), outperforming complex methods that use external proposals and memory banks. Future work is to extend our method to open vocabularies.", + "bbox": [ + 496, + 763, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "18380", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Anurag Arnab, Mostafa Dehghani, Georg Heigold, Chen Sun, Mario Lucic, and Cordelia Schmid. ViViT: A Video Vision Transformer. In ICCV, 2021. 5", + "[2] Anurag Arnab, Chen Sun, and Cordelia Schmid. Unified graph structured models for video understanding. In ICCV, 2021. 1, 2, 8", + "[3] Anurag Arnab, Xuehan Xiong, Alexey Gritsenko, Rob Romijnders, Josip Djolonga, Mostafa Dehghani, Chen Sun, Mario Lucic, and Cordelia Schmid. Beyond transfer learning: Co-finetuning for action localisation. In arXiv preprint arXiv:2207.03807, 2022. 1, 2, 6, 7, 8", + "[4] Jimmy Lei Ba, Jamie Ryan Kiros, and Geoffrey E Hinton. Layer normalization. In arXiv preprint arXiv:1607.06450, 2016. 4", + "[5] Fabien Baradel, Natalia Neverova, Christian Wolf, Julien Mille, and Greg Mori. Object level visual reasoning in videos. In ECCV, 2018. 2", + "[6] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In ECCV, 2020. 1, 2, 3, 4", + "[7] Joao Carreira, Eric Noland, Chloe Hillier, and Andrew Zisserman. A short note on the kinetics-700 human action dataset. In arXiv preprint arXiv:1907.06987, 2019. 5, 7", + "[8] Keke Chen, Xiangbo Shu, Guo-Sen Xie, Rui Yan, and Jinhui Tang. Foreground/background-masked interaction learning for spatio-temporal action detection. In ACM MM, 2023. 8", + "[9] Shoufa Chen, Peize Sun, Enze Xie, Chongjian Ge, Jiannan Wu, Lan Ma, Jiajun Shen, and Ping Luo. Watch only once: An end-to-end video action detection framework. In ICCV, 2021. 1, 2, 4, 8", + "[10] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. ImageNet: A large-scale hierarchical image database. In CVPR, 2009. 5, 7", + "[11] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. BERT: Pre-training of deep bidirectional transformers for language understanding. In NAACL, 2019. 5", + "[12] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2021. 5, 7", + "[13] Haoqi Fan, Bo Xiong, Karttikeya Mangalam, Yanghao Li, Zhicheng Yan, Jitendra Malik, and Christoph Feichtenhofer. Multiscale vision transformers. In ICCV, 2021. 6, 8", + "[14] Gueter Josmy Faure, Min-Hung Chen, and Shang-Hong Lai. Holistic interaction transformer network for action detection. In WACV, 2023. 8", + "[15] Christoph Feichtenhofer, Haoqi Fan, Jitendra Malik, and Kaiming He. Slowfast networks for video recognition. In ICCV, 2019. 1, 2, 6", + "[16] Rohit Girdhar, Joao Carreira, Carl Doersch, and Andrew Zisserman. Video action transformer network. In CVPR, 2019. 2" + ], + "bbox": [ + 78, + 114, + 470, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[17] Ross Girshick. Fast r-cnn. In ICCV, 2015. 1, 2", + "[18] Chunhui Gu, Chen Sun, David A Ross, Carl Vondrick, Caroline Pantofaru, Yeqing Li, Sudheendra Vijayanarasimhan, George Toderici, Susanna Ricco, Rahul Sukthankar, et al. Ava: A video dataset of spatio-temporally localized atomic visual actions. In CVPR, 2018. 1, 2, 4, 5, 8", + "[19] Dan Hendrycks and Kevin Gimpel. Gaussian error linear units (gelus). In arXiv preprint arXiv:1606.08415, 2016. 4", + "[20] Jonathan Huang, Vivek Rathod, Chen Sun, Menglong Zhu, Anoop Korattikara, Alireza Fathi, Ian Fischer, Zbigniew Wojna, Yang Song, Sergio Guadarrama, et al. Speed/accuracy trade-offs for modern convolutional object detectors. In CVPR, 2017. 7", + "[21] H. Jhuang, J. Gall, S. Zuffi, C. Schmid, and M. J. Black. Towards understanding action recognition. In ICCV, 2013. 1, 2, 5", + "[22] Vicky Kalogeiton, Philippe Weinzaepfel, Vittorio Ferrari, and Cordelia Schmid. Action tubelet detector for spatiotemporal action localization. In ICCV, 2017. 1, 2, 5, 8", + "[23] Will Kay, Joao Carreira, Karen Simonyan, Brian Zhang, Chloe Hillier, Sudheendra Vijayanarasimhan, Fabio Viola, Tim Green, Trevor Back, Paul Natsev, et al. The kinetics human action video dataset. In arXiv preprint arXiv:1705.06950, 2017. 5, 7", + "[24] Skanda Koppula, Yazhe Li, Evan Shelhamer, Andrew Jaegle, Nikhil Parthasarathy, Relja Arandjelovic, João Carreira, and Olivier Henaff. Where should i spend my flops? efficiency evaluations of visual pre-training methods. In arXiv preprint arXiv:2209.15589, 2022. 7", + "[25] Okan Köpukli, Xiangyu Wei, and Gerhard Rigoll. You only watch once: A unified cnn architecture for real-time spatiotemporal action localization. In arXiv preprint arXiv:1911.06644, 2019. 1, 2", + "[26] Harold W Kuhn. The hungarian method for the assignment problem. Naval research logistics quarterly, 2(1-2):83-97, 1955. 4", + "[27] Jie Lei, Tamara L Berg, and Mohit Bansal. Detecting moments and highlights in videos via natural language queries. In NeurIPS, 2021. 2", + "[28] Ang Li, Meghanata Thotakuri, David A Ross, João Carreira, Alexander Vostrikov, and Andrew Zisserman. The a-akinetics localized human actions video dataset. In arXiv preprint arXiv:2005.00214, 2020. 1, 2, 5, 8", + "[29] Dong Li, Zhaofan Qiu, Qi Dai, Ting Yao, and Tao Mei. Recurrent tubelet proposal and recognition networks for action detection. In ECCV, 2018. 2", + "[30] Kunchang Li, Yali Wang, Yinan He, Yizhuo Li, Yi Wang, Limin Wang, and Yu Qiao. Uniformerv2: Spatiotemporal learning by arming image vits with video uniformer. In arXiv preprint arXiv:2211.09552, 2022. 7", + "[31] Yixuan Li, Zixu Wang, Limin Wang, and Gangshan Wu. Actions as moving points. In ECCV, 2020. 2, 5, 8", + "[32] Yixuan Li, Lei Chen, Runyu He, Zhenzhi Wang, Gangshan Wu, and Limin Wang. Multisports: A multi-person video dataset of spatio-temporally localized sports actions. In ICCV, 2021. 2, 5" + ], + "bbox": [ + 501, + 92, + 890, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "18381", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[33] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In ECCV, 2014. 7", + "[34] Ziyi Lin, Shijie Geng, Renrui Zhang, Peng Gao, Gerard de Melo, Xiaogang Wang, Jifeng Dai, Yu Qiao, and Hongsheng Li. Frozen clip models are efficient video learners. In ECCV, 2022. 7", + "[35] Wei Liu, Dragomir Anguelov, Dumitru Erhan, Christian Szegedy, Scott Reed, Cheng-Yang Fu, and Alexander C Berg. Ssd: Single shot multibox detector. In ECCV, 2016. 2", + "[36] Xiaolong Liu, Qimeng Wang, Yao Hu, Xu Tang, Shiwei Zhang, Song Bai, and Xiang Bai. End-to-end temporal action detection with transformer. IEEE Transactions on Image Processing, 31:5427-5441, 2022. 2", + "[37] Dhruv Mahajan, Ross Girshick, Vignesh Ramanathan, Kaiming He, Manohar Paluri, Yixuan Li, Ashwin Bharambe, and Laurens Van Der Maaten. Exploring the limits of weakly supervised pretraining. In ECCV, 2018. 5, 8", + "[38] Matthias Minderer, Alexey Gritsenko, Austin Stone, Maxim Neumann, Dirk Weissenborn, Alexey Dosovitskiy, Aravindh Mahendran, Anurag Arnab, Mostafa Dehghani, Zhuoran Shen, et al. Simple Open-Vocabulary Object Detection with Vision Transformers. In ECCV, 2022. 4", + "[39] Megha Nawhal and Greg Mori. Activity graph transformer for temporal action localization. In arXiv preprint arXiv:2101.08540, 2021. 2", + "[40] Junting Pan, Siyu Chen, Mike Zheng Shou, Yu Liu, Jing Shao, and Hongsheng Li. Actor-context-actor relation network for spatio-temporal action localization. In CVPR, 2021. 1, 2, 7, 8", + "[41] Junting Pan, Ziyi Lin, Xiatian Zhu, Jing Shao, and Hongsheng Li. St-adapter: Parameter-efficient image-to-video transfer learning. In NeurIPS, 2022. 7", + "[42] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, 2021. 7", + "[43] Joseph Redmon, Santosh Divvala, Ross Girshick, and Ali Farhadi. You only look once: Unified, real-time object detection. In CVPR, 2016. 2", + "[44] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. In NeurIPS, 2015. 3, 6", + "[45] Hamid Rezatofighi, Nathan Tsoi, JunYoung Gwak, Amir Sadeghian, Ian Reid, and Silvio Savarese. Generalized intersection over union: A metric and a loss for bounding box regression. In CVPR, 2019. 4", + "[46] Bharat Singh and Larry S Davis. An analysis of scale invariance in object detection snip. In CVPR, 2018. 7", + "[47] Gurkirt Singh, Suman Saha, Michael Sapienza, Philip HS Torr, and Fabio Cuzzolin. Online real-time multiple spatiotemporal action localisation and prediction. In ICCV, 2017. 2, 5, 8", + "[48] Lin Song, Shiwei Zhang, Gang Yu, and Hongbin Sun. Tacnet: Transition-aware context network for spatio-temporal action detection. In CVPR, 2019. 2" + ], + "bbox": [ + 78, + 90, + 470, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[49] Khurram Soomro, Amir Roshan Zamir, and Mubarak Shah. Ucf101: A dataset of 101 human actions classes from videos in the wild. In arXiv preprint arXiv:1212.0402, 2012. 1, 2, 5", + "[50] Andreas Steiner, Alexander Kolesnikov, Xiaohua Zhai, Ross Wightman, Jakob Uszkoreit, and Lucas Beyer. How to train your vit? data, augmentation, and regularization in vision transformers. TMLR, 2022. 7", + "[51] Chen Sun, Abhinav Shrivastava, Carl Vondrick, Kevin Murphy, Rahul Sukthankar, and Cordelia Schmid. Actor-centric relation network. In ECCV, 2018. 1, 2", + "[52] Peize Sun, Rufeng Zhang, Yi Jiang, Tao Kong, Chenfeng Xu, Wei Zhan, Masayoshi Tomizuka, Lei Li, Zehuan Yuan, Changhu Wang, et al. Sparse r-cnn: End-to-end object detection with learnable proposals. In CVPR, 2021. 2", + "[53] Jiajun Tang, Jin Xia, Xinzhi Mu, Bo Pang, and Cewu Lu. Asynchronous interaction aggregation for action detection. In ECCV, 2020. 1, 2, 8", + "[54] Zhan Tong, Yibing Song, Jue Wang, and Limin Wang. VideoMAE: Masked autoencoders are data-efficient learners for self-supervised video pre-training. In NeurIPS, 2022. 2, 7, 8", + "[55] Du Tran, Heng Wang, Lorenzo Torresani, and Matt Feiszli. Video classification with channel-separated convolutional networks. In ICCV, 2019. 5", + "[56] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In NeurIPS, 2017. 3, 4", + "[57] Limin Wang, Bingkun Huang, Zhiyu Zhao, Zhan Tong, Yi-nan He, Yi Wang, Yali Wang, and Yu Qiao. Videomae v2: Scaling video masked autoencoders with dual masking. In arXiv preprint arXiv:2303.16727, 2023. 2, 7, 8", + "[58] Xiaolong Wang and Abhinav Gupta. Videos as space-time region graphs. In ECCV, 2018. 2", + "[59] Yuqing Wang, Zhaoliang Xu, Xinlong Wang, Chunhua Shen, Baoshan Cheng, Hao Shen, and Huaxia Xia. End-to-end video instance segmentation with transformers. In CVPR, 2021. 2", + "[60] Yi Wang, Kunchang Li, Yizhuo Li, Yinan He, Bingkun Huang, Zhiyu Zhao, Hongjie Zhang, Jilan Xu, Yi Liu, Zun Wang, et al. Internvideo: General video foundation models via generative and discriminative learning. In arXiv preprint arXiv:2212.03191, 2022. 2, 7, 8", + "[61] Chao-Yuan Wu, Christoph Feichtenhofer, Haoqi Fan, Kaiming He, Philipp Krahenbuhl, and Ross Girshick. Long-term feature banks for detailed video understanding. In CVPR, 2019. 1, 2, 5, 6, 7, 8", + "[62] Chao-Yuan Wu, Yanghao Li, Karttikeya Mangalam, Haoqi Fan, Bo Xiong, Jitendra Malik, and Christoph Feichtenhofer. MeMVit: Memory-augmented multiscale vision transformer for efficient long-term video recognition. In CVPR, 2022. 1, 2, 6, 7, 8", + "[63] Chuhan Zhang, Ankush Gupta, and Andrew Zisserman. Temporal query networks for fine-grained video understanding. In CVPR, 2021. 2", + "[64] Yubo Zhang, Pavel Tokmakov, Martial Hebert, and Cordelia Schmid. A structured model for action detection. In CVPR, 2019. 2" + ], + "bbox": [ + 501, + 92, + 890, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "18382", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[65] Jiaojiao Zhao, Yanyi Zhang, Xinyu Li, Hao Chen, Bing Shuai, Mingze Xu, Chunhui Liu, Kaustav Kundu, Yuanjun Xiong, Davide Modolo, et al. Tuber public code. https://github.com/amazon-science/tubelet-transformer, 2022.6", + "[66] Jiaojiao Zhao, Yanyi Zhang, Xinyu Li, Hao Chen, Bing Shuai, Mingze Xu, Chunhui Liu, Kaustav Kundu, Yuanjun Xiong, Davide Modolo, et al. TubeR: Tubelet Transformer for Video Action Detection. In CVPR, 2022. 1, 2, 4, 5, 7, 8", + "[67] Xingyi Zhou, Dequan Wang, and Philipp Krahenbuhl. Objects as points. In arXiv preprint arXiv:1904.07850, 2019. 2" + ], + "bbox": [ + 78, + 90, + 468, + 258 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "18383", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2024/End-to-End Spatio-Temporal Action Localisation with Video Transformers/62e0dda5-0752-47db-aad5-cb1b83945b43_model.json b/2024/End-to-End Spatio-Temporal Action Localisation with Video Transformers/62e0dda5-0752-47db-aad5-cb1b83945b43_model.json new file mode 100644 index 0000000000000000000000000000000000000000..e5c2a53d60a97da7334fdf16d9dd02a24801f6ad --- /dev/null +++ b/2024/End-to-End Spatio-Temporal Action Localisation with Video Transformers/62e0dda5-0752-47db-aad5-cb1b83945b43_model.json @@ -0,0 +1,2334 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.108, + 0.131, + 0.865, + 0.154 + ], + "angle": 0, + "content": "End-to-End Spatio-Temporal Action Localisation with Video Transformers" + }, + { + "type": "text", + "bbox": [ + 0.177, + 0.181, + 0.795, + 0.234 + ], + "angle": 0, + "content": "Alexey A. Gritsenko Xuehan Xiong Josip Djolonga Mostafa Dehghani Chen Sun Mario Lučić Cordelia Schmid Anurag Arnab Google" + }, + { + "type": "text", + "bbox": [ + 0.34, + 0.236, + 0.618, + 0.253 + ], + "angle": 0, + "content": "{agritsenko, aarnab}@google.com" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.273, + 0.314, + 0.289 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.305, + 0.474, + 0.546 + ], + "angle": 0, + "content": "The most performant spatio-temporal action localisation models use external person proposals and complex external memory banks. We propose a fully end-to-end, purely-transformer based model that directly ingests an input video, and outputs tubelets - a sequence of bounding boxes and the action classes at each frame. Our flexible model can be trained with either sparse bounding-box supervision on individual frames, or full tubelet annotations. And in both cases, it predicts coherent tubelets as the output. Moreover, our end-to-end model requires no additional pre-processing in the form of proposals, or post-processing in terms of non-maximal suppression. We perform extensive ablation experiments, and significantly advance the state-of-the-art on five different spatio-temporal action localisation benchmarks with both sparse keyframes and full tubelet annotations." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.574, + 0.21, + 0.589 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.599, + 0.47, + 0.734 + ], + "angle": 0, + "content": "Spatio-temporal action localisation is an important problem with applications in advanced video search engines, robotics and security among others. It is typically formulated in one of two ways: Firstly, predicting the bounding boxes and actions performed by an actor at a single keyframe given neighbouring frames as spatio-temporal context [18, 28]. Or alternatively, predicting a sequence of bounding boxes and actions (i.e. \"tubes\"), for each actor at each frame in the video [21, 49]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.735, + 0.471, + 0.902 + ], + "angle": 0, + "content": "The most performant models [3, 15, 40, 62], particularly for the first, keyframe-based formulation of the problem, employ a two-stage pipeline inspired by the Fast-RCNN object detector [17]: They first run a separate person detector to obtain proposals. Features from these proposals are then aggregated and classified according to the actions of interest. These models have also been supplemented with memory banks containing long-term contextual information from other frames [40, 53, 61, 62], and/or detections of other potentially relevant objects [2, 53] to capture additional scene context, achieving state-of-the-art results." + }, + { + "type": "image", + "bbox": [ + 0.523, + 0.267, + 0.87, + 0.511 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.515, + 0.895, + 0.614 + ], + "angle": 0, + "content": "Figure 1. We propose an end-to-end Spatio-Temporal Action Recognition model named STAR. Our model is end-to-end in that it does not require any external region proposals to predict tubelets – sequences of bounding boxes associated with a given person in every frame and their corresponding action classes. Our model can be trained with either sparse box annotations on selected keyframes, or full tubelet supervision." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.625, + 0.895, + 0.732 + ], + "angle": 0, + "content": "And whilst proposal-free algorithms, which do not require external person detectors, have been developed for detecting both at the keyframe-level [9, 25, 51] and tubelet-level [22, 66], their performance has typically lagged behind their proposal-based counterparts. Here, we show for the first time that an end-to-end trainable spatio-temporal model outperforms a two-stage approach." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.735, + 0.897, + 0.903 + ], + "angle": 0, + "content": "As shown in Fig. 1, we propose our Spatio-Temporal Action TransformeR (STAR) that consists of a transformer architecture, and is based on the DETR [6] detection model. Our model is \"end-to-end\" in that it does not require preprocessing in the form of proposals, nor post-processing in the form of non-maximal suppression (NMS) in contrast to the majority of prior work. The initial stage of the model is a vision encoder. This is followed by a decoder that processes learned latent queries, which represent each actor in the video, into output tubelets - a sequence of bounding boxes and action classes at each time step of" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "18373" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.227 + ], + "angle": 0, + "content": "the input video clip. Our model is versatile in that we can train it with either fully-labeled tube annotations, or with sparse keyframe annotations (when only a limited number of keyframes are labelled). In the latter case, our network still predicts tubelets, and learns to associate detections of an actor, from one frame to the next, without explicit supervision. This behaviour is facilitated by our formulation of factorised queries, decoder architecture and tubelet matching in the loss which all contain temporal inductive biases." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.228, + 0.471, + 0.56 + ], + "angle": 0, + "content": "We conduct thorough ablation studies of these modelling choices, confirming the benefit of temporal inductive biases in our model design. Informed by these experiments, we achieve state-of-the-art on both keyframe-based action localisation datasets like AVA [18] and AVA-Kinetics [28], and also tubelet-based datasets like UCF101-24 [49], JHMDB [21] and MultiSports [32]. In particular, we achieve a Frame mAP of 45.1 on AVA-Kinetics, outperforming the best previous results achieved by a massive video foundation model [57]. Moreover, our state-of-the-art results are achieved with a single forward-pass through the model, using only a video clip as input, and without any separate external person detectors providing proposals [57, 60, 62], complex memory banks [40, 62, 66], or additional object detectors [2, 53], as used by the prior works. Furthermore, we outperform these complex, prior, state-of-the-art two-stage models whilst also having additional functionality in that our model predicts tubelets, that is, temporally consistent bounding boxes at each frame of the input video clip. This capability is demonstrated by our results on tube-based datasets like UCF101-24 where we surpass the prior work [66] by 13.2 points on Video AP50." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.575, + 0.22, + 0.59 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.598, + 0.468, + 0.839 + ], + "angle": 0, + "content": "Models for spatio-temporal action localisation have typically built upon advances in object detectors for images. The most performant methods for action localisation [3, 15, 40, 53, 62] are based on \"two-stage\" detectors like Fast-RCNN [17]. These models use external, pre-computed person detections, and use them to ROI-pool features which are then classified into action classes. Although these models are cumbersome in that they require an additional model and backbone to first detect people, and therefore additional detection training data as well, they are currently the leading approaches on datasets such as AVA [18]. Such models using external proposals are also particularly suited to datasets such as AVA [18] as each person is exhaustively labelled as performing an action, and therefore there are fewer false-positives from using action-agnostic person detections compared to datasets such as UCF101 [49]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.47, + 0.901 + ], + "angle": 0, + "content": "The performance of these two-stage models has further been improved by incorporating more contextual information using feature banks extracted from additional frames in the video [40, 53, 61, 62] or by utilising detections of ad" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.166 + ], + "angle": 0, + "content": "ditional objects in the scene [2, 5, 58, 64]. Both of these cases entail significant extra computation and complexity to train additional auxiliary models, and to precompute features from them that are then used during training and inference of the localisation model." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.168, + 0.892, + 0.318 + ], + "angle": 0, + "content": "Our proposed method, in contrast, is end-to-end in that it directly produces detections without any additional inputs besides a video clip. Moreover, it outperforms these prior works without resorting to external proposals or memory banks, showing that a transformer backbone is sufficient to capture long-range dependencies in the input video. In addition, unlike previous two-stage methods, our method directly predicts tubelets: a sequence of bounding boxes and actions for each frame of the input video, and can do so even when we do not have full tubelet annotations available." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.319, + 0.892, + 0.485 + ], + "angle": 0, + "content": "A number of proposal-free action localisation models have also been developed [9, 16, 22, 25, 51, 66]. These methods are based upon alternative object detection architectures such as SSD [35], CentreNet [67], YOLO [43], DETR [6] and Sparse-RCNN [52]. However, in contrast to our approach, they have been outperformed by their proposal-based counterparts. Moreover, some of these methods [16, 25, 51] also consist of separate network backbones for learning video feature representations and proposals for a keyframe, and are thus effectively two networks trained jointly, and cannot predict tubelets either." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.486, + 0.892, + 0.651 + ], + "angle": 0, + "content": "Among prior works that do not use external proposals, and also directly predict tubelets [22, 29, 31, 47, 48], our work is most similar to TubeR [66] given that our model is also based on DETR. Our model, however, is designed with additional temporal inductive biases which improves accuracy (without using external memory banks precomputed offline like [66]. And moreover, unlike TubeR, we also demonstrate how our model can predict tubelets (i.e. predictions at every frame of the input video), even when we only have sparse keyframe supervision (i.e. ground truth annotation for a limited number of frames) available." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.652, + 0.892, + 0.713 + ], + "angle": 0, + "content": "Finally, we note that DETR has also been extended as a proposal-free method to addressing other localisation tasks in video such as instance segmentation [59], temporal localisation [36, 39, 63] and moment retrieval [27]." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.726, + 0.838, + 0.743 + ], + "angle": 0, + "content": "3. Spatio-Temporal Action Transformer" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.892, + 0.809 + ], + "angle": 0, + "content": "Our proposed model ingests a sequence of video frames, and directly predicts tubelets (a sequence of bounding boxes and action labels). No external person detections [40, 54, 57], or memory banks [62, 66], are needed." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.811, + 0.892, + 0.901 + ], + "angle": 0, + "content": "As summarised in Fig. 2, our model consists of a vision encoder (Sec. 3.1), followed by a decoder which processes learned query tokens into output tubelets (Sec. 3.2). We incorporate temporal inductive biases into our decoder to improve accuracy and tubelet prediction with weaker supervision. Our model is inspired by the DETR architecture [6]" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "18374" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.08, + 0.074, + 0.892, + 0.209 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.211, + 0.894, + 0.253 + ], + "angle": 0, + "content": "Figure 2. Our model processes a fixed-length video clip, and for each frame, outputs tubelets (i.e. linked bounding boxes with associated action class probabilities). It consists of vision encoder which outputs a video representation, \\(\\mathbf{x} \\in \\mathbb{R}^{T \\times h \\times w \\times d}\\). The video representation, along with learned queries, \\(\\mathbf{q}\\) (which are factorised into spatial \\(\\mathbf{q}^s\\) and temporal components \\(\\mathbf{q}^t\\)) are decoded into tubelets by a decoder of" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.253, + 0.441, + 0.266 + ], + "angle": 0, + "content": "\\(L\\) layers followed by shallow box and class prediction heads." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.278, + 0.47, + 0.34 + ], + "angle": 0, + "content": "for object detection in images, and is also trained with a set-based loss and Hungarian matching. We detail our loss, and how we can train with either sparse keyframe supervision or full tubelet supervision, in Sec. 3.3." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.352, + 0.232, + 0.367 + ], + "angle": 0, + "content": "3.1. Vision Encoder" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.377, + 0.473, + 0.604 + ], + "angle": 0, + "content": "The vision backbone processes an input video, \\(\\mathbf{X} \\in \\mathbb{R}^{T \\times H \\times W \\times 3}\\) to produce a feature representation of the input video \\(\\mathbf{x} \\in \\mathbb{R}^{t \\times h \\times w \\times d}\\). Here, \\(T\\), \\(H\\) and \\(W\\) are the original temporal-, height- and width-dimensions of the input video respectively, whilst \\(t\\), \\(h\\) and \\(w\\) are the spatio-temporal dimensions of their feature representation, and \\(d\\) its latent dimension. When using a transformer backbone, these spatiotemporal dimensions depend on the patch size when tokenising the input, and when using a convolutional backbone, they depend on the overall stride. To retain spatiotemporal information, we remove the spatial- and temporal-aggregation steps at the end of the original backbone. And if the temporal patch size (or stride) is larger than 1, we bilinearly upsample the final feature map along the temporal axis to maintain the original temporal resolution." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.617, + 0.24, + 0.632 + ], + "angle": 0, + "content": "3.2. Tubelet Decoder" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.64, + 0.47, + 0.748 + ], + "angle": 0, + "content": "Our decoder processes the visual features, \\(\\mathbf{x} \\in \\mathbb{R}^{T \\times h \\times w \\times c}\\), along with learned queries, \\(\\mathbf{q} \\in \\mathbb{R}^{T \\times S \\times d}\\), to output tubelets, \\(\\mathbf{y} = (\\mathbf{b}, \\mathbf{a})\\) which are a sequence of bounding boxes, \\(b \\in \\mathbb{R}^{T \\times S \\times 4}\\) and corresponding actions, \\(a \\in \\mathbb{R}^{T \\times S \\times C}\\). Here, \\(S\\) denotes the maximum number of bounding boxes per frame (padded with \"background\" as necessary) and \\(C\\) denotes the number of output classes." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.749, + 0.47, + 0.84 + ], + "angle": 0, + "content": "The idea of decoding learned queries into output detections using the transformer decoder architecture of Vaswani et al. [56] was used in DETR [6]. In summary, the decoder of [6, 56] consists of \\(L\\) layers, each performing a series of self-attention operations on the queries, and cross-attention between the queries and encoder outputs." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.47, + 0.902 + ], + "angle": 0, + "content": "We modify the queries, self-attention and cross-attention operations for our spatio-temporal localisation scenario, as shown in Fig. 2 and 3 to include additional temporal inductive biases, and to improve accuracy as detailed below." + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.258, + 0.892, + 0.43 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.44, + 0.894, + 0.538 + ], + "angle": 0, + "content": "Figure 3. Our decoder layer consists of factorised self-attention (SA) (left) and cross-attention (CA) (right) operations designed to provide a spatio-temporal inductive bias and reduce computation. Both operations restrict attention to the same spatial and temporal slices as the query token, as illustrated by the receptive field (blue) for a given query token (magenta). Factorised SA consists of two operations, whilst in Factorised CA, there is one operation." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.546, + 0.892, + 0.607 + ], + "angle": 0, + "content": "Queries Queries, q, in DETR, are decoded using the encoded visual features, x, into bounding box predictions, and are analogous to the \"anchors\" used in other detection architectures such as Faster-RCNN [44]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.609, + 0.892, + 0.655 + ], + "angle": 0, + "content": "The most straightforward way to define queries is to randomly initialise \\(\\mathbf{q} \\in \\mathbb{R}^{T \\times S \\times d}\\), where there are \\(S\\) bounding boxes at each of the \\(T\\) input frames in the video clip." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.657, + 0.892, + 0.779 + ], + "angle": 0, + "content": "However, we find it is more effective to factorise the queries into separate learned spatial, \\(\\mathbf{q}^s\\in \\mathbb{R}^{S\\times d}\\), and temporal, \\(\\mathbf{q}^{T\\times d}\\) parameters. To obtain the final tubelet queries, we simply repeat the spatial queries across all frames, and add them to their corresponding temporal embedding at each location, as shown in Fig. 2. More concretely \\(\\mathbf{q}_{ij} = \\mathbf{q}_i^t +\\mathbf{q}_j^s\\) where \\(i\\) and \\(j\\) denote the temporal and spatial indices respectively." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.781, + 0.892, + 0.901 + ], + "angle": 0, + "content": "The factorised query representation means that the same spatial embedding is used across all frames. Intuitively, this encourages the \\( i^{th} \\) spatial query embedding, \\( \\mathbf{q}_i^s \\), to bind to the same location across different frames of the video, and since objects typically have small displacements from frame to frame, may help to associate bounding boxes within a tubelet together. We verify this intuition empirically in the experimental section." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18375" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.47, + 0.168 + ], + "angle": 0, + "content": "Decoder layer The decoder layer in the original transformer [56] consists of self-attention on the queries, \\(\\mathbf{q}\\), followed by cross-attention between the queries and the outputs of the encoder, \\(\\mathbf{x}\\), and then a multilayer perceptron (MLP) layer [19, 56]:" + }, + { + "type": "equation", + "bbox": [ + 0.192, + 0.177, + 0.469, + 0.195 + ], + "angle": 0, + "content": "\\[\n\\mathbf {u} ^ {\\ell} = \\operatorname {M H S A} \\left(\\mathbf {q} ^ {\\ell}\\right) + \\mathbf {q} ^ {\\ell}, \\tag {1}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.194, + 0.197, + 0.469, + 0.216 + ], + "angle": 0, + "content": "\\[\n\\mathbf {v} ^ {\\ell} = \\operatorname {C A} \\left(\\mathbf {u} ^ {\\ell}, \\mathbf {x}\\right) + \\mathbf {u} ^ {\\ell}, \\tag {2}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.196, + 0.219, + 0.469, + 0.237 + ], + "angle": 0, + "content": "\\[\n\\mathbf {z} ^ {\\ell} = \\operatorname {M L P} \\left(\\mathbf {v} ^ {\\ell}\\right) + \\mathbf {v} ^ {\\ell}, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.249, + 0.469, + 0.325 + ], + "angle": 0, + "content": "where \\(\\mathbf{z}^{\\ell}\\) is the output of the \\(\\ell^{th}\\) decoder layer, \\(\\mathbf{u}\\) and \\(\\mathbf{v}\\) are intermediate variables, MHSA denotes multi-headed self-attention and CA cross-attention. Note that the inputs to the MLP, self- and cross-attention operations are layer-normalised [4], which we omit here for clarity." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.326, + 0.469, + 0.476 + ], + "angle": 0, + "content": "In our model, we factorise the self- and cross-attention layers across space and time respectively as shown in Fig. 3, to introduce a temporal locality inductive bias, and also to increase model efficiency. Concretely, when applying MHSA, we first compute the queries, keys and values, over which we attend twice: first independently at each time step with each frame, and then, independently along the time axis at each spatial location. Similarly, we modify the cross-attention operation so that only tubelet queries and backbone features from the same time index attend to each other." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.48, + 0.47, + 0.661 + ], + "angle": 0, + "content": "Localisation and classification heads We obtain the final predictions of the network, \\(\\mathbf{y} = (\\mathbf{b},\\mathbf{a})\\), by applying a small feed-forward network to the outputs to the decoder, \\(\\mathbf{z}\\), following DETR [6]. The sequence of bounding boxes, \\(\\mathbf{b}\\), is obtained with a 3-layer MLP, and is parameterised by the box center, width and height for each frame in the tubelet. A single-layer linear projection is used to obtain class logits, \\(\\mathbf{a}\\). As we predict a fixed number of \\(S\\) bounding boxes per frame, and \\(S\\) is more than the maximum number of ground truth instances in the frame, we also include an additional class label, \\(\\varnothing\\), which represents the \"background\" class which tubelets with no action class can be assigned to." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.669, + 0.254, + 0.687 + ], + "angle": 0, + "content": "3.3. Training objective" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.693, + 0.469, + 0.799 + ], + "angle": 0, + "content": "Our model predicts bounding boxes and action classes at each frame of the input video. Many datasets, however, such as AVA [18], are only sparsely annotated at selected keyframes of the video. In order to leverage the available annotations, we compute our training loss, Eq. 4, only at the annotated frames of the video, after having matched the predictions to the ground truth:" + }, + { + "type": "equation", + "bbox": [ + 0.165, + 0.809, + 0.469, + 0.845 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} (\\mathbf {y}, \\hat {\\mathbf {y}}) = \\frac {1}{| \\mathcal {T} |} \\sum_ {t \\in \\mathcal {T}} \\mathcal {L} _ {\\text {f r a m e}} (\\mathbf {y}, \\hat {\\mathbf {y}}), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.855, + 0.469, + 0.902 + ], + "angle": 0, + "content": "where \\(\\mathcal{T}\\) is the set of labelled frames; \\(\\mathbf{y}\\) and \\(\\hat{\\mathbf{y}}\\) denote the ground truth and predicted tubelets after matching. Following DETR [6], our training loss at each frame, \\(\\mathcal{L}_{\\mathrm{frame}}\\), is a" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.892, + 0.136 + ], + "angle": 0, + "content": "sum of an \\( L_{1} \\) regression loss on bounding boxes, the generalised IoU loss [45] on bounding boxes, and a cross-entropy loss on action labels:" + }, + { + "type": "equation", + "bbox": [ + 0.512, + 0.145, + 0.891, + 0.195 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} _ {\\mathrm {f r a m e}} \\left(\\mathbf {b} ^ {t}, \\hat {\\mathbf {b}} ^ {t}, \\mathbf {a} ^ {t}, \\hat {\\mathbf {a}} ^ {t}\\right) = \\sum_ {i} \\mathcal {L} _ {\\mathrm {b o x}} \\left(\\mathbf {b} _ {i} ^ {t}, \\hat {\\mathbf {b}} _ {i} ^ {t}\\right) + \\mathcal {L} _ {\\mathrm {i o u}} \\left(\\mathbf {b} _ {i} ^ {t}, \\hat {\\mathbf {b}} _ {i} ^ {t}\\right) \\\\ + \\mathcal {L} _ {\\text {c l a s s}} \\left(\\mathbf {a} _ {i} ^ {t}, \\hat {\\mathbf {a}} _ {i} ^ {t}\\right). \\tag {5} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.195, + 0.892, + 0.255 + ], + "angle": 0, + "content": "Matching Set-based detection models such as DETR can make predictions in any order, which is why the predictions need to be matched to the ground truth before computing the training loss." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.255, + 0.892, + 0.36 + ], + "angle": 0, + "content": "The first form of matching that we consider is to independently perform bipartite matching at each frame to align the model's predictions to the ground truth (or the \\(\\varnothing\\) background class) before computing the loss. In this case, we use the Hungarian algorithm [26] to obtain \\(T\\) permutations of \\(S\\) elements, \\(\\hat{\\pi}^t\\in \\Pi^t\\), at each frame, where the permutation at the \\(t^{th}\\) frame minimises the per-frame loss," + }, + { + "type": "equation", + "bbox": [ + 0.59, + 0.37, + 0.891, + 0.397 + ], + "angle": 0, + "content": "\\[\n\\hat {\\pi} ^ {t} = \\underset {\\pi \\in \\Pi^ {t}} {\\arg \\min } \\mathcal {L} _ {\\text {f r a m e}} \\left(\\mathbf {y} ^ {t}, \\hat {\\mathbf {y}} _ {\\pi (i)} ^ {t}\\right). \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.406, + 0.892, + 0.467 + ], + "angle": 0, + "content": "An alternative is to perform tubelet matching, where all queries with the same spatial index, \\(\\mathbf{q}^s\\), must match to the same ground truth annotation across all frames of the input video. Here the permutation is obtained over \\(S\\) elements as" + }, + { + "type": "equation", + "bbox": [ + 0.563, + 0.476, + 0.891, + 0.513 + ], + "angle": 0, + "content": "\\[\n\\hat {\\pi} = \\underset {\\pi \\in \\Pi} {\\arg \\min } \\frac {1}{| \\mathcal {T} |} \\sum_ {t \\in \\mathcal {T}} \\mathcal {L} _ {\\text {f r a m e}} \\left(\\mathbf {y} ^ {t}, \\hat {\\mathbf {y}} _ {\\pi^ {t} (i)} ^ {t}\\right). \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.521, + 0.892, + 0.628 + ], + "angle": 0, + "content": "Intuitively, tubelet matching provides stronger supervision when we have full tubelet annotations available. Note that regardless of the type of matching that we perform, the loss computation and the overall model architecture remains the same. Note that we do not weight terms in Eq. 5, for both matching and loss calculation, for simplicity, and to avoid having additional hyperparameters, as also done by [38]." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.636, + 0.618, + 0.651 + ], + "angle": 0, + "content": "3.4. Discussion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.659, + 0.892, + 0.901 + ], + "angle": 0, + "content": "As our approach is based on DETR, it does not require external proposals nor non-maximal suppression for post-processing. The idea of using DETR for action localisation has also been explored by TubeR [66] and WOO [9]. There are, however, a number of key differences: WOO does not detect tubelets at all, but only actions at the center keyframe. We also factorise our queries in the spatial and temporal dimensions (Sec. 3.2) to provide inductive biases urging spatio-temporal association. Moreover, we predict action classes separately for each time step in the tubelet, meaning that each of our queries binds to an actor in the video. TubeR, in contrast, parameterises queries such that they are each associated with separate actions (features are average-pooled over the tubelet, and then linearly classified into a single action class). This choice also means that TubeR requires an additional \"action switch\" head to predict when" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18376" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.242 + ], + "angle": 0, + "content": "tubelets start and end, which we do not require as different time steps in a tubelet can have different action classes in our model. Furthermore, we show experimentally (Tab. 1) that TubeR's parameterisation obtains lower accuracy. We also consider two types of matching in the loss computation (Sec. 3.3) unlike TubeR, with \"tubelet matching\" designed for predicting more temporally consistent tubelets. And in contrast to TubeR, we experimentally show how our decoder design allows our model to accurately predict tubelets even with weak, keyframe supervision." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.243, + 0.473, + 0.349 + ], + "angle": 0, + "content": "Finally, TubeR requires additional complexity in the form of a \"short-term context module\" [66] and the external memory bank of [61] which is computed offline using a separate model to achieve strong results. As we show experimentally in the next section, we outperform TubeR without any additional modules, meaning that our model does indeed produce tubelets in an end-to-end manner." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.359, + 0.313, + 0.376 + ], + "angle": 0, + "content": "4. Experimental Evaluation" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.381, + 0.271, + 0.397 + ], + "angle": 0, + "content": "4.1. Experimental set-up" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.402, + 0.47, + 0.462 + ], + "angle": 0, + "content": "Datasets We evaluate on four spatio-temporal action localisation benchmarks. AVA and AVA-Kinetics contain sparse annotations at each keyframe, whereas UCF101-24 and JHMDB51-21 contain full tubelet annotations." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.463, + 0.47, + 0.583 + ], + "angle": 0, + "content": "AVA [18] consists of 430, 15-minute video clips from movies. Keyframes are annotated at every second in the video, with about 210 000 labelled frames in the training set, and 57 000 in the validation set. There are 80 atomic actions labelled for every actor in the clip, of which 60 are used for evaluation [18]. Following standard practice, we report the Frame Average Precision (fAP) at an IoU threshold of 0.5 using the latest v2.2 annotations [18]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.584, + 0.47, + 0.689 + ], + "angle": 0, + "content": "AVA-Kinetics [28] is a superset of AVA, and adds detection annotations following the AVA protocol, to a subset of Kinetics 700 [7] videos. Only a single keyframe in a 10-second Kinetics clip is labelled. In total, about 140 000 labelled keyframes are added to the training set, and 32 000 to the validation sets of AVA. Once again, we follow standard practice in reporting the Frame AP at a 0.5 IoU threshold." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.689, + 0.471, + 0.902 + ], + "angle": 0, + "content": "UCF101-24 [49] is a subset of UCF101, and annotates 24 action classes with full spatio-temporal tubes in 3 207 untrimmed videos. Note that actions are not labelled exhaustively as in AVA, and there may be people present in the video who are not performing any labelled action. Following standard practice, we use the corrected annotations of [47]. We report both the Frame AP, which evaluates the predictions at each frame independently, and also the Video AP. The Video AP uses a 3D, spatio-temporal IoU to match predictions to targets. And since UCF101-24 videos are up to 900 frames long (median length of 164 frames), and our network typically processes \\( T = 32 \\) frames at a time, we link together tubelet predictions from our network into full-video-tubes using the same causal linking algo" + }, + { + "type": "table_caption", + "bbox": [ + 0.498, + 0.089, + 0.892, + 0.158 + ], + "angle": 0, + "content": "Table 1. Comparison of detection architectures on AVA controlling for the same resolution (160p) and training settings. Binding each query to a person, rather than to an action (as done in TubeR [66]) yields solid improvements. We report the mean AP for both ViViT-B and CSN-152 backbones." + }, + { + "type": "table", + "bbox": [ + 0.548, + 0.158, + 0.838, + 0.206 + ], + "angle": 0, + "content": "
ViViT-BCSN-152
Query binds to action23.625.7
Ours, query binds to person26.727.8
" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.216, + 0.747, + 0.231 + ], + "angle": 0, + "content": "rithm as [22, 31] for fair comparison." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.231, + 0.892, + 0.291 + ], + "angle": 0, + "content": "JHMDB51-21 [21] also contains full tube annotations in 928 trimmed videos. However, as the videos are shorter and at most 40 frames, we can process the entire clip with our network, and do not need to perform any linking." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.291, + 0.892, + 0.382 + ], + "angle": 0, + "content": "MultiSports [32] is a recent dataset with full tube annotations for 66 fine-grained action classes in 2129 untrimmed videos, which are up to 3676 frames long. It is challenging, as it contains fast motion with multiple actors performing different concurrent actions, and there is a large variance in action durations. Evaluation is the same as for UCF101-24." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.385, + 0.893, + 0.597 + ], + "angle": 0, + "content": "Implementation details For our vision encoder backbone, we consider both transformer-based (ViViT Factorised Encoder [1]), and convolutional (CSN [55]) backbones. For ViViT, we use the \"Base\" and \"Large\" model sizes [11, 12], which are typically first pretrained on image datasets like ImageNet-21K [10] and then finetuned on video datasets like Kinetics [23]. We also use CSN-152 pretrained on Instagram65M [37] and then Kinetics following [66]. Our models process \\( T = 32 \\) frames unless otherwise specified, with \\( S = 64 \\) spatial queries per frame and latent decoder dimensionality of \\( d = 2048 \\). Exhaustive implementation details and training hyperparameters are included in the supplement. We will also release all code and models upon acceptance." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.605, + 0.662, + 0.619 + ], + "angle": 0, + "content": "4.2. Ablation studies" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.626, + 0.892, + 0.716 + ], + "angle": 0, + "content": "We analyse the design choices in our model by conducting experiments on both AVA (with sparse per-frame supervision) and on UCF101-24 (where we can evaluate the quality of our predicted tubelets). Unless otherwise stated, our backbone is ViViT-Base pretrained on Kinetics 400, and the frame resolution is 160 pixels (160p) on the smaller side." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.892, + 0.87 + ], + "angle": 0, + "content": "Comparison of detection architectures Tab. 1 compares our model, where each query represents a person, and all of their actions (Sec. 3.2) to the approach of TubeR [66] (Sec. 3.4), where there is a separate query for each action being performed. We observe that our parameterisation has a substantial impact, with our method outperforming binding to actions by 3.1 points with a ViViT backbone, and 2.1 points with a CSN backbone on the AVA dataset, therefore motivating the design of our decoder. the supplement shows that this trend is consistent on UCF101-24 and JHMDB too." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.871, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Another architectural baseline that we can compare to is that of a two-stage Fast-RCNN model using external person" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18377" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.082, + 0.468, + 0.149 + ], + "angle": 0, + "content": "Table 2. Comparison of independent and factorised queries on the AVA, UCF101-24 and JHMDB51-21 datasets. Factorised queries are particularly beneficial for predicting tubelets, as shown by the VideoAP on UCF101-24 and JHMDB51-21 which has full tube annotations. Both models use tubelet matching in the loss." + }, + { + "type": "table", + "bbox": [ + 0.08, + 0.15, + 0.468, + 0.211 + ], + "angle": 0, + "content": "
QueryAVA fAPUCF101-24JHMDB51-21
fAPvAP20vAP50vAP50:95fAPvAP20vAP50
Independent25.285.686.359.528.985.088.585.2
Factorised26.386.587.463.429.886.989.588.2
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.218, + 0.468, + 0.299 + ], + "angle": 0, + "content": "Table 3. Comparison of independent and tubelet matching for computing the loss on AVA, UCF101-24 and JHMD51-21. Tubelet matching helps for tube-level evaluation metrics like the Video AP (vAP) on UCF101-24 and JHMDB51-21. Note that tubelet matching is actually still possible on AVA as the annotations are at 1fps with actor identities." + }, + { + "type": "table", + "bbox": [ + 0.081, + 0.3, + 0.468, + 0.357 + ], + "angle": 0, + "content": "
QueryAVA fAPUCF101-24JHMDB51-21
fAPvAP20vAP50vAP50:95fAPvAP20vAP50
Per-frame matching26.788.285.763.529.486.088.386.0
Tubelet matching26.386.587.463.429.886.989.588.2
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.37, + 0.468, + 0.536 + ], + "angle": 0, + "content": "detections from [61], as used by [3, 13, 15, 62]. This baseline using the same ViViT-B backbone achieved a mean AP of 25.2, which is still 1.5 points below our model, emphasising the promise of our end-to-end approach. Note that the proposals of [61] obtain an AP50 of 93.9 for person detection on the AVA validation set. They were obtained by first pretraining a Faster-RCNN [44] detector on COCO keypoints, and then finetuning on the person boxes from the AVA training set, using a resolution of 1333 on the longer side. Our model is end-to-end, and does not require any external proposals generated by a separate model at all." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.541, + 0.468, + 0.708 + ], + "angle": 0, + "content": "Comparison to TubeR The second row of Tab. 1 using a CSN-152 backbone corresponds to our reimplementation of TubeR. By keeping all other training hyperparameters constant, we observe that our query binding provides an improvement of \\(2.1\\mathrm{mAP}\\) points in a fair comparison. Note that we could not use the public TubeR code [65], as it does not reproduce the paper's results: A higher resolution 256p model achieved only \\(20\\mathrm{mAP}\\) when trained with the public code, whilst it is reported to achieve 31.1. Exhaustive details on our attempts to reproduce TubeR with the authors' public code is in the supplement." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.713, + 0.468, + 0.864 + ], + "angle": 0, + "content": "Query parameterisation Tab. 2 compares our independent and factorised query methods (Sec. 3.2) on AVA and UCF101-24. We observe that factorised queries consistently provide improvements on both the Frame AP and the Video AP across both datasets. As hypothesised in Sec. 3.2, we believe that this is due to the inductive bias present in this parameterisation. Note that we can only measure the Video AP on UCF101-24 as it has tubes labelled. We also show in the supplement that these observations are consistent on the JHMDB dataset too." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.468, + 0.901 + ], + "angle": 0, + "content": "Matching for loss calculation As described in Sec. 3.3, when matching the predictions to the ground truth for loss" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.082, + 0.892, + 0.149 + ], + "angle": 0, + "content": "Table 4. Our model can predict tubelets even when the ground truth annotations are sparse. We show this by subsampling training annotations from the UCF101-24 dataset. Our model sees minimal performance deterioration even when using only \\(1/24\\) or \\(4\\%\\) of the annotated frames." + }, + { + "type": "table", + "bbox": [ + 0.505, + 0.149, + 0.889, + 0.218 + ], + "angle": 0, + "content": "
SamplingLabelled framesfAPvAP20vAP50vAP50:95
All frames458 81486.587.463.429.8
Every 1239 23785.287.263.029.3
Every 2420 24384.986.863.228.1
One per video2 28470.277.148.520.4
" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.23, + 0.892, + 0.47 + ], + "angle": 0, + "content": "computation, we can either independently match the outputs at each frame to the ground truths at each frame, or, we can match entire predicted tubelets to the ground truth tubelets. Tab. 3 shows that tubelet matching does indeed improve the quality of the predicted tubelets, as shown by the Video AP on UCF101-24. However, this comes at the cost of the quality of per-frame predictions, (i.e. Frame AP). This suggests that tubelet matching improves the association of bounding boxes predicted at different frames (hence higher Video AP), but may also impair the quality of the bounding boxes predicted at each frame (Frame AP). Note that it is technically possible for us to also perform tubelet matching on AVA, since AVA is annotated at 1fps with actor identities, and our model is input 32 frames at 12.5fps (therefore 2.56 seconds of temporal context) meaning that we have sparse tubelets with 2 or 3 annotated frames." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.472, + 0.892, + 0.547 + ], + "angle": 0, + "content": "As tubelet matching helps with the overall Video AP, we use it for subsequent experiments on UCF101-24 and JHMDB51-21. For AVA, we use per-frame matching as the standard evaluation metric is the Frame AP, and annotations are sparse at 1fps." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.55, + 0.892, + 0.655 + ], + "angle": 0, + "content": "Weakly-supervised tubelet detection Our model can predict tubelets even when the ground truth annotations are sparse and only labelled at certain frames (such as the AVA dataset). We quantitatively measure this ability of our model on UCF101-24 which has full tube annotations. We do so by subsampling labels from the training set, and evaluating the full tubes on the validation set." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.657, + 0.892, + 0.867 + ], + "angle": 0, + "content": "As shown in Tab. 4, we still obtain meaningful tube predictions, with a Video AP20 of 77.1, when using only a single frame of annotation from each UCF video clip. When retaining 1 frame of supervision for every 24 labelled frames (which is roughly 1fps and corresponds to the AVA dataset's annotations), we observe minimal deterioration with respect to the fully supervised model (all Video AP metrics are within 0.7 points). Retaining 1 frame of annotation for every 12 consecutive labelled frames also performs similarly to using all frames in the video clip. These results suggest that due to the redundancy in the data (motion between frames is often limited), and the inductive bias of our model, we do not require each frame in the tube to be labelled in order to predict accurate tubelets." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Decoder design Tabs. 5 and 6 analyse the effect of the decoder depth and the type of attention in the decoder (de" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.956 + ], + "angle": 0, + "content": "18378" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.086, + 0.468, + 0.111 + ], + "angle": 0, + "content": "Table 5. Effect of decoder depth on performance on the AVA dataset. Performance saturates at \\( L = 6 \\) layers." + }, + { + "type": "table", + "bbox": [ + 0.136, + 0.111, + 0.409, + 0.148 + ], + "angle": 0, + "content": "
Layers (L)01369
mAP ↑23.424.626.226.526.7
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.151, + 0.468, + 0.192 + ], + "angle": 0, + "content": "Table 6. Effect of the type of attention used in the decoder on AVA. Factorised attention is both more accurate and efficient (almost half of the GFLOPs per decoder layer)." + }, + { + "type": "table", + "bbox": [ + 0.162, + 0.192, + 0.379, + 0.24 + ], + "angle": 0, + "content": "
Decoder attentionmAPGFLOPs
Full26.410.5
Factorised26.75.3
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.246, + 0.468, + 0.329 + ], + "angle": 0, + "content": "Table 7. Increasing the image resolution on the AVA dataset leads to consistent accuracy improvements, primarily on small objects. APs, APm and API denote the AP at 0.5 IoU threshold on small, medium and large boxes respectively following the COCO protocol [33]. AVA videos have a median aspect ratio of 16:10, and we pad the larger side when the aspect ratio is different." + }, + { + "type": "table", + "bbox": [ + 0.151, + 0.329, + 0.388, + 0.411 + ], + "angle": 0, + "content": "
ResolutionmAPAPsAPmAPI
140 × 22425.47.211.227.8
160 × 25626.711.512.528.7
220 × 35228.812.015.130.7
260 × 41629.413.315.831.0
320 × 51230.017.516.032.0
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.411, + 0.468, + 0.453 + ], + "angle": 0, + "content": "Table 8. Comparison of pretraining for our models with ViViT-B and ViViT-L backbones on AVA using a resolution of \\(160 \\times 256\\). Larger models benefit more from additional initial pretraining." + }, + { + "type": "table", + "bbox": [ + 0.136, + 0.453, + 0.406, + 0.511 + ], + "angle": 0, + "content": "
PretrainSTAR/BSTAR/L
IN21K [10] → K400 [23]26.727.0
IN21K [10] → K700 [7]27.327.6
CLIP [42] → K700 [7]30.336.2
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.518, + 0.468, + 0.653 + ], + "angle": 0, + "content": "scribed in Sec. 3.2). As seen in Tab. 5, detection accuracy on AVA increases with the number of decoder layers, plateauing at around 6 layers. It is possible to use no decoder layers too: In this case, instead of learning queries \\(\\mathbf{q}\\) (Sec. 3.2), we simply interpret the outputs of the vision encoder (Sec. 3.1), \\(\\mathbf{x}\\), as our queries and apply the localisation and classification heads directly upon them. Using decoder layers, however, can provide a performance increase of up to \\(3.3\\mathrm{mAP}\\) points (\\(14\\%\\) relative), emphasising their utility." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.655, + 0.468, + 0.715 + ], + "angle": 0, + "content": "Tab. 6 shows that factorised attention in the decoder is more accurate than standard, \"full\" attention between all queries and visual features. Moreover, it is more efficient too, using almost half of the GFLOPs at each decoder layer." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.72, + 0.468, + 0.884 + ], + "angle": 0, + "content": "Effect of resolution and pretraining Scaling up the image resolution is critical to achieving high performance for object detection in images [20, 46]. However, we are not aware of previous works studying this for video action localisation. Tab. 7 shows that we do indeed observe substantial improvements from higher resolution, improving by up to 4.6 points on AVA. As expected, higher resolutions help more for detection at small sizes, where we follow the COCO [33] convention of object sizes. Note that AVA videos have a median aspect ratio of 16:10, and we pad the larger side for videos with different aspect ratios." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.886, + 0.468, + 0.901 + ], + "angle": 0, + "content": "Similarly, Tab. 8 shows the effect of different pretraining" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.243 + ], + "angle": 0, + "content": "datasets. Video vision transformers are typically pretrained on an image dataset (like ImageNet-21K [10]), before being finetuned on a video dataset, such as Kinetics [23]. We find that the initial image checkpoint plays an important role, with CLIP [42] pretraining significantly outperforming supervised pretraining on ImageNet-21K [12, 50]. This improvement grows further when using a \"Large\" backbone. Our observations are consistent with prior works which have shown that CLIP-pretraining outperforms ImageNet-21K and even JFT pretraining [24, 30, 34, 41]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.249, + 0.892, + 0.279 + ], + "angle": 0, + "content": "Qualitative examples We include example result videos of our proposed model in the supplementary." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.291, + 0.771, + 0.307 + ], + "angle": 0, + "content": "4.3. Comparison to state-of-the-art" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.314, + 0.892, + 0.359 + ], + "angle": 0, + "content": "We compare our model to the state-of-the-art on datasets with both sparsely annotated keyframes (AVA and AVA-K), and full tubes (UCF101-24, JHMDB and MultiSports)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.366, + 0.892, + 0.607 + ], + "angle": 0, + "content": "AVA and AVA-Kinetics Tab. 9 compares to prior work on AVA and AVA-Kinetics. The best previous methods relied on external proposals [3, 54, 60] and external memory banks [40, 62] which we outperform. There are fewer end-to-end approaches, and we outperform these by an even larger margin. Note that though TubeR [66] is a proposal-free approach, their best results are actually obtained with the external memory of [61]. Consequently, we have reported the end-to-end, and external-memory versions of TubeR (\"TubeR + LTC\") separately in Tab. 9. Furthermore, as detailed in the supplement, the public TubeR training code produces significantly lower performance (20.0 mAP). Hence, to compare to it, we report results with our reimplementation in addition to the paper's results. Observe that we outperform TubeR using the same CSN-152 backbone, and then improve further using larger transformer backbones." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.608, + 0.892, + 0.834 + ], + "angle": 0, + "content": "We achieve greater relative improvements on AVA-Kinetics, showing that our end-to-end approach can leverage larger datasets more effectively. To our knowledge, we surpass the best previous results on AVA-Kinetics, achieving a Frame AP of 45.1. Notably, we outperform InternVideo [60] and VideoMAE-v2 [57], which are two recent video foundation models using more powerful backbones and larger, private, web-scale video datasets. Note that InternVideo consists of two different encoders, one of which is also initialised from CLIP. And like [60], we achieve our best AVA results by training a model on AVA-Kinetics, and then evaluating it only on the AVA validation set. Finally, we do not use any test-time augmentation, unlike previous works that ensemble results over multiple resolutions and/or left/right flips as shown by the \"Views\" column." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.892, + 0.901 + ], + "angle": 0, + "content": "UCF101-24 Tab. 10 shows that we outperform prior work on UCF101-24, both in terms of frame-level (Frame AP), and tube-level metrics (Video AP). We achieve state-of-the-art results with a CSN-152 backbone, and improve further" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18379" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.082, + 0.895, + 0.15 + ], + "angle": 0, + "content": "Table 9. Comparison to the state-of-the-art (reported with mean Average Precision; mAP ↑) on AVA v2.2 [18] and AVA-Kinetics (AVA-K) [28]. Methods using external proposals (i.e. not end-to-end) are also trained on additional object detection and human pose data. Unless otherwise stated, separate models are trained for AVA and AVA-Kinetics. * denotes the model was trained on AVA-Kinetics and evaluated on AVA. “Res.” denotes the frame resolution of the shorter side. Web-scale foundational models are denoted in grey. † We also report our reimplemented results for TubeR, as the authors’ public code does not reproduce their reported results, as detailed in the supplement." + }, + { + "type": "table", + "bbox": [ + 0.097, + 0.15, + 0.872, + 0.419 + ], + "angle": 0, + "content": "
PretrainingViewsAVAAVA-KRes.BackboneEnd-to-end
MViT-B [13]K400127.3--MViTX
Unified [2]K400627.7-320SlowFastX
AIA [53]K7001832.3-320SlowFastX
ACAR [40]K700633.336.4320SlowFastX
TubeR [66] with LTC [61]IG65M [37]→K400, COCO233.6-256CSN-152X
MeMViT [62]K700-34.4-312MViT v2X
Co-finetuning [3]IN21K→K700, MiT, SSv2132.833.1320ViViT/LX
JFT,WTS→K700, MiT, SSv2136.136.2320ViViT/LX
VideoMAE [54]SSL K700 → Sup. K700.-39.3--ViViT/LX
InternVideo* [60]7 different datasets-41.042.5-Uniformer v2X
VideoMAE v2 [57]6 different datasets-42.643.9-ViViT/gX
Action Transformer [28]K4001-23.0400I3D✓
WOO [9]K600128.3-320SlowFast✓
TubeR [66]IG65M [37]→K400, COCO131.1-256CSN-152✓
TubeR reimplemented†IG65M [37]→K400129.533.6256CSN-152✓
STAR/CSN-152 (ours)IG65M→K400131.435.8256CSN-152✓
STAR/B (ours)IN21K→K400130.036.6320ViViT/B✓
CLIP→K700133.939.1320ViViT/B✓
STAR/L (ours)CLIP→K700139.244.5320ViViT/L✓
STAR/L (ours)*CLIP→K700142.545.1420ViViT/L✓
" + }, + { + "type": "table_caption", + "bbox": [ + 0.15, + 0.421, + 0.819, + 0.433 + ], + "angle": 0, + "content": "Table 10. State-of-the-art comparison on datasets with tubelet annotations, UCF101, JHMDB51 and MultiSports." + }, + { + "type": "table", + "bbox": [ + 0.097, + 0.433, + 0.875, + 0.589 + ], + "angle": 0, + "content": "
PretrainingUCF101-24JHMDB51-21MultiSportsBackbone
fAPvAP20vAP50vAP50:95fAPvAP20vAP50fAPvAP20vAP50
ACT [22]IN1K67.177.251.425.065.774.273.7---VGG
MOC [31]IN1K→COCO78.082.853.828.370.877.377.225.212.90.6DLA34
Unified [2]K60079.3---------SlowFast
WOO [9]K600----80.5-----SlowFast
TubeR [66]IG65M→K40083.283.358.428.9-87.482.3---CSN-152
HIT [14]K70084.888.874.3-83.889.788.133.327.88.8SlowFast
FBIL [8]K70086.086.969.1----40.830.39.9SlowFast
STAR/CSN-152IG65M→K40086.787.065.430.693.596.395.445.450.118.6CSN-152
STAR/BIN21K→K40087.388.268.631.786.989.588.246.351.625.6ViViT/B
STAR/LCLIP→K70090.389.873.435.892.193.192.659.362.036.2ViViT/L
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.591, + 0.471, + 0.773 + ], + "angle": 0, + "content": "by scaling up to ViViT-Large, consistent with our results on AVA (Tab. 9). Moreover, note how we substantially outperform TubeR [66] using the same CSN-152 backbone. Our margin of improvement over TubeR grows from vAP20 (+3.7) to vAP50 (+7.0) with the same backbone, showing that our tubelets are more precise, and is in line with our visual observations in the supplement. To our knowledge, we outperform the best previous reported Video AP50 result by 13.2 points. Note that as UCF videos are up to 900 frames, and as our network processes \\( T = 32 \\) frames, we follow prior works and link together tubelets using the same causal algorithm as [2, 22, 31, 47] for fair comparison." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.78, + 0.471, + 0.903 + ], + "angle": 0, + "content": "JHMDB51-21 Tab. 10 shows that we surpass the state-of-the-art on JHMDB. Once again, we significantly outperform TubeR [66], by 13.1 vAP50, with the same CSN-152 backbone. The CSN-152 backbone outperforms ViViT in this case, possibly because this is the smallest dataset and larger backbones can overfit more easily. The videos in this dataset are trimmed (meaning that labelled actions are being performed on each frame), and also shorter. Hence the" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.591, + 0.892, + 0.636 + ], + "angle": 0, + "content": "Video AP is not as strict as it is on UCF. As videos are at most 40 frames, we set \\( T = 40 \\) in our model so that we process the entire clip at once without needing to link tubelets." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.637, + 0.892, + 0.742 + ], + "angle": 0, + "content": "MultiSports Finally, Tab. 10 shows that we improve on prior state-of-the-art on MultiSports substantially on both Frame AP and Video AP. Although this is a recent benchmark with few methods to compare against, our improvement is significant due to the challenging nature of the dataset: it contains fast motion, many simultaneous actors and large variation in action duration." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.747, + 0.771, + 0.763 + ], + "angle": 0, + "content": "5. Conclusion and Future Work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.892, + 0.901 + ], + "angle": 0, + "content": "We have presented STAR, an end-to-end spatio-temporal action localisation model that can output tubelets, when either sparse keyframe, or full tubelet annotation is available. Our approach achieves state-of-the-art results on four action localisation datasets for both frame-level and tubelet-level predictions (in particular, we obtain \\(45.1\\%\\) mAP on the challenging AVA-Kinetics dataset), outperforming complex methods that use external proposals and memory banks. Future work is to extend our method to open vocabularies." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "18380" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.156 + ], + "angle": 0, + "content": "[1] Anurag Arnab, Mostafa Dehghani, Georg Heigold, Chen Sun, Mario Lucic, and Cordelia Schmid. ViViT: A Video Vision Transformer. In ICCV, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.159, + 0.471, + 0.199 + ], + "angle": 0, + "content": "[2] Anurag Arnab, Chen Sun, and Cordelia Schmid. Unified graph structured models for video understanding. In ICCV, 2021. 1, 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.202, + 0.471, + 0.27 + ], + "angle": 0, + "content": "[3] Anurag Arnab, Xuehan Xiong, Alexey Gritsenko, Rob Romijnders, Josip Djolonga, Mostafa Dehghani, Chen Sun, Mario Lucic, and Cordelia Schmid. Beyond transfer learning: Co-finetuning for action localisation. In arXiv preprint arXiv:2207.03807, 2022. 1, 2, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.272, + 0.469, + 0.313 + ], + "angle": 0, + "content": "[4] Jimmy Lei Ba, Jamie Ryan Kiros, and Geoffrey E Hinton. Layer normalization. In arXiv preprint arXiv:1607.06450, 2016. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.316, + 0.469, + 0.357 + ], + "angle": 0, + "content": "[5] Fabien Baradel, Natalia Neverova, Christian Wolf, Julien Mille, and Greg Mori. Object level visual reasoning in videos. In ECCV, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.36, + 0.469, + 0.413 + ], + "angle": 0, + "content": "[6] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In ECCV, 2020. 1, 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.416, + 0.469, + 0.457 + ], + "angle": 0, + "content": "[7] Joao Carreira, Eric Noland, Chloe Hillier, and Andrew Zisserman. A short note on the kinetics-700 human action dataset. In arXiv preprint arXiv:1907.06987, 2019. 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.459, + 0.469, + 0.5 + ], + "angle": 0, + "content": "[8] Keke Chen, Xiangbo Shu, Guo-Sen Xie, Rui Yan, and Jinhui Tang. Foreground/background-masked interaction learning for spatio-temporal action detection. In ACM MM, 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.502, + 0.469, + 0.557 + ], + "angle": 0, + "content": "[9] Shoufa Chen, Peize Sun, Enze Xie, Chongjian Ge, Jiannan Wu, Lan Ma, Jiajun Shen, and Ping Luo. Watch only once: An end-to-end video action detection framework. In ICCV, 2021. 1, 2, 4, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.559, + 0.469, + 0.599 + ], + "angle": 0, + "content": "[10] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. ImageNet: A large-scale hierarchical image database. In CVPR, 2009. 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.602, + 0.469, + 0.643 + ], + "angle": 0, + "content": "[11] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. BERT: Pre-training of deep bidirectional transformers for language understanding. In NAACL, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.646, + 0.469, + 0.726 + ], + "angle": 0, + "content": "[12] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2021. 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.73, + 0.469, + 0.77 + ], + "angle": 0, + "content": "[13] Haoqi Fan, Bo Xiong, Karttikeya Mangalam, Yanghao Li, Zhicheng Yan, Jitendra Malik, and Christoph Feichtenhofer. Multiscale vision transformers. In ICCV, 2021. 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.773, + 0.469, + 0.813 + ], + "angle": 0, + "content": "[14] Gueter Josmy Faure, Min-Hung Chen, and Shang-Hong Lai. Holistic interaction transformer network for action detection. In WACV, 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.816, + 0.469, + 0.856 + ], + "angle": 0, + "content": "[15] Christoph Feichtenhofer, Haoqi Fan, Jitendra Malik, and Kaiming He. Slowfast networks for video recognition. In ICCV, 2019. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.859, + 0.469, + 0.899 + ], + "angle": 0, + "content": "[16] Rohit Girdhar, Joao Carreira, Carl Doersch, and Andrew Zisserman. Video action transformer network. In CVPR, 2019. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.471, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.814, + 0.106 + ], + "angle": 0, + "content": "[17] Ross Girshick. Fast r-cnn. In ICCV, 2015. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.109, + 0.892, + 0.177 + ], + "angle": 0, + "content": "[18] Chunhui Gu, Chen Sun, David A Ross, Carl Vondrick, Caroline Pantofaru, Yeqing Li, Sudheendra Vijayanarasimhan, George Toderici, Susanna Ricco, Rahul Sukthankar, et al. Ava: A video dataset of spatio-temporally localized atomic visual actions. In CVPR, 2018. 1, 2, 4, 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.18, + 0.892, + 0.208 + ], + "angle": 0, + "content": "[19] Dan Hendrycks and Kevin Gimpel. Gaussian error linear units (gelus). In arXiv preprint arXiv:1606.08415, 2016. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.21, + 0.892, + 0.278 + ], + "angle": 0, + "content": "[20] Jonathan Huang, Vivek Rathod, Chen Sun, Menglong Zhu, Anoop Korattikara, Alireza Fathi, Ian Fischer, Zbigniew Wojna, Yang Song, Sergio Guadarrama, et al. Speed/accuracy trade-offs for modern convolutional object detectors. In CVPR, 2017. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.281, + 0.892, + 0.322 + ], + "angle": 0, + "content": "[21] H. Jhuang, J. Gall, S. Zuffi, C. Schmid, and M. J. Black. Towards understanding action recognition. In ICCV, 2013. 1, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.325, + 0.892, + 0.366 + ], + "angle": 0, + "content": "[22] Vicky Kalogeiton, Philippe Weinzaepfel, Vittorio Ferrari, and Cordelia Schmid. Action tubelet detector for spatiotemporal action localization. In ICCV, 2017. 1, 2, 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.369, + 0.892, + 0.437 + ], + "angle": 0, + "content": "[23] Will Kay, Joao Carreira, Karen Simonyan, Brian Zhang, Chloe Hillier, Sudheendra Vijayanarasimhan, Fabio Viola, Tim Green, Trevor Back, Paul Natsev, et al. The kinetics human action video dataset. In arXiv preprint arXiv:1705.06950, 2017. 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.44, + 0.892, + 0.508 + ], + "angle": 0, + "content": "[24] Skanda Koppula, Yazhe Li, Evan Shelhamer, Andrew Jaegle, Nikhil Parthasarathy, Relja Arandjelovic, João Carreira, and Olivier Henaff. Where should i spend my flops? efficiency evaluations of visual pre-training methods. In arXiv preprint arXiv:2209.15589, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.511, + 0.892, + 0.566 + ], + "angle": 0, + "content": "[25] Okan Köpukli, Xiangyu Wei, and Gerhard Rigoll. You only watch once: A unified cnn architecture for real-time spatiotemporal action localization. In arXiv preprint arXiv:1911.06644, 2019. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.569, + 0.892, + 0.609 + ], + "angle": 0, + "content": "[26] Harold W Kuhn. The hungarian method for the assignment problem. Naval research logistics quarterly, 2(1-2):83-97, 1955. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.612, + 0.892, + 0.653 + ], + "angle": 0, + "content": "[27] Jie Lei, Tamara L Berg, and Mohit Bansal. Detecting moments and highlights in videos via natural language queries. In NeurIPS, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.656, + 0.892, + 0.712 + ], + "angle": 0, + "content": "[28] Ang Li, Meghanata Thotakuri, David A Ross, João Carreira, Alexander Vostrikov, and Andrew Zisserman. The a-akinetics localized human actions video dataset. In arXiv preprint arXiv:2005.00214, 2020. 1, 2, 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.714, + 0.892, + 0.755 + ], + "angle": 0, + "content": "[29] Dong Li, Zhaofan Qiu, Qi Dai, Ting Yao, and Tao Mei. Recurrent tubelet proposal and recognition networks for action detection. In ECCV, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.758, + 0.892, + 0.813 + ], + "angle": 0, + "content": "[30] Kunchang Li, Yali Wang, Yinan He, Yizhuo Li, Yi Wang, Limin Wang, and Yu Qiao. Uniformerv2: Spatiotemporal learning by arming image vits with video uniformer. In arXiv preprint arXiv:2211.09552, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.816, + 0.892, + 0.843 + ], + "angle": 0, + "content": "[31] Yixuan Li, Zixu Wang, Limin Wang, and Gangshan Wu. Actions as moving points. In ECCV, 2020. 2, 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.846, + 0.892, + 0.899 + ], + "angle": 0, + "content": "[32] Yixuan Li, Lei Chen, Runyu He, Zhenzhi Wang, Gangshan Wu, and Limin Wang. Multisports: A multi-person video dataset of spatio-temporally localized sports actions. In ICCV, 2021. 2, 5" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.518, + 0.956 + ], + "angle": 0, + "content": "18381" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.147 + ], + "angle": 0, + "content": "[33] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In ECCV, 2014. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.148, + 0.471, + 0.203 + ], + "angle": 0, + "content": "[34] Ziyi Lin, Shijie Geng, Renrui Zhang, Peng Gao, Gerard de Melo, Xiaogang Wang, Jifeng Dai, Yu Qiao, and Hongsheng Li. Frozen clip models are efficient video learners. In ECCV, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.204, + 0.471, + 0.245 + ], + "angle": 0, + "content": "[35] Wei Liu, Dragomir Anguelov, Dumitru Erhan, Christian Szegedy, Scott Reed, Cheng-Yang Fu, and Alexander C Berg. Ssd: Single shot multibox detector. In ECCV, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.246, + 0.471, + 0.3 + ], + "angle": 0, + "content": "[36] Xiaolong Liu, Qimeng Wang, Yao Hu, Xu Tang, Shiwei Zhang, Song Bai, and Xiang Bai. End-to-end temporal action detection with transformer. IEEE Transactions on Image Processing, 31:5427-5441, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.301, + 0.471, + 0.357 + ], + "angle": 0, + "content": "[37] Dhruv Mahajan, Ross Girshick, Vignesh Ramanathan, Kaiming He, Manohar Paluri, Yixuan Li, Ashwin Bharambe, and Laurens Van Der Maaten. Exploring the limits of weakly supervised pretraining. In ECCV, 2018. 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.357, + 0.471, + 0.426 + ], + "angle": 0, + "content": "[38] Matthias Minderer, Alexey Gritsenko, Austin Stone, Maxim Neumann, Dirk Weissenborn, Alexey Dosovitskiy, Aravindh Mahendran, Anurag Arnab, Mostafa Dehghani, Zhuoran Shen, et al. Simple Open-Vocabulary Object Detection with Vision Transformers. In ECCV, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.426, + 0.471, + 0.468 + ], + "angle": 0, + "content": "[39] Megha Nawhal and Greg Mori. Activity graph transformer for temporal action localization. In arXiv preprint arXiv:2101.08540, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.469, + 0.471, + 0.523 + ], + "angle": 0, + "content": "[40] Junting Pan, Siyu Chen, Mike Zheng Shou, Yu Liu, Jing Shao, and Hongsheng Li. Actor-context-actor relation network for spatio-temporal action localization. In CVPR, 2021. 1, 2, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.525, + 0.471, + 0.566 + ], + "angle": 0, + "content": "[41] Junting Pan, Ziyi Lin, Xiatian Zhu, Jing Shao, and Hongsheng Li. St-adapter: Parameter-efficient image-to-video transfer learning. In NeurIPS, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.567, + 0.471, + 0.635 + ], + "angle": 0, + "content": "[42] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.636, + 0.471, + 0.677 + ], + "angle": 0, + "content": "[43] Joseph Redmon, Santosh Divvala, Ross Girshick, and Ali Farhadi. You only look once: Unified, real-time object detection. In CVPR, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.678, + 0.471, + 0.719 + ], + "angle": 0, + "content": "[44] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. In NeurIPS, 2015. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.72, + 0.471, + 0.775 + ], + "angle": 0, + "content": "[45] Hamid Rezatofighi, Nathan Tsoi, JunYoung Gwak, Amir Sadeghian, Ian Reid, and Silvio Savarese. Generalized intersection over union: A metric and a loss for bounding box regression. In CVPR, 2019. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.776, + 0.471, + 0.803 + ], + "angle": 0, + "content": "[46] Bharat Singh and Larry S Davis. An analysis of scale invariance in object detection snip. In CVPR, 2018. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.804, + 0.471, + 0.858 + ], + "angle": 0, + "content": "[47] Gurkirt Singh, Suman Saha, Michael Sapienza, Philip HS Torr, and Fabio Cuzzolin. Online real-time multiple spatiotemporal action localisation and prediction. In ICCV, 2017. 2, 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.859, + 0.471, + 0.9 + ], + "angle": 0, + "content": "[48] Lin Song, Shiwei Zhang, Gang Yu, and Hongbin Sun. Tacnet: Transition-aware context network for spatio-temporal action detection. In CVPR, 2019. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.471, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.134 + ], + "angle": 0, + "content": "[49] Khurram Soomro, Amir Roshan Zamir, and Mubarak Shah. Ucf101: A dataset of 101 human actions classes from videos in the wild. In arXiv preprint arXiv:1212.0402, 2012. 1, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.135, + 0.892, + 0.189 + ], + "angle": 0, + "content": "[50] Andreas Steiner, Alexander Kolesnikov, Xiaohua Zhai, Ross Wightman, Jakob Uszkoreit, and Lucas Beyer. How to train your vit? data, augmentation, and regularization in vision transformers. TMLR, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.19, + 0.892, + 0.231 + ], + "angle": 0, + "content": "[51] Chen Sun, Abhinav Shrivastava, Carl Vondrick, Kevin Murphy, Rahul Sukthankar, and Cordelia Schmid. Actor-centric relation network. In ECCV, 2018. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.232, + 0.892, + 0.287 + ], + "angle": 0, + "content": "[52] Peize Sun, Rufeng Zhang, Yi Jiang, Tao Kong, Chenfeng Xu, Wei Zhan, Masayoshi Tomizuka, Lei Li, Zehuan Yuan, Changhu Wang, et al. Sparse r-cnn: End-to-end object detection with learnable proposals. In CVPR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.288, + 0.892, + 0.328 + ], + "angle": 0, + "content": "[53] Jiajun Tang, Jin Xia, Xinzhi Mu, Bo Pang, and Cewu Lu. Asynchronous interaction aggregation for action detection. In ECCV, 2020. 1, 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.329, + 0.892, + 0.383 + ], + "angle": 0, + "content": "[54] Zhan Tong, Yibing Song, Jue Wang, and Limin Wang. VideoMAE: Masked autoencoders are data-efficient learners for self-supervised video pre-training. In NeurIPS, 2022. 2, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.385, + 0.892, + 0.426 + ], + "angle": 0, + "content": "[55] Du Tran, Heng Wang, Lorenzo Torresani, and Matt Feiszli. Video classification with channel-separated convolutional networks. In ICCV, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.427, + 0.892, + 0.481 + ], + "angle": 0, + "content": "[56] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In NeurIPS, 2017. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.483, + 0.892, + 0.538 + ], + "angle": 0, + "content": "[57] Limin Wang, Bingkun Huang, Zhiyu Zhao, Zhan Tong, Yi-nan He, Yi Wang, Yali Wang, and Yu Qiao. Videomae v2: Scaling video masked autoencoders with dual masking. In arXiv preprint arXiv:2303.16727, 2023. 2, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.539, + 0.892, + 0.566 + ], + "angle": 0, + "content": "[58] Xiaolong Wang and Abhinav Gupta. Videos as space-time region graphs. In ECCV, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.567, + 0.892, + 0.621 + ], + "angle": 0, + "content": "[59] Yuqing Wang, Zhaoliang Xu, Xinlong Wang, Chunhua Shen, Baoshan Cheng, Hao Shen, and Huaxia Xia. End-to-end video instance segmentation with transformers. In CVPR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.622, + 0.892, + 0.691 + ], + "angle": 0, + "content": "[60] Yi Wang, Kunchang Li, Yizhuo Li, Yinan He, Bingkun Huang, Zhiyu Zhao, Hongjie Zhang, Jilan Xu, Yi Liu, Zun Wang, et al. Internvideo: General video foundation models via generative and discriminative learning. In arXiv preprint arXiv:2212.03191, 2022. 2, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.692, + 0.892, + 0.747 + ], + "angle": 0, + "content": "[61] Chao-Yuan Wu, Christoph Feichtenhofer, Haoqi Fan, Kaiming He, Philipp Krahenbuhl, and Ross Girshick. Long-term feature banks for detailed video understanding. In CVPR, 2019. 1, 2, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.748, + 0.892, + 0.816 + ], + "angle": 0, + "content": "[62] Chao-Yuan Wu, Yanghao Li, Karttikeya Mangalam, Haoqi Fan, Bo Xiong, Jitendra Malik, and Christoph Feichtenhofer. MeMVit: Memory-augmented multiscale vision transformer for efficient long-term video recognition. In CVPR, 2022. 1, 2, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.817, + 0.892, + 0.858 + ], + "angle": 0, + "content": "[63] Chuhan Zhang, Ankush Gupta, and Andrew Zisserman. Temporal query networks for fine-grained video understanding. In CVPR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.859, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[64] Yubo Zhang, Pavel Tokmakov, Martial Hebert, and Cordelia Schmid. A structured model for action detection. In CVPR, 2019. 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "18382" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.161 + ], + "angle": 0, + "content": "[65] Jiaojiao Zhao, Yanyi Zhang, Xinyu Li, Hao Chen, Bing Shuai, Mingze Xu, Chunhui Liu, Kaustav Kundu, Yuanjun Xiong, Davide Modolo, et al. Tuber public code. https://github.com/amazon-science/tubelet-transformer, 2022.6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.163, + 0.47, + 0.219 + ], + "angle": 0, + "content": "[66] Jiaojiao Zhao, Yanyi Zhang, Xinyu Li, Hao Chen, Bing Shuai, Mingze Xu, Chunhui Liu, Kaustav Kundu, Yuanjun Xiong, Davide Modolo, et al. TubeR: Tubelet Transformer for Video Action Detection. In CVPR, 2022. 1, 2, 4, 5, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.22, + 0.47, + 0.26 + ], + "angle": 0, + "content": "[67] Xingyi Zhou, Dequan Wang, and Philipp Krahenbuhl. Objects as points. In arXiv preprint arXiv:1904.07850, 2019. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.26 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.956 + ], + "angle": 0, + "content": "18383" + } + ] +] \ No newline at end of file diff --git a/2024/End-to-End Spatio-Temporal Action Localisation with Video Transformers/62e0dda5-0752-47db-aad5-cb1b83945b43_origin.pdf b/2024/End-to-End Spatio-Temporal Action Localisation with Video Transformers/62e0dda5-0752-47db-aad5-cb1b83945b43_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..3d4953d37aef52d34604fa64ef555135a70b3f63 --- /dev/null +++ b/2024/End-to-End Spatio-Temporal Action Localisation with Video Transformers/62e0dda5-0752-47db-aad5-cb1b83945b43_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1413b932f0e70d8a6ecd1434ba6f7cceeb313b0c6a268b969579dfe9b95a441f +size 726641 diff --git a/2024/End-to-End Spatio-Temporal Action Localisation with Video Transformers/full.md b/2024/End-to-End Spatio-Temporal Action Localisation with Video Transformers/full.md new file mode 100644 index 0000000000000000000000000000000000000000..4946b43b923479463dc2ccd413543268d01096d8 --- /dev/null +++ b/2024/End-to-End Spatio-Temporal Action Localisation with Video Transformers/full.md @@ -0,0 +1,332 @@ +# End-to-End Spatio-Temporal Action Localisation with Video Transformers + +Alexey A. Gritsenko Xuehan Xiong Josip Djolonga Mostafa Dehghani Chen Sun Mario Lučić Cordelia Schmid Anurag Arnab Google + +{agritsenko, aarnab}@google.com + +# Abstract + +The most performant spatio-temporal action localisation models use external person proposals and complex external memory banks. We propose a fully end-to-end, purely-transformer based model that directly ingests an input video, and outputs tubelets - a sequence of bounding boxes and the action classes at each frame. Our flexible model can be trained with either sparse bounding-box supervision on individual frames, or full tubelet annotations. And in both cases, it predicts coherent tubelets as the output. Moreover, our end-to-end model requires no additional pre-processing in the form of proposals, or post-processing in terms of non-maximal suppression. We perform extensive ablation experiments, and significantly advance the state-of-the-art on five different spatio-temporal action localisation benchmarks with both sparse keyframes and full tubelet annotations. + +# 1. Introduction + +Spatio-temporal action localisation is an important problem with applications in advanced video search engines, robotics and security among others. It is typically formulated in one of two ways: Firstly, predicting the bounding boxes and actions performed by an actor at a single keyframe given neighbouring frames as spatio-temporal context [18, 28]. Or alternatively, predicting a sequence of bounding boxes and actions (i.e. "tubes"), for each actor at each frame in the video [21, 49]. + +The most performant models [3, 15, 40, 62], particularly for the first, keyframe-based formulation of the problem, employ a two-stage pipeline inspired by the Fast-RCNN object detector [17]: They first run a separate person detector to obtain proposals. Features from these proposals are then aggregated and classified according to the actions of interest. These models have also been supplemented with memory banks containing long-term contextual information from other frames [40, 53, 61, 62], and/or detections of other potentially relevant objects [2, 53] to capture additional scene context, achieving state-of-the-art results. + +![](images/248495399aef562cd7b0c1d860b436c8af9430ed0e3230775834ee9027aa8a88.jpg) +Figure 1. We propose an end-to-end Spatio-Temporal Action Recognition model named STAR. Our model is end-to-end in that it does not require any external region proposals to predict tubelets – sequences of bounding boxes associated with a given person in every frame and their corresponding action classes. Our model can be trained with either sparse box annotations on selected keyframes, or full tubelet supervision. + +And whilst proposal-free algorithms, which do not require external person detectors, have been developed for detecting both at the keyframe-level [9, 25, 51] and tubelet-level [22, 66], their performance has typically lagged behind their proposal-based counterparts. Here, we show for the first time that an end-to-end trainable spatio-temporal model outperforms a two-stage approach. + +As shown in Fig. 1, we propose our Spatio-Temporal Action TransformeR (STAR) that consists of a transformer architecture, and is based on the DETR [6] detection model. Our model is "end-to-end" in that it does not require preprocessing in the form of proposals, nor post-processing in the form of non-maximal suppression (NMS) in contrast to the majority of prior work. The initial stage of the model is a vision encoder. This is followed by a decoder that processes learned latent queries, which represent each actor in the video, into output tubelets - a sequence of bounding boxes and action classes at each time step of + +the input video clip. Our model is versatile in that we can train it with either fully-labeled tube annotations, or with sparse keyframe annotations (when only a limited number of keyframes are labelled). In the latter case, our network still predicts tubelets, and learns to associate detections of an actor, from one frame to the next, without explicit supervision. This behaviour is facilitated by our formulation of factorised queries, decoder architecture and tubelet matching in the loss which all contain temporal inductive biases. + +We conduct thorough ablation studies of these modelling choices, confirming the benefit of temporal inductive biases in our model design. Informed by these experiments, we achieve state-of-the-art on both keyframe-based action localisation datasets like AVA [18] and AVA-Kinetics [28], and also tubelet-based datasets like UCF101-24 [49], JHMDB [21] and MultiSports [32]. In particular, we achieve a Frame mAP of 45.1 on AVA-Kinetics, outperforming the best previous results achieved by a massive video foundation model [57]. Moreover, our state-of-the-art results are achieved with a single forward-pass through the model, using only a video clip as input, and without any separate external person detectors providing proposals [57, 60, 62], complex memory banks [40, 62, 66], or additional object detectors [2, 53], as used by the prior works. Furthermore, we outperform these complex, prior, state-of-the-art two-stage models whilst also having additional functionality in that our model predicts tubelets, that is, temporally consistent bounding boxes at each frame of the input video clip. This capability is demonstrated by our results on tube-based datasets like UCF101-24 where we surpass the prior work [66] by 13.2 points on Video AP50. + +# 2. Related Work + +Models for spatio-temporal action localisation have typically built upon advances in object detectors for images. The most performant methods for action localisation [3, 15, 40, 53, 62] are based on "two-stage" detectors like Fast-RCNN [17]. These models use external, pre-computed person detections, and use them to ROI-pool features which are then classified into action classes. Although these models are cumbersome in that they require an additional model and backbone to first detect people, and therefore additional detection training data as well, they are currently the leading approaches on datasets such as AVA [18]. Such models using external proposals are also particularly suited to datasets such as AVA [18] as each person is exhaustively labelled as performing an action, and therefore there are fewer false-positives from using action-agnostic person detections compared to datasets such as UCF101 [49]. + +The performance of these two-stage models has further been improved by incorporating more contextual information using feature banks extracted from additional frames in the video [40, 53, 61, 62] or by utilising detections of ad + +ditional objects in the scene [2, 5, 58, 64]. Both of these cases entail significant extra computation and complexity to train additional auxiliary models, and to precompute features from them that are then used during training and inference of the localisation model. + +Our proposed method, in contrast, is end-to-end in that it directly produces detections without any additional inputs besides a video clip. Moreover, it outperforms these prior works without resorting to external proposals or memory banks, showing that a transformer backbone is sufficient to capture long-range dependencies in the input video. In addition, unlike previous two-stage methods, our method directly predicts tubelets: a sequence of bounding boxes and actions for each frame of the input video, and can do so even when we do not have full tubelet annotations available. + +A number of proposal-free action localisation models have also been developed [9, 16, 22, 25, 51, 66]. These methods are based upon alternative object detection architectures such as SSD [35], CentreNet [67], YOLO [43], DETR [6] and Sparse-RCNN [52]. However, in contrast to our approach, they have been outperformed by their proposal-based counterparts. Moreover, some of these methods [16, 25, 51] also consist of separate network backbones for learning video feature representations and proposals for a keyframe, and are thus effectively two networks trained jointly, and cannot predict tubelets either. + +Among prior works that do not use external proposals, and also directly predict tubelets [22, 29, 31, 47, 48], our work is most similar to TubeR [66] given that our model is also based on DETR. Our model, however, is designed with additional temporal inductive biases which improves accuracy (without using external memory banks precomputed offline like [66]. And moreover, unlike TubeR, we also demonstrate how our model can predict tubelets (i.e. predictions at every frame of the input video), even when we only have sparse keyframe supervision (i.e. ground truth annotation for a limited number of frames) available. + +Finally, we note that DETR has also been extended as a proposal-free method to addressing other localisation tasks in video such as instance segmentation [59], temporal localisation [36, 39, 63] and moment retrieval [27]. + +# 3. Spatio-Temporal Action Transformer + +Our proposed model ingests a sequence of video frames, and directly predicts tubelets (a sequence of bounding boxes and action labels). No external person detections [40, 54, 57], or memory banks [62, 66], are needed. + +As summarised in Fig. 2, our model consists of a vision encoder (Sec. 3.1), followed by a decoder which processes learned query tokens into output tubelets (Sec. 3.2). We incorporate temporal inductive biases into our decoder to improve accuracy and tubelet prediction with weaker supervision. Our model is inspired by the DETR architecture [6] + +![](images/f91408e60bee52cb17a7ef82c064054ea15cf6fb90c4c239741b9f8ca8c0ebfd.jpg) +Figure 2. Our model processes a fixed-length video clip, and for each frame, outputs tubelets (i.e. linked bounding boxes with associated action class probabilities). It consists of vision encoder which outputs a video representation, $\mathbf{x} \in \mathbb{R}^{T \times h \times w \times d}$ . The video representation, along with learned queries, $\mathbf{q}$ (which are factorised into spatial $\mathbf{q}^s$ and temporal components $\mathbf{q}^t$ ) are decoded into tubelets by a decoder of + +$L$ layers followed by shallow box and class prediction heads. + +for object detection in images, and is also trained with a set-based loss and Hungarian matching. We detail our loss, and how we can train with either sparse keyframe supervision or full tubelet supervision, in Sec. 3.3. + +# 3.1. Vision Encoder + +The vision backbone processes an input video, $\mathbf{X} \in \mathbb{R}^{T \times H \times W \times 3}$ to produce a feature representation of the input video $\mathbf{x} \in \mathbb{R}^{t \times h \times w \times d}$ . Here, $T$ , $H$ and $W$ are the original temporal-, height- and width-dimensions of the input video respectively, whilst $t$ , $h$ and $w$ are the spatio-temporal dimensions of their feature representation, and $d$ its latent dimension. When using a transformer backbone, these spatiotemporal dimensions depend on the patch size when tokenising the input, and when using a convolutional backbone, they depend on the overall stride. To retain spatiotemporal information, we remove the spatial- and temporal-aggregation steps at the end of the original backbone. And if the temporal patch size (or stride) is larger than 1, we bilinearly upsample the final feature map along the temporal axis to maintain the original temporal resolution. + +# 3.2. Tubelet Decoder + +Our decoder processes the visual features, $\mathbf{x} \in \mathbb{R}^{T \times h \times w \times c}$ , along with learned queries, $\mathbf{q} \in \mathbb{R}^{T \times S \times d}$ , to output tubelets, $\mathbf{y} = (\mathbf{b}, \mathbf{a})$ which are a sequence of bounding boxes, $b \in \mathbb{R}^{T \times S \times 4}$ and corresponding actions, $a \in \mathbb{R}^{T \times S \times C}$ . Here, $S$ denotes the maximum number of bounding boxes per frame (padded with "background" as necessary) and $C$ denotes the number of output classes. + +The idea of decoding learned queries into output detections using the transformer decoder architecture of Vaswani et al. [56] was used in DETR [6]. In summary, the decoder of [6, 56] consists of $L$ layers, each performing a series of self-attention operations on the queries, and cross-attention between the queries and encoder outputs. + +We modify the queries, self-attention and cross-attention operations for our spatio-temporal localisation scenario, as shown in Fig. 2 and 3 to include additional temporal inductive biases, and to improve accuracy as detailed below. + +![](images/ab85d5469b56bb6f27c3dafa2a47896ec4d5bb7af2060b6907190866c9781041.jpg) +Figure 3. Our decoder layer consists of factorised self-attention (SA) (left) and cross-attention (CA) (right) operations designed to provide a spatio-temporal inductive bias and reduce computation. Both operations restrict attention to the same spatial and temporal slices as the query token, as illustrated by the receptive field (blue) for a given query token (magenta). Factorised SA consists of two operations, whilst in Factorised CA, there is one operation. + +Queries Queries, q, in DETR, are decoded using the encoded visual features, x, into bounding box predictions, and are analogous to the "anchors" used in other detection architectures such as Faster-RCNN [44]. + +The most straightforward way to define queries is to randomly initialise $\mathbf{q} \in \mathbb{R}^{T \times S \times d}$ , where there are $S$ bounding boxes at each of the $T$ input frames in the video clip. + +However, we find it is more effective to factorise the queries into separate learned spatial, $\mathbf{q}^s\in \mathbb{R}^{S\times d}$ , and temporal, $\mathbf{q}^{T\times d}$ parameters. To obtain the final tubelet queries, we simply repeat the spatial queries across all frames, and add them to their corresponding temporal embedding at each location, as shown in Fig. 2. More concretely $\mathbf{q}_{ij} = \mathbf{q}_i^t +\mathbf{q}_j^s$ where $i$ and $j$ denote the temporal and spatial indices respectively. + +The factorised query representation means that the same spatial embedding is used across all frames. Intuitively, this encourages the $i^{th}$ spatial query embedding, $\mathbf{q}_i^s$ , to bind to the same location across different frames of the video, and since objects typically have small displacements from frame to frame, may help to associate bounding boxes within a tubelet together. We verify this intuition empirically in the experimental section. + +Decoder layer The decoder layer in the original transformer [56] consists of self-attention on the queries, $\mathbf{q}$ , followed by cross-attention between the queries and the outputs of the encoder, $\mathbf{x}$ , and then a multilayer perceptron (MLP) layer [19, 56]: + +$$ +\mathbf {u} ^ {\ell} = \operatorname {M H S A} \left(\mathbf {q} ^ {\ell}\right) + \mathbf {q} ^ {\ell}, \tag {1} +$$ + +$$ +\mathbf {v} ^ {\ell} = \operatorname {C A} \left(\mathbf {u} ^ {\ell}, \mathbf {x}\right) + \mathbf {u} ^ {\ell}, \tag {2} +$$ + +$$ +\mathbf {z} ^ {\ell} = \operatorname {M L P} \left(\mathbf {v} ^ {\ell}\right) + \mathbf {v} ^ {\ell}, \tag {3} +$$ + +where $\mathbf{z}^{\ell}$ is the output of the $\ell^{th}$ decoder layer, $\mathbf{u}$ and $\mathbf{v}$ are intermediate variables, MHSA denotes multi-headed self-attention and CA cross-attention. Note that the inputs to the MLP, self- and cross-attention operations are layer-normalised [4], which we omit here for clarity. + +In our model, we factorise the self- and cross-attention layers across space and time respectively as shown in Fig. 3, to introduce a temporal locality inductive bias, and also to increase model efficiency. Concretely, when applying MHSA, we first compute the queries, keys and values, over which we attend twice: first independently at each time step with each frame, and then, independently along the time axis at each spatial location. Similarly, we modify the cross-attention operation so that only tubelet queries and backbone features from the same time index attend to each other. + +Localisation and classification heads We obtain the final predictions of the network, $\mathbf{y} = (\mathbf{b},\mathbf{a})$ , by applying a small feed-forward network to the outputs to the decoder, $\mathbf{z}$ , following DETR [6]. The sequence of bounding boxes, $\mathbf{b}$ , is obtained with a 3-layer MLP, and is parameterised by the box center, width and height for each frame in the tubelet. A single-layer linear projection is used to obtain class logits, $\mathbf{a}$ . As we predict a fixed number of $S$ bounding boxes per frame, and $S$ is more than the maximum number of ground truth instances in the frame, we also include an additional class label, $\varnothing$ , which represents the "background" class which tubelets with no action class can be assigned to. + +# 3.3. Training objective + +Our model predicts bounding boxes and action classes at each frame of the input video. Many datasets, however, such as AVA [18], are only sparsely annotated at selected keyframes of the video. In order to leverage the available annotations, we compute our training loss, Eq. 4, only at the annotated frames of the video, after having matched the predictions to the ground truth: + +$$ +\mathcal {L} (\mathbf {y}, \hat {\mathbf {y}}) = \frac {1}{| \mathcal {T} |} \sum_ {t \in \mathcal {T}} \mathcal {L} _ {\text {f r a m e}} (\mathbf {y}, \hat {\mathbf {y}}), \tag {4} +$$ + +where $\mathcal{T}$ is the set of labelled frames; $\mathbf{y}$ and $\hat{\mathbf{y}}$ denote the ground truth and predicted tubelets after matching. Following DETR [6], our training loss at each frame, $\mathcal{L}_{\mathrm{frame}}$ , is a + +sum of an $L_{1}$ regression loss on bounding boxes, the generalised IoU loss [45] on bounding boxes, and a cross-entropy loss on action labels: + +$$ +\begin{array}{l} \mathcal {L} _ {\mathrm {f r a m e}} \left(\mathbf {b} ^ {t}, \hat {\mathbf {b}} ^ {t}, \mathbf {a} ^ {t}, \hat {\mathbf {a}} ^ {t}\right) = \sum_ {i} \mathcal {L} _ {\mathrm {b o x}} \left(\mathbf {b} _ {i} ^ {t}, \hat {\mathbf {b}} _ {i} ^ {t}\right) + \mathcal {L} _ {\mathrm {i o u}} \left(\mathbf {b} _ {i} ^ {t}, \hat {\mathbf {b}} _ {i} ^ {t}\right) \\ + \mathcal {L} _ {\text {c l a s s}} \left(\mathbf {a} _ {i} ^ {t}, \hat {\mathbf {a}} _ {i} ^ {t}\right). \tag {5} \\ \end{array} +$$ + +Matching Set-based detection models such as DETR can make predictions in any order, which is why the predictions need to be matched to the ground truth before computing the training loss. + +The first form of matching that we consider is to independently perform bipartite matching at each frame to align the model's predictions to the ground truth (or the $\varnothing$ background class) before computing the loss. In this case, we use the Hungarian algorithm [26] to obtain $T$ permutations of $S$ elements, $\hat{\pi}^t\in \Pi^t$ , at each frame, where the permutation at the $t^{th}$ frame minimises the per-frame loss, + +$$ +\hat {\pi} ^ {t} = \underset {\pi \in \Pi^ {t}} {\arg \min } \mathcal {L} _ {\text {f r a m e}} \left(\mathbf {y} ^ {t}, \hat {\mathbf {y}} _ {\pi (i)} ^ {t}\right). \tag {6} +$$ + +An alternative is to perform tubelet matching, where all queries with the same spatial index, $\mathbf{q}^s$ , must match to the same ground truth annotation across all frames of the input video. Here the permutation is obtained over $S$ elements as + +$$ +\hat {\pi} = \underset {\pi \in \Pi} {\arg \min } \frac {1}{| \mathcal {T} |} \sum_ {t \in \mathcal {T}} \mathcal {L} _ {\text {f r a m e}} \left(\mathbf {y} ^ {t}, \hat {\mathbf {y}} _ {\pi^ {t} (i)} ^ {t}\right). \tag {7} +$$ + +Intuitively, tubelet matching provides stronger supervision when we have full tubelet annotations available. Note that regardless of the type of matching that we perform, the loss computation and the overall model architecture remains the same. Note that we do not weight terms in Eq. 5, for both matching and loss calculation, for simplicity, and to avoid having additional hyperparameters, as also done by [38]. + +# 3.4. Discussion + +As our approach is based on DETR, it does not require external proposals nor non-maximal suppression for post-processing. The idea of using DETR for action localisation has also been explored by TubeR [66] and WOO [9]. There are, however, a number of key differences: WOO does not detect tubelets at all, but only actions at the center keyframe. We also factorise our queries in the spatial and temporal dimensions (Sec. 3.2) to provide inductive biases urging spatio-temporal association. Moreover, we predict action classes separately for each time step in the tubelet, meaning that each of our queries binds to an actor in the video. TubeR, in contrast, parameterises queries such that they are each associated with separate actions (features are average-pooled over the tubelet, and then linearly classified into a single action class). This choice also means that TubeR requires an additional "action switch" head to predict when + +tubelets start and end, which we do not require as different time steps in a tubelet can have different action classes in our model. Furthermore, we show experimentally (Tab. 1) that TubeR's parameterisation obtains lower accuracy. We also consider two types of matching in the loss computation (Sec. 3.3) unlike TubeR, with "tubelet matching" designed for predicting more temporally consistent tubelets. And in contrast to TubeR, we experimentally show how our decoder design allows our model to accurately predict tubelets even with weak, keyframe supervision. + +Finally, TubeR requires additional complexity in the form of a "short-term context module" [66] and the external memory bank of [61] which is computed offline using a separate model to achieve strong results. As we show experimentally in the next section, we outperform TubeR without any additional modules, meaning that our model does indeed produce tubelets in an end-to-end manner. + +# 4. Experimental Evaluation + +# 4.1. Experimental set-up + +Datasets We evaluate on four spatio-temporal action localisation benchmarks. AVA and AVA-Kinetics contain sparse annotations at each keyframe, whereas UCF101-24 and JHMDB51-21 contain full tubelet annotations. + +AVA [18] consists of 430, 15-minute video clips from movies. Keyframes are annotated at every second in the video, with about 210 000 labelled frames in the training set, and 57 000 in the validation set. There are 80 atomic actions labelled for every actor in the clip, of which 60 are used for evaluation [18]. Following standard practice, we report the Frame Average Precision (fAP) at an IoU threshold of 0.5 using the latest v2.2 annotations [18]. + +AVA-Kinetics [28] is a superset of AVA, and adds detection annotations following the AVA protocol, to a subset of Kinetics 700 [7] videos. Only a single keyframe in a 10-second Kinetics clip is labelled. In total, about 140 000 labelled keyframes are added to the training set, and 32 000 to the validation sets of AVA. Once again, we follow standard practice in reporting the Frame AP at a 0.5 IoU threshold. + +UCF101-24 [49] is a subset of UCF101, and annotates 24 action classes with full spatio-temporal tubes in 3 207 untrimmed videos. Note that actions are not labelled exhaustively as in AVA, and there may be people present in the video who are not performing any labelled action. Following standard practice, we use the corrected annotations of [47]. We report both the Frame AP, which evaluates the predictions at each frame independently, and also the Video AP. The Video AP uses a 3D, spatio-temporal IoU to match predictions to targets. And since UCF101-24 videos are up to 900 frames long (median length of 164 frames), and our network typically processes $T = 32$ frames at a time, we link together tubelet predictions from our network into full-video-tubes using the same causal linking algo + +Table 1. Comparison of detection architectures on AVA controlling for the same resolution (160p) and training settings. Binding each query to a person, rather than to an action (as done in TubeR [66]) yields solid improvements. We report the mean AP for both ViViT-B and CSN-152 backbones. + +
ViViT-BCSN-152
Query binds to action23.625.7
Ours, query binds to person26.727.8
+ +rithm as [22, 31] for fair comparison. + +JHMDB51-21 [21] also contains full tube annotations in 928 trimmed videos. However, as the videos are shorter and at most 40 frames, we can process the entire clip with our network, and do not need to perform any linking. + +MultiSports [32] is a recent dataset with full tube annotations for 66 fine-grained action classes in 2129 untrimmed videos, which are up to 3676 frames long. It is challenging, as it contains fast motion with multiple actors performing different concurrent actions, and there is a large variance in action durations. Evaluation is the same as for UCF101-24. + +Implementation details For our vision encoder backbone, we consider both transformer-based (ViViT Factorised Encoder [1]), and convolutional (CSN [55]) backbones. For ViViT, we use the "Base" and "Large" model sizes [11, 12], which are typically first pretrained on image datasets like ImageNet-21K [10] and then finetuned on video datasets like Kinetics [23]. We also use CSN-152 pretrained on Instagram65M [37] and then Kinetics following [66]. Our models process $T = 32$ frames unless otherwise specified, with $S = 64$ spatial queries per frame and latent decoder dimensionality of $d = 2048$ . Exhaustive implementation details and training hyperparameters are included in the supplement. We will also release all code and models upon acceptance. + +# 4.2. Ablation studies + +We analyse the design choices in our model by conducting experiments on both AVA (with sparse per-frame supervision) and on UCF101-24 (where we can evaluate the quality of our predicted tubelets). Unless otherwise stated, our backbone is ViViT-Base pretrained on Kinetics 400, and the frame resolution is 160 pixels (160p) on the smaller side. + +Comparison of detection architectures Tab. 1 compares our model, where each query represents a person, and all of their actions (Sec. 3.2) to the approach of TubeR [66] (Sec. 3.4), where there is a separate query for each action being performed. We observe that our parameterisation has a substantial impact, with our method outperforming binding to actions by 3.1 points with a ViViT backbone, and 2.1 points with a CSN backbone on the AVA dataset, therefore motivating the design of our decoder. the supplement shows that this trend is consistent on UCF101-24 and JHMDB too. + +Another architectural baseline that we can compare to is that of a two-stage Fast-RCNN model using external person + +Table 2. Comparison of independent and factorised queries on the AVA, UCF101-24 and JHMDB51-21 datasets. Factorised queries are particularly beneficial for predicting tubelets, as shown by the VideoAP on UCF101-24 and JHMDB51-21 which has full tube annotations. Both models use tubelet matching in the loss. + +
QueryAVA fAPUCF101-24JHMDB51-21
fAPvAP20vAP50vAP50:95fAPvAP20vAP50
Independent25.285.686.359.528.985.088.585.2
Factorised26.386.587.463.429.886.989.588.2
+ +Table 3. Comparison of independent and tubelet matching for computing the loss on AVA, UCF101-24 and JHMD51-21. Tubelet matching helps for tube-level evaluation metrics like the Video AP (vAP) on UCF101-24 and JHMDB51-21. Note that tubelet matching is actually still possible on AVA as the annotations are at 1fps with actor identities. + +
QueryAVA fAPUCF101-24JHMDB51-21
fAPvAP20vAP50vAP50:95fAPvAP20vAP50
Per-frame matching26.788.285.763.529.486.088.386.0
Tubelet matching26.386.587.463.429.886.989.588.2
+ +detections from [61], as used by [3, 13, 15, 62]. This baseline using the same ViViT-B backbone achieved a mean AP of 25.2, which is still 1.5 points below our model, emphasising the promise of our end-to-end approach. Note that the proposals of [61] obtain an AP50 of 93.9 for person detection on the AVA validation set. They were obtained by first pretraining a Faster-RCNN [44] detector on COCO keypoints, and then finetuning on the person boxes from the AVA training set, using a resolution of 1333 on the longer side. Our model is end-to-end, and does not require any external proposals generated by a separate model at all. + +Comparison to TubeR The second row of Tab. 1 using a CSN-152 backbone corresponds to our reimplementation of TubeR. By keeping all other training hyperparameters constant, we observe that our query binding provides an improvement of $2.1\mathrm{mAP}$ points in a fair comparison. Note that we could not use the public TubeR code [65], as it does not reproduce the paper's results: A higher resolution 256p model achieved only $20\mathrm{mAP}$ when trained with the public code, whilst it is reported to achieve 31.1. Exhaustive details on our attempts to reproduce TubeR with the authors' public code is in the supplement. + +Query parameterisation Tab. 2 compares our independent and factorised query methods (Sec. 3.2) on AVA and UCF101-24. We observe that factorised queries consistently provide improvements on both the Frame AP and the Video AP across both datasets. As hypothesised in Sec. 3.2, we believe that this is due to the inductive bias present in this parameterisation. Note that we can only measure the Video AP on UCF101-24 as it has tubes labelled. We also show in the supplement that these observations are consistent on the JHMDB dataset too. + +Matching for loss calculation As described in Sec. 3.3, when matching the predictions to the ground truth for loss + +Table 4. Our model can predict tubelets even when the ground truth annotations are sparse. We show this by subsampling training annotations from the UCF101-24 dataset. Our model sees minimal performance deterioration even when using only $1/24$ or $4\%$ of the annotated frames. + +
SamplingLabelled framesfAPvAP20vAP50vAP50:95
All frames458 81486.587.463.429.8
Every 1239 23785.287.263.029.3
Every 2420 24384.986.863.228.1
One per video2 28470.277.148.520.4
+ +computation, we can either independently match the outputs at each frame to the ground truths at each frame, or, we can match entire predicted tubelets to the ground truth tubelets. Tab. 3 shows that tubelet matching does indeed improve the quality of the predicted tubelets, as shown by the Video AP on UCF101-24. However, this comes at the cost of the quality of per-frame predictions, (i.e. Frame AP). This suggests that tubelet matching improves the association of bounding boxes predicted at different frames (hence higher Video AP), but may also impair the quality of the bounding boxes predicted at each frame (Frame AP). Note that it is technically possible for us to also perform tubelet matching on AVA, since AVA is annotated at 1fps with actor identities, and our model is input 32 frames at 12.5fps (therefore 2.56 seconds of temporal context) meaning that we have sparse tubelets with 2 or 3 annotated frames. + +As tubelet matching helps with the overall Video AP, we use it for subsequent experiments on UCF101-24 and JHMDB51-21. For AVA, we use per-frame matching as the standard evaluation metric is the Frame AP, and annotations are sparse at 1fps. + +Weakly-supervised tubelet detection Our model can predict tubelets even when the ground truth annotations are sparse and only labelled at certain frames (such as the AVA dataset). We quantitatively measure this ability of our model on UCF101-24 which has full tube annotations. We do so by subsampling labels from the training set, and evaluating the full tubes on the validation set. + +As shown in Tab. 4, we still obtain meaningful tube predictions, with a Video AP20 of 77.1, when using only a single frame of annotation from each UCF video clip. When retaining 1 frame of supervision for every 24 labelled frames (which is roughly 1fps and corresponds to the AVA dataset's annotations), we observe minimal deterioration with respect to the fully supervised model (all Video AP metrics are within 0.7 points). Retaining 1 frame of annotation for every 12 consecutive labelled frames also performs similarly to using all frames in the video clip. These results suggest that due to the redundancy in the data (motion between frames is often limited), and the inductive bias of our model, we do not require each frame in the tube to be labelled in order to predict accurate tubelets. + +Decoder design Tabs. 5 and 6 analyse the effect of the decoder depth and the type of attention in the decoder (de + +Table 5. Effect of decoder depth on performance on the AVA dataset. Performance saturates at $L = 6$ layers. + +
Layers (L)01369
mAP ↑23.424.626.226.526.7
+ +Table 6. Effect of the type of attention used in the decoder on AVA. Factorised attention is both more accurate and efficient (almost half of the GFLOPs per decoder layer). + +
Decoder attentionmAPGFLOPs
Full26.410.5
Factorised26.75.3
+ +Table 7. Increasing the image resolution on the AVA dataset leads to consistent accuracy improvements, primarily on small objects. APs, APm and API denote the AP at 0.5 IoU threshold on small, medium and large boxes respectively following the COCO protocol [33]. AVA videos have a median aspect ratio of 16:10, and we pad the larger side when the aspect ratio is different. + +
ResolutionmAPAPsAPmAPI
140 × 22425.47.211.227.8
160 × 25626.711.512.528.7
220 × 35228.812.015.130.7
260 × 41629.413.315.831.0
320 × 51230.017.516.032.0
+ +Table 8. Comparison of pretraining for our models with ViViT-B and ViViT-L backbones on AVA using a resolution of $160 \times 256$ . Larger models benefit more from additional initial pretraining. + +
PretrainSTAR/BSTAR/L
IN21K [10] → K400 [23]26.727.0
IN21K [10] → K700 [7]27.327.6
CLIP [42] → K700 [7]30.336.2
+ +scribed in Sec. 3.2). As seen in Tab. 5, detection accuracy on AVA increases with the number of decoder layers, plateauing at around 6 layers. It is possible to use no decoder layers too: In this case, instead of learning queries $\mathbf{q}$ (Sec. 3.2), we simply interpret the outputs of the vision encoder (Sec. 3.1), $\mathbf{x}$ , as our queries and apply the localisation and classification heads directly upon them. Using decoder layers, however, can provide a performance increase of up to $3.3\mathrm{mAP}$ points ( $14\%$ relative), emphasising their utility. + +Tab. 6 shows that factorised attention in the decoder is more accurate than standard, "full" attention between all queries and visual features. Moreover, it is more efficient too, using almost half of the GFLOPs at each decoder layer. + +Effect of resolution and pretraining Scaling up the image resolution is critical to achieving high performance for object detection in images [20, 46]. However, we are not aware of previous works studying this for video action localisation. Tab. 7 shows that we do indeed observe substantial improvements from higher resolution, improving by up to 4.6 points on AVA. As expected, higher resolutions help more for detection at small sizes, where we follow the COCO [33] convention of object sizes. Note that AVA videos have a median aspect ratio of 16:10, and we pad the larger side for videos with different aspect ratios. + +Similarly, Tab. 8 shows the effect of different pretraining + +datasets. Video vision transformers are typically pretrained on an image dataset (like ImageNet-21K [10]), before being finetuned on a video dataset, such as Kinetics [23]. We find that the initial image checkpoint plays an important role, with CLIP [42] pretraining significantly outperforming supervised pretraining on ImageNet-21K [12, 50]. This improvement grows further when using a "Large" backbone. Our observations are consistent with prior works which have shown that CLIP-pretraining outperforms ImageNet-21K and even JFT pretraining [24, 30, 34, 41]. + +Qualitative examples We include example result videos of our proposed model in the supplementary. + +# 4.3. Comparison to state-of-the-art + +We compare our model to the state-of-the-art on datasets with both sparsely annotated keyframes (AVA and AVA-K), and full tubes (UCF101-24, JHMDB and MultiSports). + +AVA and AVA-Kinetics Tab. 9 compares to prior work on AVA and AVA-Kinetics. The best previous methods relied on external proposals [3, 54, 60] and external memory banks [40, 62] which we outperform. There are fewer end-to-end approaches, and we outperform these by an even larger margin. Note that though TubeR [66] is a proposal-free approach, their best results are actually obtained with the external memory of [61]. Consequently, we have reported the end-to-end, and external-memory versions of TubeR ("TubeR + LTC") separately in Tab. 9. Furthermore, as detailed in the supplement, the public TubeR training code produces significantly lower performance (20.0 mAP). Hence, to compare to it, we report results with our reimplementation in addition to the paper's results. Observe that we outperform TubeR using the same CSN-152 backbone, and then improve further using larger transformer backbones. + +We achieve greater relative improvements on AVA-Kinetics, showing that our end-to-end approach can leverage larger datasets more effectively. To our knowledge, we surpass the best previous results on AVA-Kinetics, achieving a Frame AP of 45.1. Notably, we outperform InternVideo [60] and VideoMAE-v2 [57], which are two recent video foundation models using more powerful backbones and larger, private, web-scale video datasets. Note that InternVideo consists of two different encoders, one of which is also initialised from CLIP. And like [60], we achieve our best AVA results by training a model on AVA-Kinetics, and then evaluating it only on the AVA validation set. Finally, we do not use any test-time augmentation, unlike previous works that ensemble results over multiple resolutions and/or left/right flips as shown by the "Views" column. + +UCF101-24 Tab. 10 shows that we outperform prior work on UCF101-24, both in terms of frame-level (Frame AP), and tube-level metrics (Video AP). We achieve state-of-the-art results with a CSN-152 backbone, and improve further + +Table 9. Comparison to the state-of-the-art (reported with mean Average Precision; mAP ↑) on AVA v2.2 [18] and AVA-Kinetics (AVA-K) [28]. Methods using external proposals (i.e. not end-to-end) are also trained on additional object detection and human pose data. Unless otherwise stated, separate models are trained for AVA and AVA-Kinetics. * denotes the model was trained on AVA-Kinetics and evaluated on AVA. “Res.” denotes the frame resolution of the shorter side. Web-scale foundational models are denoted in grey. † We also report our reimplemented results for TubeR, as the authors’ public code does not reproduce their reported results, as detailed in the supplement. + +
PretrainingViewsAVAAVA-KRes.BackboneEnd-to-end
MViT-B [13]K400127.3--MViTX
Unified [2]K400627.7-320SlowFastX
AIA [53]K7001832.3-320SlowFastX
ACAR [40]K700633.336.4320SlowFastX
TubeR [66] with LTC [61]IG65M [37]→K400, COCO233.6-256CSN-152X
MeMViT [62]K700-34.4-312MViT v2X
Co-finetuning [3]IN21K→K700, MiT, SSv2132.833.1320ViViT/LX
JFT,WTS→K700, MiT, SSv2136.136.2320ViViT/LX
VideoMAE [54]SSL K700 → Sup. K700.-39.3--ViViT/LX
InternVideo* [60]7 different datasets-41.042.5-Uniformer v2X
VideoMAE v2 [57]6 different datasets-42.643.9-ViViT/gX
Action Transformer [28]K4001-23.0400I3D✓
WOO [9]K600128.3-320SlowFast✓
TubeR [66]IG65M [37]→K400, COCO131.1-256CSN-152✓
TubeR reimplemented†IG65M [37]→K400129.533.6256CSN-152✓
STAR/CSN-152 (ours)IG65M→K400131.435.8256CSN-152✓
STAR/B (ours)IN21K→K400130.036.6320ViViT/B✓
CLIP→K700133.939.1320ViViT/B✓
STAR/L (ours)CLIP→K700139.244.5320ViViT/L✓
STAR/L (ours)*CLIP→K700142.545.1420ViViT/L✓
+ +Table 10. State-of-the-art comparison on datasets with tubelet annotations, UCF101, JHMDB51 and MultiSports. + +
PretrainingUCF101-24JHMDB51-21MultiSportsBackbone
fAPvAP20vAP50vAP50:95fAPvAP20vAP50fAPvAP20vAP50
ACT [22]IN1K67.177.251.425.065.774.273.7---VGG
MOC [31]IN1K→COCO78.082.853.828.370.877.377.225.212.90.6DLA34
Unified [2]K60079.3---------SlowFast
WOO [9]K600----80.5-----SlowFast
TubeR [66]IG65M→K40083.283.358.428.9-87.482.3---CSN-152
HIT [14]K70084.888.874.3-83.889.788.133.327.88.8SlowFast
FBIL [8]K70086.086.969.1----40.830.39.9SlowFast
STAR/CSN-152IG65M→K40086.787.065.430.693.596.395.445.450.118.6CSN-152
STAR/BIN21K→K40087.388.268.631.786.989.588.246.351.625.6ViViT/B
STAR/LCLIP→K70090.389.873.435.892.193.192.659.362.036.2ViViT/L
+ +by scaling up to ViViT-Large, consistent with our results on AVA (Tab. 9). Moreover, note how we substantially outperform TubeR [66] using the same CSN-152 backbone. Our margin of improvement over TubeR grows from vAP20 (+3.7) to vAP50 (+7.0) with the same backbone, showing that our tubelets are more precise, and is in line with our visual observations in the supplement. To our knowledge, we outperform the best previous reported Video AP50 result by 13.2 points. Note that as UCF videos are up to 900 frames, and as our network processes $T = 32$ frames, we follow prior works and link together tubelets using the same causal algorithm as [2, 22, 31, 47] for fair comparison. + +JHMDB51-21 Tab. 10 shows that we surpass the state-of-the-art on JHMDB. Once again, we significantly outperform TubeR [66], by 13.1 vAP50, with the same CSN-152 backbone. The CSN-152 backbone outperforms ViViT in this case, possibly because this is the smallest dataset and larger backbones can overfit more easily. The videos in this dataset are trimmed (meaning that labelled actions are being performed on each frame), and also shorter. Hence the + +Video AP is not as strict as it is on UCF. As videos are at most 40 frames, we set $T = 40$ in our model so that we process the entire clip at once without needing to link tubelets. + +MultiSports Finally, Tab. 10 shows that we improve on prior state-of-the-art on MultiSports substantially on both Frame AP and Video AP. Although this is a recent benchmark with few methods to compare against, our improvement is significant due to the challenging nature of the dataset: it contains fast motion, many simultaneous actors and large variation in action duration. + +# 5. Conclusion and Future Work + +We have presented STAR, an end-to-end spatio-temporal action localisation model that can output tubelets, when either sparse keyframe, or full tubelet annotation is available. Our approach achieves state-of-the-art results on four action localisation datasets for both frame-level and tubelet-level predictions (in particular, we obtain $45.1\%$ mAP on the challenging AVA-Kinetics dataset), outperforming complex methods that use external proposals and memory banks. Future work is to extend our method to open vocabularies. + +# References + +[1] Anurag Arnab, Mostafa Dehghani, Georg Heigold, Chen Sun, Mario Lucic, and Cordelia Schmid. ViViT: A Video Vision Transformer. In ICCV, 2021. 5 +[2] Anurag Arnab, Chen Sun, and Cordelia Schmid. Unified graph structured models for video understanding. In ICCV, 2021. 1, 2, 8 +[3] Anurag Arnab, Xuehan Xiong, Alexey Gritsenko, Rob Romijnders, Josip Djolonga, Mostafa Dehghani, Chen Sun, Mario Lucic, and Cordelia Schmid. Beyond transfer learning: Co-finetuning for action localisation. In arXiv preprint arXiv:2207.03807, 2022. 1, 2, 6, 7, 8 +[4] Jimmy Lei Ba, Jamie Ryan Kiros, and Geoffrey E Hinton. Layer normalization. In arXiv preprint arXiv:1607.06450, 2016. 4 +[5] Fabien Baradel, Natalia Neverova, Christian Wolf, Julien Mille, and Greg Mori. Object level visual reasoning in videos. In ECCV, 2018. 2 +[6] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In ECCV, 2020. 1, 2, 3, 4 +[7] Joao Carreira, Eric Noland, Chloe Hillier, and Andrew Zisserman. A short note on the kinetics-700 human action dataset. In arXiv preprint arXiv:1907.06987, 2019. 5, 7 +[8] Keke Chen, Xiangbo Shu, Guo-Sen Xie, Rui Yan, and Jinhui Tang. Foreground/background-masked interaction learning for spatio-temporal action detection. In ACM MM, 2023. 8 +[9] Shoufa Chen, Peize Sun, Enze Xie, Chongjian Ge, Jiannan Wu, Lan Ma, Jiajun Shen, and Ping Luo. Watch only once: An end-to-end video action detection framework. In ICCV, 2021. 1, 2, 4, 8 +[10] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. ImageNet: A large-scale hierarchical image database. In CVPR, 2009. 5, 7 +[11] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. BERT: Pre-training of deep bidirectional transformers for language understanding. In NAACL, 2019. 5 +[12] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2021. 5, 7 +[13] Haoqi Fan, Bo Xiong, Karttikeya Mangalam, Yanghao Li, Zhicheng Yan, Jitendra Malik, and Christoph Feichtenhofer. Multiscale vision transformers. In ICCV, 2021. 6, 8 +[14] Gueter Josmy Faure, Min-Hung Chen, and Shang-Hong Lai. Holistic interaction transformer network for action detection. In WACV, 2023. 8 +[15] Christoph Feichtenhofer, Haoqi Fan, Jitendra Malik, and Kaiming He. Slowfast networks for video recognition. In ICCV, 2019. 1, 2, 6 +[16] Rohit Girdhar, Joao Carreira, Carl Doersch, and Andrew Zisserman. Video action transformer network. In CVPR, 2019. 2 + +[17] Ross Girshick. Fast r-cnn. In ICCV, 2015. 1, 2 +[18] Chunhui Gu, Chen Sun, David A Ross, Carl Vondrick, Caroline Pantofaru, Yeqing Li, Sudheendra Vijayanarasimhan, George Toderici, Susanna Ricco, Rahul Sukthankar, et al. Ava: A video dataset of spatio-temporally localized atomic visual actions. In CVPR, 2018. 1, 2, 4, 5, 8 +[19] Dan Hendrycks and Kevin Gimpel. Gaussian error linear units (gelus). In arXiv preprint arXiv:1606.08415, 2016. 4 +[20] Jonathan Huang, Vivek Rathod, Chen Sun, Menglong Zhu, Anoop Korattikara, Alireza Fathi, Ian Fischer, Zbigniew Wojna, Yang Song, Sergio Guadarrama, et al. Speed/accuracy trade-offs for modern convolutional object detectors. In CVPR, 2017. 7 +[21] H. Jhuang, J. Gall, S. Zuffi, C. Schmid, and M. J. Black. Towards understanding action recognition. In ICCV, 2013. 1, 2, 5 +[22] Vicky Kalogeiton, Philippe Weinzaepfel, Vittorio Ferrari, and Cordelia Schmid. Action tubelet detector for spatiotemporal action localization. In ICCV, 2017. 1, 2, 5, 8 +[23] Will Kay, Joao Carreira, Karen Simonyan, Brian Zhang, Chloe Hillier, Sudheendra Vijayanarasimhan, Fabio Viola, Tim Green, Trevor Back, Paul Natsev, et al. The kinetics human action video dataset. In arXiv preprint arXiv:1705.06950, 2017. 5, 7 +[24] Skanda Koppula, Yazhe Li, Evan Shelhamer, Andrew Jaegle, Nikhil Parthasarathy, Relja Arandjelovic, João Carreira, and Olivier Henaff. Where should i spend my flops? efficiency evaluations of visual pre-training methods. In arXiv preprint arXiv:2209.15589, 2022. 7 +[25] Okan Köpukli, Xiangyu Wei, and Gerhard Rigoll. You only watch once: A unified cnn architecture for real-time spatiotemporal action localization. In arXiv preprint arXiv:1911.06644, 2019. 1, 2 +[26] Harold W Kuhn. The hungarian method for the assignment problem. Naval research logistics quarterly, 2(1-2):83-97, 1955. 4 +[27] Jie Lei, Tamara L Berg, and Mohit Bansal. Detecting moments and highlights in videos via natural language queries. In NeurIPS, 2021. 2 +[28] Ang Li, Meghanata Thotakuri, David A Ross, João Carreira, Alexander Vostrikov, and Andrew Zisserman. The a-akinetics localized human actions video dataset. In arXiv preprint arXiv:2005.00214, 2020. 1, 2, 5, 8 +[29] Dong Li, Zhaofan Qiu, Qi Dai, Ting Yao, and Tao Mei. Recurrent tubelet proposal and recognition networks for action detection. In ECCV, 2018. 2 +[30] Kunchang Li, Yali Wang, Yinan He, Yizhuo Li, Yi Wang, Limin Wang, and Yu Qiao. Uniformerv2: Spatiotemporal learning by arming image vits with video uniformer. In arXiv preprint arXiv:2211.09552, 2022. 7 +[31] Yixuan Li, Zixu Wang, Limin Wang, and Gangshan Wu. Actions as moving points. In ECCV, 2020. 2, 5, 8 +[32] Yixuan Li, Lei Chen, Runyu He, Zhenzhi Wang, Gangshan Wu, and Limin Wang. Multisports: A multi-person video dataset of spatio-temporally localized sports actions. In ICCV, 2021. 2, 5 + +[33] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In ECCV, 2014. 7 +[34] Ziyi Lin, Shijie Geng, Renrui Zhang, Peng Gao, Gerard de Melo, Xiaogang Wang, Jifeng Dai, Yu Qiao, and Hongsheng Li. Frozen clip models are efficient video learners. In ECCV, 2022. 7 +[35] Wei Liu, Dragomir Anguelov, Dumitru Erhan, Christian Szegedy, Scott Reed, Cheng-Yang Fu, and Alexander C Berg. Ssd: Single shot multibox detector. In ECCV, 2016. 2 +[36] Xiaolong Liu, Qimeng Wang, Yao Hu, Xu Tang, Shiwei Zhang, Song Bai, and Xiang Bai. End-to-end temporal action detection with transformer. IEEE Transactions on Image Processing, 31:5427-5441, 2022. 2 +[37] Dhruv Mahajan, Ross Girshick, Vignesh Ramanathan, Kaiming He, Manohar Paluri, Yixuan Li, Ashwin Bharambe, and Laurens Van Der Maaten. Exploring the limits of weakly supervised pretraining. In ECCV, 2018. 5, 8 +[38] Matthias Minderer, Alexey Gritsenko, Austin Stone, Maxim Neumann, Dirk Weissenborn, Alexey Dosovitskiy, Aravindh Mahendran, Anurag Arnab, Mostafa Dehghani, Zhuoran Shen, et al. Simple Open-Vocabulary Object Detection with Vision Transformers. In ECCV, 2022. 4 +[39] Megha Nawhal and Greg Mori. Activity graph transformer for temporal action localization. In arXiv preprint arXiv:2101.08540, 2021. 2 +[40] Junting Pan, Siyu Chen, Mike Zheng Shou, Yu Liu, Jing Shao, and Hongsheng Li. Actor-context-actor relation network for spatio-temporal action localization. In CVPR, 2021. 1, 2, 7, 8 +[41] Junting Pan, Ziyi Lin, Xiatian Zhu, Jing Shao, and Hongsheng Li. St-adapter: Parameter-efficient image-to-video transfer learning. In NeurIPS, 2022. 7 +[42] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, 2021. 7 +[43] Joseph Redmon, Santosh Divvala, Ross Girshick, and Ali Farhadi. You only look once: Unified, real-time object detection. In CVPR, 2016. 2 +[44] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. In NeurIPS, 2015. 3, 6 +[45] Hamid Rezatofighi, Nathan Tsoi, JunYoung Gwak, Amir Sadeghian, Ian Reid, and Silvio Savarese. Generalized intersection over union: A metric and a loss for bounding box regression. In CVPR, 2019. 4 +[46] Bharat Singh and Larry S Davis. An analysis of scale invariance in object detection snip. In CVPR, 2018. 7 +[47] Gurkirt Singh, Suman Saha, Michael Sapienza, Philip HS Torr, and Fabio Cuzzolin. Online real-time multiple spatiotemporal action localisation and prediction. In ICCV, 2017. 2, 5, 8 +[48] Lin Song, Shiwei Zhang, Gang Yu, and Hongbin Sun. Tacnet: Transition-aware context network for spatio-temporal action detection. In CVPR, 2019. 2 + +[49] Khurram Soomro, Amir Roshan Zamir, and Mubarak Shah. Ucf101: A dataset of 101 human actions classes from videos in the wild. In arXiv preprint arXiv:1212.0402, 2012. 1, 2, 5 +[50] Andreas Steiner, Alexander Kolesnikov, Xiaohua Zhai, Ross Wightman, Jakob Uszkoreit, and Lucas Beyer. How to train your vit? data, augmentation, and regularization in vision transformers. TMLR, 2022. 7 +[51] Chen Sun, Abhinav Shrivastava, Carl Vondrick, Kevin Murphy, Rahul Sukthankar, and Cordelia Schmid. Actor-centric relation network. In ECCV, 2018. 1, 2 +[52] Peize Sun, Rufeng Zhang, Yi Jiang, Tao Kong, Chenfeng Xu, Wei Zhan, Masayoshi Tomizuka, Lei Li, Zehuan Yuan, Changhu Wang, et al. Sparse r-cnn: End-to-end object detection with learnable proposals. In CVPR, 2021. 2 +[53] Jiajun Tang, Jin Xia, Xinzhi Mu, Bo Pang, and Cewu Lu. Asynchronous interaction aggregation for action detection. In ECCV, 2020. 1, 2, 8 +[54] Zhan Tong, Yibing Song, Jue Wang, and Limin Wang. VideoMAE: Masked autoencoders are data-efficient learners for self-supervised video pre-training. In NeurIPS, 2022. 2, 7, 8 +[55] Du Tran, Heng Wang, Lorenzo Torresani, and Matt Feiszli. Video classification with channel-separated convolutional networks. In ICCV, 2019. 5 +[56] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In NeurIPS, 2017. 3, 4 +[57] Limin Wang, Bingkun Huang, Zhiyu Zhao, Zhan Tong, Yi-nan He, Yi Wang, Yali Wang, and Yu Qiao. Videomae v2: Scaling video masked autoencoders with dual masking. In arXiv preprint arXiv:2303.16727, 2023. 2, 7, 8 +[58] Xiaolong Wang and Abhinav Gupta. Videos as space-time region graphs. In ECCV, 2018. 2 +[59] Yuqing Wang, Zhaoliang Xu, Xinlong Wang, Chunhua Shen, Baoshan Cheng, Hao Shen, and Huaxia Xia. End-to-end video instance segmentation with transformers. In CVPR, 2021. 2 +[60] Yi Wang, Kunchang Li, Yizhuo Li, Yinan He, Bingkun Huang, Zhiyu Zhao, Hongjie Zhang, Jilan Xu, Yi Liu, Zun Wang, et al. Internvideo: General video foundation models via generative and discriminative learning. In arXiv preprint arXiv:2212.03191, 2022. 2, 7, 8 +[61] Chao-Yuan Wu, Christoph Feichtenhofer, Haoqi Fan, Kaiming He, Philipp Krahenbuhl, and Ross Girshick. Long-term feature banks for detailed video understanding. In CVPR, 2019. 1, 2, 5, 6, 7, 8 +[62] Chao-Yuan Wu, Yanghao Li, Karttikeya Mangalam, Haoqi Fan, Bo Xiong, Jitendra Malik, and Christoph Feichtenhofer. MeMVit: Memory-augmented multiscale vision transformer for efficient long-term video recognition. In CVPR, 2022. 1, 2, 6, 7, 8 +[63] Chuhan Zhang, Ankush Gupta, and Andrew Zisserman. Temporal query networks for fine-grained video understanding. In CVPR, 2021. 2 +[64] Yubo Zhang, Pavel Tokmakov, Martial Hebert, and Cordelia Schmid. A structured model for action detection. In CVPR, 2019. 2 + +[65] Jiaojiao Zhao, Yanyi Zhang, Xinyu Li, Hao Chen, Bing Shuai, Mingze Xu, Chunhui Liu, Kaustav Kundu, Yuanjun Xiong, Davide Modolo, et al. Tuber public code. https://github.com/amazon-science/tubelet-transformer, 2022.6 +[66] Jiaojiao Zhao, Yanyi Zhang, Xinyu Li, Hao Chen, Bing Shuai, Mingze Xu, Chunhui Liu, Kaustav Kundu, Yuanjun Xiong, Davide Modolo, et al. TubeR: Tubelet Transformer for Video Action Detection. In CVPR, 2022. 1, 2, 4, 5, 7, 8 +[67] Xingyi Zhou, Dequan Wang, and Philipp Krahenbuhl. Objects as points. In arXiv preprint arXiv:1904.07850, 2019. 2 \ No newline at end of file diff --git a/2024/End-to-End Spatio-Temporal Action Localisation with Video Transformers/images.zip b/2024/End-to-End Spatio-Temporal Action Localisation with Video Transformers/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..ca8475f7cc5127982984b253d4796d81686fb283 --- /dev/null +++ b/2024/End-to-End Spatio-Temporal Action Localisation with Video Transformers/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b0483455629f75aff84b539875d4a2c7fe56108c19bca4dfbaf1d75f6c3c2fe2 +size 452736 diff --git a/2024/End-to-End Spatio-Temporal Action Localisation with Video Transformers/layout.json b/2024/End-to-End Spatio-Temporal Action Localisation with Video Transformers/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..861a70c28a81e7967b9b7a972aefdc8fad0d578a --- /dev/null +++ b/2024/End-to-End Spatio-Temporal Action Localisation with Video Transformers/layout.json @@ -0,0 +1,8605 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 66, + 103, + 529, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 103, + 529, + 121 + ], + "spans": [ + { + "bbox": [ + 66, + 103, + 529, + 121 + ], + "type": "text", + "content": "End-to-End Spatio-Temporal Action Localisation with Video Transformers" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 108, + 143, + 486, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 108, + 143, + 486, + 185 + ], + "spans": [ + { + "bbox": [ + 108, + 143, + 486, + 185 + ], + "type": "text", + "content": "Alexey A. Gritsenko Xuehan Xiong Josip Djolonga Mostafa Dehghani Chen Sun Mario Lučić Cordelia Schmid Anurag Arnab Google" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 208, + 186, + 378, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 208, + 186, + 378, + 200 + ], + "spans": [ + { + "bbox": [ + 208, + 186, + 378, + 200 + ], + "type": "text", + "content": "{agritsenko, aarnab}@google.com" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 143, + 216, + 192, + 228 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 216, + 192, + 228 + ], + "spans": [ + { + "bbox": [ + 143, + 216, + 192, + 228 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 241, + 290, + 432 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 241, + 290, + 432 + ], + "spans": [ + { + "bbox": [ + 46, + 241, + 290, + 432 + ], + "type": "text", + "content": "The most performant spatio-temporal action localisation models use external person proposals and complex external memory banks. We propose a fully end-to-end, purely-transformer based model that directly ingests an input video, and outputs tubelets - a sequence of bounding boxes and the action classes at each frame. Our flexible model can be trained with either sparse bounding-box supervision on individual frames, or full tubelet annotations. And in both cases, it predicts coherent tubelets as the output. Moreover, our end-to-end model requires no additional pre-processing in the form of proposals, or post-processing in terms of non-maximal suppression. We perform extensive ablation experiments, and significantly advance the state-of-the-art on five different spatio-temporal action localisation benchmarks with both sparse keyframes and full tubelet annotations." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 454, + 128, + 466 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 454, + 128, + 466 + ], + "spans": [ + { + "bbox": [ + 47, + 454, + 128, + 466 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 474, + 287, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 474, + 287, + 581 + ], + "spans": [ + { + "bbox": [ + 46, + 474, + 287, + 581 + ], + "type": "text", + "content": "Spatio-temporal action localisation is an important problem with applications in advanced video search engines, robotics and security among others. It is typically formulated in one of two ways: Firstly, predicting the bounding boxes and actions performed by an actor at a single keyframe given neighbouring frames as spatio-temporal context [18, 28]. Or alternatively, predicting a sequence of bounding boxes and actions (i.e. \"tubes\"), for each actor at each frame in the video [21, 49]." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 582, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 582, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 582, + 288, + 714 + ], + "type": "text", + "content": "The most performant models [3, 15, 40, 62], particularly for the first, keyframe-based formulation of the problem, employ a two-stage pipeline inspired by the Fast-RCNN object detector [17]: They first run a separate person detector to obtain proposals. Features from these proposals are then aggregated and classified according to the actions of interest. These models have also been supplemented with memory banks containing long-term contextual information from other frames [40, 53, 61, 62], and/or detections of other potentially relevant objects [2, 53] to capture additional scene context, achieving state-of-the-art results." + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 320, + 211, + 532, + 404 + ], + "blocks": [ + { + "bbox": [ + 320, + 211, + 532, + 404 + ], + "lines": [ + { + "bbox": [ + 320, + 211, + 532, + 404 + ], + "spans": [ + { + "bbox": [ + 320, + 211, + 532, + 404 + ], + "type": "image", + "image_path": "248495399aef562cd7b0c1d860b436c8af9430ed0e3230775834ee9027aa8a88.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 407, + 547, + 486 + ], + "lines": [ + { + "bbox": [ + 304, + 407, + 547, + 486 + ], + "spans": [ + { + "bbox": [ + 304, + 407, + 547, + 486 + ], + "type": "text", + "content": "Figure 1. We propose an end-to-end Spatio-Temporal Action Recognition model named STAR. Our model is end-to-end in that it does not require any external region proposals to predict tubelets – sequences of bounding boxes associated with a given person in every frame and their corresponding action classes. Our model can be trained with either sparse box annotations on selected keyframes, or full tubelet supervision." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 495, + 547, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 495, + 547, + 579 + ], + "spans": [ + { + "bbox": [ + 304, + 495, + 547, + 579 + ], + "type": "text", + "content": "And whilst proposal-free algorithms, which do not require external person detectors, have been developed for detecting both at the keyframe-level [9, 25, 51] and tubelet-level [22, 66], their performance has typically lagged behind their proposal-based counterparts. Here, we show for the first time that an end-to-end trainable spatio-temporal model outperforms a two-stage approach." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 582, + 548, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 582, + 548, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 582, + 548, + 715 + ], + "type": "text", + "content": "As shown in Fig. 1, we propose our Spatio-Temporal Action TransformeR (STAR) that consists of a transformer architecture, and is based on the DETR [6] detection model. Our model is \"end-to-end\" in that it does not require preprocessing in the form of proposals, nor post-processing in the form of non-maximal suppression (NMS) in contrast to the majority of prior work. The initial stage of the model is a vision encoder. This is followed by a decoder that processes learned latent queries, which represent each actor in the video, into output tubelets - a sequence of bounding boxes and action classes at each time step of" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "18373" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 179 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 179 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 179 + ], + "type": "text", + "content": "the input video clip. Our model is versatile in that we can train it with either fully-labeled tube annotations, or with sparse keyframe annotations (when only a limited number of keyframes are labelled). In the latter case, our network still predicts tubelets, and learns to associate detections of an actor, from one frame to the next, without explicit supervision. This behaviour is facilitated by our formulation of factorised queries, decoder architecture and tubelet matching in the loss which all contain temporal inductive biases." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 180, + 288, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 180, + 288, + 443 + ], + "spans": [ + { + "bbox": [ + 46, + 180, + 288, + 443 + ], + "type": "text", + "content": "We conduct thorough ablation studies of these modelling choices, confirming the benefit of temporal inductive biases in our model design. Informed by these experiments, we achieve state-of-the-art on both keyframe-based action localisation datasets like AVA [18] and AVA-Kinetics [28], and also tubelet-based datasets like UCF101-24 [49], JHMDB [21] and MultiSports [32]. In particular, we achieve a Frame mAP of 45.1 on AVA-Kinetics, outperforming the best previous results achieved by a massive video foundation model [57]. Moreover, our state-of-the-art results are achieved with a single forward-pass through the model, using only a video clip as input, and without any separate external person detectors providing proposals [57, 60, 62], complex memory banks [40, 62, 66], or additional object detectors [2, 53], as used by the prior works. Furthermore, we outperform these complex, prior, state-of-the-art two-stage models whilst also having additional functionality in that our model predicts tubelets, that is, temporally consistent bounding boxes at each frame of the input video clip. This capability is demonstrated by our results on tube-based datasets like UCF101-24 where we surpass the prior work [66] by 13.2 points on Video AP50." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 455, + 134, + 467 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 455, + 134, + 467 + ], + "spans": [ + { + "bbox": [ + 47, + 455, + 134, + 467 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 473, + 286, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 473, + 286, + 664 + ], + "spans": [ + { + "bbox": [ + 46, + 473, + 286, + 664 + ], + "type": "text", + "content": "Models for spatio-temporal action localisation have typically built upon advances in object detectors for images. The most performant methods for action localisation [3, 15, 40, 53, 62] are based on \"two-stage\" detectors like Fast-RCNN [17]. These models use external, pre-computed person detections, and use them to ROI-pool features which are then classified into action classes. Although these models are cumbersome in that they require an additional model and backbone to first detect people, and therefore additional detection training data as well, they are currently the leading approaches on datasets such as AVA [18]. Such models using external proposals are also particularly suited to datasets such as AVA [18] as each person is exhaustively labelled as performing an action, and therefore there are fewer false-positives from using action-agnostic person detections compared to datasets such as UCF101 [49]." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 665, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 287, + 713 + ], + "type": "text", + "content": "The performance of these two-stage models has further been improved by incorporating more contextual information using feature banks extracted from additional frames in the video [40, 53, 61, 62] or by utilising detections of ad" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 72, + 545, + 131 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 131 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 131 + ], + "type": "text", + "content": "ditional objects in the scene [2, 5, 58, 64]. Both of these cases entail significant extra computation and complexity to train additional auxiliary models, and to precompute features from them that are then used during training and inference of the localisation model." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 133, + 545, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 133, + 545, + 251 + ], + "spans": [ + { + "bbox": [ + 304, + 133, + 545, + 251 + ], + "type": "text", + "content": "Our proposed method, in contrast, is end-to-end in that it directly produces detections without any additional inputs besides a video clip. Moreover, it outperforms these prior works without resorting to external proposals or memory banks, showing that a transformer backbone is sufficient to capture long-range dependencies in the input video. In addition, unlike previous two-stage methods, our method directly predicts tubelets: a sequence of bounding boxes and actions for each frame of the input video, and can do so even when we do not have full tubelet annotations available." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 252, + 545, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 252, + 545, + 384 + ], + "spans": [ + { + "bbox": [ + 304, + 252, + 545, + 384 + ], + "type": "text", + "content": "A number of proposal-free action localisation models have also been developed [9, 16, 22, 25, 51, 66]. These methods are based upon alternative object detection architectures such as SSD [35], CentreNet [67], YOLO [43], DETR [6] and Sparse-RCNN [52]. However, in contrast to our approach, they have been outperformed by their proposal-based counterparts. Moreover, some of these methods [16, 25, 51] also consist of separate network backbones for learning video feature representations and proposals for a keyframe, and are thus effectively two networks trained jointly, and cannot predict tubelets either." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 384, + 545, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 384, + 545, + 515 + ], + "spans": [ + { + "bbox": [ + 304, + 384, + 545, + 515 + ], + "type": "text", + "content": "Among prior works that do not use external proposals, and also directly predict tubelets [22, 29, 31, 47, 48], our work is most similar to TubeR [66] given that our model is also based on DETR. Our model, however, is designed with additional temporal inductive biases which improves accuracy (without using external memory banks precomputed offline like [66]. And moreover, unlike TubeR, we also demonstrate how our model can predict tubelets (i.e. predictions at every frame of the input video), even when we only have sparse keyframe supervision (i.e. ground truth annotation for a limited number of frames) available." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 516, + 545, + 564 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 516, + 545, + 564 + ], + "spans": [ + { + "bbox": [ + 304, + 516, + 545, + 564 + ], + "type": "text", + "content": "Finally, we note that DETR has also been extended as a proposal-free method to addressing other localisation tasks in video such as instance segmentation [59], temporal localisation [36, 39, 63] and moment retrieval [27]." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 574, + 512, + 588 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 574, + 512, + 588 + ], + "spans": [ + { + "bbox": [ + 305, + 574, + 512, + 588 + ], + "type": "text", + "content": "3. Spatio-Temporal Action Transformer" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 594, + 545, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 545, + 640 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 545, + 640 + ], + "type": "text", + "content": "Our proposed model ingests a sequence of video frames, and directly predicts tubelets (a sequence of bounding boxes and action labels). No external person detections [40, 54, 57], or memory banks [62, 66], are needed." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "type": "text", + "content": "As summarised in Fig. 2, our model consists of a vision encoder (Sec. 3.1), followed by a decoder which processes learned query tokens into output tubelets (Sec. 3.2). We incorporate temporal inductive biases into our decoder to improve accuracy and tubelet prediction with weaker supervision. Our model is inspired by the DETR architecture [6]" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "18374" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 48, + 58, + 545, + 165 + ], + "blocks": [ + { + "bbox": [ + 48, + 58, + 545, + 165 + ], + "lines": [ + { + "bbox": [ + 48, + 58, + 545, + 165 + ], + "spans": [ + { + "bbox": [ + 48, + 58, + 545, + 165 + ], + "type": "image", + "image_path": "f91408e60bee52cb17a7ef82c064054ea15cf6fb90c4c239741b9f8ca8c0ebfd.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 167, + 547, + 200 + ], + "lines": [ + { + "bbox": [ + 46, + 167, + 547, + 200 + ], + "spans": [ + { + "bbox": [ + 46, + 167, + 547, + 200 + ], + "type": "text", + "content": "Figure 2. Our model processes a fixed-length video clip, and for each frame, outputs tubelets (i.e. linked bounding boxes with associated action class probabilities). It consists of vision encoder which outputs a video representation, " + }, + { + "bbox": [ + 46, + 167, + 547, + 200 + ], + "type": "inline_equation", + "content": "\\mathbf{x} \\in \\mathbb{R}^{T \\times h \\times w \\times d}" + }, + { + "bbox": [ + 46, + 167, + 547, + 200 + ], + "type": "text", + "content": ". The video representation, along with learned queries, " + }, + { + "bbox": [ + 46, + 167, + 547, + 200 + ], + "type": "inline_equation", + "content": "\\mathbf{q}" + }, + { + "bbox": [ + 46, + 167, + 547, + 200 + ], + "type": "text", + "content": " (which are factorised into spatial " + }, + { + "bbox": [ + 46, + 167, + 547, + 200 + ], + "type": "inline_equation", + "content": "\\mathbf{q}^s" + }, + { + "bbox": [ + 46, + 167, + 547, + 200 + ], + "type": "text", + "content": " and temporal components " + }, + { + "bbox": [ + 46, + 167, + 547, + 200 + ], + "type": "inline_equation", + "content": "\\mathbf{q}^t" + }, + { + "bbox": [ + 46, + 167, + 547, + 200 + ], + "type": "text", + "content": ") are decoded into tubelets by a decoder of" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 200, + 269, + 210 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 200, + 269, + 210 + ], + "spans": [ + { + "bbox": [ + 47, + 200, + 269, + 210 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 47, + 200, + 269, + 210 + ], + "type": "text", + "content": " layers followed by shallow box and class prediction heads." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 220, + 287, + 269 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 220, + 287, + 269 + ], + "spans": [ + { + "bbox": [ + 46, + 220, + 287, + 269 + ], + "type": "text", + "content": "for object detection in images, and is also trained with a set-based loss and Hungarian matching. We detail our loss, and how we can train with either sparse keyframe supervision or full tubelet supervision, in Sec. 3.3." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 278, + 141, + 290 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 278, + 141, + 290 + ], + "spans": [ + { + "bbox": [ + 47, + 278, + 141, + 290 + ], + "type": "text", + "content": "3.1. Vision Encoder" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 298, + 289, + 478 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 298, + 289, + 478 + ], + "spans": [ + { + "bbox": [ + 46, + 298, + 289, + 478 + ], + "type": "text", + "content": "The vision backbone processes an input video, " + }, + { + "bbox": [ + 46, + 298, + 289, + 478 + ], + "type": "inline_equation", + "content": "\\mathbf{X} \\in \\mathbb{R}^{T \\times H \\times W \\times 3}" + }, + { + "bbox": [ + 46, + 298, + 289, + 478 + ], + "type": "text", + "content": " to produce a feature representation of the input video " + }, + { + "bbox": [ + 46, + 298, + 289, + 478 + ], + "type": "inline_equation", + "content": "\\mathbf{x} \\in \\mathbb{R}^{t \\times h \\times w \\times d}" + }, + { + "bbox": [ + 46, + 298, + 289, + 478 + ], + "type": "text", + "content": ". Here, " + }, + { + "bbox": [ + 46, + 298, + 289, + 478 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 46, + 298, + 289, + 478 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 298, + 289, + 478 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 46, + 298, + 289, + 478 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 298, + 289, + 478 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 46, + 298, + 289, + 478 + ], + "type": "text", + "content": " are the original temporal-, height- and width-dimensions of the input video respectively, whilst " + }, + { + "bbox": [ + 46, + 298, + 289, + 478 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 298, + 289, + 478 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 298, + 289, + 478 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 46, + 298, + 289, + 478 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 298, + 289, + 478 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 46, + 298, + 289, + 478 + ], + "type": "text", + "content": " are the spatio-temporal dimensions of their feature representation, and " + }, + { + "bbox": [ + 46, + 298, + 289, + 478 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 46, + 298, + 289, + 478 + ], + "type": "text", + "content": " its latent dimension. When using a transformer backbone, these spatiotemporal dimensions depend on the patch size when tokenising the input, and when using a convolutional backbone, they depend on the overall stride. To retain spatiotemporal information, we remove the spatial- and temporal-aggregation steps at the end of the original backbone. And if the temporal patch size (or stride) is larger than 1, we bilinearly upsample the final feature map along the temporal axis to maintain the original temporal resolution." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 488, + 146, + 500 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 488, + 146, + 500 + ], + "spans": [ + { + "bbox": [ + 47, + 488, + 146, + 500 + ], + "type": "text", + "content": "3.2. Tubelet Decoder" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 506, + 287, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 506, + 287, + 592 + ], + "spans": [ + { + "bbox": [ + 46, + 506, + 287, + 592 + ], + "type": "text", + "content": "Our decoder processes the visual features, " + }, + { + "bbox": [ + 46, + 506, + 287, + 592 + ], + "type": "inline_equation", + "content": "\\mathbf{x} \\in \\mathbb{R}^{T \\times h \\times w \\times c}" + }, + { + "bbox": [ + 46, + 506, + 287, + 592 + ], + "type": "text", + "content": ", along with learned queries, " + }, + { + "bbox": [ + 46, + 506, + 287, + 592 + ], + "type": "inline_equation", + "content": "\\mathbf{q} \\in \\mathbb{R}^{T \\times S \\times d}" + }, + { + "bbox": [ + 46, + 506, + 287, + 592 + ], + "type": "text", + "content": ", to output tubelets, " + }, + { + "bbox": [ + 46, + 506, + 287, + 592 + ], + "type": "inline_equation", + "content": "\\mathbf{y} = (\\mathbf{b}, \\mathbf{a})" + }, + { + "bbox": [ + 46, + 506, + 287, + 592 + ], + "type": "text", + "content": " which are a sequence of bounding boxes, " + }, + { + "bbox": [ + 46, + 506, + 287, + 592 + ], + "type": "inline_equation", + "content": "b \\in \\mathbb{R}^{T \\times S \\times 4}" + }, + { + "bbox": [ + 46, + 506, + 287, + 592 + ], + "type": "text", + "content": " and corresponding actions, " + }, + { + "bbox": [ + 46, + 506, + 287, + 592 + ], + "type": "inline_equation", + "content": "a \\in \\mathbb{R}^{T \\times S \\times C}" + }, + { + "bbox": [ + 46, + 506, + 287, + 592 + ], + "type": "text", + "content": ". Here, " + }, + { + "bbox": [ + 46, + 506, + 287, + 592 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 46, + 506, + 287, + 592 + ], + "type": "text", + "content": " denotes the maximum number of bounding boxes per frame (padded with \"background\" as necessary) and " + }, + { + "bbox": [ + 46, + 506, + 287, + 592 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 46, + 506, + 287, + 592 + ], + "type": "text", + "content": " denotes the number of output classes." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 593, + 287, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 593, + 287, + 665 + ], + "spans": [ + { + "bbox": [ + 46, + 593, + 287, + 665 + ], + "type": "text", + "content": "The idea of decoding learned queries into output detections using the transformer decoder architecture of Vaswani et al. [56] was used in DETR [6]. In summary, the decoder of [6, 56] consists of " + }, + { + "bbox": [ + 46, + 593, + 287, + 665 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 46, + 593, + 287, + 665 + ], + "type": "text", + "content": " layers, each performing a series of self-attention operations on the queries, and cross-attention between the queries and encoder outputs." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "text", + "content": "We modify the queries, self-attention and cross-attention operations for our spatio-temporal localisation scenario, as shown in Fig. 2 and 3 to include additional temporal inductive biases, and to improve accuracy as detailed below." + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 307, + 204, + 545, + 340 + ], + "blocks": [ + { + "bbox": [ + 307, + 204, + 545, + 340 + ], + "lines": [ + { + "bbox": [ + 307, + 204, + 545, + 340 + ], + "spans": [ + { + "bbox": [ + 307, + 204, + 545, + 340 + ], + "type": "image", + "image_path": "ab85d5469b56bb6f27c3dafa2a47896ec4d5bb7af2060b6907190866c9781041.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 348, + 547, + 426 + ], + "lines": [ + { + "bbox": [ + 304, + 348, + 547, + 426 + ], + "spans": [ + { + "bbox": [ + 304, + 348, + 547, + 426 + ], + "type": "text", + "content": "Figure 3. Our decoder layer consists of factorised self-attention (SA) (left) and cross-attention (CA) (right) operations designed to provide a spatio-temporal inductive bias and reduce computation. Both operations restrict attention to the same spatial and temporal slices as the query token, as illustrated by the receptive field (blue) for a given query token (magenta). Factorised SA consists of two operations, whilst in Factorised CA, there is one operation." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 432, + 545, + 480 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 432, + 545, + 480 + ], + "spans": [ + { + "bbox": [ + 304, + 432, + 545, + 480 + ], + "type": "text", + "content": "Queries Queries, q, in DETR, are decoded using the encoded visual features, x, into bounding box predictions, and are analogous to the \"anchors\" used in other detection architectures such as Faster-RCNN [44]." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 482, + 545, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 482, + 545, + 518 + ], + "spans": [ + { + "bbox": [ + 304, + 482, + 545, + 518 + ], + "type": "text", + "content": "The most straightforward way to define queries is to randomly initialise " + }, + { + "bbox": [ + 304, + 482, + 545, + 518 + ], + "type": "inline_equation", + "content": "\\mathbf{q} \\in \\mathbb{R}^{T \\times S \\times d}" + }, + { + "bbox": [ + 304, + 482, + 545, + 518 + ], + "type": "text", + "content": ", where there are " + }, + { + "bbox": [ + 304, + 482, + 545, + 518 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 304, + 482, + 545, + 518 + ], + "type": "text", + "content": " bounding boxes at each of the " + }, + { + "bbox": [ + 304, + 482, + 545, + 518 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 304, + 482, + 545, + 518 + ], + "type": "text", + "content": " input frames in the video clip." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 520, + 545, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 520, + 545, + 616 + ], + "spans": [ + { + "bbox": [ + 304, + 520, + 545, + 616 + ], + "type": "text", + "content": "However, we find it is more effective to factorise the queries into separate learned spatial, " + }, + { + "bbox": [ + 304, + 520, + 545, + 616 + ], + "type": "inline_equation", + "content": "\\mathbf{q}^s\\in \\mathbb{R}^{S\\times d}" + }, + { + "bbox": [ + 304, + 520, + 545, + 616 + ], + "type": "text", + "content": ", and temporal, " + }, + { + "bbox": [ + 304, + 520, + 545, + 616 + ], + "type": "inline_equation", + "content": "\\mathbf{q}^{T\\times d}" + }, + { + "bbox": [ + 304, + 520, + 545, + 616 + ], + "type": "text", + "content": " parameters. To obtain the final tubelet queries, we simply repeat the spatial queries across all frames, and add them to their corresponding temporal embedding at each location, as shown in Fig. 2. More concretely " + }, + { + "bbox": [ + 304, + 520, + 545, + 616 + ], + "type": "inline_equation", + "content": "\\mathbf{q}_{ij} = \\mathbf{q}_i^t +\\mathbf{q}_j^s" + }, + { + "bbox": [ + 304, + 520, + 545, + 616 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 304, + 520, + 545, + 616 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 520, + 545, + 616 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 520, + 545, + 616 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 304, + 520, + 545, + 616 + ], + "type": "text", + "content": " denote the temporal and spatial indices respectively." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 618, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 618, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 618, + 545, + 713 + ], + "type": "text", + "content": "The factorised query representation means that the same spatial embedding is used across all frames. Intuitively, this encourages the " + }, + { + "bbox": [ + 304, + 618, + 545, + 713 + ], + "type": "inline_equation", + "content": "i^{th}" + }, + { + "bbox": [ + 304, + 618, + 545, + 713 + ], + "type": "text", + "content": " spatial query embedding, " + }, + { + "bbox": [ + 304, + 618, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{q}_i^s" + }, + { + "bbox": [ + 304, + 618, + 545, + 713 + ], + "type": "text", + "content": ", to bind to the same location across different frames of the video, and since objects typically have small displacements from frame to frame, may help to associate bounding boxes within a tubelet together. We verify this intuition empirically in the experimental section." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18375" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 287, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 287, + 133 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 287, + 133 + ], + "type": "text", + "content": "Decoder layer The decoder layer in the original transformer [56] consists of self-attention on the queries, " + }, + { + "bbox": [ + 47, + 72, + 287, + 133 + ], + "type": "inline_equation", + "content": "\\mathbf{q}" + }, + { + "bbox": [ + 47, + 72, + 287, + 133 + ], + "type": "text", + "content": ", followed by cross-attention between the queries and the outputs of the encoder, " + }, + { + "bbox": [ + 47, + 72, + 287, + 133 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 47, + 72, + 287, + 133 + ], + "type": "text", + "content": ", and then a multilayer perceptron (MLP) layer [19, 56]:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 117, + 140, + 287, + 154 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 117, + 140, + 287, + 154 + ], + "spans": [ + { + "bbox": [ + 117, + 140, + 287, + 154 + ], + "type": "interline_equation", + "content": "\\mathbf {u} ^ {\\ell} = \\operatorname {M H S A} \\left(\\mathbf {q} ^ {\\ell}\\right) + \\mathbf {q} ^ {\\ell}, \\tag {1}", + "image_path": "a9bc386ed935c4c109b3a14ea3fa25f53272fc5e559784d4ff8db60c98111e75.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 118, + 156, + 287, + 171 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 118, + 156, + 287, + 171 + ], + "spans": [ + { + "bbox": [ + 118, + 156, + 287, + 171 + ], + "type": "interline_equation", + "content": "\\mathbf {v} ^ {\\ell} = \\operatorname {C A} \\left(\\mathbf {u} ^ {\\ell}, \\mathbf {x}\\right) + \\mathbf {u} ^ {\\ell}, \\tag {2}", + "image_path": "fb73c41d5dd830b6a72c78140c9002cf885361adbe98b90606d502a6a6212374.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 119, + 173, + 287, + 187 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 119, + 173, + 287, + 187 + ], + "spans": [ + { + "bbox": [ + 119, + 173, + 287, + 187 + ], + "type": "interline_equation", + "content": "\\mathbf {z} ^ {\\ell} = \\operatorname {M L P} \\left(\\mathbf {v} ^ {\\ell}\\right) + \\mathbf {v} ^ {\\ell}, \\tag {3}", + "image_path": "075b61b2455a09152b0961246698608a691825aee8ee66bf3b4cdd9db038ffee.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 197, + 287, + 257 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 197, + 287, + 257 + ], + "spans": [ + { + "bbox": [ + 47, + 197, + 287, + 257 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 197, + 287, + 257 + ], + "type": "inline_equation", + "content": "\\mathbf{z}^{\\ell}" + }, + { + "bbox": [ + 47, + 197, + 287, + 257 + ], + "type": "text", + "content": " is the output of the " + }, + { + "bbox": [ + 47, + 197, + 287, + 257 + ], + "type": "inline_equation", + "content": "\\ell^{th}" + }, + { + "bbox": [ + 47, + 197, + 287, + 257 + ], + "type": "text", + "content": " decoder layer, " + }, + { + "bbox": [ + 47, + 197, + 287, + 257 + ], + "type": "inline_equation", + "content": "\\mathbf{u}" + }, + { + "bbox": [ + 47, + 197, + 287, + 257 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 197, + 287, + 257 + ], + "type": "inline_equation", + "content": "\\mathbf{v}" + }, + { + "bbox": [ + 47, + 197, + 287, + 257 + ], + "type": "text", + "content": " are intermediate variables, MHSA denotes multi-headed self-attention and CA cross-attention. Note that the inputs to the MLP, self- and cross-attention operations are layer-normalised [4], which we omit here for clarity." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 258, + 287, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 258, + 287, + 376 + ], + "spans": [ + { + "bbox": [ + 46, + 258, + 287, + 376 + ], + "type": "text", + "content": "In our model, we factorise the self- and cross-attention layers across space and time respectively as shown in Fig. 3, to introduce a temporal locality inductive bias, and also to increase model efficiency. Concretely, when applying MHSA, we first compute the queries, keys and values, over which we attend twice: first independently at each time step with each frame, and then, independently along the time axis at each spatial location. Similarly, we modify the cross-attention operation so that only tubelet queries and backbone features from the same time index attend to each other." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 380, + 287, + 523 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 380, + 287, + 523 + ], + "spans": [ + { + "bbox": [ + 46, + 380, + 287, + 523 + ], + "type": "text", + "content": "Localisation and classification heads We obtain the final predictions of the network, " + }, + { + "bbox": [ + 46, + 380, + 287, + 523 + ], + "type": "inline_equation", + "content": "\\mathbf{y} = (\\mathbf{b},\\mathbf{a})" + }, + { + "bbox": [ + 46, + 380, + 287, + 523 + ], + "type": "text", + "content": ", by applying a small feed-forward network to the outputs to the decoder, " + }, + { + "bbox": [ + 46, + 380, + 287, + 523 + ], + "type": "inline_equation", + "content": "\\mathbf{z}" + }, + { + "bbox": [ + 46, + 380, + 287, + 523 + ], + "type": "text", + "content": ", following DETR [6]. The sequence of bounding boxes, " + }, + { + "bbox": [ + 46, + 380, + 287, + 523 + ], + "type": "inline_equation", + "content": "\\mathbf{b}" + }, + { + "bbox": [ + 46, + 380, + 287, + 523 + ], + "type": "text", + "content": ", is obtained with a 3-layer MLP, and is parameterised by the box center, width and height for each frame in the tubelet. A single-layer linear projection is used to obtain class logits, " + }, + { + "bbox": [ + 46, + 380, + 287, + 523 + ], + "type": "inline_equation", + "content": "\\mathbf{a}" + }, + { + "bbox": [ + 46, + 380, + 287, + 523 + ], + "type": "text", + "content": ". As we predict a fixed number of " + }, + { + "bbox": [ + 46, + 380, + 287, + 523 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 46, + 380, + 287, + 523 + ], + "type": "text", + "content": " bounding boxes per frame, and " + }, + { + "bbox": [ + 46, + 380, + 287, + 523 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 46, + 380, + 287, + 523 + ], + "type": "text", + "content": " is more than the maximum number of ground truth instances in the frame, we also include an additional class label, " + }, + { + "bbox": [ + 46, + 380, + 287, + 523 + ], + "type": "inline_equation", + "content": "\\varnothing" + }, + { + "bbox": [ + 46, + 380, + 287, + 523 + ], + "type": "text", + "content": ", which represents the \"background\" class which tubelets with no action class can be assigned to." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 529, + 155, + 544 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 529, + 155, + 544 + ], + "spans": [ + { + "bbox": [ + 47, + 529, + 155, + 544 + ], + "type": "text", + "content": "3.3. Training objective" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 548, + 287, + 632 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 548, + 287, + 632 + ], + "spans": [ + { + "bbox": [ + 46, + 548, + 287, + 632 + ], + "type": "text", + "content": "Our model predicts bounding boxes and action classes at each frame of the input video. Many datasets, however, such as AVA [18], are only sparsely annotated at selected keyframes of the video. In order to leverage the available annotations, we compute our training loss, Eq. 4, only at the annotated frames of the video, after having matched the predictions to the ground truth:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 100, + 640, + 287, + 669 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 100, + 640, + 287, + 669 + ], + "spans": [ + { + "bbox": [ + 100, + 640, + 287, + 669 + ], + "type": "interline_equation", + "content": "\\mathcal {L} (\\mathbf {y}, \\hat {\\mathbf {y}}) = \\frac {1}{| \\mathcal {T} |} \\sum_ {t \\in \\mathcal {T}} \\mathcal {L} _ {\\text {f r a m e}} (\\mathbf {y}, \\hat {\\mathbf {y}}), \\tag {4}", + "image_path": "ed8afe8fc284e176973f71c5805580fcb502a6ffd20cb06993b16e39bf7269c2.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 677, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 677, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 677, + 287, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 677, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{T}" + }, + { + "bbox": [ + 47, + 677, + 287, + 714 + ], + "type": "text", + "content": " is the set of labelled frames; " + }, + { + "bbox": [ + 47, + 677, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{y}" + }, + { + "bbox": [ + 47, + 677, + 287, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 677, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{y}}" + }, + { + "bbox": [ + 47, + 677, + 287, + 714 + ], + "type": "text", + "content": " denote the ground truth and predicted tubelets after matching. Following DETR [6], our training loss at each frame, " + }, + { + "bbox": [ + 47, + 677, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{frame}}" + }, + { + "bbox": [ + 47, + 677, + 287, + 714 + ], + "type": "text", + "content": ", is a" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 72, + 545, + 107 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 107 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 107 + ], + "type": "text", + "content": "sum of an " + }, + { + "bbox": [ + 305, + 72, + 545, + 107 + ], + "type": "inline_equation", + "content": "L_{1}" + }, + { + "bbox": [ + 305, + 72, + 545, + 107 + ], + "type": "text", + "content": " regression loss on bounding boxes, the generalised IoU loss [45] on bounding boxes, and a cross-entropy loss on action labels:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 114, + 545, + 154 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 114, + 545, + 154 + ], + "spans": [ + { + "bbox": [ + 313, + 114, + 545, + 154 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} _ {\\mathrm {f r a m e}} \\left(\\mathbf {b} ^ {t}, \\hat {\\mathbf {b}} ^ {t}, \\mathbf {a} ^ {t}, \\hat {\\mathbf {a}} ^ {t}\\right) = \\sum_ {i} \\mathcal {L} _ {\\mathrm {b o x}} \\left(\\mathbf {b} _ {i} ^ {t}, \\hat {\\mathbf {b}} _ {i} ^ {t}\\right) + \\mathcal {L} _ {\\mathrm {i o u}} \\left(\\mathbf {b} _ {i} ^ {t}, \\hat {\\mathbf {b}} _ {i} ^ {t}\\right) \\\\ + \\mathcal {L} _ {\\text {c l a s s}} \\left(\\mathbf {a} _ {i} ^ {t}, \\hat {\\mathbf {a}} _ {i} ^ {t}\\right). \\tag {5} \\\\ \\end{array}", + "image_path": "9e05f86736b27df839d2bb29f40f5fa44ddd88f84171843b069a5db5813c1c86.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 154, + 545, + 201 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 154, + 545, + 201 + ], + "spans": [ + { + "bbox": [ + 305, + 154, + 545, + 201 + ], + "type": "text", + "content": "Matching Set-based detection models such as DETR can make predictions in any order, which is why the predictions need to be matched to the ground truth before computing the training loss." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 201, + 545, + 285 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 201, + 545, + 285 + ], + "spans": [ + { + "bbox": [ + 305, + 201, + 545, + 285 + ], + "type": "text", + "content": "The first form of matching that we consider is to independently perform bipartite matching at each frame to align the model's predictions to the ground truth (or the " + }, + { + "bbox": [ + 305, + 201, + 545, + 285 + ], + "type": "inline_equation", + "content": "\\varnothing" + }, + { + "bbox": [ + 305, + 201, + 545, + 285 + ], + "type": "text", + "content": " background class) before computing the loss. In this case, we use the Hungarian algorithm [26] to obtain " + }, + { + "bbox": [ + 305, + 201, + 545, + 285 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 305, + 201, + 545, + 285 + ], + "type": "text", + "content": " permutations of " + }, + { + "bbox": [ + 305, + 201, + 545, + 285 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 305, + 201, + 545, + 285 + ], + "type": "text", + "content": " elements, " + }, + { + "bbox": [ + 305, + 201, + 545, + 285 + ], + "type": "inline_equation", + "content": "\\hat{\\pi}^t\\in \\Pi^t" + }, + { + "bbox": [ + 305, + 201, + 545, + 285 + ], + "type": "text", + "content": ", at each frame, where the permutation at the " + }, + { + "bbox": [ + 305, + 201, + 545, + 285 + ], + "type": "inline_equation", + "content": "t^{th}" + }, + { + "bbox": [ + 305, + 201, + 545, + 285 + ], + "type": "text", + "content": " frame minimises the per-frame loss," + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 361, + 293, + 545, + 314 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 361, + 293, + 545, + 314 + ], + "spans": [ + { + "bbox": [ + 361, + 293, + 545, + 314 + ], + "type": "interline_equation", + "content": "\\hat {\\pi} ^ {t} = \\underset {\\pi \\in \\Pi^ {t}} {\\arg \\min } \\mathcal {L} _ {\\text {f r a m e}} \\left(\\mathbf {y} ^ {t}, \\hat {\\mathbf {y}} _ {\\pi (i)} ^ {t}\\right). \\tag {6}", + "image_path": "3454525d7df96ae2e0bc05c434d24a407687e2dfe29bfb26889f4b224080973b.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 305, + 321, + 545, + 369 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 321, + 545, + 369 + ], + "spans": [ + { + "bbox": [ + 305, + 321, + 545, + 369 + ], + "type": "text", + "content": "An alternative is to perform tubelet matching, where all queries with the same spatial index, " + }, + { + "bbox": [ + 305, + 321, + 545, + 369 + ], + "type": "inline_equation", + "content": "\\mathbf{q}^s" + }, + { + "bbox": [ + 305, + 321, + 545, + 369 + ], + "type": "text", + "content": ", must match to the same ground truth annotation across all frames of the input video. Here the permutation is obtained over " + }, + { + "bbox": [ + 305, + 321, + 545, + 369 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 305, + 321, + 545, + 369 + ], + "type": "text", + "content": " elements as" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 344, + 376, + 545, + 406 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 344, + 376, + 545, + 406 + ], + "spans": [ + { + "bbox": [ + 344, + 376, + 545, + 406 + ], + "type": "interline_equation", + "content": "\\hat {\\pi} = \\underset {\\pi \\in \\Pi} {\\arg \\min } \\frac {1}{| \\mathcal {T} |} \\sum_ {t \\in \\mathcal {T}} \\mathcal {L} _ {\\text {f r a m e}} \\left(\\mathbf {y} ^ {t}, \\hat {\\mathbf {y}} _ {\\pi^ {t} (i)} ^ {t}\\right). \\tag {7}", + "image_path": "d7be6b570d91774a37fd24357e7b2a16ee9b4c449bdd17c54fa3df1bd80f6415.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 305, + 412, + 545, + 497 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 412, + 545, + 497 + ], + "spans": [ + { + "bbox": [ + 305, + 412, + 545, + 497 + ], + "type": "text", + "content": "Intuitively, tubelet matching provides stronger supervision when we have full tubelet annotations available. Note that regardless of the type of matching that we perform, the loss computation and the overall model architecture remains the same. Note that we do not weight terms in Eq. 5, for both matching and loss calculation, for simplicity, and to avoid having additional hyperparameters, as also done by [38]." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 306, + 503, + 378, + 515 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 503, + 378, + 515 + ], + "spans": [ + { + "bbox": [ + 306, + 503, + 378, + 515 + ], + "type": "text", + "content": "3.4. Discussion" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 521, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 521, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 521, + 545, + 713 + ], + "type": "text", + "content": "As our approach is based on DETR, it does not require external proposals nor non-maximal suppression for post-processing. The idea of using DETR for action localisation has also been explored by TubeR [66] and WOO [9]. There are, however, a number of key differences: WOO does not detect tubelets at all, but only actions at the center keyframe. We also factorise our queries in the spatial and temporal dimensions (Sec. 3.2) to provide inductive biases urging spatio-temporal association. Moreover, we predict action classes separately for each time step in the tubelet, meaning that each of our queries binds to an actor in the video. TubeR, in contrast, parameterises queries such that they are each associated with separate actions (features are average-pooled over the tubelet, and then linearly classified into a single action class). This choice also means that TubeR requires an additional \"action switch\" head to predict when" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18376" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 191 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 191 + ], + "type": "text", + "content": "tubelets start and end, which we do not require as different time steps in a tubelet can have different action classes in our model. Furthermore, we show experimentally (Tab. 1) that TubeR's parameterisation obtains lower accuracy. We also consider two types of matching in the loss computation (Sec. 3.3) unlike TubeR, with \"tubelet matching\" designed for predicting more temporally consistent tubelets. And in contrast to TubeR, we experimentally show how our decoder design allows our model to accurately predict tubelets even with weak, keyframe supervision." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 192, + 289, + 276 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 192, + 289, + 276 + ], + "spans": [ + { + "bbox": [ + 46, + 192, + 289, + 276 + ], + "type": "text", + "content": "Finally, TubeR requires additional complexity in the form of a \"short-term context module\" [66] and the external memory bank of [61] which is computed offline using a separate model to achieve strong results. As we show experimentally in the next section, we outperform TubeR without any additional modules, meaning that our model does indeed produce tubelets in an end-to-end manner." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 284, + 191, + 297 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 284, + 191, + 297 + ], + "spans": [ + { + "bbox": [ + 47, + 284, + 191, + 297 + ], + "type": "text", + "content": "4. Experimental Evaluation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 301, + 165, + 314 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 301, + 165, + 314 + ], + "spans": [ + { + "bbox": [ + 47, + 301, + 165, + 314 + ], + "type": "text", + "content": "4.1. Experimental set-up" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 318, + 287, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 318, + 287, + 365 + ], + "spans": [ + { + "bbox": [ + 46, + 318, + 287, + 365 + ], + "type": "text", + "content": "Datasets We evaluate on four spatio-temporal action localisation benchmarks. AVA and AVA-Kinetics contain sparse annotations at each keyframe, whereas UCF101-24 and JHMDB51-21 contain full tubelet annotations." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 366, + 287, + 461 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 366, + 287, + 461 + ], + "spans": [ + { + "bbox": [ + 46, + 366, + 287, + 461 + ], + "type": "text", + "content": "AVA [18] consists of 430, 15-minute video clips from movies. Keyframes are annotated at every second in the video, with about 210 000 labelled frames in the training set, and 57 000 in the validation set. There are 80 atomic actions labelled for every actor in the clip, of which 60 are used for evaluation [18]. Following standard practice, we report the Frame Average Precision (fAP) at an IoU threshold of 0.5 using the latest v2.2 annotations [18]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 462, + 287, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 462, + 287, + 545 + ], + "spans": [ + { + "bbox": [ + 46, + 462, + 287, + 545 + ], + "type": "text", + "content": "AVA-Kinetics [28] is a superset of AVA, and adds detection annotations following the AVA protocol, to a subset of Kinetics 700 [7] videos. Only a single keyframe in a 10-second Kinetics clip is labelled. In total, about 140 000 labelled keyframes are added to the training set, and 32 000 to the validation sets of AVA. Once again, we follow standard practice in reporting the Frame AP at a 0.5 IoU threshold." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 545, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 545, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 545, + 288, + 714 + ], + "type": "text", + "content": "UCF101-24 [49] is a subset of UCF101, and annotates 24 action classes with full spatio-temporal tubes in 3 207 untrimmed videos. Note that actions are not labelled exhaustively as in AVA, and there may be people present in the video who are not performing any labelled action. Following standard practice, we use the corrected annotations of [47]. We report both the Frame AP, which evaluates the predictions at each frame independently, and also the Video AP. The Video AP uses a 3D, spatio-temporal IoU to match predictions to targets. And since UCF101-24 videos are up to 900 frames long (median length of 164 frames), and our network typically processes " + }, + { + "bbox": [ + 46, + 545, + 288, + 714 + ], + "type": "inline_equation", + "content": "T = 32" + }, + { + "bbox": [ + 46, + 545, + 288, + 714 + ], + "type": "text", + "content": " frames at a time, we link together tubelet predictions from our network into full-video-tubes using the same causal linking algo" + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 335, + 125, + 512, + 163 + ], + "blocks": [ + { + "bbox": [ + 304, + 70, + 545, + 125 + ], + "lines": [ + { + "bbox": [ + 304, + 70, + 545, + 125 + ], + "spans": [ + { + "bbox": [ + 304, + 70, + 545, + 125 + ], + "type": "text", + "content": "Table 1. Comparison of detection architectures on AVA controlling for the same resolution (160p) and training settings. Binding each query to a person, rather than to an action (as done in TubeR [66]) yields solid improvements. We report the mean AP for both ViViT-B and CSN-152 backbones." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 335, + 125, + 512, + 163 + ], + "lines": [ + { + "bbox": [ + 335, + 125, + 512, + 163 + ], + "spans": [ + { + "bbox": [ + 335, + 125, + 512, + 163 + ], + "type": "table", + "html": "
ViViT-BCSN-152
Query binds to action23.625.7
Ours, query binds to person26.727.8
", + "image_path": "dd017bc019ad0f7434ce23d0be46a97a8f2143dd16f36a5694be1ef18f094eaa.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 171, + 457, + 182 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 171, + 457, + 182 + ], + "spans": [ + { + "bbox": [ + 305, + 171, + 457, + 182 + ], + "type": "text", + "content": "rithm as [22, 31] for fair comparison." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 182, + 545, + 230 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 182, + 545, + 230 + ], + "spans": [ + { + "bbox": [ + 304, + 182, + 545, + 230 + ], + "type": "text", + "content": "JHMDB51-21 [21] also contains full tube annotations in 928 trimmed videos. However, as the videos are shorter and at most 40 frames, we can process the entire clip with our network, and do not need to perform any linking." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 230, + 545, + 302 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 230, + 545, + 302 + ], + "spans": [ + { + "bbox": [ + 304, + 230, + 545, + 302 + ], + "type": "text", + "content": "MultiSports [32] is a recent dataset with full tube annotations for 66 fine-grained action classes in 2129 untrimmed videos, which are up to 3676 frames long. It is challenging, as it contains fast motion with multiple actors performing different concurrent actions, and there is a large variance in action durations. Evaluation is the same as for UCF101-24." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 304, + 546, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 304, + 546, + 472 + ], + "spans": [ + { + "bbox": [ + 304, + 304, + 546, + 472 + ], + "type": "text", + "content": "Implementation details For our vision encoder backbone, we consider both transformer-based (ViViT Factorised Encoder [1]), and convolutional (CSN [55]) backbones. For ViViT, we use the \"Base\" and \"Large\" model sizes [11, 12], which are typically first pretrained on image datasets like ImageNet-21K [10] and then finetuned on video datasets like Kinetics [23]. We also use CSN-152 pretrained on Instagram65M [37] and then Kinetics following [66]. Our models process " + }, + { + "bbox": [ + 304, + 304, + 546, + 472 + ], + "type": "inline_equation", + "content": "T = 32" + }, + { + "bbox": [ + 304, + 304, + 546, + 472 + ], + "type": "text", + "content": " frames unless otherwise specified, with " + }, + { + "bbox": [ + 304, + 304, + 546, + 472 + ], + "type": "inline_equation", + "content": "S = 64" + }, + { + "bbox": [ + 304, + 304, + 546, + 472 + ], + "type": "text", + "content": " spatial queries per frame and latent decoder dimensionality of " + }, + { + "bbox": [ + 304, + 304, + 546, + 472 + ], + "type": "inline_equation", + "content": "d = 2048" + }, + { + "bbox": [ + 304, + 304, + 546, + 472 + ], + "type": "text", + "content": ". Exhaustive implementation details and training hyperparameters are included in the supplement. We will also release all code and models upon acceptance." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 479, + 405, + 490 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 479, + 405, + 490 + ], + "spans": [ + { + "bbox": [ + 306, + 479, + 405, + 490 + ], + "type": "text", + "content": "4.2. Ablation studies" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 495, + 545, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 495, + 545, + 567 + ], + "spans": [ + { + "bbox": [ + 304, + 495, + 545, + 567 + ], + "type": "text", + "content": "We analyse the design choices in our model by conducting experiments on both AVA (with sparse per-frame supervision) and on UCF101-24 (where we can evaluate the quality of our predicted tubelets). Unless otherwise stated, our backbone is ViViT-Base pretrained on Kinetics 400, and the frame resolution is 160 pixels (160p) on the smaller side." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 570, + 545, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 545, + 689 + ], + "type": "text", + "content": "Comparison of detection architectures Tab. 1 compares our model, where each query represents a person, and all of their actions (Sec. 3.2) to the approach of TubeR [66] (Sec. 3.4), where there is a separate query for each action being performed. We observe that our parameterisation has a substantial impact, with our method outperforming binding to actions by 3.1 points with a ViViT backbone, and 2.1 points with a CSN backbone on the AVA dataset, therefore motivating the design of our decoder. the supplement shows that this trend is consistent on UCF101-24 and JHMDB too." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "text", + "content": "Another architectural baseline that we can compare to is that of a two-stage Fast-RCNN model using external person" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18377" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 48, + 118, + 286, + 167 + ], + "blocks": [ + { + "bbox": [ + 47, + 64, + 286, + 118 + ], + "lines": [ + { + "bbox": [ + 47, + 64, + 286, + 118 + ], + "spans": [ + { + "bbox": [ + 47, + 64, + 286, + 118 + ], + "type": "text", + "content": "Table 2. Comparison of independent and factorised queries on the AVA, UCF101-24 and JHMDB51-21 datasets. Factorised queries are particularly beneficial for predicting tubelets, as shown by the VideoAP on UCF101-24 and JHMDB51-21 which has full tube annotations. Both models use tubelet matching in the loss." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 48, + 118, + 286, + 167 + ], + "lines": [ + { + "bbox": [ + 48, + 118, + 286, + 167 + ], + "spans": [ + { + "bbox": [ + 48, + 118, + 286, + 167 + ], + "type": "table", + "html": "
QueryAVA fAPUCF101-24JHMDB51-21
fAPvAP20vAP50vAP50:95fAPvAP20vAP50
Independent25.285.686.359.528.985.088.585.2
Factorised26.386.587.463.429.886.989.588.2
", + "image_path": "3a37ca940b5122466a807fe93fadab334d14b00ae27b1ec7a33d8101a1d7d105.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 49, + 237, + 286, + 282 + ], + "blocks": [ + { + "bbox": [ + 47, + 172, + 286, + 236 + ], + "lines": [ + { + "bbox": [ + 47, + 172, + 286, + 236 + ], + "spans": [ + { + "bbox": [ + 47, + 172, + 286, + 236 + ], + "type": "text", + "content": "Table 3. Comparison of independent and tubelet matching for computing the loss on AVA, UCF101-24 and JHMD51-21. Tubelet matching helps for tube-level evaluation metrics like the Video AP (vAP) on UCF101-24 and JHMDB51-21. Note that tubelet matching is actually still possible on AVA as the annotations are at 1fps with actor identities." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 49, + 237, + 286, + 282 + ], + "lines": [ + { + "bbox": [ + 49, + 237, + 286, + 282 + ], + "spans": [ + { + "bbox": [ + 49, + 237, + 286, + 282 + ], + "type": "table", + "html": "
QueryAVA fAPUCF101-24JHMDB51-21
fAPvAP20vAP50vAP50:95fAPvAP20vAP50
Per-frame matching26.788.285.763.529.486.088.386.0
Tubelet matching26.386.587.463.429.886.989.588.2
", + "image_path": "342c2361acb77237cbac5932fae57b28f2193969e0c40bc8c18413bb969638eb.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 293, + 286, + 424 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 293, + 286, + 424 + ], + "spans": [ + { + "bbox": [ + 46, + 293, + 286, + 424 + ], + "type": "text", + "content": "detections from [61], as used by [3, 13, 15, 62]. This baseline using the same ViViT-B backbone achieved a mean AP of 25.2, which is still 1.5 points below our model, emphasising the promise of our end-to-end approach. Note that the proposals of [61] obtain an AP50 of 93.9 for person detection on the AVA validation set. They were obtained by first pretraining a Faster-RCNN [44] detector on COCO keypoints, and then finetuning on the person boxes from the AVA training set, using a resolution of 1333 on the longer side. Our model is end-to-end, and does not require any external proposals generated by a separate model at all." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 428, + 286, + 560 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 428, + 286, + 560 + ], + "spans": [ + { + "bbox": [ + 46, + 428, + 286, + 560 + ], + "type": "text", + "content": "Comparison to TubeR The second row of Tab. 1 using a CSN-152 backbone corresponds to our reimplementation of TubeR. By keeping all other training hyperparameters constant, we observe that our query binding provides an improvement of " + }, + { + "bbox": [ + 46, + 428, + 286, + 560 + ], + "type": "inline_equation", + "content": "2.1\\mathrm{mAP}" + }, + { + "bbox": [ + 46, + 428, + 286, + 560 + ], + "type": "text", + "content": " points in a fair comparison. Note that we could not use the public TubeR code [65], as it does not reproduce the paper's results: A higher resolution 256p model achieved only " + }, + { + "bbox": [ + 46, + 428, + 286, + 560 + ], + "type": "inline_equation", + "content": "20\\mathrm{mAP}" + }, + { + "bbox": [ + 46, + 428, + 286, + 560 + ], + "type": "text", + "content": " when trained with the public code, whilst it is reported to achieve 31.1. Exhaustive details on our attempts to reproduce TubeR with the authors' public code is in the supplement." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 564, + 286, + 684 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 564, + 286, + 684 + ], + "spans": [ + { + "bbox": [ + 46, + 564, + 286, + 684 + ], + "type": "text", + "content": "Query parameterisation Tab. 2 compares our independent and factorised query methods (Sec. 3.2) on AVA and UCF101-24. We observe that factorised queries consistently provide improvements on both the Frame AP and the Video AP across both datasets. As hypothesised in Sec. 3.2, we believe that this is due to the inductive bias present in this parameterisation. Note that we can only measure the Video AP on UCF101-24 as it has tubes labelled. We also show in the supplement that these observations are consistent on the JHMDB dataset too." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 689, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 286, + 713 + ], + "type": "text", + "content": "Matching for loss calculation As described in Sec. 3.3, when matching the predictions to the ground truth for loss" + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 309, + 118, + 544, + 172 + ], + "blocks": [ + { + "bbox": [ + 305, + 64, + 545, + 118 + ], + "lines": [ + { + "bbox": [ + 305, + 64, + 545, + 118 + ], + "spans": [ + { + "bbox": [ + 305, + 64, + 545, + 118 + ], + "type": "text", + "content": "Table 4. Our model can predict tubelets even when the ground truth annotations are sparse. We show this by subsampling training annotations from the UCF101-24 dataset. Our model sees minimal performance deterioration even when using only " + }, + { + "bbox": [ + 305, + 64, + 545, + 118 + ], + "type": "inline_equation", + "content": "1/24" + }, + { + "bbox": [ + 305, + 64, + 545, + 118 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 305, + 64, + 545, + 118 + ], + "type": "inline_equation", + "content": "4\\%" + }, + { + "bbox": [ + 305, + 64, + 545, + 118 + ], + "type": "text", + "content": " of the annotated frames." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 309, + 118, + 544, + 172 + ], + "lines": [ + { + "bbox": [ + 309, + 118, + 544, + 172 + ], + "spans": [ + { + "bbox": [ + 309, + 118, + 544, + 172 + ], + "type": "table", + "html": "
SamplingLabelled framesfAPvAP20vAP50vAP50:95
All frames458 81486.587.463.429.8
Every 1239 23785.287.263.029.3
Every 2420 24384.986.863.228.1
One per video2 28470.277.148.520.4
", + "image_path": "f6787ce5ae778a37e3e5ae3bb282d58ad5ea08650cc3cd9f080acce266dce6db.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 182, + 545, + 372 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 182, + 545, + 372 + ], + "spans": [ + { + "bbox": [ + 304, + 182, + 545, + 372 + ], + "type": "text", + "content": "computation, we can either independently match the outputs at each frame to the ground truths at each frame, or, we can match entire predicted tubelets to the ground truth tubelets. Tab. 3 shows that tubelet matching does indeed improve the quality of the predicted tubelets, as shown by the Video AP on UCF101-24. However, this comes at the cost of the quality of per-frame predictions, (i.e. Frame AP). This suggests that tubelet matching improves the association of bounding boxes predicted at different frames (hence higher Video AP), but may also impair the quality of the bounding boxes predicted at each frame (Frame AP). Note that it is technically possible for us to also perform tubelet matching on AVA, since AVA is annotated at 1fps with actor identities, and our model is input 32 frames at 12.5fps (therefore 2.56 seconds of temporal context) meaning that we have sparse tubelets with 2 or 3 annotated frames." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 373, + 545, + 433 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 373, + 545, + 433 + ], + "spans": [ + { + "bbox": [ + 304, + 373, + 545, + 433 + ], + "type": "text", + "content": "As tubelet matching helps with the overall Video AP, we use it for subsequent experiments on UCF101-24 and JHMDB51-21. For AVA, we use per-frame matching as the standard evaluation metric is the Frame AP, and annotations are sparse at 1fps." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 435, + 545, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 435, + 545, + 518 + ], + "spans": [ + { + "bbox": [ + 304, + 435, + 545, + 518 + ], + "type": "text", + "content": "Weakly-supervised tubelet detection Our model can predict tubelets even when the ground truth annotations are sparse and only labelled at certain frames (such as the AVA dataset). We quantitatively measure this ability of our model on UCF101-24 which has full tube annotations. We do so by subsampling labels from the training set, and evaluating the full tubes on the validation set." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 520, + 545, + 686 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 520, + 545, + 686 + ], + "spans": [ + { + "bbox": [ + 304, + 520, + 545, + 686 + ], + "type": "text", + "content": "As shown in Tab. 4, we still obtain meaningful tube predictions, with a Video AP20 of 77.1, when using only a single frame of annotation from each UCF video clip. When retaining 1 frame of supervision for every 24 labelled frames (which is roughly 1fps and corresponds to the AVA dataset's annotations), we observe minimal deterioration with respect to the fully supervised model (all Video AP metrics are within 0.7 points). Retaining 1 frame of annotation for every 12 consecutive labelled frames also performs similarly to using all frames in the video clip. These results suggest that due to the redundancy in the data (motion between frames is often limited), and the inductive bias of our model, we do not require each frame in the tube to be labelled in order to predict accurate tubelets." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "content": "Decoder design Tabs. 5 and 6 analyse the effect of the decoder depth and the type of attention in the decoder (de" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18378" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 83, + 87, + 250, + 117 + ], + "blocks": [ + { + "bbox": [ + 47, + 68, + 286, + 87 + ], + "lines": [ + { + "bbox": [ + 47, + 68, + 286, + 87 + ], + "spans": [ + { + "bbox": [ + 47, + 68, + 286, + 87 + ], + "type": "text", + "content": "Table 5. Effect of decoder depth on performance on the AVA dataset. Performance saturates at " + }, + { + "bbox": [ + 47, + 68, + 286, + 87 + ], + "type": "inline_equation", + "content": "L = 6" + }, + { + "bbox": [ + 47, + 68, + 286, + 87 + ], + "type": "text", + "content": " layers." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 83, + 87, + 250, + 117 + ], + "lines": [ + { + "bbox": [ + 83, + 87, + 250, + 117 + ], + "spans": [ + { + "bbox": [ + 83, + 87, + 250, + 117 + ], + "type": "table", + "html": "
Layers (L)01369
mAP ↑23.424.626.226.526.7
", + "image_path": "2b7a8967a6314c3128b9e78de868f8cd360cf22f2e5843f610ca0bb17a2aa49b.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 99, + 152, + 231, + 190 + ], + "blocks": [ + { + "bbox": [ + 47, + 119, + 286, + 152 + ], + "lines": [ + { + "bbox": [ + 47, + 119, + 286, + 152 + ], + "spans": [ + { + "bbox": [ + 47, + 119, + 286, + 152 + ], + "type": "text", + "content": "Table 6. Effect of the type of attention used in the decoder on AVA. Factorised attention is both more accurate and efficient (almost half of the GFLOPs per decoder layer)." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 99, + 152, + 231, + 190 + ], + "lines": [ + { + "bbox": [ + 99, + 152, + 231, + 190 + ], + "spans": [ + { + "bbox": [ + 99, + 152, + 231, + 190 + ], + "type": "table", + "html": "
Decoder attentionmAPGFLOPs
Full26.410.5
Factorised26.75.3
", + "image_path": "b81ab6741550da5be0f14da3558662c29ec7bae98d3f70b9341e106037c1b897.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 92, + 260, + 237, + 325 + ], + "blocks": [ + { + "bbox": [ + 46, + 194, + 286, + 260 + ], + "lines": [ + { + "bbox": [ + 46, + 194, + 286, + 260 + ], + "spans": [ + { + "bbox": [ + 46, + 194, + 286, + 260 + ], + "type": "text", + "content": "Table 7. Increasing the image resolution on the AVA dataset leads to consistent accuracy improvements, primarily on small objects. APs, APm and API denote the AP at 0.5 IoU threshold on small, medium and large boxes respectively following the COCO protocol [33]. AVA videos have a median aspect ratio of 16:10, and we pad the larger side when the aspect ratio is different." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 92, + 260, + 237, + 325 + ], + "lines": [ + { + "bbox": [ + 92, + 260, + 237, + 325 + ], + "spans": [ + { + "bbox": [ + 92, + 260, + 237, + 325 + ], + "type": "table", + "html": "
ResolutionmAPAPsAPmAPI
140 × 22425.47.211.227.8
160 × 25626.711.512.528.7
220 × 35228.812.015.130.7
260 × 41629.413.315.831.0
320 × 51230.017.516.032.0
", + "image_path": "902d74a4ca5ba30c8635f75f16f02d9d99c3f0fcbf8721001d1f018f672b09b0.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 83, + 358, + 248, + 404 + ], + "blocks": [ + { + "bbox": [ + 47, + 325, + 286, + 358 + ], + "lines": [ + { + "bbox": [ + 47, + 325, + 286, + 358 + ], + "spans": [ + { + "bbox": [ + 47, + 325, + 286, + 358 + ], + "type": "text", + "content": "Table 8. Comparison of pretraining for our models with ViViT-B and ViViT-L backbones on AVA using a resolution of " + }, + { + "bbox": [ + 47, + 325, + 286, + 358 + ], + "type": "inline_equation", + "content": "160 \\times 256" + }, + { + "bbox": [ + 47, + 325, + 286, + 358 + ], + "type": "text", + "content": ". Larger models benefit more from additional initial pretraining." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 83, + 358, + 248, + 404 + ], + "lines": [ + { + "bbox": [ + 83, + 358, + 248, + 404 + ], + "spans": [ + { + "bbox": [ + 83, + 358, + 248, + 404 + ], + "type": "table", + "html": "
PretrainSTAR/BSTAR/L
IN21K [10] → K400 [23]26.727.0
IN21K [10] → K700 [7]27.327.6
CLIP [42] → K700 [7]30.336.2
", + "image_path": "9f4bd6d036873a40974c2dee5b7e2a8183a5860166205d95f12a762c8caffa19.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 410, + 286, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 410, + 286, + 517 + ], + "spans": [ + { + "bbox": [ + 46, + 410, + 286, + 517 + ], + "type": "text", + "content": "scribed in Sec. 3.2). As seen in Tab. 5, detection accuracy on AVA increases with the number of decoder layers, plateauing at around 6 layers. It is possible to use no decoder layers too: In this case, instead of learning queries " + }, + { + "bbox": [ + 46, + 410, + 286, + 517 + ], + "type": "inline_equation", + "content": "\\mathbf{q}" + }, + { + "bbox": [ + 46, + 410, + 286, + 517 + ], + "type": "text", + "content": " (Sec. 3.2), we simply interpret the outputs of the vision encoder (Sec. 3.1), " + }, + { + "bbox": [ + 46, + 410, + 286, + 517 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 46, + 410, + 286, + 517 + ], + "type": "text", + "content": ", as our queries and apply the localisation and classification heads directly upon them. Using decoder layers, however, can provide a performance increase of up to " + }, + { + "bbox": [ + 46, + 410, + 286, + 517 + ], + "type": "inline_equation", + "content": "3.3\\mathrm{mAP}" + }, + { + "bbox": [ + 46, + 410, + 286, + 517 + ], + "type": "text", + "content": " points (" + }, + { + "bbox": [ + 46, + 410, + 286, + 517 + ], + "type": "inline_equation", + "content": "14\\%" + }, + { + "bbox": [ + 46, + 410, + 286, + 517 + ], + "type": "text", + "content": " relative), emphasising their utility." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 518, + 286, + 566 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 518, + 286, + 566 + ], + "spans": [ + { + "bbox": [ + 46, + 518, + 286, + 566 + ], + "type": "text", + "content": "Tab. 6 shows that factorised attention in the decoder is more accurate than standard, \"full\" attention between all queries and visual features. Moreover, it is more efficient too, using almost half of the GFLOPs at each decoder layer." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 570, + 286, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 570, + 286, + 700 + ], + "spans": [ + { + "bbox": [ + 46, + 570, + 286, + 700 + ], + "type": "text", + "content": "Effect of resolution and pretraining Scaling up the image resolution is critical to achieving high performance for object detection in images [20, 46]. However, we are not aware of previous works studying this for video action localisation. Tab. 7 shows that we do indeed observe substantial improvements from higher resolution, improving by up to 4.6 points on AVA. As expected, higher resolutions help more for detection at small sizes, where we follow the COCO [33] convention of object sizes. Note that AVA videos have a median aspect ratio of 16:10, and we pad the larger side for videos with different aspect ratios." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 59, + 701, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 701, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 59, + 701, + 286, + 713 + ], + "type": "text", + "content": "Similarly, Tab. 8 shows the effect of different pretraining" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 72, + 545, + 192 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 192 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 192 + ], + "type": "text", + "content": "datasets. Video vision transformers are typically pretrained on an image dataset (like ImageNet-21K [10]), before being finetuned on a video dataset, such as Kinetics [23]. We find that the initial image checkpoint plays an important role, with CLIP [42] pretraining significantly outperforming supervised pretraining on ImageNet-21K [12, 50]. This improvement grows further when using a \"Large\" backbone. Our observations are consistent with prior works which have shown that CLIP-pretraining outperforms ImageNet-21K and even JFT pretraining [24, 30, 34, 41]." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 197, + 545, + 220 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 197, + 545, + 220 + ], + "spans": [ + { + "bbox": [ + 304, + 197, + 545, + 220 + ], + "type": "text", + "content": "Qualitative examples We include example result videos of our proposed model in the supplementary." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 230, + 471, + 243 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 230, + 471, + 243 + ], + "spans": [ + { + "bbox": [ + 305, + 230, + 471, + 243 + ], + "type": "text", + "content": "4.3. Comparison to state-of-the-art" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 248, + 545, + 284 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 248, + 545, + 284 + ], + "spans": [ + { + "bbox": [ + 304, + 248, + 545, + 284 + ], + "type": "text", + "content": "We compare our model to the state-of-the-art on datasets with both sparsely annotated keyframes (AVA and AVA-K), and full tubes (UCF101-24, JHMDB and MultiSports)." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 289, + 545, + 480 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 289, + 545, + 480 + ], + "spans": [ + { + "bbox": [ + 304, + 289, + 545, + 480 + ], + "type": "text", + "content": "AVA and AVA-Kinetics Tab. 9 compares to prior work on AVA and AVA-Kinetics. The best previous methods relied on external proposals [3, 54, 60] and external memory banks [40, 62] which we outperform. There are fewer end-to-end approaches, and we outperform these by an even larger margin. Note that though TubeR [66] is a proposal-free approach, their best results are actually obtained with the external memory of [61]. Consequently, we have reported the end-to-end, and external-memory versions of TubeR (\"TubeR + LTC\") separately in Tab. 9. Furthermore, as detailed in the supplement, the public TubeR training code produces significantly lower performance (20.0 mAP). Hence, to compare to it, we report results with our reimplementation in addition to the paper's results. Observe that we outperform TubeR using the same CSN-152 backbone, and then improve further using larger transformer backbones." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 481, + 545, + 660 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 481, + 545, + 660 + ], + "spans": [ + { + "bbox": [ + 304, + 481, + 545, + 660 + ], + "type": "text", + "content": "We achieve greater relative improvements on AVA-Kinetics, showing that our end-to-end approach can leverage larger datasets more effectively. To our knowledge, we surpass the best previous results on AVA-Kinetics, achieving a Frame AP of 45.1. Notably, we outperform InternVideo [60] and VideoMAE-v2 [57], which are two recent video foundation models using more powerful backbones and larger, private, web-scale video datasets. Note that InternVideo consists of two different encoders, one of which is also initialised from CLIP. And like [60], we achieve our best AVA results by training a model on AVA-Kinetics, and then evaluating it only on the AVA validation set. Finally, we do not use any test-time augmentation, unlike previous works that ensemble results over multiple resolutions and/or left/right flips as shown by the \"Views\" column." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "content": "UCF101-24 Tab. 10 shows that we outperform prior work on UCF101-24, both in terms of frame-level (Frame AP), and tube-level metrics (Video AP). We achieve state-of-the-art results with a CSN-152 backbone, and improve further" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18379" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 59, + 118, + 533, + 331 + ], + "blocks": [ + { + "bbox": [ + 46, + 64, + 547, + 118 + ], + "lines": [ + { + "bbox": [ + 46, + 64, + 547, + 118 + ], + "spans": [ + { + "bbox": [ + 46, + 64, + 547, + 118 + ], + "type": "text", + "content": "Table 9. Comparison to the state-of-the-art (reported with mean Average Precision; mAP ↑) on AVA v2.2 [18] and AVA-Kinetics (AVA-K) [28]. Methods using external proposals (i.e. not end-to-end) are also trained on additional object detection and human pose data. Unless otherwise stated, separate models are trained for AVA and AVA-Kinetics. * denotes the model was trained on AVA-Kinetics and evaluated on AVA. “Res.” denotes the frame resolution of the shorter side. Web-scale foundational models are denoted in grey. † We also report our reimplemented results for TubeR, as the authors’ public code does not reproduce their reported results, as detailed in the supplement." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 59, + 118, + 533, + 331 + ], + "lines": [ + { + "bbox": [ + 59, + 118, + 533, + 331 + ], + "spans": [ + { + "bbox": [ + 59, + 118, + 533, + 331 + ], + "type": "table", + "html": "
PretrainingViewsAVAAVA-KRes.BackboneEnd-to-end
MViT-B [13]K400127.3--MViTX
Unified [2]K400627.7-320SlowFastX
AIA [53]K7001832.3-320SlowFastX
ACAR [40]K700633.336.4320SlowFastX
TubeR [66] with LTC [61]IG65M [37]→K400, COCO233.6-256CSN-152X
MeMViT [62]K700-34.4-312MViT v2X
Co-finetuning [3]IN21K→K700, MiT, SSv2132.833.1320ViViT/LX
JFT,WTS→K700, MiT, SSv2136.136.2320ViViT/LX
VideoMAE [54]SSL K700 → Sup. K700.-39.3--ViViT/LX
InternVideo* [60]7 different datasets-41.042.5-Uniformer v2X
VideoMAE v2 [57]6 different datasets-42.643.9-ViViT/gX
Action Transformer [28]K4001-23.0400I3D✓
WOO [9]K600128.3-320SlowFast✓
TubeR [66]IG65M [37]→K400, COCO131.1-256CSN-152✓
TubeR reimplemented†IG65M [37]→K400129.533.6256CSN-152✓
STAR/CSN-152 (ours)IG65M→K400131.435.8256CSN-152✓
STAR/B (ours)IN21K→K400130.036.6320ViViT/B✓
CLIP→K700133.939.1320ViViT/B✓
STAR/L (ours)CLIP→K700139.244.5320ViViT/L✓
STAR/L (ours)*CLIP→K700142.545.1420ViViT/L✓
", + "image_path": "9083e1d837640d73c685bf74f36d267bf0ca69ad9c12123ece5142adab4bc9fc.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 59, + 342, + 535, + 466 + ], + "blocks": [ + { + "bbox": [ + 91, + 333, + 501, + 342 + ], + "lines": [ + { + "bbox": [ + 91, + 333, + 501, + 342 + ], + "spans": [ + { + "bbox": [ + 91, + 333, + 501, + 342 + ], + "type": "text", + "content": "Table 10. State-of-the-art comparison on datasets with tubelet annotations, UCF101, JHMDB51 and MultiSports." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 59, + 342, + 535, + 466 + ], + "lines": [ + { + "bbox": [ + 59, + 342, + 535, + 466 + ], + "spans": [ + { + "bbox": [ + 59, + 342, + 535, + 466 + ], + "type": "table", + "html": "
PretrainingUCF101-24JHMDB51-21MultiSportsBackbone
fAPvAP20vAP50vAP50:95fAPvAP20vAP50fAPvAP20vAP50
ACT [22]IN1K67.177.251.425.065.774.273.7---VGG
MOC [31]IN1K→COCO78.082.853.828.370.877.377.225.212.90.6DLA34
Unified [2]K60079.3---------SlowFast
WOO [9]K600----80.5-----SlowFast
TubeR [66]IG65M→K40083.283.358.428.9-87.482.3---CSN-152
HIT [14]K70084.888.874.3-83.889.788.133.327.88.8SlowFast
FBIL [8]K70086.086.969.1----40.830.39.9SlowFast
STAR/CSN-152IG65M→K40086.787.065.430.693.596.395.445.450.118.6CSN-152
STAR/BIN21K→K40087.388.268.631.786.989.588.246.351.625.6ViViT/B
STAR/LCLIP→K70090.389.873.435.892.193.192.659.362.036.2ViViT/L
", + "image_path": "8d0d423272809171f9c064297de9e2e11bbdc6d2b993f8cee2d973b8488af75b.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 468, + 288, + 612 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 468, + 288, + 612 + ], + "spans": [ + { + "bbox": [ + 46, + 468, + 288, + 612 + ], + "type": "text", + "content": "by scaling up to ViViT-Large, consistent with our results on AVA (Tab. 9). Moreover, note how we substantially outperform TubeR [66] using the same CSN-152 backbone. Our margin of improvement over TubeR grows from vAP20 (+3.7) to vAP50 (+7.0) with the same backbone, showing that our tubelets are more precise, and is in line with our visual observations in the supplement. To our knowledge, we outperform the best previous reported Video AP50 result by 13.2 points. Note that as UCF videos are up to 900 frames, and as our network processes " + }, + { + "bbox": [ + 46, + 468, + 288, + 612 + ], + "type": "inline_equation", + "content": "T = 32" + }, + { + "bbox": [ + 46, + 468, + 288, + 612 + ], + "type": "text", + "content": " frames, we follow prior works and link together tubelets using the same causal algorithm as [2, 22, 31, 47] for fair comparison." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 617, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 617, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 617, + 288, + 715 + ], + "type": "text", + "content": "JHMDB51-21 Tab. 10 shows that we surpass the state-of-the-art on JHMDB. Once again, we significantly outperform TubeR [66], by 13.1 vAP50, with the same CSN-152 backbone. The CSN-152 backbone outperforms ViViT in this case, possibly because this is the smallest dataset and larger backbones can overfit more easily. The videos in this dataset are trimmed (meaning that labelled actions are being performed on each frame), and also shorter. Hence the" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 468, + 545, + 503 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 468, + 545, + 503 + ], + "spans": [ + { + "bbox": [ + 304, + 468, + 545, + 503 + ], + "type": "text", + "content": "Video AP is not as strict as it is on UCF. As videos are at most 40 frames, we set " + }, + { + "bbox": [ + 304, + 468, + 545, + 503 + ], + "type": "inline_equation", + "content": "T = 40" + }, + { + "bbox": [ + 304, + 468, + 545, + 503 + ], + "type": "text", + "content": " in our model so that we process the entire clip at once without needing to link tubelets." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 504, + 545, + 587 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 504, + 545, + 587 + ], + "spans": [ + { + "bbox": [ + 304, + 504, + 545, + 587 + ], + "type": "text", + "content": "MultiSports Finally, Tab. 10 shows that we improve on prior state-of-the-art on MultiSports substantially on both Frame AP and Video AP. Although this is a recent benchmark with few methods to compare against, our improvement is significant due to the challenging nature of the dataset: it contains fast motion, many simultaneous actors and large variation in action duration." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 591, + 471, + 604 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 591, + 471, + 604 + ], + "spans": [ + { + "bbox": [ + 305, + 591, + 471, + 604 + ], + "type": "text", + "content": "5. Conclusion and Future Work" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": "We have presented STAR, an end-to-end spatio-temporal action localisation model that can output tubelets, when either sparse keyframe, or full tubelet annotation is available. Our approach achieves state-of-the-art results on four action localisation datasets for both frame-level and tubelet-level predictions (in particular, we obtain " + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "inline_equation", + "content": "45.1\\%" + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": " mAP on the challenging AVA-Kinetics dataset), outperforming complex methods that use external proposals and memory banks. Future work is to extend our method to open vocabularies." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "18380" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 123 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 123 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 123 + ], + "type": "text", + "content": "[1] Anurag Arnab, Mostafa Dehghani, Georg Heigold, Chen Sun, Mario Lucic, and Cordelia Schmid. ViViT: A Video Vision Transformer. In ICCV, 2021. 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 125, + 288, + 157 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 125, + 288, + 157 + ], + "spans": [ + { + "bbox": [ + 53, + 125, + 288, + 157 + ], + "type": "text", + "content": "[2] Anurag Arnab, Chen Sun, and Cordelia Schmid. Unified graph structured models for video understanding. In ICCV, 2021. 1, 2, 8" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 159, + 288, + 213 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 159, + 288, + 213 + ], + "spans": [ + { + "bbox": [ + 53, + 159, + 288, + 213 + ], + "type": "text", + "content": "[3] Anurag Arnab, Xuehan Xiong, Alexey Gritsenko, Rob Romijnders, Josip Djolonga, Mostafa Dehghani, Chen Sun, Mario Lucic, and Cordelia Schmid. Beyond transfer learning: Co-finetuning for action localisation. In arXiv preprint arXiv:2207.03807, 2022. 1, 2, 6, 7, 8" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 215, + 287, + 247 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 215, + 287, + 247 + ], + "spans": [ + { + "bbox": [ + 53, + 215, + 287, + 247 + ], + "type": "text", + "content": "[4] Jimmy Lei Ba, Jamie Ryan Kiros, and Geoffrey E Hinton. Layer normalization. In arXiv preprint arXiv:1607.06450, 2016. 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 250, + 287, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 250, + 287, + 282 + ], + "spans": [ + { + "bbox": [ + 53, + 250, + 287, + 282 + ], + "type": "text", + "content": "[5] Fabien Baradel, Natalia Neverova, Christian Wolf, Julien Mille, and Greg Mori. Object level visual reasoning in videos. In ECCV, 2018. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 285, + 287, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 285, + 287, + 327 + ], + "spans": [ + { + "bbox": [ + 53, + 285, + 287, + 327 + ], + "type": "text", + "content": "[6] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In ECCV, 2020. 1, 2, 3, 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 329, + 287, + 361 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 329, + 287, + 361 + ], + "spans": [ + { + "bbox": [ + 53, + 329, + 287, + 361 + ], + "type": "text", + "content": "[7] Joao Carreira, Eric Noland, Chloe Hillier, and Andrew Zisserman. A short note on the kinetics-700 human action dataset. In arXiv preprint arXiv:1907.06987, 2019. 5, 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 363, + 287, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 363, + 287, + 396 + ], + "spans": [ + { + "bbox": [ + 53, + 363, + 287, + 396 + ], + "type": "text", + "content": "[8] Keke Chen, Xiangbo Shu, Guo-Sen Xie, Rui Yan, and Jinhui Tang. Foreground/background-masked interaction learning for spatio-temporal action detection. In ACM MM, 2023. 8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 397, + 287, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 397, + 287, + 441 + ], + "spans": [ + { + "bbox": [ + 53, + 397, + 287, + 441 + ], + "type": "text", + "content": "[9] Shoufa Chen, Peize Sun, Enze Xie, Chongjian Ge, Jiannan Wu, Lan Ma, Jiajun Shen, and Ping Luo. Watch only once: An end-to-end video action detection framework. In ICCV, 2021. 1, 2, 4, 8" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 442, + 287, + 474 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 442, + 287, + 474 + ], + "spans": [ + { + "bbox": [ + 48, + 442, + 287, + 474 + ], + "type": "text", + "content": "[10] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. ImageNet: A large-scale hierarchical image database. In CVPR, 2009. 5, 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 476, + 287, + 509 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 476, + 287, + 509 + ], + "spans": [ + { + "bbox": [ + 48, + 476, + 287, + 509 + ], + "type": "text", + "content": "[11] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. BERT: Pre-training of deep bidirectional transformers for language understanding. In NAACL, 2019. 5" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 511, + 287, + 574 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 511, + 287, + 574 + ], + "spans": [ + { + "bbox": [ + 48, + 511, + 287, + 574 + ], + "type": "text", + "content": "[12] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2021. 5, 7" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 578, + 287, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 578, + 287, + 609 + ], + "spans": [ + { + "bbox": [ + 48, + 578, + 287, + 609 + ], + "type": "text", + "content": "[13] Haoqi Fan, Bo Xiong, Karttikeya Mangalam, Yanghao Li, Zhicheng Yan, Jitendra Malik, and Christoph Feichtenhofer. Multiscale vision transformers. In ICCV, 2021. 6, 8" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 612, + 287, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 612, + 287, + 643 + ], + "spans": [ + { + "bbox": [ + 48, + 612, + 287, + 643 + ], + "type": "text", + "content": "[14] Gueter Josmy Faure, Min-Hung Chen, and Shang-Hong Lai. Holistic interaction transformer network for action detection. In WACV, 2023. 8" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 646, + 287, + 677 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 646, + 287, + 677 + ], + "spans": [ + { + "bbox": [ + 48, + 646, + 287, + 677 + ], + "type": "text", + "content": "[15] Christoph Feichtenhofer, Haoqi Fan, Jitendra Malik, and Kaiming He. Slowfast networks for video recognition. In ICCV, 2019. 1, 2, 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 48, + 680, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 680, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 680, + 287, + 712 + ], + "type": "text", + "content": "[16] Rohit Girdhar, Joao Carreira, Carl Doersch, and Andrew Zisserman. Video action transformer network. In CVPR, 2019. 2" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 34, + "blocks": [ + { + "bbox": [ + 307, + 73, + 498, + 83 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 498, + 83 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 498, + 83 + ], + "type": "text", + "content": "[17] Ross Girshick. Fast r-cnn. In ICCV, 2015. 1, 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 86, + 545, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 86, + 545, + 140 + ], + "spans": [ + { + "bbox": [ + 307, + 86, + 545, + 140 + ], + "type": "text", + "content": "[18] Chunhui Gu, Chen Sun, David A Ross, Carl Vondrick, Caroline Pantofaru, Yeqing Li, Sudheendra Vijayanarasimhan, George Toderici, Susanna Ricco, Rahul Sukthankar, et al. Ava: A video dataset of spatio-temporally localized atomic visual actions. In CVPR, 2018. 1, 2, 4, 5, 8" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 142, + 545, + 164 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 142, + 545, + 164 + ], + "spans": [ + { + "bbox": [ + 307, + 142, + 545, + 164 + ], + "type": "text", + "content": "[19] Dan Hendrycks and Kevin Gimpel. Gaussian error linear units (gelus). In arXiv preprint arXiv:1606.08415, 2016. 4" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 166, + 545, + 220 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 166, + 545, + 220 + ], + "spans": [ + { + "bbox": [ + 307, + 166, + 545, + 220 + ], + "type": "text", + "content": "[20] Jonathan Huang, Vivek Rathod, Chen Sun, Menglong Zhu, Anoop Korattikara, Alireza Fathi, Ian Fischer, Zbigniew Wojna, Yang Song, Sergio Guadarrama, et al. Speed/accuracy trade-offs for modern convolutional object detectors. In CVPR, 2017. 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 222, + 545, + 255 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 222, + 545, + 255 + ], + "spans": [ + { + "bbox": [ + 307, + 222, + 545, + 255 + ], + "type": "text", + "content": "[21] H. Jhuang, J. Gall, S. Zuffi, C. Schmid, and M. J. Black. Towards understanding action recognition. In ICCV, 2013. 1, 2, 5" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 257, + 545, + 289 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 257, + 545, + 289 + ], + "spans": [ + { + "bbox": [ + 307, + 257, + 545, + 289 + ], + "type": "text", + "content": "[22] Vicky Kalogeiton, Philippe Weinzaepfel, Vittorio Ferrari, and Cordelia Schmid. Action tubelet detector for spatiotemporal action localization. In ICCV, 2017. 1, 2, 5, 8" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 292, + 545, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 292, + 545, + 346 + ], + "spans": [ + { + "bbox": [ + 307, + 292, + 545, + 346 + ], + "type": "text", + "content": "[23] Will Kay, Joao Carreira, Karen Simonyan, Brian Zhang, Chloe Hillier, Sudheendra Vijayanarasimhan, Fabio Viola, Tim Green, Trevor Back, Paul Natsev, et al. The kinetics human action video dataset. In arXiv preprint arXiv:1705.06950, 2017. 5, 7" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 348, + 545, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 348, + 545, + 402 + ], + "spans": [ + { + "bbox": [ + 307, + 348, + 545, + 402 + ], + "type": "text", + "content": "[24] Skanda Koppula, Yazhe Li, Evan Shelhamer, Andrew Jaegle, Nikhil Parthasarathy, Relja Arandjelovic, João Carreira, and Olivier Henaff. Where should i spend my flops? efficiency evaluations of visual pre-training methods. In arXiv preprint arXiv:2209.15589, 2022. 7" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 404, + 545, + 448 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 404, + 545, + 448 + ], + "spans": [ + { + "bbox": [ + 307, + 404, + 545, + 448 + ], + "type": "text", + "content": "[25] Okan Köpukli, Xiangyu Wei, and Gerhard Rigoll. You only watch once: A unified cnn architecture for real-time spatiotemporal action localization. In arXiv preprint arXiv:1911.06644, 2019. 1, 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 450, + 545, + 482 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 450, + 545, + 482 + ], + "spans": [ + { + "bbox": [ + 307, + 450, + 545, + 482 + ], + "type": "text", + "content": "[26] Harold W Kuhn. The hungarian method for the assignment problem. Naval research logistics quarterly, 2(1-2):83-97, 1955. 4" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 484, + 545, + 517 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 484, + 545, + 517 + ], + "spans": [ + { + "bbox": [ + 307, + 484, + 545, + 517 + ], + "type": "text", + "content": "[27] Jie Lei, Tamara L Berg, and Mohit Bansal. Detecting moments and highlights in videos via natural language queries. In NeurIPS, 2021. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 519, + 545, + 563 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 519, + 545, + 563 + ], + "spans": [ + { + "bbox": [ + 307, + 519, + 545, + 563 + ], + "type": "text", + "content": "[28] Ang Li, Meghanata Thotakuri, David A Ross, João Carreira, Alexander Vostrikov, and Andrew Zisserman. The a-akinetics localized human actions video dataset. In arXiv preprint arXiv:2005.00214, 2020. 1, 2, 5, 8" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 565, + 545, + 597 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 565, + 545, + 597 + ], + "spans": [ + { + "bbox": [ + 307, + 565, + 545, + 597 + ], + "type": "text", + "content": "[29] Dong Li, Zhaofan Qiu, Qi Dai, Ting Yao, and Tao Mei. Recurrent tubelet proposal and recognition networks for action detection. In ECCV, 2018. 2" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 307, + 600, + 545, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 600, + 545, + 643 + ], + "spans": [ + { + "bbox": [ + 307, + 600, + 545, + 643 + ], + "type": "text", + "content": "[30] Kunchang Li, Yali Wang, Yinan He, Yizhuo Li, Yi Wang, Limin Wang, and Yu Qiao. Uniformerv2: Spatiotemporal learning by arming image vits with video uniformer. In arXiv preprint arXiv:2211.09552, 2022. 7" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 307, + 646, + 545, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 646, + 545, + 667 + ], + "spans": [ + { + "bbox": [ + 307, + 646, + 545, + 667 + ], + "type": "text", + "content": "[31] Yixuan Li, Zixu Wang, Limin Wang, and Gangshan Wu. Actions as moving points. In ECCV, 2020. 2, 5, 8" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 307, + 670, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 670, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 307, + 670, + 545, + 712 + ], + "type": "text", + "content": "[32] Yixuan Li, Lei Chen, Runyu He, Zhenzhi Wang, Gangshan Wu, and Limin Wang. Multisports: A multi-person video dataset of spatio-temporally localized sports actions. In ICCV, 2021. 2, 5" + } + ] + } + ], + "index": 33 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "18381" + } + ] + } + ], + "index": 35 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "text", + "content": "[33] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In ECCV, 2014. 7" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 117, + 288, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 117, + 288, + 160 + ], + "spans": [ + { + "bbox": [ + 48, + 117, + 288, + 160 + ], + "type": "text", + "content": "[34] Ziyi Lin, Shijie Geng, Renrui Zhang, Peng Gao, Gerard de Melo, Xiaogang Wang, Jifeng Dai, Yu Qiao, and Hongsheng Li. Frozen clip models are efficient video learners. In ECCV, 2022. 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 161, + 288, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 161, + 288, + 194 + ], + "spans": [ + { + "bbox": [ + 48, + 161, + 288, + 194 + ], + "type": "text", + "content": "[35] Wei Liu, Dragomir Anguelov, Dumitru Erhan, Christian Szegedy, Scott Reed, Cheng-Yang Fu, and Alexander C Berg. Ssd: Single shot multibox detector. In ECCV, 2016. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 194, + 288, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 194, + 288, + 237 + ], + "spans": [ + { + "bbox": [ + 48, + 194, + 288, + 237 + ], + "type": "text", + "content": "[36] Xiaolong Liu, Qimeng Wang, Yao Hu, Xu Tang, Shiwei Zhang, Song Bai, and Xiang Bai. End-to-end temporal action detection with transformer. IEEE Transactions on Image Processing, 31:5427-5441, 2022. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 238, + 288, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 238, + 288, + 282 + ], + "spans": [ + { + "bbox": [ + 48, + 238, + 288, + 282 + ], + "type": "text", + "content": "[37] Dhruv Mahajan, Ross Girshick, Vignesh Ramanathan, Kaiming He, Manohar Paluri, Yixuan Li, Ashwin Bharambe, and Laurens Van Der Maaten. Exploring the limits of weakly supervised pretraining. In ECCV, 2018. 5, 8" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 282, + 288, + 337 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 282, + 288, + 337 + ], + "spans": [ + { + "bbox": [ + 48, + 282, + 288, + 337 + ], + "type": "text", + "content": "[38] Matthias Minderer, Alexey Gritsenko, Austin Stone, Maxim Neumann, Dirk Weissenborn, Alexey Dosovitskiy, Aravindh Mahendran, Anurag Arnab, Mostafa Dehghani, Zhuoran Shen, et al. Simple Open-Vocabulary Object Detection with Vision Transformers. In ECCV, 2022. 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 337, + 288, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 337, + 288, + 370 + ], + "spans": [ + { + "bbox": [ + 48, + 337, + 288, + 370 + ], + "type": "text", + "content": "[39] Megha Nawhal and Greg Mori. Activity graph transformer for temporal action localization. In arXiv preprint arXiv:2101.08540, 2021. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 371, + 288, + 414 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 371, + 288, + 414 + ], + "spans": [ + { + "bbox": [ + 48, + 371, + 288, + 414 + ], + "type": "text", + "content": "[40] Junting Pan, Siyu Chen, Mike Zheng Shou, Yu Liu, Jing Shao, and Hongsheng Li. Actor-context-actor relation network for spatio-temporal action localization. In CVPR, 2021. 1, 2, 7, 8" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 415, + 288, + 448 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 415, + 288, + 448 + ], + "spans": [ + { + "bbox": [ + 48, + 415, + 288, + 448 + ], + "type": "text", + "content": "[41] Junting Pan, Ziyi Lin, Xiatian Zhu, Jing Shao, and Hongsheng Li. St-adapter: Parameter-efficient image-to-video transfer learning. In NeurIPS, 2022. 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 449, + 288, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 449, + 288, + 502 + ], + "spans": [ + { + "bbox": [ + 48, + 449, + 288, + 502 + ], + "type": "text", + "content": "[42] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, 2021. 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 503, + 288, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 503, + 288, + 536 + ], + "spans": [ + { + "bbox": [ + 48, + 503, + 288, + 536 + ], + "type": "text", + "content": "[43] Joseph Redmon, Santosh Divvala, Ross Girshick, and Ali Farhadi. You only look once: Unified, real-time object detection. In CVPR, 2016. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 536, + 288, + 569 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 536, + 288, + 569 + ], + "spans": [ + { + "bbox": [ + 48, + 536, + 288, + 569 + ], + "type": "text", + "content": "[44] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. In NeurIPS, 2015. 3, 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 570, + 288, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 570, + 288, + 613 + ], + "spans": [ + { + "bbox": [ + 48, + 570, + 288, + 613 + ], + "type": "text", + "content": "[45] Hamid Rezatofighi, Nathan Tsoi, JunYoung Gwak, Amir Sadeghian, Ian Reid, and Silvio Savarese. Generalized intersection over union: A metric and a loss for bounding box regression. In CVPR, 2019. 4" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 614, + 288, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 614, + 288, + 635 + ], + "spans": [ + { + "bbox": [ + 48, + 614, + 288, + 635 + ], + "type": "text", + "content": "[46] Bharat Singh and Larry S Davis. An analysis of scale invariance in object detection snip. In CVPR, 2018. 7" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 636, + 288, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 636, + 288, + 679 + ], + "spans": [ + { + "bbox": [ + 48, + 636, + 288, + 679 + ], + "type": "text", + "content": "[47] Gurkirt Singh, Suman Saha, Michael Sapienza, Philip HS Torr, and Fabio Cuzzolin. Online real-time multiple spatiotemporal action localisation and prediction. In ICCV, 2017. 2, 5, 8" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 680, + 288, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 680, + 288, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 680, + 288, + 712 + ], + "type": "text", + "content": "[48] Lin Song, Shiwei Zhang, Gang Yu, and Hongbin Sun. Tacnet: Transition-aware context network for spatio-temporal action detection. In CVPR, 2019. 2" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 33, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "text", + "content": "[49] Khurram Soomro, Amir Roshan Zamir, and Mubarak Shah. Ucf101: A dataset of 101 human actions classes from videos in the wild. In arXiv preprint arXiv:1212.0402, 2012. 1, 2, 5" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 106, + 545, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 106, + 545, + 149 + ], + "spans": [ + { + "bbox": [ + 307, + 106, + 545, + 149 + ], + "type": "text", + "content": "[50] Andreas Steiner, Alexander Kolesnikov, Xiaohua Zhai, Ross Wightman, Jakob Uszkoreit, and Lucas Beyer. How to train your vit? data, augmentation, and regularization in vision transformers. TMLR, 2022. 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 150, + 545, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 150, + 545, + 182 + ], + "spans": [ + { + "bbox": [ + 307, + 150, + 545, + 182 + ], + "type": "text", + "content": "[51] Chen Sun, Abhinav Shrivastava, Carl Vondrick, Kevin Murphy, Rahul Sukthankar, and Cordelia Schmid. Actor-centric relation network. In ECCV, 2018. 1, 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 183, + 545, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 183, + 545, + 227 + ], + "spans": [ + { + "bbox": [ + 307, + 183, + 545, + 227 + ], + "type": "text", + "content": "[52] Peize Sun, Rufeng Zhang, Yi Jiang, Tao Kong, Chenfeng Xu, Wei Zhan, Masayoshi Tomizuka, Lei Li, Zehuan Yuan, Changhu Wang, et al. Sparse r-cnn: End-to-end object detection with learnable proposals. In CVPR, 2021. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 228, + 545, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 228, + 545, + 259 + ], + "spans": [ + { + "bbox": [ + 307, + 228, + 545, + 259 + ], + "type": "text", + "content": "[53] Jiajun Tang, Jin Xia, Xinzhi Mu, Bo Pang, and Cewu Lu. Asynchronous interaction aggregation for action detection. In ECCV, 2020. 1, 2, 8" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 260, + 545, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 260, + 545, + 303 + ], + "spans": [ + { + "bbox": [ + 307, + 260, + 545, + 303 + ], + "type": "text", + "content": "[54] Zhan Tong, Yibing Song, Jue Wang, and Limin Wang. VideoMAE: Masked autoencoders are data-efficient learners for self-supervised video pre-training. In NeurIPS, 2022. 2, 7, 8" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 304, + 545, + 337 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 304, + 545, + 337 + ], + "spans": [ + { + "bbox": [ + 307, + 304, + 545, + 337 + ], + "type": "text", + "content": "[55] Du Tran, Heng Wang, Lorenzo Torresani, and Matt Feiszli. Video classification with channel-separated convolutional networks. In ICCV, 2019. 5" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 338, + 545, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 338, + 545, + 380 + ], + "spans": [ + { + "bbox": [ + 307, + 338, + 545, + 380 + ], + "type": "text", + "content": "[56] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In NeurIPS, 2017. 3, 4" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 382, + 545, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 382, + 545, + 426 + ], + "spans": [ + { + "bbox": [ + 307, + 382, + 545, + 426 + ], + "type": "text", + "content": "[57] Limin Wang, Bingkun Huang, Zhiyu Zhao, Zhan Tong, Yi-nan He, Yi Wang, Yali Wang, and Yu Qiao. Videomae v2: Scaling video masked autoencoders with dual masking. In arXiv preprint arXiv:2303.16727, 2023. 2, 7, 8" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 426, + 545, + 448 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 426, + 545, + 448 + ], + "spans": [ + { + "bbox": [ + 307, + 426, + 545, + 448 + ], + "type": "text", + "content": "[58] Xiaolong Wang and Abhinav Gupta. Videos as space-time region graphs. In ECCV, 2018. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 449, + 545, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 449, + 545, + 491 + ], + "spans": [ + { + "bbox": [ + 307, + 449, + 545, + 491 + ], + "type": "text", + "content": "[59] Yuqing Wang, Zhaoliang Xu, Xinlong Wang, Chunhua Shen, Baoshan Cheng, Hao Shen, and Huaxia Xia. End-to-end video instance segmentation with transformers. In CVPR, 2021. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 492, + 545, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 492, + 545, + 547 + ], + "spans": [ + { + "bbox": [ + 307, + 492, + 545, + 547 + ], + "type": "text", + "content": "[60] Yi Wang, Kunchang Li, Yizhuo Li, Yinan He, Bingkun Huang, Zhiyu Zhao, Hongjie Zhang, Jilan Xu, Yi Liu, Zun Wang, et al. Internvideo: General video foundation models via generative and discriminative learning. In arXiv preprint arXiv:2212.03191, 2022. 2, 7, 8" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 548, + 545, + 591 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 548, + 545, + 591 + ], + "spans": [ + { + "bbox": [ + 307, + 548, + 545, + 591 + ], + "type": "text", + "content": "[61] Chao-Yuan Wu, Christoph Feichtenhofer, Haoqi Fan, Kaiming He, Philipp Krahenbuhl, and Ross Girshick. Long-term feature banks for detailed video understanding. In CVPR, 2019. 1, 2, 5, 6, 7, 8" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 592, + 545, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 592, + 545, + 646 + ], + "spans": [ + { + "bbox": [ + 307, + 592, + 545, + 646 + ], + "type": "text", + "content": "[62] Chao-Yuan Wu, Yanghao Li, Karttikeya Mangalam, Haoqi Fan, Bo Xiong, Jitendra Malik, and Christoph Feichtenhofer. MeMVit: Memory-augmented multiscale vision transformer for efficient long-term video recognition. In CVPR, 2022. 1, 2, 6, 7, 8" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 307, + 647, + 545, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 647, + 545, + 679 + ], + "spans": [ + { + "bbox": [ + 307, + 647, + 545, + 679 + ], + "type": "text", + "content": "[63] Chuhan Zhang, Ankush Gupta, and Andrew Zisserman. Temporal query networks for fine-grained video understanding. In CVPR, 2021. 2" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 307, + 680, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 680, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 307, + 680, + 545, + 712 + ], + "type": "text", + "content": "[64] Yubo Zhang, Pavel Tokmakov, Martial Hebert, and Cordelia Schmid. A structured model for action detection. In CVPR, 2019. 2" + } + ] + } + ], + "index": 32 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "18382" + } + ] + } + ], + "index": 34 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 205 + ], + "type": "list", + "angle": 0, + "index": 3, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "text", + "content": "[65] Jiaojiao Zhao, Yanyi Zhang, Xinyu Li, Hao Chen, Bing Shuai, Mingze Xu, Chunhui Liu, Kaustav Kundu, Yuanjun Xiong, Davide Modolo, et al. Tuber public code. https://github.com/amazon-science/tubelet-transformer, 2022.6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 129, + 287, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 129, + 287, + 173 + ], + "spans": [ + { + "bbox": [ + 48, + 129, + 287, + 173 + ], + "type": "text", + "content": "[66] Jiaojiao Zhao, Yanyi Zhang, Xinyu Li, Hao Chen, Bing Shuai, Mingze Xu, Chunhui Liu, Kaustav Kundu, Yuanjun Xiong, Davide Modolo, et al. TubeR: Tubelet Transformer for Video Action Detection. In CVPR, 2022. 1, 2, 4, 5, 7, 8" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 174, + 287, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 174, + 287, + 205 + ], + "spans": [ + { + "bbox": [ + 48, + 174, + 287, + 205 + ], + "type": "text", + "content": "[67] Xingyi Zhou, Dequan Wang, and Philipp Krahenbuhl. Objects as points. In arXiv preprint arXiv:1904.07850, 2019. 2" + } + ] + } + ], + "index": 2 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "18383" + } + ] + } + ], + "index": 4 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/End-to-End Temporal Action Detection with 1B Parameters Across 1000 Frames/aed9af76-c260-41b9-be21-a2e11ea6ab0b_content_list.json b/2024/End-to-End Temporal Action Detection with 1B Parameters Across 1000 Frames/aed9af76-c260-41b9-be21-a2e11ea6ab0b_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..1b07a86973a9ca5166bb843fae44b45421898e55 --- /dev/null +++ b/2024/End-to-End Temporal Action Detection with 1B Parameters Across 1000 Frames/aed9af76-c260-41b9-be21-a2e11ea6ab0b_content_list.json @@ -0,0 +1,1476 @@ +[ + { + "type": "text", + "text": "End-to-End Temporal Action Detection with 1B Parameters Across 1000 Frames", + "text_level": 1, + "bbox": [ + 78, + 130, + 890, + 151 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Shuming Liu1 Chen-Lin Zhang2 Chen Zhao1* Bernard Ghanem1 \n1King Abdullah University of Science and Technology (KAUST) 24Paradigm Inc", + "bbox": [ + 151, + 179, + 812, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 251, + 312, + 266 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recently, temporal action detection (TAD) has seen significant performance improvement with end-to-end training. However, due to the memory bottleneck, only models with limited scales and limited data volumes can afford end-to-end training, which inevitably restricts TAD performance. In this paper, we reduce the memory consumption for end-to-end training, and manage to scale up the TAD backbone to 1 billion parameters and the input video to 1,536 frames, leading to significant detection performance. The key to our approach lies in our proposed temporal-informative adapter (TIA), which is a novel lightweight module that reduces training memory. Using TIA, we free the humongous backbone from learning to adapt to the TAD task by only updating the parameters in TIA. TIA also leads to better TAD representation by temporally aggregating context from adjacent frames throughout the backbone. We evaluate our model across four representative datasets. Owing to our efficient design, we are able to train end-to-end on VideoMAEv2-giant and achieve $75.4\\%$ mAP on THUMOS14, being the first end-to-end model to outperform the best feature-based methods. Code is available at https://github.com/sming256/AdaTAD.", + "bbox": [ + 73, + 282, + 473, + 614 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 642, + 209, + 657 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Temporal Action Detection (TAD) plays a vital role in the understanding of long-form videos. Its objective is to pinpoint specific action instances within untrimmed videos, identifying their start and end times, along with their respective categories [15, 20, 44, 60, 62, 67, 71]. This task is crucial for various applications, including highlight detection [36, 64], video-language grounding [38, 48], and action spotting [3]. Though innovations in the detector design have made profound progress in the past years [30, 66, 67], recent research highlights two new trends: end-to-end training [10, 32, 69], and scaling up [56, 57].", + "bbox": [ + 75, + 667, + 468, + 833 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "End-to-end training in TAD refers to jointly training the video encoder and action detector [10, 29, 62, 70]. Compared to feature-based methods, end-to-end training offers", + "bbox": [ + 75, + 834, + 468, + 878 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/dae658e7cd312bba0b207001ec1b34728d4ec6ff778d8adbd6430ab83dafccda.jpg", + "image_caption": [ + "Figure 1. AdaTAD enjoys the benefit of end-to-end training and scaling up with efficient memory usage. The bubble size represents the number of the model's learnable parameters. Using SlowFast-R50, AdaTAD achieves better performance compared to E2E-TAD [33] with less memory. When further scaling up the model to VideoMAEv2-gaint and data to 1536 frames, we achieve an impressive avg. mAP of $75.4\\%$ on THUMOS14." + ], + "image_footnote": [], + "bbox": [ + 506, + 253, + 883, + 467 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "distinct advantages. First, it can effectively bridge the gap commonly found between pretraining and fine-tuning, such as data and task discrepancy. Second, video spatial augmentations can be utilized in the end-to-end setting, leading to further performance gain.", + "bbox": [ + 496, + 582, + 890, + 657 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Scaling up refers to improving performance by increasing the model size or the input data volume, and has demonstrated its effectiveness in various domains [14, 19, 39, 53]. In TAD, offline methods have attempted to scale up the feature extraction network to reach a higher performance. A notable example includes the work by Wang et al.[56], which reports a $10\\%$ increase in mean Average Precision (mAP) by scaling from VideoMAE-S to VideoMAEv2-giant, using ActionFormer [66] on THUMOS14 [25].", + "bbox": [ + 496, + 657, + 890, + 794 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Intuitively, combining the strengths of end-to-end training and scaling up is expected to be most beneficial for improving TAD performance. However, both strategies demand substantial GPU memory, which restricts end-to-end training to a small model [29, 62, 69], or a small input volume [10, 33]. As shown in Fig. 1, the performance of previous end-to-end methods still significantly lags behind the", + "bbox": [ + 496, + 795, + 893, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author.", + "bbox": [ + 94, + 886, + 220, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "18591", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "best results achieved by feature-based approaches. Additionally, current end-to-end methods in TAD use computationally intensive full fine-tuning, risking the issues of catastrophic forgetting and overfitting during transfer learning [50, 63]. These issues can result in less competitive performance, especially when the downstream datasets are small, which is a common scenario in the TAD domain.", + "bbox": [ + 75, + 90, + 468, + 196 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we aim to overcome the above limitations by harnessing the advantages of both scaling-up and end-to-end training. To achieve this, we introduce adapter tuning for temporal action detection (AdaTAD). Our method successfully trains a TAD model in an end-to-end manner, utilizing a backbone with 1 billion parameters and processing input videos of 1,536 frames. As illustrated in Fig. 1, to the best of our knowledge, this is the first end-to-end work that outperforms the best feature-based TAD methods.", + "bbox": [ + 75, + 198, + 468, + 333 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Specifically, we employ the following strategies to enhance the TAD performance while maintaining reasonable memory consumption. First, we identify that the snippet representation commonly used in feature-based methods is excessively redundant. In response, we have adopted a more memory-efficient frame-representation scheme, establishing an effective end-to-end baseline for TAD. Second, we adopt the parameter-efficient fine-tuning technique to minimize memory usage and mitigate overfitting in transfer learning. Notably, we introduce a novel Temporal-Informative Adapter (TIA). This adapter is injected between backbone layers and is the only learnable component in the backbone during fine-tuning. Different from conventional adapters [22], TIA is tailored for the TAD task and integrates temporal depth-wise convolutions to aggregate informative context from adjacent frames. Third, for additional memory efficiency, we propose a lighter variant of our method. By positioning the TIAs alongside the original backbone, rather than inside it, backpropagation can be done through the TIAs only. This configuration can further cut down on the need for intermediate activations within the primary backbone, thereby allowing us to scale up the model size and data size to unprecedented levels.", + "bbox": [ + 75, + 335, + 470, + 681 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "AdaTAD establishes a new state-of-the-art across multiple TAD datasets. Notably, our method achieves an impressive $75.4\\%$ mAP on THUMOS14, outperforming the previous feature-based best result of $71.5\\%$ by a large margin. This achievement underscores the possible paradigm shift in TAD, i.e., moving from traditional feature extraction plus offline detectors to scaling up end-to-end TAD training. We summarize our contribution as follows:", + "bbox": [ + 75, + 684, + 468, + 805 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. We introduce an efficient end-to-end framework for TAD, scaling up the model size to 1 billion parameters and the input data to 1,536 frames. We achieve a consistent performance improvement with the scaling up, shedding light on the importance of scaling for TAD.", + "2. We propose a novel temporal-informative adapter to re" + ], + "bbox": [ + 76, + 809, + 468, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "duce memory as well as aggregate the temporal context for TAD. Different variants of these adapters are designed to trade off between performance and memory cost. To the best of our knowledge, we are the first to introduce the adapter mechanism to TAD.", + "bbox": [ + 517, + 90, + 890, + 166 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3. Our method achieves state-of-the-art performance across four TAD datasets. Remarkably, this represents the first end-to-end approach that outperforms the previous feature-based methods by a large margin.", + "bbox": [ + 500, + 167, + 890, + 227 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 238, + 640, + 253 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Temporal Action Detection. Current methods for temporal action detection, also referred to as temporal action localization, can be broadly classified into three categories based on their architectural design: one-stage, two-stage, and DETR-based methods. One-stage methods directly localize actions from a multi-scale feature pyramid, such as ActionFormer [66] and TriDet [47]. These methods integrate action classification and temporal boundary regression in a single step [45, 61, 62]. Two-stage methods, in contrast, involve an additional step of proposal feature extraction [4, 30, 40, 58, 59, 68, 73]. For instance, VSGN [67] employs boundary sampling to further refine proposals. Recently, there is a growing interest in query-based methods [34, 46, 51], which deploy a set of learned queries to interact with the feature maps and directly predict the actions' temporal boundaries and categories.", + "bbox": [ + 496, + 263, + 890, + 503 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In addition to the aforementioned categories, TAD can also be divided into feature-based and end-to-end methods. The former relies on pre-extracted RGB features and optionally incorporates optical flow features. On the other hand, end-to-end methods take raw video frames as input and jointly optimize the video encoder and action detector [31]. Due to computational constraints, AFSD [29] downsamples the input to a resolution of $96^2$ . DaoTAD [55] and E2E-TAD [33] provide evidence that high TAD performance can be achieved by relying solely on the RGB modality with various data augmentations. Further innovations came from SGP [11], TALLFormer [10], and ETAD [32], all of which introduced strategies to backpropagate only through parts of the data. Additionally, $\\mathrm{Re}^2$ TAL [69] and $\\mathrm{Dr}^2$ Net[70] design reversible network architectures to release the memory occupied by intermediate activations. Despite these advancements, all above methods follow the full fine-tuning paradigm, and none has yet surpassed the best results achieved by feature-based approaches.", + "bbox": [ + 496, + 505, + 890, + 792 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Scaling Law in Deep Learning. Scaling up model and data has been a prevalent strategy across both computer vision and natural language processing fields to achieve superior performance. The GPT series [5, 39, 41, 42] has consistently demonstrated that larger models pretrained on extensive datasets yield significant gains in language understanding capabilities. Analogously, in the realms of image", + "bbox": [ + 496, + 794, + 890, + 901 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "18592", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/e48ed53730e24a77b65a5de34100c577145e91640c4268b5d41ae7d6013d1417.jpg", + "image_caption": [ + "Figure 2. Comparative illustration of our proposed TAD framework versus popular and widely used alternatives. (a) represents the typical offline method. (b) is the traditional end-to-end method using full fine-tuning. (c) Tailored for the TAD task, our AdaTAD uses a lightweight temporal-informative adapter inside the backbone to achieve efficient transfer learning. (d) To further reduce memory usage and scale up the model/data, $\\mathrm{AdaTAD}^{\\dagger}$ uses an alternative placement for adapters outside the backbone." + ], + "image_footnote": [], + "bbox": [ + 96, + 90, + 279, + 284 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/ed78bdde4dc9859c9e785005426cb4ab21e0450f9b2c312bab69f72b787c6052.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 325, + 90, + 449, + 286 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/8707905ec3b0e326193625e371ea4aef9d949529e19d6255d017968fc020bd4e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 486, + 90, + 609, + 285 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/f000f40ce7207fe11019c00cc44ec41eccf5abf7fdd97ad02acfd15a9593980e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 645, + 90, + 869, + 285 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "and video understanding, architectures such as ViT [14] and MViT [16], have also witnessed the effectiveness of this scaling strategy. Alabdulmohsin et al.[1] present a recipe for estimating scaling law parameters reliably from learning curves in computer vision. To attain even greater performance, several studies have also scaled up image resolution [35] or video clip length [56]. In this paper, we successfully apply this principle within the domain of temporal action detection and achieve state-of-the-art results.", + "bbox": [ + 75, + 364, + 472, + 500 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Efficient Transfer Learning. Transfer learning aims to adapt a pretrained model to a new domain. In TAD, typically off-the-shelf action recognition models are employed as the backbone, such as SlowFast [17]. Traditional transfer learning adopts full fine-tuning, meaning that all parameters of the pretrained model are updated. However, studies by [43, 63] have noted that full fine-tuning may harm the pre-learned knowledge, particularly when the downstream dataset is small and less comprehensive. Moreover, as models increase in size, the computational and storage demands of full fine-tuning proportionally increase.", + "bbox": [ + 75, + 503, + 468, + 671 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In response to these challenges, several works have investigated parameter-efficient tuning (PEFT) strategies that involve fine-tuning only a fraction of the network. For instance, Adapter [22] inserts lightweight modules analogous to feedforward networks in transformers, and only tunes these elements. LoRA [23] employs low-rank matrices in each transformer layer. Prefix-tuning [28] and prompt-tuning [27] attach learnable prompt tokens at the input stage or within each layer. In computer vision, many PEFT methods [6, 7, 24, 63] have also been explored across various tasks to optimize transfer learning efficiency. Our work represents the first effort to examine the potential of the PEFT mechanism in TAD.", + "bbox": [ + 75, + 672, + 468, + 867 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Although PEFT effectively reduces the number of learnable parameters, data-intensive and computationally heavy", + "bbox": [ + 76, + 869, + 468, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "tasks like video understanding require more memory-efficient techniques. To this end, several works try to externalize the trainable components from the backbone, eliminating the need for backpropagation through the extensive original model. For example, LST [49] introduces a supplementary lightweight network that operates in parallel with the main model. Similarly, $\\mathrm{E}^3\\mathrm{VA}$ [65] leverages intermediate features with adapters to enable efficient transfer learning while minimizing memory usage. Our work is inspired by these methods yet with a streamlined and simple design.", + "bbox": [ + 496, + 364, + 890, + 517 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Methodology", + "text_level": 1, + "bbox": [ + 500, + 530, + 635, + 547 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we introduce our AdaTAD step-by-step. We first introduce notations and study the efficient video representation to establish an end-to-end TAD baseline. Next, we introduce a temporal-informative adapter designed for efficient TAD. Finally, we propose an alternative placement for adapters to further alleviate computational demands.", + "bbox": [ + 496, + 556, + 890, + 647 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Notations", + "text_level": 1, + "bbox": [ + 500, + 657, + 609, + 672 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Temporal action detection can be formulated as follows: given an untrimmed video $\\mathbf{X} \\in \\mathbb{R}^{3 \\times H \\times W \\times T}$ , where $H$ and $W$ are the height and width of each frame, and $T$ is the frame number, its temporal action annotations can be denoted as $\\Psi_{g} = \\{\\varphi_{i} = (t_{s}, t_{e}, c)\\}_{i=1}^{N}$ , where $t_{s}, t_{e}, c$ are the start, end time and category of action instance $\\varphi_{i}$ , and $N$ is the number of ground truth actions. TAD aims to predict candidate proposal set $\\Psi_{p} = \\{\\hat{\\varphi}_{i} = (\\hat{t}_{s}, \\hat{t}_{e}, \\hat{c}, s)\\}_{i=1}^{M}$ to cover $\\Psi_{g}$ , and $s$ is the confidence score.", + "bbox": [ + 496, + 681, + 890, + 821 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Frame-level representation", + "text_level": 1, + "bbox": [ + 500, + 830, + 743, + 847 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our end-to-end TAD architecture comprises two main components: feature extraction and action detection. Following previous work [69], we select ActionFormer [66] as our ac", + "bbox": [ + 496, + 854, + 890, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "18593", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "tion detection head due to its robust performance across various datasets without much hyperparameter tuning. Next, we discuss two ways of encoding raw frames into representative features (feature extraction): snippet representation and frame representation.", + "bbox": [ + 75, + 90, + 468, + 165 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Snippets Representation. Snippet-based video representations are popular choices in offline feature extraction. The whole video is divided into several short snippets (or namely clips). Each snippet has a short temporal length, e.g., 16 frames, and different snippets can have overlapping frames. Thus, the video can be conceptualized as $T$ snippets, denoted by $\\mathbf{X} \\in \\mathbb{R}^{T \\times 3 \\times 16 \\times H \\times W}$ . Each snippet is processed through the video backbone, and spatial-temporal pooling is applied to extract one snippet feature. This processing yields the feature representation $\\mathbf{F} \\in \\mathbb{R}^{T \\times C}$ , where $C$ denotes the channel dimension of the pooled features.", + "bbox": [ + 75, + 167, + 467, + 332 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Frame Representation. In contrast to snippet-based representations, frame-based video representations consider the entire video as a singular snippet or clip, represented as $\\mathbf{X} \\in \\mathbb{R}^{1 \\times 3 \\times T \\times H \\times W}$ . Then, the whole frame sequence is fed into the video backbone, and only spatial pooling is employed to extract features [10, 29, 69]. For attention-based models such as VideoMAE [54], the video is chunked into multiple shorter clips to avoid extensive temporal attention.", + "bbox": [ + 75, + 333, + 467, + 453 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Although both representations have been employed in previous studies, a fair comparison between them has not yet been performed. To address this gap, we conduct a comparative analysis of the two representations under the same setting on THUMOS14, measuring their memory usage and detection mAP. The results in Table 1 indicate that frame representation has comparable or even better performance than snippet representation, yet with much smaller memory consumption. When the feature extraction backbones are frozen, frame representation yields superior results to snippet representation for both VideoMAE [53] and SlowFast [17] backbones. Only in the end-to-end set", + "bbox": [ + 75, + 454, + 467, + 633 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/0c22a7f450caeccceb112ee96bd7f7d7551c54481c0507259bec47c03251b792.jpg", + "table_caption": [ + "Table 1. Snippets representation vs frame representation. We use the end-to-end version of ActionFormer with two representations for comparison. The snippet input is $768 \\times 3 \\times 16 \\times 160 \\times 160$ , and the frame input is $1 \\times 3 \\times 768 \\times 160 \\times 160$ . * means activation checkpointing is utilized to avoid overflowing GPU memory." + ], + "table_footnote": [], + "table_body": "
SettingBackboneRepr.Avg. mAPMem (GB)
FrozenVideoMAE-SFrame59.351.9
Snippet57.6813.2
SlowFast-R101Frame61.343.6
Snippet60.2417.2
End to EndVideoMAE-SFrame67.152.8*
Snippet68.4624.6*
SlowFast-R101Frame65.335.5*
Snippet66.7251.6*
", + "bbox": [ + 78, + 737, + 475, + 896 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "ting can the snippet representation achieve $1\\%$ mAP advantage over frame representations; however, it requires 8 times more memory consumption. Taking into account both performance and memory usage, frame-based representations could be a better choice for end-to-end TAD development.", + "bbox": [ + 496, + 90, + 890, + 166 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Therefore, we use frame representation as the default baseline to encode videos in our experiments. Following the previous TALLFormer work [10], we also incorporate activation checkpointing [8] and mixed precision training [37] to fully harness the potential of scaling.", + "bbox": [ + 496, + 167, + 890, + 242 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Temporal-Informative Adapter", + "text_level": 1, + "bbox": [ + 498, + 253, + 774, + 268 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In Section 3.2, we have built a simple end-to-end baseline using full fine-tuning. However, the baseline still suffers from two aspects: 1. Increased computational cost. In Table 1, we only use small video backbones like VideoMAES. When scaling VideoMAE-S to larger models, the computational burden and memory cost will grow rapidly. 2. Inferior transfer learning ability. More critically, the baseline follows the tradition of full fine-tuning, which may lead to inferior transfer learning. Pointed out by [50, 63], full fine-tuning may result in overfitting or forgetting, especially for large pretrained models. If downstream datasets are not sufficiently diverse, full fine-tuning can even destroy the powerful features learned from large-scale pretraining. Motivated by the above two aspects, we apply the PEFT mechanism and propose to fine-tune a plug-and-play module named Temporal-Informative Adapter (TIA) to achieve efficient and effective transfer learning for TAD.", + "bbox": [ + 496, + 277, + 890, + 532 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We first review the architecture of the standard adapter proposed by [22]. As formulated in Equation 1, the standard adapter includes a down-projection fully connected (FC) layer with parameter $\\mathbf{W}_{\\mathrm{down}} \\in \\mathbb{R}^{d \\times \\frac{d}{\\gamma}}$ , where $\\frac{d}{\\gamma}$ represents the intermediate dimension and satisfies $\\gamma > 1$ . Then, an up-projection layer $\\mathbf{W}_{\\mathrm{up}} \\in \\mathbb{R}^{\\frac{d}{\\gamma} \\times d}$ is employed to restore the channel dimension. Between these two FC layers, a non-linear activation function $\\sigma$ is inserted, such as GELU [21]. Afterward, a residual connection is added to the output of the projection layer. Note that $\\mathbf{x}$ and $\\mathbf{x}'$ are the input and output features with the same shape $\\mathbb{R}^{d \\times t \\times h \\times w}$ .", + "bbox": [ + 496, + 535, + 890, + 705 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {x} ^ {\\prime} = \\boldsymbol {W} _ {\\mathrm {u p}} ^ {\\top} \\cdot \\boldsymbol {\\sigma} \\left(\\boldsymbol {W} _ {\\mathrm {d o w n}} ^ {\\top} \\cdot \\boldsymbol {x}\\right) + \\boldsymbol {x}. \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 584, + 715, + 890, + 736 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Although the adapter has achieved great success in natural language processing and computer vision, the standard adapter only focuses on adapting channel information, which neglects the temporal context vital for the TAD task. To address this limitation, we introduce the temporalinformative adapter, as depicted in Fig. 3(b).", + "bbox": [ + 496, + 750, + 890, + 839 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The architecture of TIA follows the general bottleneck design of the standard adapter, while we integrate the temporal depth-wise convolution layers, as described in Equation 2. The temporal convolution with a kernel size of $k$ is", + "bbox": [ + 496, + 839, + 890, + 898 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "18594", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/bca175eac0c4d2d439c5b27d1a329dee4d29c7558b725de0e20f59adf28392de.jpg", + "image_caption": [ + "(a) Standard Adapter", + "Figure 3. Architecture of (a) standard adapter and (b) our temporal-informative adapter. We incorporate temporal depth-wise convolution to aggregate context from adjacent frames." + ], + "image_footnote": [], + "bbox": [ + 104, + 88, + 240, + 257 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/5b95fa8e623d1dee88ffdb06ef8f5387be369eeb3524eeb65e87242fa5349dac.jpg", + "image_caption": [ + "(b) Temporal-Informative Adapter" + ], + "image_footnote": [], + "bbox": [ + 281, + 88, + 439, + 257 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "designed to aggregate local informative context from adjacent frames and to enrich the representation of the current time step. Practically, this is achieved through the application of a 3D convolution with kernel size $(k,1,1)$ and group size $\\frac{d}{\\gamma}$ for depth-wise processing. Additionally, an FC layer with weight $W_{\\mathrm{mid}}\\in \\mathbb{R}^{\\frac{d}{\\gamma}\\times \\frac{d}{\\gamma}}$ is employed to facilitate information exchange across channels. At last, a learnable scalar $\\alpha$ is introduced to adjust the amplitude of adapter's output.", + "bbox": [ + 75, + 340, + 472, + 470 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\bar {\\boldsymbol {x}} = \\boldsymbol {\\sigma} (\\boldsymbol {W} _ {\\text {d o w n}} ^ {\\top} \\cdot \\boldsymbol {x}),\n$$\n", + "text_format": "latex", + "bbox": [ + 163, + 479, + 303, + 497 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\boldsymbol {x}} = \\boldsymbol {W} _ {\\mathrm {m i d}} ^ {\\top} \\cdot \\mathbf {D W C o n v} _ {k} (\\bar {\\boldsymbol {x}}) + \\bar {\\boldsymbol {x}}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 166, + 500, + 468, + 518 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {x} ^ {\\prime} = \\alpha \\cdot \\boldsymbol {W} _ {\\mathrm {u p}} ^ {\\top} \\cdot \\hat {\\boldsymbol {x}} + \\boldsymbol {x}.\n$$\n", + "text_format": "latex", + "bbox": [ + 163, + 521, + 315, + 540 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As shown in Fig. 2(c), TIA is designed to be inserted between different backbone layers, e.g. between each ViT block of VideoMAE or each bottleneck block of SlowFast. To ensure the newly added connection does not affect the original network at the beginning of transfer learning, the weight and bias of the adapter's last projection layer $W_{\\mathrm{up}}$ are initialized to 0. The learnable coefficient $\\alpha$ is initialized to 1. The temporal kernel size $k$ is set to 3, and the channel downsampling ratio $\\gamma$ is set to 4 by default. Under these settings, the additional trainable parameters coming from TIA only account for $4.7\\%$ of the total parameters of the original backbone. Since this backbone is frozen when TIA is used, our proposed strategy constitutes a massive reduction in trainable parameters as compared to full fine-tuning. Our experiments show that TIA can achieve better performance than full fine-tuning with less memory usage.", + "bbox": [ + 75, + 550, + 468, + 792 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Alternative Placement for Adapter", + "text_level": 1, + "bbox": [ + 76, + 801, + 379, + 816 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Although the previously described PEFT approach can reduce tunable parameters and memory usage, the gradient still needs to backpropagate over the entire backbone during training. This requirement limits our ability to scale-up the model size or input data size further. As highlighted in prior", + "bbox": [ + 75, + 824, + 468, + 902 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "works [49, 65], if we can stop the gradient backpropagation within the original backbone, additional computational savings can be achieved.", + "bbox": [ + 496, + 90, + 890, + 135 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Inspired by this insight, we propose a placement strategy for adapters that position them externally to the backbone, rather than inserting them inside. As illustrated in Fig. 2(d), we utilize the previously introduced TIA module, but its output does not feed back into the middle of the original backbone. It is directly added to the backbone's final layer. This configuration eliminates the need for backpropagation through the original network, as gradients are only tracked to the shallow lightweight adapter. To further diminish computation, we observe that adapting only the last half of backbone layers yields comparable performance while reducing half of the adaptation cost.", + "bbox": [ + 496, + 136, + 892, + 318 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To distinguish the different variants, we name the standard adaption design as AdaTAD, and the alternative placement as AdaTAD†. The latter can be considered as a lite version of the former. Compared to directly injecting adapters into the backbone, AdaTAD† may lead to a slight performance drop. However, it enables us to leverage richer models and more data, which should effectively counter this possible drop.", + "bbox": [ + 496, + 318, + 890, + 439 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 500, + 454, + 633, + 470 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Datasets and Metrics", + "text_level": 1, + "bbox": [ + 500, + 479, + 697, + 494 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We choose ActivityNet-1.3 [20], THUMOS14 [25], and Epic-Kitchens 100 [13] to evaluate our proposed approach. ActivityNet-1.3 and THUMOS14 are web-collected third-person untrimmed videos, consisting of 19,994 and 413 videos, respectively. EPIC-Kitchens 100 is collected from 700 egocentric videos. Since the action categories of EPIC-Kitchens 100 are more domain-specific and different from common pretraining data, achieving higher performance on this dataset is more challenging. Moreover, we also evaluate our method on the Ego4D-MQ [18] dataset, and the results can be found in the appendix.", + "bbox": [ + 496, + 502, + 890, + 669 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Following common practice, we report the mean Average Precision (mAP) at certain IoU thresholds and average mAP as the evaluation metrics. On ActivityNet1.3, the IoU thresholds are chosen from 0.5 to 0.95 with 10 steps. On THUMOS14, the threshold is chosen from $\\{0.3,0.4,0.5,0.6,0.7\\}$ . On EPIC-Kitchens 100, the threshold is set to $\\{0.1,0.2,0.3,0.4,0.5\\}$ .", + "bbox": [ + 496, + 670, + 890, + 776 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Implementation Details", + "text_level": 1, + "bbox": [ + 500, + 786, + 717, + 801 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We implement our method with PyTorch 2.0 and MMAc-tion2 [12] with 4 A100 GPUs. By default, mixed-precision training and activation checkpointing are adopted to save memory. We use ActionFormer [66] as our detection head, and keep the hyper-parameters unchanged on each dataset. The learning rate for the adapter in the backbone is grid", + "bbox": [ + 496, + 809, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "18595", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/4b56a1ded48a8dc87f26ec229e00c7e873d97dfc78dec537c7394e88c374aac6.jpg", + "table_caption": [ + "Table 2. Results on ActivityNet-1.3 and THUMOS14, measured by mAP (\\%) at different tIoU thresholds. E2E refers to end-to-end training, and Mem refers to memory usage (GB) per video. On ActivityNet-1.3, our prediction is combined with CUHK [72] classification results. Specifically, * means we employ stronger video-level classification results used in InternVideo [57] for a fair comparison. We report our best results in bold, and the previous best results in underline, which was achieved by the feature-based method. The last row is achieved when only the last half of backbone layers are adapted; otherwise, full-layer adaptation will lead to out-of-memory on A100-80G." + ], + "table_footnote": [], + "table_body": "
MethodBackboneE2EFlowMemActivityNet-1.3THUMOS14
0.50.750.95Avg.0.30.40.50.60.7Avg.
BMN [30]TSNX✓-50.0734.788.2933.8556.047.438.829.720.538.5
TadTR [34]I3DX✓-49.1032.608.5032.3062.457.449.237.826.346.6
ActionFormer [66]SlowFast-R50XX-54.2637.048.1336.0278.773.365.254.639.762.3
ActionFormer [66]I3DX✓-53.5036.208.2035.6082.177.871.059.443.966.8
ASL [45]I3DX✓-54.1067.408.0036.2083.179.071.759.745.867.9
TriDet [47]I3DX✓-54.7038.008.4036.8083.680.172.962.447.469.3
VideoMAEv2 [56]VideoMAEv2-gXX----------69.6
InternVideo [57]VideoMAE-H+UniformerV2XX----39.00*-----71.5
AFSD [29]I3D✓✓1252.4035.306.5034.4067.362.455.543.731.152.0
E2E-TAD [33]SlowFast-R50✓X1250.4735.9910.3335.1069.464.356.046.434.954.2
BasicTAD [62]SlowOnly-R50✓X1251.2033.417.5733.1275.570.863.550.937.459.6
TALLFormer [10]VideoSwin-B✓X2954.1036.207.9035.6076.0-63.2-34.559.2
Re2TAL [69]Re2VideoSwin-T✓X2454.7537.819.0336.8077.071.562.449.736.359.4
AdaTADSlowFast-R50✓X4.355.2838.118.8737.1181.076.269.459.044.566.0
AdaTADVideoMAE-S✓X2.556.1538.999.0737.8584.580.271.660.946.968.8
AdaTADVideoMAE-B✓X4.956.7739.359.7138.3987.082.475.363.849.271.5
AdaTADVideoMAE-L✓X11.057.6940.5610.1339.2287.784.176.766.452.473.5
AdaTADVideoMAE-H✓X19.258.0440.559.7539.3788.985.378.666.952.574.4
AdaTADVideoMAEv2-g✓X29.958.4541.1610.4539.7989.585.878.967.352.674.8
AdaTAD† (1536×2242)VideoMAEv2-g✓X43.660.8242.699.8441.15*89.685.979.467.653.875.4
AdaTAD (1536×2242)VideoMAEv2-g✓X50.661.7243.3510.8541.93*89.786.780.971.056.176.9
", + "bbox": [ + 76, + 164, + 895, + 498 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "searched from 5e-4 to 5e-5, and other parameters inside the backbone are frozen. On ActivityNet-1.3, we resize the video into a fixed length of 768 frames. On THUMOS14, we randomly truncate a window with 768 frames with a temporal stride of 4. On EPIC-Kitchens 100, we randomly truncate a window with 6144 frames with a temporal stride of 2. After the video encoder, the feature is resized to fixed lengths of 192, 768, and 768, respectively, for the three datasets. Frame resolution is set to $160^2$ by default. In all experiments, we report the training memory usage. More implementation details can be found in the appendix.", + "bbox": [ + 75, + 513, + 470, + 681 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Comparison with SoTA Methods", + "text_level": 1, + "bbox": [ + 76, + 694, + 366, + 709 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 2 compares our AdaTAD with other state-of-the-art (SoTA) methods on ActivityNet-1.3 and THUMOS14 datasets. Initially, we use SlowFast-R50 as the backbone. For comparison, we also extract corresponding offline features, utilizing the snippet representation where each snippet comprises 32 frames with $224^{2}$ resolution. We observe that end-to-end training enhances performance from $62.3\\%$ to $66.0\\%$ on THUMOS14. Notably, this architecture has also been used in E2E-TAD [33]. However, our method consumes less memory while achieving superior performance. This apple-to-apple comparison underscores the benefits of adapter tuning and the scaling-up principle.", + "bbox": [ + 75, + 719, + 470, + 902 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/2ce76e016cb6cbce31a5597cbc8d94889d85a94b1430394e3ea8532a2101fd0e.jpg", + "table_caption": [ + "Table 3. Results on EPIC-Kitchens 100 validation set. For comparison, the feature-based methods use the same SlowFast-R50." + ], + "table_footnote": [], + "table_body": "
MethodE2E0.10.20.30.40.5Avg.
Verb Task
BMN [30]✘10.88.88.47.15.68.4
G-TAD [60]✘12.111.09.48.16.59.4
ActionFormer [66]✘26.625.424.222.319.123.5
ASL [45]✘27.9-25.5-19.824.6
TriDet [47]✘28.627.426.124.220.825.4
AdaTAD (SlowFast-R50)✓26.525.723.921.717.623.1
ActionFormer (VideoMAE-L)✘32.731.629.126.723.628.7
AdaTAD (VideoMAE-L)✓33.132.230.427.523.129.3
Noun Task
BMN [30]✘10.38.36.24.53.46.5
G-TAD [60]✘11.010.08.67.05.48.4
ActionFormer [66]✘25.224.122.720.517.021.9
ASL [45]✘26.0-23.4-17.722.6
TriDet [47]✘27.426.324.622.218.323.8
AdaTAD (SlowFast-R50)✓24.523.622.320.016.521.4
ActionFormer (VideoMAE-L)✘31.329.727.225.321.326.9
AdaTAD (VideoMAE-L)✓32.431.630.127.424.629.3
", + "bbox": [ + 500, + 545, + 893, + 830 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Furthermore, when adopting the VideoMAE [53] family as our backbone and progressively scaling up the model size, the performance of AdaTAD consistently improves. Using the largest model, i.e., VideoMAEv2-giant with 1.01", + "bbox": [ + 498, + 839, + 893, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "18596", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/ab3e0d25e601d226c26ca0b696bd7093dc67e580e58494b129a98d2402f0956d.jpg", + "table_caption": [ + "Table 4. Compared to full fine-tuning, our adapter tuning can achieve better performance with less memory. Param. is the number of tunable parameters in the backbone. * means out of memory on A100-80GB, and we report the estimated number." + ], + "table_footnote": [], + "table_body": "
ModelSettingE2EParam.Mem.mAP
VideoMAE-SFeature×0-57.6
Snippet Full FT✓22M24.6G68.4
Frame Full FT✓22M2.8G67.1
AdaTAD✓1M2.5G68.8
VideoMAE-BFeature×0-64.7
Snippet Full FT✓86M87.4G*-
Frame Full FT✓86M5.6G70.1
AdaTAD✓4M4.9G71.5
VideoMAE-LFeature×0-66.5
Snippet Full FT✓304M193G*-
Frame Full FT✓304M13.1G73.0
AdaTAD✓14M11.0G73.5
", + "bbox": [ + 76, + 148, + 475, + 358 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "billion parameters, and larger input data, i.e., 1536 frames with $224^{2}$ resolution, we attain an impressive $41.9\\%$ mAP on ActivityNet-1.3 and $76.9\\%$ mAP on THUMOS14. It is noteworthy that the previous SoTA was achieved by VideoMAEv2 [56] and InternVideo [57], which utilize the same detector head as ours but with offline snippet features. Our method surpasses these in detection performance by a large margin, marking the first instance where an end-to-end TAD method can outperform SoTA feature-based results.", + "bbox": [ + 75, + 367, + 468, + 502 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We also present our results on EPIC-Kitchens 100 in Table 3. Since videos in this dataset have a longer duration, all previous methods rely solely on pre-extracted features [47, 52, 66]. Our approach is the first to adopt end-to-end training on this dataset. For fair comparison, we first utilize the same backbone as used in previous methods, i.e., SlowFast-R50 pretrained on EPIC, and we achieve comparable performance to ActionFormer [66]. Moreover, when we scale up the backbone to VideoMAE-L (it is also trained on EPIC-Kitchens 100 classification task), we achieve SoTA performance of $29.3\\%$ .", + "bbox": [ + 75, + 503, + 468, + 670 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Ablation and Analysis", + "text_level": 1, + "bbox": [ + 76, + 678, + 284, + 694 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In this section, we present a series of analyses to evaluate our proposed method and affirm the benefits of scaling up in TAD. Unless otherwise stated, our experiments utilize a standard input of 768 frames per video on THUMOS14.", + "bbox": [ + 75, + 700, + 468, + 762 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The advantage of adapter tuning. In Table 4, we compare conventional full fine-tuning with our proposed design. It is evident that end-to-end approaches significantly outperform pre-extracted features. Moreover, with full finetuning, the snippet representation slightly advances over frame representation but incurs tremendous memory costs, which aligns with our analysis in Section 3.2. However, AdaTAD uses less memory and still achieves better performance than conventional full fine-tuning. This also veri", + "bbox": [ + 75, + 763, + 468, + 901 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/3723db6c851f532d60de1a17b74361000d52030d796f488a242cb41f76d60ff4.jpg", + "table_caption": [ + "Table 5. When scaling up the input data, AdaTAD's performance consistently increases. * means snippet representation is used in offline feature extraction, and each snippet has 16 frames." + ], + "table_footnote": [], + "table_body": "
SettingModelResolutionFramesMem.mAP
FeatureVideoMAEv2-g2242768x16*-69.6
AdaTADVideoMAE-S16027682.5G68.8
160215363.8G69.7
160230726.5G70.6
22427683.8G70.7
224215366.4G71.0
2242307211.6G71.5
", + "bbox": [ + 501, + 135, + 890, + 272 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/dec7dfa864988cc16197e0482bf45af796b2fe97e4cc7f3b8394b8e997b74c3b.jpg", + "table_caption": [ + "Table 6. AdaTAD† can further push the boundaries of scaling up. OOM means out of memory on A100-80GB." + ], + "table_footnote": [], + "table_body": "
SettingModelResolutionFrameMem.mAP
AdaTADVideoMAE-L160276811.0G73.5
VideoMAEv2-g160276829.9G74.8
VideoMAEv2-g22421536OOM-
AdaTAD†VideoMAEv2-g160276822.8G73.7
VideoMAEv2-g224276830.0G74.6
VideoMAEv2-g2242153643.6G75.4
", + "bbox": [ + 501, + 311, + 890, + 430 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "fies the limitations of full fine-tuning, as discussed in Section 3.3. Specifically, our method enhances VideoMAE-S backbone with an $11.2\\%$ gain using only 1M trainable parameters. Additionally, Table 4 also demonstrates that scaling up the model size of the video backbone is an effective way to improve TAD performance.", + "bbox": [ + 496, + 439, + 890, + 531 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The advantage of scaling up the data. In addition to model scaling, Table 5 verifies the effectiveness of data scaling, which involves two aspects: frame number and frame resolution. Firstly, given the same video duration, increasing the frame number from 768 to 3072 can raise the mAP from $68.8\\%$ to $70.6\\%$ . In the meantime, the memory usage is nearly three times larger. Secondly, increasing the frame resolution from $160^2$ to $224^2$ also improves the mAP. In the end, by only scaling up the data, we elevate the mAP from $68.8\\%$ to $71.5\\%$ , already surpassing the current SoTA feature-based approach with a giant backbone model [57].", + "bbox": [ + 496, + 534, + 892, + 700 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Moreover, increasing the frame resolution from $160^2$ to $224^2$ or increasing the frame number from 768 to 1536 results in the same memory usage of 3.8G. However, the former achieves $70.7\\%$ mAP while the latter only reaches $69.7\\%$ . This suggests that frame resolution may be prioritized under the same memory budget, for the TAD task.", + "bbox": [ + 496, + 700, + 890, + 791 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The advantage of AdaTAD†. Given the effectiveness of scaling up the model or data, we further explore combining these approaches. In Table 6, using 768 frames while scaling up the model to VideoMAEv2-giant results in memory usage escalating to 29.9G. In such a scenario, further increasing the data could easily lead to memory overflow, even with the A100-80GB GPU. This indicates that adapta", + "bbox": [ + 496, + 794, + 892, + 901 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "18597", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "tion tuning reached its limit under this extreme case. Therefore, to utilize both the largest models with larger data simultaneously, $\\mathrm{AdaTAD}^{\\dagger}$ shows its advantage.", + "bbox": [ + 75, + 90, + 468, + 136 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Concretely, when switching from AdaTAD to AdaTAD†, the memory usage of the VideoMAEv2-giant model is reduced from 29.9G to 22.8G. Although a slight performance drop is observed, its reduced memory footprint enables scaling up data from 768 frames to as much as 1536 frames with a high resolution of $224^2$ . This scalability helps mitigate the performance drop and achieves a higher mAP of $75.4\\%$ .", + "bbox": [ + 75, + 136, + 467, + 242 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The ablation of the adapter design. Detailed in Table 7, we compare different adapter architectural designs. The baseline, i.e., offline snippet feature, achieves $64.7\\%$ mAP. End-to-end learning in all designs yields at least a $5\\%$ improvement. Our AdaTAD achieves $71.5\\%$ in the end. Compared to standard adapter [22], ours consumes similar memory but achieves higher mAP. This verifies that local temporal context is vital for the TAD task. In contrast to full finetuning (FT), we tune only 4M parameters using less memory. In our design, we find that removing the residual connection of the depth-wise convolution drops performance by $0.7\\%$ , and training becomes unstable. We also implement an adaptation design proposed in LongLoRA [9], which efficiently computes long-range attention and shows decent performance but requires more parameters and memory.", + "bbox": [ + 75, + 244, + 467, + 472 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/108e6f6346040b967e30f4ea1baaee4c38c22881057c907e38e94acb5b578ead.jpg", + "table_caption": [ + "Table 7. Ablation of different adapter architectural designs. VideoMAE-B is used to conduct the following experiments." + ], + "table_footnote": [], + "table_body": "
SettingE2EParam.Mem.mAPgains
Snippet FeatureX0-64.7
+ Full FT✓86M5.6G70.1+5.1
+ LongLoRA [9]✓28M6.2G71.1+6.1
+ Standard Adapter [22]✓3.6M4.8G70.2+5.2
+ AdaTAD (w/o residual)✓4.0M4.9G70.8+5.8
+ AdaTAD✓4.0M4.9G71.5+6.5
", + "bbox": [ + 78, + 515, + 460, + 636 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The necessity of end-to-end training for TAD. As previously discussed, end-to-end training can address discrepancies between the pretraining and fine-tuning stages in terms of training data and learning tasks. To corroborate this, we employ models pretrained on different datasets for the EPIC-Kitchens TAD task in Table 8. Kinetics-400 (K400) [26] represents commonly collected third-person web data and exhibits a large domain gap compared to EPIC-Kitchens 100. Using K400 for pretraining, we observe that end-to-end TAD training allows for $+5.56$ gain. Conversely, using a model already finetuned on EPIC-Kitchens still yields a $+2.32$ improvement. Unlike K400 pretraining, since this model has already adapted to the data discrepancy, we can infer that this gain leverages differences between the classification task in pretraining and the detection task in fine-tuning. Such improvements further underscore the significance of end-to-end training in TAD.", + "bbox": [ + 75, + 643, + 468, + 900 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/e6712dad19a7ac7947df5d4a5745da28320bd84010135ad315a4dd71653f4418.jpg", + "table_caption": [ + "Table 8. End-to-end TAD can alleviate the discrepancies between pretraining and finetuning. VideoMAE-L with different pretrained weights are used on the EPIC-Kitchens 100 Noun task." + ], + "table_footnote": [], + "table_body": "
Pretrain DatasetE2E0.10.30.5mAPgain
K400 [26]X18.6916.3511.5215.77
K400 [26]✓24.3322.1416.8721.33+ 5.56
K400 [26] + EPIC [13]X31.3227.2521.3326.98
K400 [26] + EPIC [13]✓32.4130.1324.5929.30+ 2.32
", + "bbox": [ + 501, + 135, + 893, + 227 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.5. Error Analyses", + "text_level": 1, + "bbox": [ + 500, + 232, + 653, + 247 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We also conduct false positive analysis at tIoU=0.5 in Fig. 4. Compared to feature-based training, learning from raw frames produces more helpful positive detections. More importantly, the percentage of wrong label error is reduced after end-to-end training, suggesting its unique advantage in classifying accurate action labels.", + "bbox": [ + 496, + 255, + 890, + 345 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/f12d6fe5149da566afa57be00d00b5ae3e7c600bcfcdc598394dfe43272757c7.jpg", + "image_caption": [ + "Figure 4. False Positive Profiling on THUMOS14 using [2]. We use VideoMAEv2-giant as the backbone, and compare end-to-end training with pre-extracted feature-based training." + ], + "image_footnote": [], + "bbox": [ + 513, + 349, + 885, + 491 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusions", + "text_level": 1, + "bbox": [ + 500, + 542, + 625, + 556 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This work introduces a memory-efficient and parameter-efficient end-to-end method named AdaTAD. Our key innovation lies in the proposed temporal-informative adapter, which is tailored for TAD with low computation costs. Furthermore, we design an alternative placement for adapters to minimize memory usage. By demonstrating the feasibility and effectiveness of scaling up end-to-end TAD, our work achieves new SoTA performance across multiple datasets. Particularly, this is the first instance of an end-to-end TAD method that surpasses the current best feature-based models. In fact, AdaTAD achieves a groundbreaking $75.4\\%$ mAP on THUMOS14. We believe our work underscores the possible paradigm shift in TAD, advocating a move away from the traditional methodology of separate feature extraction and offline detection towards a more integrated approach of scaling up end-to-end TAD training.", + "bbox": [ + 496, + 568, + 890, + 809 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgement. This work was supported by the King Abdullah University of Science and Technology (KAUST) Office of Sponsored Research through the Visual Computing Center (VCC) funding, as well as the SDAIA-KAUST Center of Excellence in Data Science and Artificial Intelligence (SDAIA-KAUST AI).", + "bbox": [ + 496, + 810, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "18598", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Ibrahim M Alabdulmohsin, Behnam Neyshabur, and Xiaohua Zhai. Revisiting neural scaling laws in language and vision. In NeurIPS, 2022. 3", + "[2] Humam Alwassel, Fabian Caba Heilbron, Victor Escorcia, and Bernard Ghanem. Diagnosing error in temporal action detectors. In ECCV, 2018. 8", + "[3] Humam Alwassel, Fabian Caba Heilbron, and Bernard Ghanem. Action search: Spotting actions in videos and its application to temporal action localization. In ECCV, 2018. 1", + "[4] Yueran Bai, Yingying Wang, Yunhai Tong, Yang Yang, Qiyue Liu, and Junhui Liu. Boundary content graph neural network for temporal action proposal generation. In ECCV, 2020. 2", + "[5] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. In NeurIPS, 2020. 2", + "[6] Hao Chen, Ran Tao, Han Zhang, Yidong Wang, Wei Ye, Jindong Wang, Guosheng Hu, and Marios Savvides. Convadapter: Exploring parameter efficient transfer learning for convnets. arXiv preprint arXiv:2208.07463, 2022. 3", + "[7] Shoufa Chen, Chongjian Ge, Zhan Tong, Jiangliu Wang, Yibing Song, Jue Wang, and Ping Luo. Adaptformer: Adapting vision transformers for scalable visual recognition. In NeurIPS, 2022. 3", + "[8] Tianqi Chen, Bing Xu, Chiyuan Zhang, and Carlos Guestrin. Training deep nets with sublinear memory cost. arXiv preprint arXiv:1604.06174, 2016. 4", + "[9] Yukang Chen, Shengju Qian, Haotian Tang, Xin Lai, Zhi-jian Liu, Song Han, and Jiaya Jia. Longlora: Efficient fin-tuning of long-context large language models. arXiv preprint arXiv:2309.12307, 2023. 8", + "[10] Feng Cheng and Gedas Bertasius. Tallformer: Temporal action localization with long-memory transformer. In ECCV, 2022. 1, 2, 4, 6", + "[11] Feng Cheng, Mingze Xu, Yuanjun Xiong, Hao Chen, Xinyu Li, Wei Li, and Wei Xia. Stochastic backpropagation: A memory efficient strategy for training video models. In CVPR, 2023. 2", + "[12] MMAction2 Contributors. Openmmlab's next generation video understanding toolbox and benchmark. https://github.com/open-mmlab/mmaction2, 2020.5", + "[13] Dima Damen, Hazel Doughty, Giovanni Maria Farinella, Sanja Fidler, Antonino Furnari, Evangelos Kazakos, Davide Moltisanti, Jonathan Munro, Toby Perrett, Will Price, et al. Scaling egocentric vision: The epic-kitchens dataset. In ECCV, 2018. 5, 8", + "[14] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2021. 1, 3", + "[15] Victor Escorcia, Fabian Caba Heilbron, Juan Carlos Niebles," + ], + "bbox": [ + 78, + 114, + 470, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "and Bernard Ghanem. DAPs: Deep action proposals for action understanding. In ECCV, 2016. 1", + "[16] Haoqi Fan, Bo Xiong, Karttikeya Mangalam, Yanghao Li, Zhicheng Yan, Jitendra Malik, and Christoph Feichtenhofer. Multiscale vision transformers. In ICCV, 2021. 3", + "[17] Christoph Feichtenhofer, Haoqi Fan, Jitendra Malik, and Kaiming He. SlowFast networks for video recognition. In ICCV, 2019. 3, 4", + "[18] Kristen Grauman, Andrew Westbury, Eugene Byrne, Zachary Chavis, Antonino Furnari, Rohit Girdhar, Jackson Hamburger, Hao Jiang, Miao Liu, Xingyu Liu, et al. Ego4d: Around the world in 3,000 hours of egocentric video. In CVPR, 2022. 5", + "[19] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, and Ross Girshick. Masked autoencoders are scalable vision learners. In CVPR, 2022. 1", + "[20] Fabian Caba Heilbron, Victor Escorcia, Bernard Ghanem, and Juan Carlos Niebles. ActivityNet: A large-scale video benchmark for human activity understanding. In CVPR, 2015. 1, 5", + "[21] Dan Hendrycks and Kevin Gimpel. Gaussian error linear units (gelus). arXiv preprint arXiv:1606.08415, 2016. 4", + "[22] Neil Houlsby, Andrei Giurgiu, Stanislaw Jastrzebski, Bruna Morrone, Quentin De Laroussilhe, Andrea Gesmundo, Mona Attariyan, and Sylvain Gelly. Parameter-efficient transfer learning for nlp. In ICML, 2019. 2, 3, 4, 8", + "[23] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. In ICML, 2021. 3", + "[24] Menglin Jia, Luming Tang, Bor-Chun Chen, Claire Cardie, Serge Belongie, Bharath Hariharan, and Ser-Nam Lim. Visual prompt tuning. In ECCV, 2022. 3", + "[25] YG Jiang, J Liu, A Roshan Zamir, G Toderici, I Laptev, M Shah, and R Sukthankar. Thumos challenge: Action recognition with a large number of classes, 2014. 1, 5", + "[26] Will Kay, Joao Carreira, Karen Simonyan, Brian Zhang, Chloe Hillier, Sudheendra Vijayanarasimhan, Fabio Viola, Tim Green, Trevor Back, Paul Natsev, et al. The kinetics human action video dataset. arXiv preprint arXiv:1705.06950, 2017.8", + "[27] Brian Lester, Rami Al-Rfou, and Noah Constant. The power of scale for parameter-efficient prompt tuning. In EMNLP, 2021. 3", + "[28] Xiang Lisa Li and Percy Liang. Prefix-tuning: Optimizing continuous prompts for generation. In ACL, 2021. 3", + "[29] Chuming Lin, Chengming Xu, Donghao Luo, Yabiao Wang, Ying Tai, Chengjie Wang, Jilin Li, Feiyue Huang, and Yanwei Fu. Learning salient boundary feature for anchor-free temporal action localization. In CVPR, 2021. 1, 2, 4, 6", + "[30] Tianwei Lin, Xiao Liu, Xin Li, Errui Ding, and Shilei Wen. BMN: boundary-matching network for temporal action proposal generation. In ICCV, 2019. 1, 2, 6", + "[31] Qinying Liu and Zilei Wang. Progressive boundary refinement network for temporal action detection. In AAAI, 2020. 2" + ], + "bbox": [ + 503, + 92, + 890, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "18599", + "bbox": [ + 480, + 945, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[32] Shuming Liu, Mengmeng Xu, Chen Zhao, Xu Zhao, and Bernard Ghanem. Etad: Training action detection end to end on a laptop. In CVPRW, 2023. 1, 2", + "[33] Xiaolong Liu, Song Bai, and Xiang Bai. An empirical study of end-to-end temporal action detection. In CVPR, 2022. 1, 2, 6", + "[34] Xiaolong Liu, Qimeng Wang, Yao Hu, Xu Tang, Shiwei Zhang, Song Bai, and Xiang Bai. End-to-end temporal action detection with transformer. IEEE Transactions on Image Processing, 31:5427-5441, 2022. 2, 6", + "[35] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In ICCV, 2021. 3", + "[36] Jinjie Mai, Abdullah Hamdi, Silvio Giancola, Chen Zhao, and Bernard Ghanem. Egoloc: Revisiting 3d object localization from egocentric videos with visual queries. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 45-57, 2023. 1", + "[37] Paulius Micikevicius, Sharan Narang, Jonah Alben, Gregory Diamos, Erich Elsen, David Garcia, Boris Ginsburg, Michael Houston, Oleksii Kuchaiev, Ganesh Venkatesh, et al. Mixed precision training. In ICLR, 2017. 4", + "[38] Jonghwan Mun, Minsu Cho, and Bohyung Han. Local-global video-text interactions for temporal grounding. In CVPR, 2020. 1", + "[39] OpenAI. Gpt-4 technical report, 2023. 1, 2", + "[40] Zhiwu Qing, Haisheng Su, Weihao Gan, Dongliang Wang, Wei Wu, Xiang Wang, Yu Qiao, Junjie Yan, Changxin Gao, and Nong Sang. Temporal context aggregation network for temporal action proposal refinement. In CVPR, 2021. 2", + "[41] Alec Radford, Karthik Narasimhan, Tim Salimans, Ilya Sutskever, et al. Improving language understanding by generative pre-training. OpenAI, 2018. 2", + "[42] Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, Ilya Sutskever, et al. Language models are unsupervised multitask learners. OpenAI blog, 1(8):9, 2019. 2", + "[43] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, 2021. 3", + "[44] Merey Ramazanova, Victor Escorcia, Fabian Caba Heilbron, Chen Zhao, and Bernard Ghanem. Owl (observe, watch, listen): Localizing actions in egocentric video via audiovisual temporal context. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshop (CVPRW), 2023. 1", + "[45] Jiayi Shao, Xiaohan Wang, Ruijie Quan, Junjun Zheng, Jiang Yang, and Yi Yang. Action sensitivity learning for temporal action localization. In ICCV, 2023. 2, 6", + "[46] Dingfeng Shi, Yujie Zhong, Qiong Cao, Jing Zhang, Lin Ma, Jia Li, and Dacheng Tao. React: Temporal action detection with relational queries. In ECCV, 2022. 2", + "[47] Dingfeng Shi, Yujie Zhong, Qiong Cao, Lin Ma, Jia Li, and Dacheng Tao. Tridet: Temporal action detection with relative boundary modeling. In CVPR, 2023. 2, 6, 7" + ], + "bbox": [ + 78, + 90, + 468, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[48] Mattia Soldan, Mengmeng Xu, Sisi Qu, Jesper Tegner, and Bernard Ghanem. VLG-Net: Video-language graph matching network for video grounding. In ICCVW, 2021. 1", + "[49] Yi-Lin Sung, Jaemin Cho, and Mohit Bansal. Lst: Ladder side-tuning for parameter and memory efficient transfer learning. In NeurIPS, 2022. 3, 5", + "[50] Yi-Lin Sung, Jaemin Cho, and Mohit Bansal. VI-adapter: Parameter-efficient transfer learning for vision-and-language tasks. In CVPR, 2022. 2, 4", + "[51] Jing Tan, Jiaqi Tang, Limin Wang, and Gangshan Wu. Relaxed transformer decoders for direct action proposal generation. In ICCV, 2021. 2", + "[52] Tuan N Tang, Kwonyoung Kim, and Kwanghoon Sohn. Temporalmaxer: Maximize temporal context with only max pooling for temporal action localization. arXiv preprint arXiv:2303.09055, 2023. 7", + "[53] Zhan Tong, Yibing Song, Jue Wang, and Limin Wang. Videomae: Masked autoencoders are data-efficient learners for self-supervised video pre-training. In NeurIPS, 2022. 1, 4, 6", + "[54] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In NeurIPS, 2017. 4", + "[55] Chenhao Wang, Hongxiang Cai, Yuxin Zou, and Yichao Xiong. Rgb stream is enough for temporal action detection. arXiv preprint arXiv:2107.04362, 2021. 2", + "[56] Limin Wang, Bingkun Huang, Zhiyu Zhao, Zhan Tong, Yi-nan He, Yi Wang, Yali Wang, and Yu Qiao. Videomae v2: Scaling video masked autoencoders with dual masking. In ICCV, 2023. 1, 3, 6, 7", + "[57] Yi Wang, Kunchang Li, Yizhuo Li, Yinan He, Bingkun Huang, Zhiyu Zhao, Hongjie Zhang, Jilan Xu, Yi Liu, Zun Wang, et al. Internvideo: General video foundation models via generative and discriminative learning. arXiv preprint arXiv:2212.03191, 2022. 1, 6, 7", + "[58] Zhenzhi Wang, Ziteng Gao, Limin Wang, Zhifeng Li, and Gangshan Wu. Boundary-aware cascade networks for temporal action segmentation. In ECCV, 2020. 2", + "[59] Kun Xia, Le Wang, Sanping Zhou, Nanning Zheng, and Wei Tang. Learning to refactor action and co-occurrence features for temporal action localization. In CVPR, 2022. 2", + "[60] Mengmeng Xu, Chen Zhao, David S Rojas, Ali Thabet, and Bernard Ghanem. G-TAD: Sub-graph localization for temporal action detection. In CVPR, 2020. 1, 6", + "[61] Le Yang, Houwen Peng, Dingwen Zhang, Jianlong Fu, and Junwei Han. Revisiting anchor mechanisms for temporal action localization. IEEE Transactions on Image Processing, 29:8535-8548, 2020. 2", + "[62] Min Yang, Guo Chen, Yin-Dong Zheng, Tong Lu, and Limin Wang. Basictad: an astounding rgb-only baseline for temporal action detection. Computer Vision and Image Understanding, 232:103692, 2023. 1, 2, 6", + "[63] Taojiannan Yang, Yi Zhu, Yusheng Xie, Aston Zhang, Chen Chen, and Mu Li. Aim: Adapting image models for efficient video action recognition. In ICLR, 2023. 2, 3, 4", + "[64] Ting Yao, Tao Mei, and Yong Rui. Highlight detection with pairwise deep ranking for first-person video summarization. In CVPR, 2016. 1" + ], + "bbox": [ + 501, + 92, + 890, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "18600", + "bbox": [ + 480, + 945, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[65] Dongshuo Yin, Xueting Han, Bin Li, Hao Feng, and Jing Bai. Parameter-efficient is not sufficient: Exploring parameter, memory, and time efficient adapter tuning for dense predictions. arXiv preprint arXiv:2306.09729, 2023. 3, 5", + "[66] Chenlin Zhang, Jianxin Wu, and Yin Li. Actionformer: Localizing moments of actions with transformers. In ECCV, 2022. 1, 2, 3, 5, 6, 7", + "[67] Chen Zhao, Ali K Thabet, and Bernard Ghanem. Video self-stitching graph network for temporal action localization. In ICCV, 2021. 1, 2", + "[68] Chen Zhao, Merey Ramazanova, Mengmeng Xu, and Bernard Ghanem. Segstad: Precise temporal action detection via semantic segmentation. In ECCVW, 2022. 2", + "[69] Chen Zhao, Shuming Liu, Karttikeya Mangalam, and Bernard Ghanem. Re $^2$ TAL: Rewiring pretrained video backbones for reversible temporal action localization. In CVPR, 2023. 1, 2, 3, 4, 6", + "[70] Chen Zhao, Shuming Liu, Karttikeya Mangalam, Guocheng Qian, Fatimah Zohra, Abdulmohsen Alghannam, Jitendra Malik, and Bernard Ghanem. $\\mathrm{Dr}^2\\mathrm{Net}$ : Dynamic reversible dual-residual networks for memory-efficient finetuning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 1, 2", + "[71] Hang Zhao, Zhicheng Yan, Lorenzo Torresani, and Antonio Torralba. HACS: Human action clips and segments dataset for recognition and temporal localization. ICCV, 2019. 1", + "[72] Y Zhao, B Zhang, Z Wu, S Yang, L Zhou, S Yan, L Wang, Y Xiong, D Lin, Y Qiao, et al. Cuhk & ethz & siat submission to activitynet challenge 2017. CVPR ActivityNet Workshop, 2017. 6", + "[73] Zixuan Zhao, Dongqi Wang, and Xu Zhao. Movement enhancement toward multi-scale video feature representation for temporal action detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 13555-13564, 2023. 2" + ], + "bbox": [ + 78, + 90, + 468, + 585 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "18601", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2024/End-to-End Temporal Action Detection with 1B Parameters Across 1000 Frames/aed9af76-c260-41b9-be21-a2e11ea6ab0b_model.json b/2024/End-to-End Temporal Action Detection with 1B Parameters Across 1000 Frames/aed9af76-c260-41b9-be21-a2e11ea6ab0b_model.json new file mode 100644 index 0000000000000000000000000000000000000000..a29025ff6ce0dc7e61af7bb5e4fb3bd2e59f2511 --- /dev/null +++ b/2024/End-to-End Temporal Action Detection with 1B Parameters Across 1000 Frames/aed9af76-c260-41b9-be21-a2e11ea6ab0b_model.json @@ -0,0 +1,2312 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.08, + 0.131, + 0.892, + 0.152 + ], + "angle": 0, + "content": "End-to-End Temporal Action Detection with 1B Parameters Across 1000 Frames" + }, + { + "type": "text", + "bbox": [ + 0.152, + 0.18, + 0.813, + 0.217 + ], + "angle": 0, + "content": "Shuming Liu1 Chen-Lin Zhang2 Chen Zhao1* Bernard Ghanem1 \n1King Abdullah University of Science and Technology (KAUST) 24Paradigm Inc" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.252, + 0.313, + 0.267 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.283, + 0.474, + 0.616 + ], + "angle": 0, + "content": "Recently, temporal action detection (TAD) has seen significant performance improvement with end-to-end training. However, due to the memory bottleneck, only models with limited scales and limited data volumes can afford end-to-end training, which inevitably restricts TAD performance. In this paper, we reduce the memory consumption for end-to-end training, and manage to scale up the TAD backbone to 1 billion parameters and the input video to 1,536 frames, leading to significant detection performance. The key to our approach lies in our proposed temporal-informative adapter (TIA), which is a novel lightweight module that reduces training memory. Using TIA, we free the humongous backbone from learning to adapt to the TAD task by only updating the parameters in TIA. TIA also leads to better TAD representation by temporally aggregating context from adjacent frames throughout the backbone. We evaluate our model across four representative datasets. Owing to our efficient design, we are able to train end-to-end on VideoMAEv2-giant and achieve \\(75.4\\%\\) mAP on THUMOS14, being the first end-to-end model to outperform the best feature-based methods. Code is available at https://github.com/sming256/AdaTAD." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.643, + 0.21, + 0.658 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.668, + 0.47, + 0.834 + ], + "angle": 0, + "content": "Temporal Action Detection (TAD) plays a vital role in the understanding of long-form videos. Its objective is to pinpoint specific action instances within untrimmed videos, identifying their start and end times, along with their respective categories [15, 20, 44, 60, 62, 67, 71]. This task is crucial for various applications, including highlight detection [36, 64], video-language grounding [38, 48], and action spotting [3]. Though innovations in the detector design have made profound progress in the past years [30, 66, 67], recent research highlights two new trends: end-to-end training [10, 32, 69], and scaling up [56, 57]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.835, + 0.47, + 0.88 + ], + "angle": 0, + "content": "End-to-end training in TAD refers to jointly training the video encoder and action detector [10, 29, 62, 70]. Compared to feature-based methods, end-to-end training offers" + }, + { + "type": "image", + "bbox": [ + 0.507, + 0.254, + 0.885, + 0.468 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.472, + 0.895, + 0.57 + ], + "angle": 0, + "content": "Figure 1. AdaTAD enjoys the benefit of end-to-end training and scaling up with efficient memory usage. The bubble size represents the number of the model's learnable parameters. Using SlowFast-R50, AdaTAD achieves better performance compared to E2E-TAD [33] with less memory. When further scaling up the model to VideoMAEv2-gaint and data to 1536 frames, we achieve an impressive avg. mAP of \\(75.4\\%\\) on THUMOS14." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.583, + 0.892, + 0.658 + ], + "angle": 0, + "content": "distinct advantages. First, it can effectively bridge the gap commonly found between pretraining and fine-tuning, such as data and task discrepancy. Second, video spatial augmentations can be utilized in the end-to-end setting, leading to further performance gain." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.659, + 0.892, + 0.795 + ], + "angle": 0, + "content": "Scaling up refers to improving performance by increasing the model size or the input data volume, and has demonstrated its effectiveness in various domains [14, 19, 39, 53]. In TAD, offline methods have attempted to scale up the feature extraction network to reach a higher performance. A notable example includes the work by Wang et al.[56], which reports a \\(10\\%\\) increase in mean Average Precision (mAP) by scaling from VideoMAE-S to VideoMAEv2-giant, using ActionFormer [66] on THUMOS14 [25]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.796, + 0.895, + 0.902 + ], + "angle": 0, + "content": "Intuitively, combining the strengths of end-to-end training and scaling up is expected to be most beneficial for improving TAD performance. However, both strategies demand substantial GPU memory, which restricts end-to-end training to a small model [29, 62, 69], or a small input volume [10, 33]. As shown in Fig. 1, the performance of previous end-to-end methods still significantly lags behind the" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.887, + 0.222, + 0.9 + ], + "angle": 0, + "content": "*Corresponding author." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "18591" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.198 + ], + "angle": 0, + "content": "best results achieved by feature-based approaches. Additionally, current end-to-end methods in TAD use computationally intensive full fine-tuning, risking the issues of catastrophic forgetting and overfitting during transfer learning [50, 63]. These issues can result in less competitive performance, especially when the downstream datasets are small, which is a common scenario in the TAD domain." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.199, + 0.47, + 0.334 + ], + "angle": 0, + "content": "In this paper, we aim to overcome the above limitations by harnessing the advantages of both scaling-up and end-to-end training. To achieve this, we introduce adapter tuning for temporal action detection (AdaTAD). Our method successfully trains a TAD model in an end-to-end manner, utilizing a backbone with 1 billion parameters and processing input videos of 1,536 frames. As illustrated in Fig. 1, to the best of our knowledge, this is the first end-to-end work that outperforms the best feature-based TAD methods." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.336, + 0.471, + 0.683 + ], + "angle": 0, + "content": "Specifically, we employ the following strategies to enhance the TAD performance while maintaining reasonable memory consumption. First, we identify that the snippet representation commonly used in feature-based methods is excessively redundant. In response, we have adopted a more memory-efficient frame-representation scheme, establishing an effective end-to-end baseline for TAD. Second, we adopt the parameter-efficient fine-tuning technique to minimize memory usage and mitigate overfitting in transfer learning. Notably, we introduce a novel Temporal-Informative Adapter (TIA). This adapter is injected between backbone layers and is the only learnable component in the backbone during fine-tuning. Different from conventional adapters [22], TIA is tailored for the TAD task and integrates temporal depth-wise convolutions to aggregate informative context from adjacent frames. Third, for additional memory efficiency, we propose a lighter variant of our method. By positioning the TIAs alongside the original backbone, rather than inside it, backpropagation can be done through the TIAs only. This configuration can further cut down on the need for intermediate activations within the primary backbone, thereby allowing us to scale up the model size and data size to unprecedented levels." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.685, + 0.47, + 0.806 + ], + "angle": 0, + "content": "AdaTAD establishes a new state-of-the-art across multiple TAD datasets. Notably, our method achieves an impressive \\(75.4\\%\\) mAP on THUMOS14, outperforming the previous feature-based best result of \\(71.5\\%\\) by a large margin. This achievement underscores the possible paradigm shift in TAD, i.e., moving from traditional feature extraction plus offline detectors to scaling up end-to-end TAD training. We summarize our contribution as follows:" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.81, + 0.47, + 0.886 + ], + "angle": 0, + "content": "1. We introduce an efficient end-to-end framework for TAD, scaling up the model size to 1 billion parameters and the input data to 1,536 frames. We achieve a consistent performance improvement with the scaling up, shedding light on the importance of scaling for TAD." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.886, + 0.469, + 0.902 + ], + "angle": 0, + "content": "2. We propose a novel temporal-informative adapter to re" + }, + { + "type": "list", + "bbox": [ + 0.078, + 0.81, + 0.47, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.092, + 0.892, + 0.167 + ], + "angle": 0, + "content": "duce memory as well as aggregate the temporal context for TAD. Different variants of these adapters are designed to trade off between performance and memory cost. To the best of our knowledge, we are the first to introduce the adapter mechanism to TAD." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.168, + 0.892, + 0.228 + ], + "angle": 0, + "content": "3. Our method achieves state-of-the-art performance across four TAD datasets. Remarkably, this represents the first end-to-end approach that outperforms the previous feature-based methods by a large margin." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.239, + 0.642, + 0.255 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.265, + 0.892, + 0.505 + ], + "angle": 0, + "content": "Temporal Action Detection. Current methods for temporal action detection, also referred to as temporal action localization, can be broadly classified into three categories based on their architectural design: one-stage, two-stage, and DETR-based methods. One-stage methods directly localize actions from a multi-scale feature pyramid, such as ActionFormer [66] and TriDet [47]. These methods integrate action classification and temporal boundary regression in a single step [45, 61, 62]. Two-stage methods, in contrast, involve an additional step of proposal feature extraction [4, 30, 40, 58, 59, 68, 73]. For instance, VSGN [67] employs boundary sampling to further refine proposals. Recently, there is a growing interest in query-based methods [34, 46, 51], which deploy a set of learned queries to interact with the feature maps and directly predict the actions' temporal boundaries and categories." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.506, + 0.892, + 0.793 + ], + "angle": 0, + "content": "In addition to the aforementioned categories, TAD can also be divided into feature-based and end-to-end methods. The former relies on pre-extracted RGB features and optionally incorporates optical flow features. On the other hand, end-to-end methods take raw video frames as input and jointly optimize the video encoder and action detector [31]. Due to computational constraints, AFSD [29] downsamples the input to a resolution of \\(96^2\\). DaoTAD [55] and E2E-TAD [33] provide evidence that high TAD performance can be achieved by relying solely on the RGB modality with various data augmentations. Further innovations came from SGP [11], TALLFormer [10], and ETAD [32], all of which introduced strategies to backpropagate only through parts of the data. Additionally, \\(\\mathrm{Re}^2\\)TAL [69] and \\(\\mathrm{Dr}^2\\)Net[70] design reversible network architectures to release the memory occupied by intermediate activations. Despite these advancements, all above methods follow the full fine-tuning paradigm, and none has yet surpassed the best results achieved by feature-based approaches." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.795, + 0.892, + 0.902 + ], + "angle": 0, + "content": "Scaling Law in Deep Learning. Scaling up model and data has been a prevalent strategy across both computer vision and natural language processing fields to achieve superior performance. The GPT series [5, 39, 41, 42] has consistently demonstrated that larger models pretrained on extensive datasets yield significant gains in language understanding capabilities. Analogously, in the realms of image" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "18592" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.097, + 0.092, + 0.28, + 0.285 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.326, + 0.091, + 0.45, + 0.287 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.487, + 0.091, + 0.611, + 0.286 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.647, + 0.091, + 0.87, + 0.286 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.293, + 0.894, + 0.351 + ], + "angle": 0, + "content": "Figure 2. Comparative illustration of our proposed TAD framework versus popular and widely used alternatives. (a) represents the typical offline method. (b) is the traditional end-to-end method using full fine-tuning. (c) Tailored for the TAD task, our AdaTAD uses a lightweight temporal-informative adapter inside the backbone to achieve efficient transfer learning. (d) To further reduce memory usage and scale up the model/data, \\(\\mathrm{AdaTAD}^{\\dagger}\\) uses an alternative placement for adapters outside the backbone." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.365, + 0.473, + 0.501 + ], + "angle": 0, + "content": "and video understanding, architectures such as ViT [14] and MViT [16], have also witnessed the effectiveness of this scaling strategy. Alabdulmohsin et al.[1] present a recipe for estimating scaling law parameters reliably from learning curves in computer vision. To attain even greater performance, several studies have also scaled up image resolution [35] or video clip length [56]. In this paper, we successfully apply this principle within the domain of temporal action detection and achieve state-of-the-art results." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.505, + 0.47, + 0.672 + ], + "angle": 0, + "content": "Efficient Transfer Learning. Transfer learning aims to adapt a pretrained model to a new domain. In TAD, typically off-the-shelf action recognition models are employed as the backbone, such as SlowFast [17]. Traditional transfer learning adopts full fine-tuning, meaning that all parameters of the pretrained model are updated. However, studies by [43, 63] have noted that full fine-tuning may harm the pre-learned knowledge, particularly when the downstream dataset is small and less comprehensive. Moreover, as models increase in size, the computational and storage demands of full fine-tuning proportionally increase." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.673, + 0.47, + 0.868 + ], + "angle": 0, + "content": "In response to these challenges, several works have investigated parameter-efficient tuning (PEFT) strategies that involve fine-tuning only a fraction of the network. For instance, Adapter [22] inserts lightweight modules analogous to feedforward networks in transformers, and only tunes these elements. LoRA [23] employs low-rank matrices in each transformer layer. Prefix-tuning [28] and prompt-tuning [27] attach learnable prompt tokens at the input stage or within each layer. In computer vision, many PEFT methods [6, 7, 24, 63] have also been explored across various tasks to optimize transfer learning efficiency. Our work represents the first effort to examine the potential of the PEFT mechanism in TAD." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.469, + 0.902 + ], + "angle": 0, + "content": "Although PEFT effectively reduces the number of learnable parameters, data-intensive and computationally heavy" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.365, + 0.892, + 0.518 + ], + "angle": 0, + "content": "tasks like video understanding require more memory-efficient techniques. To this end, several works try to externalize the trainable components from the backbone, eliminating the need for backpropagation through the extensive original model. For example, LST [49] introduces a supplementary lightweight network that operates in parallel with the main model. Similarly, \\(\\mathrm{E}^3\\mathrm{VA}\\) [65] leverages intermediate features with adapters to enable efficient transfer learning while minimizing memory usage. Our work is inspired by these methods yet with a streamlined and simple design." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.531, + 0.636, + 0.549 + ], + "angle": 0, + "content": "3. Methodology" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.557, + 0.892, + 0.648 + ], + "angle": 0, + "content": "In this section, we introduce our AdaTAD step-by-step. We first introduce notations and study the efficient video representation to establish an end-to-end TAD baseline. Next, we introduce a temporal-informative adapter designed for efficient TAD. Finally, we propose an alternative placement for adapters to further alleviate computational demands." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.658, + 0.611, + 0.673 + ], + "angle": 0, + "content": "3.1. Notations" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.682, + 0.892, + 0.822 + ], + "angle": 0, + "content": "Temporal action detection can be formulated as follows: given an untrimmed video \\(\\mathbf{X} \\in \\mathbb{R}^{3 \\times H \\times W \\times T}\\), where \\(H\\) and \\(W\\) are the height and width of each frame, and \\(T\\) is the frame number, its temporal action annotations can be denoted as \\(\\Psi_{g} = \\{\\varphi_{i} = (t_{s}, t_{e}, c)\\}_{i=1}^{N}\\), where \\(t_{s}, t_{e}, c\\) are the start, end time and category of action instance \\(\\varphi_{i}\\), and \\(N\\) is the number of ground truth actions. TAD aims to predict candidate proposal set \\(\\Psi_{p} = \\{\\hat{\\varphi}_{i} = (\\hat{t}_{s}, \\hat{t}_{e}, \\hat{c}, s)\\}_{i=1}^{M}\\) to cover \\(\\Psi_{g}\\), and \\(s\\) is the confidence score." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.832, + 0.744, + 0.848 + ], + "angle": 0, + "content": "3.2. Frame-level representation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.892, + 0.902 + ], + "angle": 0, + "content": "Our end-to-end TAD architecture comprises two main components: feature extraction and action detection. Following previous work [69], we select ActionFormer [66] as our ac" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18593" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.166 + ], + "angle": 0, + "content": "tion detection head due to its robust performance across various datasets without much hyperparameter tuning. Next, we discuss two ways of encoding raw frames into representative features (feature extraction): snippet representation and frame representation." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.168, + 0.468, + 0.333 + ], + "angle": 0, + "content": "Snippets Representation. Snippet-based video representations are popular choices in offline feature extraction. The whole video is divided into several short snippets (or namely clips). Each snippet has a short temporal length, e.g., 16 frames, and different snippets can have overlapping frames. Thus, the video can be conceptualized as \\( T \\) snippets, denoted by \\( \\mathbf{X} \\in \\mathbb{R}^{T \\times 3 \\times 16 \\times H \\times W} \\). Each snippet is processed through the video backbone, and spatial-temporal pooling is applied to extract one snippet feature. This processing yields the feature representation \\( \\mathbf{F} \\in \\mathbb{R}^{T \\times C} \\), where \\( C \\) denotes the channel dimension of the pooled features." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.334, + 0.468, + 0.454 + ], + "angle": 0, + "content": "Frame Representation. In contrast to snippet-based representations, frame-based video representations consider the entire video as a singular snippet or clip, represented as \\(\\mathbf{X} \\in \\mathbb{R}^{1 \\times 3 \\times T \\times H \\times W}\\). Then, the whole frame sequence is fed into the video backbone, and only spatial pooling is employed to extract features [10, 29, 69]. For attention-based models such as VideoMAE [54], the video is chunked into multiple shorter clips to avoid extensive temporal attention." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.455, + 0.468, + 0.635 + ], + "angle": 0, + "content": "Although both representations have been employed in previous studies, a fair comparison between them has not yet been performed. To address this gap, we conduct a comparative analysis of the two representations under the same setting on THUMOS14, measuring their memory usage and detection mAP. The results in Table 1 indicate that frame representation has comparable or even better performance than snippet representation, yet with much smaller memory consumption. When the feature extraction backbones are frozen, frame representation yields superior results to snippet representation for both VideoMAE [53] and SlowFast [17] backbones. Only in the end-to-end set" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.659, + 0.468, + 0.729 + ], + "angle": 0, + "content": "Table 1. Snippets representation vs frame representation. We use the end-to-end version of ActionFormer with two representations for comparison. The snippet input is \\(768 \\times 3 \\times 16 \\times 160 \\times 160\\), and the frame input is \\(1 \\times 3 \\times 768 \\times 160 \\times 160\\). * means activation checkpointing is utilized to avoid overflowing GPU memory." + }, + { + "type": "table", + "bbox": [ + 0.08, + 0.738, + 0.476, + 0.897 + ], + "angle": 0, + "content": "
SettingBackboneRepr.Avg. mAPMem (GB)
FrozenVideoMAE-SFrame59.351.9
Snippet57.6813.2
SlowFast-R101Frame61.343.6
Snippet60.2417.2
End to EndVideoMAE-SFrame67.152.8*
Snippet68.4624.6*
SlowFast-R101Frame65.335.5*
Snippet66.7251.6*
" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.891, + 0.167 + ], + "angle": 0, + "content": "ting can the snippet representation achieve \\(1\\%\\) mAP advantage over frame representations; however, it requires 8 times more memory consumption. Taking into account both performance and memory usage, frame-based representations could be a better choice for end-to-end TAD development." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.168, + 0.892, + 0.243 + ], + "angle": 0, + "content": "Therefore, we use frame representation as the default baseline to encode videos in our experiments. Following the previous TALLFormer work [10], we also incorporate activation checkpointing [8] and mixed precision training [37] to fully harness the potential of scaling." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.254, + 0.776, + 0.27 + ], + "angle": 0, + "content": "3.3. Temporal-Informative Adapter" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.278, + 0.892, + 0.534 + ], + "angle": 0, + "content": "In Section 3.2, we have built a simple end-to-end baseline using full fine-tuning. However, the baseline still suffers from two aspects: 1. Increased computational cost. In Table 1, we only use small video backbones like VideoMAES. When scaling VideoMAE-S to larger models, the computational burden and memory cost will grow rapidly. 2. Inferior transfer learning ability. More critically, the baseline follows the tradition of full fine-tuning, which may lead to inferior transfer learning. Pointed out by [50, 63], full fine-tuning may result in overfitting or forgetting, especially for large pretrained models. If downstream datasets are not sufficiently diverse, full fine-tuning can even destroy the powerful features learned from large-scale pretraining. Motivated by the above two aspects, we apply the PEFT mechanism and propose to fine-tune a plug-and-play module named Temporal-Informative Adapter (TIA) to achieve efficient and effective transfer learning for TAD." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.536, + 0.892, + 0.706 + ], + "angle": 0, + "content": "We first review the architecture of the standard adapter proposed by [22]. As formulated in Equation 1, the standard adapter includes a down-projection fully connected (FC) layer with parameter \\( \\mathbf{W}_{\\mathrm{down}} \\in \\mathbb{R}^{d \\times \\frac{d}{\\gamma}} \\), where \\( \\frac{d}{\\gamma} \\) represents the intermediate dimension and satisfies \\( \\gamma > 1 \\). Then, an up-projection layer \\( \\mathbf{W}_{\\mathrm{up}} \\in \\mathbb{R}^{\\frac{d}{\\gamma} \\times d} \\) is employed to restore the channel dimension. Between these two FC layers, a non-linear activation function \\( \\sigma \\) is inserted, such as GELU [21]. Afterward, a residual connection is added to the output of the projection layer. Note that \\( \\mathbf{x} \\) and \\( \\mathbf{x}' \\) are the input and output features with the same shape \\( \\mathbb{R}^{d \\times t \\times h \\times w} \\)." + }, + { + "type": "equation", + "bbox": [ + 0.586, + 0.717, + 0.891, + 0.737 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {x} ^ {\\prime} = \\boldsymbol {W} _ {\\mathrm {u p}} ^ {\\top} \\cdot \\boldsymbol {\\sigma} \\left(\\boldsymbol {W} _ {\\mathrm {d o w n}} ^ {\\top} \\cdot \\boldsymbol {x}\\right) + \\boldsymbol {x}. \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.891, + 0.84 + ], + "angle": 0, + "content": "Although the adapter has achieved great success in natural language processing and computer vision, the standard adapter only focuses on adapting channel information, which neglects the temporal context vital for the TAD task. To address this limitation, we introduce the temporalinformative adapter, as depicted in Fig. 3(b)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.892, + 0.9 + ], + "angle": 0, + "content": "The architecture of TIA follows the general bottleneck design of the standard adapter, while we integrate the temporal depth-wise convolution layers, as described in Equation 2. The temporal convolution with a kernel size of \\( k \\) is" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.956 + ], + "angle": 0, + "content": "18594" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.105, + 0.089, + 0.241, + 0.258 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.11, + 0.26, + 0.21, + 0.27 + ], + "angle": 0, + "content": "(a) Standard Adapter" + }, + { + "type": "image", + "bbox": [ + 0.282, + 0.089, + 0.44, + 0.258 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.284, + 0.259, + 0.446, + 0.27 + ], + "angle": 0, + "content": "(b) Temporal-Informative Adapter" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.274, + 0.47, + 0.316 + ], + "angle": 0, + "content": "Figure 3. Architecture of (a) standard adapter and (b) our temporal-informative adapter. We incorporate temporal depth-wise convolution to aggregate context from adjacent frames." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.342, + 0.473, + 0.471 + ], + "angle": 0, + "content": "designed to aggregate local informative context from adjacent frames and to enrich the representation of the current time step. Practically, this is achieved through the application of a 3D convolution with kernel size \\((k,1,1)\\) and group size \\(\\frac{d}{\\gamma}\\) for depth-wise processing. Additionally, an FC layer with weight \\(W_{\\mathrm{mid}}\\in \\mathbb{R}^{\\frac{d}{\\gamma}\\times \\frac{d}{\\gamma}}\\) is employed to facilitate information exchange across channels. At last, a learnable scalar \\(\\alpha\\) is introduced to adjust the amplitude of adapter's output." + }, + { + "type": "equation", + "bbox": [ + 0.164, + 0.48, + 0.304, + 0.498 + ], + "angle": 0, + "content": "\\[\n\\bar {\\boldsymbol {x}} = \\boldsymbol {\\sigma} (\\boldsymbol {W} _ {\\text {d o w n}} ^ {\\top} \\cdot \\boldsymbol {x}),\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.168, + 0.501, + 0.469, + 0.52 + ], + "angle": 0, + "content": "\\[\n\\hat {\\boldsymbol {x}} = \\boldsymbol {W} _ {\\mathrm {m i d}} ^ {\\top} \\cdot \\mathbf {D W C o n v} _ {k} (\\bar {\\boldsymbol {x}}) + \\bar {\\boldsymbol {x}}, \\tag {2}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.164, + 0.522, + 0.316, + 0.541 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {x} ^ {\\prime} = \\alpha \\cdot \\boldsymbol {W} _ {\\mathrm {u p}} ^ {\\top} \\cdot \\hat {\\boldsymbol {x}} + \\boldsymbol {x}.\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.551, + 0.47, + 0.793 + ], + "angle": 0, + "content": "As shown in Fig. 2(c), TIA is designed to be inserted between different backbone layers, e.g. between each ViT block of VideoMAE or each bottleneck block of SlowFast. To ensure the newly added connection does not affect the original network at the beginning of transfer learning, the weight and bias of the adapter's last projection layer \\( W_{\\mathrm{up}} \\) are initialized to 0. The learnable coefficient \\( \\alpha \\) is initialized to 1. The temporal kernel size \\( k \\) is set to 3, and the channel downsampling ratio \\( \\gamma \\) is set to 4 by default. Under these settings, the additional trainable parameters coming from TIA only account for \\( 4.7\\% \\) of the total parameters of the original backbone. Since this backbone is frozen when TIA is used, our proposed strategy constitutes a massive reduction in trainable parameters as compared to full fine-tuning. Our experiments show that TIA can achieve better performance than full fine-tuning with less memory usage." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.802, + 0.38, + 0.818 + ], + "angle": 0, + "content": "3.4. Alternative Placement for Adapter" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.469, + 0.903 + ], + "angle": 0, + "content": "Although the previously described PEFT approach can reduce tunable parameters and memory usage, the gradient still needs to backpropagate over the entire backbone during training. This requirement limits our ability to scale-up the model size or input data size further. As highlighted in prior" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.136 + ], + "angle": 0, + "content": "works [49, 65], if we can stop the gradient backpropagation within the original backbone, additional computational savings can be achieved." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.137, + 0.893, + 0.319 + ], + "angle": 0, + "content": "Inspired by this insight, we propose a placement strategy for adapters that position them externally to the backbone, rather than inserting them inside. As illustrated in Fig. 2(d), we utilize the previously introduced TIA module, but its output does not feed back into the middle of the original backbone. It is directly added to the backbone's final layer. This configuration eliminates the need for backpropagation through the original network, as gradients are only tracked to the shallow lightweight adapter. To further diminish computation, we observe that adapting only the last half of backbone layers yields comparable performance while reducing half of the adaptation cost." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.319, + 0.892, + 0.44 + ], + "angle": 0, + "content": "To distinguish the different variants, we name the standard adaption design as AdaTAD, and the alternative placement as AdaTAD†. The latter can be considered as a lite version of the former. Compared to directly injecting adapters into the backbone, AdaTAD† may lead to a slight performance drop. However, it enables us to leverage richer models and more data, which should effectively counter this possible drop." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.455, + 0.634, + 0.472 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.48, + 0.698, + 0.495 + ], + "angle": 0, + "content": "4.1. Datasets and Metrics" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.503, + 0.892, + 0.67 + ], + "angle": 0, + "content": "We choose ActivityNet-1.3 [20], THUMOS14 [25], and Epic-Kitchens 100 [13] to evaluate our proposed approach. ActivityNet-1.3 and THUMOS14 are web-collected third-person untrimmed videos, consisting of 19,994 and 413 videos, respectively. EPIC-Kitchens 100 is collected from 700 egocentric videos. Since the action categories of EPIC-Kitchens 100 are more domain-specific and different from common pretraining data, achieving higher performance on this dataset is more challenging. Moreover, we also evaluate our method on the Ego4D-MQ [18] dataset, and the results can be found in the appendix." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.671, + 0.892, + 0.777 + ], + "angle": 0, + "content": "Following common practice, we report the mean Average Precision (mAP) at certain IoU thresholds and average mAP as the evaluation metrics. On ActivityNet1.3, the IoU thresholds are chosen from 0.5 to 0.95 with 10 steps. On THUMOS14, the threshold is chosen from \\(\\{0.3,0.4,0.5,0.6,0.7\\}\\). On EPIC-Kitchens 100, the threshold is set to \\(\\{0.1,0.2,0.3,0.4,0.5\\}\\)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.787, + 0.718, + 0.803 + ], + "angle": 0, + "content": "4.2. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.81, + 0.892, + 0.901 + ], + "angle": 0, + "content": "We implement our method with PyTorch 2.0 and MMAc-tion2 [12] with 4 A100 GPUs. By default, mixed-precision training and activation checkpointing are adopted to save memory. We use ActionFormer [66] as our detection head, and keep the hyper-parameters unchanged on each dataset. The learning rate for the adapter in the backbone is grid" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18595" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.089, + 0.895, + 0.161 + ], + "angle": 0, + "content": "Table 2. Results on ActivityNet-1.3 and THUMOS14, measured by mAP (\\%) at different tIoU thresholds. E2E refers to end-to-end training, and Mem refers to memory usage (GB) per video. On ActivityNet-1.3, our prediction is combined with CUHK [72] classification results. Specifically, * means we employ stronger video-level classification results used in InternVideo [57] for a fair comparison. We report our best results in bold, and the previous best results in underline, which was achieved by the feature-based method. The last row is achieved when only the last half of backbone layers are adapted; otherwise, full-layer adaptation will lead to out-of-memory on A100-80G." + }, + { + "type": "table", + "bbox": [ + 0.078, + 0.165, + 0.897, + 0.499 + ], + "angle": 0, + "content": "
MethodBackboneE2EFlowMemActivityNet-1.3THUMOS14
0.50.750.95Avg.0.30.40.50.60.7Avg.
BMN [30]TSNX✓-50.0734.788.2933.8556.047.438.829.720.538.5
TadTR [34]I3DX✓-49.1032.608.5032.3062.457.449.237.826.346.6
ActionFormer [66]SlowFast-R50XX-54.2637.048.1336.0278.773.365.254.639.762.3
ActionFormer [66]I3DX✓-53.5036.208.2035.6082.177.871.059.443.966.8
ASL [45]I3DX✓-54.1067.408.0036.2083.179.071.759.745.867.9
TriDet [47]I3DX✓-54.7038.008.4036.8083.680.172.962.447.469.3
VideoMAEv2 [56]VideoMAEv2-gXX----------69.6
InternVideo [57]VideoMAE-H+UniformerV2XX----39.00*-----71.5
AFSD [29]I3D✓✓1252.4035.306.5034.4067.362.455.543.731.152.0
E2E-TAD [33]SlowFast-R50✓X1250.4735.9910.3335.1069.464.356.046.434.954.2
BasicTAD [62]SlowOnly-R50✓X1251.2033.417.5733.1275.570.863.550.937.459.6
TALLFormer [10]VideoSwin-B✓X2954.1036.207.9035.6076.0-63.2-34.559.2
Re2TAL [69]Re2VideoSwin-T✓X2454.7537.819.0336.8077.071.562.449.736.359.4
AdaTADSlowFast-R50✓X4.355.2838.118.8737.1181.076.269.459.044.566.0
AdaTADVideoMAE-S✓X2.556.1538.999.0737.8584.580.271.660.946.968.8
AdaTADVideoMAE-B✓X4.956.7739.359.7138.3987.082.475.363.849.271.5
AdaTADVideoMAE-L✓X11.057.6940.5610.1339.2287.784.176.766.452.473.5
AdaTADVideoMAE-H✓X19.258.0440.559.7539.3788.985.378.666.952.574.4
AdaTADVideoMAEv2-g✓X29.958.4541.1610.4539.7989.585.878.967.352.674.8
AdaTAD† (1536×2242)VideoMAEv2-g✓X43.660.8242.699.8441.15*89.685.979.467.653.875.4
AdaTAD (1536×2242)VideoMAEv2-g✓X50.661.7243.3510.8541.93*89.786.780.971.056.176.9
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.515, + 0.472, + 0.683 + ], + "angle": 0, + "content": "searched from 5e-4 to 5e-5, and other parameters inside the backbone are frozen. On ActivityNet-1.3, we resize the video into a fixed length of 768 frames. On THUMOS14, we randomly truncate a window with 768 frames with a temporal stride of 4. On EPIC-Kitchens 100, we randomly truncate a window with 6144 frames with a temporal stride of 2. After the video encoder, the feature is resized to fixed lengths of 192, 768, and 768, respectively, for the three datasets. Frame resolution is set to \\(160^2\\) by default. In all experiments, we report the training memory usage. More implementation details can be found in the appendix." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.695, + 0.367, + 0.71 + ], + "angle": 0, + "content": "4.3. Comparison with SoTA Methods" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.72, + 0.472, + 0.903 + ], + "angle": 0, + "content": "Table 2 compares our AdaTAD with other state-of-the-art (SoTA) methods on ActivityNet-1.3 and THUMOS14 datasets. Initially, we use SlowFast-R50 as the backbone. For comparison, we also extract corresponding offline features, utilizing the snippet representation where each snippet comprises 32 frames with \\(224^{2}\\) resolution. We observe that end-to-end training enhances performance from \\(62.3\\%\\) to \\(66.0\\%\\) on THUMOS14. Notably, this architecture has also been used in E2E-TAD [33]. However, our method consumes less memory while achieving superior performance. This apple-to-apple comparison underscores the benefits of adapter tuning and the scaling-up principle." + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.513, + 0.892, + 0.542 + ], + "angle": 0, + "content": "Table 3. Results on EPIC-Kitchens 100 validation set. For comparison, the feature-based methods use the same SlowFast-R50." + }, + { + "type": "table", + "bbox": [ + 0.501, + 0.546, + 0.895, + 0.832 + ], + "angle": 0, + "content": "
MethodE2E0.10.20.30.40.5Avg.
Verb Task
BMN [30]✘10.88.88.47.15.68.4
G-TAD [60]✘12.111.09.48.16.59.4
ActionFormer [66]✘26.625.424.222.319.123.5
ASL [45]✘27.9-25.5-19.824.6
TriDet [47]✘28.627.426.124.220.825.4
AdaTAD (SlowFast-R50)✓26.525.723.921.717.623.1
ActionFormer (VideoMAE-L)✘32.731.629.126.723.628.7
AdaTAD (VideoMAE-L)✓33.132.230.427.523.129.3
Noun Task
BMN [30]✘10.38.36.24.53.46.5
G-TAD [60]✘11.010.08.67.05.48.4
ActionFormer [66]✘25.224.122.720.517.021.9
ASL [45]✘26.0-23.4-17.722.6
TriDet [47]✘27.426.324.622.218.323.8
AdaTAD (SlowFast-R50)✓24.523.622.320.016.521.4
ActionFormer (VideoMAE-L)✘31.329.727.225.321.326.9
AdaTAD (VideoMAE-L)✓32.431.630.127.424.629.3
" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.84, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Furthermore, when adopting the VideoMAE [53] family as our backbone and progressively scaling up the model size, the performance of AdaTAD consistently improves. Using the largest model, i.e., VideoMAEv2-giant with 1.01" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "18596" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.089, + 0.472, + 0.146 + ], + "angle": 0, + "content": "Table 4. Compared to full fine-tuning, our adapter tuning can achieve better performance with less memory. Param. is the number of tunable parameters in the backbone. * means out of memory on A100-80GB, and we report the estimated number." + }, + { + "type": "table", + "bbox": [ + 0.078, + 0.15, + 0.477, + 0.359 + ], + "angle": 0, + "content": "
ModelSettingE2EParam.Mem.mAP
VideoMAE-SFeature×0-57.6
Snippet Full FT✓22M24.6G68.4
Frame Full FT✓22M2.8G67.1
AdaTAD✓1M2.5G68.8
VideoMAE-BFeature×0-64.7
Snippet Full FT✓86M87.4G*-
Frame Full FT✓86M5.6G70.1
AdaTAD✓4M4.9G71.5
VideoMAE-LFeature×0-66.5
Snippet Full FT✓304M193G*-
Frame Full FT✓304M13.1G73.0
AdaTAD✓14M11.0G73.5
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.368, + 0.47, + 0.503 + ], + "angle": 0, + "content": "billion parameters, and larger input data, i.e., 1536 frames with \\(224^{2}\\) resolution, we attain an impressive \\(41.9\\%\\) mAP on ActivityNet-1.3 and \\(76.9\\%\\) mAP on THUMOS14. It is noteworthy that the previous SoTA was achieved by VideoMAEv2 [56] and InternVideo [57], which utilize the same detector head as ours but with offline snippet features. Our method surpasses these in detection performance by a large margin, marking the first instance where an end-to-end TAD method can outperform SoTA feature-based results." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.504, + 0.47, + 0.671 + ], + "angle": 0, + "content": "We also present our results on EPIC-Kitchens 100 in Table 3. Since videos in this dataset have a longer duration, all previous methods rely solely on pre-extracted features [47, 52, 66]. Our approach is the first to adopt end-to-end training on this dataset. For fair comparison, we first utilize the same backbone as used in previous methods, i.e., SlowFast-R50 pretrained on EPIC, and we achieve comparable performance to ActionFormer [66]. Moreover, when we scale up the backbone to VideoMAE-L (it is also trained on EPIC-Kitchens 100 classification task), we achieve SoTA performance of \\(29.3\\%\\)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.679, + 0.285, + 0.695 + ], + "angle": 0, + "content": "4.4. Ablation and Analysis" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.702, + 0.47, + 0.763 + ], + "angle": 0, + "content": "In this section, we present a series of analyses to evaluate our proposed method and affirm the benefits of scaling up in TAD. Unless otherwise stated, our experiments utilize a standard input of 768 frames per video on THUMOS14." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.47, + 0.902 + ], + "angle": 0, + "content": "The advantage of adapter tuning. In Table 4, we compare conventional full fine-tuning with our proposed design. It is evident that end-to-end approaches significantly outperform pre-extracted features. Moreover, with full finetuning, the snippet representation slightly advances over frame representation but incurs tremendous memory costs, which aligns with our analysis in Section 3.2. However, AdaTAD uses less memory and still achieves better performance than conventional full fine-tuning. This also veri" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.089, + 0.892, + 0.133 + ], + "angle": 0, + "content": "Table 5. When scaling up the input data, AdaTAD's performance consistently increases. * means snippet representation is used in offline feature extraction, and each snippet has 16 frames." + }, + { + "type": "table", + "bbox": [ + 0.502, + 0.136, + 0.892, + 0.273 + ], + "angle": 0, + "content": "
SettingModelResolutionFramesMem.mAP
FeatureVideoMAEv2-g2242768x16*-69.6
AdaTADVideoMAE-S16027682.5G68.8
160215363.8G69.7
160230726.5G70.6
22427683.8G70.7
224215366.4G71.0
2242307211.6G71.5
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.279, + 0.892, + 0.309 + ], + "angle": 0, + "content": "Table 6. AdaTAD† can further push the boundaries of scaling up. OOM means out of memory on A100-80GB." + }, + { + "type": "table", + "bbox": [ + 0.502, + 0.312, + 0.891, + 0.431 + ], + "angle": 0, + "content": "
SettingModelResolutionFrameMem.mAP
AdaTADVideoMAE-L160276811.0G73.5
VideoMAEv2-g160276829.9G74.8
VideoMAEv2-g22421536OOM-
AdaTAD†VideoMAEv2-g160276822.8G73.7
VideoMAEv2-g224276830.0G74.6
VideoMAEv2-g2242153643.6G75.4
" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.44, + 0.892, + 0.532 + ], + "angle": 0, + "content": "fies the limitations of full fine-tuning, as discussed in Section 3.3. Specifically, our method enhances VideoMAE-S backbone with an \\(11.2\\%\\) gain using only 1M trainable parameters. Additionally, Table 4 also demonstrates that scaling up the model size of the video backbone is an effective way to improve TAD performance." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.535, + 0.893, + 0.701 + ], + "angle": 0, + "content": "The advantage of scaling up the data. In addition to model scaling, Table 5 verifies the effectiveness of data scaling, which involves two aspects: frame number and frame resolution. Firstly, given the same video duration, increasing the frame number from 768 to 3072 can raise the mAP from \\(68.8\\%\\) to \\(70.6\\%\\). In the meantime, the memory usage is nearly three times larger. Secondly, increasing the frame resolution from \\(160^2\\) to \\(224^2\\) also improves the mAP. In the end, by only scaling up the data, we elevate the mAP from \\(68.8\\%\\) to \\(71.5\\%\\), already surpassing the current SoTA feature-based approach with a giant backbone model [57]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.701, + 0.892, + 0.792 + ], + "angle": 0, + "content": "Moreover, increasing the frame resolution from \\(160^2\\) to \\(224^2\\) or increasing the frame number from 768 to 1536 results in the same memory usage of 3.8G. However, the former achieves \\(70.7\\%\\) mAP while the latter only reaches \\(69.7\\%\\). This suggests that frame resolution may be prioritized under the same memory budget, for the TAD task." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.795, + 0.893, + 0.902 + ], + "angle": 0, + "content": "The advantage of AdaTAD†. Given the effectiveness of scaling up the model or data, we further explore combining these approaches. In Table 6, using 768 frames while scaling up the model to VideoMAEv2-giant results in memory usage escalating to 29.9G. In such a scenario, further increasing the data could easily lead to memory overflow, even with the A100-80GB GPU. This indicates that adapta" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18597" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.137 + ], + "angle": 0, + "content": "tion tuning reached its limit under this extreme case. Therefore, to utilize both the largest models with larger data simultaneously, \\(\\mathrm{AdaTAD}^{\\dagger}\\) shows its advantage." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.137, + 0.468, + 0.243 + ], + "angle": 0, + "content": "Concretely, when switching from AdaTAD to AdaTAD†, the memory usage of the VideoMAEv2-giant model is reduced from 29.9G to 22.8G. Although a slight performance drop is observed, its reduced memory footprint enables scaling up data from 768 frames to as much as 1536 frames with a high resolution of \\(224^2\\). This scalability helps mitigate the performance drop and achieves a higher mAP of \\(75.4\\%\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.246, + 0.468, + 0.473 + ], + "angle": 0, + "content": "The ablation of the adapter design. Detailed in Table 7, we compare different adapter architectural designs. The baseline, i.e., offline snippet feature, achieves \\(64.7\\%\\) mAP. End-to-end learning in all designs yields at least a \\(5\\%\\) improvement. Our AdaTAD achieves \\(71.5\\%\\) in the end. Compared to standard adapter [22], ours consumes similar memory but achieves higher mAP. This verifies that local temporal context is vital for the TAD task. In contrast to full finetuning (FT), we tune only 4M parameters using less memory. In our design, we find that removing the residual connection of the depth-wise convolution drops performance by \\(0.7\\%\\), and training becomes unstable. We also implement an adaptation design proposed in LongLoRA [9], which efficiently computes long-range attention and shows decent performance but requires more parameters and memory." + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.487, + 0.468, + 0.514 + ], + "angle": 0, + "content": "Table 7. Ablation of different adapter architectural designs. VideoMAE-B is used to conduct the following experiments." + }, + { + "type": "table", + "bbox": [ + 0.08, + 0.516, + 0.462, + 0.637 + ], + "angle": 0, + "content": "
SettingE2EParam.Mem.mAPgains
Snippet FeatureX0-64.7
+ Full FT✓86M5.6G70.1+5.1
+ LongLoRA [9]✓28M6.2G71.1+6.1
+ Standard Adapter [22]✓3.6M4.8G70.2+5.2
+ AdaTAD (w/o residual)✓4.0M4.9G70.8+5.8
+ AdaTAD✓4.0M4.9G71.5+6.5
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.645, + 0.47, + 0.901 + ], + "angle": 0, + "content": "The necessity of end-to-end training for TAD. As previously discussed, end-to-end training can address discrepancies between the pretraining and fine-tuning stages in terms of training data and learning tasks. To corroborate this, we employ models pretrained on different datasets for the EPIC-Kitchens TAD task in Table 8. Kinetics-400 (K400) [26] represents commonly collected third-person web data and exhibits a large domain gap compared to EPIC-Kitchens 100. Using K400 for pretraining, we observe that end-to-end TAD training allows for \\(+5.56\\) gain. Conversely, using a model already finetuned on EPIC-Kitchens still yields a \\(+2.32\\) improvement. Unlike K400 pretraining, since this model has already adapted to the data discrepancy, we can infer that this gain leverages differences between the classification task in pretraining and the detection task in fine-tuning. Such improvements further underscore the significance of end-to-end training in TAD." + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.09, + 0.892, + 0.132 + ], + "angle": 0, + "content": "Table 8. End-to-end TAD can alleviate the discrepancies between pretraining and finetuning. VideoMAE-L with different pretrained weights are used on the EPIC-Kitchens 100 Noun task." + }, + { + "type": "table", + "bbox": [ + 0.502, + 0.136, + 0.895, + 0.228 + ], + "angle": 0, + "content": "
Pretrain DatasetE2E0.10.30.5mAPgain
K400 [26]X18.6916.3511.5215.77
K400 [26]✓24.3322.1416.8721.33+ 5.56
K400 [26] + EPIC [13]X31.3227.2521.3326.98
K400 [26] + EPIC [13]✓32.4130.1324.5929.30+ 2.32
" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.233, + 0.655, + 0.248 + ], + "angle": 0, + "content": "4.5. Error Analyses" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.256, + 0.892, + 0.346 + ], + "angle": 0, + "content": "We also conduct false positive analysis at tIoU=0.5 in Fig. 4. Compared to feature-based training, learning from raw frames produces more helpful positive detections. More importantly, the percentage of wrong label error is reduced after end-to-end training, suggesting its unique advantage in classifying accurate action labels." + }, + { + "type": "image", + "bbox": [ + 0.514, + 0.35, + 0.887, + 0.492 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.495, + 0.892, + 0.537 + ], + "angle": 0, + "content": "Figure 4. False Positive Profiling on THUMOS14 using [2]. We use VideoMAEv2-giant as the backbone, and compare end-to-end training with pre-extracted feature-based training." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.544, + 0.627, + 0.558 + ], + "angle": 0, + "content": "5. Conclusions" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.569, + 0.892, + 0.81 + ], + "angle": 0, + "content": "This work introduces a memory-efficient and parameter-efficient end-to-end method named AdaTAD. Our key innovation lies in the proposed temporal-informative adapter, which is tailored for TAD with low computation costs. Furthermore, we design an alternative placement for adapters to minimize memory usage. By demonstrating the feasibility and effectiveness of scaling up end-to-end TAD, our work achieves new SoTA performance across multiple datasets. Particularly, this is the first instance of an end-to-end TAD method that surpasses the current best feature-based models. In fact, AdaTAD achieves a groundbreaking \\(75.4\\%\\) mAP on THUMOS14. We believe our work underscores the possible paradigm shift in TAD, advocating a move away from the traditional methodology of separate feature extraction and offline detection towards a more integrated approach of scaling up end-to-end TAD training." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.811, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Acknowledgement. This work was supported by the King Abdullah University of Science and Technology (KAUST) Office of Sponsored Research through the Visual Computing Center (VCC) funding, as well as the SDAIA-KAUST Center of Excellence in Data Science and Artificial Intelligence (SDAIA-KAUST AI)." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18598" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.156 + ], + "angle": 0, + "content": "[1] Ibrahim M Alabdulmohsin, Behnam Neyshabur, and Xiaohua Zhai. Revisiting neural scaling laws in language and vision. In NeurIPS, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.159, + 0.47, + 0.2 + ], + "angle": 0, + "content": "[2] Humam Alwassel, Fabian Caba Heilbron, Victor Escorcia, and Bernard Ghanem. Diagnosing error in temporal action detectors. In ECCV, 2018. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.202, + 0.471, + 0.256 + ], + "angle": 0, + "content": "[3] Humam Alwassel, Fabian Caba Heilbron, and Bernard Ghanem. Action search: Spotting actions in videos and its application to temporal action localization. In ECCV, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.259, + 0.471, + 0.313 + ], + "angle": 0, + "content": "[4] Yueran Bai, Yingying Wang, Yunhai Tong, Yang Yang, Qiyue Liu, and Junhui Liu. Boundary content graph neural network for temporal action proposal generation. In ECCV, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.316, + 0.471, + 0.372 + ], + "angle": 0, + "content": "[5] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. In NeurIPS, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.374, + 0.471, + 0.429 + ], + "angle": 0, + "content": "[6] Hao Chen, Ran Tao, Han Zhang, Yidong Wang, Wei Ye, Jindong Wang, Guosheng Hu, and Marios Savvides. Convadapter: Exploring parameter efficient transfer learning for convnets. arXiv preprint arXiv:2208.07463, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.431, + 0.471, + 0.485 + ], + "angle": 0, + "content": "[7] Shoufa Chen, Chongjian Ge, Zhan Tong, Jiangliu Wang, Yibing Song, Jue Wang, and Ping Luo. Adaptformer: Adapting vision transformers for scalable visual recognition. In NeurIPS, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.488, + 0.471, + 0.529 + ], + "angle": 0, + "content": "[8] Tianqi Chen, Bing Xu, Chiyuan Zhang, and Carlos Guestrin. Training deep nets with sublinear memory cost. arXiv preprint arXiv:1604.06174, 2016. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.531, + 0.471, + 0.585 + ], + "angle": 0, + "content": "[9] Yukang Chen, Shengju Qian, Haotian Tang, Xin Lai, Zhi-jian Liu, Song Han, and Jiaya Jia. Longlora: Efficient fin-tuning of long-context large language models. arXiv preprint arXiv:2309.12307, 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.588, + 0.47, + 0.628 + ], + "angle": 0, + "content": "[10] Feng Cheng and Gedas Bertasius. Tallformer: Temporal action localization with long-memory transformer. In ECCV, 2022. 1, 2, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.631, + 0.47, + 0.685 + ], + "angle": 0, + "content": "[11] Feng Cheng, Mingze Xu, Yuanjun Xiong, Hao Chen, Xinyu Li, Wei Li, and Wei Xia. Stochastic backpropagation: A memory efficient strategy for training video models. In CVPR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.688, + 0.469, + 0.729 + ], + "angle": 0, + "content": "[12] MMAction2 Contributors. Openmmlab's next generation video understanding toolbox and benchmark. https://github.com/open-mmlab/mmaction2, 2020.5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.731, + 0.47, + 0.8 + ], + "angle": 0, + "content": "[13] Dima Damen, Hazel Doughty, Giovanni Maria Farinella, Sanja Fidler, Antonino Furnari, Evangelos Kazakos, Davide Moltisanti, Jonathan Munro, Toby Perrett, Will Price, et al. Scaling egocentric vision: The epic-kitchens dataset. In ECCV, 2018. 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.803, + 0.47, + 0.883 + ], + "angle": 0, + "content": "[14] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2021. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.887, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[15] Victor Escorcia, Fabian Caba Heilbron, Juan Carlos Niebles," + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.471, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "and Bernard Ghanem. DAPs: Deep action proposals for action understanding. In ECCV, 2016. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.123, + 0.892, + 0.164 + ], + "angle": 0, + "content": "[16] Haoqi Fan, Bo Xiong, Karttikeya Mangalam, Yanghao Li, Zhicheng Yan, Jitendra Malik, and Christoph Feichtenhofer. Multiscale vision transformers. In ICCV, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.166, + 0.892, + 0.207 + ], + "angle": 0, + "content": "[17] Christoph Feichtenhofer, Haoqi Fan, Jitendra Malik, and Kaiming He. SlowFast networks for video recognition. In ICCV, 2019. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.21, + 0.892, + 0.278 + ], + "angle": 0, + "content": "[18] Kristen Grauman, Andrew Westbury, Eugene Byrne, Zachary Chavis, Antonino Furnari, Rohit Girdhar, Jackson Hamburger, Hao Jiang, Miao Liu, Xingyu Liu, et al. Ego4d: Around the world in 3,000 hours of egocentric video. In CVPR, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.281, + 0.892, + 0.322 + ], + "angle": 0, + "content": "[19] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, and Ross Girshick. Masked autoencoders are scalable vision learners. In CVPR, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.325, + 0.892, + 0.379 + ], + "angle": 0, + "content": "[20] Fabian Caba Heilbron, Victor Escorcia, Bernard Ghanem, and Juan Carlos Niebles. ActivityNet: A large-scale video benchmark for human activity understanding. In CVPR, 2015. 1, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.382, + 0.892, + 0.41 + ], + "angle": 0, + "content": "[21] Dan Hendrycks and Kevin Gimpel. Gaussian error linear units (gelus). arXiv preprint arXiv:1606.08415, 2016. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.412, + 0.892, + 0.467 + ], + "angle": 0, + "content": "[22] Neil Houlsby, Andrei Giurgiu, Stanislaw Jastrzebski, Bruna Morrone, Quentin De Laroussilhe, Andrea Gesmundo, Mona Attariyan, and Sylvain Gelly. Parameter-efficient transfer learning for nlp. In ICML, 2019. 2, 3, 4, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.47, + 0.892, + 0.524 + ], + "angle": 0, + "content": "[23] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. In ICML, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.527, + 0.892, + 0.568 + ], + "angle": 0, + "content": "[24] Menglin Jia, Luming Tang, Bor-Chun Chen, Claire Cardie, Serge Belongie, Bharath Hariharan, and Ser-Nam Lim. Visual prompt tuning. In ECCV, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.571, + 0.892, + 0.612 + ], + "angle": 0, + "content": "[25] YG Jiang, J Liu, A Roshan Zamir, G Toderici, I Laptev, M Shah, and R Sukthankar. Thumos challenge: Action recognition with a large number of classes, 2014. 1, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.614, + 0.892, + 0.682 + ], + "angle": 0, + "content": "[26] Will Kay, Joao Carreira, Karen Simonyan, Brian Zhang, Chloe Hillier, Sudheendra Vijayanarasimhan, Fabio Viola, Tim Green, Trevor Back, Paul Natsev, et al. The kinetics human action video dataset. arXiv preprint arXiv:1705.06950, 2017.8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.686, + 0.892, + 0.725 + ], + "angle": 0, + "content": "[27] Brian Lester, Rami Al-Rfou, and Noah Constant. The power of scale for parameter-efficient prompt tuning. In EMNLP, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.729, + 0.892, + 0.756 + ], + "angle": 0, + "content": "[28] Xiang Lisa Li and Percy Liang. Prefix-tuning: Optimizing continuous prompts for generation. In ACL, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.759, + 0.892, + 0.814 + ], + "angle": 0, + "content": "[29] Chuming Lin, Chengming Xu, Donghao Luo, Yabiao Wang, Ying Tai, Chengjie Wang, Jilin Li, Feiyue Huang, and Yanwei Fu. Learning salient boundary feature for anchor-free temporal action localization. In CVPR, 2021. 1, 2, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.816, + 0.892, + 0.857 + ], + "angle": 0, + "content": "[30] Tianwei Lin, Xiao Liu, Xin Li, Errui Ding, and Shilei Wen. BMN: boundary-matching network for temporal action proposal generation. In ICCV, 2019. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.86, + 0.892, + 0.899 + ], + "angle": 0, + "content": "[31] Qinying Liu and Zilei Wang. Progressive boundary refinement network for temporal action detection. In AAAI, 2020. 2" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.892, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "18599" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.134 + ], + "angle": 0, + "content": "[32] Shuming Liu, Mengmeng Xu, Chen Zhao, Xu Zhao, and Bernard Ghanem. Etad: Training action detection end to end on a laptop. In CVPRW, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.136, + 0.47, + 0.176 + ], + "angle": 0, + "content": "[33] Xiaolong Liu, Song Bai, and Xiang Bai. An empirical study of end-to-end temporal action detection. In CVPR, 2022. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.179, + 0.469, + 0.233 + ], + "angle": 0, + "content": "[34] Xiaolong Liu, Qimeng Wang, Yao Hu, Xu Tang, Shiwei Zhang, Song Bai, and Xiang Bai. End-to-end temporal action detection with transformer. IEEE Transactions on Image Processing, 31:5427-5441, 2022. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.235, + 0.469, + 0.289 + ], + "angle": 0, + "content": "[35] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In ICCV, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.292, + 0.469, + 0.361 + ], + "angle": 0, + "content": "[36] Jinjie Mai, Abdullah Hamdi, Silvio Giancola, Chen Zhao, and Bernard Ghanem. Egoloc: Revisiting 3d object localization from egocentric videos with visual queries. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 45-57, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.362, + 0.469, + 0.417 + ], + "angle": 0, + "content": "[37] Paulius Micikevicius, Sharan Narang, Jonah Alben, Gregory Diamos, Erich Elsen, David Garcia, Boris Ginsburg, Michael Houston, Oleksii Kuchaiev, Ganesh Venkatesh, et al. Mixed precision training. In ICLR, 2017. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.42, + 0.469, + 0.46 + ], + "angle": 0, + "content": "[38] Jonghwan Mun, Minsu Cho, and Bohyung Han. Local-global video-text interactions for temporal grounding. In CVPR, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.462, + 0.37, + 0.475 + ], + "angle": 0, + "content": "[39] OpenAI. Gpt-4 technical report, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.477, + 0.469, + 0.532 + ], + "angle": 0, + "content": "[40] Zhiwu Qing, Haisheng Su, Weihao Gan, Dongliang Wang, Wei Wu, Xiang Wang, Yu Qiao, Junjie Yan, Changxin Gao, and Nong Sang. Temporal context aggregation network for temporal action proposal refinement. In CVPR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.534, + 0.469, + 0.575 + ], + "angle": 0, + "content": "[41] Alec Radford, Karthik Narasimhan, Tim Salimans, Ilya Sutskever, et al. Improving language understanding by generative pre-training. OpenAI, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.577, + 0.469, + 0.618 + ], + "angle": 0, + "content": "[42] Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, Ilya Sutskever, et al. Language models are unsupervised multitask learners. OpenAI blog, 1(8):9, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.62, + 0.469, + 0.687 + ], + "angle": 0, + "content": "[43] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.689, + 0.469, + 0.771 + ], + "angle": 0, + "content": "[44] Merey Ramazanova, Victor Escorcia, Fabian Caba Heilbron, Chen Zhao, and Bernard Ghanem. Owl (observe, watch, listen): Localizing actions in egocentric video via audiovisual temporal context. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshop (CVPRW), 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.774, + 0.469, + 0.814 + ], + "angle": 0, + "content": "[45] Jiayi Shao, Xiaohan Wang, Ruijie Quan, Junjun Zheng, Jiang Yang, and Yi Yang. Action sensitivity learning for temporal action localization. In ICCV, 2023. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.817, + 0.469, + 0.858 + ], + "angle": 0, + "content": "[46] Dingfeng Shi, Yujie Zhong, Qiong Cao, Jing Zhang, Lin Ma, Jia Li, and Dacheng Tao. React: Temporal action detection with relational queries. In ECCV, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.86, + 0.469, + 0.9 + ], + "angle": 0, + "content": "[47] Dingfeng Shi, Yujie Zhong, Qiong Cao, Lin Ma, Jia Li, and Dacheng Tao. Tridet: Temporal action detection with relative boundary modeling. In CVPR, 2023. 2, 6, 7" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.134 + ], + "angle": 0, + "content": "[48] Mattia Soldan, Mengmeng Xu, Sisi Qu, Jesper Tegner, and Bernard Ghanem. VLG-Net: Video-language graph matching network for video grounding. In ICCVW, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.136, + 0.892, + 0.176 + ], + "angle": 0, + "content": "[49] Yi-Lin Sung, Jaemin Cho, and Mohit Bansal. Lst: Ladder side-tuning for parameter and memory efficient transfer learning. In NeurIPS, 2022. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.178, + 0.892, + 0.217 + ], + "angle": 0, + "content": "[50] Yi-Lin Sung, Jaemin Cho, and Mohit Bansal. VI-adapter: Parameter-efficient transfer learning for vision-and-language tasks. In CVPR, 2022. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.219, + 0.892, + 0.259 + ], + "angle": 0, + "content": "[51] Jing Tan, Jiaqi Tang, Limin Wang, and Gangshan Wu. Relaxed transformer decoders for direct action proposal generation. In ICCV, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.261, + 0.892, + 0.314 + ], + "angle": 0, + "content": "[52] Tuan N Tang, Kwonyoung Kim, and Kwanghoon Sohn. Temporalmaxer: Maximize temporal context with only max pooling for temporal action localization. arXiv preprint arXiv:2303.09055, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.317, + 0.892, + 0.37 + ], + "angle": 0, + "content": "[53] Zhan Tong, Yibing Song, Jue Wang, and Limin Wang. Videomae: Masked autoencoders are data-efficient learners for self-supervised video pre-training. In NeurIPS, 2022. 1, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.372, + 0.892, + 0.413 + ], + "angle": 0, + "content": "[54] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In NeurIPS, 2017. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.415, + 0.892, + 0.455 + ], + "angle": 0, + "content": "[55] Chenhao Wang, Hongxiang Cai, Yuxin Zou, and Yichao Xiong. Rgb stream is enough for temporal action detection. arXiv preprint arXiv:2107.04362, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.457, + 0.892, + 0.509 + ], + "angle": 0, + "content": "[56] Limin Wang, Bingkun Huang, Zhiyu Zhao, Zhan Tong, Yi-nan He, Yi Wang, Yali Wang, and Yu Qiao. Videomae v2: Scaling video masked autoencoders with dual masking. In ICCV, 2023. 1, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.512, + 0.892, + 0.579 + ], + "angle": 0, + "content": "[57] Yi Wang, Kunchang Li, Yizhuo Li, Yinan He, Bingkun Huang, Zhiyu Zhao, Hongjie Zhang, Jilan Xu, Yi Liu, Zun Wang, et al. Internvideo: General video foundation models via generative and discriminative learning. arXiv preprint arXiv:2212.03191, 2022. 1, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.581, + 0.892, + 0.622 + ], + "angle": 0, + "content": "[58] Zhenzhi Wang, Ziteng Gao, Limin Wang, Zhifeng Li, and Gangshan Wu. Boundary-aware cascade networks for temporal action segmentation. In ECCV, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.624, + 0.892, + 0.663 + ], + "angle": 0, + "content": "[59] Kun Xia, Le Wang, Sanping Zhou, Nanning Zheng, and Wei Tang. Learning to refactor action and co-occurrence features for temporal action localization. In CVPR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.665, + 0.892, + 0.705 + ], + "angle": 0, + "content": "[60] Mengmeng Xu, Chen Zhao, David S Rojas, Ali Thabet, and Bernard Ghanem. G-TAD: Sub-graph localization for temporal action detection. In CVPR, 2020. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.707, + 0.892, + 0.76 + ], + "angle": 0, + "content": "[61] Le Yang, Houwen Peng, Dingwen Zhang, Jianlong Fu, and Junwei Han. Revisiting anchor mechanisms for temporal action localization. IEEE Transactions on Image Processing, 29:8535-8548, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.763, + 0.892, + 0.816 + ], + "angle": 0, + "content": "[62] Min Yang, Guo Chen, Yin-Dong Zheng, Tong Lu, and Limin Wang. Basictad: an astounding rgb-only baseline for temporal action detection. Computer Vision and Image Understanding, 232:103692, 2023. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.818, + 0.892, + 0.858 + ], + "angle": 0, + "content": "[63] Taojiannan Yang, Yi Zhu, Yusheng Xie, Aston Zhang, Chen Chen, and Mu Li. Aim: Adapting image models for efficient video action recognition. In ICLR, 2023. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.86, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[64] Ting Yao, Tao Mei, and Yong Rui. Highlight detection with pairwise deep ranking for first-person video summarization. In CVPR, 2016. 1" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "18600" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.147 + ], + "angle": 0, + "content": "[65] Dongshuo Yin, Xueting Han, Bin Li, Hao Feng, and Jing Bai. Parameter-efficient is not sufficient: Exploring parameter, memory, and time efficient adapter tuning for dense predictions. arXiv preprint arXiv:2306.09729, 2023. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.149, + 0.47, + 0.19 + ], + "angle": 0, + "content": "[66] Chenlin Zhang, Jianxin Wu, and Yin Li. Actionformer: Localizing moments of actions with transformers. In ECCV, 2022. 1, 2, 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.192, + 0.47, + 0.233 + ], + "angle": 0, + "content": "[67] Chen Zhao, Ali K Thabet, and Bernard Ghanem. Video self-stitching graph network for temporal action localization. In ICCV, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.235, + 0.47, + 0.275 + ], + "angle": 0, + "content": "[68] Chen Zhao, Merey Ramazanova, Mengmeng Xu, and Bernard Ghanem. Segstad: Precise temporal action detection via semantic segmentation. In ECCVW, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.278, + 0.47, + 0.332 + ], + "angle": 0, + "content": "[69] Chen Zhao, Shuming Liu, Karttikeya Mangalam, and Bernard Ghanem. Re\\(^2\\)TAL: Rewiring pretrained video backbones for reversible temporal action localization. In CVPR, 2023. 1, 2, 3, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.334, + 0.47, + 0.417 + ], + "angle": 0, + "content": "[70] Chen Zhao, Shuming Liu, Karttikeya Mangalam, Guocheng Qian, Fatimah Zohra, Abdulmohsen Alghannam, Jitendra Malik, and Bernard Ghanem. \\(\\mathrm{Dr}^2\\mathrm{Net}\\): Dynamic reversible dual-residual networks for memory-efficient finetuning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.419, + 0.47, + 0.46 + ], + "angle": 0, + "content": "[71] Hang Zhao, Zhicheng Yan, Lorenzo Torresani, and Antonio Torralba. HACS: Human action clips and segments dataset for recognition and temporal localization. ICCV, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.461, + 0.47, + 0.516 + ], + "angle": 0, + "content": "[72] Y Zhao, B Zhang, Z Wu, S Yang, L Zhou, S Yan, L Wang, Y Xiong, D Lin, Y Qiao, et al. Cuhk & ethz & siat submission to activitynet challenge 2017. CVPR ActivityNet Workshop, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.518, + 0.47, + 0.587 + ], + "angle": 0, + "content": "[73] Zixuan Zhao, Dongqi Wang, and Xu Zhao. Movement enhancement toward multi-scale video feature representation for temporal action detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 13555-13564, 2023. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.587 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.518, + 0.956 + ], + "angle": 0, + "content": "18601" + } + ] +] \ No newline at end of file diff --git a/2024/End-to-End Temporal Action Detection with 1B Parameters Across 1000 Frames/aed9af76-c260-41b9-be21-a2e11ea6ab0b_origin.pdf b/2024/End-to-End Temporal Action Detection with 1B Parameters Across 1000 Frames/aed9af76-c260-41b9-be21-a2e11ea6ab0b_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..954b7162548f975d6ffbf80d8ed10f03f67abf3a --- /dev/null +++ b/2024/End-to-End Temporal Action Detection with 1B Parameters Across 1000 Frames/aed9af76-c260-41b9-be21-a2e11ea6ab0b_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1ded5eca4fe1d596ea0be269e571ef6e41e92166e6fdfbcb4f0f4dc5f6515efe +size 726257 diff --git a/2024/End-to-End Temporal Action Detection with 1B Parameters Across 1000 Frames/full.md b/2024/End-to-End Temporal Action Detection with 1B Parameters Across 1000 Frames/full.md new file mode 100644 index 0000000000000000000000000000000000000000..2065c9e51527aac7dc34de6f9dfdcc3dc8bda524 --- /dev/null +++ b/2024/End-to-End Temporal Action Detection with 1B Parameters Across 1000 Frames/full.md @@ -0,0 +1,308 @@ +# End-to-End Temporal Action Detection with 1B Parameters Across 1000 Frames + +Shuming Liu1 Chen-Lin Zhang2 Chen Zhao1* Bernard Ghanem1 +1King Abdullah University of Science and Technology (KAUST) 24Paradigm Inc + +# Abstract + +Recently, temporal action detection (TAD) has seen significant performance improvement with end-to-end training. However, due to the memory bottleneck, only models with limited scales and limited data volumes can afford end-to-end training, which inevitably restricts TAD performance. In this paper, we reduce the memory consumption for end-to-end training, and manage to scale up the TAD backbone to 1 billion parameters and the input video to 1,536 frames, leading to significant detection performance. The key to our approach lies in our proposed temporal-informative adapter (TIA), which is a novel lightweight module that reduces training memory. Using TIA, we free the humongous backbone from learning to adapt to the TAD task by only updating the parameters in TIA. TIA also leads to better TAD representation by temporally aggregating context from adjacent frames throughout the backbone. We evaluate our model across four representative datasets. Owing to our efficient design, we are able to train end-to-end on VideoMAEv2-giant and achieve $75.4\%$ mAP on THUMOS14, being the first end-to-end model to outperform the best feature-based methods. Code is available at https://github.com/sming256/AdaTAD. + +# 1. Introduction + +Temporal Action Detection (TAD) plays a vital role in the understanding of long-form videos. Its objective is to pinpoint specific action instances within untrimmed videos, identifying their start and end times, along with their respective categories [15, 20, 44, 60, 62, 67, 71]. This task is crucial for various applications, including highlight detection [36, 64], video-language grounding [38, 48], and action spotting [3]. Though innovations in the detector design have made profound progress in the past years [30, 66, 67], recent research highlights two new trends: end-to-end training [10, 32, 69], and scaling up [56, 57]. + +End-to-end training in TAD refers to jointly training the video encoder and action detector [10, 29, 62, 70]. Compared to feature-based methods, end-to-end training offers + +![](images/dae658e7cd312bba0b207001ec1b34728d4ec6ff778d8adbd6430ab83dafccda.jpg) +Figure 1. AdaTAD enjoys the benefit of end-to-end training and scaling up with efficient memory usage. The bubble size represents the number of the model's learnable parameters. Using SlowFast-R50, AdaTAD achieves better performance compared to E2E-TAD [33] with less memory. When further scaling up the model to VideoMAEv2-gaint and data to 1536 frames, we achieve an impressive avg. mAP of $75.4\%$ on THUMOS14. + +distinct advantages. First, it can effectively bridge the gap commonly found between pretraining and fine-tuning, such as data and task discrepancy. Second, video spatial augmentations can be utilized in the end-to-end setting, leading to further performance gain. + +Scaling up refers to improving performance by increasing the model size or the input data volume, and has demonstrated its effectiveness in various domains [14, 19, 39, 53]. In TAD, offline methods have attempted to scale up the feature extraction network to reach a higher performance. A notable example includes the work by Wang et al.[56], which reports a $10\%$ increase in mean Average Precision (mAP) by scaling from VideoMAE-S to VideoMAEv2-giant, using ActionFormer [66] on THUMOS14 [25]. + +Intuitively, combining the strengths of end-to-end training and scaling up is expected to be most beneficial for improving TAD performance. However, both strategies demand substantial GPU memory, which restricts end-to-end training to a small model [29, 62, 69], or a small input volume [10, 33]. As shown in Fig. 1, the performance of previous end-to-end methods still significantly lags behind the + +best results achieved by feature-based approaches. Additionally, current end-to-end methods in TAD use computationally intensive full fine-tuning, risking the issues of catastrophic forgetting and overfitting during transfer learning [50, 63]. These issues can result in less competitive performance, especially when the downstream datasets are small, which is a common scenario in the TAD domain. + +In this paper, we aim to overcome the above limitations by harnessing the advantages of both scaling-up and end-to-end training. To achieve this, we introduce adapter tuning for temporal action detection (AdaTAD). Our method successfully trains a TAD model in an end-to-end manner, utilizing a backbone with 1 billion parameters and processing input videos of 1,536 frames. As illustrated in Fig. 1, to the best of our knowledge, this is the first end-to-end work that outperforms the best feature-based TAD methods. + +Specifically, we employ the following strategies to enhance the TAD performance while maintaining reasonable memory consumption. First, we identify that the snippet representation commonly used in feature-based methods is excessively redundant. In response, we have adopted a more memory-efficient frame-representation scheme, establishing an effective end-to-end baseline for TAD. Second, we adopt the parameter-efficient fine-tuning technique to minimize memory usage and mitigate overfitting in transfer learning. Notably, we introduce a novel Temporal-Informative Adapter (TIA). This adapter is injected between backbone layers and is the only learnable component in the backbone during fine-tuning. Different from conventional adapters [22], TIA is tailored for the TAD task and integrates temporal depth-wise convolutions to aggregate informative context from adjacent frames. Third, for additional memory efficiency, we propose a lighter variant of our method. By positioning the TIAs alongside the original backbone, rather than inside it, backpropagation can be done through the TIAs only. This configuration can further cut down on the need for intermediate activations within the primary backbone, thereby allowing us to scale up the model size and data size to unprecedented levels. + +AdaTAD establishes a new state-of-the-art across multiple TAD datasets. Notably, our method achieves an impressive $75.4\%$ mAP on THUMOS14, outperforming the previous feature-based best result of $71.5\%$ by a large margin. This achievement underscores the possible paradigm shift in TAD, i.e., moving from traditional feature extraction plus offline detectors to scaling up end-to-end TAD training. We summarize our contribution as follows: + +1. We introduce an efficient end-to-end framework for TAD, scaling up the model size to 1 billion parameters and the input data to 1,536 frames. We achieve a consistent performance improvement with the scaling up, shedding light on the importance of scaling for TAD. +2. We propose a novel temporal-informative adapter to re + +duce memory as well as aggregate the temporal context for TAD. Different variants of these adapters are designed to trade off between performance and memory cost. To the best of our knowledge, we are the first to introduce the adapter mechanism to TAD. + +3. Our method achieves state-of-the-art performance across four TAD datasets. Remarkably, this represents the first end-to-end approach that outperforms the previous feature-based methods by a large margin. + +# 2. Related Work + +Temporal Action Detection. Current methods for temporal action detection, also referred to as temporal action localization, can be broadly classified into three categories based on their architectural design: one-stage, two-stage, and DETR-based methods. One-stage methods directly localize actions from a multi-scale feature pyramid, such as ActionFormer [66] and TriDet [47]. These methods integrate action classification and temporal boundary regression in a single step [45, 61, 62]. Two-stage methods, in contrast, involve an additional step of proposal feature extraction [4, 30, 40, 58, 59, 68, 73]. For instance, VSGN [67] employs boundary sampling to further refine proposals. Recently, there is a growing interest in query-based methods [34, 46, 51], which deploy a set of learned queries to interact with the feature maps and directly predict the actions' temporal boundaries and categories. + +In addition to the aforementioned categories, TAD can also be divided into feature-based and end-to-end methods. The former relies on pre-extracted RGB features and optionally incorporates optical flow features. On the other hand, end-to-end methods take raw video frames as input and jointly optimize the video encoder and action detector [31]. Due to computational constraints, AFSD [29] downsamples the input to a resolution of $96^2$ . DaoTAD [55] and E2E-TAD [33] provide evidence that high TAD performance can be achieved by relying solely on the RGB modality with various data augmentations. Further innovations came from SGP [11], TALLFormer [10], and ETAD [32], all of which introduced strategies to backpropagate only through parts of the data. Additionally, $\mathrm{Re}^2$ TAL [69] and $\mathrm{Dr}^2$ Net[70] design reversible network architectures to release the memory occupied by intermediate activations. Despite these advancements, all above methods follow the full fine-tuning paradigm, and none has yet surpassed the best results achieved by feature-based approaches. + +Scaling Law in Deep Learning. Scaling up model and data has been a prevalent strategy across both computer vision and natural language processing fields to achieve superior performance. The GPT series [5, 39, 41, 42] has consistently demonstrated that larger models pretrained on extensive datasets yield significant gains in language understanding capabilities. Analogously, in the realms of image + +![](images/e48ed53730e24a77b65a5de34100c577145e91640c4268b5d41ae7d6013d1417.jpg) +Figure 2. Comparative illustration of our proposed TAD framework versus popular and widely used alternatives. (a) represents the typical offline method. (b) is the traditional end-to-end method using full fine-tuning. (c) Tailored for the TAD task, our AdaTAD uses a lightweight temporal-informative adapter inside the backbone to achieve efficient transfer learning. (d) To further reduce memory usage and scale up the model/data, $\mathrm{AdaTAD}^{\dagger}$ uses an alternative placement for adapters outside the backbone. + +![](images/ed78bdde4dc9859c9e785005426cb4ab21e0450f9b2c312bab69f72b787c6052.jpg) + +![](images/8707905ec3b0e326193625e371ea4aef9d949529e19d6255d017968fc020bd4e.jpg) + +![](images/f000f40ce7207fe11019c00cc44ec41eccf5abf7fdd97ad02acfd15a9593980e.jpg) + +and video understanding, architectures such as ViT [14] and MViT [16], have also witnessed the effectiveness of this scaling strategy. Alabdulmohsin et al.[1] present a recipe for estimating scaling law parameters reliably from learning curves in computer vision. To attain even greater performance, several studies have also scaled up image resolution [35] or video clip length [56]. In this paper, we successfully apply this principle within the domain of temporal action detection and achieve state-of-the-art results. + +Efficient Transfer Learning. Transfer learning aims to adapt a pretrained model to a new domain. In TAD, typically off-the-shelf action recognition models are employed as the backbone, such as SlowFast [17]. Traditional transfer learning adopts full fine-tuning, meaning that all parameters of the pretrained model are updated. However, studies by [43, 63] have noted that full fine-tuning may harm the pre-learned knowledge, particularly when the downstream dataset is small and less comprehensive. Moreover, as models increase in size, the computational and storage demands of full fine-tuning proportionally increase. + +In response to these challenges, several works have investigated parameter-efficient tuning (PEFT) strategies that involve fine-tuning only a fraction of the network. For instance, Adapter [22] inserts lightweight modules analogous to feedforward networks in transformers, and only tunes these elements. LoRA [23] employs low-rank matrices in each transformer layer. Prefix-tuning [28] and prompt-tuning [27] attach learnable prompt tokens at the input stage or within each layer. In computer vision, many PEFT methods [6, 7, 24, 63] have also been explored across various tasks to optimize transfer learning efficiency. Our work represents the first effort to examine the potential of the PEFT mechanism in TAD. + +Although PEFT effectively reduces the number of learnable parameters, data-intensive and computationally heavy + +tasks like video understanding require more memory-efficient techniques. To this end, several works try to externalize the trainable components from the backbone, eliminating the need for backpropagation through the extensive original model. For example, LST [49] introduces a supplementary lightweight network that operates in parallel with the main model. Similarly, $\mathrm{E}^3\mathrm{VA}$ [65] leverages intermediate features with adapters to enable efficient transfer learning while minimizing memory usage. Our work is inspired by these methods yet with a streamlined and simple design. + +# 3. Methodology + +In this section, we introduce our AdaTAD step-by-step. We first introduce notations and study the efficient video representation to establish an end-to-end TAD baseline. Next, we introduce a temporal-informative adapter designed for efficient TAD. Finally, we propose an alternative placement for adapters to further alleviate computational demands. + +# 3.1. Notations + +Temporal action detection can be formulated as follows: given an untrimmed video $\mathbf{X} \in \mathbb{R}^{3 \times H \times W \times T}$ , where $H$ and $W$ are the height and width of each frame, and $T$ is the frame number, its temporal action annotations can be denoted as $\Psi_{g} = \{\varphi_{i} = (t_{s}, t_{e}, c)\}_{i=1}^{N}$ , where $t_{s}, t_{e}, c$ are the start, end time and category of action instance $\varphi_{i}$ , and $N$ is the number of ground truth actions. TAD aims to predict candidate proposal set $\Psi_{p} = \{\hat{\varphi}_{i} = (\hat{t}_{s}, \hat{t}_{e}, \hat{c}, s)\}_{i=1}^{M}$ to cover $\Psi_{g}$ , and $s$ is the confidence score. + +# 3.2. Frame-level representation + +Our end-to-end TAD architecture comprises two main components: feature extraction and action detection. Following previous work [69], we select ActionFormer [66] as our ac + +tion detection head due to its robust performance across various datasets without much hyperparameter tuning. Next, we discuss two ways of encoding raw frames into representative features (feature extraction): snippet representation and frame representation. + +Snippets Representation. Snippet-based video representations are popular choices in offline feature extraction. The whole video is divided into several short snippets (or namely clips). Each snippet has a short temporal length, e.g., 16 frames, and different snippets can have overlapping frames. Thus, the video can be conceptualized as $T$ snippets, denoted by $\mathbf{X} \in \mathbb{R}^{T \times 3 \times 16 \times H \times W}$ . Each snippet is processed through the video backbone, and spatial-temporal pooling is applied to extract one snippet feature. This processing yields the feature representation $\mathbf{F} \in \mathbb{R}^{T \times C}$ , where $C$ denotes the channel dimension of the pooled features. + +Frame Representation. In contrast to snippet-based representations, frame-based video representations consider the entire video as a singular snippet or clip, represented as $\mathbf{X} \in \mathbb{R}^{1 \times 3 \times T \times H \times W}$ . Then, the whole frame sequence is fed into the video backbone, and only spatial pooling is employed to extract features [10, 29, 69]. For attention-based models such as VideoMAE [54], the video is chunked into multiple shorter clips to avoid extensive temporal attention. + +Although both representations have been employed in previous studies, a fair comparison between them has not yet been performed. To address this gap, we conduct a comparative analysis of the two representations under the same setting on THUMOS14, measuring their memory usage and detection mAP. The results in Table 1 indicate that frame representation has comparable or even better performance than snippet representation, yet with much smaller memory consumption. When the feature extraction backbones are frozen, frame representation yields superior results to snippet representation for both VideoMAE [53] and SlowFast [17] backbones. Only in the end-to-end set + +Table 1. Snippets representation vs frame representation. We use the end-to-end version of ActionFormer with two representations for comparison. The snippet input is $768 \times 3 \times 16 \times 160 \times 160$ , and the frame input is $1 \times 3 \times 768 \times 160 \times 160$ . * means activation checkpointing is utilized to avoid overflowing GPU memory. + +
SettingBackboneRepr.Avg. mAPMem (GB)
FrozenVideoMAE-SFrame59.351.9
Snippet57.6813.2
SlowFast-R101Frame61.343.6
Snippet60.2417.2
End to EndVideoMAE-SFrame67.152.8*
Snippet68.4624.6*
SlowFast-R101Frame65.335.5*
Snippet66.7251.6*
+ +ting can the snippet representation achieve $1\%$ mAP advantage over frame representations; however, it requires 8 times more memory consumption. Taking into account both performance and memory usage, frame-based representations could be a better choice for end-to-end TAD development. + +Therefore, we use frame representation as the default baseline to encode videos in our experiments. Following the previous TALLFormer work [10], we also incorporate activation checkpointing [8] and mixed precision training [37] to fully harness the potential of scaling. + +# 3.3. Temporal-Informative Adapter + +In Section 3.2, we have built a simple end-to-end baseline using full fine-tuning. However, the baseline still suffers from two aspects: 1. Increased computational cost. In Table 1, we only use small video backbones like VideoMAES. When scaling VideoMAE-S to larger models, the computational burden and memory cost will grow rapidly. 2. Inferior transfer learning ability. More critically, the baseline follows the tradition of full fine-tuning, which may lead to inferior transfer learning. Pointed out by [50, 63], full fine-tuning may result in overfitting or forgetting, especially for large pretrained models. If downstream datasets are not sufficiently diverse, full fine-tuning can even destroy the powerful features learned from large-scale pretraining. Motivated by the above two aspects, we apply the PEFT mechanism and propose to fine-tune a plug-and-play module named Temporal-Informative Adapter (TIA) to achieve efficient and effective transfer learning for TAD. + +We first review the architecture of the standard adapter proposed by [22]. As formulated in Equation 1, the standard adapter includes a down-projection fully connected (FC) layer with parameter $\mathbf{W}_{\mathrm{down}} \in \mathbb{R}^{d \times \frac{d}{\gamma}}$ , where $\frac{d}{\gamma}$ represents the intermediate dimension and satisfies $\gamma > 1$ . Then, an up-projection layer $\mathbf{W}_{\mathrm{up}} \in \mathbb{R}^{\frac{d}{\gamma} \times d}$ is employed to restore the channel dimension. Between these two FC layers, a non-linear activation function $\sigma$ is inserted, such as GELU [21]. Afterward, a residual connection is added to the output of the projection layer. Note that $\mathbf{x}$ and $\mathbf{x}'$ are the input and output features with the same shape $\mathbb{R}^{d \times t \times h \times w}$ . + +$$ +\boldsymbol {x} ^ {\prime} = \boldsymbol {W} _ {\mathrm {u p}} ^ {\top} \cdot \boldsymbol {\sigma} \left(\boldsymbol {W} _ {\mathrm {d o w n}} ^ {\top} \cdot \boldsymbol {x}\right) + \boldsymbol {x}. \tag {1} +$$ + +Although the adapter has achieved great success in natural language processing and computer vision, the standard adapter only focuses on adapting channel information, which neglects the temporal context vital for the TAD task. To address this limitation, we introduce the temporalinformative adapter, as depicted in Fig. 3(b). + +The architecture of TIA follows the general bottleneck design of the standard adapter, while we integrate the temporal depth-wise convolution layers, as described in Equation 2. The temporal convolution with a kernel size of $k$ is + +![](images/bca175eac0c4d2d439c5b27d1a329dee4d29c7558b725de0e20f59adf28392de.jpg) +(a) Standard Adapter +Figure 3. Architecture of (a) standard adapter and (b) our temporal-informative adapter. We incorporate temporal depth-wise convolution to aggregate context from adjacent frames. + +![](images/5b95fa8e623d1dee88ffdb06ef8f5387be369eeb3524eeb65e87242fa5349dac.jpg) +(b) Temporal-Informative Adapter + +designed to aggregate local informative context from adjacent frames and to enrich the representation of the current time step. Practically, this is achieved through the application of a 3D convolution with kernel size $(k,1,1)$ and group size $\frac{d}{\gamma}$ for depth-wise processing. Additionally, an FC layer with weight $W_{\mathrm{mid}}\in \mathbb{R}^{\frac{d}{\gamma}\times \frac{d}{\gamma}}$ is employed to facilitate information exchange across channels. At last, a learnable scalar $\alpha$ is introduced to adjust the amplitude of adapter's output. + +$$ +\bar {\boldsymbol {x}} = \boldsymbol {\sigma} (\boldsymbol {W} _ {\text {d o w n}} ^ {\top} \cdot \boldsymbol {x}), +$$ + +$$ +\hat {\boldsymbol {x}} = \boldsymbol {W} _ {\mathrm {m i d}} ^ {\top} \cdot \mathbf {D W C o n v} _ {k} (\bar {\boldsymbol {x}}) + \bar {\boldsymbol {x}}, \tag {2} +$$ + +$$ +\boldsymbol {x} ^ {\prime} = \alpha \cdot \boldsymbol {W} _ {\mathrm {u p}} ^ {\top} \cdot \hat {\boldsymbol {x}} + \boldsymbol {x}. +$$ + +As shown in Fig. 2(c), TIA is designed to be inserted between different backbone layers, e.g. between each ViT block of VideoMAE or each bottleneck block of SlowFast. To ensure the newly added connection does not affect the original network at the beginning of transfer learning, the weight and bias of the adapter's last projection layer $W_{\mathrm{up}}$ are initialized to 0. The learnable coefficient $\alpha$ is initialized to 1. The temporal kernel size $k$ is set to 3, and the channel downsampling ratio $\gamma$ is set to 4 by default. Under these settings, the additional trainable parameters coming from TIA only account for $4.7\%$ of the total parameters of the original backbone. Since this backbone is frozen when TIA is used, our proposed strategy constitutes a massive reduction in trainable parameters as compared to full fine-tuning. Our experiments show that TIA can achieve better performance than full fine-tuning with less memory usage. + +# 3.4. Alternative Placement for Adapter + +Although the previously described PEFT approach can reduce tunable parameters and memory usage, the gradient still needs to backpropagate over the entire backbone during training. This requirement limits our ability to scale-up the model size or input data size further. As highlighted in prior + +works [49, 65], if we can stop the gradient backpropagation within the original backbone, additional computational savings can be achieved. + +Inspired by this insight, we propose a placement strategy for adapters that position them externally to the backbone, rather than inserting them inside. As illustrated in Fig. 2(d), we utilize the previously introduced TIA module, but its output does not feed back into the middle of the original backbone. It is directly added to the backbone's final layer. This configuration eliminates the need for backpropagation through the original network, as gradients are only tracked to the shallow lightweight adapter. To further diminish computation, we observe that adapting only the last half of backbone layers yields comparable performance while reducing half of the adaptation cost. + +To distinguish the different variants, we name the standard adaption design as AdaTAD, and the alternative placement as AdaTAD†. The latter can be considered as a lite version of the former. Compared to directly injecting adapters into the backbone, AdaTAD† may lead to a slight performance drop. However, it enables us to leverage richer models and more data, which should effectively counter this possible drop. + +# 4. Experiments + +# 4.1. Datasets and Metrics + +We choose ActivityNet-1.3 [20], THUMOS14 [25], and Epic-Kitchens 100 [13] to evaluate our proposed approach. ActivityNet-1.3 and THUMOS14 are web-collected third-person untrimmed videos, consisting of 19,994 and 413 videos, respectively. EPIC-Kitchens 100 is collected from 700 egocentric videos. Since the action categories of EPIC-Kitchens 100 are more domain-specific and different from common pretraining data, achieving higher performance on this dataset is more challenging. Moreover, we also evaluate our method on the Ego4D-MQ [18] dataset, and the results can be found in the appendix. + +Following common practice, we report the mean Average Precision (mAP) at certain IoU thresholds and average mAP as the evaluation metrics. On ActivityNet1.3, the IoU thresholds are chosen from 0.5 to 0.95 with 10 steps. On THUMOS14, the threshold is chosen from $\{0.3,0.4,0.5,0.6,0.7\}$ . On EPIC-Kitchens 100, the threshold is set to $\{0.1,0.2,0.3,0.4,0.5\}$ . + +# 4.2. Implementation Details + +We implement our method with PyTorch 2.0 and MMAc-tion2 [12] with 4 A100 GPUs. By default, mixed-precision training and activation checkpointing are adopted to save memory. We use ActionFormer [66] as our detection head, and keep the hyper-parameters unchanged on each dataset. The learning rate for the adapter in the backbone is grid + +Table 2. Results on ActivityNet-1.3 and THUMOS14, measured by mAP (\%) at different tIoU thresholds. E2E refers to end-to-end training, and Mem refers to memory usage (GB) per video. On ActivityNet-1.3, our prediction is combined with CUHK [72] classification results. Specifically, * means we employ stronger video-level classification results used in InternVideo [57] for a fair comparison. We report our best results in bold, and the previous best results in underline, which was achieved by the feature-based method. The last row is achieved when only the last half of backbone layers are adapted; otherwise, full-layer adaptation will lead to out-of-memory on A100-80G. + +
MethodBackboneE2EFlowMemActivityNet-1.3THUMOS14
0.50.750.95Avg.0.30.40.50.60.7Avg.
BMN [30]TSNX✓-50.0734.788.2933.8556.047.438.829.720.538.5
TadTR [34]I3DX✓-49.1032.608.5032.3062.457.449.237.826.346.6
ActionFormer [66]SlowFast-R50XX-54.2637.048.1336.0278.773.365.254.639.762.3
ActionFormer [66]I3DX✓-53.5036.208.2035.6082.177.871.059.443.966.8
ASL [45]I3DX✓-54.1067.408.0036.2083.179.071.759.745.867.9
TriDet [47]I3DX✓-54.7038.008.4036.8083.680.172.962.447.469.3
VideoMAEv2 [56]VideoMAEv2-gXX----------69.6
InternVideo [57]VideoMAE-H+UniformerV2XX----39.00*-----71.5
AFSD [29]I3D✓✓1252.4035.306.5034.4067.362.455.543.731.152.0
E2E-TAD [33]SlowFast-R50✓X1250.4735.9910.3335.1069.464.356.046.434.954.2
BasicTAD [62]SlowOnly-R50✓X1251.2033.417.5733.1275.570.863.550.937.459.6
TALLFormer [10]VideoSwin-B✓X2954.1036.207.9035.6076.0-63.2-34.559.2
Re2TAL [69]Re2VideoSwin-T✓X2454.7537.819.0336.8077.071.562.449.736.359.4
AdaTADSlowFast-R50✓X4.355.2838.118.8737.1181.076.269.459.044.566.0
AdaTADVideoMAE-S✓X2.556.1538.999.0737.8584.580.271.660.946.968.8
AdaTADVideoMAE-B✓X4.956.7739.359.7138.3987.082.475.363.849.271.5
AdaTADVideoMAE-L✓X11.057.6940.5610.1339.2287.784.176.766.452.473.5
AdaTADVideoMAE-H✓X19.258.0440.559.7539.3788.985.378.666.952.574.4
AdaTADVideoMAEv2-g✓X29.958.4541.1610.4539.7989.585.878.967.352.674.8
AdaTAD† (1536×2242)VideoMAEv2-g✓X43.660.8242.699.8441.15*89.685.979.467.653.875.4
AdaTAD (1536×2242)VideoMAEv2-g✓X50.661.7243.3510.8541.93*89.786.780.971.056.176.9
+ +searched from 5e-4 to 5e-5, and other parameters inside the backbone are frozen. On ActivityNet-1.3, we resize the video into a fixed length of 768 frames. On THUMOS14, we randomly truncate a window with 768 frames with a temporal stride of 4. On EPIC-Kitchens 100, we randomly truncate a window with 6144 frames with a temporal stride of 2. After the video encoder, the feature is resized to fixed lengths of 192, 768, and 768, respectively, for the three datasets. Frame resolution is set to $160^2$ by default. In all experiments, we report the training memory usage. More implementation details can be found in the appendix. + +# 4.3. Comparison with SoTA Methods + +Table 2 compares our AdaTAD with other state-of-the-art (SoTA) methods on ActivityNet-1.3 and THUMOS14 datasets. Initially, we use SlowFast-R50 as the backbone. For comparison, we also extract corresponding offline features, utilizing the snippet representation where each snippet comprises 32 frames with $224^{2}$ resolution. We observe that end-to-end training enhances performance from $62.3\%$ to $66.0\%$ on THUMOS14. Notably, this architecture has also been used in E2E-TAD [33]. However, our method consumes less memory while achieving superior performance. This apple-to-apple comparison underscores the benefits of adapter tuning and the scaling-up principle. + +Table 3. Results on EPIC-Kitchens 100 validation set. For comparison, the feature-based methods use the same SlowFast-R50. + +
MethodE2E0.10.20.30.40.5Avg.
Verb Task
BMN [30]✘10.88.88.47.15.68.4
G-TAD [60]✘12.111.09.48.16.59.4
ActionFormer [66]✘26.625.424.222.319.123.5
ASL [45]✘27.9-25.5-19.824.6
TriDet [47]✘28.627.426.124.220.825.4
AdaTAD (SlowFast-R50)✓26.525.723.921.717.623.1
ActionFormer (VideoMAE-L)✘32.731.629.126.723.628.7
AdaTAD (VideoMAE-L)✓33.132.230.427.523.129.3
Noun Task
BMN [30]✘10.38.36.24.53.46.5
G-TAD [60]✘11.010.08.67.05.48.4
ActionFormer [66]✘25.224.122.720.517.021.9
ASL [45]✘26.0-23.4-17.722.6
TriDet [47]✘27.426.324.622.218.323.8
AdaTAD (SlowFast-R50)✓24.523.622.320.016.521.4
ActionFormer (VideoMAE-L)✘31.329.727.225.321.326.9
AdaTAD (VideoMAE-L)✓32.431.630.127.424.629.3
+ +Furthermore, when adopting the VideoMAE [53] family as our backbone and progressively scaling up the model size, the performance of AdaTAD consistently improves. Using the largest model, i.e., VideoMAEv2-giant with 1.01 + +Table 4. Compared to full fine-tuning, our adapter tuning can achieve better performance with less memory. Param. is the number of tunable parameters in the backbone. * means out of memory on A100-80GB, and we report the estimated number. + +
ModelSettingE2EParam.Mem.mAP
VideoMAE-SFeature×0-57.6
Snippet Full FT✓22M24.6G68.4
Frame Full FT✓22M2.8G67.1
AdaTAD✓1M2.5G68.8
VideoMAE-BFeature×0-64.7
Snippet Full FT✓86M87.4G*-
Frame Full FT✓86M5.6G70.1
AdaTAD✓4M4.9G71.5
VideoMAE-LFeature×0-66.5
Snippet Full FT✓304M193G*-
Frame Full FT✓304M13.1G73.0
AdaTAD✓14M11.0G73.5
+ +billion parameters, and larger input data, i.e., 1536 frames with $224^{2}$ resolution, we attain an impressive $41.9\%$ mAP on ActivityNet-1.3 and $76.9\%$ mAP on THUMOS14. It is noteworthy that the previous SoTA was achieved by VideoMAEv2 [56] and InternVideo [57], which utilize the same detector head as ours but with offline snippet features. Our method surpasses these in detection performance by a large margin, marking the first instance where an end-to-end TAD method can outperform SoTA feature-based results. + +We also present our results on EPIC-Kitchens 100 in Table 3. Since videos in this dataset have a longer duration, all previous methods rely solely on pre-extracted features [47, 52, 66]. Our approach is the first to adopt end-to-end training on this dataset. For fair comparison, we first utilize the same backbone as used in previous methods, i.e., SlowFast-R50 pretrained on EPIC, and we achieve comparable performance to ActionFormer [66]. Moreover, when we scale up the backbone to VideoMAE-L (it is also trained on EPIC-Kitchens 100 classification task), we achieve SoTA performance of $29.3\%$ . + +# 4.4. Ablation and Analysis + +In this section, we present a series of analyses to evaluate our proposed method and affirm the benefits of scaling up in TAD. Unless otherwise stated, our experiments utilize a standard input of 768 frames per video on THUMOS14. + +The advantage of adapter tuning. In Table 4, we compare conventional full fine-tuning with our proposed design. It is evident that end-to-end approaches significantly outperform pre-extracted features. Moreover, with full finetuning, the snippet representation slightly advances over frame representation but incurs tremendous memory costs, which aligns with our analysis in Section 3.2. However, AdaTAD uses less memory and still achieves better performance than conventional full fine-tuning. This also veri + +Table 5. When scaling up the input data, AdaTAD's performance consistently increases. * means snippet representation is used in offline feature extraction, and each snippet has 16 frames. + +
SettingModelResolutionFramesMem.mAP
FeatureVideoMAEv2-g2242768x16*-69.6
AdaTADVideoMAE-S16027682.5G68.8
160215363.8G69.7
160230726.5G70.6
22427683.8G70.7
224215366.4G71.0
2242307211.6G71.5
+ +Table 6. AdaTAD† can further push the boundaries of scaling up. OOM means out of memory on A100-80GB. + +
SettingModelResolutionFrameMem.mAP
AdaTADVideoMAE-L160276811.0G73.5
VideoMAEv2-g160276829.9G74.8
VideoMAEv2-g22421536OOM-
AdaTAD†VideoMAEv2-g160276822.8G73.7
VideoMAEv2-g224276830.0G74.6
VideoMAEv2-g2242153643.6G75.4
+ +fies the limitations of full fine-tuning, as discussed in Section 3.3. Specifically, our method enhances VideoMAE-S backbone with an $11.2\%$ gain using only 1M trainable parameters. Additionally, Table 4 also demonstrates that scaling up the model size of the video backbone is an effective way to improve TAD performance. + +The advantage of scaling up the data. In addition to model scaling, Table 5 verifies the effectiveness of data scaling, which involves two aspects: frame number and frame resolution. Firstly, given the same video duration, increasing the frame number from 768 to 3072 can raise the mAP from $68.8\%$ to $70.6\%$ . In the meantime, the memory usage is nearly three times larger. Secondly, increasing the frame resolution from $160^2$ to $224^2$ also improves the mAP. In the end, by only scaling up the data, we elevate the mAP from $68.8\%$ to $71.5\%$ , already surpassing the current SoTA feature-based approach with a giant backbone model [57]. + +Moreover, increasing the frame resolution from $160^2$ to $224^2$ or increasing the frame number from 768 to 1536 results in the same memory usage of 3.8G. However, the former achieves $70.7\%$ mAP while the latter only reaches $69.7\%$ . This suggests that frame resolution may be prioritized under the same memory budget, for the TAD task. + +The advantage of AdaTAD†. Given the effectiveness of scaling up the model or data, we further explore combining these approaches. In Table 6, using 768 frames while scaling up the model to VideoMAEv2-giant results in memory usage escalating to 29.9G. In such a scenario, further increasing the data could easily lead to memory overflow, even with the A100-80GB GPU. This indicates that adapta + +tion tuning reached its limit under this extreme case. Therefore, to utilize both the largest models with larger data simultaneously, $\mathrm{AdaTAD}^{\dagger}$ shows its advantage. + +Concretely, when switching from AdaTAD to AdaTAD†, the memory usage of the VideoMAEv2-giant model is reduced from 29.9G to 22.8G. Although a slight performance drop is observed, its reduced memory footprint enables scaling up data from 768 frames to as much as 1536 frames with a high resolution of $224^2$ . This scalability helps mitigate the performance drop and achieves a higher mAP of $75.4\%$ . + +The ablation of the adapter design. Detailed in Table 7, we compare different adapter architectural designs. The baseline, i.e., offline snippet feature, achieves $64.7\%$ mAP. End-to-end learning in all designs yields at least a $5\%$ improvement. Our AdaTAD achieves $71.5\%$ in the end. Compared to standard adapter [22], ours consumes similar memory but achieves higher mAP. This verifies that local temporal context is vital for the TAD task. In contrast to full finetuning (FT), we tune only 4M parameters using less memory. In our design, we find that removing the residual connection of the depth-wise convolution drops performance by $0.7\%$ , and training becomes unstable. We also implement an adaptation design proposed in LongLoRA [9], which efficiently computes long-range attention and shows decent performance but requires more parameters and memory. + +Table 7. Ablation of different adapter architectural designs. VideoMAE-B is used to conduct the following experiments. + +
SettingE2EParam.Mem.mAPgains
Snippet FeatureX0-64.7
+ Full FT✓86M5.6G70.1+5.1
+ LongLoRA [9]✓28M6.2G71.1+6.1
+ Standard Adapter [22]✓3.6M4.8G70.2+5.2
+ AdaTAD (w/o residual)✓4.0M4.9G70.8+5.8
+ AdaTAD✓4.0M4.9G71.5+6.5
+ +The necessity of end-to-end training for TAD. As previously discussed, end-to-end training can address discrepancies between the pretraining and fine-tuning stages in terms of training data and learning tasks. To corroborate this, we employ models pretrained on different datasets for the EPIC-Kitchens TAD task in Table 8. Kinetics-400 (K400) [26] represents commonly collected third-person web data and exhibits a large domain gap compared to EPIC-Kitchens 100. Using K400 for pretraining, we observe that end-to-end TAD training allows for $+5.56$ gain. Conversely, using a model already finetuned on EPIC-Kitchens still yields a $+2.32$ improvement. Unlike K400 pretraining, since this model has already adapted to the data discrepancy, we can infer that this gain leverages differences between the classification task in pretraining and the detection task in fine-tuning. Such improvements further underscore the significance of end-to-end training in TAD. + +Table 8. End-to-end TAD can alleviate the discrepancies between pretraining and finetuning. VideoMAE-L with different pretrained weights are used on the EPIC-Kitchens 100 Noun task. + +
Pretrain DatasetE2E0.10.30.5mAPgain
K400 [26]X18.6916.3511.5215.77
K400 [26]✓24.3322.1416.8721.33+ 5.56
K400 [26] + EPIC [13]X31.3227.2521.3326.98
K400 [26] + EPIC [13]✓32.4130.1324.5929.30+ 2.32
+ +# 4.5. Error Analyses + +We also conduct false positive analysis at tIoU=0.5 in Fig. 4. Compared to feature-based training, learning from raw frames produces more helpful positive detections. More importantly, the percentage of wrong label error is reduced after end-to-end training, suggesting its unique advantage in classifying accurate action labels. + +![](images/f12d6fe5149da566afa57be00d00b5ae3e7c600bcfcdc598394dfe43272757c7.jpg) +Figure 4. False Positive Profiling on THUMOS14 using [2]. We use VideoMAEv2-giant as the backbone, and compare end-to-end training with pre-extracted feature-based training. + +# 5. Conclusions + +This work introduces a memory-efficient and parameter-efficient end-to-end method named AdaTAD. Our key innovation lies in the proposed temporal-informative adapter, which is tailored for TAD with low computation costs. Furthermore, we design an alternative placement for adapters to minimize memory usage. By demonstrating the feasibility and effectiveness of scaling up end-to-end TAD, our work achieves new SoTA performance across multiple datasets. Particularly, this is the first instance of an end-to-end TAD method that surpasses the current best feature-based models. In fact, AdaTAD achieves a groundbreaking $75.4\%$ mAP on THUMOS14. We believe our work underscores the possible paradigm shift in TAD, advocating a move away from the traditional methodology of separate feature extraction and offline detection towards a more integrated approach of scaling up end-to-end TAD training. + +Acknowledgement. This work was supported by the King Abdullah University of Science and Technology (KAUST) Office of Sponsored Research through the Visual Computing Center (VCC) funding, as well as the SDAIA-KAUST Center of Excellence in Data Science and Artificial Intelligence (SDAIA-KAUST AI). + +# References + +[1] Ibrahim M Alabdulmohsin, Behnam Neyshabur, and Xiaohua Zhai. Revisiting neural scaling laws in language and vision. In NeurIPS, 2022. 3 +[2] Humam Alwassel, Fabian Caba Heilbron, Victor Escorcia, and Bernard Ghanem. Diagnosing error in temporal action detectors. In ECCV, 2018. 8 +[3] Humam Alwassel, Fabian Caba Heilbron, and Bernard Ghanem. Action search: Spotting actions in videos and its application to temporal action localization. In ECCV, 2018. 1 +[4] Yueran Bai, Yingying Wang, Yunhai Tong, Yang Yang, Qiyue Liu, and Junhui Liu. Boundary content graph neural network for temporal action proposal generation. In ECCV, 2020. 2 +[5] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. In NeurIPS, 2020. 2 +[6] Hao Chen, Ran Tao, Han Zhang, Yidong Wang, Wei Ye, Jindong Wang, Guosheng Hu, and Marios Savvides. Convadapter: Exploring parameter efficient transfer learning for convnets. arXiv preprint arXiv:2208.07463, 2022. 3 +[7] Shoufa Chen, Chongjian Ge, Zhan Tong, Jiangliu Wang, Yibing Song, Jue Wang, and Ping Luo. Adaptformer: Adapting vision transformers for scalable visual recognition. In NeurIPS, 2022. 3 +[8] Tianqi Chen, Bing Xu, Chiyuan Zhang, and Carlos Guestrin. Training deep nets with sublinear memory cost. arXiv preprint arXiv:1604.06174, 2016. 4 +[9] Yukang Chen, Shengju Qian, Haotian Tang, Xin Lai, Zhi-jian Liu, Song Han, and Jiaya Jia. Longlora: Efficient fin-tuning of long-context large language models. arXiv preprint arXiv:2309.12307, 2023. 8 +[10] Feng Cheng and Gedas Bertasius. Tallformer: Temporal action localization with long-memory transformer. In ECCV, 2022. 1, 2, 4, 6 +[11] Feng Cheng, Mingze Xu, Yuanjun Xiong, Hao Chen, Xinyu Li, Wei Li, and Wei Xia. Stochastic backpropagation: A memory efficient strategy for training video models. In CVPR, 2023. 2 +[12] MMAction2 Contributors. Openmmlab's next generation video understanding toolbox and benchmark. https://github.com/open-mmlab/mmaction2, 2020.5 +[13] Dima Damen, Hazel Doughty, Giovanni Maria Farinella, Sanja Fidler, Antonino Furnari, Evangelos Kazakos, Davide Moltisanti, Jonathan Munro, Toby Perrett, Will Price, et al. Scaling egocentric vision: The epic-kitchens dataset. In ECCV, 2018. 5, 8 +[14] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2021. 1, 3 +[15] Victor Escorcia, Fabian Caba Heilbron, Juan Carlos Niebles, + +and Bernard Ghanem. DAPs: Deep action proposals for action understanding. In ECCV, 2016. 1 +[16] Haoqi Fan, Bo Xiong, Karttikeya Mangalam, Yanghao Li, Zhicheng Yan, Jitendra Malik, and Christoph Feichtenhofer. Multiscale vision transformers. In ICCV, 2021. 3 +[17] Christoph Feichtenhofer, Haoqi Fan, Jitendra Malik, and Kaiming He. SlowFast networks for video recognition. In ICCV, 2019. 3, 4 +[18] Kristen Grauman, Andrew Westbury, Eugene Byrne, Zachary Chavis, Antonino Furnari, Rohit Girdhar, Jackson Hamburger, Hao Jiang, Miao Liu, Xingyu Liu, et al. Ego4d: Around the world in 3,000 hours of egocentric video. In CVPR, 2022. 5 +[19] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, and Ross Girshick. Masked autoencoders are scalable vision learners. In CVPR, 2022. 1 +[20] Fabian Caba Heilbron, Victor Escorcia, Bernard Ghanem, and Juan Carlos Niebles. ActivityNet: A large-scale video benchmark for human activity understanding. In CVPR, 2015. 1, 5 +[21] Dan Hendrycks and Kevin Gimpel. Gaussian error linear units (gelus). arXiv preprint arXiv:1606.08415, 2016. 4 +[22] Neil Houlsby, Andrei Giurgiu, Stanislaw Jastrzebski, Bruna Morrone, Quentin De Laroussilhe, Andrea Gesmundo, Mona Attariyan, and Sylvain Gelly. Parameter-efficient transfer learning for nlp. In ICML, 2019. 2, 3, 4, 8 +[23] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. In ICML, 2021. 3 +[24] Menglin Jia, Luming Tang, Bor-Chun Chen, Claire Cardie, Serge Belongie, Bharath Hariharan, and Ser-Nam Lim. Visual prompt tuning. In ECCV, 2022. 3 +[25] YG Jiang, J Liu, A Roshan Zamir, G Toderici, I Laptev, M Shah, and R Sukthankar. Thumos challenge: Action recognition with a large number of classes, 2014. 1, 5 +[26] Will Kay, Joao Carreira, Karen Simonyan, Brian Zhang, Chloe Hillier, Sudheendra Vijayanarasimhan, Fabio Viola, Tim Green, Trevor Back, Paul Natsev, et al. The kinetics human action video dataset. arXiv preprint arXiv:1705.06950, 2017.8 +[27] Brian Lester, Rami Al-Rfou, and Noah Constant. The power of scale for parameter-efficient prompt tuning. In EMNLP, 2021. 3 +[28] Xiang Lisa Li and Percy Liang. Prefix-tuning: Optimizing continuous prompts for generation. In ACL, 2021. 3 +[29] Chuming Lin, Chengming Xu, Donghao Luo, Yabiao Wang, Ying Tai, Chengjie Wang, Jilin Li, Feiyue Huang, and Yanwei Fu. Learning salient boundary feature for anchor-free temporal action localization. In CVPR, 2021. 1, 2, 4, 6 +[30] Tianwei Lin, Xiao Liu, Xin Li, Errui Ding, and Shilei Wen. BMN: boundary-matching network for temporal action proposal generation. In ICCV, 2019. 1, 2, 6 +[31] Qinying Liu and Zilei Wang. Progressive boundary refinement network for temporal action detection. In AAAI, 2020. 2 + +[32] Shuming Liu, Mengmeng Xu, Chen Zhao, Xu Zhao, and Bernard Ghanem. Etad: Training action detection end to end on a laptop. In CVPRW, 2023. 1, 2 +[33] Xiaolong Liu, Song Bai, and Xiang Bai. An empirical study of end-to-end temporal action detection. In CVPR, 2022. 1, 2, 6 +[34] Xiaolong Liu, Qimeng Wang, Yao Hu, Xu Tang, Shiwei Zhang, Song Bai, and Xiang Bai. End-to-end temporal action detection with transformer. IEEE Transactions on Image Processing, 31:5427-5441, 2022. 2, 6 +[35] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In ICCV, 2021. 3 +[36] Jinjie Mai, Abdullah Hamdi, Silvio Giancola, Chen Zhao, and Bernard Ghanem. Egoloc: Revisiting 3d object localization from egocentric videos with visual queries. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 45-57, 2023. 1 +[37] Paulius Micikevicius, Sharan Narang, Jonah Alben, Gregory Diamos, Erich Elsen, David Garcia, Boris Ginsburg, Michael Houston, Oleksii Kuchaiev, Ganesh Venkatesh, et al. Mixed precision training. In ICLR, 2017. 4 +[38] Jonghwan Mun, Minsu Cho, and Bohyung Han. Local-global video-text interactions for temporal grounding. In CVPR, 2020. 1 +[39] OpenAI. Gpt-4 technical report, 2023. 1, 2 +[40] Zhiwu Qing, Haisheng Su, Weihao Gan, Dongliang Wang, Wei Wu, Xiang Wang, Yu Qiao, Junjie Yan, Changxin Gao, and Nong Sang. Temporal context aggregation network for temporal action proposal refinement. In CVPR, 2021. 2 +[41] Alec Radford, Karthik Narasimhan, Tim Salimans, Ilya Sutskever, et al. Improving language understanding by generative pre-training. OpenAI, 2018. 2 +[42] Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, Ilya Sutskever, et al. Language models are unsupervised multitask learners. OpenAI blog, 1(8):9, 2019. 2 +[43] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, 2021. 3 +[44] Merey Ramazanova, Victor Escorcia, Fabian Caba Heilbron, Chen Zhao, and Bernard Ghanem. Owl (observe, watch, listen): Localizing actions in egocentric video via audiovisual temporal context. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshop (CVPRW), 2023. 1 +[45] Jiayi Shao, Xiaohan Wang, Ruijie Quan, Junjun Zheng, Jiang Yang, and Yi Yang. Action sensitivity learning for temporal action localization. In ICCV, 2023. 2, 6 +[46] Dingfeng Shi, Yujie Zhong, Qiong Cao, Jing Zhang, Lin Ma, Jia Li, and Dacheng Tao. React: Temporal action detection with relational queries. In ECCV, 2022. 2 +[47] Dingfeng Shi, Yujie Zhong, Qiong Cao, Lin Ma, Jia Li, and Dacheng Tao. Tridet: Temporal action detection with relative boundary modeling. In CVPR, 2023. 2, 6, 7 + +[48] Mattia Soldan, Mengmeng Xu, Sisi Qu, Jesper Tegner, and Bernard Ghanem. VLG-Net: Video-language graph matching network for video grounding. In ICCVW, 2021. 1 +[49] Yi-Lin Sung, Jaemin Cho, and Mohit Bansal. Lst: Ladder side-tuning for parameter and memory efficient transfer learning. In NeurIPS, 2022. 3, 5 +[50] Yi-Lin Sung, Jaemin Cho, and Mohit Bansal. VI-adapter: Parameter-efficient transfer learning for vision-and-language tasks. In CVPR, 2022. 2, 4 +[51] Jing Tan, Jiaqi Tang, Limin Wang, and Gangshan Wu. Relaxed transformer decoders for direct action proposal generation. In ICCV, 2021. 2 +[52] Tuan N Tang, Kwonyoung Kim, and Kwanghoon Sohn. Temporalmaxer: Maximize temporal context with only max pooling for temporal action localization. arXiv preprint arXiv:2303.09055, 2023. 7 +[53] Zhan Tong, Yibing Song, Jue Wang, and Limin Wang. Videomae: Masked autoencoders are data-efficient learners for self-supervised video pre-training. In NeurIPS, 2022. 1, 4, 6 +[54] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In NeurIPS, 2017. 4 +[55] Chenhao Wang, Hongxiang Cai, Yuxin Zou, and Yichao Xiong. Rgb stream is enough for temporal action detection. arXiv preprint arXiv:2107.04362, 2021. 2 +[56] Limin Wang, Bingkun Huang, Zhiyu Zhao, Zhan Tong, Yi-nan He, Yi Wang, Yali Wang, and Yu Qiao. Videomae v2: Scaling video masked autoencoders with dual masking. In ICCV, 2023. 1, 3, 6, 7 +[57] Yi Wang, Kunchang Li, Yizhuo Li, Yinan He, Bingkun Huang, Zhiyu Zhao, Hongjie Zhang, Jilan Xu, Yi Liu, Zun Wang, et al. Internvideo: General video foundation models via generative and discriminative learning. arXiv preprint arXiv:2212.03191, 2022. 1, 6, 7 +[58] Zhenzhi Wang, Ziteng Gao, Limin Wang, Zhifeng Li, and Gangshan Wu. Boundary-aware cascade networks for temporal action segmentation. In ECCV, 2020. 2 +[59] Kun Xia, Le Wang, Sanping Zhou, Nanning Zheng, and Wei Tang. Learning to refactor action and co-occurrence features for temporal action localization. In CVPR, 2022. 2 +[60] Mengmeng Xu, Chen Zhao, David S Rojas, Ali Thabet, and Bernard Ghanem. G-TAD: Sub-graph localization for temporal action detection. In CVPR, 2020. 1, 6 +[61] Le Yang, Houwen Peng, Dingwen Zhang, Jianlong Fu, and Junwei Han. Revisiting anchor mechanisms for temporal action localization. IEEE Transactions on Image Processing, 29:8535-8548, 2020. 2 +[62] Min Yang, Guo Chen, Yin-Dong Zheng, Tong Lu, and Limin Wang. Basictad: an astounding rgb-only baseline for temporal action detection. Computer Vision and Image Understanding, 232:103692, 2023. 1, 2, 6 +[63] Taojiannan Yang, Yi Zhu, Yusheng Xie, Aston Zhang, Chen Chen, and Mu Li. Aim: Adapting image models for efficient video action recognition. In ICLR, 2023. 2, 3, 4 +[64] Ting Yao, Tao Mei, and Yong Rui. Highlight detection with pairwise deep ranking for first-person video summarization. In CVPR, 2016. 1 + +[65] Dongshuo Yin, Xueting Han, Bin Li, Hao Feng, and Jing Bai. Parameter-efficient is not sufficient: Exploring parameter, memory, and time efficient adapter tuning for dense predictions. arXiv preprint arXiv:2306.09729, 2023. 3, 5 +[66] Chenlin Zhang, Jianxin Wu, and Yin Li. Actionformer: Localizing moments of actions with transformers. In ECCV, 2022. 1, 2, 3, 5, 6, 7 +[67] Chen Zhao, Ali K Thabet, and Bernard Ghanem. Video self-stitching graph network for temporal action localization. In ICCV, 2021. 1, 2 +[68] Chen Zhao, Merey Ramazanova, Mengmeng Xu, and Bernard Ghanem. Segstad: Precise temporal action detection via semantic segmentation. In ECCVW, 2022. 2 +[69] Chen Zhao, Shuming Liu, Karttikeya Mangalam, and Bernard Ghanem. Re $^2$ TAL: Rewiring pretrained video backbones for reversible temporal action localization. In CVPR, 2023. 1, 2, 3, 4, 6 +[70] Chen Zhao, Shuming Liu, Karttikeya Mangalam, Guocheng Qian, Fatimah Zohra, Abdulmohsen Alghannam, Jitendra Malik, and Bernard Ghanem. $\mathrm{Dr}^2\mathrm{Net}$ : Dynamic reversible dual-residual networks for memory-efficient finetuning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 1, 2 +[71] Hang Zhao, Zhicheng Yan, Lorenzo Torresani, and Antonio Torralba. HACS: Human action clips and segments dataset for recognition and temporal localization. ICCV, 2019. 1 +[72] Y Zhao, B Zhang, Z Wu, S Yang, L Zhou, S Yan, L Wang, Y Xiong, D Lin, Y Qiao, et al. Cuhk & ethz & siat submission to activitynet challenge 2017. CVPR ActivityNet Workshop, 2017. 6 +[73] Zixuan Zhao, Dongqi Wang, and Xu Zhao. Movement enhancement toward multi-scale video feature representation for temporal action detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 13555-13564, 2023. 2 \ No newline at end of file diff --git a/2024/End-to-End Temporal Action Detection with 1B Parameters Across 1000 Frames/images.zip b/2024/End-to-End Temporal Action Detection with 1B Parameters Across 1000 Frames/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..45846701ea39a3cf04b25e1cb294a90cc8257b9d --- /dev/null +++ b/2024/End-to-End Temporal Action Detection with 1B Parameters Across 1000 Frames/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3dc4da59d6de18923d7762e5472e9ffc919be09b4ada371e0e903d7306862f23 +size 645173 diff --git a/2024/End-to-End Temporal Action Detection with 1B Parameters Across 1000 Frames/layout.json b/2024/End-to-End Temporal Action Detection with 1B Parameters Across 1000 Frames/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..173edbe550b6736e95f74c9287ccc390aebdaf18 --- /dev/null +++ b/2024/End-to-End Temporal Action Detection with 1B Parameters Across 1000 Frames/layout.json @@ -0,0 +1,8681 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 48, + 103, + 545, + 120 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 103, + 545, + 120 + ], + "spans": [ + { + "bbox": [ + 48, + 103, + 545, + 120 + ], + "type": "text", + "content": "End-to-End Temporal Action Detection with 1B Parameters Across 1000 Frames" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 93, + 142, + 497, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 93, + 142, + 497, + 171 + ], + "spans": [ + { + "bbox": [ + 93, + 142, + 497, + 171 + ], + "type": "text", + "content": "Shuming Liu1 Chen-Lin Zhang2 Chen Zhao1* Bernard Ghanem1 \n1King Abdullah University of Science and Technology (KAUST) 24Paradigm Inc" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 143, + 199, + 191, + 211 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 199, + 191, + 211 + ], + "spans": [ + { + "bbox": [ + 143, + 199, + 191, + 211 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 45, + 224, + 290, + 487 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 224, + 290, + 487 + ], + "spans": [ + { + "bbox": [ + 45, + 224, + 290, + 487 + ], + "type": "text", + "content": "Recently, temporal action detection (TAD) has seen significant performance improvement with end-to-end training. However, due to the memory bottleneck, only models with limited scales and limited data volumes can afford end-to-end training, which inevitably restricts TAD performance. In this paper, we reduce the memory consumption for end-to-end training, and manage to scale up the TAD backbone to 1 billion parameters and the input video to 1,536 frames, leading to significant detection performance. The key to our approach lies in our proposed temporal-informative adapter (TIA), which is a novel lightweight module that reduces training memory. Using TIA, we free the humongous backbone from learning to adapt to the TAD task by only updating the parameters in TIA. TIA also leads to better TAD representation by temporally aggregating context from adjacent frames throughout the backbone. We evaluate our model across four representative datasets. Owing to our efficient design, we are able to train end-to-end on VideoMAEv2-giant and achieve " + }, + { + "bbox": [ + 45, + 224, + 290, + 487 + ], + "type": "inline_equation", + "content": "75.4\\%" + }, + { + "bbox": [ + 45, + 224, + 290, + 487 + ], + "type": "text", + "content": " mAP on THUMOS14, being the first end-to-end model to outperform the best feature-based methods. Code is available at https://github.com/sming256/AdaTAD." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 509, + 128, + 521 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 509, + 128, + 521 + ], + "spans": [ + { + "bbox": [ + 47, + 509, + 128, + 521 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 529, + 287, + 660 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 529, + 287, + 660 + ], + "spans": [ + { + "bbox": [ + 46, + 529, + 287, + 660 + ], + "type": "text", + "content": "Temporal Action Detection (TAD) plays a vital role in the understanding of long-form videos. Its objective is to pinpoint specific action instances within untrimmed videos, identifying their start and end times, along with their respective categories [15, 20, 44, 60, 62, 67, 71]. This task is crucial for various applications, including highlight detection [36, 64], video-language grounding [38, 48], and action spotting [3]. Though innovations in the detector design have made profound progress in the past years [30, 66, 67], recent research highlights two new trends: end-to-end training [10, 32, 69], and scaling up [56, 57]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 661, + 287, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 661, + 287, + 696 + ], + "spans": [ + { + "bbox": [ + 46, + 661, + 287, + 696 + ], + "type": "text", + "content": "End-to-end training in TAD refers to jointly training the video encoder and action detector [10, 29, 62, 70]. Compared to feature-based methods, end-to-end training offers" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 310, + 201, + 541, + 370 + ], + "blocks": [ + { + "bbox": [ + 310, + 201, + 541, + 370 + ], + "lines": [ + { + "bbox": [ + 310, + 201, + 541, + 370 + ], + "spans": [ + { + "bbox": [ + 310, + 201, + 541, + 370 + ], + "type": "image", + "image_path": "dae658e7cd312bba0b207001ec1b34728d4ec6ff778d8adbd6430ab83dafccda.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 373, + 547, + 451 + ], + "lines": [ + { + "bbox": [ + 304, + 373, + 547, + 451 + ], + "spans": [ + { + "bbox": [ + 304, + 373, + 547, + 451 + ], + "type": "text", + "content": "Figure 1. AdaTAD enjoys the benefit of end-to-end training and scaling up with efficient memory usage. The bubble size represents the number of the model's learnable parameters. Using SlowFast-R50, AdaTAD achieves better performance compared to E2E-TAD [33] with less memory. When further scaling up the model to VideoMAEv2-gaint and data to 1536 frames, we achieve an impressive avg. mAP of " + }, + { + "bbox": [ + 304, + 373, + 547, + 451 + ], + "type": "inline_equation", + "content": "75.4\\%" + }, + { + "bbox": [ + 304, + 373, + 547, + 451 + ], + "type": "text", + "content": " on THUMOS14." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 461, + 545, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 461, + 545, + 521 + ], + "spans": [ + { + "bbox": [ + 304, + 461, + 545, + 521 + ], + "type": "text", + "content": "distinct advantages. First, it can effectively bridge the gap commonly found between pretraining and fine-tuning, such as data and task discrepancy. Second, video spatial augmentations can be utilized in the end-to-end setting, leading to further performance gain." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 521, + 545, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 521, + 545, + 629 + ], + "spans": [ + { + "bbox": [ + 304, + 521, + 545, + 629 + ], + "type": "text", + "content": "Scaling up refers to improving performance by increasing the model size or the input data volume, and has demonstrated its effectiveness in various domains [14, 19, 39, 53]. In TAD, offline methods have attempted to scale up the feature extraction network to reach a higher performance. A notable example includes the work by Wang et al.[56], which reports a " + }, + { + "bbox": [ + 304, + 521, + 545, + 629 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 304, + 521, + 545, + 629 + ], + "type": "text", + "content": " increase in mean Average Precision (mAP) by scaling from VideoMAE-S to VideoMAEv2-giant, using ActionFormer [66] on THUMOS14 [25]." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 630, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 630, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 630, + 547, + 714 + ], + "type": "text", + "content": "Intuitively, combining the strengths of end-to-end training and scaling up is expected to be most beneficial for improving TAD performance. However, both strategies demand substantial GPU memory, which restricts end-to-end training to a small model [29, 62, 69], or a small input volume [10, 33]. As shown in Fig. 1, the performance of previous end-to-end methods still significantly lags behind the" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 702, + 135, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 702, + 135, + 712 + ], + "spans": [ + { + "bbox": [ + 58, + 702, + 135, + 712 + ], + "type": "text", + "content": "*Corresponding author." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18591" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 156 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 156 + ], + "type": "text", + "content": "best results achieved by feature-based approaches. Additionally, current end-to-end methods in TAD use computationally intensive full fine-tuning, risking the issues of catastrophic forgetting and overfitting during transfer learning [50, 63]. These issues can result in less competitive performance, especially when the downstream datasets are small, which is a common scenario in the TAD domain." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 157, + 287, + 264 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 157, + 287, + 264 + ], + "spans": [ + { + "bbox": [ + 46, + 157, + 287, + 264 + ], + "type": "text", + "content": "In this paper, we aim to overcome the above limitations by harnessing the advantages of both scaling-up and end-to-end training. To achieve this, we introduce adapter tuning for temporal action detection (AdaTAD). Our method successfully trains a TAD model in an end-to-end manner, utilizing a backbone with 1 billion parameters and processing input videos of 1,536 frames. As illustrated in Fig. 1, to the best of our knowledge, this is the first end-to-end work that outperforms the best feature-based TAD methods." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 266, + 288, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 266, + 288, + 540 + ], + "spans": [ + { + "bbox": [ + 46, + 266, + 288, + 540 + ], + "type": "text", + "content": "Specifically, we employ the following strategies to enhance the TAD performance while maintaining reasonable memory consumption. First, we identify that the snippet representation commonly used in feature-based methods is excessively redundant. In response, we have adopted a more memory-efficient frame-representation scheme, establishing an effective end-to-end baseline for TAD. Second, we adopt the parameter-efficient fine-tuning technique to minimize memory usage and mitigate overfitting in transfer learning. Notably, we introduce a novel Temporal-Informative Adapter (TIA). This adapter is injected between backbone layers and is the only learnable component in the backbone during fine-tuning. Different from conventional adapters [22], TIA is tailored for the TAD task and integrates temporal depth-wise convolutions to aggregate informative context from adjacent frames. Third, for additional memory efficiency, we propose a lighter variant of our method. By positioning the TIAs alongside the original backbone, rather than inside it, backpropagation can be done through the TIAs only. This configuration can further cut down on the need for intermediate activations within the primary backbone, thereby allowing us to scale up the model size and data size to unprecedented levels." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 542, + 287, + 638 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 542, + 287, + 638 + ], + "spans": [ + { + "bbox": [ + 46, + 542, + 287, + 638 + ], + "type": "text", + "content": "AdaTAD establishes a new state-of-the-art across multiple TAD datasets. Notably, our method achieves an impressive " + }, + { + "bbox": [ + 46, + 542, + 287, + 638 + ], + "type": "inline_equation", + "content": "75.4\\%" + }, + { + "bbox": [ + 46, + 542, + 287, + 638 + ], + "type": "text", + "content": " mAP on THUMOS14, outperforming the previous feature-based best result of " + }, + { + "bbox": [ + 46, + 542, + 287, + 638 + ], + "type": "inline_equation", + "content": "71.5\\%" + }, + { + "bbox": [ + 46, + 542, + 287, + 638 + ], + "type": "text", + "content": " by a large margin. This achievement underscores the possible paradigm shift in TAD, i.e., moving from traditional feature extraction plus offline detectors to scaling up end-to-end TAD training. We summarize our contribution as follows:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 641, + 287, + 714 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 47, + 641, + 287, + 701 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 641, + 287, + 701 + ], + "spans": [ + { + "bbox": [ + 47, + 641, + 287, + 701 + ], + "type": "text", + "content": "1. We introduce an efficient end-to-end framework for TAD, scaling up the model size to 1 billion parameters and the input data to 1,536 frames. We achieve a consistent performance improvement with the scaling up, shedding light on the importance of scaling for TAD." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 701, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 701, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 701, + 287, + 714 + ], + "type": "text", + "content": "2. We propose a novel temporal-informative adapter to re" + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 317, + 72, + 545, + 132 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 72, + 545, + 132 + ], + "spans": [ + { + "bbox": [ + 317, + 72, + 545, + 132 + ], + "type": "text", + "content": "duce memory as well as aggregate the temporal context for TAD. Different variants of these adapters are designed to trade off between performance and memory cost. To the best of our knowledge, we are the first to introduce the adapter mechanism to TAD." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 133, + 545, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 133, + 545, + 180 + ], + "spans": [ + { + "bbox": [ + 306, + 133, + 545, + 180 + ], + "type": "text", + "content": "3. Our method achieves state-of-the-art performance across four TAD datasets. Remarkably, this represents the first end-to-end approach that outperforms the previous feature-based methods by a large margin." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 189, + 392, + 201 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 189, + 392, + 201 + ], + "spans": [ + { + "bbox": [ + 306, + 189, + 392, + 201 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 209, + 545, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 209, + 545, + 399 + ], + "spans": [ + { + "bbox": [ + 304, + 209, + 545, + 399 + ], + "type": "text", + "content": "Temporal Action Detection. Current methods for temporal action detection, also referred to as temporal action localization, can be broadly classified into three categories based on their architectural design: one-stage, two-stage, and DETR-based methods. One-stage methods directly localize actions from a multi-scale feature pyramid, such as ActionFormer [66] and TriDet [47]. These methods integrate action classification and temporal boundary regression in a single step [45, 61, 62]. Two-stage methods, in contrast, involve an additional step of proposal feature extraction [4, 30, 40, 58, 59, 68, 73]. For instance, VSGN [67] employs boundary sampling to further refine proposals. Recently, there is a growing interest in query-based methods [34, 46, 51], which deploy a set of learned queries to interact with the feature maps and directly predict the actions' temporal boundaries and categories." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 400, + 545, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 400, + 545, + 628 + ], + "spans": [ + { + "bbox": [ + 304, + 400, + 545, + 628 + ], + "type": "text", + "content": "In addition to the aforementioned categories, TAD can also be divided into feature-based and end-to-end methods. The former relies on pre-extracted RGB features and optionally incorporates optical flow features. On the other hand, end-to-end methods take raw video frames as input and jointly optimize the video encoder and action detector [31]. Due to computational constraints, AFSD [29] downsamples the input to a resolution of " + }, + { + "bbox": [ + 304, + 400, + 545, + 628 + ], + "type": "inline_equation", + "content": "96^2" + }, + { + "bbox": [ + 304, + 400, + 545, + 628 + ], + "type": "text", + "content": ". DaoTAD [55] and E2E-TAD [33] provide evidence that high TAD performance can be achieved by relying solely on the RGB modality with various data augmentations. Further innovations came from SGP [11], TALLFormer [10], and ETAD [32], all of which introduced strategies to backpropagate only through parts of the data. Additionally, " + }, + { + "bbox": [ + 304, + 400, + 545, + 628 + ], + "type": "inline_equation", + "content": "\\mathrm{Re}^2" + }, + { + "bbox": [ + 304, + 400, + 545, + 628 + ], + "type": "text", + "content": "TAL [69] and " + }, + { + "bbox": [ + 304, + 400, + 545, + 628 + ], + "type": "inline_equation", + "content": "\\mathrm{Dr}^2" + }, + { + "bbox": [ + 304, + 400, + 545, + 628 + ], + "type": "text", + "content": "Net[70] design reversible network architectures to release the memory occupied by intermediate activations. Despite these advancements, all above methods follow the full fine-tuning paradigm, and none has yet surpassed the best results achieved by feature-based approaches." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 629, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 629, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 629, + 545, + 714 + ], + "type": "text", + "content": "Scaling Law in Deep Learning. Scaling up model and data has been a prevalent strategy across both computer vision and natural language processing fields to achieve superior performance. The GPT series [5, 39, 41, 42] has consistently demonstrated that larger models pretrained on extensive datasets yield significant gains in language understanding capabilities. Analogously, in the realms of image" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "18592" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 72, + 171, + 225 + ], + "blocks": [ + { + "bbox": [ + 59, + 72, + 171, + 225 + ], + "lines": [ + { + "bbox": [ + 59, + 72, + 171, + 225 + ], + "spans": [ + { + "bbox": [ + 59, + 72, + 171, + 225 + ], + "type": "image", + "image_path": "e48ed53730e24a77b65a5de34100c577145e91640c4268b5d41ae7d6013d1417.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 232, + 547, + 277 + ], + "lines": [ + { + "bbox": [ + 46, + 232, + 547, + 277 + ], + "spans": [ + { + "bbox": [ + 46, + 232, + 547, + 277 + ], + "type": "text", + "content": "Figure 2. Comparative illustration of our proposed TAD framework versus popular and widely used alternatives. (a) represents the typical offline method. (b) is the traditional end-to-end method using full fine-tuning. (c) Tailored for the TAD task, our AdaTAD uses a lightweight temporal-informative adapter inside the backbone to achieve efficient transfer learning. (d) To further reduce memory usage and scale up the model/data, " + }, + { + "bbox": [ + 46, + 232, + 547, + 277 + ], + "type": "inline_equation", + "content": "\\mathrm{AdaTAD}^{\\dagger}" + }, + { + "bbox": [ + 46, + 232, + 547, + 277 + ], + "type": "text", + "content": " uses an alternative placement for adapters outside the backbone." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 199, + 72, + 275, + 227 + ], + "blocks": [ + { + "bbox": [ + 199, + 72, + 275, + 227 + ], + "lines": [ + { + "bbox": [ + 199, + 72, + 275, + 227 + ], + "spans": [ + { + "bbox": [ + 199, + 72, + 275, + 227 + ], + "type": "image", + "image_path": "ed78bdde4dc9859c9e785005426cb4ab21e0450f9b2c312bab69f72b787c6052.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 298, + 72, + 373, + 226 + ], + "blocks": [ + { + "bbox": [ + 298, + 72, + 373, + 226 + ], + "lines": [ + { + "bbox": [ + 298, + 72, + 373, + 226 + ], + "spans": [ + { + "bbox": [ + 298, + 72, + 373, + 226 + ], + "type": "image", + "image_path": "8707905ec3b0e326193625e371ea4aef9d949529e19d6255d017968fc020bd4e.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 395, + 72, + 532, + 226 + ], + "blocks": [ + { + "bbox": [ + 395, + 72, + 532, + 226 + ], + "lines": [ + { + "bbox": [ + 395, + 72, + 532, + 226 + ], + "spans": [ + { + "bbox": [ + 395, + 72, + 532, + 226 + ], + "type": "image", + "image_path": "f000f40ce7207fe11019c00cc44ec41eccf5abf7fdd97ad02acfd15a9593980e.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 289, + 289, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 289, + 289, + 396 + ], + "spans": [ + { + "bbox": [ + 46, + 289, + 289, + 396 + ], + "type": "text", + "content": "and video understanding, architectures such as ViT [14] and MViT [16], have also witnessed the effectiveness of this scaling strategy. Alabdulmohsin et al.[1] present a recipe for estimating scaling law parameters reliably from learning curves in computer vision. To attain even greater performance, several studies have also scaled up image resolution [35] or video clip length [56]. In this paper, we successfully apply this principle within the domain of temporal action detection and achieve state-of-the-art results." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 399, + 287, + 532 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 399, + 287, + 532 + ], + "spans": [ + { + "bbox": [ + 46, + 399, + 287, + 532 + ], + "type": "text", + "content": "Efficient Transfer Learning. Transfer learning aims to adapt a pretrained model to a new domain. In TAD, typically off-the-shelf action recognition models are employed as the backbone, such as SlowFast [17]. Traditional transfer learning adopts full fine-tuning, meaning that all parameters of the pretrained model are updated. However, studies by [43, 63] have noted that full fine-tuning may harm the pre-learned knowledge, particularly when the downstream dataset is small and less comprehensive. Moreover, as models increase in size, the computational and storage demands of full fine-tuning proportionally increase." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 533, + 287, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 533, + 287, + 687 + ], + "spans": [ + { + "bbox": [ + 46, + 533, + 287, + 687 + ], + "type": "text", + "content": "In response to these challenges, several works have investigated parameter-efficient tuning (PEFT) strategies that involve fine-tuning only a fraction of the network. For instance, Adapter [22] inserts lightweight modules analogous to feedforward networks in transformers, and only tunes these elements. LoRA [23] employs low-rank matrices in each transformer layer. Prefix-tuning [28] and prompt-tuning [27] attach learnable prompt tokens at the input stage or within each layer. In computer vision, many PEFT methods [6, 7, 24, 63] have also been explored across various tasks to optimize transfer learning efficiency. Our work represents the first effort to examine the potential of the PEFT mechanism in TAD." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "type": "text", + "content": "Although PEFT effectively reduces the number of learnable parameters, data-intensive and computationally heavy" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 289, + 545, + 410 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 289, + 545, + 410 + ], + "spans": [ + { + "bbox": [ + 304, + 289, + 545, + 410 + ], + "type": "text", + "content": "tasks like video understanding require more memory-efficient techniques. To this end, several works try to externalize the trainable components from the backbone, eliminating the need for backpropagation through the extensive original model. For example, LST [49] introduces a supplementary lightweight network that operates in parallel with the main model. Similarly, " + }, + { + "bbox": [ + 304, + 289, + 545, + 410 + ], + "type": "inline_equation", + "content": "\\mathrm{E}^3\\mathrm{VA}" + }, + { + "bbox": [ + 304, + 289, + 545, + 410 + ], + "type": "text", + "content": " [65] leverages intermediate features with adapters to enable efficient transfer learning while minimizing memory usage. Our work is inspired by these methods yet with a streamlined and simple design." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 420, + 389, + 434 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 420, + 389, + 434 + ], + "spans": [ + { + "bbox": [ + 306, + 420, + 389, + 434 + ], + "type": "text", + "content": "3. Methodology" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 441, + 545, + 513 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 441, + 545, + 513 + ], + "spans": [ + { + "bbox": [ + 304, + 441, + 545, + 513 + ], + "type": "text", + "content": "In this section, we introduce our AdaTAD step-by-step. We first introduce notations and study the efficient video representation to establish an end-to-end TAD baseline. Next, we introduce a temporal-informative adapter designed for efficient TAD. Finally, we propose an alternative placement for adapters to further alleviate computational demands." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 521, + 373, + 533 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 521, + 373, + 533 + ], + "spans": [ + { + "bbox": [ + 306, + 521, + 373, + 533 + ], + "type": "text", + "content": "3.1. Notations" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 540, + 545, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 540, + 545, + 651 + ], + "spans": [ + { + "bbox": [ + 304, + 540, + 545, + 651 + ], + "type": "text", + "content": "Temporal action detection can be formulated as follows: given an untrimmed video " + }, + { + "bbox": [ + 304, + 540, + 545, + 651 + ], + "type": "inline_equation", + "content": "\\mathbf{X} \\in \\mathbb{R}^{3 \\times H \\times W \\times T}" + }, + { + "bbox": [ + 304, + 540, + 545, + 651 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 540, + 545, + 651 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 304, + 540, + 545, + 651 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 540, + 545, + 651 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 304, + 540, + 545, + 651 + ], + "type": "text", + "content": " are the height and width of each frame, and " + }, + { + "bbox": [ + 304, + 540, + 545, + 651 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 304, + 540, + 545, + 651 + ], + "type": "text", + "content": " is the frame number, its temporal action annotations can be denoted as " + }, + { + "bbox": [ + 304, + 540, + 545, + 651 + ], + "type": "inline_equation", + "content": "\\Psi_{g} = \\{\\varphi_{i} = (t_{s}, t_{e}, c)\\}_{i=1}^{N}" + }, + { + "bbox": [ + 304, + 540, + 545, + 651 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 540, + 545, + 651 + ], + "type": "inline_equation", + "content": "t_{s}, t_{e}, c" + }, + { + "bbox": [ + 304, + 540, + 545, + 651 + ], + "type": "text", + "content": " are the start, end time and category of action instance " + }, + { + "bbox": [ + 304, + 540, + 545, + 651 + ], + "type": "inline_equation", + "content": "\\varphi_{i}" + }, + { + "bbox": [ + 304, + 540, + 545, + 651 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 540, + 545, + 651 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 540, + 545, + 651 + ], + "type": "text", + "content": " is the number of ground truth actions. TAD aims to predict candidate proposal set " + }, + { + "bbox": [ + 304, + 540, + 545, + 651 + ], + "type": "inline_equation", + "content": "\\Psi_{p} = \\{\\hat{\\varphi}_{i} = (\\hat{t}_{s}, \\hat{t}_{e}, \\hat{c}, s)\\}_{i=1}^{M}" + }, + { + "bbox": [ + 304, + 540, + 545, + 651 + ], + "type": "text", + "content": " to cover " + }, + { + "bbox": [ + 304, + 540, + 545, + 651 + ], + "type": "inline_equation", + "content": "\\Psi_{g}" + }, + { + "bbox": [ + 304, + 540, + 545, + 651 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 540, + 545, + 651 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 304, + 540, + 545, + 651 + ], + "type": "text", + "content": " is the confidence score." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 658, + 455, + 671 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 658, + 455, + 671 + ], + "spans": [ + { + "bbox": [ + 306, + 658, + 455, + 671 + ], + "type": "text", + "content": "3.2. Frame-level representation" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 677, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 545, + 714 + ], + "type": "text", + "content": "Our end-to-end TAD architecture comprises two main components: feature extraction and action detection. Following previous work [69], we select ActionFormer [66] as our ac" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18593" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 131 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 131 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 131 + ], + "type": "text", + "content": "tion detection head due to its robust performance across various datasets without much hyperparameter tuning. Next, we discuss two ways of encoding raw frames into representative features (feature extraction): snippet representation and frame representation." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 133, + 286, + 263 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 133, + 286, + 263 + ], + "spans": [ + { + "bbox": [ + 46, + 133, + 286, + 263 + ], + "type": "text", + "content": "Snippets Representation. Snippet-based video representations are popular choices in offline feature extraction. The whole video is divided into several short snippets (or namely clips). Each snippet has a short temporal length, e.g., 16 frames, and different snippets can have overlapping frames. Thus, the video can be conceptualized as " + }, + { + "bbox": [ + 46, + 133, + 286, + 263 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 46, + 133, + 286, + 263 + ], + "type": "text", + "content": " snippets, denoted by " + }, + { + "bbox": [ + 46, + 133, + 286, + 263 + ], + "type": "inline_equation", + "content": "\\mathbf{X} \\in \\mathbb{R}^{T \\times 3 \\times 16 \\times H \\times W}" + }, + { + "bbox": [ + 46, + 133, + 286, + 263 + ], + "type": "text", + "content": ". Each snippet is processed through the video backbone, and spatial-temporal pooling is applied to extract one snippet feature. This processing yields the feature representation " + }, + { + "bbox": [ + 46, + 133, + 286, + 263 + ], + "type": "inline_equation", + "content": "\\mathbf{F} \\in \\mathbb{R}^{T \\times C}" + }, + { + "bbox": [ + 46, + 133, + 286, + 263 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 133, + 286, + 263 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 46, + 133, + 286, + 263 + ], + "type": "text", + "content": " denotes the channel dimension of the pooled features." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 264, + 286, + 359 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 264, + 286, + 359 + ], + "spans": [ + { + "bbox": [ + 46, + 264, + 286, + 359 + ], + "type": "text", + "content": "Frame Representation. In contrast to snippet-based representations, frame-based video representations consider the entire video as a singular snippet or clip, represented as " + }, + { + "bbox": [ + 46, + 264, + 286, + 359 + ], + "type": "inline_equation", + "content": "\\mathbf{X} \\in \\mathbb{R}^{1 \\times 3 \\times T \\times H \\times W}" + }, + { + "bbox": [ + 46, + 264, + 286, + 359 + ], + "type": "text", + "content": ". Then, the whole frame sequence is fed into the video backbone, and only spatial pooling is employed to extract features [10, 29, 69]. For attention-based models such as VideoMAE [54], the video is chunked into multiple shorter clips to avoid extensive temporal attention." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 360, + 286, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 360, + 286, + 502 + ], + "spans": [ + { + "bbox": [ + 46, + 360, + 286, + 502 + ], + "type": "text", + "content": "Although both representations have been employed in previous studies, a fair comparison between them has not yet been performed. To address this gap, we conduct a comparative analysis of the two representations under the same setting on THUMOS14, measuring their memory usage and detection mAP. The results in Table 1 indicate that frame representation has comparable or even better performance than snippet representation, yet with much smaller memory consumption. When the feature extraction backbones are frozen, frame representation yields superior results to snippet representation for both VideoMAE [53] and SlowFast [17] backbones. Only in the end-to-end set" + } + ] + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 48, + 584, + 291, + 710 + ], + "blocks": [ + { + "bbox": [ + 46, + 521, + 286, + 577 + ], + "lines": [ + { + "bbox": [ + 46, + 521, + 286, + 577 + ], + "spans": [ + { + "bbox": [ + 46, + 521, + 286, + 577 + ], + "type": "text", + "content": "Table 1. Snippets representation vs frame representation. We use the end-to-end version of ActionFormer with two representations for comparison. The snippet input is " + }, + { + "bbox": [ + 46, + 521, + 286, + 577 + ], + "type": "inline_equation", + "content": "768 \\times 3 \\times 16 \\times 160 \\times 160" + }, + { + "bbox": [ + 46, + 521, + 286, + 577 + ], + "type": "text", + "content": ", and the frame input is " + }, + { + "bbox": [ + 46, + 521, + 286, + 577 + ], + "type": "inline_equation", + "content": "1 \\times 3 \\times 768 \\times 160 \\times 160" + }, + { + "bbox": [ + 46, + 521, + 286, + 577 + ], + "type": "text", + "content": ". * means activation checkpointing is utilized to avoid overflowing GPU memory." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 48, + 584, + 291, + 710 + ], + "lines": [ + { + "bbox": [ + 48, + 584, + 291, + 710 + ], + "spans": [ + { + "bbox": [ + 48, + 584, + 291, + 710 + ], + "type": "table", + "html": "
SettingBackboneRepr.Avg. mAPMem (GB)
FrozenVideoMAE-SFrame59.351.9
Snippet57.6813.2
SlowFast-R101Frame61.343.6
Snippet60.2417.2
End to EndVideoMAE-SFrame67.152.8*
Snippet68.4624.6*
SlowFast-R101Frame65.335.5*
Snippet66.7251.6*
", + "image_path": "0c22a7f450caeccceb112ee96bd7f7d7551c54481c0507259bec47c03251b792.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 72, + 545, + 132 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 132 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 132 + ], + "type": "text", + "content": "ting can the snippet representation achieve " + }, + { + "bbox": [ + 304, + 72, + 545, + 132 + ], + "type": "inline_equation", + "content": "1\\%" + }, + { + "bbox": [ + 304, + 72, + 545, + 132 + ], + "type": "text", + "content": " mAP advantage over frame representations; however, it requires 8 times more memory consumption. Taking into account both performance and memory usage, frame-based representations could be a better choice for end-to-end TAD development." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 133, + 545, + 192 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 133, + 545, + 192 + ], + "spans": [ + { + "bbox": [ + 304, + 133, + 545, + 192 + ], + "type": "text", + "content": "Therefore, we use frame representation as the default baseline to encode videos in our experiments. Following the previous TALLFormer work [10], we also incorporate activation checkpointing [8] and mixed precision training [37] to fully harness the potential of scaling." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 201, + 474, + 213 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 201, + 474, + 213 + ], + "spans": [ + { + "bbox": [ + 305, + 201, + 474, + 213 + ], + "type": "text", + "content": "3.3. Temporal-Informative Adapter" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 220, + 545, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 220, + 545, + 422 + ], + "spans": [ + { + "bbox": [ + 304, + 220, + 545, + 422 + ], + "type": "text", + "content": "In Section 3.2, we have built a simple end-to-end baseline using full fine-tuning. However, the baseline still suffers from two aspects: 1. Increased computational cost. In Table 1, we only use small video backbones like VideoMAES. When scaling VideoMAE-S to larger models, the computational burden and memory cost will grow rapidly. 2. Inferior transfer learning ability. More critically, the baseline follows the tradition of full fine-tuning, which may lead to inferior transfer learning. Pointed out by [50, 63], full fine-tuning may result in overfitting or forgetting, especially for large pretrained models. If downstream datasets are not sufficiently diverse, full fine-tuning can even destroy the powerful features learned from large-scale pretraining. Motivated by the above two aspects, we apply the PEFT mechanism and propose to fine-tune a plug-and-play module named Temporal-Informative Adapter (TIA) to achieve efficient and effective transfer learning for TAD." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 424, + 545, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 424, + 545, + 559 + ], + "spans": [ + { + "bbox": [ + 304, + 424, + 545, + 559 + ], + "type": "text", + "content": "We first review the architecture of the standard adapter proposed by [22]. As formulated in Equation 1, the standard adapter includes a down-projection fully connected (FC) layer with parameter " + }, + { + "bbox": [ + 304, + 424, + 545, + 559 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_{\\mathrm{down}} \\in \\mathbb{R}^{d \\times \\frac{d}{\\gamma}}" + }, + { + "bbox": [ + 304, + 424, + 545, + 559 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 424, + 545, + 559 + ], + "type": "inline_equation", + "content": "\\frac{d}{\\gamma}" + }, + { + "bbox": [ + 304, + 424, + 545, + 559 + ], + "type": "text", + "content": " represents the intermediate dimension and satisfies " + }, + { + "bbox": [ + 304, + 424, + 545, + 559 + ], + "type": "inline_equation", + "content": "\\gamma > 1" + }, + { + "bbox": [ + 304, + 424, + 545, + 559 + ], + "type": "text", + "content": ". Then, an up-projection layer " + }, + { + "bbox": [ + 304, + 424, + 545, + 559 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_{\\mathrm{up}} \\in \\mathbb{R}^{\\frac{d}{\\gamma} \\times d}" + }, + { + "bbox": [ + 304, + 424, + 545, + 559 + ], + "type": "text", + "content": " is employed to restore the channel dimension. Between these two FC layers, a non-linear activation function " + }, + { + "bbox": [ + 304, + 424, + 545, + 559 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 304, + 424, + 545, + 559 + ], + "type": "text", + "content": " is inserted, such as GELU [21]. Afterward, a residual connection is added to the output of the projection layer. Note that " + }, + { + "bbox": [ + 304, + 424, + 545, + 559 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 304, + 424, + 545, + 559 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 424, + 545, + 559 + ], + "type": "inline_equation", + "content": "\\mathbf{x}'" + }, + { + "bbox": [ + 304, + 424, + 545, + 559 + ], + "type": "text", + "content": " are the input and output features with the same shape " + }, + { + "bbox": [ + 304, + 424, + 545, + 559 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^{d \\times t \\times h \\times w}" + }, + { + "bbox": [ + 304, + 424, + 545, + 559 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 358, + 567, + 545, + 583 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 358, + 567, + 545, + 583 + ], + "spans": [ + { + "bbox": [ + 358, + 567, + 545, + 583 + ], + "type": "interline_equation", + "content": "\\boldsymbol {x} ^ {\\prime} = \\boldsymbol {W} _ {\\mathrm {u p}} ^ {\\top} \\cdot \\boldsymbol {\\sigma} \\left(\\boldsymbol {W} _ {\\mathrm {d o w n}} ^ {\\top} \\cdot \\boldsymbol {x}\\right) + \\boldsymbol {x}. \\tag {1}", + "image_path": "8bdbae92196b5aed19a2609181b2f826263d61f36ffb6830be9304954b7512ae.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 594, + 545, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 545, + 665 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 545, + 665 + ], + "type": "text", + "content": "Although the adapter has achieved great success in natural language processing and computer vision, the standard adapter only focuses on adapting channel information, which neglects the temporal context vital for the TAD task. To address this limitation, we introduce the temporalinformative adapter, as depicted in Fig. 3(b)." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 665, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 545, + 712 + ], + "type": "text", + "content": "The architecture of TIA follows the general bottleneck design of the standard adapter, while we integrate the temporal depth-wise convolution layers, as described in Equation 2. The temporal convolution with a kernel size of " + }, + { + "bbox": [ + 304, + 665, + 545, + 712 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 665, + 545, + 712 + ], + "type": "text", + "content": " is" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18594" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 64, + 70, + 147, + 204 + ], + "blocks": [ + { + "bbox": [ + 64, + 70, + 147, + 204 + ], + "lines": [ + { + "bbox": [ + 64, + 70, + 147, + 204 + ], + "spans": [ + { + "bbox": [ + 64, + 70, + 147, + 204 + ], + "type": "image", + "image_path": "bca175eac0c4d2d439c5b27d1a329dee4d29c7558b725de0e20f59adf28392de.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 67, + 205, + 128, + 213 + ], + "lines": [ + { + "bbox": [ + 67, + 205, + 128, + 213 + ], + "spans": [ + { + "bbox": [ + 67, + 205, + 128, + 213 + ], + "type": "text", + "content": "(a) Standard Adapter" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 217, + 287, + 250 + ], + "lines": [ + { + "bbox": [ + 46, + 217, + 287, + 250 + ], + "spans": [ + { + "bbox": [ + 46, + 217, + 287, + 250 + ], + "type": "text", + "content": "Figure 3. Architecture of (a) standard adapter and (b) our temporal-informative adapter. We incorporate temporal depth-wise convolution to aggregate context from adjacent frames." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 172, + 70, + 269, + 204 + ], + "blocks": [ + { + "bbox": [ + 172, + 70, + 269, + 204 + ], + "lines": [ + { + "bbox": [ + 172, + 70, + 269, + 204 + ], + "spans": [ + { + "bbox": [ + 172, + 70, + 269, + 204 + ], + "type": "image", + "image_path": "5b95fa8e623d1dee88ffdb06ef8f5387be369eeb3524eeb65e87242fa5349dac.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 173, + 205, + 272, + 213 + ], + "lines": [ + { + "bbox": [ + 173, + 205, + 272, + 213 + ], + "spans": [ + { + "bbox": [ + 173, + 205, + 272, + 213 + ], + "type": "text", + "content": "(b) Temporal-Informative Adapter" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 270, + 289, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 270, + 289, + 373 + ], + "spans": [ + { + "bbox": [ + 46, + 270, + 289, + 373 + ], + "type": "text", + "content": "designed to aggregate local informative context from adjacent frames and to enrich the representation of the current time step. Practically, this is achieved through the application of a 3D convolution with kernel size " + }, + { + "bbox": [ + 46, + 270, + 289, + 373 + ], + "type": "inline_equation", + "content": "(k,1,1)" + }, + { + "bbox": [ + 46, + 270, + 289, + 373 + ], + "type": "text", + "content": " and group size " + }, + { + "bbox": [ + 46, + 270, + 289, + 373 + ], + "type": "inline_equation", + "content": "\\frac{d}{\\gamma}" + }, + { + "bbox": [ + 46, + 270, + 289, + 373 + ], + "type": "text", + "content": " for depth-wise processing. Additionally, an FC layer with weight " + }, + { + "bbox": [ + 46, + 270, + 289, + 373 + ], + "type": "inline_equation", + "content": "W_{\\mathrm{mid}}\\in \\mathbb{R}^{\\frac{d}{\\gamma}\\times \\frac{d}{\\gamma}}" + }, + { + "bbox": [ + 46, + 270, + 289, + 373 + ], + "type": "text", + "content": " is employed to facilitate information exchange across channels. At last, a learnable scalar " + }, + { + "bbox": [ + 46, + 270, + 289, + 373 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 46, + 270, + 289, + 373 + ], + "type": "text", + "content": " is introduced to adjust the amplitude of adapter's output." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 100, + 380, + 186, + 394 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 100, + 380, + 186, + 394 + ], + "spans": [ + { + "bbox": [ + 100, + 380, + 186, + 394 + ], + "type": "interline_equation", + "content": "\\bar {\\boldsymbol {x}} = \\boldsymbol {\\sigma} (\\boldsymbol {W} _ {\\text {d o w n}} ^ {\\top} \\cdot \\boldsymbol {x}),", + "image_path": "93d0e07ed65c79fe9650d4f176ddddf960834128fde7b19a03a5f775a6782988.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 102, + 396, + 287, + 411 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 102, + 396, + 287, + 411 + ], + "spans": [ + { + "bbox": [ + 102, + 396, + 287, + 411 + ], + "type": "interline_equation", + "content": "\\hat {\\boldsymbol {x}} = \\boldsymbol {W} _ {\\mathrm {m i d}} ^ {\\top} \\cdot \\mathbf {D W C o n v} _ {k} (\\bar {\\boldsymbol {x}}) + \\bar {\\boldsymbol {x}}, \\tag {2}", + "image_path": "cc2d413d5a5284fcb2472e4015cb2d82f0d01a54699a98dd1257d7654f7555be.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 100, + 413, + 193, + 428 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 100, + 413, + 193, + 428 + ], + "spans": [ + { + "bbox": [ + 100, + 413, + 193, + 428 + ], + "type": "interline_equation", + "content": "\\boldsymbol {x} ^ {\\prime} = \\alpha \\cdot \\boldsymbol {W} _ {\\mathrm {u p}} ^ {\\top} \\cdot \\hat {\\boldsymbol {x}} + \\boldsymbol {x}.", + "image_path": "339cf3455d4b7c944440b525e51923a6b53d7ada48d5feaeea5aa15d4fec6cba.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 436, + 287, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 436, + 287, + 628 + ], + "spans": [ + { + "bbox": [ + 46, + 436, + 287, + 628 + ], + "type": "text", + "content": "As shown in Fig. 2(c), TIA is designed to be inserted between different backbone layers, e.g. between each ViT block of VideoMAE or each bottleneck block of SlowFast. To ensure the newly added connection does not affect the original network at the beginning of transfer learning, the weight and bias of the adapter's last projection layer " + }, + { + "bbox": [ + 46, + 436, + 287, + 628 + ], + "type": "inline_equation", + "content": "W_{\\mathrm{up}}" + }, + { + "bbox": [ + 46, + 436, + 287, + 628 + ], + "type": "text", + "content": " are initialized to 0. The learnable coefficient " + }, + { + "bbox": [ + 46, + 436, + 287, + 628 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 46, + 436, + 287, + 628 + ], + "type": "text", + "content": " is initialized to 1. The temporal kernel size " + }, + { + "bbox": [ + 46, + 436, + 287, + 628 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 436, + 287, + 628 + ], + "type": "text", + "content": " is set to 3, and the channel downsampling ratio " + }, + { + "bbox": [ + 46, + 436, + 287, + 628 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 46, + 436, + 287, + 628 + ], + "type": "text", + "content": " is set to 4 by default. Under these settings, the additional trainable parameters coming from TIA only account for " + }, + { + "bbox": [ + 46, + 436, + 287, + 628 + ], + "type": "inline_equation", + "content": "4.7\\%" + }, + { + "bbox": [ + 46, + 436, + 287, + 628 + ], + "type": "text", + "content": " of the total parameters of the original backbone. Since this backbone is frozen when TIA is used, our proposed strategy constitutes a massive reduction in trainable parameters as compared to full fine-tuning. Our experiments show that TIA can achieve better performance than full fine-tuning with less memory usage." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 635, + 232, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 635, + 232, + 647 + ], + "spans": [ + { + "bbox": [ + 47, + 635, + 232, + 647 + ], + "type": "text", + "content": "3.4. Alternative Placement for Adapter" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 653, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 287, + 715 + ], + "type": "text", + "content": "Although the previously described PEFT approach can reduce tunable parameters and memory usage, the gradient still needs to backpropagate over the entire backbone during training. This requirement limits our ability to scale-up the model size or input data size further. As highlighted in prior" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 72, + 545, + 107 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 107 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 107 + ], + "type": "text", + "content": "works [49, 65], if we can stop the gradient backpropagation within the original backbone, additional computational savings can be achieved." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 108, + 546, + 252 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 108, + 546, + 252 + ], + "spans": [ + { + "bbox": [ + 304, + 108, + 546, + 252 + ], + "type": "text", + "content": "Inspired by this insight, we propose a placement strategy for adapters that position them externally to the backbone, rather than inserting them inside. As illustrated in Fig. 2(d), we utilize the previously introduced TIA module, but its output does not feed back into the middle of the original backbone. It is directly added to the backbone's final layer. This configuration eliminates the need for backpropagation through the original network, as gradients are only tracked to the shallow lightweight adapter. To further diminish computation, we observe that adapting only the last half of backbone layers yields comparable performance while reducing half of the adaptation cost." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 252, + 545, + 348 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 252, + 545, + 348 + ], + "spans": [ + { + "bbox": [ + 304, + 252, + 545, + 348 + ], + "type": "text", + "content": "To distinguish the different variants, we name the standard adaption design as AdaTAD, and the alternative placement as AdaTAD†. The latter can be considered as a lite version of the former. Compared to directly injecting adapters into the backbone, AdaTAD† may lead to a slight performance drop. However, it enables us to leverage richer models and more data, which should effectively counter this possible drop." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 360, + 388, + 373 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 360, + 388, + 373 + ], + "spans": [ + { + "bbox": [ + 306, + 360, + 388, + 373 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 380, + 427, + 392 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 380, + 427, + 392 + ], + "spans": [ + { + "bbox": [ + 306, + 380, + 427, + 392 + ], + "type": "text", + "content": "4.1. Datasets and Metrics" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 398, + 545, + 530 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 398, + 545, + 530 + ], + "spans": [ + { + "bbox": [ + 304, + 398, + 545, + 530 + ], + "type": "text", + "content": "We choose ActivityNet-1.3 [20], THUMOS14 [25], and Epic-Kitchens 100 [13] to evaluate our proposed approach. ActivityNet-1.3 and THUMOS14 are web-collected third-person untrimmed videos, consisting of 19,994 and 413 videos, respectively. EPIC-Kitchens 100 is collected from 700 egocentric videos. Since the action categories of EPIC-Kitchens 100 are more domain-specific and different from common pretraining data, achieving higher performance on this dataset is more challenging. Moreover, we also evaluate our method on the Ego4D-MQ [18] dataset, and the results can be found in the appendix." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 531, + 545, + 615 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 531, + 545, + 615 + ], + "spans": [ + { + "bbox": [ + 304, + 531, + 545, + 615 + ], + "type": "text", + "content": "Following common practice, we report the mean Average Precision (mAP) at certain IoU thresholds and average mAP as the evaluation metrics. On ActivityNet1.3, the IoU thresholds are chosen from 0.5 to 0.95 with 10 steps. On THUMOS14, the threshold is chosen from " + }, + { + "bbox": [ + 304, + 531, + 545, + 615 + ], + "type": "inline_equation", + "content": "\\{0.3,0.4,0.5,0.6,0.7\\}" + }, + { + "bbox": [ + 304, + 531, + 545, + 615 + ], + "type": "text", + "content": ". On EPIC-Kitchens 100, the threshold is set to " + }, + { + "bbox": [ + 304, + 531, + 545, + 615 + ], + "type": "inline_equation", + "content": "\\{0.1,0.2,0.3,0.4,0.5\\}" + }, + { + "bbox": [ + 304, + 531, + 545, + 615 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 306, + 623, + 439, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 623, + 439, + 635 + ], + "spans": [ + { + "bbox": [ + 306, + 623, + 439, + 635 + ], + "type": "text", + "content": "4.2. Implementation Details" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "type": "text", + "content": "We implement our method with PyTorch 2.0 and MMAc-tion2 [12] with 4 A100 GPUs. By default, mixed-precision training and activation checkpointing are adopted to save memory. We use ActionFormer [66] as our detection head, and keep the hyper-parameters unchanged on each dataset. The learning rate for the adapter in the backbone is grid" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18595" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 47, + 130, + 548, + 395 + ], + "blocks": [ + { + "bbox": [ + 46, + 70, + 547, + 127 + ], + "lines": [ + { + "bbox": [ + 46, + 70, + 547, + 127 + ], + "spans": [ + { + "bbox": [ + 46, + 70, + 547, + 127 + ], + "type": "text", + "content": "Table 2. Results on ActivityNet-1.3 and THUMOS14, measured by mAP (\\%) at different tIoU thresholds. E2E refers to end-to-end training, and Mem refers to memory usage (GB) per video. On ActivityNet-1.3, our prediction is combined with CUHK [72] classification results. Specifically, * means we employ stronger video-level classification results used in InternVideo [57] for a fair comparison. We report our best results in bold, and the previous best results in underline, which was achieved by the feature-based method. The last row is achieved when only the last half of backbone layers are adapted; otherwise, full-layer adaptation will lead to out-of-memory on A100-80G." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 47, + 130, + 548, + 395 + ], + "lines": [ + { + "bbox": [ + 47, + 130, + 548, + 395 + ], + "spans": [ + { + "bbox": [ + 47, + 130, + 548, + 395 + ], + "type": "table", + "html": "
MethodBackboneE2EFlowMemActivityNet-1.3THUMOS14
0.50.750.95Avg.0.30.40.50.60.7Avg.
BMN [30]TSNX✓-50.0734.788.2933.8556.047.438.829.720.538.5
TadTR [34]I3DX✓-49.1032.608.5032.3062.457.449.237.826.346.6
ActionFormer [66]SlowFast-R50XX-54.2637.048.1336.0278.773.365.254.639.762.3
ActionFormer [66]I3DX✓-53.5036.208.2035.6082.177.871.059.443.966.8
ASL [45]I3DX✓-54.1067.408.0036.2083.179.071.759.745.867.9
TriDet [47]I3DX✓-54.7038.008.4036.8083.680.172.962.447.469.3
VideoMAEv2 [56]VideoMAEv2-gXX----------69.6
InternVideo [57]VideoMAE-H+UniformerV2XX----39.00*-----71.5
AFSD [29]I3D✓✓1252.4035.306.5034.4067.362.455.543.731.152.0
E2E-TAD [33]SlowFast-R50✓X1250.4735.9910.3335.1069.464.356.046.434.954.2
BasicTAD [62]SlowOnly-R50✓X1251.2033.417.5733.1275.570.863.550.937.459.6
TALLFormer [10]VideoSwin-B✓X2954.1036.207.9035.6076.0-63.2-34.559.2
Re2TAL [69]Re2VideoSwin-T✓X2454.7537.819.0336.8077.071.562.449.736.359.4
AdaTADSlowFast-R50✓X4.355.2838.118.8737.1181.076.269.459.044.566.0
AdaTADVideoMAE-S✓X2.556.1538.999.0737.8584.580.271.660.946.968.8
AdaTADVideoMAE-B✓X4.956.7739.359.7138.3987.082.475.363.849.271.5
AdaTADVideoMAE-L✓X11.057.6940.5610.1339.2287.784.176.766.452.473.5
AdaTADVideoMAE-H✓X19.258.0440.559.7539.3788.985.378.666.952.574.4
AdaTADVideoMAEv2-g✓X29.958.4541.1610.4539.7989.585.878.967.352.674.8
AdaTAD† (1536×2242)VideoMAEv2-g✓X43.660.8242.699.8441.15*89.685.979.467.653.875.4
AdaTAD (1536×2242)VideoMAEv2-g✓X50.661.7243.3510.8541.93*89.786.780.971.056.176.9
", + "image_path": "4b56a1ded48a8dc87f26ec229e00c7e873d97dfc78dec537c7394e88c374aac6.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 407, + 288, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 407, + 288, + 540 + ], + "spans": [ + { + "bbox": [ + 46, + 407, + 288, + 540 + ], + "type": "text", + "content": "searched from 5e-4 to 5e-5, and other parameters inside the backbone are frozen. On ActivityNet-1.3, we resize the video into a fixed length of 768 frames. On THUMOS14, we randomly truncate a window with 768 frames with a temporal stride of 4. On EPIC-Kitchens 100, we randomly truncate a window with 6144 frames with a temporal stride of 2. After the video encoder, the feature is resized to fixed lengths of 192, 768, and 768, respectively, for the three datasets. Frame resolution is set to " + }, + { + "bbox": [ + 46, + 407, + 288, + 540 + ], + "type": "inline_equation", + "content": "160^2" + }, + { + "bbox": [ + 46, + 407, + 288, + 540 + ], + "type": "text", + "content": " by default. In all experiments, we report the training memory usage. More implementation details can be found in the appendix." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 550, + 224, + 562 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 550, + 224, + 562 + ], + "spans": [ + { + "bbox": [ + 47, + 550, + 224, + 562 + ], + "type": "text", + "content": "4.3. Comparison with SoTA Methods" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 570, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 570, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 570, + 288, + 715 + ], + "type": "text", + "content": "Table 2 compares our AdaTAD with other state-of-the-art (SoTA) methods on ActivityNet-1.3 and THUMOS14 datasets. Initially, we use SlowFast-R50 as the backbone. For comparison, we also extract corresponding offline features, utilizing the snippet representation where each snippet comprises 32 frames with " + }, + { + "bbox": [ + 46, + 570, + 288, + 715 + ], + "type": "inline_equation", + "content": "224^{2}" + }, + { + "bbox": [ + 46, + 570, + 288, + 715 + ], + "type": "text", + "content": " resolution. We observe that end-to-end training enhances performance from " + }, + { + "bbox": [ + 46, + 570, + 288, + 715 + ], + "type": "inline_equation", + "content": "62.3\\%" + }, + { + "bbox": [ + 46, + 570, + 288, + 715 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 46, + 570, + 288, + 715 + ], + "type": "inline_equation", + "content": "66.0\\%" + }, + { + "bbox": [ + 46, + 570, + 288, + 715 + ], + "type": "text", + "content": " on THUMOS14. Notably, this architecture has also been used in E2E-TAD [33]. However, our method consumes less memory while achieving superior performance. This apple-to-apple comparison underscores the benefits of adapter tuning and the scaling-up principle." + } + ] + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 306, + 432, + 547, + 658 + ], + "blocks": [ + { + "bbox": [ + 305, + 406, + 545, + 429 + ], + "lines": [ + { + "bbox": [ + 305, + 406, + 545, + 429 + ], + "spans": [ + { + "bbox": [ + 305, + 406, + 545, + 429 + ], + "type": "text", + "content": "Table 3. Results on EPIC-Kitchens 100 validation set. For comparison, the feature-based methods use the same SlowFast-R50." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 306, + 432, + 547, + 658 + ], + "lines": [ + { + "bbox": [ + 306, + 432, + 547, + 658 + ], + "spans": [ + { + "bbox": [ + 306, + 432, + 547, + 658 + ], + "type": "table", + "html": "
MethodE2E0.10.20.30.40.5Avg.
Verb Task
BMN [30]✘10.88.88.47.15.68.4
G-TAD [60]✘12.111.09.48.16.59.4
ActionFormer [66]✘26.625.424.222.319.123.5
ASL [45]✘27.9-25.5-19.824.6
TriDet [47]✘28.627.426.124.220.825.4
AdaTAD (SlowFast-R50)✓26.525.723.921.717.623.1
ActionFormer (VideoMAE-L)✘32.731.629.126.723.628.7
AdaTAD (VideoMAE-L)✓33.132.230.427.523.129.3
Noun Task
BMN [30]✘10.38.36.24.53.46.5
G-TAD [60]✘11.010.08.67.05.48.4
ActionFormer [66]✘25.224.122.720.517.021.9
ASL [45]✘26.0-23.4-17.722.6
TriDet [47]✘27.426.324.622.218.323.8
AdaTAD (SlowFast-R50)✓24.523.622.320.016.521.4
ActionFormer (VideoMAE-L)✘31.329.727.225.321.326.9
AdaTAD (VideoMAE-L)✓32.431.630.127.424.629.3
", + "image_path": "2ce76e016cb6cbce31a5597cbc8d94889d85a94b1430394e3ea8532a2101fd0e.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 665, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 665, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 305, + 665, + 547, + 715 + ], + "type": "text", + "content": "Furthermore, when adopting the VideoMAE [53] family as our backbone and progressively scaling up the model size, the performance of AdaTAD consistently improves. Using the largest model, i.e., VideoMAEv2-giant with 1.01" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "18596" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 47, + 118, + 291, + 284 + ], + "blocks": [ + { + "bbox": [ + 47, + 70, + 288, + 115 + ], + "lines": [ + { + "bbox": [ + 47, + 70, + 288, + 115 + ], + "spans": [ + { + "bbox": [ + 47, + 70, + 288, + 115 + ], + "type": "text", + "content": "Table 4. Compared to full fine-tuning, our adapter tuning can achieve better performance with less memory. Param. is the number of tunable parameters in the backbone. * means out of memory on A100-80GB, and we report the estimated number." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 47, + 118, + 291, + 284 + ], + "lines": [ + { + "bbox": [ + 47, + 118, + 291, + 284 + ], + "spans": [ + { + "bbox": [ + 47, + 118, + 291, + 284 + ], + "type": "table", + "html": "
ModelSettingE2EParam.Mem.mAP
VideoMAE-SFeature×0-57.6
Snippet Full FT✓22M24.6G68.4
Frame Full FT✓22M2.8G67.1
AdaTAD✓1M2.5G68.8
VideoMAE-BFeature×0-64.7
Snippet Full FT✓86M87.4G*-
Frame Full FT✓86M5.6G70.1
AdaTAD✓4M4.9G71.5
VideoMAE-LFeature×0-66.5
Snippet Full FT✓304M193G*-
Frame Full FT✓304M13.1G73.0
AdaTAD✓14M11.0G73.5
", + "image_path": "ab3e0d25e601d226c26ca0b696bd7093dc67e580e58494b129a98d2402f0956d.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 291, + 287, + 398 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 291, + 287, + 398 + ], + "spans": [ + { + "bbox": [ + 46, + 291, + 287, + 398 + ], + "type": "text", + "content": "billion parameters, and larger input data, i.e., 1536 frames with " + }, + { + "bbox": [ + 46, + 291, + 287, + 398 + ], + "type": "inline_equation", + "content": "224^{2}" + }, + { + "bbox": [ + 46, + 291, + 287, + 398 + ], + "type": "text", + "content": " resolution, we attain an impressive " + }, + { + "bbox": [ + 46, + 291, + 287, + 398 + ], + "type": "inline_equation", + "content": "41.9\\%" + }, + { + "bbox": [ + 46, + 291, + 287, + 398 + ], + "type": "text", + "content": " mAP on ActivityNet-1.3 and " + }, + { + "bbox": [ + 46, + 291, + 287, + 398 + ], + "type": "inline_equation", + "content": "76.9\\%" + }, + { + "bbox": [ + 46, + 291, + 287, + 398 + ], + "type": "text", + "content": " mAP on THUMOS14. It is noteworthy that the previous SoTA was achieved by VideoMAEv2 [56] and InternVideo [57], which utilize the same detector head as ours but with offline snippet features. Our method surpasses these in detection performance by a large margin, marking the first instance where an end-to-end TAD method can outperform SoTA feature-based results." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 399, + 287, + 531 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 399, + 287, + 531 + ], + "spans": [ + { + "bbox": [ + 46, + 399, + 287, + 531 + ], + "type": "text", + "content": "We also present our results on EPIC-Kitchens 100 in Table 3. Since videos in this dataset have a longer duration, all previous methods rely solely on pre-extracted features [47, 52, 66]. Our approach is the first to adopt end-to-end training on this dataset. For fair comparison, we first utilize the same backbone as used in previous methods, i.e., SlowFast-R50 pretrained on EPIC, and we achieve comparable performance to ActionFormer [66]. Moreover, when we scale up the backbone to VideoMAE-L (it is also trained on EPIC-Kitchens 100 classification task), we achieve SoTA performance of " + }, + { + "bbox": [ + 46, + 399, + 287, + 531 + ], + "type": "inline_equation", + "content": "29.3\\%" + }, + { + "bbox": [ + 46, + 399, + 287, + 531 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 537, + 174, + 550 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 537, + 174, + 550 + ], + "spans": [ + { + "bbox": [ + 47, + 537, + 174, + 550 + ], + "type": "text", + "content": "4.4. Ablation and Analysis" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 555, + 287, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 555, + 287, + 604 + ], + "spans": [ + { + "bbox": [ + 46, + 555, + 287, + 604 + ], + "type": "text", + "content": "In this section, we present a series of analyses to evaluate our proposed method and affirm the benefits of scaling up in TAD. Unless otherwise stated, our experiments utilize a standard input of 768 frames per video on THUMOS14." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 605, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 287, + 714 + ], + "type": "text", + "content": "The advantage of adapter tuning. In Table 4, we compare conventional full fine-tuning with our proposed design. It is evident that end-to-end approaches significantly outperform pre-extracted features. Moreover, with full finetuning, the snippet representation slightly advances over frame representation but incurs tremendous memory costs, which aligns with our analysis in Section 3.2. However, AdaTAD uses less memory and still achieves better performance than conventional full fine-tuning. This also veri" + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 307, + 107, + 545, + 216 + ], + "blocks": [ + { + "bbox": [ + 305, + 70, + 545, + 105 + ], + "lines": [ + { + "bbox": [ + 305, + 70, + 545, + 105 + ], + "spans": [ + { + "bbox": [ + 305, + 70, + 545, + 105 + ], + "type": "text", + "content": "Table 5. When scaling up the input data, AdaTAD's performance consistently increases. * means snippet representation is used in offline feature extraction, and each snippet has 16 frames." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 307, + 107, + 545, + 216 + ], + "lines": [ + { + "bbox": [ + 307, + 107, + 545, + 216 + ], + "spans": [ + { + "bbox": [ + 307, + 107, + 545, + 216 + ], + "type": "table", + "html": "
SettingModelResolutionFramesMem.mAP
FeatureVideoMAEv2-g2242768x16*-69.6
AdaTADVideoMAE-S16027682.5G68.8
160215363.8G69.7
160230726.5G70.6
22427683.8G70.7
224215366.4G71.0
2242307211.6G71.5
", + "image_path": "3723db6c851f532d60de1a17b74361000d52030d796f488a242cb41f76d60ff4.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 307, + 247, + 545, + 341 + ], + "blocks": [ + { + "bbox": [ + 305, + 220, + 545, + 244 + ], + "lines": [ + { + "bbox": [ + 305, + 220, + 545, + 244 + ], + "spans": [ + { + "bbox": [ + 305, + 220, + 545, + 244 + ], + "type": "text", + "content": "Table 6. AdaTAD† can further push the boundaries of scaling up. OOM means out of memory on A100-80GB." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 307, + 247, + 545, + 341 + ], + "lines": [ + { + "bbox": [ + 307, + 247, + 545, + 341 + ], + "spans": [ + { + "bbox": [ + 307, + 247, + 545, + 341 + ], + "type": "table", + "html": "
SettingModelResolutionFrameMem.mAP
AdaTADVideoMAE-L160276811.0G73.5
VideoMAEv2-g160276829.9G74.8
VideoMAEv2-g22421536OOM-
AdaTAD†VideoMAEv2-g160276822.8G73.7
VideoMAEv2-g224276830.0G74.6
VideoMAEv2-g2242153643.6G75.4
", + "image_path": "dec7dfa864988cc16197e0482bf45af796b2fe97e4cc7f3b8394b8e997b74c3b.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 348, + 545, + 421 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 348, + 545, + 421 + ], + "spans": [ + { + "bbox": [ + 304, + 348, + 545, + 421 + ], + "type": "text", + "content": "fies the limitations of full fine-tuning, as discussed in Section 3.3. Specifically, our method enhances VideoMAE-S backbone with an " + }, + { + "bbox": [ + 304, + 348, + 545, + 421 + ], + "type": "inline_equation", + "content": "11.2\\%" + }, + { + "bbox": [ + 304, + 348, + 545, + 421 + ], + "type": "text", + "content": " gain using only 1M trainable parameters. Additionally, Table 4 also demonstrates that scaling up the model size of the video backbone is an effective way to improve TAD performance." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 423, + 546, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 423, + 546, + 555 + ], + "spans": [ + { + "bbox": [ + 304, + 423, + 546, + 555 + ], + "type": "text", + "content": "The advantage of scaling up the data. In addition to model scaling, Table 5 verifies the effectiveness of data scaling, which involves two aspects: frame number and frame resolution. Firstly, given the same video duration, increasing the frame number from 768 to 3072 can raise the mAP from " + }, + { + "bbox": [ + 304, + 423, + 546, + 555 + ], + "type": "inline_equation", + "content": "68.8\\%" + }, + { + "bbox": [ + 304, + 423, + 546, + 555 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 304, + 423, + 546, + 555 + ], + "type": "inline_equation", + "content": "70.6\\%" + }, + { + "bbox": [ + 304, + 423, + 546, + 555 + ], + "type": "text", + "content": ". In the meantime, the memory usage is nearly three times larger. Secondly, increasing the frame resolution from " + }, + { + "bbox": [ + 304, + 423, + 546, + 555 + ], + "type": "inline_equation", + "content": "160^2" + }, + { + "bbox": [ + 304, + 423, + 546, + 555 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 304, + 423, + 546, + 555 + ], + "type": "inline_equation", + "content": "224^2" + }, + { + "bbox": [ + 304, + 423, + 546, + 555 + ], + "type": "text", + "content": " also improves the mAP. In the end, by only scaling up the data, we elevate the mAP from " + }, + { + "bbox": [ + 304, + 423, + 546, + 555 + ], + "type": "inline_equation", + "content": "68.8\\%" + }, + { + "bbox": [ + 304, + 423, + 546, + 555 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 304, + 423, + 546, + 555 + ], + "type": "inline_equation", + "content": "71.5\\%" + }, + { + "bbox": [ + 304, + 423, + 546, + 555 + ], + "type": "text", + "content": ", already surpassing the current SoTA feature-based approach with a giant backbone model [57]." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 555, + 545, + 627 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 555, + 545, + 627 + ], + "spans": [ + { + "bbox": [ + 304, + 555, + 545, + 627 + ], + "type": "text", + "content": "Moreover, increasing the frame resolution from " + }, + { + "bbox": [ + 304, + 555, + 545, + 627 + ], + "type": "inline_equation", + "content": "160^2" + }, + { + "bbox": [ + 304, + 555, + 545, + 627 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 304, + 555, + 545, + 627 + ], + "type": "inline_equation", + "content": "224^2" + }, + { + "bbox": [ + 304, + 555, + 545, + 627 + ], + "type": "text", + "content": " or increasing the frame number from 768 to 1536 results in the same memory usage of 3.8G. However, the former achieves " + }, + { + "bbox": [ + 304, + 555, + 545, + 627 + ], + "type": "inline_equation", + "content": "70.7\\%" + }, + { + "bbox": [ + 304, + 555, + 545, + 627 + ], + "type": "text", + "content": " mAP while the latter only reaches " + }, + { + "bbox": [ + 304, + 555, + 545, + 627 + ], + "type": "inline_equation", + "content": "69.7\\%" + }, + { + "bbox": [ + 304, + 555, + 545, + 627 + ], + "type": "text", + "content": ". This suggests that frame resolution may be prioritized under the same memory budget, for the TAD task." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 629, + 546, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 629, + 546, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 629, + 546, + 714 + ], + "type": "text", + "content": "The advantage of AdaTAD†. Given the effectiveness of scaling up the model or data, we further explore combining these approaches. In Table 6, using 768 frames while scaling up the model to VideoMAEv2-giant results in memory usage escalating to 29.9G. In such a scenario, further increasing the data could easily lead to memory overflow, even with the A100-80GB GPU. This indicates that adapta" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18597" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 108 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 108 + ], + "type": "text", + "content": "tion tuning reached its limit under this extreme case. Therefore, to utilize both the largest models with larger data simultaneously, " + }, + { + "bbox": [ + 46, + 72, + 287, + 108 + ], + "type": "inline_equation", + "content": "\\mathrm{AdaTAD}^{\\dagger}" + }, + { + "bbox": [ + 46, + 72, + 287, + 108 + ], + "type": "text", + "content": " shows its advantage." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 108, + 286, + 192 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 108, + 286, + 192 + ], + "spans": [ + { + "bbox": [ + 46, + 108, + 286, + 192 + ], + "type": "text", + "content": "Concretely, when switching from AdaTAD to AdaTAD†, the memory usage of the VideoMAEv2-giant model is reduced from 29.9G to 22.8G. Although a slight performance drop is observed, its reduced memory footprint enables scaling up data from 768 frames to as much as 1536 frames with a high resolution of " + }, + { + "bbox": [ + 46, + 108, + 286, + 192 + ], + "type": "inline_equation", + "content": "224^2" + }, + { + "bbox": [ + 46, + 108, + 286, + 192 + ], + "type": "text", + "content": ". This scalability helps mitigate the performance drop and achieves a higher mAP of " + }, + { + "bbox": [ + 46, + 108, + 286, + 192 + ], + "type": "inline_equation", + "content": "75.4\\%" + }, + { + "bbox": [ + 46, + 108, + 286, + 192 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 194, + 286, + 374 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 194, + 286, + 374 + ], + "spans": [ + { + "bbox": [ + 46, + 194, + 286, + 374 + ], + "type": "text", + "content": "The ablation of the adapter design. Detailed in Table 7, we compare different adapter architectural designs. The baseline, i.e., offline snippet feature, achieves " + }, + { + "bbox": [ + 46, + 194, + 286, + 374 + ], + "type": "inline_equation", + "content": "64.7\\%" + }, + { + "bbox": [ + 46, + 194, + 286, + 374 + ], + "type": "text", + "content": " mAP. End-to-end learning in all designs yields at least a " + }, + { + "bbox": [ + 46, + 194, + 286, + 374 + ], + "type": "inline_equation", + "content": "5\\%" + }, + { + "bbox": [ + 46, + 194, + 286, + 374 + ], + "type": "text", + "content": " improvement. Our AdaTAD achieves " + }, + { + "bbox": [ + 46, + 194, + 286, + 374 + ], + "type": "inline_equation", + "content": "71.5\\%" + }, + { + "bbox": [ + 46, + 194, + 286, + 374 + ], + "type": "text", + "content": " in the end. Compared to standard adapter [22], ours consumes similar memory but achieves higher mAP. This verifies that local temporal context is vital for the TAD task. In contrast to full finetuning (FT), we tune only 4M parameters using less memory. In our design, we find that removing the residual connection of the depth-wise convolution drops performance by " + }, + { + "bbox": [ + 46, + 194, + 286, + 374 + ], + "type": "inline_equation", + "content": "0.7\\%" + }, + { + "bbox": [ + 46, + 194, + 286, + 374 + ], + "type": "text", + "content": ", and training becomes unstable. We also implement an adaptation design proposed in LongLoRA [9], which efficiently computes long-range attention and shows decent performance but requires more parameters and memory." + } + ] + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 48, + 408, + 282, + 504 + ], + "blocks": [ + { + "bbox": [ + 47, + 385, + 286, + 407 + ], + "lines": [ + { + "bbox": [ + 47, + 385, + 286, + 407 + ], + "spans": [ + { + "bbox": [ + 47, + 385, + 286, + 407 + ], + "type": "text", + "content": "Table 7. Ablation of different adapter architectural designs. VideoMAE-B is used to conduct the following experiments." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 48, + 408, + 282, + 504 + ], + "lines": [ + { + "bbox": [ + 48, + 408, + 282, + 504 + ], + "spans": [ + { + "bbox": [ + 48, + 408, + 282, + 504 + ], + "type": "table", + "html": "
SettingE2EParam.Mem.mAPgains
Snippet FeatureX0-64.7
+ Full FT✓86M5.6G70.1+5.1
+ LongLoRA [9]✓28M6.2G71.1+6.1
+ Standard Adapter [22]✓3.6M4.8G70.2+5.2
+ AdaTAD (w/o residual)✓4.0M4.9G70.8+5.8
+ AdaTAD✓4.0M4.9G71.5+6.5
", + "image_path": "108e6f6346040b967e30f4ea1baaee4c38c22881057c907e38e94acb5b578ead.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 510, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 510, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 510, + 287, + 713 + ], + "type": "text", + "content": "The necessity of end-to-end training for TAD. As previously discussed, end-to-end training can address discrepancies between the pretraining and fine-tuning stages in terms of training data and learning tasks. To corroborate this, we employ models pretrained on different datasets for the EPIC-Kitchens TAD task in Table 8. Kinetics-400 (K400) [26] represents commonly collected third-person web data and exhibits a large domain gap compared to EPIC-Kitchens 100. Using K400 for pretraining, we observe that end-to-end TAD training allows for " + }, + { + "bbox": [ + 46, + 510, + 287, + 713 + ], + "type": "inline_equation", + "content": "+5.56" + }, + { + "bbox": [ + 46, + 510, + 287, + 713 + ], + "type": "text", + "content": " gain. Conversely, using a model already finetuned on EPIC-Kitchens still yields a " + }, + { + "bbox": [ + 46, + 510, + 287, + 713 + ], + "type": "inline_equation", + "content": "+2.32" + }, + { + "bbox": [ + 46, + 510, + 287, + 713 + ], + "type": "text", + "content": " improvement. Unlike K400 pretraining, since this model has already adapted to the data discrepancy, we can infer that this gain leverages differences between the classification task in pretraining and the detection task in fine-tuning. Such improvements further underscore the significance of end-to-end training in TAD." + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 307, + 107, + 547, + 180 + ], + "blocks": [ + { + "bbox": [ + 305, + 71, + 545, + 104 + ], + "lines": [ + { + "bbox": [ + 305, + 71, + 545, + 104 + ], + "spans": [ + { + "bbox": [ + 305, + 71, + 545, + 104 + ], + "type": "text", + "content": "Table 8. End-to-end TAD can alleviate the discrepancies between pretraining and finetuning. VideoMAE-L with different pretrained weights are used on the EPIC-Kitchens 100 Noun task." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 307, + 107, + 547, + 180 + ], + "lines": [ + { + "bbox": [ + 307, + 107, + 547, + 180 + ], + "spans": [ + { + "bbox": [ + 307, + 107, + 547, + 180 + ], + "type": "table", + "html": "
Pretrain DatasetE2E0.10.30.5mAPgain
K400 [26]X18.6916.3511.5215.77
K400 [26]✓24.3322.1416.8721.33+ 5.56
K400 [26] + EPIC [13]X31.3227.2521.3326.98
K400 [26] + EPIC [13]✓32.4130.1324.5929.30+ 2.32
", + "image_path": "e6712dad19a7ac7947df5d4a5745da28320bd84010135ad315a4dd71653f4418.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 184, + 400, + 196 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 184, + 400, + 196 + ], + "spans": [ + { + "bbox": [ + 306, + 184, + 400, + 196 + ], + "type": "text", + "content": "4.5. Error Analyses" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 202, + 545, + 274 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 202, + 545, + 274 + ], + "spans": [ + { + "bbox": [ + 304, + 202, + 545, + 274 + ], + "type": "text", + "content": "We also conduct false positive analysis at tIoU=0.5 in Fig. 4. Compared to feature-based training, learning from raw frames produces more helpful positive detections. More importantly, the percentage of wrong label error is reduced after end-to-end training, suggesting its unique advantage in classifying accurate action labels." + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 314, + 277, + 542, + 389 + ], + "blocks": [ + { + "bbox": [ + 314, + 277, + 542, + 389 + ], + "lines": [ + { + "bbox": [ + 314, + 277, + 542, + 389 + ], + "spans": [ + { + "bbox": [ + 314, + 277, + 542, + 389 + ], + "type": "image", + "image_path": "f12d6fe5149da566afa57be00d00b5ae3e7c600bcfcdc598394dfe43272757c7.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 392, + 545, + 425 + ], + "lines": [ + { + "bbox": [ + 305, + 392, + 545, + 425 + ], + "spans": [ + { + "bbox": [ + 305, + 392, + 545, + 425 + ], + "type": "text", + "content": "Figure 4. False Positive Profiling on THUMOS14 using [2]. We use VideoMAEv2-giant as the backbone, and compare end-to-end training with pre-extracted feature-based training." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 430, + 383, + 441 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 430, + 383, + 441 + ], + "spans": [ + { + "bbox": [ + 306, + 430, + 383, + 441 + ], + "type": "text", + "content": "5. Conclusions" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 450, + 545, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 450, + 545, + 641 + ], + "spans": [ + { + "bbox": [ + 304, + 450, + 545, + 641 + ], + "type": "text", + "content": "This work introduces a memory-efficient and parameter-efficient end-to-end method named AdaTAD. Our key innovation lies in the proposed temporal-informative adapter, which is tailored for TAD with low computation costs. Furthermore, we design an alternative placement for adapters to minimize memory usage. By demonstrating the feasibility and effectiveness of scaling up end-to-end TAD, our work achieves new SoTA performance across multiple datasets. Particularly, this is the first instance of an end-to-end TAD method that surpasses the current best feature-based models. In fact, AdaTAD achieves a groundbreaking " + }, + { + "bbox": [ + 304, + 450, + 545, + 641 + ], + "type": "inline_equation", + "content": "75.4\\%" + }, + { + "bbox": [ + 304, + 450, + 545, + 641 + ], + "type": "text", + "content": " mAP on THUMOS14. We believe our work underscores the possible paradigm shift in TAD, advocating a move away from the traditional methodology of separate feature extraction and offline detection towards a more integrated approach of scaling up end-to-end TAD training." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "type": "text", + "content": "Acknowledgement. This work was supported by the King Abdullah University of Science and Technology (KAUST) Office of Sponsored Research through the Visual Computing Center (VCC) funding, as well as the SDAIA-KAUST Center of Excellence in Data Science and Artificial Intelligence (SDAIA-KAUST AI)." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18598" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 123 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 123 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 123 + ], + "type": "text", + "content": "[1] Ibrahim M Alabdulmohsin, Behnam Neyshabur, and Xiaohua Zhai. Revisiting neural scaling laws in language and vision. In NeurIPS, 2022. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 125, + 287, + 158 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 125, + 287, + 158 + ], + "spans": [ + { + "bbox": [ + 53, + 125, + 287, + 158 + ], + "type": "text", + "content": "[2] Humam Alwassel, Fabian Caba Heilbron, Victor Escorcia, and Bernard Ghanem. Diagnosing error in temporal action detectors. In ECCV, 2018. 8" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 159, + 288, + 202 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 159, + 288, + 202 + ], + "spans": [ + { + "bbox": [ + 53, + 159, + 288, + 202 + ], + "type": "text", + "content": "[3] Humam Alwassel, Fabian Caba Heilbron, and Bernard Ghanem. Action search: Spotting actions in videos and its application to temporal action localization. In ECCV, 2018. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 205, + 288, + 247 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 205, + 288, + 247 + ], + "spans": [ + { + "bbox": [ + 53, + 205, + 288, + 247 + ], + "type": "text", + "content": "[4] Yueran Bai, Yingying Wang, Yunhai Tong, Yang Yang, Qiyue Liu, and Junhui Liu. Boundary content graph neural network for temporal action proposal generation. In ECCV, 2020. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 250, + 288, + 294 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 250, + 288, + 294 + ], + "spans": [ + { + "bbox": [ + 53, + 250, + 288, + 294 + ], + "type": "text", + "content": "[5] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. In NeurIPS, 2020. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 296, + 288, + 339 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 296, + 288, + 339 + ], + "spans": [ + { + "bbox": [ + 53, + 296, + 288, + 339 + ], + "type": "text", + "content": "[6] Hao Chen, Ran Tao, Han Zhang, Yidong Wang, Wei Ye, Jindong Wang, Guosheng Hu, and Marios Savvides. Convadapter: Exploring parameter efficient transfer learning for convnets. arXiv preprint arXiv:2208.07463, 2022. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 341, + 288, + 384 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 341, + 288, + 384 + ], + "spans": [ + { + "bbox": [ + 53, + 341, + 288, + 384 + ], + "type": "text", + "content": "[7] Shoufa Chen, Chongjian Ge, Zhan Tong, Jiangliu Wang, Yibing Song, Jue Wang, and Ping Luo. Adaptformer: Adapting vision transformers for scalable visual recognition. In NeurIPS, 2022. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 386, + 288, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 386, + 288, + 418 + ], + "spans": [ + { + "bbox": [ + 53, + 386, + 288, + 418 + ], + "type": "text", + "content": "[8] Tianqi Chen, Bing Xu, Chiyuan Zhang, and Carlos Guestrin. Training deep nets with sublinear memory cost. arXiv preprint arXiv:1604.06174, 2016. 4" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 420, + 288, + 463 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 420, + 288, + 463 + ], + "spans": [ + { + "bbox": [ + 53, + 420, + 288, + 463 + ], + "type": "text", + "content": "[9] Yukang Chen, Shengju Qian, Haotian Tang, Xin Lai, Zhi-jian Liu, Song Han, and Jiaya Jia. Longlora: Efficient fin-tuning of long-context large language models. arXiv preprint arXiv:2309.12307, 2023. 8" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 465, + 287, + 497 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 465, + 287, + 497 + ], + "spans": [ + { + "bbox": [ + 48, + 465, + 287, + 497 + ], + "type": "text", + "content": "[10] Feng Cheng and Gedas Bertasius. Tallformer: Temporal action localization with long-memory transformer. In ECCV, 2022. 1, 2, 4, 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 499, + 287, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 499, + 287, + 542 + ], + "spans": [ + { + "bbox": [ + 48, + 499, + 287, + 542 + ], + "type": "text", + "content": "[11] Feng Cheng, Mingze Xu, Yuanjun Xiong, Hao Chen, Xinyu Li, Wei Li, and Wei Xia. Stochastic backpropagation: A memory efficient strategy for training video models. In CVPR, 2023. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 544, + 287, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 544, + 287, + 577 + ], + "spans": [ + { + "bbox": [ + 48, + 544, + 287, + 577 + ], + "type": "text", + "content": "[12] MMAction2 Contributors. Openmmlab's next generation video understanding toolbox and benchmark. https://github.com/open-mmlab/mmaction2, 2020.5" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 578, + 287, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 578, + 287, + 633 + ], + "spans": [ + { + "bbox": [ + 48, + 578, + 287, + 633 + ], + "type": "text", + "content": "[13] Dima Damen, Hazel Doughty, Giovanni Maria Farinella, Sanja Fidler, Antonino Furnari, Evangelos Kazakos, Davide Moltisanti, Jonathan Munro, Toby Perrett, Will Price, et al. Scaling egocentric vision: The epic-kitchens dataset. In ECCV, 2018. 5, 8" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 635, + 287, + 699 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 635, + 287, + 699 + ], + "spans": [ + { + "bbox": [ + 48, + 635, + 287, + 699 + ], + "type": "text", + "content": "[14] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2021. 1, 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 702, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 702, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 702, + 287, + 713 + ], + "type": "text", + "content": "[15] Victor Escorcia, Fabian Caba Heilbron, Juan Carlos Niebles," + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 34, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "text", + "content": "and Bernard Ghanem. DAPs: Deep action proposals for action understanding. In ECCV, 2016. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 97, + 545, + 129 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 97, + 545, + 129 + ], + "spans": [ + { + "bbox": [ + 308, + 97, + 545, + 129 + ], + "type": "text", + "content": "[16] Haoqi Fan, Bo Xiong, Karttikeya Mangalam, Yanghao Li, Zhicheng Yan, Jitendra Malik, and Christoph Feichtenhofer. Multiscale vision transformers. In ICCV, 2021. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 131, + 545, + 163 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 131, + 545, + 163 + ], + "spans": [ + { + "bbox": [ + 308, + 131, + 545, + 163 + ], + "type": "text", + "content": "[17] Christoph Feichtenhofer, Haoqi Fan, Jitendra Malik, and Kaiming He. SlowFast networks for video recognition. In ICCV, 2019. 3, 4" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 166, + 545, + 220 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 166, + 545, + 220 + ], + "spans": [ + { + "bbox": [ + 308, + 166, + 545, + 220 + ], + "type": "text", + "content": "[18] Kristen Grauman, Andrew Westbury, Eugene Byrne, Zachary Chavis, Antonino Furnari, Rohit Girdhar, Jackson Hamburger, Hao Jiang, Miao Liu, Xingyu Liu, et al. Ego4d: Around the world in 3,000 hours of egocentric video. In CVPR, 2022. 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 222, + 545, + 255 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 222, + 545, + 255 + ], + "spans": [ + { + "bbox": [ + 308, + 222, + 545, + 255 + ], + "type": "text", + "content": "[19] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, and Ross Girshick. Masked autoencoders are scalable vision learners. In CVPR, 2022. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 257, + 545, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 257, + 545, + 300 + ], + "spans": [ + { + "bbox": [ + 308, + 257, + 545, + 300 + ], + "type": "text", + "content": "[20] Fabian Caba Heilbron, Victor Escorcia, Bernard Ghanem, and Juan Carlos Niebles. ActivityNet: A large-scale video benchmark for human activity understanding. In CVPR, 2015. 1, 5" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 302, + 545, + 324 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 302, + 545, + 324 + ], + "spans": [ + { + "bbox": [ + 308, + 302, + 545, + 324 + ], + "type": "text", + "content": "[21] Dan Hendrycks and Kevin Gimpel. Gaussian error linear units (gelus). arXiv preprint arXiv:1606.08415, 2016. 4" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 326, + 545, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 326, + 545, + 369 + ], + "spans": [ + { + "bbox": [ + 308, + 326, + 545, + 369 + ], + "type": "text", + "content": "[22] Neil Houlsby, Andrei Giurgiu, Stanislaw Jastrzebski, Bruna Morrone, Quentin De Laroussilhe, Andrea Gesmundo, Mona Attariyan, and Sylvain Gelly. Parameter-efficient transfer learning for nlp. In ICML, 2019. 2, 3, 4, 8" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 372, + 545, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 372, + 545, + 415 + ], + "spans": [ + { + "bbox": [ + 308, + 372, + 545, + 415 + ], + "type": "text", + "content": "[23] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. In ICML, 2021. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 417, + 545, + 449 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 417, + 545, + 449 + ], + "spans": [ + { + "bbox": [ + 308, + 417, + 545, + 449 + ], + "type": "text", + "content": "[24] Menglin Jia, Luming Tang, Bor-Chun Chen, Claire Cardie, Serge Belongie, Bharath Hariharan, and Ser-Nam Lim. Visual prompt tuning. In ECCV, 2022. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 452, + 545, + 484 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 452, + 545, + 484 + ], + "spans": [ + { + "bbox": [ + 308, + 452, + 545, + 484 + ], + "type": "text", + "content": "[25] YG Jiang, J Liu, A Roshan Zamir, G Toderici, I Laptev, M Shah, and R Sukthankar. Thumos challenge: Action recognition with a large number of classes, 2014. 1, 5" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 486, + 545, + 540 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 486, + 545, + 540 + ], + "spans": [ + { + "bbox": [ + 308, + 486, + 545, + 540 + ], + "type": "text", + "content": "[26] Will Kay, Joao Carreira, Karen Simonyan, Brian Zhang, Chloe Hillier, Sudheendra Vijayanarasimhan, Fabio Viola, Tim Green, Trevor Back, Paul Natsev, et al. The kinetics human action video dataset. arXiv preprint arXiv:1705.06950, 2017.8" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 543, + 545, + 574 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 543, + 545, + 574 + ], + "spans": [ + { + "bbox": [ + 308, + 543, + 545, + 574 + ], + "type": "text", + "content": "[27] Brian Lester, Rami Al-Rfou, and Noah Constant. The power of scale for parameter-efficient prompt tuning. In EMNLP, 2021. 3" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 308, + 577, + 545, + 598 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 577, + 545, + 598 + ], + "spans": [ + { + "bbox": [ + 308, + 577, + 545, + 598 + ], + "type": "text", + "content": "[28] Xiang Lisa Li and Percy Liang. Prefix-tuning: Optimizing continuous prompts for generation. In ACL, 2021. 3" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 308, + 601, + 545, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 601, + 545, + 644 + ], + "spans": [ + { + "bbox": [ + 308, + 601, + 545, + 644 + ], + "type": "text", + "content": "[29] Chuming Lin, Chengming Xu, Donghao Luo, Yabiao Wang, Ying Tai, Chengjie Wang, Jilin Li, Feiyue Huang, and Yanwei Fu. Learning salient boundary feature for anchor-free temporal action localization. In CVPR, 2021. 1, 2, 4, 6" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 308, + 646, + 545, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 646, + 545, + 678 + ], + "spans": [ + { + "bbox": [ + 308, + 646, + 545, + 678 + ], + "type": "text", + "content": "[30] Tianwei Lin, Xiao Liu, Xin Li, Errui Ding, and Shilei Wen. BMN: boundary-matching network for temporal action proposal generation. In ICCV, 2019. 1, 2, 6" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 308, + 681, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 681, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 308, + 681, + 545, + 712 + ], + "type": "text", + "content": "[31] Qinying Liu and Zilei Wang. Progressive boundary refinement network for temporal action detection. In AAAI, 2020. 2" + } + ] + } + ], + "index": 33 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "text", + "content": "18599" + } + ] + } + ], + "index": 35 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "type": "text", + "content": "[32] Shuming Liu, Mengmeng Xu, Chen Zhao, Xu Zhao, and Bernard Ghanem. Etad: Training action detection end to end on a laptop. In CVPRW, 2023. 1, 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 107, + 287, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 107, + 287, + 139 + ], + "spans": [ + { + "bbox": [ + 48, + 107, + 287, + 139 + ], + "type": "text", + "content": "[33] Xiaolong Liu, Song Bai, and Xiang Bai. An empirical study of end-to-end temporal action detection. In CVPR, 2022. 1, 2, 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 141, + 287, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 141, + 287, + 184 + ], + "spans": [ + { + "bbox": [ + 48, + 141, + 287, + 184 + ], + "type": "text", + "content": "[34] Xiaolong Liu, Qimeng Wang, Yao Hu, Xu Tang, Shiwei Zhang, Song Bai, and Xiang Bai. End-to-end temporal action detection with transformer. IEEE Transactions on Image Processing, 31:5427-5441, 2022. 2, 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 186, + 287, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 186, + 287, + 228 + ], + "spans": [ + { + "bbox": [ + 48, + 186, + 287, + 228 + ], + "type": "text", + "content": "[35] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In ICCV, 2021. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 231, + 287, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 231, + 287, + 285 + ], + "spans": [ + { + "bbox": [ + 48, + 231, + 287, + 285 + ], + "type": "text", + "content": "[36] Jinjie Mai, Abdullah Hamdi, Silvio Giancola, Chen Zhao, and Bernard Ghanem. Egoloc: Revisiting 3d object localization from egocentric videos with visual queries. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 45-57, 2023. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 286, + 287, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 286, + 287, + 330 + ], + "spans": [ + { + "bbox": [ + 48, + 286, + 287, + 330 + ], + "type": "text", + "content": "[37] Paulius Micikevicius, Sharan Narang, Jonah Alben, Gregory Diamos, Erich Elsen, David Garcia, Boris Ginsburg, Michael Houston, Oleksii Kuchaiev, Ganesh Venkatesh, et al. Mixed precision training. In ICLR, 2017. 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 332, + 287, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 332, + 287, + 364 + ], + "spans": [ + { + "bbox": [ + 48, + 332, + 287, + 364 + ], + "type": "text", + "content": "[38] Jonghwan Mun, Minsu Cho, and Bohyung Han. Local-global video-text interactions for temporal grounding. In CVPR, 2020. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 365, + 226, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 365, + 226, + 376 + ], + "spans": [ + { + "bbox": [ + 48, + 365, + 226, + 376 + ], + "type": "text", + "content": "[39] OpenAI. Gpt-4 technical report, 2023. 1, 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 377, + 287, + 421 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 377, + 287, + 421 + ], + "spans": [ + { + "bbox": [ + 48, + 377, + 287, + 421 + ], + "type": "text", + "content": "[40] Zhiwu Qing, Haisheng Su, Weihao Gan, Dongliang Wang, Wei Wu, Xiang Wang, Yu Qiao, Junjie Yan, Changxin Gao, and Nong Sang. Temporal context aggregation network for temporal action proposal refinement. In CVPR, 2021. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 422, + 287, + 455 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 422, + 287, + 455 + ], + "spans": [ + { + "bbox": [ + 48, + 422, + 287, + 455 + ], + "type": "text", + "content": "[41] Alec Radford, Karthik Narasimhan, Tim Salimans, Ilya Sutskever, et al. Improving language understanding by generative pre-training. OpenAI, 2018. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 456, + 287, + 489 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 456, + 287, + 489 + ], + "spans": [ + { + "bbox": [ + 48, + 456, + 287, + 489 + ], + "type": "text", + "content": "[42] Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, Ilya Sutskever, et al. Language models are unsupervised multitask learners. OpenAI blog, 1(8):9, 2019. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 491, + 287, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 491, + 287, + 544 + ], + "spans": [ + { + "bbox": [ + 48, + 491, + 287, + 544 + ], + "type": "text", + "content": "[43] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, 2021. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 545, + 287, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 545, + 287, + 610 + ], + "spans": [ + { + "bbox": [ + 48, + 545, + 287, + 610 + ], + "type": "text", + "content": "[44] Merey Ramazanova, Victor Escorcia, Fabian Caba Heilbron, Chen Zhao, and Bernard Ghanem. Owl (observe, watch, listen): Localizing actions in egocentric video via audiovisual temporal context. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshop (CVPRW), 2023. 1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 613, + 287, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 613, + 287, + 644 + ], + "spans": [ + { + "bbox": [ + 48, + 613, + 287, + 644 + ], + "type": "text", + "content": "[45] Jiayi Shao, Xiaohan Wang, Ruijie Quan, Junjun Zheng, Jiang Yang, and Yi Yang. Action sensitivity learning for temporal action localization. In ICCV, 2023. 2, 6" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 647, + 287, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 647, + 287, + 679 + ], + "spans": [ + { + "bbox": [ + 48, + 647, + 287, + 679 + ], + "type": "text", + "content": "[46] Dingfeng Shi, Yujie Zhong, Qiong Cao, Jing Zhang, Lin Ma, Jia Li, and Dacheng Tao. React: Temporal action detection with relational queries. In ECCV, 2022. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 681, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 681, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 681, + 287, + 712 + ], + "type": "text", + "content": "[47] Dingfeng Shi, Yujie Zhong, Qiong Cao, Lin Ma, Jia Li, and Dacheng Tao. Tridet: Temporal action detection with relative boundary modeling. In CVPR, 2023. 2, 6, 7" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 34, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "text", + "content": "[48] Mattia Soldan, Mengmeng Xu, Sisi Qu, Jesper Tegner, and Bernard Ghanem. VLG-Net: Video-language graph matching network for video grounding. In ICCVW, 2021. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 107, + 545, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 107, + 545, + 139 + ], + "spans": [ + { + "bbox": [ + 307, + 107, + 545, + 139 + ], + "type": "text", + "content": "[49] Yi-Lin Sung, Jaemin Cho, and Mohit Bansal. Lst: Ladder side-tuning for parameter and memory efficient transfer learning. In NeurIPS, 2022. 3, 5" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 140, + 545, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 140, + 545, + 171 + ], + "spans": [ + { + "bbox": [ + 308, + 140, + 545, + 171 + ], + "type": "text", + "content": "[50] Yi-Lin Sung, Jaemin Cho, and Mohit Bansal. VI-adapter: Parameter-efficient transfer learning for vision-and-language tasks. In CVPR, 2022. 2, 4" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 173, + 545, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 173, + 545, + 205 + ], + "spans": [ + { + "bbox": [ + 308, + 173, + 545, + 205 + ], + "type": "text", + "content": "[51] Jing Tan, Jiaqi Tang, Limin Wang, and Gangshan Wu. Relaxed transformer decoders for direct action proposal generation. In ICCV, 2021. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 206, + 545, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 206, + 545, + 248 + ], + "spans": [ + { + "bbox": [ + 308, + 206, + 545, + 248 + ], + "type": "text", + "content": "[52] Tuan N Tang, Kwonyoung Kim, and Kwanghoon Sohn. Temporalmaxer: Maximize temporal context with only max pooling for temporal action localization. arXiv preprint arXiv:2303.09055, 2023. 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 251, + 545, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 251, + 545, + 293 + ], + "spans": [ + { + "bbox": [ + 308, + 251, + 545, + 293 + ], + "type": "text", + "content": "[53] Zhan Tong, Yibing Song, Jue Wang, and Limin Wang. Videomae: Masked autoencoders are data-efficient learners for self-supervised video pre-training. In NeurIPS, 2022. 1, 4, 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 294, + 545, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 294, + 545, + 327 + ], + "spans": [ + { + "bbox": [ + 308, + 294, + 545, + 327 + ], + "type": "text", + "content": "[54] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In NeurIPS, 2017. 4" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 328, + 545, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 328, + 545, + 360 + ], + "spans": [ + { + "bbox": [ + 308, + 328, + 545, + 360 + ], + "type": "text", + "content": "[55] Chenhao Wang, Hongxiang Cai, Yuxin Zou, and Yichao Xiong. Rgb stream is enough for temporal action detection. arXiv preprint arXiv:2107.04362, 2021. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 361, + 545, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 361, + 545, + 403 + ], + "spans": [ + { + "bbox": [ + 308, + 361, + 545, + 403 + ], + "type": "text", + "content": "[56] Limin Wang, Bingkun Huang, Zhiyu Zhao, Zhan Tong, Yi-nan He, Yi Wang, Yali Wang, and Yu Qiao. Videomae v2: Scaling video masked autoencoders with dual masking. In ICCV, 2023. 1, 3, 6, 7" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 405, + 545, + 458 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 405, + 545, + 458 + ], + "spans": [ + { + "bbox": [ + 308, + 405, + 545, + 458 + ], + "type": "text", + "content": "[57] Yi Wang, Kunchang Li, Yizhuo Li, Yinan He, Bingkun Huang, Zhiyu Zhao, Hongjie Zhang, Jilan Xu, Yi Liu, Zun Wang, et al. Internvideo: General video foundation models via generative and discriminative learning. arXiv preprint arXiv:2212.03191, 2022. 1, 6, 7" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 460, + 545, + 492 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 460, + 545, + 492 + ], + "spans": [ + { + "bbox": [ + 308, + 460, + 545, + 492 + ], + "type": "text", + "content": "[58] Zhenzhi Wang, Ziteng Gao, Limin Wang, Zhifeng Li, and Gangshan Wu. Boundary-aware cascade networks for temporal action segmentation. In ECCV, 2020. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 494, + 545, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 494, + 545, + 525 + ], + "spans": [ + { + "bbox": [ + 308, + 494, + 545, + 525 + ], + "type": "text", + "content": "[59] Kun Xia, Le Wang, Sanping Zhou, Nanning Zheng, and Wei Tang. Learning to refactor action and co-occurrence features for temporal action localization. In CVPR, 2022. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 526, + 545, + 558 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 526, + 545, + 558 + ], + "spans": [ + { + "bbox": [ + 308, + 526, + 545, + 558 + ], + "type": "text", + "content": "[60] Mengmeng Xu, Chen Zhao, David S Rojas, Ali Thabet, and Bernard Ghanem. G-TAD: Sub-graph localization for temporal action detection. In CVPR, 2020. 1, 6" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 308, + 559, + 545, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 559, + 545, + 601 + ], + "spans": [ + { + "bbox": [ + 308, + 559, + 545, + 601 + ], + "type": "text", + "content": "[61] Le Yang, Houwen Peng, Dingwen Zhang, Jianlong Fu, and Junwei Han. Revisiting anchor mechanisms for temporal action localization. IEEE Transactions on Image Processing, 29:8535-8548, 2020. 2" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 308, + 604, + 545, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 604, + 545, + 646 + ], + "spans": [ + { + "bbox": [ + 308, + 604, + 545, + 646 + ], + "type": "text", + "content": "[62] Min Yang, Guo Chen, Yin-Dong Zheng, Tong Lu, and Limin Wang. Basictad: an astounding rgb-only baseline for temporal action detection. Computer Vision and Image Understanding, 232:103692, 2023. 1, 2, 6" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 308, + 647, + 545, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 647, + 545, + 679 + ], + "spans": [ + { + "bbox": [ + 308, + 647, + 545, + 679 + ], + "type": "text", + "content": "[63] Taojiannan Yang, Yi Zhu, Yusheng Xie, Aston Zhang, Chen Chen, and Mu Li. Aim: Adapting image models for efficient video action recognition. In ICLR, 2023. 2, 3, 4" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 308, + 681, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 681, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 308, + 681, + 545, + 712 + ], + "type": "text", + "content": "[64] Ting Yao, Tao Mei, and Yong Rui. Highlight detection with pairwise deep ranking for first-person video summarization. In CVPR, 2016. 1" + } + ] + } + ], + "index": 33 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "text", + "content": "18600" + } + ] + } + ], + "index": 35 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 464 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "text", + "content": "[65] Dongshuo Yin, Xueting Han, Bin Li, Hao Feng, and Jing Bai. Parameter-efficient is not sufficient: Exploring parameter, memory, and time efficient adapter tuning for dense predictions. arXiv preprint arXiv:2306.09729, 2023. 3, 5" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 118, + 287, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 118, + 287, + 150 + ], + "spans": [ + { + "bbox": [ + 49, + 118, + 287, + 150 + ], + "type": "text", + "content": "[66] Chenlin Zhang, Jianxin Wu, and Yin Li. Actionformer: Localizing moments of actions with transformers. In ECCV, 2022. 1, 2, 3, 5, 6, 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 152, + 287, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 152, + 287, + 184 + ], + "spans": [ + { + "bbox": [ + 48, + 152, + 287, + 184 + ], + "type": "text", + "content": "[67] Chen Zhao, Ali K Thabet, and Bernard Ghanem. Video self-stitching graph network for temporal action localization. In ICCV, 2021. 1, 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 186, + 287, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 186, + 287, + 217 + ], + "spans": [ + { + "bbox": [ + 48, + 186, + 287, + 217 + ], + "type": "text", + "content": "[68] Chen Zhao, Merey Ramazanova, Mengmeng Xu, and Bernard Ghanem. Segstad: Precise temporal action detection via semantic segmentation. In ECCVW, 2022. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 220, + 287, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 220, + 287, + 262 + ], + "spans": [ + { + "bbox": [ + 48, + 220, + 287, + 262 + ], + "type": "text", + "content": "[69] Chen Zhao, Shuming Liu, Karttikeya Mangalam, and Bernard Ghanem. Re" + }, + { + "bbox": [ + 48, + 220, + 287, + 262 + ], + "type": "inline_equation", + "content": "^2" + }, + { + "bbox": [ + 48, + 220, + 287, + 262 + ], + "type": "text", + "content": "TAL: Rewiring pretrained video backbones for reversible temporal action localization. In CVPR, 2023. 1, 2, 3, 4, 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 264, + 287, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 264, + 287, + 330 + ], + "spans": [ + { + "bbox": [ + 48, + 264, + 287, + 330 + ], + "type": "text", + "content": "[70] Chen Zhao, Shuming Liu, Karttikeya Mangalam, Guocheng Qian, Fatimah Zohra, Abdulmohsen Alghannam, Jitendra Malik, and Bernard Ghanem. " + }, + { + "bbox": [ + 48, + 264, + 287, + 330 + ], + "type": "inline_equation", + "content": "\\mathrm{Dr}^2\\mathrm{Net}" + }, + { + "bbox": [ + 48, + 264, + 287, + 330 + ], + "type": "text", + "content": ": Dynamic reversible dual-residual networks for memory-efficient finetuning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 1, 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 331, + 287, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 331, + 287, + 364 + ], + "spans": [ + { + "bbox": [ + 48, + 331, + 287, + 364 + ], + "type": "text", + "content": "[71] Hang Zhao, Zhicheng Yan, Lorenzo Torresani, and Antonio Torralba. HACS: Human action clips and segments dataset for recognition and temporal localization. ICCV, 2019. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 365, + 287, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 365, + 287, + 408 + ], + "spans": [ + { + "bbox": [ + 48, + 365, + 287, + 408 + ], + "type": "text", + "content": "[72] Y Zhao, B Zhang, Z Wu, S Yang, L Zhou, S Yan, L Wang, Y Xiong, D Lin, Y Qiao, et al. Cuhk & ethz & siat submission to activitynet challenge 2017. CVPR ActivityNet Workshop, 2017. 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 410, + 287, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 410, + 287, + 464 + ], + "spans": [ + { + "bbox": [ + 48, + 410, + 287, + 464 + ], + "type": "text", + "content": "[73] Zixuan Zhao, Dongqi Wang, and Xu Zhao. Movement enhancement toward multi-scale video feature representation for temporal action detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 13555-13564, 2023. 2" + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "18601" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Endow SAM with Keen Eyes_ Temporal-spatial Prompt Learning for Video Camouflaged Object Detection/e8cb1a37-360c-44d3-92de-d83afed15e31_content_list.json b/2024/Endow SAM with Keen Eyes_ Temporal-spatial Prompt Learning for Video Camouflaged Object Detection/e8cb1a37-360c-44d3-92de-d83afed15e31_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..7c6d742e05dac7d900ff1302569b6d88edff678c --- /dev/null +++ b/2024/Endow SAM with Keen Eyes_ Temporal-spatial Prompt Learning for Video Camouflaged Object Detection/e8cb1a37-360c-44d3-92de-d83afed15e31_content_list.json @@ -0,0 +1,1854 @@ +[ + { + "type": "text", + "text": "Endow SAM with Keen Eyes: Temporal-spatial Prompt Learning for Video Camouflaged Object Detection", + "text_level": 1, + "bbox": [ + 106, + 128, + 864, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Wenjun Hui $^{1,2}$ , Zhenfeng Zhu $^{1,2*}$ , Shuai Zheng $^{1,2}$ , Yao Zhao $^{1,2}$", + "bbox": [ + 238, + 202, + 727, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Institute of Information Science, Beijing Jiaotong University", + "bbox": [ + 303, + 223, + 668, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ Beijing Key Laboratory of Advanced Information Science and Network Technology", + "bbox": [ + 233, + 241, + 736, + 256 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{wenjunhui, zhfzhu, zs1997, yzhao}@bjtu.edu.cn", + "bbox": [ + 279, + 258, + 684, + 273 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 308, + 313, + 325 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The Segment Anything Model (SAM), a prompt-driven foundational model, has demonstrated remarkable performance in natural image segmentation. However, its application in video camouflaged object detection (VCOD) encounters challenges, chiefly stemming from the overlooked temporal-spatial associations and the unreliability of user-provided prompts for camouflaged objects that are difficult to discern with the naked eye. To tackle the above issues, we endow SAM with keen eyes and propose the Temporal-spatial Prompt SAM (TSP-SAM), a novel approach tailored for VCOD via an ingenious prompted learning scheme. Firstly, motion-driven self-prompt learning is employed to capture the camouflaged object, thereby bypassing the need for user-provided prompts. With the detected subtle motion cues across consecutive video frames, the overall movement of the camouflaged object is captured for more precise spatial localization. Subsequently, to eliminate the prompt bias resulting from inter-frame discontinuities, the long-range consistency within the video sequences is taken into account to promote the robustness of the self-prompts. It is also injected into the encoder of SAM to enhance the representational capabilities. Extensive experimental results on two benchmarks demonstrate that the proposed TSP-SAM achieves a significant improvement over the state-of-the-art methods. With the mIoU metric increasing by $7.8\\%$ and $9.6\\%$ , TSP-SAM emerges as a groundbreaking step forward in the field of VCOD.", + "bbox": [ + 75, + 340, + 472, + 750 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 777, + 207, + 792 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Camouflaged object detection, a pivotal task in computer vision, aims to identify the target object seamlessly blending with its surroundings in images [19, 44]. This task plays an important role in various fields, including surveillance and security [29], wildlife protection [28], and medi", + "bbox": [ + 75, + 801, + 468, + 878 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/40cbf0bea7fdfbc6bbbde8bd94b8f24a9255d95c1bc72d1a3a054d7cb961ea74.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 506, + 306, + 888, + 385 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/476016f2f056582f85d7851cf2e0cfbe1301c4e26b916b14a9285870bd52ad42.jpg", + "image_caption": [ + "Figure 1. Illustration of previous VCOD methods (a-b) and the proposed TSP-SAM (c). In contrast to previous methods with two-stage architecture, the proposed TSP-SAM is trained in an end-to-end manner, incorporating short-long term temporal-spatial relationships to collaboratively learn the reliable prompts. Meanwhile, the long-range consistency is also used to enhance the representational capability of SAM." + ], + "image_footnote": [], + "bbox": [ + 504, + 388, + 887, + 508 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "cal image processing [11, 41]. In recent years, single-image camouflaged object detection has made significant progress [16, 19, 45]. However, studies [26, 43] have discerned that observers exhibit heightened perception towards the camouflaged object in motion compared to the static scenarios. This observation has sparked a growing interest in video camouflaged object detection (VCOD) [5].", + "bbox": [ + 496, + 628, + 890, + 733 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Existing VCOD methods [1, 2, 5, 26] commonly employ a two-step framework to model the inter-frame correlations implicitly or explicitly for camouflage breaking, as shown in Fig. 1(a)(b). Despite yielding promising results, they are noticeably limited in certain aspects:", + "bbox": [ + 496, + 734, + 890, + 809 + ], + "page_idx": 0 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- (i) Primarily, their two-stage architecture exposes them to cumulative errors, thus impacting their overall accuracy.", + "- (ii) Moreover, they exhibit weak generalization ability due to the limited training data." + ], + "bbox": [ + 496, + 810, + 890, + 869 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Consequently, there is a growing demand for universal models in video camouflaged object detection.", + "bbox": [ + 496, + 869, + 890, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 807, + 44 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author", + "bbox": [ + 94, + 887, + 220, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "19058", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "With the emergence and development of the visual foundation model, the segment anything model (SAM) [25] has showcased unprecedented performance in the field of natural image segmentation. This accomplishment is attributed to the impressive representations learned on 11 million images in SA-1B [25] and 636 million parameters of ViT-H [8]. SAM not only exhibits heightened versatility in terms of model capacity, but also potentially yields more consistent results across different tasks [30]. Inspired by the remarkable generality of SAM, this work endeavors to make SAM suitable for VCOD.", + "bbox": [ + 75, + 90, + 472, + 256 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "However, it is worth noting that the performance of SAM is uncertain when applied to VCOD due to the differences between the camouflaged images and natural images. Specifically, SAM, as a prompt-driven foundational model, requires extra prompts when segmenting specific regions. Yet, the seamless blending of the camouflaged object with its surroundings renders it imperceptible to the naked eye, posing a formidable challenge in providing reliable prompts. Hence, there arises an urgent need for adaptive prompt learning to the camouflaged object.", + "bbox": [ + 75, + 258, + 472, + 409 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "It is well-known that the temporal-spatial interplay in video sequences plays a pivotal role in video-related tasks. Short-term temporal-spatial relationships accentuate the variations between frames, while long-term temporal-spatial correlations model the temporal-spatial context to characterize the dependencies between the target object and the background. Motivated by these considerations, in this work, we develop a flexible end-to-end temporal-spatial prompt SAM, called TSP-SAM, for VCOD. To the best of our knowledge, the proposed TSP-SAM is the first SAM-based camouflaged object detection framework.", + "bbox": [ + 75, + 411, + 472, + 575 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our main contributions can be summarized as follows:", + "bbox": [ + 96, + 578, + 457, + 592 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Instead of user-provided prompts, we propose motion-driven self-prompt learning to capture the camouflaged object. It perceives the overall motion of the camouflaged object by establishing inter-frame associations in the frequency domain, facilitating its spatial identification.", + "- To eliminate the prompt bias stemming from inter-frame discontinuities, robust prompt learning based on long-range consistency is proposed. By modeling the long-range temporal-spatial dependencies within video sequences, the robustness of self-prompts is well promoted.", + "- In order to enhance the representational capability of SAM, we propose temporal-spatial injection that incorporates the long-range temporal-spatial consistency into the encoder of SAM. This enhancement contributes to more precise detection within the SAM framework.", + "- Extensive experiments on MoCA-Mask and CAD2016 datasets demonstrate that the proposed TSP-SAM achieves substantial performance improvements over the state-of-the-art method, with mIoU metric increasing by $7.8\\%$ and $9.6\\%$ ." + ], + "bbox": [ + 76, + 599, + 468, + 898 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related work", + "text_level": 1, + "bbox": [ + 500, + 89, + 637, + 104 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Camouflaged object detection", + "text_level": 1, + "bbox": [ + 500, + 114, + 764, + 131 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Camouflaged object detection (COD) aims to identify objects blending into their surroundings in images. Existing CNN-based methods address this task with various strategies. Fan et al. [10, 12] introduced a two-stage process, involving a search stage for localization followed by a segmentation stage for refinement. In multi-task learning, Yang et al. [43] and Ji et al. [21] incorporated an auxiliary task to derive shared context representations for enhancing the COD task. Other strategies, such as frequency cues, have also been explored. These works [6, 45] combined the static frequency spectrum with the spatial representations to discern the subtle differences between the camouflaged object and its surroundings. In contrast to the static frequency perception, our TSP-SAM perceives the frequency energy variants to depict the implicit motion between frames, facilitating the spatial identification of the camouflaged object.", + "bbox": [ + 496, + 137, + 890, + 380 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Video camouflaged object detection", + "text_level": 1, + "bbox": [ + 500, + 390, + 808, + 406 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Compared to the static appearance of images, the motion between video frames is deemed a significant clue for breaking camouflage. Existing video camouflaged object detection methods [1, 2, 5, 26] implicitly or explicitly modeled the temporal-spatial correlations through a two-stage architecture to detect camouflaged objects. Explicit motion-based methods [1, 2, 26] segmented the optical flow field after compensation or registration into object and background. In contrast, implicit feature correspondences between frames are learned by the neural network for an effective alignment [5]. Nonetheless, their two-stage architecture leads to cumulative errors. Different from the two-stage architecture, our TSP-SAM is trained in an end-to-end manner, incorporating the short-long term temporal-spatial relationships to collaboratively learn reliable prompts for SAM.", + "bbox": [ + 496, + 412, + 890, + 640 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.3. Segment anything model", + "text_level": 1, + "bbox": [ + 500, + 650, + 725, + 666 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The segment anything model (SAM) [25], a prompt-driven large-scale foundation model, has exhibited unprecedented performance in natural image segmentation. The core components of SAM comprise an image encoder, a flexible prompt encoder and a lightweight mask decoder. Despite exhibiting excellent zero-shot segmentation capability, SAM performs unstable in application-specific tasks [4]. It fails to segment medical images [18], camouflaged objects [37] and concealed scenes [22]. To apply SAM to medical images, MedSAM [30] and SAM-Adapter [4, 40] incorporated specific domain knowledge into SAM. Yet, up to now, there has been no work on adapting SAM to be tailored for camouflaged object detection. To fill this gap, TSP-SAM explores the temporal-spatial relationships, presenting the first SAM-based camouflaged object detection framework.", + "bbox": [ + 496, + 674, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "19059", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/22814ad2fb56a81c74495463aadf4300e0bee85d21d9f69c6e6b719162192d03.jpg", + "image_caption": [ + "Figure 2. The framework of the proposed TSP-SAM. In motion-driven self-prompt learning (b), the implicit motion between consecutive frames is perceived to facilitate the spatial identification of the camouflaged object, thereby learning the dense self-prompt $P_{d}(t)$ and sparse self-prompt $P_{s}(t)$ . Subsequently, in robust prompt learning (c), the long-range consistency $X_{c}(t)$ is modeled to refine the sparse self-prompt $P_{s}(t)$ . Moreover, in temporal-spatial injection (d), the long-range consistency $X_{c}(t)$ is injected into the image embedding $X_{e}(t)$ of SAM for representation enhancement. During the training phase, SAM (a) is frozen, including the image encoder $E_{i}$ , prompt encoder $E_{p}$ and the mask decoder $E_{m}$ ." + ], + "image_footnote": [], + "bbox": [ + 166, + 90, + 803, + 311 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Overview of the proposed TSP-SAM", + "text_level": 1, + "bbox": [ + 76, + 414, + 406, + 431 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Notations", + "text_level": 1, + "bbox": [ + 76, + 440, + 187, + 455 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In the subsequent sections, we establish the key notations for clarity and consistency in our discourse. Given a sequence of camouflaged images, denote by $I(t) \\in \\mathbb{R}^{H \\times W \\times 3}$ the $t$ -th frame and $I^w(t) = [I(t - T + 1), \\dots, I(t)] \\in \\mathbb{R}^{T \\times H \\times W \\times 3}$ the corresponding windowed sequence containing $T$ frames. For a tensor $A \\in \\mathbb{R}^{H \\times W \\times C}$ , $\\mathrm{Mat}(A) \\in \\mathbb{R}^{HW \\times C}$ denotes its matrixization by spatial-wise concatenation. For a matrix $B \\in \\mathbb{R}^{HW \\times C}$ , $\\mathrm{Ten}(B) \\in \\mathbb{R}^{H \\times W \\times C}$ represents its tensorization by spatial-wise division.", + "bbox": [ + 75, + 465, + 468, + 602 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Framework", + "text_level": 1, + "bbox": [ + 76, + 614, + 205, + 628 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Fig. 2 illustrates the framework of the proposed TSP-SAM. The core components consist mainly of the following three modules:", + "bbox": [ + 75, + 638, + 468, + 683 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "(a) Motion-driven self-prompt learning. The motion-driven self-prompt learning is to use the implicit inter-frame motion $\\hat{X}_{(t - 1)\\rightarrow t}$ in the frequency domain to facilitate the spatial identification of the camouflaged object, thereby learning the self-prompts for SAM.", + "(b) Robust prompt learning based on long-range consistency. To eliminate the prompt bias stemming from underlying inter-frame discontinuities, the long-range temporal-spatial consistency $X_{c}(t)$ in video sequences is modeled to promote the robustness of the self-prompts.", + "(c) Temporal-spatial injection for representation enhancement. To enhance the representational capabilities of SAM, the long-range temporal-spatial consistency $X_{c}(t)$ is injected into the image embedding $X_{e}(t)$ of SAM, con" + ], + "bbox": [ + 75, + 686, + 468, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "attributing to more precise detection.", + "bbox": [ + 500, + 416, + 730, + 431 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4. Methodology", + "text_level": 1, + "bbox": [ + 500, + 446, + 633, + 463 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we detail the proposed TSP-SAM method for video camouflaged object detection.", + "bbox": [ + 498, + 470, + 890, + 502 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4.1. Motion-driven self-prompt learning", + "text_level": 1, + "bbox": [ + 498, + 513, + 808, + 530 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "It is well-known that SAM requires user-provided prompts when segmenting specific regions. Nevertheless, the difficulty in distinguishing the camouflaged object with the naked eye makes the requirement of providing visual prompts a significant challenge. Consequently, a compelling need arises to devise a prompt learning network, capable of adaptively identifying the camouflaged object.", + "bbox": [ + 496, + 537, + 890, + 643 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4.1.1 Frequency-based motion perception", + "text_level": 1, + "bbox": [ + 498, + 664, + 803, + 680 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Inspired by the two-stream hypothesis in neuroscience [14], which highlights the importance of motion in visual perception, we consider the motion as a crucial clue for breaking camouflage. Building on this premise, several studies [7, 23] indicate the motion between frames can be expressed exhaustively in the frequency domain through the variants in frequency energy. Hence, frequency-based motion perception is designed to facilitate the spatial identification of the camouflaged object.", + "bbox": [ + 496, + 688, + 890, + 824 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Specifically, as shown in Fig. 3, given consecutive video frames $I(t - 1) \\in \\mathbb{R}^{H \\times W \\times 3}$ and $I(t) \\in \\mathbb{R}^{H \\times W \\times 3}$ , we first map them into the frequency domain. Following [45], the image features are divided into a set of $s \\times s$ patches and each patch is processed by Discrete Cosine Transform", + "bbox": [ + 496, + 825, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "19060", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/b4f7591e9f45f69f29836617da8650d78e423ebde20e8c56290f26a1ed158b86.jpg", + "image_caption": [ + "Figure 3. The framework of the motion-driven self-prompt learning. The spatial representations $\\{X_{b}(t)\\}_{b = 1}^{N_{b}}$ are extracted by PVT [39], where $N_{b}$ is the number of attention blocks." + ], + "image_footnote": [], + "bbox": [ + 83, + 92, + 459, + 250 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "(DCT) into frequency spectrum, where $s$ is the size of patches. To group all components of the same frequency into one channel, we flatten and reshape the spectrum of these patches, forming the frequency features $\\hat{I}(t - 1) \\in \\mathbb{R}^{\\frac{H}{s} \\times \\frac{W}{s} \\times 3s^2}$ and $\\hat{I}(t) \\in \\mathbb{R}^{\\frac{H}{s} \\times \\frac{W}{s} \\times 3s^2}$ .", + "bbox": [ + 75, + 316, + 470, + 395 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In order to perceive the overall motion of the camouflaged object, the inter-frame associations are established on high frequency and low frequency bands respectively. Specifically, the frequency spectrum $\\{\\hat{I} (t - 1),\\hat{I} (t)\\}$ is divided into two groups in the channel dimension, namely low frequency bands $\\{\\hat{I}^l (t - 1),\\hat{I}^l (t)\\}$ and high frequency bands $\\{\\hat{I}^h (t - 1),\\hat{I}^h (t)\\}$ . Hence, we have:", + "bbox": [ + 76, + 397, + 468, + 503 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {X} _ {(t - 1) \\rightarrow t} ^ {l} = \\mathrm {M L P} \\big (\\operatorname {A T T N} (\\operatorname {M a t} (\\hat {I} ^ {l} (t - 1)), \\operatorname {M a t} (\\hat {I} ^ {l} (t))) \\big) \\quad (1)\n$$\n", + "text_format": "latex", + "bbox": [ + 89, + 515, + 468, + 534 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {X} _ {(t - 1) \\rightarrow t} ^ {h} = \\operatorname {M L P} (\\operatorname {A T T N} (\\operatorname {M a t} (\\hat {I} ^ {h} (t - 1)), \\operatorname {M a t} (\\hat {I} ^ {h} (t)))) \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 86, + 549, + 468, + 566 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\hat{X}_{(t - 1)\\rightarrow t}^{l}\\in \\mathbb{R}^{\\frac{HW}{s^2}\\times \\frac{3s^2}{2}}$ and $\\mathrm{MLP}(\\cdot)$ is the linear layer. $\\mathrm{ATTN}(\\cdot ,\\cdot)$ is multi-head cross attention [38] to extract the inter-frame correspondences, in which the $(t - 1)$ -th frame is taken as query and the $t$ -th frame serves as key and value. Note that Eq.1 tends to highlight the global deformations of the camouflaged object by characterizing the feature correspondences on the low-frequency bands. Meanwhile, Eq.2 implicitly reveals the local edge motion of the camouflaged object on high-frequency bands.", + "bbox": [ + 75, + 574, + 468, + 717 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Subsequently, the complete implicit motion in the frequency domain is obtained by:", + "bbox": [ + 75, + 719, + 468, + 750 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {X} _ {(t - 1) \\rightarrow t} = \\hat {I} (t) + \\operatorname {C B A} (\\operatorname {C a t} \\left(\\operatorname {T e n} \\left(\\hat {X} _ {(t - 1) \\rightarrow t} ^ {l}\\right), \\operatorname {T e n} \\left(\\hat {X} _ {(t - 1) \\rightarrow t} ^ {h}\\right)\\right)) \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 78, + 760, + 468, + 791 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathrm{CBA}(\\cdot)$ is the convolution operation followed by batch normalization and activation and $\\operatorname{Cat}(\\cdot)$ is the channel-wise concatenation. Importantly, $\\hat{X}_{(t - 1)\\rightarrow t}\\in$ $\\mathbb{R}^{\\frac{H}{s}\\times \\frac{W}{s}\\times 3s^{2}}$ implicitly depicts the overall motion of the camouflaged object by perceiving the variants in the frequency energy, enabling to facilitate the discovery of the camouflaged object.", + "bbox": [ + 75, + 791, + 468, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1.2 Implicit motion-induced spatial-frequency fusion", + "text_level": 1, + "bbox": [ + 500, + 90, + 890, + 119 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To translate the perceived motion into the spatial domain, the implicit motion-induced spatial-frequency representation fusion is established. This process involves the integration between the implicit motion $\\hat{X}_{(t - 1)\\rightarrow t}$ in the frequency domain and the spatial representations $X_{b}(t)\\in$ $\\mathbb{R}^{\\frac{H}{2^{b + 1}}\\times \\frac{W}{2^{b + 1}}\\times C},b\\in \\{1,\\dots ,N_b\\}$ extracted by PVT [39], where $N_{b}$ is the number of attention blocks. It can be expressed as follows:", + "bbox": [ + 498, + 125, + 890, + 238 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {X} _ {b} (t) = X _ {b} (t) + \\operatorname {T e n} (\\operatorname {A T T N} (\\operatorname {M a t} (\\operatorname {C B A} (\\hat {X} _ {(t - 1) \\rightarrow t})), \\operatorname {M a t} (X _ {b} (t)))) \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 506, + 247, + 890, + 263 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\hat{X}_b(t) \\in \\mathbb{R}^{\\frac{H}{2^b + 1} \\times \\frac{W}{2^b + 1} \\times C}$ . Note that the spatial perception of the camouflaged object could be strengthened by establishing cross attention between the implicit motion as query and the spatial representation as key and value.", + "bbox": [ + 498, + 271, + 890, + 335 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1.3 Self-prompt learning", + "text_level": 1, + "bbox": [ + 498, + 354, + 700, + 369 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Based on the co-representations $\\{\\hat{X}_b(t)\\}_{b=1}^{N_b}$ , two prompt heads are adopted to generate the dense self-prompt $P_d(t)$ and sparse self-prompt $P_s(t)$ , respectively. Specifically, we adopt the decoder of SLTNet [5] as dense prompt head $E_d$ as follows:", + "bbox": [ + 496, + 376, + 890, + 450 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nP _ {d} (t) = E _ {d} \\left(\\left\\{\\hat {X} _ {b} (t) \\right\\} _ {b = 1} ^ {N _ {b}}\\right) \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 607, + 450, + 890, + 470 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $P_{d}(t)$ is the dense self-prompt. Meanwhile, as illustrated in Fig. 3, to obtain the sparse prompt, we adopt the residual block of Resnet [17] to process the co-representations in a bottom-up manner:", + "bbox": [ + 496, + 474, + 890, + 534 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {X} _ {b} (t) = \\left\\{ \\begin{array}{l l} \\operatorname {R e s b l o c k} \\left(\\hat {X} _ {b} (t)\\right) & b = N _ {b} \\\\ \\operatorname {R e s b l o c k} \\left(\\hat {X} _ {b} (t) + \\tilde {X} _ {b + 1} (t)\\right) & b \\in \\{1, \\dots , N _ {b} - 1 \\} \\end{array} \\right. \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 509, + 542, + 890, + 590 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Finally, the sparse self-prompt is represented as follows:", + "bbox": [ + 500, + 590, + 870, + 606 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nP _ {s} (t) = \\operatorname {A v g p o o l} \\left(\\operatorname {C B A} \\left(\\tilde {X} _ {1} (t)\\right)\\right) \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 580, + 614, + 890, + 633 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathrm{Avgpool}(\\cdot)$ is global average pooling operation.", + "bbox": [ + 500, + 642, + 861, + 657 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.2. Robust prompt learning based on long-range consistency", + "text_level": 1, + "bbox": [ + 500, + 666, + 890, + 696 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The self-prompts of the camouflaged object are well learned by using the perceived inter-frame motion to facilitate its spatial identification. However, the susceptibility of inter-frame motion to underlying temporal-spatial discontinuities, such as occlusions, shaking, scene changes, etc., brings a potential lack of reliability in the learned self-prompts. It is well known that the long-range dependencies within the video sequences enable to characterize the temporal-spatial context to cope with the short-term temporal-spatial discontinuities. Hence, it is necessary to explore the long-range temporal-spatial consistency within the video sequences to improve the robustness of the self-prompts, which likes endowing SAM with keen eyes.", + "bbox": [ + 496, + 703, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "19061", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.2.1 Long-range consistency squeezing", + "text_level": 1, + "bbox": [ + 76, + 90, + 369, + 107 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As illustrated in Fig. 2(c), given the windowed sequence $I^{w}(t) \\in \\mathbb{R}^{T \\times H \\times W \\times 3}$ corresponding to the current frame $I(t)$ , we adopt a 3D convolution neural network $E_{t}$ to extract the long-range temporal-spatial relationships:", + "bbox": [ + 76, + 114, + 468, + 175 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nX ^ {w} (t) = E _ {t} \\left(I ^ {w} (t)\\right) \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 202, + 185, + 468, + 202 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $X^w(t) \\in \\mathbb{R}^{T \\times \\frac{H}{4} \\times \\frac{W}{4} \\times C}$ fully reveals the temporal-spatial dependencies within the video sequences. In our work, we adopt 3D Resnet18 [15] as $E_t$ .", + "bbox": [ + 76, + 210, + 468, + 258 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Subsequently, to further extract the long-range temporal-spatial consistency of the camouflaged object, we take the camouflage representation $X_{e}(t)$ from the encoder of SAM as query and the long-range temporal-spatial dependencies $X^{w}(t)$ averaged along the temporal dimension as key and value to establish the cross attention:", + "bbox": [ + 76, + 260, + 468, + 349 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nX _ {c} (t) = \\operatorname {A T T N} \\left(\\operatorname {M a t} \\left(X _ {e} (t)\\right), \\operatorname {M a t} \\left(\\operatorname {A v g} _ {\\mathrm {t}} \\left(X ^ {w} (t)\\right)\\right)\\right) \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 89, + 359, + 468, + 378 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathrm{Avg}_{\\mathrm{t}}(\\cdot)$ is an average operation to squeeze the temporal dimension. $X_{c}(t)\\in \\mathbb{R}^{\\frac{HW}{16}\\times C}$ characterizes the long-range consistency specific to the camouflaged object.", + "bbox": [ + 76, + 387, + 468, + 434 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2.2 Robust prompt learning", + "text_level": 1, + "bbox": [ + 76, + 452, + 300, + 468 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To enhance the robustness of the self-prompts to inter-frame temporal-spatial discontinuities, the long-range temporal-spatial consistency $X_{c}(t)$ should be induced into the prompt learning. Specifically, in order to map the long-range consistency into the coordinate space, a learnable query $Q_{c}(t) \\in \\mathbb{R}^{4 \\times C}$ is defined to establish cross attention with $X_{c}(t)$ . It is formulated as:", + "bbox": [ + 76, + 476, + 468, + 580 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nP _ {r} (t) = \\operatorname {A v g} _ {\\mathrm {c}} (\\operatorname {A T N} \\left(Q _ {c} (t), X _ {c} (t)\\right)) \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 145, + 592, + 468, + 609 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathrm{Avg}_c(\\cdot)$ is an average operation along the channel dimension. $P_r(t)\\in \\mathbf{R}^4$ is a refinement term derived from long-range consistency to correct the sparse self-prompt:", + "bbox": [ + 76, + 619, + 468, + 665 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nP _ {s} ^ {r} (t) = P _ {s} (t) + P _ {r} (t) \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 194, + 675, + 468, + 691 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Through incorporating long-range consistency into the prompt learning, the visual prompts could be more robust to the temporal-spatial discontinuities, which means endowing SAM with a pair of keen eyes.", + "bbox": [ + 76, + 702, + 468, + 763 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.3. Temporal-spatial injection for representation enhancement", + "text_level": 1, + "bbox": [ + 76, + 771, + 468, + 801 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "It is known that the remarkable performance of SAM relies not only on well learned robust prompts but also on the representational capability of the image encoder. However, due to the fact that the image embedding of SAM lacks the temporal-spatial information, the final predictions of consecutive frames may be inconsistent. Hence, it is necessary", + "bbox": [ + 76, + 809, + 468, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "to inject the long-range temporal-spatial consistency into SAM. As shown in Fig. 2(d), it is represented as follows:", + "bbox": [ + 498, + 90, + 890, + 121 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {X} _ {e} (t) = X _ {e} (t) + \\operatorname {T e n} \\left(X _ {c} (t)\\right) \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 594, + 131, + 890, + 148 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $X_{c}(t)$ is the long-range consistency from Eq. 9. In this way, based on the enhanced image representations and the robust prompts called keen eyes, SAM achieves more precise camouflaged object detection in videos.", + "bbox": [ + 498, + 160, + 890, + 220 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.4. Model optimization", + "text_level": 1, + "bbox": [ + 498, + 229, + 687, + 244 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We train the proposed TSP-SAM in an end-to-end manner by minimizing the joint loss below:", + "bbox": [ + 498, + 253, + 890, + 282 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} (t) = \\mathcal {L} _ {S A M} (t) + \\alpha \\mathcal {L} _ {p} ^ {s} (t) + \\beta \\mathcal {L} _ {p} ^ {d} (t) \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 566, + 292, + 890, + 311 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathcal{L}_{SAM},\\mathcal{L}_p^s$ and $\\mathcal{L}_p^d$ supervise the final prediction, the sparse prompt and the dense prompt, respectively.", + "bbox": [ + 498, + 323, + 890, + 353 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Following [3], the supervision on the sparse prompt $\\mathcal{L}_p^s (t)$ is as follows:", + "bbox": [ + 500, + 354, + 890, + 386 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {p} ^ {s} (t) = \\operatorname {B C E} \\left(P _ {s} ^ {r} (t), G (t)\\right) + \\operatorname {M S E} \\left(P _ {s} ^ {r} (t), G (t)\\right) \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 514, + 393, + 890, + 412 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where BCE is binary cross-entropy loss to supervise the coordinates and MSE is mean square error to supervise the size of the box prompt or the scale of the point prompt. For the dense prompt loss $\\mathcal{L}_p^d$ and the final prediction loss $\\mathcal{L}_{SAM}$ , we adopt the hybrid loss [13] for supervision:", + "bbox": [ + 498, + 422, + 890, + 500 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {h y b r i d}} = \\mathcal {L} _ {c e} ^ {w} + \\mathcal {L} _ {i o u} ^ {w} \\tag {15}\n$$\n", + "text_format": "latex", + "bbox": [ + 617, + 508, + 890, + 527 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathcal{L}_{ce}^{w}$ is the weighted cross-entropy loss and $\\mathcal{L}_{iou}^{w}$ is the weighted intersection-over-union loss.", + "bbox": [ + 498, + 537, + 890, + 566 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5. Experiment results and analysis", + "text_level": 1, + "bbox": [ + 500, + 580, + 790, + 598 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.1. Experiment settings", + "text_level": 1, + "bbox": [ + 500, + 604, + 689, + 622 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Datasets. Following previous methods [5], we have conducted a comprehensive evaluation on two widely used datasets: MoCA-Mask [5] and CAD2016 [1]. The MoCA-Mask dataset [5] is a video camouflaged object detection dataset, covering camouflaged animals moving in natural scenes. It contains 87 video sequences (22939 frames), of which 71 video sequences are used for training and 16 sequences for testing. It is well annotated with bounding boxes and pixel-level segmentation masks on every fifth frames. Meanwhile, Camouflaged Animal Dataset [1] (CAD2016) is a small video camouflaged object detection dataset extracted from YouTube videos. It consists of nine short video clips in total and is well annotated with pixel-level semantic masks on every fifth frames.", + "bbox": [ + 498, + 628, + 890, + 839 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Metrics. During the testing phase, we assess the prediction masks using six widely-used metrics: S-measure $(S_{\\alpha})$ [9], Weighted F-measure $(F_{\\beta}^{w})$ [31], Enhanced-alignment measure $(E_{\\phi})$ [13], Mean absolute error (MAE, $\\mathcal{M}$ ) [33],", + "bbox": [ + 498, + 839, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "19062", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/6f61b7800eef62e45a9a04519775931ef44fec0640325812467beec291ef2ecb.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodYearInputMoCA-MaskCAD2016
\\(S_{\\alpha}\\uparrow\\)\\(F^{\\text{w}}_{\\beta}\\uparrow\\)\\(E_{\\phi}\\uparrow\\)\\(\\mathcal{M}\\downarrow\\)mDice↑mIoU↑\\(S_{\\alpha}\\uparrow\\)\\(F^{\\text{w}}_{\\beta}\\uparrow\\)\\(E_{\\phi}\\uparrow\\)\\(\\mathcal{M}\\downarrow\\)mDice↑mIoU↑
SINet[10]2020-CVPRImage0.5740.1850.6550.0300.2210.1560.6010.2040.5890.0890.2890.209
SINet-v2[12]2021-TPAMIImage0.5710.1750.6080.0350.2110.1530.5440.1810.5460.0490.1700.110
ZoomNet[32]2022-CVPRImage0.5820.2110.5360.0330.2240.1670.5870.2250.5940.0630.2460.166
BGNet [36]2022-IJCAIImage0.5900.2030.6470.0230.2250.1670.6070.2030.6660.0890.3450.256
FEDERNet[16]2023-CVPRImage0.5550.1580.5420.0490.1920.1320.6070.2460.7250.0610.3610.257
FSPNet[19]2023-CVPRImage0.5940.1820.6080.0440.2380.1670.5390.2200.5530.1450.3090.212
PUENet[44]2023-TIPImage0.5940.2040.6190.0370.3020.2120.6730.4270.8030.0340.4990.389
RCRNet[42]2019-ICCVVideo0.5970.1740.5830.0250.1940.137------
PNS-Net[20]2021-MICCAIVideo0.5760.1340.5620.0380.1890.1330.6780.3690.7200.0430.4090.308
MG[43]2021-ICCVVideo0.5470.1650.5370.0950.1970.1410.4840.3140.5580.3700.3510.260
SLT-Net[5]2022-CVPRVideo0.6560.3570.7850.0210.3870.3100.6790.4200.8050.0330.4450.342
TSP-SAM(M+P)OursVideo0.6730.4000.7660.0120.4210.3450.6810.5000.8530.0310.4960.393
TSP-SAM(M+B)OursVideo0.6890.4440.8080.0080.4580.3880.7040.5240.9120.0280.5430.438
", + "bbox": [ + 80, + 89, + 893, + 295 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/a08b6536c88c2b1cd7a925b01bf31e5cd1433a7d66458caef1023c6bd65e6201.jpg", + "table_caption": [ + "Table 1. Quantitative comparisons over two benchmark datasets. The top three results are highlighted in red, green, and blue. M+P: Combination of mask and point prompts. M+B: Combination of mask and box prompts." + ], + "table_footnote": [], + "table_body": "
MethodSα↑Fwβ↑Eφ↑M↓mDice↑mIoU↑
SINet[10]0.6360.3460.7750.0410.3810.283
SINet-v2[12]0.6530.3820.7620.0390.4130.318
ZoomNet[32]0.6330.3490.6010.0330.3490.273
FEDERNet[16]0.5730.2720.6090.0510.2600.199
FSPNet[19]0.6810.4010.7160.0440.4460.332
PUENet[44]0.6910.4850.7950.0340.5140.396
RCRNet[42]0.6270.2870.6660.0480.3090.229
PNS-Net[20]0.6550.3250.6730.0480.3840.290
MG[43]0.5940.3360.6910.0590.3680.268
SLT-Net[5]0.6960.4810.8450.0300.4930.401
TSP-SAM(M+P)0.7050.5650.8360.0270.5310.422
TSP-SAM(M+B)0.7510.6280.8510.0210.6030.496
", + "bbox": [ + 80, + 340, + 467, + 527 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "meanDice (mDice), meanIoU (mIoU). Overall, a better VCOD method has larger $S_{\\alpha}$ , $F_{\\beta}^{w}$ , $E_{\\phi}$ , mDice and mIoU scores, but a smaller MAE score.", + "bbox": [ + 75, + 607, + 468, + 654 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation details. In our experiments, we freeze SAM and train our TSP-SAM in an end-to-end manner with a maximum of 20 epochs. Specifically, for a fair comparison, both the training and testing images are resize to $352 \\times 352$ . We leverage Adam [24] as our optimizer and set the weight decay to $1 \\times 10^{-5}$ . On the MoCA-Mask dataset, the learning rate is initialized to $1 \\times 10^{-5}$ for the spatial extractor of motion-driven self-prompt learning and $5 \\times 10^{-5}$ for other trainable parameters. After 10 epochs, the learning rate is reduced by $90\\%$ . On the CAD2016 dataset, we adopt a ratio of 9:1 to divide all classes into training and testing sets and report the mean value over the testing set. We set the initial learning rate to $1 \\times 10^{-4}$ for the spatial extractor of the motion-driven self-prompt learning and $5 \\times 10^{-4}$ for other trainable parameters. The other settings are the same as those on MoCA-Mask dataset.", + "bbox": [ + 75, + 659, + 472, + 900 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/f8ecbcfbb9f2b40bd20d2c5c257798c91d9e85d4c96bea791c75413e984f99c7.jpg", + "table_caption": [ + "Table 2. Generalization comparisons on CAD2016 dataset. The best results are highlighted in **bold**. M+P: Combination of mask and point prompts. M+B: Combination of mask and box prompts." + ], + "table_footnote": [], + "table_body": "
MethodTuning parameters (M)MethodTime (s) ↓Speed (fps) ↑
FSPNet274.24FSPNet568.701.31
SLT-Net253.402.94
SLT-Net82.41Overall (ours)294.992.53
TSP-SAM(ours)89.78Prompt learning73.2810.17
SAM221.713.36
", + "bbox": [ + 519, + 340, + 883, + 407 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 3. The number of Table 4. Comparisons on infertuning parameters. Hence time and speed.", + "bbox": [ + 517, + 412, + 892, + 443 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.2. Performance comparison", + "text_level": 1, + "bbox": [ + 500, + 463, + 728, + 479 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "1) Baselines: To evaluate the performance of the proposed TSP-SAM, we compare it with a range of state-of-the-art methods. These methods fall into two categories:", + "bbox": [ + 500, + 488, + 892, + 534 + ], + "page_idx": 5 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Single-image camouflaged object detection. These methods analyze the static appearance of images to discern the subtle differences between camouflaged objects and their surroundings. The compared methods in this type include SINet [10], SINet-v2 [12], ZoomNet [32], BGNet [36], FEDERNet [16], FSPNet[19], and PUENet [44].", + "- Video object segmentation. This set of methods explores the temporal-spatial relationships between frames for the purpose of segmenting objects with specific attributes. The compared methods in this type encompass RCRNet [42], PNS-Net [20], MG[43], and SLT-Net[5].", + "2) Quantitative Results: We evaluate TSP-SAM on MoCA-Mask and CAD2016 datasets. From Table 1, we can conclude that, (i) compared to the combination of mask and point prompts, the combination of mask and box prompts is more reliable for SAM. This is attributed to the limitations of the point prompt in conveying the boundary information. In contrast, bounding box excels in indicating the boundaries of the camouflaged object, even in the presence of bias. (ii) the strong contrast between the performance of TSP-SAM and single-image camouflaged object detection methods points to the importance of temporal-spatial relationships in breaking camouflage. (iii) the proposed TSP-SAM" + ], + "bbox": [ + 496, + 536, + 893, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "19063", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/8da5257c5b97459ed1fb85348d80fbdc4e58f925855d13587a69f56d3e9176f9.jpg", + "image_caption": [ + "Figure 4. Visualization of our proposed TSP-SAM and baseline methods on MoCA-Mask and CAD2016 dataset. From left to right: frames (1st column), ground truth (2nd column), prediction of our TSP-SAM with mask and box prompts (3rd column), prediction of our TSP-SAM with mask and point prompts (4th column), and predictions of compared methods (5th-7th columns)." + ], + "image_footnote": [], + "bbox": [ + 83, + 88, + 480, + 263 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a17edc19376520f40cf5d47563738915e5fab56f4b40ad5a8ccd7a594911b2da.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 488, + 88, + 883, + 263 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a3945ae3ac33cebc2a3871b4ea4bac3421814b9aa88d51fddf8307c22f32e3ca.jpg", + "image_caption": [ + "Figure 5. Visualization of several failure cases with mask and point prompts. From left to right: frames (1st column), ground truth (2nd column), mask and box prompts (3rd column), prediction with mask and box prompts (4th column), mask and point prompts (5th column), and prediction with mask and point prompts (6th column)." + ], + "image_footnote": [], + "bbox": [ + 78, + 325, + 467, + 512 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "surpasses all video object segmentation methods. Notably, TSP-SAM exhibits a $1.3\\% - 8.7\\%$ improvement over SLT-Net [5] on MoCA-Mask dataset, meaning that the temporal-spatial exploration in TSP-SAM is more effective.", + "bbox": [ + 75, + 625, + 468, + 685 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To assess the generalization performance, we train on the MoCA-Mask dataset and report the test performance on the CAD2016 dataset. It can be observed from Table 2 that the proposed TSP-SAM outperforms all baselines. Compared with SLTNet [5], TSP-SAM surpasses it by $0.9\\% - 14.7\\%$ , indicating that TSP-SAM is capable of generalizing to the unseen data more effectively. Meanwhile, we also test the performance of TSP-SAM on general video object segmentation datasets. On DAVIS2016 dataset [34], our model achieves comparable performance to the SOTA SLT-Net. While on SegTrack-v2 dataset [27] with multiple objects, our method obtains $62.84\\%$ for mIoU, surpassing SLT-Net by $9.9\\%$ . Surprisingly, our model is even comparable to the recent SMTC [35] on SegTrack-v2.", + "bbox": [ + 75, + 689, + 470, + 901 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Besides, we report the number of tuning parameters, the inference time and speed. From Table 3, compared with SLT-Net, our model achieves $9.6\\%$ improvements with a slight parameter increase. From Table 4, our model is comparable to the latest video-oriented SLT-Net. Decomposing the overall inference time, SAM consumes $75\\%$ of it.", + "bbox": [ + 496, + 328, + 890, + 417 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3) Qualitative Results: To provide more intuitive evaluations of the proposed TSP-SAM, we conduct two qualitative discussions. Firstly, as shown in Fig. 4, we present the segmentation results of several examples to intuitively compare the proposed TSP-SAM with the baselines. It can be seen that: (i) some confusing surrounding regions interfere with the baseline methods, while our TSP-SAM is more robust in identifying the camouflaged object. This demonstrates the necessity of short-long term temporal-spatial relationships. (ii) compared with the baselines, our TSP-SAM exhibits superior capabilities in boundary segmentation in terms of discriminability and details.", + "bbox": [ + 496, + 420, + 892, + 601 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Secondly, to further analyze the influence of the type of sparse prompts on overall performance, we visualize the several failure cases with mask and point prompts. From Fig. 5, it can be found that: (i) since the target object is not always a regular shape, the learned point prompt is not on the target object although it is near the center. (ii) under the condition with unclear boundaries, the point prompt is more likely to segment the regions of background. (iii) the point prompt may induce the model to segment the local areas instead of the complete target object. In contrast, the box prompt can avoid the above defects and achieve more precise segmentation results. Meanwhile, the complete localization under the learned mask and box prompts verifies the effectiveness of the temporal-spatial prompt learning.", + "bbox": [ + 496, + 603, + 893, + 816 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.3. Ablation study", + "text_level": 1, + "bbox": [ + 500, + 830, + 648, + 847 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To gain a comprehensive understanding of the efficacy of core components and the impact of hyper parameters, we delve into a detailed analysis.", + "bbox": [ + 496, + 854, + 893, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "19064", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/8b32d370e610f4f65f28997bb3c7881e37e447ae2d998d5f48f26d15e12cc098.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodSα↑Fwβ↑Eφ↑mDice↑mIoU↑
SAM+SSP0.6370.3300.7680.3670.311
SAM+MSP0.6570.3530.8230.4070.340
SAM+MSP+LCP0.6650.4090.7570.4290.366
SAM+MSP+LCP+TSI0.6890.4440.8080.4580.388
", + "bbox": [ + 80, + 88, + 467, + 167 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/a45acdd249c5cf4f00330c01d64c17a9ff13c154f770b0e6a95b19428f496075.jpg", + "image_caption": [ + "(a) mDice" + ], + "image_footnote": [], + "bbox": [ + 116, + 255, + 269, + 325 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/5f384288d61922770d1c5b2a73a8eb8a14ce557e8339c87f58b367a311632653.jpg", + "image_caption": [ + "(b) $F_{\\beta}^{w}$" + ], + "image_footnote": [], + "bbox": [ + 272, + 256, + 416, + 325 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/c288c6aa9ac6ce27b30d87c153ccab97631d5fac17596de03f8236e7ba54ad47.jpg", + "image_caption": [ + "(c) mIoU" + ], + "image_footnote": [], + "bbox": [ + 117, + 339, + 269, + 409 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/57903b9ca011a475c431aaef47d5125ea9f421d5d8f79ca6ea02148fcb79ba25.jpg", + "image_caption": [ + "(d) $E_{\\phi}$" + ], + "image_footnote": [], + "bbox": [ + 272, + 339, + 415, + 409 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Ablation analysis: Table 5 present the segmentation results by progressively applying each part in our model. Compared to the spatial-based self-prompt learning (SSP) comprising a spatial extractor and two prompt heads, motion-driven self-prompt learning (MSP) improves the mDice metric from $36.7\\%$ to $40.7\\%$ to a large degree. This affirms the effectiveness of the inter-frame motion in camouflage breaking. Subsequently, the long-range consistency-based robust prompt learning (LCP) brings $2.2\\%$ gains for mDice, indicating the effectiveness of long-range consistency in improving the robustness of self-prompts. Meanwhile, the temporal-spatial injection (TSI) further boosts the mDice metric by $2.9\\%$ , once again verifying the significance of long-range consistency.", + "bbox": [ + 75, + 474, + 468, + 686 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To further analyze the effectiveness of frequency-based motion perception in MSP, we replace it with spatial-based motion perception, frequency-based static appearance perception, and spatial-based static appearance perception to conduct the ablation studies. As can be seen from Table 6, the frequency-based motion perception yields the best performance. While verifying the significant role of interframe motion in breaking camouflage, it also demonstrates that inter-frame motion can be expressed more exhaustively in the frequency domain than in the spatial domain.", + "bbox": [ + 75, + 686, + 467, + 838 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Hyper parameter analysis: To investigate the effect of the size of the windowed sequence on segmentation performance, we conduct the ablation experiments by varying $T$ from $\\{5,6,7,8,9,10\\}$ . As illustrated in Fig. 6, the ac", + "bbox": [ + 75, + 839, + 468, + 901 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/463fada25ca77e543580f62d009e0e765cb4d6ee07d616e3593fe077c1372140.jpg", + "table_caption": [ + "Table 5. Ablation studies of the core components on MoCA-Mask dataset. SSP: Spatial-based self-prompt learning. MSP: Motion-driven self-prompt learning. LCP: Long-range consistency-based robust prompt learning. TSI: Temporal-spatial injection for representation enhancement." + ], + "table_footnote": [], + "table_body": "
MethodSα↑Fwβ↑Eφ↑mDice↑mIoU↑
Spatial appearance0.6370.3300.7680.3670.311
Frequency appearance0.6470.3440.7930.3960.328
Spatial motion0.6510.3350.7300.3960.324
Frequency motion0.6570.3530.8230.4070.340
", + "bbox": [ + 501, + 88, + 888, + 167 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 6. Ablation studies of frequency-based motion perception on MoCA-Mask dataset. Spatial appearance: Spatial-based appearance perception. Frequency appearance: Frequency-based appearance perception. Spatial motion: Spatial-based motion perception. Frequency motion: Frequency-based motion perception.", + "bbox": [ + 496, + 178, + 890, + 248 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/3def31d2b60d3effa059bbf903b374e42f8ecaf6fd254574fe74b003f1087ec9.jpg", + "image_caption": [ + "(a) mDice" + ], + "image_footnote": [], + "bbox": [ + 563, + 255, + 681, + 323 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/f25d3e7d8c5bf9e59fd7d995fcf5e440d24ef30dd60c990d708cdcddb499a2bd.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 718, + 256, + 836, + 323 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/31c48953916a6a652ba587bcaf0c52f0e7665dd5a7bdd58d6ae4db2d8dc7df05.jpg", + "image_caption": [ + "Figure 6. Sensitivity analysis of the size $T$ of the windowed sequence on MoCA-Mask dataset.", + "(c) mIoU", + "Figure 7. Sensitivity analysis of the balance factor $(\\alpha, \\beta)$ in the loss function on MoCA-Mask dataset." + ], + "image_footnote": [], + "bbox": [ + 565, + 339, + 681, + 407 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/5374b09c41621c29624e7c919a04db55bca4e7388bbc10bc109c022432701e52.jpg", + "image_caption": [ + "(b) $F_{\\beta}^{w}$", + "(d) $E_{\\phi}$" + ], + "image_footnote": [], + "bbox": [ + 720, + 338, + 836, + 407 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "curacy initially rises with increasing the size of the windowed sequence, reaches the best at $T = 7$ before declining. Moreover, we assess the sensitivity of $\\alpha$ and $\\beta$ that are used to balance the sparse and dense prompt loss in Fig. 7. The optimal segmentation performance is observed at $(\\alpha, \\beta) = (1,0.6)$ .", + "bbox": [ + 496, + 468, + 890, + 560 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 571, + 617, + 587 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we propose a flexible end-to-end temporal-spatial prompt SAM (TSP-SAM) for video camouflaged object detection. Instead of user-provided prompts, motion-driven self-prompt learning is proposed to capture the overall motion of the camouflaged object, facilitating its spatial identification. Subsequently, a robust prompt learning based on long-range consistency is proposed to promote the robustness of the self-prompts, which likes endowing SAM with keen eyes. Moreover, the long-range consistency is injected into the encoder of SAM for representation enhancement. Notably, TSP-SAM provides a new perspective, that is temporal-spatial associations, for visual prompt learning.", + "bbox": [ + 496, + 597, + 890, + 779 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "7. Acknowledgement", + "text_level": 1, + "bbox": [ + 500, + 784, + 679, + 801 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This work was supported in part by the National Key Research and Development Program of China under Grant No.2021ZD0140407, Beijing Natural Science Foundation under Grant No.7222313, and National High Level Hospital Clinical Research Funding under Grant No.2022-PUMCHC-041.", + "bbox": [ + 496, + 809, + 890, + 898 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "19065", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Pia Bideau and Erik Learned-Miller. It's moving! a probabilistic model for causal motion segmentation in moving camera videos. In ECCV, pages 433-449, 2016. 1, 2, 5", + "[2] Pia Bideau, Rakesh R Menon, and Erik Learned-Miller. Moa-net: self-supervised motion segmentation. In ECCV Workshops, pages 0–0, 2018. 1, 2", + "[3] Alexey Bochkovskiy, Chien-Yao Wang, and Hong-Yuan Mark Liao. Yolov4: Optimal speed and accuracy of object detection. arXiv preprint arXiv:2004.10934, 2020. 5", + "[4] Tianrun Chen, Lanyun Zhu, Chaotao Ding, Runlong Cao, Shangzhan Zhang, Yan Wang, Zejian Li, Lingyun Sun, Papa Mao, and Ying Zang. Sam fails to segment anything?—sam-adapter: Adapting sam in underperformed scenes: Camouflage, shadow, and more. arXiv preprint arXiv:2304.09148, 2023. 2", + "[5] Xuelian Cheng, Huan Xiong, Deng-Ping Fan, Yiran Zhong, Mehrtash Harandi, Tom Drummond, and Zongyuan Ge. Implicit motion handling for video camouflaged object detection. In CVPR, pages 13864-13873, 2022. 1, 2, 4, 5, 6, 7", + "[6] Runmin Cong, Mengyao Sun, Sanyi Zhang, Xiaofei Zhou, Wei Zhang, and Yao Zhao. Frequency perception network for camouflaged object detection. In ACM MM, pages 1179-1189, 2023. 2", + "[7] Xinyi Cui, Qingshan Liu, and Dimitris Metaxas. Temporal spectral residual: fast motion saliency detection. In ACM MM, pages 617-620, 2009. 3", + "[8] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, et al. An image is worth 16x16 words: Transformers for image recognition at scale. 2020. 2", + "[9] Deng-Ping Fan, Ming-Ming Cheng, Yun Liu, Tao Li, and Ali Borji. Structure-measure: A new way to evaluate foreground maps. In ICCV, pages 4548-4557, 2017. 5", + "[10] Deng-Ping Fan, Ge-Peng Ji, Guolei Sun, Ming-Ming Cheng, Jianbing Shen, and Ling Shao. Camouflaged object detection. In CVPR, pages 2777-2787, 2020. 2, 6", + "[11] Deng-Ping Fan, Ge-Peng Ji, Tao Zhou, Geng Chen, Huazhu Fu, Jianbing Shen, and Ling Shao. Pranet: Parallel reverse attention network for polyp segmentation. In International conference on medical image computing and computer-assisted intervention, pages 263-273, 2020. 1", + "[12] Deng-Ping Fan, Ge-Peng Ji, Ming-Ming Cheng, and Ling Shao. Concealed object detection. IEEE TPAMI, 44(10): 6024-6042, 2021. 2, 6", + "[13] Deng-Ping Fan, Ge-Peng Ji, Xuebin Qin, and Ming-Ming Cheng. Cognitive vision inspired object segmentation metric and loss function. Scientia Sinica Informationis, 6(6), 2021. 5", + "[14] Melvyn A Goodale and A David Milner. Separate visual pathways for perception and action. Trends in neurosciences, 15(1):20-25, 1992. 3", + "[15] Kensho Hara, Hirokatsu Kataoka, and Yutaka Satoh. Can spatiotemporal 3d cnns retrace the history of 2d cnns and imagenet? In CVPR, pages 6546-6555, 2018. 5" + ], + "bbox": [ + 78, + 114, + 470, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[16] Chunming He, Kai Li, Yachao Zhang, Longxiang Tang, Yulin Zhang, Zhenhua Guo, and Xiu Li. Camouflaged object detection with feature decomposition and edge reconstruction. In CVPR, pages 22046-22055, 2023. 1, 6", + "[17] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, pages 770-778, 2016. 4", + "[18] Yuhao Huang, Xin Yang, Lian Liu, Han Zhou, Ao Chang, Xinrui Zhou, Rusi Chen, Junxuan Yu, Jiongquan Chen, Chaoyu Chen, et al. Segment anything model for medical images? arXiv preprint arXiv:2304.14660, 2023. 2", + "[19] Zhou Huang, Hang Dai, Tian-Zhu Xiang, Shuo Wang, Huai-Xin Chen, Jie Qin, and Huan Xiong. Feature shrinkage pyramid for camouflaged object detection with transformers. In CVPR, pages 5557-5566, 2023. 1, 6", + "[20] Ge-Peng Ji, Yu-Cheng Chou, Deng-Ping Fan, Geng Chen, Huazhu Fu, Debesh Jha, and Ling Shao. Progressively normalized self-attention network for video polyp segmentation. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 142–152, 2021. 6", + "[21] Ge-Peng Ji, Deng-Ping Fan, Yu-Cheng Chou, Dengxin Dai, Alexander Liniger, and Luc Van Gool. Deep gradient learning for efficient camouflaged object detection. Machine Intelligence Research, 20(1):92-108, 2023. 2", + "[22] Ge-Peng Ji, Deng-Ping Fan, Peng Xu, Ming-Ming Cheng, Bowen Zhou, and Luc Van Gool. \"Sam struggles in concealed scenes—empirical study on\" segment anything\". arXiv preprint arXiv:2304.06022, 2023. 2", + "[23] Yang Jia, Jie Yuan, Jinjun Wang, Jun Fang, Qixing Zhang, and Yongming Zhang. A saliency-based method for early smoke detection in video sequences. Fire technology, 52: 1271-1292, 2016. 3", + "[24] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. ICLR, 2015. 6", + "[25] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. arXiv preprint arXiv:2304.02643, 2023. 2", + "[26] Hala Lamdouar, Charig Yang, Weidi Xie, and Andrew Zisserman. Betrayed by motion: Camouflaged object discovery via motion segmentation. In ACCV, 2020. 1, 2", + "[27] Fuxin Li, Taeyoung Kim, Ahmad Humayun, David Tsai, and James M Rehg. Video segmentation by tracking many figure-ground segments. In ICCV, pages 2192-2199, 2013. 7", + "[28] Thomas Lidbetter. Search and rescue in the face of uncertain threats. European Journal of Operational Research, 285(3): 1153-1160, 2020. 1", + "[29] Ting Liu, Yao Zhao, Yunchao Wei, Yufeng Zhao, and Shikui Wei. Concealed object detection for activate millimeter wave image. IEEE Transactions on Industrial Electronics, 66(12): 9909-9917, 2019. 1", + "[30] Jun Ma, Yuting He, Feifei Li, Lin Han, Chenyu You, and Bo Wang. Segment anything in medical images. arXiv preprint arXiv:2304.12306, 2023. 2", + "[31] Ran Margolin, Lihi Zelnik-Manor, and Ayellet Tal. How to evaluate foreground maps? In CVPR, pages 248-255, 2014. 5" + ], + "bbox": [ + 501, + 92, + 890, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "19066", + "bbox": [ + 480, + 945, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[32] Youwei Pang, Xiaoqi Zhao, Tian-Zhu Xiang, Lihe Zhang, and Huchuan Lu. Zoom in and out: A mixed-scale triplet network for camouflaged object detection. In CVPR, pages 2160-2170, 2022. 6", + "[33] Federico Perazzi, Philipp Krahenbuhl, Yael Pritch, and Alexander Hornung. Saliency filters: Contrast based filtering for salient region detection. In CVPR, pages 733-740, 2012. 5", + "[34] Federico Peruzzi, Jordi Pont-Tuset, Brian McWilliams, Luc Van Gool, Markus Gross, and Alexander Sorkine-Hornung. A benchmark dataset and evaluation methodology for video object segmentation. In CVPR, pages 724-732, 2016. 7", + "[35] Rui Qian, Shuangrui Ding, Xian Liu, and Dahua Lin. Semantics meets temporal correspondence: Self-supervised object-centric learning in videos. In ICCV, pages 16675-16687, 2023. 7", + "[36] Yujia Sun, Shuo Wang, Chenglizhao Chen, and Tian-Zhu Xiang. Boundary-guided camouflaged object detection. arXiv preprint arXiv:2207.00794, 2022. 6", + "[37] Lv Tang, Haoke Xiao, and Bo Li. Can sam segment anything? when sam meets camouflaged object detection. arXiv preprint arXiv:2304.04709, 2023. 2", + "[38] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. NeurIPS, 30, 2017. 4", + "[39] Wenhai Wang, Enze Xie, Xiang Li, Deng-Ping Fan, Kaitao Song, Ding Liang, Tong Lu, Ping Luo, and Ling Shao. Pyramid vision transformer: A versatile backbone for dense prediction without convolutions. In ICCV, pages 568-578, 2021. 4", + "[40] Junde Wu, Rao Fu, Huihui Fang, Yuanpei Liu, Zhaowei Wang, Yanwu Xu, Yueming Jin, and Tal Arbel. Medical sam adapter: Adapting segment anything model for medical image segmentation. arXiv preprint arXiv:2304.12620, 2023. 2", + "[41] Yu-Huan Wu, Shang-Hua Gao, Jie Mei, Jun Xu, Deng-Ping Fan, Rong-Guo Zhang, and Ming-Ming Cheng. Jcs: An explainable Covid-19 diagnosis system by joint classification and segmentation. IEEE TIP, 30:3113-3126, 2021. 1", + "[42] Pengxiang Yan, Guanbin Li, Yuan Xie, Zhen Li, Chuan Wang, Tianshui Chen, and Liang Lin. Semi-supervised video salient object detection using pseudo-labels. In ICCV, pages 7284-7293, 2019. 6", + "[43] Charig Yang, Hala Lamdouar, Erika Lu, Andrew Zisserman, and Weidi Xie. Self-supervised video object segmentation by motion grouping. In ICCV, pages 7177-7188, 2021. 1, 2, 6", + "[44] Yi Zhang, Jing Zhang, Wassim Hamidouche, and Olivier Deforges. Predictive uncertainty estimation for camouflaged object detection. IEEE TIP, 2023. 1, 6", + "[45] Yijie Zhong, Bo Li, Lv Tang, Senyun Kuang, Shuang Wu, and Shouhong Ding. Detecting camouflaged object in frequency domain. In CVPR, pages 4504-4513, 2022. 1, 2, 3" + ], + "bbox": [ + 78, + 90, + 468, + 867 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "19067", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2024/Endow SAM with Keen Eyes_ Temporal-spatial Prompt Learning for Video Camouflaged Object Detection/e8cb1a37-360c-44d3-92de-d83afed15e31_model.json b/2024/Endow SAM with Keen Eyes_ Temporal-spatial Prompt Learning for Video Camouflaged Object Detection/e8cb1a37-360c-44d3-92de-d83afed15e31_model.json new file mode 100644 index 0000000000000000000000000000000000000000..92dc316f8999bf39472adabf2ff4731ab39325e7 --- /dev/null +++ b/2024/Endow SAM with Keen Eyes_ Temporal-spatial Prompt Learning for Video Camouflaged Object Detection/e8cb1a37-360c-44d3-92de-d83afed15e31_model.json @@ -0,0 +1,2508 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.808, + 0.045 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.107, + 0.13, + 0.865, + 0.177 + ], + "angle": 0, + "content": "Endow SAM with Keen Eyes: Temporal-spatial Prompt Learning for Video Camouflaged Object Detection" + }, + { + "type": "text", + "bbox": [ + 0.24, + 0.203, + 0.728, + 0.222 + ], + "angle": 0, + "content": "Wenjun Hui\\(^{1,2}\\), Zhenfeng Zhu\\(^{1,2*}\\), Shuai Zheng\\(^{1,2}\\), Yao Zhao\\(^{1,2}\\)" + }, + { + "type": "text", + "bbox": [ + 0.304, + 0.224, + 0.669, + 0.24 + ], + "angle": 0, + "content": "\\(^{1}\\)Institute of Information Science, Beijing Jiaotong University" + }, + { + "type": "text", + "bbox": [ + 0.234, + 0.242, + 0.737, + 0.257 + ], + "angle": 0, + "content": "\\(^{2}\\)Beijing Key Laboratory of Advanced Information Science and Network Technology" + }, + { + "type": "text", + "bbox": [ + 0.281, + 0.26, + 0.686, + 0.275 + ], + "angle": 0, + "content": "{wenjunhui, zhfzhu, zs1997, yzhao}@bjtu.edu.cn" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.309, + 0.314, + 0.326 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.342, + 0.473, + 0.751 + ], + "angle": 0, + "content": "The Segment Anything Model (SAM), a prompt-driven foundational model, has demonstrated remarkable performance in natural image segmentation. However, its application in video camouflaged object detection (VCOD) encounters challenges, chiefly stemming from the overlooked temporal-spatial associations and the unreliability of user-provided prompts for camouflaged objects that are difficult to discern with the naked eye. To tackle the above issues, we endow SAM with keen eyes and propose the Temporal-spatial Prompt SAM (TSP-SAM), a novel approach tailored for VCOD via an ingenious prompted learning scheme. Firstly, motion-driven self-prompt learning is employed to capture the camouflaged object, thereby bypassing the need for user-provided prompts. With the detected subtle motion cues across consecutive video frames, the overall movement of the camouflaged object is captured for more precise spatial localization. Subsequently, to eliminate the prompt bias resulting from inter-frame discontinuities, the long-range consistency within the video sequences is taken into account to promote the robustness of the self-prompts. It is also injected into the encoder of SAM to enhance the representational capabilities. Extensive experimental results on two benchmarks demonstrate that the proposed TSP-SAM achieves a significant improvement over the state-of-the-art methods. With the mIoU metric increasing by \\(7.8\\%\\) and \\(9.6\\%\\), TSP-SAM emerges as a groundbreaking step forward in the field of VCOD." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.778, + 0.208, + 0.793 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.803, + 0.47, + 0.879 + ], + "angle": 0, + "content": "Camouflaged object detection, a pivotal task in computer vision, aims to identify the target object seamlessly blending with its surroundings in images [19, 44]. This task plays an important role in various fields, including surveillance and security [29], wildlife protection [28], and medi" + }, + { + "type": "image", + "bbox": [ + 0.507, + 0.308, + 0.889, + 0.386 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.506, + 0.39, + 0.888, + 0.51 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.522, + 0.892, + 0.62 + ], + "angle": 0, + "content": "Figure 1. Illustration of previous VCOD methods (a-b) and the proposed TSP-SAM (c). In contrast to previous methods with two-stage architecture, the proposed TSP-SAM is trained in an end-to-end manner, incorporating short-long term temporal-spatial relationships to collaboratively learn the reliable prompts. Meanwhile, the long-range consistency is also used to enhance the representational capability of SAM." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.629, + 0.892, + 0.734 + ], + "angle": 0, + "content": "cal image processing [11, 41]. In recent years, single-image camouflaged object detection has made significant progress [16, 19, 45]. However, studies [26, 43] have discerned that observers exhibit heightened perception towards the camouflaged object in motion compared to the static scenarios. This observation has sparked a growing interest in video camouflaged object detection (VCOD) [5]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.735, + 0.892, + 0.81 + ], + "angle": 0, + "content": "Existing VCOD methods [1, 2, 5, 26] commonly employ a two-step framework to model the inter-frame correlations implicitly or explicitly for camouflage breaking, as shown in Fig. 1(a)(b). Despite yielding promising results, they are noticeably limited in certain aspects:" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.811, + 0.892, + 0.841 + ], + "angle": 0, + "content": "- (i) Primarily, their two-stage architecture exposes them to cumulative errors, thus impacting their overall accuracy." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.841, + 0.892, + 0.871 + ], + "angle": 0, + "content": "- (ii) Moreover, they exhibit weak generalization ability due to the limited training data." + }, + { + "type": "list", + "bbox": [ + 0.498, + 0.811, + 0.892, + 0.871 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.871, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Consequently, there is a growing demand for universal models in video camouflaged object detection." + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.888, + 0.221, + 0.9 + ], + "angle": 0, + "content": "*Corresponding author" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19058" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.257 + ], + "angle": 0, + "content": "With the emergence and development of the visual foundation model, the segment anything model (SAM) [25] has showcased unprecedented performance in the field of natural image segmentation. This accomplishment is attributed to the impressive representations learned on 11 million images in SA-1B [25] and 636 million parameters of ViT-H [8]. SAM not only exhibits heightened versatility in terms of model capacity, but also potentially yields more consistent results across different tasks [30]. Inspired by the remarkable generality of SAM, this work endeavors to make SAM suitable for VCOD." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.259, + 0.473, + 0.41 + ], + "angle": 0, + "content": "However, it is worth noting that the performance of SAM is uncertain when applied to VCOD due to the differences between the camouflaged images and natural images. Specifically, SAM, as a prompt-driven foundational model, requires extra prompts when segmenting specific regions. Yet, the seamless blending of the camouflaged object with its surroundings renders it imperceptible to the naked eye, posing a formidable challenge in providing reliable prompts. Hence, there arises an urgent need for adaptive prompt learning to the camouflaged object." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.412, + 0.473, + 0.577 + ], + "angle": 0, + "content": "It is well-known that the temporal-spatial interplay in video sequences plays a pivotal role in video-related tasks. Short-term temporal-spatial relationships accentuate the variations between frames, while long-term temporal-spatial correlations model the temporal-spatial context to characterize the dependencies between the target object and the background. Motivated by these considerations, in this work, we develop a flexible end-to-end temporal-spatial prompt SAM, called TSP-SAM, for VCOD. To the best of our knowledge, the proposed TSP-SAM is the first SAM-based camouflaged object detection framework." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.579, + 0.458, + 0.593 + ], + "angle": 0, + "content": "Our main contributions can be summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.6, + 0.47, + 0.675 + ], + "angle": 0, + "content": "- Instead of user-provided prompts, we propose motion-driven self-prompt learning to capture the camouflaged object. It perceives the overall motion of the camouflaged object by establishing inter-frame associations in the frequency domain, facilitating its spatial identification." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.675, + 0.47, + 0.75 + ], + "angle": 0, + "content": "- To eliminate the prompt bias stemming from inter-frame discontinuities, robust prompt learning based on long-range consistency is proposed. By modeling the long-range temporal-spatial dependencies within video sequences, the robustness of self-prompts is well promoted." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.751, + 0.47, + 0.825 + ], + "angle": 0, + "content": "- In order to enhance the representational capability of SAM, we propose temporal-spatial injection that incorporates the long-range temporal-spatial consistency into the encoder of SAM. This enhancement contributes to more precise detection within the SAM framework." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.825, + 0.47, + 0.9 + ], + "angle": 0, + "content": "- Extensive experiments on MoCA-Mask and CAD2016 datasets demonstrate that the proposed TSP-SAM achieves substantial performance improvements over the state-of-the-art method, with mIoU metric increasing by \\(7.8\\%\\) and \\(9.6\\%\\)." + }, + { + "type": "list", + "bbox": [ + 0.078, + 0.6, + 0.47, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.09, + 0.638, + 0.106 + ], + "angle": 0, + "content": "2. Related work" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.115, + 0.765, + 0.132 + ], + "angle": 0, + "content": "2.1. Camouflaged object detection" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.138, + 0.892, + 0.381 + ], + "angle": 0, + "content": "Camouflaged object detection (COD) aims to identify objects blending into their surroundings in images. Existing CNN-based methods address this task with various strategies. Fan et al. [10, 12] introduced a two-stage process, involving a search stage for localization followed by a segmentation stage for refinement. In multi-task learning, Yang et al. [43] and Ji et al. [21] incorporated an auxiliary task to derive shared context representations for enhancing the COD task. Other strategies, such as frequency cues, have also been explored. These works [6, 45] combined the static frequency spectrum with the spatial representations to discern the subtle differences between the camouflaged object and its surroundings. In contrast to the static frequency perception, our TSP-SAM perceives the frequency energy variants to depict the implicit motion between frames, facilitating the spatial identification of the camouflaged object." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.391, + 0.809, + 0.407 + ], + "angle": 0, + "content": "2.2. Video camouflaged object detection" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.414, + 0.892, + 0.641 + ], + "angle": 0, + "content": "Compared to the static appearance of images, the motion between video frames is deemed a significant clue for breaking camouflage. Existing video camouflaged object detection methods [1, 2, 5, 26] implicitly or explicitly modeled the temporal-spatial correlations through a two-stage architecture to detect camouflaged objects. Explicit motion-based methods [1, 2, 26] segmented the optical flow field after compensation or registration into object and background. In contrast, implicit feature correspondences between frames are learned by the neural network for an effective alignment [5]. Nonetheless, their two-stage architecture leads to cumulative errors. Different from the two-stage architecture, our TSP-SAM is trained in an end-to-end manner, incorporating the short-long term temporal-spatial relationships to collaboratively learn reliable prompts for SAM." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.651, + 0.727, + 0.667 + ], + "angle": 0, + "content": "2.3. Segment anything model" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.675, + 0.892, + 0.901 + ], + "angle": 0, + "content": "The segment anything model (SAM) [25], a prompt-driven large-scale foundation model, has exhibited unprecedented performance in natural image segmentation. The core components of SAM comprise an image encoder, a flexible prompt encoder and a lightweight mask decoder. Despite exhibiting excellent zero-shot segmentation capability, SAM performs unstable in application-specific tasks [4]. It fails to segment medical images [18], camouflaged objects [37] and concealed scenes [22]. To apply SAM to medical images, MedSAM [30] and SAM-Adapter [4, 40] incorporated specific domain knowledge into SAM. Yet, up to now, there has been no work on adapting SAM to be tailored for camouflaged object detection. To fill this gap, TSP-SAM explores the temporal-spatial relationships, presenting the first SAM-based camouflaged object detection framework." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "19059" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.168, + 0.091, + 0.805, + 0.313 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.325, + 0.892, + 0.409 + ], + "angle": 0, + "content": "Figure 2. The framework of the proposed TSP-SAM. In motion-driven self-prompt learning (b), the implicit motion between consecutive frames is perceived to facilitate the spatial identification of the camouflaged object, thereby learning the dense self-prompt \\( P_{d}(t) \\) and sparse self-prompt \\( P_{s}(t) \\). Subsequently, in robust prompt learning (c), the long-range consistency \\( X_{c}(t) \\) is modeled to refine the sparse self-prompt \\( P_{s}(t) \\). Moreover, in temporal-spatial injection (d), the long-range consistency \\( X_{c}(t) \\) is injected into the image embedding \\( X_{e}(t) \\) of SAM for representation enhancement. During the training phase, SAM (a) is frozen, including the image encoder \\( E_{i} \\), prompt encoder \\( E_{p} \\) and the mask decoder \\( E_{m} \\)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.415, + 0.407, + 0.433 + ], + "angle": 0, + "content": "3. Overview of the proposed TSP-SAM" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.441, + 0.188, + 0.456 + ], + "angle": 0, + "content": "3.1. Notations" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.466, + 0.47, + 0.603 + ], + "angle": 0, + "content": "In the subsequent sections, we establish the key notations for clarity and consistency in our discourse. Given a sequence of camouflaged images, denote by \\(I(t) \\in \\mathbb{R}^{H \\times W \\times 3}\\) the \\(t\\)-th frame and \\(I^w(t) = [I(t - T + 1), \\dots, I(t)] \\in \\mathbb{R}^{T \\times H \\times W \\times 3}\\) the corresponding windowed sequence containing \\(T\\) frames. For a tensor \\(A \\in \\mathbb{R}^{H \\times W \\times C}\\), \\(\\mathrm{Mat}(A) \\in \\mathbb{R}^{HW \\times C}\\) denotes its matrixization by spatial-wise concatenation. For a matrix \\(B \\in \\mathbb{R}^{HW \\times C}\\), \\(\\mathrm{Ten}(B) \\in \\mathbb{R}^{H \\times W \\times C}\\) represents its tensorization by spatial-wise division." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.616, + 0.206, + 0.63 + ], + "angle": 0, + "content": "3.2. Framework" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.64, + 0.469, + 0.684 + ], + "angle": 0, + "content": "Fig. 2 illustrates the framework of the proposed TSP-SAM. The core components consist mainly of the following three modules:" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.687, + 0.469, + 0.762 + ], + "angle": 0, + "content": "(a) Motion-driven self-prompt learning. The motion-driven self-prompt learning is to use the implicit inter-frame motion \\(\\hat{X}_{(t - 1)\\rightarrow t}\\) in the frequency domain to facilitate the spatial identification of the camouflaged object, thereby learning the self-prompts for SAM." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.764, + 0.469, + 0.839 + ], + "angle": 0, + "content": "(b) Robust prompt learning based on long-range consistency. To eliminate the prompt bias stemming from underlying inter-frame discontinuities, the long-range temporal-spatial consistency \\( X_{c}(t) \\) in video sequences is modeled to promote the robustness of the self-prompts." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.841, + 0.469, + 0.901 + ], + "angle": 0, + "content": "(c) Temporal-spatial injection for representation enhancement. To enhance the representational capabilities of SAM, the long-range temporal-spatial consistency \\( X_{c}(t) \\) is injected into the image embedding \\( X_{e}(t) \\) of SAM, con" + }, + { + "type": "list", + "bbox": [ + 0.076, + 0.687, + 0.469, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.417, + 0.732, + 0.432 + ], + "angle": 0, + "content": "attributing to more precise detection." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.447, + 0.635, + 0.464 + ], + "angle": 0, + "content": "4. Methodology" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.472, + 0.891, + 0.503 + ], + "angle": 0, + "content": "In this section, we detail the proposed TSP-SAM method for video camouflaged object detection." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.514, + 0.81, + 0.531 + ], + "angle": 0, + "content": "4.1. Motion-driven self-prompt learning" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.538, + 0.892, + 0.644 + ], + "angle": 0, + "content": "It is well-known that SAM requires user-provided prompts when segmenting specific regions. Nevertheless, the difficulty in distinguishing the camouflaged object with the naked eye makes the requirement of providing visual prompts a significant challenge. Consequently, a compelling need arises to devise a prompt learning network, capable of adaptively identifying the camouflaged object." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.665, + 0.805, + 0.681 + ], + "angle": 0, + "content": "4.1.1 Frequency-based motion perception" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.689, + 0.892, + 0.825 + ], + "angle": 0, + "content": "Inspired by the two-stream hypothesis in neuroscience [14], which highlights the importance of motion in visual perception, we consider the motion as a crucial clue for breaking camouflage. Building on this premise, several studies [7, 23] indicate the motion between frames can be expressed exhaustively in the frequency domain through the variants in frequency energy. Hence, frequency-based motion perception is designed to facilitate the spatial identification of the camouflaged object." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.826, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Specifically, as shown in Fig. 3, given consecutive video frames \\( I(t - 1) \\in \\mathbb{R}^{H \\times W \\times 3} \\) and \\( I(t) \\in \\mathbb{R}^{H \\times W \\times 3} \\), we first map them into the frequency domain. Following [45], the image features are divided into a set of \\( s \\times s \\) patches and each patch is processed by Discrete Cosine Transform" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19060" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.084, + 0.093, + 0.46, + 0.25 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.265, + 0.471, + 0.307 + ], + "angle": 0, + "content": "Figure 3. The framework of the motion-driven self-prompt learning. The spatial representations \\(\\{X_{b}(t)\\}_{b = 1}^{N_{b}}\\) are extracted by PVT [39], where \\(N_{b}\\) is the number of attention blocks." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.318, + 0.471, + 0.396 + ], + "angle": 0, + "content": "(DCT) into frequency spectrum, where \\(s\\) is the size of patches. To group all components of the same frequency into one channel, we flatten and reshape the spectrum of these patches, forming the frequency features \\(\\hat{I}(t - 1) \\in \\mathbb{R}^{\\frac{H}{s} \\times \\frac{W}{s} \\times 3s^2}\\) and \\(\\hat{I}(t) \\in \\mathbb{R}^{\\frac{H}{s} \\times \\frac{W}{s} \\times 3s^2}\\)." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.398, + 0.47, + 0.504 + ], + "angle": 0, + "content": "In order to perceive the overall motion of the camouflaged object, the inter-frame associations are established on high frequency and low frequency bands respectively. Specifically, the frequency spectrum \\(\\{\\hat{I} (t - 1),\\hat{I} (t)\\}\\) is divided into two groups in the channel dimension, namely low frequency bands \\(\\{\\hat{I}^l (t - 1),\\hat{I}^l (t)\\}\\) and high frequency bands \\(\\{\\hat{I}^h (t - 1),\\hat{I}^h (t)\\}\\). Hence, we have:" + }, + { + "type": "equation", + "bbox": [ + 0.091, + 0.516, + 0.47, + 0.535 + ], + "angle": 0, + "content": "\\[\n\\hat {X} _ {(t - 1) \\rightarrow t} ^ {l} = \\mathrm {M L P} \\big (\\operatorname {A T T N} (\\operatorname {M a t} (\\hat {I} ^ {l} (t - 1)), \\operatorname {M a t} (\\hat {I} ^ {l} (t))) \\big) \\quad (1)\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.088, + 0.55, + 0.469, + 0.568 + ], + "angle": 0, + "content": "\\[\n\\hat {X} _ {(t - 1) \\rightarrow t} ^ {h} = \\operatorname {M L P} (\\operatorname {A T T N} (\\operatorname {M a t} (\\hat {I} ^ {h} (t - 1)), \\operatorname {M a t} (\\hat {I} ^ {h} (t)))) \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.575, + 0.469, + 0.718 + ], + "angle": 0, + "content": "where \\(\\hat{X}_{(t - 1)\\rightarrow t}^{l}\\in \\mathbb{R}^{\\frac{HW}{s^2}\\times \\frac{3s^2}{2}}\\) and \\(\\mathrm{MLP}(\\cdot)\\) is the linear layer. \\(\\mathrm{ATTN}(\\cdot ,\\cdot)\\) is multi-head cross attention [38] to extract the inter-frame correspondences, in which the \\((t - 1)\\)-th frame is taken as query and the \\(t\\)-th frame serves as key and value. Note that Eq.1 tends to highlight the global deformations of the camouflaged object by characterizing the feature correspondences on the low-frequency bands. Meanwhile, Eq.2 implicitly reveals the local edge motion of the camouflaged object on high-frequency bands." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.72, + 0.469, + 0.75 + ], + "angle": 0, + "content": "Subsequently, the complete implicit motion in the frequency domain is obtained by:" + }, + { + "type": "equation", + "bbox": [ + 0.08, + 0.761, + 0.469, + 0.792 + ], + "angle": 0, + "content": "\\[\n\\hat {X} _ {(t - 1) \\rightarrow t} = \\hat {I} (t) + \\operatorname {C B A} (\\operatorname {C a t} \\left(\\operatorname {T e n} \\left(\\hat {X} _ {(t - 1) \\rightarrow t} ^ {l}\\right), \\operatorname {T e n} \\left(\\hat {X} _ {(t - 1) \\rightarrow t} ^ {h}\\right)\\right)) \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.792, + 0.469, + 0.902 + ], + "angle": 0, + "content": "where \\(\\mathrm{CBA}(\\cdot)\\) is the convolution operation followed by batch normalization and activation and \\(\\operatorname{Cat}(\\cdot)\\) is the channel-wise concatenation. Importantly, \\(\\hat{X}_{(t - 1)\\rightarrow t}\\in\\) \\(\\mathbb{R}^{\\frac{H}{s}\\times \\frac{W}{s}\\times 3s^{2}}\\) implicitly depicts the overall motion of the camouflaged object by perceiving the variants in the frequency energy, enabling to facilitate the discovery of the camouflaged object." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.092, + 0.892, + 0.121 + ], + "angle": 0, + "content": "4.1.2 Implicit motion-induced spatial-frequency fusion" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.125, + 0.892, + 0.239 + ], + "angle": 0, + "content": "To translate the perceived motion into the spatial domain, the implicit motion-induced spatial-frequency representation fusion is established. This process involves the integration between the implicit motion \\(\\hat{X}_{(t - 1)\\rightarrow t}\\) in the frequency domain and the spatial representations \\(X_{b}(t)\\in\\) \\(\\mathbb{R}^{\\frac{H}{2^{b + 1}}\\times \\frac{W}{2^{b + 1}}\\times C},b\\in \\{1,\\dots ,N_b\\}\\) extracted by PVT [39], where \\(N_{b}\\) is the number of attention blocks. It can be expressed as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.508, + 0.248, + 0.892, + 0.265 + ], + "angle": 0, + "content": "\\[\n\\hat {X} _ {b} (t) = X _ {b} (t) + \\operatorname {T e n} (\\operatorname {A T T N} (\\operatorname {M a t} (\\operatorname {C B A} (\\hat {X} _ {(t - 1) \\rightarrow t})), \\operatorname {M a t} (X _ {b} (t)))) \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.272, + 0.892, + 0.337 + ], + "angle": 0, + "content": "where \\(\\hat{X}_b(t) \\in \\mathbb{R}^{\\frac{H}{2^b + 1} \\times \\frac{W}{2^b + 1} \\times C}\\). Note that the spatial perception of the camouflaged object could be strengthened by establishing cross attention between the implicit motion as query and the spatial representation as key and value." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.355, + 0.701, + 0.37 + ], + "angle": 0, + "content": "4.1.3 Self-prompt learning" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.377, + 0.892, + 0.451 + ], + "angle": 0, + "content": "Based on the co-representations \\(\\{\\hat{X}_b(t)\\}_{b=1}^{N_b}\\), two prompt heads are adopted to generate the dense self-prompt \\(P_d(t)\\) and sparse self-prompt \\(P_s(t)\\), respectively. Specifically, we adopt the decoder of SLTNet [5] as dense prompt head \\(E_d\\) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.609, + 0.452, + 0.892, + 0.471 + ], + "angle": 0, + "content": "\\[\nP _ {d} (t) = E _ {d} \\left(\\left\\{\\hat {X} _ {b} (t) \\right\\} _ {b = 1} ^ {N _ {b}}\\right) \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.475, + 0.892, + 0.535 + ], + "angle": 0, + "content": "where \\( P_{d}(t) \\) is the dense self-prompt. Meanwhile, as illustrated in Fig. 3, to obtain the sparse prompt, we adopt the residual block of Resnet [17] to process the co-representations in a bottom-up manner:" + }, + { + "type": "equation", + "bbox": [ + 0.511, + 0.543, + 0.891, + 0.591 + ], + "angle": 0, + "content": "\\[\n\\tilde {X} _ {b} (t) = \\left\\{ \\begin{array}{l l} \\operatorname {R e s b l o c k} \\left(\\hat {X} _ {b} (t)\\right) & b = N _ {b} \\\\ \\operatorname {R e s b l o c k} \\left(\\hat {X} _ {b} (t) + \\tilde {X} _ {b + 1} (t)\\right) & b \\in \\{1, \\dots , N _ {b} - 1 \\} \\end{array} \\right. \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.592, + 0.872, + 0.607 + ], + "angle": 0, + "content": "Finally, the sparse self-prompt is represented as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.581, + 0.615, + 0.891, + 0.634 + ], + "angle": 0, + "content": "\\[\nP _ {s} (t) = \\operatorname {A v g p o o l} \\left(\\operatorname {C B A} \\left(\\tilde {X} _ {1} (t)\\right)\\right) \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.643, + 0.862, + 0.659 + ], + "angle": 0, + "content": "where \\(\\mathrm{Avgpool}(\\cdot)\\) is global average pooling operation." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.667, + 0.891, + 0.698 + ], + "angle": 0, + "content": "4.2. Robust prompt learning based on long-range consistency" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.704, + 0.892, + 0.901 + ], + "angle": 0, + "content": "The self-prompts of the camouflaged object are well learned by using the perceived inter-frame motion to facilitate its spatial identification. However, the susceptibility of inter-frame motion to underlying temporal-spatial discontinuities, such as occlusions, shaking, scene changes, etc., brings a potential lack of reliability in the learned self-prompts. It is well known that the long-range dependencies within the video sequences enable to characterize the temporal-spatial context to cope with the short-term temporal-spatial discontinuities. Hence, it is necessary to explore the long-range temporal-spatial consistency within the video sequences to improve the robustness of the self-prompts, which likes endowing SAM with keen eyes." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "19061" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.077, + 0.092, + 0.37, + 0.108 + ], + "angle": 0, + "content": "4.2.1 Long-range consistency squeezing" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.115, + 0.47, + 0.176 + ], + "angle": 0, + "content": "As illustrated in Fig. 2(c), given the windowed sequence \\( I^{w}(t) \\in \\mathbb{R}^{T \\times H \\times W \\times 3} \\) corresponding to the current frame \\( I(t) \\), we adopt a 3D convolution neural network \\( E_{t} \\) to extract the long-range temporal-spatial relationships:" + }, + { + "type": "equation", + "bbox": [ + 0.204, + 0.186, + 0.469, + 0.203 + ], + "angle": 0, + "content": "\\[\nX ^ {w} (t) = E _ {t} \\left(I ^ {w} (t)\\right) \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.212, + 0.469, + 0.26 + ], + "angle": 0, + "content": "where \\( X^w(t) \\in \\mathbb{R}^{T \\times \\frac{H}{4} \\times \\frac{W}{4} \\times C} \\) fully reveals the temporal-spatial dependencies within the video sequences. In our work, we adopt 3D Resnet18 [15] as \\( E_t \\)." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.261, + 0.469, + 0.35 + ], + "angle": 0, + "content": "Subsequently, to further extract the long-range temporal-spatial consistency of the camouflaged object, we take the camouflage representation \\( X_{e}(t) \\) from the encoder of SAM as query and the long-range temporal-spatial dependencies \\( X^{w}(t) \\) averaged along the temporal dimension as key and value to establish the cross attention:" + }, + { + "type": "equation", + "bbox": [ + 0.09, + 0.361, + 0.469, + 0.379 + ], + "angle": 0, + "content": "\\[\nX _ {c} (t) = \\operatorname {A T T N} \\left(\\operatorname {M a t} \\left(X _ {e} (t)\\right), \\operatorname {M a t} \\left(\\operatorname {A v g} _ {\\mathrm {t}} \\left(X ^ {w} (t)\\right)\\right)\\right) \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.388, + 0.469, + 0.435 + ], + "angle": 0, + "content": "where \\(\\mathrm{Avg}_{\\mathrm{t}}(\\cdot)\\) is an average operation to squeeze the temporal dimension. \\(X_{c}(t)\\in \\mathbb{R}^{\\frac{HW}{16}\\times C}\\) characterizes the long-range consistency specific to the camouflaged object." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.453, + 0.301, + 0.469 + ], + "angle": 0, + "content": "4.2.2 Robust prompt learning" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.477, + 0.469, + 0.582 + ], + "angle": 0, + "content": "To enhance the robustness of the self-prompts to inter-frame temporal-spatial discontinuities, the long-range temporal-spatial consistency \\( X_{c}(t) \\) should be induced into the prompt learning. Specifically, in order to map the long-range consistency into the coordinate space, a learnable query \\( Q_{c}(t) \\in \\mathbb{R}^{4 \\times C} \\) is defined to establish cross attention with \\( X_{c}(t) \\). It is formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.146, + 0.593, + 0.469, + 0.611 + ], + "angle": 0, + "content": "\\[\nP _ {r} (t) = \\operatorname {A v g} _ {\\mathrm {c}} (\\operatorname {A T N} \\left(Q _ {c} (t), X _ {c} (t)\\right)) \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.62, + 0.469, + 0.666 + ], + "angle": 0, + "content": "where \\(\\mathrm{Avg}_c(\\cdot)\\) is an average operation along the channel dimension. \\(P_r(t)\\in \\mathbf{R}^4\\) is a refinement term derived from long-range consistency to correct the sparse self-prompt:" + }, + { + "type": "equation", + "bbox": [ + 0.195, + 0.676, + 0.469, + 0.693 + ], + "angle": 0, + "content": "\\[\nP _ {s} ^ {r} (t) = P _ {s} (t) + P _ {r} (t) \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.703, + 0.469, + 0.764 + ], + "angle": 0, + "content": "Through incorporating long-range consistency into the prompt learning, the visual prompts could be more robust to the temporal-spatial discontinuities, which means endowing SAM with a pair of keen eyes." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.772, + 0.469, + 0.802 + ], + "angle": 0, + "content": "4.3. Temporal-spatial injection for representation enhancement" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.81, + 0.469, + 0.901 + ], + "angle": 0, + "content": "It is known that the remarkable performance of SAM relies not only on well learned robust prompts but also on the representational capability of the image encoder. However, due to the fact that the image embedding of SAM lacks the temporal-spatial information, the final predictions of consecutive frames may be inconsistent. Hence, it is necessary" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.892, + 0.122 + ], + "angle": 0, + "content": "to inject the long-range temporal-spatial consistency into SAM. As shown in Fig. 2(d), it is represented as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.595, + 0.132, + 0.892, + 0.15 + ], + "angle": 0, + "content": "\\[\n\\hat {X} _ {e} (t) = X _ {e} (t) + \\operatorname {T e n} \\left(X _ {c} (t)\\right) \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.161, + 0.892, + 0.222 + ], + "angle": 0, + "content": "where \\( X_{c}(t) \\) is the long-range consistency from Eq. 9. In this way, based on the enhanced image representations and the robust prompts called keen eyes, SAM achieves more precise camouflaged object detection in videos." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.23, + 0.688, + 0.246 + ], + "angle": 0, + "content": "4.4. Model optimization" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.254, + 0.892, + 0.284 + ], + "angle": 0, + "content": "We train the proposed TSP-SAM in an end-to-end manner by minimizing the joint loss below:" + }, + { + "type": "equation", + "bbox": [ + 0.568, + 0.294, + 0.892, + 0.313 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} (t) = \\mathcal {L} _ {S A M} (t) + \\alpha \\mathcal {L} _ {p} ^ {s} (t) + \\beta \\mathcal {L} _ {p} ^ {d} (t) \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.324, + 0.891, + 0.354 + ], + "angle": 0, + "content": "where \\(\\mathcal{L}_{SAM},\\mathcal{L}_p^s\\) and \\(\\mathcal{L}_p^d\\) supervise the final prediction, the sparse prompt and the dense prompt, respectively." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.355, + 0.892, + 0.387 + ], + "angle": 0, + "content": "Following [3], the supervision on the sparse prompt \\(\\mathcal{L}_p^s (t)\\) is as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.515, + 0.395, + 0.892, + 0.414 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {p} ^ {s} (t) = \\operatorname {B C E} \\left(P _ {s} ^ {r} (t), G (t)\\right) + \\operatorname {M S E} \\left(P _ {s} ^ {r} (t), G (t)\\right) \\tag {14}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.424, + 0.892, + 0.5 + ], + "angle": 0, + "content": "where BCE is binary cross-entropy loss to supervise the coordinates and MSE is mean square error to supervise the size of the box prompt or the scale of the point prompt. For the dense prompt loss \\(\\mathcal{L}_p^d\\) and the final prediction loss \\(\\mathcal{L}_{SAM}\\), we adopt the hybrid loss [13] for supervision:" + }, + { + "type": "equation", + "bbox": [ + 0.619, + 0.51, + 0.892, + 0.528 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {h y b r i d}} = \\mathcal {L} _ {c e} ^ {w} + \\mathcal {L} _ {i o u} ^ {w} \\tag {15}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.538, + 0.892, + 0.568 + ], + "angle": 0, + "content": "where \\(\\mathcal{L}_{ce}^{w}\\) is the weighted cross-entropy loss and \\(\\mathcal{L}_{iou}^{w}\\) is the weighted intersection-over-union loss." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.581, + 0.792, + 0.599 + ], + "angle": 0, + "content": "5. Experiment results and analysis" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.606, + 0.691, + 0.623 + ], + "angle": 0, + "content": "5.1. Experiment settings" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.629, + 0.892, + 0.84 + ], + "angle": 0, + "content": "Datasets. Following previous methods [5], we have conducted a comprehensive evaluation on two widely used datasets: MoCA-Mask [5] and CAD2016 [1]. The MoCA-Mask dataset [5] is a video camouflaged object detection dataset, covering camouflaged animals moving in natural scenes. It contains 87 video sequences (22939 frames), of which 71 video sequences are used for training and 16 sequences for testing. It is well annotated with bounding boxes and pixel-level segmentation masks on every fifth frames. Meanwhile, Camouflaged Animal Dataset [1] (CAD2016) is a small video camouflaged object detection dataset extracted from YouTube videos. It consists of nine short video clips in total and is well annotated with pixel-level semantic masks on every fifth frames." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.84, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Metrics. During the testing phase, we assess the prediction masks using six widely-used metrics: S-measure \\((S_{\\alpha})\\) [9], Weighted F-measure \\((F_{\\beta}^{w})\\) [31], Enhanced-alignment measure \\((E_{\\phi})\\) [13], Mean absolute error (MAE, \\(\\mathcal{M}\\)) [33]," + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19062" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.081, + 0.09, + 0.894, + 0.296 + ], + "angle": 0, + "content": "
MethodYearInputMoCA-MaskCAD2016
\\(S_{\\alpha}\\uparrow\\)\\(F^{\\text{w}}_{\\beta}\\uparrow\\)\\(E_{\\phi}\\uparrow\\)\\(\\mathcal{M}\\downarrow\\)mDice↑mIoU↑\\(S_{\\alpha}\\uparrow\\)\\(F^{\\text{w}}_{\\beta}\\uparrow\\)\\(E_{\\phi}\\uparrow\\)\\(\\mathcal{M}\\downarrow\\)mDice↑mIoU↑
SINet[10]2020-CVPRImage0.5740.1850.6550.0300.2210.1560.6010.2040.5890.0890.2890.209
SINet-v2[12]2021-TPAMIImage0.5710.1750.6080.0350.2110.1530.5440.1810.5460.0490.1700.110
ZoomNet[32]2022-CVPRImage0.5820.2110.5360.0330.2240.1670.5870.2250.5940.0630.2460.166
BGNet [36]2022-IJCAIImage0.5900.2030.6470.0230.2250.1670.6070.2030.6660.0890.3450.256
FEDERNet[16]2023-CVPRImage0.5550.1580.5420.0490.1920.1320.6070.2460.7250.0610.3610.257
FSPNet[19]2023-CVPRImage0.5940.1820.6080.0440.2380.1670.5390.2200.5530.1450.3090.212
PUENet[44]2023-TIPImage0.5940.2040.6190.0370.3020.2120.6730.4270.8030.0340.4990.389
RCRNet[42]2019-ICCVVideo0.5970.1740.5830.0250.1940.137------
PNS-Net[20]2021-MICCAIVideo0.5760.1340.5620.0380.1890.1330.6780.3690.7200.0430.4090.308
MG[43]2021-ICCVVideo0.5470.1650.5370.0950.1970.1410.4840.3140.5580.3700.3510.260
SLT-Net[5]2022-CVPRVideo0.6560.3570.7850.0210.3870.3100.6790.4200.8050.0330.4450.342
TSP-SAM(M+P)OursVideo0.6730.4000.7660.0120.4210.3450.6810.5000.8530.0310.4960.393
TSP-SAM(M+B)OursVideo0.6890.4440.8080.0080.4580.3880.7040.5240.9120.0280.5430.438
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.306, + 0.893, + 0.336 + ], + "angle": 0, + "content": "Table 1. Quantitative comparisons over two benchmark datasets. The top three results are highlighted in red, green, and blue. M+P: Combination of mask and point prompts. M+B: Combination of mask and box prompts." + }, + { + "type": "table", + "bbox": [ + 0.081, + 0.342, + 0.468, + 0.528 + ], + "angle": 0, + "content": "
MethodSα↑Fwβ↑Eφ↑M↓mDice↑mIoU↑
SINet[10]0.6360.3460.7750.0410.3810.283
SINet-v2[12]0.6530.3820.7620.0390.4130.318
ZoomNet[32]0.6330.3490.6010.0330.3490.273
FEDERNet[16]0.5730.2720.6090.0510.2600.199
FSPNet[19]0.6810.4010.7160.0440.4460.332
PUENet[44]0.6910.4850.7950.0340.5140.396
RCRNet[42]0.6270.2870.6660.0480.3090.229
PNS-Net[20]0.6550.3250.6730.0480.3840.290
MG[43]0.5940.3360.6910.0590.3680.268
SLT-Net[5]0.6960.4810.8450.0300.4930.401
TSP-SAM(M+P)0.7050.5650.8360.0270.5310.422
TSP-SAM(M+B)0.7510.6280.8510.0210.6030.496
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.538, + 0.47, + 0.582 + ], + "angle": 0, + "content": "Table 2. Generalization comparisons on CAD2016 dataset. The best results are highlighted in **bold**. M+P: Combination of mask and point prompts. M+B: Combination of mask and box prompts." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.608, + 0.47, + 0.655 + ], + "angle": 0, + "content": "meanDice (mDice), meanIoU (mIoU). Overall, a better VCOD method has larger \\( S_{\\alpha} \\), \\( F_{\\beta}^{w} \\), \\( E_{\\phi} \\), mDice and mIoU scores, but a smaller MAE score." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.66, + 0.473, + 0.901 + ], + "angle": 0, + "content": "Implementation details. In our experiments, we freeze SAM and train our TSP-SAM in an end-to-end manner with a maximum of 20 epochs. Specifically, for a fair comparison, both the training and testing images are resize to \\(352 \\times 352\\). We leverage Adam [24] as our optimizer and set the weight decay to \\(1 \\times 10^{-5}\\). On the MoCA-Mask dataset, the learning rate is initialized to \\(1 \\times 10^{-5}\\) for the spatial extractor of motion-driven self-prompt learning and \\(5 \\times 10^{-5}\\) for other trainable parameters. After 10 epochs, the learning rate is reduced by \\(90\\%\\). On the CAD2016 dataset, we adopt a ratio of 9:1 to divide all classes into training and testing sets and report the mean value over the testing set. We set the initial learning rate to \\(1 \\times 10^{-4}\\) for the spatial extractor of the motion-driven self-prompt learning and \\(5 \\times 10^{-4}\\) for other trainable parameters. The other settings are the same as those on MoCA-Mask dataset." + }, + { + "type": "table", + "bbox": [ + 0.521, + 0.342, + 0.885, + 0.409 + ], + "angle": 0, + "content": "
MethodTuning parameters (M)MethodTime (s) ↓Speed (fps) ↑
FSPNet274.24FSPNet568.701.31
SLT-Net253.402.94
SLT-Net82.41Overall (ours)294.992.53
TSP-SAM(ours)89.78Prompt learning73.2810.17
SAM221.713.36
" + }, + { + "type": "table_caption", + "bbox": [ + 0.519, + 0.414, + 0.893, + 0.444 + ], + "angle": 0, + "content": "Table 3. The number of Table 4. Comparisons on infertuning parameters. Hence time and speed." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.464, + 0.73, + 0.481 + ], + "angle": 0, + "content": "5.2. Performance comparison" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.489, + 0.893, + 0.535 + ], + "angle": 0, + "content": "1) Baselines: To evaluate the performance of the proposed TSP-SAM, we compare it with a range of state-of-the-art methods. These methods fall into two categories:" + }, + { + "type": "text", + "bbox": [ + 0.501, + 0.537, + 0.894, + 0.641 + ], + "angle": 0, + "content": "- Single-image camouflaged object detection. These methods analyze the static appearance of images to discern the subtle differences between camouflaged objects and their surroundings. The compared methods in this type include SINet [10], SINet-v2 [12], ZoomNet [32], BGNet [36], FEDERNet [16], FSPNet[19], and PUENet [44]." + }, + { + "type": "text", + "bbox": [ + 0.501, + 0.643, + 0.894, + 0.717 + ], + "angle": 0, + "content": "- Video object segmentation. This set of methods explores the temporal-spatial relationships between frames for the purpose of segmenting objects with specific attributes. The compared methods in this type encompass RCRNet [42], PNS-Net [20], MG[43], and SLT-Net[5]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.895, + 0.902 + ], + "angle": 0, + "content": "2) Quantitative Results: We evaluate TSP-SAM on MoCA-Mask and CAD2016 datasets. From Table 1, we can conclude that, (i) compared to the combination of mask and point prompts, the combination of mask and box prompts is more reliable for SAM. This is attributed to the limitations of the point prompt in conveying the boundary information. In contrast, bounding box excels in indicating the boundaries of the camouflaged object, even in the presence of bias. (ii) the strong contrast between the performance of TSP-SAM and single-image camouflaged object detection methods points to the importance of temporal-spatial relationships in breaking camouflage. (iii) the proposed TSP-SAM" + }, + { + "type": "list", + "bbox": [ + 0.498, + 0.537, + 0.895, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19063" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.084, + 0.089, + 0.482, + 0.265 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.489, + 0.089, + 0.885, + 0.265 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.278, + 0.895, + 0.321 + ], + "angle": 0, + "content": "Figure 4. Visualization of our proposed TSP-SAM and baseline methods on MoCA-Mask and CAD2016 dataset. From left to right: frames (1st column), ground truth (2nd column), prediction of our TSP-SAM with mask and box prompts (3rd column), prediction of our TSP-SAM with mask and point prompts (4th column), and predictions of compared methods (5th-7th columns)." + }, + { + "type": "image", + "bbox": [ + 0.08, + 0.326, + 0.468, + 0.513 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.524, + 0.471, + 0.608 + ], + "angle": 0, + "content": "Figure 5. Visualization of several failure cases with mask and point prompts. From left to right: frames (1st column), ground truth (2nd column), mask and box prompts (3rd column), prediction with mask and box prompts (4th column), mask and point prompts (5th column), and prediction with mask and point prompts (6th column)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.625, + 0.47, + 0.686 + ], + "angle": 0, + "content": "surpasses all video object segmentation methods. Notably, TSP-SAM exhibits a \\(1.3\\% - 8.7\\%\\) improvement over SLT-Net [5] on MoCA-Mask dataset, meaning that the temporal-spatial exploration in TSP-SAM is more effective." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.69, + 0.471, + 0.902 + ], + "angle": 0, + "content": "To assess the generalization performance, we train on the MoCA-Mask dataset and report the test performance on the CAD2016 dataset. It can be observed from Table 2 that the proposed TSP-SAM outperforms all baselines. Compared with SLTNet [5], TSP-SAM surpasses it by \\(0.9\\% - 14.7\\%\\), indicating that TSP-SAM is capable of generalizing to the unseen data more effectively. Meanwhile, we also test the performance of TSP-SAM on general video object segmentation datasets. On DAVIS2016 dataset [34], our model achieves comparable performance to the SOTA SLT-Net. While on SegTrack-v2 dataset [27] with multiple objects, our method obtains \\(62.84\\%\\) for mIoU, surpassing SLT-Net by \\(9.9\\%\\). Surprisingly, our model is even comparable to the recent SMTC [35] on SegTrack-v2." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.329, + 0.892, + 0.419 + ], + "angle": 0, + "content": "Besides, we report the number of tuning parameters, the inference time and speed. From Table 3, compared with SLT-Net, our model achieves \\(9.6\\%\\) improvements with a slight parameter increase. From Table 4, our model is comparable to the latest video-oriented SLT-Net. Decomposing the overall inference time, SAM consumes \\(75\\%\\) of it." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.421, + 0.893, + 0.602 + ], + "angle": 0, + "content": "3) Qualitative Results: To provide more intuitive evaluations of the proposed TSP-SAM, we conduct two qualitative discussions. Firstly, as shown in Fig. 4, we present the segmentation results of several examples to intuitively compare the proposed TSP-SAM with the baselines. It can be seen that: (i) some confusing surrounding regions interfere with the baseline methods, while our TSP-SAM is more robust in identifying the camouflaged object. This demonstrates the necessity of short-long term temporal-spatial relationships. (ii) compared with the baselines, our TSP-SAM exhibits superior capabilities in boundary segmentation in terms of discriminability and details." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.604, + 0.895, + 0.817 + ], + "angle": 0, + "content": "Secondly, to further analyze the influence of the type of sparse prompts on overall performance, we visualize the several failure cases with mask and point prompts. From Fig. 5, it can be found that: (i) since the target object is not always a regular shape, the learned point prompt is not on the target object although it is near the center. (ii) under the condition with unclear boundaries, the point prompt is more likely to segment the regions of background. (iii) the point prompt may induce the model to segment the local areas instead of the complete target object. In contrast, the box prompt can avoid the above defects and achieve more precise segmentation results. Meanwhile, the complete localization under the learned mask and box prompts verifies the effectiveness of the temporal-spatial prompt learning." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.831, + 0.65, + 0.848 + ], + "angle": 0, + "content": "5.3. Ablation study" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.895, + 0.901 + ], + "angle": 0, + "content": "To gain a comprehensive understanding of the efficacy of core components and the impact of hyper parameters, we delve into a detailed analysis." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "19064" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.081, + 0.089, + 0.468, + 0.168 + ], + "angle": 0, + "content": "
MethodSα↑Fwβ↑Eφ↑mDice↑mIoU↑
SAM+SSP0.6370.3300.7680.3670.311
SAM+MSP0.6570.3530.8230.4070.340
SAM+MSP+LCP0.6650.4090.7570.4290.366
SAM+MSP+LCP+TSI0.6890.4440.8080.4580.388
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.178, + 0.47, + 0.248 + ], + "angle": 0, + "content": "Table 5. Ablation studies of the core components on MoCA-Mask dataset. SSP: Spatial-based self-prompt learning. MSP: Motion-driven self-prompt learning. LCP: Long-range consistency-based robust prompt learning. TSI: Temporal-spatial injection for representation enhancement." + }, + { + "type": "image", + "bbox": [ + 0.117, + 0.256, + 0.27, + 0.326 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.18, + 0.328, + 0.235, + 0.339 + ], + "angle": 0, + "content": "(a) mDice" + }, + { + "type": "image", + "bbox": [ + 0.273, + 0.257, + 0.418, + 0.326 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.336, + 0.328, + 0.377, + 0.339 + ], + "angle": 0, + "content": "(b) \\(F_{\\beta}^{w}\\)" + }, + { + "type": "image", + "bbox": [ + 0.119, + 0.34, + 0.27, + 0.41 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.183, + 0.411, + 0.232, + 0.423 + ], + "angle": 0, + "content": "(c) mIoU" + }, + { + "type": "image", + "bbox": [ + 0.274, + 0.34, + 0.416, + 0.41 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.337, + 0.411, + 0.371, + 0.425 + ], + "angle": 0, + "content": "(d) \\(E_{\\phi}\\)" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.436, + 0.47, + 0.464 + ], + "angle": 0, + "content": "Figure 6. Sensitivity analysis of the size \\( T \\) of the windowed sequence on MoCA-Mask dataset." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.475, + 0.47, + 0.687 + ], + "angle": 0, + "content": "Ablation analysis: Table 5 present the segmentation results by progressively applying each part in our model. Compared to the spatial-based self-prompt learning (SSP) comprising a spatial extractor and two prompt heads, motion-driven self-prompt learning (MSP) improves the mDice metric from \\(36.7\\%\\) to \\(40.7\\%\\) to a large degree. This affirms the effectiveness of the inter-frame motion in camouflage breaking. Subsequently, the long-range consistency-based robust prompt learning (LCP) brings \\(2.2\\%\\) gains for mDice, indicating the effectiveness of long-range consistency in improving the robustness of self-prompts. Meanwhile, the temporal-spatial injection (TSI) further boosts the mDice metric by \\(2.9\\%\\), once again verifying the significance of long-range consistency." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.688, + 0.468, + 0.839 + ], + "angle": 0, + "content": "To further analyze the effectiveness of frequency-based motion perception in MSP, we replace it with spatial-based motion perception, frequency-based static appearance perception, and spatial-based static appearance perception to conduct the ablation studies. As can be seen from Table 6, the frequency-based motion perception yields the best performance. While verifying the significant role of interframe motion in breaking camouflage, it also demonstrates that inter-frame motion can be expressed more exhaustively in the frequency domain than in the spatial domain." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.47, + 0.902 + ], + "angle": 0, + "content": "Hyper parameter analysis: To investigate the effect of the size of the windowed sequence on segmentation performance, we conduct the ablation experiments by varying \\( T \\) from \\( \\{5,6,7,8,9,10\\} \\). As illustrated in Fig. 6, the ac" + }, + { + "type": "table", + "bbox": [ + 0.503, + 0.089, + 0.89, + 0.169 + ], + "angle": 0, + "content": "
MethodSα↑Fwβ↑Eφ↑mDice↑mIoU↑
Spatial appearance0.6370.3300.7680.3670.311
Frequency appearance0.6470.3440.7930.3960.328
Spatial motion0.6510.3350.7300.3960.324
Frequency motion0.6570.3530.8230.4070.340
" + }, + { + "type": "table_caption", + "bbox": [ + 0.498, + 0.179, + 0.892, + 0.249 + ], + "angle": 0, + "content": "Table 6. Ablation studies of frequency-based motion perception on MoCA-Mask dataset. Spatial appearance: Spatial-based appearance perception. Frequency appearance: Frequency-based appearance perception. Spatial motion: Spatial-based motion perception. Frequency motion: Frequency-based motion perception." + }, + { + "type": "image", + "bbox": [ + 0.565, + 0.256, + 0.682, + 0.324 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.585, + 0.326, + 0.638, + 0.337 + ], + "angle": 0, + "content": "(a) mDice" + }, + { + "type": "image", + "bbox": [ + 0.72, + 0.257, + 0.837, + 0.324 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.759, + 0.326, + 0.797, + 0.338 + ], + "angle": 0, + "content": "(b) \\(F_{\\beta}^{w}\\)" + }, + { + "type": "image", + "bbox": [ + 0.566, + 0.34, + 0.683, + 0.408 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.587, + 0.409, + 0.636, + 0.42 + ], + "angle": 0, + "content": "(c) mIoU" + }, + { + "type": "image", + "bbox": [ + 0.721, + 0.339, + 0.837, + 0.408 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.759, + 0.408, + 0.792, + 0.423 + ], + "angle": 0, + "content": "(d) \\(E_{\\phi}\\)" + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.433, + 0.892, + 0.46 + ], + "angle": 0, + "content": "Figure 7. Sensitivity analysis of the balance factor \\((\\alpha, \\beta)\\) in the loss function on MoCA-Mask dataset." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.469, + 0.892, + 0.561 + ], + "angle": 0, + "content": "curacy initially rises with increasing the size of the windowed sequence, reaches the best at \\( T = 7 \\) before declining. Moreover, we assess the sensitivity of \\( \\alpha \\) and \\( \\beta \\) that are used to balance the sparse and dense prompt loss in Fig. 7. The optimal segmentation performance is observed at \\( (\\alpha, \\beta) = (1,0.6) \\)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.573, + 0.619, + 0.588 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.598, + 0.892, + 0.78 + ], + "angle": 0, + "content": "In this paper, we propose a flexible end-to-end temporal-spatial prompt SAM (TSP-SAM) for video camouflaged object detection. Instead of user-provided prompts, motion-driven self-prompt learning is proposed to capture the overall motion of the camouflaged object, facilitating its spatial identification. Subsequently, a robust prompt learning based on long-range consistency is proposed to promote the robustness of the self-prompts, which likes endowing SAM with keen eyes. Moreover, the long-range consistency is injected into the encoder of SAM for representation enhancement. Notably, TSP-SAM provides a new perspective, that is temporal-spatial associations, for visual prompt learning." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.785, + 0.681, + 0.802 + ], + "angle": 0, + "content": "7. Acknowledgement" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.81, + 0.892, + 0.9 + ], + "angle": 0, + "content": "This work was supported in part by the National Key Research and Development Program of China under Grant No.2021ZD0140407, Beijing Natural Science Foundation under Grant No.7222313, and National High Level Hospital Clinical Research Funding under Grant No.2022-PUMCHC-041." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19065" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.469, + 0.157 + ], + "angle": 0, + "content": "[1] Pia Bideau and Erik Learned-Miller. It's moving! a probabilistic model for causal motion segmentation in moving camera videos. In ECCV, pages 433-449, 2016. 1, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.159, + 0.47, + 0.2 + ], + "angle": 0, + "content": "[2] Pia Bideau, Rakesh R Menon, and Erik Learned-Miller. Moa-net: self-supervised motion segmentation. In ECCV Workshops, pages 0–0, 2018. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.202, + 0.471, + 0.244 + ], + "angle": 0, + "content": "[3] Alexey Bochkovskiy, Chien-Yao Wang, and Hong-Yuan Mark Liao. Yolov4: Optimal speed and accuracy of object detection. arXiv preprint arXiv:2004.10934, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.245, + 0.47, + 0.327 + ], + "angle": 0, + "content": "[4] Tianrun Chen, Lanyun Zhu, Chaotao Ding, Runlong Cao, Shangzhan Zhang, Yan Wang, Zejian Li, Lingyun Sun, Papa Mao, and Ying Zang. Sam fails to segment anything?—sam-adapter: Adapting sam in underperformed scenes: Camouflage, shadow, and more. arXiv preprint arXiv:2304.09148, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.329, + 0.47, + 0.397 + ], + "angle": 0, + "content": "[5] Xuelian Cheng, Huan Xiong, Deng-Ping Fan, Yiran Zhong, Mehrtash Harandi, Tom Drummond, and Zongyuan Ge. Implicit motion handling for video camouflaged object detection. In CVPR, pages 13864-13873, 2022. 1, 2, 4, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.401, + 0.47, + 0.455 + ], + "angle": 0, + "content": "[6] Runmin Cong, Mengyao Sun, Sanyi Zhang, Xiaofei Zhou, Wei Zhang, and Yao Zhao. Frequency perception network for camouflaged object detection. In ACM MM, pages 1179-1189, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.458, + 0.47, + 0.499 + ], + "angle": 0, + "content": "[7] Xinyi Cui, Qingshan Liu, and Dimitris Metaxas. Temporal spectral residual: fast motion saliency detection. In ACM MM, pages 617-620, 2009. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.501, + 0.469, + 0.556 + ], + "angle": 0, + "content": "[8] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, et al. An image is worth 16x16 words: Transformers for image recognition at scale. 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.559, + 0.469, + 0.6 + ], + "angle": 0, + "content": "[9] Deng-Ping Fan, Ming-Ming Cheng, Yun Liu, Tao Li, and Ali Borji. Structure-measure: A new way to evaluate foreground maps. In ICCV, pages 4548-4557, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.602, + 0.469, + 0.643 + ], + "angle": 0, + "content": "[10] Deng-Ping Fan, Ge-Peng Ji, Guolei Sun, Ming-Ming Cheng, Jianbing Shen, and Ling Shao. Camouflaged object detection. In CVPR, pages 2777-2787, 2020. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.645, + 0.469, + 0.714 + ], + "angle": 0, + "content": "[11] Deng-Ping Fan, Ge-Peng Ji, Tao Zhou, Geng Chen, Huazhu Fu, Jianbing Shen, and Ling Shao. Pranet: Parallel reverse attention network for polyp segmentation. In International conference on medical image computing and computer-assisted intervention, pages 263-273, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.716, + 0.469, + 0.757 + ], + "angle": 0, + "content": "[12] Deng-Ping Fan, Ge-Peng Ji, Ming-Ming Cheng, and Ling Shao. Concealed object detection. IEEE TPAMI, 44(10): 6024-6042, 2021. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.759, + 0.469, + 0.813 + ], + "angle": 0, + "content": "[13] Deng-Ping Fan, Ge-Peng Ji, Xuebin Qin, and Ming-Ming Cheng. Cognitive vision inspired object segmentation metric and loss function. Scientia Sinica Informationis, 6(6), 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.816, + 0.469, + 0.857 + ], + "angle": 0, + "content": "[14] Melvyn A Goodale and A David Milner. Separate visual pathways for perception and action. Trends in neurosciences, 15(1):20-25, 1992. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.859, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[15] Kensho Hara, Hirokatsu Kataoka, and Yutaka Satoh. Can spatiotemporal 3d cnns retrace the history of 2d cnns and imagenet? In CVPR, pages 6546-6555, 2018. 5" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.471, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.147 + ], + "angle": 0, + "content": "[16] Chunming He, Kai Li, Yachao Zhang, Longxiang Tang, Yulin Zhang, Zhenhua Guo, and Xiu Li. Camouflaged object detection with feature decomposition and edge reconstruction. In CVPR, pages 22046-22055, 2023. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.149, + 0.892, + 0.19 + ], + "angle": 0, + "content": "[17] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, pages 770-778, 2016. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.191, + 0.892, + 0.245 + ], + "angle": 0, + "content": "[18] Yuhao Huang, Xin Yang, Lian Liu, Han Zhou, Ao Chang, Xinrui Zhou, Rusi Chen, Junxuan Yu, Jiongquan Chen, Chaoyu Chen, et al. Segment anything model for medical images? arXiv preprint arXiv:2304.14660, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.247, + 0.892, + 0.3 + ], + "angle": 0, + "content": "[19] Zhou Huang, Hang Dai, Tian-Zhu Xiang, Shuo Wang, Huai-Xin Chen, Jie Qin, and Huan Xiong. Feature shrinkage pyramid for camouflaged object detection with transformers. In CVPR, pages 5557-5566, 2023. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.302, + 0.892, + 0.383 + ], + "angle": 0, + "content": "[20] Ge-Peng Ji, Yu-Cheng Chou, Deng-Ping Fan, Geng Chen, Huazhu Fu, Debesh Jha, and Ling Shao. Progressively normalized self-attention network for video polyp segmentation. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 142–152, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.385, + 0.892, + 0.44 + ], + "angle": 0, + "content": "[21] Ge-Peng Ji, Deng-Ping Fan, Yu-Cheng Chou, Dengxin Dai, Alexander Liniger, and Luc Van Gool. Deep gradient learning for efficient camouflaged object detection. Machine Intelligence Research, 20(1):92-108, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.442, + 0.892, + 0.496 + ], + "angle": 0, + "content": "[22] Ge-Peng Ji, Deng-Ping Fan, Peng Xu, Ming-Ming Cheng, Bowen Zhou, and Luc Van Gool. \"Sam struggles in concealed scenes—empirical study on\" segment anything\". arXiv preprint arXiv:2304.06022, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.498, + 0.892, + 0.551 + ], + "angle": 0, + "content": "[23] Yang Jia, Jie Yuan, Jinjun Wang, Jun Fang, Qixing Zhang, and Yongming Zhang. A saliency-based method for early smoke detection in video sequences. Fire technology, 52: 1271-1292, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.553, + 0.892, + 0.58 + ], + "angle": 0, + "content": "[24] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. ICLR, 2015. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.581, + 0.892, + 0.636 + ], + "angle": 0, + "content": "[25] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. arXiv preprint arXiv:2304.02643, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.637, + 0.892, + 0.677 + ], + "angle": 0, + "content": "[26] Hala Lamdouar, Charig Yang, Weidi Xie, and Andrew Zisserman. Betrayed by motion: Camouflaged object discovery via motion segmentation. In ACCV, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.679, + 0.892, + 0.719 + ], + "angle": 0, + "content": "[27] Fuxin Li, Taeyoung Kim, Ahmad Humayun, David Tsai, and James M Rehg. Video segmentation by tracking many figure-ground segments. In ICCV, pages 2192-2199, 2013. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.721, + 0.892, + 0.76 + ], + "angle": 0, + "content": "[28] Thomas Lidbetter. Search and rescue in the face of uncertain threats. European Journal of Operational Research, 285(3): 1153-1160, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.762, + 0.892, + 0.816 + ], + "angle": 0, + "content": "[29] Ting Liu, Yao Zhao, Yunchao Wei, Yufeng Zhao, and Shikui Wei. Concealed object detection for activate millimeter wave image. IEEE Transactions on Industrial Electronics, 66(12): 9909-9917, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.818, + 0.892, + 0.858 + ], + "angle": 0, + "content": "[30] Jun Ma, Yuting He, Feifei Li, Lin Han, Chenyu You, and Bo Wang. Segment anything in medical images. arXiv preprint arXiv:2304.12306, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.86, + 0.892, + 0.899 + ], + "angle": 0, + "content": "[31] Ran Margolin, Lihi Zelnik-Manor, and Ayellet Tal. How to evaluate foreground maps? In CVPR, pages 248-255, 2014. 5" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "19066" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.147 + ], + "angle": 0, + "content": "[32] Youwei Pang, Xiaoqi Zhao, Tian-Zhu Xiang, Lihe Zhang, and Huchuan Lu. Zoom in and out: A mixed-scale triplet network for camouflaged object detection. In CVPR, pages 2160-2170, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.149, + 0.47, + 0.204 + ], + "angle": 0, + "content": "[33] Federico Perazzi, Philipp Krahenbuhl, Yael Pritch, and Alexander Hornung. Saliency filters: Contrast based filtering for salient region detection. In CVPR, pages 733-740, 2012. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.206, + 0.469, + 0.274 + ], + "angle": 0, + "content": "[34] Federico Peruzzi, Jordi Pont-Tuset, Brian McWilliams, Luc Van Gool, Markus Gross, and Alexander Sorkine-Hornung. A benchmark dataset and evaluation methodology for video object segmentation. In CVPR, pages 724-732, 2016. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.277, + 0.469, + 0.33 + ], + "angle": 0, + "content": "[35] Rui Qian, Shuangrui Ding, Xian Liu, and Dahua Lin. Semantics meets temporal correspondence: Self-supervised object-centric learning in videos. In ICCV, pages 16675-16687, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.333, + 0.469, + 0.374 + ], + "angle": 0, + "content": "[36] Yujia Sun, Shuo Wang, Chenglizhao Chen, and Tian-Zhu Xiang. Boundary-guided camouflaged object detection. arXiv preprint arXiv:2207.00794, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.376, + 0.469, + 0.417 + ], + "angle": 0, + "content": "[37] Lv Tang, Haoke Xiao, and Bo Li. Can sam segment anything? when sam meets camouflaged object detection. arXiv preprint arXiv:2304.04709, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.419, + 0.469, + 0.46 + ], + "angle": 0, + "content": "[38] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. NeurIPS, 30, 2017. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.462, + 0.469, + 0.529 + ], + "angle": 0, + "content": "[39] Wenhai Wang, Enze Xie, Xiang Li, Deng-Ping Fan, Kaitao Song, Ding Liang, Tong Lu, Ping Luo, and Ling Shao. Pyramid vision transformer: A versatile backbone for dense prediction without convolutions. In ICCV, pages 568-578, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.532, + 0.469, + 0.599 + ], + "angle": 0, + "content": "[40] Junde Wu, Rao Fu, Huihui Fang, Yuanpei Liu, Zhaowei Wang, Yanwu Xu, Yueming Jin, and Tal Arbel. Medical sam adapter: Adapting segment anything model for medical image segmentation. arXiv preprint arXiv:2304.12620, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.602, + 0.469, + 0.657 + ], + "angle": 0, + "content": "[41] Yu-Huan Wu, Shang-Hua Gao, Jie Mei, Jun Xu, Deng-Ping Fan, Rong-Guo Zhang, and Ming-Ming Cheng. Jcs: An explainable Covid-19 diagnosis system by joint classification and segmentation. IEEE TIP, 30:3113-3126, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.659, + 0.469, + 0.713 + ], + "angle": 0, + "content": "[42] Pengxiang Yan, Guanbin Li, Yuan Xie, Zhen Li, Chuan Wang, Tianshui Chen, and Liang Lin. Semi-supervised video salient object detection using pseudo-labels. In ICCV, pages 7284-7293, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.716, + 0.469, + 0.769 + ], + "angle": 0, + "content": "[43] Charig Yang, Hala Lamdouar, Erika Lu, Andrew Zisserman, and Weidi Xie. Self-supervised video object segmentation by motion grouping. In ICCV, pages 7177-7188, 2021. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.772, + 0.469, + 0.813 + ], + "angle": 0, + "content": "[44] Yi Zhang, Jing Zhang, Wassim Hamidouche, and Olivier Deforges. Predictive uncertainty estimation for camouflaged object detection. IEEE TIP, 2023. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.814, + 0.469, + 0.868 + ], + "angle": 0, + "content": "[45] Yijie Zhong, Bo Li, Lv Tang, Senyun Kuang, Shuang Wu, and Shouhong Ding. Detecting camouflaged object in frequency domain. In CVPR, pages 4504-4513, 2022. 1, 2, 3" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.868 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19067" + } + ] +] \ No newline at end of file diff --git a/2024/Endow SAM with Keen Eyes_ Temporal-spatial Prompt Learning for Video Camouflaged Object Detection/e8cb1a37-360c-44d3-92de-d83afed15e31_origin.pdf b/2024/Endow SAM with Keen Eyes_ Temporal-spatial Prompt Learning for Video Camouflaged Object Detection/e8cb1a37-360c-44d3-92de-d83afed15e31_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..97b8c7f331c3f3b771c7537ce92a8545b7a00639 --- /dev/null +++ b/2024/Endow SAM with Keen Eyes_ Temporal-spatial Prompt Learning for Video Camouflaged Object Detection/e8cb1a37-360c-44d3-92de-d83afed15e31_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a863711261a73a277c1d384d692afb44b2e0c28752dbc2b8acdb21087c72f67f +size 5117356 diff --git a/2024/Endow SAM with Keen Eyes_ Temporal-spatial Prompt Learning for Video Camouflaged Object Detection/full.md b/2024/Endow SAM with Keen Eyes_ Temporal-spatial Prompt Learning for Video Camouflaged Object Detection/full.md new file mode 100644 index 0000000000000000000000000000000000000000..837bcf55fc488629f0f8a1eac0df8a7222a7aabb --- /dev/null +++ b/2024/Endow SAM with Keen Eyes_ Temporal-spatial Prompt Learning for Video Camouflaged Object Detection/full.md @@ -0,0 +1,376 @@ +# Endow SAM with Keen Eyes: Temporal-spatial Prompt Learning for Video Camouflaged Object Detection + +Wenjun Hui $^{1,2}$ , Zhenfeng Zhu $^{1,2*}$ , Shuai Zheng $^{1,2}$ , Yao Zhao $^{1,2}$ + +$^{1}$ Institute of Information Science, Beijing Jiaotong University + +$^{2}$ Beijing Key Laboratory of Advanced Information Science and Network Technology + +{wenjunhui, zhfzhu, zs1997, yzhao}@bjtu.edu.cn + +# Abstract + +The Segment Anything Model (SAM), a prompt-driven foundational model, has demonstrated remarkable performance in natural image segmentation. However, its application in video camouflaged object detection (VCOD) encounters challenges, chiefly stemming from the overlooked temporal-spatial associations and the unreliability of user-provided prompts for camouflaged objects that are difficult to discern with the naked eye. To tackle the above issues, we endow SAM with keen eyes and propose the Temporal-spatial Prompt SAM (TSP-SAM), a novel approach tailored for VCOD via an ingenious prompted learning scheme. Firstly, motion-driven self-prompt learning is employed to capture the camouflaged object, thereby bypassing the need for user-provided prompts. With the detected subtle motion cues across consecutive video frames, the overall movement of the camouflaged object is captured for more precise spatial localization. Subsequently, to eliminate the prompt bias resulting from inter-frame discontinuities, the long-range consistency within the video sequences is taken into account to promote the robustness of the self-prompts. It is also injected into the encoder of SAM to enhance the representational capabilities. Extensive experimental results on two benchmarks demonstrate that the proposed TSP-SAM achieves a significant improvement over the state-of-the-art methods. With the mIoU metric increasing by $7.8\%$ and $9.6\%$ , TSP-SAM emerges as a groundbreaking step forward in the field of VCOD. + +# 1. Introduction + +Camouflaged object detection, a pivotal task in computer vision, aims to identify the target object seamlessly blending with its surroundings in images [19, 44]. This task plays an important role in various fields, including surveillance and security [29], wildlife protection [28], and medi + +![](images/40cbf0bea7fdfbc6bbbde8bd94b8f24a9255d95c1bc72d1a3a054d7cb961ea74.jpg) + +![](images/476016f2f056582f85d7851cf2e0cfbe1301c4e26b916b14a9285870bd52ad42.jpg) +Figure 1. Illustration of previous VCOD methods (a-b) and the proposed TSP-SAM (c). In contrast to previous methods with two-stage architecture, the proposed TSP-SAM is trained in an end-to-end manner, incorporating short-long term temporal-spatial relationships to collaboratively learn the reliable prompts. Meanwhile, the long-range consistency is also used to enhance the representational capability of SAM. + +cal image processing [11, 41]. In recent years, single-image camouflaged object detection has made significant progress [16, 19, 45]. However, studies [26, 43] have discerned that observers exhibit heightened perception towards the camouflaged object in motion compared to the static scenarios. This observation has sparked a growing interest in video camouflaged object detection (VCOD) [5]. + +Existing VCOD methods [1, 2, 5, 26] commonly employ a two-step framework to model the inter-frame correlations implicitly or explicitly for camouflage breaking, as shown in Fig. 1(a)(b). Despite yielding promising results, they are noticeably limited in certain aspects: + +- (i) Primarily, their two-stage architecture exposes them to cumulative errors, thus impacting their overall accuracy. +- (ii) Moreover, they exhibit weak generalization ability due to the limited training data. + +Consequently, there is a growing demand for universal models in video camouflaged object detection. + +With the emergence and development of the visual foundation model, the segment anything model (SAM) [25] has showcased unprecedented performance in the field of natural image segmentation. This accomplishment is attributed to the impressive representations learned on 11 million images in SA-1B [25] and 636 million parameters of ViT-H [8]. SAM not only exhibits heightened versatility in terms of model capacity, but also potentially yields more consistent results across different tasks [30]. Inspired by the remarkable generality of SAM, this work endeavors to make SAM suitable for VCOD. + +However, it is worth noting that the performance of SAM is uncertain when applied to VCOD due to the differences between the camouflaged images and natural images. Specifically, SAM, as a prompt-driven foundational model, requires extra prompts when segmenting specific regions. Yet, the seamless blending of the camouflaged object with its surroundings renders it imperceptible to the naked eye, posing a formidable challenge in providing reliable prompts. Hence, there arises an urgent need for adaptive prompt learning to the camouflaged object. + +It is well-known that the temporal-spatial interplay in video sequences plays a pivotal role in video-related tasks. Short-term temporal-spatial relationships accentuate the variations between frames, while long-term temporal-spatial correlations model the temporal-spatial context to characterize the dependencies between the target object and the background. Motivated by these considerations, in this work, we develop a flexible end-to-end temporal-spatial prompt SAM, called TSP-SAM, for VCOD. To the best of our knowledge, the proposed TSP-SAM is the first SAM-based camouflaged object detection framework. + +Our main contributions can be summarized as follows: + +- Instead of user-provided prompts, we propose motion-driven self-prompt learning to capture the camouflaged object. It perceives the overall motion of the camouflaged object by establishing inter-frame associations in the frequency domain, facilitating its spatial identification. +- To eliminate the prompt bias stemming from inter-frame discontinuities, robust prompt learning based on long-range consistency is proposed. By modeling the long-range temporal-spatial dependencies within video sequences, the robustness of self-prompts is well promoted. +- In order to enhance the representational capability of SAM, we propose temporal-spatial injection that incorporates the long-range temporal-spatial consistency into the encoder of SAM. This enhancement contributes to more precise detection within the SAM framework. +- Extensive experiments on MoCA-Mask and CAD2016 datasets demonstrate that the proposed TSP-SAM achieves substantial performance improvements over the state-of-the-art method, with mIoU metric increasing by $7.8\%$ and $9.6\%$ . + +# 2. Related work + +# 2.1. Camouflaged object detection + +Camouflaged object detection (COD) aims to identify objects blending into their surroundings in images. Existing CNN-based methods address this task with various strategies. Fan et al. [10, 12] introduced a two-stage process, involving a search stage for localization followed by a segmentation stage for refinement. In multi-task learning, Yang et al. [43] and Ji et al. [21] incorporated an auxiliary task to derive shared context representations for enhancing the COD task. Other strategies, such as frequency cues, have also been explored. These works [6, 45] combined the static frequency spectrum with the spatial representations to discern the subtle differences between the camouflaged object and its surroundings. In contrast to the static frequency perception, our TSP-SAM perceives the frequency energy variants to depict the implicit motion between frames, facilitating the spatial identification of the camouflaged object. + +# 2.2. Video camouflaged object detection + +Compared to the static appearance of images, the motion between video frames is deemed a significant clue for breaking camouflage. Existing video camouflaged object detection methods [1, 2, 5, 26] implicitly or explicitly modeled the temporal-spatial correlations through a two-stage architecture to detect camouflaged objects. Explicit motion-based methods [1, 2, 26] segmented the optical flow field after compensation or registration into object and background. In contrast, implicit feature correspondences between frames are learned by the neural network for an effective alignment [5]. Nonetheless, their two-stage architecture leads to cumulative errors. Different from the two-stage architecture, our TSP-SAM is trained in an end-to-end manner, incorporating the short-long term temporal-spatial relationships to collaboratively learn reliable prompts for SAM. + +# 2.3. Segment anything model + +The segment anything model (SAM) [25], a prompt-driven large-scale foundation model, has exhibited unprecedented performance in natural image segmentation. The core components of SAM comprise an image encoder, a flexible prompt encoder and a lightweight mask decoder. Despite exhibiting excellent zero-shot segmentation capability, SAM performs unstable in application-specific tasks [4]. It fails to segment medical images [18], camouflaged objects [37] and concealed scenes [22]. To apply SAM to medical images, MedSAM [30] and SAM-Adapter [4, 40] incorporated specific domain knowledge into SAM. Yet, up to now, there has been no work on adapting SAM to be tailored for camouflaged object detection. To fill this gap, TSP-SAM explores the temporal-spatial relationships, presenting the first SAM-based camouflaged object detection framework. + +![](images/22814ad2fb56a81c74495463aadf4300e0bee85d21d9f69c6e6b719162192d03.jpg) +Figure 2. The framework of the proposed TSP-SAM. In motion-driven self-prompt learning (b), the implicit motion between consecutive frames is perceived to facilitate the spatial identification of the camouflaged object, thereby learning the dense self-prompt $P_{d}(t)$ and sparse self-prompt $P_{s}(t)$ . Subsequently, in robust prompt learning (c), the long-range consistency $X_{c}(t)$ is modeled to refine the sparse self-prompt $P_{s}(t)$ . Moreover, in temporal-spatial injection (d), the long-range consistency $X_{c}(t)$ is injected into the image embedding $X_{e}(t)$ of SAM for representation enhancement. During the training phase, SAM (a) is frozen, including the image encoder $E_{i}$ , prompt encoder $E_{p}$ and the mask decoder $E_{m}$ . + +# 3. Overview of the proposed TSP-SAM + +# 3.1. Notations + +In the subsequent sections, we establish the key notations for clarity and consistency in our discourse. Given a sequence of camouflaged images, denote by $I(t) \in \mathbb{R}^{H \times W \times 3}$ the $t$ -th frame and $I^w(t) = [I(t - T + 1), \dots, I(t)] \in \mathbb{R}^{T \times H \times W \times 3}$ the corresponding windowed sequence containing $T$ frames. For a tensor $A \in \mathbb{R}^{H \times W \times C}$ , $\mathrm{Mat}(A) \in \mathbb{R}^{HW \times C}$ denotes its matrixization by spatial-wise concatenation. For a matrix $B \in \mathbb{R}^{HW \times C}$ , $\mathrm{Ten}(B) \in \mathbb{R}^{H \times W \times C}$ represents its tensorization by spatial-wise division. + +# 3.2. Framework + +Fig. 2 illustrates the framework of the proposed TSP-SAM. The core components consist mainly of the following three modules: + +(a) Motion-driven self-prompt learning. The motion-driven self-prompt learning is to use the implicit inter-frame motion $\hat{X}_{(t - 1)\rightarrow t}$ in the frequency domain to facilitate the spatial identification of the camouflaged object, thereby learning the self-prompts for SAM. +(b) Robust prompt learning based on long-range consistency. To eliminate the prompt bias stemming from underlying inter-frame discontinuities, the long-range temporal-spatial consistency $X_{c}(t)$ in video sequences is modeled to promote the robustness of the self-prompts. +(c) Temporal-spatial injection for representation enhancement. To enhance the representational capabilities of SAM, the long-range temporal-spatial consistency $X_{c}(t)$ is injected into the image embedding $X_{e}(t)$ of SAM, con + +attributing to more precise detection. + +# 4. Methodology + +In this section, we detail the proposed TSP-SAM method for video camouflaged object detection. + +# 4.1. Motion-driven self-prompt learning + +It is well-known that SAM requires user-provided prompts when segmenting specific regions. Nevertheless, the difficulty in distinguishing the camouflaged object with the naked eye makes the requirement of providing visual prompts a significant challenge. Consequently, a compelling need arises to devise a prompt learning network, capable of adaptively identifying the camouflaged object. + +# 4.1.1 Frequency-based motion perception + +Inspired by the two-stream hypothesis in neuroscience [14], which highlights the importance of motion in visual perception, we consider the motion as a crucial clue for breaking camouflage. Building on this premise, several studies [7, 23] indicate the motion between frames can be expressed exhaustively in the frequency domain through the variants in frequency energy. Hence, frequency-based motion perception is designed to facilitate the spatial identification of the camouflaged object. + +Specifically, as shown in Fig. 3, given consecutive video frames $I(t - 1) \in \mathbb{R}^{H \times W \times 3}$ and $I(t) \in \mathbb{R}^{H \times W \times 3}$ , we first map them into the frequency domain. Following [45], the image features are divided into a set of $s \times s$ patches and each patch is processed by Discrete Cosine Transform + +![](images/b4f7591e9f45f69f29836617da8650d78e423ebde20e8c56290f26a1ed158b86.jpg) +Figure 3. The framework of the motion-driven self-prompt learning. The spatial representations $\{X_{b}(t)\}_{b = 1}^{N_{b}}$ are extracted by PVT [39], where $N_{b}$ is the number of attention blocks. + +(DCT) into frequency spectrum, where $s$ is the size of patches. To group all components of the same frequency into one channel, we flatten and reshape the spectrum of these patches, forming the frequency features $\hat{I}(t - 1) \in \mathbb{R}^{\frac{H}{s} \times \frac{W}{s} \times 3s^2}$ and $\hat{I}(t) \in \mathbb{R}^{\frac{H}{s} \times \frac{W}{s} \times 3s^2}$ . + +In order to perceive the overall motion of the camouflaged object, the inter-frame associations are established on high frequency and low frequency bands respectively. Specifically, the frequency spectrum $\{\hat{I} (t - 1),\hat{I} (t)\}$ is divided into two groups in the channel dimension, namely low frequency bands $\{\hat{I}^l (t - 1),\hat{I}^l (t)\}$ and high frequency bands $\{\hat{I}^h (t - 1),\hat{I}^h (t)\}$ . Hence, we have: + +$$ +\hat {X} _ {(t - 1) \rightarrow t} ^ {l} = \mathrm {M L P} \big (\operatorname {A T T N} (\operatorname {M a t} (\hat {I} ^ {l} (t - 1)), \operatorname {M a t} (\hat {I} ^ {l} (t))) \big) \quad (1) +$$ + +$$ +\hat {X} _ {(t - 1) \rightarrow t} ^ {h} = \operatorname {M L P} (\operatorname {A T T N} (\operatorname {M a t} (\hat {I} ^ {h} (t - 1)), \operatorname {M a t} (\hat {I} ^ {h} (t)))) \tag {2} +$$ + +where $\hat{X}_{(t - 1)\rightarrow t}^{l}\in \mathbb{R}^{\frac{HW}{s^2}\times \frac{3s^2}{2}}$ and $\mathrm{MLP}(\cdot)$ is the linear layer. $\mathrm{ATTN}(\cdot ,\cdot)$ is multi-head cross attention [38] to extract the inter-frame correspondences, in which the $(t - 1)$ -th frame is taken as query and the $t$ -th frame serves as key and value. Note that Eq.1 tends to highlight the global deformations of the camouflaged object by characterizing the feature correspondences on the low-frequency bands. Meanwhile, Eq.2 implicitly reveals the local edge motion of the camouflaged object on high-frequency bands. + +Subsequently, the complete implicit motion in the frequency domain is obtained by: + +$$ +\hat {X} _ {(t - 1) \rightarrow t} = \hat {I} (t) + \operatorname {C B A} (\operatorname {C a t} \left(\operatorname {T e n} \left(\hat {X} _ {(t - 1) \rightarrow t} ^ {l}\right), \operatorname {T e n} \left(\hat {X} _ {(t - 1) \rightarrow t} ^ {h}\right)\right)) \tag {3} +$$ + +where $\mathrm{CBA}(\cdot)$ is the convolution operation followed by batch normalization and activation and $\operatorname{Cat}(\cdot)$ is the channel-wise concatenation. Importantly, $\hat{X}_{(t - 1)\rightarrow t}\in$ $\mathbb{R}^{\frac{H}{s}\times \frac{W}{s}\times 3s^{2}}$ implicitly depicts the overall motion of the camouflaged object by perceiving the variants in the frequency energy, enabling to facilitate the discovery of the camouflaged object. + +# 4.1.2 Implicit motion-induced spatial-frequency fusion + +To translate the perceived motion into the spatial domain, the implicit motion-induced spatial-frequency representation fusion is established. This process involves the integration between the implicit motion $\hat{X}_{(t - 1)\rightarrow t}$ in the frequency domain and the spatial representations $X_{b}(t)\in$ $\mathbb{R}^{\frac{H}{2^{b + 1}}\times \frac{W}{2^{b + 1}}\times C},b\in \{1,\dots ,N_b\}$ extracted by PVT [39], where $N_{b}$ is the number of attention blocks. It can be expressed as follows: + +$$ +\hat {X} _ {b} (t) = X _ {b} (t) + \operatorname {T e n} (\operatorname {A T T N} (\operatorname {M a t} (\operatorname {C B A} (\hat {X} _ {(t - 1) \rightarrow t})), \operatorname {M a t} (X _ {b} (t)))) \tag {4} +$$ + +where $\hat{X}_b(t) \in \mathbb{R}^{\frac{H}{2^b + 1} \times \frac{W}{2^b + 1} \times C}$ . Note that the spatial perception of the camouflaged object could be strengthened by establishing cross attention between the implicit motion as query and the spatial representation as key and value. + +# 4.1.3 Self-prompt learning + +Based on the co-representations $\{\hat{X}_b(t)\}_{b=1}^{N_b}$ , two prompt heads are adopted to generate the dense self-prompt $P_d(t)$ and sparse self-prompt $P_s(t)$ , respectively. Specifically, we adopt the decoder of SLTNet [5] as dense prompt head $E_d$ as follows: + +$$ +P _ {d} (t) = E _ {d} \left(\left\{\hat {X} _ {b} (t) \right\} _ {b = 1} ^ {N _ {b}}\right) \tag {5} +$$ + +where $P_{d}(t)$ is the dense self-prompt. Meanwhile, as illustrated in Fig. 3, to obtain the sparse prompt, we adopt the residual block of Resnet [17] to process the co-representations in a bottom-up manner: + +$$ +\tilde {X} _ {b} (t) = \left\{ \begin{array}{l l} \operatorname {R e s b l o c k} \left(\hat {X} _ {b} (t)\right) & b = N _ {b} \\ \operatorname {R e s b l o c k} \left(\hat {X} _ {b} (t) + \tilde {X} _ {b + 1} (t)\right) & b \in \{1, \dots , N _ {b} - 1 \} \end{array} \right. \tag {6} +$$ + +Finally, the sparse self-prompt is represented as follows: + +$$ +P _ {s} (t) = \operatorname {A v g p o o l} \left(\operatorname {C B A} \left(\tilde {X} _ {1} (t)\right)\right) \tag {7} +$$ + +where $\mathrm{Avgpool}(\cdot)$ is global average pooling operation. + +# 4.2. Robust prompt learning based on long-range consistency + +The self-prompts of the camouflaged object are well learned by using the perceived inter-frame motion to facilitate its spatial identification. However, the susceptibility of inter-frame motion to underlying temporal-spatial discontinuities, such as occlusions, shaking, scene changes, etc., brings a potential lack of reliability in the learned self-prompts. It is well known that the long-range dependencies within the video sequences enable to characterize the temporal-spatial context to cope with the short-term temporal-spatial discontinuities. Hence, it is necessary to explore the long-range temporal-spatial consistency within the video sequences to improve the robustness of the self-prompts, which likes endowing SAM with keen eyes. + +# 4.2.1 Long-range consistency squeezing + +As illustrated in Fig. 2(c), given the windowed sequence $I^{w}(t) \in \mathbb{R}^{T \times H \times W \times 3}$ corresponding to the current frame $I(t)$ , we adopt a 3D convolution neural network $E_{t}$ to extract the long-range temporal-spatial relationships: + +$$ +X ^ {w} (t) = E _ {t} \left(I ^ {w} (t)\right) \tag {8} +$$ + +where $X^w(t) \in \mathbb{R}^{T \times \frac{H}{4} \times \frac{W}{4} \times C}$ fully reveals the temporal-spatial dependencies within the video sequences. In our work, we adopt 3D Resnet18 [15] as $E_t$ . + +Subsequently, to further extract the long-range temporal-spatial consistency of the camouflaged object, we take the camouflage representation $X_{e}(t)$ from the encoder of SAM as query and the long-range temporal-spatial dependencies $X^{w}(t)$ averaged along the temporal dimension as key and value to establish the cross attention: + +$$ +X _ {c} (t) = \operatorname {A T T N} \left(\operatorname {M a t} \left(X _ {e} (t)\right), \operatorname {M a t} \left(\operatorname {A v g} _ {\mathrm {t}} \left(X ^ {w} (t)\right)\right)\right) \tag {9} +$$ + +where $\mathrm{Avg}_{\mathrm{t}}(\cdot)$ is an average operation to squeeze the temporal dimension. $X_{c}(t)\in \mathbb{R}^{\frac{HW}{16}\times C}$ characterizes the long-range consistency specific to the camouflaged object. + +# 4.2.2 Robust prompt learning + +To enhance the robustness of the self-prompts to inter-frame temporal-spatial discontinuities, the long-range temporal-spatial consistency $X_{c}(t)$ should be induced into the prompt learning. Specifically, in order to map the long-range consistency into the coordinate space, a learnable query $Q_{c}(t) \in \mathbb{R}^{4 \times C}$ is defined to establish cross attention with $X_{c}(t)$ . It is formulated as: + +$$ +P _ {r} (t) = \operatorname {A v g} _ {\mathrm {c}} (\operatorname {A T N} \left(Q _ {c} (t), X _ {c} (t)\right)) \tag {10} +$$ + +where $\mathrm{Avg}_c(\cdot)$ is an average operation along the channel dimension. $P_r(t)\in \mathbf{R}^4$ is a refinement term derived from long-range consistency to correct the sparse self-prompt: + +$$ +P _ {s} ^ {r} (t) = P _ {s} (t) + P _ {r} (t) \tag {11} +$$ + +Through incorporating long-range consistency into the prompt learning, the visual prompts could be more robust to the temporal-spatial discontinuities, which means endowing SAM with a pair of keen eyes. + +# 4.3. Temporal-spatial injection for representation enhancement + +It is known that the remarkable performance of SAM relies not only on well learned robust prompts but also on the representational capability of the image encoder. However, due to the fact that the image embedding of SAM lacks the temporal-spatial information, the final predictions of consecutive frames may be inconsistent. Hence, it is necessary + +to inject the long-range temporal-spatial consistency into SAM. As shown in Fig. 2(d), it is represented as follows: + +$$ +\hat {X} _ {e} (t) = X _ {e} (t) + \operatorname {T e n} \left(X _ {c} (t)\right) \tag {12} +$$ + +where $X_{c}(t)$ is the long-range consistency from Eq. 9. In this way, based on the enhanced image representations and the robust prompts called keen eyes, SAM achieves more precise camouflaged object detection in videos. + +# 4.4. Model optimization + +We train the proposed TSP-SAM in an end-to-end manner by minimizing the joint loss below: + +$$ +\mathcal {L} (t) = \mathcal {L} _ {S A M} (t) + \alpha \mathcal {L} _ {p} ^ {s} (t) + \beta \mathcal {L} _ {p} ^ {d} (t) \tag {13} +$$ + +where $\mathcal{L}_{SAM},\mathcal{L}_p^s$ and $\mathcal{L}_p^d$ supervise the final prediction, the sparse prompt and the dense prompt, respectively. + +Following [3], the supervision on the sparse prompt $\mathcal{L}_p^s (t)$ is as follows: + +$$ +\mathcal {L} _ {p} ^ {s} (t) = \operatorname {B C E} \left(P _ {s} ^ {r} (t), G (t)\right) + \operatorname {M S E} \left(P _ {s} ^ {r} (t), G (t)\right) \tag {14} +$$ + +where BCE is binary cross-entropy loss to supervise the coordinates and MSE is mean square error to supervise the size of the box prompt or the scale of the point prompt. For the dense prompt loss $\mathcal{L}_p^d$ and the final prediction loss $\mathcal{L}_{SAM}$ , we adopt the hybrid loss [13] for supervision: + +$$ +\mathcal {L} _ {\text {h y b r i d}} = \mathcal {L} _ {c e} ^ {w} + \mathcal {L} _ {i o u} ^ {w} \tag {15} +$$ + +where $\mathcal{L}_{ce}^{w}$ is the weighted cross-entropy loss and $\mathcal{L}_{iou}^{w}$ is the weighted intersection-over-union loss. + +# 5. Experiment results and analysis + +# 5.1. Experiment settings + +Datasets. Following previous methods [5], we have conducted a comprehensive evaluation on two widely used datasets: MoCA-Mask [5] and CAD2016 [1]. The MoCA-Mask dataset [5] is a video camouflaged object detection dataset, covering camouflaged animals moving in natural scenes. It contains 87 video sequences (22939 frames), of which 71 video sequences are used for training and 16 sequences for testing. It is well annotated with bounding boxes and pixel-level segmentation masks on every fifth frames. Meanwhile, Camouflaged Animal Dataset [1] (CAD2016) is a small video camouflaged object detection dataset extracted from YouTube videos. It consists of nine short video clips in total and is well annotated with pixel-level semantic masks on every fifth frames. + +Metrics. During the testing phase, we assess the prediction masks using six widely-used metrics: S-measure $(S_{\alpha})$ [9], Weighted F-measure $(F_{\beta}^{w})$ [31], Enhanced-alignment measure $(E_{\phi})$ [13], Mean absolute error (MAE, $\mathcal{M}$ ) [33], + +
MethodYearInputMoCA-MaskCAD2016
\(S_{\alpha}\uparrow\)\(F^{\text{w}}_{\beta}\uparrow\)\(E_{\phi}\uparrow\)\(\mathcal{M}\downarrow\)mDice↑mIoU↑\(S_{\alpha}\uparrow\)\(F^{\text{w}}_{\beta}\uparrow\)\(E_{\phi}\uparrow\)\(\mathcal{M}\downarrow\)mDice↑mIoU↑
SINet[10]2020-CVPRImage0.5740.1850.6550.0300.2210.1560.6010.2040.5890.0890.2890.209
SINet-v2[12]2021-TPAMIImage0.5710.1750.6080.0350.2110.1530.5440.1810.5460.0490.1700.110
ZoomNet[32]2022-CVPRImage0.5820.2110.5360.0330.2240.1670.5870.2250.5940.0630.2460.166
BGNet [36]2022-IJCAIImage0.5900.2030.6470.0230.2250.1670.6070.2030.6660.0890.3450.256
FEDERNet[16]2023-CVPRImage0.5550.1580.5420.0490.1920.1320.6070.2460.7250.0610.3610.257
FSPNet[19]2023-CVPRImage0.5940.1820.6080.0440.2380.1670.5390.2200.5530.1450.3090.212
PUENet[44]2023-TIPImage0.5940.2040.6190.0370.3020.2120.6730.4270.8030.0340.4990.389
RCRNet[42]2019-ICCVVideo0.5970.1740.5830.0250.1940.137------
PNS-Net[20]2021-MICCAIVideo0.5760.1340.5620.0380.1890.1330.6780.3690.7200.0430.4090.308
MG[43]2021-ICCVVideo0.5470.1650.5370.0950.1970.1410.4840.3140.5580.3700.3510.260
SLT-Net[5]2022-CVPRVideo0.6560.3570.7850.0210.3870.3100.6790.4200.8050.0330.4450.342
TSP-SAM(M+P)OursVideo0.6730.4000.7660.0120.4210.3450.6810.5000.8530.0310.4960.393
TSP-SAM(M+B)OursVideo0.6890.4440.8080.0080.4580.3880.7040.5240.9120.0280.5430.438
+ +Table 1. Quantitative comparisons over two benchmark datasets. The top three results are highlighted in red, green, and blue. M+P: Combination of mask and point prompts. M+B: Combination of mask and box prompts. + +
MethodSα↑Fwβ↑Eφ↑M↓mDice↑mIoU↑
SINet[10]0.6360.3460.7750.0410.3810.283
SINet-v2[12]0.6530.3820.7620.0390.4130.318
ZoomNet[32]0.6330.3490.6010.0330.3490.273
FEDERNet[16]0.5730.2720.6090.0510.2600.199
FSPNet[19]0.6810.4010.7160.0440.4460.332
PUENet[44]0.6910.4850.7950.0340.5140.396
RCRNet[42]0.6270.2870.6660.0480.3090.229
PNS-Net[20]0.6550.3250.6730.0480.3840.290
MG[43]0.5940.3360.6910.0590.3680.268
SLT-Net[5]0.6960.4810.8450.0300.4930.401
TSP-SAM(M+P)0.7050.5650.8360.0270.5310.422
TSP-SAM(M+B)0.7510.6280.8510.0210.6030.496
+ +meanDice (mDice), meanIoU (mIoU). Overall, a better VCOD method has larger $S_{\alpha}$ , $F_{\beta}^{w}$ , $E_{\phi}$ , mDice and mIoU scores, but a smaller MAE score. + +Implementation details. In our experiments, we freeze SAM and train our TSP-SAM in an end-to-end manner with a maximum of 20 epochs. Specifically, for a fair comparison, both the training and testing images are resize to $352 \times 352$ . We leverage Adam [24] as our optimizer and set the weight decay to $1 \times 10^{-5}$ . On the MoCA-Mask dataset, the learning rate is initialized to $1 \times 10^{-5}$ for the spatial extractor of motion-driven self-prompt learning and $5 \times 10^{-5}$ for other trainable parameters. After 10 epochs, the learning rate is reduced by $90\%$ . On the CAD2016 dataset, we adopt a ratio of 9:1 to divide all classes into training and testing sets and report the mean value over the testing set. We set the initial learning rate to $1 \times 10^{-4}$ for the spatial extractor of the motion-driven self-prompt learning and $5 \times 10^{-4}$ for other trainable parameters. The other settings are the same as those on MoCA-Mask dataset. + +Table 2. Generalization comparisons on CAD2016 dataset. The best results are highlighted in **bold**. M+P: Combination of mask and point prompts. M+B: Combination of mask and box prompts. + +
MethodTuning parameters (M)MethodTime (s) ↓Speed (fps) ↑
FSPNet274.24FSPNet568.701.31
SLT-Net253.402.94
SLT-Net82.41Overall (ours)294.992.53
TSP-SAM(ours)89.78Prompt learning73.2810.17
SAM221.713.36
+ +Table 3. The number of Table 4. Comparisons on infertuning parameters. Hence time and speed. + +# 5.2. Performance comparison + +1) Baselines: To evaluate the performance of the proposed TSP-SAM, we compare it with a range of state-of-the-art methods. These methods fall into two categories: + +- Single-image camouflaged object detection. These methods analyze the static appearance of images to discern the subtle differences between camouflaged objects and their surroundings. The compared methods in this type include SINet [10], SINet-v2 [12], ZoomNet [32], BGNet [36], FEDERNet [16], FSPNet[19], and PUENet [44]. +- Video object segmentation. This set of methods explores the temporal-spatial relationships between frames for the purpose of segmenting objects with specific attributes. The compared methods in this type encompass RCRNet [42], PNS-Net [20], MG[43], and SLT-Net[5]. +2) Quantitative Results: We evaluate TSP-SAM on MoCA-Mask and CAD2016 datasets. From Table 1, we can conclude that, (i) compared to the combination of mask and point prompts, the combination of mask and box prompts is more reliable for SAM. This is attributed to the limitations of the point prompt in conveying the boundary information. In contrast, bounding box excels in indicating the boundaries of the camouflaged object, even in the presence of bias. (ii) the strong contrast between the performance of TSP-SAM and single-image camouflaged object detection methods points to the importance of temporal-spatial relationships in breaking camouflage. (iii) the proposed TSP-SAM + +![](images/8da5257c5b97459ed1fb85348d80fbdc4e58f925855d13587a69f56d3e9176f9.jpg) +Figure 4. Visualization of our proposed TSP-SAM and baseline methods on MoCA-Mask and CAD2016 dataset. From left to right: frames (1st column), ground truth (2nd column), prediction of our TSP-SAM with mask and box prompts (3rd column), prediction of our TSP-SAM with mask and point prompts (4th column), and predictions of compared methods (5th-7th columns). + +![](images/a17edc19376520f40cf5d47563738915e5fab56f4b40ad5a8ccd7a594911b2da.jpg) + +![](images/a3945ae3ac33cebc2a3871b4ea4bac3421814b9aa88d51fddf8307c22f32e3ca.jpg) +Figure 5. Visualization of several failure cases with mask and point prompts. From left to right: frames (1st column), ground truth (2nd column), mask and box prompts (3rd column), prediction with mask and box prompts (4th column), mask and point prompts (5th column), and prediction with mask and point prompts (6th column). + +surpasses all video object segmentation methods. Notably, TSP-SAM exhibits a $1.3\% - 8.7\%$ improvement over SLT-Net [5] on MoCA-Mask dataset, meaning that the temporal-spatial exploration in TSP-SAM is more effective. + +To assess the generalization performance, we train on the MoCA-Mask dataset and report the test performance on the CAD2016 dataset. It can be observed from Table 2 that the proposed TSP-SAM outperforms all baselines. Compared with SLTNet [5], TSP-SAM surpasses it by $0.9\% - 14.7\%$ , indicating that TSP-SAM is capable of generalizing to the unseen data more effectively. Meanwhile, we also test the performance of TSP-SAM on general video object segmentation datasets. On DAVIS2016 dataset [34], our model achieves comparable performance to the SOTA SLT-Net. While on SegTrack-v2 dataset [27] with multiple objects, our method obtains $62.84\%$ for mIoU, surpassing SLT-Net by $9.9\%$ . Surprisingly, our model is even comparable to the recent SMTC [35] on SegTrack-v2. + +Besides, we report the number of tuning parameters, the inference time and speed. From Table 3, compared with SLT-Net, our model achieves $9.6\%$ improvements with a slight parameter increase. From Table 4, our model is comparable to the latest video-oriented SLT-Net. Decomposing the overall inference time, SAM consumes $75\%$ of it. + +3) Qualitative Results: To provide more intuitive evaluations of the proposed TSP-SAM, we conduct two qualitative discussions. Firstly, as shown in Fig. 4, we present the segmentation results of several examples to intuitively compare the proposed TSP-SAM with the baselines. It can be seen that: (i) some confusing surrounding regions interfere with the baseline methods, while our TSP-SAM is more robust in identifying the camouflaged object. This demonstrates the necessity of short-long term temporal-spatial relationships. (ii) compared with the baselines, our TSP-SAM exhibits superior capabilities in boundary segmentation in terms of discriminability and details. + +Secondly, to further analyze the influence of the type of sparse prompts on overall performance, we visualize the several failure cases with mask and point prompts. From Fig. 5, it can be found that: (i) since the target object is not always a regular shape, the learned point prompt is not on the target object although it is near the center. (ii) under the condition with unclear boundaries, the point prompt is more likely to segment the regions of background. (iii) the point prompt may induce the model to segment the local areas instead of the complete target object. In contrast, the box prompt can avoid the above defects and achieve more precise segmentation results. Meanwhile, the complete localization under the learned mask and box prompts verifies the effectiveness of the temporal-spatial prompt learning. + +# 5.3. Ablation study + +To gain a comprehensive understanding of the efficacy of core components and the impact of hyper parameters, we delve into a detailed analysis. + +
MethodSα↑Fwβ↑Eφ↑mDice↑mIoU↑
SAM+SSP0.6370.3300.7680.3670.311
SAM+MSP0.6570.3530.8230.4070.340
SAM+MSP+LCP0.6650.4090.7570.4290.366
SAM+MSP+LCP+TSI0.6890.4440.8080.4580.388
+ +![](images/a45acdd249c5cf4f00330c01d64c17a9ff13c154f770b0e6a95b19428f496075.jpg) +(a) mDice + +![](images/5f384288d61922770d1c5b2a73a8eb8a14ce557e8339c87f58b367a311632653.jpg) +(b) $F_{\beta}^{w}$ + +![](images/c288c6aa9ac6ce27b30d87c153ccab97631d5fac17596de03f8236e7ba54ad47.jpg) +(c) mIoU + +![](images/57903b9ca011a475c431aaef47d5125ea9f421d5d8f79ca6ea02148fcb79ba25.jpg) +(d) $E_{\phi}$ + +Ablation analysis: Table 5 present the segmentation results by progressively applying each part in our model. Compared to the spatial-based self-prompt learning (SSP) comprising a spatial extractor and two prompt heads, motion-driven self-prompt learning (MSP) improves the mDice metric from $36.7\%$ to $40.7\%$ to a large degree. This affirms the effectiveness of the inter-frame motion in camouflage breaking. Subsequently, the long-range consistency-based robust prompt learning (LCP) brings $2.2\%$ gains for mDice, indicating the effectiveness of long-range consistency in improving the robustness of self-prompts. Meanwhile, the temporal-spatial injection (TSI) further boosts the mDice metric by $2.9\%$ , once again verifying the significance of long-range consistency. + +To further analyze the effectiveness of frequency-based motion perception in MSP, we replace it with spatial-based motion perception, frequency-based static appearance perception, and spatial-based static appearance perception to conduct the ablation studies. As can be seen from Table 6, the frequency-based motion perception yields the best performance. While verifying the significant role of interframe motion in breaking camouflage, it also demonstrates that inter-frame motion can be expressed more exhaustively in the frequency domain than in the spatial domain. + +Hyper parameter analysis: To investigate the effect of the size of the windowed sequence on segmentation performance, we conduct the ablation experiments by varying $T$ from $\{5,6,7,8,9,10\}$ . As illustrated in Fig. 6, the ac + +Table 5. Ablation studies of the core components on MoCA-Mask dataset. SSP: Spatial-based self-prompt learning. MSP: Motion-driven self-prompt learning. LCP: Long-range consistency-based robust prompt learning. TSI: Temporal-spatial injection for representation enhancement. + +
MethodSα↑Fwβ↑Eφ↑mDice↑mIoU↑
Spatial appearance0.6370.3300.7680.3670.311
Frequency appearance0.6470.3440.7930.3960.328
Spatial motion0.6510.3350.7300.3960.324
Frequency motion0.6570.3530.8230.4070.340
+ +Table 6. Ablation studies of frequency-based motion perception on MoCA-Mask dataset. Spatial appearance: Spatial-based appearance perception. Frequency appearance: Frequency-based appearance perception. Spatial motion: Spatial-based motion perception. Frequency motion: Frequency-based motion perception. + +![](images/3def31d2b60d3effa059bbf903b374e42f8ecaf6fd254574fe74b003f1087ec9.jpg) +(a) mDice + +![](images/f25d3e7d8c5bf9e59fd7d995fcf5e440d24ef30dd60c990d708cdcddb499a2bd.jpg) + +![](images/31c48953916a6a652ba587bcaf0c52f0e7665dd5a7bdd58d6ae4db2d8dc7df05.jpg) +Figure 6. Sensitivity analysis of the size $T$ of the windowed sequence on MoCA-Mask dataset. +(c) mIoU +Figure 7. Sensitivity analysis of the balance factor $(\alpha, \beta)$ in the loss function on MoCA-Mask dataset. + +![](images/5374b09c41621c29624e7c919a04db55bca4e7388bbc10bc109c022432701e52.jpg) +(b) $F_{\beta}^{w}$ +(d) $E_{\phi}$ + +curacy initially rises with increasing the size of the windowed sequence, reaches the best at $T = 7$ before declining. Moreover, we assess the sensitivity of $\alpha$ and $\beta$ that are used to balance the sparse and dense prompt loss in Fig. 7. The optimal segmentation performance is observed at $(\alpha, \beta) = (1,0.6)$ . + +# 6. Conclusion + +In this paper, we propose a flexible end-to-end temporal-spatial prompt SAM (TSP-SAM) for video camouflaged object detection. Instead of user-provided prompts, motion-driven self-prompt learning is proposed to capture the overall motion of the camouflaged object, facilitating its spatial identification. Subsequently, a robust prompt learning based on long-range consistency is proposed to promote the robustness of the self-prompts, which likes endowing SAM with keen eyes. Moreover, the long-range consistency is injected into the encoder of SAM for representation enhancement. Notably, TSP-SAM provides a new perspective, that is temporal-spatial associations, for visual prompt learning. + +# 7. Acknowledgement + +This work was supported in part by the National Key Research and Development Program of China under Grant No.2021ZD0140407, Beijing Natural Science Foundation under Grant No.7222313, and National High Level Hospital Clinical Research Funding under Grant No.2022-PUMCHC-041. + +# References + +[1] Pia Bideau and Erik Learned-Miller. It's moving! a probabilistic model for causal motion segmentation in moving camera videos. In ECCV, pages 433-449, 2016. 1, 2, 5 +[2] Pia Bideau, Rakesh R Menon, and Erik Learned-Miller. Moa-net: self-supervised motion segmentation. In ECCV Workshops, pages 0–0, 2018. 1, 2 +[3] Alexey Bochkovskiy, Chien-Yao Wang, and Hong-Yuan Mark Liao. Yolov4: Optimal speed and accuracy of object detection. arXiv preprint arXiv:2004.10934, 2020. 5 +[4] Tianrun Chen, Lanyun Zhu, Chaotao Ding, Runlong Cao, Shangzhan Zhang, Yan Wang, Zejian Li, Lingyun Sun, Papa Mao, and Ying Zang. Sam fails to segment anything?—sam-adapter: Adapting sam in underperformed scenes: Camouflage, shadow, and more. arXiv preprint arXiv:2304.09148, 2023. 2 +[5] Xuelian Cheng, Huan Xiong, Deng-Ping Fan, Yiran Zhong, Mehrtash Harandi, Tom Drummond, and Zongyuan Ge. Implicit motion handling for video camouflaged object detection. In CVPR, pages 13864-13873, 2022. 1, 2, 4, 5, 6, 7 +[6] Runmin Cong, Mengyao Sun, Sanyi Zhang, Xiaofei Zhou, Wei Zhang, and Yao Zhao. Frequency perception network for camouflaged object detection. In ACM MM, pages 1179-1189, 2023. 2 +[7] Xinyi Cui, Qingshan Liu, and Dimitris Metaxas. Temporal spectral residual: fast motion saliency detection. In ACM MM, pages 617-620, 2009. 3 +[8] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, et al. An image is worth 16x16 words: Transformers for image recognition at scale. 2020. 2 +[9] Deng-Ping Fan, Ming-Ming Cheng, Yun Liu, Tao Li, and Ali Borji. Structure-measure: A new way to evaluate foreground maps. In ICCV, pages 4548-4557, 2017. 5 +[10] Deng-Ping Fan, Ge-Peng Ji, Guolei Sun, Ming-Ming Cheng, Jianbing Shen, and Ling Shao. Camouflaged object detection. In CVPR, pages 2777-2787, 2020. 2, 6 +[11] Deng-Ping Fan, Ge-Peng Ji, Tao Zhou, Geng Chen, Huazhu Fu, Jianbing Shen, and Ling Shao. Pranet: Parallel reverse attention network for polyp segmentation. In International conference on medical image computing and computer-assisted intervention, pages 263-273, 2020. 1 +[12] Deng-Ping Fan, Ge-Peng Ji, Ming-Ming Cheng, and Ling Shao. Concealed object detection. IEEE TPAMI, 44(10): 6024-6042, 2021. 2, 6 +[13] Deng-Ping Fan, Ge-Peng Ji, Xuebin Qin, and Ming-Ming Cheng. Cognitive vision inspired object segmentation metric and loss function. Scientia Sinica Informationis, 6(6), 2021. 5 +[14] Melvyn A Goodale and A David Milner. Separate visual pathways for perception and action. Trends in neurosciences, 15(1):20-25, 1992. 3 +[15] Kensho Hara, Hirokatsu Kataoka, and Yutaka Satoh. Can spatiotemporal 3d cnns retrace the history of 2d cnns and imagenet? In CVPR, pages 6546-6555, 2018. 5 + +[16] Chunming He, Kai Li, Yachao Zhang, Longxiang Tang, Yulin Zhang, Zhenhua Guo, and Xiu Li. Camouflaged object detection with feature decomposition and edge reconstruction. In CVPR, pages 22046-22055, 2023. 1, 6 +[17] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, pages 770-778, 2016. 4 +[18] Yuhao Huang, Xin Yang, Lian Liu, Han Zhou, Ao Chang, Xinrui Zhou, Rusi Chen, Junxuan Yu, Jiongquan Chen, Chaoyu Chen, et al. Segment anything model for medical images? arXiv preprint arXiv:2304.14660, 2023. 2 +[19] Zhou Huang, Hang Dai, Tian-Zhu Xiang, Shuo Wang, Huai-Xin Chen, Jie Qin, and Huan Xiong. Feature shrinkage pyramid for camouflaged object detection with transformers. In CVPR, pages 5557-5566, 2023. 1, 6 +[20] Ge-Peng Ji, Yu-Cheng Chou, Deng-Ping Fan, Geng Chen, Huazhu Fu, Debesh Jha, and Ling Shao. Progressively normalized self-attention network for video polyp segmentation. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 142–152, 2021. 6 +[21] Ge-Peng Ji, Deng-Ping Fan, Yu-Cheng Chou, Dengxin Dai, Alexander Liniger, and Luc Van Gool. Deep gradient learning for efficient camouflaged object detection. Machine Intelligence Research, 20(1):92-108, 2023. 2 +[22] Ge-Peng Ji, Deng-Ping Fan, Peng Xu, Ming-Ming Cheng, Bowen Zhou, and Luc Van Gool. "Sam struggles in concealed scenes—empirical study on" segment anything". arXiv preprint arXiv:2304.06022, 2023. 2 +[23] Yang Jia, Jie Yuan, Jinjun Wang, Jun Fang, Qixing Zhang, and Yongming Zhang. A saliency-based method for early smoke detection in video sequences. Fire technology, 52: 1271-1292, 2016. 3 +[24] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. ICLR, 2015. 6 +[25] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. arXiv preprint arXiv:2304.02643, 2023. 2 +[26] Hala Lamdouar, Charig Yang, Weidi Xie, and Andrew Zisserman. Betrayed by motion: Camouflaged object discovery via motion segmentation. In ACCV, 2020. 1, 2 +[27] Fuxin Li, Taeyoung Kim, Ahmad Humayun, David Tsai, and James M Rehg. Video segmentation by tracking many figure-ground segments. In ICCV, pages 2192-2199, 2013. 7 +[28] Thomas Lidbetter. Search and rescue in the face of uncertain threats. European Journal of Operational Research, 285(3): 1153-1160, 2020. 1 +[29] Ting Liu, Yao Zhao, Yunchao Wei, Yufeng Zhao, and Shikui Wei. Concealed object detection for activate millimeter wave image. IEEE Transactions on Industrial Electronics, 66(12): 9909-9917, 2019. 1 +[30] Jun Ma, Yuting He, Feifei Li, Lin Han, Chenyu You, and Bo Wang. Segment anything in medical images. arXiv preprint arXiv:2304.12306, 2023. 2 +[31] Ran Margolin, Lihi Zelnik-Manor, and Ayellet Tal. How to evaluate foreground maps? In CVPR, pages 248-255, 2014. 5 + +[32] Youwei Pang, Xiaoqi Zhao, Tian-Zhu Xiang, Lihe Zhang, and Huchuan Lu. Zoom in and out: A mixed-scale triplet network for camouflaged object detection. In CVPR, pages 2160-2170, 2022. 6 +[33] Federico Perazzi, Philipp Krahenbuhl, Yael Pritch, and Alexander Hornung. Saliency filters: Contrast based filtering for salient region detection. In CVPR, pages 733-740, 2012. 5 +[34] Federico Peruzzi, Jordi Pont-Tuset, Brian McWilliams, Luc Van Gool, Markus Gross, and Alexander Sorkine-Hornung. A benchmark dataset and evaluation methodology for video object segmentation. In CVPR, pages 724-732, 2016. 7 +[35] Rui Qian, Shuangrui Ding, Xian Liu, and Dahua Lin. Semantics meets temporal correspondence: Self-supervised object-centric learning in videos. In ICCV, pages 16675-16687, 2023. 7 +[36] Yujia Sun, Shuo Wang, Chenglizhao Chen, and Tian-Zhu Xiang. Boundary-guided camouflaged object detection. arXiv preprint arXiv:2207.00794, 2022. 6 +[37] Lv Tang, Haoke Xiao, and Bo Li. Can sam segment anything? when sam meets camouflaged object detection. arXiv preprint arXiv:2304.04709, 2023. 2 +[38] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. NeurIPS, 30, 2017. 4 +[39] Wenhai Wang, Enze Xie, Xiang Li, Deng-Ping Fan, Kaitao Song, Ding Liang, Tong Lu, Ping Luo, and Ling Shao. Pyramid vision transformer: A versatile backbone for dense prediction without convolutions. In ICCV, pages 568-578, 2021. 4 +[40] Junde Wu, Rao Fu, Huihui Fang, Yuanpei Liu, Zhaowei Wang, Yanwu Xu, Yueming Jin, and Tal Arbel. Medical sam adapter: Adapting segment anything model for medical image segmentation. arXiv preprint arXiv:2304.12620, 2023. 2 +[41] Yu-Huan Wu, Shang-Hua Gao, Jie Mei, Jun Xu, Deng-Ping Fan, Rong-Guo Zhang, and Ming-Ming Cheng. Jcs: An explainable Covid-19 diagnosis system by joint classification and segmentation. IEEE TIP, 30:3113-3126, 2021. 1 +[42] Pengxiang Yan, Guanbin Li, Yuan Xie, Zhen Li, Chuan Wang, Tianshui Chen, and Liang Lin. Semi-supervised video salient object detection using pseudo-labels. In ICCV, pages 7284-7293, 2019. 6 +[43] Charig Yang, Hala Lamdouar, Erika Lu, Andrew Zisserman, and Weidi Xie. Self-supervised video object segmentation by motion grouping. In ICCV, pages 7177-7188, 2021. 1, 2, 6 +[44] Yi Zhang, Jing Zhang, Wassim Hamidouche, and Olivier Deforges. Predictive uncertainty estimation for camouflaged object detection. IEEE TIP, 2023. 1, 6 +[45] Yijie Zhong, Bo Li, Lv Tang, Senyun Kuang, Shuang Wu, and Shouhong Ding. Detecting camouflaged object in frequency domain. In CVPR, pages 4504-4513, 2022. 1, 2, 3 \ No newline at end of file diff --git a/2024/Endow SAM with Keen Eyes_ Temporal-spatial Prompt Learning for Video Camouflaged Object Detection/images.zip b/2024/Endow SAM with Keen Eyes_ Temporal-spatial Prompt Learning for Video Camouflaged Object Detection/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..eeba874e783b73114f5be2781697ce43ec1119cb --- /dev/null +++ b/2024/Endow SAM with Keen Eyes_ Temporal-spatial Prompt Learning for Video Camouflaged Object Detection/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3a6ab1120d3c87c0688a42b35580dff1450b326cca0c90dc2a4aba9c0ad00678 +size 675294 diff --git a/2024/Endow SAM with Keen Eyes_ Temporal-spatial Prompt Learning for Video Camouflaged Object Detection/layout.json b/2024/Endow SAM with Keen Eyes_ Temporal-spatial Prompt Learning for Video Camouflaged Object Detection/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..23fca1fcb0647c717def9e32ed07f355ecf391df --- /dev/null +++ b/2024/Endow SAM with Keen Eyes_ Temporal-spatial Prompt Learning for Video Camouflaged Object Detection/layout.json @@ -0,0 +1,10011 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 65, + 102, + 529, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 102, + 529, + 140 + ], + "spans": [ + { + "bbox": [ + 65, + 102, + 529, + 140 + ], + "type": "text", + "content": "Endow SAM with Keen Eyes: Temporal-spatial Prompt Learning for Video Camouflaged Object Detection" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 146, + 160, + 445, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 160, + 445, + 175 + ], + "spans": [ + { + "bbox": [ + 146, + 160, + 445, + 175 + ], + "type": "text", + "content": "Wenjun Hui" + }, + { + "bbox": [ + 146, + 160, + 445, + 175 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 146, + 160, + 445, + 175 + ], + "type": "text", + "content": ", Zhenfeng Zhu" + }, + { + "bbox": [ + 146, + 160, + 445, + 175 + ], + "type": "inline_equation", + "content": "^{1,2*}" + }, + { + "bbox": [ + 146, + 160, + 445, + 175 + ], + "type": "text", + "content": ", Shuai Zheng" + }, + { + "bbox": [ + 146, + 160, + 445, + 175 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 146, + 160, + 445, + 175 + ], + "type": "text", + "content": ", Yao Zhao" + }, + { + "bbox": [ + 146, + 160, + 445, + 175 + ], + "type": "inline_equation", + "content": "^{1,2}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 186, + 177, + 409, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 186, + 177, + 409, + 190 + ], + "spans": [ + { + "bbox": [ + 186, + 177, + 409, + 190 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 186, + 177, + 409, + 190 + ], + "type": "text", + "content": "Institute of Information Science, Beijing Jiaotong University" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 143, + 191, + 451, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 191, + 451, + 203 + ], + "spans": [ + { + "bbox": [ + 143, + 191, + 451, + 203 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 143, + 191, + 451, + 203 + ], + "type": "text", + "content": "Beijing Key Laboratory of Advanced Information Science and Network Technology" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 171, + 205, + 419, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 171, + 205, + 419, + 217 + ], + "spans": [ + { + "bbox": [ + 171, + 205, + 419, + 217 + ], + "type": "text", + "content": "{wenjunhui, zhfzhu, zs1997, yzhao}@bjtu.edu.cn" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 143, + 244, + 192, + 258 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 244, + 192, + 258 + ], + "spans": [ + { + "bbox": [ + 143, + 244, + 192, + 258 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 270, + 289, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 270, + 289, + 594 + ], + "spans": [ + { + "bbox": [ + 46, + 270, + 289, + 594 + ], + "type": "text", + "content": "The Segment Anything Model (SAM), a prompt-driven foundational model, has demonstrated remarkable performance in natural image segmentation. However, its application in video camouflaged object detection (VCOD) encounters challenges, chiefly stemming from the overlooked temporal-spatial associations and the unreliability of user-provided prompts for camouflaged objects that are difficult to discern with the naked eye. To tackle the above issues, we endow SAM with keen eyes and propose the Temporal-spatial Prompt SAM (TSP-SAM), a novel approach tailored for VCOD via an ingenious prompted learning scheme. Firstly, motion-driven self-prompt learning is employed to capture the camouflaged object, thereby bypassing the need for user-provided prompts. With the detected subtle motion cues across consecutive video frames, the overall movement of the camouflaged object is captured for more precise spatial localization. Subsequently, to eliminate the prompt bias resulting from inter-frame discontinuities, the long-range consistency within the video sequences is taken into account to promote the robustness of the self-prompts. It is also injected into the encoder of SAM to enhance the representational capabilities. Extensive experimental results on two benchmarks demonstrate that the proposed TSP-SAM achieves a significant improvement over the state-of-the-art methods. With the mIoU metric increasing by " + }, + { + "bbox": [ + 46, + 270, + 289, + 594 + ], + "type": "inline_equation", + "content": "7.8\\%" + }, + { + "bbox": [ + 46, + 270, + 289, + 594 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 270, + 289, + 594 + ], + "type": "inline_equation", + "content": "9.6\\%" + }, + { + "bbox": [ + 46, + 270, + 289, + 594 + ], + "type": "text", + "content": ", TSP-SAM emerges as a groundbreaking step forward in the field of VCOD." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 616, + 127, + 628 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 616, + 127, + 628 + ], + "spans": [ + { + "bbox": [ + 47, + 616, + 127, + 628 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 635, + 287, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 635, + 287, + 696 + ], + "spans": [ + { + "bbox": [ + 46, + 635, + 287, + 696 + ], + "type": "text", + "content": "Camouflaged object detection, a pivotal task in computer vision, aims to identify the target object seamlessly blending with its surroundings in images [19, 44]. This task plays an important role in various fields, including surveillance and security [29], wildlife protection [28], and medi" + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 310, + 243, + 544, + 305 + ], + "blocks": [ + { + "bbox": [ + 310, + 243, + 544, + 305 + ], + "lines": [ + { + "bbox": [ + 310, + 243, + 544, + 305 + ], + "spans": [ + { + "bbox": [ + 310, + 243, + 544, + 305 + ], + "type": "image", + "image_path": "40cbf0bea7fdfbc6bbbde8bd94b8f24a9255d95c1bc72d1a3a054d7cb961ea74.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 309, + 308, + 543, + 403 + ], + "blocks": [ + { + "bbox": [ + 309, + 308, + 543, + 403 + ], + "lines": [ + { + "bbox": [ + 309, + 308, + 543, + 403 + ], + "spans": [ + { + "bbox": [ + 309, + 308, + 543, + 403 + ], + "type": "image", + "image_path": "476016f2f056582f85d7851cf2e0cfbe1301c4e26b916b14a9285870bd52ad42.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 413, + 545, + 491 + ], + "lines": [ + { + "bbox": [ + 304, + 413, + 545, + 491 + ], + "spans": [ + { + "bbox": [ + 304, + 413, + 545, + 491 + ], + "type": "text", + "content": "Figure 1. Illustration of previous VCOD methods (a-b) and the proposed TSP-SAM (c). In contrast to previous methods with two-stage architecture, the proposed TSP-SAM is trained in an end-to-end manner, incorporating short-long term temporal-spatial relationships to collaboratively learn the reliable prompts. Meanwhile, the long-range consistency is also used to enhance the representational capability of SAM." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 498, + 545, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 498, + 545, + 581 + ], + "spans": [ + { + "bbox": [ + 304, + 498, + 545, + 581 + ], + "type": "text", + "content": "cal image processing [11, 41]. In recent years, single-image camouflaged object detection has made significant progress [16, 19, 45]. However, studies [26, 43] have discerned that observers exhibit heightened perception towards the camouflaged object in motion compared to the static scenarios. This observation has sparked a growing interest in video camouflaged object detection (VCOD) [5]." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 582, + 545, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 582, + 545, + 641 + ], + "spans": [ + { + "bbox": [ + 304, + 582, + 545, + 641 + ], + "type": "text", + "content": "Existing VCOD methods [1, 2, 5, 26] commonly employ a two-step framework to model the inter-frame correlations implicitly or explicitly for camouflage breaking, as shown in Fig. 1(a)(b). Despite yielding promising results, they are noticeably limited in certain aspects:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 642, + 545, + 689 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 304, + 642, + 545, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 642, + 545, + 666 + ], + "spans": [ + { + "bbox": [ + 304, + 642, + 545, + 666 + ], + "type": "text", + "content": "- (i) Primarily, their two-stage architecture exposes them to cumulative errors, thus impacting their overall accuracy." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 666, + 545, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 666, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 304, + 666, + 545, + 689 + ], + "type": "text", + "content": "- (ii) Moreover, they exhibit weak generalization ability due to the limited training data." + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 304, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 689, + 545, + 713 + ], + "type": "text", + "content": "Consequently, there is a growing demand for universal models in video camouflaged object detection." + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 494, + 35 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 494, + 35 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 494, + 35 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 703, + 135, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 703, + 135, + 712 + ], + "spans": [ + { + "bbox": [ + 58, + 703, + 135, + 712 + ], + "type": "text", + "content": "*Corresponding author" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19058" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 203 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 203 + ], + "type": "text", + "content": "With the emergence and development of the visual foundation model, the segment anything model (SAM) [25] has showcased unprecedented performance in the field of natural image segmentation. This accomplishment is attributed to the impressive representations learned on 11 million images in SA-1B [25] and 636 million parameters of ViT-H [8]. SAM not only exhibits heightened versatility in terms of model capacity, but also potentially yields more consistent results across different tasks [30]. Inspired by the remarkable generality of SAM, this work endeavors to make SAM suitable for VCOD." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 205, + 289, + 324 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 205, + 289, + 324 + ], + "spans": [ + { + "bbox": [ + 46, + 205, + 289, + 324 + ], + "type": "text", + "content": "However, it is worth noting that the performance of SAM is uncertain when applied to VCOD due to the differences between the camouflaged images and natural images. Specifically, SAM, as a prompt-driven foundational model, requires extra prompts when segmenting specific regions. Yet, the seamless blending of the camouflaged object with its surroundings renders it imperceptible to the naked eye, posing a formidable challenge in providing reliable prompts. Hence, there arises an urgent need for adaptive prompt learning to the camouflaged object." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 326, + 289, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 326, + 289, + 456 + ], + "spans": [ + { + "bbox": [ + 46, + 326, + 289, + 456 + ], + "type": "text", + "content": "It is well-known that the temporal-spatial interplay in video sequences plays a pivotal role in video-related tasks. Short-term temporal-spatial relationships accentuate the variations between frames, while long-term temporal-spatial correlations model the temporal-spatial context to characterize the dependencies between the target object and the background. Motivated by these considerations, in this work, we develop a flexible end-to-end temporal-spatial prompt SAM, called TSP-SAM, for VCOD. To the best of our knowledge, the proposed TSP-SAM is the first SAM-based camouflaged object detection framework." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 59, + 458, + 280, + 469 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 458, + 280, + 469 + ], + "spans": [ + { + "bbox": [ + 59, + 458, + 280, + 469 + ], + "type": "text", + "content": "Our main contributions can be summarized as follows:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 475, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 47, + 475, + 287, + 534 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 475, + 287, + 534 + ], + "spans": [ + { + "bbox": [ + 47, + 475, + 287, + 534 + ], + "type": "text", + "content": "- Instead of user-provided prompts, we propose motion-driven self-prompt learning to capture the camouflaged object. It perceives the overall motion of the camouflaged object by establishing inter-frame associations in the frequency domain, facilitating its spatial identification." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 534, + 287, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 534, + 287, + 594 + ], + "spans": [ + { + "bbox": [ + 47, + 534, + 287, + 594 + ], + "type": "text", + "content": "- To eliminate the prompt bias stemming from inter-frame discontinuities, robust prompt learning based on long-range consistency is proposed. By modeling the long-range temporal-spatial dependencies within video sequences, the robustness of self-prompts is well promoted." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 594, + 287, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 594, + 287, + 653 + ], + "spans": [ + { + "bbox": [ + 47, + 594, + 287, + 653 + ], + "type": "text", + "content": "- In order to enhance the representational capability of SAM, we propose temporal-spatial injection that incorporates the long-range temporal-spatial consistency into the encoder of SAM. This enhancement contributes to more precise detection within the SAM framework." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 653, + 287, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 653, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 47, + 653, + 287, + 712 + ], + "type": "text", + "content": "- Extensive experiments on MoCA-Mask and CAD2016 datasets demonstrate that the proposed TSP-SAM achieves substantial performance improvements over the state-of-the-art method, with mIoU metric increasing by " + }, + { + "bbox": [ + 47, + 653, + 287, + 712 + ], + "type": "inline_equation", + "content": "7.8\\%" + }, + { + "bbox": [ + 47, + 653, + 287, + 712 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 653, + 287, + 712 + ], + "type": "inline_equation", + "content": "9.6\\%" + }, + { + "bbox": [ + 47, + 653, + 287, + 712 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 71, + 390, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 71, + 390, + 83 + ], + "spans": [ + { + "bbox": [ + 306, + 71, + 390, + 83 + ], + "type": "text", + "content": "2. Related work" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 91, + 468, + 104 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 91, + 468, + 104 + ], + "spans": [ + { + "bbox": [ + 306, + 91, + 468, + 104 + ], + "type": "text", + "content": "2.1. Camouflaged object detection" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 109, + 545, + 301 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 109, + 545, + 301 + ], + "spans": [ + { + "bbox": [ + 304, + 109, + 545, + 301 + ], + "type": "text", + "content": "Camouflaged object detection (COD) aims to identify objects blending into their surroundings in images. Existing CNN-based methods address this task with various strategies. Fan et al. [10, 12] introduced a two-stage process, involving a search stage for localization followed by a segmentation stage for refinement. In multi-task learning, Yang et al. [43] and Ji et al. [21] incorporated an auxiliary task to derive shared context representations for enhancing the COD task. Other strategies, such as frequency cues, have also been explored. These works [6, 45] combined the static frequency spectrum with the spatial representations to discern the subtle differences between the camouflaged object and its surroundings. In contrast to the static frequency perception, our TSP-SAM perceives the frequency energy variants to depict the implicit motion between frames, facilitating the spatial identification of the camouflaged object." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 309, + 495, + 322 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 309, + 495, + 322 + ], + "spans": [ + { + "bbox": [ + 306, + 309, + 495, + 322 + ], + "type": "text", + "content": "2.2. Video camouflaged object detection" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 327, + 545, + 507 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 327, + 545, + 507 + ], + "spans": [ + { + "bbox": [ + 304, + 327, + 545, + 507 + ], + "type": "text", + "content": "Compared to the static appearance of images, the motion between video frames is deemed a significant clue for breaking camouflage. Existing video camouflaged object detection methods [1, 2, 5, 26] implicitly or explicitly modeled the temporal-spatial correlations through a two-stage architecture to detect camouflaged objects. Explicit motion-based methods [1, 2, 26] segmented the optical flow field after compensation or registration into object and background. In contrast, implicit feature correspondences between frames are learned by the neural network for an effective alignment [5]. Nonetheless, their two-stage architecture leads to cumulative errors. Different from the two-stage architecture, our TSP-SAM is trained in an end-to-end manner, incorporating the short-long term temporal-spatial relationships to collaboratively learn reliable prompts for SAM." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 515, + 444, + 528 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 515, + 444, + 528 + ], + "spans": [ + { + "bbox": [ + 306, + 515, + 444, + 528 + ], + "type": "text", + "content": "2.3. Segment anything model" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 534, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 534, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 534, + 545, + 713 + ], + "type": "text", + "content": "The segment anything model (SAM) [25], a prompt-driven large-scale foundation model, has exhibited unprecedented performance in natural image segmentation. The core components of SAM comprise an image encoder, a flexible prompt encoder and a lightweight mask decoder. Despite exhibiting excellent zero-shot segmentation capability, SAM performs unstable in application-specific tasks [4]. It fails to segment medical images [18], camouflaged objects [37] and concealed scenes [22]. To apply SAM to medical images, MedSAM [30] and SAM-Adapter [4, 40] incorporated specific domain knowledge into SAM. Yet, up to now, there has been no work on adapting SAM to be tailored for camouflaged object detection. To fill this gap, TSP-SAM explores the temporal-spatial relationships, presenting the first SAM-based camouflaged object detection framework." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "19059" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 102, + 72, + 492, + 247 + ], + "blocks": [ + { + "bbox": [ + 102, + 72, + 492, + 247 + ], + "lines": [ + { + "bbox": [ + 102, + 72, + 492, + 247 + ], + "spans": [ + { + "bbox": [ + 102, + 72, + 492, + 247 + ], + "type": "image", + "image_path": "22814ad2fb56a81c74495463aadf4300e0bee85d21d9f69c6e6b719162192d03.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 257, + 545, + 323 + ], + "lines": [ + { + "bbox": [ + 46, + 257, + 545, + 323 + ], + "spans": [ + { + "bbox": [ + 46, + 257, + 545, + 323 + ], + "type": "text", + "content": "Figure 2. The framework of the proposed TSP-SAM. In motion-driven self-prompt learning (b), the implicit motion between consecutive frames is perceived to facilitate the spatial identification of the camouflaged object, thereby learning the dense self-prompt " + }, + { + "bbox": [ + 46, + 257, + 545, + 323 + ], + "type": "inline_equation", + "content": "P_{d}(t)" + }, + { + "bbox": [ + 46, + 257, + 545, + 323 + ], + "type": "text", + "content": " and sparse self-prompt " + }, + { + "bbox": [ + 46, + 257, + 545, + 323 + ], + "type": "inline_equation", + "content": "P_{s}(t)" + }, + { + "bbox": [ + 46, + 257, + 545, + 323 + ], + "type": "text", + "content": ". Subsequently, in robust prompt learning (c), the long-range consistency " + }, + { + "bbox": [ + 46, + 257, + 545, + 323 + ], + "type": "inline_equation", + "content": "X_{c}(t)" + }, + { + "bbox": [ + 46, + 257, + 545, + 323 + ], + "type": "text", + "content": " is modeled to refine the sparse self-prompt " + }, + { + "bbox": [ + 46, + 257, + 545, + 323 + ], + "type": "inline_equation", + "content": "P_{s}(t)" + }, + { + "bbox": [ + 46, + 257, + 545, + 323 + ], + "type": "text", + "content": ". Moreover, in temporal-spatial injection (d), the long-range consistency " + }, + { + "bbox": [ + 46, + 257, + 545, + 323 + ], + "type": "inline_equation", + "content": "X_{c}(t)" + }, + { + "bbox": [ + 46, + 257, + 545, + 323 + ], + "type": "text", + "content": " is injected into the image embedding " + }, + { + "bbox": [ + 46, + 257, + 545, + 323 + ], + "type": "inline_equation", + "content": "X_{e}(t)" + }, + { + "bbox": [ + 46, + 257, + 545, + 323 + ], + "type": "text", + "content": " of SAM for representation enhancement. During the training phase, SAM (a) is frozen, including the image encoder " + }, + { + "bbox": [ + 46, + 257, + 545, + 323 + ], + "type": "inline_equation", + "content": "E_{i}" + }, + { + "bbox": [ + 46, + 257, + 545, + 323 + ], + "type": "text", + "content": ", prompt encoder " + }, + { + "bbox": [ + 46, + 257, + 545, + 323 + ], + "type": "inline_equation", + "content": "E_{p}" + }, + { + "bbox": [ + 46, + 257, + 545, + 323 + ], + "type": "text", + "content": " and the mask decoder " + }, + { + "bbox": [ + 46, + 257, + 545, + 323 + ], + "type": "inline_equation", + "content": "E_{m}" + }, + { + "bbox": [ + 46, + 257, + 545, + 323 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 328, + 249, + 342 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 328, + 249, + 342 + ], + "spans": [ + { + "bbox": [ + 47, + 328, + 249, + 342 + ], + "type": "text", + "content": "3. Overview of the proposed TSP-SAM" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 349, + 115, + 361 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 349, + 115, + 361 + ], + "spans": [ + { + "bbox": [ + 47, + 349, + 115, + 361 + ], + "type": "text", + "content": "3.1. Notations" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 369, + 287, + 477 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 369, + 287, + 477 + ], + "spans": [ + { + "bbox": [ + 46, + 369, + 287, + 477 + ], + "type": "text", + "content": "In the subsequent sections, we establish the key notations for clarity and consistency in our discourse. Given a sequence of camouflaged images, denote by " + }, + { + "bbox": [ + 46, + 369, + 287, + 477 + ], + "type": "inline_equation", + "content": "I(t) \\in \\mathbb{R}^{H \\times W \\times 3}" + }, + { + "bbox": [ + 46, + 369, + 287, + 477 + ], + "type": "text", + "content": " the " + }, + { + "bbox": [ + 46, + 369, + 287, + 477 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 369, + 287, + 477 + ], + "type": "text", + "content": "-th frame and " + }, + { + "bbox": [ + 46, + 369, + 287, + 477 + ], + "type": "inline_equation", + "content": "I^w(t) = [I(t - T + 1), \\dots, I(t)] \\in \\mathbb{R}^{T \\times H \\times W \\times 3}" + }, + { + "bbox": [ + 46, + 369, + 287, + 477 + ], + "type": "text", + "content": " the corresponding windowed sequence containing " + }, + { + "bbox": [ + 46, + 369, + 287, + 477 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 46, + 369, + 287, + 477 + ], + "type": "text", + "content": " frames. For a tensor " + }, + { + "bbox": [ + 46, + 369, + 287, + 477 + ], + "type": "inline_equation", + "content": "A \\in \\mathbb{R}^{H \\times W \\times C}" + }, + { + "bbox": [ + 46, + 369, + 287, + 477 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 369, + 287, + 477 + ], + "type": "inline_equation", + "content": "\\mathrm{Mat}(A) \\in \\mathbb{R}^{HW \\times C}" + }, + { + "bbox": [ + 46, + 369, + 287, + 477 + ], + "type": "text", + "content": " denotes its matrixization by spatial-wise concatenation. For a matrix " + }, + { + "bbox": [ + 46, + 369, + 287, + 477 + ], + "type": "inline_equation", + "content": "B \\in \\mathbb{R}^{HW \\times C}" + }, + { + "bbox": [ + 46, + 369, + 287, + 477 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 369, + 287, + 477 + ], + "type": "inline_equation", + "content": "\\mathrm{Ten}(B) \\in \\mathbb{R}^{H \\times W \\times C}" + }, + { + "bbox": [ + 46, + 369, + 287, + 477 + ], + "type": "text", + "content": " represents its tensorization by spatial-wise division." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 487, + 126, + 498 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 487, + 126, + 498 + ], + "spans": [ + { + "bbox": [ + 47, + 487, + 126, + 498 + ], + "type": "text", + "content": "3.2. Framework" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 506, + 287, + 541 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 506, + 287, + 541 + ], + "spans": [ + { + "bbox": [ + 46, + 506, + 287, + 541 + ], + "type": "text", + "content": "Fig. 2 illustrates the framework of the proposed TSP-SAM. The core components consist mainly of the following three modules:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 544, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 46, + 544, + 287, + 603 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 544, + 287, + 603 + ], + "spans": [ + { + "bbox": [ + 46, + 544, + 287, + 603 + ], + "type": "text", + "content": "(a) Motion-driven self-prompt learning. The motion-driven self-prompt learning is to use the implicit inter-frame motion " + }, + { + "bbox": [ + 46, + 544, + 287, + 603 + ], + "type": "inline_equation", + "content": "\\hat{X}_{(t - 1)\\rightarrow t}" + }, + { + "bbox": [ + 46, + 544, + 287, + 603 + ], + "type": "text", + "content": " in the frequency domain to facilitate the spatial identification of the camouflaged object, thereby learning the self-prompts for SAM." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 605, + 287, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 287, + 664 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 287, + 664 + ], + "type": "text", + "content": "(b) Robust prompt learning based on long-range consistency. To eliminate the prompt bias stemming from underlying inter-frame discontinuities, the long-range temporal-spatial consistency " + }, + { + "bbox": [ + 46, + 605, + 287, + 664 + ], + "type": "inline_equation", + "content": "X_{c}(t)" + }, + { + "bbox": [ + 46, + 605, + 287, + 664 + ], + "type": "text", + "content": " in video sequences is modeled to promote the robustness of the self-prompts." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 666, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 666, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 666, + 287, + 713 + ], + "type": "text", + "content": "(c) Temporal-spatial injection for representation enhancement. To enhance the representational capabilities of SAM, the long-range temporal-spatial consistency " + }, + { + "bbox": [ + 46, + 666, + 287, + 713 + ], + "type": "inline_equation", + "content": "X_{c}(t)" + }, + { + "bbox": [ + 46, + 666, + 287, + 713 + ], + "type": "text", + "content": " is injected into the image embedding " + }, + { + "bbox": [ + 46, + 666, + 287, + 713 + ], + "type": "inline_equation", + "content": "X_{e}(t)" + }, + { + "bbox": [ + 46, + 666, + 287, + 713 + ], + "type": "text", + "content": " of SAM, con" + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 330, + 447, + 342 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 330, + 447, + 342 + ], + "spans": [ + { + "bbox": [ + 306, + 330, + 447, + 342 + ], + "type": "text", + "content": "attributing to more precise detection." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 354, + 388, + 367 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 354, + 388, + 367 + ], + "spans": [ + { + "bbox": [ + 306, + 354, + 388, + 367 + ], + "type": "text", + "content": "4. Methodology" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 373, + 545, + 398 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 373, + 545, + 398 + ], + "spans": [ + { + "bbox": [ + 305, + 373, + 545, + 398 + ], + "type": "text", + "content": "In this section, we detail the proposed TSP-SAM method for video camouflaged object detection." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 407, + 495, + 420 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 407, + 495, + 420 + ], + "spans": [ + { + "bbox": [ + 305, + 407, + 495, + 420 + ], + "type": "text", + "content": "4.1. Motion-driven self-prompt learning" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 426, + 545, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 426, + 545, + 510 + ], + "spans": [ + { + "bbox": [ + 304, + 426, + 545, + 510 + ], + "type": "text", + "content": "It is well-known that SAM requires user-provided prompts when segmenting specific regions. Nevertheless, the difficulty in distinguishing the camouflaged object with the naked eye makes the requirement of providing visual prompts a significant challenge. Consequently, a compelling need arises to devise a prompt learning network, capable of adaptively identifying the camouflaged object." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 305, + 526, + 492, + 539 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 526, + 492, + 539 + ], + "spans": [ + { + "bbox": [ + 305, + 526, + 492, + 539 + ], + "type": "text", + "content": "4.1.1 Frequency-based motion perception" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 545, + 545, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 545, + 545, + 653 + ], + "spans": [ + { + "bbox": [ + 304, + 545, + 545, + 653 + ], + "type": "text", + "content": "Inspired by the two-stream hypothesis in neuroscience [14], which highlights the importance of motion in visual perception, we consider the motion as a crucial clue for breaking camouflage. Building on this premise, several studies [7, 23] indicate the motion between frames can be expressed exhaustively in the frequency domain through the variants in frequency energy. Hence, frequency-based motion perception is designed to facilitate the spatial identification of the camouflaged object." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "content": "Specifically, as shown in Fig. 3, given consecutive video frames " + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "inline_equation", + "content": "I(t - 1) \\in \\mathbb{R}^{H \\times W \\times 3}" + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "inline_equation", + "content": "I(t) \\in \\mathbb{R}^{H \\times W \\times 3}" + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "content": ", we first map them into the frequency domain. Following [45], the image features are divided into a set of " + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "inline_equation", + "content": "s \\times s" + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "content": " patches and each patch is processed by Discrete Cosine Transform" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19060" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 51, + 73, + 281, + 198 + ], + "blocks": [ + { + "bbox": [ + 51, + 73, + 281, + 198 + ], + "lines": [ + { + "bbox": [ + 51, + 73, + 281, + 198 + ], + "spans": [ + { + "bbox": [ + 51, + 73, + 281, + 198 + ], + "type": "image", + "image_path": "b4f7591e9f45f69f29836617da8650d78e423ebde20e8c56290f26a1ed158b86.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 209, + 288, + 243 + ], + "lines": [ + { + "bbox": [ + 46, + 209, + 288, + 243 + ], + "spans": [ + { + "bbox": [ + 46, + 209, + 288, + 243 + ], + "type": "text", + "content": "Figure 3. The framework of the motion-driven self-prompt learning. The spatial representations " + }, + { + "bbox": [ + 46, + 209, + 288, + 243 + ], + "type": "inline_equation", + "content": "\\{X_{b}(t)\\}_{b = 1}^{N_{b}}" + }, + { + "bbox": [ + 46, + 209, + 288, + 243 + ], + "type": "text", + "content": " are extracted by PVT [39], where " + }, + { + "bbox": [ + 46, + 209, + 288, + 243 + ], + "type": "inline_equation", + "content": "N_{b}" + }, + { + "bbox": [ + 46, + 209, + 288, + 243 + ], + "type": "text", + "content": " is the number of attention blocks." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 251, + 288, + 313 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 251, + 288, + 313 + ], + "spans": [ + { + "bbox": [ + 46, + 251, + 288, + 313 + ], + "type": "text", + "content": "(DCT) into frequency spectrum, where " + }, + { + "bbox": [ + 46, + 251, + 288, + 313 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 46, + 251, + 288, + 313 + ], + "type": "text", + "content": " is the size of patches. To group all components of the same frequency into one channel, we flatten and reshape the spectrum of these patches, forming the frequency features " + }, + { + "bbox": [ + 46, + 251, + 288, + 313 + ], + "type": "inline_equation", + "content": "\\hat{I}(t - 1) \\in \\mathbb{R}^{\\frac{H}{s} \\times \\frac{W}{s} \\times 3s^2}" + }, + { + "bbox": [ + 46, + 251, + 288, + 313 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 251, + 288, + 313 + ], + "type": "inline_equation", + "content": "\\hat{I}(t) \\in \\mathbb{R}^{\\frac{H}{s} \\times \\frac{W}{s} \\times 3s^2}" + }, + { + "bbox": [ + 46, + 251, + 288, + 313 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 315, + 287, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 315, + 287, + 399 + ], + "spans": [ + { + "bbox": [ + 47, + 315, + 287, + 399 + ], + "type": "text", + "content": "In order to perceive the overall motion of the camouflaged object, the inter-frame associations are established on high frequency and low frequency bands respectively. Specifically, the frequency spectrum " + }, + { + "bbox": [ + 47, + 315, + 287, + 399 + ], + "type": "inline_equation", + "content": "\\{\\hat{I} (t - 1),\\hat{I} (t)\\}" + }, + { + "bbox": [ + 47, + 315, + 287, + 399 + ], + "type": "text", + "content": " is divided into two groups in the channel dimension, namely low frequency bands " + }, + { + "bbox": [ + 47, + 315, + 287, + 399 + ], + "type": "inline_equation", + "content": "\\{\\hat{I}^l (t - 1),\\hat{I}^l (t)\\}" + }, + { + "bbox": [ + 47, + 315, + 287, + 399 + ], + "type": "text", + "content": " and high frequency bands " + }, + { + "bbox": [ + 47, + 315, + 287, + 399 + ], + "type": "inline_equation", + "content": "\\{\\hat{I}^h (t - 1),\\hat{I}^h (t)\\}" + }, + { + "bbox": [ + 47, + 315, + 287, + 399 + ], + "type": "text", + "content": ". Hence, we have:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 408, + 287, + 423 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 408, + 287, + 423 + ], + "spans": [ + { + "bbox": [ + 55, + 408, + 287, + 423 + ], + "type": "interline_equation", + "content": "\\hat {X} _ {(t - 1) \\rightarrow t} ^ {l} = \\mathrm {M L P} \\big (\\operatorname {A T T N} (\\operatorname {M a t} (\\hat {I} ^ {l} (t - 1)), \\operatorname {M a t} (\\hat {I} ^ {l} (t))) \\big) \\quad (1)", + "image_path": "faaa7fc0971685f3bf2a9ba6db8b9288d45e5611b5acc0339651a5a369dd922a.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 435, + 287, + 449 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 435, + 287, + 449 + ], + "spans": [ + { + "bbox": [ + 53, + 435, + 287, + 449 + ], + "type": "interline_equation", + "content": "\\hat {X} _ {(t - 1) \\rightarrow t} ^ {h} = \\operatorname {M L P} (\\operatorname {A T T N} (\\operatorname {M a t} (\\hat {I} ^ {h} (t - 1)), \\operatorname {M a t} (\\hat {I} ^ {h} (t)))) \\tag {2}", + "image_path": "1096d4394488f9148c56db5db58dee535c947917a484edfc131cfac056652fbe.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 455, + 287, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 455, + 287, + 568 + ], + "spans": [ + { + "bbox": [ + 46, + 455, + 287, + 568 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 455, + 287, + 568 + ], + "type": "inline_equation", + "content": "\\hat{X}_{(t - 1)\\rightarrow t}^{l}\\in \\mathbb{R}^{\\frac{HW}{s^2}\\times \\frac{3s^2}{2}}" + }, + { + "bbox": [ + 46, + 455, + 287, + 568 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 455, + 287, + 568 + ], + "type": "inline_equation", + "content": "\\mathrm{MLP}(\\cdot)" + }, + { + "bbox": [ + 46, + 455, + 287, + 568 + ], + "type": "text", + "content": " is the linear layer. " + }, + { + "bbox": [ + 46, + 455, + 287, + 568 + ], + "type": "inline_equation", + "content": "\\mathrm{ATTN}(\\cdot ,\\cdot)" + }, + { + "bbox": [ + 46, + 455, + 287, + 568 + ], + "type": "text", + "content": " is multi-head cross attention [38] to extract the inter-frame correspondences, in which the " + }, + { + "bbox": [ + 46, + 455, + 287, + 568 + ], + "type": "inline_equation", + "content": "(t - 1)" + }, + { + "bbox": [ + 46, + 455, + 287, + 568 + ], + "type": "text", + "content": "-th frame is taken as query and the " + }, + { + "bbox": [ + 46, + 455, + 287, + 568 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 455, + 287, + 568 + ], + "type": "text", + "content": "-th frame serves as key and value. Note that Eq.1 tends to highlight the global deformations of the camouflaged object by characterizing the feature correspondences on the low-frequency bands. Meanwhile, Eq.2 implicitly reveals the local edge motion of the camouflaged object on high-frequency bands." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 570, + 287, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 570, + 287, + 594 + ], + "spans": [ + { + "bbox": [ + 46, + 570, + 287, + 594 + ], + "type": "text", + "content": "Subsequently, the complete implicit motion in the frequency domain is obtained by:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 602, + 287, + 627 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 602, + 287, + 627 + ], + "spans": [ + { + "bbox": [ + 48, + 602, + 287, + 627 + ], + "type": "interline_equation", + "content": "\\hat {X} _ {(t - 1) \\rightarrow t} = \\hat {I} (t) + \\operatorname {C B A} (\\operatorname {C a t} \\left(\\operatorname {T e n} \\left(\\hat {X} _ {(t - 1) \\rightarrow t} ^ {l}\\right), \\operatorname {T e n} \\left(\\hat {X} _ {(t - 1) \\rightarrow t} ^ {h}\\right)\\right)) \\tag {3}", + "image_path": "87a66642e0145e15616518d32006a4c13c7d065f1e5e40227bed6c43c0932b95.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 627, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 627, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 627, + 287, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 627, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\mathrm{CBA}(\\cdot)" + }, + { + "bbox": [ + 46, + 627, + 287, + 714 + ], + "type": "text", + "content": " is the convolution operation followed by batch normalization and activation and " + }, + { + "bbox": [ + 46, + 627, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\operatorname{Cat}(\\cdot)" + }, + { + "bbox": [ + 46, + 627, + 287, + 714 + ], + "type": "text", + "content": " is the channel-wise concatenation. Importantly, " + }, + { + "bbox": [ + 46, + 627, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\hat{X}_{(t - 1)\\rightarrow t}\\in" + }, + { + "bbox": [ + 46, + 627, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^{\\frac{H}{s}\\times \\frac{W}{s}\\times 3s^{2}}" + }, + { + "bbox": [ + 46, + 627, + 287, + 714 + ], + "type": "text", + "content": " implicitly depicts the overall motion of the camouflaged object by perceiving the variants in the frequency energy, enabling to facilitate the discovery of the camouflaged object." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 72, + 545, + 95 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 545, + 95 + ], + "type": "text", + "content": "4.1.2 Implicit motion-induced spatial-frequency fusion" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 99, + 545, + 189 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 99, + 545, + 189 + ], + "spans": [ + { + "bbox": [ + 305, + 99, + 545, + 189 + ], + "type": "text", + "content": "To translate the perceived motion into the spatial domain, the implicit motion-induced spatial-frequency representation fusion is established. This process involves the integration between the implicit motion " + }, + { + "bbox": [ + 305, + 99, + 545, + 189 + ], + "type": "inline_equation", + "content": "\\hat{X}_{(t - 1)\\rightarrow t}" + }, + { + "bbox": [ + 305, + 99, + 545, + 189 + ], + "type": "text", + "content": " in the frequency domain and the spatial representations " + }, + { + "bbox": [ + 305, + 99, + 545, + 189 + ], + "type": "inline_equation", + "content": "X_{b}(t)\\in" + }, + { + "bbox": [ + 305, + 99, + 545, + 189 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^{\\frac{H}{2^{b + 1}}\\times \\frac{W}{2^{b + 1}}\\times C},b\\in \\{1,\\dots ,N_b\\}" + }, + { + "bbox": [ + 305, + 99, + 545, + 189 + ], + "type": "text", + "content": " extracted by PVT [39], where " + }, + { + "bbox": [ + 305, + 99, + 545, + 189 + ], + "type": "inline_equation", + "content": "N_{b}" + }, + { + "bbox": [ + 305, + 99, + 545, + 189 + ], + "type": "text", + "content": " is the number of attention blocks. It can be expressed as follows:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 310, + 196, + 545, + 209 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 310, + 196, + 545, + 209 + ], + "spans": [ + { + "bbox": [ + 310, + 196, + 545, + 209 + ], + "type": "interline_equation", + "content": "\\hat {X} _ {b} (t) = X _ {b} (t) + \\operatorname {T e n} (\\operatorname {A T T N} (\\operatorname {M a t} (\\operatorname {C B A} (\\hat {X} _ {(t - 1) \\rightarrow t})), \\operatorname {M a t} (X _ {b} (t)))) \\tag {4}", + "image_path": "b56fde7b5312e0a8b8dc152f2411ab420ba58eb280da351172ceeee185e668a8.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 215, + 545, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 215, + 545, + 266 + ], + "spans": [ + { + "bbox": [ + 305, + 215, + 545, + 266 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 215, + 545, + 266 + ], + "type": "inline_equation", + "content": "\\hat{X}_b(t) \\in \\mathbb{R}^{\\frac{H}{2^b + 1} \\times \\frac{W}{2^b + 1} \\times C}" + }, + { + "bbox": [ + 305, + 215, + 545, + 266 + ], + "type": "text", + "content": ". Note that the spatial perception of the camouflaged object could be strengthened by establishing cross attention between the implicit motion as query and the spatial representation as key and value." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 281, + 429, + 293 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 281, + 429, + 293 + ], + "spans": [ + { + "bbox": [ + 305, + 281, + 429, + 293 + ], + "type": "text", + "content": "4.1.3 Self-prompt learning" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 298, + 545, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 298, + 545, + 357 + ], + "spans": [ + { + "bbox": [ + 304, + 298, + 545, + 357 + ], + "type": "text", + "content": "Based on the co-representations " + }, + { + "bbox": [ + 304, + 298, + 545, + 357 + ], + "type": "inline_equation", + "content": "\\{\\hat{X}_b(t)\\}_{b=1}^{N_b}" + }, + { + "bbox": [ + 304, + 298, + 545, + 357 + ], + "type": "text", + "content": ", two prompt heads are adopted to generate the dense self-prompt " + }, + { + "bbox": [ + 304, + 298, + 545, + 357 + ], + "type": "inline_equation", + "content": "P_d(t)" + }, + { + "bbox": [ + 304, + 298, + 545, + 357 + ], + "type": "text", + "content": " and sparse self-prompt " + }, + { + "bbox": [ + 304, + 298, + 545, + 357 + ], + "type": "inline_equation", + "content": "P_s(t)" + }, + { + "bbox": [ + 304, + 298, + 545, + 357 + ], + "type": "text", + "content": ", respectively. Specifically, we adopt the decoder of SLTNet [5] as dense prompt head " + }, + { + "bbox": [ + 304, + 298, + 545, + 357 + ], + "type": "inline_equation", + "content": "E_d" + }, + { + "bbox": [ + 304, + 298, + 545, + 357 + ], + "type": "text", + "content": " as follows:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 372, + 357, + 545, + 373 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 372, + 357, + 545, + 373 + ], + "spans": [ + { + "bbox": [ + 372, + 357, + 545, + 373 + ], + "type": "interline_equation", + "content": "P _ {d} (t) = E _ {d} \\left(\\left\\{\\hat {X} _ {b} (t) \\right\\} _ {b = 1} ^ {N _ {b}}\\right) \\tag {5}", + "image_path": "e8c94fc2f4f2ebed527be7da1483448ae2d4a2198159daf5756042ef13400f9b.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 376, + 545, + 423 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 376, + 545, + 423 + ], + "spans": [ + { + "bbox": [ + 304, + 376, + 545, + 423 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 376, + 545, + 423 + ], + "type": "inline_equation", + "content": "P_{d}(t)" + }, + { + "bbox": [ + 304, + 376, + 545, + 423 + ], + "type": "text", + "content": " is the dense self-prompt. Meanwhile, as illustrated in Fig. 3, to obtain the sparse prompt, we adopt the residual block of Resnet [17] to process the co-representations in a bottom-up manner:" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 312, + 430, + 545, + 468 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 430, + 545, + 468 + ], + "spans": [ + { + "bbox": [ + 312, + 430, + 545, + 468 + ], + "type": "interline_equation", + "content": "\\tilde {X} _ {b} (t) = \\left\\{ \\begin{array}{l l} \\operatorname {R e s b l o c k} \\left(\\hat {X} _ {b} (t)\\right) & b = N _ {b} \\\\ \\operatorname {R e s b l o c k} \\left(\\hat {X} _ {b} (t) + \\tilde {X} _ {b + 1} (t)\\right) & b \\in \\{1, \\dots , N _ {b} - 1 \\} \\end{array} \\right. \\tag {6}", + "image_path": "209043bb8d93f7840b076fd6f53443304b615bf0260f9ff65a4e7fa6df08368c.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 306, + 468, + 533, + 480 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 468, + 533, + 480 + ], + "spans": [ + { + "bbox": [ + 306, + 468, + 533, + 480 + ], + "type": "text", + "content": "Finally, the sparse self-prompt is represented as follows:" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 355, + 487, + 545, + 502 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 355, + 487, + 545, + 502 + ], + "spans": [ + { + "bbox": [ + 355, + 487, + 545, + 502 + ], + "type": "interline_equation", + "content": "P _ {s} (t) = \\operatorname {A v g p o o l} \\left(\\operatorname {C B A} \\left(\\tilde {X} _ {1} (t)\\right)\\right) \\tag {7}", + "image_path": "6d55ec933316d1639acbb782ed58df19128d53637ce05577c434b691080b4fbf.jpg" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 306, + 509, + 527, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 509, + 527, + 521 + ], + "spans": [ + { + "bbox": [ + 306, + 509, + 527, + 521 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 306, + 509, + 527, + 521 + ], + "type": "inline_equation", + "content": "\\mathrm{Avgpool}(\\cdot)" + }, + { + "bbox": [ + 306, + 509, + 527, + 521 + ], + "type": "text", + "content": " is global average pooling operation." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 306, + 528, + 545, + 552 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 528, + 545, + 552 + ], + "spans": [ + { + "bbox": [ + 306, + 528, + 545, + 552 + ], + "type": "text", + "content": "4.2. Robust prompt learning based on long-range consistency" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 304, + 557, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 557, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 557, + 545, + 713 + ], + "type": "text", + "content": "The self-prompts of the camouflaged object are well learned by using the perceived inter-frame motion to facilitate its spatial identification. However, the susceptibility of inter-frame motion to underlying temporal-spatial discontinuities, such as occlusions, shaking, scene changes, etc., brings a potential lack of reliability in the learned self-prompts. It is well known that the long-range dependencies within the video sequences enable to characterize the temporal-spatial context to cope with the short-term temporal-spatial discontinuities. Hence, it is necessary to explore the long-range temporal-spatial consistency within the video sequences to improve the robustness of the self-prompts, which likes endowing SAM with keen eyes." + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19061" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 226, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 226, + 85 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 226, + 85 + ], + "type": "text", + "content": "4.2.1 Long-range consistency squeezing" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 91, + 287, + 139 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 91, + 287, + 139 + ], + "spans": [ + { + "bbox": [ + 47, + 91, + 287, + 139 + ], + "type": "text", + "content": "As illustrated in Fig. 2(c), given the windowed sequence " + }, + { + "bbox": [ + 47, + 91, + 287, + 139 + ], + "type": "inline_equation", + "content": "I^{w}(t) \\in \\mathbb{R}^{T \\times H \\times W \\times 3}" + }, + { + "bbox": [ + 47, + 91, + 287, + 139 + ], + "type": "text", + "content": " corresponding to the current frame " + }, + { + "bbox": [ + 47, + 91, + 287, + 139 + ], + "type": "inline_equation", + "content": "I(t)" + }, + { + "bbox": [ + 47, + 91, + 287, + 139 + ], + "type": "text", + "content": ", we adopt a 3D convolution neural network " + }, + { + "bbox": [ + 47, + 91, + 287, + 139 + ], + "type": "inline_equation", + "content": "E_{t}" + }, + { + "bbox": [ + 47, + 91, + 287, + 139 + ], + "type": "text", + "content": " to extract the long-range temporal-spatial relationships:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 124, + 147, + 287, + 160 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 124, + 147, + 287, + 160 + ], + "spans": [ + { + "bbox": [ + 124, + 147, + 287, + 160 + ], + "type": "interline_equation", + "content": "X ^ {w} (t) = E _ {t} \\left(I ^ {w} (t)\\right) \\tag {8}", + "image_path": "ebacf59e101cea03003947f54c3fb4759e1f438427d9664548e9c67ae642b2b0.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 167, + 287, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 167, + 287, + 205 + ], + "spans": [ + { + "bbox": [ + 47, + 167, + 287, + 205 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 167, + 287, + 205 + ], + "type": "inline_equation", + "content": "X^w(t) \\in \\mathbb{R}^{T \\times \\frac{H}{4} \\times \\frac{W}{4} \\times C}" + }, + { + "bbox": [ + 47, + 167, + 287, + 205 + ], + "type": "text", + "content": " fully reveals the temporal-spatial dependencies within the video sequences. In our work, we adopt 3D Resnet18 [15] as " + }, + { + "bbox": [ + 47, + 167, + 287, + 205 + ], + "type": "inline_equation", + "content": "E_t" + }, + { + "bbox": [ + 47, + 167, + 287, + 205 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 206, + 287, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 206, + 287, + 277 + ], + "spans": [ + { + "bbox": [ + 47, + 206, + 287, + 277 + ], + "type": "text", + "content": "Subsequently, to further extract the long-range temporal-spatial consistency of the camouflaged object, we take the camouflage representation " + }, + { + "bbox": [ + 47, + 206, + 287, + 277 + ], + "type": "inline_equation", + "content": "X_{e}(t)" + }, + { + "bbox": [ + 47, + 206, + 287, + 277 + ], + "type": "text", + "content": " from the encoder of SAM as query and the long-range temporal-spatial dependencies " + }, + { + "bbox": [ + 47, + 206, + 287, + 277 + ], + "type": "inline_equation", + "content": "X^{w}(t)" + }, + { + "bbox": [ + 47, + 206, + 287, + 277 + ], + "type": "text", + "content": " averaged along the temporal dimension as key and value to establish the cross attention:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 285, + 287, + 300 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 285, + 287, + 300 + ], + "spans": [ + { + "bbox": [ + 55, + 285, + 287, + 300 + ], + "type": "interline_equation", + "content": "X _ {c} (t) = \\operatorname {A T T N} \\left(\\operatorname {M a t} \\left(X _ {e} (t)\\right), \\operatorname {M a t} \\left(\\operatorname {A v g} _ {\\mathrm {t}} \\left(X ^ {w} (t)\\right)\\right)\\right) \\tag {9}", + "image_path": "ad01871199bc85317a7f300934ff170166e6ceb527cb228371ba8e86a8511cf7.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 307, + 287, + 344 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 307, + 287, + 344 + ], + "spans": [ + { + "bbox": [ + 47, + 307, + 287, + 344 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 307, + 287, + 344 + ], + "type": "inline_equation", + "content": "\\mathrm{Avg}_{\\mathrm{t}}(\\cdot)" + }, + { + "bbox": [ + 47, + 307, + 287, + 344 + ], + "type": "text", + "content": " is an average operation to squeeze the temporal dimension. " + }, + { + "bbox": [ + 47, + 307, + 287, + 344 + ], + "type": "inline_equation", + "content": "X_{c}(t)\\in \\mathbb{R}^{\\frac{HW}{16}\\times C}" + }, + { + "bbox": [ + 47, + 307, + 287, + 344 + ], + "type": "text", + "content": " characterizes the long-range consistency specific to the camouflaged object." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 358, + 184, + 371 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 358, + 184, + 371 + ], + "spans": [ + { + "bbox": [ + 47, + 358, + 184, + 371 + ], + "type": "text", + "content": "4.2.2 Robust prompt learning" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 377, + 287, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 377, + 287, + 460 + ], + "spans": [ + { + "bbox": [ + 47, + 377, + 287, + 460 + ], + "type": "text", + "content": "To enhance the robustness of the self-prompts to inter-frame temporal-spatial discontinuities, the long-range temporal-spatial consistency " + }, + { + "bbox": [ + 47, + 377, + 287, + 460 + ], + "type": "inline_equation", + "content": "X_{c}(t)" + }, + { + "bbox": [ + 47, + 377, + 287, + 460 + ], + "type": "text", + "content": " should be induced into the prompt learning. Specifically, in order to map the long-range consistency into the coordinate space, a learnable query " + }, + { + "bbox": [ + 47, + 377, + 287, + 460 + ], + "type": "inline_equation", + "content": "Q_{c}(t) \\in \\mathbb{R}^{4 \\times C}" + }, + { + "bbox": [ + 47, + 377, + 287, + 460 + ], + "type": "text", + "content": " is defined to establish cross attention with " + }, + { + "bbox": [ + 47, + 377, + 287, + 460 + ], + "type": "inline_equation", + "content": "X_{c}(t)" + }, + { + "bbox": [ + 47, + 377, + 287, + 460 + ], + "type": "text", + "content": ". It is formulated as:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 89, + 469, + 287, + 483 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 469, + 287, + 483 + ], + "spans": [ + { + "bbox": [ + 89, + 469, + 287, + 483 + ], + "type": "interline_equation", + "content": "P _ {r} (t) = \\operatorname {A v g} _ {\\mathrm {c}} (\\operatorname {A T N} \\left(Q _ {c} (t), X _ {c} (t)\\right)) \\tag {10}", + "image_path": "61336e388484bf737a2afed47a027e59052350f2d144da9e3f9efe7a744f384c.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 491, + 287, + 527 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 491, + 287, + 527 + ], + "spans": [ + { + "bbox": [ + 47, + 491, + 287, + 527 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 491, + 287, + 527 + ], + "type": "inline_equation", + "content": "\\mathrm{Avg}_c(\\cdot)" + }, + { + "bbox": [ + 47, + 491, + 287, + 527 + ], + "type": "text", + "content": " is an average operation along the channel dimension. " + }, + { + "bbox": [ + 47, + 491, + 287, + 527 + ], + "type": "inline_equation", + "content": "P_r(t)\\in \\mathbf{R}^4" + }, + { + "bbox": [ + 47, + 491, + 287, + 527 + ], + "type": "text", + "content": " is a refinement term derived from long-range consistency to correct the sparse self-prompt:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 119, + 535, + 287, + 548 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 119, + 535, + 287, + 548 + ], + "spans": [ + { + "bbox": [ + 119, + 535, + 287, + 548 + ], + "type": "interline_equation", + "content": "P _ {s} ^ {r} (t) = P _ {s} (t) + P _ {r} (t) \\tag {11}", + "image_path": "34fbb592a6ab3a2d8829f6ce75cec8153d10ad7c417a04800f16b995b4baec77.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 47, + 556, + 287, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 556, + 287, + 605 + ], + "spans": [ + { + "bbox": [ + 47, + 556, + 287, + 605 + ], + "type": "text", + "content": "Through incorporating long-range consistency into the prompt learning, the visual prompts could be more robust to the temporal-spatial discontinuities, which means endowing SAM with a pair of keen eyes." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 47, + 611, + 287, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 611, + 287, + 635 + ], + "spans": [ + { + "bbox": [ + 47, + 611, + 287, + 635 + ], + "type": "text", + "content": "4.3. Temporal-spatial injection for representation enhancement" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 47, + 641, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 641, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 641, + 287, + 713 + ], + "type": "text", + "content": "It is known that the remarkable performance of SAM relies not only on well learned robust prompts but also on the representational capability of the image encoder. However, due to the fact that the image embedding of SAM lacks the temporal-spatial information, the final predictions of consecutive frames may be inconsistent. Hence, it is necessary" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "type": "text", + "content": "to inject the long-range temporal-spatial consistency into SAM. As shown in Fig. 2(d), it is represented as follows:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 364, + 104, + 545, + 118 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 364, + 104, + 545, + 118 + ], + "spans": [ + { + "bbox": [ + 364, + 104, + 545, + 118 + ], + "type": "interline_equation", + "content": "\\hat {X} _ {e} (t) = X _ {e} (t) + \\operatorname {T e n} \\left(X _ {c} (t)\\right) \\tag {12}", + "image_path": "1dfbc21b95077429576570a01bc2e3b43940fc63d65d1a031a71546a30bc4654.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 305, + 127, + 545, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 127, + 545, + 175 + ], + "spans": [ + { + "bbox": [ + 305, + 127, + 545, + 175 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 127, + 545, + 175 + ], + "type": "inline_equation", + "content": "X_{c}(t)" + }, + { + "bbox": [ + 305, + 127, + 545, + 175 + ], + "type": "text", + "content": " is the long-range consistency from Eq. 9. In this way, based on the enhanced image representations and the robust prompts called keen eyes, SAM achieves more precise camouflaged object detection in videos." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 305, + 182, + 421, + 194 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 182, + 421, + 194 + ], + "spans": [ + { + "bbox": [ + 305, + 182, + 421, + 194 + ], + "type": "text", + "content": "4.4. Model optimization" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 305, + 201, + 545, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 201, + 545, + 224 + ], + "spans": [ + { + "bbox": [ + 305, + 201, + 545, + 224 + ], + "type": "text", + "content": "We train the proposed TSP-SAM in an end-to-end manner by minimizing the joint loss below:" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 347, + 232, + 545, + 247 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 347, + 232, + 545, + 247 + ], + "spans": [ + { + "bbox": [ + 347, + 232, + 545, + 247 + ], + "type": "interline_equation", + "content": "\\mathcal {L} (t) = \\mathcal {L} _ {S A M} (t) + \\alpha \\mathcal {L} _ {p} ^ {s} (t) + \\beta \\mathcal {L} _ {p} ^ {d} (t) \\tag {13}", + "image_path": "66c90d492f01cdde4465539e3de82c1f007ee4ae8bc9a9c2636c2605ffa7359b.jpg" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 305, + 256, + 545, + 280 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 256, + 545, + 280 + ], + "spans": [ + { + "bbox": [ + 305, + 256, + 545, + 280 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 256, + 545, + 280 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{SAM},\\mathcal{L}_p^s" + }, + { + "bbox": [ + 305, + 256, + 545, + 280 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 256, + 545, + 280 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_p^d" + }, + { + "bbox": [ + 305, + 256, + 545, + 280 + ], + "type": "text", + "content": " supervise the final prediction, the sparse prompt and the dense prompt, respectively." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 306, + 281, + 545, + 306 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 281, + 545, + 306 + ], + "spans": [ + { + "bbox": [ + 306, + 281, + 545, + 306 + ], + "type": "text", + "content": "Following [3], the supervision on the sparse prompt " + }, + { + "bbox": [ + 306, + 281, + 545, + 306 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_p^s (t)" + }, + { + "bbox": [ + 306, + 281, + 545, + 306 + ], + "type": "text", + "content": " is as follows:" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 315, + 312, + 545, + 327 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 312, + 545, + 327 + ], + "spans": [ + { + "bbox": [ + 315, + 312, + 545, + 327 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {p} ^ {s} (t) = \\operatorname {B C E} \\left(P _ {s} ^ {r} (t), G (t)\\right) + \\operatorname {M S E} \\left(P _ {s} ^ {r} (t), G (t)\\right) \\tag {14}", + "image_path": "cb3c0abff0b87d3c22f0463b24bea018f9347cb67b099139f73fc674b4c44878.jpg" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 305, + 335, + 545, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 335, + 545, + 396 + ], + "spans": [ + { + "bbox": [ + 305, + 335, + 545, + 396 + ], + "type": "text", + "content": "where BCE is binary cross-entropy loss to supervise the coordinates and MSE is mean square error to supervise the size of the box prompt or the scale of the point prompt. For the dense prompt loss " + }, + { + "bbox": [ + 305, + 335, + 545, + 396 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_p^d" + }, + { + "bbox": [ + 305, + 335, + 545, + 396 + ], + "type": "text", + "content": " and the final prediction loss " + }, + { + "bbox": [ + 305, + 335, + 545, + 396 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{SAM}" + }, + { + "bbox": [ + 305, + 335, + 545, + 396 + ], + "type": "text", + "content": ", we adopt the hybrid loss [13] for supervision:" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 378, + 403, + 545, + 418 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 378, + 403, + 545, + 418 + ], + "spans": [ + { + "bbox": [ + 378, + 403, + 545, + 418 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {h y b r i d}} = \\mathcal {L} _ {c e} ^ {w} + \\mathcal {L} _ {i o u} ^ {w} \\tag {15}", + "image_path": "0988befd6ba1648a74af4bd779d64a812f4fbd16df5025aaaf39ee6d4defd38d.jpg" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 305, + 426, + 545, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 426, + 545, + 449 + ], + "spans": [ + { + "bbox": [ + 305, + 426, + 545, + 449 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 426, + 545, + 449 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{ce}^{w}" + }, + { + "bbox": [ + 305, + 426, + 545, + 449 + ], + "type": "text", + "content": " is the weighted cross-entropy loss and " + }, + { + "bbox": [ + 305, + 426, + 545, + 449 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{iou}^{w}" + }, + { + "bbox": [ + 305, + 426, + 545, + 449 + ], + "type": "text", + "content": " is the weighted intersection-over-union loss." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 306, + 460, + 484, + 474 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 460, + 484, + 474 + ], + "spans": [ + { + "bbox": [ + 306, + 460, + 484, + 474 + ], + "type": "text", + "content": "5. Experiment results and analysis" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 306, + 479, + 422, + 493 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 479, + 422, + 493 + ], + "spans": [ + { + "bbox": [ + 306, + 479, + 422, + 493 + ], + "type": "text", + "content": "5.1. Experiment settings" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 305, + 498, + 545, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 498, + 545, + 665 + ], + "spans": [ + { + "bbox": [ + 305, + 498, + 545, + 665 + ], + "type": "text", + "content": "Datasets. Following previous methods [5], we have conducted a comprehensive evaluation on two widely used datasets: MoCA-Mask [5] and CAD2016 [1]. The MoCA-Mask dataset [5] is a video camouflaged object detection dataset, covering camouflaged animals moving in natural scenes. It contains 87 video sequences (22939 frames), of which 71 video sequences are used for training and 16 sequences for testing. It is well annotated with bounding boxes and pixel-level segmentation masks on every fifth frames. Meanwhile, Camouflaged Animal Dataset [1] (CAD2016) is a small video camouflaged object detection dataset extracted from YouTube videos. It consists of nine short video clips in total and is well annotated with pixel-level semantic masks on every fifth frames." + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 305, + 665, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 665, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 665, + 545, + 713 + ], + "type": "text", + "content": "Metrics. During the testing phase, we assess the prediction masks using six widely-used metrics: S-measure " + }, + { + "bbox": [ + 305, + 665, + 545, + 713 + ], + "type": "inline_equation", + "content": "(S_{\\alpha})" + }, + { + "bbox": [ + 305, + 665, + 545, + 713 + ], + "type": "text", + "content": " [9], Weighted F-measure " + }, + { + "bbox": [ + 305, + 665, + 545, + 713 + ], + "type": "inline_equation", + "content": "(F_{\\beta}^{w})" + }, + { + "bbox": [ + 305, + 665, + 545, + 713 + ], + "type": "text", + "content": " [31], Enhanced-alignment measure " + }, + { + "bbox": [ + 305, + 665, + 545, + 713 + ], + "type": "inline_equation", + "content": "(E_{\\phi})" + }, + { + "bbox": [ + 305, + 665, + 545, + 713 + ], + "type": "text", + "content": " [13], Mean absolute error (MAE, " + }, + { + "bbox": [ + 305, + 665, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 305, + 665, + 545, + 713 + ], + "type": "text", + "content": ") [33]," + } + ] + } + ], + "index": 30 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19062" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 49, + 71, + 547, + 234 + ], + "blocks": [ + { + "bbox": [ + 49, + 71, + 547, + 234 + ], + "lines": [ + { + "bbox": [ + 49, + 71, + 547, + 234 + ], + "spans": [ + { + "bbox": [ + 49, + 71, + 547, + 234 + ], + "type": "table", + "html": "
MethodYearInputMoCA-MaskCAD2016
\\(S_{\\alpha}\\uparrow\\)\\(F^{\\text{w}}_{\\beta}\\uparrow\\)\\(E_{\\phi}\\uparrow\\)\\(\\mathcal{M}\\downarrow\\)mDice↑mIoU↑\\(S_{\\alpha}\\uparrow\\)\\(F^{\\text{w}}_{\\beta}\\uparrow\\)\\(E_{\\phi}\\uparrow\\)\\(\\mathcal{M}\\downarrow\\)mDice↑mIoU↑
SINet[10]2020-CVPRImage0.5740.1850.6550.0300.2210.1560.6010.2040.5890.0890.2890.209
SINet-v2[12]2021-TPAMIImage0.5710.1750.6080.0350.2110.1530.5440.1810.5460.0490.1700.110
ZoomNet[32]2022-CVPRImage0.5820.2110.5360.0330.2240.1670.5870.2250.5940.0630.2460.166
BGNet [36]2022-IJCAIImage0.5900.2030.6470.0230.2250.1670.6070.2030.6660.0890.3450.256
FEDERNet[16]2023-CVPRImage0.5550.1580.5420.0490.1920.1320.6070.2460.7250.0610.3610.257
FSPNet[19]2023-CVPRImage0.5940.1820.6080.0440.2380.1670.5390.2200.5530.1450.3090.212
PUENet[44]2023-TIPImage0.5940.2040.6190.0370.3020.2120.6730.4270.8030.0340.4990.389
RCRNet[42]2019-ICCVVideo0.5970.1740.5830.0250.1940.137------
PNS-Net[20]2021-MICCAIVideo0.5760.1340.5620.0380.1890.1330.6780.3690.7200.0430.4090.308
MG[43]2021-ICCVVideo0.5470.1650.5370.0950.1970.1410.4840.3140.5580.3700.3510.260
SLT-Net[5]2022-CVPRVideo0.6560.3570.7850.0210.3870.3100.6790.4200.8050.0330.4450.342
TSP-SAM(M+P)OursVideo0.6730.4000.7660.0120.4210.3450.6810.5000.8530.0310.4960.393
TSP-SAM(M+B)OursVideo0.6890.4440.8080.0080.4580.3880.7040.5240.9120.0280.5430.438
", + "image_path": "6f61b7800eef62e45a9a04519775931ef44fec0640325812467beec291ef2ecb.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 49, + 270, + 286, + 418 + ], + "blocks": [ + { + "bbox": [ + 46, + 242, + 546, + 266 + ], + "lines": [ + { + "bbox": [ + 46, + 242, + 546, + 266 + ], + "spans": [ + { + "bbox": [ + 46, + 242, + 546, + 266 + ], + "type": "text", + "content": "Table 1. Quantitative comparisons over two benchmark datasets. The top three results are highlighted in red, green, and blue. M+P: Combination of mask and point prompts. M+B: Combination of mask and box prompts." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 49, + 270, + 286, + 418 + ], + "lines": [ + { + "bbox": [ + 49, + 270, + 286, + 418 + ], + "spans": [ + { + "bbox": [ + 49, + 270, + 286, + 418 + ], + "type": "table", + "html": "
MethodSα↑Fwβ↑Eφ↑M↓mDice↑mIoU↑
SINet[10]0.6360.3460.7750.0410.3810.283
SINet-v2[12]0.6530.3820.7620.0390.4130.318
ZoomNet[32]0.6330.3490.6010.0330.3490.273
FEDERNet[16]0.5730.2720.6090.0510.2600.199
FSPNet[19]0.6810.4010.7160.0440.4460.332
PUENet[44]0.6910.4850.7950.0340.5140.396
RCRNet[42]0.6270.2870.6660.0480.3090.229
PNS-Net[20]0.6550.3250.6730.0480.3840.290
MG[43]0.5940.3360.6910.0590.3680.268
SLT-Net[5]0.6960.4810.8450.0300.4930.401
TSP-SAM(M+P)0.7050.5650.8360.0270.5310.422
TSP-SAM(M+B)0.7510.6280.8510.0210.6030.496
", + "image_path": "a08b6536c88c2b1cd7a925b01bf31e5cd1433a7d66458caef1023c6bd65e6201.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 481, + 287, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 481, + 287, + 518 + ], + "spans": [ + { + "bbox": [ + 46, + 481, + 287, + 518 + ], + "type": "text", + "content": "meanDice (mDice), meanIoU (mIoU). Overall, a better VCOD method has larger " + }, + { + "bbox": [ + 46, + 481, + 287, + 518 + ], + "type": "inline_equation", + "content": "S_{\\alpha}" + }, + { + "bbox": [ + 46, + 481, + 287, + 518 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 481, + 287, + 518 + ], + "type": "inline_equation", + "content": "F_{\\beta}^{w}" + }, + { + "bbox": [ + 46, + 481, + 287, + 518 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 481, + 287, + 518 + ], + "type": "inline_equation", + "content": "E_{\\phi}" + }, + { + "bbox": [ + 46, + 481, + 287, + 518 + ], + "type": "text", + "content": ", mDice and mIoU scores, but a smaller MAE score." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 522, + 289, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 522, + 289, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 522, + 289, + 713 + ], + "type": "text", + "content": "Implementation details. In our experiments, we freeze SAM and train our TSP-SAM in an end-to-end manner with a maximum of 20 epochs. Specifically, for a fair comparison, both the training and testing images are resize to " + }, + { + "bbox": [ + 46, + 522, + 289, + 713 + ], + "type": "inline_equation", + "content": "352 \\times 352" + }, + { + "bbox": [ + 46, + 522, + 289, + 713 + ], + "type": "text", + "content": ". We leverage Adam [24] as our optimizer and set the weight decay to " + }, + { + "bbox": [ + 46, + 522, + 289, + 713 + ], + "type": "inline_equation", + "content": "1 \\times 10^{-5}" + }, + { + "bbox": [ + 46, + 522, + 289, + 713 + ], + "type": "text", + "content": ". On the MoCA-Mask dataset, the learning rate is initialized to " + }, + { + "bbox": [ + 46, + 522, + 289, + 713 + ], + "type": "inline_equation", + "content": "1 \\times 10^{-5}" + }, + { + "bbox": [ + 46, + 522, + 289, + 713 + ], + "type": "text", + "content": " for the spatial extractor of motion-driven self-prompt learning and " + }, + { + "bbox": [ + 46, + 522, + 289, + 713 + ], + "type": "inline_equation", + "content": "5 \\times 10^{-5}" + }, + { + "bbox": [ + 46, + 522, + 289, + 713 + ], + "type": "text", + "content": " for other trainable parameters. After 10 epochs, the learning rate is reduced by " + }, + { + "bbox": [ + 46, + 522, + 289, + 713 + ], + "type": "inline_equation", + "content": "90\\%" + }, + { + "bbox": [ + 46, + 522, + 289, + 713 + ], + "type": "text", + "content": ". On the CAD2016 dataset, we adopt a ratio of 9:1 to divide all classes into training and testing sets and report the mean value over the testing set. We set the initial learning rate to " + }, + { + "bbox": [ + 46, + 522, + 289, + 713 + ], + "type": "inline_equation", + "content": "1 \\times 10^{-4}" + }, + { + "bbox": [ + 46, + 522, + 289, + 713 + ], + "type": "text", + "content": " for the spatial extractor of the motion-driven self-prompt learning and " + }, + { + "bbox": [ + 46, + 522, + 289, + 713 + ], + "type": "inline_equation", + "content": "5 \\times 10^{-4}" + }, + { + "bbox": [ + 46, + 522, + 289, + 713 + ], + "type": "text", + "content": " for other trainable parameters. The other settings are the same as those on MoCA-Mask dataset." + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 318, + 270, + 541, + 323 + ], + "blocks": [ + { + "bbox": [ + 46, + 426, + 287, + 460 + ], + "lines": [ + { + "bbox": [ + 46, + 426, + 287, + 460 + ], + "spans": [ + { + "bbox": [ + 46, + 426, + 287, + 460 + ], + "type": "text", + "content": "Table 2. Generalization comparisons on CAD2016 dataset. The best results are highlighted in **bold**. M+P: Combination of mask and point prompts. M+B: Combination of mask and box prompts." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 318, + 270, + 541, + 323 + ], + "lines": [ + { + "bbox": [ + 318, + 270, + 541, + 323 + ], + "spans": [ + { + "bbox": [ + 318, + 270, + 541, + 323 + ], + "type": "table", + "html": "
MethodTuning parameters (M)MethodTime (s) ↓Speed (fps) ↑
FSPNet274.24FSPNet568.701.31
SLT-Net253.402.94
SLT-Net82.41Overall (ours)294.992.53
TSP-SAM(ours)89.78Prompt learning73.2810.17
SAM221.713.36
", + "image_path": "f8ecbcfbb9f2b40bd20d2c5c257798c91d9e85d4c96bea791c75413e984f99c7.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 317, + 327, + 546, + 351 + ], + "lines": [ + { + "bbox": [ + 317, + 327, + 546, + 351 + ], + "spans": [ + { + "bbox": [ + 317, + 327, + 546, + 351 + ], + "type": "text", + "content": "Table 3. The number of Table 4. Comparisons on infertuning parameters. Hence time and speed." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 306, + 367, + 446, + 380 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 367, + 446, + 380 + ], + "spans": [ + { + "bbox": [ + 306, + 367, + 446, + 380 + ], + "type": "text", + "content": "5.2. Performance comparison" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 387, + 546, + 423 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 387, + 546, + 423 + ], + "spans": [ + { + "bbox": [ + 306, + 387, + 546, + 423 + ], + "type": "text", + "content": "1) Baselines: To evaluate the performance of the proposed TSP-SAM, we compare it with a range of state-of-the-art methods. These methods fall into two categories:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 425, + 547, + 714 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 306, + 425, + 547, + 507 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 425, + 547, + 507 + ], + "spans": [ + { + "bbox": [ + 306, + 425, + 547, + 507 + ], + "type": "text", + "content": "- Single-image camouflaged object detection. These methods analyze the static appearance of images to discern the subtle differences between camouflaged objects and their surroundings. The compared methods in this type include SINet [10], SINet-v2 [12], ZoomNet [32], BGNet [36], FEDERNet [16], FSPNet[19], and PUENet [44]." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 509, + 547, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 509, + 547, + 567 + ], + "spans": [ + { + "bbox": [ + 306, + 509, + 547, + 567 + ], + "type": "text", + "content": "- Video object segmentation. This set of methods explores the temporal-spatial relationships between frames for the purpose of segmenting objects with specific attributes. The compared methods in this type encompass RCRNet [42], PNS-Net [20], MG[43], and SLT-Net[5]." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 570, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 547, + 714 + ], + "type": "text", + "content": "2) Quantitative Results: We evaluate TSP-SAM on MoCA-Mask and CAD2016 datasets. From Table 1, we can conclude that, (i) compared to the combination of mask and point prompts, the combination of mask and box prompts is more reliable for SAM. This is attributed to the limitations of the point prompt in conveying the boundary information. In contrast, bounding box excels in indicating the boundaries of the camouflaged object, even in the presence of bias. (ii) the strong contrast between the performance of TSP-SAM and single-image camouflaged object detection methods points to the importance of temporal-spatial relationships in breaking camouflage. (iii) the proposed TSP-SAM" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19063" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 51, + 70, + 294, + 209 + ], + "blocks": [ + { + "bbox": [ + 51, + 70, + 294, + 209 + ], + "lines": [ + { + "bbox": [ + 51, + 70, + 294, + 209 + ], + "spans": [ + { + "bbox": [ + 51, + 70, + 294, + 209 + ], + "type": "image", + "image_path": "8da5257c5b97459ed1fb85348d80fbdc4e58f925855d13587a69f56d3e9176f9.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 220, + 547, + 254 + ], + "lines": [ + { + "bbox": [ + 46, + 220, + 547, + 254 + ], + "spans": [ + { + "bbox": [ + 46, + 220, + 547, + 254 + ], + "type": "text", + "content": "Figure 4. Visualization of our proposed TSP-SAM and baseline methods on MoCA-Mask and CAD2016 dataset. From left to right: frames (1st column), ground truth (2nd column), prediction of our TSP-SAM with mask and box prompts (3rd column), prediction of our TSP-SAM with mask and point prompts (4th column), and predictions of compared methods (5th-7th columns)." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 299, + 70, + 541, + 209 + ], + "blocks": [ + { + "bbox": [ + 299, + 70, + 541, + 209 + ], + "lines": [ + { + "bbox": [ + 299, + 70, + 541, + 209 + ], + "spans": [ + { + "bbox": [ + 299, + 70, + 541, + 209 + ], + "type": "image", + "image_path": "a17edc19376520f40cf5d47563738915e5fab56f4b40ad5a8ccd7a594911b2da.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 48, + 258, + 286, + 406 + ], + "blocks": [ + { + "bbox": [ + 48, + 258, + 286, + 406 + ], + "lines": [ + { + "bbox": [ + 48, + 258, + 286, + 406 + ], + "spans": [ + { + "bbox": [ + 48, + 258, + 286, + 406 + ], + "type": "image", + "image_path": "a3945ae3ac33cebc2a3871b4ea4bac3421814b9aa88d51fddf8307c22f32e3ca.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 415, + 288, + 481 + ], + "lines": [ + { + "bbox": [ + 46, + 415, + 288, + 481 + ], + "spans": [ + { + "bbox": [ + 46, + 415, + 288, + 481 + ], + "type": "text", + "content": "Figure 5. Visualization of several failure cases with mask and point prompts. From left to right: frames (1st column), ground truth (2nd column), mask and box prompts (3rd column), prediction with mask and box prompts (4th column), mask and point prompts (5th column), and prediction with mask and point prompts (6th column)." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 495, + 287, + 543 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 495, + 287, + 543 + ], + "spans": [ + { + "bbox": [ + 46, + 495, + 287, + 543 + ], + "type": "text", + "content": "surpasses all video object segmentation methods. Notably, TSP-SAM exhibits a " + }, + { + "bbox": [ + 46, + 495, + 287, + 543 + ], + "type": "inline_equation", + "content": "1.3\\% - 8.7\\%" + }, + { + "bbox": [ + 46, + 495, + 287, + 543 + ], + "type": "text", + "content": " improvement over SLT-Net [5] on MoCA-Mask dataset, meaning that the temporal-spatial exploration in TSP-SAM is more effective." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 546, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 546, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 546, + 288, + 714 + ], + "type": "text", + "content": "To assess the generalization performance, we train on the MoCA-Mask dataset and report the test performance on the CAD2016 dataset. It can be observed from Table 2 that the proposed TSP-SAM outperforms all baselines. Compared with SLTNet [5], TSP-SAM surpasses it by " + }, + { + "bbox": [ + 46, + 546, + 288, + 714 + ], + "type": "inline_equation", + "content": "0.9\\% - 14.7\\%" + }, + { + "bbox": [ + 46, + 546, + 288, + 714 + ], + "type": "text", + "content": ", indicating that TSP-SAM is capable of generalizing to the unseen data more effectively. Meanwhile, we also test the performance of TSP-SAM on general video object segmentation datasets. On DAVIS2016 dataset [34], our model achieves comparable performance to the SOTA SLT-Net. While on SegTrack-v2 dataset [27] with multiple objects, our method obtains " + }, + { + "bbox": [ + 46, + 546, + 288, + 714 + ], + "type": "inline_equation", + "content": "62.84\\%" + }, + { + "bbox": [ + 46, + 546, + 288, + 714 + ], + "type": "text", + "content": " for mIoU, surpassing SLT-Net by " + }, + { + "bbox": [ + 46, + 546, + 288, + 714 + ], + "type": "inline_equation", + "content": "9.9\\%" + }, + { + "bbox": [ + 46, + 546, + 288, + 714 + ], + "type": "text", + "content": ". Surprisingly, our model is even comparable to the recent SMTC [35] on SegTrack-v2." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 260, + 545, + 331 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 260, + 545, + 331 + ], + "spans": [ + { + "bbox": [ + 304, + 260, + 545, + 331 + ], + "type": "text", + "content": "Besides, we report the number of tuning parameters, the inference time and speed. From Table 3, compared with SLT-Net, our model achieves " + }, + { + "bbox": [ + 304, + 260, + 545, + 331 + ], + "type": "inline_equation", + "content": "9.6\\%" + }, + { + "bbox": [ + 304, + 260, + 545, + 331 + ], + "type": "text", + "content": " improvements with a slight parameter increase. From Table 4, our model is comparable to the latest video-oriented SLT-Net. Decomposing the overall inference time, SAM consumes " + }, + { + "bbox": [ + 304, + 260, + 545, + 331 + ], + "type": "inline_equation", + "content": "75\\%" + }, + { + "bbox": [ + 304, + 260, + 545, + 331 + ], + "type": "text", + "content": " of it." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 333, + 546, + 476 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 333, + 546, + 476 + ], + "spans": [ + { + "bbox": [ + 304, + 333, + 546, + 476 + ], + "type": "text", + "content": "3) Qualitative Results: To provide more intuitive evaluations of the proposed TSP-SAM, we conduct two qualitative discussions. Firstly, as shown in Fig. 4, we present the segmentation results of several examples to intuitively compare the proposed TSP-SAM with the baselines. It can be seen that: (i) some confusing surrounding regions interfere with the baseline methods, while our TSP-SAM is more robust in identifying the camouflaged object. This demonstrates the necessity of short-long term temporal-spatial relationships. (ii) compared with the baselines, our TSP-SAM exhibits superior capabilities in boundary segmentation in terms of discriminability and details." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 478, + 547, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 478, + 547, + 647 + ], + "spans": [ + { + "bbox": [ + 304, + 478, + 547, + 647 + ], + "type": "text", + "content": "Secondly, to further analyze the influence of the type of sparse prompts on overall performance, we visualize the several failure cases with mask and point prompts. From Fig. 5, it can be found that: (i) since the target object is not always a regular shape, the learned point prompt is not on the target object although it is near the center. (ii) under the condition with unclear boundaries, the point prompt is more likely to segment the regions of background. (iii) the point prompt may induce the model to segment the local areas instead of the complete target object. In contrast, the box prompt can avoid the above defects and achieve more precise segmentation results. Meanwhile, the complete localization under the learned mask and box prompts verifies the effectiveness of the temporal-spatial prompt learning." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 658, + 397, + 671 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 658, + 397, + 671 + ], + "spans": [ + { + "bbox": [ + 306, + 658, + 397, + 671 + ], + "type": "text", + "content": "5.3. Ablation study" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 677, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 547, + 713 + ], + "type": "text", + "content": "To gain a comprehensive understanding of the efficacy of core components and the impact of hyper parameters, we delve into a detailed analysis." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "19064" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 49, + 70, + 286, + 133 + ], + "blocks": [ + { + "bbox": [ + 49, + 70, + 286, + 133 + ], + "lines": [ + { + "bbox": [ + 49, + 70, + 286, + 133 + ], + "spans": [ + { + "bbox": [ + 49, + 70, + 286, + 133 + ], + "type": "table", + "html": "
MethodSα↑Fwβ↑Eφ↑mDice↑mIoU↑
SAM+SSP0.6370.3300.7680.3670.311
SAM+MSP0.6570.3530.8230.4070.340
SAM+MSP+LCP0.6650.4090.7570.4290.366
SAM+MSP+LCP+TSI0.6890.4440.8080.4580.388
", + "image_path": "8b32d370e610f4f65f28997bb3c7881e37e447ae2d998d5f48f26d15e12cc098.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 71, + 202, + 165, + 258 + ], + "blocks": [ + { + "bbox": [ + 71, + 202, + 165, + 258 + ], + "lines": [ + { + "bbox": [ + 71, + 202, + 165, + 258 + ], + "spans": [ + { + "bbox": [ + 71, + 202, + 165, + 258 + ], + "type": "image", + "image_path": "a45acdd249c5cf4f00330c01d64c17a9ff13c154f770b0e6a95b19428f496075.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 110, + 259, + 143, + 268 + ], + "lines": [ + { + "bbox": [ + 110, + 259, + 143, + 268 + ], + "spans": [ + { + "bbox": [ + 110, + 259, + 143, + 268 + ], + "type": "text", + "content": "(a) mDice" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 167, + 203, + 255, + 258 + ], + "blocks": [ + { + "bbox": [ + 167, + 203, + 255, + 258 + ], + "lines": [ + { + "bbox": [ + 167, + 203, + 255, + 258 + ], + "spans": [ + { + "bbox": [ + 167, + 203, + 255, + 258 + ], + "type": "image", + "image_path": "5f384288d61922770d1c5b2a73a8eb8a14ce557e8339c87f58b367a311632653.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 205, + 259, + 230, + 268 + ], + "lines": [ + { + "bbox": [ + 205, + 259, + 230, + 268 + ], + "spans": [ + { + "bbox": [ + 205, + 259, + 230, + 268 + ], + "type": "text", + "content": "(b) " + }, + { + "bbox": [ + 205, + 259, + 230, + 268 + ], + "type": "inline_equation", + "content": "F_{\\beta}^{w}" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 72, + 269, + 165, + 324 + ], + "blocks": [ + { + "bbox": [ + 72, + 269, + 165, + 324 + ], + "lines": [ + { + "bbox": [ + 72, + 269, + 165, + 324 + ], + "spans": [ + { + "bbox": [ + 72, + 269, + 165, + 324 + ], + "type": "image", + "image_path": "c288c6aa9ac6ce27b30d87c153ccab97631d5fac17596de03f8236e7ba54ad47.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 111, + 325, + 141, + 335 + ], + "lines": [ + { + "bbox": [ + 111, + 325, + 141, + 335 + ], + "spans": [ + { + "bbox": [ + 111, + 325, + 141, + 335 + ], + "type": "text", + "content": "(c) mIoU" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 167, + 269, + 254, + 324 + ], + "blocks": [ + { + "bbox": [ + 167, + 269, + 254, + 324 + ], + "lines": [ + { + "bbox": [ + 167, + 269, + 254, + 324 + ], + "spans": [ + { + "bbox": [ + 167, + 269, + 254, + 324 + ], + "type": "image", + "image_path": "57903b9ca011a475c431aaef47d5125ea9f421d5d8f79ca6ea02148fcb79ba25.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 206, + 325, + 227, + 336 + ], + "lines": [ + { + "bbox": [ + 206, + 325, + 227, + 336 + ], + "spans": [ + { + "bbox": [ + 206, + 325, + 227, + 336 + ], + "type": "text", + "content": "(d) " + }, + { + "bbox": [ + 206, + 325, + 227, + 336 + ], + "type": "inline_equation", + "content": "E_{\\phi}" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 376, + 287, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 376, + 287, + 544 + ], + "spans": [ + { + "bbox": [ + 46, + 376, + 287, + 544 + ], + "type": "text", + "content": "Ablation analysis: Table 5 present the segmentation results by progressively applying each part in our model. Compared to the spatial-based self-prompt learning (SSP) comprising a spatial extractor and two prompt heads, motion-driven self-prompt learning (MSP) improves the mDice metric from " + }, + { + "bbox": [ + 46, + 376, + 287, + 544 + ], + "type": "inline_equation", + "content": "36.7\\%" + }, + { + "bbox": [ + 46, + 376, + 287, + 544 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 46, + 376, + 287, + 544 + ], + "type": "inline_equation", + "content": "40.7\\%" + }, + { + "bbox": [ + 46, + 376, + 287, + 544 + ], + "type": "text", + "content": " to a large degree. This affirms the effectiveness of the inter-frame motion in camouflage breaking. Subsequently, the long-range consistency-based robust prompt learning (LCP) brings " + }, + { + "bbox": [ + 46, + 376, + 287, + 544 + ], + "type": "inline_equation", + "content": "2.2\\%" + }, + { + "bbox": [ + 46, + 376, + 287, + 544 + ], + "type": "text", + "content": " gains for mDice, indicating the effectiveness of long-range consistency in improving the robustness of self-prompts. Meanwhile, the temporal-spatial injection (TSI) further boosts the mDice metric by " + }, + { + "bbox": [ + 46, + 376, + 287, + 544 + ], + "type": "inline_equation", + "content": "2.9\\%" + }, + { + "bbox": [ + 46, + 376, + 287, + 544 + ], + "type": "text", + "content": ", once again verifying the significance of long-range consistency." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 544, + 286, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 544, + 286, + 664 + ], + "spans": [ + { + "bbox": [ + 46, + 544, + 286, + 664 + ], + "type": "text", + "content": "To further analyze the effectiveness of frequency-based motion perception in MSP, we replace it with spatial-based motion perception, frequency-based static appearance perception, and spatial-based static appearance perception to conduct the ablation studies. As can be seen from Table 6, the frequency-based motion perception yields the best performance. While verifying the significant role of interframe motion in breaking camouflage, it also demonstrates that inter-frame motion can be expressed more exhaustively in the frequency domain than in the spatial domain." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "text", + "content": "Hyper parameter analysis: To investigate the effect of the size of the windowed sequence on segmentation performance, we conduct the ablation experiments by varying " + }, + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\{5,6,7,8,9,10\\}" + }, + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "text", + "content": ". As illustrated in Fig. 6, the ac" + } + ] + } + ], + "index": 13 + }, + { + "type": "table", + "bbox": [ + 307, + 70, + 544, + 133 + ], + "blocks": [ + { + "bbox": [ + 46, + 140, + 287, + 196 + ], + "lines": [ + { + "bbox": [ + 46, + 140, + 287, + 196 + ], + "spans": [ + { + "bbox": [ + 46, + 140, + 287, + 196 + ], + "type": "text", + "content": "Table 5. Ablation studies of the core components on MoCA-Mask dataset. SSP: Spatial-based self-prompt learning. MSP: Motion-driven self-prompt learning. LCP: Long-range consistency-based robust prompt learning. TSI: Temporal-spatial injection for representation enhancement." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 307, + 70, + 544, + 133 + ], + "lines": [ + { + "bbox": [ + 307, + 70, + 544, + 133 + ], + "spans": [ + { + "bbox": [ + 307, + 70, + 544, + 133 + ], + "type": "table", + "html": "
MethodSα↑Fwβ↑Eφ↑mDice↑mIoU↑
Spatial appearance0.6370.3300.7680.3670.311
Frequency appearance0.6470.3440.7930.3960.328
Spatial motion0.6510.3350.7300.3960.324
Frequency motion0.6570.3530.8230.4070.340
", + "image_path": "463fada25ca77e543580f62d009e0e765cb4d6ee07d616e3593fe077c1372140.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "table_body" + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 141, + 545, + 197 + ], + "lines": [ + { + "bbox": [ + 304, + 141, + 545, + 197 + ], + "spans": [ + { + "bbox": [ + 304, + 141, + 545, + 197 + ], + "type": "text", + "content": "Table 6. Ablation studies of frequency-based motion perception on MoCA-Mask dataset. Spatial appearance: Spatial-based appearance perception. Frequency appearance: Frequency-based appearance perception. Spatial motion: Spatial-based motion perception. Frequency motion: Frequency-based motion perception." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 345, + 202, + 417, + 256 + ], + "blocks": [ + { + "bbox": [ + 345, + 202, + 417, + 256 + ], + "lines": [ + { + "bbox": [ + 345, + 202, + 417, + 256 + ], + "spans": [ + { + "bbox": [ + 345, + 202, + 417, + 256 + ], + "type": "image", + "image_path": "3def31d2b60d3effa059bbf903b374e42f8ecaf6fd254574fe74b003f1087ec9.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 358, + 258, + 390, + 266 + ], + "lines": [ + { + "bbox": [ + 358, + 258, + 390, + 266 + ], + "spans": [ + { + "bbox": [ + 358, + 258, + 390, + 266 + ], + "type": "text", + "content": "(a) mDice" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 440, + 203, + 512, + 256 + ], + "blocks": [ + { + "bbox": [ + 440, + 203, + 512, + 256 + ], + "lines": [ + { + "bbox": [ + 440, + 203, + 512, + 256 + ], + "spans": [ + { + "bbox": [ + 440, + 203, + 512, + 256 + ], + "type": "image", + "image_path": "f25d3e7d8c5bf9e59fd7d995fcf5e440d24ef30dd60c990d708cdcddb499a2bd.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 346, + 269, + 417, + 323 + ], + "blocks": [ + { + "bbox": [ + 46, + 345, + 287, + 367 + ], + "lines": [ + { + "bbox": [ + 46, + 345, + 287, + 367 + ], + "spans": [ + { + "bbox": [ + 46, + 345, + 287, + 367 + ], + "type": "text", + "content": "Figure 6. Sensitivity analysis of the size " + }, + { + "bbox": [ + 46, + 345, + 287, + 367 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 46, + 345, + 287, + 367 + ], + "type": "text", + "content": " of the windowed sequence on MoCA-Mask dataset." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 346, + 269, + 417, + 323 + ], + "lines": [ + { + "bbox": [ + 346, + 269, + 417, + 323 + ], + "spans": [ + { + "bbox": [ + 346, + 269, + 417, + 323 + ], + "type": "image", + "image_path": "31c48953916a6a652ba587bcaf0c52f0e7665dd5a7bdd58d6ae4db2d8dc7df05.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 359, + 323, + 389, + 332 + ], + "lines": [ + { + "bbox": [ + 359, + 323, + 389, + 332 + ], + "spans": [ + { + "bbox": [ + 359, + 323, + 389, + 332 + ], + "type": "text", + "content": "(c) mIoU" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 304, + 342, + 545, + 364 + ], + "lines": [ + { + "bbox": [ + 304, + 342, + 545, + 364 + ], + "spans": [ + { + "bbox": [ + 304, + 342, + 545, + 364 + ], + "type": "text", + "content": "Figure 7. Sensitivity analysis of the balance factor " + }, + { + "bbox": [ + 304, + 342, + 545, + 364 + ], + "type": "inline_equation", + "content": "(\\alpha, \\beta)" + }, + { + "bbox": [ + 304, + 342, + 545, + 364 + ], + "type": "text", + "content": " in the loss function on MoCA-Mask dataset." + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 441, + 268, + 512, + 323 + ], + "blocks": [ + { + "bbox": [ + 464, + 258, + 487, + 267 + ], + "lines": [ + { + "bbox": [ + 464, + 258, + 487, + 267 + ], + "spans": [ + { + "bbox": [ + 464, + 258, + 487, + 267 + ], + "type": "text", + "content": "(b) " + }, + { + "bbox": [ + 464, + 258, + 487, + 267 + ], + "type": "inline_equation", + "content": "F_{\\beta}^{w}" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 441, + 268, + 512, + 323 + ], + "lines": [ + { + "bbox": [ + 441, + 268, + 512, + 323 + ], + "spans": [ + { + "bbox": [ + 441, + 268, + 512, + 323 + ], + "type": "image", + "image_path": "5374b09c41621c29624e7c919a04db55bca4e7388bbc10bc109c022432701e52.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 464, + 323, + 484, + 335 + ], + "lines": [ + { + "bbox": [ + 464, + 323, + 484, + 335 + ], + "spans": [ + { + "bbox": [ + 464, + 323, + 484, + 335 + ], + "type": "text", + "content": "(d) " + }, + { + "bbox": [ + 464, + 323, + 484, + 335 + ], + "type": "inline_equation", + "content": "E_{\\phi}" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 22 + }, + { + "bbox": [ + 304, + 371, + 545, + 444 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 371, + 545, + 444 + ], + "spans": [ + { + "bbox": [ + 304, + 371, + 545, + 444 + ], + "type": "text", + "content": "curacy initially rises with increasing the size of the windowed sequence, reaches the best at " + }, + { + "bbox": [ + 304, + 371, + 545, + 444 + ], + "type": "inline_equation", + "content": "T = 7" + }, + { + "bbox": [ + 304, + 371, + 545, + 444 + ], + "type": "text", + "content": " before declining. Moreover, we assess the sensitivity of " + }, + { + "bbox": [ + 304, + 371, + 545, + 444 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 304, + 371, + 545, + 444 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 371, + 545, + 444 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 304, + 371, + 545, + 444 + ], + "type": "text", + "content": " that are used to balance the sparse and dense prompt loss in Fig. 7. The optimal segmentation performance is observed at " + }, + { + "bbox": [ + 304, + 371, + 545, + 444 + ], + "type": "inline_equation", + "content": "(\\alpha, \\beta) = (1,0.6)" + }, + { + "bbox": [ + 304, + 371, + 545, + 444 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 306, + 453, + 378, + 465 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 453, + 378, + 465 + ], + "spans": [ + { + "bbox": [ + 306, + 453, + 378, + 465 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 304, + 473, + 545, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 473, + 545, + 617 + ], + "spans": [ + { + "bbox": [ + 304, + 473, + 545, + 617 + ], + "type": "text", + "content": "In this paper, we propose a flexible end-to-end temporal-spatial prompt SAM (TSP-SAM) for video camouflaged object detection. Instead of user-provided prompts, motion-driven self-prompt learning is proposed to capture the overall motion of the camouflaged object, facilitating its spatial identification. Subsequently, a robust prompt learning based on long-range consistency is proposed to promote the robustness of the self-prompts, which likes endowing SAM with keen eyes. Moreover, the long-range consistency is injected into the encoder of SAM for representation enhancement. Notably, TSP-SAM provides a new perspective, that is temporal-spatial associations, for visual prompt learning." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 306, + 621, + 416, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 621, + 416, + 635 + ], + "spans": [ + { + "bbox": [ + 306, + 621, + 416, + 635 + ], + "type": "text", + "content": "7. Acknowledgement" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 304, + 641, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 641, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 641, + 545, + 712 + ], + "type": "text", + "content": "This work was supported in part by the National Key Research and Development Program of China under Grant No.2021ZD0140407, Beijing Natural Science Foundation under Grant No.7222313, and National High Level Hospital Clinical Research Funding under Grant No.2022-PUMCHC-041." + } + ] + } + ], + "index": 29 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19065" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 124 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 124 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 124 + ], + "type": "text", + "content": "[1] Pia Bideau and Erik Learned-Miller. It's moving! a probabilistic model for causal motion segmentation in moving camera videos. In ECCV, pages 433-449, 2016. 1, 2, 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 125, + 287, + 158 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 125, + 287, + 158 + ], + "spans": [ + { + "bbox": [ + 53, + 125, + 287, + 158 + ], + "type": "text", + "content": "[2] Pia Bideau, Rakesh R Menon, and Erik Learned-Miller. Moa-net: self-supervised motion segmentation. In ECCV Workshops, pages 0–0, 2018. 1, 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 159, + 288, + 193 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 159, + 288, + 193 + ], + "spans": [ + { + "bbox": [ + 53, + 159, + 288, + 193 + ], + "type": "text", + "content": "[3] Alexey Bochkovskiy, Chien-Yao Wang, and Hong-Yuan Mark Liao. Yolov4: Optimal speed and accuracy of object detection. arXiv preprint arXiv:2004.10934, 2020. 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 194, + 287, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 194, + 287, + 258 + ], + "spans": [ + { + "bbox": [ + 53, + 194, + 287, + 258 + ], + "type": "text", + "content": "[4] Tianrun Chen, Lanyun Zhu, Chaotao Ding, Runlong Cao, Shangzhan Zhang, Yan Wang, Zejian Li, Lingyun Sun, Papa Mao, and Ying Zang. Sam fails to segment anything?—sam-adapter: Adapting sam in underperformed scenes: Camouflage, shadow, and more. arXiv preprint arXiv:2304.09148, 2023. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 260, + 287, + 314 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 260, + 287, + 314 + ], + "spans": [ + { + "bbox": [ + 53, + 260, + 287, + 314 + ], + "type": "text", + "content": "[5] Xuelian Cheng, Huan Xiong, Deng-Ping Fan, Yiran Zhong, Mehrtash Harandi, Tom Drummond, and Zongyuan Ge. Implicit motion handling for video camouflaged object detection. In CVPR, pages 13864-13873, 2022. 1, 2, 4, 5, 6, 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 317, + 287, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 317, + 287, + 360 + ], + "spans": [ + { + "bbox": [ + 53, + 317, + 287, + 360 + ], + "type": "text", + "content": "[6] Runmin Cong, Mengyao Sun, Sanyi Zhang, Xiaofei Zhou, Wei Zhang, and Yao Zhao. Frequency perception network for camouflaged object detection. In ACM MM, pages 1179-1189, 2023. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 362, + 287, + 395 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 362, + 287, + 395 + ], + "spans": [ + { + "bbox": [ + 53, + 362, + 287, + 395 + ], + "type": "text", + "content": "[7] Xinyi Cui, Qingshan Liu, and Dimitris Metaxas. Temporal spectral residual: fast motion saliency detection. In ACM MM, pages 617-620, 2009. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 396, + 287, + 440 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 396, + 287, + 440 + ], + "spans": [ + { + "bbox": [ + 53, + 396, + 287, + 440 + ], + "type": "text", + "content": "[8] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, et al. An image is worth 16x16 words: Transformers for image recognition at scale. 2020. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 442, + 287, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 442, + 287, + 475 + ], + "spans": [ + { + "bbox": [ + 53, + 442, + 287, + 475 + ], + "type": "text", + "content": "[9] Deng-Ping Fan, Ming-Ming Cheng, Yun Liu, Tao Li, and Ali Borji. Structure-measure: A new way to evaluate foreground maps. In ICCV, pages 4548-4557, 2017. 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 476, + 287, + 509 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 476, + 287, + 509 + ], + "spans": [ + { + "bbox": [ + 48, + 476, + 287, + 509 + ], + "type": "text", + "content": "[10] Deng-Ping Fan, Ge-Peng Ji, Guolei Sun, Ming-Ming Cheng, Jianbing Shen, and Ling Shao. Camouflaged object detection. In CVPR, pages 2777-2787, 2020. 2, 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 510, + 287, + 565 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 510, + 287, + 565 + ], + "spans": [ + { + "bbox": [ + 48, + 510, + 287, + 565 + ], + "type": "text", + "content": "[11] Deng-Ping Fan, Ge-Peng Ji, Tao Zhou, Geng Chen, Huazhu Fu, Jianbing Shen, and Ling Shao. Pranet: Parallel reverse attention network for polyp segmentation. In International conference on medical image computing and computer-assisted intervention, pages 263-273, 2020. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 567, + 287, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 567, + 287, + 599 + ], + "spans": [ + { + "bbox": [ + 48, + 567, + 287, + 599 + ], + "type": "text", + "content": "[12] Deng-Ping Fan, Ge-Peng Ji, Ming-Ming Cheng, and Ling Shao. Concealed object detection. IEEE TPAMI, 44(10): 6024-6042, 2021. 2, 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 601, + 287, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 601, + 287, + 643 + ], + "spans": [ + { + "bbox": [ + 48, + 601, + 287, + 643 + ], + "type": "text", + "content": "[13] Deng-Ping Fan, Ge-Peng Ji, Xuebin Qin, and Ming-Ming Cheng. Cognitive vision inspired object segmentation metric and loss function. Scientia Sinica Informationis, 6(6), 2021. 5" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 646, + 287, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 646, + 287, + 678 + ], + "spans": [ + { + "bbox": [ + 48, + 646, + 287, + 678 + ], + "type": "text", + "content": "[14] Melvyn A Goodale and A David Milner. Separate visual pathways for perception and action. Trends in neurosciences, 15(1):20-25, 1992. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "type": "text", + "content": "[15] Kensho Hara, Hirokatsu Kataoka, and Yutaka Satoh. Can spatiotemporal 3d cnns retrace the history of 2d cnns and imagenet? In CVPR, pages 6546-6555, 2018. 5" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 33, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "text", + "content": "[16] Chunming He, Kai Li, Yachao Zhang, Longxiang Tang, Yulin Zhang, Zhenhua Guo, and Xiu Li. Camouflaged object detection with feature decomposition and edge reconstruction. In CVPR, pages 22046-22055, 2023. 1, 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 118, + 545, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 545, + 150 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 545, + 150 + ], + "type": "text", + "content": "[17] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, pages 770-778, 2016. 4" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 151, + 545, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 151, + 545, + 194 + ], + "spans": [ + { + "bbox": [ + 308, + 151, + 545, + 194 + ], + "type": "text", + "content": "[18] Yuhao Huang, Xin Yang, Lian Liu, Han Zhou, Ao Chang, Xinrui Zhou, Rusi Chen, Junxuan Yu, Jiongquan Chen, Chaoyu Chen, et al. Segment anything model for medical images? arXiv preprint arXiv:2304.14660, 2023. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 195, + 545, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 195, + 545, + 237 + ], + "spans": [ + { + "bbox": [ + 308, + 195, + 545, + 237 + ], + "type": "text", + "content": "[19] Zhou Huang, Hang Dai, Tian-Zhu Xiang, Shuo Wang, Huai-Xin Chen, Jie Qin, and Huan Xiong. Feature shrinkage pyramid for camouflaged object detection with transformers. In CVPR, pages 5557-5566, 2023. 1, 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 239, + 545, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 239, + 545, + 303 + ], + "spans": [ + { + "bbox": [ + 308, + 239, + 545, + 303 + ], + "type": "text", + "content": "[20] Ge-Peng Ji, Yu-Cheng Chou, Deng-Ping Fan, Geng Chen, Huazhu Fu, Debesh Jha, and Ling Shao. Progressively normalized self-attention network for video polyp segmentation. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 142–152, 2021. 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 304, + 545, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 304, + 545, + 348 + ], + "spans": [ + { + "bbox": [ + 308, + 304, + 545, + 348 + ], + "type": "text", + "content": "[21] Ge-Peng Ji, Deng-Ping Fan, Yu-Cheng Chou, Dengxin Dai, Alexander Liniger, and Luc Van Gool. Deep gradient learning for efficient camouflaged object detection. Machine Intelligence Research, 20(1):92-108, 2023. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 350, + 545, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 350, + 545, + 392 + ], + "spans": [ + { + "bbox": [ + 308, + 350, + 545, + 392 + ], + "type": "text", + "content": "[22] Ge-Peng Ji, Deng-Ping Fan, Peng Xu, Ming-Ming Cheng, Bowen Zhou, and Luc Van Gool. \"Sam struggles in concealed scenes—empirical study on\" segment anything\". arXiv preprint arXiv:2304.06022, 2023. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 394, + 545, + 436 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 394, + 545, + 436 + ], + "spans": [ + { + "bbox": [ + 308, + 394, + 545, + 436 + ], + "type": "text", + "content": "[23] Yang Jia, Jie Yuan, Jinjun Wang, Jun Fang, Qixing Zhang, and Yongming Zhang. A saliency-based method for early smoke detection in video sequences. Fire technology, 52: 1271-1292, 2016. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 437, + 545, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 437, + 545, + 459 + ], + "spans": [ + { + "bbox": [ + 308, + 437, + 545, + 459 + ], + "type": "text", + "content": "[24] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. ICLR, 2015. 6" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 460, + 545, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 460, + 545, + 503 + ], + "spans": [ + { + "bbox": [ + 308, + 460, + 545, + 503 + ], + "type": "text", + "content": "[25] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. arXiv preprint arXiv:2304.02643, 2023. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 504, + 545, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 504, + 545, + 536 + ], + "spans": [ + { + "bbox": [ + 308, + 504, + 545, + 536 + ], + "type": "text", + "content": "[26] Hala Lamdouar, Charig Yang, Weidi Xie, and Andrew Zisserman. Betrayed by motion: Camouflaged object discovery via motion segmentation. In ACCV, 2020. 1, 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 537, + 545, + 569 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 537, + 545, + 569 + ], + "spans": [ + { + "bbox": [ + 308, + 537, + 545, + 569 + ], + "type": "text", + "content": "[27] Fuxin Li, Taeyoung Kim, Ahmad Humayun, David Tsai, and James M Rehg. Video segmentation by tracking many figure-ground segments. In ICCV, pages 2192-2199, 2013. 7" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 571, + 545, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 571, + 545, + 601 + ], + "spans": [ + { + "bbox": [ + 308, + 571, + 545, + 601 + ], + "type": "text", + "content": "[28] Thomas Lidbetter. Search and rescue in the face of uncertain threats. European Journal of Operational Research, 285(3): 1153-1160, 2020. 1" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 308, + 603, + 545, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 603, + 545, + 646 + ], + "spans": [ + { + "bbox": [ + 308, + 603, + 545, + 646 + ], + "type": "text", + "content": "[29] Ting Liu, Yao Zhao, Yunchao Wei, Yufeng Zhao, and Shikui Wei. Concealed object detection for activate millimeter wave image. IEEE Transactions on Industrial Electronics, 66(12): 9909-9917, 2019. 1" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 308, + 647, + 545, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 647, + 545, + 679 + ], + "spans": [ + { + "bbox": [ + 308, + 647, + 545, + 679 + ], + "type": "text", + "content": "[30] Jun Ma, Yuting He, Feifei Li, Lin Han, Chenyu You, and Bo Wang. Segment anything in medical images. arXiv preprint arXiv:2304.12306, 2023. 2" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 308, + 681, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 681, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 308, + 681, + 545, + 712 + ], + "type": "text", + "content": "[31] Ran Margolin, Lihi Zelnik-Manor, and Ayellet Tal. How to evaluate foreground maps? In CVPR, pages 248-255, 2014. 5" + } + ] + } + ], + "index": 32 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "text", + "content": "19066" + } + ] + } + ], + "index": 34 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 687 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "text", + "content": "[32] Youwei Pang, Xiaoqi Zhao, Tian-Zhu Xiang, Lihe Zhang, and Huchuan Lu. Zoom in and out: A mixed-scale triplet network for camouflaged object detection. In CVPR, pages 2160-2170, 2022. 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 118, + 287, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 118, + 287, + 161 + ], + "spans": [ + { + "bbox": [ + 48, + 118, + 287, + 161 + ], + "type": "text", + "content": "[33] Federico Perazzi, Philipp Krahenbuhl, Yael Pritch, and Alexander Hornung. Saliency filters: Contrast based filtering for salient region detection. In CVPR, pages 733-740, 2012. 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 163, + 287, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 163, + 287, + 217 + ], + "spans": [ + { + "bbox": [ + 48, + 163, + 287, + 217 + ], + "type": "text", + "content": "[34] Federico Peruzzi, Jordi Pont-Tuset, Brian McWilliams, Luc Van Gool, Markus Gross, and Alexander Sorkine-Hornung. A benchmark dataset and evaluation methodology for video object segmentation. In CVPR, pages 724-732, 2016. 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 219, + 287, + 261 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 219, + 287, + 261 + ], + "spans": [ + { + "bbox": [ + 48, + 219, + 287, + 261 + ], + "type": "text", + "content": "[35] Rui Qian, Shuangrui Ding, Xian Liu, and Dahua Lin. Semantics meets temporal correspondence: Self-supervised object-centric learning in videos. In ICCV, pages 16675-16687, 2023. 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 263, + 287, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 263, + 287, + 296 + ], + "spans": [ + { + "bbox": [ + 48, + 263, + 287, + 296 + ], + "type": "text", + "content": "[36] Yujia Sun, Shuo Wang, Chenglizhao Chen, and Tian-Zhu Xiang. Boundary-guided camouflaged object detection. arXiv preprint arXiv:2207.00794, 2022. 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 297, + 287, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 297, + 287, + 330 + ], + "spans": [ + { + "bbox": [ + 48, + 297, + 287, + 330 + ], + "type": "text", + "content": "[37] Lv Tang, Haoke Xiao, and Bo Li. Can sam segment anything? when sam meets camouflaged object detection. arXiv preprint arXiv:2304.04709, 2023. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 331, + 287, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 331, + 287, + 364 + ], + "spans": [ + { + "bbox": [ + 48, + 331, + 287, + 364 + ], + "type": "text", + "content": "[38] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. NeurIPS, 30, 2017. 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 365, + 287, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 365, + 287, + 418 + ], + "spans": [ + { + "bbox": [ + 48, + 365, + 287, + 418 + ], + "type": "text", + "content": "[39] Wenhai Wang, Enze Xie, Xiang Li, Deng-Ping Fan, Kaitao Song, Ding Liang, Tong Lu, Ping Luo, and Ling Shao. Pyramid vision transformer: A versatile backbone for dense prediction without convolutions. In ICCV, pages 568-578, 2021. 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 421, + 287, + 474 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 421, + 287, + 474 + ], + "spans": [ + { + "bbox": [ + 48, + 421, + 287, + 474 + ], + "type": "text", + "content": "[40] Junde Wu, Rao Fu, Huihui Fang, Yuanpei Liu, Zhaowei Wang, Yanwu Xu, Yueming Jin, and Tal Arbel. Medical sam adapter: Adapting segment anything model for medical image segmentation. arXiv preprint arXiv:2304.12620, 2023. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 476, + 287, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 476, + 287, + 520 + ], + "spans": [ + { + "bbox": [ + 48, + 476, + 287, + 520 + ], + "type": "text", + "content": "[41] Yu-Huan Wu, Shang-Hua Gao, Jie Mei, Jun Xu, Deng-Ping Fan, Rong-Guo Zhang, and Ming-Ming Cheng. Jcs: An explainable Covid-19 diagnosis system by joint classification and segmentation. IEEE TIP, 30:3113-3126, 2021. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 521, + 287, + 564 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 521, + 287, + 564 + ], + "spans": [ + { + "bbox": [ + 48, + 521, + 287, + 564 + ], + "type": "text", + "content": "[42] Pengxiang Yan, Guanbin Li, Yuan Xie, Zhen Li, Chuan Wang, Tianshui Chen, and Liang Lin. Semi-supervised video salient object detection using pseudo-labels. In ICCV, pages 7284-7293, 2019. 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 567, + 287, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 567, + 287, + 609 + ], + "spans": [ + { + "bbox": [ + 48, + 567, + 287, + 609 + ], + "type": "text", + "content": "[43] Charig Yang, Hala Lamdouar, Erika Lu, Andrew Zisserman, and Weidi Xie. Self-supervised video object segmentation by motion grouping. In ICCV, pages 7177-7188, 2021. 1, 2, 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 611, + 287, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 611, + 287, + 643 + ], + "spans": [ + { + "bbox": [ + 48, + 611, + 287, + 643 + ], + "type": "text", + "content": "[44] Yi Zhang, Jing Zhang, Wassim Hamidouche, and Olivier Deforges. Predictive uncertainty estimation for camouflaged object detection. IEEE TIP, 2023. 1, 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 644, + 287, + 687 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 644, + 287, + 687 + ], + "spans": [ + { + "bbox": [ + 48, + 644, + 287, + 687 + ], + "type": "text", + "content": "[45] Yijie Zhong, Bo Li, Lv Tang, Senyun Kuang, Shuang Wu, and Shouhong Ding. Detecting camouflaged object in frequency domain. In CVPR, pages 4504-4513, 2022. 1, 2, 3" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19067" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Enhance Image Classification via Inter-Class Image Mixup with Diffusion Model/d24eac86-8eb2-4146-81d8-2b9455b8e75f_content_list.json b/2024/Enhance Image Classification via Inter-Class Image Mixup with Diffusion Model/d24eac86-8eb2-4146-81d8-2b9455b8e75f_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..a211e447da5d4a17174f9a9c5f24b12fcc7474b0 --- /dev/null +++ b/2024/Enhance Image Classification via Inter-Class Image Mixup with Diffusion Model/d24eac86-8eb2-4146-81d8-2b9455b8e75f_content_list.json @@ -0,0 +1,1661 @@ +[ + { + "type": "text", + "text": "Enhance Image Classification via Inter-Class Image Mixup with Diffusion Model", + "text_level": 1, + "bbox": [ + 78, + 130, + 890, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Zhicai Wang $^{1}$ ; Longhui Wei $^{2\\dagger}$ , Tan Wang $^{3}$ , Heyu Chen $^{1}$ , Yanbin Hao $^{1}$ , Xiang Wang $^{1}$ , Xiangnan He $^{1}$ , Qi Tian $^{2}$", + "bbox": [ + 148, + 179, + 821, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1University of Science and Technology of China, 2 Huawei Inc.,", + "bbox": [ + 230, + 215, + 736, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{3}$ Nanyang Technological University", + "bbox": [ + 341, + 233, + 627, + 252 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 286, + 313, + 301 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Text-to-image (T2I) generative models have recently emerged as a powerful tool, enabling the creation of photorealistic images and giving rise to a multitude of applications. However, the effective integration of T2I models into fundamental image classification tasks remains an open question. A prevalent strategy to bolster image classification performance is through augmenting the training set with synthetic images generated by T2I models. In this study, we scrutinize the shortcomings of both current generative and conventional data augmentation techniques. Our analysis reveals that these methods struggle to produce images that are both faithful (in terms of foreground objects) and diverse (in terms of background contexts) for domain-specific concepts. To tackle this challenge, we introduce an innovative inter-class data augmentation method known as Diff-Mix1, which enriches the dataset by performing image translations between classes. Our empirical results demonstrate that Diff-Mix achieves a better balance between faithfulness and diversity, leading to a marked improvement in performance across diverse image classification scenarios, including few-shot, conventional, and long-tail classifications for domain-specific datasets.", + "bbox": [ + 75, + 316, + 473, + 650 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 676, + 209, + 694 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In comparison to GAN-based models [7, 17, 25], contemporary state-of-the-art text-to-image (T2I) diffusion models exhibit enhanced capabilities in producing high-fidelity images [12, 37, 44, 49]. With the remarkable cross-modality alignment capabilities of T2I models, there is significant potential for generative techniques to enhance image classification [2, 4]. For instance, a straightforward approach entails augmenting the existing training dataset with synthetic images generated by feeding categorical textual prompts to a T2I diffusion model. However, upon reviewing prior ap", + "bbox": [ + 75, + 703, + 468, + 854 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/7ac762f63f12a6572d6b598f0a9e1851bd95d9b404ad63da1bc377f039efb4e6.jpg", + "image_caption": [ + "Figure 1. Strategies to expand domain-specific datasets for improved classification are varied. Row 1 illustrates vanilla distillation from a pretrained text-to-image (T2I) model, which carries the risk of generating outputs with reduced faithfulness. Intra-class augmentation, depicted in Row 2, tends to yield samples with limited diversity to maintain high fidelity to the original class. Our proposed method, showcased in Rows 3 and 4, adopts an inter-class augmentation strategy. This involves introducing edits to a reference image using images from other classes within the training set, which significantly enriches the dataset with a greater diversity of samples." + ], + "image_footnote": [], + "bbox": [ + 504, + 286, + 906, + 508 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "proaches employing T2I diffusion models for image classification, it becomes evident that the challenge in generative data augmentation for domain-specific datasets is producing samples with both a faithful foreground and a diverse background. Depending on whether a reference image is used in the generative process, we divide these methods into two groups:", + "bbox": [ + 496, + 688, + 893, + 794 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "- Text-guided knowledge distillation [52, 57] involves generating new images from scratch using category-related prompts to expand the dataset. For the off-the-shelf T2I models, such vanilla distillation presume these models have comprehensive knowledge of target domain, which can be problematic for domain-specific datasets. Insufficient domain knowledge easily makes the distillation", + "bbox": [ + 500, + 795, + 893, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*This work was done during the internship in Huawei Inc..", + "bbox": [ + 93, + 862, + 403, + 875 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "$\\dagger$ Xiang Wang and Longhui Wei are both the corresponding authors.", + "bbox": [ + 98, + 875, + 450, + 887 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "1https://github.com/Zhicaiww/Diff-Mix", + "bbox": [ + 98, + 887, + 316, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "17223", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "process less effective. For example, vanilla T2I models struggle to generate images that accurately represent specific bird species based solely on their names (see Row 1 of Fig. 1).", + "bbox": [ + 89, + 90, + 467, + 151 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "- Generative data augmentation [1, 69] employs generative models to enhance existing images. Da-fusion [58], for instance, translates the source image into multiple edited versions within the same class. This strategy, termed intra-class augmentation, primarily introduces intra-class variations. While intra-class augmentation retains much of the original image's layout and visual details, it results in limited background diversity (see Row 2 of Fig. 1). However, synthetic images with constrained diversity may not sufficiently enhance the model's ability to discern foreground concepts.", + "bbox": [ + 78, + 152, + 467, + 316 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Based on these observations, a fundamental question emerges: 'Is it feasible to develop a method that optimizes both the diversity and faithfulness of synthesized data simultaneously?'", + "bbox": [ + 76, + 319, + 467, + 378 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this work, we introduce Diff-Mix, a simple yet effective data augmentation method that harnesses diffusion models to perform inter-class image interpolation, tailored for enhancing domain-specific datasets. The method encompasses two pivotal operations: personalized fine-tuning and inter-class image translation. Personalized fine-tuning [15, 46] is originally designed for customizing T2I models and enabling them to generate user-specific contents or styles. In our case, we implement the technique to tailor the model, enabling it to generate images with faithful foreground concepts. Inter-class image translation in Diff-Mix entails transforming a reference image into an edited version that incorporates prompts from different classes. This translation strategy is designed to retain the original background context while editing the foreground to align with the target concept. For instance, as depicted in the bottom rows of Fig. 1, Diff-Mix can generate images of land birds in diverse settings, such as maritime environments, enriching the dataset with a variety of counterfactual samples.", + "bbox": [ + 76, + 380, + 467, + 667 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Unlike previous non-generative augmentation methods, such as Mixup [68] and CutMix [66], Diff-Mix works in a foreground-perceivable inter-class interpolation manner and shares a different mechanism with the non-generative approaches. Our experiment under the conventional classification setting indicates that incorporating both CutMix and Diff-Mix could further enhance performance. Additionally, when compared with other generative approaches, we conduct experiments under few-shot and long-tail scenarios and observe consistent performance improvements. Our contributions can be summarized as follows:", + "bbox": [ + 76, + 669, + 467, + 834 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We pinpoint the critical factors that affect the efficacy of generative data augmentation in domain-specific image classification: namely, faithfulness and diversity.", + "- We introduce Diff-Mix, a simple yet effective generative" + ], + "bbox": [ + 76, + 839, + 467, + 898 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "data augmentation strategy that leverages fine-tuned diffusion models for inter-class image interpolation.", + "bbox": [ + 511, + 90, + 888, + 119 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "- We conduct a comparative analysis of Diff-Mix with other distillation-based and intra-class augmentation methods, as well as non-generative approaches, highlighting its unique features and benefits.", + "bbox": [ + 500, + 121, + 888, + 180 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 500, + 204, + 648, + 219 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Text-to-image diffusion models. Following pretraining on web-scale data, the T2I diffusion model has demonstrated robust capabilities in generating text-controlled images [37, 49, 53, 56]. Its versatility has led to diverse applications, including novel view synthesis [6, 63], concept learning [28, 46], and text-to-video generation [22, 55], among others. Recent advancements [29] also highlight the cross-modality features of such generative models, showcasing their ability to serve as zero-shot classifiers.", + "bbox": [ + 496, + 232, + 890, + 368 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Synthetic data for image classification. There are two perspectives on the utilization of synthetic data for image classification: knowledge distillation [2] and data augmentation [5, 51, 54, 62]. From the knowledge distillation perspective, SyntheticData [19] reports significant performance gains in both zero-shot and few-shot settings by leveraging off-the-shelf T2I models to obtain synthetic data. The work of [2] has indicated that fine-tuning the T2I model on ImageNet [47] yields improved classification accuracy by narrowing the domain gap. Some works also find that learning from the synthetic data presents strong transferability [19, 57] and robustness [4, 30, 67]. From the data augmentation standpoint, Da-fusion [58] achieves stable performance improvements on few-shot datasets by augmenting from reference images. In a related study [3], the use of StyleGAN [25] for generating interpolated images between two different domains has been shown to enhance classifier robustness for out-of-distribution data. Our work shares similarities with AMR [5], which generates realistic novel examples by interpolating between two images using GAN [16]. The distinction lies in our discussion of interpolation using the T2I diffusion model, where its noise-adding and denoising characteristics enable a smoother implementation of interpolation.", + "bbox": [ + 496, + 378, + 890, + 739 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Non-generative data augmentation. Mixup [68] and CutMix [66] stand out as two prominent non-generative data augmentation methods, serving as effective regularization techniques during training. While Mixup achieves augmented samples through a convex combination of two images, CutMix achieves augmentation by cutting and pasting parts of images. However, both methods are constrained in their ability to produce realistic images. In addressing this limitation, the utilization of generative models emerges as a potential solution to alleviate this issue.", + "bbox": [ + 496, + 750, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "17224", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 76, + 89, + 168, + 104 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Preliminary", + "text_level": 1, + "bbox": [ + 76, + 114, + 207, + 130 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Text-to-Image diffusion model. Diffusion models generate images by gradually removing noise from a Gaussian noise source [21]. In a diffusion process with a total of $T$ steps, its forward process, which gradually adds noise, is represented as a Markov chain with a Gaussian transition kernel, where $q(\\pmb{x}_t|\\pmb{x}_{t - 1}) = \\mathcal{N}(\\pmb{x}_t;\\sqrt{\\alpha_t}\\pmb{x}_{t - 1},(1 - \\alpha_t)\\mathbf{I})$ , where $x_{t}$ represents the noisy image at step $t$ . The training objective at step $t$ is to predict the noise to reconstruct $\\pmb{x}_{t - 1}$ . When training a text-conditioned diffusion model, the simplified training objective can be summarized as follows:", + "bbox": [ + 75, + 137, + 472, + 303 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbb {E} _ {\\epsilon , \\mathbf {X}, \\mathbf {c}, t} \\left[ \\| \\epsilon - \\epsilon_ {\\theta} (\\mathbf {x} _ {t}, \\mathbf {c}, t) \\| _ {2} ^ {2} \\right], \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 169, + 306, + 468, + 332 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\epsilon_{\\theta}$ represents the predicted noise, and $c$ is the encoded text caption associated with the image $\\pmb{x}$ .", + "bbox": [ + 75, + 337, + 468, + 368 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "T2I personalization. T2I personalization aims to personalize a diffusion model for generating specific concepts using a limited number of concept-oriented images [28, 40, 42]. These concepts are typically represented using identifiers (e.g., “[V]”). As a result, we formalize the constructed image-caption set as $\\pmb{x}$ , “photo of a [V]”. Various personalization methods differ in their fine-tuning strategies. For instance, Textual Inversion (TI) [15] makes the identifier learnable, but other modules are not fine-tuned, potentially sacrificing some faithfulness in image generation. On the other hand, Dreambooth (DB) [46] fine-tunes the U-Net [45] for more refined personalized generation but faces the challenge of increased computational cost.", + "bbox": [ + 75, + 373, + 472, + 570 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Image-to-image translation. Image translation enables image synthesis and editing using a reference image as guidance [24, 64, 71]. Diffusion-based image translation methods can generate fine edits, which refer to subtle modifications, with varying degrees of shift relative to the reference image [8, 35, 59]. Here, we draw inspiration from SDEdit [35] to perform edits on the reference image, where the target image $\\mathbf{x}^{\\mathrm{tar}}$ is translated from a reference image $\\mathbf{x}^{\\mathrm{ref}}$ . During translation, the reverse process does not traverse the full process but starts from a certain step $\\lfloor sT \\rfloor$ , where $s \\in [0,1]$ controls the insertion position of the reference image with noise, as follows,", + "bbox": [ + 75, + 575, + 472, + 758 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {x} _ {\\lfloor s T \\rfloor} = \\sqrt {\\tilde {\\alpha} _ {\\lfloor s T \\rfloor}} \\boldsymbol {x} _ {0} ^ {\\text {r e f}} + \\sqrt {1 - \\tilde {\\alpha} _ {\\lfloor s T \\rfloor}} \\epsilon . \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 140, + 763, + 468, + 789 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "By adjusting the strength parameter $s$ , one can strike a balance between the diversity of the generated images and their faithfulness to the reference image.", + "bbox": [ + 75, + 794, + 472, + 840 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. General Framework", + "text_level": 1, + "bbox": [ + 76, + 847, + 272, + 862 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The Diff-Mix pipeline consists of two key steps. Firstly, to produce more faithful images for domain-specific datasets,", + "bbox": [ + 75, + 869, + 472, + 902 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/679e9849bd49aaf0cc0d70513702151847cb42d3e435a39c2e741616576fd03e.jpg", + "image_caption": [ + "(a) Real images" + ], + "image_footnote": [], + "bbox": [ + 500, + 87, + 893, + 128 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/a835499778bd5de0d3f981d224126e4e499c62f5672a18da7d0a8988636b366b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 501, + 142, + 893, + 183 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/bdb3d0dfa6ae058b391cc0d7a8e02a99ee4653596d741a468d8049886392d492.jpg", + "image_caption": [ + "(b) Synthetic images (vanilla SD)" + ], + "image_footnote": [], + "bbox": [ + 501, + 196, + 892, + 236 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/77de4e4343c7bac68f53f7e49ade331e8f701868343060f55a09a66404d3c9a9.jpg", + "image_caption": [ + "(c) Synthetic images (SD fine-tuned via DB)", + "(d) Synthetic images (SD fine-tuned via $\\mathrm{TI + DB}$ )", + "Figure 2. Examples of \"American Three toed Woodpecker\". (a) Real images from the training set. (b-d) synthetic images generated using different fine-tuned models with the same number of fine-tuning steps. TI+DB indicates both text embedding and U-Net are fine-tuned. TI+DB achieves a more faithful output compared to DB alone (check the head and wing patterns of the birds)." + ], + "image_footnote": [], + "bbox": [ + 501, + 251, + 890, + 291 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "we propose treating it as a T2I personalization problem and fine-tuning the Stable Diffusion (SD). Subsequently, to enhance the diversity of synthetic data beyond the well-fitted training distribution, we employ inter-class image translation. This process produces interpolated images with increased background diversity for each class.", + "bbox": [ + 496, + 435, + 893, + 527 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.3. Fine-tune Diffusion Model", + "text_level": 1, + "bbox": [ + 500, + 537, + 741, + 551 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Vanilla distillation tends to be less effective, especially as the number of training shots increases (refer to Sec. 4.1). In order to mitigate the distribution gap, we propose finetuning Stable Diffusion in conjunction with current widely-used T2I personalization strategies.", + "bbox": [ + 496, + 561, + 893, + 637 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Dreambooth meets Textual Inversion. Many fine-grained datasets provide terminological names for their categories, like \"American Three toed Woodpecker\" and \"Pileated Woodpecker\". We could construct image-text pairs using category-related prompts and fine-tune the denoising network of SD using Eq. 1, which is analogous to Dreambooth. However, we observe that directly incorporating these specialized terms into the text during fine-tuning can impede convergence and hinder the generation of faithful images. We attribute this challenge to the semantic proximity of terminology within a fine-grained domain, where fine-tuning the vision module alone tends to be less effective at distinguishing two similar classes within the same family, like \"Woodpecker\". Inspired by Textual Inversion [15], we opt to replace the terminological name in the dataset with \"[Vi] [metaclass]\" where \"[Vi\"] is a learnable identifier, and $i$ varies from 1 to $N$ , represent-", + "bbox": [ + 496, + 643, + 893, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "17225", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/2183af4399b8471a3a3a3b7189d897524e8e43fdc30709817cab54d8704ae83d.jpg", + "image_caption": [ + "Figure 3. Fine-tuning framework of Diff-Mix operates as follows: Initially, we replace the class name with a structured identifier formatted as “[ $\\mathsf{V}^i$ ” [metaclass]}, thereby sidestepping the need for specific terminological expressions. Next, we engage in joint fine-tuning of these identifiers and the low-rank residues (LoRA) of U-Net to capture the domain-specific distribution." + ], + "image_footnote": [], + "bbox": [ + 83, + 88, + 468, + 354 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "ing the category index. The illustration of our fine-tuning strategy is presented in Fig. 3. The term “[metaclass]” is determined by the theme of the current dataset, such as “bird” for a fine-grained bird dataset. By concurrently fine-tuning the identifier and the U-Net, we empower the model to quickly adapting to the fine-grained domain, allowing it to generate faithful images using the identifier (see comparison between Row 3 and Row 4 in Fig. 2).", + "bbox": [ + 75, + 465, + 468, + 589 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Parameter efficient fine-tuning. In this context, we embrace the parameter-efficient fine-tuning strategy known as LoRA [23]. LoRA distinguishes itself by fine-tuning the residue of low-rank matrices instead of directly fine-tuning the pre-trained weight matrices. To elaborate, consider a weight matrix $\\mathbf{W} \\in \\mathbb{R}^{m \\times n}$ . The tunable residual matrix $\\Delta \\mathbf{W}$ comprises two low-rank matrices: $\\mathbf{A} \\in \\mathbb{R}^{m \\times d}$ and $\\mathbf{B} \\in \\mathbb{R}^{n \\times d}$ , defined as $\\Delta \\mathbf{W} = \\mathbf{A}\\mathbf{B}^{\\top}$ . As a default configuration, we set the rank $d$ to 10.", + "bbox": [ + 75, + 595, + 468, + 731 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.4. Data Synthesis Using Diffusion Model", + "text_level": 1, + "bbox": [ + 76, + 742, + 403, + 758 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In generating pseudo data, three strategies can be used with our fine-tuned diffusion model $^{2}$ : (1) distillation-based method Diff-Gen, (2) intra-class augmentation Diff-Aug, and (3) inter-class augmentation Diff-Mix.", + "bbox": [ + 75, + 766, + 468, + 827 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Diff-Gen and Diff-Aug. For a target class $y^i$ and its textual condition, \"photo of a $[V^i]$ [metaclass]\", both", + "bbox": [ + 76, + 833, + 468, + 864 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/cbdd6c90118c3b11ceb3757f7910278802c1de706b8a7200f87846bfc29ff2dc.jpg", + "image_caption": [ + "Figure 4. Examples of images translated using Diff-Mix and Diff-Aug across various strengths. Diff-Aug employs the same target and reference image classes, typically resulting in subtle modifications. Diff-Mix progressively adjusts the foreground to align with the target class as the translation strength increases, while preserving the background layout from the reference image." + ], + "image_footnote": [], + "bbox": [ + 501, + 88, + 888, + 354 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "methods generate synthetic samples annotated with class $i$ . Specifically, Diff-Gen generates samples from scratch by initializing with random Gaussian noise and proceeding through the full reverse diffusion process with $T$ steps. Diff-Gen can produce images aligned with its fine-tuned distribution. In contrast, Diff-Aug sacrifices a portion of diversity and generate images by editing on a reference image. Specifically, it randomly sample a image from the intra-class training set and enhances the image through image translation using Eq. 2. The term \"intra-class\" means that the conditioning prompts are constructed based on the ground truth categories of images, and such a denoising process tends to introduce less variation, particularly for the foreground concepts (see top rows of Fig. 1).", + "bbox": [ + 496, + 468, + 892, + 680 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Diff-Mix. Diff-Mix employs the same translation process as Diff-Aug, but the reference image is sampled from the full training set rather than intra-class set to enable interclass interpolation. The key difference is that Diff-Mix can generate numerous counterfactual examples, such as a blackbird in the sea (see fourth row of Fig. 1). This necessitates that downstream models make a more refined differentiation of category attributes, thereby reducing the impact of spurious correlations introduced by variations in the background. Denoting the label of the reference image as $y^{j}$ , by inserting the reference image into the reverse process with the prompt \"photo of a $[V^i]$ [metaclass]\", we can obtain interpolated images between the $i_{\\mathrm{th}}$ and $j_{\\mathrm{th}}$ categories. By controlling the intensity $s$ , we can precisely", + "bbox": [ + 496, + 688, + 892, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_footnote", + "text": "2We use the prefix \"Diff-\" denotes the T2I model is fine-tuned and \"Real-\" denotes the vanilla T2I model.", + "bbox": [ + 76, + 875, + 468, + 898 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "17226", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/568a687fbd510c9c8c70ae5df276f80603d88666330070d2558eecaa5f929a56.jpg", + "image_caption": [ + "Figure 5. A schematic explanation of Diff-Mix's effectiveness using structural casual model [41]. $\\pmb{x}_{fg}$ is the foreground that determines the real class label, $\\pmb{x}_{bg}$ denotes the background. $\\pmb{x}_{fg} \\rightarrow \\pmb{z} \\rightarrow y$ is the causal path that we are focusing and $\\pmb{x}_{fg} \\gets I \\rightarrow \\pmb{x}_{bg} \\rightarrow \\pmb{z} \\rightarrow y$ is the backdoor path that introduces spurious relations between $\\pmb{x}_{fg}$ and $y$ ." + ], + "image_footnote": [], + "bbox": [ + 86, + 99, + 254, + 186 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/d4cd3fa5cd7d74df00c0a00006795334db3edf919efef18c0fd57749cd216e2a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 281, + 97, + 449, + 183 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "manage the interpolation process. When annotating the synthetic image, unlike Mixup and Cutmix, we take into account the non-linear nature of diffusion translation. Thus the annotation function is given by", + "bbox": [ + 76, + 299, + 468, + 359 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {y} = (1 - s ^ {\\gamma}) y ^ {i} + s ^ {\\gamma} y ^ {j}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 191, + 369, + 468, + 386 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\gamma$ is a hyperparameter introducing non-linearity. Our empirical findings indicate that a $\\gamma$ smaller than 1 is favored. Additionally, in low-shot cases, the samples with higher confidence in the target class are preferred (see details in Sec. 4.5).", + "bbox": [ + 76, + 396, + 468, + 472 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Construct synthetic dataset. To construct the synthetic dataset using Diff-Mix, similar to Da-fusion [58], we adopt a randomized sampling strategy ( $s \\in \\{0.5, 0.7, 0.9\\}$ ) for the selection of translation strength. While applying the inter-class editing, we observe that Diff-Mix tends to produce more undesirable samples compared to Diff-Aug. These undesirable samples have incomplete foreground such as fragmented bird bodies. This is caused by the intrinsic shape and pose differences among classes. To mitigate this, we introduce a simple data-cleaning approach to reduce the proportion of such problematic images. We utilize the large vision language model CLIP [43] to assess the confidence in the content, serving as the filtering criterion. Further details can be found in the supplementary materials (SMs).", + "bbox": [ + 76, + 478, + 468, + 690 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Analysis. We depict the core insight of Diff-Mix in Fig. 5. To eliminate the spurious correlation introduced by $\\pmb{x}_{bg}$ , learning on the synthetic set with randomized $\\pmb{x}_{bg}$ (background) can cut off spurious correlation, forcing the classification model to infer only from the foreground. The study in Fig. 7 (b) shows that the more diverse the background (larger the referable class number), the better the performance on the CUB test set.", + "bbox": [ + 76, + 695, + 468, + 816 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 76, + 830, + 207, + 847 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this section, we investigate the effectiveness of Diff-Mix in domain-specific datasets. The key questions we aim to address are as follows:", + "bbox": [ + 76, + 854, + 468, + 898 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "Q1: Can generative inter-class augmentation lead to more significant performance gains in downstream tasks compared to those intra-class augmentation methods and distillation-based methods?", + "Q2: Is improved background diversity the secret weapon of Diff-Mix for enhancing the performance?", + "Q3: How to choose the fine-tuning strategy and annotation strategy to boost the performance gains for the inter-class augmentation?" + ], + "bbox": [ + 511, + 90, + 890, + 226 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To address Q1, we separately discuss these questions in few-shot settings in Sec. 4.1, conventional classification in Sec. 4.2, as well as long-tail classification in Sec. 4.3. Additionally, to answer Q2, we conduct a test for background robustness in Sec. 4.4 and perform an ablation study focusing on the size and diversity of synthetic data in Sec. 4.5. For Q3, we conduct an ablation study to empirically discover effective strategies for deployment in Sec. 4.5.", + "bbox": [ + 496, + 229, + 890, + 351 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Few-shot Classification", + "text_level": 1, + "bbox": [ + 500, + 363, + 712, + 378 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Experimental Setting. To investigate the impact of different data expansion methods, we conduct few-shot experiments on a domain-specific dataset Caltech-UCSD Birds (CUB) [61], with shot numbers of 1, 5, 10, and all. For augmentation-based methods, the synthetic dataset is constructed using various translation strengths $(s)$ , specifically, $s \\in \\{0.1, 0.3, \\dots, 1.0\\}$ . We expand the original training set with a multiplier of 5 and cache the synthesized dataset locally for joint training. Real data are replaced by synthetic data proportionally during training, and the replacement probability $p$ is set as 0.1, 0.2, 0.3, and 0.5 for all-shot, 10-shot, 5-shot, and 1-shot classification, respectively. All experiments use ResNet50 with an input resolution of $224 \\times 224$ . Additional details can be found in the SMs.", + "bbox": [ + 496, + 387, + 890, + 598 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Comparison methods. To unveil the trade-off between faithfulness and diversity resulting from different expansion strategies, we compared Diff-Mix with Diff-Gen and Diff-Aug. Furthermore, we conduct experiments on expansion strategies using vanilla SD: Real-Mix, Real-Gen, and Real-Aug, where 'Real' signifies that SD is not fine-tuned.", + "bbox": [ + 496, + 604, + 890, + 696 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Main results. To answer Q1 under the few-shot classification setting, we augment CUB using X-Mix, X-Aug, and X-Gen, where 'X' denotes 'Diff/Real' for simplicity. The results are shown in Fig. 6, and we can observe that:", + "bbox": [ + 496, + 703, + 890, + 762 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. Diff-Mix generally outperforms the intra-class competitor X-Aug and distillation competitor X-Gen in various few-shot scenarios. It tends to achieve higher gains when the strength $s$ is relatively large, i.e., $\\{0.5, 0.7, 0.9\\}$ , where the foreground has been edited to match the target class and the background retains similarities to the reference image.", + "2. Among the Real-X methods, distillation tends to be more effective than the augmentation method when the shot" + ], + "bbox": [ + 500, + 763, + 890, + 898 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "17227", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/a5726da6ec202f288638f381edd25aa2f9b847fb756e9a7865d7d32a5d6b4b6e.jpg", + "image_caption": [ + "(a) 1-shot" + ], + "image_footnote": [], + "bbox": [ + 81, + 90, + 269, + 200 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/6086dd31f4fd8466d946c30de710cb0dc3a65677e02b903931d3625a5b80c271.jpg", + "image_caption": [ + "(b) 5-shot" + ], + "image_footnote": [], + "bbox": [ + 274, + 89, + 464, + 200 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/610cb3645a14a320ac5979460207d35c8a49303b34cf50e08381810c67d9b1f3.jpg", + "image_caption": [ + "(c) 10-shot" + ], + "image_footnote": [], + "bbox": [ + 470, + 90, + 658, + 200 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/f79c46a60ad004f140e8205d114bd2af63b11e8700394e0fc5355e5294c44799.jpg", + "image_caption": [ + "(d) All-shot", + "Figure 6. Few-shot classification results on CUB." + ], + "image_footnote": [], + "bbox": [ + 663, + 90, + 857, + 200 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "number is low, but the trend reverses as the shot number increases (compare Real-Gen with Real-Aug). Real-Gen's samples even show less effective than Real-Aug $(s = 1.0)$ under the all-shot case3. This indicates that the importance of faithfulness in the trade-off between faithfulness and diversity increases with the shot number. Additionally, Real-Mix exhibits consistent and stable improvement over the other two methods.", + "bbox": [ + 94, + 261, + 467, + 381 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3. Diff-Gen consistently outperforms Real-Gen under four scenarios. Notably, Real-Gen's performance declines below that of the baseline as the shot numbers reach 10, showcasing the importance of the fine-tuning process which increases the faithfulness of synthetic samples.", + "bbox": [ + 76, + 382, + 468, + 458 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Conventional Classification", + "text_level": 1, + "bbox": [ + 76, + 470, + 321, + 486 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Experimental setting. To test whether Diff-Mix can further boost performance in a more challenging setting, i.e., under the all-shot scenario with high input resolution, we conduct conventional classification on five domain-specific datasets: CUB [61], Stanford Cars [27], Oxford Flowers [38], Stanford Dogs [26], and FGVC Aircraft [34]. Two backbones are employed: pretrained (ImageNet1K [49]) ResNet50 [18] with input resolution $448 \\times 448$ , and pretrained (ImageNet21K) ViT-B/16 [13] with input resolution $384 \\times 384$ . Label smoothing [36] is applied across all datasets with a confidence level of 0.9. For all expansion strategies, the expansion multiplier is 5 and the replacement probability $p$ is 0.1. Besides, we use a randomized sampling strategy $(s \\in \\{0.5, 0.7, 0.9\\})$ and a fixed $\\gamma$ (0.5, and this is specific to Diff-Mix) for all datasets.", + "bbox": [ + 76, + 496, + 468, + 722 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Comparison methods. We compare Diff-Mix with (1) Real-Filtering (RF) [19], a variation of Real-Gen that incorporates clip filtering, (2) Real-Guidance (RG) [19], which augments the dataset using intra-class image translation at low strength ( $s = 0.1$ ), (3) Da-Fusion [58], a method that solely fine-tunes the identifier to personalize each class and employs randomized sampling strategy", + "bbox": [ + 76, + 729, + 467, + 835 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "$(s \\in \\{0.25, 0.5, 0.75, 1.0\\})$ , and non-generative augmentation methods (4) CutMix [66] and (5) Mixup [68].", + "bbox": [ + 500, + 261, + 888, + 292 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Main results. We show the classification accuracy for different data expansion strategies in Table 1, our observations can be summarized: (1) Diff-Mix consistently demonstrates stable improvements across the majority of settings. Its average performance gain across the five datasets exceeds that of baselines employing intra-class augmentation methods (RG and Da-fusion), distillation method (RF), and non-generative data augmentation techniques (CutMix and Mixup). (2) Real-filtering, analogous to the discussion of Real-Gen, exhibits performance degradation on most datasets due to the distribution gap. (3) The combined use of Diff-Mix and CutMix often yields better performance gains. This is attributed to the distinct enhancement mechanisms of the two methods, i.e., vicinal risk minimization [11, 68] and foreground-background disentanglement. (4) Diff-Mix does not exhibit significant performance improvement in the dog dataset. We attribute this lack of improvement to the complexity of the dog dataset, which often contains multiple subjects in a single image, impeding effective foreground editing (refer to the SMs for visual examples).", + "bbox": [ + 496, + 297, + 890, + 601 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Long-Tail Classification", + "text_level": 1, + "bbox": [ + 500, + 613, + 718, + 628 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Experiment setting. Following the settings of previous long-tail dataset constructions [9, 32, 39], we create two domain-specific long-tail datasets, CUB-LT [50] and Flower-LT. The imbalance factor controls the exponential distribution of the imbalanced data, where a larger value indicates a more imbalanced distribution. To leverage generative models for long-tail classification, we adopt the approach of SYNAuG [65], which uniformize the imbalanced real data distribution using synthetic data. Translation strength $s$ (0.7) and $\\gamma$ (0.5) are fixed for both Diff-Mix and Real-Mix.", + "bbox": [ + 496, + 643, + 890, + 809 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Comparison methods. We compare Diff-Mix with Real-Mix, Real-Gen, the non-generative CutMix-based oversampling approach CMO [39], and its enhanced variant with two-stage deferred re-weighting [9] (CMO+DRW).", + "bbox": [ + 496, + 816, + 890, + 878 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Main results. We present the long-tail classification re", + "bbox": [ + 500, + 885, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_footnote", + "text": "$^3$ Real-Aug $(s = 1.0)$ remains analogous to the reference image (slightly higher faithfulness compared to Real-Gen) because the discrete forward process cannot approximate the ideal normal distribution within a limited number of steps $(T = 25)$ .", + "bbox": [ + 75, + 849, + 468, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "17228", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/a2c4e95fa07ac834c852e3707212523a01d6820e48d0df44751769a06cb2cb90.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
BackboneAug. MethodFT StrategyDataset
CUBAircraftFlowerCarDogAvg
ResNet50@448--86.6489.0999.2794.5487.4891.40
Cutmix[66]-87.2389.4499.2594.7387.5991.65+0.25
Mixup[68]-86.6889.4199.4094.4987.4291.48+0.08
Real-filtering [19]†X85.6088.5499.0994.5987.3091.22-0.18
Real-guidance [19]†X86.7189.0799.2594.5587.4091.59+0.19
Da-fusion [58]†TI86.3087.6499.3794.6987.3391.07-0.58
Diff-MixTI + DB87.1690.2599.5495.1287.7491.96+0.56
Diff-Mix + CutmixTI + DB87.5690.0199.4795.2187.8992.03+0.63
ViT-B/16@384--89.3783.5099.5694.2192.0691.74
Cutmix[66]-90.5283.5099.6494.8392.1392.12+0.38
Mixup[68]-90.3284.3199.7394.9892.0292.27+0.53
Real-filtering [19]†X89.4983.0799.3694.6691.9191.69-0.05
Real-guidance [19]†X89.5483.1799.5994.6592.0591.80+0.06
Da-fusion [58]†TI89.4081.8899.6194.5392.0791.50-0.24
Diff-MixTI + DB90.0584.3399.6495.0991.9992.22+0.48
Diff-Mix + CutmixTI + DB90.3585.1299.6895.2691.8992.46+0.72
", + "bbox": [ + 130, + 88, + 841, + 330 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/5ce0024502668b0e364b688abc84b46930516a74730afd3af8c8377575aa3727.jpg", + "table_caption": [ + "Table 1. Conventional classification in six fine-grained datasets. ${}^{ \\dagger }$ ,indicates our reproduced results using SD." + ], + "table_footnote": [], + "table_body": "
MethodIF=1005010
ManyMediumFewAll
CE79.1164.2813.4833.6544.8258.13
CMO [39]78.3258.5714.7832.9444.0857.62
CMO + DRW [10]78.9756.3614.6632.5746.4359.25
Real-Gen84.8865.2330.6845.8653.4361.42
Real-Mix (s=0.7)84.6366.3434.4447.7555.6762.27
Diff-Mix (s=0.7)84.0767.7936.5550.3558.1964.48
", + "bbox": [ + 81, + 369, + 464, + 467 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/f19d20d7649b4d441f67f592f722c620ea0fcd7cad0152363ef719100d6386ec.jpg", + "table_caption": [ + "Table 2. Long-tail classification in CUB-LT." + ], + "table_footnote": [], + "table_body": "
MethodIF=1005010
ManyMediumFewAll
CE99.1994.9558.1880.4390.8795.07
CMO [39]99.2595.1967.4583.9591.4395.19
CMO+ DRW [10]99.9795.0667.3184.0792.0695.92
Real-Gen98.6495.5566.1083.5691.8495.22
Real-Mix (s=0.7)99.8796.2668.5385.1992.9696.04
Diff-Mix (s=0.7)99.2596.9878.4189.4693.8696.63
", + "bbox": [ + 81, + 502, + 464, + 599 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "sults for CUB-LT in Table 2 and Flower-LT in 3. The observations are as follows: (1) Generative approaches exhibit superior performance in tackling imbalanced classification issues compared to CutMix-based methods (CMO and CMO+DRW). (2) Real-Mix surpasses Real-Gen in performance across various imbalance factors in two datasets. This indicates that tail classes can benefit from the enhanced diversity by leveraging the visual context of majority classes. (3) Diff-Mix generally achieves the best performance among the compared strategies, especially at the low-shot case, highlighting the importance of fine-tuning.", + "bbox": [ + 75, + 635, + 468, + 801 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Background Robustness", + "text_level": 1, + "bbox": [ + 76, + 815, + 299, + 830 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In this section, we aim to evaluate Diff-Mix's robustness to background shifts, specifically, whether synthesizing more diverse samples can improve the classification model's generalizability when the background is altered. To achieve", + "bbox": [ + 75, + 839, + 470, + 902 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/b98b0674ca8ecabb99c669f82edc74115567f8fdfb559649124584259afc29da.jpg", + "table_caption": [ + "Table 3. Long-tail classification in Flower-LT." + ], + "table_footnote": [], + "table_body": "
GroupBase.CutMixDA-fusionDiff-AugDiff-Mix
(waterbird, water)59.5062.4660.9061.8363.83
(waterbird, land)56.7060.1258.1060.1263.24
(landbird, land)73.4873.3972.9473.0475.64
(landbird, water)73.9774.7272.7773.5274.36
Avg.70.1971.2369.9070.2872.47
", + "bbox": [ + 509, + 371, + 877, + 445 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 4. Classification results across four groups in Waterbird [48]. Waterbird is an out-of-distribution dataset for CUB, crafted by segmenting CUB's foregrounds and paste them into the scene images from Places [70]. The constructed dataset can be divided into four groups based on the composition of foregrounds (waterbird and landbird) and backgrounds (water and land).", + "bbox": [ + 498, + 458, + 890, + 542 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "this, we utilize an out-of-distribution test set for CUB, namely Waterbird [48]. We then perform inference on the whole Waterbird set using classifiers that have been trained on either the original CUB dataset or its expanded variations. In Table 4, we present the classification accuracies across the four groups and compare Diff-Mix with other intra-class methods (Da-fusion and Diff-Aug) as well as CutMix. We observe that Diff-Mix generally outperforms its counterparts and achieves a significant performance improvement $(+6.5\\%)$ in the challenging counterfactual group (waterbirds with land backgrounds). It is important to highlight that the background scenes in the Waterbird dataset are novel to CUB, requiring the classification model to have a stronger perceptual capability for the images' foregrounds.", + "bbox": [ + 496, + 563, + 890, + 776 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.5. Discussion", + "text_level": 1, + "bbox": [ + 500, + 786, + 617, + 800 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In this section, we address Q2 by examining the impact of size and diversity on synthetic data. Furthermore, we perform an ablation study to assess the effects of fine-tuning strategies and training hyperparameters of Diff-Mix, which is aimed at answering Q3. Unless specified otherwise, our discussions are based on experiments conducted using CUB", + "bbox": [ + 496, + 809, + 893, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "17229", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/7e18fcca6a05d2485e7b5d5a97434bd0a227da1949fc4d50e9498d6edd65f57c.jpg", + "image_caption": [ + "(a)" + ], + "image_footnote": [], + "bbox": [ + 81, + 88, + 263, + 194 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/ad491210bb609c2744334dabc3e1cdbb16868909a409daa40df26f34ebb4abef.jpg", + "image_caption": [ + "(b)", + "Figure 7. Comparison of results across various (a) synthetic data sizes and (b) numbers of referable classes for each target class." + ], + "image_footnote": [], + "bbox": [ + 282, + 88, + 467, + 194 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "with ResNet50, where inputs are resized to $224 \\times 224$ .", + "bbox": [ + 76, + 273, + 439, + 287 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Size and diversity of synthetic Data. The relationship between performance gain and the size of synthetic data is depicted in Fig. 7 (a), where a classification model was trained with synthetic data of varying sizes. We observe a monotonically increasing trend as the multiplier for the synthetic dataset ranges from 2 to 10. Ideally, the combination choices of $(x_{i},y_{j})$ are in the order of $N|D^{train}|$ ( $N = 200$ for CUB). Furthermore, we limit the number of referable classes for each target class, which means the number of referable backgrounds decreases, resulting in a synthetic dataset of relatively lower diversity. The results are shown in Fig. 7 (b), and we observe a consistent improvement in performance as the number of referable classes increases. These results consistently underscore the critical role of background diversity introduced by Diff-Mix.", + "bbox": [ + 75, + 296, + 468, + 523 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Impact of fine-tuning strategy. Here we compare three different fine-tuning strategies: TI, DB, and the combined $\\mathrm{TI + DB}$ . All strategies share the same fine-tuning hyperparameters and training steps (35000). To evaluate the distribution gap, we compute the FID score [20] of synthesized images (Diff-Gen) with the training set. As illustrated in Table 5, we observe that both $\\mathrm{TI + DB}$ and TI have lower FID scores than DB. This can be attributed to the fact that semantic proximity impedes the convergence process. Additionally, while using TI alone results in a relatively low FID score, the improvements in performance are limited. This limitation stems from TI's inability to accurately reconstruct detailed concept (foreground) information, as it is primarily fine-tuned at the semantic level [28].", + "bbox": [ + 75, + 530, + 468, + 743 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Annotation function. In this section, we discuss the impact of the choices of the translation strength $s$ and the non-linear factor $\\gamma$ in Eq. 3. As shown in Table 6, we observe that as the translation strength decreases, the optimal value for $\\gamma$ also decreases, which underscores the nonlinearity of Diff-Mix. The comparison between the 5-shot and all-shot settings indicates that the model tends to prefer a more diverse synthetic dataset when the number of training shots is small ( $s = 0.9$ for 5-shot, $s = 0.7$ for all-shot). Besides, a larger confidence in the target class is pre", + "bbox": [ + 75, + 750, + 470, + 901 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/9db0d82779f31e1fe2dc3d03f45e62f7c620dbf500eb742d8b7215455c965234.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
BaselineTIDBTI + DB
5-shotFID (Diff-Gen)-18.2619.5518.43
Acc. (Diff-Mix)50.9057.6456.1159.41
All-shotFID (Diff-Gen)-14.1314.6413.99
Acc. (Diff-Mix)81.6081.8681.9982.85
", + "bbox": [ + 513, + 88, + 877, + 148 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/0f3702881ca72de16a29f76c6f0043eec4ec0c1182442166ac1bfdf7d8148c85.jpg", + "table_caption": [ + "Table 5. Comparison of distribution gap and classification accuracy across three fine-tuning strategies. TI solely fine-tunes the identifier, and DB solely fine-tunes the U-Net, and $\\mathrm{TI + DB}$" + ], + "table_footnote": [], + "table_body": "
γ5-shotAll-shot
s = 0.5s = 0.7s = 0.9s = 0.5s = 0.7s = 0.9
1.5-4.50-0.31+10.30-1.08+0.92+0.90
1.0-2.31+2.99+10.79+0.25+1.14+0.90
0.5+2.35+8.44+11.01+0.92+1.30+0.86
0.3+3.94+9.41+11.15+0.97+1.24+0.69
0.1+6.18+9.86+10.84+0.50+0.88+0.84
0.0+5.25+9.41+11.06+0.38+0.63+0.54
", + "bbox": [ + 514, + 213, + 877, + 314 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 6. Comparison of performance gain across various $\\gamma$ and translation strength $s$ . Lower $\\gamma$ indicates a higher confidence over target class, e.g. ( $\\gamma = 0.1$ , $s = 0.7$ ) results in $0.04y_{i} + 0.96y_{j}$ and ( $\\gamma = 0.5$ , $s = 0.7$ ) results in $0.16y_{i} + 0.84y_{j}$ .", + "bbox": [ + 498, + 323, + 890, + 380 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "ferred when the shot number is small ( $\\gamma = 0.1$ for 5-shot, $\\gamma = 0.5$ for all-shot). A possible explanation is that the all-shot setting is less tolerant towards unrealistic images, as discussed in Section 6. Empirically, we recommend choosing a higher translation strength ( $s \\in [0.5, 0.7, 0.9]$ ) and a smaller $\\gamma$ ( $\\gamma \\in [0.1, 0.3, 0.5]$ ) as a conservative option.", + "bbox": [ + 498, + 406, + 890, + 498 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 510, + 617, + 526 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this work, we investigate two pivotal aspects, faithfulness and diversity, that are critical for the current state-of-the-art text-to-image generative models to enhance image classification tasks. To achieve a more effective balance between these two aspects, we propose an inter-class augmentation strategy that leverages Stable Diffusion. This method enables generative models to produce a greater diversity of samples by editing images from other classes and shows consistent performance improvement across various classification tasks.", + "bbox": [ + 496, + 536, + 890, + 686 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Acknowledgement", + "text_level": 1, + "bbox": [ + 500, + 700, + 679, + 718 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This research is mainly supported by the National Natural Science Foundation of China (92270114). This work is also partially supported by the National Key R&D Program of China under Grant 2021ZD0112801.", + "bbox": [ + 498, + 726, + 893, + 786 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "17230", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 174, + 106 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Antreas Antoniou, Amos Storkey, and Harrison Edwards. Data augmentation generative adversarial networks. arXiv preprint arXiv:1711.04340, 2017. 2", + "[2] Shekoofeh Azizi, Simon Kornblith, Chitwan Sahara, Mohammad Norouzi, and David J Fleet. Synthetic data from diffusion models improves imagenet classification. arXiv preprint arXiv:2304.08466, 2023.1, 2", + "[3] Haoyue Bai, Ceyuan Yang, Yinghao Xu, S-H Gary Chan, and Bolei Zhou. Improving out-of-distribution robustness of classifiers via generative interpolation. arXiv preprint arXiv:2307.12219, 2023. 2", + "[4] Hritik Bansal and Aditya Grover. Leaving reality to imagination: Robust classification via generated datasets. arXiv preprint arXiv:2302.02503, 2023. 1, 2", + "[5] Christopher Beckham, Sina Honari, Vikas Verma, Alex M Lamb, Farnoosh Ghadiri, R Devon Hjelm, Yoshua Bengio, and Chris Pal. On adversarial mixup resynthesis. Advances in neural information processing systems, 32, 2019. 2", + "[6] Ankan Kumar Bhunia, Salman Khan, Hisham Cholakkal, Rao Muhammad Anwer, Jorma Laaksonen, Mubarak Shah, and Fahad Shahbaz Khan. Person image synthesis via denoising diffusion model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5968-5976, 2023. 2", + "[7] Andrew Brock, Jeff Donahue, and Karen Simonyan. Large scale gan training for high fidelity natural image synthesis. arXiv preprint arXiv:1809.11096, 2018. 1", + "[8] Tim Brooks, Aleksander Holynski, and Alexei A Efros. Instructpix2pix: Learning to follow image editing instructions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18392-18402, 2023. 3", + "[9] Kaidi Cao, Colin Wei, Adrien Gaidon, Nikos Arechiga, and Tengyu Ma. Learning imbalanced datasets with label-distribution-aware margin loss. Advances in neural information processing systems, 32, 2019. 6", + "[10] Kaidi Cao, Colin Wei, Adrien Gaidon, Nikos Arechiga, and Tengyu Ma. Learning imbalanced datasets with label-distribution-aware margin loss. Advances in neural information processing systems, 32, 2019. 7", + "[11] Olivier Chapelle, Jason Weston, Léon Bottou, and Vladimir Vapnik. Vicinal risk minimization. Advances in neural information processing systems, 13, 2000. 6", + "[12] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in neural information processing systems, 34:8780-8794, 2021. 1", + "[13] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 6", + "[14] M. Everingham, S. M. A. Eslami, L. Van Gool, C. K. I. Williams, J. Winn, and A. Zisserman. The pascal visual object classes challenge: A retrospective. International Journal of Computer Vision, 111(1):98-136, 2015. 1, 2, 5" + ], + "bbox": [ + 76, + 114, + 470, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[15] Rinon Gal, Yuval Alaluf, Yuval Atzmon, Or Patashnik, Amit H Bermano, Gal Chechik, and Daniel Cohen-Or. An image is worth one word: Personalizing text-to-image generation using textual inversion. arXiv preprint arXiv:2208.01618, 2022. 2, 3", + "[16] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. Advances in neural information processing systems, 27, 2014. 2", + "[17] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. Advances in neural information processing systems, 27, 2014. 1", + "[18] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 6", + "[19] Ruifei He, Shuyang Sun, Xin Yu, Chuhui Xue, Wenqing Zhang, Philip Torr, Song Bai, and Xiaojuan Qi. Is synthetic data from generative models ready for image recognition? arXiv preprint arXiv:2210.07574, 2022. 2, 6, 7, 3", + "[20] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 8", + "[21] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 3", + "[22] Jonathan Ho, William Chan, Chitwan Sahara, Jay Whang, Ruiqi Gao, Alexey Gritsenko, Diederik P Kingma, Ben Poole, Mohammad Norouzi, David J Fleet, et al. Imagen video: High definition video generation with diffusion models. arXiv preprint arXiv:2210.02303, 2022. 2", + "[23] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 4", + "[24] Phillip Isola, Jun-Yan Zhu, Tinghui Zhou, and Alexei A Efros. Image-to-image translation with conditional adversarial networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1125-1134, 2017. 3", + "[25] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4401-4410, 2019. 1, 2", + "[26] Aditya Khosla, Nityananda Jayadevaprakash, Bangpeng Yao, and Li Fei-Fei. Novel dataset for fine-grained image categorization. In First Workshop on Fine-Grained Visual Categorization, IEEE Conference on Computer Vision and Pattern Recognition, Colorado Springs, CO, 2011. 6, 2", + "[27] Jonathan Krause, Michael Stark, Jia Deng, and Li Fei-Fei. 3d object representations for fine-grained categorization. In Proceedings of the IEEE international conference on computer vision workshops, pages 554–561, 2013. 6, 2", + "[28] Nupur Kumari, Bingliang Zhang, Richard Zhang, Eli Shechtman, and Jun-Yan Zhu. Multi-concept customization" + ], + "bbox": [ + 501, + 92, + 893, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "17231", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "of text-to-image diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1931-1941, 2023. 2, 3, 8", + "[29] Alexander C Li, Mihir Prabhudesai, Shivam Duggal, Ellis Brown, and Deepak Pathak. Your diffusion model is secretly a zero-shot classifier. arXiv preprint arXiv:2303.16203, 2023. 2", + "[30] Yicong Li, Xiang Wang, Junbin Xiao, Wei Ji, and Tat seng Chua. Invariant grounding for video question answering. In CVPR, 2022. 2", + "[31] Shanchuan Lin, Bingchen Liu, Jiashi Li, and Xiao Yang. Common diffusion noise schedules and sample steps are flawed. arXiv preprint arXiv:2305.08891, 2023. 1", + "[32] Ziwei Liu, Zhongqi Miao, Xiaohang Zhan, Jiayun Wang, Boqing Gong, and Stella X Yu. Large-scale long-tailed recognition in an open world. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2537-2546, 2019. 6", + "[33] Cheng Lu, Yuhao Zhou, Fan Bao, Jianfei Chen, Chongxuan Li, and Jun Zhu. Dpm-solver: A fast ode solver for diffusion probabilistic model sampling in around 10 steps. Advances in Neural Information Processing Systems, 35:5775-5787, 2022. 3", + "[34] Subhransu Maji, Esa Rahtu, Juho Kannala, Matthew Blaschko, and Andrea Vedaldi. Fine-grained visual classification of aircraft. arXiv preprint arXiv:1306.5151, 2013. 6, 2", + "[35] Chenlin Meng, Yutong He, Yang Song, Jiaming Song, Jia-jun Wu, Jun-Yan Zhu, and Stefano Ermon. Sdedit: Guided image synthesis and editing with stochastic differential equations. arXiv preprint arXiv:2108.01073, 2021. 3", + "[36] Rafael Müller, Simon Kornblith, and Geoffrey E Hinton. When does label smoothing help? Advances in neural information processing systems, 32, 2019. 6", + "[37] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741, 2021. 1, 2", + "[38] Maria-Elena Nilsback and Andrew Zisserman. Automated flower classification over a large number of classes. In 2008 Sixth Indian conference on computer vision, graphics & image processing, pages 722-729. IEEE, 2008. 6, 2", + "[39] Seulki Park, Youngkyu Hong, Byeongho Heo, Sangdoo Yun, and Jin Young Choi. The majority can help the minority: Context-rich minority oversampling for long-tailed classification. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6887-6896, 2022. 6, 7, 3", + "[40] Maitreya Patel, Tejas Gokhale, Chitta Baral, and Yezhou Yang. Conceptbed: Evaluating concept learning abilities of text-to-image diffusion models. arXiv preprint arXiv:2306.04695, 2023. 3", + "[41] Judea Pearl. Causal inference in statistics: An overview. 2009. 5", + "[42] Zeju Qiu, Weiyang Liu, Haiwen Feng, Yuxuan Xue, Yao Feng, Zhen Liu, Dan Zhang, Adrian Weller, and Bernhard" + ], + "bbox": [ + 78, + 90, + 468, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Schölkopf. Controlling text-to-image diffusion by orthogonal finetuning. Advances in Neural Information Processing Systems, 36, 2024. 3", + "[43] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 5, 1", + "[44] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models, 2021. 1", + "[45] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical Image Computing and Computer-Assisted Intervention-MICCAI 2015: 18th International Conference, Munich, Germany, October 5-9, 2015, Proceedings, Part III 18, pages 234-241. Springer, 2015. 3", + "[46] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22500-22510, 2023. 2, 3", + "[47] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, et al. Imagenet large scale visual recognition challenge. International journal of computer vision, 115:211-252, 2015. 2", + "[48] Shiori Sagawa, Pang Wei Koh, Tatsunori B Hashimoto, and Percy Liang. Distributionally robust neural networks for group shifts: On the importance of regularization for worst-case generalization. arXiv preprint arXiv:1911.08731, 2019. 7", + "[49] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, Seyed Kamyar Seyed Ghasemipour, Burcu Karagol Ayan, S. Sara Mahdavi, Rapha Gontijo Lopes, Tim Salimans, Jonathan Ho, David J Fleet, and Mohammad Norouzi. Photorealistic text-to-image diffusion models with deep language understanding, 2022. 1, 2, 6", + "[50] Dvir Samuel, Yuval Atzmon, and Gal Chechik. From generalized zero-shot learning to long-tail with class descriptors. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 286-295, 2021. 6", + "[51] Veit Sandfort, Ke Yan, Perry J Pickhardt, and Ronald M Summers. Data augmentation using generative adversarial networks (cyclegan) to improve generalizability in ct segmentation tasks. Scientific reports, 9(1):16884, 2019. 2", + "[52] Mert Bulent Sariyildiz, Karteek Alahari, Diane Larlus, and Yannis Kalantidis. Fake it till you make it: Learning transferable representations from synthetic imagenet clones. In CVPR 2023-IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023. 1", + "[53] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training" + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "17232", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "next generation image-text models. Advances in Neural Information Processing Systems, 35:25278-25294, 2022. 2", + "[54] Jordan Shipard, Arnold Wiliem, Kien Nguyen Thanh, Wei Xiang, and Clinton Fookes. Boosting zero-shot classification with synthetic data diversity via stable diffusion. arXiv preprint arXiv:2302.03298, 2023. 2", + "[55] Uriel Singer, Adam Polyak, Thomas Hayes, Xi Yin, Jie An, Songyang Zhang, Qiyuan Hu, Harry Yang, Oron Ashual, Oran Gafni, et al. Make-a-video: Text-to-video generation without text-video data. arXiv preprint arXiv:2209.14792, 2022. 2", + "[56] Shikun Sun, Longhui Wei, Zhicai Wang, Zixuan Wang, Junliang Xing, Jia Jia, and Qi Tian. Inner classifier-free guidance and its taylor expansion for diffusion models. In The Twelfth International Conference on Learning Representations, 2023. 2", + "[57] Yonglong Tian, Lijie Fan, Phillip Isola, Huiwen Chang, and Dilip Krishnan. Stablerep: Synthetic images from text-to-image models make strong visual representation learners. arXiv preprint arXiv:2306.00984, 2023. 1, 2", + "[58] Brandon Trabucco, Kyle Doherty, Max Gurinas, and Ruslan Salakhutdinov. Effective data augmentation with diffusion models. arXiv preprint arXiv:2302.07944, 2023. 2, 5, 6, 7, 3", + "[59] Narek Tumanyan, Michal Geyer, Shai Bagon, and Tali Dekel. Plug-and-play diffusion features for text-driven image-to-image translation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1921-1930, 2023. 3", + "[60] Patrick von Platen, Suraj Patil, Anton Lozhkov, Pedro Cuenca, Nathan Lambert, Kashif Rasul, Mishig Davaadorj, and Thomas Wolf. Diffusers: State-of-the-art diffusion models. https://github.com/huggingface/ diffusers, 2022.3", + "[61] Catherine Wah, Steve Branson, Peter Welinder, Pietro Perona, and Serge Belongie. The caltech-ucsd birds-200-2011 dataset. 2011. 5, 6, 2", + "[62] Zhicai Wang, Yanbin Hao, Tingting Mu, Ouxiang Li, Shuo Wang, and Xiangnan He. Bi-directional distribution alignment for transductive zero-shot learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19893-19902, 2023. 2", + "[63] Daniel Watson, William Chan, Ricardo Martin-Brualla, Jonathan Ho, Andrea Tagliasacchi, and Mohammad Norouzi. Novel view synthesis with diffusion models. arXiv preprint arXiv:2210.04628, 2022. 2", + "[64] Zongze Wu, Dani Lischinski, and Eli Shechtman. Stylespace analysis: Disentangled controls for stylegan image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12863-12872, 2021. 3", + "[65] Moon Ye-Bin, Nam Hyeon-Woo, Wonseok Choi, Nayeong Kim, Suha Kwak, and Tae-Hyun Oh. Exploiting synthetic data for data imbalance problems: Baselines from a data perspective. arXiv preprint arXiv:2308.00994, 2023. 6", + "[66] Sangdoo Yun, Dongyoon Han, Seong Joon Oh, Sanghyuk Chun, Junsuk Choe, and Youngjoon Yoo. Cutmix: Regularization strategy to train strong classifiers with localizable" + ], + "bbox": [ + 78, + 92, + 468, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "features. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6023-6032, 2019. 2, 6, 7", + "[67] An Zhang, Wenchang Ma, Xiang Wang, and Tat seng Chua. Incorporating bias-aware margins into contrastive loss for collaborative filtering. In NeurlPS, 2022. 2", + "[68] Hongyi Zhang, Moustapha Cisse, Yann N Dauphin, and David Lopez-Paz. mixup: Beyond empirical risk minimization. arXiv preprint arXiv:1710.09412, 2017. 2, 6, 7", + "[69] Chenyu Zheng, Guoqiang Wu, and Chongxuan Li. Toward understanding generative data augmentation. Advances in Neural Information Processing Systems, 36, 2024. 2", + "[70] Bolei Zhou, Agata Lapedriza, Aditya Khosla, Aude Oliva, and Antonio Torralba. Places: A 10 million image database for scene recognition. IEEE transactions on pattern analysis and machine intelligence, 40(6):1452-1464, 2017. 7", + "[71] Jun-Yan Zhu, Taesung Park, Phillip Isola, and Alexei A Efros. Unpaired image-to-image translation using cycle-consistent adversarial networks. In Proceedings of the IEEE international conference on computer vision, pages 2223-2232, 2017. 3" + ], + "bbox": [ + 501, + 92, + 890, + 375 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "17233", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2024/Enhance Image Classification via Inter-Class Image Mixup with Diffusion Model/d24eac86-8eb2-4146-81d8-2b9455b8e75f_model.json b/2024/Enhance Image Classification via Inter-Class Image Mixup with Diffusion Model/d24eac86-8eb2-4146-81d8-2b9455b8e75f_model.json new file mode 100644 index 0000000000000000000000000000000000000000..4f91d9f08810564cc5d3603d97ab2f81af8801c8 --- /dev/null +++ b/2024/Enhance Image Classification via Inter-Class Image Mixup with Diffusion Model/d24eac86-8eb2-4146-81d8-2b9455b8e75f_model.json @@ -0,0 +1,2620 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.079, + 0.131, + 0.892, + 0.153 + ], + "angle": 0, + "content": "Enhance Image Classification via Inter-Class Image Mixup with Diffusion Model" + }, + { + "type": "text", + "bbox": [ + 0.149, + 0.18, + 0.823, + 0.217 + ], + "angle": 0, + "content": "Zhicai Wang\\(^{1}\\); Longhui Wei\\(^{2\\dagger}\\), Tan Wang\\(^{3}\\), Heyu Chen\\(^{1}\\), Yanbin Hao\\(^{1}\\), Xiang Wang\\(^{1}\\), Xiangnan He\\(^{1}\\), Qi Tian\\(^{2}\\)" + }, + { + "type": "text", + "bbox": [ + 0.232, + 0.217, + 0.737, + 0.234 + ], + "angle": 0, + "content": "1University of Science and Technology of China, 2 Huawei Inc.," + }, + { + "type": "text", + "bbox": [ + 0.342, + 0.234, + 0.629, + 0.253 + ], + "angle": 0, + "content": "\\(^{3}\\) Nanyang Technological University" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.287, + 0.314, + 0.303 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.318, + 0.474, + 0.651 + ], + "angle": 0, + "content": "Text-to-image (T2I) generative models have recently emerged as a powerful tool, enabling the creation of photorealistic images and giving rise to a multitude of applications. However, the effective integration of T2I models into fundamental image classification tasks remains an open question. A prevalent strategy to bolster image classification performance is through augmenting the training set with synthetic images generated by T2I models. In this study, we scrutinize the shortcomings of both current generative and conventional data augmentation techniques. Our analysis reveals that these methods struggle to produce images that are both faithful (in terms of foreground objects) and diverse (in terms of background contexts) for domain-specific concepts. To tackle this challenge, we introduce an innovative inter-class data augmentation method known as Diff-Mix1, which enriches the dataset by performing image translations between classes. Our empirical results demonstrate that Diff-Mix achieves a better balance between faithfulness and diversity, leading to a marked improvement in performance across diverse image classification scenarios, including few-shot, conventional, and long-tail classifications for domain-specific datasets." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.678, + 0.21, + 0.695 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.704, + 0.47, + 0.856 + ], + "angle": 0, + "content": "In comparison to GAN-based models [7, 17, 25], contemporary state-of-the-art text-to-image (T2I) diffusion models exhibit enhanced capabilities in producing high-fidelity images [12, 37, 44, 49]. With the remarkable cross-modality alignment capabilities of T2I models, there is significant potential for generative techniques to enhance image classification [2, 4]. For instance, a straightforward approach entails augmenting the existing training dataset with synthetic images generated by feeding categorical textual prompts to a T2I diffusion model. However, upon reviewing prior ap" + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.287, + 0.907, + 0.509 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.519, + 0.895, + 0.673 + ], + "angle": 0, + "content": "Figure 1. Strategies to expand domain-specific datasets for improved classification are varied. Row 1 illustrates vanilla distillation from a pretrained text-to-image (T2I) model, which carries the risk of generating outputs with reduced faithfulness. Intra-class augmentation, depicted in Row 2, tends to yield samples with limited diversity to maintain high fidelity to the original class. Our proposed method, showcased in Rows 3 and 4, adopts an inter-class augmentation strategy. This involves introducing edits to a reference image using images from other classes within the training set, which significantly enriches the dataset with a greater diversity of samples." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.689, + 0.895, + 0.795 + ], + "angle": 0, + "content": "proaches employing T2I diffusion models for image classification, it becomes evident that the challenge in generative data augmentation for domain-specific datasets is producing samples with both a faithful foreground and a diverse background. Depending on whether a reference image is used in the generative process, we divide these methods into two groups:" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.796, + 0.895, + 0.902 + ], + "angle": 0, + "content": "- Text-guided knowledge distillation [52, 57] involves generating new images from scratch using category-related prompts to expand the dataset. For the off-the-shelf T2I models, such vanilla distillation presume these models have comprehensive knowledge of target domain, which can be problematic for domain-specific datasets. Insufficient domain knowledge easily makes the distillation" + }, + { + "type": "page_footnote", + "bbox": [ + 0.094, + 0.863, + 0.405, + 0.876 + ], + "angle": 0, + "content": "*This work was done during the internship in Huawei Inc.." + }, + { + "type": "page_footnote", + "bbox": [ + 0.099, + 0.876, + 0.451, + 0.888 + ], + "angle": 0, + "content": "\\(\\dagger\\) Xiang Wang and Longhui Wei are both the corresponding authors." + }, + { + "type": "page_footnote", + "bbox": [ + 0.099, + 0.888, + 0.317, + 0.9 + ], + "angle": 0, + "content": "1https://github.com/Zhicaiww/Diff-Mix" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.863, + 0.451, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "17223" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.091, + 0.092, + 0.468, + 0.152 + ], + "angle": 0, + "content": "process less effective. For example, vanilla T2I models struggle to generate images that accurately represent specific bird species based solely on their names (see Row 1 of Fig. 1)." + }, + { + "type": "text", + "bbox": [ + 0.079, + 0.153, + 0.468, + 0.318 + ], + "angle": 0, + "content": "- Generative data augmentation [1, 69] employs generative models to enhance existing images. Da-fusion [58], for instance, translates the source image into multiple edited versions within the same class. This strategy, termed intra-class augmentation, primarily introduces intra-class variations. While intra-class augmentation retains much of the original image's layout and visual details, it results in limited background diversity (see Row 2 of Fig. 1). However, synthetic images with constrained diversity may not sufficiently enhance the model's ability to discern foreground concepts." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.32, + 0.468, + 0.379 + ], + "angle": 0, + "content": "Based on these observations, a fundamental question emerges: 'Is it feasible to develop a method that optimizes both the diversity and faithfulness of synthesized data simultaneously?'" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.381, + 0.468, + 0.668 + ], + "angle": 0, + "content": "In this work, we introduce Diff-Mix, a simple yet effective data augmentation method that harnesses diffusion models to perform inter-class image interpolation, tailored for enhancing domain-specific datasets. The method encompasses two pivotal operations: personalized fine-tuning and inter-class image translation. Personalized fine-tuning [15, 46] is originally designed for customizing T2I models and enabling them to generate user-specific contents or styles. In our case, we implement the technique to tailor the model, enabling it to generate images with faithful foreground concepts. Inter-class image translation in Diff-Mix entails transforming a reference image into an edited version that incorporates prompts from different classes. This translation strategy is designed to retain the original background context while editing the foreground to align with the target concept. For instance, as depicted in the bottom rows of Fig. 1, Diff-Mix can generate images of land birds in diverse settings, such as maritime environments, enriching the dataset with a variety of counterfactual samples." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.67, + 0.468, + 0.835 + ], + "angle": 0, + "content": "Unlike previous non-generative augmentation methods, such as Mixup [68] and CutMix [66], Diff-Mix works in a foreground-perceivable inter-class interpolation manner and shares a different mechanism with the non-generative approaches. Our experiment under the conventional classification setting indicates that incorporating both CutMix and Diff-Mix could further enhance performance. Additionally, when compared with other generative approaches, we conduct experiments under few-shot and long-tail scenarios and observe consistent performance improvements. Our contributions can be summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.84, + 0.468, + 0.884 + ], + "angle": 0, + "content": "- We pinpoint the critical factors that affect the efficacy of generative data augmentation in domain-specific image classification: namely, faithfulness and diversity." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.886, + 0.468, + 0.9 + ], + "angle": 0, + "content": "- We introduce Diff-Mix, a simple yet effective generative" + }, + { + "type": "list", + "bbox": [ + 0.078, + 0.84, + 0.468, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.89, + 0.121 + ], + "angle": 0, + "content": "data augmentation strategy that leverages fine-tuned diffusion models for inter-class image interpolation." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.122, + 0.89, + 0.181 + ], + "angle": 0, + "content": "- We conduct a comparative analysis of Diff-Mix with other distillation-based and intra-class augmentation methods, as well as non-generative approaches, highlighting its unique features and benefits." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.205, + 0.649, + 0.22 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.233, + 0.892, + 0.369 + ], + "angle": 0, + "content": "Text-to-image diffusion models. Following pretraining on web-scale data, the T2I diffusion model has demonstrated robust capabilities in generating text-controlled images [37, 49, 53, 56]. Its versatility has led to diverse applications, including novel view synthesis [6, 63], concept learning [28, 46], and text-to-video generation [22, 55], among others. Recent advancements [29] also highlight the cross-modality features of such generative models, showcasing their ability to serve as zero-shot classifiers." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.379, + 0.892, + 0.74 + ], + "angle": 0, + "content": "Synthetic data for image classification. There are two perspectives on the utilization of synthetic data for image classification: knowledge distillation [2] and data augmentation [5, 51, 54, 62]. From the knowledge distillation perspective, SyntheticData [19] reports significant performance gains in both zero-shot and few-shot settings by leveraging off-the-shelf T2I models to obtain synthetic data. The work of [2] has indicated that fine-tuning the T2I model on ImageNet [47] yields improved classification accuracy by narrowing the domain gap. Some works also find that learning from the synthetic data presents strong transferability [19, 57] and robustness [4, 30, 67]. From the data augmentation standpoint, Da-fusion [58] achieves stable performance improvements on few-shot datasets by augmenting from reference images. In a related study [3], the use of StyleGAN [25] for generating interpolated images between two different domains has been shown to enhance classifier robustness for out-of-distribution data. Our work shares similarities with AMR [5], which generates realistic novel examples by interpolating between two images using GAN [16]. The distinction lies in our discussion of interpolation using the T2I diffusion model, where its noise-adding and denoising characteristics enable a smoother implementation of interpolation." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Non-generative data augmentation. Mixup [68] and CutMix [66] stand out as two prominent non-generative data augmentation methods, serving as effective regularization techniques during training. While Mixup achieves augmented samples through a convex combination of two images, CutMix achieves augmentation by cutting and pasting parts of images. However, both methods are constrained in their ability to produce realistic images. In addressing this limitation, the utilization of generative models emerges as a potential solution to alleviate this issue." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "17224" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.078, + 0.09, + 0.169, + 0.105 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.115, + 0.208, + 0.131 + ], + "angle": 0, + "content": "3.1. Preliminary" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.138, + 0.473, + 0.304 + ], + "angle": 0, + "content": "Text-to-Image diffusion model. Diffusion models generate images by gradually removing noise from a Gaussian noise source [21]. In a diffusion process with a total of \\(T\\) steps, its forward process, which gradually adds noise, is represented as a Markov chain with a Gaussian transition kernel, where \\(q(\\pmb{x}_t|\\pmb{x}_{t - 1}) = \\mathcal{N}(\\pmb{x}_t;\\sqrt{\\alpha_t}\\pmb{x}_{t - 1},(1 - \\alpha_t)\\mathbf{I})\\), where \\(x_{t}\\) represents the noisy image at step \\(t\\). The training objective at step \\(t\\) is to predict the noise to reconstruct \\(\\pmb{x}_{t - 1}\\). When training a text-conditioned diffusion model, the simplified training objective can be summarized as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.171, + 0.307, + 0.47, + 0.333 + ], + "angle": 0, + "content": "\\[\n\\mathbb {E} _ {\\epsilon , \\mathbf {X}, \\mathbf {c}, t} \\left[ \\| \\epsilon - \\epsilon_ {\\theta} (\\mathbf {x} _ {t}, \\mathbf {c}, t) \\| _ {2} ^ {2} \\right], \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.338, + 0.47, + 0.369 + ], + "angle": 0, + "content": "where \\(\\epsilon_{\\theta}\\) represents the predicted noise, and \\(c\\) is the encoded text caption associated with the image \\(\\pmb{x}\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.374, + 0.473, + 0.571 + ], + "angle": 0, + "content": "T2I personalization. T2I personalization aims to personalize a diffusion model for generating specific concepts using a limited number of concept-oriented images [28, 40, 42]. These concepts are typically represented using identifiers (e.g., “[V]”). As a result, we formalize the constructed image-caption set as \\( \\pmb{x} \\), “photo of a [V]”. Various personalization methods differ in their fine-tuning strategies. For instance, Textual Inversion (TI) [15] makes the identifier learnable, but other modules are not fine-tuned, potentially sacrificing some faithfulness in image generation. On the other hand, Dreambooth (DB) [46] fine-tunes the U-Net [45] for more refined personalized generation but faces the challenge of increased computational cost." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.577, + 0.473, + 0.759 + ], + "angle": 0, + "content": "Image-to-image translation. Image translation enables image synthesis and editing using a reference image as guidance [24, 64, 71]. Diffusion-based image translation methods can generate fine edits, which refer to subtle modifications, with varying degrees of shift relative to the reference image [8, 35, 59]. Here, we draw inspiration from SDEdit [35] to perform edits on the reference image, where the target image \\( \\mathbf{x}^{\\mathrm{tar}} \\) is translated from a reference image \\( \\mathbf{x}^{\\mathrm{ref}} \\). During translation, the reverse process does not traverse the full process but starts from a certain step \\( \\lfloor sT \\rfloor \\), where \\( s \\in [0,1] \\) controls the insertion position of the reference image with noise, as follows," + }, + { + "type": "equation", + "bbox": [ + 0.141, + 0.765, + 0.47, + 0.79 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {x} _ {\\lfloor s T \\rfloor} = \\sqrt {\\tilde {\\alpha} _ {\\lfloor s T \\rfloor}} \\boldsymbol {x} _ {0} ^ {\\text {r e f}} + \\sqrt {1 - \\tilde {\\alpha} _ {\\lfloor s T \\rfloor}} \\epsilon . \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.795, + 0.473, + 0.841 + ], + "angle": 0, + "content": "By adjusting the strength parameter \\( s \\), one can strike a balance between the diversity of the generated images and their faithfulness to the reference image." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.848, + 0.273, + 0.863 + ], + "angle": 0, + "content": "3.2. General Framework" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.871, + 0.473, + 0.903 + ], + "angle": 0, + "content": "The Diff-Mix pipeline consists of two key steps. Firstly, to produce more faithful images for domain-specific datasets," + }, + { + "type": "image", + "bbox": [ + 0.501, + 0.088, + 0.895, + 0.13 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.655, + 0.131, + 0.74, + 0.142 + ], + "angle": 0, + "content": "(a) Real images" + }, + { + "type": "image", + "bbox": [ + 0.502, + 0.143, + 0.894, + 0.184 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.608, + 0.185, + 0.786, + 0.197 + ], + "angle": 0, + "content": "(b) Synthetic images (vanilla SD)" + }, + { + "type": "image", + "bbox": [ + 0.502, + 0.198, + 0.893, + 0.237 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.579, + 0.239, + 0.815, + 0.251 + ], + "angle": 0, + "content": "(c) Synthetic images (SD fine-tuned via DB)" + }, + { + "type": "image", + "bbox": [ + 0.502, + 0.252, + 0.892, + 0.292 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.568, + 0.294, + 0.825, + 0.306 + ], + "angle": 0, + "content": "(d) Synthetic images (SD fine-tuned via \\(\\mathrm{TI + DB}\\))" + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.318, + 0.894, + 0.416 + ], + "angle": 0, + "content": "Figure 2. Examples of \"American Three toed Woodpecker\". (a) Real images from the training set. (b-d) synthetic images generated using different fine-tuned models with the same number of fine-tuning steps. TI+DB indicates both text embedding and U-Net are fine-tuned. TI+DB achieves a more faithful output compared to DB alone (check the head and wing patterns of the birds)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.436, + 0.894, + 0.528 + ], + "angle": 0, + "content": "we propose treating it as a T2I personalization problem and fine-tuning the Stable Diffusion (SD). Subsequently, to enhance the diversity of synthetic data beyond the well-fitted training distribution, we employ inter-class image translation. This process produces interpolated images with increased background diversity for each class." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.538, + 0.742, + 0.553 + ], + "angle": 0, + "content": "3.3. Fine-tune Diffusion Model" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.562, + 0.894, + 0.638 + ], + "angle": 0, + "content": "Vanilla distillation tends to be less effective, especially as the number of training shots increases (refer to Sec. 4.1). In order to mitigate the distribution gap, we propose finetuning Stable Diffusion in conjunction with current widely-used T2I personalization strategies." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.645, + 0.895, + 0.902 + ], + "angle": 0, + "content": "Dreambooth meets Textual Inversion. Many fine-grained datasets provide terminological names for their categories, like \"American Three toed Woodpecker\" and \"Pileated Woodpecker\". We could construct image-text pairs using category-related prompts and fine-tune the denoising network of SD using Eq. 1, which is analogous to Dreambooth. However, we observe that directly incorporating these specialized terms into the text during fine-tuning can impede convergence and hinder the generation of faithful images. We attribute this challenge to the semantic proximity of terminology within a fine-grained domain, where fine-tuning the vision module alone tends to be less effective at distinguishing two similar classes within the same family, like \"Woodpecker\". Inspired by Textual Inversion [15], we opt to replace the terminological name in the dataset with \"[Vi] [metaclass]\" where \"[Vi\"] is a learnable identifier, and \\(i\\) varies from 1 to \\(N\\), represent-" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "17225" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.084, + 0.089, + 0.47, + 0.356 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.367, + 0.471, + 0.451 + ], + "angle": 0, + "content": "Figure 3. Fine-tuning framework of Diff-Mix operates as follows: Initially, we replace the class name with a structured identifier formatted as “[\\(\\mathsf{V}^i\\)” [metaclass]}, thereby sidestepping the need for specific terminological expressions. Next, we engage in joint fine-tuning of these identifiers and the low-rank residues (LoRA) of U-Net to capture the domain-specific distribution." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.467, + 0.47, + 0.59 + ], + "angle": 0, + "content": "ing the category index. The illustration of our fine-tuning strategy is presented in Fig. 3. The term “[metaclass]” is determined by the theme of the current dataset, such as “bird” for a fine-grained bird dataset. By concurrently fine-tuning the identifier and the U-Net, we empower the model to quickly adapting to the fine-grained domain, allowing it to generate faithful images using the identifier (see comparison between Row 3 and Row 4 in Fig. 2)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.596, + 0.47, + 0.732 + ], + "angle": 0, + "content": "Parameter efficient fine-tuning. In this context, we embrace the parameter-efficient fine-tuning strategy known as LoRA [23]. LoRA distinguishes itself by fine-tuning the residue of low-rank matrices instead of directly fine-tuning the pre-trained weight matrices. To elaborate, consider a weight matrix \\( \\mathbf{W} \\in \\mathbb{R}^{m \\times n} \\). The tunable residual matrix \\( \\Delta \\mathbf{W} \\) comprises two low-rank matrices: \\( \\mathbf{A} \\in \\mathbb{R}^{m \\times d} \\) and \\( \\mathbf{B} \\in \\mathbb{R}^{n \\times d} \\), defined as \\( \\Delta \\mathbf{W} = \\mathbf{A}\\mathbf{B}^{\\top} \\). As a default configuration, we set the rank \\( d \\) to 10." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.743, + 0.405, + 0.76 + ], + "angle": 0, + "content": "3.4. Data Synthesis Using Diffusion Model" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.767, + 0.469, + 0.828 + ], + "angle": 0, + "content": "In generating pseudo data, three strategies can be used with our fine-tuned diffusion model \\(^{2}\\): (1) distillation-based method Diff-Gen, (2) intra-class augmentation Diff-Aug, and (3) inter-class augmentation Diff-Mix." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.834, + 0.47, + 0.866 + ], + "angle": 0, + "content": "Diff-Gen and Diff-Aug. For a target class \\( y^i \\) and its textual condition, \"photo of a \\( [V^i] \\) [metaclass]\", both" + }, + { + "type": "page_footnote", + "bbox": [ + 0.077, + 0.875, + 0.47, + 0.9 + ], + "angle": 0, + "content": "2We use the prefix \"Diff-\" denotes the T2I model is fine-tuned and \"Real-\" denotes the vanilla T2I model." + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.089, + 0.89, + 0.356 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.367, + 0.893, + 0.452 + ], + "angle": 0, + "content": "Figure 4. Examples of images translated using Diff-Mix and Diff-Aug across various strengths. Diff-Aug employs the same target and reference image classes, typically resulting in subtle modifications. Diff-Mix progressively adjusts the foreground to align with the target class as the translation strength increases, while preserving the background layout from the reference image." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.469, + 0.893, + 0.681 + ], + "angle": 0, + "content": "methods generate synthetic samples annotated with class \\( i \\). Specifically, Diff-Gen generates samples from scratch by initializing with random Gaussian noise and proceeding through the full reverse diffusion process with \\( T \\) steps. Diff-Gen can produce images aligned with its fine-tuned distribution. In contrast, Diff-Aug sacrifices a portion of diversity and generate images by editing on a reference image. Specifically, it randomly sample a image from the intra-class training set and enhances the image through image translation using Eq. 2. The term \"intra-class\" means that the conditioning prompts are constructed based on the ground truth categories of images, and such a denoising process tends to introduce less variation, particularly for the foreground concepts (see top rows of Fig. 1)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.689, + 0.893, + 0.901 + ], + "angle": 0, + "content": "Diff-Mix. Diff-Mix employs the same translation process as Diff-Aug, but the reference image is sampled from the full training set rather than intra-class set to enable interclass interpolation. The key difference is that Diff-Mix can generate numerous counterfactual examples, such as a blackbird in the sea (see fourth row of Fig. 1). This necessitates that downstream models make a more refined differentiation of category attributes, thereby reducing the impact of spurious correlations introduced by variations in the background. Denoting the label of the reference image as \\( y^{j} \\), by inserting the reference image into the reverse process with the prompt \"photo of a \\([V^i]\\) [metaclass]\", we can obtain interpolated images between the \\( i_{\\mathrm{th}} \\) and \\( j_{\\mathrm{th}} \\) categories. By controlling the intensity \\( s \\), we can precisely" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "17226" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.088, + 0.101, + 0.255, + 0.187 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.282, + 0.098, + 0.45, + 0.184 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.205, + 0.47, + 0.289 + ], + "angle": 0, + "content": "Figure 5. A schematic explanation of Diff-Mix's effectiveness using structural casual model [41]. \\( \\pmb{x}_{fg} \\) is the foreground that determines the real class label, \\( \\pmb{x}_{bg} \\) denotes the background. \\( \\pmb{x}_{fg} \\rightarrow \\pmb{z} \\rightarrow y \\) is the causal path that we are focusing and \\( \\pmb{x}_{fg} \\gets I \\rightarrow \\pmb{x}_{bg} \\rightarrow \\pmb{z} \\rightarrow y \\) is the backdoor path that introduces spurious relations between \\( \\pmb{x}_{fg} \\) and \\( y \\)." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.3, + 0.469, + 0.36 + ], + "angle": 0, + "content": "manage the interpolation process. When annotating the synthetic image, unlike Mixup and Cutmix, we take into account the non-linear nature of diffusion translation. Thus the annotation function is given by" + }, + { + "type": "equation", + "bbox": [ + 0.192, + 0.37, + 0.469, + 0.387 + ], + "angle": 0, + "content": "\\[\n\\tilde {y} = (1 - s ^ {\\gamma}) y ^ {i} + s ^ {\\gamma} y ^ {j}, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.397, + 0.469, + 0.473 + ], + "angle": 0, + "content": "where \\(\\gamma\\) is a hyperparameter introducing non-linearity. Our empirical findings indicate that a \\(\\gamma\\) smaller than 1 is favored. Additionally, in low-shot cases, the samples with higher confidence in the target class are preferred (see details in Sec. 4.5)." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.479, + 0.469, + 0.691 + ], + "angle": 0, + "content": "Construct synthetic dataset. To construct the synthetic dataset using Diff-Mix, similar to Da-fusion [58], we adopt a randomized sampling strategy (\\(s \\in \\{0.5, 0.7, 0.9\\}\\)) for the selection of translation strength. While applying the inter-class editing, we observe that Diff-Mix tends to produce more undesirable samples compared to Diff-Aug. These undesirable samples have incomplete foreground such as fragmented bird bodies. This is caused by the intrinsic shape and pose differences among classes. To mitigate this, we introduce a simple data-cleaning approach to reduce the proportion of such problematic images. We utilize the large vision language model CLIP [43] to assess the confidence in the content, serving as the filtering criterion. Further details can be found in the supplementary materials (SMs)." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.696, + 0.469, + 0.817 + ], + "angle": 0, + "content": "Analysis. We depict the core insight of Diff-Mix in Fig. 5. To eliminate the spurious correlation introduced by \\( \\pmb{x}_{bg} \\), learning on the synthetic set with randomized \\( \\pmb{x}_{bg} \\) (background) can cut off spurious correlation, forcing the classification model to infer only from the foreground. The study in Fig. 7 (b) shows that the more diverse the background (larger the referable class number), the better the performance on the CUB test set." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.831, + 0.208, + 0.848 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.856, + 0.469, + 0.9 + ], + "angle": 0, + "content": "In this section, we investigate the effectiveness of Diff-Mix in domain-specific datasets. The key questions we aim to address are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.891, + 0.151 + ], + "angle": 0, + "content": "Q1: Can generative inter-class augmentation lead to more significant performance gains in downstream tasks compared to those intra-class augmentation methods and distillation-based methods?" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.152, + 0.891, + 0.182 + ], + "angle": 0, + "content": "Q2: Is improved background diversity the secret weapon of Diff-Mix for enhancing the performance?" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.183, + 0.891, + 0.227 + ], + "angle": 0, + "content": "Q3: How to choose the fine-tuning strategy and annotation strategy to boost the performance gains for the inter-class augmentation?" + }, + { + "type": "list", + "bbox": [ + 0.513, + 0.092, + 0.891, + 0.227 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.231, + 0.892, + 0.352 + ], + "angle": 0, + "content": "To address Q1, we separately discuss these questions in few-shot settings in Sec. 4.1, conventional classification in Sec. 4.2, as well as long-tail classification in Sec. 4.3. Additionally, to answer Q2, we conduct a test for background robustness in Sec. 4.4 and perform an ablation study focusing on the size and diversity of synthetic data in Sec. 4.5. For Q3, we conduct an ablation study to empirically discover effective strategies for deployment in Sec. 4.5." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.364, + 0.714, + 0.379 + ], + "angle": 0, + "content": "4.1. Few-shot Classification" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.388, + 0.892, + 0.599 + ], + "angle": 0, + "content": "Experimental Setting. To investigate the impact of different data expansion methods, we conduct few-shot experiments on a domain-specific dataset Caltech-UCSD Birds (CUB) [61], with shot numbers of 1, 5, 10, and all. For augmentation-based methods, the synthetic dataset is constructed using various translation strengths \\((s)\\), specifically, \\(s \\in \\{0.1, 0.3, \\dots, 1.0\\}\\). We expand the original training set with a multiplier of 5 and cache the synthesized dataset locally for joint training. Real data are replaced by synthetic data proportionally during training, and the replacement probability \\(p\\) is set as 0.1, 0.2, 0.3, and 0.5 for all-shot, 10-shot, 5-shot, and 1-shot classification, respectively. All experiments use ResNet50 with an input resolution of \\(224 \\times 224\\). Additional details can be found in the SMs." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.606, + 0.891, + 0.698 + ], + "angle": 0, + "content": "Comparison methods. To unveil the trade-off between faithfulness and diversity resulting from different expansion strategies, we compared Diff-Mix with Diff-Gen and Diff-Aug. Furthermore, we conduct experiments on expansion strategies using vanilla SD: Real-Mix, Real-Gen, and Real-Aug, where 'Real' signifies that SD is not fine-tuned." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.704, + 0.891, + 0.763 + ], + "angle": 0, + "content": "Main results. To answer Q1 under the few-shot classification setting, we augment CUB using X-Mix, X-Aug, and X-Gen, where 'X' denotes 'Diff/Real' for simplicity. The results are shown in Fig. 6, and we can observe that:" + }, + { + "type": "text", + "bbox": [ + 0.501, + 0.765, + 0.892, + 0.87 + ], + "angle": 0, + "content": "1. Diff-Mix generally outperforms the intra-class competitor X-Aug and distillation competitor X-Gen in various few-shot scenarios. It tends to achieve higher gains when the strength \\( s \\) is relatively large, i.e., \\( \\{0.5, 0.7, 0.9\\} \\), where the foreground has been edited to match the target class and the background retains similarities to the reference image." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.891, + 0.9 + ], + "angle": 0, + "content": "2. Among the Real-X methods, distillation tends to be more effective than the augmentation method when the shot" + }, + { + "type": "list", + "bbox": [ + 0.5, + 0.765, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "17227" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.091, + 0.271, + 0.202 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.15, + 0.208, + 0.203, + 0.22 + ], + "angle": 0, + "content": "(a) 1-shot" + }, + { + "type": "image", + "bbox": [ + 0.276, + 0.09, + 0.465, + 0.201 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.344, + 0.208, + 0.398, + 0.22 + ], + "angle": 0, + "content": "(b) 5-shot" + }, + { + "type": "image", + "bbox": [ + 0.471, + 0.091, + 0.659, + 0.201 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.536, + 0.208, + 0.594, + 0.22 + ], + "angle": 0, + "content": "(c) 10-shot" + }, + { + "type": "image", + "bbox": [ + 0.665, + 0.091, + 0.858, + 0.202 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.727, + 0.208, + 0.791, + 0.22 + ], + "angle": 0, + "content": "(d) All-shot" + }, + { + "type": "image_caption", + "bbox": [ + 0.335, + 0.232, + 0.63, + 0.246 + ], + "angle": 0, + "content": "Figure 6. Few-shot classification results on CUB." + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.262, + 0.468, + 0.382 + ], + "angle": 0, + "content": "number is low, but the trend reverses as the shot number increases (compare Real-Gen with Real-Aug). Real-Gen's samples even show less effective than Real-Aug \\((s = 1.0)\\) under the all-shot case3. This indicates that the importance of faithfulness in the trade-off between faithfulness and diversity increases with the shot number. Additionally, Real-Mix exhibits consistent and stable improvement over the other two methods." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.383, + 0.469, + 0.459 + ], + "angle": 0, + "content": "3. Diff-Gen consistently outperforms Real-Gen under four scenarios. Notably, Real-Gen's performance declines below that of the baseline as the shot numbers reach 10, showcasing the importance of the fine-tuning process which increases the faithfulness of synthetic samples." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.472, + 0.323, + 0.487 + ], + "angle": 0, + "content": "4.2. Conventional Classification" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.497, + 0.469, + 0.723 + ], + "angle": 0, + "content": "Experimental setting. To test whether Diff-Mix can further boost performance in a more challenging setting, i.e., under the all-shot scenario with high input resolution, we conduct conventional classification on five domain-specific datasets: CUB [61], Stanford Cars [27], Oxford Flowers [38], Stanford Dogs [26], and FGVC Aircraft [34]. Two backbones are employed: pretrained (ImageNet1K [49]) ResNet50 [18] with input resolution \\( 448 \\times 448 \\), and pretrained (ImageNet21K) ViT-B/16 [13] with input resolution \\( 384 \\times 384 \\). Label smoothing [36] is applied across all datasets with a confidence level of 0.9. For all expansion strategies, the expansion multiplier is 5 and the replacement probability \\( p \\) is 0.1. Besides, we use a randomized sampling strategy \\( (s \\in \\{0.5, 0.7, 0.9\\}) \\) and a fixed \\( \\gamma \\) (0.5, and this is specific to Diff-Mix) for all datasets." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.731, + 0.468, + 0.837 + ], + "angle": 0, + "content": "Comparison methods. We compare Diff-Mix with (1) Real-Filtering (RF) [19], a variation of Real-Gen that incorporates clip filtering, (2) Real-Guidance (RG) [19], which augments the dataset using intra-class image translation at low strength (\\(s = 0.1\\)), (3) Da-Fusion [58], a method that solely fine-tunes the identifier to personalize each class and employs randomized sampling strategy" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.262, + 0.89, + 0.293 + ], + "angle": 0, + "content": "\\((s \\in \\{0.25, 0.5, 0.75, 1.0\\})\\), and non-generative augmentation methods (4) CutMix [66] and (5) Mixup [68]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.299, + 0.892, + 0.602 + ], + "angle": 0, + "content": "Main results. We show the classification accuracy for different data expansion strategies in Table 1, our observations can be summarized: (1) Diff-Mix consistently demonstrates stable improvements across the majority of settings. Its average performance gain across the five datasets exceeds that of baselines employing intra-class augmentation methods (RG and Da-fusion), distillation method (RF), and non-generative data augmentation techniques (CutMix and Mixup). (2) Real-filtering, analogous to the discussion of Real-Gen, exhibits performance degradation on most datasets due to the distribution gap. (3) The combined use of Diff-Mix and CutMix often yields better performance gains. This is attributed to the distinct enhancement mechanisms of the two methods, i.e., vicinal risk minimization [11, 68] and foreground-background disentanglement. (4) Diff-Mix does not exhibit significant performance improvement in the dog dataset. We attribute this lack of improvement to the complexity of the dog dataset, which often contains multiple subjects in a single image, impeding effective foreground editing (refer to the SMs for visual examples)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.614, + 0.719, + 0.63 + ], + "angle": 0, + "content": "4.3. Long-Tail Classification" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.645, + 0.892, + 0.81 + ], + "angle": 0, + "content": "Experiment setting. Following the settings of previous long-tail dataset constructions [9, 32, 39], we create two domain-specific long-tail datasets, CUB-LT [50] and Flower-LT. The imbalance factor controls the exponential distribution of the imbalanced data, where a larger value indicates a more imbalanced distribution. To leverage generative models for long-tail classification, we adopt the approach of SYNAuG [65], which uniformize the imbalanced real data distribution using synthetic data. Translation strength \\( s \\) (0.7) and \\( \\gamma \\) (0.5) are fixed for both Diff-Mix and Real-Mix." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.818, + 0.892, + 0.879 + ], + "angle": 0, + "content": "Comparison methods. We compare Diff-Mix with Real-Mix, Real-Gen, the non-generative CutMix-based oversampling approach CMO [39], and its enhanced variant with two-stage deferred re-weighting [9] (CMO+DRW)." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.886, + 0.891, + 0.901 + ], + "angle": 0, + "content": "Main results. We present the long-tail classification re" + }, + { + "type": "page_footnote", + "bbox": [ + 0.076, + 0.851, + 0.469, + 0.901 + ], + "angle": 0, + "content": "\\(^3\\)Real-Aug \\((s = 1.0)\\) remains analogous to the reference image (slightly higher faithfulness compared to Real-Gen) because the discrete forward process cannot approximate the ideal normal distribution within a limited number of steps \\((T = 25)\\)." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "17228" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.132, + 0.089, + 0.842, + 0.332 + ], + "angle": 0, + "content": "
BackboneAug. MethodFT StrategyDataset
CUBAircraftFlowerCarDogAvg
ResNet50@448--86.6489.0999.2794.5487.4891.40
Cutmix[66]-87.2389.4499.2594.7387.5991.65+0.25
Mixup[68]-86.6889.4199.4094.4987.4291.48+0.08
Real-filtering [19]†X85.6088.5499.0994.5987.3091.22-0.18
Real-guidance [19]†X86.7189.0799.2594.5587.4091.59+0.19
Da-fusion [58]†TI86.3087.6499.3794.6987.3391.07-0.58
Diff-MixTI + DB87.1690.2599.5495.1287.7491.96+0.56
Diff-Mix + CutmixTI + DB87.5690.0199.4795.2187.8992.03+0.63
ViT-B/16@384--89.3783.5099.5694.2192.0691.74
Cutmix[66]-90.5283.5099.6494.8392.1392.12+0.38
Mixup[68]-90.3284.3199.7394.9892.0292.27+0.53
Real-filtering [19]†X89.4983.0799.3694.6691.9191.69-0.05
Real-guidance [19]†X89.5483.1799.5994.6592.0591.80+0.06
Da-fusion [58]†TI89.4081.8899.6194.5392.0791.50-0.24
Diff-MixTI + DB90.0584.3399.6495.0991.9992.22+0.48
Diff-Mix + CutmixTI + DB90.3585.1299.6895.2691.8992.46+0.72
" + }, + { + "type": "table_caption", + "bbox": [ + 0.157, + 0.34, + 0.809, + 0.356 + ], + "angle": 0, + "content": "Table 1. Conventional classification in six fine-grained datasets. \\( {}^{ \\dagger } \\) ,indicates our reproduced results using SD." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.371, + 0.465, + 0.468 + ], + "angle": 0, + "content": "
MethodIF=1005010
ManyMediumFewAll
CE79.1164.2813.4833.6544.8258.13
CMO [39]78.3258.5714.7832.9444.0857.62
CMO + DRW [10]78.9756.3614.6632.5746.4359.25
Real-Gen84.8865.2330.6845.8653.4361.42
Real-Mix (s=0.7)84.6366.3434.4447.7555.6762.27
Diff-Mix (s=0.7)84.0767.7936.5550.3558.1964.48
" + }, + { + "type": "table_caption", + "bbox": [ + 0.151, + 0.478, + 0.414, + 0.492 + ], + "angle": 0, + "content": "Table 2. Long-tail classification in CUB-LT." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.503, + 0.465, + 0.6 + ], + "angle": 0, + "content": "
MethodIF=1005010
ManyMediumFewAll
CE99.1994.9558.1880.4390.8795.07
CMO [39]99.2595.1967.4583.9591.4395.19
CMO+ DRW [10]99.9795.0667.3184.0792.0695.92
Real-Gen98.6495.5566.1083.5691.8495.22
Real-Mix (s=0.7)99.8796.2668.5385.1992.9696.04
Diff-Mix (s=0.7)99.2596.9878.4189.4693.8696.63
" + }, + { + "type": "table_caption", + "bbox": [ + 0.145, + 0.61, + 0.42, + 0.625 + ], + "angle": 0, + "content": "Table 3. Long-tail classification in Flower-LT." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.636, + 0.47, + 0.803 + ], + "angle": 0, + "content": "sults for CUB-LT in Table 2 and Flower-LT in 3. The observations are as follows: (1) Generative approaches exhibit superior performance in tackling imbalanced classification issues compared to CutMix-based methods (CMO and CMO+DRW). (2) Real-Mix surpasses Real-Gen in performance across various imbalance factors in two datasets. This indicates that tail classes can benefit from the enhanced diversity by leveraging the visual context of majority classes. (3) Diff-Mix generally achieves the best performance among the compared strategies, especially at the low-shot case, highlighting the importance of fine-tuning." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.816, + 0.3, + 0.832 + ], + "angle": 0, + "content": "4.4. Background Robustness" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.471, + 0.903 + ], + "angle": 0, + "content": "In this section, we aim to evaluate Diff-Mix's robustness to background shifts, specifically, whether synthesizing more diverse samples can improve the classification model's generalizability when the background is altered. To achieve" + }, + { + "type": "table", + "bbox": [ + 0.51, + 0.372, + 0.879, + 0.446 + ], + "angle": 0, + "content": "
GroupBase.CutMixDA-fusionDiff-AugDiff-Mix
(waterbird, water)59.5062.4660.9061.8363.83
(waterbird, land)56.7060.1258.1060.1263.24
(landbird, land)73.4873.3972.9473.0475.64
(landbird, water)73.9774.7272.7773.5274.36
Avg.70.1971.2369.9070.2872.47
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.459, + 0.892, + 0.543 + ], + "angle": 0, + "content": "Table 4. Classification results across four groups in Waterbird [48]. Waterbird is an out-of-distribution dataset for CUB, crafted by segmenting CUB's foregrounds and paste them into the scene images from Places [70]. The constructed dataset can be divided into four groups based on the composition of foregrounds (waterbird and landbird) and backgrounds (water and land)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.564, + 0.892, + 0.777 + ], + "angle": 0, + "content": "this, we utilize an out-of-distribution test set for CUB, namely Waterbird [48]. We then perform inference on the whole Waterbird set using classifiers that have been trained on either the original CUB dataset or its expanded variations. In Table 4, we present the classification accuracies across the four groups and compare Diff-Mix with other intra-class methods (Da-fusion and Diff-Aug) as well as CutMix. We observe that Diff-Mix generally outperforms its counterparts and achieves a significant performance improvement \\((+6.5\\%)\\) in the challenging counterfactual group (waterbirds with land backgrounds). It is important to highlight that the background scenes in the Waterbird dataset are novel to CUB, requiring the classification model to have a stronger perceptual capability for the images' foregrounds." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.787, + 0.618, + 0.801 + ], + "angle": 0, + "content": "4.5. Discussion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.81, + 0.895, + 0.903 + ], + "angle": 0, + "content": "In this section, we address Q2 by examining the impact of size and diversity on synthetic data. Furthermore, we perform an ablation study to assess the effects of fine-tuning strategies and training hyperparameters of Diff-Mix, which is aimed at answering Q3. Unless specified otherwise, our discussions are based on experiments conducted using CUB" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "17229" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.089, + 0.264, + 0.195 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.165, + 0.2, + 0.18, + 0.21 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.284, + 0.089, + 0.468, + 0.195 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.369, + 0.2, + 0.385, + 0.21 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.223, + 0.47, + 0.251 + ], + "angle": 0, + "content": "Figure 7. Comparison of results across various (a) synthetic data sizes and (b) numbers of referable classes for each target class." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.274, + 0.44, + 0.289 + ], + "angle": 0, + "content": "with ResNet50, where inputs are resized to \\(224 \\times 224\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.297, + 0.47, + 0.524 + ], + "angle": 0, + "content": "Size and diversity of synthetic Data. The relationship between performance gain and the size of synthetic data is depicted in Fig. 7 (a), where a classification model was trained with synthetic data of varying sizes. We observe a monotonically increasing trend as the multiplier for the synthetic dataset ranges from 2 to 10. Ideally, the combination choices of \\((x_{i},y_{j})\\) are in the order of \\(N|D^{train}|\\) (\\(N = 200\\) for CUB). Furthermore, we limit the number of referable classes for each target class, which means the number of referable backgrounds decreases, resulting in a synthetic dataset of relatively lower diversity. The results are shown in Fig. 7 (b), and we observe a consistent improvement in performance as the number of referable classes increases. These results consistently underscore the critical role of background diversity introduced by Diff-Mix." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.531, + 0.47, + 0.744 + ], + "angle": 0, + "content": "Impact of fine-tuning strategy. Here we compare three different fine-tuning strategies: TI, DB, and the combined \\(\\mathrm{TI + DB}\\). All strategies share the same fine-tuning hyperparameters and training steps (35000). To evaluate the distribution gap, we compute the FID score [20] of synthesized images (Diff-Gen) with the training set. As illustrated in Table 5, we observe that both \\(\\mathrm{TI + DB}\\) and TI have lower FID scores than DB. This can be attributed to the fact that semantic proximity impedes the convergence process. Additionally, while using TI alone results in a relatively low FID score, the improvements in performance are limited. This limitation stems from TI's inability to accurately reconstruct detailed concept (foreground) information, as it is primarily fine-tuned at the semantic level [28]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Annotation function. In this section, we discuss the impact of the choices of the translation strength \\( s \\) and the non-linear factor \\( \\gamma \\) in Eq. 3. As shown in Table 6, we observe that as the translation strength decreases, the optimal value for \\( \\gamma \\) also decreases, which underscores the nonlinearity of Diff-Mix. The comparison between the 5-shot and all-shot settings indicates that the model tends to prefer a more diverse synthetic dataset when the number of training shots is small (\\( s = 0.9 \\) for 5-shot, \\( s = 0.7 \\) for all-shot). Besides, a larger confidence in the target class is pre" + }, + { + "type": "table", + "bbox": [ + 0.514, + 0.089, + 0.878, + 0.149 + ], + "angle": 0, + "content": "
BaselineTIDBTI + DB
5-shotFID (Diff-Gen)-18.2619.5518.43
Acc. (Diff-Mix)50.9057.6456.1159.41
All-shotFID (Diff-Gen)-14.1314.6413.99
Acc. (Diff-Mix)81.6081.8681.9982.85
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.159, + 0.89, + 0.201 + ], + "angle": 0, + "content": "Table 5. Comparison of distribution gap and classification accuracy across three fine-tuning strategies. TI solely fine-tunes the identifier, and DB solely fine-tunes the U-Net, and \\(\\mathrm{TI + DB}\\)" + }, + { + "type": "table", + "bbox": [ + 0.515, + 0.214, + 0.878, + 0.315 + ], + "angle": 0, + "content": "
γ5-shotAll-shot
s = 0.5s = 0.7s = 0.9s = 0.5s = 0.7s = 0.9
1.5-4.50-0.31+10.30-1.08+0.92+0.90
1.0-2.31+2.99+10.79+0.25+1.14+0.90
0.5+2.35+8.44+11.01+0.92+1.30+0.86
0.3+3.94+9.41+11.15+0.97+1.24+0.69
0.1+6.18+9.86+10.84+0.50+0.88+0.84
0.0+5.25+9.41+11.06+0.38+0.63+0.54
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.324, + 0.892, + 0.381 + ], + "angle": 0, + "content": "Table 6. Comparison of performance gain across various \\(\\gamma\\) and translation strength \\(s\\). Lower \\(\\gamma\\) indicates a higher confidence over target class, e.g. (\\(\\gamma = 0.1\\), \\(s = 0.7\\)) results in \\(0.04y_{i} + 0.96y_{j}\\) and (\\(\\gamma = 0.5\\), \\(s = 0.7\\)) results in \\(0.16y_{i} + 0.84y_{j}\\)." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.407, + 0.892, + 0.499 + ], + "angle": 0, + "content": "ferred when the shot number is small (\\(\\gamma = 0.1\\) for 5-shot, \\(\\gamma = 0.5\\) for all-shot). A possible explanation is that the all-shot setting is less tolerant towards unrealistic images, as discussed in Section 6. Empirically, we recommend choosing a higher translation strength (\\(s \\in [0.5, 0.7, 0.9]\\)) and a smaller \\(\\gamma\\) (\\(\\gamma \\in [0.1, 0.3, 0.5]\\)) as a conservative option." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.511, + 0.619, + 0.527 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.537, + 0.892, + 0.687 + ], + "angle": 0, + "content": "In this work, we investigate two pivotal aspects, faithfulness and diversity, that are critical for the current state-of-the-art text-to-image generative models to enhance image classification tasks. To achieve a more effective balance between these two aspects, we propose an inter-class augmentation strategy that leverages Stable Diffusion. This method enables generative models to produce a greater diversity of samples by editing images from other classes and shows consistent performance improvement across various classification tasks." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.702, + 0.681, + 0.719 + ], + "angle": 0, + "content": "6. Acknowledgement" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.727, + 0.894, + 0.787 + ], + "angle": 0, + "content": "This research is mainly supported by the National Natural Science Foundation of China (92270114). This work is also partially supported by the National Key R&D Program of China under Grant 2021ZD0112801." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "17230" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.175, + 0.107 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.157 + ], + "angle": 0, + "content": "[1] Antreas Antoniou, Amos Storkey, and Harrison Edwards. Data augmentation generative adversarial networks. arXiv preprint arXiv:1711.04340, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.157, + 0.472, + 0.213 + ], + "angle": 0, + "content": "[2] Shekoofeh Azizi, Simon Kornblith, Chitwan Sahara, Mohammad Norouzi, and David J Fleet. Synthetic data from diffusion models improves imagenet classification. arXiv preprint arXiv:2304.08466, 2023.1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.213, + 0.472, + 0.268 + ], + "angle": 0, + "content": "[3] Haoyue Bai, Ceyuan Yang, Yinghao Xu, S-H Gary Chan, and Bolei Zhou. Improving out-of-distribution robustness of classifiers via generative interpolation. arXiv preprint arXiv:2307.12219, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.269, + 0.472, + 0.312 + ], + "angle": 0, + "content": "[4] Hritik Bansal and Aditya Grover. Leaving reality to imagination: Robust classification via generated datasets. arXiv preprint arXiv:2302.02503, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.313, + 0.472, + 0.368 + ], + "angle": 0, + "content": "[5] Christopher Beckham, Sina Honari, Vikas Verma, Alex M Lamb, Farnoosh Ghadiri, R Devon Hjelm, Yoshua Bengio, and Chris Pal. On adversarial mixup resynthesis. Advances in neural information processing systems, 32, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.368, + 0.472, + 0.452 + ], + "angle": 0, + "content": "[6] Ankan Kumar Bhunia, Salman Khan, Hisham Cholakkal, Rao Muhammad Anwer, Jorma Laaksonen, Mubarak Shah, and Fahad Shahbaz Khan. Person image synthesis via denoising diffusion model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5968-5976, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.452, + 0.472, + 0.494 + ], + "angle": 0, + "content": "[7] Andrew Brock, Jeff Donahue, and Karen Simonyan. Large scale gan training for high fidelity natural image synthesis. arXiv preprint arXiv:1809.11096, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.495, + 0.472, + 0.563 + ], + "angle": 0, + "content": "[8] Tim Brooks, Aleksander Holynski, and Alexei A Efros. Instructpix2pix: Learning to follow image editing instructions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18392-18402, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.565, + 0.472, + 0.62 + ], + "angle": 0, + "content": "[9] Kaidi Cao, Colin Wei, Adrien Gaidon, Nikos Arechiga, and Tengyu Ma. Learning imbalanced datasets with label-distribution-aware margin loss. Advances in neural information processing systems, 32, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.078, + 0.621, + 0.472, + 0.676 + ], + "angle": 0, + "content": "[10] Kaidi Cao, Colin Wei, Adrien Gaidon, Nikos Arechiga, and Tengyu Ma. Learning imbalanced datasets with label-distribution-aware margin loss. Advances in neural information processing systems, 32, 2019. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.078, + 0.677, + 0.472, + 0.718 + ], + "angle": 0, + "content": "[11] Olivier Chapelle, Jason Weston, Léon Bottou, and Vladimir Vapnik. Vicinal risk minimization. Advances in neural information processing systems, 13, 2000. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.078, + 0.719, + 0.472, + 0.761 + ], + "angle": 0, + "content": "[12] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in neural information processing systems, 34:8780-8794, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.078, + 0.761, + 0.472, + 0.844 + ], + "angle": 0, + "content": "[13] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.078, + 0.845, + 0.472, + 0.901 + ], + "angle": 0, + "content": "[14] M. Everingham, S. M. A. Eslami, L. Van Gool, C. K. I. Williams, J. Winn, and A. Zisserman. The pascal visual object classes challenge: A retrospective. International Journal of Computer Vision, 111(1):98-136, 2015. 1, 2, 5" + }, + { + "type": "list", + "bbox": [ + 0.078, + 0.115, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.161 + ], + "angle": 0, + "content": "[15] Rinon Gal, Yuval Alaluf, Yuval Atzmon, Or Patashnik, Amit H Bermano, Gal Chechik, and Daniel Cohen-Or. An image is worth one word: Personalizing text-to-image generation using textual inversion. arXiv preprint arXiv:2208.01618, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.162, + 0.894, + 0.218 + ], + "angle": 0, + "content": "[16] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. Advances in neural information processing systems, 27, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.218, + 0.894, + 0.274 + ], + "angle": 0, + "content": "[17] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. Advances in neural information processing systems, 27, 2014. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.274, + 0.894, + 0.329 + ], + "angle": 0, + "content": "[18] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.33, + 0.894, + 0.385 + ], + "angle": 0, + "content": "[19] Ruifei He, Shuyang Sun, Xin Yu, Chuhui Xue, Wenqing Zhang, Philip Torr, Song Bai, and Xiaojuan Qi. Is synthetic data from generative models ready for image recognition? arXiv preprint arXiv:2210.07574, 2022. 2, 6, 7, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.386, + 0.894, + 0.453 + ], + "angle": 0, + "content": "[20] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.455, + 0.894, + 0.497 + ], + "angle": 0, + "content": "[21] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.497, + 0.894, + 0.567 + ], + "angle": 0, + "content": "[22] Jonathan Ho, William Chan, Chitwan Sahara, Jay Whang, Ruiqi Gao, Alexey Gritsenko, Diederik P Kingma, Ben Poole, Mohammad Norouzi, David J Fleet, et al. Imagen video: High definition video generation with diffusion models. arXiv preprint arXiv:2210.02303, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.567, + 0.894, + 0.622 + ], + "angle": 0, + "content": "[23] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.622, + 0.894, + 0.691 + ], + "angle": 0, + "content": "[24] Phillip Isola, Jun-Yan Zhu, Tinghui Zhou, and Alexei A Efros. Image-to-image translation with conditional adversarial networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1125-1134, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.692, + 0.894, + 0.748 + ], + "angle": 0, + "content": "[25] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4401-4410, 2019. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.748, + 0.894, + 0.817 + ], + "angle": 0, + "content": "[26] Aditya Khosla, Nityananda Jayadevaprakash, Bangpeng Yao, and Li Fei-Fei. Novel dataset for fine-grained image categorization. In First Workshop on Fine-Grained Visual Categorization, IEEE Conference on Computer Vision and Pattern Recognition, Colorado Springs, CO, 2011. 6, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.818, + 0.894, + 0.873 + ], + "angle": 0, + "content": "[27] Jonathan Krause, Michael Stark, Jia Deng, and Li Fei-Fei. 3d object representations for fine-grained categorization. In Proceedings of the IEEE international conference on computer vision workshops, pages 554–561, 2013. 6, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.873, + 0.894, + 0.901 + ], + "angle": 0, + "content": "[28] Nupur Kumari, Bingliang Zhang, Richard Zhang, Eli Shechtman, and Jun-Yan Zhu. Multi-concept customization" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "17231" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.092, + 0.47, + 0.134 + ], + "angle": 0, + "content": "of text-to-image diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1931-1941, 2023. 2, 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.136, + 0.47, + 0.19 + ], + "angle": 0, + "content": "[29] Alexander C Li, Mihir Prabhudesai, Shivam Duggal, Ellis Brown, and Deepak Pathak. Your diffusion model is secretly a zero-shot classifier. arXiv preprint arXiv:2303.16203, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.192, + 0.47, + 0.234 + ], + "angle": 0, + "content": "[30] Yicong Li, Xiang Wang, Junbin Xiao, Wei Ji, and Tat seng Chua. Invariant grounding for video question answering. In CVPR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.236, + 0.47, + 0.277 + ], + "angle": 0, + "content": "[31] Shanchuan Lin, Bingchen Liu, Jiashi Li, and Xiao Yang. Common diffusion noise schedules and sample steps are flawed. arXiv preprint arXiv:2305.08891, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.278, + 0.47, + 0.348 + ], + "angle": 0, + "content": "[32] Ziwei Liu, Zhongqi Miao, Xiaohang Zhan, Jiayun Wang, Boqing Gong, and Stella X Yu. Large-scale long-tailed recognition in an open world. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2537-2546, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.349, + 0.47, + 0.417 + ], + "angle": 0, + "content": "[33] Cheng Lu, Yuhao Zhou, Fan Bao, Jianfei Chen, Chongxuan Li, and Jun Zhu. Dpm-solver: A fast ode solver for diffusion probabilistic model sampling in around 10 steps. Advances in Neural Information Processing Systems, 35:5775-5787, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.419, + 0.47, + 0.473 + ], + "angle": 0, + "content": "[34] Subhransu Maji, Esa Rahtu, Juho Kannala, Matthew Blaschko, and Andrea Vedaldi. Fine-grained visual classification of aircraft. arXiv preprint arXiv:1306.5151, 2013. 6, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.476, + 0.47, + 0.532 + ], + "angle": 0, + "content": "[35] Chenlin Meng, Yutong He, Yang Song, Jiaming Song, Jia-jun Wu, Jun-Yan Zhu, and Stefano Ermon. Sdedit: Guided image synthesis and editing with stochastic differential equations. arXiv preprint arXiv:2108.01073, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.533, + 0.47, + 0.575 + ], + "angle": 0, + "content": "[36] Rafael Müller, Simon Kornblith, and Geoffrey E Hinton. When does label smoothing help? Advances in neural information processing systems, 32, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.576, + 0.47, + 0.645 + ], + "angle": 0, + "content": "[37] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.646, + 0.47, + 0.702 + ], + "angle": 0, + "content": "[38] Maria-Elena Nilsback and Andrew Zisserman. Automated flower classification over a large number of classes. In 2008 Sixth Indian conference on computer vision, graphics & image processing, pages 722-729. IEEE, 2008. 6, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.703, + 0.47, + 0.785 + ], + "angle": 0, + "content": "[39] Seulki Park, Youngkyu Hong, Byeongho Heo, Sangdoo Yun, and Jin Young Choi. The majority can help the minority: Context-rich minority oversampling for long-tailed classification. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6887-6896, 2022. 6, 7, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.788, + 0.47, + 0.842 + ], + "angle": 0, + "content": "[40] Maitreya Patel, Tejas Gokhale, Chitta Baral, and Yezhou Yang. Conceptbed: Evaluating concept learning abilities of text-to-image diffusion models. arXiv preprint arXiv:2306.04695, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.845, + 0.47, + 0.871 + ], + "angle": 0, + "content": "[41] Judea Pearl. Causal inference in statistics: An overview. 2009. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.47, + 0.901 + ], + "angle": 0, + "content": "[42] Zeju Qiu, Weiyang Liu, Haiwen Feng, Yuxuan Xue, Yao Feng, Zhen Liu, Dan Zhang, Adrian Weller, and Bernhard" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.134 + ], + "angle": 0, + "content": "Schölkopf. Controlling text-to-image diffusion by orthogonal finetuning. Advances in Neural Information Processing Systems, 36, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.136, + 0.892, + 0.217 + ], + "angle": 0, + "content": "[43] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 5, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.219, + 0.892, + 0.259 + ], + "angle": 0, + "content": "[44] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.261, + 0.892, + 0.343 + ], + "angle": 0, + "content": "[45] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical Image Computing and Computer-Assisted Intervention-MICCAI 2015: 18th International Conference, Munich, Germany, October 5-9, 2015, Proceedings, Part III 18, pages 234-241. Springer, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.344, + 0.892, + 0.426 + ], + "angle": 0, + "content": "[46] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22500-22510, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.428, + 0.892, + 0.496 + ], + "angle": 0, + "content": "[47] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, et al. Imagenet large scale visual recognition challenge. International journal of computer vision, 115:211-252, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.498, + 0.892, + 0.564 + ], + "angle": 0, + "content": "[48] Shiori Sagawa, Pang Wei Koh, Tatsunori B Hashimoto, and Percy Liang. Distributionally robust neural networks for group shifts: On the importance of regularization for worst-case generalization. arXiv preprint arXiv:1911.08731, 2019. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.567, + 0.892, + 0.663 + ], + "angle": 0, + "content": "[49] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, Seyed Kamyar Seyed Ghasemipour, Burcu Karagol Ayan, S. Sara Mahdavi, Rapha Gontijo Lopes, Tim Salimans, Jonathan Ho, David J Fleet, and Mohammad Norouzi. Photorealistic text-to-image diffusion models with deep language understanding, 2022. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.665, + 0.892, + 0.72 + ], + "angle": 0, + "content": "[50] Dvir Samuel, Yuval Atzmon, and Gal Chechik. From generalized zero-shot learning to long-tail with class descriptors. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 286-295, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.721, + 0.892, + 0.775 + ], + "angle": 0, + "content": "[51] Veit Sandfort, Ke Yan, Perry J Pickhardt, and Ronald M Summers. Data augmentation using generative adversarial networks (cyclegan) to improve generalizability in ct segmentation tasks. Scientific reports, 9(1):16884, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.776, + 0.892, + 0.845 + ], + "angle": 0, + "content": "[52] Mert Bulent Sariyildiz, Karteek Alahari, Diane Larlus, and Yannis Kalantidis. Fake it till you make it: Learning transferable representations from synthetic imagenet clones. In CVPR 2023-IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.846, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[53] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "17232" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.093, + 0.468, + 0.12 + ], + "angle": 0, + "content": "next generation image-text models. Advances in Neural Information Processing Systems, 35:25278-25294, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.121, + 0.469, + 0.176 + ], + "angle": 0, + "content": "[54] Jordan Shipard, Arnold Wiliem, Kien Nguyen Thanh, Wei Xiang, and Clinton Fookes. Boosting zero-shot classification with synthetic data diversity via stable diffusion. arXiv preprint arXiv:2302.03298, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.177, + 0.469, + 0.244 + ], + "angle": 0, + "content": "[55] Uriel Singer, Adam Polyak, Thomas Hayes, Xi Yin, Jie An, Songyang Zhang, Qiyuan Hu, Harry Yang, Oron Ashual, Oran Gafni, et al. Make-a-video: Text-to-video generation without text-video data. arXiv preprint arXiv:2209.14792, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.246, + 0.469, + 0.314 + ], + "angle": 0, + "content": "[56] Shikun Sun, Longhui Wei, Zhicai Wang, Zixuan Wang, Junliang Xing, Jia Jia, and Qi Tian. Inner classifier-free guidance and its taylor expansion for diffusion models. In The Twelfth International Conference on Learning Representations, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.316, + 0.469, + 0.371 + ], + "angle": 0, + "content": "[57] Yonglong Tian, Lijie Fan, Phillip Isola, Huiwen Chang, and Dilip Krishnan. Stablerep: Synthetic images from text-to-image models make strong visual representation learners. arXiv preprint arXiv:2306.00984, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.372, + 0.469, + 0.426 + ], + "angle": 0, + "content": "[58] Brandon Trabucco, Kyle Doherty, Max Gurinas, and Ruslan Salakhutdinov. Effective data augmentation with diffusion models. arXiv preprint arXiv:2302.07944, 2023. 2, 5, 6, 7, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.428, + 0.469, + 0.496 + ], + "angle": 0, + "content": "[59] Narek Tumanyan, Michal Geyer, Shai Bagon, and Tali Dekel. Plug-and-play diffusion features for text-driven image-to-image translation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1921-1930, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.498, + 0.469, + 0.566 + ], + "angle": 0, + "content": "[60] Patrick von Platen, Suraj Patil, Anton Lozhkov, Pedro Cuenca, Nathan Lambert, Kashif Rasul, Mishig Davaadorj, and Thomas Wolf. Diffusers: State-of-the-art diffusion models. https://github.com/huggingface/ diffusers, 2022.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.567, + 0.469, + 0.607 + ], + "angle": 0, + "content": "[61] Catherine Wah, Steve Branson, Peter Welinder, Pietro Perona, and Serge Belongie. The caltech-ucsd birds-200-2011 dataset. 2011. 5, 6, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.609, + 0.47, + 0.678 + ], + "angle": 0, + "content": "[62] Zhicai Wang, Yanbin Hao, Tingting Mu, Ouxiang Li, Shuo Wang, and Xiangnan He. Bi-directional distribution alignment for transductive zero-shot learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19893-19902, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.679, + 0.469, + 0.734 + ], + "angle": 0, + "content": "[63] Daniel Watson, William Chan, Ricardo Martin-Brualla, Jonathan Ho, Andrea Tagliasacchi, and Mohammad Norouzi. Novel view synthesis with diffusion models. arXiv preprint arXiv:2210.04628, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.735, + 0.469, + 0.802 + ], + "angle": 0, + "content": "[64] Zongze Wu, Dani Lischinski, and Eli Shechtman. Stylespace analysis: Disentangled controls for stylegan image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12863-12872, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.804, + 0.469, + 0.859 + ], + "angle": 0, + "content": "[65] Moon Ye-Bin, Nam Hyeon-Woo, Wonseok Choi, Nayeong Kim, Suha Kwak, and Tae-Hyun Oh. Exploiting synthetic data for data imbalance problems: Baselines from a data perspective. arXiv preprint arXiv:2308.00994, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.86, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[66] Sangdoo Yun, Dongyoon Han, Seong Joon Oh, Sanghyuk Chun, Junsuk Choe, and Youngjoon Yoo. Cutmix: Regularization strategy to train strong classifiers with localizable" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.093, + 0.47, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "features. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6023-6032, 2019. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.122, + 0.892, + 0.163 + ], + "angle": 0, + "content": "[67] An Zhang, Wenchang Ma, Xiang Wang, and Tat seng Chua. Incorporating bias-aware margins into contrastive loss for collaborative filtering. In NeurlPS, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.165, + 0.892, + 0.207 + ], + "angle": 0, + "content": "[68] Hongyi Zhang, Moustapha Cisse, Yann N Dauphin, and David Lopez-Paz. mixup: Beyond empirical risk minimization. arXiv preprint arXiv:1710.09412, 2017. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.208, + 0.892, + 0.248 + ], + "angle": 0, + "content": "[69] Chenyu Zheng, Guoqiang Wu, and Chongxuan Li. Toward understanding generative data augmentation. Advances in Neural Information Processing Systems, 36, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.25, + 0.892, + 0.305 + ], + "angle": 0, + "content": "[70] Bolei Zhou, Agata Lapedriza, Aditya Khosla, Aude Oliva, and Antonio Torralba. Places: A 10 million image database for scene recognition. IEEE transactions on pattern analysis and machine intelligence, 40(6):1452-1464, 2017. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.307, + 0.892, + 0.375 + ], + "angle": 0, + "content": "[71] Jun-Yan Zhu, Taesung Park, Phillip Isola, and Alexei A Efros. Unpaired image-to-image translation using cycle-consistent adversarial networks. In Proceedings of the IEEE international conference on computer vision, pages 2223-2232, 2017. 3" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.375 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "17233" + } + ] +] \ No newline at end of file diff --git a/2024/Enhance Image Classification via Inter-Class Image Mixup with Diffusion Model/d24eac86-8eb2-4146-81d8-2b9455b8e75f_origin.pdf b/2024/Enhance Image Classification via Inter-Class Image Mixup with Diffusion Model/d24eac86-8eb2-4146-81d8-2b9455b8e75f_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..e9fd91e3ae5cfa78d17673978d5abe3c8b2de159 --- /dev/null +++ b/2024/Enhance Image Classification via Inter-Class Image Mixup with Diffusion Model/d24eac86-8eb2-4146-81d8-2b9455b8e75f_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9e5295141b77adb6f78532e347284e28ded049eaf54afe95fb1215487c10c02f +size 3317225 diff --git a/2024/Enhance Image Classification via Inter-Class Image Mixup with Diffusion Model/full.md b/2024/Enhance Image Classification via Inter-Class Image Mixup with Diffusion Model/full.md new file mode 100644 index 0000000000000000000000000000000000000000..f9a6a8cf728de862541491e55e299a86cb2ecb01 --- /dev/null +++ b/2024/Enhance Image Classification via Inter-Class Image Mixup with Diffusion Model/full.md @@ -0,0 +1,331 @@ +# Enhance Image Classification via Inter-Class Image Mixup with Diffusion Model + +Zhicai Wang $^{1}$ ; Longhui Wei $^{2\dagger}$ , Tan Wang $^{3}$ , Heyu Chen $^{1}$ , Yanbin Hao $^{1}$ , Xiang Wang $^{1}$ , Xiangnan He $^{1}$ , Qi Tian $^{2}$ + +1University of Science and Technology of China, 2 Huawei Inc., + +$^{3}$ Nanyang Technological University + +# Abstract + +Text-to-image (T2I) generative models have recently emerged as a powerful tool, enabling the creation of photorealistic images and giving rise to a multitude of applications. However, the effective integration of T2I models into fundamental image classification tasks remains an open question. A prevalent strategy to bolster image classification performance is through augmenting the training set with synthetic images generated by T2I models. In this study, we scrutinize the shortcomings of both current generative and conventional data augmentation techniques. Our analysis reveals that these methods struggle to produce images that are both faithful (in terms of foreground objects) and diverse (in terms of background contexts) for domain-specific concepts. To tackle this challenge, we introduce an innovative inter-class data augmentation method known as Diff-Mix1, which enriches the dataset by performing image translations between classes. Our empirical results demonstrate that Diff-Mix achieves a better balance between faithfulness and diversity, leading to a marked improvement in performance across diverse image classification scenarios, including few-shot, conventional, and long-tail classifications for domain-specific datasets. + +# 1. Introduction + +In comparison to GAN-based models [7, 17, 25], contemporary state-of-the-art text-to-image (T2I) diffusion models exhibit enhanced capabilities in producing high-fidelity images [12, 37, 44, 49]. With the remarkable cross-modality alignment capabilities of T2I models, there is significant potential for generative techniques to enhance image classification [2, 4]. For instance, a straightforward approach entails augmenting the existing training dataset with synthetic images generated by feeding categorical textual prompts to a T2I diffusion model. However, upon reviewing prior ap + +![](images/7ac762f63f12a6572d6b598f0a9e1851bd95d9b404ad63da1bc377f039efb4e6.jpg) +Figure 1. Strategies to expand domain-specific datasets for improved classification are varied. Row 1 illustrates vanilla distillation from a pretrained text-to-image (T2I) model, which carries the risk of generating outputs with reduced faithfulness. Intra-class augmentation, depicted in Row 2, tends to yield samples with limited diversity to maintain high fidelity to the original class. Our proposed method, showcased in Rows 3 and 4, adopts an inter-class augmentation strategy. This involves introducing edits to a reference image using images from other classes within the training set, which significantly enriches the dataset with a greater diversity of samples. + +proaches employing T2I diffusion models for image classification, it becomes evident that the challenge in generative data augmentation for domain-specific datasets is producing samples with both a faithful foreground and a diverse background. Depending on whether a reference image is used in the generative process, we divide these methods into two groups: + +- Text-guided knowledge distillation [52, 57] involves generating new images from scratch using category-related prompts to expand the dataset. For the off-the-shelf T2I models, such vanilla distillation presume these models have comprehensive knowledge of target domain, which can be problematic for domain-specific datasets. Insufficient domain knowledge easily makes the distillation + +process less effective. For example, vanilla T2I models struggle to generate images that accurately represent specific bird species based solely on their names (see Row 1 of Fig. 1). + +- Generative data augmentation [1, 69] employs generative models to enhance existing images. Da-fusion [58], for instance, translates the source image into multiple edited versions within the same class. This strategy, termed intra-class augmentation, primarily introduces intra-class variations. While intra-class augmentation retains much of the original image's layout and visual details, it results in limited background diversity (see Row 2 of Fig. 1). However, synthetic images with constrained diversity may not sufficiently enhance the model's ability to discern foreground concepts. + +Based on these observations, a fundamental question emerges: 'Is it feasible to develop a method that optimizes both the diversity and faithfulness of synthesized data simultaneously?' + +In this work, we introduce Diff-Mix, a simple yet effective data augmentation method that harnesses diffusion models to perform inter-class image interpolation, tailored for enhancing domain-specific datasets. The method encompasses two pivotal operations: personalized fine-tuning and inter-class image translation. Personalized fine-tuning [15, 46] is originally designed for customizing T2I models and enabling them to generate user-specific contents or styles. In our case, we implement the technique to tailor the model, enabling it to generate images with faithful foreground concepts. Inter-class image translation in Diff-Mix entails transforming a reference image into an edited version that incorporates prompts from different classes. This translation strategy is designed to retain the original background context while editing the foreground to align with the target concept. For instance, as depicted in the bottom rows of Fig. 1, Diff-Mix can generate images of land birds in diverse settings, such as maritime environments, enriching the dataset with a variety of counterfactual samples. + +Unlike previous non-generative augmentation methods, such as Mixup [68] and CutMix [66], Diff-Mix works in a foreground-perceivable inter-class interpolation manner and shares a different mechanism with the non-generative approaches. Our experiment under the conventional classification setting indicates that incorporating both CutMix and Diff-Mix could further enhance performance. Additionally, when compared with other generative approaches, we conduct experiments under few-shot and long-tail scenarios and observe consistent performance improvements. Our contributions can be summarized as follows: + +- We pinpoint the critical factors that affect the efficacy of generative data augmentation in domain-specific image classification: namely, faithfulness and diversity. +- We introduce Diff-Mix, a simple yet effective generative + +data augmentation strategy that leverages fine-tuned diffusion models for inter-class image interpolation. + +- We conduct a comparative analysis of Diff-Mix with other distillation-based and intra-class augmentation methods, as well as non-generative approaches, highlighting its unique features and benefits. + +# 2. Related Works + +Text-to-image diffusion models. Following pretraining on web-scale data, the T2I diffusion model has demonstrated robust capabilities in generating text-controlled images [37, 49, 53, 56]. Its versatility has led to diverse applications, including novel view synthesis [6, 63], concept learning [28, 46], and text-to-video generation [22, 55], among others. Recent advancements [29] also highlight the cross-modality features of such generative models, showcasing their ability to serve as zero-shot classifiers. + +Synthetic data for image classification. There are two perspectives on the utilization of synthetic data for image classification: knowledge distillation [2] and data augmentation [5, 51, 54, 62]. From the knowledge distillation perspective, SyntheticData [19] reports significant performance gains in both zero-shot and few-shot settings by leveraging off-the-shelf T2I models to obtain synthetic data. The work of [2] has indicated that fine-tuning the T2I model on ImageNet [47] yields improved classification accuracy by narrowing the domain gap. Some works also find that learning from the synthetic data presents strong transferability [19, 57] and robustness [4, 30, 67]. From the data augmentation standpoint, Da-fusion [58] achieves stable performance improvements on few-shot datasets by augmenting from reference images. In a related study [3], the use of StyleGAN [25] for generating interpolated images between two different domains has been shown to enhance classifier robustness for out-of-distribution data. Our work shares similarities with AMR [5], which generates realistic novel examples by interpolating between two images using GAN [16]. The distinction lies in our discussion of interpolation using the T2I diffusion model, where its noise-adding and denoising characteristics enable a smoother implementation of interpolation. + +Non-generative data augmentation. Mixup [68] and CutMix [66] stand out as two prominent non-generative data augmentation methods, serving as effective regularization techniques during training. While Mixup achieves augmented samples through a convex combination of two images, CutMix achieves augmentation by cutting and pasting parts of images. However, both methods are constrained in their ability to produce realistic images. In addressing this limitation, the utilization of generative models emerges as a potential solution to alleviate this issue. + +# 3. Method + +# 3.1. Preliminary + +Text-to-Image diffusion model. Diffusion models generate images by gradually removing noise from a Gaussian noise source [21]. In a diffusion process with a total of $T$ steps, its forward process, which gradually adds noise, is represented as a Markov chain with a Gaussian transition kernel, where $q(\pmb{x}_t|\pmb{x}_{t - 1}) = \mathcal{N}(\pmb{x}_t;\sqrt{\alpha_t}\pmb{x}_{t - 1},(1 - \alpha_t)\mathbf{I})$ , where $x_{t}$ represents the noisy image at step $t$ . The training objective at step $t$ is to predict the noise to reconstruct $\pmb{x}_{t - 1}$ . When training a text-conditioned diffusion model, the simplified training objective can be summarized as follows: + +$$ +\mathbb {E} _ {\epsilon , \mathbf {X}, \mathbf {c}, t} \left[ \| \epsilon - \epsilon_ {\theta} (\mathbf {x} _ {t}, \mathbf {c}, t) \| _ {2} ^ {2} \right], \tag {1} +$$ + +where $\epsilon_{\theta}$ represents the predicted noise, and $c$ is the encoded text caption associated with the image $\pmb{x}$ . + +T2I personalization. T2I personalization aims to personalize a diffusion model for generating specific concepts using a limited number of concept-oriented images [28, 40, 42]. These concepts are typically represented using identifiers (e.g., “[V]”). As a result, we formalize the constructed image-caption set as $\pmb{x}$ , “photo of a [V]”. Various personalization methods differ in their fine-tuning strategies. For instance, Textual Inversion (TI) [15] makes the identifier learnable, but other modules are not fine-tuned, potentially sacrificing some faithfulness in image generation. On the other hand, Dreambooth (DB) [46] fine-tunes the U-Net [45] for more refined personalized generation but faces the challenge of increased computational cost. + +Image-to-image translation. Image translation enables image synthesis and editing using a reference image as guidance [24, 64, 71]. Diffusion-based image translation methods can generate fine edits, which refer to subtle modifications, with varying degrees of shift relative to the reference image [8, 35, 59]. Here, we draw inspiration from SDEdit [35] to perform edits on the reference image, where the target image $\mathbf{x}^{\mathrm{tar}}$ is translated from a reference image $\mathbf{x}^{\mathrm{ref}}$ . During translation, the reverse process does not traverse the full process but starts from a certain step $\lfloor sT \rfloor$ , where $s \in [0,1]$ controls the insertion position of the reference image with noise, as follows, + +$$ +\boldsymbol {x} _ {\lfloor s T \rfloor} = \sqrt {\tilde {\alpha} _ {\lfloor s T \rfloor}} \boldsymbol {x} _ {0} ^ {\text {r e f}} + \sqrt {1 - \tilde {\alpha} _ {\lfloor s T \rfloor}} \epsilon . \tag {2} +$$ + +By adjusting the strength parameter $s$ , one can strike a balance between the diversity of the generated images and their faithfulness to the reference image. + +# 3.2. General Framework + +The Diff-Mix pipeline consists of two key steps. Firstly, to produce more faithful images for domain-specific datasets, + +![](images/679e9849bd49aaf0cc0d70513702151847cb42d3e435a39c2e741616576fd03e.jpg) +(a) Real images + +![](images/a835499778bd5de0d3f981d224126e4e499c62f5672a18da7d0a8988636b366b.jpg) + +![](images/bdb3d0dfa6ae058b391cc0d7a8e02a99ee4653596d741a468d8049886392d492.jpg) +(b) Synthetic images (vanilla SD) + +![](images/77de4e4343c7bac68f53f7e49ade331e8f701868343060f55a09a66404d3c9a9.jpg) +(c) Synthetic images (SD fine-tuned via DB) +(d) Synthetic images (SD fine-tuned via $\mathrm{TI + DB}$ ) +Figure 2. Examples of "American Three toed Woodpecker". (a) Real images from the training set. (b-d) synthetic images generated using different fine-tuned models with the same number of fine-tuning steps. TI+DB indicates both text embedding and U-Net are fine-tuned. TI+DB achieves a more faithful output compared to DB alone (check the head and wing patterns of the birds). + +we propose treating it as a T2I personalization problem and fine-tuning the Stable Diffusion (SD). Subsequently, to enhance the diversity of synthetic data beyond the well-fitted training distribution, we employ inter-class image translation. This process produces interpolated images with increased background diversity for each class. + +# 3.3. Fine-tune Diffusion Model + +Vanilla distillation tends to be less effective, especially as the number of training shots increases (refer to Sec. 4.1). In order to mitigate the distribution gap, we propose finetuning Stable Diffusion in conjunction with current widely-used T2I personalization strategies. + +Dreambooth meets Textual Inversion. Many fine-grained datasets provide terminological names for their categories, like "American Three toed Woodpecker" and "Pileated Woodpecker". We could construct image-text pairs using category-related prompts and fine-tune the denoising network of SD using Eq. 1, which is analogous to Dreambooth. However, we observe that directly incorporating these specialized terms into the text during fine-tuning can impede convergence and hinder the generation of faithful images. We attribute this challenge to the semantic proximity of terminology within a fine-grained domain, where fine-tuning the vision module alone tends to be less effective at distinguishing two similar classes within the same family, like "Woodpecker". Inspired by Textual Inversion [15], we opt to replace the terminological name in the dataset with "[Vi] [metaclass]" where "[Vi"] is a learnable identifier, and $i$ varies from 1 to $N$ , represent- + +![](images/2183af4399b8471a3a3a3b7189d897524e8e43fdc30709817cab54d8704ae83d.jpg) +Figure 3. Fine-tuning framework of Diff-Mix operates as follows: Initially, we replace the class name with a structured identifier formatted as “[ $\mathsf{V}^i$ ” [metaclass]}, thereby sidestepping the need for specific terminological expressions. Next, we engage in joint fine-tuning of these identifiers and the low-rank residues (LoRA) of U-Net to capture the domain-specific distribution. + +ing the category index. The illustration of our fine-tuning strategy is presented in Fig. 3. The term “[metaclass]” is determined by the theme of the current dataset, such as “bird” for a fine-grained bird dataset. By concurrently fine-tuning the identifier and the U-Net, we empower the model to quickly adapting to the fine-grained domain, allowing it to generate faithful images using the identifier (see comparison between Row 3 and Row 4 in Fig. 2). + +Parameter efficient fine-tuning. In this context, we embrace the parameter-efficient fine-tuning strategy known as LoRA [23]. LoRA distinguishes itself by fine-tuning the residue of low-rank matrices instead of directly fine-tuning the pre-trained weight matrices. To elaborate, consider a weight matrix $\mathbf{W} \in \mathbb{R}^{m \times n}$ . The tunable residual matrix $\Delta \mathbf{W}$ comprises two low-rank matrices: $\mathbf{A} \in \mathbb{R}^{m \times d}$ and $\mathbf{B} \in \mathbb{R}^{n \times d}$ , defined as $\Delta \mathbf{W} = \mathbf{A}\mathbf{B}^{\top}$ . As a default configuration, we set the rank $d$ to 10. + +# 3.4. Data Synthesis Using Diffusion Model + +In generating pseudo data, three strategies can be used with our fine-tuned diffusion model $^{2}$ : (1) distillation-based method Diff-Gen, (2) intra-class augmentation Diff-Aug, and (3) inter-class augmentation Diff-Mix. + +Diff-Gen and Diff-Aug. For a target class $y^i$ and its textual condition, "photo of a $[V^i]$ [metaclass]", both + +![](images/cbdd6c90118c3b11ceb3757f7910278802c1de706b8a7200f87846bfc29ff2dc.jpg) +Figure 4. Examples of images translated using Diff-Mix and Diff-Aug across various strengths. Diff-Aug employs the same target and reference image classes, typically resulting in subtle modifications. Diff-Mix progressively adjusts the foreground to align with the target class as the translation strength increases, while preserving the background layout from the reference image. + +methods generate synthetic samples annotated with class $i$ . Specifically, Diff-Gen generates samples from scratch by initializing with random Gaussian noise and proceeding through the full reverse diffusion process with $T$ steps. Diff-Gen can produce images aligned with its fine-tuned distribution. In contrast, Diff-Aug sacrifices a portion of diversity and generate images by editing on a reference image. Specifically, it randomly sample a image from the intra-class training set and enhances the image through image translation using Eq. 2. The term "intra-class" means that the conditioning prompts are constructed based on the ground truth categories of images, and such a denoising process tends to introduce less variation, particularly for the foreground concepts (see top rows of Fig. 1). + +Diff-Mix. Diff-Mix employs the same translation process as Diff-Aug, but the reference image is sampled from the full training set rather than intra-class set to enable interclass interpolation. The key difference is that Diff-Mix can generate numerous counterfactual examples, such as a blackbird in the sea (see fourth row of Fig. 1). This necessitates that downstream models make a more refined differentiation of category attributes, thereby reducing the impact of spurious correlations introduced by variations in the background. Denoting the label of the reference image as $y^{j}$ , by inserting the reference image into the reverse process with the prompt "photo of a $[V^i]$ [metaclass]", we can obtain interpolated images between the $i_{\mathrm{th}}$ and $j_{\mathrm{th}}$ categories. By controlling the intensity $s$ , we can precisely + +![](images/568a687fbd510c9c8c70ae5df276f80603d88666330070d2558eecaa5f929a56.jpg) +Figure 5. A schematic explanation of Diff-Mix's effectiveness using structural casual model [41]. $\pmb{x}_{fg}$ is the foreground that determines the real class label, $\pmb{x}_{bg}$ denotes the background. $\pmb{x}_{fg} \rightarrow \pmb{z} \rightarrow y$ is the causal path that we are focusing and $\pmb{x}_{fg} \gets I \rightarrow \pmb{x}_{bg} \rightarrow \pmb{z} \rightarrow y$ is the backdoor path that introduces spurious relations between $\pmb{x}_{fg}$ and $y$ . + +![](images/d4cd3fa5cd7d74df00c0a00006795334db3edf919efef18c0fd57749cd216e2a.jpg) + +manage the interpolation process. When annotating the synthetic image, unlike Mixup and Cutmix, we take into account the non-linear nature of diffusion translation. Thus the annotation function is given by + +$$ +\tilde {y} = (1 - s ^ {\gamma}) y ^ {i} + s ^ {\gamma} y ^ {j}, \tag {3} +$$ + +where $\gamma$ is a hyperparameter introducing non-linearity. Our empirical findings indicate that a $\gamma$ smaller than 1 is favored. Additionally, in low-shot cases, the samples with higher confidence in the target class are preferred (see details in Sec. 4.5). + +Construct synthetic dataset. To construct the synthetic dataset using Diff-Mix, similar to Da-fusion [58], we adopt a randomized sampling strategy ( $s \in \{0.5, 0.7, 0.9\}$ ) for the selection of translation strength. While applying the inter-class editing, we observe that Diff-Mix tends to produce more undesirable samples compared to Diff-Aug. These undesirable samples have incomplete foreground such as fragmented bird bodies. This is caused by the intrinsic shape and pose differences among classes. To mitigate this, we introduce a simple data-cleaning approach to reduce the proportion of such problematic images. We utilize the large vision language model CLIP [43] to assess the confidence in the content, serving as the filtering criterion. Further details can be found in the supplementary materials (SMs). + +Analysis. We depict the core insight of Diff-Mix in Fig. 5. To eliminate the spurious correlation introduced by $\pmb{x}_{bg}$ , learning on the synthetic set with randomized $\pmb{x}_{bg}$ (background) can cut off spurious correlation, forcing the classification model to infer only from the foreground. The study in Fig. 7 (b) shows that the more diverse the background (larger the referable class number), the better the performance on the CUB test set. + +# 4. Experiments + +In this section, we investigate the effectiveness of Diff-Mix in domain-specific datasets. The key questions we aim to address are as follows: + +Q1: Can generative inter-class augmentation lead to more significant performance gains in downstream tasks compared to those intra-class augmentation methods and distillation-based methods? +Q2: Is improved background diversity the secret weapon of Diff-Mix for enhancing the performance? +Q3: How to choose the fine-tuning strategy and annotation strategy to boost the performance gains for the inter-class augmentation? + +To address Q1, we separately discuss these questions in few-shot settings in Sec. 4.1, conventional classification in Sec. 4.2, as well as long-tail classification in Sec. 4.3. Additionally, to answer Q2, we conduct a test for background robustness in Sec. 4.4 and perform an ablation study focusing on the size and diversity of synthetic data in Sec. 4.5. For Q3, we conduct an ablation study to empirically discover effective strategies for deployment in Sec. 4.5. + +# 4.1. Few-shot Classification + +Experimental Setting. To investigate the impact of different data expansion methods, we conduct few-shot experiments on a domain-specific dataset Caltech-UCSD Birds (CUB) [61], with shot numbers of 1, 5, 10, and all. For augmentation-based methods, the synthetic dataset is constructed using various translation strengths $(s)$ , specifically, $s \in \{0.1, 0.3, \dots, 1.0\}$ . We expand the original training set with a multiplier of 5 and cache the synthesized dataset locally for joint training. Real data are replaced by synthetic data proportionally during training, and the replacement probability $p$ is set as 0.1, 0.2, 0.3, and 0.5 for all-shot, 10-shot, 5-shot, and 1-shot classification, respectively. All experiments use ResNet50 with an input resolution of $224 \times 224$ . Additional details can be found in the SMs. + +Comparison methods. To unveil the trade-off between faithfulness and diversity resulting from different expansion strategies, we compared Diff-Mix with Diff-Gen and Diff-Aug. Furthermore, we conduct experiments on expansion strategies using vanilla SD: Real-Mix, Real-Gen, and Real-Aug, where 'Real' signifies that SD is not fine-tuned. + +Main results. To answer Q1 under the few-shot classification setting, we augment CUB using X-Mix, X-Aug, and X-Gen, where 'X' denotes 'Diff/Real' for simplicity. The results are shown in Fig. 6, and we can observe that: + +1. Diff-Mix generally outperforms the intra-class competitor X-Aug and distillation competitor X-Gen in various few-shot scenarios. It tends to achieve higher gains when the strength $s$ is relatively large, i.e., $\{0.5, 0.7, 0.9\}$ , where the foreground has been edited to match the target class and the background retains similarities to the reference image. +2. Among the Real-X methods, distillation tends to be more effective than the augmentation method when the shot + +![](images/a5726da6ec202f288638f381edd25aa2f9b847fb756e9a7865d7d32a5d6b4b6e.jpg) +(a) 1-shot + +![](images/6086dd31f4fd8466d946c30de710cb0dc3a65677e02b903931d3625a5b80c271.jpg) +(b) 5-shot + +![](images/610cb3645a14a320ac5979460207d35c8a49303b34cf50e08381810c67d9b1f3.jpg) +(c) 10-shot + +![](images/f79c46a60ad004f140e8205d114bd2af63b11e8700394e0fc5355e5294c44799.jpg) +(d) All-shot +Figure 6. Few-shot classification results on CUB. + +number is low, but the trend reverses as the shot number increases (compare Real-Gen with Real-Aug). Real-Gen's samples even show less effective than Real-Aug $(s = 1.0)$ under the all-shot case3. This indicates that the importance of faithfulness in the trade-off between faithfulness and diversity increases with the shot number. Additionally, Real-Mix exhibits consistent and stable improvement over the other two methods. + +3. Diff-Gen consistently outperforms Real-Gen under four scenarios. Notably, Real-Gen's performance declines below that of the baseline as the shot numbers reach 10, showcasing the importance of the fine-tuning process which increases the faithfulness of synthetic samples. + +# 4.2. Conventional Classification + +Experimental setting. To test whether Diff-Mix can further boost performance in a more challenging setting, i.e., under the all-shot scenario with high input resolution, we conduct conventional classification on five domain-specific datasets: CUB [61], Stanford Cars [27], Oxford Flowers [38], Stanford Dogs [26], and FGVC Aircraft [34]. Two backbones are employed: pretrained (ImageNet1K [49]) ResNet50 [18] with input resolution $448 \times 448$ , and pretrained (ImageNet21K) ViT-B/16 [13] with input resolution $384 \times 384$ . Label smoothing [36] is applied across all datasets with a confidence level of 0.9. For all expansion strategies, the expansion multiplier is 5 and the replacement probability $p$ is 0.1. Besides, we use a randomized sampling strategy $(s \in \{0.5, 0.7, 0.9\})$ and a fixed $\gamma$ (0.5, and this is specific to Diff-Mix) for all datasets. + +Comparison methods. We compare Diff-Mix with (1) Real-Filtering (RF) [19], a variation of Real-Gen that incorporates clip filtering, (2) Real-Guidance (RG) [19], which augments the dataset using intra-class image translation at low strength ( $s = 0.1$ ), (3) Da-Fusion [58], a method that solely fine-tunes the identifier to personalize each class and employs randomized sampling strategy + +$(s \in \{0.25, 0.5, 0.75, 1.0\})$ , and non-generative augmentation methods (4) CutMix [66] and (5) Mixup [68]. + +Main results. We show the classification accuracy for different data expansion strategies in Table 1, our observations can be summarized: (1) Diff-Mix consistently demonstrates stable improvements across the majority of settings. Its average performance gain across the five datasets exceeds that of baselines employing intra-class augmentation methods (RG and Da-fusion), distillation method (RF), and non-generative data augmentation techniques (CutMix and Mixup). (2) Real-filtering, analogous to the discussion of Real-Gen, exhibits performance degradation on most datasets due to the distribution gap. (3) The combined use of Diff-Mix and CutMix often yields better performance gains. This is attributed to the distinct enhancement mechanisms of the two methods, i.e., vicinal risk minimization [11, 68] and foreground-background disentanglement. (4) Diff-Mix does not exhibit significant performance improvement in the dog dataset. We attribute this lack of improvement to the complexity of the dog dataset, which often contains multiple subjects in a single image, impeding effective foreground editing (refer to the SMs for visual examples). + +# 4.3. Long-Tail Classification + +Experiment setting. Following the settings of previous long-tail dataset constructions [9, 32, 39], we create two domain-specific long-tail datasets, CUB-LT [50] and Flower-LT. The imbalance factor controls the exponential distribution of the imbalanced data, where a larger value indicates a more imbalanced distribution. To leverage generative models for long-tail classification, we adopt the approach of SYNAuG [65], which uniformize the imbalanced real data distribution using synthetic data. Translation strength $s$ (0.7) and $\gamma$ (0.5) are fixed for both Diff-Mix and Real-Mix. + +Comparison methods. We compare Diff-Mix with Real-Mix, Real-Gen, the non-generative CutMix-based oversampling approach CMO [39], and its enhanced variant with two-stage deferred re-weighting [9] (CMO+DRW). + +Main results. We present the long-tail classification re + +
BackboneAug. MethodFT StrategyDataset
CUBAircraftFlowerCarDogAvg
ResNet50@448--86.6489.0999.2794.5487.4891.40
Cutmix[66]-87.2389.4499.2594.7387.5991.65+0.25
Mixup[68]-86.6889.4199.4094.4987.4291.48+0.08
Real-filtering [19]†X85.6088.5499.0994.5987.3091.22-0.18
Real-guidance [19]†X86.7189.0799.2594.5587.4091.59+0.19
Da-fusion [58]†TI86.3087.6499.3794.6987.3391.07-0.58
Diff-MixTI + DB87.1690.2599.5495.1287.7491.96+0.56
Diff-Mix + CutmixTI + DB87.5690.0199.4795.2187.8992.03+0.63
ViT-B/16@384--89.3783.5099.5694.2192.0691.74
Cutmix[66]-90.5283.5099.6494.8392.1392.12+0.38
Mixup[68]-90.3284.3199.7394.9892.0292.27+0.53
Real-filtering [19]†X89.4983.0799.3694.6691.9191.69-0.05
Real-guidance [19]†X89.5483.1799.5994.6592.0591.80+0.06
Da-fusion [58]†TI89.4081.8899.6194.5392.0791.50-0.24
Diff-MixTI + DB90.0584.3399.6495.0991.9992.22+0.48
Diff-Mix + CutmixTI + DB90.3585.1299.6895.2691.8992.46+0.72
+ +Table 1. Conventional classification in six fine-grained datasets. ${}^{ \dagger }$ ,indicates our reproduced results using SD. + +
MethodIF=1005010
ManyMediumFewAll
CE79.1164.2813.4833.6544.8258.13
CMO [39]78.3258.5714.7832.9444.0857.62
CMO + DRW [10]78.9756.3614.6632.5746.4359.25
Real-Gen84.8865.2330.6845.8653.4361.42
Real-Mix (s=0.7)84.6366.3434.4447.7555.6762.27
Diff-Mix (s=0.7)84.0767.7936.5550.3558.1964.48
+ +Table 2. Long-tail classification in CUB-LT. + +
MethodIF=1005010
ManyMediumFewAll
CE99.1994.9558.1880.4390.8795.07
CMO [39]99.2595.1967.4583.9591.4395.19
CMO+ DRW [10]99.9795.0667.3184.0792.0695.92
Real-Gen98.6495.5566.1083.5691.8495.22
Real-Mix (s=0.7)99.8796.2668.5385.1992.9696.04
Diff-Mix (s=0.7)99.2596.9878.4189.4693.8696.63
+ +sults for CUB-LT in Table 2 and Flower-LT in 3. The observations are as follows: (1) Generative approaches exhibit superior performance in tackling imbalanced classification issues compared to CutMix-based methods (CMO and CMO+DRW). (2) Real-Mix surpasses Real-Gen in performance across various imbalance factors in two datasets. This indicates that tail classes can benefit from the enhanced diversity by leveraging the visual context of majority classes. (3) Diff-Mix generally achieves the best performance among the compared strategies, especially at the low-shot case, highlighting the importance of fine-tuning. + +# 4.4. Background Robustness + +In this section, we aim to evaluate Diff-Mix's robustness to background shifts, specifically, whether synthesizing more diverse samples can improve the classification model's generalizability when the background is altered. To achieve + +Table 3. Long-tail classification in Flower-LT. + +
GroupBase.CutMixDA-fusionDiff-AugDiff-Mix
(waterbird, water)59.5062.4660.9061.8363.83
(waterbird, land)56.7060.1258.1060.1263.24
(landbird, land)73.4873.3972.9473.0475.64
(landbird, water)73.9774.7272.7773.5274.36
Avg.70.1971.2369.9070.2872.47
+ +Table 4. Classification results across four groups in Waterbird [48]. Waterbird is an out-of-distribution dataset for CUB, crafted by segmenting CUB's foregrounds and paste them into the scene images from Places [70]. The constructed dataset can be divided into four groups based on the composition of foregrounds (waterbird and landbird) and backgrounds (water and land). + +this, we utilize an out-of-distribution test set for CUB, namely Waterbird [48]. We then perform inference on the whole Waterbird set using classifiers that have been trained on either the original CUB dataset or its expanded variations. In Table 4, we present the classification accuracies across the four groups and compare Diff-Mix with other intra-class methods (Da-fusion and Diff-Aug) as well as CutMix. We observe that Diff-Mix generally outperforms its counterparts and achieves a significant performance improvement $(+6.5\%)$ in the challenging counterfactual group (waterbirds with land backgrounds). It is important to highlight that the background scenes in the Waterbird dataset are novel to CUB, requiring the classification model to have a stronger perceptual capability for the images' foregrounds. + +# 4.5. Discussion + +In this section, we address Q2 by examining the impact of size and diversity on synthetic data. Furthermore, we perform an ablation study to assess the effects of fine-tuning strategies and training hyperparameters of Diff-Mix, which is aimed at answering Q3. Unless specified otherwise, our discussions are based on experiments conducted using CUB + +![](images/7e18fcca6a05d2485e7b5d5a97434bd0a227da1949fc4d50e9498d6edd65f57c.jpg) +(a) + +![](images/ad491210bb609c2744334dabc3e1cdbb16868909a409daa40df26f34ebb4abef.jpg) +(b) +Figure 7. Comparison of results across various (a) synthetic data sizes and (b) numbers of referable classes for each target class. + +with ResNet50, where inputs are resized to $224 \times 224$ . + +Size and diversity of synthetic Data. The relationship between performance gain and the size of synthetic data is depicted in Fig. 7 (a), where a classification model was trained with synthetic data of varying sizes. We observe a monotonically increasing trend as the multiplier for the synthetic dataset ranges from 2 to 10. Ideally, the combination choices of $(x_{i},y_{j})$ are in the order of $N|D^{train}|$ ( $N = 200$ for CUB). Furthermore, we limit the number of referable classes for each target class, which means the number of referable backgrounds decreases, resulting in a synthetic dataset of relatively lower diversity. The results are shown in Fig. 7 (b), and we observe a consistent improvement in performance as the number of referable classes increases. These results consistently underscore the critical role of background diversity introduced by Diff-Mix. + +Impact of fine-tuning strategy. Here we compare three different fine-tuning strategies: TI, DB, and the combined $\mathrm{TI + DB}$ . All strategies share the same fine-tuning hyperparameters and training steps (35000). To evaluate the distribution gap, we compute the FID score [20] of synthesized images (Diff-Gen) with the training set. As illustrated in Table 5, we observe that both $\mathrm{TI + DB}$ and TI have lower FID scores than DB. This can be attributed to the fact that semantic proximity impedes the convergence process. Additionally, while using TI alone results in a relatively low FID score, the improvements in performance are limited. This limitation stems from TI's inability to accurately reconstruct detailed concept (foreground) information, as it is primarily fine-tuned at the semantic level [28]. + +Annotation function. In this section, we discuss the impact of the choices of the translation strength $s$ and the non-linear factor $\gamma$ in Eq. 3. As shown in Table 6, we observe that as the translation strength decreases, the optimal value for $\gamma$ also decreases, which underscores the nonlinearity of Diff-Mix. The comparison between the 5-shot and all-shot settings indicates that the model tends to prefer a more diverse synthetic dataset when the number of training shots is small ( $s = 0.9$ for 5-shot, $s = 0.7$ for all-shot). Besides, a larger confidence in the target class is pre + +
BaselineTIDBTI + DB
5-shotFID (Diff-Gen)-18.2619.5518.43
Acc. (Diff-Mix)50.9057.6456.1159.41
All-shotFID (Diff-Gen)-14.1314.6413.99
Acc. (Diff-Mix)81.6081.8681.9982.85
+ +Table 5. Comparison of distribution gap and classification accuracy across three fine-tuning strategies. TI solely fine-tunes the identifier, and DB solely fine-tunes the U-Net, and $\mathrm{TI + DB}$ + +
γ5-shotAll-shot
s = 0.5s = 0.7s = 0.9s = 0.5s = 0.7s = 0.9
1.5-4.50-0.31+10.30-1.08+0.92+0.90
1.0-2.31+2.99+10.79+0.25+1.14+0.90
0.5+2.35+8.44+11.01+0.92+1.30+0.86
0.3+3.94+9.41+11.15+0.97+1.24+0.69
0.1+6.18+9.86+10.84+0.50+0.88+0.84
0.0+5.25+9.41+11.06+0.38+0.63+0.54
+ +Table 6. Comparison of performance gain across various $\gamma$ and translation strength $s$ . Lower $\gamma$ indicates a higher confidence over target class, e.g. ( $\gamma = 0.1$ , $s = 0.7$ ) results in $0.04y_{i} + 0.96y_{j}$ and ( $\gamma = 0.5$ , $s = 0.7$ ) results in $0.16y_{i} + 0.84y_{j}$ . + +ferred when the shot number is small ( $\gamma = 0.1$ for 5-shot, $\gamma = 0.5$ for all-shot). A possible explanation is that the all-shot setting is less tolerant towards unrealistic images, as discussed in Section 6. Empirically, we recommend choosing a higher translation strength ( $s \in [0.5, 0.7, 0.9]$ ) and a smaller $\gamma$ ( $\gamma \in [0.1, 0.3, 0.5]$ ) as a conservative option. + +# 5. Conclusion + +In this work, we investigate two pivotal aspects, faithfulness and diversity, that are critical for the current state-of-the-art text-to-image generative models to enhance image classification tasks. To achieve a more effective balance between these two aspects, we propose an inter-class augmentation strategy that leverages Stable Diffusion. This method enables generative models to produce a greater diversity of samples by editing images from other classes and shows consistent performance improvement across various classification tasks. + +# 6. Acknowledgement + +This research is mainly supported by the National Natural Science Foundation of China (92270114). This work is also partially supported by the National Key R&D Program of China under Grant 2021ZD0112801. + +# References + +[1] Antreas Antoniou, Amos Storkey, and Harrison Edwards. Data augmentation generative adversarial networks. arXiv preprint arXiv:1711.04340, 2017. 2 +[2] Shekoofeh Azizi, Simon Kornblith, Chitwan Sahara, Mohammad Norouzi, and David J Fleet. Synthetic data from diffusion models improves imagenet classification. arXiv preprint arXiv:2304.08466, 2023.1, 2 +[3] Haoyue Bai, Ceyuan Yang, Yinghao Xu, S-H Gary Chan, and Bolei Zhou. Improving out-of-distribution robustness of classifiers via generative interpolation. arXiv preprint arXiv:2307.12219, 2023. 2 +[4] Hritik Bansal and Aditya Grover. Leaving reality to imagination: Robust classification via generated datasets. arXiv preprint arXiv:2302.02503, 2023. 1, 2 +[5] Christopher Beckham, Sina Honari, Vikas Verma, Alex M Lamb, Farnoosh Ghadiri, R Devon Hjelm, Yoshua Bengio, and Chris Pal. On adversarial mixup resynthesis. Advances in neural information processing systems, 32, 2019. 2 +[6] Ankan Kumar Bhunia, Salman Khan, Hisham Cholakkal, Rao Muhammad Anwer, Jorma Laaksonen, Mubarak Shah, and Fahad Shahbaz Khan. Person image synthesis via denoising diffusion model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5968-5976, 2023. 2 +[7] Andrew Brock, Jeff Donahue, and Karen Simonyan. Large scale gan training for high fidelity natural image synthesis. arXiv preprint arXiv:1809.11096, 2018. 1 +[8] Tim Brooks, Aleksander Holynski, and Alexei A Efros. Instructpix2pix: Learning to follow image editing instructions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18392-18402, 2023. 3 +[9] Kaidi Cao, Colin Wei, Adrien Gaidon, Nikos Arechiga, and Tengyu Ma. Learning imbalanced datasets with label-distribution-aware margin loss. Advances in neural information processing systems, 32, 2019. 6 +[10] Kaidi Cao, Colin Wei, Adrien Gaidon, Nikos Arechiga, and Tengyu Ma. Learning imbalanced datasets with label-distribution-aware margin loss. Advances in neural information processing systems, 32, 2019. 7 +[11] Olivier Chapelle, Jason Weston, Léon Bottou, and Vladimir Vapnik. Vicinal risk minimization. Advances in neural information processing systems, 13, 2000. 6 +[12] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in neural information processing systems, 34:8780-8794, 2021. 1 +[13] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 6 +[14] M. Everingham, S. M. A. Eslami, L. Van Gool, C. K. I. Williams, J. Winn, and A. Zisserman. The pascal visual object classes challenge: A retrospective. International Journal of Computer Vision, 111(1):98-136, 2015. 1, 2, 5 + +[15] Rinon Gal, Yuval Alaluf, Yuval Atzmon, Or Patashnik, Amit H Bermano, Gal Chechik, and Daniel Cohen-Or. An image is worth one word: Personalizing text-to-image generation using textual inversion. arXiv preprint arXiv:2208.01618, 2022. 2, 3 +[16] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. Advances in neural information processing systems, 27, 2014. 2 +[17] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. Advances in neural information processing systems, 27, 2014. 1 +[18] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 6 +[19] Ruifei He, Shuyang Sun, Xin Yu, Chuhui Xue, Wenqing Zhang, Philip Torr, Song Bai, and Xiaojuan Qi. Is synthetic data from generative models ready for image recognition? arXiv preprint arXiv:2210.07574, 2022. 2, 6, 7, 3 +[20] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 8 +[21] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 3 +[22] Jonathan Ho, William Chan, Chitwan Sahara, Jay Whang, Ruiqi Gao, Alexey Gritsenko, Diederik P Kingma, Ben Poole, Mohammad Norouzi, David J Fleet, et al. Imagen video: High definition video generation with diffusion models. arXiv preprint arXiv:2210.02303, 2022. 2 +[23] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 4 +[24] Phillip Isola, Jun-Yan Zhu, Tinghui Zhou, and Alexei A Efros. Image-to-image translation with conditional adversarial networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1125-1134, 2017. 3 +[25] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4401-4410, 2019. 1, 2 +[26] Aditya Khosla, Nityananda Jayadevaprakash, Bangpeng Yao, and Li Fei-Fei. Novel dataset for fine-grained image categorization. In First Workshop on Fine-Grained Visual Categorization, IEEE Conference on Computer Vision and Pattern Recognition, Colorado Springs, CO, 2011. 6, 2 +[27] Jonathan Krause, Michael Stark, Jia Deng, and Li Fei-Fei. 3d object representations for fine-grained categorization. In Proceedings of the IEEE international conference on computer vision workshops, pages 554–561, 2013. 6, 2 +[28] Nupur Kumari, Bingliang Zhang, Richard Zhang, Eli Shechtman, and Jun-Yan Zhu. Multi-concept customization + +of text-to-image diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1931-1941, 2023. 2, 3, 8 +[29] Alexander C Li, Mihir Prabhudesai, Shivam Duggal, Ellis Brown, and Deepak Pathak. Your diffusion model is secretly a zero-shot classifier. arXiv preprint arXiv:2303.16203, 2023. 2 +[30] Yicong Li, Xiang Wang, Junbin Xiao, Wei Ji, and Tat seng Chua. Invariant grounding for video question answering. In CVPR, 2022. 2 +[31] Shanchuan Lin, Bingchen Liu, Jiashi Li, and Xiao Yang. Common diffusion noise schedules and sample steps are flawed. arXiv preprint arXiv:2305.08891, 2023. 1 +[32] Ziwei Liu, Zhongqi Miao, Xiaohang Zhan, Jiayun Wang, Boqing Gong, and Stella X Yu. Large-scale long-tailed recognition in an open world. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2537-2546, 2019. 6 +[33] Cheng Lu, Yuhao Zhou, Fan Bao, Jianfei Chen, Chongxuan Li, and Jun Zhu. Dpm-solver: A fast ode solver for diffusion probabilistic model sampling in around 10 steps. Advances in Neural Information Processing Systems, 35:5775-5787, 2022. 3 +[34] Subhransu Maji, Esa Rahtu, Juho Kannala, Matthew Blaschko, and Andrea Vedaldi. Fine-grained visual classification of aircraft. arXiv preprint arXiv:1306.5151, 2013. 6, 2 +[35] Chenlin Meng, Yutong He, Yang Song, Jiaming Song, Jia-jun Wu, Jun-Yan Zhu, and Stefano Ermon. Sdedit: Guided image synthesis and editing with stochastic differential equations. arXiv preprint arXiv:2108.01073, 2021. 3 +[36] Rafael Müller, Simon Kornblith, and Geoffrey E Hinton. When does label smoothing help? Advances in neural information processing systems, 32, 2019. 6 +[37] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741, 2021. 1, 2 +[38] Maria-Elena Nilsback and Andrew Zisserman. Automated flower classification over a large number of classes. In 2008 Sixth Indian conference on computer vision, graphics & image processing, pages 722-729. IEEE, 2008. 6, 2 +[39] Seulki Park, Youngkyu Hong, Byeongho Heo, Sangdoo Yun, and Jin Young Choi. The majority can help the minority: Context-rich minority oversampling for long-tailed classification. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6887-6896, 2022. 6, 7, 3 +[40] Maitreya Patel, Tejas Gokhale, Chitta Baral, and Yezhou Yang. Conceptbed: Evaluating concept learning abilities of text-to-image diffusion models. arXiv preprint arXiv:2306.04695, 2023. 3 +[41] Judea Pearl. Causal inference in statistics: An overview. 2009. 5 +[42] Zeju Qiu, Weiyang Liu, Haiwen Feng, Yuxuan Xue, Yao Feng, Zhen Liu, Dan Zhang, Adrian Weller, and Bernhard + +Schölkopf. Controlling text-to-image diffusion by orthogonal finetuning. Advances in Neural Information Processing Systems, 36, 2024. 3 +[43] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 5, 1 +[44] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models, 2021. 1 +[45] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical Image Computing and Computer-Assisted Intervention-MICCAI 2015: 18th International Conference, Munich, Germany, October 5-9, 2015, Proceedings, Part III 18, pages 234-241. Springer, 2015. 3 +[46] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22500-22510, 2023. 2, 3 +[47] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, et al. Imagenet large scale visual recognition challenge. International journal of computer vision, 115:211-252, 2015. 2 +[48] Shiori Sagawa, Pang Wei Koh, Tatsunori B Hashimoto, and Percy Liang. Distributionally robust neural networks for group shifts: On the importance of regularization for worst-case generalization. arXiv preprint arXiv:1911.08731, 2019. 7 +[49] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, Seyed Kamyar Seyed Ghasemipour, Burcu Karagol Ayan, S. Sara Mahdavi, Rapha Gontijo Lopes, Tim Salimans, Jonathan Ho, David J Fleet, and Mohammad Norouzi. Photorealistic text-to-image diffusion models with deep language understanding, 2022. 1, 2, 6 +[50] Dvir Samuel, Yuval Atzmon, and Gal Chechik. From generalized zero-shot learning to long-tail with class descriptors. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 286-295, 2021. 6 +[51] Veit Sandfort, Ke Yan, Perry J Pickhardt, and Ronald M Summers. Data augmentation using generative adversarial networks (cyclegan) to improve generalizability in ct segmentation tasks. Scientific reports, 9(1):16884, 2019. 2 +[52] Mert Bulent Sariyildiz, Karteek Alahari, Diane Larlus, and Yannis Kalantidis. Fake it till you make it: Learning transferable representations from synthetic imagenet clones. In CVPR 2023-IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023. 1 +[53] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training + +next generation image-text models. Advances in Neural Information Processing Systems, 35:25278-25294, 2022. 2 +[54] Jordan Shipard, Arnold Wiliem, Kien Nguyen Thanh, Wei Xiang, and Clinton Fookes. Boosting zero-shot classification with synthetic data diversity via stable diffusion. arXiv preprint arXiv:2302.03298, 2023. 2 +[55] Uriel Singer, Adam Polyak, Thomas Hayes, Xi Yin, Jie An, Songyang Zhang, Qiyuan Hu, Harry Yang, Oron Ashual, Oran Gafni, et al. Make-a-video: Text-to-video generation without text-video data. arXiv preprint arXiv:2209.14792, 2022. 2 +[56] Shikun Sun, Longhui Wei, Zhicai Wang, Zixuan Wang, Junliang Xing, Jia Jia, and Qi Tian. Inner classifier-free guidance and its taylor expansion for diffusion models. In The Twelfth International Conference on Learning Representations, 2023. 2 +[57] Yonglong Tian, Lijie Fan, Phillip Isola, Huiwen Chang, and Dilip Krishnan. Stablerep: Synthetic images from text-to-image models make strong visual representation learners. arXiv preprint arXiv:2306.00984, 2023. 1, 2 +[58] Brandon Trabucco, Kyle Doherty, Max Gurinas, and Ruslan Salakhutdinov. Effective data augmentation with diffusion models. arXiv preprint arXiv:2302.07944, 2023. 2, 5, 6, 7, 3 +[59] Narek Tumanyan, Michal Geyer, Shai Bagon, and Tali Dekel. Plug-and-play diffusion features for text-driven image-to-image translation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1921-1930, 2023. 3 +[60] Patrick von Platen, Suraj Patil, Anton Lozhkov, Pedro Cuenca, Nathan Lambert, Kashif Rasul, Mishig Davaadorj, and Thomas Wolf. Diffusers: State-of-the-art diffusion models. https://github.com/huggingface/ diffusers, 2022.3 +[61] Catherine Wah, Steve Branson, Peter Welinder, Pietro Perona, and Serge Belongie. The caltech-ucsd birds-200-2011 dataset. 2011. 5, 6, 2 +[62] Zhicai Wang, Yanbin Hao, Tingting Mu, Ouxiang Li, Shuo Wang, and Xiangnan He. Bi-directional distribution alignment for transductive zero-shot learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19893-19902, 2023. 2 +[63] Daniel Watson, William Chan, Ricardo Martin-Brualla, Jonathan Ho, Andrea Tagliasacchi, and Mohammad Norouzi. Novel view synthesis with diffusion models. arXiv preprint arXiv:2210.04628, 2022. 2 +[64] Zongze Wu, Dani Lischinski, and Eli Shechtman. Stylespace analysis: Disentangled controls for stylegan image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12863-12872, 2021. 3 +[65] Moon Ye-Bin, Nam Hyeon-Woo, Wonseok Choi, Nayeong Kim, Suha Kwak, and Tae-Hyun Oh. Exploiting synthetic data for data imbalance problems: Baselines from a data perspective. arXiv preprint arXiv:2308.00994, 2023. 6 +[66] Sangdoo Yun, Dongyoon Han, Seong Joon Oh, Sanghyuk Chun, Junsuk Choe, and Youngjoon Yoo. Cutmix: Regularization strategy to train strong classifiers with localizable + +features. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6023-6032, 2019. 2, 6, 7 +[67] An Zhang, Wenchang Ma, Xiang Wang, and Tat seng Chua. Incorporating bias-aware margins into contrastive loss for collaborative filtering. In NeurlPS, 2022. 2 +[68] Hongyi Zhang, Moustapha Cisse, Yann N Dauphin, and David Lopez-Paz. mixup: Beyond empirical risk minimization. arXiv preprint arXiv:1710.09412, 2017. 2, 6, 7 +[69] Chenyu Zheng, Guoqiang Wu, and Chongxuan Li. Toward understanding generative data augmentation. Advances in Neural Information Processing Systems, 36, 2024. 2 +[70] Bolei Zhou, Agata Lapedriza, Aditya Khosla, Aude Oliva, and Antonio Torralba. Places: A 10 million image database for scene recognition. IEEE transactions on pattern analysis and machine intelligence, 40(6):1452-1464, 2017. 7 +[71] Jun-Yan Zhu, Taesung Park, Phillip Isola, and Alexei A Efros. Unpaired image-to-image translation using cycle-consistent adversarial networks. In Proceedings of the IEEE international conference on computer vision, pages 2223-2232, 2017. 3 \ No newline at end of file diff --git a/2024/Enhance Image Classification via Inter-Class Image Mixup with Diffusion Model/images.zip b/2024/Enhance Image Classification via Inter-Class Image Mixup with Diffusion Model/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..3da94fd5882eac589cd837e94a4df6180313fcad --- /dev/null +++ b/2024/Enhance Image Classification via Inter-Class Image Mixup with Diffusion Model/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8df811607d9cdddc9c1a560b0ae2f98bdc29903f4d566138676c70103c84c1ba +size 581409 diff --git a/2024/Enhance Image Classification via Inter-Class Image Mixup with Diffusion Model/layout.json b/2024/Enhance Image Classification via Inter-Class Image Mixup with Diffusion Model/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..38371458d804405e0f8f217f9f6a2c82ad3d3b23 --- /dev/null +++ b/2024/Enhance Image Classification via Inter-Class Image Mixup with Diffusion Model/layout.json @@ -0,0 +1,9972 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 48, + 103, + 545, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 103, + 545, + 121 + ], + "spans": [ + { + "bbox": [ + 48, + 103, + 545, + 121 + ], + "type": "text", + "content": "Enhance Image Classification via Inter-Class Image Mixup with Diffusion Model" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 91, + 142, + 503, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 142, + 503, + 171 + ], + "spans": [ + { + "bbox": [ + 91, + 142, + 503, + 171 + ], + "type": "text", + "content": "Zhicai Wang" + }, + { + "bbox": [ + 91, + 142, + 503, + 171 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 91, + 142, + 503, + 171 + ], + "type": "text", + "content": "; Longhui Wei" + }, + { + "bbox": [ + 91, + 142, + 503, + 171 + ], + "type": "inline_equation", + "content": "^{2\\dagger}" + }, + { + "bbox": [ + 91, + 142, + 503, + 171 + ], + "type": "text", + "content": ", Tan Wang" + }, + { + "bbox": [ + 91, + 142, + 503, + 171 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 91, + 142, + 503, + 171 + ], + "type": "text", + "content": ", Heyu Chen" + }, + { + "bbox": [ + 91, + 142, + 503, + 171 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 91, + 142, + 503, + 171 + ], + "type": "text", + "content": ", Yanbin Hao" + }, + { + "bbox": [ + 91, + 142, + 503, + 171 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 91, + 142, + 503, + 171 + ], + "type": "text", + "content": ", Xiang Wang" + }, + { + "bbox": [ + 91, + 142, + 503, + 171 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 91, + 142, + 503, + 171 + ], + "type": "text", + "content": ", Xiangnan He" + }, + { + "bbox": [ + 91, + 142, + 503, + 171 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 91, + 142, + 503, + 171 + ], + "type": "text", + "content": ", Qi Tian" + }, + { + "bbox": [ + 91, + 142, + 503, + 171 + ], + "type": "inline_equation", + "content": "^{2}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 141, + 171, + 451, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 171, + 451, + 185 + ], + "spans": [ + { + "bbox": [ + 141, + 171, + 451, + 185 + ], + "type": "text", + "content": "1University of Science and Technology of China, 2 Huawei Inc.," + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 209, + 185, + 384, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 209, + 185, + 384, + 200 + ], + "spans": [ + { + "bbox": [ + 209, + 185, + 384, + 200 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 209, + 185, + 384, + 200 + ], + "type": "text", + "content": " Nanyang Technological University" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 143, + 227, + 192, + 239 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 227, + 192, + 239 + ], + "spans": [ + { + "bbox": [ + 143, + 227, + 192, + 239 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 251, + 290, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 251, + 290, + 515 + ], + "spans": [ + { + "bbox": [ + 46, + 251, + 290, + 515 + ], + "type": "text", + "content": "Text-to-image (T2I) generative models have recently emerged as a powerful tool, enabling the creation of photorealistic images and giving rise to a multitude of applications. However, the effective integration of T2I models into fundamental image classification tasks remains an open question. A prevalent strategy to bolster image classification performance is through augmenting the training set with synthetic images generated by T2I models. In this study, we scrutinize the shortcomings of both current generative and conventional data augmentation techniques. Our analysis reveals that these methods struggle to produce images that are both faithful (in terms of foreground objects) and diverse (in terms of background contexts) for domain-specific concepts. To tackle this challenge, we introduce an innovative inter-class data augmentation method known as Diff-Mix1, which enriches the dataset by performing image translations between classes. Our empirical results demonstrate that Diff-Mix achieves a better balance between faithfulness and diversity, leading to a marked improvement in performance across diverse image classification scenarios, including few-shot, conventional, and long-tail classifications for domain-specific datasets." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 536, + 128, + 550 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 536, + 128, + 550 + ], + "spans": [ + { + "bbox": [ + 47, + 536, + 128, + 550 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 557, + 287, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 557, + 287, + 677 + ], + "spans": [ + { + "bbox": [ + 46, + 557, + 287, + 677 + ], + "type": "text", + "content": "In comparison to GAN-based models [7, 17, 25], contemporary state-of-the-art text-to-image (T2I) diffusion models exhibit enhanced capabilities in producing high-fidelity images [12, 37, 44, 49]. With the remarkable cross-modality alignment capabilities of T2I models, there is significant potential for generative techniques to enhance image classification [2, 4]. For instance, a straightforward approach entails augmenting the existing training dataset with synthetic images generated by feeding categorical textual prompts to a T2I diffusion model. However, upon reviewing prior ap" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 309, + 227, + 555, + 403 + ], + "blocks": [ + { + "bbox": [ + 309, + 227, + 555, + 403 + ], + "lines": [ + { + "bbox": [ + 309, + 227, + 555, + 403 + ], + "spans": [ + { + "bbox": [ + 309, + 227, + 555, + 403 + ], + "type": "image", + "image_path": "7ac762f63f12a6572d6b598f0a9e1851bd95d9b404ad63da1bc377f039efb4e6.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 411, + 547, + 533 + ], + "lines": [ + { + "bbox": [ + 304, + 411, + 547, + 533 + ], + "spans": [ + { + "bbox": [ + 304, + 411, + 547, + 533 + ], + "type": "text", + "content": "Figure 1. Strategies to expand domain-specific datasets for improved classification are varied. Row 1 illustrates vanilla distillation from a pretrained text-to-image (T2I) model, which carries the risk of generating outputs with reduced faithfulness. Intra-class augmentation, depicted in Row 2, tends to yield samples with limited diversity to maintain high fidelity to the original class. Our proposed method, showcased in Rows 3 and 4, adopts an inter-class augmentation strategy. This involves introducing edits to a reference image using images from other classes within the training set, which significantly enriches the dataset with a greater diversity of samples." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 545, + 547, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 545, + 547, + 629 + ], + "spans": [ + { + "bbox": [ + 304, + 545, + 547, + 629 + ], + "type": "text", + "content": "proaches employing T2I diffusion models for image classification, it becomes evident that the challenge in generative data augmentation for domain-specific datasets is producing samples with both a faithful foreground and a diverse background. Depending on whether a reference image is used in the generative process, we divide these methods into two groups:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 630, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 630, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 306, + 630, + 547, + 714 + ], + "type": "text", + "content": "- Text-guided knowledge distillation [52, 57] involves generating new images from scratch using category-related prompts to expand the dataset. For the off-the-shelf T2I models, such vanilla distillation presume these models have comprehensive knowledge of target domain, which can be problematic for domain-specific datasets. Insufficient domain knowledge easily makes the distillation" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 57, + 683, + 247, + 693 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 683, + 247, + 693 + ], + "spans": [ + { + "bbox": [ + 57, + 683, + 247, + 693 + ], + "type": "text", + "content": "*This work was done during the internship in Huawei Inc.." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 60, + 693, + 276, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 60, + 693, + 276, + 703 + ], + "spans": [ + { + "bbox": [ + 60, + 693, + 276, + 703 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 60, + 693, + 276, + 703 + ], + "type": "text", + "content": " Xiang Wang and Longhui Wei are both the corresponding authors." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 60, + 703, + 194, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 60, + 703, + 194, + 712 + ], + "spans": [ + { + "bbox": [ + 60, + 703, + 194, + 712 + ], + "type": "text", + "content": "1https://github.com/Zhicaiww/Diff-Mix" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17223" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 286, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 286, + 120 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 286, + 120 + ], + "type": "text", + "content": "process less effective. For example, vanilla T2I models struggle to generate images that accurately represent specific bird species based solely on their names (see Row 1 of Fig. 1)." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 121, + 286, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 121, + 286, + 251 + ], + "spans": [ + { + "bbox": [ + 48, + 121, + 286, + 251 + ], + "type": "text", + "content": "- Generative data augmentation [1, 69] employs generative models to enhance existing images. Da-fusion [58], for instance, translates the source image into multiple edited versions within the same class. This strategy, termed intra-class augmentation, primarily introduces intra-class variations. While intra-class augmentation retains much of the original image's layout and visual details, it results in limited background diversity (see Row 2 of Fig. 1). However, synthetic images with constrained diversity may not sufficiently enhance the model's ability to discern foreground concepts." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 253, + 286, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 253, + 286, + 300 + ], + "spans": [ + { + "bbox": [ + 47, + 253, + 286, + 300 + ], + "type": "text", + "content": "Based on these observations, a fundamental question emerges: 'Is it feasible to develop a method that optimizes both the diversity and faithfulness of synthesized data simultaneously?'" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 301, + 286, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 301, + 286, + 529 + ], + "spans": [ + { + "bbox": [ + 47, + 301, + 286, + 529 + ], + "type": "text", + "content": "In this work, we introduce Diff-Mix, a simple yet effective data augmentation method that harnesses diffusion models to perform inter-class image interpolation, tailored for enhancing domain-specific datasets. The method encompasses two pivotal operations: personalized fine-tuning and inter-class image translation. Personalized fine-tuning [15, 46] is originally designed for customizing T2I models and enabling them to generate user-specific contents or styles. In our case, we implement the technique to tailor the model, enabling it to generate images with faithful foreground concepts. Inter-class image translation in Diff-Mix entails transforming a reference image into an edited version that incorporates prompts from different classes. This translation strategy is designed to retain the original background context while editing the foreground to align with the target concept. For instance, as depicted in the bottom rows of Fig. 1, Diff-Mix can generate images of land birds in diverse settings, such as maritime environments, enriching the dataset with a variety of counterfactual samples." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 530, + 286, + 661 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 530, + 286, + 661 + ], + "spans": [ + { + "bbox": [ + 47, + 530, + 286, + 661 + ], + "type": "text", + "content": "Unlike previous non-generative augmentation methods, such as Mixup [68] and CutMix [66], Diff-Mix works in a foreground-perceivable inter-class interpolation manner and shares a different mechanism with the non-generative approaches. Our experiment under the conventional classification setting indicates that incorporating both CutMix and Diff-Mix could further enhance performance. Additionally, when compared with other generative approaches, we conduct experiments under few-shot and long-tail scenarios and observe consistent performance improvements. Our contributions can be summarized as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 665, + 286, + 712 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 47, + 665, + 286, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 665, + 286, + 700 + ], + "spans": [ + { + "bbox": [ + 47, + 665, + 286, + 700 + ], + "type": "text", + "content": "- We pinpoint the critical factors that affect the efficacy of generative data augmentation in domain-specific image classification: namely, faithfulness and diversity." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 701, + 286, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 701, + 286, + 712 + ], + "spans": [ + { + "bbox": [ + 47, + 701, + 286, + 712 + ], + "type": "text", + "content": "- We introduce Diff-Mix, a simple yet effective generative" + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 313, + 72, + 544, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 544, + 95 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 544, + 95 + ], + "type": "text", + "content": "data augmentation strategy that leverages fine-tuned diffusion models for inter-class image interpolation." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 96, + 544, + 143 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 96, + 544, + 143 + ], + "spans": [ + { + "bbox": [ + 306, + 96, + 544, + 143 + ], + "type": "text", + "content": "- We conduct a comparative analysis of Diff-Mix with other distillation-based and intra-class augmentation methods, as well as non-generative approaches, highlighting its unique features and benefits." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 162, + 397, + 174 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 162, + 397, + 174 + ], + "spans": [ + { + "bbox": [ + 306, + 162, + 397, + 174 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 184, + 545, + 292 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 184, + 545, + 292 + ], + "spans": [ + { + "bbox": [ + 304, + 184, + 545, + 292 + ], + "type": "text", + "content": "Text-to-image diffusion models. Following pretraining on web-scale data, the T2I diffusion model has demonstrated robust capabilities in generating text-controlled images [37, 49, 53, 56]. Its versatility has led to diverse applications, including novel view synthesis [6, 63], concept learning [28, 46], and text-to-video generation [22, 55], among others. Recent advancements [29] also highlight the cross-modality features of such generative models, showcasing their ability to serve as zero-shot classifiers." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 300, + 545, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 300, + 545, + 586 + ], + "spans": [ + { + "bbox": [ + 304, + 300, + 545, + 586 + ], + "type": "text", + "content": "Synthetic data for image classification. There are two perspectives on the utilization of synthetic data for image classification: knowledge distillation [2] and data augmentation [5, 51, 54, 62]. From the knowledge distillation perspective, SyntheticData [19] reports significant performance gains in both zero-shot and few-shot settings by leveraging off-the-shelf T2I models to obtain synthetic data. The work of [2] has indicated that fine-tuning the T2I model on ImageNet [47] yields improved classification accuracy by narrowing the domain gap. Some works also find that learning from the synthetic data presents strong transferability [19, 57] and robustness [4, 30, 67]. From the data augmentation standpoint, Da-fusion [58] achieves stable performance improvements on few-shot datasets by augmenting from reference images. In a related study [3], the use of StyleGAN [25] for generating interpolated images between two different domains has been shown to enhance classifier robustness for out-of-distribution data. Our work shares similarities with AMR [5], which generates realistic novel examples by interpolating between two images using GAN [16]. The distinction lies in our discussion of interpolation using the T2I diffusion model, where its noise-adding and denoising characteristics enable a smoother implementation of interpolation." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "text", + "content": "Non-generative data augmentation. Mixup [68] and CutMix [66] stand out as two prominent non-generative data augmentation methods, serving as effective regularization techniques during training. While Mixup achieves augmented samples through a convex combination of two images, CutMix achieves augmentation by cutting and pasting parts of images. However, both methods are constrained in their ability to produce realistic images. In addressing this limitation, the utilization of generative models emerges as a potential solution to alleviate this issue." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17224" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 71, + 103, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 71, + 103, + 83 + ], + "spans": [ + { + "bbox": [ + 47, + 71, + 103, + 83 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 91, + 127, + 103 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 91, + 127, + 103 + ], + "spans": [ + { + "bbox": [ + 47, + 91, + 127, + 103 + ], + "type": "text", + "content": "3.1. Preliminary" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 109, + 289, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 109, + 289, + 240 + ], + "spans": [ + { + "bbox": [ + 46, + 109, + 289, + 240 + ], + "type": "text", + "content": "Text-to-Image diffusion model. Diffusion models generate images by gradually removing noise from a Gaussian noise source [21]. In a diffusion process with a total of " + }, + { + "bbox": [ + 46, + 109, + 289, + 240 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 46, + 109, + 289, + 240 + ], + "type": "text", + "content": " steps, its forward process, which gradually adds noise, is represented as a Markov chain with a Gaussian transition kernel, where " + }, + { + "bbox": [ + 46, + 109, + 289, + 240 + ], + "type": "inline_equation", + "content": "q(\\pmb{x}_t|\\pmb{x}_{t - 1}) = \\mathcal{N}(\\pmb{x}_t;\\sqrt{\\alpha_t}\\pmb{x}_{t - 1},(1 - \\alpha_t)\\mathbf{I})" + }, + { + "bbox": [ + 46, + 109, + 289, + 240 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 109, + 289, + 240 + ], + "type": "inline_equation", + "content": "x_{t}" + }, + { + "bbox": [ + 46, + 109, + 289, + 240 + ], + "type": "text", + "content": " represents the noisy image at step " + }, + { + "bbox": [ + 46, + 109, + 289, + 240 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 109, + 289, + 240 + ], + "type": "text", + "content": ". The training objective at step " + }, + { + "bbox": [ + 46, + 109, + 289, + 240 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 109, + 289, + 240 + ], + "type": "text", + "content": " is to predict the noise to reconstruct " + }, + { + "bbox": [ + 46, + 109, + 289, + 240 + ], + "type": "inline_equation", + "content": "\\pmb{x}_{t - 1}" + }, + { + "bbox": [ + 46, + 109, + 289, + 240 + ], + "type": "text", + "content": ". When training a text-conditioned diffusion model, the simplified training objective can be summarized as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 104, + 243, + 287, + 263 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 104, + 243, + 287, + 263 + ], + "spans": [ + { + "bbox": [ + 104, + 243, + 287, + 263 + ], + "type": "interline_equation", + "content": "\\mathbb {E} _ {\\epsilon , \\mathbf {X}, \\mathbf {c}, t} \\left[ \\| \\epsilon - \\epsilon_ {\\theta} (\\mathbf {x} _ {t}, \\mathbf {c}, t) \\| _ {2} ^ {2} \\right], \\tag {1}", + "image_path": "a8d665f4ce970a884e9223d361946459ea2ef873b5d3ca582e96b7ca20f4ae33.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 267, + 287, + 292 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 267, + 287, + 292 + ], + "spans": [ + { + "bbox": [ + 46, + 267, + 287, + 292 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 267, + 287, + 292 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}" + }, + { + "bbox": [ + 46, + 267, + 287, + 292 + ], + "type": "text", + "content": " represents the predicted noise, and " + }, + { + "bbox": [ + 46, + 267, + 287, + 292 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 46, + 267, + 287, + 292 + ], + "type": "text", + "content": " is the encoded text caption associated with the image " + }, + { + "bbox": [ + 46, + 267, + 287, + 292 + ], + "type": "inline_equation", + "content": "\\pmb{x}" + }, + { + "bbox": [ + 46, + 267, + 287, + 292 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 296, + 289, + 452 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 296, + 289, + 452 + ], + "spans": [ + { + "bbox": [ + 46, + 296, + 289, + 452 + ], + "type": "text", + "content": "T2I personalization. T2I personalization aims to personalize a diffusion model for generating specific concepts using a limited number of concept-oriented images [28, 40, 42]. These concepts are typically represented using identifiers (e.g., “[V]”). As a result, we formalize the constructed image-caption set as " + }, + { + "bbox": [ + 46, + 296, + 289, + 452 + ], + "type": "inline_equation", + "content": "\\pmb{x}" + }, + { + "bbox": [ + 46, + 296, + 289, + 452 + ], + "type": "text", + "content": ", “photo of a [V]”. Various personalization methods differ in their fine-tuning strategies. For instance, Textual Inversion (TI) [15] makes the identifier learnable, but other modules are not fine-tuned, potentially sacrificing some faithfulness in image generation. On the other hand, Dreambooth (DB) [46] fine-tunes the U-Net [45] for more refined personalized generation but faces the challenge of increased computational cost." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 456, + 289, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 456, + 289, + 601 + ], + "spans": [ + { + "bbox": [ + 46, + 456, + 289, + 601 + ], + "type": "text", + "content": "Image-to-image translation. Image translation enables image synthesis and editing using a reference image as guidance [24, 64, 71]. Diffusion-based image translation methods can generate fine edits, which refer to subtle modifications, with varying degrees of shift relative to the reference image [8, 35, 59]. Here, we draw inspiration from SDEdit [35] to perform edits on the reference image, where the target image " + }, + { + "bbox": [ + 46, + 456, + 289, + 601 + ], + "type": "inline_equation", + "content": "\\mathbf{x}^{\\mathrm{tar}}" + }, + { + "bbox": [ + 46, + 456, + 289, + 601 + ], + "type": "text", + "content": " is translated from a reference image " + }, + { + "bbox": [ + 46, + 456, + 289, + 601 + ], + "type": "inline_equation", + "content": "\\mathbf{x}^{\\mathrm{ref}}" + }, + { + "bbox": [ + 46, + 456, + 289, + 601 + ], + "type": "text", + "content": ". During translation, the reverse process does not traverse the full process but starts from a certain step " + }, + { + "bbox": [ + 46, + 456, + 289, + 601 + ], + "type": "inline_equation", + "content": "\\lfloor sT \\rfloor" + }, + { + "bbox": [ + 46, + 456, + 289, + 601 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 456, + 289, + 601 + ], + "type": "inline_equation", + "content": "s \\in [0,1]" + }, + { + "bbox": [ + 46, + 456, + 289, + 601 + ], + "type": "text", + "content": " controls the insertion position of the reference image with noise, as follows," + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 86, + 605, + 287, + 625 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 86, + 605, + 287, + 625 + ], + "spans": [ + { + "bbox": [ + 86, + 605, + 287, + 625 + ], + "type": "interline_equation", + "content": "\\boldsymbol {x} _ {\\lfloor s T \\rfloor} = \\sqrt {\\tilde {\\alpha} _ {\\lfloor s T \\rfloor}} \\boldsymbol {x} _ {0} ^ {\\text {r e f}} + \\sqrt {1 - \\tilde {\\alpha} _ {\\lfloor s T \\rfloor}} \\epsilon . \\tag {2}", + "image_path": "33c242337471952eaf98a8ecc2f76dc728b287536053be6a3f3b118285fd9773.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 629, + 289, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 629, + 289, + 666 + ], + "spans": [ + { + "bbox": [ + 46, + 629, + 289, + 666 + ], + "type": "text", + "content": "By adjusting the strength parameter " + }, + { + "bbox": [ + 46, + 629, + 289, + 666 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 46, + 629, + 289, + 666 + ], + "type": "text", + "content": ", one can strike a balance between the diversity of the generated images and their faithfulness to the reference image." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 671, + 167, + 683 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 671, + 167, + 683 + ], + "spans": [ + { + "bbox": [ + 47, + 671, + 167, + 683 + ], + "type": "text", + "content": "3.2. General Framework" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 689, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 689, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 689, + 289, + 715 + ], + "type": "text", + "content": "The Diff-Mix pipeline consists of two key steps. Firstly, to produce more faithful images for domain-specific datasets," + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 306, + 69, + 547, + 102 + ], + "blocks": [ + { + "bbox": [ + 306, + 69, + 547, + 102 + ], + "lines": [ + { + "bbox": [ + 306, + 69, + 547, + 102 + ], + "spans": [ + { + "bbox": [ + 306, + 69, + 547, + 102 + ], + "type": "image", + "image_path": "679e9849bd49aaf0cc0d70513702151847cb42d3e435a39c2e741616576fd03e.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 400, + 103, + 452, + 112 + ], + "lines": [ + { + "bbox": [ + 400, + 103, + 452, + 112 + ], + "spans": [ + { + "bbox": [ + 400, + 103, + 452, + 112 + ], + "type": "text", + "content": "(a) Real images" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 307, + 113, + 547, + 145 + ], + "blocks": [ + { + "bbox": [ + 307, + 113, + 547, + 145 + ], + "lines": [ + { + "bbox": [ + 307, + 113, + 547, + 145 + ], + "spans": [ + { + "bbox": [ + 307, + 113, + 547, + 145 + ], + "type": "image", + "image_path": "a835499778bd5de0d3f981d224126e4e499c62f5672a18da7d0a8988636b366b.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 307, + 156, + 546, + 187 + ], + "blocks": [ + { + "bbox": [ + 372, + 146, + 481, + 156 + ], + "lines": [ + { + "bbox": [ + 372, + 146, + 481, + 156 + ], + "spans": [ + { + "bbox": [ + 372, + 146, + 481, + 156 + ], + "type": "text", + "content": "(b) Synthetic images (vanilla SD)" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 307, + 156, + 546, + 187 + ], + "lines": [ + { + "bbox": [ + 307, + 156, + 546, + 187 + ], + "spans": [ + { + "bbox": [ + 307, + 156, + 546, + 187 + ], + "type": "image", + "image_path": "bdb3d0dfa6ae058b391cc0d7a8e02a99ee4653596d741a468d8049886392d492.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 307, + 199, + 545, + 231 + ], + "blocks": [ + { + "bbox": [ + 354, + 189, + 498, + 198 + ], + "lines": [ + { + "bbox": [ + 354, + 189, + 498, + 198 + ], + "spans": [ + { + "bbox": [ + 354, + 189, + 498, + 198 + ], + "type": "text", + "content": "(c) Synthetic images (SD fine-tuned via DB)" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 307, + 199, + 545, + 231 + ], + "lines": [ + { + "bbox": [ + 307, + 199, + 545, + 231 + ], + "spans": [ + { + "bbox": [ + 307, + 199, + 545, + 231 + ], + "type": "image", + "image_path": "77de4e4343c7bac68f53f7e49ade331e8f701868343060f55a09a66404d3c9a9.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 347, + 232, + 504, + 242 + ], + "lines": [ + { + "bbox": [ + 347, + 232, + 504, + 242 + ], + "spans": [ + { + "bbox": [ + 347, + 232, + 504, + 242 + ], + "type": "text", + "content": "(d) Synthetic images (SD fine-tuned via " + }, + { + "bbox": [ + 347, + 232, + 504, + 242 + ], + "type": "inline_equation", + "content": "\\mathrm{TI + DB}" + }, + { + "bbox": [ + 347, + 232, + 504, + 242 + ], + "type": "text", + "content": ")" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 304, + 251, + 547, + 329 + ], + "lines": [ + { + "bbox": [ + 304, + 251, + 547, + 329 + ], + "spans": [ + { + "bbox": [ + 304, + 251, + 547, + 329 + ], + "type": "text", + "content": "Figure 2. Examples of \"American Three toed Woodpecker\". (a) Real images from the training set. (b-d) synthetic images generated using different fine-tuned models with the same number of fine-tuning steps. TI+DB indicates both text embedding and U-Net are fine-tuned. TI+DB achieves a more faithful output compared to DB alone (check the head and wing patterns of the birds)." + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 345, + 547, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 345, + 547, + 418 + ], + "spans": [ + { + "bbox": [ + 304, + 345, + 547, + 418 + ], + "type": "text", + "content": "we propose treating it as a T2I personalization problem and fine-tuning the Stable Diffusion (SD). Subsequently, to enhance the diversity of synthetic data beyond the well-fitted training distribution, we employ inter-class image translation. This process produces interpolated images with increased background diversity for each class." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 306, + 426, + 454, + 437 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 426, + 454, + 437 + ], + "spans": [ + { + "bbox": [ + 306, + 426, + 454, + 437 + ], + "type": "text", + "content": "3.3. Fine-tune Diffusion Model" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 304, + 445, + 547, + 505 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 445, + 547, + 505 + ], + "spans": [ + { + "bbox": [ + 304, + 445, + 547, + 505 + ], + "type": "text", + "content": "Vanilla distillation tends to be less effective, especially as the number of training shots increases (refer to Sec. 4.1). In order to mitigate the distribution gap, we propose finetuning Stable Diffusion in conjunction with current widely-used T2I personalization strategies." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 304, + 510, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 510, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 510, + 547, + 714 + ], + "type": "text", + "content": "Dreambooth meets Textual Inversion. Many fine-grained datasets provide terminological names for their categories, like \"American Three toed Woodpecker\" and \"Pileated Woodpecker\". We could construct image-text pairs using category-related prompts and fine-tune the denoising network of SD using Eq. 1, which is analogous to Dreambooth. However, we observe that directly incorporating these specialized terms into the text during fine-tuning can impede convergence and hinder the generation of faithful images. We attribute this challenge to the semantic proximity of terminology within a fine-grained domain, where fine-tuning the vision module alone tends to be less effective at distinguishing two similar classes within the same family, like \"Woodpecker\". Inspired by Textual Inversion [15], we opt to replace the terminological name in the dataset with \"[Vi] [metaclass]\" where \"[Vi\"] is a learnable identifier, and " + }, + { + "bbox": [ + 304, + 510, + 547, + 714 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 510, + 547, + 714 + ], + "type": "text", + "content": " varies from 1 to " + }, + { + "bbox": [ + 304, + 510, + 547, + 714 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 510, + 547, + 714 + ], + "type": "text", + "content": ", represent-" + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17225" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 51, + 70, + 287, + 281 + ], + "blocks": [ + { + "bbox": [ + 51, + 70, + 287, + 281 + ], + "lines": [ + { + "bbox": [ + 51, + 70, + 287, + 281 + ], + "spans": [ + { + "bbox": [ + 51, + 70, + 287, + 281 + ], + "type": "image", + "image_path": "2183af4399b8471a3a3a3b7189d897524e8e43fdc30709817cab54d8704ae83d.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 290, + 288, + 357 + ], + "lines": [ + { + "bbox": [ + 46, + 290, + 288, + 357 + ], + "spans": [ + { + "bbox": [ + 46, + 290, + 288, + 357 + ], + "type": "text", + "content": "Figure 3. Fine-tuning framework of Diff-Mix operates as follows: Initially, we replace the class name with a structured identifier formatted as “[" + }, + { + "bbox": [ + 46, + 290, + 288, + 357 + ], + "type": "inline_equation", + "content": "\\mathsf{V}^i" + }, + { + "bbox": [ + 46, + 290, + 288, + 357 + ], + "type": "text", + "content": "” [metaclass]}, thereby sidestepping the need for specific terminological expressions. Next, we engage in joint fine-tuning of these identifiers and the low-rank residues (LoRA) of U-Net to capture the domain-specific distribution." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 369, + 287, + 467 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 369, + 287, + 467 + ], + "spans": [ + { + "bbox": [ + 46, + 369, + 287, + 467 + ], + "type": "text", + "content": "ing the category index. The illustration of our fine-tuning strategy is presented in Fig. 3. The term “[metaclass]” is determined by the theme of the current dataset, such as “bird” for a fine-grained bird dataset. By concurrently fine-tuning the identifier and the U-Net, we empower the model to quickly adapting to the fine-grained domain, allowing it to generate faithful images using the identifier (see comparison between Row 3 and Row 4 in Fig. 2)." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 472, + 287, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 472, + 287, + 579 + ], + "spans": [ + { + "bbox": [ + 46, + 472, + 287, + 579 + ], + "type": "text", + "content": "Parameter efficient fine-tuning. In this context, we embrace the parameter-efficient fine-tuning strategy known as LoRA [23]. LoRA distinguishes itself by fine-tuning the residue of low-rank matrices instead of directly fine-tuning the pre-trained weight matrices. To elaborate, consider a weight matrix " + }, + { + "bbox": [ + 46, + 472, + 287, + 579 + ], + "type": "inline_equation", + "content": "\\mathbf{W} \\in \\mathbb{R}^{m \\times n}" + }, + { + "bbox": [ + 46, + 472, + 287, + 579 + ], + "type": "text", + "content": ". The tunable residual matrix " + }, + { + "bbox": [ + 46, + 472, + 287, + 579 + ], + "type": "inline_equation", + "content": "\\Delta \\mathbf{W}" + }, + { + "bbox": [ + 46, + 472, + 287, + 579 + ], + "type": "text", + "content": " comprises two low-rank matrices: " + }, + { + "bbox": [ + 46, + 472, + 287, + 579 + ], + "type": "inline_equation", + "content": "\\mathbf{A} \\in \\mathbb{R}^{m \\times d}" + }, + { + "bbox": [ + 46, + 472, + 287, + 579 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 472, + 287, + 579 + ], + "type": "inline_equation", + "content": "\\mathbf{B} \\in \\mathbb{R}^{n \\times d}" + }, + { + "bbox": [ + 46, + 472, + 287, + 579 + ], + "type": "text", + "content": ", defined as " + }, + { + "bbox": [ + 46, + 472, + 287, + 579 + ], + "type": "inline_equation", + "content": "\\Delta \\mathbf{W} = \\mathbf{A}\\mathbf{B}^{\\top}" + }, + { + "bbox": [ + 46, + 472, + 287, + 579 + ], + "type": "text", + "content": ". As a default configuration, we set the rank " + }, + { + "bbox": [ + 46, + 472, + 287, + 579 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 46, + 472, + 287, + 579 + ], + "type": "text", + "content": " to 10." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 588, + 247, + 601 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 588, + 247, + 601 + ], + "spans": [ + { + "bbox": [ + 47, + 588, + 247, + 601 + ], + "type": "text", + "content": "3.4. Data Synthesis Using Diffusion Model" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 607, + 287, + 655 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 607, + 287, + 655 + ], + "spans": [ + { + "bbox": [ + 46, + 607, + 287, + 655 + ], + "type": "text", + "content": "In generating pseudo data, three strategies can be used with our fine-tuned diffusion model " + }, + { + "bbox": [ + 46, + 607, + 287, + 655 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 46, + 607, + 287, + 655 + ], + "type": "text", + "content": ": (1) distillation-based method Diff-Gen, (2) intra-class augmentation Diff-Aug, and (3) inter-class augmentation Diff-Mix." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 660, + 287, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 660, + 287, + 685 + ], + "spans": [ + { + "bbox": [ + 47, + 660, + 287, + 685 + ], + "type": "text", + "content": "Diff-Gen and Diff-Aug. For a target class " + }, + { + "bbox": [ + 47, + 660, + 287, + 685 + ], + "type": "inline_equation", + "content": "y^i" + }, + { + "bbox": [ + 47, + 660, + 287, + 685 + ], + "type": "text", + "content": " and its textual condition, \"photo of a " + }, + { + "bbox": [ + 47, + 660, + 287, + 685 + ], + "type": "inline_equation", + "content": "[V^i]" + }, + { + "bbox": [ + 47, + 660, + 287, + 685 + ], + "type": "text", + "content": " [metaclass]\", both" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 307, + 70, + 544, + 281 + ], + "blocks": [ + { + "bbox": [ + 307, + 70, + 544, + 281 + ], + "lines": [ + { + "bbox": [ + 307, + 70, + 544, + 281 + ], + "spans": [ + { + "bbox": [ + 307, + 70, + 544, + 281 + ], + "type": "image", + "image_path": "cbdd6c90118c3b11ceb3757f7910278802c1de706b8a7200f87846bfc29ff2dc.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 290, + 546, + 357 + ], + "lines": [ + { + "bbox": [ + 305, + 290, + 546, + 357 + ], + "spans": [ + { + "bbox": [ + 305, + 290, + 546, + 357 + ], + "type": "text", + "content": "Figure 4. Examples of images translated using Diff-Mix and Diff-Aug across various strengths. Diff-Aug employs the same target and reference image classes, typically resulting in subtle modifications. Diff-Mix progressively adjusts the foreground to align with the target class as the translation strength increases, while preserving the background layout from the reference image." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 371, + 546, + 539 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 371, + 546, + 539 + ], + "spans": [ + { + "bbox": [ + 304, + 371, + 546, + 539 + ], + "type": "text", + "content": "methods generate synthetic samples annotated with class " + }, + { + "bbox": [ + 304, + 371, + 546, + 539 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 371, + 546, + 539 + ], + "type": "text", + "content": ". Specifically, Diff-Gen generates samples from scratch by initializing with random Gaussian noise and proceeding through the full reverse diffusion process with " + }, + { + "bbox": [ + 304, + 371, + 546, + 539 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 304, + 371, + 546, + 539 + ], + "type": "text", + "content": " steps. Diff-Gen can produce images aligned with its fine-tuned distribution. In contrast, Diff-Aug sacrifices a portion of diversity and generate images by editing on a reference image. Specifically, it randomly sample a image from the intra-class training set and enhances the image through image translation using Eq. 2. The term \"intra-class\" means that the conditioning prompts are constructed based on the ground truth categories of images, and such a denoising process tends to introduce less variation, particularly for the foreground concepts (see top rows of Fig. 1)." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 545, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 545, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 545, + 546, + 713 + ], + "type": "text", + "content": "Diff-Mix. Diff-Mix employs the same translation process as Diff-Aug, but the reference image is sampled from the full training set rather than intra-class set to enable interclass interpolation. The key difference is that Diff-Mix can generate numerous counterfactual examples, such as a blackbird in the sea (see fourth row of Fig. 1). This necessitates that downstream models make a more refined differentiation of category attributes, thereby reducing the impact of spurious correlations introduced by variations in the background. Denoting the label of the reference image as " + }, + { + "bbox": [ + 304, + 545, + 546, + 713 + ], + "type": "inline_equation", + "content": "y^{j}" + }, + { + "bbox": [ + 304, + 545, + 546, + 713 + ], + "type": "text", + "content": ", by inserting the reference image into the reverse process with the prompt \"photo of a " + }, + { + "bbox": [ + 304, + 545, + 546, + 713 + ], + "type": "inline_equation", + "content": "[V^i]" + }, + { + "bbox": [ + 304, + 545, + 546, + 713 + ], + "type": "text", + "content": " [metaclass]\", we can obtain interpolated images between the " + }, + { + "bbox": [ + 304, + 545, + 546, + 713 + ], + "type": "inline_equation", + "content": "i_{\\mathrm{th}}" + }, + { + "bbox": [ + 304, + 545, + 546, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 545, + 546, + 713 + ], + "type": "inline_equation", + "content": "j_{\\mathrm{th}}" + }, + { + "bbox": [ + 304, + 545, + 546, + 713 + ], + "type": "text", + "content": " categories. By controlling the intensity " + }, + { + "bbox": [ + 304, + 545, + 546, + 713 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 304, + 545, + 546, + 713 + ], + "type": "text", + "content": ", we can precisely" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 47, + 693, + 287, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 693, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 47, + 693, + 287, + 712 + ], + "type": "text", + "content": "2We use the prefix \"Diff-\" denotes the T2I model is fine-tuned and \"Real-\" denotes the vanilla T2I model." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17226" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 53, + 79, + 156, + 148 + ], + "blocks": [ + { + "bbox": [ + 53, + 79, + 156, + 148 + ], + "lines": [ + { + "bbox": [ + 53, + 79, + 156, + 148 + ], + "spans": [ + { + "bbox": [ + 53, + 79, + 156, + 148 + ], + "type": "image", + "image_path": "568a687fbd510c9c8c70ae5df276f80603d88666330070d2558eecaa5f929a56.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 47, + 162, + 287, + 228 + ], + "lines": [ + { + "bbox": [ + 47, + 162, + 287, + 228 + ], + "spans": [ + { + "bbox": [ + 47, + 162, + 287, + 228 + ], + "type": "text", + "content": "Figure 5. A schematic explanation of Diff-Mix's effectiveness using structural casual model [41]. " + }, + { + "bbox": [ + 47, + 162, + 287, + 228 + ], + "type": "inline_equation", + "content": "\\pmb{x}_{fg}" + }, + { + "bbox": [ + 47, + 162, + 287, + 228 + ], + "type": "text", + "content": " is the foreground that determines the real class label, " + }, + { + "bbox": [ + 47, + 162, + 287, + 228 + ], + "type": "inline_equation", + "content": "\\pmb{x}_{bg}" + }, + { + "bbox": [ + 47, + 162, + 287, + 228 + ], + "type": "text", + "content": " denotes the background. " + }, + { + "bbox": [ + 47, + 162, + 287, + 228 + ], + "type": "inline_equation", + "content": "\\pmb{x}_{fg} \\rightarrow \\pmb{z} \\rightarrow y" + }, + { + "bbox": [ + 47, + 162, + 287, + 228 + ], + "type": "text", + "content": " is the causal path that we are focusing and " + }, + { + "bbox": [ + 47, + 162, + 287, + 228 + ], + "type": "inline_equation", + "content": "\\pmb{x}_{fg} \\gets I \\rightarrow \\pmb{x}_{bg} \\rightarrow \\pmb{z} \\rightarrow y" + }, + { + "bbox": [ + 47, + 162, + 287, + 228 + ], + "type": "text", + "content": " is the backdoor path that introduces spurious relations between " + }, + { + "bbox": [ + 47, + 162, + 287, + 228 + ], + "type": "inline_equation", + "content": "\\pmb{x}_{fg}" + }, + { + "bbox": [ + 47, + 162, + 287, + 228 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 162, + 287, + 228 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 47, + 162, + 287, + 228 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 172, + 77, + 275, + 145 + ], + "blocks": [ + { + "bbox": [ + 172, + 77, + 275, + 145 + ], + "lines": [ + { + "bbox": [ + 172, + 77, + 275, + 145 + ], + "spans": [ + { + "bbox": [ + 172, + 77, + 275, + 145 + ], + "type": "image", + "image_path": "d4cd3fa5cd7d74df00c0a00006795334db3edf919efef18c0fd57749cd216e2a.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 237, + 287, + 285 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 237, + 287, + 285 + ], + "spans": [ + { + "bbox": [ + 47, + 237, + 287, + 285 + ], + "type": "text", + "content": "manage the interpolation process. When annotating the synthetic image, unlike Mixup and Cutmix, we take into account the non-linear nature of diffusion translation. Thus the annotation function is given by" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 117, + 293, + 287, + 306 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 117, + 293, + 287, + 306 + ], + "spans": [ + { + "bbox": [ + 117, + 293, + 287, + 306 + ], + "type": "interline_equation", + "content": "\\tilde {y} = (1 - s ^ {\\gamma}) y ^ {i} + s ^ {\\gamma} y ^ {j}, \\tag {3}", + "image_path": "09ce8f30d9fd28905683510822d6fc9e43202ee9f6a83f52a98d483bdf51a398.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 314, + 287, + 374 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 314, + 287, + 374 + ], + "spans": [ + { + "bbox": [ + 47, + 314, + 287, + 374 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 314, + 287, + 374 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 47, + 314, + 287, + 374 + ], + "type": "text", + "content": " is a hyperparameter introducing non-linearity. Our empirical findings indicate that a " + }, + { + "bbox": [ + 47, + 314, + 287, + 374 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 47, + 314, + 287, + 374 + ], + "type": "text", + "content": " smaller than 1 is favored. Additionally, in low-shot cases, the samples with higher confidence in the target class are preferred (see details in Sec. 4.5)." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 379, + 287, + 547 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 379, + 287, + 547 + ], + "spans": [ + { + "bbox": [ + 47, + 379, + 287, + 547 + ], + "type": "text", + "content": "Construct synthetic dataset. To construct the synthetic dataset using Diff-Mix, similar to Da-fusion [58], we adopt a randomized sampling strategy (" + }, + { + "bbox": [ + 47, + 379, + 287, + 547 + ], + "type": "inline_equation", + "content": "s \\in \\{0.5, 0.7, 0.9\\}" + }, + { + "bbox": [ + 47, + 379, + 287, + 547 + ], + "type": "text", + "content": ") for the selection of translation strength. While applying the inter-class editing, we observe that Diff-Mix tends to produce more undesirable samples compared to Diff-Aug. These undesirable samples have incomplete foreground such as fragmented bird bodies. This is caused by the intrinsic shape and pose differences among classes. To mitigate this, we introduce a simple data-cleaning approach to reduce the proportion of such problematic images. We utilize the large vision language model CLIP [43] to assess the confidence in the content, serving as the filtering criterion. Further details can be found in the supplementary materials (SMs)." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 551, + 287, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 551, + 287, + 647 + ], + "spans": [ + { + "bbox": [ + 47, + 551, + 287, + 647 + ], + "type": "text", + "content": "Analysis. We depict the core insight of Diff-Mix in Fig. 5. To eliminate the spurious correlation introduced by " + }, + { + "bbox": [ + 47, + 551, + 287, + 647 + ], + "type": "inline_equation", + "content": "\\pmb{x}_{bg}" + }, + { + "bbox": [ + 47, + 551, + 287, + 647 + ], + "type": "text", + "content": ", learning on the synthetic set with randomized " + }, + { + "bbox": [ + 47, + 551, + 287, + 647 + ], + "type": "inline_equation", + "content": "\\pmb{x}_{bg}" + }, + { + "bbox": [ + 47, + 551, + 287, + 647 + ], + "type": "text", + "content": " (background) can cut off spurious correlation, forcing the classification model to infer only from the foreground. The study in Fig. 7 (b) shows that the more diverse the background (larger the referable class number), the better the performance on the CUB test set." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 658, + 127, + 671 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 658, + 127, + 671 + ], + "spans": [ + { + "bbox": [ + 47, + 658, + 127, + 671 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 677, + 287, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 677, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 47, + 677, + 287, + 712 + ], + "type": "text", + "content": "In this section, we investigate the effectiveness of Diff-Mix in domain-specific datasets. The key questions we aim to address are as follows:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 72, + 545, + 179 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 313, + 72, + 545, + 119 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 545, + 119 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 545, + 119 + ], + "type": "text", + "content": "Q1: Can generative inter-class augmentation lead to more significant performance gains in downstream tasks compared to those intra-class augmentation methods and distillation-based methods?" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 120, + 545, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 120, + 545, + 144 + ], + "spans": [ + { + "bbox": [ + 313, + 120, + 545, + 144 + ], + "type": "text", + "content": "Q2: Is improved background diversity the secret weapon of Diff-Mix for enhancing the performance?" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 144, + 545, + 179 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 144, + 545, + 179 + ], + "spans": [ + { + "bbox": [ + 313, + 144, + 545, + 179 + ], + "type": "text", + "content": "Q3: How to choose the fine-tuning strategy and annotation strategy to boost the performance gains for the inter-class augmentation?" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 304, + 182, + 545, + 278 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 182, + 545, + 278 + ], + "spans": [ + { + "bbox": [ + 304, + 182, + 545, + 278 + ], + "type": "text", + "content": "To address Q1, we separately discuss these questions in few-shot settings in Sec. 4.1, conventional classification in Sec. 4.2, as well as long-tail classification in Sec. 4.3. Additionally, to answer Q2, we conduct a test for background robustness in Sec. 4.4 and perform an ablation study focusing on the size and diversity of synthetic data in Sec. 4.5. For Q3, we conduct an ablation study to empirically discover effective strategies for deployment in Sec. 4.5." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 288, + 436, + 300 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 288, + 436, + 300 + ], + "spans": [ + { + "bbox": [ + 306, + 288, + 436, + 300 + ], + "type": "text", + "content": "4.1. Few-shot Classification" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 307, + 545, + 474 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 307, + 545, + 474 + ], + "spans": [ + { + "bbox": [ + 304, + 307, + 545, + 474 + ], + "type": "text", + "content": "Experimental Setting. To investigate the impact of different data expansion methods, we conduct few-shot experiments on a domain-specific dataset Caltech-UCSD Birds (CUB) [61], with shot numbers of 1, 5, 10, and all. For augmentation-based methods, the synthetic dataset is constructed using various translation strengths " + }, + { + "bbox": [ + 304, + 307, + 545, + 474 + ], + "type": "inline_equation", + "content": "(s)" + }, + { + "bbox": [ + 304, + 307, + 545, + 474 + ], + "type": "text", + "content": ", specifically, " + }, + { + "bbox": [ + 304, + 307, + 545, + 474 + ], + "type": "inline_equation", + "content": "s \\in \\{0.1, 0.3, \\dots, 1.0\\}" + }, + { + "bbox": [ + 304, + 307, + 545, + 474 + ], + "type": "text", + "content": ". We expand the original training set with a multiplier of 5 and cache the synthesized dataset locally for joint training. Real data are replaced by synthetic data proportionally during training, and the replacement probability " + }, + { + "bbox": [ + 304, + 307, + 545, + 474 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 304, + 307, + 545, + 474 + ], + "type": "text", + "content": " is set as 0.1, 0.2, 0.3, and 0.5 for all-shot, 10-shot, 5-shot, and 1-shot classification, respectively. All experiments use ResNet50 with an input resolution of " + }, + { + "bbox": [ + 304, + 307, + 545, + 474 + ], + "type": "inline_equation", + "content": "224 \\times 224" + }, + { + "bbox": [ + 304, + 307, + 545, + 474 + ], + "type": "text", + "content": ". Additional details can be found in the SMs." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 479, + 545, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 479, + 545, + 552 + ], + "spans": [ + { + "bbox": [ + 304, + 479, + 545, + 552 + ], + "type": "text", + "content": "Comparison methods. To unveil the trade-off between faithfulness and diversity resulting from different expansion strategies, we compared Diff-Mix with Diff-Gen and Diff-Aug. Furthermore, we conduct experiments on expansion strategies using vanilla SD: Real-Mix, Real-Gen, and Real-Aug, where 'Real' signifies that SD is not fine-tuned." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 557, + 545, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 557, + 545, + 604 + ], + "spans": [ + { + "bbox": [ + 304, + 557, + 545, + 604 + ], + "type": "text", + "content": "Main results. To answer Q1 under the few-shot classification setting, we augment CUB using X-Mix, X-Aug, and X-Gen, where 'X' denotes 'Diff/Real' for simplicity. The results are shown in Fig. 6, and we can observe that:" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 306, + 605, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 21, + "blocks": [ + { + "bbox": [ + 306, + 605, + 545, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 605, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 306, + 605, + 545, + 689 + ], + "type": "text", + "content": "1. Diff-Mix generally outperforms the intra-class competitor X-Aug and distillation competitor X-Gen in various few-shot scenarios. It tends to achieve higher gains when the strength " + }, + { + "bbox": [ + 306, + 605, + 545, + 689 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 306, + 605, + 545, + 689 + ], + "type": "text", + "content": " is relatively large, i.e., " + }, + { + "bbox": [ + 306, + 605, + 545, + 689 + ], + "type": "inline_equation", + "content": "\\{0.5, 0.7, 0.9\\}" + }, + { + "bbox": [ + 306, + 605, + 545, + 689 + ], + "type": "text", + "content": ", where the foreground has been edited to match the target class and the background retains similarities to the reference image." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 306, + 689, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 545, + 712 + ], + "type": "text", + "content": "2. Among the Real-X methods, distillation tends to be more effective than the augmentation method when the shot" + } + ] + } + ], + "index": 20 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17227" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 72, + 165, + 159 + ], + "blocks": [ + { + "bbox": [ + 50, + 72, + 165, + 159 + ], + "lines": [ + { + "bbox": [ + 50, + 72, + 165, + 159 + ], + "spans": [ + { + "bbox": [ + 50, + 72, + 165, + 159 + ], + "type": "image", + "image_path": "a5726da6ec202f288638f381edd25aa2f9b847fb756e9a7865d7d32a5d6b4b6e.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 91, + 164, + 124, + 174 + ], + "lines": [ + { + "bbox": [ + 91, + 164, + 124, + 174 + ], + "spans": [ + { + "bbox": [ + 91, + 164, + 124, + 174 + ], + "type": "text", + "content": "(a) 1-shot" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 168, + 71, + 284, + 159 + ], + "blocks": [ + { + "bbox": [ + 168, + 71, + 284, + 159 + ], + "lines": [ + { + "bbox": [ + 168, + 71, + 284, + 159 + ], + "spans": [ + { + "bbox": [ + 168, + 71, + 284, + 159 + ], + "type": "image", + "image_path": "6086dd31f4fd8466d946c30de710cb0dc3a65677e02b903931d3625a5b80c271.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 210, + 164, + 243, + 174 + ], + "lines": [ + { + "bbox": [ + 210, + 164, + 243, + 174 + ], + "spans": [ + { + "bbox": [ + 210, + 164, + 243, + 174 + ], + "type": "text", + "content": "(b) 5-shot" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 288, + 72, + 403, + 159 + ], + "blocks": [ + { + "bbox": [ + 288, + 72, + 403, + 159 + ], + "lines": [ + { + "bbox": [ + 288, + 72, + 403, + 159 + ], + "spans": [ + { + "bbox": [ + 288, + 72, + 403, + 159 + ], + "type": "image", + "image_path": "610cb3645a14a320ac5979460207d35c8a49303b34cf50e08381810c67d9b1f3.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 328, + 164, + 363, + 174 + ], + "lines": [ + { + "bbox": [ + 328, + 164, + 363, + 174 + ], + "spans": [ + { + "bbox": [ + 328, + 164, + 363, + 174 + ], + "type": "text", + "content": "(c) 10-shot" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 406, + 72, + 525, + 159 + ], + "blocks": [ + { + "bbox": [ + 406, + 72, + 525, + 159 + ], + "lines": [ + { + "bbox": [ + 406, + 72, + 525, + 159 + ], + "spans": [ + { + "bbox": [ + 406, + 72, + 525, + 159 + ], + "type": "image", + "image_path": "f79c46a60ad004f140e8205d114bd2af63b11e8700394e0fc5355e5294c44799.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 444, + 164, + 484, + 174 + ], + "lines": [ + { + "bbox": [ + 444, + 164, + 484, + 174 + ], + "spans": [ + { + "bbox": [ + 444, + 164, + 484, + 174 + ], + "type": "text", + "content": "(d) All-shot" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 205, + 183, + 385, + 194 + ], + "lines": [ + { + "bbox": [ + 205, + 183, + 385, + 194 + ], + "spans": [ + { + "bbox": [ + 205, + 183, + 385, + 194 + ], + "type": "text", + "content": "Figure 6. Few-shot classification results on CUB." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 58, + 207, + 286, + 302 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 207, + 286, + 302 + ], + "spans": [ + { + "bbox": [ + 58, + 207, + 286, + 302 + ], + "type": "text", + "content": "number is low, but the trend reverses as the shot number increases (compare Real-Gen with Real-Aug). Real-Gen's samples even show less effective than Real-Aug " + }, + { + "bbox": [ + 58, + 207, + 286, + 302 + ], + "type": "inline_equation", + "content": "(s = 1.0)" + }, + { + "bbox": [ + 58, + 207, + 286, + 302 + ], + "type": "text", + "content": " under the all-shot case3. This indicates that the importance of faithfulness in the trade-off between faithfulness and diversity increases with the shot number. Additionally, Real-Mix exhibits consistent and stable improvement over the other two methods." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 303, + 287, + 363 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 303, + 287, + 363 + ], + "spans": [ + { + "bbox": [ + 47, + 303, + 287, + 363 + ], + "type": "text", + "content": "3. Diff-Gen consistently outperforms Real-Gen under four scenarios. Notably, Real-Gen's performance declines below that of the baseline as the shot numbers reach 10, showcasing the importance of the fine-tuning process which increases the faithfulness of synthetic samples." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 373, + 197, + 385 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 373, + 197, + 385 + ], + "spans": [ + { + "bbox": [ + 47, + 373, + 197, + 385 + ], + "type": "text", + "content": "4.2. Conventional Classification" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 47, + 393, + 287, + 572 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 393, + 287, + 572 + ], + "spans": [ + { + "bbox": [ + 47, + 393, + 287, + 572 + ], + "type": "text", + "content": "Experimental setting. To test whether Diff-Mix can further boost performance in a more challenging setting, i.e., under the all-shot scenario with high input resolution, we conduct conventional classification on five domain-specific datasets: CUB [61], Stanford Cars [27], Oxford Flowers [38], Stanford Dogs [26], and FGVC Aircraft [34]. Two backbones are employed: pretrained (ImageNet1K [49]) ResNet50 [18] with input resolution " + }, + { + "bbox": [ + 47, + 393, + 287, + 572 + ], + "type": "inline_equation", + "content": "448 \\times 448" + }, + { + "bbox": [ + 47, + 393, + 287, + 572 + ], + "type": "text", + "content": ", and pretrained (ImageNet21K) ViT-B/16 [13] with input resolution " + }, + { + "bbox": [ + 47, + 393, + 287, + 572 + ], + "type": "inline_equation", + "content": "384 \\times 384" + }, + { + "bbox": [ + 47, + 393, + 287, + 572 + ], + "type": "text", + "content": ". Label smoothing [36] is applied across all datasets with a confidence level of 0.9. For all expansion strategies, the expansion multiplier is 5 and the replacement probability " + }, + { + "bbox": [ + 47, + 393, + 287, + 572 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 47, + 393, + 287, + 572 + ], + "type": "text", + "content": " is 0.1. Besides, we use a randomized sampling strategy " + }, + { + "bbox": [ + 47, + 393, + 287, + 572 + ], + "type": "inline_equation", + "content": "(s \\in \\{0.5, 0.7, 0.9\\})" + }, + { + "bbox": [ + 47, + 393, + 287, + 572 + ], + "type": "text", + "content": " and a fixed " + }, + { + "bbox": [ + 47, + 393, + 287, + 572 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 47, + 393, + 287, + 572 + ], + "type": "text", + "content": " (0.5, and this is specific to Diff-Mix) for all datasets." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 47, + 578, + 286, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 578, + 286, + 662 + ], + "spans": [ + { + "bbox": [ + 47, + 578, + 286, + 662 + ], + "type": "text", + "content": "Comparison methods. We compare Diff-Mix with (1) Real-Filtering (RF) [19], a variation of Real-Gen that incorporates clip filtering, (2) Real-Guidance (RG) [19], which augments the dataset using intra-class image translation at low strength (" + }, + { + "bbox": [ + 47, + 578, + 286, + 662 + ], + "type": "inline_equation", + "content": "s = 0.1" + }, + { + "bbox": [ + 47, + 578, + 286, + 662 + ], + "type": "text", + "content": "), (3) Da-Fusion [58], a method that solely fine-tunes the identifier to personalize each class and employs randomized sampling strategy" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 207, + 544, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 207, + 544, + 232 + ], + "spans": [ + { + "bbox": [ + 306, + 207, + 544, + 232 + ], + "type": "inline_equation", + "content": "(s \\in \\{0.25, 0.5, 0.75, 1.0\\})" + }, + { + "bbox": [ + 306, + 207, + 544, + 232 + ], + "type": "text", + "content": ", and non-generative augmentation methods (4) CutMix [66] and (5) Mixup [68]." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 236, + 545, + 476 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 236, + 545, + 476 + ], + "spans": [ + { + "bbox": [ + 304, + 236, + 545, + 476 + ], + "type": "text", + "content": "Main results. We show the classification accuracy for different data expansion strategies in Table 1, our observations can be summarized: (1) Diff-Mix consistently demonstrates stable improvements across the majority of settings. Its average performance gain across the five datasets exceeds that of baselines employing intra-class augmentation methods (RG and Da-fusion), distillation method (RF), and non-generative data augmentation techniques (CutMix and Mixup). (2) Real-filtering, analogous to the discussion of Real-Gen, exhibits performance degradation on most datasets due to the distribution gap. (3) The combined use of Diff-Mix and CutMix often yields better performance gains. This is attributed to the distinct enhancement mechanisms of the two methods, i.e., vicinal risk minimization [11, 68] and foreground-background disentanglement. (4) Diff-Mix does not exhibit significant performance improvement in the dog dataset. We attribute this lack of improvement to the complexity of the dog dataset, which often contains multiple subjects in a single image, impeding effective foreground editing (refer to the SMs for visual examples)." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 486, + 440, + 498 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 486, + 440, + 498 + ], + "spans": [ + { + "bbox": [ + 306, + 486, + 440, + 498 + ], + "type": "text", + "content": "4.3. Long-Tail Classification" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 510, + 545, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 510, + 545, + 641 + ], + "spans": [ + { + "bbox": [ + 304, + 510, + 545, + 641 + ], + "type": "text", + "content": "Experiment setting. Following the settings of previous long-tail dataset constructions [9, 32, 39], we create two domain-specific long-tail datasets, CUB-LT [50] and Flower-LT. The imbalance factor controls the exponential distribution of the imbalanced data, where a larger value indicates a more imbalanced distribution. To leverage generative models for long-tail classification, we adopt the approach of SYNAuG [65], which uniformize the imbalanced real data distribution using synthetic data. Translation strength " + }, + { + "bbox": [ + 304, + 510, + 545, + 641 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 304, + 510, + 545, + 641 + ], + "type": "text", + "content": " (0.7) and " + }, + { + "bbox": [ + 304, + 510, + 545, + 641 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 304, + 510, + 545, + 641 + ], + "type": "text", + "content": " (0.5) are fixed for both Diff-Mix and Real-Mix." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 647, + 545, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 647, + 545, + 696 + ], + "spans": [ + { + "bbox": [ + 304, + 647, + 545, + 696 + ], + "type": "text", + "content": "Comparison methods. We compare Diff-Mix with Real-Mix, Real-Gen, the non-generative CutMix-based oversampling approach CMO [39], and its enhanced variant with two-stage deferred re-weighting [9] (CMO+DRW)." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 306, + 701, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 701, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 701, + 545, + 713 + ], + "type": "text", + "content": "Main results. We present the long-tail classification re" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 46, + 673, + 287, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 673, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 673, + 287, + 713 + ], + "type": "inline_equation", + "content": "^3" + }, + { + "bbox": [ + 46, + 673, + 287, + 713 + ], + "type": "text", + "content": "Real-Aug " + }, + { + "bbox": [ + 46, + 673, + 287, + 713 + ], + "type": "inline_equation", + "content": "(s = 1.0)" + }, + { + "bbox": [ + 46, + 673, + 287, + 713 + ], + "type": "text", + "content": " remains analogous to the reference image (slightly higher faithfulness compared to Real-Gen) because the discrete forward process cannot approximate the ideal normal distribution within a limited number of steps " + }, + { + "bbox": [ + 46, + 673, + 287, + 713 + ], + "type": "inline_equation", + "content": "(T = 25)" + }, + { + "bbox": [ + 46, + 673, + 287, + 713 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17228" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 80, + 70, + 515, + 262 + ], + "blocks": [ + { + "bbox": [ + 80, + 70, + 515, + 262 + ], + "lines": [ + { + "bbox": [ + 80, + 70, + 515, + 262 + ], + "spans": [ + { + "bbox": [ + 80, + 70, + 515, + 262 + ], + "type": "table", + "html": "
BackboneAug. MethodFT StrategyDataset
CUBAircraftFlowerCarDogAvg
ResNet50@448--86.6489.0999.2794.5487.4891.40
Cutmix[66]-87.2389.4499.2594.7387.5991.65+0.25
Mixup[68]-86.6889.4199.4094.4987.4291.48+0.08
Real-filtering [19]†X85.6088.5499.0994.5987.3091.22-0.18
Real-guidance [19]†X86.7189.0799.2594.5587.4091.59+0.19
Da-fusion [58]†TI86.3087.6499.3794.6987.3391.07-0.58
Diff-MixTI + DB87.1690.2599.5495.1287.7491.96+0.56
Diff-Mix + CutmixTI + DB87.5690.0199.4795.2187.8992.03+0.63
ViT-B/16@384--89.3783.5099.5694.2192.0691.74
Cutmix[66]-90.5283.5099.6494.8392.1392.12+0.38
Mixup[68]-90.3284.3199.7394.9892.0292.27+0.53
Real-filtering [19]†X89.4983.0799.3694.6691.9191.69-0.05
Real-guidance [19]†X89.5483.1799.5994.6592.0591.80+0.06
Da-fusion [58]†TI89.4081.8899.6194.5392.0791.50-0.24
Diff-MixTI + DB90.0584.3399.6495.0991.9992.22+0.48
Diff-Mix + CutmixTI + DB90.3585.1299.6895.2691.8992.46+0.72
", + "image_path": "a2c4e95fa07ac834c852e3707212523a01d6820e48d0df44751769a06cb2cb90.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 50, + 293, + 284, + 370 + ], + "blocks": [ + { + "bbox": [ + 96, + 269, + 495, + 281 + ], + "lines": [ + { + "bbox": [ + 96, + 269, + 495, + 281 + ], + "spans": [ + { + "bbox": [ + 96, + 269, + 495, + 281 + ], + "type": "text", + "content": "Table 1. Conventional classification in six fine-grained datasets. " + }, + { + "bbox": [ + 96, + 269, + 495, + 281 + ], + "type": "inline_equation", + "content": "{}^{ \\dagger }" + }, + { + "bbox": [ + 96, + 269, + 495, + 281 + ], + "type": "text", + "content": " ,indicates our reproduced results using SD." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 50, + 293, + 284, + 370 + ], + "lines": [ + { + "bbox": [ + 50, + 293, + 284, + 370 + ], + "spans": [ + { + "bbox": [ + 50, + 293, + 284, + 370 + ], + "type": "table", + "html": "
MethodIF=1005010
ManyMediumFewAll
CE79.1164.2813.4833.6544.8258.13
CMO [39]78.3258.5714.7832.9444.0857.62
CMO + DRW [10]78.9756.3614.6632.5746.4359.25
Real-Gen84.8865.2330.6845.8653.4361.42
Real-Mix (s=0.7)84.6366.3434.4447.7555.6762.27
Diff-Mix (s=0.7)84.0767.7936.5550.3558.1964.48
", + "image_path": "5ce0024502668b0e364b688abc84b46930516a74730afd3af8c8377575aa3727.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 50, + 398, + 284, + 475 + ], + "blocks": [ + { + "bbox": [ + 92, + 378, + 253, + 389 + ], + "lines": [ + { + "bbox": [ + 92, + 378, + 253, + 389 + ], + "spans": [ + { + "bbox": [ + 92, + 378, + 253, + 389 + ], + "type": "text", + "content": "Table 2. Long-tail classification in CUB-LT." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 50, + 398, + 284, + 475 + ], + "lines": [ + { + "bbox": [ + 50, + 398, + 284, + 475 + ], + "spans": [ + { + "bbox": [ + 50, + 398, + 284, + 475 + ], + "type": "table", + "html": "
MethodIF=1005010
ManyMediumFewAll
CE99.1994.9558.1880.4390.8795.07
CMO [39]99.2595.1967.4583.9591.4395.19
CMO+ DRW [10]99.9795.0667.3184.0792.0695.92
Real-Gen98.6495.5566.1083.5691.8495.22
Real-Mix (s=0.7)99.8796.2668.5385.1992.9696.04
Diff-Mix (s=0.7)99.2596.9878.4189.4693.8696.63
", + "image_path": "f19d20d7649b4d441f67f592f722c620ea0fcd7cad0152363ef719100d6386ec.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 503, + 287, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 503, + 287, + 635 + ], + "spans": [ + { + "bbox": [ + 46, + 503, + 287, + 635 + ], + "type": "text", + "content": "sults for CUB-LT in Table 2 and Flower-LT in 3. The observations are as follows: (1) Generative approaches exhibit superior performance in tackling imbalanced classification issues compared to CutMix-based methods (CMO and CMO+DRW). (2) Real-Mix surpasses Real-Gen in performance across various imbalance factors in two datasets. This indicates that tail classes can benefit from the enhanced diversity by leveraging the visual context of majority classes. (3) Diff-Mix generally achieves the best performance among the compared strategies, especially at the low-shot case, highlighting the importance of fine-tuning." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 646, + 183, + 658 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 646, + 183, + 658 + ], + "spans": [ + { + "bbox": [ + 47, + 646, + 183, + 658 + ], + "type": "text", + "content": "4.4. Background Robustness" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "type": "text", + "content": "In this section, we aim to evaluate Diff-Mix's robustness to background shifts, specifically, whether synthesizing more diverse samples can improve the classification model's generalizability when the background is altered. To achieve" + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 312, + 294, + 537, + 353 + ], + "blocks": [ + { + "bbox": [ + 88, + 483, + 257, + 495 + ], + "lines": [ + { + "bbox": [ + 88, + 483, + 257, + 495 + ], + "spans": [ + { + "bbox": [ + 88, + 483, + 257, + 495 + ], + "type": "text", + "content": "Table 3. Long-tail classification in Flower-LT." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 312, + 294, + 537, + 353 + ], + "lines": [ + { + "bbox": [ + 312, + 294, + 537, + 353 + ], + "spans": [ + { + "bbox": [ + 312, + 294, + 537, + 353 + ], + "type": "table", + "html": "
GroupBase.CutMixDA-fusionDiff-AugDiff-Mix
(waterbird, water)59.5062.4660.9061.8363.83
(waterbird, land)56.7060.1258.1060.1263.24
(landbird, land)73.4873.3972.9473.0475.64
(landbird, water)73.9774.7272.7773.5274.36
Avg.70.1971.2369.9070.2872.47
", + "image_path": "b98b0674ca8ecabb99c669f82edc74115567f8fdfb559649124584259afc29da.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 363, + 545, + 430 + ], + "lines": [ + { + "bbox": [ + 305, + 363, + 545, + 430 + ], + "spans": [ + { + "bbox": [ + 305, + 363, + 545, + 430 + ], + "type": "text", + "content": "Table 4. Classification results across four groups in Waterbird [48]. Waterbird is an out-of-distribution dataset for CUB, crafted by segmenting CUB's foregrounds and paste them into the scene images from Places [70]. The constructed dataset can be divided into four groups based on the composition of foregrounds (waterbird and landbird) and backgrounds (water and land)." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 446, + 545, + 615 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 446, + 545, + 615 + ], + "spans": [ + { + "bbox": [ + 304, + 446, + 545, + 615 + ], + "type": "text", + "content": "this, we utilize an out-of-distribution test set for CUB, namely Waterbird [48]. We then perform inference on the whole Waterbird set using classifiers that have been trained on either the original CUB dataset or its expanded variations. In Table 4, we present the classification accuracies across the four groups and compare Diff-Mix with other intra-class methods (Da-fusion and Diff-Aug) as well as CutMix. We observe that Diff-Mix generally outperforms its counterparts and achieves a significant performance improvement " + }, + { + "bbox": [ + 304, + 446, + 545, + 615 + ], + "type": "inline_equation", + "content": "(+6.5\\%)" + }, + { + "bbox": [ + 304, + 446, + 545, + 615 + ], + "type": "text", + "content": " in the challenging counterfactual group (waterbirds with land backgrounds). It is important to highlight that the background scenes in the Waterbird dataset are novel to CUB, requiring the classification model to have a stronger perceptual capability for the images' foregrounds." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 623, + 378, + 634 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 623, + 378, + 634 + ], + "spans": [ + { + "bbox": [ + 306, + 623, + 378, + 634 + ], + "type": "text", + "content": "4.5. Discussion" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 641, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 641, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 641, + 547, + 715 + ], + "type": "text", + "content": "In this section, we address Q2 by examining the impact of size and diversity on synthetic data. Furthermore, we perform an ablation study to assess the effects of fine-tuning strategies and training hyperparameters of Diff-Mix, which is aimed at answering Q3. Unless specified otherwise, our discussions are based on experiments conducted using CUB" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17229" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 70, + 161, + 154 + ], + "blocks": [ + { + "bbox": [ + 50, + 70, + 161, + 154 + ], + "lines": [ + { + "bbox": [ + 50, + 70, + 161, + 154 + ], + "spans": [ + { + "bbox": [ + 50, + 70, + 161, + 154 + ], + "type": "image", + "image_path": "7e18fcca6a05d2485e7b5d5a97434bd0a227da1949fc4d50e9498d6edd65f57c.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 100, + 158, + 110, + 166 + ], + "lines": [ + { + "bbox": [ + 100, + 158, + 110, + 166 + ], + "spans": [ + { + "bbox": [ + 100, + 158, + 110, + 166 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 173, + 70, + 286, + 154 + ], + "blocks": [ + { + "bbox": [ + 173, + 70, + 286, + 154 + ], + "lines": [ + { + "bbox": [ + 173, + 70, + 286, + 154 + ], + "spans": [ + { + "bbox": [ + 173, + 70, + 286, + 154 + ], + "type": "image", + "image_path": "ad491210bb609c2744334dabc3e1cdbb16868909a409daa40df26f34ebb4abef.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 225, + 158, + 235, + 166 + ], + "lines": [ + { + "bbox": [ + 225, + 158, + 235, + 166 + ], + "spans": [ + { + "bbox": [ + 225, + 158, + 235, + 166 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 47, + 176, + 287, + 198 + ], + "lines": [ + { + "bbox": [ + 47, + 176, + 287, + 198 + ], + "spans": [ + { + "bbox": [ + 47, + 176, + 287, + 198 + ], + "type": "text", + "content": "Figure 7. Comparison of results across various (a) synthetic data sizes and (b) numbers of referable classes for each target class." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 217, + 269, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 217, + 269, + 228 + ], + "spans": [ + { + "bbox": [ + 47, + 217, + 269, + 228 + ], + "type": "text", + "content": "with ResNet50, where inputs are resized to " + }, + { + "bbox": [ + 47, + 217, + 269, + 228 + ], + "type": "inline_equation", + "content": "224 \\times 224" + }, + { + "bbox": [ + 47, + 217, + 269, + 228 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 235, + 287, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 235, + 287, + 415 + ], + "spans": [ + { + "bbox": [ + 46, + 235, + 287, + 415 + ], + "type": "text", + "content": "Size and diversity of synthetic Data. The relationship between performance gain and the size of synthetic data is depicted in Fig. 7 (a), where a classification model was trained with synthetic data of varying sizes. We observe a monotonically increasing trend as the multiplier for the synthetic dataset ranges from 2 to 10. Ideally, the combination choices of " + }, + { + "bbox": [ + 46, + 235, + 287, + 415 + ], + "type": "inline_equation", + "content": "(x_{i},y_{j})" + }, + { + "bbox": [ + 46, + 235, + 287, + 415 + ], + "type": "text", + "content": " are in the order of " + }, + { + "bbox": [ + 46, + 235, + 287, + 415 + ], + "type": "inline_equation", + "content": "N|D^{train}|" + }, + { + "bbox": [ + 46, + 235, + 287, + 415 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 46, + 235, + 287, + 415 + ], + "type": "inline_equation", + "content": "N = 200" + }, + { + "bbox": [ + 46, + 235, + 287, + 415 + ], + "type": "text", + "content": " for CUB). Furthermore, we limit the number of referable classes for each target class, which means the number of referable backgrounds decreases, resulting in a synthetic dataset of relatively lower diversity. The results are shown in Fig. 7 (b), and we observe a consistent improvement in performance as the number of referable classes increases. These results consistently underscore the critical role of background diversity introduced by Diff-Mix." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 420, + 287, + 589 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 420, + 287, + 589 + ], + "spans": [ + { + "bbox": [ + 46, + 420, + 287, + 589 + ], + "type": "text", + "content": "Impact of fine-tuning strategy. Here we compare three different fine-tuning strategies: TI, DB, and the combined " + }, + { + "bbox": [ + 46, + 420, + 287, + 589 + ], + "type": "inline_equation", + "content": "\\mathrm{TI + DB}" + }, + { + "bbox": [ + 46, + 420, + 287, + 589 + ], + "type": "text", + "content": ". All strategies share the same fine-tuning hyperparameters and training steps (35000). To evaluate the distribution gap, we compute the FID score [20] of synthesized images (Diff-Gen) with the training set. As illustrated in Table 5, we observe that both " + }, + { + "bbox": [ + 46, + 420, + 287, + 589 + ], + "type": "inline_equation", + "content": "\\mathrm{TI + DB}" + }, + { + "bbox": [ + 46, + 420, + 287, + 589 + ], + "type": "text", + "content": " and TI have lower FID scores than DB. This can be attributed to the fact that semantic proximity impedes the convergence process. Additionally, while using TI alone results in a relatively low FID score, the improvements in performance are limited. This limitation stems from TI's inability to accurately reconstruct detailed concept (foreground) information, as it is primarily fine-tuned at the semantic level [28]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "text", + "content": "Annotation function. In this section, we discuss the impact of the choices of the translation strength " + }, + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "text", + "content": " and the non-linear factor " + }, + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "text", + "content": " in Eq. 3. As shown in Table 6, we observe that as the translation strength decreases, the optimal value for " + }, + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "text", + "content": " also decreases, which underscores the nonlinearity of Diff-Mix. The comparison between the 5-shot and all-shot settings indicates that the model tends to prefer a more diverse synthetic dataset when the number of training shots is small (" + }, + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "inline_equation", + "content": "s = 0.9" + }, + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "text", + "content": " for 5-shot, " + }, + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "inline_equation", + "content": "s = 0.7" + }, + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "text", + "content": " for all-shot). Besides, a larger confidence in the target class is pre" + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 314, + 70, + 537, + 118 + ], + "blocks": [ + { + "bbox": [ + 314, + 70, + 537, + 118 + ], + "lines": [ + { + "bbox": [ + 314, + 70, + 537, + 118 + ], + "spans": [ + { + "bbox": [ + 314, + 70, + 537, + 118 + ], + "type": "table", + "html": "
BaselineTIDBTI + DB
5-shotFID (Diff-Gen)-18.2619.5518.43
Acc. (Diff-Mix)50.9057.6456.1159.41
All-shotFID (Diff-Gen)-14.1314.6413.99
Acc. (Diff-Mix)81.6081.8681.9982.85
", + "image_path": "9db0d82779f31e1fe2dc3d03f45e62f7c620dbf500eb742d8b7215455c965234.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 315, + 169, + 537, + 249 + ], + "blocks": [ + { + "bbox": [ + 305, + 125, + 544, + 159 + ], + "lines": [ + { + "bbox": [ + 305, + 125, + 544, + 159 + ], + "spans": [ + { + "bbox": [ + 305, + 125, + 544, + 159 + ], + "type": "text", + "content": "Table 5. Comparison of distribution gap and classification accuracy across three fine-tuning strategies. TI solely fine-tunes the identifier, and DB solely fine-tunes the U-Net, and " + }, + { + "bbox": [ + 305, + 125, + 544, + 159 + ], + "type": "inline_equation", + "content": "\\mathrm{TI + DB}" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 315, + 169, + 537, + 249 + ], + "lines": [ + { + "bbox": [ + 315, + 169, + 537, + 249 + ], + "spans": [ + { + "bbox": [ + 315, + 169, + 537, + 249 + ], + "type": "table", + "html": "
γ5-shotAll-shot
s = 0.5s = 0.7s = 0.9s = 0.5s = 0.7s = 0.9
1.5-4.50-0.31+10.30-1.08+0.92+0.90
1.0-2.31+2.99+10.79+0.25+1.14+0.90
0.5+2.35+8.44+11.01+0.92+1.30+0.86
0.3+3.94+9.41+11.15+0.97+1.24+0.69
0.1+6.18+9.86+10.84+0.50+0.88+0.84
0.0+5.25+9.41+11.06+0.38+0.63+0.54
", + "image_path": "0f3702881ca72de16a29f76c6f0043eec4ec0c1182442166ac1bfdf7d8148c85.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 256, + 545, + 301 + ], + "lines": [ + { + "bbox": [ + 305, + 256, + 545, + 301 + ], + "spans": [ + { + "bbox": [ + 305, + 256, + 545, + 301 + ], + "type": "text", + "content": "Table 6. Comparison of performance gain across various " + }, + { + "bbox": [ + 305, + 256, + 545, + 301 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 305, + 256, + 545, + 301 + ], + "type": "text", + "content": " and translation strength " + }, + { + "bbox": [ + 305, + 256, + 545, + 301 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 305, + 256, + 545, + 301 + ], + "type": "text", + "content": ". Lower " + }, + { + "bbox": [ + 305, + 256, + 545, + 301 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 305, + 256, + 545, + 301 + ], + "type": "text", + "content": " indicates a higher confidence over target class, e.g. (" + }, + { + "bbox": [ + 305, + 256, + 545, + 301 + ], + "type": "inline_equation", + "content": "\\gamma = 0.1" + }, + { + "bbox": [ + 305, + 256, + 545, + 301 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 305, + 256, + 545, + 301 + ], + "type": "inline_equation", + "content": "s = 0.7" + }, + { + "bbox": [ + 305, + 256, + 545, + 301 + ], + "type": "text", + "content": ") results in " + }, + { + "bbox": [ + 305, + 256, + 545, + 301 + ], + "type": "inline_equation", + "content": "0.04y_{i} + 0.96y_{j}" + }, + { + "bbox": [ + 305, + 256, + 545, + 301 + ], + "type": "text", + "content": " and (" + }, + { + "bbox": [ + 305, + 256, + 545, + 301 + ], + "type": "inline_equation", + "content": "\\gamma = 0.5" + }, + { + "bbox": [ + 305, + 256, + 545, + 301 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 305, + 256, + 545, + 301 + ], + "type": "inline_equation", + "content": "s = 0.7" + }, + { + "bbox": [ + 305, + 256, + 545, + 301 + ], + "type": "text", + "content": ") results in " + }, + { + "bbox": [ + 305, + 256, + 545, + 301 + ], + "type": "inline_equation", + "content": "0.16y_{i} + 0.84y_{j}" + }, + { + "bbox": [ + 305, + 256, + 545, + 301 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 305, + 322, + 545, + 395 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 322, + 545, + 395 + ], + "spans": [ + { + "bbox": [ + 305, + 322, + 545, + 395 + ], + "type": "text", + "content": "ferred when the shot number is small (" + }, + { + "bbox": [ + 305, + 322, + 545, + 395 + ], + "type": "inline_equation", + "content": "\\gamma = 0.1" + }, + { + "bbox": [ + 305, + 322, + 545, + 395 + ], + "type": "text", + "content": " for 5-shot, " + }, + { + "bbox": [ + 305, + 322, + 545, + 395 + ], + "type": "inline_equation", + "content": "\\gamma = 0.5" + }, + { + "bbox": [ + 305, + 322, + 545, + 395 + ], + "type": "text", + "content": " for all-shot). A possible explanation is that the all-shot setting is less tolerant towards unrealistic images, as discussed in Section 6. Empirically, we recommend choosing a higher translation strength (" + }, + { + "bbox": [ + 305, + 322, + 545, + 395 + ], + "type": "inline_equation", + "content": "s \\in [0.5, 0.7, 0.9]" + }, + { + "bbox": [ + 305, + 322, + 545, + 395 + ], + "type": "text", + "content": ") and a smaller " + }, + { + "bbox": [ + 305, + 322, + 545, + 395 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 305, + 322, + 545, + 395 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 305, + 322, + 545, + 395 + ], + "type": "inline_equation", + "content": "\\gamma \\in [0.1, 0.3, 0.5]" + }, + { + "bbox": [ + 305, + 322, + 545, + 395 + ], + "type": "text", + "content": ") as a conservative option." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 404, + 378, + 417 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 404, + 378, + 417 + ], + "spans": [ + { + "bbox": [ + 306, + 404, + 378, + 417 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 425, + 545, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 425, + 545, + 544 + ], + "spans": [ + { + "bbox": [ + 304, + 425, + 545, + 544 + ], + "type": "text", + "content": "In this work, we investigate two pivotal aspects, faithfulness and diversity, that are critical for the current state-of-the-art text-to-image generative models to enhance image classification tasks. To achieve a more effective balance between these two aspects, we propose an inter-class augmentation strategy that leverages Stable Diffusion. This method enables generative models to produce a greater diversity of samples by editing images from other classes and shows consistent performance improvement across various classification tasks." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 555, + 416, + 569 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 555, + 416, + 569 + ], + "spans": [ + { + "bbox": [ + 306, + 555, + 416, + 569 + ], + "type": "text", + "content": "6. Acknowledgement" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 305, + 575, + 547, + 623 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 575, + 547, + 623 + ], + "spans": [ + { + "bbox": [ + 305, + 575, + 547, + 623 + ], + "type": "text", + "content": "This research is mainly supported by the National Natural Science Foundation of China (92270114). This work is also partially supported by the National Key R&D Program of China under Grant 2021ZD0112801." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17230" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 91, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 124 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 124 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 124 + ], + "type": "text", + "content": "[1] Antreas Antoniou, Amos Storkey, and Harrison Edwards. Data augmentation generative adversarial networks. arXiv preprint arXiv:1711.04340, 2017. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 124, + 288, + 168 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 124, + 288, + 168 + ], + "spans": [ + { + "bbox": [ + 53, + 124, + 288, + 168 + ], + "type": "text", + "content": "[2] Shekoofeh Azizi, Simon Kornblith, Chitwan Sahara, Mohammad Norouzi, and David J Fleet. Synthetic data from diffusion models improves imagenet classification. arXiv preprint arXiv:2304.08466, 2023.1, 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 168, + 288, + 212 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 168, + 288, + 212 + ], + "spans": [ + { + "bbox": [ + 53, + 168, + 288, + 212 + ], + "type": "text", + "content": "[3] Haoyue Bai, Ceyuan Yang, Yinghao Xu, S-H Gary Chan, and Bolei Zhou. Improving out-of-distribution robustness of classifiers via generative interpolation. arXiv preprint arXiv:2307.12219, 2023. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 213, + 288, + 247 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 213, + 288, + 247 + ], + "spans": [ + { + "bbox": [ + 53, + 213, + 288, + 247 + ], + "type": "text", + "content": "[4] Hritik Bansal and Aditya Grover. Leaving reality to imagination: Robust classification via generated datasets. arXiv preprint arXiv:2302.02503, 2023. 1, 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 247, + 288, + 291 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 247, + 288, + 291 + ], + "spans": [ + { + "bbox": [ + 53, + 247, + 288, + 291 + ], + "type": "text", + "content": "[5] Christopher Beckham, Sina Honari, Vikas Verma, Alex M Lamb, Farnoosh Ghadiri, R Devon Hjelm, Yoshua Bengio, and Chris Pal. On adversarial mixup resynthesis. Advances in neural information processing systems, 32, 2019. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 291, + 288, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 291, + 288, + 357 + ], + "spans": [ + { + "bbox": [ + 53, + 291, + 288, + 357 + ], + "type": "text", + "content": "[6] Ankan Kumar Bhunia, Salman Khan, Hisham Cholakkal, Rao Muhammad Anwer, Jorma Laaksonen, Mubarak Shah, and Fahad Shahbaz Khan. Person image synthesis via denoising diffusion model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5968-5976, 2023. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 357, + 288, + 391 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 357, + 288, + 391 + ], + "spans": [ + { + "bbox": [ + 53, + 357, + 288, + 391 + ], + "type": "text", + "content": "[7] Andrew Brock, Jeff Donahue, and Karen Simonyan. Large scale gan training for high fidelity natural image synthesis. arXiv preprint arXiv:1809.11096, 2018. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 392, + 288, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 392, + 288, + 445 + ], + "spans": [ + { + "bbox": [ + 53, + 392, + 288, + 445 + ], + "type": "text", + "content": "[8] Tim Brooks, Aleksander Holynski, and Alexei A Efros. Instructpix2pix: Learning to follow image editing instructions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18392-18402, 2023. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 447, + 288, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 447, + 288, + 491 + ], + "spans": [ + { + "bbox": [ + 53, + 447, + 288, + 491 + ], + "type": "text", + "content": "[9] Kaidi Cao, Colin Wei, Adrien Gaidon, Nikos Arechiga, and Tengyu Ma. Learning imbalanced datasets with label-distribution-aware margin loss. Advances in neural information processing systems, 32, 2019. 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 491, + 288, + 535 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 491, + 288, + 535 + ], + "spans": [ + { + "bbox": [ + 47, + 491, + 288, + 535 + ], + "type": "text", + "content": "[10] Kaidi Cao, Colin Wei, Adrien Gaidon, Nikos Arechiga, and Tengyu Ma. Learning imbalanced datasets with label-distribution-aware margin loss. Advances in neural information processing systems, 32, 2019. 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 536, + 288, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 536, + 288, + 568 + ], + "spans": [ + { + "bbox": [ + 47, + 536, + 288, + 568 + ], + "type": "text", + "content": "[11] Olivier Chapelle, Jason Weston, Léon Bottou, and Vladimir Vapnik. Vicinal risk minimization. Advances in neural information processing systems, 13, 2000. 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 47, + 569, + 288, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 569, + 288, + 602 + ], + "spans": [ + { + "bbox": [ + 47, + 569, + 288, + 602 + ], + "type": "text", + "content": "[12] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in neural information processing systems, 34:8780-8794, 2021. 1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 47, + 602, + 288, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 602, + 288, + 668 + ], + "spans": [ + { + "bbox": [ + 47, + 602, + 288, + 668 + ], + "type": "text", + "content": "[13] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 6" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 47, + 669, + 288, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 669, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 669, + 288, + 713 + ], + "type": "text", + "content": "[14] M. Everingham, S. M. A. Eslami, L. Van Gool, C. K. I. Williams, J. Winn, and A. Zisserman. The pascal visual object classes challenge: A retrospective. International Journal of Computer Vision, 111(1):98-136, 2015. 1, 2, 5" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 713 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 307, + 73, + 547, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 547, + 127 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 547, + 127 + ], + "type": "text", + "content": "[15] Rinon Gal, Yuval Alaluf, Yuval Atzmon, Or Patashnik, Amit H Bermano, Gal Chechik, and Daniel Cohen-Or. An image is worth one word: Personalizing text-to-image generation using textual inversion. arXiv preprint arXiv:2208.01618, 2022. 2, 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 128, + 547, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 128, + 547, + 172 + ], + "spans": [ + { + "bbox": [ + 307, + 128, + 547, + 172 + ], + "type": "text", + "content": "[16] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. Advances in neural information processing systems, 27, 2014. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 172, + 547, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 172, + 547, + 217 + ], + "spans": [ + { + "bbox": [ + 307, + 172, + 547, + 217 + ], + "type": "text", + "content": "[17] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. Advances in neural information processing systems, 27, 2014. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 217, + 547, + 260 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 217, + 547, + 260 + ], + "spans": [ + { + "bbox": [ + 307, + 217, + 547, + 260 + ], + "type": "text", + "content": "[18] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 261, + 547, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 261, + 547, + 304 + ], + "spans": [ + { + "bbox": [ + 307, + 261, + 547, + 304 + ], + "type": "text", + "content": "[19] Ruifei He, Shuyang Sun, Xin Yu, Chuhui Xue, Wenqing Zhang, Philip Torr, Song Bai, and Xiaojuan Qi. Is synthetic data from generative models ready for image recognition? arXiv preprint arXiv:2210.07574, 2022. 2, 6, 7, 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 305, + 547, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 305, + 547, + 358 + ], + "spans": [ + { + "bbox": [ + 307, + 305, + 547, + 358 + ], + "type": "text", + "content": "[20] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 8" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 360, + 547, + 393 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 360, + 547, + 393 + ], + "spans": [ + { + "bbox": [ + 307, + 360, + 547, + 393 + ], + "type": "text", + "content": "[21] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 393, + 547, + 449 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 393, + 547, + 449 + ], + "spans": [ + { + "bbox": [ + 307, + 393, + 547, + 449 + ], + "type": "text", + "content": "[22] Jonathan Ho, William Chan, Chitwan Sahara, Jay Whang, Ruiqi Gao, Alexey Gritsenko, Diederik P Kingma, Ben Poole, Mohammad Norouzi, David J Fleet, et al. Imagen video: High definition video generation with diffusion models. arXiv preprint arXiv:2210.02303, 2022. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 449, + 547, + 492 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 449, + 547, + 492 + ], + "spans": [ + { + "bbox": [ + 307, + 449, + 547, + 492 + ], + "type": "text", + "content": "[23] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 4" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 492, + 547, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 492, + 547, + 547 + ], + "spans": [ + { + "bbox": [ + 307, + 492, + 547, + 547 + ], + "type": "text", + "content": "[24] Phillip Isola, Jun-Yan Zhu, Tinghui Zhou, and Alexei A Efros. Image-to-image translation with conditional adversarial networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1125-1134, 2017. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 548, + 547, + 592 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 548, + 547, + 592 + ], + "spans": [ + { + "bbox": [ + 307, + 548, + 547, + 592 + ], + "type": "text", + "content": "[25] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4401-4410, 2019. 1, 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 592, + 547, + 647 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 592, + 547, + 647 + ], + "spans": [ + { + "bbox": [ + 307, + 592, + 547, + 647 + ], + "type": "text", + "content": "[26] Aditya Khosla, Nityananda Jayadevaprakash, Bangpeng Yao, and Li Fei-Fei. Novel dataset for fine-grained image categorization. In First Workshop on Fine-Grained Visual Categorization, IEEE Conference on Computer Vision and Pattern Recognition, Colorado Springs, CO, 2011. 6, 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 647, + 547, + 691 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 647, + 547, + 691 + ], + "spans": [ + { + "bbox": [ + 307, + 647, + 547, + 691 + ], + "type": "text", + "content": "[27] Jonathan Krause, Michael Stark, Jia Deng, and Li Fei-Fei. 3d object representations for fine-grained categorization. In Proceedings of the IEEE international conference on computer vision workshops, pages 554–561, 2013. 6, 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 691, + 547, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 691, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 691, + 547, + 713 + ], + "type": "text", + "content": "[28] Nupur Kumari, Bingliang Zhang, Richard Zhang, Eli Shechtman, and Jun-Yan Zhu. Multi-concept customization" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17231" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 66, + 72, + 287, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 72, + 287, + 106 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 287, + 106 + ], + "type": "text", + "content": "of text-to-image diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1931-1941, 2023. 2, 3, 8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 107, + 287, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 107, + 287, + 150 + ], + "spans": [ + { + "bbox": [ + 48, + 107, + 287, + 150 + ], + "type": "text", + "content": "[29] Alexander C Li, Mihir Prabhudesai, Shivam Duggal, Ellis Brown, and Deepak Pathak. Your diffusion model is secretly a zero-shot classifier. arXiv preprint arXiv:2303.16203, 2023. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 152, + 287, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 152, + 287, + 185 + ], + "spans": [ + { + "bbox": [ + 48, + 152, + 287, + 185 + ], + "type": "text", + "content": "[30] Yicong Li, Xiang Wang, Junbin Xiao, Wei Ji, and Tat seng Chua. Invariant grounding for video question answering. In CVPR, 2022. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 186, + 287, + 219 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 186, + 287, + 219 + ], + "spans": [ + { + "bbox": [ + 48, + 186, + 287, + 219 + ], + "type": "text", + "content": "[31] Shanchuan Lin, Bingchen Liu, Jiashi Li, and Xiao Yang. Common diffusion noise schedules and sample steps are flawed. arXiv preprint arXiv:2305.08891, 2023. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 220, + 287, + 275 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 220, + 287, + 275 + ], + "spans": [ + { + "bbox": [ + 48, + 220, + 287, + 275 + ], + "type": "text", + "content": "[32] Ziwei Liu, Zhongqi Miao, Xiaohang Zhan, Jiayun Wang, Boqing Gong, and Stella X Yu. Large-scale long-tailed recognition in an open world. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2537-2546, 2019. 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 276, + 287, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 276, + 287, + 330 + ], + "spans": [ + { + "bbox": [ + 48, + 276, + 287, + 330 + ], + "type": "text", + "content": "[33] Cheng Lu, Yuhao Zhou, Fan Bao, Jianfei Chen, Chongxuan Li, and Jun Zhu. Dpm-solver: A fast ode solver for diffusion probabilistic model sampling in around 10 steps. Advances in Neural Information Processing Systems, 35:5775-5787, 2022. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 331, + 287, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 331, + 287, + 374 + ], + "spans": [ + { + "bbox": [ + 48, + 331, + 287, + 374 + ], + "type": "text", + "content": "[34] Subhransu Maji, Esa Rahtu, Juho Kannala, Matthew Blaschko, and Andrea Vedaldi. Fine-grained visual classification of aircraft. arXiv preprint arXiv:1306.5151, 2013. 6, 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 376, + 287, + 421 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 376, + 287, + 421 + ], + "spans": [ + { + "bbox": [ + 48, + 376, + 287, + 421 + ], + "type": "text", + "content": "[35] Chenlin Meng, Yutong He, Yang Song, Jiaming Song, Jia-jun Wu, Jun-Yan Zhu, and Stefano Ermon. Sdedit: Guided image synthesis and editing with stochastic differential equations. arXiv preprint arXiv:2108.01073, 2021. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 422, + 287, + 455 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 422, + 287, + 455 + ], + "spans": [ + { + "bbox": [ + 48, + 422, + 287, + 455 + ], + "type": "text", + "content": "[36] Rafael Müller, Simon Kornblith, and Geoffrey E Hinton. When does label smoothing help? Advances in neural information processing systems, 32, 2019. 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 456, + 287, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 456, + 287, + 510 + ], + "spans": [ + { + "bbox": [ + 48, + 456, + 287, + 510 + ], + "type": "text", + "content": "[37] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741, 2021. 1, 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 511, + 287, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 511, + 287, + 555 + ], + "spans": [ + { + "bbox": [ + 48, + 511, + 287, + 555 + ], + "type": "text", + "content": "[38] Maria-Elena Nilsback and Andrew Zisserman. Automated flower classification over a large number of classes. In 2008 Sixth Indian conference on computer vision, graphics & image processing, pages 722-729. IEEE, 2008. 6, 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 556, + 287, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 556, + 287, + 621 + ], + "spans": [ + { + "bbox": [ + 48, + 556, + 287, + 621 + ], + "type": "text", + "content": "[39] Seulki Park, Youngkyu Hong, Byeongho Heo, Sangdoo Yun, and Jin Young Choi. The majority can help the minority: Context-rich minority oversampling for long-tailed classification. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6887-6896, 2022. 6, 7, 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 624, + 287, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 624, + 287, + 666 + ], + "spans": [ + { + "bbox": [ + 48, + 624, + 287, + 666 + ], + "type": "text", + "content": "[40] Maitreya Patel, Tejas Gokhale, Chitta Baral, and Yezhou Yang. Conceptbed: Evaluating concept learning abilities of text-to-image diffusion models. arXiv preprint arXiv:2306.04695, 2023. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 669, + 287, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 669, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 669, + 287, + 689 + ], + "type": "text", + "content": "[41] Judea Pearl. Causal inference in statistics: An overview. 2009. 5" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "type": "text", + "content": "[42] Zeju Qiu, Weiyang Liu, Haiwen Feng, Yuxuan Xue, Yao Feng, Zhen Liu, Dan Zhang, Adrian Weller, and Bernhard" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 106 + ], + "type": "text", + "content": "Schölkopf. Controlling text-to-image diffusion by orthogonal finetuning. Advances in Neural Information Processing Systems, 36, 2024. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 107, + 545, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 107, + 545, + 171 + ], + "spans": [ + { + "bbox": [ + 307, + 107, + 545, + 171 + ], + "type": "text", + "content": "[43] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 5, 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 173, + 545, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 173, + 545, + 205 + ], + "spans": [ + { + "bbox": [ + 307, + 173, + 545, + 205 + ], + "type": "text", + "content": "[44] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models, 2021. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 206, + 545, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 206, + 545, + 271 + ], + "spans": [ + { + "bbox": [ + 307, + 206, + 545, + 271 + ], + "type": "text", + "content": "[45] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical Image Computing and Computer-Assisted Intervention-MICCAI 2015: 18th International Conference, Munich, Germany, October 5-9, 2015, Proceedings, Part III 18, pages 234-241. Springer, 2015. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 272, + 545, + 337 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 272, + 545, + 337 + ], + "spans": [ + { + "bbox": [ + 307, + 272, + 545, + 337 + ], + "type": "text", + "content": "[46] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22500-22510, 2023. 2, 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 338, + 545, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 338, + 545, + 392 + ], + "spans": [ + { + "bbox": [ + 307, + 338, + 545, + 392 + ], + "type": "text", + "content": "[47] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, et al. Imagenet large scale visual recognition challenge. International journal of computer vision, 115:211-252, 2015. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 394, + 545, + 446 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 394, + 545, + 446 + ], + "spans": [ + { + "bbox": [ + 307, + 394, + 545, + 446 + ], + "type": "text", + "content": "[48] Shiori Sagawa, Pang Wei Koh, Tatsunori B Hashimoto, and Percy Liang. Distributionally robust neural networks for group shifts: On the importance of regularization for worst-case generalization. arXiv preprint arXiv:1911.08731, 2019. 7" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 449, + 545, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 449, + 545, + 525 + ], + "spans": [ + { + "bbox": [ + 307, + 449, + 545, + 525 + ], + "type": "text", + "content": "[49] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, Seyed Kamyar Seyed Ghasemipour, Burcu Karagol Ayan, S. Sara Mahdavi, Rapha Gontijo Lopes, Tim Salimans, Jonathan Ho, David J Fleet, and Mohammad Norouzi. Photorealistic text-to-image diffusion models with deep language understanding, 2022. 1, 2, 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 526, + 545, + 570 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 526, + 545, + 570 + ], + "spans": [ + { + "bbox": [ + 307, + 526, + 545, + 570 + ], + "type": "text", + "content": "[50] Dvir Samuel, Yuval Atzmon, and Gal Chechik. From generalized zero-shot learning to long-tail with class descriptors. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 286-295, 2021. 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 571, + 545, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 571, + 545, + 613 + ], + "spans": [ + { + "bbox": [ + 307, + 571, + 545, + 613 + ], + "type": "text", + "content": "[51] Veit Sandfort, Ke Yan, Perry J Pickhardt, and Ronald M Summers. Data augmentation using generative adversarial networks (cyclegan) to improve generalizability in ct segmentation tasks. Scientific reports, 9(1):16884, 2019. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 614, + 545, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 614, + 545, + 669 + ], + "spans": [ + { + "bbox": [ + 307, + 614, + 545, + 669 + ], + "type": "text", + "content": "[52] Mert Bulent Sariyildiz, Karteek Alahari, Diane Larlus, and Yannis Kalantidis. Fake it till you make it: Learning transferable representations from synthetic imagenet clones. In CVPR 2023-IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023. 1" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 670, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 670, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 670, + 545, + 713 + ], + "type": "text", + "content": "[53] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "17232" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 73, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 66, + 73, + 286, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 73, + 286, + 95 + ], + "spans": [ + { + "bbox": [ + 66, + 73, + 286, + 95 + ], + "type": "text", + "content": "next generation image-text models. Advances in Neural Information Processing Systems, 35:25278-25294, 2022. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 95, + 287, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 95, + 287, + 139 + ], + "spans": [ + { + "bbox": [ + 48, + 95, + 287, + 139 + ], + "type": "text", + "content": "[54] Jordan Shipard, Arnold Wiliem, Kien Nguyen Thanh, Wei Xiang, and Clinton Fookes. Boosting zero-shot classification with synthetic data diversity via stable diffusion. arXiv preprint arXiv:2302.03298, 2023. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 140, + 287, + 193 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 140, + 287, + 193 + ], + "spans": [ + { + "bbox": [ + 48, + 140, + 287, + 193 + ], + "type": "text", + "content": "[55] Uriel Singer, Adam Polyak, Thomas Hayes, Xi Yin, Jie An, Songyang Zhang, Qiyuan Hu, Harry Yang, Oron Ashual, Oran Gafni, et al. Make-a-video: Text-to-video generation without text-video data. arXiv preprint arXiv:2209.14792, 2022. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 194, + 287, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 194, + 287, + 248 + ], + "spans": [ + { + "bbox": [ + 48, + 194, + 287, + 248 + ], + "type": "text", + "content": "[56] Shikun Sun, Longhui Wei, Zhicai Wang, Zixuan Wang, Junliang Xing, Jia Jia, and Qi Tian. Inner classifier-free guidance and its taylor expansion for diffusion models. In The Twelfth International Conference on Learning Representations, 2023. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 250, + 287, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 250, + 287, + 293 + ], + "spans": [ + { + "bbox": [ + 48, + 250, + 287, + 293 + ], + "type": "text", + "content": "[57] Yonglong Tian, Lijie Fan, Phillip Isola, Huiwen Chang, and Dilip Krishnan. Stablerep: Synthetic images from text-to-image models make strong visual representation learners. arXiv preprint arXiv:2306.00984, 2023. 1, 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 294, + 287, + 337 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 294, + 287, + 337 + ], + "spans": [ + { + "bbox": [ + 48, + 294, + 287, + 337 + ], + "type": "text", + "content": "[58] Brandon Trabucco, Kyle Doherty, Max Gurinas, and Ruslan Salakhutdinov. Effective data augmentation with diffusion models. arXiv preprint arXiv:2302.07944, 2023. 2, 5, 6, 7, 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 338, + 287, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 338, + 287, + 392 + ], + "spans": [ + { + "bbox": [ + 48, + 338, + 287, + 392 + ], + "type": "text", + "content": "[59] Narek Tumanyan, Michal Geyer, Shai Bagon, and Tali Dekel. Plug-and-play diffusion features for text-driven image-to-image translation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1921-1930, 2023. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 394, + 287, + 448 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 394, + 287, + 448 + ], + "spans": [ + { + "bbox": [ + 48, + 394, + 287, + 448 + ], + "type": "text", + "content": "[60] Patrick von Platen, Suraj Patil, Anton Lozhkov, Pedro Cuenca, Nathan Lambert, Kashif Rasul, Mishig Davaadorj, and Thomas Wolf. Diffusers: State-of-the-art diffusion models. https://github.com/huggingface/ diffusers, 2022.3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 449, + 287, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 449, + 287, + 480 + ], + "spans": [ + { + "bbox": [ + 48, + 449, + 287, + 480 + ], + "type": "text", + "content": "[61] Catherine Wah, Steve Branson, Peter Welinder, Pietro Perona, and Serge Belongie. The caltech-ucsd birds-200-2011 dataset. 2011. 5, 6, 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 482, + 287, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 482, + 287, + 536 + ], + "spans": [ + { + "bbox": [ + 48, + 482, + 287, + 536 + ], + "type": "text", + "content": "[62] Zhicai Wang, Yanbin Hao, Tingting Mu, Ouxiang Li, Shuo Wang, and Xiangnan He. Bi-directional distribution alignment for transductive zero-shot learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19893-19902, 2023. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 537, + 287, + 581 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 537, + 287, + 581 + ], + "spans": [ + { + "bbox": [ + 48, + 537, + 287, + 581 + ], + "type": "text", + "content": "[63] Daniel Watson, William Chan, Ricardo Martin-Brualla, Jonathan Ho, Andrea Tagliasacchi, and Mohammad Norouzi. Novel view synthesis with diffusion models. arXiv preprint arXiv:2210.04628, 2022. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 582, + 287, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 582, + 287, + 635 + ], + "spans": [ + { + "bbox": [ + 48, + 582, + 287, + 635 + ], + "type": "text", + "content": "[64] Zongze Wu, Dani Lischinski, and Eli Shechtman. Stylespace analysis: Disentangled controls for stylegan image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12863-12872, 2021. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 636, + 287, + 680 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 636, + 287, + 680 + ], + "spans": [ + { + "bbox": [ + 48, + 636, + 287, + 680 + ], + "type": "text", + "content": "[65] Moon Ye-Bin, Nam Hyeon-Woo, Wonseok Choi, Nayeong Kim, Suha Kwak, and Tae-Hyun Oh. Exploiting synthetic data for data imbalance problems: Baselines from a data perspective. arXiv preprint arXiv:2308.00994, 2023. 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 681, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 681, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 681, + 287, + 713 + ], + "type": "text", + "content": "[66] Sangdoo Yun, Dongyoon Han, Seong Joon Oh, Sanghyuk Chun, Junsuk Choe, and Youngjoon Yoo. Cutmix: Regularization strategy to train strong classifiers with localizable" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 297 + ], + "type": "list", + "angle": 0, + "index": 21, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "text", + "content": "features. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6023-6032, 2019. 2, 6, 7" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 96, + 545, + 129 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 96, + 545, + 129 + ], + "spans": [ + { + "bbox": [ + 307, + 96, + 545, + 129 + ], + "type": "text", + "content": "[67] An Zhang, Wenchang Ma, Xiang Wang, and Tat seng Chua. Incorporating bias-aware margins into contrastive loss for collaborative filtering. In NeurlPS, 2022. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 130, + 545, + 163 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 130, + 545, + 163 + ], + "spans": [ + { + "bbox": [ + 307, + 130, + 545, + 163 + ], + "type": "text", + "content": "[68] Hongyi Zhang, Moustapha Cisse, Yann N Dauphin, and David Lopez-Paz. mixup: Beyond empirical risk minimization. arXiv preprint arXiv:1710.09412, 2017. 2, 6, 7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 164, + 545, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 164, + 545, + 196 + ], + "spans": [ + { + "bbox": [ + 307, + 164, + 545, + 196 + ], + "type": "text", + "content": "[69] Chenyu Zheng, Guoqiang Wu, and Chongxuan Li. Toward understanding generative data augmentation. Advances in Neural Information Processing Systems, 36, 2024. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 198, + 545, + 241 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 198, + 545, + 241 + ], + "spans": [ + { + "bbox": [ + 307, + 198, + 545, + 241 + ], + "type": "text", + "content": "[70] Bolei Zhou, Agata Lapedriza, Aditya Khosla, Aude Oliva, and Antonio Torralba. Places: A 10 million image database for scene recognition. IEEE transactions on pattern analysis and machine intelligence, 40(6):1452-1464, 2017. 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 243, + 545, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 243, + 545, + 297 + ], + "spans": [ + { + "bbox": [ + 307, + 243, + 545, + 297 + ], + "type": "text", + "content": "[71] Jun-Yan Zhu, Taesung Park, Phillip Isola, and Alexei A Efros. Unpaired image-to-image translation using cycle-consistent adversarial networks. In Proceedings of the IEEE international conference on computer vision, pages 2223-2232, 2017. 3" + } + ] + } + ], + "index": 20 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17233" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Enhanced Motion-Text Alignment for Image-to-Video Transfer Learning/1b431374-7f10-4d10-9e33-68850035e408_content_list.json b/2024/Enhanced Motion-Text Alignment for Image-to-Video Transfer Learning/1b431374-7f10-4d10-9e33-68850035e408_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..960b9645beb91bac67fe145d948325bfacdba713 --- /dev/null +++ b/2024/Enhanced Motion-Text Alignment for Image-to-Video Transfer Learning/1b431374-7f10-4d10-9e33-68850035e408_content_list.json @@ -0,0 +1,1525 @@ +[ + { + "type": "text", + "text": "Enhanced Motion-Text Alignment for Image-to-Video Transfer Learning", + "text_level": 1, + "bbox": [ + 117, + 130, + 852, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Wei Zhang $^{1,2*}$ , Chaoqun Wan $^{2}$ , Tongliang Liu $^{3}$ , Xinmei Tian $^{1,4\\dagger}$ , Xu Shen $^{2\\dagger}$ , Jieping Ye $^{2}$", + "bbox": [ + 143, + 179, + 864, + 198 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1University of Science and Technology of China, 2Alibaba Cloud, 3The University of Sydney Institute of Artificial Intelligence, Hefei Comprehensive National Science Center", + "bbox": [ + 138, + 199, + 872, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "zw0819@mail.ustc.edu.cn, tongliang.liu@sydney.edu.au, xinmei@ustc.edu.cn {qionlong.wcq, shenxu.sx, jieping.yjp}@alibaba-inc.com", + "bbox": [ + 89, + 234, + 923, + 271 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 304, + 313, + 320 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Extending large image-text pre-trained models (e.g., CLIP) for video understanding has made significant advancements. To enable the capability of CLIP to perceive dynamic information in videos, existing works are dedicated to equipping the visual encoder with various temporal modules. However, these methods exhibit \"asymmetry\" between the visual and textual sides, with neither temporal descriptions in input texts nor temporal modules in text encoder. This limitation hinders the potential of language supervision emphasized in CLIP, and restricts the learning of temporal features, as the text encoder has demonstrated limited proficiency in motion understanding. To address this issue, we propose leveraging \"MoTion-Enhanced Descriptions\" (MoTED) to facilitate the extraction of distinctive temporal features in videos. Specifically, we first generate discriminative motion-related descriptions via querying GPT-4 to compare easy-confusing action categories. Then, we incorporate both the visual and textual encoders with additional perception modules to process the video frames and generated descriptions, respectively. Finally, we adopt a contrastive loss to align the visual and textual motion features. Extensive experiments on five benchmarks show that MoTED surpasses state-of-the-art methods with convincing gaps, laying a solid foundation for empowering CLIP with strong temporal modeling.", + "bbox": [ + 76, + 335, + 473, + 715 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 739, + 209, + 753 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recent years have witnessed remarkable achievements in contrastive language-image pre-training models [9, 25, 36, 60, 87, 88], with CLIP [60] emerging as the front-runner. Through language supervision with a vast collection of 400 million image-text pairs, CLIP has achieved exceptional image comprehension and unprecedented zero-shot general-", + "bbox": [ + 75, + 765, + 468, + 856 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/fdb90797dde85bb2abcae4d1dd21b050d8b4d4fa3577438c3b73d612ad573c49.jpg", + "image_caption": [ + "(c) Motion-Enhanced Descriptions", + "Figure 1. (a) Previous methods [27, 49, 75, 78] align the visual embeddings with the textual embeddings derived from category names. (b) Our work symmetrically aligns the spatial/temporal visual embeddings with class descriptions and motion descriptions correspondingly. (c) The descriptions generated by large language modes (e.g., GPT-4 [50]) enhance the conceptual definitions and discriminative details of motions." + ], + "image_footnote": [], + "bbox": [ + 506, + 303, + 890, + 513 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ization. This breakthrough has opened up new possibilities for leveraging the power of large-scale pre-trained models to comprehend videos. It has also introduced a new paradigm [26, 40, 49, 52, 58, 75, 80] that endows image-based CLIP to effectively perceive dynamic information for video recognition.", + "bbox": [ + 496, + 671, + 890, + 762 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To equip CLIP with motion perception, existing works propose to incorporate various temporal modules into the visual encoder. These temporal modules are additional tunable parameters, including designed efficient units between pre-trained transformer blocks [49, 75], and parallel structures to disentangle the temporal modeling out of the spatial modeling [40, 58]. However, while considerable effort has been put into capturing temporal visual features from videos, these methods have paid little attention to the in", + "bbox": [ + 496, + 763, + 892, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*This work was done when the author was visiting Alibaba as a research intern.", + "bbox": [ + 76, + 862, + 468, + 886 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "† Corresponding author.", + "bbox": [ + 96, + 887, + 220, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "18504", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "put text and text encoder, leading to an imbalance between the visual and textual alignment. When transferring to the classification task of video recognition (Fig. 1(a)), the only available text is the \"category names\" of the actions, e.g., \"clapping\" and \"slapping\". These coarse-grained word-level descriptors lack clear descriptions and explanations, making it difficult to distinguish between them. For instance, \"slapping\" refers to a quick hand movement striking against something, while \"clapping\" refers to the act of striking both hands together rapidly. The scarcity of textual motion information contradicts the language supervision principle of CLIP. Furthermore, it's observed that the CLIP text encoder delivers a strong bias towards spatial concepts (e.g., nouns) [5, 30, 47, 63], with a weak understanding of temporal cues (e.g., verbs). These issues significantly limit the effectiveness of extending CLIP with motion modeling for video understanding.", + "bbox": [ + 75, + 90, + 472, + 349 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To overcome the limitations of language supervision on temporal modeling, we propose a novel approach that symmetrically aligns the spatial/temporal visual embeddings with class descriptions (\"slapping\") and motion descriptions correspondingly (\"a quick hand movement striking against something\"). We introduce the MoTion-Enhanced Descriptions (MoTED) as the language supervision for the visual temporal modules. In the implementation, we encounter two main challenges: i) how to generate motion-related descriptions? ii) how to effectively utilize descriptions as supervision? To tackle the first challenge, we aim for clear, detailed, and distinctive descriptions in the text. With the recent advancements in large language models (LLMs) [4, 11, 50], it has become possible to automatically generate conceptual descriptions for corresponding actions. As for the second challenge, we disentangle the spatial and temporal learning by constructing a temporal encoder, in parallel to the CLIP image encoder, to extract the temporal features aligned with motion-enhanced descriptions.", + "bbox": [ + 75, + 352, + 472, + 640 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To be specific, we employ GPT-4 [50] to generate descriptions by posing a query: “Q: what is the motion concept of ? A:” (as shown in Fig. 1(c)). However, our empirical findings indicate that only the concept descriptions offer marginal assistance in motion perception, particularly on datasets such as Something-Something-V2 (SSv2) [20]) that necessitate the differentiation of similar classes using intricate motion cues. To mitigate this issue, we propose to incorporate related top-k actions as context for the LLMs to generate more discriminative descriptions. These generated texts are then passed through the CLIP text encoder to extract textual motion representations, where an adapter is utilized to eliminate the biases in the original model. Correspondingly, the temporal module in the visual encoder processes the video to extract dynamic features and aligns them with the generated textual descriptions through contrastive learning. Finally, these two sets of", + "bbox": [ + 75, + 643, + 472, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "features are independently fused with the original CLIP's text and image features using cross-attention, and subsequently aligned. Evaluated on two supervised video recognition benchmarks, i.e., Kinetics-400 [28] and SSv2 [20], as well as three zero-shot benchmarks, i.e., Kinetics-600 [12], HMDB51 [24], UCF101 [65], the proposed MoTED surpasses state-of-the-art methods with convincing gaps, indicating the effectiveness of aligning enhanced motion descriptions with the temporal embedding of input videos.", + "bbox": [ + 496, + 90, + 890, + 226 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We summarize the contributions as follows:", + "bbox": [ + 517, + 227, + 808, + 241 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We present a new perspective that underscores the significance of textual side on par with visual side. By delving into in-depth distinguishing descriptions of actions, we make the first attempt to reveal the potential of language supervision as emphasized in CLIP.", + "- We propose MoTED that leverages LLMs to automatically generate action descriptions, and mining distinctive descriptions among similar actions. Then, we use a parallel path for both visual and textual motion modeling.", + "- We evaluate our approach on supervised as well as generalization tasks. Extensive experiments demonstrate the superiority and good generalization ability of the proposed method." + ], + "bbox": [ + 500, + 242, + 890, + 438 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 452, + 640, + 468 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Vision-Language Pre-training. In recent years, Vision-Language Pre-training (VLP) [9, 25, 36, 45, 46, 60, 76, 87, 88] has made remarkable progress. One of the most remarkable and influential works is CLIP [60], which adopts the contrastive language-image pretraining paradigm. Following that, this paradigm has shown impressive zero-shot generalization capabilities on various image-related tasks [31, 35, 37, 43, 82, 90]. However, pre-training a language-video model [34, 81] is prohibitively expensive, as it requires large-scale video-text data and extensive training resources (e.g., thousands of GPU days). Meanwhile, transferring language-image pre-trained models to the video domain [10, 26, 40, 49, 52, 58, 75, 78] has captured significant attention due to its striking performance and training efficiency. For instance, X-CLIP [49] integrates the CLIP image encoder with a cross-frame attention module for temporal modeling. Vita-CLIP [78] utilizes multi-modal prompting techniques to learn video and text-specific context vectors. DiST [58] is the most related work that also disentangles spatial and temporal learning in the visual side. In contrast, our MoTED highlights the effectiveness of temporal disentangling both in the visual and textual sides, supervised by the detailed descriptions of motions.", + "bbox": [ + 496, + 477, + 890, + 824 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Video Recognition. The conventional approaches in video recognition primarily focus on spatio-temporal learning under fully-supervised settings, where all categories are predefined. These approaches have achieved remarkable performance using various architectures, including convolution", + "bbox": [ + 496, + 825, + 893, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "18505", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/c777a177803b74fa4a8ecf60d1633d0ccd35e865c08bb75d94524f870af32ca1.jpg", + "image_caption": [ + "Figure 2. The overall framework of MoTED. Building upon the dual image and text encoders initialized by CLIP [60], we extend the capability to perceive motion information symmetrically in both sides. In the text side, the motion-related conceptual and discriminative descriptions are generated via querying LLMs (e.g., GPT-4) using category names as input. In the vision side, a motion encoder is built to extract temporal dynamics given densely-sampled frames and integrated with the middle-layer features from image encoder. To align the motion embeddings and video embeddings from both sides, it employs two contrastive losses, respectively." + ], + "image_footnote": [], + "bbox": [ + 86, + 88, + 893, + 306 + ], + "page_idx": 2 + }, + { + "type": "table", + "img_path": "images/2f381b127262b50e7eb47ac0abfb3ef16675d4a7770d19462f06583fe810e3a5.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Related workTaskAction annotationExample
Dist. Sup. [39]pretrainevent description‘insert window mounting bolts’
Weakly Sup. [14]pretrainevent description‘take up the iron clamp’
LAVILA [89]pretrainevent description‘a lady walks past a car’
VFC [47]VL transfer learningevent description‘two brown horses eating grass’
MoTED (Ours)VL transfer learningmotion concept description‘slapping: a quick hand movement striking against something’
", + "bbox": [ + 78, + 414, + 475, + 537 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "networks [7, 17, 21, 59, 64, 67–69, 71, 73], and vision transformers [1, 2, 6, 16, 32, 33, 38, 42, 48, 54, 55, 62]. In addition to the architecture design, self-supervised video representation learning [13, 18, 19, 23, 53, 57, 66, 74, 77, 79] has also gained popularity recently. However, these methods operate purely within the visual domain and meet bottlenecks of recognizing unseen or unfamiliar categories in real-world applications. Fortunately, the advance of large language models (LLMs) [4, 51] provides opportunities to mitigate this issue, due to their powerful capabilities of encoding world knowledge [29]. Recent studies [44, 56, 83, 85] have verified that factual sentences generated by LLMs can improve zero-shot image recognition accuracy. VFC [47] leverages PaLM [11] to create verb-focused hard negatives to enhance the understanding of verbs in video models. LSS [61] integrates language-based action concepts with self-supervised learning to adapt an image model to video domain. Our MoTED employs the descriptions that delineate the labeled classes and further accentuate the useful traits distinguishing similar classes.", + "bbox": [ + 75, + 598, + 472, + 902 + ], + "page_idx": 2 + }, + { + "type": "table", + "img_path": "images/99f02099f7781b12232bad5334b13f7743bfc3bcda7809c4c8383677e7b3a374.jpg", + "table_caption": [ + "Table 1. Comparison of our method with related works focused on action annotations. 'VL' is the abbreviation of 'vision-language'." + ], + "table_footnote": [], + "table_body": "
Related workLLMs roleExample
Chatvideo [72]manager& summarizeInput: 'Summarize the activity in video' \nOutput: 'A person is cooking in the kitchen'
MM-REACT [86]execution& summarizeInput: 'Please create a summary of the video' \nOutput: 'The speaker is making a BLT sandwich ...'
MiniGPT4 [91]/LLAVA [41]KB& data cleanInput: 'Describe this image in detail.' \nOutput: 'The image shows a group of musicians ...'
MoTED (Ours)generate descriptionsInput: 'What is the motion concept of slapping' \nOutput: 'A quick hand movement striking against ...'
", + "bbox": [ + 501, + 414, + 898, + 541 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Table 2. Comparison of our method with related works that take LLMs as a knowledge base (KB) and automatically annotated tool.", + "bbox": [ + 498, + 545, + 893, + 574 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 500, + 592, + 589, + 608 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The generic pipeline of MoTED is presented in Fig. 2, which consists of three steps: (1) Motion Description Generation to obtain conceptual descriptions $\\mathcal{D}_{con}$ and discriminative descriptions $\\mathcal{D}_{dis}$ , via querying language models given the category names of the target video dataset in Sec. 3.1; (2) Textual Motion Adaptation to obtain the text embeddings that extract the motion semantics given the generated descriptions $\\mathcal{D}_{con}$ and $\\mathcal{D}_{dis}$ in Sec. 3.2; (3) Visual Motion Extraction to obtain the visual embeddings that extract the motion semantics given input video frames in Sec. 3.3. Overall, the framework is trained in an end-to-end contrastive manner as illustrated in Sec. 3.4.", + "bbox": [ + 496, + 619, + 890, + 800 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Motion Description Generation", + "text_level": 1, + "bbox": [ + 500, + 815, + 777, + 830 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The first step is to obtain a set of appropriate descriptions for each category. Given a video dataset with multiple different categories (e.g., Kinetics400 [28] with 400 action classes), the descriptions are generated automati", + "bbox": [ + 496, + 839, + 893, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "18506", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "cally utilizing GPT-4 [50]. Note that, the generation process is agnostic to this choice and other LLMs can be used instead. For each category name, we query GPT-4 to provide the motion concepts using the following prompt: “Q: What is the motion concept in a video of ?” A: As shown in Fig. 1, the generated concept descriptions often cover moving objects, object interactions, moving directions and speeds, etc. But the output of LLMs can also be open-ended duplicates or anything in natural language (as illustrated in the Appendix). To control the generated descriptions to be concise and motion-related, we adopt a two-shot prompt in which we include two exemplars of question-answer for the same operation that is being queried.", + "bbox": [ + 75, + 90, + 472, + 303 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Besides the conceptual descriptions $\\mathcal{D}_{con}$ , we further investigate into the confusing actions that are characterized by their distinct differences from similar categories. To this end, we first compute the cosine distance of text embeddings for every two categories. Then we select top-k similar classes (e.g., $k = 5$ ) and query GPT-4 to generate motion characteristics for distinguishing the similar classes using the following prompt: \"Q: What are the useful features for distinguishing the original class from the similar class ", + "bbox": [ + 75, + 303, + 472, + 468 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "A:\" Different from the conceptual descriptions $\\mathcal{D}_{con}$ that are mainly determined by the general knowledge of LLMs, the descriptions $\\mathcal{D}_{dis}$ further contain task-specific information that the downstream task emphasizes.", + "bbox": [ + 76, + 469, + 468, + 530 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this way, given a dataset of $N$ categories, we can obtain the motion-enhanced $N * (k + 1)$ descriptions, consisting of 1 conceptual description and $k$ discriminative descriptions for each category. The above generation process is completed before the training of whole framework.", + "bbox": [ + 75, + 531, + 468, + 606 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The detailed comparisons of action annotations and LLM-based generations are presented in Tab. 1 and Tab. 2, respectively. Tab. 1 shows that existing works employ annotations that describe the action events in a subject-verb-object manner, which are object-centric. In contrast, our generated annotations are motion-related, object-agnostic and to supervise the learning of visual motion features with the world knowledge of motion concepts. In Tab. 2, LLMs are mainly applied in generative tasks, to summarize query results and generate general responses, or to obtain instruction data to align with human preference. In contrast, we apply LLMs to obtain descriptions of motion concept, aligned with the temporal visual features.", + "bbox": [ + 75, + 607, + 470, + 804 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Textual Motion Adaptation", + "text_level": 1, + "bbox": [ + 76, + 815, + 323, + 832 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "After generating the motion-enhanced descriptions, the second step is to perceive the motion cues within the descriptions and aggregate them into a compact motion-enhanced embedding. Given a set of motion-enhanced descriptions $\\mathcal{D}$", + "bbox": [ + 75, + 839, + 470, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "$= \\left\\{\\mathcal{D}_{con}^{i},\\mathcal{D}_{dis}^{i,j}\\right\\}$ where $i\\in [1,N]$ and $j\\in [1,k]$ , we extract normalized feature embedding $e_i$ by using the text encoder $f_{txt}$ : $e_i = f_{txt}(d_i)$ , where $d_{i}$ is a motion-enhanced description sampled from $\\mathcal{D}$ . In this way, we obtain the text embeddings $E_{txt}\\in \\mathbb{R}^{N\\times (k + 1)\\times C}$ , where $C$ denotes the channel number of each embedding. Noticeably, the parameters of text encoder is frozen and initialized by CLIP [60] to inherit its capacity of encoding visual-aligned semantics.", + "bbox": [ + 496, + 88, + 890, + 217 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "However, as CLIP is pre-trained on image-text paired corpus, the text encoder has a strong bias towards spatial appearance of objects and backgrounds, instead of temporal motions [5, 30, 47, 63]. To adapt the text encoder to understand the motion-enhanced descriptions, we introduce a Text Motion Adapter, which consists of a multi-head self-attention (MHSA) [70]. The adapter is parameter-tunable and takes $E_{txt}$ as input to learn the information dependencies between motion-enhanced descriptions. To aggregate the motion semantics for each category, we perform adaptive pooling to obtain the motion-enhanced embeddings $E_{txt}^{m} \\in \\mathbb{R}^{N \\times 1 \\times C}$ .", + "bbox": [ + 496, + 217, + 892, + 398 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Visual Motion Extraction", + "text_level": 1, + "bbox": [ + 498, + 405, + 732, + 420 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In the visual side, as shown in Fig. 3, it comprises of two components: (i) The image encoder adopts the CLIP pre-trained Vision Transformer (ViT), which extracts frozen features for sparse frames with powerful spatial semantics. (ii) The motion encoder takes dense frames as input to capture the local temporal cues, integrated with the global temporal dynamics from middle-layer image features.", + "bbox": [ + 496, + 429, + 890, + 535 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3.1 Image Encoder", + "text_level": 1, + "bbox": [ + 498, + 551, + 661, + 566 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Given a video clip $\\mathbf{X}_S\\in \\mathbb{R}^{T\\times H\\times W\\times 3}$ ( $T$ , $H$ , and $W$ represent the frame number, height, and width, respectively), the image features are extracted individually for several sparse frames. Following ViT [15], each frame is divided into $K = \\frac{H}{P}\\times \\frac{W}{P}$ patches, and the size of each patch is denoted as $P\\times P$ . These patches are then projected using a fully connected layer, referred to as the 2D stem in Fig. 3. This projection generates a sequence of patch embeddings $[\\mathbf{x}_{t,\\mathrm{cls}}^{(0)},\\mathbf{x}_{t,1}^{(0)},\\mathbf{x}_{t,2}^{(0)},\\dots ,\\mathbf{x}_{t,K}^{(0)}] + \\mathbf{e}^{\\mathrm{spatial}}$ , where $t = \\{1,\\dots ,T\\}$ , $\\mathbf{x}_{\\mathrm{cls}}$ is an additional learnable token (termed as \"class embedding\"), and $\\mathbf{e}^{\\mathrm{spatial}}$ denotes spatial position embedding. Assuming that the spatial encoder has $L$ Transformer blocks, the features of the $l_{\\mathrm{th}}$ layer for the $t_\\mathrm{th}$ frame can be extracted by:", + "bbox": [ + 496, + 574, + 890, + 789 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {X} _ {t} ^ {(l)} = \\operatorname {T r a n s f o r m e r} ^ {(l)} \\left(\\mathbf {X} _ {t} ^ {(l - 1)}\\right) \\in \\mathbb {R} ^ {(K + 1) \\times C}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 540, + 795, + 890, + 816 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $l = \\{1, \\dots, L\\}$ denotes the layer index. The class embeddings of $T$ frames in the $l_{\\mathrm{th}}$ layer is termed as $\\mathbf{X}^{(l)} = [\\mathbf{X}_1^{(l)}, \\dots, \\mathbf{X}_T^{(l)}] \\in \\mathbb{R}^{T \\times 1 \\times C}$ . Benefit from the CLIP pretrained parameters, the class embeddings $\\mathbf{X}^{(L)}$ aggregate powerful spatial semantics within each frame.", + "bbox": [ + 496, + 821, + 890, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "18507", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/e322011f8dfecfe46dbb4134236f27bde2cc8040bd79533ffd81132398abc644.jpg", + "image_caption": [ + "Figure 3. The structural details of MoTED in the vision side." + ], + "image_footnote": [], + "bbox": [ + 83, + 90, + 467, + 281 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3.2 Motion Encoder", + "text_level": 1, + "bbox": [ + 76, + 337, + 245, + 351 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To fully extract the motion information in videos, the temporal input $\\mathbf{X}_T\\in \\mathbb{R}^{\\alpha T\\times H\\times W\\times 3}$ for the motion encoder is sampled around the spatial input $\\mathbf{X}_S$ by $\\alpha$ times. In this study, we set $\\alpha = 2$ by default, as empirically studied in [17, 58]. Then, $\\mathbf{X}_T$ is projected by a 3D convolution, i.e., the 3D stem in Fig. 3, for patch embedding. The kernel size and stride of the 3D stem are both $P$ in spatial dimension and $\\alpha$ in temporal dimension. Thus the number of temporal patch tokens are the same with that in the image encoder, which makes it convenient to integrate the spatial and temporal features. Thus the projected temporal features can be formulated as: $\\mathbf{Z}^{(0)} = \\mathrm{Conv3d}(\\mathbf{X}_T)\\in \\mathbb{R}^{T\\times K\\times C}$ , where $K = \\frac{H}{P}\\times \\frac{W}{P}$ . Then, a series of Temporal Blocks are designed to extract motion patterns, which can be written as:", + "bbox": [ + 75, + 361, + 472, + 573 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {Z} ^ {(l)} = \\operatorname {T e m p - B l o c k} ^ {(l)} \\left(\\mathbf {Z} ^ {(l - 1)}, \\mathbf {X} ^ {(l - 1)}\\right) \\in \\mathbb {R} ^ {T \\times K \\times C}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 88, + 597, + 468, + 616 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where the function Temp-Block $(\\cdot)$ performs temporal modeling integrated with the spatial features $X^{(l - 1)}$", + "bbox": [ + 75, + 628, + 468, + 657 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Following the effective and efficient design philosophy in [32, 33], the temporal block consists of the following three modules: (i) Local Motion Extraction (LME). Since 3D convolution can capture detailed and local spatiotemporal features, by processing each pixel with context from a small 3D neighborhood (e.g., $2 \\times 16 \\times 16$ ), the temporal patch tokens $\\mathbf{Z}^{(0)}$ remain the nature of local motion extraction. Then these tokens are flattened to a sequence $\\mathbb{R}^{TK \\times C}$ and perform spatiotemporal learning via joint self-attention modules [66] and token merging for efficiency [3]. (ii) Global Motion Extraction (GME). Given the class embeddings $\\mathbf{X}^{(l)}$ that aggregate powerful spatial semantics for each individual frame, we apply a cross-frame self-attention module [49] to learn the global inter-frame interactions. (iii) Feature Integration. After obtaining the local and global motion features, we adopt a cross-attention mod", + "bbox": [ + 75, + 659, + 472, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "ule, which takes local features as query and global features as key/value, to complement the features for better temporal modeling. In this way, the output of the final temporal block $\\mathbf{Z}^{(L)}$ aggregates powerful temporal dynamics within the dense frames.", + "bbox": [ + 496, + 90, + 890, + 167 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Training Loss", + "text_level": 1, + "bbox": [ + 498, + 176, + 643, + 193 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Following CLIP [60], we first apply adaptive pooling to obtain a video-level motion embedding $\\mathbf{Z}_{avg}^{(L)} \\in \\mathbb{R}^{1 \\times C}$ . Then we introduce a contrastive loss to align the visual and textual motion embeddings:", + "bbox": [ + 496, + 200, + 890, + 263 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {m o t i o n}} = - \\log \\frac {\\exp \\left(\\sin \\left(\\mathbf {Z} _ {\\text {a v g}} ^ {(L)} , \\mathbf {e} _ {i}\\right) / \\tau\\right)}{\\sum_ {c = 1} ^ {N} \\exp \\left(\\sin \\left(\\mathbf {Z} _ {\\text {a v g}} ^ {(L)} , \\mathbf {e} _ {c}\\right) / \\tau\\right)}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 540, + 275, + 890, + 316 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathrm{sim}(\\cdot ,\\cdot)$ is the normalized cosine similarity, $\\tau$ refers to the temperature parameter. $\\mathbf{e}_i$ is a motion-enhanced text embedding for the $i_{\\mathrm{th}}$ category.", + "bbox": [ + 496, + 325, + 890, + 372 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Moreover, we adopt a cross-attention module to fuse the visual motion features (as key/value) and image features (as query) to obtain the compact video embedding $\\mathbf{v}_i$ . Similarly, the fused text embedding $\\mathbf{e}_i'$ is acquired via a cross-attention module. We introduce a contrastive loss to align the visual and textual video-level embeddings:", + "bbox": [ + 496, + 372, + 890, + 463 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {v i d e o}} = - \\log \\frac {\\exp \\left(\\operatorname {s i m} \\left(\\mathbf {v} _ {i} , \\mathbf {e} _ {i} ^ {\\prime}\\right) / \\tau\\right)}{\\sum_ {c = 1} ^ {N} \\exp \\left(\\operatorname {s i m} \\left(\\mathbf {v} _ {i} , \\mathbf {e} _ {c} ^ {\\prime}\\right) / \\tau\\right)}. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 553, + 474, + 890, + 511 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Compatible with CLIP training paradigm, the overall framework is trained based on these two contrastive losses:", + "bbox": [ + 498, + 521, + 890, + 551 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\mathcal {L} _ {\\text {m o t i o n}} + \\mathcal {L} _ {\\text {v i d e o}}. \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 614, + 569, + 890, + 584 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 500, + 598, + 633, + 616 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Dataset and Implementation.", + "text_level": 1, + "bbox": [ + 498, + 633, + 759, + 651 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Datasets. In the supervised setting, we train on the train set of Kinetics-400 (K400) [28] and Something-Something-V2 (SSv2) [20] and report supervised performance against existing methods on the validation sets of K400 and SSv2. In the zero-shot setting, we train on the Kinetics-400 training set and evaluate on three datasets: Kinetics-600 (K600) [12], HMDB51 [24] and UCF101 [65]. For zero-shot evaluation on K600, following [12], we use the 220 new categories outside of K400 for evaluation, and conduct evaluation three times, each time randomly sampling 160 categories for evaluation from the 220 categories. For zero-shot evaluation on HMDB51 and UCF101, we follow [49] and report average top-1 accuracy and standard deviation on three splits of the test set.", + "bbox": [ + 496, + 657, + 890, + 869 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Implementation Details. Following previous work [40, 58], we use the CLIP [60] pre-trained ViT-B/16, ViT-L/14", + "bbox": [ + 498, + 869, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "18508", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/2572135649c923680bf865fff5a1719aa4d2bcbb4b89b24b70a7daa5b2c3f68d.jpg", + "table_caption": [], + "table_footnote": [ + "(a) \"Concept.\" is the abbreviation of \"conceptual descriptions\". \"Discr.\" is the abbreviation of \"discriminative descriptions\"." + ], + "table_body": "
Class NamesConcept.Discr.SSv2K400
✓XX65.782.9
✓✓X66.084.5
✓X✓69.984.0
✓✓✓70.185.1
", + "bbox": [ + 84, + 102, + 349, + 176 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/ae714a9a820bffee25b47662d333fc16c93f09a192b3cc5d4da707bac866c676.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
M.Enc.Adapter.SSv2K400
XX64.381.5
✓X68.683.4
X✓66.283.8
✓✓70.185.1
", + "bbox": [ + 393, + 88, + 581, + 165 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "(b) \"M.Enc.\" is the separated motion encoder where motion modules are parallel with CLIP visual encoder. \"Adapter\" is the text motion adapter.", + "bbox": [ + 351, + 180, + 617, + 218 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/a6cc3563377997bd74c48c520c757f80c31036009299dc37a544ef7a7c5411fe.jpg", + "table_caption": [], + "table_footnote": [ + "(c) \"Text.Fuse.\" indicates feature fusion in the textual side. \"Vision.Fuse.\" indicates feature fusion in the visual side." + ], + "table_body": "
Text.Fuse.Vision.Fuse.SSv2K400
XX68.383.6
✓X68.984.2
X✓69.584.5
✓✓70.185.1
", + "bbox": [ + 645, + 89, + 874, + 165 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 3. Ablations on Something-Something-V2 and Kinetics-400. The spatial encoder is a 8-frame vanilla ViT-B/16 pre-trained by CLIP [60]. The inference protocol of all models and datasets are 3 clips × 1 center crop.", + "bbox": [ + 75, + 223, + 892, + 252 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/b848dc7327e56cbfbc7307b0faf989346b6773822fa7005d57c405e14d0bbb5d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 78, + 266, + 478, + 327 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/661bdcac5cfcf91f68631a263b297c87e0fb3d9b585739073bc9f1af0cc3a846.jpg", + "image_caption": [ + "Relevance between text and motion features", + "(a) case 1" + ], + "image_footnote": [], + "bbox": [ + 78, + 329, + 478, + 382 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/4aef931383e4376f9db7a3352bfcfbf88e3155f28f48a8ad5b5c84f4a2ce5541.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 491, + 265, + 890, + 327 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/1812d5779aff347aaf381d9b414dd1265f58885d7ae41061d6641f136837685f.jpg", + "image_caption": [ + "Relevance between text and motion features", + "(b) case2", + "Figure 4. Two cases to visualize the relevance [8] between text and image/motion features to highlight the information relevant to the prediction. The different \"regions of interest\" and \"words of importance\" indicate that the motion and image features could be disentangled." + ], + "image_footnote": [], + "bbox": [ + 491, + 329, + 890, + 382 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "and ViT-L/14-336p as our image encoder. Unless otherwise specified, we mark the default settings in the temporal encoder in gray in Sec. 4.2. For simplicity, the ablation studies are conducted based on ViT-B/16 with a low-rate sampled 8 frames. We conduct the experiments with the NVIDIA 32G V100 GPUs. More implementation details (e.g., training and testing hyper-parameters) are described in the Appendix.", + "bbox": [ + 75, + 462, + 468, + 583 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Ablation and Analysis", + "text_level": 1, + "bbox": [ + 76, + 590, + 284, + 606 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Language Supervision. In this section, we aim to demonstrate the significance of language supervision by comparing different types of text descriptions. Tab. 3a presents a comparison of three types of texts: category names, motion concept descriptions, motion discriminative descriptions. It reveals that the performance of action category names, which contain the least information, is notably lower than other results, particularly on the SSv2 dataset with abundant temporal information. Although descriptions with only action concepts show a slight improvement of $0.3\\% /1.6\\%$ , they lack the vital information regarding the distinguishing characteristics of the actions. As a result, the model exhibits poor classification accuracy for similar categories. On the other hand, when the discriminative descriptions are adopted, the accuracy on SSv2 is improved noticeably $(+4.2\\%)$ , revealing the significance of discriminative characteristics for fine-grained datasets. Additionally, when the full descriptions are utilized, the model's performance obtain the further gains of $4.4\\% /2.2\\%$ .", + "bbox": [ + 75, + 613, + 472, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Disentangled Motion Modeling. Our method has an advantage in allowing the learning of dynamic information without interference, while preserving CLIP's original spatially transferable representation capabilities. To validate this, we compared two different structures in Tab. 3b. For the serial structure, additional dynamic modules are inserted between the Transformer Blocks of each layer of the visual encoder, resulting in a unified video representation. All the texts are combined through the text encoder to obtain text representations for alignment. It can be observed that the parallel structure outperforms the serial structure, with improvements of $4.3\\%$ and $1.9\\%$ respectively. This demonstrates the rationality of the parallel approach. Tab. 3b also presents a comparison of introducing an adapter in the text encoder. The model's classification performance improved by $1.9\\%$ and $2.3\\%$ after adding the adapter, indicating its effectiveness in reducing bias in the text encoder.", + "bbox": [ + 496, + 462, + 892, + 718 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Feature Fusion Direction. As shown in Tab. 3c, both directions of information integration can improve performances. The combination of the two can boost accuracy more significantly $+1.8\\% / + 1.5\\%$ on SSv2/K400, respectively. This verifies the importance of spatio-temporal blending for the parallel architectures. In our opinion, independent learning of motion can effectively avoid excessive reliance of the model on the previously learned spatial information. By using cross attention for fusion, it is possible to effectively integrate features from two different dimensions, ultimately forming the features of the target video and achieving the best results.", + "bbox": [ + 496, + 719, + 893, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "18509", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/5a47145889b60ab18a3b0c5a8a3ee11b1d046b4d3fdaa2a509dfadfa3b270383.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodPre-trainArchitectureInput SizeFLOPs×Cr.×Cl. (T)Param (M)FrozenTop-1Top-5
SlowFast [17]ImageNet-21KR101+NL16 × 22420.1 × 3 × 160X63.187.6
ViViT FE [1]IN21K+K400ViT-L16 × 22421.0 × 3 × 4612X65.489.8
MTV-B(320p) [84]IN21K+K400-32 × 22420.9 × 3 × 4310X68.590.4
MViT [16]Kinetics-600MViT-B-2432 × 22420.2 × 3 × 153X68.791.5
Video Swin [42]IN21K+K400Swin-B32 × 22420.3 × 3 × 160X69.692.7
TAdaConvNeXtV2 [22]IN1K+K400ConvNeXt-S32 × 22420.2 × 3 × 282X70.092.0
EVL* [40]CLIP-400MViT-B32 × 22420.68 × 1 × 3175✓62.4-
ST-Adapter* [52]CLIP-400MViT-B32 × 22420.61 × 1 × 393✓69.592.6
DiST*CLIP-400MViT-B32 × 22420.65 × 1 × 3105✓70.992.1
MoTED*CLIP-400MViT-B8 × 22420.18 × 1 × 3112✓70.191.8
MoTED*CLIP-400MViT-B16 × 22420.34 × 1 × 3112✓71.292.4
MoTED*CLIP-400MViT-B32 × 22420.68 × 1 × 3112✓71.992.7
UnifromerV2 [32]CLIP-400MViT-L32 × 22421.73 × 1 × 3574X73.094.5
TAdaFormer [22]CLIP-400MViT-L32 × 22421.70 × 2 × 3364X73.6-
EVL* [40]CLIP-400MViT-L32 × 22423.21 × 1 × 3654✓66.7-
EVL* [40]CLIP-400MViT-L32 × 33628.08 × 1 × 3654✓68.0-
ST-Adapter* [52]CLIP-400MViT-L32 × 22422.75 × 1 × 3347✓72.393.9
DiST*CLIP-400MViT-L32 × 22422.83 × 1 × 3336✓73.193.2
MoTED*CLIP-400MViT-L8 × 22420.78 × 1 × 3346✓71.592.6
MoTED*CLIP-400MViT-L16 × 22421.49 × 1 × 3346✓73.093.4
MoTED*CLIP-400MViT-L32 × 22422.89 × 1 × 3346✓73.893.8
", + "bbox": [ + 91, + 88, + 879, + 429 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Motion Modeling Visualization. In Fig. 4, we also perform the analysis of vision features generated by image and motion encoder to investigate the learned patterns from language supervision. Based on the reasoning tool [8], we depict the attention maps of class token from the final transformer block of the image/motion stream encoder w.r.t. the text encoder. It is observed that, for the case1 in Fig. 4, the CLIP image encoder attends to both motion-relevant foreground and motion-irrelevant background with a major focus on \"surface\". In contrast, features extracted from the motion modules concentrate on the motion-relevant regions of the moving object, and emphasize the motion-related words \"sliding\". This phenomenon reveals that the motion features can complement the static spatial semantics of objects in the image features via modeling object motions. More cases can be accessed in the Appendix.", + "bbox": [ + 75, + 494, + 472, + 738 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Fully-supervised Experiments", + "text_level": 1, + "bbox": [ + 76, + 753, + 344, + 771 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In the supervised setting, the results on SSv2 and K400 are presented in Tab. 4 and Tab. 6, respectively. Compared with AVL, our proposed MoTED introduces a similar temporal module for CLIP visual encoder, but our method has a significant improvement compared to AVL, with an improvement of $9.5\\% / 2.0\\%$ and $5.8\\% / 1.5\\%$ on ViT-B and ViT-L respectively on SSv2/K400. The performance gains greatly demonstrate the rationality and effectiveness of using lan", + "bbox": [ + 75, + 779, + 470, + 902 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "guage supervision. Interestingly, the performance gains are particularly pronounced on SSv2. This is because SSv2 is a fine-grained action classification task requiring stronger action discrimination and perception of dynamic information.", + "bbox": [ + 496, + 494, + 892, + 556 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/4b9db167afe472ddeb2c93536b105974dd8fa356d8b2b50e7459f56759f806bc.jpg", + "table_caption": [ + "Table 4. Comparison with the state-of-the-art methods on Something-Something-V2. \"Cr.\" and \"Cl.\" are the abbreviation for \"spatial crops\" and \"temporal clips\". \"Frozen\" indicates freezing the CLIP pre-trained parameters." + ], + "table_footnote": [], + "table_body": "
Related workVisual disentangleTextual disentangleK400 (Acc / Δ)SSv2 (Acc / Δ)
Previous SOTAXX84.2 / -69.5 / -
DiST [58]✓X85.0 / +0.870.9 / +1.4
MoTED (Ours)✓✓86.2 / +1.271.9 / +1.0
", + "bbox": [ + 501, + 570, + 888, + 637 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 5. Comparison of our method with previous state-of-the-art (SOTA) and the latest related work DiST.", + "bbox": [ + 500, + 642, + 890, + 670 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Disentangling spatio-temporal learning is an effective approach to endow the model with temporal capability. As shown in Tab. 5, DiST focuses on temporal disentangling in the visual side with gains of $+0.8\\% / +1.4\\%$ on K400/SSv2. Our study highlights the effectiveness of temporal disentangling in the textual side with detailed descriptions of motions, with further gains of $+1.2\\% / +1.0\\%$ . This result reveals that disentangling textual encoder is equally effective w.r.t. disentangling visual encoder for vision-language transfer learning.", + "bbox": [ + 496, + 683, + 893, + 835 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Zero-shot Experiments", + "text_level": 1, + "bbox": [ + 500, + 845, + 714, + 863 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Zero-shot generalization is an attractive characteristic of CLIP-extended models, making them more practical in real", + "bbox": [ + 498, + 869, + 893, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "18510", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/db809d50ed249d841c3bea2cccfac414162b015dcdb087160197aa5123ec8e7d.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodPre-trainArchitectureInput SizeTFLOPs×Cr.×Cl.Param (M)FrozenTop-1Top-5
SlowFast [17]-R101+NL16 × 22420.4 × 3 × 1060X79.893.9
TimeSformer [2]ImageNet-21KViT-L96 × 22428.4 × 3 × 1430X80.794.7
MViT [16]-MViT-B64 × 22420.5 × 1 × 537X81.295.1
ViViT FE [1]ImageNet-21KViT-L128 × 22424.0 × 3 × 1N/AX81.793.8
Video Swin [42]ImageNet-21KSwin-L32 × 22420.6 × 3 × 4197X83.195.9
TAdaConvNeXtV2 [22]ImageNet-21KConvNeXt-B32 × 22420.3 × 3 × 4146X83.7-
X-CLIP [49]CLIP-400MViT-B16 × 22420.28 × 3 × 4128X84.796.8
ST-Adapter* [52]CLIP-400MViT-B32 × 22420.61 × 1 × 393✓82.796.2
EVL* [40]CLIP-400MViT-B32 × 22420.59 × 1 × 3115✓84.2-
DiST*CLIP-400MViT-B32 × 22420.65 × 1 × 3112✓85.097.0
MoTED*CLIP-400MViT-B8 × 22420.18 × 1 × 3116✓85.197.0
MoTED*CLIP-400MViT-B16 × 22420.34 × 1 × 3116✓85.497.2
MoTED*CLIP-400MViT-B32 × 22420.68 × 1 × 3116✓86.297.5
UnifromerV2 [32]CLIP-400M+K710ViT-L32 × 22422.66 × 2 × 3354X89.398.2
TAdaFormer [22]CLIP-400M+K710ViT-L32 × 22421.41 × 4 × 3364X89.5-
ST-Adapter* [52]CLIP-400MViT-L32 × 22422.75 × 1 × 3347✓87.297.6
EVL* [40]CLIP-400MViT-L32 × 22422.70 × 1 × 3363✓87.3-
DiST*CLIP-400MViT-L32 × 22422.83 × 1 × 3343✓88.097.9
MoTED*CLIP-400MViT-L8 × 22420.78 × 1 × 3349✓87.497.8
MoTED*CLIP-400MViT-L16 × 22421.49 × 1 × 3349✓88.098.0
MoTED*CLIP-400MViT-L32 × 22422.89 × 1 × 3349✓88.898.2
", + "bbox": [ + 94, + 87, + 877, + 430 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/8ee63430ed2f55d74a3a4f0f4d58974805df2f3860cf317afb32f184c179b72b.jpg", + "table_caption": [ + "Table 6. Comparison with state-of-the-arts on Kinetics-400. \"Cr.\" and \"Cl.\" are the abbreviation for \"spatial crops\" and \"temporal clips\". \"Frozen\" indicates freezing the CLIP pre-trained parameters." + ], + "table_footnote": [], + "table_body": "
MethodModelHMDB51UCF101K600
ActionCLIP [75]B/1640.8±5.458.3±3.4-
X-CLIP [49]B/1644.6±5.272.0±2.365.2±0.4
Vita-CLIP [78]B/1648.6±0.675.0±0.667.4±0.5
DiST*B/1655.4±1.272.3±0.6-
MoTED*B/1658.2±1.178.3±0.669.9 ± 0.5
", + "bbox": [ + 78, + 493, + 468, + 584 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 7. Comparison of zero-shot accuracy with the state-of-the-art CLIP-based methods on three datasets (e.g., HMDB51 [24], UCF101 [65], and K600 [12]). \"\": frozen backbone.", + "bbox": [ + 76, + 594, + 468, + 638 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "world applications. For zero-shot settings, we evaluate MoTED on three widely used benchmarks. Following prior work [49], we train the networks on the K400 training set, then conduct the zero-shot evaluation on three unseen datasets (i.e., UCF101, HMDB51 and K600), as shown in Tab. 7. All models share the same architecture of \"ViT-B\", with 32 frames during inference. Compared with other vison-language methods, MoTED achieves the better zero-shot performances with a significant margin on HMDB51 $(+3.8\\%)$ , UCF101 $(+6.0\\%)$ and K600 $(+2.5\\%)$ . Different from X-CLIP [49], DiST [58], Vita-CLIP [78] that learns motion representation with the supervision of action category names solely, the proposed MoTED makes full use of conceptual and discriminative descriptions and learn general motion representations with the aid of language supervision. In addition, our method also has a relatively small", + "bbox": [ + 75, + 657, + 472, + 902 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "variance, only about $1\\%$ . We assume this is due to the benefits brought by the rich content of the text, as detailed descriptions can promote stable dynamic feature learning.", + "bbox": [ + 498, + 494, + 890, + 542 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 558, + 619, + 574 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this study, we aim to overcome the limitations of existing methods in extending large image-text pre-trained models for video understanding. The proposed MoTED framework introduces Motion-Enhanced Descriptions, which are applied to facilitate the extraction of unique temporal features in videos. By generating motion-related descriptions and incorporating perception modules, MoTED aligns visual and textual motion features using a contrastive loss. Experimental results on five benchmarks demonstrate that MoTED provides a strong basis for enhancing CLIP with robust temporal modeling. In future works, we hope to further dedicate to exploring the potential of language supervision and combining it with more powerful dynamic information perception modules to achieve higher performance in video recognition and make it truly practical.", + "bbox": [ + 496, + 584, + 890, + 811 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Acknowledgement", + "text_level": 1, + "bbox": [ + 500, + 828, + 679, + 845 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This work was supported in part by NSFC No. 62222117, the Fundamental Research Funds for the Central Universities under contract WK3490000005, and KY2100000117.", + "bbox": [ + 498, + 854, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "18511", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Anurag Arnab, Mostafa Dehghani, Georg Heigold, Chen Sun, Mario Lučić, and Cordelia Schmid. Vivit: A video vision transformer. In ICCV, pages 6836-6846, 2021. 3, 7, 8", + "[2] Gedas Bertasius, Heng Wang, and Lorenzo Torresani. Is space-time attention all you need for video understanding? In ICML, page 4, 2021. 3, 8", + "[3] Daniel Bolya, Cheng-Yang Fu, Xiaoliang Dai, Peizhao Zhang, Christoph Feichtenhofer, and Judy Hoffman. Token merging: Your vit but faster. In ICLR, 2023. 5", + "[4] Tom B. Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-Voss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel M. Ziegler, Jeffrey Wu, Clemens Winter, Christopher Hesse, Mark Chen, Eric Sigler, Mateusz Litwin, Scott Gray, Benjamin Chess, Jack Clark, Christopher Berner, Sam McCandlish, Alec Radford, Ilya Sutskever, and Dario Amodei. Language models are few-shot learners. In NeurIPS, 2020. 2, 3", + "[5] Shyamal Buch, Cristóbal Eyzaguirre, Adrien Gaidon, Jiajun Wu, Li Fei-Fei, and Juan Carlos Niebles. Revisiting the \"video\" in video-language understanding. In CVPR, pages 2907–2917, 2022. 2, 4", + "[6] Adrian Bulat, Juan Manuel Perez Rua, Swathikiran Sudhakaran, Brais Martinez, and Georgios Tzimiropoulos. Space-time mixing attention for video transformer. NeurIPS, 34:19594-19607, 2021. 3", + "[7] Joao Carreira and Andrew Zisserman. Quo vadis, action recognition? a new model and the kinetics dataset. In CVPR, pages 6299-6308, 2017. 3", + "[8] Hila Chefer, Shir Gur, and Lior Wolf. Generic attention-model explainability for interpreting bi-modal and encoder-decoder transformers. In ICCV, pages 387-396, 2021. 6, 7", + "[9] Xi Chen, Xiao Wang, Soravit Changpinyo, A. J. Piergiovanni, Piotr Padlewski, Daniel Salz, Sebastian Goodman, Adam Grycner, Basil Mustafa, Lucas Beyer, Alexander Kolesnikov, Joan Puigcerver, Nan Ding, Keran Rong, Hassan Akbari, Gaurav Mishra, Linting Xue, Ashish V. Thapliyal, James Bradbury, and Weicheng Kuo. Pali: A jointly scaled multilingual language-image model. In ICLR, 2023. 1, 2", + "[10] Xing Cheng, Hezheng Lin, Xiangyu Wu, Fan Yang, and Dong Shen. Improving video-text retrieval by multi-stream corpus alignment and dual softmax loss. arXiv preprint arXiv:2109.04290, 2021. 2", + "[11] Aakanksha Chowdhery, Sharan Narang, Jacob Devlin, Maarten Bosma, Gaurav Mishra, Adam Roberts, Paul Barham, Hyung Won Chung, Charles Sutton, Sebastian Gehrmann, Parker Schuh, Kensen Shi, Sasha Tsvyashchenko, Joshua Maynez, Abhishek Rao, Parker Barnes, Yi Tay, Noam Shazeer, Vinodkumar Prabhakaran, Emily Reif, Nan Du, Ben Hutchinson, Reiner Pope, James Bradbury, Jacob Austin, Michael Isard, Guy Gur-Ari, Pengcheng Yin, Toju Duke, Anselm Levskaya, Sanjay" + ], + "bbox": [ + 78, + 114, + 468, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Ghemawat, Sunipa Dev, Henryk Michalewski, Xavier Garcia, Vedant Misra, Kevin Robinson, Liam Fedus, Denny Zhou, Daphne Ippolito, David Luan, Hyeontaek Lim, Barret Zoph, Alexander Spiridonov, Ryan Sepassi, David Dohan, Shivani Agrawal, Mark Omernick, Andrew M. Dai, Thanumalayan Sankaranarayana Pillai, Marie Pellat, Aitor Lewkowycz, Erica Moreira, Rewon Child, Oleksandr Polozov, Katherine Lee, Zongwei Zhou, Xuezhi Wang, Brennan Saeta, Mark Diaz, Orhan Firat, Michele Catasta, Jason Wei, Kathy Meier-Hellstern, Douglas Eck, Jeff Dean, Slav Petrov, and Noah Fiedel. Palm: Scaling language modeling with pathways. J. Mach. Learn. Res., 24:240:1-240:113, 2023. 2, 3", + "[12] Dima Damen, Hazel Doughty, Giovanni Maria Farinella, Antonino Furnari, Evangelos Kazakos, Jian Ma, Davide Moltisanti, Jonathan Munro, Toby Perrett, Will Price, et al. Rescaling egocentric vision. arXiv preprint arXiv:2006.13256, 2020. 2, 5, 8", + "[13] Ali Diba, Vivek Sharma, Reza Safdari, Dariush Lotfi, Saquib Sarfraz, Rainer Stiefelhagen, and Luc Van Gool. Vi2cIr: Video and image for visual contrastive learning of representation. In ICCV, pages 1502-1512, 2021. 3", + "[14] Sixun Dong, Huazhang Hu, Dongze Lian, Weixin Luo, Yicheng Qian, and Shenghua Gao. Weakly supervised video representation learning with unaligned text for sequential videos. In CVPR, pages 2437-2447. IEEE, 2023. 3", + "[15] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2020. 4", + "[16] Haoqi Fan, Bo Xiong, Karttikeya Mangalam, Yanghao Li, Zhicheng Yan, Jitendra Malik, and Christoph Feichtenhofer. Multiscale vision transformers. In ICCV, pages 6824-6835, 2021. 3, 7, 8", + "[17] Christoph Feichtenhofer, Haoqi Fan, Jitendra Malik, and Kaiming He. Slowfast networks for video recognition. In ICCV, pages 6202-6211, 2019. 3, 5, 7, 8", + "[18] Christoph Feichtenhofer, Haoqi Fan, Bo Xiong, Ross Girshick, and Kaiming He. A large-scale study on unsupervised spatiotemporal representation learning. In CVPR, pages 3299-3309, 2021. 3", + "[19] Christoph Feichtenhofer, Haoqi Fan, Yanghao Li, and Kaiming He. Masked autoencoders as spatiotemporal learners. NeurIPS, 2022. 3", + "[20] Raghav Goyal, Samira Ebrahimi Kahou, Vincent Michalski, Joanna Materzynska, Susanne Westphal, Heuna Kim, Valentin Haenel, Ingo Fruend, Peter Yianilos, Moritz Mueller-Freitag, et al. The\" something something\" video database for learning and evaluating visual common sense. In ICCV, pages 5842-5850, 2017. 2, 5", + "[21] Ziyuan Huang, Shiwei Zhang, Liang Pan, Zhiwu Qing, Mingqian Tang, Ziwei Liu, and Marcelo H Ang Jr. Tada! temporally-adaptive convolutions for video understanding. In ICLR, 2022. 3", + "[22] Ziyuan Huang, Shiwei Zhang, Liang Pan, Zhiwu Qing, Yingya Zhang, Ziwei Liu, and Marcelo H Ang Jr." + ], + "bbox": [ + 503, + 92, + 890, + 901 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "18512", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Temporally-adaptive models for efficient video understanding. arXiv preprint arXiv:2308.05787, 2023. 7, 8", + "[23] Simon Jenni and Hailin Jin. Time-equivariant contrastive video representation learning. In ICCV, pages 9970-9980, 2021. 3", + "[24] H Jhuang, H Garrote, E Poggio, T Serre, and T Hmdb. A large video database for human motion recognition. In ICCV, page 6, 2011. 2, 5, 8", + "[25] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In ICML, pages 4904-4916. PMLR, 2021. 1, 2", + "[26] Chen Ju, Tengda Han, Kunhao Zheng, Ya Zhang, and Weidi Xie. Prompting visual-language models for efficient video understanding. In ECCV, pages 105-124. Springer, 2022. 1, 2", + "[27] Chen Ju, Tengda Han, Kunhao Zheng, Ya Zhang, and Weidi Xie. Prompting visual-language models for efficient video understanding. In ECCV, pages 105-124. Springer, 2022. 1", + "[28] Will Kay, Joao Carreira, Karen Simonyan, Brian Zhang, Chloe Hillier, Sudheendra Vijayanarasimhan, Fabio Viola, Tim Green, Trevor Back, Paul Natsev, et al. The kinetics human action video dataset. arXiv preprint arXiv:1705.06950, 2017. 2, 3, 5", + "[29] Jan Kocon, Igor Cichecki, Oliwier Kaszyca, Mateusz Kochanek, Dominika Szydlo, Joanna Baran, Julita Bielaniewicz, Marcin Gruza, Arkadiusz Janz, Kamil Kanclerz, Anna Kocon, Bartlomiej Koptyra, Wiktoria Mieleszczenko-Kowszewicz, Piotr Milkowski, Marcin Oleksy, Maciej Piasecki, Lukasz Radlinski, Konrad Wojtasik, Stanislaw Wozniak, and Przemyslaw Kazienko. Chatgpt: Jack of all trades, master of none. Inf. Fusion, 99: 101861, 2023. 3", + "[30] Jie Lei, Tamara L. Berg, and Mohit Bansal. Revealing single frame bias for video-and-language learning. In ACL, pages 487-507, 2023. 2, 4", + "[31] Chunyuan Li, Haotian Liu, Liunian Harold Li, Pengchuan Zhang, Jyoti Aneja, Jianwei Yang, Ping Jin, Houdong Hu, Zicheng Liu, Yong Jae Lee, and Jianfeng Gao. ELEVATER: A benchmark and toolkit for evaluating language-augmented visual models. In NeurIPS, 2022. 2", + "[32] Kunchang Li, Yali Wang, Yinan He, Yizhuo Li, Yi Wang, Limin Wang, and Yu Qiao. Uniformerv2: Spatiotemporal learning by arming image vits with video uniformer. arXiv preprint arXiv:2211.09552, 2022. 3, 5, 7, 8", + "[33] Kunchang Li, Yali Wang, Junhao Zhang, Peng Gao, Guanglu Song, Yu Liu, Hongsheng Li, and Yu Qiao. Uniformer: Unifying convolution and self-attention for visual recognition. TPAMI, 2023. 3, 5", + "[34] Linjie Li, Yen-Chun Chen, Yu Cheng, Zhe Gan, Licheng Yu, and Jingjing Liu. HERO: hierarchical encoder for video+language omni-representation pre-training. In EMNLP, pages 2046-2065, 2020. 2", + "[35] Liunian Harold Li, Pengchuan Zhang, Haotian Zhang, Jianwei Yang, Chunyuan Li, Yiwu Zhong, Lijuan Wang, Lu Yuan, Lei Zhang, Jenq-Neng Hwang, Kai-Wei Chang, and" + ], + "bbox": [ + 78, + 90, + 468, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Jianfeng Gao. Grounded language-image pre-training. In CVPR, pages 10955-10965, 2022. 2", + "[36] Yanghao Li, Haoqi Fan, Ronghang Hu, Christoph Feichtenhofer, and Kaiming He. Scaling language-image pre-training via masking. arXiv preprint arXiv:2212.00794, 2022. 1, 2", + "[37] Yangguang Li, Feng Liang, Lichen Zhao, Yufeng Cui, Wanli Ouyang, Jing Shao, Fengwei Yu, and Junjie Yan. Supervision exists everywhere: A data efficient contrastive language-image pre-training paradigm. In ICLR, 2022. 2", + "[38] Yanghao Li, Chao-Yuan Wu, Haoqi Fan, Karttikeya Mangalam, Bo Xiong, Jitendra Malik, and Christoph Feichtenhofer. Mvitv2: Improved multiscale vision transformers for classification and detection. In CVPR, pages 4804-4814, 2022. 3", + "[39] Xudong Lin, Fabio Petroni, Gedas Bertasius, Marcus Rohrbach, Shih-Fu Chang, and Lorenzo Torresani. Learning to recognize procedural activities with distant supervision. In CVPR, pages 13843-13853. IEEE, 2022. 3", + "[40] Ziyi Lin, Shijie Geng, Renrui Zhang, Peng Gao, Gerard de Melo, Xiaogang Wang, Jifeng Dai, Yu Qiao, and Hongsheng Li. Frozen clip models are efficient video learners. In ECCV, pages 388-404. Springer, 2022. 1, 2, 5, 7, 8", + "[41] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. In NeurIPS, 2023. 3", + "[42] Ze Liu, Jia Ning, Yue Cao, Yixuan Wei, Zheng Zhang, Stephen Lin, and Han Hu. Video swim transformer. In CVPR, pages 3202-3211, 2022. 3, 7, 8", + "[43] Timo Lüddecke and Alexander S. Ecker. Image segmentation using text and image prompts. In CVPR, pages 7076-7086, 2022. 2", + "[44] Sachit Menon and Carl Vondrick. Visual classification via description from large language models. In ICLR, 2023. 3", + "[45] Antoine Miech, Dimitri Zhukov, Jean-Baptiste Alayrac, Makarand Tapaswi, Ivan Laptev, and Josef Sivic. Howto100m: Learning a text-video embedding by watching hundred million narrated video clips. In ICCV, pages 2630-2640, 2019. 2", + "[46] Antoine Miech, Jean-Baptiste Alayrac, Lucas Smaira, Ivan Laptev, Josef Sivic, and Andrew Zisserman. End-to-end learning of visual representations from uncurated instructional videos. In CVPR, pages 9879-9889, 2020. 2", + "[47] Liliane Momeni, Mathilde Caron, Arsha Nagrani, Andrew Zisserman, and Cordelia Schmid. Verbs in action: Improving verb understanding in video-language models. ICCV, 2023. 2, 3, 4", + "[48] Daniel Neimark, Omri Bar, Maya Zohar, and Dotan Asselmann. Video transformer network. In ICCV, pages 3163-3172, 2021. 3", + "[49] Bolin Ni, Houwen Peng, Minghao Chen, Songyang Zhang, Gaofeng Meng, Jianlong Fu, Shiming Xiang, and Haibin Ling. Expanding language-image pretrained models for general video recognition. In ECCV, pages 1-18. Springer, 2022. 1, 2, 5, 8", + "[50] OpenAI. Gpt-4 technical report, 2023. 1, 2, 4", + "[51] TB OpenAI. Chatgpt: Optimizing language models for dialogue. OpenAI, 2022. 3" + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "18513", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[52] Junting Pan, Ziyi Lin, Xiatian Zhu, Jing Shao, and Hongsheng Li. Parameter-efficient image-to-video transfer learning. arXiv e-prints, pages arXiv-2206, 2022. 1, 2, 7, 8", + "[53] Tian Pan, Yibing Song, Tianyu Yang, Wenhao Jiang, and Wei Liu. Videomoco: Contrastive video representation learning with temporally adversarial examples. In CVPR, pages 11205-11214, 2021. 3", + "[54] Jungin Park, Jiyoung Lee, and Kwanghoon Sohn. Dual-path adaptation from image to video transformers. CVPR, 2023. 3", + "[55] Mandela Patrick, Dylan Campbell, Yuki Asano, Ishan Misra, Florian Metze, Christoph Feichtenhofer, Andrea Vedaldi, and João F Henriques. Keeping your eye on the ball: Trajectory attention in video transformers. NeurIPS, 34:12493-12506, 2021. 3", + "[56] Sarah M. Pratt, Rosanne Liu, and Ali Farhadi. What does a platypus look like? generating customized prompts for zero-shot image classification. ICCV, 2022. 3", + "[57] Rui Qian, Tianjian Meng, Boqing Gong, Ming-Hsuan Yang, Huisheng Wang, Serge Belongie, and Yin Cui. Spatiotemporal contrastive video representation learning. In CVPR, pages 6964-6974, 2021. 3", + "[58] Zhiwu Qing, Shiwei Zhang, Ziyuan Huang, Yingya Zhang, Changxin Gao, Deli Zhao, and Nong Sang. Disentangling spatial and temporal learning for efficient image-to-video transfer learning. ICCV, 2023. 1, 2, 5, 7, 8", + "[59] Zhaofan Qiu, Ting Yao, and Tao Mei. Learning spatiotemporal representation with pseudo-3d residual networks. In ICCV, pages 5533–5541, 2017. 3", + "[60] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, pages 8748-8763. PMLR, 2021. 1, 2, 3, 4, 5, 6", + "[61] Kanchana Ranasinghe and Michael S. Ryoo. Language-based action concept spaces improve video self-supervised learning. CoRR, abs/2307.10922, 2023. 3", + "[62] Michael Ryoo, AJ Piergiovanni, Anurag Arnab, Mostafa Dehghani, and Anelia Angelova. Tokenlearner: Adaptive space-time tokenization for videos. *NeurIPS*, 34:12786-12797, 2021. 3", + "[63] Laura Sevilla-Lara, Shengxin Zha, Zhicheng Yan, Vedanuj Goswami, Matt Feiszli, and Lorenzo Torresani. Only time can tell: Discovering temporal data for temporal modeling. In WACV, pages 535-544, 2021. 2, 4", + "[64] Karen Simonyan and Andrew Zisserman. Two-stream convolutional networks for action recognition in videos. NeurIPS, 27, 2014. 3", + "[65] Khurram Soomro, Amir Roshan Zamir, and Mubarak Shah. Ucf101: A dataset of 101 human actions classes from videos in the wild. arXiv preprint arXiv:1212.0402, 2012. 2, 5, 8", + "[66] Zhan Tong, Yibing Song, Jue Wang, and Limin Wang. Videomae: Masked autoencoders are data-efficient learners for self-supervised video pre-training. NeurIPS, 2022. 3, 5", + "[67] Du Tran, Lubomir Bourdev, Rob Fergus, Lorenzo Torresani, and Manohar Paluri. Learning spatiotemporal features" + ], + "bbox": [ + 78, + 90, + 468, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "with 3d convolutional networks. In ICCV, pages 4489-4497, 2015. 3", + "[68] Du Tran, Heng Wang, Lorenzo Torresani, Jamie Ray, Yann LeCun, and Manohar Paluri. A closer look at spatiotemporal convolutions for action recognition. In CVPR, pages 6450-6459, 2018.", + "[69] Du Tran, Heng Wang, Lorenzo Torresani, and Matt Feiszli. Video classification with channel-separated convolutional networks. In ICCV, pages 5552-5561, 2019. 3", + "[70] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In NeurIPS, pages 5998-6008, 2017. 4", + "[71] Heng Wang, Du Tran, Lorenzo Torresani, and Matt Feiszli. Video modeling with correlation networks. In CVPR, pages 352-361, 2020. 3", + "[72] Junke Wang, Dongdong Chen, Chong Luo, Xiyang Dai, Lu Yuan, Zuxuan Wu, and Yu-Gang Jiang. Chatvideo: A tracklet-centric multimodal and versatile video understanding system. CoRR, abs/2304.14407, 2023. 3", + "[73] Limin Wang, Yuanjun Xiong, Zhe Wang, Yu Qiao, Dahua Lin, Xiaou Tang, and Luc Van Gool. Temporal segment networks for action recognition in videos. TPAMI, 41(11): 2740-2755, 2018. 3", + "[74] Limin Wang, Bingkun Huang, Zhiyu Zhao, Zhan Tong, Yi-nan He, Yi Wang, Yali Wang, and Yu Qiao. Videomae V2: scaling video masked autoencoders with dual masking. In CVPR, pages 14549-14560, 2023. 3", + "[75] Mengmeng Wang, Jiazheng Xing, and Yong Liu. Actionclip: A new paradigm for video action recognition. arXiv preprint arXiv:2109.08472, 2021. 1, 2, 8", + "[76] Peng Wang, An Yang, Rui Men, Junyang Lin, Shuai Bai, Zhikang Li, Jianxin Ma, Chang Zhou, Jingren Zhou, and Hongxia Yang. OFA: unifying architectures, tasks, and modalities through a simple sequence-to-sequence learning framework. In ICML, pages 23318-23340, 2022. 2", + "[77] Rui Wang, Dongdong Chen, Zuxuan Wu, Yinpeng Chen, Xiyang Dai, Mengchen Liu, Yu-Gang Jiang, Luowei Zhou, and Lu Yuan. Bevt: Bert pretraining of video transformers. In CVPR, pages 14733-14743, 2022. 3", + "[78] Syed Talal Wasim, Muzammal Naseer, Salman H. Khan, Fahad Shahbaz Khan, and Mubarak Shah. Vita-clip: Video and text adaptive CLIP via multimodal prompting. CVPR, 2023. 1, 2, 8", + "[79] Chen Wei, Haoqi Fan, Saining Xie, Chao-Yuan Wu, Alan Yuille, and Christoph Feichtenhofer. Masked feature prediction for self-supervised visual pre-training. In CVPR, pages 14668-14678, 2022. 3", + "[80] Wenhao Wu, Zhun Sun, and Wanli Ouyang. Revisiting classifier: Transferring vision-language models for video recognition. In AAAI, pages 2847-2855, 2023. 1", + "[81] Hu Xu, Gargi Ghosh, Po-Yao Huang, Dmytro Okhonko, Armen Aghajanyan, Florian Metze, Luke Zettlemoyer, and Christoph Feichtenhofer. Videoclip: Contrastive pre-training for zero-shot video-text understanding. In EMNLP, pages 6787-6800, 2021. 2" + ], + "bbox": [ + 503, + 92, + 890, + 898 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "18514", + "bbox": [ + 480, + 945, + 519, + 955 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[82] Jiarui Xu, Shalini De Mello, Sifei Liu, Wonmin Byeon, Thomas M. Breuel, Jan Kautz, and Xiaolong Wang. Groupvit: Semantic segmentation emerges from text supervision. In CVPR, pages 18113-18123, 2022. 2", + "[83] An Yan, Yu Wang, Yiwu Zhong, Chengyu Dong, Zexue He, Yujie Lu, William Wang, Jingbo Shang, and Julian J. McAuley. Learning concise and descriptive attributes for visual recognition. ICCV, 2023. 3", + "[84] Shen Yan, Xuehan Xiong, Anurag Arnab, Zhichao Lu, Mi Zhang, Chen Sun, and Cordelia Schmid. Multiview transformers for video recognition. In CVPR, pages 3333-3343, 2022. 7", + "[85] Yue Yang, Artemis Panagopoulou, Shenghao Zhou, Daniel Jin, Chris Callison-Burch, and Mark Yatskar. Language in a bottle: Language model guided concept bottlenecks for interpretable image classification. In CVPR, pages 19187-19197, 2023. 3", + "[86] Zhengyuan Yang, Linjie Li, Jianfeng Wang, Kevin Lin, Ehsan Azarnasab, Faisal Ahmed, Zicheng Liu, Ce Liu, Michael Zeng, and Lijuan Wang. MM-REACT: prompting chatgpt for multimodal reasoning and action. CoRR, abs/2303.11381, 2023. 3", + "[87] Jiahui Yu, Zirui Wang, Vijay Vasudevan, Legg Yeung, Mojtaba Seyedhosseini, and Yonghui Wu. Coca: Contrastive captioners are image-text foundation models. arXiv preprint arXiv:2205.01917, 2022. 1, 2", + "[88] Lu Yuan, Dongdong Chen, Yi-Ling Chen, Noel Codella, Xiyang Dai, Jianfeng Gao, Houdong Hu, Xuedong Huang, Boxin Li, Chunyuan Li, et al. Florence: A new foundation model for computer vision. arXiv preprint arXiv:2111.11432, 2021. 1, 2", + "[89] Yue Zhao, Ishan Misra, Philipp Krahenbuhl, and Rohit Girdhar. Learning video representations from large language models. In CVPR, pages 6586-6597. IEEE, 2023. 3", + "[90] Yiwu Zhong, Jianwei Yang, Pengchuan Zhang, Chunyuan Li, Noel Codella, Liunian Harold Li, Luowei Zhou, Xiyang Dai, Lu Yuan, Yin Li, and Jianfeng Gao. Regionclip: Region-based language-image pretraining. In CVPR, pages 16772-16782, 2022. 2", + "[91] Deyao Zhu, Jun Chen, Xiaoqian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models. CoRR, abs/2304.10592, 2023. 3" + ], + "bbox": [ + 78, + 90, + 468, + 696 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "18515", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 11 + } +] \ No newline at end of file diff --git a/2024/Enhanced Motion-Text Alignment for Image-to-Video Transfer Learning/1b431374-7f10-4d10-9e33-68850035e408_model.json b/2024/Enhanced Motion-Text Alignment for Image-to-Video Transfer Learning/1b431374-7f10-4d10-9e33-68850035e408_model.json new file mode 100644 index 0000000000000000000000000000000000000000..babdf0ef48df4634ec3ee22701530dd4243316c4 --- /dev/null +++ b/2024/Enhanced Motion-Text Alignment for Image-to-Video Transfer Learning/1b431374-7f10-4d10-9e33-68850035e408_model.json @@ -0,0 +1,2589 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.119, + 0.131, + 0.853, + 0.154 + ], + "angle": 0, + "content": "Enhanced Motion-Text Alignment for Image-to-Video Transfer Learning" + }, + { + "type": "text", + "bbox": [ + 0.145, + 0.18, + 0.865, + 0.199 + ], + "angle": 0, + "content": "Wei Zhang\\(^{1,2*}\\), Chaoqun Wan\\(^{2}\\), Tongliang Liu\\(^{3}\\), Xinmei Tian\\(^{1,4\\dagger}\\), Xu Shen\\(^{2\\dagger}\\), Jieping Ye\\(^{2}\\)" + }, + { + "type": "text", + "bbox": [ + 0.139, + 0.2, + 0.873, + 0.234 + ], + "angle": 0, + "content": "1University of Science and Technology of China, 2Alibaba Cloud, 3The University of Sydney Institute of Artificial Intelligence, Hefei Comprehensive National Science Center" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.235, + 0.924, + 0.272 + ], + "angle": 0, + "content": "zw0819@mail.ustc.edu.cn, tongliang.liu@sydney.edu.au, xinmei@ustc.edu.cn {qionlong.wcq, shenxu.sx, jieping.yjp}@alibaba-inc.com" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.305, + 0.314, + 0.321 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.336, + 0.474, + 0.716 + ], + "angle": 0, + "content": "Extending large image-text pre-trained models (e.g., CLIP) for video understanding has made significant advancements. To enable the capability of CLIP to perceive dynamic information in videos, existing works are dedicated to equipping the visual encoder with various temporal modules. However, these methods exhibit \"asymmetry\" between the visual and textual sides, with neither temporal descriptions in input texts nor temporal modules in text encoder. This limitation hinders the potential of language supervision emphasized in CLIP, and restricts the learning of temporal features, as the text encoder has demonstrated limited proficiency in motion understanding. To address this issue, we propose leveraging \"MoTion-Enhanced Descriptions\" (MoTED) to facilitate the extraction of distinctive temporal features in videos. Specifically, we first generate discriminative motion-related descriptions via querying GPT-4 to compare easy-confusing action categories. Then, we incorporate both the visual and textual encoders with additional perception modules to process the video frames and generated descriptions, respectively. Finally, we adopt a contrastive loss to align the visual and textual motion features. Extensive experiments on five benchmarks show that MoTED surpasses state-of-the-art methods with convincing gaps, laying a solid foundation for empowering CLIP with strong temporal modeling." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.74, + 0.21, + 0.755 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.766, + 0.47, + 0.857 + ], + "angle": 0, + "content": "Recent years have witnessed remarkable achievements in contrastive language-image pre-training models [9, 25, 36, 60, 87, 88], with CLIP [60] emerging as the front-runner. Through language supervision with a vast collection of 400 million image-text pairs, CLIP has achieved exceptional image comprehension and unprecedented zero-shot general-" + }, + { + "type": "page_footnote", + "bbox": [ + 0.077, + 0.863, + 0.47, + 0.887 + ], + "angle": 0, + "content": "*This work was done when the author was visiting Alibaba as a research intern." + }, + { + "type": "page_footnote", + "bbox": [ + 0.098, + 0.888, + 0.222, + 0.9 + ], + "angle": 0, + "content": "† Corresponding author." + }, + { + "type": "list", + "bbox": [ + 0.077, + 0.863, + 0.47, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.304, + 0.891, + 0.515 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.613, + 0.517, + 0.778, + 0.529 + ], + "angle": 0, + "content": "(c) Motion-Enhanced Descriptions" + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.545, + 0.894, + 0.643 + ], + "angle": 0, + "content": "Figure 1. (a) Previous methods [27, 49, 75, 78] align the visual embeddings with the textual embeddings derived from category names. (b) Our work symmetrically aligns the spatial/temporal visual embeddings with class descriptions and motion descriptions correspondingly. (c) The descriptions generated by large language modes (e.g., GPT-4 [50]) enhance the conceptual definitions and discriminative details of motions." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.672, + 0.892, + 0.763 + ], + "angle": 0, + "content": "ization. This breakthrough has opened up new possibilities for leveraging the power of large-scale pre-trained models to comprehend videos. It has also introduced a new paradigm [26, 40, 49, 52, 58, 75, 80] that endows image-based CLIP to effectively perceive dynamic information for video recognition." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.893, + 0.902 + ], + "angle": 0, + "content": "To equip CLIP with motion perception, existing works propose to incorporate various temporal modules into the visual encoder. These temporal modules are additional tunable parameters, including designed efficient units between pre-trained transformer blocks [49, 75], and parallel structures to disentangle the temporal modeling out of the spatial modeling [40, 58]. However, while considerable effort has been put into capturing temporal visual features from videos, these methods have paid little attention to the in" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "18504" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.35 + ], + "angle": 0, + "content": "put text and text encoder, leading to an imbalance between the visual and textual alignment. When transferring to the classification task of video recognition (Fig. 1(a)), the only available text is the \"category names\" of the actions, e.g., \"clapping\" and \"slapping\". These coarse-grained word-level descriptors lack clear descriptions and explanations, making it difficult to distinguish between them. For instance, \"slapping\" refers to a quick hand movement striking against something, while \"clapping\" refers to the act of striking both hands together rapidly. The scarcity of textual motion information contradicts the language supervision principle of CLIP. Furthermore, it's observed that the CLIP text encoder delivers a strong bias towards spatial concepts (e.g., nouns) [5, 30, 47, 63], with a weak understanding of temporal cues (e.g., verbs). These issues significantly limit the effectiveness of extending CLIP with motion modeling for video understanding." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.353, + 0.473, + 0.641 + ], + "angle": 0, + "content": "To overcome the limitations of language supervision on temporal modeling, we propose a novel approach that symmetrically aligns the spatial/temporal visual embeddings with class descriptions (\"slapping\") and motion descriptions correspondingly (\"a quick hand movement striking against something\"). We introduce the MoTion-Enhanced Descriptions (MoTED) as the language supervision for the visual temporal modules. In the implementation, we encounter two main challenges: i) how to generate motion-related descriptions? ii) how to effectively utilize descriptions as supervision? To tackle the first challenge, we aim for clear, detailed, and distinctive descriptions in the text. With the recent advancements in large language models (LLMs) [4, 11, 50], it has become possible to automatically generate conceptual descriptions for corresponding actions. As for the second challenge, we disentangle the spatial and temporal learning by constructing a temporal encoder, in parallel to the CLIP image encoder, to extract the temporal features aligned with motion-enhanced descriptions." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.645, + 0.473, + 0.903 + ], + "angle": 0, + "content": "To be specific, we employ GPT-4 [50] to generate descriptions by posing a query: “Q: what is the motion concept of ? A:” (as shown in Fig. 1(c)). However, our empirical findings indicate that only the concept descriptions offer marginal assistance in motion perception, particularly on datasets such as Something-Something-V2 (SSv2) [20]) that necessitate the differentiation of similar classes using intricate motion cues. To mitigate this issue, we propose to incorporate related top-k actions as context for the LLMs to generate more discriminative descriptions. These generated texts are then passed through the CLIP text encoder to extract textual motion representations, where an adapter is utilized to eliminate the biases in the original model. Correspondingly, the temporal module in the visual encoder processes the video to extract dynamic features and aligns them with the generated textual descriptions through contrastive learning. Finally, these two sets of" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.227 + ], + "angle": 0, + "content": "features are independently fused with the original CLIP's text and image features using cross-attention, and subsequently aligned. Evaluated on two supervised video recognition benchmarks, i.e., Kinetics-400 [28] and SSv2 [20], as well as three zero-shot benchmarks, i.e., Kinetics-600 [12], HMDB51 [24], UCF101 [65], the proposed MoTED surpasses state-of-the-art methods with convincing gaps, indicating the effectiveness of aligning enhanced motion descriptions with the temporal embedding of input videos." + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.228, + 0.81, + 0.242 + ], + "angle": 0, + "content": "We summarize the contributions as follows:" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.243, + 0.892, + 0.317 + ], + "angle": 0, + "content": "- We present a new perspective that underscores the significance of textual side on par with visual side. By delving into in-depth distinguishing descriptions of actions, we make the first attempt to reveal the potential of language supervision as emphasized in CLIP." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.318, + 0.891, + 0.378 + ], + "angle": 0, + "content": "- We propose MoTED that leverages LLMs to automatically generate action descriptions, and mining distinctive descriptions among similar actions. Then, we use a parallel path for both visual and textual motion modeling." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.379, + 0.892, + 0.439 + ], + "angle": 0, + "content": "- We evaluate our approach on supervised as well as generalization tasks. Extensive experiments demonstrate the superiority and good generalization ability of the proposed method." + }, + { + "type": "list", + "bbox": [ + 0.5, + 0.243, + 0.892, + 0.439 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.453, + 0.642, + 0.469 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.478, + 0.892, + 0.825 + ], + "angle": 0, + "content": "Vision-Language Pre-training. In recent years, Vision-Language Pre-training (VLP) [9, 25, 36, 45, 46, 60, 76, 87, 88] has made remarkable progress. One of the most remarkable and influential works is CLIP [60], which adopts the contrastive language-image pretraining paradigm. Following that, this paradigm has shown impressive zero-shot generalization capabilities on various image-related tasks [31, 35, 37, 43, 82, 90]. However, pre-training a language-video model [34, 81] is prohibitively expensive, as it requires large-scale video-text data and extensive training resources (e.g., thousands of GPU days). Meanwhile, transferring language-image pre-trained models to the video domain [10, 26, 40, 49, 52, 58, 75, 78] has captured significant attention due to its striking performance and training efficiency. For instance, X-CLIP [49] integrates the CLIP image encoder with a cross-frame attention module for temporal modeling. Vita-CLIP [78] utilizes multi-modal prompting techniques to learn video and text-specific context vectors. DiST [58] is the most related work that also disentangles spatial and temporal learning in the visual side. In contrast, our MoTED highlights the effectiveness of temporal disentangling both in the visual and textual sides, supervised by the detailed descriptions of motions." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.826, + 0.894, + 0.903 + ], + "angle": 0, + "content": "Video Recognition. The conventional approaches in video recognition primarily focus on spatio-temporal learning under fully-supervised settings, where all categories are predefined. These approaches have achieved remarkable performance using various architectures, including convolution" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18505" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.088, + 0.089, + 0.895, + 0.308 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.321, + 0.895, + 0.392 + ], + "angle": 0, + "content": "Figure 2. The overall framework of MoTED. Building upon the dual image and text encoders initialized by CLIP [60], we extend the capability to perceive motion information symmetrically in both sides. In the text side, the motion-related conceptual and discriminative descriptions are generated via querying LLMs (e.g., GPT-4) using category names as input. In the vision side, a motion encoder is built to extract temporal dynamics given densely-sampled frames and integrated with the middle-layer features from image encoder. To align the motion embeddings and video embeddings from both sides, it employs two contrastive losses, respectively." + }, + { + "type": "table", + "bbox": [ + 0.079, + 0.415, + 0.477, + 0.539 + ], + "angle": 0, + "content": "
Related workTaskAction annotationExample
Dist. Sup. [39]pretrainevent description‘insert window mounting bolts’
Weakly Sup. [14]pretrainevent description‘take up the iron clamp’
LAVILA [89]pretrainevent description‘a lady walks past a car’
VFC [47]VL transfer learningevent description‘two brown horses eating grass’
MoTED (Ours)VL transfer learningmotion concept description‘slapping: a quick hand movement striking against something’
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.543, + 0.471, + 0.571 + ], + "angle": 0, + "content": "Table 1. Comparison of our method with related works focused on action annotations. 'VL' is the abbreviation of 'vision-language'." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.599, + 0.473, + 0.903 + ], + "angle": 0, + "content": "networks [7, 17, 21, 59, 64, 67–69, 71, 73], and vision transformers [1, 2, 6, 16, 32, 33, 38, 42, 48, 54, 55, 62]. In addition to the architecture design, self-supervised video representation learning [13, 18, 19, 23, 53, 57, 66, 74, 77, 79] has also gained popularity recently. However, these methods operate purely within the visual domain and meet bottlenecks of recognizing unseen or unfamiliar categories in real-world applications. Fortunately, the advance of large language models (LLMs) [4, 51] provides opportunities to mitigate this issue, due to their powerful capabilities of encoding world knowledge [29]. Recent studies [44, 56, 83, 85] have verified that factual sentences generated by LLMs can improve zero-shot image recognition accuracy. VFC [47] leverages PaLM [11] to create verb-focused hard negatives to enhance the understanding of verbs in video models. LSS [61] integrates language-based action concepts with self-supervised learning to adapt an image model to video domain. Our MoTED employs the descriptions that delineate the labeled classes and further accentuate the useful traits distinguishing similar classes." + }, + { + "type": "table", + "bbox": [ + 0.502, + 0.415, + 0.899, + 0.542 + ], + "angle": 0, + "content": "
Related workLLMs roleExample
Chatvideo [72]manager& summarizeInput: 'Summarize the activity in video' \nOutput: 'A person is cooking in the kitchen'
MM-REACT [86]execution& summarizeInput: 'Please create a summary of the video' \nOutput: 'The speaker is making a BLT sandwich ...'
MiniGPT4 [91]/LLAVA [41]KB& data cleanInput: 'Describe this image in detail.' \nOutput: 'The image shows a group of musicians ...'
MoTED (Ours)generate descriptionsInput: 'What is the motion concept of slapping' \nOutput: 'A quick hand movement striking against ...'
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.546, + 0.895, + 0.575 + ], + "angle": 0, + "content": "Table 2. Comparison of our method with related works that take LLMs as a knowledge base (KB) and automatically annotated tool." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.593, + 0.591, + 0.609 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.62, + 0.892, + 0.801 + ], + "angle": 0, + "content": "The generic pipeline of MoTED is presented in Fig. 2, which consists of three steps: (1) Motion Description Generation to obtain conceptual descriptions \\(\\mathcal{D}_{con}\\) and discriminative descriptions \\(\\mathcal{D}_{dis}\\), via querying language models given the category names of the target video dataset in Sec. 3.1; (2) Textual Motion Adaptation to obtain the text embeddings that extract the motion semantics given the generated descriptions \\(\\mathcal{D}_{con}\\) and \\(\\mathcal{D}_{dis}\\) in Sec. 3.2; (3) Visual Motion Extraction to obtain the visual embeddings that extract the motion semantics given input video frames in Sec. 3.3. Overall, the framework is trained in an end-to-end contrastive manner as illustrated in Sec. 3.4." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.816, + 0.779, + 0.832 + ], + "angle": 0, + "content": "3.1. Motion Description Generation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.894, + 0.903 + ], + "angle": 0, + "content": "The first step is to obtain a set of appropriate descriptions for each category. Given a video dataset with multiple different categories (e.g., Kinetics400 [28] with 400 action classes), the descriptions are generated automati" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "18506" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.091, + 0.473, + 0.304 + ], + "angle": 0, + "content": "cally utilizing GPT-4 [50]. Note that, the generation process is agnostic to this choice and other LLMs can be used instead. For each category name, we query GPT-4 to provide the motion concepts using the following prompt: “Q: What is the motion concept in a video of ?” A: As shown in Fig. 1, the generated concept descriptions often cover moving objects, object interactions, moving directions and speeds, etc. But the output of LLMs can also be open-ended duplicates or anything in natural language (as illustrated in the Appendix). To control the generated descriptions to be concise and motion-related, we adopt a two-shot prompt in which we include two exemplars of question-answer for the same operation that is being queried." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.304, + 0.473, + 0.469 + ], + "angle": 0, + "content": "Besides the conceptual descriptions \\(\\mathcal{D}_{con}\\), we further investigate into the confusing actions that are characterized by their distinct differences from similar categories. To this end, we first compute the cosine distance of text embeddings for every two categories. Then we select top-k similar classes (e.g., \\(k = 5\\)) and query GPT-4 to generate motion characteristics for distinguishing the similar classes using the following prompt: \"Q: What are the useful features for distinguishing the original class from the similar class " + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.47, + 0.47, + 0.531 + ], + "angle": 0, + "content": "A:\" Different from the conceptual descriptions \\(\\mathcal{D}_{con}\\) that are mainly determined by the general knowledge of LLMs, the descriptions \\(\\mathcal{D}_{dis}\\) further contain task-specific information that the downstream task emphasizes." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.532, + 0.47, + 0.607 + ], + "angle": 0, + "content": "In this way, given a dataset of \\(N\\) categories, we can obtain the motion-enhanced \\(N * (k + 1)\\) descriptions, consisting of 1 conceptual description and \\(k\\) discriminative descriptions for each category. The above generation process is completed before the training of whole framework." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.608, + 0.471, + 0.805 + ], + "angle": 0, + "content": "The detailed comparisons of action annotations and LLM-based generations are presented in Tab. 1 and Tab. 2, respectively. Tab. 1 shows that existing works employ annotations that describe the action events in a subject-verb-object manner, which are object-centric. In contrast, our generated annotations are motion-related, object-agnostic and to supervise the learning of visual motion features with the world knowledge of motion concepts. In Tab. 2, LLMs are mainly applied in generative tasks, to summarize query results and generate general responses, or to obtain instruction data to align with human preference. In contrast, we apply LLMs to obtain descriptions of motion concept, aligned with the temporal visual features." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.816, + 0.325, + 0.833 + ], + "angle": 0, + "content": "3.2. Textual Motion Adaptation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.471, + 0.902 + ], + "angle": 0, + "content": "After generating the motion-enhanced descriptions, the second step is to perceive the motion cues within the descriptions and aggregate them into a compact motion-enhanced embedding. Given a set of motion-enhanced descriptions \\(\\mathcal{D}\\)" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.089, + 0.892, + 0.218 + ], + "angle": 0, + "content": "\\(= \\left\\{\\mathcal{D}_{con}^{i},\\mathcal{D}_{dis}^{i,j}\\right\\}\\) where \\(i\\in [1,N]\\) and \\(j\\in [1,k]\\), we extract normalized feature embedding \\(e_i\\) by using the text encoder \\(f_{txt}\\): \\(e_i = f_{txt}(d_i)\\), where \\(d_{i}\\) is a motion-enhanced description sampled from \\(\\mathcal{D}\\). In this way, we obtain the text embeddings \\(E_{txt}\\in \\mathbb{R}^{N\\times (k + 1)\\times C}\\), where \\(C\\) denotes the channel number of each embedding. Noticeably, the parameters of text encoder is frozen and initialized by CLIP [60] to inherit its capacity of encoding visual-aligned semantics." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.218, + 0.893, + 0.4 + ], + "angle": 0, + "content": "However, as CLIP is pre-trained on image-text paired corpus, the text encoder has a strong bias towards spatial appearance of objects and backgrounds, instead of temporal motions [5, 30, 47, 63]. To adapt the text encoder to understand the motion-enhanced descriptions, we introduce a Text Motion Adapter, which consists of a multi-head self-attention (MHSA) [70]. The adapter is parameter-tunable and takes \\( E_{txt} \\) as input to learn the information dependencies between motion-enhanced descriptions. To aggregate the motion semantics for each category, we perform adaptive pooling to obtain the motion-enhanced embeddings \\( E_{txt}^{m} \\in \\mathbb{R}^{N \\times 1 \\times C} \\)." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.406, + 0.733, + 0.421 + ], + "angle": 0, + "content": "3.3. Visual Motion Extraction" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.43, + 0.892, + 0.536 + ], + "angle": 0, + "content": "In the visual side, as shown in Fig. 3, it comprises of two components: (i) The image encoder adopts the CLIP pre-trained Vision Transformer (ViT), which extracts frozen features for sparse frames with powerful spatial semantics. (ii) The motion encoder takes dense frames as input to capture the local temporal cues, integrated with the global temporal dynamics from middle-layer image features." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.553, + 0.662, + 0.568 + ], + "angle": 0, + "content": "3.3.1 Image Encoder" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.575, + 0.892, + 0.79 + ], + "angle": 0, + "content": "Given a video clip \\(\\mathbf{X}_S\\in \\mathbb{R}^{T\\times H\\times W\\times 3}\\) (\\(T\\), \\(H\\), and \\(W\\) represent the frame number, height, and width, respectively), the image features are extracted individually for several sparse frames. Following ViT [15], each frame is divided into \\(K = \\frac{H}{P}\\times \\frac{W}{P}\\) patches, and the size of each patch is denoted as \\(P\\times P\\). These patches are then projected using a fully connected layer, referred to as the 2D stem in Fig. 3. This projection generates a sequence of patch embeddings \\([\\mathbf{x}_{t,\\mathrm{cls}}^{(0)},\\mathbf{x}_{t,1}^{(0)},\\mathbf{x}_{t,2}^{(0)},\\dots ,\\mathbf{x}_{t,K}^{(0)}] + \\mathbf{e}^{\\mathrm{spatial}}\\), where \\(t = \\{1,\\dots ,T\\}\\), \\(\\mathbf{x}_{\\mathrm{cls}}\\) is an additional learnable token (termed as \"class embedding\"), and \\(\\mathbf{e}^{\\mathrm{spatial}}\\) denotes spatial position embedding. Assuming that the spatial encoder has \\(L\\) Transformer blocks, the features of the \\(l_{\\mathrm{th}}\\) layer for the \\(t_\\mathrm{th}\\) frame can be extracted by:" + }, + { + "type": "equation", + "bbox": [ + 0.542, + 0.796, + 0.892, + 0.817 + ], + "angle": 0, + "content": "\\[\n\\mathbf {X} _ {t} ^ {(l)} = \\operatorname {T r a n s f o r m e r} ^ {(l)} \\left(\\mathbf {X} _ {t} ^ {(l - 1)}\\right) \\in \\mathbb {R} ^ {(K + 1) \\times C}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.823, + 0.892, + 0.902 + ], + "angle": 0, + "content": "where \\( l = \\{1, \\dots, L\\} \\) denotes the layer index. The class embeddings of \\( T \\) frames in the \\( l_{\\mathrm{th}} \\) layer is termed as \\( \\mathbf{X}^{(l)} = [\\mathbf{X}_1^{(l)}, \\dots, \\mathbf{X}_T^{(l)}] \\in \\mathbb{R}^{T \\times 1 \\times C} \\). Benefit from the CLIP pretrained parameters, the class embeddings \\( \\mathbf{X}^{(L)} \\) aggregate powerful spatial semantics within each frame." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18507" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.084, + 0.091, + 0.468, + 0.282 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.092, + 0.296, + 0.454, + 0.31 + ], + "angle": 0, + "content": "Figure 3. The structural details of MoTED in the vision side." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.338, + 0.246, + 0.352 + ], + "angle": 0, + "content": "3.3.2 Motion Encoder" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.362, + 0.473, + 0.574 + ], + "angle": 0, + "content": "To fully extract the motion information in videos, the temporal input \\(\\mathbf{X}_T\\in \\mathbb{R}^{\\alpha T\\times H\\times W\\times 3}\\) for the motion encoder is sampled around the spatial input \\(\\mathbf{X}_S\\) by \\(\\alpha\\) times. In this study, we set \\(\\alpha = 2\\) by default, as empirically studied in [17, 58]. Then, \\(\\mathbf{X}_T\\) is projected by a 3D convolution, i.e., the 3D stem in Fig. 3, for patch embedding. The kernel size and stride of the 3D stem are both \\(P\\) in spatial dimension and \\(\\alpha\\) in temporal dimension. Thus the number of temporal patch tokens are the same with that in the image encoder, which makes it convenient to integrate the spatial and temporal features. Thus the projected temporal features can be formulated as: \\(\\mathbf{Z}^{(0)} = \\mathrm{Conv3d}(\\mathbf{X}_T)\\in \\mathbb{R}^{T\\times K\\times C}\\), where \\(K = \\frac{H}{P}\\times \\frac{W}{P}\\). Then, a series of Temporal Blocks are designed to extract motion patterns, which can be written as:" + }, + { + "type": "equation", + "bbox": [ + 0.089, + 0.598, + 0.47, + 0.617 + ], + "angle": 0, + "content": "\\[\n\\mathbf {Z} ^ {(l)} = \\operatorname {T e m p - B l o c k} ^ {(l)} \\left(\\mathbf {Z} ^ {(l - 1)}, \\mathbf {X} ^ {(l - 1)}\\right) \\in \\mathbb {R} ^ {T \\times K \\times C}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.629, + 0.47, + 0.659 + ], + "angle": 0, + "content": "where the function Temp-Block \\((\\cdot)\\) performs temporal modeling integrated with the spatial features \\(X^{(l - 1)}\\)" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.66, + 0.473, + 0.902 + ], + "angle": 0, + "content": "Following the effective and efficient design philosophy in [32, 33], the temporal block consists of the following three modules: (i) Local Motion Extraction (LME). Since 3D convolution can capture detailed and local spatiotemporal features, by processing each pixel with context from a small 3D neighborhood (e.g., \\(2 \\times 16 \\times 16\\)), the temporal patch tokens \\(\\mathbf{Z}^{(0)}\\) remain the nature of local motion extraction. Then these tokens are flattened to a sequence \\(\\mathbb{R}^{TK \\times C}\\) and perform spatiotemporal learning via joint self-attention modules [66] and token merging for efficiency [3]. (ii) Global Motion Extraction (GME). Given the class embeddings \\(\\mathbf{X}^{(l)}\\) that aggregate powerful spatial semantics for each individual frame, we apply a cross-frame self-attention module [49] to learn the global inter-frame interactions. (iii) Feature Integration. After obtaining the local and global motion features, we adopt a cross-attention mod" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.168 + ], + "angle": 0, + "content": "ule, which takes local features as query and global features as key/value, to complement the features for better temporal modeling. In this way, the output of the final temporal block \\(\\mathbf{Z}^{(L)}\\) aggregates powerful temporal dynamics within the dense frames." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.178, + 0.644, + 0.194 + ], + "angle": 0, + "content": "3.4. Training Loss" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.201, + 0.892, + 0.264 + ], + "angle": 0, + "content": "Following CLIP [60], we first apply adaptive pooling to obtain a video-level motion embedding \\(\\mathbf{Z}_{avg}^{(L)} \\in \\mathbb{R}^{1 \\times C}\\). Then we introduce a contrastive loss to align the visual and textual motion embeddings:" + }, + { + "type": "equation", + "bbox": [ + 0.541, + 0.276, + 0.892, + 0.318 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {m o t i o n}} = - \\log \\frac {\\exp \\left(\\sin \\left(\\mathbf {Z} _ {\\text {a v g}} ^ {(L)} , \\mathbf {e} _ {i}\\right) / \\tau\\right)}{\\sum_ {c = 1} ^ {N} \\exp \\left(\\sin \\left(\\mathbf {Z} _ {\\text {a v g}} ^ {(L)} , \\mathbf {e} _ {c}\\right) / \\tau\\right)}, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.327, + 0.892, + 0.373 + ], + "angle": 0, + "content": "where \\(\\mathrm{sim}(\\cdot ,\\cdot)\\) is the normalized cosine similarity, \\(\\tau\\) refers to the temperature parameter. \\(\\mathbf{e}_i\\) is a motion-enhanced text embedding for the \\(i_{\\mathrm{th}}\\) category." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.373, + 0.892, + 0.464 + ], + "angle": 0, + "content": "Moreover, we adopt a cross-attention module to fuse the visual motion features (as key/value) and image features (as query) to obtain the compact video embedding \\(\\mathbf{v}_i\\). Similarly, the fused text embedding \\(\\mathbf{e}_i'\\) is acquired via a cross-attention module. We introduce a contrastive loss to align the visual and textual video-level embeddings:" + }, + { + "type": "equation", + "bbox": [ + 0.555, + 0.475, + 0.892, + 0.512 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {v i d e o}} = - \\log \\frac {\\exp \\left(\\operatorname {s i m} \\left(\\mathbf {v} _ {i} , \\mathbf {e} _ {i} ^ {\\prime}\\right) / \\tau\\right)}{\\sum_ {c = 1} ^ {N} \\exp \\left(\\operatorname {s i m} \\left(\\mathbf {v} _ {i} , \\mathbf {e} _ {c} ^ {\\prime}\\right) / \\tau\\right)}. \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.522, + 0.892, + 0.553 + ], + "angle": 0, + "content": "Compatible with CLIP training paradigm, the overall framework is trained based on these two contrastive losses:" + }, + { + "type": "equation", + "bbox": [ + 0.616, + 0.57, + 0.892, + 0.585 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\mathcal {L} _ {\\text {m o t i o n}} + \\mathcal {L} _ {\\text {v i d e o}}. \\tag {5}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.599, + 0.634, + 0.617 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.635, + 0.76, + 0.652 + ], + "angle": 0, + "content": "4.1. Dataset and Implementation." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.659, + 0.892, + 0.87 + ], + "angle": 0, + "content": "Datasets. In the supervised setting, we train on the train set of Kinetics-400 (K400) [28] and Something-Something-V2 (SSv2) [20] and report supervised performance against existing methods on the validation sets of K400 and SSv2. In the zero-shot setting, we train on the Kinetics-400 training set and evaluate on three datasets: Kinetics-600 (K600) [12], HMDB51 [24] and UCF101 [65]. For zero-shot evaluation on K600, following [12], we use the 220 new categories outside of K400 for evaluation, and conduct evaluation three times, each time randomly sampling 160 categories for evaluation from the 220 categories. For zero-shot evaluation on HMDB51 and UCF101, we follow [49] and report average top-1 accuracy and standard deviation on three splits of the test set." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.871, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Implementation Details. Following previous work [40, 58], we use the CLIP [60] pre-trained ViT-B/16, ViT-L/14" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18508" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.086, + 0.103, + 0.351, + 0.178 + ], + "angle": 0, + "content": "
Class NamesConcept.Discr.SSv2K400
✓XX65.782.9
✓✓X66.084.5
✓X✓69.984.0
✓✓✓70.185.1
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.08, + 0.182, + 0.345, + 0.219 + ], + "angle": 0, + "content": "(a) \"Concept.\" is the abbreviation of \"conceptual descriptions\". \"Discr.\" is the abbreviation of \"discriminative descriptions\"." + }, + { + "type": "table", + "bbox": [ + 0.394, + 0.089, + 0.583, + 0.166 + ], + "angle": 0, + "content": "
M.Enc.Adapter.SSv2K400
XX64.381.5
✓X68.683.4
X✓66.283.8
✓✓70.185.1
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.352, + 0.181, + 0.618, + 0.219 + ], + "angle": 0, + "content": "(b) \"M.Enc.\" is the separated motion encoder where motion modules are parallel with CLIP visual encoder. \"Adapter\" is the text motion adapter." + }, + { + "type": "table", + "bbox": [ + 0.647, + 0.09, + 0.875, + 0.166 + ], + "angle": 0, + "content": "
Text.Fuse.Vision.Fuse.SSv2K400
XX68.383.6
✓X68.984.2
X✓69.584.5
✓✓70.185.1
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.626, + 0.181, + 0.89, + 0.218 + ], + "angle": 0, + "content": "(c) \"Text.Fuse.\" indicates feature fusion in the textual side. \"Vision.Fuse.\" indicates feature fusion in the visual side." + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.224, + 0.893, + 0.253 + ], + "angle": 0, + "content": "Table 3. Ablations on Something-Something-V2 and Kinetics-400. The spatial encoder is a 8-frame vanilla ViT-B/16 pre-trained by CLIP [60]. The inference protocol of all models and datasets are 3 clips × 1 center crop." + }, + { + "type": "image", + "bbox": [ + 0.079, + 0.267, + 0.48, + 0.328 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.079, + 0.33, + 0.48, + 0.383 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.187, + 0.384, + 0.381, + 0.392 + ], + "angle": 0, + "content": "Relevance between text and motion features" + }, + { + "type": "image_caption", + "bbox": [ + 0.227, + 0.394, + 0.267, + 0.403 + ], + "angle": 0, + "content": "(a) case 1" + }, + { + "type": "image", + "bbox": [ + 0.492, + 0.266, + 0.892, + 0.328 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.492, + 0.33, + 0.892, + 0.383 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.582, + 0.383, + 0.776, + 0.391 + ], + "angle": 0, + "content": "Relevance between text and motion features" + }, + { + "type": "image_caption", + "bbox": [ + 0.658, + 0.394, + 0.7, + 0.403 + ], + "angle": 0, + "content": "(b) case2" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.408, + 0.894, + 0.437 + ], + "angle": 0, + "content": "Figure 4. Two cases to visualize the relevance [8] between text and image/motion features to highlight the information relevant to the prediction. The different \"regions of interest\" and \"words of importance\" indicate that the motion and image features could be disentangled." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.463, + 0.47, + 0.584 + ], + "angle": 0, + "content": "and ViT-L/14-336p as our image encoder. Unless otherwise specified, we mark the default settings in the temporal encoder in gray in Sec. 4.2. For simplicity, the ablation studies are conducted based on ViT-B/16 with a low-rate sampled 8 frames. We conduct the experiments with the NVIDIA 32G V100 GPUs. More implementation details (e.g., training and testing hyper-parameters) are described in the Appendix." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.591, + 0.285, + 0.607 + ], + "angle": 0, + "content": "4.2. Ablation and Analysis" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.614, + 0.473, + 0.902 + ], + "angle": 0, + "content": "Language Supervision. In this section, we aim to demonstrate the significance of language supervision by comparing different types of text descriptions. Tab. 3a presents a comparison of three types of texts: category names, motion concept descriptions, motion discriminative descriptions. It reveals that the performance of action category names, which contain the least information, is notably lower than other results, particularly on the SSv2 dataset with abundant temporal information. Although descriptions with only action concepts show a slight improvement of \\(0.3\\% /1.6\\%\\), they lack the vital information regarding the distinguishing characteristics of the actions. As a result, the model exhibits poor classification accuracy for similar categories. On the other hand, when the discriminative descriptions are adopted, the accuracy on SSv2 is improved noticeably \\((+4.2\\%)\\), revealing the significance of discriminative characteristics for fine-grained datasets. Additionally, when the full descriptions are utilized, the model's performance obtain the further gains of \\(4.4\\% /2.2\\%\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.463, + 0.893, + 0.719 + ], + "angle": 0, + "content": "Disentangled Motion Modeling. Our method has an advantage in allowing the learning of dynamic information without interference, while preserving CLIP's original spatially transferable representation capabilities. To validate this, we compared two different structures in Tab. 3b. For the serial structure, additional dynamic modules are inserted between the Transformer Blocks of each layer of the visual encoder, resulting in a unified video representation. All the texts are combined through the text encoder to obtain text representations for alignment. It can be observed that the parallel structure outperforms the serial structure, with improvements of \\(4.3\\%\\) and \\(1.9\\%\\) respectively. This demonstrates the rationality of the parallel approach. Tab. 3b also presents a comparison of introducing an adapter in the text encoder. The model's classification performance improved by \\(1.9\\%\\) and \\(2.3\\%\\) after adding the adapter, indicating its effectiveness in reducing bias in the text encoder." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.895, + 0.901 + ], + "angle": 0, + "content": "Feature Fusion Direction. As shown in Tab. 3c, both directions of information integration can improve performances. The combination of the two can boost accuracy more significantly \\(+1.8\\% / + 1.5\\%\\) on SSv2/K400, respectively. This verifies the importance of spatio-temporal blending for the parallel architectures. In our opinion, independent learning of motion can effectively avoid excessive reliance of the model on the previously learned spatial information. By using cross attention for fusion, it is possible to effectively integrate features from two different dimensions, ultimately forming the features of the target video and achieving the best results." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "18509" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.093, + 0.089, + 0.88, + 0.43 + ], + "angle": 0, + "content": "
MethodPre-trainArchitectureInput SizeFLOPs×Cr.×Cl. (T)Param (M)FrozenTop-1Top-5
SlowFast [17]ImageNet-21KR101+NL16 × 22420.1 × 3 × 160X63.187.6
ViViT FE [1]IN21K+K400ViT-L16 × 22421.0 × 3 × 4612X65.489.8
MTV-B(320p) [84]IN21K+K400-32 × 22420.9 × 3 × 4310X68.590.4
MViT [16]Kinetics-600MViT-B-2432 × 22420.2 × 3 × 153X68.791.5
Video Swin [42]IN21K+K400Swin-B32 × 22420.3 × 3 × 160X69.692.7
TAdaConvNeXtV2 [22]IN1K+K400ConvNeXt-S32 × 22420.2 × 3 × 282X70.092.0
EVL* [40]CLIP-400MViT-B32 × 22420.68 × 1 × 3175✓62.4-
ST-Adapter* [52]CLIP-400MViT-B32 × 22420.61 × 1 × 393✓69.592.6
DiST*CLIP-400MViT-B32 × 22420.65 × 1 × 3105✓70.992.1
MoTED*CLIP-400MViT-B8 × 22420.18 × 1 × 3112✓70.191.8
MoTED*CLIP-400MViT-B16 × 22420.34 × 1 × 3112✓71.292.4
MoTED*CLIP-400MViT-B32 × 22420.68 × 1 × 3112✓71.992.7
UnifromerV2 [32]CLIP-400MViT-L32 × 22421.73 × 1 × 3574X73.094.5
TAdaFormer [22]CLIP-400MViT-L32 × 22421.70 × 2 × 3364X73.6-
EVL* [40]CLIP-400MViT-L32 × 22423.21 × 1 × 3654✓66.7-
EVL* [40]CLIP-400MViT-L32 × 33628.08 × 1 × 3654✓68.0-
ST-Adapter* [52]CLIP-400MViT-L32 × 22422.75 × 1 × 3347✓72.393.9
DiST*CLIP-400MViT-L32 × 22422.83 × 1 × 3336✓73.193.2
MoTED*CLIP-400MViT-L8 × 22420.78 × 1 × 3346✓71.592.6
MoTED*CLIP-400MViT-L16 × 22421.49 × 1 × 3346✓73.093.4
MoTED*CLIP-400MViT-L32 × 22422.89 × 1 × 3346✓73.893.8
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.441, + 0.893, + 0.471 + ], + "angle": 0, + "content": "Table 4. Comparison with the state-of-the-art methods on Something-Something-V2. \"Cr.\" and \"Cl.\" are the abbreviation for \"spatial crops\" and \"temporal clips\". \"Frozen\" indicates freezing the CLIP pre-trained parameters." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.496, + 0.473, + 0.739 + ], + "angle": 0, + "content": "Motion Modeling Visualization. In Fig. 4, we also perform the analysis of vision features generated by image and motion encoder to investigate the learned patterns from language supervision. Based on the reasoning tool [8], we depict the attention maps of class token from the final transformer block of the image/motion stream encoder w.r.t. the text encoder. It is observed that, for the case1 in Fig. 4, the CLIP image encoder attends to both motion-relevant foreground and motion-irrelevant background with a major focus on \"surface\". In contrast, features extracted from the motion modules concentrate on the motion-relevant regions of the moving object, and emphasize the motion-related words \"sliding\". This phenomenon reveals that the motion features can complement the static spatial semantics of objects in the image features via modeling object motions. More cases can be accessed in the Appendix." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.754, + 0.345, + 0.772 + ], + "angle": 0, + "content": "4.3. Fully-supervised Experiments" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.78, + 0.471, + 0.903 + ], + "angle": 0, + "content": "In the supervised setting, the results on SSv2 and K400 are presented in Tab. 4 and Tab. 6, respectively. Compared with AVL, our proposed MoTED introduces a similar temporal module for CLIP visual encoder, but our method has a significant improvement compared to AVL, with an improvement of \\(9.5\\% / 2.0\\%\\) and \\(5.8\\% / 1.5\\%\\) on ViT-B and ViT-L respectively on SSv2/K400. The performance gains greatly demonstrate the rationality and effectiveness of using lan" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.496, + 0.893, + 0.558 + ], + "angle": 0, + "content": "guage supervision. Interestingly, the performance gains are particularly pronounced on SSv2. This is because SSv2 is a fine-grained action classification task requiring stronger action discrimination and perception of dynamic information." + }, + { + "type": "table", + "bbox": [ + 0.503, + 0.571, + 0.89, + 0.638 + ], + "angle": 0, + "content": "
Related workVisual disentangleTextual disentangleK400 (Acc / Δ)SSv2 (Acc / Δ)
Previous SOTAXX84.2 / -69.5 / -
DiST [58]✓X85.0 / +0.870.9 / +1.4
MoTED (Ours)✓✓86.2 / +1.271.9 / +1.0
" + }, + { + "type": "table_caption", + "bbox": [ + 0.5, + 0.643, + 0.892, + 0.671 + ], + "angle": 0, + "content": "Table 5. Comparison of our method with previous state-of-the-art (SOTA) and the latest related work DiST." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.684, + 0.895, + 0.837 + ], + "angle": 0, + "content": "Disentangling spatio-temporal learning is an effective approach to endow the model with temporal capability. As shown in Tab. 5, DiST focuses on temporal disentangling in the visual side with gains of \\(+0.8\\% / +1.4\\%\\) on K400/SSv2. Our study highlights the effectiveness of temporal disentangling in the textual side with detailed descriptions of motions, with further gains of \\(+1.2\\% / +1.0\\%\\). This result reveals that disentangling textual encoder is equally effective w.r.t. disentangling visual encoder for vision-language transfer learning." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.847, + 0.715, + 0.864 + ], + "angle": 0, + "content": "4.4. Zero-shot Experiments" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.871, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Zero-shot generalization is an attractive characteristic of CLIP-extended models, making them more practical in real" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "18510" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.095, + 0.088, + 0.878, + 0.431 + ], + "angle": 0, + "content": "
MethodPre-trainArchitectureInput SizeTFLOPs×Cr.×Cl.Param (M)FrozenTop-1Top-5
SlowFast [17]-R101+NL16 × 22420.4 × 3 × 1060X79.893.9
TimeSformer [2]ImageNet-21KViT-L96 × 22428.4 × 3 × 1430X80.794.7
MViT [16]-MViT-B64 × 22420.5 × 1 × 537X81.295.1
ViViT FE [1]ImageNet-21KViT-L128 × 22424.0 × 3 × 1N/AX81.793.8
Video Swin [42]ImageNet-21KSwin-L32 × 22420.6 × 3 × 4197X83.195.9
TAdaConvNeXtV2 [22]ImageNet-21KConvNeXt-B32 × 22420.3 × 3 × 4146X83.7-
X-CLIP [49]CLIP-400MViT-B16 × 22420.28 × 3 × 4128X84.796.8
ST-Adapter* [52]CLIP-400MViT-B32 × 22420.61 × 1 × 393✓82.796.2
EVL* [40]CLIP-400MViT-B32 × 22420.59 × 1 × 3115✓84.2-
DiST*CLIP-400MViT-B32 × 22420.65 × 1 × 3112✓85.097.0
MoTED*CLIP-400MViT-B8 × 22420.18 × 1 × 3116✓85.197.0
MoTED*CLIP-400MViT-B16 × 22420.34 × 1 × 3116✓85.497.2
MoTED*CLIP-400MViT-B32 × 22420.68 × 1 × 3116✓86.297.5
UnifromerV2 [32]CLIP-400M+K710ViT-L32 × 22422.66 × 2 × 3354X89.398.2
TAdaFormer [22]CLIP-400M+K710ViT-L32 × 22421.41 × 4 × 3364X89.5-
ST-Adapter* [52]CLIP-400MViT-L32 × 22422.75 × 1 × 3347✓87.297.6
EVL* [40]CLIP-400MViT-L32 × 22422.70 × 1 × 3363✓87.3-
DiST*CLIP-400MViT-L32 × 22422.83 × 1 × 3343✓88.097.9
MoTED*CLIP-400MViT-L8 × 22420.78 × 1 × 3349✓87.497.8
MoTED*CLIP-400MViT-L16 × 22421.49 × 1 × 3349✓88.098.0
MoTED*CLIP-400MViT-L32 × 22422.89 × 1 × 3349✓88.898.2
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.441, + 0.892, + 0.471 + ], + "angle": 0, + "content": "Table 6. Comparison with state-of-the-arts on Kinetics-400. \"Cr.\" and \"Cl.\" are the abbreviation for \"spatial crops\" and \"temporal clips\". \"Frozen\" indicates freezing the CLIP pre-trained parameters." + }, + { + "type": "table", + "bbox": [ + 0.08, + 0.494, + 0.469, + 0.585 + ], + "angle": 0, + "content": "
MethodModelHMDB51UCF101K600
ActionCLIP [75]B/1640.8±5.458.3±3.4-
X-CLIP [49]B/1644.6±5.272.0±2.365.2±0.4
Vita-CLIP [78]B/1648.6±0.675.0±0.667.4±0.5
DiST*B/1655.4±1.272.3±0.6-
MoTED*B/1658.2±1.178.3±0.669.9 ± 0.5
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.595, + 0.47, + 0.639 + ], + "angle": 0, + "content": "Table 7. Comparison of zero-shot accuracy with the state-of-the-art CLIP-based methods on three datasets (e.g., HMDB51 [24], UCF101 [65], and K600 [12]). \"\": frozen backbone." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.659, + 0.473, + 0.903 + ], + "angle": 0, + "content": "world applications. For zero-shot settings, we evaluate MoTED on three widely used benchmarks. Following prior work [49], we train the networks on the K400 training set, then conduct the zero-shot evaluation on three unseen datasets (i.e., UCF101, HMDB51 and K600), as shown in Tab. 7. All models share the same architecture of \"ViT-B\", with 32 frames during inference. Compared with other vison-language methods, MoTED achieves the better zero-shot performances with a significant margin on HMDB51 \\((+3.8\\%)\\), UCF101 \\((+6.0\\%)\\) and K600 \\((+2.5\\%)\\). Different from X-CLIP [49], DiST [58], Vita-CLIP [78] that learns motion representation with the supervision of action category names solely, the proposed MoTED makes full use of conceptual and discriminative descriptions and learn general motion representations with the aid of language supervision. In addition, our method also has a relatively small" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.496, + 0.892, + 0.543 + ], + "angle": 0, + "content": "variance, only about \\(1\\%\\). We assume this is due to the benefits brought by the rich content of the text, as detailed descriptions can promote stable dynamic feature learning." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.559, + 0.62, + 0.575 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.585, + 0.892, + 0.813 + ], + "angle": 0, + "content": "In this study, we aim to overcome the limitations of existing methods in extending large image-text pre-trained models for video understanding. The proposed MoTED framework introduces Motion-Enhanced Descriptions, which are applied to facilitate the extraction of unique temporal features in videos. By generating motion-related descriptions and incorporating perception modules, MoTED aligns visual and textual motion features using a contrastive loss. Experimental results on five benchmarks demonstrate that MoTED provides a strong basis for enhancing CLIP with robust temporal modeling. In future works, we hope to further dedicate to exploring the potential of language supervision and combining it with more powerful dynamic information perception modules to achieve higher performance in video recognition and make it truly practical." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.829, + 0.681, + 0.846 + ], + "angle": 0, + "content": "6. Acknowledgement" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.856, + 0.892, + 0.901 + ], + "angle": 0, + "content": "This work was supported in part by NSFC No. 62222117, the Fundamental Research Funds for the Central Universities under contract WK3490000005, and KY2100000117." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "18511" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.169 + ], + "angle": 0, + "content": "[1] Anurag Arnab, Mostafa Dehghani, Georg Heigold, Chen Sun, Mario Lučić, and Cordelia Schmid. Vivit: A video vision transformer. In ICCV, pages 6836-6846, 2021. 3, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.172, + 0.47, + 0.213 + ], + "angle": 0, + "content": "[2] Gedas Bertasius, Heng Wang, and Lorenzo Torresani. Is space-time attention all you need for video understanding? In ICML, page 4, 2021. 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.215, + 0.47, + 0.257 + ], + "angle": 0, + "content": "[3] Daniel Bolya, Cheng-Yang Fu, Xiaoliang Dai, Peizhao Zhang, Christoph Feichtenhofer, and Judy Hoffman. Token merging: Your vit but faster. In ICLR, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.258, + 0.47, + 0.395 + ], + "angle": 0, + "content": "[4] Tom B. Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-Voss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel M. Ziegler, Jeffrey Wu, Clemens Winter, Christopher Hesse, Mark Chen, Eric Sigler, Mateusz Litwin, Scott Gray, Benjamin Chess, Jack Clark, Christopher Berner, Sam McCandlish, Alec Radford, Ilya Sutskever, and Dario Amodei. Language models are few-shot learners. In NeurIPS, 2020. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.397, + 0.47, + 0.451 + ], + "angle": 0, + "content": "[5] Shyamal Buch, Cristóbal Eyzaguirre, Adrien Gaidon, Jiajun Wu, Li Fei-Fei, and Juan Carlos Niebles. Revisiting the \"video\" in video-language understanding. In CVPR, pages 2907–2917, 2022. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.453, + 0.469, + 0.507 + ], + "angle": 0, + "content": "[6] Adrian Bulat, Juan Manuel Perez Rua, Swathikiran Sudhakaran, Brais Martinez, and Georgios Tzimiropoulos. Space-time mixing attention for video transformer. NeurIPS, 34:19594-19607, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.509, + 0.469, + 0.55 + ], + "angle": 0, + "content": "[7] Joao Carreira and Andrew Zisserman. Quo vadis, action recognition? a new model and the kinetics dataset. In CVPR, pages 6299-6308, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.552, + 0.469, + 0.605 + ], + "angle": 0, + "content": "[8] Hila Chefer, Shir Gur, and Lior Wolf. Generic attention-model explainability for interpreting bi-modal and encoder-decoder transformers. In ICCV, pages 387-396, 2021. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.608, + 0.469, + 0.718 + ], + "angle": 0, + "content": "[9] Xi Chen, Xiao Wang, Soravit Changpinyo, A. J. Piergiovanni, Piotr Padlewski, Daniel Salz, Sebastian Goodman, Adam Grycner, Basil Mustafa, Lucas Beyer, Alexander Kolesnikov, Joan Puigcerver, Nan Ding, Keran Rong, Hassan Akbari, Gaurav Mishra, Linting Xue, Ashish V. Thapliyal, James Bradbury, and Weicheng Kuo. Pali: A jointly scaled multilingual language-image model. In ICLR, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.72, + 0.469, + 0.774 + ], + "angle": 0, + "content": "[10] Xing Cheng, Hezheng Lin, Xiangyu Wu, Fan Yang, and Dong Shen. Improving video-text retrieval by multi-stream corpus alignment and dual softmax loss. arXiv preprint arXiv:2109.04290, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.776, + 0.469, + 0.902 + ], + "angle": 0, + "content": "[11] Aakanksha Chowdhery, Sharan Narang, Jacob Devlin, Maarten Bosma, Gaurav Mishra, Adam Roberts, Paul Barham, Hyung Won Chung, Charles Sutton, Sebastian Gehrmann, Parker Schuh, Kensen Shi, Sasha Tsvyashchenko, Joshua Maynez, Abhishek Rao, Parker Barnes, Yi Tay, Noam Shazeer, Vinodkumar Prabhakaran, Emily Reif, Nan Du, Ben Hutchinson, Reiner Pope, James Bradbury, Jacob Austin, Michael Isard, Guy Gur-Ari, Pengcheng Yin, Toju Duke, Anselm Levskaya, Sanjay" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.47, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.272 + ], + "angle": 0, + "content": "Ghemawat, Sunipa Dev, Henryk Michalewski, Xavier Garcia, Vedant Misra, Kevin Robinson, Liam Fedus, Denny Zhou, Daphne Ippolito, David Luan, Hyeontaek Lim, Barret Zoph, Alexander Spiridonov, Ryan Sepassi, David Dohan, Shivani Agrawal, Mark Omernick, Andrew M. Dai, Thanumalayan Sankaranarayana Pillai, Marie Pellat, Aitor Lewkowycz, Erica Moreira, Rewon Child, Oleksandr Polozov, Katherine Lee, Zongwei Zhou, Xuezhi Wang, Brennan Saeta, Mark Diaz, Orhan Firat, Michele Catasta, Jason Wei, Kathy Meier-Hellstern, Douglas Eck, Jeff Dean, Slav Petrov, and Noah Fiedel. Palm: Scaling language modeling with pathways. J. Mach. Learn. Res., 24:240:1-240:113, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.274, + 0.892, + 0.343 + ], + "angle": 0, + "content": "[12] Dima Damen, Hazel Doughty, Giovanni Maria Farinella, Antonino Furnari, Evangelos Kazakos, Jian Ma, Davide Moltisanti, Jonathan Munro, Toby Perrett, Will Price, et al. Rescaling egocentric vision. arXiv preprint arXiv:2006.13256, 2020. 2, 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.345, + 0.892, + 0.401 + ], + "angle": 0, + "content": "[13] Ali Diba, Vivek Sharma, Reza Safdari, Dariush Lotfi, Saquib Sarfraz, Rainer Stiefelhagen, and Luc Van Gool. Vi2cIr: Video and image for visual contrastive learning of representation. In ICCV, pages 1502-1512, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.402, + 0.892, + 0.457 + ], + "angle": 0, + "content": "[14] Sixun Dong, Huazhang Hu, Dongze Lian, Weixin Luo, Yicheng Qian, and Shenghua Gao. Weakly supervised video representation learning with unaligned text for sequential videos. In CVPR, pages 2437-2447. IEEE, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.459, + 0.892, + 0.528 + ], + "angle": 0, + "content": "[15] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.53, + 0.892, + 0.585 + ], + "angle": 0, + "content": "[16] Haoqi Fan, Bo Xiong, Karttikeya Mangalam, Yanghao Li, Zhicheng Yan, Jitendra Malik, and Christoph Feichtenhofer. Multiscale vision transformers. In ICCV, pages 6824-6835, 2021. 3, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.587, + 0.892, + 0.629 + ], + "angle": 0, + "content": "[17] Christoph Feichtenhofer, Haoqi Fan, Jitendra Malik, and Kaiming He. Slowfast networks for video recognition. In ICCV, pages 6202-6211, 2019. 3, 5, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.631, + 0.892, + 0.685 + ], + "angle": 0, + "content": "[18] Christoph Feichtenhofer, Haoqi Fan, Bo Xiong, Ross Girshick, and Kaiming He. A large-scale study on unsupervised spatiotemporal representation learning. In CVPR, pages 3299-3309, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.688, + 0.892, + 0.728 + ], + "angle": 0, + "content": "[19] Christoph Feichtenhofer, Haoqi Fan, Yanghao Li, and Kaiming He. Masked autoencoders as spatiotemporal learners. NeurIPS, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.731, + 0.892, + 0.814 + ], + "angle": 0, + "content": "[20] Raghav Goyal, Samira Ebrahimi Kahou, Vincent Michalski, Joanna Materzynska, Susanne Westphal, Heuna Kim, Valentin Haenel, Ingo Fruend, Peter Yianilos, Moritz Mueller-Freitag, et al. The\" something something\" video database for learning and evaluating visual common sense. In ICCV, pages 5842-5850, 2017. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.816, + 0.892, + 0.87 + ], + "angle": 0, + "content": "[21] Ziyuan Huang, Shiwei Zhang, Liang Pan, Zhiwu Qing, Mingqian Tang, Ziwei Liu, and Marcelo H Ang Jr. Tada! temporally-adaptive convolutions for video understanding. In ICLR, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.873, + 0.892, + 0.902 + ], + "angle": 0, + "content": "[22] Ziyuan Huang, Shiwei Zhang, Liang Pan, Zhiwu Qing, Yingya Zhang, Ziwei Liu, and Marcelo H Ang Jr." + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.892, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "18512" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.092, + 0.468, + 0.12 + ], + "angle": 0, + "content": "Temporally-adaptive models for efficient video understanding. arXiv preprint arXiv:2308.05787, 2023. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.122, + 0.469, + 0.162 + ], + "angle": 0, + "content": "[23] Simon Jenni and Hailin Jin. Time-equivariant contrastive video representation learning. In ICCV, pages 9970-9980, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.165, + 0.469, + 0.207 + ], + "angle": 0, + "content": "[24] H Jhuang, H Garrote, E Poggio, T Serre, and T Hmdb. A large video database for human motion recognition. In ICCV, page 6, 2011. 2, 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.208, + 0.469, + 0.276 + ], + "angle": 0, + "content": "[25] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In ICML, pages 4904-4916. PMLR, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.279, + 0.469, + 0.332 + ], + "angle": 0, + "content": "[26] Chen Ju, Tengda Han, Kunhao Zheng, Ya Zhang, and Weidi Xie. Prompting visual-language models for efficient video understanding. In ECCV, pages 105-124. Springer, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.335, + 0.469, + 0.377 + ], + "angle": 0, + "content": "[27] Chen Ju, Tengda Han, Kunhao Zheng, Ya Zhang, and Weidi Xie. Prompting visual-language models for efficient video understanding. In ECCV, pages 105-124. Springer, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.379, + 0.469, + 0.447 + ], + "angle": 0, + "content": "[28] Will Kay, Joao Carreira, Karen Simonyan, Brian Zhang, Chloe Hillier, Sudheendra Vijayanarasimhan, Fabio Viola, Tim Green, Trevor Back, Paul Natsev, et al. The kinetics human action video dataset. arXiv preprint arXiv:1705.06950, 2017. 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.449, + 0.469, + 0.572 + ], + "angle": 0, + "content": "[29] Jan Kocon, Igor Cichecki, Oliwier Kaszyca, Mateusz Kochanek, Dominika Szydlo, Joanna Baran, Julita Bielaniewicz, Marcin Gruza, Arkadiusz Janz, Kamil Kanclerz, Anna Kocon, Bartlomiej Koptyra, Wiktoria Mieleszczenko-Kowszewicz, Piotr Milkowski, Marcin Oleksy, Maciej Piasecki, Lukasz Radlinski, Konrad Wojtasik, Stanislaw Wozniak, and Przemyslaw Kazienko. Chatgpt: Jack of all trades, master of none. Inf. Fusion, 99: 101861, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.575, + 0.469, + 0.616 + ], + "angle": 0, + "content": "[30] Jie Lei, Tamara L. Berg, and Mohit Bansal. Revealing single frame bias for video-and-language learning. In ACL, pages 487-507, 2023. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.618, + 0.469, + 0.687 + ], + "angle": 0, + "content": "[31] Chunyuan Li, Haotian Liu, Liunian Harold Li, Pengchuan Zhang, Jyoti Aneja, Jianwei Yang, Ping Jin, Houdong Hu, Zicheng Liu, Yong Jae Lee, and Jianfeng Gao. ELEVATER: A benchmark and toolkit for evaluating language-augmented visual models. In NeurIPS, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.689, + 0.469, + 0.744 + ], + "angle": 0, + "content": "[32] Kunchang Li, Yali Wang, Yinan He, Yizhuo Li, Yi Wang, Limin Wang, and Yu Qiao. Uniformerv2: Spatiotemporal learning by arming image vits with video uniformer. arXiv preprint arXiv:2211.09552, 2022. 3, 5, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.746, + 0.469, + 0.8 + ], + "angle": 0, + "content": "[33] Kunchang Li, Yali Wang, Junhao Zhang, Peng Gao, Guanglu Song, Yu Liu, Hongsheng Li, and Yu Qiao. Uniformer: Unifying convolution and self-attention for visual recognition. TPAMI, 2023. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.802, + 0.469, + 0.857 + ], + "angle": 0, + "content": "[34] Linjie Li, Yen-Chun Chen, Yu Cheng, Zhe Gan, Licheng Yu, and Jingjing Liu. HERO: hierarchical encoder for video+language omni-representation pre-training. In EMNLP, pages 2046-2065, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.859, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[35] Liunian Harold Li, Pengchuan Zhang, Haotian Zhang, Jianwei Yang, Chunyuan Li, Yiwu Zhong, Lijuan Wang, Lu Yuan, Lei Zhang, Jenq-Neng Hwang, Kai-Wei Chang, and" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.469, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "Jianfeng Gao. Grounded language-image pre-training. In CVPR, pages 10955-10965, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.122, + 0.892, + 0.164 + ], + "angle": 0, + "content": "[36] Yanghao Li, Haoqi Fan, Ronghang Hu, Christoph Feichtenhofer, and Kaiming He. Scaling language-image pre-training via masking. arXiv preprint arXiv:2212.00794, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.166, + 0.892, + 0.222 + ], + "angle": 0, + "content": "[37] Yangguang Li, Feng Liang, Lichen Zhao, Yufeng Cui, Wanli Ouyang, Jing Shao, Fengwei Yu, and Junjie Yan. Supervision exists everywhere: A data efficient contrastive language-image pre-training paradigm. In ICLR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.224, + 0.892, + 0.291 + ], + "angle": 0, + "content": "[38] Yanghao Li, Chao-Yuan Wu, Haoqi Fan, Karttikeya Mangalam, Bo Xiong, Jitendra Malik, and Christoph Feichtenhofer. Mvitv2: Improved multiscale vision transformers for classification and detection. In CVPR, pages 4804-4814, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.295, + 0.892, + 0.35 + ], + "angle": 0, + "content": "[39] Xudong Lin, Fabio Petroni, Gedas Bertasius, Marcus Rohrbach, Shih-Fu Chang, and Lorenzo Torresani. Learning to recognize procedural activities with distant supervision. In CVPR, pages 13843-13853. IEEE, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.352, + 0.892, + 0.408 + ], + "angle": 0, + "content": "[40] Ziyi Lin, Shijie Geng, Renrui Zhang, Peng Gao, Gerard de Melo, Xiaogang Wang, Jifeng Dai, Yu Qiao, and Hongsheng Li. Frozen clip models are efficient video learners. In ECCV, pages 388-404. Springer, 2022. 1, 2, 5, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.41, + 0.892, + 0.437 + ], + "angle": 0, + "content": "[41] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. In NeurIPS, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.439, + 0.892, + 0.481 + ], + "angle": 0, + "content": "[42] Ze Liu, Jia Ning, Yue Cao, Yixuan Wei, Zheng Zhang, Stephen Lin, and Han Hu. Video swim transformer. In CVPR, pages 3202-3211, 2022. 3, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.483, + 0.892, + 0.523 + ], + "angle": 0, + "content": "[43] Timo Lüddecke and Alexander S. Ecker. Image segmentation using text and image prompts. In CVPR, pages 7076-7086, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.526, + 0.892, + 0.554 + ], + "angle": 0, + "content": "[44] Sachit Menon and Carl Vondrick. Visual classification via description from large language models. In ICLR, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.556, + 0.892, + 0.624 + ], + "angle": 0, + "content": "[45] Antoine Miech, Dimitri Zhukov, Jean-Baptiste Alayrac, Makarand Tapaswi, Ivan Laptev, and Josef Sivic. Howto100m: Learning a text-video embedding by watching hundred million narrated video clips. In ICCV, pages 2630-2640, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.627, + 0.892, + 0.683 + ], + "angle": 0, + "content": "[46] Antoine Miech, Jean-Baptiste Alayrac, Lucas Smaira, Ivan Laptev, Josef Sivic, and Andrew Zisserman. End-to-end learning of visual representations from uncurated instructional videos. In CVPR, pages 9879-9889, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.685, + 0.892, + 0.739 + ], + "angle": 0, + "content": "[47] Liliane Momeni, Mathilde Caron, Arsha Nagrani, Andrew Zisserman, and Cordelia Schmid. Verbs in action: Improving verb understanding in video-language models. ICCV, 2023. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.742, + 0.892, + 0.783 + ], + "angle": 0, + "content": "[48] Daniel Neimark, Omri Bar, Maya Zohar, and Dotan Asselmann. Video transformer network. In ICCV, pages 3163-3172, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.786, + 0.892, + 0.855 + ], + "angle": 0, + "content": "[49] Bolin Ni, Houwen Peng, Minghao Chen, Songyang Zhang, Gaofeng Meng, Jianlong Fu, Shiming Xiang, and Haibin Ling. Expanding language-image pretrained models for general video recognition. In ECCV, pages 1-18. Springer, 2022. 1, 2, 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.858, + 0.805, + 0.871 + ], + "angle": 0, + "content": "[50] OpenAI. Gpt-4 technical report, 2023. 1, 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.873, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[51] TB OpenAI. Chatgpt: Optimizing language models for dialogue. OpenAI, 2022. 3" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18513" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.134 + ], + "angle": 0, + "content": "[52] Junting Pan, Ziyi Lin, Xiatian Zhu, Jing Shao, and Hongsheng Li. Parameter-efficient image-to-video transfer learning. arXiv e-prints, pages arXiv-2206, 2022. 1, 2, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.136, + 0.47, + 0.19 + ], + "angle": 0, + "content": "[53] Tian Pan, Yibing Song, Tianyu Yang, Wenhao Jiang, and Wei Liu. Videomoco: Contrastive video representation learning with temporally adversarial examples. In CVPR, pages 11205-11214, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.192, + 0.469, + 0.233 + ], + "angle": 0, + "content": "[54] Jungin Park, Jiyoung Lee, and Kwanghoon Sohn. Dual-path adaptation from image to video transformers. CVPR, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.235, + 0.469, + 0.303 + ], + "angle": 0, + "content": "[55] Mandela Patrick, Dylan Campbell, Yuki Asano, Ishan Misra, Florian Metze, Christoph Feichtenhofer, Andrea Vedaldi, and João F Henriques. Keeping your eye on the ball: Trajectory attention in video transformers. NeurIPS, 34:12493-12506, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.305, + 0.469, + 0.347 + ], + "angle": 0, + "content": "[56] Sarah M. Pratt, Rosanne Liu, and Ali Farhadi. What does a platypus look like? generating customized prompts for zero-shot image classification. ICCV, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.349, + 0.469, + 0.403 + ], + "angle": 0, + "content": "[57] Rui Qian, Tianjian Meng, Boqing Gong, Ming-Hsuan Yang, Huisheng Wang, Serge Belongie, and Yin Cui. Spatiotemporal contrastive video representation learning. In CVPR, pages 6964-6974, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.405, + 0.469, + 0.46 + ], + "angle": 0, + "content": "[58] Zhiwu Qing, Shiwei Zhang, Ziyuan Huang, Yingya Zhang, Changxin Gao, Deli Zhao, and Nong Sang. Disentangling spatial and temporal learning for efficient image-to-video transfer learning. ICCV, 2023. 1, 2, 5, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.462, + 0.469, + 0.503 + ], + "angle": 0, + "content": "[59] Zhaofan Qiu, Ting Yao, and Tao Mei. Learning spatiotemporal representation with pseudo-3d residual networks. In ICCV, pages 5533–5541, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.504, + 0.469, + 0.586 + ], + "angle": 0, + "content": "[60] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, pages 8748-8763. PMLR, 2021. 1, 2, 3, 4, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.589, + 0.469, + 0.63 + ], + "angle": 0, + "content": "[61] Kanchana Ranasinghe and Michael S. Ryoo. Language-based action concept spaces improve video self-supervised learning. CoRR, abs/2307.10922, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.632, + 0.469, + 0.685 + ], + "angle": 0, + "content": "[62] Michael Ryoo, AJ Piergiovanni, Anurag Arnab, Mostafa Dehghani, and Anelia Angelova. Tokenlearner: Adaptive space-time tokenization for videos. *NeurIPS*, 34:12786-12797, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.688, + 0.469, + 0.743 + ], + "angle": 0, + "content": "[63] Laura Sevilla-Lara, Shengxin Zha, Zhicheng Yan, Vedanuj Goswami, Matt Feiszli, and Lorenzo Torresani. Only time can tell: Discovering temporal data for temporal modeling. In WACV, pages 535-544, 2021. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.745, + 0.469, + 0.786 + ], + "angle": 0, + "content": "[64] Karen Simonyan and Andrew Zisserman. Two-stream convolutional networks for action recognition in videos. NeurIPS, 27, 2014. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.788, + 0.469, + 0.829 + ], + "angle": 0, + "content": "[65] Khurram Soomro, Amir Roshan Zamir, and Mubarak Shah. Ucf101: A dataset of 101 human actions classes from videos in the wild. arXiv preprint arXiv:1212.0402, 2012. 2, 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.831, + 0.469, + 0.872 + ], + "angle": 0, + "content": "[66] Zhan Tong, Yibing Song, Jue Wang, and Limin Wang. Videomae: Masked autoencoders are data-efficient learners for self-supervised video pre-training. NeurIPS, 2022. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.874, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[67] Du Tran, Lubomir Bourdev, Rob Fergus, Lorenzo Torresani, and Manohar Paluri. Learning spatiotemporal features" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.892, + 0.119 + ], + "angle": 0, + "content": "with 3d convolutional networks. In ICCV, pages 4489-4497, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.123, + 0.892, + 0.178 + ], + "angle": 0, + "content": "[68] Du Tran, Heng Wang, Lorenzo Torresani, Jamie Ray, Yann LeCun, and Manohar Paluri. A closer look at spatiotemporal convolutions for action recognition. In CVPR, pages 6450-6459, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.181, + 0.892, + 0.223 + ], + "angle": 0, + "content": "[69] Du Tran, Heng Wang, Lorenzo Torresani, and Matt Feiszli. Video classification with channel-separated convolutional networks. In ICCV, pages 5552-5561, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.224, + 0.892, + 0.279 + ], + "angle": 0, + "content": "[70] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In NeurIPS, pages 5998-6008, 2017. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.282, + 0.892, + 0.323 + ], + "angle": 0, + "content": "[71] Heng Wang, Du Tran, Lorenzo Torresani, and Matt Feiszli. Video modeling with correlation networks. In CVPR, pages 352-361, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.326, + 0.892, + 0.381 + ], + "angle": 0, + "content": "[72] Junke Wang, Dongdong Chen, Chong Luo, Xiyang Dai, Lu Yuan, Zuxuan Wu, and Yu-Gang Jiang. Chatvideo: A tracklet-centric multimodal and versatile video understanding system. CoRR, abs/2304.14407, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.384, + 0.892, + 0.438 + ], + "angle": 0, + "content": "[73] Limin Wang, Yuanjun Xiong, Zhe Wang, Yu Qiao, Dahua Lin, Xiaou Tang, and Luc Van Gool. Temporal segment networks for action recognition in videos. TPAMI, 41(11): 2740-2755, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.442, + 0.892, + 0.496 + ], + "angle": 0, + "content": "[74] Limin Wang, Bingkun Huang, Zhiyu Zhao, Zhan Tong, Yi-nan He, Yi Wang, Yali Wang, and Yu Qiao. Videomae V2: scaling video masked autoencoders with dual masking. In CVPR, pages 14549-14560, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.499, + 0.892, + 0.54 + ], + "angle": 0, + "content": "[75] Mengmeng Wang, Jiazheng Xing, and Yong Liu. Actionclip: A new paradigm for video action recognition. arXiv preprint arXiv:2109.08472, 2021. 1, 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.543, + 0.892, + 0.612 + ], + "angle": 0, + "content": "[76] Peng Wang, An Yang, Rui Men, Junyang Lin, Shuai Bai, Zhikang Li, Jianxin Ma, Chang Zhou, Jingren Zhou, and Hongxia Yang. OFA: unifying architectures, tasks, and modalities through a simple sequence-to-sequence learning framework. In ICML, pages 23318-23340, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.614, + 0.892, + 0.669 + ], + "angle": 0, + "content": "[77] Rui Wang, Dongdong Chen, Zuxuan Wu, Yinpeng Chen, Xiyang Dai, Mengchen Liu, Yu-Gang Jiang, Luowei Zhou, and Lu Yuan. Bevt: Bert pretraining of video transformers. In CVPR, pages 14733-14743, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.672, + 0.892, + 0.727 + ], + "angle": 0, + "content": "[78] Syed Talal Wasim, Muzammal Naseer, Salman H. Khan, Fahad Shahbaz Khan, and Mubarak Shah. Vita-clip: Video and text adaptive CLIP via multimodal prompting. CVPR, 2023. 1, 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.73, + 0.892, + 0.784 + ], + "angle": 0, + "content": "[79] Chen Wei, Haoqi Fan, Saining Xie, Chao-Yuan Wu, Alan Yuille, and Christoph Feichtenhofer. Masked feature prediction for self-supervised visual pre-training. In CVPR, pages 14668-14678, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.788, + 0.892, + 0.83 + ], + "angle": 0, + "content": "[80] Wenhao Wu, Zhun Sun, and Wanli Ouyang. Revisiting classifier: Transferring vision-language models for video recognition. In AAAI, pages 2847-2855, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.832, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[81] Hu Xu, Gargi Ghosh, Po-Yao Huang, Dmytro Okhonko, Armen Aghajanyan, Florian Metze, Luke Zettlemoyer, and Christoph Feichtenhofer. Videoclip: Contrastive pre-training for zero-shot video-text understanding. In EMNLP, pages 6787-6800, 2021. 2" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "18514" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.147 + ], + "angle": 0, + "content": "[82] Jiarui Xu, Shalini De Mello, Sifei Liu, Wonmin Byeon, Thomas M. Breuel, Jan Kautz, and Xiaolong Wang. Groupvit: Semantic segmentation emerges from text supervision. In CVPR, pages 18113-18123, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.149, + 0.468, + 0.204 + ], + "angle": 0, + "content": "[83] An Yan, Yu Wang, Yiwu Zhong, Chengyu Dong, Zexue He, Yujie Lu, William Wang, Jingbo Shang, and Julian J. McAuley. Learning concise and descriptive attributes for visual recognition. ICCV, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.207, + 0.468, + 0.26 + ], + "angle": 0, + "content": "[84] Shen Yan, Xuehan Xiong, Anurag Arnab, Zhichao Lu, Mi Zhang, Chen Sun, and Cordelia Schmid. Multiview transformers for video recognition. In CVPR, pages 3333-3343, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.263, + 0.468, + 0.33 + ], + "angle": 0, + "content": "[85] Yue Yang, Artemis Panagopoulou, Shenghao Zhou, Daniel Jin, Chris Callison-Burch, and Mark Yatskar. Language in a bottle: Language model guided concept bottlenecks for interpretable image classification. In CVPR, pages 19187-19197, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.333, + 0.468, + 0.401 + ], + "angle": 0, + "content": "[86] Zhengyuan Yang, Linjie Li, Jianfeng Wang, Kevin Lin, Ehsan Azarnasab, Faisal Ahmed, Zicheng Liu, Ce Liu, Michael Zeng, and Lijuan Wang. MM-REACT: prompting chatgpt for multimodal reasoning and action. CoRR, abs/2303.11381, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.404, + 0.468, + 0.458 + ], + "angle": 0, + "content": "[87] Jiahui Yu, Zirui Wang, Vijay Vasudevan, Legg Yeung, Mojtaba Seyedhosseini, and Yonghui Wu. Coca: Contrastive captioners are image-text foundation models. arXiv preprint arXiv:2205.01917, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.461, + 0.468, + 0.528 + ], + "angle": 0, + "content": "[88] Lu Yuan, Dongdong Chen, Yi-Ling Chen, Noel Codella, Xiyang Dai, Jianfeng Gao, Houdong Hu, Xuedong Huang, Boxin Li, Chunyuan Li, et al. Florence: A new foundation model for computer vision. arXiv preprint arXiv:2111.11432, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.531, + 0.468, + 0.572 + ], + "angle": 0, + "content": "[89] Yue Zhao, Ishan Misra, Philipp Krahenbuhl, and Rohit Girdhar. Learning video representations from large language models. In CVPR, pages 6586-6597. IEEE, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.574, + 0.468, + 0.641 + ], + "angle": 0, + "content": "[90] Yiwu Zhong, Jianwei Yang, Pengchuan Zhang, Chunyuan Li, Noel Codella, Liunian Harold Li, Luowei Zhou, Xiyang Dai, Lu Yuan, Yin Li, and Jianfeng Gao. Regionclip: Region-based language-image pretraining. In CVPR, pages 16772-16782, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.644, + 0.468, + 0.698 + ], + "angle": 0, + "content": "[91] Deyao Zhu, Jun Chen, Xiaoqian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models. CoRR, abs/2304.10592, 2023. 3" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.698 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.956 + ], + "angle": 0, + "content": "18515" + } + ] +] \ No newline at end of file diff --git a/2024/Enhanced Motion-Text Alignment for Image-to-Video Transfer Learning/1b431374-7f10-4d10-9e33-68850035e408_origin.pdf b/2024/Enhanced Motion-Text Alignment for Image-to-Video Transfer Learning/1b431374-7f10-4d10-9e33-68850035e408_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..ee6a8f324e39167bdd9556ef29310579fc350819 --- /dev/null +++ b/2024/Enhanced Motion-Text Alignment for Image-to-Video Transfer Learning/1b431374-7f10-4d10-9e33-68850035e408_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:fb93e4d37ebfeffd372f3d270adf54303c0afc40e5af7e6ed9f4cc44bff79ec1 +size 1749560 diff --git a/2024/Enhanced Motion-Text Alignment for Image-to-Video Transfer Learning/full.md b/2024/Enhanced Motion-Text Alignment for Image-to-Video Transfer Learning/full.md new file mode 100644 index 0000000000000000000000000000000000000000..98377cfa37b2dacfd85127fd9214280c51aa870b --- /dev/null +++ b/2024/Enhanced Motion-Text Alignment for Image-to-Video Transfer Learning/full.md @@ -0,0 +1,330 @@ +# Enhanced Motion-Text Alignment for Image-to-Video Transfer Learning + +Wei Zhang $^{1,2*}$ , Chaoqun Wan $^{2}$ , Tongliang Liu $^{3}$ , Xinmei Tian $^{1,4\dagger}$ , Xu Shen $^{2\dagger}$ , Jieping Ye $^{2}$ + +1University of Science and Technology of China, 2Alibaba Cloud, 3The University of Sydney Institute of Artificial Intelligence, Hefei Comprehensive National Science Center + +zw0819@mail.ustc.edu.cn, tongliang.liu@sydney.edu.au, xinmei@ustc.edu.cn {qionlong.wcq, shenxu.sx, jieping.yjp}@alibaba-inc.com + +# Abstract + +Extending large image-text pre-trained models (e.g., CLIP) for video understanding has made significant advancements. To enable the capability of CLIP to perceive dynamic information in videos, existing works are dedicated to equipping the visual encoder with various temporal modules. However, these methods exhibit "asymmetry" between the visual and textual sides, with neither temporal descriptions in input texts nor temporal modules in text encoder. This limitation hinders the potential of language supervision emphasized in CLIP, and restricts the learning of temporal features, as the text encoder has demonstrated limited proficiency in motion understanding. To address this issue, we propose leveraging "MoTion-Enhanced Descriptions" (MoTED) to facilitate the extraction of distinctive temporal features in videos. Specifically, we first generate discriminative motion-related descriptions via querying GPT-4 to compare easy-confusing action categories. Then, we incorporate both the visual and textual encoders with additional perception modules to process the video frames and generated descriptions, respectively. Finally, we adopt a contrastive loss to align the visual and textual motion features. Extensive experiments on five benchmarks show that MoTED surpasses state-of-the-art methods with convincing gaps, laying a solid foundation for empowering CLIP with strong temporal modeling. + +# 1. Introduction + +Recent years have witnessed remarkable achievements in contrastive language-image pre-training models [9, 25, 36, 60, 87, 88], with CLIP [60] emerging as the front-runner. Through language supervision with a vast collection of 400 million image-text pairs, CLIP has achieved exceptional image comprehension and unprecedented zero-shot general- + +![](images/fdb90797dde85bb2abcae4d1dd21b050d8b4d4fa3577438c3b73d612ad573c49.jpg) +(c) Motion-Enhanced Descriptions +Figure 1. (a) Previous methods [27, 49, 75, 78] align the visual embeddings with the textual embeddings derived from category names. (b) Our work symmetrically aligns the spatial/temporal visual embeddings with class descriptions and motion descriptions correspondingly. (c) The descriptions generated by large language modes (e.g., GPT-4 [50]) enhance the conceptual definitions and discriminative details of motions. + +ization. This breakthrough has opened up new possibilities for leveraging the power of large-scale pre-trained models to comprehend videos. It has also introduced a new paradigm [26, 40, 49, 52, 58, 75, 80] that endows image-based CLIP to effectively perceive dynamic information for video recognition. + +To equip CLIP with motion perception, existing works propose to incorporate various temporal modules into the visual encoder. These temporal modules are additional tunable parameters, including designed efficient units between pre-trained transformer blocks [49, 75], and parallel structures to disentangle the temporal modeling out of the spatial modeling [40, 58]. However, while considerable effort has been put into capturing temporal visual features from videos, these methods have paid little attention to the in + +put text and text encoder, leading to an imbalance between the visual and textual alignment. When transferring to the classification task of video recognition (Fig. 1(a)), the only available text is the "category names" of the actions, e.g., "clapping" and "slapping". These coarse-grained word-level descriptors lack clear descriptions and explanations, making it difficult to distinguish between them. For instance, "slapping" refers to a quick hand movement striking against something, while "clapping" refers to the act of striking both hands together rapidly. The scarcity of textual motion information contradicts the language supervision principle of CLIP. Furthermore, it's observed that the CLIP text encoder delivers a strong bias towards spatial concepts (e.g., nouns) [5, 30, 47, 63], with a weak understanding of temporal cues (e.g., verbs). These issues significantly limit the effectiveness of extending CLIP with motion modeling for video understanding. + +To overcome the limitations of language supervision on temporal modeling, we propose a novel approach that symmetrically aligns the spatial/temporal visual embeddings with class descriptions ("slapping") and motion descriptions correspondingly ("a quick hand movement striking against something"). We introduce the MoTion-Enhanced Descriptions (MoTED) as the language supervision for the visual temporal modules. In the implementation, we encounter two main challenges: i) how to generate motion-related descriptions? ii) how to effectively utilize descriptions as supervision? To tackle the first challenge, we aim for clear, detailed, and distinctive descriptions in the text. With the recent advancements in large language models (LLMs) [4, 11, 50], it has become possible to automatically generate conceptual descriptions for corresponding actions. As for the second challenge, we disentangle the spatial and temporal learning by constructing a temporal encoder, in parallel to the CLIP image encoder, to extract the temporal features aligned with motion-enhanced descriptions. + +To be specific, we employ GPT-4 [50] to generate descriptions by posing a query: “Q: what is the motion concept of ? A:” (as shown in Fig. 1(c)). However, our empirical findings indicate that only the concept descriptions offer marginal assistance in motion perception, particularly on datasets such as Something-Something-V2 (SSv2) [20]) that necessitate the differentiation of similar classes using intricate motion cues. To mitigate this issue, we propose to incorporate related top-k actions as context for the LLMs to generate more discriminative descriptions. These generated texts are then passed through the CLIP text encoder to extract textual motion representations, where an adapter is utilized to eliminate the biases in the original model. Correspondingly, the temporal module in the visual encoder processes the video to extract dynamic features and aligns them with the generated textual descriptions through contrastive learning. Finally, these two sets of + +features are independently fused with the original CLIP's text and image features using cross-attention, and subsequently aligned. Evaluated on two supervised video recognition benchmarks, i.e., Kinetics-400 [28] and SSv2 [20], as well as three zero-shot benchmarks, i.e., Kinetics-600 [12], HMDB51 [24], UCF101 [65], the proposed MoTED surpasses state-of-the-art methods with convincing gaps, indicating the effectiveness of aligning enhanced motion descriptions with the temporal embedding of input videos. + +We summarize the contributions as follows: + +- We present a new perspective that underscores the significance of textual side on par with visual side. By delving into in-depth distinguishing descriptions of actions, we make the first attempt to reveal the potential of language supervision as emphasized in CLIP. +- We propose MoTED that leverages LLMs to automatically generate action descriptions, and mining distinctive descriptions among similar actions. Then, we use a parallel path for both visual and textual motion modeling. +- We evaluate our approach on supervised as well as generalization tasks. Extensive experiments demonstrate the superiority and good generalization ability of the proposed method. + +# 2. Related Work + +Vision-Language Pre-training. In recent years, Vision-Language Pre-training (VLP) [9, 25, 36, 45, 46, 60, 76, 87, 88] has made remarkable progress. One of the most remarkable and influential works is CLIP [60], which adopts the contrastive language-image pretraining paradigm. Following that, this paradigm has shown impressive zero-shot generalization capabilities on various image-related tasks [31, 35, 37, 43, 82, 90]. However, pre-training a language-video model [34, 81] is prohibitively expensive, as it requires large-scale video-text data and extensive training resources (e.g., thousands of GPU days). Meanwhile, transferring language-image pre-trained models to the video domain [10, 26, 40, 49, 52, 58, 75, 78] has captured significant attention due to its striking performance and training efficiency. For instance, X-CLIP [49] integrates the CLIP image encoder with a cross-frame attention module for temporal modeling. Vita-CLIP [78] utilizes multi-modal prompting techniques to learn video and text-specific context vectors. DiST [58] is the most related work that also disentangles spatial and temporal learning in the visual side. In contrast, our MoTED highlights the effectiveness of temporal disentangling both in the visual and textual sides, supervised by the detailed descriptions of motions. + +Video Recognition. The conventional approaches in video recognition primarily focus on spatio-temporal learning under fully-supervised settings, where all categories are predefined. These approaches have achieved remarkable performance using various architectures, including convolution + +![](images/c777a177803b74fa4a8ecf60d1633d0ccd35e865c08bb75d94524f870af32ca1.jpg) +Figure 2. The overall framework of MoTED. Building upon the dual image and text encoders initialized by CLIP [60], we extend the capability to perceive motion information symmetrically in both sides. In the text side, the motion-related conceptual and discriminative descriptions are generated via querying LLMs (e.g., GPT-4) using category names as input. In the vision side, a motion encoder is built to extract temporal dynamics given densely-sampled frames and integrated with the middle-layer features from image encoder. To align the motion embeddings and video embeddings from both sides, it employs two contrastive losses, respectively. + +
Related workTaskAction annotationExample
Dist. Sup. [39]pretrainevent description‘insert window mounting bolts’
Weakly Sup. [14]pretrainevent description‘take up the iron clamp’
LAVILA [89]pretrainevent description‘a lady walks past a car’
VFC [47]VL transfer learningevent description‘two brown horses eating grass’
MoTED (Ours)VL transfer learningmotion concept description‘slapping: a quick hand movement striking against something’
+ +networks [7, 17, 21, 59, 64, 67–69, 71, 73], and vision transformers [1, 2, 6, 16, 32, 33, 38, 42, 48, 54, 55, 62]. In addition to the architecture design, self-supervised video representation learning [13, 18, 19, 23, 53, 57, 66, 74, 77, 79] has also gained popularity recently. However, these methods operate purely within the visual domain and meet bottlenecks of recognizing unseen or unfamiliar categories in real-world applications. Fortunately, the advance of large language models (LLMs) [4, 51] provides opportunities to mitigate this issue, due to their powerful capabilities of encoding world knowledge [29]. Recent studies [44, 56, 83, 85] have verified that factual sentences generated by LLMs can improve zero-shot image recognition accuracy. VFC [47] leverages PaLM [11] to create verb-focused hard negatives to enhance the understanding of verbs in video models. LSS [61] integrates language-based action concepts with self-supervised learning to adapt an image model to video domain. Our MoTED employs the descriptions that delineate the labeled classes and further accentuate the useful traits distinguishing similar classes. + +Table 1. Comparison of our method with related works focused on action annotations. 'VL' is the abbreviation of 'vision-language'. + +
Related workLLMs roleExample
Chatvideo [72]manager& summarizeInput: 'Summarize the activity in video' +Output: 'A person is cooking in the kitchen'
MM-REACT [86]execution& summarizeInput: 'Please create a summary of the video' +Output: 'The speaker is making a BLT sandwich ...'
MiniGPT4 [91]/LLAVA [41]KB& data cleanInput: 'Describe this image in detail.' +Output: 'The image shows a group of musicians ...'
MoTED (Ours)generate descriptionsInput: 'What is the motion concept of slapping' +Output: 'A quick hand movement striking against ...'
+ +Table 2. Comparison of our method with related works that take LLMs as a knowledge base (KB) and automatically annotated tool. + +# 3. Method + +The generic pipeline of MoTED is presented in Fig. 2, which consists of three steps: (1) Motion Description Generation to obtain conceptual descriptions $\mathcal{D}_{con}$ and discriminative descriptions $\mathcal{D}_{dis}$ , via querying language models given the category names of the target video dataset in Sec. 3.1; (2) Textual Motion Adaptation to obtain the text embeddings that extract the motion semantics given the generated descriptions $\mathcal{D}_{con}$ and $\mathcal{D}_{dis}$ in Sec. 3.2; (3) Visual Motion Extraction to obtain the visual embeddings that extract the motion semantics given input video frames in Sec. 3.3. Overall, the framework is trained in an end-to-end contrastive manner as illustrated in Sec. 3.4. + +# 3.1. Motion Description Generation + +The first step is to obtain a set of appropriate descriptions for each category. Given a video dataset with multiple different categories (e.g., Kinetics400 [28] with 400 action classes), the descriptions are generated automati + +cally utilizing GPT-4 [50]. Note that, the generation process is agnostic to this choice and other LLMs can be used instead. For each category name, we query GPT-4 to provide the motion concepts using the following prompt: “Q: What is the motion concept in a video of ?” A: As shown in Fig. 1, the generated concept descriptions often cover moving objects, object interactions, moving directions and speeds, etc. But the output of LLMs can also be open-ended duplicates or anything in natural language (as illustrated in the Appendix). To control the generated descriptions to be concise and motion-related, we adopt a two-shot prompt in which we include two exemplars of question-answer for the same operation that is being queried. + +Besides the conceptual descriptions $\mathcal{D}_{con}$ , we further investigate into the confusing actions that are characterized by their distinct differences from similar categories. To this end, we first compute the cosine distance of text embeddings for every two categories. Then we select top-k similar classes (e.g., $k = 5$ ) and query GPT-4 to generate motion characteristics for distinguishing the similar classes using the following prompt: "Q: What are the useful features for distinguishing the original class from the similar class + +A:" Different from the conceptual descriptions $\mathcal{D}_{con}$ that are mainly determined by the general knowledge of LLMs, the descriptions $\mathcal{D}_{dis}$ further contain task-specific information that the downstream task emphasizes. + +In this way, given a dataset of $N$ categories, we can obtain the motion-enhanced $N * (k + 1)$ descriptions, consisting of 1 conceptual description and $k$ discriminative descriptions for each category. The above generation process is completed before the training of whole framework. + +The detailed comparisons of action annotations and LLM-based generations are presented in Tab. 1 and Tab. 2, respectively. Tab. 1 shows that existing works employ annotations that describe the action events in a subject-verb-object manner, which are object-centric. In contrast, our generated annotations are motion-related, object-agnostic and to supervise the learning of visual motion features with the world knowledge of motion concepts. In Tab. 2, LLMs are mainly applied in generative tasks, to summarize query results and generate general responses, or to obtain instruction data to align with human preference. In contrast, we apply LLMs to obtain descriptions of motion concept, aligned with the temporal visual features. + +# 3.2. Textual Motion Adaptation + +After generating the motion-enhanced descriptions, the second step is to perceive the motion cues within the descriptions and aggregate them into a compact motion-enhanced embedding. Given a set of motion-enhanced descriptions $\mathcal{D}$ + +$= \left\{\mathcal{D}_{con}^{i},\mathcal{D}_{dis}^{i,j}\right\}$ where $i\in [1,N]$ and $j\in [1,k]$ , we extract normalized feature embedding $e_i$ by using the text encoder $f_{txt}$ : $e_i = f_{txt}(d_i)$ , where $d_{i}$ is a motion-enhanced description sampled from $\mathcal{D}$ . In this way, we obtain the text embeddings $E_{txt}\in \mathbb{R}^{N\times (k + 1)\times C}$ , where $C$ denotes the channel number of each embedding. Noticeably, the parameters of text encoder is frozen and initialized by CLIP [60] to inherit its capacity of encoding visual-aligned semantics. + +However, as CLIP is pre-trained on image-text paired corpus, the text encoder has a strong bias towards spatial appearance of objects and backgrounds, instead of temporal motions [5, 30, 47, 63]. To adapt the text encoder to understand the motion-enhanced descriptions, we introduce a Text Motion Adapter, which consists of a multi-head self-attention (MHSA) [70]. The adapter is parameter-tunable and takes $E_{txt}$ as input to learn the information dependencies between motion-enhanced descriptions. To aggregate the motion semantics for each category, we perform adaptive pooling to obtain the motion-enhanced embeddings $E_{txt}^{m} \in \mathbb{R}^{N \times 1 \times C}$ . + +# 3.3. Visual Motion Extraction + +In the visual side, as shown in Fig. 3, it comprises of two components: (i) The image encoder adopts the CLIP pre-trained Vision Transformer (ViT), which extracts frozen features for sparse frames with powerful spatial semantics. (ii) The motion encoder takes dense frames as input to capture the local temporal cues, integrated with the global temporal dynamics from middle-layer image features. + +# 3.3.1 Image Encoder + +Given a video clip $\mathbf{X}_S\in \mathbb{R}^{T\times H\times W\times 3}$ ( $T$ , $H$ , and $W$ represent the frame number, height, and width, respectively), the image features are extracted individually for several sparse frames. Following ViT [15], each frame is divided into $K = \frac{H}{P}\times \frac{W}{P}$ patches, and the size of each patch is denoted as $P\times P$ . These patches are then projected using a fully connected layer, referred to as the 2D stem in Fig. 3. This projection generates a sequence of patch embeddings $[\mathbf{x}_{t,\mathrm{cls}}^{(0)},\mathbf{x}_{t,1}^{(0)},\mathbf{x}_{t,2}^{(0)},\dots ,\mathbf{x}_{t,K}^{(0)}] + \mathbf{e}^{\mathrm{spatial}}$ , where $t = \{1,\dots ,T\}$ , $\mathbf{x}_{\mathrm{cls}}$ is an additional learnable token (termed as "class embedding"), and $\mathbf{e}^{\mathrm{spatial}}$ denotes spatial position embedding. Assuming that the spatial encoder has $L$ Transformer blocks, the features of the $l_{\mathrm{th}}$ layer for the $t_\mathrm{th}$ frame can be extracted by: + +$$ +\mathbf {X} _ {t} ^ {(l)} = \operatorname {T r a n s f o r m e r} ^ {(l)} \left(\mathbf {X} _ {t} ^ {(l - 1)}\right) \in \mathbb {R} ^ {(K + 1) \times C}, \tag {1} +$$ + +where $l = \{1, \dots, L\}$ denotes the layer index. The class embeddings of $T$ frames in the $l_{\mathrm{th}}$ layer is termed as $\mathbf{X}^{(l)} = [\mathbf{X}_1^{(l)}, \dots, \mathbf{X}_T^{(l)}] \in \mathbb{R}^{T \times 1 \times C}$ . Benefit from the CLIP pretrained parameters, the class embeddings $\mathbf{X}^{(L)}$ aggregate powerful spatial semantics within each frame. + +![](images/e322011f8dfecfe46dbb4134236f27bde2cc8040bd79533ffd81132398abc644.jpg) +Figure 3. The structural details of MoTED in the vision side. + +# 3.3.2 Motion Encoder + +To fully extract the motion information in videos, the temporal input $\mathbf{X}_T\in \mathbb{R}^{\alpha T\times H\times W\times 3}$ for the motion encoder is sampled around the spatial input $\mathbf{X}_S$ by $\alpha$ times. In this study, we set $\alpha = 2$ by default, as empirically studied in [17, 58]. Then, $\mathbf{X}_T$ is projected by a 3D convolution, i.e., the 3D stem in Fig. 3, for patch embedding. The kernel size and stride of the 3D stem are both $P$ in spatial dimension and $\alpha$ in temporal dimension. Thus the number of temporal patch tokens are the same with that in the image encoder, which makes it convenient to integrate the spatial and temporal features. Thus the projected temporal features can be formulated as: $\mathbf{Z}^{(0)} = \mathrm{Conv3d}(\mathbf{X}_T)\in \mathbb{R}^{T\times K\times C}$ , where $K = \frac{H}{P}\times \frac{W}{P}$ . Then, a series of Temporal Blocks are designed to extract motion patterns, which can be written as: + +$$ +\mathbf {Z} ^ {(l)} = \operatorname {T e m p - B l o c k} ^ {(l)} \left(\mathbf {Z} ^ {(l - 1)}, \mathbf {X} ^ {(l - 1)}\right) \in \mathbb {R} ^ {T \times K \times C}, \tag {2} +$$ + +where the function Temp-Block $(\cdot)$ performs temporal modeling integrated with the spatial features $X^{(l - 1)}$ + +Following the effective and efficient design philosophy in [32, 33], the temporal block consists of the following three modules: (i) Local Motion Extraction (LME). Since 3D convolution can capture detailed and local spatiotemporal features, by processing each pixel with context from a small 3D neighborhood (e.g., $2 \times 16 \times 16$ ), the temporal patch tokens $\mathbf{Z}^{(0)}$ remain the nature of local motion extraction. Then these tokens are flattened to a sequence $\mathbb{R}^{TK \times C}$ and perform spatiotemporal learning via joint self-attention modules [66] and token merging for efficiency [3]. (ii) Global Motion Extraction (GME). Given the class embeddings $\mathbf{X}^{(l)}$ that aggregate powerful spatial semantics for each individual frame, we apply a cross-frame self-attention module [49] to learn the global inter-frame interactions. (iii) Feature Integration. After obtaining the local and global motion features, we adopt a cross-attention mod + +ule, which takes local features as query and global features as key/value, to complement the features for better temporal modeling. In this way, the output of the final temporal block $\mathbf{Z}^{(L)}$ aggregates powerful temporal dynamics within the dense frames. + +# 3.4. Training Loss + +Following CLIP [60], we first apply adaptive pooling to obtain a video-level motion embedding $\mathbf{Z}_{avg}^{(L)} \in \mathbb{R}^{1 \times C}$ . Then we introduce a contrastive loss to align the visual and textual motion embeddings: + +$$ +\mathcal {L} _ {\text {m o t i o n}} = - \log \frac {\exp \left(\sin \left(\mathbf {Z} _ {\text {a v g}} ^ {(L)} , \mathbf {e} _ {i}\right) / \tau\right)}{\sum_ {c = 1} ^ {N} \exp \left(\sin \left(\mathbf {Z} _ {\text {a v g}} ^ {(L)} , \mathbf {e} _ {c}\right) / \tau\right)}, \tag {3} +$$ + +where $\mathrm{sim}(\cdot ,\cdot)$ is the normalized cosine similarity, $\tau$ refers to the temperature parameter. $\mathbf{e}_i$ is a motion-enhanced text embedding for the $i_{\mathrm{th}}$ category. + +Moreover, we adopt a cross-attention module to fuse the visual motion features (as key/value) and image features (as query) to obtain the compact video embedding $\mathbf{v}_i$ . Similarly, the fused text embedding $\mathbf{e}_i'$ is acquired via a cross-attention module. We introduce a contrastive loss to align the visual and textual video-level embeddings: + +$$ +\mathcal {L} _ {\text {v i d e o}} = - \log \frac {\exp \left(\operatorname {s i m} \left(\mathbf {v} _ {i} , \mathbf {e} _ {i} ^ {\prime}\right) / \tau\right)}{\sum_ {c = 1} ^ {N} \exp \left(\operatorname {s i m} \left(\mathbf {v} _ {i} , \mathbf {e} _ {c} ^ {\prime}\right) / \tau\right)}. \tag {4} +$$ + +Compatible with CLIP training paradigm, the overall framework is trained based on these two contrastive losses: + +$$ +\mathcal {L} = \mathcal {L} _ {\text {m o t i o n}} + \mathcal {L} _ {\text {v i d e o}}. \tag {5} +$$ + +# 4. Experiments + +# 4.1. Dataset and Implementation. + +Datasets. In the supervised setting, we train on the train set of Kinetics-400 (K400) [28] and Something-Something-V2 (SSv2) [20] and report supervised performance against existing methods on the validation sets of K400 and SSv2. In the zero-shot setting, we train on the Kinetics-400 training set and evaluate on three datasets: Kinetics-600 (K600) [12], HMDB51 [24] and UCF101 [65]. For zero-shot evaluation on K600, following [12], we use the 220 new categories outside of K400 for evaluation, and conduct evaluation three times, each time randomly sampling 160 categories for evaluation from the 220 categories. For zero-shot evaluation on HMDB51 and UCF101, we follow [49] and report average top-1 accuracy and standard deviation on three splits of the test set. + +Implementation Details. Following previous work [40, 58], we use the CLIP [60] pre-trained ViT-B/16, ViT-L/14 + +
Class NamesConcept.Discr.SSv2K400
✓XX65.782.9
✓✓X66.084.5
✓X✓69.984.0
✓✓✓70.185.1
+ +(a) "Concept." is the abbreviation of "conceptual descriptions". "Discr." is the abbreviation of "discriminative descriptions". + +
M.Enc.Adapter.SSv2K400
XX64.381.5
✓X68.683.4
X✓66.283.8
✓✓70.185.1
+ +(b) "M.Enc." is the separated motion encoder where motion modules are parallel with CLIP visual encoder. "Adapter" is the text motion adapter. + +
Text.Fuse.Vision.Fuse.SSv2K400
XX68.383.6
✓X68.984.2
X✓69.584.5
✓✓70.185.1
+ +(c) "Text.Fuse." indicates feature fusion in the textual side. "Vision.Fuse." indicates feature fusion in the visual side. + +Table 3. Ablations on Something-Something-V2 and Kinetics-400. The spatial encoder is a 8-frame vanilla ViT-B/16 pre-trained by CLIP [60]. The inference protocol of all models and datasets are 3 clips × 1 center crop. + +![](images/b848dc7327e56cbfbc7307b0faf989346b6773822fa7005d57c405e14d0bbb5d.jpg) + +![](images/661bdcac5cfcf91f68631a263b297c87e0fb3d9b585739073bc9f1af0cc3a846.jpg) +Relevance between text and motion features +(a) case 1 + +![](images/4aef931383e4376f9db7a3352bfcfbf88e3155f28f48a8ad5b5c84f4a2ce5541.jpg) + +![](images/1812d5779aff347aaf381d9b414dd1265f58885d7ae41061d6641f136837685f.jpg) +Relevance between text and motion features +(b) case2 +Figure 4. Two cases to visualize the relevance [8] between text and image/motion features to highlight the information relevant to the prediction. The different "regions of interest" and "words of importance" indicate that the motion and image features could be disentangled. + +and ViT-L/14-336p as our image encoder. Unless otherwise specified, we mark the default settings in the temporal encoder in gray in Sec. 4.2. For simplicity, the ablation studies are conducted based on ViT-B/16 with a low-rate sampled 8 frames. We conduct the experiments with the NVIDIA 32G V100 GPUs. More implementation details (e.g., training and testing hyper-parameters) are described in the Appendix. + +# 4.2. Ablation and Analysis + +Language Supervision. In this section, we aim to demonstrate the significance of language supervision by comparing different types of text descriptions. Tab. 3a presents a comparison of three types of texts: category names, motion concept descriptions, motion discriminative descriptions. It reveals that the performance of action category names, which contain the least information, is notably lower than other results, particularly on the SSv2 dataset with abundant temporal information. Although descriptions with only action concepts show a slight improvement of $0.3\% /1.6\%$ , they lack the vital information regarding the distinguishing characteristics of the actions. As a result, the model exhibits poor classification accuracy for similar categories. On the other hand, when the discriminative descriptions are adopted, the accuracy on SSv2 is improved noticeably $(+4.2\%)$ , revealing the significance of discriminative characteristics for fine-grained datasets. Additionally, when the full descriptions are utilized, the model's performance obtain the further gains of $4.4\% /2.2\%$ . + +Disentangled Motion Modeling. Our method has an advantage in allowing the learning of dynamic information without interference, while preserving CLIP's original spatially transferable representation capabilities. To validate this, we compared two different structures in Tab. 3b. For the serial structure, additional dynamic modules are inserted between the Transformer Blocks of each layer of the visual encoder, resulting in a unified video representation. All the texts are combined through the text encoder to obtain text representations for alignment. It can be observed that the parallel structure outperforms the serial structure, with improvements of $4.3\%$ and $1.9\%$ respectively. This demonstrates the rationality of the parallel approach. Tab. 3b also presents a comparison of introducing an adapter in the text encoder. The model's classification performance improved by $1.9\%$ and $2.3\%$ after adding the adapter, indicating its effectiveness in reducing bias in the text encoder. + +Feature Fusion Direction. As shown in Tab. 3c, both directions of information integration can improve performances. The combination of the two can boost accuracy more significantly $+1.8\% / + 1.5\%$ on SSv2/K400, respectively. This verifies the importance of spatio-temporal blending for the parallel architectures. In our opinion, independent learning of motion can effectively avoid excessive reliance of the model on the previously learned spatial information. By using cross attention for fusion, it is possible to effectively integrate features from two different dimensions, ultimately forming the features of the target video and achieving the best results. + +
MethodPre-trainArchitectureInput SizeFLOPs×Cr.×Cl. (T)Param (M)FrozenTop-1Top-5
SlowFast [17]ImageNet-21KR101+NL16 × 22420.1 × 3 × 160X63.187.6
ViViT FE [1]IN21K+K400ViT-L16 × 22421.0 × 3 × 4612X65.489.8
MTV-B(320p) [84]IN21K+K400-32 × 22420.9 × 3 × 4310X68.590.4
MViT [16]Kinetics-600MViT-B-2432 × 22420.2 × 3 × 153X68.791.5
Video Swin [42]IN21K+K400Swin-B32 × 22420.3 × 3 × 160X69.692.7
TAdaConvNeXtV2 [22]IN1K+K400ConvNeXt-S32 × 22420.2 × 3 × 282X70.092.0
EVL* [40]CLIP-400MViT-B32 × 22420.68 × 1 × 3175✓62.4-
ST-Adapter* [52]CLIP-400MViT-B32 × 22420.61 × 1 × 393✓69.592.6
DiST*CLIP-400MViT-B32 × 22420.65 × 1 × 3105✓70.992.1
MoTED*CLIP-400MViT-B8 × 22420.18 × 1 × 3112✓70.191.8
MoTED*CLIP-400MViT-B16 × 22420.34 × 1 × 3112✓71.292.4
MoTED*CLIP-400MViT-B32 × 22420.68 × 1 × 3112✓71.992.7
UnifromerV2 [32]CLIP-400MViT-L32 × 22421.73 × 1 × 3574X73.094.5
TAdaFormer [22]CLIP-400MViT-L32 × 22421.70 × 2 × 3364X73.6-
EVL* [40]CLIP-400MViT-L32 × 22423.21 × 1 × 3654✓66.7-
EVL* [40]CLIP-400MViT-L32 × 33628.08 × 1 × 3654✓68.0-
ST-Adapter* [52]CLIP-400MViT-L32 × 22422.75 × 1 × 3347✓72.393.9
DiST*CLIP-400MViT-L32 × 22422.83 × 1 × 3336✓73.193.2
MoTED*CLIP-400MViT-L8 × 22420.78 × 1 × 3346✓71.592.6
MoTED*CLIP-400MViT-L16 × 22421.49 × 1 × 3346✓73.093.4
MoTED*CLIP-400MViT-L32 × 22422.89 × 1 × 3346✓73.893.8
+ +Motion Modeling Visualization. In Fig. 4, we also perform the analysis of vision features generated by image and motion encoder to investigate the learned patterns from language supervision. Based on the reasoning tool [8], we depict the attention maps of class token from the final transformer block of the image/motion stream encoder w.r.t. the text encoder. It is observed that, for the case1 in Fig. 4, the CLIP image encoder attends to both motion-relevant foreground and motion-irrelevant background with a major focus on "surface". In contrast, features extracted from the motion modules concentrate on the motion-relevant regions of the moving object, and emphasize the motion-related words "sliding". This phenomenon reveals that the motion features can complement the static spatial semantics of objects in the image features via modeling object motions. More cases can be accessed in the Appendix. + +# 4.3. Fully-supervised Experiments + +In the supervised setting, the results on SSv2 and K400 are presented in Tab. 4 and Tab. 6, respectively. Compared with AVL, our proposed MoTED introduces a similar temporal module for CLIP visual encoder, but our method has a significant improvement compared to AVL, with an improvement of $9.5\% / 2.0\%$ and $5.8\% / 1.5\%$ on ViT-B and ViT-L respectively on SSv2/K400. The performance gains greatly demonstrate the rationality and effectiveness of using lan + +guage supervision. Interestingly, the performance gains are particularly pronounced on SSv2. This is because SSv2 is a fine-grained action classification task requiring stronger action discrimination and perception of dynamic information. + +Table 4. Comparison with the state-of-the-art methods on Something-Something-V2. "Cr." and "Cl." are the abbreviation for "spatial crops" and "temporal clips". "Frozen" indicates freezing the CLIP pre-trained parameters. + +
Related workVisual disentangleTextual disentangleK400 (Acc / Δ)SSv2 (Acc / Δ)
Previous SOTAXX84.2 / -69.5 / -
DiST [58]✓X85.0 / +0.870.9 / +1.4
MoTED (Ours)✓✓86.2 / +1.271.9 / +1.0
+ +Table 5. Comparison of our method with previous state-of-the-art (SOTA) and the latest related work DiST. + +Disentangling spatio-temporal learning is an effective approach to endow the model with temporal capability. As shown in Tab. 5, DiST focuses on temporal disentangling in the visual side with gains of $+0.8\% / +1.4\%$ on K400/SSv2. Our study highlights the effectiveness of temporal disentangling in the textual side with detailed descriptions of motions, with further gains of $+1.2\% / +1.0\%$ . This result reveals that disentangling textual encoder is equally effective w.r.t. disentangling visual encoder for vision-language transfer learning. + +# 4.4. Zero-shot Experiments + +Zero-shot generalization is an attractive characteristic of CLIP-extended models, making them more practical in real + +
MethodPre-trainArchitectureInput SizeTFLOPs×Cr.×Cl.Param (M)FrozenTop-1Top-5
SlowFast [17]-R101+NL16 × 22420.4 × 3 × 1060X79.893.9
TimeSformer [2]ImageNet-21KViT-L96 × 22428.4 × 3 × 1430X80.794.7
MViT [16]-MViT-B64 × 22420.5 × 1 × 537X81.295.1
ViViT FE [1]ImageNet-21KViT-L128 × 22424.0 × 3 × 1N/AX81.793.8
Video Swin [42]ImageNet-21KSwin-L32 × 22420.6 × 3 × 4197X83.195.9
TAdaConvNeXtV2 [22]ImageNet-21KConvNeXt-B32 × 22420.3 × 3 × 4146X83.7-
X-CLIP [49]CLIP-400MViT-B16 × 22420.28 × 3 × 4128X84.796.8
ST-Adapter* [52]CLIP-400MViT-B32 × 22420.61 × 1 × 393✓82.796.2
EVL* [40]CLIP-400MViT-B32 × 22420.59 × 1 × 3115✓84.2-
DiST*CLIP-400MViT-B32 × 22420.65 × 1 × 3112✓85.097.0
MoTED*CLIP-400MViT-B8 × 22420.18 × 1 × 3116✓85.197.0
MoTED*CLIP-400MViT-B16 × 22420.34 × 1 × 3116✓85.497.2
MoTED*CLIP-400MViT-B32 × 22420.68 × 1 × 3116✓86.297.5
UnifromerV2 [32]CLIP-400M+K710ViT-L32 × 22422.66 × 2 × 3354X89.398.2
TAdaFormer [22]CLIP-400M+K710ViT-L32 × 22421.41 × 4 × 3364X89.5-
ST-Adapter* [52]CLIP-400MViT-L32 × 22422.75 × 1 × 3347✓87.297.6
EVL* [40]CLIP-400MViT-L32 × 22422.70 × 1 × 3363✓87.3-
DiST*CLIP-400MViT-L32 × 22422.83 × 1 × 3343✓88.097.9
MoTED*CLIP-400MViT-L8 × 22420.78 × 1 × 3349✓87.497.8
MoTED*CLIP-400MViT-L16 × 22421.49 × 1 × 3349✓88.098.0
MoTED*CLIP-400MViT-L32 × 22422.89 × 1 × 3349✓88.898.2
+ +Table 6. Comparison with state-of-the-arts on Kinetics-400. "Cr." and "Cl." are the abbreviation for "spatial crops" and "temporal clips". "Frozen" indicates freezing the CLIP pre-trained parameters. + +
MethodModelHMDB51UCF101K600
ActionCLIP [75]B/1640.8±5.458.3±3.4-
X-CLIP [49]B/1644.6±5.272.0±2.365.2±0.4
Vita-CLIP [78]B/1648.6±0.675.0±0.667.4±0.5
DiST*B/1655.4±1.272.3±0.6-
MoTED*B/1658.2±1.178.3±0.669.9 ± 0.5
+ +Table 7. Comparison of zero-shot accuracy with the state-of-the-art CLIP-based methods on three datasets (e.g., HMDB51 [24], UCF101 [65], and K600 [12]). "": frozen backbone. + +world applications. For zero-shot settings, we evaluate MoTED on three widely used benchmarks. Following prior work [49], we train the networks on the K400 training set, then conduct the zero-shot evaluation on three unseen datasets (i.e., UCF101, HMDB51 and K600), as shown in Tab. 7. All models share the same architecture of "ViT-B", with 32 frames during inference. Compared with other vison-language methods, MoTED achieves the better zero-shot performances with a significant margin on HMDB51 $(+3.8\%)$ , UCF101 $(+6.0\%)$ and K600 $(+2.5\%)$ . Different from X-CLIP [49], DiST [58], Vita-CLIP [78] that learns motion representation with the supervision of action category names solely, the proposed MoTED makes full use of conceptual and discriminative descriptions and learn general motion representations with the aid of language supervision. In addition, our method also has a relatively small + +variance, only about $1\%$ . We assume this is due to the benefits brought by the rich content of the text, as detailed descriptions can promote stable dynamic feature learning. + +# 5. Conclusion + +In this study, we aim to overcome the limitations of existing methods in extending large image-text pre-trained models for video understanding. The proposed MoTED framework introduces Motion-Enhanced Descriptions, which are applied to facilitate the extraction of unique temporal features in videos. By generating motion-related descriptions and incorporating perception modules, MoTED aligns visual and textual motion features using a contrastive loss. Experimental results on five benchmarks demonstrate that MoTED provides a strong basis for enhancing CLIP with robust temporal modeling. In future works, we hope to further dedicate to exploring the potential of language supervision and combining it with more powerful dynamic information perception modules to achieve higher performance in video recognition and make it truly practical. + +# 6. Acknowledgement + +This work was supported in part by NSFC No. 62222117, the Fundamental Research Funds for the Central Universities under contract WK3490000005, and KY2100000117. + +# References + +[1] Anurag Arnab, Mostafa Dehghani, Georg Heigold, Chen Sun, Mario Lučić, and Cordelia Schmid. Vivit: A video vision transformer. In ICCV, pages 6836-6846, 2021. 3, 7, 8 +[2] Gedas Bertasius, Heng Wang, and Lorenzo Torresani. Is space-time attention all you need for video understanding? In ICML, page 4, 2021. 3, 8 +[3] Daniel Bolya, Cheng-Yang Fu, Xiaoliang Dai, Peizhao Zhang, Christoph Feichtenhofer, and Judy Hoffman. Token merging: Your vit but faster. In ICLR, 2023. 5 +[4] Tom B. Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-Voss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel M. Ziegler, Jeffrey Wu, Clemens Winter, Christopher Hesse, Mark Chen, Eric Sigler, Mateusz Litwin, Scott Gray, Benjamin Chess, Jack Clark, Christopher Berner, Sam McCandlish, Alec Radford, Ilya Sutskever, and Dario Amodei. Language models are few-shot learners. In NeurIPS, 2020. 2, 3 +[5] Shyamal Buch, Cristóbal Eyzaguirre, Adrien Gaidon, Jiajun Wu, Li Fei-Fei, and Juan Carlos Niebles. Revisiting the "video" in video-language understanding. In CVPR, pages 2907–2917, 2022. 2, 4 +[6] Adrian Bulat, Juan Manuel Perez Rua, Swathikiran Sudhakaran, Brais Martinez, and Georgios Tzimiropoulos. Space-time mixing attention for video transformer. NeurIPS, 34:19594-19607, 2021. 3 +[7] Joao Carreira and Andrew Zisserman. Quo vadis, action recognition? a new model and the kinetics dataset. In CVPR, pages 6299-6308, 2017. 3 +[8] Hila Chefer, Shir Gur, and Lior Wolf. Generic attention-model explainability for interpreting bi-modal and encoder-decoder transformers. In ICCV, pages 387-396, 2021. 6, 7 +[9] Xi Chen, Xiao Wang, Soravit Changpinyo, A. J. Piergiovanni, Piotr Padlewski, Daniel Salz, Sebastian Goodman, Adam Grycner, Basil Mustafa, Lucas Beyer, Alexander Kolesnikov, Joan Puigcerver, Nan Ding, Keran Rong, Hassan Akbari, Gaurav Mishra, Linting Xue, Ashish V. Thapliyal, James Bradbury, and Weicheng Kuo. Pali: A jointly scaled multilingual language-image model. In ICLR, 2023. 1, 2 +[10] Xing Cheng, Hezheng Lin, Xiangyu Wu, Fan Yang, and Dong Shen. Improving video-text retrieval by multi-stream corpus alignment and dual softmax loss. arXiv preprint arXiv:2109.04290, 2021. 2 +[11] Aakanksha Chowdhery, Sharan Narang, Jacob Devlin, Maarten Bosma, Gaurav Mishra, Adam Roberts, Paul Barham, Hyung Won Chung, Charles Sutton, Sebastian Gehrmann, Parker Schuh, Kensen Shi, Sasha Tsvyashchenko, Joshua Maynez, Abhishek Rao, Parker Barnes, Yi Tay, Noam Shazeer, Vinodkumar Prabhakaran, Emily Reif, Nan Du, Ben Hutchinson, Reiner Pope, James Bradbury, Jacob Austin, Michael Isard, Guy Gur-Ari, Pengcheng Yin, Toju Duke, Anselm Levskaya, Sanjay + +Ghemawat, Sunipa Dev, Henryk Michalewski, Xavier Garcia, Vedant Misra, Kevin Robinson, Liam Fedus, Denny Zhou, Daphne Ippolito, David Luan, Hyeontaek Lim, Barret Zoph, Alexander Spiridonov, Ryan Sepassi, David Dohan, Shivani Agrawal, Mark Omernick, Andrew M. Dai, Thanumalayan Sankaranarayana Pillai, Marie Pellat, Aitor Lewkowycz, Erica Moreira, Rewon Child, Oleksandr Polozov, Katherine Lee, Zongwei Zhou, Xuezhi Wang, Brennan Saeta, Mark Diaz, Orhan Firat, Michele Catasta, Jason Wei, Kathy Meier-Hellstern, Douglas Eck, Jeff Dean, Slav Petrov, and Noah Fiedel. Palm: Scaling language modeling with pathways. J. Mach. Learn. Res., 24:240:1-240:113, 2023. 2, 3 +[12] Dima Damen, Hazel Doughty, Giovanni Maria Farinella, Antonino Furnari, Evangelos Kazakos, Jian Ma, Davide Moltisanti, Jonathan Munro, Toby Perrett, Will Price, et al. Rescaling egocentric vision. arXiv preprint arXiv:2006.13256, 2020. 2, 5, 8 +[13] Ali Diba, Vivek Sharma, Reza Safdari, Dariush Lotfi, Saquib Sarfraz, Rainer Stiefelhagen, and Luc Van Gool. Vi2cIr: Video and image for visual contrastive learning of representation. In ICCV, pages 1502-1512, 2021. 3 +[14] Sixun Dong, Huazhang Hu, Dongze Lian, Weixin Luo, Yicheng Qian, and Shenghua Gao. Weakly supervised video representation learning with unaligned text for sequential videos. In CVPR, pages 2437-2447. IEEE, 2023. 3 +[15] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2020. 4 +[16] Haoqi Fan, Bo Xiong, Karttikeya Mangalam, Yanghao Li, Zhicheng Yan, Jitendra Malik, and Christoph Feichtenhofer. Multiscale vision transformers. In ICCV, pages 6824-6835, 2021. 3, 7, 8 +[17] Christoph Feichtenhofer, Haoqi Fan, Jitendra Malik, and Kaiming He. Slowfast networks for video recognition. In ICCV, pages 6202-6211, 2019. 3, 5, 7, 8 +[18] Christoph Feichtenhofer, Haoqi Fan, Bo Xiong, Ross Girshick, and Kaiming He. A large-scale study on unsupervised spatiotemporal representation learning. In CVPR, pages 3299-3309, 2021. 3 +[19] Christoph Feichtenhofer, Haoqi Fan, Yanghao Li, and Kaiming He. Masked autoencoders as spatiotemporal learners. NeurIPS, 2022. 3 +[20] Raghav Goyal, Samira Ebrahimi Kahou, Vincent Michalski, Joanna Materzynska, Susanne Westphal, Heuna Kim, Valentin Haenel, Ingo Fruend, Peter Yianilos, Moritz Mueller-Freitag, et al. The" something something" video database for learning and evaluating visual common sense. In ICCV, pages 5842-5850, 2017. 2, 5 +[21] Ziyuan Huang, Shiwei Zhang, Liang Pan, Zhiwu Qing, Mingqian Tang, Ziwei Liu, and Marcelo H Ang Jr. Tada! temporally-adaptive convolutions for video understanding. In ICLR, 2022. 3 +[22] Ziyuan Huang, Shiwei Zhang, Liang Pan, Zhiwu Qing, Yingya Zhang, Ziwei Liu, and Marcelo H Ang Jr. + +Temporally-adaptive models for efficient video understanding. arXiv preprint arXiv:2308.05787, 2023. 7, 8 +[23] Simon Jenni and Hailin Jin. Time-equivariant contrastive video representation learning. In ICCV, pages 9970-9980, 2021. 3 +[24] H Jhuang, H Garrote, E Poggio, T Serre, and T Hmdb. A large video database for human motion recognition. In ICCV, page 6, 2011. 2, 5, 8 +[25] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In ICML, pages 4904-4916. PMLR, 2021. 1, 2 +[26] Chen Ju, Tengda Han, Kunhao Zheng, Ya Zhang, and Weidi Xie. Prompting visual-language models for efficient video understanding. In ECCV, pages 105-124. Springer, 2022. 1, 2 +[27] Chen Ju, Tengda Han, Kunhao Zheng, Ya Zhang, and Weidi Xie. Prompting visual-language models for efficient video understanding. In ECCV, pages 105-124. Springer, 2022. 1 +[28] Will Kay, Joao Carreira, Karen Simonyan, Brian Zhang, Chloe Hillier, Sudheendra Vijayanarasimhan, Fabio Viola, Tim Green, Trevor Back, Paul Natsev, et al. The kinetics human action video dataset. arXiv preprint arXiv:1705.06950, 2017. 2, 3, 5 +[29] Jan Kocon, Igor Cichecki, Oliwier Kaszyca, Mateusz Kochanek, Dominika Szydlo, Joanna Baran, Julita Bielaniewicz, Marcin Gruza, Arkadiusz Janz, Kamil Kanclerz, Anna Kocon, Bartlomiej Koptyra, Wiktoria Mieleszczenko-Kowszewicz, Piotr Milkowski, Marcin Oleksy, Maciej Piasecki, Lukasz Radlinski, Konrad Wojtasik, Stanislaw Wozniak, and Przemyslaw Kazienko. Chatgpt: Jack of all trades, master of none. Inf. Fusion, 99: 101861, 2023. 3 +[30] Jie Lei, Tamara L. Berg, and Mohit Bansal. Revealing single frame bias for video-and-language learning. In ACL, pages 487-507, 2023. 2, 4 +[31] Chunyuan Li, Haotian Liu, Liunian Harold Li, Pengchuan Zhang, Jyoti Aneja, Jianwei Yang, Ping Jin, Houdong Hu, Zicheng Liu, Yong Jae Lee, and Jianfeng Gao. ELEVATER: A benchmark and toolkit for evaluating language-augmented visual models. In NeurIPS, 2022. 2 +[32] Kunchang Li, Yali Wang, Yinan He, Yizhuo Li, Yi Wang, Limin Wang, and Yu Qiao. Uniformerv2: Spatiotemporal learning by arming image vits with video uniformer. arXiv preprint arXiv:2211.09552, 2022. 3, 5, 7, 8 +[33] Kunchang Li, Yali Wang, Junhao Zhang, Peng Gao, Guanglu Song, Yu Liu, Hongsheng Li, and Yu Qiao. Uniformer: Unifying convolution and self-attention for visual recognition. TPAMI, 2023. 3, 5 +[34] Linjie Li, Yen-Chun Chen, Yu Cheng, Zhe Gan, Licheng Yu, and Jingjing Liu. HERO: hierarchical encoder for video+language omni-representation pre-training. In EMNLP, pages 2046-2065, 2020. 2 +[35] Liunian Harold Li, Pengchuan Zhang, Haotian Zhang, Jianwei Yang, Chunyuan Li, Yiwu Zhong, Lijuan Wang, Lu Yuan, Lei Zhang, Jenq-Neng Hwang, Kai-Wei Chang, and + +Jianfeng Gao. Grounded language-image pre-training. In CVPR, pages 10955-10965, 2022. 2 +[36] Yanghao Li, Haoqi Fan, Ronghang Hu, Christoph Feichtenhofer, and Kaiming He. Scaling language-image pre-training via masking. arXiv preprint arXiv:2212.00794, 2022. 1, 2 +[37] Yangguang Li, Feng Liang, Lichen Zhao, Yufeng Cui, Wanli Ouyang, Jing Shao, Fengwei Yu, and Junjie Yan. Supervision exists everywhere: A data efficient contrastive language-image pre-training paradigm. In ICLR, 2022. 2 +[38] Yanghao Li, Chao-Yuan Wu, Haoqi Fan, Karttikeya Mangalam, Bo Xiong, Jitendra Malik, and Christoph Feichtenhofer. Mvitv2: Improved multiscale vision transformers for classification and detection. In CVPR, pages 4804-4814, 2022. 3 +[39] Xudong Lin, Fabio Petroni, Gedas Bertasius, Marcus Rohrbach, Shih-Fu Chang, and Lorenzo Torresani. Learning to recognize procedural activities with distant supervision. In CVPR, pages 13843-13853. IEEE, 2022. 3 +[40] Ziyi Lin, Shijie Geng, Renrui Zhang, Peng Gao, Gerard de Melo, Xiaogang Wang, Jifeng Dai, Yu Qiao, and Hongsheng Li. Frozen clip models are efficient video learners. In ECCV, pages 388-404. Springer, 2022. 1, 2, 5, 7, 8 +[41] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. In NeurIPS, 2023. 3 +[42] Ze Liu, Jia Ning, Yue Cao, Yixuan Wei, Zheng Zhang, Stephen Lin, and Han Hu. Video swim transformer. In CVPR, pages 3202-3211, 2022. 3, 7, 8 +[43] Timo Lüddecke and Alexander S. Ecker. Image segmentation using text and image prompts. In CVPR, pages 7076-7086, 2022. 2 +[44] Sachit Menon and Carl Vondrick. Visual classification via description from large language models. In ICLR, 2023. 3 +[45] Antoine Miech, Dimitri Zhukov, Jean-Baptiste Alayrac, Makarand Tapaswi, Ivan Laptev, and Josef Sivic. Howto100m: Learning a text-video embedding by watching hundred million narrated video clips. In ICCV, pages 2630-2640, 2019. 2 +[46] Antoine Miech, Jean-Baptiste Alayrac, Lucas Smaira, Ivan Laptev, Josef Sivic, and Andrew Zisserman. End-to-end learning of visual representations from uncurated instructional videos. In CVPR, pages 9879-9889, 2020. 2 +[47] Liliane Momeni, Mathilde Caron, Arsha Nagrani, Andrew Zisserman, and Cordelia Schmid. Verbs in action: Improving verb understanding in video-language models. ICCV, 2023. 2, 3, 4 +[48] Daniel Neimark, Omri Bar, Maya Zohar, and Dotan Asselmann. Video transformer network. In ICCV, pages 3163-3172, 2021. 3 +[49] Bolin Ni, Houwen Peng, Minghao Chen, Songyang Zhang, Gaofeng Meng, Jianlong Fu, Shiming Xiang, and Haibin Ling. Expanding language-image pretrained models for general video recognition. In ECCV, pages 1-18. Springer, 2022. 1, 2, 5, 8 +[50] OpenAI. Gpt-4 technical report, 2023. 1, 2, 4 +[51] TB OpenAI. Chatgpt: Optimizing language models for dialogue. OpenAI, 2022. 3 + +[52] Junting Pan, Ziyi Lin, Xiatian Zhu, Jing Shao, and Hongsheng Li. Parameter-efficient image-to-video transfer learning. arXiv e-prints, pages arXiv-2206, 2022. 1, 2, 7, 8 +[53] Tian Pan, Yibing Song, Tianyu Yang, Wenhao Jiang, and Wei Liu. Videomoco: Contrastive video representation learning with temporally adversarial examples. In CVPR, pages 11205-11214, 2021. 3 +[54] Jungin Park, Jiyoung Lee, and Kwanghoon Sohn. Dual-path adaptation from image to video transformers. CVPR, 2023. 3 +[55] Mandela Patrick, Dylan Campbell, Yuki Asano, Ishan Misra, Florian Metze, Christoph Feichtenhofer, Andrea Vedaldi, and João F Henriques. Keeping your eye on the ball: Trajectory attention in video transformers. NeurIPS, 34:12493-12506, 2021. 3 +[56] Sarah M. Pratt, Rosanne Liu, and Ali Farhadi. What does a platypus look like? generating customized prompts for zero-shot image classification. ICCV, 2022. 3 +[57] Rui Qian, Tianjian Meng, Boqing Gong, Ming-Hsuan Yang, Huisheng Wang, Serge Belongie, and Yin Cui. Spatiotemporal contrastive video representation learning. In CVPR, pages 6964-6974, 2021. 3 +[58] Zhiwu Qing, Shiwei Zhang, Ziyuan Huang, Yingya Zhang, Changxin Gao, Deli Zhao, and Nong Sang. Disentangling spatial and temporal learning for efficient image-to-video transfer learning. ICCV, 2023. 1, 2, 5, 7, 8 +[59] Zhaofan Qiu, Ting Yao, and Tao Mei. Learning spatiotemporal representation with pseudo-3d residual networks. In ICCV, pages 5533–5541, 2017. 3 +[60] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, pages 8748-8763. PMLR, 2021. 1, 2, 3, 4, 5, 6 +[61] Kanchana Ranasinghe and Michael S. Ryoo. Language-based action concept spaces improve video self-supervised learning. CoRR, abs/2307.10922, 2023. 3 +[62] Michael Ryoo, AJ Piergiovanni, Anurag Arnab, Mostafa Dehghani, and Anelia Angelova. Tokenlearner: Adaptive space-time tokenization for videos. *NeurIPS*, 34:12786-12797, 2021. 3 +[63] Laura Sevilla-Lara, Shengxin Zha, Zhicheng Yan, Vedanuj Goswami, Matt Feiszli, and Lorenzo Torresani. Only time can tell: Discovering temporal data for temporal modeling. In WACV, pages 535-544, 2021. 2, 4 +[64] Karen Simonyan and Andrew Zisserman. Two-stream convolutional networks for action recognition in videos. NeurIPS, 27, 2014. 3 +[65] Khurram Soomro, Amir Roshan Zamir, and Mubarak Shah. Ucf101: A dataset of 101 human actions classes from videos in the wild. arXiv preprint arXiv:1212.0402, 2012. 2, 5, 8 +[66] Zhan Tong, Yibing Song, Jue Wang, and Limin Wang. Videomae: Masked autoencoders are data-efficient learners for self-supervised video pre-training. NeurIPS, 2022. 3, 5 +[67] Du Tran, Lubomir Bourdev, Rob Fergus, Lorenzo Torresani, and Manohar Paluri. Learning spatiotemporal features + +with 3d convolutional networks. In ICCV, pages 4489-4497, 2015. 3 +[68] Du Tran, Heng Wang, Lorenzo Torresani, Jamie Ray, Yann LeCun, and Manohar Paluri. A closer look at spatiotemporal convolutions for action recognition. In CVPR, pages 6450-6459, 2018. +[69] Du Tran, Heng Wang, Lorenzo Torresani, and Matt Feiszli. Video classification with channel-separated convolutional networks. In ICCV, pages 5552-5561, 2019. 3 +[70] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In NeurIPS, pages 5998-6008, 2017. 4 +[71] Heng Wang, Du Tran, Lorenzo Torresani, and Matt Feiszli. Video modeling with correlation networks. In CVPR, pages 352-361, 2020. 3 +[72] Junke Wang, Dongdong Chen, Chong Luo, Xiyang Dai, Lu Yuan, Zuxuan Wu, and Yu-Gang Jiang. Chatvideo: A tracklet-centric multimodal and versatile video understanding system. CoRR, abs/2304.14407, 2023. 3 +[73] Limin Wang, Yuanjun Xiong, Zhe Wang, Yu Qiao, Dahua Lin, Xiaou Tang, and Luc Van Gool. Temporal segment networks for action recognition in videos. TPAMI, 41(11): 2740-2755, 2018. 3 +[74] Limin Wang, Bingkun Huang, Zhiyu Zhao, Zhan Tong, Yi-nan He, Yi Wang, Yali Wang, and Yu Qiao. Videomae V2: scaling video masked autoencoders with dual masking. In CVPR, pages 14549-14560, 2023. 3 +[75] Mengmeng Wang, Jiazheng Xing, and Yong Liu. Actionclip: A new paradigm for video action recognition. arXiv preprint arXiv:2109.08472, 2021. 1, 2, 8 +[76] Peng Wang, An Yang, Rui Men, Junyang Lin, Shuai Bai, Zhikang Li, Jianxin Ma, Chang Zhou, Jingren Zhou, and Hongxia Yang. OFA: unifying architectures, tasks, and modalities through a simple sequence-to-sequence learning framework. In ICML, pages 23318-23340, 2022. 2 +[77] Rui Wang, Dongdong Chen, Zuxuan Wu, Yinpeng Chen, Xiyang Dai, Mengchen Liu, Yu-Gang Jiang, Luowei Zhou, and Lu Yuan. Bevt: Bert pretraining of video transformers. In CVPR, pages 14733-14743, 2022. 3 +[78] Syed Talal Wasim, Muzammal Naseer, Salman H. Khan, Fahad Shahbaz Khan, and Mubarak Shah. Vita-clip: Video and text adaptive CLIP via multimodal prompting. CVPR, 2023. 1, 2, 8 +[79] Chen Wei, Haoqi Fan, Saining Xie, Chao-Yuan Wu, Alan Yuille, and Christoph Feichtenhofer. Masked feature prediction for self-supervised visual pre-training. In CVPR, pages 14668-14678, 2022. 3 +[80] Wenhao Wu, Zhun Sun, and Wanli Ouyang. Revisiting classifier: Transferring vision-language models for video recognition. In AAAI, pages 2847-2855, 2023. 1 +[81] Hu Xu, Gargi Ghosh, Po-Yao Huang, Dmytro Okhonko, Armen Aghajanyan, Florian Metze, Luke Zettlemoyer, and Christoph Feichtenhofer. Videoclip: Contrastive pre-training for zero-shot video-text understanding. In EMNLP, pages 6787-6800, 2021. 2 + +[82] Jiarui Xu, Shalini De Mello, Sifei Liu, Wonmin Byeon, Thomas M. Breuel, Jan Kautz, and Xiaolong Wang. Groupvit: Semantic segmentation emerges from text supervision. In CVPR, pages 18113-18123, 2022. 2 +[83] An Yan, Yu Wang, Yiwu Zhong, Chengyu Dong, Zexue He, Yujie Lu, William Wang, Jingbo Shang, and Julian J. McAuley. Learning concise and descriptive attributes for visual recognition. ICCV, 2023. 3 +[84] Shen Yan, Xuehan Xiong, Anurag Arnab, Zhichao Lu, Mi Zhang, Chen Sun, and Cordelia Schmid. Multiview transformers for video recognition. In CVPR, pages 3333-3343, 2022. 7 +[85] Yue Yang, Artemis Panagopoulou, Shenghao Zhou, Daniel Jin, Chris Callison-Burch, and Mark Yatskar. Language in a bottle: Language model guided concept bottlenecks for interpretable image classification. In CVPR, pages 19187-19197, 2023. 3 +[86] Zhengyuan Yang, Linjie Li, Jianfeng Wang, Kevin Lin, Ehsan Azarnasab, Faisal Ahmed, Zicheng Liu, Ce Liu, Michael Zeng, and Lijuan Wang. MM-REACT: prompting chatgpt for multimodal reasoning and action. CoRR, abs/2303.11381, 2023. 3 +[87] Jiahui Yu, Zirui Wang, Vijay Vasudevan, Legg Yeung, Mojtaba Seyedhosseini, and Yonghui Wu. Coca: Contrastive captioners are image-text foundation models. arXiv preprint arXiv:2205.01917, 2022. 1, 2 +[88] Lu Yuan, Dongdong Chen, Yi-Ling Chen, Noel Codella, Xiyang Dai, Jianfeng Gao, Houdong Hu, Xuedong Huang, Boxin Li, Chunyuan Li, et al. Florence: A new foundation model for computer vision. arXiv preprint arXiv:2111.11432, 2021. 1, 2 +[89] Yue Zhao, Ishan Misra, Philipp Krahenbuhl, and Rohit Girdhar. Learning video representations from large language models. In CVPR, pages 6586-6597. IEEE, 2023. 3 +[90] Yiwu Zhong, Jianwei Yang, Pengchuan Zhang, Chunyuan Li, Noel Codella, Liunian Harold Li, Luowei Zhou, Xiyang Dai, Lu Yuan, Yin Li, and Jianfeng Gao. Regionclip: Region-based language-image pretraining. In CVPR, pages 16772-16782, 2022. 2 +[91] Deyao Zhu, Jun Chen, Xiaoqian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models. CoRR, abs/2304.10592, 2023. 3 \ No newline at end of file diff --git a/2024/Enhanced Motion-Text Alignment for Image-to-Video Transfer Learning/images.zip b/2024/Enhanced Motion-Text Alignment for Image-to-Video Transfer Learning/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..6fe1a21b217db221f230bb2885d6e4eebd729a71 --- /dev/null +++ b/2024/Enhanced Motion-Text Alignment for Image-to-Video Transfer Learning/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b55cd19f150ea4249b9314481936e3d97a3d1e17d46cbb63efcd0c076d88e071 +size 708319 diff --git a/2024/Enhanced Motion-Text Alignment for Image-to-Video Transfer Learning/layout.json b/2024/Enhanced Motion-Text Alignment for Image-to-Video Transfer Learning/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..91479dbc4eba915835e3a1f54f0792b914e0dccd --- /dev/null +++ b/2024/Enhanced Motion-Text Alignment for Image-to-Video Transfer Learning/layout.json @@ -0,0 +1,9548 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 72, + 103, + 522, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 72, + 103, + 522, + 121 + ], + "spans": [ + { + "bbox": [ + 72, + 103, + 522, + 121 + ], + "type": "text", + "content": "Enhanced Motion-Text Alignment for Image-to-Video Transfer Learning" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 88, + 142, + 529, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 88, + 142, + 529, + 157 + ], + "spans": [ + { + "bbox": [ + 88, + 142, + 529, + 157 + ], + "type": "text", + "content": "Wei Zhang" + }, + { + "bbox": [ + 88, + 142, + 529, + 157 + ], + "type": "inline_equation", + "content": "^{1,2*}" + }, + { + "bbox": [ + 88, + 142, + 529, + 157 + ], + "type": "text", + "content": ", Chaoqun Wan" + }, + { + "bbox": [ + 88, + 142, + 529, + 157 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 88, + 142, + 529, + 157 + ], + "type": "text", + "content": ", Tongliang Liu" + }, + { + "bbox": [ + 88, + 142, + 529, + 157 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 88, + 142, + 529, + 157 + ], + "type": "text", + "content": ", Xinmei Tian" + }, + { + "bbox": [ + 88, + 142, + 529, + 157 + ], + "type": "inline_equation", + "content": "^{1,4\\dagger}" + }, + { + "bbox": [ + 88, + 142, + 529, + 157 + ], + "type": "text", + "content": ", Xu Shen" + }, + { + "bbox": [ + 88, + 142, + 529, + 157 + ], + "type": "inline_equation", + "content": "^{2\\dagger}" + }, + { + "bbox": [ + 88, + 142, + 529, + 157 + ], + "type": "text", + "content": ", Jieping Ye" + }, + { + "bbox": [ + 88, + 142, + 529, + 157 + ], + "type": "inline_equation", + "content": "^{2}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 85, + 158, + 534, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 85, + 158, + 534, + 185 + ], + "spans": [ + { + "bbox": [ + 85, + 158, + 534, + 185 + ], + "type": "text", + "content": "1University of Science and Technology of China, 2Alibaba Cloud, 3The University of Sydney Institute of Artificial Intelligence, Hefei Comprehensive National Science Center" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 186, + 565, + 215 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 186, + 565, + 215 + ], + "spans": [ + { + "bbox": [ + 55, + 186, + 565, + 215 + ], + "type": "text", + "content": "zw0819@mail.ustc.edu.cn, tongliang.liu@sydney.edu.au, xinmei@ustc.edu.cn {qionlong.wcq, shenxu.sx, jieping.yjp}@alibaba-inc.com" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 143, + 241, + 192, + 254 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 241, + 192, + 254 + ], + "spans": [ + { + "bbox": [ + 143, + 241, + 192, + 254 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 266, + 290, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 266, + 290, + 567 + ], + "spans": [ + { + "bbox": [ + 47, + 266, + 290, + 567 + ], + "type": "text", + "content": "Extending large image-text pre-trained models (e.g., CLIP) for video understanding has made significant advancements. To enable the capability of CLIP to perceive dynamic information in videos, existing works are dedicated to equipping the visual encoder with various temporal modules. However, these methods exhibit \"asymmetry\" between the visual and textual sides, with neither temporal descriptions in input texts nor temporal modules in text encoder. This limitation hinders the potential of language supervision emphasized in CLIP, and restricts the learning of temporal features, as the text encoder has demonstrated limited proficiency in motion understanding. To address this issue, we propose leveraging \"MoTion-Enhanced Descriptions\" (MoTED) to facilitate the extraction of distinctive temporal features in videos. Specifically, we first generate discriminative motion-related descriptions via querying GPT-4 to compare easy-confusing action categories. Then, we incorporate both the visual and textual encoders with additional perception modules to process the video frames and generated descriptions, respectively. Finally, we adopt a contrastive loss to align the visual and textual motion features. Extensive experiments on five benchmarks show that MoTED surpasses state-of-the-art methods with convincing gaps, laying a solid foundation for empowering CLIP with strong temporal modeling." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 586, + 128, + 597 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 586, + 128, + 597 + ], + "spans": [ + { + "bbox": [ + 47, + 586, + 128, + 597 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 606, + 287, + 678 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 606, + 287, + 678 + ], + "spans": [ + { + "bbox": [ + 46, + 606, + 287, + 678 + ], + "type": "text", + "content": "Recent years have witnessed remarkable achievements in contrastive language-image pre-training models [9, 25, 36, 60, 87, 88], with CLIP [60] emerging as the front-runner. Through language supervision with a vast collection of 400 million image-text pairs, CLIP has achieved exceptional image comprehension and unprecedented zero-shot general-" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 310, + 240, + 545, + 407 + ], + "blocks": [ + { + "bbox": [ + 310, + 240, + 545, + 407 + ], + "lines": [ + { + "bbox": [ + 310, + 240, + 545, + 407 + ], + "spans": [ + { + "bbox": [ + 310, + 240, + 545, + 407 + ], + "type": "image", + "image_path": "fdb90797dde85bb2abcae4d1dd21b050d8b4d4fa3577438c3b73d612ad573c49.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 375, + 409, + 476, + 418 + ], + "lines": [ + { + "bbox": [ + 375, + 409, + 476, + 418 + ], + "spans": [ + { + "bbox": [ + 375, + 409, + 476, + 418 + ], + "type": "text", + "content": "(c) Motion-Enhanced Descriptions" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 305, + 431, + 547, + 509 + ], + "lines": [ + { + "bbox": [ + 305, + 431, + 547, + 509 + ], + "spans": [ + { + "bbox": [ + 305, + 431, + 547, + 509 + ], + "type": "text", + "content": "Figure 1. (a) Previous methods [27, 49, 75, 78] align the visual embeddings with the textual embeddings derived from category names. (b) Our work symmetrically aligns the spatial/temporal visual embeddings with class descriptions and motion descriptions correspondingly. (c) The descriptions generated by large language modes (e.g., GPT-4 [50]) enhance the conceptual definitions and discriminative details of motions." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 532, + 545, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 532, + 545, + 604 + ], + "spans": [ + { + "bbox": [ + 304, + 532, + 545, + 604 + ], + "type": "text", + "content": "ization. This breakthrough has opened up new possibilities for leveraging the power of large-scale pre-trained models to comprehend videos. It has also introduced a new paradigm [26, 40, 49, 52, 58, 75, 80] that endows image-based CLIP to effectively perceive dynamic information for video recognition." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 605, + 546, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 546, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 546, + 714 + ], + "type": "text", + "content": "To equip CLIP with motion perception, existing works propose to incorporate various temporal modules into the visual encoder. These temporal modules are additional tunable parameters, including designed efficient units between pre-trained transformer blocks [49, 75], and parallel structures to disentangle the temporal modeling out of the spatial modeling [40, 58]. However, while considerable effort has been put into capturing temporal visual features from videos, these methods have paid little attention to the in" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 683, + 287, + 702 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 683, + 287, + 702 + ], + "spans": [ + { + "bbox": [ + 47, + 683, + 287, + 702 + ], + "type": "text", + "content": "*This work was done when the author was visiting Alibaba as a research intern." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 59, + 703, + 135, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 703, + 135, + 712 + ], + "spans": [ + { + "bbox": [ + 59, + 703, + 135, + 712 + ], + "type": "text", + "content": "† Corresponding author." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "18504" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 277 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 277 + ], + "type": "text", + "content": "put text and text encoder, leading to an imbalance between the visual and textual alignment. When transferring to the classification task of video recognition (Fig. 1(a)), the only available text is the \"category names\" of the actions, e.g., \"clapping\" and \"slapping\". These coarse-grained word-level descriptors lack clear descriptions and explanations, making it difficult to distinguish between them. For instance, \"slapping\" refers to a quick hand movement striking against something, while \"clapping\" refers to the act of striking both hands together rapidly. The scarcity of textual motion information contradicts the language supervision principle of CLIP. Furthermore, it's observed that the CLIP text encoder delivers a strong bias towards spatial concepts (e.g., nouns) [5, 30, 47, 63], with a weak understanding of temporal cues (e.g., verbs). These issues significantly limit the effectiveness of extending CLIP with motion modeling for video understanding." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 279, + 289, + 507 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 279, + 289, + 507 + ], + "spans": [ + { + "bbox": [ + 46, + 279, + 289, + 507 + ], + "type": "text", + "content": "To overcome the limitations of language supervision on temporal modeling, we propose a novel approach that symmetrically aligns the spatial/temporal visual embeddings with class descriptions (\"slapping\") and motion descriptions correspondingly (\"a quick hand movement striking against something\"). We introduce the MoTion-Enhanced Descriptions (MoTED) as the language supervision for the visual temporal modules. In the implementation, we encounter two main challenges: i) how to generate motion-related descriptions? ii) how to effectively utilize descriptions as supervision? To tackle the first challenge, we aim for clear, detailed, and distinctive descriptions in the text. With the recent advancements in large language models (LLMs) [4, 11, 50], it has become possible to automatically generate conceptual descriptions for corresponding actions. As for the second challenge, we disentangle the spatial and temporal learning by constructing a temporal encoder, in parallel to the CLIP image encoder, to extract the temporal features aligned with motion-enhanced descriptions." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 510, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 510, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 510, + 289, + 715 + ], + "type": "text", + "content": "To be specific, we employ GPT-4 [50] to generate descriptions by posing a query: “Q: what is the motion concept of ? A:” (as shown in Fig. 1(c)). However, our empirical findings indicate that only the concept descriptions offer marginal assistance in motion perception, particularly on datasets such as Something-Something-V2 (SSv2) [20]) that necessitate the differentiation of similar classes using intricate motion cues. To mitigate this issue, we propose to incorporate related top-k actions as context for the LLMs to generate more discriminative descriptions. These generated texts are then passed through the CLIP text encoder to extract textual motion representations, where an adapter is utilized to eliminate the biases in the original model. Correspondingly, the temporal module in the visual encoder processes the video to extract dynamic features and aligns them with the generated textual descriptions through contrastive learning. Finally, these two sets of" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 304, + 72, + 545, + 179 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 179 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 179 + ], + "type": "text", + "content": "features are independently fused with the original CLIP's text and image features using cross-attention, and subsequently aligned. Evaluated on two supervised video recognition benchmarks, i.e., Kinetics-400 [28] and SSv2 [20], as well as three zero-shot benchmarks, i.e., Kinetics-600 [12], HMDB51 [24], UCF101 [65], the proposed MoTED surpasses state-of-the-art methods with convincing gaps, indicating the effectiveness of aligning enhanced motion descriptions with the temporal embedding of input videos." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 317, + 180, + 495, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 180, + 495, + 191 + ], + "spans": [ + { + "bbox": [ + 317, + 180, + 495, + 191 + ], + "type": "text", + "content": "We summarize the contributions as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 306, + 192, + 545, + 347 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 306, + 192, + 545, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 192, + 545, + 251 + ], + "spans": [ + { + "bbox": [ + 306, + 192, + 545, + 251 + ], + "type": "text", + "content": "- We present a new perspective that underscores the significance of textual side on par with visual side. By delving into in-depth distinguishing descriptions of actions, we make the first attempt to reveal the potential of language supervision as emphasized in CLIP." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 251, + 545, + 299 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 251, + 545, + 299 + ], + "spans": [ + { + "bbox": [ + 306, + 251, + 545, + 299 + ], + "type": "text", + "content": "- We propose MoTED that leverages LLMs to automatically generate action descriptions, and mining distinctive descriptions among similar actions. Then, we use a parallel path for both visual and textual motion modeling." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 300, + 545, + 347 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 300, + 545, + 347 + ], + "spans": [ + { + "bbox": [ + 306, + 300, + 545, + 347 + ], + "type": "text", + "content": "- We evaluate our approach on supervised as well as generalization tasks. Extensive experiments demonstrate the superiority and good generalization ability of the proposed method." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 358, + 392, + 371 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 358, + 392, + 371 + ], + "spans": [ + { + "bbox": [ + 306, + 358, + 392, + 371 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 378, + 545, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 378, + 545, + 653 + ], + "spans": [ + { + "bbox": [ + 304, + 378, + 545, + 653 + ], + "type": "text", + "content": "Vision-Language Pre-training. In recent years, Vision-Language Pre-training (VLP) [9, 25, 36, 45, 46, 60, 76, 87, 88] has made remarkable progress. One of the most remarkable and influential works is CLIP [60], which adopts the contrastive language-image pretraining paradigm. Following that, this paradigm has shown impressive zero-shot generalization capabilities on various image-related tasks [31, 35, 37, 43, 82, 90]. However, pre-training a language-video model [34, 81] is prohibitively expensive, as it requires large-scale video-text data and extensive training resources (e.g., thousands of GPU days). Meanwhile, transferring language-image pre-trained models to the video domain [10, 26, 40, 49, 52, 58, 75, 78] has captured significant attention due to its striking performance and training efficiency. For instance, X-CLIP [49] integrates the CLIP image encoder with a cross-frame attention module for temporal modeling. Vita-CLIP [78] utilizes multi-modal prompting techniques to learn video and text-specific context vectors. DiST [58] is the most related work that also disentangles spatial and temporal learning in the visual side. In contrast, our MoTED highlights the effectiveness of temporal disentangling both in the visual and textual sides, supervised by the detailed descriptions of motions." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 654, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 547, + 715 + ], + "type": "text", + "content": "Video Recognition. The conventional approaches in video recognition primarily focus on spatio-temporal learning under fully-supervised settings, where all categories are predefined. These approaches have achieved remarkable performance using various architectures, including convolution" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18505" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 53, + 70, + 547, + 243 + ], + "blocks": [ + { + "bbox": [ + 53, + 70, + 547, + 243 + ], + "lines": [ + { + "bbox": [ + 53, + 70, + 547, + 243 + ], + "spans": [ + { + "bbox": [ + 53, + 70, + 547, + 243 + ], + "type": "image", + "image_path": "c777a177803b74fa4a8ecf60d1633d0ccd35e865c08bb75d94524f870af32ca1.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 254, + 547, + 310 + ], + "lines": [ + { + "bbox": [ + 46, + 254, + 547, + 310 + ], + "spans": [ + { + "bbox": [ + 46, + 254, + 547, + 310 + ], + "type": "text", + "content": "Figure 2. The overall framework of MoTED. Building upon the dual image and text encoders initialized by CLIP [60], we extend the capability to perceive motion information symmetrically in both sides. In the text side, the motion-related conceptual and discriminative descriptions are generated via querying LLMs (e.g., GPT-4) using category names as input. In the vision side, a motion encoder is built to extract temporal dynamics given densely-sampled frames and integrated with the middle-layer features from image encoder. To align the motion embeddings and video embeddings from both sides, it employs two contrastive losses, respectively." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 48, + 328, + 291, + 426 + ], + "blocks": [ + { + "bbox": [ + 48, + 328, + 291, + 426 + ], + "lines": [ + { + "bbox": [ + 48, + 328, + 291, + 426 + ], + "spans": [ + { + "bbox": [ + 48, + 328, + 291, + 426 + ], + "type": "table", + "html": "
Related workTaskAction annotationExample
Dist. Sup. [39]pretrainevent description‘insert window mounting bolts’
Weakly Sup. [14]pretrainevent description‘take up the iron clamp’
LAVILA [89]pretrainevent description‘a lady walks past a car’
VFC [47]VL transfer learningevent description‘two brown horses eating grass’
MoTED (Ours)VL transfer learningmotion concept description‘slapping: a quick hand movement striking against something’
", + "image_path": "2f381b127262b50e7eb47ac0abfb3ef16675d4a7770d19462f06583fe810e3a5.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 474, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 474, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 474, + 289, + 715 + ], + "type": "text", + "content": "networks [7, 17, 21, 59, 64, 67–69, 71, 73], and vision transformers [1, 2, 6, 16, 32, 33, 38, 42, 48, 54, 55, 62]. In addition to the architecture design, self-supervised video representation learning [13, 18, 19, 23, 53, 57, 66, 74, 77, 79] has also gained popularity recently. However, these methods operate purely within the visual domain and meet bottlenecks of recognizing unseen or unfamiliar categories in real-world applications. Fortunately, the advance of large language models (LLMs) [4, 51] provides opportunities to mitigate this issue, due to their powerful capabilities of encoding world knowledge [29]. Recent studies [44, 56, 83, 85] have verified that factual sentences generated by LLMs can improve zero-shot image recognition accuracy. VFC [47] leverages PaLM [11] to create verb-focused hard negatives to enhance the understanding of verbs in video models. LSS [61] integrates language-based action concepts with self-supervised learning to adapt an image model to video domain. Our MoTED employs the descriptions that delineate the labeled classes and further accentuate the useful traits distinguishing similar classes." + } + ] + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 307, + 328, + 550, + 429 + ], + "blocks": [ + { + "bbox": [ + 47, + 430, + 288, + 452 + ], + "lines": [ + { + "bbox": [ + 47, + 430, + 288, + 452 + ], + "spans": [ + { + "bbox": [ + 47, + 430, + 288, + 452 + ], + "type": "text", + "content": "Table 1. Comparison of our method with related works focused on action annotations. 'VL' is the abbreviation of 'vision-language'." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 307, + 328, + 550, + 429 + ], + "lines": [ + { + "bbox": [ + 307, + 328, + 550, + 429 + ], + "spans": [ + { + "bbox": [ + 307, + 328, + 550, + 429 + ], + "type": "table", + "html": "
Related workLLMs roleExample
Chatvideo [72]manager& summarizeInput: 'Summarize the activity in video' \nOutput: 'A person is cooking in the kitchen'
MM-REACT [86]execution& summarizeInput: 'Please create a summary of the video' \nOutput: 'The speaker is making a BLT sandwich ...'
MiniGPT4 [91]/LLAVA [41]KB& data cleanInput: 'Describe this image in detail.' \nOutput: 'The image shows a group of musicians ...'
MoTED (Ours)generate descriptionsInput: 'What is the motion concept of slapping' \nOutput: 'A quick hand movement striking against ...'
", + "image_path": "99f02099f7781b12232bad5334b13f7743bfc3bcda7809c4c8383677e7b3a374.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 305, + 432, + 547, + 455 + ], + "lines": [ + { + "bbox": [ + 305, + 432, + 547, + 455 + ], + "spans": [ + { + "bbox": [ + 305, + 432, + 547, + 455 + ], + "type": "text", + "content": "Table 2. Comparison of our method with related works that take LLMs as a knowledge base (KB) and automatically annotated tool." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 306, + 469, + 361, + 482 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 469, + 361, + 482 + ], + "spans": [ + { + "bbox": [ + 306, + 469, + 361, + 482 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 491, + 545, + 634 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 491, + 545, + 634 + ], + "spans": [ + { + "bbox": [ + 304, + 491, + 545, + 634 + ], + "type": "text", + "content": "The generic pipeline of MoTED is presented in Fig. 2, which consists of three steps: (1) Motion Description Generation to obtain conceptual descriptions " + }, + { + "bbox": [ + 304, + 491, + 545, + 634 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{con}" + }, + { + "bbox": [ + 304, + 491, + 545, + 634 + ], + "type": "text", + "content": " and discriminative descriptions " + }, + { + "bbox": [ + 304, + 491, + 545, + 634 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{dis}" + }, + { + "bbox": [ + 304, + 491, + 545, + 634 + ], + "type": "text", + "content": ", via querying language models given the category names of the target video dataset in Sec. 3.1; (2) Textual Motion Adaptation to obtain the text embeddings that extract the motion semantics given the generated descriptions " + }, + { + "bbox": [ + 304, + 491, + 545, + 634 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{con}" + }, + { + "bbox": [ + 304, + 491, + 545, + 634 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 491, + 545, + 634 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{dis}" + }, + { + "bbox": [ + 304, + 491, + 545, + 634 + ], + "type": "text", + "content": " in Sec. 3.2; (3) Visual Motion Extraction to obtain the visual embeddings that extract the motion semantics given input video frames in Sec. 3.3. Overall, the framework is trained in an end-to-end contrastive manner as illustrated in Sec. 3.4." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 646, + 476, + 658 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 646, + 476, + 658 + ], + "spans": [ + { + "bbox": [ + 306, + 646, + 476, + 658 + ], + "type": "text", + "content": "3.1. Motion Description Generation" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 665, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 547, + 715 + ], + "type": "text", + "content": "The first step is to obtain a set of appropriate descriptions for each category. Given a video dataset with multiple different categories (e.g., Kinetics400 [28] with 400 action classes), the descriptions are generated automati" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "18506" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 240 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 240 + ], + "type": "text", + "content": "cally utilizing GPT-4 [50]. Note that, the generation process is agnostic to this choice and other LLMs can be used instead. For each category name, we query GPT-4 to provide the motion concepts using the following prompt: “Q: What is the motion concept in a video of ?” A: As shown in Fig. 1, the generated concept descriptions often cover moving objects, object interactions, moving directions and speeds, etc. But the output of LLMs can also be open-ended duplicates or anything in natural language (as illustrated in the Appendix). To control the generated descriptions to be concise and motion-related, we adopt a two-shot prompt in which we include two exemplars of question-answer for the same operation that is being queried." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 240, + 289, + 371 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 240, + 289, + 371 + ], + "spans": [ + { + "bbox": [ + 46, + 240, + 289, + 371 + ], + "type": "text", + "content": "Besides the conceptual descriptions " + }, + { + "bbox": [ + 46, + 240, + 289, + 371 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{con}" + }, + { + "bbox": [ + 46, + 240, + 289, + 371 + ], + "type": "text", + "content": ", we further investigate into the confusing actions that are characterized by their distinct differences from similar categories. To this end, we first compute the cosine distance of text embeddings for every two categories. Then we select top-k similar classes (e.g., " + }, + { + "bbox": [ + 46, + 240, + 289, + 371 + ], + "type": "inline_equation", + "content": "k = 5" + }, + { + "bbox": [ + 46, + 240, + 289, + 371 + ], + "type": "text", + "content": ") and query GPT-4 to generate motion characteristics for distinguishing the similar classes using the following prompt: \"Q: What are the useful features for distinguishing the original class from the similar class " + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 372, + 287, + 420 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 372, + 287, + 420 + ], + "spans": [ + { + "bbox": [ + 47, + 372, + 287, + 420 + ], + "type": "text", + "content": "A:\" Different from the conceptual descriptions " + }, + { + "bbox": [ + 47, + 372, + 287, + 420 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{con}" + }, + { + "bbox": [ + 47, + 372, + 287, + 420 + ], + "type": "text", + "content": " that are mainly determined by the general knowledge of LLMs, the descriptions " + }, + { + "bbox": [ + 47, + 372, + 287, + 420 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{dis}" + }, + { + "bbox": [ + 47, + 372, + 287, + 420 + ], + "type": "text", + "content": " further contain task-specific information that the downstream task emphasizes." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 421, + 287, + 480 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 421, + 287, + 480 + ], + "spans": [ + { + "bbox": [ + 46, + 421, + 287, + 480 + ], + "type": "text", + "content": "In this way, given a dataset of " + }, + { + "bbox": [ + 46, + 421, + 287, + 480 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 421, + 287, + 480 + ], + "type": "text", + "content": " categories, we can obtain the motion-enhanced " + }, + { + "bbox": [ + 46, + 421, + 287, + 480 + ], + "type": "inline_equation", + "content": "N * (k + 1)" + }, + { + "bbox": [ + 46, + 421, + 287, + 480 + ], + "type": "text", + "content": " descriptions, consisting of 1 conceptual description and " + }, + { + "bbox": [ + 46, + 421, + 287, + 480 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 421, + 287, + 480 + ], + "type": "text", + "content": " discriminative descriptions for each category. The above generation process is completed before the training of whole framework." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 481, + 288, + 637 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 481, + 288, + 637 + ], + "spans": [ + { + "bbox": [ + 46, + 481, + 288, + 637 + ], + "type": "text", + "content": "The detailed comparisons of action annotations and LLM-based generations are presented in Tab. 1 and Tab. 2, respectively. Tab. 1 shows that existing works employ annotations that describe the action events in a subject-verb-object manner, which are object-centric. In contrast, our generated annotations are motion-related, object-agnostic and to supervise the learning of visual motion features with the world knowledge of motion concepts. In Tab. 2, LLMs are mainly applied in generative tasks, to summarize query results and generate general responses, or to obtain instruction data to align with human preference. In contrast, we apply LLMs to obtain descriptions of motion concept, aligned with the temporal visual features." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 646, + 198, + 659 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 646, + 198, + 659 + ], + "spans": [ + { + "bbox": [ + 47, + 646, + 198, + 659 + ], + "type": "text", + "content": "3.2. Textual Motion Adaptation" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 665, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 288, + 714 + ], + "type": "text", + "content": "After generating the motion-enhanced descriptions, the second step is to perceive the motion cues within the descriptions and aggregate them into a compact motion-enhanced embedding. Given a set of motion-enhanced descriptions " + }, + { + "bbox": [ + 46, + 665, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 70, + 545, + 172 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 70, + 545, + 172 + ], + "spans": [ + { + "bbox": [ + 304, + 70, + 545, + 172 + ], + "type": "inline_equation", + "content": "= \\left\\{\\mathcal{D}_{con}^{i},\\mathcal{D}_{dis}^{i,j}\\right\\}" + }, + { + "bbox": [ + 304, + 70, + 545, + 172 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 304, + 70, + 545, + 172 + ], + "type": "inline_equation", + "content": "i\\in [1,N]" + }, + { + "bbox": [ + 304, + 70, + 545, + 172 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 70, + 545, + 172 + ], + "type": "inline_equation", + "content": "j\\in [1,k]" + }, + { + "bbox": [ + 304, + 70, + 545, + 172 + ], + "type": "text", + "content": ", we extract normalized feature embedding " + }, + { + "bbox": [ + 304, + 70, + 545, + 172 + ], + "type": "inline_equation", + "content": "e_i" + }, + { + "bbox": [ + 304, + 70, + 545, + 172 + ], + "type": "text", + "content": " by using the text encoder " + }, + { + "bbox": [ + 304, + 70, + 545, + 172 + ], + "type": "inline_equation", + "content": "f_{txt}" + }, + { + "bbox": [ + 304, + 70, + 545, + 172 + ], + "type": "text", + "content": ": " + }, + { + "bbox": [ + 304, + 70, + 545, + 172 + ], + "type": "inline_equation", + "content": "e_i = f_{txt}(d_i)" + }, + { + "bbox": [ + 304, + 70, + 545, + 172 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 70, + 545, + 172 + ], + "type": "inline_equation", + "content": "d_{i}" + }, + { + "bbox": [ + 304, + 70, + 545, + 172 + ], + "type": "text", + "content": " is a motion-enhanced description sampled from " + }, + { + "bbox": [ + 304, + 70, + 545, + 172 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 304, + 70, + 545, + 172 + ], + "type": "text", + "content": ". In this way, we obtain the text embeddings " + }, + { + "bbox": [ + 304, + 70, + 545, + 172 + ], + "type": "inline_equation", + "content": "E_{txt}\\in \\mathbb{R}^{N\\times (k + 1)\\times C}" + }, + { + "bbox": [ + 304, + 70, + 545, + 172 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 70, + 545, + 172 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 304, + 70, + 545, + 172 + ], + "type": "text", + "content": " denotes the channel number of each embedding. Noticeably, the parameters of text encoder is frozen and initialized by CLIP [60] to inherit its capacity of encoding visual-aligned semantics." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 172, + 546, + 316 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 172, + 546, + 316 + ], + "spans": [ + { + "bbox": [ + 304, + 172, + 546, + 316 + ], + "type": "text", + "content": "However, as CLIP is pre-trained on image-text paired corpus, the text encoder has a strong bias towards spatial appearance of objects and backgrounds, instead of temporal motions [5, 30, 47, 63]. To adapt the text encoder to understand the motion-enhanced descriptions, we introduce a Text Motion Adapter, which consists of a multi-head self-attention (MHSA) [70]. The adapter is parameter-tunable and takes " + }, + { + "bbox": [ + 304, + 172, + 546, + 316 + ], + "type": "inline_equation", + "content": "E_{txt}" + }, + { + "bbox": [ + 304, + 172, + 546, + 316 + ], + "type": "text", + "content": " as input to learn the information dependencies between motion-enhanced descriptions. To aggregate the motion semantics for each category, we perform adaptive pooling to obtain the motion-enhanced embeddings " + }, + { + "bbox": [ + 304, + 172, + 546, + 316 + ], + "type": "inline_equation", + "content": "E_{txt}^{m} \\in \\mathbb{R}^{N \\times 1 \\times C}" + }, + { + "bbox": [ + 304, + 172, + 546, + 316 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 321, + 448, + 333 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 321, + 448, + 333 + ], + "spans": [ + { + "bbox": [ + 305, + 321, + 448, + 333 + ], + "type": "text", + "content": "3.3. Visual Motion Extraction" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 340, + 545, + 424 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 340, + 545, + 424 + ], + "spans": [ + { + "bbox": [ + 304, + 340, + 545, + 424 + ], + "type": "text", + "content": "In the visual side, as shown in Fig. 3, it comprises of two components: (i) The image encoder adopts the CLIP pre-trained Vision Transformer (ViT), which extracts frozen features for sparse frames with powerful spatial semantics. (ii) The motion encoder takes dense frames as input to capture the local temporal cues, integrated with the global temporal dynamics from middle-layer image features." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 437, + 405, + 449 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 437, + 405, + 449 + ], + "spans": [ + { + "bbox": [ + 305, + 437, + 405, + 449 + ], + "type": "text", + "content": "3.3.1 Image Encoder" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 455, + 545, + 625 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 455, + 545, + 625 + ], + "spans": [ + { + "bbox": [ + 304, + 455, + 545, + 625 + ], + "type": "text", + "content": "Given a video clip " + }, + { + "bbox": [ + 304, + 455, + 545, + 625 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_S\\in \\mathbb{R}^{T\\times H\\times W\\times 3}" + }, + { + "bbox": [ + 304, + 455, + 545, + 625 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 304, + 455, + 545, + 625 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 304, + 455, + 545, + 625 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 455, + 545, + 625 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 304, + 455, + 545, + 625 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 455, + 545, + 625 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 304, + 455, + 545, + 625 + ], + "type": "text", + "content": " represent the frame number, height, and width, respectively), the image features are extracted individually for several sparse frames. Following ViT [15], each frame is divided into " + }, + { + "bbox": [ + 304, + 455, + 545, + 625 + ], + "type": "inline_equation", + "content": "K = \\frac{H}{P}\\times \\frac{W}{P}" + }, + { + "bbox": [ + 304, + 455, + 545, + 625 + ], + "type": "text", + "content": " patches, and the size of each patch is denoted as " + }, + { + "bbox": [ + 304, + 455, + 545, + 625 + ], + "type": "inline_equation", + "content": "P\\times P" + }, + { + "bbox": [ + 304, + 455, + 545, + 625 + ], + "type": "text", + "content": ". These patches are then projected using a fully connected layer, referred to as the 2D stem in Fig. 3. This projection generates a sequence of patch embeddings " + }, + { + "bbox": [ + 304, + 455, + 545, + 625 + ], + "type": "inline_equation", + "content": "[\\mathbf{x}_{t,\\mathrm{cls}}^{(0)},\\mathbf{x}_{t,1}^{(0)},\\mathbf{x}_{t,2}^{(0)},\\dots ,\\mathbf{x}_{t,K}^{(0)}] + \\mathbf{e}^{\\mathrm{spatial}}" + }, + { + "bbox": [ + 304, + 455, + 545, + 625 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 455, + 545, + 625 + ], + "type": "inline_equation", + "content": "t = \\{1,\\dots ,T\\}" + }, + { + "bbox": [ + 304, + 455, + 545, + 625 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 455, + 545, + 625 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{\\mathrm{cls}}" + }, + { + "bbox": [ + 304, + 455, + 545, + 625 + ], + "type": "text", + "content": " is an additional learnable token (termed as \"class embedding\"), and " + }, + { + "bbox": [ + 304, + 455, + 545, + 625 + ], + "type": "inline_equation", + "content": "\\mathbf{e}^{\\mathrm{spatial}}" + }, + { + "bbox": [ + 304, + 455, + 545, + 625 + ], + "type": "text", + "content": " denotes spatial position embedding. Assuming that the spatial encoder has " + }, + { + "bbox": [ + 304, + 455, + 545, + 625 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 304, + 455, + 545, + 625 + ], + "type": "text", + "content": " Transformer blocks, the features of the " + }, + { + "bbox": [ + 304, + 455, + 545, + 625 + ], + "type": "inline_equation", + "content": "l_{\\mathrm{th}}" + }, + { + "bbox": [ + 304, + 455, + 545, + 625 + ], + "type": "text", + "content": " layer for the " + }, + { + "bbox": [ + 304, + 455, + 545, + 625 + ], + "type": "inline_equation", + "content": "t_\\mathrm{th}" + }, + { + "bbox": [ + 304, + 455, + 545, + 625 + ], + "type": "text", + "content": " frame can be extracted by:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 331, + 630, + 545, + 647 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 331, + 630, + 545, + 647 + ], + "spans": [ + { + "bbox": [ + 331, + 630, + 545, + 647 + ], + "type": "interline_equation", + "content": "\\mathbf {X} _ {t} ^ {(l)} = \\operatorname {T r a n s f o r m e r} ^ {(l)} \\left(\\mathbf {X} _ {t} ^ {(l - 1)}\\right) \\in \\mathbb {R} ^ {(K + 1) \\times C}, \\tag {1}", + "image_path": "0ff47b239084c34a9867b6890029e360415359e7e27724c1faf029e0ec543b85.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 651, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 651, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 651, + 545, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 651, + 545, + 714 + ], + "type": "inline_equation", + "content": "l = \\{1, \\dots, L\\}" + }, + { + "bbox": [ + 304, + 651, + 545, + 714 + ], + "type": "text", + "content": " denotes the layer index. The class embeddings of " + }, + { + "bbox": [ + 304, + 651, + 545, + 714 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 304, + 651, + 545, + 714 + ], + "type": "text", + "content": " frames in the " + }, + { + "bbox": [ + 304, + 651, + 545, + 714 + ], + "type": "inline_equation", + "content": "l_{\\mathrm{th}}" + }, + { + "bbox": [ + 304, + 651, + 545, + 714 + ], + "type": "text", + "content": " layer is termed as " + }, + { + "bbox": [ + 304, + 651, + 545, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{X}^{(l)} = [\\mathbf{X}_1^{(l)}, \\dots, \\mathbf{X}_T^{(l)}] \\in \\mathbb{R}^{T \\times 1 \\times C}" + }, + { + "bbox": [ + 304, + 651, + 545, + 714 + ], + "type": "text", + "content": ". Benefit from the CLIP pretrained parameters, the class embeddings " + }, + { + "bbox": [ + 304, + 651, + 545, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{X}^{(L)}" + }, + { + "bbox": [ + 304, + 651, + 545, + 714 + ], + "type": "text", + "content": " aggregate powerful spatial semantics within each frame." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18507" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 51, + 72, + 286, + 223 + ], + "blocks": [ + { + "bbox": [ + 51, + 72, + 286, + 223 + ], + "lines": [ + { + "bbox": [ + 51, + 72, + 286, + 223 + ], + "spans": [ + { + "bbox": [ + 51, + 72, + 286, + 223 + ], + "type": "image", + "image_path": "e322011f8dfecfe46dbb4134236f27bde2cc8040bd79533ffd81132398abc644.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 56, + 234, + 277, + 245 + ], + "lines": [ + { + "bbox": [ + 56, + 234, + 277, + 245 + ], + "spans": [ + { + "bbox": [ + 56, + 234, + 277, + 245 + ], + "type": "text", + "content": "Figure 3. The structural details of MoTED in the vision side." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 267, + 150, + 278 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 267, + 150, + 278 + ], + "spans": [ + { + "bbox": [ + 47, + 267, + 150, + 278 + ], + "type": "text", + "content": "3.3.2 Motion Encoder" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 286, + 289, + 454 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 286, + 289, + 454 + ], + "spans": [ + { + "bbox": [ + 46, + 286, + 289, + 454 + ], + "type": "text", + "content": "To fully extract the motion information in videos, the temporal input " + }, + { + "bbox": [ + 46, + 286, + 289, + 454 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_T\\in \\mathbb{R}^{\\alpha T\\times H\\times W\\times 3}" + }, + { + "bbox": [ + 46, + 286, + 289, + 454 + ], + "type": "text", + "content": " for the motion encoder is sampled around the spatial input " + }, + { + "bbox": [ + 46, + 286, + 289, + 454 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_S" + }, + { + "bbox": [ + 46, + 286, + 289, + 454 + ], + "type": "text", + "content": " by " + }, + { + "bbox": [ + 46, + 286, + 289, + 454 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 46, + 286, + 289, + 454 + ], + "type": "text", + "content": " times. In this study, we set " + }, + { + "bbox": [ + 46, + 286, + 289, + 454 + ], + "type": "inline_equation", + "content": "\\alpha = 2" + }, + { + "bbox": [ + 46, + 286, + 289, + 454 + ], + "type": "text", + "content": " by default, as empirically studied in [17, 58]. Then, " + }, + { + "bbox": [ + 46, + 286, + 289, + 454 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_T" + }, + { + "bbox": [ + 46, + 286, + 289, + 454 + ], + "type": "text", + "content": " is projected by a 3D convolution, i.e., the 3D stem in Fig. 3, for patch embedding. The kernel size and stride of the 3D stem are both " + }, + { + "bbox": [ + 46, + 286, + 289, + 454 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 46, + 286, + 289, + 454 + ], + "type": "text", + "content": " in spatial dimension and " + }, + { + "bbox": [ + 46, + 286, + 289, + 454 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 46, + 286, + 289, + 454 + ], + "type": "text", + "content": " in temporal dimension. Thus the number of temporal patch tokens are the same with that in the image encoder, which makes it convenient to integrate the spatial and temporal features. Thus the projected temporal features can be formulated as: " + }, + { + "bbox": [ + 46, + 286, + 289, + 454 + ], + "type": "inline_equation", + "content": "\\mathbf{Z}^{(0)} = \\mathrm{Conv3d}(\\mathbf{X}_T)\\in \\mathbb{R}^{T\\times K\\times C}" + }, + { + "bbox": [ + 46, + 286, + 289, + 454 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 286, + 289, + 454 + ], + "type": "inline_equation", + "content": "K = \\frac{H}{P}\\times \\frac{W}{P}" + }, + { + "bbox": [ + 46, + 286, + 289, + 454 + ], + "type": "text", + "content": ". Then, a series of Temporal Blocks are designed to extract motion patterns, which can be written as:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 473, + 287, + 488 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 473, + 287, + 488 + ], + "spans": [ + { + "bbox": [ + 54, + 473, + 287, + 488 + ], + "type": "interline_equation", + "content": "\\mathbf {Z} ^ {(l)} = \\operatorname {T e m p - B l o c k} ^ {(l)} \\left(\\mathbf {Z} ^ {(l - 1)}, \\mathbf {X} ^ {(l - 1)}\\right) \\in \\mathbb {R} ^ {T \\times K \\times C}, \\tag {2}", + "image_path": "08b117b1c691c1f7c3a31a1185e24d1eee6a31270fce1550acd17c91e8956b30.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 498, + 287, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 498, + 287, + 521 + ], + "spans": [ + { + "bbox": [ + 46, + 498, + 287, + 521 + ], + "type": "text", + "content": "where the function Temp-Block " + }, + { + "bbox": [ + 46, + 498, + 287, + 521 + ], + "type": "inline_equation", + "content": "(\\cdot)" + }, + { + "bbox": [ + 46, + 498, + 287, + 521 + ], + "type": "text", + "content": " performs temporal modeling integrated with the spatial features " + }, + { + "bbox": [ + 46, + 498, + 287, + 521 + ], + "type": "inline_equation", + "content": "X^{(l - 1)}" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 522, + 289, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 522, + 289, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 522, + 289, + 714 + ], + "type": "text", + "content": "Following the effective and efficient design philosophy in [32, 33], the temporal block consists of the following three modules: (i) Local Motion Extraction (LME). Since 3D convolution can capture detailed and local spatiotemporal features, by processing each pixel with context from a small 3D neighborhood (e.g., " + }, + { + "bbox": [ + 46, + 522, + 289, + 714 + ], + "type": "inline_equation", + "content": "2 \\times 16 \\times 16" + }, + { + "bbox": [ + 46, + 522, + 289, + 714 + ], + "type": "text", + "content": "), the temporal patch tokens " + }, + { + "bbox": [ + 46, + 522, + 289, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{Z}^{(0)}" + }, + { + "bbox": [ + 46, + 522, + 289, + 714 + ], + "type": "text", + "content": " remain the nature of local motion extraction. Then these tokens are flattened to a sequence " + }, + { + "bbox": [ + 46, + 522, + 289, + 714 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^{TK \\times C}" + }, + { + "bbox": [ + 46, + 522, + 289, + 714 + ], + "type": "text", + "content": " and perform spatiotemporal learning via joint self-attention modules [66] and token merging for efficiency [3]. (ii) Global Motion Extraction (GME). Given the class embeddings " + }, + { + "bbox": [ + 46, + 522, + 289, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{X}^{(l)}" + }, + { + "bbox": [ + 46, + 522, + 289, + 714 + ], + "type": "text", + "content": " that aggregate powerful spatial semantics for each individual frame, we apply a cross-frame self-attention module [49] to learn the global inter-frame interactions. (iii) Feature Integration. After obtaining the local and global motion features, we adopt a cross-attention mod" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "text", + "content": "ule, which takes local features as query and global features as key/value, to complement the features for better temporal modeling. In this way, the output of the final temporal block " + }, + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "inline_equation", + "content": "\\mathbf{Z}^{(L)}" + }, + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "text", + "content": " aggregates powerful temporal dynamics within the dense frames." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 140, + 394, + 153 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 140, + 394, + 153 + ], + "spans": [ + { + "bbox": [ + 305, + 140, + 394, + 153 + ], + "type": "text", + "content": "3.4. Training Loss" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 159, + 545, + 209 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 159, + 545, + 209 + ], + "spans": [ + { + "bbox": [ + 304, + 159, + 545, + 209 + ], + "type": "text", + "content": "Following CLIP [60], we first apply adaptive pooling to obtain a video-level motion embedding " + }, + { + "bbox": [ + 304, + 159, + 545, + 209 + ], + "type": "inline_equation", + "content": "\\mathbf{Z}_{avg}^{(L)} \\in \\mathbb{R}^{1 \\times C}" + }, + { + "bbox": [ + 304, + 159, + 545, + 209 + ], + "type": "text", + "content": ". Then we introduce a contrastive loss to align the visual and textual motion embeddings:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 331, + 218, + 545, + 251 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 331, + 218, + 545, + 251 + ], + "spans": [ + { + "bbox": [ + 331, + 218, + 545, + 251 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {m o t i o n}} = - \\log \\frac {\\exp \\left(\\sin \\left(\\mathbf {Z} _ {\\text {a v g}} ^ {(L)} , \\mathbf {e} _ {i}\\right) / \\tau\\right)}{\\sum_ {c = 1} ^ {N} \\exp \\left(\\sin \\left(\\mathbf {Z} _ {\\text {a v g}} ^ {(L)} , \\mathbf {e} _ {c}\\right) / \\tau\\right)}, \\tag {3}", + "image_path": "afa6f15d89e88519381ed301e357042f3d9e97223fd73c0b1277812eaf2234af.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 258, + 545, + 295 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 258, + 545, + 295 + ], + "spans": [ + { + "bbox": [ + 304, + 258, + 545, + 295 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 258, + 545, + 295 + ], + "type": "inline_equation", + "content": "\\mathrm{sim}(\\cdot ,\\cdot)" + }, + { + "bbox": [ + 304, + 258, + 545, + 295 + ], + "type": "text", + "content": " is the normalized cosine similarity, " + }, + { + "bbox": [ + 304, + 258, + 545, + 295 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 304, + 258, + 545, + 295 + ], + "type": "text", + "content": " refers to the temperature parameter. " + }, + { + "bbox": [ + 304, + 258, + 545, + 295 + ], + "type": "inline_equation", + "content": "\\mathbf{e}_i" + }, + { + "bbox": [ + 304, + 258, + 545, + 295 + ], + "type": "text", + "content": " is a motion-enhanced text embedding for the " + }, + { + "bbox": [ + 304, + 258, + 545, + 295 + ], + "type": "inline_equation", + "content": "i_{\\mathrm{th}}" + }, + { + "bbox": [ + 304, + 258, + 545, + 295 + ], + "type": "text", + "content": " category." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 295, + 545, + 367 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 295, + 545, + 367 + ], + "spans": [ + { + "bbox": [ + 304, + 295, + 545, + 367 + ], + "type": "text", + "content": "Moreover, we adopt a cross-attention module to fuse the visual motion features (as key/value) and image features (as query) to obtain the compact video embedding " + }, + { + "bbox": [ + 304, + 295, + 545, + 367 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_i" + }, + { + "bbox": [ + 304, + 295, + 545, + 367 + ], + "type": "text", + "content": ". Similarly, the fused text embedding " + }, + { + "bbox": [ + 304, + 295, + 545, + 367 + ], + "type": "inline_equation", + "content": "\\mathbf{e}_i'" + }, + { + "bbox": [ + 304, + 295, + 545, + 367 + ], + "type": "text", + "content": " is acquired via a cross-attention module. We introduce a contrastive loss to align the visual and textual video-level embeddings:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 339, + 376, + 545, + 405 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 339, + 376, + 545, + 405 + ], + "spans": [ + { + "bbox": [ + 339, + 376, + 545, + 405 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {v i d e o}} = - \\log \\frac {\\exp \\left(\\operatorname {s i m} \\left(\\mathbf {v} _ {i} , \\mathbf {e} _ {i} ^ {\\prime}\\right) / \\tau\\right)}{\\sum_ {c = 1} ^ {N} \\exp \\left(\\operatorname {s i m} \\left(\\mathbf {v} _ {i} , \\mathbf {e} _ {c} ^ {\\prime}\\right) / \\tau\\right)}. \\tag {4}", + "image_path": "f48bd4627b4e2573ee0c87f5283d9225f07be5f6737a9dae08a609a2a836cddc.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 413, + 545, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 413, + 545, + 437 + ], + "spans": [ + { + "bbox": [ + 305, + 413, + 545, + 437 + ], + "type": "text", + "content": "Compatible with CLIP training paradigm, the overall framework is trained based on these two contrastive losses:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 376, + 451, + 545, + 463 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 376, + 451, + 545, + 463 + ], + "spans": [ + { + "bbox": [ + 376, + 451, + 545, + 463 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\mathcal {L} _ {\\text {m o t i o n}} + \\mathcal {L} _ {\\text {v i d e o}}. \\tag {5}", + "image_path": "68cacaa612ab238f57b623e7be1fb3a82e770356311f0f66141aa0b4f12adfaa.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 474, + 388, + 488 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 474, + 388, + 488 + ], + "spans": [ + { + "bbox": [ + 306, + 474, + 388, + 488 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 305, + 502, + 465, + 516 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 502, + 465, + 516 + ], + "spans": [ + { + "bbox": [ + 305, + 502, + 465, + 516 + ], + "type": "text", + "content": "4.1. Dataset and Implementation." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 521, + 545, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 521, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 304, + 521, + 545, + 689 + ], + "type": "text", + "content": "Datasets. In the supervised setting, we train on the train set of Kinetics-400 (K400) [28] and Something-Something-V2 (SSv2) [20] and report supervised performance against existing methods on the validation sets of K400 and SSv2. In the zero-shot setting, we train on the Kinetics-400 training set and evaluate on three datasets: Kinetics-600 (K600) [12], HMDB51 [24] and UCF101 [65]. For zero-shot evaluation on K600, following [12], we use the 220 new categories outside of K400 for evaluation, and conduct evaluation three times, each time randomly sampling 160 categories for evaluation from the 220 categories. For zero-shot evaluation on HMDB51 and UCF101, we follow [49] and report average top-1 accuracy and standard deviation on three splits of the test set." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "text", + "content": "Implementation Details. Following previous work [40, 58], we use the CLIP [60] pre-trained ViT-B/16, ViT-L/14" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18508" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 52, + 81, + 214, + 140 + ], + "blocks": [ + { + "bbox": [ + 52, + 81, + 214, + 140 + ], + "lines": [ + { + "bbox": [ + 52, + 81, + 214, + 140 + ], + "spans": [ + { + "bbox": [ + 52, + 81, + 214, + 140 + ], + "type": "table", + "html": "
Class NamesConcept.Discr.SSv2K400
✓XX65.782.9
✓✓X66.084.5
✓X✓69.984.0
✓✓✓70.185.1
", + "image_path": "2572135649c923680bf865fff5a1719aa4d2bcbb4b89b24b70a7daa5b2c3f68d.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 48, + 144, + 211, + 173 + ], + "lines": [ + { + "bbox": [ + 48, + 144, + 211, + 173 + ], + "spans": [ + { + "bbox": [ + 48, + 144, + 211, + 173 + ], + "type": "text", + "content": "(a) \"Concept.\" is the abbreviation of \"conceptual descriptions\". \"Discr.\" is the abbreviation of \"discriminative descriptions\"." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 241, + 70, + 356, + 131 + ], + "blocks": [ + { + "bbox": [ + 241, + 70, + 356, + 131 + ], + "lines": [ + { + "bbox": [ + 241, + 70, + 356, + 131 + ], + "spans": [ + { + "bbox": [ + 241, + 70, + 356, + 131 + ], + "type": "table", + "html": "
M.Enc.Adapter.SSv2K400
XX64.381.5
✓X68.683.4
X✓66.283.8
✓✓70.185.1
", + "image_path": "ae714a9a820bffee25b47662d333fc16c93f09a192b3cc5d4da707bac866c676.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 215, + 143, + 378, + 173 + ], + "lines": [ + { + "bbox": [ + 215, + 143, + 378, + 173 + ], + "spans": [ + { + "bbox": [ + 215, + 143, + 378, + 173 + ], + "type": "text", + "content": "(b) \"M.Enc.\" is the separated motion encoder where motion modules are parallel with CLIP visual encoder. \"Adapter\" is the text motion adapter." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "type": "table", + "bbox": [ + 395, + 71, + 535, + 131 + ], + "blocks": [ + { + "bbox": [ + 395, + 71, + 535, + 131 + ], + "lines": [ + { + "bbox": [ + 395, + 71, + 535, + 131 + ], + "spans": [ + { + "bbox": [ + 395, + 71, + 535, + 131 + ], + "type": "table", + "html": "
Text.Fuse.Vision.Fuse.SSv2K400
XX68.383.6
✓X68.984.2
X✓69.584.5
✓✓70.185.1
", + "image_path": "a6cc3563377997bd74c48c520c757f80c31036009299dc37a544ef7a7c5411fe.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 383, + 143, + 544, + 172 + ], + "lines": [ + { + "bbox": [ + 383, + 143, + 544, + 172 + ], + "spans": [ + { + "bbox": [ + 383, + 143, + 544, + 172 + ], + "type": "text", + "content": "(c) \"Text.Fuse.\" indicates feature fusion in the textual side. \"Vision.Fuse.\" indicates feature fusion in the visual side." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 177, + 546, + 200 + ], + "lines": [ + { + "bbox": [ + 46, + 177, + 546, + 200 + ], + "spans": [ + { + "bbox": [ + 46, + 177, + 546, + 200 + ], + "type": "text", + "content": "Table 3. Ablations on Something-Something-V2 and Kinetics-400. The spatial encoder is a 8-frame vanilla ViT-B/16 pre-trained by CLIP [60]. The inference protocol of all models and datasets are 3 clips × 1 center crop." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 48, + 211, + 293, + 259 + ], + "blocks": [ + { + "bbox": [ + 48, + 211, + 293, + 259 + ], + "lines": [ + { + "bbox": [ + 48, + 211, + 293, + 259 + ], + "spans": [ + { + "bbox": [ + 48, + 211, + 293, + 259 + ], + "type": "image", + "image_path": "b848dc7327e56cbfbc7307b0faf989346b6773822fa7005d57c405e14d0bbb5d.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 48, + 261, + 293, + 303 + ], + "blocks": [ + { + "bbox": [ + 48, + 261, + 293, + 303 + ], + "lines": [ + { + "bbox": [ + 48, + 261, + 293, + 303 + ], + "spans": [ + { + "bbox": [ + 48, + 261, + 293, + 303 + ], + "type": "image", + "image_path": "661bdcac5cfcf91f68631a263b297c87e0fb3d9b585739073bc9f1af0cc3a846.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 114, + 304, + 233, + 310 + ], + "lines": [ + { + "bbox": [ + 114, + 304, + 233, + 310 + ], + "spans": [ + { + "bbox": [ + 114, + 304, + 233, + 310 + ], + "type": "text", + "content": "Relevance between text and motion features" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 138, + 312, + 163, + 319 + ], + "lines": [ + { + "bbox": [ + 138, + 312, + 163, + 319 + ], + "spans": [ + { + "bbox": [ + 138, + 312, + 163, + 319 + ], + "type": "text", + "content": "(a) case 1" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 301, + 210, + 545, + 259 + ], + "blocks": [ + { + "bbox": [ + 301, + 210, + 545, + 259 + ], + "lines": [ + { + "bbox": [ + 301, + 210, + 545, + 259 + ], + "spans": [ + { + "bbox": [ + 301, + 210, + 545, + 259 + ], + "type": "image", + "image_path": "4aef931383e4376f9db7a3352bfcfbf88e3155f28f48a8ad5b5c84f4a2ce5541.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 301, + 261, + 545, + 303 + ], + "blocks": [ + { + "bbox": [ + 301, + 261, + 545, + 303 + ], + "lines": [ + { + "bbox": [ + 301, + 261, + 545, + 303 + ], + "spans": [ + { + "bbox": [ + 301, + 261, + 545, + 303 + ], + "type": "image", + "image_path": "1812d5779aff347aaf381d9b414dd1265f58885d7ae41061d6641f136837685f.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 356, + 303, + 474, + 309 + ], + "lines": [ + { + "bbox": [ + 356, + 303, + 474, + 309 + ], + "spans": [ + { + "bbox": [ + 356, + 303, + 474, + 309 + ], + "type": "text", + "content": "Relevance between text and motion features" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 402, + 312, + 428, + 319 + ], + "lines": [ + { + "bbox": [ + 402, + 312, + 428, + 319 + ], + "spans": [ + { + "bbox": [ + 402, + 312, + 428, + 319 + ], + "type": "text", + "content": "(b) case2" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 323, + 547, + 346 + ], + "lines": [ + { + "bbox": [ + 46, + 323, + 547, + 346 + ], + "spans": [ + { + "bbox": [ + 46, + 323, + 547, + 346 + ], + "type": "text", + "content": "Figure 4. Two cases to visualize the relevance [8] between text and image/motion features to highlight the information relevant to the prediction. The different \"regions of interest\" and \"words of importance\" indicate that the motion and image features could be disentangled." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "bbox": [ + 46, + 366, + 287, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 366, + 287, + 462 + ], + "spans": [ + { + "bbox": [ + 46, + 366, + 287, + 462 + ], + "type": "text", + "content": "and ViT-L/14-336p as our image encoder. Unless otherwise specified, we mark the default settings in the temporal encoder in gray in Sec. 4.2. For simplicity, the ablation studies are conducted based on ViT-B/16 with a low-rate sampled 8 frames. We conduct the experiments with the NVIDIA 32G V100 GPUs. More implementation details (e.g., training and testing hyper-parameters) are described in the Appendix." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 47, + 468, + 174, + 480 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 468, + 174, + 480 + ], + "spans": [ + { + "bbox": [ + 47, + 468, + 174, + 480 + ], + "type": "text", + "content": "4.2. Ablation and Analysis" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 46, + 486, + 289, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 486, + 289, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 486, + 289, + 714 + ], + "type": "text", + "content": "Language Supervision. In this section, we aim to demonstrate the significance of language supervision by comparing different types of text descriptions. Tab. 3a presents a comparison of three types of texts: category names, motion concept descriptions, motion discriminative descriptions. It reveals that the performance of action category names, which contain the least information, is notably lower than other results, particularly on the SSv2 dataset with abundant temporal information. Although descriptions with only action concepts show a slight improvement of " + }, + { + "bbox": [ + 46, + 486, + 289, + 714 + ], + "type": "inline_equation", + "content": "0.3\\% /1.6\\%" + }, + { + "bbox": [ + 46, + 486, + 289, + 714 + ], + "type": "text", + "content": ", they lack the vital information regarding the distinguishing characteristics of the actions. As a result, the model exhibits poor classification accuracy for similar categories. On the other hand, when the discriminative descriptions are adopted, the accuracy on SSv2 is improved noticeably " + }, + { + "bbox": [ + 46, + 486, + 289, + 714 + ], + "type": "inline_equation", + "content": "(+4.2\\%)" + }, + { + "bbox": [ + 46, + 486, + 289, + 714 + ], + "type": "text", + "content": ", revealing the significance of discriminative characteristics for fine-grained datasets. Additionally, when the full descriptions are utilized, the model's performance obtain the further gains of " + }, + { + "bbox": [ + 46, + 486, + 289, + 714 + ], + "type": "inline_equation", + "content": "4.4\\% /2.2\\%" + }, + { + "bbox": [ + 46, + 486, + 289, + 714 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 366, + 546, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 366, + 546, + 569 + ], + "spans": [ + { + "bbox": [ + 304, + 366, + 546, + 569 + ], + "type": "text", + "content": "Disentangled Motion Modeling. Our method has an advantage in allowing the learning of dynamic information without interference, while preserving CLIP's original spatially transferable representation capabilities. To validate this, we compared two different structures in Tab. 3b. For the serial structure, additional dynamic modules are inserted between the Transformer Blocks of each layer of the visual encoder, resulting in a unified video representation. All the texts are combined through the text encoder to obtain text representations for alignment. It can be observed that the parallel structure outperforms the serial structure, with improvements of " + }, + { + "bbox": [ + 304, + 366, + 546, + 569 + ], + "type": "inline_equation", + "content": "4.3\\%" + }, + { + "bbox": [ + 304, + 366, + 546, + 569 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 366, + 546, + 569 + ], + "type": "inline_equation", + "content": "1.9\\%" + }, + { + "bbox": [ + 304, + 366, + 546, + 569 + ], + "type": "text", + "content": " respectively. This demonstrates the rationality of the parallel approach. Tab. 3b also presents a comparison of introducing an adapter in the text encoder. The model's classification performance improved by " + }, + { + "bbox": [ + 304, + 366, + 546, + 569 + ], + "type": "inline_equation", + "content": "1.9\\%" + }, + { + "bbox": [ + 304, + 366, + 546, + 569 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 366, + 546, + 569 + ], + "type": "inline_equation", + "content": "2.3\\%" + }, + { + "bbox": [ + 304, + 366, + 546, + 569 + ], + "type": "text", + "content": " after adding the adapter, indicating its effectiveness in reducing bias in the text encoder." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 570, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 547, + 713 + ], + "type": "text", + "content": "Feature Fusion Direction. As shown in Tab. 3c, both directions of information integration can improve performances. The combination of the two can boost accuracy more significantly " + }, + { + "bbox": [ + 304, + 570, + 547, + 713 + ], + "type": "inline_equation", + "content": "+1.8\\% / + 1.5\\%" + }, + { + "bbox": [ + 304, + 570, + 547, + 713 + ], + "type": "text", + "content": " on SSv2/K400, respectively. This verifies the importance of spatio-temporal blending for the parallel architectures. In our opinion, independent learning of motion can effectively avoid excessive reliance of the model on the previously learned spatial information. By using cross attention for fusion, it is possible to effectively integrate features from two different dimensions, ultimately forming the features of the target video and achieving the best results." + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "18509" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 70, + 538, + 340 + ], + "blocks": [ + { + "bbox": [ + 56, + 70, + 538, + 340 + ], + "lines": [ + { + "bbox": [ + 56, + 70, + 538, + 340 + ], + "spans": [ + { + "bbox": [ + 56, + 70, + 538, + 340 + ], + "type": "table", + "html": "
MethodPre-trainArchitectureInput SizeFLOPs×Cr.×Cl. (T)Param (M)FrozenTop-1Top-5
SlowFast [17]ImageNet-21KR101+NL16 × 22420.1 × 3 × 160X63.187.6
ViViT FE [1]IN21K+K400ViT-L16 × 22421.0 × 3 × 4612X65.489.8
MTV-B(320p) [84]IN21K+K400-32 × 22420.9 × 3 × 4310X68.590.4
MViT [16]Kinetics-600MViT-B-2432 × 22420.2 × 3 × 153X68.791.5
Video Swin [42]IN21K+K400Swin-B32 × 22420.3 × 3 × 160X69.692.7
TAdaConvNeXtV2 [22]IN1K+K400ConvNeXt-S32 × 22420.2 × 3 × 282X70.092.0
EVL* [40]CLIP-400MViT-B32 × 22420.68 × 1 × 3175✓62.4-
ST-Adapter* [52]CLIP-400MViT-B32 × 22420.61 × 1 × 393✓69.592.6
DiST*CLIP-400MViT-B32 × 22420.65 × 1 × 3105✓70.992.1
MoTED*CLIP-400MViT-B8 × 22420.18 × 1 × 3112✓70.191.8
MoTED*CLIP-400MViT-B16 × 22420.34 × 1 × 3112✓71.292.4
MoTED*CLIP-400MViT-B32 × 22420.68 × 1 × 3112✓71.992.7
UnifromerV2 [32]CLIP-400MViT-L32 × 22421.73 × 1 × 3574X73.094.5
TAdaFormer [22]CLIP-400MViT-L32 × 22421.70 × 2 × 3364X73.6-
EVL* [40]CLIP-400MViT-L32 × 22423.21 × 1 × 3654✓66.7-
EVL* [40]CLIP-400MViT-L32 × 33628.08 × 1 × 3654✓68.0-
ST-Adapter* [52]CLIP-400MViT-L32 × 22422.75 × 1 × 3347✓72.393.9
DiST*CLIP-400MViT-L32 × 22422.83 × 1 × 3336✓73.193.2
MoTED*CLIP-400MViT-L8 × 22420.78 × 1 × 3346✓71.592.6
MoTED*CLIP-400MViT-L16 × 22421.49 × 1 × 3346✓73.093.4
MoTED*CLIP-400MViT-L32 × 22422.89 × 1 × 3346✓73.893.8
", + "image_path": "5a47145889b60ab18a3b0c5a8a3ee11b1d046b4d3fdaa2a509dfadfa3b270383.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 392, + 289, + 585 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 392, + 289, + 585 + ], + "spans": [ + { + "bbox": [ + 46, + 392, + 289, + 585 + ], + "type": "text", + "content": "Motion Modeling Visualization. In Fig. 4, we also perform the analysis of vision features generated by image and motion encoder to investigate the learned patterns from language supervision. Based on the reasoning tool [8], we depict the attention maps of class token from the final transformer block of the image/motion stream encoder w.r.t. the text encoder. It is observed that, for the case1 in Fig. 4, the CLIP image encoder attends to both motion-relevant foreground and motion-irrelevant background with a major focus on \"surface\". In contrast, features extracted from the motion modules concentrate on the motion-relevant regions of the moving object, and emphasize the motion-related words \"sliding\". This phenomenon reveals that the motion features can complement the static spatial semantics of objects in the image features via modeling object motions. More cases can be accessed in the Appendix." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 597, + 211, + 611 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 597, + 211, + 611 + ], + "spans": [ + { + "bbox": [ + 47, + 597, + 211, + 611 + ], + "type": "text", + "content": "4.3. Fully-supervised Experiments" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 617, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 617, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 617, + 288, + 715 + ], + "type": "text", + "content": "In the supervised setting, the results on SSv2 and K400 are presented in Tab. 4 and Tab. 6, respectively. Compared with AVL, our proposed MoTED introduces a similar temporal module for CLIP visual encoder, but our method has a significant improvement compared to AVL, with an improvement of " + }, + { + "bbox": [ + 46, + 617, + 288, + 715 + ], + "type": "inline_equation", + "content": "9.5\\% / 2.0\\%" + }, + { + "bbox": [ + 46, + 617, + 288, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 617, + 288, + 715 + ], + "type": "inline_equation", + "content": "5.8\\% / 1.5\\%" + }, + { + "bbox": [ + 46, + 617, + 288, + 715 + ], + "type": "text", + "content": " on ViT-B and ViT-L respectively on SSv2/K400. The performance gains greatly demonstrate the rationality and effectiveness of using lan" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 392, + 546, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 392, + 546, + 441 + ], + "spans": [ + { + "bbox": [ + 304, + 392, + 546, + 441 + ], + "type": "text", + "content": "guage supervision. Interestingly, the performance gains are particularly pronounced on SSv2. This is because SSv2 is a fine-grained action classification task requiring stronger action discrimination and perception of dynamic information." + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 307, + 452, + 544, + 505 + ], + "blocks": [ + { + "bbox": [ + 46, + 349, + 546, + 373 + ], + "lines": [ + { + "bbox": [ + 46, + 349, + 546, + 373 + ], + "spans": [ + { + "bbox": [ + 46, + 349, + 546, + 373 + ], + "type": "text", + "content": "Table 4. Comparison with the state-of-the-art methods on Something-Something-V2. \"Cr.\" and \"Cl.\" are the abbreviation for \"spatial crops\" and \"temporal clips\". \"Frozen\" indicates freezing the CLIP pre-trained parameters." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 307, + 452, + 544, + 505 + ], + "lines": [ + { + "bbox": [ + 307, + 452, + 544, + 505 + ], + "spans": [ + { + "bbox": [ + 307, + 452, + 544, + 505 + ], + "type": "table", + "html": "
Related workVisual disentangleTextual disentangleK400 (Acc / Δ)SSv2 (Acc / Δ)
Previous SOTAXX84.2 / -69.5 / -
DiST [58]✓X85.0 / +0.870.9 / +1.4
MoTED (Ours)✓✓86.2 / +1.271.9 / +1.0
", + "image_path": "4b9db167afe472ddeb2c93536b105974dd8fa356d8b2b50e7459f56759f806bc.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 509, + 545, + 531 + ], + "lines": [ + { + "bbox": [ + 306, + 509, + 545, + 531 + ], + "spans": [ + { + "bbox": [ + 306, + 509, + 545, + 531 + ], + "type": "text", + "content": "Table 5. Comparison of our method with previous state-of-the-art (SOTA) and the latest related work DiST." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 541, + 547, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 541, + 547, + 662 + ], + "spans": [ + { + "bbox": [ + 304, + 541, + 547, + 662 + ], + "type": "text", + "content": "Disentangling spatio-temporal learning is an effective approach to endow the model with temporal capability. As shown in Tab. 5, DiST focuses on temporal disentangling in the visual side with gains of " + }, + { + "bbox": [ + 304, + 541, + 547, + 662 + ], + "type": "inline_equation", + "content": "+0.8\\% / +1.4\\%" + }, + { + "bbox": [ + 304, + 541, + 547, + 662 + ], + "type": "text", + "content": " on K400/SSv2. Our study highlights the effectiveness of temporal disentangling in the textual side with detailed descriptions of motions, with further gains of " + }, + { + "bbox": [ + 304, + 541, + 547, + 662 + ], + "type": "inline_equation", + "content": "+1.2\\% / +1.0\\%" + }, + { + "bbox": [ + 304, + 541, + 547, + 662 + ], + "type": "text", + "content": ". This result reveals that disentangling textual encoder is equally effective w.r.t. disentangling visual encoder for vision-language transfer learning." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 670, + 437, + 684 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 670, + 437, + 684 + ], + "spans": [ + { + "bbox": [ + 306, + 670, + 437, + 684 + ], + "type": "text", + "content": "4.4. Zero-shot Experiments" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 689, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 689, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 305, + 689, + 547, + 715 + ], + "type": "text", + "content": "Zero-shot generalization is an attractive characteristic of CLIP-extended models, making them more practical in real" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "18510" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 58, + 69, + 537, + 341 + ], + "blocks": [ + { + "bbox": [ + 58, + 69, + 537, + 341 + ], + "lines": [ + { + "bbox": [ + 58, + 69, + 537, + 341 + ], + "spans": [ + { + "bbox": [ + 58, + 69, + 537, + 341 + ], + "type": "table", + "html": "
MethodPre-trainArchitectureInput SizeTFLOPs×Cr.×Cl.Param (M)FrozenTop-1Top-5
SlowFast [17]-R101+NL16 × 22420.4 × 3 × 1060X79.893.9
TimeSformer [2]ImageNet-21KViT-L96 × 22428.4 × 3 × 1430X80.794.7
MViT [16]-MViT-B64 × 22420.5 × 1 × 537X81.295.1
ViViT FE [1]ImageNet-21KViT-L128 × 22424.0 × 3 × 1N/AX81.793.8
Video Swin [42]ImageNet-21KSwin-L32 × 22420.6 × 3 × 4197X83.195.9
TAdaConvNeXtV2 [22]ImageNet-21KConvNeXt-B32 × 22420.3 × 3 × 4146X83.7-
X-CLIP [49]CLIP-400MViT-B16 × 22420.28 × 3 × 4128X84.796.8
ST-Adapter* [52]CLIP-400MViT-B32 × 22420.61 × 1 × 393✓82.796.2
EVL* [40]CLIP-400MViT-B32 × 22420.59 × 1 × 3115✓84.2-
DiST*CLIP-400MViT-B32 × 22420.65 × 1 × 3112✓85.097.0
MoTED*CLIP-400MViT-B8 × 22420.18 × 1 × 3116✓85.197.0
MoTED*CLIP-400MViT-B16 × 22420.34 × 1 × 3116✓85.497.2
MoTED*CLIP-400MViT-B32 × 22420.68 × 1 × 3116✓86.297.5
UnifromerV2 [32]CLIP-400M+K710ViT-L32 × 22422.66 × 2 × 3354X89.398.2
TAdaFormer [22]CLIP-400M+K710ViT-L32 × 22421.41 × 4 × 3364X89.5-
ST-Adapter* [52]CLIP-400MViT-L32 × 22422.75 × 1 × 3347✓87.297.6
EVL* [40]CLIP-400MViT-L32 × 22422.70 × 1 × 3363✓87.3-
DiST*CLIP-400MViT-L32 × 22422.83 × 1 × 3343✓88.097.9
MoTED*CLIP-400MViT-L8 × 22420.78 × 1 × 3349✓87.497.8
MoTED*CLIP-400MViT-L16 × 22421.49 × 1 × 3349✓88.098.0
MoTED*CLIP-400MViT-L32 × 22422.89 × 1 × 3349✓88.898.2
", + "image_path": "db809d50ed249d841c3bea2cccfac414162b015dcdb087160197aa5123ec8e7d.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 48, + 391, + 287, + 463 + ], + "blocks": [ + { + "bbox": [ + 47, + 349, + 545, + 373 + ], + "lines": [ + { + "bbox": [ + 47, + 349, + 545, + 373 + ], + "spans": [ + { + "bbox": [ + 47, + 349, + 545, + 373 + ], + "type": "text", + "content": "Table 6. Comparison with state-of-the-arts on Kinetics-400. \"Cr.\" and \"Cl.\" are the abbreviation for \"spatial crops\" and \"temporal clips\". \"Frozen\" indicates freezing the CLIP pre-trained parameters." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 48, + 391, + 287, + 463 + ], + "lines": [ + { + "bbox": [ + 48, + 391, + 287, + 463 + ], + "spans": [ + { + "bbox": [ + 48, + 391, + 287, + 463 + ], + "type": "table", + "html": "
MethodModelHMDB51UCF101K600
ActionCLIP [75]B/1640.8±5.458.3±3.4-
X-CLIP [49]B/1644.6±5.272.0±2.365.2±0.4
Vita-CLIP [78]B/1648.6±0.675.0±0.667.4±0.5
DiST*B/1655.4±1.272.3±0.6-
MoTED*B/1658.2±1.178.3±0.669.9 ± 0.5
", + "image_path": "8ee63430ed2f55d74a3a4f0f4d58974805df2f3860cf317afb32f184c179b72b.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 471, + 287, + 506 + ], + "lines": [ + { + "bbox": [ + 47, + 471, + 287, + 506 + ], + "spans": [ + { + "bbox": [ + 47, + 471, + 287, + 506 + ], + "type": "text", + "content": "Table 7. Comparison of zero-shot accuracy with the state-of-the-art CLIP-based methods on three datasets (e.g., HMDB51 [24], UCF101 [65], and K600 [12]). \"\": frozen backbone." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 521, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 521, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 521, + 289, + 715 + ], + "type": "text", + "content": "world applications. For zero-shot settings, we evaluate MoTED on three widely used benchmarks. Following prior work [49], we train the networks on the K400 training set, then conduct the zero-shot evaluation on three unseen datasets (i.e., UCF101, HMDB51 and K600), as shown in Tab. 7. All models share the same architecture of \"ViT-B\", with 32 frames during inference. Compared with other vison-language methods, MoTED achieves the better zero-shot performances with a significant margin on HMDB51 " + }, + { + "bbox": [ + 46, + 521, + 289, + 715 + ], + "type": "inline_equation", + "content": "(+3.8\\%)" + }, + { + "bbox": [ + 46, + 521, + 289, + 715 + ], + "type": "text", + "content": ", UCF101 " + }, + { + "bbox": [ + 46, + 521, + 289, + 715 + ], + "type": "inline_equation", + "content": "(+6.0\\%)" + }, + { + "bbox": [ + 46, + 521, + 289, + 715 + ], + "type": "text", + "content": " and K600 " + }, + { + "bbox": [ + 46, + 521, + 289, + 715 + ], + "type": "inline_equation", + "content": "(+2.5\\%)" + }, + { + "bbox": [ + 46, + 521, + 289, + 715 + ], + "type": "text", + "content": ". Different from X-CLIP [49], DiST [58], Vita-CLIP [78] that learns motion representation with the supervision of action category names solely, the proposed MoTED makes full use of conceptual and discriminative descriptions and learn general motion representations with the aid of language supervision. In addition, our method also has a relatively small" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 305, + 392, + 545, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 392, + 545, + 430 + ], + "spans": [ + { + "bbox": [ + 305, + 392, + 545, + 430 + ], + "type": "text", + "content": "variance, only about " + }, + { + "bbox": [ + 305, + 392, + 545, + 430 + ], + "type": "inline_equation", + "content": "1\\%" + }, + { + "bbox": [ + 305, + 392, + 545, + 430 + ], + "type": "text", + "content": ". We assume this is due to the benefits brought by the rich content of the text, as detailed descriptions can promote stable dynamic feature learning." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 442, + 379, + 455 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 442, + 379, + 455 + ], + "spans": [ + { + "bbox": [ + 306, + 442, + 379, + 455 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 463, + 545, + 643 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 463, + 545, + 643 + ], + "spans": [ + { + "bbox": [ + 304, + 463, + 545, + 643 + ], + "type": "text", + "content": "In this study, we aim to overcome the limitations of existing methods in extending large image-text pre-trained models for video understanding. The proposed MoTED framework introduces Motion-Enhanced Descriptions, which are applied to facilitate the extraction of unique temporal features in videos. By generating motion-related descriptions and incorporating perception modules, MoTED aligns visual and textual motion features using a contrastive loss. Experimental results on five benchmarks demonstrate that MoTED provides a strong basis for enhancing CLIP with robust temporal modeling. In future works, we hope to further dedicate to exploring the potential of language supervision and combining it with more powerful dynamic information perception modules to achieve higher performance in video recognition and make it truly practical." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 656, + 416, + 670 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 656, + 416, + 670 + ], + "spans": [ + { + "bbox": [ + 306, + 656, + 416, + 670 + ], + "type": "text", + "content": "6. Acknowledgement" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 677, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 677, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 677, + 545, + 713 + ], + "type": "text", + "content": "This work was supported in part by NSFC No. 62222117, the Fundamental Research Funds for the Central Universities under contract WK3490000005, and KY2100000117." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18511" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 287, + 714 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 133 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 133 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 133 + ], + "type": "text", + "content": "[1] Anurag Arnab, Mostafa Dehghani, Georg Heigold, Chen Sun, Mario Lučić, and Cordelia Schmid. Vivit: A video vision transformer. In ICCV, pages 6836-6846, 2021. 3, 7, 8" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 136, + 287, + 168 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 136, + 287, + 168 + ], + "spans": [ + { + "bbox": [ + 53, + 136, + 287, + 168 + ], + "type": "text", + "content": "[2] Gedas Bertasius, Heng Wang, and Lorenzo Torresani. Is space-time attention all you need for video understanding? In ICML, page 4, 2021. 3, 8" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 170, + 287, + 203 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 170, + 287, + 203 + ], + "spans": [ + { + "bbox": [ + 53, + 170, + 287, + 203 + ], + "type": "text", + "content": "[3] Daniel Bolya, Cheng-Yang Fu, Xiaoliang Dai, Peizhao Zhang, Christoph Feichtenhofer, and Judy Hoffman. Token merging: Your vit but faster. In ICLR, 2023. 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 204, + 287, + 312 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 204, + 287, + 312 + ], + "spans": [ + { + "bbox": [ + 53, + 204, + 287, + 312 + ], + "type": "text", + "content": "[4] Tom B. Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-Voss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel M. Ziegler, Jeffrey Wu, Clemens Winter, Christopher Hesse, Mark Chen, Eric Sigler, Mateusz Litwin, Scott Gray, Benjamin Chess, Jack Clark, Christopher Berner, Sam McCandlish, Alec Radford, Ilya Sutskever, and Dario Amodei. Language models are few-shot learners. In NeurIPS, 2020. 2, 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 314, + 287, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 314, + 287, + 357 + ], + "spans": [ + { + "bbox": [ + 53, + 314, + 287, + 357 + ], + "type": "text", + "content": "[5] Shyamal Buch, Cristóbal Eyzaguirre, Adrien Gaidon, Jiajun Wu, Li Fei-Fei, and Juan Carlos Niebles. Revisiting the \"video\" in video-language understanding. In CVPR, pages 2907–2917, 2022. 2, 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 358, + 287, + 401 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 358, + 287, + 401 + ], + "spans": [ + { + "bbox": [ + 53, + 358, + 287, + 401 + ], + "type": "text", + "content": "[6] Adrian Bulat, Juan Manuel Perez Rua, Swathikiran Sudhakaran, Brais Martinez, and Georgios Tzimiropoulos. Space-time mixing attention for video transformer. NeurIPS, 34:19594-19607, 2021. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 403, + 287, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 403, + 287, + 435 + ], + "spans": [ + { + "bbox": [ + 53, + 403, + 287, + 435 + ], + "type": "text", + "content": "[7] Joao Carreira and Andrew Zisserman. Quo vadis, action recognition? a new model and the kinetics dataset. In CVPR, pages 6299-6308, 2017. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 437, + 287, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 437, + 287, + 479 + ], + "spans": [ + { + "bbox": [ + 53, + 437, + 287, + 479 + ], + "type": "text", + "content": "[8] Hila Chefer, Shir Gur, and Lior Wolf. Generic attention-model explainability for interpreting bi-modal and encoder-decoder transformers. In ICCV, pages 387-396, 2021. 6, 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 481, + 287, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 481, + 287, + 568 + ], + "spans": [ + { + "bbox": [ + 53, + 481, + 287, + 568 + ], + "type": "text", + "content": "[9] Xi Chen, Xiao Wang, Soravit Changpinyo, A. J. Piergiovanni, Piotr Padlewski, Daniel Salz, Sebastian Goodman, Adam Grycner, Basil Mustafa, Lucas Beyer, Alexander Kolesnikov, Joan Puigcerver, Nan Ding, Keran Rong, Hassan Akbari, Gaurav Mishra, Linting Xue, Ashish V. Thapliyal, James Bradbury, and Weicheng Kuo. Pali: A jointly scaled multilingual language-image model. In ICLR, 2023. 1, 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 570, + 287, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 570, + 287, + 613 + ], + "spans": [ + { + "bbox": [ + 48, + 570, + 287, + 613 + ], + "type": "text", + "content": "[10] Xing Cheng, Hezheng Lin, Xiangyu Wu, Fan Yang, and Dong Shen. Improving video-text retrieval by multi-stream corpus alignment and dual softmax loss. arXiv preprint arXiv:2109.04290, 2021. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 614, + 287, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 614, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 48, + 614, + 287, + 714 + ], + "type": "text", + "content": "[11] Aakanksha Chowdhery, Sharan Narang, Jacob Devlin, Maarten Bosma, Gaurav Mishra, Adam Roberts, Paul Barham, Hyung Won Chung, Charles Sutton, Sebastian Gehrmann, Parker Schuh, Kensen Shi, Sasha Tsvyashchenko, Joshua Maynez, Abhishek Rao, Parker Barnes, Yi Tay, Noam Shazeer, Vinodkumar Prabhakaran, Emily Reif, Nan Du, Ben Hutchinson, Reiner Pope, James Bradbury, Jacob Austin, Michael Isard, Guy Gur-Ari, Pengcheng Yin, Toju Duke, Anselm Levskaya, Sanjay" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 545, + 714 + ], + "type": "list", + "angle": 0, + "index": 25, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 215 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 215 + ], + "type": "text", + "content": "Ghemawat, Sunipa Dev, Henryk Michalewski, Xavier Garcia, Vedant Misra, Kevin Robinson, Liam Fedus, Denny Zhou, Daphne Ippolito, David Luan, Hyeontaek Lim, Barret Zoph, Alexander Spiridonov, Ryan Sepassi, David Dohan, Shivani Agrawal, Mark Omernick, Andrew M. Dai, Thanumalayan Sankaranarayana Pillai, Marie Pellat, Aitor Lewkowycz, Erica Moreira, Rewon Child, Oleksandr Polozov, Katherine Lee, Zongwei Zhou, Xuezhi Wang, Brennan Saeta, Mark Diaz, Orhan Firat, Michele Catasta, Jason Wei, Kathy Meier-Hellstern, Douglas Eck, Jeff Dean, Slav Petrov, and Noah Fiedel. Palm: Scaling language modeling with pathways. J. Mach. Learn. Res., 24:240:1-240:113, 2023. 2, 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 308, + 217, + 545, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 217, + 545, + 271 + ], + "spans": [ + { + "bbox": [ + 308, + 217, + 545, + 271 + ], + "type": "text", + "content": "[12] Dima Damen, Hazel Doughty, Giovanni Maria Farinella, Antonino Furnari, Evangelos Kazakos, Jian Ma, Davide Moltisanti, Jonathan Munro, Toby Perrett, Will Price, et al. Rescaling egocentric vision. arXiv preprint arXiv:2006.13256, 2020. 2, 5, 8" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 308, + 273, + 545, + 317 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 273, + 545, + 317 + ], + "spans": [ + { + "bbox": [ + 308, + 273, + 545, + 317 + ], + "type": "text", + "content": "[13] Ali Diba, Vivek Sharma, Reza Safdari, Dariush Lotfi, Saquib Sarfraz, Rainer Stiefelhagen, and Luc Van Gool. Vi2cIr: Video and image for visual contrastive learning of representation. In ICCV, pages 1502-1512, 2021. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 318, + 545, + 361 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 318, + 545, + 361 + ], + "spans": [ + { + "bbox": [ + 308, + 318, + 545, + 361 + ], + "type": "text", + "content": "[14] Sixun Dong, Huazhang Hu, Dongze Lian, Weixin Luo, Yicheng Qian, and Shenghua Gao. Weakly supervised video representation learning with unaligned text for sequential videos. In CVPR, pages 2437-2447. IEEE, 2023. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 363, + 545, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 363, + 545, + 418 + ], + "spans": [ + { + "bbox": [ + 308, + 363, + 545, + 418 + ], + "type": "text", + "content": "[15] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2020. 4" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 419, + 545, + 463 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 419, + 545, + 463 + ], + "spans": [ + { + "bbox": [ + 308, + 419, + 545, + 463 + ], + "type": "text", + "content": "[16] Haoqi Fan, Bo Xiong, Karttikeya Mangalam, Yanghao Li, Zhicheng Yan, Jitendra Malik, and Christoph Feichtenhofer. Multiscale vision transformers. In ICCV, pages 6824-6835, 2021. 3, 7, 8" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 464, + 545, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 464, + 545, + 498 + ], + "spans": [ + { + "bbox": [ + 308, + 464, + 545, + 498 + ], + "type": "text", + "content": "[17] Christoph Feichtenhofer, Haoqi Fan, Jitendra Malik, and Kaiming He. Slowfast networks for video recognition. In ICCV, pages 6202-6211, 2019. 3, 5, 7, 8" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 499, + 545, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 499, + 545, + 542 + ], + "spans": [ + { + "bbox": [ + 308, + 499, + 545, + 542 + ], + "type": "text", + "content": "[18] Christoph Feichtenhofer, Haoqi Fan, Bo Xiong, Ross Girshick, and Kaiming He. A large-scale study on unsupervised spatiotemporal representation learning. In CVPR, pages 3299-3309, 2021. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 544, + 545, + 576 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 544, + 545, + 576 + ], + "spans": [ + { + "bbox": [ + 308, + 544, + 545, + 576 + ], + "type": "text", + "content": "[19] Christoph Feichtenhofer, Haoqi Fan, Yanghao Li, and Kaiming He. Masked autoencoders as spatiotemporal learners. NeurIPS, 2022. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 578, + 545, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 578, + 545, + 644 + ], + "spans": [ + { + "bbox": [ + 308, + 578, + 545, + 644 + ], + "type": "text", + "content": "[20] Raghav Goyal, Samira Ebrahimi Kahou, Vincent Michalski, Joanna Materzynska, Susanne Westphal, Heuna Kim, Valentin Haenel, Ingo Fruend, Peter Yianilos, Moritz Mueller-Freitag, et al. The\" something something\" video database for learning and evaluating visual common sense. In ICCV, pages 5842-5850, 2017. 2, 5" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 646, + 545, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 646, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 308, + 646, + 545, + 689 + ], + "type": "text", + "content": "[21] Ziyuan Huang, Shiwei Zhang, Liang Pan, Zhiwu Qing, Mingqian Tang, Ziwei Liu, and Marcelo H Ang Jr. Tada! temporally-adaptive convolutions for video understanding. In ICLR, 2022. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 691, + 545, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 691, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 308, + 691, + 545, + 714 + ], + "type": "text", + "content": "[22] Ziyuan Huang, Shiwei Zhang, Liang Pan, Zhiwu Qing, Yingya Zhang, Ziwei Liu, and Marcelo H Ang Jr." + } + ] + } + ], + "index": 24 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "18512" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 66, + 72, + 286, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 72, + 286, + 95 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 286, + 95 + ], + "type": "text", + "content": "Temporally-adaptive models for efficient video understanding. arXiv preprint arXiv:2308.05787, 2023. 7, 8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 96, + 287, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 96, + 287, + 128 + ], + "spans": [ + { + "bbox": [ + 48, + 96, + 287, + 128 + ], + "type": "text", + "content": "[23] Simon Jenni and Hailin Jin. Time-equivariant contrastive video representation learning. In ICCV, pages 9970-9980, 2021. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 130, + 287, + 163 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 130, + 287, + 163 + ], + "spans": [ + { + "bbox": [ + 48, + 130, + 287, + 163 + ], + "type": "text", + "content": "[24] H Jhuang, H Garrote, E Poggio, T Serre, and T Hmdb. A large video database for human motion recognition. In ICCV, page 6, 2011. 2, 5, 8" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 164, + 287, + 218 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 164, + 287, + 218 + ], + "spans": [ + { + "bbox": [ + 48, + 164, + 287, + 218 + ], + "type": "text", + "content": "[25] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In ICML, pages 4904-4916. PMLR, 2021. 1, 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 220, + 287, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 220, + 287, + 262 + ], + "spans": [ + { + "bbox": [ + 48, + 220, + 287, + 262 + ], + "type": "text", + "content": "[26] Chen Ju, Tengda Han, Kunhao Zheng, Ya Zhang, and Weidi Xie. Prompting visual-language models for efficient video understanding. In ECCV, pages 105-124. Springer, 2022. 1, 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 265, + 287, + 298 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 265, + 287, + 298 + ], + "spans": [ + { + "bbox": [ + 48, + 265, + 287, + 298 + ], + "type": "text", + "content": "[27] Chen Ju, Tengda Han, Kunhao Zheng, Ya Zhang, and Weidi Xie. Prompting visual-language models for efficient video understanding. In ECCV, pages 105-124. Springer, 2022. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 300, + 287, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 300, + 287, + 354 + ], + "spans": [ + { + "bbox": [ + 48, + 300, + 287, + 354 + ], + "type": "text", + "content": "[28] Will Kay, Joao Carreira, Karen Simonyan, Brian Zhang, Chloe Hillier, Sudheendra Vijayanarasimhan, Fabio Viola, Tim Green, Trevor Back, Paul Natsev, et al. The kinetics human action video dataset. arXiv preprint arXiv:1705.06950, 2017. 2, 3, 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 355, + 287, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 355, + 287, + 453 + ], + "spans": [ + { + "bbox": [ + 48, + 355, + 287, + 453 + ], + "type": "text", + "content": "[29] Jan Kocon, Igor Cichecki, Oliwier Kaszyca, Mateusz Kochanek, Dominika Szydlo, Joanna Baran, Julita Bielaniewicz, Marcin Gruza, Arkadiusz Janz, Kamil Kanclerz, Anna Kocon, Bartlomiej Koptyra, Wiktoria Mieleszczenko-Kowszewicz, Piotr Milkowski, Marcin Oleksy, Maciej Piasecki, Lukasz Radlinski, Konrad Wojtasik, Stanislaw Wozniak, and Przemyslaw Kazienko. Chatgpt: Jack of all trades, master of none. Inf. Fusion, 99: 101861, 2023. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 455, + 287, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 455, + 287, + 487 + ], + "spans": [ + { + "bbox": [ + 48, + 455, + 287, + 487 + ], + "type": "text", + "content": "[30] Jie Lei, Tamara L. Berg, and Mohit Bansal. Revealing single frame bias for video-and-language learning. In ACL, pages 487-507, 2023. 2, 4" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 489, + 287, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 489, + 287, + 544 + ], + "spans": [ + { + "bbox": [ + 48, + 489, + 287, + 544 + ], + "type": "text", + "content": "[31] Chunyuan Li, Haotian Liu, Liunian Harold Li, Pengchuan Zhang, Jyoti Aneja, Jianwei Yang, Ping Jin, Houdong Hu, Zicheng Liu, Yong Jae Lee, and Jianfeng Gao. ELEVATER: A benchmark and toolkit for evaluating language-augmented visual models. In NeurIPS, 2022. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 545, + 287, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 545, + 287, + 589 + ], + "spans": [ + { + "bbox": [ + 48, + 545, + 287, + 589 + ], + "type": "text", + "content": "[32] Kunchang Li, Yali Wang, Yinan He, Yizhuo Li, Yi Wang, Limin Wang, and Yu Qiao. Uniformerv2: Spatiotemporal learning by arming image vits with video uniformer. arXiv preprint arXiv:2211.09552, 2022. 3, 5, 7, 8" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 590, + 287, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 590, + 287, + 633 + ], + "spans": [ + { + "bbox": [ + 48, + 590, + 287, + 633 + ], + "type": "text", + "content": "[33] Kunchang Li, Yali Wang, Junhao Zhang, Peng Gao, Guanglu Song, Yu Liu, Hongsheng Li, and Yu Qiao. Uniformer: Unifying convolution and self-attention for visual recognition. TPAMI, 2023. 3, 5" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 635, + 287, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 635, + 287, + 678 + ], + "spans": [ + { + "bbox": [ + 48, + 635, + 287, + 678 + ], + "type": "text", + "content": "[34] Linjie Li, Yen-Chun Chen, Yu Cheng, Zhe Gan, Licheng Yu, and Jingjing Liu. HERO: hierarchical encoder for video+language omni-representation pre-training. In EMNLP, pages 2046-2065, 2020. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "type": "text", + "content": "[35] Liunian Harold Li, Pengchuan Zhang, Haotian Zhang, Jianwei Yang, Chunyuan Li, Yiwu Zhong, Lijuan Wang, Lu Yuan, Lei Zhang, Jenq-Neng Hwang, Kai-Wei Chang, and" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 32, + "blocks": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "type": "text", + "content": "Jianfeng Gao. Grounded language-image pre-training. In CVPR, pages 10955-10965, 2022. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 96, + 545, + 129 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 96, + 545, + 129 + ], + "spans": [ + { + "bbox": [ + 307, + 96, + 545, + 129 + ], + "type": "text", + "content": "[36] Yanghao Li, Haoqi Fan, Ronghang Hu, Christoph Feichtenhofer, and Kaiming He. Scaling language-image pre-training via masking. arXiv preprint arXiv:2212.00794, 2022. 1, 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 131, + 545, + 175 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 131, + 545, + 175 + ], + "spans": [ + { + "bbox": [ + 307, + 131, + 545, + 175 + ], + "type": "text", + "content": "[37] Yangguang Li, Feng Liang, Lichen Zhao, Yufeng Cui, Wanli Ouyang, Jing Shao, Fengwei Yu, and Junjie Yan. Supervision exists everywhere: A data efficient contrastive language-image pre-training paradigm. In ICLR, 2022. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 177, + 545, + 230 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 177, + 545, + 230 + ], + "spans": [ + { + "bbox": [ + 307, + 177, + 545, + 230 + ], + "type": "text", + "content": "[38] Yanghao Li, Chao-Yuan Wu, Haoqi Fan, Karttikeya Mangalam, Bo Xiong, Jitendra Malik, and Christoph Feichtenhofer. Mvitv2: Improved multiscale vision transformers for classification and detection. In CVPR, pages 4804-4814, 2022. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 233, + 545, + 277 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 233, + 545, + 277 + ], + "spans": [ + { + "bbox": [ + 307, + 233, + 545, + 277 + ], + "type": "text", + "content": "[39] Xudong Lin, Fabio Petroni, Gedas Bertasius, Marcus Rohrbach, Shih-Fu Chang, and Lorenzo Torresani. Learning to recognize procedural activities with distant supervision. In CVPR, pages 13843-13853. IEEE, 2022. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 278, + 545, + 323 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 278, + 545, + 323 + ], + "spans": [ + { + "bbox": [ + 307, + 278, + 545, + 323 + ], + "type": "text", + "content": "[40] Ziyi Lin, Shijie Geng, Renrui Zhang, Peng Gao, Gerard de Melo, Xiaogang Wang, Jifeng Dai, Yu Qiao, and Hongsheng Li. Frozen clip models are efficient video learners. In ECCV, pages 388-404. Springer, 2022. 1, 2, 5, 7, 8" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 324, + 545, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 324, + 545, + 346 + ], + "spans": [ + { + "bbox": [ + 307, + 324, + 545, + 346 + ], + "type": "text", + "content": "[41] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. In NeurIPS, 2023. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 347, + 545, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 347, + 545, + 380 + ], + "spans": [ + { + "bbox": [ + 307, + 347, + 545, + 380 + ], + "type": "text", + "content": "[42] Ze Liu, Jia Ning, Yue Cao, Yixuan Wei, Zheng Zhang, Stephen Lin, and Han Hu. Video swim transformer. In CVPR, pages 3202-3211, 2022. 3, 7, 8" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 382, + 545, + 414 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 382, + 545, + 414 + ], + "spans": [ + { + "bbox": [ + 307, + 382, + 545, + 414 + ], + "type": "text", + "content": "[43] Timo Lüddecke and Alexander S. Ecker. Image segmentation using text and image prompts. In CVPR, pages 7076-7086, 2022. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 416, + 545, + 438 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 416, + 545, + 438 + ], + "spans": [ + { + "bbox": [ + 307, + 416, + 545, + 438 + ], + "type": "text", + "content": "[44] Sachit Menon and Carl Vondrick. Visual classification via description from large language models. In ICLR, 2023. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 440, + 545, + 494 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 440, + 545, + 494 + ], + "spans": [ + { + "bbox": [ + 307, + 440, + 545, + 494 + ], + "type": "text", + "content": "[45] Antoine Miech, Dimitri Zhukov, Jean-Baptiste Alayrac, Makarand Tapaswi, Ivan Laptev, and Josef Sivic. Howto100m: Learning a text-video embedding by watching hundred million narrated video clips. In ICCV, pages 2630-2640, 2019. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 496, + 545, + 540 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 496, + 545, + 540 + ], + "spans": [ + { + "bbox": [ + 307, + 496, + 545, + 540 + ], + "type": "text", + "content": "[46] Antoine Miech, Jean-Baptiste Alayrac, Lucas Smaira, Ivan Laptev, Josef Sivic, and Andrew Zisserman. End-to-end learning of visual representations from uncurated instructional videos. In CVPR, pages 9879-9889, 2020. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 542, + 545, + 585 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 542, + 545, + 585 + ], + "spans": [ + { + "bbox": [ + 307, + 542, + 545, + 585 + ], + "type": "text", + "content": "[47] Liliane Momeni, Mathilde Caron, Arsha Nagrani, Andrew Zisserman, and Cordelia Schmid. Verbs in action: Improving verb understanding in video-language models. ICCV, 2023. 2, 3, 4" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 587, + 545, + 620 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 587, + 545, + 620 + ], + "spans": [ + { + "bbox": [ + 307, + 587, + 545, + 620 + ], + "type": "text", + "content": "[48] Daniel Neimark, Omri Bar, Maya Zohar, and Dotan Asselmann. Video transformer network. In ICCV, pages 3163-3172, 2021. 3" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 622, + 545, + 677 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 622, + 545, + 677 + ], + "spans": [ + { + "bbox": [ + 307, + 622, + 545, + 677 + ], + "type": "text", + "content": "[49] Bolin Ni, Houwen Peng, Minghao Chen, Songyang Zhang, Gaofeng Meng, Jianlong Fu, Shiming Xiang, and Haibin Ling. Expanding language-image pretrained models for general video recognition. In ECCV, pages 1-18. Springer, 2022. 1, 2, 5, 8" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 679, + 492, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 679, + 492, + 689 + ], + "spans": [ + { + "bbox": [ + 307, + 679, + 492, + 689 + ], + "type": "text", + "content": "[50] OpenAI. Gpt-4 technical report, 2023. 1, 2, 4" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 307, + 691, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 691, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 691, + 545, + 713 + ], + "type": "text", + "content": "[51] TB OpenAI. Chatgpt: Optimizing language models for dialogue. OpenAI, 2022. 3" + } + ] + } + ], + "index": 31 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18513" + } + ] + } + ], + "index": 33 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "type": "text", + "content": "[52] Junting Pan, Ziyi Lin, Xiatian Zhu, Jing Shao, and Hongsheng Li. Parameter-efficient image-to-video transfer learning. arXiv e-prints, pages arXiv-2206, 2022. 1, 2, 7, 8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 107, + 287, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 107, + 287, + 150 + ], + "spans": [ + { + "bbox": [ + 49, + 107, + 287, + 150 + ], + "type": "text", + "content": "[53] Tian Pan, Yibing Song, Tianyu Yang, Wenhao Jiang, and Wei Liu. Videomoco: Contrastive video representation learning with temporally adversarial examples. In CVPR, pages 11205-11214, 2021. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 152, + 287, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 152, + 287, + 184 + ], + "spans": [ + { + "bbox": [ + 49, + 152, + 287, + 184 + ], + "type": "text", + "content": "[54] Jungin Park, Jiyoung Lee, and Kwanghoon Sohn. Dual-path adaptation from image to video transformers. CVPR, 2023. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 186, + 287, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 186, + 287, + 239 + ], + "spans": [ + { + "bbox": [ + 49, + 186, + 287, + 239 + ], + "type": "text", + "content": "[55] Mandela Patrick, Dylan Campbell, Yuki Asano, Ishan Misra, Florian Metze, Christoph Feichtenhofer, Andrea Vedaldi, and João F Henriques. Keeping your eye on the ball: Trajectory attention in video transformers. NeurIPS, 34:12493-12506, 2021. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 49, + 241, + 287, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 241, + 287, + 274 + ], + "spans": [ + { + "bbox": [ + 49, + 241, + 287, + 274 + ], + "type": "text", + "content": "[56] Sarah M. Pratt, Rosanne Liu, and Ali Farhadi. What does a platypus look like? generating customized prompts for zero-shot image classification. ICCV, 2022. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 49, + 276, + 287, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 276, + 287, + 319 + ], + "spans": [ + { + "bbox": [ + 49, + 276, + 287, + 319 + ], + "type": "text", + "content": "[57] Rui Qian, Tianjian Meng, Boqing Gong, Ming-Hsuan Yang, Huisheng Wang, Serge Belongie, and Yin Cui. Spatiotemporal contrastive video representation learning. In CVPR, pages 6964-6974, 2021. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 49, + 320, + 287, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 320, + 287, + 364 + ], + "spans": [ + { + "bbox": [ + 49, + 320, + 287, + 364 + ], + "type": "text", + "content": "[58] Zhiwu Qing, Shiwei Zhang, Ziyuan Huang, Yingya Zhang, Changxin Gao, Deli Zhao, and Nong Sang. Disentangling spatial and temporal learning for efficient image-to-video transfer learning. ICCV, 2023. 1, 2, 5, 7, 8" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 49, + 365, + 287, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 365, + 287, + 398 + ], + "spans": [ + { + "bbox": [ + 49, + 365, + 287, + 398 + ], + "type": "text", + "content": "[59] Zhaofan Qiu, Ting Yao, and Tao Mei. Learning spatiotemporal representation with pseudo-3d residual networks. In ICCV, pages 5533–5541, 2017. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 49, + 399, + 287, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 399, + 287, + 464 + ], + "spans": [ + { + "bbox": [ + 49, + 399, + 287, + 464 + ], + "type": "text", + "content": "[60] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, pages 8748-8763. PMLR, 2021. 1, 2, 3, 4, 5, 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 49, + 466, + 287, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 466, + 287, + 498 + ], + "spans": [ + { + "bbox": [ + 49, + 466, + 287, + 498 + ], + "type": "text", + "content": "[61] Kanchana Ranasinghe and Michael S. Ryoo. Language-based action concept spaces improve video self-supervised learning. CoRR, abs/2307.10922, 2023. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 49, + 500, + 287, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 500, + 287, + 542 + ], + "spans": [ + { + "bbox": [ + 49, + 500, + 287, + 542 + ], + "type": "text", + "content": "[62] Michael Ryoo, AJ Piergiovanni, Anurag Arnab, Mostafa Dehghani, and Anelia Angelova. Tokenlearner: Adaptive space-time tokenization for videos. *NeurIPS*, 34:12786-12797, 2021. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 49, + 544, + 287, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 544, + 287, + 588 + ], + "spans": [ + { + "bbox": [ + 49, + 544, + 287, + 588 + ], + "type": "text", + "content": "[63] Laura Sevilla-Lara, Shengxin Zha, Zhicheng Yan, Vedanuj Goswami, Matt Feiszli, and Lorenzo Torresani. Only time can tell: Discovering temporal data for temporal modeling. In WACV, pages 535-544, 2021. 2, 4" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 49, + 590, + 287, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 590, + 287, + 622 + ], + "spans": [ + { + "bbox": [ + 49, + 590, + 287, + 622 + ], + "type": "text", + "content": "[64] Karen Simonyan and Andrew Zisserman. Two-stream convolutional networks for action recognition in videos. NeurIPS, 27, 2014. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 49, + 624, + 287, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 624, + 287, + 656 + ], + "spans": [ + { + "bbox": [ + 49, + 624, + 287, + 656 + ], + "type": "text", + "content": "[65] Khurram Soomro, Amir Roshan Zamir, and Mubarak Shah. Ucf101: A dataset of 101 human actions classes from videos in the wild. arXiv preprint arXiv:1212.0402, 2012. 2, 5, 8" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 49, + 658, + 287, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 658, + 287, + 690 + ], + "spans": [ + { + "bbox": [ + 49, + 658, + 287, + 690 + ], + "type": "text", + "content": "[66] Zhan Tong, Yibing Song, Jue Wang, and Limin Wang. Videomae: Masked autoencoders are data-efficient learners for self-supervised video pre-training. NeurIPS, 2022. 3, 5" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 49, + 692, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 692, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 49, + 692, + 287, + 713 + ], + "type": "text", + "content": "[67] Du Tran, Lubomir Bourdev, Rob Fergus, Lorenzo Torresani, and Manohar Paluri. Learning spatiotemporal features" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 32, + "blocks": [ + { + "bbox": [ + 327, + 73, + 545, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 545, + 94 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 545, + 94 + ], + "type": "text", + "content": "with 3d convolutional networks. In ICCV, pages 4489-4497, 2015. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 97, + 545, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 97, + 545, + 140 + ], + "spans": [ + { + "bbox": [ + 308, + 97, + 545, + 140 + ], + "type": "text", + "content": "[68] Du Tran, Heng Wang, Lorenzo Torresani, Jamie Ray, Yann LeCun, and Manohar Paluri. A closer look at spatiotemporal convolutions for action recognition. In CVPR, pages 6450-6459, 2018." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 143, + 545, + 176 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 143, + 545, + 176 + ], + "spans": [ + { + "bbox": [ + 308, + 143, + 545, + 176 + ], + "type": "text", + "content": "[69] Du Tran, Heng Wang, Lorenzo Torresani, and Matt Feiszli. Video classification with channel-separated convolutional networks. In ICCV, pages 5552-5561, 2019. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 177, + 545, + 220 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 177, + 545, + 220 + ], + "spans": [ + { + "bbox": [ + 308, + 177, + 545, + 220 + ], + "type": "text", + "content": "[70] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In NeurIPS, pages 5998-6008, 2017. 4" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 223, + 545, + 255 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 223, + 545, + 255 + ], + "spans": [ + { + "bbox": [ + 308, + 223, + 545, + 255 + ], + "type": "text", + "content": "[71] Heng Wang, Du Tran, Lorenzo Torresani, and Matt Feiszli. Video modeling with correlation networks. In CVPR, pages 352-361, 2020. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 258, + 545, + 301 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 258, + 545, + 301 + ], + "spans": [ + { + "bbox": [ + 308, + 258, + 545, + 301 + ], + "type": "text", + "content": "[72] Junke Wang, Dongdong Chen, Chong Luo, Xiyang Dai, Lu Yuan, Zuxuan Wu, and Yu-Gang Jiang. Chatvideo: A tracklet-centric multimodal and versatile video understanding system. CoRR, abs/2304.14407, 2023. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 304, + 545, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 304, + 545, + 346 + ], + "spans": [ + { + "bbox": [ + 308, + 304, + 545, + 346 + ], + "type": "text", + "content": "[73] Limin Wang, Yuanjun Xiong, Zhe Wang, Yu Qiao, Dahua Lin, Xiaou Tang, and Luc Van Gool. Temporal segment networks for action recognition in videos. TPAMI, 41(11): 2740-2755, 2018. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 350, + 545, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 350, + 545, + 392 + ], + "spans": [ + { + "bbox": [ + 308, + 350, + 545, + 392 + ], + "type": "text", + "content": "[74] Limin Wang, Bingkun Huang, Zhiyu Zhao, Zhan Tong, Yi-nan He, Yi Wang, Yali Wang, and Yu Qiao. Videomae V2: scaling video masked autoencoders with dual masking. In CVPR, pages 14549-14560, 2023. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 395, + 545, + 427 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 395, + 545, + 427 + ], + "spans": [ + { + "bbox": [ + 308, + 395, + 545, + 427 + ], + "type": "text", + "content": "[75] Mengmeng Wang, Jiazheng Xing, and Yong Liu. Actionclip: A new paradigm for video action recognition. arXiv preprint arXiv:2109.08472, 2021. 1, 2, 8" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 430, + 545, + 484 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 430, + 545, + 484 + ], + "spans": [ + { + "bbox": [ + 308, + 430, + 545, + 484 + ], + "type": "text", + "content": "[76] Peng Wang, An Yang, Rui Men, Junyang Lin, Shuai Bai, Zhikang Li, Jianxin Ma, Chang Zhou, Jingren Zhou, and Hongxia Yang. OFA: unifying architectures, tasks, and modalities through a simple sequence-to-sequence learning framework. In ICML, pages 23318-23340, 2022. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 486, + 545, + 529 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 486, + 545, + 529 + ], + "spans": [ + { + "bbox": [ + 308, + 486, + 545, + 529 + ], + "type": "text", + "content": "[77] Rui Wang, Dongdong Chen, Zuxuan Wu, Yinpeng Chen, Xiyang Dai, Mengchen Liu, Yu-Gang Jiang, Luowei Zhou, and Lu Yuan. Bevt: Bert pretraining of video transformers. In CVPR, pages 14733-14743, 2022. 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 532, + 545, + 575 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 532, + 545, + 575 + ], + "spans": [ + { + "bbox": [ + 308, + 532, + 545, + 575 + ], + "type": "text", + "content": "[78] Syed Talal Wasim, Muzammal Naseer, Salman H. Khan, Fahad Shahbaz Khan, and Mubarak Shah. Vita-clip: Video and text adaptive CLIP via multimodal prompting. CVPR, 2023. 1, 2, 8" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 578, + 545, + 620 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 578, + 545, + 620 + ], + "spans": [ + { + "bbox": [ + 308, + 578, + 545, + 620 + ], + "type": "text", + "content": "[79] Chen Wei, Haoqi Fan, Saining Xie, Chao-Yuan Wu, Alan Yuille, and Christoph Feichtenhofer. Masked feature prediction for self-supervised visual pre-training. In CVPR, pages 14668-14678, 2022. 3" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 308, + 624, + 545, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 624, + 545, + 657 + ], + "spans": [ + { + "bbox": [ + 308, + 624, + 545, + 657 + ], + "type": "text", + "content": "[80] Wenhao Wu, Zhun Sun, and Wanli Ouyang. Revisiting classifier: Transferring vision-language models for video recognition. In AAAI, pages 2847-2855, 2023. 1" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 308, + 658, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 658, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 308, + 658, + 545, + 712 + ], + "type": "text", + "content": "[81] Hu Xu, Gargi Ghosh, Po-Yao Huang, Dmytro Okhonko, Armen Aghajanyan, Florian Metze, Luke Zettlemoyer, and Christoph Feichtenhofer. Videoclip: Contrastive pre-training for zero-shot video-text understanding. In EMNLP, pages 6787-6800, 2021. 2" + } + ] + } + ], + "index": 31 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "text", + "content": "18514" + } + ] + } + ], + "index": 33 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 552 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "text", + "content": "[82] Jiarui Xu, Shalini De Mello, Sifei Liu, Wonmin Byeon, Thomas M. Breuel, Jan Kautz, and Xiaolong Wang. Groupvit: Semantic segmentation emerges from text supervision. In CVPR, pages 18113-18123, 2022. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 118, + 286, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 118, + 286, + 161 + ], + "spans": [ + { + "bbox": [ + 49, + 118, + 286, + 161 + ], + "type": "text", + "content": "[83] An Yan, Yu Wang, Yiwu Zhong, Chengyu Dong, Zexue He, Yujie Lu, William Wang, Jingbo Shang, and Julian J. McAuley. Learning concise and descriptive attributes for visual recognition. ICCV, 2023. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 163, + 286, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 163, + 286, + 205 + ], + "spans": [ + { + "bbox": [ + 49, + 163, + 286, + 205 + ], + "type": "text", + "content": "[84] Shen Yan, Xuehan Xiong, Anurag Arnab, Zhichao Lu, Mi Zhang, Chen Sun, and Cordelia Schmid. Multiview transformers for video recognition. In CVPR, pages 3333-3343, 2022. 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 208, + 286, + 261 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 208, + 286, + 261 + ], + "spans": [ + { + "bbox": [ + 49, + 208, + 286, + 261 + ], + "type": "text", + "content": "[85] Yue Yang, Artemis Panagopoulou, Shenghao Zhou, Daniel Jin, Chris Callison-Burch, and Mark Yatskar. Language in a bottle: Language model guided concept bottlenecks for interpretable image classification. In CVPR, pages 19187-19197, 2023. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 49, + 263, + 286, + 317 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 263, + 286, + 317 + ], + "spans": [ + { + "bbox": [ + 49, + 263, + 286, + 317 + ], + "type": "text", + "content": "[86] Zhengyuan Yang, Linjie Li, Jianfeng Wang, Kevin Lin, Ehsan Azarnasab, Faisal Ahmed, Zicheng Liu, Ce Liu, Michael Zeng, and Lijuan Wang. MM-REACT: prompting chatgpt for multimodal reasoning and action. CoRR, abs/2303.11381, 2023. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 49, + 319, + 286, + 362 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 319, + 286, + 362 + ], + "spans": [ + { + "bbox": [ + 49, + 319, + 286, + 362 + ], + "type": "text", + "content": "[87] Jiahui Yu, Zirui Wang, Vijay Vasudevan, Legg Yeung, Mojtaba Seyedhosseini, and Yonghui Wu. Coca: Contrastive captioners are image-text foundation models. arXiv preprint arXiv:2205.01917, 2022. 1, 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 49, + 365, + 286, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 365, + 286, + 418 + ], + "spans": [ + { + "bbox": [ + 49, + 365, + 286, + 418 + ], + "type": "text", + "content": "[88] Lu Yuan, Dongdong Chen, Yi-Ling Chen, Noel Codella, Xiyang Dai, Jianfeng Gao, Houdong Hu, Xuedong Huang, Boxin Li, Chunyuan Li, et al. Florence: A new foundation model for computer vision. arXiv preprint arXiv:2111.11432, 2021. 1, 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 49, + 420, + 286, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 420, + 286, + 453 + ], + "spans": [ + { + "bbox": [ + 49, + 420, + 286, + 453 + ], + "type": "text", + "content": "[89] Yue Zhao, Ishan Misra, Philipp Krahenbuhl, and Rohit Girdhar. Learning video representations from large language models. In CVPR, pages 6586-6597. IEEE, 2023. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 49, + 454, + 286, + 507 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 454, + 286, + 507 + ], + "spans": [ + { + "bbox": [ + 49, + 454, + 286, + 507 + ], + "type": "text", + "content": "[90] Yiwu Zhong, Jianwei Yang, Pengchuan Zhang, Chunyuan Li, Noel Codella, Liunian Harold Li, Luowei Zhou, Xiyang Dai, Lu Yuan, Yin Li, and Jianfeng Gao. Regionclip: Region-based language-image pretraining. In CVPR, pages 16772-16782, 2022. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 49, + 510, + 286, + 552 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 510, + 286, + 552 + ], + "spans": [ + { + "bbox": [ + 49, + 510, + 286, + 552 + ], + "type": "text", + "content": "[91] Deyao Zhu, Jun Chen, Xiaoqian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models. CoRR, abs/2304.10592, 2023. 3" + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "18515" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file